On Mon, Dec 18, 2023 at 02:12:00PM -0800, Darrick J. Wong wrote: > [cc willy and fsdevel] > > On Tue, Dec 19, 2023 at 07:21:37AM +1100, Dave Chinner wrote: > > [cc Jan Kara] > > > > On Mon, Dec 18, 2023 at 07:29:34AM -0500, Brian Foster wrote: > > > On Sat, Dec 16, 2023 at 09:39:51AM -0800, Darrick J. Wong wrote: > > > > On Sat, Dec 16, 2023 at 07:55:59PM +0800, Long Li wrote: > > > > > While fsstress + drop cache test, we get following warning: > > > > > > > > > > ------------[ cut here ]------------ > > > > > WARNING: CPU: 2 PID: 1003 at fs/iomap/buffered-io.c:1182 iomap_file_buffered_write_punch_delalloc+0x691/0x730 > > > > > Modules linked in: > > > > > CPU: 2 PID: 1003 Comm: fsstress Not tainted 6.7.0-rc5-06945-g3ba9b31d6bf3-dirty #256 > > > > > Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS ?-20190727_073836-buildvm-ppc64le-16.ppc.fedoraproject.org-3.fc31 04/01/2014 > > > > > RIP: 0010:iomap_file_buffered_write_punch_delalloc+0x691/0x730 > > > > > Code: d1 0b 01 0f 0b 48 83 05 14 a2 d1 0b 01 48 89 05 35 a1 d1 0b 49 39 ec 0f 8c 09 fb ff ff e9 b6 fd ff ff 48 83 05 df a1 d1 0b 01 <0f> 0b 48 83 05 dd a1 d1 0b 01 48 39 6c 24 10 7c c0 48 89 05 07 a1 > > > > > RSP: 0018:ffffc900005b7b08 EFLAGS: 00010202 > > > > > RAX: 0000000000000001 RBX: ffff888102363d40 RCX: 0000000000000001 > > > > > RDX: 0000000000000000 RSI: 0000000000000000 RDI: ffff888108080000 > > > > > RBP: 0000000000050000 R08: ffff888108084eb8 R09: ffff888108084eb8 > > > > > R10: 000000000000005c R11: 0000000000000059 R12: 0000000000050000 > > > > > R13: ffffffff8c978ef0 R14: 0000000000050000 R15: 000000000005a000 > > > > > FS: 00007efc04c63b40(0000) GS:ffff88813bd00000(0000) knlGS:0000000000000000 > > > > > CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 > > > > > CR2: 00007efc0375c000 CR3: 0000000105a4d000 CR4: 00000000000006f0 > > > > > DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 > > > > > DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 > > > > > Call Trace: > > > > > <TASK> > > > > > xfs_buffered_write_iomap_end+0x40/0xb0 > > > > > iomap_iter+0x8e/0x5f0 > > > > > iomap_file_buffered_write+0xa4/0x460 > > > > > xfs_file_buffered_write+0x156/0x3d0 > > > > > xfs_file_write_iter+0xb2/0x1c0 > > > > > do_iter_readv_writev+0x19b/0x1f0 > > > > > vfs_writev+0x114/0x4f0 > > > > > do_writev+0x7f/0x1c0 > > > > > __x64_sys_writev+0x24/0x30 > > > > > do_syscall_64+0x3f/0xe0 > > > > > entry_SYSCALL_64_after_hwframe+0x62/0x6a > > > > > RIP: 0033:0x7efc03b06610 > > > > > Code: 73 01 c3 48 8b 0d 78 88 2c 00 f7 d8 64 89 01 48 83 c8 ff c3 66 0f 1f 44 00 00 83 3d d9 e0 2c 00 00 75 10 b8 14 00 00 00 0f 05 <48> 3d 01 f0 ff ff 73 31 c3 48 83 ec 08 e8 5e 8e 01 00 48 89 04 24 > > > > > RSP: 002b:00007ffdf8f426d8 EFLAGS: 00000246 ORIG_RAX: 0000000000000014 > > > > > RAX: ffffffffffffffda RBX: 000000000000007a RCX: 00007efc03b06610 > > > > > RDX: 00000000000002c4 RSI: 00000000012f5580 RDI: 0000000000000003 > > > > > RBP: 0000000000000003 R08: 00000000012f53a0 R09: 0000000000000077 > > > > > R10: 000000000000007c R11: 0000000000000246 R12: 00000000000002c4 > > > > > R13: 00000000012dba50 R14: 00000000012f5580 R15: 0000000000000094 > > > > > > > > > > The warning occurred in the following code of iomap_write_delalloc_release(). > > > > > After analyzing vmcore, I found that the reason for the warning is that > > > > > data_end was equal to start_byte. > > > > > > > > > > WARN_ON_ONCE(data_end <= start_byte); > > > > > > > > > > If some delay is added between seeking for data and seeking for hole > > > > > in iomap_write_delalloc_release(), the problem can be reproduced quickly. > > > > > The root cause of the problem is that clean data page was dropped between > > > > > two seeking in the page cache. As a result, data_end may be equal to > > > > > start_byte. > > > > > > > > > > buffered write drop cache > > > > > --------------------------- --------------------------- > > > > > xfs_buffered_write_iomap_end > > > > > iomap_file_buffered_write_punch_delalloc > > > > > iomap_write_delalloc_release > > > > > start_byte = mapping_seek_hole_data(SEEK_DATA) > > > > > > > > > > drop_pagecache_sb > > > > > invalidate_mapping_pages > > > > > mapping_try_invalidate > > > > > mapping_evict_folio > > > > > remove_mapping > > > > > > > > > > data_end = mapping_seek_hole_data(SEEK_HOLE) > > > > > WARN_ON_ONCE(data_end <= start_byte) > > > > > > > > > > In my investigation, I found that clean data pages was alloced and added > > > > > to page cache when reading the file's hole. After that, while buffered > > > > > write and goes into delalloc release, we seek for data, it will find > > > > > the start offset of the clean data pages. If the clean data pages was > > > > > dropped, when we seek for hole, it will find the same offset as the > > > > > previous seek. > > > > > > > > iomap_write_delalloc_release holds the invalidation lock, shouldn't that > > > > be sufficient to prevent folios from being dropped? > > > > > > > > > > I believe it's the other way around. The invalidate lock prevents new > > > folios from being added (i.e. to serialize an invalidation sequence). > > > IIRC it won't necessarily prevent folios from going away. > > > > It also serialises operations doing invalidations against other > > operations doing invalidations (e.g. truncate, hole punch, direct > > IO, etc) that require the invalidation lock to be held exclusively. > > > > Having looked a little deeper, it appears we have a large number of > > invalidation paths that don't actually hold the invalidation lock at > > all. drop_pagecache_sb() is just one of them. Is that a bug? > > Clearly I think so. ;) +1 > > If not, then this code needs to prevent any mapping tree change from > > occurring whilst it is walking the range of the write that needs to > > be scanned. The write has partially failed at this point, and we're > > trying to clean up the mess the partial write has left behind. We > > really need exclusive access to that range of the mapping whilst > > cleanup is done. > > > > IOWs, if the invalidation lock is not sufficient to serialise > > against other invalidations being run concurrently, what mapping > > lock can we hold whilst we call into the filesystem, take sleeping > > FS locks and modify the extent tree that will serialise against > > other invalidations? > > I don't know of any other locks that would provide that; also, I thought > invalidate_lock (fka MMAPLOCK_EXCL) was what xfs was supposed to take to > serialize against folios getting ripped out of the mapping and page > faults. Well, originally it was just fs operations vs page faults, but it also had the (XFS internal) property of serialise operations against other fs-driven invalidations and breaking layout leases (e.g. xfs_break_dax_layouts(), xfs_ilock2_io_mmap() from xfs_reflink_remap_prep(), etc) Details are a bit fuzzy, but ISTR that I wrote this iomap code with the assumption that holding the invalidate_lock also serialises against other invalidation operations - it was originally based on the XFS_MMAPLOCK_EXCL being held, so that likely bled over. > Though now that I think about this further -- have dropcaches and xfs > post-write error delalloc punching always lacked coordination? I don't think drop_caches has proper serialisation against anything. If we go back to before we have background inode inactivation, and drop caches could deadlock against filesystem freeze because it calls the shrinker which runs transactions whilst holding the s_umount lock so thaw can't unfreeze the filesystem. Any filesystem that does an operation in ->evict or ->destroy_inode as called from memory reclaim context that can be blocked by a frozen fs state can trip over this. In reality, drop_caches was only supposed to be a debug-only interface and so it's never really been designed/coded to be properly safe against anything. Hence I think this is drop_caches bug more than it is a iomap bug... > I think > for most pagecache operations that old "lock and recheck mapping/offset" > strategy works ok; it's just here where we might've been touching > multiple folios that things go awry...? That old mechanism never really worked reliably. Just look at the constant stream of "we broke truncate" bugs that we used to have to deal with. They've dropped right off now that most filesystems use the invalidate lock for truncate serialisation rather than the old "lock, check mapping, check i_size, unlock, retry" loop that people were also struggling to get right. -Dave -- Dave Chinner david@xxxxxxxxxxxxx
