More adaptive ARC eviction

Traditionally ARC adaptation was limited to MRU/MFU distribution.  But
for years people with metadata-centric workload demanded mechanisms to
also manage data/metadata distribution, that in original ZFS was just
a FIFO.  As result ZFS effectively got separate states for data and
metadata, minimum and maximum metadata limits etc, but it all required
manual tuning, was not adaptive and in its heart remained a bad FIFO.

This change removes most of existing eviction logic, rewriting it from
scratch.  This makes MRU/MFU adaptation individual for data and meta-
data, same as the distribution between data and metadata themselves.
Since most of required states separation was already done, it only
required to make arcs_size state field specific per data/metadata.

The adaptation logic is still based on previous concept of ghost hits,
just now it balances ARC capacity between 4 states: MRU data, MRU
metadata, MFU data and MFU metadata.  To simplify arc_c changes instead
of arc_p measured in bytes, this code uses 3 variable arc_meta, arc_pd
and arc_pm, representing ARC balance between metadata and data, MRU and
MFU for data, and MRU and MFU for metadata respectively as 32-bit fixed
point fractions.  Since we care about the math result only when need to
evict, this moves all the logic from arc_adapt() to arc_evict(), that
reduces per-block overhead, since per-block operations are limited to
stats collection, now moved from arc_adapt() to arc_access() and using
cheaper wmsums.  This also allows to remove ugly ARC_HDR_DO_ADAPT flag
from many places.

This change also removes number of metadata specific tunables, part of
which were actually not functioning correctly, since not all metadata
are equal and some (like L2ARC headers) are not really evictable.
Instead it introduced single opaque knob zfs_arc_meta_balance, tuning
ARC's reaction on ghost hits, allowing administrator give more or less
preference to metadata without setting strict limits.

Some of old code parts like arc_evict_meta() are just removed, because
since introduction of ABD ARC they really make no sense: only headers
referenced by small number of buffers are not evictable, and they are
really not evictable no matter what this code do.  Instead just call
arc_prune_async() if too much metadata appear not evictable.

Reviewed-by: Brian Behlendorf <[email protected]>
Reviewed-by: Allan Jude <[email protected]>
Signed-off-by: Alexander Motin <[email protected]>
Sponsored by: iXsystems, Inc.
Closes #14359

Author: amotin

cmd/arc_summary

@@ -270,16 +270,14 @@ def draw_graph(kstats_dict):
     arc_perc = f_perc(arc_stats['size'], arc_stats['c_max'])
     mfu_size = f_bytes(arc_stats['mfu_size'])
     mru_size = f_bytes(arc_stats['mru_size'])
-    meta_limit = f_bytes(arc_stats['arc_meta_limit'])
     meta_size = f_bytes(arc_stats['arc_meta_used'])
     dnode_limit = f_bytes(arc_stats['arc_dnode_limit'])
     dnode_size = f_bytes(arc_stats['dnode_size'])
 
-    info_form = ('ARC: {0} ({1})  MFU: {2}  MRU: {3}  META: {4} ({5}) '
-                 'DNODE {6} ({7})')
+    info_form = ('ARC: {0} ({1})  MFU: {2}  MRU: {3}  META: {4} '
+                 'DNODE {5} ({6})')
     info_line = info_form.format(arc_size, arc_perc, mfu_size, mru_size,
-                                 meta_size, meta_limit, dnode_size,
-                                 dnode_limit)
+                                 meta_size, dnode_size, dnode_limit)
     info_spc = ' '*int((GRAPH_WIDTH-len(info_line))/2)
     info_line = GRAPH_INDENT+info_spc+info_line
 
@@ -558,16 +556,28 @@ def section_arc(kstats_dict):
     arc_target_size = arc_stats['c']
     arc_max = arc_stats['c_max']
     arc_min = arc_stats['c_min']
-    anon_size = arc_stats['anon_size']
-    mfu_size = arc_stats['mfu_size']
-    mru_size = arc_stats['mru_size']
-    mfug_size = arc_stats['mfu_ghost_size']
-    mrug_size = arc_stats['mru_ghost_size']
-    unc_size = arc_stats['uncached_size']
-    meta_limit = arc_stats['arc_meta_limit']
-    meta_size = arc_stats['arc_meta_used']
+    meta = arc_stats['meta']
+    pd = arc_stats['pd']
+    pm = arc_stats['pm']
+    anon_data = arc_stats['anon_data']
+    anon_metadata = arc_stats['anon_metadata']
+    mfu_data = arc_stats['mfu_data']
+    mfu_metadata = arc_stats['mfu_metadata']
+    mru_data = arc_stats['mru_data']
+    mru_metadata = arc_stats['mru_metadata']
+    mfug_data = arc_stats['mfu_ghost_data']
+    mfug_metadata = arc_stats['mfu_ghost_metadata']
+    mrug_data = arc_stats['mru_ghost_data']
+    mrug_metadata = arc_stats['mru_ghost_metadata']
+    unc_data = arc_stats['uncached_data']
+    unc_metadata = arc_stats['uncached_metadata']
+    bonus_size = arc_stats['bonus_size']
     dnode_limit = arc_stats['arc_dnode_limit']
     dnode_size = arc_stats['dnode_size']
+    dbuf_size = arc_stats['dbuf_size']
+    hdr_size = arc_stats['hdr_size']
+    l2_hdr_size = arc_stats['l2_hdr_size']
+    abd_chunk_waste_size = arc_stats['abd_chunk_waste_size']
     target_size_ratio = '{0}:1'.format(int(arc_max) // int(arc_min))
 
     prt_2('ARC size (current):',
@@ -578,25 +588,56 @@ def section_arc(kstats_dict):
            f_perc(arc_min, arc_max), f_bytes(arc_min))
     prt_i2('Max size (high water):',
            target_size_ratio, f_bytes(arc_max))
-    caches_size = int(anon_size)+int(mfu_size)+int(mru_size)+int(unc_size)
-    prt_i2('Anonymouns data size:',
-           f_perc(anon_size, caches_size), f_bytes(anon_size))
-    prt_i2('Most Frequently Used (MFU) cache size:',
-           f_perc(mfu_size, caches_size), f_bytes(mfu_size))
-    prt_i2('Most Recently Used (MRU) cache size:',
-           f_perc(mru_size, caches_size), f_bytes(mru_size))
-    prt_i1('Most Frequently Used (MFU) ghost size:', f_bytes(mfug_size))
-    prt_i1('Most Recently Used (MRU) ghost size:', f_bytes(mrug_size))
+    caches_size = int(anon_data)+int(anon_metadata)+\
+        int(mfu_data)+int(mfu_metadata)+int(mru_data)+int(mru_metadata)+\
+        int(unc_data)+int(unc_metadata)
+    prt_i2('Anonymous data size:',
+           f_perc(anon_data, caches_size), f_bytes(anon_data))
+    prt_i2('Anonymous metadata size:',
+           f_perc(anon_metadata, caches_size), f_bytes(anon_metadata))
+    s = 4294967296
+    v = (s-int(pd))*(s-int(meta))/s
+    prt_i2('MFU data target:', f_perc(v, s),
+        f_bytes(v / 65536 * caches_size / 65536))
+    prt_i2('MFU data size:',
+           f_perc(mfu_data, caches_size), f_bytes(mfu_data))
+    prt_i1('MFU ghost data size:', f_bytes(mfug_data))
+    v = (s-int(pm))*int(meta)/s
+    prt_i2('MFU metadata target:', f_perc(v, s),
+        f_bytes(v / 65536 * caches_size / 65536))
+    prt_i2('MFU metadata size:',
+           f_perc(mfu_metadata, caches_size), f_bytes(mfu_metadata))
+    prt_i1('MFU ghost metadata size:', f_bytes(mfug_metadata))
+    v = int(pd)*(s-int(meta))/s
+    prt_i2('MRU data target:', f_perc(v, s),
+        f_bytes(v / 65536 * caches_size / 65536))
+    prt_i2('MRU data size:',
+           f_perc(mru_data, caches_size), f_bytes(mru_data))
+    prt_i1('MRU ghost data size:', f_bytes(mrug_data))
+    v = int(pm)*int(meta)/s
+    prt_i2('MRU metadata target:', f_perc(v, s),
+        f_bytes(v / 65536 * caches_size / 65536))
+    prt_i2('MRU metadata size:',
+           f_perc(mru_metadata, caches_size), f_bytes(mru_metadata))
+    prt_i1('MRU ghost metadata size:', f_bytes(mrug_metadata))
     prt_i2('Uncached data size:',
-           f_perc(unc_size, caches_size), f_bytes(unc_size))
-    prt_i2('Metadata cache size (hard limit):',
-           f_perc(meta_limit, arc_max), f_bytes(meta_limit))
-    prt_i2('Metadata cache size (current):',
-           f_perc(meta_size, meta_limit), f_bytes(meta_size))
-    prt_i2('Dnode cache size (hard limit):',
-           f_perc(dnode_limit, meta_limit), f_bytes(dnode_limit))
-    prt_i2('Dnode cache size (current):',
+           f_perc(unc_data, caches_size), f_bytes(unc_data))
+    prt_i2('Uncached metadata size:',
+           f_perc(unc_metadata, caches_size), f_bytes(unc_metadata))
+    prt_i2('Bonus size:',
+           f_perc(bonus_size, arc_size), f_bytes(bonus_size))
+    prt_i2('Dnode cache target:',
+           f_perc(dnode_limit, arc_max), f_bytes(dnode_limit))
+    prt_i2('Dnode cache size:',
            f_perc(dnode_size, dnode_limit), f_bytes(dnode_size))
+    prt_i2('Dbuf size:',
+           f_perc(dbuf_size, arc_size), f_bytes(dbuf_size))
+    prt_i2('Header size:',
+           f_perc(hdr_size, arc_size), f_bytes(hdr_size))
+    prt_i2('L2 header size:',
+           f_perc(l2_hdr_size, arc_size), f_bytes(l2_hdr_size))
+    prt_i2('ABD chunk waste size:',
+           f_perc(abd_chunk_waste_size, arc_size), f_bytes(abd_chunk_waste_size))
     print()
 
     print('ARC hash breakdown:')

cmd/zdb/zdb.c

@@ -117,7 +117,6 @@ zdb_ot_name(dmu_object_type_t type)
 
 extern int reference_tracking_enable;
 extern int zfs_recover;
-extern unsigned long zfs_arc_meta_min, zfs_arc_meta_limit;
 extern uint_t zfs_vdev_async_read_max_active;
 extern boolean_t spa_load_verify_dryrun;
 extern boolean_t spa_mode_readable_spacemaps;
@@ -8809,8 +8808,8 @@ main(int argc, char **argv)
 	 * ZDB does not typically re-read blocks; therefore limit the ARC
 	 * to 256 MB, which can be used entirely for metadata.
 	 */
-	zfs_arc_min = zfs_arc_meta_min = 2ULL << SPA_MAXBLOCKSHIFT;
-	zfs_arc_max = zfs_arc_meta_limit = 256 * 1024 * 1024;
+	zfs_arc_min = 2ULL << SPA_MAXBLOCKSHIFT;
+	zfs_arc_max = 256 * 1024 * 1024;
 #endif
 
 	/*

include/sys/arc.h

@@ -200,7 +200,6 @@ struct arc_buf {
 };
 
 typedef enum arc_buf_contents {
-	ARC_BUFC_INVALID,			/* invalid type */
 	ARC_BUFC_DATA,				/* buffer contains data */
 	ARC_BUFC_METADATA,			/* buffer contains metadata */
 	ARC_BUFC_NUMTYPES

include/sys/arc_impl.h

@@ -82,15 +82,18 @@ typedef struct arc_state {
 	 * supports the "dbufs" kstat
 	 */
 	arc_state_type_t arcs_state;
+	/*
+	 * total amount of data in this state.
+	 */
+	zfs_refcount_t arcs_size[ARC_BUFC_NUMTYPES] ____cacheline_aligned;
 	/*
 	 * total amount of evictable data in this state
 	 */
-	zfs_refcount_t arcs_esize[ARC_BUFC_NUMTYPES] ____cacheline_aligned;
+	zfs_refcount_t arcs_esize[ARC_BUFC_NUMTYPES];
 	/*
-	 * total amount of data in this state; this includes: evictable,
-	 * non-evictable, ARC_BUFC_DATA, and ARC_BUFC_METADATA.
+	 * amount of hit bytes for this state (counted only for ghost states)
 	 */
-	zfs_refcount_t arcs_size;
+	wmsum_t arcs_hits[ARC_BUFC_NUMTYPES];
 } arc_state_t;
 
 typedef struct arc_callback arc_callback_t;
@@ -358,8 +361,9 @@ typedef struct l2arc_lb_ptr_buf {
 #define	L2BLK_SET_PREFETCH(field, x)	BF64_SET((field), 39, 1, x)
 #define	L2BLK_GET_CHECKSUM(field)	BF64_GET((field), 40, 8)
 #define	L2BLK_SET_CHECKSUM(field, x)	BF64_SET((field), 40, 8, x)
-#define	L2BLK_GET_TYPE(field)		BF64_GET((field), 48, 8)
-#define	L2BLK_SET_TYPE(field, x)	BF64_SET((field), 48, 8, x)
+/* +/- 1 here are to keep compatibility after ARC_BUFC_INVALID removal. */
+#define	L2BLK_GET_TYPE(field)		(BF64_GET((field), 48, 8) - 1)
+#define	L2BLK_SET_TYPE(field, x)	BF64_SET((field), 48, 8, (x) + 1)
 #define	L2BLK_GET_PROTECTED(field)	BF64_GET((field), 56, 1)
 #define	L2BLK_SET_PROTECTED(field, x)	BF64_SET((field), 56, 1, x)
 #define	L2BLK_GET_STATE(field)		BF64_GET((field), 57, 4)
@@ -582,7 +586,9 @@ typedef struct arc_stats {
 	kstat_named_t arcstat_hash_collisions;
 	kstat_named_t arcstat_hash_chains;
 	kstat_named_t arcstat_hash_chain_max;
-	kstat_named_t arcstat_p;
+	kstat_named_t arcstat_meta;
+	kstat_named_t arcstat_pd;
+	kstat_named_t arcstat_pm;
 	kstat_named_t arcstat_c;
 	kstat_named_t arcstat_c_min;
 	kstat_named_t arcstat_c_max;
@@ -655,6 +661,8 @@ typedef struct arc_stats {
 	 * are all included in this value.
 	 */
 	kstat_named_t arcstat_anon_size;
+	kstat_named_t arcstat_anon_data;
+	kstat_named_t arcstat_anon_metadata;
 	/*
 	 * Number of bytes consumed by ARC buffers that meet the
 	 * following criteria: backing buffers of type ARC_BUFC_DATA,
@@ -676,6 +684,8 @@ typedef struct arc_stats {
 	 * are all included in this value.
 	 */
 	kstat_named_t arcstat_mru_size;
+	kstat_named_t arcstat_mru_data;
+	kstat_named_t arcstat_mru_metadata;
 	/*
 	 * Number of bytes consumed by ARC buffers that meet the
 	 * following criteria: backing buffers of type ARC_BUFC_DATA,
@@ -700,6 +710,8 @@ typedef struct arc_stats {
 	 * buffers *would have* consumed this number of bytes.
 	 */
 	kstat_named_t arcstat_mru_ghost_size;
+	kstat_named_t arcstat_mru_ghost_data;
+	kstat_named_t arcstat_mru_ghost_metadata;
 	/*
 	 * Number of bytes that *would have been* consumed by ARC
 	 * buffers that are eligible for eviction, of type
@@ -719,6 +731,8 @@ typedef struct arc_stats {
 	 * are all included in this value.
 	 */
 	kstat_named_t arcstat_mfu_size;
+	kstat_named_t arcstat_mfu_data;
+	kstat_named_t arcstat_mfu_metadata;
 	/*
 	 * Number of bytes consumed by ARC buffers that are eligible for
 	 * eviction, of type ARC_BUFC_DATA, and reside in the arc_mfu
@@ -737,6 +751,8 @@ typedef struct arc_stats {
 	 * arcstat_mru_ghost_size for more details.
 	 */
 	kstat_named_t arcstat_mfu_ghost_size;
+	kstat_named_t arcstat_mfu_ghost_data;
+	kstat_named_t arcstat_mfu_ghost_metadata;
 	/*
 	 * Number of bytes that *would have been* consumed by ARC
 	 * buffers that are eligible for eviction, of type
@@ -754,6 +770,8 @@ typedef struct arc_stats {
 	 * ARC_FLAG_UNCACHED being set.
 	 */
 	kstat_named_t arcstat_uncached_size;
+	kstat_named_t arcstat_uncached_data;
+	kstat_named_t arcstat_uncached_metadata;
 	/*
 	 * Number of data bytes that are going to be evicted from ARC due to
 	 * ARC_FLAG_UNCACHED being set.
@@ -876,10 +894,7 @@ typedef struct arc_stats {
 	kstat_named_t arcstat_loaned_bytes;
 	kstat_named_t arcstat_prune;
 	kstat_named_t arcstat_meta_used;
-	kstat_named_t arcstat_meta_limit;
 	kstat_named_t arcstat_dnode_limit;
-	kstat_named_t arcstat_meta_max;
-	kstat_named_t arcstat_meta_min;
 	kstat_named_t arcstat_async_upgrade_sync;
 	/* Number of predictive prefetch requests. */
 	kstat_named_t arcstat_predictive_prefetch;
@@ -942,7 +957,7 @@ typedef struct arc_sums {
 	wmsum_t arcstat_data_size;
 	wmsum_t arcstat_metadata_size;
 	wmsum_t arcstat_dbuf_size;
-	aggsum_t arcstat_dnode_size;
+	wmsum_t arcstat_dnode_size;
 	wmsum_t arcstat_bonus_size;
 	wmsum_t arcstat_l2_hits;
 	wmsum_t arcstat_l2_misses;
@@ -987,7 +1002,7 @@ typedef struct arc_sums {
 	wmsum_t arcstat_memory_direct_count;
 	wmsum_t arcstat_memory_indirect_count;
 	wmsum_t arcstat_prune;
-	aggsum_t arcstat_meta_used;
+	wmsum_t arcstat_meta_used;
 	wmsum_t arcstat_async_upgrade_sync;
 	wmsum_t arcstat_predictive_prefetch;
 	wmsum_t arcstat_demand_hit_predictive_prefetch;
@@ -1015,7 +1030,9 @@ typedef struct arc_evict_waiter {
 #define	ARCSTAT_BUMPDOWN(stat)	ARCSTAT_INCR(stat, -1)
 
 #define	arc_no_grow	ARCSTAT(arcstat_no_grow) /* do not grow cache size */
-#define	arc_p		ARCSTAT(arcstat_p)	/* target size of MRU */
+#define	arc_meta	ARCSTAT(arcstat_meta)	/* target frac of metadata */
+#define	arc_pd		ARCSTAT(arcstat_pd)	/* target frac of data MRU */
+#define	arc_pm		ARCSTAT(arcstat_pm)	/* target frac of meta MRU */
 #define	arc_c		ARCSTAT(arcstat_c)	/* target size of cache */
 #define	arc_c_min	ARCSTAT(arcstat_c_min)	/* min target cache size */
 #define	arc_c_max	ARCSTAT(arcstat_c_max)	/* max target cache size */