svn commit: r322242 - vendor-sys/illumos/dist/uts/common/fs/zfs

[Andriy Gapon]

Author: avg
Date: Tue Aug  8 11:24:13 2017
New Revision: 322242
URL: https://svnweb.freebsd.org/changeset/base/322242

Log:
  8373 TXG_WAIT in ZIL commit path
  
  illumos/illumos-gate at d28671a3b094af696bea87f52272d4c4d89321c7
  https://github.com/illumos/illumos-gate/commit/d28671a3b094af696bea87f52272d4c4d89321c7
  
  https://www.illumos.org/issues/8373
    The code that writes ZIL blocks uses dmu_tx_assign(TXG_WAIT) to assign a
    transaction to a transaction group.
    That seems to be logically incorrect as writing of the ZIL block does not
    introduce any new dirty data.
    Also, when there is a lot of dirty data, the call can introduce significant
    delays into the ZIL commit path,
    thus affecting all synchronous writes. Additionally, ARC throttling may affect
    the ZIL writing.
    We probably need a new mechanism similar to dmu_tx_create_assigned to assign
    ZIL transactions.
    (Ab)using TXG_WAITED does not seem to be sufficient.
  
  Reviewed by: Matthew Ahrens <mahrens at delphix.com>
  Reviewed by: Prakash Surya <prakash.surya at delphix.com>
  Approved by: Dan McDonald <danmcd at joyent.com>
  Author: Andriy Gapon <avg at FreeBSD.org>

Modified:
  vendor-sys/illumos/dist/uts/common/fs/zfs/zil.c

Modified: vendor-sys/illumos/dist/uts/common/fs/zfs/zil.c
==============================================================================
--- vendor-sys/illumos/dist/uts/common/fs/zfs/zil.c	Tue Aug  8 11:21:58 2017	(r322241)
+++ vendor-sys/illumos/dist/uts/common/fs/zfs/zil.c	Tue Aug  8 11:24:13 2017	(r322242)
@@ -974,7 +974,24 @@ zil_lwb_write_start(zilog_t *zilog, lwb_t *lwb)
 	 * to clean up in the event of allocation failure or I/O failure.
 	 */
 	tx = dmu_tx_create(zilog->zl_os);
-	VERIFY(dmu_tx_assign(tx, TXG_WAIT) == 0);
+
+	/*
+	 * Since we are not going to create any new dirty data and we can even
+	 * help with clearing the existing dirty data, we should not be subject
+	 * to the dirty data based delays.
+	 * We (ab)use TXG_WAITED to bypass the delay mechanism.
+	 * One side effect from using TXG_WAITED is that dmu_tx_assign() can
+	 * fail if the pool is suspended.  Those are dramatic circumstances,
+	 * so we return NULL to signal that the normal ZIL processing is not
+	 * possible and txg_wait_synced() should be used to ensure that the data
+	 * is on disk.
+	 */
+	error = dmu_tx_assign(tx, TXG_WAITED);
+	if (error != 0) {
+		ASSERT3S(error, ==, EIO);
+		dmu_tx_abort(tx);
+		return (NULL);
+	}
 	dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
 	txg = dmu_tx_get_txg(tx);