ext4: fix race between sync and completed io work
authorJeff Moyer <jmoyer@redhat.com>
Mon, 5 Mar 2012 15:29:52 +0000 (10:29 -0500)
committerTheodore Ts'o <tytso@mit.edu>
Mon, 5 Mar 2012 15:29:52 +0000 (10:29 -0500)
The following command line will leave the aio-stress process unkillable
on an ext4 file system (in my case, mounted on /mnt/test):

aio-stress -t 20 -s 10 -O -S -o 2 -I 1000 /mnt/test/aiostress.3561.4 /mnt/test/aiostress.3561.4.20 /mnt/test/aiostress.3561.4.19 /mnt/test/aiostress.3561.4.18 /mnt/test/aiostress.3561.4.17 /mnt/test/aiostress.3561.4.16 /mnt/test/aiostress.3561.4.15 /mnt/test/aiostress.3561.4.14 /mnt/test/aiostress.3561.4.13 /mnt/test/aiostress.3561.4.12 /mnt/test/aiostress.3561.4.11 /mnt/test/aiostress.3561.4.10 /mnt/test/aiostress.3561.4.9 /mnt/test/aiostress.3561.4.8 /mnt/test/aiostress.3561.4.7 /mnt/test/aiostress.3561.4.6 /mnt/test/aiostress.3561.4.5 /mnt/test/aiostress.3561.4.4 /mnt/test/aiostress.3561.4.3 /mnt/test/aiostress.3561.4.2

This is using the aio-stress program from the xfstests test suite.
That particular command line tells aio-stress to do random writes to
20 files from 20 threads (one thread per file).  The files are NOT
preallocated, so you will get writes to random offsets within the
file, thus creating holes and extending i_size.  It also opens the
file with O_DIRECT and O_SYNC.

On to the problem.  When an I/O requires unwritten extent conversion,
it is queued onto the completed_io_list for the ext4 inode.  Two code
paths will pull work items from this list.  The first is the
ext4_end_io_work routine, and the second is ext4_flush_completed_IO,
which is called via the fsync path (and O_SYNC handling, as well).
There are two issues I've found in these code paths.  First, if the
fsync path beats the work routine to a particular I/O, the work
routine will free the io_end structure!  It does not take into account
the fact that the io_end may still be in use by the fsync path.  I've
fixed this issue by adding yet another IO_END flag, indicating that
the io_end is being processed by the fsync path.

The second problem is that the work routine will make an assignment to
io->flag outside of the lock.  I have witnessed this result in a hang
at umount.  Moving the flag setting inside the lock resolved that
problem.

The problem was introduced by commit b82e384c7b ("ext4: optimize
locking for end_io extent conversion"), which first appeared in 3.2.
As such, the fix should be backported to that release (probably along
with the unwritten extent conversion race fix).

Signed-off-by: Jeff Moyer <jmoyer@redhat.com>
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
CC: stable@kernel.org
fs/ext4/ext4.h
fs/ext4/fsync.c
fs/ext4/page-io.c

index 6481e3ca35285c16502e90985d58274ca1cc665d..37e7d8b66c99567520033e0fb4b7b8af9f76d51f 100644 (file)
@@ -185,6 +185,7 @@ struct mpage_da_data {
 #define EXT4_IO_END_ERROR      0x0002
 #define EXT4_IO_END_QUEUED     0x0004
 #define EXT4_IO_END_DIRECT     0x0008
+#define EXT4_IO_END_IN_FSYNC   0x0010
 
 struct ext4_io_page {
        struct page     *p_page;
index 00a2cb753efdeca63d6bbba4912729dc1c854182..bb6c7d8113134ee7796a141065d9ed602650f10e 100644 (file)
@@ -89,6 +89,7 @@ int ext4_flush_completed_IO(struct inode *inode)
                io = list_entry(ei->i_completed_io_list.next,
                                ext4_io_end_t, list);
                list_del_init(&io->list);
+               io->flag |= EXT4_IO_END_IN_FSYNC;
                /*
                 * Calling ext4_end_io_nolock() to convert completed
                 * IO to written.
@@ -108,6 +109,7 @@ int ext4_flush_completed_IO(struct inode *inode)
                if (ret < 0)
                        ret2 = ret;
                spin_lock_irqsave(&ei->i_completed_io_lock, flags);
+               io->flag &= ~EXT4_IO_END_IN_FSYNC;
        }
        spin_unlock_irqrestore(&ei->i_completed_io_lock, flags);
        return (ret2 < 0) ? ret2 : 0;
index 9e1b8eb1e7ac44bb0e12628d0973bfb4fe8da269..dcdeef169a69811cf995cd34c37bbdcb6d79fd52 100644 (file)
@@ -129,12 +129,18 @@ static void ext4_end_io_work(struct work_struct *work)
        unsigned long           flags;
 
        spin_lock_irqsave(&ei->i_completed_io_lock, flags);
+       if (io->flag & EXT4_IO_END_IN_FSYNC)
+               goto requeue;
        if (list_empty(&io->list)) {
                spin_unlock_irqrestore(&ei->i_completed_io_lock, flags);
                goto free;
        }
 
        if (!mutex_trylock(&inode->i_mutex)) {
+               bool was_queued;
+requeue:
+               was_queued = !!(io->flag & EXT4_IO_END_QUEUED);
+               io->flag |= EXT4_IO_END_QUEUED;
                spin_unlock_irqrestore(&ei->i_completed_io_lock, flags);
                /*
                 * Requeue the work instead of waiting so that the work
@@ -147,9 +153,8 @@ static void ext4_end_io_work(struct work_struct *work)
                 * yield the cpu if it sees an end_io request that has already
                 * been requeued.
                 */
-               if (io->flag & EXT4_IO_END_QUEUED)
+               if (was_queued)
                        yield();
-               io->flag |= EXT4_IO_END_QUEUED;
                return;
        }
        list_del_init(&io->list);