diff options
Diffstat (limited to 'sys/cddl/contrib/opensolaris/uts/common/fs/zfs/zio.c')
| -rw-r--r-- | sys/cddl/contrib/opensolaris/uts/common/fs/zfs/zio.c | 2470 |
1 files changed, 1449 insertions, 1021 deletions
diff --git a/sys/cddl/contrib/opensolaris/uts/common/fs/zfs/zio.c b/sys/cddl/contrib/opensolaris/uts/common/fs/zfs/zio.c index b5dd35f5599e..4650d42b7c2f 100644 --- a/sys/cddl/contrib/opensolaris/uts/common/fs/zfs/zio.c +++ b/sys/cddl/contrib/opensolaris/uts/common/fs/zfs/zio.c @@ -19,12 +19,10 @@ * CDDL HEADER END */ /* - * Copyright 2007 Sun Microsystems, Inc. All rights reserved. + * Copyright 2008 Sun Microsystems, Inc. All rights reserved. * Use is subject to license terms. */ -#pragma ident "%Z%%M% %I% %E% SMI" - #include <sys/zfs_context.h> #include <sys/fm/fs/zfs.h> #include <sys/spa.h> @@ -61,23 +59,9 @@ uint8_t zio_priority_table[ZIO_PRIORITY_TABLE_SIZE] = { char *zio_type_name[ZIO_TYPES] = { "null", "read", "write", "free", "claim", "ioctl" }; -/* At or above this size, force gang blocking - for testing */ -uint64_t zio_gang_bang = SPA_MAXBLOCKSIZE + 1; - -/* Force an allocation failure when non-zero */ -uint16_t zio_zil_fail_shift = 0; - -typedef struct zio_sync_pass { - int zp_defer_free; /* defer frees after this pass */ - int zp_dontcompress; /* don't compress after this pass */ - int zp_rewrite; /* rewrite new bps after this pass */ -} zio_sync_pass_t; - -zio_sync_pass_t zio_sync_pass = { - 1, /* zp_defer_free */ - 4, /* zp_dontcompress */ - 1, /* zp_rewrite */ -}; +#define SYNC_PASS_DEFERRED_FREE 1 /* defer frees after this pass */ +#define SYNC_PASS_DONT_COMPRESS 4 /* don't compress after this pass */ +#define SYNC_PASS_REWRITE 1 /* rewrite new bps after this pass */ /* * ========================================================================== @@ -94,6 +78,13 @@ kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT]; extern vmem_t *zio_alloc_arena; #endif +/* + * An allocating zio is one that either currently has the DVA allocate + * stage set or will have it later in its lifetime. + */ +#define IO_IS_ALLOCATING(zio) \ + ((zio)->io_orig_pipeline & (1U << ZIO_STAGE_DVA_ALLOCATE)) + void zio_init(void) { @@ -107,7 +98,6 @@ zio_init(void) data_alloc_arena = zio_alloc_arena; #endif #endif - zio_cache = kmem_cache_create("zio_cache", sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0); @@ -144,9 +134,6 @@ zio_init(void) zio_data_buf_cache[c] = kmem_cache_create(name, size, align, NULL, NULL, NULL, NULL, data_alloc_arena, KMC_NODEBUG); - - dprintf("creating cache for size %5lx align %5lx\n", - size, align); } } @@ -212,7 +199,7 @@ zio_buf_alloc(size_t size) ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); - return (kmem_cache_alloc(zio_buf_cache[c], KM_SLEEP)); + return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE)); #else return (kmem_alloc(size, KM_SLEEP)); #endif @@ -232,7 +219,7 @@ zio_data_buf_alloc(size_t size) ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); - return (kmem_cache_alloc(zio_data_buf_cache[c], KM_SLEEP)); + return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE)); #else return (kmem_alloc(size, KM_SLEEP)); #endif @@ -272,13 +259,15 @@ zio_data_buf_free(void *buf, size_t size) * ========================================================================== */ static void -zio_push_transform(zio_t *zio, void *data, uint64_t size, uint64_t bufsize) +zio_push_transform(zio_t *zio, void *data, uint64_t size, uint64_t bufsize, + zio_transform_func_t *transform) { zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP); - zt->zt_data = data; - zt->zt_size = size; + zt->zt_orig_data = zio->io_data; + zt->zt_orig_size = zio->io_size; zt->zt_bufsize = bufsize; + zt->zt_transform = transform; zt->zt_next = zio->io_transform_stack; zio->io_transform_stack = zt; @@ -288,128 +277,233 @@ zio_push_transform(zio_t *zio, void *data, uint64_t size, uint64_t bufsize) } static void -zio_pop_transform(zio_t *zio, void **data, uint64_t *size, uint64_t *bufsize) +zio_pop_transforms(zio_t *zio) { - zio_transform_t *zt = zio->io_transform_stack; + zio_transform_t *zt; + + while ((zt = zio->io_transform_stack) != NULL) { + if (zt->zt_transform != NULL) + zt->zt_transform(zio, + zt->zt_orig_data, zt->zt_orig_size); - *data = zt->zt_data; - *size = zt->zt_size; - *bufsize = zt->zt_bufsize; + zio_buf_free(zio->io_data, zt->zt_bufsize); - zio->io_transform_stack = zt->zt_next; - kmem_free(zt, sizeof (zio_transform_t)); + zio->io_data = zt->zt_orig_data; + zio->io_size = zt->zt_orig_size; + zio->io_transform_stack = zt->zt_next; - if ((zt = zio->io_transform_stack) != NULL) { - zio->io_data = zt->zt_data; - zio->io_size = zt->zt_size; + kmem_free(zt, sizeof (zio_transform_t)); } } +/* + * ========================================================================== + * I/O transform callbacks for subblocks and decompression + * ========================================================================== + */ +static void +zio_subblock(zio_t *zio, void *data, uint64_t size) +{ + ASSERT(zio->io_size > size); + + if (zio->io_type == ZIO_TYPE_READ) + bcopy(zio->io_data, data, size); +} + static void -zio_clear_transform_stack(zio_t *zio) +zio_decompress(zio_t *zio, void *data, uint64_t size) { - void *data; - uint64_t size, bufsize; + if (zio->io_error == 0 && + zio_decompress_data(BP_GET_COMPRESS(zio->io_bp), + zio->io_data, zio->io_size, data, size) != 0) + zio->io_error = EIO; +} - ASSERT(zio->io_transform_stack != NULL); +/* + * ========================================================================== + * I/O parent/child relationships and pipeline interlocks + * ========================================================================== + */ - zio_pop_transform(zio, &data, &size, &bufsize); - while (zio->io_transform_stack != NULL) { - zio_buf_free(data, bufsize); - zio_pop_transform(zio, &data, &size, &bufsize); +static void +zio_add_child(zio_t *pio, zio_t *zio) +{ + mutex_enter(&pio->io_lock); + if (zio->io_stage < ZIO_STAGE_READY) + pio->io_children[zio->io_child_type][ZIO_WAIT_READY]++; + if (zio->io_stage < ZIO_STAGE_DONE) + pio->io_children[zio->io_child_type][ZIO_WAIT_DONE]++; + zio->io_sibling_prev = NULL; + zio->io_sibling_next = pio->io_child; + if (pio->io_child != NULL) + pio->io_child->io_sibling_prev = zio; + pio->io_child = zio; + zio->io_parent = pio; + mutex_exit(&pio->io_lock); +} + +static void +zio_remove_child(zio_t *pio, zio_t *zio) +{ + zio_t *next, *prev; + + ASSERT(zio->io_parent == pio); + + mutex_enter(&pio->io_lock); + next = zio->io_sibling_next; + prev = zio->io_sibling_prev; + if (next != NULL) + next->io_sibling_prev = prev; + if (prev != NULL) + prev->io_sibling_next = next; + if (pio->io_child == zio) + pio->io_child = next; + mutex_exit(&pio->io_lock); +} + +static boolean_t +zio_wait_for_children(zio_t *zio, enum zio_child child, enum zio_wait_type wait) +{ + uint64_t *countp = &zio->io_children[child][wait]; + boolean_t waiting = B_FALSE; + + mutex_enter(&zio->io_lock); + ASSERT(zio->io_stall == NULL); + if (*countp != 0) { + zio->io_stage--; + zio->io_stall = countp; + waiting = B_TRUE; + } + mutex_exit(&zio->io_lock); + + return (waiting); +} + +static void +zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait) +{ + uint64_t *countp = &pio->io_children[zio->io_child_type][wait]; + int *errorp = &pio->io_child_error[zio->io_child_type]; + + mutex_enter(&pio->io_lock); + if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE)) + *errorp = zio_worst_error(*errorp, zio->io_error); + pio->io_reexecute |= zio->io_reexecute; + ASSERT3U(*countp, >, 0); + if (--*countp == 0 && pio->io_stall == countp) { + pio->io_stall = NULL; + mutex_exit(&pio->io_lock); + zio_execute(pio); + } else { + mutex_exit(&pio->io_lock); } } +static void +zio_inherit_child_errors(zio_t *zio, enum zio_child c) +{ + if (zio->io_child_error[c] != 0 && zio->io_error == 0) + zio->io_error = zio->io_child_error[c]; +} + /* * ========================================================================== - * Create the various types of I/O (read, write, free) + * Create the various types of I/O (read, write, free, etc) * ========================================================================== */ static zio_t * zio_create(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, void *data, uint64_t size, zio_done_func_t *done, void *private, - zio_type_t type, int priority, int flags, uint8_t stage, uint32_t pipeline) + zio_type_t type, int priority, int flags, vdev_t *vd, uint64_t offset, + const zbookmark_t *zb, uint8_t stage, uint32_t pipeline) { zio_t *zio; ASSERT3U(size, <=, SPA_MAXBLOCKSIZE); ASSERT(P2PHASE(size, SPA_MINBLOCKSIZE) == 0); + ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0); + + ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER)); + ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER)); + ASSERT(vd || stage == ZIO_STAGE_OPEN); zio = kmem_cache_alloc(zio_cache, KM_SLEEP); bzero(zio, sizeof (zio_t)); - zio->io_parent = pio; - zio->io_spa = spa; - zio->io_txg = txg; + + mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL); + cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL); + + if (vd != NULL) + zio->io_child_type = ZIO_CHILD_VDEV; + else if (flags & ZIO_FLAG_GANG_CHILD) + zio->io_child_type = ZIO_CHILD_GANG; + else + zio->io_child_type = ZIO_CHILD_LOGICAL; + if (bp != NULL) { zio->io_bp = bp; zio->io_bp_copy = *bp; zio->io_bp_orig = *bp; + if (type != ZIO_TYPE_WRITE) + zio->io_bp = &zio->io_bp_copy; /* so caller can free */ + if (zio->io_child_type == ZIO_CHILD_LOGICAL) { + if (BP_IS_GANG(bp)) + pipeline |= ZIO_GANG_STAGES; + zio->io_logical = zio; + } } + + zio->io_spa = spa; + zio->io_txg = txg; + zio->io_data = data; + zio->io_size = size; zio->io_done = done; zio->io_private = private; zio->io_type = type; zio->io_priority = priority; - zio->io_stage = stage; - zio->io_pipeline = pipeline; - zio->io_async_stages = ZIO_ASYNC_PIPELINE_STAGES; - zio->io_timestamp = lbolt64; - zio->io_flags = flags; - mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL); - cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL); - zio_push_transform(zio, data, size, size); + zio->io_vd = vd; + zio->io_offset = offset; + zio->io_orig_flags = zio->io_flags = flags; + zio->io_orig_stage = zio->io_stage = stage; + zio->io_orig_pipeline = zio->io_pipeline = pipeline; - /* - * Note on config lock: - * - * If CONFIG_HELD is set, then the caller already has the config - * lock, so we don't need it for this io. - * - * We set CONFIG_GRABBED to indicate that we have grabbed the - * config lock on behalf of this io, so it should be released - * in zio_done. - * - * Unless CONFIG_HELD is set, we will grab the config lock for - * any top-level (parent-less) io, *except* NULL top-level ios. - * The NULL top-level ios rarely have any children, so we delay - * grabbing the lock until the first child is added (but it is - * still grabbed on behalf of the top-level i/o, so additional - * children don't need to also grab it). This greatly reduces - * contention on the config lock. - */ - if (pio == NULL) { - if (type != ZIO_TYPE_NULL && - !(flags & ZIO_FLAG_CONFIG_HELD)) { - spa_config_enter(zio->io_spa, RW_READER, zio); - zio->io_flags |= ZIO_FLAG_CONFIG_GRABBED; - } - zio->io_root = zio; - } else { - zio->io_root = pio->io_root; - if (!(flags & ZIO_FLAG_NOBOOKMARK)) + if (zb != NULL) + zio->io_bookmark = *zb; + + if (pio != NULL) { + /* + * Logical I/Os can have logical, gang, or vdev children. + * Gang I/Os can have gang or vdev children. + * Vdev I/Os can only have vdev children. + * The following ASSERT captures all of these constraints. + */ + ASSERT(zio->io_child_type <= pio->io_child_type); + if (zio->io_logical == NULL) zio->io_logical = pio->io_logical; - mutex_enter(&pio->io_lock); - if (pio->io_parent == NULL && - pio->io_type == ZIO_TYPE_NULL && - !(pio->io_flags & ZIO_FLAG_CONFIG_GRABBED) && - !(pio->io_flags & ZIO_FLAG_CONFIG_HELD)) { - pio->io_flags |= ZIO_FLAG_CONFIG_GRABBED; - spa_config_enter(zio->io_spa, RW_READER, pio); - } - if (stage < ZIO_STAGE_READY) - pio->io_children_notready++; - pio->io_children_notdone++; - zio->io_sibling_next = pio->io_child; - zio->io_sibling_prev = NULL; - if (pio->io_child != NULL) - pio->io_child->io_sibling_prev = zio; - pio->io_child = zio; - zio->io_ndvas = pio->io_ndvas; - mutex_exit(&pio->io_lock); + zio_add_child(pio, zio); } return (zio); } +static void +zio_destroy(zio_t *zio) +{ + spa_t *spa = zio->io_spa; + uint8_t async_root = zio->io_async_root; + + mutex_destroy(&zio->io_lock); + cv_destroy(&zio->io_cv); + kmem_cache_free(zio_cache, zio); + + if (async_root) { + mutex_enter(&spa->spa_async_root_lock); + if (--spa->spa_async_root_count == 0) + cv_broadcast(&spa->spa_async_root_cv); + mutex_exit(&spa->spa_async_root_lock); + } +} + zio_t * zio_null(zio_t *pio, spa_t *spa, zio_done_func_t *done, void *private, int flags) @@ -417,8 +511,8 @@ zio_null(zio_t *pio, spa_t *spa, zio_done_func_t *done, void *private, zio_t *zio; zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private, - ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, ZIO_STAGE_OPEN, - ZIO_WAIT_FOR_CHILDREN_PIPELINE); + ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, NULL, 0, NULL, + ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE); return (zio); } @@ -430,160 +524,89 @@ zio_root(spa_t *spa, zio_done_func_t *done, void *private, int flags) } zio_t * -zio_read(zio_t *pio, spa_t *spa, blkptr_t *bp, void *data, - uint64_t size, zio_done_func_t *done, void *private, - int priority, int flags, zbookmark_t *zb) +zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, + void *data, uint64_t size, zio_done_func_t *done, void *private, + int priority, int flags, const zbookmark_t *zb) { zio_t *zio; - ASSERT3U(size, ==, BP_GET_LSIZE(bp)); - - zio = zio_create(pio, spa, bp->blk_birth, bp, data, size, done, private, - ZIO_TYPE_READ, priority, flags | ZIO_FLAG_USER, + zio = zio_create(pio, spa, bp->blk_birth, (blkptr_t *)bp, + data, size, done, private, + ZIO_TYPE_READ, priority, flags, NULL, 0, zb, ZIO_STAGE_OPEN, ZIO_READ_PIPELINE); - zio->io_bookmark = *zb; - - zio->io_logical = zio; - - /* - * Work off our copy of the bp so the caller can free it. - */ - zio->io_bp = &zio->io_bp_copy; - - if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF) { - uint64_t csize = BP_GET_PSIZE(bp); - void *cbuf = zio_buf_alloc(csize); - - zio_push_transform(zio, cbuf, csize, csize); - zio->io_pipeline |= 1U << ZIO_STAGE_READ_DECOMPRESS; - } - - if (BP_IS_GANG(bp)) { - uint64_t gsize = SPA_GANGBLOCKSIZE; - void *gbuf = zio_buf_alloc(gsize); - - zio_push_transform(zio, gbuf, gsize, gsize); - zio->io_pipeline |= 1U << ZIO_STAGE_READ_GANG_MEMBERS; - } return (zio); } zio_t * -zio_write(zio_t *pio, spa_t *spa, int checksum, int compress, int ncopies, - uint64_t txg, blkptr_t *bp, void *data, uint64_t size, - zio_done_func_t *ready, zio_done_func_t *done, void *private, int priority, - int flags, zbookmark_t *zb) +zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, + void *data, uint64_t size, zio_prop_t *zp, + zio_done_func_t *ready, zio_done_func_t *done, void *private, + int priority, int flags, const zbookmark_t *zb) { zio_t *zio; - ASSERT(checksum >= ZIO_CHECKSUM_OFF && - checksum < ZIO_CHECKSUM_FUNCTIONS); - - ASSERT(compress >= ZIO_COMPRESS_OFF && - compress < ZIO_COMPRESS_FUNCTIONS); + ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF && + zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS && + zp->zp_compress >= ZIO_COMPRESS_OFF && + zp->zp_compress < ZIO_COMPRESS_FUNCTIONS && + zp->zp_type < DMU_OT_NUMTYPES && + zp->zp_level < 32 && + zp->zp_ndvas > 0 && + zp->zp_ndvas <= spa_max_replication(spa)); + ASSERT(ready != NULL); zio = zio_create(pio, spa, txg, bp, data, size, done, private, - ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_USER, + ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb, ZIO_STAGE_OPEN, ZIO_WRITE_PIPELINE); zio->io_ready = ready; - - zio->io_bookmark = *zb; - - zio->io_logical = zio; - - zio->io_checksum = checksum; - zio->io_compress = compress; - zio->io_ndvas = ncopies; - - if (compress != ZIO_COMPRESS_OFF) - zio->io_async_stages |= 1U << ZIO_STAGE_WRITE_COMPRESS; - - if (bp->blk_birth != txg) { - /* XXX the bp usually (always?) gets re-zeroed later */ - BP_ZERO(bp); - BP_SET_LSIZE(bp, size); - BP_SET_PSIZE(bp, size); - } else { - /* Make sure someone doesn't change their mind on overwrites */ - ASSERT(MIN(zio->io_ndvas + BP_IS_GANG(bp), - spa_max_replication(spa)) == BP_GET_NDVAS(bp)); - } + zio->io_prop = *zp; return (zio); } zio_t * -zio_rewrite(zio_t *pio, spa_t *spa, int checksum, - uint64_t txg, blkptr_t *bp, void *data, uint64_t size, - zio_done_func_t *done, void *private, int priority, int flags, - zbookmark_t *zb) +zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, void *data, + uint64_t size, zio_done_func_t *done, void *private, int priority, + int flags, zbookmark_t *zb) { zio_t *zio; zio = zio_create(pio, spa, txg, bp, data, size, done, private, - ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_USER, + ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb, ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE); - zio->io_bookmark = *zb; - zio->io_checksum = checksum; - zio->io_compress = ZIO_COMPRESS_OFF; - - if (pio != NULL) - ASSERT3U(zio->io_ndvas, <=, BP_GET_NDVAS(bp)); - - return (zio); -} - -static zio_t * -zio_write_allocate(zio_t *pio, spa_t *spa, int checksum, - uint64_t txg, blkptr_t *bp, void *data, uint64_t size, - zio_done_func_t *done, void *private, int priority, int flags) -{ - zio_t *zio; - - BP_ZERO(bp); - BP_SET_LSIZE(bp, size); - BP_SET_PSIZE(bp, size); - BP_SET_COMPRESS(bp, ZIO_COMPRESS_OFF); - - zio = zio_create(pio, spa, txg, bp, data, size, done, private, - ZIO_TYPE_WRITE, priority, flags, - ZIO_STAGE_OPEN, ZIO_WRITE_ALLOCATE_PIPELINE); - - zio->io_checksum = checksum; - zio->io_compress = ZIO_COMPRESS_OFF; - return (zio); } zio_t * zio_free(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, - zio_done_func_t *done, void *private) + zio_done_func_t *done, void *private, int flags) { zio_t *zio; ASSERT(!BP_IS_HOLE(bp)); + if (bp->blk_fill == BLK_FILL_ALREADY_FREED) + return (zio_null(pio, spa, NULL, NULL, flags)); + if (txg == spa->spa_syncing_txg && - spa->spa_sync_pass > zio_sync_pass.zp_defer_free) { + spa_sync_pass(spa) > SYNC_PASS_DEFERRED_FREE) { bplist_enqueue_deferred(&spa->spa_sync_bplist, bp); - return (zio_null(pio, spa, NULL, NULL, 0)); + return (zio_null(pio, spa, NULL, NULL, flags)); } - zio = zio_create(pio, spa, txg, bp, NULL, 0, done, private, - ZIO_TYPE_FREE, ZIO_PRIORITY_FREE, ZIO_FLAG_USER, - ZIO_STAGE_OPEN, ZIO_FREE_PIPELINE); - - zio->io_bp = &zio->io_bp_copy; + zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp), + done, private, ZIO_TYPE_FREE, ZIO_PRIORITY_FREE, flags, + NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_FREE_PIPELINE); return (zio); } zio_t * zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, - zio_done_func_t *done, void *private) + zio_done_func_t *done, void *private, int flags) { zio_t *zio; @@ -601,11 +624,9 @@ zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa)); ASSERT3U(spa_first_txg(spa), <=, txg); - zio = zio_create(pio, spa, txg, bp, NULL, 0, done, private, - ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, 0, - ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE); - - zio->io_bp = &zio->io_bp_copy; + zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp), + done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, flags, + NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE); return (zio); } @@ -619,10 +640,9 @@ zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd, if (vd->vdev_children == 0) { zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private, - ZIO_TYPE_IOCTL, priority, flags, + ZIO_TYPE_IOCTL, priority, flags, vd, 0, NULL, ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE); - zio->io_vd = vd; zio->io_cmd = cmd; } else { zio = zio_null(pio, spa, NULL, NULL, flags); @@ -635,54 +655,23 @@ zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd, return (zio); } -static void -zio_phys_bp_init(vdev_t *vd, blkptr_t *bp, uint64_t offset, uint64_t size, - int checksum) -{ - ASSERT(vd->vdev_children == 0); - - ASSERT(size <= SPA_MAXBLOCKSIZE); - ASSERT(P2PHASE(size, SPA_MINBLOCKSIZE) == 0); - ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0); - - ASSERT(offset + size <= VDEV_LABEL_START_SIZE || - offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE); - ASSERT3U(offset + size, <=, vd->vdev_psize); - - BP_ZERO(bp); - - BP_SET_LSIZE(bp, size); - BP_SET_PSIZE(bp, size); - - BP_SET_CHECKSUM(bp, checksum); - BP_SET_COMPRESS(bp, ZIO_COMPRESS_OFF); - BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER); - - if (checksum != ZIO_CHECKSUM_OFF) - ZIO_SET_CHECKSUM(&bp->blk_cksum, offset, 0, 0, 0); -} - zio_t * zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size, void *data, int checksum, zio_done_func_t *done, void *private, - int priority, int flags) + int priority, int flags, boolean_t labels) { zio_t *zio; - blkptr_t blk; - zio_phys_bp_init(vd, &blk, offset, size, checksum); + ASSERT(vd->vdev_children == 0); + ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE || + offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE); + ASSERT3U(offset + size, <=, vd->vdev_psize); - zio = zio_create(pio, vd->vdev_spa, 0, &blk, data, size, done, private, - ZIO_TYPE_READ, priority, flags | ZIO_FLAG_PHYSICAL, + zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private, + ZIO_TYPE_READ, priority, flags, vd, offset, NULL, ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE); - zio->io_vd = vd; - zio->io_offset = offset; - - /* - * Work off our copy of the bp so the caller can free it. - */ - zio->io_bp = &zio->io_bp_copy; + zio->io_prop.zp_checksum = checksum; return (zio); } @@ -690,53 +679,49 @@ zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size, zio_t * zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size, void *data, int checksum, zio_done_func_t *done, void *private, - int priority, int flags) + int priority, int flags, boolean_t labels) { - zio_block_tail_t *zbt; - void *wbuf; zio_t *zio; - blkptr_t blk; - zio_phys_bp_init(vd, &blk, offset, size, checksum); + ASSERT(vd->vdev_children == 0); + ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE || + offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE); + ASSERT3U(offset + size, <=, vd->vdev_psize); - zio = zio_create(pio, vd->vdev_spa, 0, &blk, data, size, done, private, - ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_PHYSICAL, + zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private, + ZIO_TYPE_WRITE, priority, flags, vd, offset, NULL, ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE); - zio->io_vd = vd; - zio->io_offset = offset; - - zio->io_bp = &zio->io_bp_copy; - zio->io_checksum = checksum; + zio->io_prop.zp_checksum = checksum; if (zio_checksum_table[checksum].ci_zbt) { /* * zbt checksums are necessarily destructive -- they modify - * one word of the write buffer to hold the verifier/checksum. + * the end of the write buffer to hold the verifier/checksum. * Therefore, we must make a local copy in case the data is - * being written to multiple places. + * being written to multiple places in parallel. */ - wbuf = zio_buf_alloc(size); + void *wbuf = zio_buf_alloc(size); bcopy(data, wbuf, size); - zio_push_transform(zio, wbuf, size, size); - - zbt = (zio_block_tail_t *)((char *)wbuf + size) - 1; - zbt->zbt_cksum = blk.blk_cksum; + zio_push_transform(zio, wbuf, size, size, NULL); } return (zio); } /* - * Create a child I/O to do some work for us. It has no associated bp. + * Create a child I/O to do some work for us. */ zio_t * -zio_vdev_child_io(zio_t *zio, blkptr_t *bp, vdev_t *vd, uint64_t offset, +zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset, void *data, uint64_t size, int type, int priority, int flags, zio_done_func_t *done, void *private) { uint32_t pipeline = ZIO_VDEV_CHILD_PIPELINE; - zio_t *cio; + zio_t *zio; + + ASSERT(vd->vdev_parent == + (pio->io_vd ? pio->io_vd : pio->io_spa->spa_root_vdev)); if (type == ZIO_TYPE_READ && bp != NULL) { /* @@ -746,517 +731,754 @@ zio_vdev_child_io(zio_t *zio, blkptr_t *bp, vdev_t *vd, uint64_t offset, * eliminates redundant checksums in the interior nodes. */ pipeline |= 1U << ZIO_STAGE_CHECKSUM_VERIFY; - zio->io_pipeline &= ~(1U << ZIO_STAGE_CHECKSUM_VERIFY); + pio->io_pipeline &= ~(1U << ZIO_STAGE_CHECKSUM_VERIFY); } - cio = zio_create(zio, zio->io_spa, zio->io_txg, bp, data, size, + if (vd->vdev_children == 0) + offset += VDEV_LABEL_START_SIZE; + + zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size, done, private, type, priority, - (zio->io_flags & ZIO_FLAG_VDEV_INHERIT) | ZIO_FLAG_CANFAIL | flags, + (pio->io_flags & ZIO_FLAG_VDEV_INHERIT) | + ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | flags, + vd, offset, &pio->io_bookmark, ZIO_STAGE_VDEV_IO_START - 1, pipeline); - cio->io_vd = vd; - cio->io_offset = offset; - - return (cio); + return (zio); } -/* - * ========================================================================== - * Initiate I/O, either sync or async - * ========================================================================== - */ -int -zio_wait(zio_t *zio) +zio_t * +zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, void *data, uint64_t size, + int type, int priority, int flags, zio_done_func_t *done, void *private) { - int error; - - ASSERT(zio->io_stage == ZIO_STAGE_OPEN); - - zio->io_waiter = curthread; + zio_t *zio; - zio_next_stage_async(zio); + ASSERT(vd->vdev_ops->vdev_op_leaf); - mutex_enter(&zio->io_lock); - while (zio->io_stalled != ZIO_STAGE_DONE) - cv_wait(&zio->io_cv, &zio->io_lock); - mutex_exit(&zio->io_lock); + zio = zio_create(NULL, vd->vdev_spa, 0, NULL, + data, size, done, private, type, priority, + flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY, + vd, offset, NULL, + ZIO_STAGE_VDEV_IO_START - 1, ZIO_VDEV_CHILD_PIPELINE); - error = zio->io_error; - cv_destroy(&zio->io_cv); - mutex_destroy(&zio->io_lock); - kmem_cache_free(zio_cache, zio); - - return (error); + return (zio); } void -zio_nowait(zio_t *zio) +zio_flush(zio_t *zio, vdev_t *vd) { - zio_next_stage_async(zio); + zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE, + NULL, NULL, ZIO_PRIORITY_NOW, + ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY)); } /* * ========================================================================== - * I/O pipeline interlocks: parent/child dependency scoreboarding + * Prepare to read and write logical blocks * ========================================================================== */ -static void -zio_wait_for_children(zio_t *zio, uint32_t stage, uint64_t *countp) + +static int +zio_read_bp_init(zio_t *zio) { - mutex_enter(&zio->io_lock); - if (*countp == 0) { - ASSERT(zio->io_stalled == 0); - mutex_exit(&zio->io_lock); - zio_next_stage(zio); - } else { - zio->io_stalled = stage; - mutex_exit(&zio->io_lock); + blkptr_t *bp = zio->io_bp; + + if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF && zio->io_logical == zio) { + uint64_t csize = BP_GET_PSIZE(bp); + void *cbuf = zio_buf_alloc(csize); + + zio_push_transform(zio, cbuf, csize, csize, zio_decompress); } + + if (!dmu_ot[BP_GET_TYPE(bp)].ot_metadata && BP_GET_LEVEL(bp) == 0) + zio->io_flags |= ZIO_FLAG_DONT_CACHE; + + return (ZIO_PIPELINE_CONTINUE); } -static void -zio_notify_parent(zio_t *zio, uint32_t stage, uint64_t *countp) +static int +zio_write_bp_init(zio_t *zio) { - zio_t *pio = zio->io_parent; + zio_prop_t *zp = &zio->io_prop; + int compress = zp->zp_compress; + blkptr_t *bp = zio->io_bp; + void *cbuf; + uint64_t lsize = zio->io_size; + uint64_t csize = lsize; + uint64_t cbufsize = 0; + int pass = 1; - mutex_enter(&pio->io_lock); - if (pio->io_error == 0 && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE)) - pio->io_error = zio->io_error; - if (--*countp == 0 && pio->io_stalled == stage) { - pio->io_stalled = 0; - mutex_exit(&pio->io_lock); - zio_next_stage_async(pio); + /* + * If our children haven't all reached the ready stage, + * wait for them and then repeat this pipeline stage. + */ + if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) || + zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_READY)) + return (ZIO_PIPELINE_STOP); + + if (!IO_IS_ALLOCATING(zio)) + return (ZIO_PIPELINE_CONTINUE); + + ASSERT(compress != ZIO_COMPRESS_INHERIT); + + if (bp->blk_birth == zio->io_txg) { + /* + * We're rewriting an existing block, which means we're + * working on behalf of spa_sync(). For spa_sync() to + * converge, it must eventually be the case that we don't + * have to allocate new blocks. But compression changes + * the blocksize, which forces a reallocate, and makes + * convergence take longer. Therefore, after the first + * few passes, stop compressing to ensure convergence. + */ + pass = spa_sync_pass(zio->io_spa); + ASSERT(pass > 1); + + if (pass > SYNC_PASS_DONT_COMPRESS) + compress = ZIO_COMPRESS_OFF; + + /* + * Only MOS (objset 0) data should need to be rewritten. + */ + ASSERT(zio->io_logical->io_bookmark.zb_objset == 0); + + /* Make sure someone doesn't change their mind on overwrites */ + ASSERT(MIN(zp->zp_ndvas + BP_IS_GANG(bp), + spa_max_replication(zio->io_spa)) == BP_GET_NDVAS(bp)); + } + + if (compress != ZIO_COMPRESS_OFF) { + if (!zio_compress_data(compress, zio->io_data, zio->io_size, + &cbuf, &csize, &cbufsize)) { + compress = ZIO_COMPRESS_OFF; + } else if (csize != 0) { + zio_push_transform(zio, cbuf, csize, cbufsize, NULL); + } + } + + /* + * The final pass of spa_sync() must be all rewrites, but the first + * few passes offer a trade-off: allocating blocks defers convergence, + * but newly allocated blocks are sequential, so they can be written + * to disk faster. Therefore, we allow the first few passes of + * spa_sync() to allocate new blocks, but force rewrites after that. + * There should only be a handful of blocks after pass 1 in any case. + */ + if (bp->blk_birth == zio->io_txg && BP_GET_PSIZE(bp) == csize && + pass > SYNC_PASS_REWRITE) { + ASSERT(csize != 0); + uint32_t gang_stages = zio->io_pipeline & ZIO_GANG_STAGES; + zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages; + zio->io_flags |= ZIO_FLAG_IO_REWRITE; } else { - mutex_exit(&pio->io_lock); + BP_ZERO(bp); + zio->io_pipeline = ZIO_WRITE_PIPELINE; + } + + if (csize == 0) { + zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; + } else { + ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER); + BP_SET_LSIZE(bp, lsize); + BP_SET_PSIZE(bp, csize); + BP_SET_COMPRESS(bp, compress); + BP_SET_CHECKSUM(bp, zp->zp_checksum); + BP_SET_TYPE(bp, zp->zp_type); + BP_SET_LEVEL(bp, zp->zp_level); + BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER); } + + return (ZIO_PIPELINE_CONTINUE); } +/* + * ========================================================================== + * Execute the I/O pipeline + * ========================================================================== + */ + static void -zio_wait_children_ready(zio_t *zio) +zio_taskq_dispatch(zio_t *zio, enum zio_taskq_type q) { - zio_wait_for_children(zio, ZIO_STAGE_WAIT_CHILDREN_READY, - &zio->io_children_notready); -} + zio_type_t t = zio->io_type; -void -zio_wait_children_done(zio_t *zio) -{ - zio_wait_for_children(zio, ZIO_STAGE_WAIT_CHILDREN_DONE, - &zio->io_children_notdone); + /* + * If we're a config writer, the normal issue and interrupt threads + * may all be blocked waiting for the config lock. In this case, + * select the otherwise-unused taskq for ZIO_TYPE_NULL. + */ + if (zio->io_flags & ZIO_FLAG_CONFIG_WRITER) + t = ZIO_TYPE_NULL; + + /* + * A similar issue exists for the L2ARC write thread until L2ARC 2.0. + */ + if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux) + t = ZIO_TYPE_NULL; + + (void) taskq_dispatch(zio->io_spa->spa_zio_taskq[t][q], + (task_func_t *)zio_execute, zio, TQ_SLEEP); } -static void -zio_ready(zio_t *zio) +static boolean_t +zio_taskq_member(zio_t *zio, enum zio_taskq_type q) { - zio_t *pio = zio->io_parent; + kthread_t *executor = zio->io_executor; + spa_t *spa = zio->io_spa; - if (zio->io_ready) - zio->io_ready(zio); + for (zio_type_t t = 0; t < ZIO_TYPES; t++) + if (taskq_member(spa->spa_zio_taskq[t][q], executor)) + return (B_TRUE); - if (pio != NULL) - zio_notify_parent(zio, ZIO_STAGE_WAIT_CHILDREN_READY, - &pio->io_children_notready); + return (B_FALSE); +} - if (zio->io_bp) - zio->io_bp_copy = *zio->io_bp; +static int +zio_issue_async(zio_t *zio) +{ + zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE); - zio_next_stage(zio); + return (ZIO_PIPELINE_STOP); } -static void -zio_done(zio_t *zio) +void +zio_interrupt(zio_t *zio) { - zio_t *pio = zio->io_parent; - spa_t *spa = zio->io_spa; - blkptr_t *bp = zio->io_bp; - vdev_t *vd = zio->io_vd; + zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT); +} - ASSERT(zio->io_children_notready == 0); - ASSERT(zio->io_children_notdone == 0); +/* + * Execute the I/O pipeline until one of the following occurs: + * (1) the I/O completes; (2) the pipeline stalls waiting for + * dependent child I/Os; (3) the I/O issues, so we're waiting + * for an I/O completion interrupt; (4) the I/O is delegated by + * vdev-level caching or aggregation; (5) the I/O is deferred + * due to vdev-level queueing; (6) the I/O is handed off to + * another thread. In all cases, the pipeline stops whenever + * there's no CPU work; it never burns a thread in cv_wait(). + * + * There's no locking on io_stage because there's no legitimate way + * for multiple threads to be attempting to process the same I/O. + */ +static zio_pipe_stage_t *zio_pipeline[ZIO_STAGES]; - if (bp != NULL) { - ASSERT(bp->blk_pad[0] == 0); - ASSERT(bp->blk_pad[1] == 0); - ASSERT(bp->blk_pad[2] == 0); - ASSERT(bcmp(bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0); - if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(bp) && - !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) { - ASSERT(!BP_SHOULD_BYTESWAP(bp)); - if (zio->io_ndvas != 0) - ASSERT3U(zio->io_ndvas, <=, BP_GET_NDVAS(bp)); - ASSERT(BP_COUNT_GANG(bp) == 0 || - (BP_COUNT_GANG(bp) == BP_GET_NDVAS(bp))); - } - } +void +zio_execute(zio_t *zio) +{ + zio->io_executor = curthread; - if (vd != NULL) - vdev_stat_update(zio); + while (zio->io_stage < ZIO_STAGE_DONE) { + uint32_t pipeline = zio->io_pipeline; + zio_stage_t stage = zio->io_stage; + int rv; - if (zio->io_error) { - /* - * If this I/O is attached to a particular vdev, - * generate an error message describing the I/O failure - * at the block level. We ignore these errors if the - * device is currently unavailable. - */ - if (zio->io_error != ECKSUM && vd != NULL && !vdev_is_dead(vd)) - zfs_ereport_post(FM_EREPORT_ZFS_IO, - zio->io_spa, vd, zio, 0, 0); + ASSERT(!MUTEX_HELD(&zio->io_lock)); - if ((zio->io_error == EIO || - !(zio->io_flags & ZIO_FLAG_SPECULATIVE)) && - zio->io_logical == zio) { - /* - * For root I/O requests, tell the SPA to log the error - * appropriately. Also, generate a logical data - * ereport. - */ - spa_log_error(zio->io_spa, zio); + while (((1U << ++stage) & pipeline) == 0) + continue; - zfs_ereport_post(FM_EREPORT_ZFS_DATA, - zio->io_spa, NULL, zio, 0, 0); - } + ASSERT(stage <= ZIO_STAGE_DONE); + ASSERT(zio->io_stall == NULL); /* - * For I/O requests that cannot fail, panic appropriately. + * If we are in interrupt context and this pipeline stage + * will grab a config lock that is held across I/O, + * issue async to avoid deadlock. */ - if (!(zio->io_flags & ZIO_FLAG_CANFAIL)) { - char *blkbuf; - - blkbuf = kmem_alloc(BP_SPRINTF_LEN, KM_NOSLEEP); - if (blkbuf) { - sprintf_blkptr(blkbuf, BP_SPRINTF_LEN, - bp ? bp : &zio->io_bp_copy); - } - panic("ZFS: %s (%s on %s off %llx: zio %p %s): error " - "%d", zio->io_error == ECKSUM ? - "bad checksum" : "I/O failure", - zio_type_name[zio->io_type], - vdev_description(vd), - (u_longlong_t)zio->io_offset, - zio, blkbuf ? blkbuf : "", zio->io_error); + if (((1U << stage) & ZIO_CONFIG_LOCK_BLOCKING_STAGES) && + zio->io_vd == NULL && + zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) { + zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE); + return; } + + zio->io_stage = stage; + rv = zio_pipeline[stage](zio); + + if (rv == ZIO_PIPELINE_STOP) + return; + + ASSERT(rv == ZIO_PIPELINE_CONTINUE); } - zio_clear_transform_stack(zio); +} - if (zio->io_done) - zio->io_done(zio); +/* + * ========================================================================== + * Initiate I/O, either sync or async + * ========================================================================== + */ +int +zio_wait(zio_t *zio) +{ + int error; - ASSERT(zio->io_delegate_list == NULL); - ASSERT(zio->io_delegate_next == NULL); + ASSERT(zio->io_stage == ZIO_STAGE_OPEN); + ASSERT(zio->io_executor == NULL); - if (pio != NULL) { - zio_t *next, *prev; + zio->io_waiter = curthread; - mutex_enter(&pio->io_lock); - next = zio->io_sibling_next; - prev = zio->io_sibling_prev; - if (next != NULL) - next->io_sibling_prev = prev; - if (prev != NULL) - prev->io_sibling_next = next; - if (pio->io_child == zio) - pio->io_child = next; - mutex_exit(&pio->io_lock); + zio_execute(zio); - zio_notify_parent(zio, ZIO_STAGE_WAIT_CHILDREN_DONE, - &pio->io_children_notdone); - } + mutex_enter(&zio->io_lock); + while (zio->io_executor != NULL) + cv_wait(&zio->io_cv, &zio->io_lock); + mutex_exit(&zio->io_lock); - /* - * Note: this I/O is now done, and will shortly be freed, so there is no - * need to clear this (or any other) flag. - */ - if (zio->io_flags & ZIO_FLAG_CONFIG_GRABBED) - spa_config_exit(spa, zio); + error = zio->io_error; + zio_destroy(zio); - if (zio->io_waiter != NULL) { - mutex_enter(&zio->io_lock); - ASSERT(zio->io_stage == ZIO_STAGE_DONE); - zio->io_stalled = zio->io_stage; - cv_broadcast(&zio->io_cv); - mutex_exit(&zio->io_lock); - } else { - cv_destroy(&zio->io_cv); - mutex_destroy(&zio->io_lock); - kmem_cache_free(zio_cache, zio); + return (error); +} + +void +zio_nowait(zio_t *zio) +{ + ASSERT(zio->io_executor == NULL); + + if (zio->io_parent == NULL && zio->io_child_type == ZIO_CHILD_LOGICAL) { + /* + * This is a logical async I/O with no parent to wait for it. + * Attach it to the pool's global async root zio so that + * spa_unload() has a way of waiting for async I/O to finish. + */ + spa_t *spa = zio->io_spa; + zio->io_async_root = B_TRUE; + mutex_enter(&spa->spa_async_root_lock); + spa->spa_async_root_count++; + mutex_exit(&spa->spa_async_root_lock); } + + zio_execute(zio); } /* * ========================================================================== - * Compression support + * Reexecute or suspend/resume failed I/O * ========================================================================== */ + static void -zio_write_compress(zio_t *zio) +zio_reexecute(zio_t *pio) { - int compress = zio->io_compress; - blkptr_t *bp = zio->io_bp; - void *cbuf; - uint64_t lsize = zio->io_size; - uint64_t csize = lsize; - uint64_t cbufsize = 0; - int pass; + zio_t *zio, *zio_next; - if (bp->blk_birth == zio->io_txg) { + pio->io_flags = pio->io_orig_flags; + pio->io_stage = pio->io_orig_stage; + pio->io_pipeline = pio->io_orig_pipeline; + pio->io_reexecute = 0; + pio->io_error = 0; + for (int c = 0; c < ZIO_CHILD_TYPES; c++) + pio->io_child_error[c] = 0; + + if (IO_IS_ALLOCATING(pio)) { /* - * We're rewriting an existing block, which means we're - * working on behalf of spa_sync(). For spa_sync() to - * converge, it must eventually be the case that we don't - * have to allocate new blocks. But compression changes - * the blocksize, which forces a reallocate, and makes - * convergence take longer. Therefore, after the first - * few passes, stop compressing to ensure convergence. + * Remember the failed bp so that the io_ready() callback + * can update its accounting upon reexecution. The block + * was already freed in zio_done(); we indicate this with + * a fill count of -1 so that zio_free() knows to skip it. */ - pass = spa_sync_pass(zio->io_spa); - if (pass > zio_sync_pass.zp_dontcompress) - compress = ZIO_COMPRESS_OFF; - } else { - ASSERT(BP_IS_HOLE(bp)); - pass = 1; + blkptr_t *bp = pio->io_bp; + ASSERT(bp->blk_birth == 0 || bp->blk_birth == pio->io_txg); + bp->blk_fill = BLK_FILL_ALREADY_FREED; + pio->io_bp_orig = *bp; + BP_ZERO(bp); } - if (compress != ZIO_COMPRESS_OFF) - if (!zio_compress_data(compress, zio->io_data, zio->io_size, - &cbuf, &csize, &cbufsize)) - compress = ZIO_COMPRESS_OFF; - - if (compress != ZIO_COMPRESS_OFF && csize != 0) - zio_push_transform(zio, cbuf, csize, cbufsize); + /* + * As we reexecute pio's children, new children could be created. + * New children go to the head of the io_child list, however, + * so we will (correctly) not reexecute them. The key is that + * the remainder of the io_child list, from 'zio_next' onward, + * cannot be affected by any side effects of reexecuting 'zio'. + */ + for (zio = pio->io_child; zio != NULL; zio = zio_next) { + zio_next = zio->io_sibling_next; + mutex_enter(&pio->io_lock); + pio->io_children[zio->io_child_type][ZIO_WAIT_READY]++; + pio->io_children[zio->io_child_type][ZIO_WAIT_DONE]++; + mutex_exit(&pio->io_lock); + zio_reexecute(zio); + } /* - * The final pass of spa_sync() must be all rewrites, but the first - * few passes offer a trade-off: allocating blocks defers convergence, - * but newly allocated blocks are sequential, so they can be written - * to disk faster. Therefore, we allow the first few passes of - * spa_sync() to reallocate new blocks, but force rewrites after that. - * There should only be a handful of blocks after pass 1 in any case. + * Now that all children have been reexecuted, execute the parent. */ - if (bp->blk_birth == zio->io_txg && BP_GET_PSIZE(bp) == csize && - pass > zio_sync_pass.zp_rewrite) { - ASSERT(csize != 0); - BP_SET_LSIZE(bp, lsize); - BP_SET_COMPRESS(bp, compress); - zio->io_pipeline = ZIO_REWRITE_PIPELINE; - } else { - if (bp->blk_birth == zio->io_txg) - BP_ZERO(bp); - if (csize == 0) { - BP_ZERO(bp); - zio->io_pipeline = ZIO_WAIT_FOR_CHILDREN_PIPELINE; - } else { - ASSERT3U(BP_GET_NDVAS(bp), ==, 0); - BP_SET_LSIZE(bp, lsize); - BP_SET_PSIZE(bp, csize); - BP_SET_COMPRESS(bp, compress); - zio->io_pipeline = ZIO_WRITE_ALLOCATE_PIPELINE; - } + zio_execute(pio); +} + +void +zio_suspend(spa_t *spa, zio_t *zio) +{ + if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC) + fm_panic("Pool '%s' has encountered an uncorrectable I/O " + "failure and the failure mode property for this pool " + "is set to panic.", spa_name(spa)); + + zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0); + + mutex_enter(&spa->spa_suspend_lock); + + if (spa->spa_suspend_zio_root == NULL) + spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL, 0); + + spa->spa_suspended = B_TRUE; + + if (zio != NULL) { + ASSERT(zio != spa->spa_suspend_zio_root); + ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); + ASSERT(zio->io_parent == NULL); + ASSERT(zio->io_stage == ZIO_STAGE_DONE); + zio_add_child(spa->spa_suspend_zio_root, zio); } - zio_next_stage(zio); + mutex_exit(&spa->spa_suspend_lock); } -static void -zio_read_decompress(zio_t *zio) +void +zio_resume(spa_t *spa) { - blkptr_t *bp = zio->io_bp; - void *data; - uint64_t size; - uint64_t bufsize; - int compress = BP_GET_COMPRESS(bp); + zio_t *pio, *zio; - ASSERT(compress != ZIO_COMPRESS_OFF); + /* + * Reexecute all previously suspended i/o. + */ + mutex_enter(&spa->spa_suspend_lock); + spa->spa_suspended = B_FALSE; + cv_broadcast(&spa->spa_suspend_cv); + pio = spa->spa_suspend_zio_root; + spa->spa_suspend_zio_root = NULL; + mutex_exit(&spa->spa_suspend_lock); + + if (pio == NULL) + return; - zio_pop_transform(zio, &data, &size, &bufsize); + while ((zio = pio->io_child) != NULL) { + zio_remove_child(pio, zio); + zio->io_parent = NULL; + zio_reexecute(zio); + } - if (zio_decompress_data(compress, data, size, - zio->io_data, zio->io_size)) - zio->io_error = EIO; + ASSERT(pio->io_children[ZIO_CHILD_LOGICAL][ZIO_WAIT_DONE] == 0); - zio_buf_free(data, bufsize); + (void) zio_wait(pio); +} - zio_next_stage(zio); +void +zio_resume_wait(spa_t *spa) +{ + mutex_enter(&spa->spa_suspend_lock); + while (spa_suspended(spa)) + cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock); + mutex_exit(&spa->spa_suspend_lock); } /* * ========================================================================== - * Gang block support + * Gang blocks. + * + * A gang block is a collection of small blocks that looks to the DMU + * like one large block. When zio_dva_allocate() cannot find a block + * of the requested size, due to either severe fragmentation or the pool + * being nearly full, it calls zio_write_gang_block() to construct the + * block from smaller fragments. + * + * A gang block consists of a gang header (zio_gbh_phys_t) and up to + * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like + * an indirect block: it's an array of block pointers. It consumes + * only one sector and hence is allocatable regardless of fragmentation. + * The gang header's bps point to its gang members, which hold the data. + * + * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg> + * as the verifier to ensure uniqueness of the SHA256 checksum. + * Critically, the gang block bp's blk_cksum is the checksum of the data, + * not the gang header. This ensures that data block signatures (needed for + * deduplication) are independent of how the block is physically stored. + * + * Gang blocks can be nested: a gang member may itself be a gang block. + * Thus every gang block is a tree in which root and all interior nodes are + * gang headers, and the leaves are normal blocks that contain user data. + * The root of the gang tree is called the gang leader. + * + * To perform any operation (read, rewrite, free, claim) on a gang block, + * zio_gang_assemble() first assembles the gang tree (minus data leaves) + * in the io_gang_tree field of the original logical i/o by recursively + * reading the gang leader and all gang headers below it. This yields + * an in-core tree containing the contents of every gang header and the + * bps for every constituent of the gang block. + * + * With the gang tree now assembled, zio_gang_issue() just walks the gang tree + * and invokes a callback on each bp. To free a gang block, zio_gang_issue() + * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp. + * zio_claim_gang() provides a similarly trivial wrapper for zio_claim(). + * zio_read_gang() is a wrapper around zio_read() that omits reading gang + * headers, since we already have those in io_gang_tree. zio_rewrite_gang() + * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite() + * of the gang header plus zio_checksum_compute() of the data to update the + * gang header's blk_cksum as described above. + * + * The two-phase assemble/issue model solves the problem of partial failure -- + * what if you'd freed part of a gang block but then couldn't read the + * gang header for another part? Assembling the entire gang tree first + * ensures that all the necessary gang header I/O has succeeded before + * starting the actual work of free, claim, or write. Once the gang tree + * is assembled, free and claim are in-memory operations that cannot fail. + * + * In the event that a gang write fails, zio_dva_unallocate() walks the + * gang tree to immediately free (i.e. insert back into the space map) + * everything we've allocated. This ensures that we don't get ENOSPC + * errors during repeated suspend/resume cycles due to a flaky device. + * + * Gang rewrites only happen during sync-to-convergence. If we can't assemble + * the gang tree, we won't modify the block, so we can safely defer the free + * (knowing that the block is still intact). If we *can* assemble the gang + * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free + * each constituent bp and we can allocate a new block on the next sync pass. + * + * In all cases, the gang tree allows complete recovery from partial failure. * ========================================================================== */ -static void -zio_gang_pipeline(zio_t *zio) + +static zio_t * +zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data) { - /* - * By default, the pipeline assumes that we're dealing with a gang - * block. If we're not, strip out any gang-specific stages. - */ - if (!BP_IS_GANG(zio->io_bp)) - zio->io_pipeline &= ~ZIO_GANG_STAGES; + if (gn != NULL) + return (pio); - zio_next_stage(zio); + return (zio_read(pio, pio->io_spa, bp, data, BP_GET_PSIZE(bp), + NULL, NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), + &pio->io_bookmark)); } -static void -zio_gang_byteswap(zio_t *zio) +zio_t * +zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data) { - ASSERT(zio->io_size == SPA_GANGBLOCKSIZE); + zio_t *zio; - if (BP_SHOULD_BYTESWAP(zio->io_bp)) - byteswap_uint64_array(zio->io_data, zio->io_size); + if (gn != NULL) { + zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp, + gn->gn_gbh, SPA_GANGBLOCKSIZE, NULL, NULL, pio->io_priority, + ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark); + /* + * As we rewrite each gang header, the pipeline will compute + * a new gang block header checksum for it; but no one will + * compute a new data checksum, so we do that here. The one + * exception is the gang leader: the pipeline already computed + * its data checksum because that stage precedes gang assembly. + * (Presently, nothing actually uses interior data checksums; + * this is just good hygiene.) + */ + if (gn != pio->io_logical->io_gang_tree) { + zio_checksum_compute(zio, BP_GET_CHECKSUM(bp), + data, BP_GET_PSIZE(bp)); + } + } else { + zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp, + data, BP_GET_PSIZE(bp), NULL, NULL, pio->io_priority, + ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark); + } + + return (zio); } -static void -zio_get_gang_header(zio_t *zio) +/* ARGSUSED */ +zio_t * +zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data) { - blkptr_t *bp = zio->io_bp; - uint64_t gsize = SPA_GANGBLOCKSIZE; - void *gbuf = zio_buf_alloc(gsize); + return (zio_free(pio, pio->io_spa, pio->io_txg, bp, + NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio))); +} - ASSERT(BP_IS_GANG(bp)); +/* ARGSUSED */ +zio_t * +zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data) +{ + return (zio_claim(pio, pio->io_spa, pio->io_txg, bp, + NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio))); +} - zio_push_transform(zio, gbuf, gsize, gsize); +static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = { + NULL, + zio_read_gang, + zio_rewrite_gang, + zio_free_gang, + zio_claim_gang, + NULL +}; - zio_nowait(zio_create(zio, zio->io_spa, bp->blk_birth, bp, gbuf, gsize, - NULL, NULL, ZIO_TYPE_READ, zio->io_priority, - zio->io_flags & ZIO_FLAG_GANG_INHERIT, - ZIO_STAGE_OPEN, ZIO_READ_PIPELINE)); +static void zio_gang_tree_assemble_done(zio_t *zio); - zio_wait_children_done(zio); +static zio_gang_node_t * +zio_gang_node_alloc(zio_gang_node_t **gnpp) +{ + zio_gang_node_t *gn; + + ASSERT(*gnpp == NULL); + + gn = kmem_zalloc(sizeof (*gn), KM_SLEEP); + gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE); + *gnpp = gn; + + return (gn); } static void -zio_read_gang_members(zio_t *zio) +zio_gang_node_free(zio_gang_node_t **gnpp) { - zio_gbh_phys_t *gbh; - uint64_t gsize, gbufsize, loff, lsize; - int i; + zio_gang_node_t *gn = *gnpp; - ASSERT(BP_IS_GANG(zio->io_bp)); + for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) + ASSERT(gn->gn_child[g] == NULL); - zio_gang_byteswap(zio); - zio_pop_transform(zio, (void **)&gbh, &gsize, &gbufsize); + zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE); + kmem_free(gn, sizeof (*gn)); + *gnpp = NULL; +} - for (loff = 0, i = 0; loff != zio->io_size; loff += lsize, i++) { - blkptr_t *gbp = &gbh->zg_blkptr[i]; - lsize = BP_GET_PSIZE(gbp); +static void +zio_gang_tree_free(zio_gang_node_t **gnpp) +{ + zio_gang_node_t *gn = *gnpp; - ASSERT(BP_GET_COMPRESS(gbp) == ZIO_COMPRESS_OFF); - ASSERT3U(lsize, ==, BP_GET_LSIZE(gbp)); - ASSERT3U(loff + lsize, <=, zio->io_size); - ASSERT(i < SPA_GBH_NBLKPTRS); - ASSERT(!BP_IS_HOLE(gbp)); + if (gn == NULL) + return; - zio_nowait(zio_read(zio, zio->io_spa, gbp, - (char *)zio->io_data + loff, lsize, NULL, NULL, - zio->io_priority, zio->io_flags & ZIO_FLAG_GANG_INHERIT, - &zio->io_bookmark)); - } + for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) + zio_gang_tree_free(&gn->gn_child[g]); - zio_buf_free(gbh, gbufsize); - zio_wait_children_done(zio); + zio_gang_node_free(gnpp); } static void -zio_rewrite_gang_members(zio_t *zio) +zio_gang_tree_assemble(zio_t *lio, blkptr_t *bp, zio_gang_node_t **gnpp) { - zio_gbh_phys_t *gbh; - uint64_t gsize, gbufsize, loff, lsize; - int i; + zio_gang_node_t *gn = zio_gang_node_alloc(gnpp); - ASSERT(BP_IS_GANG(zio->io_bp)); - ASSERT3U(zio->io_size, ==, SPA_GANGBLOCKSIZE); + ASSERT(lio->io_logical == lio); + ASSERT(BP_IS_GANG(bp)); + + zio_nowait(zio_read(lio, lio->io_spa, bp, gn->gn_gbh, + SPA_GANGBLOCKSIZE, zio_gang_tree_assemble_done, gn, + lio->io_priority, ZIO_GANG_CHILD_FLAGS(lio), &lio->io_bookmark)); +} - zio_gang_byteswap(zio); - zio_pop_transform(zio, (void **)&gbh, &gsize, &gbufsize); +static void +zio_gang_tree_assemble_done(zio_t *zio) +{ + zio_t *lio = zio->io_logical; + zio_gang_node_t *gn = zio->io_private; + blkptr_t *bp = zio->io_bp; - ASSERT(gsize == gbufsize); + ASSERT(zio->io_parent == lio); + ASSERT(zio->io_child == NULL); - for (loff = 0, i = 0; loff != zio->io_size; loff += lsize, i++) { - blkptr_t *gbp = &gbh->zg_blkptr[i]; - lsize = BP_GET_PSIZE(gbp); + if (zio->io_error) + return; - ASSERT(BP_GET_COMPRESS(gbp) == ZIO_COMPRESS_OFF); - ASSERT3U(lsize, ==, BP_GET_LSIZE(gbp)); - ASSERT3U(loff + lsize, <=, zio->io_size); - ASSERT(i < SPA_GBH_NBLKPTRS); - ASSERT(!BP_IS_HOLE(gbp)); + if (BP_SHOULD_BYTESWAP(bp)) + byteswap_uint64_array(zio->io_data, zio->io_size); - zio_nowait(zio_rewrite(zio, zio->io_spa, zio->io_checksum, - zio->io_txg, gbp, (char *)zio->io_data + loff, lsize, - NULL, NULL, zio->io_priority, zio->io_flags, - &zio->io_bookmark)); - } + ASSERT(zio->io_data == gn->gn_gbh); + ASSERT(zio->io_size == SPA_GANGBLOCKSIZE); + ASSERT(gn->gn_gbh->zg_tail.zbt_magic == ZBT_MAGIC); - zio_push_transform(zio, gbh, gsize, gbufsize); - zio_wait_children_ready(zio); + for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) { + blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g]; + if (!BP_IS_GANG(gbp)) + continue; + zio_gang_tree_assemble(lio, gbp, &gn->gn_child[g]); + } } static void -zio_free_gang_members(zio_t *zio) +zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, void *data) { - zio_gbh_phys_t *gbh; - uint64_t gsize, gbufsize; - int i; + zio_t *lio = pio->io_logical; + zio_t *zio; - ASSERT(BP_IS_GANG(zio->io_bp)); + ASSERT(BP_IS_GANG(bp) == !!gn); + ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(lio->io_bp)); + ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == lio->io_gang_tree); - zio_gang_byteswap(zio); - zio_pop_transform(zio, (void **)&gbh, &gsize, &gbufsize); + /* + * If you're a gang header, your data is in gn->gn_gbh. + * If you're a gang member, your data is in 'data' and gn == NULL. + */ + zio = zio_gang_issue_func[lio->io_type](pio, bp, gn, data); - for (i = 0; i < SPA_GBH_NBLKPTRS; i++) { - blkptr_t *gbp = &gbh->zg_blkptr[i]; + if (gn != NULL) { + ASSERT(gn->gn_gbh->zg_tail.zbt_magic == ZBT_MAGIC); - if (BP_IS_HOLE(gbp)) - continue; - zio_nowait(zio_free(zio, zio->io_spa, zio->io_txg, - gbp, NULL, NULL)); + for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) { + blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g]; + if (BP_IS_HOLE(gbp)) + continue; + zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data); + data = (char *)data + BP_GET_PSIZE(gbp); + } } - zio_buf_free(gbh, gbufsize); - zio_next_stage(zio); + if (gn == lio->io_gang_tree) + ASSERT3P((char *)lio->io_data + lio->io_size, ==, data); + + if (zio != pio) + zio_nowait(zio); } -static void -zio_claim_gang_members(zio_t *zio) +static int +zio_gang_assemble(zio_t *zio) { - zio_gbh_phys_t *gbh; - uint64_t gsize, gbufsize; - int i; + blkptr_t *bp = zio->io_bp; - ASSERT(BP_IS_GANG(zio->io_bp)); + ASSERT(BP_IS_GANG(bp) && zio == zio->io_logical); - zio_gang_byteswap(zio); - zio_pop_transform(zio, (void **)&gbh, &gsize, &gbufsize); + zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree); - for (i = 0; i < SPA_GBH_NBLKPTRS; i++) { - blkptr_t *gbp = &gbh->zg_blkptr[i]; - if (BP_IS_HOLE(gbp)) - continue; - zio_nowait(zio_claim(zio, zio->io_spa, zio->io_txg, - gbp, NULL, NULL)); - } + return (ZIO_PIPELINE_CONTINUE); +} + +static int +zio_gang_issue(zio_t *zio) +{ + zio_t *lio = zio->io_logical; + blkptr_t *bp = zio->io_bp; + + if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE)) + return (ZIO_PIPELINE_STOP); + + ASSERT(BP_IS_GANG(bp) && zio == lio); + + if (zio->io_child_error[ZIO_CHILD_GANG] == 0) + zio_gang_tree_issue(lio, lio->io_gang_tree, bp, lio->io_data); + else + zio_gang_tree_free(&lio->io_gang_tree); + + zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; - zio_buf_free(gbh, gbufsize); - zio_next_stage(zio); + return (ZIO_PIPELINE_CONTINUE); } static void -zio_write_allocate_gang_member_done(zio_t *zio) +zio_write_gang_member_ready(zio_t *zio) { zio_t *pio = zio->io_parent; + zio_t *lio = zio->io_logical; dva_t *cdva = zio->io_bp->blk_dva; dva_t *pdva = pio->io_bp->blk_dva; uint64_t asize; - int d; - ASSERT3U(pio->io_ndvas, ==, zio->io_ndvas); + if (BP_IS_HOLE(zio->io_bp)) + return; + + ASSERT(BP_IS_HOLE(&zio->io_bp_orig)); + + ASSERT(zio->io_child_type == ZIO_CHILD_GANG); + ASSERT3U(zio->io_prop.zp_ndvas, ==, lio->io_prop.zp_ndvas); + ASSERT3U(zio->io_prop.zp_ndvas, <=, BP_GET_NDVAS(zio->io_bp)); + ASSERT3U(pio->io_prop.zp_ndvas, <=, BP_GET_NDVAS(pio->io_bp)); ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp)); - ASSERT3U(zio->io_ndvas, <=, BP_GET_NDVAS(zio->io_bp)); - ASSERT3U(pio->io_ndvas, <=, BP_GET_NDVAS(pio->io_bp)); mutex_enter(&pio->io_lock); - for (d = 0; d < BP_GET_NDVAS(pio->io_bp); d++) { + for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) { ASSERT(DVA_GET_GANG(&pdva[d])); asize = DVA_GET_ASIZE(&pdva[d]); asize += DVA_GET_ASIZE(&cdva[d]); @@ -1265,97 +1487,77 @@ zio_write_allocate_gang_member_done(zio_t *zio) mutex_exit(&pio->io_lock); } -static void -zio_write_allocate_gang_members(zio_t *zio) +static int +zio_write_gang_block(zio_t *pio) { - blkptr_t *bp = zio->io_bp; - dva_t *dva = bp->blk_dva; - spa_t *spa = zio->io_spa; + spa_t *spa = pio->io_spa; + blkptr_t *bp = pio->io_bp; + zio_t *lio = pio->io_logical; + zio_t *zio; + zio_gang_node_t *gn, **gnpp; zio_gbh_phys_t *gbh; - uint64_t txg = zio->io_txg; - uint64_t resid = zio->io_size; - uint64_t maxalloc = P2ROUNDUP(zio->io_size >> 1, SPA_MINBLOCKSIZE); - uint64_t gsize, loff, lsize; - uint32_t gbps_left; - int ndvas = zio->io_ndvas; + uint64_t txg = pio->io_txg; + uint64_t resid = pio->io_size; + uint64_t lsize; + int ndvas = lio->io_prop.zp_ndvas; int gbh_ndvas = MIN(ndvas + 1, spa_max_replication(spa)); + zio_prop_t zp; int error; - int i, d; - - gsize = SPA_GANGBLOCKSIZE; - gbps_left = SPA_GBH_NBLKPTRS; - - error = metaslab_alloc(spa, gsize, bp, gbh_ndvas, txg, NULL, B_FALSE); - if (error == ENOSPC) - panic("can't allocate gang block header"); - ASSERT(error == 0); - - for (d = 0; d < gbh_ndvas; d++) - DVA_SET_GANG(&dva[d], 1); - - bp->blk_birth = txg; - - gbh = zio_buf_alloc(gsize); - bzero(gbh, gsize); - /* We need to test multi-level gang blocks */ - if (maxalloc >= zio_gang_bang && (LBOLT & 0x1) == 0) - maxalloc = MAX(maxalloc >> 2, SPA_MINBLOCKSIZE); + error = metaslab_alloc(spa, spa->spa_normal_class, SPA_GANGBLOCKSIZE, + bp, gbh_ndvas, txg, pio == lio ? NULL : lio->io_bp, + METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER); + if (error) { + pio->io_error = error; + return (ZIO_PIPELINE_CONTINUE); + } - for (loff = 0, i = 0; loff != zio->io_size; - loff += lsize, resid -= lsize, gbps_left--, i++) { - blkptr_t *gbp = &gbh->zg_blkptr[i]; - dva = gbp->blk_dva; + if (pio == lio) { + gnpp = &lio->io_gang_tree; + } else { + gnpp = pio->io_private; + ASSERT(pio->io_ready == zio_write_gang_member_ready); + } - ASSERT(gbps_left != 0); - maxalloc = MIN(maxalloc, resid); + gn = zio_gang_node_alloc(gnpp); + gbh = gn->gn_gbh; + bzero(gbh, SPA_GANGBLOCKSIZE); - while (resid <= maxalloc * gbps_left) { - error = metaslab_alloc(spa, maxalloc, gbp, ndvas, - txg, bp, B_FALSE); - if (error == 0) - break; - ASSERT3U(error, ==, ENOSPC); - if (maxalloc == SPA_MINBLOCKSIZE) - panic("really out of space"); - maxalloc = P2ROUNDUP(maxalloc >> 1, SPA_MINBLOCKSIZE); - } + /* + * Create the gang header. + */ + zio = zio_rewrite(pio, spa, txg, bp, gbh, SPA_GANGBLOCKSIZE, NULL, NULL, + pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark); - if (resid <= maxalloc * gbps_left) { - lsize = maxalloc; - BP_SET_LSIZE(gbp, lsize); - BP_SET_PSIZE(gbp, lsize); - BP_SET_COMPRESS(gbp, ZIO_COMPRESS_OFF); - gbp->blk_birth = txg; - zio_nowait(zio_rewrite(zio, spa, - zio->io_checksum, txg, gbp, - (char *)zio->io_data + loff, lsize, - zio_write_allocate_gang_member_done, NULL, - zio->io_priority, zio->io_flags, - &zio->io_bookmark)); - } else { - lsize = P2ROUNDUP(resid / gbps_left, SPA_MINBLOCKSIZE); - ASSERT(lsize != SPA_MINBLOCKSIZE); - zio_nowait(zio_write_allocate(zio, spa, - zio->io_checksum, txg, gbp, - (char *)zio->io_data + loff, lsize, - zio_write_allocate_gang_member_done, NULL, - zio->io_priority, zio->io_flags)); - } + /* + * Create and nowait the gang children. + */ + for (int g = 0; resid != 0; resid -= lsize, g++) { + lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g), + SPA_MINBLOCKSIZE); + ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid); + + zp.zp_checksum = lio->io_prop.zp_checksum; + zp.zp_compress = ZIO_COMPRESS_OFF; + zp.zp_type = DMU_OT_NONE; + zp.zp_level = 0; + zp.zp_ndvas = lio->io_prop.zp_ndvas; + + zio_nowait(zio_write(zio, spa, txg, &gbh->zg_blkptr[g], + (char *)pio->io_data + (pio->io_size - resid), lsize, &zp, + zio_write_gang_member_ready, NULL, &gn->gn_child[g], + pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), + &pio->io_bookmark)); } - ASSERT(resid == 0 && loff == zio->io_size); - - zio->io_pipeline |= 1U << ZIO_STAGE_GANG_CHECKSUM_GENERATE; - - zio_push_transform(zio, gbh, gsize, gsize); /* - * As much as we'd like this to be zio_wait_children_ready(), - * updating our ASIZE doesn't happen until the io_done callback, - * so we have to wait for that to finish in order for our BP - * to be stable. + * Set pio's pipeline to just wait for zio to finish. */ - zio_wait_children_done(zio); + pio->io_pipeline = ZIO_INTERLOCK_PIPELINE; + + zio_nowait(zio); + + return (ZIO_PIPELINE_CONTINUE); } /* @@ -1363,59 +1565,139 @@ zio_write_allocate_gang_members(zio_t *zio) * Allocate and free blocks * ========================================================================== */ -static void + +static int zio_dva_allocate(zio_t *zio) { + spa_t *spa = zio->io_spa; + metaslab_class_t *mc = spa->spa_normal_class; blkptr_t *bp = zio->io_bp; int error; ASSERT(BP_IS_HOLE(bp)); ASSERT3U(BP_GET_NDVAS(bp), ==, 0); - ASSERT3U(zio->io_ndvas, >, 0); - ASSERT3U(zio->io_ndvas, <=, spa_max_replication(zio->io_spa)); - - /* For testing, make some blocks above a certain size be gang blocks */ - if (zio->io_size >= zio_gang_bang && (LBOLT & 0x3) == 0) { - zio_write_allocate_gang_members(zio); - return; - } - + ASSERT3U(zio->io_prop.zp_ndvas, >, 0); + ASSERT3U(zio->io_prop.zp_ndvas, <=, spa_max_replication(spa)); ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp)); - error = metaslab_alloc(zio->io_spa, zio->io_size, bp, zio->io_ndvas, - zio->io_txg, NULL, B_FALSE); + error = metaslab_alloc(spa, mc, zio->io_size, bp, + zio->io_prop.zp_ndvas, zio->io_txg, NULL, 0); - if (error == 0) { - bp->blk_birth = zio->io_txg; - } else if (error == ENOSPC) { - if (zio->io_size == SPA_MINBLOCKSIZE) - panic("really, truly out of space"); - zio_write_allocate_gang_members(zio); - return; - } else { + if (error) { + if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE) + return (zio_write_gang_block(zio)); zio->io_error = error; } - zio_next_stage(zio); + + return (ZIO_PIPELINE_CONTINUE); } -static void +static int zio_dva_free(zio_t *zio) { - blkptr_t *bp = zio->io_bp; + metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE); - metaslab_free(zio->io_spa, bp, zio->io_txg, B_FALSE); + return (ZIO_PIPELINE_CONTINUE); +} - BP_ZERO(bp); +static int +zio_dva_claim(zio_t *zio) +{ + int error; - zio_next_stage(zio); + error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg); + if (error) + zio->io_error = error; + + return (ZIO_PIPELINE_CONTINUE); } +/* + * Undo an allocation. This is used by zio_done() when an I/O fails + * and we want to give back the block we just allocated. + * This handles both normal blocks and gang blocks. + */ static void -zio_dva_claim(zio_t *zio) +zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp) { - zio->io_error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg); + spa_t *spa = zio->io_spa; + boolean_t now = !(zio->io_flags & ZIO_FLAG_IO_REWRITE); + + ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp)); + + if (zio->io_bp == bp && !now) { + /* + * This is a rewrite for sync-to-convergence. + * We can't do a metaslab_free(NOW) because bp wasn't allocated + * during this sync pass, which means that metaslab_sync() + * already committed the allocation. + */ + ASSERT(DVA_EQUAL(BP_IDENTITY(bp), + BP_IDENTITY(&zio->io_bp_orig))); + ASSERT(spa_sync_pass(spa) > 1); - zio_next_stage(zio); + if (BP_IS_GANG(bp) && gn == NULL) { + /* + * This is a gang leader whose gang header(s) we + * couldn't read now, so defer the free until later. + * The block should still be intact because without + * the headers, we'd never even start the rewrite. + */ + bplist_enqueue_deferred(&spa->spa_sync_bplist, bp); + return; + } + } + + if (!BP_IS_HOLE(bp)) + metaslab_free(spa, bp, bp->blk_birth, now); + + if (gn != NULL) { + for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) { + zio_dva_unallocate(zio, gn->gn_child[g], + &gn->gn_gbh->zg_blkptr[g]); + } + } +} + +/* + * Try to allocate an intent log block. Return 0 on success, errno on failure. + */ +int +zio_alloc_blk(spa_t *spa, uint64_t size, blkptr_t *new_bp, blkptr_t *old_bp, + uint64_t txg) +{ + int error; + + error = metaslab_alloc(spa, spa->spa_log_class, size, + new_bp, 1, txg, old_bp, METASLAB_HINTBP_AVOID); + + if (error) + error = metaslab_alloc(spa, spa->spa_normal_class, size, + new_bp, 1, txg, old_bp, METASLAB_HINTBP_AVOID); + + if (error == 0) { + BP_SET_LSIZE(new_bp, size); + BP_SET_PSIZE(new_bp, size); + BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF); + BP_SET_CHECKSUM(new_bp, ZIO_CHECKSUM_ZILOG); + BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG); + BP_SET_LEVEL(new_bp, 0); + BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER); + } + + return (error); +} + +/* + * Free an intent log block. We know it can't be a gang block, so there's + * nothing to do except metaslab_free() it. + */ +void +zio_free_blk(spa_t *spa, blkptr_t *bp, uint64_t txg) +{ + ASSERT(!BP_IS_GANG(bp)); + + metaslab_free(spa, bp, txg, B_FALSE); } /* @@ -1425,150 +1707,223 @@ zio_dva_claim(zio_t *zio) */ static void -zio_vdev_io_start(zio_t *zio) +zio_vdev_io_probe_done(zio_t *zio) +{ + zio_t *dio; + vdev_t *vd = zio->io_private; + + mutex_enter(&vd->vdev_probe_lock); + ASSERT(vd->vdev_probe_zio == zio); + vd->vdev_probe_zio = NULL; + mutex_exit(&vd->vdev_probe_lock); + + while ((dio = zio->io_delegate_list) != NULL) { + zio->io_delegate_list = dio->io_delegate_next; + dio->io_delegate_next = NULL; + if (!vdev_accessible(vd, dio)) + dio->io_error = ENXIO; + zio_execute(dio); + } +} + +/* + * Probe the device to determine whether I/O failure is specific to this + * zio (e.g. a bad sector) or affects the entire vdev (e.g. unplugged). + */ +static int +zio_vdev_io_probe(zio_t *zio) { vdev_t *vd = zio->io_vd; - vdev_t *tvd = vd ? vd->vdev_top : NULL; - blkptr_t *bp = zio->io_bp; - uint64_t align; + zio_t *pio = NULL; + boolean_t created_pio = B_FALSE; - if (vd == NULL) { - /* The mirror_ops handle multiple DVAs in a single BP */ - vdev_mirror_ops.vdev_op_io_start(zio); - return; + /* + * Don't probe the probe. + */ + if (zio->io_flags & ZIO_FLAG_PROBE) + return (ZIO_PIPELINE_CONTINUE); + + /* + * To prevent 'probe storms' when a device fails, we create + * just one probe i/o at a time. All zios that want to probe + * this vdev will join the probe zio's io_delegate_list. + */ + mutex_enter(&vd->vdev_probe_lock); + + if ((pio = vd->vdev_probe_zio) == NULL) { + vd->vdev_probe_zio = pio = zio_root(zio->io_spa, + zio_vdev_io_probe_done, vd, ZIO_FLAG_CANFAIL); + created_pio = B_TRUE; + vd->vdev_probe_wanted = B_TRUE; + spa_async_request(zio->io_spa, SPA_ASYNC_PROBE); } - align = 1ULL << tvd->vdev_ashift; + zio->io_delegate_next = pio->io_delegate_list; + pio->io_delegate_list = zio; + + mutex_exit(&vd->vdev_probe_lock); - if (zio->io_retries == 0 && vd == tvd) - zio->io_flags |= ZIO_FLAG_FAILFAST; + if (created_pio) { + zio_nowait(vdev_probe(vd, pio)); + zio_nowait(pio); + } + + return (ZIO_PIPELINE_STOP); +} + +static int +zio_vdev_io_start(zio_t *zio) +{ + vdev_t *vd = zio->io_vd; + uint64_t align; + spa_t *spa = zio->io_spa; + + ASSERT(zio->io_error == 0); + ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0); + + if (vd == NULL) { + if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER)) + spa_config_enter(spa, SCL_ZIO, zio, RW_READER); - if (!(zio->io_flags & ZIO_FLAG_PHYSICAL) && - vd->vdev_children == 0) { - zio->io_flags |= ZIO_FLAG_PHYSICAL; - zio->io_offset += VDEV_LABEL_START_SIZE; + /* + * The mirror_ops handle multiple DVAs in a single BP. + */ + return (vdev_mirror_ops.vdev_op_io_start(zio)); } + align = 1ULL << vd->vdev_top->vdev_ashift; + if (P2PHASE(zio->io_size, align) != 0) { uint64_t asize = P2ROUNDUP(zio->io_size, align); char *abuf = zio_buf_alloc(asize); - ASSERT(vd == tvd); + ASSERT(vd == vd->vdev_top); if (zio->io_type == ZIO_TYPE_WRITE) { bcopy(zio->io_data, abuf, zio->io_size); bzero(abuf + zio->io_size, asize - zio->io_size); } - zio_push_transform(zio, abuf, asize, asize); - ASSERT(!(zio->io_flags & ZIO_FLAG_SUBBLOCK)); - zio->io_flags |= ZIO_FLAG_SUBBLOCK; + zio_push_transform(zio, abuf, asize, asize, zio_subblock); } ASSERT(P2PHASE(zio->io_offset, align) == 0); ASSERT(P2PHASE(zio->io_size, align) == 0); - ASSERT(bp == NULL || - P2ROUNDUP(ZIO_GET_IOSIZE(zio), align) == zio->io_size); ASSERT(zio->io_type != ZIO_TYPE_WRITE || (spa_mode & FWRITE)); - vdev_io_start(zio); + if (vd->vdev_ops->vdev_op_leaf && + (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE)) { - /* zio_next_stage_async() gets called from io completion interrupt */ -} + if (zio->io_type == ZIO_TYPE_READ && vdev_cache_read(zio) == 0) + return (ZIO_PIPELINE_STOP); -static void -zio_vdev_io_done(zio_t *zio) -{ - if (zio->io_vd == NULL) - /* The mirror_ops handle multiple DVAs in a single BP */ - vdev_mirror_ops.vdev_op_io_done(zio); - else - vdev_io_done(zio); + if ((zio = vdev_queue_io(zio)) == NULL) + return (ZIO_PIPELINE_STOP); + + if (!vdev_accessible(vd, zio)) { + zio->io_error = ENXIO; + zio_interrupt(zio); + return (ZIO_PIPELINE_STOP); + } + + } + + return (vd->vdev_ops->vdev_op_io_start(zio)); } -/* XXPOLICY */ -boolean_t -zio_should_retry(zio_t *zio) +static int +zio_vdev_io_done(zio_t *zio) { vdev_t *vd = zio->io_vd; + vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops; + boolean_t unexpected_error = B_FALSE; - if (zio->io_error == 0) - return (B_FALSE); - if (zio->io_delegate_list != NULL) - return (B_FALSE); - if (vd && vd != vd->vdev_top) - return (B_FALSE); - if (zio->io_flags & ZIO_FLAG_DONT_RETRY) - return (B_FALSE); - if (zio->io_retries > 0) - return (B_FALSE); + if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE)) + return (ZIO_PIPELINE_STOP); - return (B_TRUE); + ASSERT(zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE); + + if (vd != NULL && vd->vdev_ops->vdev_op_leaf) { + + vdev_queue_io_done(zio); + + if (zio->io_type == ZIO_TYPE_WRITE) + vdev_cache_write(zio); + + if (zio_injection_enabled && zio->io_error == 0) + zio->io_error = zio_handle_device_injection(vd, EIO); + + if (zio_injection_enabled && zio->io_error == 0) + zio->io_error = zio_handle_label_injection(zio, EIO); + + if (zio->io_error) { + if (!vdev_accessible(vd, zio)) { + zio->io_error = ENXIO; + } else { + unexpected_error = B_TRUE; + } + } + } + + ops->vdev_op_io_done(zio); + + if (unexpected_error) + return (zio_vdev_io_probe(zio)); + + return (ZIO_PIPELINE_CONTINUE); } -static void +static int zio_vdev_io_assess(zio_t *zio) { vdev_t *vd = zio->io_vd; - vdev_t *tvd = vd ? vd->vdev_top : NULL; - - ASSERT(zio->io_vsd == NULL); - - if (zio->io_flags & ZIO_FLAG_SUBBLOCK) { - void *abuf; - uint64_t asize; - ASSERT(vd == tvd); - zio_pop_transform(zio, &abuf, &asize, &asize); - if (zio->io_type == ZIO_TYPE_READ) - bcopy(abuf, zio->io_data, zio->io_size); - zio_buf_free(abuf, asize); - zio->io_flags &= ~ZIO_FLAG_SUBBLOCK; + + if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE)) + return (ZIO_PIPELINE_STOP); + + if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER)) + spa_config_exit(zio->io_spa, SCL_ZIO, zio); + + if (zio->io_vsd != NULL) { + zio->io_vsd_free(zio); + zio->io_vsd = NULL; } - if (zio_injection_enabled && !zio->io_error) + if (zio_injection_enabled && zio->io_error == 0) zio->io_error = zio_handle_fault_injection(zio, EIO); /* * If the I/O failed, determine whether we should attempt to retry it. */ - /* XXPOLICY */ - if (zio_should_retry(zio)) { - ASSERT(tvd == vd); - - zio->io_retries++; + if (zio->io_error && vd == NULL && + !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) { + ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */ + ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */ zio->io_error = 0; - zio->io_flags &= ZIO_FLAG_VDEV_INHERIT | - ZIO_FLAG_CONFIG_GRABBED; - /* XXPOLICY */ - zio->io_flags &= ~ZIO_FLAG_FAILFAST; - zio->io_flags |= ZIO_FLAG_DONT_CACHE; + zio->io_flags |= ZIO_FLAG_IO_RETRY | + ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE; zio->io_stage = ZIO_STAGE_VDEV_IO_START - 1; + zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE); + return (ZIO_PIPELINE_STOP); + } - dprintf("retry #%d for %s to %s offset %llx\n", - zio->io_retries, zio_type_name[zio->io_type], - vdev_description(vd), zio->io_offset); + /* + * If we got an error on a leaf device, convert it to ENXIO + * if the device is not accessible at all. + */ + if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf && + !vdev_accessible(vd, zio)) + zio->io_error = ENXIO; - zio_next_stage_async(zio); - return; - } + /* + * If we can't write to an interior vdev (mirror or RAID-Z), + * set vdev_cant_write so that we stop trying to allocate from it. + */ + if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE && + vd != NULL && !vd->vdev_ops->vdev_op_leaf) + vd->vdev_cant_write = B_TRUE; - if (zio->io_error != 0 && zio->io_error != ECKSUM && - !(zio->io_flags & ZIO_FLAG_SPECULATIVE) && vd) { - /* - * Poor man's hotplug support. Even if we're done retrying this - * I/O, try to reopen the vdev to see if it's still attached. - * To avoid excessive thrashing, we only try it once a minute. - * This also has the effect of detecting when missing devices - * have come back, by polling the device once a minute. - * - * We need to do this asynchronously because we can't grab - * all the necessary locks way down here. - */ - if (gethrtime() - vd->vdev_last_try > 60ULL * NANOSEC) { - vd->vdev_last_try = gethrtime(); - tvd->vdev_reopen_wanted = 1; - spa_async_request(vd->vdev_spa, SPA_ASYNC_REOPEN); - } - } + if (zio->io_error) + zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; - zio_next_stage(zio); + return (ZIO_PIPELINE_CONTINUE); } void @@ -1603,49 +1958,63 @@ zio_vdev_io_bypass(zio_t *zio) * Generate and verify checksums * ========================================================================== */ -static void +static int zio_checksum_generate(zio_t *zio) { - int checksum = zio->io_checksum; blkptr_t *bp = zio->io_bp; + enum zio_checksum checksum; - ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp)); + if (bp == NULL) { + /* + * This is zio_write_phys(). + * We're either generating a label checksum, or none at all. + */ + checksum = zio->io_prop.zp_checksum; - BP_SET_CHECKSUM(bp, checksum); - BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER); + if (checksum == ZIO_CHECKSUM_OFF) + return (ZIO_PIPELINE_CONTINUE); - zio_checksum(checksum, &bp->blk_cksum, zio->io_data, zio->io_size); + ASSERT(checksum == ZIO_CHECKSUM_LABEL); + } else { + if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) { + ASSERT(!IO_IS_ALLOCATING(zio)); + checksum = ZIO_CHECKSUM_GANG_HEADER; + } else { + checksum = BP_GET_CHECKSUM(bp); + } + } - zio_next_stage(zio); + zio_checksum_compute(zio, checksum, zio->io_data, zio->io_size); + + return (ZIO_PIPELINE_CONTINUE); } -static void -zio_gang_checksum_generate(zio_t *zio) +static int +zio_checksum_verify(zio_t *zio) { - zio_cksum_t zc; - zio_gbh_phys_t *gbh = zio->io_data; - - ASSERT(BP_IS_GANG(zio->io_bp)); - ASSERT3U(zio->io_size, ==, SPA_GANGBLOCKSIZE); - - zio_set_gang_verifier(zio, &gbh->zg_tail.zbt_cksum); + blkptr_t *bp = zio->io_bp; + int error; - zio_checksum(ZIO_CHECKSUM_GANG_HEADER, &zc, zio->io_data, zio->io_size); + if (bp == NULL) { + /* + * This is zio_read_phys(). + * We're either verifying a label checksum, or nothing at all. + */ + if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF) + return (ZIO_PIPELINE_CONTINUE); - zio_next_stage(zio); -} + ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL); + } -static void -zio_checksum_verify(zio_t *zio) -{ - if (zio->io_bp != NULL) { - zio->io_error = zio_checksum_error(zio); - if (zio->io_error && !(zio->io_flags & ZIO_FLAG_SPECULATIVE)) + if ((error = zio_checksum_error(zio)) != 0) { + zio->io_error = error; + if (!(zio->io_flags & ZIO_FLAG_SPECULATIVE)) { zfs_ereport_post(FM_EREPORT_ZFS_CHECKSUM, zio->io_spa, zio->io_vd, zio, 0, 0); + } } - zio_next_stage(zio); + return (ZIO_PIPELINE_CONTINUE); } /* @@ -1658,204 +2027,263 @@ zio_checksum_verified(zio_t *zio) } /* - * Set the external verifier for a gang block based on stuff in the bp + * ========================================================================== + * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other. + * An error of 0 indictes success. ENXIO indicates whole-device failure, + * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO + * indicate errors that are specific to one I/O, and most likely permanent. + * Any other error is presumed to be worse because we weren't expecting it. + * ========================================================================== */ -void -zio_set_gang_verifier(zio_t *zio, zio_cksum_t *zcp) +int +zio_worst_error(int e1, int e2) { - blkptr_t *bp = zio->io_bp; + static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO }; + int r1, r2; - zcp->zc_word[0] = DVA_GET_VDEV(BP_IDENTITY(bp)); - zcp->zc_word[1] = DVA_GET_OFFSET(BP_IDENTITY(bp)); - zcp->zc_word[2] = bp->blk_birth; - zcp->zc_word[3] = 0; + for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++) + if (e1 == zio_error_rank[r1]) + break; + + for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++) + if (e2 == zio_error_rank[r2]) + break; + + return (r1 > r2 ? e1 : e2); } /* * ========================================================================== - * Define the pipeline + * I/O completion * ========================================================================== */ -typedef void zio_pipe_stage_t(zio_t *zio); - -static void -zio_badop(zio_t *zio) +static int +zio_ready(zio_t *zio) { - panic("Invalid I/O pipeline stage %u for zio %p", zio->io_stage, zio); -} - -zio_pipe_stage_t *zio_pipeline[ZIO_STAGE_DONE + 2] = { - zio_badop, - zio_wait_children_ready, - zio_write_compress, - zio_checksum_generate, - zio_gang_pipeline, - zio_get_gang_header, - zio_rewrite_gang_members, - zio_free_gang_members, - zio_claim_gang_members, - zio_dva_allocate, - zio_dva_free, - zio_dva_claim, - zio_gang_checksum_generate, - zio_ready, - zio_vdev_io_start, - zio_vdev_io_done, - zio_vdev_io_assess, - zio_wait_children_done, - zio_checksum_verify, - zio_read_gang_members, - zio_read_decompress, - zio_done, - zio_badop -}; + blkptr_t *bp = zio->io_bp; + zio_t *pio = zio->io_parent; -/* - * Move an I/O to the next stage of the pipeline and execute that stage. - * There's no locking on io_stage because there's no legitimate way for - * multiple threads to be attempting to process the same I/O. - */ -void -zio_next_stage(zio_t *zio) -{ - uint32_t pipeline = zio->io_pipeline; + if (zio->io_ready) { + if (BP_IS_GANG(bp) && + zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY)) + return (ZIO_PIPELINE_STOP); - ASSERT(!MUTEX_HELD(&zio->io_lock)); + ASSERT(IO_IS_ALLOCATING(zio)); + ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp)); + ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0); - if (zio->io_error) { - dprintf("zio %p vdev %s offset %llx stage %d error %d\n", - zio, vdev_description(zio->io_vd), - zio->io_offset, zio->io_stage, zio->io_error); - if (((1U << zio->io_stage) & ZIO_VDEV_IO_PIPELINE) == 0) - pipeline &= ZIO_ERROR_PIPELINE_MASK; + zio->io_ready(zio); } - while (((1U << ++zio->io_stage) & pipeline) == 0) - continue; + if (bp != NULL && bp != &zio->io_bp_copy) + zio->io_bp_copy = *bp; - ASSERT(zio->io_stage <= ZIO_STAGE_DONE); - ASSERT(zio->io_stalled == 0); + if (zio->io_error) + zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; - /* - * See the comment in zio_next_stage_async() about per-CPU taskqs. - */ - if (((1U << zio->io_stage) & zio->io_async_stages) && - (zio->io_stage == ZIO_STAGE_WRITE_COMPRESS) && - !(zio->io_flags & ZIO_FLAG_METADATA)) { - taskq_t *tq = zio->io_spa->spa_zio_issue_taskq[zio->io_type]; - (void) taskq_dispatch(tq, - (task_func_t *)zio_pipeline[zio->io_stage], zio, TQ_SLEEP); - } else { - zio_pipeline[zio->io_stage](zio); - } + if (pio != NULL) + zio_notify_parent(pio, zio, ZIO_WAIT_READY); + + return (ZIO_PIPELINE_CONTINUE); } -void -zio_next_stage_async(zio_t *zio) +static int +zio_done(zio_t *zio) { - taskq_t *tq; - uint32_t pipeline = zio->io_pipeline; - - ASSERT(!MUTEX_HELD(&zio->io_lock)); + spa_t *spa = zio->io_spa; + zio_t *pio = zio->io_parent; + zio_t *lio = zio->io_logical; + blkptr_t *bp = zio->io_bp; + vdev_t *vd = zio->io_vd; + uint64_t psize = zio->io_size; - if (zio->io_error) { - dprintf("zio %p vdev %s offset %llx stage %d error %d\n", - zio, vdev_description(zio->io_vd), - zio->io_offset, zio->io_stage, zio->io_error); - if (((1U << zio->io_stage) & ZIO_VDEV_IO_PIPELINE) == 0) - pipeline &= ZIO_ERROR_PIPELINE_MASK; - } + /* + * If our of children haven't all completed, + * wait for them and then repeat this pipeline stage. + */ + if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE) || + zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE) || + zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_DONE)) + return (ZIO_PIPELINE_STOP); - while (((1U << ++zio->io_stage) & pipeline) == 0) - continue; + for (int c = 0; c < ZIO_CHILD_TYPES; c++) + for (int w = 0; w < ZIO_WAIT_TYPES; w++) + ASSERT(zio->io_children[c][w] == 0); - ASSERT(zio->io_stage <= ZIO_STAGE_DONE); - ASSERT(zio->io_stalled == 0); + if (bp != NULL) { + ASSERT(bp->blk_pad[0] == 0); + ASSERT(bp->blk_pad[1] == 0); + ASSERT(bp->blk_pad[2] == 0); + ASSERT(bcmp(bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 || + (pio != NULL && bp == pio->io_bp)); + if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(bp) && + !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) { + ASSERT(!BP_SHOULD_BYTESWAP(bp)); + ASSERT3U(zio->io_prop.zp_ndvas, <=, BP_GET_NDVAS(bp)); + ASSERT(BP_COUNT_GANG(bp) == 0 || + (BP_COUNT_GANG(bp) == BP_GET_NDVAS(bp))); + } + } /* - * For performance, we'll probably want two sets of task queues: - * per-CPU issue taskqs and per-CPU completion taskqs. The per-CPU - * part is for read performance: since we have to make a pass over - * the data to checksum it anyway, we want to do this on the same CPU - * that issued the read, because (assuming CPU scheduling affinity) - * that thread is probably still there. Getting this optimization - * right avoids performance-hostile cache-to-cache transfers. - * - * Note that having two sets of task queues is also necessary for - * correctness: if all of the issue threads get bogged down waiting - * for dependent reads (e.g. metaslab freelist) to complete, then - * there won't be any threads available to service I/O completion - * interrupts. + * If there were child vdev or gang errors, they apply to us now. */ - if ((1U << zio->io_stage) & zio->io_async_stages) { - if (zio->io_stage < ZIO_STAGE_VDEV_IO_DONE) - tq = zio->io_spa->spa_zio_issue_taskq[zio->io_type]; - else - tq = zio->io_spa->spa_zio_intr_taskq[zio->io_type]; - (void) taskq_dispatch(tq, - (task_func_t *)zio_pipeline[zio->io_stage], zio, TQ_SLEEP); - } else { - zio_pipeline[zio->io_stage](zio); - } -} + zio_inherit_child_errors(zio, ZIO_CHILD_VDEV); + zio_inherit_child_errors(zio, ZIO_CHILD_GANG); -static boolean_t -zio_alloc_should_fail(void) -{ - static uint16_t allocs = 0; + zio_pop_transforms(zio); /* note: may set zio->io_error */ - return (P2PHASE(allocs++, 1U<<zio_zil_fail_shift) == 0); -} + vdev_stat_update(zio, psize); -/* - * Try to allocate an intent log block. Return 0 on success, errno on failure. - */ -int -zio_alloc_blk(spa_t *spa, uint64_t size, blkptr_t *new_bp, blkptr_t *old_bp, - uint64_t txg) -{ - int error; + if (zio->io_error) { + /* + * If this I/O is attached to a particular vdev, + * generate an error message describing the I/O failure + * at the block level. We ignore these errors if the + * device is currently unavailable. + */ + if (zio->io_error != ECKSUM && vd != NULL && !vdev_is_dead(vd)) + zfs_ereport_post(FM_EREPORT_ZFS_IO, spa, vd, zio, 0, 0); - spa_config_enter(spa, RW_READER, FTAG); + if ((zio->io_error == EIO || + !(zio->io_flags & ZIO_FLAG_SPECULATIVE)) && zio == lio) { + /* + * For logical I/O requests, tell the SPA to log the + * error and generate a logical data ereport. + */ + spa_log_error(spa, zio); + zfs_ereport_post(FM_EREPORT_ZFS_DATA, spa, NULL, zio, + 0, 0); + } + } - if (zio_zil_fail_shift && zio_alloc_should_fail()) { - spa_config_exit(spa, FTAG); - return (ENOSPC); + if (zio->io_error && zio == lio) { + /* + * Determine whether zio should be reexecuted. This will + * propagate all the way to the root via zio_notify_parent(). + */ + ASSERT(vd == NULL && bp != NULL); + + if (IO_IS_ALLOCATING(zio)) + if (zio->io_error != ENOSPC) + zio->io_reexecute |= ZIO_REEXECUTE_NOW; + else + zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND; + + if ((zio->io_type == ZIO_TYPE_READ || + zio->io_type == ZIO_TYPE_FREE) && + zio->io_error == ENXIO && + spa_get_failmode(spa) != ZIO_FAILURE_MODE_CONTINUE) + zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND; + + if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute) + zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND; } /* - * We were passed the previous log blocks dva_t in bp->blk_dva[0]. + * If there were logical child errors, they apply to us now. + * We defer this until now to avoid conflating logical child + * errors with errors that happened to the zio itself when + * updating vdev stats and reporting FMA events above. */ - error = metaslab_alloc(spa, size, new_bp, 1, txg, old_bp, B_TRUE); + zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL); - if (error == 0) { - BP_SET_LSIZE(new_bp, size); - BP_SET_PSIZE(new_bp, size); - BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF); - BP_SET_CHECKSUM(new_bp, ZIO_CHECKSUM_ZILOG); - BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG); - BP_SET_LEVEL(new_bp, 0); - BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER); - new_bp->blk_birth = txg; + if (zio->io_reexecute) { + /* + * This is a logical I/O that wants to reexecute. + * + * Reexecute is top-down. When an i/o fails, if it's not + * the root, it simply notifies its parent and sticks around. + * The parent, seeing that it still has children in zio_done(), + * does the same. This percolates all the way up to the root. + * The root i/o will reexecute or suspend the entire tree. + * + * This approach ensures that zio_reexecute() honors + * all the original i/o dependency relationships, e.g. + * parents not executing until children are ready. + */ + ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); + + if (IO_IS_ALLOCATING(zio)) + zio_dva_unallocate(zio, zio->io_gang_tree, bp); + + zio_gang_tree_free(&zio->io_gang_tree); + + if (pio != NULL) { + /* + * We're not a root i/o, so there's nothing to do + * but notify our parent. Don't propagate errors + * upward since we haven't permanently failed yet. + */ + zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE; + zio_notify_parent(pio, zio, ZIO_WAIT_DONE); + } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) { + /* + * We'd fail again if we reexecuted now, so suspend + * until conditions improve (e.g. device comes online). + */ + zio_suspend(spa, zio); + } else { + /* + * Reexecution is potentially a huge amount of work. + * Hand it off to the otherwise-unused claim taskq. + */ + (void) taskq_dispatch( + spa->spa_zio_taskq[ZIO_TYPE_CLAIM][ZIO_TASKQ_ISSUE], + (task_func_t *)zio_reexecute, zio, TQ_SLEEP); + } + return (ZIO_PIPELINE_STOP); } - spa_config_exit(spa, FTAG); + ASSERT(zio->io_child == NULL); + ASSERT(zio->io_reexecute == 0); + ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL)); - return (error); -} + if (zio->io_done) + zio->io_done(zio); -/* - * Free an intent log block. We know it can't be a gang block, so there's - * nothing to do except metaslab_free() it. - */ -void -zio_free_blk(spa_t *spa, blkptr_t *bp, uint64_t txg) -{ - ASSERT(!BP_IS_GANG(bp)); + zio_gang_tree_free(&zio->io_gang_tree); - spa_config_enter(spa, RW_READER, FTAG); + ASSERT(zio->io_delegate_list == NULL); + ASSERT(zio->io_delegate_next == NULL); - metaslab_free(spa, bp, txg, B_FALSE); + if (pio != NULL) { + zio_remove_child(pio, zio); + zio_notify_parent(pio, zio, ZIO_WAIT_DONE); + } + + if (zio->io_waiter != NULL) { + mutex_enter(&zio->io_lock); + zio->io_executor = NULL; + cv_broadcast(&zio->io_cv); + mutex_exit(&zio->io_lock); + } else { + zio_destroy(zio); + } - spa_config_exit(spa, FTAG); + return (ZIO_PIPELINE_STOP); } + +/* + * ========================================================================== + * I/O pipeline definition + * ========================================================================== + */ +static zio_pipe_stage_t *zio_pipeline[ZIO_STAGES] = { + NULL, + zio_issue_async, + zio_read_bp_init, + zio_write_bp_init, + zio_checksum_generate, + zio_gang_assemble, + zio_gang_issue, + zio_dva_allocate, + zio_dva_free, + zio_dva_claim, + zio_ready, + zio_vdev_io_start, + zio_vdev_io_done, + zio_vdev_io_assess, + zio_checksum_verify, + zio_done +}; |