mm/filemap: optimize filemap folio adding

Instead of doing multiple tree walks, do one optimism range check with
lock hold, and exit if raced with another insertion.  If a shadow exists,
check it with a new xas_get_order helper before releasing the lock to
avoid redundant tree walks for getting its order.

Drop the lock and do the allocation only if a split is needed.

In the best case, it only need to walk the tree once.  If it needs to
alloc and split, 3 walks are issued (One for first ranged conflict check
and order retrieving, one for the second check after allocation, one for
the insert after split).

Testing with 4K pages, in an 8G cgroup, with 16G brd as block device:

  echo 3 > /proc/sys/vm/drop_caches

  fio -name=cached --numjobs=16 --filename=/mnt/test.img \
    --buffered=1 --ioengine=mmap --rw=randread --time_based \
    --ramp_time=30s --runtime=5m --group_reporting

Before:
bw (  MiB/s): min= 1027, max= 3520, per=100.00%, avg=2445.02, stdev=18.90, samples=8691
iops        : min=263001, max=901288, avg=625924.36, stdev=4837.28, samples=8691

After (+7.3%):
bw (  MiB/s): min=  493, max= 3947, per=100.00%, avg=2625.56, stdev=25.74, samples=8651
iops        : min=126454, max=1010681, avg=672142.61, stdev=6590.48, samples=8651

Test result with THP (do a THP randread then switch to 4K page in hope it
issues a lot of splitting):

  echo 3 > /proc/sys/vm/drop_caches

  fio -name=cached --numjobs=16 --filename=/mnt/test.img \
      --buffered=1 --ioengine=mmap -thp=1 --readonly \
      --rw=randread --time_based --ramp_time=30s --runtime=10m \
      --group_reporting

  fio -name=cached --numjobs=16 --filename=/mnt/test.img \
      --buffered=1 --ioengine=mmap \
      --rw=randread --time_based --runtime=5s --group_reporting

Before:
bw (  KiB/s): min= 4141, max=14202, per=100.00%, avg=7935.51, stdev=96.85, samples=18976
iops        : min= 1029, max= 3548, avg=1979.52, stdev=24.23, samples=18976·

READ: bw=4545B/s (4545B/s), 4545B/s-4545B/s (4545B/s-4545B/s), io=64.0KiB (65.5kB), run=14419-14419msec

After (+12.5%):
bw (  KiB/s): min= 4611, max=15370, per=100.00%, avg=8928.74, stdev=105.17, samples=19146
iops        : min= 1151, max= 3842, avg=2231.27, stdev=26.29, samples=19146

READ: bw=4635B/s (4635B/s), 4635B/s-4635B/s (4635B/s-4635B/s), io=64.0KiB (65.5kB), run=14137-14137msec

The performance is better for both 4K (+7.5%) and THP (+12.5%) cached read.

Link: https://lkml.kernel.org/r/20240415171857.19244-5-ryncsn@gmail.com
Signed-off-by: Kairui Song <kasong@tencent.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>

1 files changed, 41 insertions, 15 deletions

diff --git a/mm/filemap.c b/mm/filemap.c
index 17a66ea544e7..7b0b2229d4ed 100644
--- a/mm/filemap.c
+++ b/mm/filemap.c
@@ -852,7 +852,9 @@ noinline int __filemap_add_folio(struct address_space *mapping,
 		struct folio *folio, pgoff_t index, gfp_t gfp, void **shadowp)
 {
 	XA_STATE(xas, &mapping->i_pages, index);
-	bool huge = folio_test_hugetlb(folio);
+	void *alloced_shadow = NULL;
+	int alloced_order = 0;
+	bool huge;
 	long nr;
 
 	VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
@@ -861,6 +863,7 @@ noinline int __filemap_add_folio(struct address_space *mapping,
 
 	VM_BUG_ON_FOLIO(index & (folio_nr_pages(folio) - 1), folio);
 	xas_set_order(&xas, index, folio_order(folio));
+	huge = folio_test_hugetlb(folio);
 	nr = folio_nr_pages(folio);
 
 	gfp &= GFP_RECLAIM_MASK;
@@ -868,16 +871,10 @@ noinline int __filemap_add_folio(struct address_space *mapping,
 	folio->mapping = mapping;
 	folio->index = xas.xa_index;
 
-	do {
-		unsigned int order = xa_get_order(xas.xa, xas.xa_index);
+	for (;;) {
+		int order = -1, split_order = 0;
 		void *entry, *old = NULL;
 
-		if (order > folio_order(folio)) {
-			xas_split_alloc(&xas, xa_load(xas.xa, xas.xa_index),
-					order, gfp);
-			if (xas_error(&xas))
-				goto error;
-		}
 		xas_lock_irq(&xas);
 		xas_for_each_conflict(&xas, entry) {
 			old = entry;
@@ -885,19 +882,33 @@ noinline int __filemap_add_folio(struct address_space *mapping,
 				xas_set_err(&xas, -EEXIST);
 				goto unlock;
 			}
+			/*
+			 * If a larger entry exists,
+			 * it will be the first and only entry iterated.
+			 */
+			if (order == -1)
+				order = xas_get_order(&xas);
+		}
+
+		/* entry may have changed before we re-acquire the lock */
+		if (alloced_order && (old != alloced_shadow || order != alloced_order)) {
+			xas_destroy(&xas);
+			alloced_order = 0;
 		}
 
 		if (old) {
-			if (shadowp)
-				*shadowp = old;
-			/* entry may have been split before we acquired lock */
-			order = xa_get_order(xas.xa, xas.xa_index);
-			if (order > folio_order(folio)) {
+			if (order > 0 && order > folio_order(folio)) {
 				/* How to handle large swap entries? */
 				BUG_ON(shmem_mapping(mapping));
+				if (!alloced_order) {
+					split_order = order;
+					goto unlock;
+				}
 				xas_split(&xas, old, order);
 				xas_reset(&xas);
 			}
+			if (shadowp)
+				*shadowp = old;
 		}
 
 		xas_store(&xas, folio);
@@ -913,9 +924,24 @@ noinline int __filemap_add_folio(struct address_space *mapping,
 				__lruvec_stat_mod_folio(folio,
 						NR_FILE_THPS, nr);
 		}
+
 unlock:
 		xas_unlock_irq(&xas);
-	} while (xas_nomem(&xas, gfp));
+
+		/* split needed, alloc here and retry. */
+		if (split_order) {
+			xas_split_alloc(&xas, old, split_order, gfp);
+			if (xas_error(&xas))
+				goto error;
+			alloced_shadow = old;
+			alloced_order = split_order;
+			xas_reset(&xas);
+			continue;
+		}
+
+		if (!xas_nomem(&xas, gfp))
+			break;
+	}
 
 	if (xas_error(&xas))
 		goto error;