zdb: show BRT statistics and dump its contents

cmd/zdb/zdb.c

@@ -34,6 +34,7 @@
  * Copyright (c) 2021 Allan Jude
  * Copyright (c) 2021 Toomas Soome <tsoome@me.com>
  * Copyright (c) 2023, Klara Inc.
+ * Copyright (c) 2023, Rob Norris <robn@despairlabs.com>
  */
 
 #include <stdio.h>
@@ -80,6 +81,7 @@
 #include <sys/dsl_scan.h>
 #include <sys/btree.h>
 #include <sys/brt.h>
+#include <sys/brt_impl.h>
 #include <zfs_comutil.h>
 #include <sys/zstd/zstd.h>
 
@@ -899,6 +901,8 @@ usage(void)
 	    "don't print label contents\n");
 	(void) fprintf(stderr, "        -t --txg=INTEGER             "
 	    "highest txg to use when searching for uberblocks\n");
+	(void) fprintf(stderr, "        -T --brt-stats               "
+	    "BRT statistics\n");
 	(void) fprintf(stderr, "        -u --uberblock               "
 	    "uberblock\n");
 	(void) fprintf(stderr, "        -U --cachefile=PATH          "
@@ -999,6 +1003,15 @@ zdb_nicenum(uint64_t num, char *buf, size_t buflen)
 		nicenum(num, buf, buflen);
 }
 
+static void
+zdb_nicebytes(uint64_t bytes, char *buf, size_t buflen)
+{
+	if (dump_opt['P'])
+		(void) snprintf(buf, buflen, "%llu", (longlong_t)bytes);
+	else
+		zfs_nicebytes(bytes, buf, buflen);
+}
+
 static const char histo_stars[] = "****************************************";
 static const uint64_t histo_width = sizeof (histo_stars) - 1;
 
@@ -2081,6 +2094,76 @@ dump_all_ddts(spa_t *spa)
 	dump_dedup_ratio(&dds_total);
 }
 
+static void
+dump_brt(spa_t *spa)
+{
+	if (!spa_feature_is_enabled(spa, SPA_FEATURE_BLOCK_CLONING)) {
+		printf("BRT: unsupported on this pool\n");
+		return;
+	}
+
+	if (!spa_feature_is_active(spa, SPA_FEATURE_BLOCK_CLONING)) {
+		printf("BRT: empty\n");
+		return;
+	}
+
+	brt_t *brt = spa->spa_brt;
+	VERIFY(brt);
+
+	char count[32], used[32], saved[32];
+	zdb_nicebytes(brt_get_used(spa), used, sizeof (used));
+	zdb_nicebytes(brt_get_saved(spa), saved, sizeof (saved));
+	uint64_t ratio = brt_get_ratio(spa);
+	printf("BRT: used %s; saved %s; ratio %llu.%02llux\n", used, saved,
+	    (u_longlong_t)(ratio / 100), (u_longlong_t)(ratio % 100));
+
+	if (dump_opt['T'] < 2)
+		return;
+
+	for (uint64_t vdevid = 0; vdevid < brt->brt_nvdevs; vdevid++) {
+		brt_vdev_t *brtvd = &brt->brt_vdevs[vdevid];
+		if (brtvd == NULL)
+			continue;
+
+		if (!brtvd->bv_initiated) {
+			printf("BRT: vdev %lu: empty\n", vdevid);
+			continue;
+		}
+
+		zdb_nicenum(brtvd->bv_totalcount, count, sizeof (count));
+		zdb_nicebytes(brtvd->bv_usedspace, used, sizeof (used));
+		zdb_nicebytes(brtvd->bv_savedspace, saved, sizeof (saved));
+		printf("BRT: vdev %lu: refcnt %s; used %s; saved %s\n",
+		    vdevid, count, used, saved);
+	}
+
+	if (dump_opt['T'] < 3)
+		return;
+
+	char dva[64];
+	printf("\n%-16s %-10s\n", "DVA", "REFCNT");
+
+	for (uint64_t vdevid = 0; vdevid < brt->brt_nvdevs; vdevid++) {
+		brt_vdev_t *brtvd = &brt->brt_vdevs[vdevid];
+		if (brtvd == NULL || !brtvd->bv_initiated)
+			continue;
+
+		zap_cursor_t zc;
+		zap_attribute_t za;
+		for (zap_cursor_init(&zc, brt->brt_mos, brtvd->bv_mos_entries);
+		    zap_cursor_retrieve(&zc, &za) == 0;
+		    zap_cursor_advance(&zc)) {
+			uint64_t offset = *(uint64_t *)za.za_name;
+			uint64_t refcnt = za.za_first_integer;
+
+			snprintf(dva, sizeof (dva), "%lu:%llx", vdevid,
+			    (u_longlong_t)offset);
+			printf("%-16s %-10llu\n", dva, (u_longlong_t)refcnt);
+		}
+		zap_cursor_fini(&zc);
+	}
+}
+
 static void
 dump_dtl_seg(void *arg, uint64_t start, uint64_t size)
 {
@@ -8108,6 +8191,9 @@ dump_zpool(spa_t *spa)
 	if (dump_opt['D'])
 		dump_all_ddts(spa);
 
+	if (dump_opt['T'])
+		dump_brt(spa);
+
 	if (dump_opt['d'] > 2 || dump_opt['m'])
 		dump_metaslabs(spa);
 	if (dump_opt['M'])
@@ -8894,6 +8980,7 @@ main(int argc, char **argv)
 		{"io-stats",		no_argument,		NULL, 's'},
 		{"simulate-dedup",	no_argument,		NULL, 'S'},
 		{"txg",			required_argument,	NULL, 't'},
+		{"brt-stats",		no_argument,		NULL, 'T'},
 		{"uberblock",		no_argument,		NULL, 'u'},
 		{"cachefile",		required_argument,	NULL, 'U'},
 		{"verbose",		no_argument,		NULL, 'v'},
@@ -8907,7 +8994,7 @@ main(int argc, char **argv)
 	};
 
 	while ((c = getopt_long(argc, argv,
-	    "AbBcCdDeEFGhiI:kK:lLmMNo:Op:PqrRsSt:uU:vVx:XYyZ",
+	    "AbBcCdDeEFGhiI:kK:lLmMNo:Op:PqrRsSt:TuU:vVx:XYyZ",
 	    long_options, NULL)) != -1) {
 		switch (c) {
 		case 'b':
@@ -8929,6 +9016,7 @@ main(int argc, char **argv)
 		case 'R':
 		case 's':
 		case 'S':
+		case 'T':
 		case 'u':
 		case 'y':
 		case 'Z':

include/Makefile.am

@@ -33,6 +33,7 @@ COMMON_H = \
 	sys/bqueue.h \
 	sys/btree.h \
 	sys/brt.h \
+	sys/brt_impl.h \
 	sys/dataset_kstats.h \
 	sys/dbuf.h \
 	sys/ddt.h \

include/sys/brt_impl.h

@@ -0,0 +1,199 @@
+/*
+ * CDDL HEADER START
+ *
+ * The contents of this file are subject to the terms of the
+ * Common Development and Distribution License (the "License").
+ * You may not use this file except in compliance with the License.
+ *
+ * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
+ * or https://opensource.org/licenses/CDDL-1.0.
+ * See the License for the specific language governing permissions
+ * and limitations under the License.
+ *
+ * When distributing Covered Code, include this CDDL HEADER in each
+ * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
+ * If applicable, add the following below this CDDL HEADER, with the
+ * fields enclosed by brackets "[]" replaced with your own identifying
+ * information: Portions Copyright [yyyy] [name of copyright owner]
+ *
+ * CDDL HEADER END
+ */
+/*
+ * Copyright (c) 2020, 2021, 2022 by Pawel Jakub Dawidek
+ */
+
+#ifndef _SYS_BRT_IMPL_H
+#define	_SYS_BRT_IMPL_H
+
+#ifdef	__cplusplus
+extern "C" {
+#endif
+
+/*
+ * BRT - Block Reference Table.
+ */
+#define	BRT_OBJECT_VDEV_PREFIX	"com.fudosecurity:brt:vdev:"
+
+/*
+ * We divide each VDEV into 16MB chunks. Each chunk is represented in memory
+ * by a 16bit counter, thus 1TB VDEV requires 128kB of memory: (1TB / 16MB) * 2B
+ * Each element in this array represents how many BRT entries do we have in this
+ * chunk of storage. We always load this entire array into memory and update as
+ * needed. By having it in memory we can quickly tell (during zio_free()) if
+ * there are any BRT entries that we might need to update.
+ *
+ * This value cannot be larger than 16MB, at least as long as we support
+ * 512 byte block sizes. With 512 byte block size we can have exactly
+ * 32768 blocks in 16MB. In 32MB we could have 65536 blocks, which is one too
+ * many for a 16bit counter.
+ */
+#define	BRT_RANGESIZE	(16 * 1024 * 1024)
+_Static_assert(BRT_RANGESIZE / SPA_MINBLOCKSIZE <= UINT16_MAX,
+	"BRT_RANGESIZE is too large.");
+/*
+ * We don't want to update the whole structure every time. Maintain bitmap
+ * of dirty blocks within the regions, so that a single bit represents a
+ * block size of entcounts. For example if we have a 1PB vdev then all
+ * entcounts take 128MB of memory ((64TB / 16MB) * 2B). We can divide this
+ * 128MB array of entcounts into 32kB disk blocks, as we don't want to update
+ * the whole 128MB on disk when we have updated only a single entcount.
+ * We maintain a bitmap where each 32kB disk block within 128MB entcounts array
+ * is represented by a single bit. This gives us 4096 bits. A set bit in the
+ * bitmap means that we had a change in at least one of the 16384 entcounts
+ * that reside on a 32kB disk block (32kB / sizeof (uint16_t)).
+ */
+#define	BRT_BLOCKSIZE	(32 * 1024)
+#define	BRT_RANGESIZE_TO_NBLOCKS(size)					\
+	(((size) - 1) / BRT_BLOCKSIZE / sizeof (uint16_t) + 1)
+
+#define	BRT_LITTLE_ENDIAN	0
+#define	BRT_BIG_ENDIAN		1
+#ifdef _ZFS_LITTLE_ENDIAN
+#define	BRT_NATIVE_BYTEORDER		BRT_LITTLE_ENDIAN
+#define	BRT_NON_NATIVE_BYTEORDER	BRT_BIG_ENDIAN
+#else
+#define	BRT_NATIVE_BYTEORDER		BRT_BIG_ENDIAN
+#define	BRT_NON_NATIVE_BYTEORDER	BRT_LITTLE_ENDIAN
+#endif
+
+typedef struct brt_vdev_phys {
+	uint64_t	bvp_mos_entries;
+	uint64_t	bvp_size;
+	uint64_t	bvp_byteorder;
+	uint64_t	bvp_totalcount;
+	uint64_t	bvp_rangesize;
+	uint64_t	bvp_usedspace;
+	uint64_t	bvp_savedspace;
+} brt_vdev_phys_t;
+
+typedef struct brt_vdev {
+	/*
+	 * VDEV id.
+	 */
+	uint64_t	bv_vdevid;
+	/*
+	 * Is the structure initiated?
+	 * (bv_entcount and bv_bitmap are allocated?)
+	 */
+	boolean_t	bv_initiated;
+	/*
+	 * Object number in the MOS for the entcount array and brt_vdev_phys.
+	 */
+	uint64_t	bv_mos_brtvdev;
+	/*
+	 * Object number in the MOS for the entries table.
+	 */
+	uint64_t	bv_mos_entries;
+	/*
+	 * Entries to sync.
+	 */
+	avl_tree_t	bv_tree;
+	/*
+	 * Does the bv_entcount[] array needs byte swapping?
+	 */
+	boolean_t	bv_need_byteswap;
+	/*
+	 * Number of entries in the bv_entcount[] array.
+	 */
+	uint64_t	bv_size;
+	/*
+	 * This is the array with BRT entry count per BRT_RANGESIZE.
+	 */
+	uint16_t	*bv_entcount;
+	/*
+	 * Sum of all bv_entcount[]s.
+	 */
+	uint64_t	bv_totalcount;
+	/*
+	 * Space on disk occupied by cloned blocks (without compression).
+	 */
+	uint64_t	bv_usedspace;
+	/*
+	 * How much additional space would be occupied without block cloning.
+	 */
+	uint64_t	bv_savedspace;
+	/*
+	 * brt_vdev_phys needs updating on disk.
+	 */
+	boolean_t	bv_meta_dirty;
+	/*
+	 * bv_entcount[] needs updating on disk.
+	 */
+	boolean_t	bv_entcount_dirty;
+	/*
+	 * bv_entcount[] potentially can be a bit too big to sychronize it all
+	 * when we just changed few entcounts. The fields below allow us to
+	 * track updates to bv_entcount[] array since the last sync.
+	 * A single bit in the bv_bitmap represents as many entcounts as can
+	 * fit into a single BRT_BLOCKSIZE.
+	 * For example we have 65536 entcounts in the bv_entcount array
+	 * (so the whole array is 128kB). We updated bv_entcount[2] and
+	 * bv_entcount[5]. In that case only first bit in the bv_bitmap will
+	 * be set and we will write only first BRT_BLOCKSIZE out of 128kB.
+	 */
+	ulong_t		*bv_bitmap;
+	uint64_t	bv_nblocks;
+} brt_vdev_t;
+
+/*
+ * In-core brt
+ */
+typedef struct brt {
+	krwlock_t	brt_lock;
+	spa_t		*brt_spa;
+#define	brt_mos		brt_spa->spa_meta_objset
+	uint64_t	brt_rangesize;
+	uint64_t	brt_usedspace;
+	uint64_t	brt_savedspace;
+	avl_tree_t	brt_pending_tree[TXG_SIZE];
+	kmutex_t	brt_pending_lock[TXG_SIZE];
+	/* Sum of all entries across all bv_trees. */
+	uint64_t	brt_nentries;
+	brt_vdev_t	*brt_vdevs;
+	uint64_t	brt_nvdevs;
+} brt_t;
+
+/* Size of bre_offset / sizeof (uint64_t). */
+#define	BRT_KEY_WORDS	(1)
+
+/*
+ * In-core brt entry.
+ * On-disk we use bre_offset as the key and bre_refcount as the value.
+ */
+typedef struct brt_entry {
+	uint64_t	bre_offset;
+	uint64_t	bre_refcount;
+	avl_node_t	bre_node;
+} brt_entry_t;
+
+typedef struct brt_pending_entry {
+	blkptr_t	bpe_bp;
+	int		bpe_count;
+	avl_node_t	bpe_node;
+} brt_pending_entry_t;
+
+#ifdef	__cplusplus
+}
+#endif
+
+#endif	/* _SYS_BRT_IMPL_H */

man/man8/zdb.8

@@ -14,7 +14,7 @@
 .\" Copyright (c) 2017 Lawrence Livermore National Security, LLC.
 .\" Copyright (c) 2017 Intel Corporation.
 .\"
-.Dd June 27, 2023
+.Dd November 18, 2023
 .Dt ZDB 8
 .Os
 .
@@ -23,7 +23,7 @@
 .Nd display ZFS storage pool debugging and consistency information
 .Sh SYNOPSIS
 .Nm
-.Op Fl AbcdDFGhikLMNPsvXYy
+.Op Fl AbcdDFGhikLMNPsTvXYy
 .Op Fl e Oo Fl V Oc Oo Fl p Ar path Oc Ns …
 .Op Fl I Ar inflight-I/O-ops
 .Oo Fl o Ar var Ns = Ns Ar value Oc Ns …
@@ -403,6 +403,13 @@ Display operation counts, bandwidth, and error counts of I/O to the pool from
 Simulate the effects of deduplication, constructing a DDT and then display
 that DDT as with
 .Fl DD .
+.It Fl T , -brt-stats
+Display block reference table (BRT) statistics, including the size of uniques
+blocks cloned, the space saving as a result of cloning, and the saving ratio.
+.It Fl TT
+Display the per-vdev BRT statistics, including total references.
+.It Fl TTT
+Dump the contents of the block reference tables.
 .It Fl u , -uberblock
 Display the current uberblock.
 .El