KVM: PPC: Book3S HV: Prevent POWER7/8 TLB flush flushing SLB
Subsystems (2)
Nicholas Piggin
reviewerPATCH 1/1 KVM: PPC: Book3S HV: Prevent POWER7/8 TLB flush flushing SLB
The POWER9 ERAT flush instruction is a SLBIA with IH=7, which is a reserved value on POWER7/8. On POWER8 this invalidates the SLB entries above index 0, similarly to SLBIA IH=0.
If the SLB entries are invalidated, and then the guest is bypassed, the host SLB does not get re-loaded, so the bolted entries above 0 will be lost. This can result in kernel stack access causing a SLB fault.
Kernel stack access causing a SLB fault was responsible for the infamous mega bug (search "Fix SLB reload bug"). Although since commit 48e7b7695745 ("powerpc/64s/hash: Convert SLB miss handlers to C") that starts using the kernel stack in the SLB miss handler, it might only result in an infinite loop of SLB faults. In any case it's a bug.
Fix this by only executing the instruction on >= POWER9 where IH=7 is defined not to invalidate the SLB. POWER7/8 don't require this ERAT flush.
--- arch/powerpc/kvm/book3s_hv_builtin.c | 5 ++++- 1 file changed, 4 insertions(+), 1 deletion(-)
arch/powerpc/kvm/book3s_hv_builtin.c +4 -1
@@ -695,6 +695,7 @@ static void flush_guest_tlb(struct kvm *kvm) "r" (0) : "memory"); } asmvolatile ("ptesync" : : :"memory" ); +// POWER9 congruence-class TLBIEL leaves ERAT. Flush it now. asm volatile(PPC_RADIX_INVALIDATE_ERAT_GUEST : : :"memory"); }else {for (set =0 ; set < kvm->arch.tlb_sets; ++set) {
@@ -705,7 +706,9 @@ static void flush_guest_tlb(struct kvm *kvm) rb +=PPC_BIT (51 );/* increment set number */ } asmvolatile ("ptesync" : : :"memory" ); - asm volatile(PPC_ISA_3_0_INVALIDATE_ERAT : : :"memory"); +// POWER9 congruence-class TLBIEL leaves ERAT. Flush it now. +if (cpu_has_feature (CPU_FTR_ARCH_300 )) + asm volatile(PPC_ISA_3_0_INVALIDATE_ERAT : : :"memory"); } }
Signature · 1 line
2.23.0
Quoted 14 lines · Nicholas Piggin
The POWER9 ERAT flush instruction is a SLBIA with IH=7, which is a reserved value on POWER7/8. On POWER8 this invalidates the SLB entries above index 0, similarly to SLBIA IH=0.
If the SLB entries are invalidated, and then the guest is bypassed, the host SLB does not get re-loaded, so the bolted entries above 0 will be lost. This can result in kernel stack access causing a SLB fault.
Kernel stack access causing a SLB fault was responsible for the infamous mega bug (search "Fix SLB reload bug"). Although since commit
48e7b7695745("powerpc/64s/hash: Convert SLB miss handlers to C") that starts using the kernel stack in the SLB miss handler, it might only result in an infinite loop of SLB faults. In any case it's a bug.
Quoted 6 lines
Fix this by only executing the instruction on >= POWER9 where IH=7 is defined not to invalidate the SLB. POWER7/8 don't require this ERAT flush.
Fixes 5008711259201 ("KVM: PPC: Book3S HV: Invalidate ERAT when flushing guest TLB entries")Signed-off-by Nicholas Piggin <[email protected]>
Quoted 26 lines
--- arch/powerpc/kvm/book3s_hv_builtin.c | 5 ++++- 1 file changed, 4 insertions(+), 1 deletion(-)arch/powerpc/kvm/book3s_hv_builtin.c +4
-1@@ -695,6 +695,7 @@ static void flush_guest_tlb(struct kvm *kvm) "r" (0) : "memory"); } asmvolatile ("ptesync" : : :"memory" ); +// POWER9 congruence-class TLBIEL leaves ERAT. Flush it now. asm volatile(PPC_RADIX_INVALIDATE_ERAT_GUEST : : :"memory"); }else {for (set =0 ; set < kvm->arch.tlb_sets; ++set) {@@ -705,7 +706,9 @@ static void flush_guest_tlb(struct kvm *kvm) rb +=PPC_BIT (51 );/* increment set number */ } asmvolatile ("ptesync" : : :"memory" ); - asm volatile(PPC_ISA_3_0_INVALIDATE_ERAT : : :"memory"); +// POWER9 congruence-class TLBIEL leaves ERAT. Flush it now. +if (cpu_has_feature (CPU_FTR_ARCH_300 )) + asm volatile(PPC_ISA_3_0_INVALIDATE_ERAT : : :"memory"); } }
Michael Ellerman
AppliedQuoted 9 lines · Nicholas Piggin
The POWER9 ERAT flush instruction is a SLBIA with IH=7, which is a reserved value on POWER7/8. On POWER8 this invalidates the SLB entries above index 0, similarly to SLBIA IH=0.
If the SLB entries are invalidated, and then the guest is bypassed, the host SLB does not get re-loaded, so the bolted entries above 0 will be lost. This can result in kernel stack access causing a SLB fault.
[...]
Applied to powerpc/fixes.
[1/1] KVM: PPC: Book3S HV: Prevent POWER7/8 TLB flush flushing SLB
https://git.kernel.org/powerpc/c/cf0b0e3712f7af90006f8317ff27278094c2c128
cheers