Return-Path: linux-nfs-owner@vger.kernel.org Received: from cantor2.suse.de ([195.135.220.15]:53234 "EHLO mx2.suse.de" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1750813AbaIPFca (ORCPT ); Tue, 16 Sep 2014 01:32:30 -0400 From: NeilBrown To: Peter Zijlstra , Andrew Morton , Trond Myklebust , Ingo Molnar Date: Tue, 16 Sep 2014 15:31:34 +1000 Subject: [PATCH 0/4] Remove possible deadlocks in nfs_release_page() Cc: linux-fsdevel@vger.kernel.org, linux-mm@kvack.org, linux-nfs@vger.kernel.org, linux-kernel@vger.kernel.org, Jeff Layton Message-ID: <20140916051911.22257.24658.stgit@notabene.brown> MIME-Version: 1.0 Content-Type: text/plain; charset="utf-8" Sender: linux-nfs-owner@vger.kernel.org List-ID: Because nfs_release_page() submits a 'COMMIT' nfs request and waits for it to complete, and does this during memory reclaim, it is susceptible to deadlocks if memory allocation happens anywhere in sending the COMMIT message. If the NFS server is on the same host (i.e. loop-back NFS), then any memory allocations in the NFS server can also cause deadlocks. nfs_release_page() already has some code to avoid deadlocks in some circumstances, but these are not sufficient for loopback NFS. This patch set changes the approach to deadlock avoidance. Rather than detecting cases that could deadlock and avoiding the COMMIT, it always tries the COMMIT, but only waits a short time (1 second). This avoid any deadlock possibility at the expense of not waiting longer than 1 second even if no deadlock is pending. nfs_release_page() does not *need* to wait longer - all callers that matter handle a failure gracefully - they move on to other pages. This set: - adds some "_timeout()" functions to "wait_on_bit". Only a wait_on_page version is actually used. - exports page wake_up support. NFS knows that the COMMIT is complete when PG_private is clear. So nfs_release_page will use wait_on_page_bit_killable_timeout to wait for the bit to clear, and needs access to wake_up_page() - changes nfs_release_page() to use wait_on_page_bit_killable_timeout() - removes the other deadlock avoidance mechanisms from nfs_release_page, so that PF_FSTRANS is again only used by XFS. As such, it needs buy-in from sched people, mm people, and NFS people. Assuming I get that buy-in, suggests for how these patches can flow into mainline would be appreciated ... I daren't hope they can all go in through one tree.... Thanks, NeilBrown --- NeilBrown (4): SCHED: add some "wait..on_bit...timeout()" interfaces. MM: export page_wakeup functions NFS: avoid deadlocks with loop-back mounted NFS filesystems. NFS/SUNRPC: Remove other deadlock-avoidance mechanisms in nfs_release_page() fs/nfs/file.c | 22 ++++++++++++---------- fs/nfs/write.c | 2 ++ include/linux/pagemap.h | 12 ++++++++++-- include/linux/wait.h | 5 ++++- kernel/sched/wait.c | 36 ++++++++++++++++++++++++++++++++++++ mm/filemap.c | 21 +++++++++++++++------ net/sunrpc/sched.c | 2 -- net/sunrpc/xprtrdma/transport.c | 2 -- net/sunrpc/xprtsock.c | 10 ---------- 9 files changed, 79 insertions(+), 33 deletions(-) -- Signature