Re: [Question]: major faults are still triggered after mlockall when numa balancing

[Yin, Fengwei]

On 11/14/2023 9:41 AM, Yang Shi wrote:
> On Thu, Nov 9, 2023 at 5:57 PM Yin, Fengwei <fengwei.yin@xxxxxxxxx> wrote:
>>
>>
>>
>> On 11/10/2023 6:54 AM, Yang Shi wrote:
>>> On Thu, Nov 9, 2023 at 5:48 AM zhangpeng (AS) <zhangpeng362@xxxxxxxxxx> wrote:
>>>>
>>>> Hi everyone,
>>>>
>>>> There is a performance issue that has been bothering us recently.
>>>> This problem can reproduce in the latest mainline version (Linux 6.6).
>>>>
>>>> We use mlockall(MCL_CURRENT | MCL_FUTURE) in the user mode process
>>>> to avoid performance problems caused by major fault.
>>>>
>>>> There is a stage in numa fault which will set pte as 0 in do_numa_page() :
>>>> ptep_modify_prot_start() will clear the vmf->pte, until
>>>> ptep_modify_prot_commit() assign a value to the vmf->pte.
>>>>
>>>> For the data segment of the user-mode program, the global variable area
>>>> is a private mapping. After the pagecache is loaded, the private
>>>> anonymous page is generated after the COW is triggered. Mlockall can
>>>> lock COW pages (anonymous pages), but the original file pages cannot
>>>> be locked and may be reclaimed. If the global variable (private anon page)
>>>> is accessed when vmf->pte is zero which is concurrently set by numa fault,
>>>> a file page fault will be triggered.
>>>>
>>>> At this time, the original private file page may have been reclaimed.
>>>> If the page cache is not available at this time, a major fault will be
>>>> triggered and the file will be read, causing additional overhead.
>>>>
>>>> Our problem scenario is as follows:
>>>>
>>>> task 1                      task 2
>>>> ------                      ------
>>>> /* scan global variables */
>>>> do_numa_page()
>>>>    spin_lock(vmf->ptl)
>>>>    ptep_modify_prot_start()
>>>>    /* set vmf->pte as null */
>>>>                              /* Access global variables */
>>>>                              handle_pte_fault()
>>>>                                /* no pte lock */
>>>>                                do_pte_missing()
>>>>                                  do_fault()
>>>>                                    do_read_fault()
>>>>    ptep_modify_prot_commit()
>>>>    /* ptep update done */
>>>>    pte_unmap_unlock(vmf->pte, vmf->ptl)
>>>>                                      do_fault_around()
>>>>                                      __do_fault()
>>>>                                        filemap_fault()
>>>>                                          /* page cache is not available
>>>>                                          and a major fault is triggered */
>>>>                                          do_sync_mmap_readahead()
>>>>                                          /* page_not_uptodate and goto
>>>>                                          out_retry. */
>>>>
>>>> Is there any way to avoid such a major fault?
>>>
>>> IMHO I don't think it is a bug. The man page quoted by Willy says "All
>>> mapped pages are guaranteed to be resident in RAM when the call
>>> returns successfully", but the later COW already made the file page
>>> unmapped, right? The PTE pointed to the COW'ed anon page.
>>> Hypothetically if we kept the file page mlocked and unmapped,
>>> munlock() would have not munlocked the file page at all, it would be
>>> mlocked in memory forever.
>> But in this case, even the COW page is mlocked. There is small window
>> that PTE is set to null in do_numa_page(). data segment access (it's to
>> COW page which has nothing to do with original page cache) happens in
>> this small window will trigger filemap_fault() to fault in original
>> page cache.
> 
> Yes, my point is this may not break the mlockall, but the potential
> optimization by avoiding the major fault may still stand.
Totally agree.

Regards
Yin, Fengwei

> 
>>
>> I had thought to do double check whether vmf->pte is NULL in do_read_fault().
>> But it's not reliable enough.
>>
>> Matthew's idea to use protnone to block both hardware accessing and
>> do_pte_missing() looks more promising to me.
>>
>>
>> Regards
>> Yin, Fengwei
>>
>>>
>>>>
>>>> --
>>>> Best Regards,
>>>> Peng
>>>>