Btrfs: fix hang on error (such as ENOSPC) when writing extent pages
authorFilipe Manana <fdmanana@gmail.com>
Fri, 9 May 2014 16:17:40 +0000 (17:17 +0100)
committerChris Mason <clm@fb.com>
Tue, 10 Jun 2014 00:20:20 +0000 (17:20 -0700)
When running low on available disk space and having several processes
doing buffered file IO, I got the following trace in dmesg:

[ 4202.720152] INFO: task kworker/u8:1:5450 blocked for more than 120 seconds.
[ 4202.720401]       Not tainted 3.13.0-fdm-btrfs-next-26+ #1
[ 4202.720596] "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
[ 4202.720874] kworker/u8:1    D 0000000000000001     0  5450      2 0x00000000
[ 4202.720904] Workqueue: btrfs-flush_delalloc normal_work_helper [btrfs]
[ 4202.720908]  ffff8801f62ddc38 0000000000000082 ffff880203ac2490 00000000001d3f40
[ 4202.720913]  ffff8801f62ddfd8 00000000001d3f40 ffff8800c4f0c920 ffff880203ac2490
[ 4202.720918]  00000000001d4a40 ffff88020fe85a40 ffff88020fe85ab8 0000000000000001
[ 4202.720922] Call Trace:
[ 4202.720931]  [<ffffffff816a3cb9>] schedule+0x29/0x70
[ 4202.720950]  [<ffffffffa01ec48d>] btrfs_start_ordered_extent+0x6d/0x110 [btrfs]
[ 4202.720956]  [<ffffffff8108e620>] ? bit_waitqueue+0xc0/0xc0
[ 4202.720972]  [<ffffffffa01ec559>] btrfs_run_ordered_extent_work+0x29/0x40 [btrfs]
[ 4202.720988]  [<ffffffffa0201987>] normal_work_helper+0x137/0x2c0 [btrfs]
[ 4202.720994]  [<ffffffff810680e5>] process_one_work+0x1f5/0x530
(...)
[ 4202.721027] 2 locks held by kworker/u8:1/5450:
[ 4202.721028]  #0:  (%s-%s){++++..}, at: [<ffffffff81068083>] process_one_work+0x193/0x530
[ 4202.721037]  #1:  ((&work->normal_work)){+.+...}, at: [<ffffffff81068083>] process_one_work+0x193/0x530
[ 4202.721054] INFO: task btrfs:7891 blocked for more than 120 seconds.
[ 4202.721258]       Not tainted 3.13.0-fdm-btrfs-next-26+ #1
[ 4202.721444] "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
[ 4202.721699] btrfs           D 0000000000000001     0  7891   7890 0x00000001
[ 4202.721704]  ffff88018c2119e8 0000000000000086 ffff8800a33d2490 00000000001d3f40
[ 4202.721710]  ffff88018c211fd8 00000000001d3f40 ffff8802144b0000 ffff8800a33d2490
[ 4202.721714]  ffff8800d8576640 ffff88020fe85bc0 ffff88020fe85bc8 7fffffffffffffff
[ 4202.721718] Call Trace:
[ 4202.721723]  [<ffffffff816a3cb9>] schedule+0x29/0x70
[ 4202.721727]  [<ffffffff816a2ebc>] schedule_timeout+0x1dc/0x270
[ 4202.721732]  [<ffffffff8109bd79>] ? mark_held_locks+0xb9/0x140
[ 4202.721736]  [<ffffffff816a90c0>] ? _raw_spin_unlock_irq+0x30/0x40
[ 4202.721740]  [<ffffffff8109bf0d>] ? trace_hardirqs_on_caller+0x10d/0x1d0
[ 4202.721744]  [<ffffffff816a488f>] wait_for_completion+0xdf/0x120
[ 4202.721749]  [<ffffffff8107fa90>] ? try_to_wake_up+0x310/0x310
[ 4202.721765]  [<ffffffffa01ebee4>] btrfs_wait_ordered_extents+0x1f4/0x280 [btrfs]
[ 4202.721781]  [<ffffffffa020526e>] btrfs_mksubvol.isra.62+0x30e/0x5a0 [btrfs]
[ 4202.721786]  [<ffffffff8108e620>] ? bit_waitqueue+0xc0/0xc0
[ 4202.721799]  [<ffffffffa02056a9>] btrfs_ioctl_snap_create_transid+0x1a9/0x1b0 [btrfs]
[ 4202.721813]  [<ffffffffa020583a>] btrfs_ioctl_snap_create_v2+0x10a/0x170 [btrfs]
(...)

It turns out that extent_io.c:__extent_writepage(), which ends up being called
through filemap_fdatawrite_range() in btrfs_start_ordered_extent(), was getting
-ENOSPC when calling the fill_delalloc callback. In this situation, it returned
without the writepage_end_io_hook callback (inode.c:btrfs_writepage_end_io_hook)
ever being called for the respective page, which prevents the ordered extent's
bytes_left count from ever reaching 0, and therefore a finish_ordered_fn work
is never queued into the endio_write_workers queue. This makes the task that
called btrfs_start_ordered_extent() hang forever on the wait queue of the ordered
extent.

This is fairly easy to reproduce using a small filesystem and fsstress on
a quad core vm:

    mkfs.btrfs -f -b `expr 2100 \* 1024 \* 1024` /dev/sdd
    mount /dev/sdd /mnt

    fsstress -p 6 -d /mnt -n 100000 -x \
        "btrfs subvolume snapshot -r /mnt /mnt/mysnap" \
    -f allocsp=0 \
    -f bulkstat=0 \
    -f bulkstat1=0 \
    -f chown=0 \
    -f creat=1 \
    -f dread=0 \
    -f dwrite=0 \
    -f fallocate=1 \
    -f fdatasync=0 \
    -f fiemap=0 \
    -f freesp=0 \
    -f fsync=0 \
    -f getattr=0 \
    -f getdents=0 \
    -f link=0 \
    -f mkdir=0 \
    -f mknod=0 \
    -f punch=1 \
    -f read=0 \
    -f readlink=0 \
    -f rename=0 \
    -f resvsp=0 \
    -f rmdir=0 \
    -f setxattr=0 \
    -f stat=0 \
    -f symlink=0 \
    -f sync=0 \
    -f truncate=1 \
    -f unlink=0 \
    -f unresvsp=0 \
    -f write=4

So just ensure that if an error happens while writing the extent page
we call the writepage_end_io_hook callback. Also make it return the
error code and ensure the caller (extent_write_cache_pages) processes
all pages in the page vector even if an error happens only for some
of them, so that ordered extents end up released.

Signed-off-by: Filipe David Borba Manana <fdmanana@gmail.com>
Signed-off-by: Chris Mason <clm@fb.com>
fs/btrfs/extent_io.c

index 3955e475ceece295ea9db55def2178a319a817cc..fa31c8d2c095589b379b32f77d469f24366a1763 100644 (file)
@@ -3278,6 +3278,7 @@ static int __extent_writepage(struct page *page, struct writeback_control *wbc,
                                     end - cur + 1, 1);
                if (IS_ERR_OR_NULL(em)) {
                        SetPageError(page);
+                       ret = PTR_ERR_OR_ZERO(em);
                        break;
                }
 
@@ -3364,13 +3365,17 @@ done:
                set_page_writeback(page);
                end_page_writeback(page);
        }
+       if (PageError(page)) {
+               ret = ret < 0 ? ret : -EIO;
+               end_extent_writepage(page, ret, start, page_end);
+       }
        unlock_page(page);
 
 done_unlocked:
 
        /* drop our reference on any cached states */
        free_extent_state(cached_state);
-       return 0;
+       return ret;
 }
 
 static int eb_wait(void *word)
@@ -3690,6 +3695,7 @@ static int extent_write_cache_pages(struct extent_io_tree *tree,
        struct inode *inode = mapping->host;
        int ret = 0;
        int done = 0;
+       int err = 0;
        int nr_to_write_done = 0;
        struct pagevec pvec;
        int nr_pages;
@@ -3776,8 +3782,8 @@ retry:
                                unlock_page(page);
                                ret = 0;
                        }
-                       if (ret)
-                               done = 1;
+                       if (!err && ret < 0)
+                               err = ret;
 
                        /*
                         * the filesystem may choose to bump up nr_to_write.
@@ -3789,7 +3795,7 @@ retry:
                pagevec_release(&pvec);
                cond_resched();
        }
-       if (!scanned && !done) {
+       if (!scanned && !done && !err) {
                /*
                 * We hit the last page and there is more work to be done: wrap
                 * back to the start of the file
@@ -3799,7 +3805,7 @@ retry:
                goto retry;
        }
        btrfs_add_delayed_iput(inode);
-       return ret;
+       return err;
 }
 
 static void flush_epd_write_bio(struct extent_page_data *epd)