Message-ID: <80148272-208A-4306-A8EC-98C79BBBDCA0@nvidia.com>
Date: Fri, 26 Apr 2024 10:49:51 -0400
From: Zi Yan <ziy@...dia.com>
To: Ryan Roberts <ryan.roberts@....com>
Cc: Anshuman Khandual <anshuman.khandual@....com>,
 Andrew Morton <akpm@...ux-foundation.org>,
 "Aneesh Kumar K.V" <aneesh.kumar@...nel.org>, linux-mm@...ck.org,
 linux-kernel@...r.kernel.org
Subject: Re: [PATCH v1] mm: Fix race between __split_huge_pmd_locked() and
 GUP-fast

On 26 Apr 2024, at 3:43, Ryan Roberts wrote:

> On 26/04/2024 05:19, Anshuman Khandual wrote:
>> On 4/25/24 22:37, Ryan Roberts wrote:
>>> __split_huge_pmd_locked() can be called for a present THP, devmap or
>>> (non-present) migration entry. It calls pmdp_invalidate()
>>> unconditionally on the pmdp and only determines if it is present or not
>>> based on the returned old pmd. This is a problem for the migration entry
>>> case because pmd_mkinvalid(), called by pmdp_invalidate() must only be
>>> called for a present pmd.
>>
>> pmdp_invalidate() must be called only for present PMD - is this expected
>> by core MM ? Does this cause any problem otherwise ?
>
> I'm saying that only calling pmdp_invalidate() on a pte_present()==true pte is
> the only semantic that makes sense. And, yes, it causes a problem if called on a
> pte_present()==false pte - that's exactly what I'm describing in this commit log.
>
> To labour the point, this is the logical type hierachy of PTEs (and block-mapped
> PMDs) as I see it:
>
> ---8<----
>
> pte
>  |- present
>  |   |- valid
>  |   |- invalid
>  |
>  |- not_present
>      |- none
>      |- swap_pte
>
> present: All fields must be interpretted the way the HW sees them. e.g.
>          pte_pfn(), pte_write(), pte_dirty(), pte_young(), pte_mkwrite(),
>          pte_mkold() can all be legitimately used to query and modify the pte.
>
>   valid: The HW may access the pte, interpret the fields and create a TLB entry,
>          etc.
>
>   invalid: The HW will never access the pte or create a TLB entry for it.
>
> not_present: The fields are SW-defined. HW never accesses the PTE.
>
>   none: Unused; represents a hole
>
>   swap_pte: Contains a swap entry and swap pte bits. The contained swap entry
>             may 1 of a few different types e.g. actual swap entry, migration
>             entry, hw poison, etc.
>
> ---8<----
>
> We test present vs not_present with pte_present()
>
> We test none vs swap_pte with pte_none()
>
> valid vs invalid is slightly more vague. The core-mm can move a PMD from valid
> -> invalid by calling pmd_mkinvalid(). But it can't query the state. And it
> can't do this generically for a PTE.
>
>
> Based on that lot, it makes no sense to me that we should permit calling
> pmd_mkinvalid() on a non-present pte. Indeed, we don't permit calling
> pte_mkwrite() etc on a non-present pte. And those functions are not defensive;
> they don't check that the pte is present before making the change. They just
> trust that the core-mm will not call them for non-present ptes.

I am OK with disallowing to call pmd_mkinvalid() on a non-present entry, but
would like to know how to enforce it or document it. Because x86, risc-v, mips,
and loongarch can call pmd_mkinvalid() on a non-present entry without causing
any issue, any developer who work on these arches but arm64 can use pmd_mkinvalid()
improperly until someone else tests it on arm64. You will need to add VM_WARM_ON()
to all arch versions of pmd_mkinvalid().

>
> The alternative approach would be to make pmdp_invalidate() defensive so that it
> checks the pmd is present before making any changes. But it doesn't semantically
> make sense to invalidate a non-present pmd in the first place so why call it
> under these circumstances? There is also a practical problem in that some arches
> implement their own pmdp_invalidate() so you would want to make all those
> defensive too, which would grow the size of the change.

Like I said above, if you do not do this, other arches developers can break arm64
without knowing it, since their pmd_mkinvalid() do not change a non-present
PMD to a present one.

>>
>>>
>>> On arm64 at least, pmd_mkinvalid() will mark the pmd such that any
>>> future call to pmd_present() will return true. And therefore any
>>
>> IIRC the following semantics needs to be followed as expected by core MM.
>>
>> -------------------------------------------------------------------------
>> |	PMD states	|	pmd_present	|	pmd_trans_huge	|
>> -------------------------------------------------------------------------
>> |	Mapped		|	Yes		|	Yes		|
>> -------------------------------------------------------------------------
>> |	Splitting	|	Yes		|	Yes		|
>> -------------------------------------------------------------------------
>> |	Migration/Swap	|	No		|	No		|
>> -------------------------------------------------------------------------
>
> Indeed, the problem, as I see it, is if pmd_mkinvalid() is called on a
> "Migration/Swap" pmd, then a future call to pmd_present() will return Yes, which
> is clearly wrong. pmd_trans_huge() will also return Yes due to:
>
> static inline int pmd_trans_huge(pmd_t pmd)
> {
> 	return pmd_val(pmd) && pmd_present(pmd) && !(pmd_val(pmd) & PMD_TABLE_BIT);
> }
>
> At least this happens for arm64. Although Zi suggests other arches look like
> they will do this too in the other email.
>
> The reason is that arm64's pmd_mkinvalid() unconditionally sets
> PMD_PRESENT_INVALID (bit 59) and clears PMD_SECT_VALID (bit 0) in the pte. So
> next time pmd_present() is called it will see PMD_PRESENT_INVALID is set and
> return true.
>
>>
>>
>>> lockless pgtable walker could see the migration entry pmd in this state
>>> and start interpretting the fields as if it were present, leading to
>>> BadThings (TM). GUP-fast appears to be one such lockless pgtable walker.
>>
>> Could you please explain how bad things might happen ?
>
> See 2 places where pmdp_get_lockless() is called in gup.c, without the PTL.
> These could both return the swap pte for which pmd_mkinvalid() has been called.
> In both cases, this would lead to the pmd_present() check eroneously returning
> true, eventually causing incorrect interpretation of the pte fields. e.g.:
>
> gup_pmd_range()
>   pmd_t pmd = pmdp_get_lockless(pmdp);
>   gup_huge_pmd(pmd, ...)
>     page = nth_page(pmd_page(orig), (addr & ~PMD_MASK) >> PAGE_SHIFT);
>
> page is guff.
>
> Let me know what you think!
>
> Thanks,
> Ryan
>
>
>>
>>> I suspect the same is possible on other architectures.
>>>
>>> Fix this by only calling pmdp_invalidate() for a present pmd. And for
>>> good measure let's add a warning to the generic implementation of
>>> pmdp_invalidate(). I've manually reviewed all other
>>> pmdp_invalidate[_ad]() call sites and believe all others to be
>>> conformant.
>>>
>>> This is a theoretical bug found during code review. I don't have any
>>> test case to trigger it in practice.
>>>
>>> Fixes: 84c3fc4e9c56 ("mm: thp: check pmd migration entry in common path")
>>> Signed-off-by: Ryan Roberts <ryan.roberts@....com>
>>> ---
>>>
>>> Applies on top of v6.9-rc5. Passes all the mm selftests on arm64.
>>>
>>> Thanks,
>>> Ryan
>>>
>>>
>>>  mm/huge_memory.c     | 5 +++--
>>>  mm/pgtable-generic.c | 2 ++
>>>  2 files changed, 5 insertions(+), 2 deletions(-)
>>>
>>> diff --git a/mm/huge_memory.c b/mm/huge_memory.c
>>> index 89f58c7603b2..80939ad00718 100644
>>> --- a/mm/huge_memory.c
>>> +++ b/mm/huge_memory.c
>>> @@ -2513,12 +2513,12 @@ static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
>>>  	 * for this pmd), then we flush the SMP TLB and finally we write the
>>>  	 * non-huge version of the pmd entry with pmd_populate.
>>>  	 */
>>> -	old_pmd = pmdp_invalidate(vma, haddr, pmd);
>>>
>>> -	pmd_migration = is_pmd_migration_entry(old_pmd);
>>> +	pmd_migration = is_pmd_migration_entry(*pmd);
>>>  	if (unlikely(pmd_migration)) {
>>>  		swp_entry_t entry;
>>>
>>> +		old_pmd = *pmd;
>>>  		entry = pmd_to_swp_entry(old_pmd);
>>>  		page = pfn_swap_entry_to_page(entry);
>>>  		write = is_writable_migration_entry(entry);
>>> @@ -2529,6 +2529,7 @@ static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
>>>  		soft_dirty = pmd_swp_soft_dirty(old_pmd);
>>>  		uffd_wp = pmd_swp_uffd_wp(old_pmd);
>>>  	} else {
>>> +		old_pmd = pmdp_invalidate(vma, haddr, pmd);
>>>  		page = pmd_page(old_pmd);
>>>  		folio = page_folio(page);
>>>  		if (pmd_dirty(old_pmd)) {
>>> diff --git a/mm/pgtable-generic.c b/mm/pgtable-generic.c
>>> index 4fcd959dcc4d..74e34ea90656 100644
>>> --- a/mm/pgtable-generic.c
>>> +++ b/mm/pgtable-generic.c
>>> @@ -198,6 +198,7 @@ pgtable_t pgtable_trans_huge_withdraw(struct mm_struct *mm, pmd_t *pmdp)
>>>  pmd_t pmdp_invalidate(struct vm_area_struct *vma, unsigned long address,
>>>  		     pmd_t *pmdp)
>>>  {
>>> +	VM_WARN_ON(!pmd_present(*pmdp));
>>>  	pmd_t old = pmdp_establish(vma, address, pmdp, pmd_mkinvalid(*pmdp));
>>>  	flush_pmd_tlb_range(vma, address, address + HPAGE_PMD_SIZE);
>>>  	return old;
>>> @@ -208,6 +209,7 @@ pmd_t pmdp_invalidate(struct vm_area_struct *vma, unsigned long address,
>>>  pmd_t pmdp_invalidate_ad(struct vm_area_struct *vma, unsigned long address,
>>>  			 pmd_t *pmdp)
>>>  {
>>> +	VM_WARN_ON(!pmd_present(*pmdp));
>>>  	return pmdp_invalidate(vma, address, pmdp);
>>>  }
>>>  #endif
>>> --
>>> 2.25.1
>>>
>>>


--
Best Regards,
Yan, Zi

Download attachment "signature.asc" of type "application/pgp-signature" (855 bytes)

Powered by blists - more mailing lists