Speculative prefetch for reordered requests

Before this change speculative prefetcher was able to detect a stream
only if all of its accesses are perfectly sequential.  It was easy to
implement and is perfectly fine for single-threaded applications.
Unfortunately multi-threaded network servers, such as iSCSI, SMB or
NFS usually have plenty of threads and may often reorder requests,
preventing successful speculation and prefetch.

This change allows speculative prefetcher to detect streams even if
requests are reordered by introducing a list of 9 non-contiguous
ranges up to 16MB ahead of current stream position and filling the
gaps as more requests arrive.  It also allows stream to proceed
even with holes up to a certain configurable threshold (25%).

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by:	Alexander Motin <mav@FreeBSD.org>
Sponsored by:	iXsystems, Inc.
Closes openzfs#16022

cmd/arc_summary

@@ -793,18 +793,27 @@ def section_dmu(kstats_dict):
 
     zfetch_stats = isolate_section('zfetchstats', kstats_dict)
 
-    zfetch_access_total = int(zfetch_stats['hits'])+int(zfetch_stats['misses'])
+    zfetch_access_total = int(zfetch_stats['hits']) +\
+        int(zfetch_stats['future']) + int(zfetch_stats['stride']) +\
+        int(zfetch_stats['past']) + int(zfetch_stats['misses'])
 
     prt_1('DMU predictive prefetcher calls:', f_hits(zfetch_access_total))
     prt_i2('Stream hits:',
            f_perc(zfetch_stats['hits'], zfetch_access_total),
            f_hits(zfetch_stats['hits']))
+    future = int(zfetch_stats['future']) + int(zfetch_stats['stride'])
+    prt_i2('Hits ahead of stream:', f_perc(future, zfetch_access_total),
+           f_hits(future))
+    prt_i2('Hits behind stream:',
+           f_perc(zfetch_stats['past'], zfetch_access_total),
+           f_hits(zfetch_stats['past']))
     prt_i2('Stream misses:',
            f_perc(zfetch_stats['misses'], zfetch_access_total),
            f_hits(zfetch_stats['misses']))
     prt_i2('Streams limit reached:',
            f_perc(zfetch_stats['max_streams'], zfetch_stats['misses']),
            f_hits(zfetch_stats['max_streams']))
+    prt_i1('Stream strides:', f_hits(zfetch_stats['stride']))
     prt_i1('Prefetches issued', f_hits(zfetch_stats['io_issued']))
     print()
 

include/sys/dmu_zfetch.h

@@ -45,18 +45,24 @@ typedef struct zfetch {
 	int		zf_numstreams;	/* number of zstream_t's */
 } zfetch_t;
 
+typedef struct zsrange {
+	uint16_t	start;
+	uint16_t	end;
+} zsrange_t;
+
+#define	ZFETCH_RANGES	9		/* Fits zstream_t into 128 bytes */
+
 typedef struct zstream {
+	list_node_t	zs_node;	/* link for zf_stream */
 	uint64_t	zs_blkid;	/* expect next access at this blkid */
+	uint_t		zs_atime;	/* time last prefetch issued */
+	zsrange_t	zs_ranges[ZFETCH_RANGES]; /* ranges from future */
 	unsigned int	zs_pf_dist;	/* data prefetch distance in bytes */
 	unsigned int	zs_ipf_dist;	/* L1 prefetch distance in bytes */
 	uint64_t	zs_pf_start;	/* first data block to prefetch */
 	uint64_t	zs_pf_end;	/* data block to prefetch up to */
 	uint64_t	zs_ipf_start;	/* first data block to prefetch L1 */
 	uint64_t	zs_ipf_end;	/* data block to prefetch L1 up to */
-
-	list_node_t	zs_node;	/* link for zf_stream */
-	hrtime_t	zs_atime;	/* time last prefetch issued */
-	zfetch_t	*zs_fetch;	/* parent fetch */
 	boolean_t	zs_missed;	/* stream saw cache misses */
 	boolean_t	zs_more;	/* need more distant prefetch */
 	zfs_refcount_t	zs_callers;	/* number of pending callers */
@@ -74,7 +80,7 @@ void		dmu_zfetch_init(zfetch_t *, struct dnode *);
 void		dmu_zfetch_fini(zfetch_t *);
 zstream_t	*dmu_zfetch_prepare(zfetch_t *, uint64_t, uint64_t, boolean_t,
     boolean_t);
-void		dmu_zfetch_run(zstream_t *, boolean_t, boolean_t);
+void		dmu_zfetch_run(zfetch_t *, zstream_t *, boolean_t, boolean_t);
 void		dmu_zfetch(zfetch_t *, uint64_t, uint64_t, boolean_t, boolean_t,
     boolean_t);
 

man/man4/zfs.4

@@ -564,6 +564,10 @@ However, this is limited by
 Maximum micro ZAP size.
 A micro ZAP is upgraded to a fat ZAP, once it grows beyond the specified size.
 .
+.It Sy zfetch_hole_shift Ns = Ns Sy 2 Pq uint
+Log2 fraction of holes in speculative prefetch stream allowed for it to
+proceed.
+.
 .It Sy zfetch_min_distance Ns = Ns Sy 4194304 Ns B Po 4 MiB Pc Pq uint
 Min bytes to prefetch per stream.
 Prefetch distance starts from the demand access size and quickly grows to
@@ -578,6 +582,13 @@ Max bytes to prefetch per stream.
 .It Sy zfetch_max_idistance Ns = Ns Sy 67108864 Ns B Po 64 MiB Pc Pq uint
 Max bytes to prefetch indirects for per stream.
 .
+.It Sy zfetch_max_reorder Ns = Ns Sy 16777216 Ns B Po 16 MiB Pc Pq uint
+Requests within this byte distance from the current prefetch stream position
+are considered parts of the stream, reordered due to parallel processing.
+Such requests do not advance the stream position immediately unless
+.Sy zfetch_hole_shift
+fill threshold is reached, but saved to fill holes in the stream later.
+.
 .It Sy zfetch_max_streams Ns = Ns Sy 8 Pq uint
 Max number of streams per zfetch (prefetch streams per file).
 .

module/zfs/dmu.c

@@ -569,8 +569,10 @@ dmu_buf_hold_array_by_dnode(dnode_t *dn, uint64_t offset, uint64_t length,
 	for (i = 0; i < nblks; i++) {
 		dmu_buf_impl_t *db = dbuf_hold(dn, blkid + i, tag);
 		if (db == NULL) {
-			if (zs)
-				dmu_zfetch_run(zs, missed, B_TRUE);
+			if (zs) {
+				dmu_zfetch_run(&dn->dn_zfetch, zs, missed,
+				    B_TRUE);
+			}
 			rw_exit(&dn->dn_struct_rwlock);
 			dmu_buf_rele_array(dbp, nblks, tag);
 			if (read)
@@ -606,7 +608,7 @@ dmu_buf_hold_array_by_dnode(dnode_t *dn, uint64_t offset, uint64_t length,
 		zfs_racct_write(length, nblks);
 
 	if (zs)
-		dmu_zfetch_run(zs, missed, B_TRUE);
+		dmu_zfetch_run(&dn->dn_zfetch, zs, missed, B_TRUE);
 	rw_exit(&dn->dn_struct_rwlock);
 
 	if (read) {