Re: [PATCH] btrfs: Always try all copies when reading extent buffers

[Nikolay Borisov <nborisov@suse.com>]

On 6.11.18 г. 16:53 ч., Qu Wenruo wrote:
> 
> 
> On 2018/11/6 下午10:40, Nikolay Borisov wrote:
>> When a metadata read is served the endio routine btree_readpage_end_io_hook
>> is called which eventually runs the tree-checker. If tree-checker fails
>> to validate the read eb then it sets EXTENT_BUFFER_CORRUPT flag. This
>> leads to btree_read_extent_buffer_pages wrongly assuming that all
>> available copies of this extent buffer are wrong and failing prematurely.
>> Fix this modify btree_read_extent_buffer_pages to read all copies of
>> the data.
>>
>> This failure was exhibitted in xfstests btrfs/124 which would
>> spuriously fail its balance operations. The reason was that when balance
>> was run following re-introduction of the missing raid1 disk
>> __btrfs_map_block would map the read request to stripe 0, which
>> corresponded to devid 2 (the disk which is being removed in the test):
>>
>>     item 2 key (FIRST_CHUNK_TREE CHUNK_ITEM 3553624064) itemoff 15975 itemsize 112
>> 	length 1073741824 owner 2 stripe_len 65536 type DATA|RAID1
>> 	io_align 65536 io_width 65536 sector_size 4096
>> 	num_stripes 2 sub_stripes 1
>> 		stripe 0 devid 2 offset 2156920832
>> 		dev_uuid 8466c350-ed0c-4c3b-b17d-6379b445d5c8
>> 		stripe 1 devid 1 offset 3553624064
>> 		dev_uuid 1265d8db-5596-477e-af03-df08eb38d2ca
>>
>> This caused read requests for a checksum item that to be routed to the
>> stale disk which triggered the aforementioned logic involving
>> EXTENT_BUFFER_CORRUPT flag. This then triggered cascading failures of
>> the balance operation.
>>
>> Signed-off-by: Nikolay Borisov <nborisov@suse.com>
>> Suggested-by: Qu Wenruo <wqu@suse.com>
> 
> Reviewed-by: Qu Wenruo <wqu@suse.com>
> 
> However there is still a tiny piece missing.
> 
> Tree checker is done after some basic checks, including:
> 1) bytenr
> 2) level
> 3) fsid
> 4) csum
> 
> 1~2) can be easily skipped just by pure luck.
> 
> But 3) and especially 4) are not that easy to hit.
> Not to mention meeting both 3) and 4), since csum range covers fsid.
> 
> So I must say you're a really super lucky guy!

s/lucky/unlucky/ :)

I wonder if reads just land in some random memory on the stale disk
hence it's really a matter of what's "there" so that reads fail with
random reasons?

> 
> Thanks,
> Qu
> 
> 
>> Fixes: a826d6dcb32d ("Btrfs: check items for correctness as we search")
>> ---
>>  fs/btrfs/disk-io.c | 11 +----------
>>  1 file changed, 1 insertion(+), 10 deletions(-)
>>
>> diff --git a/fs/btrfs/disk-io.c b/fs/btrfs/disk-io.c
>> index 00ee5e37e989..279c6dbcc736 100644
>> --- a/fs/btrfs/disk-io.c
>> +++ b/fs/btrfs/disk-io.c
>> @@ -477,9 +477,9 @@ static int btree_read_extent_buffer_pages(struct btrfs_fs_info *fs_info,
>>  	int mirror_num = 0;
>>  	int failed_mirror = 0;
>>  
>> -	clear_bit(EXTENT_BUFFER_CORRUPT, &eb->bflags);
>>  	io_tree = &BTRFS_I(fs_info->btree_inode)->io_tree;
>>  	while (1) {
>> +		clear_bit(EXTENT_BUFFER_CORRUPT, &eb->bflags);
>>  		ret = read_extent_buffer_pages(io_tree, eb, WAIT_COMPLETE,
>>  					       mirror_num);
>>  		if (!ret) {
>> @@ -493,15 +493,6 @@ static int btree_read_extent_buffer_pages(struct btrfs_fs_info *fs_info,
>>  				break;
>>  		}
>>  
>> -		/*
>> -		 * This buffer's crc is fine, but its contents are corrupted, so
>> -		 * there is no reason to read the other copies, they won't be
>> -		 * any less wrong.
>> -		 */
>> -		if (test_bit(EXTENT_BUFFER_CORRUPT, &eb->bflags) ||
>> -		    ret == -EUCLEAN)
>> -			break;
>> -
>>  		num_copies = btrfs_num_copies(fs_info,
>>  					      eb->start, eb->len);
>>  		if (num_copies == 1)
>>
>