Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1751579AbcKFGp6 (ORCPT ); Sun, 6 Nov 2016 01:45:58 -0500 Received: from mail-wm0-f43.google.com ([74.125.82.43]:35959 "EHLO mail-wm0-f43.google.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1751375AbcKFGp5 (ORCPT ); Sun, 6 Nov 2016 01:45:57 -0500 MIME-Version: 1.0 In-Reply-To: <20161105213411.GA32353@quack2.suse.cz> References: <20161102220851.GA1839@veci.piliscsaba.szeredi.hu> <20161105213411.GA32353@quack2.suse.cz> From: Amir Goldstein Date: Sun, 6 Nov 2016 08:45:54 +0200 Message-ID: Subject: Re: fsnotify_mark_srcu wtf? To: Jan Kara Cc: Miklos Szeredi , Eric Paris , linux-fsdevel , linux-kernel Content-Type: text/plain; charset=UTF-8 Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org Content-Length: 2381 Lines: 53 On Sat, Nov 5, 2016 at 11:34 PM, Jan Kara wrote: > On Wed 02-11-16 23:09:26, Miklos Szeredi wrote: >> We've got a report where a fanotify daemon that implements permission checks >> screws up and doesn't send a reply. This then causes widespread hangs due to >> fsnotify_mark_srcu read side lock being held and thus causing synchronize_srcu() >> called from e.g. inotify_release()-> fsnotify_destroy_group()-> >> fsnotify_mark_destroy_list() to block. > > Yes. But if a program implementing permission checks does not reply, your > system is likely hosed anyway. We can only try to somewhat limit the > damage... > That was my initial thought as well, but at least with the sample code Miklos sent the only thing that gets hosed is the one process watching that one file. You could think of a use case of fanotify being used to watch over files in a specific user directory, where the damage on the entire system should/could be limited. No? >> Below program demonstrates the issue. It should output a single line: >> >> close(inotify_fd): success >> >> Instead it outputs nothing, which means that close(inotify_fd) got blocked by >> the waiting permission event. >> >> Wouldn't making the srcu per-group fix this? Would that be too expensive? > > Per-group would be IMHO too expensive. You can have lots of groups and I'm > not sure srcu would scale to that. Furthermore the SRCU protects the list > of groups that need to get notification so it would not even be easily > possible. Also Amir's solution is buggy - I'll comment on that as a reply > to his patch. I'll try to find something to improve the situation but so > far I have no good idea... > Yes, very much buggy indeed :/ Anyway, the reason I drafted it quickly was to highlight the fact that the marks only need to live to the point of decision whether or not the event should be sent to the group and afterwards, its sufficient to grab the group reference, without having impact on the entire system. Yet another possible ugly (but less buggy) solution would be to iterate all marks under SRCU read protection. If any group is about to block (either by suggested return value EAGAIN or another by using a new op should_handle_event_deferred), defer event handling to post marks iteration, by keeping a few group references on stack. But hopefully, you'll find some less ugly solution. Amir.