X-Git-Url: https://git.camperquake.de/gitweb.cgi?a=blobdiff_plain;f=module%2Fzfs%2Fzio.c;h=0022c64cc5c05e490fa987f70c1d41a1bcc9a2c4;hb=f5fc4acaa77e2c1782a9495bbf1a39884b4c3940;hp=5ff35764aed92a58db637498d4ea9bcac8855cbe;hpb=c409e4647f221ab724a0bd10c480ac95447203c3;p=zfs.git

diff --git a/module/zfs/zio.c b/module/zfs/zio.c
index 5ff3576..0022c64 100644
--- a/module/zfs/zio.c
+++ b/module/zfs/zio.c
@@ -20,6 +20,7 @@
  */
 /*
  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
+ * Copyright (c) 2011 by Delphix. All rights reserved.
  */
 
 #include <sys/zfs_context.h>
@@ -79,6 +80,7 @@ int zio_delay_max = ZIO_DELAY_MAX;
 #ifdef _KERNEL
 extern vmem_t *zio_alloc_arena;
 #endif
+extern int zfs_mg_alloc_failures;
 
 /*
  * An allocating zio is one that either currently has the DVA allocate
@@ -158,6 +160,12 @@ zio_init(void)
 			zio_data_buf_cache[c - 1] = zio_data_buf_cache[c];
 	}
 
+	/*
+	 * The zio write taskqs have 1 thread per cpu, allow 1/2 of the taskqs
+	 * to fail 3 times per txg or 8 failures, whichever is greater.
+	 */
+	zfs_mg_alloc_failures = MAX((3 * max_ncpus / 2), 8);
+
 	zio_inject_init();
 }
 
@@ -1150,6 +1158,8 @@ __zio_execute(zio_t *zio)
 	while (zio->io_stage < ZIO_STAGE_DONE) {
 		enum zio_stage pipeline = zio->io_pipeline;
 		enum zio_stage stage = zio->io_stage;
+		dsl_pool_t *dsl;
+		boolean_t cut;
 		int rv;
 
 		ASSERT(!MUTEX_HELD(&zio->io_lock));
@@ -1162,19 +1172,26 @@ __zio_execute(zio_t *zio)
 
 		ASSERT(stage <= ZIO_STAGE_DONE);
 
+		dsl = spa_get_dsl(zio->io_spa);
+		cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
+		    zio_requeue_io_start_cut_in_line : B_FALSE;
+
 		/*
 		 * If we are in interrupt context and this pipeline stage
 		 * will grab a config lock that is held across I/O,
 		 * or may wait for an I/O that needs an interrupt thread
 		 * to complete, issue async to avoid deadlock.
 		 *
+		 * If we are in the txg_sync_thread or being called
+		 * during pool init issue async to minimize stack depth.
+		 * Both of these call paths may be recursively called.
+		 *
 		 * For VDEV_IO_START, we cut in line so that the io will
 		 * be sent to disk promptly.
 		 */
-		if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
-		    zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
-			boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
-			    zio_requeue_io_start_cut_in_line : B_FALSE;
+		if (((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
+		    zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) ||
+		    (dsl != NULL && dsl_pool_sync_context(dsl))) {
 			zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
 			return;
 		}
@@ -2142,6 +2159,7 @@ zio_dva_allocate(zio_t *zio)
 	metaslab_class_t *mc = spa_normal_class(spa);
 	blkptr_t *bp = zio->io_bp;
 	int error;
+	int flags = 0;
 
 	if (zio->io_gang_leader == NULL) {
 		ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
@@ -2154,10 +2172,21 @@ zio_dva_allocate(zio_t *zio)
 	ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
 	ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
 
+	/*
+	 * The dump device does not support gang blocks so allocation on
+	 * behalf of the dump device (i.e. ZIO_FLAG_NODATA) must avoid
+	 * the "fast" gang feature.
+	 */
+	flags |= (zio->io_flags & ZIO_FLAG_NODATA) ? METASLAB_GANG_AVOID : 0;
+	flags |= (zio->io_flags & ZIO_FLAG_GANG_CHILD) ?
+	    METASLAB_GANG_CHILD : 0;
 	error = metaslab_alloc(spa, mc, zio->io_size, bp,
-	    zio->io_prop.zp_copies, zio->io_txg, NULL, 0);
+	    zio->io_prop.zp_copies, zio->io_txg, NULL, flags);
 
 	if (error) {
+		spa_dbgmsg(spa, "%s: metaslab allocation failure: zio %p, "
+		    "size %llu, error %d", spa_name(spa), zio, zio->io_size,
+		    error);
 		if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
 			return (zio_write_gang_block(zio));
 		zio->io_error = error;