Re: [PATCH] mm/migrate_device: fix folio refcount leak on folio_split_unmapped failure

[Usama Arif <usama.arif@linux.dev>]

On 05/03/2026 16:39, Zi Yan wrote:
> On 5 Mar 2026, at 11:36, Usama Arif wrote:
> 
>> On 05/03/2026 12:09, Mika Penttilä wrote:
>>> On 3/5/26 13:44, Usama Arif wrote:
>>>
>>>>
>>>> On 05/03/2026 06:09, Mika Penttilä wrote:
>>>>> Hi!
>>>>>
>>>>> On 3/5/26 01:28, Usama Arif wrote:
>>>>>
>>>>>> On 04/03/2026 22:09, Balbir Singh wrote:
>>>>>>> On 3/5/26 08:54, Zi Yan wrote:
>>>>>>>> On 4 Mar 2026, at 16:48, Balbir Singh wrote:
>>>>>>>>
>>>>>>>>> On 3/5/26 02:17, Zi Yan wrote:
>>>>>>>>>> On 4 Mar 2026, at 7:01, Usama Arif wrote:
>>>>>>>>>>
>>>>>>>>>>> From: Usama Arif <usama.arif@linux.dev>
>>>>>>>>>>>
>>>>>>>>>>> migrate_vma_split_unmapped_folio() takes an extra reference via
>>>>>>>>>>> folio_get() before calling folio_split_unmapped().  On success, the
>>>>>>>>>>> split consumes this reference: __folio_freeze_and_split_unmapped()
>>>>>>>>>>> expects the +1 in its folio_ref_freeze() check, and distributes it
>>>>>>>>>>> across the resulting sub-folios via folio_ref_unfreeze(...+1), which
>>>>>>>>>>> are later balanced by folio_put() calls in __migrate_device_finalize().
>>>>>>>>>>>
>>>>>>>>>>> If folio_split_unmapped() fails (e.g., unexpected pinning returns
>>>>>>>>>>> -EAGAIN), the function returns without calling folio_put().  The extra
>>>>>>>>>>> reference is never released.
>>>>>>>>>>>
>>>>>>>>>>> Add the missing folio_put() on the error path.
>>>>>>>>>>>
>>>>>>>>>>> Fixes: 4265d67e405a4 ("mm/migrate_device: add THP splitting during migration")
>>>>>>>>>>> Closes: https://lore.kernel.org/all/CAA1CXcDyqPPwf_-W7B+PFQtL8HdoJGCEqVsVxq7DhOUB=L4PQA@mail.gmail.com/
>>>>>>>>>>> Reported-by: Nico Pache <npache@redhat.com>
>>>>>>>>>>> Signed-off-by: Usama Arif <usama.arif@linux.dev>
>>>>>>>>>>> ---
>>>>>>>>>>>  mm/migrate_device.c | 4 +++-
>>>>>>>>>>>  1 file changed, 3 insertions(+), 1 deletion(-)
>>>>>>>>>>>
>>>>>>>>>>> diff --git a/mm/migrate_device.c b/mm/migrate_device.c
>>>>>>>>>>> index 0a8b31939640f..351ecd9065d13 100644
>>>>>>>>>>> --- a/mm/migrate_device.c
>>>>>>>>>>> +++ b/mm/migrate_device.c
>>>>>>>>>>> @@ -917,8 +917,10 @@ static int migrate_vma_split_unmapped_folio(struct migrate_vma *migrate,
>>>>>>>>>>>  	folio_get(folio);
>>>>>>>>>>>  	split_huge_pmd_address(migrate->vma, addr, true);
>>>>>>>>>>>  	ret = folio_split_unmapped(folio, 0);
>>>>>>>>>>> -	if (ret)
>>>>>>>>>>> +	if (ret) {
>>>>>>>>>>> +		folio_put(folio);
>>>>>>>>>>>  		return ret;
>>>>>>>>>>> +	}
>>>>>>>>>>>  	migrate->src[idx] &= ~MIGRATE_PFN_COMPOUND;
>>>>>>>>>>>  	flags = migrate->src[idx] & ((1UL << MIGRATE_PFN_SHIFT) - 1);
>>>>>>>>>>>  	pfn = migrate->src[idx] >> MIGRATE_PFN_SHIFT;
>>>>>>>>>>> -- 
>>>>>>>>>>> 2.47.3
>>>>>>>>>> Add Balbir, who wrote the code, to comment on this.
>>>>>>>>>>
>>>>>>>>> Thanks Zi!
>>>>>>>>>
>>>>>>>>> Just wondering if there is a reproducer for the issue and how the fix was tested?
>>>>>>>>> I expect migrate_vma_finalize() to be called for folios, even when split failed and
>>>>>>>>> drop the lock.
>>>>>>>> Does migrate_vma_finalize() do folio_put() for failed-to-split folios?
>>>>>>>> If so, how does it distinguish between split folios and failed-to-split folios?
>>>>>>>> By comparing source and destination folio orders?
>>>>>>>>
>>>>>>> We reset the MIGRATE_PFN_MIGRATE flag for failing to migrate pfns.  We do a folio_put
>>>>>>> on the src in finalize, if it is split then on all the split folios as well.
>>>>>>>
>>>>>>>> What we see from migrate_vma_split_unmapped_folio() is that
>>>>>>>> it adds a refcount for all input folios, but only drops a refcount
>>>>>>>> for the split folio. Isn’t it cause failed-to-split folios to have
>>>>>>>> additional refcount?
>>>>>>>>
>>>>>> Hello!
>>>>>>
>>>>>> Thanks for reviewing everyone. So its very difficult to create a reproducer I think
>>>>>> the extra reference would need to appear after migrate_device_unmap() but before
>>>>>> folio_split_unmapped() in migrate_vma_pages()? That's hard to trigger reliably from
>>>>>> userspace.
>>>>>>
>>>>>> The fix came about when Nico indicated there might be an issue if split_huge_pmd_address
>>>>>> fails in my patch [1].
>>>>>>
>>>>>> Below is my understanding of how refcounting is working over here step by step. I
>>>>>> might very well be wrong on this, and the refcounting is a bit all over the place
>>>>>> and I might miss a reference change somewhere so would really appreciate if someone
>>>>>> can confirm this!
>>>>>>
>>>>>>
>>>>>> 1. migrate_vma_collect_huge_pmd():
>>>>>>   a) folio_get(folio) -> +1 (collect reference)
>>>>>> 2. migrate_device_unmap():
>>>>>>   a) folio_isolate_lru() -> +1 (isolation reference)
>>>>>>   b) folio_put() -> -1 (drops the collect reference)
>>>>>>
>>>>>>
>>>>>> Without this patch fix:
>>>>>>
>>>>>> 3. migrate_vma_split_unmapped_folio():
>>>>>>   a) folio_get(folio) -> +1 (split reference)
>>>>>>   b) folio_split_unmapped() -> fails
>>>>>>   c) Returns error — without folio_put() which is the fix
>>>>>> 4. Caller in migrate_vma_pages(): clears MIGRATE_PFN_MIGRATE | MIGRATE_PFN_COMPOUND
>>>>>> 5. __migrate_device_finalize(): sees !(src_pfns[i] & MIGRATE_PFN_MIGRATE), restores the folio:
>>>>>>   a) remove_migration_ptes(src, src) — re-establishes user PTEs
>>>>>>   b) folio_unlock(src)
>>>>>>   c) folio_put(src) -> -1 (drops the isolation reference)
>>>>>>
>>>>>> The split reference in 3.a is never released and the folio has a permanently elevated refcount.
>>>>>> Unless I missed a folio_put somewhere for the refcount increase in folio_isolate_lru() (2.b)?
>>>>>>
>>>>>> Please let me know if this makes sense!
>>>>>>
>>>>>> [1] https://lore.kernel.org/all/CAA1CXcDyqPPwf_-W7B+PFQtL8HdoJGCEqVsVxq7DhOUB=L4PQA@mail.gmail.com/
>>>>>>
>>>>>>> Thanks! Yes, the patch makes sense
>>>>>>>
>>>>>>> Acked-by: Balbir Singh <balbirs@nvidia.com>
>>>>>>>
>>>>>>> Balbir
>>>>> I remember stumbling on this while ago also. The folio_get() in migrate_vma_split_unmapped_folio()
>>>>> is balanced with put_page() in __split_huge_pmd_locked() (freeze = true), can't fail for device pages.
>>>>> Folios at this point are unmapped but have 1 refcount from "collecting".
>>>>> After folio_split_unmapped() the refcount(s) is still 1.
>>>>>
>>>>> So it seems the code is good as is? A comment though would be good for the extra folio_get..
>>>>>
>>>> hmm I dont think the put_page() in __split_huge_pmd_locked() is there to balance the folio_get() in
>>>> migrate_vma_split_unmapped_folio(). There are other points where split_huge_pmd_locked() is called
>>>> with freeze = true [1] and they don't get a reference before calling split_huge_pmd.
>>>>
>>>> I think the folio_put() in __split_huge_pmd_locked() freeze = true case is there as migration
>>>> entries are being installed?
>>>>
>>>> [1] https://elixir.bootlin.com/linux/v6.19.3/source/mm/rmap.c#L2334
>>>>
>>>>
>>> Yes normally you want to drop the reference when installing migration entries but in this context
>>> you have already done the collecting for the THP folio and you want to balance with the folio_get()
>>> the put_page() to keep the refs unchanged. Is that right Balbir?
>>>
>>> --Mika
>>>
>>
>> Hi Mika,
>>
>> You are right, This patch is wrong. I tried the below diff to force folio_split_unmapped to return
>> -EAGAIN. I ran tools/testing/selftests/mm/hmm-tests -r hmm.hmm_device_private.migrate_anon_huge_err
>> to trigger the path for folio_split_unmapped.
>>
>> diff --git a/mm/huge_memory.c b/mm/huge_memory.c
>> index 8e2746ea74adf..6df33b4990a13 100644
>> --- a/mm/huge_memory.c
>> +++ b/mm/huge_memory.c
>> @@ -4140,6 +4140,8 @@ int folio_split_unmapped(struct folio *folio, unsigned int new_order)
>>         if (folio_expected_ref_count(folio) != folio_ref_count(folio) - 1)
>>                 return -EAGAIN;
>>
>> +       return -EAGAIN;
>> +
>>         local_irq_disable();
>>         ret = __folio_freeze_and_split_unmapped(folio, new_order, &folio->page, NULL,
>>                                                 NULL, false, NULL, SPLIT_TYPE_UNIFORM,
>>
>>
>>
>> I inserted a lot of traces to keep track of refcounts [1]. Without this patch, I get
>> ....
>>        hmm-tests-129     [000] .....     1.476233: __migrate_device_pages: SPLIT_UNMAPPED: folio=ffc536e2c4100000 refcount=0 AFTER error NO folio_put
>>        hmm-tests-129     [000] .....     1.476234: __migrate_device_pages: PAGES: split FAILED folio=ffc536e2c4100000 refcount=0
>>        hmm-tests-129     [000] .....     1.476236: __migrate_device_finalize: FINALIZE[0]: src=ffc536e2c4100000 dst=ffc536e2c4100000 src==dst=1 refcount_src=1 mapcount_src=0 order_src=0 migrate=0 BEFORE remove_migration_ptes
>>        hmm-tests-129     [000] .....     1.476237: __migrate_device_finalize: FINALIZE[0]: src=ffc536e2c4100000 refcount=1 mapcount=0 AFTER remove_migration_ptes
>>        hmm-tests-129     [000] .....     1.476237: __migrate_device_finalize: FINALIZE[0]: src=ffc536e2c4100000 refcount=0 AFTER folio_put(src)
>>
>> i.e. refcount = 512, which is correct as split_huge_pmd_address was successful. Full output is
>> at [2].
>>
>> With this patch, I get:
>>
>> BUG: Bad rss-counter state mm:00000000cfe88d5e type:MM_FILEPAGES val:-511 Comm:bash Pid:63
>> BUG: Bad rss-counter state mm:00000000cfe88d5e type:MM_ANONPAGES val:511 Comm:bash Pid:63
>> ...
>>        hmm-tests-129     [000] .....     1.468315: __migrate_device_pages: SPLIT_UNMAPPED: folio=ffed210c840f0000 refcount=1 AFTER error folio_put FIX PRESENT
>>        hmm-tests-129     [000] .....     1.468315: __migrate_device_pages: PAGES: split FAILED folio=ffed210c840f0000 refcount=1
>>        hmm-tests-129     [000] .....     1.468318: __migrate_device_finalize: FINALIZE[0]: src=ffed210c840f0000 dst=ffed210c840f0000 src==dst=1 refcount_src=1 mapcount_src=0 order_src=9 migrate=0 BEFORE remove_migration_ptes
>>        hmm-tests-129     [000] .....     1.468357: __migrate_device_finalize: FINALIZE[0]: src=ffed210c840f0000 refcount=513 mapcount=512 AFTER remove_migration_ptes
>>        hmm-tests-129     [000] .....     1.468357: __migrate_device_finalize: FINALIZE[0]: src=ffed210c840f0000 refcount=512 AFTER folio_put(src)
>>
>> refcount=0 means the folio would be freed which is not correct. The full output is at [3].
>>
>> Thank you for clearing this up!
> 
> Thank you for doing the investigation. Can you send a patch to add a comment
> in migrate_vma_split_unmapped_folio() about this to avoid the confusion
> in the future?
> 

Yeah this was really confusing.

Does something like below look good?

diff --git a/mm/migrate_device.c b/mm/migrate_device.c
index 78c7acf024615..a302f9d3ce921 100644
--- a/mm/migrate_device.c
+++ b/mm/migrate_device.c
@@ -910,6 +910,11 @@ static int migrate_vma_split_unmapped_folio(struct migrate_vma *migrate,
 
        folio_get(folio);
        split_huge_pmd_address(migrate->vma, addr, true);
+       /*
+        * split_huge_pmd_address consumes the folio_get reference above.
+        * Therefore no folio_put is needed on the folio_split_unmapped
+        * error path.
+        */
        ret = folio_split_unmapped(folio, 0);
        if (ret)
                return ret;

>>
>>
>> [1] https://gist.github.com/uarif1/65e1e816af7aa0ae38dd6ec64d62a993
>> [2] https://gist.github.com/uarif1/79ea9500667daa4e2ef09cb5d308f041
>> [3] https://gist.github.com/uarif1/8a35a6c65ba8b3a1c1dfe72dc30e821d
> 
> 
> Best Regards,
> Yan, Zi