On 2018/8/29 下午11:38, Nikolay Borisov wrote:
>
>
> On 3.08.2018 08:50, Qu Wenruo wrote:
>> [BUG]
>> During fuzz/007 we hit the following error:
>> ------
>> ====== RUN MAYFAIL /home/adam/btrfs/btrfs-progs/btrfs rescue super-recover
>> -y -v
>> /home/adam/btrfs/btrfs-progs/tests//fuzz-tests/images/bko-200409.raw.restored.scratch
>> ERROR: tree_root block unaligned: 33554431
>> ERROR: superblock checksum matches but it has invalid members
>> ERROR: tree_root block unaligned: 33554431
>> ERROR: superblock checksum matches but it has invalid members
>> ERROR: tree_root block unaligned: 33554431
>> ERROR: superblock checksum matches but it has invalid members
>> ERROR: failed to add chunk map start=12582912 len=8454144: -17 (File exists)
>> Couldn't read chunk tree
>> failed (ignored, ret=139): /home/adam/btrfs/btrfs-progs/btrfs rescue
>> super-recover -y -v
>> /home/adam/btrfs/btrfs-progs/tests//fuzz-tests/images/bko-200409.raw.restored.scratch
>> mayfail: returned code 139 (SEGFAULT), not ignored
>> test failed for case 007-simple-super-recover
>> ------
>>
>> [CAUSE]
>> In __open_ctree_fd(), if we have valid @open_ctree_flags and
>> btrfs_scan_fs_devices() successes without problem, no matter what
>> happens we will call btrfs_close_devices(), thus free all related
>> devices.
>
> Why do you think it's _always_ going to be called? Looking into that
> function it seems this can happen if
> btrfs_setup_chunk_tree_and_device_map fails.
No need to reach btrfS_setup_chunk_tree_and_device_map().
As long as we could reach btrfs_open_devices(), no matter whether if
succeeded or not, we will call btrfs_close_devices() to cleanup the
@fs_devices.
If btrfs_open_devices() fails, we goto fail label in
btrfs_open_devices() which calls btrfs_close_devices().
Or we succeeded in btrfs_open_devices(), then next error label is
out_devices in __open_ctree_fd(), and will call btrfs_close_devices() too.
And since in super recovery we have already called
btrfs_scan_fs_devices() so in __open_ctree_fd() it shouldn't fail.
So either we will hit btrfs_close_devices(), no matter whatever happens.
>>
>> In super-recover, before we call open_ctree(), we have called
>> btrfs_scan_fs_devices() already, so btrfs_scan_fs_devices() should not
>> fail in open_ctree(), fs_devices will always be freed in open_ctree() or
>> close_ctree().
>
> Isn't the actual issue just that we call close_ctree. So the actual life
> time of fs_devices is :
No, no need to call close_ctree().
Just as described above, even failed __open_ctree_fd() could call
btrfs_close_devices() and free @fs_devices.
Thanks,
Qu
>
> 1. Create in btrfs_scan_fs_devices called from btrfs_recover_superblocks
> 2. All other references to those fs_devices will just return the same
> reference.
> 3. Calling close_ctree frees fs_devices.
>
>>
>> [FIX]
>> So in super-recover.c, we should not call btrfs_close_devices(), or we
>> will find fs_devices->list get poisoned, and trigger segfault when
>> exiting.
>>
>> Signed-off-by: Qu Wenruo <w...@suse.com>
>> ---
>> super-recover.c | 3 ---
>> 1 file changed, 3 deletions(-)
>>
>> diff --git a/super-recover.c b/super-recover.c
>> index 880fd7712546..86b3df9867dc 100644
>> --- a/super-recover.c
>> +++ b/super-recover.c
>> @@ -292,9 +292,6 @@ int btrfs_recover_superblocks(const char *dname,
>> no_recover:
>> recover_err_str(ret);
>> free_recover_superblock(&recover);
>> - /* check if we have freed fs_devices in close_ctree() */
>> - if (!root)
>> - btrfs_close_devices(recover.fs_devices);
>> return ret;
>> }
>>
>>
