On 2022/4/15 11:07, ying.huang@xxxxxxxxx wrote: > On Wed, 2022-04-13 at 09:26 +0800, ying.huang@xxxxxxxxx wrote: >> On Tue, 2022-04-12 at 16:59 +0200, David Hildenbrand wrote: >>> On 12.04.22 15:42, Miaohe Lin wrote: >>>> On 2022/4/12 16:59, Oscar Salvador wrote: >>>>> On Sat, Apr 09, 2022 at 05:34:53PM +0800, Miaohe Lin wrote: >>>>>> We don't need to check can_split_folio() because folio_maybe_dma_pinned() >>>>>> is checked before. It will avoid the long term pinned pages to be swapped >>>>>> out. And we can live with short term pinned pages. Without can_split_folio >>>>>> checking we can simplify the code. Also activate_locked can be changed to >>>>>> keep_locked as it's just short term pinning. >>>>> >>>>> What do you mean by "we can live with short term pinned pages"? >>>>> Does it mean that it was not pinned when we check >>>>> folio_maybe_dma_pinned() but now it is? >>>>> >>>>> To me it looks like the pinning is fluctuating and we rely on >>>>> split_folio_to_list() to see whether we succeed or not, and if not >>>>> we give it another spin in the next round? >>>> >>>> Yes. Short term pinned pages is relative to long term pinned pages and these pages won't be >>>> pinned for a noticeable time. So it's expected to split the folio successfully in the next >>>> round as the pinning is really fluctuating. Or am I miss something? >>>> >>> >>> Just so we're on the same page. folio_maybe_dma_pinned() only capture >>> FOLL_PIN, but not FOLL_GET. You can have long-term FOLL_GET right now >>> via vmsplice(). >> >> Per my original understanding, folio_maybe_dma_pinned() can be used to >> detect long-term pinned pages. And it seems reasonable to skip the >> long-term pinned pages and try short-term pinned pages during page >> reclaiming. But as you pointed out, vmsplice() doesn't use FOLL_PIN. >> So if vmsplice() is expected to pin pages for long time, and we have no >> way to detect it, then we should keep can_split_folio() in the original >> code. >> >> Copying more people who have worked on long-term pinning for comments. > > Checked the discussion in the following thread, > > https://lore.kernel.org/lkml/CA+CK2bBffHBxjmb9jmSKacm0fJMinyt3Nhk8Nx6iudcQSj80_w@xxxxxxxxxxxxxx/ > > It seems that from the practical point of view, folio_maybe_dma_pinned() > can identify most long-term pinned pages that may block memory hot- > remove or CMA allocation. Although as David pointed out, some pages may > still be GUPed for long time (e.g. via vmsplice) even if > !folio_maybe_dma_pinned(). > > But from another point of view, can_split_huge_page() is cheap and THP > swapout is expensive (swap space, disk IO, and hard to be recovered), so > it may be better to keep can_split_huge_page() in shink_page_list(). Many thanks for your explanation. Looks convincing for me. Is it worth a comment about the above stuff? Anyway, will drop this patch. Thanks! > > Best Regards, > Huang, Ying > >> >>> can_split_folio() is more precise then folio_maybe_dma_pinned(), but >>> both are racy as long as the page is still mapped. >>> >>> >> > > > . >
