From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: from mail-qk0-f199.google.com (mail-qk0-f199.google.com [209.85.220.199]) by kanga.kvack.org (Postfix) with ESMTP id E598F6B0003 for ; Sat, 17 Feb 2018 20:06:49 -0500 (EST) Received: by mail-qk0-f199.google.com with SMTP id b67so6147695qkh.5 for ; Sat, 17 Feb 2018 17:06:49 -0800 (PST) Received: from mail-sor-f65.google.com (mail-sor-f65.google.com. [209.85.220.65]) by mx.google.com with SMTPS id h89sor2107505qtd.103.2018.02.17.17.06.48 for (Google Transport Security); Sat, 17 Feb 2018 17:06:48 -0800 (PST) MIME-Version: 1.0 In-Reply-To: <20180216153823.ad74f1d2c157adc67ed2c970@linux-foundation.org> References: <20180213014220.2464-1-ying.huang@intel.com> <20180213154123.9f4ef9e406ea8365ca46d9c5@linux-foundation.org> <87fu64jthz.fsf@yhuang-dev.intel.com> <20180216153823.ad74f1d2c157adc67ed2c970@linux-foundation.org> From: huang ying Date: Sun, 18 Feb 2018 09:06:47 +0800 Message-ID: Subject: Re: [PATCH -mm -v5 RESEND] mm, swap: Fix race between swapoff and some swap operations Content-Type: text/plain; charset="UTF-8" Sender: owner-linux-mm@kvack.org List-ID: To: Andrew Morton Cc: "Huang, Ying" , linux-mm@kvack.org, LKML , Hugh Dickins , "Paul E . McKenney" , Minchan Kim , Johannes Weiner , Tim Chen , Shaohua Li , Mel Gorman , jglisse@redhat.com, Michal Hocko , Andrea Arcangeli , David Rientjes , Rik van Riel , Jan Kara , Dave Jiang , Aaron Lu On Sat, Feb 17, 2018 at 7:38 AM, Andrew Morton wrote: > On Wed, 14 Feb 2018 08:38:00 +0800 "Huang\, Ying" wrote: > >> Andrew Morton writes: >> >> > On Tue, 13 Feb 2018 09:42:20 +0800 "Huang, Ying" wrote: >> > >> >> From: Huang Ying >> >> >> >> When the swapin is performed, after getting the swap entry information >> >> from the page table, system will swap in the swap entry, without any >> >> lock held to prevent the swap device from being swapoff. This may >> >> cause the race like below, >> > >> > Sigh. In terms of putting all the work into the swapoff path and >> > avoiding overheads in the hot paths, I guess this is about as good as >> > it will get. >> > >> > It's a very low-priority fix so I'd prefer to keep the patch in -mm >> > until Hugh has had an opportunity to think about it. >> > >> >> ... >> >> >> >> +/* >> >> + * Check whether swap entry is valid in the swap device. If so, >> >> + * return pointer to swap_info_struct, and keep the swap entry valid >> >> + * via preventing the swap device from being swapoff, until >> >> + * put_swap_device() is called. Otherwise return NULL. >> >> + */ >> >> +struct swap_info_struct *get_swap_device(swp_entry_t entry) >> >> +{ >> >> + struct swap_info_struct *si; >> >> + unsigned long type, offset; >> >> + >> >> + if (!entry.val) >> >> + goto out; >> >> + type = swp_type(entry); >> >> + if (type >= nr_swapfiles) >> >> + goto bad_nofile; >> >> + si = swap_info[type]; >> >> + >> >> + preempt_disable(); >> > >> > This preempt_disable() is later than I'd expect. If a well-timed race >> > occurs, `si' could now be pointing at a defunct entry. If that >> > well-timed race include a swapoff AND a swapon, `si' could be pointing >> > at the info for a new device? >> >> struct swap_info_struct pointed to by swap_info[] will never be freed. >> During swapoff, we only free the memory pointed to by the fields of >> struct swap_info_struct. And when swapon, we will always reuse >> swap_info[type] if it's not NULL. So it should be safe to dereference >> swap_info[type] with preemption enabled. > > That's my point. If there's a race window during which there is a > parallel swapoff+swapon, this swap_info_struct may now be in use for a > different device? Yes. It's possible. And the caller of get_swap_device() can live with it if the swap_info_struct has been fully initialized. For example, for the race in the patch description, do_swap_page swapin_readahead __read_swap_cache_async swapcache_prepare __swap_duplicate in __swap_duplicate(), it's possible that the swap device returned by get_swap_device() is different from the swap device when __swap_duplicate() call get_swap_device(). But the struct_info_struct has been fully initialized, so __swap_duplicate() can reference si->swap_map[] safely. And we will check si->swap_map[] before any further operation. Even if the swap entry is swapped out again for the new swap device, we will check the page table again in do_swap_page(). So there is no functionality problem. Best Regards, Huang, Ying -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to majordomo@kvack.org. For more info on Linux MM, see: http://www.linux-mm.org/ . Don't email: email@kvack.org