Re: [PATCH v6 0/6] mm / virtio: Provide support for unused page reporting

[Alexander Duyck <alexander.h.duyck@linux.intel.com>]

On Fri, 2019-08-23 at 01:16 -0400, Pankaj Gupta wrote:
> > On Thu, 2019-08-22 at 06:43 -0400, Pankaj Gupta wrote:
> > > > This series provides an asynchronous means of reporting to a hypervisor
> > > > that a guest page is no longer in use and can have the data associated
> > > > with it dropped. To do this I have implemented functionality that allows
> > > > for what I am referring to as unused page reporting
> > > > 
> > > > The functionality for this is fairly simple. When enabled it will
> > > > allocate
> > > > statistics to track the number of reported pages in a given free area.
> > > > When the number of free pages exceeds this value plus a high water value,
> > > > currently 32, it will begin performing page reporting which consists of
> > > > pulling pages off of free list and placing them into a scatter list. The
> > > > scatterlist is then given to the page reporting device and it will
> > > > perform
> > > > the required action to make the pages "reported", in the case of
> > > > virtio-balloon this results in the pages being madvised as MADV_DONTNEED
> > > > and as such they are forced out of the guest. After this they are placed
> > > > back on the free list, and an additional bit is added if they are not
> > > > merged indicating that they are a reported buddy page instead of a
> > > > standard buddy page. The cycle then repeats with additional non-reported
> > > > pages being pulled until the free areas all consist of reported pages.
> > > > 
> > > > I am leaving a number of things hard-coded such as limiting the lowest
> > > > order processed to PAGEBLOCK_ORDER, and have left it up to the guest to
> > > > determine what the limit is on how many pages it wants to allocate to
> > > > process the hints. The upper limit for this is based on the size of the
> > > > queue used to store the scattergather list.
> > > > 
> > > > My primary testing has just been to verify the memory is being freed
> > > > after
> > > > allocation by running memhog 40g on a 40g guest and watching the total
> > > > free memory via /proc/meminfo on the host. With this I have verified most
> > > > of the memory is freed after each iteration.
> > > 
> > > I tried to go through the entire patch series. I can see you reported a
> > > -3.27 drop from the baseline. If its because of re-faulting the page after
> > > host has freed them? Can we avoid freeing all the pages from the guest
> > > free_area
> > > and keep some pages(maybe some mixed order), so that next allocation is
> > > done from
> > > the guest itself than faulting to host. This will work with real workload
> > > where
> > > allocation and deallocation happen at regular intervals.
> > > 
> > > This can be further optimized based on other factors like host memory
> > > pressure etc.
> > > 
> > > Thanks,
> > > Pankaj
> > 
> > When I originally started implementing and testing this code I was seeing
> > less than a 1% regression. I didn't feel like that was really an accurate
> > result since it wasn't putting much stress on the changed code so I have
> > modified my tests and kernel so that I have memory shuffting and THP
> > enabled. In addition I have gone out of my way to lock things down to a
> > single NUMA node on my host system as the code I had would sometimes
> > perform better than baseline when running the test due to the fact that
> > memory was being freed back to the hose and then reallocated which
> > actually allowed for better NUMA locality.
> > 
> > The general idea was I wanted to know what the worst case penalty would be
> > for running this code, and it turns out most of that is just the cost of
> > faulting back in the pages. By enabling memory shuffling I am forcing the
> > memory to churn as pages are added to both the head and tail of the
> > free_list. The test itself was modified so that it didn't allocate order 0
> > pages and instead was allocating transparent huge pages so the effects
> > were as visible as possible. Without that the page faulting overhead would
> > mostly fall into the noise of having to allocate the memory as order 0
> > pages, that is what I had essentially seen earlier when I was running the
> > stock page_fault1 test.
> 
> Right. I think the reason is this test is allocating THP's in guest, host side
> you are still using order 0 pages, I assume?

No, on host side they should be huge pages as well. Most of the cost for
the fault is the page zeroing I believe since we are having to zero a 2MB
page twice, once in the host and once in the guest.

Basically if I disable THP in the guest the results are roughly half what
they are with THP enabled, and the difference between the patchset and
baseline drops to less than 1%.

> > This code does no hinting on anything smaller than either MAX_ORDER - 1 or
> > HUGETLB_PAGE_ORDER pages, and it only starts when there are at least 32 of
> > them available to hint on. This results in us not starting to perform the
> > hinting until there is 64MB to 128MB of memory sitting in the higher order
> > regions of the zone.
> 
> o.k
> 
> > The hinting itself stops as soon as we run out of unhinted pages to pull
> > from. When this occurs we let any pages that are freed after that
> > accumulate until we get back to 32 pages being free in a given order.
> > During this time we should build up the cache of warm pages that you
> > mentioned, assuming that shuffling is not enabled.
> 
> I was thinking about something like retaining pages to a lower watermark here.
> Looks like we still might have few lower order pages in free list if they are
> not merged to orders which are hinted. 

Right. We should have everything below the reporting order untouched and
as such it will not be faulted. It is only if the page gets merged back up
to reporting order that we will reported it, and only if we have at least
32 of them available.

> > As far as further optimizations I don't think there is anything here that
> > prevents us from doing that. For now I am focused on just getting the
> > basics in place so we have a foundation to start from.
> 
> Agree. Thanks for explaining.
> 
> Best rgards,
> Pankaj

Thanks.

- Alex