Add DDT prune command

Requires the new 'flat' physical data which has the start
time for a class entry.

The amount to prune can be based on a target percentage of
the unique entries or based on the age (i.e., every entry
older than N days).

Sponsored-by: Klara, Inc.
Sponsored-by: iXsystems, Inc.

Signed-off-by: Don Brady <don.brady@klarasystems.com>

cmd/zdb/zdb.c

@@ -2045,7 +2045,7 @@ dump_all_ddts(spa_t *spa)
 
 	for (enum zio_checksum c = 0; c < ZIO_CHECKSUM_FUNCTIONS; c++) {
 		ddt_t *ddt = spa->spa_ddt[c];
-		if (!ddt)
+		if (!ddt || ddt->ddt_version == DDT_VERSION_UNCONFIGURED)
 			continue;
 		for (ddt_type_t type = 0; type < DDT_TYPES; type++) {
 			for (ddt_class_t class = 0; class < DDT_CLASSES;
@@ -2072,6 +2072,32 @@ dump_all_ddts(spa_t *spa)
 	}
 
 	dump_dedup_ratio(&dds_total);
+
+	/*
+	 * Dump a histogram of unique class entry age
+	 */
+	if (dump_opt['D'] == 3 && getenv("ZDB_DDT_UNIQUE_AGE_HIST") != NULL) {
+		ddt_age_histo_t histogram;
+
+		(void) printf("DDT walk unique, building age histogram...\n");
+		ddt_prune_walk(spa, 0, &histogram);
+
+		/*
+		 * print out histogram for unique entry class birth
+		 */
+		if (histogram.dah_entries > 0) {
+			(void) printf("%5s  %9s  %4s\n",
+			    "age", "blocks", "amnt");
+			(void) printf("%5s  %9s  %4s\n",
+			    "-----", "---------", "----");
+			for (int i = 0; i < HIST_BINS; i++) {
+				(void) printf("%5d  %9d %4d%%\n", 1 << i,
+				    (int)histogram.dah_age_histo[i],
+				    (int)((histogram.dah_age_histo[i] * 100) /
+				    histogram.dah_entries));
+			}
+		}
+	}
 }
 
 static void
@@ -5749,12 +5775,17 @@ zdb_count_block(zdb_cb_t *zcb, zilog_t *zilog, const blkptr_t *bp,
 		ddt_entry_t *dde = ddt_lookup(ddt, bp);
 
 		/*
-		 * ddt_lookup() can only return NULL if this block didn't exist
+		 * ddt_lookup() can return NULL if this block didn't exist
 		 * in the DDT and creating it would take the DDT over its
 		 * quota. Since we got the block from disk, it must exist in
-		 * the DDT, so this can't happen.
+		 * the DDT, so this can't happen. However, when unique entries
+		 * are pruned, the dedup bit can be set with no corresponding
+		 * entry in the DDT.
 		 */
-		VERIFY3P(dde, !=, NULL);
+		if (dde == NULL) {
+			ddt_exit(ddt);
+			goto skipped;
+		}
 
 		/* Get the phys for this variant */
 		ddt_phys_variant_t v = ddt_phys_select(ddt, dde, bp);
@@ -5774,8 +5805,8 @@ zdb_count_block(zdb_cb_t *zcb, zilog_t *zilog, const blkptr_t *bp,
 			    (void *)(((uintptr_t)dde->dde_io) | (1 << v));
 
 		/* Consume a reference for this block. */
-		VERIFY3U(ddt_phys_total_refcnt(ddt, dde->dde_phys), >, 0);
-		ddt_phys_decref(dde->dde_phys, v);
+		if (ddt_phys_total_refcnt(ddt, dde->dde_phys) > 0)
+			ddt_phys_decref(dde->dde_phys, v);
 
 		/*
 		 * If this entry has a single flat phys, it may have been
@@ -5864,6 +5895,7 @@ zdb_count_block(zdb_cb_t *zcb, zilog_t *zilog, const blkptr_t *bp,
 		}
 	}
 
+skipped:
 	for (i = 0; i < 4; i++) {
 		int l = (i < 2) ? BP_GET_LEVEL(bp) : ZB_TOTAL;
 		int t = (i & 1) ? type : ZDB_OT_TOTAL;
@@ -8138,7 +8170,7 @@ dump_mos_leaks(spa_t *spa)
 
 	for (uint64_t c = 0; c < ZIO_CHECKSUM_FUNCTIONS; c++) {
 		ddt_t *ddt = spa->spa_ddt[c];
-		if (!ddt)
+		if (!ddt || ddt->ddt_version == DDT_VERSION_UNCONFIGURED)
 			continue;
 
 		/* DDT store objects */
@@ -8150,11 +8182,14 @@ dump_mos_leaks(spa_t *spa)
 		}
 
 		/* FDT container */
-		mos_obj_refd(ddt->ddt_dir_object);
+		if (ddt->ddt_version == DDT_VERSION_FDT)
+			mos_obj_refd(ddt->ddt_dir_object);
 
 		/* FDT log objects */
-		mos_obj_refd(ddt->ddt_log[0].ddl_object);
-		mos_obj_refd(ddt->ddt_log[1].ddl_object);
+		if (ddt->ddt_flags & DDT_FLAG_LOG) {
+			mos_obj_refd(ddt->ddt_log[0].ddl_object);
+			mos_obj_refd(ddt->ddt_log[1].ddl_object);
+		}
 	}
 
 	if (spa->spa_brt != NULL) {

cmd/zpool/zpool_main.c

@@ -130,6 +130,8 @@ static int zpool_do_version(int, char **);
 
 static int zpool_do_wait(int, char **);
 
+static int zpool_do_ddt_prune(int, char **);
+
 static int zpool_do_help(int argc, char **argv);
 
 static zpool_compat_status_t zpool_do_load_compat(
@@ -170,6 +172,7 @@ typedef enum {
 	HELP_CLEAR,
 	HELP_CREATE,
 	HELP_CHECKPOINT,
+	HELP_DDT_PRUNE,
 	HELP_DESTROY,
 	HELP_DETACH,
 	HELP_EXPORT,
@@ -426,6 +429,8 @@ static zpool_command_t command_table[] = {
 	{ "sync",	zpool_do_sync,		HELP_SYNC		},
 	{ NULL },
 	{ "wait",	zpool_do_wait,		HELP_WAIT		},
+	{ NULL },
+	{ "ddtprune",	zpool_do_ddt_prune,	HELP_DDT_PRUNE		},
 };
 
 #define	NCOMMAND	(ARRAY_SIZE(command_table))
@@ -545,6 +550,8 @@ get_usage(zpool_help_t idx)
 	case HELP_WAIT:
 		return (gettext("\twait [-Hp] [-T d|u] [-t <activity>[,...]] "
 		    "<pool> [interval]\n"));
+	case HELP_DDT_PRUNE:
+		return (gettext("\tddtprune -d|-p <amount> <pool>\n"));
 	default:
 		__builtin_unreachable();
 	}
@@ -13342,6 +13349,88 @@ found:;
 	return (error);
 }
 
+/*
+ * zpool ddtprune -d|-p <amount> <pool>
+ *
+ *       -d <days>	Prune entries <days> old and older
+ *       -p <percent>	Prune <percent> amount of entries
+ *
+ * Prune single reference entries from DDT to satisfy the amount specified.
+ */
+int
+zpool_do_ddt_prune(int argc, char **argv)
+{
+	zpool_ddt_prune_unit_t unit = ZPOOL_DDT_PRUNE_NONE;
+	uint64_t amount = 0;
+	zpool_handle_t *zhp;
+	char *endptr;
+	int c;
+
+	while ((c = getopt(argc, argv, "d:p:")) != -1) {
+		switch (c) {
+		case 'd':
+			if (unit == ZPOOL_DDT_PRUNE_PERCENTAGE) {
+				(void) fprintf(stderr, gettext("-d cannot be "
+				    "combined with -p option\n"));
+				usage(B_FALSE);
+			}
+			errno = 0;
+			amount = strtoull(optarg, &endptr, 0);
+			if (errno != 0 || *endptr != '\0' || amount == 0) {
+				(void) fprintf(stderr,
+				    gettext("invalid days value\n"));
+				usage(B_FALSE);
+			}
+			amount *= 86400;	/* convert days to seconds */
+			unit = ZPOOL_DDT_PRUNE_AGE;
+			break;
+		case 'p':
+			if (unit == ZPOOL_DDT_PRUNE_AGE) {
+				(void) fprintf(stderr, gettext("-p cannot be "
+				    "combined with -d option\n"));
+				usage(B_FALSE);
+			}
+			errno = 0;
+			amount = strtoull(optarg, &endptr, 0);
+			if (errno != 0 || *endptr != '\0' ||
+			    amount == 0 || amount > 100) {
+				(void) fprintf(stderr,
+				    gettext("invalid percentage value\n"));
+				usage(B_FALSE);
+			}
+			unit = ZPOOL_DDT_PRUNE_PERCENTAGE;
+			break;
+		case '?':
+			(void) fprintf(stderr, gettext("invalid option '%c'\n"),
+			    optopt);
+			usage(B_FALSE);
+		}
+	}
+	argc -= optind;
+	argv += optind;
+
+	if (unit == ZPOOL_DDT_PRUNE_NONE) {
+		(void) fprintf(stderr,
+		    gettext("missing amount option (-d|-p <value>)\n"));
+		usage(B_FALSE);
+	} else if (argc < 1) {
+		(void) fprintf(stderr, gettext("missing pool argument\n"));
+		usage(B_FALSE);
+	} else if (argc > 1) {
+		(void) fprintf(stderr, gettext("too many arguments\n"));
+		usage(B_FALSE);
+	}
+	zhp = zpool_open(g_zfs, argv[0]);
+	if (zhp == NULL)
+		return (-1);
+
+	int error = zpool_ddt_prune(zhp, unit, amount);
+
+	zpool_close(zhp);
+
+	return (error);
+}
+
 static int
 find_command_idx(const char *command, int *idx)
 {

cmd/ztest.c

@@ -276,6 +276,8 @@ extern unsigned long zio_decompress_fail_fraction;
 extern unsigned long zfs_reconstruct_indirect_damage_fraction;
 extern uint64_t raidz_expand_max_reflow_bytes;
 extern uint_t raidz_expand_pause_point;
+extern boolean_t ddt_prune_artificial_age;
+extern boolean_t ddt_dump_prune_histogram;
 
 
 static ztest_shared_opts_t *ztest_shared_opts;
@@ -446,6 +448,7 @@ ztest_func_t ztest_fletcher;
 ztest_func_t ztest_fletcher_incr;
 ztest_func_t ztest_verify_dnode_bt;
 ztest_func_t ztest_pool_prefetch_ddt;
+ztest_func_t ztest_ddt_prune;
 
 static uint64_t zopt_always = 0ULL * NANOSEC;		/* all the time */
 static uint64_t zopt_incessant = 1ULL * NANOSEC / 10;	/* every 1/10 second */
@@ -502,6 +505,7 @@ static ztest_info_t ztest_info[] = {
 	ZTI_INIT(ztest_fletcher_incr, 1, &zopt_rarely),
 	ZTI_INIT(ztest_verify_dnode_bt, 1, &zopt_sometimes),
 	ZTI_INIT(ztest_pool_prefetch_ddt, 1, &zopt_rarely),
+	ZTI_INIT(ztest_ddt_prune, 1, &zopt_rarely),
 };
 
 #define	ZTEST_FUNCS	(sizeof (ztest_info) / sizeof (ztest_info_t))
@@ -7288,6 +7292,17 @@ ztest_trim(ztest_ds_t *zd, uint64_t id)
 	mutex_exit(&ztest_vdev_lock);
 }
 
+void
+ztest_ddt_prune(ztest_ds_t *zd, uint64_t id)
+{
+	(void) zd, (void) id;
+
+	spa_t *spa = ztest_spa;
+	uint64_t pct = ztest_random(15) + 1;
+
+	(void) ddt_prune_unique_entries(spa, ZPOOL_DDT_PRUNE_PERCENTAGE, pct);
+}
+
 /*
  * Verify pool integrity by running zdb.
  */
@@ -7469,6 +7484,13 @@ ztest_resume_thread(void *arg)
 {
 	spa_t *spa = arg;
 
+	/*
+	 * Synthesize aged DDT entries for ddt prune testing
+	 */
+	ddt_prune_artificial_age = B_TRUE;
+	if (ztest_opts.zo_verbose >= 3)
+		ddt_dump_prune_histogram = B_TRUE;
+
 	while (!ztest_exiting) {
 		if (spa_suspended(spa))
 			ztest_resume(spa);
@@ -8587,6 +8609,12 @@ ztest_init(ztest_shared_t *zs)
 		if (i == SPA_FEATURE_LOG_SPACEMAP && ztest_random(4) == 0)
 			continue;
 
+		/*
+		 * split 50/50 between legacy and fast dedup
+		 */
+		if (i == SPA_FEATURE_FAST_DEDUP && ztest_random(2) != 0)
+			continue;
+
 		VERIFY3S(-1, !=, asprintf(&buf, "feature@%s",
 		    spa_feature_table[i].fi_uname));
 		fnvlist_add_uint64(props, buf, 0);

contrib/debian/openzfs-zfsutils.install

@@ -100,6 +100,7 @@ usr/share/man/man8/zpool-clear.8
 usr/share/man/man8/zpool-create.8
 usr/share/man/man8/zpool-destroy.8
 usr/share/man/man8/zpool-detach.8
+usr/share/man/man8/zpool-ddtprune.8
 usr/share/man/man8/zpool-events.8
 usr/share/man/man8/zpool-export.8
 usr/share/man/man8/zpool-get.8

include/libzfs.h

@@ -305,6 +305,9 @@ _LIBZFS_H int zpool_reopen_one(zpool_handle_t *, void *);
 
 _LIBZFS_H int zpool_sync_one(zpool_handle_t *, void *);
 
+_LIBZFS_H int zpool_ddt_prune(zpool_handle_t *, zpool_ddt_prune_unit_t,
+    uint64_t);
+
 _LIBZFS_H int zpool_vdev_online(zpool_handle_t *, const char *, int,
     vdev_state_t *);
 _LIBZFS_H int zpool_vdev_offline(zpool_handle_t *, const char *, boolean_t);

include/libzfs_core.h

@@ -161,6 +161,9 @@ _LIBZFS_CORE_H int lzc_set_vdev_prop(const char *, nvlist_t *, nvlist_t **);
 
 _LIBZFS_CORE_H int lzc_scrub(zfs_ioc_t, const char *, nvlist_t *, nvlist_t **);
 
+_LIBZFS_CORE_H int lzc_ddt_prune(const char *, zpool_ddt_prune_unit_t,
+    uint64_t);
+
 #ifdef	__cplusplus
 }
 #endif

include/sys/ddt.h

@@ -405,6 +405,9 @@ extern int ddt_walk(spa_t *spa, ddt_bookmark_t *ddb,
 
 extern boolean_t ddt_addref(spa_t *spa, const blkptr_t *bp);
 
+extern int ddt_prune_unique_entries(spa_t *spa, zpool_ddt_prune_unit_t unit,
+    uint64_t amount);
+
 #ifdef	__cplusplus
 }
 #endif