Sequential scrub and resilvers

Currently, scrubs and resilvers can take an extremely
long time to complete. This is largely due to the fact
that zfs scans process pools in logical order, as
determined by each block's bookmark. This makes sense
from a simplicity perspective, but blocks in zfs are
often scattered randomly across disks, particularly
due to zfs's copy-on-write mechanisms.

This patch improves performance by splitting scrubs
and resilvers into a metadata scanning phase and an IO
issuing phase. The metadata scan reads through the
structure of the pool and gathers an in-memory queue
of I/Os, sorted by size and offset on disk. The issuing
phase will then issue the scrub I/Os as sequentially as
possible, greatly improving performance.

This patch also updates and cleans up some of the scan
code which has not been updated in several years.

Authored-by: Saso Kiselkov <saso.kiselkov@nexenta.com>
Authored-by: Alek Pinchuk <apinchuk@datto.com>
Authored-by: Tom Caputi <tcaputi@datto.com>
Signed-off-by: Tom Caputi <tcaputi@datto.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>

cmd/zdb/zdb.c

@@ -2226,8 +2226,6 @@ dump_dir(objset_t *os)
 		max_slot_used = object + dnode_slots - 1;
 	}
 
-	ASSERT3U(object_count, ==, usedobjs);
-
 	(void) printf("\n");
 
 	(void) printf("    Dnode slots:\n");
@@ -2245,6 +2243,8 @@ dump_dir(objset_t *os)
 		(void) fprintf(stderr, "dmu_object_next() = %d\n", error);
 		abort();
 	}
+
+	ASSERT3U(object_count, ==, usedobjs);
 }
 
 static void
@@ -3089,7 +3089,7 @@ zdb_blkptr_done(zio_t *zio)
 	abd_free(zio->io_abd);
 
 	mutex_enter(&spa->spa_scrub_lock);
-	spa->spa_scrub_inflight--;
+	spa->spa_load_verify_ios--;
 	cv_broadcast(&spa->spa_scrub_io_cv);
 
 	if (ioerr && !(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
@@ -3160,9 +3160,9 @@ zdb_blkptr_cb(spa_t *spa, zilog_t *zilog, const blkptr_t *bp,
 			flags |= ZIO_FLAG_SPECULATIVE;
 
 		mutex_enter(&spa->spa_scrub_lock);
-		while (spa->spa_scrub_inflight > max_inflight)
+		while (spa->spa_load_verify_ios > max_inflight)
 			cv_wait(&spa->spa_scrub_io_cv, &spa->spa_scrub_lock);
-		spa->spa_scrub_inflight++;
+		spa->spa_load_verify_ios++;
 		mutex_exit(&spa->spa_scrub_lock);
 
 		zio_nowait(zio_read(NULL, spa, bp, abd, size,

cmd/zpool/zpool_main.c

@@ -57,6 +57,8 @@
 #include <sys/fm/protocol.h>
 #include <sys/zfs_ioctl.h>
 #include <sys/mount.h>
+#include <sys/sysmacros.h>
+
 #include <math.h>
 
 #include <libzfs.h>
@@ -1761,7 +1763,7 @@ print_status_config(zpool_handle_t *zhp, status_cbdata_t *cb, const char *name,
 	(void) nvlist_lookup_uint64_array(nv, ZPOOL_CONFIG_SCAN_STATS,
 	    (uint64_t **)&ps, &c);
 
-	if (ps && ps->pss_state == DSS_SCANNING &&
+	if (ps != NULL && ps->pss_state == DSS_SCANNING &&
 	    vs->vs_scan_processed != 0 && children == 0) {
 		(void) printf(gettext("  (%s)"),
 		    (ps->pss_func == POOL_SCAN_RESILVER) ?
@@ -5967,11 +5969,13 @@ void
 print_scan_status(pool_scan_stat_t *ps)
 {
 	time_t start, end, pause;
-	uint64_t elapsed, mins_left, hours_left;
-	uint64_t pass_exam, examined, total;
-	uint_t rate;
+	uint64_t total_secs_left;
+	uint64_t elapsed, secs_left, mins_left, hours_left, days_left;
+	uint64_t pass_scanned, scanned, pass_issued, issued, total;
+	uint_t scan_rate, issue_rate;
 	double fraction_done;
-	char processed_buf[7], examined_buf[7], total_buf[7], rate_buf[7];
+	char processed_buf[7], scanned_buf[7], issued_buf[7], total_buf[7];
+	char srate_buf[7], irate_buf[7];
 
 	(void) printf(gettext("  scan: "));
 
@@ -5985,30 +5989,35 @@ print_scan_status(pool_scan_stat_t *ps)
 	start = ps->pss_start_time;
 	end = ps->pss_end_time;
 	pause = ps->pss_pass_scrub_pause;
+
 	zfs_nicebytes(ps->pss_processed, processed_buf, sizeof (processed_buf));
 
 	assert(ps->pss_func == POOL_SCAN_SCRUB ||
 	    ps->pss_func == POOL_SCAN_RESILVER);
-	/*
-	 * Scan is finished or canceled.
-	 */
+
+	/* Scan is finished or canceled. */
 	if (ps->pss_state == DSS_FINISHED) {
-		uint64_t minutes_taken = (end - start) / 60;
-		char *fmt = NULL;
+		total_secs_left = end - start;
+		days_left = total_secs_left / 60 / 60 / 24;
+		hours_left = (total_secs_left / 60 / 60) % 24;
+		mins_left = (total_secs_left / 60) % 60;
+		secs_left = (total_secs_left % 60);
 
 		if (ps->pss_func == POOL_SCAN_SCRUB) {
-			fmt = gettext("scrub repaired %s in %lluh%um with "
-			    "%llu errors on %s");
+			(void) printf(gettext("scrub repaired %s "
+			    "in %llu days %02llu:%02llu:%02llu "
+			    "with %llu errors on %s"), processed_buf,
+			    (u_longlong_t)days_left, (u_longlong_t)hours_left,
+			    (u_longlong_t)mins_left, (u_longlong_t)secs_left,
+			    (u_longlong_t)ps->pss_errors, ctime(&end));
 		} else if (ps->pss_func == POOL_SCAN_RESILVER) {
-			fmt = gettext("resilvered %s in %lluh%um with "
-			    "%llu errors on %s");
+			(void) printf(gettext("resilvered %s "
+			    "in %llu days %02llu:%02llu:%02llu "
+			    "with %llu errors on %s"), processed_buf,
+			    (u_longlong_t)days_left, (u_longlong_t)hours_left,
+			    (u_longlong_t)mins_left, (u_longlong_t)secs_left,
+			    (u_longlong_t)ps->pss_errors, ctime(&end));
 		}
-		/* LINTED */
-		(void) printf(fmt, processed_buf,
-		    (u_longlong_t)(minutes_taken / 60),
-		    (uint_t)(minutes_taken % 60),
-		    (u_longlong_t)ps->pss_errors,
-		    ctime((time_t *)&end));
 		return;
 	} else if (ps->pss_state == DSS_CANCELED) {
 		if (ps->pss_func == POOL_SCAN_SCRUB) {
@@ -6023,70 +6032,84 @@ print_scan_status(pool_scan_stat_t *ps)
 
 	assert(ps->pss_state == DSS_SCANNING);
 
-	/*
-	 * Scan is in progress.
-	 */
+	/* Scan is in progress. Resilvers can't be paused. */
 	if (ps->pss_func == POOL_SCAN_SCRUB) {
 		if (pause == 0) {
 			(void) printf(gettext("scrub in progress since %s"),
 			    ctime(&start));
 		} else {
-			char buf[32];
-			struct tm *p = localtime(&pause);
-			(void) strftime(buf, sizeof (buf), "%a %b %e %T %Y", p);
-			(void) printf(gettext("scrub paused since %s\n"), buf);
-			(void) printf(gettext("\tscrub started on   %s"),
+			(void) printf(gettext("scrub paused since %s"),
+			    ctime(&pause));
+			(void) printf(gettext("\tscrub started on %s"),
 			    ctime(&start));
 		}
 	} else if (ps->pss_func == POOL_SCAN_RESILVER) {
 		(void) printf(gettext("resilver in progress since %s"),
 		    ctime(&start));
 	}
 
-	examined = ps->pss_examined ? ps->pss_examined : 1;
+	scanned = ps->pss_examined;
+	pass_scanned = ps->pss_pass_exam;
+	issued = ps->pss_issued;
+	pass_issued = ps->pss_pass_issued;
 	total = ps->pss_to_examine;
-	fraction_done = (double)examined / total;
 
-	/* elapsed time for this pass */
+	/* we are only done with a block once we have issued the IO for it */
+	fraction_done = (double)issued / total;
+
+	/* elapsed time for this pass, rounding up to 1 if it's 0 */
 	elapsed = time(NULL) - ps->pss_pass_start;
 	elapsed -= ps->pss_pass_scrub_spent_paused;
-	elapsed = elapsed ? elapsed : 1;
-	pass_exam = ps->pss_pass_exam ? ps->pss_pass_exam : 1;
-	rate = pass_exam / elapsed;
-	rate = rate ? rate : 1;
-	mins_left = ((total - examined) / rate) / 60;
-	hours_left = mins_left / 60;
-
-	zfs_nicebytes(examined, examined_buf, sizeof (examined_buf));
+	elapsed = (elapsed != 0) ? elapsed : 1;
+
+	scan_rate = pass_scanned / elapsed;
+	issue_rate = pass_issued / elapsed;
+	total_secs_left = (issue_rate != 0) ?
+	    ((total - issued) / issue_rate) : UINT64_MAX;
+
+	days_left = total_secs_left / 60 / 60 / 24;
+	hours_left = (total_secs_left / 60 / 60) % 24;
+	mins_left = (total_secs_left / 60) % 60;
+	secs_left = (total_secs_left % 60);
+
+	/* format all of the numbers we will be reporting */
+	zfs_nicebytes(scanned, scanned_buf, sizeof (scanned_buf));
+	zfs_nicebytes(issued, issued_buf, sizeof (issued_buf));
 	zfs_nicebytes(total, total_buf, sizeof (total_buf));
+	zfs_nicebytes(scan_rate, srate_buf, sizeof (srate_buf));
+	zfs_nicebytes(issue_rate, irate_buf, sizeof (irate_buf));
 
-	/*
-	 * do not print estimated time if hours_left is more than 30 days
-	 * or we have a paused scrub
-	 */
+	/* do not print estimated time if we have a paused scrub */
 	if (pause == 0) {
-		zfs_nicebytes(rate, rate_buf, sizeof (rate_buf));
-		(void) printf(gettext("\t%s scanned out of %s at %s/s"),
-		    examined_buf, total_buf, rate_buf);
-		if (hours_left < (30 * 24)) {
-			(void) printf(gettext(", %lluh%um to go\n"),
-			    (u_longlong_t)hours_left, (uint_t)(mins_left % 60));
-		} else {
-			(void) printf(gettext(
-			    ", (scan is slow, no estimated time)\n"));
-		}
+		(void) printf(gettext("\t%s scanned at %s/s, "
+		    "%s issued at %s/s, %s total\n"),
+		    scanned_buf, srate_buf, issued_buf, irate_buf, total_buf);
 	} else {
-		(void) printf(gettext("\t%s scanned out of %s\n"),
-		    examined_buf, total_buf);
+		(void) printf(gettext("\t%s scanned, %s issued, %s total\n"),
+		    scanned_buf, issued_buf, total_buf);
 	}
 
 	if (ps->pss_func == POOL_SCAN_RESILVER) {
-		(void) printf(gettext("\t%s resilvered, %.2f%% done\n"),
+		(void) printf(gettext("\t%s resilvered, %.2f%% done"),
 		    processed_buf, 100 * fraction_done);
 	} else if (ps->pss_func == POOL_SCAN_SCRUB) {
-		(void) printf(gettext("\t%s repaired, %.2f%% done\n"),
+		(void) printf(gettext("\t%s repaired, %.2f%% done"),
 		    processed_buf, 100 * fraction_done);
 	}
+
+	if (pause == 0) {
+		if (issue_rate >= 10 * 1024 * 1024) {
+			(void) printf(gettext(", %llu days "
+			    "%02llu:%02llu:%02llu to go\n"),
+			    (u_longlong_t)days_left, (u_longlong_t)hours_left,
+			    (u_longlong_t)mins_left, (u_longlong_t)secs_left);
+		} else {
+			(void) printf(gettext(", no estimated "
+			    "completion time\n"));
+		}
+	} else {
+		(void) printf(gettext("\n"));
+	}
 }
 
 static void

include/sys/arc.h

@@ -66,11 +66,11 @@ typedef struct arc_prune arc_prune_t;
  * while transforming data into its desired format - specifically, when
  * decrypting, the key may not be present, or the HMAC may not be correct
  * which signifies deliberate tampering with the on-disk state
- * (assuming that the checksum was correct). The "error" parameter will be
- * nonzero in this case, even if there is no associated zio.
+ * (assuming that the checksum was correct). If any error occurs, the "buf"
+ * parameter will be NULL.
  */
-typedef void arc_read_done_func_t(zio_t *zio, int error, arc_buf_t *buf,
-    void *private);
+typedef void arc_read_done_func_t(zio_t *zio, const zbookmark_phys_t *zb,
+    const blkptr_t *bp, arc_buf_t *buf, void *private);
 typedef void arc_write_done_func_t(zio_t *zio, arc_buf_t *buf, void *private);
 typedef void arc_prune_func_t(int64_t bytes, void *private);
 
@@ -106,44 +106,45 @@ typedef enum arc_flags
 	ARC_FLAG_CACHED			= 1 << 3,	/* I/O was in cache */
 	ARC_FLAG_L2CACHE		= 1 << 4,	/* cache in L2ARC */
 	ARC_FLAG_PREDICTIVE_PREFETCH	= 1 << 5,	/* I/O from zfetch */
+	ARC_FLAG_PRESCIENT_PREFETCH	= 1 << 6,	/* long min lifespan */
 
 	/*
 	 * Private ARC flags.  These flags are private ARC only flags that
 	 * will show up in b_flags in the arc_hdr_buf_t. These flags should
 	 * only be set by ARC code.
 	 */
-	ARC_FLAG_IN_HASH_TABLE		= 1 << 6,	/* buffer is hashed */
-	ARC_FLAG_IO_IN_PROGRESS		= 1 << 7,	/* I/O in progress */
-	ARC_FLAG_IO_ERROR		= 1 << 8,	/* I/O failed for buf */
-	ARC_FLAG_INDIRECT		= 1 << 9,	/* indirect block */
+	ARC_FLAG_IN_HASH_TABLE		= 1 << 7,	/* buffer is hashed */
+	ARC_FLAG_IO_IN_PROGRESS		= 1 << 8,	/* I/O in progress */
+	ARC_FLAG_IO_ERROR		= 1 << 9,	/* I/O failed for buf */
+	ARC_FLAG_INDIRECT		= 1 << 10,	/* indirect block */
 	/* Indicates that block was read with ASYNC priority. */
-	ARC_FLAG_PRIO_ASYNC_READ	= 1 << 10,
-	ARC_FLAG_L2_WRITING		= 1 << 11,	/* write in progress */
-	ARC_FLAG_L2_EVICTED		= 1 << 12,	/* evicted during I/O */
-	ARC_FLAG_L2_WRITE_HEAD		= 1 << 13,	/* head of write list */
+	ARC_FLAG_PRIO_ASYNC_READ	= 1 << 11,
+	ARC_FLAG_L2_WRITING		= 1 << 12,	/* write in progress */
+	ARC_FLAG_L2_EVICTED		= 1 << 13,	/* evicted during I/O */
+	ARC_FLAG_L2_WRITE_HEAD		= 1 << 14,	/* head of write list */
 	/*
 	 * Encrypted or authenticated on disk (may be plaintext in memory).
 	 * This header has b_crypt_hdr allocated. Does not include indirect
 	 * blocks with checksums of MACs which will also have their X
 	 * (encrypted) bit set in the bp.
 	 */
-	ARC_FLAG_PROTECTED		= 1 << 14,
+	ARC_FLAG_PROTECTED		= 1 << 15,
 	/* data has not been authenticated yet */
-	ARC_FLAG_NOAUTH			= 1 << 15,
+	ARC_FLAG_NOAUTH			= 1 << 16,
 	/* indicates that the buffer contains metadata (otherwise, data) */
-	ARC_FLAG_BUFC_METADATA		= 1 << 16,
+	ARC_FLAG_BUFC_METADATA		= 1 << 17,
 
 	/* Flags specifying whether optional hdr struct fields are defined */
-	ARC_FLAG_HAS_L1HDR		= 1 << 17,
-	ARC_FLAG_HAS_L2HDR		= 1 << 18,
+	ARC_FLAG_HAS_L1HDR		= 1 << 18,
+	ARC_FLAG_HAS_L2HDR		= 1 << 19,
 
 	/*
 	 * Indicates the arc_buf_hdr_t's b_pdata matches the on-disk data.
 	 * This allows the l2arc to use the blkptr's checksum to verify
 	 * the data without having to store the checksum in the hdr.
 	 */
-	ARC_FLAG_COMPRESSED_ARC		= 1 << 19,
-	ARC_FLAG_SHARED_DATA		= 1 << 20,
+	ARC_FLAG_COMPRESSED_ARC		= 1 << 20,
+	ARC_FLAG_SHARED_DATA		= 1 << 21,
 
 	/*
 	 * The arc buffer's compression mode is stored in the top 7 bits of the

include/sys/dsl_pool.h

@@ -80,6 +80,7 @@ typedef struct zfs_blkstat {
 
 typedef struct zfs_all_blkstats {
 	zfs_blkstat_t	zab_type[DN_MAX_LEVELS + 1][DMU_OT_TOTAL + 1];
+	kmutex_t	zab_lock;
 } zfs_all_blkstats_t;