From patchwork Mon Jun 25 13:31:57 2018 Content-Type: text/plain; charset="utf-8" MIME-Version: 1.0 Content-Transfer-Encoding: 8bit X-Patchwork-Submitter: Michal Hocko X-Patchwork-Id: 10486161 Return-Path: Received: from mail.wl.linuxfoundation.org (pdx-wl-mail.web.codeaurora.org [172.30.200.125]) by pdx-korg-patchwork.web.codeaurora.org (Postfix) with ESMTP id 8667A6031B for ; Mon, 25 Jun 2018 13:32:19 +0000 (UTC) Received: from mail.wl.linuxfoundation.org (localhost [127.0.0.1]) by mail.wl.linuxfoundation.org (Postfix) with ESMTP id 7933D27F92 for ; Mon, 25 Jun 2018 13:32:19 +0000 (UTC) Received: by mail.wl.linuxfoundation.org (Postfix, from userid 486) id 6D8FF27FA9; Mon, 25 Jun 2018 13:32:19 +0000 (UTC) X-Spam-Checker-Version: SpamAssassin 3.3.1 (2010-03-16) on pdx-wl-mail.web.codeaurora.org X-Spam-Level: X-Spam-Status: No, score=-7.9 required=2.0 tests=BAYES_00, MAILING_LIST_MULTI, RCVD_IN_DNSWL_HI autolearn=unavailable version=3.3.1 Received: from vger.kernel.org (vger.kernel.org [209.132.180.67]) by mail.wl.linuxfoundation.org (Postfix) with ESMTP id 8A38F27FA1 for ; Mon, 25 Jun 2018 13:32:14 +0000 (UTC) Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S934049AbeFYNcD (ORCPT ); Mon, 25 Jun 2018 09:32:03 -0400 Received: from mx2.suse.de ([195.135.220.15]:50182 "EHLO mx2.suse.de" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S933583AbeFYNcB (ORCPT ); Mon, 25 Jun 2018 09:32:01 -0400 X-Virus-Scanned: by amavisd-new at test-mx.suse.de Received: from relay1.suse.de (charybdis-ext-too.suse.de [195.135.220.254]) by mx2.suse.de (Postfix) with ESMTP id 0AC1AAEE3; Mon, 25 Jun 2018 13:31:59 +0000 (UTC) Date: Mon, 25 Jun 2018 15:31:57 +0200 From: Michal Hocko To: Felix Kuehling Cc: Christian =?iso-8859-1?Q?K=F6nig?= , LKML , "David (ChunMing) Zhou" , Paolo Bonzini , Radim =?utf-8?B?S3LEjW3DocWZ?= , Alex Deucher , David Airlie , Jani Nikula , Joonas Lahtinen , Rodrigo Vivi , Doug Ledford , Jason Gunthorpe , Mike Marciniszyn , Dennis Dalessandro , Sudeep Dutt , Ashutosh Dixit , Dimitri Sivanich , Boris Ostrovsky , Juergen Gross , =?iso-8859-1?B?Suly9G1l?= Glisse , Andrea Arcangeli , kvm@vger.kernel.org, amd-gfx@lists.freedesktop.org, dri-devel@lists.freedesktop.org, intel-gfx@lists.freedesktop.org, linux-rdma@vger.kernel.org, xen-devel@lists.xenproject.org, linux-mm@kvack.org, David Rientjes Subject: Re: [RFC PATCH] mm, oom: distinguish blockable mode for mmu notifiers Message-ID: <20180625133157.GL28965@dhcp22.suse.cz> References: <20180622150242.16558-1-mhocko@kernel.org> <0aa9f695-5702-6704-9462-7779cbfdb3fd@amd.com> <20180622152444.GC10465@dhcp22.suse.cz> <20180625080103.GB28965@dhcp22.suse.cz> MIME-Version: 1.0 Content-Disposition: inline In-Reply-To: <20180625080103.GB28965@dhcp22.suse.cz> User-Agent: Mutt/1.9.5 (2018-04-13) Sender: kvm-owner@vger.kernel.org Precedence: bulk List-ID: X-Mailing-List: kvm@vger.kernel.org X-Virus-Scanned: ClamAV using ClamSMTP On Mon 25-06-18 10:01:03, Michal Hocko wrote: > On Fri 22-06-18 16:09:06, Felix Kuehling wrote: > > On 2018-06-22 11:24 AM, Michal Hocko wrote: > > > On Fri 22-06-18 17:13:02, Christian König wrote: > > >> Hi Michal, > > >> > > >> [Adding Felix as well] > > >> > > >> Well first of all you have a misconception why at least the AMD graphics > > >> driver need to be able to sleep in an MMU notifier: We need to sleep because > > >> we need to wait for hardware operations to finish and *NOT* because we need > > >> to wait for locks. > > >> > > >> I'm not sure if your flag now means that you generally can't sleep in MMU > > >> notifiers any more, but if that's the case at least AMD hardware will break > > >> badly. In our case the approach of waiting for a short time for the process > > >> to be reaped and then select another victim actually sounds like the right > > >> thing to do. > > > Well, I do not need to make the notifier code non blocking all the time. > > > All I need is to ensure that it won't sleep if the flag says so and > > > return -EAGAIN instead. > > > > > > So here is what I do for amdgpu: > > > > In the case of KFD we also need to take the DQM lock: > > > > amdgpu_mn_invalidate_range_start_hsa -> amdgpu_amdkfd_evict_userptr -> > > kgd2kfd_quiesce_mm -> kfd_process_evict_queues -> evict_process_queues_cpsch > > > > So we'd need to pass the blockable parameter all the way through that > > call chain. > > Thanks, I have missed that part. So I guess I will start with something > similar to intel-gfx and back off when the current range needs some > treatment. So this on top. Does it look correct? > > diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_mn.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_mn.c > index d138a526feff..e2d422b3eb0b 100644 > --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_mn.c > +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_mn.c > @@ -266,6 +266,11 @@ static int amdgpu_mn_invalidate_range_start_hsa(struct mmu_notifier *mn, > struct amdgpu_mn_node *node; > struct amdgpu_bo *bo; > > + if (!blockable) { > + amdgpu_mn_read_unlock(); > + return -EAGAIN; > + } > + > node = container_of(it, struct amdgpu_mn_node, it); > it = interval_tree_iter_next(it, start, end); Ble, just noticed that half of the change didn't get to git index... This is what I have commit c4701b36ac2802b903db3d05cf77c030fccce3a8 Author: Michal Hocko Date: Mon Jun 25 15:24:03 2018 +0200 fold me - amd gpu notifiers can sleep deeper in the callchain (evict_process_queues_cpsch on a lock and amdgpu_mn_invalidate_node on unbound timeout) make sure we bail out when we have an intersecting range for starter diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_mn.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_mn.c index d138a526feff..3399a4a927fb 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_mn.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_mn.c @@ -225,6 +225,11 @@ static int amdgpu_mn_invalidate_range_start_gfx(struct mmu_notifier *mn, while (it) { struct amdgpu_mn_node *node; + if (!blockable) { + amdgpu_mn_read_unlock(rmn); + return -EAGAIN; + } + node = container_of(it, struct amdgpu_mn_node, it); it = interval_tree_iter_next(it, start, end); @@ -266,6 +271,11 @@ static int amdgpu_mn_invalidate_range_start_hsa(struct mmu_notifier *mn, struct amdgpu_mn_node *node; struct amdgpu_bo *bo; + if (!blockable) { + amdgpu_mn_read_unlock(rmn); + return -EAGAIN; + } + node = container_of(it, struct amdgpu_mn_node, it); it = interval_tree_iter_next(it, start, end);