[PATCH 1/5] ceph: Do not propagate page array emplacement errors as batch errors

[PATCH 1/5] ceph: Do not propagate page array emplacement errors as batch errors

[Sam Edwards]

When fscrypt is enabled, move_dirty_folio_in_page_array() may fail
because it needs to allocate bounce buffers to store the encrypted
versions of each folio. Each folio beyond the first allocates its bounce
buffer with GFP_NOWAIT. Failures are common (and expected) under this
allocation mode; they should flush (not abort) the batch.

However, ceph_process_folio_batch() uses the same `rc` variable for its
own return code and for capturing the return codes of its routine calls;
failing to reset `rc` back to 0 results in the error being propagated
out to the main writeback loop, which cannot actually tolerate any
errors here: once `ceph_wbc.pages` is allocated, it must be passed to
ceph_submit_write() to be freed. If it survives until the next iteration
(e.g. due to the goto being followed), ceph_allocate_page_array()'s
BUG_ON() will oops the worker. (Subsequent patches in this series make
the loop more robust.)

Note that this failure mode is currently masked due to another bug
(addressed later in this series) that prevents multiple encrypted folios
from being selected for the same write.

For now, just reset `rc` when redirtying the folio and prevent the
error from propagating. After this change, ceph_process_folio_batch() no
longer returns errors; its only remaining failure indicator is
`locked_pages == 0`, which the caller already handles correctly. The
next patch in this series addresses this.

Fixes: ce80b76dd327 ("ceph: introduce ceph_process_folio_batch() method")
Cc: stable@vger.kernel.org
Signed-off-by: Sam Edwards <CFSworks@gmail.com>
---
 fs/ceph/addr.c | 1 +
 1 file changed, 1 insertion(+)

diff --git a/fs/ceph/addr.c b/fs/ceph/addr.c
index 63b75d214210..3462df35d245 100644
--- a/fs/ceph/addr.c
+++ b/fs/ceph/addr.c
@@ -1369,6 +1369,7 @@ int ceph_process_folio_batch(struct address_space *mapping,
 		rc = move_dirty_folio_in_page_array(mapping, wbc, ceph_wbc,
 				folio);
 		if (rc) {
+			rc = 0;
 			folio_redirty_for_writepage(wbc, folio);
 			folio_unlock(folio);
 			break;
-- 
2.51.2

[PATCH 3/5] ceph: Free page array when ceph_submit_write fails

[Sam Edwards]

If `locked_pages` is zero, the page array must not be allocated:
ceph_process_folio_batch() uses `locked_pages` to decide when to
allocate `pages`, and redundant allocations trigger
ceph_allocate_page_array()'s BUG_ON(), resulting in a worker oops (and
writeback stall) or even a kernel panic. Consequently, the main loop in
ceph_writepages_start() assumes that the lifetime of `pages` is confined
to a single iteration.

The ceph_submit_write() function claims ownership of the page array on
success. But failures only redirty/unlock the pages and fail to free the
array, making the failure case in ceph_submit_write() fatal.

Free the page array in ceph_submit_write()'s error-handling 'if' block
so that the caller's invariant (that the array does not outlive the
iteration) is maintained unconditionally, allowing failures in
ceph_submit_write() to be recoverable as originally intended.

Fixes: 1551ec61dc55 ("ceph: introduce ceph_submit_write() method")
Cc: stable@vger.kernel.org
Signed-off-by: Sam Edwards <CFSworks@gmail.com>
---
 fs/ceph/addr.c | 7 +++++++
 1 file changed, 7 insertions(+)

diff --git a/fs/ceph/addr.c b/fs/ceph/addr.c
index 2b722916fb9b..91cc43950162 100644
--- a/fs/ceph/addr.c
+++ b/fs/ceph/addr.c
@@ -1466,6 +1466,13 @@ int ceph_submit_write(struct address_space *mapping,
 			unlock_page(page);
 		}
 
+		if (ceph_wbc->from_pool) {
+			mempool_free(ceph_wbc->pages, ceph_wb_pagevec_pool);
+			ceph_wbc->from_pool = false;
+		} else
+			kfree(ceph_wbc->pages);
+		ceph_wbc->pages = NULL;
+
 		ceph_osdc_put_request(req);
 		return -EIO;
 	}
-- 
2.51.2

[PATCH 5/5] ceph: Fix write storm on fscrypted files

[Sam Edwards]

CephFS stores file data across multiple RADOS objects. An object is the
atomic unit of storage, so the writeback code must clean only folios
that belong to the same object with each OSD request.

CephFS also supports RAID0-style striping of file contents: if enabled,
each object stores multiple unbroken "stripe units" covering different
portions of the file; if disabled, a "stripe unit" is simply the whole
object. The stripe unit is (usually) reported as the inode's block size.

Though the writeback logic could, in principle, lock all dirty folios
belonging to the same object, its current design is to lock only a
single stripe unit at a time. Ever since this code was first written,
it has determined this size by checking the inode's block size.
However, the relatively-new fscrypt support needed to reduce the block
size for encrypted inodes to the crypto block size (see 'fixes' commit),
which causes an unnecessarily high number of write operations (~1024x as
many, with 4MiB objects) and grossly degraded performance.

Fix this (and clarify intent) by using i_layout.stripe_unit directly in
ceph_define_write_size() so that encrypted inodes are written back with
the same number of operations as if they were unencrypted.

Fixes: 94af0470924c ("ceph: add some fscrypt guardrails")
Cc: stable@vger.kernel.org
Signed-off-by: Sam Edwards <CFSworks@gmail.com>
---
 fs/ceph/addr.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/fs/ceph/addr.c b/fs/ceph/addr.c
index b3569d44d510..cb1da8e27c2b 100644
--- a/fs/ceph/addr.c
+++ b/fs/ceph/addr.c
@@ -1000,7 +1000,8 @@ unsigned int ceph_define_write_size(struct address_space *mapping)
 {
 	struct inode *inode = mapping->host;
 	struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
-	unsigned int wsize = i_blocksize(inode);
+	struct ceph_inode_info *ci = ceph_inode(inode);
+	unsigned int wsize = ci->i_layout.stripe_unit;
 
 	if (fsc->mount_options->wsize < wsize)
 		wsize = fsc->mount_options->wsize;
-- 
2.51.2

Re: [PATCH 1/5] ceph: Do not propagate page array emplacement errors as batch errors

[Viacheslav Dubeyko]

On Tue, 2025-12-30 at 18:43 -0800, Sam Edwards wrote:
> When fscrypt is enabled, move_dirty_folio_in_page_array() may fail
> because it needs to allocate bounce buffers to store the encrypted
> versions of each folio. Each folio beyond the first allocates its bounce
> buffer with GFP_NOWAIT. Failures are common (and expected) under this
> allocation mode; they should flush (not abort) the batch.
> 
> However, ceph_process_folio_batch() uses the same `rc` variable for its
> own return code and for capturing the return codes of its routine calls;
> failing to reset `rc` back to 0 results in the error being propagated
> out to the main writeback loop, which cannot actually tolerate any
> errors here: once `ceph_wbc.pages` is allocated, it must be passed to
> ceph_submit_write() to be freed. If it survives until the next iteration
> (e.g. due to the goto being followed), ceph_allocate_page_array()'s
> BUG_ON() will oops the worker. (Subsequent patches in this series make
> the loop more robust.)

I think you are right with the fix. We have the loop here and if we already
moved some dirty folios, then we should flush it. But what if we failed on the
first one folio, then should we return no error code in this case?

> 
> Note that this failure mode is currently masked due to another bug
> (addressed later in this series) that prevents multiple encrypted folios
> from being selected for the same write.

So, maybe, this patch has been not correctly placed in the order? It will be
good to see the reproduction of the issue and which symptoms we have for this
issue. Do you have the reproduction script and call trace of the issue?

> 
> For now, just reset `rc` when redirtying the folio and prevent the
> error from propagating. After this change, ceph_process_folio_batch() no
> longer returns errors; its only remaining failure indicator is
> `locked_pages == 0`, which the caller already handles correctly. The
> next patch in this series addresses this.
> 
> Fixes: ce80b76dd327 ("ceph: introduce ceph_process_folio_batch() method")
> Cc: stable@vger.kernel.org
> Signed-off-by: Sam Edwards <CFSworks@gmail.com>
> ---
>  fs/ceph/addr.c | 1 +
>  1 file changed, 1 insertion(+)
> 
> diff --git a/fs/ceph/addr.c b/fs/ceph/addr.c
> index 63b75d214210..3462df35d245 100644
> --- a/fs/ceph/addr.c
> +++ b/fs/ceph/addr.c
> @@ -1369,6 +1369,7 @@ int ceph_process_folio_batch(struct address_space *mapping,
>  		rc = move_dirty_folio_in_page_array(mapping, wbc, ceph_wbc,
>  				folio);
>  		if (rc) {
> +			rc = 0;

I like the fix but I would like to clarify the above questions at first.

Thanks,
Slava. 

>  			folio_redirty_for_writepage(wbc, folio);
>  			folio_unlock(folio);
>  			break;

Re: [PATCH 3/5] ceph: Free page array when ceph_submit_write fails

[Viacheslav Dubeyko]

On Tue, 2025-12-30 at 18:43 -0800, Sam Edwards wrote:
> If `locked_pages` is zero, the page array must not be allocated:
> ceph_process_folio_batch() uses `locked_pages` to decide when to
> allocate `pages`, 
> 

I don't quite follow how this statement is relevant to the issue. If
`locked_pages` is zero, then ceph_submit_write() will not to be called. Do I
miss something here?

> and redundant allocations trigger
> ceph_allocate_page_array()'s BUG_ON(), resulting in a worker oops (and
> writeback stall) or even a kernel panic. Consequently, the main loop in
> ceph_writepages_start() assumes that the lifetime of `pages` is confined
> to a single iteration.

It will be great to see the reproducer script or application and call trace of
the issue. Could you please share the reproduction path and the call trace of
the issue?

> 
> The ceph_submit_write() function claims ownership of the page array on
> success. 
> 

As far as I can see, writepages_finish() should free the page array on success.

> But failures only redirty/unlock the pages and fail to free the
> array, making the failure case in ceph_submit_write() fatal.
> 
> Free the page array in ceph_submit_write()'s error-handling 'if' block
> so that the caller's invariant (that the array does not outlive the
> iteration) is maintained unconditionally, allowing failures in
> ceph_submit_write() to be recoverable as originally intended.
> 
> Fixes: 1551ec61dc55 ("ceph: introduce ceph_submit_write() method")
> Cc: stable@vger.kernel.org
> Signed-off-by: Sam Edwards <CFSworks@gmail.com>
> ---
>  fs/ceph/addr.c | 7 +++++++
>  1 file changed, 7 insertions(+)
> 
> diff --git a/fs/ceph/addr.c b/fs/ceph/addr.c
> index 2b722916fb9b..91cc43950162 100644
> --- a/fs/ceph/addr.c
> +++ b/fs/ceph/addr.c
> @@ -1466,6 +1466,13 @@ int ceph_submit_write(struct address_space *mapping,
>  			unlock_page(page);
>  		}
>  
> +		if (ceph_wbc->from_pool) {
> +			mempool_free(ceph_wbc->pages, ceph_wb_pagevec_pool);
> +			ceph_wbc->from_pool = false;
> +		} else
> +			kfree(ceph_wbc->pages);
> +		ceph_wbc->pages = NULL;

Probably, it makes sense to introduce a method ceph_free_page_array likewise to
__ceph_allocate_page_array() and to use for freeing page array in all places.

Could ceph_wbc->locked_pages be greater than zero but ceph_wbc->pages == NULL?

Thanks,
Slava.

> +
>  		ceph_osdc_put_request(req);
>  		return -EIO;
>  	}

Re: [PATCH 5/5] ceph: Fix write storm on fscrypted files

[Viacheslav Dubeyko]

On Tue, 2025-12-30 at 18:43 -0800, Sam Edwards wrote:
> CephFS stores file data across multiple RADOS objects. An object is the
> atomic unit of storage, so the writeback code must clean only folios
> that belong to the same object with each OSD request.
> 
> CephFS also supports RAID0-style striping of file contents: if enabled,
> each object stores multiple unbroken "stripe units" covering different
> portions of the file; if disabled, a "stripe unit" is simply the whole
> object. The stripe unit is (usually) reported as the inode's block size.
> 
> Though the writeback logic could, in principle, lock all dirty folios
> belonging to the same object, its current design is to lock only a
> single stripe unit at a time. Ever since this code was first written,
> it has determined this size by checking the inode's block size.
> However, the relatively-new fscrypt support needed to reduce the block
> size for encrypted inodes to the crypto block size (see 'fixes' commit),
> which causes an unnecessarily high number of write operations (~1024x as
> many, with 4MiB objects) and grossly degraded performance.

Do you have any benchmarking results that prove your point?

Thanks,
Slava.

> 
> Fix this (and clarify intent) by using i_layout.stripe_unit directly in
> ceph_define_write_size() so that encrypted inodes are written back with
> the same number of operations as if they were unencrypted.
> 
> Fixes: 94af0470924c ("ceph: add some fscrypt guardrails")
> Cc: stable@vger.kernel.org
> Signed-off-by: Sam Edwards <CFSworks@gmail.com>
> ---
>  fs/ceph/addr.c | 3 ++-
>  1 file changed, 2 insertions(+), 1 deletion(-)
> 
> diff --git a/fs/ceph/addr.c b/fs/ceph/addr.c
> index b3569d44d510..cb1da8e27c2b 100644
> --- a/fs/ceph/addr.c
> +++ b/fs/ceph/addr.c
> @@ -1000,7 +1000,8 @@ unsigned int ceph_define_write_size(struct address_space *mapping)
>  {
>  	struct inode *inode = mapping->host;
>  	struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
> -	unsigned int wsize = i_blocksize(inode);
> +	struct ceph_inode_info *ci = ceph_inode(inode);
> +	unsigned int wsize = ci->i_layout.stripe_unit;
>  
>  	if (fsc->mount_options->wsize < wsize)
>  		wsize = fsc->mount_options->wsize;

Re: [PATCH 1/5] ceph: Do not propagate page array emplacement errors as batch errors

[Sam Edwards]

On Mon, Jan 5, 2026 at 12:24 PM Viacheslav Dubeyko
<Slava.Dubeyko@ibm.com> wrote:
>
> On Tue, 2025-12-30 at 18:43 -0800, Sam Edwards wrote:
> > When fscrypt is enabled, move_dirty_folio_in_page_array() may fail
> > because it needs to allocate bounce buffers to store the encrypted
> > versions of each folio. Each folio beyond the first allocates its bounce
> > buffer with GFP_NOWAIT. Failures are common (and expected) under this
> > allocation mode; they should flush (not abort) the batch.
> >
> > However, ceph_process_folio_batch() uses the same `rc` variable for its
> > own return code and for capturing the return codes of its routine calls;
> > failing to reset `rc` back to 0 results in the error being propagated
> > out to the main writeback loop, which cannot actually tolerate any
> > errors here: once `ceph_wbc.pages` is allocated, it must be passed to
> > ceph_submit_write() to be freed. If it survives until the next iteration
> > (e.g. due to the goto being followed), ceph_allocate_page_array()'s
> > BUG_ON() will oops the worker. (Subsequent patches in this series make
> > the loop more robust.)
>

Hi Slava,

> I think you are right with the fix. We have the loop here and if we already
> moved some dirty folios, then we should flush it. But what if we failed on the
> first one folio, then should we return no error code in this case?

The case you ask about, where move_dirty_folio_in_page_array() returns
an error for the first folio, is currently not possible:
1) The only error code that move_dirty_folio_in_page_array() can
propagate is from fscrypt_encrypt_pagecache_blocks(), which it calls
with GFP_NOFS for the first folio. The latter function's doc comment
outright states:
 * The bounce page allocation is mempool-backed, so it will always succeed when
 * @gfp_flags includes __GFP_DIRECT_RECLAIM, e.g. when it's GFP_NOFS.
2) The error return isn't even reachable for the first folio because
of the BUG_ON(ceph_wbc->locked_pages == 0); line.

>
> >
> > Note that this failure mode is currently masked due to another bug
> > (addressed later in this series) that prevents multiple encrypted folios
> > from being selected for the same write.
>
> So, maybe, this patch has been not correctly placed in the order?

This crash is unmasked by patch 5 of this series. (It allows multiple
folios to be batched when fscrypt is enabled.) Patch 5 has no hard
dependency on anything else in this series, so it could -- in
principle -- be ordered first as you suggest. However, that ordering
would deliberately cause a regression in kernel stability, even if
only briefly. That's not considered good practice in my view, as it
may affect people who are trying to bisect and regression test. So the
ordering of this series is: fix the crash in the unused code first,
then fix the bug that makes it unused.

> It will be
> good to see the reproduction of the issue and which symptoms we have for this
> issue. Do you have the reproduction script and call trace of the issue?

Fair point!

Function inlining makes the call trace not very interesting:
Call trace:
 ceph_writepages_start+0x16ec/0x18e0 [ceph] ()
 do_writepages+0xb0/0x1c0
 __writeback_single_inode+0x4c/0x4d8
 writeback_sb_inodes+0x238/0x4c8
 __writeback_inodes_wb+0x64/0x120
 wb_writeback+0x320/0x3e8
 wb_workfn+0x42c/0x518
 process_one_work+0x17c/0x428
 worker_thread+0x260/0x390
 kthread+0x148/0x240
 ret_from_fork+0x10/0x20
Code: 34ffdee0 52800020 3903e7e0 17fffef4 (d4210000)
---[ end trace 0000000000000000 ]---
Kernel panic - not syncing: Oops - BUG: Fatal exception

ceph_writepages_start+0x16ec corresponds to linux-6.18.2/fs/ceph/addr.c:1222

However, these repro steps should work:
1) Apply patch 5 from this series (and no other patches)
2) Mount CephFS and activate fscrypt
3) Copy a large directory into the CephFS mount
4) After dozens of GBs transferred, you should observe the above kernel oops

Warm regards,
Sam




>
> >
> > For now, just reset `rc` when redirtying the folio and prevent the
> > error from propagating. After this change, ceph_process_folio_batch() no
> > longer returns errors; its only remaining failure indicator is
> > `locked_pages == 0`, which the caller already handles correctly. The
> > next patch in this series addresses this.
> >
> > Fixes: ce80b76dd327 ("ceph: introduce ceph_process_folio_batch() method")
> > Cc: stable@vger.kernel.org
> > Signed-off-by: Sam Edwards <CFSworks@gmail.com>
> > ---
> >  fs/ceph/addr.c | 1 +
> >  1 file changed, 1 insertion(+)
> >
> > diff --git a/fs/ceph/addr.c b/fs/ceph/addr.c
> > index 63b75d214210..3462df35d245 100644
> > --- a/fs/ceph/addr.c
> > +++ b/fs/ceph/addr.c
> > @@ -1369,6 +1369,7 @@ int ceph_process_folio_batch(struct address_space *mapping,
> >               rc = move_dirty_folio_in_page_array(mapping, wbc, ceph_wbc,
> >                               folio);
> >               if (rc) {
> > +                     rc = 0;
>
> I like the fix but I would like to clarify the above questions at first.
>
> Thanks,
> Slava.
>
> >                       folio_redirty_for_writepage(wbc, folio);
> >                       folio_unlock(folio);
> >                       break;

Re: [PATCH 3/5] ceph: Free page array when ceph_submit_write fails

[Sam Edwards]

On Mon, Jan 5, 2026 at 1:09 PM Viacheslav Dubeyko <Slava.Dubeyko@ibm.com> wrote:
>
> On Tue, 2025-12-30 at 18:43 -0800, Sam Edwards wrote:
> > If `locked_pages` is zero, the page array must not be allocated:
> > ceph_process_folio_batch() uses `locked_pages` to decide when to
> > allocate `pages`,
> >
>

Hi Slava,

> I don't quite follow how this statement is relevant to the issue. If
> `locked_pages` is zero, then ceph_submit_write() will not to be called. Do I
> miss something here?

That statement is only informing that ceph_process_folio_batch() will
BUG() when locked_pages == 0 && pages != NULL. It establishes why
`pages` must be freed/NULLed before the next iteration of
ceph_writepages_start()'s loop (which sets locked_pages = 0).

>
> > and redundant allocations trigger
> > ceph_allocate_page_array()'s BUG_ON(), resulting in a worker oops (and
> > writeback stall) or even a kernel panic. Consequently, the main loop in
> > ceph_writepages_start() assumes that the lifetime of `pages` is confined
> > to a single iteration.
>
> It will be great to see the reproducer script or application and call trace of
> the issue. Could you please share the reproduction path and the call trace of
> the issue?

It's difficult to reproduce organically. It arises when
`!ceph_inc_osd_stopping_blocker(fsc->mdsc)`, which I understand can
only happen in a race. I used the fault injection framework to force
ceph_inc_osd_stopping_blocker() to fail.

The call trace is disinteresting. See my reply to your comments on
patch 1: it's the same trace.

>
> >
> > The ceph_submit_write() function claims ownership of the page array on
> > success.
> >
>
> As far as I can see, writepages_finish() should free the page array on success.

That's my understanding too; by "claims ownership of the page array" I
only mean that ceph_writepages_start() isn't responsible for cleaning
it up, once it calls ceph_submit_write().

>
> > But failures only redirty/unlock the pages and fail to free the
> > array, making the failure case in ceph_submit_write() fatal.
> >
> > Free the page array in ceph_submit_write()'s error-handling 'if' block
> > so that the caller's invariant (that the array does not outlive the
> > iteration) is maintained unconditionally, allowing failures in
> > ceph_submit_write() to be recoverable as originally intended.
> >
> > Fixes: 1551ec61dc55 ("ceph: introduce ceph_submit_write() method")
> > Cc: stable@vger.kernel.org
> > Signed-off-by: Sam Edwards <CFSworks@gmail.com>
> > ---
> >  fs/ceph/addr.c | 7 +++++++
> >  1 file changed, 7 insertions(+)
> >
> > diff --git a/fs/ceph/addr.c b/fs/ceph/addr.c
> > index 2b722916fb9b..91cc43950162 100644
> > --- a/fs/ceph/addr.c
> > +++ b/fs/ceph/addr.c
> > @@ -1466,6 +1466,13 @@ int ceph_submit_write(struct address_space *mapping,
> >                       unlock_page(page);
> >               }
> >
> > +             if (ceph_wbc->from_pool) {
> > +                     mempool_free(ceph_wbc->pages, ceph_wb_pagevec_pool);
> > +                     ceph_wbc->from_pool = false;
> > +             } else
> > +                     kfree(ceph_wbc->pages);
> > +             ceph_wbc->pages = NULL;
>
> Probably, it makes sense to introduce a method ceph_free_page_array likewise to
> __ceph_allocate_page_array() and to use for freeing page array in all places.

I like the suggestion but not the name. Instead of
ceph_free_page_array(), it should probably be called
ceph_discard_page_array(), because it is also redirtying the pages and
must not be used after successful writeback. (To me, "free" implies
success while "discard" implies failure.)

> Could ceph_wbc->locked_pages be greater than zero but ceph_wbc->pages == NULL?

ceph_wbc->locked_pages is the current array index into
ceph_wbc->pages, so they both need to be reset sometime before the
next iteration of ceph_writepages_start()'s loop.

Warm regards,
Sam



>
> Thanks,
> Slava.
>
> > +
> >               ceph_osdc_put_request(req);
> >               return -EIO;
> >       }

Re: [PATCH 5/5] ceph: Fix write storm on fscrypted files

[Sam Edwards]

On Mon, Jan 5, 2026 at 2:34 PM Viacheslav Dubeyko <Slava.Dubeyko@ibm.com> wrote:
>
> On Tue, 2025-12-30 at 18:43 -0800, Sam Edwards wrote:
> > CephFS stores file data across multiple RADOS objects. An object is the
> > atomic unit of storage, so the writeback code must clean only folios
> > that belong to the same object with each OSD request.
> >
> > CephFS also supports RAID0-style striping of file contents: if enabled,
> > each object stores multiple unbroken "stripe units" covering different
> > portions of the file; if disabled, a "stripe unit" is simply the whole
> > object. The stripe unit is (usually) reported as the inode's block size.
> >
> > Though the writeback logic could, in principle, lock all dirty folios
> > belonging to the same object, its current design is to lock only a
> > single stripe unit at a time. Ever since this code was first written,
> > it has determined this size by checking the inode's block size.
> > However, the relatively-new fscrypt support needed to reduce the block
> > size for encrypted inodes to the crypto block size (see 'fixes' commit),
> > which causes an unnecessarily high number of write operations (~1024x as
> > many, with 4MiB objects) and grossly degraded performance.

Hi Slava,

> Do you have any benchmarking results that prove your point?

I haven't done any "real" benchmarking for this change. On my setup
(closer to a home server than a typical Ceph deployment), sequential
write throughput increased from ~1.7 to ~66 MB/s with this patch
applied. I don't consider this single datapoint representative, so
rather than presenting it as a general benchmark in the commit
message, I chose the qualitative wording "grossly degraded
performance." Actual impact will vary depending on workload, disk
type, OSD count, etc.

Those curious about the bug's performance impact in their environment
can find out without enabling fscrypt, using: mount -o wsize=4096

However, the core rationale for my claim is based on principles, not
on measurements: batching writes into fewer operations necessarily
spreads per-operation overhead across more bytes. So this change
removes an artificial per-op bottleneck on sequential write
performance. The exact impact varies, but the patch does improve
(fscrypt-enabled) write throughput in nearly every case.

Warm regards,
Sam


>
> Thanks,
> Slava.
>
> >
> > Fix this (and clarify intent) by using i_layout.stripe_unit directly in
> > ceph_define_write_size() so that encrypted inodes are written back with
> > the same number of operations as if they were unencrypted.
> >
> > Fixes: 94af0470924c ("ceph: add some fscrypt guardrails")
> > Cc: stable@vger.kernel.org
> > Signed-off-by: Sam Edwards <CFSworks@gmail.com>
> > ---
> >  fs/ceph/addr.c | 3 ++-
> >  1 file changed, 2 insertions(+), 1 deletion(-)
> >
> > diff --git a/fs/ceph/addr.c b/fs/ceph/addr.c
> > index b3569d44d510..cb1da8e27c2b 100644
> > --- a/fs/ceph/addr.c
> > +++ b/fs/ceph/addr.c
> > @@ -1000,7 +1000,8 @@ unsigned int ceph_define_write_size(struct address_space *mapping)
> >  {
> >       struct inode *inode = mapping->host;
> >       struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
> > -     unsigned int wsize = i_blocksize(inode);
> > +     struct ceph_inode_info *ci = ceph_inode(inode);
> > +     unsigned int wsize = ci->i_layout.stripe_unit;
> >
> >       if (fsc->mount_options->wsize < wsize)
> >               wsize = fsc->mount_options->wsize;

RE: [PATCH 1/5] ceph: Do not propagate page array emplacement errors as batch errors

[Viacheslav Dubeyko]

On Mon, 2026-01-05 at 22:52 -0800, Sam Edwards wrote:
> On Mon, Jan 5, 2026 at 12:24 PM Viacheslav Dubeyko
> <Slava.Dubeyko@ibm.com> wrote:
> > 
> > On Tue, 2025-12-30 at 18:43 -0800, Sam Edwards wrote:
> > > When fscrypt is enabled, move_dirty_folio_in_page_array() may fail
> > > because it needs to allocate bounce buffers to store the encrypted
> > > versions of each folio. Each folio beyond the first allocates its bounce
> > > buffer with GFP_NOWAIT. Failures are common (and expected) under this
> > > allocation mode; they should flush (not abort) the batch.
> > > 
> > > However, ceph_process_folio_batch() uses the same `rc` variable for its
> > > own return code and for capturing the return codes of its routine calls;
> > > failing to reset `rc` back to 0 results in the error being propagated
> > > out to the main writeback loop, which cannot actually tolerate any
> > > errors here: once `ceph_wbc.pages` is allocated, it must be passed to
> > > ceph_submit_write() to be freed. If it survives until the next iteration
> > > (e.g. due to the goto being followed), ceph_allocate_page_array()'s
> > > BUG_ON() will oops the worker. (Subsequent patches in this series make
> > > the loop more robust.)
> > 
> 
> Hi Slava,
> 
> > I think you are right with the fix. We have the loop here and if we already
> > moved some dirty folios, then we should flush it. But what if we failed on the
> > first one folio, then should we return no error code in this case?
> 
> The case you ask about, where move_dirty_folio_in_page_array() returns
> an error for the first folio, is currently not possible:
> 1) The only error code that move_dirty_folio_in_page_array() can
> propagate is from fscrypt_encrypt_pagecache_blocks(), which it calls
> with GFP_NOFS for the first folio. The latter function's doc comment
> outright states:
>  * The bounce page allocation is mempool-backed, so it will always succeed when
>  * @gfp_flags includes __GFP_DIRECT_RECLAIM, e.g. when it's GFP_NOFS.
> 2) The error return isn't even reachable for the first folio because
> of the BUG_ON(ceph_wbc->locked_pages == 0); line.
> 

Unfortunately, the kernel code is not something completely stable. We cannot
rely on particular state of the code. The code should be stable, robust enough,
and ready for different situations. The mentioned BUG_ON() could be removed
somehow during refactoring because we already have comment there "better not
fail on first page!". Also, the behavior of fscrypt_encrypt_pagecache_blocks()
could be changed too. So, we need to expect any bad situation and this is why I
prefer to manage such potential (and maybe not so potential) erroneous
situation(s).

> > 
> > > 
> > > Note that this failure mode is currently masked due to another bug
> > > (addressed later in this series) that prevents multiple encrypted folios
> > > from being selected for the same write.
> > 
> > So, maybe, this patch has been not correctly placed in the order?
> 
> This crash is unmasked by patch 5 of this series. (It allows multiple
> folios to be batched when fscrypt is enabled.) Patch 5 has no hard
> dependency on anything else in this series, so it could -- in
> principle -- be ordered first as you suggest. However, that ordering
> would deliberately cause a regression in kernel stability, even if
> only briefly. That's not considered good practice in my view, as it
> may affect people who are trying to bisect and regression test. So the
> ordering of this series is: fix the crash in the unused code first,
> then fix the bug that makes it unused.
> 

OK, your point sounds confusing, frankly speaking. If we cannot reproduce the
issue because another bug hides the issue, then no such issue exists. And we
don't need to fix something. So, from the logical point of view, we need to fix
the first bug, then we can reproduce the hidden issue, and, finally, the fix
makes sense.

I didn't suggest too make the patch 5th as the first one. But I believe that
this patch should follow to the patch 5th. 

> > It will be
> > good to see the reproduction of the issue and which symptoms we have for this
> > issue. Do you have the reproduction script and call trace of the issue?
> 
> Fair point!
> 
> Function inlining makes the call trace not very interesting:
> Call trace:
>  ceph_writepages_start+0x16ec/0x18e0 [ceph] ()
>  do_writepages+0xb0/0x1c0
>  __writeback_single_inode+0x4c/0x4d8
>  writeback_sb_inodes+0x238/0x4c8
>  __writeback_inodes_wb+0x64/0x120
>  wb_writeback+0x320/0x3e8
>  wb_workfn+0x42c/0x518
>  process_one_work+0x17c/0x428
>  worker_thread+0x260/0x390
>  kthread+0x148/0x240
>  ret_from_fork+0x10/0x20
> Code: 34ffdee0 52800020 3903e7e0 17fffef4 (d4210000)
> ---[ end trace 0000000000000000 ]---
> Kernel panic - not syncing: Oops - BUG: Fatal exception
> 
> ceph_writepages_start+0x16ec corresponds to linux-6.18.2/fs/ceph/addr.c:1222
> 
> However, these repro steps should work:
> 1) Apply patch 5 from this series (and no other patches)
> 2) Mount CephFS and activate fscrypt
> 3) Copy a large directory into the CephFS mount
> 4) After dozens of GBs transferred, you should observe the above kernel oops

Could we have all of these details in the commit message?

Thanks,
Slava.

> 
> Warm regards,
> Sam
> 
> 
> 
> 
> > 
> > > 
> > > For now, just reset `rc` when redirtying the folio and prevent the
> > > error from propagating. After this change, ceph_process_folio_batch() no
> > > longer returns errors; its only remaining failure indicator is
> > > `locked_pages == 0`, which the caller already handles correctly. The
> > > next patch in this series addresses this.
> > > 
> > > Fixes: ce80b76dd327 ("ceph: introduce ceph_process_folio_batch() method")
> > > Cc: stable@vger.kernel.org
> > > Signed-off-by: Sam Edwards <CFSworks@gmail.com>
> > > ---
> > >  fs/ceph/addr.c | 1 +
> > >  1 file changed, 1 insertion(+)
> > > 
> > > diff --git a/fs/ceph/addr.c b/fs/ceph/addr.c
> > > index 63b75d214210..3462df35d245 100644
> > > --- a/fs/ceph/addr.c
> > > +++ b/fs/ceph/addr.c
> > > @@ -1369,6 +1369,7 @@ int ceph_process_folio_batch(struct address_space *mapping,
> > >               rc = move_dirty_folio_in_page_array(mapping, wbc, ceph_wbc,
> > >                               folio);
> > >               if (rc) {
> > > +                     rc = 0;
> > 
> > I like the fix but I would like to clarify the above questions at first.
> > 
> > Thanks,
> > Slava.
> > 
> > >                       folio_redirty_for_writepage(wbc, folio);
> > >                       folio_unlock(folio);
> > >                       break;

RE: [PATCH 5/5] ceph: Fix write storm on fscrypted files

[Viacheslav Dubeyko]

On Mon, 2026-01-05 at 22:53 -0800, Sam Edwards wrote:
> On Mon, Jan 5, 2026 at 2:34 PM Viacheslav Dubeyko <Slava.Dubeyko@ibm.com> wrote:
> > 
> > On Tue, 2025-12-30 at 18:43 -0800, Sam Edwards wrote:
> > > CephFS stores file data across multiple RADOS objects. An object is the
> > > atomic unit of storage, so the writeback code must clean only folios
> > > that belong to the same object with each OSD request.
> > > 
> > > CephFS also supports RAID0-style striping of file contents: if enabled,
> > > each object stores multiple unbroken "stripe units" covering different
> > > portions of the file; if disabled, a "stripe unit" is simply the whole
> > > object. The stripe unit is (usually) reported as the inode's block size.
> > > 
> > > Though the writeback logic could, in principle, lock all dirty folios
> > > belonging to the same object, its current design is to lock only a
> > > single stripe unit at a time. Ever since this code was first written,
> > > it has determined this size by checking the inode's block size.
> > > However, the relatively-new fscrypt support needed to reduce the block
> > > size for encrypted inodes to the crypto block size (see 'fixes' commit),
> > > which causes an unnecessarily high number of write operations (~1024x as
> > > many, with 4MiB objects) and grossly degraded performance.
> 
> Hi Slava,
> 
> > Do you have any benchmarking results that prove your point?
> 
> I haven't done any "real" benchmarking for this change. On my setup
> (closer to a home server than a typical Ceph deployment), sequential
> write throughput increased from ~1.7 to ~66 MB/s with this patch
> applied. I don't consider this single datapoint representative, so
> rather than presenting it as a general benchmark in the commit
> message, I chose the qualitative wording "grossly degraded
> performance." Actual impact will vary depending on workload, disk
> type, OSD count, etc.
> 
> Those curious about the bug's performance impact in their environment
> can find out without enabling fscrypt, using: mount -o wsize=4096
> 
> However, the core rationale for my claim is based on principles, not
> on measurements: batching writes into fewer operations necessarily
> spreads per-operation overhead across more bytes. So this change
> removes an artificial per-op bottleneck on sequential write
> performance. The exact impact varies, but the patch does improve
> (fscrypt-enabled) write throughput in nearly every case.
> 

If you claim in commit message that "this patch fixes performance degradation",
then you MUST share the evidence (benchmarking results). Otherwise, you cannot
make such statements in commit message. Yes, it could be a good fix but please
don't make a promise of performance improvement. Because, end-users have very
different environments and workloads. And what could work on your side may not
work for other use-cases and environments. Potentially, you could describe your
environment, workload and to share your estimation/expectation of potential
performance improvement.

Thanks,
Slava.

> Warm regards,
> Sam
> 
> 
> > 
> > Thanks,
> > Slava.
> > 
> > > 
> > > Fix this (and clarify intent) by using i_layout.stripe_unit directly in
> > > ceph_define_write_size() so that encrypted inodes are written back with
> > > the same number of operations as if they were unencrypted.
> > > 
> > > Fixes: 94af0470924c ("ceph: add some fscrypt guardrails")
> > > Cc: stable@vger.kernel.org
> > > Signed-off-by: Sam Edwards <CFSworks@gmail.com>
> > > ---
> > >  fs/ceph/addr.c | 3 ++-
> > >  1 file changed, 2 insertions(+), 1 deletion(-)
> > > 
> > > diff --git a/fs/ceph/addr.c b/fs/ceph/addr.c
> > > index b3569d44d510..cb1da8e27c2b 100644
> > > --- a/fs/ceph/addr.c
> > > +++ b/fs/ceph/addr.c
> > > @@ -1000,7 +1000,8 @@ unsigned int ceph_define_write_size(struct address_space *mapping)
> > >  {
> > >       struct inode *inode = mapping->host;
> > >       struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
> > > -     unsigned int wsize = i_blocksize(inode);
> > > +     struct ceph_inode_info *ci = ceph_inode(inode);
> > > +     unsigned int wsize = ci->i_layout.stripe_unit;
> > > 
> > >       if (fsc->mount_options->wsize < wsize)
> > >               wsize = fsc->mount_options->wsize;

Re: [PATCH 1/5] ceph: Do not propagate page array emplacement errors as batch errors

[Sam Edwards]

On Tue, Jan 6, 2026 at 1:08 PM Viacheslav Dubeyko <Slava.Dubeyko@ibm.com> wrote:
>
> On Mon, 2026-01-05 at 22:52 -0800, Sam Edwards wrote:
> > On Mon, Jan 5, 2026 at 12:24 PM Viacheslav Dubeyko
> > <Slava.Dubeyko@ibm.com> wrote:
> > >
> > > On Tue, 2025-12-30 at 18:43 -0800, Sam Edwards wrote:
> > > > When fscrypt is enabled, move_dirty_folio_in_page_array() may fail
> > > > because it needs to allocate bounce buffers to store the encrypted
> > > > versions of each folio. Each folio beyond the first allocates its bounce
> > > > buffer with GFP_NOWAIT. Failures are common (and expected) under this
> > > > allocation mode; they should flush (not abort) the batch.
> > > >
> > > > However, ceph_process_folio_batch() uses the same `rc` variable for its
> > > > own return code and for capturing the return codes of its routine calls;
> > > > failing to reset `rc` back to 0 results in the error being propagated
> > > > out to the main writeback loop, which cannot actually tolerate any
> > > > errors here: once `ceph_wbc.pages` is allocated, it must be passed to
> > > > ceph_submit_write() to be freed. If it survives until the next iteration
> > > > (e.g. due to the goto being followed), ceph_allocate_page_array()'s
> > > > BUG_ON() will oops the worker. (Subsequent patches in this series make
> > > > the loop more robust.)
> > >
> >
> > Hi Slava,
> >
> > > I think you are right with the fix. We have the loop here and if we already
> > > moved some dirty folios, then we should flush it. But what if we failed on the
> > > first one folio, then should we return no error code in this case?
> >
> > The case you ask about, where move_dirty_folio_in_page_array() returns
> > an error for the first folio, is currently not possible:
> > 1) The only error code that move_dirty_folio_in_page_array() can
> > propagate is from fscrypt_encrypt_pagecache_blocks(), which it calls
> > with GFP_NOFS for the first folio. The latter function's doc comment
> > outright states:
> >  * The bounce page allocation is mempool-backed, so it will always succeed when
> >  * @gfp_flags includes __GFP_DIRECT_RECLAIM, e.g. when it's GFP_NOFS.
> > 2) The error return isn't even reachable for the first folio because
> > of the BUG_ON(ceph_wbc->locked_pages == 0); line.
> >
>

Good day Slava,

> Unfortunately, the kernel code is not something completely stable. We cannot
> rely on particular state of the code. The code should be stable, robust enough,
> and ready for different situations.

You describe "defensive programming." I fully agree and am a strong
advocate for it, but each defensive measure comes with a complexity
cost. A skilled defensive programmer evaluates the likelihood of each
failure and invests that cost only where it's most warranted.

> The mentioned BUG_ON() could be removed
> somehow during refactoring because we already have comment there "better not
> fail on first page!".

If the question is "What happens if the first folio fails when the
BUG_ON is removed?" then my answer is: that is the responsibility of
the person removing it. I am leaving the BUG_ON() in place.

> Also, the behavior of fscrypt_encrypt_pagecache_blocks()
> could be changed too.

Changing that would alter the contract of
fscrypt_encrypt_pagecache_blocks(). Contracts can evolve, but anyone
making such a change must audit all call sites to ensure nothing
breaks. Today, this is purely hypothetical; the function is not being
changed. Speculating about behavior under a different, unimplemented
contract is not a basis for complicating the current code.

> So, we need to expect any bad situation and this is why I
> prefer to manage such potential (and maybe not so potential) erroneous
> situation(s).

This point is moot. Even if move_dirty_folio_in_page_array() somehow
returned nonzero on the first folio, ceph_process_folio_batch() would
simply lock zero folios, which ceph_writepages_start() already handles
gracefully.

>
> > >
> > > >
> > > > Note that this failure mode is currently masked due to another bug
> > > > (addressed later in this series) that prevents multiple encrypted folios
> > > > from being selected for the same write.
> > >
> > > So, maybe, this patch has been not correctly placed in the order?
> >
> > This crash is unmasked by patch 5 of this series. (It allows multiple
> > folios to be batched when fscrypt is enabled.) Patch 5 has no hard
> > dependency on anything else in this series, so it could -- in
> > principle -- be ordered first as you suggest. However, that ordering
> > would deliberately cause a regression in kernel stability, even if
> > only briefly. That's not considered good practice in my view, as it
> > may affect people who are trying to bisect and regression test. So the
> > ordering of this series is: fix the crash in the unused code first,
> > then fix the bug that makes it unused.
> >
>
> OK, your point sounds confusing, frankly speaking. If we cannot reproduce the
> issue because another bug hides the issue, then no such issue exists. And we
> don't need to fix something. So, from the logical point of view, we need to fix
> the first bug, then we can reproduce the hidden issue, and, finally, the fix
> makes sense.

With respect, that reasoning is flawed and not appropriate for a
technical discussion. The crash in question cannot currently occur,
but that does *not* make the fix unnecessary. Patch #5 in this series
will re-enable the code path, at which point the crash becomes
possible. Addressing it now ensures correctness and avoids introducing
a regression. Attempting to "see it happen in the wild" before fixing
it is neither required nor acceptable practice.

We are not uncertain about the crash: I have provided steps to
reproduce it. You can apply patch #5 before #1 *in your own tree* to
observe the crash if that helps you evaluate the patches. *But under
no circumstances should this be done in mainline!* Introducing a crash
upstream, even transiently, is strictly disallowed, and suggesting
otherwise is not appropriate behavior for a Linux kernel developer.

>
> I didn't suggest too make the patch 5th as the first one. But I believe that
> this patch should follow to the patch 5th.

As I explained, putting patch #5 before this one would deliberately
introduce a regression -- a crash. Triggering this in mainline is not
allowed by kernel development policy [1]; there is no exception for
"transient regressions that are fixed immediately afterward." A
regression is a regression.

>
> > > It will be
> > > good to see the reproduction of the issue and which symptoms we have for this
> > > issue. Do you have the reproduction script and call trace of the issue?
> >
> > Fair point!
> >
> > Function inlining makes the call trace not very interesting:
> > Call trace:
> >  ceph_writepages_start+0x16ec/0x18e0 [ceph] ()
> >  do_writepages+0xb0/0x1c0
> >  __writeback_single_inode+0x4c/0x4d8
> >  writeback_sb_inodes+0x238/0x4c8
> >  __writeback_inodes_wb+0x64/0x120
> >  wb_writeback+0x320/0x3e8
> >  wb_workfn+0x42c/0x518
> >  process_one_work+0x17c/0x428
> >  worker_thread+0x260/0x390
> >  kthread+0x148/0x240
> >  ret_from_fork+0x10/0x20
> > Code: 34ffdee0 52800020 3903e7e0 17fffef4 (d4210000)
> > ---[ end trace 0000000000000000 ]---
> > Kernel panic - not syncing: Oops - BUG: Fatal exception
> >
> > ceph_writepages_start+0x16ec corresponds to linux-6.18.2/fs/ceph/addr.c:1222
> >
> > However, these repro steps should work:
> > 1) Apply patch 5 from this series (and no other patches)
> > 2) Mount CephFS and activate fscrypt
> > 3) Copy a large directory into the CephFS mount
> > 4) After dozens of GBs transferred, you should observe the above kernel oops
>
> Could we have all of these details in the commit message?

Would this truly help future readers, or just create noise? The commit
message already explains the exact execution path to the
BUG_ON()/oops, which is what really matters; call traces are
secondary. I did not want to imply that readers cannot understand the
seriousness of the issue without a crash log. I will include these
details if the group consensus prefers it, but I am otherwise opposed.

Hope you and yours are well,
Sam

[1] See the "no regressions" rule:
https://docs.kernel.org/admin-guide/reporting-regressions.html

>
> Thanks,
> Slava.
>
> >
> > Warm regards,
> > Sam
> >
> >
> >
> >
> > >
> > > >
> > > > For now, just reset `rc` when redirtying the folio and prevent the
> > > > error from propagating. After this change, ceph_process_folio_batch() no
> > > > longer returns errors; its only remaining failure indicator is
> > > > `locked_pages == 0`, which the caller already handles correctly. The
> > > > next patch in this series addresses this.
> > > >
> > > > Fixes: ce80b76dd327 ("ceph: introduce ceph_process_folio_batch() method")
> > > > Cc: stable@vger.kernel.org
> > > > Signed-off-by: Sam Edwards <CFSworks@gmail.com>
> > > > ---
> > > >  fs/ceph/addr.c | 1 +
> > > >  1 file changed, 1 insertion(+)
> > > >
> > > > diff --git a/fs/ceph/addr.c b/fs/ceph/addr.c
> > > > index 63b75d214210..3462df35d245 100644
> > > > --- a/fs/ceph/addr.c
> > > > +++ b/fs/ceph/addr.c
> > > > @@ -1369,6 +1369,7 @@ int ceph_process_folio_batch(struct address_space *mapping,
> > > >               rc = move_dirty_folio_in_page_array(mapping, wbc, ceph_wbc,
> > > >                               folio);
> > > >               if (rc) {
> > > > +                     rc = 0;
> > >
> > > I like the fix but I would like to clarify the above questions at first.
> > >
> > > Thanks,
> > > Slava.
> > >
> > > >                       folio_redirty_for_writepage(wbc, folio);
> > > >                       folio_unlock(folio);
> > > >                       break;

Re: [PATCH 5/5] ceph: Fix write storm on fscrypted files

[Sam Edwards]

On Tue, Jan 6, 2026 at 3:11 PM Viacheslav Dubeyko <Slava.Dubeyko@ibm.com> wrote:
>
> On Mon, 2026-01-05 at 22:53 -0800, Sam Edwards wrote:
> > On Mon, Jan 5, 2026 at 2:34 PM Viacheslav Dubeyko <Slava.Dubeyko@ibm.com> wrote:
> > >
> > > On Tue, 2025-12-30 at 18:43 -0800, Sam Edwards wrote:
> > > > CephFS stores file data across multiple RADOS objects. An object is the
> > > > atomic unit of storage, so the writeback code must clean only folios
> > > > that belong to the same object with each OSD request.
> > > >
> > > > CephFS also supports RAID0-style striping of file contents: if enabled,
> > > > each object stores multiple unbroken "stripe units" covering different
> > > > portions of the file; if disabled, a "stripe unit" is simply the whole
> > > > object. The stripe unit is (usually) reported as the inode's block size.
> > > >
> > > > Though the writeback logic could, in principle, lock all dirty folios
> > > > belonging to the same object, its current design is to lock only a
> > > > single stripe unit at a time. Ever since this code was first written,
> > > > it has determined this size by checking the inode's block size.
> > > > However, the relatively-new fscrypt support needed to reduce the block
> > > > size for encrypted inodes to the crypto block size (see 'fixes' commit),
> > > > which causes an unnecessarily high number of write operations (~1024x as
> > > > many, with 4MiB objects) and grossly degraded performance.
> >
> > Hi Slava,
> >
> > > Do you have any benchmarking results that prove your point?
> >
> > I haven't done any "real" benchmarking for this change. On my setup
> > (closer to a home server than a typical Ceph deployment), sequential
> > write throughput increased from ~1.7 to ~66 MB/s with this patch
> > applied. I don't consider this single datapoint representative, so
> > rather than presenting it as a general benchmark in the commit
> > message, I chose the qualitative wording "grossly degraded
> > performance." Actual impact will vary depending on workload, disk
> > type, OSD count, etc.
> >
> > Those curious about the bug's performance impact in their environment
> > can find out without enabling fscrypt, using: mount -o wsize=4096
> >
> > However, the core rationale for my claim is based on principles, not
> > on measurements: batching writes into fewer operations necessarily
> > spreads per-operation overhead across more bytes. So this change
> > removes an artificial per-op bottleneck on sequential write
> > performance. The exact impact varies, but the patch does improve
> > (fscrypt-enabled) write throughput in nearly every case.
> >
>

Hi Slava,

> If you claim in commit message that "this patch fixes performance degradation",
> then you MUST share the evidence (benchmarking results). Otherwise, you cannot
> make such statements in commit message. Yes, it could be a good fix but please
> don't make a promise of performance improvement. Because, end-users have very
> different environments and workloads. And what could work on your side may not
> work for other use-cases and environments.

I agree with the underlying concern: I do not have a representative
environment, and it would be irresponsible to promise or quantify a
specific speedup. For that reason, the commit message does not claim
any particular improvement factor.

What it does say is that this patch fixes a known performance
degradation that artificially limits the write batch size. But that
statement is about correctness relative to previous behavior, not
about guaranteeing any specific performance outcome for end users.

> Potentially, you could describe your
> environment, workload and to share your estimation/expectation of potential
> performance improvement.

I don’t think that would be useful here. As you pointed out, any such
numbers would be highly workload- and environment-specific and would
not be representative or actionable. The purpose of this patch is
simply to remove an unintentional limit, not to advertise or promise
measurable gains.

Best,
Sam

>
> Thanks,
> Slava.
>
> > Warm regards,
> > Sam
> >
> >
> > >
> > > Thanks,
> > > Slava.
> > >
> > > >
> > > > Fix this (and clarify intent) by using i_layout.stripe_unit directly in
> > > > ceph_define_write_size() so that encrypted inodes are written back with
> > > > the same number of operations as if they were unencrypted.
> > > >
> > > > Fixes: 94af0470924c ("ceph: add some fscrypt guardrails")
> > > > Cc: stable@vger.kernel.org
> > > > Signed-off-by: Sam Edwards <CFSworks@gmail.com>
> > > > ---
> > > >  fs/ceph/addr.c | 3 ++-
> > > >  1 file changed, 2 insertions(+), 1 deletion(-)
> > > >
> > > > diff --git a/fs/ceph/addr.c b/fs/ceph/addr.c
> > > > index b3569d44d510..cb1da8e27c2b 100644
> > > > --- a/fs/ceph/addr.c
> > > > +++ b/fs/ceph/addr.c
> > > > @@ -1000,7 +1000,8 @@ unsigned int ceph_define_write_size(struct address_space *mapping)
> > > >  {
> > > >       struct inode *inode = mapping->host;
> > > >       struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
> > > > -     unsigned int wsize = i_blocksize(inode);
> > > > +     struct ceph_inode_info *ci = ceph_inode(inode);
> > > > +     unsigned int wsize = ci->i_layout.stripe_unit;
> > > >
> > > >       if (fsc->mount_options->wsize < wsize)
> > > >               wsize = fsc->mount_options->wsize;