Commit 87c6a9b253520b66e7f5e8f67a37a701eaa51cee

Authored by Jens Axboe
1 parent 3542a5c0de

writeback: make balance_dirty_pages() gradually back more off

Currently it just sleeps for a very short time, just 1 jiffy. If
we keep looping in there, continually delay for a little longer
of up to 100msec in total. That was the old limit for congestion
wait.

Signed-off-by: Jens Axboe <jens.axboe@oracle.com>

Showing 1 changed file with 10 additions and 1 deletions Side-by-side Diff

... ... @@ -485,6 +485,7 @@
485 485 unsigned long bdi_thresh;
486 486 unsigned long pages_written = 0;
487 487 unsigned long write_chunk = sync_writeback_pages();
  488 + unsigned long pause = 1;
488 489  
489 490 struct backing_dev_info *bdi = mapping->backing_dev_info;
490 491  
... ... @@ -561,7 +562,15 @@
561 562 if (pages_written >= write_chunk)
562 563 break; /* We've done our duty */
563 564  
564   - schedule_timeout_interruptible(1);
  565 + schedule_timeout_interruptible(pause);
  566 +
  567 + /*
  568 + * Increase the delay for each loop, up to our previous
  569 + * default of taking a 100ms nap.
  570 + */
  571 + pause <<= 1;
  572 + if (pause > HZ / 10)
  573 + pause = HZ / 10;
565 574 }
566 575  
567 576 if (bdi_nr_reclaimable + bdi_nr_writeback < bdi_thresh &&