[Gluster-devel] New locking strategy
Xavier Hernandez
xhernandez at datalab.es
Tue Apr 3 15:20:59 UTC 2012
On 04/03/2012 04:53 PM, Anand Avati wrote:
>
>
> On Tue, Apr 3, 2012 at 6:20 AM, Xavier Hernandez
> <xhernandez at datalab.es <mailto:xhernandez at datalab.es>> wrote:
>
> Hello developers,
>
> I'm currently trying to implement a new method for managing inode
> and entry modifications that will be faster (I hope) than the
> current method for the most common cases. To do so I need to know
> exactly how the locking mechanism works. I have been browsing the
> source code and doing some tests, and I would like to be sure that
> I have understood it correctly before continuing.
>
> All information is based on latest qa releases from 3.3 branch.
>
> My understanding is this:
>
> - There are three locking fops: lk, inodelk and entrylk.
> - Client application locks created using fcntl() are received by
> the translators as lk requests.
> - All other functionalities of lk fop are not currently used by
> any translator (I mean F_RESLK_LCK, F_RESLK_LCKW, F_RESLK_UNLCK
> and F_GETLK_FD).
> - inodelk and entrlylk are only used by AFR to lock inodes or
> directory entries before modification.
> - Translators don't generate lk requests internally.
> - Client application requests cannot directly generate an inodelk
> or entrylk requests.
>
>
> Correct so far.
>
> - inodelk and entrylk locks are always mandatory.
>
>
> inodelk and entrylk are always advisory. Never mandatory (at least so far)
Sorry, I was thinking one thing and said another. Never mind, it's as I
expected: no request will be blocked by any of these locks (except
themselves)
> - lk locks may be mandatory or advisory.
>
>
> mandatory mode is not tested for a very long time.
Ok, I won't waste much time on mandatory locks.
> - lk and inodelk are independent from each other, meaning that a
> lock using lk will not be visible to inodelk and will not block
> it. inodelk won't block lk requests neither.
> - User requests can only be blocked by lk created locks
>
>
> Correct, and only if mandatory locks are enabled (which isn't tested
> right now)
Ok.
>
> (if a write request from user is allowed to pass without using
> inodelk, it won't be blocked by a previous inodelk).
>
>
> inodelks are never mandatory, so the question does not apply
Ok.
Thank you very much
Xavi
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://supercolony.gluster.org/pipermail/gluster-devel/attachments/20120403/45302d70/attachment-0003.html>
More information about the Gluster-devel
mailing list