writeback: avoid livelocking WB_SYNC_ALL writeback
authorJan Kara <jack@suse.cz>
Thu, 13 Jan 2011 23:45:48 +0000 (15:45 -0800)
committerLinus Torvalds <torvalds@linux-foundation.org>
Fri, 14 Jan 2011 01:32:32 +0000 (17:32 -0800)
When wb_writeback() is called in WB_SYNC_ALL mode, work->nr_to_write is
usually set to LONG_MAX.  The logic in wb_writeback() then calls
__writeback_inodes_sb() with nr_to_write == MAX_WRITEBACK_PAGES and we
easily end up with non-positive nr_to_write after the function returns, if
the inode has more than MAX_WRITEBACK_PAGES dirty pages at the moment.

When nr_to_write is <= 0 wb_writeback() decides we need another round of
writeback but this is wrong in some cases!  For example when a single
large file is continuously dirtied, we would never finish syncing it
because each pass would be able to write MAX_WRITEBACK_PAGES and inode
dirty timestamp never gets updated (as inode is never completely clean).
Thus __writeback_inodes_sb() would write the redirtied inode again and
again.

Fix the issue by setting nr_to_write to LONG_MAX in WB_SYNC_ALL mode.  We
do not need nr_to_write in WB_SYNC_ALL mode anyway since
write_cache_pages() does livelock avoidance using page tagging in
WB_SYNC_ALL mode.

This makes wb_writeback() call __writeback_inodes_sb() only once on
WB_SYNC_ALL.  The latter function won't livelock because it works on

- a finite set of files by doing queue_io() once at the beginning
- a finite set of pages by PAGECACHE_TAG_TOWRITE page tagging

After this patch, program from http://lkml.org/lkml/2010/10/24/154 is no
longer able to stall sync forever.

[fengguang.wu@intel.com: fix locking comment]
Signed-off-by: Jan Kara <jack@suse.cz>
Signed-off-by: Wu Fengguang <fengguang.wu@intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Christoph Hellwig <hch@lst.de>
Cc: Jan Engelhardt <jengelh@medozas.de>
Cc: Jens Axboe <axboe@kernel.dk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
fs/fs-writeback.c

index 9e72d04e706e06dfd0885432cec6dc20e15374c8..e8063c938dd23ec1a7bd10db1b7143f11a96819b 100644 (file)
@@ -630,6 +630,7 @@ static long wb_writeback(struct bdi_writeback *wb,
        };
        unsigned long oldest_jif;
        long wrote = 0;
+       long write_chunk;
        struct inode *inode;
 
        if (wbc.for_kupdate) {
@@ -642,6 +643,24 @@ static long wb_writeback(struct bdi_writeback *wb,
                wbc.range_end = LLONG_MAX;
        }
 
+       /*
+        * WB_SYNC_ALL mode does livelock avoidance by syncing dirty
+        * inodes/pages in one big loop. Setting wbc.nr_to_write=LONG_MAX
+        * here avoids calling into writeback_inodes_wb() more than once.
+        *
+        * The intended call sequence for WB_SYNC_ALL writeback is:
+        *
+        *      wb_writeback()
+        *          __writeback_inodes_sb()     <== called only once
+        *              write_cache_pages()     <== called once for each inode
+        *                   (quickly) tag currently dirty pages
+        *                   (maybe slowly) sync all tagged pages
+        */
+       if (wbc.sync_mode == WB_SYNC_NONE)
+               write_chunk = MAX_WRITEBACK_PAGES;
+       else
+               write_chunk = LONG_MAX;
+
        wbc.wb_start = jiffies; /* livelock avoidance */
        for (;;) {
                /*
@@ -668,7 +687,7 @@ static long wb_writeback(struct bdi_writeback *wb,
                        break;
 
                wbc.more_io = 0;
-               wbc.nr_to_write = MAX_WRITEBACK_PAGES;
+               wbc.nr_to_write = write_chunk;
                wbc.pages_skipped = 0;
 
                trace_wbc_writeback_start(&wbc, wb->bdi);
@@ -678,8 +697,8 @@ static long wb_writeback(struct bdi_writeback *wb,
                        writeback_inodes_wb(wb, &wbc);
                trace_wbc_writeback_written(&wbc, wb->bdi);
 
-               work->nr_pages -= MAX_WRITEBACK_PAGES - wbc.nr_to_write;
-               wrote += MAX_WRITEBACK_PAGES - wbc.nr_to_write;
+               work->nr_pages -= write_chunk - wbc.nr_to_write;
+               wrote += write_chunk - wbc.nr_to_write;
 
                /*
                 * If we consumed everything, see if we have more
@@ -694,7 +713,7 @@ static long wb_writeback(struct bdi_writeback *wb,
                /*
                 * Did we write something? Try for more
                 */
-               if (wbc.nr_to_write < MAX_WRITEBACK_PAGES)
+               if (wbc.nr_to_write < write_chunk)
                        continue;
                /*
                 * Nothing written. Wait for some inode to