Re: [PATCH] mm/mprotect: Only reference swap pfn page if type match

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On 26.08.22 18:04, Peter Xu wrote:
> On Fri, Aug 26, 2022 at 04:39:08PM +0200, David Hildenbrand wrote:
>> On 26.08.22 16:25, Peter Xu wrote:
>>> On Fri, Aug 26, 2022 at 12:49:37PM +0200, David Hildenbrand wrote:
>>>> On 24.08.22 00:11, Peter Xu wrote:
>>>>> Yu Zhao reported a bug after the commit "mm/swap: Add swp_offset_pfn() to
>>>>> fetch PFN from swap entry" added a check in swp_offset_pfn() for swap type [1]:
>>>>>
>>>>>   kernel BUG at include/linux/swapops.h:117!
>>>>>   CPU: 46 PID: 5245 Comm: EventManager_De Tainted: G S         O L 6.0.0-dbg-DEV #2
>>>>>   RIP: 0010:pfn_swap_entry_to_page+0x72/0xf0
>>>>>   Code: c6 48 8b 36 48 83 fe ff 74 53 48 01 d1 48 83 c1 08 48 8b 09 f6
>>>>>   c1 01 75 7b 66 90 48 89 c1 48 8b 09 f6 c1 01 74 74 5d c3 eb 9e <0f> 0b
>>>>>   48 ba ff ff ff ff 03 00 00 00 eb ae a9 ff 0f 00 00 75 13 48
>>>>>   RSP: 0018:ffffa59e73fabb80 EFLAGS: 00010282
>>>>>   RAX: 00000000ffffffe8 RBX: 0c00000000000000 RCX: ffffcd5440000000
>>>>>   RDX: 1ffffffffff7a80a RSI: 0000000000000000 RDI: 0c0000000000042b
>>>>>   RBP: ffffa59e73fabb80 R08: ffff9965ca6e8bb8 R09: 0000000000000000
>>>>>   R10: ffffffffa5a2f62d R11: 0000030b372e9fff R12: ffff997b79db5738
>>>>>   R13: 000000000000042b R14: 0c0000000000042b R15: 1ffffffffff7a80a
>>>>>   FS:  00007f549d1bb700(0000) GS:ffff99d3cf680000(0000) knlGS:0000000000000000
>>>>>   CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
>>>>>   CR2: 0000440d035b3180 CR3: 0000002243176004 CR4: 00000000003706e0
>>>>>   DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
>>>>>   DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400
>>>>>   Call Trace:
>>>>>    <TASK>
>>>>>    change_pte_range+0x36e/0x880
>>>>>    change_p4d_range+0x2e8/0x670
>>>>>    change_protection_range+0x14e/0x2c0
>>>>>    mprotect_fixup+0x1ee/0x330
>>>>>    do_mprotect_pkey+0x34c/0x440
>>>>>    __x64_sys_mprotect+0x1d/0x30
>>>>>
>>>>> It triggers because pfn_swap_entry_to_page() could be called upon e.g. a
>>>>> genuine swap entry.
>>>>>
>>>>> Fix it by only calling it when it's a write migration entry where the page*
>>>>> is used.
>>>>>
>>>>> [1] https://lore.kernel.org/lkml/CAOUHufaVC2Za-p8m0aiHw6YkheDcrO-C3wRGixwDS32VTS+k1w@xxxxxxxxxxxxxx/
>>>>>
>>>>> Fixes: 6c287605fd56 ("mm: remember exclusively mapped anonymous pages with PG_anon_exclusive")
>>>>> Cc: David Hildenbrand <david@xxxxxxxxxx>
>>>>> Cc: <stable@xxxxxxxxxxxxxxx>
>>>>> Reported-by: Yu Zhao <yuzhao@xxxxxxxxxx>
>>>>> Signed-off-by: Peter Xu <peterx@xxxxxxxxxx>
>>>>> ---
>>>>>  mm/mprotect.c | 3 ++-
>>>>>  1 file changed, 2 insertions(+), 1 deletion(-)
>>>>>
>>>>> diff --git a/mm/mprotect.c b/mm/mprotect.c
>>>>> index f2b9b1da9083..4549f5945ebe 100644
>>>>> --- a/mm/mprotect.c
>>>>> +++ b/mm/mprotect.c
>>>>> @@ -203,10 +203,11 @@ static unsigned long change_pte_range(struct mmu_gather *tlb,
>>>>>  			pages++;
>>>>>  		} else if (is_swap_pte(oldpte)) {
>>>>>  			swp_entry_t entry = pte_to_swp_entry(oldpte);
>>>>> -			struct page *page = pfn_swap_entry_to_page(entry);
>>>>>  			pte_t newpte;
>>>>>  
>>>>>  			if (is_writable_migration_entry(entry)) {
>>>>> +				struct page *page = pfn_swap_entry_to_page(entry);
>>>>> +
>>>>>  				/*
>>>>>  				 * A protection check is difficult so
>>>>>  				 * just be safe and disable write
>>>>
>>>>
>>>> Stumbling over the THP code, I was wondering if we also want to adjust change_huge_pmd()
>>>> and hugetlb_change_protection. There are no actual swap entries, so I assume we're fine.
>>>>
>>>>
>>>> diff --git a/mm/huge_memory.c b/mm/huge_memory.c
>>>> index 482c1826e723..466364e7fc5f 100644
>>>> --- a/mm/huge_memory.c
>>>> +++ b/mm/huge_memory.c
>>>> @@ -1798,10 +1798,10 @@ int change_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
>>>>  #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
>>>>         if (is_swap_pmd(*pmd)) {
>>>>                 swp_entry_t entry = pmd_to_swp_entry(*pmd);
>>>> -               struct page *page = pfn_swap_entry_to_page(entry);
>>>>  
>>>>                 VM_BUG_ON(!is_pmd_migration_entry(*pmd));
>>>>                 if (is_writable_migration_entry(entry)) {
>>>> +                       struct page *page = pfn_swap_entry_to_page(entry);
>>>>                         pmd_t newpmd;
>>>>                         /*
>>>>                          * A protection check is difficult so
>>>> diff --git a/mm/hugetlb.c b/mm/hugetlb.c
>>>> index 2480ba627aa5..559465fae5cd 100644
>>>> --- a/mm/hugetlb.c
>>>> +++ b/mm/hugetlb.c
>>>> @@ -6370,9 +6370,9 @@ unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
>>>>                 }
>>>>                 if (unlikely(is_hugetlb_entry_migration(pte))) {
>>>>                         swp_entry_t entry = pte_to_swp_entry(pte);
>>>> -                       struct page *page = pfn_swap_entry_to_page(entry);
>>>>  
>>>>                         if (!is_readable_migration_entry(entry)) {
>>>> +                               struct page *page = pfn_swap_entry_to_page(entry);
>>>>                                 pte_t newpte;
>>>>  
>>>>                                 if (PageAnon(page))
>>>>
>>>>
>>>> @Peter, what's your thought?
>>>
>>> IMHO they're not needed?
>>>
>>> The rule is simple in my mind: we should only pass in a pfn-typed swap
>>> entry into pfn_swap_entry_to_page() (or the new swp_offset_pfn()), or it's
>>> a violation of the API.  In these two cases they do not violate the API and
>>> they're always safe because they're guaranteed to be pfn swap entries when
>>> calling.
>>
>> I was wondering about extreme corner cases regarding the struct page.
>>
>> Assume we have a hwpoison_entry that pointed at a valid struct page. We
>> can succeed in offlining+removing the section it's located on (I was
>> recently challenging if we want to keep that behavior as it's really
>> shaky already), freeing the relevant memmap entry and the memory section.
>>
>> pfn_swap_entry_to_page() -> pfn_to_page() would be problematic if there
>> is no memmap anymore.
>>
>>
>> I assume it's ok to always call it for is_pfn_swap_entry(), but in the
>> PMD case we only check for is_swap_pmd()? Isn't that problematic?
> 
> I don't know extensively enough on hwpoison on validity of fetching page
> from pfn inside on online/offline ops, but.. if the only concern is about
> hwpoison entry existance here I think its fine?  Because iirc we'l split
> thp when any of the subpage got poisoned, so we should never hit a hwpoison
> entry in thp path.

Ah right, so we're good. Thanks!

-- 
Thanks,

David / dhildenb





[Index of Archives]     [Linux ARM Kernel]     [Linux ARM]     [Linux Omap]     [Fedora ARM]     [IETF Annouce]     [Bugtraq]     [Linux OMAP]     [Linux MIPS]     [eCos]     [Asterisk Internet PBX]     [Linux API]

  Powered by Linux