Re: [PATCH] Btrfs: avoid deadlock with memory reclaim due to allocation of devices

[Nikolay Borisov <nborisov@suse.com>]

On 9.01.19 г. 20:26 ч., David Sterba wrote:
> On Thu, Dec 13, 2018 at 09:17:25PM +0000, fdmanana@kernel.org wrote:
>> -	if (list_empty(&fs_devices->resized_devices))
>> -		return;
>> -
>> -	mutex_lock(&fs_devices->device_list_mutex);
>>  	mutex_lock(&fs_info->chunk_mutex);
>>  	list_for_each_entry_safe(curr, next, &fs_devices->resized_devices,
>>  				 resized_list) {
>> @@ -7309,7 +7306,6 @@ void btrfs_update_commit_device_size(struct btrfs_fs_info *fs_info)
>>  		curr->commit_total_bytes = curr->disk_total_bytes;
> 
> I'm not sure about removing the device_list_mutex that's said to protect
> the commit_total_bytes (comment in struct btrfs_device).
> 
> Otherwise the logic is ok, the double lock could happen as you describe.
> 
> btrfs_update_commit_device_size is called from btrfs_commit_transaction,
> at the same time as commit_bytes_used. The latter is handled in a
> similar way in btrfs_update_commit_device_bytes_used, but does not take
> the device_list_mutex.
> 
> commit_total_bytes is checked several times (eg. in write_dev_supers) to
> see if writing the superblock copy is still within the device range.
> 
> So, without the protected change, it's theoretically possible that a
> stale value is used for the test and the superblock is either written
> though it should not, and the other way around.

But can it really, btrfs_[grow|shrink]_device happen under transaction
and their modification of the device_disk_total_bytes (the value
assigned to commit_total_bytes) always happen under chunk_mutex. Also
the updates to both values are really owned by the transaction, so even
if grow/shrink modify those value they will queue those changes in a new
transaction, hence write_dev_super will see consistent value in the
current transaction.


I have a patch from Jeff (which is part of a bigger work) that actually
unifies the resize/device size change lists into a single single and
makes the code a bit easier to grok, nevertheless the above explanation
is still correct even without this patch.


> 
> This would require a resize racing at the time of the check. Grow and
> shrink seem to take chunk_mutex while adjusting all the total/size
> values, but it's not actually easy to follow as sometimes there are
> helpers like btrfs_device_set_total_bytes used and sometimes it's direct
> access.
> 
> That the device_list_mutex can be safely dropped probably follows from
> the simple fact that btrfs_update_commit_device_bytes_used is called
> before write_dev_supers in the same context.
> 
> But this sounds too simple, given that there are locks taken and
> released and btrfs_write_and_wait_transaction called between.
> 
> Referencing this code:
> 
> 2201         btrfs_update_commit_device_size(fs_info);
> 2202         btrfs_update_commit_device_bytes_used(cur_trans);
> 2203
> 2204         clear_bit(BTRFS_FS_LOG1_ERR, &fs_info->flags);
> 2205         clear_bit(BTRFS_FS_LOG2_ERR, &fs_info->flags);
> 2206
> 2207         btrfs_trans_release_chunk_metadata(trans);
> 2208
> 2209         spin_lock(&fs_info->trans_lock);
> 2210         cur_trans->state = TRANS_STATE_UNBLOCKED;
> 2211         fs_info->running_transaction = NULL;
> 2212         spin_unlock(&fs_info->trans_lock);
> 2213         mutex_unlock(&fs_info->reloc_mutex);
> 2214
> 2215         wake_up(&fs_info->transaction_wait);
> 2216
> 2217         ret = btrfs_write_and_wait_transaction(trans);
> 2218         if (ret) {
> 2219                 btrfs_handle_fs_error(fs_info, ret,
> 2220                                       "Error while writing out transaction");
> 2221                 mutex_unlock(&fs_info->tree_log_mutex);
> 2222                 goto scrub_continue;
> 2223         }
> 2224
> 2225         ret = write_all_supers(fs_info, 0);
>