On Wed, Nov 16, 2022 at 08:57:19AM -0500, Brian Foster wrote: > On Tue, Nov 15, 2022 at 12:30:39PM +1100, Dave Chinner wrote: > > From: Dave Chinner <dchinner@xxxxxxxxxx> > > > ... > > > > Signed-off-by: Dave Chinner <dchinner@xxxxxxxxxx> > > --- > > fs/xfs/xfs_iomap.c | 151 ++++++++++++++++++++++++++++++++++++++++++--- > > 1 file changed, 141 insertions(+), 10 deletions(-) > > > > diff --git a/fs/xfs/xfs_iomap.c b/fs/xfs/xfs_iomap.c > > index 7bb55dbc19d3..2d48fcc7bd6f 100644 > > --- a/fs/xfs/xfs_iomap.c > > +++ b/fs/xfs/xfs_iomap.c > > @@ -1134,6 +1134,146 @@ xfs_buffered_write_delalloc_punch( > > end_fsb - start_fsb); > > } > > > ... > > +/* > > + * Punch out all the delalloc blocks in the range given except for those that > > + * have dirty data still pending in the page cache - those are going to be > > + * written and so must still retain the delalloc backing for writeback. > > + * > > + * As we are scanning the page cache for data, we don't need to reimplement the > > + * wheel - mapping_seek_hole_data() does exactly what we need to identify the > > + * start and end of data ranges correctly even for sub-folio block sizes. This > > + * byte range based iteration is especially convenient because it means we don't > > + * have to care about variable size folios, nor where the start or end of the > > + * data range lies within a folio, if they lie within the same folio or even if > > + * there are multiple discontiguous data ranges within the folio. > > + */ > > +static int > > +xfs_buffered_write_delalloc_release( > > + struct inode *inode, > > + loff_t start_byte, > > + loff_t end_byte) > > +{ > > + loff_t punch_start_byte = start_byte; > > + int error = 0; > > + > > + /* > > + * Lock the mapping to avoid races with page faults re-instantiating > > + * folios and dirtying them via ->page_mkwrite whilst we walk the > > + * cache and perform delalloc extent removal. Failing to do this can > > + * leave dirty pages with no space reservation in the cache. > > + */ > > + filemap_invalidate_lock(inode->i_mapping); > > + while (start_byte < end_byte) { > > + loff_t data_end; > > + > > + start_byte = mapping_seek_hole_data(inode->i_mapping, > > + start_byte, end_byte, SEEK_DATA); > > FWIW, the fact that mapping seek data is based on uptodate status means > that seek behavior can change based on prior reads. Yup. It should be obvious that any page cache scan based algorithm will change based on changing page cache residency. > For example, see how > seek hole/data presents reads of unwritten ranges as data [1]. The same > thing isn't observable for holes because iomap doesn't check the mapping > in that case, but underlying iop state is the same and that is what this > code is looking at. Well, yes. That's the fundamental, underlying issue that this patchset is addressing for the write() operation: that the page cache contents and the underlying filesystem extent map are not guaranteed to be coherent and can be changed independently of each other. The whole problem with looking exclusively at filesystem level extent state (and hence FIEMAP) is that the extent state doesn't tell us whether the is uncommitted data over the range of the extent in the page cache. The filesystem extent state and page cache data *can't be coherent* in a writeback caching environment. This is the fundamental difference between what the filesystem extent map tells us (FIEMAP) and what querying the page cache tells us (SEEK_DATA/SEEK_HOLE). This is also the underlying problem with iomap_truncate_page() - it fails to query the page cache for data over unwritten extents, so fails to zero the post-EOF part of dirty folios over unwritten extents and so it all goes wrong... > The filtering being done here means we essentially only care about dirty > pages backed by delalloc blocks. That means if you get here with a dirty > page and the portion of the page affected by this failed write is > uptodate, this won't punch an underlying delalloc block even though > nothing else may have written to it in the meantime. Hmmm. IOMAP_F_NEW implies that the newly allocated delalloc iomap will not span ranges that have pre-existing *dirty* data in the page cache. Those *must* already have (del)allocated extents, hence the iomap for the newly allocated delalloc extent will always end before pre-existing dirty data in the page cache starts. Hence the seek scan range over an IOMAP_F_NEW IOMAP_DELALLOC map precludes stepping into ranges that have pre-existing cached dirty data. We also can't get a racing write() to the same range right now because this is all under IOLOCK_EXCL, hence we only ever see dirty folios as a result of race with page faults. page faults zero the entire folio they insert into the page cache and iomap_folio_mkwrite_iter() asserts that the entire folio is marked up to date. Hence if we find a dirty folio outside the range the write() dirtied, we are guaranteed that the entire dirty folio is up to date.... Yes, there can be pre-existing *clean* folios (and clean partially up to date folios) in the page cache, but we won't have dirty partially up to date pages in the middle of the range we are scanning. Hence we only need to care about the edge cases (folios that overlap start and ends). We skip the partially written start block, and we always punch up to the end block if it is different from the last block we punched up to. If the end of the data spans into a dirty folio, we know that dirty range is up to date because the seek scan only returns ranges that are up to date. Hence we don't punch those partial blocks out.... Regardless, let's assume we have a racing write that has partially updated and dirtied a folio (because we've moved to XFS_IOLOCK_SHARED locking for writes). This case is already handled by the mapping_seek_hole_data() based iteration. That is, the mapping_seek_hole_data() loop provides us with *discrete ranges of up to date data* that are independent of folio size, up-to-date range granularity, dirty range tracking, filesystem block size, etc. Hence if the next discrete range we discover is in the same dirty folio as the previous discrete range of up to date data, we know we have a sub-folio sized hole in the data that is not up to date. Because there is no data over this range, we have to punch out the underlying delalloc extent over that range. IOWs, the dirty state of the folio and/or the granularity of the dirty range tracking is irrelevant here - we know there was no data in the cache (dlean or dirty) over this range because it is discontiguous with the previous range of data returned. IOWs, if we have this "up to date" map on a dirty folio like this: Data +-------+UUUUUUU+-------+UUUUUUU+-------+ Extent map +DDDDDDD+DDDDDDD+DDDDDDD+DDDDDDD+DDDDDDD+ Then the unrolled iteration and punching we do would look like this: First iteration of the range: punch_start: V +-------+UUUUUUU+-------+UUUUUUU+-------+ SEEK_DATA: V +-------+UUUUUUU+-------+UUUUUUU+-------+ SEEK_HOLE: ^ Data range: +UUUUUUU+ Punch range: +-------+ Extent map: +-------+DDDDDDD+DDDDDDD+DDDDDDD+DDDDDDD+ Second iteration: punch_start V +-------+UUUUUUU+-------+UUUUUUU+-------+ SEEK_DATA: V +-------+UUUUUUU+-------+UUUUUUU+-------+ SEEK_HOLE: ^ Data range: +UUUUUUU+ Punch range: +-------+ Extent map: +-------+DDDDDDD+-------+DDDDDDD+DDDDDDD+ Third iteration: punch_start V +-------+UUUUUUU+-------+UUUUUUU+-------+ SEEK_DATA: - moves into next folio in cache .... Punch range: +-------+ ...... Extent map: +-------+DDDDDDD+-------+DDDDDDD+-------+ ...... (to end of scan range or start of next data) As you can see, this scan does not care about folio size, sub-folio range granularity or filesystem block sizes. It also matches exactly how writeback handles dirty, partially up to date folios, so there's no stray delalloc blocks left around to be tripped over after failed or short writes occur. Indeed, if we move to sub-folio dirty range tracking, we can simply add a mapping_seek_hole_data() variant that walks dirty ranges in the page cache rather than up to date ranges. Then we can remove the inner loop from this code that looks up folios to determine dirty state. The above algorithm does not change - we just walk from discrete range to discrete range punching the gaps between them.... IOWs, the algorithm is largely future proof - the only thing that needs to change if we change iomap to track sub-folio dirty ranges is how we check the data range for being dirty. That should be no surprise, really, the surprise should be that we can make some simple mods to page cache seek to remove the need for checking dirty state in this code altogether.... > That sort of state > can be created by a prior read of the range on a sub-page block size fs, > or perhaps a racing async readahead (via read fault of a lower > offset..?), etc. Yup, generic/346 exercises this racing unaligned, sub-folio mmap write vs write() case. This test, specifically, was the reason I moved to using mapping_seek_hole_data() - g/346 found an endless stream of bugs in the sub-multi-page-folio range iteration code I kept trying to write.... > I suspect this is not a serious error because the page is dirty > and writeback will thus convert the block. The only exception to > that I can see is if the block is beyond EOF (consider a mapped > read to a page that straddles EOF, followed by a post-eof write > that fails), writeback won't actually map the block directly. I don't think that can happen. iomap_write_failed() calls truncate_pagecache_range() to remove any newly instantiated cached data beyond the original EOF. Hence the delalloc punch will remove everything beyond the original EOF that was allocated for the failed write. Hence when we get to writeback we're not going to find any up-to-date data beyond the EOF block in the page cache, nor any stray delalloc blocks way beyond EOF.... > It may convert if contiguous with delalloc blocks inside EOF (and > sufficiently sized physical extents exist), or even if not, should > still otherwise be cleaned up by the various other means we > already have to manage post-eof blocks. > > So IOW there's a tradeoff being made here for possible spurious > allocation and I/O and a subtle dependency on writeback that > should probably be documented somewhere. As per above, I don't think there is any spurious/stale allocation left behind by the punch code, nor is there any dependency on writeback to ignore it such issues. > The larger concern is that if > writeback eventually changes based on dirty range tracking in a way that > breaks this dependency, that introduces yet another stale delalloc block > landmine associated with this error handling code (regardless of whether > you want to call that a bug in this code, seek data, whatever), and > those problems are difficult enough to root cause as it is. If iomap changes how it tracks dirty ranges, this punch code only needs small changes to work with that correctly. There aren't any unknown landmines here - if we change dirty tracking, we know that we have to update the code that depends on the existing dirty tracking mechanisms to work correctly with the new infrastructure... Cheers, Dave. -- Dave Chinner david@xxxxxxxxxxxxx