From mboxrd@z Thu Jan 1 00:00:00 1970 From: Cyrill Gorcunov Subject: Re: [patch 3/7] fs, notify: Add file handle entry into inotify_inode_mark Date: Wed, 14 Nov 2012 15:03:31 +0400 Message-ID: <20121114110331.GI16685@moon> References: <20121112101440.665694060@openvz.org> <2092535.8S9EcgmZCZ@deuteros> <50A37692.1010500@parallels.com> <1491483.8kFV7tRC1p@deuteros> <50A378C0.70406@parallels.com> Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii Cc: Tvrtko Ursulin , David Rientjes , Andrew Morton , linux-kernel@vger.kernel.org, linux-fsdevel@vger.kernel.org, Al Viro , Alexey Dobriyan , James Bottomley , Matthew Helsley , aneesh.kumar@linux.vnet.ibm.com, bfields@fieldses.org To: Pavel Emelyanov Return-path: Content-Disposition: inline In-Reply-To: <50A378C0.70406@parallels.com> Sender: linux-kernel-owner@vger.kernel.org List-Id: linux-fsdevel.vger.kernel.org On Wed, Nov 14, 2012 at 02:56:00PM +0400, Pavel Emelyanov wrote: > >>> How much space does a typical file system need to encode a handle? Am I > >>> right that for must it is just a few bytes? (I just glanced at the code > >>> so I might be wrong.) In which case, could the handle buffer be allocated > >>> dynamically depending on the underlying filesystem? Perhaps adding a > >>> facility to query a filesystem about its maximum handle buffer needs? Do > >>> you think the saving would justify this extra work? > >> > >> Well, the MAX_HANDLE_SZ is taken from NFSv4 and is 128 bytes which is quite > >> big for inotify extension indeed. The good news is that this amount of bytes > >> seem to be required for the most descriptive fhandle -- with info about > >> parent, etc. We don't need such, we can live with shorter handle, people > >> said that 40 bytes was enough for that. > >> > >> However, your idea about determining the handle size dynamically seems > >> promising. As far as I can see from the code we can call for encode_fh with > >> size equals zero and filesystem would report back the amount of bytes it > >> requires for a handle. > >> > >> We can try going this route, what do you think? > > > > Sounds much better since that would only add one pointer to the watch > > structure in the normal case. > > > > Also at checkpoint time it will use only a few bytes (compared to 64) for the > > encode buffer for most filesystems. This part is probably not that important > > but still a win. > > No, the thing is -- we need to know the handle _before_ we start checkpoint. > More exactly -- at the time the inotify_add_watch is called. So the memory save > would be not that big. for the worst case (NFSv4) it'll be 128 bytes + 8|4 byte pointer, but for more common cases such as extXfs it'll be shrinked down to 8 byte handle + 8|4 byte pointer, which is a pretty good i think.