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

Groups > linux.kernel > #1386349 > unrolled thread

[PATCH 00/30] DBRD updates

Back to article view | Back to linux.kernel

Contents

  [PATCH 00/30] DBRD updates Philipp Reisner <philipp.reisner@linbit.com> - 2016-04-25 14:30 +0200
    [PATCH 04/30] drbd: Implement handling of thinly provisioned storage on resync target nodes Philipp Reisner <philipp.reisner@linbit.com> - 2016-04-25 14:30 +0200
      Re: [Drbd-dev] [PATCH 04/30] drbd: Implement handling of thinly provisioned storage on resync target nodes Philipp Reisner <philipp.reisner@linbit.com> - 2016-04-25 18:50 +0200
    [PATCH 15/30] drbd: finish resync on sync source only by notification from sync target Philipp Reisner <philipp.reisner@linbit.com> - 2016-04-25 14:30 +0200
    [PATCH 17/30] drbd: don't forget error completion when "unsuspending" IO Philipp Reisner <philipp.reisner@linbit.com> - 2016-04-25 14:30 +0200
    [PATCH 05/30] drbd: Introduce new disk config option rs-discard-granularity Philipp Reisner <philipp.reisner@linbit.com> - 2016-04-25 14:30 +0200
      Re: [Drbd-dev] [PATCH 05/30] drbd: Introduce new disk config option rs-discard-granularity Philipp Reisner <philipp.reisner@linbit.com> - 2016-04-25 18:50 +0200
        Re: [Drbd-dev] [PATCH 05/30] drbd: Introduce new disk config option rs-discard-granularity Philipp Reisner <philipp.reisner@linbit.com> - 2016-04-25 21:50 +0200
          [PATCH 05/30] drbd: Introduce new disk config option  rs-discard-granularity Lars Ellenberg <lars.ellenberg@linbit.com> - 2016-04-25 22:40 +0200
    [PATCH 28/30] drbd: finally report ms, not jiffies, in log message Philipp Reisner <philipp.reisner@linbit.com> - 2016-04-25 14:30 +0200
    [PATCH 11/30] drbd: when receiving P_TRIM, zero-out partial unaligned chunks Philipp Reisner <philipp.reisner@linbit.com> - 2016-04-25 14:30 +0200
      [PATCH 11/30] drbd: when receiving P_TRIM, zero-out partial  unaligned chunks Lars Ellenberg <lars.ellenberg@linbit.com> - 2016-04-25 22:40 +0200
        Re: [PATCH 11/30] drbd: when receiving P_TRIM, zero-out partial  unaligned chunks Lars Ellenberg <lars.ellenberg@linbit.com> - 2016-04-25 23:30 +0200
    [PATCH 24/30] drbd: disallow promotion during resync handshake, avoid deadlock and hard reset Philipp Reisner <philipp.reisner@linbit.com> - 2016-04-25 14:30 +0200
    [PATCH 29/30] drbd: al_write_transaction: skip re-scanning of bitmap page pointer array Philipp Reisner <philipp.reisner@linbit.com> - 2016-04-25 14:30 +0200
    [PATCH 10/30] drbd: allow parallel flushes for multi-volume resources Philipp Reisner <philipp.reisner@linbit.com> - 2016-04-25 14:30 +0200

#1386349 — [PATCH 00/30] DBRD updates

Hi Jens,

apart from the usual maintenance and bug fixes this time comes
support for WRITE_SAME and lots of improvemnts for DISCARD.

Overview:
As replication technology we want to use DISCARDs only if they
really zero the backing storage. Thin LVM does it but claims
not to do it. To make that reasonably usable with DRBD we
added this "discard_zeroes_if_aligned" hack^H^H^H^H configure
option. Please see the commit messages for all the details.

Please add it to your for-4.7/drivers branch.
Thanks!


Fabian Frederick (1):
  drbd: code cleanups without semantic changes

Lars Ellenberg (24):
  drbd: bitmap bulk IO: do not always suspend IO
  drbd: change bitmap write-out when leaving resync states
  drbd: adjust assert in w_bitmap_io to account for
    BM_LOCKED_CHANGE_ALLOWED
  drbd: fix regression: protocol A sometimes synchronous, C sometimes
    double-latency
  drbd: fix for truncated minor number in callback command line
  drbd: allow parallel flushes for multi-volume resources
  drbd: when receiving P_TRIM, zero-out partial unaligned chunks
  drbd: possibly disable discard support, if backend has
    discard_zeroes_data=0
  drbd: zero-out partial unaligned discards on local backend
  drbd: allow larger max_discard_sectors
  drbd: finish resync on sync source only by notification from sync
    target
  drbd: introduce unfence-peer handler
  drbd: don't forget error completion when "unsuspending" IO
  drbd: if there is no good data accessible, writes should be IO errors
  drbd: only restart frozen disk io when D_UP_TO_DATE
  drbd: discard_zeroes_if_aligned allows "thin" resync for
    discard_zeroes_data=0
  drbd: report sizes if rejecting too small peer disk
  drbd: introduce WRITE_SAME support
  drbd: sync_handshake: handle identical uuids with current (frozen)
    Primary
  drbd: disallow promotion during resync handshake, avoid deadlock and
    hard reset
  drbd: bump current uuid when resuming IO with diskless peer
  drbd: finally report ms, not jiffies, in log message
  drbd: al_write_transaction: skip re-scanning of bitmap page pointer
    array
  drbd: correctly handle failed crypto_alloc_hash

Philipp Reisner (4):
  drbd: Kill code duplication
  drbd: Implement handling of thinly provisioned storage on resync
    target nodes
  drbd: Introduce new disk config option rs-discard-granularity
  drbd: Create the protocol feature THIN_RESYNC

Roland Kammerer (1):
  drbd: get rid of empty statement in is_valid_state

 drivers/block/drbd/drbd_actlog.c   |  29 +-
 drivers/block/drbd/drbd_bitmap.c   |  84 ++++--
 drivers/block/drbd/drbd_debugfs.c  |  13 +-
 drivers/block/drbd/drbd_int.h      |  49 +++-
 drivers/block/drbd/drbd_interval.h |  14 +-
 drivers/block/drbd/drbd_main.c     | 115 +++++++-
 drivers/block/drbd/drbd_nl.c       | 282 +++++++++++++++-----
 drivers/block/drbd/drbd_proc.c     |  30 +--
 drivers/block/drbd/drbd_protocol.h |  77 +++++-
 drivers/block/drbd/drbd_receiver.c | 534 ++++++++++++++++++++++++++++++-------
 drivers/block/drbd/drbd_req.c      |  84 ++++--
 drivers/block/drbd/drbd_req.h      |   5 +-
 drivers/block/drbd/drbd_state.c    |  61 ++++-
 drivers/block/drbd/drbd_state.h    |   2 +-
 drivers/block/drbd/drbd_strings.c  |   8 +-
 drivers/block/drbd/drbd_worker.c   |  85 +++++-
 include/linux/drbd.h               |  10 +-
 include/linux/drbd_genl.h          |   7 +-
 include/linux/drbd_limits.h        |  15 +-
 19 files changed, 1205 insertions(+), 299 deletions(-)

-- 
1.9.1

[toc] | [next] | [standalone]

#1386350 — [PATCH 04/30] drbd: Implement handling of thinly provisioned storage on resync target nodes

If during resync we read only zeroes for a range of sectors assume
that these secotors can be discarded on the sync target node.

Signed-off-by: Philipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: Lars Ellenberg <lars.ellenberg@linbit.com>
---
 drivers/block/drbd/drbd_int.h      |  5 +++
 drivers/block/drbd/drbd_main.c     | 18 ++++++++
 drivers/block/drbd/drbd_protocol.h |  4 ++
 drivers/block/drbd/drbd_receiver.c | 88 ++++++++++++++++++++++++++++++++++++--
 drivers/block/drbd/drbd_worker.c   | 29 ++++++++++++-
 5 files changed, 140 insertions(+), 4 deletions(-)

diff --git a/drivers/block/drbd/drbd_int.h b/drivers/block/drbd/drbd_int.h
index 7a1cf7e..1a93f4f 100644
--- a/drivers/block/drbd/drbd_int.h
+++ b/drivers/block/drbd/drbd_int.h
@@ -471,6 +471,9 @@ enum {
 	/* this originates from application on peer
 	 * (not some resync or verify or other DRBD internal request) */
 	__EE_APPLICATION,
+
+	/* If it contains only 0 bytes, send back P_RS_DEALLOCATED */
+	__EE_RS_THIN_REQ,
 };
 #define EE_CALL_AL_COMPLETE_IO (1<<__EE_CALL_AL_COMPLETE_IO)
 #define EE_MAY_SET_IN_SYNC     (1<<__EE_MAY_SET_IN_SYNC)
@@ -485,6 +488,7 @@ enum {
 #define EE_SUBMITTED		(1<<__EE_SUBMITTED)
 #define EE_WRITE		(1<<__EE_WRITE)
 #define EE_APPLICATION		(1<<__EE_APPLICATION)
+#define EE_RS_THIN_REQ		(1<<__EE_RS_THIN_REQ)
 
 /* flag bits per device */
 enum {
@@ -1123,6 +1127,7 @@ extern int drbd_send_ov_request(struct drbd_peer_device *, sector_t sector, int
 extern int drbd_send_bitmap(struct drbd_device *device);
 extern void drbd_send_sr_reply(struct drbd_peer_device *, enum drbd_state_rv retcode);
 extern void conn_send_sr_reply(struct drbd_connection *connection, enum drbd_state_rv retcode);
+extern int drbd_send_rs_deallocated(struct drbd_peer_device *, struct drbd_peer_request *);
 extern void drbd_backing_dev_free(struct drbd_device *device, struct drbd_backing_dev *ldev);
 extern void drbd_device_cleanup(struct drbd_device *device);
 void drbd_print_uuids(struct drbd_device *device, const char *text);
diff --git a/drivers/block/drbd/drbd_main.c b/drivers/block/drbd/drbd_main.c
index 802d729..3cecc4f 100644
--- a/drivers/block/drbd/drbd_main.c
+++ b/drivers/block/drbd/drbd_main.c
@@ -1377,6 +1377,22 @@ int drbd_send_ack_ex(struct drbd_peer_device *peer_device, enum drbd_packet cmd,
 			      cpu_to_be64(block_id));
 }
 
+int drbd_send_rs_deallocated(struct drbd_peer_device *peer_device,
+			     struct drbd_peer_request *peer_req)
+{
+	struct drbd_socket *sock;
+	struct p_block_desc *p;
+
+	sock = &peer_device->connection->data;
+	p = drbd_prepare_command(peer_device, sock);
+	if (!p)
+		return -EIO;
+	p->sector = cpu_to_be64(peer_req->i.sector);
+	p->blksize = cpu_to_be32(peer_req->i.size);
+	p->pad = 0;
+	return drbd_send_command(peer_device, sock, P_RS_DEALLOCATED, sizeof(*p), NULL, 0);
+}
+
 int drbd_send_drequest(struct drbd_peer_device *peer_device, int cmd,
 		       sector_t sector, int size, u64 block_id)
 {
@@ -3681,6 +3697,8 @@ const char *cmdname(enum drbd_packet cmd)
 		[P_CONN_ST_CHG_REPLY]	= "conn_st_chg_reply",
 		[P_RETRY_WRITE]		= "retry_write",
 		[P_PROTOCOL_UPDATE]	= "protocol_update",
+		[P_RS_THIN_REQ]         = "rs_thin_req",
+		[P_RS_DEALLOCATED]      = "rs_deallocated",
 
 		/* enum drbd_packet, but not commands - obsoleted flags:
 		 *	P_MAY_IGNORE
diff --git a/drivers/block/drbd/drbd_protocol.h b/drivers/block/drbd/drbd_protocol.h
index ef92453..e5e74e3 100644
--- a/drivers/block/drbd/drbd_protocol.h
+++ b/drivers/block/drbd/drbd_protocol.h
@@ -60,6 +60,10 @@ enum drbd_packet {
 	 * which is why I chose TRIM here, to disambiguate. */
 	P_TRIM                = 0x31,
 
+	/* Only use these two if both support FF_THIN_RESYNC */
+	P_RS_THIN_REQ         = 0x32, /* Request a block for resync or reply P_RS_DEALLOCATED */
+	P_RS_DEALLOCATED      = 0x33, /* Contains only zeros on sync source node */
+
 	P_MAY_IGNORE	      = 0x100, /* Flag to test if (cmd > P_MAY_IGNORE) ... */
 	P_MAX_OPT_CMD	      = 0x101,
 
diff --git a/drivers/block/drbd/drbd_receiver.c b/drivers/block/drbd/drbd_receiver.c
index 8b30ab5..3a6c2ec 100644
--- a/drivers/block/drbd/drbd_receiver.c
+++ b/drivers/block/drbd/drbd_receiver.c
@@ -1417,9 +1417,15 @@ int drbd_submit_peer_request(struct drbd_device *device,
 		 * so we can find it to present it in debugfs */
 		peer_req->submit_jif = jiffies;
 		peer_req->flags |= EE_SUBMITTED;
-		spin_lock_irq(&device->resource->req_lock);
-		list_add_tail(&peer_req->w.list, &device->active_ee);
-		spin_unlock_irq(&device->resource->req_lock);
+
+		/* If this was a resync request from receive_rs_deallocated(),
+		 * it is already on the sync_ee list */
+		if (list_empty(&peer_req->w.list)) {
+			spin_lock_irq(&device->resource->req_lock);
+			list_add_tail(&peer_req->w.list, &device->active_ee);
+			spin_unlock_irq(&device->resource->req_lock);
+		}
+
 		if (blkdev_issue_zeroout(device->ldev->backing_bdev,
 			sector, data_size >> 9, GFP_NOIO, false))
 			peer_req->flags |= EE_WAS_ERROR;
@@ -2574,6 +2580,7 @@ static int receive_DataRequest(struct drbd_connection *connection, struct packet
 		case P_DATA_REQUEST:
 			drbd_send_ack_rp(peer_device, P_NEG_DREPLY, p);
 			break;
+		case P_RS_THIN_REQ:
 		case P_RS_DATA_REQUEST:
 		case P_CSUM_RS_REQUEST:
 		case P_OV_REQUEST:
@@ -2613,6 +2620,12 @@ static int receive_DataRequest(struct drbd_connection *connection, struct packet
 		peer_req->flags |= EE_APPLICATION;
 		goto submit;
 
+	case P_RS_THIN_REQ:
+		/* If at some point in the future we have a smart way to
+		   find out if this data block is completely deallocated,
+		   then we would do something smarter here than reading
+		   the block... */
+		peer_req->flags |= EE_RS_THIN_REQ;
 	case P_RS_DATA_REQUEST:
 		peer_req->w.cb = w_e_end_rsdata_req;
 		fault_type = DRBD_FAULT_RS_RD;
@@ -4587,6 +4600,72 @@ static int receive_out_of_sync(struct drbd_connection *connection, struct packet
 	return 0;
 }
 
+static int receive_rs_deallocated(struct drbd_connection *connection, struct packet_info *pi)
+{
+	struct drbd_peer_device *peer_device;
+	struct p_block_desc *p = pi->data;
+	struct drbd_device *device;
+	sector_t sector;
+	int size, err = 0;
+
+	peer_device = conn_peer_device(connection, pi->vnr);
+	if (!peer_device)
+		return -EIO;
+	device = peer_device->device;
+
+	sector = be64_to_cpu(p->sector);
+	size = be32_to_cpu(p->blksize);
+
+	dec_rs_pending(device);
+
+	if (get_ldev(device)) {
+		struct drbd_peer_request *peer_req;
+		const int rw = REQ_WRITE | REQ_DISCARD;
+
+		peer_req = drbd_alloc_peer_req(peer_device, ID_SYNCER, sector,
+					       size, false, GFP_NOIO);
+		if (!peer_req) {
+			put_ldev(device);
+			return -ENOMEM;
+		}
+
+		peer_req->w.cb = e_end_resync_block;
+		peer_req->submit_jif = jiffies;
+		peer_req->flags |= EE_IS_TRIM;
+
+		spin_lock_irq(&device->resource->req_lock);
+		list_add_tail(&peer_req->w.list, &device->sync_ee);
+		spin_unlock_irq(&device->resource->req_lock);
+
+		atomic_add(pi->size >> 9, &device->rs_sect_ev);
+		err = drbd_submit_peer_request(device, peer_req, rw, DRBD_FAULT_RS_WR);
+
+		if (err) {
+			spin_lock_irq(&device->resource->req_lock);
+			list_del(&peer_req->w.list);
+			spin_unlock_irq(&device->resource->req_lock);
+
+			drbd_free_peer_req(device, peer_req);
+			put_ldev(device);
+			err = 0;
+			goto fail;
+		}
+
+		inc_unacked(device);
+
+		/* No put_ldev() here. Gets called in drbd_endio_write_sec_final(),
+		   as well as drbd_rs_complete_io() */
+	} else {
+	fail:
+		drbd_rs_complete_io(device, sector);
+		drbd_send_ack_ex(peer_device, P_NEG_ACK, sector, size, ID_SYNCER);
+	}
+
+	atomic_add(size >> 9, &device->rs_sect_in);
+
+	return err;
+}
+
 struct data_cmd {
 	int expect_payload;
 	size_t pkt_size;
@@ -4614,11 +4693,14 @@ static struct data_cmd drbd_cmd_handler[] = {
 	[P_OV_REQUEST]      = { 0, sizeof(struct p_block_req), receive_DataRequest },
 	[P_OV_REPLY]        = { 1, sizeof(struct p_block_req), receive_DataRequest },
 	[P_CSUM_RS_REQUEST] = { 1, sizeof(struct p_block_req), receive_DataRequest },
+	[P_RS_THIN_REQ]     = { 0, sizeof(struct p_block_req), receive_DataRequest },
 	[P_DELAY_PROBE]     = { 0, sizeof(struct p_delay_probe93), receive_skip },
 	[P_OUT_OF_SYNC]     = { 0, sizeof(struct p_block_desc), receive_out_of_sync },
 	[P_CONN_ST_CHG_REQ] = { 0, sizeof(struct p_req_state), receive_req_conn_state },
 	[P_PROTOCOL_UPDATE] = { 1, sizeof(struct p_protocol), receive_protocol },
 	[P_TRIM]	    = { 0, sizeof(struct p_trim), receive_Data },
+	[P_RS_DEALLOCATED]  = { 0, sizeof(struct p_block_desc), receive_rs_deallocated },
+
 };
 
 static void drbdd(struct drbd_connection *connection)
diff --git a/drivers/block/drbd/drbd_worker.c b/drivers/block/drbd/drbd_worker.c
index 4d87499..01d74ee2 100644
--- a/drivers/block/drbd/drbd_worker.c
+++ b/drivers/block/drbd/drbd_worker.c
@@ -1035,6 +1035,30 @@ int w_e_end_data_req(struct drbd_work *w, int cancel)
 	return err;
 }
 
+static bool all_zero(struct drbd_peer_request *peer_req)
+{
+	struct page *page = peer_req->pages;
+	unsigned int len = peer_req->i.size;
+
+	page_chain_for_each(page) {
+		unsigned int l = min_t(unsigned int, len, PAGE_SIZE);
+		unsigned int i, words = l / sizeof(long);
+		unsigned long *d;
+
+		d = kmap_atomic(page);
+		for (i = 0; i < words; i++) {
+			if (d[i]) {
+				kunmap_atomic(d);
+				return false;
+			}
+		}
+		kunmap_atomic(d);
+		len -= l;
+	}
+
+	return true;
+}
+
 /**
  * w_e_end_rsdata_req() - Worker callback to send a P_RS_DATA_REPLY packet in response to a P_RS_DATA_REQUEST
  * @w:		work object.
@@ -1063,7 +1087,10 @@ int w_e_end_rsdata_req(struct drbd_work *w, int cancel)
 	} else if (likely((peer_req->flags & EE_WAS_ERROR) == 0)) {
 		if (likely(device->state.pdsk >= D_INCONSISTENT)) {
 			inc_rs_pending(device);
-			err = drbd_send_block(peer_device, P_RS_DATA_REPLY, peer_req);
+			if (peer_req->flags & EE_RS_THIN_REQ && all_zero(peer_req))
+				err = drbd_send_rs_deallocated(peer_device, peer_req);
+			else
+				err = drbd_send_block(peer_device, P_RS_DATA_REPLY, peer_req);
 		} else {
 			if (__ratelimit(&drbd_ratelimit_state))
 				drbd_err(device, "Not sending RSDataReply, "
-- 
1.9.1

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

#1386664 — Re: [Drbd-dev] [PATCH 04/30] drbd: Implement handling of thinly provisioned storage on resync target nodes

Am Montag, 25. April 2016, 08:28:45 schrieb Bart Van Assche:
> On 04/25/2016 05:10 AM, Philipp Reisner wrote:
> > If during resync we read only zeroes for a range of sectors assume
> > that these secotors can be discarded on the sync target node.
> 
> Hello Phil,
> 
> With which interconnect(s) has this patch been tested? I'm afraid that
> for high-speed interconnects this patch will slow down I/O instead of
> making it faster because all_zero() examines all data before it is sent.
> 

Hi Bart,

that it might make things slower is true for sure. The benefit it
provides is to de-allocate blocks on the secondary obviously.
The whole feature is optional and it is off by default.

Obviously we want to have a generic interface like SEEK_HOLE/
SEEK_DATA for block devices, but that does not exist as of today.

best regards,
 Phil

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

#1386351 — [PATCH 15/30] drbd: finish resync on sync source only by notification from sync target

From: Lars Ellenberg <lars@linbit.com>

If the replication link breaks exactly during "resync finished" detection,
finishing too early on the sync source could again lead to UUIDs rotated
too fast, and potentially a spurious full resync on next handshake.

Always wait for explicit resync finished state change notification from
the sync target.

Signed-off-by: Philipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: Lars Ellenberg <lars.ellenberg@linbit.com>
---
 drivers/block/drbd/drbd_actlog.c | 16 ++++++++++++----
 drivers/block/drbd/drbd_int.h    | 19 ++++++++++++++-----
 2 files changed, 26 insertions(+), 9 deletions(-)

diff --git a/drivers/block/drbd/drbd_actlog.c b/drivers/block/drbd/drbd_actlog.c
index 1664762..4e07cff 100644
--- a/drivers/block/drbd/drbd_actlog.c
+++ b/drivers/block/drbd/drbd_actlog.c
@@ -768,10 +768,18 @@ static bool lazy_bitmap_update_due(struct drbd_device *device)
 
 static void maybe_schedule_on_disk_bitmap_update(struct drbd_device *device, bool rs_done)
 {
-	if (rs_done)
-		set_bit(RS_DONE, &device->flags);
-		/* and also set RS_PROGRESS below */
-	else if (!lazy_bitmap_update_due(device))
+	if (rs_done) {
+		struct drbd_connection *connection = first_peer_device(device)->connection;
+		if (connection->agreed_pro_version <= 95 ||
+		    is_sync_target_state(device->state.conn))
+			set_bit(RS_DONE, &device->flags);
+			/* and also set RS_PROGRESS below */
+
+		/* Else: rather wait for explicit notification via receive_state,
+		 * to avoid uuids-rotated-too-fast causing full resync
+		 * in next handshake, in case the replication link breaks
+		 * at the most unfortunate time... */
+	} else if (!lazy_bitmap_update_due(device))
 		return;
 
 	drbd_device_post_work(device, RS_PROGRESS);
diff --git a/drivers/block/drbd/drbd_int.h b/drivers/block/drbd/drbd_int.h
index d82e531..451a745 100644
--- a/drivers/block/drbd/drbd_int.h
+++ b/drivers/block/drbd/drbd_int.h
@@ -2102,13 +2102,22 @@ static inline void _sub_unacked(struct drbd_device *device, int n, const char *f
 	ERR_IF_CNT_IS_NEGATIVE(unacked_cnt, func, line);
 }
 
+static inline bool is_sync_target_state(enum drbd_conns connection_state)
+{
+	return	connection_state == C_SYNC_TARGET ||
+		connection_state == C_PAUSED_SYNC_T;
+}
+
+static inline bool is_sync_source_state(enum drbd_conns connection_state)
+{
+	return	connection_state == C_SYNC_SOURCE ||
+		connection_state == C_PAUSED_SYNC_S;
+}
+
 static inline bool is_sync_state(enum drbd_conns connection_state)
 {
-	return
-	   (connection_state == C_SYNC_SOURCE
-	||  connection_state == C_SYNC_TARGET
-	||  connection_state == C_PAUSED_SYNC_S
-	||  connection_state == C_PAUSED_SYNC_T);
+	return	is_sync_source_state(connection_state) ||
+		is_sync_target_state(connection_state);
 }
 
 /**
-- 
1.9.1

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

#1386353 — [PATCH 17/30] drbd: don't forget error completion when "unsuspending" IO

From: Lars Ellenberg <lars.ellenberg@linbit.com>

Possibly sequence of events:
SyncTarget is made Primary, then loses replication link
(only path to good data on SyncSource).

Behavior is then controlled by the on-no-data-accessible policy,
which defaults to OND_IO_ERROR (may be set to OND_SUSPEND_IO).

If OND_IO_ERROR is in fact the current policy, we clear the susp_fen
(IO suspended due to fencing policy) flag, do NOT set the susp_nod
(IO suspended due to no data) flag.

But we forgot to call the IO error completion for all pending,
suspended, requests.

While at it, also add a race check for a theoretically possible
race with a new handshake (network hickup), we may be able to
re-send requests, and can avoid passing IO errors up the stack.

Signed-off-by: Philipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: Lars Ellenberg <lars.ellenberg@linbit.com>
---
 drivers/block/drbd/drbd_nl.c | 48 +++++++++++++++++++++++++++++---------------
 1 file changed, 32 insertions(+), 16 deletions(-)

diff --git a/drivers/block/drbd/drbd_nl.c b/drivers/block/drbd/drbd_nl.c
index f16084a..a703a0e 100644
--- a/drivers/block/drbd/drbd_nl.c
+++ b/drivers/block/drbd/drbd_nl.c
@@ -442,19 +442,17 @@ static enum drbd_fencing_p highest_fencing_policy(struct drbd_connection *connec
 	}
 	rcu_read_unlock();
 
-	if (fp == FP_NOT_AVAIL) {
-		/* IO Suspending works on the whole resource.
-		   Do it only for one device. */
-		vnr = 0;
-		peer_device = idr_get_next(&connection->peer_devices, &vnr);
-		drbd_change_state(peer_device->device, CS_VERBOSE | CS_HARD, NS(susp_fen, 0));
-	}
-
 	return fp;
 }
 
+static bool resource_is_supended(struct drbd_resource *resource)
+{
+	return resource->susp || resource->susp_fen || resource->susp_nod;
+}
+
 bool conn_try_outdate_peer(struct drbd_connection *connection)
 {
+	struct drbd_resource * const resource = connection->resource;
 	unsigned int connect_cnt;
 	union drbd_state mask = { };
 	union drbd_state val = { };
@@ -462,21 +460,41 @@ bool conn_try_outdate_peer(struct drbd_connection *connection)
 	char *ex_to_string;
 	int r;
 
-	spin_lock_irq(&connection->resource->req_lock);
+	spin_lock_irq(&resource->req_lock);
 	if (connection->cstate >= C_WF_REPORT_PARAMS) {
 		drbd_err(connection, "Expected cstate < C_WF_REPORT_PARAMS\n");
-		spin_unlock_irq(&connection->resource->req_lock);
+		spin_unlock_irq(&resource->req_lock);
 		return false;
 	}
 
 	connect_cnt = connection->connect_cnt;
-	spin_unlock_irq(&connection->resource->req_lock);
+	spin_unlock_irq(&resource->req_lock);
 
 	fp = highest_fencing_policy(connection);
 	switch (fp) {
 	case FP_NOT_AVAIL:
 		drbd_warn(connection, "Not fencing peer, I'm not even Consistent myself.\n");
-		goto out;
+		spin_lock_irq(&resource->req_lock);
+		if (connection->cstate < C_WF_REPORT_PARAMS) {
+			_conn_request_state(connection,
+					    (union drbd_state) { { .susp_fen = 1 } },
+					    (union drbd_state) { { .susp_fen = 0 } },
+					    CS_VERBOSE | CS_HARD | CS_DC_SUSP);
+			/* We are no longer suspended due to the fencing policy.
+			 * We may still be suspended due to the on-no-data-accessible policy.
+			 * If that was OND_IO_ERROR, fail pending requests. */
+			if (!resource_is_supended(resource))
+				_tl_restart(connection, CONNECTION_LOST_WHILE_PENDING);
+		}
+		/* Else: in case we raced with a connection handshake,
+		 * let the handshake figure out if we maybe can RESEND,
+		 * and do not resume/fail pending requests here.
+		 * Worst case is we stay suspended for now, which may be
+		 * resolved by either re-establishing the replication link, or
+		 * the next link failure, or eventually the administrator.  */
+		spin_unlock_irq(&resource->req_lock);
+		return false;
+
 	case FP_DONT_CARE:
 		return true;
 	default: ;
@@ -529,13 +547,11 @@ bool conn_try_outdate_peer(struct drbd_connection *connection)
 	drbd_info(connection, "fence-peer helper returned %d (%s)\n",
 		  (r>>8) & 0xff, ex_to_string);
 
- out:
-
 	/* Not using
 	   conn_request_state(connection, mask, val, CS_VERBOSE);
 	   here, because we might were able to re-establish the connection in the
 	   meantime. */
-	spin_lock_irq(&connection->resource->req_lock);
+	spin_lock_irq(&resource->req_lock);
 	if (connection->cstate < C_WF_REPORT_PARAMS && !test_bit(STATE_SENT, &connection->flags)) {
 		if (connection->connect_cnt != connect_cnt)
 			/* In case the connection was established and droped
@@ -544,7 +560,7 @@ bool conn_try_outdate_peer(struct drbd_connection *connection)
 		else
 			_conn_request_state(connection, mask, val, CS_VERBOSE);
 	}
-	spin_unlock_irq(&connection->resource->req_lock);
+	spin_unlock_irq(&resource->req_lock);
 
 	return conn_highest_pdsk(connection) <= D_OUTDATED;
 }
-- 
1.9.1

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

#1386354 — [PATCH 05/30] drbd: Introduce new disk config option rs-discard-granularity

As long as the value is 0 the feature is disabled. With setting
it to a positive value, DRBD limits and aligns its resync requests
to the rs-discard-granularity setting. If the sync source detects
all zeros in such a block, the resync target discards the range
on disk.

Signed-off-by: Philipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: Lars Ellenberg <lars.ellenberg@linbit.com>
---
 drivers/block/drbd/drbd_nl.c | 32 +++++++++++++++++++++++++++++---
 include/linux/drbd_genl.h    |  6 +++---
 include/linux/drbd_limits.h  |  6 ++++++
 3 files changed, 38 insertions(+), 6 deletions(-)

diff --git a/drivers/block/drbd/drbd_nl.c b/drivers/block/drbd/drbd_nl.c
index bb71e73..ee4642f 100644
--- a/drivers/block/drbd/drbd_nl.c
+++ b/drivers/block/drbd/drbd_nl.c
@@ -1348,12 +1348,38 @@ static bool write_ordering_changed(struct disk_conf *a, struct disk_conf *b)
 		a->disk_drain != b->disk_drain;
 }
 
-static void sanitize_disk_conf(struct disk_conf *disk_conf, struct drbd_backing_dev *nbc)
+static void sanitize_disk_conf(struct drbd_device *device, struct disk_conf *disk_conf,
+			       struct drbd_backing_dev *nbc)
 {
+	struct request_queue * const q = nbc->backing_bdev->bd_disk->queue;
+
 	if (disk_conf->al_extents < DRBD_AL_EXTENTS_MIN)
 		disk_conf->al_extents = DRBD_AL_EXTENTS_MIN;
 	if (disk_conf->al_extents > drbd_al_extents_max(nbc))
 		disk_conf->al_extents = drbd_al_extents_max(nbc);
+
+	if (!blk_queue_discard(q) || !q->limits.discard_zeroes_data) {
+		disk_conf->rs_discard_granularity = 0; /* disable feature */
+		drbd_info(device, "rs_discard_granularity feature disabled\n");
+	}
+
+	if (disk_conf->rs_discard_granularity) {
+		int orig_value = disk_conf->rs_discard_granularity;
+		int remainder;
+
+		if (q->limits.discard_granularity > disk_conf->rs_discard_granularity)
+			disk_conf->rs_discard_granularity = q->limits.discard_granularity;
+
+		remainder = disk_conf->rs_discard_granularity % q->limits.discard_granularity;
+		disk_conf->rs_discard_granularity += remainder;
+
+		if (disk_conf->rs_discard_granularity > q->limits.max_discard_sectors << 9)
+			disk_conf->rs_discard_granularity = q->limits.max_discard_sectors << 9;
+
+		if (disk_conf->rs_discard_granularity != orig_value)
+			drbd_info(device, "rs_discard_granularity changed to %d\n",
+				  disk_conf->rs_discard_granularity);
+	}
 }
 
 int drbd_adm_disk_opts(struct sk_buff *skb, struct genl_info *info)
@@ -1403,7 +1429,7 @@ int drbd_adm_disk_opts(struct sk_buff *skb, struct genl_info *info)
 	if (!expect(new_disk_conf->resync_rate >= 1))
 		new_disk_conf->resync_rate = 1;
 
-	sanitize_disk_conf(new_disk_conf, device->ldev);
+	sanitize_disk_conf(device, new_disk_conf, device->ldev);
 
 	if (new_disk_conf->c_plan_ahead > DRBD_C_PLAN_AHEAD_MAX)
 		new_disk_conf->c_plan_ahead = DRBD_C_PLAN_AHEAD_MAX;
@@ -1698,7 +1724,7 @@ int drbd_adm_attach(struct sk_buff *skb, struct genl_info *info)
 	if (retcode != NO_ERROR)
 		goto fail;
 
-	sanitize_disk_conf(new_disk_conf, nbc);
+	sanitize_disk_conf(device, new_disk_conf, nbc);
 
 	if (drbd_get_max_capacity(nbc) < new_disk_conf->disk_size) {
 		drbd_err(device, "max capacity %llu smaller than disk size %llu\n",
diff --git a/include/linux/drbd_genl.h b/include/linux/drbd_genl.h
index 2d0e5ad..ab649d8 100644
--- a/include/linux/drbd_genl.h
+++ b/include/linux/drbd_genl.h
@@ -123,14 +123,14 @@ GENL_struct(DRBD_NLA_DISK_CONF, 3, disk_conf,
 	__u32_field_def(13,	DRBD_GENLA_F_MANDATORY,	c_fill_target, DRBD_C_FILL_TARGET_DEF)
 	__u32_field_def(14,	DRBD_GENLA_F_MANDATORY,	c_max_rate, DRBD_C_MAX_RATE_DEF)
 	__u32_field_def(15,	DRBD_GENLA_F_MANDATORY,	c_min_rate, DRBD_C_MIN_RATE_DEF)
+	__u32_field_def(20,     DRBD_GENLA_F_MANDATORY, disk_timeout, DRBD_DISK_TIMEOUT_DEF)
+	__u32_field_def(21,     0 /* OPTIONAL */,       read_balancing, DRBD_READ_BALANCING_DEF)
+	__u32_field_def(25,     0 /* OPTIONAL */,       rs_discard_granularity, DRBD_RS_DISCARD_GRANULARITY_DEF)
 
 	__flg_field_def(16, DRBD_GENLA_F_MANDATORY,	disk_barrier, DRBD_DISK_BARRIER_DEF)
 	__flg_field_def(17, DRBD_GENLA_F_MANDATORY,	disk_flushes, DRBD_DISK_FLUSHES_DEF)
 	__flg_field_def(18, DRBD_GENLA_F_MANDATORY,	disk_drain, DRBD_DISK_DRAIN_DEF)
 	__flg_field_def(19, DRBD_GENLA_F_MANDATORY,	md_flushes, DRBD_MD_FLUSHES_DEF)
-	__u32_field_def(20,	DRBD_GENLA_F_MANDATORY,	disk_timeout, DRBD_DISK_TIMEOUT_DEF)
-	__u32_field_def(21,	0 /* OPTIONAL */,       read_balancing, DRBD_READ_BALANCING_DEF)
-	/* 9: __u32_field_def(22,	DRBD_GENLA_F_MANDATORY,	unplug_watermark, DRBD_UNPLUG_WATERMARK_DEF) */
 	__flg_field_def(23,     0 /* OPTIONAL */,	al_updates, DRBD_AL_UPDATES_DEF)
 )
 
diff --git a/include/linux/drbd_limits.h b/include/linux/drbd_limits.h
index 8ac8c5d..2e4aad8 100644
--- a/include/linux/drbd_limits.h
+++ b/include/linux/drbd_limits.h
@@ -230,4 +230,10 @@
 #define DRBD_SOCKET_CHECK_TIMEO_MAX DRBD_PING_TIMEO_MAX
 #define DRBD_SOCKET_CHECK_TIMEO_DEF 0
 #define DRBD_SOCKET_CHECK_TIMEO_SCALE '1'
+
+#define DRBD_RS_DISCARD_GRANULARITY_MIN 0
+#define DRBD_RS_DISCARD_GRANULARITY_MAX (1<<20)  /* 1MiByte */
+#define DRBD_RS_DISCARD_GRANULARITY_DEF 0     /* disabled by default */
+#define DRBD_RS_DISCARD_GRANULARITY_SCALE '1' /* bytes */
+
 #endif
-- 
1.9.1

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

#1386667 — Re: [Drbd-dev] [PATCH 05/30] drbd: Introduce new disk config option rs-discard-granularity

Am Montag, 25. April 2016, 08:35:26 schrieb Bart Van Assche:
> On 04/25/2016 05:10 AM, Philipp Reisner wrote:
> > As long as the value is 0 the feature is disabled. With setting
> > it to a positive value, DRBD limits and aligns its resync requests
> > to the rs-discard-granularity setting. If the sync source detects
> > all zeros in such a block, the resync target discards the range
> > on disk.
> 
> Hello Phil,
> 
> Can you explain why rs-discard-granularity is configurable instead of
> e.g. setting it to the least common multiple of the discard
> granularities of the underlying block devices at both sides?
> 
> Thanks,
> 

Hi Bart,

we had this idea as well. It seems that real world devices like larger
discards better than smaller discards. The other motivation was that
a device mapper logical volume might change it on the fly...
So we think it is best to delegate the decision on the discard chunk
size to user space.

best regards,
 Phil
	

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

#1386801 — Re: [Drbd-dev] [PATCH 05/30] drbd: Introduce new disk config option rs-discard-granularity

Am Montag, 25. April 2016, 11:48:30 schrieb Bart Van Assche:
> On 04/25/2016 09:42 AM, Philipp Reisner wrote:
> > Am Montag, 25. April 2016, 08:35:26 schrieb Bart Van Assche:
> >> On 04/25/2016 05:10 AM, Philipp Reisner wrote:
> >>> As long as the value is 0 the feature is disabled. With setting
> >>> it to a positive value, DRBD limits and aligns its resync requests
> >>> to the rs-discard-granularity setting. If the sync source detects
> >>> all zeros in such a block, the resync target discards the range
> >>> on disk.
> >> 
> >> Can you explain why rs-discard-granularity is configurable instead of
> >> e.g. setting it to the least common multiple of the discard
> >> granularities of the underlying block devices at both sides?
> > 
> > we had this idea as well. It seems that real world devices like larger
> > discards better than smaller discards. The other motivation was that
> > a device mapper logical volume might change it on the fly...
> > So we think it is best to delegate the decision on the discard chunk
> > size to user space.
> 
> Hello Phil,
> 
> Are you aware that for aligned discard requests the discard granularity
> does not affect the size of discard requests at all?
> 
> Regarding LVM volumes: if the discard granularity for such volumes can
> change on the fly, shouldn't I/O be quiesced by the LVM kernel driver
> before it changes the discard granularity? I think that increasing
> discard granularity while I/O is in progress should be considered as a bug.
> 
> Bart.

Hi Bart,

I worked on this about 6 month ago, sorry for not having all the details
at the top of my head immediately. I think it came back now:
We need to announce the discard granularity when we create the device/minor.
At might it might be that there is no connection to the peer node. So we
are left with information about the discard granularity of the local
backing device only.
Therefore we decided to delegate it to the user/admin to provide the
discard granularity for the resync process.

best regards,
 phil 

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

#1386872 — [PATCH 05/30] drbd: Introduce new disk config option rs-discard-granularity

On Mon, Apr 25, 2016 at 02:49:11PM -0500, Philipp Reisner wrote:
> Am Montag, 25. April 2016, 11:48:30 schrieb Bart Van Assche:
> > On 04/25/2016 09:42 AM, Philipp Reisner wrote:
> > > Am Montag, 25. April 2016, 08:35:26 schrieb Bart Van Assche:
> > >> On 04/25/2016 05:10 AM, Philipp Reisner wrote:
> > >>> As long as the value is 0 the feature is disabled. With setting
> > >>> it to a positive value, DRBD limits and aligns its resync requests
> > >>> to the rs-discard-granularity setting. If the sync source detects
> > >>> all zeros in such a block, the resync target discards the range
> > >>> on disk.
> > >> 
> > >> Can you explain why rs-discard-granularity is configurable instead of
> > >> e.g. setting it to the least common multiple of the discard
> > >> granularities of the underlying block devices at both sides?
> > > 
> > > we had this idea as well. It seems that real world devices like larger
> > > discards better than smaller discards. The other motivation was that
> > > a device mapper logical volume might change it on the fly...
> > > So we think it is best to delegate the decision on the discard chunk
> > > size to user space.
> > 
> > Hello Phil,
> > 
> > Are you aware that for aligned discard requests the discard granularity
> > does not affect the size of discard requests at all?
> > 
> > Regarding LVM volumes: if the discard granularity for such volumes can
> > change on the fly, shouldn't I/O be quiesced by the LVM kernel driver
> > before it changes the discard granularity? I think that increasing
> > discard granularity while I/O is in progress should be considered as a bug.
> > 
> > Bart.
> 
> Hi Bart,
> 
> I worked on this about 6 month ago, sorry for not having all the details
> at the top of my head immediately. I think it came back now:
> We need to announce the discard granularity when we create the device/minor.
> At might it might be that there is no connection to the peer node. So we
> are left with information about the discard granularity of the local
> backing device only.
> Therefore we decided to delegate it to the user/admin to provide the
> discard granularity for the resync process.

Also, even though it may be technically possible to discard at 512 Byte
granularity, you may want to have the resync only actually do it for
bigger chunks. for $reasons.

    Lars

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

#1386355 — [PATCH 28/30] drbd: finally report ms, not jiffies, in log message

From: Lars Ellenberg <lars.ellenberg@linbit.com>

Also skip the message unless bitmap IO took longer than 5 ms.

Signed-off-by: Philipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: Lars Ellenberg <lars.ellenberg@linbit.com>
---
 drivers/block/drbd/drbd_bitmap.c | 12 ++++++++----
 1 file changed, 8 insertions(+), 4 deletions(-)

diff --git a/drivers/block/drbd/drbd_bitmap.c b/drivers/block/drbd/drbd_bitmap.c
index 17e5e60..801b8f3 100644
--- a/drivers/block/drbd/drbd_bitmap.c
+++ b/drivers/block/drbd/drbd_bitmap.c
@@ -1121,10 +1121,14 @@ static int bm_rw(struct drbd_device *device, const unsigned int flags, unsigned
 		kref_put(&ctx->kref, &drbd_bm_aio_ctx_destroy);
 
 	/* summary for global bitmap IO */
-	if (flags == 0)
-		drbd_info(device, "bitmap %s of %u pages took %lu jiffies\n",
-			 (flags & BM_AIO_READ) ? "READ" : "WRITE",
-			 count, jiffies - now);
+	if (flags == 0) {
+		unsigned int ms = jiffies_to_msecs(jiffies - now);
+		if (ms > 5) {
+			drbd_info(device, "bitmap %s of %u pages took %u ms\n",
+				 (flags & BM_AIO_READ) ? "READ" : "WRITE",
+				 count, ms);
+		}
+	}
 
 	if (ctx->error) {
 		drbd_alert(device, "we had at least one MD IO ERROR during bitmap IO\n");
-- 
1.9.1

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

#1386356 — [PATCH 11/30] drbd: when receiving P_TRIM, zero-out partial unaligned chunks

From: Lars Ellenberg <lars.ellenberg@linbit.com>

We can avoid spurious data divergence caused by partially-ignored
discards on certain backends with discard_zeroes_data=0, if we
translate partial unaligned discard requests into explicit zero-out.

The relevant use case is LVM/DM thin.

If on different nodes, DRBD is backed by devices with differing
discard characteristics, discards may lead to data divergence
(old data or garbage left over on one backend, zeroes due to
unmapped areas on the other backend). Online verify would now
potentially report tons of spurious differences.

While probably harmless for most use cases (fstrim on a file system),
DRBD cannot have that, it would violate our promise to upper layers
that our data instances on the nodes are identical.

To be correct and play safe (make sure data is identical on both copies),
we would have to disable discard support, if our local backend (on a
Primary) does not support "discard_zeroes_data=true".

We'd also have to translate discards to explicit zero-out on the
receiving (typically: Secondary) side, unless the receiving side
supports "discard_zeroes_data=true".

Which both would allocate those blocks, instead of unmapping them,
in contrast with expectations.

LVM/DM thin does set discard_zeroes_data=0,
because it silently ignores discards to partial chunks.

We can work around this by checking the alignment first.
For unaligned (wrt. alignment and granularity) or too small discards,
we zero-out the initial (and/or) trailing unaligned partial chunks,
but discard all the aligned full chunks.

At least for LVM/DM thin, the result is effectively "discard_zeroes_data=1".

Arguably it should behave this way internally, by default,
and we'll try to make that happen.

But our workaround is still valid for already deployed setups,
and for other devices that may behave this way.

Setting discard-zeroes-if-aligned=yes will allow DRBD to use
discards, and to announce discard_zeroes_data=true, even on
backends that announce discard_zeroes_data=false.

Setting discard-zeroes-if-aligned=no will cause DRBD to always
fall-back to zero-out on the receiving side, and to not even
announce discard capabilities on the Primary, if the respective
backend announces discard_zeroes_data=false.

We used to ignore the discard_zeroes_data setting completely.
To not break established and expected behaviour, and suddenly
cause fstrim on thin-provisioned LVs to run out-of-space,
instead of freeing up space, the default value is "yes".

Signed-off-by: Philipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: Lars Ellenberg <lars.ellenberg@linbit.com>
---
 drivers/block/drbd/drbd_int.h      |   2 +-
 drivers/block/drbd/drbd_nl.c       |  15 ++--
 drivers/block/drbd/drbd_receiver.c | 140 ++++++++++++++++++++++++++++++-------
 include/linux/drbd_genl.h          |   1 +
 include/linux/drbd_limits.h        |   6 ++
 5 files changed, 134 insertions(+), 30 deletions(-)

diff --git a/drivers/block/drbd/drbd_int.h b/drivers/block/drbd/drbd_int.h
index 1a93f4f..8cc2955 100644
--- a/drivers/block/drbd/drbd_int.h
+++ b/drivers/block/drbd/drbd_int.h
@@ -1488,7 +1488,7 @@ enum determine_dev_size {
 extern enum determine_dev_size
 drbd_determine_dev_size(struct drbd_device *, enum dds_flags, struct resize_parms *) __must_hold(local);
 extern void resync_after_online_grow(struct drbd_device *);
-extern void drbd_reconsider_max_bio_size(struct drbd_device *device, struct drbd_backing_dev *bdev);
+extern void drbd_reconsider_queue_parameters(struct drbd_device *device, struct drbd_backing_dev *bdev);
 extern enum drbd_state_rv drbd_set_role(struct drbd_device *device,
 					enum drbd_role new_role,
 					int force);
diff --git a/drivers/block/drbd/drbd_nl.c b/drivers/block/drbd/drbd_nl.c
index e63c5c4..4a0b184 100644
--- a/drivers/block/drbd/drbd_nl.c
+++ b/drivers/block/drbd/drbd_nl.c
@@ -1161,13 +1161,17 @@ static void drbd_setup_queue_param(struct drbd_device *device, struct drbd_backi
 	unsigned int max_hw_sectors = max_bio_size >> 9;
 	unsigned int max_segments = 0;
 	struct request_queue *b = NULL;
+	struct disk_conf *dc;
+	bool discard_zeroes_if_aligned = true;
 
 	if (bdev) {
 		b = bdev->backing_bdev->bd_disk->queue;
 
 		max_hw_sectors = min(queue_max_hw_sectors(b), max_bio_size >> 9);
 		rcu_read_lock();
-		max_segments = rcu_dereference(device->ldev->disk_conf)->max_bio_bvecs;
+		dc = rcu_dereference(device->ldev->disk_conf);
+		max_segments = dc->max_bio_bvecs;
+		discard_zeroes_if_aligned = dc->discard_zeroes_if_aligned;
 		rcu_read_unlock();
 
 		blk_set_stacking_limits(&q->limits);
@@ -1185,7 +1189,7 @@ static void drbd_setup_queue_param(struct drbd_device *device, struct drbd_backi
 
 		blk_queue_max_discard_sectors(q, DRBD_MAX_DISCARD_SECTORS);
 
-		if (blk_queue_discard(b) &&
+		if (blk_queue_discard(b) && (b->limits.discard_zeroes_data || discard_zeroes_if_aligned) &&
 		    (connection->cstate < C_CONNECTED || connection->agreed_features & FF_TRIM)) {
 			/* We don't care, stacking below should fix it for the local device.
 			 * Whether or not it is a suitable granularity on the remote device
@@ -1216,7 +1220,7 @@ static void drbd_setup_queue_param(struct drbd_device *device, struct drbd_backi
 	}
 }
 
-void drbd_reconsider_max_bio_size(struct drbd_device *device, struct drbd_backing_dev *bdev)
+void drbd_reconsider_queue_parameters(struct drbd_device *device, struct drbd_backing_dev *bdev)
 {
 	unsigned int now, new, local, peer;
 
@@ -1488,6 +1492,9 @@ int drbd_adm_disk_opts(struct sk_buff *skb, struct genl_info *info)
 	if (write_ordering_changed(old_disk_conf, new_disk_conf))
 		drbd_bump_write_ordering(device->resource, NULL, WO_BDEV_FLUSH);
 
+	if (old_disk_conf->discard_zeroes_if_aligned != new_disk_conf->discard_zeroes_if_aligned)
+		drbd_reconsider_queue_parameters(device, device->ldev);
+
 	drbd_md_sync(device);
 
 	if (device->state.conn >= C_CONNECTED) {
@@ -1866,7 +1873,7 @@ int drbd_adm_attach(struct sk_buff *skb, struct genl_info *info)
 	device->read_cnt = 0;
 	device->writ_cnt = 0;
 
-	drbd_reconsider_max_bio_size(device, device->ldev);
+	drbd_reconsider_queue_parameters(device, device->ldev);
 
 	/* If I am currently not R_PRIMARY,
 	 * but meta data primary indicator is set,
diff --git a/drivers/block/drbd/drbd_receiver.c b/drivers/block/drbd/drbd_receiver.c
index a2e7ba9..078c4d98 100644
--- a/drivers/block/drbd/drbd_receiver.c
+++ b/drivers/block/drbd/drbd_receiver.c
@@ -1442,6 +1442,108 @@ void drbd_bump_write_ordering(struct drbd_resource *resource, struct drbd_backin
 		drbd_info(resource, "Method to ensure write ordering: %s\n", write_ordering_str[resource->write_ordering]);
 }
 
+/*
+ * We *may* ignore the discard-zeroes-data setting, if so configured.
+ *
+ * Assumption is that it "discard_zeroes_data=0" is only because the backend
+ * may ignore partial unaligned discards.
+ *
+ * LVM/DM thin as of at least
+ *   LVM version:     2.02.115(2)-RHEL7 (2015-01-28)
+ *   Library version: 1.02.93-RHEL7 (2015-01-28)
+ *   Driver version:  4.29.0
+ * still behaves this way.
+ *
+ * For unaligned (wrt. alignment and granularity) or too small discards,
+ * we zero-out the initial (and/or) trailing unaligned partial chunks,
+ * but discard all the aligned full chunks.
+ *
+ * At least for LVM/DM thin, the result is effectively "discard_zeroes_data=1".
+ */
+int drbd_issue_discard_or_zero_out(struct drbd_device *device, sector_t start, unsigned int nr_sectors, bool discard)
+{
+	struct block_device *bdev = device->ldev->backing_bdev;
+	struct request_queue *q = bdev_get_queue(bdev);
+	sector_t tmp, nr;
+	unsigned int max_discard_sectors, granularity;
+	int alignment;
+	int err = 0;
+
+	if (!discard)
+		goto zero_out;
+
+	/* Zero-sector (unknown) and one-sector granularities are the same.  */
+	granularity = max(q->limits.discard_granularity >> 9, 1U);
+	alignment = (bdev_discard_alignment(bdev) >> 9) % granularity;
+
+	max_discard_sectors = min(q->limits.max_discard_sectors, (1U << 22));
+	max_discard_sectors -= max_discard_sectors % granularity;
+	if (unlikely(!max_discard_sectors))
+		goto zero_out;
+
+	if (nr_sectors < granularity)
+		goto zero_out;
+
+	tmp = start;
+	if (sector_div(tmp, granularity) != alignment) {
+		if (nr_sectors < 2*granularity)
+			goto zero_out;
+		/* start + gran - (start + gran - align) % gran */
+		tmp = start + granularity - alignment;
+		tmp = start + granularity - sector_div(tmp, granularity);
+
+		nr = tmp - start;
+		err |= blkdev_issue_zeroout(bdev, start, nr, GFP_NOIO, 0);
+		nr_sectors -= nr;
+		start = tmp;
+	}
+	while (nr_sectors >= granularity) {
+		nr = min_t(sector_t, nr_sectors, max_discard_sectors);
+		err |= blkdev_issue_discard(bdev, start, nr, GFP_NOIO, 0);
+		nr_sectors -= nr;
+		start += nr;
+	}
+ zero_out:
+	if (nr_sectors) {
+		err |= blkdev_issue_zeroout(bdev, start, nr_sectors, GFP_NOIO, 0);
+	}
+	return err != 0;
+}
+
+static bool can_do_reliable_discards(struct drbd_device *device)
+{
+	struct request_queue *q = bdev_get_queue(device->ldev->backing_bdev);
+	struct disk_conf *dc;
+	bool can_do;
+
+	if (!blk_queue_discard(q))
+		return false;
+
+	if (q->limits.discard_zeroes_data)
+		return true;
+
+	rcu_read_lock();
+	dc = rcu_dereference(device->ldev->disk_conf);
+	can_do = dc->discard_zeroes_if_aligned;
+	rcu_read_unlock();
+	return can_do;
+}
+
+void drbd_issue_peer_discard(struct drbd_device *device, struct drbd_peer_request *peer_req)
+{
+	/* If the backend cannot discard, or does not guarantee
+	 * read-back zeroes in discarded ranges, we fall back to
+	 * zero-out.  Unless configuration specifically requested
+	 * otherwise. */
+	if (!can_do_reliable_discards(device))
+		peer_req->flags |= EE_IS_TRIM_USE_ZEROOUT;
+
+	if (drbd_issue_discard_or_zero_out(device, peer_req->i.sector,
+	    peer_req->i.size >> 9, !(peer_req->flags & EE_IS_TRIM_USE_ZEROOUT)))
+		peer_req->flags |= EE_WAS_ERROR;
+	drbd_endio_write_sec_final(peer_req);
+}
+
 /**
  * drbd_submit_peer_request()
  * @device:	DRBD device.
@@ -1472,7 +1574,13 @@ int drbd_submit_peer_request(struct drbd_device *device,
 	unsigned nr_pages = (data_size + PAGE_SIZE -1) >> PAGE_SHIFT;
 	int err = -ENOMEM;
 
-	if (peer_req->flags & EE_IS_TRIM_USE_ZEROOUT) {
+	/* TRIM/DISCARD: for now, always use the helper function
+	 * blkdev_issue_zeroout(..., discard=true).
+	 * It's synchronous, but it does the right thing wrt. bio splitting.
+	 * Correctness first, performance later.  Next step is to code an
+	 * asynchronous variant of the same.
+	 */
+	if (peer_req->flags & EE_IS_TRIM) {
 		/* wait for all pending IO completions, before we start
 		 * zeroing things out. */
 		conn_wait_active_ee_empty(peer_req->peer_device->connection);
@@ -1489,19 +1597,10 @@ int drbd_submit_peer_request(struct drbd_device *device,
 			spin_unlock_irq(&device->resource->req_lock);
 		}
 
-		if (blkdev_issue_zeroout(device->ldev->backing_bdev,
-			sector, data_size >> 9, GFP_NOIO, false))
-			peer_req->flags |= EE_WAS_ERROR;
-		drbd_endio_write_sec_final(peer_req);
+		drbd_issue_peer_discard(device, peer_req);
 		return 0;
 	}
 
-	/* Discards don't have any payload.
-	 * But the scsi layer still expects a bio_vec it can use internally,
-	 * see sd_setup_discard_cmnd() and blk_add_request_payload(). */
-	if (peer_req->flags & EE_IS_TRIM)
-		nr_pages = 1;
-
 	/* In most cases, we will only need one bio.  But in case the lower
 	 * level restrictions happen to be different at this offset on this
 	 * side than those of the sending peer, we may need to submit the
@@ -1527,11 +1626,6 @@ next_bio:
 	bios = bio;
 	++n_bios;
 
-	if (rw & REQ_DISCARD) {
-		bio->bi_iter.bi_size = data_size;
-		goto submit;
-	}
-
 	page_chain_for_each(page) {
 		unsigned len = min_t(unsigned, data_size, PAGE_SIZE);
 		if (!bio_add_page(bio, page, len, 0)) {
@@ -1553,7 +1647,6 @@ next_bio:
 		--nr_pages;
 	}
 	D_ASSERT(device, data_size == 0);
-submit:
 	D_ASSERT(device, page == NULL);
 
 	atomic_set(&peer_req->pending_bios, n_bios);
@@ -2413,10 +2506,7 @@ static int receive_Data(struct drbd_connection *connection, struct packet_info *
 	dp_flags = be32_to_cpu(p->dp_flags);
 	rw |= wire_flags_to_bio(dp_flags);
 	if (pi->cmd == P_TRIM) {
-		struct request_queue *q = bdev_get_queue(device->ldev->backing_bdev);
 		peer_req->flags |= EE_IS_TRIM;
-		if (!blk_queue_discard(q))
-			peer_req->flags |= EE_IS_TRIM_USE_ZEROOUT;
 		D_ASSERT(peer_device, peer_req->i.size > 0);
 		D_ASSERT(peer_device, rw & REQ_DISCARD);
 		D_ASSERT(peer_device, peer_req->pages == NULL);
@@ -2487,7 +2577,7 @@ static int receive_Data(struct drbd_connection *connection, struct packet_info *
 	 * and we wait for all pending requests, respectively wait for
 	 * active_ee to become empty in drbd_submit_peer_request();
 	 * better not add ourselves here. */
-	if ((peer_req->flags & EE_IS_TRIM_USE_ZEROOUT) == 0)
+	if ((peer_req->flags & EE_IS_TRIM) == 0)
 		list_add_tail(&peer_req->w.list, &device->active_ee);
 	spin_unlock_irq(&device->resource->req_lock);
 
@@ -3903,14 +3993,14 @@ static int receive_sizes(struct drbd_connection *connection, struct packet_info
 	}
 
 	device->peer_max_bio_size = be32_to_cpu(p->max_bio_size);
-	/* Leave drbd_reconsider_max_bio_size() before drbd_determine_dev_size().
+	/* Leave drbd_reconsider_queue_parameters() before drbd_determine_dev_size().
 	   In case we cleared the QUEUE_FLAG_DISCARD from our queue in
-	   drbd_reconsider_max_bio_size(), we can be sure that after
+	   drbd_reconsider_queue_parameters(), we can be sure that after
 	   drbd_determine_dev_size() no REQ_DISCARDs are in the queue. */
 
 	ddsf = be16_to_cpu(p->dds_flags);
 	if (get_ldev(device)) {
-		drbd_reconsider_max_bio_size(device, device->ldev);
+		drbd_reconsider_queue_parameters(device, device->ldev);
 		dd = drbd_determine_dev_size(device, ddsf, NULL);
 		put_ldev(device);
 		if (dd == DS_ERROR)
@@ -3930,7 +4020,7 @@ static int receive_sizes(struct drbd_connection *connection, struct packet_info
 		 * However, if he sends a zero current size,
 		 * take his (user-capped or) backing disk size anyways.
 		 */
-		drbd_reconsider_max_bio_size(device, NULL);
+		drbd_reconsider_queue_parameters(device, NULL);
 		drbd_set_my_capacity(device, p_csize ?: p_usize ?: p_size);
 	}
 
diff --git a/include/linux/drbd_genl.h b/include/linux/drbd_genl.h
index ab649d8..c934d3a 100644
--- a/include/linux/drbd_genl.h
+++ b/include/linux/drbd_genl.h
@@ -132,6 +132,7 @@ GENL_struct(DRBD_NLA_DISK_CONF, 3, disk_conf,
 	__flg_field_def(18, DRBD_GENLA_F_MANDATORY,	disk_drain, DRBD_DISK_DRAIN_DEF)
 	__flg_field_def(19, DRBD_GENLA_F_MANDATORY,	md_flushes, DRBD_MD_FLUSHES_DEF)
 	__flg_field_def(23,     0 /* OPTIONAL */,	al_updates, DRBD_AL_UPDATES_DEF)
+	__flg_field_def(24,     0 /* OPTIONAL */,	discard_zeroes_if_aligned, DRBD_DISCARD_ZEROES_IF_ALIGNED)
 )
 
 GENL_struct(DRBD_NLA_RESOURCE_OPTS, 4, res_opts,
diff --git a/include/linux/drbd_limits.h b/include/linux/drbd_limits.h
index 2e4aad8..a351c40 100644
--- a/include/linux/drbd_limits.h
+++ b/include/linux/drbd_limits.h
@@ -210,6 +210,12 @@
 #define DRBD_MD_FLUSHES_DEF	1
 #define DRBD_TCP_CORK_DEF	1
 #define DRBD_AL_UPDATES_DEF     1
+/* We used to ignore the discard_zeroes_data setting.
+ * To not change established (and expected) behaviour,
+ * by default assume that, for discard_zeroes_data=0,
+ * we can make that an effective discard_zeroes_data=1,
+ * if we only explicitly zero-out unaligned partial chunks. */
+#define DRBD_DISCARD_ZEROES_IF_ALIGNED 1
 
 #define DRBD_ALLOW_TWO_PRIMARIES_DEF	0
 #define DRBD_ALWAYS_ASBP_DEF	0
-- 
1.9.1

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

#1386873 — [PATCH 11/30] drbd: when receiving P_TRIM, zero-out partial unaligned chunks

On Mon, Apr 25, 2016 at 09:37:28AM -0700, Bart Van Assche wrote:
> On 04/25/2016 05:13 AM, Philipp Reisner wrote:
> >+	while (nr_sectors >= granularity) {
> >+		nr = min_t(sector_t, nr_sectors, max_discard_sectors);
> >+		err |= blkdev_issue_discard(bdev, start, nr, GFP_NOIO, 0);
> >+		nr_sectors -= nr;
> >+		start += nr;
> >+	}
> 
> Hello Phil,
> 
> In blk_bio_discard_split() the following statement protects against
> block drivers for which max_discard_sectors is not a multiple of the
> discard granularity:
> 
> 	max_discard_sectors -= max_discard_sectors % granularity;
> 
> Do we need something similar in the above loop?

Right. Did not realize that was "legal".

> To Jens: should the drbd_issue_discard_or_zero_out() function go
> upstream or should rather what this function does be integrated in
> blkdev_issue_zeroout() as is done by the patch series
> "[PATCH v2 0/6] Make blkdev_issue_discard() submit aligned discard requests"
> (http://thread.gmane.org/gmane.linux.kernel.device-mapper.devel/23801)?

Thanks,
    Lars

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

#1386911 — Re: [PATCH 11/30] drbd: when receiving P_TRIM, zero-out partial unaligned chunks

On Mon, Apr 25, 2016 at 02:07:38PM +0200, Philipp Reisner wrote:
> +	max_discard_sectors = min(q->limits.max_discard_sectors, (1U << 22));
> +	max_discard_sectors -= max_discard_sectors % granularity;

^^^^ there

On Mon, Apr 25, 2016 at 10:32:17PM +0200, Lars Ellenberg wrote:
> On Mon, Apr 25, 2016 at 09:37:28AM -0700, Bart Van Assche wrote:
> > On 04/25/2016 05:13 AM, Philipp Reisner wrote:
> > >+	while (nr_sectors >= granularity) {
> > >+		nr = min_t(sector_t, nr_sectors, max_discard_sectors);
> > >+		err |= blkdev_issue_discard(bdev, start, nr, GFP_NOIO, 0);
> > >+		nr_sectors -= nr;
> > >+		start += nr;
> > >+	}
> > 
> > Hello Phil,
> > 
> > In blk_bio_discard_split() the following statement protects against
> > block drivers for which max_discard_sectors is not a multiple of the
> > discard granularity:
> > 
> > 	max_discard_sectors -= max_discard_sectors % granularity;
> > 
> > Do we need something similar in the above loop?
> 
> Right. Did not realize that was "legal".

Wait. We already have that. see above.

    Lars

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

#1386357 — [PATCH 24/30] drbd: disallow promotion during resync handshake, avoid deadlock and hard reset

From: Lars Ellenberg <lars.ellenberg@linbit.com>

We already serialize connection state changes,
and other, non-connection state changes (role changes)
while we are establishing a connection.

But if we have an established connection,
then trigger a resync handshake (by primary --force or similar),
until now we just had to be "lucky".

Consider this sequence (e.g. deployment scenario):
create-md; up;
  -> Connected Secondary/Secondary Inconsistent/Inconsistent
then do a racy primary --force on both peers.

 block drbd0: drbd_sync_handshake:
 block drbd0: self 0000000000000004:0000000000000000:0000000000000000:0000000000000000 bits:25590 flags:0
 block drbd0: peer 0000000000000004:0000000000000000:0000000000000000:0000000000000000 bits:25590 flags:0
 block drbd0: peer( Unknown -> Secondary ) conn( WFReportParams -> Connected ) pdsk( DUnknown -> Inconsistent )
 block drbd0: peer( Secondary -> Primary ) pdsk( Inconsistent -> UpToDate )
  *** HERE things go wrong. ***
 block drbd0: role( Secondary -> Primary )
 block drbd0: drbd_sync_handshake:
 block drbd0: self 0000000000000005:0000000000000000:0000000000000000:0000000000000000 bits:25590 flags:0
 block drbd0: peer C90D2FC716D232AB:0000000000000004:0000000000000000:0000000000000000 bits:25590 flags:0
 block drbd0: Becoming sync target due to disk states.
 block drbd0: Writing the whole bitmap, full sync required after drbd_sync_handshake.
 block drbd0: Remote failed to finish a request within 6007ms > ko-count (2) * timeout (30 * 0.1s)
 drbd s0: peer( Primary -> Unknown ) conn( Connected -> Timeout ) pdsk( UpToDate -> DUnknown )

The problem here is that the local promotion happens before the sync handshake
triggered by the remote promotion was completed.  Some assumptions elsewhere
become wrong, and when the expected resync handshake is then received and
processed, we get stuck in a deadlock, which can only be recovered by reboot :-(

Fix: if we know the peer has good data,
and our own disk is present, but NOT good,
and there is no resync going on yet,
we expect a sync handshake to happen "soon".
So reject a racy promotion with SS_IN_TRANSIENT_STATE.

Result:
 ... as above ...
 block drbd0: peer( Secondary -> Primary ) pdsk( Inconsistent -> UpToDate )
  *** local promotion being postponed until ... ***
 block drbd0: drbd_sync_handshake:
 block drbd0: self 0000000000000004:0000000000000000:0000000000000000:0000000000000000 bits:25590 flags:0
 block drbd0: peer 77868BDA836E12A5:0000000000000004:0000000000000000:0000000000000000 bits:25590 flags:0
  ...
 block drbd0: conn( WFBitMapT -> WFSyncUUID )
 block drbd0: updated sync uuid 85D06D0E8887AD44:0000000000000000:0000000000000000:0000000000000000
 block drbd0: conn( WFSyncUUID -> SyncTarget )
  *** ... after the resync handshake ***
 block drbd0: role( Secondary -> Primary )

Signed-off-by: Philipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: Lars Ellenberg <lars.ellenberg@linbit.com>
---
 drivers/block/drbd/drbd_state.c | 9 +++++++++
 1 file changed, 9 insertions(+)

diff --git a/drivers/block/drbd/drbd_state.c b/drivers/block/drbd/drbd_state.c
index 24422e8..7562c5c 100644
--- a/drivers/block/drbd/drbd_state.c
+++ b/drivers/block/drbd/drbd_state.c
@@ -906,6 +906,15 @@ is_valid_soft_transition(union drbd_state os, union drbd_state ns, struct drbd_c
 	      (ns.conn >= C_CONNECTED && os.conn == C_WF_REPORT_PARAMS)))
 		rv = SS_IN_TRANSIENT_STATE;
 
+	/* Do not promote during resync handshake triggered by "force primary".
+	 * This is a hack. It should really be rejected by the peer during the
+	 * cluster wide state change request. */
+	if (os.role != R_PRIMARY && ns.role == R_PRIMARY
+		&& ns.pdsk == D_UP_TO_DATE
+		&& ns.disk != D_UP_TO_DATE && ns.disk != D_DISKLESS
+		&& (ns.conn <= C_WF_SYNC_UUID || ns.conn != os.conn))
+			rv = SS_IN_TRANSIENT_STATE;
+
 	if ((ns.conn == C_VERIFY_S || ns.conn == C_VERIFY_T) && os.conn < C_CONNECTED)
 		rv = SS_NEED_CONNECTION;
 
-- 
1.9.1

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

#1386358 — [PATCH 29/30] drbd: al_write_transaction: skip re-scanning of bitmap page pointer array

From: Lars Ellenberg <lars.ellenberg@linbit.com>

For larger devices, the array of bitmap page pointers can grow very
large (8000 pointers per TB of storage).

For each activity log transaction, we need to flush the associated
bitmap pages to stable storage. Currently, we just "mark" the respective
pages while setting up the transaction, then tell the bitmap code to
write out all marked pages, but skip unchanged pages.

But one such transaction can affect only a small number of bitmap pages,
there is no need to scan the full array of several (ten-)thousand
page pointers to find the few marked ones.

Instead, remember the index numbers of the few affected pages,
and later only re-check those to skip duplicates and unchanged ones.

Signed-off-by: Philipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: Lars Ellenberg <lars.ellenberg@linbit.com>
---
 drivers/block/drbd/drbd_actlog.c |  2 ++
 drivers/block/drbd/drbd_bitmap.c | 66 +++++++++++++++++++++++++++++++---------
 drivers/block/drbd/drbd_int.h    |  1 +
 3 files changed, 54 insertions(+), 15 deletions(-)

diff --git a/drivers/block/drbd/drbd_actlog.c b/drivers/block/drbd/drbd_actlog.c
index 99a2b92..8305615 100644
--- a/drivers/block/drbd/drbd_actlog.c
+++ b/drivers/block/drbd/drbd_actlog.c
@@ -339,6 +339,8 @@ static int __al_write_transaction(struct drbd_device *device, struct al_transact
 
 	i = 0;
 
+	drbd_bm_reset_al_hints(device);
+
 	/* Even though no one can start to change this list
 	 * once we set the LC_LOCKED -- from drbd_al_begin_io(),
 	 * lc_try_lock_for_transaction() --, someone may still
diff --git a/drivers/block/drbd/drbd_bitmap.c b/drivers/block/drbd/drbd_bitmap.c
index 801b8f3..b1c2a57 100644
--- a/drivers/block/drbd/drbd_bitmap.c
+++ b/drivers/block/drbd/drbd_bitmap.c
@@ -96,6 +96,13 @@ struct drbd_bitmap {
 	struct page **bm_pages;
 	spinlock_t bm_lock;
 
+	/* exclusively to be used by __al_write_transaction(),
+	 * drbd_bm_mark_for_writeout() and
+	 * and drbd_bm_write_hinted() -> bm_rw() called from there.
+	 */
+	unsigned int n_bitmap_hints;
+	unsigned int al_bitmap_hints[AL_UPDATES_PER_TRANSACTION];
+
 	/* see LIMITATIONS: above */
 
 	unsigned long bm_set;       /* nr of set bits; THINK maybe atomic_t? */
@@ -242,6 +249,11 @@ static void bm_set_page_need_writeout(struct page *page)
 	set_bit(BM_PAGE_NEED_WRITEOUT, &page_private(page));
 }
 
+void drbd_bm_reset_al_hints(struct drbd_device *device)
+{
+	device->bitmap->n_bitmap_hints = 0;
+}
+
 /**
  * drbd_bm_mark_for_writeout() - mark a page with a "hint" to be considered for writeout
  * @device:	DRBD device.
@@ -253,6 +265,7 @@ static void bm_set_page_need_writeout(struct page *page)
  */
 void drbd_bm_mark_for_writeout(struct drbd_device *device, int page_nr)
 {
+	struct drbd_bitmap *b = device->bitmap;
 	struct page *page;
 	if (page_nr >= device->bitmap->bm_number_of_pages) {
 		drbd_warn(device, "BAD: page_nr: %u, number_of_pages: %u\n",
@@ -260,7 +273,9 @@ void drbd_bm_mark_for_writeout(struct drbd_device *device, int page_nr)
 		return;
 	}
 	page = device->bitmap->bm_pages[page_nr];
-	set_bit(BM_PAGE_HINT_WRITEOUT, &page_private(page));
+	BUG_ON(b->n_bitmap_hints >= ARRAY_SIZE(b->al_bitmap_hints));
+	if (!test_and_set_bit(BM_PAGE_HINT_WRITEOUT, &page_private(page)))
+		b->al_bitmap_hints[b->n_bitmap_hints++] = page_nr;
 }
 
 static int bm_test_page_unchanged(struct page *page)
@@ -1030,7 +1045,7 @@ static int bm_rw(struct drbd_device *device, const unsigned int flags, unsigned
 {
 	struct drbd_bm_aio_ctx *ctx;
 	struct drbd_bitmap *b = device->bitmap;
-	int num_pages, i, count = 0;
+	unsigned int num_pages, i, count = 0;
 	unsigned long now;
 	char ppb[10];
 	int err = 0;
@@ -1078,16 +1093,37 @@ static int bm_rw(struct drbd_device *device, const unsigned int flags, unsigned
 	now = jiffies;
 
 	/* let the layers below us try to merge these bios... */
-	for (i = 0; i < num_pages; i++) {
-		/* ignore completely unchanged pages */
-		if (lazy_writeout_upper_idx && i == lazy_writeout_upper_idx)
-			break;
-		if (!(flags & BM_AIO_READ)) {
-			if ((flags & BM_AIO_WRITE_HINTED) &&
-			    !test_and_clear_bit(BM_PAGE_HINT_WRITEOUT,
-				    &page_private(b->bm_pages[i])))
-				continue;
 
+	if (flags & BM_AIO_READ) {
+		for (i = 0; i < num_pages; i++) {
+			atomic_inc(&ctx->in_flight);
+			bm_page_io_async(ctx, i);
+			++count;
+			cond_resched();
+		}
+	} else if (flags & BM_AIO_WRITE_HINTED) {
+		/* ASSERT: BM_AIO_WRITE_ALL_PAGES is not set. */
+		unsigned int hint;
+		for (hint = 0; hint < b->n_bitmap_hints; hint++) {
+			i = b->al_bitmap_hints[hint];
+			if (i >= num_pages) /* == -1U: no hint here. */
+				continue;
+			/* Several AL-extents may point to the same page. */
+			if (!test_and_clear_bit(BM_PAGE_HINT_WRITEOUT,
+			    &page_private(b->bm_pages[i])))
+				continue;
+			/* Has it even changed? */
+			if (bm_test_page_unchanged(b->bm_pages[i]))
+				continue;
+			atomic_inc(&ctx->in_flight);
+			bm_page_io_async(ctx, i);
+			++count;
+		}
+	} else {
+		for (i = 0; i < num_pages; i++) {
+			/* ignore completely unchanged pages */
+			if (lazy_writeout_upper_idx && i == lazy_writeout_upper_idx)
+				break;
 			if (!(flags & BM_AIO_WRITE_ALL_PAGES) &&
 			    bm_test_page_unchanged(b->bm_pages[i])) {
 				dynamic_drbd_dbg(device, "skipped bm write for idx %u\n", i);
@@ -1100,11 +1136,11 @@ static int bm_rw(struct drbd_device *device, const unsigned int flags, unsigned
 				dynamic_drbd_dbg(device, "skipped bm lazy write for idx %u\n", i);
 				continue;
 			}
+			atomic_inc(&ctx->in_flight);
+			bm_page_io_async(ctx, i);
+			++count;
+			cond_resched();
 		}
-		atomic_inc(&ctx->in_flight);
-		bm_page_io_async(ctx, i);
-		++count;
-		cond_resched();
 	}
 
 	/*
diff --git a/drivers/block/drbd/drbd_int.h b/drivers/block/drbd/drbd_int.h
index acb1462..c83393c 100644
--- a/drivers/block/drbd/drbd_int.h
+++ b/drivers/block/drbd/drbd_int.h
@@ -1378,6 +1378,7 @@ extern int  drbd_bm_e_weight(struct drbd_device *device, unsigned long enr);
 extern int  drbd_bm_read(struct drbd_device *device) __must_hold(local);
 extern void drbd_bm_mark_for_writeout(struct drbd_device *device, int page_nr);
 extern int  drbd_bm_write(struct drbd_device *device) __must_hold(local);
+extern void drbd_bm_reset_al_hints(struct drbd_device *device) __must_hold(local);
 extern int  drbd_bm_write_hinted(struct drbd_device *device) __must_hold(local);
 extern int  drbd_bm_write_lazy(struct drbd_device *device, unsigned upper_idx) __must_hold(local);
 extern int drbd_bm_write_all(struct drbd_device *device) __must_hold(local);
-- 
1.9.1

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

#1386359 — [PATCH 10/30] drbd: allow parallel flushes for multi-volume resources

From: Lars Ellenberg <lars.ellenberg@linbit.com>

To maintain write-order fidelity accros all volumes in a DRBD resource,
the receiver of a P_BARRIER needs to issue flushes to all volumes.
We used to do this by calling blkdev_issue_flush(), synchronously,
one volume at a time.

We now submit all flushes to all volumes in parallel, then wait for all
completions, to reduce worst-case latencies on multi-volume resources.

Signed-off-by: Philipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: Lars Ellenberg <lars.ellenberg@linbit.com>
---
 drivers/block/drbd/drbd_receiver.c | 113 +++++++++++++++++++++++++++++--------
 1 file changed, 88 insertions(+), 25 deletions(-)

diff --git a/drivers/block/drbd/drbd_receiver.c b/drivers/block/drbd/drbd_receiver.c
index 4cfc721..a2e7ba9 100644
--- a/drivers/block/drbd/drbd_receiver.c
+++ b/drivers/block/drbd/drbd_receiver.c
@@ -1204,13 +1204,83 @@ static int drbd_recv_header(struct drbd_connection *connection, struct packet_in
 	return err;
 }
 
-static void drbd_flush(struct drbd_connection *connection)
+/* This is blkdev_issue_flush, but asynchronous.
+ * We want to submit to all component volumes in parallel,
+ * then wait for all completions.
+ */
+struct issue_flush_context {
+	atomic_t pending;
+	int error;
+	struct completion done;
+};
+struct one_flush_context {
+	struct drbd_device *device;
+	struct issue_flush_context *ctx;
+};
+
+void one_flush_endio(struct bio *bio)
 {
-	int rv;
-	struct drbd_peer_device *peer_device;
-	int vnr;
+	struct one_flush_context *octx = bio->bi_private;
+	struct drbd_device *device = octx->device;
+	struct issue_flush_context *ctx = octx->ctx;
+
+	if (bio->bi_error) {
+		ctx->error = bio->bi_error;
+		drbd_info(device, "local disk FLUSH FAILED with status %d\n", bio->bi_error);
+	}
+	kfree(octx);
+	bio_put(bio);
+
+	clear_bit(FLUSH_PENDING, &device->flags);
+	put_ldev(device);
+	kref_put(&device->kref, drbd_destroy_device);
+
+	if (atomic_dec_and_test(&ctx->pending))
+		complete(&ctx->done);
+}
+
+static void submit_one_flush(struct drbd_device *device, struct issue_flush_context *ctx)
+{
+	struct bio *bio = bio_alloc(GFP_NOIO, 0);
+	struct one_flush_context *octx = kmalloc(sizeof(*octx), GFP_NOIO);
+	if (!bio || !octx) {
+		drbd_warn(device, "Could not allocate a bio, CANNOT ISSUE FLUSH\n");
+		/* FIXME: what else can I do now?  disconnecting or detaching
+		 * really does not help to improve the state of the world, either.
+		 */
+		kfree(octx);
+		if (bio)
+			bio_put(bio);
 
+		ctx->error = -ENOMEM;
+		put_ldev(device);
+		kref_put(&device->kref, drbd_destroy_device);
+		return;
+	}
+
+	octx->device = device;
+	octx->ctx = ctx;
+	bio->bi_bdev = device->ldev->backing_bdev;
+	bio->bi_private = octx;
+	bio->bi_end_io = one_flush_endio;
+
+	device->flush_jif = jiffies;
+	set_bit(FLUSH_PENDING, &device->flags);
+	atomic_inc(&ctx->pending);
+	submit_bio(WRITE_FLUSH, bio);
+}
+
+static void drbd_flush(struct drbd_connection *connection)
+{
 	if (connection->resource->write_ordering >= WO_BDEV_FLUSH) {
+		struct drbd_peer_device *peer_device;
+		struct issue_flush_context ctx;
+		int vnr;
+
+		atomic_set(&ctx.pending, 1);
+		ctx.error = 0;
+		init_completion(&ctx.done);
+
 		rcu_read_lock();
 		idr_for_each_entry(&connection->peer_devices, peer_device, vnr) {
 			struct drbd_device *device = peer_device->device;
@@ -1220,31 +1290,24 @@ static void drbd_flush(struct drbd_connection *connection)
 			kref_get(&device->kref);
 			rcu_read_unlock();
 
-			/* Right now, we have only this one synchronous code path
-			 * for flushes between request epochs.
-			 * We may want to make those asynchronous,
-			 * or at least parallelize the flushes to the volume devices.
-			 */
-			device->flush_jif = jiffies;
-			set_bit(FLUSH_PENDING, &device->flags);
-			rv = blkdev_issue_flush(device->ldev->backing_bdev,
-					GFP_NOIO, NULL);
-			clear_bit(FLUSH_PENDING, &device->flags);
-			if (rv) {
-				drbd_info(device, "local disk flush failed with status %d\n", rv);
-				/* would rather check on EOPNOTSUPP, but that is not reliable.
-				 * don't try again for ANY return value != 0
-				 * if (rv == -EOPNOTSUPP) */
-				drbd_bump_write_ordering(connection->resource, NULL, WO_DRAIN_IO);
-			}
-			put_ldev(device);
-			kref_put(&device->kref, drbd_destroy_device);
+			submit_one_flush(device, &ctx);
 
 			rcu_read_lock();
-			if (rv)
-				break;
 		}
 		rcu_read_unlock();
+
+		/* Do we want to add a timeout,
+		 * if disk-timeout is set? */
+		if (!atomic_dec_and_test(&ctx.pending))
+			wait_for_completion(&ctx.done);
+
+		if (ctx.error) {
+			/* would rather check on EOPNOTSUPP, but that is not reliable.
+			 * don't try again for ANY return value != 0
+			 * if (rv == -EOPNOTSUPP) */
+			/* Any error is already reported by bio_endio callback. */
+			drbd_bump_write_ordering(connection->resource, NULL, WO_DRAIN_IO);
+		}
 	}
 }
 
-- 
1.9.1

[toc] | [prev] | [standalone]

Back to top | Article view | linux.kernel

csiph-web