Hey Zi and Jason, Thanks a lot for reaching out! On Thu, May 9, 2024 at 12:35 AM Jason Gunthorpe <jgg@xxxxxxxxxx> wrote: > > On Wed, May 08, 2024 at 12:22:08PM -0400, Zi Yan wrote: > > On 8 May 2024, at 11:52, Jason Gunthorpe wrote: > > > > > On Wed, May 08, 2024 at 10:56:34AM -0400, Zi Yan wrote: > > > > > >> Lance is improving try_to_unmap_one() to support unmapping PMD THP as a whole, > > >> so he moves split_huge_pmd_address() inside while (page_vma_mapped_walk(&pvmw)) > > >> and after mmu_notifier_invalidate_range_start() as split_huge_pmd_locked() > > >> and does not include the mmu notifier ops inside split_huge_pmd_address(). IMO, It might be reasonable to exclude the mmu notifier ops in split_huge_pmd_locked(). IIUC, before acquiring the PTL, callers need to tear down the secondary mappings via mmu_notifier_invalidate_range_start() with the range aligned to HPAGE_PMD_SIZE. > > >> I wonder if that could cause issues, since the mmu_notifier_invalidate_range_start() > > >> before the while loop only has range of the original address and > > >> split huge pmd can affect the entire PMD address range and these two ranges > > >> might not be the same. As Baolin mentioned [1] before: "For a PMD mapped THP, I think the address is already THP size alignment returned from vma_address(&folio->page, vma)." Given this, perhaps we don't need to re-align the input address after starting the pagewalk? IMO, if any corner cases arise, we should catch them by using VM_WARN_ON_ONCE() in split_huge_pmd_locked(). Zi, what do you think? [1] https://lore.kernel.org/linux-mm/cc9fd23f-7d87-48a7-a737-acbea8e95fb7@xxxxxxxxxxxxxxxxx/ > > > > > > That does not sound entirely good.. > > > > > > I suppose it depends on what split does, if the MM page table has the > > > same translation before and after split then perhaps no invalidation > > > is even necessary. > > > > Before split, it is a PMD mapping to a PMD THP (order-9). After split, > > they are 512 PTEs mapping to the same THP. Unless the secondary TLB > > does not support PMD mapping and use 512 PTEs instead, it seems to > > be an issue from my understanding. > > I may not recall fully, but I don't think any secondaries are > so sensitive to the PMD/PTE distinction.. At least the ones using > hmm_range_fault() are not. > > When the PTE eventually comes up for invalidation then the secondary > should wipe out any granual they may have captured. > > Though, perhaps KVM should be checked carefully. > > > In terms of two mmu_notifier ranges, first is in the split_huge_pmd_address()[1] > > and second is in try_to_unmap_one()[2]. When try_to_unmap_one() is unmapping > > a subpage in the middle of a PMD THP, the former notifies about the PMD range > > change due to one PMD split into 512 PTEs and the latter only needs to notify > > about the invalidation of the unmapped PTE. I do not think the latter can > > replace the former, although a potential optimization can be that the latter > > can be removed as it is included in the range of the former. > > I think we probably don't need both, either size might be fine, but > the larger size is definately fine.. > > > Regarding Lance's current code change, is it OK to change mmu_notifier range > > after mmu_notifier_invalidate_range_start()? > > No, it cannot be changed during a start/stop transaction. I understood and will keep that in mind - thanks! Thanks again for clarifying! Lance > > Jason > >
