On 12.02.19 10:03, Wang, Wei W wrote: > On Tuesday, February 5, 2019 4:19 AM, Nitesh Narayan Lal wrote: >> The following patch-set proposes an efficient mechanism for handing freed >> memory between the guest and the host. It enables the guests with no page >> cache to rapidly free and reclaims memory to and from the host respectively. >> >> Benefit: >> With this patch-series, in our test-case, executed on a single system and >> single NUMA node with 15GB memory, we were able to successfully launch >> atleast 5 guests when page hinting was enabled and 3 without it. (Detailed >> explanation of the test procedure is provided at the bottom). >> >> Changelog in V8: >> In this patch-series, the earlier approach [1] which was used to capture and >> scan the pages freed by the guest has been changed. The new approach is >> briefly described below: >> >> The patch-set still leverages the existing arch_free_page() to add this >> functionality. It maintains a per CPU array which is used to store the pages >> freed by the guest. The maximum number of entries which it can hold is >> defined by MAX_FGPT_ENTRIES(1000). When the array is completely filled, it >> is scanned and only the pages which are available in the buddy are stored. >> This process continues until the array is filled with pages which are part of >> the buddy free list. After which it wakes up a kernel per-cpu-thread. >> This kernel per-cpu-thread rescans the per-cpu-array for any re-allocation >> and if the page is not reallocated and present in the buddy, the kernel >> thread attempts to isolate it from the buddy. If it is successfully isolated, the >> page is added to another per-cpu array. Once the entire scanning process is >> complete, all the isolated pages are reported to the host through an existing >> virtio-balloon driver. > > Hi Nitesh, > > Have you guys thought about something like below, which would be simpler: Responding because I'm the first to stumble over this mail, hah! :) > > - use bitmaps to record free pages, e.g. xbitmap: https://lkml.org/lkml/2018/1/9/304. > The bitmap can be indexed by the guest pfn, and it's globally accessed by all the CPUs; Global means all VCPUs will be competing potentially for a single lock when freeing/allocating a page, no? What if you have 64VCPUs allocating/freeing memory like crazy? (I assume some kind of locking is required even if the bitmap would be atomic. Also, doesn't xbitmap mean that we eventually have to allocate memory at places where we don't want to - e.g. from arch_free_page ?) That's the big benefit of taking the pages of the buddy free list. Other VCPUs won't stumble over them, waiting for them to get freed in the hypervisor. > - arch_free_page(): set the bits of the freed pages from the bitmap > (no per-CPU array with hardcoded fixed length and no per-cpu scanning thread) > - arch_alloc_page(): clear the related bits from the bitmap > - expose 2 APIs for the callers: > -- unsigned long get_free_page_hints(unsigned long pfn_start, unsigned int nr); > This API searches for the next free page chunk (@nr of pages), starting from @pfn_start. > Bits of those free pages will be cleared after this function returns. > -- void put_free_page_hints(unsigned long pfn_start, unsigned int nr); > This API sets the @nr continuous bits starting from pfn_start. > > Usage example with balloon: > 1) host requests to start ballooning; > 2) balloon driver get_free_page_hints and report the hints to host via report_vq; > 3) host calls madvise(pfn_start, DONTNEED) for each reported chunk of free pages and put back pfn_start to ack_vq; > 4) balloon driver receives pfn_start and calls put_free_page_hints(pfn_start) to have the related bits from the bitmap to be set, indicating that those free pages are ready to be allocated. This sounds more like "the host requests to get free pages once in a while" compared to "the host is always informed about free pages". At the time where the host actually has to ask the guest (e.g. because the host is low on memory), it might be to late to wait for guest action. Nitesh uses MADV_FREE here (as far as I recall :) ), to only mark pages as candidates for removal and if the host is low on memory, only scanning the guest page tables is sufficient to free up memory. But both points might just be an implementation detail in the example you describe. > > In above 2), get_free_page_hints clears the bits which indicates that those pages are not ready to be used by the guest yet. Why? > This is because 3) will unmap the underlying physical pages from EPT. Normally, when guest re-visits those pages, EPT violations and QEMU page faults will get a new host page to set up the related EPT entry. If guest uses that page before the page gets unmapped (i.e. right before step 3), no EPT violation happens and the guest will use the same physical page that will be unmapped and given to other host threads. So we need to make sure that the guest free page is usable only after step 3 finishes. > > Back to arch_alloc_page(), it needs to check if the allocated pages have "1" set in the bitmap, if that's true, just clear the bits. Otherwise, it means step 2) above has happened and step 4) hasn't been reached. In this case, we can either have arch_alloc_page() busywaiting a bit till 4) is done for that page > Or better to have a balloon callback which prioritize 3) and 4) to make this page usable by the guest. Regarding the latter, the VCPU allocating a page cannot do anything if the page (along with other pages) is just being freed by the hypervisor. It has to busy-wait, no chance to prioritize. > > Using bitmaps to record free page hints don't need to take the free pages off the buddy list and return them later, which needs to go through the long allocation/free code path. > Yes, but it means that any process is able to get stuck on such a page for as long as it takes to report the free pages to the hypervisor and for it to call madvise(pfn_start, DONTNEED) on any such page. Nice idea, but I think we definitely need something the can potentially be implemented per-cpu without any global locks involved. Thanks! > Best, > Wei > -- Thanks, David / dhildenb
