[PATCH RFC v2 0/2] block: enable RWF_DONTCACHE for block devices

[PATCH RFC v2 0/2] block: enable RWF_DONTCACHE for block devices

[Tal Zussman]

Add support for using RWF_DONTCACHE with block devices and other
buffer_head-based I/O.

Dropbehind pruning needs to be done in non-IRQ context, but block
devices complete writeback in IRQ context. To fix this, we first defer
dropbehind completion initiated from IRQ context by scheduling a work
item on the system workqueue to process a batch of folios.

Then, fix up the block_write_begin() interface to allow issuing
RWF_DONTCACHE I/Os.

This support is useful for databases that operate on raw block devices,
among other userspace applications.

I tested this (with CONFIG_BUFFER_HEAD=y) for reads and writes on a
single block device on a VM, so results may be noisy.

Reads were tested on the root partition with a 45GB range (~2x RAM).
Writes were tested on a disabled swap parition (~1GB) in a memcg of size
244MB to force reclaim pressure.

Results: 

===== READS (/dev/nvme0n1p2) =====
 sec   normal MB/s  dontcache MB/s
----  ------------  --------------
   1         993.9          1799.6
   2         992.8          1693.8
   3         923.4          2565.9
   4        1013.5          3917.3
   5        1557.9          2438.2
   6        2363.4          1844.3
   7        1447.9          2048.6
   8         899.4          1951.7
   9        1246.8          1756.1
  10        1139.0          1665.6
  11        1089.7          1707.7
  12        1270.4          1736.5
  13        1244.0          1756.3
  14        1389.7          1566.2
----  ------------  --------------
 avg        1258.0          2005.4  (+59%)

==== WRITES (/dev/nvme0n1p3) =====
 sec   normal MB/s  dontcache MB/s
----  ------------  --------------
   1        2396.1          9670.6
   2        8444.8          9391.5
   3         770.8          9400.8
   4          61.5          9565.9
   5        7701.0          8832.6
   6        8634.3          9912.9
   7         469.2          9835.4
   8        8588.5          9587.2
   9        8602.2          9334.8
  10         591.1          8678.8
  11        8528.7          3847.0
----  ------------  --------------
 avg        4981.7          8914.3  (+79%)

---
Changes in v2:
- Add R-b from Jan Kara for 2/2.
- Add patch to defer dropbehind completion from IRQ context via a work
  item (1/2).
- Add initial performance numbers to cover letter.
- Link to v1: https://lore.kernel.org/r/20260218-blk-dontcache-v1-1-fad6675ef71f@columbia.edu

---
Tal Zussman (2):
      filemap: defer dropbehind invalidation from IRQ context
      block: enable RWF_DONTCACHE for block devices

 block/fops.c                |  4 +--
 fs/bfs/file.c               |  2 +-
 fs/buffer.c                 | 12 ++++---
 fs/exfat/inode.c            |  2 +-
 fs/ext2/inode.c             |  2 +-
 fs/jfs/inode.c              |  2 +-
 fs/minix/inode.c            |  2 +-
 fs/nilfs2/inode.c           |  2 +-
 fs/nilfs2/recovery.c        |  2 +-
 fs/ntfs3/inode.c            |  2 +-
 fs/omfs/file.c              |  2 +-
 fs/udf/inode.c              |  2 +-
 fs/ufs/inode.c              |  2 +-
 include/linux/buffer_head.h |  5 +--
 mm/filemap.c                | 84 ++++++++++++++++++++++++++++++++++++++++++---
 15 files changed, 103 insertions(+), 24 deletions(-)
---
base-commit: 05f7e89ab9731565d8a62e3b5d1ec206485eeb0b
change-id: 20260218-blk-dontcache-338133dd045e

Best regards,
-- 
Tal Zussman <tz2294@columbia.edu>

[PATCH RFC v2 1/2] filemap: defer dropbehind invalidation from IRQ context

[Tal Zussman]

folio_end_dropbehind() is called from folio_end_writeback(), which can
run in IRQ context through buffer_head completion.

Previously, when folio_end_dropbehind() detected !in_task(), it skipped
the invalidation entirely. This meant that folios marked for dropbehind
via RWF_DONTCACHE would remain in the page cache after writeback when
completed from IRQ context, defeating the purpose of using it.

Fix this by deferring the dropbehind invalidation to a work item.  When
folio_end_dropbehind() is called from IRQ context, the folio is added to
a global folio_batch and the work item is scheduled. The worker drains
the batch, locking each folio and calling filemap_end_dropbehind(), and
re-drains if new folios arrived while processing.

This unblocks enabling RWF_UNCACHED for block devices and other
buffer_head-based I/O.

Signed-off-by: Tal Zussman <tz2294@columbia.edu>
---
 mm/filemap.c | 84 ++++++++++++++++++++++++++++++++++++++++++++++++++++++++----
 1 file changed, 79 insertions(+), 5 deletions(-)

diff --git a/mm/filemap.c b/mm/filemap.c
index ebd75684cb0a..6263f35c5d13 100644
--- a/mm/filemap.c
+++ b/mm/filemap.c
@@ -1085,6 +1085,8 @@ static const struct ctl_table filemap_sysctl_table[] = {
 	}
 };
 
+static void __init dropbehind_init(void);
+
 void __init pagecache_init(void)
 {
 	int i;
@@ -1092,6 +1094,7 @@ void __init pagecache_init(void)
 	for (i = 0; i < PAGE_WAIT_TABLE_SIZE; i++)
 		init_waitqueue_head(&folio_wait_table[i]);
 
+	dropbehind_init();
 	page_writeback_init();
 	register_sysctl_init("vm", filemap_sysctl_table);
 }
@@ -1613,23 +1616,94 @@ static void filemap_end_dropbehind(struct folio *folio)
  * If folio was marked as dropbehind, then pages should be dropped when writeback
  * completes. Do that now. If we fail, it's likely because of a big folio -
  * just reset dropbehind for that case and latter completions should invalidate.
+ *
+ * When called from IRQ context (e.g. buffer_head completion), we cannot lock
+ * the folio and invalidate. Defer to a workqueue so that callers like
+ * end_buffer_async_write() that complete in IRQ context still get their folios
+ * pruned.
  */
+static DEFINE_SPINLOCK(dropbehind_lock);
+static struct folio_batch dropbehind_fbatch;
+static struct work_struct dropbehind_work;
+
+static void dropbehind_work_fn(struct work_struct *w)
+{
+	struct folio_batch fbatch;
+
+again:
+	spin_lock_irq(&dropbehind_lock);
+	fbatch = dropbehind_fbatch;
+	folio_batch_reinit(&dropbehind_fbatch);
+	spin_unlock_irq(&dropbehind_lock);
+
+	for (int i = 0; i < folio_batch_count(&fbatch); i++) {
+		struct folio *folio = fbatch.folios[i];
+
+		if (folio_trylock(folio)) {
+			filemap_end_dropbehind(folio);
+			folio_unlock(folio);
+		}
+		folio_put(folio);
+	}
+
+	/* Drain folios that were added while we were processing. */
+	spin_lock_irq(&dropbehind_lock);
+	if (folio_batch_count(&dropbehind_fbatch)) {
+		spin_unlock_irq(&dropbehind_lock);
+		goto again;
+	}
+	spin_unlock_irq(&dropbehind_lock);
+}
+
+static void __init dropbehind_init(void)
+{
+	folio_batch_init(&dropbehind_fbatch);
+	INIT_WORK(&dropbehind_work, dropbehind_work_fn);
+}
+
+static void folio_end_dropbehind_irq(struct folio *folio)
+{
+	unsigned long flags;
+
+	spin_lock_irqsave(&dropbehind_lock, flags);
+
+	/* If there is no space in the folio_batch, skip the invalidation. */
+	if (!folio_batch_space(&dropbehind_fbatch)) {
+		spin_unlock_irqrestore(&dropbehind_lock, flags);
+		return;
+	}
+
+	folio_get(folio);
+	folio_batch_add(&dropbehind_fbatch, folio);
+	spin_unlock_irqrestore(&dropbehind_lock, flags);
+
+	schedule_work(&dropbehind_work);
+}
+
 void folio_end_dropbehind(struct folio *folio)
 {
 	if (!folio_test_dropbehind(folio))
 		return;
 
 	/*
-	 * Hitting !in_task() should not happen off RWF_DONTCACHE writeback,
-	 * but can happen if normal writeback just happens to find dirty folios
-	 * that were created as part of uncached writeback, and that writeback
-	 * would otherwise not need non-IRQ handling. Just skip the
-	 * invalidation in that case.
+	 * Hitting !in_task() can happen for IO completed from IRQ contexts or
+	 * if normal writeback just happens to find dirty folios that were
+	 * created as part of uncached writeback, and that writeback would
+	 * otherwise not need non-IRQ handling.
 	 */
 	if (in_task() && folio_trylock(folio)) {
 		filemap_end_dropbehind(folio);
 		folio_unlock(folio);
+		return;
 	}
+
+	/*
+	 * In IRQ context we cannot lock the folio or call into the
+	 * invalidation path. Defer to a workqueue. This happens for
+	 * buffer_head-based writeback which runs from bio IRQ context.
+	 */
+	if (!in_task())
+		folio_end_dropbehind_irq(folio);
 }
 EXPORT_SYMBOL_GPL(folio_end_dropbehind);
 

-- 
2.39.5

Re: [PATCH RFC v2 1/2] filemap: defer dropbehind invalidation from IRQ context

[Jens Axboe]

On 2/25/26 3:40 PM, Tal Zussman wrote:
> folio_end_dropbehind() is called from folio_end_writeback(), which can
> run in IRQ context through buffer_head completion.
> 
> Previously, when folio_end_dropbehind() detected !in_task(), it skipped
> the invalidation entirely. This meant that folios marked for dropbehind
> via RWF_DONTCACHE would remain in the page cache after writeback when
> completed from IRQ context, defeating the purpose of using it.
> 
> Fix this by deferring the dropbehind invalidation to a work item.  When
> folio_end_dropbehind() is called from IRQ context, the folio is added to
> a global folio_batch and the work item is scheduled. The worker drains
> the batch, locking each folio and calling filemap_end_dropbehind(), and
> re-drains if new folios arrived while processing.
> 
> This unblocks enabling RWF_UNCACHED for block devices and other
> buffer_head-based I/O.
> 
> Signed-off-by: Tal Zussman <tz2294@columbia.edu>
> ---
>  mm/filemap.c | 84 ++++++++++++++++++++++++++++++++++++++++++++++++++++++++----
>  1 file changed, 79 insertions(+), 5 deletions(-)
> 
> diff --git a/mm/filemap.c b/mm/filemap.c
> index ebd75684cb0a..6263f35c5d13 100644
> --- a/mm/filemap.c
> +++ b/mm/filemap.c
> @@ -1085,6 +1085,8 @@ static const struct ctl_table filemap_sysctl_table[] = {
>  	}
>  };
>  
> +static void __init dropbehind_init(void);
> +
>  void __init pagecache_init(void)
>  {
>  	int i;
> @@ -1092,6 +1094,7 @@ void __init pagecache_init(void)
>  	for (i = 0; i < PAGE_WAIT_TABLE_SIZE; i++)
>  		init_waitqueue_head(&folio_wait_table[i]);
>  
> +	dropbehind_init();
>  	page_writeback_init();
>  	register_sysctl_init("vm", filemap_sysctl_table);
>  }
> @@ -1613,23 +1616,94 @@ static void filemap_end_dropbehind(struct folio *folio)
>   * If folio was marked as dropbehind, then pages should be dropped when writeback
>   * completes. Do that now. If we fail, it's likely because of a big folio -
>   * just reset dropbehind for that case and latter completions should invalidate.
> + *
> + * When called from IRQ context (e.g. buffer_head completion), we cannot lock
> + * the folio and invalidate. Defer to a workqueue so that callers like
> + * end_buffer_async_write() that complete in IRQ context still get their folios
> + * pruned.
>   */
> +static DEFINE_SPINLOCK(dropbehind_lock);
> +static struct folio_batch dropbehind_fbatch;
> +static struct work_struct dropbehind_work;
> +
> +static void dropbehind_work_fn(struct work_struct *w)
> +{
> +	struct folio_batch fbatch;
> +
> +again:
> +	spin_lock_irq(&dropbehind_lock);
> +	fbatch = dropbehind_fbatch;
> +	folio_batch_reinit(&dropbehind_fbatch);
> +	spin_unlock_irq(&dropbehind_lock);
> +
> +	for (int i = 0; i < folio_batch_count(&fbatch); i++) {
> +		struct folio *folio = fbatch.folios[i];
> +
> +		if (folio_trylock(folio)) {
> +			filemap_end_dropbehind(folio);
> +			folio_unlock(folio);
> +		}
> +		folio_put(folio);
> +	}
> +
> +	/* Drain folios that were added while we were processing. */
> +	spin_lock_irq(&dropbehind_lock);
> +	if (folio_batch_count(&dropbehind_fbatch)) {
> +		spin_unlock_irq(&dropbehind_lock);
> +		goto again;
> +	}
> +	spin_unlock_irq(&dropbehind_lock);
> +}
> +
> +static void __init dropbehind_init(void)
> +{
> +	folio_batch_init(&dropbehind_fbatch);
> +	INIT_WORK(&dropbehind_work, dropbehind_work_fn);
> +}
> +
> +static void folio_end_dropbehind_irq(struct folio *folio)
> +{
> +	unsigned long flags;
> +
> +	spin_lock_irqsave(&dropbehind_lock, flags);
> +
> +	/* If there is no space in the folio_batch, skip the invalidation. */
> +	if (!folio_batch_space(&dropbehind_fbatch)) {
> +		spin_unlock_irqrestore(&dropbehind_lock, flags);
> +		return;
> +	}
> +
> +	folio_get(folio);
> +	folio_batch_add(&dropbehind_fbatch, folio);
> +	spin_unlock_irqrestore(&dropbehind_lock, flags);
> +
> +	schedule_work(&dropbehind_work);
> +}

How well does this scale? I did a patch basically the same as this, but
not using a folio batch though. But the main sticking point was
dropbehind_lock contention, to the point where I left it alone and
thought "ok maybe we just do this when we're done with the awful
buffer_head stuff". What happens if you have N threads doing IO at the
same time to N block devices? I suspect it'll look absolutely terrible,
as each thread will be banging on that dropbehind_lock.

One solution could potentially be to use per-cpu lists for this. If you
have N threads working on separate block devices, they will tend to be
sticky to their CPU anyway.

tldr - I don't believe the above will work well enough to scale
appropriately.

Let me know if you want me to test this on my big box, it's got a bunch
of drives and CPUs to match.

I did a patch exactly matching this, youc an probably find it 

>  void folio_end_dropbehind(struct folio *folio)
>  {
>  	if (!folio_test_dropbehind(folio))
>  		return;
>  
>  	/*
> -	 * Hitting !in_task() should not happen off RWF_DONTCACHE writeback,
> -	 * but can happen if normal writeback just happens to find dirty folios
> -	 * that were created as part of uncached writeback, and that writeback
> -	 * would otherwise not need non-IRQ handling. Just skip the
> -	 * invalidation in that case.
> +	 * Hitting !in_task() can happen for IO completed from IRQ contexts or
> +	 * if normal writeback just happens to find dirty folios that were
> +	 * created as part of uncached writeback, and that writeback would
> +	 * otherwise not need non-IRQ handling.
>  	 */
>  	if (in_task() && folio_trylock(folio)) {
>  		filemap_end_dropbehind(folio);
>  		folio_unlock(folio);
> +		return;
>  	}
> +
> +	/*
> +	 * In IRQ context we cannot lock the folio or call into the
> +	 * invalidation path. Defer to a workqueue. This happens for
> +	 * buffer_head-based writeback which runs from bio IRQ context.
> +	 */
> +	if (!in_task())
> +		folio_end_dropbehind_irq(folio);
>  }

Ideally we'd have the caller be responsible for this, rather than put it
inside folio_end_dropbehind().

-- 
Jens Axboe

Re: [PATCH RFC v2 1/2] filemap: defer dropbehind invalidation from IRQ context

[Tal Zussman]

On Wed, Feb 25, 2026 at 5:52 PM Jens Axboe <axboe@kernel.dk> wrote:
> On 2/25/26 3:40 PM, Tal Zussman wrote:
> > folio_end_dropbehind() is called from folio_end_writeback(), which can
> > run in IRQ context through buffer_head completion.
> >
> > Previously, when folio_end_dropbehind() detected !in_task(), it skipped
> > the invalidation entirely. This meant that folios marked for dropbehind
> > via RWF_DONTCACHE would remain in the page cache after writeback when
> > completed from IRQ context, defeating the purpose of using it.
> >
> > Fix this by deferring the dropbehind invalidation to a work item.  When
> > folio_end_dropbehind() is called from IRQ context, the folio is added to
> > a global folio_batch and the work item is scheduled. The worker drains
> > the batch, locking each folio and calling filemap_end_dropbehind(), and
> > re-drains if new folios arrived while processing.
> >
> > This unblocks enabling RWF_UNCACHED for block devices and other
> > buffer_head-based I/O.
> >
> > Signed-off-by: Tal Zussman <tz2294@columbia.edu>
> > ---
> >  mm/filemap.c | 84 ++++++++++++++++++++++++++++++++++++++++++++++++++++++++----
> >  1 file changed, 79 insertions(+), 5 deletions(-)
> >
> > diff --git a/mm/filemap.c b/mm/filemap.c
> > index ebd75684cb0a..6263f35c5d13 100644
> > --- a/mm/filemap.c
> > +++ b/mm/filemap.c
> > @@ -1085,6 +1085,8 @@ static const struct ctl_table filemap_sysctl_table[] = {
> >   }
> >  };
> >
> > +static void __init dropbehind_init(void);
> > +
> >  void __init pagecache_init(void)
> >  {
> >   int i;
> > @@ -1092,6 +1094,7 @@ void __init pagecache_init(void)
> >   for (i = 0; i < PAGE_WAIT_TABLE_SIZE; i++)
> >   init_waitqueue_head(&folio_wait_table[i]);
> >
> > + dropbehind_init();
> >   page_writeback_init();
> >   register_sysctl_init("vm", filemap_sysctl_table);
> >  }
> > @@ -1613,23 +1616,94 @@ static void filemap_end_dropbehind(struct folio *folio)
> >   * If folio was marked as dropbehind, then pages should be dropped when writeback
> >   * completes. Do that now. If we fail, it's likely because of a big folio -
> >   * just reset dropbehind for that case and latter completions should invalidate.
> > + *
> > + * When called from IRQ context (e.g. buffer_head completion), we cannot lock
> > + * the folio and invalidate. Defer to a workqueue so that callers like
> > + * end_buffer_async_write() that complete in IRQ context still get their folios
> > + * pruned.
> >   */
> > +static DEFINE_SPINLOCK(dropbehind_lock);
> > +static struct folio_batch dropbehind_fbatch;
> > +static struct work_struct dropbehind_work;
> > +
> > +static void dropbehind_work_fn(struct work_struct *w)
> > +{
> > + struct folio_batch fbatch;
> > +
> > +again:
> > + spin_lock_irq(&dropbehind_lock);
> > + fbatch = dropbehind_fbatch;
> > + folio_batch_reinit(&dropbehind_fbatch);
> > + spin_unlock_irq(&dropbehind_lock);
> > +
> > + for (int i = 0; i < folio_batch_count(&fbatch); i++) {
> > + struct folio *folio = fbatch.folios[i];
> > +
> > + if (folio_trylock(folio)) {
> > + filemap_end_dropbehind(folio);
> > + folio_unlock(folio);
> > + }
> > + folio_put(folio);
> > + }
> > +
> > + /* Drain folios that were added while we were processing. */
> > + spin_lock_irq(&dropbehind_lock);
> > + if (folio_batch_count(&dropbehind_fbatch)) {
> > + spin_unlock_irq(&dropbehind_lock);
> > + goto again;
> > + }
> > + spin_unlock_irq(&dropbehind_lock);
> > +}
> > +
> > +static void __init dropbehind_init(void)
> > +{
> > + folio_batch_init(&dropbehind_fbatch);
> > + INIT_WORK(&dropbehind_work, dropbehind_work_fn);
> > +}
> > +
> > +static void folio_end_dropbehind_irq(struct folio *folio)
> > +{
> > + unsigned long flags;
> > +
> > + spin_lock_irqsave(&dropbehind_lock, flags);
> > +
> > + /* If there is no space in the folio_batch, skip the invalidation. */
> > + if (!folio_batch_space(&dropbehind_fbatch)) {
> > + spin_unlock_irqrestore(&dropbehind_lock, flags);
> > + return;
> > + }
> > +
> > + folio_get(folio);
> > + folio_batch_add(&dropbehind_fbatch, folio);
> > + spin_unlock_irqrestore(&dropbehind_lock, flags);
> > +
> > + schedule_work(&dropbehind_work);
> > +}
>
> How well does this scale? I did a patch basically the same as this, but
> not using a folio batch though. But the main sticking point was
> dropbehind_lock contention, to the point where I left it alone and
> thought "ok maybe we just do this when we're done with the awful
> buffer_head stuff". What happens if you have N threads doing IO at the
> same time to N block devices? I suspect it'll look absolutely terrible,
> as each thread will be banging on that dropbehind_lock.
>
> One solution could potentially be to use per-cpu lists for this. If you
> have N threads working on separate block devices, they will tend to be
> sticky to their CPU anyway.
>
> tldr - I don't believe the above will work well enough to scale
> appropriately.
>
> Let me know if you want me to test this on my big box, it's got a bunch
> of drives and CPUs to match.
>
> I did a patch exactly matching this, youc an probably find it

Yep, that makes sense. I think a per-cpu folio_batch, spinlock, and
work_struct would solve this (assuming that's what you meant by per-cpu lists)
and would be simple enough to implement. I can put that together and send it
tomorrow. I'll see if I can find your patch too.

Any testing you can do on that version would be very appreciated! I'm
unfortunately disk-limited for the moment...

> >  void folio_end_dropbehind(struct folio *folio)
> >  {
> >   if (!folio_test_dropbehind(folio))
> >   return;
> >
> >   /*
> > - * Hitting !in_task() should not happen off RWF_DONTCACHE writeback,
> > - * but can happen if normal writeback just happens to find dirty folios
> > - * that were created as part of uncached writeback, and that writeback
> > - * would otherwise not need non-IRQ handling. Just skip the
> > - * invalidation in that case.
> > + * Hitting !in_task() can happen for IO completed from IRQ contexts or
> > + * if normal writeback just happens to find dirty folios that were
> > + * created as part of uncached writeback, and that writeback would
> > + * otherwise not need non-IRQ handling.
> >   */
> >   if (in_task() && folio_trylock(folio)) {
> >   filemap_end_dropbehind(folio);
> >   folio_unlock(folio);
> > + return;
> >   }
> > +
> > + /*
> > + * In IRQ context we cannot lock the folio or call into the
> > + * invalidation path. Defer to a workqueue. This happens for
> > + * buffer_head-based writeback which runs from bio IRQ context.
> > + */
> > + if (!in_task())
> > + folio_end_dropbehind_irq(folio);
> >  }
>
> Ideally we'd have the caller be responsible for this, rather than put it
> inside folio_end_dropbehind().
>
> --
> Jens Axboe

Re: [PATCH RFC v2 1/2] filemap: defer dropbehind invalidation from IRQ context

[Jens Axboe]

On 2/25/26 6:38 PM, Tal Zussman wrote:
> On Wed, Feb 25, 2026 at 5:52?PM Jens Axboe <axboe@kernel.dk> wrote:
>> On 2/25/26 3:40 PM, Tal Zussman wrote:
>>> folio_end_dropbehind() is called from folio_end_writeback(), which can
>>> run in IRQ context through buffer_head completion.
>>>
>>> Previously, when folio_end_dropbehind() detected !in_task(), it skipped
>>> the invalidation entirely. This meant that folios marked for dropbehind
>>> via RWF_DONTCACHE would remain in the page cache after writeback when
>>> completed from IRQ context, defeating the purpose of using it.
>>>
>>> Fix this by deferring the dropbehind invalidation to a work item.  When
>>> folio_end_dropbehind() is called from IRQ context, the folio is added to
>>> a global folio_batch and the work item is scheduled. The worker drains
>>> the batch, locking each folio and calling filemap_end_dropbehind(), and
>>> re-drains if new folios arrived while processing.
>>>
>>> This unblocks enabling RWF_UNCACHED for block devices and other
>>> buffer_head-based I/O.
>>>
>>> Signed-off-by: Tal Zussman <tz2294@columbia.edu>
>>> ---
>>>  mm/filemap.c | 84 ++++++++++++++++++++++++++++++++++++++++++++++++++++++++----
>>>  1 file changed, 79 insertions(+), 5 deletions(-)
>>>
>>> diff --git a/mm/filemap.c b/mm/filemap.c
>>> index ebd75684cb0a..6263f35c5d13 100644
>>> --- a/mm/filemap.c
>>> +++ b/mm/filemap.c
>>> @@ -1085,6 +1085,8 @@ static const struct ctl_table filemap_sysctl_table[] = {
>>>   }
>>>  };
>>>
>>> +static void __init dropbehind_init(void);
>>> +
>>>  void __init pagecache_init(void)
>>>  {
>>>   int i;
>>> @@ -1092,6 +1094,7 @@ void __init pagecache_init(void)
>>>   for (i = 0; i < PAGE_WAIT_TABLE_SIZE; i++)
>>>   init_waitqueue_head(&folio_wait_table[i]);
>>>
>>> + dropbehind_init();
>>>   page_writeback_init();
>>>   register_sysctl_init("vm", filemap_sysctl_table);
>>>  }
>>> @@ -1613,23 +1616,94 @@ static void filemap_end_dropbehind(struct folio *folio)
>>>   * If folio was marked as dropbehind, then pages should be dropped when writeback
>>>   * completes. Do that now. If we fail, it's likely because of a big folio -
>>>   * just reset dropbehind for that case and latter completions should invalidate.
>>> + *
>>> + * When called from IRQ context (e.g. buffer_head completion), we cannot lock
>>> + * the folio and invalidate. Defer to a workqueue so that callers like
>>> + * end_buffer_async_write() that complete in IRQ context still get their folios
>>> + * pruned.
>>>   */
>>> +static DEFINE_SPINLOCK(dropbehind_lock);
>>> +static struct folio_batch dropbehind_fbatch;
>>> +static struct work_struct dropbehind_work;
>>> +
>>> +static void dropbehind_work_fn(struct work_struct *w)
>>> +{
>>> + struct folio_batch fbatch;
>>> +
>>> +again:
>>> + spin_lock_irq(&dropbehind_lock);
>>> + fbatch = dropbehind_fbatch;
>>> + folio_batch_reinit(&dropbehind_fbatch);
>>> + spin_unlock_irq(&dropbehind_lock);
>>> +
>>> + for (int i = 0; i < folio_batch_count(&fbatch); i++) {
>>> + struct folio *folio = fbatch.folios[i];
>>> +
>>> + if (folio_trylock(folio)) {
>>> + filemap_end_dropbehind(folio);
>>> + folio_unlock(folio);
>>> + }
>>> + folio_put(folio);
>>> + }
>>> +
>>> + /* Drain folios that were added while we were processing. */
>>> + spin_lock_irq(&dropbehind_lock);
>>> + if (folio_batch_count(&dropbehind_fbatch)) {
>>> + spin_unlock_irq(&dropbehind_lock);
>>> + goto again;
>>> + }
>>> + spin_unlock_irq(&dropbehind_lock);
>>> +}
>>> +
>>> +static void __init dropbehind_init(void)
>>> +{
>>> + folio_batch_init(&dropbehind_fbatch);
>>> + INIT_WORK(&dropbehind_work, dropbehind_work_fn);
>>> +}
>>> +
>>> +static void folio_end_dropbehind_irq(struct folio *folio)
>>> +{
>>> + unsigned long flags;
>>> +
>>> + spin_lock_irqsave(&dropbehind_lock, flags);
>>> +
>>> + /* If there is no space in the folio_batch, skip the invalidation. */
>>> + if (!folio_batch_space(&dropbehind_fbatch)) {
>>> + spin_unlock_irqrestore(&dropbehind_lock, flags);
>>> + return;
>>> + }
>>> +
>>> + folio_get(folio);
>>> + folio_batch_add(&dropbehind_fbatch, folio);
>>> + spin_unlock_irqrestore(&dropbehind_lock, flags);
>>> +
>>> + schedule_work(&dropbehind_work);
>>> +}
>>
>> How well does this scale? I did a patch basically the same as this, but
>> not using a folio batch though. But the main sticking point was
>> dropbehind_lock contention, to the point where I left it alone and
>> thought "ok maybe we just do this when we're done with the awful
>> buffer_head stuff". What happens if you have N threads doing IO at the
>> same time to N block devices? I suspect it'll look absolutely terrible,
>> as each thread will be banging on that dropbehind_lock.
>>
>> One solution could potentially be to use per-cpu lists for this. If you
>> have N threads working on separate block devices, they will tend to be
>> sticky to their CPU anyway.
>>
>> tldr - I don't believe the above will work well enough to scale
>> appropriately.
>>
>> Let me know if you want me to test this on my big box, it's got a bunch
>> of drives and CPUs to match.
>>
>> I did a patch exactly matching this, youc an probably find it
> 
> Yep, that makes sense. I think a per-cpu folio_batch, spinlock, and
> work_struct would solve this (assuming that's what you meant by
> per-cpu lists) and would be simple enough to implement. I can put that
> together and send it tomorrow. I'll see if I can find your patch too.

Was just looking for my patch as well... I don't think I ever posted it,
because I didn't like it very much. It's probably sitting in my git tree
somewhere.

But it looks very much the same as yours, modulo the folio batching.

One thing to keep in mind with per-cpu lists and then a per-cpu work
item is that you will potentially have all of them running. Hopefully
they can do that without burning too much CPU. However, might be more
useful to have one per node or something like that, provided it can keep
up, and just have that worker iterate the lists in that node. But we can
experiment with that, I'd say just do the naive version first which is
basically this patch but turned into a per-cpu collection of
lock/list/work_item.

> Any testing you can do on that version would be very appreciated! I'm
> unfortunately disk-limited for the moment...

No problem - I've got 32 drives in that box, and can hit about
230-240GB/sec of bandwidth off those drives. It'll certainly spot any
issues with scaling this and having many threads running uncached IO.

-- 
Jens Axboe

Re: [PATCH RFC v2 1/2] filemap: defer dropbehind invalidation from IRQ context

[Matthew Wilcox]

On Wed, Feb 25, 2026 at 03:52:41PM -0700, Jens Axboe wrote:
> How well does this scale? I did a patch basically the same as this, but
> not using a folio batch though. But the main sticking point was
> dropbehind_lock contention, to the point where I left it alone and
> thought "ok maybe we just do this when we're done with the awful
> buffer_head stuff". What happens if you have N threads doing IO at the
> same time to N block devices? I suspect it'll look absolutely terrible,
> as each thread will be banging on that dropbehind_lock.
> 
> One solution could potentially be to use per-cpu lists for this. If you
> have N threads working on separate block devices, they will tend to be
> sticky to their CPU anyway.

Back in 2021, I had Vishal look at switching the page cache from using
hardirq-disabling locks to softirq-disabling locks [1].  Some of the
feedback (which doesn't seem to be entirely findable on the lists ...)
was that we'd be better off punting writeback completion from interrupt
context to task context and going from spin_lock_irq() to spin_lock()
rather than going to spin_lock_bh().

I recently saw something (possibly XFS?) promoting this idea again.
And now there's this.  Perhaps the time has come to process all
write-completions in task context, rather than everyone coming up with
their own workqueues to solve their little piece of the problem?

[1] https://lore.kernel.org/linux-block/20210730213630.44891-1-vishal.moola@gmail.com/

Re: [PATCH RFC v2 1/2] filemap: defer dropbehind invalidation from IRQ context

[Jens Axboe]

On 2/25/26 7:55 PM, Matthew Wilcox wrote:
> On Wed, Feb 25, 2026 at 03:52:41PM -0700, Jens Axboe wrote:
>> How well does this scale? I did a patch basically the same as this, but
>> not using a folio batch though. But the main sticking point was
>> dropbehind_lock contention, to the point where I left it alone and
>> thought "ok maybe we just do this when we're done with the awful
>> buffer_head stuff". What happens if you have N threads doing IO at the
>> same time to N block devices? I suspect it'll look absolutely terrible,
>> as each thread will be banging on that dropbehind_lock.
>>
>> One solution could potentially be to use per-cpu lists for this. If you
>> have N threads working on separate block devices, they will tend to be
>> sticky to their CPU anyway.
> 
> Back in 2021, I had Vishal look at switching the page cache from using
> hardirq-disabling locks to softirq-disabling locks [1].  Some of the
> feedback (which doesn't seem to be entirely findable on the lists ...)
> was that we'd be better off punting writeback completion from interrupt
> context to task context and going from spin_lock_irq() to spin_lock()
> rather than going to spin_lock_bh().
> 
> I recently saw something (possibly XFS?) promoting this idea again.
> And now there's this.  Perhaps the time has come to process all
> write-completions in task context, rather than everyone coming up with
> their own workqueues to solve their little piece of the problem?

Perhaps, even though the punting tends to suck... One idea I toyed with
but had to abandon due to fs freezeing was letting callers that process
completions in task context anyway just do the necessary work at that
time. There's literally nothing worse than having part of a completion
happen in IRQ, then punt parts of that to a worker, and need to wait for
the worker to finish whatever it needs to do - only to then wake the
target task. We can trivially do this in io_uring, as the actual
completion is posted from the task itself anyway. We just need to have
the task do the bottom half of the completion as well, rather than some
unrelated kthread worker.

I'd be worried a generic solution would be the worst of all worlds, as
it prevents optimizations that happen in eg iomap and other spots, where
only completions that absolutely need to happen in task context get
punted. There's a big difference between handling a completion inline vs
needing a round-trip to some worker to do it.

-- 
Jens Axboe

[PATCH RFC v2 2/2] block: enable RWF_DONTCACHE for block devices

[Tal Zussman]

Block device buffered reads and writes already pass through
filemap_read() and iomap_file_buffered_write() respectively, both of
which handle IOCB_DONTCACHE. Enable RWF_DONTCACHE for block device files
by setting FOP_DONTCACHE in def_blk_fops.

For CONFIG_BUFFER_HEAD paths, thread the kiocb through
block_write_begin() so that buffer_head-based I/O can use DONTCACHE
behavior as well. Callers without a kiocb context (e.g. nilfs2 recovery)
pass NULL, which preserves the existing behavior.

This support is useful for databases that operate on raw block devices,
among other userspace applications.

Reviewed-by: Jan Kara <jack@suse.cz>
Signed-off-by: Tal Zussman <tz2294@columbia.edu>
---
 block/fops.c                |  4 ++--
 fs/bfs/file.c               |  2 +-
 fs/buffer.c                 | 12 ++++++++----
 fs/exfat/inode.c            |  2 +-
 fs/ext2/inode.c             |  2 +-
 fs/jfs/inode.c              |  2 +-
 fs/minix/inode.c            |  2 +-
 fs/nilfs2/inode.c           |  2 +-
 fs/nilfs2/recovery.c        |  2 +-
 fs/ntfs3/inode.c            |  2 +-
 fs/omfs/file.c              |  2 +-
 fs/udf/inode.c              |  2 +-
 fs/ufs/inode.c              |  2 +-
 include/linux/buffer_head.h |  5 +++--
 14 files changed, 24 insertions(+), 19 deletions(-)

diff --git a/block/fops.c b/block/fops.c
index 4d32785b31d9..6bc727f8b252 100644
--- a/block/fops.c
+++ b/block/fops.c
@@ -505,7 +505,7 @@ static int blkdev_write_begin(const struct kiocb *iocb,
 			      unsigned len, struct folio **foliop,
 			      void **fsdata)
 {
-	return block_write_begin(mapping, pos, len, foliop, blkdev_get_block);
+	return block_write_begin(iocb, mapping, pos, len, foliop, blkdev_get_block);
 }
 
 static int blkdev_write_end(const struct kiocb *iocb,
@@ -967,7 +967,7 @@ const struct file_operations def_blk_fops = {
 	.splice_write	= iter_file_splice_write,
 	.fallocate	= blkdev_fallocate,
 	.uring_cmd	= blkdev_uring_cmd,
-	.fop_flags	= FOP_BUFFER_RASYNC,
+	.fop_flags	= FOP_BUFFER_RASYNC | FOP_DONTCACHE,
 };
 
 static __init int blkdev_init(void)
diff --git a/fs/bfs/file.c b/fs/bfs/file.c
index d33d6bde992b..f2804e38b8a7 100644
--- a/fs/bfs/file.c
+++ b/fs/bfs/file.c
@@ -177,7 +177,7 @@ static int bfs_write_begin(const struct kiocb *iocb,
 {
 	int ret;
 
-	ret = block_write_begin(mapping, pos, len, foliop, bfs_get_block);
+	ret = block_write_begin(iocb, mapping, pos, len, foliop, bfs_get_block);
 	if (unlikely(ret))
 		bfs_write_failed(mapping, pos + len);
 
diff --git a/fs/buffer.c b/fs/buffer.c
index 838c0c571022..33c3580b85d8 100644
--- a/fs/buffer.c
+++ b/fs/buffer.c
@@ -2241,14 +2241,18 @@ EXPORT_SYMBOL(block_commit_write);
  *
  * The filesystem needs to handle block truncation upon failure.
  */
-int block_write_begin(struct address_space *mapping, loff_t pos, unsigned len,
-		struct folio **foliop, get_block_t *get_block)
+int block_write_begin(const struct kiocb *iocb, struct address_space *mapping,
+		loff_t pos, unsigned len, struct folio **foliop, get_block_t *get_block)
 {
 	pgoff_t index = pos >> PAGE_SHIFT;
+	fgf_t fgp_flags = FGP_WRITEBEGIN;
 	struct folio *folio;
 	int status;
 
-	folio = __filemap_get_folio(mapping, index, FGP_WRITEBEGIN,
+	if (iocb && iocb->ki_flags & IOCB_DONTCACHE)
+		fgp_flags |= FGP_DONTCACHE;
+
+	folio = __filemap_get_folio(mapping, index, fgp_flags,
 			mapping_gfp_mask(mapping));
 	if (IS_ERR(folio))
 		return PTR_ERR(folio);
@@ -2591,7 +2595,7 @@ int cont_write_begin(const struct kiocb *iocb, struct address_space *mapping,
 		(*bytes)++;
 	}
 
-	return block_write_begin(mapping, pos, len, foliop, get_block);
+	return block_write_begin(iocb, mapping, pos, len, foliop, get_block);
 }
 EXPORT_SYMBOL(cont_write_begin);
 
diff --git a/fs/exfat/inode.c b/fs/exfat/inode.c
index f9501c3a3666..39d36e8fdfd6 100644
--- a/fs/exfat/inode.c
+++ b/fs/exfat/inode.c
@@ -456,7 +456,7 @@ static int exfat_write_begin(const struct kiocb *iocb,
 	if (unlikely(exfat_forced_shutdown(mapping->host->i_sb)))
 		return -EIO;
 
-	ret = block_write_begin(mapping, pos, len, foliop, exfat_get_block);
+	ret = block_write_begin(iocb, mapping, pos, len, foliop, exfat_get_block);
 
 	if (ret < 0)
 		exfat_write_failed(mapping, pos+len);
diff --git a/fs/ext2/inode.c b/fs/ext2/inode.c
index dbfe9098a124..11aab03de752 100644
--- a/fs/ext2/inode.c
+++ b/fs/ext2/inode.c
@@ -930,7 +930,7 @@ ext2_write_begin(const struct kiocb *iocb, struct address_space *mapping,
 {
 	int ret;
 
-	ret = block_write_begin(mapping, pos, len, foliop, ext2_get_block);
+	ret = block_write_begin(iocb, mapping, pos, len, foliop, ext2_get_block);
 	if (ret < 0)
 		ext2_write_failed(mapping, pos + len);
 	return ret;
diff --git a/fs/jfs/inode.c b/fs/jfs/inode.c
index 4709762713ef..ae52db437771 100644
--- a/fs/jfs/inode.c
+++ b/fs/jfs/inode.c
@@ -303,7 +303,7 @@ static int jfs_write_begin(const struct kiocb *iocb,
 {
 	int ret;
 
-	ret = block_write_begin(mapping, pos, len, foliop, jfs_get_block);
+	ret = block_write_begin(iocb, mapping, pos, len, foliop, jfs_get_block);
 	if (unlikely(ret))
 		jfs_write_failed(mapping, pos + len);
 
diff --git a/fs/minix/inode.c b/fs/minix/inode.c
index 51ea9bdc813f..9075c0ba2f20 100644
--- a/fs/minix/inode.c
+++ b/fs/minix/inode.c
@@ -465,7 +465,7 @@ static int minix_write_begin(const struct kiocb *iocb,
 {
 	int ret;
 
-	ret = block_write_begin(mapping, pos, len, foliop, minix_get_block);
+	ret = block_write_begin(iocb, mapping, pos, len, foliop, minix_get_block);
 	if (unlikely(ret))
 		minix_write_failed(mapping, pos + len);
 
diff --git a/fs/nilfs2/inode.c b/fs/nilfs2/inode.c
index 51bde45d5865..d9d57eeecc5d 100644
--- a/fs/nilfs2/inode.c
+++ b/fs/nilfs2/inode.c
@@ -230,7 +230,7 @@ static int nilfs_write_begin(const struct kiocb *iocb,
 	if (unlikely(err))
 		return err;
 
-	err = block_write_begin(mapping, pos, len, foliop, nilfs_get_block);
+	err = block_write_begin(iocb, mapping, pos, len, foliop, nilfs_get_block);
 	if (unlikely(err)) {
 		nilfs_write_failed(mapping, pos + len);
 		nilfs_transaction_abort(inode->i_sb);
diff --git a/fs/nilfs2/recovery.c b/fs/nilfs2/recovery.c
index a9c61d0492cb..2f5fe44bf736 100644
--- a/fs/nilfs2/recovery.c
+++ b/fs/nilfs2/recovery.c
@@ -541,7 +541,7 @@ static int nilfs_recover_dsync_blocks(struct the_nilfs *nilfs,
 		}
 
 		pos = rb->blkoff << inode->i_blkbits;
-		err = block_write_begin(inode->i_mapping, pos, blocksize,
+		err = block_write_begin(NULL, inode->i_mapping, pos, blocksize,
 					&folio, nilfs_get_block);
 		if (unlikely(err)) {
 			loff_t isize = inode->i_size;
diff --git a/fs/ntfs3/inode.c b/fs/ntfs3/inode.c
index 0a9ac5efeb67..8c788feb319e 100644
--- a/fs/ntfs3/inode.c
+++ b/fs/ntfs3/inode.c
@@ -966,7 +966,7 @@ int ntfs_write_begin(const struct kiocb *iocb, struct address_space *mapping,
 			goto out;
 	}
 
-	err = block_write_begin(mapping, pos, len, foliop,
+	err = block_write_begin(iocb, mapping, pos, len, foliop,
 				ntfs_get_block_write_begin);
 
 out:
diff --git a/fs/omfs/file.c b/fs/omfs/file.c
index 49a1de5a827f..3bade632e36e 100644
--- a/fs/omfs/file.c
+++ b/fs/omfs/file.c
@@ -317,7 +317,7 @@ static int omfs_write_begin(const struct kiocb *iocb,
 {
 	int ret;
 
-	ret = block_write_begin(mapping, pos, len, foliop, omfs_get_block);
+	ret = block_write_begin(iocb, mapping, pos, len, foliop, omfs_get_block);
 	if (unlikely(ret))
 		omfs_write_failed(mapping, pos + len);
 
diff --git a/fs/udf/inode.c b/fs/udf/inode.c
index 7fae8002344a..aec9cdc938be 100644
--- a/fs/udf/inode.c
+++ b/fs/udf/inode.c
@@ -259,7 +259,7 @@ static int udf_write_begin(const struct kiocb *iocb,
 	int ret;
 
 	if (iinfo->i_alloc_type != ICBTAG_FLAG_AD_IN_ICB) {
-		ret = block_write_begin(mapping, pos, len, foliop,
+		ret = block_write_begin(iocb, mapping, pos, len, foliop,
 					udf_get_block);
 		if (unlikely(ret))
 			udf_write_failed(mapping, pos + len);
diff --git a/fs/ufs/inode.c b/fs/ufs/inode.c
index e2b0a35de2a7..dfba985265a8 100644
--- a/fs/ufs/inode.c
+++ b/fs/ufs/inode.c
@@ -481,7 +481,7 @@ static int ufs_write_begin(const struct kiocb *iocb,
 {
 	int ret;
 
-	ret = block_write_begin(mapping, pos, len, foliop, ufs_getfrag_block);
+	ret = block_write_begin(iocb, mapping, pos, len, foliop, ufs_getfrag_block);
 	if (unlikely(ret))
 		ufs_write_failed(mapping, pos + len);
 
diff --git a/include/linux/buffer_head.h b/include/linux/buffer_head.h
index b16b88bfbc3e..4b07dec5f8eb 100644
--- a/include/linux/buffer_head.h
+++ b/include/linux/buffer_head.h
@@ -258,8 +258,9 @@ int __block_write_full_folio(struct inode *inode, struct folio *folio,
 		get_block_t *get_block, struct writeback_control *wbc);
 int block_read_full_folio(struct folio *, get_block_t *);
 bool block_is_partially_uptodate(struct folio *, size_t from, size_t count);
-int block_write_begin(struct address_space *mapping, loff_t pos, unsigned len,
-		struct folio **foliop, get_block_t *get_block);
+int block_write_begin(const struct kiocb *iocb, struct address_space *mapping,
+		loff_t pos, unsigned len, struct folio **foliop,
+		get_block_t *get_block);
 int __block_write_begin(struct folio *folio, loff_t pos, unsigned len,
 		get_block_t *get_block);
 int block_write_end(loff_t pos, unsigned len, unsigned copied, struct folio *);

-- 
2.39.5