From mboxrd@z Thu Jan 1 00:00:00 1970 From: Al Viro Subject: Re: [RFC] revoke(2) and generic handling of things like remove_proc_entry() Date: Sat, 6 Apr 2013 06:00:49 +0100 Message-ID: <20130406050049.GI4068@ZenIV.linux.org.uk> References: <20130405042932.GB4068@ZenIV.linux.org.uk> Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii Cc: linux-kernel@vger.kernel.org, Linus Torvalds , Greg Kroah-Hartman , Rik van Riel , Andrew Morton , Alexey Dobriyan , Takashi Iwai To: linux-fsdevel@vger.kernel.org Return-path: Received: from zeniv.linux.org.uk ([195.92.253.2]:54460 "EHLO ZenIV.linux.org.uk" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1752054Ab3DFFAx (ORCPT ); Sat, 6 Apr 2013 01:00:53 -0400 Content-Disposition: inline In-Reply-To: <20130405042932.GB4068@ZenIV.linux.org.uk> Sender: linux-fsdevel-owner@vger.kernel.org List-ID: On Fri, Apr 05, 2013 at 05:29:32AM +0100, Al Viro wrote: > 4) nasty semantics issue - mmap() vs. revoke (of any sort, including > remove_proc_entry(), etc.). Suppose a revokable file had been mmapped; > now it's going away. What should we do to its VMAs? Right now sysfs > and procfs get away with that, but only because there's only one thing > that has ->mmap() there - /proc/bus/pci and sysfs equivalents. I've > no idea how does pci_mmap_page_range() interact with PCI hotplug (and > I'm not at all sure that whatever it does isn't racy wrt device removal), > but I suspect that it strongly depends on lack of ->fault() for those > VMAs, which makes killing all PTEs pointing to pages in question enough. > How generic do we want to make it? Anybody wanting to add more files > that could be mmapped in procfs/sysfs/debugfs deserves to be hurt, but > if we start playing with revoke(2), restriction might become inconvenient. > I'm not sure what kind of behaviour do we want there - *BSD at least > used to have revoke(2) only for character devices that had no mmap()... Actually, after looking at what sysfs does... We might get away with the following * new vma flag - VM_REVOKABLE; set by mmap() if ->f_revoke is non-NULL. We are short on spare bits there, but there still are some... * start_using_vma(vma) that checks the presence of that flag, returns true if it's absent and __start_using(vma->vm_file->f_revoke) otherwise; a matching stop_using_vma(vma) as well. * surround vma method calls with start_using_vma/stop_using_vma, similar to file ones. Do what fs/sysfs/bin.c wrappers do for revoked ones - VM_FAULT_SIGBUS for ->fault() and ->page_mkwrite(), -EINVAL for ->access() and ->set_policy(), vma->vm_policy for ->get_policy(), 0 for ->migrate(), "do nothing" for ->open() (and I'm not at all sure that this one is correct), hell knows what for ->close(). Note that the *only* instance with ->open and without ->close is sysfs pile of wrappers itself... Hell knows... We have few enough call sites for ->vm_op->foo() to make it feasible and overhead would be trivial. OTOH, I'm not sure what's the right behaviour for mmap of something like drm after revoke(2) - leaving writable pages there looks wrong... BTW, snd_card_disconnect() doesn't do anything to existing mappings; smells like a bug, and there we do have ones with non-trivial ->mmap(). Could ALSA folks comment? One note about the mockup implementation upthread - __release_revoke() should suck in a bit more than just ->release() - turning fasync off should also go there.