Illumos 5027 - zfs large block support

5027 zfs large block support
Reviewed by: Alek Pinchuk <pinchuk.alek@gmail.com>
Reviewed by: George Wilson <george.wilson@delphix.com>
Reviewed by: Josef 'Jeff' Sipek <josef.sipek@nexenta.com>
Reviewed by: Richard Elling <richard.elling@richardelling.com>
Reviewed by: Saso Kiselkov <skiselkov.ml@gmail.com>
Reviewed by: Brian Behlendorf <behlendorf1@llnl.gov>
Approved by: Dan McDonald <danmcd@omniti.com>

References:
  https://www.illumos.org/issues/5027
  https://github.com/illumos/illumos-gate/commit/b515258

Porting Notes:

* Included in this patch is a tiny ISP2() cleanup in zio_init() from
Illumos 5255.

* Unlike the upstream Illumos commit this patch does not impose an
arbitrary 128K block size limit on volumes.  Volumes, like filesystems,
are limited by the zfs_max_recordsize=1M module option.

* By default the maximum record size is limited to 1M by the module
option zfs_max_recordsize.  This value may be safely increased up to
16M which is the largest block size supported by the on-disk format.
At the moment, 1M blocks clearly offer a significant performance
improvement but the benefits of going beyond this for the majority
of workloads are less clear.

* The illumos version of this patch increased DMU_MAX_ACCESS to 32M.
This was determined not to be large enough when using 16M blocks
because the zfs_make_xattrdir() function will fail (EFBIG) when
assigning a TX.  This was immediately observed under Linux because
all newly created files must have a security xattr created and
that was failing.  Therefore, we've set DMU_MAX_ACCESS to 64M.

* On 32-bit platforms a hard limit of 1M is set for blocks due
to the limited virtual address space.  We should be able to relax
this one the ABD patches are merged.

Ported-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes #354

cmd/zdb/zdb.c

@@ -2185,6 +2185,8 @@ dump_label(const char *dev)
 	(void) close(fd);
 }
 
+static uint64_t num_large_blocks;
+
 /*ARGSUSED*/
 static int
 dump_one_dir(const char *dsname, void *arg)
@@ -2197,6 +2199,8 @@ dump_one_dir(const char *dsname, void *arg)
 		(void) printf("Could not open %s, error %d\n", dsname, error);
 		return (0);
 	}
+	if (dmu_objset_ds(os)->ds_large_blocks)
+		num_large_blocks++;
 	dump_dir(os);
 	dmu_objset_disown(os, FTAG);
 	fuid_table_destroy();
@@ -2207,7 +2211,7 @@ dump_one_dir(const char *dsname, void *arg)
 /*
  * Block statistics.
  */
-#define	PSIZE_HISTO_SIZE (SPA_MAXBLOCKSIZE / SPA_MINBLOCKSIZE + 1)
+#define	PSIZE_HISTO_SIZE (SPA_OLD_MAXBLOCKSIZE / SPA_MINBLOCKSIZE + 2)
 typedef struct zdb_blkstats {
 	uint64_t zb_asize;
 	uint64_t zb_lsize;
@@ -2273,7 +2277,15 @@ zdb_count_block(zdb_cb_t *zcb, zilog_t *zilog, const blkptr_t *bp,
 		zb->zb_lsize += BP_GET_LSIZE(bp);
 		zb->zb_psize += BP_GET_PSIZE(bp);
 		zb->zb_count++;
-		zb->zb_psize_histogram[BP_GET_PSIZE(bp) >> SPA_MINBLOCKSHIFT]++;
+
+		/*
+		 * The histogram is only big enough to record blocks up to
+		 * SPA_OLD_MAXBLOCKSIZE; larger blocks go into the last,
+		 * "other", bucket.
+		 */
+		int idx = BP_GET_PSIZE(bp) >> SPA_MINBLOCKSHIFT;
+		idx = MIN(idx, SPA_OLD_MAXBLOCKSIZE / SPA_MINBLOCKSIZE + 1);
+		zb->zb_psize_histogram[idx]++;
 
 		zb->zb_gangs += BP_COUNT_GANG(bp);
 
@@ -2979,6 +2991,7 @@ dump_zpool(spa_t *spa)
 		dump_metaslab_groups(spa);
 
 	if (dump_opt['d'] || dump_opt['i']) {
+		uint64_t refcount;
 		dump_dir(dp->dp_meta_objset);
 		if (dump_opt['d'] >= 3) {
 			dump_bpobj(&spa->spa_deferred_bpobj,
@@ -2998,8 +3011,21 @@ dump_zpool(spa_t *spa)
 		}
 		(void) dmu_objset_find(spa_name(spa), dump_one_dir,
 		    NULL, DS_FIND_SNAPSHOTS | DS_FIND_CHILDREN);
+
+		(void) feature_get_refcount(spa,
+		    &spa_feature_table[SPA_FEATURE_LARGE_BLOCKS], &refcount);
+		if (num_large_blocks != refcount) {
+			(void) printf("large_blocks feature refcount mismatch: "
+			    "expected %lld != actual %lld\n",
+			    (longlong_t)num_large_blocks,
+			    (longlong_t)refcount);
+			rc = 2;
+		} else {
+			(void) printf("Verified large_blocks feature refcount "
+			    "is correct (%llu)\n", (longlong_t)refcount);
 		}
-	if (dump_opt['b'] || dump_opt['c'])
+	}
+	if (rc == 0 && (dump_opt['b'] || dump_opt['c']))
 		rc = dump_block_stats(spa);
 
 	if (rc == 0)

cmd/zfs/zfs_main.c

@@ -258,9 +258,9 @@ get_usage(zfs_help_t idx)
 	case HELP_ROLLBACK:
 		return (gettext("\trollback [-rRf] <snapshot>\n"));
 	case HELP_SEND:
-		return (gettext("\tsend [-DnPpRrve] [-[iI] snapshot] "
+		return (gettext("\tsend [-DnPpRvLe] [-[iI] snapshot] "
 		    "<snapshot>\n"
-		    "\tsend [-e] [-i snapshot|bookmark] "
+		    "\tsend [-Le] [-i snapshot|bookmark] "
 		    "<filesystem|volume|snapshot>\n"));
 	case HELP_SET:
 		return (gettext("\tset <property=value> "
@@ -3683,7 +3683,7 @@ zfs_do_send(int argc, char **argv)
 	boolean_t extraverbose = B_FALSE;
 
 	/* check options */
-	while ((c = getopt(argc, argv, ":i:I:RDpvnPe")) != -1) {
+	while ((c = getopt(argc, argv, ":i:I:RDpvnPLe")) != -1) {
 		switch (c) {
 		case 'i':
 			if (fromname)
@@ -3718,6 +3718,9 @@ zfs_do_send(int argc, char **argv)
 		case 'n':
 			flags.dryrun = B_TRUE;
 			break;
+		case 'L':
+			flags.largeblock = B_TRUE;
+			break;
 		case 'e':
 			flags.embed_data = B_TRUE;
 			break;
@@ -3774,6 +3777,8 @@ zfs_do_send(int argc, char **argv)
 		if (zhp == NULL)
 			return (1);
 
+		if (flags.largeblock)
+			lzc_flags |= LZC_SEND_FLAG_LARGE_BLOCK;
 		if (flags.embed_data)
 			lzc_flags |= LZC_SEND_FLAG_EMBED_DATA;
 

cmd/zstreamdump/zstreamdump.c

@@ -56,7 +56,6 @@ uint64_t total_stream_len = 0;
 FILE *send_stream = 0;
 boolean_t do_byteswap = B_FALSE;
 boolean_t do_cksum = B_TRUE;
-#define	INITIAL_BUFLEN (1<<20)
 
 static void
 usage(void)
@@ -69,6 +68,18 @@ usage(void)
 	exit(1);
 }
 
+static void *
+safe_malloc(size_t size)
+{
+	void *rv = malloc(size);
+	if (rv == NULL) {
+		(void) fprintf(stderr, "ERROR; failed to allocate %u bytes\n",
+		    (unsigned)size);
+		abort();
+	}
+	return (rv);
+}
+
 /*
  * ssread - send stream read.
  *
@@ -160,7 +171,7 @@ print_block(char *buf, int length)
 int
 main(int argc, char *argv[])
 {
-	char *buf = malloc(INITIAL_BUFLEN);
+	char *buf = safe_malloc(SPA_MAXBLOCKSIZE);
 	uint64_t drr_record_count[DRR_NUMTYPES] = { 0 };
 	uint64_t total_records = 0;
 	dmu_replay_record_t thedrr;
@@ -308,9 +319,9 @@ main(int argc, char *argv[])
 				nvlist_t *nv;
 				int sz = drr->drr_payloadlen;
 
-				if (sz > INITIAL_BUFLEN) {
+				if (sz > SPA_MAXBLOCKSIZE) {
 					free(buf);
-					buf = malloc(sz);
+					buf = safe_malloc(sz);
 				}
 				(void) ssread(buf, sz, &zc);
 				if (ferror(send_stream))

cmd/ztest/ztest.c

@@ -1040,9 +1040,14 @@ ztest_spa_get_ashift(void) {
 static int
 ztest_random_blocksize(void)
 {
-	// Choose a block size >= the ashift.
+	/*
-	uint64_t block_shift =
+	 * Choose a block size >= the ashift.
-	    ztest_random(SPA_MAXBLOCKSHIFT - ztest_spa_get_ashift() + 1);
+	 * If the SPA supports new MAXBLOCKSIZE, test up to 1MB blocks.
+	 */
+	int maxbs = SPA_OLD_MAXBLOCKSHIFT;
+	if (spa_maxblocksize(ztest_spa) == SPA_MAXBLOCKSIZE)
+		maxbs = 20;
+	uint64_t block_shift = ztest_random(maxbs - ztest_spa_get_ashift() + 1);
 	return (1 << (SPA_MINBLOCKSHIFT + block_shift));
 }
 
@@ -4972,7 +4977,7 @@ ztest_fault_inject(ztest_ds_t *zd, uint64_t id)
 	char *path0;
 	char *pathrand;
 	size_t fsize;
-	int bshift = SPA_MAXBLOCKSHIFT + 2;	/* don't scrog all labels */
+	int bshift = SPA_OLD_MAXBLOCKSHIFT + 2;	/* don't scrog all labels */
 	int iters = 1000;
 	int maxfaults;
 	int mirror_save;

include/libzfs.h

@@ -617,6 +617,9 @@ typedef struct sendflags {
 	/* show progress (ie. -v) */
 	boolean_t progress;
 
+	/* large blocks (>128K) are permitted */
+	boolean_t largeblock;
+
 	/* WRITE_EMBEDDED records of type DATA are permitted */
 	boolean_t embed_data;
 } sendflags_t;

include/libzfs_core.h

@@ -53,7 +53,8 @@ int lzc_release(nvlist_t *, nvlist_t **);
 int lzc_get_holds(const char *, nvlist_t **);
 
 enum lzc_send_flags {
-	LZC_SEND_FLAG_EMBED_DATA = 1 << 0
+	LZC_SEND_FLAG_EMBED_DATA = 1 << 0,
+	LZC_SEND_FLAG_LARGE_BLOCK = 1 << 1
 };
 
 int lzc_send(const char *, const char *, int, enum lzc_send_flags);

include/sys/dmu.h

@@ -245,7 +245,7 @@ void zfs_znode_byteswap(void *buf, size_t size);
  * The maximum number of bytes that can be accessed as part of one
  * operation, including metadata.
  */
-#define	DMU_MAX_ACCESS (10<<20) /* 10MB */
+#define	DMU_MAX_ACCESS (64 * 1024 * 1024) /* 64MB */
 #define	DMU_MAX_DELETEBLKCNT (20480) /* ~5MB of indirect blocks */
 
 #define	DMU_USERUSED_OBJECT	(-1ULL)
@@ -732,6 +732,7 @@ void xuio_stat_wbuf_copied(void);
 void xuio_stat_wbuf_nocopy(void);
 
 extern int zfs_prefetch_disable;
+extern int zfs_max_recordsize;
 
 /*
  * Asynchronously try to read in the data.

include/sys/dmu_objset.h

@@ -99,6 +99,7 @@ struct objset {
 	zfs_cache_type_t os_secondary_cache;
 	zfs_sync_type_t os_sync;
 	zfs_redundant_metadata_type_t os_redundant_metadata;
+	int os_recordsize;
 
 	/* no lock needed: */
 	struct dmu_tx *os_synctx; /* XXX sketchy */

include/sys/dmu_send.h

@@ -37,12 +37,14 @@ struct dsl_dataset;
 struct drr_begin;
 struct avl_tree;
 
-int dmu_send(const char *tosnap, const char *fromsnap, boolean_t embedok,
+int dmu_send(const char *tosnap, const char *fromsnap,
+    boolean_t embedok, boolean_t large_block_ok,
     int outfd, struct vnode *vp, offset_t *off);
 int dmu_send_estimate(struct dsl_dataset *ds, struct dsl_dataset *fromds,
     uint64_t *sizep);
 int dmu_send_obj(const char *pool, uint64_t tosnap, uint64_t fromsnap,
-    boolean_t embedok, int outfd, vnode_t *vp, offset_t *off);
+    boolean_t embedok, boolean_t large_block_ok,
+    int outfd, struct vnode *vp, offset_t *off);
 
 typedef struct dmu_recv_cookie {
 	struct dsl_dataset *drc_ds;

include/sys/dsl_dataset.h

@@ -83,6 +83,13 @@ struct dsl_pool;
  */
 #define	DS_FIELD_BOOKMARK_NAMES "com.delphix:bookmarks"
 
+/*
+ * This field is present (with value=0) if this dataset may contain large
+ * blocks (>128KB).  If it is present, then this dataset
+ * is counted in the refcount of the SPA_FEATURE_LARGE_BLOCKS feature.
+ */
+#define	DS_FIELD_LARGE_BLOCKS "org.open-zfs:large_blocks"
+
 /*
  * DS_FLAG_CI_DATASET is set if the dataset contains a file system whose
  * name lookups should be performed case-insensitively.
@@ -138,6 +145,8 @@ typedef struct dsl_dataset {
 	/* only used in syncing context, only valid for non-snapshots: */
 	struct dsl_dataset *ds_prev;
 	uint64_t ds_bookmarks;  /* DMU_OTN_ZAP_METADATA */
+	boolean_t ds_large_blocks;
+	boolean_t ds_need_large_blocks;
 
 	/* has internal locking: */
 	dsl_deadlist_t ds_deadlist;
@@ -252,6 +261,8 @@ int dsl_dataset_space_written(dsl_dataset_t *oldsnap, dsl_dataset_t *new,
 int dsl_dataset_space_wouldfree(dsl_dataset_t *firstsnap, dsl_dataset_t *last,
     uint64_t *usedp, uint64_t *compp, uint64_t *uncompp);
 boolean_t dsl_dataset_is_dirty(dsl_dataset_t *ds);
+int dsl_dataset_activate_large_blocks(const char *dsname);
+void dsl_dataset_activate_large_blocks_sync_impl(uint64_t dsobj, dmu_tx_t *tx);
 
 int dsl_dsobj_to_dsname(char *pname, uint64_t obj, char *buf);
 

include/sys/fs/zfs.h

@@ -200,6 +200,7 @@ typedef enum {
 	ZPOOL_PROP_FREEING,
 	ZPOOL_PROP_FRAGMENTATION,
 	ZPOOL_PROP_LEAKED,
+	ZPOOL_PROP_MAXBLOCKSIZE,
 	ZPOOL_PROP_TNAME,
 	ZPOOL_NUM_PROPS
 } zpool_prop_t;

include/sys/spa.h

@@ -98,17 +98,26 @@ _NOTE(CONSTCOND) } while (0)
 _NOTE(CONSTCOND) } while (0)
 
 /*
- * We currently support nine block sizes, from 512 bytes to 128K.
+ * We currently support block sizes from 512 bytes to 16MB.
- * We could go higher, but the benefits are near-zero and the cost
+ * The benefits of larger blocks, and thus larger IO, need to be weighed
- * of COWing a giant block to modify one byte would become excessive.
+ * against the cost of COWing a giant block to modify one byte, and the
+ * large latency of reading or writing a large block.
+ *
+ * Note that although blocks up to 16MB are supported, the recordsize
+ * property can not be set larger than zfs_max_recordsize (default 1MB).
+ * See the comment near zfs_max_recordsize in dsl_dataset.c for details.
+ *
+ * Note that although the LSIZE field of the blkptr_t can store sizes up
+ * to 32MB, the dnode's dn_datablkszsec can only store sizes up to
+ * 32MB - 512 bytes.  Therefore, we limit SPA_MAXBLOCKSIZE to 16MB.
  */
 #define	SPA_MINBLOCKSHIFT	9
-#define	SPA_MAXBLOCKSHIFT	17
+#define	SPA_OLD_MAXBLOCKSHIFT	17
+#define	SPA_MAXBLOCKSHIFT	24
 #define	SPA_MINBLOCKSIZE	(1ULL << SPA_MINBLOCKSHIFT)
+#define	SPA_OLD_MAXBLOCKSIZE	(1ULL << SPA_OLD_MAXBLOCKSHIFT)
 #define	SPA_MAXBLOCKSIZE	(1ULL << SPA_MAXBLOCKSHIFT)
 
-#define	SPA_BLOCKSIZES		(SPA_MAXBLOCKSHIFT - SPA_MINBLOCKSHIFT + 1)
-
 /*
  * Size of block to hold the configuration data (a packed nvlist)
  */
@@ -830,6 +839,7 @@ extern boolean_t spa_has_slogs(spa_t *spa);
 extern boolean_t spa_is_root(spa_t *spa);
 extern boolean_t spa_writeable(spa_t *spa);
 extern boolean_t spa_has_pending_synctask(spa_t *spa);
+extern int spa_maxblocksize(spa_t *spa);
 extern void zfs_blkptr_verify(spa_t *spa, const blkptr_t *bp);
 
 extern int spa_mode(spa_t *spa);

include/sys/zap_impl.h

@@ -42,8 +42,7 @@ extern int fzap_default_block_shift;
 
 #define	MZAP_ENT_LEN		64
 #define	MZAP_NAME_LEN		(MZAP_ENT_LEN - 8 - 4 - 2)
-#define	MZAP_MAX_BLKSHIFT	SPA_MAXBLOCKSHIFT
+#define	MZAP_MAX_BLKSZ		SPA_OLD_MAXBLOCKSIZE
-#define	MZAP_MAX_BLKSZ		(1 << MZAP_MAX_BLKSHIFT)
 
 #define	ZAP_NEED_CD		(-1U)
 

include/sys/zfs_ioctl.h

@@ -96,13 +96,16 @@ typedef enum drr_headertype {
 /* flags #3 - #15 are reserved for incompatible closed-source implementations */
 #define	DMU_BACKUP_FEATURE_EMBED_DATA		(1<<16)
 #define	DMU_BACKUP_FEATURE_EMBED_DATA_LZ4	(1<<17)
+/* flag #18 is reserved for a Delphix feature */
+#define	DMU_BACKUP_FEATURE_LARGE_BLOCKS		(1<<19)
 
 /*
  * Mask of all supported backup features
  */
 #define	DMU_BACKUP_FEATURE_MASK	(DMU_BACKUP_FEATURE_DEDUP | \
     DMU_BACKUP_FEATURE_DEDUPPROPS | DMU_BACKUP_FEATURE_SA_SPILL | \
-    DMU_BACKUP_FEATURE_EMBED_DATA | DMU_BACKUP_FEATURE_EMBED_DATA_LZ4)
+    DMU_BACKUP_FEATURE_EMBED_DATA | DMU_BACKUP_FEATURE_EMBED_DATA_LZ4 | \
+    DMU_BACKUP_FEATURE_LARGE_BLOCKS)
 
 /* Are all features in the given flag word currently supported? */
 #define	DMU_STREAM_SUPPORTED(x)	(!((x) & ~DMU_BACKUP_FEATURE_MASK))

include/sys/zfs_sa.h

@@ -129,7 +129,7 @@ typedef struct znode_phys {
 #ifdef _KERNEL
 
 #define	DXATTR_MAX_ENTRY_SIZE	(32768)
-#define	DXATTR_MAX_SA_SIZE	(SPA_MAXBLOCKSIZE >> 1)
+#define	DXATTR_MAX_SA_SIZE	(SPA_OLD_MAXBLOCKSIZE >> 1)
 
 int zfs_sa_readlink(struct znode *, uio_t *);
 void zfs_sa_symlink(struct znode *, char *link, int len, dmu_tx_t *);

include/sys/zfs_znode.h

@@ -137,8 +137,6 @@ extern "C" {
 #define	ZFS_SHARES_DIR		"SHARES"
 #define	ZFS_SA_ATTRS		"SA_ATTRS"
 
-#define	ZFS_MAX_BLOCKSIZE	(SPA_MAXBLOCKSIZE)
-
 /*
  * Path component length
  *

include/sys/zil.h

@@ -90,7 +90,6 @@ typedef struct zil_chain {
 } zil_chain_t;
 
 #define	ZIL_MIN_BLKSZ	4096ULL
-#define	ZIL_MAX_BLKSZ	SPA_MAXBLOCKSIZE
 
 /*
  * The words of a log block checksum.

include/sys/zil_impl.h

@@ -140,7 +140,7 @@ typedef struct zil_bp_node {
 	avl_node_t	zn_node;
 } zil_bp_node_t;
 
-#define	ZIL_MAX_LOG_DATA (SPA_MAXBLOCKSIZE - sizeof (zil_chain_t) - \
+#define	ZIL_MAX_LOG_DATA (SPA_OLD_MAXBLOCKSIZE - sizeof (zil_chain_t) - \
     sizeof (lr_write_t))
 
 #ifdef	__cplusplus

include/zfeature_common.h

@@ -49,6 +49,7 @@ typedef enum spa_feature {
 	SPA_FEATURE_EMBEDDED_DATA,
 	SPA_FEATURE_BOOKMARKS,
 	SPA_FEATURE_FS_SS_LIMIT,
+	SPA_FEATURE_LARGE_BLOCKS,
 	SPA_FEATURES
 } spa_feature_t;
 

lib/libzfs/libzfs_dataset.c

@@ -1055,21 +1055,28 @@ zfs_valid_proplist(libzfs_handle_t *hdl, zfs_type_t type, nvlist_t *nvl,
 			break;
 		}
 
-		case ZFS_PROP_RECORDSIZE:
 		case ZFS_PROP_VOLBLOCKSIZE:
-			/* must be power of two within SPA_{MIN,MAX}BLOCKSIZE */
+		case ZFS_PROP_RECORDSIZE:
+		{
+			int maxbs = SPA_MAXBLOCKSIZE;
+			if (zhp != NULL) {
+				maxbs = zpool_get_prop_int(zhp->zpool_hdl,
+				    ZPOOL_PROP_MAXBLOCKSIZE, NULL);
+			}
+			/*
+			 * The value must be a power of two between
+			 * SPA_MINBLOCKSIZE and maxbs.
+			 */
 			if (intval < SPA_MINBLOCKSIZE ||
-			    intval > SPA_MAXBLOCKSIZE || !ISP2(intval)) {
+			    intval > maxbs || !ISP2(intval)) {
 				zfs_error_aux(hdl, dgettext(TEXT_DOMAIN,
-				    "'%s' must be power of 2 from %u "
+				    "'%s' must be power of 2 from 512B "
-				    "to %uk"), propname,
+				    "to %uKB"), propname, maxbs >> 10);
-				    (uint_t)SPA_MINBLOCKSIZE,
-				    (uint_t)SPA_MAXBLOCKSIZE >> 10);
 				(void) zfs_error(hdl, EZFS_BADPROP, errbuf);
 				goto error;
 			}
 			break;
-
+		}
 		case ZFS_PROP_MLSLABEL:
 		{
 #ifdef HAVE_MLSLABEL
@@ -1446,7 +1453,8 @@ zfs_setprop_error(libzfs_handle_t *hdl, zfs_prop_t prop, int err,
 		break;
 
 	case ERANGE:
-		if (prop == ZFS_PROP_COMPRESSION) {
+		if (prop == ZFS_PROP_COMPRESSION ||
+		    prop == ZFS_PROP_RECORDSIZE) {
 			(void) zfs_error_aux(hdl, dgettext(TEXT_DOMAIN,
 			    "property setting is not allowed on "
 			    "bootable datasets"));
@@ -3212,9 +3220,7 @@ zfs_create(libzfs_handle_t *hdl, const char *path, zfs_type_t type,
 		case EDOM:
 			zfs_error_aux(hdl, dgettext(TEXT_DOMAIN,
 			    "volume block size must be power of 2 from "
-			    "%u to %uk"),
+			    "512B to %uKB"), zfs_max_recordsize >> 10);
-			    (uint_t)SPA_MINBLOCKSIZE,
-			    (uint_t)SPA_MAXBLOCKSIZE >> 10);
 
 			return (zfs_error(hdl, EZFS_BADPROP, errbuf));
 

lib/libzfs/libzfs_sendrecv.c

@@ -214,7 +214,7 @@ static void *
 cksummer(void *arg)
 {
 	dedup_arg_t *dda = arg;
-	char *buf = malloc(1<<20);
+	char *buf = zfs_alloc(dda->dedup_hdl, SPA_MAXBLOCKSIZE);
 	dmu_replay_record_t thedrr;
 	dmu_replay_record_t *drr = &thedrr;
 	struct drr_begin *drrb = &thedrr.drr_u.drr_begin;
@@ -279,9 +279,9 @@ cksummer(void *arg)
 			    DMU_COMPOUNDSTREAM && drr->drr_payloadlen != 0) {
 				int sz = drr->drr_payloadlen;
 
-				if (sz > 1<<20) {
+				if (sz > SPA_MAXBLOCKSIZE) {
-					free(buf);
+					buf = zfs_realloc(dda->dedup_hdl, buf,
-					buf = malloc(sz);
+					    SPA_MAXBLOCKSIZE, sz);
 				}
 				(void) ssread(buf, sz, ofp);
 				if (ferror(stdin))
@@ -834,7 +834,7 @@ typedef struct send_dump_data {
 	char prevsnap[ZFS_MAXNAMELEN];
 	uint64_t prevsnap_obj;
 	boolean_t seenfrom, seento, replicate, doall, fromorigin;
-	boolean_t verbose, dryrun, parsable, progress, embed_data;
+	boolean_t verbose, dryrun, parsable, progress, embed_data, large_block;
 	int outfd;
 	boolean_t err;
 	nvlist_t *fss;
@@ -1181,6 +1181,8 @@ dump_snapshot(zfs_handle_t *zhp, void *arg)
 		}
 
 		enum lzc_send_flags flags = 0;
+		if (sdd->large_block)
+			flags |= LZC_SEND_FLAG_LARGE_BLOCK;
 		if (sdd->embed_data)
 			flags |= LZC_SEND_FLAG_EMBED_DATA;
 
@@ -1529,6 +1531,7 @@ zfs_send(zfs_handle_t *zhp, const char *fromsnap, const char *tosnap,
 	sdd.parsable = flags->parsable;
 	sdd.progress = flags->progress;
 	sdd.dryrun = flags->dryrun;
+	sdd.large_block = flags->largeblock;
 	sdd.embed_data = flags->embed_data;
 	sdd.filter_cb = filter_func;
 	sdd.filter_cb_arg = cb_arg;
@@ -2564,7 +2567,7 @@ static int
 recv_skip(libzfs_handle_t *hdl, int fd, boolean_t byteswap)
 {
 	dmu_replay_record_t *drr;
-	void *buf = malloc(1<<20);
+	void *buf = zfs_alloc(hdl, SPA_MAXBLOCKSIZE);
 	char errbuf[1024];
 
 	(void) snprintf(errbuf, sizeof (errbuf), dgettext(TEXT_DOMAIN,

lib/libzfs_core/libzfs_core.c

@@ -455,6 +455,10 @@ lzc_get_holds(const char *snapname, nvlist_t **holdsp)
  *
  * "fd" is the file descriptor to write the send stream to.
  *
+ * If "flags" contains LZC_SEND_FLAG_LARGE_BLOCK, the stream is permitted
+ * to contain DRR_WRITE records with drr_length > 128K, and DRR_OBJECT
+ * records with drr_blksz > 128K.
+ *
  * If "flags" contains LZC_SEND_FLAG_EMBED_DATA, the stream is permitted
  * to contain DRR_WRITE_EMBEDDED records with drr_etype==BP_EMBEDDED_TYPE_DATA,
  * which the receiving system must support (as indicated by support
@@ -471,6 +475,8 @@ lzc_send(const char *snapname, const char *from, int fd,
 	fnvlist_add_int32(args, "fd", fd);
 	if (from != NULL)
 		fnvlist_add_string(args, "fromsnap", from);
+	if (flags & LZC_SEND_FLAG_LARGE_BLOCK)
+		fnvlist_add_boolean(args, "largeblockok");
 	if (flags & LZC_SEND_FLAG_EMBED_DATA)
 		fnvlist_add_boolean(args, "embedok");
 	err = lzc_ioctl(ZFS_IOC_SEND_NEW, snapname, args, NULL);

man/man5/zfs-module-parameters.5

@@ -945,6 +945,24 @@ Largest data block to write to zil
 Default value: \fB32,768\fR.
 .RE
 
+.sp
+.ne 2
+.na
+\fBzfs_max_recordsize\fR (int)
+.ad
+.RS 12n
+We currently support block sizes from 512 bytes to 16MB.  The benefits of
+larger blocks, and thus larger IO, need to be weighed against the cost of
+COWing a giant block to modify one byte.  Additionally, very large blocks
+can have an impact on i/o latency, and also potentially on the memory
+allocator.  Therefore, we do not allow the recordsize to be set larger than
+zfs_max_recordsize (default 1MB).  Larger blocks can be created by changing
+this tunable, and pools with larger blocks can always be imported and used,
+regardless of this setting.
+.sp
+Default value: \fB1,048,576\fR.
+.RE
+
 .sp
 .ne 2
 .na

man/man5/zpool-features.5

@@ -411,5 +411,26 @@ never return to being \fBenabled\fR.
 
 .RE
 
+.sp
+.ne 2
+.na
+\fB\fBlarge_blocks\fR\fR
+.ad
+.RS 4n
+.TS
+l l .
+GUID	org.open-zfs:large_block
+READ\-ONLY COMPATIBLE	no
+DEPENDENCIES	extensible_dataset
+.TE
+
+The \fBlarge_block\fR feature allows the record size on a dataset to be
+set larger than 128KB.
+
+This feature becomes \fBactive\fR once a \fBrecordsize\fR property has been
+set larger than 128KB, and will return to being \fBenabled\fR once all
+filesystems that have ever had their recordsize larger than 128KB are destroyed.
+.RE
+
 .SH "SEE ALSO"
 \fBzpool\fR(8)

man/man8/zfs.8

@@ -174,12 +174,12 @@ zfs \- configures ZFS file systems
 
 .LP
 .nf
-\fBzfs\fR \fBsend\fR [\fB-DnPpRve\fR] [\fB-\fR[\fBiI\fR] \fIsnapshot\fR] \fIsnapshot\fR
+\fBzfs\fR \fBsend\fR [\fB-DnPpRveL\fR] [\fB-\fR[\fBiI\fR] \fIsnapshot\fR] \fIsnapshot\fR
 .fi
 
 .LP
 .nf
-\fBzfs\fR \fBsend\fR [\fB-e\fR] [\fB-i \fIsnapshot\fR|\fIbookmark\fR]\fR \fIfilesystem\fR|\fIvolume\fR|\fIsnapshot\fR
+\fBzfs\fR \fBsend\fR [\fB-eL\fR] [\fB-i \fIsnapshot\fR|\fIbookmark\fR]\fR \fIfilesystem\fR|\fIvolume\fR|\fIsnapshot\fR
 .fi
 
 .LP
@@ -2706,7 +2706,7 @@ See \fBzpool-features\fR(5) for details on ZFS feature flags and the
 .sp
 .ne 2
 .na
-\fBzfs send\fR [\fB-DnPpRve\fR] [\fB-\fR[\fBiI\fR] \fIsnapshot\fR] \fIsnapshot\fR
+\fBzfs send\fR [\fB-DnPpRveL\fR] [\fB-\fR[\fBiI\fR] \fIsnapshot\fR] \fIsnapshot\fR
 .ad
 .sp .6
 .RS 4n
@@ -2759,6 +2759,22 @@ If the \fB-i\fR or \fB-I\fR flags are used in conjunction with the \fB-R\fR flag
 Generate a deduplicated stream. Blocks which would have been sent multiple times in the send stream will only be sent once. The receiving system must also support this feature to receive a deduplicated stream.  This flag can be used regardless of the dataset's dedup  property, but performance will be much better if the filesystem uses a dedup-capable checksum (eg.  sha256).
 .RE
 
+.sp
+.ne 2
+.mk
+.na
+\fB\fB-L\fR\fR
+.ad
+.sp .6
+.RS 4n
+Generate a stream which may contain blocks larger than 128KB.  This flag
+has no effect if the \fBlarge_blocks\fR pool feature is disabled, or if
+the \fRrecordsize\fR property of this filesystem has never been set above
+128KB.  The receiving system must have the \fBlarge_blocks\fR pool feature
+enabled as well.  See \fBzpool-features\fR(5) for details on ZFS feature
+flags and the \fBlarge_blocks\fR feature.
+.RE
+
 .sp
 .ne 2
 .mk
@@ -2828,7 +2844,7 @@ The format of the stream is committed. You will be able to receive your streams
 .sp
 .ne 2
 .na
-\fBzfs send\fR [\fB-e\fR] [\fB-i\fR \fIsnapshot\fR|\fIbookmark\fR] \fIfilesystem\fR|\fIvolume\fR|\fIsnapshot\fR
+\fBzfs send\fR [\fB-eL\fR] [\fB-i\fR \fIsnapshot\fR|\fIbookmark\fR] \fIfilesystem\fR|\fIvolume\fR|\fIsnapshot\fR
 .ad
 .sp .6
 .RS 4n
@@ -2856,6 +2872,22 @@ be the origin snapshot, or an earlier snapshot in the origin's filesystem,
 or the origin's origin, etc.
 .RE
 
+.sp
+.ne 2
+.mk
+.na
+\fB\fB-L\fR\fR
+.ad
+.sp .6
+.RS 4n
+Generate a stream which may contain blocks larger than 128KB.  This flag
+has no effect if the \fBlarge_blocks\fR pool feature is disabled, or if
+the \fRrecordsize\fR property of this filesystem has never been set above
+128KB.  The receiving system must have the \fBlarge_blocks\fR pool feature
+enabled as well.  See \fBzpool-features\fR(5) for details on ZFS feature
+flags and the \fBlarge_blocks\fR feature.
+.RE
+
 .sp
 .ne 2
 .mk
@@ -2909,7 +2941,6 @@ The \fB-d\fR and \fB-e\fR options cause the file system name of the target snaps
 Discard the first element of the sent snapshot's file system name, using the remaining elements to determine the name of the target file system for the new snapshot as described in the paragraph above.
 .RE
 
-
 .sp
 .ne 2
 .na

module/zcommon/zfs_prop.c

@@ -426,8 +426,8 @@ zfs_prop_init(void)
 
 	/* inherit number properties */
 	zprop_register_number(ZFS_PROP_RECORDSIZE, "recordsize",
-	    SPA_MAXBLOCKSIZE, PROP_INHERIT,
+	    SPA_OLD_MAXBLOCKSIZE, PROP_INHERIT,
-	    ZFS_TYPE_FILESYSTEM, "512 to 128k, power of 2", "RECSIZE");
+	    ZFS_TYPE_FILESYSTEM, "512 to 1M, power of 2", "RECSIZE");
 
 	/* hidden properties */
 	zprop_register_hidden(ZFS_PROP_CREATETXG, "createtxg", PROP_TYPE_NUMBER,

module/zcommon/zpool_prop.c

@@ -131,6 +131,8 @@ zpool_prop_init(void)
 	/* hidden properties */
 	zprop_register_hidden(ZPOOL_PROP_NAME, "name", PROP_TYPE_STRING,
 	    PROP_READONLY, ZFS_TYPE_POOL, "NAME");
+	zprop_register_hidden(ZPOOL_PROP_MAXBLOCKSIZE, "maxblocksize",
+	    PROP_TYPE_NUMBER, PROP_READONLY, ZFS_TYPE_POOL, "MAXBLOCKSIZE");
 	zprop_register_hidden(ZPOOL_PROP_TNAME, "tname", PROP_TYPE_STRING,
 	    PROP_ONETIME, ZFS_TYPE_POOL, "TNAME");
 }

module/zfs/arc.c

@@ -1329,7 +1329,7 @@ arc_buf_alloc(spa_t *spa, uint64_t size, void *tag, arc_buf_contents_t type)
 	arc_buf_hdr_t *hdr;
 	arc_buf_t *buf;
 
-	VERIFY3U(size, <=, SPA_MAXBLOCKSIZE);
+	VERIFY3U(size, <=, spa_maxblocksize(spa));
 	hdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE);
 	ASSERT(BUF_EMPTY(hdr));
 	hdr->b_size = size;
@@ -3289,7 +3289,7 @@ top:
 		 * Gracefully handle a damaged logical block size as a
 		 * checksum error by passing a dummy zio to the done callback.
 		 */
-		if (size > SPA_MAXBLOCKSIZE) {
+		if (size > spa_maxblocksize(spa)) {
 			if (done) {
 				rzio = zio_null(pio, spa, NULL,
 				    NULL, NULL, zio_flags);

module/zfs/bpobj.c

@@ -43,7 +43,7 @@ bpobj_alloc_empty(objset_t *os, int blocksize, dmu_tx_t *tx)
 		if (!spa_feature_is_active(spa, SPA_FEATURE_EMPTY_BPOBJ)) {
 			ASSERT0(dp->dp_empty_bpobj);
 			dp->dp_empty_bpobj =
-			    bpobj_alloc(os, SPA_MAXBLOCKSIZE, tx);
+			    bpobj_alloc(os, SPA_OLD_MAXBLOCKSIZE, tx);
 			VERIFY(zap_add(os,
 			    DMU_POOL_DIRECTORY_OBJECT,
 			    DMU_POOL_EMPTY_BPOBJ, sizeof (uint64_t), 1,
@@ -399,7 +399,8 @@ bpobj_enqueue_subobj(bpobj_t *bpo, uint64_t subobj, dmu_tx_t *tx)
 	dmu_buf_will_dirty(bpo->bpo_dbuf, tx);
 	if (bpo->bpo_phys->bpo_subobjs == 0) {
 		bpo->bpo_phys->bpo_subobjs = dmu_object_alloc(bpo->bpo_os,
-		    DMU_OT_BPOBJ_SUBOBJ, SPA_MAXBLOCKSIZE, DMU_OT_NONE, 0, tx);
+		    DMU_OT_BPOBJ_SUBOBJ, SPA_OLD_MAXBLOCKSIZE,
+		    DMU_OT_NONE, 0, tx);
 	}
 
 	ASSERT0(dmu_object_info(bpo->bpo_os, bpo->bpo_phys->bpo_subobjs, &doi));

module/zfs/bptree.c

@@ -65,7 +65,7 @@ bptree_alloc(objset_t *os, dmu_tx_t *tx)
 	bptree_phys_t *bt;
 
 	obj = dmu_object_alloc(os, DMU_OTN_UINT64_METADATA,
-	    SPA_MAXBLOCKSIZE, DMU_OTN_UINT64_METADATA,
+	    SPA_OLD_MAXBLOCKSIZE, DMU_OTN_UINT64_METADATA,
 	    sizeof (bptree_phys_t), tx);
 
 	/*

module/zfs/dbuf.c

@@ -2216,9 +2216,7 @@ dbuf_spill_set_blksz(dmu_buf_t *db_fake, uint64_t blksz, dmu_tx_t *tx)
 		return (SET_ERROR(ENOTSUP));
 	if (blksz == 0)
 		blksz = SPA_MINBLOCKSIZE;
-	if (blksz > SPA_MAXBLOCKSIZE)
+	ASSERT3U(blksz, <=, spa_maxblocksize(dmu_objset_spa(db->db_objset)));
-		blksz = SPA_MAXBLOCKSIZE;
-	else
 	blksz = P2ROUNDUP(blksz, SPA_MINBLOCKSIZE);
 
 	DB_DNODE_ENTER(db);

module/zfs/dmu_objset.c

@@ -256,6 +256,14 @@ logbias_changed_cb(void *arg, uint64_t newval)
 		zil_set_logbias(os->os_zil, newval);
 }
 
+static void
+recordsize_changed_cb(void *arg, uint64_t newval)
+{
+	objset_t *os = arg;
+
+	os->os_recordsize = newval;
+}
+
 void
 dmu_objset_byteswap(void *buf, size_t size)
 {
@@ -385,6 +393,11 @@ dmu_objset_open_impl(spa_t *spa, dsl_dataset_t *ds, blkptr_t *bp,
 				    ZFS_PROP_REDUNDANT_METADATA),
 				    redundant_metadata_changed_cb, os);
 			}
+			if (err == 0) {
+				err = dsl_prop_register(ds,
+				    zfs_prop_to_name(ZFS_PROP_RECORDSIZE),
+				    recordsize_changed_cb, os);
+			}
 		}
 		if (err != 0) {
 			VERIFY(arc_buf_remove_ref(os->os_phys_buf,
@@ -660,6 +673,9 @@ dmu_objset_evict(objset_t *os)
 			VERIFY0(dsl_prop_unregister(ds,
 			    zfs_prop_to_name(ZFS_PROP_REDUNDANT_METADATA),
 			    redundant_metadata_changed_cb, os));
+			VERIFY0(dsl_prop_unregister(ds,
+			    zfs_prop_to_name(ZFS_PROP_RECORDSIZE),
+			    recordsize_changed_cb, os));
 		}
 		VERIFY0(dsl_prop_unregister(ds,
 		    zfs_prop_to_name(ZFS_PROP_PRIMARYCACHE),

module/zfs/dmu_send.c

@@ -234,11 +234,12 @@ dump_write(dmu_sendarg_t *dsp, dmu_object_type_t type,
 	drrw->drr_offset = offset;
 	drrw->drr_length = blksz;
 	drrw->drr_toguid = dsp->dsa_toguid;
-	if (BP_IS_EMBEDDED(bp)) {
+	if (bp == NULL || BP_IS_EMBEDDED(bp)) {
 		/*
-		 * There's no pre-computed checksum of embedded BP's, so
+		 * There's no pre-computed checksum for partial-block
-		 * (like fletcher4-checkummed blocks) userland will have
+		 * writes or embedded BP's, so (like
-		 * to compute a dedup-capable checksum itself.
+		 * fletcher4-checkummed blocks) userland will have to
+		 * compute a dedup-capable checksum itself.
 		 */
 		drrw->drr_checksumtype = ZIO_CHECKSUM_OFF;
 	} else {
@@ -400,6 +401,10 @@ dump_dnode(dmu_sendarg_t *dsp, uint64_t object, dnode_phys_t *dnp)
 	drro->drr_compress = dnp->dn_compress;
 	drro->drr_toguid = dsp->dsa_toguid;
 
+	if (!(dsp->dsa_featureflags & DMU_BACKUP_FEATURE_LARGE_BLOCKS) &&
+	    drro->drr_blksz > SPA_OLD_MAXBLOCKSIZE)
+		drro->drr_blksz = SPA_OLD_MAXBLOCKSIZE;
+
 	if (dump_bytes(dsp, dsp->dsa_drr, sizeof (dmu_replay_record_t)) != 0)
 		return (SET_ERROR(EINTR));
 
@@ -517,6 +522,7 @@ backup_cb(spa_t *spa, zilog_t *zilog, const blkptr_t *bp,
 		    zb->zb_blkid * blksz, blksz, bp);
 	} else { /* it's a level-0 block of a regular object */
 		uint32_t aflags = ARC_WAIT;
+		uint64_t offset;
 		arc_buf_t *abuf;
 		int blksz = BP_GET_LSIZE(bp);
 
@@ -539,8 +545,24 @@ backup_cb(spa_t *spa, zilog_t *zilog, const blkptr_t *bp,
 			}
 		}
 
-		err = dump_write(dsp, type, zb->zb_object, zb->zb_blkid * blksz,
+		offset = zb->zb_blkid * blksz;
-		    blksz, bp, abuf->b_data);
+
+		if (!(dsp->dsa_featureflags &
+		    DMU_BACKUP_FEATURE_LARGE_BLOCKS) &&
+		    blksz > SPA_OLD_MAXBLOCKSIZE) {
+			char *buf = abuf->b_data;
+			while (blksz > 0 && err == 0) {
+				int n = MIN(blksz, SPA_OLD_MAXBLOCKSIZE);
+				err = dump_write(dsp, type, zb->zb_object,
+				    offset, n, NULL, buf);
+				offset += n;
+				buf += n;
+				blksz -= n;
+			}
+		} else {
+			err = dump_write(dsp, type, zb->zb_object,
+			    offset, blksz, bp, abuf->b_data);
+		}
 		(void) arc_buf_remove_ref(abuf, &abuf);
 	}
 
@@ -554,7 +576,7 @@ backup_cb(spa_t *spa, zilog_t *zilog, const blkptr_t *bp,
 static int
 dmu_send_impl(void *tag, dsl_pool_t *dp, dsl_dataset_t *ds,
     zfs_bookmark_phys_t *fromzb, boolean_t is_clone, boolean_t embedok,
-    int outfd, vnode_t *vp, offset_t *off)
+    boolean_t large_block_ok, int outfd, vnode_t *vp, offset_t *off)
 {
 	objset_t *os;
 	dmu_replay_record_t *drr;
@@ -589,6 +611,8 @@ dmu_send_impl(void *tag, dsl_pool_t *dp, dsl_dataset_t *ds,
 	}
 #endif
 
+	if (large_block_ok && ds->ds_large_blocks)
+		featureflags |= DMU_BACKUP_FEATURE_LARGE_BLOCKS;
 	if (embedok &&
 	    spa_feature_is_active(dp->dp_spa, SPA_FEATURE_EMBEDDED_DATA)) {
 		featureflags |= DMU_BACKUP_FEATURE_EMBED_DATA;
@@ -684,7 +708,8 @@ out:
 
 int
 dmu_send_obj(const char *pool, uint64_t tosnap, uint64_t fromsnap,
-    boolean_t embedok, int outfd, vnode_t *vp, offset_t *off)
+    boolean_t embedok, boolean_t large_block_ok,
+    int outfd, vnode_t *vp, offset_t *off)
 {
 	dsl_pool_t *dp;
 	dsl_dataset_t *ds;
@@ -719,18 +744,19 @@ dmu_send_obj(const char *pool, uint64_t tosnap, uint64_t fromsnap,
 		zb.zbm_guid = dsl_dataset_phys(fromds)->ds_guid;
 		is_clone = (fromds->ds_dir != ds->ds_dir);
 		dsl_dataset_rele(fromds, FTAG);
-		err = dmu_send_impl(FTAG, dp, ds, &zb, is_clone, embedok,
+		err = dmu_send_impl(FTAG, dp, ds, &zb, is_clone,
-		    outfd, vp, off);
+		    embedok, large_block_ok, outfd, vp, off);
 	} else {
-		err = dmu_send_impl(FTAG, dp, ds, NULL, B_FALSE, embedok,
+		err = dmu_send_impl(FTAG, dp, ds, NULL, B_FALSE,
-		    outfd, vp, off);
+		    embedok, large_block_ok, outfd, vp, off);
 	}
 	dsl_dataset_rele(ds, FTAG);
 	return (err);
 }
 
 int
-dmu_send(const char *tosnap, const char *fromsnap, boolean_t embedok,
+dmu_send(const char *tosnap, const char *fromsnap,
+    boolean_t embedok, boolean_t large_block_ok,
     int outfd, vnode_t *vp, offset_t *off)
 {
 	dsl_pool_t *dp;
@@ -797,11 +823,11 @@ dmu_send(const char *tosnap, const char *fromsnap, boolean_t embedok,
 			dsl_pool_rele(dp, FTAG);
 			return (err);
 		}
-		err = dmu_send_impl(FTAG, dp, ds, &zb, is_clone, embedok,
+		err = dmu_send_impl(FTAG, dp, ds, &zb, is_clone,
-		    outfd, vp, off);
+		    embedok, large_block_ok, outfd, vp, off);
 	} else {
-		err = dmu_send_impl(FTAG, dp, ds, NULL, B_FALSE, embedok,
+		err = dmu_send_impl(FTAG, dp, ds, NULL, B_FALSE,
-		    outfd, vp, off);
+		    embedok, large_block_ok, outfd, vp, off);
 	}
 	if (owned)
 		dsl_dataset_disown(ds, FTAG);
@@ -1000,6 +1026,15 @@ dmu_recv_begin_check(void *arg, dmu_tx_t *tx)
 	    !spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_LZ4_COMPRESS))
 		return (SET_ERROR(ENOTSUP));
 
+	/*
+	 * The receiving code doesn't know how to translate large blocks
+	 * to smaller ones, so the pool must have the LARGE_BLOCKS
+	 * feature enabled if the stream has LARGE_BLOCKS.
+	 */
+	if ((featureflags & DMU_BACKUP_FEATURE_LARGE_BLOCKS) &&
+	    !spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_LARGE_BLOCKS))
+		return (SET_ERROR(ENOTSUP));
+
 	error = dsl_dataset_hold(dp, tofs, FTAG, &ds);
 	if (error == 0) {
 		/* target fs already exists; recv into temp clone */
@@ -1125,6 +1160,13 @@ dmu_recv_begin_sync(void *arg, dmu_tx_t *tx)
 	}
 	VERIFY0(dsl_dataset_own_obj(dp, dsobj, dmu_recv_tag, &newds));
 
+	if ((DMU_GET_FEATUREFLAGS(drrb->drr_versioninfo) &
+	    DMU_BACKUP_FEATURE_LARGE_BLOCKS) &&
+	    !newds->ds_large_blocks) {
+		dsl_dataset_activate_large_blocks_sync_impl(dsobj, tx);
+		newds->ds_large_blocks = B_TRUE;
+	}
+
 	dmu_buf_will_dirty(newds->ds_dbuf, tx);
 	dsl_dataset_phys(newds)->ds_flags |= DS_FLAG_INCONSISTENT;
 
@@ -1250,6 +1292,7 @@ restore_read(struct restorearg *ra, int len, char *buf)
 
 	/* some things will require 8-byte alignment, so everything must */
 	ASSERT0(len % 8);
+	ASSERT3U(len, <=, ra->bufsize);
 
 	while (done < len) {
 		ssize_t resid;
@@ -1391,7 +1434,7 @@ restore_object(struct restorearg *ra, objset_t *os, struct drr_object *drro)
 	    drro->drr_compress >= ZIO_COMPRESS_FUNCTIONS ||
 	    P2PHASE(drro->drr_blksz, SPA_MINBLOCKSIZE) ||
 	    drro->drr_blksz < SPA_MINBLOCKSIZE ||
-	    drro->drr_blksz > SPA_MAXBLOCKSIZE ||
+	    drro->drr_blksz > spa_maxblocksize(dmu_objset_spa(os)) ||
 	    drro->drr_bonuslen > DN_MAX_BONUSLEN) {
 		return (SET_ERROR(EINVAL));
 	}
@@ -1665,7 +1708,7 @@ restore_spill(struct restorearg *ra, objset_t *os, struct drr_spill *drrs)
 	int err;
 
 	if (drrs->drr_length < SPA_MINBLOCKSIZE ||
-	    drrs->drr_length > SPA_MAXBLOCKSIZE)
+	    drrs->drr_length > spa_maxblocksize(dmu_objset_spa(os)))
 		return (SET_ERROR(EINVAL));
 
 	data = restore_read(ra, drrs->drr_length, NULL);
@@ -1752,7 +1795,7 @@ dmu_recv_stream(dmu_recv_cookie_t *drc, vnode_t *vp, offset_t *voffp,
 	ra.cksum = drc->drc_cksum;
 	ra.vp = vp;
 	ra.voff = *voffp;
-	ra.bufsize = 1<<20;
+	ra.bufsize = SPA_MAXBLOCKSIZE;
 	ra.buf = vmem_alloc(ra.bufsize, KM_SLEEP);
 
 	/* these were verified in dmu_recv_begin */

module/zfs/dmu_tx.c

@@ -241,7 +241,7 @@ dmu_tx_count_write(dmu_tx_hold_t *txh, uint64_t off, uint64_t len)
 		return;
 
 	min_bs = SPA_MINBLOCKSHIFT;
-	max_bs = SPA_MAXBLOCKSHIFT;
+	max_bs = highbit64(txh->txh_tx->tx_objset->os_recordsize) - 1;
 	min_ibs = DN_MIN_INDBLKSHIFT;
 	max_ibs = DN_MAX_INDBLKSHIFT;
 
@@ -310,6 +310,14 @@ dmu_tx_count_write(dmu_tx_hold_t *txh, uint64_t off, uint64_t len)
 			 */
 			ASSERT(dn->dn_datablkshift != 0);
 			min_bs = max_bs = dn->dn_datablkshift;
+		} else {
+			/*
+			 * The blocksize can increase up to the recordsize,
+			 * or if it is already more than the recordsize,
+			 * up to the next power of 2.
+			 */
+			min_bs = highbit64(dn->dn_datablksz - 1);
+			max_bs = MAX(max_bs, highbit64(dn->dn_datablksz - 1));
 		}
 
 		/*
@@ -745,11 +753,11 @@ dmu_tx_hold_zap(dmu_tx_t *tx, uint64_t object, int add, const char *name)
 		bp = &dn->dn_phys->dn_blkptr[0];
 		if (dsl_dataset_block_freeable(dn->dn_objset->os_dsl_dataset,
 		    bp, bp->blk_birth))
-			txh->txh_space_tooverwrite += SPA_MAXBLOCKSIZE;
+			txh->txh_space_tooverwrite += MZAP_MAX_BLKSZ;
 		else
-			txh->txh_space_towrite += SPA_MAXBLOCKSIZE;
+			txh->txh_space_towrite += MZAP_MAX_BLKSZ;
 		if (!BP_IS_HOLE(bp))
-			txh->txh_space_tounref += SPA_MAXBLOCKSIZE;
+			txh->txh_space_tounref += MZAP_MAX_BLKSZ;
 		return;
 	}
 
@@ -1546,18 +1554,18 @@ dmu_tx_hold_spill(dmu_tx_t *tx, uint64_t object)
 
 	/* If blkptr doesn't exist then add space to towrite */
 	if (!(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR)) {
-		txh->txh_space_towrite += SPA_MAXBLOCKSIZE;
+		txh->txh_space_towrite += SPA_OLD_MAXBLOCKSIZE;
 	} else {
 		blkptr_t *bp;
 
 		bp = &dn->dn_phys->dn_spill;
 		if (dsl_dataset_block_freeable(dn->dn_objset->os_dsl_dataset,
 		    bp, bp->blk_birth))
-			txh->txh_space_tooverwrite += SPA_MAXBLOCKSIZE;
+			txh->txh_space_tooverwrite += SPA_OLD_MAXBLOCKSIZE;
 		else
-			txh->txh_space_towrite += SPA_MAXBLOCKSIZE;
+			txh->txh_space_towrite += SPA_OLD_MAXBLOCKSIZE;
 		if (!BP_IS_HOLE(bp))
-			txh->txh_space_tounref += SPA_MAXBLOCKSIZE;
+			txh->txh_space_tounref += SPA_OLD_MAXBLOCKSIZE;
 	}
 }
 

module/zfs/dnode.c

@@ -540,10 +540,10 @@ dnode_allocate(dnode_t *dn, dmu_object_type_t ot, int blocksize, int ibs,
 {
 	int i;
 
+	ASSERT3U(blocksize, <=,
+	    spa_maxblocksize(dmu_objset_spa(dn->dn_objset)));
 	if (blocksize == 0)
 		blocksize = 1 << zfs_default_bs;
-	else if (blocksize > SPA_MAXBLOCKSIZE)
-		blocksize = SPA_MAXBLOCKSIZE;
 	else
 		blocksize = P2ROUNDUP(blocksize, SPA_MINBLOCKSIZE);
 
@@ -624,7 +624,8 @@ dnode_reallocate(dnode_t *dn, dmu_object_type_t ot, int blocksize,
 	int nblkptr;
 
 	ASSERT3U(blocksize, >=, SPA_MINBLOCKSIZE);
-	ASSERT3U(blocksize, <=, SPA_MAXBLOCKSIZE);
+	ASSERT3U(blocksize, <=,
+	    spa_maxblocksize(dmu_objset_spa(dn->dn_objset)));
 	ASSERT0(blocksize % SPA_MINBLOCKSIZE);
 	ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT || dmu_tx_private_ok(tx));
 	ASSERT(tx->tx_txg != 0);
@@ -1377,10 +1378,9 @@ dnode_set_blksz(dnode_t *dn, uint64_t size, int ibs, dmu_tx_t *tx)
 	dmu_buf_impl_t *db;
 	int err;
 
+	ASSERT3U(size, <=, spa_maxblocksize(dmu_objset_spa(dn->dn_objset)));
 	if (size == 0)
 		size = SPA_MINBLOCKSIZE;
-	if (size > SPA_MAXBLOCKSIZE)
-		size = SPA_MAXBLOCKSIZE;
 	else
 		size = P2ROUNDUP(size, SPA_MINBLOCKSIZE);
 

module/zfs/dsl_dataset.c

@@ -51,6 +51,17 @@
 #include <sys/dsl_userhold.h>
 #include <sys/dsl_bookmark.h>
 
+/*
+ * The SPA supports block sizes up to 16MB.  However, very large blocks
+ * can have an impact on i/o latency (e.g. tying up a spinning disk for
+ * ~300ms), and also potentially on the memory allocator.  Therefore,
+ * we do not allow the recordsize to be set larger than zfs_max_recordsize
+ * (default 1MB).  Larger blocks can be created by changing this tunable,
+ * and pools with larger blocks can always be imported and used, regardless
+ * of this setting.
+ */
+int zfs_max_recordsize = 1 * 1024 * 1024;
+
 #define	SWITCH64(x, y) \
 	{ \
 		uint64_t __tmp = (x); \
@@ -60,8 +71,6 @@
 
 #define	DS_REF_MAX	(1ULL << 62)
 
-#define	DSL_DEADLIST_BLOCKSIZE	SPA_MAXBLOCKSIZE
-
 extern inline dsl_dataset_phys_t *dsl_dataset_phys(dsl_dataset_t *ds);
 
 /*
@@ -117,6 +126,8 @@ dsl_dataset_block_born(dsl_dataset_t *ds, const blkptr_t *bp, dmu_tx_t *tx)
 	dsl_dataset_phys(ds)->ds_compressed_bytes += compressed;
 	dsl_dataset_phys(ds)->ds_uncompressed_bytes += uncompressed;
 	dsl_dataset_phys(ds)->ds_unique_bytes += used;
+	if (BP_GET_LSIZE(bp) > SPA_OLD_MAXBLOCKSIZE)
+		ds->ds_need_large_blocks = B_TRUE;
 	mutex_exit(&ds->ds_lock);
 	dsl_dir_diduse_space(ds->ds_dir, DD_USED_HEAD, delta,
 	    compressed, uncompressed, tx);
@@ -414,6 +425,14 @@ dsl_dataset_hold_obj(dsl_pool_t *dp, uint64_t dsobj, void *tag,
 		list_create(&ds->ds_sendstreams, sizeof (dmu_sendarg_t),
 		    offsetof(dmu_sendarg_t, dsa_link));
 
+		if (doi.doi_type == DMU_OTN_ZAP_METADATA) {
+			err = zap_contains(mos, dsobj, DS_FIELD_LARGE_BLOCKS);
+			if (err == 0)
+				ds->ds_large_blocks = B_TRUE;
+			else
+				ASSERT3U(err, ==, ENOENT);
+		}
+
 		if (err == 0) {
 			err = dsl_dir_hold_obj(dp,
 			    dsl_dataset_phys(ds)->ds_dir_obj, NULL, ds,
@@ -730,6 +749,9 @@ dsl_dataset_create_sync_dd(dsl_dir_t *dd, dsl_dataset_t *origin,
 		dsphys->ds_flags |= dsl_dataset_phys(origin)->ds_flags &
 		    (DS_FLAG_INCONSISTENT | DS_FLAG_CI_DATASET);
 
+		if (origin->ds_large_blocks)
+			dsl_dataset_activate_large_blocks_sync_impl(dsobj, tx);
+
 		dmu_buf_will_dirty(origin->ds_dbuf, tx);
 		dsl_dataset_phys(origin)->ds_num_children++;
 
@@ -1253,6 +1275,9 @@ dsl_dataset_snapshot_sync_impl(dsl_dataset_t *ds, const char *snapname,
 	dsphys->ds_bp = dsl_dataset_phys(ds)->ds_bp;
 	dmu_buf_rele(dbuf, FTAG);
 
+	if (ds->ds_large_blocks)
+		dsl_dataset_activate_large_blocks_sync_impl(dsobj, tx);
+
 	ASSERT3U(ds->ds_prev != 0, ==,
 	    dsl_dataset_phys(ds)->ds_prev_snap_obj != 0);
 	if (ds->ds_prev) {
@@ -1541,6 +1566,11 @@ dsl_dataset_sync(dsl_dataset_t *ds, zio_t *zio, dmu_tx_t *tx)
 	dsl_dataset_phys(ds)->ds_fsid_guid = ds->ds_fsid_guid;
 
 	dmu_objset_sync(ds->ds_objset, zio, tx);
+
+	if (ds->ds_need_large_blocks && !ds->ds_large_blocks) {
+		dsl_dataset_activate_large_blocks_sync_impl(ds->ds_object, tx);
+		ds->ds_large_blocks = B_TRUE;
+	}
 }
 
 static void
@@ -3222,6 +3252,77 @@ dsl_dataset_space_wouldfree(dsl_dataset_t *firstsnap,
 	return (err);
 }
 
+static int
+dsl_dataset_activate_large_blocks_check(void *arg, dmu_tx_t *tx)
+{
+	const char *dsname = arg;
+	dsl_dataset_t *ds;
+	dsl_pool_t *dp = dmu_tx_pool(tx);
+	int error = 0;
+
+	if (!spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_LARGE_BLOCKS))
+		return (SET_ERROR(ENOTSUP));
+
+	ASSERT(spa_feature_is_enabled(dp->dp_spa,
+	    SPA_FEATURE_EXTENSIBLE_DATASET));
+
+	error = dsl_dataset_hold(dp, dsname, FTAG, &ds);
+	if (error != 0)
+		return (error);
+
+	if (ds->ds_large_blocks)
+		error = EALREADY;
+	dsl_dataset_rele(ds, FTAG);
+
+	return (error);
+}
+
+void
+dsl_dataset_activate_large_blocks_sync_impl(uint64_t dsobj, dmu_tx_t *tx)
+{
+	spa_t *spa = dmu_tx_pool(tx)->dp_spa;
+	objset_t *mos = dmu_tx_pool(tx)->dp_meta_objset;
+	uint64_t zero = 0;
+
+	spa_feature_incr(spa, SPA_FEATURE_LARGE_BLOCKS, tx);
+	dmu_object_zapify(mos, dsobj, DMU_OT_DSL_DATASET, tx);
+
+	VERIFY0(zap_add(mos, dsobj, DS_FIELD_LARGE_BLOCKS,
+	    sizeof (zero), 1, &zero, tx));
+}
+
+static void
+dsl_dataset_activate_large_blocks_sync(void *arg, dmu_tx_t *tx)
+{
+	const char *dsname = arg;
+	dsl_dataset_t *ds;
+
+	VERIFY0(dsl_dataset_hold(dmu_tx_pool(tx), dsname, FTAG, &ds));
+
+	dsl_dataset_activate_large_blocks_sync_impl(ds->ds_object, tx);
+	ASSERT(!ds->ds_large_blocks);
+	ds->ds_large_blocks = B_TRUE;
+	dsl_dataset_rele(ds, FTAG);
+}
+
+int
+dsl_dataset_activate_large_blocks(const char *dsname)
+{
+	int error;
+
+	error = dsl_sync_task(dsname,
+	    dsl_dataset_activate_large_blocks_check,
+	    dsl_dataset_activate_large_blocks_sync, (void *)dsname,
+	    1, ZFS_SPACE_CHECK_RESERVED);
+
+	/*
+	 * EALREADY indicates that this dataset already supports large blocks.
+	 */
+	if (error == EALREADY)
+		error = 0;
+	return (error);
+}
+
 /*
  * Return TRUE if 'earlier' is an earlier snapshot in 'later's timeline.
  * For example, they could both be snapshots of the same filesystem, and
@@ -3275,6 +3376,15 @@ dsl_dataset_zapify(dsl_dataset_t *ds, dmu_tx_t *tx)
 }
 
 #if defined(_KERNEL) && defined(HAVE_SPL)
+#if defined(_LP64)
+module_param(zfs_max_recordsize, int, 0644);
+MODULE_PARM_DESC(zfs_max_recordsize, "Max allowed record size");
+#else
+/* Limited to 1M on 32-bit platforms due to lack of virtual address space */
+module_param(zfs_max_recordsize, int, 0444);
+MODULE_PARM_DESC(zfs_max_recordsize, "Max allowed record size");
+#endif
+
 EXPORT_SYMBOL(dsl_dataset_hold);
 EXPORT_SYMBOL(dsl_dataset_hold_obj);
 EXPORT_SYMBOL(dsl_dataset_own);

module/zfs/dsl_deadlist.c

@@ -148,7 +148,7 @@ uint64_t
 dsl_deadlist_alloc(objset_t *os, dmu_tx_t *tx)
 {
 	if (spa_version(dmu_objset_spa(os)) < SPA_VERSION_DEADLISTS)
-		return (bpobj_alloc(os, SPA_MAXBLOCKSIZE, tx));
+		return (bpobj_alloc(os, SPA_OLD_MAXBLOCKSIZE, tx));
 	return (zap_create(os, DMU_OT_DEADLIST, DMU_OT_DEADLIST_HDR,
 	    sizeof (dsl_deadlist_phys_t), tx));
 }
@@ -185,7 +185,7 @@ dle_enqueue(dsl_deadlist_t *dl, dsl_deadlist_entry_t *dle,
 {
 	if (dle->dle_bpobj.bpo_object ==
 	    dmu_objset_pool(dl->dl_os)->dp_empty_bpobj) {
-		uint64_t obj = bpobj_alloc(dl->dl_os, SPA_MAXBLOCKSIZE, tx);
+		uint64_t obj = bpobj_alloc(dl->dl_os, SPA_OLD_MAXBLOCKSIZE, tx);
 		bpobj_close(&dle->dle_bpobj);
 		bpobj_decr_empty(dl->dl_os, tx);
 		VERIFY3U(0, ==, bpobj_open(&dle->dle_bpobj, dl->dl_os, obj));
@@ -259,7 +259,7 @@ dsl_deadlist_add_key(dsl_deadlist_t *dl, uint64_t mintxg, dmu_tx_t *tx)
 
 	dle = kmem_alloc(sizeof (*dle), KM_SLEEP);
 	dle->dle_mintxg = mintxg;
-	obj = bpobj_alloc_empty(dl->dl_os, SPA_MAXBLOCKSIZE, tx);
+	obj = bpobj_alloc_empty(dl->dl_os, SPA_OLD_MAXBLOCKSIZE, tx);
 	VERIFY3U(0, ==, bpobj_open(&dle->dle_bpobj, dl->dl_os, obj));
 	avl_add(&dl->dl_tree, dle);
 
@@ -344,7 +344,7 @@ dsl_deadlist_clone(dsl_deadlist_t *dl, uint64_t maxtxg,
 		if (dle->dle_mintxg >= maxtxg)
 			break;
 
-		obj = bpobj_alloc_empty(dl->dl_os, SPA_MAXBLOCKSIZE, tx);
+		obj = bpobj_alloc_empty(dl->dl_os, SPA_OLD_MAXBLOCKSIZE, tx);
 		VERIFY3U(0, ==, zap_add_int_key(dl->dl_os, newobj,
 		    dle->dle_mintxg, obj, tx));
 	}

module/zfs/dsl_destroy.c

@@ -277,6 +277,10 @@ dsl_destroy_snapshot_sync_impl(dsl_dataset_t *ds, boolean_t defer, dmu_tx_t *tx)
 
 	obj = ds->ds_object;
 
+	if (ds->ds_large_blocks) {
+		ASSERT0(zap_contains(mos, obj, DS_FIELD_LARGE_BLOCKS));
+		spa_feature_decr(dp->dp_spa, SPA_FEATURE_LARGE_BLOCKS, tx);
+	}
 	if (dsl_dataset_phys(ds)->ds_prev_snap_obj != 0) {
 		ASSERT3P(ds->ds_prev, ==, NULL);
 		VERIFY0(dsl_dataset_hold_obj(dp,
@@ -738,6 +742,9 @@ dsl_destroy_head_sync_impl(dsl_dataset_t *ds, dmu_tx_t *tx)
 		ASSERT0(ds->ds_reserved);
 	}
 
+	if (ds->ds_large_blocks)
+		spa_feature_decr(dp->dp_spa, SPA_FEATURE_LARGE_BLOCKS, tx);
+
 	dsl_scan_ds_destroyed(ds, tx);
 
 	obj = ds->ds_object;

module/zfs/dsl_pool.c

@@ -372,7 +372,7 @@ dsl_pool_create(spa_t *spa, nvlist_t *zplprops, uint64_t txg)
 		    FREE_DIR_NAME, &dp->dp_free_dir));
 
 		/* create and open the free_bplist */
-		obj = bpobj_alloc(dp->dp_meta_objset, SPA_MAXBLOCKSIZE, tx);
+		obj = bpobj_alloc(dp->dp_meta_objset, SPA_OLD_MAXBLOCKSIZE, tx);
 		VERIFY(zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 		    DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj, tx) == 0);
 		VERIFY0(bpobj_open(&dp->dp_free_bpobj,
@@ -804,7 +804,7 @@ dsl_pool_upgrade_dir_clones(dsl_pool_t *dp, dmu_tx_t *tx)
 	 * subobj support.  So call dmu_object_alloc() directly.
 	 */
 	obj = dmu_object_alloc(dp->dp_meta_objset, DMU_OT_BPOBJ,
-	    SPA_MAXBLOCKSIZE, DMU_OT_BPOBJ_HDR, sizeof (bpobj_phys_t), tx);
+	    SPA_OLD_MAXBLOCKSIZE, DMU_OT_BPOBJ_HDR, sizeof (bpobj_phys_t), tx);
 	VERIFY0(zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 	    DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj, tx));
 	VERIFY0(bpobj_open(&dp->dp_free_bpobj, dp->dp_meta_objset, obj));

module/zfs/sa.c

@@ -504,7 +504,7 @@ sa_resize_spill(sa_handle_t *hdl, uint32_t size, dmu_tx_t *tx)
 
 	if (size == 0) {
 		blocksize = SPA_MINBLOCKSIZE;
-	} else if (size > SPA_MAXBLOCKSIZE) {
+	} else if (size > SPA_OLD_MAXBLOCKSIZE) {
 		ASSERT(0);
 		return (SET_ERROR(EFBIG));
 	} else {
@@ -693,7 +693,7 @@ sa_build_layouts(sa_handle_t *hdl, sa_bulk_attr_t *attr_desc, int attr_count,
 	hdrsize = sa_find_sizes(sa, attr_desc, attr_count, hdl->sa_bonus,
 	    SA_BONUS, &spill_idx, &used, &spilling);
 
-	if (used > SPA_MAXBLOCKSIZE)
+	if (used > SPA_OLD_MAXBLOCKSIZE)
 		return (SET_ERROR(EFBIG));
 
 	VERIFY(0 == dmu_set_bonus(hdl->sa_bonus, spilling ?
@@ -717,7 +717,7 @@ sa_build_layouts(sa_handle_t *hdl, sa_bulk_attr_t *attr_desc, int attr_count,
 		    attr_count - spill_idx, hdl->sa_spill, SA_SPILL, &i,
 		    &spill_used, &dummy);
 
-		if (spill_used > SPA_MAXBLOCKSIZE)
+		if (spill_used > SPA_OLD_MAXBLOCKSIZE)
 			return (SET_ERROR(EFBIG));
 
 		if (BUF_SPACE_NEEDED(spill_used, spillhdrsize) >

module/zfs/spa.c

@@ -266,6 +266,14 @@ spa_prop_get_config(spa_t *spa, nvlist_t **nvp)
 		spa_prop_add_list(*nvp, ZPOOL_PROP_ALTROOT, spa->spa_root,
 		    0, ZPROP_SRC_LOCAL);
 
+	if (spa_feature_is_enabled(spa, SPA_FEATURE_LARGE_BLOCKS)) {
+		spa_prop_add_list(*nvp, ZPOOL_PROP_MAXBLOCKSIZE, NULL,
+		    MIN(zfs_max_recordsize, SPA_MAXBLOCKSIZE), ZPROP_SRC_NONE);
+	} else {
+		spa_prop_add_list(*nvp, ZPOOL_PROP_MAXBLOCKSIZE, NULL,
+		    SPA_OLD_MAXBLOCKSIZE, ZPROP_SRC_NONE);
+	}
+
 	if ((dp = list_head(&spa->spa_config_list)) != NULL) {
 		if (dp->scd_path == NULL) {
 			spa_prop_add_list(*nvp, ZPOOL_PROP_CACHEFILE,
@@ -482,7 +490,7 @@ spa_prop_validate(spa_t *spa, nvlist_t *props)
 
 			if (!error) {
 				objset_t *os;
-				uint64_t compress;
+				uint64_t propval;
 
 				if (strval == NULL || strval[0] == '\0') {
 					objnum = zpool_prop_default_numeric(
@@ -494,15 +502,25 @@ spa_prop_validate(spa_t *spa, nvlist_t *props)
 				if (error)
 					break;
 
-				/* Must be ZPL and not gzip compressed. */
+				/*
+				 * Must be ZPL, and its property settings
+				 * must be supported by GRUB (compression
+				 * is not gzip, and large blocks are not used).
+				 */
 
 				if (dmu_objset_type(os) != DMU_OST_ZFS) {
 					error = SET_ERROR(ENOTSUP);
 				} else if ((error =
 				    dsl_prop_get_int_ds(dmu_objset_ds(os),
 				    zfs_prop_to_name(ZFS_PROP_COMPRESSION),
-				    &compress)) == 0 &&
+				    &propval)) == 0 &&
-				    !BOOTFS_COMPRESS_VALID(compress)) {
+				    !BOOTFS_COMPRESS_VALID(propval)) {
+					error = SET_ERROR(ENOTSUP);
+				} else if ((error =
+				    dsl_prop_get_int_ds(dmu_objset_ds(os),
+				    zfs_prop_to_name(ZFS_PROP_RECORDSIZE),
+				    &propval)) == 0 &&
+				    propval > SPA_OLD_MAXBLOCKSIZE) {
 					error = SET_ERROR(ENOTSUP);
 				} else {
 					objnum = dmu_objset_id(os);

module/zfs/spa_history.c

@@ -89,7 +89,7 @@ spa_history_create_obj(spa_t *spa, dmu_tx_t *tx)
 
 	ASSERT(spa->spa_history == 0);
 	spa->spa_history = dmu_object_alloc(mos, DMU_OT_SPA_HISTORY,
-	    SPA_MAXBLOCKSIZE, DMU_OT_SPA_HISTORY_OFFSETS,
+	    SPA_OLD_MAXBLOCKSIZE, DMU_OT_SPA_HISTORY_OFFSETS,
 	    sizeof (spa_history_phys_t), tx);
 
 	VERIFY(zap_add(mos, DMU_POOL_DIRECTORY_OBJECT,

module/zfs/spa_misc.c

@@ -1985,6 +1985,15 @@ spa_debug_enabled(spa_t *spa)
 	return (spa->spa_debug);
 }
 
+int
+spa_maxblocksize(spa_t *spa)
+{
+	if (spa_feature_is_enabled(spa, SPA_FEATURE_LARGE_BLOCKS))
+		return (SPA_MAXBLOCKSIZE);
+	else
+		return (SPA_OLD_MAXBLOCKSIZE);
+}
+
 #if defined(_KERNEL) && defined(HAVE_SPL)
 /* Namespace manipulation */
 EXPORT_SYMBOL(spa_lookup);
@@ -2040,6 +2049,7 @@ EXPORT_SYMBOL(spa_suspended);
 EXPORT_SYMBOL(spa_bootfs);
 EXPORT_SYMBOL(spa_delegation);
 EXPORT_SYMBOL(spa_meta_objset);
+EXPORT_SYMBOL(spa_maxblocksize);
 
 /* Miscellaneous support routines */
 EXPORT_SYMBOL(spa_rename);

module/zfs/vdev.c

@@ -847,9 +847,9 @@ vdev_metaslab_init(vdev_t *vd, uint64_t txg)
 
 	/*
 	 * Compute the raidz-deflation ratio.  Note, we hard-code
-	 * in 128k (1 << 17) because it is the current "typical" blocksize.
+	 * in 128k (1 << 17) because it is the "typical" blocksize.
-	 * Even if SPA_MAXBLOCKSIZE changes, this algorithm must never change,
+	 * Even though SPA_MAXBLOCKSIZE changed, this algorithm can not change,
-	 * or we will inconsistently account for existing bp's.
+	 * otherwise it would inconsistently account for existing bp's.
 	 */
 	vd->vdev_deflate_ratio = (1 << 17) /
 	    (vdev_psize_to_asize(vd, 1 << 17) >> SPA_MINBLOCKSHIFT);

module/zfs/vdev_disk.c

@@ -552,9 +552,9 @@ retry:
 			goto retry;
 		}
 
-		dr->dr_bio[i] = bio_alloc(GFP_NOIO,
-		    bio_nr_pages(bio_ptr, bio_size));
 		/* bio_alloc() with __GFP_WAIT never returns NULL */
+		dr->dr_bio[i] = bio_alloc(GFP_NOIO,
+		    MIN(bio_nr_pages(bio_ptr, bio_size), BIO_MAX_PAGES));
 		if (unlikely(dr->dr_bio[i] == NULL)) {
 			vdev_disk_dio_free(dr);
 			return (ENOMEM);

module/zfs/vdev_queue.c

@@ -167,7 +167,7 @@ int zfs_vdev_async_write_active_max_dirty_percent = 60;
  * we include spans of optional I/Os to aid aggregation at the disk even when
  * they aren't able to help us aggregate at this level.
  */
-int zfs_vdev_aggregation_limit = SPA_MAXBLOCKSIZE;
+int zfs_vdev_aggregation_limit = SPA_OLD_MAXBLOCKSIZE;
 int zfs_vdev_read_gap_limit = 32 << 10;
 int zfs_vdev_write_gap_limit = 4 << 10;
 

module/zfs/zap_micro.c

@@ -34,6 +34,7 @@
 #include <sys/zap_leaf.h>
 #include <sys/avl.h>
 #include <sys/arc.h>
+#include <sys/dmu_objset.h>
 
 #ifdef _KERNEL
 #include <sys/sunddi.h>
@@ -654,9 +655,9 @@ zap_create_flags(objset_t *os, int normflags, zap_flags_t flags,
 	uint64_t obj = dmu_object_alloc(os, ot, 0, bonustype, bonuslen, tx);
 
 	ASSERT(leaf_blockshift >= SPA_MINBLOCKSHIFT &&
-	    leaf_blockshift <= SPA_MAXBLOCKSHIFT &&
+	    leaf_blockshift <= SPA_OLD_MAXBLOCKSHIFT &&
 	    indirect_blockshift >= SPA_MINBLOCKSHIFT &&
-	    indirect_blockshift <= SPA_MAXBLOCKSHIFT);
+	    indirect_blockshift <= SPA_OLD_MAXBLOCKSHIFT);
 
 	VERIFY(dmu_object_set_blocksize(os, obj,
 	    1ULL << leaf_blockshift, indirect_blockshift, tx) == 0);
@@ -1347,7 +1348,6 @@ zap_count_write(objset_t *os, uint64_t zapobj, const char *name, int add,
 	zap_t *zap;
 	int err = 0;
 
-
 	/*
 	 * Since, we don't have a name, we cannot figure out which blocks will
 	 * be affected in this operation. So, account for the worst case :
@@ -1360,7 +1360,7 @@ zap_count_write(objset_t *os, uint64_t zapobj, const char *name, int add,
 	 * large microzap results in a promotion to fatzap.
 	 */
 	if (name == NULL) {
-		*towrite += (3 + (add ? 4 : 0)) * SPA_MAXBLOCKSIZE;
+		*towrite += (3 + (add ? 4 : 0)) * SPA_OLD_MAXBLOCKSIZE;
 		return (err);
 	}
 
@@ -1384,7 +1384,7 @@ zap_count_write(objset_t *os, uint64_t zapobj, const char *name, int add,
 			/*
 			 * We treat this case as similar to (name == NULL)
 			 */
-			*towrite += (3 + (add ? 4 : 0)) * SPA_MAXBLOCKSIZE;
+			*towrite += (3 + (add ? 4 : 0)) * SPA_OLD_MAXBLOCKSIZE;
 		}
 	} else {
 		/*
@@ -1403,12 +1403,12 @@ zap_count_write(objset_t *os, uint64_t zapobj, const char *name, int add,
 		 *			ptrtbl blocks
 		 */
 		if (dmu_buf_freeable(zap->zap_dbuf))
-			*tooverwrite += SPA_MAXBLOCKSIZE;
+			*tooverwrite += MZAP_MAX_BLKSZ;
 		else
-			*towrite += SPA_MAXBLOCKSIZE;
+			*towrite += MZAP_MAX_BLKSZ;
 
 		if (add) {
-			*towrite += 4 * SPA_MAXBLOCKSIZE;
+			*towrite += 4 * MZAP_MAX_BLKSZ;
 		}
 	}
 

module/zfs/zfeature_common.c

@@ -56,7 +56,8 @@ valid_char(char c, boolean_t after_colon)
 {
 	return ((c >= 'a' && c <= 'z') ||
 	    (c >= '0' && c <= '9') ||
-	    c == (after_colon ? '_' : '.'));
+	    (after_colon && c == '_') ||
+	    (!after_colon && (c == '.' || c == '-')));
 }
 
 /*
@@ -230,4 +231,15 @@ zpool_feature_init(void)
 	    "com.delphix:embedded_data", "embedded_data",
 	    "Blocks which compress very well use even less space.",
 	    B_FALSE, B_TRUE, B_TRUE, NULL);
+
+	{
+	static const spa_feature_t large_blocks_deps[] = {
+		SPA_FEATURE_EXTENSIBLE_DATASET,
+		SPA_FEATURE_NONE
+	};
+	zfeature_register(SPA_FEATURE_LARGE_BLOCKS,
+	    "org.open-zfs:large_blocks", "large_blocks",
+	    "Support for blocks larger than 128KB.", B_FALSE, B_FALSE, B_FALSE,
+	    large_blocks_deps);
+	}
 }

module/zfs/zfs_ioctl.c

@@ -2392,7 +2392,7 @@ zfs_prop_set_special(const char *dsname, zprop_source_t source,
 	const char *propname = nvpair_name(pair);
 	zfs_prop_t prop = zfs_name_to_prop(propname);
 	uint64_t intval;
-	int err;
+	int err = -1;
 
 	if (prop == ZPROP_INVAL) {
 		if (zfs_prop_userquota(propname))
@@ -3790,8 +3790,7 @@ zfs_check_settable(const char *dsname, nvpair_t *pair, cred_t *cr)
 		 * the SPA supports it. We ignore any errors here since
 		 * we'll catch them later.
 		 */
-		if (nvpair_type(pair) == DATA_TYPE_UINT64 &&
+		if (nvpair_value_uint64(pair, &intval) == 0) {
-		    nvpair_value_uint64(pair, &intval) == 0) {
 			if (intval >= ZIO_COMPRESS_GZIP_1 &&
 			    intval <= ZIO_COMPRESS_GZIP_9 &&
 			    zfs_earlier_version(dsname,
@@ -3842,6 +3841,42 @@ zfs_check_settable(const char *dsname, nvpair_t *pair, cred_t *cr)
 			return (SET_ERROR(ENOTSUP));
 		break;
 
+	case ZFS_PROP_RECORDSIZE:
+		/* Record sizes above 128k need the feature to be enabled */
+		if (nvpair_value_uint64(pair, &intval) == 0 &&
+		    intval > SPA_OLD_MAXBLOCKSIZE) {
+			spa_t *spa;
+
+			/*
+			 * If this is a bootable dataset then
+			 * the we don't allow large (>128K) blocks,
+			 * because GRUB doesn't support them.
+			 */
+			if (zfs_is_bootfs(dsname) &&
+			    intval > SPA_OLD_MAXBLOCKSIZE) {
+				return (SET_ERROR(EDOM));
+			}
+
+			/*
+			 * We don't allow setting the property above 1MB,
+			 * unless the tunable has been changed.
+			 */
+			if (intval > zfs_max_recordsize ||
+			    intval > SPA_MAXBLOCKSIZE)
+				return (SET_ERROR(EDOM));
+
+			if ((err = spa_open(dsname, &spa, FTAG)) != 0)
+				return (err);
+
+			if (!spa_feature_is_enabled(spa,
+			    SPA_FEATURE_LARGE_BLOCKS)) {
+				spa_close(spa, FTAG);
+				return (SET_ERROR(ENOTSUP));
+			}
+			spa_close(spa, FTAG);
+		}
+		break;
+
 	case ZFS_PROP_SHARESMB:
 		if (zpl_earlier_version(dsname, ZPL_VERSION_FUID))
 			return (SET_ERROR(ENOTSUP));
@@ -4221,7 +4256,7 @@ out:
  * zc_fromobj	objsetid of incremental fromsnap (may be zero)
  * zc_guid	if set, estimate size of stream only.  zc_cookie is ignored.
  *		output size in zc_objset_type.
- * zc_flags	if =1, WRITE_EMBEDDED records are permitted
+ * zc_flags	lzc_send_flags
  *
  * outputs:
  * zc_objset_type	estimated size, if zc_guid is set
@@ -4233,6 +4268,7 @@ zfs_ioc_send(zfs_cmd_t *zc)
 	offset_t off;
 	boolean_t estimate = (zc->zc_guid != 0);
 	boolean_t embedok = (zc->zc_flags & 0x1);
+	boolean_t large_block_ok = (zc->zc_flags & 0x2);
 
 	if (zc->zc_obj != 0) {
 		dsl_pool_t *dp;
@@ -4294,7 +4330,8 @@ zfs_ioc_send(zfs_cmd_t *zc)
 
 		off = fp->f_offset;
 		error = dmu_send_obj(zc->zc_name, zc->zc_sendobj,
-		    zc->zc_fromobj, embedok, zc->zc_cookie, fp->f_vnode, &off);
+		    zc->zc_fromobj, embedok, large_block_ok,
+		    zc->zc_cookie, fp->f_vnode, &off);
 
 		if (VOP_SEEK(fp->f_vnode, fp->f_offset, &off, NULL) == 0)
 			fp->f_offset = off;
@@ -5160,6 +5197,8 @@ zfs_ioc_space_snaps(const char *lastsnap, nvlist_t *innvl, nvlist_t *outnvl)
  * innvl: {
  *     "fd" -> file descriptor to write stream to (int32)
  *     (optional) "fromsnap" -> full snap name to send an incremental from
+ *     (optional) "largeblockok" -> (value ignored)
+ *         indicates that blocks > 128KB are permitted
  *     (optional) "embedok" -> (value ignored)
  *         presence indicates DRR_WRITE_EMBEDDED records are permitted
  * }
@@ -5175,6 +5214,7 @@ zfs_ioc_send_new(const char *snapname, nvlist_t *innvl, nvlist_t *outnvl)
 	char *fromname = NULL;
 	int fd;
 	file_t *fp;
+	boolean_t largeblockok;
 	boolean_t embedok;
 
 	error = nvlist_lookup_int32(innvl, "fd", &fd);
@@ -5183,13 +5223,15 @@ zfs_ioc_send_new(const char *snapname, nvlist_t *innvl, nvlist_t *outnvl)
 
 	(void) nvlist_lookup_string(innvl, "fromsnap", &fromname);
 
+	largeblockok = nvlist_exists(innvl, "largeblockok");
 	embedok = nvlist_exists(innvl, "embedok");
 
 	if ((fp = getf(fd)) == NULL)
 		return (SET_ERROR(EBADF));
 
 	off = fp->f_offset;
-	error = dmu_send(snapname, fromname, embedok, fd, fp->f_vnode, &off);
+	error = dmu_send(snapname, fromname, embedok, largeblockok,
+	    fd, fp->f_vnode, &off);
 
 	if (VOP_SEEK(fp->f_vnode, fp->f_offset, &off, NULL) == 0)
 		fp->f_offset = off;

module/zfs/zfs_log.c

@@ -492,7 +492,7 @@ zfs_log_write(zilog_t *zilog, dmu_tx_t *tx, int txtype,
 		 * If the write would overflow the largest block then split it.
 		 */
 		if (write_state != WR_INDIRECT && resid > ZIL_MAX_LOG_DATA)
-			len = SPA_MAXBLOCKSIZE >> 1;
+			len = SPA_OLD_MAXBLOCKSIZE >> 1;
 		else
 			len = resid;
 

module/zfs/zfs_vfsops.c

@@ -188,10 +188,9 @@ static void
 blksz_changed_cb(void *arg, uint64_t newval)
 {
 	zfs_sb_t *zsb = arg;
-
+	ASSERT3U(newval, <=, spa_maxblocksize(dmu_objset_spa(zsb->z_os)));
-	if (newval < SPA_MINBLOCKSIZE ||
+	ASSERT3U(newval, >=, SPA_MINBLOCKSIZE);
-	    newval > SPA_MAXBLOCKSIZE || !ISP2(newval))
+	ASSERT(ISP2(newval));
-		newval = SPA_MAXBLOCKSIZE;
 
 	zsb->z_max_blksz = newval;
 }
@@ -672,7 +671,7 @@ zfs_sb_create(const char *osname, zfs_sb_t **zsbp)
 	 */
 	zsb->z_sb = NULL;
 	zsb->z_parent = zsb;
-	zsb->z_max_blksz = SPA_MAXBLOCKSIZE;
+	zsb->z_max_blksz = SPA_OLD_MAXBLOCKSIZE;
 	zsb->z_show_ctldir = ZFS_SNAPDIR_VISIBLE;
 	zsb->z_os = os;
 

module/zfs/zfs_vnops.c

@@ -771,8 +771,14 @@ zfs_write(struct inode *ip, uio_t *uio, int ioflag, cred_t *cr)
 			uint64_t new_blksz;
 
 			if (zp->z_blksz > max_blksz) {
+				/*
+				 * File's blocksize is already larger than the
+				 * "recordsize" property.  Only let it grow to
+				 * the next power of 2.
+				 */
 				ASSERT(!ISP2(zp->z_blksz));
-				new_blksz = MIN(end_size, SPA_MAXBLOCKSIZE);
+				new_blksz = MIN(end_size,
+				    1 << highbit64(zp->z_blksz));
 			} else {
 				new_blksz = MIN(end_size, max_blksz);
 			}

module/zfs/zfs_znode.c

@@ -61,6 +61,7 @@
 #endif /* _KERNEL */
 
 #include <sys/dmu.h>
+#include <sys/dmu_objset.h>
 #include <sys/refcount.h>
 #include <sys/stat.h>
 #include <sys/zap.h>
@@ -1304,8 +1305,13 @@ zfs_extend(znode_t *zp, uint64_t end)
 		 * We are growing the file past the current block size.
 		 */
 		if (zp->z_blksz > ZTOZSB(zp)->z_max_blksz) {
+			/*
+			 * File's blocksize is already larger than the
+			 * "recordsize" property.  Only let it grow to
+			 * the next power of 2.
+			 */
 			ASSERT(!ISP2(zp->z_blksz));
-			newblksz = MIN(end, SPA_MAXBLOCKSIZE);
+			newblksz = MIN(end, 1 << highbit64(zp->z_blksz));
 		} else {
 			newblksz = MIN(end, ZTOZSB(zp)->z_max_blksz);
 		}

module/zfs/zil.c

@@ -243,6 +243,7 @@ zil_read_log_block(zilog_t *zilog, const blkptr_t *bp, blkptr_t *nbp, void *dst,
 			    sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk)) {
 				error = SET_ERROR(ECKSUM);
 			} else {
+				ASSERT3U(len, <=, SPA_OLD_MAXBLOCKSIZE);
 				bcopy(lr, dst, len);
 				*end = (char *)dst + len;
 				*nbp = zilc->zc_next_blk;
@@ -257,6 +258,8 @@ zil_read_log_block(zilog_t *zilog, const blkptr_t *bp, blkptr_t *nbp, void *dst,
 			    (zilc->zc_nused > (size - sizeof (*zilc)))) {
 				error = SET_ERROR(ECKSUM);
 			} else {
+				ASSERT3U(zilc->zc_nused, <=,
+				    SPA_OLD_MAXBLOCKSIZE);
 				bcopy(lr, dst, zilc->zc_nused);
 				*end = (char *)dst + zilc->zc_nused;
 				*nbp = zilc->zc_next_blk;
@@ -342,7 +345,7 @@ zil_parse(zilog_t *zilog, zil_parse_blk_func_t *parse_blk_func,
 	 * If the log has been claimed, stop if we encounter a sequence
 	 * number greater than the highest claimed sequence number.
 	 */
-	lrbuf = zio_buf_alloc(SPA_MAXBLOCKSIZE);
+	lrbuf = zio_buf_alloc(SPA_OLD_MAXBLOCKSIZE);
 	zil_bp_tree_init(zilog);
 
 	for (blk = zh->zh_log; !BP_IS_HOLE(&blk); blk = next_blk) {
@@ -389,7 +392,7 @@ done:
 	    (max_blk_seq == claim_blk_seq && max_lr_seq == claim_lr_seq));
 
 	zil_bp_tree_fini(zilog);
-	zio_buf_free(lrbuf, SPA_MAXBLOCKSIZE);
+	zio_buf_free(lrbuf, SPA_OLD_MAXBLOCKSIZE);
 
 	return (error);
 }
@@ -941,7 +944,7 @@ zil_lwb_write_init(zilog_t *zilog, lwb_t *lwb)
  *
  * These must be a multiple of 4KB. Note only the amount used (again
  * aligned to 4KB) actually gets written. However, we can't always just
- * allocate SPA_MAXBLOCKSIZE as the slog space could be exhausted.
+ * allocate SPA_OLD_MAXBLOCKSIZE as the slog space could be exhausted.
  */
 uint64_t zil_block_buckets[] = {
     4096,		/* non TX_WRITE */
@@ -1023,7 +1026,7 @@ zil_lwb_write_start(zilog_t *zilog, lwb_t *lwb)
 		continue;
 	zil_blksz = zil_block_buckets[i];
 	if (zil_blksz == UINT64_MAX)
-		zil_blksz = SPA_MAXBLOCKSIZE;
+		zil_blksz = SPA_OLD_MAXBLOCKSIZE;
 	zilog->zl_prev_blks[zilog->zl_prev_rotor] = zil_blksz;
 	for (i = 0; i < ZIL_PREV_BLKS; i++)
 		zil_blksz = MAX(zil_blksz, zilog->zl_prev_blks[i]);

module/zfs/zio.c

@@ -24,6 +24,7 @@
  * Copyright (c) 2011 Nexenta Systems, Inc. All rights reserved.
  */
 
+#include <sys/sysmacros.h>
 #include <sys/zfs_context.h>
 #include <sys/fm/fs/zfs.h>
 #include <sys/spa.h>
@@ -107,9 +108,8 @@ zio_init(void)
 
 	/*
 	 * For small buffers, we want a cache for each multiple of
-	 * SPA_MINBLOCKSIZE.  For medium-size buffers, we want a cache
+	 * SPA_MINBLOCKSIZE.  For larger buffers, we want a cache
-	 * for each quarter-power of 2.  For large buffers, we want
+	 * for each quarter-power of 2.
-	 * a cache for each multiple of PAGESIZE.
 	 */
 	for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
 		size_t size = (c + 1) << SPA_MINBLOCKSHIFT;
@@ -117,7 +117,16 @@ zio_init(void)
 		size_t align = 0;
 		size_t cflags = (size > zio_buf_debug_limit) ? KMC_NODEBUG : 0;
 
-		while (p2 & (p2 - 1))
+#ifdef _ILP32
+		/*
+		 * Cache size limited to 1M on 32-bit platforms until ARC
+		 * buffers no longer require virtual address space.
+		 */
+		if (size > zfs_max_recordsize)
+			break;
+#endif
+
+		while (!ISP2(p2))
 			p2 &= p2 - 1;
 
 #ifndef _KERNEL
@@ -132,10 +141,8 @@ zio_init(void)
 #endif
 		if (size <= 4 * SPA_MINBLOCKSIZE) {
 			align = SPA_MINBLOCKSIZE;
-		} else if (IS_P2ALIGNED(size, PAGESIZE)) {
-			align = PAGESIZE;
 		} else if (IS_P2ALIGNED(size, p2 >> 2)) {
-			align = p2 >> 2;
+			align = MIN(p2 >> 2, PAGESIZE);
 		}
 
 		if (align != 0) {
@@ -174,6 +181,14 @@ zio_fini(void)
 	kmem_cache_t *last_data_cache = NULL;
 
 	for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
+#ifdef _ILP32
+		/*
+		 * Cache size limited to 1M on 32-bit platforms until ARC
+		 * buffers no longer require virtual address space.
+		 */
+		if (((c + 1) << SPA_MINBLOCKSHIFT) > zfs_max_recordsize)
+			break;
+#endif
 		if (zio_buf_cache[c] != last_cache) {
 			last_cache = zio_buf_cache[c];
 			kmem_cache_destroy(zio_buf_cache[c]);