btrfs: set_page_extent_mapped after read_folio in btrfs_cont_expand
authorJosef Bacik <josef@toxicpanda.com>
Wed, 12 Jul 2023 16:44:12 +0000 (12:44 -0400)
committerDavid Sterba <dsterba@suse.com>
Tue, 18 Jul 2023 01:13:58 +0000 (03:13 +0200)
While trying to get the subpage blocksize tests running, I hit the
following panic on generic/476

  assertion failed: PagePrivate(page) && page->private, in fs/btrfs/subpage.c:229
  kernel BUG at fs/btrfs/subpage.c:229!
  Internal error: Oops - BUG: 00000000f2000800 [#1] SMP
  CPU: 1 PID: 1453 Comm: fsstress Not tainted 6.4.0-rc7+ #12
  Hardware name: QEMU KVM Virtual Machine, BIOS edk2-20230301gitf80f052277c8-26.fc38 03/01/2023
  pstate: 61400005 (nZCv daif +PAN -UAO -TCO +DIT -SSBS BTYPE=--)
  pc : btrfs_subpage_assert+0xbc/0xf0
  lr : btrfs_subpage_assert+0xbc/0xf0
  Call trace:
   btrfs_subpage_assert+0xbc/0xf0
   btrfs_subpage_clear_checked+0x38/0xc0
   btrfs_page_clear_checked+0x48/0x98
   btrfs_truncate_block+0x5d0/0x6a8
   btrfs_cont_expand+0x5c/0x528
   btrfs_write_check.isra.0+0xf8/0x150
   btrfs_buffered_write+0xb4/0x760
   btrfs_do_write_iter+0x2f8/0x4b0
   btrfs_file_write_iter+0x1c/0x30
   do_iter_readv_writev+0xc8/0x158
   do_iter_write+0x9c/0x210
   vfs_iter_write+0x24/0x40
   iter_file_splice_write+0x224/0x390
   direct_splice_actor+0x38/0x68
   splice_direct_to_actor+0x12c/0x260
   do_splice_direct+0x90/0xe8
   generic_copy_file_range+0x50/0x90
   vfs_copy_file_range+0x29c/0x470
   __arm64_sys_copy_file_range+0xcc/0x498
   invoke_syscall.constprop.0+0x80/0xd8
   do_el0_svc+0x6c/0x168
   el0_svc+0x50/0x1b0
   el0t_64_sync_handler+0x114/0x120
   el0t_64_sync+0x194/0x198

This happens because during btrfs_cont_expand we'll get a page, set it
as mapped, and if it's not Uptodate we'll read it.  However between the
read and re-locking the page we could have called release_folio() on the
page, but left the page in the file mapping.  release_folio() can clear
the page private, and thus further down we blow up when we go to modify
the subpage bits.

Fix this by putting the set_page_extent_mapped() after the read.  This
is safe because read_folio() will call set_page_extent_mapped() before
it does the read, and then if we clear page private but leave it on the
mapping we're completely safe re-setting set_page_extent_mapped().  With
this patch I can now run generic/476 without panicing.

CC: stable@vger.kernel.org # 6.1+
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: David Sterba <dsterba@suse.com>
fs/btrfs/inode.c

index b12850b..1f58deb 100644 (file)
@@ -4863,9 +4863,6 @@ again:
                ret = -ENOMEM;
                goto out;
        }
-       ret = set_page_extent_mapped(page);
-       if (ret < 0)
-               goto out_unlock;
 
        if (!PageUptodate(page)) {
                ret = btrfs_read_folio(NULL, page_folio(page));
@@ -4880,6 +4877,17 @@ again:
                        goto out_unlock;
                }
        }
+
+       /*
+        * We unlock the page after the io is completed and then re-lock it
+        * above.  release_folio() could have come in between that and cleared
+        * PagePrivate(), but left the page in the mapping.  Set the page mapped
+        * here to make sure it's properly set for the subpage stuff.
+        */
+       ret = set_page_extent_mapped(page);
+       if (ret < 0)
+               goto out_unlock;
+
        wait_on_page_writeback(page);
 
        lock_extent(io_tree, block_start, block_end, &cached_state);