Hello Sean, On 1/11/23 20:05, Sean Christopherson wrote: > On Thu, Aug 18, 2022, Christian König wrote: >> Am 18.08.22 um 01:13 schrieb Dmitry Osipenko: >>> On 8/18/22 01:57, Dmitry Osipenko wrote: >>>> On 8/15/22 18:54, Dmitry Osipenko wrote: >>>>> On 8/15/22 17:57, Dmitry Osipenko wrote: >>>>>> On 8/15/22 16:53, Christian König wrote: >>>>>>> Am 15.08.22 um 15:45 schrieb Dmitry Osipenko: >>>>>>>> [SNIP] >>>>>>>>> Well that comment sounds like KVM is doing the right thing, so I'm >>>>>>>>> wondering what exactly is going on here. >>>>>>>> KVM actually doesn't hold the page reference, it takes the temporal >>>>>>>> reference during page fault and then drops the reference once page is >>>>>>>> mapped, IIUC. Is it still illegal for TTM? Or there is a possibility for >>>>>>>> a race condition here? >>>>>>>> >>>>>>> Well the question is why does KVM grab the page reference in the first >>>>>>> place? >>>>>>> >>>>>>> If that is to prevent the mapping from changing then yes that's illegal >>>>>>> and won't work. It can always happen that you grab the address, solve >>>>>>> the fault and then immediately fault again because the address you just >>>>>>> grabbed is invalidated. >>>>>>> >>>>>>> If it's for some other reason than we should probably investigate if we >>>>>>> shouldn't stop doing this. > > ... > >>>>> If we need to bump the refcount only for VM_MIXEDMAP and not for >>>>> VM_PFNMAP, then perhaps we could add a flag for that to the kvm_main >>>>> code that will denote to kvm_release_page_clean whether it needs to put >>>>> the page? >>>> The other variant that kind of works is to mark TTM pages reserved using >>>> SetPageReserved/ClearPageReserved, telling KVM not to mess with the page >>>> struct. But the potential consequences of doing this are unclear to me. >>>> >>>> Christian, do you think we can do it? >>> Although, no. It also doesn't work with KVM without additional changes >>> to KVM. >> >> Well my fundamental problem is that I can't fit together why KVM is grabing >> a page reference in the first place. > > It's to workaround a deficiency in KVM. > >> See the idea of the page reference is that you have one reference is that >> you count the reference so that the memory is not reused while you access >> it, e.g. for I/O or mapping it into different address spaces etc... >> >> But none of those use cases seem to apply to KVM. If I'm not totally >> mistaken in KVM you want to make sure that the address space mapping, e.g. >> the translation between virtual and physical address, don't change while you >> handle it, but grabbing a page reference is the completely wrong approach >> for that. > > TL;DR: 100% agree, and we're working on fixing this in KVM, but were still months > away from a full solution. > > Yep. KVM uses mmu_notifiers to react to mapping changes, with a few caveats that > we are (slowly) fixing, though those caveats are only tangentially related. > > The deficiency in KVM is that KVM's internal APIs to translate a virtual address > to a physical address spit out only the resulting host PFN. The details of _how_ > that PFN was acquired are not captured. Specifically, KVM loses track of whether > or not a PFN was acquired via gup() or follow_pte() (KVM is very permissive when > it comes to backing guest memory). > > Because gup() gifts the caller a reference, that means KVM also loses track of > whether or not KVM holds a page refcount. To avoid pinning guest memory, KVM does > quickly put the reference gifted by gup(), but because KVM doesn't _know_ if it > holds a reference, KVM uses a heuristic, which is essentially "is the PFN associated > with a 'normal' struct page?". > > /* > * Returns a 'struct page' if the pfn is "valid" and backed by a refcounted > * page, NULL otherwise. Note, the list of refcounted PG_reserved page types > * is likely incomplete, it has been compiled purely through people wanting to > * back guest with a certain type of memory and encountering issues. > */ > struct page *kvm_pfn_to_refcounted_page(kvm_pfn_t pfn) > > That heuristic also triggers if follow_pte() resolves to a PFN that is associated > with a "struct page", and so to avoid putting a reference it doesn't own, KVM does > the silly thing of manually getting a reference immediately after follow_pte(). > > And that in turn gets tripped up non-refcounted tail pages because KVM sees a > normal, valid "struct page" and assumes it's refcounted. To fudge around that > issue, KVM requires "struct page" memory to be refcounted. > > The long-term solution is to refactor KVM to precisely track whether or not KVM > holds a reference. Patches have been prosposed to do exactly that[1], but they > were put on hold due to the aforementioned caveats with mmu_notifiers. The > caveats are that most flows where KVM plumbs a physical address into hardware > structures aren't wired up to KVM's mmu_notifier. > > KVM could support non-refcounted struct page memory without first fixing the > mmu_notifier issues, but I was (and still am) concerned that that would create an > even larger hole in KVM until the mmu_notifier issues are sorted out[2]. > > [1] https://lore.kernel.org/all/20211129034317.2964790-1-stevensd@xxxxxxxxxx > [2] https://lore.kernel.org/all/Ydhq5aHW+JFo15UF@xxxxxxxxxx Thanks for the summary! Indeed, it's the KVM side that needs to be patched. Couple months ago I found that a non-TTM i915 driver also suffers from the same problem because it uses huge pages that we want map to a guest. So we definitely will need to fix the KVM side. -- Best regards, Dmitry
