Re: [RFC PATCH v2 39/51] KVM: guest_memfd: Merge and truncate on fallocate(PUNCH_HOLE)

[Ackerley Tng <ackerleytng@google.com>]

Yan Zhao <yan.y.zhao@intel.com> writes:

> On Wed, May 28, 2025 at 09:39:35AM -0700, Ackerley Tng wrote:
>> Yan Zhao <yan.y.zhao@intel.com> writes:
>> 
>> > On Wed, May 14, 2025 at 04:42:18PM -0700, Ackerley Tng wrote:
>> >> Merge and truncate on fallocate(PUNCH_HOLE), but if the file is being
>> >> closed, defer merging to folio_put() callback.
>> >> 
>> >> Change-Id: Iae26987756e70c83f3b121edbc0ed0bc105eec0d
>> >> Signed-off-by: Ackerley Tng <ackerleytng@google.com>
>> >> ---
>> >>  virt/kvm/guest_memfd.c | 76 +++++++++++++++++++++++++++++++++++++-----
>> >>  1 file changed, 68 insertions(+), 8 deletions(-)
>> >> 
>> >> diff --git a/virt/kvm/guest_memfd.c b/virt/kvm/guest_memfd.c
>> >> index cb426c1dfef8..04b1513c2998 100644
>> >> --- a/virt/kvm/guest_memfd.c
>> >> +++ b/virt/kvm/guest_memfd.c
>> >> @@ -859,6 +859,35 @@ static int kvm_gmem_restructure_folios_in_range(struct inode *inode,
>> >>  	return ret;
>> >>  }
>> >>  
>> >> +static long kvm_gmem_merge_truncate_indices(struct inode *inode, pgoff_t index,
>> >> +					   size_t nr_pages)
>> >> +{
>> >> +	struct folio *f;
>> >> +	pgoff_t unused;
>> >> +	long num_freed;
>> >> +
>> >> +	unmap_mapping_pages(inode->i_mapping, index, nr_pages, false);
>> >> +
>> >> +	if (!kvm_gmem_has_safe_refcount(inode->i_mapping, index, nr_pages, &unused))
>> 
>> Yan, thank you for your reviews!
>> 

Thanks again for your reviews. I would like to respond to this since I'm
finally getting back to this part.

>> > Why is kvm_gmem_has_safe_refcount() checked here, but not in
>> > kvm_gmem_zero_range() within kvm_gmem_truncate_inode_range() in patch 33?
>> >
>> 
>> The contract for guest_memfd with HugeTLB pages is that if holes are
>> punched in any ranges less than a full huge page, no pages are removed
>> from the filemap. Those ranges are only zeroed.
>> 
>> In kvm_gmem_zero_range(), we never remove any folios, and so there is no
>> need to merge. If there's no need to merge, then we don't need to check
>> for a safe refcount, and can just proceed to zero.
> However, if there are still extra ref count to a shared page, its content will
> be zeroed out.
>

I believe this topic is kind of overtaken by events. IIUC the current
upstream stance is that for guest_memfd we're not allowing hole-punching
of pages for huge pages, so once a HugeTLB guest_memfd is requested,
hole punching of less than the requested HugeTLB size will result in
-EINVAL being returned to userspace.

>> 
>> kvm_gmem_merge_truncate_indices() is only used during hole punching and
>> not when the file is closed. Hole punch vs file closure is checked using
>> mapping_exiting(inode->i_mapping).
>> 
>> During a hole punch, we will only allow truncation if there are no
>> unexpected refcounts on any subpages, hence this
>> kvm_gmem_has_safe_refcount() check.
> Hmm, I couldn't find a similar refcount check in hugetlbfs_punch_hole().
> Did I overlook it?
>
> So, why does guest_memfd require this check when punching a hole?
>

There's no equivalent check in HugeTLBfs.

For guest_memfd, we want to defer merging to the kernel worker as little
as possible, hence we want to merge before truncating. Checking for
elevated refcounts is a prerequisite for merging, not directly for hole
punching.

>> >> +		return -EAGAIN;
>> >> +
>> >
>> > Rather than merging the folios, could we simply call kvm_gmem_truncate_indices()
>> > instead?
>> >
>> > num_freed = kvm_gmem_truncate_indices(inode->i_mapping, index, nr_pages);
>> > return num_freed;
>> >
>> 
>> We could do this too, but then that would be deferring the huge page
>> merging to the folio_put() callback and eventually the kernel worker
>> thread.
> With deferring the huge page merging to folio_put(), a benefit is that
> __kvm_gmem_filemap_add_folio() can be saved for the merged folio. This function
> is possible to fail and is unnecessary for punch hole as the folio will be
> removed immediately from the filemap in truncate_inode_folio().
>
>

That is a good point! Definitely sounds good to defer this to folio_put().

>> My goal here is to try to not to defer merging and freeing as much as
>> possible so that most of the page/memory operations are
>> synchronous, because synchronous operations are more predictable.
>> 
>> As an example of improving predictability, in one of the selftests, I do
>> a hole punch and then try to allocate again. Because the merging and
>> freeing of the HugeTLB page sometimes takes too long, the allocation
>> sometimes fails: the guest_memfd's subpool hadn't yet received the freed
>> page back. With a synchronous truncation, the truncation may take
>> longer, but the selftest predictably passes.
> Maybe check if guestmem_hugetlb_handle_folio_put() is invoked in the
> interrupt context, and, if not, invoke the guestmem_hugetlb_cleanup_folio()
> synchronously?
>
>

I think this is a good idea. I would like to pursue this in a future
revision of a patch series.

It seems like the use of in_atomic() is strongly discouraged, do you
have any tips on how to determine if folio_put() is being called from
atomic context?

>> 
>> [...snip...]
>>