Re: [PATCH v9 0/8] stg mail -e --version=v9 \

[Nitesh Narayan Lal <nitesh@redhat.com>]

On 9/12/19 3:47 AM, David Hildenbrand wrote:
> On 12.09.19 09:16, Michal Hocko wrote:
>> On Wed 11-09-19 18:09:18, David Hildenbrand wrote:
>>> On 11.09.19 15:51, Michal Hocko wrote:
>>>> On Wed 11-09-19 15:20:02, Michal Hocko wrote:
>>>> [...]
>>>>>> 4. Continuously report, not the "one time report everything" approach.
>>>>> So you mean the allocator reporting this rather than an external code to
>>>>> poll right? I do not know, how much this is nice to have than must have?
>>>> Another idea that I haven't really thought through so it might turned
>>>> out to be completely bogus but let's try anyway. Your "report everything"
>>>> just made me look and realize that free_pages_prepare already performs
>>>> stuff that actually does something similar yet unrelated.
>>>>
>>>> We do report to special page poisoning, zeroying or
>>>> CONFIG_DEBUG_PAGEALLOC to unmap the address from the kernel address
>>>> space. This sounds like something fitting your model no?
>>>>
>>> AFAIKS, the poisoning/unmapping is done whenever a page is freed. I
>>> don't quite see yet how that would help to remember if a page was
>>> already reported.
>> Do you still have to differ that state when each page is reported?
> Ah, very good point. I can see that the reason for this was not
> discussed in this thread so far. (Alexander, Nitesh, please correct me
> if I am wrong). It's buried in the long history of free page
> hinting/reporting.
>
> Some early patch sets tried to report during every free synchronously.
> Free a page, report them to the hypervisor. This resulted in some issues
> (especially, locking-related and the virtio + the hypervisor being
> involved, resulting in unpredictable delays, quite some overhead ...).
> It was no good.

+1
If I remember correctly then Alexander had posted a patch-set
prior to this series where he was reporting every page of a fixed
order from __free_one_page(). But as you said it will be costly as
it will involve one hypercall per page of reporting_order.

>
> One design decision then was to not report single pages, but a bunch of
> pages at once. This made it necessary to "remember" the pages to be
> reported and to temporarily block them from getting allocated while
> reporting.

Until my v7 posting [1] I was doing this. We did not proceed with
this as blocking allocation was not recommended for reporting.

>
> Nitesh implemented (at least) two "capture PFNs of free pages in an
> array when freeing" approaches. One being synchronous from the freeing
> CPU once the list was full (having similar issues as plain synchronous
> reporting) and one being asynchronous by a separate thread (which solved
> many locking issues).

One issue with asynchronous + array approach was that it could have lead
to false OOMs due to several pages being isolated at the same time.

>
> Turned out the a simple array can quickly lead to us having to drop
> "reports" to the hypervisor because the array is full and the reporting
> thread was not able to keep up. Not good as well. Especially, if some
> process frees a lot of memory this can happen quickly and Nitesh wa
> sable to trigger this scenario frequently.

+1

>
> Finally, Nitesh decided to use the bitmap instead to keep track of pages
> to report. I'd like to note that this approach could still be combined
> with an "array of potentially free PFNs". Only when the array/circular
> buffer runs out of entries ("reporting thread cannot keep up"), we would
> have to go back to scanning the bitmap.

I will have to think about it.

> That was also the point where Alexander decided to look into integrating
> tracking/handling reported/unreported pages directly in the buddy.
>
>>> After reporting the page we would have to switch some
>>> state (Nitesh: bitmap bit, Alexander: page flag) to identify that.
>> Yes, you can either store the state somewhere.
>>
>>> Of course, we could map the page and treat that as "the state" when we
>>> reported it, but I am not sure that's such a good idea :)
>>>
>>> As always, I might be very wrong ...
>> I still do not fully understand the usecase so I might be equally wrong.
>> My thinking is along these lines. Why should you scan free pages when
>> you can effectively capture each freed page? If you go one step further
>> then post_alloc_hook would be the counterpart to know that your page has
>> been allocated.
> I'd like to note that Nitesh's patch set contains the following hunk,
> which is roughly what you were thinking :)
>
>
> -static inline void __free_one_page(struct page *page,
> +inline void __free_one_page(struct page *page,
>  		unsigned long pfn,
>  		struct zone *zone, unsigned int order,
> -		int migratetype)
> +		int migratetype, bool hint)
>  {
>  	unsigned long combined_pfn;
>  	unsigned long uninitialized_var(buddy_pfn);
> @@ -980,7 +981,8 @@ static inline void __free_one_page(struct page *page,
>  				migratetype);
>  	else
>  		add_to_free_area(page, &zone->free_area[order], migratetype);
> -
> +	if (hint)
> +		page_hinting_enqueue(page, order);
>  }
>
>
> (ignore the hint parameter, when he would switch to a isolate vs.
> alloc/free, that can go away and all we left is the enqueue part)

Precisely.
Although, there would be a scenario where the allocation will
take place for a page whose order would be < REPORTING_ORDER.
In that case, if I decide to ignore all the remaining pages and clear
the previously head free page bit, I might end
up losing the reporting opportunity.

But I can certainly look into this.

>
>
> Inside that callback we can remember the pages any way we want. Right
> now in a bitmap. Maybe later in a array + bitmap (as discussed above).
> Another idea I had was to simply go over all pages and report them when
> running into this "array full" condition. But I am not yet sure about
> the performance implications on rather large machines. So the bitmap
> idea might have some other limitations but seems to do its job.

That's correct, I was actually trying to come up with a basic framework.
Which is acceptable in terms of benefits and performance and that can fit into
most of the use-cases (if not all).
After which my plan was to further optimize it.

>
> Hoe that makes things clearer and am not missing something.

Thanks for explaining.

>
-- 
Thanks
Nitesh