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. 23f78cdc34SPaul Dagnelie * Copyright (c) 2011, 2018 by Delphix. All rights reserved. 245aeb9474SGarrett D'Amore * Copyright (c) 2011 Nexenta Systems, Inc. All rights reserved. 25c3d26abcSMatthew Ahrens * Copyright (c) 2014 Integros [integros.com] 26663207adSDon Brady * Copyright (c) 2017, Intel Corporation. 27fa9e4066Sahrens */ 28fa9e4066Sahrens 29de710d24SJosef 'Jeff' Sipek #include <sys/sysmacros.h> 30fa9e4066Sahrens #include <sys/zfs_context.h> 31ea8dc4b6Seschrock #include <sys/fm/fs/zfs.h> 32fa9e4066Sahrens #include <sys/spa.h> 33fa9e4066Sahrens #include <sys/txg.h> 34fa9e4066Sahrens #include <sys/spa_impl.h> 35fa9e4066Sahrens #include <sys/vdev_impl.h> 36fa9e4066Sahrens #include <sys/zio_impl.h> 37fa9e4066Sahrens #include <sys/zio_compress.h> 38fa9e4066Sahrens #include <sys/zio_checksum.h> 39b24ab676SJeff Bonwick #include <sys/dmu_objset.h> 40b24ab676SJeff Bonwick #include <sys/arc.h> 41b24ab676SJeff Bonwick #include <sys/ddt.h> 425d7b4d43SMatthew Ahrens #include <sys/blkptr.h> 4343466aaeSMax Grossman #include <sys/zfeature.h> 44*a3874b8bSToomas Soome #include <sys/dsl_scan.h> 450f7643c7SGeorge Wilson #include <sys/metaslab_impl.h> 46770499e1SDan Kimmel #include <sys/abd.h> 47f78cdc34SPaul Dagnelie #include <sys/cityhash.h> 48fa9e4066Sahrens 49fa9e4066Sahrens /* 50fa9e4066Sahrens * ========================================================================== 51fa9e4066Sahrens * I/O type descriptions 52fa9e4066Sahrens * ========================================================================== 53fa9e4066Sahrens */ 5469962b56SMatthew Ahrens const char *zio_type_name[ZIO_TYPES] = { 5580eb36f2SGeorge Wilson "zio_null", "zio_read", "zio_write", "zio_free", "zio_claim", 5680eb36f2SGeorge Wilson "zio_ioctl" 5780eb36f2SGeorge Wilson }; 58fa9e4066Sahrens 590f7643c7SGeorge Wilson boolean_t zio_dva_throttle_enabled = B_TRUE; 600f7643c7SGeorge Wilson 61fa9e4066Sahrens /* 62fa9e4066Sahrens * ========================================================================== 63fa9e4066Sahrens * I/O kmem caches 64fa9e4066Sahrens * ========================================================================== 65fa9e4066Sahrens */ 66ccae0b50Seschrock kmem_cache_t *zio_cache; 67a3f829aeSBill Moore kmem_cache_t *zio_link_cache; 68fa9e4066Sahrens kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT]; 69ad23a2dbSjohansen kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT]; 70ad23a2dbSjohansen 71ad23a2dbSjohansen #ifdef _KERNEL 72ad23a2dbSjohansen extern vmem_t *zio_alloc_arena; 73ad23a2dbSjohansen #endif 74fa9e4066Sahrens 75738f37bcSGeorge Wilson #define ZIO_PIPELINE_CONTINUE 0x100 76738f37bcSGeorge Wilson #define ZIO_PIPELINE_STOP 0x101 77738f37bcSGeorge Wilson 78a2cdcdd2SPaul Dagnelie #define BP_SPANB(indblkshift, level) \ 79a2cdcdd2SPaul Dagnelie (((uint64_t)1) << ((level) * ((indblkshift) - SPA_BLKPTRSHIFT))) 80a2cdcdd2SPaul Dagnelie #define COMPARE_META_LEVEL 0x80000000ul 8101f55e48SGeorge Wilson /* 8201f55e48SGeorge Wilson * The following actions directly effect the spa's sync-to-convergence logic. 8301f55e48SGeorge Wilson * The values below define the sync pass when we start performing the action. 8401f55e48SGeorge Wilson * Care should be taken when changing these values as they directly impact 8501f55e48SGeorge Wilson * spa_sync() performance. Tuning these values may introduce subtle performance 8601f55e48SGeorge Wilson * pathologies and should only be done in the context of performance analysis. 8701f55e48SGeorge Wilson * These tunables will eventually be removed and replaced with #defines once 8801f55e48SGeorge Wilson * enough analysis has been done to determine optimal values. 8901f55e48SGeorge Wilson * 9001f55e48SGeorge Wilson * The 'zfs_sync_pass_deferred_free' pass must be greater than 1 to ensure that 9101f55e48SGeorge Wilson * regular blocks are not deferred. 9201f55e48SGeorge Wilson */ 9301f55e48SGeorge Wilson int zfs_sync_pass_deferred_free = 2; /* defer frees starting in this pass */ 9401f55e48SGeorge Wilson int zfs_sync_pass_dont_compress = 5; /* don't compress starting in this pass */ 9501f55e48SGeorge Wilson int zfs_sync_pass_rewrite = 2; /* rewrite new bps starting in this pass */ 9601f55e48SGeorge Wilson 970a4e9518Sgw /* 98e14bb325SJeff Bonwick * An allocating zio is one that either currently has the DVA allocate 99e14bb325SJeff Bonwick * stage set or will have it later in its lifetime. 1000a4e9518Sgw */ 101b24ab676SJeff Bonwick #define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE) 102b24ab676SJeff Bonwick 10335a5a358SJonathan Adams boolean_t zio_requeue_io_start_cut_in_line = B_TRUE; 10435a5a358SJonathan Adams 105b24ab676SJeff Bonwick #ifdef ZFS_DEBUG 106b24ab676SJeff Bonwick int zio_buf_debug_limit = 16384; 107b24ab676SJeff Bonwick #else 108b24ab676SJeff Bonwick int zio_buf_debug_limit = 0; 109b24ab676SJeff Bonwick #endif 1100a4e9518Sgw 1110f7643c7SGeorge Wilson static void zio_taskq_dispatch(zio_t *, zio_taskq_type_t, boolean_t); 1120f7643c7SGeorge Wilson 113fa9e4066Sahrens void 114fa9e4066Sahrens zio_init(void) 115fa9e4066Sahrens { 116fa9e4066Sahrens size_t c; 117ad23a2dbSjohansen vmem_t *data_alloc_arena = NULL; 118ad23a2dbSjohansen 119ad23a2dbSjohansen #ifdef _KERNEL 120ad23a2dbSjohansen data_alloc_arena = zio_alloc_arena; 121ad23a2dbSjohansen #endif 122a3f829aeSBill Moore zio_cache = kmem_cache_create("zio_cache", 123a3f829aeSBill Moore sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0); 124a3f829aeSBill Moore zio_link_cache = kmem_cache_create("zio_link_cache", 125a3f829aeSBill Moore sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0); 126ccae0b50Seschrock 127fa9e4066Sahrens /* 128fa9e4066Sahrens * For small buffers, we want a cache for each multiple of 129b5152584SMatthew Ahrens * SPA_MINBLOCKSIZE. For larger buffers, we want a cache 130b5152584SMatthew Ahrens * for each quarter-power of 2. 131fa9e4066Sahrens */ 132fa9e4066Sahrens for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) { 133fa9e4066Sahrens size_t size = (c + 1) << SPA_MINBLOCKSHIFT; 134fa9e4066Sahrens size_t p2 = size; 135fa9e4066Sahrens size_t align = 0; 136e291592aSJonathan Adams size_t cflags = (size > zio_buf_debug_limit) ? KMC_NODEBUG : 0; 137fa9e4066Sahrens 138de710d24SJosef 'Jeff' Sipek while (!ISP2(p2)) 139fa9e4066Sahrens p2 &= p2 - 1; 140fa9e4066Sahrens 141cd1c8b85SMatthew Ahrens #ifndef _KERNEL 142cd1c8b85SMatthew Ahrens /* 143cd1c8b85SMatthew Ahrens * If we are using watchpoints, put each buffer on its own page, 144cd1c8b85SMatthew Ahrens * to eliminate the performance overhead of trapping to the 145cd1c8b85SMatthew Ahrens * kernel when modifying a non-watched buffer that shares the 146cd1c8b85SMatthew Ahrens * page with a watched buffer. 147cd1c8b85SMatthew Ahrens */ 148cd1c8b85SMatthew Ahrens if (arc_watch && !IS_P2ALIGNED(size, PAGESIZE)) 149cd1c8b85SMatthew Ahrens continue; 150cd1c8b85SMatthew Ahrens #endif 151fa9e4066Sahrens if (size <= 4 * SPA_MINBLOCKSIZE) { 152fa9e4066Sahrens align = SPA_MINBLOCKSIZE; 153cd1c8b85SMatthew Ahrens } else if (IS_P2ALIGNED(size, p2 >> 2)) { 154b5152584SMatthew Ahrens align = MIN(p2 >> 2, PAGESIZE); 155fa9e4066Sahrens } 156fa9e4066Sahrens 157fa9e4066Sahrens if (align != 0) { 158ad23a2dbSjohansen char name[36]; 1595ad82045Snd (void) sprintf(name, "zio_buf_%lu", (ulong_t)size); 160fa9e4066Sahrens zio_buf_cache[c] = kmem_cache_create(name, size, 161e291592aSJonathan Adams align, NULL, NULL, NULL, NULL, NULL, cflags); 162ad23a2dbSjohansen 163e291592aSJonathan Adams /* 164e291592aSJonathan Adams * Since zio_data bufs do not appear in crash dumps, we 165e291592aSJonathan Adams * pass KMC_NOTOUCH so that no allocator metadata is 166e291592aSJonathan Adams * stored with the buffers. 167e291592aSJonathan Adams */ 168ad23a2dbSjohansen (void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size); 169ad23a2dbSjohansen zio_data_buf_cache[c] = kmem_cache_create(name, size, 170ad23a2dbSjohansen align, NULL, NULL, NULL, NULL, data_alloc_arena, 171e291592aSJonathan Adams cflags | KMC_NOTOUCH); 172fa9e4066Sahrens } 173fa9e4066Sahrens } 174fa9e4066Sahrens 175fa9e4066Sahrens while (--c != 0) { 176fa9e4066Sahrens ASSERT(zio_buf_cache[c] != NULL); 177fa9e4066Sahrens if (zio_buf_cache[c - 1] == NULL) 178fa9e4066Sahrens zio_buf_cache[c - 1] = zio_buf_cache[c]; 179ad23a2dbSjohansen 180ad23a2dbSjohansen ASSERT(zio_data_buf_cache[c] != NULL); 181ad23a2dbSjohansen if (zio_data_buf_cache[c - 1] == NULL) 182ad23a2dbSjohansen zio_data_buf_cache[c - 1] = zio_data_buf_cache[c]; 183fa9e4066Sahrens } 184ea8dc4b6Seschrock 185ea8dc4b6Seschrock zio_inject_init(); 186fa9e4066Sahrens } 187fa9e4066Sahrens 188fa9e4066Sahrens void 189fa9e4066Sahrens zio_fini(void) 190fa9e4066Sahrens { 191fa9e4066Sahrens size_t c; 192fa9e4066Sahrens kmem_cache_t *last_cache = NULL; 193ad23a2dbSjohansen kmem_cache_t *last_data_cache = NULL; 194fa9e4066Sahrens 195fa9e4066Sahrens for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) { 196fa9e4066Sahrens if (zio_buf_cache[c] != last_cache) { 197fa9e4066Sahrens last_cache = zio_buf_cache[c]; 198fa9e4066Sahrens kmem_cache_destroy(zio_buf_cache[c]); 199fa9e4066Sahrens } 200fa9e4066Sahrens zio_buf_cache[c] = NULL; 201ad23a2dbSjohansen 202ad23a2dbSjohansen if (zio_data_buf_cache[c] != last_data_cache) { 203ad23a2dbSjohansen last_data_cache = zio_data_buf_cache[c]; 204ad23a2dbSjohansen kmem_cache_destroy(zio_data_buf_cache[c]); 205ad23a2dbSjohansen } 206ad23a2dbSjohansen zio_data_buf_cache[c] = NULL; 207fa9e4066Sahrens } 208ea8dc4b6Seschrock 209a3f829aeSBill Moore kmem_cache_destroy(zio_link_cache); 210ccae0b50Seschrock kmem_cache_destroy(zio_cache); 211ccae0b50Seschrock 212ea8dc4b6Seschrock zio_inject_fini(); 213fa9e4066Sahrens } 214fa9e4066Sahrens 215fa9e4066Sahrens /* 216fa9e4066Sahrens * ========================================================================== 217fa9e4066Sahrens * Allocate and free I/O buffers 218fa9e4066Sahrens * ========================================================================== 219fa9e4066Sahrens */ 220ad23a2dbSjohansen 221ad23a2dbSjohansen /* 222ad23a2dbSjohansen * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a 223ad23a2dbSjohansen * crashdump if the kernel panics, so use it judiciously. Obviously, it's 224ad23a2dbSjohansen * useful to inspect ZFS metadata, but if possible, we should avoid keeping 225ad23a2dbSjohansen * excess / transient data in-core during a crashdump. 226ad23a2dbSjohansen */ 227fa9e4066Sahrens void * 228fa9e4066Sahrens zio_buf_alloc(size_t size) 229fa9e4066Sahrens { 230fa9e4066Sahrens size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; 231fa9e4066Sahrens 232f63ab3d5SMatthew Ahrens VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); 233fa9e4066Sahrens 2341ab7f2deSmaybee return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE)); 235fa9e4066Sahrens } 236fa9e4066Sahrens 237ad23a2dbSjohansen /* 238ad23a2dbSjohansen * Use zio_data_buf_alloc to allocate data. The data will not appear in a 239ad23a2dbSjohansen * crashdump if the kernel panics. This exists so that we will limit the amount 240ad23a2dbSjohansen * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount 241ad23a2dbSjohansen * of kernel heap dumped to disk when the kernel panics) 242ad23a2dbSjohansen */ 243ad23a2dbSjohansen void * 244ad23a2dbSjohansen zio_data_buf_alloc(size_t size) 245ad23a2dbSjohansen { 246ad23a2dbSjohansen size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; 247ad23a2dbSjohansen 248f63ab3d5SMatthew Ahrens VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); 249ad23a2dbSjohansen 2501ab7f2deSmaybee return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE)); 251ad23a2dbSjohansen } 252ad23a2dbSjohansen 253fa9e4066Sahrens void 254fa9e4066Sahrens zio_buf_free(void *buf, size_t size) 255fa9e4066Sahrens { 256fa9e4066Sahrens size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; 257fa9e4066Sahrens 258f63ab3d5SMatthew Ahrens VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); 259fa9e4066Sahrens 260fa9e4066Sahrens kmem_cache_free(zio_buf_cache[c], buf); 261fa9e4066Sahrens } 262fa9e4066Sahrens 263ad23a2dbSjohansen void 264ad23a2dbSjohansen zio_data_buf_free(void *buf, size_t size) 265ad23a2dbSjohansen { 266ad23a2dbSjohansen size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; 267ad23a2dbSjohansen 268f63ab3d5SMatthew Ahrens VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); 269ad23a2dbSjohansen 270ad23a2dbSjohansen kmem_cache_free(zio_data_buf_cache[c], buf); 271ad23a2dbSjohansen } 272b3995adbSahrens 273fa9e4066Sahrens /* 274fa9e4066Sahrens * ========================================================================== 275fa9e4066Sahrens * Push and pop I/O transform buffers 276fa9e4066Sahrens * ========================================================================== 277fa9e4066Sahrens */ 278dcbf3bd6SGeorge Wilson void 279770499e1SDan Kimmel zio_push_transform(zio_t *zio, abd_t *data, uint64_t size, uint64_t bufsize, 2809a686fbcSPaul Dagnelie zio_transform_func_t *transform) 281fa9e4066Sahrens { 282fa9e4066Sahrens zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP); 283fa9e4066Sahrens 284770499e1SDan Kimmel /* 285770499e1SDan Kimmel * Ensure that anyone expecting this zio to contain a linear ABD isn't 286770499e1SDan Kimmel * going to get a nasty surprise when they try to access the data. 287770499e1SDan Kimmel */ 288770499e1SDan Kimmel IMPLY(abd_is_linear(zio->io_abd), abd_is_linear(data)); 289770499e1SDan Kimmel 290770499e1SDan Kimmel zt->zt_orig_abd = zio->io_abd; 291e14bb325SJeff Bonwick zt->zt_orig_size = zio->io_size; 292fa9e4066Sahrens zt->zt_bufsize = bufsize; 293e14bb325SJeff Bonwick zt->zt_transform = transform; 294fa9e4066Sahrens 295fa9e4066Sahrens zt->zt_next = zio->io_transform_stack; 296fa9e4066Sahrens zio->io_transform_stack = zt; 297fa9e4066Sahrens 298770499e1SDan Kimmel zio->io_abd = data; 299fa9e4066Sahrens zio->io_size = size; 300fa9e4066Sahrens } 301fa9e4066Sahrens 302dcbf3bd6SGeorge Wilson void 303e14bb325SJeff Bonwick zio_pop_transforms(zio_t *zio) 304fa9e4066Sahrens { 305e14bb325SJeff Bonwick zio_transform_t *zt; 306e14bb325SJeff Bonwick 307e14bb325SJeff Bonwick while ((zt = zio->io_transform_stack) != NULL) { 308e14bb325SJeff Bonwick if (zt->zt_transform != NULL) 309e14bb325SJeff Bonwick zt->zt_transform(zio, 310770499e1SDan Kimmel zt->zt_orig_abd, zt->zt_orig_size); 311fa9e4066Sahrens 312b24ab676SJeff Bonwick if (zt->zt_bufsize != 0) 313770499e1SDan Kimmel abd_free(zio->io_abd); 314fa9e4066Sahrens 315770499e1SDan Kimmel zio->io_abd = zt->zt_orig_abd; 316e14bb325SJeff Bonwick zio->io_size = zt->zt_orig_size; 317e14bb325SJeff Bonwick zio->io_transform_stack = zt->zt_next; 318fa9e4066Sahrens 319e14bb325SJeff Bonwick kmem_free(zt, sizeof (zio_transform_t)); 320fa9e4066Sahrens } 321fa9e4066Sahrens } 322fa9e4066Sahrens 323e14bb325SJeff Bonwick /* 324e14bb325SJeff Bonwick * ========================================================================== 325e14bb325SJeff Bonwick * I/O transform callbacks for subblocks and decompression 326e14bb325SJeff Bonwick * ========================================================================== 327e14bb325SJeff Bonwick */ 328e14bb325SJeff Bonwick static void 329770499e1SDan Kimmel zio_subblock(zio_t *zio, abd_t *data, uint64_t size) 330e14bb325SJeff Bonwick { 331e14bb325SJeff Bonwick ASSERT(zio->io_size > size); 332e14bb325SJeff Bonwick 333e14bb325SJeff Bonwick if (zio->io_type == ZIO_TYPE_READ) 334770499e1SDan Kimmel abd_copy(data, zio->io_abd, size); 335e14bb325SJeff Bonwick } 336e14bb325SJeff Bonwick 337e14bb325SJeff Bonwick static void 338770499e1SDan Kimmel zio_decompress(zio_t *zio, abd_t *data, uint64_t size) 339e14bb325SJeff Bonwick { 340770499e1SDan Kimmel if (zio->io_error == 0) { 341770499e1SDan Kimmel void *tmp = abd_borrow_buf(data, size); 342770499e1SDan Kimmel int ret = zio_decompress_data(BP_GET_COMPRESS(zio->io_bp), 343770499e1SDan Kimmel zio->io_abd, tmp, zio->io_size, size); 344770499e1SDan Kimmel abd_return_buf_copy(data, tmp, size); 345770499e1SDan Kimmel 346770499e1SDan Kimmel if (ret != 0) 347770499e1SDan Kimmel zio->io_error = SET_ERROR(EIO); 348770499e1SDan Kimmel } 349e14bb325SJeff Bonwick } 350e14bb325SJeff Bonwick 351e14bb325SJeff Bonwick /* 352e14bb325SJeff Bonwick * ========================================================================== 353e14bb325SJeff Bonwick * I/O parent/child relationships and pipeline interlocks 354e14bb325SJeff Bonwick * ========================================================================== 355e14bb325SJeff Bonwick */ 356a3f829aeSBill Moore zio_t * 3570f7643c7SGeorge Wilson zio_walk_parents(zio_t *cio, zio_link_t **zl) 358a3f829aeSBill Moore { 359a3f829aeSBill Moore list_t *pl = &cio->io_parent_list; 360e14bb325SJeff Bonwick 3610f7643c7SGeorge Wilson *zl = (*zl == NULL) ? list_head(pl) : list_next(pl, *zl); 3620f7643c7SGeorge Wilson if (*zl == NULL) 363a3f829aeSBill Moore return (NULL); 364a3f829aeSBill Moore 3650f7643c7SGeorge Wilson ASSERT((*zl)->zl_child == cio); 3660f7643c7SGeorge Wilson return ((*zl)->zl_parent); 367a3f829aeSBill Moore } 368a3f829aeSBill Moore 369a3f829aeSBill Moore zio_t * 3700f7643c7SGeorge Wilson zio_walk_children(zio_t *pio, zio_link_t **zl) 371a3f829aeSBill Moore { 372a3f829aeSBill Moore list_t *cl = &pio->io_child_list; 373a3f829aeSBill Moore 374*a3874b8bSToomas Soome ASSERT(MUTEX_HELD(&pio->io_lock)); 375*a3874b8bSToomas Soome 3760f7643c7SGeorge Wilson *zl = (*zl == NULL) ? list_head(cl) : list_next(cl, *zl); 3770f7643c7SGeorge Wilson if (*zl == NULL) 378a3f829aeSBill Moore return (NULL); 379a3f829aeSBill Moore 3800f7643c7SGeorge Wilson ASSERT((*zl)->zl_parent == pio); 3810f7643c7SGeorge Wilson return ((*zl)->zl_child); 382a3f829aeSBill Moore } 383a3f829aeSBill Moore 384a3f829aeSBill Moore zio_t * 385a3f829aeSBill Moore zio_unique_parent(zio_t *cio) 386a3f829aeSBill Moore { 3870f7643c7SGeorge Wilson zio_link_t *zl = NULL; 3880f7643c7SGeorge Wilson zio_t *pio = zio_walk_parents(cio, &zl); 389a3f829aeSBill Moore 3900f7643c7SGeorge Wilson VERIFY3P(zio_walk_parents(cio, &zl), ==, NULL); 391a3f829aeSBill Moore return (pio); 392a3f829aeSBill Moore } 393a3f829aeSBill Moore 394a3f829aeSBill Moore void 395a3f829aeSBill Moore zio_add_child(zio_t *pio, zio_t *cio) 396e14bb325SJeff Bonwick { 397a3f829aeSBill Moore zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP); 398a3f829aeSBill Moore 399a3f829aeSBill Moore /* 400a3f829aeSBill Moore * Logical I/Os can have logical, gang, or vdev children. 401a3f829aeSBill Moore * Gang I/Os can have gang or vdev children. 402a3f829aeSBill Moore * Vdev I/Os can only have vdev children. 403a3f829aeSBill Moore * The following ASSERT captures all of these constraints. 404a3f829aeSBill Moore */ 4051271e4b1SPrakash Surya ASSERT3S(cio->io_child_type, <=, pio->io_child_type); 406a3f829aeSBill Moore 407a3f829aeSBill Moore zl->zl_parent = pio; 408a3f829aeSBill Moore zl->zl_child = cio; 409a3f829aeSBill Moore 410e14bb325SJeff Bonwick mutex_enter(&pio->io_lock); 411*a3874b8bSToomas Soome mutex_enter(&cio->io_lock); 412a3f829aeSBill Moore 413a3f829aeSBill Moore ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0); 414a3f829aeSBill Moore 415a3f829aeSBill Moore for (int w = 0; w < ZIO_WAIT_TYPES; w++) 416a3f829aeSBill Moore pio->io_children[cio->io_child_type][w] += !cio->io_state[w]; 417a3f829aeSBill Moore 418a3f829aeSBill Moore list_insert_head(&pio->io_child_list, zl); 419a3f829aeSBill Moore list_insert_head(&cio->io_parent_list, zl); 420a3f829aeSBill Moore 421b24ab676SJeff Bonwick pio->io_child_count++; 422b24ab676SJeff Bonwick cio->io_parent_count++; 423b24ab676SJeff Bonwick 424a3f829aeSBill Moore mutex_exit(&cio->io_lock); 425*a3874b8bSToomas Soome mutex_exit(&pio->io_lock); 426e14bb325SJeff Bonwick } 427e14bb325SJeff Bonwick 428fa9e4066Sahrens static void 429a3f829aeSBill Moore zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl) 430e14bb325SJeff Bonwick { 431a3f829aeSBill Moore ASSERT(zl->zl_parent == pio); 432a3f829aeSBill Moore ASSERT(zl->zl_child == cio); 433e14bb325SJeff Bonwick 434e14bb325SJeff Bonwick mutex_enter(&pio->io_lock); 435*a3874b8bSToomas Soome mutex_enter(&cio->io_lock); 436a3f829aeSBill Moore 437a3f829aeSBill Moore list_remove(&pio->io_child_list, zl); 438a3f829aeSBill Moore list_remove(&cio->io_parent_list, zl); 439a3f829aeSBill Moore 440b24ab676SJeff Bonwick pio->io_child_count--; 441b24ab676SJeff Bonwick cio->io_parent_count--; 442b24ab676SJeff Bonwick 443a3f829aeSBill Moore mutex_exit(&cio->io_lock); 444*a3874b8bSToomas Soome mutex_exit(&pio->io_lock); 445a3f829aeSBill Moore 446a3f829aeSBill Moore kmem_cache_free(zio_link_cache, zl); 447e14bb325SJeff Bonwick } 448e14bb325SJeff Bonwick 449e14bb325SJeff Bonwick static boolean_t 450d6e1c446SGeorge Wilson zio_wait_for_children(zio_t *zio, uint8_t childbits, enum zio_wait_type wait) 451fa9e4066Sahrens { 452e14bb325SJeff Bonwick boolean_t waiting = B_FALSE; 453e14bb325SJeff Bonwick 454e14bb325SJeff Bonwick mutex_enter(&zio->io_lock); 455e14bb325SJeff Bonwick ASSERT(zio->io_stall == NULL); 456d6e1c446SGeorge Wilson for (int c = 0; c < ZIO_CHILD_TYPES; c++) { 457d6e1c446SGeorge Wilson if (!(ZIO_CHILD_BIT_IS_SET(childbits, c))) 458d6e1c446SGeorge Wilson continue; 459d6e1c446SGeorge Wilson 460d6e1c446SGeorge Wilson uint64_t *countp = &zio->io_children[c][wait]; 461d6e1c446SGeorge Wilson if (*countp != 0) { 462d6e1c446SGeorge Wilson zio->io_stage >>= 1; 463d6e1c446SGeorge Wilson ASSERT3U(zio->io_stage, !=, ZIO_STAGE_OPEN); 464d6e1c446SGeorge Wilson zio->io_stall = countp; 465d6e1c446SGeorge Wilson waiting = B_TRUE; 466d6e1c446SGeorge Wilson break; 467d6e1c446SGeorge Wilson } 468e14bb325SJeff Bonwick } 469e14bb325SJeff Bonwick mutex_exit(&zio->io_lock); 470e14bb325SJeff Bonwick return (waiting); 471e14bb325SJeff Bonwick } 472fa9e4066Sahrens 473e14bb325SJeff Bonwick static void 474e14bb325SJeff Bonwick zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait) 475e14bb325SJeff Bonwick { 476e14bb325SJeff Bonwick uint64_t *countp = &pio->io_children[zio->io_child_type][wait]; 477e14bb325SJeff Bonwick int *errorp = &pio->io_child_error[zio->io_child_type]; 478fa9e4066Sahrens 479e14bb325SJeff Bonwick mutex_enter(&pio->io_lock); 480e14bb325SJeff Bonwick if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE)) 481e14bb325SJeff Bonwick *errorp = zio_worst_error(*errorp, zio->io_error); 482e14bb325SJeff Bonwick pio->io_reexecute |= zio->io_reexecute; 483e14bb325SJeff Bonwick ASSERT3U(*countp, >, 0); 48469962b56SMatthew Ahrens 48569962b56SMatthew Ahrens (*countp)--; 48669962b56SMatthew Ahrens 48769962b56SMatthew Ahrens if (*countp == 0 && pio->io_stall == countp) { 4880f7643c7SGeorge Wilson zio_taskq_type_t type = 4890f7643c7SGeorge Wilson pio->io_stage < ZIO_STAGE_VDEV_IO_START ? ZIO_TASKQ_ISSUE : 4900f7643c7SGeorge Wilson ZIO_TASKQ_INTERRUPT; 491e14bb325SJeff Bonwick pio->io_stall = NULL; 492e14bb325SJeff Bonwick mutex_exit(&pio->io_lock); 4930f7643c7SGeorge Wilson /* 4940f7643c7SGeorge Wilson * Dispatch the parent zio in its own taskq so that 4950f7643c7SGeorge Wilson * the child can continue to make progress. This also 4960f7643c7SGeorge Wilson * prevents overflowing the stack when we have deeply nested 4970f7643c7SGeorge Wilson * parent-child relationships. 4980f7643c7SGeorge Wilson */ 4990f7643c7SGeorge Wilson zio_taskq_dispatch(pio, type, B_FALSE); 500e14bb325SJeff Bonwick } else { 501e14bb325SJeff Bonwick mutex_exit(&pio->io_lock); 502fa9e4066Sahrens } 503fa9e4066Sahrens } 504fa9e4066Sahrens 505e14bb325SJeff Bonwick static void 506e14bb325SJeff Bonwick zio_inherit_child_errors(zio_t *zio, enum zio_child c) 507e14bb325SJeff Bonwick { 508e14bb325SJeff Bonwick if (zio->io_child_error[c] != 0 && zio->io_error == 0) 509e14bb325SJeff Bonwick zio->io_error = zio->io_child_error[c]; 510e14bb325SJeff Bonwick } 511e14bb325SJeff Bonwick 5120f7643c7SGeorge Wilson int 51394c2d0ebSMatthew Ahrens zio_bookmark_compare(const void *x1, const void *x2) 5140f7643c7SGeorge Wilson { 5150f7643c7SGeorge Wilson const zio_t *z1 = x1; 5160f7643c7SGeorge Wilson const zio_t *z2 = x2; 5170f7643c7SGeorge Wilson 51894c2d0ebSMatthew Ahrens if (z1->io_bookmark.zb_objset < z2->io_bookmark.zb_objset) 5190f7643c7SGeorge Wilson return (-1); 52094c2d0ebSMatthew Ahrens if (z1->io_bookmark.zb_objset > z2->io_bookmark.zb_objset) 5210f7643c7SGeorge Wilson return (1); 5220f7643c7SGeorge Wilson 52394c2d0ebSMatthew Ahrens if (z1->io_bookmark.zb_object < z2->io_bookmark.zb_object) 5240f7643c7SGeorge Wilson return (-1); 52594c2d0ebSMatthew Ahrens if (z1->io_bookmark.zb_object > z2->io_bookmark.zb_object) 52694c2d0ebSMatthew Ahrens return (1); 52794c2d0ebSMatthew Ahrens 52894c2d0ebSMatthew Ahrens if (z1->io_bookmark.zb_level < z2->io_bookmark.zb_level) 52994c2d0ebSMatthew Ahrens return (-1); 53094c2d0ebSMatthew Ahrens if (z1->io_bookmark.zb_level > z2->io_bookmark.zb_level) 53194c2d0ebSMatthew Ahrens return (1); 53294c2d0ebSMatthew Ahrens 53394c2d0ebSMatthew Ahrens if (z1->io_bookmark.zb_blkid < z2->io_bookmark.zb_blkid) 53494c2d0ebSMatthew Ahrens return (-1); 53594c2d0ebSMatthew Ahrens if (z1->io_bookmark.zb_blkid > z2->io_bookmark.zb_blkid) 5360f7643c7SGeorge Wilson return (1); 5370f7643c7SGeorge Wilson 5380f7643c7SGeorge Wilson if (z1 < z2) 5390f7643c7SGeorge Wilson return (-1); 5400f7643c7SGeorge Wilson if (z1 > z2) 5410f7643c7SGeorge Wilson return (1); 5420f7643c7SGeorge Wilson 5430f7643c7SGeorge Wilson return (0); 5440f7643c7SGeorge Wilson } 5450f7643c7SGeorge Wilson 546fa9e4066Sahrens /* 547fa9e4066Sahrens * ========================================================================== 548e14bb325SJeff Bonwick * Create the various types of I/O (read, write, free, etc) 549fa9e4066Sahrens * ========================================================================== 550fa9e4066Sahrens */ 551fa9e4066Sahrens static zio_t * 552b24ab676SJeff Bonwick zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp, 553770499e1SDan Kimmel abd_t *data, uint64_t lsize, uint64_t psize, zio_done_func_t *done, 5545602294fSDan Kimmel void *private, zio_type_t type, zio_priority_t priority, 5555602294fSDan Kimmel enum zio_flag flags, vdev_t *vd, uint64_t offset, 5565602294fSDan Kimmel const zbookmark_phys_t *zb, enum zio_stage stage, enum zio_stage pipeline) 557fa9e4066Sahrens { 558fa9e4066Sahrens zio_t *zio; 559fa9e4066Sahrens 5605602294fSDan Kimmel ASSERT3U(psize, <=, SPA_MAXBLOCKSIZE); 5615602294fSDan Kimmel ASSERT(P2PHASE(psize, SPA_MINBLOCKSIZE) == 0); 562e14bb325SJeff Bonwick ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0); 563fa9e4066Sahrens 564e14bb325SJeff Bonwick ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER)); 565e14bb325SJeff Bonwick ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER)); 566e14bb325SJeff Bonwick ASSERT(vd || stage == ZIO_STAGE_OPEN); 567088f3894Sahrens 5685602294fSDan Kimmel IMPLY(lsize != psize, (flags & ZIO_FLAG_RAW) != 0); 5695602294fSDan Kimmel 570ccae0b50Seschrock zio = kmem_cache_alloc(zio_cache, KM_SLEEP); 571ccae0b50Seschrock bzero(zio, sizeof (zio_t)); 572e14bb325SJeff Bonwick 573e14bb325SJeff Bonwick mutex_init(&zio->io_lock, NULL, MUTEX_DEFAULT, NULL); 574e14bb325SJeff Bonwick cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL); 575e14bb325SJeff Bonwick 576a3f829aeSBill Moore list_create(&zio->io_parent_list, sizeof (zio_link_t), 577a3f829aeSBill Moore offsetof(zio_link_t, zl_parent_node)); 578a3f829aeSBill Moore list_create(&zio->io_child_list, sizeof (zio_link_t), 579a3f829aeSBill Moore offsetof(zio_link_t, zl_child_node)); 5808363e80aSGeorge Wilson metaslab_trace_init(&zio->io_alloc_list); 581a3f829aeSBill Moore 582e14bb325SJeff Bonwick if (vd != NULL) 583e14bb325SJeff Bonwick zio->io_child_type = ZIO_CHILD_VDEV; 584e14bb325SJeff Bonwick else if (flags & ZIO_FLAG_GANG_CHILD) 585e14bb325SJeff Bonwick zio->io_child_type = ZIO_CHILD_GANG; 586b24ab676SJeff Bonwick else if (flags & ZIO_FLAG_DDT_CHILD) 587b24ab676SJeff Bonwick zio->io_child_type = ZIO_CHILD_DDT; 588e14bb325SJeff Bonwick else 589e14bb325SJeff Bonwick zio->io_child_type = ZIO_CHILD_LOGICAL; 590e14bb325SJeff Bonwick 591fa9e4066Sahrens if (bp != NULL) { 592b24ab676SJeff Bonwick zio->io_bp = (blkptr_t *)bp; 593fa9e4066Sahrens zio->io_bp_copy = *bp; 594fa9e4066Sahrens zio->io_bp_orig = *bp; 595b24ab676SJeff Bonwick if (type != ZIO_TYPE_WRITE || 596b24ab676SJeff Bonwick zio->io_child_type == ZIO_CHILD_DDT) 597e14bb325SJeff Bonwick zio->io_bp = &zio->io_bp_copy; /* so caller can free */ 598f5383399SBill Moore if (zio->io_child_type == ZIO_CHILD_LOGICAL) 599e14bb325SJeff Bonwick zio->io_logical = zio; 600f5383399SBill Moore if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp)) 601f5383399SBill Moore pipeline |= ZIO_GANG_STAGES; 602fa9e4066Sahrens } 603e14bb325SJeff Bonwick 604e14bb325SJeff Bonwick zio->io_spa = spa; 605e14bb325SJeff Bonwick zio->io_txg = txg; 606fa9e4066Sahrens zio->io_done = done; 607fa9e4066Sahrens zio->io_private = private; 608fa9e4066Sahrens zio->io_type = type; 609fa9e4066Sahrens zio->io_priority = priority; 610e14bb325SJeff Bonwick zio->io_vd = vd; 611e14bb325SJeff Bonwick zio->io_offset = offset; 612770499e1SDan Kimmel zio->io_orig_abd = zio->io_abd = data; 6135602294fSDan Kimmel zio->io_orig_size = zio->io_size = psize; 6145602294fSDan Kimmel zio->io_lsize = lsize; 615e14bb325SJeff Bonwick zio->io_orig_flags = zio->io_flags = flags; 616e14bb325SJeff Bonwick zio->io_orig_stage = zio->io_stage = stage; 617e14bb325SJeff Bonwick zio->io_orig_pipeline = zio->io_pipeline = pipeline; 6180f7643c7SGeorge Wilson zio->io_pipeline_trace = ZIO_STAGE_OPEN; 619fa9e4066Sahrens 620a3f829aeSBill Moore zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY); 621a3f829aeSBill Moore zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE); 622a3f829aeSBill Moore 623e14bb325SJeff Bonwick if (zb != NULL) 624e14bb325SJeff Bonwick zio->io_bookmark = *zb; 625e14bb325SJeff Bonwick 626e14bb325SJeff Bonwick if (pio != NULL) { 627663207adSDon Brady if (zio->io_metaslab_class == NULL) 628663207adSDon Brady zio->io_metaslab_class = pio->io_metaslab_class; 629e14bb325SJeff Bonwick if (zio->io_logical == NULL) 630ea8dc4b6Seschrock zio->io_logical = pio->io_logical; 631f5383399SBill Moore if (zio->io_child_type == ZIO_CHILD_GANG) 632f5383399SBill Moore zio->io_gang_leader = pio->io_gang_leader; 633e14bb325SJeff Bonwick zio_add_child(pio, zio); 634fa9e4066Sahrens } 635fa9e4066Sahrens 636fa9e4066Sahrens return (zio); 637fa9e4066Sahrens } 638fa9e4066Sahrens 6390a4e9518Sgw static void 640e14bb325SJeff Bonwick zio_destroy(zio_t *zio) 6410a4e9518Sgw { 6428363e80aSGeorge Wilson metaslab_trace_fini(&zio->io_alloc_list); 643a3f829aeSBill Moore list_destroy(&zio->io_parent_list); 644a3f829aeSBill Moore list_destroy(&zio->io_child_list); 645e14bb325SJeff Bonwick mutex_destroy(&zio->io_lock); 646e14bb325SJeff Bonwick cv_destroy(&zio->io_cv); 647e14bb325SJeff Bonwick kmem_cache_free(zio_cache, zio); 6480a4e9518Sgw } 6490a4e9518Sgw 650fa9e4066Sahrens zio_t * 651a3f829aeSBill Moore zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done, 652b24ab676SJeff Bonwick void *private, enum zio_flag flags) 653fa9e4066Sahrens { 654fa9e4066Sahrens zio_t *zio; 655fa9e4066Sahrens 6565602294fSDan Kimmel zio = zio_create(pio, spa, 0, NULL, NULL, 0, 0, done, private, 657a3f829aeSBill Moore ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL, 658e14bb325SJeff Bonwick ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE); 659fa9e4066Sahrens 660fa9e4066Sahrens return (zio); 661fa9e4066Sahrens } 662fa9e4066Sahrens 663fa9e4066Sahrens zio_t * 664b24ab676SJeff Bonwick zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags) 665fa9e4066Sahrens { 666a3f829aeSBill Moore return (zio_null(NULL, spa, NULL, done, private, flags)); 667fa9e4066Sahrens } 668fa9e4066Sahrens 669f63ab3d5SMatthew Ahrens void 670f63ab3d5SMatthew Ahrens zfs_blkptr_verify(spa_t *spa, const blkptr_t *bp) 671f63ab3d5SMatthew Ahrens { 672f63ab3d5SMatthew Ahrens if (!DMU_OT_IS_VALID(BP_GET_TYPE(bp))) { 673f63ab3d5SMatthew Ahrens zfs_panic_recover("blkptr at %p has invalid TYPE %llu", 674f63ab3d5SMatthew Ahrens bp, (longlong_t)BP_GET_TYPE(bp)); 675f63ab3d5SMatthew Ahrens } 676f63ab3d5SMatthew Ahrens if (BP_GET_CHECKSUM(bp) >= ZIO_CHECKSUM_FUNCTIONS || 677f63ab3d5SMatthew Ahrens BP_GET_CHECKSUM(bp) <= ZIO_CHECKSUM_ON) { 678f63ab3d5SMatthew Ahrens zfs_panic_recover("blkptr at %p has invalid CHECKSUM %llu", 679f63ab3d5SMatthew Ahrens bp, (longlong_t)BP_GET_CHECKSUM(bp)); 680f63ab3d5SMatthew Ahrens } 681f63ab3d5SMatthew Ahrens if (BP_GET_COMPRESS(bp) >= ZIO_COMPRESS_FUNCTIONS || 682f63ab3d5SMatthew Ahrens BP_GET_COMPRESS(bp) <= ZIO_COMPRESS_ON) { 683f63ab3d5SMatthew Ahrens zfs_panic_recover("blkptr at %p has invalid COMPRESS %llu", 684f63ab3d5SMatthew Ahrens bp, (longlong_t)BP_GET_COMPRESS(bp)); 685f63ab3d5SMatthew Ahrens } 686f63ab3d5SMatthew Ahrens if (BP_GET_LSIZE(bp) > SPA_MAXBLOCKSIZE) { 687f63ab3d5SMatthew Ahrens zfs_panic_recover("blkptr at %p has invalid LSIZE %llu", 688f63ab3d5SMatthew Ahrens bp, (longlong_t)BP_GET_LSIZE(bp)); 689f63ab3d5SMatthew Ahrens } 690f63ab3d5SMatthew Ahrens if (BP_GET_PSIZE(bp) > SPA_MAXBLOCKSIZE) { 691f63ab3d5SMatthew Ahrens zfs_panic_recover("blkptr at %p has invalid PSIZE %llu", 692f63ab3d5SMatthew Ahrens bp, (longlong_t)BP_GET_PSIZE(bp)); 693f63ab3d5SMatthew Ahrens } 694f63ab3d5SMatthew Ahrens 695f63ab3d5SMatthew Ahrens if (BP_IS_EMBEDDED(bp)) { 696f63ab3d5SMatthew Ahrens if (BPE_GET_ETYPE(bp) > NUM_BP_EMBEDDED_TYPES) { 697f63ab3d5SMatthew Ahrens zfs_panic_recover("blkptr at %p has invalid ETYPE %llu", 698f63ab3d5SMatthew Ahrens bp, (longlong_t)BPE_GET_ETYPE(bp)); 699f63ab3d5SMatthew Ahrens } 700f63ab3d5SMatthew Ahrens } 701f63ab3d5SMatthew Ahrens 7026f793812SPavel Zakharov /* 7036f793812SPavel Zakharov * Do not verify individual DVAs if the config is not trusted. This 7046f793812SPavel Zakharov * will be done once the zio is executed in vdev_mirror_map_alloc. 7056f793812SPavel Zakharov */ 7066f793812SPavel Zakharov if (!spa->spa_trust_config) 7076f793812SPavel Zakharov return; 7086f793812SPavel Zakharov 709f63ab3d5SMatthew Ahrens /* 710f63ab3d5SMatthew Ahrens * Pool-specific checks. 711f63ab3d5SMatthew Ahrens * 712f63ab3d5SMatthew Ahrens * Note: it would be nice to verify that the blk_birth and 713f63ab3d5SMatthew Ahrens * BP_PHYSICAL_BIRTH() are not too large. However, spa_freeze() 714f63ab3d5SMatthew Ahrens * allows the birth time of log blocks (and dmu_sync()-ed blocks 715f63ab3d5SMatthew Ahrens * that are in the log) to be arbitrarily large. 716f63ab3d5SMatthew Ahrens */ 717f63ab3d5SMatthew Ahrens for (int i = 0; i < BP_GET_NDVAS(bp); i++) { 718f63ab3d5SMatthew Ahrens uint64_t vdevid = DVA_GET_VDEV(&bp->blk_dva[i]); 719f63ab3d5SMatthew Ahrens if (vdevid >= spa->spa_root_vdev->vdev_children) { 720f63ab3d5SMatthew Ahrens zfs_panic_recover("blkptr at %p DVA %u has invalid " 721f63ab3d5SMatthew Ahrens "VDEV %llu", 722f63ab3d5SMatthew Ahrens bp, i, (longlong_t)vdevid); 7235897eb49SJustin Gibbs continue; 724f63ab3d5SMatthew Ahrens } 725f63ab3d5SMatthew Ahrens vdev_t *vd = spa->spa_root_vdev->vdev_child[vdevid]; 726f63ab3d5SMatthew Ahrens if (vd == NULL) { 727f63ab3d5SMatthew Ahrens zfs_panic_recover("blkptr at %p DVA %u has invalid " 728f63ab3d5SMatthew Ahrens "VDEV %llu", 729f63ab3d5SMatthew Ahrens bp, i, (longlong_t)vdevid); 7305897eb49SJustin Gibbs continue; 731f63ab3d5SMatthew Ahrens } 732f63ab3d5SMatthew Ahrens if (vd->vdev_ops == &vdev_hole_ops) { 733f63ab3d5SMatthew Ahrens zfs_panic_recover("blkptr at %p DVA %u has hole " 734f63ab3d5SMatthew Ahrens "VDEV %llu", 735f63ab3d5SMatthew Ahrens bp, i, (longlong_t)vdevid); 7365897eb49SJustin Gibbs continue; 737f63ab3d5SMatthew Ahrens } 738f63ab3d5SMatthew Ahrens if (vd->vdev_ops == &vdev_missing_ops) { 739f63ab3d5SMatthew Ahrens /* 740f63ab3d5SMatthew Ahrens * "missing" vdevs are valid during import, but we 741f63ab3d5SMatthew Ahrens * don't have their detailed info (e.g. asize), so 742f63ab3d5SMatthew Ahrens * we can't perform any more checks on them. 743f63ab3d5SMatthew Ahrens */ 744f63ab3d5SMatthew Ahrens continue; 745f63ab3d5SMatthew Ahrens } 746f63ab3d5SMatthew Ahrens uint64_t offset = DVA_GET_OFFSET(&bp->blk_dva[i]); 747f63ab3d5SMatthew Ahrens uint64_t asize = DVA_GET_ASIZE(&bp->blk_dva[i]); 748f63ab3d5SMatthew Ahrens if (BP_IS_GANG(bp)) 749f63ab3d5SMatthew Ahrens asize = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE); 750f63ab3d5SMatthew Ahrens if (offset + asize > vd->vdev_asize) { 751f63ab3d5SMatthew Ahrens zfs_panic_recover("blkptr at %p DVA %u has invalid " 752f63ab3d5SMatthew Ahrens "OFFSET %llu", 753f63ab3d5SMatthew Ahrens bp, i, (longlong_t)offset); 754f63ab3d5SMatthew Ahrens } 755f63ab3d5SMatthew Ahrens } 756f63ab3d5SMatthew Ahrens } 757f63ab3d5SMatthew Ahrens 7586f793812SPavel Zakharov boolean_t 7596f793812SPavel Zakharov zfs_dva_valid(spa_t *spa, const dva_t *dva, const blkptr_t *bp) 7606f793812SPavel Zakharov { 7616f793812SPavel Zakharov uint64_t vdevid = DVA_GET_VDEV(dva); 7626f793812SPavel Zakharov 7636f793812SPavel Zakharov if (vdevid >= spa->spa_root_vdev->vdev_children) 7646f793812SPavel Zakharov return (B_FALSE); 7656f793812SPavel Zakharov 7666f793812SPavel Zakharov vdev_t *vd = spa->spa_root_vdev->vdev_child[vdevid]; 7676f793812SPavel Zakharov if (vd == NULL) 7686f793812SPavel Zakharov return (B_FALSE); 7696f793812SPavel Zakharov 7706f793812SPavel Zakharov if (vd->vdev_ops == &vdev_hole_ops) 7716f793812SPavel Zakharov return (B_FALSE); 7726f793812SPavel Zakharov 7736f793812SPavel Zakharov if (vd->vdev_ops == &vdev_missing_ops) { 7746f793812SPavel Zakharov return (B_FALSE); 7756f793812SPavel Zakharov } 7766f793812SPavel Zakharov 7776f793812SPavel Zakharov uint64_t offset = DVA_GET_OFFSET(dva); 7786f793812SPavel Zakharov uint64_t asize = DVA_GET_ASIZE(dva); 7796f793812SPavel Zakharov 7806f793812SPavel Zakharov if (BP_IS_GANG(bp)) 7816f793812SPavel Zakharov asize = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE); 7826f793812SPavel Zakharov if (offset + asize > vd->vdev_asize) 7836f793812SPavel Zakharov return (B_FALSE); 7846f793812SPavel Zakharov 7856f793812SPavel Zakharov return (B_TRUE); 7866f793812SPavel Zakharov } 7876f793812SPavel Zakharov 788fa9e4066Sahrens zio_t * 789e14bb325SJeff Bonwick zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, 790770499e1SDan Kimmel abd_t *data, uint64_t size, zio_done_func_t *done, void *private, 7917802d7bfSMatthew Ahrens zio_priority_t priority, enum zio_flag flags, const zbookmark_phys_t *zb) 792fa9e4066Sahrens { 793fa9e4066Sahrens zio_t *zio; 794fa9e4066Sahrens 795f63ab3d5SMatthew Ahrens zfs_blkptr_verify(spa, bp); 796f63ab3d5SMatthew Ahrens 797b24ab676SJeff Bonwick zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp, 7985602294fSDan Kimmel data, size, size, done, private, 799e14bb325SJeff Bonwick ZIO_TYPE_READ, priority, flags, NULL, 0, zb, 800b24ab676SJeff Bonwick ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ? 801b24ab676SJeff Bonwick ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE); 802fa9e4066Sahrens 803fa9e4066Sahrens return (zio); 804fa9e4066Sahrens } 805fa9e4066Sahrens 806fa9e4066Sahrens zio_t * 807e14bb325SJeff Bonwick zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, 808770499e1SDan Kimmel abd_t *data, uint64_t lsize, uint64_t psize, const zio_prop_t *zp, 8098df0bcf0SPaul Dagnelie zio_done_func_t *ready, zio_done_func_t *children_ready, 8108df0bcf0SPaul Dagnelie zio_done_func_t *physdone, zio_done_func_t *done, 8118df0bcf0SPaul Dagnelie void *private, zio_priority_t priority, enum zio_flag flags, 8128df0bcf0SPaul Dagnelie const zbookmark_phys_t *zb) 813fa9e4066Sahrens { 814fa9e4066Sahrens zio_t *zio; 815fa9e4066Sahrens 816e14bb325SJeff Bonwick ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF && 817e14bb325SJeff Bonwick zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS && 818e14bb325SJeff Bonwick zp->zp_compress >= ZIO_COMPRESS_OFF && 819e14bb325SJeff Bonwick zp->zp_compress < ZIO_COMPRESS_FUNCTIONS && 820ad135b5dSChristopher Siden DMU_OT_IS_VALID(zp->zp_type) && 821e14bb325SJeff Bonwick zp->zp_level < 32 && 822b24ab676SJeff Bonwick zp->zp_copies > 0 && 82380901aeaSGeorge Wilson zp->zp_copies <= spa_max_replication(spa)); 8240a4e9518Sgw 8255602294fSDan Kimmel zio = zio_create(pio, spa, txg, bp, data, lsize, psize, done, private, 826e14bb325SJeff Bonwick ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb, 827b24ab676SJeff Bonwick ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ? 828b24ab676SJeff Bonwick ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE); 829fa9e4066Sahrens 830c717a561Smaybee zio->io_ready = ready; 8318df0bcf0SPaul Dagnelie zio->io_children_ready = children_ready; 83269962b56SMatthew Ahrens zio->io_physdone = physdone; 833e14bb325SJeff Bonwick zio->io_prop = *zp; 834fa9e4066Sahrens 8355d7b4d43SMatthew Ahrens /* 8365d7b4d43SMatthew Ahrens * Data can be NULL if we are going to call zio_write_override() to 8375d7b4d43SMatthew Ahrens * provide the already-allocated BP. But we may need the data to 8385d7b4d43SMatthew Ahrens * verify a dedup hit (if requested). In this case, don't try to 8395d7b4d43SMatthew Ahrens * dedup (just take the already-allocated BP verbatim). 8405d7b4d43SMatthew Ahrens */ 8415d7b4d43SMatthew Ahrens if (data == NULL && zio->io_prop.zp_dedup_verify) { 8425d7b4d43SMatthew Ahrens zio->io_prop.zp_dedup = zio->io_prop.zp_dedup_verify = B_FALSE; 8435d7b4d43SMatthew Ahrens } 8445d7b4d43SMatthew Ahrens 845fa9e4066Sahrens return (zio); 846fa9e4066Sahrens } 847fa9e4066Sahrens 848fa9e4066Sahrens zio_t * 849770499e1SDan Kimmel zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, abd_t *data, 85069962b56SMatthew Ahrens uint64_t size, zio_done_func_t *done, void *private, 8517802d7bfSMatthew Ahrens zio_priority_t priority, enum zio_flag flags, zbookmark_phys_t *zb) 852fa9e4066Sahrens { 853fa9e4066Sahrens zio_t *zio; 854fa9e4066Sahrens 8555602294fSDan Kimmel zio = zio_create(pio, spa, txg, bp, data, size, size, done, private, 8560f7643c7SGeorge Wilson ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_IO_REWRITE, NULL, 0, zb, 857e14bb325SJeff Bonwick ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE); 858fa9e4066Sahrens 859fa9e4066Sahrens return (zio); 860fa9e4066Sahrens } 861fa9e4066Sahrens 862b24ab676SJeff Bonwick void 86380901aeaSGeorge Wilson zio_write_override(zio_t *zio, blkptr_t *bp, int copies, boolean_t nopwrite) 864b24ab676SJeff Bonwick { 865b24ab676SJeff Bonwick ASSERT(zio->io_type == ZIO_TYPE_WRITE); 866b24ab676SJeff Bonwick ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 867b24ab676SJeff Bonwick ASSERT(zio->io_stage == ZIO_STAGE_OPEN); 868b24ab676SJeff Bonwick ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa)); 869b24ab676SJeff Bonwick 87080901aeaSGeorge Wilson /* 87180901aeaSGeorge Wilson * We must reset the io_prop to match the values that existed 87280901aeaSGeorge Wilson * when the bp was first written by dmu_sync() keeping in mind 87380901aeaSGeorge Wilson * that nopwrite and dedup are mutually exclusive. 87480901aeaSGeorge Wilson */ 87580901aeaSGeorge Wilson zio->io_prop.zp_dedup = nopwrite ? B_FALSE : zio->io_prop.zp_dedup; 87680901aeaSGeorge Wilson zio->io_prop.zp_nopwrite = nopwrite; 877b24ab676SJeff Bonwick zio->io_prop.zp_copies = copies; 878b24ab676SJeff Bonwick zio->io_bp_override = bp; 879b24ab676SJeff Bonwick } 880b24ab676SJeff Bonwick 881b24ab676SJeff Bonwick void 882b24ab676SJeff Bonwick zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp) 883b24ab676SJeff Bonwick { 8845d7b4d43SMatthew Ahrens 8855cabbc6bSPrashanth Sreenivasa zfs_blkptr_verify(spa, bp); 8865cabbc6bSPrashanth Sreenivasa 8875d7b4d43SMatthew Ahrens /* 8885d7b4d43SMatthew Ahrens * The check for EMBEDDED is a performance optimization. We 8895d7b4d43SMatthew Ahrens * process the free here (by ignoring it) rather than 8905d7b4d43SMatthew Ahrens * putting it on the list and then processing it in zio_free_sync(). 8915d7b4d43SMatthew Ahrens */ 8925d7b4d43SMatthew Ahrens if (BP_IS_EMBEDDED(bp)) 8935d7b4d43SMatthew Ahrens return; 8943b2aab18SMatthew Ahrens metaslab_check_free(spa, bp); 8959cb154a3SMatthew Ahrens 8969cb154a3SMatthew Ahrens /* 8979cb154a3SMatthew Ahrens * Frees that are for the currently-syncing txg, are not going to be 8989cb154a3SMatthew Ahrens * deferred, and which will not need to do a read (i.e. not GANG or 8999cb154a3SMatthew Ahrens * DEDUP), can be processed immediately. Otherwise, put them on the 9009cb154a3SMatthew Ahrens * in-memory list for later processing. 9019cb154a3SMatthew Ahrens */ 9029cb154a3SMatthew Ahrens if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp) || 9039cb154a3SMatthew Ahrens txg != spa->spa_syncing_txg || 9049cb154a3SMatthew Ahrens spa_sync_pass(spa) >= zfs_sync_pass_deferred_free) { 9059cb154a3SMatthew Ahrens bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp); 9069cb154a3SMatthew Ahrens } else { 9079cb154a3SMatthew Ahrens VERIFY0(zio_wait(zio_free_sync(NULL, spa, txg, bp, 0))); 9089cb154a3SMatthew Ahrens } 909b24ab676SJeff Bonwick } 910b24ab676SJeff Bonwick 911fa9e4066Sahrens zio_t * 912b24ab676SJeff Bonwick zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp, 913b24ab676SJeff Bonwick enum zio_flag flags) 914fa9e4066Sahrens { 915fa9e4066Sahrens zio_t *zio; 9169cb154a3SMatthew Ahrens enum zio_stage stage = ZIO_FREE_PIPELINE; 917fa9e4066Sahrens 918fa9e4066Sahrens ASSERT(!BP_IS_HOLE(bp)); 919b24ab676SJeff Bonwick ASSERT(spa_syncing_txg(spa) == txg); 92001f55e48SGeorge Wilson ASSERT(spa_sync_pass(spa) < zfs_sync_pass_deferred_free); 921fa9e4066Sahrens 9225d7b4d43SMatthew Ahrens if (BP_IS_EMBEDDED(bp)) 9235d7b4d43SMatthew Ahrens return (zio_null(pio, spa, NULL, NULL, NULL, 0)); 9245d7b4d43SMatthew Ahrens 9253b2aab18SMatthew Ahrens metaslab_check_free(spa, bp); 9266e6d5868SMatthew Ahrens arc_freed(spa, bp); 927*a3874b8bSToomas Soome dsl_scan_freed(spa, bp); 9283b2aab18SMatthew Ahrens 9299cb154a3SMatthew Ahrens /* 9309cb154a3SMatthew Ahrens * GANG and DEDUP blocks can induce a read (for the gang block header, 9319cb154a3SMatthew Ahrens * or the DDT), so issue them asynchronously so that this thread is 9329cb154a3SMatthew Ahrens * not tied up. 9339cb154a3SMatthew Ahrens */ 9349cb154a3SMatthew Ahrens if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp)) 9359cb154a3SMatthew Ahrens stage |= ZIO_STAGE_ISSUE_ASYNC; 9369cb154a3SMatthew Ahrens 937e14bb325SJeff Bonwick zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp), 9385602294fSDan Kimmel BP_GET_PSIZE(bp), NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_NOW, 9395602294fSDan Kimmel flags, NULL, 0, NULL, ZIO_STAGE_OPEN, stage); 9409cb154a3SMatthew Ahrens 941fa9e4066Sahrens return (zio); 942fa9e4066Sahrens } 943fa9e4066Sahrens 944fa9e4066Sahrens zio_t * 945b24ab676SJeff Bonwick zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp, 946b24ab676SJeff Bonwick zio_done_func_t *done, void *private, enum zio_flag flags) 947fa9e4066Sahrens { 948fa9e4066Sahrens zio_t *zio; 949fa9e4066Sahrens 9505cabbc6bSPrashanth Sreenivasa zfs_blkptr_verify(spa, bp); 9515d7b4d43SMatthew Ahrens 9525d7b4d43SMatthew Ahrens if (BP_IS_EMBEDDED(bp)) 9535d7b4d43SMatthew Ahrens return (zio_null(pio, spa, NULL, NULL, NULL, 0)); 9545d7b4d43SMatthew Ahrens 955fa9e4066Sahrens /* 956fa9e4066Sahrens * A claim is an allocation of a specific block. Claims are needed 957fa9e4066Sahrens * to support immediate writes in the intent log. The issue is that 958fa9e4066Sahrens * immediate writes contain committed data, but in a txg that was 959fa9e4066Sahrens * *not* committed. Upon opening the pool after an unclean shutdown, 960fa9e4066Sahrens * the intent log claims all blocks that contain immediate write data 961fa9e4066Sahrens * so that the SPA knows they're in use. 962fa9e4066Sahrens * 963fa9e4066Sahrens * All claims *must* be resolved in the first txg -- before the SPA 964fa9e4066Sahrens * starts allocating blocks -- so that nothing is allocated twice. 965b24ab676SJeff Bonwick * If txg == 0 we just verify that the block is claimable. 966fa9e4066Sahrens */ 96786714001SSerapheim Dimitropoulos ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, 96886714001SSerapheim Dimitropoulos spa_min_claim_txg(spa)); 96986714001SSerapheim Dimitropoulos ASSERT(txg == spa_min_claim_txg(spa) || txg == 0); 970b24ab676SJeff Bonwick ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(1M) */ 971fa9e4066Sahrens 972e14bb325SJeff Bonwick zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp), 9735602294fSDan Kimmel BP_GET_PSIZE(bp), done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, 9745602294fSDan Kimmel flags, NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE); 9750f7643c7SGeorge Wilson ASSERT0(zio->io_queued_timestamp); 976fa9e4066Sahrens 977fa9e4066Sahrens return (zio); 978fa9e4066Sahrens } 979fa9e4066Sahrens 980fa9e4066Sahrens zio_t * 981fa9e4066Sahrens zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd, 98269962b56SMatthew Ahrens zio_done_func_t *done, void *private, enum zio_flag flags) 983fa9e4066Sahrens { 984fa9e4066Sahrens zio_t *zio; 985fa9e4066Sahrens int c; 986fa9e4066Sahrens 987fa9e4066Sahrens if (vd->vdev_children == 0) { 9885602294fSDan Kimmel zio = zio_create(pio, spa, 0, NULL, NULL, 0, 0, done, private, 98969962b56SMatthew Ahrens ZIO_TYPE_IOCTL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL, 990fa9e4066Sahrens ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE); 991fa9e4066Sahrens 992fa9e4066Sahrens zio->io_cmd = cmd; 993fa9e4066Sahrens } else { 994a3f829aeSBill Moore zio = zio_null(pio, spa, NULL, NULL, NULL, flags); 995fa9e4066Sahrens 996fa9e4066Sahrens for (c = 0; c < vd->vdev_children; c++) 997fa9e4066Sahrens zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd, 99869962b56SMatthew Ahrens done, private, flags)); 999fa9e4066Sahrens } 1000fa9e4066Sahrens 1001fa9e4066Sahrens return (zio); 1002fa9e4066Sahrens } 1003fa9e4066Sahrens 1004fa9e4066Sahrens zio_t * 1005fa9e4066Sahrens zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size, 1006770499e1SDan Kimmel abd_t *data, int checksum, zio_done_func_t *done, void *private, 100769962b56SMatthew Ahrens zio_priority_t priority, enum zio_flag flags, boolean_t labels) 1008fa9e4066Sahrens { 1009fa9e4066Sahrens zio_t *zio; 10100a4e9518Sgw 1011e14bb325SJeff Bonwick ASSERT(vd->vdev_children == 0); 1012e14bb325SJeff Bonwick ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE || 1013e14bb325SJeff Bonwick offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE); 1014e14bb325SJeff Bonwick ASSERT3U(offset + size, <=, vd->vdev_psize); 1015fa9e4066Sahrens 10165602294fSDan Kimmel zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, size, done, 10175602294fSDan Kimmel private, ZIO_TYPE_READ, priority, flags | ZIO_FLAG_PHYSICAL, vd, 10185602294fSDan Kimmel offset, NULL, ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE); 1019fa9e4066Sahrens 1020e14bb325SJeff Bonwick zio->io_prop.zp_checksum = checksum; 1021fa9e4066Sahrens 1022fa9e4066Sahrens return (zio); 1023fa9e4066Sahrens } 1024fa9e4066Sahrens 1025fa9e4066Sahrens zio_t * 1026fa9e4066Sahrens zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size, 1027770499e1SDan Kimmel abd_t *data, int checksum, zio_done_func_t *done, void *private, 102869962b56SMatthew Ahrens zio_priority_t priority, enum zio_flag flags, boolean_t labels) 1029fa9e4066Sahrens { 1030fa9e4066Sahrens zio_t *zio; 10310a4e9518Sgw 1032e14bb325SJeff Bonwick ASSERT(vd->vdev_children == 0); 1033e14bb325SJeff Bonwick ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE || 1034e14bb325SJeff Bonwick offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE); 1035e14bb325SJeff Bonwick ASSERT3U(offset + size, <=, vd->vdev_psize); 1036fa9e4066Sahrens 10375602294fSDan Kimmel zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, size, done, 10385602294fSDan Kimmel private, ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_PHYSICAL, vd, 10395602294fSDan Kimmel offset, NULL, ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE); 1040fa9e4066Sahrens 1041e14bb325SJeff Bonwick zio->io_prop.zp_checksum = checksum; 1042fa9e4066Sahrens 104345818ee1SMatthew Ahrens if (zio_checksum_table[checksum].ci_flags & ZCHECKSUM_FLAG_EMBEDDED) { 1044fa9e4066Sahrens /* 10456e1f5caaSNeil Perrin * zec checksums are necessarily destructive -- they modify 1046e14bb325SJeff Bonwick * the end of the write buffer to hold the verifier/checksum. 1047fa9e4066Sahrens * Therefore, we must make a local copy in case the data is 1048e14bb325SJeff Bonwick * being written to multiple places in parallel. 1049fa9e4066Sahrens */ 1050770499e1SDan Kimmel abd_t *wbuf = abd_alloc_sametype(data, size); 1051770499e1SDan Kimmel abd_copy(wbuf, data, size); 1052770499e1SDan Kimmel 1053e14bb325SJeff Bonwick zio_push_transform(zio, wbuf, size, size, NULL); 1054fa9e4066Sahrens } 1055fa9e4066Sahrens 1056fa9e4066Sahrens return (zio); 1057fa9e4066Sahrens } 1058fa9e4066Sahrens 1059fa9e4066Sahrens /* 1060e14bb325SJeff Bonwick * Create a child I/O to do some work for us. 1061fa9e4066Sahrens */ 1062fa9e4066Sahrens zio_t * 1063e14bb325SJeff Bonwick zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset, 1064770499e1SDan Kimmel abd_t *data, uint64_t size, int type, zio_priority_t priority, 1065dcbf3bd6SGeorge Wilson enum zio_flag flags, zio_done_func_t *done, void *private) 1066fa9e4066Sahrens { 1067b24ab676SJeff Bonwick enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE; 1068e14bb325SJeff Bonwick zio_t *zio; 1069e14bb325SJeff Bonwick 10705cabbc6bSPrashanth Sreenivasa /* 10715cabbc6bSPrashanth Sreenivasa * vdev child I/Os do not propagate their error to the parent. 10725cabbc6bSPrashanth Sreenivasa * Therefore, for correct operation the caller *must* check for 10735cabbc6bSPrashanth Sreenivasa * and handle the error in the child i/o's done callback. 10745cabbc6bSPrashanth Sreenivasa * The only exceptions are i/os that we don't care about 10755cabbc6bSPrashanth Sreenivasa * (OPTIONAL or REPAIR). 10765cabbc6bSPrashanth Sreenivasa */ 10775cabbc6bSPrashanth Sreenivasa ASSERT((flags & ZIO_FLAG_OPTIONAL) || (flags & ZIO_FLAG_IO_REPAIR) || 10785cabbc6bSPrashanth Sreenivasa done != NULL); 10795cabbc6bSPrashanth Sreenivasa 1080fa9e4066Sahrens if (type == ZIO_TYPE_READ && bp != NULL) { 1081fa9e4066Sahrens /* 1082fa9e4066Sahrens * If we have the bp, then the child should perform the 1083fa9e4066Sahrens * checksum and the parent need not. This pushes error 1084fa9e4066Sahrens * detection as close to the leaves as possible and 1085fa9e4066Sahrens * eliminates redundant checksums in the interior nodes. 1086fa9e4066Sahrens */ 1087b24ab676SJeff Bonwick pipeline |= ZIO_STAGE_CHECKSUM_VERIFY; 1088b24ab676SJeff Bonwick pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY; 1089fa9e4066Sahrens } 1090fa9e4066Sahrens 10915cabbc6bSPrashanth Sreenivasa if (vd->vdev_ops->vdev_op_leaf) { 10925cabbc6bSPrashanth Sreenivasa ASSERT0(vd->vdev_children); 1093e14bb325SJeff Bonwick offset += VDEV_LABEL_START_SIZE; 10945cabbc6bSPrashanth Sreenivasa } 1095e14bb325SJeff Bonwick 10965cabbc6bSPrashanth Sreenivasa flags |= ZIO_VDEV_CHILD_FLAGS(pio); 1097b24ab676SJeff Bonwick 1098b24ab676SJeff Bonwick /* 1099b24ab676SJeff Bonwick * If we've decided to do a repair, the write is not speculative -- 1100b24ab676SJeff Bonwick * even if the original read was. 1101b24ab676SJeff Bonwick */ 1102b24ab676SJeff Bonwick if (flags & ZIO_FLAG_IO_REPAIR) 1103b24ab676SJeff Bonwick flags &= ~ZIO_FLAG_SPECULATIVE; 1104b24ab676SJeff Bonwick 11050f7643c7SGeorge Wilson /* 11060f7643c7SGeorge Wilson * If we're creating a child I/O that is not associated with a 11070f7643c7SGeorge Wilson * top-level vdev, then the child zio is not an allocating I/O. 11080f7643c7SGeorge Wilson * If this is a retried I/O then we ignore it since we will 11090f7643c7SGeorge Wilson * have already processed the original allocating I/O. 11100f7643c7SGeorge Wilson */ 11110f7643c7SGeorge Wilson if (flags & ZIO_FLAG_IO_ALLOCATING && 11120f7643c7SGeorge Wilson (vd != vd->vdev_top || (flags & ZIO_FLAG_IO_RETRY))) { 1113663207adSDon Brady ASSERT(pio->io_metaslab_class != NULL); 1114663207adSDon Brady ASSERT(pio->io_metaslab_class->mc_alloc_throttle_enabled); 11150f7643c7SGeorge Wilson ASSERT(type == ZIO_TYPE_WRITE); 11160f7643c7SGeorge Wilson ASSERT(priority == ZIO_PRIORITY_ASYNC_WRITE); 11170f7643c7SGeorge Wilson ASSERT(!(flags & ZIO_FLAG_IO_REPAIR)); 11180f7643c7SGeorge Wilson ASSERT(!(pio->io_flags & ZIO_FLAG_IO_REWRITE) || 11190f7643c7SGeorge Wilson pio->io_child_type == ZIO_CHILD_GANG); 11200f7643c7SGeorge Wilson 11210f7643c7SGeorge Wilson flags &= ~ZIO_FLAG_IO_ALLOCATING; 11220f7643c7SGeorge Wilson } 11230f7643c7SGeorge Wilson 11245602294fSDan Kimmel zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size, size, 1125b24ab676SJeff Bonwick done, private, type, priority, flags, vd, offset, &pio->io_bookmark, 1126b24ab676SJeff Bonwick ZIO_STAGE_VDEV_IO_START >> 1, pipeline); 11270f7643c7SGeorge Wilson ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_VDEV); 1128fa9e4066Sahrens 112969962b56SMatthew Ahrens zio->io_physdone = pio->io_physdone; 113069962b56SMatthew Ahrens if (vd->vdev_ops->vdev_op_leaf && zio->io_logical != NULL) 113169962b56SMatthew Ahrens zio->io_logical->io_phys_children++; 113269962b56SMatthew Ahrens 1133e14bb325SJeff Bonwick return (zio); 113432b87932Sek } 113532b87932Sek 1136e14bb325SJeff Bonwick zio_t * 1137770499e1SDan Kimmel zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, abd_t *data, uint64_t size, 11383a4b1be9SMatthew Ahrens zio_type_t type, zio_priority_t priority, enum zio_flag flags, 11399a686fbcSPaul Dagnelie zio_done_func_t *done, void *private) 1140fa9e4066Sahrens { 1141e14bb325SJeff Bonwick zio_t *zio; 1142fa9e4066Sahrens 1143e14bb325SJeff Bonwick ASSERT(vd->vdev_ops->vdev_op_leaf); 1144fa9e4066Sahrens 1145e14bb325SJeff Bonwick zio = zio_create(NULL, vd->vdev_spa, 0, NULL, 11465602294fSDan Kimmel data, size, size, done, private, type, priority, 114769962b56SMatthew Ahrens flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY | ZIO_FLAG_DELEGATED, 1148e14bb325SJeff Bonwick vd, offset, NULL, 1149b24ab676SJeff Bonwick ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE); 1150fa9e4066Sahrens 1151e14bb325SJeff Bonwick return (zio); 1152e05725b1Sbonwick } 1153e05725b1Sbonwick 1154e05725b1Sbonwick void 1155e14bb325SJeff Bonwick zio_flush(zio_t *zio, vdev_t *vd) 1156e05725b1Sbonwick { 1157e14bb325SJeff Bonwick zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE, 115869962b56SMatthew Ahrens NULL, NULL, 1159e14bb325SJeff Bonwick ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY)); 1160fa9e4066Sahrens } 1161fa9e4066Sahrens 11626e1f5caaSNeil Perrin void 11636e1f5caaSNeil Perrin zio_shrink(zio_t *zio, uint64_t size) 11646e1f5caaSNeil Perrin { 11651271e4b1SPrakash Surya ASSERT3P(zio->io_executor, ==, NULL); 11661271e4b1SPrakash Surya ASSERT3P(zio->io_orig_size, ==, zio->io_size); 11671271e4b1SPrakash Surya ASSERT3U(size, <=, zio->io_size); 11686e1f5caaSNeil Perrin 11696e1f5caaSNeil Perrin /* 11706e1f5caaSNeil Perrin * We don't shrink for raidz because of problems with the 11716e1f5caaSNeil Perrin * reconstruction when reading back less than the block size. 11726e1f5caaSNeil Perrin * Note, BP_IS_RAIDZ() assumes no compression. 11736e1f5caaSNeil Perrin */ 11746e1f5caaSNeil Perrin ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF); 11755602294fSDan Kimmel if (!BP_IS_RAIDZ(zio->io_bp)) { 11765602294fSDan Kimmel /* we are not doing a raw write */ 11775602294fSDan Kimmel ASSERT3U(zio->io_size, ==, zio->io_lsize); 11785602294fSDan Kimmel zio->io_orig_size = zio->io_size = zio->io_lsize = size; 11795602294fSDan Kimmel } 11806e1f5caaSNeil Perrin } 11816e1f5caaSNeil Perrin 1182fa9e4066Sahrens /* 1183fa9e4066Sahrens * ========================================================================== 1184e14bb325SJeff Bonwick * Prepare to read and write logical blocks 1185fa9e4066Sahrens * ========================================================================== 1186fa9e4066Sahrens */ 1187e14bb325SJeff Bonwick 1188e05725b1Sbonwick static int 1189e14bb325SJeff Bonwick zio_read_bp_init(zio_t *zio) 1190fa9e4066Sahrens { 1191e14bb325SJeff Bonwick blkptr_t *bp = zio->io_bp; 1192e05725b1Sbonwick 11935cabbc6bSPrashanth Sreenivasa ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy); 11945cabbc6bSPrashanth Sreenivasa 119503361682SJeff Bonwick if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF && 1196f5383399SBill Moore zio->io_child_type == ZIO_CHILD_LOGICAL && 1197f5383399SBill Moore !(zio->io_flags & ZIO_FLAG_RAW)) { 11985d7b4d43SMatthew Ahrens uint64_t psize = 11995d7b4d43SMatthew Ahrens BP_IS_EMBEDDED(bp) ? BPE_GET_PSIZE(bp) : BP_GET_PSIZE(bp); 1200770499e1SDan Kimmel zio_push_transform(zio, abd_alloc_sametype(zio->io_abd, psize), 1201770499e1SDan Kimmel psize, psize, zio_decompress); 1202e14bb325SJeff Bonwick } 1203fa9e4066Sahrens 12045d7b4d43SMatthew Ahrens if (BP_IS_EMBEDDED(bp) && BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA) { 12055d7b4d43SMatthew Ahrens zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 1206770499e1SDan Kimmel 1207770499e1SDan Kimmel int psize = BPE_GET_PSIZE(bp); 1208770499e1SDan Kimmel void *data = abd_borrow_buf(zio->io_abd, psize); 1209770499e1SDan Kimmel decode_embedded_bp_compressed(bp, data); 1210770499e1SDan Kimmel abd_return_buf_copy(zio->io_abd, data, psize); 12115d7b4d43SMatthew Ahrens } else { 12125d7b4d43SMatthew Ahrens ASSERT(!BP_IS_EMBEDDED(bp)); 12135cabbc6bSPrashanth Sreenivasa ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy); 12145d7b4d43SMatthew Ahrens } 12155d7b4d43SMatthew Ahrens 1216ad135b5dSChristopher Siden if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp)) && BP_GET_LEVEL(bp) == 0) 1217e14bb325SJeff Bonwick zio->io_flags |= ZIO_FLAG_DONT_CACHE; 1218fa9e4066Sahrens 1219bbfd46c4SJeff Bonwick if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP) 1220bbfd46c4SJeff Bonwick zio->io_flags |= ZIO_FLAG_DONT_CACHE; 1221bbfd46c4SJeff Bonwick 1222b24ab676SJeff Bonwick if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL) 1223b24ab676SJeff Bonwick zio->io_pipeline = ZIO_DDT_READ_PIPELINE; 1224b24ab676SJeff Bonwick 1225e14bb325SJeff Bonwick return (ZIO_PIPELINE_CONTINUE); 1226fa9e4066Sahrens } 1227fa9e4066Sahrens 1228e05725b1Sbonwick static int 1229e14bb325SJeff Bonwick zio_write_bp_init(zio_t *zio) 12300a4e9518Sgw { 1231e14bb325SJeff Bonwick if (!IO_IS_ALLOCATING(zio)) 1232e14bb325SJeff Bonwick return (ZIO_PIPELINE_CONTINUE); 12330a4e9518Sgw 1234b24ab676SJeff Bonwick ASSERT(zio->io_child_type != ZIO_CHILD_DDT); 1235b24ab676SJeff Bonwick 1236b24ab676SJeff Bonwick if (zio->io_bp_override) { 12370f7643c7SGeorge Wilson blkptr_t *bp = zio->io_bp; 12380f7643c7SGeorge Wilson zio_prop_t *zp = &zio->io_prop; 12390f7643c7SGeorge Wilson 1240b24ab676SJeff Bonwick ASSERT(bp->blk_birth != zio->io_txg); 1241b24ab676SJeff Bonwick ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0); 1242b24ab676SJeff Bonwick 1243b24ab676SJeff Bonwick *bp = *zio->io_bp_override; 1244b24ab676SJeff Bonwick zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 1245b24ab676SJeff Bonwick 12465d7b4d43SMatthew Ahrens if (BP_IS_EMBEDDED(bp)) 12475d7b4d43SMatthew Ahrens return (ZIO_PIPELINE_CONTINUE); 12485d7b4d43SMatthew Ahrens 124980901aeaSGeorge Wilson /* 125080901aeaSGeorge Wilson * If we've been overridden and nopwrite is set then 125180901aeaSGeorge Wilson * set the flag accordingly to indicate that a nopwrite 125280901aeaSGeorge Wilson * has already occurred. 125380901aeaSGeorge Wilson */ 125480901aeaSGeorge Wilson if (!BP_IS_HOLE(bp) && zp->zp_nopwrite) { 125580901aeaSGeorge Wilson ASSERT(!zp->zp_dedup); 12560f7643c7SGeorge Wilson ASSERT3U(BP_GET_CHECKSUM(bp), ==, zp->zp_checksum); 125780901aeaSGeorge Wilson zio->io_flags |= ZIO_FLAG_NOPWRITE; 125880901aeaSGeorge Wilson return (ZIO_PIPELINE_CONTINUE); 125980901aeaSGeorge Wilson } 126080901aeaSGeorge Wilson 126180901aeaSGeorge Wilson ASSERT(!zp->zp_nopwrite); 126280901aeaSGeorge Wilson 1263b24ab676SJeff Bonwick if (BP_IS_HOLE(bp) || !zp->zp_dedup) 1264b24ab676SJeff Bonwick return (ZIO_PIPELINE_CONTINUE); 1265b24ab676SJeff Bonwick 126645818ee1SMatthew Ahrens ASSERT((zio_checksum_table[zp->zp_checksum].ci_flags & 126745818ee1SMatthew Ahrens ZCHECKSUM_FLAG_DEDUP) || zp->zp_dedup_verify); 1268b24ab676SJeff Bonwick 1269b24ab676SJeff Bonwick if (BP_GET_CHECKSUM(bp) == zp->zp_checksum) { 1270b24ab676SJeff Bonwick BP_SET_DEDUP(bp, 1); 1271b24ab676SJeff Bonwick zio->io_pipeline |= ZIO_STAGE_DDT_WRITE; 1272b24ab676SJeff Bonwick return (ZIO_PIPELINE_CONTINUE); 1273b24ab676SJeff Bonwick } 12740f7643c7SGeorge Wilson 12750f7643c7SGeorge Wilson /* 12760f7643c7SGeorge Wilson * We were unable to handle this as an override bp, treat 12770f7643c7SGeorge Wilson * it as a regular write I/O. 12780f7643c7SGeorge Wilson */ 1279b39b744bSMatthew Ahrens zio->io_bp_override = NULL; 12800f7643c7SGeorge Wilson *bp = zio->io_bp_orig; 12810f7643c7SGeorge Wilson zio->io_pipeline = zio->io_orig_pipeline; 1282b24ab676SJeff Bonwick } 12830a4e9518Sgw 12840f7643c7SGeorge Wilson return (ZIO_PIPELINE_CONTINUE); 12850f7643c7SGeorge Wilson } 12860f7643c7SGeorge Wilson 12870f7643c7SGeorge Wilson static int 12880f7643c7SGeorge Wilson zio_write_compress(zio_t *zio) 12890f7643c7SGeorge Wilson { 12900f7643c7SGeorge Wilson spa_t *spa = zio->io_spa; 12910f7643c7SGeorge Wilson zio_prop_t *zp = &zio->io_prop; 12920f7643c7SGeorge Wilson enum zio_compress compress = zp->zp_compress; 12930f7643c7SGeorge Wilson blkptr_t *bp = zio->io_bp; 12945602294fSDan Kimmel uint64_t lsize = zio->io_lsize; 12955602294fSDan Kimmel uint64_t psize = zio->io_size; 12960f7643c7SGeorge Wilson int pass = 1; 12970f7643c7SGeorge Wilson 12985602294fSDan Kimmel EQUIV(lsize != psize, (zio->io_flags & ZIO_FLAG_RAW) != 0); 12995602294fSDan Kimmel 13000f7643c7SGeorge Wilson /* 13010f7643c7SGeorge Wilson * If our children haven't all reached the ready stage, 13020f7643c7SGeorge Wilson * wait for them and then repeat this pipeline stage. 13030f7643c7SGeorge Wilson */ 1304d6e1c446SGeorge Wilson if (zio_wait_for_children(zio, ZIO_CHILD_LOGICAL_BIT | 1305d6e1c446SGeorge Wilson ZIO_CHILD_GANG_BIT, ZIO_WAIT_READY)) { 13060f7643c7SGeorge Wilson return (ZIO_PIPELINE_STOP); 1307d6e1c446SGeorge Wilson } 13080f7643c7SGeorge Wilson 13090f7643c7SGeorge Wilson if (!IO_IS_ALLOCATING(zio)) 13100f7643c7SGeorge Wilson return (ZIO_PIPELINE_CONTINUE); 13110f7643c7SGeorge Wilson 13120f7643c7SGeorge Wilson if (zio->io_children_ready != NULL) { 13130f7643c7SGeorge Wilson /* 13140f7643c7SGeorge Wilson * Now that all our children are ready, run the callback 13150f7643c7SGeorge Wilson * associated with this zio in case it wants to modify the 13160f7643c7SGeorge Wilson * data to be written. 13170f7643c7SGeorge Wilson */ 13180f7643c7SGeorge Wilson ASSERT3U(zp->zp_level, >, 0); 13190f7643c7SGeorge Wilson zio->io_children_ready(zio); 13200f7643c7SGeorge Wilson } 13210f7643c7SGeorge Wilson 13220f7643c7SGeorge Wilson ASSERT(zio->io_child_type != ZIO_CHILD_DDT); 13230f7643c7SGeorge Wilson ASSERT(zio->io_bp_override == NULL); 13240f7643c7SGeorge Wilson 132543466aaeSMax Grossman if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg) { 1326e14bb325SJeff Bonwick /* 1327e14bb325SJeff Bonwick * We're rewriting an existing block, which means we're 1328e14bb325SJeff Bonwick * working on behalf of spa_sync(). For spa_sync() to 1329e14bb325SJeff Bonwick * converge, it must eventually be the case that we don't 1330e14bb325SJeff Bonwick * have to allocate new blocks. But compression changes 1331e14bb325SJeff Bonwick * the blocksize, which forces a reallocate, and makes 1332e14bb325SJeff Bonwick * convergence take longer. Therefore, after the first 1333e14bb325SJeff Bonwick * few passes, stop compressing to ensure convergence. 1334e14bb325SJeff Bonwick */ 1335b24ab676SJeff Bonwick pass = spa_sync_pass(spa); 1336b24ab676SJeff Bonwick 1337b24ab676SJeff Bonwick ASSERT(zio->io_txg == spa_syncing_txg(spa)); 1338b24ab676SJeff Bonwick ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 1339b24ab676SJeff Bonwick ASSERT(!BP_GET_DEDUP(bp)); 1340e05725b1Sbonwick 134101f55e48SGeorge Wilson if (pass >= zfs_sync_pass_dont_compress) 1342e14bb325SJeff Bonwick compress = ZIO_COMPRESS_OFF; 1343e05725b1Sbonwick 1344e14bb325SJeff Bonwick /* Make sure someone doesn't change their mind on overwrites */ 13455d7b4d43SMatthew Ahrens ASSERT(BP_IS_EMBEDDED(bp) || MIN(zp->zp_copies + BP_IS_GANG(bp), 1346b24ab676SJeff Bonwick spa_max_replication(spa)) == BP_GET_NDVAS(bp)); 1347e14bb325SJeff Bonwick } 1348fa9e4066Sahrens 13495602294fSDan Kimmel /* If it's a compressed write that is not raw, compress the buffer. */ 13505602294fSDan Kimmel if (compress != ZIO_COMPRESS_OFF && psize == lsize) { 1351b24ab676SJeff Bonwick void *cbuf = zio_buf_alloc(lsize); 1352770499e1SDan Kimmel psize = zio_compress_data(compress, zio->io_abd, cbuf, lsize); 1353b24ab676SJeff Bonwick if (psize == 0 || psize == lsize) { 1354e14bb325SJeff Bonwick compress = ZIO_COMPRESS_OFF; 1355b24ab676SJeff Bonwick zio_buf_free(cbuf, lsize); 13565d7b4d43SMatthew Ahrens } else if (!zp->zp_dedup && psize <= BPE_PAYLOAD_SIZE && 13575d7b4d43SMatthew Ahrens zp->zp_level == 0 && !DMU_OT_HAS_FILL(zp->zp_type) && 13585d7b4d43SMatthew Ahrens spa_feature_is_enabled(spa, SPA_FEATURE_EMBEDDED_DATA)) { 13595d7b4d43SMatthew Ahrens encode_embedded_bp_compressed(bp, 13605d7b4d43SMatthew Ahrens cbuf, compress, lsize, psize); 13615d7b4d43SMatthew Ahrens BPE_SET_ETYPE(bp, BP_EMBEDDED_TYPE_DATA); 13625d7b4d43SMatthew Ahrens BP_SET_TYPE(bp, zio->io_prop.zp_type); 13635d7b4d43SMatthew Ahrens BP_SET_LEVEL(bp, zio->io_prop.zp_level); 13645d7b4d43SMatthew Ahrens zio_buf_free(cbuf, lsize); 13655d7b4d43SMatthew Ahrens bp->blk_birth = zio->io_txg; 13665d7b4d43SMatthew Ahrens zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 13675d7b4d43SMatthew Ahrens ASSERT(spa_feature_is_active(spa, 13685d7b4d43SMatthew Ahrens SPA_FEATURE_EMBEDDED_DATA)); 13695d7b4d43SMatthew Ahrens return (ZIO_PIPELINE_CONTINUE); 1370b24ab676SJeff Bonwick } else { 13715d7b4d43SMatthew Ahrens /* 137281cd5c55SMatthew Ahrens * Round up compressed size up to the ashift 137381cd5c55SMatthew Ahrens * of the smallest-ashift device, and zero the tail. 137481cd5c55SMatthew Ahrens * This ensures that the compressed size of the BP 137581cd5c55SMatthew Ahrens * (and thus compressratio property) are correct, 137681cd5c55SMatthew Ahrens * in that we charge for the padding used to fill out 137781cd5c55SMatthew Ahrens * the last sector. 13785d7b4d43SMatthew Ahrens */ 137981cd5c55SMatthew Ahrens ASSERT3U(spa->spa_min_ashift, >=, SPA_MINBLOCKSHIFT); 138081cd5c55SMatthew Ahrens size_t rounded = (size_t)P2ROUNDUP(psize, 138181cd5c55SMatthew Ahrens 1ULL << spa->spa_min_ashift); 138281cd5c55SMatthew Ahrens if (rounded >= lsize) { 13835d7b4d43SMatthew Ahrens compress = ZIO_COMPRESS_OFF; 13845d7b4d43SMatthew Ahrens zio_buf_free(cbuf, lsize); 138581cd5c55SMatthew Ahrens psize = lsize; 13865d7b4d43SMatthew Ahrens } else { 1387770499e1SDan Kimmel abd_t *cdata = abd_get_from_buf(cbuf, lsize); 1388770499e1SDan Kimmel abd_take_ownership_of_buf(cdata, B_TRUE); 1389770499e1SDan Kimmel abd_zero_off(cdata, psize, rounded - psize); 139081cd5c55SMatthew Ahrens psize = rounded; 1391770499e1SDan Kimmel zio_push_transform(zio, cdata, 13925d7b4d43SMatthew Ahrens psize, lsize, NULL); 13935d7b4d43SMatthew Ahrens } 1394e14bb325SJeff Bonwick } 13950f7643c7SGeorge Wilson 13960f7643c7SGeorge Wilson /* 13970f7643c7SGeorge Wilson * We were unable to handle this as an override bp, treat 13980f7643c7SGeorge Wilson * it as a regular write I/O. 13990f7643c7SGeorge Wilson */ 14000f7643c7SGeorge Wilson zio->io_bp_override = NULL; 14010f7643c7SGeorge Wilson *bp = zio->io_bp_orig; 14020f7643c7SGeorge Wilson zio->io_pipeline = zio->io_orig_pipeline; 14035602294fSDan Kimmel } else { 14045602294fSDan Kimmel ASSERT3U(psize, !=, 0); 1405e14bb325SJeff Bonwick } 1406c717a561Smaybee 1407e14bb325SJeff Bonwick /* 1408e14bb325SJeff Bonwick * The final pass of spa_sync() must be all rewrites, but the first 1409e14bb325SJeff Bonwick * few passes offer a trade-off: allocating blocks defers convergence, 1410e14bb325SJeff Bonwick * but newly allocated blocks are sequential, so they can be written 1411e14bb325SJeff Bonwick * to disk faster. Therefore, we allow the first few passes of 1412e14bb325SJeff Bonwick * spa_sync() to allocate new blocks, but force rewrites after that. 1413e14bb325SJeff Bonwick * There should only be a handful of blocks after pass 1 in any case. 1414e14bb325SJeff Bonwick */ 141543466aaeSMax Grossman if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg && 141643466aaeSMax Grossman BP_GET_PSIZE(bp) == psize && 141701f55e48SGeorge Wilson pass >= zfs_sync_pass_rewrite) { 1418663207adSDon Brady VERIFY3U(psize, !=, 0); 1419b24ab676SJeff Bonwick enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES; 1420663207adSDon Brady 1421e14bb325SJeff Bonwick zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages; 1422e14bb325SJeff Bonwick zio->io_flags |= ZIO_FLAG_IO_REWRITE; 1423e14bb325SJeff Bonwick } else { 1424e14bb325SJeff Bonwick BP_ZERO(bp); 1425e14bb325SJeff Bonwick zio->io_pipeline = ZIO_WRITE_PIPELINE; 1426e14bb325SJeff Bonwick } 1427fa9e4066Sahrens 1428b24ab676SJeff Bonwick if (psize == 0) { 142943466aaeSMax Grossman if (zio->io_bp_orig.blk_birth != 0 && 143043466aaeSMax Grossman spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) { 143143466aaeSMax Grossman BP_SET_LSIZE(bp, lsize); 143243466aaeSMax Grossman BP_SET_TYPE(bp, zp->zp_type); 143343466aaeSMax Grossman BP_SET_LEVEL(bp, zp->zp_level); 143443466aaeSMax Grossman BP_SET_BIRTH(bp, zio->io_txg, 0); 143543466aaeSMax Grossman } 1436e14bb325SJeff Bonwick zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 1437e14bb325SJeff Bonwick } else { 1438e14bb325SJeff Bonwick ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER); 1439e14bb325SJeff Bonwick BP_SET_LSIZE(bp, lsize); 144043466aaeSMax Grossman BP_SET_TYPE(bp, zp->zp_type); 144143466aaeSMax Grossman BP_SET_LEVEL(bp, zp->zp_level); 1442b24ab676SJeff Bonwick BP_SET_PSIZE(bp, psize); 1443e14bb325SJeff Bonwick BP_SET_COMPRESS(bp, compress); 1444e14bb325SJeff Bonwick BP_SET_CHECKSUM(bp, zp->zp_checksum); 1445b24ab676SJeff Bonwick BP_SET_DEDUP(bp, zp->zp_dedup); 1446e14bb325SJeff Bonwick BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER); 1447b24ab676SJeff Bonwick if (zp->zp_dedup) { 1448b24ab676SJeff Bonwick ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 1449b24ab676SJeff Bonwick ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE)); 1450b24ab676SJeff Bonwick zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE; 1451b24ab676SJeff Bonwick } 145280901aeaSGeorge Wilson if (zp->zp_nopwrite) { 145380901aeaSGeorge Wilson ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 145480901aeaSGeorge Wilson ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE)); 145580901aeaSGeorge Wilson zio->io_pipeline |= ZIO_STAGE_NOP_WRITE; 145680901aeaSGeorge Wilson } 1457b24ab676SJeff Bonwick } 1458b24ab676SJeff Bonwick return (ZIO_PIPELINE_CONTINUE); 1459b24ab676SJeff Bonwick } 1460b24ab676SJeff Bonwick 1461b24ab676SJeff Bonwick static int 1462b24ab676SJeff Bonwick zio_free_bp_init(zio_t *zio) 1463b24ab676SJeff Bonwick { 1464b24ab676SJeff Bonwick blkptr_t *bp = zio->io_bp; 1465b24ab676SJeff Bonwick 1466b24ab676SJeff Bonwick if (zio->io_child_type == ZIO_CHILD_LOGICAL) { 1467b24ab676SJeff Bonwick if (BP_GET_DEDUP(bp)) 1468b24ab676SJeff Bonwick zio->io_pipeline = ZIO_DDT_FREE_PIPELINE; 1469e14bb325SJeff Bonwick } 1470fa9e4066Sahrens 14715cabbc6bSPrashanth Sreenivasa ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy); 14725cabbc6bSPrashanth Sreenivasa 1473e05725b1Sbonwick return (ZIO_PIPELINE_CONTINUE); 1474fa9e4066Sahrens } 1475fa9e4066Sahrens 1476e14bb325SJeff Bonwick /* 1477e14bb325SJeff Bonwick * ========================================================================== 1478e14bb325SJeff Bonwick * Execute the I/O pipeline 1479e14bb325SJeff Bonwick * ========================================================================== 1480e14bb325SJeff Bonwick */ 1481e14bb325SJeff Bonwick 1482e14bb325SJeff Bonwick static void 1483ec94d322SAdam Leventhal zio_taskq_dispatch(zio_t *zio, zio_taskq_type_t q, boolean_t cutinline) 1484fa9e4066Sahrens { 148580eb36f2SGeorge Wilson spa_t *spa = zio->io_spa; 1486e14bb325SJeff Bonwick zio_type_t t = zio->io_type; 14875aeb9474SGarrett D'Amore int flags = (cutinline ? TQ_FRONT : 0); 14880a4e9518Sgw 14890a4e9518Sgw /* 1490bbe36defSGeorge Wilson * If we're a config writer or a probe, the normal issue and 1491bbe36defSGeorge Wilson * interrupt threads may all be blocked waiting for the config lock. 1492bbe36defSGeorge Wilson * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL. 14930a4e9518Sgw */ 1494bbe36defSGeorge Wilson if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE)) 1495e14bb325SJeff Bonwick t = ZIO_TYPE_NULL; 14960a4e9518Sgw 14970a4e9518Sgw /* 1498e14bb325SJeff Bonwick * A similar issue exists for the L2ARC write thread until L2ARC 2.0. 14990a4e9518Sgw */ 1500e14bb325SJeff Bonwick if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux) 1501e14bb325SJeff Bonwick t = ZIO_TYPE_NULL; 15020a4e9518Sgw 150380eb36f2SGeorge Wilson /* 1504ec94d322SAdam Leventhal * If this is a high priority I/O, then use the high priority taskq if 1505ec94d322SAdam Leventhal * available. 150680eb36f2SGeorge Wilson */ 15072258ad0bSGeorge Wilson if ((zio->io_priority == ZIO_PRIORITY_NOW || 15082258ad0bSGeorge Wilson zio->io_priority == ZIO_PRIORITY_SYNC_WRITE) && 1509ec94d322SAdam Leventhal spa->spa_zio_taskq[t][q + 1].stqs_count != 0) 151080eb36f2SGeorge Wilson q++; 151180eb36f2SGeorge Wilson 151280eb36f2SGeorge Wilson ASSERT3U(q, <, ZIO_TASKQ_TYPES); 15135aeb9474SGarrett D'Amore 15145aeb9474SGarrett D'Amore /* 15155aeb9474SGarrett D'Amore * NB: We are assuming that the zio can only be dispatched 15165aeb9474SGarrett D'Amore * to a single taskq at a time. It would be a grievous error 15175aeb9474SGarrett D'Amore * to dispatch the zio to another taskq at the same time. 15185aeb9474SGarrett D'Amore */ 15195aeb9474SGarrett D'Amore ASSERT(zio->io_tqent.tqent_next == NULL); 1520ec94d322SAdam Leventhal spa_taskq_dispatch_ent(spa, t, q, (task_func_t *)zio_execute, zio, 1521ec94d322SAdam Leventhal flags, &zio->io_tqent); 1522e14bb325SJeff Bonwick } 15230a4e9518Sgw 1524e14bb325SJeff Bonwick static boolean_t 1525ec94d322SAdam Leventhal zio_taskq_member(zio_t *zio, zio_taskq_type_t q) 1526e14bb325SJeff Bonwick { 1527e14bb325SJeff Bonwick kthread_t *executor = zio->io_executor; 1528e14bb325SJeff Bonwick spa_t *spa = zio->io_spa; 15290a4e9518Sgw 1530ec94d322SAdam Leventhal for (zio_type_t t = 0; t < ZIO_TYPES; t++) { 1531ec94d322SAdam Leventhal spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q]; 1532ec94d322SAdam Leventhal uint_t i; 1533ec94d322SAdam Leventhal for (i = 0; i < tqs->stqs_count; i++) { 1534ec94d322SAdam Leventhal if (taskq_member(tqs->stqs_taskq[i], executor)) 1535ec94d322SAdam Leventhal return (B_TRUE); 1536ec94d322SAdam Leventhal } 1537ec94d322SAdam Leventhal } 15380a4e9518Sgw 1539e14bb325SJeff Bonwick return (B_FALSE); 1540e14bb325SJeff Bonwick } 1541e05725b1Sbonwick 1542e14bb325SJeff Bonwick static int 1543e14bb325SJeff Bonwick zio_issue_async(zio_t *zio) 1544e14bb325SJeff Bonwick { 154535a5a358SJonathan Adams zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE); 1546e14bb325SJeff Bonwick 1547e14bb325SJeff Bonwick return (ZIO_PIPELINE_STOP); 15480a4e9518Sgw } 15490a4e9518Sgw 1550e14bb325SJeff Bonwick void 1551e14bb325SJeff Bonwick zio_interrupt(zio_t *zio) 15520a4e9518Sgw { 155335a5a358SJonathan Adams zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE); 1554e14bb325SJeff Bonwick } 15550a4e9518Sgw 155697e81309SPrakash Surya void 155797e81309SPrakash Surya zio_delay_interrupt(zio_t *zio) 155897e81309SPrakash Surya { 155997e81309SPrakash Surya /* 156097e81309SPrakash Surya * The timeout_generic() function isn't defined in userspace, so 156197e81309SPrakash Surya * rather than trying to implement the function, the zio delay 156297e81309SPrakash Surya * functionality has been disabled for userspace builds. 156397e81309SPrakash Surya */ 156497e81309SPrakash Surya 156597e81309SPrakash Surya #ifdef _KERNEL 156697e81309SPrakash Surya /* 156797e81309SPrakash Surya * If io_target_timestamp is zero, then no delay has been registered 156897e81309SPrakash Surya * for this IO, thus jump to the end of this function and "skip" the 156997e81309SPrakash Surya * delay; issuing it directly to the zio layer. 157097e81309SPrakash Surya */ 157197e81309SPrakash Surya if (zio->io_target_timestamp != 0) { 157297e81309SPrakash Surya hrtime_t now = gethrtime(); 157397e81309SPrakash Surya 157497e81309SPrakash Surya if (now >= zio->io_target_timestamp) { 157597e81309SPrakash Surya /* 157697e81309SPrakash Surya * This IO has already taken longer than the target 157797e81309SPrakash Surya * delay to complete, so we don't want to delay it 157897e81309SPrakash Surya * any longer; we "miss" the delay and issue it 157997e81309SPrakash Surya * directly to the zio layer. This is likely due to 158097e81309SPrakash Surya * the target latency being set to a value less than 158197e81309SPrakash Surya * the underlying hardware can satisfy (e.g. delay 158297e81309SPrakash Surya * set to 1ms, but the disks take 10ms to complete an 158397e81309SPrakash Surya * IO request). 158497e81309SPrakash Surya */ 158597e81309SPrakash Surya 158697e81309SPrakash Surya DTRACE_PROBE2(zio__delay__miss, zio_t *, zio, 158797e81309SPrakash Surya hrtime_t, now); 158897e81309SPrakash Surya 158997e81309SPrakash Surya zio_interrupt(zio); 159097e81309SPrakash Surya } else { 159197e81309SPrakash Surya hrtime_t diff = zio->io_target_timestamp - now; 159297e81309SPrakash Surya 159397e81309SPrakash Surya DTRACE_PROBE3(zio__delay__hit, zio_t *, zio, 159497e81309SPrakash Surya hrtime_t, now, hrtime_t, diff); 159597e81309SPrakash Surya 159697e81309SPrakash Surya (void) timeout_generic(CALLOUT_NORMAL, 159797e81309SPrakash Surya (void (*)(void *))zio_interrupt, zio, diff, 1, 0); 159897e81309SPrakash Surya } 159997e81309SPrakash Surya 160097e81309SPrakash Surya return; 160197e81309SPrakash Surya } 160297e81309SPrakash Surya #endif 160397e81309SPrakash Surya 160497e81309SPrakash Surya DTRACE_PROBE1(zio__delay__skip, zio_t *, zio); 160597e81309SPrakash Surya zio_interrupt(zio); 160697e81309SPrakash Surya } 160797e81309SPrakash Surya 1608e14bb325SJeff Bonwick /* 1609e14bb325SJeff Bonwick * Execute the I/O pipeline until one of the following occurs: 1610f7170741SWill Andrews * 1611f7170741SWill Andrews * (1) the I/O completes 1612f7170741SWill Andrews * (2) the pipeline stalls waiting for dependent child I/Os 1613f7170741SWill Andrews * (3) the I/O issues, so we're waiting for an I/O completion interrupt 1614f7170741SWill Andrews * (4) the I/O is delegated by vdev-level caching or aggregation 1615f7170741SWill Andrews * (5) the I/O is deferred due to vdev-level queueing 1616f7170741SWill Andrews * (6) the I/O is handed off to another thread. 1617f7170741SWill Andrews * 1618f7170741SWill Andrews * In all cases, the pipeline stops whenever there's no CPU work; it never 1619f7170741SWill Andrews * burns a thread in cv_wait(). 1620e14bb325SJeff Bonwick * 1621e14bb325SJeff Bonwick * There's no locking on io_stage because there's no legitimate way 1622e14bb325SJeff Bonwick * for multiple threads to be attempting to process the same I/O. 1623e14bb325SJeff Bonwick */ 1624b24ab676SJeff Bonwick static zio_pipe_stage_t *zio_pipeline[]; 16250a4e9518Sgw 1626e14bb325SJeff Bonwick void 1627e14bb325SJeff Bonwick zio_execute(zio_t *zio) 1628e14bb325SJeff Bonwick { 1629e14bb325SJeff Bonwick zio->io_executor = curthread; 16300a4e9518Sgw 16310f7643c7SGeorge Wilson ASSERT3U(zio->io_queued_timestamp, >, 0); 16320f7643c7SGeorge Wilson 1633e14bb325SJeff Bonwick while (zio->io_stage < ZIO_STAGE_DONE) { 1634b24ab676SJeff Bonwick enum zio_stage pipeline = zio->io_pipeline; 1635b24ab676SJeff Bonwick enum zio_stage stage = zio->io_stage; 1636e14bb325SJeff Bonwick int rv; 16370a4e9518Sgw 1638e14bb325SJeff Bonwick ASSERT(!MUTEX_HELD(&zio->io_lock)); 1639b24ab676SJeff Bonwick ASSERT(ISP2(stage)); 1640b24ab676SJeff Bonwick ASSERT(zio->io_stall == NULL); 16410a4e9518Sgw 1642b24ab676SJeff Bonwick do { 1643b24ab676SJeff Bonwick stage <<= 1; 1644b24ab676SJeff Bonwick } while ((stage & pipeline) == 0); 1645e14bb325SJeff Bonwick 1646e14bb325SJeff Bonwick ASSERT(stage <= ZIO_STAGE_DONE); 16470a4e9518Sgw 16480a4e9518Sgw /* 1649e14bb325SJeff Bonwick * If we are in interrupt context and this pipeline stage 1650e14bb325SJeff Bonwick * will grab a config lock that is held across I/O, 1651b24ab676SJeff Bonwick * or may wait for an I/O that needs an interrupt thread 1652b24ab676SJeff Bonwick * to complete, issue async to avoid deadlock. 165335a5a358SJonathan Adams * 165435a5a358SJonathan Adams * For VDEV_IO_START, we cut in line so that the io will 165535a5a358SJonathan Adams * be sent to disk promptly. 16560a4e9518Sgw */ 1657b24ab676SJeff Bonwick if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL && 1658e14bb325SJeff Bonwick zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) { 165935a5a358SJonathan Adams boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ? 166035a5a358SJonathan Adams zio_requeue_io_start_cut_in_line : B_FALSE; 166135a5a358SJonathan Adams zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut); 1662e14bb325SJeff Bonwick return; 16630a4e9518Sgw } 16640a4e9518Sgw 1665e14bb325SJeff Bonwick zio->io_stage = stage; 16660f7643c7SGeorge Wilson zio->io_pipeline_trace |= zio->io_stage; 1667bf16b11eSMatthew Ahrens rv = zio_pipeline[highbit64(stage) - 1](zio); 16680a4e9518Sgw 1669e14bb325SJeff Bonwick if (rv == ZIO_PIPELINE_STOP) 1670e14bb325SJeff Bonwick return; 16710a4e9518Sgw 1672e14bb325SJeff Bonwick ASSERT(rv == ZIO_PIPELINE_CONTINUE); 1673e14bb325SJeff Bonwick } 16740a4e9518Sgw } 16750a4e9518Sgw 1676e14bb325SJeff Bonwick /* 1677e14bb325SJeff Bonwick * ========================================================================== 1678e14bb325SJeff Bonwick * Initiate I/O, either sync or async 1679e14bb325SJeff Bonwick * ========================================================================== 1680e14bb325SJeff Bonwick */ 1681e14bb325SJeff Bonwick int 1682e14bb325SJeff Bonwick zio_wait(zio_t *zio) 16830a4e9518Sgw { 1684e14bb325SJeff Bonwick int error; 16850a4e9518Sgw 16861271e4b1SPrakash Surya ASSERT3P(zio->io_stage, ==, ZIO_STAGE_OPEN); 16871271e4b1SPrakash Surya ASSERT3P(zio->io_executor, ==, NULL); 16880a4e9518Sgw 1689e14bb325SJeff Bonwick zio->io_waiter = curthread; 16900f7643c7SGeorge Wilson ASSERT0(zio->io_queued_timestamp); 16910f7643c7SGeorge Wilson zio->io_queued_timestamp = gethrtime(); 1692e05725b1Sbonwick 1693e14bb325SJeff Bonwick zio_execute(zio); 16940a4e9518Sgw 1695e14bb325SJeff Bonwick mutex_enter(&zio->io_lock); 1696e14bb325SJeff Bonwick while (zio->io_executor != NULL) 1697e14bb325SJeff Bonwick cv_wait(&zio->io_cv, &zio->io_lock); 1698e14bb325SJeff Bonwick mutex_exit(&zio->io_lock); 169932b87932Sek 1700e14bb325SJeff Bonwick error = zio->io_error; 1701e14bb325SJeff Bonwick zio_destroy(zio); 170232b87932Sek 1703e14bb325SJeff Bonwick return (error); 170432b87932Sek } 170532b87932Sek 1706e14bb325SJeff Bonwick void 1707e14bb325SJeff Bonwick zio_nowait(zio_t *zio) 17080a4e9518Sgw { 17091271e4b1SPrakash Surya ASSERT3P(zio->io_executor, ==, NULL); 1710fa9e4066Sahrens 1711a3f829aeSBill Moore if (zio->io_child_type == ZIO_CHILD_LOGICAL && 1712a3f829aeSBill Moore zio_unique_parent(zio) == NULL) { 1713ea8dc4b6Seschrock /* 1714e14bb325SJeff Bonwick * This is a logical async I/O with no parent to wait for it. 171554d692b7SGeorge Wilson * We add it to the spa_async_root_zio "Godfather" I/O which 171654d692b7SGeorge Wilson * will ensure they complete prior to unloading the pool. 1717ea8dc4b6Seschrock */ 1718e14bb325SJeff Bonwick spa_t *spa = zio->io_spa; 171954d692b7SGeorge Wilson 17206f834bc1SMatthew Ahrens zio_add_child(spa->spa_async_zio_root[CPU_SEQID], zio); 1721e14bb325SJeff Bonwick } 1722ea8dc4b6Seschrock 17230f7643c7SGeorge Wilson ASSERT0(zio->io_queued_timestamp); 17240f7643c7SGeorge Wilson zio->io_queued_timestamp = gethrtime(); 1725e14bb325SJeff Bonwick zio_execute(zio); 1726e14bb325SJeff Bonwick } 1727ea8dc4b6Seschrock 1728e14bb325SJeff Bonwick /* 1729e14bb325SJeff Bonwick * ========================================================================== 17301271e4b1SPrakash Surya * Reexecute, cancel, or suspend/resume failed I/O 1731e14bb325SJeff Bonwick * ========================================================================== 1732e14bb325SJeff Bonwick */ 1733fa9e4066Sahrens 1734e14bb325SJeff Bonwick static void 1735e14bb325SJeff Bonwick zio_reexecute(zio_t *pio) 1736e14bb325SJeff Bonwick { 1737a3f829aeSBill Moore zio_t *cio, *cio_next; 1738a3f829aeSBill Moore 1739a3f829aeSBill Moore ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL); 1740a3f829aeSBill Moore ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN); 1741f5383399SBill Moore ASSERT(pio->io_gang_leader == NULL); 1742f5383399SBill Moore ASSERT(pio->io_gang_tree == NULL); 1743e05725b1Sbonwick 1744e14bb325SJeff Bonwick pio->io_flags = pio->io_orig_flags; 1745e14bb325SJeff Bonwick pio->io_stage = pio->io_orig_stage; 1746e14bb325SJeff Bonwick pio->io_pipeline = pio->io_orig_pipeline; 1747e14bb325SJeff Bonwick pio->io_reexecute = 0; 174880901aeaSGeorge Wilson pio->io_flags |= ZIO_FLAG_REEXECUTED; 17490f7643c7SGeorge Wilson pio->io_pipeline_trace = 0; 1750e14bb325SJeff Bonwick pio->io_error = 0; 1751a3f829aeSBill Moore for (int w = 0; w < ZIO_WAIT_TYPES; w++) 1752a3f829aeSBill Moore pio->io_state[w] = 0; 1753e14bb325SJeff Bonwick for (int c = 0; c < ZIO_CHILD_TYPES; c++) 1754e14bb325SJeff Bonwick pio->io_child_error[c] = 0; 17550a4e9518Sgw 1756b24ab676SJeff Bonwick if (IO_IS_ALLOCATING(pio)) 1757b24ab676SJeff Bonwick BP_ZERO(pio->io_bp); 1758d58459f4Sek 1759e14bb325SJeff Bonwick /* 1760e14bb325SJeff Bonwick * As we reexecute pio's children, new children could be created. 1761a3f829aeSBill Moore * New children go to the head of pio's io_child_list, however, 1762e14bb325SJeff Bonwick * so we will (correctly) not reexecute them. The key is that 1763a3f829aeSBill Moore * the remainder of pio's io_child_list, from 'cio_next' onward, 1764a3f829aeSBill Moore * cannot be affected by any side effects of reexecuting 'cio'. 1765e14bb325SJeff Bonwick */ 17660f7643c7SGeorge Wilson zio_link_t *zl = NULL; 1767*a3874b8bSToomas Soome mutex_enter(&pio->io_lock); 17680f7643c7SGeorge Wilson for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) { 17690f7643c7SGeorge Wilson cio_next = zio_walk_children(pio, &zl); 1770a3f829aeSBill Moore for (int w = 0; w < ZIO_WAIT_TYPES; w++) 1771a3f829aeSBill Moore pio->io_children[cio->io_child_type][w]++; 1772e14bb325SJeff Bonwick mutex_exit(&pio->io_lock); 1773a3f829aeSBill Moore zio_reexecute(cio); 1774*a3874b8bSToomas Soome mutex_enter(&pio->io_lock); 1775fa9e4066Sahrens } 1776*a3874b8bSToomas Soome mutex_exit(&pio->io_lock); 1777e05725b1Sbonwick 1778e14bb325SJeff Bonwick /* 1779e14bb325SJeff Bonwick * Now that all children have been reexecuted, execute the parent. 178054d692b7SGeorge Wilson * We don't reexecute "The Godfather" I/O here as it's the 178148bbca81SDaniel Hoffman * responsibility of the caller to wait on it. 1782e14bb325SJeff Bonwick */ 17830f7643c7SGeorge Wilson if (!(pio->io_flags & ZIO_FLAG_GODFATHER)) { 17840f7643c7SGeorge Wilson pio->io_queued_timestamp = gethrtime(); 178554d692b7SGeorge Wilson zio_execute(pio); 17860f7643c7SGeorge Wilson } 17870a4e9518Sgw } 17880a4e9518Sgw 1789e14bb325SJeff Bonwick void 1790e0f1c0afSOlaf Faaland zio_suspend(spa_t *spa, zio_t *zio, zio_suspend_reason_t reason) 17910a4e9518Sgw { 1792e14bb325SJeff Bonwick if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC) 1793e14bb325SJeff Bonwick fm_panic("Pool '%s' has encountered an uncorrectable I/O " 1794e14bb325SJeff Bonwick "failure and the failure mode property for this pool " 1795e14bb325SJeff Bonwick "is set to panic.", spa_name(spa)); 17960a4e9518Sgw 1797e14bb325SJeff Bonwick zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0); 17980a4e9518Sgw 1799e14bb325SJeff Bonwick mutex_enter(&spa->spa_suspend_lock); 1800fa9e4066Sahrens 1801e14bb325SJeff Bonwick if (spa->spa_suspend_zio_root == NULL) 180254d692b7SGeorge Wilson spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL, 180354d692b7SGeorge Wilson ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | 180454d692b7SGeorge Wilson ZIO_FLAG_GODFATHER); 1805fa9e4066Sahrens 1806e0f1c0afSOlaf Faaland spa->spa_suspended = reason; 1807fa9e4066Sahrens 1808e14bb325SJeff Bonwick if (zio != NULL) { 180954d692b7SGeorge Wilson ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER)); 1810e14bb325SJeff Bonwick ASSERT(zio != spa->spa_suspend_zio_root); 1811e14bb325SJeff Bonwick ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 1812a3f829aeSBill Moore ASSERT(zio_unique_parent(zio) == NULL); 1813e14bb325SJeff Bonwick ASSERT(zio->io_stage == ZIO_STAGE_DONE); 1814e14bb325SJeff Bonwick zio_add_child(spa->spa_suspend_zio_root, zio); 1815e14bb325SJeff Bonwick } 1816fa9e4066Sahrens 1817e14bb325SJeff Bonwick mutex_exit(&spa->spa_suspend_lock); 1818e14bb325SJeff Bonwick } 1819fa9e4066Sahrens 182054d692b7SGeorge Wilson int 1821e14bb325SJeff Bonwick zio_resume(spa_t *spa) 1822e14bb325SJeff Bonwick { 182354d692b7SGeorge Wilson zio_t *pio; 1824fa9e4066Sahrens 1825b3995adbSahrens /* 1826e14bb325SJeff Bonwick * Reexecute all previously suspended i/o. 1827b3995adbSahrens */ 1828e14bb325SJeff Bonwick mutex_enter(&spa->spa_suspend_lock); 1829e0f1c0afSOlaf Faaland spa->spa_suspended = ZIO_SUSPEND_NONE; 1830e14bb325SJeff Bonwick cv_broadcast(&spa->spa_suspend_cv); 1831e14bb325SJeff Bonwick pio = spa->spa_suspend_zio_root; 1832e14bb325SJeff Bonwick spa->spa_suspend_zio_root = NULL; 1833e14bb325SJeff Bonwick mutex_exit(&spa->spa_suspend_lock); 1834e14bb325SJeff Bonwick 1835e14bb325SJeff Bonwick if (pio == NULL) 183654d692b7SGeorge Wilson return (0); 1837e14bb325SJeff Bonwick 183854d692b7SGeorge Wilson zio_reexecute(pio); 183954d692b7SGeorge Wilson return (zio_wait(pio)); 1840e14bb325SJeff Bonwick } 1841e14bb325SJeff Bonwick 1842e14bb325SJeff Bonwick void 1843e14bb325SJeff Bonwick zio_resume_wait(spa_t *spa) 1844e14bb325SJeff Bonwick { 1845e14bb325SJeff Bonwick mutex_enter(&spa->spa_suspend_lock); 1846e14bb325SJeff Bonwick while (spa_suspended(spa)) 1847e14bb325SJeff Bonwick cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock); 1848e14bb325SJeff Bonwick mutex_exit(&spa->spa_suspend_lock); 1849fa9e4066Sahrens } 1850fa9e4066Sahrens 1851fa9e4066Sahrens /* 1852fa9e4066Sahrens * ========================================================================== 1853e14bb325SJeff Bonwick * Gang blocks. 1854e14bb325SJeff Bonwick * 1855e14bb325SJeff Bonwick * A gang block is a collection of small blocks that looks to the DMU 1856e14bb325SJeff Bonwick * like one large block. When zio_dva_allocate() cannot find a block 1857e14bb325SJeff Bonwick * of the requested size, due to either severe fragmentation or the pool 1858e14bb325SJeff Bonwick * being nearly full, it calls zio_write_gang_block() to construct the 1859e14bb325SJeff Bonwick * block from smaller fragments. 1860e14bb325SJeff Bonwick * 1861e14bb325SJeff Bonwick * A gang block consists of a gang header (zio_gbh_phys_t) and up to 1862e14bb325SJeff Bonwick * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like 1863e14bb325SJeff Bonwick * an indirect block: it's an array of block pointers. It consumes 1864e14bb325SJeff Bonwick * only one sector and hence is allocatable regardless of fragmentation. 1865e14bb325SJeff Bonwick * The gang header's bps point to its gang members, which hold the data. 1866e14bb325SJeff Bonwick * 1867e14bb325SJeff Bonwick * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg> 1868e14bb325SJeff Bonwick * as the verifier to ensure uniqueness of the SHA256 checksum. 1869e14bb325SJeff Bonwick * Critically, the gang block bp's blk_cksum is the checksum of the data, 1870e14bb325SJeff Bonwick * not the gang header. This ensures that data block signatures (needed for 1871e14bb325SJeff Bonwick * deduplication) are independent of how the block is physically stored. 1872e14bb325SJeff Bonwick * 1873e14bb325SJeff Bonwick * Gang blocks can be nested: a gang member may itself be a gang block. 1874e14bb325SJeff Bonwick * Thus every gang block is a tree in which root and all interior nodes are 1875e14bb325SJeff Bonwick * gang headers, and the leaves are normal blocks that contain user data. 1876e14bb325SJeff Bonwick * The root of the gang tree is called the gang leader. 1877e14bb325SJeff Bonwick * 1878e14bb325SJeff Bonwick * To perform any operation (read, rewrite, free, claim) on a gang block, 1879e14bb325SJeff Bonwick * zio_gang_assemble() first assembles the gang tree (minus data leaves) 1880e14bb325SJeff Bonwick * in the io_gang_tree field of the original logical i/o by recursively 1881e14bb325SJeff Bonwick * reading the gang leader and all gang headers below it. This yields 1882e14bb325SJeff Bonwick * an in-core tree containing the contents of every gang header and the 1883e14bb325SJeff Bonwick * bps for every constituent of the gang block. 1884e14bb325SJeff Bonwick * 1885e14bb325SJeff Bonwick * With the gang tree now assembled, zio_gang_issue() just walks the gang tree 1886e14bb325SJeff Bonwick * and invokes a callback on each bp. To free a gang block, zio_gang_issue() 1887e14bb325SJeff Bonwick * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp. 1888e14bb325SJeff Bonwick * zio_claim_gang() provides a similarly trivial wrapper for zio_claim(). 1889e14bb325SJeff Bonwick * zio_read_gang() is a wrapper around zio_read() that omits reading gang 1890e14bb325SJeff Bonwick * headers, since we already have those in io_gang_tree. zio_rewrite_gang() 1891e14bb325SJeff Bonwick * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite() 1892e14bb325SJeff Bonwick * of the gang header plus zio_checksum_compute() of the data to update the 1893e14bb325SJeff Bonwick * gang header's blk_cksum as described above. 1894e14bb325SJeff Bonwick * 1895e14bb325SJeff Bonwick * The two-phase assemble/issue model solves the problem of partial failure -- 1896e14bb325SJeff Bonwick * what if you'd freed part of a gang block but then couldn't read the 1897e14bb325SJeff Bonwick * gang header for another part? Assembling the entire gang tree first 1898e14bb325SJeff Bonwick * ensures that all the necessary gang header I/O has succeeded before 1899e14bb325SJeff Bonwick * starting the actual work of free, claim, or write. Once the gang tree 1900e14bb325SJeff Bonwick * is assembled, free and claim are in-memory operations that cannot fail. 1901e14bb325SJeff Bonwick * 1902e14bb325SJeff Bonwick * In the event that a gang write fails, zio_dva_unallocate() walks the 1903e14bb325SJeff Bonwick * gang tree to immediately free (i.e. insert back into the space map) 1904e14bb325SJeff Bonwick * everything we've allocated. This ensures that we don't get ENOSPC 1905e14bb325SJeff Bonwick * errors during repeated suspend/resume cycles due to a flaky device. 1906e14bb325SJeff Bonwick * 1907e14bb325SJeff Bonwick * Gang rewrites only happen during sync-to-convergence. If we can't assemble 1908e14bb325SJeff Bonwick * the gang tree, we won't modify the block, so we can safely defer the free 1909e14bb325SJeff Bonwick * (knowing that the block is still intact). If we *can* assemble the gang 1910e14bb325SJeff Bonwick * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free 1911e14bb325SJeff Bonwick * each constituent bp and we can allocate a new block on the next sync pass. 1912e14bb325SJeff Bonwick * 1913e14bb325SJeff Bonwick * In all cases, the gang tree allows complete recovery from partial failure. 1914fa9e4066Sahrens * ========================================================================== 1915fa9e4066Sahrens */ 1916e14bb325SJeff Bonwick 1917770499e1SDan Kimmel static void 1918770499e1SDan Kimmel zio_gang_issue_func_done(zio_t *zio) 1919770499e1SDan Kimmel { 1920770499e1SDan Kimmel abd_put(zio->io_abd); 1921770499e1SDan Kimmel } 1922770499e1SDan Kimmel 1923e14bb325SJeff Bonwick static zio_t * 1924770499e1SDan Kimmel zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data, 1925770499e1SDan Kimmel uint64_t offset) 1926fa9e4066Sahrens { 1927e14bb325SJeff Bonwick if (gn != NULL) 1928e14bb325SJeff Bonwick return (pio); 1929fa9e4066Sahrens 1930770499e1SDan Kimmel return (zio_read(pio, pio->io_spa, bp, abd_get_offset(data, offset), 1931770499e1SDan Kimmel BP_GET_PSIZE(bp), zio_gang_issue_func_done, 1932770499e1SDan Kimmel NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), 1933e14bb325SJeff Bonwick &pio->io_bookmark)); 1934e14bb325SJeff Bonwick } 1935e14bb325SJeff Bonwick 1936770499e1SDan Kimmel static zio_t * 1937770499e1SDan Kimmel zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data, 1938770499e1SDan Kimmel uint64_t offset) 1939e14bb325SJeff Bonwick { 1940e14bb325SJeff Bonwick zio_t *zio; 1941e14bb325SJeff Bonwick 1942e14bb325SJeff Bonwick if (gn != NULL) { 1943770499e1SDan Kimmel abd_t *gbh_abd = 1944770499e1SDan Kimmel abd_get_from_buf(gn->gn_gbh, SPA_GANGBLOCKSIZE); 1945e14bb325SJeff Bonwick zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp, 1946770499e1SDan Kimmel gbh_abd, SPA_GANGBLOCKSIZE, zio_gang_issue_func_done, NULL, 1947770499e1SDan Kimmel pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), 1948770499e1SDan Kimmel &pio->io_bookmark); 1949fa9e4066Sahrens /* 1950e14bb325SJeff Bonwick * As we rewrite each gang header, the pipeline will compute 1951e14bb325SJeff Bonwick * a new gang block header checksum for it; but no one will 1952e14bb325SJeff Bonwick * compute a new data checksum, so we do that here. The one 1953e14bb325SJeff Bonwick * exception is the gang leader: the pipeline already computed 1954e14bb325SJeff Bonwick * its data checksum because that stage precedes gang assembly. 1955e14bb325SJeff Bonwick * (Presently, nothing actually uses interior data checksums; 1956e14bb325SJeff Bonwick * this is just good hygiene.) 1957fa9e4066Sahrens */ 1958f5383399SBill Moore if (gn != pio->io_gang_leader->io_gang_tree) { 1959770499e1SDan Kimmel abd_t *buf = abd_get_offset(data, offset); 1960770499e1SDan Kimmel 1961e14bb325SJeff Bonwick zio_checksum_compute(zio, BP_GET_CHECKSUM(bp), 1962770499e1SDan Kimmel buf, BP_GET_PSIZE(bp)); 1963770499e1SDan Kimmel 1964770499e1SDan Kimmel abd_put(buf); 1965e14bb325SJeff Bonwick } 1966b24ab676SJeff Bonwick /* 1967b24ab676SJeff Bonwick * If we are here to damage data for testing purposes, 1968b24ab676SJeff Bonwick * leave the GBH alone so that we can detect the damage. 1969b24ab676SJeff Bonwick */ 1970b24ab676SJeff Bonwick if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE) 1971b24ab676SJeff Bonwick zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES; 1972fa9e4066Sahrens } else { 1973e14bb325SJeff Bonwick zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp, 1974770499e1SDan Kimmel abd_get_offset(data, offset), BP_GET_PSIZE(bp), 1975770499e1SDan Kimmel zio_gang_issue_func_done, NULL, pio->io_priority, 1976e14bb325SJeff Bonwick ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark); 1977fa9e4066Sahrens } 1978fa9e4066Sahrens 1979e14bb325SJeff Bonwick return (zio); 1980e14bb325SJeff Bonwick } 1981fa9e4066Sahrens 1982e14bb325SJeff Bonwick /* ARGSUSED */ 1983770499e1SDan Kimmel static zio_t * 1984770499e1SDan Kimmel zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data, 1985770499e1SDan Kimmel uint64_t offset) 1986e14bb325SJeff Bonwick { 1987b24ab676SJeff Bonwick return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp, 1988b24ab676SJeff Bonwick ZIO_GANG_CHILD_FLAGS(pio))); 1989fa9e4066Sahrens } 1990fa9e4066Sahrens 1991e14bb325SJeff Bonwick /* ARGSUSED */ 1992770499e1SDan Kimmel static zio_t * 1993770499e1SDan Kimmel zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data, 1994770499e1SDan Kimmel uint64_t offset) 1995fa9e4066Sahrens { 1996e14bb325SJeff Bonwick return (zio_claim(pio, pio->io_spa, pio->io_txg, bp, 1997e14bb325SJeff Bonwick NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio))); 1998e14bb325SJeff Bonwick } 1999fa9e4066Sahrens 2000e14bb325SJeff Bonwick static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = { 2001e14bb325SJeff Bonwick NULL, 2002e14bb325SJeff Bonwick zio_read_gang, 2003e14bb325SJeff Bonwick zio_rewrite_gang, 2004e14bb325SJeff Bonwick zio_free_gang, 2005e14bb325SJeff Bonwick zio_claim_gang, 2006e14bb325SJeff Bonwick NULL 2007e14bb325SJeff Bonwick }; 2008fa9e4066Sahrens 2009e14bb325SJeff Bonwick static void zio_gang_tree_assemble_done(zio_t *zio); 2010fa9e4066Sahrens 2011e14bb325SJeff Bonwick static zio_gang_node_t * 2012e14bb325SJeff Bonwick zio_gang_node_alloc(zio_gang_node_t **gnpp) 2013e14bb325SJeff Bonwick { 2014e14bb325SJeff Bonwick zio_gang_node_t *gn; 2015fa9e4066Sahrens 2016e14bb325SJeff Bonwick ASSERT(*gnpp == NULL); 2017fa9e4066Sahrens 2018e14bb325SJeff Bonwick gn = kmem_zalloc(sizeof (*gn), KM_SLEEP); 2019e14bb325SJeff Bonwick gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE); 2020e14bb325SJeff Bonwick *gnpp = gn; 2021e14bb325SJeff Bonwick 2022e14bb325SJeff Bonwick return (gn); 2023fa9e4066Sahrens } 2024fa9e4066Sahrens 2025fa9e4066Sahrens static void 2026e14bb325SJeff Bonwick zio_gang_node_free(zio_gang_node_t **gnpp) 2027fa9e4066Sahrens { 2028e14bb325SJeff Bonwick zio_gang_node_t *gn = *gnpp; 2029fa9e4066Sahrens 2030e14bb325SJeff Bonwick for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) 2031e14bb325SJeff Bonwick ASSERT(gn->gn_child[g] == NULL); 2032e14bb325SJeff Bonwick 2033e14bb325SJeff Bonwick zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE); 2034e14bb325SJeff Bonwick kmem_free(gn, sizeof (*gn)); 2035e14bb325SJeff Bonwick *gnpp = NULL; 2036fa9e4066Sahrens } 2037fa9e4066Sahrens 2038e14bb325SJeff Bonwick static void 2039e14bb325SJeff Bonwick zio_gang_tree_free(zio_gang_node_t **gnpp) 2040fa9e4066Sahrens { 2041e14bb325SJeff Bonwick zio_gang_node_t *gn = *gnpp; 2042fa9e4066Sahrens 2043e14bb325SJeff Bonwick if (gn == NULL) 2044e14bb325SJeff Bonwick return; 2045fa9e4066Sahrens 2046e14bb325SJeff Bonwick for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) 2047e14bb325SJeff Bonwick zio_gang_tree_free(&gn->gn_child[g]); 2048fa9e4066Sahrens 2049e14bb325SJeff Bonwick zio_gang_node_free(gnpp); 2050fa9e4066Sahrens } 2051fa9e4066Sahrens 2052e14bb325SJeff Bonwick static void 2053f5383399SBill Moore zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp) 2054fa9e4066Sahrens { 2055e14bb325SJeff Bonwick zio_gang_node_t *gn = zio_gang_node_alloc(gnpp); 2056770499e1SDan Kimmel abd_t *gbh_abd = abd_get_from_buf(gn->gn_gbh, SPA_GANGBLOCKSIZE); 2057e14bb325SJeff Bonwick 2058f5383399SBill Moore ASSERT(gio->io_gang_leader == gio); 2059e14bb325SJeff Bonwick ASSERT(BP_IS_GANG(bp)); 2060fa9e4066Sahrens 2061770499e1SDan Kimmel zio_nowait(zio_read(gio, gio->io_spa, bp, gbh_abd, SPA_GANGBLOCKSIZE, 2062770499e1SDan Kimmel zio_gang_tree_assemble_done, gn, gio->io_priority, 2063770499e1SDan Kimmel ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark)); 2064e14bb325SJeff Bonwick } 2065fa9e4066Sahrens 2066e14bb325SJeff Bonwick static void 2067e14bb325SJeff Bonwick zio_gang_tree_assemble_done(zio_t *zio) 2068e14bb325SJeff Bonwick { 2069f5383399SBill Moore zio_t *gio = zio->io_gang_leader; 2070e14bb325SJeff Bonwick zio_gang_node_t *gn = zio->io_private; 2071e14bb325SJeff Bonwick blkptr_t *bp = zio->io_bp; 2072fa9e4066Sahrens 2073f5383399SBill Moore ASSERT(gio == zio_unique_parent(zio)); 2074b24ab676SJeff Bonwick ASSERT(zio->io_child_count == 0); 2075fa9e4066Sahrens 2076e14bb325SJeff Bonwick if (zio->io_error) 2077e14bb325SJeff Bonwick return; 2078fa9e4066Sahrens 2079770499e1SDan Kimmel /* this ABD was created from a linear buf in zio_gang_tree_assemble */ 2080e14bb325SJeff Bonwick if (BP_SHOULD_BYTESWAP(bp)) 2081770499e1SDan Kimmel byteswap_uint64_array(abd_to_buf(zio->io_abd), zio->io_size); 2082fa9e4066Sahrens 2083770499e1SDan Kimmel ASSERT3P(abd_to_buf(zio->io_abd), ==, gn->gn_gbh); 2084e14bb325SJeff Bonwick ASSERT(zio->io_size == SPA_GANGBLOCKSIZE); 20856e1f5caaSNeil Perrin ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC); 2086e05725b1Sbonwick 2087770499e1SDan Kimmel abd_put(zio->io_abd); 2088770499e1SDan Kimmel 2089e14bb325SJeff Bonwick for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) { 2090e14bb325SJeff Bonwick blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g]; 2091e14bb325SJeff Bonwick if (!BP_IS_GANG(gbp)) 2092e14bb325SJeff Bonwick continue; 2093f5383399SBill Moore zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]); 2094e14bb325SJeff Bonwick } 2095fa9e4066Sahrens } 2096fa9e4066Sahrens 2097e14bb325SJeff Bonwick static void 2098770499e1SDan Kimmel zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, abd_t *data, 2099770499e1SDan Kimmel uint64_t offset) 2100fa9e4066Sahrens { 2101f5383399SBill Moore zio_t *gio = pio->io_gang_leader; 2102e14bb325SJeff Bonwick zio_t *zio; 2103fa9e4066Sahrens 2104e14bb325SJeff Bonwick ASSERT(BP_IS_GANG(bp) == !!gn); 2105f5383399SBill Moore ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp)); 2106f5383399SBill Moore ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree); 2107fa9e4066Sahrens 2108e14bb325SJeff Bonwick /* 2109e14bb325SJeff Bonwick * If you're a gang header, your data is in gn->gn_gbh. 2110e14bb325SJeff Bonwick * If you're a gang member, your data is in 'data' and gn == NULL. 2111e14bb325SJeff Bonwick */ 2112770499e1SDan Kimmel zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data, offset); 2113fa9e4066Sahrens 2114e14bb325SJeff Bonwick if (gn != NULL) { 21156e1f5caaSNeil Perrin ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC); 2116fa9e4066Sahrens 2117e14bb325SJeff Bonwick for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) { 2118e14bb325SJeff Bonwick blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g]; 2119e14bb325SJeff Bonwick if (BP_IS_HOLE(gbp)) 2120e14bb325SJeff Bonwick continue; 2121770499e1SDan Kimmel zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data, 2122770499e1SDan Kimmel offset); 2123770499e1SDan Kimmel offset += BP_GET_PSIZE(gbp); 2124e14bb325SJeff Bonwick } 2125fa9e4066Sahrens } 2126fa9e4066Sahrens 2127f5383399SBill Moore if (gn == gio->io_gang_tree) 2128770499e1SDan Kimmel ASSERT3U(gio->io_size, ==, offset); 2129e05725b1Sbonwick 2130e14bb325SJeff Bonwick if (zio != pio) 2131e14bb325SJeff Bonwick zio_nowait(zio); 2132fa9e4066Sahrens } 2133fa9e4066Sahrens 2134e05725b1Sbonwick static int 2135e14bb325SJeff Bonwick zio_gang_assemble(zio_t *zio) 2136fa9e4066Sahrens { 2137e14bb325SJeff Bonwick blkptr_t *bp = zio->io_bp; 2138fa9e4066Sahrens 2139f5383399SBill Moore ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL); 2140f5383399SBill Moore ASSERT(zio->io_child_type > ZIO_CHILD_GANG); 2141f5383399SBill Moore 2142f5383399SBill Moore zio->io_gang_leader = zio; 2143fa9e4066Sahrens 2144e14bb325SJeff Bonwick zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree); 2145e05725b1Sbonwick 2146e05725b1Sbonwick return (ZIO_PIPELINE_CONTINUE); 2147fa9e4066Sahrens } 2148fa9e4066Sahrens 2149e05725b1Sbonwick static int 2150e14bb325SJeff Bonwick zio_gang_issue(zio_t *zio) 2151fa9e4066Sahrens { 2152e14bb325SJeff Bonwick blkptr_t *bp = zio->io_bp; 2153fa9e4066Sahrens 2154d6e1c446SGeorge Wilson if (zio_wait_for_children(zio, ZIO_CHILD_GANG_BIT, ZIO_WAIT_DONE)) { 2155e14bb325SJeff Bonwick return (ZIO_PIPELINE_STOP); 2156d6e1c446SGeorge Wilson } 2157fa9e4066Sahrens 2158f5383399SBill Moore ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio); 2159f5383399SBill Moore ASSERT(zio->io_child_type > ZIO_CHILD_GANG); 2160fa9e4066Sahrens 2161e14bb325SJeff Bonwick if (zio->io_child_error[ZIO_CHILD_GANG] == 0) 2162770499e1SDan Kimmel zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_abd, 2163770499e1SDan Kimmel 0); 2164e14bb325SJeff Bonwick else 2165f5383399SBill Moore zio_gang_tree_free(&zio->io_gang_tree); 2166fa9e4066Sahrens 2167e14bb325SJeff Bonwick zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 2168e05725b1Sbonwick 2169e05725b1Sbonwick return (ZIO_PIPELINE_CONTINUE); 2170fa9e4066Sahrens } 2171fa9e4066Sahrens 2172fa9e4066Sahrens static void 2173e14bb325SJeff Bonwick zio_write_gang_member_ready(zio_t *zio) 2174fa9e4066Sahrens { 2175a3f829aeSBill Moore zio_t *pio = zio_unique_parent(zio); 2176f5383399SBill Moore zio_t *gio = zio->io_gang_leader; 217744cd46caSbillm dva_t *cdva = zio->io_bp->blk_dva; 217844cd46caSbillm dva_t *pdva = pio->io_bp->blk_dva; 2179fa9e4066Sahrens uint64_t asize; 2180fa9e4066Sahrens 2181e14bb325SJeff Bonwick if (BP_IS_HOLE(zio->io_bp)) 2182e14bb325SJeff Bonwick return; 2183e14bb325SJeff Bonwick 2184e14bb325SJeff Bonwick ASSERT(BP_IS_HOLE(&zio->io_bp_orig)); 2185e14bb325SJeff Bonwick 2186e14bb325SJeff Bonwick ASSERT(zio->io_child_type == ZIO_CHILD_GANG); 2187b24ab676SJeff Bonwick ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies); 2188b24ab676SJeff Bonwick ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp)); 2189b24ab676SJeff Bonwick ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp)); 219044cd46caSbillm ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp)); 2191fa9e4066Sahrens 2192fa9e4066Sahrens mutex_enter(&pio->io_lock); 2193e14bb325SJeff Bonwick for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) { 219444cd46caSbillm ASSERT(DVA_GET_GANG(&pdva[d])); 219544cd46caSbillm asize = DVA_GET_ASIZE(&pdva[d]); 219644cd46caSbillm asize += DVA_GET_ASIZE(&cdva[d]); 219744cd46caSbillm DVA_SET_ASIZE(&pdva[d], asize); 219844cd46caSbillm } 2199fa9e4066Sahrens mutex_exit(&pio->io_lock); 2200fa9e4066Sahrens } 2201fa9e4066Sahrens 2202770499e1SDan Kimmel static void 2203770499e1SDan Kimmel zio_write_gang_done(zio_t *zio) 2204770499e1SDan Kimmel { 22057341a7deSBrad Lewis /* 22067341a7deSBrad Lewis * The io_abd field will be NULL for a zio with no data. The io_flags 22077341a7deSBrad Lewis * will initially have the ZIO_FLAG_NODATA bit flag set, but we can't 22087341a7deSBrad Lewis * check for it here as it is cleared in zio_ready. 22097341a7deSBrad Lewis */ 22107341a7deSBrad Lewis if (zio->io_abd != NULL) 22117341a7deSBrad Lewis abd_put(zio->io_abd); 2212770499e1SDan Kimmel } 2213770499e1SDan Kimmel 22140a4e9518Sgw static int 2215e14bb325SJeff Bonwick zio_write_gang_block(zio_t *pio) 2216fa9e4066Sahrens { 2217e14bb325SJeff Bonwick spa_t *spa = pio->io_spa; 22180f7643c7SGeorge Wilson metaslab_class_t *mc = spa_normal_class(spa); 2219e14bb325SJeff Bonwick blkptr_t *bp = pio->io_bp; 2220f5383399SBill Moore zio_t *gio = pio->io_gang_leader; 2221e14bb325SJeff Bonwick zio_t *zio; 2222e14bb325SJeff Bonwick zio_gang_node_t *gn, **gnpp; 2223fa9e4066Sahrens zio_gbh_phys_t *gbh; 2224770499e1SDan Kimmel abd_t *gbh_abd; 2225e14bb325SJeff Bonwick uint64_t txg = pio->io_txg; 2226e14bb325SJeff Bonwick uint64_t resid = pio->io_size; 2227e14bb325SJeff Bonwick uint64_t lsize; 2228b24ab676SJeff Bonwick int copies = gio->io_prop.zp_copies; 2229b24ab676SJeff Bonwick int gbh_copies = MIN(copies + 1, spa_max_replication(spa)); 2230e14bb325SJeff Bonwick zio_prop_t zp; 2231fa9e4066Sahrens int error; 22327341a7deSBrad Lewis boolean_t has_data = !(pio->io_flags & ZIO_FLAG_NODATA); 2233fa9e4066Sahrens 22340f7643c7SGeorge Wilson int flags = METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER; 22350f7643c7SGeorge Wilson if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) { 22360f7643c7SGeorge Wilson ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE); 22377341a7deSBrad Lewis ASSERT(has_data); 22380f7643c7SGeorge Wilson 22390f7643c7SGeorge Wilson flags |= METASLAB_ASYNC_ALLOC; 2240e914ace2STim Schumacher VERIFY(zfs_refcount_held(&mc->mc_alloc_slots[pio->io_allocator], 2241f78cdc34SPaul Dagnelie pio)); 22420f7643c7SGeorge Wilson 22430f7643c7SGeorge Wilson /* 22440f7643c7SGeorge Wilson * The logical zio has already placed a reservation for 22450f7643c7SGeorge Wilson * 'copies' allocation slots but gang blocks may require 22460f7643c7SGeorge Wilson * additional copies. These additional copies 22470f7643c7SGeorge Wilson * (i.e. gbh_copies - copies) are guaranteed to succeed 22480f7643c7SGeorge Wilson * since metaslab_class_throttle_reserve() always allows 22490f7643c7SGeorge Wilson * additional reservations for gang blocks. 22500f7643c7SGeorge Wilson */ 22510f7643c7SGeorge Wilson VERIFY(metaslab_class_throttle_reserve(mc, gbh_copies - copies, 2252f78cdc34SPaul Dagnelie pio->io_allocator, pio, flags)); 22530f7643c7SGeorge Wilson } 22540f7643c7SGeorge Wilson 22550f7643c7SGeorge Wilson error = metaslab_alloc(spa, mc, SPA_GANGBLOCKSIZE, 22568363e80aSGeorge Wilson bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp, flags, 2257f78cdc34SPaul Dagnelie &pio->io_alloc_list, pio, pio->io_allocator); 2258e05725b1Sbonwick if (error) { 22590f7643c7SGeorge Wilson if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) { 22600f7643c7SGeorge Wilson ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE); 22617341a7deSBrad Lewis ASSERT(has_data); 22620f7643c7SGeorge Wilson 22630f7643c7SGeorge Wilson /* 22640f7643c7SGeorge Wilson * If we failed to allocate the gang block header then 22650f7643c7SGeorge Wilson * we remove any additional allocation reservations that 22660f7643c7SGeorge Wilson * we placed here. The original reservation will 22670f7643c7SGeorge Wilson * be removed when the logical I/O goes to the ready 22680f7643c7SGeorge Wilson * stage. 22690f7643c7SGeorge Wilson */ 22700f7643c7SGeorge Wilson metaslab_class_throttle_unreserve(mc, 2271f78cdc34SPaul Dagnelie gbh_copies - copies, pio->io_allocator, pio); 22720f7643c7SGeorge Wilson } 2273e14bb325SJeff Bonwick pio->io_error = error; 2274e05725b1Sbonwick return (ZIO_PIPELINE_CONTINUE); 2275e05725b1Sbonwick } 2276fa9e4066Sahrens 2277f5383399SBill Moore if (pio == gio) { 2278f5383399SBill Moore gnpp = &gio->io_gang_tree; 2279e14bb325SJeff Bonwick } else { 2280e14bb325SJeff Bonwick gnpp = pio->io_private; 2281e14bb325SJeff Bonwick ASSERT(pio->io_ready == zio_write_gang_member_ready); 2282fa9e4066Sahrens } 2283fa9e4066Sahrens 2284e14bb325SJeff Bonwick gn = zio_gang_node_alloc(gnpp); 2285e14bb325SJeff Bonwick gbh = gn->gn_gbh; 2286e14bb325SJeff Bonwick bzero(gbh, SPA_GANGBLOCKSIZE); 2287770499e1SDan Kimmel gbh_abd = abd_get_from_buf(gbh, SPA_GANGBLOCKSIZE); 2288fa9e4066Sahrens 2289e14bb325SJeff Bonwick /* 2290e14bb325SJeff Bonwick * Create the gang header. 2291e14bb325SJeff Bonwick */ 2292770499e1SDan Kimmel zio = zio_rewrite(pio, spa, txg, bp, gbh_abd, SPA_GANGBLOCKSIZE, 2293770499e1SDan Kimmel zio_write_gang_done, NULL, pio->io_priority, 2294770499e1SDan Kimmel ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark); 2295fa9e4066Sahrens 2296e14bb325SJeff Bonwick /* 2297e14bb325SJeff Bonwick * Create and nowait the gang children. 2298e14bb325SJeff Bonwick */ 2299e14bb325SJeff Bonwick for (int g = 0; resid != 0; resid -= lsize, g++) { 2300e14bb325SJeff Bonwick lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g), 2301e14bb325SJeff Bonwick SPA_MINBLOCKSIZE); 2302e14bb325SJeff Bonwick ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid); 2303e14bb325SJeff Bonwick 2304f5383399SBill Moore zp.zp_checksum = gio->io_prop.zp_checksum; 2305e14bb325SJeff Bonwick zp.zp_compress = ZIO_COMPRESS_OFF; 2306e14bb325SJeff Bonwick zp.zp_type = DMU_OT_NONE; 2307e14bb325SJeff Bonwick zp.zp_level = 0; 2308b24ab676SJeff Bonwick zp.zp_copies = gio->io_prop.zp_copies; 230980901aeaSGeorge Wilson zp.zp_dedup = B_FALSE; 231080901aeaSGeorge Wilson zp.zp_dedup_verify = B_FALSE; 231180901aeaSGeorge Wilson zp.zp_nopwrite = B_FALSE; 2312e14bb325SJeff Bonwick 23130f7643c7SGeorge Wilson zio_t *cio = zio_write(zio, spa, txg, &gbh->zg_blkptr[g], 23147341a7deSBrad Lewis has_data ? abd_get_offset(pio->io_abd, pio->io_size - 23157341a7deSBrad Lewis resid) : NULL, lsize, lsize, &zp, 23167341a7deSBrad Lewis zio_write_gang_member_ready, NULL, NULL, 2317770499e1SDan Kimmel zio_write_gang_done, &gn->gn_child[g], pio->io_priority, 23180f7643c7SGeorge Wilson ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark); 23190f7643c7SGeorge Wilson 23200f7643c7SGeorge Wilson if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) { 23210f7643c7SGeorge Wilson ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE); 23227341a7deSBrad Lewis ASSERT(has_data); 23230f7643c7SGeorge Wilson 23240f7643c7SGeorge Wilson /* 23250f7643c7SGeorge Wilson * Gang children won't throttle but we should 23260f7643c7SGeorge Wilson * account for their work, so reserve an allocation 23270f7643c7SGeorge Wilson * slot for them here. 23280f7643c7SGeorge Wilson */ 23290f7643c7SGeorge Wilson VERIFY(metaslab_class_throttle_reserve(mc, 2330f78cdc34SPaul Dagnelie zp.zp_copies, cio->io_allocator, cio, flags)); 23310f7643c7SGeorge Wilson } 23320f7643c7SGeorge Wilson zio_nowait(cio); 2333e14bb325SJeff Bonwick } 2334e05725b1Sbonwick 233544cd46caSbillm /* 2336e14bb325SJeff Bonwick * Set pio's pipeline to just wait for zio to finish. 233744cd46caSbillm */ 2338e14bb325SJeff Bonwick pio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 2339e14bb325SJeff Bonwick 2340e14bb325SJeff Bonwick zio_nowait(zio); 2341e14bb325SJeff Bonwick 2342e14bb325SJeff Bonwick return (ZIO_PIPELINE_CONTINUE); 2343fa9e4066Sahrens } 2344fa9e4066Sahrens 234580901aeaSGeorge Wilson /* 234645818ee1SMatthew Ahrens * The zio_nop_write stage in the pipeline determines if allocating a 234745818ee1SMatthew Ahrens * new bp is necessary. The nopwrite feature can handle writes in 234845818ee1SMatthew Ahrens * either syncing or open context (i.e. zil writes) and as a result is 234945818ee1SMatthew Ahrens * mutually exclusive with dedup. 235045818ee1SMatthew Ahrens * 235145818ee1SMatthew Ahrens * By leveraging a cryptographically secure checksum, such as SHA256, we 235245818ee1SMatthew Ahrens * can compare the checksums of the new data and the old to determine if 235345818ee1SMatthew Ahrens * allocating a new block is required. Note that our requirements for 235445818ee1SMatthew Ahrens * cryptographic strength are fairly weak: there can't be any accidental 235545818ee1SMatthew Ahrens * hash collisions, but we don't need to be secure against intentional 235645818ee1SMatthew Ahrens * (malicious) collisions. To trigger a nopwrite, you have to be able 235745818ee1SMatthew Ahrens * to write the file to begin with, and triggering an incorrect (hash 235845818ee1SMatthew Ahrens * collision) nopwrite is no worse than simply writing to the file. 235945818ee1SMatthew Ahrens * That said, there are no known attacks against the checksum algorithms 236045818ee1SMatthew Ahrens * used for nopwrite, assuming that the salt and the checksums 236145818ee1SMatthew Ahrens * themselves remain secret. 236280901aeaSGeorge Wilson */ 236380901aeaSGeorge Wilson static int 236480901aeaSGeorge Wilson zio_nop_write(zio_t *zio) 236580901aeaSGeorge Wilson { 236680901aeaSGeorge Wilson blkptr_t *bp = zio->io_bp; 236780901aeaSGeorge Wilson blkptr_t *bp_orig = &zio->io_bp_orig; 236880901aeaSGeorge Wilson zio_prop_t *zp = &zio->io_prop; 236980901aeaSGeorge Wilson 237080901aeaSGeorge Wilson ASSERT(BP_GET_LEVEL(bp) == 0); 237180901aeaSGeorge Wilson ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE)); 237280901aeaSGeorge Wilson ASSERT(zp->zp_nopwrite); 237380901aeaSGeorge Wilson ASSERT(!zp->zp_dedup); 237480901aeaSGeorge Wilson ASSERT(zio->io_bp_override == NULL); 237580901aeaSGeorge Wilson ASSERT(IO_IS_ALLOCATING(zio)); 237680901aeaSGeorge Wilson 237780901aeaSGeorge Wilson /* 237880901aeaSGeorge Wilson * Check to see if the original bp and the new bp have matching 237980901aeaSGeorge Wilson * characteristics (i.e. same checksum, compression algorithms, etc). 238080901aeaSGeorge Wilson * If they don't then just continue with the pipeline which will 238180901aeaSGeorge Wilson * allocate a new bp. 238280901aeaSGeorge Wilson */ 238380901aeaSGeorge Wilson if (BP_IS_HOLE(bp_orig) || 238445818ee1SMatthew Ahrens !(zio_checksum_table[BP_GET_CHECKSUM(bp)].ci_flags & 238545818ee1SMatthew Ahrens ZCHECKSUM_FLAG_NOPWRITE) || 238680901aeaSGeorge Wilson BP_GET_CHECKSUM(bp) != BP_GET_CHECKSUM(bp_orig) || 238780901aeaSGeorge Wilson BP_GET_COMPRESS(bp) != BP_GET_COMPRESS(bp_orig) || 238880901aeaSGeorge Wilson BP_GET_DEDUP(bp) != BP_GET_DEDUP(bp_orig) || 238980901aeaSGeorge Wilson zp->zp_copies != BP_GET_NDVAS(bp_orig)) 239080901aeaSGeorge Wilson return (ZIO_PIPELINE_CONTINUE); 239180901aeaSGeorge Wilson 239280901aeaSGeorge Wilson /* 239380901aeaSGeorge Wilson * If the checksums match then reset the pipeline so that we 239480901aeaSGeorge Wilson * avoid allocating a new bp and issuing any I/O. 239580901aeaSGeorge Wilson */ 239680901aeaSGeorge Wilson if (ZIO_CHECKSUM_EQUAL(bp->blk_cksum, bp_orig->blk_cksum)) { 239745818ee1SMatthew Ahrens ASSERT(zio_checksum_table[zp->zp_checksum].ci_flags & 239845818ee1SMatthew Ahrens ZCHECKSUM_FLAG_NOPWRITE); 239980901aeaSGeorge Wilson ASSERT3U(BP_GET_PSIZE(bp), ==, BP_GET_PSIZE(bp_orig)); 240080901aeaSGeorge Wilson ASSERT3U(BP_GET_LSIZE(bp), ==, BP_GET_LSIZE(bp_orig)); 240180901aeaSGeorge Wilson ASSERT(zp->zp_compress != ZIO_COMPRESS_OFF); 240280901aeaSGeorge Wilson ASSERT(bcmp(&bp->blk_prop, &bp_orig->blk_prop, 240380901aeaSGeorge Wilson sizeof (uint64_t)) == 0); 240480901aeaSGeorge Wilson 240580901aeaSGeorge Wilson *bp = *bp_orig; 240680901aeaSGeorge Wilson zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 240780901aeaSGeorge Wilson zio->io_flags |= ZIO_FLAG_NOPWRITE; 240880901aeaSGeorge Wilson } 240980901aeaSGeorge Wilson 241080901aeaSGeorge Wilson return (ZIO_PIPELINE_CONTINUE); 241180901aeaSGeorge Wilson } 241280901aeaSGeorge Wilson 2413fa9e4066Sahrens /* 2414fa9e4066Sahrens * ========================================================================== 2415b24ab676SJeff Bonwick * Dedup 2416fa9e4066Sahrens * ========================================================================== 2417fa9e4066Sahrens */ 2418b24ab676SJeff Bonwick static void 2419b24ab676SJeff Bonwick zio_ddt_child_read_done(zio_t *zio) 2420b24ab676SJeff Bonwick { 2421b24ab676SJeff Bonwick blkptr_t *bp = zio->io_bp; 2422b24ab676SJeff Bonwick ddt_entry_t *dde = zio->io_private; 2423b24ab676SJeff Bonwick ddt_phys_t *ddp; 2424b24ab676SJeff Bonwick zio_t *pio = zio_unique_parent(zio); 2425b24ab676SJeff Bonwick 2426b24ab676SJeff Bonwick mutex_enter(&pio->io_lock); 2427b24ab676SJeff Bonwick ddp = ddt_phys_select(dde, bp); 2428b24ab676SJeff Bonwick if (zio->io_error == 0) 2429b24ab676SJeff Bonwick ddt_phys_clear(ddp); /* this ddp doesn't need repair */ 2430770499e1SDan Kimmel 2431770499e1SDan Kimmel if (zio->io_error == 0 && dde->dde_repair_abd == NULL) 2432770499e1SDan Kimmel dde->dde_repair_abd = zio->io_abd; 2433b24ab676SJeff Bonwick else 2434770499e1SDan Kimmel abd_free(zio->io_abd); 2435b24ab676SJeff Bonwick mutex_exit(&pio->io_lock); 2436b24ab676SJeff Bonwick } 2437b24ab676SJeff Bonwick 2438b24ab676SJeff Bonwick static int 2439b24ab676SJeff Bonwick zio_ddt_read_start(zio_t *zio) 2440b24ab676SJeff Bonwick { 2441b24ab676SJeff Bonwick blkptr_t *bp = zio->io_bp; 2442b24ab676SJeff Bonwick 2443b24ab676SJeff Bonwick ASSERT(BP_GET_DEDUP(bp)); 2444b24ab676SJeff Bonwick ASSERT(BP_GET_PSIZE(bp) == zio->io_size); 2445b24ab676SJeff Bonwick ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 2446b24ab676SJeff Bonwick 2447b24ab676SJeff Bonwick if (zio->io_child_error[ZIO_CHILD_DDT]) { 2448b24ab676SJeff Bonwick ddt_t *ddt = ddt_select(zio->io_spa, bp); 2449b24ab676SJeff Bonwick ddt_entry_t *dde = ddt_repair_start(ddt, bp); 2450b24ab676SJeff Bonwick ddt_phys_t *ddp = dde->dde_phys; 2451b24ab676SJeff Bonwick ddt_phys_t *ddp_self = ddt_phys_select(dde, bp); 2452b24ab676SJeff Bonwick blkptr_t blk; 2453b24ab676SJeff Bonwick 2454b24ab676SJeff Bonwick ASSERT(zio->io_vsd == NULL); 2455b24ab676SJeff Bonwick zio->io_vsd = dde; 2456b24ab676SJeff Bonwick 2457b24ab676SJeff Bonwick if (ddp_self == NULL) 2458b24ab676SJeff Bonwick return (ZIO_PIPELINE_CONTINUE); 2459b24ab676SJeff Bonwick 2460b24ab676SJeff Bonwick for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) { 2461b24ab676SJeff Bonwick if (ddp->ddp_phys_birth == 0 || ddp == ddp_self) 2462b24ab676SJeff Bonwick continue; 2463bbfd46c4SJeff Bonwick ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp, 2464bbfd46c4SJeff Bonwick &blk); 2465b24ab676SJeff Bonwick zio_nowait(zio_read(zio, zio->io_spa, &blk, 2466770499e1SDan Kimmel abd_alloc_for_io(zio->io_size, B_TRUE), 2467770499e1SDan Kimmel zio->io_size, zio_ddt_child_read_done, dde, 2468770499e1SDan Kimmel zio->io_priority, ZIO_DDT_CHILD_FLAGS(zio) | 2469770499e1SDan Kimmel ZIO_FLAG_DONT_PROPAGATE, &zio->io_bookmark)); 2470b24ab676SJeff Bonwick } 2471b24ab676SJeff Bonwick return (ZIO_PIPELINE_CONTINUE); 2472b24ab676SJeff Bonwick } 2473b24ab676SJeff Bonwick 2474b24ab676SJeff Bonwick zio_nowait(zio_read(zio, zio->io_spa, bp, 2475770499e1SDan Kimmel zio->io_abd, zio->io_size, NULL, NULL, zio->io_priority, 2476b24ab676SJeff Bonwick ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark)); 2477b24ab676SJeff Bonwick 2478b24ab676SJeff Bonwick return (ZIO_PIPELINE_CONTINUE); 2479b24ab676SJeff Bonwick } 2480e14bb325SJeff Bonwick 2481b24ab676SJeff Bonwick static int 2482b24ab676SJeff Bonwick zio_ddt_read_done(zio_t *zio) 2483b24ab676SJeff Bonwick { 2484b24ab676SJeff Bonwick blkptr_t *bp = zio->io_bp; 2485b24ab676SJeff Bonwick 2486d6e1c446SGeorge Wilson if (zio_wait_for_children(zio, ZIO_CHILD_DDT_BIT, ZIO_WAIT_DONE)) { 2487b24ab676SJeff Bonwick return (ZIO_PIPELINE_STOP); 2488d6e1c446SGeorge Wilson } 2489b24ab676SJeff Bonwick 2490b24ab676SJeff Bonwick ASSERT(BP_GET_DEDUP(bp)); 2491b24ab676SJeff Bonwick ASSERT(BP_GET_PSIZE(bp) == zio->io_size); 2492b24ab676SJeff Bonwick ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 2493b24ab676SJeff Bonwick 2494b24ab676SJeff Bonwick if (zio->io_child_error[ZIO_CHILD_DDT]) { 2495b24ab676SJeff Bonwick ddt_t *ddt = ddt_select(zio->io_spa, bp); 2496b24ab676SJeff Bonwick ddt_entry_t *dde = zio->io_vsd; 2497b24ab676SJeff Bonwick if (ddt == NULL) { 2498b16da2e2SGeorge Wilson ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE); 2499b24ab676SJeff Bonwick return (ZIO_PIPELINE_CONTINUE); 2500b24ab676SJeff Bonwick } 2501b24ab676SJeff Bonwick if (dde == NULL) { 2502b24ab676SJeff Bonwick zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1; 250335a5a358SJonathan Adams zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE); 2504b24ab676SJeff Bonwick return (ZIO_PIPELINE_STOP); 2505b24ab676SJeff Bonwick } 2506770499e1SDan Kimmel if (dde->dde_repair_abd != NULL) { 2507770499e1SDan Kimmel abd_copy(zio->io_abd, dde->dde_repair_abd, 2508770499e1SDan Kimmel zio->io_size); 2509b24ab676SJeff Bonwick zio->io_child_error[ZIO_CHILD_DDT] = 0; 2510b24ab676SJeff Bonwick } 2511b24ab676SJeff Bonwick ddt_repair_done(ddt, dde); 2512b24ab676SJeff Bonwick zio->io_vsd = NULL; 2513b24ab676SJeff Bonwick } 2514b24ab676SJeff Bonwick 2515b24ab676SJeff Bonwick ASSERT(zio->io_vsd == NULL); 2516b24ab676SJeff Bonwick 2517b24ab676SJeff Bonwick return (ZIO_PIPELINE_CONTINUE); 2518b24ab676SJeff Bonwick } 2519b24ab676SJeff Bonwick 2520b24ab676SJeff Bonwick static boolean_t 2521b24ab676SJeff Bonwick zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde) 2522b24ab676SJeff Bonwick { 2523b24ab676SJeff Bonwick spa_t *spa = zio->io_spa; 25245602294fSDan Kimmel boolean_t do_raw = (zio->io_flags & ZIO_FLAG_RAW); 25255602294fSDan Kimmel 25265602294fSDan Kimmel /* We should never get a raw, override zio */ 25275602294fSDan Kimmel ASSERT(!(zio->io_bp_override && do_raw)); 2528b24ab676SJeff Bonwick 2529b24ab676SJeff Bonwick /* 2530b24ab676SJeff Bonwick * Note: we compare the original data, not the transformed data, 2531b24ab676SJeff Bonwick * because when zio->io_bp is an override bp, we will not have 2532b24ab676SJeff Bonwick * pushed the I/O transforms. That's an important optimization 2533b24ab676SJeff Bonwick * because otherwise we'd compress/encrypt all dmu_sync() data twice. 2534b24ab676SJeff Bonwick */ 2535b24ab676SJeff Bonwick for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) { 2536b24ab676SJeff Bonwick zio_t *lio = dde->dde_lead_zio[p]; 2537b24ab676SJeff Bonwick 2538b24ab676SJeff Bonwick if (lio != NULL) { 2539b24ab676SJeff Bonwick return (lio->io_orig_size != zio->io_orig_size || 2540770499e1SDan Kimmel abd_cmp(zio->io_orig_abd, lio->io_orig_abd, 2541b24ab676SJeff Bonwick zio->io_orig_size) != 0); 2542b24ab676SJeff Bonwick } 2543b24ab676SJeff Bonwick } 2544b24ab676SJeff Bonwick 2545b24ab676SJeff Bonwick for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) { 2546b24ab676SJeff Bonwick ddt_phys_t *ddp = &dde->dde_phys[p]; 2547b24ab676SJeff Bonwick 2548b24ab676SJeff Bonwick if (ddp->ddp_phys_birth != 0) { 2549b24ab676SJeff Bonwick arc_buf_t *abuf = NULL; 25507adb730bSGeorge Wilson arc_flags_t aflags = ARC_FLAG_WAIT; 25515602294fSDan Kimmel int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE; 2552b24ab676SJeff Bonwick blkptr_t blk = *zio->io_bp; 2553b24ab676SJeff Bonwick int error; 2554b24ab676SJeff Bonwick 2555b24ab676SJeff Bonwick ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth); 2556b24ab676SJeff Bonwick 2557b24ab676SJeff Bonwick ddt_exit(ddt); 2558b24ab676SJeff Bonwick 25595602294fSDan Kimmel /* 25605602294fSDan Kimmel * Intuitively, it would make more sense to compare 2561770499e1SDan Kimmel * io_abd than io_orig_abd in the raw case since you 25625602294fSDan Kimmel * don't want to look at any transformations that have 25635602294fSDan Kimmel * happened to the data. However, for raw I/Os the 2564770499e1SDan Kimmel * data will actually be the same in io_abd and 2565770499e1SDan Kimmel * io_orig_abd, so all we have to do is issue this as 25665602294fSDan Kimmel * a raw ARC read. 25675602294fSDan Kimmel */ 25685602294fSDan Kimmel if (do_raw) { 25695602294fSDan Kimmel zio_flags |= ZIO_FLAG_RAW; 25705602294fSDan Kimmel ASSERT3U(zio->io_size, ==, zio->io_orig_size); 2571770499e1SDan Kimmel ASSERT0(abd_cmp(zio->io_abd, zio->io_orig_abd, 25725602294fSDan Kimmel zio->io_size)); 25735602294fSDan Kimmel ASSERT3P(zio->io_transform_stack, ==, NULL); 25745602294fSDan Kimmel } 25755602294fSDan Kimmel 25761b912ec7SGeorge Wilson error = arc_read(NULL, spa, &blk, 2577b24ab676SJeff Bonwick arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ, 25785602294fSDan Kimmel zio_flags, &aflags, &zio->io_bookmark); 2579b24ab676SJeff Bonwick 2580b24ab676SJeff Bonwick if (error == 0) { 2581b24ab676SJeff Bonwick if (arc_buf_size(abuf) != zio->io_orig_size || 2582770499e1SDan Kimmel abd_cmp_buf(zio->io_orig_abd, abuf->b_data, 2583b24ab676SJeff Bonwick zio->io_orig_size) != 0) 2584be6fd75aSMatthew Ahrens error = SET_ERROR(EEXIST); 2585dcbf3bd6SGeorge Wilson arc_buf_destroy(abuf, &abuf); 2586b24ab676SJeff Bonwick } 2587b24ab676SJeff Bonwick 2588b24ab676SJeff Bonwick ddt_enter(ddt); 2589b24ab676SJeff Bonwick return (error != 0); 2590b24ab676SJeff Bonwick } 2591b24ab676SJeff Bonwick } 2592b24ab676SJeff Bonwick 2593b24ab676SJeff Bonwick return (B_FALSE); 2594b24ab676SJeff Bonwick } 2595b24ab676SJeff Bonwick 2596b24ab676SJeff Bonwick static void 2597b24ab676SJeff Bonwick zio_ddt_child_write_ready(zio_t *zio) 2598b24ab676SJeff Bonwick { 2599b24ab676SJeff Bonwick int p = zio->io_prop.zp_copies; 2600b24ab676SJeff Bonwick ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp); 2601b24ab676SJeff Bonwick ddt_entry_t *dde = zio->io_private; 2602b24ab676SJeff Bonwick ddt_phys_t *ddp = &dde->dde_phys[p]; 2603b24ab676SJeff Bonwick zio_t *pio; 2604b24ab676SJeff Bonwick 2605b24ab676SJeff Bonwick if (zio->io_error) 2606b24ab676SJeff Bonwick return; 2607b24ab676SJeff Bonwick 2608b24ab676SJeff Bonwick ddt_enter(ddt); 2609b24ab676SJeff Bonwick 2610b24ab676SJeff Bonwick ASSERT(dde->dde_lead_zio[p] == zio); 2611b24ab676SJeff Bonwick 2612b24ab676SJeff Bonwick ddt_phys_fill(ddp, zio->io_bp); 2613b24ab676SJeff Bonwick 26140f7643c7SGeorge Wilson zio_link_t *zl = NULL; 26150f7643c7SGeorge Wilson while ((pio = zio_walk_parents(zio, &zl)) != NULL) 2616b24ab676SJeff Bonwick ddt_bp_fill(ddp, pio->io_bp, zio->io_txg); 2617b24ab676SJeff Bonwick 2618b24ab676SJeff Bonwick ddt_exit(ddt); 2619b24ab676SJeff Bonwick } 2620b24ab676SJeff Bonwick 2621b24ab676SJeff Bonwick static void 2622b24ab676SJeff Bonwick zio_ddt_child_write_done(zio_t *zio) 2623b24ab676SJeff Bonwick { 2624b24ab676SJeff Bonwick int p = zio->io_prop.zp_copies; 2625b24ab676SJeff Bonwick ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp); 2626b24ab676SJeff Bonwick ddt_entry_t *dde = zio->io_private; 2627b24ab676SJeff Bonwick ddt_phys_t *ddp = &dde->dde_phys[p]; 2628b24ab676SJeff Bonwick 2629b24ab676SJeff Bonwick ddt_enter(ddt); 2630b24ab676SJeff Bonwick 2631b24ab676SJeff Bonwick ASSERT(ddp->ddp_refcnt == 0); 2632b24ab676SJeff Bonwick ASSERT(dde->dde_lead_zio[p] == zio); 2633b24ab676SJeff Bonwick dde->dde_lead_zio[p] = NULL; 2634b24ab676SJeff Bonwick 2635b24ab676SJeff Bonwick if (zio->io_error == 0) { 26360f7643c7SGeorge Wilson zio_link_t *zl = NULL; 26370f7643c7SGeorge Wilson while (zio_walk_parents(zio, &zl) != NULL) 2638b24ab676SJeff Bonwick ddt_phys_addref(ddp); 2639b24ab676SJeff Bonwick } else { 2640b24ab676SJeff Bonwick ddt_phys_clear(ddp); 2641b24ab676SJeff Bonwick } 2642b24ab676SJeff Bonwick 2643b24ab676SJeff Bonwick ddt_exit(ddt); 2644b24ab676SJeff Bonwick } 2645b24ab676SJeff Bonwick 2646b24ab676SJeff Bonwick static void 2647b24ab676SJeff Bonwick zio_ddt_ditto_write_done(zio_t *zio) 2648b24ab676SJeff Bonwick { 2649b24ab676SJeff Bonwick int p = DDT_PHYS_DITTO; 2650b24ab676SJeff Bonwick zio_prop_t *zp = &zio->io_prop; 2651b24ab676SJeff Bonwick blkptr_t *bp = zio->io_bp; 2652b24ab676SJeff Bonwick ddt_t *ddt = ddt_select(zio->io_spa, bp); 2653b24ab676SJeff Bonwick ddt_entry_t *dde = zio->io_private; 2654b24ab676SJeff Bonwick ddt_phys_t *ddp = &dde->dde_phys[p]; 2655b24ab676SJeff Bonwick ddt_key_t *ddk = &dde->dde_key; 2656b24ab676SJeff Bonwick 2657b24ab676SJeff Bonwick ddt_enter(ddt); 2658b24ab676SJeff Bonwick 2659b24ab676SJeff Bonwick ASSERT(ddp->ddp_refcnt == 0); 2660b24ab676SJeff Bonwick ASSERT(dde->dde_lead_zio[p] == zio); 2661b24ab676SJeff Bonwick dde->dde_lead_zio[p] = NULL; 2662b24ab676SJeff Bonwick 2663b24ab676SJeff Bonwick if (zio->io_error == 0) { 2664b24ab676SJeff Bonwick ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum)); 2665b24ab676SJeff Bonwick ASSERT(zp->zp_copies < SPA_DVAS_PER_BP); 2666b24ab676SJeff Bonwick ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp)); 2667b24ab676SJeff Bonwick if (ddp->ddp_phys_birth != 0) 2668b24ab676SJeff Bonwick ddt_phys_free(ddt, ddk, ddp, zio->io_txg); 2669b24ab676SJeff Bonwick ddt_phys_fill(ddp, bp); 2670b24ab676SJeff Bonwick } 2671b24ab676SJeff Bonwick 2672b24ab676SJeff Bonwick ddt_exit(ddt); 2673b24ab676SJeff Bonwick } 2674b24ab676SJeff Bonwick 2675b24ab676SJeff Bonwick static int 2676b24ab676SJeff Bonwick zio_ddt_write(zio_t *zio) 2677b24ab676SJeff Bonwick { 2678b24ab676SJeff Bonwick spa_t *spa = zio->io_spa; 2679b24ab676SJeff Bonwick blkptr_t *bp = zio->io_bp; 2680b24ab676SJeff Bonwick uint64_t txg = zio->io_txg; 2681b24ab676SJeff Bonwick zio_prop_t *zp = &zio->io_prop; 2682b24ab676SJeff Bonwick int p = zp->zp_copies; 2683b24ab676SJeff Bonwick int ditto_copies; 2684b24ab676SJeff Bonwick zio_t *cio = NULL; 2685b24ab676SJeff Bonwick zio_t *dio = NULL; 2686b24ab676SJeff Bonwick ddt_t *ddt = ddt_select(spa, bp); 2687b24ab676SJeff Bonwick ddt_entry_t *dde; 2688b24ab676SJeff Bonwick ddt_phys_t *ddp; 2689b24ab676SJeff Bonwick 2690b24ab676SJeff Bonwick ASSERT(BP_GET_DEDUP(bp)); 2691b24ab676SJeff Bonwick ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum); 2692b24ab676SJeff Bonwick ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override); 26935602294fSDan Kimmel ASSERT(!(zio->io_bp_override && (zio->io_flags & ZIO_FLAG_RAW))); 2694b24ab676SJeff Bonwick 2695b24ab676SJeff Bonwick ddt_enter(ddt); 2696b24ab676SJeff Bonwick dde = ddt_lookup(ddt, bp, B_TRUE); 2697b24ab676SJeff Bonwick ddp = &dde->dde_phys[p]; 2698b24ab676SJeff Bonwick 2699b24ab676SJeff Bonwick if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) { 2700b24ab676SJeff Bonwick /* 2701b24ab676SJeff Bonwick * If we're using a weak checksum, upgrade to a strong checksum 2702b24ab676SJeff Bonwick * and try again. If we're already using a strong checksum, 2703b24ab676SJeff Bonwick * we can't resolve it, so just convert to an ordinary write. 2704b24ab676SJeff Bonwick * (And automatically e-mail a paper to Nature?) 2705b24ab676SJeff Bonwick */ 270645818ee1SMatthew Ahrens if (!(zio_checksum_table[zp->zp_checksum].ci_flags & 270745818ee1SMatthew Ahrens ZCHECKSUM_FLAG_DEDUP)) { 2708b24ab676SJeff Bonwick zp->zp_checksum = spa_dedup_checksum(spa); 2709b24ab676SJeff Bonwick zio_pop_transforms(zio); 2710b24ab676SJeff Bonwick zio->io_stage = ZIO_STAGE_OPEN; 2711b24ab676SJeff Bonwick BP_ZERO(bp); 2712b24ab676SJeff Bonwick } else { 271380901aeaSGeorge Wilson zp->zp_dedup = B_FALSE; 27145602294fSDan Kimmel BP_SET_DEDUP(bp, B_FALSE); 2715b24ab676SJeff Bonwick } 27165602294fSDan Kimmel ASSERT(!BP_GET_DEDUP(bp)); 2717b24ab676SJeff Bonwick zio->io_pipeline = ZIO_WRITE_PIPELINE; 2718b24ab676SJeff Bonwick ddt_exit(ddt); 2719b24ab676SJeff Bonwick return (ZIO_PIPELINE_CONTINUE); 2720b24ab676SJeff Bonwick } 2721b24ab676SJeff Bonwick 2722b24ab676SJeff Bonwick ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp); 2723b24ab676SJeff Bonwick ASSERT(ditto_copies < SPA_DVAS_PER_BP); 2724b24ab676SJeff Bonwick 2725b24ab676SJeff Bonwick if (ditto_copies > ddt_ditto_copies_present(dde) && 2726b24ab676SJeff Bonwick dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) { 2727b24ab676SJeff Bonwick zio_prop_t czp = *zp; 2728b24ab676SJeff Bonwick 2729b24ab676SJeff Bonwick czp.zp_copies = ditto_copies; 2730b24ab676SJeff Bonwick 2731b24ab676SJeff Bonwick /* 2732b24ab676SJeff Bonwick * If we arrived here with an override bp, we won't have run 2733b24ab676SJeff Bonwick * the transform stack, so we won't have the data we need to 2734b24ab676SJeff Bonwick * generate a child i/o. So, toss the override bp and restart. 2735b24ab676SJeff Bonwick * This is safe, because using the override bp is just an 2736b24ab676SJeff Bonwick * optimization; and it's rare, so the cost doesn't matter. 2737b24ab676SJeff Bonwick */ 2738b24ab676SJeff Bonwick if (zio->io_bp_override) { 2739b24ab676SJeff Bonwick zio_pop_transforms(zio); 2740b24ab676SJeff Bonwick zio->io_stage = ZIO_STAGE_OPEN; 2741b24ab676SJeff Bonwick zio->io_pipeline = ZIO_WRITE_PIPELINE; 2742b24ab676SJeff Bonwick zio->io_bp_override = NULL; 2743b24ab676SJeff Bonwick BP_ZERO(bp); 2744b24ab676SJeff Bonwick ddt_exit(ddt); 2745b24ab676SJeff Bonwick return (ZIO_PIPELINE_CONTINUE); 2746b24ab676SJeff Bonwick } 2747b24ab676SJeff Bonwick 2748770499e1SDan Kimmel dio = zio_write(zio, spa, txg, bp, zio->io_orig_abd, 27495602294fSDan Kimmel zio->io_orig_size, zio->io_orig_size, &czp, NULL, NULL, 27508df0bcf0SPaul Dagnelie NULL, zio_ddt_ditto_write_done, dde, zio->io_priority, 2751b24ab676SJeff Bonwick ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark); 2752b24ab676SJeff Bonwick 2753770499e1SDan Kimmel zio_push_transform(dio, zio->io_abd, zio->io_size, 0, NULL); 2754b24ab676SJeff Bonwick dde->dde_lead_zio[DDT_PHYS_DITTO] = dio; 2755b24ab676SJeff Bonwick } 2756b24ab676SJeff Bonwick 2757b24ab676SJeff Bonwick if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) { 2758b24ab676SJeff Bonwick if (ddp->ddp_phys_birth != 0) 2759b24ab676SJeff Bonwick ddt_bp_fill(ddp, bp, txg); 2760b24ab676SJeff Bonwick if (dde->dde_lead_zio[p] != NULL) 2761b24ab676SJeff Bonwick zio_add_child(zio, dde->dde_lead_zio[p]); 2762b24ab676SJeff Bonwick else 2763b24ab676SJeff Bonwick ddt_phys_addref(ddp); 2764b24ab676SJeff Bonwick } else if (zio->io_bp_override) { 2765b24ab676SJeff Bonwick ASSERT(bp->blk_birth == txg); 2766b24ab676SJeff Bonwick ASSERT(BP_EQUAL(bp, zio->io_bp_override)); 2767b24ab676SJeff Bonwick ddt_phys_fill(ddp, bp); 2768b24ab676SJeff Bonwick ddt_phys_addref(ddp); 2769b24ab676SJeff Bonwick } else { 2770770499e1SDan Kimmel cio = zio_write(zio, spa, txg, bp, zio->io_orig_abd, 27715602294fSDan Kimmel zio->io_orig_size, zio->io_orig_size, zp, 27728df0bcf0SPaul Dagnelie zio_ddt_child_write_ready, NULL, NULL, 2773b24ab676SJeff Bonwick zio_ddt_child_write_done, dde, zio->io_priority, 2774b24ab676SJeff Bonwick ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark); 2775b24ab676SJeff Bonwick 2776770499e1SDan Kimmel zio_push_transform(cio, zio->io_abd, zio->io_size, 0, NULL); 2777b24ab676SJeff Bonwick dde->dde_lead_zio[p] = cio; 2778b24ab676SJeff Bonwick } 2779b24ab676SJeff Bonwick 2780b24ab676SJeff Bonwick ddt_exit(ddt); 2781b24ab676SJeff Bonwick 2782b24ab676SJeff Bonwick if (cio) 2783b24ab676SJeff Bonwick zio_nowait(cio); 2784b24ab676SJeff Bonwick if (dio) 2785b24ab676SJeff Bonwick zio_nowait(dio); 2786b24ab676SJeff Bonwick 2787b24ab676SJeff Bonwick return (ZIO_PIPELINE_CONTINUE); 2788b24ab676SJeff Bonwick } 2789b24ab676SJeff Bonwick 27903f9d6ad7SLin Ling ddt_entry_t *freedde; /* for debugging */ 27913f9d6ad7SLin Ling 2792b24ab676SJeff Bonwick static int 2793b24ab676SJeff Bonwick zio_ddt_free(zio_t *zio) 2794b24ab676SJeff Bonwick { 2795b24ab676SJeff Bonwick spa_t *spa = zio->io_spa; 2796b24ab676SJeff Bonwick blkptr_t *bp = zio->io_bp; 2797b24ab676SJeff Bonwick ddt_t *ddt = ddt_select(spa, bp); 2798b24ab676SJeff Bonwick ddt_entry_t *dde; 2799b24ab676SJeff Bonwick ddt_phys_t *ddp; 2800b24ab676SJeff Bonwick 2801b24ab676SJeff Bonwick ASSERT(BP_GET_DEDUP(bp)); 2802b24ab676SJeff Bonwick ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 2803b24ab676SJeff Bonwick 2804b24ab676SJeff Bonwick ddt_enter(ddt); 28053f9d6ad7SLin Ling freedde = dde = ddt_lookup(ddt, bp, B_TRUE); 2806b24ab676SJeff Bonwick ddp = ddt_phys_select(dde, bp); 2807b24ab676SJeff Bonwick ddt_phys_decref(ddp); 2808b24ab676SJeff Bonwick ddt_exit(ddt); 2809b24ab676SJeff Bonwick 2810b24ab676SJeff Bonwick return (ZIO_PIPELINE_CONTINUE); 2811b24ab676SJeff Bonwick } 2812b24ab676SJeff Bonwick 2813b24ab676SJeff Bonwick /* 2814b24ab676SJeff Bonwick * ========================================================================== 2815b24ab676SJeff Bonwick * Allocate and free blocks 2816b24ab676SJeff Bonwick * ========================================================================== 2817b24ab676SJeff Bonwick */ 28180f7643c7SGeorge Wilson 28190f7643c7SGeorge Wilson static zio_t * 2820f78cdc34SPaul Dagnelie zio_io_to_allocate(spa_t *spa, int allocator) 28210f7643c7SGeorge Wilson { 28220f7643c7SGeorge Wilson zio_t *zio; 28230f7643c7SGeorge Wilson 2824f78cdc34SPaul Dagnelie ASSERT(MUTEX_HELD(&spa->spa_alloc_locks[allocator])); 28250f7643c7SGeorge Wilson 2826f78cdc34SPaul Dagnelie zio = avl_first(&spa->spa_alloc_trees[allocator]); 28270f7643c7SGeorge Wilson if (zio == NULL) 28280f7643c7SGeorge Wilson return (NULL); 28290f7643c7SGeorge Wilson 28300f7643c7SGeorge Wilson ASSERT(IO_IS_ALLOCATING(zio)); 28310f7643c7SGeorge Wilson 28320f7643c7SGeorge Wilson /* 28330f7643c7SGeorge Wilson * Try to place a reservation for this zio. If we're unable to 28340f7643c7SGeorge Wilson * reserve then we throttle. 28350f7643c7SGeorge Wilson */ 2836f78cdc34SPaul Dagnelie ASSERT3U(zio->io_allocator, ==, allocator); 2837663207adSDon Brady if (!metaslab_class_throttle_reserve(zio->io_metaslab_class, 2838f78cdc34SPaul Dagnelie zio->io_prop.zp_copies, zio->io_allocator, zio, 0)) { 28390f7643c7SGeorge Wilson return (NULL); 28400f7643c7SGeorge Wilson } 28410f7643c7SGeorge Wilson 2842f78cdc34SPaul Dagnelie avl_remove(&spa->spa_alloc_trees[allocator], zio); 28430f7643c7SGeorge Wilson ASSERT3U(zio->io_stage, <, ZIO_STAGE_DVA_ALLOCATE); 28440f7643c7SGeorge Wilson 28450f7643c7SGeorge Wilson return (zio); 28460f7643c7SGeorge Wilson } 28470f7643c7SGeorge Wilson 28480f7643c7SGeorge Wilson static int 28490f7643c7SGeorge Wilson zio_dva_throttle(zio_t *zio) 28500f7643c7SGeorge Wilson { 28510f7643c7SGeorge Wilson spa_t *spa = zio->io_spa; 28520f7643c7SGeorge Wilson zio_t *nio; 2853663207adSDon Brady metaslab_class_t *mc; 2854663207adSDon Brady 2855663207adSDon Brady /* locate an appropriate allocation class */ 2856663207adSDon Brady mc = spa_preferred_class(spa, zio->io_size, zio->io_prop.zp_type, 2857663207adSDon Brady zio->io_prop.zp_level, zio->io_prop.zp_zpl_smallblk); 28580f7643c7SGeorge Wilson 28590f7643c7SGeorge Wilson if (zio->io_priority == ZIO_PRIORITY_SYNC_WRITE || 2860663207adSDon Brady !mc->mc_alloc_throttle_enabled || 28610f7643c7SGeorge Wilson zio->io_child_type == ZIO_CHILD_GANG || 28620f7643c7SGeorge Wilson zio->io_flags & ZIO_FLAG_NODATA) { 28630f7643c7SGeorge Wilson return (ZIO_PIPELINE_CONTINUE); 28640f7643c7SGeorge Wilson } 28650f7643c7SGeorge Wilson 28660f7643c7SGeorge Wilson ASSERT(zio->io_child_type > ZIO_CHILD_GANG); 28670f7643c7SGeorge Wilson 28680f7643c7SGeorge Wilson ASSERT3U(zio->io_queued_timestamp, >, 0); 28690f7643c7SGeorge Wilson ASSERT(zio->io_stage == ZIO_STAGE_DVA_THROTTLE); 28700f7643c7SGeorge Wilson 2871f78cdc34SPaul Dagnelie zbookmark_phys_t *bm = &zio->io_bookmark; 2872f78cdc34SPaul Dagnelie /* 2873f78cdc34SPaul Dagnelie * We want to try to use as many allocators as possible to help improve 2874f78cdc34SPaul Dagnelie * performance, but we also want logically adjacent IOs to be physically 2875f78cdc34SPaul Dagnelie * adjacent to improve sequential read performance. We chunk each object 2876f78cdc34SPaul Dagnelie * into 2^20 block regions, and then hash based on the objset, object, 2877f78cdc34SPaul Dagnelie * level, and region to accomplish both of these goals. 2878f78cdc34SPaul Dagnelie */ 2879f78cdc34SPaul Dagnelie zio->io_allocator = cityhash4(bm->zb_objset, bm->zb_object, 2880f78cdc34SPaul Dagnelie bm->zb_level, bm->zb_blkid >> 20) % spa->spa_alloc_count; 2881f78cdc34SPaul Dagnelie mutex_enter(&spa->spa_alloc_locks[zio->io_allocator]); 28820f7643c7SGeorge Wilson ASSERT(zio->io_type == ZIO_TYPE_WRITE); 2883663207adSDon Brady zio->io_metaslab_class = mc; 2884f78cdc34SPaul Dagnelie avl_add(&spa->spa_alloc_trees[zio->io_allocator], zio); 2885663207adSDon Brady nio = zio_io_to_allocate(spa, zio->io_allocator); 2886f78cdc34SPaul Dagnelie mutex_exit(&spa->spa_alloc_locks[zio->io_allocator]); 28870f7643c7SGeorge Wilson 28880f7643c7SGeorge Wilson if (nio == zio) 28890f7643c7SGeorge Wilson return (ZIO_PIPELINE_CONTINUE); 28900f7643c7SGeorge Wilson 28910f7643c7SGeorge Wilson if (nio != NULL) { 28920f7643c7SGeorge Wilson ASSERT(nio->io_stage == ZIO_STAGE_DVA_THROTTLE); 28930f7643c7SGeorge Wilson /* 28940f7643c7SGeorge Wilson * We are passing control to a new zio so make sure that 28950f7643c7SGeorge Wilson * it is processed by a different thread. We do this to 28960f7643c7SGeorge Wilson * avoid stack overflows that can occur when parents are 28970f7643c7SGeorge Wilson * throttled and children are making progress. We allow 28980f7643c7SGeorge Wilson * it to go to the head of the taskq since it's already 28990f7643c7SGeorge Wilson * been waiting. 29000f7643c7SGeorge Wilson */ 29010f7643c7SGeorge Wilson zio_taskq_dispatch(nio, ZIO_TASKQ_ISSUE, B_TRUE); 29020f7643c7SGeorge Wilson } 29030f7643c7SGeorge Wilson return (ZIO_PIPELINE_STOP); 29040f7643c7SGeorge Wilson } 29050f7643c7SGeorge Wilson 2906663207adSDon Brady static void 2907f78cdc34SPaul Dagnelie zio_allocate_dispatch(spa_t *spa, int allocator) 29080f7643c7SGeorge Wilson { 29090f7643c7SGeorge Wilson zio_t *zio; 29100f7643c7SGeorge Wilson 2911f78cdc34SPaul Dagnelie mutex_enter(&spa->spa_alloc_locks[allocator]); 2912f78cdc34SPaul Dagnelie zio = zio_io_to_allocate(spa, allocator); 2913f78cdc34SPaul Dagnelie mutex_exit(&spa->spa_alloc_locks[allocator]); 29140f7643c7SGeorge Wilson if (zio == NULL) 29150f7643c7SGeorge Wilson return; 29160f7643c7SGeorge Wilson 29170f7643c7SGeorge Wilson ASSERT3U(zio->io_stage, ==, ZIO_STAGE_DVA_THROTTLE); 29180f7643c7SGeorge Wilson ASSERT0(zio->io_error); 29190f7643c7SGeorge Wilson zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_TRUE); 29200f7643c7SGeorge Wilson } 29210f7643c7SGeorge Wilson 2922e05725b1Sbonwick static int 2923fa9e4066Sahrens zio_dva_allocate(zio_t *zio) 2924fa9e4066Sahrens { 29258654d025Sperrin spa_t *spa = zio->io_spa; 2926663207adSDon Brady metaslab_class_t *mc; 2927fa9e4066Sahrens blkptr_t *bp = zio->io_bp; 2928fa9e4066Sahrens int error; 292909c9d376SGeorge Wilson int flags = 0; 2930fa9e4066Sahrens 2931f5383399SBill Moore if (zio->io_gang_leader == NULL) { 2932f5383399SBill Moore ASSERT(zio->io_child_type > ZIO_CHILD_GANG); 2933f5383399SBill Moore zio->io_gang_leader = zio; 2934f5383399SBill Moore } 2935f5383399SBill Moore 2936fa9e4066Sahrens ASSERT(BP_IS_HOLE(bp)); 2937fb09f5aaSMadhav Suresh ASSERT0(BP_GET_NDVAS(bp)); 2938b24ab676SJeff Bonwick ASSERT3U(zio->io_prop.zp_copies, >, 0); 2939b24ab676SJeff Bonwick ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa)); 2940fa9e4066Sahrens ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp)); 2941fa9e4066Sahrens 2942663207adSDon Brady if (zio->io_flags & ZIO_FLAG_NODATA) 29430f7643c7SGeorge Wilson flags |= METASLAB_DONT_THROTTLE; 2944663207adSDon Brady if (zio->io_flags & ZIO_FLAG_GANG_CHILD) 29450f7643c7SGeorge Wilson flags |= METASLAB_GANG_CHILD; 2946663207adSDon Brady if (zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE) 29470f7643c7SGeorge Wilson flags |= METASLAB_ASYNC_ALLOC; 2948663207adSDon Brady 2949663207adSDon Brady /* 2950663207adSDon Brady * if not already chosen, locate an appropriate allocation class 2951663207adSDon Brady */ 2952663207adSDon Brady mc = zio->io_metaslab_class; 2953663207adSDon Brady if (mc == NULL) { 2954663207adSDon Brady mc = spa_preferred_class(spa, zio->io_size, 2955663207adSDon Brady zio->io_prop.zp_type, zio->io_prop.zp_level, 2956663207adSDon Brady zio->io_prop.zp_zpl_smallblk); 2957663207adSDon Brady zio->io_metaslab_class = mc; 29580f7643c7SGeorge Wilson } 29590f7643c7SGeorge Wilson 2960e14bb325SJeff Bonwick error = metaslab_alloc(spa, mc, zio->io_size, bp, 29618363e80aSGeorge Wilson zio->io_prop.zp_copies, zio->io_txg, NULL, flags, 2962f78cdc34SPaul Dagnelie &zio->io_alloc_list, zio, zio->io_allocator); 2963fa9e4066Sahrens 2964663207adSDon Brady /* 2965663207adSDon Brady * Fallback to normal class when an alloc class is full 2966663207adSDon Brady */ 2967663207adSDon Brady if (error == ENOSPC && mc != spa_normal_class(spa)) { 2968663207adSDon Brady /* 2969663207adSDon Brady * If throttling, transfer reservation over to normal class. 2970663207adSDon Brady * The io_allocator slot can remain the same even though we 2971663207adSDon Brady * are switching classes. 2972663207adSDon Brady */ 2973663207adSDon Brady if (mc->mc_alloc_throttle_enabled && 2974663207adSDon Brady (zio->io_flags & ZIO_FLAG_IO_ALLOCATING)) { 2975663207adSDon Brady metaslab_class_throttle_unreserve(mc, 2976663207adSDon Brady zio->io_prop.zp_copies, zio->io_allocator, zio); 2977663207adSDon Brady zio->io_flags &= ~ZIO_FLAG_IO_ALLOCATING; 2978663207adSDon Brady 2979663207adSDon Brady mc = spa_normal_class(spa); 2980663207adSDon Brady VERIFY(metaslab_class_throttle_reserve(mc, 2981663207adSDon Brady zio->io_prop.zp_copies, zio->io_allocator, zio, 2982663207adSDon Brady flags | METASLAB_MUST_RESERVE)); 2983663207adSDon Brady } else { 2984663207adSDon Brady mc = spa_normal_class(spa); 2985663207adSDon Brady } 2986663207adSDon Brady zio->io_metaslab_class = mc; 2987663207adSDon Brady 2988663207adSDon Brady error = metaslab_alloc(spa, mc, zio->io_size, bp, 2989663207adSDon Brady zio->io_prop.zp_copies, zio->io_txg, NULL, flags, 2990663207adSDon Brady &zio->io_alloc_list, zio, zio->io_allocator); 2991663207adSDon Brady } 2992663207adSDon Brady 29930f7643c7SGeorge Wilson if (error != 0) { 299421f7c81cSMatthew Ahrens zfs_dbgmsg("%s: metaslab allocation failure: zio %p, " 299509c9d376SGeorge Wilson "size %llu, error %d", spa_name(spa), zio, zio->io_size, 299609c9d376SGeorge Wilson error); 2997e14bb325SJeff Bonwick if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE) 2998e14bb325SJeff Bonwick return (zio_write_gang_block(zio)); 2999fa9e4066Sahrens zio->io_error = error; 3000fa9e4066Sahrens } 3001e05725b1Sbonwick 3002e05725b1Sbonwick return (ZIO_PIPELINE_CONTINUE); 3003fa9e4066Sahrens } 3004fa9e4066Sahrens 3005e05725b1Sbonwick static int 3006fa9e4066Sahrens zio_dva_free(zio_t *zio) 3007fa9e4066Sahrens { 3008e14bb325SJeff Bonwick metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE); 3009fa9e4066Sahrens 3010e05725b1Sbonwick return (ZIO_PIPELINE_CONTINUE); 3011fa9e4066Sahrens } 3012fa9e4066Sahrens 3013e05725b1Sbonwick static int 3014fa9e4066Sahrens zio_dva_claim(zio_t *zio) 3015fa9e4066Sahrens { 3016e14bb325SJeff Bonwick int error; 3017e14bb325SJeff Bonwick 3018e14bb325SJeff Bonwick error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg); 3019e14bb325SJeff Bonwick if (error) 3020e14bb325SJeff Bonwick zio->io_error = error; 3021fa9e4066Sahrens 3022e05725b1Sbonwick return (ZIO_PIPELINE_CONTINUE); 3023fa9e4066Sahrens } 3024fa9e4066Sahrens 3025e14bb325SJeff Bonwick /* 3026e14bb325SJeff Bonwick * Undo an allocation. This is used by zio_done() when an I/O fails 3027e14bb325SJeff Bonwick * and we want to give back the block we just allocated. 3028e14bb325SJeff Bonwick * This handles both normal blocks and gang blocks. 3029e14bb325SJeff Bonwick */ 3030e14bb325SJeff Bonwick static void 3031e14bb325SJeff Bonwick zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp) 3032e14bb325SJeff Bonwick { 3033e14bb325SJeff Bonwick ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp)); 3034b24ab676SJeff Bonwick ASSERT(zio->io_bp_override == NULL); 3035e14bb325SJeff Bonwick 3036e14bb325SJeff Bonwick if (!BP_IS_HOLE(bp)) 3037b24ab676SJeff Bonwick metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE); 3038e14bb325SJeff Bonwick 3039e14bb325SJeff Bonwick if (gn != NULL) { 3040e14bb325SJeff Bonwick for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) { 3041e14bb325SJeff Bonwick zio_dva_unallocate(zio, gn->gn_child[g], 3042e14bb325SJeff Bonwick &gn->gn_gbh->zg_blkptr[g]); 3043e14bb325SJeff Bonwick } 3044e14bb325SJeff Bonwick } 3045e14bb325SJeff Bonwick } 3046e14bb325SJeff Bonwick 3047e14bb325SJeff Bonwick /* 3048e14bb325SJeff Bonwick * Try to allocate an intent log block. Return 0 on success, errno on failure. 3049e14bb325SJeff Bonwick */ 3050e14bb325SJeff Bonwick int 3051f78cdc34SPaul Dagnelie zio_alloc_zil(spa_t *spa, uint64_t objset, uint64_t txg, blkptr_t *new_bp, 3052f78cdc34SPaul Dagnelie blkptr_t *old_bp, uint64_t size, boolean_t *slog) 3053e14bb325SJeff Bonwick { 3054e09fa4daSNeil Perrin int error = 1; 30558363e80aSGeorge Wilson zio_alloc_list_t io_alloc_list; 3056e14bb325SJeff Bonwick 3057b24ab676SJeff Bonwick ASSERT(txg > spa_syncing_txg(spa)); 3058b24ab676SJeff Bonwick 30598363e80aSGeorge Wilson metaslab_trace_init(&io_alloc_list); 3060663207adSDon Brady 3061663207adSDon Brady /* 3062663207adSDon Brady * Block pointer fields are useful to metaslabs for stats and debugging. 3063663207adSDon Brady * Fill in the obvious ones before calling into metaslab_alloc(). 3064663207adSDon Brady */ 3065663207adSDon Brady BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG); 3066663207adSDon Brady BP_SET_PSIZE(new_bp, size); 3067663207adSDon Brady BP_SET_LEVEL(new_bp, 0); 3068663207adSDon Brady 3069f78cdc34SPaul Dagnelie /* 3070f78cdc34SPaul Dagnelie * When allocating a zil block, we don't have information about 3071f78cdc34SPaul Dagnelie * the final destination of the block except the objset it's part 3072f78cdc34SPaul Dagnelie * of, so we just hash the objset ID to pick the allocator to get 3073f78cdc34SPaul Dagnelie * some parallelism. 3074f78cdc34SPaul Dagnelie */ 3075c5ee4681SAlexander Motin error = metaslab_alloc(spa, spa_log_class(spa), size, new_bp, 1, 3076f78cdc34SPaul Dagnelie txg, old_bp, METASLAB_HINTBP_AVOID, &io_alloc_list, NULL, 3077f78cdc34SPaul Dagnelie cityhash4(0, 0, 0, objset) % spa->spa_alloc_count); 3078c5ee4681SAlexander Motin if (error == 0) { 3079c5ee4681SAlexander Motin *slog = TRUE; 3080c5ee4681SAlexander Motin } else { 3081b24ab676SJeff Bonwick error = metaslab_alloc(spa, spa_normal_class(spa), size, 30828363e80aSGeorge Wilson new_bp, 1, txg, old_bp, METASLAB_HINTBP_AVOID, 3083f78cdc34SPaul Dagnelie &io_alloc_list, NULL, cityhash4(0, 0, 0, objset) % 3084f78cdc34SPaul Dagnelie spa->spa_alloc_count); 3085c5ee4681SAlexander Motin if (error == 0) 3086c5ee4681SAlexander Motin *slog = FALSE; 3087840345f6SGeorge Wilson } 30888363e80aSGeorge Wilson metaslab_trace_fini(&io_alloc_list); 3089e14bb325SJeff Bonwick 3090e14bb325SJeff Bonwick if (error == 0) { 3091e14bb325SJeff Bonwick BP_SET_LSIZE(new_bp, size); 3092e14bb325SJeff Bonwick BP_SET_PSIZE(new_bp, size); 3093e14bb325SJeff Bonwick BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF); 30946e1f5caaSNeil Perrin BP_SET_CHECKSUM(new_bp, 30956e1f5caaSNeil Perrin spa_version(spa) >= SPA_VERSION_SLIM_ZIL 30966e1f5caaSNeil Perrin ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG); 3097e14bb325SJeff Bonwick BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG); 3098e14bb325SJeff Bonwick BP_SET_LEVEL(new_bp, 0); 3099b24ab676SJeff Bonwick BP_SET_DEDUP(new_bp, 0); 3100e14bb325SJeff Bonwick BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER); 31011271e4b1SPrakash Surya } else { 31021271e4b1SPrakash Surya zfs_dbgmsg("%s: zil block allocation failure: " 31031271e4b1SPrakash Surya "size %llu, error %d", spa_name(spa), size, error); 3104e14bb325SJeff Bonwick } 3105e14bb325SJeff Bonwick 3106e14bb325SJeff Bonwick return (error); 3107e14bb325SJeff Bonwick } 3108e14bb325SJeff Bonwick 3109fa9e4066Sahrens /* 3110fa9e4066Sahrens * ========================================================================== 3111fa9e4066Sahrens * Read and write to physical devices 3112fa9e4066Sahrens * ========================================================================== 3113fa9e4066Sahrens */ 3114738f37bcSGeorge Wilson 3115738f37bcSGeorge Wilson 3116738f37bcSGeorge Wilson /* 3117738f37bcSGeorge Wilson * Issue an I/O to the underlying vdev. Typically the issue pipeline 3118738f37bcSGeorge Wilson * stops after this stage and will resume upon I/O completion. 3119738f37bcSGeorge Wilson * However, there are instances where the vdev layer may need to 3120738f37bcSGeorge Wilson * continue the pipeline when an I/O was not issued. Since the I/O 3121738f37bcSGeorge Wilson * that was sent to the vdev layer might be different than the one 3122738f37bcSGeorge Wilson * currently active in the pipeline (see vdev_queue_io()), we explicitly 3123738f37bcSGeorge Wilson * force the underlying vdev layers to call either zio_execute() or 3124738f37bcSGeorge Wilson * zio_interrupt() to ensure that the pipeline continues with the correct I/O. 3125738f37bcSGeorge Wilson */ 3126e05725b1Sbonwick static int 312744cd46caSbillm zio_vdev_io_start(zio_t *zio) 3128fa9e4066Sahrens { 3129fa9e4066Sahrens vdev_t *vd = zio->io_vd; 313044cd46caSbillm uint64_t align; 31310a4e9518Sgw spa_t *spa = zio->io_spa; 31320a4e9518Sgw 3133e14bb325SJeff Bonwick ASSERT(zio->io_error == 0); 3134e14bb325SJeff Bonwick ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0); 3135fa9e4066Sahrens 3136e14bb325SJeff Bonwick if (vd == NULL) { 3137e14bb325SJeff Bonwick if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER)) 3138e14bb325SJeff Bonwick spa_config_enter(spa, SCL_ZIO, zio, RW_READER); 3139fa9e4066Sahrens 3140e14bb325SJeff Bonwick /* 3141e14bb325SJeff Bonwick * The mirror_ops handle multiple DVAs in a single BP. 3142e14bb325SJeff Bonwick */ 3143738f37bcSGeorge Wilson vdev_mirror_ops.vdev_op_io_start(zio); 3144738f37bcSGeorge Wilson return (ZIO_PIPELINE_STOP); 3145fa9e4066Sahrens } 3146fa9e4066Sahrens 31470f7643c7SGeorge Wilson ASSERT3P(zio->io_logical, !=, zio); 31486f793812SPavel Zakharov if (zio->io_type == ZIO_TYPE_WRITE) { 31496f793812SPavel Zakharov ASSERT(spa->spa_trust_config); 31506f793812SPavel Zakharov 3151*a3874b8bSToomas Soome /* 3152*a3874b8bSToomas Soome * Note: the code can handle other kinds of writes, 3153*a3874b8bSToomas Soome * but we don't expect them. 3154*a3874b8bSToomas Soome */ 31556f793812SPavel Zakharov if (zio->io_vd->vdev_removing) { 31566f793812SPavel Zakharov ASSERT(zio->io_flags & 31576f793812SPavel Zakharov (ZIO_FLAG_PHYSICAL | ZIO_FLAG_SELF_HEAL | 31583a4b1be9SMatthew Ahrens ZIO_FLAG_RESILVER | ZIO_FLAG_INDUCE_DAMAGE)); 31596f793812SPavel Zakharov } 31605cabbc6bSPrashanth Sreenivasa } 31610f7643c7SGeorge Wilson 3162e14bb325SJeff Bonwick align = 1ULL << vd->vdev_top->vdev_ashift; 3163e14bb325SJeff Bonwick 31642a104a52SAlex Reece if (!(zio->io_flags & ZIO_FLAG_PHYSICAL) && 31652a104a52SAlex Reece P2PHASE(zio->io_size, align) != 0) { 31662a104a52SAlex Reece /* Transform logical writes to be a full physical block size. */ 3167ecc2d604Sbonwick uint64_t asize = P2ROUNDUP(zio->io_size, align); 3168770499e1SDan Kimmel abd_t *abuf = abd_alloc_sametype(zio->io_abd, asize); 3169e14bb325SJeff Bonwick ASSERT(vd == vd->vdev_top); 3170ecc2d604Sbonwick if (zio->io_type == ZIO_TYPE_WRITE) { 3171770499e1SDan Kimmel abd_copy(abuf, zio->io_abd, zio->io_size); 3172770499e1SDan Kimmel abd_zero_off(abuf, zio->io_size, asize - zio->io_size); 3173ecc2d604Sbonwick } 3174e14bb325SJeff Bonwick zio_push_transform(zio, abuf, asize, asize, zio_subblock); 3175ecc2d604Sbonwick } 3176ecc2d604Sbonwick 31772a104a52SAlex Reece /* 31782a104a52SAlex Reece * If this is not a physical io, make sure that it is properly aligned 31792a104a52SAlex Reece * before proceeding. 31802a104a52SAlex Reece */ 31812a104a52SAlex Reece if (!(zio->io_flags & ZIO_FLAG_PHYSICAL)) { 31822a104a52SAlex Reece ASSERT0(P2PHASE(zio->io_offset, align)); 31832a104a52SAlex Reece ASSERT0(P2PHASE(zio->io_size, align)); 31842a104a52SAlex Reece } else { 31852a104a52SAlex Reece /* 31862a104a52SAlex Reece * For physical writes, we allow 512b aligned writes and assume 31872a104a52SAlex Reece * the device will perform a read-modify-write as necessary. 31882a104a52SAlex Reece */ 31892a104a52SAlex Reece ASSERT0(P2PHASE(zio->io_offset, SPA_MINBLOCKSIZE)); 31902a104a52SAlex Reece ASSERT0(P2PHASE(zio->io_size, SPA_MINBLOCKSIZE)); 31912a104a52SAlex Reece } 31922a104a52SAlex Reece 3193f9af39baSGeorge Wilson VERIFY(zio->io_type != ZIO_TYPE_WRITE || spa_writeable(spa)); 31948ad4d6ddSJeff Bonwick 31958ad4d6ddSJeff Bonwick /* 31968ad4d6ddSJeff Bonwick * If this is a repair I/O, and there's no self-healing involved -- 31978ad4d6ddSJeff Bonwick * that is, we're just resilvering what we expect to resilver -- 31988ad4d6ddSJeff Bonwick * then don't do the I/O unless zio's txg is actually in vd's DTL. 31993a4b1be9SMatthew Ahrens * This prevents spurious resilvering. 32003a4b1be9SMatthew Ahrens * 32013a4b1be9SMatthew Ahrens * There are a few ways that we can end up creating these spurious 32023a4b1be9SMatthew Ahrens * resilver i/os: 32033a4b1be9SMatthew Ahrens * 32043a4b1be9SMatthew Ahrens * 1. A resilver i/o will be issued if any DVA in the BP has a 32053a4b1be9SMatthew Ahrens * dirty DTL. The mirror code will issue resilver writes to 32063a4b1be9SMatthew Ahrens * each DVA, including the one(s) that are not on vdevs with dirty 32073a4b1be9SMatthew Ahrens * DTLs. 32083a4b1be9SMatthew Ahrens * 32093a4b1be9SMatthew Ahrens * 2. With nested replication, which happens when we have a 32103a4b1be9SMatthew Ahrens * "replacing" or "spare" vdev that's a child of a mirror or raidz. 32113a4b1be9SMatthew Ahrens * For example, given mirror(replacing(A+B), C), it's likely that 32123a4b1be9SMatthew Ahrens * only A is out of date (it's the new device). In this case, we'll 32133a4b1be9SMatthew Ahrens * read from C, then use the data to resilver A+B -- but we don't 32143a4b1be9SMatthew Ahrens * actually want to resilver B, just A. The top-level mirror has no 32153a4b1be9SMatthew Ahrens * way to know this, so instead we just discard unnecessary repairs 32163a4b1be9SMatthew Ahrens * as we work our way down the vdev tree. 32173a4b1be9SMatthew Ahrens * 32183a4b1be9SMatthew Ahrens * 3. ZTEST also creates mirrors of mirrors, mirrors of raidz, etc. 32193a4b1be9SMatthew Ahrens * The same logic applies to any form of nested replication: ditto 32203a4b1be9SMatthew Ahrens * + mirror, RAID-Z + replacing, etc. 32213a4b1be9SMatthew Ahrens * 32223a4b1be9SMatthew Ahrens * However, indirect vdevs point off to other vdevs which may have 32233a4b1be9SMatthew Ahrens * DTL's, so we never bypass them. The child i/os on concrete vdevs 32243a4b1be9SMatthew Ahrens * will be properly bypassed instead. 32258ad4d6ddSJeff Bonwick */ 32268ad4d6ddSJeff Bonwick if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) && 32278ad4d6ddSJeff Bonwick !(zio->io_flags & ZIO_FLAG_SELF_HEAL) && 32288ad4d6ddSJeff Bonwick zio->io_txg != 0 && /* not a delegated i/o */ 32293a4b1be9SMatthew Ahrens vd->vdev_ops != &vdev_indirect_ops && 32308ad4d6ddSJeff Bonwick !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) { 32318ad4d6ddSJeff Bonwick ASSERT(zio->io_type == ZIO_TYPE_WRITE); 32328ad4d6ddSJeff Bonwick zio_vdev_io_bypass(zio); 32338ad4d6ddSJeff Bonwick return (ZIO_PIPELINE_CONTINUE); 32348ad4d6ddSJeff Bonwick } 3235fa9e4066Sahrens 3236e14bb325SJeff Bonwick if (vd->vdev_ops->vdev_op_leaf && 3237e14bb325SJeff Bonwick (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE)) { 3238e14bb325SJeff Bonwick 323943466aaeSMax Grossman if (zio->io_type == ZIO_TYPE_READ && vdev_cache_read(zio)) 3240a3f829aeSBill Moore return (ZIO_PIPELINE_CONTINUE); 3241e14bb325SJeff Bonwick 3242e14bb325SJeff Bonwick if ((zio = vdev_queue_io(zio)) == NULL) 3243e14bb325SJeff Bonwick return (ZIO_PIPELINE_STOP); 3244e14bb325SJeff Bonwick 3245e14bb325SJeff Bonwick if (!vdev_accessible(vd, zio)) { 3246be6fd75aSMatthew Ahrens zio->io_error = SET_ERROR(ENXIO); 3247e14bb325SJeff Bonwick zio_interrupt(zio); 3248e14bb325SJeff Bonwick return (ZIO_PIPELINE_STOP); 3249e14bb325SJeff Bonwick } 3250e14bb325SJeff Bonwick } 3251e14bb325SJeff Bonwick 3252738f37bcSGeorge Wilson vd->vdev_ops->vdev_op_io_start(zio); 3253738f37bcSGeorge Wilson return (ZIO_PIPELINE_STOP); 3254fa9e4066Sahrens } 3255fa9e4066Sahrens 3256e05725b1Sbonwick static int 3257fa9e4066Sahrens zio_vdev_io_done(zio_t *zio) 3258fa9e4066Sahrens { 3259e14bb325SJeff Bonwick vdev_t *vd = zio->io_vd; 3260e14bb325SJeff Bonwick vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops; 3261e14bb325SJeff Bonwick boolean_t unexpected_error = B_FALSE; 3262e05725b1Sbonwick 3263d6e1c446SGeorge Wilson if (zio_wait_for_children(zio, ZIO_CHILD_VDEV_BIT, ZIO_WAIT_DONE)) { 3264e14bb325SJeff Bonwick return (ZIO_PIPELINE_STOP); 3265d6e1c446SGeorge Wilson } 3266fa9e4066Sahrens 3267e14bb325SJeff Bonwick ASSERT(zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE); 3268e14bb325SJeff Bonwick 3269e14bb325SJeff Bonwick if (vd != NULL && vd->vdev_ops->vdev_op_leaf) { 3270e14bb325SJeff Bonwick 3271e14bb325SJeff Bonwick vdev_queue_io_done(zio); 3272fa9e4066Sahrens 3273e14bb325SJeff Bonwick if (zio->io_type == ZIO_TYPE_WRITE) 3274e14bb325SJeff Bonwick vdev_cache_write(zio); 3275e14bb325SJeff Bonwick 3276e14bb325SJeff Bonwick if (zio_injection_enabled && zio->io_error == 0) 32778956713aSEric Schrock zio->io_error = zio_handle_device_injection(vd, 32788956713aSEric Schrock zio, EIO); 3279e14bb325SJeff Bonwick 3280e14bb325SJeff Bonwick if (zio_injection_enabled && zio->io_error == 0) 3281e14bb325SJeff Bonwick zio->io_error = zio_handle_label_injection(zio, EIO); 3282e14bb325SJeff Bonwick 3283e14bb325SJeff Bonwick if (zio->io_error) { 3284e14bb325SJeff Bonwick if (!vdev_accessible(vd, zio)) { 3285be6fd75aSMatthew Ahrens zio->io_error = SET_ERROR(ENXIO); 3286e14bb325SJeff Bonwick } else { 3287e14bb325SJeff Bonwick unexpected_error = B_TRUE; 3288e14bb325SJeff Bonwick } 3289e14bb325SJeff Bonwick } 329051ece835Seschrock } 3291fa9e4066Sahrens 3292e14bb325SJeff Bonwick ops->vdev_op_io_done(zio); 3293e14bb325SJeff Bonwick 3294e14bb325SJeff Bonwick if (unexpected_error) 3295a3f829aeSBill Moore VERIFY(vdev_probe(vd, zio) == NULL); 3296e14bb325SJeff Bonwick 3297e14bb325SJeff Bonwick return (ZIO_PIPELINE_CONTINUE); 3298fa9e4066Sahrens } 3299fa9e4066Sahrens 3300*a3874b8bSToomas Soome /* 3301*a3874b8bSToomas Soome * This function is used to change the priority of an existing zio that is 3302*a3874b8bSToomas Soome * currently in-flight. This is used by the arc to upgrade priority in the 3303*a3874b8bSToomas Soome * event that a demand read is made for a block that is currently queued 3304*a3874b8bSToomas Soome * as a scrub or async read IO. Otherwise, the high priority read request 3305*a3874b8bSToomas Soome * would end up having to wait for the lower priority IO. 3306*a3874b8bSToomas Soome */ 3307*a3874b8bSToomas Soome void 3308*a3874b8bSToomas Soome zio_change_priority(zio_t *pio, zio_priority_t priority) 3309*a3874b8bSToomas Soome { 3310*a3874b8bSToomas Soome zio_t *cio, *cio_next; 3311*a3874b8bSToomas Soome zio_link_t *zl = NULL; 3312*a3874b8bSToomas Soome 3313*a3874b8bSToomas Soome ASSERT3U(priority, <, ZIO_PRIORITY_NUM_QUEUEABLE); 3314*a3874b8bSToomas Soome 3315*a3874b8bSToomas Soome if (pio->io_vd != NULL && pio->io_vd->vdev_ops->vdev_op_leaf) { 3316*a3874b8bSToomas Soome vdev_queue_change_io_priority(pio, priority); 3317*a3874b8bSToomas Soome } else { 3318*a3874b8bSToomas Soome pio->io_priority = priority; 3319*a3874b8bSToomas Soome } 3320*a3874b8bSToomas Soome 3321*a3874b8bSToomas Soome mutex_enter(&pio->io_lock); 3322*a3874b8bSToomas Soome for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) { 3323*a3874b8bSToomas Soome cio_next = zio_walk_children(pio, &zl); 3324*a3874b8bSToomas Soome zio_change_priority(cio, priority); 3325*a3874b8bSToomas Soome } 3326*a3874b8bSToomas Soome mutex_exit(&pio->io_lock); 3327*a3874b8bSToomas Soome } 3328*a3874b8bSToomas Soome 332922fe2c88SJonathan Adams /* 333022fe2c88SJonathan Adams * For non-raidz ZIOs, we can just copy aside the bad data read from the 333122fe2c88SJonathan Adams * disk, and use that to finish the checksum ereport later. 333222fe2c88SJonathan Adams */ 333322fe2c88SJonathan Adams static void 333422fe2c88SJonathan Adams zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr, 333522fe2c88SJonathan Adams const void *good_buf) 333622fe2c88SJonathan Adams { 333722fe2c88SJonathan Adams /* no processing needed */ 333822fe2c88SJonathan Adams zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE); 333922fe2c88SJonathan Adams } 334022fe2c88SJonathan Adams 334122fe2c88SJonathan Adams /*ARGSUSED*/ 334222fe2c88SJonathan Adams void 334322fe2c88SJonathan Adams zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored) 334422fe2c88SJonathan Adams { 334522fe2c88SJonathan Adams void *buf = zio_buf_alloc(zio->io_size); 334622fe2c88SJonathan Adams 3347770499e1SDan Kimmel abd_copy_to_buf(buf, zio->io_abd, zio->io_size); 334822fe2c88SJonathan Adams 334922fe2c88SJonathan Adams zcr->zcr_cbinfo = zio->io_size; 335022fe2c88SJonathan Adams zcr->zcr_cbdata = buf; 335122fe2c88SJonathan Adams zcr->zcr_finish = zio_vsd_default_cksum_finish; 335222fe2c88SJonathan Adams zcr->zcr_free = zio_buf_free; 335322fe2c88SJonathan Adams } 335422fe2c88SJonathan Adams 3355e05725b1Sbonwick static int 3356fa9e4066Sahrens zio_vdev_io_assess(zio_t *zio) 3357fa9e4066Sahrens { 3358fa9e4066Sahrens vdev_t *vd = zio->io_vd; 3359e14bb325SJeff Bonwick 3360d6e1c446SGeorge Wilson if (zio_wait_for_children(zio, ZIO_CHILD_VDEV_BIT, ZIO_WAIT_DONE)) { 3361e14bb325SJeff Bonwick return (ZIO_PIPELINE_STOP); 3362d6e1c446SGeorge Wilson } 3363e14bb325SJeff Bonwick 3364e14bb325SJeff Bonwick if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER)) 3365e14bb325SJeff Bonwick spa_config_exit(zio->io_spa, SCL_ZIO, zio); 3366e14bb325SJeff Bonwick 3367e14bb325SJeff Bonwick if (zio->io_vsd != NULL) { 336822fe2c88SJonathan Adams zio->io_vsd_ops->vsd_free(zio); 3369e14bb325SJeff Bonwick zio->io_vsd = NULL; 3370ecc2d604Sbonwick } 3371ecc2d604Sbonwick 3372e14bb325SJeff Bonwick if (zio_injection_enabled && zio->io_error == 0) 3373ea8dc4b6Seschrock zio->io_error = zio_handle_fault_injection(zio, EIO); 3374ea8dc4b6Seschrock 3375fa9e4066Sahrens /* 3376fa9e4066Sahrens * If the I/O failed, determine whether we should attempt to retry it. 337735a5a358SJonathan Adams * 337835a5a358SJonathan Adams * On retry, we cut in line in the issue queue, since we don't want 337935a5a358SJonathan Adams * compression/checksumming/etc. work to prevent our (cheap) IO reissue. 3380fa9e4066Sahrens */ 3381e14bb325SJeff Bonwick if (zio->io_error && vd == NULL && 3382e14bb325SJeff Bonwick !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) { 3383e14bb325SJeff Bonwick ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */ 3384e14bb325SJeff Bonwick ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */ 3385fa9e4066Sahrens zio->io_error = 0; 3386e14bb325SJeff Bonwick zio->io_flags |= ZIO_FLAG_IO_RETRY | 3387e14bb325SJeff Bonwick ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE; 3388b24ab676SJeff Bonwick zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1; 338935a5a358SJonathan Adams zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, 339035a5a358SJonathan Adams zio_requeue_io_start_cut_in_line); 3391e14bb325SJeff Bonwick return (ZIO_PIPELINE_STOP); 3392ea8dc4b6Seschrock } 3393fa9e4066Sahrens 3394e14bb325SJeff Bonwick /* 3395e14bb325SJeff Bonwick * If we got an error on a leaf device, convert it to ENXIO 3396e14bb325SJeff Bonwick * if the device is not accessible at all. 3397e14bb325SJeff Bonwick */ 3398e14bb325SJeff Bonwick if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf && 3399e14bb325SJeff Bonwick !vdev_accessible(vd, zio)) 3400be6fd75aSMatthew Ahrens zio->io_error = SET_ERROR(ENXIO); 3401e14bb325SJeff Bonwick 3402e14bb325SJeff Bonwick /* 3403e14bb325SJeff Bonwick * If we can't write to an interior vdev (mirror or RAID-Z), 3404e14bb325SJeff Bonwick * set vdev_cant_write so that we stop trying to allocate from it. 3405e14bb325SJeff Bonwick */ 3406e14bb325SJeff Bonwick if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE && 34073b2aab18SMatthew Ahrens vd != NULL && !vd->vdev_ops->vdev_op_leaf) { 3408e14bb325SJeff Bonwick vd->vdev_cant_write = B_TRUE; 34093b2aab18SMatthew Ahrens } 3410e14bb325SJeff Bonwick 3411295438baSHans Rosenfeld /* 3412295438baSHans Rosenfeld * If a cache flush returns ENOTSUP or ENOTTY, we know that no future 3413295438baSHans Rosenfeld * attempts will ever succeed. In this case we set a persistent bit so 3414295438baSHans Rosenfeld * that we don't bother with it in the future. 3415295438baSHans Rosenfeld */ 3416295438baSHans Rosenfeld if ((zio->io_error == ENOTSUP || zio->io_error == ENOTTY) && 3417295438baSHans Rosenfeld zio->io_type == ZIO_TYPE_IOCTL && 3418295438baSHans Rosenfeld zio->io_cmd == DKIOCFLUSHWRITECACHE && vd != NULL) 3419295438baSHans Rosenfeld vd->vdev_nowritecache = B_TRUE; 3420295438baSHans Rosenfeld 3421e14bb325SJeff Bonwick if (zio->io_error) 3422e14bb325SJeff Bonwick zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 3423e14bb325SJeff Bonwick 342469962b56SMatthew Ahrens if (vd != NULL && vd->vdev_ops->vdev_op_leaf && 342569962b56SMatthew Ahrens zio->io_physdone != NULL) { 342669962b56SMatthew Ahrens ASSERT(!(zio->io_flags & ZIO_FLAG_DELEGATED)); 342769962b56SMatthew Ahrens ASSERT(zio->io_child_type == ZIO_CHILD_VDEV); 342869962b56SMatthew Ahrens zio->io_physdone(zio->io_logical); 342969962b56SMatthew Ahrens } 343069962b56SMatthew Ahrens 3431e05725b1Sbonwick return (ZIO_PIPELINE_CONTINUE); 3432fa9e4066Sahrens } 3433fa9e4066Sahrens 3434fa9e4066Sahrens void 3435fa9e4066Sahrens zio_vdev_io_reissue(zio_t *zio) 3436fa9e4066Sahrens { 3437fa9e4066Sahrens ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START); 3438fa9e4066Sahrens ASSERT(zio->io_error == 0); 3439fa9e4066Sahrens 3440b24ab676SJeff Bonwick zio->io_stage >>= 1; 3441fa9e4066Sahrens } 3442fa9e4066Sahrens 3443fa9e4066Sahrens void 3444fa9e4066Sahrens zio_vdev_io_redone(zio_t *zio) 3445fa9e4066Sahrens { 3446fa9e4066Sahrens ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE); 3447fa9e4066Sahrens 3448b24ab676SJeff Bonwick zio->io_stage >>= 1; 3449fa9e4066Sahrens } 3450fa9e4066Sahrens 3451fa9e4066Sahrens void 3452fa9e4066Sahrens zio_vdev_io_bypass(zio_t *zio) 3453fa9e4066Sahrens { 3454fa9e4066Sahrens ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START); 3455fa9e4066Sahrens ASSERT(zio->io_error == 0); 3456fa9e4066Sahrens 3457fa9e4066Sahrens zio->io_flags |= ZIO_FLAG_IO_BYPASS; 3458b24ab676SJeff Bonwick zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1; 3459fa9e4066Sahrens } 3460fa9e4066Sahrens 3461fa9e4066Sahrens /* 3462fa9e4066Sahrens * ========================================================================== 3463fa9e4066Sahrens * Generate and verify checksums 3464fa9e4066Sahrens * ========================================================================== 3465fa9e4066Sahrens */ 3466e05725b1Sbonwick static int 3467fa9e4066Sahrens zio_checksum_generate(zio_t *zio) 3468fa9e4066Sahrens { 3469fa9e4066Sahrens blkptr_t *bp = zio->io_bp; 3470e14bb325SJeff Bonwick enum zio_checksum checksum; 3471fa9e4066Sahrens 3472e14bb325SJeff Bonwick if (bp == NULL) { 3473e14bb325SJeff Bonwick /* 3474e14bb325SJeff Bonwick * This is zio_write_phys(). 3475e14bb325SJeff Bonwick * We're either generating a label checksum, or none at all. 3476e14bb325SJeff Bonwick */ 3477e14bb325SJeff Bonwick checksum = zio->io_prop.zp_checksum; 3478e14bb325SJeff Bonwick 3479e14bb325SJeff Bonwick if (checksum == ZIO_CHECKSUM_OFF) 3480e14bb325SJeff Bonwick return (ZIO_PIPELINE_CONTINUE); 3481fa9e4066Sahrens 3482e14bb325SJeff Bonwick ASSERT(checksum == ZIO_CHECKSUM_LABEL); 3483e14bb325SJeff Bonwick } else { 3484e14bb325SJeff Bonwick if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) { 3485e14bb325SJeff Bonwick ASSERT(!IO_IS_ALLOCATING(zio)); 3486e14bb325SJeff Bonwick checksum = ZIO_CHECKSUM_GANG_HEADER; 3487e14bb325SJeff Bonwick } else { 3488e14bb325SJeff Bonwick checksum = BP_GET_CHECKSUM(bp); 3489e14bb325SJeff Bonwick } 3490e14bb325SJeff Bonwick } 3491fa9e4066Sahrens 3492770499e1SDan Kimmel zio_checksum_compute(zio, checksum, zio->io_abd, zio->io_size); 3493fa9e4066Sahrens 3494e05725b1Sbonwick return (ZIO_PIPELINE_CONTINUE); 3495fa9e4066Sahrens } 3496fa9e4066Sahrens 3497e05725b1Sbonwick static int 3498e14bb325SJeff Bonwick zio_checksum_verify(zio_t *zio) 3499fa9e4066Sahrens { 350022fe2c88SJonathan Adams zio_bad_cksum_t info; 3501e14bb325SJeff Bonwick blkptr_t *bp = zio->io_bp; 3502e14bb325SJeff Bonwick int error; 3503fa9e4066Sahrens 3504b24ab676SJeff Bonwick ASSERT(zio->io_vd != NULL); 3505b24ab676SJeff Bonwick 3506e14bb325SJeff Bonwick if (bp == NULL) { 3507e14bb325SJeff Bonwick /* 3508e14bb325SJeff Bonwick * This is zio_read_phys(). 3509e14bb325SJeff Bonwick * We're either verifying a label checksum, or nothing at all. 3510e14bb325SJeff Bonwick */ 3511e14bb325SJeff Bonwick if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF) 3512e14bb325SJeff Bonwick return (ZIO_PIPELINE_CONTINUE); 3513fa9e4066Sahrens 3514e14bb325SJeff Bonwick ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL); 3515e14bb325SJeff Bonwick } 3516fa9e4066Sahrens 351722fe2c88SJonathan Adams if ((error = zio_checksum_error(zio, &info)) != 0) { 3518e14bb325SJeff Bonwick zio->io_error = error; 3519373dc1cfSMatthew Ahrens if (error == ECKSUM && 3520373dc1cfSMatthew Ahrens !(zio->io_flags & ZIO_FLAG_SPECULATIVE)) { 352122fe2c88SJonathan Adams zfs_ereport_start_checksum(zio->io_spa, 352222fe2c88SJonathan Adams zio->io_vd, zio, zio->io_offset, 352322fe2c88SJonathan Adams zio->io_size, NULL, &info); 3524e14bb325SJeff Bonwick } 3525fa9e4066Sahrens } 3526fa9e4066Sahrens 3527e05725b1Sbonwick return (ZIO_PIPELINE_CONTINUE); 3528fa9e4066Sahrens } 3529fa9e4066Sahrens 3530fa9e4066Sahrens /* 3531fa9e4066Sahrens * Called by RAID-Z to ensure we don't compute the checksum twice. 3532fa9e4066Sahrens */ 3533fa9e4066Sahrens void 3534fa9e4066Sahrens zio_checksum_verified(zio_t *zio) 3535fa9e4066Sahrens { 3536b24ab676SJeff Bonwick zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY; 3537fa9e4066Sahrens } 3538fa9e4066Sahrens 3539fa9e4066Sahrens /* 3540e14bb325SJeff Bonwick * ========================================================================== 3541e14bb325SJeff Bonwick * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other. 35425d7b4d43SMatthew Ahrens * An error of 0 indicates success. ENXIO indicates whole-device failure, 3543e14bb325SJeff Bonwick * which may be transient (e.g. unplugged) or permament. ECKSUM and EIO 3544e14bb325SJeff Bonwick * indicate errors that are specific to one I/O, and most likely permanent. 3545e14bb325SJeff Bonwick * Any other error is presumed to be worse because we weren't expecting it. 3546e14bb325SJeff Bonwick * ========================================================================== 3547fa9e4066Sahrens */ 3548e14bb325SJeff Bonwick int 3549e14bb325SJeff Bonwick zio_worst_error(int e1, int e2) 3550fa9e4066Sahrens { 3551e14bb325SJeff Bonwick static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO }; 3552e14bb325SJeff Bonwick int r1, r2; 3553e14bb325SJeff Bonwick 3554e14bb325SJeff Bonwick for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++) 3555e14bb325SJeff Bonwick if (e1 == zio_error_rank[r1]) 3556e14bb325SJeff Bonwick break; 3557e14bb325SJeff Bonwick 3558e14bb325SJeff Bonwick for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++) 3559e14bb325SJeff Bonwick if (e2 == zio_error_rank[r2]) 3560e14bb325SJeff Bonwick break; 356144cd46caSbillm 3562e14bb325SJeff Bonwick return (r1 > r2 ? e1 : e2); 3563fa9e4066Sahrens } 3564fa9e4066Sahrens 3565fa9e4066Sahrens /* 3566fa9e4066Sahrens * ========================================================================== 3567e14bb325SJeff Bonwick * I/O completion 3568fa9e4066Sahrens * ========================================================================== 3569fa9e4066Sahrens */ 3570e14bb325SJeff Bonwick static int 3571e14bb325SJeff Bonwick zio_ready(zio_t *zio) 3572fa9e4066Sahrens { 3573e14bb325SJeff Bonwick blkptr_t *bp = zio->io_bp; 3574a3f829aeSBill Moore zio_t *pio, *pio_next; 35750f7643c7SGeorge Wilson zio_link_t *zl = NULL; 3576fa9e4066Sahrens 3577d6e1c446SGeorge Wilson if (zio_wait_for_children(zio, ZIO_CHILD_GANG_BIT | ZIO_CHILD_DDT_BIT, 3578d6e1c446SGeorge Wilson ZIO_WAIT_READY)) { 3579f5383399SBill Moore return (ZIO_PIPELINE_STOP); 3580d6e1c446SGeorge Wilson } 3581fa9e4066Sahrens 3582f5383399SBill Moore if (zio->io_ready) { 3583e14bb325SJeff Bonwick ASSERT(IO_IS_ALLOCATING(zio)); 358480901aeaSGeorge Wilson ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp) || 358580901aeaSGeorge Wilson (zio->io_flags & ZIO_FLAG_NOPWRITE)); 3586e14bb325SJeff Bonwick ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0); 3587fa9e4066Sahrens 3588e14bb325SJeff Bonwick zio->io_ready(zio); 3589e14bb325SJeff Bonwick } 3590fa9e4066Sahrens 3591e14bb325SJeff Bonwick if (bp != NULL && bp != &zio->io_bp_copy) 3592e14bb325SJeff Bonwick zio->io_bp_copy = *bp; 3593fa9e4066Sahrens 35940f7643c7SGeorge Wilson if (zio->io_error != 0) { 3595e14bb325SJeff Bonwick zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; 3596fa9e4066Sahrens 35970f7643c7SGeorge Wilson if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING) { 35980f7643c7SGeorge Wilson ASSERT(IO_IS_ALLOCATING(zio)); 35990f7643c7SGeorge Wilson ASSERT(zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE); 3600663207adSDon Brady ASSERT(zio->io_metaslab_class != NULL); 3601663207adSDon Brady 36020f7643c7SGeorge Wilson /* 36030f7643c7SGeorge Wilson * We were unable to allocate anything, unreserve and 36040f7643c7SGeorge Wilson * issue the next I/O to allocate. 36050f7643c7SGeorge Wilson */ 36060f7643c7SGeorge Wilson metaslab_class_throttle_unreserve( 3607663207adSDon Brady zio->io_metaslab_class, zio->io_prop.zp_copies, 3608663207adSDon Brady zio->io_allocator, zio); 3609f78cdc34SPaul Dagnelie zio_allocate_dispatch(zio->io_spa, zio->io_allocator); 36100f7643c7SGeorge Wilson } 36110f7643c7SGeorge Wilson } 36120f7643c7SGeorge Wilson 3613a3f829aeSBill Moore mutex_enter(&zio->io_lock); 3614a3f829aeSBill Moore zio->io_state[ZIO_WAIT_READY] = 1; 36150f7643c7SGeorge Wilson pio = zio_walk_parents(zio, &zl); 3616a3f829aeSBill Moore mutex_exit(&zio->io_lock); 3617a3f829aeSBill Moore 3618a3f829aeSBill Moore /* 3619a3f829aeSBill Moore * As we notify zio's parents, new parents could be added. 3620a3f829aeSBill Moore * New parents go to the head of zio's io_parent_list, however, 3621a3f829aeSBill Moore * so we will (correctly) not notify them. The remainder of zio's 3622a3f829aeSBill Moore * io_parent_list, from 'pio_next' onward, cannot change because 3623a3f829aeSBill Moore * all parents must wait for us to be done before they can be done. 3624a3f829aeSBill Moore */ 3625a3f829aeSBill Moore for (; pio != NULL; pio = pio_next) { 36260f7643c7SGeorge Wilson pio_next = zio_walk_parents(zio, &zl); 3627e14bb325SJeff Bonwick zio_notify_parent(pio, zio, ZIO_WAIT_READY); 3628a3f829aeSBill Moore } 3629fa9e4066Sahrens 3630b24ab676SJeff Bonwick if (zio->io_flags & ZIO_FLAG_NODATA) { 3631b24ab676SJeff Bonwick if (BP_IS_GANG(bp)) { 3632b24ab676SJeff Bonwick zio->io_flags &= ~ZIO_FLAG_NODATA; 3633b24ab676SJeff Bonwick } else { 3634770499e1SDan Kimmel ASSERT((uintptr_t)zio->io_abd < SPA_MAXBLOCKSIZE); 3635b24ab676SJeff Bonwick zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES; 3636b24ab676SJeff Bonwick } 3637b24ab676SJeff Bonwick } 3638b24ab676SJeff Bonwick 3639a33cae98STim Haley if (zio_injection_enabled && 3640a33cae98STim Haley zio->io_spa->spa_syncing_txg == zio->io_txg) 3641a33cae98STim Haley zio_handle_ignored_writes(zio); 3642a33cae98STim Haley 3643e14bb325SJeff Bonwick return (ZIO_PIPELINE_CONTINUE); 3644fa9e4066Sahrens } 3645fa9e4066Sahrens 36460f7643c7SGeorge Wilson /* 36470f7643c7SGeorge Wilson * Update the allocation throttle accounting. 36480f7643c7SGeorge Wilson */ 36490f7643c7SGeorge Wilson static void 36500f7643c7SGeorge Wilson zio_dva_throttle_done(zio_t *zio) 36510f7643c7SGeorge Wilson { 36520f7643c7SGeorge Wilson zio_t *lio = zio->io_logical; 36530f7643c7SGeorge Wilson zio_t *pio = zio_unique_parent(zio); 36540f7643c7SGeorge Wilson vdev_t *vd = zio->io_vd; 36550f7643c7SGeorge Wilson int flags = METASLAB_ASYNC_ALLOC; 36560f7643c7SGeorge Wilson 36570f7643c7SGeorge Wilson ASSERT3P(zio->io_bp, !=, NULL); 36580f7643c7SGeorge Wilson ASSERT3U(zio->io_type, ==, ZIO_TYPE_WRITE); 36590f7643c7SGeorge Wilson ASSERT3U(zio->io_priority, ==, ZIO_PRIORITY_ASYNC_WRITE); 36600f7643c7SGeorge Wilson ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_VDEV); 36610f7643c7SGeorge Wilson ASSERT(vd != NULL); 36620f7643c7SGeorge Wilson ASSERT3P(vd, ==, vd->vdev_top); 36630f7643c7SGeorge Wilson ASSERT(!(zio->io_flags & (ZIO_FLAG_IO_REPAIR | ZIO_FLAG_IO_RETRY))); 36640f7643c7SGeorge Wilson ASSERT(zio->io_flags & ZIO_FLAG_IO_ALLOCATING); 36650f7643c7SGeorge Wilson ASSERT(!(lio->io_flags & ZIO_FLAG_IO_REWRITE)); 36660f7643c7SGeorge Wilson ASSERT(!(lio->io_orig_flags & ZIO_FLAG_NODATA)); 36670f7643c7SGeorge Wilson 36680f7643c7SGeorge Wilson /* 36690f7643c7SGeorge Wilson * Parents of gang children can have two flavors -- ones that 36700f7643c7SGeorge Wilson * allocated the gang header (will have ZIO_FLAG_IO_REWRITE set) 36710f7643c7SGeorge Wilson * and ones that allocated the constituent blocks. The allocation 36720f7643c7SGeorge Wilson * throttle needs to know the allocating parent zio so we must find 36730f7643c7SGeorge Wilson * it here. 36740f7643c7SGeorge Wilson */ 36750f7643c7SGeorge Wilson if (pio->io_child_type == ZIO_CHILD_GANG) { 36760f7643c7SGeorge Wilson /* 36770f7643c7SGeorge Wilson * If our parent is a rewrite gang child then our grandparent 36780f7643c7SGeorge Wilson * would have been the one that performed the allocation. 36790f7643c7SGeorge Wilson */ 36800f7643c7SGeorge Wilson if (pio->io_flags & ZIO_FLAG_IO_REWRITE) 36810f7643c7SGeorge Wilson pio = zio_unique_parent(pio); 36820f7643c7SGeorge Wilson flags |= METASLAB_GANG_CHILD; 36830f7643c7SGeorge Wilson } 36840f7643c7SGeorge Wilson 36850f7643c7SGeorge Wilson ASSERT(IO_IS_ALLOCATING(pio)); 36860f7643c7SGeorge Wilson ASSERT3P(zio, !=, zio->io_logical); 36870f7643c7SGeorge Wilson ASSERT(zio->io_logical != NULL); 36880f7643c7SGeorge Wilson ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REPAIR)); 36890f7643c7SGeorge Wilson ASSERT0(zio->io_flags & ZIO_FLAG_NOPWRITE); 3690663207adSDon Brady ASSERT(zio->io_metaslab_class != NULL); 36910f7643c7SGeorge Wilson 36920f7643c7SGeorge Wilson mutex_enter(&pio->io_lock); 3693f78cdc34SPaul Dagnelie metaslab_group_alloc_decrement(zio->io_spa, vd->vdev_id, pio, flags, 3694f78cdc34SPaul Dagnelie pio->io_allocator, B_TRUE); 36950f7643c7SGeorge Wilson mutex_exit(&pio->io_lock); 36960f7643c7SGeorge Wilson 3697663207adSDon Brady metaslab_class_throttle_unreserve(zio->io_metaslab_class, 1, 3698663207adSDon Brady pio->io_allocator, pio); 36990f7643c7SGeorge Wilson 37000f7643c7SGeorge Wilson /* 37010f7643c7SGeorge Wilson * Call into the pipeline to see if there is more work that 37020f7643c7SGeorge Wilson * needs to be done. If there is work to be done it will be 37030f7643c7SGeorge Wilson * dispatched to another taskq thread. 37040f7643c7SGeorge Wilson */ 3705f78cdc34SPaul Dagnelie zio_allocate_dispatch(zio->io_spa, pio->io_allocator); 37060f7643c7SGeorge Wilson } 37070f7643c7SGeorge Wilson 3708e14bb325SJeff Bonwick static int 3709e14bb325SJeff Bonwick zio_done(zio_t *zio) 3710d63d470bSgw { 3711e14bb325SJeff Bonwick spa_t *spa = zio->io_spa; 3712e14bb325SJeff Bonwick zio_t *lio = zio->io_logical; 3713e14bb325SJeff Bonwick blkptr_t *bp = zio->io_bp; 3714e14bb325SJeff Bonwick vdev_t *vd = zio->io_vd; 3715e14bb325SJeff Bonwick uint64_t psize = zio->io_size; 3716a3f829aeSBill Moore zio_t *pio, *pio_next; 37170f7643c7SGeorge Wilson zio_link_t *zl = NULL; 3718d63d470bSgw 3719e14bb325SJeff Bonwick /* 3720f5383399SBill Moore * If our children haven't all completed, 3721e14bb325SJeff Bonwick * wait for them and then repeat this pipeline stage. 3722e14bb325SJeff Bonwick */ 3723d6e1c446SGeorge Wilson if (zio_wait_for_children(zio, ZIO_CHILD_ALL_BITS, ZIO_WAIT_DONE)) { 3724e14bb325SJeff Bonwick return (ZIO_PIPELINE_STOP); 3725d6e1c446SGeorge Wilson } 3726d63d470bSgw 37270f7643c7SGeorge Wilson /* 37280f7643c7SGeorge Wilson * If the allocation throttle is enabled, then update the accounting. 37290f7643c7SGeorge Wilson * We only track child I/Os that are part of an allocating async 37300f7643c7SGeorge Wilson * write. We must do this since the allocation is performed 37310f7643c7SGeorge Wilson * by the logical I/O but the actual write is done by child I/Os. 37320f7643c7SGeorge Wilson */ 37330f7643c7SGeorge Wilson if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING && 37340f7643c7SGeorge Wilson zio->io_child_type == ZIO_CHILD_VDEV) { 3735663207adSDon Brady ASSERT(zio->io_metaslab_class != NULL); 3736663207adSDon Brady ASSERT(zio->io_metaslab_class->mc_alloc_throttle_enabled); 37370f7643c7SGeorge Wilson zio_dva_throttle_done(zio); 37380f7643c7SGeorge Wilson } 37390f7643c7SGeorge Wilson 37400f7643c7SGeorge Wilson /* 37410f7643c7SGeorge Wilson * If the allocation throttle is enabled, verify that 37420f7643c7SGeorge Wilson * we have decremented the refcounts for every I/O that was throttled. 37430f7643c7SGeorge Wilson */ 37440f7643c7SGeorge Wilson if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING) { 37450f7643c7SGeorge Wilson ASSERT(zio->io_type == ZIO_TYPE_WRITE); 37460f7643c7SGeorge Wilson ASSERT(zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE); 37470f7643c7SGeorge Wilson ASSERT(bp != NULL); 3748663207adSDon Brady 3749f78cdc34SPaul Dagnelie metaslab_group_alloc_verify(spa, zio->io_bp, zio, 3750f78cdc34SPaul Dagnelie zio->io_allocator); 3751e914ace2STim Schumacher VERIFY(zfs_refcount_not_held( 3752663207adSDon Brady &zio->io_metaslab_class->mc_alloc_slots[zio->io_allocator], 3753663207adSDon Brady zio)); 37540f7643c7SGeorge Wilson } 37550f7643c7SGeorge Wilson 3756e14bb325SJeff Bonwick for (int c = 0; c < ZIO_CHILD_TYPES; c++) 3757e14bb325SJeff Bonwick for (int w = 0; w < ZIO_WAIT_TYPES; w++) 3758e14bb325SJeff Bonwick ASSERT(zio->io_children[c][w] == 0); 3759e14bb325SJeff Bonwick 37605d7b4d43SMatthew Ahrens if (bp != NULL && !BP_IS_EMBEDDED(bp)) { 3761e14bb325SJeff Bonwick ASSERT(bp->blk_pad[0] == 0); 3762e14bb325SJeff Bonwick ASSERT(bp->blk_pad[1] == 0); 3763e14bb325SJeff Bonwick ASSERT(bcmp(bp, &zio->io_bp_copy, sizeof (blkptr_t)) == 0 || 3764a3f829aeSBill Moore (bp == zio_unique_parent(zio)->io_bp)); 3765e14bb325SJeff Bonwick if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(bp) && 3766b24ab676SJeff Bonwick zio->io_bp_override == NULL && 3767e14bb325SJeff Bonwick !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) { 3768e14bb325SJeff Bonwick ASSERT(!BP_SHOULD_BYTESWAP(bp)); 3769b24ab676SJeff Bonwick ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(bp)); 3770e14bb325SJeff Bonwick ASSERT(BP_COUNT_GANG(bp) == 0 || 3771e14bb325SJeff Bonwick (BP_COUNT_GANG(bp) == BP_GET_NDVAS(bp))); 3772e14bb325SJeff Bonwick } 377380901aeaSGeorge Wilson if (zio->io_flags & ZIO_FLAG_NOPWRITE) 377480901aeaSGeorge Wilson VERIFY(BP_EQUAL(bp, &zio->io_bp_orig)); 3775e14bb325SJeff Bonwick } 3776fa9e4066Sahrens 3777e14bb325SJeff Bonwick /* 3778b24ab676SJeff Bonwick * If there were child vdev/gang/ddt errors, they apply to us now. 3779e14bb325SJeff Bonwick */ 3780e14bb325SJeff Bonwick zio_inherit_child_errors(zio, ZIO_CHILD_VDEV); 3781e14bb325SJeff Bonwick zio_inherit_child_errors(zio, ZIO_CHILD_GANG); 3782b24ab676SJeff Bonwick zio_inherit_child_errors(zio, ZIO_CHILD_DDT); 3783b24ab676SJeff Bonwick 3784b24ab676SJeff Bonwick /* 3785b24ab676SJeff Bonwick * If the I/O on the transformed data was successful, generate any 3786b24ab676SJeff Bonwick * checksum reports now while we still have the transformed data. 3787b24ab676SJeff Bonwick */ 3788b24ab676SJeff Bonwick if (zio->io_error == 0) { 3789b24ab676SJeff Bonwick while (zio->io_cksum_report != NULL) { 3790b24ab676SJeff Bonwick zio_cksum_report_t *zcr = zio->io_cksum_report; 3791b24ab676SJeff Bonwick uint64_t align = zcr->zcr_align; 3792b24ab676SJeff Bonwick uint64_t asize = P2ROUNDUP(psize, align); 3793770499e1SDan Kimmel char *abuf = NULL; 3794770499e1SDan Kimmel abd_t *adata = zio->io_abd; 3795b24ab676SJeff Bonwick 3796b24ab676SJeff Bonwick if (asize != psize) { 3797770499e1SDan Kimmel adata = abd_alloc_linear(asize, B_TRUE); 3798770499e1SDan Kimmel abd_copy(adata, zio->io_abd, psize); 3799770499e1SDan Kimmel abd_zero_off(adata, psize, asize - psize); 3800b24ab676SJeff Bonwick } 3801b24ab676SJeff Bonwick 3802770499e1SDan Kimmel if (adata != NULL) 3803770499e1SDan Kimmel abuf = abd_borrow_buf_copy(adata, asize); 3804770499e1SDan Kimmel 3805b24ab676SJeff Bonwick zio->io_cksum_report = zcr->zcr_next; 3806b24ab676SJeff Bonwick zcr->zcr_next = NULL; 3807b24ab676SJeff Bonwick zcr->zcr_finish(zcr, abuf); 3808b24ab676SJeff Bonwick zfs_ereport_free_checksum(zcr); 3809b24ab676SJeff Bonwick 3810770499e1SDan Kimmel if (adata != NULL) 3811770499e1SDan Kimmel abd_return_buf(adata, abuf, asize); 3812770499e1SDan Kimmel 3813b24ab676SJeff Bonwick if (asize != psize) 3814770499e1SDan Kimmel abd_free(adata); 3815b24ab676SJeff Bonwick } 3816b24ab676SJeff Bonwick } 3817e14bb325SJeff Bonwick 3818e14bb325SJeff Bonwick zio_pop_transforms(zio); /* note: may set zio->io_error */ 3819e14bb325SJeff Bonwick 3820e14bb325SJeff Bonwick vdev_stat_update(zio, psize); 3821e14bb325SJeff Bonwick 3822e14bb325SJeff Bonwick if (zio->io_error) { 3823e14bb325SJeff Bonwick /* 3824e14bb325SJeff Bonwick * If this I/O is attached to a particular vdev, 3825e14bb325SJeff Bonwick * generate an error message describing the I/O failure 3826e14bb325SJeff Bonwick * at the block level. We ignore these errors if the 3827e14bb325SJeff Bonwick * device is currently unavailable. 3828e14bb325SJeff Bonwick */ 3829e14bb325SJeff Bonwick if (zio->io_error != ECKSUM && vd != NULL && !vdev_is_dead(vd)) 3830e14bb325SJeff Bonwick zfs_ereport_post(FM_EREPORT_ZFS_IO, spa, vd, zio, 0, 0); 3831e14bb325SJeff Bonwick 38328f18d1faSGeorge Wilson if ((zio->io_error == EIO || !(zio->io_flags & 38338f18d1faSGeorge Wilson (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) && 38348f18d1faSGeorge Wilson zio == lio) { 3835e14bb325SJeff Bonwick /* 3836e14bb325SJeff Bonwick * For logical I/O requests, tell the SPA to log the 3837e14bb325SJeff Bonwick * error and generate a logical data ereport. 3838e14bb325SJeff Bonwick */ 3839e14bb325SJeff Bonwick spa_log_error(spa, zio); 3840e14bb325SJeff Bonwick zfs_ereport_post(FM_EREPORT_ZFS_DATA, spa, NULL, zio, 3841e14bb325SJeff Bonwick 0, 0); 3842e14bb325SJeff Bonwick } 3843e14bb325SJeff Bonwick } 3844fa9e4066Sahrens 3845e14bb325SJeff Bonwick if (zio->io_error && zio == lio) { 3846e14bb325SJeff Bonwick /* 3847e14bb325SJeff Bonwick * Determine whether zio should be reexecuted. This will 3848e14bb325SJeff Bonwick * propagate all the way to the root via zio_notify_parent(). 3849e14bb325SJeff Bonwick */ 3850e14bb325SJeff Bonwick ASSERT(vd == NULL && bp != NULL); 3851b24ab676SJeff Bonwick ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 3852e14bb325SJeff Bonwick 3853b24ab676SJeff Bonwick if (IO_IS_ALLOCATING(zio) && 3854b24ab676SJeff Bonwick !(zio->io_flags & ZIO_FLAG_CANFAIL)) { 3855e14bb325SJeff Bonwick if (zio->io_error != ENOSPC) 3856e14bb325SJeff Bonwick zio->io_reexecute |= ZIO_REEXECUTE_NOW; 3857e14bb325SJeff Bonwick else 3858e14bb325SJeff Bonwick zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND; 3859b24ab676SJeff Bonwick } 3860e14bb325SJeff Bonwick 3861e14bb325SJeff Bonwick if ((zio->io_type == ZIO_TYPE_READ || 3862e14bb325SJeff Bonwick zio->io_type == ZIO_TYPE_FREE) && 386344ecc532SGeorge Wilson !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && 3864e14bb325SJeff Bonwick zio->io_error == ENXIO && 3865b16da2e2SGeorge Wilson spa_load_state(spa) == SPA_LOAD_NONE && 3866e14bb325SJeff Bonwick spa_get_failmode(spa) != ZIO_FAILURE_MODE_CONTINUE) 3867e14bb325SJeff Bonwick zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND; 3868e14bb325SJeff Bonwick 3869e14bb325SJeff Bonwick if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute) 3870e14bb325SJeff Bonwick zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND; 387122fe2c88SJonathan Adams 387222fe2c88SJonathan Adams /* 387322fe2c88SJonathan Adams * Here is a possibly good place to attempt to do 387422fe2c88SJonathan Adams * either combinatorial reconstruction or error correction 387522fe2c88SJonathan Adams * based on checksums. It also might be a good place 387622fe2c88SJonathan Adams * to send out preliminary ereports before we suspend 387722fe2c88SJonathan Adams * processing. 387822fe2c88SJonathan Adams */ 3879d63d470bSgw } 3880d63d470bSgw 388167bd71c6Sperrin /* 3882e14bb325SJeff Bonwick * If there were logical child errors, they apply to us now. 3883e14bb325SJeff Bonwick * We defer this until now to avoid conflating logical child 3884e14bb325SJeff Bonwick * errors with errors that happened to the zio itself when 3885e14bb325SJeff Bonwick * updating vdev stats and reporting FMA events above. 388667bd71c6Sperrin */ 3887e14bb325SJeff Bonwick zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL); 38888654d025Sperrin 3889b24ab676SJeff Bonwick if ((zio->io_error || zio->io_reexecute) && 3890b24ab676SJeff Bonwick IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio && 389180901aeaSGeorge Wilson !(zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE))) 3892f5383399SBill Moore zio_dva_unallocate(zio, zio->io_gang_tree, bp); 3893f5383399SBill Moore 3894f5383399SBill Moore zio_gang_tree_free(&zio->io_gang_tree); 3895f5383399SBill Moore 389633a372edSGeorge Wilson /* 389733a372edSGeorge Wilson * Godfather I/Os should never suspend. 389833a372edSGeorge Wilson */ 389933a372edSGeorge Wilson if ((zio->io_flags & ZIO_FLAG_GODFATHER) && 390033a372edSGeorge Wilson (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) 390133a372edSGeorge Wilson zio->io_reexecute = 0; 390233a372edSGeorge Wilson 390333a372edSGeorge Wilson if (zio->io_reexecute) { 3904e14bb325SJeff Bonwick /* 3905e14bb325SJeff Bonwick * This is a logical I/O that wants to reexecute. 3906e14bb325SJeff Bonwick * 3907e14bb325SJeff Bonwick * Reexecute is top-down. When an i/o fails, if it's not 3908e14bb325SJeff Bonwick * the root, it simply notifies its parent and sticks around. 3909e14bb325SJeff Bonwick * The parent, seeing that it still has children in zio_done(), 3910e14bb325SJeff Bonwick * does the same. This percolates all the way up to the root. 3911e14bb325SJeff Bonwick * The root i/o will reexecute or suspend the entire tree. 3912e14bb325SJeff Bonwick * 3913e14bb325SJeff Bonwick * This approach ensures that zio_reexecute() honors 3914e14bb325SJeff Bonwick * all the original i/o dependency relationships, e.g. 3915e14bb325SJeff Bonwick * parents not executing until children are ready. 3916e14bb325SJeff Bonwick */ 3917e14bb325SJeff Bonwick ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); 3918fa9e4066Sahrens 3919f5383399SBill Moore zio->io_gang_leader = NULL; 3920e14bb325SJeff Bonwick 3921a3f829aeSBill Moore mutex_enter(&zio->io_lock); 3922a3f829aeSBill Moore zio->io_state[ZIO_WAIT_DONE] = 1; 3923a3f829aeSBill Moore mutex_exit(&zio->io_lock); 3924a3f829aeSBill Moore 392554d692b7SGeorge Wilson /* 392654d692b7SGeorge Wilson * "The Godfather" I/O monitors its children but is 392754d692b7SGeorge Wilson * not a true parent to them. It will track them through 392854d692b7SGeorge Wilson * the pipeline but severs its ties whenever they get into 392954d692b7SGeorge Wilson * trouble (e.g. suspended). This allows "The Godfather" 393054d692b7SGeorge Wilson * I/O to return status without blocking. 393154d692b7SGeorge Wilson */ 39320f7643c7SGeorge Wilson zl = NULL; 39330f7643c7SGeorge Wilson for (pio = zio_walk_parents(zio, &zl); pio != NULL; 39340f7643c7SGeorge Wilson pio = pio_next) { 39350f7643c7SGeorge Wilson zio_link_t *remove_zl = zl; 39360f7643c7SGeorge Wilson pio_next = zio_walk_parents(zio, &zl); 393754d692b7SGeorge Wilson 393854d692b7SGeorge Wilson if ((pio->io_flags & ZIO_FLAG_GODFATHER) && 393954d692b7SGeorge Wilson (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) { 39400f7643c7SGeorge Wilson zio_remove_child(pio, zio, remove_zl); 394154d692b7SGeorge Wilson zio_notify_parent(pio, zio, ZIO_WAIT_DONE); 394254d692b7SGeorge Wilson } 394354d692b7SGeorge Wilson } 394454d692b7SGeorge Wilson 3945a3f829aeSBill Moore if ((pio = zio_unique_parent(zio)) != NULL) { 3946e14bb325SJeff Bonwick /* 3947e14bb325SJeff Bonwick * We're not a root i/o, so there's nothing to do 3948e14bb325SJeff Bonwick * but notify our parent. Don't propagate errors 3949e14bb325SJeff Bonwick * upward since we haven't permanently failed yet. 3950e14bb325SJeff Bonwick */ 395133a372edSGeorge Wilson ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER)); 3952e14bb325SJeff Bonwick zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE; 3953e14bb325SJeff Bonwick zio_notify_parent(pio, zio, ZIO_WAIT_DONE); 3954e14bb325SJeff Bonwick } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) { 3955e14bb325SJeff Bonwick /* 3956e14bb325SJeff Bonwick * We'd fail again if we reexecuted now, so suspend 3957e14bb325SJeff Bonwick * until conditions improve (e.g. device comes online). 3958e14bb325SJeff Bonwick */ 3959e0f1c0afSOlaf Faaland zio_suspend(zio->io_spa, zio, ZIO_SUSPEND_IOERR); 3960e14bb325SJeff Bonwick } else { 3961e14bb325SJeff Bonwick /* 3962e14bb325SJeff Bonwick * Reexecution is potentially a huge amount of work. 3963e14bb325SJeff Bonwick * Hand it off to the otherwise-unused claim taskq. 3964e14bb325SJeff Bonwick */ 39655aeb9474SGarrett D'Amore ASSERT(zio->io_tqent.tqent_next == NULL); 3966ec94d322SAdam Leventhal spa_taskq_dispatch_ent(spa, ZIO_TYPE_CLAIM, 3967ec94d322SAdam Leventhal ZIO_TASKQ_ISSUE, (task_func_t *)zio_reexecute, zio, 3968ec94d322SAdam Leventhal 0, &zio->io_tqent); 3969e14bb325SJeff Bonwick } 3970e14bb325SJeff Bonwick return (ZIO_PIPELINE_STOP); 3971fa9e4066Sahrens } 3972fa9e4066Sahrens 3973b24ab676SJeff Bonwick ASSERT(zio->io_child_count == 0); 397433a372edSGeorge Wilson ASSERT(zio->io_reexecute == 0); 3975e14bb325SJeff Bonwick ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL)); 3976fa9e4066Sahrens 3977b24ab676SJeff Bonwick /* 3978b24ab676SJeff Bonwick * Report any checksum errors, since the I/O is complete. 3979b24ab676SJeff Bonwick */ 398022fe2c88SJonathan Adams while (zio->io_cksum_report != NULL) { 3981b24ab676SJeff Bonwick zio_cksum_report_t *zcr = zio->io_cksum_report; 3982b24ab676SJeff Bonwick zio->io_cksum_report = zcr->zcr_next; 3983b24ab676SJeff Bonwick zcr->zcr_next = NULL; 3984b24ab676SJeff Bonwick zcr->zcr_finish(zcr, NULL); 3985b24ab676SJeff Bonwick zfs_ereport_free_checksum(zcr); 398622fe2c88SJonathan Adams } 398722fe2c88SJonathan Adams 3988a3f829aeSBill Moore /* 3989a3f829aeSBill Moore * It is the responsibility of the done callback to ensure that this 3990a3f829aeSBill Moore * particular zio is no longer discoverable for adoption, and as 3991a3f829aeSBill Moore * such, cannot acquire any new parents. 3992a3f829aeSBill Moore */ 3993e14bb325SJeff Bonwick if (zio->io_done) 3994e14bb325SJeff Bonwick zio->io_done(zio); 3995fa9e4066Sahrens 3996a3f829aeSBill Moore mutex_enter(&zio->io_lock); 3997a3f829aeSBill Moore zio->io_state[ZIO_WAIT_DONE] = 1; 3998a3f829aeSBill Moore mutex_exit(&zio->io_lock); 3999fa9e4066Sahrens 40000f7643c7SGeorge Wilson zl = NULL; 40010f7643c7SGeorge Wilson for (pio = zio_walk_parents(zio, &zl); pio != NULL; pio = pio_next) { 40020f7643c7SGeorge Wilson zio_link_t *remove_zl = zl; 40030f7643c7SGeorge Wilson pio_next = zio_walk_parents(zio, &zl); 40040f7643c7SGeorge Wilson zio_remove_child(pio, zio, remove_zl); 4005e14bb325SJeff Bonwick zio_notify_parent(pio, zio, ZIO_WAIT_DONE); 4006e14bb325SJeff Bonwick } 4007fa9e4066Sahrens 4008e14bb325SJeff Bonwick if (zio->io_waiter != NULL) { 4009e14bb325SJeff Bonwick mutex_enter(&zio->io_lock); 4010e14bb325SJeff Bonwick zio->io_executor = NULL; 4011e14bb325SJeff Bonwick cv_broadcast(&zio->io_cv); 4012e14bb325SJeff Bonwick mutex_exit(&zio->io_lock); 4013e14bb325SJeff Bonwick } else { 4014e14bb325SJeff Bonwick zio_destroy(zio); 4015e14bb325SJeff Bonwick } 4016fa9e4066Sahrens 4017e14bb325SJeff Bonwick return (ZIO_PIPELINE_STOP); 4018fa9e4066Sahrens } 401946341222Sperrin 402046341222Sperrin /* 4021e14bb325SJeff Bonwick * ========================================================================== 4022e14bb325SJeff Bonwick * I/O pipeline definition 4023e14bb325SJeff Bonwick * ========================================================================== 402446341222Sperrin */ 4025b24ab676SJeff Bonwick static zio_pipe_stage_t *zio_pipeline[] = { 4026e14bb325SJeff Bonwick NULL, 4027e14bb325SJeff Bonwick zio_read_bp_init, 40280f7643c7SGeorge Wilson zio_write_bp_init, 4029b24ab676SJeff Bonwick zio_free_bp_init, 4030b24ab676SJeff Bonwick zio_issue_async, 40310f7643c7SGeorge Wilson zio_write_compress, 4032e14bb325SJeff Bonwick zio_checksum_generate, 403380901aeaSGeorge Wilson zio_nop_write, 4034b24ab676SJeff Bonwick zio_ddt_read_start, 4035b24ab676SJeff Bonwick zio_ddt_read_done, 4036b24ab676SJeff Bonwick zio_ddt_write, 4037b24ab676SJeff Bonwick zio_ddt_free, 4038e14bb325SJeff Bonwick zio_gang_assemble, 4039e14bb325SJeff Bonwick zio_gang_issue, 40400f7643c7SGeorge Wilson zio_dva_throttle, 4041e14bb325SJeff Bonwick zio_dva_allocate, 4042e14bb325SJeff Bonwick zio_dva_free, 4043e14bb325SJeff Bonwick zio_dva_claim, 4044e14bb325SJeff Bonwick zio_ready, 4045e14bb325SJeff Bonwick zio_vdev_io_start, 4046e14bb325SJeff Bonwick zio_vdev_io_done, 4047e14bb325SJeff Bonwick zio_vdev_io_assess, 4048e14bb325SJeff Bonwick zio_checksum_verify, 4049e14bb325SJeff Bonwick zio_done 4050e14bb325SJeff Bonwick }; 4051ad135b5dSChristopher Siden 4052ad135b5dSChristopher Siden 4053ad135b5dSChristopher Siden 4054ad135b5dSChristopher Siden 4055a2cdcdd2SPaul Dagnelie /* 4056a2cdcdd2SPaul Dagnelie * Compare two zbookmark_phys_t's to see which we would reach first in a 4057a2cdcdd2SPaul Dagnelie * pre-order traversal of the object tree. 4058a2cdcdd2SPaul Dagnelie * 4059a2cdcdd2SPaul Dagnelie * This is simple in every case aside from the meta-dnode object. For all other 4060a2cdcdd2SPaul Dagnelie * objects, we traverse them in order (object 1 before object 2, and so on). 4061a2cdcdd2SPaul Dagnelie * However, all of these objects are traversed while traversing object 0, since 4062a2cdcdd2SPaul Dagnelie * the data it points to is the list of objects. Thus, we need to convert to a 4063a2cdcdd2SPaul Dagnelie * canonical representation so we can compare meta-dnode bookmarks to 4064a2cdcdd2SPaul Dagnelie * non-meta-dnode bookmarks. 4065a2cdcdd2SPaul Dagnelie * 4066a2cdcdd2SPaul Dagnelie * We do this by calculating "equivalents" for each field of the zbookmark. 4067a2cdcdd2SPaul Dagnelie * zbookmarks outside of the meta-dnode use their own object and level, and 4068a2cdcdd2SPaul Dagnelie * calculate the level 0 equivalent (the first L0 blkid that is contained in the 4069a2cdcdd2SPaul Dagnelie * blocks this bookmark refers to) by multiplying their blkid by their span 4070a2cdcdd2SPaul Dagnelie * (the number of L0 blocks contained within one block at their level). 4071a2cdcdd2SPaul Dagnelie * zbookmarks inside the meta-dnode calculate their object equivalent 4072a2cdcdd2SPaul Dagnelie * (which is L0equiv * dnodes per data block), use 0 for their L0equiv, and use 4073a2cdcdd2SPaul Dagnelie * level + 1<<31 (any value larger than a level could ever be) for their level. 4074a2cdcdd2SPaul Dagnelie * This causes them to always compare before a bookmark in their object 4075a2cdcdd2SPaul Dagnelie * equivalent, compare appropriately to bookmarks in other objects, and to 4076a2cdcdd2SPaul Dagnelie * compare appropriately to other bookmarks in the meta-dnode. 4077a2cdcdd2SPaul Dagnelie */ 4078a2cdcdd2SPaul Dagnelie int 4079a2cdcdd2SPaul Dagnelie zbookmark_compare(uint16_t dbss1, uint8_t ibs1, uint16_t dbss2, uint8_t ibs2, 4080a2cdcdd2SPaul Dagnelie const zbookmark_phys_t *zb1, const zbookmark_phys_t *zb2) 4081a2cdcdd2SPaul Dagnelie { 4082a2cdcdd2SPaul Dagnelie /* 4083a2cdcdd2SPaul Dagnelie * These variables represent the "equivalent" values for the zbookmark, 4084a2cdcdd2SPaul Dagnelie * after converting zbookmarks inside the meta dnode to their 4085a2cdcdd2SPaul Dagnelie * normal-object equivalents. 4086a2cdcdd2SPaul Dagnelie */ 4087a2cdcdd2SPaul Dagnelie uint64_t zb1obj, zb2obj; 4088a2cdcdd2SPaul Dagnelie uint64_t zb1L0, zb2L0; 4089a2cdcdd2SPaul Dagnelie uint64_t zb1level, zb2level; 4090ad135b5dSChristopher Siden 4091a2cdcdd2SPaul Dagnelie if (zb1->zb_object == zb2->zb_object && 4092a2cdcdd2SPaul Dagnelie zb1->zb_level == zb2->zb_level && 4093a2cdcdd2SPaul Dagnelie zb1->zb_blkid == zb2->zb_blkid) 4094a2cdcdd2SPaul Dagnelie return (0); 4095a2cdcdd2SPaul Dagnelie 4096a2cdcdd2SPaul Dagnelie /* 4097a2cdcdd2SPaul Dagnelie * BP_SPANB calculates the span in blocks. 4098a2cdcdd2SPaul Dagnelie */ 4099a2cdcdd2SPaul Dagnelie zb1L0 = (zb1->zb_blkid) * BP_SPANB(ibs1, zb1->zb_level); 4100a2cdcdd2SPaul Dagnelie zb2L0 = (zb2->zb_blkid) * BP_SPANB(ibs2, zb2->zb_level); 4101ad135b5dSChristopher Siden 4102ad135b5dSChristopher Siden if (zb1->zb_object == DMU_META_DNODE_OBJECT) { 4103a2cdcdd2SPaul Dagnelie zb1obj = zb1L0 * (dbss1 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT)); 4104a2cdcdd2SPaul Dagnelie zb1L0 = 0; 4105a2cdcdd2SPaul Dagnelie zb1level = zb1->zb_level + COMPARE_META_LEVEL; 4106a2cdcdd2SPaul Dagnelie } else { 4107a2cdcdd2SPaul Dagnelie zb1obj = zb1->zb_object; 4108a2cdcdd2SPaul Dagnelie zb1level = zb1->zb_level; 4109ad135b5dSChristopher Siden } 4110ad135b5dSChristopher Siden 4111a2cdcdd2SPaul Dagnelie if (zb2->zb_object == DMU_META_DNODE_OBJECT) { 4112a2cdcdd2SPaul Dagnelie zb2obj = zb2L0 * (dbss2 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT)); 4113a2cdcdd2SPaul Dagnelie zb2L0 = 0; 4114a2cdcdd2SPaul Dagnelie zb2level = zb2->zb_level + COMPARE_META_LEVEL; 4115a2cdcdd2SPaul Dagnelie } else { 4116a2cdcdd2SPaul Dagnelie zb2obj = zb2->zb_object; 4117a2cdcdd2SPaul Dagnelie zb2level = zb2->zb_level; 4118a2cdcdd2SPaul Dagnelie } 4119a2cdcdd2SPaul Dagnelie 4120a2cdcdd2SPaul Dagnelie /* Now that we have a canonical representation, do the comparison. */ 4121a2cdcdd2SPaul Dagnelie if (zb1obj != zb2obj) 4122a2cdcdd2SPaul Dagnelie return (zb1obj < zb2obj ? -1 : 1); 4123a2cdcdd2SPaul Dagnelie else if (zb1L0 != zb2L0) 4124a2cdcdd2SPaul Dagnelie return (zb1L0 < zb2L0 ? -1 : 1); 4125a2cdcdd2SPaul Dagnelie else if (zb1level != zb2level) 4126a2cdcdd2SPaul Dagnelie return (zb1level > zb2level ? -1 : 1); 4127a2cdcdd2SPaul Dagnelie /* 4128a2cdcdd2SPaul Dagnelie * This can (theoretically) happen if the bookmarks have the same object 4129a2cdcdd2SPaul Dagnelie * and level, but different blkids, if the block sizes are not the same. 4130a2cdcdd2SPaul Dagnelie * There is presently no way to change the indirect block sizes 4131a2cdcdd2SPaul Dagnelie */ 4132a2cdcdd2SPaul Dagnelie return (0); 4133a2cdcdd2SPaul Dagnelie } 4134a2cdcdd2SPaul Dagnelie 4135a2cdcdd2SPaul Dagnelie /* 4136a2cdcdd2SPaul Dagnelie * This function checks the following: given that last_block is the place that 4137a2cdcdd2SPaul Dagnelie * our traversal stopped last time, does that guarantee that we've visited 4138a2cdcdd2SPaul Dagnelie * every node under subtree_root? Therefore, we can't just use the raw output 4139a2cdcdd2SPaul Dagnelie * of zbookmark_compare. We have to pass in a modified version of 4140a2cdcdd2SPaul Dagnelie * subtree_root; by incrementing the block id, and then checking whether 4141a2cdcdd2SPaul Dagnelie * last_block is before or equal to that, we can tell whether or not having 4142a2cdcdd2SPaul Dagnelie * visited last_block implies that all of subtree_root's children have been 4143a2cdcdd2SPaul Dagnelie * visited. 4144a2cdcdd2SPaul Dagnelie */ 4145a2cdcdd2SPaul Dagnelie boolean_t 4146a2cdcdd2SPaul Dagnelie zbookmark_subtree_completed(const dnode_phys_t *dnp, 4147a2cdcdd2SPaul Dagnelie const zbookmark_phys_t *subtree_root, const zbookmark_phys_t *last_block) 4148a2cdcdd2SPaul Dagnelie { 4149a2cdcdd2SPaul Dagnelie zbookmark_phys_t mod_zb = *subtree_root; 4150a2cdcdd2SPaul Dagnelie mod_zb.zb_blkid++; 4151a2cdcdd2SPaul Dagnelie ASSERT(last_block->zb_level == 0); 4152a2cdcdd2SPaul Dagnelie 4153a2cdcdd2SPaul Dagnelie /* The objset_phys_t isn't before anything. */ 4154a2cdcdd2SPaul Dagnelie if (dnp == NULL) 4155ad135b5dSChristopher Siden return (B_FALSE); 4156a2cdcdd2SPaul Dagnelie 4157a2cdcdd2SPaul Dagnelie /* 4158a2cdcdd2SPaul Dagnelie * We pass in 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT) for the 4159a2cdcdd2SPaul Dagnelie * data block size in sectors, because that variable is only used if 4160a2cdcdd2SPaul Dagnelie * the bookmark refers to a block in the meta-dnode. Since we don't 4161a2cdcdd2SPaul Dagnelie * know without examining it what object it refers to, and there's no 4162a2cdcdd2SPaul Dagnelie * harm in passing in this value in other cases, we always pass it in. 4163a2cdcdd2SPaul Dagnelie * 4164a2cdcdd2SPaul Dagnelie * We pass in 0 for the indirect block size shift because zb2 must be 4165a2cdcdd2SPaul Dagnelie * level 0. The indirect block size is only used to calculate the span 4166a2cdcdd2SPaul Dagnelie * of the bookmark, but since the bookmark must be level 0, the span is 4167a2cdcdd2SPaul Dagnelie * always 1, so the math works out. 4168a2cdcdd2SPaul Dagnelie * 4169a2cdcdd2SPaul Dagnelie * If you make changes to how the zbookmark_compare code works, be sure 4170a2cdcdd2SPaul Dagnelie * to make sure that this code still works afterwards. 4171a2cdcdd2SPaul Dagnelie */ 4172a2cdcdd2SPaul Dagnelie return (zbookmark_compare(dnp->dn_datablkszsec, dnp->dn_indblkshift, 4173a2cdcdd2SPaul Dagnelie 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT), 0, &mod_zb, 4174a2cdcdd2SPaul Dagnelie last_block) <= 0); 4175ad135b5dSChristopher Siden } 4176