fat zap should prefetch when iterating

When iterating over a ZAP object, we're almost always certain to iterate
over the entire object. If there are multiple leaf blocks, we can
realize a performance win by issuing reads for all the leaf blocks in
parallel when the iteration begins.

For example, if we have 10,000 snapshots, "zfs destroy -nv
pool/fs@1%9999" can take 30 minutes when the cache is cold. This change
provides a >3x performance improvement, by issuing the reads for all ~64
blocks of each ZAP object in parallel.

Reviewed-by: Andreas Dilger <andreas.dilger@whamcloud.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Matthew Ahrens <mahrens@delphix.com>
External-issue: DLPX-58347
Closes openzfs#8862

include/sys/zap.h

@@ -21,7 +21,7 @@
 
 /*
  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
- * Copyright (c) 2013 by Delphix. All rights reserved.
+ * Copyright (c) 2012, 2018 by Delphix. All rights reserved.
  * Copyright 2017 Nexenta Systems, Inc.
  */
 
@@ -350,6 +350,7 @@ typedef struct zap_cursor {
 	uint64_t zc_serialized;
 	uint64_t zc_hash;
 	uint32_t zc_cd;
+	boolean_t zc_prefetch;
 } zap_cursor_t;
 
 typedef struct {
@@ -375,7 +376,9 @@ typedef struct {
  * Initialize a zap cursor, pointing to the "first" attribute of the
  * zapobj.  You must _fini the cursor when you are done with it.
  */
-void zap_cursor_init(zap_cursor_t *zc, objset_t *ds, uint64_t zapobj);
+void zap_cursor_init(zap_cursor_t *zc, objset_t *os, uint64_t zapobj);
+void zap_cursor_init_noprefetch(zap_cursor_t *zc, objset_t *os,
+    uint64_t zapobj);
 void zap_cursor_fini(zap_cursor_t *zc);
 
 /*

man/man5/zfs-module-parameters.5

@@ -104,6 +104,18 @@ to a log2 fraction of the target arc size.
 Default value: \fB6\fR.
 .RE
 
+.sp
+.ne 2
+.na
+\fBdmu_prefetch_max\fR (int)
+.ad
+.RS 12n
+Limit the amount we can prefetch with one call to this amount (in bytes).
+This helps to limit the amount of memory that can be used by prefetching.
+.sp
+Default value: \fB134,217,728\fR (128MB).
+.RE
+
 .sp
 .ne 2
 .na
@@ -502,6 +514,19 @@ regular reads (but there's no reason it has to be the same).
 Default value: \fB32,768\fR.
 .RE
 
+.sp
+.ne 2
+.na
+\fBzap_iterate_prefetch\fR (int)
+.ad
+.RS 12n
+If this is set, when we start iterating over a ZAP object, zfs will prefetch
+the entire object (all leaf blocks).  However, this is limited by
+\fBdmu_prefetch_max\fR.
+.sp
+Use \fB1\fR for on (default) and \fB0\fR for off.
+.RE
+
 .sp
 .ne 2
 .na

module/zfs/ddt_zap.c

@@ -21,6 +21,7 @@
 
 /*
  * Copyright (c) 2009, 2010, Oracle and/or its affiliates. All rights reserved.
+ * Copyright (c) 2018 by Delphix. All rights reserved.
  */
 
 #include <sys/zfs_context.h>
@@ -117,7 +118,18 @@ ddt_zap_walk(objset_t *os, uint64_t object, ddt_entry_t *dde, uint64_t *walk)
 	zap_attribute_t za;
 	int error;
 
-	zap_cursor_init_serialized(&zc, os, object, *walk);
+	if (*walk == 0) {
+		/*
+		 * We don't want to prefetch the entire ZAP object, because
+		 * it can be enormous.  Also the primary use of DDT iteration
+		 * is for scrubbing, in which case we will be issuing many
+		 * scrub I/Os for each ZAP block that we read in, so
+		 * reading the ZAP is unlikely to be the bottleneck.
+		 */
+		zap_cursor_init_noprefetch(&zc, os, object);
+	} else {
+		zap_cursor_init_serialized(&zc, os, object, *walk);
+	}
 	if ((error = zap_cursor_retrieve(&zc, &za)) == 0) {
 		uchar_t cbuf[sizeof (dde->dde_phys) + 1];
 		uint64_t csize = za.za_num_integers;

module/zfs/dmu.c

@@ -81,6 +81,13 @@ int zfs_dmu_offset_next_sync = 0;
  */
 int zfs_object_remap_one_indirect_delay_ms = 0;
 
+/*
+ * Limit the amount we can prefetch with one call to this amount.  This
+ * helps to limit the amount of memory that can be used by prefetching.
+ * Larger objects should be prefetched a bit at a time.
+ */
+int dmu_prefetch_max = 8 * SPA_MAXBLOCKSIZE;
+
 const dmu_object_type_info_t dmu_ot[DMU_OT_NUMTYPES] = {
 	{DMU_BSWAP_UINT8,  TRUE,  FALSE, FALSE, "unallocated"		},
 	{DMU_BSWAP_ZAP,    TRUE,  TRUE,  FALSE, "object directory"	},
@@ -667,6 +674,11 @@ dmu_prefetch(objset_t *os, uint64_t object, int64_t level, uint64_t offset,
 		return;
 	}
 
+	/*
+	 * See comment before the definition of dmu_prefetch_max.
+	 */
+	len = MIN(len, dmu_prefetch_max);
+
 	/*
 	 * XXX - Note, if the dnode for the requested object is not
 	 * already cached, we will do a *synchronous* read in the
@@ -2629,6 +2641,10 @@ module_param(zfs_dmu_offset_next_sync, int, 0644);
 MODULE_PARM_DESC(zfs_dmu_offset_next_sync,
 	"Enable forcing txg sync to find holes");
 
+module_param(dmu_prefetch_max, int, 0644);
+MODULE_PARM_DESC(dmu_prefetch_max,
+	"Limit one prefetch call to this size");
+
 /* END CSTYLED */
 
 #endif

module/zfs/zap.c

@@ -20,7 +20,7 @@
  */
 /*
  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
- * Copyright (c) 2012, 2016 by Delphix. All rights reserved.
+ * Copyright (c) 2012, 2018 by Delphix. All rights reserved.
  * Copyright (c) 2014 Spectra Logic Corporation, All rights reserved.
  */
 
@@ -49,6 +49,36 @@
 #include <sys/zap_impl.h>
 #include <sys/zap_leaf.h>
 
+/*
+ * If zap_iterate_prefetch is set, we will prefetch the entire ZAP object
+ * (all leaf blocks) when we start iterating over it.
+ *
+ * For zap_cursor_init(), the callers all intend to iterate through all the
+ * entries.  There are a few cases where an error (typically i/o error) could
+ * cause it to bail out early.
+ *
+ * For zap_cursor_init_serialized(), there are callers that do the iteration
+ * outside of ZFS.  Typically they would iterate over everything, but we
+ * don't have control of that.  E.g. zfs_ioc_snapshot_list_next(),
+ * zcp_snapshots_iter(), and other iterators over things in the MOS - these
+ * are called by /sbin/zfs and channel programs.  The other example is
+ * zfs_readdir() which iterates over directory entries for the getdents()
+ * syscall.  /sbin/ls iterates to the end (unless it receives a signal), but
+ * userland doesn't have to.
+ *
+ * Given that the ZAP entries aren't returned in a specific order, the only
+ * legitimate use cases for partial iteration would be:
+ *
+ * 1. Pagination: e.g. you only want to display 100 entries at a time, so you
+ *    get the first 100 and then wait for the user to hit "next page", which
+ *    they may never do).
+ *
+ * 2. You want to know if there are more than X entries, without relying on
+ *    the zfs-specific implementation of the directory's st_size (which is
+ *    the number of entries).
+ */
+int zap_iterate_prefetch = B_TRUE;
+
 int fzap_default_block_shift = 14; /* 16k blocksize */
 
 extern inline zap_phys_t *zap_f_phys(zap_t *zap);
@@ -1189,6 +1219,21 @@ fzap_cursor_retrieve(zap_t *zap, zap_cursor_t *zc, zap_attribute_t *za)
 	/* retrieve the next entry at or after zc_hash/zc_cd */
 	/* if no entry, return ENOENT */
 
+	/*
+	 * If we are reading from the beginning, we're almost certain to
+	 * iterate over the entire ZAP object.  If there are multiple leaf
+	 * blocks (freeblk > 2), prefetch the whole object (up to
+	 * dmu_prefetch_max bytes), so that we read the leaf blocks
+	 * concurrently. (Unless noprefetch was requested via
+	 * zap_cursor_init_noprefetch()).
+	 */
+	if (zc->zc_hash == 0 && zap_iterate_prefetch &&
+	    zc->zc_prefetch && zap_f_phys(zap)->zap_freeblk > 2) {
+		dmu_prefetch(zc->zc_objset, zc->zc_zapobj, 0, 0,
+		    zap_f_phys(zap)->zap_freeblk << FZAP_BLOCK_SHIFT(zap),
+		    ZIO_PRIORITY_ASYNC_READ);
+	}
+
 	if (zc->zc_leaf &&
 	    (ZAP_HASH_IDX(zc->zc_hash,
 	    zap_leaf_phys(zc->zc_leaf)->l_hdr.lh_prefix_len) !=
@@ -1333,3 +1378,12 @@ fzap_get_stats(zap_t *zap, zap_stats_t *zs)
 		}
 	}
 }
+
+#if defined(_KERNEL)
+/* BEGIN CSTYLED */
+module_param(zap_iterate_prefetch, int, 0644);
+MODULE_PARM_DESC(zap_iterate_prefetch,
+	"When iterating ZAP object, prefetch it");
+
+/* END CSTYLED */
+#endif

module/zfs/zap_micro.c

@@ -21,7 +21,7 @@
 
 /*
  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
- * Copyright (c) 2011, 2017 by Delphix. All rights reserved.
+ * Copyright (c) 2011, 2018 by Delphix. All rights reserved.
  * Copyright (c) 2014 Spectra Logic Corporation, All rights reserved.
  * Copyright 2017 Nexenta Systems, Inc.
  */
@@ -1472,9 +1472,9 @@ zap_remove_uint64(objset_t *os, uint64_t zapobj, const uint64_t *key,
  * Routines for iterating over the attributes.
  */
 
-void
-zap_cursor_init_serialized(zap_cursor_t *zc, objset_t *os, uint64_t zapobj,
-    uint64_t serialized)
+static void
+zap_cursor_init_impl(zap_cursor_t *zc, objset_t *os, uint64_t zapobj,
+    uint64_t serialized, boolean_t prefetch)
 {
 	zc->zc_objset = os;
 	zc->zc_zap = NULL;
@@ -1483,12 +1483,33 @@ zap_cursor_init_serialized(zap_cursor_t *zc, objset_t *os, uint64_t zapobj,
 	zc->zc_serialized = serialized;
 	zc->zc_hash = 0;
 	zc->zc_cd = 0;
+	zc->zc_prefetch = prefetch;
+}
+void
+zap_cursor_init_serialized(zap_cursor_t *zc, objset_t *os, uint64_t zapobj,
+    uint64_t serialized)
+{
+	zap_cursor_init_impl(zc, os, zapobj, serialized, B_TRUE);
 }
 
+/*
+ * Initialize a cursor at the beginning of the ZAP object.  The entire
+ * ZAP object will be prefetched.
+ */
 void
 zap_cursor_init(zap_cursor_t *zc, objset_t *os, uint64_t zapobj)
 {
-	zap_cursor_init_serialized(zc, os, zapobj, 0);
+	zap_cursor_init_impl(zc, os, zapobj, 0, B_TRUE);
+}
+
+/*
+ * Initialize a cursor at the beginning, but request that we not prefetch
+ * the entire ZAP object.
+ */
+void
+zap_cursor_init_noprefetch(zap_cursor_t *zc, objset_t *os, uint64_t zapobj)
+{
+	zap_cursor_init_impl(zc, os, zapobj, 0, B_FALSE);
 }
 
 void