"Yan, Zheng" <zyan@xxxxxxxxxx> writes:
>> On Sep 8, 2018, at 00:03, Luis Henriques <lhenriques@xxxxxxxx> wrote:
>>
>> "Yan, Zheng" <ukernel@xxxxxxxxx> writes:
>>
>>> On Fri, Sep 7, 2018 at 12:08 AM Luis Henriques <lhenriques@xxxxxxxx> wrote:
>> <...>
>>
>>>> + if (objoff || (copy_len < src_ci->i_layout.object_size)) {
>>>> + /* Do not copy beyond this object */
>>>> + if (copy_len > objlen)
>>>> + copy_len = objlen;
>>>> + err = do_splice_direct(src_file, &src_off, dst_file,
>>>> + &dst_off, copy_len, flags);
>>>
>>> We should release caps refs before calling do_splice_direct().
>>> do_splice_direct calls read_iter and write_iter, which will get caps
>>> refs again.
>>
>> That's... annoying. I could release caps, call do_slice_direct and get
>> the caps again but this would introduce a race window.
>>
>> I could also re-write this loop so that it would:
>>
>> - do all the required do_splice_direct calls
>> - get the caps
>> - loop doing all the remote object copies
>> - release caps
>>
>> But the whole operation would still not be atomic. Do you have any
>> suggestion on how to make it atomic?
>
> The default implementation (using do_splice_direct) is not atomic. I
> think we just need to ensure write on single object is atomic.
Well... yep, that probably makes sense. In that case, I'll rework the
code to do the get+release caps in the right places. Again, thanks for
your valuable input!
Cheers,
--
Luis
