[PATCH v11 0/7] mm, iomap, NFSD: stop DONTCACHE writes reading back shared pages

[PATCH v11 0/7] mm, iomap, NFSD: stop DONTCACHE writes reading back shared pages

[Mike Snitzer]

Hi,

An RWF_DONTCACHE write is buffered IO: its data goes through the page
cache, and writeback drops each folio once the folio is written back.
A write that starts or ends inside a folio covers that folio only
partially.  Usually another write covers the rest of the folio.  If
writeback drops the folio before the other write writes its part, the
other write must first read the folio back from disk.  32 threads
writing interleaved records to one XFS file, 4 KiB pages, before and
after 5/7:

  record  pwritev2() flags          4 KiB device       MiB/s
  bytes                             reads per write
                                    before   after     before  after
  47008   RWF_DONTCACHE             0.97     0.0005    831     2281
  47008   RWF_DONTCACHE|RWF_DSYNC   0.99     0.0005    601     1227
  6000    RWF_DONTCACHE             0.97     0.0001    121      488
  6000    RWF_DONTCACHE|RWF_DSYNC   0.99     0.0002     90      206

47008 bytes is the transfer size of IO500's ior-hard, in which each
rank writes interleaved records to one shared file.  Any record size
that is not a multiple of the page size has the same problem.  6000
bytes is a second example.

The test program and script: https://lore.kernel.org/linux-nfs/atqJBS8MlYfeUU-t@kernel.org/

- 1/7 moves code into a helper, with no functional change.
- 2/7 fixes 88d6f128d06d ("mm: track DONTCACHE dirty pages per
  bdi_writeback"), in v7.3: WB_DONTCACHE_DIRTY could drift.
- 3/7 reports WB_DONTCACHE_DIRTY in the bdi debugfs stats.
- 4/7 adds folio_mark_dropbehind() for a folio that is already dirty.
- 5/7 keeps a partially covered folio until the write that completes it.
- 6/7 lets a writer that syncs its own write complete those folios
  itself, after the sync.
- 7/7 makes NFSD do that for the misaligned ends of a direct-mode WRITE.

Merging: Chuck suggested taking the series through the mm tree.  1/7 to
6/7 apply cleanly to each of these mm branches:

  mm-unstable  93d7a816a2df ("fs/proc/task_mmu: fix lost maps entries above 2^63")
  mm-new       488f62a72b8e ("riscv: mm: fix concurrency in mark_new_valid_map()")
  mm-stable    fa4ad798f6f6 ("Merge tag 'for-7.4-rc1-tag' of git://git.kernel.org/pub/scm/linux/kernel/git/kdave/linux")

7/7 needs the NFSD direct-mode changes that are in cel/nfsd-testing.
The whole series applies to cel/nfsd-testing at f7ef2ce48812 ("SUNRPC:
Load the GSS module by OID only when the mechanism is missing"), where
each patch builds with W=1 with no warnings.  So 7/7 could follow in
the nfsd tree once the mm part is merged.

Changes since v10:
- Two fixes, from sashiko's review of v10 6/7
  (https://lore.kernel.org/r/sashiko-outbox-198089@kernel.org):
  - 5/7: iomap decides whether a unit is an edge from what is left of
    the whole write (iter->len), not of the current extent.  With blocks
    smaller than the folio, an extent boundary can split a unit that the
    write covers in full.  v10 then looked the unit up as two edges.
    With IOCB_DONTCACHE_DEFER, nothing marked the unit, and it stayed
    cached until reclaim.
  - 6/7: filemap_dontcache_edge_complete() leaves a clean mapped folio
    unmarked, as filemap_end_dropbehind() leaves a mapped folio.  v10
    marked it dropbehind, but no writeback was to clean it.
- The commit messages and code comments use different terms.  They say
  RWF_DONTCACHE, DONTCACHE or dropbehind, and "the other write that
  covers the rest of the folio".
- 5/7 says that an RWF_DONTCACHE write goes through the page cache.  It
  says that the kernel cannot know if another write will come, and what
  stays cached when no other write comes.  It names IO500's ior-hard as
  the source of the 47008-byte records, and adds the 6000-byte result.
- 6/7 and 7/7 say which WRITEs NFSD syncs: a WRITE that the client
  sends as FILE_SYNC or DATA_SYNC gets one vfs_fsync_range() over the
  whole WRITE.  An UNSTABLE WRITE gets none.  7/7 also says that the
  Linux client sends an O_DIRECT write that fits in one WRITE as
  FILE_SYNC.
- Christoph's Reviewed-by on 1/7.  Jeff's Reviewed-by of v10 is carried
  on 2/7, 3/7, 4/7 and 7/7, where only comments changed.  It is dropped
  from 5/7 and 6/7, which have the fixes above.
- The review notes below the --- line are dropped.
- Rebased on the current cel/nfsd-testing.

v10: https://lore.kernel.org/linux-nfs/20261103211754.35352-1-snitzer@kernel.org/
v9: https://lore.kernel.org/linux-nfs/20261024221152.6965-1-snitzer@kernel.org/
v8: https://lore.kernel.org/linux-nfs/20261023024548.71123-1-snitzer@kernel.org/
v7: https://lore.kernel.org/linux-nfs/20261020220346.28020-1-snitzer@kernel.org/
v6: https://lore.kernel.org/linux-nfs/atbfEbgMU6UGkn2X@kernel.org/
v5: https://lore.kernel.org/linux-nfs/20261019150725.1507-1-snitzer@kernel.org/
v4: https://lore.kernel.org/linux-nfs/20261015233309.42999-1-snitzer@kernel.org/
v3: https://lore.kernel.org/linux-nfs/20261001045502.48381-1-snitzer@kernel.org/
v2: https://lore.kernel.org/linux-nfs/20260929231329.22018-1-snitzer@kernel.org/
v1: https://lore.kernel.org/linux-nfs/20260929173423.16149-1-snitzer@kernel.org/

Thanks,
Mike

Mike Snitzer (7):
  mm: move clearing the dropbehind flag of a found folio into a helper
  mm: take a folio off WB_DONTCACHE_DIRTY only if it was counted there
  mm: expose WB_DONTCACHE_DIRTY in bdi debugfs stats
  mm: add folio_mark_dropbehind() to mark a folio dropbehind and account it
  mm, iomap: keep a partially covered page until a DONTCACHE write completes it
  mm, iomap: let a DONTCACHE writer complete its shared pages itself
  NFSD: drop a shared boundary page only once both WRITEs have written it

 .../filesystems/nfs/nfsd-io-modes.rst         |  18 ++
 fs/iomap/buffered-io.c                        |  30 ++-
 fs/iomap/trace.h                              |   3 +-
 fs/nfsd/vfs.c                                 | 107 ++++++++-
 include/linux/fs.h                            |   5 +-
 include/linux/iomap.h                         |   1 +
 include/linux/pagemap.h                       |  12 +
 mm/backing-dev.c                              |   6 +
 mm/filemap.c                                  | 207 ++++++++++++++++--
 9 files changed, 358 insertions(+), 31 deletions(-)


base-commit: f7ef2ce48812080b87572a2f0fdec93cb2d2a629
-- 
2.52.0

[PATCH v11 1/7] mm: move clearing the dropbehind flag of a found folio into a helper

[Mike Snitzer]

Move the code in __filemap_get_folio_mpol() that clears PG_dropbehind
on a folio found by a cached lookup, and takes the folio off
WB_DONTCACHE_DIRTY, into filemap_clear_dropbehind().  No functional
change.  The next patch changes the helper.

Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Mike Snitzer <snitzer@kernel.org>
---
 mm/filemap.c | 33 ++++++++++++++++++++-------------
 1 file changed, 20 insertions(+), 13 deletions(-)

diff --git a/mm/filemap.c b/mm/filemap.c
index b04be64a5069f..437113addbddf 100644
--- a/mm/filemap.c
+++ b/mm/filemap.c
@@ -1956,6 +1956,24 @@ void *filemap_get_entry(struct address_space *mapping, pgoff_t index)
 	return folio;
 }
 
+/* Clear the dropbehind flag of a folio that a cached lookup found. */
+static void filemap_clear_dropbehind(struct address_space *mapping,
+				     struct folio *folio)
+{
+	if (!folio_test_clear_dropbehind(folio))
+		return;
+	if (folio_test_dirty(folio) && mapping_can_writeback(mapping)) {
+		struct inode *inode = mapping->host;
+		struct bdi_writeback *wb;
+		struct wb_lock_cookie cookie = {};
+		long nr = folio_nr_pages(folio);
+
+		wb = unlocked_inode_to_wb_begin(inode, &cookie);
+		wb_stat_mod(wb, WB_DONTCACHE_DIRTY, -nr);
+		unlocked_inode_to_wb_end(inode, &cookie);
+	}
+}
+
 /**
  * __filemap_get_folio_mpol - Find and get a reference to a folio.
  * @mapping: The address_space to search.
@@ -2087,19 +2105,8 @@ struct folio *__filemap_get_folio_mpol(struct address_space *mapping,
 	if (!folio)
 		return ERR_PTR(-ENOENT);
 	/* not an uncached lookup, clear uncached if set */
-	if (!(fgp_flags & FGP_DONTCACHE) && folio_test_clear_dropbehind(folio)) {
-		if (folio_test_dirty(folio) &&
-		    mapping_can_writeback(mapping)) {
-			struct inode *inode = mapping->host;
-			struct bdi_writeback *wb;
-			struct wb_lock_cookie cookie = {};
-			long nr = folio_nr_pages(folio);
-
-			wb = unlocked_inode_to_wb_begin(inode, &cookie);
-			wb_stat_mod(wb, WB_DONTCACHE_DIRTY, -nr);
-			unlocked_inode_to_wb_end(inode, &cookie);
-		}
-	}
+	if (!(fgp_flags & FGP_DONTCACHE))
+		filemap_clear_dropbehind(mapping, folio);
 	return folio;
 }
 EXPORT_SYMBOL(__filemap_get_folio_mpol);
-- 
2.52.0

[PATCH v11 2/7] mm: take a folio off WB_DONTCACHE_DIRTY only if it was counted there

[Mike Snitzer]

A cached lookup that finds a dropbehind folio clears the flag and
subtracts the folio from WB_DONTCACHE_DIRTY if PG_dirty is set.  That
test races with the dirtying of the folio.  filemap_dirty_folio() sets
PG_dirty before folio_account_dirtied() counts the folio, so a lookup in
between subtracts a folio that was never added.  The lookup also does
not hold the folio lock.  Every other path that subtracts a dirty folio
from the counter holds it, so the lookup can subtract the folio a second
time.  Either way the counter drifts, and once it is negative the
DONTCACHE writeback kick writes nothing.

Fix this by having filemap_clear_dropbehind() clear the dropbehind flag
of a folio that a cached lookup found, and take the folio off
WB_DONTCACHE_DIRTY if it is counted there.  With the folio locked and
the i_pages lock held, the folio is counted if both PG_dirty and the
dirty tag are set.  A folio that another task has locked keeps the
flag.

Fixes: 88d6f128d06d ("mm: track DONTCACHE dirty pages per bdi_writeback")
Reviewed-by: Jeff Layton <jlayton@kernel.org>
Signed-off-by: Mike Snitzer <snitzer@kernel.org>
---
 mm/filemap.c | 42 ++++++++++++++++++++++++++++--------------
 1 file changed, 28 insertions(+), 14 deletions(-)

diff --git a/mm/filemap.c b/mm/filemap.c
index 437113addbddf..98ce955404127 100644
--- a/mm/filemap.c
+++ b/mm/filemap.c
@@ -1956,22 +1956,36 @@ void *filemap_get_entry(struct address_space *mapping, pgoff_t index)
 	return folio;
 }
 
-/* Clear the dropbehind flag of a folio that a cached lookup found. */
+/*
+ * Clear the dropbehind flag of a folio that a cached lookup found, and take
+ * the folio off WB_DONTCACHE_DIRTY if it is counted there.  With the folio
+ * locked and the i_pages lock held, the folio is counted if both PG_dirty and
+ * the dirty tag are set.  A folio that another task has locked keeps the flag.
+ */
 static void filemap_clear_dropbehind(struct address_space *mapping,
-				     struct folio *folio)
+				     struct folio *folio, bool locked)
 {
-	if (!folio_test_clear_dropbehind(folio))
-		return;
-	if (folio_test_dirty(folio) && mapping_can_writeback(mapping)) {
-		struct inode *inode = mapping->host;
-		struct bdi_writeback *wb;
-		struct wb_lock_cookie cookie = {};
-		long nr = folio_nr_pages(folio);
+	unsigned long flags;
 
-		wb = unlocked_inode_to_wb_begin(inode, &cookie);
-		wb_stat_mod(wb, WB_DONTCACHE_DIRTY, -nr);
-		unlocked_inode_to_wb_end(inode, &cookie);
+	if (!locked) {
+		if (!folio_trylock(folio))
+			return;
+		if (unlikely(folio->mapping != mapping))
+			goto out;
 	}
+	if (!mapping_can_writeback(mapping)) {
+		folio_clear_dropbehind(folio);
+		goto out;
+	}
+	xa_lock_irqsave(&mapping->i_pages, flags);
+	if (folio_test_clear_dropbehind(folio) && folio_test_dirty(folio) &&
+	    xa_get_mark(&mapping->i_pages, folio->index, PAGECACHE_TAG_DIRTY))
+		wb_stat_mod(inode_to_wb(mapping->host), WB_DONTCACHE_DIRTY,
+			    -folio_nr_pages(folio));
+	xa_unlock_irqrestore(&mapping->i_pages, flags);
+out:
+	if (!locked)
+		folio_unlock(folio);
 }
 
 /**
@@ -2105,8 +2119,8 @@ struct folio *__filemap_get_folio_mpol(struct address_space *mapping,
 	if (!folio)
 		return ERR_PTR(-ENOENT);
 	/* not an uncached lookup, clear uncached if set */
-	if (!(fgp_flags & FGP_DONTCACHE))
-		filemap_clear_dropbehind(mapping, folio);
+	if (!(fgp_flags & FGP_DONTCACHE) && folio_test_dropbehind(folio))
+		filemap_clear_dropbehind(mapping, folio, fgp_flags & FGP_LOCK);
 	return folio;
 }
 EXPORT_SYMBOL(__filemap_get_folio_mpol);
-- 
2.52.0

[PATCH v11 3/7] mm: expose WB_DONTCACHE_DIRTY in bdi debugfs stats

[Mike Snitzer]

The per-bdi_writeback WB_DONTCACHE_DIRTY counter drives how much the
flusher writes back in response to an IOCB_DONTCACHE kick, but it is
not reported anywhere, so the DONTCACHE dirty backlog can only be
inferred from MemFree/Cached movement.

Report it as BdiDontcacheDirty in /sys/kernel/debug/bdi/<dev>/stats
and as WbDontcacheDirty in the per-wb stats, next to the
Reclaimable line.  Use wb_stat_sum() rather than wb_stat() to fold in
the per-cpu batches: the flusher targets the summed value and small
backlogs would otherwise read as zero.

Fixes: 88d6f128d06d ("mm: track DONTCACHE dirty pages per bdi_writeback")
Reviewed-by: Jeff Layton <jlayton@kernel.org>
Signed-off-by: Mike Snitzer <snitzer@kernel.org>
---
 mm/backing-dev.c | 6 ++++++
 1 file changed, 6 insertions(+)

diff --git a/mm/backing-dev.c b/mm/backing-dev.c
index fad1495d3b200..a09c21759fb9f 100644
--- a/mm/backing-dev.c
+++ b/mm/backing-dev.c
@@ -46,6 +46,7 @@ struct wb_stats {
 	unsigned long nr_dirty_time;
 	unsigned long nr_writeback;
 	unsigned long nr_reclaimable;
+	unsigned long nr_dontcache_dirty;
 	unsigned long nr_dirtied;
 	unsigned long nr_written;
 	unsigned long dirty_thresh;
@@ -78,6 +79,7 @@ static void collect_wb_stats(struct wb_stats *stats,
 
 	stats->nr_writeback += wb_stat(wb, WB_WRITEBACK);
 	stats->nr_reclaimable += wb_stat(wb, WB_RECLAIMABLE);
+	stats->nr_dontcache_dirty += wb_stat_sum(wb, WB_DONTCACHE_DIRTY);
 	stats->nr_dirtied += wb_stat(wb, WB_DIRTIED);
 	stats->nr_written += wb_stat(wb, WB_WRITTEN);
 	stats->wb_thresh += wb_calc_thresh(wb, stats->dirty_thresh);
@@ -125,6 +127,7 @@ static int bdi_debug_stats_show(struct seq_file *m, void *v)
 	seq_printf(m,
 		   "BdiWriteback:       %10lu kB\n"
 		   "BdiReclaimable:     %10lu kB\n"
+		   "BdiDontcacheDirty:  %10lu kB\n"
 		   "BdiDirtyThresh:     %10lu kB\n"
 		   "DirtyThresh:        %10lu kB\n"
 		   "BackgroundThresh:   %10lu kB\n"
@@ -139,6 +142,7 @@ static int bdi_debug_stats_show(struct seq_file *m, void *v)
 		   "state:              %10lx\n",
 		   K(stats.nr_writeback),
 		   K(stats.nr_reclaimable),
+		   K(stats.nr_dontcache_dirty),
 		   K(stats.wb_thresh),
 		   K(dirty_thresh),
 		   K(background_thresh),
@@ -163,6 +167,7 @@ static void wb_stats_show(struct seq_file *m, struct bdi_writeback *wb,
 		   "WbCgIno:           %10lu\n"
 		   "WbWriteback:       %10lu kB\n"
 		   "WbReclaimable:     %10lu kB\n"
+		   "WbDontcacheDirty:  %10lu kB\n"
 		   "WbDirtyThresh:     %10lu kB\n"
 		   "WbDirtied:         %10lu kB\n"
 		   "WbWritten:         %10lu kB\n"
@@ -179,6 +184,7 @@ static void wb_stats_show(struct seq_file *m, struct bdi_writeback *wb,
 #endif
 		   K(stats->nr_writeback),
 		   K(stats->nr_reclaimable),
+		   K(stats->nr_dontcache_dirty),
 		   K(stats->wb_thresh),
 		   K(stats->nr_dirtied),
 		   K(stats->nr_written),
-- 
2.52.0

[PATCH v11 4/7] mm: add folio_mark_dropbehind() to mark a folio dropbehind and account it

[Mike Snitzer]

WB_DONTCACHE_DIRTY counts the dirty folios that are marked dropbehind.
When a dropbehind folio is dirtied, folio_account_dirtied() adds it to
the counter.  When the folio is cleaned, folio_clear_dirty_for_io() or
folio_account_cleaned() subtracts it again.  This balances only if the
folio is marked before it is dirtied, as it is for the folios that an
RWF_DONTCACHE write creates.

To prepare for NFSD to mark a folio dropbehind once both writes that
share it have written their parts, handle a folio that is already dirty
when it is marked.  folio_set_dropbehind() only sets the flag.  With it,
the clean would subtract the folio from WB_DONTCACHE_DIRTY although it
was never added, and the counter would go negative.

Add folio_mark_dropbehind(), the inverse of filemap_clear_dropbehind().
With the folio locked, it sets PG_dropbehind.  Under the i_pages lock,
it adds the folio to WB_DONTCACHE_DIRTY if both PG_dirty and the dirty
tag are set.  A dirtier that has set PG_dirty but not yet counted the
folio finds the flag set and counts the folio itself.

The name follows folio_set_dirty() and folio_mark_dirty(): the set
variant only sets the flag, the mark variant also accounts for it.

Reviewed-by: Jeff Layton <jlayton@kernel.org>
Signed-off-by: Mike Snitzer <snitzer@kernel.org>
---
 include/linux/pagemap.h |  1 +
 mm/filemap.c            | 32 ++++++++++++++++++++++++++++++++
 2 files changed, 33 insertions(+)

diff --git a/include/linux/pagemap.h b/include/linux/pagemap.h
index 73af18a373670..3f5295d6fd73b 100644
--- a/include/linux/pagemap.h
+++ b/include/linux/pagemap.h
@@ -1306,6 +1306,7 @@ void end_page_writeback(struct page *page);
 void folio_end_writeback(struct folio *folio);
 void folio_end_writeback_no_dropbehind(struct folio *folio);
 void folio_end_dropbehind(struct folio *folio);
+void folio_mark_dropbehind(struct folio *folio);
 void folio_wait_stable(struct folio *folio);
 void __folio_mark_dirty(struct folio *folio, struct address_space *, int warn);
 void folio_account_cleaned(struct folio *folio, struct bdi_writeback *wb);
diff --git a/mm/filemap.c b/mm/filemap.c
index 98ce955404127..a54d0fbd37222 100644
--- a/mm/filemap.c
+++ b/mm/filemap.c
@@ -1654,6 +1654,38 @@ void folio_end_dropbehind(struct folio *folio)
 }
 EXPORT_SYMBOL_GPL(folio_end_dropbehind);
 
+/**
+ * folio_mark_dropbehind - mark an existing folio dropbehind
+ * @folio: the folio, locked by the caller
+ *
+ * Mark @folio so that the writeback that cleans it drops it.  If the folio
+ * is already counted as dirty, which with the i_pages lock held means both
+ * PG_dirty and the dirty tag are set, add it to WB_DONTCACHE_DIRTY so that
+ * the clean that subtracts it leaves the counter balanced.
+ */
+void folio_mark_dropbehind(struct folio *folio)
+{
+	struct address_space *mapping = folio->mapping;
+	unsigned long flags;
+
+	VM_WARN_ON_ONCE_FOLIO(!folio_test_locked(folio), folio);
+
+	if (folio_test_dropbehind(folio))
+		return;
+	if (!mapping || !mapping_can_writeback(mapping)) {
+		folio_set_dropbehind(folio);
+		return;
+	}
+	xa_lock_irqsave(&mapping->i_pages, flags);
+	folio_set_dropbehind(folio);
+	if (folio_test_dirty(folio) &&
+	    xa_get_mark(&mapping->i_pages, folio->index, PAGECACHE_TAG_DIRTY))
+		wb_stat_mod(inode_to_wb(mapping->host), WB_DONTCACHE_DIRTY,
+			    folio_nr_pages(folio));
+	xa_unlock_irqrestore(&mapping->i_pages, flags);
+}
+EXPORT_SYMBOL_GPL(folio_mark_dropbehind);
+
 /**
  * folio_end_writeback_no_dropbehind - End writeback against a folio.
  * @folio: The folio.
-- 
2.52.0

[PATCH v11 5/7] mm, iomap: keep a partially covered page until a DONTCACHE write completes it

[Mike Snitzer]

An RWF_DONTCACHE buffered write goes through the page cache like any
buffered write.  It marks every folio it creates dropbehind, so
writeback drops the folio as soon as the data of the write is on disk.
A write that starts or ends inside a folio covers that folio only
partially.  Usually another write covers the rest of the folio.  An
example is the next record of a file written in records that are not a
multiple of the page size.  Writeback can drop the folio before the
other write writes its part.  Then the other write must first read the
folio back from disk: one read per misaligned write.

Add FGP_DONTCACHE_EDGE for the lookup of such a folio.  iomap sets it
when a write starts or ends inside the smallest folio the mapping
allows.  A folio created for an edge is not marked dropbehind.  A folio
found uptodate already holds the data of the other write, so this write
completes it.  Mark it with folio_mark_dropbehind() under the folio
lock, and the writeback that cleans the data of both writes drops it.
With blocks smaller than the folio, the edge is not uptodate yet, so
folio_dontcache_edge_written() also accepts a folio whose blocks outside
this write are uptodate.  A larger folio, such as one from readahead, is
left alone.

The kernel cannot know if another write will come, and this commit does
not assume it.  It reads only the uptodate state of the folio.  An edge
that no other write completes, such as the last page of a file, stays
cached until reclaim.  Writeback drops all other folios of the write, as
before.

With this commit, 32 threads write interleaved 47008-byte records to
one XFS file with RWF_DONTCACHE.  Device reads per write fell from 0.97
to 0.0005, and throughput rose from 831 to 2281 MiB/s.  47008 bytes is
the transfer size of IO500's ior-hard, in which each rank writes
interleaved records to one shared file.  Any record size that is not a
multiple of the page size has the same problem.  With 6000-byte records,
device reads per write fell from 0.97 to 0.0001, and throughput rose
from 121 to 488 MiB/s.

Signed-off-by: Mike Snitzer <snitzer@kernel.org>
---
 fs/iomap/buffered-io.c  | 26 +++++++++++++++---
 include/linux/pagemap.h |  5 ++++
 mm/filemap.c            | 59 ++++++++++++++++++++++++++++++++++++++++-
 3 files changed, 86 insertions(+), 4 deletions(-)

diff --git a/fs/iomap/buffered-io.c b/fs/iomap/buffered-io.c
index d6e2f1ded5ca6..28431192fb2e3 100644
--- a/fs/iomap/buffered-io.c
+++ b/fs/iomap/buffered-io.c
@@ -798,16 +798,36 @@ EXPORT_SYMBOL_GPL(iomap_is_partially_uptodate);
  */
 struct folio *iomap_get_folio(struct iomap_iter *iter, loff_t pos, size_t len)
 {
+	struct address_space *mapping = iter->inode->i_mapping;
 	fgf_t fgp = FGP_WRITEBEGIN;
+	struct folio *folio;
 
 	if (iter->flags & IOMAP_NOWAIT)
 		fgp |= FGP_NOWAIT;
-	if (iter->flags & IOMAP_DONTCACHE)
+	if (iter->flags & IOMAP_DONTCACHE) {
+		size_t unit = mapping_min_folio_nrbytes(mapping);
+		size_t off = pos & (unit - 1);
+
 		fgp |= FGP_DONTCACHE;
+		/* Another write may cover the rest of this unit. */
+		if (off || iter->len < unit) {
+			fgp |= FGP_DONTCACHE_EDGE;
+			len = min_t(size_t, len, unit - off);
+		}
+	}
 	fgp |= fgf_set_order(len);
 
-	return __filemap_get_folio(iter->inode->i_mapping, pos >> PAGE_SHIFT,
-			fgp, mapping_gfp_mask(iter->inode->i_mapping));
+	folio = __filemap_get_folio(mapping, pos >> PAGE_SHIFT, fgp,
+				    mapping_gfp_mask(mapping));
+	/*
+	 * With blocks smaller than the folio, an edge is not uptodate
+	 * until this write completes it.
+	 */
+	if ((fgp & FGP_DONTCACHE_EDGE) &&
+	    !IS_ERR(folio) && !folio_test_uptodate(folio) &&
+	    folio_dontcache_edge_written(folio, pos, len))
+		folio_mark_dropbehind(folio);
+	return folio;
 }
 EXPORT_SYMBOL_GPL(iomap_get_folio);
 
diff --git a/include/linux/pagemap.h b/include/linux/pagemap.h
index 3f5295d6fd73b..07f7f5190c24c 100644
--- a/include/linux/pagemap.h
+++ b/include/linux/pagemap.h
@@ -698,6 +698,9 @@ pgoff_t page_cache_prev_miss(struct address_space *mapping,
  * * %FGP_NOWAIT - Don't block on the folio lock.
  * * %FGP_STABLE - Wait for the folio to be stable (finished writeback)
  * * %FGP_DONTCACHE - Uncached buffered IO
+ * * %FGP_DONTCACHE_EDGE - With %FGP_DONTCACHE: the write covers the folio
+ *   only partially.  A new folio is not marked dropbehind, one found
+ *   uptodate is.
  * * %FGP_WRITEBEGIN - The flags to use in a filesystem write_begin()
  *   implementation.
  */
@@ -712,6 +715,7 @@ typedef unsigned int __bitwise fgf_t;
 #define FGP_FOR_MMAP		((__force fgf_t)0x00000040)
 #define FGP_STABLE		((__force fgf_t)0x00000080)
 #define FGP_DONTCACHE		((__force fgf_t)0x00000100)
+#define FGP_DONTCACHE_EDGE	((__force fgf_t)0x00000200)
 #define FGF_GET_ORDER(fgf)	(((__force unsigned)fgf) >> 26)	/* top 6 bits */
 
 #define FGP_WRITEBEGIN		(FGP_LOCK | FGP_WRITE | FGP_CREAT | FGP_STABLE)
@@ -1307,6 +1311,7 @@ void folio_end_writeback(struct folio *folio);
 void folio_end_writeback_no_dropbehind(struct folio *folio);
 void folio_end_dropbehind(struct folio *folio);
 void folio_mark_dropbehind(struct folio *folio);
+bool folio_dontcache_edge_written(struct folio *folio, loff_t pos, size_t len);
 void folio_wait_stable(struct folio *folio);
 void __folio_mark_dirty(struct folio *folio, struct address_space *, int warn);
 void folio_account_cleaned(struct folio *folio, struct bdi_writeback *wb);
diff --git a/mm/filemap.c b/mm/filemap.c
index a54d0fbd37222..afba94b33dcc6 100644
--- a/mm/filemap.c
+++ b/mm/filemap.c
@@ -1686,6 +1686,52 @@ void folio_mark_dropbehind(struct folio *folio)
 }
 EXPORT_SYMBOL_GPL(folio_mark_dropbehind);
 
+/* The unit two writes share is the smallest folio the mapping allows. */
+static bool filemap_dontcache_edge_unit(struct address_space *mapping,
+					struct folio *folio)
+{
+	return folio_nr_pages(folio) == mapping_min_folio_nrpages(mapping);
+}
+
+/**
+ * folio_dontcache_edge_written - has another write written the rest?
+ * @folio: locked edge folio
+ * @pos: file offset of the first byte this write covers in @folio
+ * @len: bytes this write covers in @folio
+ *
+ * Return: true if @folio is an edge unit that is uptodate, or whose blocks
+ * that this write does not cover in full are uptodate.
+ */
+bool folio_dontcache_edge_written(struct folio *folio, loff_t pos, size_t len)
+{
+	struct address_space *mapping = folio->mapping;
+	size_t size = folio_size(folio);
+	size_t from, to, bs;
+
+	VM_WARN_ON_ONCE_FOLIO(!folio_test_locked(folio), folio);
+
+	if (!mapping || !filemap_dontcache_edge_unit(mapping, folio))
+		return false;
+	if (folio_test_uptodate(folio))
+		return true;
+	if (!mapping->a_ops->is_partially_uptodate)
+		return false;
+
+	/* The blocks this write covers in full are [from, to). */
+	bs = i_blocksize(mapping->host);
+	from = offset_in_folio(folio, pos);
+	to = min(from + len, size);
+	from = round_up(from, bs);
+	to = round_down(to, bs);
+	if (from && !mapping->a_ops->is_partially_uptodate(folio, 0, from))
+		return false;
+	if (to < size &&
+	    !mapping->a_ops->is_partially_uptodate(folio, to, size - to))
+		return false;
+	return true;
+}
+EXPORT_SYMBOL_GPL(folio_dontcache_edge_written);
+
 /**
  * folio_end_writeback_no_dropbehind - End writeback against a folio.
  * @folio: The folio.
@@ -2078,6 +2124,15 @@ struct folio *__filemap_get_folio_mpol(struct address_space *mapping,
 
 	if (fgp_flags & FGP_STABLE)
 		folio_wait_stable(folio);
+
+	/*
+	 * An uptodate edge holds the part of the other write, so this write
+	 * completes it.
+	 */
+	if ((fgp_flags & FGP_DONTCACHE) && (fgp_flags & FGP_DONTCACHE_EDGE) &&
+	    (fgp_flags & FGP_LOCK) && folio_test_uptodate(folio) &&
+	    filemap_dontcache_edge_unit(mapping, folio))
+		folio_mark_dropbehind(folio);
 no_page:
 	if (!folio && (fgp_flags & FGP_CREAT)) {
 		unsigned int min_order = mapping_min_folio_order(mapping);
@@ -2115,7 +2170,9 @@ struct folio *__filemap_get_folio_mpol(struct address_space *mapping,
 			/* Init accessed so avoid atomic mark_page_accessed later */
 			if (fgp_flags & FGP_ACCESSED)
 				__folio_set_referenced(folio);
-			if (fgp_flags & FGP_DONTCACHE)
+			/* A new edge waits for the write that completes it. */
+			if ((fgp_flags & FGP_DONTCACHE) &&
+			    !(fgp_flags & FGP_DONTCACHE_EDGE))
 				__folio_set_dropbehind(folio);
 
 			err = filemap_add_folio(mapping, folio, index, gfp);
-- 
2.52.0

[PATCH v11 6/7] mm, iomap: let a DONTCACHE writer complete its shared pages itself

[Mike Snitzer]

With the previous commit, the write that completes a shared folio marks
it while it is still dirty.  The folio then counts as DONTCACHE dirty
and wakes the DONTCACHE writeback kick.  Some callers sync once, after
writing several segments.  NFSD does this for a WRITE that the client
sends as FILE_SYNC or DATA_SYNC.  Such a caller wakes the kick once per
segment, for pages that its own sync is about to write.

Add IOCB_DONTCACHE_DEFER (with IOMAP_DONTCACHE_DEFER and
FGP_DONTCACHE_DEFER).  The write leaves the folios it covers only
partially unmarked, and the caller completes them itself.  Called right
before the write, filemap_dontcache_edge_completes() tells the caller
whether the other write that shares the folio already wrote its part.
Called once the write and its sync are done,
filemap_dontcache_edge_complete() waits for writeback in flight.  Then
it marks a dirty folio and kicks writeback, or drops a clean unmapped
folio at once.  A clean mapped folio stays cached, as
filemap_end_dropbehind() leaves a mapped folio.

Two writes that arrive together may both find the folio absent.  The
folio then stays cached until reclaim.

IOCB_DONTCACHE_DEFER enables NFSD, in a later commit, to avoid waking
the DONTCACHE writeback kick.

Signed-off-by: Mike Snitzer <snitzer@kernel.org>
---
 fs/iomap/buffered-io.c  |  6 +++-
 fs/iomap/trace.h        |  3 +-
 include/linux/fs.h      |  5 ++-
 include/linux/iomap.h   |  1 +
 include/linux/pagemap.h |  6 ++++
 mm/filemap.c            | 73 +++++++++++++++++++++++++++++++++++++++--
 6 files changed, 89 insertions(+), 5 deletions(-)

diff --git a/fs/iomap/buffered-io.c b/fs/iomap/buffered-io.c
index 28431192fb2e3..053fe39dec595 100644
--- a/fs/iomap/buffered-io.c
+++ b/fs/iomap/buffered-io.c
@@ -812,6 +812,8 @@ struct folio *iomap_get_folio(struct iomap_iter *iter, loff_t pos, size_t len)
 		/* Another write may cover the rest of this unit. */
 		if (off || iter->len < unit) {
 			fgp |= FGP_DONTCACHE_EDGE;
+			if (iter->flags & IOMAP_DONTCACHE_DEFER)
+				fgp |= FGP_DONTCACHE_DEFER;
 			len = min_t(size_t, len, unit - off);
 		}
 	}
@@ -823,7 +825,7 @@ struct folio *iomap_get_folio(struct iomap_iter *iter, loff_t pos, size_t len)
 	 * With blocks smaller than the folio, an edge is not uptodate
 	 * until this write completes it.
 	 */
-	if ((fgp & FGP_DONTCACHE_EDGE) &&
+	if ((fgp & FGP_DONTCACHE_EDGE) && !(fgp & FGP_DONTCACHE_DEFER) &&
 	    !IS_ERR(folio) && !folio_test_uptodate(folio) &&
 	    folio_dontcache_edge_written(folio, pos, len))
 		folio_mark_dropbehind(folio);
@@ -1333,6 +1335,8 @@ iomap_file_buffered_write(struct kiocb *iocb, struct iov_iter *i,
 		iter.flags |= IOMAP_NOWAIT;
 	if (iocb->ki_flags & IOCB_DONTCACHE)
 		iter.flags |= IOMAP_DONTCACHE;
+	if (iocb->ki_flags & IOCB_DONTCACHE_DEFER)
+		iter.flags |= IOMAP_DONTCACHE_DEFER;
 
 	while ((ret = iomap_iter(&iter, ops)) > 0)
 		iter.status = iomap_write_iter(&iter, i, write_ops);
diff --git a/fs/iomap/trace.h b/fs/iomap/trace.h
index e1ea8392cf47d..4f0343eeed1b4 100644
--- a/fs/iomap/trace.h
+++ b/fs/iomap/trace.h
@@ -104,7 +104,8 @@ DEFINE_RANGE_EVENT(iomap_zero_iter);
 	{ IOMAP_UNSHARE,	"UNSHARE" }, \
 	{ IOMAP_DAX,		"DAX" }, \
 	{ IOMAP_ATOMIC,		"ATOMIC" }, \
-	{ IOMAP_DONTCACHE,	"DONTCACHE" }
+	{ IOMAP_DONTCACHE,	"DONTCACHE" }, \
+	{ IOMAP_DONTCACHE_DEFER, "DONTCACHE_DEFER" }
 
 #define IOMAP_F_FLAGS_STRINGS \
 	{ IOMAP_F_NEW,		"NEW" }, \
diff --git a/include/linux/fs.h b/include/linux/fs.h
index 99caef43ff3c4..19aaba0bf41f9 100644
--- a/include/linux/fs.h
+++ b/include/linux/fs.h
@@ -357,6 +357,8 @@ struct readahead_control;
 /* kiocb is a read or write operation submitted by fs/aio.c. */
 #define IOCB_AIO_RW		(1 << 22)
 #define IOCB_HAS_METADATA	(1 << 23)
+/* With IOCB_DONTCACHE: see filemap_dontcache_edge_complete() */
+#define IOCB_DONTCACHE_DEFER	(1 << 24)
 
 /* for use in trace events */
 #define TRACE_IOCB_STRINGS \
@@ -374,7 +376,8 @@ struct readahead_control;
 	{ IOCB_NOIO,		"NOIO" }, \
 	{ IOCB_ALLOC_CACHE,	"ALLOC_CACHE" }, \
 	{ IOCB_AIO_RW,		"AIO_RW" }, \
-	{ IOCB_HAS_METADATA,	"AIO_HAS_METADATA" }
+	{ IOCB_HAS_METADATA,	"AIO_HAS_METADATA" }, \
+	{ IOCB_DONTCACHE_DEFER,	"DONTCACHE_DEFER" }
 
 struct kiocb {
 	struct file		*ki_filp;
diff --git a/include/linux/iomap.h b/include/linux/iomap.h
index 5254430844b2c..fcef97e64a0bd 100644
--- a/include/linux/iomap.h
+++ b/include/linux/iomap.h
@@ -212,6 +212,7 @@ struct iomap_write_ops {
 #endif /* CONFIG_FS_DAX */
 #define IOMAP_ATOMIC		(1 << 9) /* torn-write protection */
 #define IOMAP_DONTCACHE		(1 << 10)
+#define IOMAP_DONTCACHE_DEFER	(1 << 11) /* see IOCB_DONTCACHE_DEFER */
 
 /*
  * Return the existing mapping at pos, or reserve space starting at pos for up
diff --git a/include/linux/pagemap.h b/include/linux/pagemap.h
index 07f7f5190c24c..74ae4914b5820 100644
--- a/include/linux/pagemap.h
+++ b/include/linux/pagemap.h
@@ -701,6 +701,8 @@ pgoff_t page_cache_prev_miss(struct address_space *mapping,
  * * %FGP_DONTCACHE_EDGE - With %FGP_DONTCACHE: the write covers the folio
  *   only partially.  A new folio is not marked dropbehind, one found
  *   uptodate is.
+ * * %FGP_DONTCACHE_DEFER - With %FGP_DONTCACHE_EDGE: a folio found uptodate is
+ *   not marked either, the caller completes it.
  * * %FGP_WRITEBEGIN - The flags to use in a filesystem write_begin()
  *   implementation.
  */
@@ -716,6 +718,7 @@ typedef unsigned int __bitwise fgf_t;
 #define FGP_STABLE		((__force fgf_t)0x00000080)
 #define FGP_DONTCACHE		((__force fgf_t)0x00000100)
 #define FGP_DONTCACHE_EDGE	((__force fgf_t)0x00000200)
+#define FGP_DONTCACHE_DEFER	((__force fgf_t)0x00000400)
 #define FGF_GET_ORDER(fgf)	(((__force unsigned)fgf) >> 26)	/* top 6 bits */
 
 #define FGP_WRITEBEGIN		(FGP_LOCK | FGP_WRITE | FGP_CREAT | FGP_STABLE)
@@ -1312,6 +1315,9 @@ void folio_end_writeback_no_dropbehind(struct folio *folio);
 void folio_end_dropbehind(struct folio *folio);
 void folio_mark_dropbehind(struct folio *folio);
 bool folio_dontcache_edge_written(struct folio *folio, loff_t pos, size_t len);
+bool filemap_dontcache_edge_completes(struct address_space *mapping,
+				      loff_t pos, size_t len);
+void filemap_dontcache_edge_complete(struct address_space *mapping, loff_t pos);
 void folio_wait_stable(struct folio *folio);
 void __folio_mark_dirty(struct folio *folio, struct address_space *, int warn);
 void folio_account_cleaned(struct folio *folio, struct bdi_writeback *wb);
diff --git a/mm/filemap.c b/mm/filemap.c
index afba94b33dcc6..b8d34abe55671 100644
--- a/mm/filemap.c
+++ b/mm/filemap.c
@@ -1732,6 +1732,75 @@ bool folio_dontcache_edge_written(struct folio *folio, loff_t pos, size_t len)
 }
 EXPORT_SYMBOL_GPL(folio_dontcache_edge_written);
 
+/**
+ * filemap_dontcache_edge_completes - will this write complete a shared folio?
+ * @mapping: the file's mapping
+ * @pos: file offset of the first byte the write covers in the folio
+ * @len: bytes the write covers in the folio
+ *
+ * For an IOCB_DONTCACHE_DEFER write: call right before the write.
+ *
+ * Return: true if the caller must call filemap_dontcache_edge_complete() once
+ * the write, and any sync of it, is done.
+ */
+bool filemap_dontcache_edge_completes(struct address_space *mapping,
+				      loff_t pos, size_t len)
+{
+	struct folio *folio;
+	bool ret;
+
+	/* FGP_DONTCACHE keeps a mark, FGP_LOCK guards the block state. */
+	folio = __filemap_get_folio(mapping, pos >> PAGE_SHIFT,
+				    FGP_LOCK | FGP_DONTCACHE, 0);
+	if (IS_ERR(folio))
+		return false;
+	ret = folio_dontcache_edge_written(folio, pos, len);
+	folio_unlock(folio);
+	folio_put(folio);
+	return ret;
+}
+EXPORT_SYMBOL_GPL(filemap_dontcache_edge_completes);
+
+/**
+ * filemap_dontcache_edge_complete - drop a folio both writes have written
+ * @mapping: the file's mapping
+ * @pos: any byte offset within the folio
+ *
+ * Call once the write, and any sync of it, is done.  A dirty folio is marked
+ * for the writeback that cleans it, and a clean unmapped folio is dropped now.
+ * A clean mapped folio stays cached, as filemap_end_dropbehind() leaves it.
+ */
+void filemap_dontcache_edge_complete(struct address_space *mapping, loff_t pos)
+{
+	struct folio *folio;
+	bool kick = false;
+
+	folio = __filemap_get_folio(mapping, pos >> PAGE_SHIFT,
+				    FGP_LOCK | FGP_DONTCACHE, 0);
+	if (IS_ERR(folio))
+		return;
+	if (!filemap_dontcache_edge_unit(mapping, folio))
+		goto out;
+	/* Writeback in flight was submitted unmarked and would not drop it. */
+	folio_wait_writeback(folio);
+	if (folio_test_dirty(folio)) {
+		folio_mark_dropbehind(folio);
+		kick = true;
+	} else if (!folio_mapped(folio)) {
+		/*
+		 * Drop it now, unmarked: a mark cleared outside the i_pages
+		 * lock could miss a store that counts the folio.
+		 */
+		folio_unmap_invalidate(mapping, folio, 0);
+	}
+out:
+	folio_unlock(folio);
+	folio_put(folio);
+	if (kick)
+		filemap_dontcache_kick_writeback(mapping);
+}
+EXPORT_SYMBOL_GPL(filemap_dontcache_edge_complete);
+
 /**
  * folio_end_writeback_no_dropbehind - End writeback against a folio.
  * @folio: The folio.
@@ -2130,8 +2199,8 @@ struct folio *__filemap_get_folio_mpol(struct address_space *mapping,
 	 * completes it.
 	 */
 	if ((fgp_flags & FGP_DONTCACHE) && (fgp_flags & FGP_DONTCACHE_EDGE) &&
-	    (fgp_flags & FGP_LOCK) && folio_test_uptodate(folio) &&
-	    filemap_dontcache_edge_unit(mapping, folio))
+	    !(fgp_flags & FGP_DONTCACHE_DEFER) && (fgp_flags & FGP_LOCK) &&
+	    folio_test_uptodate(folio) && filemap_dontcache_edge_unit(mapping, folio))
 		folio_mark_dropbehind(folio);
 no_page:
 	if (!folio && (fgp_flags & FGP_CREAT)) {
-- 
2.52.0

[PATCH v11 7/7] NFSD: drop a shared boundary page only once both WRITEs have written it

[Mike Snitzer]

NFSD splits a misaligned direct-mode WRITE into a direct middle and
buffered DONTCACHE ends.  Each end covers its first or last unit, the
smallest folio the mapping allows, only partially.  The other WRITE that
covers the rest of that unit shares it.

Issue the buffered ends with IOCB_DONTCACHE_DEFER.  Before writing an
end, ask filemap_dontcache_edge_completes() whether the other WRITE has
written its part of each shared unit.  Once the WRITE is done, complete
those units with filemap_dontcache_edge_complete().  If the client sent
the WRITE as FILE_SYNC or DATA_SYNC, that is after the one
vfs_fsync_range() that NFSD runs over the whole WRITE.  An UNSTABLE
WRITE has no fsync, so its units are completed while still dirty.  A
short write completes a unit only if it wrote its own part of it.

With this commit, 32 interleaved writers of 47008-byte records through
NFSD DIRECT read 0.0001 to 0.0002 blocks per WRITE.  Before, they read
0.99 blocks per WRITE.  FILE_SYNC WRITEs no longer wake the DONTCACHE
writeback kick, where they woke it 23,000 to 65,000 times a run without
IOCB_DONTCACHE_DEFER.  On a four-server pNFS flexfiles rig, the page
cache grew by 2.9 GiB during the write phase instead of about 11 GiB.

47008 bytes is the transfer size of IO500's ior-hard, in which each rank
writes interleaved records to one shared file with O_DIRECT.  The Linux
client sends an O_DIRECT write that fits in one WRITE as FILE_SYNC.

Reported-by: Jonathan Flynn <jonathan.flynn@hammerspace.com>
Reviewed-by: Jeff Layton <jlayton@kernel.org>
Signed-off-by: Mike Snitzer <snitzer@kernel.org>
---
 .../filesystems/nfs/nfsd-io-modes.rst         |  18 +++
 fs/nfsd/vfs.c                                 | 107 ++++++++++++++++--
 2 files changed, 113 insertions(+), 12 deletions(-)

diff --git a/Documentation/filesystems/nfs/nfsd-io-modes.rst b/Documentation/filesystems/nfs/nfsd-io-modes.rst
index d57b7916f6c2e..6f5941e9ea7b6 100644
--- a/Documentation/filesystems/nfs/nfsd-io-modes.rst
+++ b/Documentation/filesystems/nfs/nfsd-io-modes.rst
@@ -197,6 +197,24 @@ Misaligned WRITE:
     above. Set it to N for a workload that reads back or rewrites what
     it just wrote.
 
+    The page holding a start or end segment is shared with the other
+    WRITE that covers the rest of the page, which may arrive before,
+    after or at the same time, from any client. So that the second WRITE
+    to such a page does not have to read it back from disk, the page
+    cache keeps a page a DONTCACHE write covers only partially,
+    unmarked, for the write that completes it. NFSD asks, immediately
+    before writing a start or end segment, whether the other WRITE has
+    already written its part; if it has, NFSD marks the page DONTCACHE
+    once its own WRITE, including its fsync, is complete, and the page
+    is dropped. A page that no other WRITE completes, or that the other
+    WRITE writes at the same time, stays cached until it is reclaimed.
+    The first and last pages of a WRITE that is not split are handled
+    the same way.
+
+    None of this applies when the start and end segments are normal
+    buffered IO, or for a filesystem that does not use iomap for
+    buffered writes.
+
 Tracing:
     The nfsd_read_direct trace event marks a READ that NFSD issues as
     O_DIRECT. It records the offset and length the client requested.
diff --git a/fs/nfsd/vfs.c b/fs/nfsd/vfs.c
index 10ea334cfdab6..5d265b2294b00 100644
--- a/fs/nfsd/vfs.c
+++ b/fs/nfsd/vfs.c
@@ -1432,6 +1432,7 @@ static int wait_for_concurrent_writes(struct file *file)
 struct nfsd_write_dio_seg {
 	struct iov_iter			iter;
 	int				flags;
+	bool				claim;		/* completes shared pages */
 };
 
 static unsigned long
@@ -1451,6 +1452,79 @@ nfsd_write_dio_seg_init(struct nfsd_write_dio_seg *segment,
 		iov_iter_advance(&segment->iter, start);
 	iov_iter_truncate(&segment->iter, len);
 	segment->flags = iocb->ki_flags;
+	segment->claim = false;
+}
+
+/* The shared pages a segment completes, and how much of it was written. */
+struct nfsd_write_dio_hold {
+	loff_t	first, last;
+	size_t	first_len;	/* the segment's bytes in its first unit */
+	bool	complete_first, complete_last;
+	size_t	written;
+};
+
+/*
+ * A page that a buffered DONTCACHE segment covers only partially is shared
+ * with the other WRITE that covers the rest of the page.  The page cache keeps
+ * it, unmarked, for the write that completes it; the segment is issued
+ * IOCB_DONTCACHE_DEFER so that this WRITE, not the page cache, marks it, once
+ * the whole WRITE, its fsync included, is complete.  Ask whether the other
+ * WRITE has already written its part immediately before writing @seg: asking
+ * earlier gives the other WRITE time to write its part and find this WRITE's
+ * part missing.
+ */
+static void
+nfsd_write_dio_seg_hold(struct file *file, const struct nfsd_write_dio_seg *seg,
+			loff_t pos, size_t len, struct nfsd_write_dio_hold *hold)
+{
+	struct address_space *mapping = file->f_mapping;
+	size_t unit, last_len;
+	loff_t mask, last_unit;
+	bool one_unit, head, tail;
+
+	hold->first = pos;
+	hold->last = pos + len - 1;
+	hold->complete_first = false;
+	hold->complete_last = false;
+	hold->written = 0;
+	if (!seg->claim || !len)
+		return;
+
+	/* The shared unit is the smallest folio the mapping allows. */
+	unit = mapping_min_folio_nrbytes(mapping);
+	mask = ~(loff_t)(unit - 1);
+	one_unit = (hold->first & mask) == (hold->last & mask);
+	head = hold->first & (unit - 1);
+	tail = (hold->last + 1) & (unit - 1);
+	last_unit = hold->last & mask;
+	last_len = (hold->last & (unit - 1)) + 1;
+
+	hold->first_len = min_t(size_t, len,
+				unit - (hold->first & (unit - 1)));
+	if (head || (one_unit && tail))
+		hold->complete_first =
+			filemap_dontcache_edge_completes(mapping, hold->first,
+							 hold->first_len);
+	if (!one_unit && tail)
+		hold->complete_last =
+			filemap_dontcache_edge_completes(mapping, last_unit,
+							 last_len);
+}
+
+/* Drop the shared pages this WRITE was the second of two to write. */
+static void
+nfsd_write_dio_seg_release(struct file *file,
+			   const struct nfsd_write_dio_hold *hold)
+{
+	struct address_space *mapping = file->f_mapping;
+
+	/* A short write may have stopped inside the first unit. */
+	if (hold->complete_first && hold->written >= hold->first_len)
+		filemap_dontcache_edge_complete(mapping, hold->first);
+	/* A short write did not reach the last unit. */
+	if (hold->complete_last &&
+	    hold->written == hold->last - hold->first + 1)
+		filemap_dontcache_edge_complete(mapping, hold->last);
 }
 
 static unsigned int
@@ -1473,7 +1547,9 @@ nfsd_write_dio_iters_init(struct svc_rqst *rqstp, struct svc_fh *fhp,
 
 	if (nf->nf_file->f_op->fop_flags & FOP_DONTCACHE)
 		dontcache_flags = IOCB_DONTCACHE;
-	buffered_flags = nfsd_direct_misaligned_dontcache ? dontcache_flags : 0;
+	/* NFSD marks their shared pages itself, see nfsd_write_dio_seg_hold() */
+	buffered_flags = nfsd_direct_misaligned_dontcache && dontcache_flags ?
+			 dontcache_flags | IOCB_DONTCACHE_DEFER : 0;
 
 	/*
 	 * Check if direct I/O is feasible for this write request.
@@ -1527,7 +1603,8 @@ nfsd_write_dio_iters_init(struct svc_rqst *rqstp, struct svc_fh *fhp,
 	if (prefix) {
 		nfsd_write_dio_seg_init(&segments[nsegs], bvec,
 					nvecs, total, 0, prefix, iocb);
-		segments[nsegs++].flags |= buffered_flags;
+		segments[nsegs].flags |= buffered_flags;
+		segments[nsegs++].claim = !!buffered_flags;
 	}
 
 	nfsd_write_dio_seg_init(&segments[nsegs], bvec, nvecs,
@@ -1560,7 +1637,8 @@ nfsd_write_dio_iters_init(struct svc_rqst *rqstp, struct svc_fh *fhp,
 	if (suffix) {
 		nfsd_write_dio_seg_init(&segments[nsegs], bvec, nvecs, total,
 					prefix + middle, suffix, iocb);
-		segments[nsegs++].flags |= buffered_flags;
+		segments[nsegs].flags |= buffered_flags;
+		segments[nsegs++].claim = !!buffered_flags;
 	}
 	return nsegs;
 
@@ -1569,6 +1647,7 @@ nfsd_write_dio_iters_init(struct svc_rqst *rqstp, struct svc_fh *fhp,
 	nfsd_write_dio_seg_init(&segments[0], bvec, nvecs, total, 0,
 				total, iocb);
 	segments[0].flags |= buffered_flags;
+	segments[0].claim = !!buffered_flags;
 	return 1;
 }
 
@@ -1577,12 +1656,13 @@ nfsd_direct_write(struct svc_rqst *rqstp, struct svc_fh *fhp,
 		  struct nfsd_file *nf, unsigned int nvecs,
 		  unsigned long *cnt, struct kiocb *kiocb)
 {
+	struct nfsd_write_dio_hold holds[3];
 	struct nfsd_write_dio_seg segments[3];
 	struct file *file = nf->nf_file;
 	loff_t start = kiocb->ki_pos;
-	unsigned int nsegs, i;
+	unsigned int nsegs, nheld = 0, i;
+	ssize_t host_err = 0;
 	int sync_flags;
-	ssize_t host_err;
 	size_t expected;
 
 	nsegs = nfsd_write_dio_iters_init(rqstp, fhp, nf, rqstp->rq_bvec,
@@ -1612,23 +1692,26 @@ nfsd_direct_write(struct svc_rqst *rqstp, struct svc_fh *fhp,
 					kiocb->ki_flags);
 
 		expected = iov_iter_count(&segments[i].iter);
-
+		nfsd_write_dio_seg_hold(file, &segments[i], kiocb->ki_pos,
+					expected, &holds[nheld]);
 		host_err = vfs_iocb_iter_write(file, kiocb, &segments[i].iter);
 		if (host_err < 0)
-			return host_err;
+			break;
+		holds[nheld++].written = host_err;
 		*cnt += host_err;
 		if (host_err < (ssize_t)expected)
 			break;	/* partial write */
 	}
 
-	if (sync_flags && *cnt) {
+	if (host_err >= 0 && sync_flags && *cnt)
 		host_err = vfs_fsync_range(file, start, start + *cnt - 1,
 					   !(sync_flags & IOCB_SYNC));
-		if (host_err < 0)
-			return host_err;
-	}
 
-	return 0;
+	/* Complete the shared units once the WRITE and its fsync are done. */
+	for (i = 0; i < nheld; i++)
+		nfsd_write_dio_seg_release(file, &holds[i]);
+
+	return host_err < 0 ? host_err : 0;
 }
 
 /**
-- 
2.52.0