From mboxrd@z Thu Jan 1 00:00:00 1970 From: Christoph Hellwig Subject: Re: [GIT PULL] Btrfs pull request Date: Wed, 9 Nov 2011 05:34:15 -0500 Message-ID: <20111109103415.GA11064@infradead.org> References: <20111106183851.GA4339@shiny> <4EB7A6CC.5030502@gmx.net> <4EB7D2D8.2060505@gmx.net> <4EB8DEAC.7000906@gmx.net> <20111109010701.GI4149@shiny> <20111109074847.GA16742@infradead.org> <4EBA560B.1070509@jan-o-sch.net> Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii Cc: linux-btrfs , Christoph Hellwig , Chris Mason , Arne Jansen , Andrea Gelmini , LKML To: Jan Schmidt Return-path: In-Reply-To: <4EBA560B.1070509@jan-o-sch.net> List-ID: On Wed, Nov 09, 2011 at 11:29:31AM +0100, Jan Schmidt wrote: > Am 09.11.2011 08:48, schrieb Christoph Hellwig: > >On Tue, Nov 08, 2011 at 08:07:01PM -0500, Chris Mason wrote: > >>Looks like bio_add_page() is failing and we're getting the enomem from > >>there. LVM is only letting us put one page in each bio. > > > >Yes, at the moment all bio based DM targets only allow single page I/O. > > Wait. If I got that correctly, each bio_add_page needs special > ENOMEM treatment (assuming the target could always be a device > mapper target), right? Each bio_add_page caller needs to expect it can't add more than a page worth of data. If you look at callers what write large amounts of data (XFS, mpage code) you'll always see a pattern of: bio = bio_alloc(); while (bytes_left) { len = min(bytes_left, PAGE_SIZE); bytes = bio_add_page(io, ..., len); if (bytes < len) { submit_bio(bio); bio = bio_alloc(); } update_indices(); } submit_bio(bio);