writeback: fix bandwidth estimate for spiky workload

Michael Stapelberg has reported that for workload with short big spikes of
writes (GCC linker seem to trigger this frequently) the write throughput
is heavily underestimated and tends to steadily sink until it reaches
zero.  This has rather bad impact on writeback throttling (causing
stalls).  The problem is that writeback throughput estimate gets updated
at most once per 200 ms.  One update happens early after we submit pages
for writeback (at that point writeout of only small fraction of pages is
completed and thus observed throughput is tiny).  Next update happens only
during the next write spike (updates happen only from inode writeback and
dirty throttling code) and if that is more than 1s after previous spike,
we decide system was idle and just ignore whatever was written until this
moment.

Fix the problem by making sure writeback throughput estimate is also
updated shortly after writeback completes to get reasonable estimate of
throughput for spiky workloads.

[jack@suse.cz: avoid division by 0 in wb_update_dirty_ratelimit()]

Link: https://lore.kernel.org/lkml/20210617095309.3542373-1-stapelberg+linux@google.com
Link: https://lkml.kernel.org/r/20210713104716.22868-3-jack@suse.cz
Signed-off-by: Jan Kara <jack@suse.cz>
Reported-by: Michael Stapelberg <stapelberg+linux@google.com>
Tested-by: Michael Stapelberg <stapelberg+linux@google.com>
Cc: Wu Fengguang <fengguang.wu@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

1 files changed, 25 insertions, 14 deletions

diff --git a/mm/page-writeback.c b/mm/page-writeback.c
index e4a381b8944b..156f5888c09d 100644
--- a/mm/page-writeback.c
+++ b/mm/page-writeback.c
@@ -1336,18 +1336,19 @@ static void __wb_update_bandwidth(struct dirty_throttle_control *gdtc,
 {
 	struct bdi_writeback *wb = gdtc->wb;
 	unsigned long now = jiffies;
-	unsigned long elapsed = now - wb->bw_time_stamp;
+	unsigned long elapsed;
 	unsigned long dirtied;
 	unsigned long written;
 
-	lockdep_assert_held(&wb->list_lock);
+	spin_lock(&wb->list_lock);
 
 	/*
-	 * rate-limit, only update once every 200ms.
+	 * Lockless checks for elapsed time are racy and delayed update after
+	 * IO completion doesn't do it at all (to make sure written pages are
+	 * accounted reasonably quickly). Make sure elapsed >= 1 to avoid
+	 * division errors.
 	 */
-	if (elapsed < BANDWIDTH_INTERVAL)
-		return;
-
+	elapsed = max(now - wb->bw_time_stamp, 1UL);
 	dirtied = percpu_counter_read(&wb->stat[WB_DIRTIED]);
 	written = percpu_counter_read(&wb->stat[WB_WRITTEN]);
 
@@ -1369,15 +1370,14 @@ static void __wb_update_bandwidth(struct dirty_throttle_control *gdtc,
 	wb->dirtied_stamp = dirtied;
 	wb->written_stamp = written;
 	wb->bw_time_stamp = now;
+	spin_unlock(&wb->list_lock);
 }
 
-static void wb_update_bandwidth(struct bdi_writeback *wb)
+void wb_update_bandwidth(struct bdi_writeback *wb)
 {
 	struct dirty_throttle_control gdtc = { GDTC_INIT(wb) };
 
-	spin_lock(&wb->list_lock);
 	__wb_update_bandwidth(&gdtc, NULL, false);
-	spin_unlock(&wb->list_lock);
 }
 
 /* Interval after which we consider wb idle and don't estimate bandwidth */
@@ -1722,11 +1722,8 @@ free_running:
 			wb->dirty_exceeded = 1;
 
 		if (time_is_before_jiffies(wb->bw_time_stamp +
-					   BANDWIDTH_INTERVAL)) {
-			spin_lock(&wb->list_lock);
+					   BANDWIDTH_INTERVAL))
 			__wb_update_bandwidth(gdtc, mdtc, true);
-			spin_unlock(&wb->list_lock);
-		}
 
 		/* throttle according to the chosen dtc */
 		dirty_ratelimit = wb->dirty_ratelimit;
@@ -2374,7 +2371,13 @@ int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
 		cond_resched();
 		congestion_wait(BLK_RW_ASYNC, HZ/50);
 	}
-	wb_update_bandwidth(wb);
+	/*
+	 * Usually few pages are written by now from those we've just submitted
+	 * but if there's constant writeback being submitted, this makes sure
+	 * writeback bandwidth is updated once in a while.
+	 */
+	if (time_is_before_jiffies(wb->bw_time_stamp + BANDWIDTH_INTERVAL))
+		wb_update_bandwidth(wb);
 	return ret;
 }
 
@@ -2754,6 +2757,14 @@ static void wb_inode_writeback_start(struct bdi_writeback *wb)
 static void wb_inode_writeback_end(struct bdi_writeback *wb)
 {
 	atomic_dec(&wb->writeback_inodes);
+	/*
+	 * Make sure estimate of writeback throughput gets updated after
+	 * writeback completed. We delay the update by BANDWIDTH_INTERVAL
+	 * (which is the interval other bandwidth updates use for batching) so
+	 * that if multiple inodes end writeback at a similar time, they get
+	 * batched into one bandwidth update.
+	 */
+	queue_delayed_work(bdi_wq, &wb->bw_dwork, BANDWIDTH_INTERVAL);
 }
 
 int test_clear_page_writeback(struct page *page)