Groups | Search | Server Info | Keyboard shortcuts | Login | Register [http] [https] [nntp] [nntps]

Groups > linux.kernel > #1742944 > unrolled thread

[Patch v4 00/22] CIFS: Implement SMBDirect

Back to article view | Back to linux.kernel

Contents

  [Patch v4 00/22] CIFS: Implement SMBDirect Long Li <longli@exchange.microsoft.com> - 2017-10-02 04:40 +0200
    [Patch v4 15/22] CIFS: SMBD: Upper layer sends data via RDMA send Long Li <longli@exchange.microsoft.com> - 2017-10-02 04:40 +0200
    [Patch v4 21/22] CIFS: SMBD: Upper layer performs SMB read via RDMA write through memory registration Long Li <longli@exchange.microsoft.com> - 2017-10-02 04:40 +0200
    [Patch v4 09/22] CIFS: SMBD: Implement function to destroy a SMBDirect connection Long Li <longli@exchange.microsoft.com> - 2017-10-02 04:40 +0200
    [Patch v4 18/22] CIFS: SMBD: Upper layer performs SMB write via RDMA read through memory registration Long Li <longli@exchange.microsoft.com> - 2017-10-02 04:40 +0200
    [Patch v4 13/22] CIFS: SMBD: Upper layer receives data via RDMA receive Long Li <longli@exchange.microsoft.com> - 2017-10-02 04:40 +0200
    [Patch v4 10/22] CIFS: SMBD: Upper layer destroys SMBDirect session on shutdown or umount Long Li <longli@exchange.microsoft.com> - 2017-10-02 04:40 +0200
    [Patch v4 06/22] CIFS: SMBD: Upper layer connects to SMBDirect session Long Li <longli@exchange.microsoft.com> - 2017-10-02 04:40 +0200
    [Patch v4 19/22] CIFS: SMBD: Add parameter rdata to smb2_new_read_req Long Li <longli@exchange.microsoft.com> - 2017-10-02 04:40 +0200
    [Patch v4 17/22] CIFS: SMBD: Implement RDMA memory registration Long Li <longli@exchange.microsoft.com> - 2017-10-02 04:40 +0200

#1742944 — [Patch v4 00/22] CIFS: Implement SMBDirect

From: Long Li <longli@microsoft.com>

Starting with SMB2 dialect 3.0, Microsoft introduced SMBDirect transport
protocol for transferring upper layer (SMB2) payload over RDMA via Infiniband,
RoCE or iWARP. The prococol is published in [MS-SMBD]
(https://msdn.microsoft.com/en-us/library/hh536346.aspx).

Patch v2 added RDMA read/write via memory registration, and addressed
feedbacks on v1.

Patch v3 improved performance by introducing an additional queue for handling
empty packets and reducing lock contention on IRQ path. Also added light
weight profiling by reading TSC and addressed feedbacks on v2.

Patch v4 fixed connectivity issues with iWAPR devices and addressed comments.

Long Li (22):
  CIFS: SMBD: Add SMBDirect protocol initial values and constants
  CIFS: SMBD: Establish SMBDirect connection
  CIFS: SMBD: export protocol initial values
  CIFS: SMBD: Add rdma mount option
  CIFS: SMBD: Implement function to create a SMBDirect connection
  CIFS: SMBD: Upper layer connects to SMBDirect session
  CIFS: SMBD: Implement function to reconnect to a SMBDirect transport
  CIFS: SMBD: Upper layer reconnects to SMBDirect session
  CIFS: SMBD: Implement function to destroy a SMBDirect connection
  CIFS: SMBD: Upper layer destroys SMBDirect session on shutdown or
    umount
  CIFS: SMBD: Set SMBDirect maximum read or write size for I/O
  CIFS: SMBD: Implement function to receive data via RDMA receive
  CIFS: SMBD: Upper layer receives data via RDMA receive
  CIFS: SMBD: Implement function to send data via RDMA send
  CIFS: SMBD: Upper layer sends data via RDMA send
  CIFS: SMBD: Fix the definition for SMB2_CHANNEL_RDMA_V1_INVALIDATE
  CIFS: SMBD: Implement RDMA memory registration
  CIFS: SMBD: Upper layer performs SMB write via RDMA read through
    memory registration
  CIFS: SMBD: Add parameter rdata to smb2_new_read_req
  CIFS: SMBD: Read correct returned data length for RDMA write (SMB
    read) I/O
  CIFS: SMBD: Upper layer performs SMB read via RDMA write through
    memory registration
  CIFS: SMBD: Add SMBDirect debug counters

 fs/cifs/Makefile     |    2 +-
 fs/cifs/cifs_debug.c |  159 +++
 fs/cifs/cifsfs.c     |    2 +
 fs/cifs/cifsglob.h   |   17 +-
 fs/cifs/cifssmb.c    |   10 +-
 fs/cifs/connect.c    |   46 +-
 fs/cifs/file.c       |   10 +
 fs/cifs/smb1ops.c    |    2 +-
 fs/cifs/smb2ops.c    |   21 +-
 fs/cifs/smb2pdu.c    |  114 ++-
 fs/cifs/smb2pdu.h    |    2 +-
 fs/cifs/smbdirect.c  | 2651 ++++++++++++++++++++++++++++++++++++++++++++++++++
 fs/cifs/smbdirect.h  |  325 +++++++
 fs/cifs/transport.c  |    7 +
 14 files changed, 3348 insertions(+), 20 deletions(-)
 create mode 100644 fs/cifs/smbdirect.c
 create mode 100644 fs/cifs/smbdirect.h

-- 
2.7.4

[toc] | [next] | [standalone]

#1742945 — [Patch v4 15/22] CIFS: SMBD: Upper layer sends data via RDMA send

From: Long Li <longli@microsoft.com>

With SMBDirect connected, use it for sending data via RDMA send.

Signed-off-by: Long Li <longli@microsoft.com>
---
 fs/cifs/transport.c | 7 +++++++
 1 file changed, 7 insertions(+)

diff --git a/fs/cifs/transport.c b/fs/cifs/transport.c
index 7efbab0..3a9b5a0 100644
--- a/fs/cifs/transport.c
+++ b/fs/cifs/transport.c
@@ -37,6 +37,7 @@
 #include "cifsglob.h"
 #include "cifsproto.h"
 #include "cifs_debug.h"
+#include "smbdirect.h"
 
 void
 cifs_wake_up_task(struct mid_q_entry *mid)
@@ -230,6 +231,11 @@ __smb_send_rqst(struct TCP_Server_Info *server, struct smb_rqst *rqst)
 	struct msghdr smb_msg;
 	int val = 1;
 
+	if (server->smbd_conn) {
+		rc = smbd_send(server->smbd_conn, rqst);
+		goto done;
+	}
+
 	if (ssocket == NULL)
 		return -ENOTSOCK;
 
@@ -299,6 +305,7 @@ __smb_send_rqst(struct TCP_Server_Info *server, struct smb_rqst *rqst)
 		server->tcpStatus = CifsNeedReconnect;
 	}
 
+done:
 	if (rc < 0 && rc != -EINTR)
 		cifs_dbg(VFS, "Error %d sending data on socket to server\n",
 			 rc);
-- 
2.7.4

[toc] | [prev] | [next] | [standalone]

#1742946 — [Patch v4 21/22] CIFS: SMBD: Upper layer performs SMB read via RDMA write through memory registration

From: Long Li <longli@microsoft.com>

If I/O size is larger than rdma_readwrite_threshold, use RDMA write for
SMB read by specifying channel SMB2_CHANNEL_RDMA_V1 or
SMB2_CHANNEL_RDMA_V1_INVALIDATE in the SMB packet, depending on SMB dialect
used. Append a smbd_buffer_descriptor_v1 to the end of the SMB packet and fill
in other values to indicate this SMB read uses RDMA write.

There is no need to read from the transport for incoming payload. At the time
SMB read response comes back, the data is already transfered and placed in the
pages by RDMA hardware.

When SMB read is finished, deregister the memory regions if RDMA write is used
for this SMB read. smbd_deregister_mr may need to do local invalidation and
sleep, if server remote invalidation is not used.

There are situations where the MID may not be created on I/O failure, under
which memory region is deregistered when read data context is released.

Signed-off-by: Long Li <longli@microsoft.com>
---
 fs/cifs/cifsglob.h |  1 +
 fs/cifs/file.c     | 10 ++++++++++
 fs/cifs/smb2pdu.c  | 43 +++++++++++++++++++++++++++++++++++++++++++
 3 files changed, 54 insertions(+)

diff --git a/fs/cifs/cifsglob.h b/fs/cifs/cifsglob.h
index f851b50..30b99a5 100644
--- a/fs/cifs/cifsglob.h
+++ b/fs/cifs/cifsglob.h
@@ -1152,6 +1152,7 @@ struct cifs_readdata {
 				struct cifs_readdata *rdata,
 				struct iov_iter *iter);
 	struct kvec			iov[2];
+	struct smbd_mr			*mr;
 	unsigned int			pagesz;
 	unsigned int			tailsz;
 	unsigned int			credits;
diff --git a/fs/cifs/file.c b/fs/cifs/file.c
index 0786f19..8396f1e 100644
--- a/fs/cifs/file.c
+++ b/fs/cifs/file.c
@@ -42,6 +42,7 @@
 #include "cifs_debug.h"
 #include "cifs_fs_sb.h"
 #include "fscache.h"
+#include "smbdirect.h"
 
 
 static inline int cifs_convert_flags(unsigned int flags)
@@ -2909,6 +2910,11 @@ cifs_readdata_release(struct kref *refcount)
 	struct cifs_readdata *rdata = container_of(refcount,
 					struct cifs_readdata, refcount);
 
+	if (rdata->mr) {
+		smbd_deregister_mr(rdata->mr);
+		rdata->mr = NULL;
+	}
+
 	if (rdata->cfile)
 		cifsFileInfo_put(rdata->cfile);
 
@@ -3037,6 +3043,8 @@ uncached_fill_pages(struct TCP_Server_Info *server,
 		}
 		if (iter)
 			result = copy_page_from_iter(page, 0, n, iter);
+		else if (rdata->mr)
+			result = n;
 		else
 			result = cifs_read_page_from_socket(server, page, n);
 		if (result < 0)
@@ -3606,6 +3614,8 @@ readpages_fill_pages(struct TCP_Server_Info *server,
 
 		if (iter)
 			result = copy_page_from_iter(page, 0, n, iter);
+		else if (rdata->mr)
+			result = n;
 		else
 			result = cifs_read_page_from_socket(server, page, n);
 		if (result < 0)
diff --git a/fs/cifs/smb2pdu.c b/fs/cifs/smb2pdu.c
index 7053db9..31dcee0 100644
--- a/fs/cifs/smb2pdu.c
+++ b/fs/cifs/smb2pdu.c
@@ -2380,6 +2380,39 @@ smb2_new_read_req(void **buf, unsigned int *total_len,
 	req->Length = cpu_to_le32(io_parms->length);
 	req->Offset = cpu_to_le64(io_parms->offset);
 
+	/*
+	 * If we want to do a RDMA write, fill in and append
+	 * smbd_buffer_descriptor_v1 to the end of read request
+	 */
+	if (server->rdma && rdata &&
+		rdata->bytes >= server->smbd_conn->rdma_readwrite_threshold) {
+
+		struct smbd_buffer_descriptor_v1 *v1;
+		bool need_invalidate =
+			io_parms->tcon->ses->server->dialect == SMB30_PROT_ID;
+
+		rdata->mr = smbd_register_mr(
+				server->smbd_conn, rdata->pages,
+				rdata->nr_pages, rdata->tailsz,
+				true, need_invalidate);
+		if (!rdata->mr)
+			return -ENOBUFS;
+
+		req->Channel = SMB2_CHANNEL_RDMA_V1_INVALIDATE;
+		if (need_invalidate)
+			req->Channel = SMB2_CHANNEL_RDMA_V1;
+		req->ReadChannelInfoOffset =
+			offsetof(struct smb2_read_plain_req, Buffer);
+		req->ReadChannelInfoLength =
+			sizeof(struct smbd_buffer_descriptor_v1);
+		v1 = (struct smbd_buffer_descriptor_v1 *) &req->Buffer[0];
+		v1->offset = rdata->mr->mr->iova;
+		v1->token = rdata->mr->mr->rkey;
+		v1->length = rdata->mr->mr->length;
+
+		*total_len += sizeof(*v1) - 1;
+	}
+
 	if (request_type & CHAINED_REQUEST) {
 		if (!(request_type & END_OF_CHAIN)) {
 			/* next 8-byte aligned request */
@@ -2459,6 +2492,16 @@ smb2_readv_callback(struct mid_q_entry *mid)
 			rdata->result = -EIO;
 	}
 
+	/*
+	 * If this rdata has a memmory registered, the MR can be freed
+	 * MR needs to be freed as soon as I/O finishes to prevent deadlock
+	 * because they have limited number and are used for future I/Os
+	 */
+	if (rdata->mr) {
+		smbd_deregister_mr(rdata->mr);
+		rdata->mr = NULL;
+	}
+
 	if (rdata->result)
 		cifs_stats_fail_inc(tcon, SMB2_READ_HE);
 
-- 
2.7.4

[toc] | [prev] | [next] | [standalone]

#1742947 — [Patch v4 09/22] CIFS: SMBD: Implement function to destroy a SMBDirect connection

From: Long Li <longli@microsoft.com>

Add function to tear down a SMBDirect connection. This is used by upper layer
to free all SMBDirect connection and transport resources.

Signed-off-by: Long Li <longli@microsoft.com>
---
 fs/cifs/smbdirect.c | 16 ++++++++++++++++
 fs/cifs/smbdirect.h |  3 +++
 2 files changed, 19 insertions(+)

diff --git a/fs/cifs/smbdirect.c b/fs/cifs/smbdirect.c
index 1f0f33c..cb129c2 100644
--- a/fs/cifs/smbdirect.c
+++ b/fs/cifs/smbdirect.c
@@ -1416,6 +1416,22 @@ static void idle_connection_timer(struct work_struct *work)
 			info->keep_alive_interval*HZ);
 }
 
+/* Destroy this SMBD connection, called from upper layer */
+void smbd_destroy(struct smbd_connection *info)
+{
+	log_rdma_event(INFO, "destroying rdma session\n");
+
+	/* Kick off the disconnection process */
+	smbd_disconnect_rdma_connection(info);
+
+	log_rdma_event(INFO, "wait for transport being destroyed\n");
+	wait_event(info->wait_destroy,
+		info->transport_status == SMBD_DESTROYED);
+
+	destroy_workqueue(info->workqueue);
+	kfree(info);
+}
+
 /*
  * Reconnect this SMBD connection, called from upper layer
  * return value: 0 on success, or actual error code
diff --git a/fs/cifs/smbdirect.h b/fs/cifs/smbdirect.h
index 9818852..d14a484 100644
--- a/fs/cifs/smbdirect.h
+++ b/fs/cifs/smbdirect.h
@@ -252,6 +252,9 @@ struct smbd_connection *smbd_get_connection(
 /* Reconnect SMBDirect session */
 int smbd_reconnect(struct TCP_Server_Info *server);
 
+/* Destroy SMBDirect session */
+void smbd_destroy(struct smbd_connection *info);
+
 void profiling_display_histogram(
 	struct seq_file *m, unsigned long long array[]);
 #endif
-- 
2.7.4

[toc] | [prev] | [next] | [standalone]

#1742948 — [Patch v4 18/22] CIFS: SMBD: Upper layer performs SMB write via RDMA read through memory registration

From: Long Li <longli@microsoft.com>

When sending I/O, if size is larger than rdma_readwrite_threshold we prepare
to send SMB write packet for a RDMA read via memory registration. The actual
I/O is done by remote peer through local RDMA hardware. Modify the relevant
fields in the packet accordingly, and append a smbd_buffer_descriptor_v1 to
the end of the SMB write packet.

On write I/O finish, deregister the memory region if this was for a RDMA read.
If remote invalidation is not used, the call to smbd_deregister_mr will do
local invalidation and possibly wait. Memory region is normally deregistered
in MID callback as soon as it's used. There are situations where the MID may
not be created on I/O failure, under which memory region is deregistered when
write data context is released.

Signed-off-by: Long Li <longli@microsoft.com>
---
 fs/cifs/cifsglob.h |  1 +
 fs/cifs/cifssmb.c  |  6 ++++++
 fs/cifs/smb2pdu.c  | 57 +++++++++++++++++++++++++++++++++++++++++++++++++++++-
 3 files changed, 63 insertions(+), 1 deletion(-)

diff --git a/fs/cifs/cifsglob.h b/fs/cifs/cifsglob.h
index 5585516..bcb6df1 100644
--- a/fs/cifs/cifsglob.h
+++ b/fs/cifs/cifsglob.h
@@ -1168,6 +1168,7 @@ struct cifs_writedata {
 	pid_t				pid;
 	unsigned int			bytes;
 	int				result;
+	struct smbd_mr			*mr;
 	unsigned int			pagesz;
 	unsigned int			tailsz;
 	unsigned int			credits;
diff --git a/fs/cifs/cifssmb.c b/fs/cifs/cifssmb.c
index 5857009..0e29ecf 100644
--- a/fs/cifs/cifssmb.c
+++ b/fs/cifs/cifssmb.c
@@ -43,6 +43,7 @@
 #include "cifs_unicode.h"
 #include "cifs_debug.h"
 #include "fscache.h"
+#include "smbdirect.h"
 
 #ifdef CONFIG_CIFS_POSIX
 static struct {
@@ -1912,6 +1913,11 @@ cifs_writedata_release(struct kref *refcount)
 	struct cifs_writedata *wdata = container_of(refcount,
 					struct cifs_writedata, refcount);
 
+	if (wdata->mr) {
+		smbd_deregister_mr(wdata->mr);
+		wdata->mr = NULL;
+	}
+
 	if (wdata->cfile)
 		cifsFileInfo_put(wdata->cfile);
 
diff --git a/fs/cifs/smb2pdu.c b/fs/cifs/smb2pdu.c
index bab3da6..6089957 100644
--- a/fs/cifs/smb2pdu.c
+++ b/fs/cifs/smb2pdu.c
@@ -48,6 +48,7 @@
 #include "smb2glob.h"
 #include "cifspdu.h"
 #include "cifs_spnego.h"
+#include "smbdirect.h"
 
 /*
  *  The following table defines the expected "StructureSize" of SMB2 requests
@@ -2653,6 +2654,18 @@ smb2_writev_callback(struct mid_q_entry *mid)
 		break;
 	}
 
+	/*
+	 * If this wdata has a memory registered, the MR can be freed
+	 * The number of MRs available is limited, it's important to recover
+	 * used MR as soon as I/O is finished. Hold MR longer in the later
+	 * I/O process can possibly result in I/O deadlock due to lack of MR
+	 * to send request on I/O retry
+	 */
+	if (wdata->mr) {
+		smbd_deregister_mr(wdata->mr);
+		wdata->mr = NULL;
+	}
+
 	if (wdata->result)
 		cifs_stats_fail_inc(tcon, SMB2_WRITE_HE);
 
@@ -2704,6 +2717,41 @@ smb2_async_writev(struct cifs_writedata *wdata,
 				offsetof(struct smb2_write_req, Buffer) - 4);
 	req->RemainingBytes = 0;
 
+	/*
+	 * If we want to do a server RDMA read, fill in and append
+	 * smbd_buffer_descriptor_v1 to the end of write request
+	 */
+	if (server->rdma && wdata->bytes >=
+		server->smbd_conn->rdma_readwrite_threshold) {
+
+		struct smbd_buffer_descriptor_v1 *v1;
+		bool need_invalidate = server->dialect == SMB30_PROT_ID;
+
+		wdata->mr = smbd_register_mr(
+				server->smbd_conn, wdata->pages,
+				wdata->nr_pages, wdata->tailsz,
+				false, need_invalidate);
+		if (!wdata->mr) {
+			rc = -ENOBUFS;
+			goto async_writev_out;
+		}
+		req->Length = 0;
+		req->DataOffset = 0;
+		req->RemainingBytes =
+			(wdata->nr_pages-1)*PAGE_SIZE + wdata->tailsz;
+		req->Channel = SMB2_CHANNEL_RDMA_V1_INVALIDATE;
+		if (need_invalidate)
+			req->Channel = SMB2_CHANNEL_RDMA_V1;
+		req->WriteChannelInfoOffset =
+			offsetof(struct smb2_write_req, Buffer) - 4;
+		req->WriteChannelInfoLength =
+			sizeof(struct smbd_buffer_descriptor_v1);
+		v1 = (struct smbd_buffer_descriptor_v1 *) &req->Buffer[0];
+		v1->offset = wdata->mr->mr->iova;
+		v1->token = wdata->mr->mr->rkey;
+		v1->length = wdata->mr->mr->length;
+	}
+
 	/* 4 for rfc1002 length field and 1 for Buffer */
 	iov[0].iov_len = 4;
 	iov[0].iov_base = req;
@@ -2717,10 +2765,17 @@ smb2_async_writev(struct cifs_writedata *wdata,
 	rqst.rq_pagesz = wdata->pagesz;
 	rqst.rq_tailsz = wdata->tailsz;
 
+	if (wdata->mr) {
+		iov[1].iov_len += sizeof(struct smbd_buffer_descriptor_v1);
+		rqst.rq_npages = 0;
+	}
+
 	cifs_dbg(FYI, "async write at %llu %u bytes\n",
 		 wdata->offset, wdata->bytes);
 
-	req->Length = cpu_to_le32(wdata->bytes);
+	/* For RDMA read, I/O size is in RemainingBytes not in Length */
+	if (!wdata->mr)
+		req->Length = cpu_to_le32(wdata->bytes);
 
 	inc_rfc1001_len(&req->hdr, wdata->bytes - 1 /* Buffer */);
 
-- 
2.7.4

[toc] | [prev] | [next] | [standalone]

#1742949 — [Patch v4 13/22] CIFS: SMBD: Upper layer receives data via RDMA receive

From: Long Li <longli@microsoft.com>

With SMBDirect connected, use it for receiving data via RDMA receive.

Signed-off-by: Long Li <longli@microsoft.com>
---
 fs/cifs/connect.c | 5 ++++-
 1 file changed, 4 insertions(+), 1 deletion(-)

diff --git a/fs/cifs/connect.c b/fs/cifs/connect.c
index 1a9f22f..8026682 100644
--- a/fs/cifs/connect.c
+++ b/fs/cifs/connect.c
@@ -542,7 +542,10 @@ cifs_readv_from_socket(struct TCP_Server_Info *server, struct msghdr *smb_msg)
 		if (server_unresponsive(server))
 			return -ECONNABORTED;
 
-		length = sock_recvmsg(server->ssocket, smb_msg, 0);
+		if (server->smbd_conn)
+			length = smbd_recv(server->smbd_conn, smb_msg);
+		else
+			length = sock_recvmsg(server->ssocket, smb_msg, 0);
 
 		if (server->tcpStatus == CifsExiting)
 			return -ESHUTDOWN;
-- 
2.7.4

[toc] | [prev] | [next] | [standalone]

#1742950 — [Patch v4 10/22] CIFS: SMBD: Upper layer destroys SMBDirect session on shutdown or umount

From: Long Li <longli@microsoft.com>

When CIFS wants to umount, call shutdown on transport when SMBDirect is used.

Signed-off-by: Long Li <longli@microsoft.com>
---
 fs/cifs/connect.c | 5 +++++
 1 file changed, 5 insertions(+)

diff --git a/fs/cifs/connect.c b/fs/cifs/connect.c
index 26ad706..1a9f22f 100644
--- a/fs/cifs/connect.c
+++ b/fs/cifs/connect.c
@@ -704,6 +704,11 @@ static void clean_demultiplex_info(struct TCP_Server_Info *server)
 	/* give those requests time to exit */
 	msleep(125);
 
+	if (server->smbd_conn) {
+		smbd_destroy(server->smbd_conn);
+		server->smbd_conn = NULL;
+	}
+
 	if (server->ssocket) {
 		sock_release(server->ssocket);
 		server->ssocket = NULL;
-- 
2.7.4

[toc] | [prev] | [next] | [standalone]

#1742951 — [Patch v4 06/22] CIFS: SMBD: Upper layer connects to SMBDirect session

From: Long Li <longli@microsoft.com>

When "rdma" is specified in the mount option, CIFS attempts to connect to
SMBDirect instead of TCP socket.

Signed-off-by: Long Li <longli@microsoft.com>
---
 fs/cifs/connect.c | 15 +++++++++++++++
 1 file changed, 15 insertions(+)

diff --git a/fs/cifs/connect.c b/fs/cifs/connect.c
index b5a575f..94b6357 100644
--- a/fs/cifs/connect.c
+++ b/fs/cifs/connect.c
@@ -45,6 +45,7 @@
 #include <linux/parser.h>
 #include <linux/bvec.h>
 
+#include "smbdirect.h"
 #include "cifspdu.h"
 #include "cifsglob.h"
 #include "cifsproto.h"
@@ -2280,12 +2281,26 @@ cifs_get_tcp_session(struct smb_vol *volume_info)
 	else
 		tcp_ses->echo_interval = SMB_ECHO_INTERVAL_DEFAULT * HZ;
 
+	if (tcp_ses->rdma) {
+		tcp_ses->smbd_conn = smbd_get_connection(
+			tcp_ses, (struct sockaddr *)&volume_info->dstaddr);
+		if (tcp_ses->smbd_conn) {
+			cifs_dbg(VFS, "RDMA transport established\n");
+			rc = 0;
+			goto connected;
+		} else {
+			rc = -ENOENT;
+			goto out_err_crypto_release;
+		}
+	}
+
 	rc = ip_connect(tcp_ses);
 	if (rc < 0) {
 		cifs_dbg(VFS, "Error connecting to socket. Aborting operation.\n");
 		goto out_err_crypto_release;
 	}
 
+connected:
 	/*
 	 * since we're in a cifs function already, we know that
 	 * this will succeed. No need for try_module_get().
-- 
2.7.4

[toc] | [prev] | [next] | [standalone]

#1742952 — [Patch v4 19/22] CIFS: SMBD: Add parameter rdata to smb2_new_read_req

From: Long Li <longli@microsoft.com>

This patch is for preparing upper layer for doing SMB read via RDMA write.

When we assemble the SMB read packet header, we need to know the I/O layout
if this request is to use a RDMA write. rdata has all the information we need
for memory registration. Add rdata to smb2_new_read_req.

Signed-off-by: Long Li <longli@microsoft.com>
---
 fs/cifs/smb2pdu.c | 14 +++++++++-----
 1 file changed, 9 insertions(+), 5 deletions(-)

diff --git a/fs/cifs/smb2pdu.c b/fs/cifs/smb2pdu.c
index 6089957..7053db9 100644
--- a/fs/cifs/smb2pdu.c
+++ b/fs/cifs/smb2pdu.c
@@ -2351,18 +2351,21 @@ SMB2_flush(const unsigned int xid, struct cifs_tcon *tcon, u64 persistent_fid,
  */
 static int
 smb2_new_read_req(void **buf, unsigned int *total_len,
-		  struct cifs_io_parms *io_parms, unsigned int remaining_bytes,
-		  int request_type)
+	struct cifs_io_parms *io_parms, struct cifs_readdata *rdata,
+	unsigned int remaining_bytes, int request_type)
 {
 	int rc = -EACCES;
 	struct smb2_read_plain_req *req = NULL;
 	struct smb2_sync_hdr *shdr;
+	struct TCP_Server_Info *server;
 
 	rc = smb2_plain_req_init(SMB2_READ, io_parms->tcon, (void **) &req,
 				 total_len);
 	if (rc)
 		return rc;
-	if (io_parms->tcon->ses->server == NULL)
+
+	server = io_parms->tcon->ses->server;
+	if (server == NULL)
 		return -ECONNABORTED;
 
 	shdr = &req->sync_hdr;
@@ -2490,7 +2493,8 @@ smb2_async_readv(struct cifs_readdata *rdata)
 
 	server = io_parms.tcon->ses->server;
 
-	rc = smb2_new_read_req((void **) &buf, &total_len, &io_parms, 0, 0);
+	rc = smb2_new_read_req(
+		(void **) &buf, &total_len, &io_parms, rdata, 0, 0);
 	if (rc) {
 		if (rc == -EAGAIN && rdata->credits) {
 			/* credits was reset by reconnect */
@@ -2558,7 +2562,7 @@ SMB2_read(const unsigned int xid, struct cifs_io_parms *io_parms,
 	struct cifs_ses *ses = io_parms->tcon->ses;
 
 	*nbytes = 0;
-	rc = smb2_new_read_req((void **)&req, &total_len, io_parms, 0, 0);
+	rc = smb2_new_read_req((void **)&req, &total_len, io_parms, NULL, 0, 0);
 	if (rc)
 		return rc;
 
-- 
2.7.4

[toc] | [prev] | [next] | [standalone]

#1742953 — [Patch v4 17/22] CIFS: SMBD: Implement RDMA memory registration

From: Long Li <longli@microsoft.com>

Memory registration is used for transferring payload via RDMA read or write.
After I/O is done, memory registrations are recovered and reused. This
process can be time consuming and is done in a work queue.

Signed-off-by: Long Li <longli@microsoft.com>
---
 fs/cifs/smbdirect.c | 428 ++++++++++++++++++++++++++++++++++++++++++++++++++++
 fs/cifs/smbdirect.h |  55 +++++++
 2 files changed, 483 insertions(+)

diff --git a/fs/cifs/smbdirect.c b/fs/cifs/smbdirect.c
index 90e2c94..3f2de48 100644
--- a/fs/cifs/smbdirect.c
+++ b/fs/cifs/smbdirect.c
@@ -49,6 +49,9 @@ static int smbd_post_send_page(struct smbd_connection *info,
 		struct page *page, unsigned long offset,
 		size_t size, int remaining_data_length);
 
+static void destroy_mr_list(struct smbd_connection *info);
+static int allocate_mr_list(struct smbd_connection *info);
+
 /* SMBD version number */
 #define SMBD_V1	0x0100
 
@@ -219,6 +222,12 @@ static void smbd_destroy_rdma_work(struct work_struct *work)
 	wait_event(info->wait_send_payload_pending,
 		atomic_read(&info->send_payload_pending) == 0);
 
+	log_rdma_event(INFO, "freeing mr list\n");
+	wake_up_interruptible_all(&info->wait_mr);
+	wait_event(info->wait_for_mr_cleanup,
+		atomic_read(&info->mr_used_count) == 0);
+	destroy_mr_list(info);
+
 	/* It's not posssible for upper layer to get to reassembly */
 	log_rdma_event(INFO, "drain the reassembly queue\n");
 	do {
@@ -475,6 +484,16 @@ static bool process_negotiation_response(
 	}
 	info->max_fragmented_send_size =
 		le32_to_cpu(packet->max_fragmented_size);
+	info->rdma_readwrite_threshold =
+		rdma_readwrite_threshold > info->max_fragmented_send_size ?
+		info->max_fragmented_send_size :
+		rdma_readwrite_threshold;
+
+
+	info->max_readwrite_size = min_t(u32,
+			le32_to_cpu(packet->max_readwrite_size),
+			info->max_frmr_depth * PAGE_SIZE);
+	info->max_frmr_depth = info->max_readwrite_size / PAGE_SIZE;
 
 	return true;
 }
@@ -773,6 +792,12 @@ static int smbd_ia_open(
 		rc = -EPROTONOSUPPORT;
 		goto out2;
 	}
+	info->max_frmr_depth = min_t(int,
+		smbd_max_frmr_depth,
+		info->id->device->attrs.max_fast_reg_page_list_len);
+	info->mr_type = IB_MR_TYPE_MEM_REG;
+	if (info->id->device->attrs.device_cap_flags & IB_DEVICE_SG_GAPS_REG)
+		info->mr_type = IB_MR_TYPE_SG_GAPS;
 
 	info->pd = ib_alloc_pd(info->id->device, 0);
 	if (IS_ERR(info->pd)) {
@@ -1610,6 +1635,8 @@ struct smbd_connection *_smbd_get_connection(
 	struct rdma_conn_param conn_param;
 	struct ib_qp_init_attr qp_attr;
 	struct sockaddr_in *addr_in = (struct sockaddr_in *) dstaddr;
+	struct ib_port_immutable port_immutable;
+	u32 ird_ord_hdr[2];
 
 	info = kzalloc(sizeof(struct smbd_connection), GFP_KERNEL);
 	if (!info)
@@ -1698,6 +1725,28 @@ struct smbd_connection *_smbd_get_connection(
 	memset(&conn_param, 0, sizeof(conn_param));
 	conn_param.initiator_depth = 0;
 
+	conn_param.responder_resources =
+		info->id->device->attrs.max_qp_rd_atom
+			< SMBD_CM_RESPONDER_RESOURCES ?
+		info->id->device->attrs.max_qp_rd_atom :
+		SMBD_CM_RESPONDER_RESOURCES;
+	info->responder_resources = conn_param.responder_resources;
+	log_rdma_mr(INFO, "responder_resources=%d\n",
+		info->responder_resources);
+
+	/* Need to send IRD/ORD in private data for iWARP */
+	info->id->device->get_port_immutable(
+		info->id->device, info->id->port_num, &port_immutable);
+	if (port_immutable.core_cap_flags & RDMA_CORE_PORT_IWARP) {
+		ird_ord_hdr[0] = info->responder_resources;
+		ird_ord_hdr[1] = 1;
+		conn_param.private_data = ird_ord_hdr;
+		conn_param.private_data_len = sizeof(ird_ord_hdr);
+	} else {
+		conn_param.private_data = NULL;
+		conn_param.private_data_len = 0;
+	}
+
 	conn_param.retry_count = SMBD_CM_RETRY;
 	conn_param.rnr_retry_count = SMBD_CM_RNR_RETRY;
 	conn_param.flow_control = 0;
@@ -1762,8 +1811,19 @@ struct smbd_connection *_smbd_get_connection(
 		goto negotiation_failed;
 	}
 
+	rc = allocate_mr_list(info);
+	if (rc) {
+		log_rdma_mr(ERR, "memory registration allocation failed\n");
+		goto allocate_mr_failed;
+	}
+
 	return info;
 
+allocate_mr_failed:
+	/* At this point, need to a full transport shutdown */
+	smbd_destroy(info);
+	return NULL;
+
 negotiation_failed:
 	cancel_delayed_work_sync(&info->idle_timer_work);
 	destroy_caches_and_workqueue(info);
@@ -2221,3 +2281,371 @@ int smbd_send(struct smbd_connection *info, struct smb_rqst *rqst)
 
 	return rc;
 }
+
+static void register_mr_done(struct ib_cq *cq, struct ib_wc *wc)
+{
+	struct smbd_mr *mr;
+	struct ib_cqe *cqe;
+
+	if (wc->status) {
+		log_rdma_mr(ERR, "status=%d\n", wc->status);
+		cqe = wc->wr_cqe;
+		mr = container_of(cqe, struct smbd_mr, cqe);
+		smbd_disconnect_rdma_connection(mr->conn);
+	}
+}
+
+/*
+ * The work queue function that recovers MRs
+ * We need to call ib_dereg_mr() and ib_alloc_mr() before this MR can be used
+ * again. Both calls are slow, so finish them in a workqueue. This will not
+ * block I/O path.
+ * There is one workqueue that recovers MRs, there is no need to lock as the
+ * I/O requests calling smbd_register_mr will never update the links in the
+ * mr_list.
+ */
+static void smbd_mr_recovery_work(struct work_struct *work)
+{
+	struct smbd_connection *info =
+		container_of(work, struct smbd_connection, mr_recovery_work);
+	struct smbd_mr *smbdirect_mr;
+	int rc;
+
+	list_for_each_entry(smbdirect_mr, &info->mr_list, list) {
+		if (smbdirect_mr->state == MR_INVALIDATED ||
+			smbdirect_mr->state == MR_ERROR) {
+
+			if (smbdirect_mr->state == MR_INVALIDATED) {
+				ib_dma_unmap_sg(
+					info->id->device, smbdirect_mr->sgl,
+					smbdirect_mr->sgl_count,
+					smbdirect_mr->dir);
+				smbdirect_mr->state = MR_READY;
+			} else if (smbdirect_mr->state == MR_ERROR) {
+
+				/* recover this MR entry */
+				rc = ib_dereg_mr(smbdirect_mr->mr);
+				if (rc) {
+					log_rdma_mr(ERR,
+						"ib_dereg_mr faield rc=%x\n",
+						rc);
+					smbd_disconnect_rdma_connection(info);
+				}
+
+				smbdirect_mr->mr = ib_alloc_mr(
+					info->pd, info->mr_type,
+					info->max_frmr_depth);
+				if (IS_ERR(smbdirect_mr->mr)) {
+					log_rdma_mr(ERR,
+						"ib_alloc_mr failed mr_type=%x "
+						"max_frmr_depth=%x\n",
+						info->mr_type,
+						info->max_frmr_depth);
+					smbd_disconnect_rdma_connection(info);
+				}
+
+				smbdirect_mr->state = MR_READY;
+			}
+			/* smbdirect_mr->state is updated by this function
+			 * and is read and updated by I/O issuing CPUs trying
+			 * to get a MR, the call to atomic_inc_return
+			 * implicates a memory barrier and guarantees this
+			 * value is updated before waking up any calls to
+			 * get_mr() from the I/O issuing CPUs
+			 */
+			if (atomic_inc_return(&info->mr_ready_count) == 1)
+				wake_up_interruptible(&info->wait_mr);
+		}
+	}
+}
+
+static void destroy_mr_list(struct smbd_connection *info)
+{
+	struct smbd_mr *mr, *tmp;
+
+	cancel_work_sync(&info->mr_recovery_work);
+	list_for_each_entry_safe(mr, tmp, &info->mr_list, list) {
+		if (mr->state == MR_INVALIDATED)
+			ib_dma_unmap_sg(info->id->device, mr->sgl,
+				mr->sgl_count, mr->dir);
+		ib_dereg_mr(mr->mr);
+		kfree(mr->sgl);
+		kfree(mr);
+	}
+}
+
+/*
+ * Allocate MRs used for RDMA read/write
+ * The number of MRs will not exceed hardware capability in responder_resources
+ * All MRs are kept in mr_list. The MR can be recovered after it's used
+ * Recovery is done in smbd_mr_recovery_work. The content of list entry changes
+ * as MRs are used and recovered for I/O, but the list links will not change
+ */
+static int allocate_mr_list(struct smbd_connection *info)
+{
+	int i;
+	struct smbd_mr *smbdirect_mr, *tmp;
+
+	INIT_LIST_HEAD(&info->mr_list);
+	init_waitqueue_head(&info->wait_mr);
+	spin_lock_init(&info->mr_list_lock);
+	atomic_set(&info->mr_ready_count, 0);
+	atomic_set(&info->mr_used_count, 0);
+	init_waitqueue_head(&info->wait_for_mr_cleanup);
+	/* Allocate more MRs (2x) than hardware responder_resources */
+	for (i = 0; i < info->responder_resources * 2; i++) {
+		smbdirect_mr = kzalloc(sizeof(*smbdirect_mr), GFP_KERNEL);
+		if (!smbdirect_mr)
+			goto out;
+		smbdirect_mr->mr = ib_alloc_mr(info->pd, info->mr_type,
+					info->max_frmr_depth);
+		if (IS_ERR(smbdirect_mr->mr)) {
+			log_rdma_mr(ERR, "ib_alloc_mr failed mr_type=%x "
+				"max_frmr_depth=%x\n",
+				info->mr_type, info->max_frmr_depth);
+			goto out;
+		}
+		smbdirect_mr->sgl = kcalloc(
+					info->max_frmr_depth,
+					sizeof(struct scatterlist),
+					GFP_KERNEL);
+		if (!smbdirect_mr->sgl) {
+			log_rdma_mr(ERR, "failed to allocate sgl\n");
+			ib_dereg_mr(smbdirect_mr->mr);
+			goto out;
+		}
+		smbdirect_mr->state = MR_READY;
+		smbdirect_mr->conn = info;
+
+		list_add_tail(&smbdirect_mr->list, &info->mr_list);
+		atomic_inc(&info->mr_ready_count);
+	}
+	INIT_WORK(&info->mr_recovery_work, smbd_mr_recovery_work);
+	return 0;
+
+out:
+	kfree(smbdirect_mr);
+
+	list_for_each_entry_safe(smbdirect_mr, tmp, &info->mr_list, list) {
+		ib_dereg_mr(smbdirect_mr->mr);
+		kfree(smbdirect_mr->sgl);
+		kfree(smbdirect_mr);
+	}
+	return -ENOMEM;
+}
+
+/*
+ * Get a MR from mr_list. This function waits until there is at least one
+ * MR available in the list. It may access the list while the
+ * smbd_mr_recovery_work is recovering the MR list. This doesn't need a lock
+ * as they never modify the same places. However, there may be several CPUs
+ * issueing I/O trying to get MR at the same time, mr_list_lock is used to
+ * protect this situation.
+ */
+static struct smbd_mr *get_mr(struct smbd_connection *info)
+{
+	struct smbd_mr *ret;
+	int rc;
+again:
+	rc = wait_event_interruptible(info->wait_mr,
+		atomic_read(&info->mr_ready_count) ||
+		info->transport_status != SMBD_CONNECTED);
+	if (rc) {
+		log_rdma_mr(ERR, "wait_event_interruptible rc=%x\n", rc);
+		return NULL;
+	}
+
+	if (info->transport_status != SMBD_CONNECTED) {
+		log_rdma_mr(ERR, "info->transport_status=%x\n",
+			info->transport_status);
+		return NULL;
+	}
+
+	spin_lock(&info->mr_list_lock);
+	list_for_each_entry(ret, &info->mr_list, list) {
+		if (ret->state == MR_READY) {
+			ret->state = MR_REGISTERED;
+			spin_unlock(&info->mr_list_lock);
+			atomic_dec(&info->mr_ready_count);
+			atomic_inc(&info->mr_used_count);
+			return ret;
+		}
+	}
+
+	spin_unlock(&info->mr_list_lock);
+	/*
+	 * It is possible that we can get a MR because other processes may try
+	 * to acquire a MR at the same time. If this is the case, retry it.
+	 */
+	goto again;
+}
+
+/*
+ * Register memory for RDMA read/write
+ * pages[]: the list of pages to register memory with
+ * num_pages: the number of pages to register
+ * tailsz: if non-zero, the bytes to register in the last page
+ * writing: true if this is a RDMA write (SMB read), false for RDMA read
+ * need_invalidate: true if this MR needs to be locally invalidated after I/O
+ * return value: the MR registered, NULL if failed.
+ */
+struct smbd_mr *smbd_register_mr(
+	struct smbd_connection *info, struct page *pages[], int num_pages,
+	int tailsz, bool writing, bool need_invalidate)
+{
+	struct smbd_mr *smbdirect_mr;
+	int rc, i;
+	enum dma_data_direction dir;
+	struct ib_reg_wr *reg_wr;
+	struct ib_send_wr *bad_wr;
+	unsigned long long t1 = rdtsc();
+
+	if (num_pages > info->max_frmr_depth) {
+		log_rdma_mr(ERR, "num_pages=%d max_frmr_depth=%d\n",
+			num_pages, info->max_frmr_depth);
+		return NULL;
+	}
+
+	smbdirect_mr = get_mr(info);
+	if (!smbdirect_mr) {
+		log_rdma_mr(ERR, "get_mr returning NULL\n");
+		return NULL;
+	}
+	smbdirect_mr->need_invalidate = need_invalidate;
+	smbdirect_mr->sgl_count = num_pages;
+	sg_init_table(smbdirect_mr->sgl, num_pages);
+
+	for (i = 0; i < num_pages - 1; i++)
+		sg_set_page(&smbdirect_mr->sgl[i], pages[i], PAGE_SIZE, 0);
+
+	sg_set_page(&smbdirect_mr->sgl[i], pages[i],
+		tailsz ? tailsz : PAGE_SIZE, 0);
+
+	dir = writing ? DMA_FROM_DEVICE : DMA_TO_DEVICE;
+	smbdirect_mr->dir = dir;
+	rc = ib_dma_map_sg(info->id->device, smbdirect_mr->sgl, num_pages, dir);
+	if (!rc) {
+		log_rdma_mr(INFO, "ib_dma_map_sg num_pages=%x dir=%x rc=%x\n",
+			num_pages, dir, rc);
+		goto dma_map_error;
+	}
+
+	rc = ib_map_mr_sg(smbdirect_mr->mr, smbdirect_mr->sgl, num_pages,
+		NULL, PAGE_SIZE);
+	if (rc != num_pages) {
+		log_rdma_mr(INFO,
+			"ib_map_mr_sg failed rc = %x num_pages = %x\n",
+			rc, num_pages);
+		goto map_mr_error;
+	}
+
+	ib_update_fast_reg_key(smbdirect_mr->mr,
+		ib_inc_rkey(smbdirect_mr->mr->rkey));
+	reg_wr = &smbdirect_mr->wr;
+	reg_wr->wr.opcode = IB_WR_REG_MR;
+	smbdirect_mr->cqe.done = register_mr_done;
+	reg_wr->wr.wr_cqe = &smbdirect_mr->cqe;
+	reg_wr->wr.num_sge = 0;
+	reg_wr->wr.send_flags = IB_SEND_SIGNALED;
+	reg_wr->mr = smbdirect_mr->mr;
+	reg_wr->key = smbdirect_mr->mr->rkey;
+	reg_wr->access = writing ?
+			IB_ACCESS_REMOTE_WRITE | IB_ACCESS_LOCAL_WRITE :
+			IB_ACCESS_REMOTE_READ;
+
+	/*
+	 * There is no need for waiting for complemtion on ib_post_send
+	 * on IB_WR_REG_MR. Hardware enforces a barrier and order of execution
+	 * on the next ib_post_send when we actaully send I/O to remote peer
+	 */
+	rc = ib_post_send(info->id->qp, &reg_wr->wr, &bad_wr);
+	if (!rc) {
+		profiling_add_histogram(
+			rdtsc()-t1, info->smbd_register_mr_cycles);
+		return smbdirect_mr;
+	}
+
+	log_rdma_mr(ERR, "ib_post_send failed rc=%x reg_wr->key=%x\n",
+		rc, reg_wr->key);
+
+	/* If all failed, attempt to recover this MR by setting it MR_ERROR*/
+map_mr_error:
+	ib_dma_unmap_sg(info->id->device, smbdirect_mr->sgl,
+		smbdirect_mr->sgl_count, smbdirect_mr->dir);
+
+dma_map_error:
+	smbdirect_mr->state = MR_ERROR;
+	if (atomic_dec_and_test(&info->mr_used_count))
+		wake_up(&info->wait_for_mr_cleanup);
+
+	return NULL;
+}
+
+static void local_inv_done(struct ib_cq *cq, struct ib_wc *wc)
+{
+	struct smbd_mr *smbdirect_mr;
+	struct ib_cqe *cqe;
+
+	cqe = wc->wr_cqe;
+	smbdirect_mr = container_of(cqe, struct smbd_mr, cqe);
+	smbdirect_mr->state = MR_INVALIDATED;
+	if (wc->status != IB_WC_SUCCESS) {
+		log_rdma_mr(ERR, "invalidate failed status=%x\n", wc->status);
+		smbdirect_mr->state = MR_ERROR;
+	}
+	complete(&smbdirect_mr->invalidate_done);
+}
+
+/*
+ * Deregister a MR after I/O is done
+ * This function may wait if remote invalidation is not used
+ * and we have to locally invalidate the buffer to prevent data is being
+ * modified by remote peer after upper layer consumes it
+ */
+int smbd_deregister_mr(struct smbd_mr *smbdirect_mr)
+{
+	struct ib_send_wr *wr, *bad_wr;
+	struct smbd_connection *info = smbdirect_mr->conn;
+	int rc = 0;
+	unsigned long long t1 = rdtsc();
+
+	if (smbdirect_mr->need_invalidate) {
+		/* Need to finish local invalidation before returning */
+		wr = &smbdirect_mr->inv_wr;
+		wr->opcode = IB_WR_LOCAL_INV;
+		smbdirect_mr->cqe.done = local_inv_done;
+		wr->wr_cqe = &smbdirect_mr->cqe;
+		wr->num_sge = 0;
+		wr->ex.invalidate_rkey = smbdirect_mr->mr->rkey;
+		wr->send_flags = IB_SEND_SIGNALED;
+
+		init_completion(&smbdirect_mr->invalidate_done);
+		rc = ib_post_send(info->id->qp, wr, &bad_wr);
+		if (rc) {
+			log_rdma_mr(ERR, "ib_post_send failed rc=%x\n", rc);
+			smbd_disconnect_rdma_connection(info);
+			goto done;
+		}
+		wait_for_completion(&smbdirect_mr->invalidate_done);
+		smbdirect_mr->need_invalidate = false;
+	} else
+		/*
+		 * For remote invalidation, just set it to MR_INVALIDATED
+		 * and defer to mr_recovery_work to recover the MR for next use
+		 */
+		smbdirect_mr->state = MR_INVALIDATED;
+
+	/*
+	 * Schedule the work to do MR recovery for future I/Os
+	 * MR recovery is slow and we don't want it to block the current I/O
+	 */
+	queue_work(info->workqueue, &info->mr_recovery_work);
+
+	profiling_add_histogram(rdtsc()-t1, info->smbd_deregister_mr_cycles);
+
+done:
+	if (atomic_dec_and_test(&info->mr_used_count))
+		wake_up(&info->wait_for_mr_cleanup);
+
+	return rc;
+}
diff --git a/fs/cifs/smbdirect.h b/fs/cifs/smbdirect.h
index e9bd938..b8ddfe9 100644
--- a/fs/cifs/smbdirect.h
+++ b/fs/cifs/smbdirect.h
@@ -87,6 +87,29 @@ struct smbd_connection {
 	int receive_credit_target;
 	int fragment_reassembly_remaining;
 
+	/* Memory registrations */
+	/* Maximum number of RDMA read/write outstanding on this connection */
+	int responder_resources;
+	/* Maximum number of SGEs in a RDMA write/read */
+	int max_frmr_depth;
+	/*
+	 * If payload is less than or equal to the threshold,
+	 * use RDMA send/recv to send upper layer I/O.
+	 * If payload is more than the threshold,
+	 * use RDMA read/write through memory registration for I/O.
+	 */
+	int rdma_readwrite_threshold;
+	enum ib_mr_type mr_type;
+	struct list_head mr_list;
+	spinlock_t mr_list_lock;
+	/* The number of available MRs ready for memory registration */
+	atomic_t mr_ready_count;
+	atomic_t mr_used_count;
+	wait_queue_head_t wait_mr;
+	struct work_struct mr_recovery_work;
+	/* Used by transport to wait until all MRs are returned */
+	wait_queue_head_t wait_for_mr_cleanup;
+
 	/* Activity accoutning */
 	/* Pending reqeusts issued from upper layer */
 	int smbd_send_pending;
@@ -157,6 +180,8 @@ struct smbd_connection {
 	unsigned int count_send_empty;
 
 #define SIZE_LONG_LONG	64
+	unsigned long long smbd_register_mr_cycles[SIZE_LONG_LONG];
+	unsigned long long smbd_deregister_mr_cycles[SIZE_LONG_LONG];
 	unsigned long long smbd_write_cycles[SIZE_LONG_LONG];
 	unsigned long long smbd_recv_cycles[SIZE_LONG_LONG];
 	unsigned long long recv_done_cycles[SIZE_LONG_LONG];
@@ -265,6 +290,36 @@ void smbd_destroy(struct smbd_connection *info);
 int smbd_recv(struct smbd_connection *info, struct msghdr *msg);
 int smbd_send(struct smbd_connection *info, struct smb_rqst *rqst);
 
+enum mr_state {
+	MR_READY,
+	MR_REGISTERED,
+	MR_INVALIDATED,
+	MR_ERROR
+};
+
+struct smbd_mr {
+	struct smbd_connection	*conn;
+	struct list_head	list;
+	enum mr_state		state;
+	struct ib_mr		*mr;
+	struct scatterlist	*sgl;
+	int			sgl_count;
+	enum dma_data_direction	dir;
+	union {
+		struct ib_reg_wr	wr;
+		struct ib_send_wr	inv_wr;
+	};
+	struct ib_cqe		cqe;
+	bool			need_invalidate;
+	struct completion	invalidate_done;
+};
+
+/* Interfaces to register and deregister MR for RDMA read/write */
+struct smbd_mr *smbd_register_mr(
+	struct smbd_connection *info, struct page *pages[], int num_pages,
+	int tailsz, bool writing, bool need_invalidate);
+int smbd_deregister_mr(struct smbd_mr *mr);
+
 void profiling_display_histogram(
 	struct seq_file *m, unsigned long long array[]);
 #endif
-- 
2.7.4

[toc] | [prev] | [standalone]

Back to top | Article view | linux.kernel

csiph-web