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

[Joe Thornber <thornber@redhat.com>]

On Tue, Oct 29, 2019 at 05:07:30AM +0000, 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)

Correct, the data will not be released until the client calls bcache_abort_fd(), to
indicate that it's giving up on the write.  That way the client is free to retry
io, eg, see this unit test:

 689│static void test_write_bad_issue_stops_flush(void *context)
   1│{
   2│        struct fixture *f = context;
   3│        struct mock_engine *me = f->me;
   4│        struct bcache *cache = f->cache;
   5│        struct block *b;
   6│        int fd = 17;
   7│ 
   8│        T_ASSERT(bcache_get(cache, fd, 0, GF_ZERO, &b));
   9│        _expect_write_bad_issue(me, fd, 0);
  10│        bcache_put(b);
  11│        T_ASSERT(!bcache_flush(cache));
  12│
  13│        // we'll let it succeed the second time
  14│        _expect_write(me, fd, 0);
  15│        _expect(me, E_WAIT);
  16│        T_ASSERT(bcache_flush(cache));
  17│}


> 2. when dev_write_bytes failed, call dev_unset_last_byte with "fd=-1" is wrong.

Quite possibly, this unset_last_byte stuff is a hack that Dave put in.  I'll look some more.


> 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;
>      }
>   }

Yep, this is good. Added.


> -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;
>   }

No, the wait functions just wait for io to complete, they're not interested
in whether it succeeded.  That's what b->error is for.


> 
>   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;
>   }

I don't understand how this changes the behaviour from just checking the
size of cache->errored.

- Joe