On Mon, 12 Aug 2013 17:16:28 +0400, Vyacheslav Dubeyko wrote:
>> Otherwise, your change of nilfs_segbuf_submit_bio() is just a
>> equivalent transformation and doesn't fix problem, that is, a mismatch
>> of the number of calls between complete() and wait_for_completion().
>>
>> In your patch, nilfs_end_bio_write() function calls the complete()
>> routine as before even if it received an EOPNOTSUPP error.
>>
>> In that case, segbuf->sb_nbio must be incremented to call
>> wait_for_completion() the right number of times.
>>
>
> I moved incrementing of sb_nbio nearly after submit_bio(mode, bio). So,
> firstly, we call submit_bio(mode, bio) and segbuf->sb_nbio++. Then,
> because of asynchronous nature of nilfs_end_bio_write() call, we have
> two alternatives: (1) nilfs_end_bio_write() detects that BIO_EOPNOTSUPP
> flag is set; (2) BIO_EOPNOTSUPP flag isn't set because bio is not
> processed yet or we haven't error.
>
> If nilfs_end_bio_write() method would detect -EOPNOTSUPP error then it
> set BIO_EOPNOTSUPP flag, increment sb_err field and to call complete().
>
> So, if we detect that BIO_EOPNOTSUPP flag is set in
> nilfs_segbuf_submit_bio() then we decrement sb_nbio and to return from
> nilfs_segbuf_submit_bio() with error. And if we had submitted
> successfully some bios before then it will need to wait theirs
sb_nbio must be incremented exactly the same number of times as
complete() function was called (or will be called) because
nilfs_segbuf_wait() will call wail_for_completion() for the number of
times set to sb_nbio:
do {
wait_for_completion(&segbuf->sb_bio_event);
} while (--segbuf->sb_nbio > 0);
Two functions complete() and wait_for_completion() must be called the
same number of times for the same sb_bio_event. Otherwise,
wait_for_completion() will hang or leak.
In your patch, nilfs_end_bio_write() will call complete() even in the
case "(1) nilfs_end_bio_write() detects that BIO_EOPNOTSUPP flag is
set" (like the current implementation). Therefore, sb_nbio must be
incremented for that count.
This looks another issue of the current implemention and the second
bug can be fixed as follows:
submit_bio(mode, bio);
+ segbuf->sb_nbio++;
if (bio_flagged(bio, BIO_EOPNOTSUPP)) {
bio_put(bio);
err = -EOPNOTSUPP;
goto failed;
}
- segbuf->sb_nbio++;
bio_put(bio);
> By the way, why sb_err has atomic_t type but sb_nbio is simple int type?
> Maybe, sb_nbio should be atomic_t type too?
No, the current implementation is correct.
sb_err is incremented asynchronously by nilfs_end_bio_write(), so it
is implemented with atomic_t type. On the other hand, sb_nbio is
operated sequentially and therefore it doesn't need to use atomic_t
type.
Regards,
Ryusuke Konishi
--
To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
