Re: [PATCH] mm: remove redundant lru_add_drain() prior to unmapping pages

[Jianfeng Wang <jianfeng.w.wang@oracle.com>]

On 12/14/23 9:57 AM, Tim Chen wrote:
> On Wed, 2023-12-13 at 17:03 -0800, Jianfeng Wang wrote:
>> On 12/13/23 2:57 PM, Tim Chen wrote:
>>> On Tue, 2023-12-12 at 23:28 -0800, Jianfeng Wang wrote:
>>>> When unmapping VMA pages, pages will be gathered in batch and released by
>>>> tlb_finish_mmu() if CONFIG_MMU_GATHER_NO_GATHER is not set. The function
>>>> tlb_finish_mmu() is responsible for calling free_pages_and_swap_cache(),
>>>> which calls lru_add_drain() to drain cached pages in folio_batch before
>>>> releasing gathered pages. Thus, it is redundant to call lru_add_drain()
>>>> before gathering pages, if CONFIG_MMU_GATHER_NO_GATHER is not set.
>>>>
>>>> Remove lru_add_drain() prior to gathering and unmapping pages in
>>>> exit_mmap() and unmap_region() if CONFIG_MMU_GATHER_NO_GATHER is not set.
>>>>
>>>> Note that the page unmapping process in oom_killer (e.g., in
>>>> __oom_reap_task_mm()) also uses tlb_finish_mmu() and does not have
>>>> redundant lru_add_drain(). So, this commit makes the code more consistent.
>>>>
>>>> Signed-off-by: Jianfeng Wang <jianfeng.w.wang@oracle.com>
>>>> ---
>>>>  mm/mmap.c | 4 ++++
>>>>  1 file changed, 4 insertions(+)
>>>>
>>>> diff --git a/mm/mmap.c b/mm/mmap.c
>>>> index 1971bfffcc03..0451285dee4f 100644
>>>> --- a/mm/mmap.c
>>>> +++ b/mm/mmap.c
>>>> @@ -2330,7 +2330,9 @@ static void unmap_region(struct mm_struct *mm, struct ma_state *mas,
>>>>  	struct mmu_gather tlb;
>>>>  	unsigned long mt_start = mas->index;
>>>>  
>>>> +#ifdef CONFIG_MMU_GATHER_NO_GATHER
>>>
>>> In your comment you say skip lru_add_drain() when CONFIG_MMU_GATHER_NO_GATHER
>>> is *not* set.  So shouldn't this be
>>>
>>> #ifndef CONFIG_MMU_GATHER_NO_GATHER ?
>>>
>> Hi Tim,
>>
>> The mmu_gather feature is used to gather pages produced by unmap_vmas() and
>> release them in batch in tlb_finish_mmu(). The feature is *on* if
>> CONFIG_MMU_GATHER_NO_GATHER is *not* set. Note that: tlb_finish_mmu() will call
>> free_pages_and_swap_cache()/lru_add_drain() only when the feature is on.
> 
> Thanks for the explanation.
> 
> Looking at the code, lru_add_drain() is executed for #ifndef CONFIG_MMU_GATHER_NO_GATHER
> in tlb_finish_mmu(). So the logic of your patch is fine.
>  
> The #ifndef CONFIG_MMU_GATHER_NO_GATHER means
> mmu_gather feature is on.  The double negative throws me off on in my first read
> of your commit log.
> 
> Suggest that you add a comment in code to make it easier for
> future code maintenence:
> 
> /* defer lru_add_drain() to tlb_finish_mmu() for ifndef CONFIG_MMU_GATHER_NO_GATHER */
> 
> Is your change of skipping the extra lru_add_drain() motivated by some performance reason
> in a workload? Wonder whether it is worth adding an extra ifdef in the code.
> 
> Tim
> 

Okay, great suggestion.

We observe heavy contention on the LRU lock, introduced by lru_add_drain() and
release_pages() for a prod workload, and we're trying to reduce the level of
contention.

lru_add_drain() is a complex function that first takes a local CPU lock and
iterate through *all* folio_batches to see if there are pages to be moved to
and between LRU lists. At that point, any page in these folio_batches will
trigger acquiring the per-LRU spin lock and increase the level of lock
contention. Applying the change can avoid calling lru_add_drain() unnecessarily,
which is a source of lock contention. Together with the comment line suggested
by you, I believe this also increases code readability to clarify the
mmu_gather feature.

- Jianfeng

>>
>> Yes, this commit aims to skip lru_add_drain() when CONFIG_MMU_GATHER_NO_GATHER
>> is *not* set (i.e. when the mmu_gather feature is on) because it is redundant. 
>>
>> If CONFIG_MMU_GATHER_NO_GATHER is set, pages will be released in unmap_vmas().
>> tlb_finish_mmu() will not call lru_add_drain(). So, it is still necessary to
>> keep the lru_add_drain() call to clear cached pages before unmap_vmas(), as
>> folio_batchs hold a reference count for pages in them.
>>
>> The same applies to the other case.
>>
>> Thanks,
>> - Jianfeng
>>
>>>>  	lru_add_drain();
>>>> +#endif
>>>>  	tlb_gather_mmu(&tlb, mm);
>>>>  	update_hiwater_rss(mm);
>>>>  	unmap_vmas(&tlb, mas, vma, start, end, tree_end, mm_wr_locked);
>>>> @@ -3300,7 +3302,9 @@ void exit_mmap(struct mm_struct *mm)
>>>>  		return;
>>>>  	}
>>>>  
>>>> +#ifdef CONFIG_MMU_GATHER_NO_GATHER
>>>
>>> same question as above.
>>>
>>>>  	lru_add_drain();
>>>> +#endif
>>>>  	flush_cache_mm(mm);
>>>>  	tlb_gather_mmu_fullmm(&tlb, mm);
>>>>  	/* update_hiwater_rss(mm) here? but nobody should be looking */
>>>
>