On Tue, Aug 27, 2024 at 3:57 PM Peter Xu <peterx@xxxxxxxxxx> wrote: > > On Tue, Aug 27, 2024 at 03:36:07PM -0700, Jiaqi Yan wrote: > > Hi Peter, > > Hi, Jiaqi, > > > I am curious if there is any work needed for unmap_mapping_range? If a > > driver hugely remap_pfn_range()ed at 1G granularity, can the driver > > unmap at PAGE_SIZE granularity? For example, when handling a PFN is > > Yes it can, but it'll invoke the split_huge_pud() which default routes to > removal of the whole pud right now (currently only covers either DAX > mappings or huge pfnmaps; it won't for anonymous if it comes, for example). > > In that case it'll rely on the driver providing proper fault() / > huge_fault() to refault things back with smaller sizes later when accessed > again. I see, so the driver needs to drive the recovery process, and code needs to be in the driver. But it seems to me the recovery process will be more or less the same to different drivers? In that case does it make sense that memory_failure do the common things for all drivers? Instead of removing the whole pud, can driver or memory_failure do something similar to non-struct-page-version of split_huge_page? So driver doesn't need to re-fault good pages back? > > > poisoned in the 1G mapping, it would be great if the mapping can be > > splitted to 2M mappings + 4k mappings, so only the single poisoned PFN > > is lost. (Pretty much like the past proposal* to use HGM** to improve > > hugetlb's memory failure handling). > > Note that we're only talking about MMIO mappings here, in which case the > PFN doesn't even have a struct page, so the whole poison idea shouldn't > apply, afaiu. Yes, there won't be any struct page. Ankit proposed this patchset* for handling poisoning. I wonder if someday the vfio-nvgrace-gpu-pci driver adopts your change via new remap_pfn_range (install PMD/PUD instead of PTE), and memory_failure_pfn still unmap_mapping_range(pfn_space->mapping, pfn << PAGE_SHIFT, PAGE_SIZE, 0), can it somehow just work and no re-fault needed? * https://lore.kernel.org/lkml/20231123003513.24292-2-ankita@xxxxxxxxxx/#t > > Thanks, > > -- > Peter Xu >
