Re: [PATCH 2/2] fuse: wait for writeback in fuse_file_fallocate() -v2

[Maxim Patlasov <mpatlasov@parallels.com>]

On 09/11/2013 02:12 PM, Miklos Szeredi wrote:
> On Fri, Aug 30, 2013 at 1:33 PM, Maxim Patlasov <mpatlasov@parallels.com> wrote:
>> 08/30/2013 01:13 PM, Miklos Szeredi пишет:
>>
>>> On Thu, Aug 29, 2013 at 6:41 PM, Miklos Szeredi <miklos@szeredi.hu> wrote:
>>>> BTW, isn't it enough to do the filemap_write_and_wait() *plus* the
>>>> fuse_set_nowrite()?
>>> Thought about it a bit and I think this should do fine.
>>>
>>> Any writes before the fallocate will go trough before the fallocate.
>>> i_mutex guarantees that only one instance of fuse_set_nowrite() is
>>> running.  Any mmaped writes during the fallocate() will go after the
>>> fallocate request and the page cache truncation and that's fine too.
>>> Page cache is consistent since it doens't contain pages for those
>>> writes to the hole.  Subsequent reads to that area will fill them in.
>>>
>>> Any other concerns?
>>
>> No. What you suggest looks as a neat and correct solution. I'll resend the
>> updated patch after some testing (since now till Monday).
>>
>> As for proof-of-correctness, all you wrote above is correct, but the first
>> point had been boiling my mind for a while. I came to the following
>> reasoning (hopefully it is what you meant):
>>
>> The fact that filemap_write_and_wait() returned infers that
>> end_page_writeback() was called for all relevant pages. And fuse doesn't
>> call it before adding request to fi->queued_writes and calling
>> fuse_flush_writepages(). And the latter, in turn, guarantees proper
>> accounting of request in fi->writectr. Here, of course, it's crucial that we
>> can't have concurrent fuse_set_nowrite(), as you explained. Hence, so far as
>> fi->writectr was bumped, fuse_set_nowrite() we call after
>> filemap_write_and_wait() would wait until all changes have gone to the
>> server.
> Any news about this?

Testing updated patch revealed a problem (fsx caught data corruption). 
Then I instrumented debug version to get a cue. The debug version 
survived several days of testing, but now I discovered that that test 
setup was not fully correct. I'll re-run it now.

Thanks,
Maxim