From mboxrd@z Thu Jan 1 00:00:00 1970 From: Jens Axboe Subject: Re: [patch v3] splice: fix race with page invalidation Date: Wed, 30 Jul 2008 19:54:07 +0200 Message-ID: <20080730175406.GN20055@kernel.dk> References: Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii Cc: torvalds@linux-foundation.org, akpm@linux-foundation.org, nickpiggin@yahoo.com.au, linux-fsdevel@vger.kernel.org, linux-kernel@vger.kernel.org, linux-mm@kvack.org To: Miklos Szeredi Return-path: Received: from brick.kernel.dk ([87.55.233.238]:2693 "EHLO kernel.dk" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1752317AbYG3RyL (ORCPT ); Wed, 30 Jul 2008 13:54:11 -0400 Content-Disposition: inline In-Reply-To: Sender: linux-fsdevel-owner@vger.kernel.org List-ID: On Wed, Jul 30 2008, Miklos Szeredi wrote: > On Wed, 30 Jul 2008, Linus Torvalds wrote: > > On Wed, 30 Jul 2008, Miklos Szeredi wrote: > > > > > > There are no real disadvantages: splice() from a file was > > > originally meant to be asynchronous, but in reality it only did > > > that for non-readahead pages, which happen rarely. > > > > I still don't like this. I still don't see the point, and I still > > think there is something fundamentally wrong elsewhere. You snipped the part where Linus objected to dismissing the async nature, I fully agree with that part. > We discussed the possible solutions with Nick, and came to the > conclusion, that short term (i.e. 2.6.27) this is probably the best > solution. Ehm where? Nick also said that he didn't like removing the ->confirm() bits as they are completely related to the async nature of splice. You already submitted this exact patch earlier and it was nak'ed. > Long term sure, I have no problem with implementing async splice. > > In fact, I may even have personal interest in looking at splice, > because people are asking for a zero-copy interface for fuse. > > But that is definitely not 2.6.27, so I think you should reconsider > taking this patch, which is obviously correct due to its simplicity, > and won't cause any performance regressions either. Then please just fix the issue, instead of removing the bits that make this possible. -- Jens Axboe