| Message ID | 20170221025211.30007-1-david@gibson.dropbear.id.au (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
On Tue, Feb 21, 2017 at 01:52:11PM +1100, David Gibson wrote: > ppc_hash64_store_hpte() is used to update HPTEs in the hashed page table > (HPT) for 64-bit machines. This is used when the (emulated) CPU needs to > update the referenced (R) or changed (C) bits in the HPTE. > > Some time ago this was converted to take an HPTE index, instead of a > raw offset to the HPTE within the HPT (similar functions for 32-bit still > take an offset). In the process a serious bug was introduced: we're > still using the index parameter as though it was an offset, failing to > multiply by the size of an HPTE, so it will update bits in the wrong part > of the HPT. This can corrupt the guests's HPT, causing crashes or data > loss. > > AFAICT the reason we haven't noticed this error earlier is that for 64-bit > machines we've been testing almost exclusively with Linux guests. Linux > on ppc does not make use of the hardware R & C bits, so this writeback > will never be triggered. It also occurs only on TCG, not KVM, guests. > > Signed-off-by: David Gibson <david@gibson.dropbear.id.au> Self NACK. Sorry. In my panic, I managed to miss the multiply a few lines above. There isn't actually a bug here, just some confusing variable naming. > --- > target/ppc/mmu-hash64.c | 11 +++++++---- > 1 file changed, 7 insertions(+), 4 deletions(-) > > diff --git a/target/ppc/mmu-hash64.c b/target/ppc/mmu-hash64.c > index bb78fb5..dc3b5f7 100644 > --- a/target/ppc/mmu-hash64.c > +++ b/target/ppc/mmu-hash64.c > @@ -894,12 +894,15 @@ void ppc_hash64_store_hpte(PowerPCCPU *cpu, > > pte_index *= HASH_PTE_SIZE_64; > if (env->external_htab) { > - stq_p(env->external_htab + pte_index, pte0); > - stq_p(env->external_htab + pte_index + HASH_PTE_SIZE_64 / 2, pte1); > + stq_p(env->external_htab + pte_index * HASH_PTE_SIZE_64, pte0); > + stq_p(env->external_htab + pte_index * HASH_PTE_SIZE_64 > + + HASH_PTE_SIZE_64 / 2, pte1); > } else { > - stq_phys(CPU(cpu)->as, env->htab_base + pte_index, pte0); > stq_phys(CPU(cpu)->as, > - env->htab_base + pte_index + HASH_PTE_SIZE_64 / 2, pte1); > + env->htab_base + pte_index * HASH_PTE_SIZE_64, pte0); > + stq_phys(CPU(cpu)->as, > + env->htab_base + pte_index * HASH_PTE_SIZE_64 > + + HASH_PTE_SIZE_64 / 2, pte1); > } > } >
diff --git a/target/ppc/mmu-hash64.c b/target/ppc/mmu-hash64.c index bb78fb5..dc3b5f7 100644 --- a/target/ppc/mmu-hash64.c +++ b/target/ppc/mmu-hash64.c @@ -894,12 +894,15 @@ void ppc_hash64_store_hpte(PowerPCCPU *cpu, pte_index *= HASH_PTE_SIZE_64; if (env->external_htab) { - stq_p(env->external_htab + pte_index, pte0); - stq_p(env->external_htab + pte_index + HASH_PTE_SIZE_64 / 2, pte1); + stq_p(env->external_htab + pte_index * HASH_PTE_SIZE_64, pte0); + stq_p(env->external_htab + pte_index * HASH_PTE_SIZE_64 + + HASH_PTE_SIZE_64 / 2, pte1); } else { - stq_phys(CPU(cpu)->as, env->htab_base + pte_index, pte0); stq_phys(CPU(cpu)->as, - env->htab_base + pte_index + HASH_PTE_SIZE_64 / 2, pte1); + env->htab_base + pte_index * HASH_PTE_SIZE_64, pte0); + stq_phys(CPU(cpu)->as, + env->htab_base + pte_index * HASH_PTE_SIZE_64 + + HASH_PTE_SIZE_64 / 2, pte1); } }
ppc_hash64_store_hpte() is used to update HPTEs in the hashed page table (HPT) for 64-bit machines. This is used when the (emulated) CPU needs to update the referenced (R) or changed (C) bits in the HPTE. Some time ago this was converted to take an HPTE index, instead of a raw offset to the HPTE within the HPT (similar functions for 32-bit still take an offset). In the process a serious bug was introduced: we're still using the index parameter as though it was an offset, failing to multiply by the size of an HPTE, so it will update bits in the wrong part of the HPT. This can corrupt the guests's HPT, causing crashes or data loss. AFAICT the reason we haven't noticed this error earlier is that for 64-bit machines we've been testing almost exclusively with Linux guests. Linux on ppc does not make use of the hardware R & C bits, so this writeback will never be triggered. It also occurs only on TCG, not KVM, guests. Signed-off-by: David Gibson <david@gibson.dropbear.id.au> --- target/ppc/mmu-hash64.c | 11 +++++++---- 1 file changed, 7 insertions(+), 4 deletions(-)