[PATCH v2 01/14] mm: pgtable: introduce pte_offset_map_{ro|rw}_nolock()

[David Hildenbrand]

On 29.08.24 12:59, Qi Zheng wrote:
> 
> 
> On 2024/8/28 18:48, David Hildenbrand wrote:
>> On 27.08.24 06:33, Qi Zheng wrote:
> 
> [...]
> 
>>> sufficient AFAIUK.
>>
>> Drop the "AFAIUK" :)
>>
>> "For R/O access this is sufficient."
>>
>>>
>>> pte_offset_map_rw_nolock(mm, pmd, addr, pmdvalp, ptlp), above, is like
>>> pte_offset_map_ro_nolock(); but when successful, it also outputs the
>>> pdmval. For R/W access, the callers can not accept that the page table
>>> it sees has been unmapped and is about to get freed. The pmdval can help
>>> callers to recheck pmd_same() to identify this case once the spinlock is
>>> taken. For some cases where exclusivity is already guaranteed, such as
>>> holding the write lock of mmap_lock, or in cases where checking is
>>> sufficient, such as a !pte_none() pte will be rechecked after the
>>> spinlock is taken, there is no need to recheck pdmval.
>>
>> Right, using pte_same() one can achieve a similar result, assuming that
>> the freed page table gets all ptes set to pte_none().
>>
>> page_table_check_pte_clear_range() before pte_free_defer() in
>> retract_page_tables/collapse_pte_mapped_thp() sanity checks that I think.
> 
> Since commit 1d65b771bc08, retract_page_tables() only holds the
> i_mmap_lock_read(mapping) but not mmap_lock, so it seems that
> holding the write lock of mmap_lock cannot guarantee the stability
> of the PTE page.

Guess it depends. khugepaged on anonymous memory will block any page 
table walkers (like anon THP collapse does) -- per-VMA lock, mmap lock, 
mapping lock/RMAP lock ... so it *should* be sufficient to hold any of 
these, right?

So at least for now, these (anonymous memory) cases would be ok. Likely 
that will change when reclaiming empty page tables.

> 
> IIUC, I will also perform a pmd_same() check on the case where the
> write lock of mmap_lock is held in v3. Or do I miss something?

Can you spell out the instances where you think it might be required.

-- 
Cheers,

David / dhildenb