Groups | Search | Server Info | Keyboard shortcuts | Login | Register [http] [https] [nntp] [nntps]
Groups > linux.kernel > #1193568 > unrolled thread
| Started by | Ashutosh Dixit <ashutosh.dixit@intel.com> |
|---|---|
| First post | 2015-07-28 01:30 +0200 |
| Last post | 2015-07-28 02:00 +0200 |
| Articles | 19 — 4 participants |
Back to article view | Back to linux.kernel
[PATCH char-misc-next 00/19] misc: mic: Enable COSM and remaining SCIF functionality Ashutosh Dixit <ashutosh.dixit@intel.com> - 2015-07-28 01:30 +0200
[PATCH char-misc-next 01/19] misc: mic: SCIF poll Ashutosh Dixit <ashutosh.dixit@intel.com> - 2015-07-28 01:30 +0200
[PATCH char-misc-next 02/19] misc: mic: Add support for kernel mode SCIF clients Ashutosh Dixit <ashutosh.dixit@intel.com> - 2015-07-28 01:30 +0200
[PATCH char-misc-next 19/19] misc: mic: SCIF RMA nodeqp and minor miscellaneous changes Ashutosh Dixit <ashutosh.dixit@intel.com> - 2015-07-28 01:40 +0200
[PATCH char-misc-next 18/19] misc: mic: SCIF fence Ashutosh Dixit <ashutosh.dixit@intel.com> - 2015-07-28 01:40 +0200
[PATCH char-misc-next 12/19] misc: mic: SCIF RMA header file and IOCTL changes Ashutosh Dixit <ashutosh.dixit@intel.com> - 2015-07-28 01:40 +0200
[PATCH char-misc-next 10/19] lib: convert iova.c into a library Ashutosh Dixit <ashutosh.dixit@intel.com> - 2015-07-28 01:40 +0200
Re: [PATCH char-misc-next 10/19] lib: convert iova.c into a library Joerg Roedel <joro@8bytes.org> - 2015-07-28 12:10 +0200
Re: [PATCH char-misc-next 10/19] lib: convert iova.c into a library Robin Murphy <robin.murphy@arm.com> - 2015-07-28 12:50 +0200
Re: [PATCH char-misc-next 10/19] lib: convert iova.c into a library Andrew Morton <akpm@linux-foundation.org> - 2015-07-28 22:50 +0200
[PATCH char-misc-next 11/19] dma: Add support to program MIC x100 status descriptiors Ashutosh Dixit <ashutosh.dixit@intel.com> - 2015-07-28 01:40 +0200
[PATCH char-misc-next 16/19] misc: mic: SCIF remote memory map/unmap interface Ashutosh Dixit <ashutosh.dixit@intel.com> - 2015-07-28 01:40 +0200
[PATCH char-misc-next 15/19] misc: mic: SCIF RMA list operations Ashutosh Dixit <ashutosh.dixit@intel.com> - 2015-07-28 01:40 +0200
[PATCH char-misc-next 04/19] misc: mic: Coprocessor State Management (COSM) driver Ashutosh Dixit <ashutosh.dixit@intel.com> - 2015-07-28 01:40 +0200
[PATCH char-misc-next 13/19] misc: mic: SCIF RMA header file Ashutosh Dixit <ashutosh.dixit@intel.com> - 2015-07-28 01:40 +0200
[PATCH char-misc-next 05/19] misc: mic: COSM SCIF server Ashutosh Dixit <ashutosh.dixit@intel.com> - 2015-07-28 01:40 +0200
[PATCH char-misc-next 03/19] misc: mic: MIC COSM bus Ashutosh Dixit <ashutosh.dixit@intel.com> - 2015-07-28 01:40 +0200
[PATCH char-misc-next 08/19] misc: mic: Remove COSM functionality from the MIC card driver Ashutosh Dixit <ashutosh.dixit@intel.com> - 2015-07-28 01:50 +0200
[PATCH char-misc-next 06/19] misc: mic: COSM client driver Ashutosh Dixit <ashutosh.dixit@intel.com> - 2015-07-28 02:00 +0200
| From | Ashutosh Dixit <ashutosh.dixit@intel.com> |
|---|---|
| Date | 2015-07-28 01:30 +0200 |
| Subject | [PATCH char-misc-next 00/19] misc: mic: Enable COSM and remaining SCIF functionality |
| Message-ID | <pR00h-8jZ-7@gated-at.bofh.it> |
The Symmetric Communication Interface (SCIF) API was accepted upstream
during the v4.2 merge window @ https://lkml.org/lkml/2015/3/30/865.
This patch series completes the implementation of the SCIF API and
also introduces the MIC Coprocessor State Management (COSM) drivers.
The first patch introduces SCIF poll which allows user and kernel mode
clients to block for events on a SCIF endpoint. The next patch
introduces support for SCIF kernel mode clients. SCIF clients are
probed when new SCIF nodes come online. Similarly the client remove
method is called as SCIF nodes go offline. The SCIF client framework
makes use of the SCIF peer bus infrastructure submitted previously.
The next seven patches are a re-factoring of functionality for boot,
shutdown and reset of MIC cards. This functionality, called COSM, was
previously present in the MIC host driver but has now been re-factored
into a separate driver so that it can be shared across multiple
generations of MIC products. The MIC sysfs interface documented
earlier has been updated. Further, the COSM host driver communicates
with a COSM client on the card using SCIF. COSM is therefore the first
kernel mode SCIF client and demonstrates the use of a subset of the
SCIF API.
Patch 10 converts the iova implementation used by INTEL_IOMMU into a
library without any implementation changes so that it can be reused by
SCIF to generate registered offsets.
Patch 11 onward completes the SCIF implementation by enabling the
following SCIF Remote Memory Access (RMA) functionality:
a. Memory registration to pin and unpin pages
b. Remote Memory mapping for low latency CPU accesses
c. Remote DMA (RDMA) for high bandwidth DMA transfers
d. Fence mechanisms to synchronize RDMAs
Documentation/mic/scif_overview.txt contains more information about
SCIF. This patch series is divided into 19 patches as follows:
1) SCIF poll
2) Support for kernel mode SCIF clients
3) MIC COSM bus
4) Coprocessor State Management (COSM) driver
5) COSM SCIF server
6) COSM client driver
7) Remove COSM functionality from the MIC host driver
8) Remove COSM functionality from the MIC card driver
9) Update MIC host daemon with COSM changes
10) Convert iova.c into a library
11) Add support to program MIC x100 status descriptors
12) SCIF RMA header file and IOCTL changes
13) SCIF RMA header file
14) SCIF memory registration and unregistration
15) SCIF RMA list operations
16) SCIF remote memory map/unmap interface
17) SCIF DMA and CPU copy interface
18) SCIF fence
19) SCIF RMA nodeqp and minor miscellaneous changes
These patches have also been scanned by Fengguang Wu's 0-day
infrastructure and no issues have been reported. The IOVA patch (patch
10) is being sent to char-misc because of the dependency between that
patch and SCIF patches 12 through 19.
Ashutosh Dixit (9):
misc: mic: SCIF poll
misc: mic: Add support for kernel mode SCIF clients
misc: mic: MIC COSM bus
misc: mic: Coprocessor State Management (COSM) driver
misc: mic: COSM SCIF server
misc: mic: COSM client driver
misc: mic: Remove COSM functionality from the MIC host driver
misc: mic: Remove COSM functionality from the MIC card driver
misc: mic: Update MIC host daemon with COSM changes
Harish Chegondi (1):
lib: convert iova.c into a library
Siva Yerramreddy (1):
dma: Add support to program MIC x100 status descriptiors
Sudeep Dutt (8):
misc: mic: SCIF RMA header file and IOCTL changes
misc: mic: SCIF RMA header file
misc: mic: SCIF memory registration and unregistration
misc: mic: SCIF RMA list operations
misc: mic: SCIF remote memory map/unmap interface
misc: mic: SCIF DMA and CPU copy interface
misc: mic: SCIF fence
misc: mic: SCIF RMA nodeqp and minor miscellaneous changes
Documentation/ABI/testing/sysfs-class-mic.txt | 29 +-
Documentation/mic/mic_overview.txt | 31 +-
Documentation/mic/mpssd/mpss | 4 +-
Documentation/mic/mpssd/mpssd.c | 362 +++--
Documentation/mic/mpssd/mpssd.h | 1 +
drivers/dma/mic_x100_dma.c | 44 +-
drivers/iommu/Kconfig | 5 +-
drivers/iommu/Makefile | 1 -
drivers/misc/mic/Kconfig | 23 +-
drivers/misc/mic/Makefile | 2 +
drivers/misc/mic/bus/Makefile | 1 +
drivers/misc/mic/bus/cosm_bus.c | 141 ++
drivers/misc/mic/bus/cosm_bus.h | 134 ++
drivers/misc/mic/bus/mic_bus.c | 22 +-
drivers/misc/mic/bus/scif_bus.c | 7 +-
drivers/misc/mic/bus/scif_bus.h | 6 +-
drivers/misc/mic/card/mic_device.c | 88 +-
drivers/misc/mic/card/mic_x100.c | 2 +-
drivers/misc/mic/common/mic_dev.h | 13 +
drivers/misc/mic/cosm/Makefile | 10 +
drivers/misc/mic/cosm/cosm_debugfs.c | 155 ++
drivers/misc/mic/cosm/cosm_main.c | 388 +++++
drivers/misc/mic/cosm/cosm_main.h | 70 +
drivers/misc/mic/cosm/cosm_scif_server.c | 405 +++++
drivers/misc/mic/cosm/cosm_sysfs.c | 461 ++++++
drivers/misc/mic/cosm_client/Makefile | 7 +
drivers/misc/mic/cosm_client/cosm_scif_client.c | 275 ++++
drivers/misc/mic/host/Makefile | 1 -
drivers/misc/mic/host/mic_boot.c | 317 +---
drivers/misc/mic/host/mic_debugfs.c | 114 +-
drivers/misc/mic/host/mic_device.h | 88 +-
drivers/misc/mic/host/mic_fops.c | 4 +-
drivers/misc/mic/host/mic_intr.c | 46 +-
drivers/misc/mic/host/mic_main.c | 223 +--
drivers/misc/mic/host/mic_smpt.c | 30 +-
drivers/misc/mic/host/mic_sysfs.c | 459 ------
drivers/misc/mic/host/mic_virtio.c | 17 +-
drivers/misc/mic/host/mic_virtio.h | 2 +-
drivers/misc/mic/host/mic_x100.c | 46 +-
drivers/misc/mic/scif/Makefile | 5 +
drivers/misc/mic/scif/scif_api.c | 223 ++-
drivers/misc/mic/scif/scif_debugfs.c | 85 +-
drivers/misc/mic/scif/scif_dma.c | 1979 +++++++++++++++++++++++
drivers/misc/mic/scif/scif_epd.c | 9 +-
drivers/misc/mic/scif/scif_epd.h | 50 +
drivers/misc/mic/scif/scif_fd.c | 168 ++
drivers/misc/mic/scif/scif_fence.c | 773 +++++++++
drivers/misc/mic/scif/scif_main.c | 107 +-
drivers/misc/mic/scif/scif_main.h | 35 +-
drivers/misc/mic/scif/scif_map.h | 25 +-
drivers/misc/mic/scif/scif_mmap.c | 711 ++++++++
drivers/misc/mic/scif/scif_nm.c | 20 +-
drivers/misc/mic/scif/scif_nodeqp.c | 118 +-
drivers/misc/mic/scif/scif_nodeqp.h | 42 +-
drivers/misc/mic/scif/scif_peer_bus.c | 179 +-
drivers/misc/mic/scif/scif_peer_bus.h | 42 +-
drivers/misc/mic/scif/scif_rma.c | 1761 ++++++++++++++++++++
drivers/misc/mic/scif/scif_rma.h | 464 ++++++
drivers/misc/mic/scif/scif_rma_list.c | 291 ++++
drivers/misc/mic/scif/scif_rma_list.h | 57 +
include/linux/mic_bus.h | 3 +-
include/linux/scif.h | 364 ++++-
include/uapi/linux/mic_common.h | 16 +-
include/uapi/linux/scif_ioctl.h | 85 +
lib/Kconfig | 6 +
lib/Makefile | 2 +
{drivers/iommu => lib}/iova.c | 9 +
67 files changed, 9985 insertions(+), 1678 deletions(-)
create mode 100644 drivers/misc/mic/bus/cosm_bus.c
create mode 100644 drivers/misc/mic/bus/cosm_bus.h
create mode 100644 drivers/misc/mic/cosm/Makefile
create mode 100644 drivers/misc/mic/cosm/cosm_debugfs.c
create mode 100644 drivers/misc/mic/cosm/cosm_main.c
create mode 100644 drivers/misc/mic/cosm/cosm_main.h
create mode 100644 drivers/misc/mic/cosm/cosm_scif_server.c
create mode 100644 drivers/misc/mic/cosm/cosm_sysfs.c
create mode 100644 drivers/misc/mic/cosm_client/Makefile
create mode 100644 drivers/misc/mic/cosm_client/cosm_scif_client.c
delete mode 100644 drivers/misc/mic/host/mic_sysfs.c
create mode 100644 drivers/misc/mic/scif/scif_dma.c
create mode 100644 drivers/misc/mic/scif/scif_fence.c
create mode 100644 drivers/misc/mic/scif/scif_mmap.c
create mode 100644 drivers/misc/mic/scif/scif_rma.c
create mode 100644 drivers/misc/mic/scif/scif_rma.h
create mode 100644 drivers/misc/mic/scif/scif_rma_list.c
create mode 100644 drivers/misc/mic/scif/scif_rma_list.h
rename {drivers/iommu => lib}/iova.c (97%)
--
2.0.0.rc3.2.g998f840
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
[toc] | [next] | [standalone]
| From | Ashutosh Dixit <ashutosh.dixit@intel.com> |
|---|---|
| Date | 2015-07-28 01:30 +0200 |
| Subject | [PATCH char-misc-next 01/19] misc: mic: SCIF poll |
| Message-ID | <pR00i-8jZ-21@gated-at.bofh.it> |
| In reply to | #1193568 |
SCIF poll allows both user and kernel mode clients to wait on
events on a SCIF endpoint. These events include availability of
space or data in the SCIF ring buffer, availability of connection
requests on a listening endpoint and completion of connections
when using async connects.
Reviewed-by: Nikhil Rao <nikhil.rao@intel.com>
Reviewed-by: Sudeep Dutt <sudeep.dutt@intel.com>
Signed-off-by: Ashutosh Dixit <ashutosh.dixit@intel.com>
---
drivers/misc/mic/scif/scif_api.c | 158 +++++++++++++++++++++++++++++++++++++-
drivers/misc/mic/scif/scif_epd.h | 22 ++++++
drivers/misc/mic/scif/scif_fd.c | 9 +++
drivers/misc/mic/scif/scif_main.h | 2 +
include/linux/scif.h | 74 ++++++++++++++++++
5 files changed, 264 insertions(+), 1 deletion(-)
diff --git a/drivers/misc/mic/scif/scif_api.c b/drivers/misc/mic/scif/scif_api.c
index f39d313..bf2d70f 100644
--- a/drivers/misc/mic/scif/scif_api.c
+++ b/drivers/misc/mic/scif/scif_api.c
@@ -37,9 +37,21 @@ enum conn_async_state {
ASYNC_CONN_FLUSH_WORK /* async work flush in progress */
};
+/*
+ * File operations for anonymous inode file associated with a SCIF endpoint,
+ * used in kernel mode SCIF poll. Kernel mode SCIF poll calls portions of the
+ * poll API in the kernel and these take in a struct file *. Since a struct
+ * file is not available to kernel mode SCIF, it uses an anonymous file for
+ * this purpose.
+ */
+const struct file_operations scif_anon_fops = {
+ .owner = THIS_MODULE,
+};
+
scif_epd_t scif_open(void)
{
struct scif_endpt *ep;
+ int err;
might_sleep();
ep = kzalloc(sizeof(*ep), GFP_KERNEL);
@@ -50,6 +62,10 @@ scif_epd_t scif_open(void)
if (!ep->qp_info.qp)
goto err_qp_alloc;
+ err = scif_anon_inode_getfile(ep);
+ if (err)
+ goto err_anon_inode;
+
spin_lock_init(&ep->lock);
mutex_init(&ep->sendlock);
mutex_init(&ep->recvlock);
@@ -59,6 +75,8 @@ scif_epd_t scif_open(void)
"SCIFAPI open: ep %p success\n", ep);
return ep;
+err_anon_inode:
+ kfree(ep->qp_info.qp);
err_qp_alloc:
kfree(ep);
err_ep_alloc:
@@ -279,6 +297,7 @@ int scif_close(scif_epd_t epd)
}
}
scif_put_port(ep->port.port);
+ scif_anon_inode_fput(ep);
scif_teardown_ep(ep);
scif_add_epd_to_zombie_list(ep, !SCIF_EPLOCK_HELD);
return 0;
@@ -558,8 +577,10 @@ void scif_conn_handler(struct work_struct *work)
list_del(&ep->conn_list);
}
spin_unlock(&scif_info.nb_connect_lock);
- if (ep)
+ if (ep) {
ep->conn_err = scif_conn_func(ep);
+ wake_up_interruptible(&ep->conn_pend_wq);
+ }
} while (ep);
}
@@ -660,6 +681,7 @@ int __scif_connect(scif_epd_t epd, struct scif_port_id *dst, bool non_block)
ep->remote_dev = &scif_dev[dst->node];
ep->qp_info.qp->magic = SCIFEP_MAGIC;
if (ep->conn_async_state == ASYNC_CONN_INPROGRESS) {
+ init_waitqueue_head(&ep->conn_pend_wq);
spin_lock(&scif_info.nb_connect_lock);
list_add_tail(&ep->conn_list, &scif_info.nb_connect_list);
spin_unlock(&scif_info.nb_connect_lock);
@@ -788,6 +810,10 @@ retry_connection:
goto scif_accept_error_qpalloc;
}
+ err = scif_anon_inode_getfile(cep);
+ if (err)
+ goto scif_accept_error_anon_inode;
+
cep->qp_info.qp->magic = SCIFEP_MAGIC;
spdev = scif_get_peer_dev(cep->remote_dev);
if (IS_ERR(spdev)) {
@@ -858,6 +884,8 @@ retry:
spin_unlock(&cep->lock);
return 0;
scif_accept_error_map:
+ scif_anon_inode_fput(cep);
+scif_accept_error_anon_inode:
scif_teardown_ep(cep);
scif_accept_error_qpalloc:
kfree(cep);
@@ -1247,6 +1275,134 @@ int scif_recv(scif_epd_t epd, void *msg, int len, int flags)
}
EXPORT_SYMBOL_GPL(scif_recv);
+static inline void _scif_poll_wait(struct file *f, wait_queue_head_t *wq,
+ poll_table *p, struct scif_endpt *ep)
+{
+ /*
+ * Because poll_wait makes a GFP_KERNEL allocation, give up the lock
+ * and regrab it afterwards. Because the endpoint state might have
+ * changed while the lock was given up, the state must be checked
+ * again after re-acquiring the lock. The code in __scif_pollfd(..)
+ * does this.
+ */
+ spin_unlock(&ep->lock);
+ poll_wait(f, wq, p);
+ spin_lock(&ep->lock);
+}
+
+unsigned int
+__scif_pollfd(struct file *f, poll_table *wait, struct scif_endpt *ep)
+{
+ unsigned int mask = 0;
+
+ dev_dbg(scif_info.mdev.this_device,
+ "SCIFAPI pollfd: ep %p %s\n", ep, scif_ep_states[ep->state]);
+
+ spin_lock(&ep->lock);
+
+ /* Endpoint is waiting for a non-blocking connect to complete */
+ if (ep->conn_async_state == ASYNC_CONN_INPROGRESS) {
+ _scif_poll_wait(f, &ep->conn_pend_wq, wait, ep);
+ if (ep->conn_async_state == ASYNC_CONN_INPROGRESS) {
+ if (ep->state == SCIFEP_CONNECTED ||
+ ep->state == SCIFEP_DISCONNECTED ||
+ ep->conn_err)
+ mask |= POLLOUT;
+ goto exit;
+ }
+ }
+
+ /* Endpoint is listening for incoming connection requests */
+ if (ep->state == SCIFEP_LISTENING) {
+ _scif_poll_wait(f, &ep->conwq, wait, ep);
+ if (ep->state == SCIFEP_LISTENING) {
+ if (ep->conreqcnt)
+ mask |= POLLIN;
+ goto exit;
+ }
+ }
+
+ /* Endpoint is connected or disconnected */
+ if (ep->state == SCIFEP_CONNECTED || ep->state == SCIFEP_DISCONNECTED) {
+ if (poll_requested_events(wait) & POLLIN)
+ _scif_poll_wait(f, &ep->recvwq, wait, ep);
+ if (poll_requested_events(wait) & POLLOUT)
+ _scif_poll_wait(f, &ep->sendwq, wait, ep);
+ if (ep->state == SCIFEP_CONNECTED ||
+ ep->state == SCIFEP_DISCONNECTED) {
+ /* Data can be read without blocking */
+ if (scif_rb_count(&ep->qp_info.qp->inbound_q, 1))
+ mask |= POLLIN;
+ /* Data can be written without blocking */
+ if (scif_rb_space(&ep->qp_info.qp->outbound_q))
+ mask |= POLLOUT;
+ /* Return POLLHUP if endpoint is disconnected */
+ if (ep->state == SCIFEP_DISCONNECTED)
+ mask |= POLLHUP;
+ goto exit;
+ }
+ }
+
+ /* Return POLLERR if the endpoint is in none of the above states */
+ mask |= POLLERR;
+exit:
+ spin_unlock(&ep->lock);
+ return mask;
+}
+
+/**
+ * scif_poll() - Kernel mode SCIF poll
+ * @ufds: Array of scif_pollepd structures containing the end points
+ * and events to poll on
+ * @nfds: Size of the ufds array
+ * @timeout_msecs: Timeout in msecs, -ve implies infinite timeout
+ *
+ * The code flow in this function is based on do_poll(..) in select.c
+ *
+ * Returns the number of endpoints which have pending events or 0 in
+ * the event of a timeout. If a signal is used for wake up, -EINTR is
+ * returned.
+ */
+int
+scif_poll(struct scif_pollepd *ufds, unsigned int nfds, long timeout_msecs)
+{
+ struct poll_wqueues table;
+ poll_table *pt;
+ int i, mask, count = 0, timed_out = timeout_msecs == 0;
+ u64 timeout = timeout_msecs < 0 ? MAX_SCHEDULE_TIMEOUT
+ : msecs_to_jiffies(timeout_msecs);
+
+ poll_initwait(&table);
+ pt = &table.pt;
+ while (1) {
+ for (i = 0; i < nfds; i++) {
+ pt->_key = ufds[i].events | POLLERR | POLLHUP;
+ mask = __scif_pollfd(ufds[i].epd->anon,
+ pt, ufds[i].epd);
+ mask &= ufds[i].events | POLLERR | POLLHUP;
+ if (mask) {
+ count++;
+ pt->_qproc = NULL;
+ }
+ ufds[i].revents = mask;
+ }
+ pt->_qproc = NULL;
+ if (!count) {
+ count = table.error;
+ if (signal_pending(current))
+ count = -EINTR;
+ }
+ if (count || timed_out)
+ break;
+
+ if (!schedule_timeout_interruptible(timeout))
+ timed_out = 1;
+ }
+ poll_freewait(&table);
+ return count;
+}
+EXPORT_SYMBOL_GPL(scif_poll);
+
int scif_get_node_ids(u16 *nodes, int len, u16 *self)
{
int online = 0;
diff --git a/drivers/misc/mic/scif/scif_epd.h b/drivers/misc/mic/scif/scif_epd.h
index 331322a..4dc4ccd 100644
--- a/drivers/misc/mic/scif/scif_epd.h
+++ b/drivers/misc/mic/scif/scif_epd.h
@@ -96,7 +96,9 @@ struct scif_endpt_qp_info {
* @conn_port: Connection port
* @conn_err: Errors during connection
* @conn_async_state: Async connection
+ * @conn_pend_wq: Used by poll while waiting for incoming connections
* @conn_list: List of async connection requests
+ * @anon: anonymous file for use in kernel mode scif poll
*/
struct scif_endpt {
enum scif_epd_state state;
@@ -125,7 +127,9 @@ struct scif_endpt {
struct scif_port_id conn_port;
int conn_err;
int conn_async_state;
+ wait_queue_head_t conn_pend_wq;
struct list_head conn_list;
+ struct file *anon;
};
static inline int scifdev_alive(struct scif_endpt *ep)
@@ -133,6 +137,22 @@ static inline int scifdev_alive(struct scif_endpt *ep)
return _scifdev_alive(ep->remote_dev);
}
+static inline int scif_anon_inode_getfile(scif_epd_t epd)
+{
+ epd->anon = anon_inode_getfile("scif", &scif_anon_fops, NULL, 0);
+ if (IS_ERR(epd->anon))
+ return PTR_ERR(epd->anon);
+ return 0;
+}
+
+static inline void scif_anon_inode_fput(scif_epd_t epd)
+{
+ if (epd->anon) {
+ fput(epd->anon);
+ epd->anon = NULL;
+ }
+}
+
void scif_cleanup_zombie_epd(void);
void scif_teardown_ep(void *endpt);
void scif_cleanup_ep_qp(struct scif_endpt *ep);
@@ -157,4 +177,6 @@ void scif_clientsend(struct scif_dev *scifdev, struct scifmsg *msg);
void scif_clientrcvd(struct scif_dev *scifdev, struct scifmsg *msg);
int __scif_connect(scif_epd_t epd, struct scif_port_id *dst, bool non_block);
int __scif_flush(scif_epd_t epd);
+unsigned int __scif_pollfd(struct file *f, poll_table *wait,
+ struct scif_endpt *ep);
#endif /* SCIF_EPD_H */
diff --git a/drivers/misc/mic/scif/scif_fd.c b/drivers/misc/mic/scif/scif_fd.c
index eccf7e7..24e47f7 100644
--- a/drivers/misc/mic/scif/scif_fd.c
+++ b/drivers/misc/mic/scif/scif_fd.c
@@ -34,6 +34,13 @@ static int scif_fdclose(struct inode *inode, struct file *f)
return scif_close(priv);
}
+static unsigned int scif_fdpoll(struct file *f, poll_table *wait)
+{
+ struct scif_endpt *priv = f->private_data;
+
+ return __scif_pollfd(f, wait, priv);
+}
+
static int scif_fdflush(struct file *f, fl_owner_t id)
{
struct scif_endpt *ep = f->private_data;
@@ -193,6 +200,7 @@ static long scif_fdioctl(struct file *f, unsigned int cmd, unsigned long arg)
spin_unlock(&scif_info.eplock);
/* Free the resources automatically created from the open. */
+ scif_anon_inode_fput(priv);
scif_teardown_ep(priv);
scif_add_epd_to_zombie_list(priv, !SCIF_EPLOCK_HELD);
f->private_data = newep;
@@ -298,6 +306,7 @@ const struct file_operations scif_fops = {
.open = scif_fdopen,
.release = scif_fdclose,
.unlocked_ioctl = scif_fdioctl,
+ .poll = scif_fdpoll,
.flush = scif_fdflush,
.owner = THIS_MODULE,
};
diff --git a/drivers/misc/mic/scif/scif_main.h b/drivers/misc/mic/scif/scif_main.h
index 580bc63..87c1327 100644
--- a/drivers/misc/mic/scif/scif_main.h
+++ b/drivers/misc/mic/scif/scif_main.h
@@ -22,6 +22,7 @@
#include <linux/pci.h>
#include <linux/miscdevice.h>
#include <linux/dmaengine.h>
+#include <linux/anon_inodes.h>
#include <linux/file.h>
#include <linux/scif.h>
@@ -184,6 +185,7 @@ extern struct scif_info scif_info;
extern struct idr scif_ports;
extern struct scif_dev *scif_dev;
extern const struct file_operations scif_fops;
+extern const struct file_operations scif_anon_fops;
/* Size of the RB for the Node QP */
#define SCIF_NODE_QP_SIZE 0x10000
diff --git a/include/linux/scif.h b/include/linux/scif.h
index 44f4f38..b1923ad 100644
--- a/include/linux/scif.h
+++ b/include/linux/scif.h
@@ -93,6 +93,18 @@ enum {
typedef struct scif_endpt *scif_epd_t;
+/**
+ * struct scif_pollepd - SCIF endpoint to be monitored via scif_poll
+ * @epd: SCIF endpoint
+ * @events: requested events
+ * @revents: returned events
+ */
+struct scif_pollepd {
+ scif_epd_t epd;
+ short events;
+ short revents;
+};
+
#define SCIF_OPEN_FAILED ((scif_epd_t)-1)
#define SCIF_REGISTER_FAILED ((off_t)-1)
#define SCIF_MMAP_FAILED ((void *)-1)
@@ -990,4 +1002,66 @@ int scif_fence_signal(scif_epd_t epd, off_t loff, u64 lval, off_t roff,
*/
int scif_get_node_ids(u16 *nodes, int len, u16 *self);
+/**
+ * scif_poll() - Wait for some event on an endpoint
+ * @epds: Array of endpoint descriptors
+ * @nepds: Length of epds
+ * @timeout: Upper limit on time for which scif_poll() will block
+ *
+ * scif_poll() waits for one of a set of endpoints to become ready to perform
+ * an I/O operation.
+ *
+ * The epds argument specifies the endpoint descriptors to be examined and the
+ * events of interest for each endpoint descriptor. epds is a pointer to an
+ * array with one member for each open endpoint descriptor of interest.
+ *
+ * The number of items in the epds array is specified in nepds. The epd field
+ * of scif_pollepd is an endpoint descriptor of an open endpoint. The field
+ * events is a bitmask specifying the events which the application is
+ * interested in. The field revents is an output parameter, filled by the
+ * kernel with the events that actually occurred. The bits returned in revents
+ * can include any of those specified in events, or one of the values POLLERR,
+ * POLLHUP, or POLLNVAL. (These three bits are meaningless in the events
+ * field, and will be set in the revents field whenever the corresponding
+ * condition is true.)
+ *
+ * If none of the events requested (and no error) has occurred for any of the
+ * endpoint descriptors, then scif_poll() blocks until one of the events occurs.
+ *
+ * The timeout argument specifies an upper limit on the time for which
+ * scif_poll() will block, in milliseconds. Specifying a negative value in
+ * timeout means an infinite timeout.
+ *
+ * The following bits may be set in events and returned in revents.
+ * POLLIN - Data may be received without blocking. For a connected
+ * endpoint, this means that scif_recv() may be called without blocking. For a
+ * listening endpoint, this means that scif_accept() may be called without
+ * blocking.
+ * POLLOUT - Data may be sent without blocking. For a connected endpoint, this
+ * means that scif_send() may be called without blocking. POLLOUT may also be
+ * used to block waiting for a non-blocking connect to complete. This bit value
+ * has no meaning for a listening endpoint and is ignored if specified.
+ *
+ * The following bits are only returned in revents, and are ignored if set in
+ * events.
+ * POLLERR - An error occurred on the endpoint
+ * POLLHUP - The connection to the peer endpoint was disconnected
+ * POLLNVAL - The specified endpoint descriptor is invalid.
+ *
+ * Return:
+ * Upon successful completion, scif_poll() returns a non-negative value. A
+ * positive value indicates the total number of endpoint descriptors that have
+ * been selected (that is, endpoint descriptors for which the revents member is
+ * non-zero). A value of 0 indicates that the call timed out and no endpoint
+ * descriptors have been selected. Otherwise in user mode -1 is returned and
+ * errno is set to indicate the error; in kernel mode the negative of one of
+ * the following errors is returned.
+ *
+ * Errors:
+ * EINTR - A signal occurred before any requested event
+ * EINVAL - The nepds argument is greater than {OPEN_MAX}
+ * ENOMEM - There was no space to allocate file descriptor tables
+ */
+int scif_poll(struct scif_pollepd *epds, unsigned int nepds, long timeout);
+
#endif /* __SCIF_H__ */
--
2.0.0.rc3.2.g998f840
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
[toc] | [prev] | [next] | [standalone]
| From | Ashutosh Dixit <ashutosh.dixit@intel.com> |
|---|---|
| Date | 2015-07-28 01:30 +0200 |
| Subject | [PATCH char-misc-next 02/19] misc: mic: Add support for kernel mode SCIF clients |
| Message-ID | <pR00i-8jZ-37@gated-at.bofh.it> |
| In reply to | #1193568 |
Add support for registration/de-registration of kernel mode SCIF
clients. SCIF clients are probed with new and existing SCIF peer
devices. Similarly the client remove method is called when SCIF
peer devices are removed.
Changes to SCIF peer device framework necessitated by supporting
kernel mode SCIF clients are also included in this patch.
Reviewed-by: Nikhil Rao <nikhil.rao@intel.com>
Reviewed-by: Sudeep Dutt <sudeep.dutt@intel.com>
Signed-off-by: Ashutosh Dixit <ashutosh.dixit@intel.com>
---
drivers/misc/mic/scif/scif_api.c | 44 +++++++++
drivers/misc/mic/scif/scif_main.c | 88 ++++--------------
drivers/misc/mic/scif/scif_main.h | 5 +-
drivers/misc/mic/scif/scif_nm.c | 10 +--
drivers/misc/mic/scif/scif_nodeqp.c | 65 +++++---------
drivers/misc/mic/scif/scif_peer_bus.c | 165 +++++++++++++++++++++-------------
drivers/misc/mic/scif/scif_peer_bus.h | 42 +--------
include/linux/scif.h | 58 ++++++++++++
8 files changed, 256 insertions(+), 221 deletions(-)
diff --git a/drivers/misc/mic/scif/scif_api.c b/drivers/misc/mic/scif/scif_api.c
index bf2d70f..900c86d 100644
--- a/drivers/misc/mic/scif/scif_api.c
+++ b/drivers/misc/mic/scif/scif_api.c
@@ -1430,3 +1430,47 @@ int scif_get_node_ids(u16 *nodes, int len, u16 *self)
return online;
}
EXPORT_SYMBOL_GPL(scif_get_node_ids);
+
+static int scif_add_client_dev(struct device *dev, struct subsys_interface *si)
+{
+ struct scif_client *client =
+ container_of(si, struct scif_client, si);
+ struct scif_peer_dev *spdev =
+ container_of(dev, struct scif_peer_dev, dev);
+
+ if (client->probe)
+ client->probe(spdev);
+ return 0;
+}
+
+static int scif_remove_client_dev(struct device *dev,
+ struct subsys_interface *si)
+{
+ struct scif_client *client =
+ container_of(si, struct scif_client, si);
+ struct scif_peer_dev *spdev =
+ container_of(dev, struct scif_peer_dev, dev);
+
+ if (client->remove)
+ client->remove(spdev);
+ return 0;
+}
+
+void scif_client_unregister(struct scif_client *client)
+{
+ subsys_interface_unregister(&client->si);
+}
+EXPORT_SYMBOL_GPL(scif_client_unregister);
+
+int scif_client_register(struct scif_client *client)
+{
+ struct subsys_interface *si = &client->si;
+
+ si->name = client->name;
+ si->subsys = &scif_peer_bus;
+ si->add_dev = scif_add_client_dev;
+ si->remove_dev = scif_remove_client_dev;
+
+ return subsys_interface_register(&client->si);
+}
+EXPORT_SYMBOL_GPL(scif_client_register);
diff --git a/drivers/misc/mic/scif/scif_main.c b/drivers/misc/mic/scif/scif_main.c
index 6ce851f..f90bd06 100644
--- a/drivers/misc/mic/scif/scif_main.c
+++ b/drivers/misc/mic/scif/scif_main.c
@@ -80,35 +80,6 @@ irqreturn_t scif_intr_handler(int irq, void *data)
return IRQ_HANDLED;
}
-static int scif_peer_probe(struct scif_peer_dev *spdev)
-{
- struct scif_dev *scifdev = &scif_dev[spdev->dnode];
-
- mutex_lock(&scif_info.conflock);
- scif_info.total++;
- scif_info.maxid = max_t(u32, spdev->dnode, scif_info.maxid);
- mutex_unlock(&scif_info.conflock);
- rcu_assign_pointer(scifdev->spdev, spdev);
-
- /* In the future SCIF kernel client devices will be added here */
- return 0;
-}
-
-static void scif_peer_remove(struct scif_peer_dev *spdev)
-{
- struct scif_dev *scifdev = &scif_dev[spdev->dnode];
-
- /* In the future SCIF kernel client devices will be removed here */
- spdev = rcu_dereference(scifdev->spdev);
- if (spdev)
- RCU_INIT_POINTER(scifdev->spdev, NULL);
- synchronize_rcu();
-
- mutex_lock(&scif_info.conflock);
- scif_info.total--;
- mutex_unlock(&scif_info.conflock);
-}
-
static void scif_qp_setup_handler(struct work_struct *work)
{
struct scif_dev *scifdev = container_of(work, struct scif_dev,
@@ -139,20 +110,13 @@ static void scif_qp_setup_handler(struct work_struct *work)
}
}
-static int scif_setup_scifdev(struct scif_hw_dev *sdev)
+static int scif_setup_scifdev(void)
{
+ /* We support a maximum of 129 SCIF nodes including the mgmt node */
+#define MAX_SCIF_NODES 129
int i;
- u8 num_nodes;
+ u8 num_nodes = MAX_SCIF_NODES;
- if (sdev->snode) {
- struct mic_bootparam __iomem *bp = sdev->rdp;
-
- num_nodes = ioread8(&bp->tot_nodes);
- } else {
- struct mic_bootparam *bp = sdev->dp;
-
- num_nodes = bp->tot_nodes;
- }
scif_dev = kcalloc(num_nodes, sizeof(*scif_dev), GFP_KERNEL);
if (!scif_dev)
return -ENOMEM;
@@ -163,7 +127,7 @@ static int scif_setup_scifdev(struct scif_hw_dev *sdev)
scifdev->exit = OP_IDLE;
init_waitqueue_head(&scifdev->disconn_wq);
mutex_init(&scifdev->lock);
- INIT_WORK(&scifdev->init_msg_work, scif_qp_response_ack);
+ INIT_WORK(&scifdev->peer_add_work, scif_add_peer_device);
INIT_DELAYED_WORK(&scifdev->p2p_dwork,
scif_poll_qp_state);
INIT_DELAYED_WORK(&scifdev->qp_dwork,
@@ -181,27 +145,21 @@ static void scif_destroy_scifdev(void)
static int scif_probe(struct scif_hw_dev *sdev)
{
- struct scif_dev *scifdev;
+ struct scif_dev *scifdev = &scif_dev[sdev->dnode];
int rc;
dev_set_drvdata(&sdev->dev, sdev);
+ scifdev->sdev = sdev;
+
if (1 == atomic_add_return(1, &g_loopb_cnt)) {
- struct scif_dev *loopb_dev;
+ struct scif_dev *loopb_dev = &scif_dev[sdev->snode];
- rc = scif_setup_scifdev(sdev);
- if (rc)
- goto exit;
- scifdev = &scif_dev[sdev->dnode];
- scifdev->sdev = sdev;
- loopb_dev = &scif_dev[sdev->snode];
loopb_dev->sdev = sdev;
rc = scif_setup_loopback_qp(loopb_dev);
if (rc)
- goto free_sdev;
- } else {
- scifdev = &scif_dev[sdev->dnode];
- scifdev->sdev = sdev;
+ goto exit;
}
+
rc = scif_setup_intr_wq(scifdev);
if (rc)
goto destroy_loopb;
@@ -237,8 +195,6 @@ destroy_intr:
destroy_loopb:
if (atomic_dec_and_test(&g_loopb_cnt))
scif_destroy_loopback_qp(&scif_dev[sdev->snode]);
-free_sdev:
- scif_destroy_scifdev();
exit:
return rc;
}
@@ -290,13 +246,6 @@ static void scif_remove(struct scif_hw_dev *sdev)
scifdev->sdev = NULL;
}
-static struct scif_peer_driver scif_peer_driver = {
- .driver.name = KBUILD_MODNAME,
- .driver.owner = THIS_MODULE,
- .probe = scif_peer_probe,
- .remove = scif_peer_remove,
-};
-
static struct scif_hw_dev_id id_table[] = {
{ MIC_SCIF_DEV, SCIF_DEV_ANY_ID },
{ 0 },
@@ -312,6 +261,8 @@ static struct scif_driver scif_driver = {
static int _scif_init(void)
{
+ int rc;
+
spin_lock_init(&scif_info.eplock);
spin_lock_init(&scif_info.nb_connect_lock);
spin_lock_init(&scif_info.port_lock);
@@ -326,10 +277,15 @@ static int _scif_init(void)
init_waitqueue_head(&scif_info.exitwq);
scif_info.en_msg_log = 0;
scif_info.p2p_enable = 1;
+ rc = scif_setup_scifdev();
+ if (rc)
+ goto error;
INIT_WORK(&scif_info.misc_work, scif_misc_handler);
INIT_WORK(&scif_info.conn_work, scif_conn_handler);
idr_init(&scif_ports);
return 0;
+error:
+ return rc;
}
static void _scif_exit(void)
@@ -347,12 +303,9 @@ static int __init scif_init(void)
rc = scif_peer_bus_init();
if (rc)
goto exit;
- rc = scif_peer_register_driver(&scif_peer_driver);
- if (rc)
- goto peer_bus_exit;
rc = scif_register_driver(&scif_driver);
if (rc)
- goto unreg_scif_peer;
+ goto peer_bus_exit;
rc = misc_register(mdev);
if (rc)
goto unreg_scif;
@@ -360,8 +313,6 @@ static int __init scif_init(void)
return 0;
unreg_scif:
scif_unregister_driver(&scif_driver);
-unreg_scif_peer:
- scif_peer_unregister_driver(&scif_peer_driver);
peer_bus_exit:
scif_peer_bus_exit();
exit:
@@ -374,7 +325,6 @@ static void __exit scif_exit(void)
scif_exit_debugfs();
misc_deregister(&scif_info.mdev);
scif_unregister_driver(&scif_driver);
- scif_peer_unregister_driver(&scif_peer_driver);
scif_peer_bus_exit();
_scif_exit();
}
diff --git a/drivers/misc/mic/scif/scif_main.h b/drivers/misc/mic/scif/scif_main.h
index 87c1327..b0795b0 100644
--- a/drivers/misc/mic/scif/scif_main.h
+++ b/drivers/misc/mic/scif/scif_main.h
@@ -140,7 +140,7 @@ struct scif_p2p_info {
* @db: doorbell the peer will trigger to generate an interrupt on self
* @rdb: Doorbell to trigger on the peer to generate an interrupt on the peer
* @cookie: Cookie received while registering the interrupt handler
- * init_msg_work: work scheduled for SCIF_INIT message processing
+ * @peer_add_work: Work for handling device_add for peer devices
* @p2p_dwork: Delayed work to enable polling for P2P state
* @qp_dwork: Delayed work for enabling polling for remote QP information
* @p2p_retry: Number of times to retry polling of P2P state
@@ -166,7 +166,7 @@ struct scif_dev {
int db;
int rdb;
struct mic_irq *cookie;
- struct work_struct init_msg_work;
+ struct work_struct peer_add_work;
struct delayed_work p2p_dwork;
struct delayed_work qp_dwork;
int p2p_retry;
@@ -183,6 +183,7 @@ struct scif_dev {
extern struct scif_info scif_info;
extern struct idr scif_ports;
+extern struct bus_type scif_peer_bus;
extern struct scif_dev *scif_dev;
extern const struct file_operations scif_fops;
extern const struct file_operations scif_anon_fops;
diff --git a/drivers/misc/mic/scif/scif_nm.c b/drivers/misc/mic/scif/scif_nm.c
index 9b4c538..f1b1b97 100644
--- a/drivers/misc/mic/scif/scif_nm.c
+++ b/drivers/misc/mic/scif/scif_nm.c
@@ -147,14 +147,8 @@ void scif_cleanup_scifdev(struct scif_dev *dev)
void scif_handle_remove_node(int node)
{
struct scif_dev *scifdev = &scif_dev[node];
- struct scif_peer_dev *spdev;
-
- rcu_read_lock();
- spdev = rcu_dereference(scifdev->spdev);
- rcu_read_unlock();
- if (spdev)
- scif_peer_unregister_device(spdev);
- else
+
+ if (scif_peer_unregister_device(scifdev))
scif_send_acks(scifdev);
}
diff --git a/drivers/misc/mic/scif/scif_nodeqp.c b/drivers/misc/mic/scif/scif_nodeqp.c
index 6dfdae3..5cdb9f0 100644
--- a/drivers/misc/mic/scif/scif_nodeqp.c
+++ b/drivers/misc/mic/scif/scif_nodeqp.c
@@ -260,6 +260,11 @@ int scif_setup_qp_connect_response(struct scif_dev *scifdev,
r_buf,
get_count_order(remote_size));
/*
+ * Because the node QP may already be processing an INIT message, set
+ * the read pointer so the cached read offset isn't lost
+ */
+ qp->remote_qp->local_read = qp->inbound_q.current_read_offset;
+ /*
* resetup the inbound_q now that we know where the
* inbound_read really is.
*/
@@ -529,27 +534,6 @@ static void scif_p2p_setup(void)
}
}
-void scif_qp_response_ack(struct work_struct *work)
-{
- struct scif_dev *scifdev = container_of(work, struct scif_dev,
- init_msg_work);
- struct scif_peer_dev *spdev;
-
- /* Drop the INIT message if it has already been received */
- if (_scifdev_alive(scifdev))
- return;
-
- spdev = scif_peer_register_device(scifdev);
- if (IS_ERR(spdev))
- return;
-
- if (scif_is_mgmt_node()) {
- mutex_lock(&scif_info.conflock);
- scif_p2p_setup();
- mutex_unlock(&scif_info.conflock);
- }
-}
-
static char *message_types[] = {"BAD",
"INIT",
"EXIT",
@@ -682,13 +666,14 @@ scif_init(struct scif_dev *scifdev, struct scifmsg *msg)
* address to complete initializing the inbound_q.
*/
flush_delayed_work(&scifdev->qp_dwork);
- /*
- * Delegate the peer device registration to a workqueue, otherwise if
- * SCIF client probe (called during peer device registration) calls
- * scif_connect(..), it will block the message processing thread causing
- * a deadlock.
- */
- schedule_work(&scifdev->init_msg_work);
+
+ scif_peer_register_device(scifdev);
+
+ if (scif_is_mgmt_node()) {
+ mutex_lock(&scif_info.conflock);
+ scif_p2p_setup();
+ mutex_unlock(&scif_info.conflock);
+ }
}
/**
@@ -838,13 +823,13 @@ void scif_poll_qp_state(struct work_struct *work)
msecs_to_jiffies(SCIF_NODE_QP_TIMEOUT));
return;
}
- scif_peer_register_device(peerdev);
return;
timeout:
dev_err(&peerdev->sdev->dev,
"%s %d remote node %d offline, state = 0x%x\n",
__func__, __LINE__, peerdev->node, qp->qp_state);
qp->remote_qp->qp_state = SCIF_QP_OFFLINE;
+ scif_peer_unregister_device(peerdev);
scif_cleanup_scifdev(peerdev);
}
@@ -894,6 +879,9 @@ scif_node_add_ack(struct scif_dev *scifdev, struct scifmsg *msg)
goto local_error;
peerdev->rdb = msg->payload[2];
qp->remote_qp->qp_state = SCIF_QP_ONLINE;
+
+ scif_peer_register_device(peerdev);
+
schedule_delayed_work(&peerdev->p2p_dwork, 0);
return;
local_error:
@@ -1169,7 +1157,6 @@ int scif_setup_loopback_qp(struct scif_dev *scifdev)
int err = 0;
void *local_q;
struct scif_qp *qp;
- struct scif_peer_dev *spdev;
err = scif_setup_intr_wq(scifdev);
if (err)
@@ -1216,15 +1203,11 @@ int scif_setup_loopback_qp(struct scif_dev *scifdev)
&qp->local_write,
local_q, get_count_order(SCIF_NODE_QP_SIZE));
scif_info.nodeid = scifdev->node;
- spdev = scif_peer_register_device(scifdev);
- if (IS_ERR(spdev)) {
- err = PTR_ERR(spdev);
- goto free_local_q;
- }
+
+ scif_peer_register_device(scifdev);
+
scif_info.loopb_dev = scifdev;
return err;
-free_local_q:
- kfree(local_q);
free_qpairs:
kfree(scifdev->qpairs);
destroy_loopb_wq:
@@ -1243,13 +1226,7 @@ exit:
*/
int scif_destroy_loopback_qp(struct scif_dev *scifdev)
{
- struct scif_peer_dev *spdev;
-
- rcu_read_lock();
- spdev = rcu_dereference(scifdev->spdev);
- rcu_read_unlock();
- if (spdev)
- scif_peer_unregister_device(spdev);
+ scif_peer_unregister_device(scifdev);
destroy_workqueue(scif_info.loopb_wq);
scif_destroy_intr_wq(scifdev);
kfree(scifdev->qpairs->outbound_q.rb_base);
diff --git a/drivers/misc/mic/scif/scif_peer_bus.c b/drivers/misc/mic/scif/scif_peer_bus.c
index 589ae9a..547bf7b 100644
--- a/drivers/misc/mic/scif/scif_peer_bus.c
+++ b/drivers/misc/mic/scif/scif_peer_bus.c
@@ -24,93 +24,138 @@ dev_to_scif_peer(struct device *dev)
return container_of(dev, struct scif_peer_dev, dev);
}
-static inline struct scif_peer_driver *
-drv_to_scif_peer(struct device_driver *drv)
-{
- return container_of(drv, struct scif_peer_driver, driver);
-}
+struct bus_type scif_peer_bus = {
+ .name = "scif_peer_bus",
+};
-static int scif_peer_dev_match(struct device *dv, struct device_driver *dr)
+static void scif_peer_release_dev(struct device *d)
{
- return !strncmp(dev_name(dv), dr->name, 4);
+ struct scif_peer_dev *sdev = dev_to_scif_peer(d);
+ struct scif_dev *scifdev = &scif_dev[sdev->dnode];
+
+ scif_cleanup_scifdev(scifdev);
+ kfree(sdev);
}
-static int scif_peer_dev_probe(struct device *d)
+static int scif_peer_initialize_device(struct scif_dev *scifdev)
{
- struct scif_peer_dev *dev = dev_to_scif_peer(d);
- struct scif_peer_driver *drv = drv_to_scif_peer(dev->dev.driver);
+ struct scif_peer_dev *spdev;
+ int ret;
- return drv->probe(dev);
-}
+ spdev = kzalloc(sizeof(*spdev), GFP_KERNEL);
+ if (!spdev) {
+ ret = -ENOMEM;
+ goto err;
+ }
-static int scif_peer_dev_remove(struct device *d)
-{
- struct scif_peer_dev *dev = dev_to_scif_peer(d);
- struct scif_peer_driver *drv = drv_to_scif_peer(dev->dev.driver);
+ spdev->dev.parent = scifdev->sdev->dev.parent;
+ spdev->dev.release = scif_peer_release_dev;
+ spdev->dnode = scifdev->node;
+ spdev->dev.bus = &scif_peer_bus;
+ dev_set_name(&spdev->dev, "scif_peer-dev%u", spdev->dnode);
+
+ device_initialize(&spdev->dev);
+ get_device(&spdev->dev);
+ rcu_assign_pointer(scifdev->spdev, spdev);
- drv->remove(dev);
+ mutex_lock(&scif_info.conflock);
+ scif_info.total++;
+ scif_info.maxid = max_t(u32, spdev->dnode, scif_info.maxid);
+ mutex_unlock(&scif_info.conflock);
return 0;
+err:
+ dev_err(&scifdev->sdev->dev,
+ "dnode %d: initialize_device rc %d\n", scifdev->node, ret);
+ return ret;
}
-static struct bus_type scif_peer_bus = {
- .name = "scif_peer_bus",
- .match = scif_peer_dev_match,
- .probe = scif_peer_dev_probe,
- .remove = scif_peer_dev_remove,
-};
-
-int scif_peer_register_driver(struct scif_peer_driver *driver)
+static int scif_peer_add_device(struct scif_dev *scifdev)
{
- driver->driver.bus = &scif_peer_bus;
- return driver_register(&driver->driver);
+ struct scif_peer_dev *spdev = rcu_dereference(scifdev->spdev);
+ int ret;
+
+ ret = device_add(&spdev->dev);
+ put_device(&spdev->dev);
+ if (ret) {
+ dev_err(&scifdev->sdev->dev,
+ "dnode %d: peer device_add failed\n", scifdev->node);
+ goto put_spdev;
+ }
+ dev_dbg(&spdev->dev, "Added peer dnode %d\n", spdev->dnode);
+ return 0;
+put_spdev:
+ RCU_INIT_POINTER(scifdev->spdev, NULL);
+ synchronize_rcu();
+ put_device(&spdev->dev);
+
+ mutex_lock(&scif_info.conflock);
+ scif_info.total--;
+ mutex_unlock(&scif_info.conflock);
+ return ret;
}
-void scif_peer_unregister_driver(struct scif_peer_driver *driver)
+void scif_add_peer_device(struct work_struct *work)
{
- driver_unregister(&driver->driver);
+ struct scif_dev *scifdev = container_of(work, struct scif_dev,
+ peer_add_work);
+
+ scif_peer_add_device(scifdev);
}
-static void scif_peer_release_dev(struct device *d)
+/*
+ * Peer device registration is split into a device_initialize and a device_add.
+ * The reason for doing this is as follows: First, peer device registration
+ * itself cannot be done in the message processing thread and must be delegated
+ * to another workqueue, otherwise if SCIF client probe, called during peer
+ * device registration, calls scif_connect(..), it will block the message
+ * processing thread causing a deadlock. Next, device_initialize is done in the
+ * "top-half" message processing thread and device_add in the "bottom-half"
+ * workqueue. If this is not done, SCIF_CNCT_REQ message processing executing
+ * concurrently with SCIF_INIT message processing is unable to get a reference
+ * on the peer device, thereby failing the connect request.
+ */
+void scif_peer_register_device(struct scif_dev *scifdev)
{
- struct scif_peer_dev *sdev = dev_to_scif_peer(d);
- struct scif_dev *scifdev = &scif_dev[sdev->dnode];
+ int ret;
- scif_cleanup_scifdev(scifdev);
- kfree(sdev);
+ mutex_lock(&scifdev->lock);
+ ret = scif_peer_initialize_device(scifdev);
+ if (ret)
+ goto exit;
+ schedule_work(&scifdev->peer_add_work);
+exit:
+ mutex_unlock(&scifdev->lock);
}
-struct scif_peer_dev *
-scif_peer_register_device(struct scif_dev *scifdev)
+int scif_peer_unregister_device(struct scif_dev *scifdev)
{
- int ret;
struct scif_peer_dev *spdev;
- spdev = kzalloc(sizeof(*spdev), GFP_KERNEL);
- if (!spdev)
- return ERR_PTR(-ENOMEM);
-
- spdev->dev.parent = scifdev->sdev->dev.parent;
- spdev->dev.release = scif_peer_release_dev;
- spdev->dnode = scifdev->node;
- spdev->dev.bus = &scif_peer_bus;
+ mutex_lock(&scifdev->lock);
+ /* Flush work to ensure device register is complete */
+ flush_work(&scifdev->peer_add_work);
- dev_set_name(&spdev->dev, "scif_peer-dev%u", spdev->dnode);
/*
- * device_register() causes the bus infrastructure to look for a
- * matching driver.
+ * Continue holding scifdev->lock since theoretically unregister_device
+ * can be called simultaneously from multiple threads
*/
- ret = device_register(&spdev->dev);
- if (ret)
- goto free_spdev;
- return spdev;
-free_spdev:
- kfree(spdev);
- return ERR_PTR(ret);
-}
-
-void scif_peer_unregister_device(struct scif_peer_dev *sdev)
-{
- device_unregister(&sdev->dev);
+ spdev = rcu_dereference(scifdev->spdev);
+ if (!spdev) {
+ mutex_unlock(&scifdev->lock);
+ return -ENODEV;
+ }
+
+ RCU_INIT_POINTER(scifdev->spdev, NULL);
+ synchronize_rcu();
+ mutex_unlock(&scifdev->lock);
+
+ dev_dbg(&spdev->dev, "Removing peer dnode %d\n", spdev->dnode);
+ device_unregister(&spdev->dev);
+
+ mutex_lock(&scif_info.conflock);
+ scif_info.total--;
+ mutex_unlock(&scif_info.conflock);
+ return 0;
}
int scif_peer_bus_init(void)
diff --git a/drivers/misc/mic/scif/scif_peer_bus.h b/drivers/misc/mic/scif/scif_peer_bus.h
index 33f0dbb..a3b8dd2 100644
--- a/drivers/misc/mic/scif/scif_peer_bus.h
+++ b/drivers/misc/mic/scif/scif_peer_bus.h
@@ -19,47 +19,13 @@
#include <linux/device.h>
#include <linux/mic_common.h>
-
-/*
- * Peer devices show up as PCIe devices for the mgmt node but not the cards.
- * The mgmt node discovers all the cards on the PCIe bus and informs the other
- * cards about their peers. Upon notification of a peer a node adds a peer
- * device to the peer bus to maintain symmetry in the way devices are
- * discovered across all nodes in the SCIF network.
- */
-/**
- * scif_peer_dev - representation of a peer SCIF device
- * @dev: underlying device
- * @dnode - The destination node which this device will communicate with.
- */
-struct scif_peer_dev {
- struct device dev;
- u8 dnode;
-};
-
-/**
- * scif_peer_driver - operations for a scif_peer I/O driver
- * @driver: underlying device driver (populate name and owner).
- * @id_table: the ids serviced by this driver.
- * @probe: the function to call when a device is found. Returns 0 or -errno.
- * @remove: the function to call when a device is removed.
- */
-struct scif_peer_driver {
- struct device_driver driver;
- const struct scif_peer_dev_id *id_table;
-
- int (*probe)(struct scif_peer_dev *dev);
- void (*remove)(struct scif_peer_dev *dev);
-};
+#include <linux/scif.h>
struct scif_dev;
-int scif_peer_register_driver(struct scif_peer_driver *driver);
-void scif_peer_unregister_driver(struct scif_peer_driver *driver);
-
-struct scif_peer_dev *scif_peer_register_device(struct scif_dev *sdev);
-void scif_peer_unregister_device(struct scif_peer_dev *sdev);
-
+void scif_add_peer_device(struct work_struct *work);
+void scif_peer_register_device(struct scif_dev *sdev);
+int scif_peer_unregister_device(struct scif_dev *scifdev);
int scif_peer_bus_init(void);
void scif_peer_bus_exit(void);
#endif /* _SCIF_PEER_BUS_H */
diff --git a/include/linux/scif.h b/include/linux/scif.h
index b1923ad..fd62c05 100644
--- a/include/linux/scif.h
+++ b/include/linux/scif.h
@@ -55,6 +55,7 @@
#include <linux/types.h>
#include <linux/poll.h>
+#include <linux/device.h>
#include <linux/scif_ioctl.h>
#define SCIF_ACCEPT_SYNC 1
@@ -105,6 +106,37 @@ struct scif_pollepd {
short revents;
};
+/**
+ * scif_peer_dev - representation of a peer SCIF device
+ *
+ * Peer devices show up as PCIe devices for the mgmt node but not the cards.
+ * The mgmt node discovers all the cards on the PCIe bus and informs the other
+ * cards about their peers. Upon notification of a peer a node adds a peer
+ * device to the peer bus to maintain symmetry in the way devices are
+ * discovered across all nodes in the SCIF network.
+ *
+ * @dev: underlying device
+ * @dnode - The destination node which this device will communicate with.
+ */
+struct scif_peer_dev {
+ struct device dev;
+ u8 dnode;
+};
+
+/**
+ * scif_client - representation of a SCIF client
+ * @name: client name
+ * @probe - client method called when a peer device is registered
+ * @remove - client method called when a peer device is unregistered
+ * @si - subsys_interface used internally for implementing SCIF clients
+ */
+struct scif_client {
+ const char *name;
+ void (*probe)(struct scif_peer_dev *spdev);
+ void (*remove)(struct scif_peer_dev *spdev);
+ struct subsys_interface si;
+};
+
#define SCIF_OPEN_FAILED ((scif_epd_t)-1)
#define SCIF_REGISTER_FAILED ((off_t)-1)
#define SCIF_MMAP_FAILED ((void *)-1)
@@ -1064,4 +1096,30 @@ int scif_get_node_ids(u16 *nodes, int len, u16 *self);
*/
int scif_poll(struct scif_pollepd *epds, unsigned int nepds, long timeout);
+/**
+ * scif_client_register() - Register a SCIF client
+ * @client: client to be registered
+ *
+ * scif_client_register() registers a SCIF client. The probe() method
+ * of the client is called when SCIF peer devices come online and the
+ * remove() method is called when the peer devices disappear.
+ *
+ * Return:
+ * Upon successful completion, scif_client_register() returns a non-negative
+ * value. Otherwise the return value is the same as subsys_interface_register()
+ * in the kernel.
+ */
+int scif_client_register(struct scif_client *client);
+
+/**
+ * scif_client_unregister() - Unregister a SCIF client
+ * @client: client to be unregistered
+ *
+ * scif_client_unregister() unregisters a SCIF client.
+ *
+ * Return:
+ * None
+ */
+void scif_client_unregister(struct scif_client *client);
+
#endif /* __SCIF_H__ */
--
2.0.0.rc3.2.g998f840
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
[toc] | [prev] | [next] | [standalone]
| From | Ashutosh Dixit <ashutosh.dixit@intel.com> |
|---|---|
| Date | 2015-07-28 01:40 +0200 |
| Subject | [PATCH char-misc-next 19/19] misc: mic: SCIF RMA nodeqp and minor miscellaneous changes |
| Message-ID | <pR09X-8vE-1@gated-at.bofh.it> |
| In reply to | #1193568 |
From: Sudeep Dutt <sudeep.dutt@intel.com>
This patch adds the SCIF kernel node QP control messages required to
enable SCIF RMAs. Examples of such node QP control messages include
registration, unregistration, remote memory allocation requests,
remote memory unmap and SCIF remote fence requests.
The patch also updates the SCIF driver with minor changes required to
enable SCIF RMAs by adding the new files to the build, initializing
RMA specific information during SCIF endpoint creation, reserving SCIF
DMA channels, initializing SCIF RMA specific global data structures,
adding the IOCTL hooks required for SCIF RMAs and updating RMA
specific debugfs hooks.
Reviewed-by: Ashutosh Dixit <ashutosh.dixit@intel.com>
Reviewed-by: Nikhil Rao <nikhil.rao@intel.com>
Signed-off-by: Sudeep Dutt <sudeep.dutt@intel.com>
---
drivers/misc/mic/Kconfig | 1 +
drivers/misc/mic/scif/Makefile | 5 ++
drivers/misc/mic/scif/scif_api.c | 21 +++++
drivers/misc/mic/scif/scif_debugfs.c | 85 +++++++++++++++++-
drivers/misc/mic/scif/scif_epd.c | 9 +-
drivers/misc/mic/scif/scif_epd.h | 28 ++++++
drivers/misc/mic/scif/scif_fd.c | 159 ++++++++++++++++++++++++++++++++++
drivers/misc/mic/scif/scif_main.c | 19 ++++
drivers/misc/mic/scif/scif_main.h | 28 +++++-
drivers/misc/mic/scif/scif_map.h | 25 +++++-
drivers/misc/mic/scif/scif_nm.c | 10 ++-
drivers/misc/mic/scif/scif_nodeqp.c | 53 +++++++++++-
drivers/misc/mic/scif/scif_nodeqp.h | 42 ++++++++-
drivers/misc/mic/scif/scif_peer_bus.c | 14 +++
14 files changed, 484 insertions(+), 15 deletions(-)
diff --git a/drivers/misc/mic/Kconfig b/drivers/misc/mic/Kconfig
index 1488232..e3bf95c 100644
--- a/drivers/misc/mic/Kconfig
+++ b/drivers/misc/mic/Kconfig
@@ -75,6 +75,7 @@ comment "SCIF Driver"
config SCIF
tristate "SCIF Driver"
depends on 64BIT && PCI && X86 && SCIF_BUS
+ select IOVA
help
This enables SCIF Driver support for the Intel Many Integrated
Core (MIC) family of PCIe form factor coprocessor devices that
diff --git a/drivers/misc/mic/scif/Makefile b/drivers/misc/mic/scif/Makefile
index bf10bb7..29cfc3e 100644
--- a/drivers/misc/mic/scif/Makefile
+++ b/drivers/misc/mic/scif/Makefile
@@ -13,3 +13,8 @@ scif-objs += scif_epd.o
scif-objs += scif_rb.o
scif-objs += scif_nodeqp.o
scif-objs += scif_nm.o
+scif-objs += scif_dma.o
+scif-objs += scif_fence.o
+scif-objs += scif_mmap.o
+scif-objs += scif_rma.o
+scif-objs += scif_rma_list.o
diff --git a/drivers/misc/mic/scif/scif_api.c b/drivers/misc/mic/scif/scif_api.c
index 900c86d..7fd0fe3 100644
--- a/drivers/misc/mic/scif/scif_api.c
+++ b/drivers/misc/mic/scif/scif_api.c
@@ -70,6 +70,7 @@ scif_epd_t scif_open(void)
mutex_init(&ep->sendlock);
mutex_init(&ep->recvlock);
+ scif_rma_ep_init(ep);
ep->state = SCIFEP_UNBOUND;
dev_dbg(scif_info.mdev.this_device,
"SCIFAPI open: ep %p success\n", ep);
@@ -184,8 +185,11 @@ int scif_close(scif_epd_t epd)
switch (oldstate) {
case SCIFEP_ZOMBIE:
+ dev_err(scif_info.mdev.this_device,
+ "SCIFAPI close: zombie state unexpected\n");
case SCIFEP_DISCONNECTED:
spin_unlock(&ep->lock);
+ scif_unregister_all_windows(epd);
/* Remove from the disconnected list */
mutex_lock(&scif_info.connlock);
list_for_each_safe(pos, tmpq, &scif_info.disconnected) {
@@ -207,6 +211,7 @@ int scif_close(scif_epd_t epd)
case SCIFEP_CLOSING:
{
spin_unlock(&ep->lock);
+ scif_unregister_all_windows(epd);
scif_disconnect_ep(ep);
break;
}
@@ -469,6 +474,13 @@ static int scif_conn_func(struct scif_endpt *ep)
struct scifmsg msg;
struct device *spdev;
+ err = scif_reserve_dma_chan(ep);
+ if (err) {
+ dev_err(&ep->remote_dev->sdev->dev,
+ "%s %d err %d\n", __func__, __LINE__, err);
+ ep->state = SCIFEP_BOUND;
+ goto connect_error_simple;
+ }
/* Initiate the first part of the endpoint QP setup */
err = scif_setup_qp_connect(ep->qp_info.qp, &ep->qp_info.qp_offset,
SCIF_ENDPT_QP_SIZE, ep->remote_dev);
@@ -804,6 +816,15 @@ retry_connection:
cep->remote_dev = &scif_dev[peer->node];
cep->remote_ep = conreq->msg.payload[0];
+ scif_rma_ep_init(cep);
+
+ err = scif_reserve_dma_chan(cep);
+ if (err) {
+ dev_err(scif_info.mdev.this_device,
+ "%s %d err %d\n", __func__, __LINE__, err);
+ goto scif_accept_error_qpalloc;
+ }
+
cep->qp_info.qp = kzalloc(sizeof(*cep->qp_info.qp), GFP_KERNEL);
if (!cep->qp_info.qp) {
err = -ENOMEM;
diff --git a/drivers/misc/mic/scif/scif_debugfs.c b/drivers/misc/mic/scif/scif_debugfs.c
index 51f14e2..c08d657 100644
--- a/drivers/misc/mic/scif/scif_debugfs.c
+++ b/drivers/misc/mic/scif/scif_debugfs.c
@@ -62,10 +62,87 @@ static const struct file_operations scif_dev_ops = {
.release = scif_dev_test_release
};
-void __init scif_init_debugfs(void)
+static void scif_display_window(struct scif_window *window, struct seq_file *s)
+{
+ int j;
+ struct scatterlist *sg;
+ scif_pinned_pages_t pin = window->pinned_pages;
+
+ seq_printf(s, "window %p type %d temp %d offset 0x%llx ",
+ window, window->type, window->temp, window->offset);
+ seq_printf(s, "nr_pages 0x%llx nr_contig_chunks 0x%x prot %d ",
+ window->nr_pages, window->nr_contig_chunks, window->prot);
+ seq_printf(s, "ref_count %d magic 0x%llx peer_window 0x%llx ",
+ window->ref_count, window->magic, window->peer_window);
+ seq_printf(s, "unreg_state 0x%x va_for_temp 0x%lx\n",
+ window->unreg_state, window->va_for_temp);
+
+ for (j = 0; j < window->nr_contig_chunks; j++)
+ seq_printf(s, "page[%d] dma_addr 0x%llx num_pages 0x%llx\n", j,
+ window->dma_addr[j], window->num_pages[j]);
+
+ if (SCIF_WINDOW_SELF == window->type && pin)
+ for (j = 0; j < window->nr_pages; j++)
+ seq_printf(s, "page[%d] = pinned_pages %p address %p\n",
+ j, pin->pages[j],
+ page_address(pin->pages[j]));
+
+ if (window->st)
+ for_each_sg(window->st->sgl, sg, window->st->nents, j)
+ seq_printf(s, "sg[%d] dma addr 0x%llx length 0x%x\n",
+ j, sg_dma_address(sg), sg_dma_len(sg));
+}
+
+static void scif_display_all_windows(struct list_head *head, struct seq_file *s)
{
- struct dentry *d;
+ struct list_head *item;
+ struct scif_window *window;
+ list_for_each(item, head) {
+ window = list_entry(item, struct scif_window, list);
+ scif_display_window(window, s);
+ }
+}
+
+static int scif_rma_test(struct seq_file *s, void *unused)
+{
+ struct scif_endpt *ep;
+ struct list_head *pos;
+
+ mutex_lock(&scif_info.connlock);
+ list_for_each(pos, &scif_info.connected) {
+ ep = list_entry(pos, struct scif_endpt, list);
+ seq_printf(s, "ep %p self windows\n", ep);
+ mutex_lock(&ep->rma_info.rma_lock);
+ scif_display_all_windows(&ep->rma_info.reg_list, s);
+ seq_printf(s, "ep %p remote windows\n", ep);
+ scif_display_all_windows(&ep->rma_info.remote_reg_list, s);
+ mutex_unlock(&ep->rma_info.rma_lock);
+ }
+ mutex_unlock(&scif_info.connlock);
+ return 0;
+}
+
+static int scif_rma_test_open(struct inode *inode, struct file *file)
+{
+ return single_open(file, scif_rma_test, inode->i_private);
+}
+
+static int scif_rma_test_release(struct inode *inode, struct file *file)
+{
+ return single_release(inode, file);
+}
+
+static const struct file_operations scif_rma_ops = {
+ .owner = THIS_MODULE,
+ .open = scif_rma_test_open,
+ .read = seq_read,
+ .llseek = seq_lseek,
+ .release = scif_rma_test_release
+};
+
+void __init scif_init_debugfs(void)
+{
scif_dbg = debugfs_create_dir(KBUILD_MODNAME, NULL);
if (!scif_dbg) {
dev_err(scif_info.mdev.this_device,
@@ -73,8 +150,8 @@ void __init scif_init_debugfs(void)
return;
}
- d = debugfs_create_file("scif_dev", 0444, scif_dbg,
- NULL, &scif_dev_ops);
+ debugfs_create_file("scif_dev", 0444, scif_dbg, NULL, &scif_dev_ops);
+ debugfs_create_file("scif_rma", 0444, scif_dbg, NULL, &scif_rma_ops);
debugfs_create_u8("en_msg_log", 0666, scif_dbg, &scif_info.en_msg_log);
debugfs_create_u8("p2p_enable", 0666, scif_dbg, &scif_info.p2p_enable);
}
diff --git a/drivers/misc/mic/scif/scif_epd.c b/drivers/misc/mic/scif/scif_epd.c
index b4bfbb0..34fdc2a 100644
--- a/drivers/misc/mic/scif/scif_epd.c
+++ b/drivers/misc/mic/scif/scif_epd.c
@@ -102,9 +102,12 @@ void scif_cleanup_zombie_epd(void)
spin_lock(&scif_info.eplock);
list_for_each_safe(pos, tmpq, &scif_info.zombie) {
ep = list_entry(pos, struct scif_endpt, list);
- list_del(pos);
- scif_info.nr_zombies--;
- kfree(ep);
+ if (scif_rma_ep_can_uninit(ep)) {
+ list_del(pos);
+ scif_info.nr_zombies--;
+ put_iova_domain(&ep->rma_info.iovad);
+ kfree(ep);
+ }
}
spin_unlock(&scif_info.eplock);
}
diff --git a/drivers/misc/mic/scif/scif_epd.h b/drivers/misc/mic/scif/scif_epd.h
index 4dc4ccd..1771d7a 100644
--- a/drivers/misc/mic/scif/scif_epd.h
+++ b/drivers/misc/mic/scif/scif_epd.h
@@ -98,6 +98,8 @@ struct scif_endpt_qp_info {
* @conn_async_state: Async connection
* @conn_pend_wq: Used by poll while waiting for incoming connections
* @conn_list: List of async connection requests
+ * @rma_info: Information for triggering SCIF RMA and DMA operations
+ * @mmu_list: link to list of MMU notifier cleanup work
* @anon: anonymous file for use in kernel mode scif poll
*/
struct scif_endpt {
@@ -129,6 +131,8 @@ struct scif_endpt {
int conn_async_state;
wait_queue_head_t conn_pend_wq;
struct list_head conn_list;
+ struct scif_endpt_rma_info rma_info;
+ struct list_head mmu_list;
struct file *anon;
};
@@ -137,6 +141,27 @@ static inline int scifdev_alive(struct scif_endpt *ep)
return _scifdev_alive(ep->remote_dev);
}
+/*
+ * scif_verify_epd:
+ * ep: SCIF endpoint
+ *
+ * Checks several generic error conditions and returns the
+ * appropriate error.
+ */
+static inline int scif_verify_epd(struct scif_endpt *ep)
+{
+ if (ep->state == SCIFEP_DISCONNECTED)
+ return -ECONNRESET;
+
+ if (ep->state != SCIFEP_CONNECTED)
+ return -ENOTCONN;
+
+ if (!scifdev_alive(ep))
+ return -ENODEV;
+
+ return 0;
+}
+
static inline int scif_anon_inode_getfile(scif_epd_t epd)
{
epd->anon = anon_inode_getfile("scif", &scif_anon_fops, NULL, 0);
@@ -177,6 +202,9 @@ void scif_clientsend(struct scif_dev *scifdev, struct scifmsg *msg);
void scif_clientrcvd(struct scif_dev *scifdev, struct scifmsg *msg);
int __scif_connect(scif_epd_t epd, struct scif_port_id *dst, bool non_block);
int __scif_flush(scif_epd_t epd);
+int scif_mmap(struct vm_area_struct *vma, scif_epd_t epd);
unsigned int __scif_pollfd(struct file *f, poll_table *wait,
struct scif_endpt *ep);
+int __scif_pin_pages(void *addr, size_t len, int *out_prot,
+ int map_flags, scif_pinned_pages_t *pages);
#endif /* SCIF_EPD_H */
diff --git a/drivers/misc/mic/scif/scif_fd.c b/drivers/misc/mic/scif/scif_fd.c
index 24e47f7..7203d9a 100644
--- a/drivers/misc/mic/scif/scif_fd.c
+++ b/drivers/misc/mic/scif/scif_fd.c
@@ -34,6 +34,13 @@ static int scif_fdclose(struct inode *inode, struct file *f)
return scif_close(priv);
}
+static int scif_fdmmap(struct file *f, struct vm_area_struct *vma)
+{
+ struct scif_endpt *priv = f->private_data;
+
+ return scif_mmap(vma, priv);
+}
+
static unsigned int scif_fdpoll(struct file *f, poll_table *wait)
{
struct scif_endpt *priv = f->private_data;
@@ -298,6 +305,157 @@ getnodes_err1:
getnodes_err2:
return err;
}
+ case SCIF_REG:
+ {
+ struct scif_endpt *priv = f->private_data;
+ struct scifioctl_reg reg;
+ off_t ret;
+
+ if (copy_from_user(®, argp, sizeof(reg))) {
+ err = -EFAULT;
+ goto reg_err;
+ }
+ if (reg.flags & SCIF_MAP_KERNEL) {
+ err = -EINVAL;
+ goto reg_err;
+ }
+ ret = scif_register(priv, (void *)reg.addr, reg.len,
+ reg.offset, reg.prot, reg.flags);
+ if (ret < 0) {
+ err = (int)ret;
+ goto reg_err;
+ }
+
+ if (copy_to_user(&((struct scifioctl_reg __user *)argp)
+ ->out_offset, &ret, sizeof(reg.out_offset))) {
+ err = -EFAULT;
+ goto reg_err;
+ }
+ err = 0;
+reg_err:
+ scif_err_debug(err, "scif_register");
+ return err;
+ }
+ case SCIF_UNREG:
+ {
+ struct scif_endpt *priv = f->private_data;
+ struct scifioctl_unreg unreg;
+
+ if (copy_from_user(&unreg, argp, sizeof(unreg))) {
+ err = -EFAULT;
+ goto unreg_err;
+ }
+ err = scif_unregister(priv, unreg.offset, unreg.len);
+unreg_err:
+ scif_err_debug(err, "scif_unregister");
+ return err;
+ }
+ case SCIF_READFROM:
+ {
+ struct scif_endpt *priv = f->private_data;
+ struct scifioctl_copy copy;
+
+ if (copy_from_user(©, argp, sizeof(copy))) {
+ err = -EFAULT;
+ goto readfrom_err;
+ }
+ err = scif_readfrom(priv, copy.loffset, copy.len, copy.roffset,
+ copy.flags);
+readfrom_err:
+ scif_err_debug(err, "scif_readfrom");
+ return err;
+ }
+ case SCIF_WRITETO:
+ {
+ struct scif_endpt *priv = f->private_data;
+ struct scifioctl_copy copy;
+
+ if (copy_from_user(©, argp, sizeof(copy))) {
+ err = -EFAULT;
+ goto writeto_err;
+ }
+ err = scif_writeto(priv, copy.loffset, copy.len, copy.roffset,
+ copy.flags);
+writeto_err:
+ scif_err_debug(err, "scif_writeto");
+ return err;
+ }
+ case SCIF_VREADFROM:
+ {
+ struct scif_endpt *priv = f->private_data;
+ struct scifioctl_copy copy;
+
+ if (copy_from_user(©, argp, sizeof(copy))) {
+ err = -EFAULT;
+ goto vreadfrom_err;
+ }
+ err = scif_vreadfrom(priv, (void __force *)copy.addr, copy.len,
+ copy.roffset, copy.flags);
+vreadfrom_err:
+ scif_err_debug(err, "scif_vreadfrom");
+ return err;
+ }
+ case SCIF_VWRITETO:
+ {
+ struct scif_endpt *priv = f->private_data;
+ struct scifioctl_copy copy;
+
+ if (copy_from_user(©, argp, sizeof(copy))) {
+ err = -EFAULT;
+ goto vwriteto_err;
+ }
+ err = scif_vwriteto(priv, (void __force *)copy.addr, copy.len,
+ copy.roffset, copy.flags);
+vwriteto_err:
+ scif_err_debug(err, "scif_vwriteto");
+ return err;
+ }
+ case SCIF_FENCE_MARK:
+ {
+ struct scif_endpt *priv = f->private_data;
+ struct scifioctl_fence_mark mark;
+ int tmp_mark = 0;
+
+ if (copy_from_user(&mark, argp, sizeof(mark))) {
+ err = -EFAULT;
+ goto fence_mark_err;
+ }
+ err = scif_fence_mark(priv, mark.flags, &tmp_mark);
+ if (err)
+ goto fence_mark_err;
+ if (copy_to_user((void __user *)mark.mark, &tmp_mark,
+ sizeof(tmp_mark))) {
+ err = -EFAULT;
+ goto fence_mark_err;
+ }
+fence_mark_err:
+ scif_err_debug(err, "scif_fence_mark");
+ return err;
+ }
+ case SCIF_FENCE_WAIT:
+ {
+ struct scif_endpt *priv = f->private_data;
+
+ err = scif_fence_wait(priv, arg);
+ scif_err_debug(err, "scif_fence_wait");
+ return err;
+ }
+ case SCIF_FENCE_SIGNAL:
+ {
+ struct scif_endpt *priv = f->private_data;
+ struct scifioctl_fence_signal signal;
+
+ if (copy_from_user(&signal, argp, sizeof(signal))) {
+ err = -EFAULT;
+ goto fence_signal_err;
+ }
+
+ err = scif_fence_signal(priv, signal.loff, signal.lval,
+ signal.roff, signal.rval, signal.flags);
+fence_signal_err:
+ scif_err_debug(err, "scif_fence_signal");
+ return err;
+ }
}
return -EINVAL;
}
@@ -306,6 +464,7 @@ const struct file_operations scif_fops = {
.open = scif_fdopen,
.release = scif_fdclose,
.unlocked_ioctl = scif_fdioctl,
+ .mmap = scif_fdmmap,
.poll = scif_fdpoll,
.flush = scif_fdflush,
.owner = THIS_MODULE,
diff --git a/drivers/misc/mic/scif/scif_main.c b/drivers/misc/mic/scif/scif_main.c
index f90bd06..c9f3446 100644
--- a/drivers/misc/mic/scif/scif_main.c
+++ b/drivers/misc/mic/scif/scif_main.c
@@ -34,6 +34,7 @@ struct scif_info scif_info = {
};
struct scif_dev *scif_dev;
+struct kmem_cache *unaligned_cache;
static atomic_t g_loopb_cnt;
/* Runs in the context of intr_wq */
@@ -264,26 +265,43 @@ static int _scif_init(void)
int rc;
spin_lock_init(&scif_info.eplock);
+ spin_lock_init(&scif_info.rmalock);
spin_lock_init(&scif_info.nb_connect_lock);
spin_lock_init(&scif_info.port_lock);
mutex_init(&scif_info.conflock);
mutex_init(&scif_info.connlock);
+ mutex_init(&scif_info.fencelock);
INIT_LIST_HEAD(&scif_info.uaccept);
INIT_LIST_HEAD(&scif_info.listen);
INIT_LIST_HEAD(&scif_info.zombie);
INIT_LIST_HEAD(&scif_info.connected);
INIT_LIST_HEAD(&scif_info.disconnected);
+ INIT_LIST_HEAD(&scif_info.rma);
+ INIT_LIST_HEAD(&scif_info.rma_tc);
+ INIT_LIST_HEAD(&scif_info.mmu_notif_cleanup);
+ INIT_LIST_HEAD(&scif_info.fence);
INIT_LIST_HEAD(&scif_info.nb_connect_list);
init_waitqueue_head(&scif_info.exitwq);
+ scif_info.rma_tc_limit = SCIF_RMA_TEMP_CACHE_LIMIT;
scif_info.en_msg_log = 0;
scif_info.p2p_enable = 1;
rc = scif_setup_scifdev();
if (rc)
goto error;
+ unaligned_cache = kmem_cache_create("Unaligned_DMA",
+ SCIF_KMEM_UNALIGNED_BUF_SIZE,
+ 0, SLAB_HWCACHE_ALIGN, NULL);
+ if (!unaligned_cache) {
+ rc = -ENOMEM;
+ goto free_sdev;
+ }
INIT_WORK(&scif_info.misc_work, scif_misc_handler);
+ INIT_WORK(&scif_info.mmu_notif_work, scif_mmu_notif_handler);
INIT_WORK(&scif_info.conn_work, scif_conn_handler);
idr_init(&scif_ports);
return 0;
+free_sdev:
+ scif_destroy_scifdev();
error:
return rc;
}
@@ -291,6 +309,7 @@ error:
static void _scif_exit(void)
{
idr_destroy(&scif_ports);
+ kmem_cache_destroy(unaligned_cache);
scif_destroy_scifdev();
}
diff --git a/drivers/misc/mic/scif/scif_main.h b/drivers/misc/mic/scif/scif_main.h
index b0795b0..b9d1194 100644
--- a/drivers/misc/mic/scif/scif_main.h
+++ b/drivers/misc/mic/scif/scif_main.h
@@ -22,16 +22,18 @@
#include <linux/pci.h>
#include <linux/miscdevice.h>
#include <linux/dmaengine.h>
+#include <linux/iova.h>
#include <linux/anon_inodes.h>
#include <linux/file.h>
+#include <linux/vmalloc.h>
#include <linux/scif.h>
-
#include "../common/mic_dev.h"
#define SCIF_MGMT_NODE 0
#define SCIF_DEFAULT_WATCHDOG_TO 30
#define SCIF_NODE_ACCEPT_TIMEOUT (3 * HZ)
#define SCIF_NODE_ALIVE_TIMEOUT (SCIF_DEFAULT_WATCHDOG_TO * HZ)
+#define SCIF_RMA_TEMP_CACHE_LIMIT 0x20000
/*
* Generic state used for certain node QP message exchanges
@@ -74,6 +76,14 @@ enum scif_msg_state {
* @loopb_work: Used for submitting work to loopb_wq
* @loopb_recv_q: List of messages received on the loopb_wq
* @card_initiated_exit: set when the card has initiated the exit
+ * @rmalock: Synchronize access to RMA operations
+ * @fencelock: Synchronize access to list of remote fences requested.
+ * @rma: List of temporary registered windows to be destroyed.
+ * @rma_tc: List of temporary registered & cached Windows to be destroyed
+ * @fence: List of remote fence requests
+ * @mmu_notif_work: Work for registration caching MMU notifier workqueue
+ * @mmu_notif_cleanup: List of temporary cached windows for reg cache
+ * @rma_tc_limit: RMA temporary cache limit
*/
struct scif_info {
u8 nodeid;
@@ -103,6 +113,14 @@ struct scif_info {
struct work_struct loopb_work;
struct list_head loopb_recv_q;
bool card_initiated_exit;
+ spinlock_t rmalock;
+ struct mutex fencelock;
+ struct list_head rma;
+ struct list_head rma_tc;
+ struct list_head fence;
+ struct work_struct mmu_notif_work;
+ struct list_head mmu_notif_cleanup;
+ unsigned long rma_tc_limit;
};
/*
@@ -153,6 +171,8 @@ struct scif_p2p_info {
* @disconn_rescnt: Keeps track of number of node remove requests sent
* @exit: Status of exit message
* @qp_dma_addr: Queue pair DMA address passed to the peer
+ * @dma_ch_idx: Round robin index for DMA channels
+ * @signal_pool: DMA pool used for scheduling scif_fence_signal DMA's
*/
struct scif_dev {
u8 node;
@@ -179,8 +199,12 @@ struct scif_dev {
atomic_t disconn_rescnt;
enum scif_msg_state exit;
dma_addr_t qp_dma_addr;
+ int dma_ch_idx;
+ struct dma_pool *signal_pool;
};
+extern bool scif_reg_cache_enable;
+extern bool scif_ulimit_check;
extern struct scif_info scif_info;
extern struct idr scif_ports;
extern struct bus_type scif_peer_bus;
@@ -192,6 +216,8 @@ extern const struct file_operations scif_anon_fops;
#define SCIF_NODE_QP_SIZE 0x10000
#include "scif_nodeqp.h"
+#include "scif_rma.h"
+#include "scif_rma_list.h"
/*
* scifdev_self:
diff --git a/drivers/misc/mic/scif/scif_map.h b/drivers/misc/mic/scif/scif_map.h
index 20e50b4..3e86360 100644
--- a/drivers/misc/mic/scif/scif_map.h
+++ b/drivers/misc/mic/scif/scif_map.h
@@ -80,7 +80,7 @@ scif_unmap_single(dma_addr_t local, struct scif_dev *scifdev,
size_t size)
{
if (!scifdev_self(scifdev)) {
- if (scifdev_is_p2p(scifdev) && local > scifdev->base_addr)
+ if (scifdev_is_p2p(scifdev))
local = local - scifdev->base_addr;
dma_unmap_single(&scifdev->sdev->dev, local,
size, DMA_BIDIRECTIONAL);
@@ -110,4 +110,27 @@ scif_iounmap(void *virt, size_t len, struct scif_dev *scifdev)
sdev->hw_ops->iounmap(sdev, (void __force __iomem *)virt);
}
}
+
+static __always_inline int
+scif_map_page(dma_addr_t *dma_handle, struct page *page,
+ struct scif_dev *scifdev)
+{
+ int err = 0;
+
+ if (scifdev_self(scifdev)) {
+ *dma_handle = page_to_phys(page);
+ } else {
+ struct scif_hw_dev *sdev = scifdev->sdev;
+ *dma_handle = dma_map_page(&sdev->dev,
+ page, 0x0, PAGE_SIZE,
+ DMA_BIDIRECTIONAL);
+ if (dma_mapping_error(&sdev->dev, *dma_handle))
+ err = -ENOMEM;
+ else if (scifdev_is_p2p(scifdev))
+ *dma_handle = *dma_handle + scifdev->base_addr;
+ }
+ if (err)
+ *dma_handle = 0;
+ return err;
+}
#endif /* SCIF_MAP_H */
diff --git a/drivers/misc/mic/scif/scif_nm.c b/drivers/misc/mic/scif/scif_nm.c
index f1b1b97..79f26a0 100644
--- a/drivers/misc/mic/scif/scif_nm.c
+++ b/drivers/misc/mic/scif/scif_nm.c
@@ -34,6 +34,7 @@ static void scif_invalidate_ep(int node)
list_for_each_safe(pos, tmpq, &scif_info.disconnected) {
ep = list_entry(pos, struct scif_endpt, list);
if (ep->remote_dev->node == node) {
+ scif_unmap_all_windows(ep);
spin_lock(&ep->lock);
scif_cleanup_ep_qp(ep);
spin_unlock(&ep->lock);
@@ -50,6 +51,7 @@ static void scif_invalidate_ep(int node)
wake_up_interruptible(&ep->sendwq);
wake_up_interruptible(&ep->recvwq);
spin_unlock(&ep->lock);
+ scif_unmap_all_windows(ep);
}
}
mutex_unlock(&scif_info.connlock);
@@ -61,8 +63,8 @@ void scif_free_qp(struct scif_dev *scifdev)
if (!qp)
return;
- scif_free_coherent((void *)qp->inbound_q.rb_base,
- qp->local_buf, scifdev, qp->inbound_q.size);
+ scif_unmap_single(qp->local_buf, scifdev, qp->inbound_q.size);
+ kfree(qp->inbound_q.rb_base);
scif_unmap_single(qp->local_qp, scifdev, sizeof(struct scif_qp));
kfree(scifdev->qpairs);
scifdev->qpairs = NULL;
@@ -125,8 +127,12 @@ void scif_cleanup_scifdev(struct scif_dev *dev)
}
scif_destroy_intr_wq(dev);
}
+ flush_work(&scif_info.misc_work);
scif_destroy_p2p(dev);
scif_invalidate_ep(dev->node);
+ scif_zap_mmaps(dev->node);
+ scif_cleanup_rma_for_zombies(dev->node);
+ flush_work(&scif_info.misc_work);
scif_send_acks(dev);
if (!dev->node && scif_info.card_initiated_exit) {
/*
diff --git a/drivers/misc/mic/scif/scif_nodeqp.c b/drivers/misc/mic/scif/scif_nodeqp.c
index 5cdb9f0..e278917 100644
--- a/drivers/misc/mic/scif/scif_nodeqp.c
+++ b/drivers/misc/mic/scif/scif_nodeqp.c
@@ -552,7 +552,29 @@ static char *message_types[] = {"BAD",
"DISCNT_ACK",
"CLIENT_SENT",
"CLIENT_RCVD",
- "SCIF_GET_NODE_INFO"};
+ "SCIF_GET_NODE_INFO",
+ "REGISTER",
+ "REGISTER_ACK",
+ "REGISTER_NACK",
+ "UNREGISTER",
+ "UNREGISTER_ACK",
+ "UNREGISTER_NACK",
+ "ALLOC_REQ",
+ "ALLOC_GNT",
+ "ALLOC_REJ",
+ "FREE_PHYS",
+ "FREE_VIRT",
+ "MUNMAP",
+ "MARK",
+ "MARK_ACK",
+ "MARK_NACK",
+ "WAIT",
+ "WAIT_ACK",
+ "WAIT_NACK",
+ "SIGNAL_LOCAL",
+ "SIGNAL_REMOTE",
+ "SIG_ACK",
+ "SIG_NACK"};
static void
scif_display_message(struct scif_dev *scifdev, struct scifmsg *msg,
@@ -646,10 +668,16 @@ int scif_nodeqp_send(struct scif_dev *scifdev, struct scifmsg *msg)
*
* Work queue handler for servicing miscellaneous SCIF tasks.
* Examples include:
- * 1) Cleanup of zombie endpoints.
+ * 1) Remote fence requests.
+ * 2) Destruction of temporary registered windows
+ * created during scif_vreadfrom()/scif_vwriteto().
+ * 3) Cleanup of zombie endpoints.
*/
void scif_misc_handler(struct work_struct *work)
{
+ scif_rma_handle_remote_fences();
+ scif_rma_destroy_windows();
+ scif_rma_destroy_tcw_invalid();
scif_cleanup_zombie_epd();
}
@@ -995,6 +1023,27 @@ static void (*scif_intr_func[SCIF_MAX_MSG + 1])
scif_clientsend, /* SCIF_CLIENT_SENT */
scif_clientrcvd, /* SCIF_CLIENT_RCVD */
scif_get_node_info_resp,/* SCIF_GET_NODE_INFO */
+ scif_recv_reg, /* SCIF_REGISTER */
+ scif_recv_reg_ack, /* SCIF_REGISTER_ACK */
+ scif_recv_reg_nack, /* SCIF_REGISTER_NACK */
+ scif_recv_unreg, /* SCIF_UNREGISTER */
+ scif_recv_unreg_ack, /* SCIF_UNREGISTER_ACK */
+ scif_recv_unreg_nack, /* SCIF_UNREGISTER_NACK */
+ scif_alloc_req, /* SCIF_ALLOC_REQ */
+ scif_alloc_gnt_rej, /* SCIF_ALLOC_GNT */
+ scif_alloc_gnt_rej, /* SCIF_ALLOC_REJ */
+ scif_free_virt, /* SCIF_FREE_VIRT */
+ scif_recv_munmap, /* SCIF_MUNMAP */
+ scif_recv_mark, /* SCIF_MARK */
+ scif_recv_mark_resp, /* SCIF_MARK_ACK */
+ scif_recv_mark_resp, /* SCIF_MARK_NACK */
+ scif_recv_wait, /* SCIF_WAIT */
+ scif_recv_wait_resp, /* SCIF_WAIT_ACK */
+ scif_recv_wait_resp, /* SCIF_WAIT_NACK */
+ scif_recv_sig_local, /* SCIF_SIG_LOCAL */
+ scif_recv_sig_remote, /* SCIF_SIG_REMOTE */
+ scif_recv_sig_resp, /* SCIF_SIG_ACK */
+ scif_recv_sig_resp, /* SCIF_SIG_NACK */
};
/**
diff --git a/drivers/misc/mic/scif/scif_nodeqp.h b/drivers/misc/mic/scif/scif_nodeqp.h
index 6c0ed67..9589627 100644
--- a/drivers/misc/mic/scif/scif_nodeqp.h
+++ b/drivers/misc/mic/scif/scif_nodeqp.h
@@ -74,7 +74,28 @@
#define SCIF_CLIENT_SENT 16 /* Notify the peer that data has been written */
#define SCIF_CLIENT_RCVD 17 /* Notify the peer that data has been read */
#define SCIF_GET_NODE_INFO 18 /* Get current node mask from the mgmt node*/
-#define SCIF_MAX_MSG SCIF_GET_NODE_INFO
+#define SCIF_REGISTER 19 /* Tell peer about a new registered window */
+#define SCIF_REGISTER_ACK 20 /* Notify peer about unregistration success */
+#define SCIF_REGISTER_NACK 21 /* Notify peer about registration success */
+#define SCIF_UNREGISTER 22 /* Tell peer about unregistering a window */
+#define SCIF_UNREGISTER_ACK 23 /* Notify peer about registration failure */
+#define SCIF_UNREGISTER_NACK 24 /* Notify peer about unregistration failure */
+#define SCIF_ALLOC_REQ 25 /* Request a mapped buffer */
+#define SCIF_ALLOC_GNT 26 /* Notify peer about allocation success */
+#define SCIF_ALLOC_REJ 27 /* Notify peer about allocation failure */
+#define SCIF_FREE_VIRT 28 /* Free previously allocated virtual memory */
+#define SCIF_MUNMAP 29 /* Acknowledgment for a SCIF_MMAP request */
+#define SCIF_MARK 30 /* SCIF Remote Fence Mark Request */
+#define SCIF_MARK_ACK 31 /* SCIF Remote Fence Mark Success */
+#define SCIF_MARK_NACK 32 /* SCIF Remote Fence Mark Failure */
+#define SCIF_WAIT 33 /* SCIF Remote Fence Wait Request */
+#define SCIF_WAIT_ACK 34 /* SCIF Remote Fence Wait Success */
+#define SCIF_WAIT_NACK 35 /* SCIF Remote Fence Wait Failure */
+#define SCIF_SIG_LOCAL 36 /* SCIF Remote Fence Local Signal Request */
+#define SCIF_SIG_REMOTE 37 /* SCIF Remote Fence Remote Signal Request */
+#define SCIF_SIG_ACK 38 /* SCIF Remote Fence Remote Signal Success */
+#define SCIF_SIG_NACK 39 /* SCIF Remote Fence Remote Signal Failure */
+#define SCIF_MAX_MSG SCIF_SIG_NACK
/*
* struct scifmsg - Node QP message format
@@ -92,6 +113,24 @@ struct scifmsg {
} __packed;
/*
+ * struct scif_allocmsg - Used with SCIF_ALLOC_REQ to request
+ * the remote note to allocate memory
+ *
+ * phys_addr: Physical address of the buffer
+ * vaddr: Virtual address of the buffer
+ * size: Size of the buffer
+ * state: Current state
+ * allocwq: wait queue for status
+ */
+struct scif_allocmsg {
+ dma_addr_t phys_addr;
+ unsigned long vaddr;
+ size_t size;
+ enum scif_msg_state state;
+ wait_queue_head_t allocwq;
+};
+
+/*
* struct scif_qp - Node Queue Pair
*
* Interesting structure -- a little difficult because we can only
@@ -158,7 +197,6 @@ int scif_setup_qp_connect_response(struct scif_dev *scifdev,
int scif_setup_loopback_qp(struct scif_dev *scifdev);
int scif_destroy_loopback_qp(struct scif_dev *scifdev);
void scif_poll_qp_state(struct work_struct *work);
-void scif_qp_response_ack(struct work_struct *work);
void scif_destroy_p2p(struct scif_dev *scifdev);
void scif_send_exit(struct scif_dev *scifdev);
static inline struct device *scif_get_peer_dev(struct scif_dev *scifdev)
diff --git a/drivers/misc/mic/scif/scif_peer_bus.c b/drivers/misc/mic/scif/scif_peer_bus.c
index 547bf7b..6ffa3bd 100644
--- a/drivers/misc/mic/scif/scif_peer_bus.c
+++ b/drivers/misc/mic/scif/scif_peer_bus.c
@@ -72,6 +72,7 @@ err:
static int scif_peer_add_device(struct scif_dev *scifdev)
{
struct scif_peer_dev *spdev = rcu_dereference(scifdev->spdev);
+ char pool_name[16];
int ret;
ret = device_add(&spdev->dev);
@@ -81,8 +82,21 @@ static int scif_peer_add_device(struct scif_dev *scifdev)
"dnode %d: peer device_add failed\n", scifdev->node);
goto put_spdev;
}
+
+ scnprintf(pool_name, sizeof(pool_name), "scif-%d", spdev->dnode);
+ scifdev->signal_pool = dmam_pool_create(pool_name, &scifdev->sdev->dev,
+ sizeof(struct scif_status), 1,
+ 0);
+ if (!scifdev->signal_pool) {
+ dev_err(&scifdev->sdev->dev,
+ "dnode %d: dmam_pool_create failed\n", scifdev->node);
+ ret = -ENOMEM;
+ goto del_spdev;
+ }
dev_dbg(&spdev->dev, "Added peer dnode %d\n", spdev->dnode);
return 0;
+del_spdev:
+ device_del(&spdev->dev);
put_spdev:
RCU_INIT_POINTER(scifdev->spdev, NULL);
synchronize_rcu();
--
2.0.0.rc3.2.g998f840
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
[toc] | [prev] | [next] | [standalone]
| From | Ashutosh Dixit <ashutosh.dixit@intel.com> |
|---|---|
| Date | 2015-07-28 01:40 +0200 |
| Subject | [PATCH char-misc-next 18/19] misc: mic: SCIF fence |
| Message-ID | <pR09Y-8vE-13@gated-at.bofh.it> |
| In reply to | #1193568 |
From: Sudeep Dutt <sudeep.dutt@intel.com>
This patch implements the fence APIs required to synchronize
DMAs. SCIF provides an interface to return a "mark" for all DMAs
programmed at the instant the API was called. Users can then "wait" on
the mark provided previously by blocking inside the kernel. Upon
receipt of a DMA completion interrupt the waiting thread is woken
up. There is also an interface to signal DMA completion by polling for
a location to be updated via a "signal" cookie to avoid the interrupt
overhead in the mark/wait interface. SCIF allows programming fences on
both the local and the remote node for both the mark/wait or the fence
signal APIs.
Reviewed-by: Ashutosh Dixit <ashutosh.dixit@intel.com>
Reviewed-by: Nikhil Rao <nikhil.rao@intel.com>
Signed-off-by: Sudeep Dutt <sudeep.dutt@intel.com>
---
drivers/misc/mic/scif/scif_fence.c | 773 +++++++++++++++++++++++++++++++++++++
1 file changed, 773 insertions(+)
create mode 100644 drivers/misc/mic/scif/scif_fence.c
diff --git a/drivers/misc/mic/scif/scif_fence.c b/drivers/misc/mic/scif/scif_fence.c
new file mode 100644
index 0000000..1620aaa
--- /dev/null
+++ b/drivers/misc/mic/scif/scif_fence.c
@@ -0,0 +1,773 @@
+/*
+ * Intel MIC Platform Software Stack (MPSS)
+ *
+ * Copyright(c) 2015 Intel Corporation.
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License, version 2, as
+ * published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful, but
+ * WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+ * General Public License for more details.
+ *
+ * Intel SCIF driver.
+ *
+ */
+
+#include "scif_main.h"
+
+/**
+ * scif_recv_mark: Handle SCIF_MARK request
+ * @msg: Interrupt message
+ *
+ * The peer has requested a mark.
+ */
+void scif_recv_mark(struct scif_dev *scifdev, struct scifmsg *msg)
+{
+ struct scif_endpt *ep = (struct scif_endpt *)msg->payload[0];
+ int mark, err;
+
+ err = _scif_fence_mark(ep, &mark);
+ if (err)
+ msg->uop = SCIF_MARK_NACK;
+ else
+ msg->uop = SCIF_MARK_ACK;
+ msg->payload[0] = ep->remote_ep;
+ msg->payload[2] = mark;
+ scif_nodeqp_send(ep->remote_dev, msg);
+}
+
+/**
+ * scif_recv_mark_resp: Handle SCIF_MARK_(N)ACK messages.
+ * @msg: Interrupt message
+ *
+ * The peer has responded to a SCIF_MARK message.
+ */
+void scif_recv_mark_resp(struct scif_dev *scifdev, struct scifmsg *msg)
+{
+ struct scif_endpt *ep = (struct scif_endpt *)msg->payload[0];
+ struct scif_fence_info *fence_req =
+ (struct scif_fence_info *)msg->payload[1];
+
+ mutex_lock(&ep->rma_info.rma_lock);
+ if (SCIF_MARK_ACK == msg->uop) {
+ fence_req->state = OP_COMPLETED;
+ fence_req->dma_mark = (int)msg->payload[2];
+ } else {
+ fence_req->state = OP_FAILED;
+ }
+ mutex_unlock(&ep->rma_info.rma_lock);
+ complete(&fence_req->comp);
+}
+
+/**
+ * scif_recv_wait: Handle SCIF_WAIT request
+ * @msg: Interrupt message
+ *
+ * The peer has requested waiting on a fence.
+ */
+void scif_recv_wait(struct scif_dev *scifdev, struct scifmsg *msg)
+{
+ struct scif_endpt *ep = (struct scif_endpt *)msg->payload[0];
+ struct scif_remote_fence_info *fence;
+
+ /*
+ * Allocate structure for remote fence information and
+ * send a NACK if the allocation failed. The peer will
+ * return ENOMEM upon receiving a NACK.
+ */
+ fence = kmalloc(sizeof(*fence), GFP_KERNEL);
+ if (!fence) {
+ msg->payload[0] = ep->remote_ep;
+ msg->uop = SCIF_WAIT_NACK;
+ scif_nodeqp_send(ep->remote_dev, msg);
+ return;
+ }
+
+ /* Prepare the fence request */
+ memcpy(&fence->msg, msg, sizeof(struct scifmsg));
+ INIT_LIST_HEAD(&fence->list);
+
+ /* Insert to the global remote fence request list */
+ mutex_lock(&scif_info.fencelock);
+ atomic_inc(&ep->rma_info.fence_refcount);
+ list_add_tail(&fence->list, &scif_info.fence);
+ mutex_unlock(&scif_info.fencelock);
+
+ schedule_work(&scif_info.misc_work);
+}
+
+/**
+ * scif_recv_wait_resp: Handle SCIF_WAIT_(N)ACK messages.
+ * @msg: Interrupt message
+ *
+ * The peer has responded to a SCIF_WAIT message.
+ */
+void scif_recv_wait_resp(struct scif_dev *scifdev, struct scifmsg *msg)
+{
+ struct scif_endpt *ep = (struct scif_endpt *)msg->payload[0];
+ struct scif_fence_info *fence_req =
+ (struct scif_fence_info *)msg->payload[1];
+
+ mutex_lock(&ep->rma_info.rma_lock);
+ if (SCIF_WAIT_ACK == msg->uop)
+ fence_req->state = OP_COMPLETED;
+ else
+ fence_req->state = OP_FAILED;
+ mutex_unlock(&ep->rma_info.rma_lock);
+ complete(&fence_req->comp);
+}
+
+/**
+ * scif_recv_sig_local: Handle SCIF_SIG_LOCAL request
+ * @msg: Interrupt message
+ *
+ * The peer has requested a signal on a local offset.
+ */
+void scif_recv_sig_local(struct scif_dev *scifdev, struct scifmsg *msg)
+{
+ struct scif_endpt *ep = (struct scif_endpt *)msg->payload[0];
+ int err;
+
+ err = scif_prog_signal(ep, msg->payload[1], msg->payload[2],
+ SCIF_WINDOW_SELF);
+ if (err)
+ msg->uop = SCIF_SIG_NACK;
+ else
+ msg->uop = SCIF_SIG_ACK;
+ msg->payload[0] = ep->remote_ep;
+ scif_nodeqp_send(ep->remote_dev, msg);
+}
+
+/**
+ * scif_recv_sig_remote: Handle SCIF_SIGNAL_REMOTE request
+ * @msg: Interrupt message
+ *
+ * The peer has requested a signal on a remote offset.
+ */
+void scif_recv_sig_remote(struct scif_dev *scifdev, struct scifmsg *msg)
+{
+ struct scif_endpt *ep = (struct scif_endpt *)msg->payload[0];
+ int err;
+
+ err = scif_prog_signal(ep, msg->payload[1], msg->payload[2],
+ SCIF_WINDOW_PEER);
+ if (err)
+ msg->uop = SCIF_SIG_NACK;
+ else
+ msg->uop = SCIF_SIG_ACK;
+ msg->payload[0] = ep->remote_ep;
+ scif_nodeqp_send(ep->remote_dev, msg);
+}
+
+/**
+ * scif_recv_sig_resp: Handle SCIF_SIG_(N)ACK messages.
+ * @msg: Interrupt message
+ *
+ * The peer has responded to a signal request.
+ */
+void scif_recv_sig_resp(struct scif_dev *scifdev, struct scifmsg *msg)
+{
+ struct scif_endpt *ep = (struct scif_endpt *)msg->payload[0];
+ struct scif_fence_info *fence_req =
+ (struct scif_fence_info *)msg->payload[3];
+
+ mutex_lock(&ep->rma_info.rma_lock);
+ if (SCIF_SIG_ACK == msg->uop)
+ fence_req->state = OP_COMPLETED;
+ else
+ fence_req->state = OP_FAILED;
+ mutex_unlock(&ep->rma_info.rma_lock);
+ complete(&fence_req->comp);
+}
+
+static inline void *scif_get_local_va(off_t off, struct scif_window *window)
+{
+ struct page **pages = window->pinned_pages->pages;
+ int page_nr = (off - window->offset) >> PAGE_SHIFT;
+ off_t page_off = off & ~PAGE_MASK;
+
+ return page_address(pages[page_nr]) + page_off;
+}
+
+static void scif_prog_signal_cb(void *arg)
+{
+ struct scif_status *status = arg;
+
+ dma_pool_free(status->ep->remote_dev->signal_pool, status,
+ status->src_dma_addr);
+}
+
+static int _scif_prog_signal(scif_epd_t epd, dma_addr_t dst, u64 val)
+{
+ struct scif_endpt *ep = (struct scif_endpt *)epd;
+ struct dma_chan *chan = ep->rma_info.dma_chan;
+ struct dma_device *ddev = chan->device;
+ bool x100 = !is_dma_copy_aligned(chan->device, 1, 1, 1);
+ struct dma_async_tx_descriptor *tx;
+ struct scif_status *status = NULL;
+ dma_addr_t src;
+ dma_cookie_t cookie;
+ int err, flags = 0;
+
+ if (x100) {
+ /*
+ * For X100 use the status descriptor to write the value to
+ * the destination. For this the value itself is passed as
+ * the source address into the DMA engine API
+ */
+ src = val;
+ } else {
+ status = dma_pool_alloc(ep->remote_dev->signal_pool, GFP_KERNEL,
+ &src);
+ if (!status) {
+ err = -ENOMEM;
+ dev_err(&ep->remote_dev->sdev->dev, "%s %d err %d\n",
+ __func__, __LINE__, err);
+ goto alloc_fail;
+ }
+ status->val = val;
+ status->src_dma_addr = src;
+ status->ep = ep;
+ src += offsetof(struct scif_status, val);
+ flags = DMA_PREP_INTERRUPT;
+ }
+
+ tx = ddev->device_prep_dma_memcpy(chan, 0, 0, 0, DMA_PREP_FENCE);
+ if (!tx) {
+ err = -ENOMEM;
+ dev_err(&ep->remote_dev->sdev->dev, "%s %d err %d\n",
+ __func__, __LINE__, err);
+ goto dma_fail;
+ }
+ cookie = tx->tx_submit(tx);
+ if (dma_submit_error(cookie)) {
+ err = (int)cookie;
+ dev_err(&ep->remote_dev->sdev->dev, "%s %d err %d\n",
+ __func__, __LINE__, err);
+ goto dma_fail;
+ }
+ dma_async_issue_pending(chan);
+ tx = ddev->device_prep_dma_memcpy(chan, dst, src, sizeof(val), flags);
+ if (!tx) {
+ err = -ENOMEM;
+ dev_err(&ep->remote_dev->sdev->dev, "%s %d err %d\n",
+ __func__, __LINE__, err);
+ goto dma_fail;
+ }
+ if (!x100) {
+ tx->callback = scif_prog_signal_cb;
+ tx->callback_param = status;
+ }
+ cookie = tx->tx_submit(tx);
+ if (dma_submit_error(cookie)) {
+ err = -EIO;
+ dev_err(&ep->remote_dev->sdev->dev, "%s %d err %d\n",
+ __func__, __LINE__, err);
+ goto dma_fail;
+ }
+ dma_async_issue_pending(chan);
+ return 0;
+dma_fail:
+ if (!x100)
+ dma_pool_free(status->ep->remote_dev->signal_pool, status,
+ status->src_dma_addr);
+alloc_fail:
+ return err;
+}
+
+/*
+ * scif_prog_signal:
+ * @epd - Endpoint Descriptor
+ * @offset - registered address to write @val to
+ * @val - Value to be written at @offset
+ * @type - Type of the window.
+ *
+ * Arrange to write a value to the registered offset after ensuring that the
+ * offset provided is indeed valid.
+ */
+int scif_prog_signal(scif_epd_t epd, off_t offset, u64 val,
+ enum scif_window_type type)
+{
+ struct scif_endpt *ep = (struct scif_endpt *)epd;
+ struct scif_window *window = NULL;
+ struct scif_rma_req req;
+ dma_addr_t dst_dma_addr;
+ int err;
+
+ mutex_lock(&ep->rma_info.rma_lock);
+ req.out_window = &window;
+ req.offset = offset;
+ req.nr_bytes = sizeof(u64);
+ req.prot = SCIF_PROT_WRITE;
+ req.type = SCIF_WINDOW_SINGLE;
+ if (SCIF_WINDOW_SELF == type)
+ req.head = &ep->rma_info.reg_list;
+ else
+ req.head = &ep->rma_info.remote_reg_list;
+ /* Does a valid window exist? */
+ err = scif_query_window(&req);
+ if (err) {
+ dev_err(scif_info.mdev.this_device,
+ "%s %d err %d\n", __func__, __LINE__, err);
+ goto unlock_ret;
+ }
+
+ if (scif_is_mgmt_node() && scifdev_self(ep->remote_dev)) {
+ u64 *dst_virt;
+
+ if (SCIF_WINDOW_SELF == type)
+ dst_virt = scif_get_local_va(offset, window);
+ else
+ dst_virt =
+ scif_get_local_va(offset, (struct scif_window *)
+ window->peer_window);
+ *dst_virt = val;
+ } else {
+ dst_dma_addr = __scif_off_to_dma_addr(window, offset);
+ err = _scif_prog_signal(epd, dst_dma_addr, val);
+ }
+unlock_ret:
+ mutex_unlock(&ep->rma_info.rma_lock);
+ return err;
+}
+
+static int _scif_fence_wait(scif_epd_t epd, int mark)
+{
+ struct scif_endpt *ep = (struct scif_endpt *)epd;
+ dma_cookie_t cookie = mark & ~SCIF_REMOTE_FENCE;
+ int err;
+
+ /* Wait for DMA callback in scif_fence_mark_cb(..) */
+ err = wait_event_interruptible_timeout(ep->rma_info.markwq,
+ dma_async_is_tx_complete(
+ ep->rma_info.dma_chan,
+ cookie, NULL, NULL) ==
+ DMA_COMPLETE,
+ SCIF_NODE_ALIVE_TIMEOUT);
+ if (!err)
+ err = -ETIMEDOUT;
+ else if (err > 0)
+ err = 0;
+ return err;
+}
+
+/**
+ * scif_rma_handle_remote_fences:
+ *
+ * This routine services remote fence requests.
+ */
+void scif_rma_handle_remote_fences(void)
+{
+ struct list_head *item, *tmp;
+ struct scif_remote_fence_info *fence;
+ struct scif_endpt *ep;
+ int mark, err;
+
+ might_sleep();
+ mutex_lock(&scif_info.fencelock);
+ list_for_each_safe(item, tmp, &scif_info.fence) {
+ fence = list_entry(item, struct scif_remote_fence_info,
+ list);
+ /* Remove fence from global list */
+ list_del(&fence->list);
+
+ /* Initiate the fence operation */
+ ep = (struct scif_endpt *)fence->msg.payload[0];
+ mark = fence->msg.payload[2];
+ err = _scif_fence_wait(ep, mark);
+ if (err)
+ fence->msg.uop = SCIF_WAIT_NACK;
+ else
+ fence->msg.uop = SCIF_WAIT_ACK;
+ fence->msg.payload[0] = ep->remote_ep;
+ scif_nodeqp_send(ep->remote_dev, &fence->msg);
+ kfree(fence);
+ if (!atomic_sub_return(1, &ep->rma_info.fence_refcount))
+ schedule_work(&scif_info.misc_work);
+ }
+ mutex_unlock(&scif_info.fencelock);
+}
+
+static int _scif_send_fence(scif_epd_t epd, int uop, int mark, int *out_mark)
+{
+ int err;
+ struct scifmsg msg;
+ struct scif_fence_info *fence_req;
+ struct scif_endpt *ep = (struct scif_endpt *)epd;
+
+ fence_req = kmalloc(sizeof(*fence_req), GFP_KERNEL);
+ if (!fence_req) {
+ err = -ENOMEM;
+ goto error;
+ }
+
+ fence_req->state = OP_IN_PROGRESS;
+ init_completion(&fence_req->comp);
+
+ msg.src = ep->port;
+ msg.uop = uop;
+ msg.payload[0] = ep->remote_ep;
+ msg.payload[1] = (u64)fence_req;
+ if (SCIF_WAIT == uop)
+ msg.payload[2] = mark;
+ spin_lock(&ep->lock);
+ if (SCIFEP_CONNECTED == ep->state)
+ err = scif_nodeqp_send(ep->remote_dev, &msg);
+ else
+ err = -ENOTCONN;
+ spin_unlock(&ep->lock);
+ if (err)
+ goto error_free;
+retry:
+ /* Wait for a SCIF_WAIT_(N)ACK message */
+ err = wait_for_completion_timeout(&fence_req->comp,
+ SCIF_NODE_ALIVE_TIMEOUT);
+ if (!err && scifdev_alive(ep))
+ goto retry;
+ if (!err)
+ err = -ENODEV;
+ if (err > 0)
+ err = 0;
+ mutex_lock(&ep->rma_info.rma_lock);
+ if (err < 0) {
+ if (OP_IN_PROGRESS == fence_req->state)
+ fence_req->state = OP_FAILED;
+ }
+ if (OP_FAILED == fence_req->state && !err)
+ err = -ENOMEM;
+ if (SCIF_MARK == uop && OP_COMPLETED == fence_req->state)
+ *out_mark = SCIF_REMOTE_FENCE | fence_req->dma_mark;
+ mutex_unlock(&ep->rma_info.rma_lock);
+error_free:
+ kfree(fence_req);
+error:
+ return err;
+}
+
+/**
+ * scif_send_fence_mark:
+ * @epd: end point descriptor.
+ * @out_mark: Output DMA mark reported by peer.
+ *
+ * Send a remote fence mark request.
+ */
+static int scif_send_fence_mark(scif_epd_t epd, int *out_mark)
+{
+ return _scif_send_fence(epd, SCIF_MARK, 0, out_mark);
+}
+
+/**
+ * scif_send_fence_wait:
+ * @epd: end point descriptor.
+ * @mark: DMA mark to wait for.
+ *
+ * Send a remote fence wait request.
+ */
+static int scif_send_fence_wait(scif_epd_t epd, int mark)
+{
+ return _scif_send_fence(epd, SCIF_WAIT, mark, NULL);
+}
+
+static int _scif_send_fence_signal_wait(struct scif_endpt *ep,
+ struct scif_fence_info *fence_req)
+{
+ int err;
+
+retry:
+ /* Wait for a SCIF_SIG_(N)ACK message */
+ err = wait_for_completion_timeout(&fence_req->comp,
+ SCIF_NODE_ALIVE_TIMEOUT);
+ if (!err && scifdev_alive(ep))
+ goto retry;
+ if (!err)
+ err = -ENODEV;
+ if (err > 0)
+ err = 0;
+ if (err < 0) {
+ mutex_lock(&ep->rma_info.rma_lock);
+ if (OP_IN_PROGRESS == fence_req->state)
+ fence_req->state = OP_FAILED;
+ mutex_unlock(&ep->rma_info.rma_lock);
+ }
+ if (OP_FAILED == fence_req->state && !err)
+ err = -ENXIO;
+ return err;
+}
+
+/**
+ * scif_send_fence_signal:
+ * @epd - endpoint descriptor
+ * @loff - local offset
+ * @lval - local value to write to loffset
+ * @roff - remote offset
+ * @rval - remote value to write to roffset
+ * @flags - flags
+ *
+ * Sends a remote fence signal request
+ */
+static int scif_send_fence_signal(scif_epd_t epd, off_t roff, u64 rval,
+ off_t loff, u64 lval, int flags)
+{
+ int err = 0;
+ struct scifmsg msg;
+ struct scif_fence_info *fence_req;
+ struct scif_endpt *ep = (struct scif_endpt *)epd;
+
+ fence_req = kmalloc(sizeof(*fence_req), GFP_KERNEL);
+ if (!fence_req) {
+ err = -ENOMEM;
+ goto error;
+ }
+
+ fence_req->state = OP_IN_PROGRESS;
+ init_completion(&fence_req->comp);
+ msg.src = ep->port;
+ if (flags & SCIF_SIGNAL_LOCAL) {
+ msg.uop = SCIF_SIG_LOCAL;
+ msg.payload[0] = ep->remote_ep;
+ msg.payload[1] = roff;
+ msg.payload[2] = rval;
+ msg.payload[3] = (u64)fence_req;
+ spin_lock(&ep->lock);
+ if (SCIFEP_CONNECTED == ep->state)
+ err = scif_nodeqp_send(ep->remote_dev, &msg);
+ else
+ err = -ENOTCONN;
+ spin_unlock(&ep->lock);
+ if (err)
+ goto error_free;
+ err = _scif_send_fence_signal_wait(ep, fence_req);
+ if (err)
+ goto error_free;
+ }
+ fence_req->state = OP_IN_PROGRESS;
+
+ if (flags & SCIF_SIGNAL_REMOTE) {
+ msg.uop = SCIF_SIG_REMOTE;
+ msg.payload[0] = ep->remote_ep;
+ msg.payload[1] = loff;
+ msg.payload[2] = lval;
+ msg.payload[3] = (u64)fence_req;
+ spin_lock(&ep->lock);
+ if (SCIFEP_CONNECTED == ep->state)
+ err = scif_nodeqp_send(ep->remote_dev, &msg);
+ else
+ err = -ENOTCONN;
+ spin_unlock(&ep->lock);
+ if (err)
+ goto error_free;
+ err = _scif_send_fence_signal_wait(ep, fence_req);
+ }
+error_free:
+ kfree(fence_req);
+error:
+ return err;
+}
+
+static void scif_fence_mark_cb(void *arg)
+{
+ struct scif_endpt *ep = (struct scif_endpt *)arg;
+
+ wake_up_interruptible(&ep->rma_info.markwq);
+ atomic_dec(&ep->rma_info.fence_refcount);
+}
+
+/*
+ * _scif_fence_mark:
+ *
+ * @epd - endpoint descriptor
+ * Set up a mark for this endpoint and return the value of the mark.
+ */
+int _scif_fence_mark(scif_epd_t epd, int *mark)
+{
+ struct scif_endpt *ep = (struct scif_endpt *)epd;
+ struct dma_chan *chan = ep->rma_info.dma_chan;
+ struct dma_device *ddev = chan->device;
+ struct dma_async_tx_descriptor *tx;
+ dma_cookie_t cookie;
+ int err;
+
+ tx = ddev->device_prep_dma_memcpy(chan, 0, 0, 0, DMA_PREP_FENCE);
+ if (!tx) {
+ err = -ENOMEM;
+ dev_err(&ep->remote_dev->sdev->dev, "%s %d err %d\n",
+ __func__, __LINE__, err);
+ return err;
+ }
+ cookie = tx->tx_submit(tx);
+ if (dma_submit_error(cookie)) {
+ err = (int)cookie;
+ dev_err(&ep->remote_dev->sdev->dev, "%s %d err %d\n",
+ __func__, __LINE__, err);
+ return err;
+ }
+ dma_async_issue_pending(chan);
+ tx = ddev->device_prep_dma_interrupt(chan, DMA_PREP_INTERRUPT);
+ if (!tx) {
+ err = -ENOMEM;
+ dev_err(&ep->remote_dev->sdev->dev, "%s %d err %d\n",
+ __func__, __LINE__, err);
+ return err;
+ }
+ tx->callback = scif_fence_mark_cb;
+ tx->callback_param = ep;
+ *mark = cookie = tx->tx_submit(tx);
+ if (dma_submit_error(cookie)) {
+ err = (int)cookie;
+ dev_err(&ep->remote_dev->sdev->dev, "%s %d err %d\n",
+ __func__, __LINE__, err);
+ return err;
+ }
+ atomic_inc(&ep->rma_info.fence_refcount);
+ dma_async_issue_pending(chan);
+ return 0;
+}
+
+#define SCIF_LOOPB_MAGIC_MARK 0xdead
+
+int scif_fence_mark(scif_epd_t epd, int flags, int *mark)
+{
+ struct scif_endpt *ep = (struct scif_endpt *)epd;
+ int err = 0;
+
+ dev_dbg(scif_info.mdev.this_device,
+ "SCIFAPI fence_mark: ep %p flags 0x%x mark 0x%x\n",
+ ep, flags, *mark);
+ err = scif_verify_epd(ep);
+ if (err)
+ return err;
+
+ /* Invalid flags? */
+ if (flags & ~(SCIF_FENCE_INIT_SELF | SCIF_FENCE_INIT_PEER))
+ return -EINVAL;
+
+ /* At least one of init self or peer RMA should be set */
+ if (!(flags & (SCIF_FENCE_INIT_SELF | SCIF_FENCE_INIT_PEER)))
+ return -EINVAL;
+
+ /* Exactly one of init self or peer RMA should be set but not both */
+ if ((flags & SCIF_FENCE_INIT_SELF) && (flags & SCIF_FENCE_INIT_PEER))
+ return -EINVAL;
+
+ /*
+ * Management node loopback does not need to use DMA.
+ * Return a valid mark to be symmetric.
+ */
+ if (scifdev_self(ep->remote_dev) && scif_is_mgmt_node()) {
+ *mark = SCIF_LOOPB_MAGIC_MARK;
+ return 0;
+ }
+
+ if (flags & SCIF_FENCE_INIT_SELF)
+ err = _scif_fence_mark(epd, mark);
+ else
+ err = scif_send_fence_mark(ep, mark);
+
+ if (err)
+ dev_err(scif_info.mdev.this_device,
+ "%s %d err %d\n", __func__, __LINE__, err);
+ dev_dbg(scif_info.mdev.this_device,
+ "SCIFAPI fence_mark: ep %p flags 0x%x mark 0x%x err %d\n",
+ ep, flags, *mark, err);
+ return err;
+}
+EXPORT_SYMBOL_GPL(scif_fence_mark);
+
+int scif_fence_wait(scif_epd_t epd, int mark)
+{
+ struct scif_endpt *ep = (struct scif_endpt *)epd;
+ int err = 0;
+
+ dev_dbg(scif_info.mdev.this_device,
+ "SCIFAPI fence_wait: ep %p mark 0x%x\n",
+ ep, mark);
+ err = scif_verify_epd(ep);
+ if (err)
+ return err;
+ /*
+ * Management node loopback does not need to use DMA.
+ * The only valid mark provided is 0 so simply
+ * return success if the mark is valid.
+ */
+ if (scifdev_self(ep->remote_dev) && scif_is_mgmt_node()) {
+ if (SCIF_LOOPB_MAGIC_MARK == mark)
+ return 0;
+ else
+ return -EINVAL;
+ }
+ if (mark & SCIF_REMOTE_FENCE)
+ err = scif_send_fence_wait(epd, mark);
+ else
+ err = _scif_fence_wait(epd, mark);
+ if (err < 0)
+ dev_err(scif_info.mdev.this_device,
+ "%s %d err %d\n", __func__, __LINE__, err);
+ return err;
+}
+EXPORT_SYMBOL_GPL(scif_fence_wait);
+
+int scif_fence_signal(scif_epd_t epd, off_t loff, u64 lval,
+ off_t roff, u64 rval, int flags)
+{
+ struct scif_endpt *ep = (struct scif_endpt *)epd;
+ int err = 0;
+
+ dev_dbg(scif_info.mdev.this_device,
+ "SCIFAPI fence_signal: ep %p loff 0x%lx lval 0x%llx roff 0x%lx rval 0x%llx flags 0x%x\n",
+ ep, loff, lval, roff, rval, flags);
+ err = scif_verify_epd(ep);
+ if (err)
+ return err;
+
+ /* Invalid flags? */
+ if (flags & ~(SCIF_FENCE_INIT_SELF | SCIF_FENCE_INIT_PEER |
+ SCIF_SIGNAL_LOCAL | SCIF_SIGNAL_REMOTE))
+ return -EINVAL;
+
+ /* At least one of init self or peer RMA should be set */
+ if (!(flags & (SCIF_FENCE_INIT_SELF | SCIF_FENCE_INIT_PEER)))
+ return -EINVAL;
+
+ /* Exactly one of init self or peer RMA should be set but not both */
+ if ((flags & SCIF_FENCE_INIT_SELF) && (flags & SCIF_FENCE_INIT_PEER))
+ return -EINVAL;
+
+ /* At least one of SCIF_SIGNAL_LOCAL or SCIF_SIGNAL_REMOTE required */
+ if (!(flags & (SCIF_SIGNAL_LOCAL | SCIF_SIGNAL_REMOTE)))
+ return -EINVAL;
+
+ /* Only Dword offsets allowed */
+ if ((flags & SCIF_SIGNAL_LOCAL) && (loff & (sizeof(u32) - 1)))
+ return -EINVAL;
+
+ /* Only Dword aligned offsets allowed */
+ if ((flags & SCIF_SIGNAL_REMOTE) && (roff & (sizeof(u32) - 1)))
+ return -EINVAL;
+
+ if (flags & SCIF_FENCE_INIT_PEER) {
+ err = scif_send_fence_signal(epd, roff, rval, loff,
+ lval, flags);
+ } else {
+ /* Local Signal in Local RAS */
+ if (flags & SCIF_SIGNAL_LOCAL) {
+ err = scif_prog_signal(epd, loff, lval,
+ SCIF_WINDOW_SELF);
+ if (err)
+ goto error_ret;
+ }
+
+ /* Signal in Remote RAS */
+ if (flags & SCIF_SIGNAL_REMOTE)
+ err = scif_prog_signal(epd, roff,
+ rval, SCIF_WINDOW_PEER);
+ }
+error_ret:
+ if (err)
+ dev_err(scif_info.mdev.this_device,
+ "%s %d err %d\n", __func__, __LINE__, err);
+ return err;
+}
+EXPORT_SYMBOL_GPL(scif_fence_signal);
--
2.0.0.rc3.2.g998f840
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
[toc] | [prev] | [next] | [standalone]
| From | Ashutosh Dixit <ashutosh.dixit@intel.com> |
|---|---|
| Date | 2015-07-28 01:40 +0200 |
| Subject | [PATCH char-misc-next 12/19] misc: mic: SCIF RMA header file and IOCTL changes |
| Message-ID | <pR09Y-8vE-27@gated-at.bofh.it> |
| In reply to | #1193568 |
From: Sudeep Dutt <sudeep.dutt@intel.com>
This patch updates the SCIF header file and IOCTL interface with the
changes required to support RMAs. APIs added include the ability to
pin pages and register those pages with SCIF. SCIF kernel clients can
also add references to remote registered pages and access them via the
CPU. The user space IOCTL interface has been updated to enable SCIF
registration, RDMA/CPU copies and fence APIs for RDMA synchronization.
Reviewed-by: Nikhil Rao <nikhil.rao@intel.com>
Reviewed-by: Ashutosh Dixit <ashutosh.dixit@intel.com>
Signed-off-by: Sudeep Dutt <sudeep.dutt@intel.com>
---
include/linux/scif.h | 232 ++++++++++++++++++++++++++++++++++++++--
include/uapi/linux/scif_ioctl.h | 85 +++++++++++++++
2 files changed, 307 insertions(+), 10 deletions(-)
diff --git a/include/linux/scif.h b/include/linux/scif.h
index fd62c05..bd7b722 100644
--- a/include/linux/scif.h
+++ b/include/linux/scif.h
@@ -93,6 +93,25 @@ enum {
#define SCIF_PORT_RSVD 1088
typedef struct scif_endpt *scif_epd_t;
+typedef struct scif_pinned_pages *scif_pinned_pages_t;
+
+/**
+ * struct scif_range - SCIF registered range used in kernel mode
+ * @nr_pages: number of pages of PAGE_SIZE
+ * @prot_flags: R/W protection
+ * @phys_addr: Array of bus addresses
+ * @va: Array of kernel virtual addresses backed by the pages in the phys_addr
+ * array. The va is populated only when called on the host for a remote
+ * SCIF connection on MIC. This is required to support the use case of DMA
+ * between MIC and another device which is not a SCIF node e.g., an IB or
+ * ethernet NIC.
+ */
+struct scif_range {
+ int nr_pages;
+ int prot_flags;
+ dma_addr_t *phys_addr;
+ void __iomem **va;
+};
/**
* struct scif_pollepd - SCIF endpoint to be monitored via scif_poll
@@ -389,7 +408,6 @@ int scif_close(scif_epd_t epd);
* Errors:
* EBADF, ENOTTY - epd is not a valid endpoint descriptor
* ECONNRESET - Connection reset by peer
- * EFAULT - An invalid address was specified for a parameter
* EINVAL - flags is invalid, or len is negative
* ENODEV - The remote node is lost or existed, but is not currently in the
* network since it may have crashed
@@ -442,7 +460,6 @@ int scif_send(scif_epd_t epd, void *msg, int len, int flags);
* EAGAIN - The destination node is returning from a low power state
* EBADF, ENOTTY - epd is not a valid endpoint descriptor
* ECONNRESET - Connection reset by peer
- * EFAULT - An invalid address was specified for a parameter
* EINVAL - flags is invalid, or len is negative
* ENODEV - The remote node is lost or existed, but is not currently in the
* network since it may have crashed
@@ -505,9 +522,6 @@ int scif_recv(scif_epd_t epd, void *msg, int len, int flags);
* SCIF_PROT_READ - allow read operations from the window
* SCIF_PROT_WRITE - allow write operations to the window
*
- * The map_flags argument can be set to SCIF_MAP_FIXED which interprets a
- * fixed offset.
- *
* Return:
* Upon successful completion, scif_register() returns the offset at which the
* mapping was placed (po); otherwise in user mode SCIF_REGISTER_FAILED (that
@@ -520,7 +534,6 @@ int scif_recv(scif_epd_t epd, void *msg, int len, int flags);
* EAGAIN - The mapping could not be performed due to lack of resources
* EBADF, ENOTTY - epd is not a valid endpoint descriptor
* ECONNRESET - Connection reset by peer
- * EFAULT - Addresses in the range [addr, addr + len - 1] are invalid
* EINVAL - map_flags is invalid, or prot_flags is invalid, or SCIF_MAP_FIXED is
* set in flags, and offset is not a multiple of the page size, or addr is not a
* multiple of the page size, or len is not a multiple of the page size, or is
@@ -803,7 +816,6 @@ int scif_writeto(scif_epd_t epd, off_t loffset, size_t len, off_t
* EACCESS - Attempt to write to a read-only range
* EBADF, ENOTTY - epd is not a valid endpoint descriptor
* ECONNRESET - Connection reset by peer
- * EFAULT - Addresses in the range [addr, addr + len - 1] are invalid
* EINVAL - rma_flags is invalid
* ENODEV - The remote node is lost or existed, but is not currently in the
* network since it may have crashed
@@ -884,7 +896,6 @@ int scif_vreadfrom(scif_epd_t epd, void *addr, size_t len, off_t roffset,
* EACCESS - Attempt to write to a read-only range
* EBADF, ENOTTY - epd is not a valid endpoint descriptor
* ECONNRESET - Connection reset by peer
- * EFAULT - Addresses in the range [addr, addr + len - 1] are invalid
* EINVAL - rma_flags is invalid
* ENODEV - The remote node is lost or existed, but is not currently in the
* network since it may have crashed
@@ -1028,11 +1039,212 @@ int scif_fence_signal(scif_epd_t epd, off_t loff, u64 lval, off_t roff,
* online nodes in the SCIF network including 'self'; otherwise in user mode
* -1 is returned and errno is set to indicate the error; in kernel mode no
* errors are returned.
+ */
+int scif_get_node_ids(u16 *nodes, int len, u16 *self);
+
+/**
+ * scif_pin_pages() - Pin a set of pages
+ * @addr: Virtual address of range to pin
+ * @len: Length of range to pin
+ * @prot_flags: Page protection flags
+ * @map_flags: Page classification flags
+ * @pinned_pages: Handle to pinned pages
+ *
+ * scif_pin_pages() pins (locks in physical memory) the physical pages which
+ * back the range of virtual address pages starting at addr and continuing for
+ * len bytes. addr and len are constrained to be multiples of the page size. A
+ * successful scif_pin_pages() call returns a handle to pinned_pages which may
+ * be used in subsequent calls to scif_register_pinned_pages().
+ *
+ * The pages will remain pinned as long as there is a reference against the
+ * scif_pinned_pages_t value returned by scif_pin_pages() and until
+ * scif_unpin_pages() is called, passing the scif_pinned_pages_t value. A
+ * reference is added to a scif_pinned_pages_t value each time a window is
+ * created by calling scif_register_pinned_pages() and passing the
+ * scif_pinned_pages_t value. A reference is removed from a
+ * scif_pinned_pages_t value each time such a window is deleted.
+ *
+ * Subsequent operations which change the memory pages to which virtual
+ * addresses are mapped (such as mmap(), munmap()) have no effect on the
+ * scif_pinned_pages_t value or windows created against it.
+ *
+ * If the process will fork(), it is recommended that the registered
+ * virtual address range be marked with MADV_DONTFORK. Doing so will prevent
+ * problems due to copy-on-write semantics.
+ *
+ * The prot_flags argument is formed by OR'ing together one or more of the
+ * following values.
+ * SCIF_PROT_READ - allow read operations against the pages
+ * SCIF_PROT_WRITE - allow write operations against the pages
+ * The map_flags argument can be set as SCIF_MAP_KERNEL to interpret addr as a
+ * kernel space address. By default, addr is interpreted as a user space
+ * address.
+ *
+ * Return:
+ * Upon successful completion, scif_pin_pages() returns 0; otherwise the
+ * negative of one of the following errors is returned.
*
* Errors:
- * EFAULT - Bad address
+ * EINVAL - prot_flags is invalid, map_flags is invalid, or offset is negative
+ * ENOMEM - Not enough space
*/
-int scif_get_node_ids(u16 *nodes, int len, u16 *self);
+int scif_pin_pages(void *addr, size_t len, int prot_flags, int map_flags,
+ scif_pinned_pages_t *pinned_pages);
+
+/**
+ * scif_unpin_pages() - Unpin a set of pages
+ * @pinned_pages: Handle to pinned pages to be unpinned
+ *
+ * scif_unpin_pages() prevents scif_register_pinned_pages() from registering new
+ * windows against pinned_pages. The physical pages represented by pinned_pages
+ * will remain pinned until all windows previously registered against
+ * pinned_pages are deleted (the window is scif_unregister()'d and all
+ * references to the window are removed (see scif_unregister()).
+ *
+ * pinned_pages must have been obtain from a previous call to scif_pin_pages().
+ * After calling scif_unpin_pages(), it is an error to pass pinned_pages to
+ * scif_register_pinned_pages().
+ *
+ * Return:
+ * Upon successful completion, scif_unpin_pages() returns 0; otherwise the
+ * negative of one of the following errors is returned.
+ *
+ * Errors:
+ * EINVAL - pinned_pages is not valid
+ */
+int scif_unpin_pages(scif_pinned_pages_t pinned_pages);
+
+/**
+ * scif_register_pinned_pages() - Mark a memory region for remote access.
+ * @epd: endpoint descriptor
+ * @pinned_pages: Handle to pinned pages
+ * @offset: Registered address space offset
+ * @map_flags: Flags which control where pages are mapped
+ *
+ * The scif_register_pinned_pages() function opens a window, a range of whole
+ * pages of the registered address space of the endpoint epd, starting at
+ * offset po. The value of po, further described below, is a function of the
+ * parameters offset and pinned_pages, and the value of map_flags. Each page of
+ * the window represents a corresponding physical memory page of the range
+ * represented by pinned_pages; the length of the window is the same as the
+ * length of range represented by pinned_pages. A successful
+ * scif_register_pinned_pages() call returns po as the return value.
+ *
+ * When SCIF_MAP_FIXED is set in the map_flags argument, po will be offset
+ * exactly, and offset is constrained to be a multiple of the page size. The
+ * mapping established by scif_register_pinned_pages() will not replace any
+ * existing registration; an error is returned if any page of the new window
+ * would intersect an existing window.
+ *
+ * When SCIF_MAP_FIXED is not set, the implementation uses offset in an
+ * implementation-defined manner to arrive at po. The po so chosen will be an
+ * area of the registered address space that the implementation deems suitable
+ * for a mapping of the required size. An offset value of 0 is interpreted as
+ * granting the implementation complete freedom in selecting po, subject to
+ * constraints described below. A non-zero value of offset is taken to be a
+ * suggestion of an offset near which the mapping should be placed. When the
+ * implementation selects a value for po, it does not replace any extant
+ * window. In all cases, po will be a multiple of the page size.
+ *
+ * The physical pages which are so represented by a window are available for
+ * access in calls to scif_get_pages(), scif_readfrom(), scif_writeto(),
+ * scif_vreadfrom(), and scif_vwriteto(). While a window is registered, the
+ * physical pages represented by the window will not be reused by the memory
+ * subsystem for any other purpose. Note that the same physical page may be
+ * represented by multiple windows.
+ *
+ * Windows created by scif_register_pinned_pages() are unregistered by
+ * scif_unregister().
+ *
+ * The map_flags argument can be set to SCIF_MAP_FIXED which interprets a
+ * fixed offset.
+ *
+ * Return:
+ * Upon successful completion, scif_register_pinned_pages() returns the offset
+ * at which the mapping was placed (po); otherwise the negative of one of the
+ * following errors is returned.
+ *
+ * Errors:
+ * EADDRINUSE - SCIF_MAP_FIXED is set in map_flags and pages in the new window
+ * would intersect an existing window
+ * EAGAIN - The mapping could not be performed due to lack of resources
+ * ECONNRESET - Connection reset by peer
+ * EINVAL - map_flags is invalid, or SCIF_MAP_FIXED is set in map_flags, and
+ * offset is not a multiple of the page size, or offset is negative
+ * ENODEV - The remote node is lost or existed, but is not currently in the
+ * network since it may have crashed
+ * ENOMEM - Not enough space
+ * ENOTCONN - The endpoint is not connected
+ */
+off_t scif_register_pinned_pages(scif_epd_t epd,
+ scif_pinned_pages_t pinned_pages,
+ off_t offset, int map_flags);
+
+/**
+ * scif_get_pages() - Add references to remote registered pages
+ * @epd: endpoint descriptor
+ * @offset: remote registered offset
+ * @len: length of range of pages
+ * @pages: returned scif_range structure
+ *
+ * scif_get_pages() returns the addresses of the physical pages represented by
+ * those pages of the registered address space of the peer of epd, starting at
+ * offset and continuing for len bytes. offset and len are constrained to be
+ * multiples of the page size.
+ *
+ * All of the pages in the specified range [offset, offset + len - 1] must be
+ * within a single window of the registered address space of the peer of epd.
+ *
+ * The addresses are returned as a virtually contiguous array pointed to by the
+ * phys_addr component of the scif_range structure whose address is returned in
+ * pages. The nr_pages component of scif_range is the length of the array. The
+ * prot_flags component of scif_range holds the protection flag value passed
+ * when the pages were registered.
+ *
+ * Each physical page whose address is returned by scif_get_pages() remains
+ * available and will not be released for reuse until the scif_range structure
+ * is returned in a call to scif_put_pages(). The scif_range structure returned
+ * by scif_get_pages() must be unmodified.
+ *
+ * It is an error to call scif_close() on an endpoint on which a scif_range
+ * structure of that endpoint has not been returned to scif_put_pages().
+ *
+ * Return:
+ * Upon successful completion, scif_get_pages() returns 0; otherwise the
+ * negative of one of the following errors is returned.
+ * Errors:
+ * ECONNRESET - Connection reset by peer.
+ * EINVAL - offset is not a multiple of the page size, or offset is negative, or
+ * len is not a multiple of the page size
+ * ENODEV - The remote node is lost or existed, but is not currently in the
+ * network since it may have crashed
+ * ENOTCONN - The endpoint is not connected
+ * ENXIO - Offsets in the range [offset, offset + len - 1] are invalid
+ * for the registered address space of the peer epd
+ */
+int scif_get_pages(scif_epd_t epd, off_t offset, size_t len,
+ struct scif_range **pages);
+
+/**
+ * scif_put_pages() - Remove references from remote registered pages
+ * @pages: pages to be returned
+ *
+ * scif_put_pages() releases a scif_range structure previously obtained by
+ * calling scif_get_pages(). The physical pages represented by pages may
+ * be reused when the window which represented those pages is unregistered.
+ * Therefore, those pages must not be accessed after calling scif_put_pages().
+ *
+ * Return:
+ * Upon successful completion, scif_put_pages() returns 0; otherwise the
+ * negative of one of the following errors is returned.
+ * Errors:
+ * EINVAL - pages does not point to a valid scif_range structure, or
+ * the scif_range structure pointed to by pages was already returned
+ * ENODEV - The remote node is lost or existed, but is not currently in the
+ * network since it may have crashed
+ * ENOTCONN - The endpoint is not connected
+ */
+int scif_put_pages(struct scif_range *pages);
/**
* scif_poll() - Wait for some event on an endpoint
diff --git a/include/uapi/linux/scif_ioctl.h b/include/uapi/linux/scif_ioctl.h
index 4a94d91..d904891 100644
--- a/include/uapi/linux/scif_ioctl.h
+++ b/include/uapi/linux/scif_ioctl.h
@@ -107,6 +107,82 @@ struct scifioctl_msg {
};
/**
+ * struct scifioctl_reg - used for SCIF_REG IOCTL
+ * @addr: starting virtual address
+ * @len: length of range
+ * @offset: offset of window
+ * @prot: read/write protection
+ * @flags: flags
+ * @out_offset: offset returned
+ */
+struct scifioctl_reg {
+ __u64 addr;
+ __u64 len;
+ __s64 offset;
+ __s32 prot;
+ __s32 flags;
+ __s64 out_offset;
+};
+
+/**
+ * struct scifioctl_unreg - used for SCIF_UNREG IOCTL
+ * @offset: start of range to unregister
+ * @len: length of range to unregister
+ */
+struct scifioctl_unreg {
+ __s64 offset;
+ __u64 len;
+};
+
+/**
+ * struct scifioctl_copy - used for SCIF DMA copy IOCTLs
+ *
+ * @loffset: offset in local registered address space to/from
+ * which to copy
+ * @len: length of range to copy
+ * @roffset: offset in remote registered address space to/from
+ * which to copy
+ * @addr: user virtual address to/from which to copy
+ * @flags: flags
+ *
+ * This structure is used for SCIF_READFROM, SCIF_WRITETO, SCIF_VREADFROM
+ * and SCIF_VREADFROM IOCTL's.
+ */
+struct scifioctl_copy {
+ __s64 loffset;
+ __u64 len;
+ __s64 roffset;
+ __u64 addr;
+ __s32 flags;
+};
+
+/**
+ * struct scifioctl_fence_mark - used for SCIF_FENCE_MARK IOCTL
+ * @flags: flags
+ * @mark: fence handle which is a pointer to a __s32
+ */
+struct scifioctl_fence_mark {
+ __s32 flags;
+ __u64 mark;
+};
+
+/**
+ * struct scifioctl_fence_signal - used for SCIF_FENCE_SIGNAL IOCTL
+ * @loff: local offset
+ * @lval: value to write to loffset
+ * @roff: remote offset
+ * @rval: value to write to roffset
+ * @flags: flags
+ */
+struct scifioctl_fence_signal {
+ __s64 loff;
+ __u64 lval;
+ __s64 roff;
+ __u64 rval;
+ __s32 flags;
+};
+
+/**
* struct scifioctl_node_ids - used for SCIF_GET_NODEIDS IOCTL
* @nodes: pointer to an array of node_ids
* @self: ID of the current node
@@ -125,6 +201,15 @@ struct scifioctl_node_ids {
#define SCIF_ACCEPTREG _IOWR('s', 5, __u64)
#define SCIF_SEND _IOWR('s', 6, struct scifioctl_msg)
#define SCIF_RECV _IOWR('s', 7, struct scifioctl_msg)
+#define SCIF_REG _IOWR('s', 8, struct scifioctl_reg)
+#define SCIF_UNREG _IOWR('s', 9, struct scifioctl_unreg)
+#define SCIF_READFROM _IOWR('s', 10, struct scifioctl_copy)
+#define SCIF_WRITETO _IOWR('s', 11, struct scifioctl_copy)
+#define SCIF_VREADFROM _IOWR('s', 12, struct scifioctl_copy)
+#define SCIF_VWRITETO _IOWR('s', 13, struct scifioctl_copy)
#define SCIF_GET_NODEIDS _IOWR('s', 14, struct scifioctl_node_ids)
+#define SCIF_FENCE_MARK _IOWR('s', 15, struct scifioctl_fence_mark)
+#define SCIF_FENCE_WAIT _IOWR('s', 16, __s32)
+#define SCIF_FENCE_SIGNAL _IOWR('s', 17, struct scifioctl_fence_signal)
#endif /* SCIF_IOCTL_H */
--
2.0.0.rc3.2.g998f840
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
[toc] | [prev] | [next] | [standalone]
| From | Ashutosh Dixit <ashutosh.dixit@intel.com> |
|---|---|
| Date | 2015-07-28 01:40 +0200 |
| Subject | [PATCH char-misc-next 10/19] lib: convert iova.c into a library |
| Message-ID | <pR09Y-8vE-29@gated-at.bofh.it> |
| In reply to | #1193568 |
From: Harish Chegondi <harish.chegondi@intel.com>
This patch converts iova.c into a library, moving it from
drivers/iommu/ to lib/, and exports its virtual address allocation and
management functions so that other modules can reuse them.
Cc: Joerg Roedel <joro@8bytes.org>
Reviewed-by: Anil S Keshavamurthy <anil.s.keshavamurthy@intel.com>
Reviewed-by: Sudeep Dutt <sudeep.dutt@intel.com>
Signed-off-by: Harish Chegondi <harish.chegondi@intel.com>
---
drivers/iommu/Kconfig | 5 +----
drivers/iommu/Makefile | 1 -
lib/Kconfig | 6 ++++++
lib/Makefile | 2 ++
{drivers/iommu => lib}/iova.c | 9 +++++++++
5 files changed, 18 insertions(+), 5 deletions(-)
rename {drivers/iommu => lib}/iova.c (97%)
diff --git a/drivers/iommu/Kconfig b/drivers/iommu/Kconfig
index f1fb1d3..c711889 100644
--- a/drivers/iommu/Kconfig
+++ b/drivers/iommu/Kconfig
@@ -41,9 +41,6 @@ config IOMMU_IO_PGTABLE_LPAE_SELFTEST
endmenu
-config IOMMU_IOVA
- bool
-
config OF_IOMMU
def_bool y
depends on OF && IOMMU_API
@@ -125,8 +122,8 @@ config INTEL_IOMMU
bool "Support for Intel IOMMU using DMA Remapping Devices"
depends on PCI_MSI && ACPI && (X86 || IA64_GENERIC)
select IOMMU_API
- select IOMMU_IOVA
select DMAR_TABLE
+ select IOVA
help
DMA remapping (DMAR) devices support enables independent address
translations for Direct Memory Access (DMA) from devices.
diff --git a/drivers/iommu/Makefile b/drivers/iommu/Makefile
index c6dcc51..fbb1372 100644
--- a/drivers/iommu/Makefile
+++ b/drivers/iommu/Makefile
@@ -3,7 +3,6 @@ obj-$(CONFIG_IOMMU_API) += iommu-traces.o
obj-$(CONFIG_IOMMU_API) += iommu-sysfs.o
obj-$(CONFIG_IOMMU_IO_PGTABLE) += io-pgtable.o
obj-$(CONFIG_IOMMU_IO_PGTABLE_LPAE) += io-pgtable-arm.o
-obj-$(CONFIG_IOMMU_IOVA) += iova.o
obj-$(CONFIG_OF_IOMMU) += of_iommu.o
obj-$(CONFIG_MSM_IOMMU) += msm_iommu.o msm_iommu_dev.o
obj-$(CONFIG_AMD_IOMMU) += amd_iommu.o amd_iommu_init.o
diff --git a/lib/Kconfig b/lib/Kconfig
index 3a2ef67..62ba145 100644
--- a/lib/Kconfig
+++ b/lib/Kconfig
@@ -62,6 +62,12 @@ config ARCH_USE_CMPXCHG_LOCKREF
config ARCH_HAS_FAST_MULTIPLIER
bool
+config IOVA
+ bool
+ help
+ This option enables a 64 bit virtual address allocation and management
+ library.
+
config CRC_CCITT
tristate "CRC-CCITT functions"
help
diff --git a/lib/Makefile b/lib/Makefile
index 6897b52..058722d 100644
--- a/lib/Makefile
+++ b/lib/Makefile
@@ -205,3 +205,5 @@ quiet_cmd_build_OID_registry = GEN $@
clean-files += oid_registry_data.c
obj-$(CONFIG_UCS2_STRING) += ucs2_string.o
+
+obj-$(CONFIG_IOVA) += iova.o
diff --git a/drivers/iommu/iova.c b/lib/iova.c
similarity index 97%
rename from drivers/iommu/iova.c
rename to lib/iova.c
index b7c3d92..7753006 100644
--- a/drivers/iommu/iova.c
+++ b/lib/iova.c
@@ -72,6 +72,7 @@ init_iova_domain(struct iova_domain *iovad, unsigned long granule,
iovad->start_pfn = start_pfn;
iovad->dma_32bit_pfn = pfn_32bit;
}
+EXPORT_SYMBOL_GPL(init_iova_domain);
static struct rb_node *
__get_cached_rbnode(struct iova_domain *iovad, unsigned long *limit_pfn)
@@ -281,6 +282,7 @@ alloc_iova(struct iova_domain *iovad, unsigned long size,
return new_iova;
}
+EXPORT_SYMBOL_GPL(alloc_iova);
/**
* find_iova - find's an iova for a given pfn
@@ -321,6 +323,7 @@ struct iova *find_iova(struct iova_domain *iovad, unsigned long pfn)
spin_unlock_irqrestore(&iovad->iova_rbtree_lock, flags);
return NULL;
}
+EXPORT_SYMBOL_GPL(find_iova);
/**
* __free_iova - frees the given iova
@@ -339,6 +342,7 @@ __free_iova(struct iova_domain *iovad, struct iova *iova)
spin_unlock_irqrestore(&iovad->iova_rbtree_lock, flags);
free_iova_mem(iova);
}
+EXPORT_SYMBOL_GPL(__free_iova);
/**
* free_iova - finds and frees the iova for a given pfn
@@ -356,6 +360,7 @@ free_iova(struct iova_domain *iovad, unsigned long pfn)
__free_iova(iovad, iova);
}
+EXPORT_SYMBOL_GPL(free_iova);
/**
* put_iova_domain - destroys the iova doamin
@@ -378,6 +383,7 @@ void put_iova_domain(struct iova_domain *iovad)
}
spin_unlock_irqrestore(&iovad->iova_rbtree_lock, flags);
}
+EXPORT_SYMBOL_GPL(put_iova_domain);
static int
__is_range_overlap(struct rb_node *node,
@@ -467,6 +473,7 @@ finish:
spin_unlock_irqrestore(&iovad->iova_rbtree_lock, flags);
return iova;
}
+EXPORT_SYMBOL_GPL(reserve_iova);
/**
* copy_reserved_iova - copies the reserved between domains
@@ -493,6 +500,7 @@ copy_reserved_iova(struct iova_domain *from, struct iova_domain *to)
}
spin_unlock_irqrestore(&from->iova_rbtree_lock, flags);
}
+EXPORT_SYMBOL_GPL(copy_reserved_iova);
struct iova *
split_and_remove_iova(struct iova_domain *iovad, struct iova *iova,
@@ -534,3 +542,4 @@ error:
free_iova_mem(prev);
return NULL;
}
+EXPORT_SYMBOL_GPL(split_and_remove_iova);
--
2.0.0.rc3.2.g998f840
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
[toc] | [prev] | [next] | [standalone]
| From | Joerg Roedel <joro@8bytes.org> |
|---|---|
| Date | 2015-07-28 12:10 +0200 |
| Subject | Re: [PATCH char-misc-next 10/19] lib: convert iova.c into a library |
| Message-ID | <pR9ZE-61X-31@gated-at.bofh.it> |
| In reply to | #1193585 |
On Mon, Jul 27, 2015 at 04:57:32PM -0700, Ashutosh Dixit wrote: > From: Harish Chegondi <harish.chegondi@intel.com> > > This patch converts iova.c into a library, moving it from > drivers/iommu/ to lib/, and exports its virtual address allocation and > management functions so that other modules can reuse them. > > Cc: Joerg Roedel <joro@8bytes.org> > Reviewed-by: Anil S Keshavamurthy <anil.s.keshavamurthy@intel.com> > Reviewed-by: Sudeep Dutt <sudeep.dutt@intel.com> > Signed-off-by: Harish Chegondi <harish.chegondi@intel.com> Where is this going to be used outside of the IOMMU world? Joerg -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
[toc] | [prev] | [next] | [standalone]
| From | Robin Murphy <robin.murphy@arm.com> |
|---|---|
| Date | 2015-07-28 12:50 +0200 |
| Subject | Re: [PATCH char-misc-next 10/19] lib: convert iova.c into a library |
| Message-ID | <pRaCl-6MU-1@gated-at.bofh.it> |
| In reply to | #1194001 |
On 28/07/15 11:03, Joerg Roedel wrote: > On Mon, Jul 27, 2015 at 04:57:32PM -0700, Ashutosh Dixit wrote: >> From: Harish Chegondi <harish.chegondi@intel.com> >> >> This patch converts iova.c into a library, moving it from >> drivers/iommu/ to lib/, and exports its virtual address allocation and >> management functions so that other modules can reuse them. >> >> Cc: Joerg Roedel <joro@8bytes.org> >> Reviewed-by: Anil S Keshavamurthy <anil.s.keshavamurthy@intel.com> >> Reviewed-by: Sudeep Dutt <sudeep.dutt@intel.com> >> Signed-off-by: Harish Chegondi <harish.chegondi@intel.com> > > Where is this going to be used outside of the IOMMU world? > ...and how does it relate to the patches from Sakari (+CC) doing much the same thing[1]? Having gone and fished out the main LKML thread ([2], to help anyone else missing it), I don't see any obvious dependency on the Intel IOMMU driver - what happens here if that is compiled out and hasn't called iommu_iova_cache_init() first? Robin. [1]:http://thread.gmane.org/gmane.linux.kernel.iommu/10142 [2]:http://thread.gmane.org/gmane.linux.kernel/2005895 -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
[toc] | [prev] | [next] | [standalone]
| From | Andrew Morton <akpm@linux-foundation.org> |
|---|---|
| Date | 2015-07-28 22:50 +0200 |
| Subject | Re: [PATCH char-misc-next 10/19] lib: convert iova.c into a library |
| Message-ID | <pRjZ0-3vU-23@gated-at.bofh.it> |
| In reply to | #1193585 |
On Mon, 27 Jul 2015 16:57:32 -0700 Ashutosh Dixit <ashutosh.dixit@intel.com> wrote:
> From: Harish Chegondi <harish.chegondi@intel.com>
>
> This patch converts iova.c into a library, moving it from
> drivers/iommu/ to lib/, and exports its virtual address allocation and
> management functions so that other modules can reuse them.
From the following emails it is unclear to me whether we'll actually be
going ahead with this, but whatever. It's a chance to do some code
reading.
> {drivers/iommu => lib}/iova.c | 9 +++++++++
Except the code isn't here. Stupid git. Here 'tis:
> /*
> * Copyright __ 2006-2009, Intel Corporation.
Non-ascii thing which I can't get to display correctly in anything.
Give it up, it'll never work, use "(c)".
> * This program is free software; you can redistribute it and/or modify it
> * under the terms and conditions of the GNU General Public License,
> * version 2, as published by the Free Software Foundation.
> *
> * This program is distributed in the hope it will be useful, but WITHOUT
> * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
> * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for
> * more details.
> *
> * You should have received a copy of the GNU General Public License along with
> * this program; if not, write to the Free Software Foundation, Inc., 59 Temple
> * Place - Suite 330, Boston, MA 02111-1307 USA.
> *
> * Author: Anil S Keshavamurthy <anil.s.keshavamurthy@intel.com>
> */
>
> #include <linux/iova.h>
> #include <linux/slab.h>
Wow. This file only compiles by sheer dumb luck. kernel.h?
spinlock.h? bitops.h? Some of these things come in thanks to iova.h,
others don't.
> static struct kmem_cache *iommu_iova_cache;
>
> int iommu_iova_cache_init(void)
Should actually be called iommu_iova_cache_create(). Because it's a
wrapper around kmem_cache_create(), and "create" is what it does.
> {
> int ret = 0;
>
> iommu_iova_cache = kmem_cache_create("iommu_iova",
> sizeof(struct iova),
> 0,
> SLAB_HWCACHE_ALIGN,
> NULL);
Could use KMEM_CACHE().
> if (!iommu_iova_cache) {
> pr_err("Couldn't create iova cache\n");
> ret = -ENOMEM;
> }
>
> return ret;
> }
>
> void iommu_iova_cache_destroy(void)
The naming throughout this driver is a mess. Sometimes it's
iommu_iova_<operation>. Other times it's <operation>_iova. Other
times it's <operation>_iova_<operation>. There's no thought and no
consistency.
The usual standard will work well here. Stick with iova_<operation>
and use it consistently.
So
iova_cache_create
iova_cache_destroy
iova_mem_free
iova_init_domain
iova_alloc
iova_find
iova_etcetera
> {
> kmem_cache_destroy(iommu_iova_cache);
> }
>
> struct iova *alloc_iova_mem(void)
> {
> return kmem_cache_alloc(iommu_iova_cache, GFP_ATOMIC);
> }
GFP_ATOMIC is unreliable and undesirably depletes page allocator
reserves. If it is really really the case that no caller can ever
allocate an iova in any other context then OK, but that's an
extraordinary thing and I suggest it should be well documented in code
comments.
However I suspect the thing to do here is to convert this to take a
gfp_t argument to permit callers to use the stronger memory allocation
strategies.
> void free_iova_mem(struct iova *iova)
> {
> kmem_cache_free(iommu_iova_cache, iova);
> }
>
> void
> init_iova_domain(struct iova_domain *iovad, unsigned long granule,
> unsigned long start_pfn, unsigned long pfn_32bit)
> {
> /*
> * IOVA granularity will normally be equal to the smallest
> * supported IOMMU page size; both *must* be capable of
> * representing individual CPU pages exactly.
> */
> BUG_ON((granule > PAGE_SIZE) || !is_power_of_2(granule));
>
> spin_lock_init(&iovad->iova_rbtree_lock);
> iovad->rbroot = RB_ROOT;
> iovad->cached32_node = NULL;
> iovad->granule = granule;
> iovad->start_pfn = start_pfn;
> iovad->dma_32bit_pfn = pfn_32bit;
> }
>
> static struct rb_node *
> __get_cached_rbnode(struct iova_domain *iovad, unsigned long *limit_pfn)
> {
> if ((*limit_pfn != iovad->dma_32bit_pfn) ||
> (iovad->cached32_node == NULL))
> return rb_last(&iovad->rbroot);
> else {
> struct rb_node *prev_node = rb_prev(iovad->cached32_node);
> struct iova *curr_iova =
> container_of(iovad->cached32_node, struct iova, node);
> *limit_pfn = curr_iova->pfn_lo - 1;
> return prev_node;
> }
> }
>
> static void
> __cached_rbnode_insert_update(struct iova_domain *iovad,
> unsigned long limit_pfn, struct iova *new)
> {
> if (limit_pfn != iovad->dma_32bit_pfn)
> return;
> iovad->cached32_node = &new->node;
> }
>
> static void
> __cached_rbnode_delete_update(struct iova_domain *iovad, struct iova *free)
> {
> struct iova *cached_iova;
> struct rb_node *curr;
>
> if (!iovad->cached32_node)
> return;
> curr = iovad->cached32_node;
> cached_iova = container_of(curr, struct iova, node);
>
> if (free->pfn_lo >= cached_iova->pfn_lo) {
> struct rb_node *node = rb_next(&free->node);
> struct iova *iova = container_of(node, struct iova, node);
>
> /* only cache if it's below 32bit pfn */
> if (node && iova->pfn_lo < iovad->dma_32bit_pfn)
> iovad->cached32_node = node;
> else
> iovad->cached32_node = NULL;
> }
> }
>
> /* Computes the padding size required, to make the
> * the start address naturally aligned on its size
> */
Comment has unusual layout, doesn't use all the display and has "the
the". Like this:
/*
* Computes the padding size required, to make the start address naturally
* aligned on its size.
*/
> static int
> iova_get_pad_size(int size, unsigned int limit_pfn)
iova_<operation>. Yay.
> {
> unsigned int pad_size = 0;
> unsigned int order = ilog2(size);
>
> if (order)
> pad_size = (limit_pfn + 1) % (1 << order);
>
> return pad_size;
> }
>
> static int __alloc_and_insert_iova_range(struct iova_domain *iovad,
> unsigned long size, unsigned long limit_pfn,
> struct iova *new, bool size_aligned)
> {
> struct rb_node *prev, *curr = NULL;
> unsigned long flags;
> unsigned long saved_pfn;
> unsigned int pad_size = 0;
>
> /* Walk the tree backwards */
> spin_lock_irqsave(&iovad->iova_rbtree_lock, flags);
> saved_pfn = limit_pfn;
> curr = __get_cached_rbnode(iovad, &limit_pfn);
> prev = curr;
> while (curr) {
> struct iova *curr_iova = container_of(curr, struct iova, node);
>
> if (limit_pfn < curr_iova->pfn_lo)
> goto move_left;
> else if (limit_pfn < curr_iova->pfn_hi)
> goto adjust_limit_pfn;
> else {
> if (size_aligned)
> pad_size = iova_get_pad_size(size, limit_pfn);
> if ((curr_iova->pfn_hi + size + pad_size) <= limit_pfn)
> break; /* found a free slot */
> }
> adjust_limit_pfn:
> limit_pfn = curr_iova->pfn_lo - 1;
> move_left:
> prev = curr;
> curr = rb_prev(curr);
> }
>
> if (!curr) {
> if (size_aligned)
> pad_size = iova_get_pad_size(size, limit_pfn);
> if ((iovad->start_pfn + size + pad_size) > limit_pfn) {
> spin_unlock_irqrestore(&iovad->iova_rbtree_lock, flags);
> return -ENOMEM;
> }
> }
>
> /* pfn_lo will point to size aligned address if size_aligned is set */
> new->pfn_lo = limit_pfn - (size + pad_size) + 1;
> new->pfn_hi = new->pfn_lo + size - 1;
>
> /* Insert the new_iova into domain rbtree by holding writer lock */
> /* Add new node and rebalance tree. */
> {
> struct rb_node **entry, *parent = NULL;
>
> /* If we have 'prev', it's a valid place to start the
> insertion. Otherwise, start from the root. */
> if (prev)
> entry = &prev;
> else
> entry = &iovad->rbroot.rb_node;
>
> /* Figure out where to put new node */
> while (*entry) {
> struct iova *this = container_of(*entry,
> struct iova, node);
Could do
struct iova *this;
this = container_of(*entry, struct iova, node);
and avoid 80-col weirdnesses. This happens a lot.
> parent = *entry;
>
> if (new->pfn_lo < this->pfn_lo)
> entry = &((*entry)->rb_left);
> else if (new->pfn_lo > this->pfn_lo)
> entry = &((*entry)->rb_right);
> else
> BUG(); /* this should not happen */
> }
>
> /* Add new node and rebalance tree. */
> rb_link_node(&new->node, parent, entry);
> rb_insert_color(&new->node, &iovad->rbroot);
> }
> __cached_rbnode_insert_update(iovad, saved_pfn, new);
>
> spin_unlock_irqrestore(&iovad->iova_rbtree_lock, flags);
>
>
> return 0;
> }
>
> static void
> iova_insert_rbtree(struct rb_root *root, struct iova *iova)
> {
> struct rb_node **new = &(root->rb_node), *parent = NULL;
> /* Figure out where to put new node */
> while (*new) {
> struct iova *this = container_of(*new, struct iova, node);
>
> parent = *new;
>
> if (iova->pfn_lo < this->pfn_lo)
> new = &((*new)->rb_left);
> else if (iova->pfn_lo > this->pfn_lo)
> new = &((*new)->rb_right);
> else
> BUG(); /* this should not happen */
> }
> /* Add new node and rebalance tree. */
> rb_link_node(&iova->node, parent, new);
> rb_insert_color(&iova->node, root);
> }
>
> /**
> * alloc_iova - allocates an iova
> * @iovad: - iova domain in question
> * @size: - size of page frames to allocate
> * @limit_pfn: - max limit address
> * @size_aligned: - set if size_aligned address range is required
The "-" on the function arguments isn't correct kerneldoc, I think.
> * This function allocates an iova in the range iovad->start_pfn to limit_pfn,
> * searching top-down from limit_pfn to iovad->start_pfn. If the size_aligned
> * flag is set then the allocated address iova->pfn_lo will be naturally
> * aligned on roundup_power_of_two(size).
> */
> struct iova *
> alloc_iova(struct iova_domain *iovad, unsigned long size,
> unsigned long limit_pfn,
> bool size_aligned)
Layout of the arguments is strange and straggly.
> {
> struct iova *new_iova;
> int ret;
>
> new_iova = alloc_iova_mem();
> if (!new_iova)
> return NULL;
>
> /* If size aligned is set then round the size to
> * to next power of two.
> */
That's a pretty useless comment!
> if (size_aligned)
> size = __roundup_pow_of_two(size);
>
> ret = __alloc_and_insert_iova_range(iovad, size, limit_pfn,
> new_iova, size_aligned);
>
> if (ret) {
> free_iova_mem(new_iova);
> return NULL;
> }
>
> return new_iova;
> }
>
> /**
> * find_iova - find's an iova for a given pfn
> * @iovad: - iova domain in question.
> * @pfn: - page frame number
> * This function finds and returns an iova belonging to the
> * given doamin which matches the given pfn.
> */
> struct iova *find_iova(struct iova_domain *iovad, unsigned long pfn)
> {
> unsigned long flags;
> struct rb_node *node;
>
> /* Take the lock so that no other thread is manipulating the rbtree */
> spin_lock_irqsave(&iovad->iova_rbtree_lock, flags);
> node = iovad->rbroot.rb_node;
> while (node) {
> struct iova *iova = container_of(node, struct iova, node);
>
> /* If pfn falls within iova's range, return iova */
> if ((pfn >= iova->pfn_lo) && (pfn <= iova->pfn_hi)) {
> spin_unlock_irqrestore(&iovad->iova_rbtree_lock, flags);
> /* We are not holding the lock while this iova
> * is referenced by the caller as the same thread
> * which called this function also calls __free_iova()
> * and it is by design that only one thread can possibly
> * reference a particular iova and hence no conflict.
> */
Well that sounds fishy.
I'm generally assuming that this code isn't going to be worked on by
the lay masses - much of the design resides within a few people's heads
and that is where it will stay. Otherwise a whole lotta documetation
will need to be added. eg, what is this "design" of which you speak.
> return iova;
> }
>
> if (pfn < iova->pfn_lo)
> node = node->rb_left;
> else if (pfn > iova->pfn_lo)
> node = node->rb_right;
> }
>
> spin_unlock_irqrestore(&iovad->iova_rbtree_lock, flags);
> return NULL;
> }
>
> /**
> * __free_iova - frees the given iova
> * @iovad: iova domain in question.
> * @iova: iova in question.
> * Frees the given iova belonging to the giving domain
> */
> void
> __free_iova(struct iova_domain *iovad, struct iova *iova)
> {
> unsigned long flags;
>
> spin_lock_irqsave(&iovad->iova_rbtree_lock, flags);
> __cached_rbnode_delete_update(iovad, iova);
> rb_erase(&iova->node, &iovad->rbroot);
> spin_unlock_irqrestore(&iovad->iova_rbtree_lock, flags);
> free_iova_mem(iova);
> }
>
> /**
> * free_iova - finds and frees the iova for a given pfn
> * @iovad: - iova domain in question.
> * @pfn: - pfn that is allocated previously
> * This functions finds an iova for a given pfn and then
> * frees the iova from that domain.
> */
> void
> free_iova(struct iova_domain *iovad, unsigned long pfn)
> {
> struct iova *iova = find_iova(iovad, pfn);
>
> if (iova)
> __free_iova(iovad, iova);
>
> }
>
> /**
> * put_iova_domain - destroys the iova doamin
> * @iovad: - iova domain in question.
> * All the iova's in that domain are destroyed.
> */
> void put_iova_domain(struct iova_domain *iovad)
This is not a "put" function. In Linux, "put" means "decrement the
refcount and free if it fell to zero". A better name would be
iova_domain_free().
> {
> struct rb_node *node;
> unsigned long flags;
>
> spin_lock_irqsave(&iovad->iova_rbtree_lock, flags);
> node = rb_first(&iovad->rbroot);
> while (node) {
> struct iova *iova = container_of(node, struct iova, node);
>
> rb_erase(node, &iovad->rbroot);
> free_iova_mem(iova);
> node = rb_first(&iovad->rbroot);
> }
> spin_unlock_irqrestore(&iovad->iova_rbtree_lock, flags);
> }
>
> static int
> __is_range_overlap(struct rb_node *node,
> unsigned long pfn_lo, unsigned long pfn_hi)
> {
> struct iova *iova = container_of(node, struct iova, node);
>
> if ((pfn_lo <= iova->pfn_hi) && (pfn_hi >= iova->pfn_lo))
> return 1;
> return 0;
return (pfn_lo <= iova->pfn_hi) && (pfn_hi >= iova->pfn_lo);
> }
>
> static inline struct iova *
> alloc_and_init_iova(unsigned long pfn_lo, unsigned long pfn_hi)
> {
> struct iova *iova;
>
> iova = alloc_iova_mem();
> if (iova) {
> iova->pfn_lo = pfn_lo;
> iova->pfn_hi = pfn_hi;
> }
>
> return iova;
> }
>
> static struct iova *
> __insert_new_range(struct iova_domain *iovad,
> unsigned long pfn_lo, unsigned long pfn_hi)
> {
> struct iova *iova;
>
> iova = alloc_and_init_iova(pfn_lo, pfn_hi);
> if (iova)
> iova_insert_rbtree(&iovad->rbroot, iova);
>
> return iova;
> }
>
> static void
> __adjust_overlap_range(struct iova *iova,
> unsigned long *pfn_lo, unsigned long *pfn_hi)
> {
> if (*pfn_lo < iova->pfn_lo)
> iova->pfn_lo = *pfn_lo;
> if (*pfn_hi > iova->pfn_hi)
> *pfn_lo = iova->pfn_hi + 1;
> }
>
> /**
> * reserve_iova - reserves an iova in the given range
> * @iovad: - iova domain pointer
> * @pfn_lo: - lower page frame address
> * @pfn_hi:- higher pfn adderss
More kerneldoc funnies.
s/adderss/address/
> * This function allocates reserves the address range from pfn_lo to pfn_hi so
Sentence needs an edit
> * that this address is not dished out as part of alloc_iova.
> */
> struct iova *
> reserve_iova(struct iova_domain *iovad,
> unsigned long pfn_lo, unsigned long pfn_hi)
> {
> struct rb_node *node;
> unsigned long flags;
> struct iova *iova;
> unsigned int overlap = 0;
>
> spin_lock_irqsave(&iovad->iova_rbtree_lock, flags);
> for (node = rb_first(&iovad->rbroot); node; node = rb_next(node)) {
> if (__is_range_overlap(node, pfn_lo, pfn_hi)) {
> iova = container_of(node, struct iova, node);
> __adjust_overlap_range(iova, &pfn_lo, &pfn_hi);
> if ((pfn_lo >= iova->pfn_lo) &&
> (pfn_hi <= iova->pfn_hi))
> goto finish;
> overlap = 1;
>
> } else if (overlap)
> break;
indenting
> }
>
> /* We are here either because this is the first reserver node
> * or need to insert remaining non overlap addr range
> */
> iova = __insert_new_range(iovad, pfn_lo, pfn_hi);
> finish:
>
> spin_unlock_irqrestore(&iovad->iova_rbtree_lock, flags);
> return iova;
> }
>
> /**
> * copy_reserved_iova - copies the reserved between domains
> * @from: - source doamin from where to copy
> * @to: - destination domin where to copy
> * This function copies reserved iova's from one doamin to
> * other.
> */
> void
> copy_reserved_iova(struct iova_domain *from, struct iova_domain *to)
> {
> unsigned long flags;
> struct rb_node *node;
>
> spin_lock_irqsave(&from->iova_rbtree_lock, flags);
> for (node = rb_first(&from->rbroot); node; node = rb_next(node)) {
> struct iova *iova = container_of(node, struct iova, node);
> struct iova *new_iova;
>
> new_iova = reserve_iova(to, iova->pfn_lo, iova->pfn_hi);
> if (!new_iova)
> printk(KERN_ERR "Reserve iova range %lx@%lx failed\n",
> iova->pfn_lo, iova->pfn_lo);
> }
> spin_unlock_irqrestore(&from->iova_rbtree_lock, flags);
> }
>
> struct iova *
> split_and_remove_iova(struct iova_domain *iovad, struct iova *iova,
> unsigned long pfn_lo, unsigned long pfn_hi)
It would be nice to complete the kerneldoc, at least on the exported
interface.
> {
> unsigned long flags;
> struct iova *prev = NULL, *next = NULL;
>
> spin_lock_irqsave(&iovad->iova_rbtree_lock, flags);
> if (iova->pfn_lo < pfn_lo) {
> prev = alloc_and_init_iova(iova->pfn_lo, pfn_lo - 1);
> if (prev == NULL)
> goto error;
> }
> if (iova->pfn_hi > pfn_hi) {
> next = alloc_and_init_iova(pfn_hi + 1, iova->pfn_hi);
> if (next == NULL)
> goto error;
> }
>
> __cached_rbnode_delete_update(iovad, iova);
> rb_erase(&iova->node, &iovad->rbroot);
>
> if (prev) {
> iova_insert_rbtree(&iovad->rbroot, prev);
> iova->pfn_lo = pfn_lo;
> }
> if (next) {
> iova_insert_rbtree(&iovad->rbroot, next);
> iova->pfn_hi = pfn_hi;
> }
> spin_unlock_irqrestore(&iovad->iova_rbtree_lock, flags);
>
> return iova;
>
> error:
> spin_unlock_irqrestore(&iovad->iova_rbtree_lock, flags);
> if (prev)
> free_iova_mem(prev);
> return NULL;
> }
>
>
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
[toc] | [prev] | [next] | [standalone]
| From | Ashutosh Dixit <ashutosh.dixit@intel.com> |
|---|---|
| Date | 2015-07-28 01:40 +0200 |
| Subject | [PATCH char-misc-next 11/19] dma: Add support to program MIC x100 status descriptiors |
| Message-ID | <pR09Z-8vE-49@gated-at.bofh.it> |
| In reply to | #1193568 |
From: Siva Yerramreddy <yshivakrishna@gmail.com>
The MIC X100 DMA engine has a special status descriptor which writes
an 8 byte value to a destination location. This is used to signal
completion of all DMA descriptors prior to the status descriptor. This
patch uses the src field as a value since the DMA engine API does not
allow passing a value as an argument, if the length is 8 bytes. This
special case works since the MIC x100 DMA engine requires cache line
(64 byte) aligned src/dst/len so we can use the 8 byte length as a
special case to enable programming status descriptors without
requiring a change to the DMA engine API.
Reviewed-by: Nikhil Rao <nikhil.rao@intel.com>
Reviewed-by: Ashutosh Dixit <ashutosh.dixit@intel.com>
Reviewed-by: Sudeep Dutt <sudeep.dutt@intel.com>
Signed-off-by: Siva Yerramreddy <yshivakrishna@gmail.com>
---
drivers/dma/mic_x100_dma.c | 44 +++++++++++++++++++++++++++++++++++++++++++-
1 file changed, 43 insertions(+), 1 deletion(-)
diff --git a/drivers/dma/mic_x100_dma.c b/drivers/dma/mic_x100_dma.c
index 74d9db0..8d8acfc 100644
--- a/drivers/dma/mic_x100_dma.c
+++ b/drivers/dma/mic_x100_dma.c
@@ -193,8 +193,16 @@ static void mic_dma_prog_intr(struct mic_dma_chan *ch)
static int mic_dma_do_dma(struct mic_dma_chan *ch, int flags, dma_addr_t src,
dma_addr_t dst, size_t len)
{
- if (-ENOMEM == mic_dma_prog_memcpy_desc(ch, src, dst, len))
+ if (len && -ENOMEM == mic_dma_prog_memcpy_desc(ch, src, dst, len)) {
return -ENOMEM;
+ } else {
+ /* 3 is the maximum number of status descriptors */
+ int ret = mic_dma_avail_desc_ring_space(ch, 3);
+
+ if (ret < 0)
+ return ret;
+ }
+
/* Above mic_dma_prog_memcpy_desc() makes sure we have enough space */
if (flags & DMA_PREP_FENCE) {
mic_dma_prep_status_desc(&ch->desc_ring[ch->head], 0,
@@ -208,6 +216,21 @@ static int mic_dma_do_dma(struct mic_dma_chan *ch, int flags, dma_addr_t src,
return 0;
}
+/* Program a status descriptor with phys as address and value to be written */
+static int mic_dma_do_status_update(struct mic_dma_chan *ch, dma_addr_t phys,
+ u64 value)
+{
+ int ret = mic_dma_avail_desc_ring_space(ch, 4);
+
+ if (ret < 0)
+ return ret;
+ ret = 0;
+ mic_dma_prep_status_desc(&ch->desc_ring[ch->head],
+ value, phys, false);
+ mic_dma_hw_ring_inc_head(ch);
+ return ret;
+}
+
static inline void mic_dma_issue_pending(struct dma_chan *ch)
{
struct mic_dma_chan *mic_ch = to_mic_dma_chan(ch);
@@ -287,9 +310,28 @@ mic_dma_prep_memcpy_lock(struct dma_chan *ch, dma_addr_t dma_dest,
return NULL;
spin_lock(&mic_ch->prep_lock);
+ if (len == 8) {
+ /*
+ * This is a hack to program status descriptor since
+ * DMA engine API doesn't have support for this.
+ */
+ result = mic_dma_do_status_update(mic_ch, dma_dest, dma_src);
+ if (result)
+ goto error;
+
+ result = mic_dma_do_dma(mic_ch, DMA_PREP_FENCE, 0, 0, 0);
+ if (result < 0)
+ goto error;
+ result = mic_dma_do_dma(mic_ch, flags, 0, 0, 0);
+ if (result < 0)
+ goto error;
+ return allocate_tx(mic_ch);
+ }
+
result = mic_dma_do_dma(mic_ch, flags, dma_src, dma_dest, len);
if (result >= 0)
return allocate_tx(mic_ch);
+error:
dev_err(dev, "Error enqueueing dma, error=%d\n", result);
spin_unlock(&mic_ch->prep_lock);
return NULL;
--
2.0.0.rc3.2.g998f840
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
[toc] | [prev] | [next] | [standalone]
| From | Ashutosh Dixit <ashutosh.dixit@intel.com> |
|---|---|
| Date | 2015-07-28 01:40 +0200 |
| Subject | [PATCH char-misc-next 16/19] misc: mic: SCIF remote memory map/unmap interface |
| Message-ID | <pR0a0-8vE-69@gated-at.bofh.it> |
| In reply to | #1193568 |
From: Sudeep Dutt <sudeep.dutt@intel.com>
This patch implements the SCIF mmap/munmap interface. A similar
capability is provided to kernel clients via the
scif_get_pages()/scif_put_pages() APIs. The SCIF mmap interface
queries to check if a window is valid and then remaps the local
virtual address to the remote physical pages. These mappings are
subsequently destroyed upon receipt of the VMA close operation or
scif_get_pages(). This functionality allows SCIF users to directly
access remote memory without any driver interaction once the mappings
are created thereby providing bare-metal PCIe latency. These mappings
are zapped to avoid RMA accesses from user space, if a Coprocessor is
reset.
Reviewed-by: Ashutosh Dixit <ashutosh.dixit@intel.com>
Reviewed-by: Nikhil Rao <nikhil.rao@intel.com>
Signed-off-by: Sudeep Dutt <sudeep.dutt@intel.com>
---
drivers/misc/mic/scif/scif_mmap.c | 711 ++++++++++++++++++++++++++++++++++++++
1 file changed, 711 insertions(+)
create mode 100644 drivers/misc/mic/scif/scif_mmap.c
diff --git a/drivers/misc/mic/scif/scif_mmap.c b/drivers/misc/mic/scif/scif_mmap.c
new file mode 100644
index 0000000..c6c728e
--- /dev/null
+++ b/drivers/misc/mic/scif/scif_mmap.c
@@ -0,0 +1,711 @@
+/*
+ * Intel MIC Platform Software Stack (MPSS)
+ *
+ * Copyright(c) 2015 Intel Corporation.
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License, version 2, as
+ * published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful, but
+ * WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+ * General Public License for more details.
+ *
+ * Intel SCIF driver.
+ *
+ */
+#include "scif_main.h"
+
+/*
+ * struct scif_vma_info - Information about a remote memory mapping
+ * created via scif_mmap(..)
+ * @vma: VM area struct
+ * @list: link to list of active vmas
+ */
+struct scif_vma_info {
+ struct vm_area_struct *vma;
+ struct list_head list;
+};
+
+void scif_recv_munmap(struct scif_dev *scifdev, struct scifmsg *msg)
+{
+ struct scif_rma_req req;
+ struct scif_window *window = NULL;
+ struct scif_window *recv_window =
+ (struct scif_window *)msg->payload[0];
+ struct scif_endpt *ep;
+
+ ep = (struct scif_endpt *)recv_window->ep;
+ req.out_window = &window;
+ req.offset = recv_window->offset;
+ req.prot = recv_window->prot;
+ req.nr_bytes = recv_window->nr_pages << PAGE_SHIFT;
+ req.type = SCIF_WINDOW_FULL;
+ req.head = &ep->rma_info.reg_list;
+ msg->payload[0] = ep->remote_ep;
+
+ mutex_lock(&ep->rma_info.rma_lock);
+ /* Does a valid window exist? */
+ if (scif_query_window(&req)) {
+ dev_err(&scifdev->sdev->dev,
+ "%s %d -ENXIO\n", __func__, __LINE__);
+ msg->uop = SCIF_UNREGISTER_ACK;
+ goto error;
+ }
+
+ scif_put_window(window, window->nr_pages);
+
+ if (!window->ref_count) {
+ atomic_inc(&ep->rma_info.tw_refcount);
+ ep->rma_info.async_list_del = 1;
+ list_del_init(&window->list);
+ scif_free_window_offset(ep, window, window->offset);
+ }
+error:
+ mutex_unlock(&ep->rma_info.rma_lock);
+ if (window && !window->ref_count)
+ scif_queue_for_cleanup(window, &scif_info.rma);
+}
+
+/*
+ * Remove valid remote memory mappings created via scif_mmap(..) from the
+ * process address space since the remote node is lost
+ */
+static void __scif_zap_mmaps(struct scif_endpt *ep)
+{
+ struct list_head *item;
+ struct scif_vma_info *info;
+ struct vm_area_struct *vma;
+ unsigned long size;
+
+ spin_lock(&ep->lock);
+ list_for_each(item, &ep->rma_info.vma_list) {
+ info = list_entry(item, struct scif_vma_info, list);
+ vma = info->vma;
+ size = vma->vm_end - vma->vm_start;
+ zap_vma_ptes(vma, vma->vm_start, size);
+ dev_dbg(scif_info.mdev.this_device,
+ "%s ep %p zap vma %p size 0x%lx\n",
+ __func__, ep, info->vma, size);
+ }
+ spin_unlock(&ep->lock);
+}
+
+/*
+ * Traverse the list of endpoints for a particular remote node and
+ * zap valid remote memory mappings since the remote node is lost
+ */
+static void _scif_zap_mmaps(int node, struct list_head *head)
+{
+ struct scif_endpt *ep;
+ struct list_head *item;
+
+ mutex_lock(&scif_info.connlock);
+ list_for_each(item, head) {
+ ep = list_entry(item, struct scif_endpt, list);
+ if (ep->remote_dev->node == node)
+ __scif_zap_mmaps(ep);
+ }
+ mutex_unlock(&scif_info.connlock);
+}
+
+/*
+ * Wrapper for removing remote memory mappings for a particular node. This API
+ * is called by peer nodes as part of handling a lost node.
+ */
+void scif_zap_mmaps(int node)
+{
+ _scif_zap_mmaps(node, &scif_info.connected);
+ _scif_zap_mmaps(node, &scif_info.disconnected);
+}
+
+/*
+ * This API is only called while handling a lost node:
+ * a) Remote node is dead.
+ * b) Remote memory mappings have been zapped
+ * So we can traverse the remote_reg_list without any locks. Since
+ * the window has not yet been unregistered we can drop the ref count
+ * and queue it to the cleanup thread.
+ */
+static void __scif_cleanup_rma_for_zombies(struct scif_endpt *ep)
+{
+ struct list_head *pos, *tmp;
+ struct scif_window *window;
+
+ list_for_each_safe(pos, tmp, &ep->rma_info.remote_reg_list) {
+ window = list_entry(pos, struct scif_window, list);
+ if (window->ref_count)
+ scif_put_window(window, window->nr_pages);
+ else
+ dev_err(scif_info.mdev.this_device,
+ "%s %d unexpected\n",
+ __func__, __LINE__);
+ if (!window->ref_count) {
+ atomic_inc(&ep->rma_info.tw_refcount);
+ list_del_init(&window->list);
+ scif_queue_for_cleanup(window, &scif_info.rma);
+ }
+ }
+}
+
+/* Cleanup remote registration lists for zombie endpoints */
+void scif_cleanup_rma_for_zombies(int node)
+{
+ struct scif_endpt *ep;
+ struct list_head *item;
+
+ spin_lock(&scif_info.eplock);
+ list_for_each(item, &scif_info.zombie) {
+ ep = list_entry(item, struct scif_endpt, list);
+ if (ep->remote_dev && ep->remote_dev->node == node)
+ __scif_cleanup_rma_for_zombies(ep);
+ }
+ spin_unlock(&scif_info.eplock);
+ flush_work(&scif_info.misc_work);
+}
+
+/* Insert the VMA into the per endpoint VMA list */
+static int scif_insert_vma(struct scif_endpt *ep, struct vm_area_struct *vma)
+{
+ struct scif_vma_info *info;
+ int err = 0;
+
+ info = kzalloc(sizeof(*info), GFP_KERNEL);
+ if (!info) {
+ err = -ENOMEM;
+ goto done;
+ }
+ info->vma = vma;
+ spin_lock(&ep->lock);
+ list_add_tail(&info->list, &ep->rma_info.vma_list);
+ spin_unlock(&ep->lock);
+done:
+ return err;
+}
+
+/* Delete the VMA from the per endpoint VMA list */
+static void scif_delete_vma(struct scif_endpt *ep, struct vm_area_struct *vma)
+{
+ struct list_head *item;
+ struct scif_vma_info *info;
+
+ spin_lock(&ep->lock);
+ list_for_each(item, &ep->rma_info.vma_list) {
+ info = list_entry(item, struct scif_vma_info, list);
+ if (info->vma == vma) {
+ list_del(&info->list);
+ kfree(info);
+ break;
+ }
+ }
+ spin_unlock(&ep->lock);
+}
+
+static phys_addr_t scif_get_phys(phys_addr_t phys, struct scif_endpt *ep)
+{
+ struct scif_dev *scifdev = (struct scif_dev *)ep->remote_dev;
+ struct scif_hw_dev *sdev = scifdev->sdev;
+ phys_addr_t out_phys, apt_base = 0;
+
+ /*
+ * If the DMA address is card relative then we need to add the
+ * aperture base for mmap to work correctly
+ */
+ if (!scifdev_self(scifdev) && sdev->aper && sdev->card_rel_da)
+ apt_base = sdev->aper->pa;
+ out_phys = apt_base + phys;
+ return out_phys;
+}
+
+struct _scif_range {
+ void *cookie;
+ struct scif_range *sr;
+};
+
+int scif_get_pages(scif_epd_t epd, off_t offset, size_t len,
+ struct scif_range **pages)
+{
+ struct scif_endpt *ep = (struct scif_endpt *)epd;
+ struct scif_rma_req req;
+ struct scif_window *window = NULL;
+ int nr_pages, err, i;
+ struct scif_range *sr = NULL;
+ struct _scif_range *_sr = NULL;
+
+ dev_dbg(scif_info.mdev.this_device,
+ "SCIFAPI get_pinned_pages: ep %p offset 0x%lx len 0x%lx\n",
+ ep, offset, len);
+ err = scif_verify_epd(ep);
+ if (err)
+ return err;
+
+ if (!len || (offset < 0) ||
+ (offset + len < offset) ||
+ (ALIGN(offset, PAGE_SIZE) != offset) ||
+ (ALIGN(len, PAGE_SIZE) != len))
+ return -EINVAL;
+
+ nr_pages = len >> PAGE_SHIFT;
+
+ req.out_window = &window;
+ req.offset = offset;
+ req.prot = 0;
+ req.nr_bytes = len;
+ req.type = SCIF_WINDOW_SINGLE;
+ req.head = &ep->rma_info.remote_reg_list;
+
+ mutex_lock(&ep->rma_info.rma_lock);
+ /* Does a valid window exist? */
+ err = scif_query_window(&req);
+ if (err) {
+ dev_err(&ep->remote_dev->sdev->dev,
+ "%s %d err %d\n", __func__, __LINE__, err);
+ goto error;
+ }
+
+ /* Allocate scif_range */
+ sr = kzalloc(sizeof(*sr), GFP_KERNEL);
+ if (!sr) {
+ err = -ENOMEM;
+ goto error;
+ }
+
+ /* Allocate _scif_range */
+ _sr = kzalloc(sizeof(*_sr), GFP_KERNEL);
+ if (!_sr) {
+ err = -ENOMEM;
+ goto error;
+ }
+
+ /* Allocate phys addr array */
+ sr->phys_addr = scif_zalloc(nr_pages * sizeof(dma_addr_t));
+ if (!sr->phys_addr) {
+ err = -ENOMEM;
+ goto error;
+ }
+
+ if (scif_is_mgmt_node() && !scifdev_self(ep->remote_dev)) {
+ /* Allocate virtual address array */
+ sr->va = scif_zalloc(nr_pages * sizeof(void *));
+ if (!sr->va) {
+ err = -ENOMEM;
+ goto error;
+ }
+ }
+ /* Populate the values */
+ _sr->cookie = window;
+ _sr->sr = sr;
+ sr->nr_pages = nr_pages;
+ sr->prot_flags = window->prot;
+
+ for (i = 0; i < nr_pages; i++) {
+ sr->phys_addr[i] = __scif_off_to_dma_addr(window, offset +
+ (i * PAGE_SIZE));
+ sr->phys_addr[i] = scif_get_phys(sr->phys_addr[i], ep);
+ if (scif_is_mgmt_node() && !scifdev_self(ep->remote_dev))
+ sr->va[i] = ep->remote_dev->sdev->aper->va +
+ sr->phys_addr[i] -
+ ep->remote_dev->sdev->aper->pa;
+ }
+
+ scif_get_window(window, nr_pages);
+ *pages = sr;
+error:
+ mutex_unlock(&ep->rma_info.rma_lock);
+ if (err) {
+ if (sr) {
+ scif_free(sr->phys_addr, nr_pages * sizeof(dma_addr_t));
+ scif_free(sr->va, nr_pages * sizeof(void *));
+ kfree(sr);
+ }
+ kfree(_sr);
+ dev_err(&ep->remote_dev->sdev->dev,
+ "%s %d err %d\n", __func__, __LINE__, err);
+ }
+ return err;
+}
+EXPORT_SYMBOL_GPL(scif_get_pages);
+
+int scif_put_pages(struct scif_range *pages)
+{
+ struct scif_endpt *ep;
+ struct scif_window *window;
+ struct scifmsg msg;
+ struct _scif_range *_sr = container_of(&pages, struct _scif_range, sr);
+
+ if (!pages)
+ return -EINVAL;
+
+ window = _sr->cookie;
+
+ if (!window || window->magic != SCIFEP_MAGIC)
+ return -EINVAL;
+
+ ep = (struct scif_endpt *)window->ep;
+ /*
+ * If the state is SCIFEP_CONNECTED or SCIFEP_DISCONNECTED then the
+ * callee should be allowed to release references to the pages,
+ * else the endpoint was not connected in the first place,
+ * hence the ENOTCONN.
+ */
+ if (ep->state != SCIFEP_CONNECTED && ep->state != SCIFEP_DISCONNECTED)
+ return -ENOTCONN;
+
+ mutex_lock(&ep->rma_info.rma_lock);
+
+ scif_put_window(window, pages->nr_pages);
+
+ /* Initiate window destruction if ref count is zero */
+ if (!window->ref_count) {
+ list_del(&window->list);
+ mutex_unlock(&ep->rma_info.rma_lock);
+ scif_drain_dma_intr(ep->remote_dev->sdev,
+ ep->rma_info.dma_chan);
+ /* Inform the peer about this window being destroyed. */
+ msg.uop = SCIF_MUNMAP;
+ msg.src = ep->port;
+ msg.payload[0] = window->peer_window;
+ /* No error handling for notification messages */
+ scif_nodeqp_send(ep->remote_dev, &msg);
+ /* Destroy this window from the peer's registered AS */
+ scif_destroy_remote_window(window);
+ } else {
+ mutex_unlock(&ep->rma_info.rma_lock);
+ }
+
+ scif_free(pages->phys_addr, pages->nr_pages * sizeof(dma_addr_t));
+ scif_free(pages->va, pages->nr_pages * sizeof(void *));
+ kfree(pages);
+ return 0;
+}
+EXPORT_SYMBOL_GPL(scif_put_pages);
+
+/*
+ * scif_rma_list_mmap:
+ *
+ * Traverse the remote registration list starting from start_window:
+ * 1) Create VtoP mappings via remap_pfn_range(..)
+ * 2) Once step 1) and 2) complete successfully then traverse the range of
+ * windows again and bump the reference count.
+ * RMA lock must be held.
+ */
+static int scif_rma_list_mmap(struct scif_window *start_window, s64 offset,
+ int nr_pages, struct vm_area_struct *vma)
+{
+ s64 end_offset, loop_offset = offset;
+ struct scif_window *window = start_window;
+ int loop_nr_pages, nr_pages_left = nr_pages;
+ struct scif_endpt *ep = (struct scif_endpt *)start_window->ep;
+ struct list_head *head = &ep->rma_info.remote_reg_list;
+ int i, err = 0;
+ dma_addr_t phys_addr;
+ struct scif_window_iter src_win_iter;
+ size_t contig_bytes = 0;
+
+ might_sleep();
+ list_for_each_entry_from(window, head, list) {
+ end_offset = window->offset +
+ (window->nr_pages << PAGE_SHIFT);
+ loop_nr_pages = min_t(int,
+ (end_offset - loop_offset) >> PAGE_SHIFT,
+ nr_pages_left);
+ scif_init_window_iter(window, &src_win_iter);
+ for (i = 0; i < loop_nr_pages; i++) {
+ phys_addr = scif_off_to_dma_addr(window, loop_offset,
+ &contig_bytes,
+ &src_win_iter);
+ phys_addr = scif_get_phys(phys_addr, ep);
+ err = remap_pfn_range(vma,
+ vma->vm_start +
+ loop_offset - offset,
+ phys_addr >> PAGE_SHIFT,
+ PAGE_SIZE,
+ vma->vm_page_prot);
+ if (err)
+ goto error;
+ loop_offset += PAGE_SIZE;
+ }
+ nr_pages_left -= loop_nr_pages;
+ if (!nr_pages_left)
+ break;
+ }
+ /*
+ * No more failures expected. Bump up the ref count for all
+ * the windows. Another traversal from start_window required
+ * for handling errors encountered across windows during
+ * remap_pfn_range(..).
+ */
+ loop_offset = offset;
+ nr_pages_left = nr_pages;
+ window = start_window;
+ head = &ep->rma_info.remote_reg_list;
+ list_for_each_entry_from(window, head, list) {
+ end_offset = window->offset +
+ (window->nr_pages << PAGE_SHIFT);
+ loop_nr_pages = min_t(int,
+ (end_offset - loop_offset) >> PAGE_SHIFT,
+ nr_pages_left);
+ scif_get_window(window, loop_nr_pages);
+ nr_pages_left -= loop_nr_pages;
+ loop_offset += (loop_nr_pages << PAGE_SHIFT);
+ if (!nr_pages_left)
+ break;
+ }
+error:
+ if (err)
+ dev_err(scif_info.mdev.this_device,
+ "%s %d err %d\n", __func__, __LINE__, err);
+ return err;
+}
+
+/*
+ * scif_rma_list_munmap:
+ *
+ * Traverse the remote registration list starting from window:
+ * 1) Decrement ref count.
+ * 2) If the ref count drops to zero then send a SCIF_MUNMAP message to peer.
+ * RMA lock must be held.
+ */
+static void scif_rma_list_munmap(struct scif_window *start_window,
+ s64 offset, int nr_pages)
+{
+ struct scifmsg msg;
+ s64 loop_offset = offset, end_offset;
+ int loop_nr_pages, nr_pages_left = nr_pages;
+ struct scif_endpt *ep = (struct scif_endpt *)start_window->ep;
+ struct list_head *head = &ep->rma_info.remote_reg_list;
+ struct scif_window *window = start_window, *_window;
+
+ msg.uop = SCIF_MUNMAP;
+ msg.src = ep->port;
+ loop_offset = offset;
+ nr_pages_left = nr_pages;
+ list_for_each_entry_safe_from(window, _window, head, list) {
+ end_offset = window->offset +
+ (window->nr_pages << PAGE_SHIFT);
+ loop_nr_pages = min_t(int,
+ (end_offset - loop_offset) >> PAGE_SHIFT,
+ nr_pages_left);
+ scif_put_window(window, loop_nr_pages);
+ if (!window->ref_count) {
+ struct scif_dev *rdev = ep->remote_dev;
+
+ scif_drain_dma_intr(rdev->sdev,
+ ep->rma_info.dma_chan);
+ /* Inform the peer about this munmap */
+ msg.payload[0] = window->peer_window;
+ /* No error handling for Notification messages. */
+ scif_nodeqp_send(ep->remote_dev, &msg);
+ list_del(&window->list);
+ /* Destroy this window from the peer's registered AS */
+ scif_destroy_remote_window(window);
+ }
+ nr_pages_left -= loop_nr_pages;
+ loop_offset += (loop_nr_pages << PAGE_SHIFT);
+ if (!nr_pages_left)
+ break;
+ }
+}
+
+/*
+ * The private data field of each VMA used to mmap a remote window
+ * points to an instance of struct vma_pvt
+ */
+struct vma_pvt {
+ struct scif_endpt *ep; /* End point for remote window */
+ s64 offset; /* offset within remote window */
+ bool valid_offset; /* offset is valid only if the original
+ * mmap request was for a single page
+ * else the offset within the vma is
+ * the correct offset
+ */
+ struct kref ref;
+};
+
+static void vma_pvt_release(struct kref *ref)
+{
+ struct vma_pvt *vmapvt = container_of(ref, struct vma_pvt, ref);
+
+ kfree(vmapvt);
+}
+
+/**
+ * scif_vma_open - VMA open driver callback
+ * @vma: VMM memory area.
+ * The open method is called by the kernel to allow the subsystem implementing
+ * the VMA to initialize the area. This method is invoked any time a new
+ * reference to the VMA is made (when a process forks, for example).
+ * The one exception happens when the VMA is first created by mmap;
+ * in this case, the driver's mmap method is called instead.
+ * This function is also invoked when an existing VMA is split by the kernel
+ * due to a call to munmap on a subset of the VMA resulting in two VMAs.
+ * The kernel invokes this function only on one of the two VMAs.
+ */
+static void scif_vma_open(struct vm_area_struct *vma)
+{
+ struct vma_pvt *vmapvt = vma->vm_private_data;
+
+ dev_dbg(scif_info.mdev.this_device,
+ "SCIFAPI vma open: vma_start 0x%lx vma_end 0x%lx\n",
+ vma->vm_start, vma->vm_end);
+ scif_insert_vma(vmapvt->ep, vma);
+ kref_get(&vmapvt->ref);
+}
+
+/**
+ * scif_munmap - VMA close driver callback.
+ * @vma: VMM memory area.
+ * When an area is destroyed, the kernel calls its close operation.
+ * Note that there's no usage count associated with VMA's; the area
+ * is opened and closed exactly once by each process that uses it.
+ */
+static void scif_munmap(struct vm_area_struct *vma)
+{
+ struct scif_endpt *ep;
+ struct vma_pvt *vmapvt = vma->vm_private_data;
+ int nr_pages = (vma->vm_end - vma->vm_start) >> PAGE_SHIFT;
+ s64 offset;
+ struct scif_rma_req req;
+ struct scif_window *window = NULL;
+ int err;
+
+ might_sleep();
+ dev_dbg(scif_info.mdev.this_device,
+ "SCIFAPI munmap: vma_start 0x%lx vma_end 0x%lx\n",
+ vma->vm_start, vma->vm_end);
+ ep = vmapvt->ep;
+ offset = vmapvt->valid_offset ? vmapvt->offset :
+ (vma->vm_pgoff) << PAGE_SHIFT;
+ dev_dbg(scif_info.mdev.this_device,
+ "SCIFAPI munmap: ep %p nr_pages 0x%x offset 0x%llx\n",
+ ep, nr_pages, offset);
+ req.out_window = &window;
+ req.offset = offset;
+ req.nr_bytes = vma->vm_end - vma->vm_start;
+ req.prot = vma->vm_flags & (VM_READ | VM_WRITE);
+ req.type = SCIF_WINDOW_PARTIAL;
+ req.head = &ep->rma_info.remote_reg_list;
+
+ mutex_lock(&ep->rma_info.rma_lock);
+
+ err = scif_query_window(&req);
+ if (err)
+ dev_err(scif_info.mdev.this_device,
+ "%s %d err %d\n", __func__, __LINE__, err);
+ else
+ scif_rma_list_munmap(window, offset, nr_pages);
+
+ mutex_unlock(&ep->rma_info.rma_lock);
+ /*
+ * The kernel probably zeroes these out but we still want
+ * to clean up our own mess just in case.
+ */
+ vma->vm_ops = NULL;
+ vma->vm_private_data = NULL;
+ kref_put(&vmapvt->ref, vma_pvt_release);
+ scif_delete_vma(ep, vma);
+}
+
+static const struct vm_operations_struct scif_vm_ops = {
+ .open = scif_vma_open,
+ .close = scif_munmap,
+};
+
+/**
+ * scif_mmap - Map pages in virtual address space to a remote window.
+ * @vma: VMM memory area.
+ * @epd: endpoint descriptor
+ *
+ * Return: Upon successful completion, scif_mmap() returns zero
+ * else an apt error is returned as documented in scif.h
+ */
+int scif_mmap(struct vm_area_struct *vma, scif_epd_t epd)
+{
+ struct scif_rma_req req;
+ struct scif_window *window = NULL;
+ struct scif_endpt *ep = (struct scif_endpt *)epd;
+ s64 start_offset = vma->vm_pgoff << PAGE_SHIFT;
+ int nr_pages = (vma->vm_end - vma->vm_start) >> PAGE_SHIFT;
+ int err;
+ struct vma_pvt *vmapvt;
+
+ dev_dbg(scif_info.mdev.this_device,
+ "SCIFAPI mmap: ep %p start_offset 0x%llx nr_pages 0x%x\n",
+ ep, start_offset, nr_pages);
+ err = scif_verify_epd(ep);
+ if (err)
+ return err;
+
+ might_sleep();
+
+ err = scif_insert_vma(ep, vma);
+ if (err)
+ return err;
+
+ vmapvt = kzalloc(sizeof(*vmapvt), GFP_KERNEL);
+ if (!vmapvt) {
+ scif_delete_vma(ep, vma);
+ return -ENOMEM;
+ }
+
+ vmapvt->ep = ep;
+ kref_init(&vmapvt->ref);
+
+ req.out_window = &window;
+ req.offset = start_offset;
+ req.nr_bytes = vma->vm_end - vma->vm_start;
+ req.prot = vma->vm_flags & (VM_READ | VM_WRITE);
+ req.type = SCIF_WINDOW_PARTIAL;
+ req.head = &ep->rma_info.remote_reg_list;
+
+ mutex_lock(&ep->rma_info.rma_lock);
+ /* Does a valid window exist? */
+ err = scif_query_window(&req);
+ if (err) {
+ dev_err(&ep->remote_dev->sdev->dev,
+ "%s %d err %d\n", __func__, __LINE__, err);
+ goto error_unlock;
+ }
+
+ /* Default prot for loopback */
+ if (!scifdev_self(ep->remote_dev))
+ vma->vm_page_prot = pgprot_writecombine(vma->vm_page_prot);
+
+ /*
+ * VM_DONTCOPY - Do not copy this vma on fork
+ * VM_DONTEXPAND - Cannot expand with mremap()
+ * VM_RESERVED - Count as reserved_vm like IO
+ * VM_PFNMAP - Page-ranges managed without "struct page"
+ * VM_IO - Memory mapped I/O or similar
+ *
+ * We do not want to copy this VMA automatically on a fork(),
+ * expand this VMA due to mremap() or swap out these pages since
+ * the VMA is actually backed by physical pages in the remote
+ * node's physical memory and not via a struct page.
+ */
+ vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND | VM_DONTDUMP;
+
+ if (!scifdev_self(ep->remote_dev))
+ vma->vm_flags |= VM_IO | VM_PFNMAP;
+
+ /* Map this range of windows */
+ err = scif_rma_list_mmap(window, start_offset, nr_pages, vma);
+ if (err) {
+ dev_err(&ep->remote_dev->sdev->dev,
+ "%s %d err %d\n", __func__, __LINE__, err);
+ goto error_unlock;
+ }
+ /* Set up the driver call back */
+ vma->vm_ops = &scif_vm_ops;
+ vma->vm_private_data = vmapvt;
+error_unlock:
+ mutex_unlock(&ep->rma_info.rma_lock);
+ if (err) {
+ kfree(vmapvt);
+ dev_err(&ep->remote_dev->sdev->dev,
+ "%s %d err %d\n", __func__, __LINE__, err);
+ scif_delete_vma(ep, vma);
+ }
+ return err;
+}
--
2.0.0.rc3.2.g998f840
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
[toc] | [prev] | [next] | [standalone]
| From | Ashutosh Dixit <ashutosh.dixit@intel.com> |
|---|---|
| Date | 2015-07-28 01:40 +0200 |
| Subject | [PATCH char-misc-next 15/19] misc: mic: SCIF RMA list operations |
| Message-ID | <pR0a0-8vE-71@gated-at.bofh.it> |
| In reply to | #1193568 |
From: Sudeep Dutt <sudeep.dutt@intel.com>
This patch adds the implementation for operations performed on the
list of SCIF windows. Examples of such operations includes adding the
windows to the list of registered (or cached) windows, querying the
list of self or remote windows and unregistering windows. The query
operation is used by SCIF APIs which initiate DMAs, CPU copies or
fences to ensure that a window remains valid during a transfer.
Reviewed-by: Ashutosh Dixit <ashutosh.dixit@intel.com>
Signed-off-by: Nikhil Rao <nikhil.rao@intel.com>
Signed-off-by: Sudeep Dutt <sudeep.dutt@intel.com>
---
drivers/misc/mic/scif/scif_rma_list.c | 291 ++++++++++++++++++++++++++++++++++
drivers/misc/mic/scif/scif_rma_list.h | 57 +++++++
2 files changed, 348 insertions(+)
create mode 100644 drivers/misc/mic/scif/scif_rma_list.c
create mode 100644 drivers/misc/mic/scif/scif_rma_list.h
diff --git a/drivers/misc/mic/scif/scif_rma_list.c b/drivers/misc/mic/scif/scif_rma_list.c
new file mode 100644
index 0000000..e1ef8da
--- /dev/null
+++ b/drivers/misc/mic/scif/scif_rma_list.c
@@ -0,0 +1,291 @@
+/*
+ * Intel MIC Platform Software Stack (MPSS)
+ *
+ * Copyright(c) 2015 Intel Corporation.
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License, version 2, as
+ * published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful, but
+ * WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+ * General Public License for more details.
+ *
+ * Intel SCIF driver.
+ *
+ */
+#include "scif_main.h"
+#include <linux/mmu_notifier.h>
+#include <linux/highmem.h>
+
+/*
+ * scif_insert_tcw:
+ *
+ * Insert a temp window to the temp registration list sorted by va_for_temp.
+ * RMA lock must be held.
+ */
+void scif_insert_tcw(struct scif_window *window, struct list_head *head)
+{
+ struct scif_window *curr = NULL;
+ struct scif_window *prev = list_entry(head, struct scif_window, list);
+ struct list_head *item;
+
+ INIT_LIST_HEAD(&window->list);
+ /* Compare with tail and if the entry is new tail add it to the end */
+ if (!list_empty(head)) {
+ curr = list_entry(head->prev, struct scif_window, list);
+ if (curr->va_for_temp < window->va_for_temp) {
+ list_add_tail(&window->list, head);
+ return;
+ }
+ }
+ list_for_each(item, head) {
+ curr = list_entry(item, struct scif_window, list);
+ if (curr->va_for_temp > window->va_for_temp)
+ break;
+ prev = curr;
+ }
+ list_add(&window->list, &prev->list);
+}
+
+/*
+ * scif_insert_window:
+ *
+ * Insert a window to the self registration list sorted by offset.
+ * RMA lock must be held.
+ */
+void scif_insert_window(struct scif_window *window, struct list_head *head)
+{
+ struct scif_window *curr = NULL, *prev = NULL;
+ struct list_head *item;
+
+ INIT_LIST_HEAD(&window->list);
+ list_for_each(item, head) {
+ curr = list_entry(item, struct scif_window, list);
+ if (curr->offset > window->offset)
+ break;
+ prev = curr;
+ }
+ if (!prev)
+ list_add(&window->list, head);
+ else
+ list_add(&window->list, &prev->list);
+ scif_set_window_ref(window, window->nr_pages);
+}
+
+/*
+ * scif_query_tcw:
+ *
+ * Query the temp cached registration list of ep for an overlapping window
+ * in case of permission mismatch, destroy the previous window. if permissions
+ * match and overlap is partial, destroy the window but return the new range
+ * RMA lock must be held.
+ */
+int scif_query_tcw(struct scif_endpt *ep, struct scif_rma_req *req)
+{
+ struct list_head *item, *temp, *head = req->head;
+ struct scif_window *window;
+ u64 start_va_window, start_va_req = req->va_for_temp;
+ u64 end_va_window, end_va_req = start_va_req + req->nr_bytes;
+
+ if (!req->nr_bytes)
+ return -EINVAL;
+ /*
+ * Avoid traversing the entire list to find out that there
+ * is no entry that matches
+ */
+ if (!list_empty(head)) {
+ window = list_last_entry(head, struct scif_window, list);
+ end_va_window = window->va_for_temp +
+ (window->nr_pages << PAGE_SHIFT);
+ if (start_va_req > end_va_window)
+ return -ENXIO;
+ }
+ list_for_each_safe(item, temp, head) {
+ window = list_entry(item, struct scif_window, list);
+ start_va_window = window->va_for_temp;
+ end_va_window = window->va_for_temp +
+ (window->nr_pages << PAGE_SHIFT);
+ if (start_va_req < start_va_window &&
+ end_va_req < start_va_window)
+ break;
+ if (start_va_req >= end_va_window)
+ continue;
+ if ((window->prot & req->prot) == req->prot) {
+ if (start_va_req >= start_va_window &&
+ end_va_req <= end_va_window) {
+ *req->out_window = window;
+ return 0;
+ }
+ /* expand window */
+ if (start_va_req < start_va_window) {
+ req->nr_bytes +=
+ start_va_window - start_va_req;
+ req->va_for_temp = start_va_window;
+ }
+ if (end_va_req >= end_va_window)
+ req->nr_bytes += end_va_window - end_va_req;
+ }
+ /* Destroy the old window to create a new one */
+ __scif_rma_destroy_tcw_helper(window);
+ break;
+ }
+ return -ENXIO;
+}
+
+/*
+ * scif_query_window:
+ *
+ * Query the registration list and check if a valid contiguous
+ * range of windows exist.
+ * RMA lock must be held.
+ */
+int scif_query_window(struct scif_rma_req *req)
+{
+ struct list_head *item;
+ struct scif_window *window;
+ s64 end_offset, offset = req->offset;
+ u64 tmp_min, nr_bytes_left = req->nr_bytes;
+
+ if (!req->nr_bytes)
+ return -EINVAL;
+
+ list_for_each(item, req->head) {
+ window = list_entry(item, struct scif_window, list);
+ end_offset = window->offset +
+ (window->nr_pages << PAGE_SHIFT);
+ if (offset < window->offset)
+ /* Offset not found! */
+ return -ENXIO;
+ if (offset >= end_offset)
+ continue;
+ /* Check read/write protections. */
+ if ((window->prot & req->prot) != req->prot)
+ return -EPERM;
+ if (nr_bytes_left == req->nr_bytes)
+ /* Store the first window */
+ *req->out_window = window;
+ tmp_min = min((u64)end_offset - offset, nr_bytes_left);
+ nr_bytes_left -= tmp_min;
+ offset += tmp_min;
+ /*
+ * Range requested encompasses
+ * multiple windows contiguously.
+ */
+ if (!nr_bytes_left) {
+ /* Done for partial window */
+ if (req->type == SCIF_WINDOW_PARTIAL ||
+ req->type == SCIF_WINDOW_SINGLE)
+ return 0;
+ /* Extra logic for full windows */
+ if (offset == end_offset)
+ /* Spanning multiple whole windows */
+ return 0;
+ /* Not spanning multiple whole windows */
+ return -ENXIO;
+ }
+ if (req->type == SCIF_WINDOW_SINGLE)
+ break;
+ }
+ dev_err(scif_info.mdev.this_device,
+ "%s %d ENXIO\n", __func__, __LINE__);
+ return -ENXIO;
+}
+
+/*
+ * scif_rma_list_unregister:
+ *
+ * Traverse the self registration list starting from window:
+ * 1) Call scif_unregister_window(..)
+ * RMA lock must be held.
+ */
+int scif_rma_list_unregister(struct scif_window *window,
+ s64 offset, int nr_pages)
+{
+ struct scif_endpt *ep = (struct scif_endpt *)window->ep;
+ struct list_head *head = &ep->rma_info.reg_list;
+ s64 end_offset;
+ int err = 0;
+ int loop_nr_pages;
+ struct scif_window *_window;
+
+ list_for_each_entry_safe_from(window, _window, head, list) {
+ end_offset = window->offset + (window->nr_pages << PAGE_SHIFT);
+ loop_nr_pages = min((int)((end_offset - offset) >> PAGE_SHIFT),
+ nr_pages);
+ err = scif_unregister_window(window);
+ if (err)
+ return err;
+ nr_pages -= loop_nr_pages;
+ offset += (loop_nr_pages << PAGE_SHIFT);
+ if (!nr_pages)
+ break;
+ }
+ return 0;
+}
+
+/*
+ * scif_unmap_all_window:
+ *
+ * Traverse all the windows in the self registration list and:
+ * 1) Delete any DMA mappings created
+ */
+void scif_unmap_all_windows(scif_epd_t epd)
+{
+ struct list_head *item, *tmp;
+ struct scif_window *window;
+ struct scif_endpt *ep = (struct scif_endpt *)epd;
+ struct list_head *head = &ep->rma_info.reg_list;
+
+ mutex_lock(&ep->rma_info.rma_lock);
+ list_for_each_safe(item, tmp, head) {
+ window = list_entry(item, struct scif_window, list);
+ scif_unmap_window(ep->remote_dev, window);
+ }
+ mutex_unlock(&ep->rma_info.rma_lock);
+}
+
+/*
+ * scif_unregister_all_window:
+ *
+ * Traverse all the windows in the self registration list and:
+ * 1) Call scif_unregister_window(..)
+ * RMA lock must be held.
+ */
+int scif_unregister_all_windows(scif_epd_t epd)
+{
+ struct list_head *item, *tmp;
+ struct scif_window *window;
+ struct scif_endpt *ep = (struct scif_endpt *)epd;
+ struct list_head *head = &ep->rma_info.reg_list;
+ int err = 0;
+
+ mutex_lock(&ep->rma_info.rma_lock);
+retry:
+ item = NULL;
+ tmp = NULL;
+ list_for_each_safe(item, tmp, head) {
+ window = list_entry(item, struct scif_window, list);
+ ep->rma_info.async_list_del = 0;
+ err = scif_unregister_window(window);
+ if (err)
+ dev_err(scif_info.mdev.this_device,
+ "%s %d err %d\n",
+ __func__, __LINE__, err);
+ /*
+ * Need to restart list traversal if there has been
+ * an asynchronous list entry deletion.
+ */
+ if (ACCESS_ONCE(ep->rma_info.async_list_del))
+ goto retry;
+ }
+ mutex_unlock(&ep->rma_info.rma_lock);
+ if (!list_empty(&ep->rma_info.mmn_list)) {
+ spin_lock(&scif_info.rmalock);
+ list_add_tail(&ep->mmu_list, &scif_info.mmu_notif_cleanup);
+ spin_unlock(&scif_info.rmalock);
+ schedule_work(&scif_info.mmu_notif_work);
+ }
+ return err;
+}
diff --git a/drivers/misc/mic/scif/scif_rma_list.h b/drivers/misc/mic/scif/scif_rma_list.h
new file mode 100644
index 0000000..7d58d1d
--- /dev/null
+++ b/drivers/misc/mic/scif/scif_rma_list.h
@@ -0,0 +1,57 @@
+/*
+ * Intel MIC Platform Software Stack (MPSS)
+ *
+ * Copyright(c) 2015 Intel Corporation.
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License, version 2, as
+ * published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful, but
+ * WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+ * General Public License for more details.
+ *
+ * Intel SCIF driver.
+ *
+ */
+#ifndef SCIF_RMA_LIST_H
+#define SCIF_RMA_LIST_H
+
+/*
+ * struct scif_rma_req - Self Registration list RMA Request query
+ *
+ * @out_window - Returns the window if found
+ * @offset: Starting offset
+ * @nr_bytes: number of bytes
+ * @prot: protection requested i.e. read or write or both
+ * @type: Specify single, partial or multiple windows
+ * @head: Head of list on which to search
+ * @va_for_temp: VA for searching temporary cached windows
+ */
+struct scif_rma_req {
+ struct scif_window **out_window;
+ union {
+ s64 offset;
+ unsigned long va_for_temp;
+ };
+ size_t nr_bytes;
+ int prot;
+ enum scif_window_type type;
+ struct list_head *head;
+};
+
+/* Insert */
+void scif_insert_window(struct scif_window *window, struct list_head *head);
+void scif_insert_tcw(struct scif_window *window,
+ struct list_head *head);
+/* Query */
+int scif_query_window(struct scif_rma_req *request);
+int scif_query_tcw(struct scif_endpt *ep, struct scif_rma_req *request);
+/* Called from close to unregister all self windows */
+int scif_unregister_all_windows(scif_epd_t epd);
+void scif_unmap_all_windows(scif_epd_t epd);
+/* Traverse list and unregister */
+int scif_rma_list_unregister(struct scif_window *window, s64 offset,
+ int nr_pages);
+#endif /* SCIF_RMA_LIST_H */
--
2.0.0.rc3.2.g998f840
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
[toc] | [prev] | [next] | [standalone]
| From | Ashutosh Dixit <ashutosh.dixit@intel.com> |
|---|---|
| Date | 2015-07-28 01:40 +0200 |
| Subject | [PATCH char-misc-next 04/19] misc: mic: Coprocessor State Management (COSM) driver |
| Message-ID | <pR0a0-8vE-85@gated-at.bofh.it> |
| In reply to | #1193568 |
The COSM driver allows boot, shutdown and reset of Intel MIC devices
via sysfs. This functionality was previously present in the Intel MIC
host driver but has now been taken out into a separate driver so that
it can be shared between multiple generations of Intel MIC products.
The sysfs kernel ABI used by the COSM driver is the same as that
defined originally for the MIC host driver in
Documentation/ABI/testing/sysfs-class-mic.txt.
The COSM driver also contains support for dumping the MIC card log_buf
and doing a "force reset" for the card via debugfs. The OSPM support
present in the MIC host driver has now largely been moved to user
space and only a small required OSPM functionality is now present in
the driver.
Reviewed-by: Nikhil Rao <nikhil.rao@intel.com>
Reviewed-by: Sudeep Dutt <sudeep.dutt@intel.com>
Signed-off-by: Dasaratharaman Chandramouli <dasaratharaman.chandramouli@intel.com>
Signed-off-by: Ashutosh Dixit <ashutosh.dixit@intel.com>
---
drivers/misc/mic/cosm/Makefile | 10 +
drivers/misc/mic/cosm/cosm_debugfs.c | 155 ++++++++++++
drivers/misc/mic/cosm/cosm_main.c | 388 +++++++++++++++++++++++++++++
drivers/misc/mic/cosm/cosm_main.h | 70 ++++++
drivers/misc/mic/cosm/cosm_sysfs.c | 461 +++++++++++++++++++++++++++++++++++
5 files changed, 1084 insertions(+)
create mode 100644 drivers/misc/mic/cosm/Makefile
create mode 100644 drivers/misc/mic/cosm/cosm_debugfs.c
create mode 100644 drivers/misc/mic/cosm/cosm_main.c
create mode 100644 drivers/misc/mic/cosm/cosm_main.h
create mode 100644 drivers/misc/mic/cosm/cosm_sysfs.c
diff --git a/drivers/misc/mic/cosm/Makefile b/drivers/misc/mic/cosm/Makefile
new file mode 100644
index 0000000..b85d4d4
--- /dev/null
+++ b/drivers/misc/mic/cosm/Makefile
@@ -0,0 +1,10 @@
+#
+# Makefile - Intel MIC Coprocessor State Management (COSM) Driver
+# Copyright(c) 2015, Intel Corporation.
+#
+obj-$(CONFIG_MIC_COSM) += mic_cosm.o
+
+mic_cosm-objs := cosm_main.o
+mic_cosm-objs += cosm_debugfs.o
+mic_cosm-objs += cosm_sysfs.o
+mic_cosm-objs += cosm_scif_server.o
diff --git a/drivers/misc/mic/cosm/cosm_debugfs.c b/drivers/misc/mic/cosm/cosm_debugfs.c
new file mode 100644
index 0000000..30d953d
--- /dev/null
+++ b/drivers/misc/mic/cosm/cosm_debugfs.c
@@ -0,0 +1,155 @@
+/*
+ * Intel MIC Platform Software Stack (MPSS)
+ *
+ * Copyright(c) 2015 Intel Corporation.
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License, version 2, as
+ * published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful, but
+ * WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+ * General Public License for more details.
+ *
+ * The full GNU General Public License is included in this distribution in
+ * the file called "COPYING".
+ *
+ * Intel MIC Coprocessor State Management (COSM) Driver
+ *
+ */
+
+#include <linux/debugfs.h>
+#include <linux/slab.h>
+#include "cosm_main.h"
+
+/* Debugfs parent dir */
+static struct dentry *cosm_dbg;
+
+/**
+ * cosm_log_buf_show - Display MIC kernel log buffer
+ *
+ * log_buf addr/len is read from System.map by user space
+ * and populated in sysfs entries.
+ */
+static int cosm_log_buf_show(struct seq_file *s, void *unused)
+{
+ void __iomem *log_buf_va;
+ int __iomem *log_buf_len_va;
+ struct cosm_device *cdev = s->private;
+ void *kva;
+ int size;
+ u64 aper_offset;
+
+ if (!cdev || !cdev->log_buf_addr || !cdev->log_buf_len)
+ goto done;
+
+ mutex_lock(&cdev->cosm_mutex);
+ switch (cdev->state) {
+ case MIC_BOOTING:
+ case MIC_ONLINE:
+ case MIC_SHUTTING_DOWN:
+ break;
+ default:
+ goto unlock;
+ }
+
+ /*
+ * Card kernel will never be relocated and any kernel text/data mapping
+ * can be translated to phys address by subtracting __START_KERNEL_map.
+ */
+ aper_offset = (u64)cdev->log_buf_len - __START_KERNEL_map;
+ log_buf_len_va = cdev->hw_ops->aper(cdev)->va + aper_offset;
+ aper_offset = (u64)cdev->log_buf_addr - __START_KERNEL_map;
+ log_buf_va = cdev->hw_ops->aper(cdev)->va + aper_offset;
+
+ size = ioread32(log_buf_len_va);
+ kva = kmalloc(size, GFP_KERNEL);
+ if (!kva)
+ goto unlock;
+
+ memcpy_fromio(kva, log_buf_va, size);
+ seq_write(s, kva, size);
+ kfree(kva);
+unlock:
+ mutex_unlock(&cdev->cosm_mutex);
+done:
+ return 0;
+}
+
+static int cosm_log_buf_open(struct inode *inode, struct file *file)
+{
+ return single_open(file, cosm_log_buf_show, inode->i_private);
+}
+
+static const struct file_operations log_buf_ops = {
+ .owner = THIS_MODULE,
+ .open = cosm_log_buf_open,
+ .read = seq_read,
+ .llseek = seq_lseek,
+ .release = single_release
+};
+
+/**
+ * cosm_force_reset_show - Force MIC reset
+ *
+ * Invokes the force_reset COSM bus op instead of the standard reset
+ * op in case a force reset of the MIC device is required
+ */
+static int cosm_force_reset_show(struct seq_file *s, void *pos)
+{
+ struct cosm_device *cdev = s->private;
+
+ cosm_stop(cdev, true);
+ return 0;
+}
+
+static int cosm_force_reset_debug_open(struct inode *inode, struct file *file)
+{
+ return single_open(file, cosm_force_reset_show, inode->i_private);
+}
+
+static const struct file_operations force_reset_ops = {
+ .owner = THIS_MODULE,
+ .open = cosm_force_reset_debug_open,
+ .read = seq_read,
+ .llseek = seq_lseek,
+ .release = single_release
+};
+
+void cosm_create_debug_dir(struct cosm_device *cdev)
+{
+ char name[16];
+
+ if (!cosm_dbg)
+ return;
+
+ scnprintf(name, sizeof(name), "mic%d", cdev->index);
+ cdev->dbg_dir = debugfs_create_dir(name, cosm_dbg);
+ if (!cdev->dbg_dir)
+ return;
+
+ debugfs_create_file("log_buf", 0444, cdev->dbg_dir, cdev, &log_buf_ops);
+ debugfs_create_file("force_reset", 0444, cdev->dbg_dir, cdev,
+ &force_reset_ops);
+}
+
+void cosm_delete_debug_dir(struct cosm_device *cdev)
+{
+ if (!cdev->dbg_dir)
+ return;
+
+ debugfs_remove_recursive(cdev->dbg_dir);
+}
+
+void cosm_init_debugfs(void)
+{
+ cosm_dbg = debugfs_create_dir(KBUILD_MODNAME, NULL);
+ if (!cosm_dbg)
+ pr_err("can't create debugfs dir\n");
+}
+
+void cosm_exit_debugfs(void)
+{
+ debugfs_remove(cosm_dbg);
+}
diff --git a/drivers/misc/mic/cosm/cosm_main.c b/drivers/misc/mic/cosm/cosm_main.c
new file mode 100644
index 0000000..ee463c8
--- /dev/null
+++ b/drivers/misc/mic/cosm/cosm_main.c
@@ -0,0 +1,388 @@
+/*
+ * Intel MIC Platform Software Stack (MPSS)
+ *
+ * Copyright(c) 2015 Intel Corporation.
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License, version 2, as
+ * published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful, but
+ * WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+ * General Public License for more details.
+ *
+ * The full GNU General Public License is included in this distribution in
+ * the file called "COPYING".
+ *
+ * Intel MIC Coprocessor State Management (COSM) Driver
+ *
+ */
+
+#include <linux/module.h>
+#include <linux/delay.h>
+#include <linux/idr.h>
+#include <linux/slab.h>
+#include <linux/cred.h>
+#include "cosm_main.h"
+
+static const char cosm_driver_name[] = "mic";
+
+/* COSM ID allocator */
+static struct ida g_cosm_ida;
+/* Class of MIC devices for sysfs accessibility. */
+static struct class *g_cosm_class;
+/* Number of MIC devices */
+static atomic_t g_num_dev;
+
+/**
+ * cosm_hw_reset - Issue a HW reset for the MIC device
+ * @cdev: pointer to cosm_device instance
+ */
+static void cosm_hw_reset(struct cosm_device *cdev, bool force)
+{
+ int i;
+
+#define MIC_RESET_TO (45)
+ if (force && cdev->hw_ops->force_reset)
+ cdev->hw_ops->force_reset(cdev);
+ else
+ cdev->hw_ops->reset(cdev);
+
+ for (i = 0; i < MIC_RESET_TO; i++) {
+ if (cdev->hw_ops->ready(cdev)) {
+ cosm_set_state(cdev, MIC_READY);
+ return;
+ }
+ /*
+ * Resets typically take 10s of seconds to complete.
+ * Since an MMIO read is required to check if the
+ * firmware is ready or not, a 1 second delay works nicely.
+ */
+ msleep(1000);
+ }
+ cosm_set_state(cdev, MIC_RESET_FAILED);
+}
+
+/**
+ * cosm_start - Start the MIC
+ * @cdev: pointer to cosm_device instance
+ *
+ * This function prepares an MIC for boot and initiates boot.
+ * RETURNS: An appropriate -ERRNO error value on error, or 0 for success.
+ */
+int cosm_start(struct cosm_device *cdev)
+{
+ const struct cred *orig_cred;
+ struct cred *override_cred;
+ int rc;
+
+ mutex_lock(&cdev->cosm_mutex);
+ if (!cdev->bootmode) {
+ dev_err(&cdev->dev, "%s %d bootmode not set\n",
+ __func__, __LINE__);
+ rc = -EINVAL;
+ goto unlock_ret;
+ }
+retry:
+ if (MIC_READY != cdev->state) {
+ dev_err(&cdev->dev, "%s %d MIC state not READY\n",
+ __func__, __LINE__);
+ rc = -EINVAL;
+ goto unlock_ret;
+ }
+ if (!cdev->hw_ops->ready(cdev)) {
+ cosm_hw_reset(cdev, false);
+ /*
+ * The state will either be MIC_READY if the reset succeeded
+ * or MIC_RESET_FAILED if the firmware reset failed.
+ */
+ goto retry;
+ }
+
+ /*
+ * Set credentials to root to allow non-root user to download initramsfs
+ * with 600 permissions
+ */
+ override_cred = prepare_creds();
+ if (!override_cred) {
+ dev_err(&cdev->dev, "%s %d prepare_creds failed\n",
+ __func__, __LINE__);
+ rc = -ENOMEM;
+ goto unlock_ret;
+ }
+ override_cred->fsuid = GLOBAL_ROOT_UID;
+ orig_cred = override_creds(override_cred);
+
+ rc = cdev->hw_ops->start(cdev, cdev->index);
+
+ revert_creds(orig_cred);
+ put_cred(override_cred);
+ if (rc)
+ goto unlock_ret;
+
+ /*
+ * If linux is being booted, card is treated 'online' only
+ * when the scif interface in the card is up. If anything else
+ * is booted, we set card to 'online' immediately.
+ */
+ if (!strcmp(cdev->bootmode, "linux"))
+ cosm_set_state(cdev, MIC_BOOTING);
+ else
+ cosm_set_state(cdev, MIC_ONLINE);
+unlock_ret:
+ mutex_unlock(&cdev->cosm_mutex);
+ if (rc)
+ dev_err(&cdev->dev, "cosm_start failed rc %d\n", rc);
+ return rc;
+}
+
+/**
+ * cosm_stop - Prepare the MIC for reset and trigger reset
+ * @cdev: pointer to cosm_device instance
+ * @force: force a MIC to reset even if it is already reset and ready.
+ *
+ * RETURNS: None
+ */
+void cosm_stop(struct cosm_device *cdev, bool force)
+{
+ mutex_lock(&cdev->cosm_mutex);
+ if (MIC_READY != cdev->state || force) {
+ /*
+ * Don't call hw_ops if they have been called previously.
+ * stop(..) calls device_unregister and will crash the system if
+ * called multiple times.
+ */
+ bool call_hw_ops = MIC_RESET_FAILED != cdev->state &&
+ MIC_READY != cdev->state;
+
+ if (MIC_RESETTING != cdev->state)
+ cosm_set_state(cdev, MIC_RESETTING);
+ cdev->heartbeat_watchdog_enable = false;
+ if (call_hw_ops)
+ cdev->hw_ops->stop(cdev, force);
+ cosm_hw_reset(cdev, force);
+ cosm_set_shutdown_status(cdev, MIC_NOP);
+ if (call_hw_ops && cdev->hw_ops->post_reset)
+ cdev->hw_ops->post_reset(cdev, cdev->state);
+ }
+ mutex_unlock(&cdev->cosm_mutex);
+ flush_work(&cdev->scif_work);
+}
+
+/**
+ * cosm_reset_trigger_work - Trigger MIC reset
+ * @work: The work structure
+ *
+ * This work is scheduled whenever the host wants to reset the MIC.
+ */
+static void cosm_reset_trigger_work(struct work_struct *work)
+{
+ struct cosm_device *cdev = container_of(work, struct cosm_device,
+ reset_trigger_work);
+ cosm_stop(cdev, false);
+}
+
+/**
+ * cosm_reset - Schedule MIC reset
+ * @cdev: pointer to cosm_device instance
+ *
+ * RETURNS: An -EINVAL if the card is already READY or 0 for success.
+ */
+int cosm_reset(struct cosm_device *cdev)
+{
+ int rc = 0;
+
+ mutex_lock(&cdev->cosm_mutex);
+ if (MIC_READY != cdev->state) {
+ cosm_set_state(cdev, MIC_RESETTING);
+ schedule_work(&cdev->reset_trigger_work);
+ } else {
+ dev_err(&cdev->dev, "%s %d MIC is READY\n", __func__, __LINE__);
+ rc = -EINVAL;
+ }
+ mutex_unlock(&cdev->cosm_mutex);
+ return rc;
+}
+
+/**
+ * cosm_shutdown - Initiate MIC shutdown.
+ * @cdev: pointer to cosm_device instance
+ *
+ * RETURNS: None
+ */
+int cosm_shutdown(struct cosm_device *cdev)
+{
+ struct cosm_msg msg = { .id = COSM_MSG_SHUTDOWN };
+ int rc = 0;
+
+ mutex_lock(&cdev->cosm_mutex);
+ if (MIC_ONLINE != cdev->state) {
+ rc = -EINVAL;
+ dev_err(&cdev->dev, "%s %d skipping shutdown in state: %s\n",
+ __func__, __LINE__, cosm_state_string[cdev->state]);
+ goto err;
+ }
+
+ if (!cdev->epd) {
+ rc = -ENOTCONN;
+ dev_err(&cdev->dev, "%s %d scif endpoint not connected rc %d\n",
+ __func__, __LINE__, rc);
+ goto err;
+ }
+
+ rc = scif_send(cdev->epd, &msg, sizeof(msg), SCIF_SEND_BLOCK);
+ if (rc < 0) {
+ dev_err(&cdev->dev, "%s %d scif_send failed rc %d\n",
+ __func__, __LINE__, rc);
+ goto err;
+ }
+ cdev->heartbeat_watchdog_enable = false;
+ cosm_set_state(cdev, MIC_SHUTTING_DOWN);
+ rc = 0;
+err:
+ mutex_unlock(&cdev->cosm_mutex);
+ return rc;
+}
+
+static int cosm_driver_probe(struct cosm_device *cdev)
+{
+ int rc;
+
+ /* Initialize SCIF server at first probe */
+ if (1 == atomic_add_return(1, &g_num_dev)) {
+ rc = cosm_scif_init();
+ if (rc)
+ goto scif_exit;
+ }
+ mutex_init(&cdev->cosm_mutex);
+ INIT_WORK(&cdev->reset_trigger_work, cosm_reset_trigger_work);
+ INIT_WORK(&cdev->scif_work, cosm_scif_work);
+ cdev->sysfs_heartbeat_enable = true;
+ cosm_sysfs_init(cdev);
+ cdev->sdev = device_create_with_groups(g_cosm_class, cdev->dev.parent,
+ MKDEV(0, cdev->index), cdev, cdev->attr_group,
+ "mic%d", cdev->index);
+ if (IS_ERR(cdev->sdev)) {
+ rc = PTR_ERR(cdev->sdev);
+ dev_err(&cdev->dev, "device_create_with_groups failed rc %d\n",
+ rc);
+ goto scif_exit;
+ }
+
+ cdev->state_sysfs = sysfs_get_dirent(cdev->sdev->kobj.sd,
+ "state");
+ if (!cdev->state_sysfs) {
+ rc = -ENODEV;
+ dev_err(&cdev->dev, "sysfs_get_dirent failed rc %d\n", rc);
+ goto destroy_device;
+ }
+ cosm_create_debug_dir(cdev);
+ return 0;
+destroy_device:
+ device_destroy(g_cosm_class, MKDEV(0, cdev->index));
+scif_exit:
+ if (atomic_dec_and_test(&g_num_dev))
+ cosm_scif_exit();
+ return rc;
+}
+
+static void cosm_driver_remove(struct cosm_device *cdev)
+{
+ cosm_delete_debug_dir(cdev);
+ sysfs_put(cdev->state_sysfs);
+ device_destroy(g_cosm_class, MKDEV(0, cdev->index));
+ flush_work(&cdev->reset_trigger_work);
+ cosm_stop(cdev, false);
+ if (atomic_dec_and_test(&g_num_dev))
+ cosm_scif_exit();
+
+ /* These sysfs entries might have allocated */
+ kfree(cdev->cmdline);
+ kfree(cdev->firmware);
+ kfree(cdev->ramdisk);
+ kfree(cdev->bootmode);
+}
+
+static int cosm_suspend(struct device *dev)
+{
+ struct cosm_device *cdev = dev_to_cosm(dev);
+
+ mutex_lock(&cdev->cosm_mutex);
+ switch (cdev->state) {
+ /**
+ * Suspend/freeze hooks in userspace have already shutdown the card.
+ * Card should be 'ready' in most cases. It is however possible that
+ * some userspace application initiated a boot. In those cases, we
+ * simply reset the card.
+ */
+ case MIC_ONLINE:
+ case MIC_BOOTING:
+ case MIC_SHUTTING_DOWN:
+ mutex_unlock(&cdev->cosm_mutex);
+ cosm_stop(cdev, false);
+ break;
+ default:
+ mutex_unlock(&cdev->cosm_mutex);
+ break;
+ }
+ return 0;
+}
+
+static const struct dev_pm_ops cosm_pm_ops = {
+ .suspend = cosm_suspend,
+ .freeze = cosm_suspend
+};
+
+static struct cosm_driver cosm_driver = {
+ .driver = {
+ .name = KBUILD_MODNAME,
+ .owner = THIS_MODULE,
+ .pm = &cosm_pm_ops,
+ },
+ .probe = cosm_driver_probe,
+ .remove = cosm_driver_remove
+};
+
+static int __init cosm_init(void)
+{
+ int ret;
+
+ cosm_init_debugfs();
+
+ g_cosm_class = class_create(THIS_MODULE, cosm_driver_name);
+ if (IS_ERR(g_cosm_class)) {
+ ret = PTR_ERR(g_cosm_class);
+ pr_err("class_create failed ret %d\n", ret);
+ goto cleanup_debugfs;
+ }
+
+ ida_init(&g_cosm_ida);
+ ret = cosm_register_driver(&cosm_driver);
+ if (ret) {
+ pr_err("cosm_register_driver failed ret %d\n", ret);
+ goto ida_destroy;
+ }
+ return 0;
+ida_destroy:
+ ida_destroy(&g_cosm_ida);
+ class_destroy(g_cosm_class);
+cleanup_debugfs:
+ cosm_exit_debugfs();
+ return ret;
+}
+
+static void __exit cosm_exit(void)
+{
+ cosm_unregister_driver(&cosm_driver);
+ ida_destroy(&g_cosm_ida);
+ class_destroy(g_cosm_class);
+ cosm_exit_debugfs();
+}
+
+module_init(cosm_init);
+module_exit(cosm_exit);
+
+MODULE_AUTHOR("Intel Corporation");
+MODULE_DESCRIPTION("Intel(R) MIC Coprocessor State Management (COSM) Driver");
+MODULE_LICENSE("GPL v2");
diff --git a/drivers/misc/mic/cosm/cosm_main.h b/drivers/misc/mic/cosm/cosm_main.h
new file mode 100644
index 0000000..f01156f
--- /dev/null
+++ b/drivers/misc/mic/cosm/cosm_main.h
@@ -0,0 +1,70 @@
+/*
+ * Intel MIC Platform Software Stack (MPSS)
+ *
+ * Copyright(c) 2015 Intel Corporation.
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License, version 2, as
+ * published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful, but
+ * WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+ * General Public License for more details.
+ *
+ * The full GNU General Public License is included in this distribution in
+ * the file called "COPYING".
+ *
+ * Intel MIC Coprocessor State Management (COSM) Driver
+ *
+ */
+#ifndef _COSM_COSM_H_
+#define _COSM_COSM_H_
+
+#include <linux/scif.h>
+#include "../bus/cosm_bus.h"
+
+#define COSM_HEARTBEAT_SEND_SEC 30
+#define SCIF_COSM_LISTEN_PORT 201
+
+/**
+ * enum COSM msg id's
+ * @COSM_MSG_SHUTDOWN: host->card trigger shutdown
+ * @COSM_MSG_SYNC_TIME: host->card send host time to card to sync time
+ * @COSM_MSG_HEARTBEAT: card->host heartbeat
+ * @COSM_MSG_SHUTDOWN_STATUS: card->host with shutdown status as payload
+ */
+enum cosm_msg_id {
+ COSM_MSG_SHUTDOWN,
+ COSM_MSG_SYNC_TIME,
+ COSM_MSG_HEARTBEAT,
+ COSM_MSG_SHUTDOWN_STATUS,
+};
+
+struct cosm_msg {
+ u64 id;
+ union {
+ u64 shutdown_status;
+ struct timespec64 timespec;
+ };
+};
+
+extern const char * const cosm_state_string[];
+extern const char * const cosm_shutdown_status_string[];
+
+void cosm_sysfs_init(struct cosm_device *cdev);
+int cosm_start(struct cosm_device *cdev);
+void cosm_stop(struct cosm_device *cdev, bool force);
+int cosm_reset(struct cosm_device *cdev);
+int cosm_shutdown(struct cosm_device *cdev);
+void cosm_set_state(struct cosm_device *cdev, u8 state);
+void cosm_set_shutdown_status(struct cosm_device *cdev, u8 status);
+void cosm_init_debugfs(void);
+void cosm_exit_debugfs(void);
+void cosm_create_debug_dir(struct cosm_device *cdev);
+void cosm_delete_debug_dir(struct cosm_device *cdev);
+int cosm_scif_init(void);
+void cosm_scif_exit(void);
+void cosm_scif_work(struct work_struct *work);
+
+#endif
diff --git a/drivers/misc/mic/cosm/cosm_sysfs.c b/drivers/misc/mic/cosm/cosm_sysfs.c
new file mode 100644
index 0000000..aea3f8b
--- /dev/null
+++ b/drivers/misc/mic/cosm/cosm_sysfs.c
@@ -0,0 +1,461 @@
+/*
+ * Intel MIC Platform Software Stack (MPSS)
+ *
+ * Copyright(c) 2015 Intel Corporation.
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License, version 2, as
+ * published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful, but
+ * WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+ * General Public License for more details.
+ *
+ * The full GNU General Public License is included in this distribution in
+ * the file called "COPYING".
+ *
+ * Intel MIC Coprocessor State Management (COSM) Driver
+ *
+ */
+#include <linux/slab.h>
+#include "cosm_main.h"
+
+/*
+ * A state-to-string lookup table, for exposing a human readable state
+ * via sysfs. Always keep in sync with enum cosm_states
+ */
+const char * const cosm_state_string[] = {
+ [MIC_READY] = "ready",
+ [MIC_BOOTING] = "booting",
+ [MIC_ONLINE] = "online",
+ [MIC_SHUTTING_DOWN] = "shutting_down",
+ [MIC_RESETTING] = "resetting",
+ [MIC_RESET_FAILED] = "reset_failed",
+};
+
+/*
+ * A shutdown-status-to-string lookup table, for exposing a human
+ * readable state via sysfs. Always keep in sync with enum cosm_shutdown_status
+ */
+const char * const cosm_shutdown_status_string[] = {
+ [MIC_NOP] = "nop",
+ [MIC_CRASHED] = "crashed",
+ [MIC_HALTED] = "halted",
+ [MIC_POWER_OFF] = "poweroff",
+ [MIC_RESTART] = "restart",
+};
+
+void cosm_set_shutdown_status(struct cosm_device *cdev, u8 shutdown_status)
+{
+ dev_dbg(&cdev->dev, "Shutdown Status %s -> %s\n",
+ cosm_shutdown_status_string[cdev->shutdown_status],
+ cosm_shutdown_status_string[shutdown_status]);
+ cdev->shutdown_status = shutdown_status;
+}
+
+void cosm_set_state(struct cosm_device *cdev, u8 state)
+{
+ dev_dbg(&cdev->dev, "State %s -> %s\n",
+ cosm_state_string[cdev->state],
+ cosm_state_string[state]);
+ cdev->state = state;
+ sysfs_notify_dirent(cdev->state_sysfs);
+}
+
+static ssize_t
+family_show(struct device *dev, struct device_attribute *attr, char *buf)
+{
+ struct cosm_device *cdev = dev_get_drvdata(dev);
+
+ if (!cdev)
+ return -EINVAL;
+
+ return cdev->hw_ops->family(cdev, buf);
+}
+static DEVICE_ATTR_RO(family);
+
+static ssize_t
+stepping_show(struct device *dev, struct device_attribute *attr, char *buf)
+{
+ struct cosm_device *cdev = dev_get_drvdata(dev);
+
+ if (!cdev)
+ return -EINVAL;
+
+ return cdev->hw_ops->stepping(cdev, buf);
+}
+static DEVICE_ATTR_RO(stepping);
+
+static ssize_t
+state_show(struct device *dev, struct device_attribute *attr, char *buf)
+{
+ struct cosm_device *cdev = dev_get_drvdata(dev);
+
+ if (!cdev || cdev->state >= MIC_LAST)
+ return -EINVAL;
+
+ return scnprintf(buf, PAGE_SIZE, "%s\n",
+ cosm_state_string[cdev->state]);
+}
+
+static ssize_t
+state_store(struct device *dev, struct device_attribute *attr,
+ const char *buf, size_t count)
+{
+ struct cosm_device *cdev = dev_get_drvdata(dev);
+ int rc;
+
+ if (!cdev)
+ return -EINVAL;
+
+ if (sysfs_streq(buf, "boot")) {
+ rc = cosm_start(cdev);
+ goto done;
+ }
+ if (sysfs_streq(buf, "reset")) {
+ rc = cosm_reset(cdev);
+ goto done;
+ }
+
+ if (sysfs_streq(buf, "shutdown")) {
+ rc = cosm_shutdown(cdev);
+ goto done;
+ }
+ rc = -EINVAL;
+done:
+ if (rc)
+ count = rc;
+ return count;
+}
+static DEVICE_ATTR_RW(state);
+
+static ssize_t shutdown_status_show(struct device *dev,
+ struct device_attribute *attr, char *buf)
+{
+ struct cosm_device *cdev = dev_get_drvdata(dev);
+
+ if (!cdev || cdev->shutdown_status >= MIC_STATUS_LAST)
+ return -EINVAL;
+
+ return scnprintf(buf, PAGE_SIZE, "%s\n",
+ cosm_shutdown_status_string[cdev->shutdown_status]);
+}
+static DEVICE_ATTR_RO(shutdown_status);
+
+static ssize_t
+heartbeat_enable_show(struct device *dev,
+ struct device_attribute *attr, char *buf)
+{
+ struct cosm_device *cdev = dev_get_drvdata(dev);
+
+ if (!cdev)
+ return -EINVAL;
+
+ return scnprintf(buf, PAGE_SIZE, "%d\n", cdev->sysfs_heartbeat_enable);
+}
+
+static ssize_t
+heartbeat_enable_store(struct device *dev,
+ struct device_attribute *attr,
+ const char *buf, size_t count)
+{
+ struct cosm_device *cdev = dev_get_drvdata(dev);
+ int enable;
+ int ret;
+
+ if (!cdev)
+ return -EINVAL;
+
+ mutex_lock(&cdev->cosm_mutex);
+ ret = kstrtoint(buf, 10, &enable);
+ if (ret)
+ goto unlock;
+
+ cdev->sysfs_heartbeat_enable = enable;
+ /* if state is not online, cdev->heartbeat_watchdog_enable is 0 */
+ if (MIC_ONLINE == cdev->state)
+ cdev->heartbeat_watchdog_enable = enable;
+ ret = count;
+unlock:
+ mutex_unlock(&cdev->cosm_mutex);
+ return ret;
+}
+static DEVICE_ATTR_RW(heartbeat_enable);
+
+static ssize_t
+cmdline_show(struct device *dev, struct device_attribute *attr, char *buf)
+{
+ struct cosm_device *cdev = dev_get_drvdata(dev);
+ char *cmdline;
+
+ if (!cdev)
+ return -EINVAL;
+
+ cmdline = cdev->cmdline;
+
+ if (cmdline)
+ return scnprintf(buf, PAGE_SIZE, "%s\n", cmdline);
+ return 0;
+}
+
+static ssize_t
+cmdline_store(struct device *dev, struct device_attribute *attr,
+ const char *buf, size_t count)
+{
+ struct cosm_device *cdev = dev_get_drvdata(dev);
+
+ if (!cdev)
+ return -EINVAL;
+
+ mutex_lock(&cdev->cosm_mutex);
+ kfree(cdev->cmdline);
+
+ cdev->cmdline = kmalloc(count + 1, GFP_KERNEL);
+ if (!cdev->cmdline) {
+ count = -ENOMEM;
+ goto unlock;
+ }
+
+ strncpy(cdev->cmdline, buf, count);
+
+ if (cdev->cmdline[count - 1] == '\n')
+ cdev->cmdline[count - 1] = '\0';
+ else
+ cdev->cmdline[count] = '\0';
+unlock:
+ mutex_unlock(&cdev->cosm_mutex);
+ return count;
+}
+static DEVICE_ATTR_RW(cmdline);
+
+static ssize_t
+firmware_show(struct device *dev, struct device_attribute *attr, char *buf)
+{
+ struct cosm_device *cdev = dev_get_drvdata(dev);
+ char *firmware;
+
+ if (!cdev)
+ return -EINVAL;
+
+ firmware = cdev->firmware;
+
+ if (firmware)
+ return scnprintf(buf, PAGE_SIZE, "%s\n", firmware);
+ return 0;
+}
+
+static ssize_t
+firmware_store(struct device *dev, struct device_attribute *attr,
+ const char *buf, size_t count)
+{
+ struct cosm_device *cdev = dev_get_drvdata(dev);
+
+ if (!cdev)
+ return -EINVAL;
+
+ mutex_lock(&cdev->cosm_mutex);
+ kfree(cdev->firmware);
+
+ cdev->firmware = kmalloc(count + 1, GFP_KERNEL);
+ if (!cdev->firmware) {
+ count = -ENOMEM;
+ goto unlock;
+ }
+ strncpy(cdev->firmware, buf, count);
+
+ if (cdev->firmware[count - 1] == '\n')
+ cdev->firmware[count - 1] = '\0';
+ else
+ cdev->firmware[count] = '\0';
+unlock:
+ mutex_unlock(&cdev->cosm_mutex);
+ return count;
+}
+static DEVICE_ATTR_RW(firmware);
+
+static ssize_t
+ramdisk_show(struct device *dev, struct device_attribute *attr, char *buf)
+{
+ struct cosm_device *cdev = dev_get_drvdata(dev);
+ char *ramdisk;
+
+ if (!cdev)
+ return -EINVAL;
+
+ ramdisk = cdev->ramdisk;
+
+ if (ramdisk)
+ return scnprintf(buf, PAGE_SIZE, "%s\n", ramdisk);
+ return 0;
+}
+
+static ssize_t
+ramdisk_store(struct device *dev, struct device_attribute *attr,
+ const char *buf, size_t count)
+{
+ struct cosm_device *cdev = dev_get_drvdata(dev);
+
+ if (!cdev)
+ return -EINVAL;
+
+ mutex_lock(&cdev->cosm_mutex);
+ kfree(cdev->ramdisk);
+
+ cdev->ramdisk = kmalloc(count + 1, GFP_KERNEL);
+ if (!cdev->ramdisk) {
+ count = -ENOMEM;
+ goto unlock;
+ }
+
+ strncpy(cdev->ramdisk, buf, count);
+
+ if (cdev->ramdisk[count - 1] == '\n')
+ cdev->ramdisk[count - 1] = '\0';
+ else
+ cdev->ramdisk[count] = '\0';
+unlock:
+ mutex_unlock(&cdev->cosm_mutex);
+ return count;
+}
+static DEVICE_ATTR_RW(ramdisk);
+
+static ssize_t
+bootmode_show(struct device *dev, struct device_attribute *attr, char *buf)
+{
+ struct cosm_device *cdev = dev_get_drvdata(dev);
+ char *bootmode;
+
+ if (!cdev)
+ return -EINVAL;
+
+ bootmode = cdev->bootmode;
+
+ if (bootmode)
+ return scnprintf(buf, PAGE_SIZE, "%s\n", bootmode);
+ return 0;
+}
+
+static ssize_t
+bootmode_store(struct device *dev, struct device_attribute *attr,
+ const char *buf, size_t count)
+{
+ struct cosm_device *cdev = dev_get_drvdata(dev);
+
+ if (!cdev)
+ return -EINVAL;
+
+ if (!sysfs_streq(buf, "linux") && !sysfs_streq(buf, "flash"))
+ return -EINVAL;
+
+ mutex_lock(&cdev->cosm_mutex);
+ kfree(cdev->bootmode);
+
+ cdev->bootmode = kmalloc(count + 1, GFP_KERNEL);
+ if (!cdev->bootmode) {
+ count = -ENOMEM;
+ goto unlock;
+ }
+
+ strncpy(cdev->bootmode, buf, count);
+
+ if (cdev->bootmode[count - 1] == '\n')
+ cdev->bootmode[count - 1] = '\0';
+ else
+ cdev->bootmode[count] = '\0';
+unlock:
+ mutex_unlock(&cdev->cosm_mutex);
+ return count;
+}
+static DEVICE_ATTR_RW(bootmode);
+
+static ssize_t
+log_buf_addr_show(struct device *dev, struct device_attribute *attr,
+ char *buf)
+{
+ struct cosm_device *cdev = dev_get_drvdata(dev);
+
+ if (!cdev)
+ return -EINVAL;
+
+ return scnprintf(buf, PAGE_SIZE, "%p\n", cdev->log_buf_addr);
+}
+
+static ssize_t
+log_buf_addr_store(struct device *dev, struct device_attribute *attr,
+ const char *buf, size_t count)
+{
+ struct cosm_device *cdev = dev_get_drvdata(dev);
+ int ret;
+ unsigned long addr;
+
+ if (!cdev)
+ return -EINVAL;
+
+ ret = kstrtoul(buf, 16, &addr);
+ if (ret)
+ goto exit;
+
+ cdev->log_buf_addr = (void *)addr;
+ ret = count;
+exit:
+ return ret;
+}
+static DEVICE_ATTR_RW(log_buf_addr);
+
+static ssize_t
+log_buf_len_show(struct device *dev, struct device_attribute *attr,
+ char *buf)
+{
+ struct cosm_device *cdev = dev_get_drvdata(dev);
+
+ if (!cdev)
+ return -EINVAL;
+
+ return scnprintf(buf, PAGE_SIZE, "%p\n", cdev->log_buf_len);
+}
+
+static ssize_t
+log_buf_len_store(struct device *dev, struct device_attribute *attr,
+ const char *buf, size_t count)
+{
+ struct cosm_device *cdev = dev_get_drvdata(dev);
+ int ret;
+ unsigned long addr;
+
+ if (!cdev)
+ return -EINVAL;
+
+ ret = kstrtoul(buf, 16, &addr);
+ if (ret)
+ goto exit;
+
+ cdev->log_buf_len = (int *)addr;
+ ret = count;
+exit:
+ return ret;
+}
+static DEVICE_ATTR_RW(log_buf_len);
+
+static struct attribute *cosm_default_attrs[] = {
+ &dev_attr_family.attr,
+ &dev_attr_stepping.attr,
+ &dev_attr_state.attr,
+ &dev_attr_shutdown_status.attr,
+ &dev_attr_heartbeat_enable.attr,
+ &dev_attr_cmdline.attr,
+ &dev_attr_firmware.attr,
+ &dev_attr_ramdisk.attr,
+ &dev_attr_bootmode.attr,
+ &dev_attr_log_buf_addr.attr,
+ &dev_attr_log_buf_len.attr,
+
+ NULL
+};
+
+ATTRIBUTE_GROUPS(cosm_default);
+
+void cosm_sysfs_init(struct cosm_device *cdev)
+{
+ cdev->attr_group = cosm_default_groups;
+}
--
2.0.0.rc3.2.g998f840
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
[toc] | [prev] | [next] | [standalone]
| From | Ashutosh Dixit <ashutosh.dixit@intel.com> |
|---|---|
| Date | 2015-07-28 01:40 +0200 |
| Subject | [PATCH char-misc-next 13/19] misc: mic: SCIF RMA header file |
| Message-ID | <pR0a1-8vE-89@gated-at.bofh.it> |
| In reply to | #1193568 |
From: Sudeep Dutt <sudeep.dutt@intel.com>
This patch adds the internal data structures required to perform SCIF
RMAs. The data structures required to maintain per SCIF endpoint, RMA
information are contained in scif_endpt_rma_info. scif_pinned_pages
describes a set of SCIF pinned pages maintained locally. The
scif_window is a data structure which contains all the fields required
to describe a SCIF registered window on self and remote nodes. It
contains an offset which is used as a key to perform SCIF DMAs and CPU
copies between self and remote registered windows.
Reviewed-by: Nikhil Rao <nikhil.rao@intel.com>
Reviewed-by: Ashutosh Dixit <ashutosh.dixit@intel.com>
Signed-off-by: Sudeep Dutt <sudeep.dutt@intel.com>
---
drivers/misc/mic/scif/scif_rma.h | 464 +++++++++++++++++++++++++++++++++++++++
1 file changed, 464 insertions(+)
create mode 100644 drivers/misc/mic/scif/scif_rma.h
diff --git a/drivers/misc/mic/scif/scif_rma.h b/drivers/misc/mic/scif/scif_rma.h
new file mode 100644
index 0000000..fa67222
--- /dev/null
+++ b/drivers/misc/mic/scif/scif_rma.h
@@ -0,0 +1,464 @@
+/*
+ * Intel MIC Platform Software Stack (MPSS)
+ *
+ * This file is provided under a dual BSD/GPLv2 license. When using or
+ * redistributing this file, you may do so under either license.
+ *
+ * GPL LICENSE SUMMARY
+ *
+ * Copyright(c) 2015 Intel Corporation.
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of version 2 of the GNU General Public License as
+ * published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful, but
+ * WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+ * General Public License for more details.
+ *
+ * BSD LICENSE
+ *
+ * Copyright(c) 2015 Intel Corporation.
+ *
+ * Redistribution and use in source and binary forms, with or without
+ * modification, are permitted provided that the following conditions
+ * are met:
+ *
+ * * Redistributions of source code must retain the above copyright
+ * notice, this list of conditions and the following disclaimer.
+ * * Redistributions in binary form must reproduce the above copyright
+ * notice, this list of conditions and the following disclaimer in
+ * the documentation and/or other materials provided with the
+ * distribution.
+ * * Neither the name of Intel Corporation nor the names of its
+ * contributors may be used to endorse or promote products derived
+ * from this software without specific prior written permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+ * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+ * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+ * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+ * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+ * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+ * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+ * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+ * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ * Intel SCIF driver.
+ *
+ */
+#ifndef SCIF_RMA_H
+#define SCIF_RMA_H
+
+#include <linux/dma_remapping.h>
+#include <linux/mmu_notifier.h>
+
+#include "../bus/scif_bus.h"
+
+/* If this bit is set then the mark is a remote fence mark */
+#define SCIF_REMOTE_FENCE_BIT 31
+/* Magic value used to indicate a remote fence request */
+#define SCIF_REMOTE_FENCE BIT_ULL(SCIF_REMOTE_FENCE_BIT)
+
+#define SCIF_MAX_UNALIGNED_BUF_SIZE (1024 * 1024ULL)
+#define SCIF_KMEM_UNALIGNED_BUF_SIZE (SCIF_MAX_UNALIGNED_BUF_SIZE + \
+ (L1_CACHE_BYTES << 1))
+
+#define SCIF_IOVA_START_PFN (1)
+#define SCIF_IOVA_PFN(addr) ((addr) >> PAGE_SHIFT)
+#define SCIF_DMA_64BIT_PFN SCIF_IOVA_PFN(DMA_BIT_MASK(64))
+#define SCIF_DMA_63BIT_PFN SCIF_IOVA_PFN(DMA_BIT_MASK(63))
+
+/*
+ * struct scif_endpt_rma_info - Per Endpoint Remote Memory Access Information
+ *
+ * @reg_list: List of registration windows for self
+ * @remote_reg_list: List of registration windows for peer
+ * @iovad: Offset generator
+ * @rma_lock: Synchronizes access to self/remote list and also protects the
+ * window from being destroyed while RMAs are in progress.
+ * @tc_lock: Synchronizes access to temporary cached windows list
+ * for SCIF Registration Caching.
+ * @mmn_lock: Synchronizes access to the list of MMU notifiers registered
+ * @tw_refcount: Keeps track of number of outstanding temporary registered
+ * windows created by scif_vreadfrom/scif_vwriteto which have
+ * not been destroyed.
+ * @tcw_refcount: Same as tw_refcount but for temporary cached windows
+ * @tcw_total_pages: Same as tcw_refcount but in terms of pages pinned
+ * @mmn_list: MMU notifier so that we can destroy the windows when required
+ * @fence_refcount: Keeps track of number of outstanding remote fence
+ * requests which have been received by the peer.
+ * @dma_chan: DMA channel used for all DMA transfers for this endpoint.
+ * @async_list_del: Detect asynchronous list entry deletion
+ * @vma_list: List of vmas with remote memory mappings
+ * @markwq: Wait queue used for scif_fence_mark/scif_fence_wait
+*/
+struct scif_endpt_rma_info {
+ struct list_head reg_list;
+ struct list_head remote_reg_list;
+ struct iova_domain iovad;
+ struct mutex rma_lock;
+ spinlock_t tc_lock;
+ struct mutex mmn_lock;
+ atomic_t tw_refcount;
+ atomic_t tcw_refcount;
+ atomic_t tcw_total_pages;
+ struct list_head mmn_list;
+ atomic_t fence_refcount;
+ struct dma_chan *dma_chan;
+ int async_list_del;
+ struct list_head vma_list;
+ wait_queue_head_t markwq;
+};
+
+/*
+ * struct scif_fence_info - used for tracking fence requests
+ *
+ * @state: State of this transfer
+ * @wq: Fences wait on this queue
+ * @dma_mark: Used for storing the DMA mark
+ */
+struct scif_fence_info {
+ enum scif_msg_state state;
+ struct completion comp;
+ int dma_mark;
+};
+
+/*
+ * struct scif_remote_fence_info - used for tracking remote fence requests
+ *
+ * @msg: List of SCIF node QP fence messages
+ * @list: Link to list of remote fence requests
+ */
+struct scif_remote_fence_info {
+ struct scifmsg msg;
+ struct list_head list;
+};
+
+/*
+ * Specifies whether an RMA operation can span across partial windows, a single
+ * window or multiple contiguous windows. Mmaps can span across partial windows.
+ * Unregistration can span across complete windows. scif_get_pages() can span a
+ * single window. A window can also be of type self or peer.
+ */
+enum scif_window_type {
+ SCIF_WINDOW_PARTIAL,
+ SCIF_WINDOW_SINGLE,
+ SCIF_WINDOW_FULL,
+ SCIF_WINDOW_SELF,
+ SCIF_WINDOW_PEER
+};
+
+/* The number of physical addresses that can be stored in a PAGE. */
+#define SCIF_NR_ADDR_IN_PAGE (0x1000 >> 3)
+
+/*
+ * struct scif_rma_lookup - RMA lookup data structure for page list transfers
+ *
+ * Store an array of lookup offsets. Each offset in this array maps
+ * one 4K page containing 512 physical addresses i.e. 2MB. 512 such
+ * offsets in a 4K page will correspond to 1GB of registered address space.
+
+ * @lookup: Array of offsets
+ * @offset: DMA offset of lookup array
+ */
+struct scif_rma_lookup {
+ dma_addr_t *lookup;
+ dma_addr_t offset;
+};
+
+/*
+ * struct scif_pinned_pages - A set of pinned pages obtained with
+ * scif_pin_pages() which could be part of multiple registered
+ * windows across different end points.
+ *
+ * @nr_pages: Number of pages which is defined as a s64 instead of an int
+ * to avoid sign extension with buffers >= 2GB
+ * @prot: read/write protections
+ * @map_flags: Flags specified during the pin operation
+ * @ref_count: Reference count bumped in terms of number of pages
+ * @magic: A magic value
+ * @pages: Array of pointers to struct pages populated with get_user_pages(..)
+ */
+struct scif_pinned_pages {
+ s64 nr_pages;
+ int prot;
+ int map_flags;
+ atomic_t ref_count;
+ u64 magic;
+ struct page **pages;
+};
+
+/*
+ * struct scif_status - Stores DMA status update information
+ *
+ * @src_dma_addr: Source buffer DMA address
+ * @val: src location for value to be written to the destination
+ * @ep: SCIF endpoint
+ */
+struct scif_status {
+ dma_addr_t src_dma_addr;
+ u64 val;
+ struct scif_endpt *ep;
+};
+
+/*
+ * struct scif_window - Registration Window for Self and Remote
+ *
+ * @nr_pages: Number of pages which is defined as a s64 instead of an int
+ * to avoid sign extension with buffers >= 2GB
+ * @nr_contig_chunks: Number of contiguous physical chunks
+ * @prot: read/write protections
+ * @ref_count: reference count in terms of number of pages
+ * @magic: Cookie to detect corruption
+ * @offset: registered offset
+ * @va_for_temp: va address that this window represents
+ * @dma_mark: Used to determine if all DMAs against the window are done
+ * @ep: Pointer to EP. Useful for passing EP around with messages to
+ avoid expensive list traversals.
+ * @list: link to list of windows for the endpoint
+ * @type: self or peer window
+ * @peer_window: Pointer to peer window. Useful for sending messages to peer
+ * without requiring an extra list traversal
+ * @unreg_state: unregistration state
+ * @offset_freed: True if the offset has been freed
+ * @temp: True for temporary windows created via scif_vreadfrom/scif_vwriteto
+ * @mm: memory descriptor for the task_struct which initiated the RMA
+ * @st: scatter gather table for DMA mappings with IOMMU enabled
+ * @pinned_pages: The set of pinned_pages backing this window
+ * @alloc_handle: Handle for sending ALLOC_REQ
+ * @regwq: Wait Queue for an registration (N)ACK
+ * @reg_state: Registration state
+ * @unregwq: Wait Queue for an unregistration (N)ACK
+ * @dma_addr_lookup: Lookup for physical addresses used for DMA
+ * @nr_lookup: Number of entries in lookup
+ * @mapped_offset: Offset used to map the window by the peer
+ * @dma_addr: Array of physical addresses used for Mgmt node & MIC initiated DMA
+ * @num_pages: Array specifying number of pages for each physical address
+ */
+struct scif_window {
+ s64 nr_pages;
+ int nr_contig_chunks;
+ int prot;
+ int ref_count;
+ u64 magic;
+ s64 offset;
+ unsigned long va_for_temp;
+ int dma_mark;
+ u64 ep;
+ struct list_head list;
+ enum scif_window_type type;
+ u64 peer_window;
+ enum scif_msg_state unreg_state;
+ bool offset_freed;
+ bool temp;
+ struct mm_struct *mm;
+ struct sg_table *st;
+ union {
+ struct {
+ struct scif_pinned_pages *pinned_pages;
+ struct scif_allocmsg alloc_handle;
+ wait_queue_head_t regwq;
+ enum scif_msg_state reg_state;
+ wait_queue_head_t unregwq;
+ };
+ struct {
+ struct scif_rma_lookup dma_addr_lookup;
+ struct scif_rma_lookup num_pages_lookup;
+ int nr_lookup;
+ dma_addr_t mapped_offset;
+ };
+ };
+ dma_addr_t *dma_addr;
+ u64 *num_pages;
+} __packed;
+
+/*
+ * scif_mmu_notif - SCIF mmu notifier information
+ *
+ * @mmu_notifier ep_mmu_notifier: MMU notifier operations
+ * @tc_reg_list: List of temp registration windows for self
+ * @mm: memory descriptor for the task_struct which initiated the RMA
+ * @ep: SCIF endpoint
+ * @list: link to list of MMU notifier information
+ */
+struct scif_mmu_notif {
+#ifdef CONFIG_MMU_NOTIFIER
+ struct mmu_notifier ep_mmu_notifier;
+#endif
+ struct list_head tc_reg_list;
+ struct mm_struct *mm;
+ struct scif_endpt *ep;
+ struct list_head list;
+};
+
+enum scif_rma_dir {
+ SCIF_LOCAL_TO_REMOTE,
+ SCIF_REMOTE_TO_LOCAL
+};
+
+extern struct kmem_cache *unaligned_cache;
+/* Initialize RMA for this EP */
+void scif_rma_ep_init(struct scif_endpt *ep);
+/* Check if epd can be uninitialized */
+int scif_rma_ep_can_uninit(struct scif_endpt *ep);
+/* Obtain a new offset. Callee must grab RMA lock */
+int scif_get_window_offset(struct scif_endpt *ep, int flags,
+ s64 offset, int nr_pages, s64 *out_offset);
+/* Free offset. Callee must grab RMA lock */
+void scif_free_window_offset(struct scif_endpt *ep,
+ struct scif_window *window, s64 offset);
+/* Create self registration window */
+struct scif_window *scif_create_window(struct scif_endpt *ep, int nr_pages,
+ s64 offset, bool temp);
+/* Destroy self registration window.*/
+int scif_destroy_window(struct scif_endpt *ep, struct scif_window *window);
+void scif_unmap_window(struct scif_dev *remote_dev, struct scif_window *window);
+/* Map pages of self window to Aperture/PCI */
+int scif_map_window(struct scif_dev *remote_dev,
+ struct scif_window *window);
+/* Unregister a self window */
+int scif_unregister_window(struct scif_window *window);
+/* Destroy remote registration window */
+void
+scif_destroy_remote_window(struct scif_window *window);
+/* remove valid remote memory mappings from process address space */
+void scif_zap_mmaps(int node);
+/* Query if any applications have remote memory mappings */
+bool scif_rma_do_apps_have_mmaps(int node);
+/* Cleanup remote registration lists for zombie endpoints */
+void scif_cleanup_rma_for_zombies(int node);
+/* Reserve a DMA channel for a particular endpoint */
+int scif_reserve_dma_chan(struct scif_endpt *ep);
+/* Setup a DMA mark for an endpoint */
+int _scif_fence_mark(scif_epd_t epd, int *mark);
+int scif_prog_signal(scif_epd_t epd, off_t offset, u64 val,
+ enum scif_window_type type);
+void scif_alloc_req(struct scif_dev *scifdev, struct scifmsg *msg);
+void scif_alloc_gnt_rej(struct scif_dev *scifdev, struct scifmsg *msg);
+void scif_free_virt(struct scif_dev *scifdev, struct scifmsg *msg);
+void scif_recv_reg(struct scif_dev *scifdev, struct scifmsg *msg);
+void scif_recv_unreg(struct scif_dev *scifdev, struct scifmsg *msg);
+void scif_recv_reg_ack(struct scif_dev *scifdev, struct scifmsg *msg);
+void scif_recv_reg_nack(struct scif_dev *scifdev, struct scifmsg *msg);
+void scif_recv_unreg_ack(struct scif_dev *scifdev, struct scifmsg *msg);
+void scif_recv_unreg_nack(struct scif_dev *scifdev, struct scifmsg *msg);
+void scif_recv_munmap(struct scif_dev *scifdev, struct scifmsg *msg);
+void scif_recv_mark(struct scif_dev *scifdev, struct scifmsg *msg);
+void scif_recv_mark_resp(struct scif_dev *scifdev, struct scifmsg *msg);
+void scif_recv_wait(struct scif_dev *scifdev, struct scifmsg *msg);
+void scif_recv_wait_resp(struct scif_dev *scifdev, struct scifmsg *msg);
+void scif_recv_sig_local(struct scif_dev *scifdev, struct scifmsg *msg);
+void scif_recv_sig_remote(struct scif_dev *scifdev, struct scifmsg *msg);
+void scif_recv_sig_resp(struct scif_dev *scifdev, struct scifmsg *msg);
+void scif_mmu_notif_handler(struct work_struct *work);
+void scif_rma_handle_remote_fences(void);
+void scif_rma_destroy_windows(void);
+void scif_rma_destroy_tcw_invalid(void);
+int scif_drain_dma_intr(struct scif_hw_dev *sdev, struct dma_chan *chan);
+
+struct scif_window_iter {
+ s64 offset;
+ int index;
+};
+
+static inline void
+scif_init_window_iter(struct scif_window *window, struct scif_window_iter *iter)
+{
+ iter->offset = window->offset;
+ iter->index = 0;
+}
+
+dma_addr_t scif_off_to_dma_addr(struct scif_window *window, s64 off,
+ size_t *nr_bytes,
+ struct scif_window_iter *iter);
+static inline
+dma_addr_t __scif_off_to_dma_addr(struct scif_window *window, s64 off)
+{
+ return scif_off_to_dma_addr(window, off, NULL, NULL);
+}
+
+static inline bool scif_unaligned(off_t src_offset, off_t dst_offset)
+{
+ src_offset = src_offset & (L1_CACHE_BYTES - 1);
+ dst_offset = dst_offset & (L1_CACHE_BYTES - 1);
+ return !(src_offset == dst_offset);
+}
+
+/*
+ * scif_zalloc:
+ * @size: Size of the allocation request.
+ *
+ * Helper API which attempts to allocate zeroed pages via
+ * __get_free_pages(..) first and then falls back on
+ * vzalloc(..) if that fails.
+ */
+static inline void *scif_zalloc(size_t size)
+{
+ void *ret = NULL;
+ size_t align = ALIGN(size, PAGE_SIZE);
+
+ if (align && get_order(align) < MAX_ORDER)
+ ret = (void *)__get_free_pages(GFP_KERNEL | __GFP_ZERO,
+ get_order(align));
+ return ret ? ret : vzalloc(align);
+}
+
+/*
+ * scif_free:
+ * @addr: Address to be freed.
+ * @size: Size of the allocation.
+ * Helper API which frees memory allocated via scif_zalloc().
+ */
+static inline void scif_free(void *addr, size_t size)
+{
+ size_t align = ALIGN(size, PAGE_SIZE);
+
+ if (is_vmalloc_addr(addr))
+ vfree(addr);
+ else
+ free_pages((unsigned long)addr, get_order(align));
+}
+
+static inline void scif_get_window(struct scif_window *window, int nr_pages)
+{
+ window->ref_count += nr_pages;
+}
+
+static inline void scif_put_window(struct scif_window *window, int nr_pages)
+{
+ window->ref_count -= nr_pages;
+}
+
+static inline void scif_set_window_ref(struct scif_window *window, int nr_pages)
+{
+ window->ref_count = nr_pages;
+}
+
+static inline void
+scif_queue_for_cleanup(struct scif_window *window, struct list_head *list)
+{
+ spin_lock(&scif_info.rmalock);
+ list_add_tail(&window->list, list);
+ spin_unlock(&scif_info.rmalock);
+ schedule_work(&scif_info.misc_work);
+}
+
+static inline void __scif_rma_destroy_tcw_helper(struct scif_window *window)
+{
+ list_del_init(&window->list);
+ scif_queue_for_cleanup(window, &scif_info.rma_tc);
+}
+
+static inline bool scif_is_iommu_enabled(void)
+{
+#ifdef CONFIG_INTEL_IOMMU
+ return intel_iommu_enabled;
+#else
+ return false;
+#endif
+}
+#endif /* SCIF_RMA_H */
--
2.0.0.rc3.2.g998f840
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
[toc] | [prev] | [next] | [standalone]
| From | Ashutosh Dixit <ashutosh.dixit@intel.com> |
|---|---|
| Date | 2015-07-28 01:40 +0200 |
| Subject | [PATCH char-misc-next 05/19] misc: mic: COSM SCIF server |
| Message-ID | <pR0a1-8vE-93@gated-at.bofh.it> |
| In reply to | #1193568 |
The COSM driver communicates with the MIC cards over SCIF. A SCIF
"server" listens for incoming connections from "client" MIC cards as
they boot. After the connection is accepted a separate work item is
scheduled for each MIC card. This work item normally stays blocked in
scif_poll but wakes up to process messages from the card.
The SCIF connection between the host and card COSM components is used
to (a) send the command to shut down the card (b) receive shutdown
status back from the card upon completion of shutdown (c) receive
periodic heartbeat messages to detect card crashes (d) send host time
to the card to enable the card to sync its time to the host.
Reviewed-by: Nikhil Rao <nikhil.rao@intel.com>
Reviewed-by: Sudeep Dutt <sudeep.dutt@intel.com>
Signed-off-by: Ashutosh Dixit <ashutosh.dixit@intel.com>
---
drivers/misc/mic/cosm/cosm_scif_server.c | 405 +++++++++++++++++++++++++++++++
1 file changed, 405 insertions(+)
create mode 100644 drivers/misc/mic/cosm/cosm_scif_server.c
diff --git a/drivers/misc/mic/cosm/cosm_scif_server.c b/drivers/misc/mic/cosm/cosm_scif_server.c
new file mode 100644
index 0000000..62a73fb
--- /dev/null
+++ b/drivers/misc/mic/cosm/cosm_scif_server.c
@@ -0,0 +1,405 @@
+/*
+ * Intel MIC Platform Software Stack (MPSS)
+ *
+ * Copyright(c) 2015 Intel Corporation.
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License, version 2, as
+ * published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful, but
+ * WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+ * General Public License for more details.
+ *
+ * The full GNU General Public License is included in this distribution in
+ * the file called "COPYING".
+ *
+ * Intel MIC Coprocessor State Management (COSM) Driver
+ *
+ */
+#include <linux/kthread.h>
+#include "cosm_main.h"
+
+/*
+ * The COSM driver uses SCIF to communicate between the management node and the
+ * MIC cards. SCIF is used to (a) Send a shutdown command to the card (b)
+ * receive a shutdown status back from the card upon completion of shutdown and
+ * (c) receive periodic heartbeat messages from the card used to deduce if the
+ * card has crashed.
+ *
+ * A COSM server consisting of a SCIF listening endpoint waits for incoming
+ * connections from the card. Upon acceptance of the connection, a separate
+ * work-item is scheduled to handle SCIF message processing for that card. The
+ * life-time of this work-item is therefore the time from which the connection
+ * from a card is accepted to the time at which the connection is closed. A new
+ * work-item starts each time the card boots and is alive till the card (a)
+ * shuts down (b) is reset (c) crashes (d) cosm_client driver on the card is
+ * unloaded.
+ *
+ * From the point of view of COSM interactions with SCIF during card
+ * shutdown, reset and crash are as follows:
+ *
+ * Card shutdown
+ * -------------
+ * 1. COSM client on the card invokes orderly_poweroff() in response to SHUTDOWN
+ * message from the host.
+ * 2. Card driver shutdown callback invokes scif_unregister_device(..) resulting
+ * in scif_remove(..) getting called on the card
+ * 3. scif_remove -> scif_stop -> scif_handle_remove_node ->
+ * scif_peer_unregister_device -> device_unregister for the host peer device
+ * 4. During device_unregister remove(..) method of cosm_client is invoked which
+ * closes the COSM SCIF endpoint on the card. This results in a SCIF_DISCNCT
+ * message being sent to host SCIF. SCIF_DISCNCT message processing on the
+ * host SCIF sets the host COSM SCIF endpoint state to DISCONNECTED and wakes
+ * up the host COSM thread blocked in scif_poll(..) resulting in
+ * scif_poll(..) returning POLLHUP.
+ * 5. On the card, scif_peer_release_dev is next called which results in an
+ * SCIF_EXIT message being sent to the host and after receiving the
+ * SCIF_EXIT_ACK from the host the peer device teardown on the card is
+ * complete.
+ * 6. As part of the SCIF_EXIT message processing on the host, host sends a
+ * SCIF_REMOVE_NODE to itself corresponding to the card being removed. This
+ * starts a similar SCIF peer device teardown sequence on the host
+ * corresponding to the card being shut down.
+ *
+ * Card reset
+ * ----------
+ * The case of interest here is when the card has not been previously shut down
+ * since most of the steps below are skipped in that case:
+
+ * 1. cosm_stop(..) invokes hw_ops->stop(..) method of the base PCIe driver
+ * which unregisters the SCIF HW device resulting in scif_remove(..) being
+ * called on the host.
+ * 2. scif_remove(..) calls scif_disconnect_node(..) which results in a
+ * SCIF_EXIT message being sent to the card.
+ * 3. The card executes scif_stop() as part of SCIF_EXIT message
+ * processing. This results in the COSM endpoint on the card being closed and
+ * the SCIF host peer device on the card getting unregistered similar to
+ * steps 3, 4 and 5 for the card shutdown case above. scif_poll(..) on the
+ * host returns POLLHUP as a result.
+ * 4. On the host, card peer device unregister and SCIF HW remove(..) also
+ * subsequently complete.
+ *
+ * Card crash
+ * ----------
+ * If a reset is issued after the card has crashed, there is no SCIF_DISCNT
+ * message from the card which would result in scif_poll(..) returning
+ * POLLHUP. In this case when the host SCIF driver sends a SCIF_REMOVE_NODE
+ * message to itself resulting in the card SCIF peer device being unregistered,
+ * this results in a scif_peer_release_dev -> scif_cleanup_scifdev->
+ * scif_invalidate_ep call sequence which sets the endpoint state to
+ * DISCONNECTED and results in scif_poll(..) returning POLLHUP.
+ */
+
+#define COSM_SCIF_BACKLOG 16
+#define COSM_HEARTBEAT_CHECK_DELTA_SEC 10
+#define COSM_HEARTBEAT_TIMEOUT_SEC \
+ (COSM_HEARTBEAT_SEND_SEC + COSM_HEARTBEAT_CHECK_DELTA_SEC)
+#define COSM_HEARTBEAT_TIMEOUT_MSEC (COSM_HEARTBEAT_TIMEOUT_SEC * MSEC_PER_SEC)
+
+static struct task_struct *server_thread;
+static scif_epd_t listen_epd;
+
+/* Publish MIC card's shutdown status to user space MIC daemon */
+static void cosm_update_mic_status(struct cosm_device *cdev)
+{
+ if (MIC_NOP != cdev->shutdown_status_int) {
+ cosm_set_shutdown_status(cdev, cdev->shutdown_status_int);
+ cdev->shutdown_status_int = MIC_NOP;
+ }
+}
+
+/* Store MIC card's shutdown status internally when it is received */
+static void cosm_shutdown_status_int(struct cosm_device *cdev,
+ enum mic_status shutdown_status)
+{
+ switch (shutdown_status) {
+ case MIC_HALTED:
+ case MIC_POWER_OFF:
+ case MIC_RESTART:
+ case MIC_CRASHED:
+ break;
+ default:
+ dev_err(&cdev->dev, "%s %d Unexpected shutdown_status %d\n",
+ __func__, __LINE__, shutdown_status);
+ return;
+ };
+ cdev->shutdown_status_int = shutdown_status;
+ cdev->heartbeat_watchdog_enable = false;
+
+ if (MIC_SHUTTING_DOWN != cdev->state)
+ cosm_set_state(cdev, MIC_SHUTTING_DOWN);
+}
+
+/* Non-blocking recv. Read and process all available messages */
+static void cosm_scif_recv(struct cosm_device *cdev)
+{
+ struct cosm_msg msg;
+ int rc;
+
+ while (1) {
+ rc = scif_recv(cdev->epd, &msg, sizeof(msg), 0);
+ if (!rc) {
+ break;
+ } else if (rc < 0) {
+ dev_dbg(&cdev->dev, "%s: %d rc %d\n",
+ __func__, __LINE__, rc);
+ break;
+ }
+ dev_dbg(&cdev->dev, "%s: %d rc %d id 0x%llx\n",
+ __func__, __LINE__, rc, msg.id);
+
+ switch (msg.id) {
+ case COSM_MSG_SHUTDOWN_STATUS:
+ cosm_shutdown_status_int(cdev, msg.shutdown_status);
+ break;
+ case COSM_MSG_HEARTBEAT:
+ /* Nothing to do, heartbeat only unblocks scif_poll */
+ break;
+ default:
+ dev_err(&cdev->dev, "%s: %d unknown msg.id %lld\n",
+ __func__, __LINE__, msg.id);
+ break;
+ }
+ }
+}
+
+/* Publish crashed status for this MIC card */
+static void cosm_set_crashed(struct cosm_device *cdev)
+{
+ dev_err(&cdev->dev, "node alive timeout\n");
+ cosm_shutdown_status_int(cdev, MIC_CRASHED);
+ cosm_update_mic_status(cdev);
+}
+
+/* Send host time to the MIC card to sync system time between host and MIC */
+static void cosm_send_time(struct cosm_device *cdev)
+{
+ struct cosm_msg msg = { .id = COSM_MSG_SYNC_TIME };
+ int rc;
+
+ getnstimeofday64(&msg.timespec);
+ rc = scif_send(cdev->epd, &msg, sizeof(msg), SCIF_SEND_BLOCK);
+ if (rc < 0)
+ dev_err(&cdev->dev, "%s %d scif_send failed rc %d\n",
+ __func__, __LINE__, rc);
+}
+
+/*
+ * Close this cosm_device's endpoint after its peer endpoint on the card has
+ * been closed. In all cases except MIC card crash POLLHUP on the host is
+ * triggered by the client's endpoint being closed.
+ */
+static void cosm_scif_close(struct cosm_device *cdev)
+{
+ /*
+ * Because SHUTDOWN_STATUS message is sent by the MIC cards in the
+ * reboot notifier when shutdown is still not complete, we notify mpssd
+ * to reset the card when SCIF endpoint is closed.
+ */
+ cosm_update_mic_status(cdev);
+ scif_close(cdev->epd);
+ cdev->epd = NULL;
+ dev_dbg(&cdev->dev, "%s %d\n", __func__, __LINE__);
+}
+
+/*
+ * Set card state to ONLINE when a new SCIF connection from a MIC card is
+ * received. Normally the state is BOOTING when the connection comes in, but can
+ * be ONLINE if cosm_client driver on the card was unloaded and then reloaded.
+ */
+static int cosm_set_online(struct cosm_device *cdev)
+{
+ int rc = 0;
+
+ if (MIC_BOOTING == cdev->state || MIC_ONLINE == cdev->state) {
+ cdev->heartbeat_watchdog_enable = cdev->sysfs_heartbeat_enable;
+ cdev->epd = cdev->newepd;
+ if (MIC_BOOTING == cdev->state)
+ cosm_set_state(cdev, MIC_ONLINE);
+ cosm_send_time(cdev);
+ dev_dbg(&cdev->dev, "%s %d\n", __func__, __LINE__);
+ } else {
+ dev_warn(&cdev->dev, "%s %d not going online in state: %s\n",
+ __func__, __LINE__, cosm_state_string[cdev->state]);
+ rc = -EINVAL;
+ }
+ /* Drop reference acquired by bus_find_device in the server thread */
+ put_device(&cdev->dev);
+ return rc;
+}
+
+/*
+ * Work function for handling work for a SCIF connection from a particular MIC
+ * card. It first sets the card state to ONLINE and then calls scif_poll to
+ * block on activity such as incoming messages on the SCIF endpoint. When the
+ * endpoint is closed, the work function exits, completing its life cycle, from
+ * MIC card boot to card shutdown/reset/crash.
+ */
+void cosm_scif_work(struct work_struct *work)
+{
+ struct cosm_device *cdev = container_of(work, struct cosm_device,
+ scif_work);
+ struct scif_pollepd pollepd;
+ int rc;
+
+ mutex_lock(&cdev->cosm_mutex);
+ if (cosm_set_online(cdev))
+ goto exit;
+
+ while (1) {
+ pollepd.epd = cdev->epd;
+ pollepd.events = POLLIN;
+
+ /* Drop the mutex before blocking in scif_poll(..) */
+ mutex_unlock(&cdev->cosm_mutex);
+ /* poll(..) with timeout on our endpoint */
+ rc = scif_poll(&pollepd, 1, COSM_HEARTBEAT_TIMEOUT_MSEC);
+ mutex_lock(&cdev->cosm_mutex);
+ if (rc < 0) {
+ dev_err(&cdev->dev, "%s %d scif_poll rc %d\n",
+ __func__, __LINE__, rc);
+ continue;
+ }
+
+ /* There is a message from the card */
+ if (pollepd.revents & POLLIN)
+ cosm_scif_recv(cdev);
+
+ /* The peer endpoint is closed or this endpoint disconnected */
+ if (pollepd.revents & POLLHUP) {
+ cosm_scif_close(cdev);
+ break;
+ }
+
+ /* Did we timeout from poll? */
+ if (!rc && cdev->heartbeat_watchdog_enable)
+ cosm_set_crashed(cdev);
+ }
+exit:
+ dev_dbg(&cdev->dev, "%s %d exiting\n", __func__, __LINE__);
+ mutex_unlock(&cdev->cosm_mutex);
+}
+
+/*
+ * COSM SCIF server thread function. Accepts incoming SCIF connections from MIC
+ * cards, finds the correct cosm_device to associate that connection with and
+ * schedules individual work items for each MIC card.
+ */
+static int cosm_scif_server(void *unused)
+{
+ struct cosm_device *cdev;
+ scif_epd_t newepd;
+ struct scif_port_id port_id;
+ int rc;
+
+ allow_signal(SIGKILL);
+
+ while (!kthread_should_stop()) {
+ rc = scif_accept(listen_epd, &port_id, &newepd,
+ SCIF_ACCEPT_SYNC);
+ if (rc < 0) {
+ if (-ERESTARTSYS != rc)
+ pr_err("%s %d rc %d\n", __func__, __LINE__, rc);
+ continue;
+ }
+
+ /*
+ * Associate the incoming connection with a particular
+ * cosm_device, COSM device ID == SCIF node ID - 1
+ */
+ cdev = cosm_find_cdev_by_id(port_id.node - 1);
+ if (!cdev)
+ continue;
+ cdev->newepd = newepd;
+ schedule_work(&cdev->scif_work);
+ }
+
+ pr_debug("%s %d Server thread stopped\n", __func__, __LINE__);
+ return 0;
+}
+
+static int cosm_scif_listen(void)
+{
+ int rc;
+
+ listen_epd = scif_open();
+ if (!listen_epd) {
+ pr_err("%s %d scif_open failed\n", __func__, __LINE__);
+ return -ENOMEM;
+ }
+
+ rc = scif_bind(listen_epd, SCIF_COSM_LISTEN_PORT);
+ if (rc < 0) {
+ pr_err("%s %d scif_bind failed rc %d\n",
+ __func__, __LINE__, rc);
+ goto err;
+ }
+
+ rc = scif_listen(listen_epd, COSM_SCIF_BACKLOG);
+ if (rc < 0) {
+ pr_err("%s %d scif_listen rc %d\n", __func__, __LINE__, rc);
+ goto err;
+ }
+ pr_debug("%s %d listen_epd set up\n", __func__, __LINE__);
+ return 0;
+err:
+ scif_close(listen_epd);
+ listen_epd = NULL;
+ return rc;
+}
+
+static void cosm_scif_listen_exit(void)
+{
+ pr_debug("%s %d closing listen_epd\n", __func__, __LINE__);
+ if (listen_epd) {
+ scif_close(listen_epd);
+ listen_epd = NULL;
+ }
+}
+
+/*
+ * Create a listening SCIF endpoint and a server kthread which accepts incoming
+ * SCIF connections from MIC cards
+ */
+int cosm_scif_init(void)
+{
+ int rc = cosm_scif_listen();
+
+ if (rc) {
+ pr_err("%s %d cosm_scif_listen rc %d\n",
+ __func__, __LINE__, rc);
+ goto err;
+ }
+
+ server_thread = kthread_run(cosm_scif_server, NULL, "cosm_server");
+ if (IS_ERR(server_thread)) {
+ rc = PTR_ERR(server_thread);
+ pr_err("%s %d kthread_run rc %d\n", __func__, __LINE__, rc);
+ goto listen_exit;
+ }
+ return 0;
+listen_exit:
+ cosm_scif_listen_exit();
+err:
+ return rc;
+}
+
+/* Stop the running server thread and close the listening SCIF endpoint */
+void cosm_scif_exit(void)
+{
+ int rc;
+
+ if (!IS_ERR_OR_NULL(server_thread)) {
+ rc = send_sig(SIGKILL, server_thread, 0);
+ if (rc) {
+ pr_err("%s %d send_sig rc %d\n",
+ __func__, __LINE__, rc);
+ return;
+ }
+ kthread_stop(server_thread);
+ }
+
+ cosm_scif_listen_exit();
+}
--
2.0.0.rc3.2.g998f840
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
[toc] | [prev] | [next] | [standalone]
| From | Ashutosh Dixit <ashutosh.dixit@intel.com> |
|---|---|
| Date | 2015-07-28 01:40 +0200 |
| Subject | [PATCH char-misc-next 03/19] misc: mic: MIC COSM bus |
| Message-ID | <pR0a2-8vE-97@gated-at.bofh.it> |
| In reply to | #1193568 |
The MIC COSM bus allows the co-processor state management (COSM)
functionality to be shared between multiple generations of Intel MIC
products. The COSM driver registers itself on the COSM bus. The base
PCIe drivers implement the bus ops and register COSM devices on the
bus, resulting in the COSM driver being probed with the COSM devices.
COSM bus ops, e.g. start, stop, ready, reset, therefore abstract out
common functionality from its specific implementation for individual
generations of MIC products.
Reviewed-by: Nikhil Rao <nikhil.rao@intel.com>
Reviewed-by: Sudeep Dutt <sudeep.dutt@intel.com>
Signed-off-by: Ashutosh Dixit <ashutosh.dixit@intel.com>
---
drivers/misc/mic/Kconfig | 22 ++++++-
drivers/misc/mic/bus/Makefile | 1 +
drivers/misc/mic/bus/cosm_bus.c | 141 ++++++++++++++++++++++++++++++++++++++++
drivers/misc/mic/bus/cosm_bus.h | 134 ++++++++++++++++++++++++++++++++++++++
4 files changed, 296 insertions(+), 2 deletions(-)
create mode 100644 drivers/misc/mic/bus/cosm_bus.c
create mode 100644 drivers/misc/mic/bus/cosm_bus.h
diff --git a/drivers/misc/mic/Kconfig b/drivers/misc/mic/Kconfig
index e9f2f56..1488232 100644
--- a/drivers/misc/mic/Kconfig
+++ b/drivers/misc/mic/Kconfig
@@ -36,7 +36,7 @@ comment "Intel MIC Host Driver"
config INTEL_MIC_HOST
tristate "Intel MIC Host Driver"
- depends on 64BIT && PCI && X86 && INTEL_MIC_BUS && SCIF_BUS
+ depends on 64BIT && PCI && X86 && INTEL_MIC_BUS && SCIF_BUS && MIC_COSM
select VHOST_RING
help
This enables Host Driver support for the Intel Many Integrated
@@ -56,7 +56,7 @@ comment "Intel MIC Card Driver"
config INTEL_MIC_CARD
tristate "Intel MIC Card Driver"
- depends on 64BIT && X86 && INTEL_MIC_BUS && SCIF_BUS
+ depends on 64BIT && X86 && INTEL_MIC_BUS && SCIF_BUS && MIC_COSM
select VIRTIO
help
This enables card driver support for the Intel Many Integrated
@@ -88,3 +88,21 @@ config SCIF
More information about the Intel MIC family as well as the Linux
OS and tools for MIC to use with this driver are available from
<http://software.intel.com/en-us/mic-developer>.
+
+comment "Intel MIC Coprocessor State Management (COSM) Drivers"
+
+config MIC_COSM
+ tristate "Intel MIC Coprocessor State Management (COSM) Drivers"
+ depends on 64BIT && PCI && X86 && SCIF
+ help
+ This enables COSM driver support for the Intel Many
+ Integrated Core (MIC) family of PCIe form factor coprocessor
+ devices. COSM drivers implement functions such as boot,
+ shutdown, reset and reboot of MIC devices.
+
+ If you are building a host kernel with an Intel MIC device then
+ say M (recommended) or Y, else say N. If unsure say N.
+
+ More information about the Intel MIC family as well as the Linux
+ OS and tools for MIC to use with this driver are available from
+ <http://software.intel.com/en-us/mic-developer>.
diff --git a/drivers/misc/mic/bus/Makefile b/drivers/misc/mic/bus/Makefile
index 1ed37e2..761842b 100644
--- a/drivers/misc/mic/bus/Makefile
+++ b/drivers/misc/mic/bus/Makefile
@@ -4,3 +4,4 @@
#
obj-$(CONFIG_INTEL_MIC_BUS) += mic_bus.o
obj-$(CONFIG_SCIF_BUS) += scif_bus.o
+obj-$(CONFIG_MIC_COSM) += cosm_bus.o
diff --git a/drivers/misc/mic/bus/cosm_bus.c b/drivers/misc/mic/bus/cosm_bus.c
new file mode 100644
index 0000000..1e36830
--- /dev/null
+++ b/drivers/misc/mic/bus/cosm_bus.c
@@ -0,0 +1,141 @@
+/*
+ * Intel MIC Platform Software Stack (MPSS)
+ *
+ * Copyright(c) 2015 Intel Corporation.
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License, version 2, as
+ * published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful, but
+ * WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+ * General Public License for more details.
+ *
+ * The full GNU General Public License is included in this distribution in
+ * the file called "COPYING".
+ *
+ * Intel MIC COSM Bus Driver
+ */
+#include <linux/slab.h>
+#include <linux/module.h>
+#include <linux/idr.h>
+#include "cosm_bus.h"
+
+/* Unique numbering for cosm devices. */
+static DEFINE_IDA(cosm_index_ida);
+
+static int cosm_dev_probe(struct device *d)
+{
+ struct cosm_device *dev = dev_to_cosm(d);
+ struct cosm_driver *drv = drv_to_cosm(dev->dev.driver);
+
+ return drv->probe(dev);
+}
+
+static int cosm_dev_remove(struct device *d)
+{
+ struct cosm_device *dev = dev_to_cosm(d);
+ struct cosm_driver *drv = drv_to_cosm(dev->dev.driver);
+
+ drv->remove(dev);
+ return 0;
+}
+
+static struct bus_type cosm_bus = {
+ .name = "cosm_bus",
+ .probe = cosm_dev_probe,
+ .remove = cosm_dev_remove,
+};
+
+int cosm_register_driver(struct cosm_driver *driver)
+{
+ driver->driver.bus = &cosm_bus;
+ return driver_register(&driver->driver);
+}
+EXPORT_SYMBOL_GPL(cosm_register_driver);
+
+void cosm_unregister_driver(struct cosm_driver *driver)
+{
+ driver_unregister(&driver->driver);
+}
+EXPORT_SYMBOL_GPL(cosm_unregister_driver);
+
+static inline void cosm_release_dev(struct device *d)
+{
+ struct cosm_device *cdev = dev_to_cosm(d);
+
+ kfree(cdev);
+}
+
+struct cosm_device *
+cosm_register_device(struct device *pdev, struct cosm_hw_ops *hw_ops)
+{
+ struct cosm_device *cdev;
+ int ret;
+
+ cdev = kzalloc(sizeof(*cdev), GFP_KERNEL);
+ if (!cdev)
+ return ERR_PTR(-ENOMEM);
+
+ cdev->dev.parent = pdev;
+ cdev->dev.release = cosm_release_dev;
+ cdev->hw_ops = hw_ops;
+ dev_set_drvdata(&cdev->dev, cdev);
+ cdev->dev.bus = &cosm_bus;
+
+ /* Assign a unique device index and hence name */
+ ret = ida_simple_get(&cosm_index_ida, 0, 0, GFP_KERNEL);
+ if (ret < 0)
+ goto free_cdev;
+
+ cdev->index = ret;
+ cdev->dev.id = ret;
+ dev_set_name(&cdev->dev, "cosm-dev%u", cdev->index);
+
+ ret = device_register(&cdev->dev);
+ if (ret)
+ goto ida_remove;
+ return cdev;
+ida_remove:
+ ida_simple_remove(&cosm_index_ida, cdev->index);
+free_cdev:
+ kfree(cdev);
+ return ERR_PTR(ret);
+}
+EXPORT_SYMBOL_GPL(cosm_register_device);
+
+void cosm_unregister_device(struct cosm_device *dev)
+{
+ int index = dev->index; /* save for after device release */
+
+ device_unregister(&dev->dev);
+ ida_simple_remove(&cosm_index_ida, index);
+}
+EXPORT_SYMBOL_GPL(cosm_unregister_device);
+
+struct cosm_device *cosm_find_cdev_by_id(int id)
+{
+ struct device *dev = subsys_find_device_by_id(&cosm_bus, id, NULL);
+
+ return dev ? container_of(dev, struct cosm_device, dev) : NULL;
+}
+EXPORT_SYMBOL_GPL(cosm_find_cdev_by_id);
+
+static int __init cosm_init(void)
+{
+ return bus_register(&cosm_bus);
+}
+
+static void __exit cosm_exit(void)
+{
+ bus_unregister(&cosm_bus);
+ ida_destroy(&cosm_index_ida);
+}
+
+core_initcall(cosm_init);
+module_exit(cosm_exit);
+
+MODULE_AUTHOR("Intel Corporation");
+MODULE_DESCRIPTION("Intel(R) MIC card OS state management bus driver");
+MODULE_LICENSE("GPL v2");
diff --git a/drivers/misc/mic/bus/cosm_bus.h b/drivers/misc/mic/bus/cosm_bus.h
new file mode 100644
index 0000000..f7c57f2
--- /dev/null
+++ b/drivers/misc/mic/bus/cosm_bus.h
@@ -0,0 +1,134 @@
+/*
+ * Intel MIC Platform Software Stack (MPSS)
+ *
+ * Copyright(c) 2015 Intel Corporation.
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License, version 2, as
+ * published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful, but
+ * WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+ * General Public License for more details.
+ *
+ * The full GNU General Public License is included in this distribution in
+ * the file called "COPYING".
+ *
+ * Intel MIC COSM Bus Driver
+ */
+#ifndef _COSM_BUS_H_
+#define _COSM_BUS_H_
+
+#include <linux/scif.h>
+#include <linux/mic_common.h>
+#include "../common/mic_dev.h"
+
+/**
+ * cosm_device - representation of a cosm device
+ *
+ * @attr_group: Pointer to list of sysfs attribute groups.
+ * @sdev: Device for sysfs entries.
+ * @state: MIC state.
+ * @shutdown_status: MIC status reported by card for shutdown/crashes.
+ * @shutdown_status_int: Internal shutdown status maintained by the driver
+ * @cosm_mutex: Mutex for synchronizing access to data structures.
+ * @reset_trigger_work: Work for triggering reset requests.
+ * @scif_work: Work for handling per device SCIF connections
+ * @cmdline: Kernel command line.
+ * @firmware: Firmware file name.
+ * @ramdisk: Ramdisk file name.
+ * @bootmode: Boot mode i.e. "linux" or "elf" for flash updates.
+ * @log_buf_addr: Log buffer address for MIC.
+ * @log_buf_len: Log buffer length address for MIC.
+ * @state_sysfs: Sysfs dirent for notifying ring 3 about MIC state changes.
+ * @hw_ops: the hardware bus ops for this device.
+ * @dev: underlying device.
+ * @index: unique position on the cosm bus
+ * @dbg_dir: debug fs directory
+ * @newepd: new endpoint from scif accept to be assigned to this cdev
+ * @epd: SCIF endpoint for this cdev
+ * @heartbeat_watchdog_enable: if heartbeat watchdog is enabled for this cdev
+ * @sysfs_heartbeat_enable: sysfs setting for disabling heartbeat notification
+ */
+struct cosm_device {
+ const struct attribute_group **attr_group;
+ struct device *sdev;
+ u8 state;
+ u8 shutdown_status;
+ u8 shutdown_status_int;
+ struct mutex cosm_mutex;
+ struct work_struct reset_trigger_work;
+ struct work_struct scif_work;
+ char *cmdline;
+ char *firmware;
+ char *ramdisk;
+ char *bootmode;
+ void *log_buf_addr;
+ int *log_buf_len;
+ struct kernfs_node *state_sysfs;
+ struct cosm_hw_ops *hw_ops;
+ struct device dev;
+ int index;
+ struct dentry *dbg_dir;
+ scif_epd_t newepd;
+ scif_epd_t epd;
+ bool heartbeat_watchdog_enable;
+ bool sysfs_heartbeat_enable;
+};
+
+/**
+ * cosm_driver - operations for a cosm driver
+ *
+ * @driver: underlying device driver (populate name and owner).
+ * @probe: the function to call when a device is found. Returns 0 or -errno.
+ * @remove: the function to call when a device is removed.
+ */
+struct cosm_driver {
+ struct device_driver driver;
+ int (*probe)(struct cosm_device *dev);
+ void (*remove)(struct cosm_device *dev);
+};
+
+/**
+ * cosm_hw_ops - cosm bus ops
+ *
+ * @reset: trigger MIC reset
+ * @force_reset: force MIC reset
+ * @post_reset: inform MIC reset is complete
+ * @ready: is MIC ready for OS download
+ * @start: boot MIC
+ * @stop: prepare MIC for reset
+ * @family: return MIC HW family string
+ * @stepping: return MIC HW stepping string
+ * @aper: return MIC PCIe aperture
+ */
+struct cosm_hw_ops {
+ void (*reset)(struct cosm_device *cdev);
+ void (*force_reset)(struct cosm_device *cdev);
+ void (*post_reset)(struct cosm_device *cdev, enum mic_states state);
+ bool (*ready)(struct cosm_device *cdev);
+ int (*start)(struct cosm_device *cdev, int id);
+ void (*stop)(struct cosm_device *cdev, bool force);
+ ssize_t (*family)(struct cosm_device *cdev, char *buf);
+ ssize_t (*stepping)(struct cosm_device *cdev, char *buf);
+ struct mic_mw *(*aper)(struct cosm_device *cdev);
+};
+
+struct cosm_device *
+cosm_register_device(struct device *pdev, struct cosm_hw_ops *hw_ops);
+void cosm_unregister_device(struct cosm_device *dev);
+int cosm_register_driver(struct cosm_driver *drv);
+void cosm_unregister_driver(struct cosm_driver *drv);
+struct cosm_device *cosm_find_cdev_by_id(int id);
+
+static inline struct cosm_device *dev_to_cosm(struct device *dev)
+{
+ return container_of(dev, struct cosm_device, dev);
+}
+
+static inline struct cosm_driver *drv_to_cosm(struct device_driver *drv)
+{
+ return container_of(drv, struct cosm_driver, driver);
+}
+#endif /* _COSM_BUS_H */
--
2.0.0.rc3.2.g998f840
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
[toc] | [prev] | [next] | [standalone]
| From | Ashutosh Dixit <ashutosh.dixit@intel.com> |
|---|---|
| Date | 2015-07-28 01:50 +0200 |
| Subject | [PATCH char-misc-next 08/19] misc: mic: Remove COSM functionality from the MIC card driver |
| Message-ID | <pR0jD-fw-3@gated-at.bofh.it> |
| In reply to | #1193568 |
Since card side COSM functionality, to trigger MIC device shutdowns
and communicate shutdown status to the host, is now moved into a
separate COSM client driver, this patch removes this functionality
from the base MIC card driver. The mic_bus driver is also updated to
use the device index provided by COSM rather than maintain its own
device index.
Reviewed-by: Nikhil Rao <nikhil.rao@intel.com>
Reviewed-by: Sudeep Dutt <sudeep.dutt@intel.com>
Signed-off-by: Ashutosh Dixit <ashutosh.dixit@intel.com>
---
drivers/misc/mic/bus/mic_bus.c | 22 ++--------
drivers/misc/mic/card/mic_device.c | 88 ++------------------------------------
drivers/misc/mic/card/mic_x100.c | 2 +-
drivers/misc/mic/common/mic_dev.h | 13 ++++++
include/linux/mic_bus.h | 3 +-
5 files changed, 23 insertions(+), 105 deletions(-)
diff --git a/drivers/misc/mic/bus/mic_bus.c b/drivers/misc/mic/bus/mic_bus.c
index 961ae90..c64955d 100644
--- a/drivers/misc/mic/bus/mic_bus.c
+++ b/drivers/misc/mic/bus/mic_bus.c
@@ -25,9 +25,6 @@
#include <linux/idr.h>
#include <linux/mic_bus.h>
-/* Unique numbering for mbus devices. */
-static DEFINE_IDA(mbus_index_ida);
-
static ssize_t device_show(struct device *d,
struct device_attribute *attr, char *buf)
{
@@ -147,7 +144,8 @@ static void mbus_release_dev(struct device *d)
struct mbus_device *
mbus_register_device(struct device *pdev, int id, struct dma_map_ops *dma_ops,
- struct mbus_hw_ops *hw_ops, void __iomem *mmio_va)
+ struct mbus_hw_ops *hw_ops, int index,
+ void __iomem *mmio_va)
{
int ret;
struct mbus_device *mbdev;
@@ -166,13 +164,7 @@ mbus_register_device(struct device *pdev, int id, struct dma_map_ops *dma_ops,
mbdev->dev.release = mbus_release_dev;
mbdev->hw_ops = hw_ops;
mbdev->dev.bus = &mic_bus;
-
- /* Assign a unique device index and hence name. */
- ret = ida_simple_get(&mbus_index_ida, 0, 0, GFP_KERNEL);
- if (ret < 0)
- goto free_mbdev;
-
- mbdev->index = ret;
+ mbdev->index = index;
dev_set_name(&mbdev->dev, "mbus-dev%u", mbdev->index);
/*
* device_register() causes the bus infrastructure to look for a
@@ -180,10 +172,8 @@ mbus_register_device(struct device *pdev, int id, struct dma_map_ops *dma_ops,
*/
ret = device_register(&mbdev->dev);
if (ret)
- goto ida_remove;
+ goto free_mbdev;
return mbdev;
-ida_remove:
- ida_simple_remove(&mbus_index_ida, mbdev->index);
free_mbdev:
kfree(mbdev);
return ERR_PTR(ret);
@@ -192,10 +182,7 @@ EXPORT_SYMBOL_GPL(mbus_register_device);
void mbus_unregister_device(struct mbus_device *mbdev)
{
- int index = mbdev->index; /* save for after device release */
-
device_unregister(&mbdev->dev);
- ida_simple_remove(&mbus_index_ida, index);
}
EXPORT_SYMBOL_GPL(mbus_unregister_device);
@@ -207,7 +194,6 @@ static int __init mbus_init(void)
static void __exit mbus_exit(void)
{
bus_unregister(&mic_bus);
- ida_destroy(&mbus_index_ida);
}
core_initcall(mbus_init);
diff --git a/drivers/misc/mic/card/mic_device.c b/drivers/misc/mic/card/mic_device.c
index 6338908..d0edaf7 100644
--- a/drivers/misc/mic/card/mic_device.c
+++ b/drivers/misc/mic/card/mic_device.c
@@ -37,71 +37,6 @@
#include "mic_virtio.h"
static struct mic_driver *g_drv;
-static struct mic_irq *shutdown_cookie;
-
-static void mic_notify_host(u8 state)
-{
- struct mic_driver *mdrv = g_drv;
- struct mic_bootparam __iomem *bootparam = mdrv->dp;
-
- iowrite8(state, &bootparam->shutdown_status);
- dev_dbg(mdrv->dev, "%s %d system_state %d\n",
- __func__, __LINE__, state);
- mic_send_intr(&mdrv->mdev, ioread8(&bootparam->c2h_shutdown_db));
-}
-
-static int mic_panic_event(struct notifier_block *this, unsigned long event,
- void *ptr)
-{
- struct mic_driver *mdrv = g_drv;
- struct mic_bootparam __iomem *bootparam = mdrv->dp;
-
- iowrite8(-1, &bootparam->h2c_config_db);
- iowrite8(-1, &bootparam->h2c_shutdown_db);
- mic_notify_host(MIC_CRASHED);
- return NOTIFY_DONE;
-}
-
-static struct notifier_block mic_panic = {
- .notifier_call = mic_panic_event,
-};
-
-static irqreturn_t mic_shutdown_isr(int irq, void *data)
-{
- struct mic_driver *mdrv = g_drv;
- struct mic_bootparam __iomem *bootparam = mdrv->dp;
-
- mic_ack_interrupt(&g_drv->mdev);
- if (ioread8(&bootparam->shutdown_card))
- orderly_poweroff(true);
- return IRQ_HANDLED;
-}
-
-static int mic_shutdown_init(void)
-{
- int rc = 0;
- struct mic_driver *mdrv = g_drv;
- struct mic_bootparam __iomem *bootparam = mdrv->dp;
- int shutdown_db;
-
- shutdown_db = mic_next_card_db();
- shutdown_cookie = mic_request_card_irq(mic_shutdown_isr, NULL,
- "Shutdown", mdrv, shutdown_db);
- if (IS_ERR(shutdown_cookie))
- rc = PTR_ERR(shutdown_cookie);
- else
- iowrite8(shutdown_db, &bootparam->h2c_shutdown_db);
- return rc;
-}
-
-static void mic_shutdown_uninit(void)
-{
- struct mic_driver *mdrv = g_drv;
- struct mic_bootparam __iomem *bootparam = mdrv->dp;
-
- iowrite8(-1, &bootparam->h2c_shutdown_db);
- mic_free_card_irq(shutdown_cookie, mdrv);
-}
static int __init mic_dp_init(void)
{
@@ -359,11 +294,7 @@ int __init mic_driver_init(struct mic_driver *mdrv)
u8 node_id;
g_drv = mdrv;
- /*
- * Unloading the card module is not supported. The MIC card module
- * handles fundamental operations like host/card initiated shutdowns
- * and informing the host about card crashes and cannot be unloaded.
- */
+ /* Unloading the card module is not supported. */
if (!try_module_get(mdrv->dev->driver->owner)) {
rc = -ENODEV;
goto done;
@@ -374,12 +305,9 @@ int __init mic_driver_init(struct mic_driver *mdrv)
rc = mic_init_irq();
if (rc)
goto dp_uninit;
- rc = mic_shutdown_init();
- if (rc)
- goto irq_uninit;
if (!mic_request_dma_chans(mdrv)) {
rc = -ENODEV;
- goto shutdown_uninit;
+ goto irq_uninit;
}
rc = mic_devices_init(mdrv);
if (rc)
@@ -390,21 +318,18 @@ int __init mic_driver_init(struct mic_driver *mdrv)
NULL, &scif_hw_ops,
0, node_id, &mdrv->mdev.mmio, NULL,
NULL, mdrv->dp, mdrv->dma_ch,
- mdrv->num_dma_ch);
+ mdrv->num_dma_ch, true);
if (IS_ERR(mdrv->scdev)) {
rc = PTR_ERR(mdrv->scdev);
goto device_uninit;
}
mic_create_card_debug_dir(mdrv);
- atomic_notifier_chain_register(&panic_notifier_list, &mic_panic);
done:
return rc;
device_uninit:
mic_devices_uninit(mdrv);
dma_free:
mic_free_dma_chans(mdrv);
-shutdown_uninit:
- mic_shutdown_uninit();
irq_uninit:
mic_uninit_irq();
dp_uninit:
@@ -425,13 +350,6 @@ void mic_driver_uninit(struct mic_driver *mdrv)
scif_unregister_device(mdrv->scdev);
mic_devices_uninit(mdrv);
mic_free_dma_chans(mdrv);
- /*
- * Inform the host about the shutdown status i.e. poweroff/restart etc.
- * The module cannot be unloaded so the only code path to call
- * mic_devices_uninit(..) is the shutdown callback.
- */
- mic_notify_host(system_state);
- mic_shutdown_uninit();
mic_uninit_irq();
mic_dp_uninit();
module_put(mdrv->dev->driver->owner);
diff --git a/drivers/misc/mic/card/mic_x100.c b/drivers/misc/mic/card/mic_x100.c
index 77fd41781..b2958ce 100644
--- a/drivers/misc/mic/card/mic_x100.c
+++ b/drivers/misc/mic/card/mic_x100.c
@@ -261,7 +261,7 @@ static int __init mic_probe(struct platform_device *pdev)
mic_hw_intr_init(mdrv);
platform_set_drvdata(pdev, mdrv);
mdrv->dma_mbdev = mbus_register_device(mdrv->dev, MBUS_DEV_DMA_MIC,
- NULL, &mbus_hw_ops,
+ NULL, &mbus_hw_ops, 0,
mdrv->mdev.mmio.va);
if (IS_ERR(mdrv->dma_mbdev)) {
rc = PTR_ERR(mdrv->dma_mbdev);
diff --git a/drivers/misc/mic/common/mic_dev.h b/drivers/misc/mic/common/mic_dev.h
index 0b58c46..5077677 100644
--- a/drivers/misc/mic/common/mic_dev.h
+++ b/drivers/misc/mic/common/mic_dev.h
@@ -21,6 +21,19 @@
#ifndef __MIC_DEV_H__
#define __MIC_DEV_H__
+/* The maximum number of MIC devices supported in a single host system. */
+#define MIC_MAX_NUM_DEVS 128
+
+/**
+ * enum mic_hw_family - The hardware family to which a device belongs.
+ */
+enum mic_hw_family {
+ MIC_FAMILY_X100 = 0,
+ MIC_FAMILY_X200,
+ MIC_FAMILY_UNKNOWN,
+ MIC_FAMILY_LAST
+};
+
/**
* struct mic_mw - MIC memory window
*
diff --git a/include/linux/mic_bus.h b/include/linux/mic_bus.h
index d5b5f76..27d7c95 100644
--- a/include/linux/mic_bus.h
+++ b/include/linux/mic_bus.h
@@ -91,7 +91,8 @@ struct mbus_hw_ops {
struct mbus_device *
mbus_register_device(struct device *pdev, int id, struct dma_map_ops *dma_ops,
- struct mbus_hw_ops *hw_ops, void __iomem *mmio_va);
+ struct mbus_hw_ops *hw_ops, int index,
+ void __iomem *mmio_va);
void mbus_unregister_device(struct mbus_device *mbdev);
int mbus_register_driver(struct mbus_driver *drv);
--
2.0.0.rc3.2.g998f840
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
[toc] | [prev] | [next] | [standalone]
| From | Ashutosh Dixit <ashutosh.dixit@intel.com> |
|---|---|
| Date | 2015-07-28 02:00 +0200 |
| Subject | [PATCH char-misc-next 06/19] misc: mic: COSM client driver |
| Message-ID | <pR0tk-r2-19@gated-at.bofh.it> |
| In reply to | #1193568 |
The COSM client driver running on the MIC cards is implemented as a
kernel mode SCIF client. It responds to a "shutdown" message from the
host by triggering a card shutdown and also communicates the shutdown
or reboot status back the host. It is also responsible for syncing the
card time to that of the host. Because SCIF messaging cannot be used
in a panic context, the COSM client driver also periodically sends a
heartbeat SCIF message to the host thereby enabling the host to detect
card crashes.
Reviewed-by: Nikhil Rao <nikhil.rao@intel.com>
Reviewed-by: Sudeep Dutt <sudeep.dutt@intel.com>
Signed-off-by: Ashutosh Dixit <ashutosh.dixit@intel.com>
---
drivers/misc/mic/cosm_client/Makefile | 7 +
drivers/misc/mic/cosm_client/cosm_scif_client.c | 275 ++++++++++++++++++++++++
2 files changed, 282 insertions(+)
create mode 100644 drivers/misc/mic/cosm_client/Makefile
create mode 100644 drivers/misc/mic/cosm_client/cosm_scif_client.c
diff --git a/drivers/misc/mic/cosm_client/Makefile b/drivers/misc/mic/cosm_client/Makefile
new file mode 100644
index 0000000..6f751a5
--- /dev/null
+++ b/drivers/misc/mic/cosm_client/Makefile
@@ -0,0 +1,7 @@
+#
+# Makefile - Intel MIC COSM Client Driver
+# Copyright(c) 2015, Intel Corporation.
+#
+obj-$(CONFIG_MIC_COSM) += cosm_client.o
+
+cosm_client-objs += cosm_scif_client.o
diff --git a/drivers/misc/mic/cosm_client/cosm_scif_client.c b/drivers/misc/mic/cosm_client/cosm_scif_client.c
new file mode 100644
index 0000000..03e98bf
--- /dev/null
+++ b/drivers/misc/mic/cosm_client/cosm_scif_client.c
@@ -0,0 +1,275 @@
+/*
+ * Intel MIC Platform Software Stack (MPSS)
+ *
+ * Copyright(c) 2015 Intel Corporation.
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License, version 2, as
+ * published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful, but
+ * WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+ * General Public License for more details.
+ *
+ * The full GNU General Public License is included in this distribution in
+ * the file called "COPYING".
+ *
+ * Intel MIC COSM Client Driver
+ *
+ */
+#include <linux/module.h>
+#include <linux/delay.h>
+#include <linux/reboot.h>
+#include <linux/kthread.h>
+#include "../cosm/cosm_main.h"
+
+#define COSM_SCIF_MAX_RETRIES 10
+#define COSM_HEARTBEAT_SEND_MSEC (COSM_HEARTBEAT_SEND_SEC * MSEC_PER_SEC)
+
+static struct task_struct *client_thread;
+static scif_epd_t client_epd;
+static struct scif_peer_dev *client_spdev;
+
+/*
+ * Reboot notifier: receives shutdown status from the OS and communicates it
+ * back to the COSM process on the host
+ */
+static int cosm_reboot_event(struct notifier_block *this, unsigned long event,
+ void *ptr)
+{
+ struct cosm_msg msg = { .id = COSM_MSG_SHUTDOWN_STATUS };
+ int rc;
+
+ event = (event == SYS_RESTART) ? SYSTEM_RESTART : event;
+ dev_info(&client_spdev->dev, "%s %d received event %ld\n",
+ __func__, __LINE__, event);
+
+ msg.shutdown_status = event;
+ rc = scif_send(client_epd, &msg, sizeof(msg), SCIF_SEND_BLOCK);
+ if (rc < 0)
+ dev_err(&client_spdev->dev, "%s %d scif_send rc %d\n",
+ __func__, __LINE__, rc);
+
+ return NOTIFY_DONE;
+}
+
+static struct notifier_block cosm_reboot = {
+ .notifier_call = cosm_reboot_event,
+};
+
+/* Set system time from timespec value received from the host */
+static void cosm_set_time(struct cosm_msg *msg)
+{
+ int rc = do_settimeofday64(&msg->timespec);
+
+ if (rc)
+ dev_err(&client_spdev->dev, "%s: %d settimeofday rc %d\n",
+ __func__, __LINE__, rc);
+}
+
+/* COSM client receive message processing */
+static void cosm_client_recv(void)
+{
+ struct cosm_msg msg;
+ int rc;
+
+ while (1) {
+ rc = scif_recv(client_epd, &msg, sizeof(msg), 0);
+ if (!rc) {
+ return;
+ } else if (rc < 0) {
+ dev_err(&client_spdev->dev, "%s: %d rc %d\n",
+ __func__, __LINE__, rc);
+ return;
+ }
+
+ dev_dbg(&client_spdev->dev, "%s: %d rc %d id 0x%llx\n",
+ __func__, __LINE__, rc, msg.id);
+
+ switch (msg.id) {
+ case COSM_MSG_SYNC_TIME:
+ cosm_set_time(&msg);
+ break;
+ case COSM_MSG_SHUTDOWN:
+ orderly_poweroff(true);
+ break;
+ default:
+ dev_err(&client_spdev->dev, "%s: %d unknown id %lld\n",
+ __func__, __LINE__, msg.id);
+ break;
+ }
+ }
+}
+
+/* Initiate connection to the COSM server on the host */
+static int cosm_scif_connect(void)
+{
+ struct scif_port_id port_id;
+ int i, rc;
+
+ client_epd = scif_open();
+ if (!client_epd) {
+ dev_err(&client_spdev->dev, "%s %d scif_open failed\n",
+ __func__, __LINE__);
+ return -ENOMEM;
+ }
+
+ port_id.node = 0;
+ port_id.port = SCIF_COSM_LISTEN_PORT;
+
+ for (i = 0; i < COSM_SCIF_MAX_RETRIES; i++) {
+ rc = scif_connect(client_epd, &port_id);
+ if (rc < 0)
+ msleep(1000);
+ else
+ break;
+ }
+
+ if (rc < 0) {
+ dev_err(&client_spdev->dev, "%s %d scif_connect rc %d\n",
+ __func__, __LINE__, rc);
+ scif_close(client_epd);
+ client_epd = NULL;
+ }
+ return rc < 0 ? rc : 0;
+}
+
+/* Close host SCIF connection */
+static void cosm_scif_connect_exit(void)
+{
+ if (client_epd) {
+ scif_close(client_epd);
+ client_epd = NULL;
+ }
+}
+
+/*
+ * COSM SCIF client thread function: waits for messages from the host and sends
+ * a heartbeat to the host
+ */
+static int cosm_scif_client(void *unused)
+{
+ struct cosm_msg msg = { .id = COSM_MSG_HEARTBEAT };
+ struct scif_pollepd pollepd;
+ int rc;
+
+ allow_signal(SIGKILL);
+
+ while (!kthread_should_stop()) {
+ pollepd.epd = client_epd;
+ pollepd.events = POLLIN;
+
+ rc = scif_poll(&pollepd, 1, COSM_HEARTBEAT_SEND_MSEC);
+ if (rc < 0) {
+ if (-EINTR != rc)
+ dev_err(&client_spdev->dev,
+ "%s %d scif_poll rc %d\n",
+ __func__, __LINE__, rc);
+ continue;
+ }
+
+ if (pollepd.revents & POLLIN)
+ cosm_client_recv();
+
+ msg.id = COSM_MSG_HEARTBEAT;
+ rc = scif_send(client_epd, &msg, sizeof(msg), SCIF_SEND_BLOCK);
+ if (rc < 0)
+ dev_err(&client_spdev->dev, "%s %d scif_send rc %d\n",
+ __func__, __LINE__, rc);
+ }
+
+ dev_dbg(&client_spdev->dev, "%s %d Client thread stopped\n",
+ __func__, __LINE__);
+ return 0;
+}
+
+static void cosm_scif_probe(struct scif_peer_dev *spdev)
+{
+ int rc;
+
+ dev_dbg(&spdev->dev, "%s %d: dnode %d\n",
+ __func__, __LINE__, spdev->dnode);
+
+ /* We are only interested in the host with spdev->dnode == 0 */
+ if (spdev->dnode)
+ return;
+
+ client_spdev = spdev;
+ rc = cosm_scif_connect();
+ if (rc)
+ goto exit;
+
+ rc = register_reboot_notifier(&cosm_reboot);
+ if (rc) {
+ dev_err(&spdev->dev,
+ "reboot notifier registration failed rc %d\n", rc);
+ goto connect_exit;
+ }
+
+ client_thread = kthread_run(cosm_scif_client, NULL, "cosm_client");
+ if (IS_ERR(client_thread)) {
+ rc = PTR_ERR(client_thread);
+ dev_err(&spdev->dev, "%s %d kthread_run rc %d\n",
+ __func__, __LINE__, rc);
+ goto unreg_reboot;
+ }
+ return;
+unreg_reboot:
+ unregister_reboot_notifier(&cosm_reboot);
+connect_exit:
+ cosm_scif_connect_exit();
+exit:
+ client_spdev = NULL;
+}
+
+static void cosm_scif_remove(struct scif_peer_dev *spdev)
+{
+ int rc;
+
+ dev_dbg(&spdev->dev, "%s %d: dnode %d\n",
+ __func__, __LINE__, spdev->dnode);
+
+ if (spdev->dnode)
+ return;
+
+ if (!IS_ERR_OR_NULL(client_thread)) {
+ rc = send_sig(SIGKILL, client_thread, 0);
+ if (rc) {
+ pr_err("%s %d send_sig rc %d\n",
+ __func__, __LINE__, rc);
+ return;
+ }
+ kthread_stop(client_thread);
+ }
+ unregister_reboot_notifier(&cosm_reboot);
+ cosm_scif_connect_exit();
+ client_spdev = NULL;
+}
+
+static struct scif_client scif_client_cosm = {
+ .name = KBUILD_MODNAME,
+ .probe = cosm_scif_probe,
+ .remove = cosm_scif_remove,
+};
+
+static int __init cosm_client_init(void)
+{
+ int rc = scif_client_register(&scif_client_cosm);
+
+ if (rc)
+ pr_err("scif_client_register failed rc %d\n", rc);
+ return rc;
+}
+
+static void __exit cosm_client_exit(void)
+{
+ scif_client_unregister(&scif_client_cosm);
+}
+
+module_init(cosm_client_init);
+module_exit(cosm_client_exit);
+
+MODULE_AUTHOR("Intel Corporation");
+MODULE_DESCRIPTION("Intel(R) MIC card OS state management client driver");
+MODULE_LICENSE("GPL v2");
--
2.0.0.rc3.2.g998f840
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
[toc] | [prev] | [standalone]
Back to top | Article view | linux.kernel
csiph-web