Several improvements to ARC shrinking

 - When receiving memory pressure signal from OS be more strict
trying to free some memory.  Otherwise kernel may come again and
request much more.  Return as result how much arc_c was actually
reduced due to this request, that may be less than requested.
 - On Linux when receiving direct reclaim from some file system
(that may be ZFS) instead of ignoring request completely, just
shrink the ARC, but do not wait for eviction.  Waiting there may
cause deadlock.  Ignoring it as before may put extra pressure on
other caches and/or swap, and cause OOM if nothing help.  While
not waiting may result in more ARC evicted later, and may be too
late if OOM killer activate right now, but I hope it to be better
than doing nothing at all.
 - On Linux set arc_no_grow before waiting for reclaim, not after,
or it may grow back while we are waiting.
 - On Linux add new parameter zfs_arc_shrinker_seeks to balance
ARC eviction cost, relative to page cache and other subsystems.
 - Slightly update Linux arc_set_sys_free() math for new kernels.

Signed-off-by:	Alexander Motin <[email protected]>
Sponsored by:	iXsystems, Inc.

include/sys/arc_impl.h

@@ -1058,10 +1058,10 @@ extern uint_t arc_lotsfree_percent;
 extern uint64_t zfs_arc_min;
 extern uint64_t zfs_arc_max;
 
-extern void arc_reduce_target_size(int64_t to_free);
+extern uint64_t arc_reduce_target_size(uint64_t to_free);
 extern boolean_t arc_reclaim_needed(void);
 extern void arc_kmem_reap_soon(void);
-extern void arc_wait_for_eviction(uint64_t, boolean_t);
+extern void arc_wait_for_eviction(uint64_t, boolean_t, boolean_t);
 
 extern void arc_lowmem_init(void);
 extern void arc_lowmem_fini(void);

man/man4/zfs.4

@@ -831,6 +831,13 @@ even with a small average compressed block size of ~8 KiB.
 The parameter can be set to 0 (zero) to disable the limit,
 and only applies on Linux.
 .
+.It Sy zfs_arc_shrinker_seeks Ns = Ns Sy 2 Pq int
+Relative cost of ARC eviction on Linux, AKA number of seeks needed to
+restore evicted page.
+Bigger values make ARC more precious and evictions smaller, comparing to
+other kernel subsystems.
+Value of 4 means parity with page cache.
+.
 .It Sy zfs_arc_sys_free Ns = Ns Sy 0 Ns B Pq u64
 The target number of bytes the ARC should leave as free memory on the system.
 If zero, equivalent to the bigger of

module/os/freebsd/zfs/arc_os.c

@@ -149,26 +149,25 @@ static eventhandler_tag arc_event_lowmem = NULL;
 static void
 arc_lowmem(void *arg __unused, int howto __unused)
 {
-	int64_t free_memory, to_free;
+	int64_t can_free, free_memory, to_free;
 
 	arc_no_grow = B_TRUE;
 	arc_warm = B_TRUE;
 	arc_growtime = gethrtime() + SEC2NSEC(arc_grow_retry);
+
 	free_memory = arc_available_memory();
-	int64_t can_free = arc_c - arc_c_min;
-	if (can_free <= 0)
-		return;
-	to_free = (can_free >> arc_shrink_shift) - MIN(free_memory, 0);
+	can_free = arc_c - arc_c_min;
+	to_free = (MAX(can_free, 0) >> arc_shrink_shift) - MIN(free_memory, 0);
 	DTRACE_PROBE2(arc__needfree, int64_t, free_memory, int64_t, to_free);
-	arc_reduce_target_size(to_free);
+	to_free = arc_reduce_target_size(to_free);
 
 	/*
 	 * It is unsafe to block here in arbitrary threads, because we can come
 	 * here from ARC itself and may hold ARC locks and thus risk a deadlock
 	 * with ARC reclaim thread.
 	 */
 	if (curproc == pageproc)
-		arc_wait_for_eviction(to_free, B_FALSE);
+		arc_wait_for_eviction(to_free, B_FALSE, B_FALSE);
 }
 
 void

module/os/linux/zfs/arc_os.c

@@ -49,6 +49,7 @@
 #include <linux/page_compat.h>
 #include <linux/notifier.h>
 #include <linux/memory.h>
+#include <linux/version.h>
 #endif
 #include <sys/callb.h>
 #include <sys/kstat.h>
@@ -58,6 +59,7 @@
 #include <sys/trace_zfs.h>
 #include <sys/aggsum.h>
 
+#ifdef _KERNEL
 /*
  * This is a limit on how many pages the ARC shrinker makes available for
  * eviction in response to one page allocation attempt.  Note that in
@@ -72,11 +74,20 @@
  * See also the comment in arc_shrinker_count().
  * Set to 0 to disable limit.
  */
-int zfs_arc_shrinker_limit = 10000;
+static int zfs_arc_shrinker_limit = 10000;
+
+/*
+ * Relative cost of ARC eviction, AKA number of seeks needed to restore evicted
+ * page.  Bigger values make ARC more precious and evictions smaller comparing
+ * to other kernel subsystems.  Value of 4 means parity with page cache,
+ * according to my reading of kernel's do_shrink_slab() and other code.
+ */
+static int zfs_arc_shrinker_seeks = DEFAULT_SEEKS;
 
 #ifdef CONFIG_MEMORY_HOTPLUG
 static struct notifier_block arc_hotplug_callback_mem_nb;
 #endif
+#endif
 
 /*
  * Return a default max arc size based on the amount of physical memory.
@@ -170,22 +181,7 @@ static unsigned long
 arc_shrinker_count(struct shrinker *shrink, struct shrink_control *sc)
 {
 	/*
-	 * __GFP_FS won't be set if we are called from ZFS code (see
-	 * kmem_flags_convert(), which removes it).  To avoid a deadlock, we
-	 * don't allow evicting in this case.  We return 0 rather than
-	 * SHRINK_STOP so that the shrinker logic doesn't accumulate a
-	 * deficit against us.
-	 */
-	if (!(sc->gfp_mask & __GFP_FS)) {
-		return (0);
-	}
-
-	/*
-	 * This code is reached in the "direct reclaim" case, where the
-	 * kernel (outside ZFS) is trying to allocate a page, and the system
-	 * is low on memory.
-	 *
-	 * The kernel's shrinker code doesn't understand how many pages the
+	 * The kernel's shrinker code may not understand how many pages the
 	 * ARC's callback actually frees, so it may ask the ARC to shrink a
 	 * lot for one page allocation. This is problematic because it may
 	 * take a long time, thus delaying the page allocation, and because
@@ -204,40 +200,44 @@ arc_shrinker_count(struct shrinker *shrink, struct shrink_control *sc)
 	 *
 	 * See also the comment above zfs_arc_shrinker_limit.
 	 */
+	int64_t can_free = btop(arc_evictable_memory());
 	int64_t limit = zfs_arc_shrinker_limit != 0 ?
 	    zfs_arc_shrinker_limit : INT64_MAX;
-	return (MIN(limit, btop((int64_t)arc_evictable_memory())));
+	return (MIN(can_free, limit));
 }
 
 static unsigned long
 arc_shrinker_scan(struct shrinker *shrink, struct shrink_control *sc)
 {
-	ASSERT((sc->gfp_mask & __GFP_FS) != 0);
-
 	/* The arc is considered warm once reclaim has occurred */
 	if (unlikely(arc_warm == B_FALSE))
 		arc_warm = B_TRUE;
 
+	/*
+	 * We are experiencing memory pressure which the arc_evict_zthr was
+	 * unable to keep up with.  Set arc_no_grow to briefly pause ARC
+	 * growth to avoid compounding the memory pressure.
+	 */
+	arc_no_grow = B_TRUE;
+
 	/*
 	 * Evict the requested number of pages by reducing arc_c and waiting
-	 * for the requested amount of data to be evicted.
+	 * for the requested amount of data to be evicted.  To avoid deadlock
+	 * do not wait for eviction if we may be called from ZFS itself (see
+	 * kmem_flags_convert() removing __GFP_FS).  It may cause excessive
+	 * eviction later if many evictions are accumulated, but just skipping
+	 * the eviction is not good either if most of memory is used by ARC.
 	 */
-	arc_reduce_target_size(ptob(sc->nr_to_scan));
-	arc_wait_for_eviction(ptob(sc->nr_to_scan), B_FALSE);
+	uint64_t to_free = arc_reduce_target_size(ptob(sc->nr_to_scan));
+	if (sc->gfp_mask & __GFP_FS)
+		arc_wait_for_eviction(to_free, B_FALSE, B_FALSE);
 	if (current->reclaim_state != NULL)
 #ifdef	HAVE_RECLAIM_STATE_RECLAIMED
-		current->reclaim_state->reclaimed += sc->nr_to_scan;
+		current->reclaim_state->reclaimed += btop(to_free);
 #else
-		current->reclaim_state->reclaimed_slab += sc->nr_to_scan;
+		current->reclaim_state->reclaimed_slab += btop(to_free);
 #endif
 
-	/*
-	 * We are experiencing memory pressure which the arc_evict_zthr was
-	 * unable to keep up with. Set arc_no_grow to briefly pause arc
-	 * growth to avoid compounding the memory pressure.
-	 */
-	arc_no_grow = B_TRUE;
-
 	/*
 	 * When direct reclaim is observed it usually indicates a rapid
 	 * increase in memory pressure.  This occurs because the kswapd
@@ -250,7 +250,7 @@ arc_shrinker_scan(struct shrinker *shrink, struct shrink_control *sc)
 		ARCSTAT_BUMP(arcstat_memory_direct_count);
 	}
 
-	return (sc->nr_to_scan);
+	return (btop(to_free));
 }
 
 static struct shrinker *arc_shrinker = NULL;
@@ -304,9 +304,7 @@ arc_set_sys_free(uint64_t allmem)
 	 * arc_wait_for_eviction() will wait until at least the
 	 * high_wmark_pages() are free (see arc_evict_state_impl()).
 	 *
-	 * Note: Even when the system is very low on memory, the kernel's
-	 * shrinker code may only ask for one "batch" of pages (512KB) to be
-	 * evicted.  If concurrent allocations consume these pages, there may
+	 * Note: If concurrent allocations consume these pages, there may
 	 * still be insufficient free pages, and the OOM killer takes action.
 	 *
 	 * By setting arc_sys_free large enough, and having
@@ -318,20 +316,26 @@ arc_set_sys_free(uint64_t allmem)
 	 * It's hard to iterate the zones from a linux kernel module, which
 	 * makes it difficult to determine the watermark dynamically. Instead
 	 * we compute the maximum high watermark for this system, based
-	 * on the amount of memory, assuming default parameters on Linux kernel
-	 * 5.3.
+	 * on the amount of memory, using the same method as the kernel uses
+	 * to calculate its internal `min_free_kbytes` variable.  See
+	 * torvalds/linux@ee8eb9a5fe86 for the change in the upper clamp value
+	 * from 64M to 256M.
 	 */
 
 	/*
 	 * Base wmark_low is 4 * the square root of Kbytes of RAM.
 	 */
-	long wmark = 4 * int_sqrt(allmem/1024) * 1024;
+	long wmark = int_sqrt(allmem / 1024 * 16) * 1024;
 
 	/*
-	 * Clamp to between 128K and 64MB.
+	 * Clamp to between 128K and 256/64MB.
 	 */
 	wmark = MAX(wmark, 128 * 1024);
+#if LINUX_VERSION_CODE >= KERNEL_VERSION(5, 7, 0)
+	wmark = MIN(wmark, 256 * 1024 * 1024);
+#else
 	wmark = MIN(wmark, 64 * 1024 * 1024);
+#endif
 
 	/*
 	 * watermark_boost can increase the wmark by up to 150%.
@@ -357,7 +361,7 @@ arc_lowmem_init(void)
 	 * swapping out pages when it is preferable to shrink the arc.
 	 */
 	arc_shrinker = spl_register_shrinker("zfs-arc-shrinker",
-	    arc_shrinker_count, arc_shrinker_scan, DEFAULT_SEEKS);
+	    arc_shrinker_count, arc_shrinker_scan, zfs_arc_shrinker_seeks);
 	VERIFY(arc_shrinker);
 
 	arc_set_sys_free(allmem);
@@ -500,3 +504,5 @@ arc_unregister_hotplug(void)
 
 ZFS_MODULE_PARAM(zfs_arc, zfs_arc_, shrinker_limit, INT, ZMOD_RW,
 	"Limit on number of pages that ARC shrinker can reclaim at once");
+ZFS_MODULE_PARAM(zfs_arc, zfs_arc_, shrinker_seeks, INT, ZMOD_RD,
+	"Relative cost of ARC eviction vs other kernel subsystems");