On 09.04.19 15:37, Michael S. Tsirkin wrote: > On Tue, Apr 09, 2019 at 03:36:08PM +0200, David Hildenbrand wrote: >> On 09.04.19 15:31, Michael S. Tsirkin wrote: >>> On Tue, Apr 09, 2019 at 11:20:36AM +0200, David Hildenbrand wrote: >>>> BTW I like the idea of allocating pages that have already been hinted as >>>> last "choice", allocating pages that have not been hinted yet first. >>> >>> OK I guess but note this is just a small window during which >>> not all pages have been hinted. >> >> Yes, good point. It might sound desirable but might be completely >> irrelevant in practice. >> >>> >>> So if we actually think this has value then we need >>> to design something that will desist and not drop pages >>> in steady state too. >> >> By dropping, you mean dropping hints of e.g. MAX_ORDER - 1 or e.g. not >> reporting MAX_ORDER - 3? > > I mean the issue is host unmaps the pages from guest right? That is > what makes hinted pages slower than non-hinted ones. If we do not want > that to happen for some pages, then either host can defer acting on the > hint, or we can defer hinting. Ah right, I think what Alex mentioned is that pages processed in the hypervisor via MADVISE_FREE will be set RO, so the kernel can detect if they will actually be used again, resulting int 1. A pagefault if written and the page(s) have not been reused by the host 2. A pagefault if read/written if the page(s) have been reused by the host Now, assuming hinting is fast, most pages will be hinted right away and therefore result in pagefaults. I think this is what you meant. Deferring processing in the hypervisor cannot be done after a request has been marked as processed and handed back to the guest. (otherwise pages might already get reused) So pages would have to "age" in the guest instead before they might be worth hinting. Marking pages as "Offline" alone won't help. Agreed. -- Thanks, David / dhildenb
