[linux-lvm] resend patch - bcache may mistakenly write data to another disk when writes error

[Heming Zhao]

Add another comment.

The key of commit 2938b4dcc & 6b0d969b is function _invalidate_fd().
But from my analysis, it looks your codes do not fix this issue.

_invalidate_fd calls bcache_invalidate_fd & bcache_abort_fd.
bcache_invalidate_fd work as before, only return true or false to indicate there is or isn't fd in cache->errored.
Then bcache_abort_fd calls _abort_v, _abort_v calls _unlink_block & _free_block.
These two functions only delist block from currently cache->errored & cache->free list.
The data still in radix tree with flags BF_DIRTY.
So when next lvm reuses closed error fd, and calls _lookup_or_read_block (with legacy fd), the data will reuse again.
The bug will be trigger.

Thanks,
zhm

On 10/29/19 1:07 PM, heming.zhao wrote:
> Hello Joe,
> 
> Please check my comments for your commit 2938b4dcc & 6b0d969b
> 
> 1. b->ref_count is non-zero, and write error happens, the data never release?
>     (no place to call _unlink_block & _free_block)
> 
> 2. when dev_write_bytes failed, call dev_unset_last_byte with "fd=-1" is wrong.
> 
> 3. I still think below error handling should be added.
>     Below base on stable-2.02, but the core idea is same, should access the return value of io_submit & io_getevents.
> ```
> static bool _async_issue(struct io_engine *ioe, enum dir d, int fd,
>     ... ...
>      if (r < 0) {
>          _cb_free(e->cbs, cb);
> +       ((struct block *)context)->error = r; <== assign errno & print warning
> +       log_warn("io_submit <%c> off %llu bytes %llu return %d:%s",
> +           (d == DIR_READ) ? 'R' : 'W', (long long unsigned)offset,
> +           (long long unsigned)nbytes, r, strerror(-r));
>          return false;
>      }
> 
> static void _issue_low_level(struct block *b, enum dir d)
>     ... ...
>      dm_list_move(&cache->io_pending, &b->list);
> 
>      if (!cache->engine->issue(cache->engine, d, b->fd, sb, se, b->data, b)) {
> -       /* FIXME: if io_submit() set an errno, return that instead of EIO? */
> -       _complete_io(b, -EIO);
> +       _complete_io(b, b->error); <=== this pass the right errno to caller.
>          return;
>      }
>   }
> 
> -static void _wait_all(struct bcache *cache)
> +static bool _wait_all(struct bcache *cache) <=== change to return error
>   {
> +   bool ret = true;
>      while (!dm_list_empty(&cache->io_pending))
> -       _wait_io(cache);
> +       ret = _wait_io(cache);
> +   return ret;
>   }
> 
> -static void _wait_specific(struct block *b)
> +static bool _wait_specific(struct block *b) <=== change to return error
>   {
> +   bool ret = true;
>      while (_test_flags(b, BF_IO_PENDING))
> -       _wait_io(b->cache);
> +       ret = _wait_io(b->cache);
> +   return ret;
>   }
> 
>   bool bcache_flush(struct bcache *cache) <==== add more error handling
>   {
> +   bool write_ret = true, wait_ret = true;
> 
>      ... ...
>          _issue_write(b);
> +       if (b->error) write_ret = false;
>      }
> 
> -   _wait_all(cache);
> +   wait_ret = _wait_all(cache);
> 
> -   return dm_list_empty(&cache->errored);
> +   if (write_ret == false || wait_ret == false ||
> +           !dm_list_empty(&cache->errored))
> +       return false;
> +   else
> +       return true;
>   }
> ```
> 
> Thanks,
> zhm
> 
> On 10/28/19 11:43 PM, Joe Thornber wrote:
>> On Thu, Oct 24, 2019 at 03:06:18AM +0000, Heming Zhao wrote:
>>> First fail is in bcache_flush, then bcache_invalidate_fd does nothing because the data
>>> in cache->errored, which not belongs to dirty & clean list. Then the data mistakenly
>>> move from cache->errored into cache->dirty by "bcache_get => _lookup_or_read_block"
>>> (because the data holds BF_DIRTY flag).
>>
>> I just pushed a couple of patches that will hopefully fix this issue for you:
>>
>> commit 6b0d969b2a85ba69046afa26af4d7bcddddbccd5 (HEAD -> master, origin/master, origin/HEAD, 2019-10-11-bcache-purge)
>> Author: Joe Thornber <ejt at redhat.com>
>> Date:   Mon Oct 28 15:01:47 2019 +0000
>>
>>      [label] Use bcache_abort_fd() to ensure blocks are no longer in the cache.
>>      The return value from bcache_invalidate_fd() was not being checked.
>>      So I've introduced a little function, _invalidate_fd() that always
>>      calls bcache_abort_fd() if the write fails.
>>
>> commit 2938b4dcca0a1df661758abfab7f402ea7aab018
>> Author: Joe Thornber <ejt at redhat.com>
>> Date:   Mon Oct 28 14:29:47 2019 +0000
>>
>>      [bcache] add bcache_abort()
>>      This gives us a way to cope with write failures.
>>
>>
>>
>>
>> Also there are big changes to bcache coming, that remove file descriptors from
>> the interface completely.  See the branch 2019-09-05-add-io-manager for more info
>> (bcache has been renamed io_manager).
>>
>> - Joe
>>
>> _______________________________________________
>> linux-lvm mailing list
>> linux-lvm at redhat.com
>> https://www.redhat.com/mailman/listinfo/linux-lvm
>> read the LVM HOW-TO at http://tldp.org/HOWTO/LVM-HOWTO/
>>