Received: from eggs.gnu.org ([2001:4830:134:3::10]:49594) by lists.gnu.org with esmtp (Exim 4.71) (envelope-from ) id 1cejSM-000367-PA for qemu-devel@nongnu.org; Fri, 17 Feb 2017 09:24:19 -0500 Received: from Debian-exim by eggs.gnu.org with spam-scanned (Exim 4.71) (envelope-from ) id 1cejSL-0006wD-RV for qemu-devel@nongnu.org; Fri, 17 Feb 2017 09:24:18 -0500 Date: Fri, 17 Feb 2017 15:24:03 +0100 From: Kevin Wolf Message-ID: <20170217142403.GJ5338@noname.redhat.com> References: <1487153430-17260-9-git-send-email-vsementsov@virtuozzo.com> <20170216112506.GB4869@noname.redhat.com> <20170216114944.GD4869@noname.redhat.com> <8d7a5d17-7b1b-578a-67fc-c79beb196a4c@virtuozzo.com> <20170217120935.GC5338@noname.redhat.com> <20170217124842.GE5338@noname.redhat.com> <3a827d65-1028-d0fd-c3aa-cabe61dafbd9@openvz.org> <20170217133430.GI5338@noname.redhat.com> <635d98d9-0f65-4fa6-2208-02ff7d23a2e9@openvz.org> MIME-Version: 1.0 Content-Type: text/plain; charset=us-ascii Content-Disposition: inline In-Reply-To: <635d98d9-0f65-4fa6-2208-02ff7d23a2e9@openvz.org> Subject: Re: [Qemu-devel] [PATCH v15 08/25] block: introduce auto-loading bitmaps List-Id: List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , To: "Denis V. Lunev" Cc: Vladimir Sementsov-Ogievskiy , qemu-block@nongnu.org, qemu-devel@nongnu.org, mreitz@redhat.com, armbru@redhat.com, eblake@redhat.com, jsnow@redhat.com, famz@redhat.com, stefanha@redhat.com, pbonzini@redhat.com Am 17.02.2017 um 14:48 hat Denis V. Lunev geschrieben: > On 02/17/2017 04:34 PM, Kevin Wolf wrote: > > Am 17.02.2017 um 14:22 hat Denis V. Lunev geschrieben: > >> But for sure this is bad from the downtime point of view. > >> On migrate you will have to write to the image and re-read > >> it again on the target. This would be very slow. This will > >> not help for the migration with non-shared disk too. > >> > >> That is why we have specifically worked in a migration, > >> which for a good does not influence downtime at all now. > >> > >> With a write we are issuing several write requests + sync. > >> Our measurements shows that bdrv_drain could take around > >> a second on an averagely loaded conventional system, which > >> seems unacceptable addition to me. > > I'm not arguing against optimising migration, I fully agree with you. I > > just think that we should start with a correct if slow base version and > > then add optimisation to that, instead of starting with a broken base > > version and adding to that. > > > > Look, whether you do the expensive I/O on open/close and make that a > > slow operation or whether you do it on invalidate_cache/inactivate > > doesn't really make a difference in term of slowness because in general > > both operations are called exactly once. But it does make a difference > > in terms of correctness. > > > > Once you do the optimisation, of course, you'll skip writing those > > bitmaps that you transfer using a different channel, no matter whether > > you skip it in bdrv_close() or in bdrv_inactivate(). > > > > Kevin > I do not understand this point as in order to optimize this > we will have to create specific code path or option from > the migration code and keep this as an ugly kludge forever. The point that I don't understand is why it makes any difference for the follow-up migration series whether the writeout is in bdrv_close() or bdrv_inactivate(). I don't really see the difference between the two from a migration POV; both need to be skipped if we transfer the bitmap using a different channel. Maybe I would see the reason if I could find the time to look at the migration patches first, but unfortunately I don't have this time at the moment. My point is just that generally we want to have a correctly working qemu after every single patch, and even more importantly after every series. As the migration series is separate from this, I don't think it's a good excuse for doing worse than we could easily do here. Kevin