From mboxrd@z Thu Jan 1 00:00:00 1970 From: Kent Overstreet Subject: Re: [PATCH RFC] fs/aio: fix sleeping while TASK_INTERRUPTIBLE Date: Wed, 24 Dec 2014 19:29:36 -0800 Message-ID: <20141225032936.GC4415@kmo-pixel> References: <20141223001619.GA26385@ret.masoncoding.com> <20141223184328.GB17185@kvack.org> <1419360926.13012.12@mail.thefacebook.com> <20141223215847.GD17185@kvack.org> <20141225025958.GA30938@moria.home.lan> <20141225031134.GI17185@kvack.org> Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii Cc: Chris Mason , linux-fsdevel@vger.kernel.org, linux-aio@kvack.org, Peter Zijlstra To: Benjamin LaHaise Return-path: Content-Disposition: inline In-Reply-To: <20141225031134.GI17185@kvack.org> Sender: owner-linux-aio@kvack.org List-Id: linux-fsdevel.vger.kernel.org On Wed, Dec 24, 2014 at 10:11:34PM -0500, Benjamin LaHaise wrote: > On Wed, Dec 24, 2014 at 06:59:58PM -0800, Kent Overstreet wrote: > > On Tue, Dec 23, 2014 at 04:58:47PM -0500, Benjamin LaHaise wrote: > > > Hi Chris, > > > > > > On Tue, Dec 23, 2014 at 01:55:26PM -0500, Chris Mason wrote: > > > > Works for me, the patch is mostly a (somewhat commented) list of all > > > > the places we're currently doing it wrong. > > > > > > I think the following change should suffice to fix this issue, and it's a > > > lot easier to review, too. I've given this a quick test, and it works for > > > me. I do have one concern: is it safe to call mutex_lock() when the current > > > task is already on other wait queues? If the answer is no, then it may be > > > necessary to convert ->ring_lock back into spinlock as it was prior to 3.10 > > > to avoid using mutex_lock(). The same question applies to kmap() and > > > copy_to_user(), and those concerns might have implications across the rest > > > of the kernel. Thoughts? > > > > > > + __set_current_state(state); > > > > I don't think this is safe - if we race, and another thread wakes us up, we're > > setting our state back to TASK_INTERRUPTIBLE _without_ us being on the waitlist. > > It should be -- the conditions for going to sleep are checked after current's > state is set here. Ah - and then you set the task state back to TASK_RUNNING if _any_ events were found... yeah, I guess that seems safe. Probably worth a few comments, though :) -- To unsubscribe, send a message with 'unsubscribe linux-aio' in the body to majordomo@kvack.org. For more info on Linux AIO, see: http://www.kvack.org/aio/ Don't email: aart@kvack.org