sys/cddl/contrib/opensolaris/uts/common/fs/zfs/dsl_scan.c

   1 /*
   2  * CDDL HEADER START
   3  *
   4  * The contents of this file are subject to the terms of the
   5  * Common Development and Distribution License (the "License").
   6  * You may not use this file except in compliance with the License.
   7  *
   8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
   9  * or http://www.opensolaris.org/os/licensing.
  10  * See the License for the specific language governing permissions
  11  * and limitations under the License.
  12  *
  13  * When distributing Covered Code, include this CDDL HEADER in each
  14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
  15  * If applicable, add the following below this CDDL HEADER, with the
  16  * fields enclosed by brackets "[]" replaced with your own identifying
  17  * information: Portions Copyright [yyyy] [name of copyright owner]
  18  *
  19  * CDDL HEADER END
  20  */
  21 /*
  22  * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
  23  * Copyright (c) 2011, 2017 by Delphix. All rights reserved.
  24  * Copyright 2016 Gary Mills
  25  * Copyright (c) 2011, 2017 by Delphix. All rights reserved.
  26  * Copyright 2017 Joyent, Inc.
  27  * Copyright (c) 2017 Datto Inc.
  28  */
  29
  30 #include <sys/dsl_scan.h>
  31 #include <sys/dsl_pool.h>
  32 #include <sys/dsl_dataset.h>
  33 #include <sys/dsl_prop.h>
  34 #include <sys/dsl_dir.h>
  35 #include <sys/dsl_synctask.h>
  36 #include <sys/dnode.h>
  37 #include <sys/dmu_tx.h>
  38 #include <sys/dmu_objset.h>
  39 #include <sys/arc.h>
  40 #include <sys/zap.h>
  41 #include <sys/zio.h>
  42 #include <sys/zfs_context.h>
  43 #include <sys/fs/zfs.h>
  44 #include <sys/zfs_znode.h>
  45 #include <sys/spa_impl.h>
  46 #include <sys/vdev_impl.h>
  47 #include <sys/zil_impl.h>
  48 #include <sys/zio_checksum.h>
  49 #include <sys/ddt.h>
  50 #include <sys/sa.h>
  51 #include <sys/sa_impl.h>
  52 #include <sys/zfeature.h>
  53 #include <sys/abd.h>
  54 #include <sys/range_tree.h>
  55 #ifdef _KERNEL
  56 #include <sys/zfs_vfsops.h>
  57 #endif
  58
  59 /*
  60  * Grand theory statement on scan queue sorting
  61  *
  62  * Scanning is implemented by recursively traversing all indirection levels
  63  * in an object and reading all blocks referenced from said objects. This
  64  * results in us approximately traversing the object from lowest logical
  65  * offset to the highest. For best performance, we would want the logical
  66  * blocks to be physically contiguous. However, this is frequently not the
  67  * case with pools given the allocation patterns of copy-on-write filesystems.
  68  * So instead, we put the I/Os into a reordering queue and issue them in a
  69  * way that will most benefit physical disks (LBA-order).
  70  *
  71  * Queue management:
  72  *
  73  * Ideally, we would want to scan all metadata and queue up all block I/O
  74  * prior to starting to issue it, because that allows us to do an optimal
  75  * sorting job. This can however consume large amounts of memory. Therefore
  76  * we continuously monitor the size of the queues and constrain them to 5%
  77  * (zfs_scan_mem_lim_fact) of physmem. If the queues grow larger than this
  78  * limit, we clear out a few of the largest extents at the head of the queues
  79  * to make room for more scanning. Hopefully, these extents will be fairly
  80  * large and contiguous, allowing us to approach sequential I/O throughput
  81  * even without a fully sorted tree.
  82  *
  83  * Metadata scanning takes place in dsl_scan_visit(), which is called from
  84  * dsl_scan_sync() every spa_sync(). If we have either fully scanned all
  85  * metadata on the pool, or we need to make room in memory because our
  86  * queues are too large, dsl_scan_visit() is postponed and
  87  * scan_io_queues_run() is called from dsl_scan_sync() instead. This implies
  88  * that metadata scanning and queued I/O issuing are mutually exclusive. This
  89  * allows us to provide maximum sequential I/O throughput for the majority of
  90  * I/O's issued since sequential I/O performance is significantly negatively
  91  * impacted if it is interleaved with random I/O.
  92  *
  93  * Implementation Notes
  94  *
  95  * One side effect of the queued scanning algorithm is that the scanning code
  96  * needs to be notified whenever a block is freed. This is needed to allow
  97  * the scanning code to remove these I/Os from the issuing queue. Additionally,
  98  * we do not attempt to queue gang blocks to be issued sequentially since this
  99  * is very hard to do and would have an extremely limitted performance benefit.
 100  * Instead, we simply issue gang I/Os as soon as we find them using the legacy
 101  * algorithm.
 102  *
 103  * Backwards compatibility
 104  *
 105  * This new algorithm is backwards compatible with the legacy on-disk data
 106  * structures (and therefore does not require a new feature flag).
 107  * Periodically during scanning (see zfs_scan_checkpoint_intval), the scan
 108  * will stop scanning metadata (in logical order) and wait for all outstanding
 109  * sorted I/O to complete. Once this is done, we write out a checkpoint
 110  * bookmark, indicating that we have scanned everything logically before it.
 111  * If the pool is imported on a machine without the new sorting algorithm,
 112  * the scan simply resumes from the last checkpoint using the legacy algorithm.
 113  */
 114
 115 typedef int (scan_cb_t)(dsl_pool_t *, const blkptr_t *,
 116     const zbookmark_phys_t *);
 117
 118 static scan_cb_t dsl_scan_scrub_cb;
 119
 120 static int scan_ds_queue_compare(const void *a, const void *b);
 121 static int scan_prefetch_queue_compare(const void *a, const void *b);
 122 static void scan_ds_queue_clear(dsl_scan_t *scn);
 123 static boolean_t scan_ds_queue_contains(dsl_scan_t *scn, uint64_t dsobj,
 124     uint64_t *txg);
 125 static void scan_ds_queue_insert(dsl_scan_t *scn, uint64_t dsobj, uint64_t txg);
 126 static void scan_ds_queue_remove(dsl_scan_t *scn, uint64_t dsobj);
 127 static void scan_ds_queue_sync(dsl_scan_t *scn, dmu_tx_t *tx);
 128
 129 extern int zfs_vdev_async_write_active_min_dirty_percent;
 130
 131 /*
 132  * By default zfs will check to ensure it is not over the hard memory
 133  * limit before each txg. If finer-grained control of this is needed
 134  * this value can be set to 1 to enable checking before scanning each
 135  * block.
 136  */
 137 int zfs_scan_strict_mem_lim = B_FALSE;
 138
 139 /*
 140  * Maximum number of parallelly executing I/Os per top-level vdev.
 141  * Tune with care. Very high settings (hundreds) are known to trigger
 142  * some firmware bugs and resets on certain SSDs.
 143  */
 144 int zfs_top_maxinflight = 32;           /* maximum I/Os per top-level */
 145 unsigned int zfs_resilver_delay = 2;    /* number of ticks to delay resilver -- 2 is a good number */
 146 unsigned int zfs_scrub_delay = 4;       /* number of ticks to delay scrub -- 4 is a good number */
 147 unsigned int zfs_scan_idle = 50;        /* idle window in clock ticks */
 148
 149 /*
 150  * Maximum number of parallelly executed bytes per leaf vdev. We attempt
 151  * to strike a balance here between keeping the vdev queues full of I/Os
 152  * at all times and not overflowing the queues to cause long latency,
 153  * which would cause long txg sync times. No matter what, we will not
 154  * overload the drives with I/O, since that is protected by
 155  * zfs_vdev_scrub_max_active.
 156  */
 157 unsigned long zfs_scan_vdev_limit = 4 << 20;
 158
 159 int zfs_scan_issue_strategy = 0;
 160 int zfs_scan_legacy = B_FALSE;  /* don't queue & sort zios, go direct */
 161 uint64_t zfs_scan_max_ext_gap = 2 << 20;        /* in bytes */
 162
 163 unsigned int zfs_scan_checkpoint_intval = 7200; /* seconds */
 164 #define ZFS_SCAN_CHECKPOINT_INTVAL      SEC_TO_TICK(zfs_scan_checkpoint_intval)
 165
 166 /*
 167  * fill_weight is non-tunable at runtime, so we copy it at module init from
 168  * zfs_scan_fill_weight. Runtime adjustments to zfs_scan_fill_weight would
 169  * break queue sorting.
 170  */
 171 uint64_t zfs_scan_fill_weight = 3;
 172 static uint64_t fill_weight;
 173
 174 /* See dsl_scan_should_clear() for details on the memory limit tunables */
 175 uint64_t zfs_scan_mem_lim_min = 16 << 20;       /* bytes */
 176 uint64_t zfs_scan_mem_lim_soft_max = 128 << 20; /* bytes */
 177 int zfs_scan_mem_lim_fact = 20;         /* fraction of physmem */
 178 int zfs_scan_mem_lim_soft_fact = 20;    /* fraction of mem lim above */
 179
 180 unsigned int zfs_scrub_min_time_ms = 1000; /* min millisecs to scrub per txg */
 181 unsigned int zfs_free_min_time_ms = 1000; /* min millisecs to free per txg */
 182 unsigned int zfs_obsolete_min_time_ms = 500; /* min millisecs to obsolete per txg */
 183 unsigned int zfs_resilver_min_time_ms = 3000; /* min millisecs to resilver per txg */
 184 boolean_t zfs_no_scrub_io = B_FALSE; /* set to disable scrub i/o */
 185 boolean_t zfs_no_scrub_prefetch = B_FALSE; /* set to disable scrub prefetch */
 186
 187 SYSCTL_DECL(_vfs_zfs);
 188 SYSCTL_UINT(_vfs_zfs, OID_AUTO, top_maxinflight, CTLFLAG_RWTUN,
 189     &zfs_top_maxinflight, 0, "Maximum I/Os per top-level vdev");
 190 SYSCTL_UINT(_vfs_zfs, OID_AUTO, resilver_delay, CTLFLAG_RWTUN,
 191     &zfs_resilver_delay, 0, "Number of ticks to delay resilver");
 192 SYSCTL_UINT(_vfs_zfs, OID_AUTO, scrub_delay, CTLFLAG_RWTUN,
 193     &zfs_scrub_delay, 0, "Number of ticks to delay scrub");
 194 SYSCTL_UINT(_vfs_zfs, OID_AUTO, scan_idle, CTLFLAG_RWTUN,
 195     &zfs_scan_idle, 0, "Idle scan window in clock ticks");
 196 SYSCTL_UINT(_vfs_zfs, OID_AUTO, scan_min_time_ms, CTLFLAG_RWTUN,
 197     &zfs_scrub_min_time_ms, 0, "Min millisecs to scrub per txg");
 198 SYSCTL_UINT(_vfs_zfs, OID_AUTO, free_min_time_ms, CTLFLAG_RWTUN,
 199     &zfs_free_min_time_ms, 0, "Min millisecs to free per txg");
 200 SYSCTL_UINT(_vfs_zfs, OID_AUTO, resilver_min_time_ms, CTLFLAG_RWTUN,
 201     &zfs_resilver_min_time_ms, 0, "Min millisecs to resilver per txg");
 202 SYSCTL_INT(_vfs_zfs, OID_AUTO, no_scrub_io, CTLFLAG_RWTUN,
 203     &zfs_no_scrub_io, 0, "Disable scrub I/O");
 204 SYSCTL_INT(_vfs_zfs, OID_AUTO, no_scrub_prefetch, CTLFLAG_RWTUN,
 205     &zfs_no_scrub_prefetch, 0, "Disable scrub prefetching");
 206 SYSCTL_UINT(_vfs_zfs, OID_AUTO, zfs_scan_legacy, CTLFLAG_RWTUN,
 207     &zfs_scan_legacy, 0, "Scrub using legacy non-sequential method");
 208 SYSCTL_UINT(_vfs_zfs, OID_AUTO, zfs_scan_checkpoint_interval, CTLFLAG_RWTUN,
 209     &zfs_scan_checkpoint_intval, 0, "Scan progress on-disk checkpointing interval");
 210
 211 enum ddt_class zfs_scrub_ddt_class_max = DDT_CLASS_DUPLICATE;
 212 /* max number of blocks to free in a single TXG */
 213 uint64_t zfs_async_block_max_blocks = UINT64_MAX;
 214 SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, free_max_blocks, CTLFLAG_RWTUN,
 215     &zfs_async_block_max_blocks, 0, "Maximum number of blocks to free in one TXG");
 216
 217 /*
 218  * We wait a few txgs after importing a pool to begin scanning so that
 219  * the import / mounting code isn't held up by scrub / resilver IO.
 220  * Unfortunately, it is a bit difficult to determine exactly how long
 221  * this will take since userspace will trigger fs mounts asynchronously
 222  * and the kernel will create zvol minors asynchronously. As a result,
 223  * the value provided here is a bit arbitrary, but represents a
 224  * reasonable estimate of how many txgs it will take to finish fully
 225  * importing a pool
 226  */
 227 #define        SCAN_IMPORT_WAIT_TXGS           5
 228
 229
 230 #define DSL_SCAN_IS_SCRUB_RESILVER(scn) \
 231         ((scn)->scn_phys.scn_func == POOL_SCAN_SCRUB || \
 232         (scn)->scn_phys.scn_func == POOL_SCAN_RESILVER)
 233
 234 extern int zfs_txg_timeout;
 235
 236 /*
 237  * Enable/disable the processing of the free_bpobj object.
 238  */
 239 boolean_t zfs_free_bpobj_enabled = B_TRUE;
 240
 241 SYSCTL_INT(_vfs_zfs, OID_AUTO, free_bpobj_enabled, CTLFLAG_RWTUN,
 242     &zfs_free_bpobj_enabled, 0, "Enable free_bpobj processing");
 243
 244 /* the order has to match pool_scan_type */
 245 static scan_cb_t *scan_funcs[POOL_SCAN_FUNCS] = {
 246         NULL,
 247         dsl_scan_scrub_cb,      /* POOL_SCAN_SCRUB */
 248         dsl_scan_scrub_cb,      /* POOL_SCAN_RESILVER */
 249 };
 250
 251 /* In core node for the scn->scn_queue. Represents a dataset to be scanned */
 252 typedef struct {
 253         uint64_t        sds_dsobj;
 254         uint64_t        sds_txg;
 255         avl_node_t      sds_node;
 256 } scan_ds_t;
 257
 258 /*
 259  * This controls what conditions are placed on dsl_scan_sync_state():
 260  * SYNC_OPTIONAL) write out scn_phys iff scn_bytes_pending == 0
 261  * SYNC_MANDATORY) write out scn_phys always. scn_bytes_pending must be 0.
 262  * SYNC_CACHED) if scn_bytes_pending == 0, write out scn_phys. Otherwise
 263  *      write out the scn_phys_cached version.
 264  * See dsl_scan_sync_state for details.
 265  */
 266 typedef enum {
 267         SYNC_OPTIONAL,
 268         SYNC_MANDATORY,
 269         SYNC_CACHED
 270 } state_sync_type_t;
 271
 272 /*
 273  * This struct represents the minimum information needed to reconstruct a
 274  * zio for sequential scanning. This is useful because many of these will
 275  * accumulate in the sequential IO queues before being issued, so saving
 276  * memory matters here.
 277  */
 278 typedef struct scan_io {
 279         /* fields from blkptr_t */
 280         uint64_t                sio_offset;
 281         uint64_t                sio_blk_prop;
 282         uint64_t                sio_phys_birth;
 283         uint64_t                sio_birth;
 284         zio_cksum_t             sio_cksum;
 285         uint32_t                sio_asize;
 286
 287         /* fields from zio_t */
 288         int                     sio_flags;
 289         zbookmark_phys_t        sio_zb;
 290
 291         /* members for queue sorting */
 292         union {
 293                 avl_node_t      sio_addr_node; /* link into issueing queue */
 294                 list_node_t     sio_list_node; /* link for issuing to disk */
 295         } sio_nodes;
 296 } scan_io_t;
 297
 298 struct dsl_scan_io_queue {
 299         dsl_scan_t      *q_scn; /* associated dsl_scan_t */
 300         vdev_t          *q_vd; /* top-level vdev that this queue represents */
 301
 302         /* trees used for sorting I/Os and extents of I/Os */
 303         range_tree_t    *q_exts_by_addr;
 304         avl_tree_t      q_exts_by_size;
 305         avl_tree_t      q_sios_by_addr;
 306
 307         /* members for zio rate limiting */
 308         uint64_t        q_maxinflight_bytes;
 309         uint64_t        q_inflight_bytes;
 310         kcondvar_t      q_zio_cv; /* used under vd->vdev_scan_io_queue_lock */
 311
 312         /* per txg statistics */
 313         uint64_t        q_total_seg_size_this_txg;
 314         uint64_t        q_segs_this_txg;
 315         uint64_t        q_total_zio_size_this_txg;
 316         uint64_t        q_zios_this_txg;
 317 };
 318
 319 /* private data for dsl_scan_prefetch_cb() */
 320 typedef struct scan_prefetch_ctx {
 321         refcount_t spc_refcnt;          /* refcount for memory management */
 322         dsl_scan_t *spc_scn;            /* dsl_scan_t for the pool */
 323         boolean_t spc_root;             /* is this prefetch for an objset? */
 324         uint8_t spc_indblkshift;        /* dn_indblkshift of current dnode */
 325         uint16_t spc_datablkszsec;      /* dn_idatablkszsec of current dnode */
 326 } scan_prefetch_ctx_t;
 327
 328 /* private data for dsl_scan_prefetch() */
 329 typedef struct scan_prefetch_issue_ctx {
 330         avl_node_t spic_avl_node;       /* link into scn->scn_prefetch_queue */
 331         scan_prefetch_ctx_t *spic_spc;  /* spc for the callback */
 332         blkptr_t spic_bp;               /* bp to prefetch */
 333         zbookmark_phys_t spic_zb;       /* bookmark to prefetch */
 334 } scan_prefetch_issue_ctx_t;
 335
 336 static void scan_exec_io(dsl_pool_t *dp, const blkptr_t *bp, int zio_flags,
 337     const zbookmark_phys_t *zb, dsl_scan_io_queue_t *queue);
 338 static void scan_io_queue_insert_impl(dsl_scan_io_queue_t *queue,
 339     scan_io_t *sio);
 340
 341 static dsl_scan_io_queue_t *scan_io_queue_create(vdev_t *vd);
 342 static void scan_io_queues_destroy(dsl_scan_t *scn);
 343
 344 static kmem_cache_t *sio_cache;
 345
 346 void
 347 scan_init(void)
 348 {
 349         /*
 350          * This is used in ext_size_compare() to weight segments
 351          * based on how sparse they are. This cannot be changed
 352          * mid-scan and the tree comparison functions don't currently
 353          * have a mechansim for passing additional context to the
 354          * compare functions. Thus we store this value globally and
 355          * we only allow it to be set at module intiailization time
 356          */
 357         fill_weight = zfs_scan_fill_weight;
 358
 359         sio_cache = kmem_cache_create("sio_cache",
 360             sizeof (scan_io_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
 361 }
 362
 363 void
 364 scan_fini(void)
 365 {
 366         kmem_cache_destroy(sio_cache);
 367 }
 368
 369 static inline boolean_t
 370 dsl_scan_is_running(const dsl_scan_t *scn)
 371 {
 372         return (scn->scn_phys.scn_state == DSS_SCANNING);
 373 }
 374
 375 boolean_t
 376 dsl_scan_resilvering(dsl_pool_t *dp)
 377 {
 378         return (dsl_scan_is_running(dp->dp_scan) &&
 379             dp->dp_scan->scn_phys.scn_func == POOL_SCAN_RESILVER);
 380 }
 381
 382 static inline void
 383 sio2bp(const scan_io_t *sio, blkptr_t *bp, uint64_t vdev_id)
 384 {
 385         bzero(bp, sizeof (*bp));
 386         DVA_SET_ASIZE(&bp->blk_dva[0], sio->sio_asize);
 387         DVA_SET_VDEV(&bp->blk_dva[0], vdev_id);
 388         DVA_SET_OFFSET(&bp->blk_dva[0], sio->sio_offset);
 389         bp->blk_prop = sio->sio_blk_prop;
 390         bp->blk_phys_birth = sio->sio_phys_birth;
 391         bp->blk_birth = sio->sio_birth;
 392         bp->blk_fill = 1;       /* we always only work with data pointers */
 393         bp->blk_cksum = sio->sio_cksum;
 394 }
 395
 396 static inline void
 397 bp2sio(const blkptr_t *bp, scan_io_t *sio, int dva_i)
 398 {
 399         /* we discard the vdev id, since we can deduce it from the queue */
 400         sio->sio_offset = DVA_GET_OFFSET(&bp->blk_dva[dva_i]);
 401         sio->sio_asize = DVA_GET_ASIZE(&bp->blk_dva[dva_i]);
 402         sio->sio_blk_prop = bp->blk_prop;
 403         sio->sio_phys_birth = bp->blk_phys_birth;
 404         sio->sio_birth = bp->blk_birth;
 405         sio->sio_cksum = bp->blk_cksum;
 406 }
 407
 408 void
 409 dsl_scan_global_init(void)
 410 {
 411         /*
 412          * This is used in ext_size_compare() to weight segments
 413          * based on how sparse they are. This cannot be changed
 414          * mid-scan and the tree comparison functions don't currently
 415          * have a mechansim for passing additional context to the
 416          * compare functions. Thus we store this value globally and
 417          * we only allow it to be set at module intiailization time
 418          */
 419         fill_weight = zfs_scan_fill_weight;
 420 }
 421
 422 int
 423 dsl_scan_init(dsl_pool_t *dp, uint64_t txg)
 424 {
 425         int err;
 426         dsl_scan_t *scn;
 427         spa_t *spa = dp->dp_spa;
 428         uint64_t f;
 429
 430         scn = dp->dp_scan = kmem_zalloc(sizeof (dsl_scan_t), KM_SLEEP);
 431         scn->scn_dp = dp;
 432
 433         /*
 434          * It's possible that we're resuming a scan after a reboot so
 435          * make sure that the scan_async_destroying flag is initialized
 436          * appropriately.
 437          */
 438         ASSERT(!scn->scn_async_destroying);
 439         scn->scn_async_destroying = spa_feature_is_active(dp->dp_spa,
 440             SPA_FEATURE_ASYNC_DESTROY);
 441
 442         bcopy(&scn->scn_phys, &scn->scn_phys_cached, sizeof (scn->scn_phys));
 443         avl_create(&scn->scn_queue, scan_ds_queue_compare, sizeof (scan_ds_t),
 444             offsetof(scan_ds_t, sds_node));
 445         avl_create(&scn->scn_prefetch_queue, scan_prefetch_queue_compare,
 446             sizeof (scan_prefetch_issue_ctx_t),
 447             offsetof(scan_prefetch_issue_ctx_t, spic_avl_node));
 448
 449         err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 450             "scrub_func", sizeof (uint64_t), 1, &f);
 451         if (err == 0) {
 452                 /*
 453                  * There was an old-style scrub in progress.  Restart a
 454                  * new-style scrub from the beginning.
 455                  */
 456                 scn->scn_restart_txg = txg;
 457                 zfs_dbgmsg("old-style scrub was in progress; "
 458                     "restarting new-style scrub in txg %llu",
 459                     (longlong_t)scn->scn_restart_txg);
 460
 461                 /*
 462                  * Load the queue obj from the old location so that it
 463                  * can be freed by dsl_scan_done().
 464                  */
 465                 (void) zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 466                     "scrub_queue", sizeof (uint64_t), 1,
 467                     &scn->scn_phys.scn_queue_obj);
 468         } else {
 469                 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 470                     DMU_POOL_SCAN, sizeof (uint64_t), SCAN_PHYS_NUMINTS,
 471                     &scn->scn_phys);
 472                 if (err == ENOENT)
 473                         return (0);
 474                 else if (err)
 475                         return (err);
 476
 477                 /*
 478                  * We might be restarting after a reboot, so jump the issued
 479                  * counter to how far we've scanned. We know we're consistent
 480                  * up to here.
 481                  */
 482                 scn->scn_issued_before_pass = scn->scn_phys.scn_examined;
 483
 484                 if (dsl_scan_is_running(scn) &&
 485                     spa_prev_software_version(dp->dp_spa) < SPA_VERSION_SCAN) {
 486                         /*
 487                          * A new-type scrub was in progress on an old
 488                          * pool, and the pool was accessed by old
 489                          * software.  Restart from the beginning, since
 490                          * the old software may have changed the pool in
 491                          * the meantime.
 492                          */
 493                         scn->scn_restart_txg = txg;
 494                         zfs_dbgmsg("new-style scrub was modified "
 495                             "by old software; restarting in txg %llu",
 496                             (longlong_t)scn->scn_restart_txg);
 497                 }
 498         }
 499
 500         /* reload the queue into the in-core state */
 501         if (scn->scn_phys.scn_queue_obj != 0) {
 502                 zap_cursor_t zc;
 503                 zap_attribute_t za;
 504
 505                 for (zap_cursor_init(&zc, dp->dp_meta_objset,
 506                     scn->scn_phys.scn_queue_obj);
 507                     zap_cursor_retrieve(&zc, &za) == 0;
 508                     (void) zap_cursor_advance(&zc)) {
 509                         scan_ds_queue_insert(scn,
 510                             zfs_strtonum(za.za_name, NULL),
 511                             za.za_first_integer);
 512                 }
 513                 zap_cursor_fini(&zc);
 514         }
 515
 516         spa_scan_stat_init(spa);
 517         return (0);
 518 }
 519
 520 void
 521 dsl_scan_fini(dsl_pool_t *dp)
 522 {
 523         if (dp->dp_scan != NULL) {
 524                 dsl_scan_t *scn = dp->dp_scan;
 525
 526                 if (scn->scn_taskq != NULL)
 527                         taskq_destroy(scn->scn_taskq);
 528                 scan_ds_queue_clear(scn);
 529                 avl_destroy(&scn->scn_queue);
 530                 avl_destroy(&scn->scn_prefetch_queue);
 531
 532                 kmem_free(dp->dp_scan, sizeof (dsl_scan_t));
 533                 dp->dp_scan = NULL;
 534         }
 535 }
 536
 537 static boolean_t
 538 dsl_scan_restarting(dsl_scan_t *scn, dmu_tx_t *tx)
 539 {
 540         return (scn->scn_restart_txg != 0 &&
 541             scn->scn_restart_txg <= tx->tx_txg);
 542 }
 543
 544 boolean_t
 545 dsl_scan_scrubbing(const dsl_pool_t *dp)
 546 {
 547         dsl_scan_phys_t *scn_phys = &dp->dp_scan->scn_phys;
 548
 549         return (scn_phys->scn_state == DSS_SCANNING &&
 550             scn_phys->scn_func == POOL_SCAN_SCRUB);
 551 }
 552
 553 boolean_t
 554 dsl_scan_is_paused_scrub(const dsl_scan_t *scn)
 555 {
 556         return (dsl_scan_scrubbing(scn->scn_dp) &&
 557             scn->scn_phys.scn_flags & DSF_SCRUB_PAUSED);
 558 }
 559
 560 /*
 561  * Writes out a persistent dsl_scan_phys_t record to the pool directory.
 562  * Because we can be running in the block sorting algorithm, we do not always
 563  * want to write out the record, only when it is "safe" to do so. This safety
 564  * condition is achieved by making sure that the sorting queues are empty
 565  * (scn_bytes_pending == 0). When this condition is not true, the sync'd state
 566  * is inconsistent with how much actual scanning progress has been made. The
 567  * kind of sync to be performed is specified by the sync_type argument. If the
 568  * sync is optional, we only sync if the queues are empty. If the sync is
 569  * mandatory, we do a hard ASSERT to make sure that the queues are empty. The
 570  * third possible state is a "cached" sync. This is done in response to:
 571  * 1) The dataset that was in the last sync'd dsl_scan_phys_t having been
 572  *      destroyed, so we wouldn't be able to restart scanning from it.
 573  * 2) The snapshot that was in the last sync'd dsl_scan_phys_t having been
 574  *      superseded by a newer snapshot.
 575  * 3) The dataset that was in the last sync'd dsl_scan_phys_t having been
 576  *      swapped with its clone.
 577  * In all cases, a cached sync simply rewrites the last record we've written,
 578  * just slightly modified. For the modifications that are performed to the
 579  * last written dsl_scan_phys_t, see dsl_scan_ds_destroyed,
 580  * dsl_scan_ds_snapshotted and dsl_scan_ds_clone_swapped.
 581  */
 582 static void
 583 dsl_scan_sync_state(dsl_scan_t *scn, dmu_tx_t *tx, state_sync_type_t sync_type)
 584 {
 585         int i;
 586         spa_t *spa = scn->scn_dp->dp_spa;
 587
 588         ASSERT(sync_type != SYNC_MANDATORY || scn->scn_bytes_pending == 0);
 589         if (scn->scn_bytes_pending == 0) {
 590                 for (i = 0; i < spa->spa_root_vdev->vdev_children; i++) {
 591                         vdev_t *vd = spa->spa_root_vdev->vdev_child[i];
 592                         dsl_scan_io_queue_t *q = vd->vdev_scan_io_queue;
 593
 594                         if (q == NULL)
 595                                 continue;
 596
 597                         mutex_enter(&vd->vdev_scan_io_queue_lock);
 598                         ASSERT3P(avl_first(&q->q_sios_by_addr), ==, NULL);
 599                         ASSERT3P(avl_first(&q->q_exts_by_size), ==, NULL);
 600                         ASSERT3P(range_tree_first(q->q_exts_by_addr), ==, NULL);
 601                         mutex_exit(&vd->vdev_scan_io_queue_lock);
 602                 }
 603
 604                 if (scn->scn_phys.scn_queue_obj != 0)
 605                         scan_ds_queue_sync(scn, tx);
 606                 VERIFY0(zap_update(scn->scn_dp->dp_meta_objset,
 607                     DMU_POOL_DIRECTORY_OBJECT,
 608                     DMU_POOL_SCAN, sizeof (uint64_t), SCAN_PHYS_NUMINTS,
 609                     &scn->scn_phys, tx));
 610                 bcopy(&scn->scn_phys, &scn->scn_phys_cached,
 611                     sizeof (scn->scn_phys));
 612
 613                 if (scn->scn_checkpointing)
 614                         zfs_dbgmsg("finish scan checkpoint");
 615
 616                 scn->scn_checkpointing = B_FALSE;
 617                 scn->scn_last_checkpoint = ddi_get_lbolt();
 618         } else if (sync_type == SYNC_CACHED) {
 619                 VERIFY0(zap_update(scn->scn_dp->dp_meta_objset,
 620                     DMU_POOL_DIRECTORY_OBJECT,
 621                     DMU_POOL_SCAN, sizeof (uint64_t), SCAN_PHYS_NUMINTS,
 622                     &scn->scn_phys_cached, tx));
 623         }
 624 }
 625
 626 /* ARGSUSED */
 627 static int
 628 dsl_scan_setup_check(void *arg, dmu_tx_t *tx)
 629 {
 630         dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
 631
 632         if (dsl_scan_is_running(scn))
 633                 return (SET_ERROR(EBUSY));
 634
 635         return (0);
 636 }
 637
 638 static void
 639 dsl_scan_setup_sync(void *arg, dmu_tx_t *tx)
 640 {
 641         dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
 642         pool_scan_func_t *funcp = arg;
 643         dmu_object_type_t ot = 0;
 644         dsl_pool_t *dp = scn->scn_dp;
 645         spa_t *spa = dp->dp_spa;
 646
 647         ASSERT(!dsl_scan_is_running(scn));
 648         ASSERT(*funcp > POOL_SCAN_NONE && *funcp < POOL_SCAN_FUNCS);
 649         bzero(&scn->scn_phys, sizeof (scn->scn_phys));
 650         scn->scn_phys.scn_func = *funcp;
 651         scn->scn_phys.scn_state = DSS_SCANNING;
 652         scn->scn_phys.scn_min_txg = 0;
 653         scn->scn_phys.scn_max_txg = tx->tx_txg;
 654         scn->scn_phys.scn_ddt_class_max = DDT_CLASSES - 1; /* the entire DDT */
 655         scn->scn_phys.scn_start_time = gethrestime_sec();
 656         scn->scn_phys.scn_errors = 0;
 657         scn->scn_phys.scn_to_examine = spa->spa_root_vdev->vdev_stat.vs_alloc;
 658         scn->scn_issued_before_pass = 0;
 659         scn->scn_restart_txg = 0;
 660         scn->scn_done_txg = 0;
 661         scn->scn_last_checkpoint = 0;
 662         scn->scn_checkpointing = B_FALSE;
 663         spa_scan_stat_init(spa);
 664
 665         if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
 666                 scn->scn_phys.scn_ddt_class_max = zfs_scrub_ddt_class_max;
 667
 668                 /* rewrite all disk labels */
 669                 vdev_config_dirty(spa->spa_root_vdev);
 670
 671                 if (vdev_resilver_needed(spa->spa_root_vdev,
 672                     &scn->scn_phys.scn_min_txg, &scn->scn_phys.scn_max_txg)) {
 673                         spa_event_notify(spa, NULL, NULL,
 674                             ESC_ZFS_RESILVER_START);
 675                 } else {
 676                         spa_event_notify(spa, NULL, NULL, ESC_ZFS_SCRUB_START);
 677                 }
 678
 679                 spa->spa_scrub_started = B_TRUE;
 680                 /*
 681                  * If this is an incremental scrub, limit the DDT scrub phase
 682                  * to just the auto-ditto class (for correctness); the rest
 683                  * of the scrub should go faster using top-down pruning.
 684                  */
 685                 if (scn->scn_phys.scn_min_txg > TXG_INITIAL)
 686                         scn->scn_phys.scn_ddt_class_max = DDT_CLASS_DITTO;
 687
 688         }
 689
 690         /* back to the generic stuff */
 691
 692         if (dp->dp_blkstats == NULL) {
 693                 dp->dp_blkstats =
 694                     kmem_alloc(sizeof (zfs_all_blkstats_t), KM_SLEEP);
 695                 mutex_init(&dp->dp_blkstats->zab_lock, NULL,
 696                     MUTEX_DEFAULT, NULL);
 697         }
 698         bzero(&dp->dp_blkstats->zab_type, sizeof (dp->dp_blkstats->zab_type));
 699
 700         if (spa_version(spa) < SPA_VERSION_DSL_SCRUB)
 701                 ot = DMU_OT_ZAP_OTHER;
 702
 703         scn->scn_phys.scn_queue_obj = zap_create(dp->dp_meta_objset,
 704             ot ? ot : DMU_OT_SCAN_QUEUE, DMU_OT_NONE, 0, tx);
 705
 706         bcopy(&scn->scn_phys, &scn->scn_phys_cached, sizeof (scn->scn_phys));
 707
 708         dsl_scan_sync_state(scn, tx, SYNC_MANDATORY);
 709
 710         spa_history_log_internal(spa, "scan setup", tx,
 711             "func=%u mintxg=%llu maxtxg=%llu",
 712             *funcp, scn->scn_phys.scn_min_txg, scn->scn_phys.scn_max_txg);
 713 }
 714
 715 /*
 716  * Called by the ZFS_IOC_POOL_SCAN ioctl to start a scrub or resilver.
 717  * Can also be called to resume a paused scrub.
 718  */
 719 int
 720 dsl_scan(dsl_pool_t *dp, pool_scan_func_t func)
 721 {
 722         spa_t *spa = dp->dp_spa;
 723         dsl_scan_t *scn = dp->dp_scan;
 724
 725         /*
 726          * Purge all vdev caches and probe all devices.  We do this here
 727          * rather than in sync context because this requires a writer lock
 728          * on the spa_config lock, which we can't do from sync context.  The
 729          * spa_scrub_reopen flag indicates that vdev_open() should not
 730          * attempt to start another scrub.
 731          */
 732         spa_vdev_state_enter(spa, SCL_NONE);
 733         spa->spa_scrub_reopen = B_TRUE;
 734         vdev_reopen(spa->spa_root_vdev);
 735         spa->spa_scrub_reopen = B_FALSE;
 736         (void) spa_vdev_state_exit(spa, NULL, 0);
 737
 738         if (func == POOL_SCAN_SCRUB && dsl_scan_is_paused_scrub(scn)) {
 739                 /* got scrub start cmd, resume paused scrub */
 740                 int err = dsl_scrub_set_pause_resume(scn->scn_dp,
 741                     POOL_SCRUB_NORMAL);
 742                 if (err == 0) {
 743                         spa_event_notify(spa, NULL, NULL, ESC_ZFS_SCRUB_RESUME);
 744                         return (ECANCELED);
 745                 }
 746                 return (SET_ERROR(err));
 747         }
 748
 749         return (dsl_sync_task(spa_name(spa), dsl_scan_setup_check,
 750             dsl_scan_setup_sync, &func, 0, ZFS_SPACE_CHECK_EXTRA_RESERVED));
 751 }
 752
 753 /* ARGSUSED */
 754 static void
 755 dsl_scan_done(dsl_scan_t *scn, boolean_t complete, dmu_tx_t *tx)
 756 {
 757         static const char *old_names[] = {
 758                 "scrub_bookmark",
 759                 "scrub_ddt_bookmark",
 760                 "scrub_ddt_class_max",
 761                 "scrub_queue",
 762                 "scrub_min_txg",
 763                 "scrub_max_txg",
 764                 "scrub_func",
 765                 "scrub_errors",
 766                 NULL
 767         };
 768
 769         dsl_pool_t *dp = scn->scn_dp;
 770         spa_t *spa = dp->dp_spa;
 771         int i;
 772
 773         /* Remove any remnants of an old-style scrub. */
 774         for (i = 0; old_names[i]; i++) {
 775                 (void) zap_remove(dp->dp_meta_objset,
 776                     DMU_POOL_DIRECTORY_OBJECT, old_names[i], tx);
 777         }
 778
 779         if (scn->scn_phys.scn_queue_obj != 0) {
 780                 VERIFY0(dmu_object_free(dp->dp_meta_objset,
 781                     scn->scn_phys.scn_queue_obj, tx));
 782                 scn->scn_phys.scn_queue_obj = 0;
 783         }
 784         scan_ds_queue_clear(scn);
 785
 786         scn->scn_phys.scn_flags &= ~DSF_SCRUB_PAUSED;
 787
 788         /*
 789          * If we were "restarted" from a stopped state, don't bother
 790          * with anything else.
 791          */
 792         if (!dsl_scan_is_running(scn)) {
 793                 ASSERT(!scn->scn_is_sorted);
 794                 return;
 795         }
 796
 797         if (scn->scn_is_sorted) {
 798                 scan_io_queues_destroy(scn);
 799                 scn->scn_is_sorted = B_FALSE;
 800
 801                 if (scn->scn_taskq != NULL) {
 802                         taskq_destroy(scn->scn_taskq);
 803                         scn->scn_taskq = NULL;
 804                 }
 805         }
 806
 807         scn->scn_phys.scn_state = complete ? DSS_FINISHED : DSS_CANCELED;
 808
 809         if (dsl_scan_restarting(scn, tx))
 810                 spa_history_log_internal(spa, "scan aborted, restarting", tx,
 811                     "errors=%llu", spa_get_errlog_size(spa));
 812         else if (!complete)
 813                 spa_history_log_internal(spa, "scan cancelled", tx,
 814                     "errors=%llu", spa_get_errlog_size(spa));
 815         else
 816                 spa_history_log_internal(spa, "scan done", tx,
 817                     "errors=%llu", spa_get_errlog_size(spa));
 818
 819         if (DSL_SCAN_IS_SCRUB_RESILVER(scn)) {
 820                 spa->spa_scrub_started = B_FALSE;
 821                 spa->spa_scrub_active = B_FALSE;
 822
 823                 /*
 824                  * If the scrub/resilver completed, update all DTLs to
 825                  * reflect this.  Whether it succeeded or not, vacate
 826                  * all temporary scrub DTLs.
 827                  *
 828                  * As the scrub does not currently support traversing
 829                  * data that have been freed but are part of a checkpoint,
 830                  * we don't mark the scrub as done in the DTLs as faults
 831                  * may still exist in those vdevs.
 832                  */
 833                 if (complete &&
 834                     !spa_feature_is_active(spa, SPA_FEATURE_POOL_CHECKPOINT)) {
 835                         vdev_dtl_reassess(spa->spa_root_vdev, tx->tx_txg,
 836                             scn->scn_phys.scn_max_txg, B_TRUE);
 837
 838                         spa_event_notify(spa, NULL, NULL,
 839                             scn->scn_phys.scn_min_txg ?
 840                             ESC_ZFS_RESILVER_FINISH : ESC_ZFS_SCRUB_FINISH);
 841                 } else {
 842                         vdev_dtl_reassess(spa->spa_root_vdev, tx->tx_txg,
 843                             0, B_TRUE);
 844                 }
 845                 spa_errlog_rotate(spa);
 846
 847                 /*
 848                  * We may have finished replacing a device.
 849                  * Let the async thread assess this and handle the detach.
 850                  */
 851                 spa_async_request(spa, SPA_ASYNC_RESILVER_DONE);
 852         }
 853
 854         scn->scn_phys.scn_end_time = gethrestime_sec();
 855
 856         ASSERT(!dsl_scan_is_running(scn));
 857 }
 858
 859 /* ARGSUSED */
 860 static int
 861 dsl_scan_cancel_check(void *arg, dmu_tx_t *tx)
 862 {
 863         dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
 864
 865         if (!dsl_scan_is_running(scn))
 866                 return (SET_ERROR(ENOENT));
 867         return (0);
 868 }
 869
 870 /* ARGSUSED */
 871 static void
 872 dsl_scan_cancel_sync(void *arg, dmu_tx_t *tx)
 873 {
 874         dsl_scan_t *scn = dmu_tx_pool(tx)->dp_scan;
 875
 876         dsl_scan_done(scn, B_FALSE, tx);
 877         dsl_scan_sync_state(scn, tx, SYNC_MANDATORY);
 878         spa_event_notify(scn->scn_dp->dp_spa, NULL, NULL, ESC_ZFS_SCRUB_ABORT);
 879 }
 880
 881 int
 882 dsl_scan_cancel(dsl_pool_t *dp)
 883 {
 884         return (dsl_sync_task(spa_name(dp->dp_spa), dsl_scan_cancel_check,
 885             dsl_scan_cancel_sync, NULL, 3, ZFS_SPACE_CHECK_RESERVED));
 886 }
 887
 888 static int
 889 dsl_scrub_pause_resume_check(void *arg, dmu_tx_t *tx)
 890 {
 891         pool_scrub_cmd_t *cmd = arg;
 892         dsl_pool_t *dp = dmu_tx_pool(tx);
 893         dsl_scan_t *scn = dp->dp_scan;
 894
 895         if (*cmd == POOL_SCRUB_PAUSE) {
 896                 /* can't pause a scrub when there is no in-progress scrub */
 897                 if (!dsl_scan_scrubbing(dp))
 898                         return (SET_ERROR(ENOENT));
 899
 900                 /* can't pause a paused scrub */
 901                 if (dsl_scan_is_paused_scrub(scn))
 902                         return (SET_ERROR(EBUSY));
 903         } else if (*cmd != POOL_SCRUB_NORMAL) {
 904                 return (SET_ERROR(ENOTSUP));
 905         }
 906
 907         return (0);
 908 }
 909
 910 static void
 911 dsl_scrub_pause_resume_sync(void *arg, dmu_tx_t *tx)
 912 {
 913         pool_scrub_cmd_t *cmd = arg;
 914         dsl_pool_t *dp = dmu_tx_pool(tx);
 915         spa_t *spa = dp->dp_spa;
 916         dsl_scan_t *scn = dp->dp_scan;
 917
 918         if (*cmd == POOL_SCRUB_PAUSE) {
 919                 /* can't pause a scrub when there is no in-progress scrub */
 920                 spa->spa_scan_pass_scrub_pause = gethrestime_sec();
 921                 scn->scn_phys.scn_flags |= DSF_SCRUB_PAUSED;
 922                 dsl_scan_sync_state(scn, tx, SYNC_CACHED);
 923                 spa_event_notify(spa, NULL, NULL, ESC_ZFS_SCRUB_PAUSED);
 924         } else {
 925                 ASSERT3U(*cmd, ==, POOL_SCRUB_NORMAL);
 926                 if (dsl_scan_is_paused_scrub(scn)) {
 927                         /*
 928                          * We need to keep track of how much time we spend
 929                          * paused per pass so that we can adjust the scrub rate
 930                          * shown in the output of 'zpool status'
 931                          */
 932                         spa->spa_scan_pass_scrub_spent_paused +=
 933                             gethrestime_sec() - spa->spa_scan_pass_scrub_pause;
 934                         spa->spa_scan_pass_scrub_pause = 0;
 935                         scn->scn_phys.scn_flags &= ~DSF_SCRUB_PAUSED;
 936                         dsl_scan_sync_state(scn, tx, SYNC_CACHED);
 937                 }
 938         }
 939 }
 940
 941 /*
 942  * Set scrub pause/resume state if it makes sense to do so
 943  */
 944 int
 945 dsl_scrub_set_pause_resume(const dsl_pool_t *dp, pool_scrub_cmd_t cmd)
 946 {
 947         return (dsl_sync_task(spa_name(dp->dp_spa),
 948             dsl_scrub_pause_resume_check, dsl_scrub_pause_resume_sync, &cmd, 3,
 949             ZFS_SPACE_CHECK_RESERVED));
 950 }
 951
 952
 953 /* start a new scan, or restart an existing one. */
 954 void
 955 dsl_resilver_restart(dsl_pool_t *dp, uint64_t txg)
 956 {
 957         if (txg == 0) {
 958                 dmu_tx_t *tx;
 959                 tx = dmu_tx_create_dd(dp->dp_mos_dir);
 960                 VERIFY(0 == dmu_tx_assign(tx, TXG_WAIT));
 961
 962                 txg = dmu_tx_get_txg(tx);
 963                 dp->dp_scan->scn_restart_txg = txg;
 964                 dmu_tx_commit(tx);
 965         } else {
 966                 dp->dp_scan->scn_restart_txg = txg;
 967         }
 968         zfs_dbgmsg("restarting resilver txg=%llu", txg);
 969 }
 970
 971 void
 972 dsl_free(dsl_pool_t *dp, uint64_t txg, const blkptr_t *bp)
 973 {
 974         zio_free(dp->dp_spa, txg, bp);
 975 }
 976
 977 void
 978 dsl_free_sync(zio_t *pio, dsl_pool_t *dp, uint64_t txg, const blkptr_t *bpp)
 979 {
 980         ASSERT(dsl_pool_sync_context(dp));
 981         zio_nowait(zio_free_sync(pio, dp->dp_spa, txg, bpp, BP_GET_PSIZE(bpp),
 982             pio->io_flags));
 983 }
 984
 985 static int
 986 scan_ds_queue_compare(const void *a, const void *b)
 987 {
 988         const scan_ds_t *sds_a = a, *sds_b = b;
 989
 990         if (sds_a->sds_dsobj < sds_b->sds_dsobj)
 991                 return (-1);
 992         if (sds_a->sds_dsobj == sds_b->sds_dsobj)
 993                 return (0);
 994         return (1);
 995 }
 996
 997 static void
 998 scan_ds_queue_clear(dsl_scan_t *scn)
 999 {
1000         void *cookie = NULL;
1001         scan_ds_t *sds;
1002         while ((sds = avl_destroy_nodes(&scn->scn_queue, &cookie)) != NULL) {
1003                 kmem_free(sds, sizeof (*sds));
1004         }
1005 }
1006
1007 static boolean_t
1008 scan_ds_queue_contains(dsl_scan_t *scn, uint64_t dsobj, uint64_t *txg)
1009 {
1010         scan_ds_t srch, *sds;
1011
1012         srch.sds_dsobj = dsobj;
1013         sds = avl_find(&scn->scn_queue, &srch, NULL);
1014         if (sds != NULL && txg != NULL)
1015                 *txg = sds->sds_txg;
1016         return (sds != NULL);
1017 }
1018
1019 static void
1020 scan_ds_queue_insert(dsl_scan_t *scn, uint64_t dsobj, uint64_t txg)
1021 {
1022         scan_ds_t *sds;
1023         avl_index_t where;
1024
1025         sds = kmem_zalloc(sizeof (*sds), KM_SLEEP);
1026         sds->sds_dsobj = dsobj;
1027         sds->sds_txg = txg;
1028
1029         VERIFY3P(avl_find(&scn->scn_queue, sds, &where), ==, NULL);
1030         avl_insert(&scn->scn_queue, sds, where);
1031 }
1032
1033 static void
1034 scan_ds_queue_remove(dsl_scan_t *scn, uint64_t dsobj)
1035 {
1036         scan_ds_t srch, *sds;
1037
1038         srch.sds_dsobj = dsobj;
1039
1040         sds = avl_find(&scn->scn_queue, &srch, NULL);
1041         VERIFY(sds != NULL);
1042         avl_remove(&scn->scn_queue, sds);
1043         kmem_free(sds, sizeof (*sds));
1044 }
1045
1046 static void
1047 scan_ds_queue_sync(dsl_scan_t *scn, dmu_tx_t *tx)
1048 {
1049         dsl_pool_t *dp = scn->scn_dp;
1050         spa_t *spa = dp->dp_spa;
1051         dmu_object_type_t ot = (spa_version(spa) >= SPA_VERSION_DSL_SCRUB) ?
1052             DMU_OT_SCAN_QUEUE : DMU_OT_ZAP_OTHER;
1053
1054         ASSERT0(scn->scn_bytes_pending);
1055         ASSERT(scn->scn_phys.scn_queue_obj != 0);
1056
1057         VERIFY0(dmu_object_free(dp->dp_meta_objset,
1058             scn->scn_phys.scn_queue_obj, tx));
1059         scn->scn_phys.scn_queue_obj = zap_create(dp->dp_meta_objset, ot,
1060             DMU_OT_NONE, 0, tx);
1061         for (scan_ds_t *sds = avl_first(&scn->scn_queue);
1062             sds != NULL; sds = AVL_NEXT(&scn->scn_queue, sds)) {
1063                 VERIFY0(zap_add_int_key(dp->dp_meta_objset,
1064                     scn->scn_phys.scn_queue_obj, sds->sds_dsobj,
1065                     sds->sds_txg, tx));
1066         }
1067 }
1068
1069 /*
1070  * Computes the memory limit state that we're currently in. A sorted scan
1071  * needs quite a bit of memory to hold the sorting queue, so we need to
1072  * reasonably constrain the size so it doesn't impact overall system
1073  * performance. We compute two limits:
1074  * 1) Hard memory limit: if the amount of memory used by the sorting
1075  *      queues on a pool gets above this value, we stop the metadata
1076  *      scanning portion and start issuing the queued up and sorted
1077  *      I/Os to reduce memory usage.
1078  *      This limit is calculated as a fraction of physmem (by default 5%).
1079  *      We constrain the lower bound of the hard limit to an absolute
1080  *      minimum of zfs_scan_mem_lim_min (default: 16 MiB). We also constrain
1081  *      the upper bound to 5% of the total pool size - no chance we'll
1082  *      ever need that much memory, but just to keep the value in check.
1083  * 2) Soft memory limit: once we hit the hard memory limit, we start
1084  *      issuing I/O to reduce queue memory usage, but we don't want to
1085  *      completely empty out the queues, since we might be able to find I/Os
1086  *      that will fill in the gaps of our non-sequential IOs at some point
1087  *      in the future. So we stop the issuing of I/Os once the amount of
1088  *      memory used drops below the soft limit (at which point we stop issuing
1089  *      I/O and start scanning metadata again).
1090  *
1091  *      This limit is calculated by subtracting a fraction of the hard
1092  *      limit from the hard limit. By default this fraction is 5%, so
1093  *      the soft limit is 95% of the hard limit. We cap the size of the
1094  *      difference between the hard and soft limits at an absolute
1095  *      maximum of zfs_scan_mem_lim_soft_max (default: 128 MiB) - this is
1096  *      sufficient to not cause too frequent switching between the
1097  *      metadata scan and I/O issue (even at 2k recordsize, 128 MiB's
1098  *      worth of queues is about 1.2 GiB of on-pool data, so scanning
1099  *      that should take at least a decent fraction of a second).
1100  */
1101 static boolean_t
1102 dsl_scan_should_clear(dsl_scan_t *scn)
1103 {
1104         vdev_t *rvd = scn->scn_dp->dp_spa->spa_root_vdev;
1105         uint64_t mlim_hard, mlim_soft, mused;
1106         uint64_t alloc = metaslab_class_get_alloc(spa_normal_class(
1107             scn->scn_dp->dp_spa));
1108
1109         mlim_hard = MAX((physmem / zfs_scan_mem_lim_fact) * PAGESIZE,
1110             zfs_scan_mem_lim_min);
1111         mlim_hard = MIN(mlim_hard, alloc / 20);
1112         mlim_soft = mlim_hard - MIN(mlim_hard / zfs_scan_mem_lim_soft_fact,
1113             zfs_scan_mem_lim_soft_max);
1114         mused = 0;
1115         for (uint64_t i = 0; i < rvd->vdev_children; i++) {
1116                 vdev_t *tvd = rvd->vdev_child[i];
1117                 dsl_scan_io_queue_t *queue;
1118
1119                 mutex_enter(&tvd->vdev_scan_io_queue_lock);
1120                 queue = tvd->vdev_scan_io_queue;
1121                 if (queue != NULL) {
1122                         /* #extents in exts_by_size = # in exts_by_addr */
1123                         mused += avl_numnodes(&queue->q_exts_by_size) *
1124                             sizeof (range_seg_t) +
1125                             avl_numnodes(&queue->q_sios_by_addr) *
1126                             sizeof (scan_io_t);
1127                 }
1128                 mutex_exit(&tvd->vdev_scan_io_queue_lock);
1129         }
1130
1131         dprintf("current scan memory usage: %llu bytes\n", (longlong_t)mused);
1132
1133         if (mused == 0)
1134                 ASSERT0(scn->scn_bytes_pending);
1135
1136         /*
1137          * If we are above our hard limit, we need to clear out memory.
1138          * If we are below our soft limit, we need to accumulate sequential IOs.
1139          * Otherwise, we should keep doing whatever we are currently doing.
1140          */
1141         if (mused >= mlim_hard)
1142                 return (B_TRUE);
1143         else if (mused < mlim_soft)
1144                 return (B_FALSE);
1145         else
1146                 return (scn->scn_clearing);
1147 }
1148
1149 static boolean_t
1150 dsl_scan_check_suspend(dsl_scan_t *scn, const zbookmark_phys_t *zb)
1151 {
1152         /* we never skip user/group accounting objects */
1153         if (zb && (int64_t)zb->zb_object < 0)
1154                 return (B_FALSE);
1155
1156         if (scn->scn_suspending)
1157                 return (B_TRUE); /* we're already suspending */
1158
1159         if (!ZB_IS_ZERO(&scn->scn_phys.scn_bookmark))
1160                 return (B_FALSE); /* we're resuming */
1161
1162         /* We only know how to resume from level-0 blocks. */
1163         if (zb && zb->zb_level != 0)
1164                 return (B_FALSE);
1165
1166         /*
1167          * We suspend if:
1168          *  - we have scanned for at least the minimum time (default 1 sec
1169          *    for scrub, 3 sec for resilver), and either we have sufficient
1170          *    dirty data that we are starting to write more quickly
1171          *    (default 30%), or someone is explicitly waiting for this txg
1172          *    to complete.
1173          *  or
1174          *  - the spa is shutting down because this pool is being exported
1175          *    or the machine is rebooting.
1176          *  or
1177          *  - the scan queue has reached its memory use limit
1178          */
1179         uint64_t elapsed_nanosecs = gethrtime();
1180         uint64_t curr_time_ns = gethrtime();
1181         uint64_t scan_time_ns = curr_time_ns - scn->scn_sync_start_time;
1182         uint64_t sync_time_ns = curr_time_ns -
1183             scn->scn_dp->dp_spa->spa_sync_starttime;
1184
1185         int dirty_pct = scn->scn_dp->dp_dirty_total * 100 / zfs_dirty_data_max;
1186         int mintime = (scn->scn_phys.scn_func == POOL_SCAN_RESILVER) ?
1187             zfs_resilver_min_time_ms : zfs_scrub_min_time_ms;
1188
1189         if ((NSEC2MSEC(scan_time_ns) > mintime &&
1190             (dirty_pct >= zfs_vdev_async_write_active_min_dirty_percent ||
1191             txg_sync_waiting(scn->scn_dp) ||
1192             NSEC2SEC(sync_time_ns) >= zfs_txg_timeout)) ||
1193             spa_shutting_down(scn->scn_dp->dp_spa) ||
1194             (zfs_scan_strict_mem_lim && dsl_scan_should_clear(scn))) {
1195                 if (zb) {
1196                         dprintf("suspending at bookmark %llx/%llx/%llx/%llx\n",
1197                             (longlong_t)zb->zb_objset,
1198                             (longlong_t)zb->zb_object,
1199                             (longlong_t)zb->zb_level,
1200                             (longlong_t)zb->zb_blkid);
1201                         scn->scn_phys.scn_bookmark = *zb;
1202                 } else {
1203                         dsl_scan_phys_t *scnp = &scn->scn_phys;
1204
1205                         dprintf("suspending at at DDT bookmark "
1206                             "%llx/%llx/%llx/%llx\n",
1207                             (longlong_t)scnp->scn_ddt_bookmark.ddb_class,
1208                             (longlong_t)scnp->scn_ddt_bookmark.ddb_type,
1209                             (longlong_t)scnp->scn_ddt_bookmark.ddb_checksum,
1210                             (longlong_t)scnp->scn_ddt_bookmark.ddb_cursor);
1211                 }
1212                 scn->scn_suspending = B_TRUE;
1213                 return (B_TRUE);
1214         }
1215         return (B_FALSE);
1216 }
1217
1218 typedef struct zil_scan_arg {
1219         dsl_pool_t      *zsa_dp;
1220         zil_header_t    *zsa_zh;
1221 } zil_scan_arg_t;
1222
1223 /* ARGSUSED */
1224 static int
1225 dsl_scan_zil_block(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
1226 {
1227         zil_scan_arg_t *zsa = arg;
1228         dsl_pool_t *dp = zsa->zsa_dp;
1229         dsl_scan_t *scn = dp->dp_scan;
1230         zil_header_t *zh = zsa->zsa_zh;
1231         zbookmark_phys_t zb;
1232
1233         if (BP_IS_HOLE(bp) || bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
1234                 return (0);
1235
1236         /*
1237          * One block ("stubby") can be allocated a long time ago; we
1238          * want to visit that one because it has been allocated
1239          * (on-disk) even if it hasn't been claimed (even though for
1240          * scrub there's nothing to do to it).
1241          */
1242         if (claim_txg == 0 && bp->blk_birth >= spa_min_claim_txg(dp->dp_spa))
1243                 return (0);
1244
1245         SET_BOOKMARK(&zb, zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET],
1246             ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
1247
1248         VERIFY(0 == scan_funcs[scn->scn_phys.scn_func](dp, bp, &zb));
1249         return (0);
1250 }
1251
1252 /* ARGSUSED */
1253 static int
1254 dsl_scan_zil_record(zilog_t *zilog, lr_t *lrc, void *arg, uint64_t claim_txg)
1255 {
1256         if (lrc->lrc_txtype == TX_WRITE) {
1257                 zil_scan_arg_t *zsa = arg;
1258                 dsl_pool_t *dp = zsa->zsa_dp;
1259                 dsl_scan_t *scn = dp->dp_scan;
1260                 zil_header_t *zh = zsa->zsa_zh;
1261                 lr_write_t *lr = (lr_write_t *)lrc;
1262                 blkptr_t *bp = &lr->lr_blkptr;
1263                 zbookmark_phys_t zb;
1264
1265                 if (BP_IS_HOLE(bp) ||
1266                     bp->blk_birth <= scn->scn_phys.scn_cur_min_txg)
1267                         return (0);
1268
1269                 /*
1270                  * birth can be < claim_txg if this record's txg is
1271                  * already txg sync'ed (but this log block contains
1272                  * other records that are not synced)
1273                  */
1274                 if (claim_txg == 0 || bp->blk_birth < claim_txg)
1275                         return (0);
1276
1277                 SET_BOOKMARK(&zb, zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET],
1278                     lr->lr_foid, ZB_ZIL_LEVEL,
1279                     lr->lr_offset / BP_GET_LSIZE(bp));
1280
1281                 VERIFY(0 == scan_funcs[scn->scn_phys.scn_func](dp, bp, &zb));
1282         }
1283         return (0);
1284 }
1285
1286 static void
1287 dsl_scan_zil(dsl_pool_t *dp, zil_header_t *zh)
1288 {
1289         uint64_t claim_txg = zh->zh_claim_txg;
1290         zil_scan_arg_t zsa = { dp, zh };
1291         zilog_t *zilog;
1292
1293         ASSERT(spa_writeable(dp->dp_spa));
1294
1295         /*
1296          * We only want to visit blocks that have been claimed
1297          * but not yet replayed.
1298          */
1299         if (claim_txg == 0)
1300                 return;
1301
1302         zilog = zil_alloc(dp->dp_meta_objset, zh);
1303
1304         (void) zil_parse(zilog, dsl_scan_zil_block, dsl_scan_zil_record, &zsa,
1305             claim_txg);
1306
1307         zil_free(zilog);
1308 }
1309
1310 /*
1311  * We compare scan_prefetch_issue_ctx_t's based on their bookmarks. The idea
1312  * here is to sort the AVL tree by the order each block will be needed.
1313  */
1314 static int
1315 scan_prefetch_queue_compare(const void *a, const void *b)
1316 {
1317         const scan_prefetch_issue_ctx_t *spic_a = a, *spic_b = b;
1318         const scan_prefetch_ctx_t *spc_a = spic_a->spic_spc;
1319         const scan_prefetch_ctx_t *spc_b = spic_b->spic_spc;
1320
1321         return (zbookmark_compare(spc_a->spc_datablkszsec,
1322             spc_a->spc_indblkshift, spc_b->spc_datablkszsec,
1323             spc_b->spc_indblkshift, &spic_a->spic_zb, &spic_b->spic_zb));
1324 }
1325
1326 static void
1327 scan_prefetch_ctx_rele(scan_prefetch_ctx_t *spc, void *tag)
1328 {
1329         if (refcount_remove(&spc->spc_refcnt, tag) == 0) {
1330                 refcount_destroy(&spc->spc_refcnt);
1331                 kmem_free(spc, sizeof (scan_prefetch_ctx_t));
1332         }
1333 }
1334
1335 static scan_prefetch_ctx_t *
1336 scan_prefetch_ctx_create(dsl_scan_t *scn, dnode_phys_t *dnp, void *tag)
1337 {
1338         scan_prefetch_ctx_t *spc;
1339
1340         spc = kmem_alloc(sizeof (scan_prefetch_ctx_t), KM_SLEEP);
1341         refcount_create(&spc->spc_refcnt);
1342         refcount_add(&spc->spc_refcnt, tag);
1343         spc->spc_scn = scn;
1344         if (dnp != NULL) {
1345                 spc->spc_datablkszsec = dnp->dn_datablkszsec;
1346                 spc->spc_indblkshift = dnp->dn_indblkshift;
1347                 spc->spc_root = B_FALSE;
1348         } else {
1349                 spc->spc_datablkszsec = 0;
1350                 spc->spc_indblkshift = 0;
1351                 spc->spc_root = B_TRUE;
1352         }
1353
1354         return (spc);
1355 }
1356
1357 static void
1358 scan_prefetch_ctx_add_ref(scan_prefetch_ctx_t *spc, void *tag)
1359 {
1360         refcount_add(&spc->spc_refcnt, tag);
1361 }
1362
1363 static boolean_t
1364 dsl_scan_check_prefetch_resume(scan_prefetch_ctx_t *spc,
1365     const zbookmark_phys_t *zb)
1366 {
1367         zbookmark_phys_t *last_zb = &spc->spc_scn->scn_prefetch_bookmark;
1368         dnode_phys_t tmp_dnp;
1369         dnode_phys_t *dnp = (spc->spc_root) ? NULL : &tmp_dnp;
1370
1371         if (zb->zb_objset != last_zb->zb_objset)
1372                 return (B_TRUE);
1373         if ((int64_t)zb->zb_object < 0)
1374                 return (B_FALSE);
1375
1376         tmp_dnp.dn_datablkszsec = spc->spc_datablkszsec;
1377         tmp_dnp.dn_indblkshift = spc->spc_indblkshift;
1378
1379         if (zbookmark_subtree_completed(dnp, zb, last_zb))
1380                 return (B_TRUE);
1381
1382         return (B_FALSE);
1383 }
1384
1385 static void
1386 dsl_scan_prefetch(scan_prefetch_ctx_t *spc, blkptr_t *bp, zbookmark_phys_t *zb)
1387 {
1388         avl_index_t idx;
1389         dsl_scan_t *scn = spc->spc_scn;
1390         spa_t *spa = scn->scn_dp->dp_spa;
1391         scan_prefetch_issue_ctx_t *spic;
1392
1393         if (zfs_no_scrub_prefetch)
1394                 return;
1395
1396         if (BP_IS_HOLE(bp) || bp->blk_birth <= scn->scn_phys.scn_cur_min_txg ||
1397             (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_DNODE &&
1398             BP_GET_TYPE(bp) != DMU_OT_OBJSET))
1399                 return;
1400
1401         if (dsl_scan_check_prefetch_resume(spc, zb))
1402                 return;
1403
1404         scan_prefetch_ctx_add_ref(spc, scn);
1405         spic = kmem_alloc(sizeof (scan_prefetch_issue_ctx_t), KM_SLEEP);
1406         spic->spic_spc = spc;
1407         spic->spic_bp = *bp;
1408         spic->spic_zb = *zb;
1409
1410         /*
1411          * Add the IO to the queue of blocks to prefetch. This allows us to
1412          * prioritize blocks that we will need first for the main traversal
1413          * thread.
1414          */
1415         mutex_enter(&spa->spa_scrub_lock);
1416         if (avl_find(&scn->scn_prefetch_queue, spic, &idx) != NULL) {
1417                 /* this block is already queued for prefetch */
1418                 kmem_free(spic, sizeof (scan_prefetch_issue_ctx_t));
1419                 scan_prefetch_ctx_rele(spc, scn);
1420                 mutex_exit(&spa->spa_scrub_lock);
1421                 return;
1422         }
1423
1424         avl_insert(&scn->scn_prefetch_queue, spic, idx);
1425         cv_broadcast(&spa->spa_scrub_io_cv);
1426         mutex_exit(&spa->spa_scrub_lock);
1427 }
1428
1429 static void
1430 dsl_scan_prefetch_dnode(dsl_scan_t *scn, dnode_phys_t *dnp,
1431     uint64_t objset, uint64_t object)
1432 {
1433         int i;
1434         zbookmark_phys_t zb;
1435         scan_prefetch_ctx_t *spc;
1436
1437         if (dnp->dn_nblkptr == 0 && !(dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR))
1438                 return;
1439
1440         SET_BOOKMARK(&zb, objset, object, 0, 0);
1441
1442         spc = scan_prefetch_ctx_create(scn, dnp, FTAG);
1443
1444         for (i = 0; i < dnp->dn_nblkptr; i++) {
1445                 zb.zb_level = BP_GET_LEVEL(&dnp->dn_blkptr[i]);
1446                 zb.zb_blkid = i;
1447                 dsl_scan_prefetch(spc, &dnp->dn_blkptr[i], &zb);
1448         }
1449
1450         if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
1451                 zb.zb_level = 0;
1452                 zb.zb_blkid = DMU_SPILL_BLKID;
1453                 dsl_scan_prefetch(spc, &dnp->dn_spill, &zb);
1454         }
1455
1456         scan_prefetch_ctx_rele(spc, FTAG);
1457 }
1458
1459 void
1460 dsl_scan_prefetch_cb(zio_t *zio, const zbookmark_phys_t *zb, const blkptr_t *bp,
1461     arc_buf_t *buf, void *private)
1462 {
1463         scan_prefetch_ctx_t *spc = private;
1464         dsl_scan_t *scn = spc->spc_scn;
1465         spa_t *spa = scn->scn_dp->dp_spa;
1466
1467         /* broadcast that the IO has completed for rate limitting purposes */
1468         mutex_enter(&spa->spa_scrub_lock);
1469         ASSERT3U(spa->spa_scrub_inflight, >=, BP_GET_PSIZE(bp));
1470         spa->spa_scrub_inflight -= BP_GET_PSIZE(bp);
1471         cv_broadcast(&spa->spa_scrub_io_cv);
1472         mutex_exit(&spa->spa_scrub_lock);
1473
1474         /* if there was an error or we are done prefetching, just cleanup */
1475         if (buf == NULL || scn->scn_suspending)
1476                 goto out;
1477
1478         if (BP_GET_LEVEL(bp) > 0) {
1479                 int i;
1480                 blkptr_t *cbp;
1481                 int epb = BP_GET_LSIZE(bp) >> SPA_BLKPTRSHIFT;
1482                 zbookmark_phys_t czb;
1483
1484                 for (i = 0, cbp = buf->b_data; i < epb; i++, cbp++) {
1485                         SET_BOOKMARK(&czb, zb->zb_objset, zb->zb_object,
1486                             zb->zb_level - 1, zb->zb_blkid * epb + i);
1487                         dsl_scan_prefetch(spc, cbp, &czb);
1488                 }
1489         } else if (BP_GET_TYPE(bp) == DMU_OT_DNODE) {
1490                 dnode_phys_t *cdnp = buf->b_data;
1491                 int i;
1492                 int epb = BP_GET_LSIZE(bp) >> DNODE_SHIFT;
1493
1494                 for (i = 0, cdnp = buf->b_data; i < epb; i++, cdnp++) {
1495                         dsl_scan_prefetch_dnode(scn, cdnp,
1496                                                 zb->zb_objset, zb->zb_blkid * epb + i);
1497                 }
1498         } else if (BP_GET_TYPE(bp) == DMU_OT_OBJSET) {
1499                 objset_phys_t *osp = buf->b_data;
1500
1501                 dsl_scan_prefetch_dnode(scn, &osp->os_meta_dnode,
1502                     zb->zb_objset, DMU_META_DNODE_OBJECT);
1503
1504                 if (OBJSET_BUF_HAS_USERUSED(buf)) {
1505                         dsl_scan_prefetch_dnode(scn,
1506                             &osp->os_groupused_dnode, zb->zb_objset,
1507                             DMU_GROUPUSED_OBJECT);
1508                         dsl_scan_prefetch_dnode(scn,
1509                             &osp->os_userused_dnode, zb->zb_objset,
1510                             DMU_USERUSED_OBJECT);
1511                 }
1512         }
1513
1514 out:
1515         if (buf != NULL)
1516                 arc_buf_destroy(buf, private);
1517         scan_prefetch_ctx_rele(spc, scn);
1518 }
1519
1520 /* ARGSUSED */
1521 static void
1522 dsl_scan_prefetch_thread(void *arg)
1523 {
1524         dsl_scan_t *scn = arg;
1525         spa_t *spa = scn->scn_dp->dp_spa;
1526         vdev_t *rvd = spa->spa_root_vdev;
1527         uint64_t maxinflight = rvd->vdev_children * zfs_top_maxinflight;
1528         scan_prefetch_issue_ctx_t *spic;
1529
1530         /* loop until we are told to stop */
1531         while (!scn->scn_prefetch_stop) {
1532                 arc_flags_t flags = ARC_FLAG_NOWAIT |
1533                     ARC_FLAG_PRESCIENT_PREFETCH | ARC_FLAG_PREFETCH;
1534                 int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SCAN_THREAD;
1535
1536                 mutex_enter(&spa->spa_scrub_lock);
1537
1538                 /*
1539                  * Wait until we have an IO to issue and are not above our
1540                  * maximum in flight limit.
1541                  */
1542                 while (!scn->scn_prefetch_stop &&
1543                     (avl_numnodes(&scn->scn_prefetch_queue) == 0 ||
1544                     spa->spa_scrub_inflight >= scn->scn_maxinflight_bytes)) {
1545                         cv_wait(&spa->spa_scrub_io_cv, &spa->spa_scrub_lock);
1546                 }
1547
1548                 /* recheck if we should stop since we waited for the cv */
1549                 if (scn->scn_prefetch_stop) {
1550                         mutex_exit(&spa->spa_scrub_lock);
1551                         break;
1552                 }
1553
1554                 /* remove the prefetch IO from the tree */
1555                 spic = avl_first(&scn->scn_prefetch_queue);
1556                 spa->spa_scrub_inflight += BP_GET_PSIZE(&spic->spic_bp);
1557                 avl_remove(&scn->scn_prefetch_queue, spic);
1558
1559                 mutex_exit(&spa->spa_scrub_lock);
1560
1561                 /* issue the prefetch asynchronously */
1562                 (void) arc_read(scn->scn_zio_root, scn->scn_dp->dp_spa,
1563                     &spic->spic_bp, dsl_scan_prefetch_cb, spic->spic_spc,
1564                     ZIO_PRIORITY_SCRUB, zio_flags, &flags, &spic->spic_zb);
1565
1566                 kmem_free(spic, sizeof (scan_prefetch_issue_ctx_t));
1567         }
1568
1569         ASSERT(scn->scn_prefetch_stop);
1570
1571         /* free any prefetches we didn't get to complete */
1572         mutex_enter(&spa->spa_scrub_lock);
1573         while ((spic = avl_first(&scn->scn_prefetch_queue)) != NULL) {
1574                 avl_remove(&scn->scn_prefetch_queue, spic);
1575                 scan_prefetch_ctx_rele(spic->spic_spc, scn);
1576                 kmem_free(spic, sizeof (scan_prefetch_issue_ctx_t));
1577         }
1578         ASSERT0(avl_numnodes(&scn->scn_prefetch_queue));
1579         mutex_exit(&spa->spa_scrub_lock);
1580 }
1581
1582 static boolean_t
1583 dsl_scan_check_resume(dsl_scan_t *scn, const dnode_phys_t *dnp,
1584     const zbookmark_phys_t *zb)
1585 {
1586         /*
1587          * We never skip over user/group accounting objects (obj<0)
1588          */
1589         if (!ZB_IS_ZERO(&scn->scn_phys.scn_bookmark) &&
1590             (int64_t)zb->zb_object >= 0) {
1591                 /*
1592                  * If we already visited this bp & everything below (in
1593                  * a prior txg sync), don't bother doing it again.
1594                  */
1595                 if (zbookmark_subtree_completed(dnp, zb,
1596                     &scn->scn_phys.scn_bookmark))
1597                         return (B_TRUE);
1598
1599                 /*
1600                  * If we found the block we're trying to resume from, or
1601                  * we went past it to a different object, zero it out to
1602                  * indicate that it's OK to start checking for suspending
1603                  * again.
1604                  */
1605                 if (bcmp(zb, &scn->scn_phys.scn_bookmark, sizeof (*zb)) == 0 ||
1606                     zb->zb_object > scn->scn_phys.scn_bookmark.zb_object) {
1607                         dprintf("resuming at %llx/%llx/%llx/%llx\n",
1608                             (longlong_t)zb->zb_objset,
1609                             (longlong_t)zb->zb_object,
1610                             (longlong_t)zb->zb_level,
1611                             (longlong_t)zb->zb_blkid);
1612                         bzero(&scn->scn_phys.scn_bookmark, sizeof (*zb));
1613                 }
1614         }
1615         return (B_FALSE);
1616 }
1617
1618 static void dsl_scan_visitbp(blkptr_t *bp, const zbookmark_phys_t *zb,
1619     dnode_phys_t *dnp, dsl_dataset_t *ds, dsl_scan_t *scn,
1620     dmu_objset_type_t ostype, dmu_tx_t *tx);
1621 static void dsl_scan_visitdnode(
1622     dsl_scan_t *, dsl_dataset_t *ds, dmu_objset_type_t ostype,
1623     dnode_phys_t *dnp, uint64_t object, dmu_tx_t *tx);
1624
1625 /*
1626  * Return nonzero on i/o error.
1627  * Return new buf to write out in *bufp.
1628  */
1629 static int
1630 dsl_scan_recurse(dsl_scan_t *scn, dsl_dataset_t *ds, dmu_objset_type_t ostype,
1631     dnode_phys_t *dnp, const blkptr_t *bp,
1632     const zbookmark_phys_t *zb, dmu_tx_t *tx)
1633 {
1634         dsl_pool_t *dp = scn->scn_dp;
1635         int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SCAN_THREAD;
1636         int err;
1637
1638         if (BP_GET_LEVEL(bp) > 0) {
1639                 arc_flags_t flags = ARC_FLAG_WAIT;
1640                 int i;
1641                 blkptr_t *cbp;
1642                 int epb = BP_GET_LSIZE(bp) >> SPA_BLKPTRSHIFT;
1643                 arc_buf_t *buf;
1644
1645                 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, &buf,
1646                     ZIO_PRIORITY_SCRUB, zio_flags, &flags, zb);
1647                 if (err) {
1648                         scn->scn_phys.scn_errors++;
1649                         return (err);
1650                 }
1651                 for (i = 0, cbp = buf->b_data; i < epb; i++, cbp++) {
1652                         zbookmark_phys_t czb;
1653
1654                         SET_BOOKMARK(&czb, zb->zb_objset, zb->zb_object,
1655                             zb->zb_level - 1,
1656                             zb->zb_blkid * epb + i);
1657                         dsl_scan_visitbp(cbp, &czb, dnp,
1658                             ds, scn, ostype, tx);
1659                 }
1660                 arc_buf_destroy(buf, &buf);
1661         } else if (BP_GET_TYPE(bp) == DMU_OT_DNODE) {
1662                 arc_flags_t flags = ARC_FLAG_WAIT;
1663                 dnode_phys_t *cdnp;
1664                 int i;
1665                 int epb = BP_GET_LSIZE(bp) >> DNODE_SHIFT;
1666                 arc_buf_t *buf;
1667
1668                 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, &buf,
1669                     ZIO_PRIORITY_SCRUB, zio_flags, &flags, zb);
1670                 if (err) {
1671                         scn->scn_phys.scn_errors++;
1672                         return (err);
1673                 }
1674                 for (i = 0, cdnp = buf->b_data; i < epb; i++, cdnp++) {
1675                         dsl_scan_visitdnode(scn, ds, ostype,
1676                             cdnp, zb->zb_blkid * epb + i, tx);
1677                 }
1678
1679                 arc_buf_destroy(buf, &buf);
1680         } else if (BP_GET_TYPE(bp) == DMU_OT_OBJSET) {
1681                 arc_flags_t flags = ARC_FLAG_WAIT;
1682                 objset_phys_t *osp;
1683                 arc_buf_t *buf;
1684
1685                 err = arc_read(NULL, dp->dp_spa, bp, arc_getbuf_func, &buf,
1686                     ZIO_PRIORITY_SCRUB, zio_flags, &flags, zb);
1687                 if (err) {
1688                         scn->scn_phys.scn_errors++;
1689                         return (err);
1690                 }
1691
1692                 osp = buf->b_data;
1693
1694                 dsl_scan_visitdnode(scn, ds, osp->os_type,
1695                     &osp->os_meta_dnode, DMU_META_DNODE_OBJECT, tx);
1696
1697                 if (OBJSET_BUF_HAS_USERUSED(buf)) {
1698                         /*
1699                          * We also always visit user/group accounting
1700                          * objects, and never skip them, even if we are
1701                          * suspending.  This is necessary so that the space
1702                          * deltas from this txg get integrated.
1703                          */
1704                         dsl_scan_visitdnode(scn, ds, osp->os_type,
1705                             &osp->os_groupused_dnode,
1706                             DMU_GROUPUSED_OBJECT, tx);
1707                         dsl_scan_visitdnode(scn, ds, osp->os_type,
1708                             &osp->os_userused_dnode,
1709                             DMU_USERUSED_OBJECT, tx);
1710                 }
1711                 arc_buf_destroy(buf, &buf);
1712         }
1713
1714         return (0);
1715 }
1716
1717 static void
1718 dsl_scan_visitdnode(dsl_scan_t *scn, dsl_dataset_t *ds,
1719     dmu_objset_type_t ostype, dnode_phys_t *dnp,
1720     uint64_t object, dmu_tx_t *tx)
1721 {
1722         int j;
1723
1724         for (j = 0; j < dnp->dn_nblkptr; j++) {
1725                 zbookmark_phys_t czb;
1726
1727                 SET_BOOKMARK(&czb, ds ? ds->ds_object : 0, object,
1728                     dnp->dn_nlevels - 1, j);
1729                 dsl_scan_visitbp(&dnp->dn_blkptr[j],
1730                     &czb, dnp, ds, scn, ostype, tx);
1731         }
1732
1733         if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
1734                 zbookmark_phys_t czb;
1735                 SET_BOOKMARK(&czb, ds ? ds->ds_object : 0, object,
1736                     0, DMU_SPILL_BLKID);
1737                 dsl_scan_visitbp(&dnp->dn_spill,
1738                     &czb, dnp, ds, scn, ostype, tx);
1739         }
1740 }
1741
1742 /*
1743  * The arguments are in this order because mdb can only print the
1744  * first 5; we want them to be useful.
1745  */
1746 static void
1747 dsl_scan_visitbp(blkptr_t *bp, const zbookmark_phys_t *zb,
1748     dnode_phys_t *dnp, dsl_dataset_t *ds, dsl_scan_t *scn,
1749     dmu_objset_type_t ostype, dmu_tx_t *tx)
1750 {
1751         dsl_pool_t *dp = scn->scn_dp;
1752         blkptr_t *bp_toread = NULL;
1753
1754         if (dsl_scan_check_suspend(scn, zb))
1755                 return;
1756
1757         if (dsl_scan_check_resume(scn, dnp, zb))
1758                 return;
1759
1760         scn->scn_visited_this_txg++;
1761
1762         dprintf_bp(bp,
1763             "visiting ds=%p/%llu zb=%llx/%llx/%llx/%llx bp=%p",
1764             ds, ds ? ds->ds_object : 0,
1765             zb->zb_objset, zb->zb_object, zb->zb_level, zb->zb_blkid,
1766             bp);
1767
1768         if (BP_IS_HOLE(bp)) {
1769                 scn->scn_holes_this_txg++;
1770                 return;
1771         }
1772
1773         if (bp->blk_birth <= scn->scn_phys.scn_cur_min_txg) {
1774                 scn->scn_lt_min_this_txg++;
1775                 return;
1776         }
1777
1778         bp_toread = kmem_alloc(sizeof (blkptr_t), KM_SLEEP);
1779         *bp_toread = *bp;
1780
1781         if (dsl_scan_recurse(scn, ds, ostype, dnp, bp_toread, zb, tx) != 0)
1782                 return;
1783
1784         /*
1785          * If dsl_scan_ddt() has already visited this block, it will have
1786          * already done any translations or scrubbing, so don't call the
1787          * callback again.
1788          */
1789         if (ddt_class_contains(dp->dp_spa,
1790             scn->scn_phys.scn_ddt_class_max, bp)) {
1791                 scn->scn_ddt_contained_this_txg++;
1792                 goto out;
1793         }
1794
1795         /*
1796          * If this block is from the future (after cur_max_txg), then we
1797          * are doing this on behalf of a deleted snapshot, and we will
1798          * revisit the future block on the next pass of this dataset.
1799          * Don't scan it now unless we need to because something
1800          * under it was modified.
1801          */
1802         if (BP_PHYSICAL_BIRTH(bp) > scn->scn_phys.scn_cur_max_txg) {
1803                 scn->scn_gt_max_this_txg++;
1804                 goto out;
1805         }
1806
1807         scan_funcs[scn->scn_phys.scn_func](dp, bp, zb);
1808 out:
1809         kmem_free(bp_toread, sizeof (blkptr_t));
1810 }
1811
1812 static void
1813 dsl_scan_visit_rootbp(dsl_scan_t *scn, dsl_dataset_t *ds, blkptr_t *bp,
1814     dmu_tx_t *tx)
1815 {
1816         zbookmark_phys_t zb;
1817         scan_prefetch_ctx_t *spc;
1818
1819         SET_BOOKMARK(&zb, ds ? ds->ds_object : DMU_META_OBJSET,
1820             ZB_ROOT_OBJECT, ZB_ROOT_LEVEL, ZB_ROOT_BLKID);
1821
1822         if (ZB_IS_ZERO(&scn->scn_phys.scn_bookmark)) {
1823                 SET_BOOKMARK(&scn->scn_prefetch_bookmark,
1824                     zb.zb_objset, 0, 0, 0);
1825         } else {
1826                 scn->scn_prefetch_bookmark = scn->scn_phys.scn_bookmark;
1827         }
1828
1829         scn->scn_objsets_visited_this_txg++;
1830
1831         spc = scan_prefetch_ctx_create(scn, NULL, FTAG);
1832         dsl_scan_prefetch(spc, bp, &zb);
1833         scan_prefetch_ctx_rele(spc, FTAG);
1834
1835         dsl_scan_visitbp(bp, &zb, NULL, ds, scn, DMU_OST_NONE, tx);
1836
1837         dprintf_ds(ds, "finished scan%s", "");
1838 }
1839
1840 static void
1841 ds_destroyed_scn_phys(dsl_dataset_t *ds, dsl_scan_phys_t *scn_phys)
1842 {
1843         if (scn_phys->scn_bookmark.zb_objset == ds->ds_object) {
1844                 if (ds->ds_is_snapshot) {
1845                         /*
1846                          * Note:
1847                          *  - scn_cur_{min,max}_txg stays the same.
1848                          *  - Setting the flag is not really necessary if
1849                          *    scn_cur_max_txg == scn_max_txg, because there
1850                          *    is nothing after this snapshot that we care
1851                          *    about.  However, we set it anyway and then
1852                          *    ignore it when we retraverse it in
1853                          *    dsl_scan_visitds().
1854                          */
1855                         scn_phys->scn_bookmark.zb_objset =
1856                             dsl_dataset_phys(ds)->ds_next_snap_obj;
1857                         zfs_dbgmsg("destroying ds %llu; currently traversing; "
1858                             "reset zb_objset to %llu",
1859                             (u_longlong_t)ds->ds_object,
1860                             (u_longlong_t)dsl_dataset_phys(ds)->
1861                             ds_next_snap_obj);
1862                         scn_phys->scn_flags |= DSF_VISIT_DS_AGAIN;
1863                 } else {
1864                         SET_BOOKMARK(&scn_phys->scn_bookmark,
1865                             ZB_DESTROYED_OBJSET, 0, 0, 0);
1866                         zfs_dbgmsg("destroying ds %llu; currently traversing; "
1867                             "reset bookmark to -1,0,0,0",
1868                             (u_longlong_t)ds->ds_object);
1869                 }
1870         }
1871 }
1872
1873 /*
1874  * Invoked when a dataset is destroyed. We need to make sure that:
1875  *
1876  * 1) If it is the dataset that was currently being scanned, we write
1877  *      a new dsl_scan_phys_t and marking the objset reference in it
1878  *      as destroyed.
1879  * 2) Remove it from the work queue, if it was present.
1880  *
1881  * If the dataset was actually a snapshot, instead of marking the dataset
1882  * as destroyed, we instead substitute the next snapshot in line.
1883  */
1884 void
1885 dsl_scan_ds_destroyed(dsl_dataset_t *ds, dmu_tx_t *tx)
1886 {
1887         dsl_pool_t *dp = ds->ds_dir->dd_pool;
1888         dsl_scan_t *scn = dp->dp_scan;
1889         uint64_t mintxg;
1890
1891         if (!dsl_scan_is_running(scn))
1892                 return;
1893
1894         ds_destroyed_scn_phys(ds, &scn->scn_phys);
1895         ds_destroyed_scn_phys(ds, &scn->scn_phys_cached);
1896
1897         if (scan_ds_queue_contains(scn, ds->ds_object, &mintxg)) {
1898                 scan_ds_queue_remove(scn, ds->ds_object);
1899                 if (ds->ds_is_snapshot)
1900                         scan_ds_queue_insert(scn,
1901                             dsl_dataset_phys(ds)->ds_next_snap_obj, mintxg);
1902         }
1903
1904         if (zap_lookup_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj,
1905             ds->ds_object, &mintxg) == 0) {
1906                 ASSERT3U(dsl_dataset_phys(ds)->ds_num_children, <=, 1);
1907                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
1908                     scn->scn_phys.scn_queue_obj, ds->ds_object, tx));
1909                 if (ds->ds_is_snapshot) {
1910                         /*
1911                          * We keep the same mintxg; it could be >
1912                          * ds_creation_txg if the previous snapshot was
1913                          * deleted too.
1914                          */
1915                         VERIFY(zap_add_int_key(dp->dp_meta_objset,
1916                             scn->scn_phys.scn_queue_obj,
1917                             dsl_dataset_phys(ds)->ds_next_snap_obj,
1918                             mintxg, tx) == 0);
1919                         zfs_dbgmsg("destroying ds %llu; in queue; "
1920                             "replacing with %llu",
1921                             (u_longlong_t)ds->ds_object,
1922                             (u_longlong_t)dsl_dataset_phys(ds)->
1923                             ds_next_snap_obj);
1924                 } else {
1925                         zfs_dbgmsg("destroying ds %llu; in queue; removing",
1926                             (u_longlong_t)ds->ds_object);
1927                 }
1928         }
1929
1930         /*
1931          * dsl_scan_sync() should be called after this, and should sync
1932          * out our changed state, but just to be safe, do it here.
1933          */
1934         dsl_scan_sync_state(scn, tx, SYNC_CACHED);
1935 }
1936
1937 static void
1938 ds_snapshotted_bookmark(dsl_dataset_t *ds, zbookmark_phys_t *scn_bookmark)
1939 {
1940         if (scn_bookmark->zb_objset == ds->ds_object) {
1941                 scn_bookmark->zb_objset =
1942                     dsl_dataset_phys(ds)->ds_prev_snap_obj;
1943                 zfs_dbgmsg("snapshotting ds %llu; currently traversing; "
1944                     "reset zb_objset to %llu",
1945                     (u_longlong_t)ds->ds_object,
1946                     (u_longlong_t)dsl_dataset_phys(ds)->ds_prev_snap_obj);
1947         }
1948 }
1949
1950 /*
1951  * Called when a dataset is snapshotted. If we were currently traversing
1952  * this snapshot, we reset our bookmark to point at the newly created
1953  * snapshot. We also modify our work queue to remove the old snapshot and
1954  * replace with the new one.
1955  */
1956 void
1957 dsl_scan_ds_snapshotted(dsl_dataset_t *ds, dmu_tx_t *tx)
1958 {
1959         dsl_pool_t *dp = ds->ds_dir->dd_pool;
1960         dsl_scan_t *scn = dp->dp_scan;
1961         uint64_t mintxg;
1962
1963         if (!dsl_scan_is_running(scn))
1964                 return;
1965
1966         ASSERT(dsl_dataset_phys(ds)->ds_prev_snap_obj != 0);
1967
1968         ds_snapshotted_bookmark(ds, &scn->scn_phys.scn_bookmark);
1969         ds_snapshotted_bookmark(ds, &scn->scn_phys_cached.scn_bookmark);
1970
1971         if (scan_ds_queue_contains(scn, ds->ds_object, &mintxg)) {
1972                 scan_ds_queue_remove(scn, ds->ds_object);
1973                 scan_ds_queue_insert(scn,
1974                     dsl_dataset_phys(ds)->ds_prev_snap_obj, mintxg);
1975         }
1976
1977         if (zap_lookup_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj,
1978             ds->ds_object, &mintxg) == 0) {
1979                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
1980                     scn->scn_phys.scn_queue_obj, ds->ds_object, tx));
1981                 VERIFY(zap_add_int_key(dp->dp_meta_objset,
1982                     scn->scn_phys.scn_queue_obj,
1983                     dsl_dataset_phys(ds)->ds_prev_snap_obj, mintxg, tx) == 0);
1984                 zfs_dbgmsg("snapshotting ds %llu; in queue; "
1985                     "replacing with %llu",
1986                     (u_longlong_t)ds->ds_object,
1987                     (u_longlong_t)dsl_dataset_phys(ds)->ds_prev_snap_obj);
1988         }
1989
1990         dsl_scan_sync_state(scn, tx, SYNC_CACHED);
1991 }
1992
1993 static void
1994 ds_clone_swapped_bookmark(dsl_dataset_t *ds1, dsl_dataset_t *ds2,
1995     zbookmark_phys_t *scn_bookmark)
1996 {
1997         if (scn_bookmark->zb_objset == ds1->ds_object) {
1998                 scn_bookmark->zb_objset = ds2->ds_object;
1999                 zfs_dbgmsg("clone_swap ds %llu; currently traversing; "
2000                     "reset zb_objset to %llu",
2001                     (u_longlong_t)ds1->ds_object,
2002                     (u_longlong_t)ds2->ds_object);
2003         } else if (scn_bookmark->zb_objset == ds2->ds_object) {
2004                 scn_bookmark->zb_objset = ds1->ds_object;
2005                 zfs_dbgmsg("clone_swap ds %llu; currently traversing; "
2006                     "reset zb_objset to %llu",
2007                     (u_longlong_t)ds2->ds_object,
2008                     (u_longlong_t)ds1->ds_object);
2009         }
2010 }
2011
2012 /*
2013  * Called when a parent dataset and its clone are swapped. If we were
2014  * currently traversing the dataset, we need to switch to traversing the
2015  * newly promoted parent.
2016  */
2017 void
2018 dsl_scan_ds_clone_swapped(dsl_dataset_t *ds1, dsl_dataset_t *ds2, dmu_tx_t *tx)
2019 {
2020         dsl_pool_t *dp = ds1->ds_dir->dd_pool;
2021         dsl_scan_t *scn = dp->dp_scan;
2022         uint64_t mintxg;
2023
2024         if (!dsl_scan_is_running(scn))
2025                 return;
2026
2027         ds_clone_swapped_bookmark(ds1, ds2, &scn->scn_phys.scn_bookmark);
2028         ds_clone_swapped_bookmark(ds1, ds2, &scn->scn_phys_cached.scn_bookmark);
2029
2030         if (scan_ds_queue_contains(scn, ds1->ds_object, &mintxg)) {
2031                 scan_ds_queue_remove(scn, ds1->ds_object);
2032                 scan_ds_queue_insert(scn, ds2->ds_object, mintxg);
2033         }
2034         if (scan_ds_queue_contains(scn, ds2->ds_object, &mintxg)) {
2035                 scan_ds_queue_remove(scn, ds2->ds_object);
2036                 scan_ds_queue_insert(scn, ds1->ds_object, mintxg);
2037         }
2038
2039         if (zap_lookup_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj,
2040             ds1->ds_object, &mintxg) == 0) {
2041                 int err;
2042                 ASSERT3U(mintxg, ==, dsl_dataset_phys(ds1)->ds_prev_snap_txg);
2043                 ASSERT3U(mintxg, ==, dsl_dataset_phys(ds2)->ds_prev_snap_txg);
2044                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
2045                     scn->scn_phys.scn_queue_obj, ds1->ds_object, tx));
2046                 err = zap_add_int_key(dp->dp_meta_objset,
2047                     scn->scn_phys.scn_queue_obj, ds2->ds_object, mintxg, tx);
2048                 VERIFY(err == 0 || err == EEXIST);
2049                 if (err == EEXIST) {
2050                         /* Both were there to begin with */
2051                         VERIFY(0 == zap_add_int_key(dp->dp_meta_objset,
2052                             scn->scn_phys.scn_queue_obj,
2053                             ds1->ds_object, mintxg, tx));
2054                 }
2055                 zfs_dbgmsg("clone_swap ds %llu; in queue; "
2056                     "replacing with %llu",
2057                     (u_longlong_t)ds1->ds_object,
2058                     (u_longlong_t)ds2->ds_object);
2059         }
2060         if (zap_lookup_int_key(dp->dp_meta_objset, scn->scn_phys.scn_queue_obj,
2061             ds2->ds_object, &mintxg) == 0) {
2062                 ASSERT3U(mintxg, ==, dsl_dataset_phys(ds1)->ds_prev_snap_txg);
2063                 ASSERT3U(mintxg, ==, dsl_dataset_phys(ds2)->ds_prev_snap_txg);
2064                 VERIFY3U(0, ==, zap_remove_int(dp->dp_meta_objset,
2065                     scn->scn_phys.scn_queue_obj, ds2->ds_object, tx));
2066                 VERIFY(0 == zap_add_int_key(dp->dp_meta_objset,
2067                     scn->scn_phys.scn_queue_obj, ds1->ds_object, mintxg, tx));
2068                 zfs_dbgmsg("clone_swap ds %llu; in queue; "
2069                     "replacing with %llu",
2070                     (u_longlong_t)ds2->ds_object,
2071                     (u_longlong_t)ds1->ds_object);
2072         }
2073
2074         dsl_scan_sync_state(scn, tx, SYNC_CACHED);
2075 }
2076
2077 /* ARGSUSED */
2078 static int
2079 enqueue_clones_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
2080 {
2081         uint64_t originobj = *(uint64_t *)arg;
2082         dsl_dataset_t *ds;
2083         int err;
2084         dsl_scan_t *scn = dp->dp_scan;
2085
2086         if (dsl_dir_phys(hds->ds_dir)->dd_origin_obj != originobj)
2087                 return (0);
2088
2089         err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
2090         if (err)
2091                 return (err);
2092
2093         while (dsl_dataset_phys(ds)->ds_prev_snap_obj != originobj) {
2094                 dsl_dataset_t *prev;
2095                 err = dsl_dataset_hold_obj(dp,
2096                     dsl_dataset_phys(ds)->ds_prev_snap_obj, FTAG, &prev);
2097
2098                 dsl_dataset_rele(ds, FTAG);
2099                 if (err)
2100                         return (err);
2101                 ds = prev;
2102         }
2103         scan_ds_queue_insert(scn, ds->ds_object,
2104             dsl_dataset_phys(ds)->ds_prev_snap_txg);
2105         dsl_dataset_rele(ds, FTAG);
2106         return (0);
2107 }
2108
2109 static void
2110 dsl_scan_visitds(dsl_scan_t *scn, uint64_t dsobj, dmu_tx_t *tx)
2111 {
2112         dsl_pool_t *dp = scn->scn_dp;
2113         dsl_dataset_t *ds;
2114
2115         VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
2116
2117         if (scn->scn_phys.scn_cur_min_txg >=
2118             scn->scn_phys.scn_max_txg) {
2119                 /*
2120                  * This can happen if this snapshot was created after the
2121                  * scan started, and we already completed a previous snapshot
2122                  * that was created after the scan started.  This snapshot
2123                  * only references blocks with:
2124                  *
2125                  *      birth < our ds_creation_txg
2126                  *      cur_min_txg is no less than ds_creation_txg.
2127                  *      We have already visited these blocks.
2128                  * or
2129                  *      birth > scn_max_txg
2130                  *      The scan requested not to visit these blocks.
2131                  *
2132                  * Subsequent snapshots (and clones) can reference our
2133                  * blocks, or blocks with even higher birth times.
2134                  * Therefore we do not need to visit them either,
2135                  * so we do not add them to the work queue.
2136                  *
2137                  * Note that checking for cur_min_txg >= cur_max_txg
2138                  * is not sufficient, because in that case we may need to
2139                  * visit subsequent snapshots.  This happens when min_txg > 0,
2140                  * which raises cur_min_txg.  In this case we will visit
2141                  * this dataset but skip all of its blocks, because the
2142                  * rootbp's birth time is < cur_min_txg.  Then we will
2143                  * add the next snapshots/clones to the work queue.
2144                  */
2145                 char *dsname = kmem_alloc(MAXNAMELEN, KM_SLEEP);
2146                 dsl_dataset_name(ds, dsname);
2147                 zfs_dbgmsg("scanning dataset %llu (%s) is unnecessary because "
2148                     "cur_min_txg (%llu) >= max_txg (%llu)",
2149                     (longlong_t)dsobj, dsname,
2150                     (longlong_t)scn->scn_phys.scn_cur_min_txg,
2151                     (longlong_t)scn->scn_phys.scn_max_txg);
2152                 kmem_free(dsname, MAXNAMELEN);
2153
2154                 goto out;
2155         }
2156
2157         /*
2158          * Only the ZIL in the head (non-snapshot) is valid. Even though
2159          * snapshots can have ZIL block pointers (which may be the same
2160          * BP as in the head), they must be ignored. In addition, $ORIGIN
2161          * doesn't have a objset (i.e. its ds_bp is a hole) so we don't
2162          * need to look for a ZIL in it either. So we traverse the ZIL here,
2163          * rather than in scan_recurse(), because the regular snapshot
2164          * block-sharing rules don't apply to it.
2165          */
2166         if (DSL_SCAN_IS_SCRUB_RESILVER(scn) && !dsl_dataset_is_snapshot(ds) &&
2167             ds->ds_dir != dp->dp_origin_snap->ds_dir) {
2168                 objset_t *os;
2169                 if (dmu_objset_from_ds(ds, &os) != 0) {
2170                         goto out;
2171                 }
2172                 dsl_scan_zil(dp, &os->os_zil_header);
2173         }
2174
2175         /*
2176          * Iterate over the bps in this ds.
2177          */
2178         dmu_buf_will_dirty(ds->ds_dbuf, tx);
2179         rrw_enter(&ds->ds_bp_rwlock, RW_READER, FTAG);
2180         dsl_scan_visit_rootbp(scn, ds, &dsl_dataset_phys(ds)->ds_bp, tx);
2181         rrw_exit(&ds->ds_bp_rwlock, FTAG);
2182
2183         char *dsname = kmem_alloc(ZFS_MAX_DATASET_NAME_LEN, KM_SLEEP);
2184         dsl_dataset_name(ds, dsname);
2185         zfs_dbgmsg("scanned dataset %llu (%s) with min=%llu max=%llu; "
2186             "suspending=%u",
2187             (longlong_t)dsobj, dsname,
2188             (longlong_t)scn->scn_phys.scn_cur_min_txg,
2189             (longlong_t)scn->scn_phys.scn_cur_max_txg,
2190             (int)scn->scn_suspending);
2191         kmem_free(dsname, ZFS_MAX_DATASET_NAME_LEN);
2192
2193         if (scn->scn_suspending)
2194                 goto out;
2195
2196         /*
2197          * We've finished this pass over this dataset.
2198          */
2199
2200         /*
2201          * If we did not completely visit this dataset, do another pass.
2202          */
2203         if (scn->scn_phys.scn_flags & DSF_VISIT_DS_AGAIN) {
2204                 zfs_dbgmsg("incomplete pass; visiting again");
2205                 scn->scn_phys.scn_flags &= ~DSF_VISIT_DS_AGAIN;
2206                 scan_ds_queue_insert(scn, ds->ds_object,
2207                     scn->scn_phys.scn_cur_max_txg);
2208                 goto out;
2209         }
2210
2211         /*
2212          * Add descendent datasets to work queue.
2213          */
2214         if (dsl_dataset_phys(ds)->ds_next_snap_obj != 0) {
2215                 scan_ds_queue_insert(scn,
2216                     dsl_dataset_phys(ds)->ds_next_snap_obj,
2217                     dsl_dataset_phys(ds)->ds_creation_txg);
2218         }
2219         if (dsl_dataset_phys(ds)->ds_num_children > 1) {
2220                 boolean_t usenext = B_FALSE;
2221                 if (dsl_dataset_phys(ds)->ds_next_clones_obj != 0) {
2222                         uint64_t count;
2223                         /*
2224                          * A bug in a previous version of the code could
2225                          * cause upgrade_clones_cb() to not set
2226                          * ds_next_snap_obj when it should, leading to a
2227                          * missing entry.  Therefore we can only use the
2228                          * next_clones_obj when its count is correct.
2229                          */
2230                         int err = zap_count(dp->dp_meta_objset,
2231                             dsl_dataset_phys(ds)->ds_next_clones_obj, &count);
2232                         if (err == 0 &&
2233                             count == dsl_dataset_phys(ds)->ds_num_children - 1)
2234                                 usenext = B_TRUE;
2235                 }
2236
2237                 if (usenext) {
2238                         zap_cursor_t zc;
2239                         zap_attribute_t za;
2240                         for (zap_cursor_init(&zc, dp->dp_meta_objset,
2241                             dsl_dataset_phys(ds)->ds_next_clones_obj);
2242                             zap_cursor_retrieve(&zc, &za) == 0;
2243                             (void) zap_cursor_advance(&zc)) {
2244                                 scan_ds_queue_insert(scn,
2245                                     zfs_strtonum(za.za_name, NULL),
2246                                     dsl_dataset_phys(ds)->ds_creation_txg);
2247                         }
2248                         zap_cursor_fini(&zc);
2249                 } else {
2250                         VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
2251                             enqueue_clones_cb, &ds->ds_object,
2252                             DS_FIND_CHILDREN));
2253                 }
2254         }
2255
2256 out:
2257         dsl_dataset_rele(ds, FTAG);
2258 }
2259
2260 /* ARGSUSED */
2261 static int
2262 enqueue_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
2263 {
2264         dsl_dataset_t *ds;
2265         int err;
2266         dsl_scan_t *scn = dp->dp_scan;
2267
2268         err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
2269         if (err)
2270                 return (err);
2271
2272         while (dsl_dataset_phys(ds)->ds_prev_snap_obj != 0) {
2273                 dsl_dataset_t *prev;
2274                 err = dsl_dataset_hold_obj(dp,
2275                     dsl_dataset_phys(ds)->ds_prev_snap_obj, FTAG, &prev);
2276                 if (err) {
2277                         dsl_dataset_rele(ds, FTAG);
2278                         return (err);
2279                 }
2280
2281                 /*
2282                  * If this is a clone, we don't need to worry about it for now.
2283                  */
2284                 if (dsl_dataset_phys(prev)->ds_next_snap_obj != ds->ds_object) {
2285                         dsl_dataset_rele(ds, FTAG);
2286                         dsl_dataset_rele(prev, FTAG);
2287                         return (0);
2288                 }
2289                 dsl_dataset_rele(ds, FTAG);
2290                 ds = prev;
2291         }
2292
2293         scan_ds_queue_insert(scn, ds->ds_object,
2294             dsl_dataset_phys(ds)->ds_prev_snap_txg);
2295         dsl_dataset_rele(ds, FTAG);
2296         return (0);
2297 }
2298
2299 /* ARGSUSED */
2300 void
2301 dsl_scan_ddt_entry(dsl_scan_t *scn, enum zio_checksum checksum,
2302     ddt_entry_t *dde, dmu_tx_t *tx)
2303 {
2304         const ddt_key_t *ddk = &dde->dde_key;
2305         ddt_phys_t *ddp = dde->dde_phys;
2306         blkptr_t bp;
2307         zbookmark_phys_t zb = { 0 };
2308         int p;
2309
2310         if (scn->scn_phys.scn_state != DSS_SCANNING)
2311                 return;
2312
2313         for (p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
2314                 if (ddp->ddp_phys_birth == 0 ||
2315                     ddp->ddp_phys_birth > scn->scn_phys.scn_max_txg)
2316                         continue;
2317                 ddt_bp_create(checksum, ddk, ddp, &bp);
2318
2319                 scn->scn_visited_this_txg++;
2320                 scan_funcs[scn->scn_phys.scn_func](scn->scn_dp, &bp, &zb);
2321         }
2322 }
2323
2324 /*
2325  * Scrub/dedup interaction.
2326  *
2327  * If there are N references to a deduped block, we don't want to scrub it
2328  * N times -- ideally, we should scrub it exactly once.
2329  *
2330  * We leverage the fact that the dde's replication class (enum ddt_class)
2331  * is ordered from highest replication class (DDT_CLASS_DITTO) to lowest
2332  * (DDT_CLASS_UNIQUE) so that we may walk the DDT in that order.
2333  *
2334  * To prevent excess scrubbing, the scrub begins by walking the DDT
2335  * to find all blocks with refcnt > 1, and scrubs each of these once.
2336  * Since there are two replication classes which contain blocks with
2337  * refcnt > 1, we scrub the highest replication class (DDT_CLASS_DITTO) first.
2338  * Finally the top-down scrub begins, only visiting blocks with refcnt == 1.
2339  *
2340  * There would be nothing more to say if a block's refcnt couldn't change
2341  * during a scrub, but of course it can so we must account for changes
2342  * in a block's replication class.
2343  *
2344  * Here's an example of what can occur:
2345  *
2346  * If a block has refcnt > 1 during the DDT scrub phase, but has refcnt == 1
2347  * when visited during the top-down scrub phase, it will be scrubbed twice.
2348  * This negates our scrub optimization, but is otherwise harmless.
2349  *
2350  * If a block has refcnt == 1 during the DDT scrub phase, but has refcnt > 1
2351  * on each visit during the top-down scrub phase, it will never be scrubbed.
2352  * To catch this, ddt_sync_entry() notifies the scrub code whenever a block's
2353  * reference class transitions to a higher level (i.e DDT_CLASS_UNIQUE to
2354  * DDT_CLASS_DUPLICATE); if it transitions from refcnt == 1 to refcnt > 1
2355  * while a scrub is in progress, it scrubs the block right then.
2356  */
2357 static void
2358 dsl_scan_ddt(dsl_scan_t *scn, dmu_tx_t *tx)
2359 {
2360         ddt_bookmark_t *ddb = &scn->scn_phys.scn_ddt_bookmark;
2361         ddt_entry_t dde = { 0 };
2362         int error;
2363         uint64_t n = 0;
2364
2365         while ((error = ddt_walk(scn->scn_dp->dp_spa, ddb, &dde)) == 0) {
2366                 ddt_t *ddt;
2367
2368                 if (ddb->ddb_class > scn->scn_phys.scn_ddt_class_max)
2369                         break;
2370                 dprintf("visiting ddb=%llu/%llu/%llu/%llx\n",
2371                     (longlong_t)ddb->ddb_class,
2372                     (longlong_t)ddb->ddb_type,
2373                     (longlong_t)ddb->ddb_checksum,
2374                     (longlong_t)ddb->ddb_cursor);
2375
2376                 /* There should be no pending changes to the dedup table */
2377                 ddt = scn->scn_dp->dp_spa->spa_ddt[ddb->ddb_checksum];
2378                 ASSERT(avl_first(&ddt->ddt_tree) == NULL);
2379
2380                 dsl_scan_ddt_entry(scn, ddb->ddb_checksum, &dde, tx);
2381                 n++;
2382
2383                 if (dsl_scan_check_suspend(scn, NULL))
2384                         break;
2385         }
2386
2387         zfs_dbgmsg("scanned %llu ddt entries with class_max = %u; "
2388             "suspending=%u", (longlong_t)n,
2389             (int)scn->scn_phys.scn_ddt_class_max, (int)scn->scn_suspending);
2390
2391         ASSERT(error == 0 || error == ENOENT);
2392         ASSERT(error != ENOENT ||
2393             ddb->ddb_class > scn->scn_phys.scn_ddt_class_max);
2394 }
2395
2396 static uint64_t
2397 dsl_scan_ds_maxtxg(dsl_dataset_t *ds)
2398 {
2399         uint64_t smt = ds->ds_dir->dd_pool->dp_scan->scn_phys.scn_max_txg;
2400         if (ds->ds_is_snapshot)
2401                 return (MIN(smt, dsl_dataset_phys(ds)->ds_creation_txg));
2402         return (smt);
2403 }
2404
2405 static void
2406 dsl_scan_visit(dsl_scan_t *scn, dmu_tx_t *tx)
2407 {
2408         scan_ds_t *sds;
2409         dsl_pool_t *dp = scn->scn_dp;
2410
2411         if (scn->scn_phys.scn_ddt_bookmark.ddb_class <=
2412             scn->scn_phys.scn_ddt_class_max) {
2413                 scn->scn_phys.scn_cur_min_txg = scn->scn_phys.scn_min_txg;
2414                 scn->scn_phys.scn_cur_max_txg = scn->scn_phys.scn_max_txg;
2415                 dsl_scan_ddt(scn, tx);
2416                 if (scn->scn_suspending)
2417                         return;
2418         }
2419
2420         if (scn->scn_phys.scn_bookmark.zb_objset == DMU_META_OBJSET) {
2421                 /* First do the MOS & ORIGIN */
2422
2423                 scn->scn_phys.scn_cur_min_txg = scn->scn_phys.scn_min_txg;
2424                 scn->scn_phys.scn_cur_max_txg = scn->scn_phys.scn_max_txg;
2425                 dsl_scan_visit_rootbp(scn, NULL,
2426                     &dp->dp_meta_rootbp, tx);
2427                 spa_set_rootblkptr(dp->dp_spa, &dp->dp_meta_rootbp);
2428                 if (scn->scn_suspending)
2429                         return;
2430
2431                 if (spa_version(dp->dp_spa) < SPA_VERSION_DSL_SCRUB) {
2432                         VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
2433                             enqueue_cb, NULL, DS_FIND_CHILDREN));
2434                 } else {
2435                         dsl_scan_visitds(scn,
2436                             dp->dp_origin_snap->ds_object, tx);
2437                 }
2438                 ASSERT(!scn->scn_suspending);
2439         } else if (scn->scn_phys.scn_bookmark.zb_objset !=
2440             ZB_DESTROYED_OBJSET) {
2441                 uint64_t dsobj = scn->scn_phys.scn_bookmark.zb_objset;
2442                 /*
2443                  * If we were suspended, continue from here. Note if the
2444                  * ds we were suspended on was deleted, the zb_objset may
2445                  * be -1, so we will skip this and find a new objset
2446                  * below.
2447                  */
2448                 dsl_scan_visitds(scn, dsobj, tx);
2449                 if (scn->scn_suspending)
2450                         return;
2451         }
2452
2453         /*
2454          * In case we suspended right at the end of the ds, zero the
2455          * bookmark so we don't think that we're still trying to resume.
2456          */
2457         bzero(&scn->scn_phys.scn_bookmark, sizeof (zbookmark_phys_t));
2458
2459         /*
2460          * Keep pulling things out of the dataset avl queue. Updates to the
2461          * persistent zap-object-as-queue happen only at checkpoints.
2462          */
2463         while ((sds = avl_first(&scn->scn_queue)) != NULL) {
2464                 dsl_dataset_t *ds;
2465                 uint64_t dsobj = sds->sds_dsobj;
2466                 uint64_t txg = sds->sds_txg;
2467
2468                 /* dequeue and free the ds from the queue */
2469                 scan_ds_queue_remove(scn, dsobj);
2470                 sds = NULL;     /* must not be touched after removal */
2471
2472                 /* Set up min / max txg */
2473                 VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
2474                 if (txg != 0) {
2475                         scn->scn_phys.scn_cur_min_txg =
2476                             MAX(scn->scn_phys.scn_min_txg, txg);
2477                 } else {
2478                         scn->scn_phys.scn_cur_min_txg =
2479                             MAX(scn->scn_phys.scn_min_txg,
2480                             dsl_dataset_phys(ds)->ds_prev_snap_txg);
2481                 }
2482                 scn->scn_phys.scn_cur_max_txg = dsl_scan_ds_maxtxg(ds);
2483                 dsl_dataset_rele(ds, FTAG);
2484
2485                 dsl_scan_visitds(scn, dsobj, tx);
2486                 if (scn->scn_suspending)
2487                         return;
2488         }
2489         /* No more objsets to fetch, we're done */
2490         scn->scn_phys.scn_bookmark.zb_objset = ZB_DESTROYED_OBJSET;
2491         ASSERT0(scn->scn_suspending);
2492 }
2493
2494 static uint64_t
2495 dsl_scan_count_leaves(vdev_t *vd)
2496 {
2497         uint64_t i, leaves = 0;
2498
2499         /* we only count leaves that belong to the main pool and are readable */
2500         if (vd->vdev_islog || vd->vdev_isspare ||
2501             vd->vdev_isl2cache || !vdev_readable(vd))
2502                 return (0);
2503
2504         if (vd->vdev_ops->vdev_op_leaf)
2505                 return (1);
2506
2507         for (i = 0; i < vd->vdev_children; i++) {
2508                 leaves += dsl_scan_count_leaves(vd->vdev_child[i]);
2509         }
2510
2511         return (leaves);
2512 }
2513
2514
2515 static void
2516 scan_io_queues_update_zio_stats(dsl_scan_io_queue_t *q, const blkptr_t *bp)
2517 {
2518         int i;
2519         uint64_t cur_size = 0;
2520
2521         for (i = 0; i < BP_GET_NDVAS(bp); i++) {
2522                 cur_size += DVA_GET_ASIZE(&bp->blk_dva[i]);
2523         }
2524
2525         q->q_total_zio_size_this_txg += cur_size;
2526         q->q_zios_this_txg++;
2527 }
2528
2529 static void
2530 scan_io_queues_update_seg_stats(dsl_scan_io_queue_t *q, uint64_t start,
2531     uint64_t end)
2532 {
2533         q->q_total_seg_size_this_txg += end - start;
2534         q->q_segs_this_txg++;
2535 }
2536
2537 static boolean_t
2538 scan_io_queue_check_suspend(dsl_scan_t *scn)
2539 {
2540         /* See comment in dsl_scan_check_suspend() */
2541         uint64_t curr_time_ns = gethrtime();
2542         uint64_t scan_time_ns = curr_time_ns - scn->scn_sync_start_time;
2543         uint64_t sync_time_ns = curr_time_ns -
2544             scn->scn_dp->dp_spa->spa_sync_starttime;
2545         int dirty_pct = scn->scn_dp->dp_dirty_total * 100 / zfs_dirty_data_max;
2546         int mintime = (scn->scn_phys.scn_func == POOL_SCAN_RESILVER) ?
2547             zfs_resilver_min_time_ms : zfs_scrub_min_time_ms;
2548
2549         return ((NSEC2MSEC(scan_time_ns) > mintime &&
2550             (dirty_pct >= zfs_vdev_async_write_active_min_dirty_percent ||
2551             txg_sync_waiting(scn->scn_dp) ||
2552             NSEC2SEC(sync_time_ns) >= zfs_txg_timeout)) ||
2553             spa_shutting_down(scn->scn_dp->dp_spa));
2554 }
2555
2556 /*
2557  * Given a list of scan_io_t's in io_list, this issues the io's out to
2558  * disk. This consumes the io_list and frees the scan_io_t's. This is
2559  * called when emptying queues, either when we're up against the memory
2560  * limit or when we have finished scanning. Returns B_TRUE if we stopped
2561  * processing the list before we finished. Any zios that were not issued
2562  * will remain in the io_list.
2563  */
2564 static boolean_t
2565 scan_io_queue_issue(dsl_scan_io_queue_t *queue, list_t *io_list)
2566 {
2567         dsl_scan_t *scn = queue->q_scn;
2568         scan_io_t *sio;
2569         int64_t bytes_issued = 0;
2570         boolean_t suspended = B_FALSE;
2571
2572         while ((sio = list_head(io_list)) != NULL) {
2573                 blkptr_t bp;
2574
2575                 if (scan_io_queue_check_suspend(scn)) {
2576                         suspended = B_TRUE;
2577                         break;
2578                 }
2579
2580                 sio2bp(sio, &bp, queue->q_vd->vdev_id);
2581                 bytes_issued += sio->sio_asize;
2582                 scan_exec_io(scn->scn_dp, &bp, sio->sio_flags,
2583                     &sio->sio_zb, queue);
2584                 (void) list_remove_head(io_list);
2585                 scan_io_queues_update_zio_stats(queue, &bp);
2586                 kmem_free(sio, sizeof (*sio));
2587         }
2588
2589         atomic_add_64(&scn->scn_bytes_pending, -bytes_issued);
2590
2591         return (suspended);
2592 }
2593
2594 /*
2595  * Given a range_seg_t (extent) and a list, this function passes over a
2596  * scan queue and gathers up the appropriate ios which fit into that
2597  * scan seg (starting from lowest LBA). At the end, we remove the segment
2598  * from the q_exts_by_addr range tree.
2599  */
2600 static boolean_t
2601 scan_io_queue_gather(dsl_scan_io_queue_t *queue, range_seg_t *rs, list_t *list)
2602 {
2603         scan_io_t srch_sio, *sio, *next_sio;
2604         avl_index_t idx;
2605         uint_t num_sios = 0;
2606         int64_t bytes_issued = 0;
2607
2608         ASSERT(rs != NULL);
2609         ASSERT(MUTEX_HELD(&queue->q_vd->vdev_scan_io_queue_lock));
2610
2611         srch_sio.sio_offset = rs->rs_start;
2612
2613         /*
2614          * The exact start of the extent might not contain any matching zios,
2615          * so if that's the case, examine the next one in the tree.
2616          */
2617         sio = avl_find(&queue->q_sios_by_addr, &srch_sio, &idx);
2618         if (sio == NULL)
2619                 sio = avl_nearest(&queue->q_sios_by_addr, idx, AVL_AFTER);
2620
2621         while (sio != NULL && sio->sio_offset < rs->rs_end && num_sios <= 32) {
2622                 ASSERT3U(sio->sio_offset, >=, rs->rs_start);
2623                 ASSERT3U(sio->sio_offset + sio->sio_asize, <=, rs->rs_end);
2624
2625                 next_sio = AVL_NEXT(&queue->q_sios_by_addr, sio);
2626                 avl_remove(&queue->q_sios_by_addr, sio);
2627
2628                 bytes_issued += sio->sio_asize;
2629                 num_sios++;
2630                 list_insert_tail(list, sio);
2631                 sio = next_sio;
2632         }
2633
2634         /*
2635          * We limit the number of sios we process at once to 32 to avoid
2636          * biting off more than we can chew. If we didn't take everything
2637          * in the segment we update it to reflect the work we were able to
2638          * complete. Otherwise, we remove it from the range tree entirely.
2639          */
2640         if (sio != NULL && sio->sio_offset < rs->rs_end) {
2641                 range_tree_adjust_fill(queue->q_exts_by_addr, rs,
2642                     -bytes_issued);
2643                 range_tree_resize_segment(queue->q_exts_by_addr, rs,
2644                     sio->sio_offset, rs->rs_end - sio->sio_offset);
2645
2646                 return (B_TRUE);
2647         } else {
2648                 range_tree_remove(queue->q_exts_by_addr, rs->rs_start,
2649                     rs->rs_end - rs->rs_start);
2650                 return (B_FALSE);
2651         }
2652 }
2653
2654
2655 /*
2656  * This is called from the queue emptying thread and selects the next
2657  * extent from which we are to issue io's. The behavior of this function
2658  * depends on the state of the scan, the current memory consumption and
2659  * whether or not we are performing a scan shutdown.
2660  * 1) We select extents in an elevator algorithm (LBA-order) if the scan
2661  *      needs to perform a checkpoint
2662  * 2) We select the largest available extent if we are up against the
2663  *      memory limit.
2664  * 3) Otherwise we don't select any extents.
2665  */
2666 static const range_seg_t *
2667 scan_io_queue_fetch_ext(dsl_scan_io_queue_t *queue)
2668 {
2669         dsl_scan_t *scn = queue->q_scn;
2670
2671         ASSERT(MUTEX_HELD(&queue->q_vd->vdev_scan_io_queue_lock));
2672         ASSERT(scn->scn_is_sorted);
2673
2674         /* handle tunable overrides */
2675         if (scn->scn_checkpointing || scn->scn_clearing) {
2676                 if (zfs_scan_issue_strategy == 1) {
2677                         return (range_tree_first(queue->q_exts_by_addr));
2678                 } else if (zfs_scan_issue_strategy == 2) {
2679                         return (avl_first(&queue->q_exts_by_size));
2680                 }
2681         }
2682
2683         /*
2684          * During normal clearing, we want to issue our largest segments
2685          * first, keeping IO as sequential as possible, and leaving the
2686          * smaller extents for later with the hope that they might eventually
2687          * grow to larger sequential segments. However, when the scan is
2688          * checkpointing, no new extents will be added to the sorting queue,
2689          * so the way we are sorted now is as good as it will ever get.
2690          * In this case, we instead switch to issuing extents in LBA order.
2691          */
2692         if (scn->scn_checkpointing) {
2693                 return (range_tree_first(queue->q_exts_by_addr));
2694         } else if (scn->scn_clearing) {
2695                 return (avl_first(&queue->q_exts_by_size));
2696         } else {
2697                 return (NULL);
2698         }
2699 }
2700
2701 static void
2702 scan_io_queues_run_one(void *arg)
2703 {
2704         dsl_scan_io_queue_t *queue = arg;
2705         kmutex_t *q_lock = &queue->q_vd->vdev_scan_io_queue_lock;
2706         boolean_t suspended = B_FALSE;
2707         range_seg_t *rs = NULL;
2708         scan_io_t *sio = NULL;
2709         list_t sio_list;
2710         uint64_t bytes_per_leaf = zfs_scan_vdev_limit;
2711         uint64_t nr_leaves = dsl_scan_count_leaves(queue->q_vd);
2712
2713         ASSERT(queue->q_scn->scn_is_sorted);
2714
2715         list_create(&sio_list, sizeof (scan_io_t),
2716             offsetof(scan_io_t, sio_nodes.sio_list_node));
2717         mutex_enter(q_lock);
2718
2719         /* calculate maximum in-flight bytes for this txg (min 1MB) */
2720         queue->q_maxinflight_bytes =
2721             MAX(nr_leaves * bytes_per_leaf, 1ULL << 20);
2722
2723         /* reset per-queue scan statistics for this txg */
2724         queue->q_total_seg_size_this_txg = 0;
2725         queue->q_segs_this_txg = 0;
2726         queue->q_total_zio_size_this_txg = 0;
2727         queue->q_zios_this_txg = 0;
2728
2729         /* loop until we have run out of time or sios */
2730         while ((rs = (range_seg_t*)scan_io_queue_fetch_ext(queue)) != NULL) {
2731                 uint64_t seg_start = 0, seg_end = 0;
2732                 boolean_t more_left = B_TRUE;
2733
2734                 ASSERT(list_is_empty(&sio_list));
2735
2736                 /* loop while we still have sios left to process in this rs */
2737                 while (more_left) {
2738                         scan_io_t *first_sio, *last_sio;
2739
2740                         /*
2741                          * We have selected which extent needs to be
2742                          * processed next. Gather up the corresponding sios.
2743                          */
2744                         more_left = scan_io_queue_gather(queue, rs, &sio_list);
2745                         ASSERT(!list_is_empty(&sio_list));
2746                         first_sio = list_head(&sio_list);
2747                         last_sio = list_tail(&sio_list);
2748
2749                         seg_end = last_sio->sio_offset + last_sio->sio_asize;
2750                         if (seg_start == 0)
2751                                 seg_start = first_sio->sio_offset;
2752
2753                         /*
2754                          * Issuing sios can take a long time so drop the
2755                          * queue lock. The sio queue won't be updated by
2756                          * other threads since we're in syncing context so
2757                          * we can be sure that our trees will remain exactly
2758                          * as we left them.
2759                          */
2760                         mutex_exit(q_lock);
2761                         suspended = scan_io_queue_issue(queue, &sio_list);
2762                         mutex_enter(q_lock);
2763
2764                         if (suspended)
2765                                 break;
2766                 }
2767                 /* update statistics for debugging purposes */
2768                 scan_io_queues_update_seg_stats(queue, seg_start, seg_end);
2769
2770                 if (suspended)
2771                         break;
2772         }
2773
2774
2775         /* If we were suspended in the middle of processing,
2776          * requeue any unfinished sios and exit.
2777          */
2778         while ((sio = list_head(&sio_list)) != NULL) {
2779                 list_remove(&sio_list, sio);
2780                 scan_io_queue_insert_impl(queue, sio);
2781         }
2782
2783         mutex_exit(q_lock);
2784         list_destroy(&sio_list);
2785 }
2786
2787 /*
2788  * Performs an emptying run on all scan queues in the pool. This just
2789  * punches out one thread per top-level vdev, each of which processes
2790  * only that vdev's scan queue. We can parallelize the I/O here because
2791  * we know that each queue's io's only affect its own top-level vdev.
2792  *
2793  * This function waits for the queue runs to complete, and must be
2794  * called from dsl_scan_sync (or in general, syncing context).
2795  */
2796 static void
2797 scan_io_queues_run(dsl_scan_t *scn)
2798 {
2799         spa_t *spa = scn->scn_dp->dp_spa;
2800
2801         ASSERT(scn->scn_is_sorted);
2802         ASSERT(spa_config_held(spa, SCL_CONFIG, RW_READER));
2803
2804         if (scn->scn_bytes_pending == 0)
2805                 return;
2806
2807         if (scn->scn_taskq == NULL) {
2808                 char *tq_name = kmem_zalloc(ZFS_MAX_DATASET_NAME_LEN + 16,
2809                     KM_SLEEP);
2810                 int nthreads = spa->spa_root_vdev->vdev_children;
2811
2812                 /*
2813                  * We need to make this taskq *always* execute as many
2814                  * threads in parallel as we have top-level vdevs and no
2815                  * less, otherwise strange serialization of the calls to
2816                  * scan_io_queues_run_one can occur during spa_sync runs
2817                  * and that significantly impacts performance.
2818                  */
2819                 (void) snprintf(tq_name, ZFS_MAX_DATASET_NAME_LEN + 16,
2820                     "dsl_scan_tq_%s", spa->spa_name);
2821                 scn->scn_taskq = taskq_create(tq_name, nthreads, minclsyspri,
2822                     nthreads, nthreads, TASKQ_PREPOPULATE);
2823                 kmem_free(tq_name, ZFS_MAX_DATASET_NAME_LEN + 16);
2824         }
2825
2826         for (uint64_t i = 0; i < spa->spa_root_vdev->vdev_children; i++) {
2827                 vdev_t *vd = spa->spa_root_vdev->vdev_child[i];
2828
2829                 mutex_enter(&vd->vdev_scan_io_queue_lock);
2830                 if (vd->vdev_scan_io_queue != NULL) {
2831                         VERIFY(taskq_dispatch(scn->scn_taskq,
2832                             scan_io_queues_run_one, vd->vdev_scan_io_queue,
2833                             TQ_SLEEP) != TASKQID_INVALID);
2834                 }
2835                 mutex_exit(&vd->vdev_scan_io_queue_lock);
2836         }
2837
2838         /*
2839          * Wait for the queues to finish issuing thir IOs for this run
2840          * before we return. There may still be IOs in flight at this
2841          * point.
2842          */
2843         taskq_wait(scn->scn_taskq);
2844 }
2845
2846 static boolean_t
2847 dsl_scan_async_block_should_pause(dsl_scan_t *scn)
2848 {
2849         uint64_t elapsed_nanosecs;
2850
2851         if (zfs_recover)
2852                 return (B_FALSE);
2853
2854         if (scn->scn_visited_this_txg >= zfs_async_block_max_blocks)
2855                 return (B_TRUE);
2856
2857         elapsed_nanosecs = gethrtime() - scn->scn_sync_start_time;
2858         return (elapsed_nanosecs / NANOSEC > zfs_txg_timeout ||
2859             (NSEC2MSEC(elapsed_nanosecs) > scn->scn_async_block_min_time_ms &&
2860             txg_sync_waiting(scn->scn_dp)) ||
2861             spa_shutting_down(scn->scn_dp->dp_spa));
2862 }
2863
2864 static int
2865 dsl_scan_free_block_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
2866 {
2867         dsl_scan_t *scn = arg;
2868
2869         if (!scn->scn_is_bptree ||
2870             (BP_GET_LEVEL(bp) == 0 && BP_GET_TYPE(bp) != DMU_OT_OBJSET)) {
2871                 if (dsl_scan_async_block_should_pause(scn))
2872                         return (SET_ERROR(ERESTART));
2873         }
2874
2875         zio_nowait(zio_free_sync(scn->scn_zio_root, scn->scn_dp->dp_spa,
2876             dmu_tx_get_txg(tx), bp, BP_GET_PSIZE(bp), 0));
2877         dsl_dir_diduse_space(tx->tx_pool->dp_free_dir, DD_USED_HEAD,
2878             -bp_get_dsize_sync(scn->scn_dp->dp_spa, bp),
2879             -BP_GET_PSIZE(bp), -BP_GET_UCSIZE(bp), tx);
2880         scn->scn_visited_this_txg++;
2881         return (0);
2882 }
2883
2884 static void
2885 dsl_scan_update_stats(dsl_scan_t *scn)
2886 {
2887         spa_t *spa = scn->scn_dp->dp_spa;
2888         uint64_t i;
2889         uint64_t seg_size_total = 0, zio_size_total = 0;
2890         uint64_t seg_count_total = 0, zio_count_total = 0;
2891
2892         for (i = 0; i < spa->spa_root_vdev->vdev_children; i++) {
2893                 vdev_t *vd = spa->spa_root_vdev->vdev_child[i];
2894                 dsl_scan_io_queue_t *queue = vd->vdev_scan_io_queue;
2895
2896                 if (queue == NULL)
2897                         continue;
2898
2899                 seg_size_total += queue->q_total_seg_size_this_txg;
2900                 zio_size_total += queue->q_total_zio_size_this_txg;
2901                 seg_count_total += queue->q_segs_this_txg;
2902                 zio_count_total += queue->q_zios_this_txg;
2903         }
2904
2905         if (seg_count_total == 0 || zio_count_total == 0) {
2906                 scn->scn_avg_seg_size_this_txg = 0;
2907                 scn->scn_avg_zio_size_this_txg = 0;
2908                 scn->scn_segs_this_txg = 0;
2909                 scn->scn_zios_this_txg = 0;
2910                 return;
2911         }
2912
2913         scn->scn_avg_seg_size_this_txg = seg_size_total / seg_count_total;
2914         scn->scn_avg_zio_size_this_txg = zio_size_total / zio_count_total;
2915         scn->scn_segs_this_txg = seg_count_total;
2916         scn->scn_zios_this_txg = zio_count_total;
2917 }
2918
2919 static int
2920 dsl_scan_obsolete_block_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
2921 {
2922         dsl_scan_t *scn = arg;
2923         const dva_t *dva = &bp->blk_dva[0];
2924
2925         if (dsl_scan_async_block_should_pause(scn))
2926                 return (SET_ERROR(ERESTART));
2927
2928         spa_vdev_indirect_mark_obsolete(scn->scn_dp->dp_spa,
2929             DVA_GET_VDEV(dva), DVA_GET_OFFSET(dva),
2930             DVA_GET_ASIZE(dva), tx);
2931         scn->scn_visited_this_txg++;
2932         return (0);
2933 }
2934
2935 boolean_t
2936 dsl_scan_active(dsl_scan_t *scn)
2937 {
2938         spa_t *spa = scn->scn_dp->dp_spa;
2939         uint64_t used = 0, comp, uncomp;
2940
2941         if (spa->spa_load_state != SPA_LOAD_NONE)
2942                 return (B_FALSE);
2943         if (spa_shutting_down(spa))
2944                 return (B_FALSE);
2945         if ((dsl_scan_is_running(scn) && !dsl_scan_is_paused_scrub(scn)) ||
2946             (scn->scn_async_destroying && !scn->scn_async_stalled))
2947                 return (B_TRUE);
2948
2949         if (spa_version(scn->scn_dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
2950                 (void) bpobj_space(&scn->scn_dp->dp_free_bpobj,
2951                     &used, &comp, &uncomp);
2952         }
2953         return (used != 0);
2954 }
2955
2956 static boolean_t
2957 dsl_scan_need_resilver(spa_t *spa, const dva_t *dva, size_t psize,
2958     uint64_t phys_birth)
2959 {
2960         vdev_t *vd;
2961
2962         if (vd->vdev_ops == &vdev_indirect_ops) {
2963                 /*
2964                  * The indirect vdev can point to multiple
2965                  * vdevs.  For simplicity, always create
2966                  * the resilver zio_t. zio_vdev_io_start()
2967                  * will bypass the child resilver i/o's if
2968                  * they are on vdevs that don't have DTL's.
2969                  */
2970                 return (B_TRUE);
2971         }
2972         if (DVA_GET_GANG(dva)) {
2973                 /*
2974                  * Gang members may be spread across multiple
2975                  * vdevs, so the best estimate we have is the
2976                  * scrub range, which has already been checked.
2977                  * XXX -- it would be better to change our
2978                  * allocation policy to ensure that all
2979                  * gang members reside on the same vdev.
2980                  */
2981                 return (B_TRUE);
2982         }
2983
2984         vd = vdev_lookup_top(spa, DVA_GET_VDEV(dva));
2985
2986         /*
2987          * Check if the txg falls within the range which must be
2988          * resilvered.  DVAs outside this range can always be skipped.
2989          */
2990         if (!vdev_dtl_contains(vd, DTL_PARTIAL, phys_birth, 1))
2991                 return (B_FALSE);
2992
2993         /*
2994          * Check if the top-level vdev must resilver this offset.
2995          * When the offset does not intersect with a dirty leaf DTL
2996          * then it may be possible to skip the resilver IO.  The psize
2997          * is provided instead of asize to simplify the check for RAIDZ.
2998          */
2999         if (!vdev_dtl_need_resilver(vd, DVA_GET_OFFSET(dva), psize))
3000                 return (B_FALSE);
3001
3002         return (B_TRUE);
3003 }
3004
3005 static int
3006 dsl_process_async_destroys(dsl_pool_t *dp, dmu_tx_t *tx)
3007 {
3008         int err = 0;
3009         dsl_scan_t *scn = dp->dp_scan;
3010         spa_t *spa = dp->dp_spa;
3011
3012         if (spa_suspend_async_destroy(spa))
3013                 return (0);
3014
3015         if (zfs_free_bpobj_enabled &&
3016             spa_version(spa) >= SPA_VERSION_DEADLISTS) {
3017                 scn->scn_is_bptree = B_FALSE;
3018                 scn->scn_async_block_min_time_ms = zfs_free_min_time_ms;
3019                 scn->scn_zio_root = zio_root(spa, NULL,
3020                     NULL, ZIO_FLAG_MUSTSUCCEED);
3021                 err = bpobj_iterate(&dp->dp_free_bpobj,
3022                     dsl_scan_free_block_cb, scn, tx);
3023                 VERIFY0(zio_wait(scn->scn_zio_root));
3024                 scn->scn_zio_root = NULL;
3025
3026                 if (err != 0 && err != ERESTART)
3027                         zfs_panic_recover("error %u from bpobj_iterate()", err);
3028         }
3029
3030         if (err == 0 && spa_feature_is_active(spa, SPA_FEATURE_ASYNC_DESTROY)) {
3031                 ASSERT(scn->scn_async_destroying);
3032                 scn->scn_is_bptree = B_TRUE;
3033                 scn->scn_zio_root = zio_root(spa, NULL,
3034                     NULL, ZIO_FLAG_MUSTSUCCEED);
3035                 err = bptree_iterate(dp->dp_meta_objset,
3036                     dp->dp_bptree_obj, B_TRUE, dsl_scan_free_block_cb, scn, tx);
3037                 VERIFY0(zio_wait(scn->scn_zio_root));
3038                 scn->scn_zio_root = NULL;
3039
3040                 if (err == EIO || err == ECKSUM) {
3041                         err = 0;
3042                 } else if (err != 0 && err != ERESTART) {
3043                         zfs_panic_recover("error %u from "
3044                             "traverse_dataset_destroyed()", err);
3045                 }
3046
3047                 if (bptree_is_empty(dp->dp_meta_objset, dp->dp_bptree_obj)) {
3048                         /* finished; deactivate async destroy feature */
3049                         spa_feature_decr(spa, SPA_FEATURE_ASYNC_DESTROY, tx);
3050                         ASSERT(!spa_feature_is_active(spa,
3051                             SPA_FEATURE_ASYNC_DESTROY));
3052                         VERIFY0(zap_remove(dp->dp_meta_objset,
3053                             DMU_POOL_DIRECTORY_OBJECT,
3054                             DMU_POOL_BPTREE_OBJ, tx));
3055                         VERIFY0(bptree_free(dp->dp_meta_objset,
3056                             dp->dp_bptree_obj, tx));
3057                         dp->dp_bptree_obj = 0;
3058                         scn->scn_async_destroying = B_FALSE;
3059                         scn->scn_async_stalled = B_FALSE;
3060                 } else {
3061                         /*
3062                          * If we didn't make progress, mark the async
3063                          * destroy as stalled, so that we will not initiate
3064                          * a spa_sync() on its behalf.  Note that we only
3065                          * check this if we are not finished, because if the
3066                          * bptree had no blocks for us to visit, we can
3067                          * finish without "making progress".
3068                          */
3069                         scn->scn_async_stalled =
3070                             (scn->scn_visited_this_txg == 0);
3071                 }
3072         }
3073         if (scn->scn_visited_this_txg) {
3074                 zfs_dbgmsg("freed %llu blocks in %llums from "
3075                     "free_bpobj/bptree txg %llu; err=%d",
3076                     (longlong_t)scn->scn_visited_this_txg,
3077                     (longlong_t)
3078                     NSEC2MSEC(gethrtime() - scn->scn_sync_start_time),
3079                     (longlong_t)tx->tx_txg, err);
3080                 scn->scn_visited_this_txg = 0;
3081
3082                 /*
3083                  * Write out changes to the DDT that may be required as a
3084                  * result of the blocks freed.  This ensures that the DDT
3085                  * is clean when a scrub/resilver runs.
3086                  */
3087                 ddt_sync(spa, tx->tx_txg);
3088         }
3089         if (err != 0)
3090                 return (err);
3091         if (dp->dp_free_dir != NULL && !scn->scn_async_destroying &&
3092             zfs_free_leak_on_eio &&
3093             (dsl_dir_phys(dp->dp_free_dir)->dd_used_bytes != 0 ||
3094             dsl_dir_phys(dp->dp_free_dir)->dd_compressed_bytes != 0 ||
3095             dsl_dir_phys(dp->dp_free_dir)->dd_uncompressed_bytes != 0)) {
3096                 /*
3097                  * We have finished background destroying, but there is still
3098                  * some space left in the dp_free_dir. Transfer this leaked
3099                  * space to the dp_leak_dir.
3100                  */
3101                 if (dp->dp_leak_dir == NULL) {
3102                         rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG);
3103                         (void) dsl_dir_create_sync(dp, dp->dp_root_dir,
3104                             LEAK_DIR_NAME, tx);
3105                         VERIFY0(dsl_pool_open_special_dir(dp,
3106                             LEAK_DIR_NAME, &dp->dp_leak_dir));
3107                         rrw_exit(&dp->dp_config_rwlock, FTAG);
3108                 }
3109                 dsl_dir_diduse_space(dp->dp_leak_dir, DD_USED_HEAD,
3110                     dsl_dir_phys(dp->dp_free_dir)->dd_used_bytes,
3111                     dsl_dir_phys(dp->dp_free_dir)->dd_compressed_bytes,
3112                     dsl_dir_phys(dp->dp_free_dir)->dd_uncompressed_bytes, tx);
3113                 dsl_dir_diduse_space(dp->dp_free_dir, DD_USED_HEAD,
3114                     -dsl_dir_phys(dp->dp_free_dir)->dd_used_bytes,
3115                     -dsl_dir_phys(dp->dp_free_dir)->dd_compressed_bytes,
3116                     -dsl_dir_phys(dp->dp_free_dir)->dd_uncompressed_bytes, tx);
3117         }
3118
3119         if (dp->dp_free_dir != NULL && !scn->scn_async_destroying) {
3120                 /* finished; verify that space accounting went to zero */
3121                 ASSERT0(dsl_dir_phys(dp->dp_free_dir)->dd_used_bytes);
3122                 ASSERT0(dsl_dir_phys(dp->dp_free_dir)->dd_compressed_bytes);
3123                 ASSERT0(dsl_dir_phys(dp->dp_free_dir)->dd_uncompressed_bytes);
3124         }
3125
3126         EQUIV(bpobj_is_open(&dp->dp_obsolete_bpobj),
3127             0 == zap_contains(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
3128             DMU_POOL_OBSOLETE_BPOBJ));
3129         if (err == 0 && bpobj_is_open(&dp->dp_obsolete_bpobj)) {
3130                 ASSERT(spa_feature_is_active(dp->dp_spa,
3131                     SPA_FEATURE_OBSOLETE_COUNTS));
3132
3133                 scn->scn_is_bptree = B_FALSE;
3134                 scn->scn_async_block_min_time_ms = zfs_obsolete_min_time_ms;
3135                 err = bpobj_iterate(&dp->dp_obsolete_bpobj,
3136                     dsl_scan_obsolete_block_cb, scn, tx);
3137                 if (err != 0 && err != ERESTART)
3138                         zfs_panic_recover("error %u from bpobj_iterate()", err);
3139
3140                 if (bpobj_is_empty(&dp->dp_obsolete_bpobj))
3141                         dsl_pool_destroy_obsolete_bpobj(dp, tx);
3142         }
3143
3144         return (0);
3145 }
3146
3147 /*
3148  * This is the primary entry point for scans that is called from syncing
3149  * context. Scans must happen entirely during syncing context so that we
3150  * cna guarantee that blocks we are currently scanning will not change out
3151  * from under us. While a scan is active, this funciton controls how quickly
3152  * transaction groups proceed, instead of the normal handling provided by
3153  * txg_sync_thread().
3154  */
3155 void
3156 dsl_scan_sync(dsl_pool_t *dp, dmu_tx_t *tx)
3157 {
3158         dsl_scan_t *scn = dp->dp_scan;
3159         spa_t *spa = dp->dp_spa;
3160         int err = 0;
3161         state_sync_type_t sync_type = SYNC_OPTIONAL;
3162
3163         /*
3164          * Check for scn_restart_txg before checking spa_load_state, so
3165          * that we can restart an old-style scan while the pool is being
3166          * imported (see dsl_scan_init).
3167          */
3168         if (dsl_scan_restarting(scn, tx)) {
3169                 pool_scan_func_t func = POOL_SCAN_SCRUB;
3170                 dsl_scan_done(scn, B_FALSE, tx);
3171                 if (vdev_resilver_needed(spa->spa_root_vdev, NULL, NULL))
3172                         func = POOL_SCAN_RESILVER;
3173                 zfs_dbgmsg("restarting scan func=%u txg=%llu",
3174                     func, (longlong_t)tx->tx_txg);
3175                 dsl_scan_setup_sync(&func, tx);
3176         }
3177
3178         /*
3179          * Only process scans in sync pass 1.
3180          */
3181         if (spa_sync_pass(dp->dp_spa) > 1)
3182                 return;
3183
3184         /*
3185          * If the spa is shutting down, then stop scanning. This will
3186          * ensure that the scan does not dirty any new data during the
3187          * shutdown phase.
3188          */
3189         if (spa_shutting_down(spa))
3190                 return;
3191
3192         /*
3193          * If the scan is inactive due to a stalled async destroy, try again.
3194          */
3195         if (!scn->scn_async_stalled && !dsl_scan_active(scn))
3196                 return;
3197
3198         /* reset scan statistics */
3199         scn->scn_visited_this_txg = 0;
3200         scn->scn_holes_this_txg = 0;
3201         scn->scn_lt_min_this_txg = 0;
3202         scn->scn_gt_max_this_txg = 0;
3203         scn->scn_ddt_contained_this_txg = 0;
3204         scn->scn_objsets_visited_this_txg = 0;
3205         scn->scn_avg_seg_size_this_txg = 0;
3206         scn->scn_segs_this_txg = 0;
3207         scn->scn_avg_zio_size_this_txg = 0;
3208         scn->scn_zios_this_txg = 0;
3209         scn->scn_suspending = B_FALSE;
3210         scn->scn_sync_start_time = gethrtime();
3211         spa->spa_scrub_active = B_TRUE;
3212
3213         /*
3214          * First process the async destroys.  If we pause, don't do
3215          * any scrubbing or resilvering.  This ensures that there are no
3216          * async destroys while we are scanning, so the scan code doesn't
3217          * have to worry about traversing it.  It is also faster to free the
3218          * blocks than to scrub them.
3219          */
3220         err = dsl_process_async_destroys(dp, tx);
3221         if (err != 0)
3222                 return;
3223
3224         if (!dsl_scan_is_running(scn) || dsl_scan_is_paused_scrub(scn))
3225                 return;
3226
3227         /*
3228          * Wait a few txgs after importing to begin scanning so that
3229          * we can get the pool imported quickly.
3230          */
3231         if (spa->spa_syncing_txg < spa->spa_first_txg + SCAN_IMPORT_WAIT_TXGS)
3232                 return;
3233
3234         /*
3235          * It is possible to switch from unsorted to sorted at any time,
3236          * but afterwards the scan will remain sorted unless reloaded from
3237          * a checkpoint after a reboot.
3238          */
3239         if (!zfs_scan_legacy) {
3240                 scn->scn_is_sorted = B_TRUE;
3241                 if (scn->scn_last_checkpoint == 0)
3242                         scn->scn_last_checkpoint = ddi_get_lbolt();
3243         }
3244
3245         /*
3246          * For sorted scans, determine what kind of work we will be doing
3247          * this txg based on our memory limitations and whether or not we
3248          * need to perform a checkpoint.
3249          */
3250         if (scn->scn_is_sorted) {
3251                 /*
3252                  * If we are over our checkpoint interval, set scn_clearing
3253                  * so that we can begin checkpointing immediately. The
3254                  * checkpoint allows us to save a consisent bookmark
3255                  * representing how much data we have scrubbed so far.
3256                  * Otherwise, use the memory limit to determine if we should
3257                  * scan for metadata or start issue scrub IOs. We accumulate
3258                  * metadata until we hit our hard memory limit at which point
3259                  * we issue scrub IOs until we are at our soft memory limit.
3260                  */
3261                 if (scn->scn_checkpointing ||
3262                     ddi_get_lbolt() - scn->scn_last_checkpoint >
3263                     SEC_TO_TICK(zfs_scan_checkpoint_intval)) {
3264                         if (!scn->scn_checkpointing)
3265                                 zfs_dbgmsg("begin scan checkpoint");
3266
3267                         scn->scn_checkpointing = B_TRUE;
3268                         scn->scn_clearing = B_TRUE;
3269                 } else {
3270                         boolean_t should_clear = dsl_scan_should_clear(scn);
3271                         if (should_clear && !scn->scn_clearing) {
3272                                 zfs_dbgmsg("begin scan clearing");
3273                                 scn->scn_clearing = B_TRUE;
3274                         } else if (!should_clear && scn->scn_clearing) {
3275                                 zfs_dbgmsg("finish scan clearing");
3276                                 scn->scn_clearing = B_FALSE;
3277                         }
3278                 }
3279         } else {
3280                 ASSERT0(scn->scn_checkpointing);
3281                 ASSERT0(scn->scn_clearing);
3282         }
3283
3284         if (!scn->scn_clearing && scn->scn_done_txg == 0) {
3285                 /* Need to scan metadata for more blocks to scrub */
3286                 dsl_scan_phys_t *scnp = &scn->scn_phys;
3287                 taskqid_t prefetch_tqid;
3288                 uint64_t bytes_per_leaf = zfs_scan_vdev_limit;
3289                 uint64_t nr_leaves = dsl_scan_count_leaves(spa->spa_root_vdev);
3290
3291                 /*
3292                  * Calculate the max number of in-flight bytes for pool-wide
3293                  * scanning operations (minimum 1MB). Limits for the issuing
3294                  * phase are done per top-level vdev and are handled separately.
3295                  */
3296                 scn->scn_maxinflight_bytes =
3297                     MAX(nr_leaves * bytes_per_leaf, 1ULL << 20);
3298
3299                 if (scnp->scn_ddt_bookmark.ddb_class <=
3300                     scnp->scn_ddt_class_max) {
3301                         ASSERT(ZB_IS_ZERO(&scnp->scn_bookmark));
3302                         zfs_dbgmsg("doing scan sync txg %llu; "
3303                             "ddt bm=%llu/%llu/%llu/%llx",
3304                             (longlong_t)tx->tx_txg,
3305                             (longlong_t)scnp->scn_ddt_bookmark.ddb_class,
3306                             (longlong_t)scnp->scn_ddt_bookmark.ddb_type,
3307                             (longlong_t)scnp->scn_ddt_bookmark.ddb_checksum,
3308                             (longlong_t)scnp->scn_ddt_bookmark.ddb_cursor);
3309                 } else {
3310                         zfs_dbgmsg("doing scan sync txg %llu; "
3311                             "bm=%llu/%llu/%llu/%llu",
3312                             (longlong_t)tx->tx_txg,
3313                             (longlong_t)scnp->scn_bookmark.zb_objset,
3314                             (longlong_t)scnp->scn_bookmark.zb_object,
3315                             (longlong_t)scnp->scn_bookmark.zb_level,
3316                             (longlong_t)scnp->scn_bookmark.zb_blkid);
3317                 }
3318
3319                 scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
3320                     NULL, ZIO_FLAG_CANFAIL);
3321
3322                 scn->scn_prefetch_stop = B_FALSE;
3323                 prefetch_tqid = taskq_dispatch(dp->dp_sync_taskq,
3324                     dsl_scan_prefetch_thread, scn, TQ_SLEEP);
3325                 ASSERT(prefetch_tqid != TASKQID_INVALID);
3326
3327                 dsl_pool_config_enter(dp, FTAG);
3328                 dsl_scan_visit(scn, tx);
3329                 dsl_pool_config_exit(dp, FTAG);
3330
3331                 mutex_enter(&dp->dp_spa->spa_scrub_lock);
3332                 scn->scn_prefetch_stop = B_TRUE;
3333                 cv_broadcast(&spa->spa_scrub_io_cv);
3334                 mutex_exit(&dp->dp_spa->spa_scrub_lock);
3335
3336                 taskq_wait_id(dp->dp_sync_taskq, prefetch_tqid);
3337                 (void) zio_wait(scn->scn_zio_root);
3338                 scn->scn_zio_root = NULL;
3339
3340                 zfs_dbgmsg("scan visited %llu blocks in %llums "
3341                     "(%llu os's, %llu holes, %llu < mintxg, "
3342                     "%llu in ddt, %llu > maxtxg)",
3343                     (longlong_t)scn->scn_visited_this_txg,
3344                     (longlong_t)NSEC2MSEC(gethrtime() -
3345                     scn->scn_sync_start_time),
3346                     (longlong_t)scn->scn_objsets_visited_this_txg,
3347                     (longlong_t)scn->scn_holes_this_txg,
3348                     (longlong_t)scn->scn_lt_min_this_txg,
3349                     (longlong_t)scn->scn_ddt_contained_this_txg,
3350                     (longlong_t)scn->scn_gt_max_this_txg);
3351
3352                 if (!scn->scn_suspending) {
3353                         ASSERT0(avl_numnodes(&scn->scn_queue));
3354                         scn->scn_done_txg = tx->tx_txg + 1;
3355                         if (scn->scn_is_sorted) {
3356                                 scn->scn_checkpointing = B_TRUE;
3357                                 scn->scn_clearing = B_TRUE;
3358                         }
3359                         zfs_dbgmsg("scan complete txg %llu",
3360                                    (longlong_t)tx->tx_txg);
3361                 }
3362         } else if (scn->scn_is_sorted && scn->scn_bytes_pending != 0) {
3363                 /* need to issue scrubbing IOs from per-vdev queues */
3364                 scn->scn_zio_root = zio_root(dp->dp_spa, NULL,
3365                     NULL, ZIO_FLAG_CANFAIL);
3366                 scan_io_queues_run(scn);
3367                 (void) zio_wait(scn->scn_zio_root);
3368                 scn->scn_zio_root = NULL;
3369
3370                 /* calculate and dprintf the current memory usage */
3371                 (void) dsl_scan_should_clear(scn);
3372                 dsl_scan_update_stats(scn);
3373
3374                 zfs_dbgmsg("scrubbed %llu blocks (%llu segs) in %llums "
3375                     "(avg_block_size = %llu, avg_seg_size = %llu)",
3376                     (longlong_t)scn->scn_zios_this_txg,
3377                     (longlong_t)scn->scn_segs_this_txg,
3378                     (longlong_t)NSEC2MSEC(gethrtime() -
3379                     scn->scn_sync_start_time),
3380                     (longlong_t)scn->scn_avg_zio_size_this_txg,
3381                     (longlong_t)scn->scn_avg_seg_size_this_txg);
3382         } else if (scn->scn_done_txg != 0 && scn->scn_done_txg <= tx->tx_txg) {
3383                 /* Finished with everything. Mark the scrub as complete */
3384                 zfs_dbgmsg("scan issuing complete txg %llu",
3385                     (longlong_t)tx->tx_txg);
3386                 ASSERT3U(scn->scn_done_txg, !=, 0);
3387                 ASSERT0(spa->spa_scrub_inflight);
3388                 ASSERT0(scn->scn_bytes_pending);
3389                 dsl_scan_done(scn, B_TRUE, tx);
3390                 sync_type = SYNC_MANDATORY;
3391         }
3392
3393         dsl_scan_sync_state(scn, tx, sync_type);
3394 }
3395
3396 static void
3397 count_block(dsl_scan_t *scn, zfs_all_blkstats_t *zab, const blkptr_t *bp)
3398 {
3399         int i;
3400
3401         /* update the spa's stats on how many bytes we have issued */
3402         for (i = 0; i < BP_GET_NDVAS(bp); i++) {
3403                 atomic_add_64(&scn->scn_dp->dp_spa->spa_scan_pass_issued,
3404                     DVA_GET_ASIZE(&bp->blk_dva[i]));
3405         }
3406
3407         /*
3408          * If we resume after a reboot, zab will be NULL; don't record
3409          * incomplete stats in that case.
3410          */
3411         if (zab == NULL)
3412                 return;
3413
3414         mutex_enter(&zab->zab_lock);
3415
3416         for (i = 0; i < 4; i++) {
3417                 int l = (i < 2) ? BP_GET_LEVEL(bp) : DN_MAX_LEVELS;
3418                 int t = (i & 1) ? BP_GET_TYPE(bp) : DMU_OT_TOTAL;
3419                 if (t & DMU_OT_NEWTYPE)
3420                         t = DMU_OT_OTHER;
3421                 zfs_blkstat_t *zb = &zab->zab_type[l][t];
3422                 int equal;
3423
3424                 zb->zb_count++;
3425                 zb->zb_asize += BP_GET_ASIZE(bp);
3426                 zb->zb_lsize += BP_GET_LSIZE(bp);
3427                 zb->zb_psize += BP_GET_PSIZE(bp);
3428                 zb->zb_gangs += BP_COUNT_GANG(bp);
3429
3430                 switch (BP_GET_NDVAS(bp)) {
3431                 case 2:
3432                         if (DVA_GET_VDEV(&bp->blk_dva[0]) ==
3433                             DVA_GET_VDEV(&bp->blk_dva[1]))
3434                                 zb->zb_ditto_2_of_2_samevdev++;
3435                         break;
3436                 case 3:
3437                         equal = (DVA_GET_VDEV(&bp->blk_dva[0]) ==
3438                             DVA_GET_VDEV(&bp->blk_dva[1])) +
3439                             (DVA_GET_VDEV(&bp->blk_dva[0]) ==
3440                             DVA_GET_VDEV(&bp->blk_dva[2])) +
3441                             (DVA_GET_VDEV(&bp->blk_dva[1]) ==
3442                             DVA_GET_VDEV(&bp->blk_dva[2]));
3443                         if (equal == 1)
3444                                 zb->zb_ditto_2_of_3_samevdev++;
3445                         else if (equal == 3)
3446                                 zb->zb_ditto_3_of_3_samevdev++;
3447                         break;
3448                 }
3449         }
3450
3451         mutex_exit(&zab->zab_lock);
3452 }
3453
3454 static void
3455 scan_io_queue_insert_impl(dsl_scan_io_queue_t *queue, scan_io_t *sio)
3456 {
3457         avl_index_t idx;
3458         int64_t asize = sio->sio_asize;
3459         dsl_scan_t *scn = queue->q_scn;
3460
3461         ASSERT(MUTEX_HELD(&queue->q_vd->vdev_scan_io_queue_lock));
3462
3463         if (avl_find(&queue->q_sios_by_addr, sio, &idx) != NULL) {
3464                 /* block is already scheduled for reading */
3465                 atomic_add_64(&scn->scn_bytes_pending, -asize);
3466                 kmem_free(sio, sizeof (*sio));
3467                 return;
3468         }
3469         avl_insert(&queue->q_sios_by_addr, sio, idx);
3470         range_tree_add(queue->q_exts_by_addr, sio->sio_offset, asize);
3471 }
3472
3473 /*
3474  * Given all the info we got from our metadata scanning process, we
3475  * construct a scan_io_t and insert it into the scan sorting queue. The
3476  * I/O must already be suitable for us to process. This is controlled
3477  * by dsl_scan_enqueue().
3478  */
3479 static void
3480 scan_io_queue_insert(dsl_scan_io_queue_t *queue, const blkptr_t *bp, int dva_i,
3481     int zio_flags, const zbookmark_phys_t *zb)
3482 {
3483         dsl_scan_t *scn = queue->q_scn;
3484         scan_io_t *sio = kmem_zalloc(sizeof (*sio), KM_SLEEP);
3485
3486         ASSERT0(BP_IS_GANG(bp));
3487         ASSERT(MUTEX_HELD(&queue->q_vd->vdev_scan_io_queue_lock));
3488
3489         bp2sio(bp, sio, dva_i);
3490         sio->sio_flags = zio_flags;
3491         sio->sio_zb = *zb;
3492
3493         /*
3494          * Increment the bytes pending counter now so that we can't
3495          * get an integer underflow in case the worker processes the
3496          * zio before we get to incrementing this counter.
3497          */
3498         atomic_add_64(&scn->scn_bytes_pending, sio->sio_asize);
3499
3500         scan_io_queue_insert_impl(queue, sio);
3501 }
3502
3503 /*
3504  * Given a set of I/O parameters as discovered by the metadata traversal
3505  * process, attempts to place the I/O into the sorted queues (if allowed),
3506  * or immediately executes the I/O.
3507  */
3508 static void
3509 dsl_scan_enqueue(dsl_pool_t *dp, const blkptr_t *bp, int zio_flags,
3510     const zbookmark_phys_t *zb)
3511 {
3512         spa_t *spa = dp->dp_spa;
3513
3514         ASSERT(!BP_IS_EMBEDDED(bp));
3515
3516         /*
3517          * Gang blocks are hard to issue sequentially, so we just issue them
3518          * here immediately instead of queuing them.
3519          */
3520         if (!dp->dp_scan->scn_is_sorted || BP_IS_GANG(bp)) {
3521                 scan_exec_io(dp, bp, zio_flags, zb, NULL);
3522                 return;
3523         }
3524         for (int i = 0; i < BP_GET_NDVAS(bp); i++) {
3525                 dva_t dva;
3526                 vdev_t *vdev;
3527
3528                 dva = bp->blk_dva[i];
3529                 vdev = vdev_lookup_top(spa, DVA_GET_VDEV(&dva));
3530                 ASSERT(vdev != NULL);
3531
3532                 mutex_enter(&vdev->vdev_scan_io_queue_lock);
3533                 if (vdev->vdev_scan_io_queue == NULL)
3534                         vdev->vdev_scan_io_queue = scan_io_queue_create(vdev);
3535                 ASSERT(dp->dp_scan != NULL);
3536                 scan_io_queue_insert(vdev->vdev_scan_io_queue, bp,
3537                     i, zio_flags, zb);
3538                 mutex_exit(&vdev->vdev_scan_io_queue_lock);
3539         }
3540 }
3541
3542 static int
3543 dsl_scan_scrub_cb(dsl_pool_t *dp,
3544     const blkptr_t *bp, const zbookmark_phys_t *zb)
3545 {
3546         dsl_scan_t *scn = dp->dp_scan;
3547         spa_t *spa = dp->dp_spa;
3548         uint64_t phys_birth = BP_PHYSICAL_BIRTH(bp);
3549         size_t psize = BP_GET_PSIZE(bp);
3550         boolean_t needs_io;
3551         int zio_flags = ZIO_FLAG_SCAN_THREAD | ZIO_FLAG_RAW | ZIO_FLAG_CANFAIL;
3552         int d;
3553
3554         if (phys_birth <= scn->scn_phys.scn_min_txg ||
3555             phys_birth >= scn->scn_phys.scn_max_txg)
3556                 return (0);
3557
3558         if (BP_IS_EMBEDDED(bp)) {
3559                 count_block(scn, dp->dp_blkstats, bp);
3560                 return (0);
3561         }
3562
3563         ASSERT(DSL_SCAN_IS_SCRUB_RESILVER(scn));
3564         if (scn->scn_phys.scn_func == POOL_SCAN_SCRUB) {
3565                 zio_flags |= ZIO_FLAG_SCRUB;
3566                 needs_io = B_TRUE;
3567         } else {
3568                 ASSERT3U(scn->scn_phys.scn_func, ==, POOL_SCAN_RESILVER);
3569                 zio_flags |= ZIO_FLAG_RESILVER;
3570                 needs_io = B_FALSE;
3571         }
3572
3573         /* If it's an intent log block, failure is expected. */
3574         if (zb->zb_level == ZB_ZIL_LEVEL)
3575                 zio_flags |= ZIO_FLAG_SPECULATIVE;
3576
3577         for (d = 0; d < BP_GET_NDVAS(bp); d++) {
3578                 const dva_t *dva = &bp->blk_dva[d];
3579
3580                 /*
3581                  * Keep track of how much data we've examined so that
3582                  * zpool(1M) status can make useful progress reports.
3583                  */
3584                 scn->scn_phys.scn_examined += DVA_GET_ASIZE(dva);
3585                 spa->spa_scan_pass_exam += DVA_GET_ASIZE(dva);
3586
3587                 /* if it's a resilver, this may not be in the target range */
3588                 if (!needs_io)
3589                         needs_io = dsl_scan_need_resilver(spa, dva, psize,
3590                             phys_birth);
3591         }
3592
3593         if (needs_io && !zfs_no_scrub_io) {
3594                 dsl_scan_enqueue(dp, bp, zio_flags, zb);
3595         } else {
3596                 count_block(scn, dp->dp_blkstats, bp);
3597         }
3598
3599         /* do not relocate this block */
3600         return (0);
3601 }
3602
3603 static void
3604 dsl_scan_scrub_done(zio_t *zio)
3605 {
3606         spa_t *spa = zio->io_spa;
3607         blkptr_t *bp = zio->io_bp;
3608         dsl_scan_io_queue_t *queue = zio->io_private;
3609
3610         abd_free(zio->io_abd);
3611
3612         if (queue == NULL) {
3613                 mutex_enter(&spa->spa_scrub_lock);
3614                 ASSERT3U(spa->spa_scrub_inflight, >=, BP_GET_PSIZE(bp));
3615                 spa->spa_scrub_inflight -= BP_GET_PSIZE(bp);
3616                 cv_broadcast(&spa->spa_scrub_io_cv);
3617                 mutex_exit(&spa->spa_scrub_lock);
3618         } else {
3619                 mutex_enter(&queue->q_vd->vdev_scan_io_queue_lock);
3620                 ASSERT3U(queue->q_inflight_bytes, >=, BP_GET_PSIZE(bp));
3621                 queue->q_inflight_bytes -= BP_GET_PSIZE(bp);
3622                 cv_broadcast(&queue->q_zio_cv);
3623                 mutex_exit(&queue->q_vd->vdev_scan_io_queue_lock);
3624         }
3625
3626         if (zio->io_error && (zio->io_error != ECKSUM ||
3627             !(zio->io_flags & ZIO_FLAG_SPECULATIVE))) {
3628                 atomic_inc_64(&spa->spa_dsl_pool->dp_scan->scn_phys.scn_errors);
3629         }
3630 }
3631
3632 /*
3633  * Given a scanning zio's information, executes the zio. The zio need
3634  * not necessarily be only sortable, this function simply executes the
3635  * zio, no matter what it is. The optional queue argument allows the
3636  * caller to specify that they want per top level vdev IO rate limiting
3637  * instead of the legacy global limiting.
3638  */
3639 static void
3640 scan_exec_io(dsl_pool_t *dp, const blkptr_t *bp, int zio_flags,
3641     const zbookmark_phys_t *zb, dsl_scan_io_queue_t *queue)
3642 {
3643         spa_t *spa = dp->dp_spa;
3644         dsl_scan_t *scn = dp->dp_scan;
3645         size_t size = BP_GET_PSIZE(bp);
3646         abd_t *data = abd_alloc_for_io(size, B_FALSE);
3647         unsigned int scan_delay = 0;
3648
3649         if (queue == NULL) {
3650                 mutex_enter(&spa->spa_scrub_lock);
3651                 while (spa->spa_scrub_inflight >= scn->scn_maxinflight_bytes)
3652                         cv_wait(&spa->spa_scrub_io_cv, &spa->spa_scrub_lock);
3653                 spa->spa_scrub_inflight += BP_GET_PSIZE(bp);
3654                 mutex_exit(&spa->spa_scrub_lock);
3655         } else {
3656                 kmutex_t *q_lock = &queue->q_vd->vdev_scan_io_queue_lock;
3657
3658                 mutex_enter(q_lock);
3659                 while (queue->q_inflight_bytes >= queue->q_maxinflight_bytes)
3660                         cv_wait(&queue->q_zio_cv, q_lock);
3661                 queue->q_inflight_bytes += BP_GET_PSIZE(bp);
3662                 mutex_exit(q_lock);
3663         }
3664
3665         if (zio_flags & ZIO_FLAG_RESILVER)
3666                 scan_delay = zfs_resilver_delay;
3667         else {
3668                 ASSERT(zio_flags & ZIO_FLAG_SCRUB);
3669                 scan_delay = zfs_scrub_delay;
3670         }
3671
3672         if (scan_delay && (ddi_get_lbolt64() - spa->spa_last_io <= zfs_scan_idle))
3673                 delay(MAX((int)scan_delay, 0));
3674
3675         count_block(dp->dp_scan, dp->dp_blkstats, bp);
3676         zio_nowait(zio_read(dp->dp_scan->scn_zio_root, spa, bp, data, size,
3677             dsl_scan_scrub_done, queue, ZIO_PRIORITY_SCRUB, zio_flags, zb));
3678 }
3679
3680 /*
3681  * This is the primary extent sorting algorithm. We balance two parameters:
3682  * 1) how many bytes of I/O are in an extent
3683  * 2) how well the extent is filled with I/O (as a fraction of its total size)
3684  * Since we allow extents to have gaps between their constituent I/Os, it's
3685  * possible to have a fairly large extent that contains the same amount of
3686  * I/O bytes than a much smaller extent, which just packs the I/O more tightly.
3687  * The algorithm sorts based on a score calculated from the extent's size,
3688  * the relative fill volume (in %) and a "fill weight" parameter that controls
3689  * the split between whether we prefer larger extents or more well populated
3690  * extents:
3691  *
3692  * SCORE = FILL_IN_BYTES + (FILL_IN_PERCENT * FILL_IN_BYTES * FILL_WEIGHT)
3693  *
3694  * Example:
3695  * 1) assume extsz = 64 MiB
3696  * 2) assume fill = 32 MiB (extent is half full)
3697  * 3) assume fill_weight = 3
3698  * 4)   SCORE = 32M + (((32M * 100) / 64M) * 3 * 32M) / 100
3699  *      SCORE = 32M + (50 * 3 * 32M) / 100
3700  *      SCORE = 32M + (4800M / 100)
3701  *      SCORE = 32M + 48M
3702  *               ^     ^
3703  *               |     +--- final total relative fill-based score
3704  *               +--------- final total fill-based score
3705  *      SCORE = 80M
3706  *
3707  * As can be seen, at fill_ratio=3, the algorithm is slightly biased towards
3708  * extents that are more completely filled (in a 3:2 ratio) vs just larger.
3709  * Note that as an optimization, we replace multiplication and division by
3710  * 100 with bitshifting by 7 (which effecitvely multiplies and divides by 128).
3711  */
3712 static int
3713 ext_size_compare(const void *x, const void *y)
3714 {
3715         const range_seg_t *rsa = x, *rsb = y;
3716         uint64_t sa = rsa->rs_end - rsa->rs_start,
3717             sb = rsb->rs_end - rsb->rs_start;
3718         uint64_t score_a, score_b;
3719
3720         score_a = rsa->rs_fill + ((((rsa->rs_fill << 7) / sa) *
3721             fill_weight * rsa->rs_fill) >> 7);
3722         score_b = rsb->rs_fill + ((((rsb->rs_fill << 7) / sb) *
3723             fill_weight * rsb->rs_fill) >> 7);
3724
3725         if (score_a > score_b)
3726                 return (-1);
3727         if (score_a == score_b) {
3728                 if (rsa->rs_start < rsb->rs_start)
3729                         return (-1);
3730                 if (rsa->rs_start == rsb->rs_start)
3731                         return (0);
3732                 return (1);
3733         }
3734         return (1);
3735 }
3736
3737 /*
3738  * Comparator for the q_sios_by_addr tree. Sorting is simply performed
3739  * based on LBA-order (from lowest to highest).
3740  */
3741 static int
3742 io_addr_compare(const void *x, const void *y)
3743 {
3744         const scan_io_t *a = x, *b = y;
3745
3746         if (a->sio_offset < b->sio_offset)
3747                 return (-1);
3748         if (a->sio_offset == b->sio_offset)
3749                 return (0);
3750         return (1);
3751 }
3752
3753 /* IO queues are created on demand when they are needed. */
3754 static dsl_scan_io_queue_t *
3755 scan_io_queue_create(vdev_t *vd)
3756 {
3757         dsl_scan_t *scn = vd->vdev_spa->spa_dsl_pool->dp_scan;
3758         dsl_scan_io_queue_t *q = kmem_zalloc(sizeof (*q), KM_SLEEP);
3759
3760         q->q_scn = scn;
3761         q->q_vd = vd;
3762         cv_init(&q->q_zio_cv, NULL, CV_DEFAULT, NULL);
3763         q->q_exts_by_addr = range_tree_create_impl(&rt_avl_ops,
3764             &q->q_exts_by_size, ext_size_compare, zfs_scan_max_ext_gap);
3765         avl_create(&q->q_sios_by_addr, io_addr_compare,
3766             sizeof (scan_io_t), offsetof(scan_io_t, sio_nodes.sio_addr_node));
3767
3768         return (q);
3769 }
3770
3771 /*
3772  * Destroys a scan queue and all segments and scan_io_t's contained in it.
3773  * No further execution of I/O occurs, anything pending in the queue is
3774  * simply freed without being executed.
3775  */
3776 void
3777 dsl_scan_io_queue_destroy(dsl_scan_io_queue_t *queue)
3778 {
3779         dsl_scan_t *scn = queue->q_scn;
3780         scan_io_t *sio;
3781         void *cookie = NULL;
3782         int64_t bytes_dequeued = 0;
3783
3784         ASSERT(MUTEX_HELD(&queue->q_vd->vdev_scan_io_queue_lock));
3785
3786         while ((sio = avl_destroy_nodes(&queue->q_sios_by_addr, &cookie)) !=
3787             NULL) {
3788                 ASSERT(range_tree_contains(queue->q_exts_by_addr,
3789                     sio->sio_offset, sio->sio_asize));
3790                 bytes_dequeued += sio->sio_asize;
3791                 kmem_free(sio, sizeof (*sio));
3792         }
3793
3794         atomic_add_64(&scn->scn_bytes_pending, -bytes_dequeued);
3795         range_tree_vacate(queue->q_exts_by_addr, NULL, queue);
3796         range_tree_destroy(queue->q_exts_by_addr);
3797         avl_destroy(&queue->q_sios_by_addr);
3798         cv_destroy(&queue->q_zio_cv);
3799
3800         kmem_free(queue, sizeof (*queue));
3801 }
3802
3803 /*
3804  * Properly transfers a dsl_scan_queue_t from `svd' to `tvd'. This is
3805  * called on behalf of vdev_top_transfer when creating or destroying
3806  * a mirror vdev due to zpool attach/detach.
3807  */
3808 void
3809 dsl_scan_io_queue_vdev_xfer(vdev_t *svd, vdev_t *tvd)
3810 {
3811         mutex_enter(&svd->vdev_scan_io_queue_lock);
3812         mutex_enter(&tvd->vdev_scan_io_queue_lock);
3813
3814         VERIFY3P(tvd->vdev_scan_io_queue, ==, NULL);
3815         tvd->vdev_scan_io_queue = svd->vdev_scan_io_queue;
3816         svd->vdev_scan_io_queue = NULL;
3817         if (tvd->vdev_scan_io_queue != NULL)
3818                 tvd->vdev_scan_io_queue->q_vd = tvd;
3819
3820         mutex_exit(&tvd->vdev_scan_io_queue_lock);
3821         mutex_exit(&svd->vdev_scan_io_queue_lock);
3822 }
3823
3824 static void
3825 scan_io_queues_destroy(dsl_scan_t *scn)
3826 {
3827         vdev_t *rvd = scn->scn_dp->dp_spa->spa_root_vdev;
3828
3829         for (uint64_t i = 0; i < rvd->vdev_children; i++) {
3830                 vdev_t *tvd = rvd->vdev_child[i];
3831
3832                 mutex_enter(&tvd->vdev_scan_io_queue_lock);
3833                 if (tvd->vdev_scan_io_queue != NULL)
3834                         dsl_scan_io_queue_destroy(tvd->vdev_scan_io_queue);
3835                 tvd->vdev_scan_io_queue = NULL;
3836                 mutex_exit(&tvd->vdev_scan_io_queue_lock);
3837         }
3838 }
3839
3840 static void
3841 dsl_scan_freed_dva(spa_t *spa, const blkptr_t *bp, int dva_i)
3842 {
3843         dsl_pool_t *dp = spa->spa_dsl_pool;
3844         dsl_scan_t *scn = dp->dp_scan;
3845         vdev_t *vdev;
3846         kmutex_t *q_lock;
3847         dsl_scan_io_queue_t *queue;
3848         scan_io_t srch, *sio;
3849         avl_index_t idx;
3850         uint64_t start, size;
3851
3852         vdev = vdev_lookup_top(spa, DVA_GET_VDEV(&bp->blk_dva[dva_i]));
3853         ASSERT(vdev != NULL);
3854         q_lock = &vdev->vdev_scan_io_queue_lock;
3855         queue = vdev->vdev_scan_io_queue;
3856
3857         mutex_enter(q_lock);
3858         if (queue == NULL) {
3859                 mutex_exit(q_lock);
3860                 return;
3861         }
3862
3863         bp2sio(bp, &srch, dva_i);
3864         start = srch.sio_offset;
3865         size = srch.sio_asize;
3866
3867         /*
3868          * We can find the zio in two states:
3869          * 1) Cold, just sitting in the queue of zio's to be issued at
3870          *      some point in the future. In this case, all we do is
3871          *      remove the zio from the q_sios_by_addr tree, decrement
3872          *      its data volume from the containing range_seg_t and
3873          *      resort the q_exts_by_size tree to reflect that the
3874          *      range_seg_t has lost some of its 'fill'. We don't shorten
3875          *      the range_seg_t - this is usually rare enough not to be
3876          *      worth the extra hassle of trying keep track of precise
3877          *      extent boundaries.
3878          * 2) Hot, where the zio is currently in-flight in
3879          *      dsl_scan_issue_ios. In this case, we can't simply
3880          *      reach in and stop the in-flight zio's, so we instead
3881          *      block the caller. Eventually, dsl_scan_issue_ios will
3882          *      be done with issuing the zio's it gathered and will
3883          *      signal us.
3884          */
3885         sio = avl_find(&queue->q_sios_by_addr, &srch, &idx);
3886         if (sio != NULL) {
3887                 int64_t asize = sio->sio_asize;
3888                 blkptr_t tmpbp;
3889
3890                 /* Got it while it was cold in the queue */
3891                 ASSERT3U(start, ==, sio->sio_offset);
3892                 ASSERT3U(size, ==, asize);
3893                 avl_remove(&queue->q_sios_by_addr, sio);
3894
3895                 ASSERT(range_tree_contains(queue->q_exts_by_addr, start, size));
3896                 range_tree_remove_fill(queue->q_exts_by_addr, start, size);
3897
3898                 /*
3899                  * We only update scn_bytes_pending in the cold path,
3900                  * otherwise it will already have been accounted for as
3901                  * part of the zio's execution.
3902                  */
3903                 atomic_add_64(&scn->scn_bytes_pending, -asize);
3904
3905                 /* count the block as though we issued it */
3906                 sio2bp(sio, &tmpbp, dva_i);
3907                 count_block(scn, dp->dp_blkstats, &tmpbp);
3908
3909                 kmem_free(sio, sizeof (*sio));
3910         }
3911         mutex_exit(q_lock);
3912 }
3913
3914 /*
3915  * Callback invoked when a zio_free() zio is executing. This needs to be
3916  * intercepted to prevent the zio from deallocating a particular portion
3917  * of disk space and it then getting reallocated and written to, while we
3918  * still have it queued up for processing.
3919  */
3920 void
3921 dsl_scan_freed(spa_t *spa, const blkptr_t *bp)
3922 {
3923         dsl_pool_t *dp = spa->spa_dsl_pool;
3924         dsl_scan_t *scn = dp->dp_scan;
3925
3926         ASSERT(!BP_IS_EMBEDDED(bp));
3927         ASSERT(scn != NULL);
3928         if (!dsl_scan_is_running(scn))
3929                 return;
3930
3931         for (int i = 0; i < BP_GET_NDVAS(bp); i++)
3932                 dsl_scan_freed_dva(spa, bp, i);
3933 }