From mboxrd@z Thu Jan 1 00:00:00 1970 From: Andy Lutomirski Subject: [RFC PATCH 0/4] Rework mtime and ctime updates on mmaped writes Date: Thu, 20 Dec 2012 15:10:08 -0800 Message-ID: References: Cc: Dave Chinner , Al Viro , Jan Kara , Andy Lutomirski To: linux-kernel@vger.kernel.org, Linux FS Devel Return-path: Received: from mail-pb0-f45.google.com ([209.85.160.45]:48401 "EHLO mail-pb0-f45.google.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1751800Ab2LTXQS (ORCPT ); Thu, 20 Dec 2012 18:16:18 -0500 Received: by mail-pb0-f45.google.com with SMTP id mc8so2291725pbc.4 for ; Thu, 20 Dec 2012 15:16:17 -0800 (PST) In-Reply-To: Sender: linux-fsdevel-owner@vger.kernel.org List-ID: Writes via mmap currently update mtime and ctime in ->page_mkwrite. This is unfortunate from a performance and a correctness point of view. The file times should be updated after writes, not before (so that every write eventually results in a fresh timestamp). More importantly (for me), ->page_mkwrite is called periodically even on mlocked pages, and some filesystems can sleep in mark_inode_dirty. This patchset attempts to fix both issues at once. It adds a new address_space flag AS_CMTIME that is set atomically whenever the system transfers a pte dirty bit to a struct page backed by the address_space. This can happen with various locks held and when low on memory. Later on, whenever syncing an inode (which happens indirectly in msync) or whenever a vma is torn down, if AS_CMTIME is set, then the file times are updated. This happens in a context from which (I think) it's safe to dirty inodes. One nice property of this approach is that it requires no fs-specific work. It's actually quite a bit simpler than I expected. I've tested this, and mtime and ctime are updated on munmap, exit, MS_SYNC, and fsync after writing via mmap. The times are also updated 30 seconds after writing, all by themselves :) Lockdep has no complaints. NB: I am not at all an expert in anything fs or pagecache related. Please help me find things that may be wrong with these patches. Andy Lutomirski (4): mm: Explicitly track when the page dirty bit is transferred from a pte mm: Update file times when inodes are written after mmaped writes Remove file_update_time from all mkwrite paths ext4: Fix an incorrect comment about i_mutex fs/9p/vfs_file.c | 3 --- fs/btrfs/inode.c | 4 +--- fs/buffer.c | 6 ------ fs/ceph/addr.c | 3 --- fs/ext4/fsync.c | 2 -- fs/ext4/inode.c | 1 - fs/gfs2/file.c | 3 --- fs/inode.c | 37 +++++++++++++++++++++++++++++++++++++ fs/nilfs2/file.c | 1 - fs/sysfs/bin.c | 2 -- include/linux/fs.h | 1 + include/linux/mm.h | 1 + include/linux/pagemap.h | 3 +++ mm/filemap.c | 1 - mm/memory-failure.c | 4 +--- mm/memory.c | 5 +---- mm/mmap.c | 4 ++++ mm/page-writeback.c | 42 ++++++++++++++++++++++++++++++++++++++++-- mm/rmap.c | 9 ++++++--- 19 files changed, 95 insertions(+), 37 deletions(-) -- 1.7.11.7