| Message ID | 20241108161312.28365-1-jgross@suse.com (mailing list archive) |
|---|---|
| State | New |
| Headers | show |
| Series | KVM/x86: don't use a literal 1 instead of RET_PF_RETRY | expand |
Queued, thanks. Paolo
On Fri, Nov 08, 2024, Paolo Bonzini wrote:
> Queued, thanks.
Noooo! Can you un-queue?
The return from kvm_mmu_page_fault() is NOT RET_PF_xxx, it's KVM outer 0/1/-errno.
I.e. '1' is saying "resume the guest", it has *nothing* to do with RET_PF_RETRY.
E.g. that path also handles RET_PF_FIXED, RET_PF_SPURIOUS, etc.
On 08.11.24 19:44, Sean Christopherson wrote: > On Fri, Nov 08, 2024, Paolo Bonzini wrote: >> Queued, thanks. > > Noooo! Can you un-queue? > > The return from kvm_mmu_page_fault() is NOT RET_PF_xxx, it's KVM outer 0/1/-errno. > I.e. '1' is saying "resume the guest", it has *nothing* to do with RET_PF_RETRY. > E.g. that path also handles RET_PF_FIXED, RET_PF_SPURIOUS, etc. And what about the existing "return RET_PF_RETRY" further up? Juergen
On Fri, Nov 08, 2024, Jürgen Groß wrote: > On 08.11.24 19:44, Sean Christopherson wrote: > > On Fri, Nov 08, 2024, Paolo Bonzini wrote: > > > Queued, thanks. > > > > Noooo! Can you un-queue? > > > > The return from kvm_mmu_page_fault() is NOT RET_PF_xxx, it's KVM outer 0/1/-errno. > > I.e. '1' is saying "resume the guest", it has *nothing* to do with RET_PF_RETRY. > > E.g. that path also handles RET_PF_FIXED, RET_PF_SPURIOUS, etc. > > And what about the existing "return RET_PF_RETRY" further up? Oof. Works by coincidence. The intent in that case is to retry the fault, but the fact that RET_PF_RETRY happens to be '1' is mostly luck. Returning a postive value other than '1' should work, but as called out by the comments for the enum, using '0' for CONTINUE isn't a hard requirement. E.g. if for some reason we used '0' for RET_PF_RETRY, this code would break. * Note, all values must be greater than or equal to zero so as not to encroach * on -errno return values. Somewhat arbitrarily use '0' for CONTINUE, which * will allow for efficient machine code when checking for CONTINUE, e.g. * "TEST %rax, %rax, JNZ", as all "stop!" values are non-zero. FWIW, you are far from the first person to complain about KVM's mostly-undocumented 0/1/-errno return encoding scheme. The problems is that it's so pervasive throughout KVM, that in some cases it's not easy to understand if a function is actually using that scheme, or just happens to return similar values. I.e. converting to enums (or #defines) would require a lot of work and churn.
On 08.11.24 23:12, Sean Christopherson wrote: > On Fri, Nov 08, 2024, Jürgen Groß wrote: >> On 08.11.24 19:44, Sean Christopherson wrote: >>> On Fri, Nov 08, 2024, Paolo Bonzini wrote: >>>> Queued, thanks. >>> >>> Noooo! Can you un-queue? >>> >>> The return from kvm_mmu_page_fault() is NOT RET_PF_xxx, it's KVM outer 0/1/-errno. >>> I.e. '1' is saying "resume the guest", it has *nothing* to do with RET_PF_RETRY. >>> E.g. that path also handles RET_PF_FIXED, RET_PF_SPURIOUS, etc. >> >> And what about the existing "return RET_PF_RETRY" further up? > > Oof. Works by coincidence. The intent in that case is to retry the fault, but > the fact that RET_PF_RETRY happens to be '1' is mostly luck. Returning a postive > value other than '1' should work, but as called out by the comments for the enum, > using '0' for CONTINUE isn't a hard requirement. E.g. if for some reason we used > '0' for RET_PF_RETRY, this code would break. I think this function is an especially awful case, as it seems to be natural to return a RET_PF_ value from a function named kvm_mmu_page_fault(). > > * Note, all values must be greater than or equal to zero so as not to encroach > * on -errno return values. Somewhat arbitrarily use '0' for CONTINUE, which > * will allow for efficient machine code when checking for CONTINUE, e.g. > * "TEST %rax, %rax, JNZ", as all "stop!" values are non-zero. > > FWIW, you are far from the first person to complain about KVM's mostly-undocumented > 0/1/-errno return encoding scheme. The problems is that it's so pervasive > throughout KVM, that in some cases it's not easy to understand if a function is > actually using that scheme, or just happens to return similar values. I.e. > converting to enums (or #defines) would require a lot of work and churn. I think it would be helpful to at least add comments to the functions returning the 0/1/-errno value. And it would be even better to have #defines for the 0 and 1. New use cases should use the #defines, and whether we convert current users is another question (I'd go for it, as it is only a minor additional work when adding the comments anyway). If you are fine with that, I can start the effort. Juergen
On 11/8/24 19:44, Sean Christopherson wrote: > On Fri, Nov 08, 2024, Paolo Bonzini wrote: >> Queued, thanks. > > Noooo! Can you un-queue? Yes, I hadn't even pushed it to kvm/queue. I applied it out of a whim but then realized that it wasn't really -rc7 material. > The return from kvm_mmu_page_fault() is NOT RET_PF_xxx, it's KVM outer 0/1/-errno. > I.e. '1' is saying "resume the guest", it has *nothing* to do with RET_PF_RETRY. > E.g. that path also handles RET_PF_FIXED, RET_PF_SPURIOUS, etc. Gah, I even checked the function and was messed up by the other "return RET_PF_RETRY". If you add X86EMUL_* to the mix, it's even worse. I had to read this three times to understand that it was *not* returning X86EMUL_CONTINUE by mistake. Can I haz strongly-typed enums like in C++?... r = kvm_check_emulate_insn(vcpu, emulation_type, insn, insn_len); if (r != X86EMUL_CONTINUE) { ... } if (!(emulation_type & EMULTYPE_NO_DECODE)) { kvm_clear_exception_queue(vcpu); if (kvm_vcpu_check_code_breakpoint(vcpu, emulation_type, &r)) return r; ... } So yeah this really has to be fixed the right way, after all even RET_PF_* started out as a conversion from 0/1. Obligatory bikeshedding, how do KVM_RET_USER and KVM_RET_GUEST sound like? Paolo
On 09.11.24 09:03, Paolo Bonzini wrote: > On 11/8/24 19:44, Sean Christopherson wrote: >> On Fri, Nov 08, 2024, Paolo Bonzini wrote: >>> Queued, thanks. >> >> Noooo! Can you un-queue? > > Yes, I hadn't even pushed it to kvm/queue. I applied it out of a whim but then > realized that it wasn't really -rc7 material. > >> The return from kvm_mmu_page_fault() is NOT RET_PF_xxx, it's KVM outer 0/1/- >> errno. >> I.e. '1' is saying "resume the guest", it has *nothing* to do with RET_PF_RETRY. >> E.g. that path also handles RET_PF_FIXED, RET_PF_SPURIOUS, etc. > > Gah, I even checked the function and was messed up by the other "return > RET_PF_RETRY". > > If you add X86EMUL_* to the mix, it's even worse. I had to read this three > times to understand that it was *not* returning X86EMUL_CONTINUE by mistake. > Can I haz strongly-typed enums like in C++?... > > r = kvm_check_emulate_insn(vcpu, emulation_type, insn, insn_len); > if (r != X86EMUL_CONTINUE) { > ... > } > > if (!(emulation_type & EMULTYPE_NO_DECODE)) { > kvm_clear_exception_queue(vcpu); > if (kvm_vcpu_check_code_breakpoint(vcpu, emulation_type, &r)) > return r; > ... > } > > So yeah this really has to be fixed the right way, after all even RET_PF_* > started out as a conversion from 0/1. > > Obligatory bikeshedding, how do KVM_RET_USER and KVM_RET_GUEST sound like? +1 Juergen
diff --git a/arch/x86/kvm/mmu/mmu.c b/arch/x86/kvm/mmu/mmu.c index 8e853a5fc867..d4a9f845b373 100644 --- a/arch/x86/kvm/mmu/mmu.c +++ b/arch/x86/kvm/mmu/mmu.c @@ -6157,7 +6157,7 @@ int noinline kvm_mmu_page_fault(struct kvm_vcpu *vcpu, gpa_t cr2_or_gpa, u64 err vcpu->stat.pf_spurious++; if (r != RET_PF_EMULATE) - return 1; + return RET_PF_RETRY; emulate: return x86_emulate_instruction(vcpu, cr2_or_gpa, emulation_type, insn,
Using a literal 1 instead of RET_PF_RETRY is not nice, fix that. Signed-off-by: Juergen Gross <jgross@suse.com> --- arch/x86/kvm/mmu/mmu.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)