Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1946932AbbGaUu6 (ORCPT ); Fri, 31 Jul 2015 16:50:58 -0400 Received: from nautica.notk.org ([91.121.71.147]:48316 "EHLO nautica.notk.org" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1754518AbbGaUuy (ORCPT ); Fri, 31 Jul 2015 16:50:54 -0400 Date: Fri, 31 Jul 2015 22:50:36 +0200 From: Dominique Martinet To: Hugh Dickins Cc: Linus Torvalds , "J. Bruce Fields" , Dominique Martinet , Al Viro , Linux Kernel Mailing List , linux-fsdevel Subject: Re: v4.2-rc dcache regression, probably 75a6f82a0d10 Message-ID: <20150731205036.GA3752@nautica> References: MIME-Version: 1.0 Content-Type: text/plain; charset=utf-8 Content-Disposition: inline In-Reply-To: User-Agent: Mutt/1.5.21 (2010-09-15) Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org Content-Length: 1762 Lines: 42 Hugh Dickins wrote on Fri, Jul 31, 2015: > It will indeed be weird and odd if it confirms that DCACHE_DISCONNECTED > revert is good. I agree that Dominique's 4bf46a272647 seems now more > likely, if still unlikely; but that was included in v4.1, and I saw > no problem with v4.1 once the rmap_walk() skip was fixed. I think it could, actually, and that neither commits are actually bad -- just that they affect timing enough to raise an issue between d_delete (I guess?) and link_path_walk (see last mail in other thread[1]) It's probably an old race that was very hard to hit because of cache coherency. Basically, before the wmb/rmb, the dentry was always updated closely to its flags, so the other CPU would "usually" get both updates at the same time; the barriers make it so the updates are split and it's possible to get it, and would explain why I could pick 4bf46a2726 as "the one" I'm not sure why the problem wouldn't arise on tmpfs though. Hugh, could you try the reproducer I gave in the other thread[2] on both filesystems maybe? I need to let the thing run for a while, might need to tune params as well. I was trying to fine tune cpu affinity with less threads but it's not getting anywhere. I'll also check if it's getting even easier to reproduce with 75a6f82a0d10 (or a recent kernel), who knows... How fast do you hit the bug with the commit? Thanks, -- Dominique [1] https://marc.info/?l=linux-fsdevel&m=143835651005259&w=2 [2] https://marc.info/?l=linux-fsdevel&m=143825706609188&w=2 -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/