On 2022/6/14 18:33, Joao Martins wrote:
> [was out the past couple days, hence the late response]
>
> On 6/12/22 16:44, Muchun Song wrote:
>> On Sat, Jun 11, 2022 at 10:13:52AM +0800, Miaohe Lin wrote:
>>> Since commit 5232c63f46fd ("mm: Make compound_pincount always available"),
>>> compound_pincount_ptr is stored at first tail page now. So we should call
>>> prep_compound_head() after the first tail page is initialized to take
>>> advantage of the likelihood of that tail struct page being cached given
>>> that we will read them right after in prep_compound_head().
>>>
>>> Signed-off-by: Miaohe Lin <linmiaohe@xxxxxxxxxx>
>>> Cc: Joao Martins <joao.m.martins@xxxxxxxxxx>
>>> ---
>>> v2:
>>> Don't move prep_compound_head() outside loop per Joao.
>>> ---
>>> mm/page_alloc.c | 17 +++++++++++------
>>> 1 file changed, 11 insertions(+), 6 deletions(-)
>>>
>>> diff --git a/mm/page_alloc.c b/mm/page_alloc.c
>>> index 4c7d99ee58b4..048df5d78add 100644
>>> --- a/mm/page_alloc.c
>>> +++ b/mm/page_alloc.c
>>> @@ -6771,13 +6771,18 @@ static void __ref memmap_init_compound(struct page *head,
>>> set_page_count(page, 0);
>>>
>>> /*
>>> - * The first tail page stores compound_mapcount_ptr() and
>>> - * compound_order() and the second tail page stores
>>> - * compound_pincount_ptr(). Call prep_compound_head() after
>>> - * the first and second tail pages have been initialized to
>>> - * not have the data overwritten.
>>> + * The first tail page stores compound_mapcount_ptr(),
>>> + * compound_order() and compound_pincount_ptr(). Call
>>> + * prep_compound_head() after the first tail page have
>>> + * been initialized to not have the data overwritten.
>>> + *
>>> + * Note the idea to make this right after we initialize
>>> + * the offending tail pages is trying to take advantage
>>> + * of the likelihood of those tail struct pages being
>>> + * cached given that we will read them right after in
>>> + * prep_compound_head().
>>> */
>>> - if (pfn == head_pfn + 2)
>>> + if (unlikely(pfn == head_pfn + 1))
>>> prep_compound_head(head, order);
>>
>> For me it is weird not to put this out of the loop. I saw the reason
>> is because of the caching suggested by Joao. But I think this is not
>> a hot path and putting it out of the loop may be more intuitive at least
>> for me. Maybe this optimization is unnecessary (maybe I am wrong).
>
> So, depending on your setup, this might actually sit in the boot path. Yes, it is at
> bringup/teardown of new memory, so it does not sit in a 'hot path' and struct pages are
> cold. But it is part of a function that initialiazes a whole DIMM worth of struct pages.
> And PMEM dimms can be denser than RAM ones IIRC. In my case we usually have 128G PMEM
> DIMMs in our servers.
>
>> And it will be consistent with prep_compound_page() (at least it does
>> not do the similar optimization) if we drop this optimization.
>>
>> Hi Joao,
>>
>> I am wondering is it a significant optimization for zone device memory?
>> I found this code existed from the 1st version you introduced.
>
> Not quite. It did not existed in the RFC. As a matter of fact the very first
> version was totally ignoring anything cache related (i.e. just calling
> prep_compound_page() in the loop for all struct pages after all the struct pages were
> inited) until Dan suggested I fix that part because people in the past have spent time
> optimizing it.
>
>> So
>> I suspect maybe you have some numbers, would you like to share with us?
>>
>
> 128G DIMMs /with struct pages placed in DRAM/ with 2M page size used to take around
> 250-400ms. Now once you placed the struct pages in PMEM those numbers go up to 4 secs all
> the way up to 8secs (there's a lot of high variance). Now imagine 12 dimms and those
> numbers can get in the ranges of 3 - 4.8secs for DRAM-struct-pages, and with
> PMEM-struct-pages to more than 48secs.
>
> Note that initializing as compound pages brought those numbers closer in the middle
> of the interval given that we need to do more work other than just initializing the
> raw struct page. With DRAM struct pages with the vmemmap deduplication trick (which is now
> default used) these got decreased down to 80-110ms per DIMM. But I actually got started
> with numbers in the order of ~180-190ms per pmem DIMM (ignore cache effects). I should
> note that I haven't measured /exactly/ the benefit of prep_compound_head() early calling.
> But the other numbers help gauging the cache effects in this path.
>
> Earlier (in v1) I merely expressed a minor concern. /Maybe/ this matters or maybe the cost
Many thanks for your detailed explanation. In v1, I thought you do have the numbers that show
the cache-miss avoidance of the succeeding two tail page cache-lines per 2M page does matter.
That's my bad. Sorry.
> of prep_compound_head() outweighs the cache-miss avoidance of the succeeding two tail page
> cache-lines per 2M page. Well, now it's one tail page. Nonetheless, I would expect that
> this is part of the testing the submitter performs, given that this is not one of those
Am I supposed to provide the numbers that show how cache effects? The number I can provide now
will be based on the emulated pmem device due to lacking of real pmem device (because we're
under control, that's a pity :( ). That number might not be wanted because the struct pages will
always be placed in DRAM. Any suggestions? It's very kind of you if you can help provide
this number. :)
> 'no functional change' patches as written in v1 commit message :( Should that be the case,
> then let's go with v1 as that certainly brings consistency with prep_compound_page().
Many thanks!
> .
>
