Re: [PATCH 1/4] mm: pagewalk: assert write mmap lock only for walking the user page tables

[Kefeng Wang <wangkefeng.wang@huawei.com>]

On 2023/12/2 16:08, Muchun Song wrote:
> 
> 
>> On Dec 1, 2023, at 19:09, Kefeng Wang <wangkefeng.wang@huawei.com> wrote:
>>
>>
>>
>> On 2023/11/27 16:46, Muchun Song wrote:
>>> The 8782fb61cc848 ("mm: pagewalk: Fix race between unmap and page walker")
>>> introduces an assertion to walk_page_range_novma() to make all the users
>>> of page table walker is safe. However, the race only exists for walking the
>>> user page tables. And it is ridiculous to hold a particular user mmap write
>>> lock against the changes of the kernel page tables. So only assert at least
>>> mmap read lock when walking the kernel page tables. And some users matching
>>> this case could downgrade to a mmap read lock to relief the contention of
>>> mmap lock of init_mm, it will be nicer in hugetlb (only holding mmap read
>>> lock) in the next patch.
>>> Signed-off-by: Muchun Song <songmuchun@bytedance.com>
>>> ---
>>>   mm/pagewalk.c | 29 ++++++++++++++++++++++++++++-
>>>   1 file changed, 28 insertions(+), 1 deletion(-)
>>> diff --git a/mm/pagewalk.c b/mm/pagewalk.c
>>> index b7d7e4fcfad7a..f46c80b18ce4f 100644
>>> --- a/mm/pagewalk.c
>>> +++ b/mm/pagewalk.c
>>> @@ -539,6 +539,11 @@ int walk_page_range(struct mm_struct *mm, unsigned long start,
>>>    * not backed by VMAs. Because 'unusual' entries may be walked this function
>>>    * will also not lock the PTEs for the pte_entry() callback. This is useful for
>>>    * walking the kernel pages tables or page tables for firmware.
>>> + *
>>> + * Note: Be careful to walk the kernel pages tables, the caller may be need to
>>> + * take other effective approache (mmap lock may be insufficient) to prevent
>>> + * the intermediate kernel page tables belonging to the specified address range
>>> + * from being freed (e.g. memory hot-remove).
>>>    */
>>>   int walk_page_range_novma(struct mm_struct *mm, unsigned long start,
>>>     unsigned long end, const struct mm_walk_ops *ops,
>>> @@ -556,7 +561,29 @@ int walk_page_range_novma(struct mm_struct *mm, unsigned long start,
>>>    if (start >= end || !walk.mm)
>>>    return -EINVAL;
>>>   - mmap_assert_write_locked(walk.mm);
>>> + /*
>>> + * 1) For walking the user virtual address space:
>>> + *
>>> + * The mmap lock protects the page walker from changes to the page
>>> + * tables during the walk.  However a read lock is insufficient to
>>> + * protect those areas which don't have a VMA as munmap() detaches
>>> + * the VMAs before downgrading to a read lock and actually tearing
>>> + * down PTEs/page tables. In which case, the mmap write lock should
>>> + * be hold.
>>> + *
>>> + * 2) For walking the kernel virtual address space:
>>> + *
>>> + * The kernel intermediate page tables usually do not be freed, so
>>> + * the mmap map read lock is sufficient. But there are some exceptions.
>>> + * E.g. memory hot-remove. In which case, the mmap lock is insufficient
>>> + * to prevent the intermediate kernel pages tables belonging to the
>>> + * specified address range from being freed. The caller should take
>>> + * other actions to prevent this race.
>>> + */
>>> + if (mm == &init_mm)
>>> + 	mmap_assert_locked(walk.mm);
>>> + else
>>> + 	mmap_assert_write_locked(walk.mm);
>>
>> Maybe just use process_mm_walk_lock() and set correct page_walk_lock in struct mm_walk_ops?
> 
> No. You also need to make sure the users do not pass the wrong
> walk_lock, so you also need to add something like following:
> 

But all other walk_page_XX has been converted，see more from commit
49b0638502da "mm: enable page walking API to lock vmas during the walk"),
there's nothing special about this one, the calls must pass the right
page_walk_lock to mm_walk_ops,

> if (mm == &init_mm)
> 	VM_BUG_ON(walk_lock != PGWALK_RDLOCK);
> else
> 	VM_BUG_ON(walk_lock == PGWALK_RDLOCK);
> 
> I do not think the code will be simple.

or adding the above lock check into process_mm_walk_lock too.