On Thu, 22 Dec 2022 20:58:40 +0000, Oliver Upton <oliver.upton@xxxxxxxxx> wrote: > > On Thu, Dec 22, 2022 at 09:01:15AM +0000, Marc Zyngier wrote: > > On Wed, 21 Dec 2022 17:46:24 +0000, Oliver Upton <oliver.upton@xxxxxxxxx> wrote: > > > - When UFFD is in use, translation faults are reported to userspace as > > > writes when from a RW memslot and reads when from an RO memslot. > > > > Not quite: translation faults are reported as reads if TCR_EL1.HA > > isn't set, and as writes if it is. Ignoring TCR_EL1.HD for a moment, > > this matches exactly the behaviour of the page-table walker, which > > will update the S1 PTs only if this bit is set. > > My bad, yes you're right. I conflated the use case here with the > architectural state. > > I'm probably being way too pedantic, but I just wanted to make sure we > agree about the ensuing subtlety. More below: > > > Or is it what userfaultfd does on its own? That'd be confusing... > > > > > > > > - S1 page table memory is spuriously marked as dirty, as we presume a > > > write immediately follows the translation fault. That isn't entirely > > > senseless, as it would mean both the target page and the S1 PT that > > > maps it are both old. This is nothing new I suppose, just weird. > > > > s/old/young/ ? > > > > I think you're confusing the PT access with the access that caused the > > PT access (I'll have that printed on a t-shirt, thank you very much). > > I'd buy it! > > > Here, we're not considering the cause of the PT access anymore. If > > TCR_EL1.HA is set, the S1 PT page will be marked as accessed even on a > > read, and only that page. > > I think this is where the disconnect might be. TCR_EL1.HA == 1 suggests > a write could possibly follow, but I don't think it requires it. The > page table walker must first load the S1 PTE before writing to it. Ah, you're talking of the write to the PTE. Too many writes! My reasoning is based on Rule LFTXR in DDI0487I.a, which says: "When the PE performs a hardware update of the AF, it sets the AF to 1 in the corresponding descriptor in memory, in a coherent manner, using an atomic read-modify-write of that descriptor." An atomic-or operation fits this description, and I cannot see anything in the architecture that would prevent the write of a PTE even if AF is already set, such as mandating something like a test-and-set or compare-and-swap. I'm not saying this is the only possible implementation, or even a good one. But I don't think this is incompatible with what the architecture mandates. > > From AArch64.S1Translate() (DDI0487H.a): > > (fault, descaddress, walkstate, descriptor) = AArch64.S1Walk(fault, walkparams, va, regime, > ss, acctype, iswrite, ispriv); > > [...] > > new_desc = descriptor; > if walkparams.ha == '1' && AArch64.FaultAllowsSetAccessFlag(fault) then > // Set descriptor AF bit > new_desc<10> = '1'; > > [...] > > // Either the access flag was clear or AP<2> is set > if new_desc != descriptor then > if regime == Regime_EL10 && EL2Enabled() then > s1aarch64 = TRUE; > s2fs1walk = TRUE; > aligned = TRUE; > iswrite = TRUE; > (s2fault, descupdateaddress) = AArch64.S2Translate(fault, descaddress, s1aarch64, > ss, s2fs1walk, AccType_ATOMICRW, > aligned, iswrite, ispriv); > > if s2fault.statuscode != Fault_None then > return (s2fault, AddressDescriptor UNKNOWN); > else > descupdateaddress = descaddress; > > (fault, mem_desc) = AArch64.MemSwapTableDesc(fault, descriptor, new_desc, > walkparams.ee, descupdateaddress) > > Buried in AArch64.S1Walk() is a stage-2 walk for a read to fetch the > descriptor. The second stage-2 walk for write is conditioned on having > already fetched the stage-1 descriptor and determining the AF needs > to be set. The question is whether this is one possible implementation, or the only possible implementation. My bet is on the former. > Relating back to UFFD: if we expect KVM to do exactly what hardware > does, UFFD should see an attempted read when the first walk fails > because of an S2 translation fault. Based on this patch, though, we'd > promote it to a write if TCR_EL1.HA == 1. > > This has the additional nuance of marking the S1 PT's IPA as dirty, even > though it might not actually have been written to. Having said that, > the false positive rate should be negligible given that S1 PTs ought to > account for a small amount of guest memory. > > Like I said before, I'm probably being unnecessarily pedantic :) It just > seems to me that the view we're giving userspace of S1PTW aborts isn't > exactly architectural and I want to make sure that is explicitly > intentional. I think it is perfectly fine to be pedantic about these things, because they really matter. I still think that this change doesn't violate the architecture. But at the same time, I see some value in strictly following what the HW does, specially given that this only optimises the case where: - S1 PTs do not have a pre-existing S2 mapping (either placed there by the VMM, or swapped out) - TCR_EL1.HA==1 which is such a corner case that nobody will loose any sleep over it (and I'll buy beer to anyone who can come up with a real workload where this optimisation actually matters). So my suggestion is to drop the change altogether, and stick with the original fix. Thanks, M. -- Without deviation from the norm, progress is not possible.
