From mboxrd@z Thu Jan 1 00:00:00 1970 From: Andrew Morton Subject: Re: [patch 9/9] mm: fix pagecache write deadlocks Date: Sun, 4 Feb 2007 01:44:45 -0800 Message-ID: <20070204014445.88e6c8c7.akpm@linux-foundation.org> References: <20070204063707.23659.20741.sendpatchset@linux.site> <20070204063833.23659.55105.sendpatchset@linux.site> Mime-Version: 1.0 Content-Type: text/plain; charset=US-ASCII Content-Transfer-Encoding: 7bit Cc: Linux Kernel , Linux Filesystems , Linux Memory Management To: Nick Piggin Return-path: Received: from smtp.osdl.org ([65.172.181.24]:34854 "EHLO smtp.osdl.org" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1752196AbXBDJoy (ORCPT ); Sun, 4 Feb 2007 04:44:54 -0500 In-Reply-To: <20070204063833.23659.55105.sendpatchset@linux.site> Sender: linux-fsdevel-owner@vger.kernel.org List-Id: linux-fsdevel.vger.kernel.org On Sun, 4 Feb 2007 09:51:07 +0100 (CET) Nick Piggin wrote: > 2. If we find the destination page is non uptodate, unlock it (this could be > made slightly more optimal), then find and pin the source page with > get_user_pages. Relock the destination page and continue with the copy. > However, instead of a usercopy (which might take a fault), copy the data > via the kernel address space. argh. We just can't go adding all this gunk into the write() path. mmap_sem, a full pte-walk, taking of pte-page locks, etc. For every page. Even single-process write() will suffer, let along multithreaded stuff, where mmap_sem contention may be the bigger problem. I was going to do some quick measurements of this, but the code oopses on power4 (http://userweb.kernel.org/~akpm/s5000402.jpg) There's a build error in filemap_xip.c btw. We need to think different. What happened to the idea of doing an atomic copy into the non-uptodate page and handling it somehow? Another option might be to effectively pin the whole mm during the copy: down_read(¤t->mm->unpaging_lock); get_user(addr); /* Fault the page in */ ... copy_from_user() up_read(¤t->mm->unpaging_lock); then, anyone who wants to unmap pages from this mm requires write_lock(unpaging_lock). So we know the results of that get_user() cannot be undone. Or perhaps something like this can be done on a per-vma basis. Just something to tell the VM "hey, you're not allowed to unmap this page right now"?