On Thu 12-09-19 09:47:30, 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. It would really be preferable to summarize such a previous discussion ideally with some references. > 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. > > 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. > > 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). > > 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. > > 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. > > That was also the point where Alexander decided to look into integrating > tracking/handling reported/unreported pages directly in the buddy. OK, this gives at least some background which is really appreciated. Explaining _why_ you need something in the core MM is essential to move forward. > >> 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) > > > 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. > > Hoe that makes things clearer and am not missing something. It certainly helped me to get a better idea. I have commented on my reservations regarding the approach in this thread elsewhere but at least I _think_ I am getting a point of what you guys try to achieve. Thanks! -- Michal Hocko SUSE Labs