Hi all, This is my adaptation of Dave's last RFC. Dave's patches are unchanged except for exporting mapping_seek_hole_data to fix a compilation error. The last seven patches of the series are where I change things up. The first two patches refactor ->iomap_begin and ->iomap_end to receive a const pointer to the iterator, which reduces the argument count and makes it possible for ->iomap_begin to access the iter->private pointer. The third new patch changes the iomap pagecache write functions to enable the filesystem to set iter->private, similar to iomap_dio_rw. Having done that, I converted the xfs code to stuff the data/cow sequence counters in an iter->private object instead of bit stuffing them into the iomap->private pointer. Maybe it would've been smarter to make filesystems tell iomap about their notions of how large struct iomap objects should be (thereby enabling each fs to cram extra data along in the iomap) but that seemed like more work. So having replaced the iomap sequence counters with an explicit object, I then made the validator check the cow fork. Not sure if it's really necessary, but paranoia on my part. I /think/ it's the case that updates to the cow fork don't affect writing to the page cache, but I've wondered if the same validation rules might apply to other things (like directio writes). Lastly, I added a couple of write/writeback delay knobs so that I could write some tests that simulate race conditions and check that slow threads encounter iomap invalidation midway through an operation. I haven't gotten to analyzing Brian's eofblock truncate fixes yet, but I wanted to push this out for comments since it's now survived an overnight fstests run. NOTE: I don't have RH's original reproducer, so I have no idea if this series really fixes that corruption problem. ---- Recently a customer workload encountered a data corruption in a specific multi-threaded write operation. The workload combined racing unaligned adjacent buffered writes with low memory conditions that caused both writeback and memory reclaim to race with the writes. The result of this was random partial blocks containing zeroes instead of the correct data. The underlying problem is that iomap caches the write iomap for the duration of the write() operation, but it fails to take into account that the extent underlying the iomap can change whilst the write is in progress. The short story is that an iomap can span mutliple folios, and so under low memory writeback can be cleaning folios the write() overlaps. Whilst the overlapping data is cached in memory, this isn't a problem, but because the folios are now clean they can be reclaimed. Once reclaimed, the write() does the wrong thing when re-instantiating partial folios because the iomap no longer reflects the underlying state of the extent. e.g. it thinks the extent is unwritten, so it zeroes the partial range, when in fact the underlying extent is now written and so it should have read the data from disk. This is how we get random zero ranges in the file instead of the correct data. The gory details of the race condition can be found here: https://lore.kernel.org/linux-xfs/20220817093627.GZ3600936@xxxxxxxxxxxxxxxxxxx/ Fixing the problem has two aspects. The first aspect of the problem is ensuring that iomap can detect a stale cached iomap during a write in a race-free manner. We already do this stale iomap detection in the writeback path, so we have a mechanism for detecting that the iomap backing the data range may have changed and needs to be remapped. In the case of the write() path, we have to ensure that the iomap is validated at a point in time when the page cache is stable and cannot be reclaimed from under us. We also need to validate the extent before we start performing any modifications to the folio state or contents. Combine these two requirements together, and the only "safe" place to validate the iomap is after we have looked up and locked the folio we are going to copy the data into, but before we've performed any initialisation operations on that folio. If the iomap fails validation, we then mark it stale, unlock the folio and end the write. This effectively means a stale iomap results in a short write. Filesystems should already be able to handle this, as write operations can end short for many reasons and need to iterate through another mapping cycle to be completed. Hence the iomap changes needed to detect and handle stale iomaps during write() operations is relatively simple.... However, the assumption is that filesystems should already be able to handle write failures safely, and that's where the second (first?) part of the problem exists. That is, handling a partial write is harder than just "punching out the unused delayed allocation extent". This is because mmap() based faults can race with writes, and if they land in the delalloc region that the write allocated, then punching out the delalloc region can cause data corruption. This data corruption problem is exposed by generic/346 when iomap is converted to detect stale iomaps during write() operations. Hence write failure in the filesytems needs to handle the fact that the write() in progress doesn't necessarily own the data in the page cache over the range of the delalloc extent it just allocated. As a result, we can't just truncate the page cache over the range the write() didn't reach and punch all the delalloc extent. We have to walk the page cache over the untouched range and skip over any dirty data region in the cache in that range. Which is .... non-trivial. That is, iterating the page cache has to handle partially populated folios (i.e. block size < page size) that contain data. The data might be discontiguous within a folio. Indeed, there might be *multiple* discontiguous data regions within a single folio. And to make matters more complex, multi-page folios mean we just don't know how many sub-folio regions we might have to iterate to find all these regions. All the corner cases between the conversions and rounding between filesystem block size, folio size and multi-page folio size combined with unaligned write offsets kept breaking my brain. Eventually, I realised that if the XFS code tracked the processed write regions by byte ranges instead of fileysetm block or page cache index, we could simply use mapping_seek_hole_data() to find the start and end of each discrete data region within the range we needed to scan. SEEK_DATA finds the start of the cached data region, SEEK_HOLE finds the end of the region. THese are byte based interfaces that understand partially uptodate folio regions, and so can iterate discrete sub-folio data regions directly. This largely solved the problem of discovering the dirty regions we need to keep the delalloc extent over. Of course, now xfs/196 fails. This is a error injection test that is supposed to exercise the delalloc extent recover code that the above fixes just completely reworked. the error injection assumes that it can just truncate the page cache over the write and then punch out the delalloc extent completely. This is fundamentally broken, and only has been working by chance - the chance is that writes are page aligned and page aligned writes don't install large folios in the page cache. IOWs, with sub-folio block size, and not know what size folios are in the cache, we can't actually guarantee that we can remove the cached dirty folios from the cache via truncation, and hence the new code will not remove the delalloc extents under those dirty folios. As a result the error injection results is writing zeroes to disk rather that removing the delalloc extents from memory. I can't make this error injection to work the way it was intended, so I removed it. The code that it is supposed to exercise is now exercised every time we detect a stale iomap, so we have much better coverage of the failed write error handling than the error injection provides us with, anyway.... So, this passes fstests on 1kb and 4kb block sizes and the data corruption reproducer does not detect data corruption, so this set of fixes is /finally/ something I'd consider ready for merge. Comments and testing welcome! -Dave. Version 2: - refactor iomap code a lot, track data/cow sequence counters separately, add debugging knobs so we can test the revalidation [djwong] Version 1: - complete rework of iomap stale detection - complete rework of XFS partial delalloc write error handling. Original RFC: - https://lore.kernel.org/linux-xfs/20220921082959.1411675-1-david@xxxxxxxxxxxxx/ If you're going to start using this mess, you probably ought to just pull from my git trees, which are linked below. This is an extraordinary way to destroy everything. Enjoy! Comments and questions are, as always, welcome. --D kernel git tree: https://git.kernel.org/cgit/linux/kernel/git/djwong/xfs-linux.git/log/?h=iomap-write-races-6.2 --- fs/btrfs/inode.c | 18 +- fs/erofs/data.c | 12 + fs/erofs/zmap.c | 6 - fs/ext2/inode.c | 16 +- fs/ext4/extents.c | 5 - fs/ext4/inode.c | 38 +++- fs/f2fs/data.c | 9 + fs/fuse/dax.c | 14 +- fs/gfs2/bmap.c | 28 ++- fs/gfs2/file.c | 2 fs/hpfs/file.c | 8 + fs/iomap/buffered-io.c | 67 ++++++-- fs/iomap/iter.c | 27 ++- fs/xfs/libxfs/xfs_bmap.c | 4 fs/xfs/libxfs/xfs_errortag.h | 18 +- fs/xfs/xfs_aops.c | 12 + fs/xfs/xfs_error.c | 43 ++++- fs/xfs/xfs_error.h | 22 ++ fs/xfs/xfs_file.c | 5 - fs/xfs/xfs_iomap.c | 371 ++++++++++++++++++++++++++++++++++-------- fs/xfs/xfs_iomap.h | 7 + fs/xfs/xfs_reflink.c | 3 fs/zonefs/super.c | 27 ++- include/linux/iomap.h | 34 +++- mm/filemap.c | 1 25 files changed, 610 insertions(+), 187 deletions(-)