Tiered early abort, zstd edition

It turns out that "do LZ4 and zstd-1 both fail" is a great heuristic
for "don't even bother trying higher zstd tiers".

By way of illustration:
$ cat /incompress | mbuffer | zfs recv -o compression=zstd-12 evenfaster/lowcomp_1M_zstd12_normal
summary: 39.8 GiByte in  3min 40.2sec - average of  185 MiB/s
$ echo 3 | sudo tee /sys/module/zzstd/parameters/zstd_lz4_pass
3
$ cat /incompress | mbuffer -m 4G | zfs recv -o compression=zstd-12 evenfaster/lowcomp_1M_zstd12_patched
summary: 39.8 GiByte in 48.6sec - average of  839 MiB/s
$ sudo zfs list -p -o name,used,lused,ratio evenfaster/lowcomp_1M_zstd12_normal evenfaster/lowcomp_1M_zstd12_patched
NAME                                         USED        LUSED  RATIO
evenfaster/lowcomp_1M_zstd12_normal   39549931520  42721221632   1.08
evenfaster/lowcomp_1M_zstd12_patched  39626399744  42721217536   1.07
$ python3 -c "print(39626399744 - 39549931520)"
76468224
$

I'll take 76 MB out of 42 GB for > 4x speedup.

Reviewed-by: Allan Jude <allan@klarasystems.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: George Melikov <mail@gmelikov.ru>
Reviewed-by: Kjeld Schouten <kjeld@schouten-lebbing.nl>
Reviewed-by: Ahelenia Ziemiańska <nabijaczleweli@nabijaczleweli.xyz>
Signed-off-by: Rich Ercolani <rincebrain@gmail.com>
Closes #13244

include/sys/zstd/zstd.h

@@ -78,6 +78,8 @@ typedef struct zfs_zstd_meta {
  * kstat helper macros
  */
 #define	ZSTDSTAT(stat)		(zstd_stats.stat.value.ui64)
+#define	ZSTDSTAT_ZERO(stat)	\
+	(atomic_store_64(&zstd_stats.stat.value.ui64, 0))
 #define	ZSTDSTAT_ADD(stat, val) \
 	atomic_add_64(&zstd_stats.stat.value.ui64, (val))
 #define	ZSTDSTAT_SUB(stat, val) \
@@ -90,6 +92,8 @@ void zstd_fini(void);
 
 size_t zfs_zstd_compress(void *s_start, void *d_start, size_t s_len,
     size_t d_len, int level);
+size_t zfs_zstd_compress_wrap(void *s_start, void *d_start, size_t s_len,
+    size_t d_len, int level);
 int zfs_zstd_get_level(void *s_start, size_t s_len, uint8_t *level);
 int zfs_zstd_decompress_level(void *s_start, void *d_start, size_t s_len,
     size_t d_len, uint8_t *level);

man/man4/zfs.4

@@ -2129,6 +2129,14 @@ However, if there are fewer than
 metaslabs in the vdev, this functionality is disabled.
 This ensures that we don't set aside an unreasonable amount of space for the ZIL.
 .
+.It Sy zfs_zstd_earlyabort_pass Ns = Ns Sy 1 Pq int
+Whether heuristic for detection of incompressible data with zstd levels >= 3
+using LZ4 and zstd-1 passes is enabled.
+.
+.It Sy zfs_zstd_abort_size Ns = Ns Sy 131072 Pq int
+Minimal uncompressed size (inclusive) of a record before the early abort
+heuristic will be attempted.
+.
 .It Sy zio_deadman_log_all Ns = Ns Sy 0 Ns | Ns 1 Pq int
 If non-zero, the zio deadman will produce debugging messages
 .Pq see Sy zfs_dbgmsg_enable

module/zfs/zio_compress.c

@@ -66,7 +66,7 @@ zio_compress_info_t zio_compress_table[ZIO_COMPRESS_FUNCTIONS] = {
 	{"gzip-9",	9,	gzip_compress,	gzip_decompress, NULL},
 	{"zle",		64,	zle_compress,	zle_decompress, NULL},
 	{"lz4",		0,	lz4_compress_zfs, lz4_decompress_zfs, NULL},
-	{"zstd",	ZIO_ZSTD_LEVEL_DEFAULT,	zfs_zstd_compress,
+	{"zstd",	ZIO_ZSTD_LEVEL_DEFAULT,	zfs_zstd_compress_wrap,
 	    zfs_zstd_decompress, zfs_zstd_decompress_level},
 };
 

module/zstd/zfs_zstd.c

@@ -50,6 +50,10 @@
 #include "lib/zstd.h"
 #include "lib/common/zstd_errors.h"
 
+static int zstd_earlyabort_pass = 1;
+static int zstd_cutoff_level = ZIO_ZSTD_LEVEL_3;
+static unsigned int zstd_abort_size = (128 * 1024);
+
 static kstat_t *zstd_ksp = NULL;
 
 typedef struct zstd_stats {
@@ -62,6 +66,21 @@ typedef struct zstd_stats {
 	kstat_named_t	zstd_stat_dec_header_inval;
 	kstat_named_t	zstd_stat_com_fail;
 	kstat_named_t	zstd_stat_dec_fail;
+	/*
+	 * LZ4 first-pass early abort verdict
+	 */
+	kstat_named_t	zstd_stat_lz4pass_allowed;
+	kstat_named_t	zstd_stat_lz4pass_rejected;
+	/*
+	 * zstd-1 second-pass early abort verdict
+	 */
+	kstat_named_t	zstd_stat_zstdpass_allowed;
+	kstat_named_t	zstd_stat_zstdpass_rejected;
+	/*
+	 * We excluded this from early abort for some reason
+	 */
+	kstat_named_t	zstd_stat_passignored;
+	kstat_named_t	zstd_stat_passignored_size;
 	kstat_named_t	zstd_stat_buffers;
 	kstat_named_t	zstd_stat_size;
 } zstd_stats_t;
@@ -76,10 +95,44 @@ static zstd_stats_t zstd_stats = {
 	{ "decompress_header_invalid",	KSTAT_DATA_UINT64 },
 	{ "compress_failed",		KSTAT_DATA_UINT64 },
 	{ "decompress_failed",		KSTAT_DATA_UINT64 },
+	{ "lz4pass_allowed",		KSTAT_DATA_UINT64 },
+	{ "lz4pass_rejected",		KSTAT_DATA_UINT64 },
+	{ "zstdpass_allowed",		KSTAT_DATA_UINT64 },
+	{ "zstdpass_rejected",		KSTAT_DATA_UINT64 },
+	{ "passignored",		KSTAT_DATA_UINT64 },
+	{ "passignored_size",		KSTAT_DATA_UINT64 },
 	{ "buffers",			KSTAT_DATA_UINT64 },
 	{ "size",			KSTAT_DATA_UINT64 },
 };
 
+#ifdef _KERNEL
+static int
+kstat_zstd_update(kstat_t *ksp, int rw)
+{
+	ASSERT(ksp != NULL);
+
+	if (rw == KSTAT_WRITE && ksp == zstd_ksp) {
+		ZSTDSTAT_ZERO(zstd_stat_alloc_fail);
+		ZSTDSTAT_ZERO(zstd_stat_alloc_fallback);
+		ZSTDSTAT_ZERO(zstd_stat_com_alloc_fail);
+		ZSTDSTAT_ZERO(zstd_stat_dec_alloc_fail);
+		ZSTDSTAT_ZERO(zstd_stat_com_inval);
+		ZSTDSTAT_ZERO(zstd_stat_dec_inval);
+		ZSTDSTAT_ZERO(zstd_stat_dec_header_inval);
+		ZSTDSTAT_ZERO(zstd_stat_com_fail);
+		ZSTDSTAT_ZERO(zstd_stat_dec_fail);
+		ZSTDSTAT_ZERO(zstd_stat_lz4pass_allowed);
+		ZSTDSTAT_ZERO(zstd_stat_lz4pass_rejected);
+		ZSTDSTAT_ZERO(zstd_stat_zstdpass_allowed);
+		ZSTDSTAT_ZERO(zstd_stat_zstdpass_rejected);
+		ZSTDSTAT_ZERO(zstd_stat_passignored);
+		ZSTDSTAT_ZERO(zstd_stat_passignored_size);
+	}
+
+	return (0);
+}
+#endif
+
 /* Enums describing the allocator type specified by kmem_type in zstd_kmem */
 enum zstd_kmem_type {
 	ZSTD_KMEM_UNKNOWN = 0,
@@ -377,6 +430,64 @@ zstd_enum_to_level(enum zio_zstd_levels level, int16_t *zstd_level)
 }
 
 
+size_t
+zfs_zstd_compress_wrap(void *s_start, void *d_start, size_t s_len, size_t d_len,
+    int level)
+{
+	int16_t zstd_level;
+	if (zstd_enum_to_level(level, &zstd_level)) {
+		ZSTDSTAT_BUMP(zstd_stat_com_inval);
+		return (s_len);
+	}
+	/*
+	 * A zstd early abort heuristic.
+	 *
+	 * - Zeroth, if this is <= zstd-3, or < zstd_abort_size (currently
+	 *   128k), don't try any of this, just go.
+	 *   (because experimentally that was a reasonable cutoff for a perf win
+	 *   with tiny ratio change)
+	 * - First, we try LZ4 compression, and if it doesn't early abort, we
+	 *   jump directly to whatever compression level we intended to try.
+	 * - Second, we try zstd-1 - if that errors out (usually, but not
+	 *   exclusively, if it would overflow), we give up early.
+	 *
+	 *   If it works, instead we go on and compress anyway.
+	 *
+	 * Why two passes? LZ4 alone gets you a lot of the way, but on highly
+	 * compressible data, it was losing up to 8.5% of the compressed
+	 * savings versus no early abort, and all the zstd-fast levels are
+	 * worse indications on their own than LZ4, and don't improve the LZ4
+	 * pass noticably if stacked like this.
+	 */
+	size_t actual_abort_size = zstd_abort_size;
+	if (zstd_earlyabort_pass > 0 && zstd_level >= zstd_cutoff_level &&
+	    s_len >= actual_abort_size) {
+		int pass_len = 1;
+		pass_len = lz4_compress_zfs(s_start, d_start, s_len, d_len, 0);
+		if (pass_len < d_len) {
+			ZSTDSTAT_BUMP(zstd_stat_lz4pass_allowed);
+			goto keep_trying;
+		}
+		ZSTDSTAT_BUMP(zstd_stat_lz4pass_rejected);
+
+		pass_len = zfs_zstd_compress(s_start, d_start, s_len, d_len,
+		    ZIO_ZSTD_LEVEL_1);
+		if (pass_len == s_len || pass_len <= 0 || pass_len > d_len) {
+			ZSTDSTAT_BUMP(zstd_stat_zstdpass_rejected);
+			return (s_len);
+		}
+		ZSTDSTAT_BUMP(zstd_stat_zstdpass_allowed);
+	} else {
+		ZSTDSTAT_BUMP(zstd_stat_passignored);
+		if (s_len < actual_abort_size) {
+			ZSTDSTAT_BUMP(zstd_stat_passignored_size);
+		}
+	}
+keep_trying:
+	return (zfs_zstd_compress(s_start, d_start, s_len, d_len, level));
+
+}
+
 /* Compress block using zstd */
 size_t
 zfs_zstd_compress(void *s_start, void *d_start, size_t s_len, size_t d_len,
@@ -437,8 +548,10 @@ zfs_zstd_compress(void *s_start, void *d_start, size_t s_len, size_t d_len,
 		 * too small, that is not a failure. Everything else is a
 		 * failure, so increment the compression failure counter.
 		 */
-		if (ZSTD_getErrorCode(c_len) != ZSTD_error_dstSize_tooSmall) {
+		int err = ZSTD_getErrorCode(c_len);
+		if (err != ZSTD_error_dstSize_tooSmall) {
 			ZSTDSTAT_BUMP(zstd_stat_com_fail);
+			dprintf("Error: %s", ZSTD_getErrorString(err));
 		}
 		return (s_len);
 	}
@@ -753,6 +866,9 @@ zstd_init(void)
 	if (zstd_ksp != NULL) {
 		zstd_ksp->ks_data = &zstd_stats;
 		kstat_install(zstd_ksp);
+#ifdef _KERNEL
+		zstd_ksp->ks_update = kstat_zstd_update;
+#endif
 	}
 
 	return (0);
@@ -781,8 +897,8 @@ module_init(zstd_init);
 module_exit(zstd_fini);
 #endif
 
-EXPORT_SYMBOL(zfs_zstd_compress);
+ZFS_MODULE_PARAM(zfs, zstd_, earlyabort_pass, INT, ZMOD_RW,
-EXPORT_SYMBOL(zfs_zstd_decompress_level);
+	"Enable early abort attempts when using zstd");
-EXPORT_SYMBOL(zfs_zstd_decompress);
+ZFS_MODULE_PARAM(zfs, zstd_, abort_size, UINT, ZMOD_RW,
-EXPORT_SYMBOL(zfs_zstd_cache_reap_now);
+	"Minimal size of block to attempt early abort");
 #endif