Btrfs: fix memory corruption on failure to submit bio for direct IO
authorFilipe Manana <fdmanana@suse.com>
Wed, 1 Jul 2015 11:13:10 +0000 (12:13 +0100)
committerChris Mason <clm@fb.com>
Thu, 2 Jul 2015 00:17:18 +0000 (17:17 -0700)
If we fail to submit a bio for a direct IO request, we were grabbing the
corresponding ordered extent and decrementing its reference count twice,
once for our lookup reference and once for the ordered tree reference.
This was a problem because it caused the ordered extent to be freed
without removing it from the ordered tree and any lists it might be
attached to, leaving dangling pointers to the ordered extent around.
Example trace with CONFIG_DEBUG_PAGEALLOC=y:

[161779.858707] BUG: unable to handle kernel paging request at 0000000087654330
[161779.859983] IP: [<ffffffff8124ca68>] rb_prev+0x22/0x3b
[161779.860636] PGD 34d818067 PUD 0
[161779.860636] Oops: 0000 [#1] PREEMPT SMP DEBUG_PAGEALLOC
(...)
[161779.860636] Call Trace:
[161779.860636]  [<ffffffffa06b36a6>] __tree_search+0xd9/0xf9 [btrfs]
[161779.860636]  [<ffffffffa06b3708>] tree_search+0x42/0x63 [btrfs]
[161779.860636]  [<ffffffffa06b4868>] ? btrfs_lookup_ordered_range+0x2d/0xa5 [btrfs]
[161779.860636]  [<ffffffffa06b4873>] btrfs_lookup_ordered_range+0x38/0xa5 [btrfs]
[161779.860636]  [<ffffffffa06aab8e>] btrfs_get_blocks_direct+0x11b/0x615 [btrfs]
[161779.860636]  [<ffffffff8119727f>] do_blockdev_direct_IO+0x5ff/0xb43
[161779.860636]  [<ffffffffa06aaa73>] ? btrfs_page_exists_in_range+0x1ad/0x1ad [btrfs]
[161779.860636]  [<ffffffffa06a2c9a>] ? btrfs_get_extent_fiemap+0x1bc/0x1bc [btrfs]
[161779.860636]  [<ffffffff811977f5>] __blockdev_direct_IO+0x32/0x34
[161779.860636]  [<ffffffffa06a2c9a>] ? btrfs_get_extent_fiemap+0x1bc/0x1bc [btrfs]
[161779.860636]  [<ffffffffa06a10ae>] btrfs_direct_IO+0x198/0x21f [btrfs]
[161779.860636]  [<ffffffffa06a2c9a>] ? btrfs_get_extent_fiemap+0x1bc/0x1bc [btrfs]
[161779.860636]  [<ffffffff81112ca1>] generic_file_direct_write+0xb3/0x128
[161779.860636]  [<ffffffffa06affaa>] ? btrfs_file_write_iter+0x15f/0x3e0 [btrfs]
[161779.860636]  [<ffffffffa06b004c>] btrfs_file_write_iter+0x201/0x3e0 [btrfs]
(...)

We were also not freeing the btrfs_dio_private we allocated previously,
which kmemleak reported with the following trace in its sysfs file:

unreferenced object 0xffff8803f553bf80 (size 96):
  comm "xfs_io", pid 4501, jiffies 4295039588 (age 173.936s)
  hex dump (first 32 bytes):
    88 6c 9b f5 02 88 ff ff 00 00 00 00 00 00 00 00  .l..............
    00 00 00 00 00 00 00 00 00 00 c4 00 00 00 00 00  ................
  backtrace:
    [<ffffffff81161ffe>] create_object+0x172/0x29a
    [<ffffffff8145870f>] kmemleak_alloc+0x25/0x41
    [<ffffffff81154e64>] kmemleak_alloc_recursive.constprop.40+0x16/0x18
    [<ffffffff811579ed>] kmem_cache_alloc_trace+0xfb/0x148
    [<ffffffffa03d8cff>] btrfs_submit_direct+0x65/0x16a [btrfs]
    [<ffffffff811968dc>] dio_bio_submit+0x62/0x8f
    [<ffffffff811975fe>] do_blockdev_direct_IO+0x97e/0xb43
    [<ffffffff811977f5>] __blockdev_direct_IO+0x32/0x34
    [<ffffffffa03d70ae>] btrfs_direct_IO+0x198/0x21f [btrfs]
    [<ffffffff81112ca1>] generic_file_direct_write+0xb3/0x128
    [<ffffffffa03e604d>] btrfs_file_write_iter+0x201/0x3e0 [btrfs]
    [<ffffffff8116586a>] __vfs_write+0x7c/0xa5
    [<ffffffff81165da9>] vfs_write+0xa0/0xe4
    [<ffffffff81166675>] SyS_pwrite64+0x64/0x82
    [<ffffffff81464fd7>] system_call_fastpath+0x12/0x6f
    [<ffffffffffffffff>] 0xffffffffffffffff

For read requests we weren't doing any cleanup either (none of the work
done by btrfs_endio_direct_read()), so a failure submitting a bio for a
read request would leave a range in the inode's io_tree locked forever,
blocking any future operations (both reads and writes) against that range.

So fix this by making sure we do the same cleanup that we do for the case
where the bio submission succeeds.

Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>
fs/btrfs/inode.c
fs/btrfs/ordered-data.c

index 855935f6671ae59b1b025c3916d553bc689191ef..c0b2b6b51b2a29db5c05841e0a96ba67dc515a2a 100644 (file)
@@ -8163,9 +8163,8 @@ out_err:
 static void btrfs_submit_direct(int rw, struct bio *dio_bio,
                                struct inode *inode, loff_t file_offset)
 {
-       struct btrfs_root *root = BTRFS_I(inode)->root;
-       struct btrfs_dio_private *dip;
-       struct bio *io_bio;
+       struct btrfs_dio_private *dip = NULL;
+       struct bio *io_bio = NULL;
        struct btrfs_io_bio *btrfs_bio;
        int skip_sum;
        int write = rw & REQ_WRITE;
@@ -8182,7 +8181,7 @@ static void btrfs_submit_direct(int rw, struct bio *dio_bio,
        dip = kzalloc(sizeof(*dip), GFP_NOFS);
        if (!dip) {
                ret = -ENOMEM;
-               goto free_io_bio;
+               goto free_ordered;
        }
 
        dip->private = dio_bio->bi_private;
@@ -8210,25 +8209,55 @@ static void btrfs_submit_direct(int rw, struct bio *dio_bio,
 
        if (btrfs_bio->end_io)
                btrfs_bio->end_io(btrfs_bio, ret);
-free_io_bio:
-       bio_put(io_bio);
 
 free_ordered:
        /*
-        * If this is a write, we need to clean up the reserved space and kill
-        * the ordered extent.
+        * If we arrived here it means either we failed to submit the dip
+        * or we either failed to clone the dio_bio or failed to allocate the
+        * dip. If we cloned the dio_bio and allocated the dip, we can just
+        * call bio_endio against our io_bio so that we get proper resource
+        * cleanup if we fail to submit the dip, otherwise, we must do the
+        * same as btrfs_endio_direct_[write|read] because we can't call these
+        * callbacks - they require an allocated dip and a clone of dio_bio.
         */
-       if (write) {
-               struct btrfs_ordered_extent *ordered;
-               ordered = btrfs_lookup_ordered_extent(inode, file_offset);
-               if (!test_bit(BTRFS_ORDERED_PREALLOC, &ordered->flags) &&
-                   !test_bit(BTRFS_ORDERED_NOCOW, &ordered->flags))
-                       btrfs_free_reserved_extent(root, ordered->start,
-                                                  ordered->disk_len, 1);
-               btrfs_put_ordered_extent(ordered);
-               btrfs_put_ordered_extent(ordered);
+       if (io_bio && dip) {
+               bio_endio(io_bio, ret);
+               /*
+                * The end io callbacks free our dip, do the final put on io_bio
+                * and all the cleanup and final put for dio_bio (through
+                * dio_end_io()).
+                */
+               dip = NULL;
+               io_bio = NULL;
+       } else {
+               if (write) {
+                       struct btrfs_ordered_extent *ordered;
+
+                       ordered = btrfs_lookup_ordered_extent(inode,
+                                                             file_offset);
+                       set_bit(BTRFS_ORDERED_IOERR, &ordered->flags);
+                       /*
+                        * Decrements our ref on the ordered extent and removes
+                        * the ordered extent from the inode's ordered tree,
+                        * doing all the proper resource cleanup such as for the
+                        * reserved space and waking up any waiters for this
+                        * ordered extent (through btrfs_remove_ordered_extent).
+                        */
+                       btrfs_finish_ordered_io(ordered);
+               } else {
+                       unlock_extent(&BTRFS_I(inode)->io_tree, file_offset,
+                             file_offset + dio_bio->bi_iter.bi_size - 1);
+               }
+               clear_bit(BIO_UPTODATE, &dio_bio->bi_flags);
+               /*
+                * Releases and cleans up our dio_bio, no need to bio_put()
+                * nor bio_endio()/bio_io_error() against dio_bio.
+                */
+               dio_end_io(dio_bio, ret);
        }
-       bio_endio(dio_bio, ret);
+       if (io_bio)
+               bio_put(io_bio);
+       kfree(dip);
 }
 
 static ssize_t check_direct_IO(struct btrfs_root *root, struct kiocb *iocb,
index 89656d799ff6fa3772b52cb17d535ebdc1cbb910..52170cf1757e3d25192a6cf6211a9d2edaa52751 100644 (file)
@@ -552,6 +552,10 @@ void btrfs_put_ordered_extent(struct btrfs_ordered_extent *entry)
        trace_btrfs_ordered_extent_put(entry->inode, entry);
 
        if (atomic_dec_and_test(&entry->refs)) {
+               ASSERT(list_empty(&entry->log_list));
+               ASSERT(list_empty(&entry->trans_list));
+               ASSERT(list_empty(&entry->root_extent_list));
+               ASSERT(RB_EMPTY_NODE(&entry->rb_node));
                if (entry->inode)
                        btrfs_add_delayed_iput(entry->inode);
                while (!list_empty(&entry->list)) {
@@ -579,6 +583,7 @@ void btrfs_remove_ordered_extent(struct inode *inode,
        spin_lock_irq(&tree->lock);
        node = &entry->rb_node;
        rb_erase(node, &tree->tree);
+       RB_CLEAR_NODE(node);
        if (tree->last == node)
                tree->last = NULL;
        set_bit(BTRFS_ORDERED_COMPLETE, &entry->flags);