ZIL: Cleanup sync and commit handling

ZVOL:
 - Mark all ZVOL ZIL transactions as sync.  Since ZVOLs have only
one object, it makes no sense to maintain async queue and on each
commit merge it into sync. Single sync queue is just cheaper, while
it changes nothing until actual commit request arrives.
 - Remove zsd_sync_cnt and the zil_async_to_sync() calls since we
are no longer switching between sync and async queues.

ZFS:
 - Mark write transactions as sync based only on number of sync
opens (z_sync_cnt).  We can not randomly jump between sync and
async unless we want data corruptions due to writes reordering.
 - When file first opened with O_SYNC (z_sync_cnt incremented to 1)
call zil_async_to_sync() for it to preserve correct ordering between
past and future writes.
 - Drop zfs_fsyncer_key logic.  Looks like it was an optimization
for workloads heavily intermixing async writes with tons of fsyncs.
But first it was broken 8 years ago due to Linux tsd implementation
not allowing data storage between syscalls, and second, I doubt it
is safe to switch from async to sync so often and without calling
zil_async_to_sync().

 - Rename sync argument of *_log_write() into commit, now only
signalling caller's intent to call zil_commit() soon after.  It
allows WR_COPIED optimizations without extra other meanings.

Reviewed-by: Brian Behlendorf <[email protected]>
Reviewed-by: George Wilson <[email protected]>
Signed-off-by:	Alexander Motin <[email protected]>
Sponsored by:	iXsystems, Inc.
Closes #15366 

2 files changed, 24 insertions, 14 deletions

diff --git a/module/os/linux/zfs/zfs_vnops_os.c b/module/os/linux/zfs/zfs_vnops_os.c
index 1770e2372..6ad75ace0 100644
--- a/module/os/linux/zfs/zfs_vnops_os.c
+++ b/module/os/linux/zfs/zfs_vnops_os.c
@@ -192,9 +192,15 @@ zfs_open(struct inode *ip, int mode, int flag, cred_t *cr)
 		return (SET_ERROR(EPERM));
 	}
 
-	/* Keep a count of the synchronous opens in the znode */
-	if (flag & O_SYNC)
-		atomic_inc_32(&zp->z_sync_cnt);
+	/*
+	 * Keep a count of the synchronous opens in the znode.  On first
+	 * synchronous open we must convert all previous async transactions
+	 * into sync to keep correct ordering.
+	 */
+	if (flag & O_SYNC) {
+		if (atomic_inc_32_nv(&zp->z_sync_cnt) == 1)
+			zil_async_to_sync(zfsvfs->z_log, zp->z_id);
+	}
 
 	zfs_exit(zfsvfs, FTAG);
 	return (0);
@@ -3826,21 +3832,14 @@ zfs_putpage(struct inode *ip, struct page *pp, struct writeback_control *wbc,
 
 	err = sa_bulk_update(zp->z_sa_hdl, bulk, cnt, tx);
 
-	zfs_log_write(zfsvfs->z_log, tx, TX_WRITE, zp, pgoff, pglen, 0,
-	    for_sync ? zfs_putpage_sync_commit_cb :
-	    zfs_putpage_async_commit_cb, pp);
-
-	dmu_tx_commit(tx);
-
-	zfs_rangelock_exit(lr);
-
+	boolean_t commit = B_FALSE;
 	if (wbc->sync_mode != WB_SYNC_NONE) {
 		/*
 		 * Note that this is rarely called under writepages(), because
 		 * writepages() normally handles the entire commit for
 		 * performance reasons.
 		 */
-		zil_commit(zfsvfs->z_log, zp->z_id);
+		commit = B_TRUE;
 	} else if (!for_sync && atomic_load_32(&zp->z_sync_writes_cnt) > 0) {
 		/*
 		 * If the caller does not intend to wait synchronously
@@ -3850,9 +3849,20 @@ zfs_putpage(struct inode *ip, struct page *pp, struct writeback_control *wbc,
 		 * our writeback to complete. Refer to the comment in
 		 * zpl_fsync() (when HAVE_FSYNC_RANGE is defined) for details.
 		 */
-		zil_commit(zfsvfs->z_log, zp->z_id);
+		commit = B_TRUE;
 	}
 
+	zfs_log_write(zfsvfs->z_log, tx, TX_WRITE, zp, pgoff, pglen, commit,
+	    for_sync ? zfs_putpage_sync_commit_cb :
+	    zfs_putpage_async_commit_cb, pp);
+
+	dmu_tx_commit(tx);
+
+	zfs_rangelock_exit(lr);
+
+	if (commit)
+		zil_commit(zfsvfs->z_log, zp->z_id);
+
 	dataset_kstats_update_write_kstats(&zfsvfs->z_kstat, pglen);
 
 	zfs_exit(zfsvfs, FTAG);
diff --git a/module/os/linux/zfs/zvol_os.c b/module/os/linux/zfs/zvol_os.c
index f94ce69fb..3dc0bb388 100644
--- a/module/os/linux/zfs/zvol_os.c
+++ b/module/os/linux/zfs/zvol_os.c
@@ -387,7 +387,7 @@ zvol_discard(zv_request_t *zvr)
 	if (error != 0) {
 		dmu_tx_abort(tx);
 	} else {
-		zvol_log_truncate(zv, tx, start, size, B_TRUE);
+		zvol_log_truncate(zv, tx, start, size);
 		dmu_tx_commit(tx);
 		error = dmu_free_long_range(zv->zv_objset,
 		    ZVOL_OBJ, start, size);