Message-ID: <e1f7bfa3-7418-4b4f-9339-c37e7e699c5e@redhat.com>
Date: Thu, 24 Apr 2025 14:21:34 +0200
From: David Hildenbrand <david@...hat.com>
To: Ming Wang <wangming01@...ngson.cn>,
 Andrew Morton <akpm@...ux-foundation.org>,
 Matthew Wilcox <willy@...radead.org>, Oscar Salvador <osalvador@...e.de>,
 Andrii Nakryiko <andrii@...nel.org>, Ryan Roberts <ryan.roberts@....com>,
 Christophe Leroy <christophe.leroy@...roup.eu>,
 Naoya Horiguchi <nao.horiguchi@...il.com>, Michal Hocko <mhocko@...e.cz>,
 David Rientjes <rientjes@...gle.com>, Joern Engel <joern@...fs.org>,
 Hugh Dickins <hughd@...gle.com>, linux-kernel@...r.kernel.org,
 linux-fsdevel@...r.kernel.org
Cc: Huacai Chen <chenhuacai@...nel.org>, lixuefeng@...ngson.cn,
 Hongchen Zhang <zhanghongchen@...ngson.cn>
Subject: Re: [PATCH] smaps: Fix crash in smaps_hugetlb_range for non-present
 hugetlb entries

On 23.04.25 10:14, Ming Wang wrote:
> 
> 
> On 4/23/25 15:07, David Hildenbrand wrote:
>> On 23.04.25 03:03, Ming Wang wrote:
>>> When reading /proc/pid/smaps for a process that has mapped a hugetlbfs
>>> file with MAP_PRIVATE, the kernel might crash inside
>>> pfn_swap_entry_to_page.
>>> This occurs on LoongArch under specific conditions.
>>>
>>> The root cause involves several steps:
>>> 1. When the hugetlbfs file is mapped (MAP_PRIVATE), the initial PMD
>>>      (or relevant level) entry is often populated by the kernel during
>>> mmap()
>>>      with a non-present entry pointing to the architecture's
>>> invalid_pte_table
>>>      On the affected LoongArch system, this address was observed to
>>>      be 0x90000000031e4000.
>>> 2. The smaps walker (walk_hugetlb_range -> smaps_hugetlb_range) reads
>>>      this entry.
>>> 3. The generic is_swap_pte() macro checks `!pte_present() && !
>>> pte_none()`.
>>>      The entry (invalid_pte_table address) is not present. Crucially,
>>>      the generic pte_none() check (`!(pte_val(pte) & ~_PAGE_GLOBAL)`)
>>>      returns false because the invalid_pte_table address is non-zero.
>>>      Therefore, is_swap_pte() incorrectly returns true.
>>> 4. The code enters the `else if (is_swap_pte(...))` block.
>>> 5. Inside this block, it checks `is_pfn_swap_entry()`. Due to a bit
>>>      pattern coincidence in the invalid_pte_table address on LoongArch,
>>>      the embedded generic `is_migration_entry()` check happens to return
>>>      true (misinterpreting parts of the address as a migration type).
>>> 6. This leads to a call to pfn_swap_entry_to_page() with the bogus
>>>      swap entry derived from the invalid table address.
>>> 7. pfn_swap_entry_to_page() extracts a meaningless PFN, finds an
>>>      unrelated struct page, checks its lock status (unlocked), and hits
>>>      the `BUG_ON(is_migration_entry(entry) && !PageLocked(p))` assertion.
>>>
>>> The original code's intent in the `else if` block seems aimed at handling
>>> potential migration entries, as indicated by the inner
>>> `is_pfn_swap_entry()`
>>> check. The issue arises because the outer `is_swap_pte()` check
>>> incorrectly
>>> includes the invalid table pointer case on LoongArch.
>>
>> This has a big loongarch smell to it.
>>
>> If we end up passing !pte_present() && !pte_none(), then loongarch must
>> be fixed to filter out these weird non-present entries.
>>
>> is_swap_pte() must not succeed on something that is not an actual swap pte.
>>
> 
> Hi David,
> 
> Thanks a lot for your feedback and insightful analysis!
> 
> You're absolutely right, the core issue here stems from how the generic
> is_swap_pte() macro interacts with the specific value of
> invalid_pte_table (or the equivalent invalid table entries for PMD) on
> the LoongArch architecture. I agree that this has a strong LoongArch
> characteristic.
> 
> On the affected LoongArch system, the address used for invalid_pte_table
> (observed as 0x90000000031e4000 in the vmcore) happens to satisfy both
> !pte_present() and !pte_none() conditions. This is because:
> 1. It lacks the _PAGE_PRESENT and _PAGE_PROTNONE bits (correct for an
> invalid entry).
> 2. The generic pte_none() check (`!(pte_val(pte) & ~_PAGE_GLOBAL)`)
> returns false, as the address value itself is non-zero and doesn't match
> the all-zero (except global bit) pattern.
> This causes is_swap_pte() to incorrectly return true for these
> non-mapped, initial entries set up during mmap().
> 
> The reason my proposed patch changes the condition in
> smaps_hugetlb_range() from is_swap_pte(ptent) to
> is_hugetlb_entry_migration(pte) is precisely to leverage an
> **architecture-level filtering mechanism**, as you suggested LoongArch
> should provide.
> 
> This works because is_hugetlb_entry_migration() internally calls
> `huge_pte_none()`. LoongArch **already provides** an
> architecture-specific override for huge_pte_none() (via
> `__HAVE_ARCH_HUGE_PTE_NONE`), which is defined as follows in
> arch/loongarch/include/asm/pgtable.h:
> 
> ```
> static inline int huge_pte_none(pte_t pte)
> {
>       unsigned long val = pte_val(pte) & ~_PAGE_GLOBAL;
>       /* Check for all zeros (except global) OR if it points to
> invalid_pte_table */
>       return !val || (val == (unsigned long)invalid_pte_table);
> }
> ```

There is now an alternative fix on the list, right?

https://lore.kernel.org/loongarch/20250424083037.2226732-1-wangming01@loongson.cn/T/#u

-- 
Cheers,

David / dhildenb

Powered by blists - more mailing lists