> -----Original Message-----
> From: Dmitri.Lenev@stripped [mailto:Dmitri.Lenev@stripped] On Behalf Of Dmitry Lenev
> Sent: Thursday, September 09, 2010 9:36 PM
> To: Vladislav Vaintroub
> Cc: commits@stripped
> Subject: Re: bzr commit into mysql-trunk branch (vvaintroub:3194) Bug#56585
>
> Hello Vlad!
>
Hello Dmitri,
> I think I have found yet another issue with rw_pr_lock
> implementation:
<skip>
> Sounds like a deadlock? Or do I miss something?
Nope, you're right, thanks for spotting. There is a possibility for deadlock.
> It seems to me that problem can be solved if we ensure that decrement
> of reader_count and setting of allow_writer event will happen atomically
> (within scope of reader_cs critical section).
> And, yes, this probably means that unlock will get more expensive in
> case of multiply readers case, but as we have discussed this is not
> the most common case in MDL.
>
> BTW if both increment and decrement of reader_count will be protected
> by the same critical section they no longer have to be done as
> Interlocked operations.
>
> What do you think?
I think it makes sense. I just measured what does Enter/Leave critical section on unlock()
costs for sysbench readonly scenarios,
and the difference is not measurable. And unless there is a big number of readers locking
and unlocking at the very same time(which
you say does not happen), enter/leave critical section will the same overhead as an atomic
operation.
My last commit on this topic implements it just like you suggest.
Thanks again for looking!
Vlad
