[PATCH] secretmem: optimize page_is_secretmem()

[David Hildenbrand]

On 19.04.21 11:38, David Hildenbrand wrote:
> On 19.04.21 11:36, Mike Rapoport wrote:
>> On Mon, Apr 19, 2021 at 11:15:02AM +0200, David Hildenbrand wrote:
>>> On 19.04.21 10:42, Mike Rapoport wrote:
>>>> From: Mike Rapoport <rppt at linux.ibm.com>
>>>>
>>>> Kernel test robot reported -4.2% regression of will-it-scale.per_thread_ops
>>>> due to commit "mm: introduce memfd_secret system call to create "secret"
>>>> memory areas".
>>>>
>>>> The perf profile of the test indicated that the regression is caused by
>>>> page_is_secretmem() called from gup_pte_range() (inlined by gup_pgd_range):
>>>>
>>>>     27.76  +2.5  30.23       perf-profile.children.cycles-pp.gup_pgd_range
>>>>      0.00  +3.2   3.19 ± 2%  perf-profile.children.cycles-pp.page_mapping
>>>>      0.00  +3.7   3.66 ± 2%  perf-profile.children.cycles-pp.page_is_secretmem
>>>>
>>>> Further analysis showed that the slow down happens because neither
>>>> page_is_secretmem() nor page_mapping() are not inline and moreover,
>>>> multiple page flags checks in page_mapping() involve calling
>>>> compound_head() several times for the same page.
>>>>
>>>> Make page_is_secretmem() inline and replace page_mapping() with page flag
>>>> checks that do not imply page-to-head conversion.
>>>>
>>>> Reported-by: kernel test robot <oliver.sang at intel.com>
>>>> Signed-off-by: Mike Rapoport <rppt at linux.ibm.com>
>>>> ---
>>>>
>>>> @Andrew,
>>>> The patch is vs v5.12-rc7-mmots-2021-04-15-16-28, I'd appreciate if it would
>>>> be added as a fixup to the memfd_secret series.
>>>>
>>>>     include/linux/secretmem.h | 26 +++++++++++++++++++++++++-
>>>>     mm/secretmem.c            | 12 +-----------
>>>>     2 files changed, 26 insertions(+), 12 deletions(-)
>>>>
>>>> diff --git a/include/linux/secretmem.h b/include/linux/secretmem.h
>>>> index 907a6734059c..b842b38cbeb1 100644
>>>> --- a/include/linux/secretmem.h
>>>> +++ b/include/linux/secretmem.h
>>>> @@ -4,8 +4,32 @@
>>>>     #ifdef CONFIG_SECRETMEM
>>>> +extern const struct address_space_operations secretmem_aops;
>>>> +
>>>> +static inline bool page_is_secretmem(struct page *page)
>>>> +{
>>>> +	struct address_space *mapping;
>>>> +
>>>> +	/*
>>>> +	 * Using page_mapping() is quite slow because of the actual call
>>>> +	 * instruction and repeated compound_head(page) inside the
>>>> +	 * page_mapping() function.
>>>> +	 * We know that secretmem pages are not compound and LRU so we can
>>>> +	 * save a couple of cycles here.
>>>> +	 */
>>>> +	if (PageCompound(page) || !PageLRU(page))
>>>> +		return false;
>>>
>>> I'd assume secretmem pages are rare in basically every setup out there. So
>>> maybe throwing in a couple of likely()/unlikely() might make sense.
>>
>> I'd say we could do unlikely(page_is_secretmem()) at call sites. Here I can
>> hardly estimate which pages are going to be checked.
>>    
>>>> +
>>>> +	mapping = (struct address_space *)
>>>> +		((unsigned long)page->mapping & ~PAGE_MAPPING_FLAGS);
>>>> +
>>>
>>> Not sure if open-coding page_mapping is really a good idea here -- or even
>>> necessary after the fast path above is in place. Anyhow, just my 2 cents.
>>
>> Well, most if the -4.2% of the performance regression kbuild reported were
>> due to repeated compount_head(page) in page_mapping(). So the whole point
>> of this patch is to avoid calling page_mapping().
> 
> I would have thought the fast path "(PageCompound(page) ||
> !PageLRU(page))" would already avoid calling page_mapping() in many cases.

(and I do wonder if a generic page_mapping() optimization would make 
sense instead)

Willy can most probably give the best advise here :)

-- 
Thanks,

David / dhildenb