[PATCH RFC] mm/readahead: improve randread performance with readahead disabled

[Yu Kuai <yukuai1@huaweicloud.com>]

From: Yu Kuai <yukuai3@huawei.com>

We have a workload of random 4k-128k read on a HDD, from iostat we observed
that average request size is 256k+ and bandwidth is 100MB+, this is because
readahead waste lots of disk bandwidth. Hence we disable readahead and
performance from user side is indeed much better(2x+), however, from
iostat we observed request size is just 4k and bandwidth is just around
40MB.

Then we do a simple dd test and found out if readahead is disabled,
page_cache_sync_ra() will force to read one page at a time, and this
really doesn't make sense because we can just issue user requested size
request to disk.

Fix this problem by removing the limit to read one page at a time from
page_cache_sync_ra(), this way the random read workload can get better
performance with readahead disabled.

PS: I'm not sure if I miss anything, so this version is RFC
Signed-off-by: Yu Kuai <yukuai3@huawei.com>
---
 mm/readahead.c | 12 ++++++++++--
 1 file changed, 10 insertions(+), 2 deletions(-)

diff --git a/mm/readahead.c b/mm/readahead.c
index 20d36d6b055e..1df85ccba575 100644
--- a/mm/readahead.c
+++ b/mm/readahead.c
@@ -561,13 +561,21 @@ void page_cache_sync_ra(struct readahead_control *ractl,
 	 * Even if readahead is disabled, issue this request as readahead
 	 * as we'll need it to satisfy the requested range. The forced
 	 * readahead will do the right thing and limit the read to just the
-	 * requested range, which we'll set to 1 page for this case.
+	 * requested range.
 	 */
-	if (!ra->ra_pages || blk_cgroup_congested()) {
+	if (blk_cgroup_congested()) {
 		if (!ractl->file)
 			return;
+		/*
+		 * If the cgroup is congested, ensure to do at least 1 page of
+		 * readahead to make progress on the read.
+		 */
 		req_count = 1;
 		do_forced_ra = true;
+	} else if (!ra->ra_pages) {
+		if (!ractl->file)
+			return;
+		do_forced_ra = true;
 	}
 
 	/* be dumb */
-- 
2.39.2