X-Git-Url: https://git.camperquake.de/gitweb.cgi?a=blobdiff_plain;f=module%2Fzfs%2Fzio.c;h=6b03be6f3bbf62e8d7c442e145e8bd5fd8ca90cb;hb=591fb62f19ee2431983a4cbeb0d200b1b8e7daf5;hp=0fa823687b51531df188a3c536af55fbd80d87c2;hpb=2fac4c2a74f1abc54d82068e3ae9a62e9a0468b0;p=zfs.git

diff --git a/module/zfs/zio.c b/module/zfs/zio.c
index 0fa8236..6b03be6 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
@@ -106,9 +108,9 @@ zio_init(void)
 	data_alloc_arena = zio_alloc_arena;
 #endif
 	zio_cache = kmem_cache_create("zio_cache",
-	    sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
+	    sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, KMC_KMEM);
 	zio_link_cache = kmem_cache_create("zio_link_cache",
-	    sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
+	    sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, KMC_KMEM);
 
 	/*
 	 * For small buffers, we want a cache for each multiple of
@@ -134,17 +136,27 @@ zio_init(void)
 
 		if (align != 0) {
 			char name[36];
+			int flags = zio_bulk_flags;
+
+			/*
+			 * The smallest buffers (512b) are heavily used and
+			 * experience a lot of churn.  The slabs allocated
+			 * for them are also relatively small (32K).  Thus
+			 * in over to avoid expensive calls to vmalloc() we
+			 * make an exception to the usual slab allocation
+			 * policy and force these buffers to be kmem backed.
+			 */
+			if (size == (1 << SPA_MINBLOCKSHIFT))
+				flags |= KMC_KMEM;
+
 			(void) sprintf(name, "zio_buf_%lu", (ulong_t)size);
 			zio_buf_cache[c] = kmem_cache_create(name, size,
-			    align, NULL, NULL, NULL, NULL, NULL,
-			    (size > zio_buf_debug_limit ? KMC_NODEBUG : 0) |
-			    zio_bulk_flags);
+			    align, NULL, NULL, NULL, NULL, NULL, flags);
 
 			(void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size);
 			zio_data_buf_cache[c] = kmem_cache_create(name, size,
-			    align, NULL, NULL, NULL, NULL, data_alloc_arena,
-			    (size > zio_buf_debug_limit ? KMC_NODEBUG : 0) |
-			    zio_bulk_flags);
+			    align, NULL, NULL, NULL, NULL,
+			    data_alloc_arena, flags);
 		}
 	}
 
@@ -158,6 +170,12 @@ 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();
 }
 
@@ -2151,6 +2169,7 @@ 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);
@@ -2163,10 +2182,21 @@ 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;