Re: [PATCH v2 2/5] mm: limit pagecache folio refcount when splicing into pipes

["David Hildenbrand (Arm)" <david@kernel.org>]

On 11/9/26 10:30, David Hildenbrand (Arm) wrote:
> On 11/9/26 09:57, Jose A. Perez de Azpillaga wrote:
>> On 11/7/26 11:23 PM, David Hildenbrand (Arm) wrote:
>>> Similar to mapping pagecache folios into page tables, let's reject
>>> additional references when splicing folios into a pipe if the refcount
>>> is at its limit, similar to how we handle it during GUP.
>>>
>>> This will prevent more cases of possible folio refcount overflows with
>>> large pagecache folios.
>>>
>>> Let's keep it simple and check in the two callers of
>>> splice_folio_into_pipe(). Return -EOVERFLOW if the limit prevents any
>>> progress, so do_splice_read() can fall back to copy_splice_read() when
>>> read_iter is available. This keeps sendfile() and other splice users
>>> working when ordinary reads would still succeed.
>>>
>>> Also fall back to copying in FUSE when taking another reference through
>>> fuse_ref_folio() would exceed the limit.
>>>
>>> Also check in try_get_page() to cover pipe-buffer duplication by tee()
>>> and partial pipe-to-pipe splicing.
>>>
>>> Fixes: 800d8c63b2e9 ("shmem: add huge pages support")
>>> Fixes: 99cb0dbd47a1 ("mm,thp: add read-only THP support for (non-shmem) FS")
>>> Fixes: 793917d997df ("mm/readahead: Add large folio readahead")
>>> Cc: stable@vger.kernel.org
>>> Signed-off-by: David Hildenbrand (Arm) <david@kernel.org>
>>> ---
>>>  fs/fuse/dev.c      |  5 +++--
>>>  fs/splice.c        |  8 +++++++-
>>>  include/linux/mm.h |  9 ++++++---
>>>  mm/filemap.c       | 11 ++++++++---
>>>  mm/shmem.c         |  4 ++++
>>>  5 files changed, 28 insertions(+), 9 deletions(-)
>>>
>>> diff --git a/fs/fuse/dev.c b/fs/fuse/dev.c
>>> index 3641abb90d6bb..4f6b7b42bfd3e 100644
>>> --- a/fs/fuse/dev.c
>>> +++ b/fs/fuse/dev.c
>>> @@ -1265,9 +1265,10 @@ int fuse_copy_folio(struct fuse_copy_state *cs, struct folio **foliop,
>>>  		if (cs->write && cs->pipebufs && folio) {
>>>  			/*
>>>  			 * Can't control lifetime of pipe buffers, so always
>>> -			 * copy user pages.
>>> +			 * copy user pages. Also copy folios at their refcount limit.
>>>  			 */
>>> -			if (cs->req->args->user_pages) {
>>> +			if (cs->req->args->user_pages ||
>>> +			    unlikely(folio_ref_count_over_limit(folio))) {
>>>  				err = fuse_copy_fill(cs);
>>>  				if (err)
>>>  					return err;
>>> diff --git a/fs/splice.c b/fs/splice.c
>>> index e6d43329ed3f0..df73a4e65118b 100644
>>> --- a/fs/splice.c
>>> +++ b/fs/splice.c
>>> @@ -956,6 +956,7 @@ static ssize_t do_splice_read(struct file *in, loff_t *ppos,
>>>  			      unsigned int flags)
>>>  {
>>>  	unsigned int p_space;
>>> +	ssize_t ret;
>>>  
>>>  	if (unlikely(!(in->f_mode & FMODE_READ)))
>>>  		return -EBADF;
>>> @@ -977,7 +978,12 @@ static ssize_t do_splice_read(struct file *in, loff_t *ppos,
>>>  	 */
>>>  	if ((in->f_flags & O_DIRECT) || IS_DAX(in->f_mapping->host))
>>>  		return copy_splice_read(in, ppos, pipe, len, flags);
>>> -	return in->f_op->splice_read(in, ppos, pipe, len, flags);
>>> +	ret = in->f_op->splice_read(in, ppos, pipe, len, flags);
>>> +
>>> +	/* Fallback to copying when hitting folio refcount limits. */
>>> +	if (unlikely(ret == -EOVERFLOW) && in->f_op->read_iter)
>>> +		ret = copy_splice_read(in, ppos, pipe, len, flags);
>>
>> here it can't tell where the -EOVERFLOW came from, if another file
>> returns -EOVERFLOW (i.e. tty_fops uses copy_splice_read, and
>> n_hdlc_tty_read() returns -EOVERFLOW after dropping the frame), it now
>> gets a second read.
> 
> 
> That's what claude told me during review, and when asked whether that actually
> happened in practice it said no ;)
> 
> So is that actually a problem?

Double checked, claude managed to confuse itself.

n_hdlc_tty_read() can indeed return -EOVERFLOW.

I should just have let splice keep failing, dealing with copy fallback starts to
get annoying.

I guess switching to -ETOOMANYREFS should work (and might be a better fitting error)

-- 
Cheers,

David