2022-07-15 10:06:17

by Zhou Guanghui

Subject: [PATCH] mm/huge_memory: Return from zap_huge_pmd after WARN_ONCE.

After WARN_ONCE is processed, the subsequent page judgment results
in NULL pointer access. It is more reasonable to return from the
function here.

Signed-off-by: Zhou Guanghui <[email protected]>
---
mm/huge_memory.c | 5 ++++-
1 file changed, 4 insertions(+), 1 deletion(-)

diff --git a/mm/huge_memory.c b/mm/huge_memory.c
index 834f288b3769..7f5ccca6792a 100644
--- a/mm/huge_memory.c
+++ b/mm/huge_memory.c
@@ -1601,8 +1601,11 @@ int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
entry = pmd_to_swp_entry(orig_pmd);
page = pfn_swap_entry_to_page(entry);
flush_needed = 0;
- } else
+ } else {
+ spin_unlock(ptl);
WARN_ONCE(1, "Non present huge pmd without pmd migration enabled!");
+ return 1;
+ }

if (PageAnon(page)) {
zap_deposited_table(tlb->mm, pmd);
--
2.17.1

2022-07-15 17:56:57

by Matthew Wilcox

Subject: Re: [PATCH] mm/huge_memory: Return from zap_huge_pmd after WARN_ONCE.

On Fri, Jul 15, 2022 at 09:22:38AM +0000, Zhou Guanghui wrote:
> After WARN_ONCE is processed, the subsequent page judgment results
> in NULL pointer access. It is more reasonable to return from the
> function here.

I'm not sure this is a good idea. Probably better to crash than
continue.

> Signed-off-by: Zhou Guanghui <[email protected]>
> ---
> mm/huge_memory.c | 5 ++++-
> 1 file changed, 4 insertions(+), 1 deletion(-)
>
> diff --git a/mm/huge_memory.c b/mm/huge_memory.c
> index 834f288b3769..7f5ccca6792a 100644
> --- a/mm/huge_memory.c
> +++ b/mm/huge_memory.c
> @@ -1601,8 +1601,11 @@ int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
> entry = pmd_to_swp_entry(orig_pmd);
> page = pfn_swap_entry_to_page(entry);
> flush_needed = 0;
> - } else
> + } else {
> + spin_unlock(ptl);
> WARN_ONCE(1, "Non present huge pmd without pmd migration enabled!");
> + return 1;
> + }
>
> if (PageAnon(page)) {
> zap_deposited_table(tlb->mm, pmd);
> --
> 2.17.1
>
>

2022-07-18 07:14:00

by Zhou Guanghui

Subject: Re: [PATCH] mm/huge_memory: Return from zap_huge_pmd after WARN_ONCE.

On 2022/7/16 1:50, Matthew Wilcox wrote:
> On Fri, Jul 15, 2022 at 09:22:38AM +0000, Zhou Guanghui wrote:
>> After WARN_ONCE is processed, the subsequent page judgment results
>> in NULL pointer access. It is more reasonable to return from the
>> function here.
>
> I'm not sure this is a good idea. Probably better to crash than
> continue.

Except present pmd and pmd migration entry, there should be no other possible scenarios. Whether crash or warn is unnecessary. However, the current process that crashes after a warning is reported is not reasonable.

Like this:
if (pmd_present(orig_pmd)) {
xxx
} else {
swp_entry_t entry;

VM_BUG_ON(!is_pmd_migration_entry(orig_pmd));
entry = pmd_to_swp_entry(orig_pmd);
page = pfn_to_page(swp_offset(entry));
flush_needed = 0;
}

Thanks.

>
>> Signed-off-by: Zhou Guanghui <[email protected]>
>> ---
>> mm/huge_memory.c | 5 ++++-
>> 1 file changed, 4 insertions(+), 1 deletion(-)
>>
>> diff --git a/mm/huge_memory.c b/mm/huge_memory.c
>> index 834f288b3769..7f5ccca6792a 100644
>> --- a/mm/huge_memory.c
>> +++ b/mm/huge_memory.c
>> @@ -1601,8 +1601,11 @@ int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
>> entry = pmd_to_swp_entry(orig_pmd);
>> page = pfn_swap_entry_to_page(entry);
>> flush_needed = 0;
>> - } else
>> + } else {
>> + spin_unlock(ptl);
>> WARN_ONCE(1, "Non present huge pmd without pmd migration enabled!");
>> + return 1;
>> + }
>>
>> if (PageAnon(page)) {
>> zap_deposited_table(tlb->mm, pmd);
>> --
>> 2.17.1
>>
>>
>
> .