1fa9e4066Sahrens /* 2fa9e4066Sahrens * CDDL HEADER START 3fa9e4066Sahrens * 4fa9e4066Sahrens * The contents of this file are subject to the terms of the 5ea8dc4b6Seschrock * Common Development and Distribution License (the "License"). 6ea8dc4b6Seschrock * You may not use this file except in compliance with the License. 7fa9e4066Sahrens * 8fa9e4066Sahrens * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 9fa9e4066Sahrens * or http://www.opensolaris.org/os/licensing. 10fa9e4066Sahrens * See the License for the specific language governing permissions 11fa9e4066Sahrens * and limitations under the License. 12fa9e4066Sahrens * 13fa9e4066Sahrens * When distributing Covered Code, include this CDDL HEADER in each 14fa9e4066Sahrens * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 15fa9e4066Sahrens * If applicable, add the following below this CDDL HEADER, with the 16fa9e4066Sahrens * fields enclosed by brackets "[]" replaced with your own identifying 17fa9e4066Sahrens * information: Portions Copyright [yyyy] [name of copyright owner] 18fa9e4066Sahrens * 19fa9e4066Sahrens * CDDL HEADER END 20fa9e4066Sahrens */ 21fa9e4066Sahrens /* 223f9d6ad7SLin Ling * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved. 23bf16b11eSMatthew Ahrens * Copyright (c) 2011, 2014 by Delphix. All rights reserved. 245aeb9474SGarrett D'Amore * Copyright (c) 2011 Nexenta Systems, Inc. All rights reserved. 25fa9e4066Sahrens */ 26fa9e4066Sahrens 27fa9e4066Sahrens #include <sys/zfs_context.h> 28ea8dc4b6Seschrock #include <sys/fm/fs/zfs.h> 29fa9e4066Sahrens #include <sys/spa.h> 30fa9e4066Sahrens #include <sys/txg.h> 31fa9e4066Sahrens #include <sys/spa_impl.h> 32fa9e4066Sahrens #include <sys/vdev_impl.h> 33fa9e4066Sahrens #include <sys/zio_impl.h> 34fa9e4066Sahrens #include <sys/zio_compress.h> 35fa9e4066Sahrens #include <sys/zio_checksum.h> 36b24ab676SJeff Bonwick #include <sys/dmu_objset.h> 37b24ab676SJeff Bonwick #include <sys/arc.h> 38b24ab676SJeff Bonwick #include <sys/ddt.h> 395d7b4d43SMatthew Ahrens #include <sys/blkptr.h> 4043466aaeSMax Grossman #include <sys/zfeature.h> 41fa9e4066Sahrens 42fa9e4066Sahrens /* 43fa9e4066Sahrens * ========================================================================== 44fa9e4066Sahrens * I/O type descriptions 45fa9e4066Sahrens * ========================================================================== 46fa9e4066Sahrens */ 4769962b56SMatthew Ahrens const char *zio_type_name[ZIO_TYPES] = { 4880eb36f2SGeorge Wilson "zio_null", "zio_read", "zio_write", "zio_free", "zio_claim", 4980eb36f2SGeorge Wilson "zio_ioctl" 5080eb36f2SGeorge Wilson }; 51fa9e4066Sahrens 52fa9e4066Sahrens /* 53fa9e4066Sahrens * ========================================================================== 54fa9e4066Sahrens * I/O kmem caches 55fa9e4066Sahrens * ========================================================================== 56fa9e4066Sahrens */ 57ccae0b50Seschrock kmem_cache_t *zio_cache; 58a3f829aeSBill Moore kmem_cache_t *zio_link_cache; 59fa9e4066Sahrens kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT]; 60ad23a2dbSjohansen kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT]; 61ad23a2dbSjohansen 62ad23a2dbSjohansen #ifdef _KERNEL 63ad23a2dbSjohansen extern vmem_t *zio_alloc_arena; 64ad23a2dbSjohansen #endif 65fa9e4066Sahrens 66*738f37bcSGeorge Wilson #define ZIO_PIPELINE_CONTINUE 0x100 67*738f37bcSGeorge Wilson #define ZIO_PIPELINE_STOP 0x101 68*738f37bcSGeorge Wilson 6901f55e48SGeorge Wilson /* 7001f55e48SGeorge Wilson * The following actions directly effect the spa's sync-to-convergence logic. 7101f55e48SGeorge Wilson * The values below define the sync pass when we start performing the action. 7201f55e48SGeorge Wilson * Care should be taken when changing these values as they directly impact 7301f55e48SGeorge Wilson * spa_sync() performance. Tuning these values may introduce subtle performance 7401f55e48SGeorge Wilson * pathologies and should only be done in the context of performance analysis. 7501f55e48SGeorge Wilson * These tunables will eventually be removed and replaced with #defines once 7601f55e48SGeorge Wilson * enough analysis has been done to determine optimal values. 7701f55e48SGeorge Wilson * 7801f55e48SGeorge Wilson * The 'zfs_sync_pass_deferred_free' pass must be greater than 1 to ensure that 7901f55e48SGeorge Wilson * regular blocks are not deferred. 8001f55e48SGeorge Wilson */ 8101f55e48SGeorge Wilson int zfs_sync_pass_deferred_free = 2; /* defer frees starting in this pass */ 8201f55e48SGeorge Wilson int zfs_sync_pass_dont_compress = 5; /* don't compress starting in this pass */ 8301f55e48SGeorge Wilson int zfs_sync_pass_rewrite = 2; /* rewrite new bps starting in this pass */ 8401f55e48SGeorge Wilson 850a4e9518Sgw /* 86e14bb325SJeff Bonwick * An allocating zio is one that either currently has the DVA allocate 87e14bb325SJeff Bonwick * stage set or will have it later in its lifetime. 880a4e9518Sgw */ 89b24ab676SJeff Bonwick #define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE) 90b24ab676SJeff Bonwick 9135a5a358SJonathan Adams boolean_t zio_requeue_io_start_cut_in_line = B_TRUE; 9235a5a358SJonathan Adams 93b24ab676SJeff Bonwick #ifdef ZFS_DEBUG 94b24ab676SJeff Bonwick int zio_buf_debug_limit = 16384; 95b24ab676SJeff Bonwick #else 96b24ab676SJeff Bonwick int zio_buf_debug_limit = 0; 97b24ab676SJeff Bonwick #endif 980a4e9518Sgw 99fa9e4066Sahrens void 100fa9e4066Sahrens zio_init(void) 101fa9e4066Sahrens { 102fa9e4066Sahrens size_t c; 103ad23a2dbSjohansen vmem_t *data_alloc_arena = NULL; 104ad23a2dbSjohansen 105ad23a2dbSjohansen #ifdef _KERNEL 106ad23a2dbSjohansen data_alloc_arena = zio_alloc_arena; 107ad23a2dbSjohansen #endif 108a3f829aeSBill Moore zio_cache = kmem_cache_create("zio_cache", 109a3f829aeSBill Moore sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0); 110a3f829aeSBill Moore zio_link_cache = kmem_cache_create("zio_link_cache", 111a3f829aeSBill Moore sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0); 112ccae0b50Seschrock 113fa9e4066Sahrens /* 114fa9e4066Sahrens * For small buffers, we want a cache for each multiple of 115fa9e4066Sahrens * SPA_MINBLOCKSIZE. For medium-size buffers, we want a cache 116fa9e4066Sahrens * for each quarter-power of 2. For large buffers, we want 117fa9e4066Sahrens * a cache for each multiple of PAGESIZE. 118fa9e4066Sahrens */ 119fa9e4066Sahrens for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) { 120fa9e4066Sahrens size_t size = (c + 1) << SPA_MINBLOCKSHIFT; 121fa9e4066Sahrens size_t p2 = size; 122fa9e4066Sahrens size_t align = 0; 123e291592aSJonathan Adams size_t cflags = (size > zio_buf_debug_limit) ? KMC_NODEBUG : 0; 124fa9e4066Sahrens 125fa9e4066Sahrens while (p2 & (p2 - 1)) 126fa9e4066Sahrens p2 &= p2 - 1; 127fa9e4066Sahrens 128cd1c8b85SMatthew Ahrens #ifndef _KERNEL 129cd1c8b85SMatthew Ahrens /* 130cd1c8b85SMatthew Ahrens * If we are using watchpoints, put each buffer on its own page, 131cd1c8b85SMatthew Ahrens * to eliminate the performance overhead of trapping to the 132cd1c8b85SMatthew Ahrens * kernel when modifying a non-watched buffer that shares the 133cd1c8b85SMatthew Ahrens * page with a watched buffer. 134cd1c8b85SMatthew Ahrens */ 135cd1c8b85SMatthew Ahrens if (arc_watch && !IS_P2ALIGNED(size, PAGESIZE)) 136cd1c8b85SMatthew Ahrens continue; 137cd1c8b85SMatthew Ahrens #endif 138fa9e4066Sahrens if (size <= 4 * SPA_MINBLOCKSIZE) { 139fa9e4066Sahrens align = SPA_MINBLOCKSIZE; 140cd1c8b85SMatthew Ahrens } else if (IS_P2ALIGNED(size, PAGESIZE)) { 141fa9e4066Sahrens align = PAGESIZE; 142cd1c8b85SMatthew Ahrens } else if (IS_P2ALIGNED(size, p2 >> 2)) { 143fa9e4066Sahrens align = p2 >> 2; 144fa9e4066Sahrens } 145fa9e4066Sahrens 146fa9e4066Sahrens if (align != 0) { 147ad23a2dbSjohansen char name[36]; 1485ad82045Snd (void) sprintf(name, "zio_buf_%lu", (ulong_t)size); 149fa9e4066Sahrens zio_buf_cache[c] = kmem_cache_create(name, size, 150e291592aSJonathan Adams align, NULL, NULL, NULL, NULL, NULL, cflags); 151ad23a2dbSjohansen 152e291592aSJonathan Adams /* 153e291592aSJonathan Adams * Since zio_data bufs do not appear in crash dumps, we 154e291592aSJonathan Adams * pass KMC_NOTOUCH so that no allocator metadata is 155e291592aSJonathan Adams * stored with the buffers. 156e291592aSJonathan Adams */ 157ad23a2dbSjohansen (void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size); 158ad23a2dbSjohansen zio_data_buf_cache[c] = kmem_cache_create(name, size, 159ad23a2dbSjohansen align, NULL, NULL, NULL, NULL, data_alloc_arena, 160e291592aSJonathan Adams cflags | KMC_NOTOUCH); 161fa9e4066Sahrens } 162fa9e4066Sahrens } 163fa9e4066Sahrens 164fa9e4066Sahrens while (--c != 0) { 165fa9e4066Sahrens ASSERT(zio_buf_cache[c] != NULL); 166fa9e4066Sahrens if (zio_buf_cache[c - 1] == NULL) 167fa9e4066Sahrens zio_buf_cache[c - 1] = zio_buf_cache[c]; 168ad23a2dbSjohansen 169ad23a2dbSjohansen ASSERT(zio_data_buf_cache[c] != NULL); 170ad23a2dbSjohansen if (zio_data_buf_cache[c - 1] == NULL) 171ad23a2dbSjohansen zio_data_buf_cache[c - 1] = zio_data_buf_cache[c]; 172fa9e4066Sahrens } 173ea8dc4b6Seschrock 174ea8dc4b6Seschrock zio_inject_init(); 175fa9e4066Sahrens } 176fa9e4066Sahrens 177fa9e4066Sahrens void 178fa9e4066Sahrens zio_fini(void) 179fa9e4066Sahrens { 180fa9e4066Sahrens size_t c; 181fa9e4066Sahrens kmem_cache_t *last_cache = NULL; 182ad23a2dbSjohansen kmem_cache_t *last_data_cache = NULL; 183fa9e4066Sahrens 184fa9e4066Sahrens for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) { 185fa9e4066Sahrens if (zio_buf_cache[c] != last_cache) { 186fa9e4066Sahrens last_cache = zio_buf_cache[c]; 187fa9e4066Sahrens kmem_cache_destroy(zio_buf_cache[c]); 188fa9e4066Sahrens } 189fa9e4066Sahrens zio_buf_cache[c] = NULL; 190ad23a2dbSjohansen 191ad23a2dbSjohansen if (zio_data_buf_cache[c] != last_data_cache) { 192ad23a2dbSjohansen last_data_cache = zio_data_buf_cache[c]; 193ad23a2dbSjohansen kmem_cache_destroy(zio_data_buf_cache[c]); 194ad23a2dbSjohansen } 195ad23a2dbSjohansen zio_data_buf_cache[c] = NULL; 196fa9e4066Sahrens } 197ea8dc4b6Seschrock 198a3f829aeSBill Moore kmem_cache_destroy(zio_link_cache); 199ccae0b50Seschrock kmem_cache_destroy(zio_cache); 200ccae0b50Seschrock 201ea8dc4b6Seschrock zio_inject_fini(); 202fa9e4066Sahrens } 203fa9e4066Sahrens 204fa9e4066Sahrens /* 205fa9e4066Sahrens * ========================================================================== 206fa9e4066Sahrens * Allocate and free I/O buffers 207fa9e4066Sahrens * ========================================================================== 208fa9e4066Sahrens */ 209ad23a2dbSjohansen 210ad23a2dbSjohansen /* 211ad23a2dbSjohansen * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a 212ad23a2dbSjohansen * crashdump if the kernel panics, so use it judiciously. Obviously, it's 213ad23a2dbSjohansen * useful to inspect ZFS metadata, but if possible, we should avoid keeping 214ad23a2dbSjohansen * excess / transient data in-core during a crashdump. 215ad23a2dbSjohansen */ 216fa9e4066Sahrens void * 217fa9e4066Sahrens zio_buf_alloc(size_t size) 218fa9e4066Sahrens { 219fa9e4066Sahrens size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; 220fa9e4066Sahrens 2215d7b4d43SMatthew Ahrens ASSERT3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); 222fa9e4066Sahrens 2231ab7f2deSmaybee return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE)); 224fa9e4066Sahrens } 225fa9e4066Sahrens 226ad23a2dbSjohansen /* 227ad23a2dbSjohansen * Use zio_data_buf_alloc to allocate data. The data will not appear in a 228ad23a2dbSjohansen * crashdump if the kernel panics. This exists so that we will limit the amount 229ad23a2dbSjohansen * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount 230ad23a2dbSjohansen * of kernel heap dumped to disk when the kernel panics) 231ad23a2dbSjohansen */ 232ad23a2dbSjohansen void * 233ad23a2dbSjohansen zio_data_buf_alloc(size_t size) 234ad23a2dbSjohansen { 235ad23a2dbSjohansen size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; 236ad23a2dbSjohansen 237ad23a2dbSjohansen ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); 238ad23a2dbSjohansen 2391ab7f2deSmaybee return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE)); 240ad23a2dbSjohansen } 241ad23a2dbSjohansen 242fa9e4066Sahrens void 243fa9e4066Sahrens zio_buf_free(void *buf, size_t size) 244fa9e4066Sahrens { 245fa9e4066Sahrens size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; 246fa9e4066Sahrens 247fa9e4066Sahrens ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); 248fa9e4066Sahrens 249fa9e4066Sahrens kmem_cache_free(zio_buf_cache[c], buf); 250fa9e4066Sahrens } 251fa9e4066Sahrens 252ad23a2dbSjohansen void 253ad23a2dbSjohansen zio_data_buf_free(void *buf, size_t size) 254ad23a2dbSjohansen { 255ad23a2dbSjohansen size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; 256ad23a2dbSjohansen 257ad23a2dbSjohansen ASSERT(c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); 258ad23a2dbSjohansen 259ad23a2dbSjohansen kmem_cache_free(zio_data_buf_cache[c], buf); 260ad23a2dbSjohansen } 261b3995adbSahrens 262fa9e4066Sahrens /* 263fa9e4066Sahrens * ========================================================================== 264fa9e4066Sahrens * Push and pop I/O transform buffers 265fa9e4066Sahrens * ========================================================================== 266fa9e4066Sahrens */ 267fa9e4066Sahrens static void 268e14bb325SJeff Bonwick zio_push_transform(zio_t *zio, void *data, uint64_t size, uint64_t bufsize, 269e14bb325SJeff Bonwick zio_transform_func_t *transform) 270fa9e4066Sahrens { 271fa9e4066Sahrens zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP); 272fa9e4066Sahrens 273e14bb325SJeff Bonwick zt->zt_orig_data = zio->io_data; 274e14bb325SJeff Bonwick zt->zt_orig_size = zio->io_size; 275fa9e4066Sahrens zt->zt_bufsize = bufsize; 276e14bb325SJeff Bonwick zt->zt_transform = transform; 277fa9e4066Sahrens 278fa9e4066Sahrens zt->zt_next = zio->io_transform_stack; 279fa9e4066Sahrens zio->io_transform_stack = zt; 280fa9e4066Sahrens 281fa9e4066Sahrens zio->io_data = data; 282fa9e4066Sahrens zio->io_size = size; 283fa9e4066Sahrens } 284fa9e4066Sahrens 285fa9e4066Sahrens static void 286e14bb325SJeff Bonwick zio_pop_transforms(zio_t *zio) 287fa9e4066Sahrens { 288e14bb325SJeff Bonwick zio_transform_t *zt; 289e14bb325SJeff Bonwick 290e14bb325SJeff Bonwick while ((zt = zio->io_transform_stack) != NULL) { 291e14bb325SJeff Bonwick if (zt->zt_transform != NULL) 292e14bb325SJeff Bonwick zt->zt_transform(zio, 293e14bb325SJeff Bonwick zt->zt_orig_data, zt->zt_orig_size); 294fa9e4066Sahrens 295b24ab676SJeff Bonwick if (zt->zt_bufsize != 0) 296b24ab676SJeff Bonwick zio_buf_free(zio->io_data, zt->zt_bufsize); 297fa9e4066Sahrens 298e14bb325SJeff Bonwick zio->io_data = zt->zt_orig_data; 299e14bb325SJeff Bonwick zio->io_size = zt->zt_orig_size; 300e14bb325SJeff Bonwick zio->io_transform_stack = zt->zt_next; 301fa9e4066Sahrens 302e14bb325SJeff Bonwick kmem_free(zt, sizeof (zio_transform_t)); 303fa9e4066Sahrens } 304fa9e4066Sahrens } 305fa9e4066Sahrens 306e14bb325SJeff Bonwick /* 307e14bb325SJeff Bonwick * ========================================================================== 308e14bb325SJeff Bonwick * I/O transform callbacks for subblocks and decompression 309e14bb325SJeff Bonwick * ========================================================================== 310e14bb325SJeff Bonwick */ 311e14bb325SJeff Bonwick static void 312e14bb325SJeff Bonwick zio_subblock(zio_t *zio, void *data, uint64_t size) 313e14bb325SJeff Bonwick { 314e14bb325SJeff Bonwick ASSERT(zio->io_size > size); 315e14bb325SJeff Bonwick 316e14bb325SJeff Bonwick if (zio->io_type == ZIO_TYPE_READ) 317e14bb325SJeff Bonwick bcopy(zio->io_data, data, size); 318e14bb325SJeff Bonwick } 319e14bb325SJeff Bonwick 320e14bb325SJeff Bonwick static void 321e14bb325SJeff Bonwick zio_decompress(zio_t *zio, void *data, uint64_t size) 322e14bb325SJeff Bonwick { 323e14bb325SJeff Bonwick if (zio->io_error == 0 && 324e14bb325SJeff Bonwick zio_decompress_data(BP_GET_COMPRESS(zio->io_bp), 325b24ab676SJeff Bonwick zio->io_data, data, zio->io_size, size) != 0) 326be6fd75aSMatthew Ahrens zio->io_error = SET_ERROR(EIO); 327e14bb325SJeff Bonwick } 328e14bb325SJeff Bonwick 329e14bb325SJeff Bonwick /* 330e14bb325SJeff Bonwick * ========================================================================== 331e14bb325SJeff Bonwick * I/O parent/child relationships and pipeline interlocks 332e14bb325SJeff Bonwick * ========================================================================== 333e14bb325SJeff Bonwick */ 334a3f829aeSBill Moore /* 335a3f829aeSBill Moore * NOTE - Callers to zio_walk_parents() and zio_walk_children must 336a3f829aeSBill Moore * continue calling these functions until they return NULL. 337a3f829aeSBill Moore * Otherwise, the next caller will pick up the list walk in 338a3f829aeSBill Moore * some indeterminate state. (Otherwise every caller would 339a3f829aeSBill Moore * have to pass in a cookie to keep the state represented by 340a3f829aeSBill Moore * io_walk_link, which gets annoying.) 341a3f829aeSBill Moore */ 342a3f829aeSBill Moore zio_t * 343a3f829aeSBill Moore zio_walk_parents(zio_t *cio) 344a3f829aeSBill Moore { 345a3f829aeSBill Moore zio_link_t *zl = cio->io_walk_link; 346a3f829aeSBill Moore list_t *pl = &cio->io_parent_list; 347e14bb325SJeff Bonwick 348a3f829aeSBill Moore zl = (zl == NULL) ? list_head(pl) : list_next(pl, zl); 349a3f829aeSBill Moore cio->io_walk_link = zl; 350a3f829aeSBill Moore 351a3f829aeSBill Moore if (zl == NULL) 352a3f829aeSBill Moore return (NULL); 353a3f829aeSBill Moore 354a3f829aeSBill Moore ASSERT(zl->zl_child == cio); 355a3f829aeSBill Moore return (zl->zl_parent); 356a3f829aeSBill Moore } 357a3f829aeSBill Moore 358a3f829aeSBill Moore zio_t * 359a3f829aeSBill Moore zio_walk_children(zio_t *pio) 360a3f829aeSBill Moore { 361a3f829aeSBill Moore zio_link_t *zl = pio->io_walk_link; 362a3f829aeSBill Moore list_t *cl = &pio->io_child_list; 363a3f829aeSBill Moore 364a3f829aeSBill Moore zl = (zl == NULL) ? list_head(cl) : list_next(cl, zl); 365a3f829aeSBill Moore pio->io_walk_link = zl; 366a3f829aeSBill Moore 367a3f829aeSBill Moore if (zl == NULL) 368a3f829aeSBill Moore return (NULL); 369a3f829aeSBill Moore 370a3f829aeSBill Moore ASSERT(zl->zl_parent == pio); 371a3f829aeSBill Moore return (zl->zl_child); 372a3f829aeSBill Moore } 373a3f829aeSBill Moore 374a3f829aeSBill Moore zio_t * 375a3f829aeSBill Moore zio_unique_parent(zio_t *cio) 376a3f829aeSBill Moore { 377a3f829aeSBill Moore zio_t *pio = zio_walk_parents(cio); 378a3f829aeSBill Moore 379a3f829aeSBill Moore VERIFY(zio_walk_parents(cio) == NULL); 380a3f829aeSBill Moore return (pio); 381a3f829aeSBill Moore } 382a3f829aeSBill Moore 383a3f829aeSBill Moore void 384a3f829aeSBill Moore zio_add_child(zio_t *pio, zio_t *cio) 385e14bb325SJeff Bonwick { 386a3f829aeSBill Moore zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP); 387a3f829aeSBill Moore 388a3f829aeSBill Moore /* 389a3f829aeSBill Moore * Logical I/Os can have logical, gang, or vdev children. 390a3f829aeSBill Moore * Gang I/Os can have gang or vdev children. 391a3f829aeSBill Moore * Vdev I/Os can only have vdev children. 392a3f829aeSBill Moore * The following ASSERT captures all of these constraints. 393a3f829aeSBill Moore */ 394a3f829aeSBill Moore ASSERT(cio->io_child_type <= pio->io_child_type); 395a3f829aeSBill Moore 396a3f829aeSBill Moore zl->zl_parent = pio; 397a3f829aeSBill Moore zl->zl_child = cio; 398a3f829aeSBill Moore 399a3f829aeSBill Moore mutex_enter(&cio->io_lock); 400e14bb325SJeff Bonwick mutex_enter(&pio->io_lock); 401a3f829aeSBill Moore 402a3f829aeSBill Moore ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0); 403a3f829aeSBill Moore 404a3f829aeSBill Moore for (int w = 0; w < ZIO_WAIT_TYPES; w++) 405a3f829aeSBill Moore pio->io_children[cio->io_child_type][w] += !cio->io_state[w]; 406a3f829aeSBill Moore 407a3f829aeSBill Moore list_insert_head(&pio->io_child_list, zl); 408a3f829aeSBill Moore list_insert_head(&cio->io_parent_list, zl); 409a3f829aeSBill Moore 410b24ab676SJeff Bonwick pio->io_child_count++; 411b24ab676SJeff Bonwick cio->io_parent_count++; 412b24ab676SJeff Bonwick 413e14bb325SJeff Bonwick mutex_exit(&pio->io_lock); 414a3f829aeSBill Moore mutex_exit(&cio->io_lock); 415e14bb325SJeff Bonwick } 416e14bb325SJeff Bonwick 417fa9e4066Sahrens static void 418a3f829aeSBill Moore zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl) 419e14bb325SJeff Bonwick { 420a3f829aeSBill Moore ASSERT(zl->zl_parent == pio); 421a3f829aeSBill Moore ASSERT(zl->zl_child == cio); 422e14bb325SJeff Bonwick 423a3f829aeSBill Moore mutex_enter(&cio->io_lock); 424e14bb325SJeff Bonwick mutex_enter(&pio->io_lock); 425a3f829aeSBill Moore 426a3f829aeSBill Moore list_remove(&pio->io_child_list, zl); 427a3f829aeSBill Moore list_remove(&cio->io_parent_list, zl); 428a3f829aeSBill Moore 429b24ab676SJeff Bonwick pio->io_child_count--; 430b24ab676SJeff Bonwick cio->io_parent_count--; 431b24ab676SJeff Bonwick 432e14bb325SJeff Bonwick mutex_exit(&pio->io_lock); 433a3f829aeSBill Moore mutex_exit(&cio->io_lock); 434a3f829aeSBill Moore 435a3f829aeSBill Moore kmem_cache_free(zio_link_cache, zl); 436e14bb325SJeff Bonwick } 437e14bb325SJeff Bonwick 438e14bb325SJeff Bonwick static boolean_t 439e14bb325SJeff Bonwick zio_wait_for_children(zio_t *zio, enum zio_child child, enum zio_wait_type wait) 440fa9e4066Sahrens { 441e14bb325SJeff Bonwick uint64_t *countp = &zio->io_children[child][wait]; 442e14bb325SJeff Bonwick boolean_t waiting = B_FALSE; 443e14bb325SJeff Bonwick 444e14bb325SJeff Bonwick mutex_enter(&zio->io_lock); 445e14bb325SJeff Bonwick ASSERT(zio->io_stall == NULL); 446e14bb325SJeff Bonwick if (*countp != 0) { 447b24ab676SJeff Bonwick zio->io_stage >>= 1; 448e14bb325SJeff Bonwick zio->io_stall = countp; 449e14bb325SJeff Bonwick waiting = B_TRUE; 450e14bb325SJeff Bonwick } 451e14bb325SJeff Bonwick mutex_exit(&zio->io_lock); 452e14bb325SJeff Bonwick 453e14bb325SJeff Bonwick return (waiting); 454e14bb325SJeff Bonwick } 455fa9e4066Sahrens 456e14bb325SJeff Bonwick static void 457e14bb325SJeff Bonwick zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait) 458e14bb325SJeff Bonwick { 459e14bb325SJeff Bonwick uint64_t *countp = &pio->io_children[zio->io_child_type][wait]; 460e14bb325SJeff Bonwick int *errorp = &pio->io_child_error[zio->io_child_type]; 461fa9e4066Sahrens 462e14bb325SJeff Bonwick mutex_enter(&pio->io_lock); 463e14bb325SJeff Bonwick if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE)) 464e14bb325SJeff Bonwick *errorp = zio_worst_error(*errorp, zio->io_error); 465e14bb325SJeff Bonwick pio->io_reexecute |= zio->io_reexecute; 466e14bb325SJeff Bonwick ASSERT3U(*countp, >, 0); 46769962b56SMatthew Ahrens 46869962b56SMatthew Ahrens (*countp)--; 46969962b56SMatthew Ahrens 47069962b56SMatthew Ahrens if (*countp == 0 && pio->io_stall == countp) { 471e14bb325SJeff Bonwick pio->io_stall = NULL; 472e14bb325SJeff Bonwick mutex_exit(&pio->io_lock); 473e14bb325SJeff Bonwick zio_execute(pio); 474e14bb325SJeff Bonwick } else { 475e14bb325SJeff Bonwick mutex_exit(&pio->io_lock); 476fa9e4066Sahrens } 477fa9e4066Sahrens } 478fa9e4066Sahrens 479e14bb325SJeff Bonwick static void 480e14bb325SJeff Bonwick zio_inherit_child_errors(zio_t *zio, enum zio_child c) 481e14bb325SJeff Bonwick { 482e14bb325SJeff Bonwick if (zio->io_child_error[c] != 0 && zio->io_error == 0) 483e14bb325SJeff Bonwick zio->io_error = zio->io_child_error[c]; 484e14bb325SJeff Bonwick } 485e14bb325SJeff Bonwick 486fa9e4066Sahrens /* 487fa9e4066Sahrens * ========================================================================== 488e14bb325SJeff Bonwick * Create the various types of I/O (read, write, free, etc) 489fa9e4066Sahrens * ========================================================================== 490fa9e4066Sahrens */ 491fa9e4066Sahrens static zio_t * 492b24ab676SJeff Bonwick zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp, 493fa9e4066Sahrens void *data, uint64_t size, zio_done_func_t *done, void *private, 49469962b56SMatthew Ahrens zio_type_t type, zio_priority_t priority, enum zio_flag flags, 4957802d7bfSMatthew Ahrens vdev_t *vd, uint64_t offset, const zbookmark_phys_t *zb, 496b24ab676SJeff Bonwick enum zio_stage stage, enum zio_stage pipeline) 497fa9e4066Sahrens { 498fa9e4066Sahrens zio_t *zio; 499fa9e4066Sahrens 500fa9e4066Sahrens ASSERT3U(size, <=, SPA_MAXBLOCKSIZE); 501fa9e4066Sahrens ASSERT(P2PHASE(size, SPA_MINBLOCKSIZE) == 0); 502e14bb325SJeff Bonwick ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0); 503fa9e4066Sahrens 504e14bb325SJeff Bonwick ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER)); 505e14bb325SJeff Bonwick ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER)); 506e14bb325SJeff Bonwick ASSERT(vd || stage == ZIO_STAGE_OPEN); 507088f3894Sahrens 508ccae0b50Seschrock zio = kmem_cache_alloc(zio_cache, KM_SLEEP); 509ccae0b50Seschrock bzero(zio, sizeof (zio_t)); 510e14bb325SJeff Bonwick 511e14bb325SJeff Bonwick mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL); 512e14bb325SJeff Bonwick cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL); 513e14bb325SJeff Bonwick 514a3f829aeSBill Moore list_create(&zio->io_parent_list, sizeof (zio_link_t), 515a3f829aeSBill Moore offsetof(zio_link_t, zl_parent_node)); 516a3f829aeSBill Moore list_create(&zio->io_child_list, sizeof (zio_link_t), 517a3f829aeSBill Moore offsetof(zio_link_t, zl_child_node)); 518a3f829aeSBill Moore 519e14bb325SJeff Bonwick if (vd != NULL) 520e14bb325SJeff Bonwick zio->io_child_type = ZIO_CHILD_VDEV; 521e14bb325SJeff Bonwick else if (flags & ZIO_FLAG_GANG_CHILD) 522e14bb325SJeff Bonwick zio->io_child_type = ZIO_CHILD_GANG; 523b24ab676SJeff Bonwick else if (flags & ZIO_FLAG_DDT_CHILD) 524b24ab676SJeff Bonwick zio->io_child_type = ZIO_CHILD_DDT; 525e14bb325SJeff Bonwick else 526e14bb325SJeff Bonwick zio->io_child_type = ZIO_CHILD_LOGICAL; 527e14bb325SJeff Bonwick 528fa9e4066Sahrens if (bp != NULL) { 529b24ab676SJeff Bonwick zio->io_bp = (blkptr_t *)bp; 530fa9e4066Sahrens zio->io_bp_copy = *bp; 531fa9e4066Sahrens zio->io_bp_orig = *bp; 532b24ab676SJeff Bonwick if (type != ZIO_TYPE_WRITE || 533b24ab676SJeff Bonwick zio->io_child_type == ZIO_CHILD_DDT) 534e14bb325SJeff Bonwick zio->io_bp = &zio->io_bp_copy; /* so caller can free */ 535f5383399SBill Moore if (zio->io_child_type == ZIO_CHILD_LOGICAL) 536e14bb325SJeff Bonwick zio->io_logical = zio; 537f5383399SBill Moore if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp)) 538f5383399SBill Moore pipeline |= ZIO_GANG_STAGES; 539fa9e4066Sahrens } 540e14bb325SJeff Bonwick 541e14bb325SJeff Bonwick zio->io_spa = spa; 542e14bb325SJeff Bonwick zio->io_txg = txg; 543fa9e4066Sahrens zio->io_done = done; 544fa9e4066Sahrens zio->io_private = private; 545fa9e4066Sahrens zio->io_type = type; 546fa9e4066Sahrens zio->io_priority = priority; 547e14bb325SJeff Bonwick zio->io_vd = vd; 548e14bb325SJeff Bonwick zio->io_offset = offset; 549b24ab676SJeff Bonwick zio->io_orig_data = zio->io_data = data; 550b24ab676SJeff Bonwick zio->io_orig_size = zio->io_size = size; 551e14bb325SJeff Bonwick zio->io_orig_flags = zio->io_flags = flags; 552e14bb325SJeff Bonwick zio->io_orig_stage = zio->io_stage = stage; 553e14bb325SJeff Bonwick zio->io_orig_pipeline = zio->io_pipeline = pipeline; 554fa9e4066Sahrens 555a3f829aeSBill Moore zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY); 556a3f829aeSBill Moore zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE); 557a3f829aeSBill Moore 558e14bb325SJeff Bonwick if (zb != NULL) 559e14bb325SJeff Bonwick zio->io_bookmark = *zb; 560e14bb325SJeff Bonwick 561e14bb325SJeff Bonwick if (pio != NULL) { 562e14bb325SJeff Bonwick if (zio->io_logical == NULL) 563ea8dc4b6Seschrock zio->io_logical = pio->io_logical; 564f5383399SBill Moore if (zio->io_child_type == ZIO_CHILD_GANG) 565f5383399SBill Moore zio->io_gang_leader = pio->io_gang_leader; 566e14bb325SJeff Bonwick zio_add_child(pio, zio); 567fa9e4066Sahrens } 568fa9e4066Sahrens 569fa9e4066Sahrens return (zio); 570fa9e4066Sahrens } 571fa9e4066Sahrens 5720a4e9518Sgw static void 573e14bb325SJeff Bonwick zio_destroy(zio_t *zio) 5740a4e9518Sgw { 575a3f829aeSBill Moore list_destroy(&zio->io_parent_list); 576a3f829aeSBill Moore list_destroy(&zio->io_child_list); 577e14bb325SJeff Bonwick mutex_destroy(&zio->io_lock); 578e14bb325SJeff Bonwick cv_destroy(&zio->io_cv); 579e14bb325SJeff Bonwick kmem_cache_free(zio_cache, zio); 5800a4e9518Sgw } 5810a4e9518Sgw 582fa9e4066Sahrens zio_t * 583a3f829aeSBill Moore zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done, 584b24ab676SJeff Bonwick void *private, enum zio_flag flags) 585fa9e4066Sahrens { 586fa9e4066Sahrens zio_t *zio; 587fa9e4066Sahrens 588fa9e4066Sahrens zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private, 589a3f829aeSBill Moore ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL, 590e14bb325SJeff Bonwick ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE); 591fa9e4066Sahrens 592fa9e4066Sahrens return (zio); 593fa9e4066Sahrens } 594fa9e4066Sahrens 595fa9e4066Sahrens zio_t * 596b24ab676SJeff Bonwick zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags) 597fa9e4066Sahrens { 598a3f829aeSBill Moore return (zio_null(NULL, spa, NULL, done, private, flags)); 599fa9e4066Sahrens } 600fa9e4066Sahrens 601fa9e4066Sahrens zio_t * 602e14bb325SJeff Bonwick zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, 603e14bb325SJeff Bonwick void *data, uint64_t size, zio_done_func_t *done, void *private, 6047802d7bfSMatthew Ahrens zio_priority_t priority, enum zio_flag flags, const zbookmark_phys_t *zb) 605fa9e4066Sahrens { 606fa9e4066Sahrens zio_t *zio; 607fa9e4066Sahrens 608b24ab676SJeff Bonwick zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp, 609088f3894Sahrens data, size, done, private, 610e14bb325SJeff Bonwick ZIO_TYPE_READ, priority, flags, NULL, 0, zb, 611b24ab676SJeff Bonwick ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ? 612b24ab676SJeff Bonwick ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE); 613fa9e4066Sahrens 614fa9e4066Sahrens return (zio); 615fa9e4066Sahrens } 616fa9e4066Sahrens 617fa9e4066Sahrens zio_t * 618e14bb325SJeff Bonwick zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, 619b24ab676SJeff Bonwick void *data, uint64_t size, const zio_prop_t *zp, 62069962b56SMatthew Ahrens zio_done_func_t *ready, zio_done_func_t *physdone, zio_done_func_t *done, 62169962b56SMatthew Ahrens void *private, 6227802d7bfSMatthew Ahrens zio_priority_t priority, enum zio_flag flags, const zbookmark_phys_t *zb) 623fa9e4066Sahrens { 624fa9e4066Sahrens zio_t *zio; 625fa9e4066Sahrens 626e14bb325SJeff Bonwick ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF && 627e14bb325SJeff Bonwick zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS && 628e14bb325SJeff Bonwick zp->zp_compress >= ZIO_COMPRESS_OFF && 629e14bb325SJeff Bonwick zp->zp_compress < ZIO_COMPRESS_FUNCTIONS && 630ad135b5dSChristopher Siden DMU_OT_IS_VALID(zp->zp_type) && 631e14bb325SJeff Bonwick zp->zp_level < 32 && 632b24ab676SJeff Bonwick zp->zp_copies > 0 && 63380901aeaSGeorge Wilson zp->zp_copies <= spa_max_replication(spa)); 6340a4e9518Sgw 635fa9e4066Sahrens zio = zio_create(pio, spa, txg, bp, data, size, done, private, 636e14bb325SJeff Bonwick ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb, 637b24ab676SJeff Bonwick ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ? 638b24ab676SJeff Bonwick ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE); 639fa9e4066Sahrens 640c717a561Smaybee zio->io_ready = ready; 64169962b56SMatthew Ahrens zio->io_physdone = physdone; 642e14bb325SJeff Bonwick zio->io_prop = *zp; 643fa9e4066Sahrens 6445d7b4d43SMatthew Ahrens /* 6455d7b4d43SMatthew Ahrens * Data can be NULL if we are going to call zio_write_override() to 6465d7b4d43SMatthew Ahrens * provide the already-allocated BP. But we may need the data to 6475d7b4d43SMatthew Ahrens * verify a dedup hit (if requested). In this case, don't try to 6485d7b4d43SMatthew Ahrens * dedup (just take the already-allocated BP verbatim). 6495d7b4d43SMatthew Ahrens */ 6505d7b4d43SMatthew Ahrens if (data == NULL && zio->io_prop.zp_dedup_verify) { 6515d7b4d43SMatthew Ahrens zio->io_prop.zp_dedup = zio->io_prop.zp_dedup_verify = B_FALSE; 6525d7b4d43SMatthew Ahrens } 6535d7b4d43SMatthew Ahrens 654fa9e4066Sahrens return (zio); 655fa9e4066Sahrens } 656fa9e4066Sahrens 657fa9e4066Sahrens zio_t * 658e14bb325SJeff Bonwick zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, void *data, 65969962b56SMatthew Ahrens uint64_t size, zio_done_func_t *done, void *private, 6607802d7bfSMatthew Ahrens zio_priority_t priority, enum zio_flag flags, zbookmark_phys_t *zb) 661fa9e4066Sahrens { 662fa9e4066Sahrens zio_t *zio; 663fa9e4066Sahrens 664fa9e4066Sahrens zio = zio_create(pio, spa, txg, bp, data, size, done, private, 665e14bb325SJeff Bonwick ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb, 666e14bb325SJeff Bonwick ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE); 667fa9e4066Sahrens 668fa9e4066Sahrens return (zio); 669fa9e4066Sahrens } 670fa9e4066Sahrens 671b24ab676SJeff Bonwick void 67280901aeaSGeorge Wilson zio_write_override(zio_t *zio, blkptr_t *bp, int copies, boolean_t nopwrite) 673b24ab676SJeff Bonwick { 674b24ab676SJeff Bonwick ASSERT(zio->io_type == ZIO_TYPE_WRITE); 675b24ab676SJeff Bonwick ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 676b24ab676SJeff Bonwick ASSERT(zio->io_stage == ZIO_STAGE_OPEN); 677b24ab676SJeff Bonwick ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa)); 678b24ab676SJeff Bonwick 67980901aeaSGeorge Wilson /* 68080901aeaSGeorge Wilson * We must reset the io_prop to match the values that existed 68180901aeaSGeorge Wilson * when the bp was first written by dmu_sync() keeping in mind 68280901aeaSGeorge Wilson * that nopwrite and dedup are mutually exclusive. 68380901aeaSGeorge Wilson */ 68480901aeaSGeorge Wilson zio->io_prop.zp_dedup = nopwrite ? B_FALSE : zio->io_prop.zp_dedup; 68580901aeaSGeorge Wilson zio->io_prop.zp_nopwrite = nopwrite; 686b24ab676SJeff Bonwick zio->io_prop.zp_copies = copies; 687b24ab676SJeff Bonwick zio->io_bp_override = bp; 688b24ab676SJeff Bonwick } 689b24ab676SJeff Bonwick 690b24ab676SJeff Bonwick void 691b24ab676SJeff Bonwick zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp) 692b24ab676SJeff Bonwick { 6935d7b4d43SMatthew Ahrens 6945d7b4d43SMatthew Ahrens /* 6955d7b4d43SMatthew Ahrens * The check for EMBEDDED is a performance optimization. We 6965d7b4d43SMatthew Ahrens * process the free here (by ignoring it) rather than 6975d7b4d43SMatthew Ahrens * putting it on the list and then processing it in zio_free_sync(). 6985d7b4d43SMatthew Ahrens */ 6995d7b4d43SMatthew Ahrens if (BP_IS_EMBEDDED(bp)) 7005d7b4d43SMatthew Ahrens return; 7013b2aab18SMatthew Ahrens metaslab_check_free(spa, bp); 7029cb154a3SMatthew Ahrens 7039cb154a3SMatthew Ahrens /* 7049cb154a3SMatthew Ahrens * Frees that are for the currently-syncing txg, are not going to be 7059cb154a3SMatthew Ahrens * deferred, and which will not need to do a read (i.e. not GANG or 7069cb154a3SMatthew Ahrens * DEDUP), can be processed immediately. Otherwise, put them on the 7079cb154a3SMatthew Ahrens * in-memory list for later processing. 7089cb154a3SMatthew Ahrens */ 7099cb154a3SMatthew Ahrens if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp) || 7109cb154a3SMatthew Ahrens txg != spa->spa_syncing_txg || 7119cb154a3SMatthew Ahrens spa_sync_pass(spa) >= zfs_sync_pass_deferred_free) { 7129cb154a3SMatthew Ahrens bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp); 7139cb154a3SMatthew Ahrens } else { 7149cb154a3SMatthew Ahrens VERIFY0(zio_wait(zio_free_sync(NULL, spa, txg, bp, 0))); 7159cb154a3SMatthew Ahrens } 716b24ab676SJeff Bonwick } 717b24ab676SJeff Bonwick 718fa9e4066Sahrens zio_t * 719b24ab676SJeff Bonwick zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp, 720b24ab676SJeff Bonwick enum zio_flag flags) 721fa9e4066Sahrens { 722fa9e4066Sahrens zio_t *zio; 7239cb154a3SMatthew Ahrens enum zio_stage stage = ZIO_FREE_PIPELINE; 724fa9e4066Sahrens 725fa9e4066Sahrens ASSERT(!BP_IS_HOLE(bp)); 726b24ab676SJeff Bonwick ASSERT(spa_syncing_txg(spa) == txg); 72701f55e48SGeorge Wilson ASSERT(spa_sync_pass(spa) < zfs_sync_pass_deferred_free); 728fa9e4066Sahrens 7295d7b4d43SMatthew Ahrens if (BP_IS_EMBEDDED(bp)) 7305d7b4d43SMatthew Ahrens return (zio_null(pio, spa, NULL, NULL, NULL, 0)); 7315d7b4d43SMatthew Ahrens 7323b2aab18SMatthew Ahrens metaslab_check_free(spa, bp); 7336e6d5868SMatthew Ahrens arc_freed(spa, bp); 7343b2aab18SMatthew Ahrens 7359cb154a3SMatthew Ahrens /* 7369cb154a3SMatthew Ahrens * GANG and DEDUP blocks can induce a read (for the gang block header, 7379cb154a3SMatthew Ahrens * or the DDT), so issue them asynchronously so that this thread is 7389cb154a3SMatthew Ahrens * not tied up. 7399cb154a3SMatthew Ahrens */ 7409cb154a3SMatthew Ahrens if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp)) 7419cb154a3SMatthew Ahrens stage |= ZIO_STAGE_ISSUE_ASYNC; 7429cb154a3SMatthew Ahrens 743e14bb325SJeff Bonwick zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp), 74469962b56SMatthew Ahrens NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_NOW, flags, 7459cb154a3SMatthew Ahrens NULL, 0, NULL, ZIO_STAGE_OPEN, stage); 7469cb154a3SMatthew Ahrens 747fa9e4066Sahrens return (zio); 748fa9e4066Sahrens } 749fa9e4066Sahrens 750fa9e4066Sahrens zio_t * 751b24ab676SJeff Bonwick zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp, 752b24ab676SJeff Bonwick zio_done_func_t *done, void *private, enum zio_flag flags) 753fa9e4066Sahrens { 754fa9e4066Sahrens zio_t *zio; 755fa9e4066Sahrens 7565d7b4d43SMatthew Ahrens dprintf_bp(bp, "claiming in txg %llu", txg); 7575d7b4d43SMatthew Ahrens 7585d7b4d43SMatthew Ahrens if (BP_IS_EMBEDDED(bp)) 7595d7b4d43SMatthew Ahrens return (zio_null(pio, spa, NULL, NULL, NULL, 0)); 7605d7b4d43SMatthew Ahrens 761fa9e4066Sahrens /* 762fa9e4066Sahrens * A claim is an allocation of a specific block. Claims are needed 763fa9e4066Sahrens * to support immediate writes in the intent log. The issue is that 764fa9e4066Sahrens * immediate writes contain committed data, but in a txg that was 765fa9e4066Sahrens * *not* committed. Upon opening the pool after an unclean shutdown, 766fa9e4066Sahrens * the intent log claims all blocks that contain immediate write data 767fa9e4066Sahrens * so that the SPA knows they're in use. 768fa9e4066Sahrens * 769fa9e4066Sahrens * All claims *must* be resolved in the first txg -- before the SPA 770fa9e4066Sahrens * starts allocating blocks -- so that nothing is allocated twice. 771b24ab676SJeff Bonwick * If txg == 0 we just verify that the block is claimable. 772fa9e4066Sahrens */ 773fa9e4066Sahrens ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa)); 774b24ab676SJeff Bonwick ASSERT(txg == spa_first_txg(spa) || txg == 0); 775b24ab676SJeff Bonwick ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(1M) */ 776fa9e4066Sahrens 777e14bb325SJeff Bonwick zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp), 778e14bb325SJeff Bonwick done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, flags, 779e14bb325SJeff Bonwick NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE); 780fa9e4066Sahrens 781fa9e4066Sahrens return (zio); 782fa9e4066Sahrens } 783fa9e4066Sahrens 784fa9e4066Sahrens zio_t * 785fa9e4066Sahrens zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd, 78669962b56SMatthew Ahrens zio_done_func_t *done, void *private, enum zio_flag flags) 787fa9e4066Sahrens { 788fa9e4066Sahrens zio_t *zio; 789fa9e4066Sahrens int c; 790fa9e4066Sahrens 791fa9e4066Sahrens if (vd->vdev_children == 0) { 792fa9e4066Sahrens zio = zio_create(pio, spa, 0, NULL, NULL, 0, done, private, 79369962b56SMatthew Ahrens ZIO_TYPE_IOCTL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL, 794fa9e4066Sahrens ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE); 795fa9e4066Sahrens 796fa9e4066Sahrens zio->io_cmd = cmd; 797fa9e4066Sahrens } else { 798a3f829aeSBill Moore zio = zio_null(pio, spa, NULL, NULL, NULL, flags); 799fa9e4066Sahrens 800fa9e4066Sahrens for (c = 0; c < vd->vdev_children; c++) 801fa9e4066Sahrens zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd, 80269962b56SMatthew Ahrens done, private, flags)); 803fa9e4066Sahrens } 804fa9e4066Sahrens 805fa9e4066Sahrens return (zio); 806fa9e4066Sahrens } 807fa9e4066Sahrens 808fa9e4066Sahrens zio_t * 809fa9e4066Sahrens zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size, 810fa9e4066Sahrens void *data, int checksum, zio_done_func_t *done, void *private, 81169962b56SMatthew Ahrens zio_priority_t priority, enum zio_flag flags, boolean_t labels) 812fa9e4066Sahrens { 813fa9e4066Sahrens zio_t *zio; 8140a4e9518Sgw 815e14bb325SJeff Bonwick ASSERT(vd->vdev_children == 0); 816e14bb325SJeff Bonwick ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE || 817e14bb325SJeff Bonwick offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE); 818e14bb325SJeff Bonwick ASSERT3U(offset + size, <=, vd->vdev_psize); 819fa9e4066Sahrens 820e14bb325SJeff Bonwick zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private, 8212a104a52SAlex Reece ZIO_TYPE_READ, priority, flags | ZIO_FLAG_PHYSICAL, vd, offset, 8222a104a52SAlex Reece NULL, ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE); 823fa9e4066Sahrens 824e14bb325SJeff Bonwick zio->io_prop.zp_checksum = checksum; 825fa9e4066Sahrens 826fa9e4066Sahrens return (zio); 827fa9e4066Sahrens } 828fa9e4066Sahrens 829fa9e4066Sahrens zio_t * 830fa9e4066Sahrens zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size, 831fa9e4066Sahrens void *data, int checksum, zio_done_func_t *done, void *private, 83269962b56SMatthew Ahrens zio_priority_t priority, enum zio_flag flags, boolean_t labels) 833fa9e4066Sahrens { 834fa9e4066Sahrens zio_t *zio; 8350a4e9518Sgw 836e14bb325SJeff Bonwick ASSERT(vd->vdev_children == 0); 837e14bb325SJeff Bonwick ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE || 838e14bb325SJeff Bonwick offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE); 839e14bb325SJeff Bonwick ASSERT3U(offset + size, <=, vd->vdev_psize); 840fa9e4066Sahrens 841e14bb325SJeff Bonwick zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, done, private, 8422a104a52SAlex Reece ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_PHYSICAL, vd, offset, 8432a104a52SAlex Reece NULL, ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE); 844fa9e4066Sahrens 845e14bb325SJeff Bonwick zio->io_prop.zp_checksum = checksum; 846fa9e4066Sahrens 8476e1f5caaSNeil Perrin if (zio_checksum_table[checksum].ci_eck) { 848fa9e4066Sahrens /* 8496e1f5caaSNeil Perrin * zec checksums are necessarily destructive -- they modify 850e14bb325SJeff Bonwick * the end of the write buffer to hold the verifier/checksum. 851fa9e4066Sahrens * Therefore, we must make a local copy in case the data is 852e14bb325SJeff Bonwick * being written to multiple places in parallel. 853fa9e4066Sahrens */ 854e14bb325SJeff Bonwick void *wbuf = zio_buf_alloc(size); 855fa9e4066Sahrens bcopy(data, wbuf, size); 856e14bb325SJeff Bonwick zio_push_transform(zio, wbuf, size, size, NULL); 857fa9e4066Sahrens } 858fa9e4066Sahrens 859fa9e4066Sahrens return (zio); 860fa9e4066Sahrens } 861fa9e4066Sahrens 862fa9e4066Sahrens /* 863e14bb325SJeff Bonwick * Create a child I/O to do some work for us. 864fa9e4066Sahrens */ 865fa9e4066Sahrens zio_t * 866e14bb325SJeff Bonwick zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset, 86769962b56SMatthew Ahrens void *data, uint64_t size, int type, zio_priority_t priority, 86869962b56SMatthew Ahrens enum zio_flag flags, zio_done_func_t *done, void *private) 869fa9e4066Sahrens { 870b24ab676SJeff Bonwick enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE; 871e14bb325SJeff Bonwick zio_t *zio; 872e14bb325SJeff Bonwick 873e14bb325SJeff Bonwick ASSERT(vd->vdev_parent == 874e14bb325SJeff Bonwick (pio->io_vd ? pio->io_vd : pio->io_spa->spa_root_vdev)); 875fa9e4066Sahrens 876fa9e4066Sahrens if (type == ZIO_TYPE_READ && bp != NULL) { 877fa9e4066Sahrens /* 878fa9e4066Sahrens * If we have the bp, then the child should perform the 879fa9e4066Sahrens * checksum and the parent need not. This pushes error 880fa9e4066Sahrens * detection as close to the leaves as possible and 881fa9e4066Sahrens * eliminates redundant checksums in the interior nodes. 882fa9e4066Sahrens */ 883b24ab676SJeff Bonwick pipeline |= ZIO_STAGE_CHECKSUM_VERIFY; 884b24ab676SJeff Bonwick pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY; 885fa9e4066Sahrens } 886fa9e4066Sahrens 887e14bb325SJeff Bonwick if (vd->vdev_children == 0) 888e14bb325SJeff Bonwick offset += VDEV_LABEL_START_SIZE; 889e14bb325SJeff Bonwick 890b24ab676SJeff Bonwick flags |= ZIO_VDEV_CHILD_FLAGS(pio) | ZIO_FLAG_DONT_PROPAGATE; 891b24ab676SJeff Bonwick 892b24ab676SJeff Bonwick /* 893b24ab676SJeff Bonwick * If we've decided to do a repair, the write is not speculative -- 894b24ab676SJeff Bonwick * even if the original read was. 895b24ab676SJeff Bonwick */ 896b24ab676SJeff Bonwick if (flags & ZIO_FLAG_IO_REPAIR) 897b24ab676SJeff Bonwick flags &= ~ZIO_FLAG_SPECULATIVE; 898b24ab676SJeff Bonwick 899e14bb325SJeff Bonwick zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size, 900b24ab676SJeff Bonwick done, private, type, priority, flags, vd, offset, &pio->io_bookmark, 901b24ab676SJeff Bonwick ZIO_STAGE_VDEV_IO_START >> 1, pipeline); 902fa9e4066Sahrens 90369962b56SMatthew Ahrens zio->io_physdone = pio->io_physdone; 90469962b56SMatthew Ahrens if (vd->vdev_ops->vdev_op_leaf && zio->io_logical != NULL) 90569962b56SMatthew Ahrens zio->io_logical->io_phys_children++; 90669962b56SMatthew Ahrens 907e14bb325SJeff Bonwick return (zio); 90832b87932Sek } 90932b87932Sek 910e14bb325SJeff Bonwick zio_t * 911e14bb325SJeff Bonwick zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, void *data, uint64_t size, 91269962b56SMatthew Ahrens int type, zio_priority_t priority, enum zio_flag flags, 913b24ab676SJeff Bonwick zio_done_func_t *done, void *private) 914fa9e4066Sahrens { 915e14bb325SJeff Bonwick zio_t *zio; 916fa9e4066Sahrens 917e14bb325SJeff Bonwick ASSERT(vd->vdev_ops->vdev_op_leaf); 918fa9e4066Sahrens 919e14bb325SJeff Bonwick zio = zio_create(NULL, vd->vdev_spa, 0, NULL, 920e14bb325SJeff Bonwick data, size, done, private, type, priority, 92169962b56SMatthew Ahrens flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY | ZIO_FLAG_DELEGATED, 922e14bb325SJeff Bonwick vd, offset, NULL, 923b24ab676SJeff Bonwick ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE); 924fa9e4066Sahrens 925e14bb325SJeff Bonwick return (zio); 926e05725b1Sbonwick } 927e05725b1Sbonwick 928e05725b1Sbonwick void 929e14bb325SJeff Bonwick zio_flush(zio_t *zio, vdev_t *vd) 930e05725b1Sbonwick { 931e14bb325SJeff Bonwick zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE, 93269962b56SMatthew Ahrens NULL, NULL, 933e14bb325SJeff Bonwick ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY)); 934fa9e4066Sahrens } 935fa9e4066Sahrens 9366e1f5caaSNeil Perrin void 9376e1f5caaSNeil Perrin zio_shrink(zio_t *zio, uint64_t size) 9386e1f5caaSNeil Perrin { 9396e1f5caaSNeil Perrin ASSERT(zio->io_executor == NULL); 9406e1f5caaSNeil Perrin ASSERT(zio->io_orig_size == zio->io_size); 9416e1f5caaSNeil Perrin ASSERT(size <= zio->io_size); 9426e1f5caaSNeil Perrin 9436e1f5caaSNeil Perrin /* 9446e1f5caaSNeil Perrin * We don't shrink for raidz because of problems with the 9456e1f5caaSNeil Perrin * reconstruction when reading back less than the block size. 9466e1f5caaSNeil Perrin * Note, BP_IS_RAIDZ() assumes no compression. 9476e1f5caaSNeil Perrin */ 9486e1f5caaSNeil Perrin ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF); 9496e1f5caaSNeil Perrin if (!BP_IS_RAIDZ(zio->io_bp)) 9506e1f5caaSNeil Perrin zio->io_orig_size = zio->io_size = size; 9516e1f5caaSNeil Perrin } 9526e1f5caaSNeil Perrin 953fa9e4066Sahrens /* 954fa9e4066Sahrens * ========================================================================== 955e14bb325SJeff Bonwick * Prepare to read and write logical blocks 956fa9e4066Sahrens * ========================================================================== 957fa9e4066Sahrens */ 958e14bb325SJeff Bonwick 959e05725b1Sbonwick static int 960e14bb325SJeff Bonwick zio_read_bp_init(zio_t *zio) 961fa9e4066Sahrens { 962e14bb325SJeff Bonwick blkptr_t *bp = zio->io_bp; 963e05725b1Sbonwick 96403361682SJeff Bonwick if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF && 965f5383399SBill Moore zio->io_child_type == ZIO_CHILD_LOGICAL && 966f5383399SBill Moore !(zio->io_flags & ZIO_FLAG_RAW)) { 9675d7b4d43SMatthew Ahrens uint64_t psize = 9685d7b4d43SMatthew Ahrens BP_IS_EMBEDDED(bp) ? BPE_GET_PSIZE(bp) : BP_GET_PSIZE(bp); 969b24ab676SJeff Bonwick void *cbuf = zio_buf_alloc(psize); 970e05725b1Sbonwick 971b24ab676SJeff Bonwick zio_push_transform(zio, cbuf, psize, psize, zio_decompress); 972e14bb325SJeff Bonwick } 973fa9e4066Sahrens 9745d7b4d43SMatthew Ahrens if (BP_IS_EMBEDDED(bp) && BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA) { 9755d7b4d43SMatthew Ahrens zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 9765d7b4d43SMatthew Ahrens decode_embedded_bp_compressed(bp, zio->io_data); 9775d7b4d43SMatthew Ahrens } else { 9785d7b4d43SMatthew Ahrens ASSERT(!BP_IS_EMBEDDED(bp)); 9795d7b4d43SMatthew Ahrens } 9805d7b4d43SMatthew Ahrens 981ad135b5dSChristopher Siden if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp)) && BP_GET_LEVEL(bp) == 0) 982e14bb325SJeff Bonwick zio->io_flags |= ZIO_FLAG_DONT_CACHE; 983fa9e4066Sahrens 984bbfd46c4SJeff Bonwick if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP) 985bbfd46c4SJeff Bonwick zio->io_flags |= ZIO_FLAG_DONT_CACHE; 986bbfd46c4SJeff Bonwick 987b24ab676SJeff Bonwick if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL) 988b24ab676SJeff Bonwick zio->io_pipeline = ZIO_DDT_READ_PIPELINE; 989b24ab676SJeff Bonwick 990e14bb325SJeff Bonwick return (ZIO_PIPELINE_CONTINUE); 991fa9e4066Sahrens } 992fa9e4066Sahrens 993e05725b1Sbonwick static int 994e14bb325SJeff Bonwick zio_write_bp_init(zio_t *zio) 9950a4e9518Sgw { 996b24ab676SJeff Bonwick spa_t *spa = zio->io_spa; 997e14bb325SJeff Bonwick zio_prop_t *zp = &zio->io_prop; 998b24ab676SJeff Bonwick enum zio_compress compress = zp->zp_compress; 999e05725b1Sbonwick blkptr_t *bp = zio->io_bp; 1000e14bb325SJeff Bonwick uint64_t lsize = zio->io_size; 1001b24ab676SJeff Bonwick uint64_t psize = lsize; 1002e14bb325SJeff Bonwick int pass = 1; 1003e05725b1Sbonwick 1004e14bb325SJeff Bonwick /* 1005e14bb325SJeff Bonwick * If our children haven't all reached the ready stage, 1006e14bb325SJeff Bonwick * wait for them and then repeat this pipeline stage. 1007e14bb325SJeff Bonwick */ 1008e14bb325SJeff Bonwick if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) || 1009e14bb325SJeff Bonwick zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_READY)) 1010e14bb325SJeff Bonwick return (ZIO_PIPELINE_STOP); 10110a4e9518Sgw 1012e14bb325SJeff Bonwick if (!IO_IS_ALLOCATING(zio)) 1013e14bb325SJeff Bonwick return (ZIO_PIPELINE_CONTINUE); 10140a4e9518Sgw 1015b24ab676SJeff Bonwick ASSERT(zio->io_child_type != ZIO_CHILD_DDT); 1016b24ab676SJeff Bonwick 1017b24ab676SJeff Bonwick if (zio->io_bp_override) { 1018b24ab676SJeff Bonwick ASSERT(bp->blk_birth != zio->io_txg); 1019b24ab676SJeff Bonwick ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0); 1020b24ab676SJeff Bonwick 1021b24ab676SJeff Bonwick *bp = *zio->io_bp_override; 1022b24ab676SJeff Bonwick zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 1023b24ab676SJeff Bonwick 10245d7b4d43SMatthew Ahrens if (BP_IS_EMBEDDED(bp)) 10255d7b4d43SMatthew Ahrens return (ZIO_PIPELINE_CONTINUE); 10265d7b4d43SMatthew Ahrens 102780901aeaSGeorge Wilson /* 102880901aeaSGeorge Wilson * If we've been overridden and nopwrite is set then 102980901aeaSGeorge Wilson * set the flag accordingly to indicate that a nopwrite 103080901aeaSGeorge Wilson * has already occurred. 103180901aeaSGeorge Wilson */ 103280901aeaSGeorge Wilson if (!BP_IS_HOLE(bp) && zp->zp_nopwrite) { 103380901aeaSGeorge Wilson ASSERT(!zp->zp_dedup); 103480901aeaSGeorge Wilson zio->io_flags |= ZIO_FLAG_NOPWRITE; 103580901aeaSGeorge Wilson return (ZIO_PIPELINE_CONTINUE); 103680901aeaSGeorge Wilson } 103780901aeaSGeorge Wilson 103880901aeaSGeorge Wilson ASSERT(!zp->zp_nopwrite); 103980901aeaSGeorge Wilson 1040b24ab676SJeff Bonwick if (BP_IS_HOLE(bp) || !zp->zp_dedup) 1041b24ab676SJeff Bonwick return (ZIO_PIPELINE_CONTINUE); 1042b24ab676SJeff Bonwick 1043b24ab676SJeff Bonwick ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup || 1044b24ab676SJeff Bonwick zp->zp_dedup_verify); 1045b24ab676SJeff Bonwick 1046b24ab676SJeff Bonwick if (BP_GET_CHECKSUM(bp) == zp->zp_checksum) { 1047b24ab676SJeff Bonwick BP_SET_DEDUP(bp, 1); 1048b24ab676SJeff Bonwick zio->io_pipeline |= ZIO_STAGE_DDT_WRITE; 1049b24ab676SJeff Bonwick return (ZIO_PIPELINE_CONTINUE); 1050b24ab676SJeff Bonwick } 1051b24ab676SJeff Bonwick zio->io_bp_override = NULL; 1052b24ab676SJeff Bonwick BP_ZERO(bp); 1053b24ab676SJeff Bonwick } 10540a4e9518Sgw 105543466aaeSMax Grossman if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg) { 1056e14bb325SJeff Bonwick /* 1057e14bb325SJeff Bonwick * We're rewriting an existing block, which means we're 1058e14bb325SJeff Bonwick * working on behalf of spa_sync(). For spa_sync() to 1059e14bb325SJeff Bonwick * converge, it must eventually be the case that we don't 1060e14bb325SJeff Bonwick * have to allocate new blocks. But compression changes 1061e14bb325SJeff Bonwick * the blocksize, which forces a reallocate, and makes 1062e14bb325SJeff Bonwick * convergence take longer. Therefore, after the first 1063e14bb325SJeff Bonwick * few passes, stop compressing to ensure convergence. 1064e14bb325SJeff Bonwick */ 1065b24ab676SJeff Bonwick pass = spa_sync_pass(spa); 1066b24ab676SJeff Bonwick 1067b24ab676SJeff Bonwick ASSERT(zio->io_txg == spa_syncing_txg(spa)); 1068b24ab676SJeff Bonwick ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 1069b24ab676SJeff Bonwick ASSERT(!BP_GET_DEDUP(bp)); 1070e05725b1Sbonwick 107101f55e48SGeorge Wilson if (pass >= zfs_sync_pass_dont_compress) 1072e14bb325SJeff Bonwick compress = ZIO_COMPRESS_OFF; 1073e05725b1Sbonwick 1074e14bb325SJeff Bonwick /* Make sure someone doesn't change their mind on overwrites */ 10755d7b4d43SMatthew Ahrens ASSERT(BP_IS_EMBEDDED(bp) || MIN(zp->zp_copies + BP_IS_GANG(bp), 1076b24ab676SJeff Bonwick spa_max_replication(spa)) == BP_GET_NDVAS(bp)); 1077e14bb325SJeff Bonwick } 1078fa9e4066Sahrens 1079e14bb325SJeff Bonwick if (compress != ZIO_COMPRESS_OFF) { 1080b24ab676SJeff Bonwick void *cbuf = zio_buf_alloc(lsize); 1081b24ab676SJeff Bonwick psize = zio_compress_data(compress, zio->io_data, cbuf, lsize); 1082b24ab676SJeff Bonwick if (psize == 0 || psize == lsize) { 1083e14bb325SJeff Bonwick compress = ZIO_COMPRESS_OFF; 1084b24ab676SJeff Bonwick zio_buf_free(cbuf, lsize); 10855d7b4d43SMatthew Ahrens } else if (!zp->zp_dedup && psize <= BPE_PAYLOAD_SIZE && 10865d7b4d43SMatthew Ahrens zp->zp_level == 0 && !DMU_OT_HAS_FILL(zp->zp_type) && 10875d7b4d43SMatthew Ahrens spa_feature_is_enabled(spa, SPA_FEATURE_EMBEDDED_DATA)) { 10885d7b4d43SMatthew Ahrens encode_embedded_bp_compressed(bp, 10895d7b4d43SMatthew Ahrens cbuf, compress, lsize, psize); 10905d7b4d43SMatthew Ahrens BPE_SET_ETYPE(bp, BP_EMBEDDED_TYPE_DATA); 10915d7b4d43SMatthew Ahrens BP_SET_TYPE(bp, zio->io_prop.zp_type); 10925d7b4d43SMatthew Ahrens BP_SET_LEVEL(bp, zio->io_prop.zp_level); 10935d7b4d43SMatthew Ahrens zio_buf_free(cbuf, lsize); 10945d7b4d43SMatthew Ahrens bp->blk_birth = zio->io_txg; 10955d7b4d43SMatthew Ahrens zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 10965d7b4d43SMatthew Ahrens ASSERT(spa_feature_is_active(spa, 10975d7b4d43SMatthew Ahrens SPA_FEATURE_EMBEDDED_DATA)); 10985d7b4d43SMatthew Ahrens return (ZIO_PIPELINE_CONTINUE); 1099b24ab676SJeff Bonwick } else { 11005d7b4d43SMatthew Ahrens /* 11015d7b4d43SMatthew Ahrens * Round up compressed size to MINBLOCKSIZE and 11025d7b4d43SMatthew Ahrens * zero the tail. 11035d7b4d43SMatthew Ahrens */ 11045d7b4d43SMatthew Ahrens size_t rounded = 11055d7b4d43SMatthew Ahrens P2ROUNDUP(psize, (size_t)SPA_MINBLOCKSIZE); 11065d7b4d43SMatthew Ahrens if (rounded > psize) { 11075d7b4d43SMatthew Ahrens bzero((char *)cbuf + psize, rounded - psize); 11085d7b4d43SMatthew Ahrens psize = rounded; 11095d7b4d43SMatthew Ahrens } 11105d7b4d43SMatthew Ahrens if (psize == lsize) { 11115d7b4d43SMatthew Ahrens compress = ZIO_COMPRESS_OFF; 11125d7b4d43SMatthew Ahrens zio_buf_free(cbuf, lsize); 11135d7b4d43SMatthew Ahrens } else { 11145d7b4d43SMatthew Ahrens zio_push_transform(zio, cbuf, 11155d7b4d43SMatthew Ahrens psize, lsize, NULL); 11165d7b4d43SMatthew Ahrens } 1117e14bb325SJeff Bonwick } 1118e14bb325SJeff Bonwick } 1119c717a561Smaybee 1120e14bb325SJeff Bonwick /* 1121e14bb325SJeff Bonwick * The final pass of spa_sync() must be all rewrites, but the first 1122e14bb325SJeff Bonwick * few passes offer a trade-off: allocating blocks defers convergence, 1123e14bb325SJeff Bonwick * but newly allocated blocks are sequential, so they can be written 1124e14bb325SJeff Bonwick * to disk faster. Therefore, we allow the first few passes of 1125e14bb325SJeff Bonwick * spa_sync() to allocate new blocks, but force rewrites after that. 1126e14bb325SJeff Bonwick * There should only be a handful of blocks after pass 1 in any case. 1127e14bb325SJeff Bonwick */ 112843466aaeSMax Grossman if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg && 112943466aaeSMax Grossman BP_GET_PSIZE(bp) == psize && 113001f55e48SGeorge Wilson pass >= zfs_sync_pass_rewrite) { 1131b24ab676SJeff Bonwick ASSERT(psize != 0); 1132b24ab676SJeff Bonwick enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES; 1133e14bb325SJeff Bonwick zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages; 1134e14bb325SJeff Bonwick zio->io_flags |= ZIO_FLAG_IO_REWRITE; 1135e14bb325SJeff Bonwick } else { 1136e14bb325SJeff Bonwick BP_ZERO(bp); 1137e14bb325SJeff Bonwick zio->io_pipeline = ZIO_WRITE_PIPELINE; 1138e14bb325SJeff Bonwick } 1139fa9e4066Sahrens 1140b24ab676SJeff Bonwick if (psize == 0) { 114143466aaeSMax Grossman if (zio->io_bp_orig.blk_birth != 0 && 114243466aaeSMax Grossman spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) { 114343466aaeSMax Grossman BP_SET_LSIZE(bp, lsize); 114443466aaeSMax Grossman BP_SET_TYPE(bp, zp->zp_type); 114543466aaeSMax Grossman BP_SET_LEVEL(bp, zp->zp_level); 114643466aaeSMax Grossman BP_SET_BIRTH(bp, zio->io_txg, 0); 114743466aaeSMax Grossman } 1148e14bb325SJeff Bonwick zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 1149e14bb325SJeff Bonwick } else { 1150e14bb325SJeff Bonwick ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER); 1151e14bb325SJeff Bonwick BP_SET_LSIZE(bp, lsize); 115243466aaeSMax Grossman BP_SET_TYPE(bp, zp->zp_type); 115343466aaeSMax Grossman BP_SET_LEVEL(bp, zp->zp_level); 1154b24ab676SJeff Bonwick BP_SET_PSIZE(bp, psize); 1155e14bb325SJeff Bonwick BP_SET_COMPRESS(bp, compress); 1156e14bb325SJeff Bonwick BP_SET_CHECKSUM(bp, zp->zp_checksum); 1157b24ab676SJeff Bonwick BP_SET_DEDUP(bp, zp->zp_dedup); 1158e14bb325SJeff Bonwick BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER); 1159b24ab676SJeff Bonwick if (zp->zp_dedup) { 1160b24ab676SJeff Bonwick ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 1161b24ab676SJeff Bonwick ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE)); 1162b24ab676SJeff Bonwick zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE; 1163b24ab676SJeff Bonwick } 116480901aeaSGeorge Wilson if (zp->zp_nopwrite) { 116580901aeaSGeorge Wilson ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 116680901aeaSGeorge Wilson ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE)); 116780901aeaSGeorge Wilson zio->io_pipeline |= ZIO_STAGE_NOP_WRITE; 116880901aeaSGeorge Wilson } 1169b24ab676SJeff Bonwick } 1170b24ab676SJeff Bonwick 1171b24ab676SJeff Bonwick return (ZIO_PIPELINE_CONTINUE); 1172b24ab676SJeff Bonwick } 1173b24ab676SJeff Bonwick 1174b24ab676SJeff Bonwick static int 1175b24ab676SJeff Bonwick zio_free_bp_init(zio_t *zio) 1176b24ab676SJeff Bonwick { 1177b24ab676SJeff Bonwick blkptr_t *bp = zio->io_bp; 1178b24ab676SJeff Bonwick 1179b24ab676SJeff Bonwick if (zio->io_child_type == ZIO_CHILD_LOGICAL) { 1180b24ab676SJeff Bonwick if (BP_GET_DEDUP(bp)) 1181b24ab676SJeff Bonwick zio->io_pipeline = ZIO_DDT_FREE_PIPELINE; 1182e14bb325SJeff Bonwick } 1183fa9e4066Sahrens 1184e05725b1Sbonwick return (ZIO_PIPELINE_CONTINUE); 1185fa9e4066Sahrens } 1186fa9e4066Sahrens 1187e14bb325SJeff Bonwick /* 1188e14bb325SJeff Bonwick * ========================================================================== 1189e14bb325SJeff Bonwick * Execute the I/O pipeline 1190e14bb325SJeff Bonwick * ========================================================================== 1191e14bb325SJeff Bonwick */ 1192e14bb325SJeff Bonwick 1193e14bb325SJeff Bonwick static void 1194ec94d322SAdam Leventhal zio_taskq_dispatch(zio_t *zio, zio_taskq_type_t q, boolean_t cutinline) 1195fa9e4066Sahrens { 119680eb36f2SGeorge Wilson spa_t *spa = zio->io_spa; 1197e14bb325SJeff Bonwick zio_type_t t = zio->io_type; 11985aeb9474SGarrett D'Amore int flags = (cutinline ? TQ_FRONT : 0); 11990a4e9518Sgw 12000a4e9518Sgw /* 1201bbe36defSGeorge Wilson * If we're a config writer or a probe, the normal issue and 1202bbe36defSGeorge Wilson * interrupt threads may all be blocked waiting for the config lock. 1203bbe36defSGeorge Wilson * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL. 12040a4e9518Sgw */ 1205bbe36defSGeorge Wilson if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE)) 1206e14bb325SJeff Bonwick t = ZIO_TYPE_NULL; 12070a4e9518Sgw 12080a4e9518Sgw /* 1209e14bb325SJeff Bonwick * A similar issue exists for the L2ARC write thread until L2ARC 2.0. 12100a4e9518Sgw */ 1211e14bb325SJeff Bonwick if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux) 1212e14bb325SJeff Bonwick t = ZIO_TYPE_NULL; 12130a4e9518Sgw 121480eb36f2SGeorge Wilson /* 1215ec94d322SAdam Leventhal * If this is a high priority I/O, then use the high priority taskq if 1216ec94d322SAdam Leventhal * available. 121780eb36f2SGeorge Wilson */ 121880eb36f2SGeorge Wilson if (zio->io_priority == ZIO_PRIORITY_NOW && 1219ec94d322SAdam Leventhal spa->spa_zio_taskq[t][q + 1].stqs_count != 0) 122080eb36f2SGeorge Wilson q++; 122180eb36f2SGeorge Wilson 122280eb36f2SGeorge Wilson ASSERT3U(q, <, ZIO_TASKQ_TYPES); 12235aeb9474SGarrett D'Amore 12245aeb9474SGarrett D'Amore /* 12255aeb9474SGarrett D'Amore * NB: We are assuming that the zio can only be dispatched 12265aeb9474SGarrett D'Amore * to a single taskq at a time. It would be a grievous error 12275aeb9474SGarrett D'Amore * to dispatch the zio to another taskq at the same time. 12285aeb9474SGarrett D'Amore */ 12295aeb9474SGarrett D'Amore ASSERT(zio->io_tqent.tqent_next == NULL); 1230ec94d322SAdam Leventhal spa_taskq_dispatch_ent(spa, t, q, (task_func_t *)zio_execute, zio, 1231ec94d322SAdam Leventhal flags, &zio->io_tqent); 1232e14bb325SJeff Bonwick } 12330a4e9518Sgw 1234e14bb325SJeff Bonwick static boolean_t 1235ec94d322SAdam Leventhal zio_taskq_member(zio_t *zio, zio_taskq_type_t q) 1236e14bb325SJeff Bonwick { 1237e14bb325SJeff Bonwick kthread_t *executor = zio->io_executor; 1238e14bb325SJeff Bonwick spa_t *spa = zio->io_spa; 12390a4e9518Sgw 1240ec94d322SAdam Leventhal for (zio_type_t t = 0; t < ZIO_TYPES; t++) { 1241ec94d322SAdam Leventhal spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q]; 1242ec94d322SAdam Leventhal uint_t i; 1243ec94d322SAdam Leventhal for (i = 0; i < tqs->stqs_count; i++) { 1244ec94d322SAdam Leventhal if (taskq_member(tqs->stqs_taskq[i], executor)) 1245ec94d322SAdam Leventhal return (B_TRUE); 1246ec94d322SAdam Leventhal } 1247ec94d322SAdam Leventhal } 12480a4e9518Sgw 1249e14bb325SJeff Bonwick return (B_FALSE); 1250e14bb325SJeff Bonwick } 1251e05725b1Sbonwick 1252e14bb325SJeff Bonwick static int 1253e14bb325SJeff Bonwick zio_issue_async(zio_t *zio) 1254e14bb325SJeff Bonwick { 125535a5a358SJonathan Adams zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE); 1256e14bb325SJeff Bonwick 1257e14bb325SJeff Bonwick return (ZIO_PIPELINE_STOP); 12580a4e9518Sgw } 12590a4e9518Sgw 1260e14bb325SJeff Bonwick void 1261e14bb325SJeff Bonwick zio_interrupt(zio_t *zio) 12620a4e9518Sgw { 126335a5a358SJonathan Adams zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE); 1264e14bb325SJeff Bonwick } 12650a4e9518Sgw 1266e14bb325SJeff Bonwick /* 1267e14bb325SJeff Bonwick * Execute the I/O pipeline until one of the following occurs: 1268f7170741SWill Andrews * 1269f7170741SWill Andrews * (1) the I/O completes 1270f7170741SWill Andrews * (2) the pipeline stalls waiting for dependent child I/Os 1271f7170741SWill Andrews * (3) the I/O issues, so we're waiting for an I/O completion interrupt 1272f7170741SWill Andrews * (4) the I/O is delegated by vdev-level caching or aggregation 1273f7170741SWill Andrews * (5) the I/O is deferred due to vdev-level queueing 1274f7170741SWill Andrews * (6) the I/O is handed off to another thread. 1275f7170741SWill Andrews * 1276f7170741SWill Andrews * In all cases, the pipeline stops whenever there's no CPU work; it never 1277f7170741SWill Andrews * burns a thread in cv_wait(). 1278e14bb325SJeff Bonwick * 1279e14bb325SJeff Bonwick * There's no locking on io_stage because there's no legitimate way 1280e14bb325SJeff Bonwick * for multiple threads to be attempting to process the same I/O. 1281e14bb325SJeff Bonwick */ 1282b24ab676SJeff Bonwick static zio_pipe_stage_t *zio_pipeline[]; 12830a4e9518Sgw 1284e14bb325SJeff Bonwick void 1285e14bb325SJeff Bonwick zio_execute(zio_t *zio) 1286e14bb325SJeff Bonwick { 1287e14bb325SJeff Bonwick zio->io_executor = curthread; 12880a4e9518Sgw 1289e14bb325SJeff Bonwick while (zio->io_stage < ZIO_STAGE_DONE) { 1290b24ab676SJeff Bonwick enum zio_stage pipeline = zio->io_pipeline; 1291b24ab676SJeff Bonwick enum zio_stage stage = zio->io_stage; 1292e14bb325SJeff Bonwick int rv; 12930a4e9518Sgw 1294e14bb325SJeff Bonwick ASSERT(!MUTEX_HELD(&zio->io_lock)); 1295b24ab676SJeff Bonwick ASSERT(ISP2(stage)); 1296b24ab676SJeff Bonwick ASSERT(zio->io_stall == NULL); 12970a4e9518Sgw 1298b24ab676SJeff Bonwick do { 1299b24ab676SJeff Bonwick stage <<= 1; 1300b24ab676SJeff Bonwick } while ((stage & pipeline) == 0); 1301e14bb325SJeff Bonwick 1302e14bb325SJeff Bonwick ASSERT(stage <= ZIO_STAGE_DONE); 13030a4e9518Sgw 13040a4e9518Sgw /* 1305e14bb325SJeff Bonwick * If we are in interrupt context and this pipeline stage 1306e14bb325SJeff Bonwick * will grab a config lock that is held across I/O, 1307b24ab676SJeff Bonwick * or may wait for an I/O that needs an interrupt thread 1308b24ab676SJeff Bonwick * to complete, issue async to avoid deadlock. 130935a5a358SJonathan Adams * 131035a5a358SJonathan Adams * For VDEV_IO_START, we cut in line so that the io will 131135a5a358SJonathan Adams * be sent to disk promptly. 13120a4e9518Sgw */ 1313b24ab676SJeff Bonwick if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL && 1314e14bb325SJeff Bonwick zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) { 131535a5a358SJonathan Adams boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ? 131635a5a358SJonathan Adams zio_requeue_io_start_cut_in_line : B_FALSE; 131735a5a358SJonathan Adams zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut); 1318e14bb325SJeff Bonwick return; 13190a4e9518Sgw } 13200a4e9518Sgw 1321e14bb325SJeff Bonwick zio->io_stage = stage; 1322bf16b11eSMatthew Ahrens rv = zio_pipeline[highbit64(stage) - 1](zio); 13230a4e9518Sgw 1324e14bb325SJeff Bonwick if (rv == ZIO_PIPELINE_STOP) 1325e14bb325SJeff Bonwick return; 13260a4e9518Sgw 1327e14bb325SJeff Bonwick ASSERT(rv == ZIO_PIPELINE_CONTINUE); 1328e14bb325SJeff Bonwick } 13290a4e9518Sgw } 13300a4e9518Sgw 1331e14bb325SJeff Bonwick /* 1332e14bb325SJeff Bonwick * ========================================================================== 1333e14bb325SJeff Bonwick * Initiate I/O, either sync or async 1334e14bb325SJeff Bonwick * ========================================================================== 1335e14bb325SJeff Bonwick */ 1336e14bb325SJeff Bonwick int 1337e14bb325SJeff Bonwick zio_wait(zio_t *zio) 13380a4e9518Sgw { 1339e14bb325SJeff Bonwick int error; 13400a4e9518Sgw 1341e14bb325SJeff Bonwick ASSERT(zio->io_stage == ZIO_STAGE_OPEN); 1342e14bb325SJeff Bonwick ASSERT(zio->io_executor == NULL); 13430a4e9518Sgw 1344e14bb325SJeff Bonwick zio->io_waiter = curthread; 1345e05725b1Sbonwick 1346e14bb325SJeff Bonwick zio_execute(zio); 13470a4e9518Sgw 1348e14bb325SJeff Bonwick mutex_enter(&zio->io_lock); 1349e14bb325SJeff Bonwick while (zio->io_executor != NULL) 1350e14bb325SJeff Bonwick cv_wait(&zio->io_cv, &zio->io_lock); 1351e14bb325SJeff Bonwick mutex_exit(&zio->io_lock); 135232b87932Sek 1353e14bb325SJeff Bonwick error = zio->io_error; 1354e14bb325SJeff Bonwick zio_destroy(zio); 135532b87932Sek 1356e14bb325SJeff Bonwick return (error); 135732b87932Sek } 135832b87932Sek 1359e14bb325SJeff Bonwick void 1360e14bb325SJeff Bonwick zio_nowait(zio_t *zio) 13610a4e9518Sgw { 1362e14bb325SJeff Bonwick ASSERT(zio->io_executor == NULL); 1363fa9e4066Sahrens 1364a3f829aeSBill Moore if (zio->io_child_type == ZIO_CHILD_LOGICAL && 1365a3f829aeSBill Moore zio_unique_parent(zio) == NULL) { 1366ea8dc4b6Seschrock /* 1367e14bb325SJeff Bonwick * This is a logical async I/O with no parent to wait for it. 136854d692b7SGeorge Wilson * We add it to the spa_async_root_zio "Godfather" I/O which 136954d692b7SGeorge Wilson * will ensure they complete prior to unloading the pool. 1370ea8dc4b6Seschrock */ 1371e14bb325SJeff Bonwick spa_t *spa = zio->io_spa; 137254d692b7SGeorge Wilson 13736f834bc1SMatthew Ahrens zio_add_child(spa->spa_async_zio_root[CPU_SEQID], zio); 1374e14bb325SJeff Bonwick } 1375ea8dc4b6Seschrock 1376e14bb325SJeff Bonwick zio_execute(zio); 1377e14bb325SJeff Bonwick } 1378ea8dc4b6Seschrock 1379e14bb325SJeff Bonwick /* 1380e14bb325SJeff Bonwick * ========================================================================== 1381e14bb325SJeff Bonwick * Reexecute or suspend/resume failed I/O 1382e14bb325SJeff Bonwick * ========================================================================== 1383e14bb325SJeff Bonwick */ 1384fa9e4066Sahrens 1385e14bb325SJeff Bonwick static void 1386e14bb325SJeff Bonwick zio_reexecute(zio_t *pio) 1387e14bb325SJeff Bonwick { 1388a3f829aeSBill Moore zio_t *cio, *cio_next; 1389a3f829aeSBill Moore 1390a3f829aeSBill Moore ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL); 1391a3f829aeSBill Moore ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN); 1392f5383399SBill Moore ASSERT(pio->io_gang_leader == NULL); 1393f5383399SBill Moore ASSERT(pio->io_gang_tree == NULL); 1394e05725b1Sbonwick 1395e14bb325SJeff Bonwick pio->io_flags = pio->io_orig_flags; 1396e14bb325SJeff Bonwick pio->io_stage = pio->io_orig_stage; 1397e14bb325SJeff Bonwick pio->io_pipeline = pio->io_orig_pipeline; 1398e14bb325SJeff Bonwick pio->io_reexecute = 0; 139980901aeaSGeorge Wilson pio->io_flags |= ZIO_FLAG_REEXECUTED; 1400e14bb325SJeff Bonwick pio->io_error = 0; 1401a3f829aeSBill Moore for (int w = 0; w < ZIO_WAIT_TYPES; w++) 1402a3f829aeSBill Moore pio->io_state[w] = 0; 1403e14bb325SJeff Bonwick for (int c = 0; c < ZIO_CHILD_TYPES; c++) 1404e14bb325SJeff Bonwick pio->io_child_error[c] = 0; 14050a4e9518Sgw 1406b24ab676SJeff Bonwick if (IO_IS_ALLOCATING(pio)) 1407b24ab676SJeff Bonwick BP_ZERO(pio->io_bp); 1408d58459f4Sek 1409e14bb325SJeff Bonwick /* 1410e14bb325SJeff Bonwick * As we reexecute pio's children, new children could be created. 1411a3f829aeSBill Moore * New children go to the head of pio's io_child_list, however, 1412e14bb325SJeff Bonwick * so we will (correctly) not reexecute them. The key is that 1413a3f829aeSBill Moore * the remainder of pio's io_child_list, from 'cio_next' onward, 1414a3f829aeSBill Moore * cannot be affected by any side effects of reexecuting 'cio'. 1415e14bb325SJeff Bonwick */ 1416a3f829aeSBill Moore for (cio = zio_walk_children(pio); cio != NULL; cio = cio_next) { 1417a3f829aeSBill Moore cio_next = zio_walk_children(pio); 1418e14bb325SJeff Bonwick mutex_enter(&pio->io_lock); 1419a3f829aeSBill Moore for (int w = 0; w < ZIO_WAIT_TYPES; w++) 1420a3f829aeSBill Moore pio->io_children[cio->io_child_type][w]++; 1421e14bb325SJeff Bonwick mutex_exit(&pio->io_lock); 1422a3f829aeSBill Moore zio_reexecute(cio); 1423fa9e4066Sahrens } 1424e05725b1Sbonwick 1425e14bb325SJeff Bonwick /* 1426e14bb325SJeff Bonwick * Now that all children have been reexecuted, execute the parent. 142754d692b7SGeorge Wilson * We don't reexecute "The Godfather" I/O here as it's the 142854d692b7SGeorge Wilson * responsibility of the caller to wait on him. 1429e14bb325SJeff Bonwick */ 143054d692b7SGeorge Wilson if (!(pio->io_flags & ZIO_FLAG_GODFATHER)) 143154d692b7SGeorge Wilson zio_execute(pio); 14320a4e9518Sgw } 14330a4e9518Sgw 1434e14bb325SJeff Bonwick void 1435e14bb325SJeff Bonwick zio_suspend(spa_t *spa, zio_t *zio) 14360a4e9518Sgw { 1437e14bb325SJeff Bonwick if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC) 1438e14bb325SJeff Bonwick fm_panic("Pool '%s' has encountered an uncorrectable I/O " 1439e14bb325SJeff Bonwick "failure and the failure mode property for this pool " 1440e14bb325SJeff Bonwick "is set to panic.", spa_name(spa)); 14410a4e9518Sgw 1442e14bb325SJeff Bonwick zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0); 14430a4e9518Sgw 1444e14bb325SJeff Bonwick mutex_enter(&spa->spa_suspend_lock); 1445fa9e4066Sahrens 1446e14bb325SJeff Bonwick if (spa->spa_suspend_zio_root == NULL) 144754d692b7SGeorge Wilson spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL, 144854d692b7SGeorge Wilson ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | 144954d692b7SGeorge Wilson ZIO_FLAG_GODFATHER); 1450fa9e4066Sahrens 1451e14bb325SJeff Bonwick spa->spa_suspended = B_TRUE; 1452fa9e4066Sahrens 1453e14bb325SJeff Bonwick if (zio != NULL) { 145454d692b7SGeorge Wilson ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER)); 1455e14bb325SJeff Bonwick ASSERT(zio != spa->spa_suspend_zio_root); 1456e14bb325SJeff Bonwick ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 1457a3f829aeSBill Moore ASSERT(zio_unique_parent(zio) == NULL); 1458e14bb325SJeff Bonwick ASSERT(zio->io_stage == ZIO_STAGE_DONE); 1459e14bb325SJeff Bonwick zio_add_child(spa->spa_suspend_zio_root, zio); 1460e14bb325SJeff Bonwick } 1461fa9e4066Sahrens 1462e14bb325SJeff Bonwick mutex_exit(&spa->spa_suspend_lock); 1463e14bb325SJeff Bonwick } 1464fa9e4066Sahrens 146554d692b7SGeorge Wilson int 1466e14bb325SJeff Bonwick zio_resume(spa_t *spa) 1467e14bb325SJeff Bonwick { 146854d692b7SGeorge Wilson zio_t *pio; 1469fa9e4066Sahrens 1470b3995adbSahrens /* 1471e14bb325SJeff Bonwick * Reexecute all previously suspended i/o. 1472b3995adbSahrens */ 1473e14bb325SJeff Bonwick mutex_enter(&spa->spa_suspend_lock); 1474e14bb325SJeff Bonwick spa->spa_suspended = B_FALSE; 1475e14bb325SJeff Bonwick cv_broadcast(&spa->spa_suspend_cv); 1476e14bb325SJeff Bonwick pio = spa->spa_suspend_zio_root; 1477e14bb325SJeff Bonwick spa->spa_suspend_zio_root = NULL; 1478e14bb325SJeff Bonwick mutex_exit(&spa->spa_suspend_lock); 1479e14bb325SJeff Bonwick 1480e14bb325SJeff Bonwick if (pio == NULL) 148154d692b7SGeorge Wilson return (0); 1482e14bb325SJeff Bonwick 148354d692b7SGeorge Wilson zio_reexecute(pio); 148454d692b7SGeorge Wilson return (zio_wait(pio)); 1485e14bb325SJeff Bonwick } 1486e14bb325SJeff Bonwick 1487e14bb325SJeff Bonwick void 1488e14bb325SJeff Bonwick zio_resume_wait(spa_t *spa) 1489e14bb325SJeff Bonwick { 1490e14bb325SJeff Bonwick mutex_enter(&spa->spa_suspend_lock); 1491e14bb325SJeff Bonwick while (spa_suspended(spa)) 1492e14bb325SJeff Bonwick cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock); 1493e14bb325SJeff Bonwick mutex_exit(&spa->spa_suspend_lock); 1494fa9e4066Sahrens } 1495fa9e4066Sahrens 1496fa9e4066Sahrens /* 1497fa9e4066Sahrens * ========================================================================== 1498e14bb325SJeff Bonwick * Gang blocks. 1499e14bb325SJeff Bonwick * 1500e14bb325SJeff Bonwick * A gang block is a collection of small blocks that looks to the DMU 1501e14bb325SJeff Bonwick * like one large block. When zio_dva_allocate() cannot find a block 1502e14bb325SJeff Bonwick * of the requested size, due to either severe fragmentation or the pool 1503e14bb325SJeff Bonwick * being nearly full, it calls zio_write_gang_block() to construct the 1504e14bb325SJeff Bonwick * block from smaller fragments. 1505e14bb325SJeff Bonwick * 1506e14bb325SJeff Bonwick * A gang block consists of a gang header (zio_gbh_phys_t) and up to 1507e14bb325SJeff Bonwick * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like 1508e14bb325SJeff Bonwick * an indirect block: it's an array of block pointers. It consumes 1509e14bb325SJeff Bonwick * only one sector and hence is allocatable regardless of fragmentation. 1510e14bb325SJeff Bonwick * The gang header's bps point to its gang members, which hold the data. 1511e14bb325SJeff Bonwick * 1512e14bb325SJeff Bonwick * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg> 1513e14bb325SJeff Bonwick * as the verifier to ensure uniqueness of the SHA256 checksum. 1514e14bb325SJeff Bonwick * Critically, the gang block bp's blk_cksum is the checksum of the data, 1515e14bb325SJeff Bonwick * not the gang header. This ensures that data block signatures (needed for 1516e14bb325SJeff Bonwick * deduplication) are independent of how the block is physically stored. 1517e14bb325SJeff Bonwick * 1518e14bb325SJeff Bonwick * Gang blocks can be nested: a gang member may itself be a gang block. 1519e14bb325SJeff Bonwick * Thus every gang block is a tree in which root and all interior nodes are 1520e14bb325SJeff Bonwick * gang headers, and the leaves are normal blocks that contain user data. 1521e14bb325SJeff Bonwick * The root of the gang tree is called the gang leader. 1522e14bb325SJeff Bonwick * 1523e14bb325SJeff Bonwick * To perform any operation (read, rewrite, free, claim) on a gang block, 1524e14bb325SJeff Bonwick * zio_gang_assemble() first assembles the gang tree (minus data leaves) 1525e14bb325SJeff Bonwick * in the io_gang_tree field of the original logical i/o by recursively 1526e14bb325SJeff Bonwick * reading the gang leader and all gang headers below it. This yields 1527e14bb325SJeff Bonwick * an in-core tree containing the contents of every gang header and the 1528e14bb325SJeff Bonwick * bps for every constituent of the gang block. 1529e14bb325SJeff Bonwick * 1530e14bb325SJeff Bonwick * With the gang tree now assembled, zio_gang_issue() just walks the gang tree 1531e14bb325SJeff Bonwick * and invokes a callback on each bp. To free a gang block, zio_gang_issue() 1532e14bb325SJeff Bonwick * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp. 1533e14bb325SJeff Bonwick * zio_claim_gang() provides a similarly trivial wrapper for zio_claim(). 1534e14bb325SJeff Bonwick * zio_read_gang() is a wrapper around zio_read() that omits reading gang 1535e14bb325SJeff Bonwick * headers, since we already have those in io_gang_tree. zio_rewrite_gang() 1536e14bb325SJeff Bonwick * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite() 1537e14bb325SJeff Bonwick * of the gang header plus zio_checksum_compute() of the data to update the 1538e14bb325SJeff Bonwick * gang header's blk_cksum as described above. 1539e14bb325SJeff Bonwick * 1540e14bb325SJeff Bonwick * The two-phase assemble/issue model solves the problem of partial failure -- 1541e14bb325SJeff Bonwick * what if you'd freed part of a gang block but then couldn't read the 1542e14bb325SJeff Bonwick * gang header for another part? Assembling the entire gang tree first 1543e14bb325SJeff Bonwick * ensures that all the necessary gang header I/O has succeeded before 1544e14bb325SJeff Bonwick * starting the actual work of free, claim, or write. Once the gang tree 1545e14bb325SJeff Bonwick * is assembled, free and claim are in-memory operations that cannot fail. 1546e14bb325SJeff Bonwick * 1547e14bb325SJeff Bonwick * In the event that a gang write fails, zio_dva_unallocate() walks the 1548e14bb325SJeff Bonwick * gang tree to immediately free (i.e. insert back into the space map) 1549e14bb325SJeff Bonwick * everything we've allocated. This ensures that we don't get ENOSPC 1550e14bb325SJeff Bonwick * errors during repeated suspend/resume cycles due to a flaky device. 1551e14bb325SJeff Bonwick * 1552e14bb325SJeff Bonwick * Gang rewrites only happen during sync-to-convergence. If we can't assemble 1553e14bb325SJeff Bonwick * the gang tree, we won't modify the block, so we can safely defer the free 1554e14bb325SJeff Bonwick * (knowing that the block is still intact). If we *can* assemble the gang 1555e14bb325SJeff Bonwick * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free 1556e14bb325SJeff Bonwick * each constituent bp and we can allocate a new block on the next sync pass. 1557e14bb325SJeff Bonwick * 1558e14bb325SJeff Bonwick * In all cases, the gang tree allows complete recovery from partial failure. 1559fa9e4066Sahrens * ========================================================================== 1560fa9e4066Sahrens */ 1561e14bb325SJeff Bonwick 1562e14bb325SJeff Bonwick static zio_t * 1563e14bb325SJeff Bonwick zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data) 1564fa9e4066Sahrens { 1565e14bb325SJeff Bonwick if (gn != NULL) 1566e14bb325SJeff Bonwick return (pio); 1567fa9e4066Sahrens 1568e14bb325SJeff Bonwick return (zio_read(pio, pio->io_spa, bp, data, BP_GET_PSIZE(bp), 1569e14bb325SJeff Bonwick NULL, NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), 1570e14bb325SJeff Bonwick &pio->io_bookmark)); 1571e14bb325SJeff Bonwick } 1572e14bb325SJeff Bonwick 1573e14bb325SJeff Bonwick zio_t * 1574e14bb325SJeff Bonwick zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data) 1575e14bb325SJeff Bonwick { 1576e14bb325SJeff Bonwick zio_t *zio; 1577e14bb325SJeff Bonwick 1578e14bb325SJeff Bonwick if (gn != NULL) { 1579e14bb325SJeff Bonwick zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp, 1580e14bb325SJeff Bonwick gn->gn_gbh, SPA_GANGBLOCKSIZE, NULL, NULL, pio->io_priority, 1581e14bb325SJeff Bonwick ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark); 1582fa9e4066Sahrens /* 1583e14bb325SJeff Bonwick * As we rewrite each gang header, the pipeline will compute 1584e14bb325SJeff Bonwick * a new gang block header checksum for it; but no one will 1585e14bb325SJeff Bonwick * compute a new data checksum, so we do that here. The one 1586e14bb325SJeff Bonwick * exception is the gang leader: the pipeline already computed 1587e14bb325SJeff Bonwick * its data checksum because that stage precedes gang assembly. 1588e14bb325SJeff Bonwick * (Presently, nothing actually uses interior data checksums; 1589e14bb325SJeff Bonwick * this is just good hygiene.) 1590fa9e4066Sahrens */ 1591f5383399SBill Moore if (gn != pio->io_gang_leader->io_gang_tree) { 1592e14bb325SJeff Bonwick zio_checksum_compute(zio, BP_GET_CHECKSUM(bp), 1593e14bb325SJeff Bonwick data, BP_GET_PSIZE(bp)); 1594e14bb325SJeff Bonwick } 1595b24ab676SJeff Bonwick /* 1596b24ab676SJeff Bonwick * If we are here to damage data for testing purposes, 1597b24ab676SJeff Bonwick * leave the GBH alone so that we can detect the damage. 1598b24ab676SJeff Bonwick */ 1599b24ab676SJeff Bonwick if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE) 1600b24ab676SJeff Bonwick zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES; 1601fa9e4066Sahrens } else { 1602e14bb325SJeff Bonwick zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp, 1603e14bb325SJeff Bonwick data, BP_GET_PSIZE(bp), NULL, NULL, pio->io_priority, 1604e14bb325SJeff Bonwick ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark); 1605fa9e4066Sahrens } 1606fa9e4066Sahrens 1607e14bb325SJeff Bonwick return (zio); 1608e14bb325SJeff Bonwick } 1609fa9e4066Sahrens 1610e14bb325SJeff Bonwick /* ARGSUSED */ 1611e14bb325SJeff Bonwick zio_t * 1612e14bb325SJeff Bonwick zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data) 1613e14bb325SJeff Bonwick { 1614b24ab676SJeff Bonwick return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp, 1615b24ab676SJeff Bonwick ZIO_GANG_CHILD_FLAGS(pio))); 1616fa9e4066Sahrens } 1617fa9e4066Sahrens 1618e14bb325SJeff Bonwick /* ARGSUSED */ 1619e14bb325SJeff Bonwick zio_t * 1620e14bb325SJeff Bonwick zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data) 1621fa9e4066Sahrens { 1622e14bb325SJeff Bonwick return (zio_claim(pio, pio->io_spa, pio->io_txg, bp, 1623e14bb325SJeff Bonwick NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio))); 1624e14bb325SJeff Bonwick } 1625fa9e4066Sahrens 1626e14bb325SJeff Bonwick static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = { 1627e14bb325SJeff Bonwick NULL, 1628e14bb325SJeff Bonwick zio_read_gang, 1629e14bb325SJeff Bonwick zio_rewrite_gang, 1630e14bb325SJeff Bonwick zio_free_gang, 1631e14bb325SJeff Bonwick zio_claim_gang, 1632e14bb325SJeff Bonwick NULL 1633e14bb325SJeff Bonwick }; 1634fa9e4066Sahrens 1635e14bb325SJeff Bonwick static void zio_gang_tree_assemble_done(zio_t *zio); 1636fa9e4066Sahrens 1637e14bb325SJeff Bonwick static zio_gang_node_t * 1638e14bb325SJeff Bonwick zio_gang_node_alloc(zio_gang_node_t **gnpp) 1639e14bb325SJeff Bonwick { 1640e14bb325SJeff Bonwick zio_gang_node_t *gn; 1641fa9e4066Sahrens 1642e14bb325SJeff Bonwick ASSERT(*gnpp == NULL); 1643fa9e4066Sahrens 1644e14bb325SJeff Bonwick gn = kmem_zalloc(sizeof (*gn), KM_SLEEP); 1645e14bb325SJeff Bonwick gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE); 1646e14bb325SJeff Bonwick *gnpp = gn; 1647e14bb325SJeff Bonwick 1648e14bb325SJeff Bonwick return (gn); 1649fa9e4066Sahrens } 1650fa9e4066Sahrens 1651fa9e4066Sahrens static void 1652e14bb325SJeff Bonwick zio_gang_node_free(zio_gang_node_t **gnpp) 1653fa9e4066Sahrens { 1654e14bb325SJeff Bonwick zio_gang_node_t *gn = *gnpp; 1655fa9e4066Sahrens 1656e14bb325SJeff Bonwick for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) 1657e14bb325SJeff Bonwick ASSERT(gn->gn_child[g] == NULL); 1658e14bb325SJeff Bonwick 1659e14bb325SJeff Bonwick zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE); 1660e14bb325SJeff Bonwick kmem_free(gn, sizeof (*gn)); 1661e14bb325SJeff Bonwick *gnpp = NULL; 1662fa9e4066Sahrens } 1663fa9e4066Sahrens 1664e14bb325SJeff Bonwick static void 1665e14bb325SJeff Bonwick zio_gang_tree_free(zio_gang_node_t **gnpp) 1666fa9e4066Sahrens { 1667e14bb325SJeff Bonwick zio_gang_node_t *gn = *gnpp; 1668fa9e4066Sahrens 1669e14bb325SJeff Bonwick if (gn == NULL) 1670e14bb325SJeff Bonwick return; 1671fa9e4066Sahrens 1672e14bb325SJeff Bonwick for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) 1673e14bb325SJeff Bonwick zio_gang_tree_free(&gn->gn_child[g]); 1674fa9e4066Sahrens 1675e14bb325SJeff Bonwick zio_gang_node_free(gnpp); 1676fa9e4066Sahrens } 1677fa9e4066Sahrens 1678e14bb325SJeff Bonwick static void 1679f5383399SBill Moore zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp) 1680fa9e4066Sahrens { 1681e14bb325SJeff Bonwick zio_gang_node_t *gn = zio_gang_node_alloc(gnpp); 1682e14bb325SJeff Bonwick 1683f5383399SBill Moore ASSERT(gio->io_gang_leader == gio); 1684e14bb325SJeff Bonwick ASSERT(BP_IS_GANG(bp)); 1685fa9e4066Sahrens 1686f5383399SBill Moore zio_nowait(zio_read(gio, gio->io_spa, bp, gn->gn_gbh, 1687e14bb325SJeff Bonwick SPA_GANGBLOCKSIZE, zio_gang_tree_assemble_done, gn, 1688f5383399SBill Moore gio->io_priority, ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark)); 1689e14bb325SJeff Bonwick } 1690fa9e4066Sahrens 1691e14bb325SJeff Bonwick static void 1692e14bb325SJeff Bonwick zio_gang_tree_assemble_done(zio_t *zio) 1693e14bb325SJeff Bonwick { 1694f5383399SBill Moore zio_t *gio = zio->io_gang_leader; 1695e14bb325SJeff Bonwick zio_gang_node_t *gn = zio->io_private; 1696e14bb325SJeff Bonwick blkptr_t *bp = zio->io_bp; 1697fa9e4066Sahrens 1698f5383399SBill Moore ASSERT(gio == zio_unique_parent(zio)); 1699b24ab676SJeff Bonwick ASSERT(zio->io_child_count == 0); 1700fa9e4066Sahrens 1701e14bb325SJeff Bonwick if (zio->io_error) 1702e14bb325SJeff Bonwick return; 1703fa9e4066Sahrens 1704e14bb325SJeff Bonwick if (BP_SHOULD_BYTESWAP(bp)) 1705e14bb325SJeff Bonwick byteswap_uint64_array(zio->io_data, zio->io_size); 1706fa9e4066Sahrens 1707e14bb325SJeff Bonwick ASSERT(zio->io_data == gn->gn_gbh); 1708e14bb325SJeff Bonwick ASSERT(zio->io_size == SPA_GANGBLOCKSIZE); 17096e1f5caaSNeil Perrin ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC); 1710e05725b1Sbonwick 1711e14bb325SJeff Bonwick for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) { 1712e14bb325SJeff Bonwick blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g]; 1713e14bb325SJeff Bonwick if (!BP_IS_GANG(gbp)) 1714e14bb325SJeff Bonwick continue; 1715f5383399SBill Moore zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]); 1716e14bb325SJeff Bonwick } 1717fa9e4066Sahrens } 1718fa9e4066Sahrens 1719e14bb325SJeff Bonwick static void 1720e14bb325SJeff Bonwick zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, void *data) 1721fa9e4066Sahrens { 1722f5383399SBill Moore zio_t *gio = pio->io_gang_leader; 1723e14bb325SJeff Bonwick zio_t *zio; 1724fa9e4066Sahrens 1725e14bb325SJeff Bonwick ASSERT(BP_IS_GANG(bp) == !!gn); 1726f5383399SBill Moore ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp)); 1727f5383399SBill Moore ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree); 1728fa9e4066Sahrens 1729e14bb325SJeff Bonwick /* 1730e14bb325SJeff Bonwick * If you're a gang header, your data is in gn->gn_gbh. 1731e14bb325SJeff Bonwick * If you're a gang member, your data is in 'data' and gn == NULL. 1732e14bb325SJeff Bonwick */ 1733f5383399SBill Moore zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data); 1734fa9e4066Sahrens 1735e14bb325SJeff Bonwick if (gn != NULL) { 17366e1f5caaSNeil Perrin ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC); 1737fa9e4066Sahrens 1738e14bb325SJeff Bonwick for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) { 1739e14bb325SJeff Bonwick blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g]; 1740e14bb325SJeff Bonwick if (BP_IS_HOLE(gbp)) 1741e14bb325SJeff Bonwick continue; 1742e14bb325SJeff Bonwick zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data); 1743e14bb325SJeff Bonwick data = (char *)data + BP_GET_PSIZE(gbp); 1744e14bb325SJeff Bonwick } 1745fa9e4066Sahrens } 1746fa9e4066Sahrens 1747f5383399SBill Moore if (gn == gio->io_gang_tree) 1748f5383399SBill Moore ASSERT3P((char *)gio->io_data + gio->io_size, ==, data); 1749e05725b1Sbonwick 1750e14bb325SJeff Bonwick if (zio != pio) 1751e14bb325SJeff Bonwick zio_nowait(zio); 1752fa9e4066Sahrens } 1753fa9e4066Sahrens 1754e05725b1Sbonwick static int 1755e14bb325SJeff Bonwick zio_gang_assemble(zio_t *zio) 1756fa9e4066Sahrens { 1757e14bb325SJeff Bonwick blkptr_t *bp = zio->io_bp; 1758fa9e4066Sahrens 1759f5383399SBill Moore ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL); 1760f5383399SBill Moore ASSERT(zio->io_child_type > ZIO_CHILD_GANG); 1761f5383399SBill Moore 1762f5383399SBill Moore zio->io_gang_leader = zio; 1763fa9e4066Sahrens 1764e14bb325SJeff Bonwick zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree); 1765e05725b1Sbonwick 1766e05725b1Sbonwick return (ZIO_PIPELINE_CONTINUE); 1767fa9e4066Sahrens } 1768fa9e4066Sahrens 1769e05725b1Sbonwick static int 1770e14bb325SJeff Bonwick zio_gang_issue(zio_t *zio) 1771fa9e4066Sahrens { 1772e14bb325SJeff Bonwick blkptr_t *bp = zio->io_bp; 1773fa9e4066Sahrens 1774e14bb325SJeff Bonwick if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE)) 1775e14bb325SJeff Bonwick return (ZIO_PIPELINE_STOP); 1776fa9e4066Sahrens 1777f5383399SBill Moore ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio); 1778f5383399SBill Moore ASSERT(zio->io_child_type > ZIO_CHILD_GANG); 1779fa9e4066Sahrens 1780e14bb325SJeff Bonwick if (zio->io_child_error[ZIO_CHILD_GANG] == 0) 1781f5383399SBill Moore zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_data); 1782e14bb325SJeff Bonwick else 1783f5383399SBill Moore zio_gang_tree_free(&zio->io_gang_tree); 1784fa9e4066Sahrens 1785e14bb325SJeff Bonwick zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 1786e05725b1Sbonwick 1787e05725b1Sbonwick return (ZIO_PIPELINE_CONTINUE); 1788fa9e4066Sahrens } 1789fa9e4066Sahrens 1790fa9e4066Sahrens static void 1791e14bb325SJeff Bonwick zio_write_gang_member_ready(zio_t *zio) 1792fa9e4066Sahrens { 1793a3f829aeSBill Moore zio_t *pio = zio_unique_parent(zio); 1794f5383399SBill Moore zio_t *gio = zio->io_gang_leader; 179544cd46caSbillm dva_t *cdva = zio->io_bp->blk_dva; 179644cd46caSbillm dva_t *pdva = pio->io_bp->blk_dva; 1797fa9e4066Sahrens uint64_t asize; 1798fa9e4066Sahrens 1799e14bb325SJeff Bonwick if (BP_IS_HOLE(zio->io_bp)) 1800e14bb325SJeff Bonwick return; 1801e14bb325SJeff Bonwick 1802e14bb325SJeff Bonwick ASSERT(BP_IS_HOLE(&zio->io_bp_orig)); 1803e14bb325SJeff Bonwick 1804e14bb325SJeff Bonwick ASSERT(zio->io_child_type == ZIO_CHILD_GANG); 1805b24ab676SJeff Bonwick ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies); 1806b24ab676SJeff Bonwick ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp)); 1807b24ab676SJeff Bonwick ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp)); 180844cd46caSbillm ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp)); 1809fa9e4066Sahrens 1810fa9e4066Sahrens mutex_enter(&pio->io_lock); 1811e14bb325SJeff Bonwick for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) { 181244cd46caSbillm ASSERT(DVA_GET_GANG(&pdva[d])); 181344cd46caSbillm asize = DVA_GET_ASIZE(&pdva[d]); 181444cd46caSbillm asize += DVA_GET_ASIZE(&cdva[d]); 181544cd46caSbillm DVA_SET_ASIZE(&pdva[d], asize); 181644cd46caSbillm } 1817fa9e4066Sahrens mutex_exit(&pio->io_lock); 1818fa9e4066Sahrens } 1819fa9e4066Sahrens 18200a4e9518Sgw static int 1821e14bb325SJeff Bonwick zio_write_gang_block(zio_t *pio) 1822fa9e4066Sahrens { 1823e14bb325SJeff Bonwick spa_t *spa = pio->io_spa; 1824e14bb325SJeff Bonwick blkptr_t *bp = pio->io_bp; 1825f5383399SBill Moore zio_t *gio = pio->io_gang_leader; 1826e14bb325SJeff Bonwick zio_t *zio; 1827e14bb325SJeff Bonwick zio_gang_node_t *gn, **gnpp; 1828fa9e4066Sahrens zio_gbh_phys_t *gbh; 1829e14bb325SJeff Bonwick uint64_t txg = pio->io_txg; 1830e14bb325SJeff Bonwick uint64_t resid = pio->io_size; 1831e14bb325SJeff Bonwick uint64_t lsize; 1832b24ab676SJeff Bonwick int copies = gio->io_prop.zp_copies; 1833b24ab676SJeff Bonwick int gbh_copies = MIN(copies + 1, spa_max_replication(spa)); 1834e14bb325SJeff Bonwick zio_prop_t zp; 1835fa9e4066Sahrens int error; 1836fa9e4066Sahrens 1837b24ab676SJeff Bonwick error = metaslab_alloc(spa, spa_normal_class(spa), SPA_GANGBLOCKSIZE, 1838b24ab676SJeff Bonwick bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp, 1839e14bb325SJeff Bonwick METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER); 1840e05725b1Sbonwick if (error) { 1841e14bb325SJeff Bonwick pio->io_error = error; 1842e05725b1Sbonwick return (ZIO_PIPELINE_CONTINUE); 1843e05725b1Sbonwick } 1844fa9e4066Sahrens 1845f5383399SBill Moore if (pio == gio) { 1846f5383399SBill Moore gnpp = &gio->io_gang_tree; 1847e14bb325SJeff Bonwick } else { 1848e14bb325SJeff Bonwick gnpp = pio->io_private; 1849e14bb325SJeff Bonwick ASSERT(pio->io_ready == zio_write_gang_member_ready); 1850fa9e4066Sahrens } 1851fa9e4066Sahrens 1852e14bb325SJeff Bonwick gn = zio_gang_node_alloc(gnpp); 1853e14bb325SJeff Bonwick gbh = gn->gn_gbh; 1854e14bb325SJeff Bonwick bzero(gbh, SPA_GANGBLOCKSIZE); 1855fa9e4066Sahrens 1856e14bb325SJeff Bonwick /* 1857e14bb325SJeff Bonwick * Create the gang header. 1858e14bb325SJeff Bonwick */ 1859e14bb325SJeff Bonwick zio = zio_rewrite(pio, spa, txg, bp, gbh, SPA_GANGBLOCKSIZE, NULL, NULL, 1860e14bb325SJeff Bonwick pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark); 1861fa9e4066Sahrens 1862e14bb325SJeff Bonwick /* 1863e14bb325SJeff Bonwick * Create and nowait the gang children. 1864e14bb325SJeff Bonwick */ 1865e14bb325SJeff Bonwick for (int g = 0; resid != 0; resid -= lsize, g++) { 1866e14bb325SJeff Bonwick lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g), 1867e14bb325SJeff Bonwick SPA_MINBLOCKSIZE); 1868e14bb325SJeff Bonwick ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid); 1869e14bb325SJeff Bonwick 1870f5383399SBill Moore zp.zp_checksum = gio->io_prop.zp_checksum; 1871e14bb325SJeff Bonwick zp.zp_compress = ZIO_COMPRESS_OFF; 1872e14bb325SJeff Bonwick zp.zp_type = DMU_OT_NONE; 1873e14bb325SJeff Bonwick zp.zp_level = 0; 1874b24ab676SJeff Bonwick zp.zp_copies = gio->io_prop.zp_copies; 187580901aeaSGeorge Wilson zp.zp_dedup = B_FALSE; 187680901aeaSGeorge Wilson zp.zp_dedup_verify = B_FALSE; 187780901aeaSGeorge Wilson zp.zp_nopwrite = B_FALSE; 1878e14bb325SJeff Bonwick 1879e14bb325SJeff Bonwick zio_nowait(zio_write(zio, spa, txg, &gbh->zg_blkptr[g], 1880e14bb325SJeff Bonwick (char *)pio->io_data + (pio->io_size - resid), lsize, &zp, 188169962b56SMatthew Ahrens zio_write_gang_member_ready, NULL, NULL, &gn->gn_child[g], 1882e14bb325SJeff Bonwick pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), 1883e14bb325SJeff Bonwick &pio->io_bookmark)); 1884e14bb325SJeff Bonwick } 1885e05725b1Sbonwick 188644cd46caSbillm /* 1887e14bb325SJeff Bonwick * Set pio's pipeline to just wait for zio to finish. 188844cd46caSbillm */ 1889e14bb325SJeff Bonwick pio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 1890e14bb325SJeff Bonwick 1891e14bb325SJeff Bonwick zio_nowait(zio); 1892e14bb325SJeff Bonwick 1893e14bb325SJeff Bonwick return (ZIO_PIPELINE_CONTINUE); 1894fa9e4066Sahrens } 1895fa9e4066Sahrens 189680901aeaSGeorge Wilson /* 189780901aeaSGeorge Wilson * The zio_nop_write stage in the pipeline determines if allocating 189880901aeaSGeorge Wilson * a new bp is necessary. By leveraging a cryptographically secure checksum, 189980901aeaSGeorge Wilson * such as SHA256, we can compare the checksums of the new data and the old 190080901aeaSGeorge Wilson * to determine if allocating a new block is required. The nopwrite 190180901aeaSGeorge Wilson * feature can handle writes in either syncing or open context (i.e. zil 190280901aeaSGeorge Wilson * writes) and as a result is mutually exclusive with dedup. 190380901aeaSGeorge Wilson */ 190480901aeaSGeorge Wilson static int 190580901aeaSGeorge Wilson zio_nop_write(zio_t *zio) 190680901aeaSGeorge Wilson { 190780901aeaSGeorge Wilson blkptr_t *bp = zio->io_bp; 190880901aeaSGeorge Wilson blkptr_t *bp_orig = &zio->io_bp_orig; 190980901aeaSGeorge Wilson zio_prop_t *zp = &zio->io_prop; 191080901aeaSGeorge Wilson 191180901aeaSGeorge Wilson ASSERT(BP_GET_LEVEL(bp) == 0); 191280901aeaSGeorge Wilson ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE)); 191380901aeaSGeorge Wilson ASSERT(zp->zp_nopwrite); 191480901aeaSGeorge Wilson ASSERT(!zp->zp_dedup); 191580901aeaSGeorge Wilson ASSERT(zio->io_bp_override == NULL); 191680901aeaSGeorge Wilson ASSERT(IO_IS_ALLOCATING(zio)); 191780901aeaSGeorge Wilson 191880901aeaSGeorge Wilson /* 191980901aeaSGeorge Wilson * Check to see if the original bp and the new bp have matching 192080901aeaSGeorge Wilson * characteristics (i.e. same checksum, compression algorithms, etc). 192180901aeaSGeorge Wilson * If they don't then just continue with the pipeline which will 192280901aeaSGeorge Wilson * allocate a new bp. 192380901aeaSGeorge Wilson */ 192480901aeaSGeorge Wilson if (BP_IS_HOLE(bp_orig) || 192580901aeaSGeorge Wilson !zio_checksum_table[BP_GET_CHECKSUM(bp)].ci_dedup || 192680901aeaSGeorge Wilson BP_GET_CHECKSUM(bp) != BP_GET_CHECKSUM(bp_orig) || 192780901aeaSGeorge Wilson BP_GET_COMPRESS(bp) != BP_GET_COMPRESS(bp_orig) || 192880901aeaSGeorge Wilson BP_GET_DEDUP(bp) != BP_GET_DEDUP(bp_orig) || 192980901aeaSGeorge Wilson zp->zp_copies != BP_GET_NDVAS(bp_orig)) 193080901aeaSGeorge Wilson return (ZIO_PIPELINE_CONTINUE); 193180901aeaSGeorge Wilson 193280901aeaSGeorge Wilson /* 193380901aeaSGeorge Wilson * If the checksums match then reset the pipeline so that we 193480901aeaSGeorge Wilson * avoid allocating a new bp and issuing any I/O. 193580901aeaSGeorge Wilson */ 193680901aeaSGeorge Wilson if (ZIO_CHECKSUM_EQUAL(bp->blk_cksum, bp_orig->blk_cksum)) { 193780901aeaSGeorge Wilson ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup); 193880901aeaSGeorge Wilson ASSERT3U(BP_GET_PSIZE(bp), ==, BP_GET_PSIZE(bp_orig)); 193980901aeaSGeorge Wilson ASSERT3U(BP_GET_LSIZE(bp), ==, BP_GET_LSIZE(bp_orig)); 194080901aeaSGeorge Wilson ASSERT(zp->zp_compress != ZIO_COMPRESS_OFF); 194180901aeaSGeorge Wilson ASSERT(bcmp(&bp->blk_prop, &bp_orig->blk_prop, 194280901aeaSGeorge Wilson sizeof (uint64_t)) == 0); 194380901aeaSGeorge Wilson 194480901aeaSGeorge Wilson *bp = *bp_orig; 194580901aeaSGeorge Wilson zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 194680901aeaSGeorge Wilson zio->io_flags |= ZIO_FLAG_NOPWRITE; 194780901aeaSGeorge Wilson } 194880901aeaSGeorge Wilson 194980901aeaSGeorge Wilson return (ZIO_PIPELINE_CONTINUE); 195080901aeaSGeorge Wilson } 195180901aeaSGeorge Wilson 1952fa9e4066Sahrens /* 1953fa9e4066Sahrens * ========================================================================== 1954b24ab676SJeff Bonwick * Dedup 1955fa9e4066Sahrens * ========================================================================== 1956fa9e4066Sahrens */ 1957b24ab676SJeff Bonwick static void 1958b24ab676SJeff Bonwick zio_ddt_child_read_done(zio_t *zio) 1959b24ab676SJeff Bonwick { 1960b24ab676SJeff Bonwick blkptr_t *bp = zio->io_bp; 1961b24ab676SJeff Bonwick ddt_entry_t *dde = zio->io_private; 1962b24ab676SJeff Bonwick ddt_phys_t *ddp; 1963b24ab676SJeff Bonwick zio_t *pio = zio_unique_parent(zio); 1964b24ab676SJeff Bonwick 1965b24ab676SJeff Bonwick mutex_enter(&pio->io_lock); 1966b24ab676SJeff Bonwick ddp = ddt_phys_select(dde, bp); 1967b24ab676SJeff Bonwick if (zio->io_error == 0) 1968b24ab676SJeff Bonwick ddt_phys_clear(ddp); /* this ddp doesn't need repair */ 1969b24ab676SJeff Bonwick if (zio->io_error == 0 && dde->dde_repair_data == NULL) 1970b24ab676SJeff Bonwick dde->dde_repair_data = zio->io_data; 1971b24ab676SJeff Bonwick else 1972b24ab676SJeff Bonwick zio_buf_free(zio->io_data, zio->io_size); 1973b24ab676SJeff Bonwick mutex_exit(&pio->io_lock); 1974b24ab676SJeff Bonwick } 1975b24ab676SJeff Bonwick 1976b24ab676SJeff Bonwick static int 1977b24ab676SJeff Bonwick zio_ddt_read_start(zio_t *zio) 1978b24ab676SJeff Bonwick { 1979b24ab676SJeff Bonwick blkptr_t *bp = zio->io_bp; 1980b24ab676SJeff Bonwick 1981b24ab676SJeff Bonwick ASSERT(BP_GET_DEDUP(bp)); 1982b24ab676SJeff Bonwick ASSERT(BP_GET_PSIZE(bp) == zio->io_size); 1983b24ab676SJeff Bonwick ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 1984b24ab676SJeff Bonwick 1985b24ab676SJeff Bonwick if (zio->io_child_error[ZIO_CHILD_DDT]) { 1986b24ab676SJeff Bonwick ddt_t *ddt = ddt_select(zio->io_spa, bp); 1987b24ab676SJeff Bonwick ddt_entry_t *dde = ddt_repair_start(ddt, bp); 1988b24ab676SJeff Bonwick ddt_phys_t *ddp = dde->dde_phys; 1989b24ab676SJeff Bonwick ddt_phys_t *ddp_self = ddt_phys_select(dde, bp); 1990b24ab676SJeff Bonwick blkptr_t blk; 1991b24ab676SJeff Bonwick 1992b24ab676SJeff Bonwick ASSERT(zio->io_vsd == NULL); 1993b24ab676SJeff Bonwick zio->io_vsd = dde; 1994b24ab676SJeff Bonwick 1995b24ab676SJeff Bonwick if (ddp_self == NULL) 1996b24ab676SJeff Bonwick return (ZIO_PIPELINE_CONTINUE); 1997b24ab676SJeff Bonwick 1998b24ab676SJeff Bonwick for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) { 1999b24ab676SJeff Bonwick if (ddp->ddp_phys_birth == 0 || ddp == ddp_self) 2000b24ab676SJeff Bonwick continue; 2001bbfd46c4SJeff Bonwick ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp, 2002bbfd46c4SJeff Bonwick &blk); 2003b24ab676SJeff Bonwick zio_nowait(zio_read(zio, zio->io_spa, &blk, 2004b24ab676SJeff Bonwick zio_buf_alloc(zio->io_size), zio->io_size, 2005b24ab676SJeff Bonwick zio_ddt_child_read_done, dde, zio->io_priority, 2006b24ab676SJeff Bonwick ZIO_DDT_CHILD_FLAGS(zio) | ZIO_FLAG_DONT_PROPAGATE, 2007b24ab676SJeff Bonwick &zio->io_bookmark)); 2008b24ab676SJeff Bonwick } 2009b24ab676SJeff Bonwick return (ZIO_PIPELINE_CONTINUE); 2010b24ab676SJeff Bonwick } 2011b24ab676SJeff Bonwick 2012b24ab676SJeff Bonwick zio_nowait(zio_read(zio, zio->io_spa, bp, 2013b24ab676SJeff Bonwick zio->io_data, zio->io_size, NULL, NULL, zio->io_priority, 2014b24ab676SJeff Bonwick ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark)); 2015b24ab676SJeff Bonwick 2016b24ab676SJeff Bonwick return (ZIO_PIPELINE_CONTINUE); 2017b24ab676SJeff Bonwick } 2018e14bb325SJeff Bonwick 2019b24ab676SJeff Bonwick static int 2020b24ab676SJeff Bonwick zio_ddt_read_done(zio_t *zio) 2021b24ab676SJeff Bonwick { 2022b24ab676SJeff Bonwick blkptr_t *bp = zio->io_bp; 2023b24ab676SJeff Bonwick 2024b24ab676SJeff Bonwick if (zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE)) 2025b24ab676SJeff Bonwick return (ZIO_PIPELINE_STOP); 2026b24ab676SJeff Bonwick 2027b24ab676SJeff Bonwick ASSERT(BP_GET_DEDUP(bp)); 2028b24ab676SJeff Bonwick ASSERT(BP_GET_PSIZE(bp) == zio->io_size); 2029b24ab676SJeff Bonwick ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 2030b24ab676SJeff Bonwick 2031b24ab676SJeff Bonwick if (zio->io_child_error[ZIO_CHILD_DDT]) { 2032b24ab676SJeff Bonwick ddt_t *ddt = ddt_select(zio->io_spa, bp); 2033b24ab676SJeff Bonwick ddt_entry_t *dde = zio->io_vsd; 2034b24ab676SJeff Bonwick if (ddt == NULL) { 2035b16da2e2SGeorge Wilson ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE); 2036b24ab676SJeff Bonwick return (ZIO_PIPELINE_CONTINUE); 2037b24ab676SJeff Bonwick } 2038b24ab676SJeff Bonwick if (dde == NULL) { 2039b24ab676SJeff Bonwick zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1; 204035a5a358SJonathan Adams zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE); 2041b24ab676SJeff Bonwick return (ZIO_PIPELINE_STOP); 2042b24ab676SJeff Bonwick } 2043b24ab676SJeff Bonwick if (dde->dde_repair_data != NULL) { 2044b24ab676SJeff Bonwick bcopy(dde->dde_repair_data, zio->io_data, zio->io_size); 2045b24ab676SJeff Bonwick zio->io_child_error[ZIO_CHILD_DDT] = 0; 2046b24ab676SJeff Bonwick } 2047b24ab676SJeff Bonwick ddt_repair_done(ddt, dde); 2048b24ab676SJeff Bonwick zio->io_vsd = NULL; 2049b24ab676SJeff Bonwick } 2050b24ab676SJeff Bonwick 2051b24ab676SJeff Bonwick ASSERT(zio->io_vsd == NULL); 2052b24ab676SJeff Bonwick 2053b24ab676SJeff Bonwick return (ZIO_PIPELINE_CONTINUE); 2054b24ab676SJeff Bonwick } 2055b24ab676SJeff Bonwick 2056b24ab676SJeff Bonwick static boolean_t 2057b24ab676SJeff Bonwick zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde) 2058b24ab676SJeff Bonwick { 2059b24ab676SJeff Bonwick spa_t *spa = zio->io_spa; 2060b24ab676SJeff Bonwick 2061b24ab676SJeff Bonwick /* 2062b24ab676SJeff Bonwick * Note: we compare the original data, not the transformed data, 2063b24ab676SJeff Bonwick * because when zio->io_bp is an override bp, we will not have 2064b24ab676SJeff Bonwick * pushed the I/O transforms. That's an important optimization 2065b24ab676SJeff Bonwick * because otherwise we'd compress/encrypt all dmu_sync() data twice. 2066b24ab676SJeff Bonwick */ 2067b24ab676SJeff Bonwick for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) { 2068b24ab676SJeff Bonwick zio_t *lio = dde->dde_lead_zio[p]; 2069b24ab676SJeff Bonwick 2070b24ab676SJeff Bonwick if (lio != NULL) { 2071b24ab676SJeff Bonwick return (lio->io_orig_size != zio->io_orig_size || 2072b24ab676SJeff Bonwick bcmp(zio->io_orig_data, lio->io_orig_data, 2073b24ab676SJeff Bonwick zio->io_orig_size) != 0); 2074b24ab676SJeff Bonwick } 2075b24ab676SJeff Bonwick } 2076b24ab676SJeff Bonwick 2077b24ab676SJeff Bonwick for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) { 2078b24ab676SJeff Bonwick ddt_phys_t *ddp = &dde->dde_phys[p]; 2079b24ab676SJeff Bonwick 2080b24ab676SJeff Bonwick if (ddp->ddp_phys_birth != 0) { 2081b24ab676SJeff Bonwick arc_buf_t *abuf = NULL; 2082b24ab676SJeff Bonwick uint32_t aflags = ARC_WAIT; 2083b24ab676SJeff Bonwick blkptr_t blk = *zio->io_bp; 2084b24ab676SJeff Bonwick int error; 2085b24ab676SJeff Bonwick 2086b24ab676SJeff Bonwick ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth); 2087b24ab676SJeff Bonwick 2088b24ab676SJeff Bonwick ddt_exit(ddt); 2089b24ab676SJeff Bonwick 20901b912ec7SGeorge Wilson error = arc_read(NULL, spa, &blk, 2091b24ab676SJeff Bonwick arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ, 2092b24ab676SJeff Bonwick ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, 2093b24ab676SJeff Bonwick &aflags, &zio->io_bookmark); 2094b24ab676SJeff Bonwick 2095b24ab676SJeff Bonwick if (error == 0) { 2096b24ab676SJeff Bonwick if (arc_buf_size(abuf) != zio->io_orig_size || 2097b24ab676SJeff Bonwick bcmp(abuf->b_data, zio->io_orig_data, 2098b24ab676SJeff Bonwick zio->io_orig_size) != 0) 2099be6fd75aSMatthew Ahrens error = SET_ERROR(EEXIST); 21003b2aab18SMatthew Ahrens VERIFY(arc_buf_remove_ref(abuf, &abuf)); 2101b24ab676SJeff Bonwick } 2102b24ab676SJeff Bonwick 2103b24ab676SJeff Bonwick ddt_enter(ddt); 2104b24ab676SJeff Bonwick return (error != 0); 2105b24ab676SJeff Bonwick } 2106b24ab676SJeff Bonwick } 2107b24ab676SJeff Bonwick 2108b24ab676SJeff Bonwick return (B_FALSE); 2109b24ab676SJeff Bonwick } 2110b24ab676SJeff Bonwick 2111b24ab676SJeff Bonwick static void 2112b24ab676SJeff Bonwick zio_ddt_child_write_ready(zio_t *zio) 2113b24ab676SJeff Bonwick { 2114b24ab676SJeff Bonwick int p = zio->io_prop.zp_copies; 2115b24ab676SJeff Bonwick ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp); 2116b24ab676SJeff Bonwick ddt_entry_t *dde = zio->io_private; 2117b24ab676SJeff Bonwick ddt_phys_t *ddp = &dde->dde_phys[p]; 2118b24ab676SJeff Bonwick zio_t *pio; 2119b24ab676SJeff Bonwick 2120b24ab676SJeff Bonwick if (zio->io_error) 2121b24ab676SJeff Bonwick return; 2122b24ab676SJeff Bonwick 2123b24ab676SJeff Bonwick ddt_enter(ddt); 2124b24ab676SJeff Bonwick 2125b24ab676SJeff Bonwick ASSERT(dde->dde_lead_zio[p] == zio); 2126b24ab676SJeff Bonwick 2127b24ab676SJeff Bonwick ddt_phys_fill(ddp, zio->io_bp); 2128b24ab676SJeff Bonwick 2129b24ab676SJeff Bonwick while ((pio = zio_walk_parents(zio)) != NULL) 2130b24ab676SJeff Bonwick ddt_bp_fill(ddp, pio->io_bp, zio->io_txg); 2131b24ab676SJeff Bonwick 2132b24ab676SJeff Bonwick ddt_exit(ddt); 2133b24ab676SJeff Bonwick } 2134b24ab676SJeff Bonwick 2135b24ab676SJeff Bonwick static void 2136b24ab676SJeff Bonwick zio_ddt_child_write_done(zio_t *zio) 2137b24ab676SJeff Bonwick { 2138b24ab676SJeff Bonwick int p = zio->io_prop.zp_copies; 2139b24ab676SJeff Bonwick ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp); 2140b24ab676SJeff Bonwick ddt_entry_t *dde = zio->io_private; 2141b24ab676SJeff Bonwick ddt_phys_t *ddp = &dde->dde_phys[p]; 2142b24ab676SJeff Bonwick 2143b24ab676SJeff Bonwick ddt_enter(ddt); 2144b24ab676SJeff Bonwick 2145b24ab676SJeff Bonwick ASSERT(ddp->ddp_refcnt == 0); 2146b24ab676SJeff Bonwick ASSERT(dde->dde_lead_zio[p] == zio); 2147b24ab676SJeff Bonwick dde->dde_lead_zio[p] = NULL; 2148b24ab676SJeff Bonwick 2149b24ab676SJeff Bonwick if (zio->io_error == 0) { 2150b24ab676SJeff Bonwick while (zio_walk_parents(zio) != NULL) 2151b24ab676SJeff Bonwick ddt_phys_addref(ddp); 2152b24ab676SJeff Bonwick } else { 2153b24ab676SJeff Bonwick ddt_phys_clear(ddp); 2154b24ab676SJeff Bonwick } 2155b24ab676SJeff Bonwick 2156b24ab676SJeff Bonwick ddt_exit(ddt); 2157b24ab676SJeff Bonwick } 2158b24ab676SJeff Bonwick 2159b24ab676SJeff Bonwick static void 2160b24ab676SJeff Bonwick zio_ddt_ditto_write_done(zio_t *zio) 2161b24ab676SJeff Bonwick { 2162b24ab676SJeff Bonwick int p = DDT_PHYS_DITTO; 2163b24ab676SJeff Bonwick zio_prop_t *zp = &zio->io_prop; 2164b24ab676SJeff Bonwick blkptr_t *bp = zio->io_bp; 2165b24ab676SJeff Bonwick ddt_t *ddt = ddt_select(zio->io_spa, bp); 2166b24ab676SJeff Bonwick ddt_entry_t *dde = zio->io_private; 2167b24ab676SJeff Bonwick ddt_phys_t *ddp = &dde->dde_phys[p]; 2168b24ab676SJeff Bonwick ddt_key_t *ddk = &dde->dde_key; 2169b24ab676SJeff Bonwick 2170b24ab676SJeff Bonwick ddt_enter(ddt); 2171b24ab676SJeff Bonwick 2172b24ab676SJeff Bonwick ASSERT(ddp->ddp_refcnt == 0); 2173b24ab676SJeff Bonwick ASSERT(dde->dde_lead_zio[p] == zio); 2174b24ab676SJeff Bonwick dde->dde_lead_zio[p] = NULL; 2175b24ab676SJeff Bonwick 2176b24ab676SJeff Bonwick if (zio->io_error == 0) { 2177b24ab676SJeff Bonwick ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum)); 2178b24ab676SJeff Bonwick ASSERT(zp->zp_copies < SPA_DVAS_PER_BP); 2179b24ab676SJeff Bonwick ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp)); 2180b24ab676SJeff Bonwick if (ddp->ddp_phys_birth != 0) 2181b24ab676SJeff Bonwick ddt_phys_free(ddt, ddk, ddp, zio->io_txg); 2182b24ab676SJeff Bonwick ddt_phys_fill(ddp, bp); 2183b24ab676SJeff Bonwick } 2184b24ab676SJeff Bonwick 2185b24ab676SJeff Bonwick ddt_exit(ddt); 2186b24ab676SJeff Bonwick } 2187b24ab676SJeff Bonwick 2188b24ab676SJeff Bonwick static int 2189b24ab676SJeff Bonwick zio_ddt_write(zio_t *zio) 2190b24ab676SJeff Bonwick { 2191b24ab676SJeff Bonwick spa_t *spa = zio->io_spa; 2192b24ab676SJeff Bonwick blkptr_t *bp = zio->io_bp; 2193b24ab676SJeff Bonwick uint64_t txg = zio->io_txg; 2194b24ab676SJeff Bonwick zio_prop_t *zp = &zio->io_prop; 2195b24ab676SJeff Bonwick int p = zp->zp_copies; 2196b24ab676SJeff Bonwick int ditto_copies; 2197b24ab676SJeff Bonwick zio_t *cio = NULL; 2198b24ab676SJeff Bonwick zio_t *dio = NULL; 2199b24ab676SJeff Bonwick ddt_t *ddt = ddt_select(spa, bp); 2200b24ab676SJeff Bonwick ddt_entry_t *dde; 2201b24ab676SJeff Bonwick ddt_phys_t *ddp; 2202b24ab676SJeff Bonwick 2203b24ab676SJeff Bonwick ASSERT(BP_GET_DEDUP(bp)); 2204b24ab676SJeff Bonwick ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum); 2205b24ab676SJeff Bonwick ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override); 2206b24ab676SJeff Bonwick 2207b24ab676SJeff Bonwick ddt_enter(ddt); 2208b24ab676SJeff Bonwick dde = ddt_lookup(ddt, bp, B_TRUE); 2209b24ab676SJeff Bonwick ddp = &dde->dde_phys[p]; 2210b24ab676SJeff Bonwick 2211b24ab676SJeff Bonwick if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) { 2212b24ab676SJeff Bonwick /* 2213b24ab676SJeff Bonwick * If we're using a weak checksum, upgrade to a strong checksum 2214b24ab676SJeff Bonwick * and try again. If we're already using a strong checksum, 2215b24ab676SJeff Bonwick * we can't resolve it, so just convert to an ordinary write. 2216b24ab676SJeff Bonwick * (And automatically e-mail a paper to Nature?) 2217b24ab676SJeff Bonwick */ 2218b24ab676SJeff Bonwick if (!zio_checksum_table[zp->zp_checksum].ci_dedup) { 2219b24ab676SJeff Bonwick zp->zp_checksum = spa_dedup_checksum(spa); 2220b24ab676SJeff Bonwick zio_pop_transforms(zio); 2221b24ab676SJeff Bonwick zio->io_stage = ZIO_STAGE_OPEN; 2222b24ab676SJeff Bonwick BP_ZERO(bp); 2223b24ab676SJeff Bonwick } else { 222480901aeaSGeorge Wilson zp->zp_dedup = B_FALSE; 2225b24ab676SJeff Bonwick } 2226b24ab676SJeff Bonwick zio->io_pipeline = ZIO_WRITE_PIPELINE; 2227b24ab676SJeff Bonwick ddt_exit(ddt); 2228b24ab676SJeff Bonwick return (ZIO_PIPELINE_CONTINUE); 2229b24ab676SJeff Bonwick } 2230b24ab676SJeff Bonwick 2231b24ab676SJeff Bonwick ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp); 2232b24ab676SJeff Bonwick ASSERT(ditto_copies < SPA_DVAS_PER_BP); 2233b24ab676SJeff Bonwick 2234b24ab676SJeff Bonwick if (ditto_copies > ddt_ditto_copies_present(dde) && 2235b24ab676SJeff Bonwick dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) { 2236b24ab676SJeff Bonwick zio_prop_t czp = *zp; 2237b24ab676SJeff Bonwick 2238b24ab676SJeff Bonwick czp.zp_copies = ditto_copies; 2239b24ab676SJeff Bonwick 2240b24ab676SJeff Bonwick /* 2241b24ab676SJeff Bonwick * If we arrived here with an override bp, we won't have run 2242b24ab676SJeff Bonwick * the transform stack, so we won't have the data we need to 2243b24ab676SJeff Bonwick * generate a child i/o. So, toss the override bp and restart. 2244b24ab676SJeff Bonwick * This is safe, because using the override bp is just an 2245b24ab676SJeff Bonwick * optimization; and it's rare, so the cost doesn't matter. 2246b24ab676SJeff Bonwick */ 2247b24ab676SJeff Bonwick if (zio->io_bp_override) { 2248b24ab676SJeff Bonwick zio_pop_transforms(zio); 2249b24ab676SJeff Bonwick zio->io_stage = ZIO_STAGE_OPEN; 2250b24ab676SJeff Bonwick zio->io_pipeline = ZIO_WRITE_PIPELINE; 2251b24ab676SJeff Bonwick zio->io_bp_override = NULL; 2252b24ab676SJeff Bonwick BP_ZERO(bp); 2253b24ab676SJeff Bonwick ddt_exit(ddt); 2254b24ab676SJeff Bonwick return (ZIO_PIPELINE_CONTINUE); 2255b24ab676SJeff Bonwick } 2256b24ab676SJeff Bonwick 2257b24ab676SJeff Bonwick dio = zio_write(zio, spa, txg, bp, zio->io_orig_data, 225869962b56SMatthew Ahrens zio->io_orig_size, &czp, NULL, NULL, 2259b24ab676SJeff Bonwick zio_ddt_ditto_write_done, dde, zio->io_priority, 2260b24ab676SJeff Bonwick ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark); 2261b24ab676SJeff Bonwick 2262b24ab676SJeff Bonwick zio_push_transform(dio, zio->io_data, zio->io_size, 0, NULL); 2263b24ab676SJeff Bonwick dde->dde_lead_zio[DDT_PHYS_DITTO] = dio; 2264b24ab676SJeff Bonwick } 2265b24ab676SJeff Bonwick 2266b24ab676SJeff Bonwick if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) { 2267b24ab676SJeff Bonwick if (ddp->ddp_phys_birth != 0) 2268b24ab676SJeff Bonwick ddt_bp_fill(ddp, bp, txg); 2269b24ab676SJeff Bonwick if (dde->dde_lead_zio[p] != NULL) 2270b24ab676SJeff Bonwick zio_add_child(zio, dde->dde_lead_zio[p]); 2271b24ab676SJeff Bonwick else 2272b24ab676SJeff Bonwick ddt_phys_addref(ddp); 2273b24ab676SJeff Bonwick } else if (zio->io_bp_override) { 2274b24ab676SJeff Bonwick ASSERT(bp->blk_birth == txg); 2275b24ab676SJeff Bonwick ASSERT(BP_EQUAL(bp, zio->io_bp_override)); 2276b24ab676SJeff Bonwick ddt_phys_fill(ddp, bp); 2277b24ab676SJeff Bonwick ddt_phys_addref(ddp); 2278b24ab676SJeff Bonwick } else { 2279b24ab676SJeff Bonwick cio = zio_write(zio, spa, txg, bp, zio->io_orig_data, 228069962b56SMatthew Ahrens zio->io_orig_size, zp, zio_ddt_child_write_ready, NULL, 2281b24ab676SJeff Bonwick zio_ddt_child_write_done, dde, zio->io_priority, 2282b24ab676SJeff Bonwick ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark); 2283b24ab676SJeff Bonwick 2284b24ab676SJeff Bonwick zio_push_transform(cio, zio->io_data, zio->io_size, 0, NULL); 2285b24ab676SJeff Bonwick dde->dde_lead_zio[p] = cio; 2286b24ab676SJeff Bonwick } 2287b24ab676SJeff Bonwick 2288b24ab676SJeff Bonwick ddt_exit(ddt); 2289b24ab676SJeff Bonwick 2290b24ab676SJeff Bonwick if (cio) 2291b24ab676SJeff Bonwick zio_nowait(cio); 2292b24ab676SJeff Bonwick if (dio) 2293b24ab676SJeff Bonwick zio_nowait(dio); 2294b24ab676SJeff Bonwick 2295b24ab676SJeff Bonwick return (ZIO_PIPELINE_CONTINUE); 2296b24ab676SJeff Bonwick } 2297b24ab676SJeff Bonwick 22983f9d6ad7SLin Ling ddt_entry_t *freedde; /* for debugging */ 22993f9d6ad7SLin Ling 2300b24ab676SJeff Bonwick static int 2301b24ab676SJeff Bonwick zio_ddt_free(zio_t *zio) 2302b24ab676SJeff Bonwick { 2303b24ab676SJeff Bonwick spa_t *spa = zio->io_spa; 2304b24ab676SJeff Bonwick blkptr_t *bp = zio->io_bp; 2305b24ab676SJeff Bonwick ddt_t *ddt = ddt_select(spa, bp); 2306b24ab676SJeff Bonwick ddt_entry_t *dde; 2307b24ab676SJeff Bonwick ddt_phys_t *ddp; 2308b24ab676SJeff Bonwick 2309b24ab676SJeff Bonwick ASSERT(BP_GET_DEDUP(bp)); 2310b24ab676SJeff Bonwick ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 2311b24ab676SJeff Bonwick 2312b24ab676SJeff Bonwick ddt_enter(ddt); 23133f9d6ad7SLin Ling freedde = dde = ddt_lookup(ddt, bp, B_TRUE); 2314b24ab676SJeff Bonwick ddp = ddt_phys_select(dde, bp); 2315b24ab676SJeff Bonwick ddt_phys_decref(ddp); 2316b24ab676SJeff Bonwick ddt_exit(ddt); 2317b24ab676SJeff Bonwick 2318b24ab676SJeff Bonwick return (ZIO_PIPELINE_CONTINUE); 2319b24ab676SJeff Bonwick } 2320b24ab676SJeff Bonwick 2321b24ab676SJeff Bonwick /* 2322b24ab676SJeff Bonwick * ========================================================================== 2323b24ab676SJeff Bonwick * Allocate and free blocks 2324b24ab676SJeff Bonwick * ========================================================================== 2325b24ab676SJeff Bonwick */ 2326e05725b1Sbonwick static int 2327fa9e4066Sahrens zio_dva_allocate(zio_t *zio) 2328fa9e4066Sahrens { 23298654d025Sperrin spa_t *spa = zio->io_spa; 2330b24ab676SJeff Bonwick metaslab_class_t *mc = spa_normal_class(spa); 2331fa9e4066Sahrens blkptr_t *bp = zio->io_bp; 2332fa9e4066Sahrens int error; 233309c9d376SGeorge Wilson int flags = 0; 2334fa9e4066Sahrens 2335f5383399SBill Moore if (zio->io_gang_leader == NULL) { 2336f5383399SBill Moore ASSERT(zio->io_child_type > ZIO_CHILD_GANG); 2337f5383399SBill Moore zio->io_gang_leader = zio; 2338f5383399SBill Moore } 2339f5383399SBill Moore 2340fa9e4066Sahrens ASSERT(BP_IS_HOLE(bp)); 2341fb09f5aaSMadhav Suresh ASSERT0(BP_GET_NDVAS(bp)); 2342b24ab676SJeff Bonwick ASSERT3U(zio->io_prop.zp_copies, >, 0); 2343b24ab676SJeff Bonwick ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa)); 2344fa9e4066Sahrens ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp)); 2345fa9e4066Sahrens 234609c9d376SGeorge Wilson /* 234709c9d376SGeorge Wilson * The dump device does not support gang blocks so allocation on 234809c9d376SGeorge Wilson * behalf of the dump device (i.e. ZIO_FLAG_NODATA) must avoid 234909c9d376SGeorge Wilson * the "fast" gang feature. 235009c9d376SGeorge Wilson */ 235109c9d376SGeorge Wilson flags |= (zio->io_flags & ZIO_FLAG_NODATA) ? METASLAB_GANG_AVOID : 0; 235209c9d376SGeorge Wilson flags |= (zio->io_flags & ZIO_FLAG_GANG_CHILD) ? 235309c9d376SGeorge Wilson METASLAB_GANG_CHILD : 0; 2354e14bb325SJeff Bonwick error = metaslab_alloc(spa, mc, zio->io_size, bp, 235509c9d376SGeorge Wilson zio->io_prop.zp_copies, zio->io_txg, NULL, flags); 2356fa9e4066Sahrens 2357e14bb325SJeff Bonwick if (error) { 235809c9d376SGeorge Wilson spa_dbgmsg(spa, "%s: metaslab allocation failure: zio %p, " 235909c9d376SGeorge Wilson "size %llu, error %d", spa_name(spa), zio, zio->io_size, 236009c9d376SGeorge Wilson error); 2361e14bb325SJeff Bonwick if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE) 2362e14bb325SJeff Bonwick return (zio_write_gang_block(zio)); 2363fa9e4066Sahrens zio->io_error = error; 2364fa9e4066Sahrens } 2365e05725b1Sbonwick 2366e05725b1Sbonwick return (ZIO_PIPELINE_CONTINUE); 2367fa9e4066Sahrens } 2368fa9e4066Sahrens 2369e05725b1Sbonwick static int 2370fa9e4066Sahrens zio_dva_free(zio_t *zio) 2371fa9e4066Sahrens { 2372e14bb325SJeff Bonwick metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE); 2373fa9e4066Sahrens 2374e05725b1Sbonwick return (ZIO_PIPELINE_CONTINUE); 2375fa9e4066Sahrens } 2376fa9e4066Sahrens 2377e05725b1Sbonwick static int 2378fa9e4066Sahrens zio_dva_claim(zio_t *zio) 2379fa9e4066Sahrens { 2380e14bb325SJeff Bonwick int error; 2381e14bb325SJeff Bonwick 2382e14bb325SJeff Bonwick error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg); 2383e14bb325SJeff Bonwick if (error) 2384e14bb325SJeff Bonwick zio->io_error = error; 2385fa9e4066Sahrens 2386e05725b1Sbonwick return (ZIO_PIPELINE_CONTINUE); 2387fa9e4066Sahrens } 2388fa9e4066Sahrens 2389e14bb325SJeff Bonwick /* 2390e14bb325SJeff Bonwick * Undo an allocation. This is used by zio_done() when an I/O fails 2391e14bb325SJeff Bonwick * and we want to give back the block we just allocated. 2392e14bb325SJeff Bonwick * This handles both normal blocks and gang blocks. 2393e14bb325SJeff Bonwick */ 2394e14bb325SJeff Bonwick static void 2395e14bb325SJeff Bonwick zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp) 2396e14bb325SJeff Bonwick { 2397e14bb325SJeff Bonwick ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp)); 2398b24ab676SJeff Bonwick ASSERT(zio->io_bp_override == NULL); 2399e14bb325SJeff Bonwick 2400e14bb325SJeff Bonwick if (!BP_IS_HOLE(bp)) 2401b24ab676SJeff Bonwick metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE); 2402e14bb325SJeff Bonwick 2403e14bb325SJeff Bonwick if (gn != NULL) { 2404e14bb325SJeff Bonwick for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) { 2405e14bb325SJeff Bonwick zio_dva_unallocate(zio, gn->gn_child[g], 2406e14bb325SJeff Bonwick &gn->gn_gbh->zg_blkptr[g]); 2407e14bb325SJeff Bonwick } 2408e14bb325SJeff Bonwick } 2409e14bb325SJeff Bonwick } 2410e14bb325SJeff Bonwick 2411e14bb325SJeff Bonwick /* 2412e14bb325SJeff Bonwick * Try to allocate an intent log block. Return 0 on success, errno on failure. 2413e14bb325SJeff Bonwick */ 2414e14bb325SJeff Bonwick int 2415b24ab676SJeff Bonwick zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, blkptr_t *old_bp, 2416b24ab676SJeff Bonwick uint64_t size, boolean_t use_slog) 2417e14bb325SJeff Bonwick { 2418e09fa4daSNeil Perrin int error = 1; 2419e14bb325SJeff Bonwick 2420b24ab676SJeff Bonwick ASSERT(txg > spa_syncing_txg(spa)); 2421b24ab676SJeff Bonwick 2422840345f6SGeorge Wilson /* 2423840345f6SGeorge Wilson * ZIL blocks are always contiguous (i.e. not gang blocks) so we 2424840345f6SGeorge Wilson * set the METASLAB_GANG_AVOID flag so that they don't "fast gang" 2425840345f6SGeorge Wilson * when allocating them. 2426840345f6SGeorge Wilson */ 2427840345f6SGeorge Wilson if (use_slog) { 2428b24ab676SJeff Bonwick error = metaslab_alloc(spa, spa_log_class(spa), size, 2429840345f6SGeorge Wilson new_bp, 1, txg, old_bp, 2430840345f6SGeorge Wilson METASLAB_HINTBP_AVOID | METASLAB_GANG_AVOID); 2431840345f6SGeorge Wilson } 2432e14bb325SJeff Bonwick 2433840345f6SGeorge Wilson if (error) { 2434b24ab676SJeff Bonwick error = metaslab_alloc(spa, spa_normal_class(spa), size, 2435840345f6SGeorge Wilson new_bp, 1, txg, old_bp, 243622e30981SGeorge Wilson METASLAB_HINTBP_AVOID); 2437840345f6SGeorge Wilson } 2438e14bb325SJeff Bonwick 2439e14bb325SJeff Bonwick if (error == 0) { 2440e14bb325SJeff Bonwick BP_SET_LSIZE(new_bp, size); 2441e14bb325SJeff Bonwick BP_SET_PSIZE(new_bp, size); 2442e14bb325SJeff Bonwick BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF); 24436e1f5caaSNeil Perrin BP_SET_CHECKSUM(new_bp, 24446e1f5caaSNeil Perrin spa_version(spa) >= SPA_VERSION_SLIM_ZIL 24456e1f5caaSNeil Perrin ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG); 2446e14bb325SJeff Bonwick BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG); 2447e14bb325SJeff Bonwick BP_SET_LEVEL(new_bp, 0); 2448b24ab676SJeff Bonwick BP_SET_DEDUP(new_bp, 0); 2449e14bb325SJeff Bonwick BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER); 2450e14bb325SJeff Bonwick } 2451e14bb325SJeff Bonwick 2452e14bb325SJeff Bonwick return (error); 2453e14bb325SJeff Bonwick } 2454e14bb325SJeff Bonwick 2455e14bb325SJeff Bonwick /* 2456b24ab676SJeff Bonwick * Free an intent log block. 2457e14bb325SJeff Bonwick */ 2458e14bb325SJeff Bonwick void 2459b24ab676SJeff Bonwick zio_free_zil(spa_t *spa, uint64_t txg, blkptr_t *bp) 2460e14bb325SJeff Bonwick { 2461b24ab676SJeff Bonwick ASSERT(BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG); 2462e14bb325SJeff Bonwick ASSERT(!BP_IS_GANG(bp)); 2463e14bb325SJeff Bonwick 2464b24ab676SJeff Bonwick zio_free(spa, txg, bp); 2465e14bb325SJeff Bonwick } 2466e14bb325SJeff Bonwick 2467fa9e4066Sahrens /* 2468fa9e4066Sahrens * ========================================================================== 2469fa9e4066Sahrens * Read and write to physical devices 2470fa9e4066Sahrens * ========================================================================== 2471fa9e4066Sahrens */ 2472*738f37bcSGeorge Wilson 2473*738f37bcSGeorge Wilson 2474*738f37bcSGeorge Wilson /* 2475*738f37bcSGeorge Wilson * Issue an I/O to the underlying vdev. Typically the issue pipeline 2476*738f37bcSGeorge Wilson * stops after this stage and will resume upon I/O completion. 2477*738f37bcSGeorge Wilson * However, there are instances where the vdev layer may need to 2478*738f37bcSGeorge Wilson * continue the pipeline when an I/O was not issued. Since the I/O 2479*738f37bcSGeorge Wilson * that was sent to the vdev layer might be different than the one 2480*738f37bcSGeorge Wilson * currently active in the pipeline (see vdev_queue_io()), we explicitly 2481*738f37bcSGeorge Wilson * force the underlying vdev layers to call either zio_execute() or 2482*738f37bcSGeorge Wilson * zio_interrupt() to ensure that the pipeline continues with the correct I/O. 2483*738f37bcSGeorge Wilson */ 2484e05725b1Sbonwick static int 248544cd46caSbillm zio_vdev_io_start(zio_t *zio) 2486fa9e4066Sahrens { 2487fa9e4066Sahrens vdev_t *vd = zio->io_vd; 248844cd46caSbillm uint64_t align; 24890a4e9518Sgw spa_t *spa = zio->io_spa; 24900a4e9518Sgw 2491e14bb325SJeff Bonwick ASSERT(zio->io_error == 0); 2492e14bb325SJeff Bonwick ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0); 2493fa9e4066Sahrens 2494e14bb325SJeff Bonwick if (vd == NULL) { 2495e14bb325SJeff Bonwick if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER)) 2496e14bb325SJeff Bonwick spa_config_enter(spa, SCL_ZIO, zio, RW_READER); 2497fa9e4066Sahrens 2498e14bb325SJeff Bonwick /* 2499e14bb325SJeff Bonwick * The mirror_ops handle multiple DVAs in a single BP. 2500e14bb325SJeff Bonwick */ 2501*738f37bcSGeorge Wilson vdev_mirror_ops.vdev_op_io_start(zio); 2502*738f37bcSGeorge Wilson return (ZIO_PIPELINE_STOP); 2503fa9e4066Sahrens } 2504fa9e4066Sahrens 250544ecc532SGeorge Wilson /* 250644ecc532SGeorge Wilson * We keep track of time-sensitive I/Os so that the scan thread 250744ecc532SGeorge Wilson * can quickly react to certain workloads. In particular, we care 250844ecc532SGeorge Wilson * about non-scrubbing, top-level reads and writes with the following 250944ecc532SGeorge Wilson * characteristics: 2510*738f37bcSGeorge Wilson * - synchronous writes of user data to non-slog devices 251144ecc532SGeorge Wilson * - any reads of user data 251244ecc532SGeorge Wilson * When these conditions are met, adjust the timestamp of spa_last_io 251344ecc532SGeorge Wilson * which allows the scan thread to adjust its workload accordingly. 251444ecc532SGeorge Wilson */ 251544ecc532SGeorge Wilson if (!(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && zio->io_bp != NULL && 251644ecc532SGeorge Wilson vd == vd->vdev_top && !vd->vdev_islog && 251744ecc532SGeorge Wilson zio->io_bookmark.zb_objset != DMU_META_OBJSET && 251844ecc532SGeorge Wilson zio->io_txg != spa_syncing_txg(spa)) { 251944ecc532SGeorge Wilson uint64_t old = spa->spa_last_io; 252044ecc532SGeorge Wilson uint64_t new = ddi_get_lbolt64(); 252144ecc532SGeorge Wilson if (old != new) 252244ecc532SGeorge Wilson (void) atomic_cas_64(&spa->spa_last_io, old, new); 252344ecc532SGeorge Wilson } 252444ecc532SGeorge Wilson 2525e14bb325SJeff Bonwick align = 1ULL << vd->vdev_top->vdev_ashift; 2526e14bb325SJeff Bonwick 25272a104a52SAlex Reece if (!(zio->io_flags & ZIO_FLAG_PHYSICAL) && 25282a104a52SAlex Reece P2PHASE(zio->io_size, align) != 0) { 25292a104a52SAlex Reece /* Transform logical writes to be a full physical block size. */ 2530ecc2d604Sbonwick uint64_t asize = P2ROUNDUP(zio->io_size, align); 2531ecc2d604Sbonwick char *abuf = zio_buf_alloc(asize); 2532e14bb325SJeff Bonwick ASSERT(vd == vd->vdev_top); 2533ecc2d604Sbonwick if (zio->io_type == ZIO_TYPE_WRITE) { 2534ecc2d604Sbonwick bcopy(zio->io_data, abuf, zio->io_size); 2535ecc2d604Sbonwick bzero(abuf + zio->io_size, asize - zio->io_size); 2536ecc2d604Sbonwick } 2537e14bb325SJeff Bonwick zio_push_transform(zio, abuf, asize, asize, zio_subblock); 2538ecc2d604Sbonwick } 2539ecc2d604Sbonwick 25402a104a52SAlex Reece /* 25412a104a52SAlex Reece * If this is not a physical io, make sure that it is properly aligned 25422a104a52SAlex Reece * before proceeding. 25432a104a52SAlex Reece */ 25442a104a52SAlex Reece if (!(zio->io_flags & ZIO_FLAG_PHYSICAL)) { 25452a104a52SAlex Reece ASSERT0(P2PHASE(zio->io_offset, align)); 25462a104a52SAlex Reece ASSERT0(P2PHASE(zio->io_size, align)); 25472a104a52SAlex Reece } else { 25482a104a52SAlex Reece /* 25492a104a52SAlex Reece * For physical writes, we allow 512b aligned writes and assume 25502a104a52SAlex Reece * the device will perform a read-modify-write as necessary. 25512a104a52SAlex Reece */ 25522a104a52SAlex Reece ASSERT0(P2PHASE(zio->io_offset, SPA_MINBLOCKSIZE)); 25532a104a52SAlex Reece ASSERT0(P2PHASE(zio->io_size, SPA_MINBLOCKSIZE)); 25542a104a52SAlex Reece } 25552a104a52SAlex Reece 2556f9af39baSGeorge Wilson VERIFY(zio->io_type != ZIO_TYPE_WRITE || spa_writeable(spa)); 25578ad4d6ddSJeff Bonwick 25588ad4d6ddSJeff Bonwick /* 25598ad4d6ddSJeff Bonwick * If this is a repair I/O, and there's no self-healing involved -- 25608ad4d6ddSJeff Bonwick * that is, we're just resilvering what we expect to resilver -- 25618ad4d6ddSJeff Bonwick * then don't do the I/O unless zio's txg is actually in vd's DTL. 25628ad4d6ddSJeff Bonwick * This prevents spurious resilvering with nested replication. 25638ad4d6ddSJeff Bonwick * For example, given a mirror of mirrors, (A+B)+(C+D), if only 25648ad4d6ddSJeff Bonwick * A is out of date, we'll read from C+D, then use the data to 25658ad4d6ddSJeff Bonwick * resilver A+B -- but we don't actually want to resilver B, just A. 25668ad4d6ddSJeff Bonwick * The top-level mirror has no way to know this, so instead we just 25678ad4d6ddSJeff Bonwick * discard unnecessary repairs as we work our way down the vdev tree. 25688ad4d6ddSJeff Bonwick * The same logic applies to any form of nested replication: 25698ad4d6ddSJeff Bonwick * ditto + mirror, RAID-Z + replacing, etc. This covers them all. 25708ad4d6ddSJeff Bonwick */ 25718ad4d6ddSJeff Bonwick if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) && 25728ad4d6ddSJeff Bonwick !(zio->io_flags & ZIO_FLAG_SELF_HEAL) && 25738ad4d6ddSJeff Bonwick zio->io_txg != 0 && /* not a delegated i/o */ 25748ad4d6ddSJeff Bonwick !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) { 25758ad4d6ddSJeff Bonwick ASSERT(zio->io_type == ZIO_TYPE_WRITE); 25768ad4d6ddSJeff Bonwick zio_vdev_io_bypass(zio); 25778ad4d6ddSJeff Bonwick return (ZIO_PIPELINE_CONTINUE); 25788ad4d6ddSJeff Bonwick } 2579fa9e4066Sahrens 2580e14bb325SJeff Bonwick if (vd->vdev_ops->vdev_op_leaf && 2581e14bb325SJeff Bonwick (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE)) { 2582e14bb325SJeff Bonwick 258343466aaeSMax Grossman if (zio->io_type == ZIO_TYPE_READ && vdev_cache_read(zio)) 2584a3f829aeSBill Moore return (ZIO_PIPELINE_CONTINUE); 2585e14bb325SJeff Bonwick 2586e14bb325SJeff Bonwick if ((zio = vdev_queue_io(zio)) == NULL) 2587e14bb325SJeff Bonwick return (ZIO_PIPELINE_STOP); 2588e14bb325SJeff Bonwick 2589e14bb325SJeff Bonwick if (!vdev_accessible(vd, zio)) { 2590be6fd75aSMatthew Ahrens zio->io_error = SET_ERROR(ENXIO); 2591e14bb325SJeff Bonwick zio_interrupt(zio); 2592e14bb325SJeff Bonwick return (ZIO_PIPELINE_STOP); 2593e14bb325SJeff Bonwick } 2594e14bb325SJeff Bonwick } 2595e14bb325SJeff Bonwick 2596*738f37bcSGeorge Wilson vd->vdev_ops->vdev_op_io_start(zio); 2597*738f37bcSGeorge Wilson return (ZIO_PIPELINE_STOP); 2598fa9e4066Sahrens } 2599fa9e4066Sahrens 2600e05725b1Sbonwick static int 2601fa9e4066Sahrens zio_vdev_io_done(zio_t *zio) 2602fa9e4066Sahrens { 2603e14bb325SJeff Bonwick vdev_t *vd = zio->io_vd; 2604e14bb325SJeff Bonwick vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops; 2605e14bb325SJeff Bonwick boolean_t unexpected_error = B_FALSE; 2606e05725b1Sbonwick 2607e14bb325SJeff Bonwick if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE)) 2608e14bb325SJeff Bonwick return (ZIO_PIPELINE_STOP); 2609fa9e4066Sahrens 2610e14bb325SJeff Bonwick ASSERT(zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE); 2611e14bb325SJeff Bonwick 2612e14bb325SJeff Bonwick if (vd != NULL && vd->vdev_ops->vdev_op_leaf) { 2613e14bb325SJeff Bonwick 2614e14bb325SJeff Bonwick vdev_queue_io_done(zio); 2615fa9e4066Sahrens 2616e14bb325SJeff Bonwick if (zio->io_type == ZIO_TYPE_WRITE) 2617e14bb325SJeff Bonwick vdev_cache_write(zio); 2618e14bb325SJeff Bonwick 2619e14bb325SJeff Bonwick if (zio_injection_enabled && zio->io_error == 0) 26208956713aSEric Schrock zio->io_error = zio_handle_device_injection(vd, 26218956713aSEric Schrock zio, EIO); 2622e14bb325SJeff Bonwick 2623e14bb325SJeff Bonwick if (zio_injection_enabled && zio->io_error == 0) 2624e14bb325SJeff Bonwick zio->io_error = zio_handle_label_injection(zio, EIO); 2625e14bb325SJeff Bonwick 2626e14bb325SJeff Bonwick if (zio->io_error) { 2627e14bb325SJeff Bonwick if (!vdev_accessible(vd, zio)) { 2628be6fd75aSMatthew Ahrens zio->io_error = SET_ERROR(ENXIO); 2629e14bb325SJeff Bonwick } else { 2630e14bb325SJeff Bonwick unexpected_error = B_TRUE; 2631e14bb325SJeff Bonwick } 2632e14bb325SJeff Bonwick } 263351ece835Seschrock } 2634fa9e4066Sahrens 2635e14bb325SJeff Bonwick ops->vdev_op_io_done(zio); 2636e14bb325SJeff Bonwick 2637e14bb325SJeff Bonwick if (unexpected_error) 2638a3f829aeSBill Moore VERIFY(vdev_probe(vd, zio) == NULL); 2639e14bb325SJeff Bonwick 2640e14bb325SJeff Bonwick return (ZIO_PIPELINE_CONTINUE); 2641fa9e4066Sahrens } 2642fa9e4066Sahrens 264322fe2c88SJonathan Adams /* 264422fe2c88SJonathan Adams * For non-raidz ZIOs, we can just copy aside the bad data read from the 264522fe2c88SJonathan Adams * disk, and use that to finish the checksum ereport later. 264622fe2c88SJonathan Adams */ 264722fe2c88SJonathan Adams static void 264822fe2c88SJonathan Adams zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr, 264922fe2c88SJonathan Adams const void *good_buf) 265022fe2c88SJonathan Adams { 265122fe2c88SJonathan Adams /* no processing needed */ 265222fe2c88SJonathan Adams zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE); 265322fe2c88SJonathan Adams } 265422fe2c88SJonathan Adams 265522fe2c88SJonathan Adams /*ARGSUSED*/ 265622fe2c88SJonathan Adams void 265722fe2c88SJonathan Adams zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored) 265822fe2c88SJonathan Adams { 265922fe2c88SJonathan Adams void *buf = zio_buf_alloc(zio->io_size); 266022fe2c88SJonathan Adams 266122fe2c88SJonathan Adams bcopy(zio->io_data, buf, zio->io_size); 266222fe2c88SJonathan Adams 266322fe2c88SJonathan Adams zcr->zcr_cbinfo = zio->io_size; 266422fe2c88SJonathan Adams zcr->zcr_cbdata = buf; 266522fe2c88SJonathan Adams zcr->zcr_finish = zio_vsd_default_cksum_finish; 266622fe2c88SJonathan Adams zcr->zcr_free = zio_buf_free; 266722fe2c88SJonathan Adams } 266822fe2c88SJonathan Adams 2669e05725b1Sbonwick static int 2670fa9e4066Sahrens zio_vdev_io_assess(zio_t *zio) 2671fa9e4066Sahrens { 2672fa9e4066Sahrens vdev_t *vd = zio->io_vd; 2673e14bb325SJeff Bonwick 2674e14bb325SJeff Bonwick if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE)) 2675e14bb325SJeff Bonwick return (ZIO_PIPELINE_STOP); 2676e14bb325SJeff Bonwick 2677e14bb325SJeff Bonwick if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER)) 2678e14bb325SJeff Bonwick spa_config_exit(zio->io_spa, SCL_ZIO, zio); 2679e14bb325SJeff Bonwick 2680e14bb325SJeff Bonwick if (zio->io_vsd != NULL) { 268122fe2c88SJonathan Adams zio->io_vsd_ops->vsd_free(zio); 2682e14bb325SJeff Bonwick zio->io_vsd = NULL; 2683ecc2d604Sbonwick } 2684ecc2d604Sbonwick 2685e14bb325SJeff Bonwick if (zio_injection_enabled && zio->io_error == 0) 2686ea8dc4b6Seschrock zio->io_error = zio_handle_fault_injection(zio, EIO); 2687ea8dc4b6Seschrock 2688fa9e4066Sahrens /* 2689fa9e4066Sahrens * If the I/O failed, determine whether we should attempt to retry it. 269035a5a358SJonathan Adams * 269135a5a358SJonathan Adams * On retry, we cut in line in the issue queue, since we don't want 269235a5a358SJonathan Adams * compression/checksumming/etc. work to prevent our (cheap) IO reissue. 2693fa9e4066Sahrens */ 2694e14bb325SJeff Bonwick if (zio->io_error && vd == NULL && 2695e14bb325SJeff Bonwick !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) { 2696e14bb325SJeff Bonwick ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */ 2697e14bb325SJeff Bonwick ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */ 2698fa9e4066Sahrens zio->io_error = 0; 2699e14bb325SJeff Bonwick zio->io_flags |= ZIO_FLAG_IO_RETRY | 2700e14bb325SJeff Bonwick ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE; 2701b24ab676SJeff Bonwick zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1; 270235a5a358SJonathan Adams zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, 270335a5a358SJonathan Adams zio_requeue_io_start_cut_in_line); 2704e14bb325SJeff Bonwick return (ZIO_PIPELINE_STOP); 2705ea8dc4b6Seschrock } 2706fa9e4066Sahrens 2707e14bb325SJeff Bonwick /* 2708e14bb325SJeff Bonwick * If we got an error on a leaf device, convert it to ENXIO 2709e14bb325SJeff Bonwick * if the device is not accessible at all. 2710e14bb325SJeff Bonwick */ 2711e14bb325SJeff Bonwick if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf && 2712e14bb325SJeff Bonwick !vdev_accessible(vd, zio)) 2713be6fd75aSMatthew Ahrens zio->io_error = SET_ERROR(ENXIO); 2714e14bb325SJeff Bonwick 2715e14bb325SJeff Bonwick /* 2716e14bb325SJeff Bonwick * If we can't write to an interior vdev (mirror or RAID-Z), 2717e14bb325SJeff Bonwick * set vdev_cant_write so that we stop trying to allocate from it. 2718e14bb325SJeff Bonwick */ 2719e14bb325SJeff Bonwick if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE && 27203b2aab18SMatthew Ahrens vd != NULL && !vd->vdev_ops->vdev_op_leaf) { 2721e14bb325SJeff Bonwick vd->vdev_cant_write = B_TRUE; 27223b2aab18SMatthew Ahrens } 2723e14bb325SJeff Bonwick 2724e14bb325SJeff Bonwick if (zio->io_error) 2725e14bb325SJeff Bonwick zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 2726e14bb325SJeff Bonwick 272769962b56SMatthew Ahrens if (vd != NULL && vd->vdev_ops->vdev_op_leaf && 272869962b56SMatthew Ahrens zio->io_physdone != NULL) { 272969962b56SMatthew Ahrens ASSERT(!(zio->io_flags & ZIO_FLAG_DELEGATED)); 273069962b56SMatthew Ahrens ASSERT(zio->io_child_type == ZIO_CHILD_VDEV); 273169962b56SMatthew Ahrens zio->io_physdone(zio->io_logical); 273269962b56SMatthew Ahrens } 273369962b56SMatthew Ahrens 2734e05725b1Sbonwick return (ZIO_PIPELINE_CONTINUE); 2735fa9e4066Sahrens } 2736fa9e4066Sahrens 2737fa9e4066Sahrens void 2738fa9e4066Sahrens zio_vdev_io_reissue(zio_t *zio) 2739fa9e4066Sahrens { 2740fa9e4066Sahrens ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START); 2741fa9e4066Sahrens ASSERT(zio->io_error == 0); 2742fa9e4066Sahrens 2743b24ab676SJeff Bonwick zio->io_stage >>= 1; 2744fa9e4066Sahrens } 2745fa9e4066Sahrens 2746fa9e4066Sahrens void 2747fa9e4066Sahrens zio_vdev_io_redone(zio_t *zio) 2748fa9e4066Sahrens { 2749fa9e4066Sahrens ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE); 2750fa9e4066Sahrens 2751b24ab676SJeff Bonwick zio->io_stage >>= 1; 2752fa9e4066Sahrens } 2753fa9e4066Sahrens 2754fa9e4066Sahrens void 2755fa9e4066Sahrens zio_vdev_io_bypass(zio_t *zio) 2756fa9e4066Sahrens { 2757fa9e4066Sahrens ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START); 2758fa9e4066Sahrens ASSERT(zio->io_error == 0); 2759fa9e4066Sahrens 2760fa9e4066Sahrens zio->io_flags |= ZIO_FLAG_IO_BYPASS; 2761b24ab676SJeff Bonwick zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1; 2762fa9e4066Sahrens } 2763fa9e4066Sahrens 2764fa9e4066Sahrens /* 2765fa9e4066Sahrens * ========================================================================== 2766fa9e4066Sahrens * Generate and verify checksums 2767fa9e4066Sahrens * ========================================================================== 2768fa9e4066Sahrens */ 2769e05725b1Sbonwick static int 2770fa9e4066Sahrens zio_checksum_generate(zio_t *zio) 2771fa9e4066Sahrens { 2772fa9e4066Sahrens blkptr_t *bp = zio->io_bp; 2773e14bb325SJeff Bonwick enum zio_checksum checksum; 2774fa9e4066Sahrens 2775e14bb325SJeff Bonwick if (bp == NULL) { 2776e14bb325SJeff Bonwick /* 2777e14bb325SJeff Bonwick * This is zio_write_phys(). 2778e14bb325SJeff Bonwick * We're either generating a label checksum, or none at all. 2779e14bb325SJeff Bonwick */ 2780e14bb325SJeff Bonwick checksum = zio->io_prop.zp_checksum; 2781e14bb325SJeff Bonwick 2782e14bb325SJeff Bonwick if (checksum == ZIO_CHECKSUM_OFF) 2783e14bb325SJeff Bonwick return (ZIO_PIPELINE_CONTINUE); 2784fa9e4066Sahrens 2785e14bb325SJeff Bonwick ASSERT(checksum == ZIO_CHECKSUM_LABEL); 2786e14bb325SJeff Bonwick } else { 2787e14bb325SJeff Bonwick if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) { 2788e14bb325SJeff Bonwick ASSERT(!IO_IS_ALLOCATING(zio)); 2789e14bb325SJeff Bonwick checksum = ZIO_CHECKSUM_GANG_HEADER; 2790e14bb325SJeff Bonwick } else { 2791e14bb325SJeff Bonwick checksum = BP_GET_CHECKSUM(bp); 2792e14bb325SJeff Bonwick } 2793e14bb325SJeff Bonwick } 2794fa9e4066Sahrens 2795e14bb325SJeff Bonwick zio_checksum_compute(zio, checksum, zio->io_data, zio->io_size); 2796fa9e4066Sahrens 2797e05725b1Sbonwick return (ZIO_PIPELINE_CONTINUE); 2798fa9e4066Sahrens } 2799fa9e4066Sahrens 2800e05725b1Sbonwick static int 2801e14bb325SJeff Bonwick zio_checksum_verify(zio_t *zio) 2802fa9e4066Sahrens { 280322fe2c88SJonathan Adams zio_bad_cksum_t info; 2804e14bb325SJeff Bonwick blkptr_t *bp = zio->io_bp; 2805e14bb325SJeff Bonwick int error; 2806fa9e4066Sahrens 2807b24ab676SJeff Bonwick ASSERT(zio->io_vd != NULL); 2808b24ab676SJeff Bonwick 2809e14bb325SJeff Bonwick if (bp == NULL) { 2810e14bb325SJeff Bonwick /* 2811e14bb325SJeff Bonwick * This is zio_read_phys(). 2812e14bb325SJeff Bonwick * We're either verifying a label checksum, or nothing at all. 2813e14bb325SJeff Bonwick */ 2814e14bb325SJeff Bonwick if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF) 2815e14bb325SJeff Bonwick return (ZIO_PIPELINE_CONTINUE); 2816fa9e4066Sahrens 2817e14bb325SJeff Bonwick ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL); 2818e14bb325SJeff Bonwick } 2819fa9e4066Sahrens 282022fe2c88SJonathan Adams if ((error = zio_checksum_error(zio, &info)) != 0) { 2821e14bb325SJeff Bonwick zio->io_error = error; 2822e14bb325SJeff Bonwick if (!(zio->io_flags & ZIO_FLAG_SPECULATIVE)) { 282322fe2c88SJonathan Adams zfs_ereport_start_checksum(zio->io_spa, 282422fe2c88SJonathan Adams zio->io_vd, zio, zio->io_offset, 282522fe2c88SJonathan Adams zio->io_size, NULL, &info); 2826e14bb325SJeff Bonwick } 2827fa9e4066Sahrens } 2828fa9e4066Sahrens 2829e05725b1Sbonwick return (ZIO_PIPELINE_CONTINUE); 2830fa9e4066Sahrens } 2831fa9e4066Sahrens 2832fa9e4066Sahrens /* 2833fa9e4066Sahrens * Called by RAID-Z to ensure we don't compute the checksum twice. 2834fa9e4066Sahrens */ 2835fa9e4066Sahrens void 2836fa9e4066Sahrens zio_checksum_verified(zio_t *zio) 2837fa9e4066Sahrens { 2838b24ab676SJeff Bonwick zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY; 2839fa9e4066Sahrens } 2840fa9e4066Sahrens 2841fa9e4066Sahrens /* 2842e14bb325SJeff Bonwick * ========================================================================== 2843e14bb325SJeff Bonwick * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other. 28445d7b4d43SMatthew Ahrens * An error of 0 indicates success. ENXIO indicates whole-device failure, 2845e14bb325SJeff Bonwick * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO 2846e14bb325SJeff Bonwick * indicate errors that are specific to one I/O, and most likely permanent. 2847e14bb325SJeff Bonwick * Any other error is presumed to be worse because we weren't expecting it. 2848e14bb325SJeff Bonwick * ========================================================================== 2849fa9e4066Sahrens */ 2850e14bb325SJeff Bonwick int 2851e14bb325SJeff Bonwick zio_worst_error(int e1, int e2) 2852fa9e4066Sahrens { 2853e14bb325SJeff Bonwick static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO }; 2854e14bb325SJeff Bonwick int r1, r2; 2855e14bb325SJeff Bonwick 2856e14bb325SJeff Bonwick for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++) 2857e14bb325SJeff Bonwick if (e1 == zio_error_rank[r1]) 2858e14bb325SJeff Bonwick break; 2859e14bb325SJeff Bonwick 2860e14bb325SJeff Bonwick for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++) 2861e14bb325SJeff Bonwick if (e2 == zio_error_rank[r2]) 2862e14bb325SJeff Bonwick break; 286344cd46caSbillm 2864e14bb325SJeff Bonwick return (r1 > r2 ? e1 : e2); 2865fa9e4066Sahrens } 2866fa9e4066Sahrens 2867fa9e4066Sahrens /* 2868fa9e4066Sahrens * ========================================================================== 2869e14bb325SJeff Bonwick * I/O completion 2870fa9e4066Sahrens * ========================================================================== 2871fa9e4066Sahrens */ 2872e14bb325SJeff Bonwick static int 2873e14bb325SJeff Bonwick zio_ready(zio_t *zio) 2874fa9e4066Sahrens { 2875e14bb325SJeff Bonwick blkptr_t *bp = zio->io_bp; 2876a3f829aeSBill Moore zio_t *pio, *pio_next; 2877fa9e4066Sahrens 2878b24ab676SJeff Bonwick if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) || 2879b24ab676SJeff Bonwick zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_READY)) 2880f5383399SBill Moore return (ZIO_PIPELINE_STOP); 2881fa9e4066Sahrens 2882f5383399SBill Moore if (zio->io_ready) { 2883e14bb325SJeff Bonwick ASSERT(IO_IS_ALLOCATING(zio)); 288480901aeaSGeorge Wilson ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp) || 288580901aeaSGeorge Wilson (zio->io_flags & ZIO_FLAG_NOPWRITE)); 2886e14bb325SJeff Bonwick ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0); 2887fa9e4066Sahrens 2888e14bb325SJeff Bonwick zio->io_ready(zio); 2889e14bb325SJeff Bonwick } 2890fa9e4066Sahrens 2891e14bb325SJeff Bonwick if (bp != NULL && bp != &zio->io_bp_copy) 2892e14bb325SJeff Bonwick zio->io_bp_copy = *bp; 2893fa9e4066Sahrens 2894e14bb325SJeff Bonwick if (zio->io_error) 2895e14bb325SJeff Bonwick zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 2896fa9e4066Sahrens 2897a3f829aeSBill Moore mutex_enter(&zio->io_lock); 2898a3f829aeSBill Moore zio->io_state[ZIO_WAIT_READY] = 1; 2899a3f829aeSBill Moore pio = zio_walk_parents(zio); 2900a3f829aeSBill Moore mutex_exit(&zio->io_lock); 2901a3f829aeSBill Moore 2902a3f829aeSBill Moore /* 2903a3f829aeSBill Moore * As we notify zio's parents, new parents could be added. 2904a3f829aeSBill Moore * New parents go to the head of zio's io_parent_list, however, 2905a3f829aeSBill Moore * so we will (correctly) not notify them. The remainder of zio's 2906a3f829aeSBill Moore * io_parent_list, from 'pio_next' onward, cannot change because 2907a3f829aeSBill Moore * all parents must wait for us to be done before they can be done. 2908a3f829aeSBill Moore */ 2909a3f829aeSBill Moore for (; pio != NULL; pio = pio_next) { 2910a3f829aeSBill Moore pio_next = zio_walk_parents(zio); 2911e14bb325SJeff Bonwick zio_notify_parent(pio, zio, ZIO_WAIT_READY); 2912a3f829aeSBill Moore } 2913fa9e4066Sahrens 2914b24ab676SJeff Bonwick if (zio->io_flags & ZIO_FLAG_NODATA) { 2915b24ab676SJeff Bonwick if (BP_IS_GANG(bp)) { 2916b24ab676SJeff Bonwick zio->io_flags &= ~ZIO_FLAG_NODATA; 2917b24ab676SJeff Bonwick } else { 2918b24ab676SJeff Bonwick ASSERT((uintptr_t)zio->io_data < SPA_MAXBLOCKSIZE); 2919b24ab676SJeff Bonwick zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES; 2920b24ab676SJeff Bonwick } 2921b24ab676SJeff Bonwick } 2922b24ab676SJeff Bonwick 2923a33cae98STim Haley if (zio_injection_enabled && 2924a33cae98STim Haley zio->io_spa->spa_syncing_txg == zio->io_txg) 2925a33cae98STim Haley zio_handle_ignored_writes(zio); 2926a33cae98STim Haley 2927e14bb325SJeff Bonwick return (ZIO_PIPELINE_CONTINUE); 2928fa9e4066Sahrens } 2929fa9e4066Sahrens 2930e14bb325SJeff Bonwick static int 2931e14bb325SJeff Bonwick zio_done(zio_t *zio) 2932d63d470bSgw { 2933e14bb325SJeff Bonwick spa_t *spa = zio->io_spa; 2934e14bb325SJeff Bonwick zio_t *lio = zio->io_logical; 2935e14bb325SJeff Bonwick blkptr_t *bp = zio->io_bp; 2936e14bb325SJeff Bonwick vdev_t *vd = zio->io_vd; 2937e14bb325SJeff Bonwick uint64_t psize = zio->io_size; 2938a3f829aeSBill Moore zio_t *pio, *pio_next; 2939d63d470bSgw 2940e14bb325SJeff Bonwick /* 2941f5383399SBill Moore * If our children haven't all completed, 2942e14bb325SJeff Bonwick * wait for them and then repeat this pipeline stage. 2943e14bb325SJeff Bonwick */ 2944e14bb325SJeff Bonwick if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE) || 2945e14bb325SJeff Bonwick zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE) || 2946b24ab676SJeff Bonwick zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE) || 2947e14bb325SJeff Bonwick zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_DONE)) 2948e14bb325SJeff Bonwick return (ZIO_PIPELINE_STOP); 2949d63d470bSgw 2950e14bb325SJeff Bonwick for (int c = 0; c < ZIO_CHILD_TYPES; c++) 2951e14bb325SJeff Bonwick for (int w = 0; w < ZIO_WAIT_TYPES; w++) 2952e14bb325SJeff Bonwick ASSERT(zio->io_children[c][w] == 0); 2953e14bb325SJeff Bonwick 29545d7b4d43SMatthew Ahrens if (bp != NULL && !BP_IS_EMBEDDED(bp)) { 2955e14bb325SJeff Bonwick ASSERT(bp->blk_pad[0] == 0); 2956e14bb325SJeff Bonwick ASSERT(bp->blk_pad[1] == 0); 2957e14bb325SJeff Bonwick ASSERT(bcmp(bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 || 2958a3f829aeSBill Moore (bp == zio_unique_parent(zio)->io_bp)); 2959e14bb325SJeff Bonwick if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(bp) && 2960b24ab676SJeff Bonwick zio->io_bp_override == NULL && 2961e14bb325SJeff Bonwick !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) { 2962e14bb325SJeff Bonwick ASSERT(!BP_SHOULD_BYTESWAP(bp)); 2963b24ab676SJeff Bonwick ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(bp)); 2964e14bb325SJeff Bonwick ASSERT(BP_COUNT_GANG(bp) == 0 || 2965e14bb325SJeff Bonwick (BP_COUNT_GANG(bp) == BP_GET_NDVAS(bp))); 2966e14bb325SJeff Bonwick } 296780901aeaSGeorge Wilson if (zio->io_flags & ZIO_FLAG_NOPWRITE) 296880901aeaSGeorge Wilson VERIFY(BP_EQUAL(bp, &zio->io_bp_orig)); 2969e14bb325SJeff Bonwick } 2970fa9e4066Sahrens 2971e14bb325SJeff Bonwick /* 2972b24ab676SJeff Bonwick * If there were child vdev/gang/ddt errors, they apply to us now. 2973e14bb325SJeff Bonwick */ 2974e14bb325SJeff Bonwick zio_inherit_child_errors(zio, ZIO_CHILD_VDEV); 2975e14bb325SJeff Bonwick zio_inherit_child_errors(zio, ZIO_CHILD_GANG); 2976b24ab676SJeff Bonwick zio_inherit_child_errors(zio, ZIO_CHILD_DDT); 2977b24ab676SJeff Bonwick 2978b24ab676SJeff Bonwick /* 2979b24ab676SJeff Bonwick * If the I/O on the transformed data was successful, generate any 2980b24ab676SJeff Bonwick * checksum reports now while we still have the transformed data. 2981b24ab676SJeff Bonwick */ 2982b24ab676SJeff Bonwick if (zio->io_error == 0) { 2983b24ab676SJeff Bonwick while (zio->io_cksum_report != NULL) { 2984b24ab676SJeff Bonwick zio_cksum_report_t *zcr = zio->io_cksum_report; 2985b24ab676SJeff Bonwick uint64_t align = zcr->zcr_align; 2986b24ab676SJeff Bonwick uint64_t asize = P2ROUNDUP(psize, align); 2987b24ab676SJeff Bonwick char *abuf = zio->io_data; 2988b24ab676SJeff Bonwick 2989b24ab676SJeff Bonwick if (asize != psize) { 2990b24ab676SJeff Bonwick abuf = zio_buf_alloc(asize); 2991b24ab676SJeff Bonwick bcopy(zio->io_data, abuf, psize); 2992b24ab676SJeff Bonwick bzero(abuf + psize, asize - psize); 2993b24ab676SJeff Bonwick } 2994b24ab676SJeff Bonwick 2995b24ab676SJeff Bonwick zio->io_cksum_report = zcr->zcr_next; 2996b24ab676SJeff Bonwick zcr->zcr_next = NULL; 2997b24ab676SJeff Bonwick zcr->zcr_finish(zcr, abuf); 2998b24ab676SJeff Bonwick zfs_ereport_free_checksum(zcr); 2999b24ab676SJeff Bonwick 3000b24ab676SJeff Bonwick if (asize != psize) 3001b24ab676SJeff Bonwick zio_buf_free(abuf, asize); 3002b24ab676SJeff Bonwick } 3003b24ab676SJeff Bonwick } 3004e14bb325SJeff Bonwick 3005e14bb325SJeff Bonwick zio_pop_transforms(zio); /* note: may set zio->io_error */ 3006e14bb325SJeff Bonwick 3007e14bb325SJeff Bonwick vdev_stat_update(zio, psize); 3008e14bb325SJeff Bonwick 3009e14bb325SJeff Bonwick if (zio->io_error) { 3010e14bb325SJeff Bonwick /* 3011e14bb325SJeff Bonwick * If this I/O is attached to a particular vdev, 3012e14bb325SJeff Bonwick * generate an error message describing the I/O failure 3013e14bb325SJeff Bonwick * at the block level. We ignore these errors if the 3014e14bb325SJeff Bonwick * device is currently unavailable. 3015e14bb325SJeff Bonwick */ 3016e14bb325SJeff Bonwick if (zio->io_error != ECKSUM && vd != NULL && !vdev_is_dead(vd)) 3017e14bb325SJeff Bonwick zfs_ereport_post(FM_EREPORT_ZFS_IO, spa, vd, zio, 0, 0); 3018e14bb325SJeff Bonwick 30198f18d1faSGeorge Wilson if ((zio->io_error == EIO || !(zio->io_flags & 30208f18d1faSGeorge Wilson (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) && 30218f18d1faSGeorge Wilson zio == lio) { 3022e14bb325SJeff Bonwick /* 3023e14bb325SJeff Bonwick * For logical I/O requests, tell the SPA to log the 3024e14bb325SJeff Bonwick * error and generate a logical data ereport. 3025e14bb325SJeff Bonwick */ 3026e14bb325SJeff Bonwick spa_log_error(spa, zio); 3027e14bb325SJeff Bonwick zfs_ereport_post(FM_EREPORT_ZFS_DATA, spa, NULL, zio, 3028e14bb325SJeff Bonwick 0, 0); 3029e14bb325SJeff Bonwick } 3030e14bb325SJeff Bonwick } 3031fa9e4066Sahrens 3032e14bb325SJeff Bonwick if (zio->io_error && zio == lio) { 3033e14bb325SJeff Bonwick /* 3034e14bb325SJeff Bonwick * Determine whether zio should be reexecuted. This will 3035e14bb325SJeff Bonwick * propagate all the way to the root via zio_notify_parent(). 3036e14bb325SJeff Bonwick */ 3037e14bb325SJeff Bonwick ASSERT(vd == NULL && bp != NULL); 3038b24ab676SJeff Bonwick ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 3039e14bb325SJeff Bonwick 3040b24ab676SJeff Bonwick if (IO_IS_ALLOCATING(zio) && 3041b24ab676SJeff Bonwick !(zio->io_flags & ZIO_FLAG_CANFAIL)) { 3042e14bb325SJeff Bonwick if (zio->io_error != ENOSPC) 3043e14bb325SJeff Bonwick zio->io_reexecute |= ZIO_REEXECUTE_NOW; 3044e14bb325SJeff Bonwick else 3045e14bb325SJeff Bonwick zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND; 3046b24ab676SJeff Bonwick } 3047e14bb325SJeff Bonwick 3048e14bb325SJeff Bonwick if ((zio->io_type == ZIO_TYPE_READ || 3049e14bb325SJeff Bonwick zio->io_type == ZIO_TYPE_FREE) && 305044ecc532SGeorge Wilson !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && 3051e14bb325SJeff Bonwick zio->io_error == ENXIO && 3052b16da2e2SGeorge Wilson spa_load_state(spa) == SPA_LOAD_NONE && 3053e14bb325SJeff Bonwick spa_get_failmode(spa) != ZIO_FAILURE_MODE_CONTINUE) 3054e14bb325SJeff Bonwick zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND; 3055e14bb325SJeff Bonwick 3056e14bb325SJeff Bonwick if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute) 3057e14bb325SJeff Bonwick zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND; 305822fe2c88SJonathan Adams 305922fe2c88SJonathan Adams /* 306022fe2c88SJonathan Adams * Here is a possibly good place to attempt to do 306122fe2c88SJonathan Adams * either combinatorial reconstruction or error correction 306222fe2c88SJonathan Adams * based on checksums. It also might be a good place 306322fe2c88SJonathan Adams * to send out preliminary ereports before we suspend 306422fe2c88SJonathan Adams * processing. 306522fe2c88SJonathan Adams */ 3066d63d470bSgw } 3067d63d470bSgw 306867bd71c6Sperrin /* 3069e14bb325SJeff Bonwick * If there were logical child errors, they apply to us now. 3070e14bb325SJeff Bonwick * We defer this until now to avoid conflating logical child 3071e14bb325SJeff Bonwick * errors with errors that happened to the zio itself when 3072e14bb325SJeff Bonwick * updating vdev stats and reporting FMA events above. 307367bd71c6Sperrin */ 3074e14bb325SJeff Bonwick zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL); 30758654d025Sperrin 3076b24ab676SJeff Bonwick if ((zio->io_error || zio->io_reexecute) && 3077b24ab676SJeff Bonwick IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio && 307880901aeaSGeorge Wilson !(zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE))) 3079f5383399SBill Moore zio_dva_unallocate(zio, zio->io_gang_tree, bp); 3080f5383399SBill Moore 3081f5383399SBill Moore zio_gang_tree_free(&zio->io_gang_tree); 3082f5383399SBill Moore 308333a372edSGeorge Wilson /* 308433a372edSGeorge Wilson * Godfather I/Os should never suspend. 308533a372edSGeorge Wilson */ 308633a372edSGeorge Wilson if ((zio->io_flags & ZIO_FLAG_GODFATHER) && 308733a372edSGeorge Wilson (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) 308833a372edSGeorge Wilson zio->io_reexecute = 0; 308933a372edSGeorge Wilson 309033a372edSGeorge Wilson if (zio->io_reexecute) { 3091e14bb325SJeff Bonwick /* 3092e14bb325SJeff Bonwick * This is a logical I/O that wants to reexecute. 3093e14bb325SJeff Bonwick * 3094e14bb325SJeff Bonwick * Reexecute is top-down. When an i/o fails, if it's not 3095e14bb325SJeff Bonwick * the root, it simply notifies its parent and sticks around. 3096e14bb325SJeff Bonwick * The parent, seeing that it still has children in zio_done(), 3097e14bb325SJeff Bonwick * does the same. This percolates all the way up to the root. 3098e14bb325SJeff Bonwick * The root i/o will reexecute or suspend the entire tree. 3099e14bb325SJeff Bonwick * 3100e14bb325SJeff Bonwick * This approach ensures that zio_reexecute() honors 3101e14bb325SJeff Bonwick * all the original i/o dependency relationships, e.g. 3102e14bb325SJeff Bonwick * parents not executing until children are ready. 3103e14bb325SJeff Bonwick */ 3104e14bb325SJeff Bonwick ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 3105fa9e4066Sahrens 3106f5383399SBill Moore zio->io_gang_leader = NULL; 3107e14bb325SJeff Bonwick 3108a3f829aeSBill Moore mutex_enter(&zio->io_lock); 3109a3f829aeSBill Moore zio->io_state[ZIO_WAIT_DONE] = 1; 3110a3f829aeSBill Moore mutex_exit(&zio->io_lock); 3111a3f829aeSBill Moore 311254d692b7SGeorge Wilson /* 311354d692b7SGeorge Wilson * "The Godfather" I/O monitors its children but is 311454d692b7SGeorge Wilson * not a true parent to them. It will track them through 311554d692b7SGeorge Wilson * the pipeline but severs its ties whenever they get into 311654d692b7SGeorge Wilson * trouble (e.g. suspended). This allows "The Godfather" 311754d692b7SGeorge Wilson * I/O to return status without blocking. 311854d692b7SGeorge Wilson */ 311954d692b7SGeorge Wilson for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) { 312054d692b7SGeorge Wilson zio_link_t *zl = zio->io_walk_link; 312154d692b7SGeorge Wilson pio_next = zio_walk_parents(zio); 312254d692b7SGeorge Wilson 312354d692b7SGeorge Wilson if ((pio->io_flags & ZIO_FLAG_GODFATHER) && 312454d692b7SGeorge Wilson (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) { 312554d692b7SGeorge Wilson zio_remove_child(pio, zio, zl); 312654d692b7SGeorge Wilson zio_notify_parent(pio, zio, ZIO_WAIT_DONE); 312754d692b7SGeorge Wilson } 312854d692b7SGeorge Wilson } 312954d692b7SGeorge Wilson 3130a3f829aeSBill Moore if ((pio = zio_unique_parent(zio)) != NULL) { 3131e14bb325SJeff Bonwick /* 3132e14bb325SJeff Bonwick * We're not a root i/o, so there's nothing to do 3133e14bb325SJeff Bonwick * but notify our parent. Don't propagate errors 3134e14bb325SJeff Bonwick * upward since we haven't permanently failed yet. 3135e14bb325SJeff Bonwick */ 313633a372edSGeorge Wilson ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER)); 3137e14bb325SJeff Bonwick zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE; 3138e14bb325SJeff Bonwick zio_notify_parent(pio, zio, ZIO_WAIT_DONE); 3139e14bb325SJeff Bonwick } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) { 3140e14bb325SJeff Bonwick /* 3141e14bb325SJeff Bonwick * We'd fail again if we reexecuted now, so suspend 3142e14bb325SJeff Bonwick * until conditions improve (e.g. device comes online). 3143e14bb325SJeff Bonwick */ 3144e14bb325SJeff Bonwick zio_suspend(spa, zio); 3145e14bb325SJeff Bonwick } else { 3146e14bb325SJeff Bonwick /* 3147e14bb325SJeff Bonwick * Reexecution is potentially a huge amount of work. 3148e14bb325SJeff Bonwick * Hand it off to the otherwise-unused claim taskq. 3149e14bb325SJeff Bonwick */ 31505aeb9474SGarrett D'Amore ASSERT(zio->io_tqent.tqent_next == NULL); 3151ec94d322SAdam Leventhal spa_taskq_dispatch_ent(spa, ZIO_TYPE_CLAIM, 3152ec94d322SAdam Leventhal ZIO_TASKQ_ISSUE, (task_func_t *)zio_reexecute, zio, 3153ec94d322SAdam Leventhal 0, &zio->io_tqent); 3154e14bb325SJeff Bonwick } 3155e14bb325SJeff Bonwick return (ZIO_PIPELINE_STOP); 3156fa9e4066Sahrens } 3157fa9e4066Sahrens 3158b24ab676SJeff Bonwick ASSERT(zio->io_child_count == 0); 315933a372edSGeorge Wilson ASSERT(zio->io_reexecute == 0); 3160e14bb325SJeff Bonwick ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL)); 3161fa9e4066Sahrens 3162b24ab676SJeff Bonwick /* 3163b24ab676SJeff Bonwick * Report any checksum errors, since the I/O is complete. 3164b24ab676SJeff Bonwick */ 316522fe2c88SJonathan Adams while (zio->io_cksum_report != NULL) { 3166b24ab676SJeff Bonwick zio_cksum_report_t *zcr = zio->io_cksum_report; 3167b24ab676SJeff Bonwick zio->io_cksum_report = zcr->zcr_next; 3168b24ab676SJeff Bonwick zcr->zcr_next = NULL; 3169b24ab676SJeff Bonwick zcr->zcr_finish(zcr, NULL); 3170b24ab676SJeff Bonwick zfs_ereport_free_checksum(zcr); 317122fe2c88SJonathan Adams } 317222fe2c88SJonathan Adams 3173a3f829aeSBill Moore /* 3174a3f829aeSBill Moore * It is the responsibility of the done callback to ensure that this 3175a3f829aeSBill Moore * particular zio is no longer discoverable for adoption, and as 3176a3f829aeSBill Moore * such, cannot acquire any new parents. 3177a3f829aeSBill Moore */ 3178e14bb325SJeff Bonwick if (zio->io_done) 3179e14bb325SJeff Bonwick zio->io_done(zio); 3180fa9e4066Sahrens 3181a3f829aeSBill Moore mutex_enter(&zio->io_lock); 3182a3f829aeSBill Moore zio->io_state[ZIO_WAIT_DONE] = 1; 3183a3f829aeSBill Moore mutex_exit(&zio->io_lock); 3184fa9e4066Sahrens 3185a3f829aeSBill Moore for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) { 3186a3f829aeSBill Moore zio_link_t *zl = zio->io_walk_link; 3187a3f829aeSBill Moore pio_next = zio_walk_parents(zio); 3188a3f829aeSBill Moore zio_remove_child(pio, zio, zl); 3189e14bb325SJeff Bonwick zio_notify_parent(pio, zio, ZIO_WAIT_DONE); 3190e14bb325SJeff Bonwick } 3191fa9e4066Sahrens 3192e14bb325SJeff Bonwick if (zio->io_waiter != NULL) { 3193e14bb325SJeff Bonwick mutex_enter(&zio->io_lock); 3194e14bb325SJeff Bonwick zio->io_executor = NULL; 3195e14bb325SJeff Bonwick cv_broadcast(&zio->io_cv); 3196e14bb325SJeff Bonwick mutex_exit(&zio->io_lock); 3197e14bb325SJeff Bonwick } else { 3198e14bb325SJeff Bonwick zio_destroy(zio); 3199e14bb325SJeff Bonwick } 3200fa9e4066Sahrens 3201e14bb325SJeff Bonwick return (ZIO_PIPELINE_STOP); 3202fa9e4066Sahrens } 320346341222Sperrin 320446341222Sperrin /* 3205e14bb325SJeff Bonwick * ========================================================================== 3206e14bb325SJeff Bonwick * I/O pipeline definition 3207e14bb325SJeff Bonwick * ========================================================================== 320846341222Sperrin */ 3209b24ab676SJeff Bonwick static zio_pipe_stage_t *zio_pipeline[] = { 3210e14bb325SJeff Bonwick NULL, 3211e14bb325SJeff Bonwick zio_read_bp_init, 3212b24ab676SJeff Bonwick zio_free_bp_init, 3213b24ab676SJeff Bonwick zio_issue_async, 3214e14bb325SJeff Bonwick zio_write_bp_init, 3215e14bb325SJeff Bonwick zio_checksum_generate, 321680901aeaSGeorge Wilson zio_nop_write, 3217b24ab676SJeff Bonwick zio_ddt_read_start, 3218b24ab676SJeff Bonwick zio_ddt_read_done, 3219b24ab676SJeff Bonwick zio_ddt_write, 3220b24ab676SJeff Bonwick zio_ddt_free, 3221e14bb325SJeff Bonwick zio_gang_assemble, 3222e14bb325SJeff Bonwick zio_gang_issue, 3223e14bb325SJeff Bonwick zio_dva_allocate, 3224e14bb325SJeff Bonwick zio_dva_free, 3225e14bb325SJeff Bonwick zio_dva_claim, 3226e14bb325SJeff Bonwick zio_ready, 3227e14bb325SJeff Bonwick zio_vdev_io_start, 3228e14bb325SJeff Bonwick zio_vdev_io_done, 3229e14bb325SJeff Bonwick zio_vdev_io_assess, 3230e14bb325SJeff Bonwick zio_checksum_verify, 3231e14bb325SJeff Bonwick zio_done 3232e14bb325SJeff Bonwick }; 3233ad135b5dSChristopher Siden 3234ad135b5dSChristopher Siden /* dnp is the dnode for zb1->zb_object */ 3235ad135b5dSChristopher Siden boolean_t 32367802d7bfSMatthew Ahrens zbookmark_is_before(const dnode_phys_t *dnp, const zbookmark_phys_t *zb1, 32377802d7bfSMatthew Ahrens const zbookmark_phys_t *zb2) 3238ad135b5dSChristopher Siden { 3239ad135b5dSChristopher Siden uint64_t zb1nextL0, zb2thisobj; 3240ad135b5dSChristopher Siden 3241ad135b5dSChristopher Siden ASSERT(zb1->zb_objset == zb2->zb_objset); 3242ad135b5dSChristopher Siden ASSERT(zb2->zb_level == 0); 3243ad135b5dSChristopher Siden 3244ad135b5dSChristopher Siden /* The objset_phys_t isn't before anything. */ 3245ad135b5dSChristopher Siden if (dnp == NULL) 3246ad135b5dSChristopher Siden return (B_FALSE); 3247ad135b5dSChristopher Siden 3248ad135b5dSChristopher Siden zb1nextL0 = (zb1->zb_blkid + 1) << 3249ad135b5dSChristopher Siden ((zb1->zb_level) * (dnp->dn_indblkshift - SPA_BLKPTRSHIFT)); 3250ad135b5dSChristopher Siden 3251ad135b5dSChristopher Siden zb2thisobj = zb2->zb_object ? zb2->zb_object : 3252ad135b5dSChristopher Siden zb2->zb_blkid << (DNODE_BLOCK_SHIFT - DNODE_SHIFT); 3253ad135b5dSChristopher Siden 3254ad135b5dSChristopher Siden if (zb1->zb_object == DMU_META_DNODE_OBJECT) { 3255ad135b5dSChristopher Siden uint64_t nextobj = zb1nextL0 * 3256ad135b5dSChristopher Siden (dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT) >> DNODE_SHIFT; 3257ad135b5dSChristopher Siden return (nextobj <= zb2thisobj); 3258ad135b5dSChristopher Siden } 3259ad135b5dSChristopher Siden 3260ad135b5dSChristopher Siden if (zb1->zb_object < zb2thisobj) 3261ad135b5dSChristopher Siden return (B_TRUE); 3262ad135b5dSChristopher Siden if (zb1->zb_object > zb2thisobj) 3263ad135b5dSChristopher Siden return (B_FALSE); 3264ad135b5dSChristopher Siden if (zb2->zb_object == DMU_META_DNODE_OBJECT) 3265ad135b5dSChristopher Siden return (B_FALSE); 3266ad135b5dSChristopher Siden return (zb1nextL0 <= zb2->zb_blkid); 3267ad135b5dSChristopher Siden } 3268