target/ppc: Fix slbia TLB invalidation gap
PATCH 1/2 target/ppc: Fix slbia TLB invalidation gap
slbia must invalidate TLBs even if it does not remove a valid SLB entry, because slbmte can overwrite valid entries without removing their TLBs.
As the architecture says, slbia invalidates all lookaside information, not conditionally based on if it removed valid entries.
It does not seem possible for POWER8 or earlier Linux kernels to hit this bug because it never changes its kernel SLB translations, and it should always have valid entries if any accesses are made to usespace regions. However other operating systems which may modify SLB entry 0 or do more fancy things with segments might be affected.
When POWER9 slbia support is added in the next patch, this becomes a real problem because some new slbia variants don't invalidate all non-zero entries.
--- target/ppc/mmu-hash64.c | 21 +++++++++++++++------ 1 file changed, 15 insertions(+), 6 deletions(-)
target/ppc/mmu-hash64.c +15 -6
@@ -100,20 +100,29 @@ void helper_slbia(CPUPPCState *env)PowerPCCPU *cpu =env_archcpu (env);int n; +/* +* slbia must always flush all TLB (which is equivalent to ERAT in ppc +* architecture). Matching on SLB_ESID_V is not good enough, because slbmte +* can overwrite a valid SLB without flushing its lookaside information. +* +* It would be possible to keep the TLB in synch with the SLB by flushing +* when a valid entry is overwritten by slbmte, and therefore slbia would +* not have to flush unless it evicts a valid SLB entry. However it is +* expected that slbmte is more common than slbia, and slbia is usually +* going to evict valid SLB entries, so that tradeoff is unlikely to be a +* good one. +*/ +/* XXX: Warning: slbia never invalidates the first segment */ for (n =1 ; n < cpu->hash64_opts->slb_size; n++) {ppc_slb_t *slb = &env->slb[n];if (slb->esid &SLB_ESID_V ) { slb->esid &= ~SLB_ESID_V ; -/* -* XXX: given the fact that segment size is 256 MB or 1TB, -* and we still don't have a tlb_flush_mask(env, n, mask) -* in QEMU, we just invalidate all TLBs -*/ - env->tlb_need_flush |=TLB_NEED_LOCAL_FLUSH ; } } + + env->tlb_need_flush |=TLB_NEED_LOCAL_FLUSH ; } static void __helper_slbie(CPUPPCState *env, target_ulong addr,
Signature · 1 line
2.23.0
PATCH 2/2 target/ppc: Fix ISA v3.0 (POWER9) slbia implementation
Linux using the hash MMU ("disable_radix" command line) on a POWER9 machine quickly hits translation bugs due to using v3.0 slbia features that are not implemented in TCG. Add them.
--- target/ppc/helper.h | 2 +- target/ppc/mmu-hash64.c | 57 ++++++++++++++++++++++++++++++++++++----- target/ppc/translate.c | 5 +++- 3 files changed, 55 insertions(+), 9 deletions(-)
target/ppc/helper.h +1 -1
@@ -615,7 +615,7 @@ DEF_HELPER_FLAGS_3(store_slb, TCG_CALL_NO_RWG, void, env, tl, tl) DEF_HELPER_2(load_slb_esid, tl, env, tl) DEF_HELPER_2(load_slb_vsid, tl, env, tl) DEF_HELPER_2(find_slb_vsid, tl, env, tl) -DEF_HELPER_FLAGS_1(slbia, TCG_CALL_NO_RWG, void, env) +DEF_HELPER_FLAGS_2(slbia, TCG_CALL_NO_RWG, void, env, i32) DEF_HELPER_FLAGS_2(slbie, TCG_CALL_NO_RWG, void, env, tl) DEF_HELPER_FLAGS_2(slbieg, TCG_CALL_NO_RWG, void, env, tl)#endif
target/ppc/mmu-hash64.c +50 -7
83 lines not on this page · raw message
target/ppc/translate.c +4 -1
@@ -5027,12 +5027,15 @@ static void gen_tlbsync(DisasContext *ctx)/* slbia */ static void gen_slbia (DisasContext *ctx) { +uint32_t ih = (ctx->opcode >>21 ) &0x7 ; +TCGv_i32 t0 =tcg_const_i32 (ih); +#if defined(CONFIG_USER_ONLY )GEN_PRIV ;#else CHK_SV ; -gen_helper_slbia (cpu_env); +gen_helper_slbia (cpu_env, t0);#endif /* defined(CONFIG_USER_ONLY) */ }
Signature · 1 line
2.23.0
Quoted 4 lines · Nicholas Piggin
slbia must invalidate TLBs even if it does not remove a valid SLB entry, because slbmte can overwrite valid entries without removing their TLBs.
Quoted 8 lines
As the architecture says, slbia invalidates all lookaside information, not conditionally based on if it removed valid entries.
It does not seem possible for POWER8 or earlier Linux kernels to hit this bug because it never changes its kernel SLB translations, and it should always have valid entries if any accesses are made to usespace regions. However other operating systems which may modify SLB entry 0 or do more fancy things with segments might be affected.
Did you hit the bug on the other OS ?
Quoted 5 lines
When POWER9 slbia support is added in the next patch, this becomes a real problem because some new slbia variants don't invalidate all non-zero entries.
Signed-off-by Nicholas Piggin <[email protected]>
Looks correct.
Thanks,
C.
Quoted 45 lines
--- target/ppc/mmu-hash64.c | 21 +++++++++++++++------ 1 file changed, 15 insertions(+), 6 deletions(-)target/ppc/mmu-hash64.c +15
-6@@ -100,20 +100,29 @@ void helper_slbia(CPUPPCState *env)PowerPCCPU *cpu =env_archcpu (env);int n; +/* +* slbia must always flush all TLB (which is equivalent to ERAT in ppc +* architecture). Matching on SLB_ESID_V is not good enough, because slbmte +* can overwrite a valid SLB without flushing its lookaside information. +* +* It would be possible to keep the TLB in synch with the SLB by flushing +* when a valid entry is overwritten by slbmte, and therefore slbia would +* not have to flush unless it evicts a valid SLB entry. However it is +* expected that slbmte is more common than slbia, and slbia is usually +* going to evict valid SLB entries, so that tradeoff is unlikely to be a +* good one. +*/ +/* XXX: Warning: slbia never invalidates the first segment */ for (n =1 ; n < cpu->hash64_opts->slb_size; n++) {ppc_slb_t *slb = &env->slb[n];if (slb->esid &SLB_ESID_V ) { slb->esid &= ~SLB_ESID_V ; -/* -* XXX: given the fact that segment size is 256 MB or 1TB, -* and we still don't have a tlb_flush_mask(env, n, mask) -* in QEMU, we just invalidate all TLBs -*/ - env->tlb_need_flush |=TLB_NEED_LOCAL_FLUSH ; } } + + env->tlb_need_flush |=TLB_NEED_LOCAL_FLUSH ; } static void __helper_slbie(CPUPPCState *env, target_ulong addr,
Quoted 10 lines · On Wed, 18 Mar 2020 14:41:34 +1000 Nicholas Piggin <[email protected]> wrote:
slbia must invalidate TLBs even if it does not remove a valid SLB entry, because slbmte can overwrite valid entries without removing their TLBs.
As the architecture says, slbia invalidates all lookaside information, not conditionally based on if it removed valid entries.
It does not seem possible for POWER8 or earlier Linux kernels to hit this bug because it never changes its kernel SLB translations, and it should always have valid entries if any accesses are made to usespace
s/usespace/userspace
Quoted 9 lines
regions. However other operating systems which may modify SLB entry 0 or do more fancy things with segments might be affected.
When POWER9 slbia support is added in the next patch, this becomes a real problem because some new slbia variants don't invalidate all non-zero entries.
Signed-off-by Nicholas Piggin <[email protected]>---
LGTM
Quoted 43 lines
target/ppc/mmu-hash64.c | 21 +++++++++++++++------ 1 file changed, 15 insertions(+), 6 deletions(-)target/ppc/mmu-hash64.c +15
-6@@ -100,20 +100,29 @@ void helper_slbia(CPUPPCState *env)PowerPCCPU *cpu =env_archcpu (env);int n; +/* +* slbia must always flush all TLB (which is equivalent to ERAT in ppc +* architecture). Matching on SLB_ESID_V is not good enough, because slbmte +* can overwrite a valid SLB without flushing its lookaside information. +* +* It would be possible to keep the TLB in synch with the SLB by flushing +* when a valid entry is overwritten by slbmte, and therefore slbia would +* not have to flush unless it evicts a valid SLB entry. However it is +* expected that slbmte is more common than slbia, and slbia is usually +* going to evict valid SLB entries, so that tradeoff is unlikely to be a +* good one. +*/ +/* XXX: Warning: slbia never invalidates the first segment */ for (n =1 ; n < cpu->hash64_opts->slb_size; n++) {ppc_slb_t *slb = &env->slb[n];if (slb->esid &SLB_ESID_V ) { slb->esid &= ~SLB_ESID_V ; -/* -* XXX: given the fact that segment size is 256 MB or 1TB, -* and we still don't have a tlb_flush_mask(env, n, mask) -* in QEMU, we just invalidate all TLBs -*/ - env->tlb_need_flush |=TLB_NEED_LOCAL_FLUSH ; } } + + env->tlb_need_flush |=TLB_NEED_LOCAL_FLUSH ; } static void __helper_slbie(CPUPPCState *env, target_ulong addr,
Re: [EXTERNAL] [PATCH 2/2] target/ppc: Fix ISA v3.0 (POWER9) slbia implementation
In reply to Nicholas PigginQuoted 3 lines · Nicholas Piggin
Linux using the hash MMU ("
disable_radix" command line) on a POWER9 machine quickly hits translation bugs due to using v3.0 slbia features that are not implemented in TCG. Add them.
I checked the ISA books and this looks OK but you are also modifying slbie.
Thanks,
C.
Quoted 131 lines
Signed-off-by Nicholas Piggin <[email protected]>--- target/ppc/helper.h | 2 +- target/ppc/mmu-hash64.c | 57 ++++++++++++++++++++++++++++++++++++----- target/ppc/translate.c | 5 +++- 3 files changed, 55 insertions(+), 9 deletions(-)target/ppc/helper.h +1
-1@@ -615,7 +615,7 @@ DEF_HELPER_FLAGS_3(store_slb, TCG_CALL_NO_RWG, void, env, tl, tl) DEF_HELPER_2(load_slb_esid, tl, env, tl) DEF_HELPER_2(load_slb_vsid, tl, env, tl) DEF_HELPER_2(find_slb_vsid, tl, env, tl) -DEF_HELPER_FLAGS_1(slbia, TCG_CALL_NO_RWG, void, env) +DEF_HELPER_FLAGS_2(slbia, TCG_CALL_NO_RWG, void, env, i32) DEF_HELPER_FLAGS_2(slbie, TCG_CALL_NO_RWG, void, env, tl) DEF_HELPER_FLAGS_2(slbieg, TCG_CALL_NO_RWG, void, env, tl)#endif target/ppc/mmu-hash64.c +50
-7@@ -95,9 +95,10 @@ void dump_slb(PowerPCCPU *cpu) } } -void helper_slbia (CPUPPCState *env) +void helper_slbia (CPUPPCState *env,uint32_t ih) {PowerPCCPU *cpu =env_archcpu (env); +int starting_entry;int n;/* @@ -111,18 +112,59 @@ void helper_slbia(CPUPPCState *env)* expected that slbmte is more common than slbia, and slbia is usually * going to evict valid SLB entries, so that tradeoff is unlikely to be a * good one. +* +* ISA v2.05 introduced IH field with values 0,1,2,6. These all invalidate +* the same SLB entries (everything but entry 0), but differ in what +* "lookaside information" is invalidated. TCG can ignore this and flush +* everything. +* +* ISA v3.0 introduced additional values 3,4,7, which change what SLBs are +* invalidated. */ -/* XXX: Warning: slbia never invalidates the first segment */ -for (n =1 ; n < cpu->hash64_opts->slb_size; n++) { -ppc_slb_t *slb = &env->slb[n]; + env->tlb_need_flush |=TLB_NEED_LOCAL_FLUSH ; + + starting_entry =1 ;/* default for IH=0,1,2,6 */ + +if (env->mmu_model ==POWERPC_MMU_3_00 ) { +switch (ih) { +case 0x7 : +/* invalidate no SLBs, but all lookaside information */ +return ; -if (slb->esid &SLB_ESID_V ) { - slb->esid &= ~SLB_ESID_V ; +case 0x3 : +case 0x4 : +/* also considers SLB entry 0 */ + starting_entry =0 ; +break ; + +case 0x5 : +/* treat undefined values as ih==0, and warn */ +qemu_log_mask (LOG_GUEST_ERROR , +"slbia undefined IH field %u.\n" , ih); +break ; + +default : +/* 0,1,2,6 */ +break ; } } - env->tlb_need_flush |=TLB_NEED_LOCAL_FLUSH ; +for (n = starting_entry; n < cpu->hash64_opts->slb_size; n++) { +ppc_slb_t *slb = &env->slb[n]; + +if (!(slb->esid &SLB_ESID_V )) { +continue ; + } +if (env->mmu_model ==POWERPC_MMU_3_00 ) { +if (ih ==0x3 && (slb->vsid &SLB_VSID_C ) ==0 ) { +/* preserves entries with a class value of 0 */ +continue ; + } + } + + slb->esid &= ~SLB_ESID_V ; + } } static void __helper_slbie(CPUPPCState *env, target_ulong addr,@@ -136,6 +178,7 @@ static void __helper_slbie(CPUPPCState *env, target_ulong addr,return ; } + env->tlb_need_flush |=TLB_NEED_LOCAL_FLUSH ;if (slb->esid &SLB_ESID_V ) { slb->esid &= ~SLB_ESID_V ;target/ppc/translate.c +4
-1@@ -5027,12 +5027,15 @@ static void gen_tlbsync(DisasContext *ctx)/* slbia */ static void gen_slbia (DisasContext *ctx) { +uint32_t ih = (ctx->opcode >>21 ) &0x7 ; +TCGv_i32 t0 =tcg_const_i32 (ih); +#if defined(CONFIG_USER_ONLY )GEN_PRIV ;#else CHK_SV ; -gen_helper_slbia (cpu_env); +gen_helper_slbia (cpu_env, t0);#endif /* defined(CONFIG_USER_ONLY) */ }
Re: [EXTERNAL] [PATCH 2/2] target/ppc: Fix ISA v3.0 (POWER9) slbia implementation
Quoted 7 lines · Cédric Le Goater
Quoted 3 lines · Nicholas Piggin
Linux using the hash MMU ("
disable_radix" command line) on a POWER9 machine quickly hits translation bugs due to using v3.0 slbia features that are not implemented in TCG. Add them.I checked the ISA books and this looks OK but you are also modifying slbie.
For the same reason, I believe slbie needs to invalidate caches even if the entry isn't present.
The kernel will under some circumstances overwrite SLB entries without invalidating (because the translation itself isn't invalid, it's just that the SLB is full, so anything cached in the ERAT is still technically ok).
However, when those things get really invalidated, they need to be taken out, even if they no longer have a corresponding SLB entry.
Cheers,
Ben.
Quoted 147 lines
Thanks,
C.
Quoted 142 lines
Signed-off-by Nicholas Piggin <[email protected]>--- target/ppc/helper.h | 2 +- target/ppc/mmu-hash64.c | 57 ++++++++++++++++++++++++++++++++++++- ---- target/ppc/translate.c | 5 +++- 3 files changed, 55 insertions(+), 9 deletions(-)target/ppc/helper.h +1
-1@@ -615,7 +615,7 @@ DEF_HELPER_FLAGS_3(store_slb, TCG_CALL_NO_RWG, void, env, tl, tl) DEF_HELPER_2(load_slb_esid, tl, env, tl) DEF_HELPER_2(load_slb_vsid, tl, env, tl) DEF_HELPER_2(find_slb_vsid, tl, env, tl) -DEF_HELPER_FLAGS_1(slbia, TCG_CALL_NO_RWG, void, env) +DEF_HELPER_FLAGS_2(slbia, TCG_CALL_NO_RWG, void, env, i32) DEF_HELPER_FLAGS_2(slbie, TCG_CALL_NO_RWG, void, env, tl) DEF_HELPER_FLAGS_2(slbieg, TCG_CALL_NO_RWG, void, env, tl)#endif
target/ppc/mmu-hash64.c +2
-1@@ -95,9 +95,10 @@ void dump_slb(PowerPCCPU *cpu) } } -void helper_slbia (CPUPPCState *env) +void helper_slbia (CPUPPCState *env,uint32_t ih) {PowerPCCPU *cpu =env_archcpu (env); +int starting_entry;int n;/* @@ -111,18 +112,59 @@ void helper_slbia(CPUPPCState *env)* expected that slbmte is more common than slbia, and slbia is usually
* going to evict valid SLB entries, so that tradeoff is
unlikely to be a
* good one.target/ppc/mmu-hash64.c +2
-0+* +* ISA v2.05 introduced IH field with values 0,1,2,6. These all invalidate
+ * the same SLB entries (everything but entry 0), but differ
in what
+ * "lookaside information" is invalidated. TCG can ignore this
and flushtarget/ppc/mmu-hash64.c +3
-0+* everything. +* +* ISA v3.0 introduced additional values 3,4,7, which change what SLBs are
target/ppc/mmu-hash64.c +9
-3+* invalidated. */ -/* XXX: Warning: slbia never invalidates the first segment */ -for (n =1 ; n < cpu->hash64_opts->slb_size; n++) { -ppc_slb_t *slb = &env->slb[n]; + env->tlb_need_flush |=TLB_NEED_LOCAL_FLUSH ; + + starting_entry =1 ;/* default for IH=0,1,2,6 */ + +if (env->mmu_model ==POWERPC_MMU_3_00 ) { +switch (ih) { +case 0x7 : +/* invalidate no SLBs, but all lookaside information */
target/ppc/mmu-hash64.c +17
-3+return ; -if (slb->esid &SLB_ESID_V ) { - slb->esid &= ~SLB_ESID_V ; +case 0x3 : +case 0x4 : +/* also considers SLB entry 0 */ + starting_entry =0 ; +break ; + +case 0x5 : +/* treat undefined values as ih==0, and warn */ +qemu_log_mask (LOG_GUEST_ERROR , +"slbia undefined IH field %u.\n" , ih); +break ; + +default : +/* 0,1,2,6 */ +break ; } } - env->tlb_need_flush |=TLB_NEED_LOCAL_FLUSH ; +for (n = starting_entry; n < cpu->hash64_opts->slb_size; n++){
target/ppc/mmu-hash64.c +15
-0+ppc_slb_t *slb = &env->slb[n]; + +if (!(slb->esid &SLB_ESID_V )) { +continue ; + } +if (env->mmu_model ==POWERPC_MMU_3_00 ) { +if (ih ==0x3 && (slb->vsid &SLB_VSID_C ) ==0 ) { +/* preserves entries with a class value of 0 */ +continue ; + } + } + + slb->esid &= ~SLB_ESID_V ; + } } static void __helper_slbie(CPUPPCState *env, target_ulong addr,@@ -136,6 +178,7 @@ static void __helper_slbie(CPUPPCState *env,target_ulong addr, return; } + env->tlb_need_flush |=TLB_NEED_LOCAL_FLUSH ;if (slb->esid &SLB_ESID_V ) { slb->esid &= ~SLB_ESID_V ;target/ppc/translate.c +4
-1@@ -5027,12 +5027,15 @@ static void gen_tlbsync(DisasContext *ctx)/* slbia */ static void gen_slbia (DisasContext *ctx) { +uint32_t ih = (ctx->opcode >>21 ) &0x7 ; +TCGv_i32 t0 =tcg_const_i32 (ih); +#if defined(CONFIG_USER_ONLY )GEN_PRIV ;#else CHK_SV ; -gen_helper_slbia (cpu_env); +gen_helper_slbia (cpu_env, t0);#endif /* defined(CONFIG_USER_ONLY) */ }
Cédric Le Goater's on March 19, 2020 2:45 am:
Quoted 8 lines · Nicholas Piggin
Quoted 7 lines
slbia must invalidate TLBs even if it does not remove a valid SLB entry, because slbmte can overwrite valid entries without removing their TLBs.
As the architecture says, slbia invalidates all lookaside information, not conditionally based on if it removed valid entries.
Quoted 7 lines
Quoted 5 lines
It does not seem possible for POWER8 or earlier Linux kernels to hit this bug because it never changes its kernel SLB translations, and it should always have valid entries if any accesses are made to usespace regions. However other operating systems which may modify SLB entry 0 or do more fancy things with segments might be affected.
Did you hit the bug on the other OS ?
No, hit it when fixing POWER9 hash.
Quoted 10 lines
Quoted 6 lines
When POWER9 slbia support is added in the next patch, this becomes a real problem because some new slbia variants don't invalidate all non-zero entries.
Signed-off-by Nicholas Piggin <[email protected]>Looks correct.
Reviewed-by Cédric Le Goater <[email protected]>
Thanks,
Nick
Re: [EXTERNAL] [PATCH 2/2] target/ppc: Fix ISA v3.0 (POWER9) slbia implementation
In reply to Benjamin HerrenschmidtBenjamin Herrenschmidt's on March 19, 2020 6:46 am:
Quoted 8 lines · Cédric Le Goater
Quoted 7 lines · Nicholas Piggin
Quoted 3 lines
Linux using the hash MMU ("
disable_radix" command line) on a POWER9 machine quickly hits translation bugs due to using v3.0 slbia features that are not implemented in TCG. Add them.I checked the ISA books and this looks OK but you are also modifying slbie.
That was a mistake that leaked in from debugging the crashes.
Quoted 2 lines
For the same reason, I believe slbie needs to invalidate caches even if the entry isn't present.
I don't think it does per the ISA. If we overwrite it then we can only invalidate with slbia. That's why there is that slb insertion cache for pre-POWER9 that lets us use slbies to context switch so long as none have been overwritten.
Quoted 7 lines
The kernel will under some circumstances overwrite SLB entries without invalidating (because the translation itself isn't invalid, it's just that the SLB is full, so anything cached in the ERAT is still technically ok).
However, when those things get really invalidated, they need to be taken out, even if they no longer have a corresponding SLB entry.
Yeah we track that and do slbia in that case.
Thanks,
Nick
On Wed, Mar 18, 2020 at 05:52:32PM +0100, Greg Kurz wrote:
65;5803;1c> On Wed, 18 Mar 2020 14:41:34 +1000
Quoted 6 lines
Quoted 4 lines · Nicholas Piggin
slbia must invalidate TLBs even if it does not remove a valid SLB entry, because slbmte can overwrite valid entries without removing their TLBs.
Quoted 8 lines
Quoted 6 lines
As the architecture says, slbia invalidates all lookaside information, not conditionally based on if it removed valid entries.
It does not seem possible for POWER8 or earlier Linux kernels to hit this bug because it never changes its kernel SLB translations, and it should always have valid entries if any accesses are made to usespace
s/usespace/userspace
Corrected in my tree, thanks.
Quoted 59 lines
Quoted 10 lines
regions. However other operating systems which may modify SLB entry 0 or do more fancy things with segments might be affected.
When POWER9 slbia support is added in the next patch, this becomes a real problem because some new slbia variants don't invalidate all non-zero entries.
Signed-off-by Nicholas Piggin <[email protected]>---
LGTM
Reviewed-by Greg Kurz <[email protected]>Quoted 43 lines
target/ppc/mmu-hash64.c | 21 +++++++++++++++------ 1 file changed, 15 insertions(+), 6 deletions(-)target/ppc/mmu-hash64.c +15
-6@@ -100,20 +100,29 @@ void helper_slbia(CPUPPCState *env)PowerPCCPU *cpu =env_archcpu (env);int n; +/* +* slbia must always flush all TLB (which is equivalent to ERAT in ppc +* architecture). Matching on SLB_ESID_V is not good enough, because slbmte +* can overwrite a valid SLB without flushing its lookaside information. +* +* It would be possible to keep the TLB in synch with the SLB by flushing +* when a valid entry is overwritten by slbmte, and therefore slbia would +* not have to flush unless it evicts a valid SLB entry. However it is +* expected that slbmte is more common than slbia, and slbia is usually +* going to evict valid SLB entries, so that tradeoff is unlikely to be a +* good one. +*/ +/* XXX: Warning: slbia never invalidates the first segment */ for (n =1 ; n < cpu->hash64_opts->slb_size; n++) {ppc_slb_t *slb = &env->slb[n];if (slb->esid &SLB_ESID_V ) { slb->esid &= ~SLB_ESID_V ; -/* -* XXX: given the fact that segment size is 256 MB or 1TB, -* and we still don't have a tlb_flush_mask(env, n, mask) -* in QEMU, we just invalidate all TLBs -*/ - env->tlb_need_flush |=TLB_NEED_LOCAL_FLUSH ; } } + + env->tlb_need_flush |=TLB_NEED_LOCAL_FLUSH ; } static void __helper_slbie(CPUPPCState *env, target_ulong addr,
Signature · 4 lines
David Gibson | I'll have my music baroque, and my code david AT gibson.dropbear.id.au | minimalist, thank you. NOT _the_ _other_ | _way_ _around_! http://www.ozlabs.org/~dgibson
Re: [EXTERNAL] [PATCH 2/2] target/ppc: Fix ISA v3.0 (POWER9) slbia implementation
In reply to Benjamin HerrenschmidtQuoted 12 lines · Benjamin Herrenschmidt
Quoted 7 lines · Cédric Le Goater
Quoted 3 lines · Nicholas Piggin
Linux using the hash MMU ("
disable_radix" command line) on a POWER9 machine quickly hits translation bugs due to using v3.0 slbia features that are not implemented in TCG. Add them.I checked the ISA books and this looks OK but you are also modifying slbie.
For the same reason, I believe slbie needs to invalidate caches even if the entry isn't present.
Quoted 7 lines
The kernel will under some circumstances overwrite SLB entries without invalidating (because the translation itself isn't invalid, it's just that the SLB is full, so anything cached in the ERAT is still technically ok).
However, when those things get really invalidated, they need to be taken out, even if they no longer have a corresponding SLB entry.
Right, the slbie change is certainly correct, but it doesn't match what the commit message says this is doing. Nick, can you split that out please.
Signature · 4 lines
David Gibson | I'll have my music baroque, and my code david AT gibson.dropbear.id.au | minimalist, thank you. NOT _the_ _other_ | _way_ _around_! http://www.ozlabs.org/~dgibson
David Gibson
AppliedQuoted 13 lines · Nicholas Piggin
slbia must invalidate TLBs even if it does not remove a valid SLB entry, because slbmte can overwrite valid entries without removing their TLBs.
As the architecture says, slbia invalidates all lookaside information, not conditionally based on if it removed valid entries.
It does not seem possible for POWER8 or earlier Linux kernels to hit this bug because it never changes its kernel SLB translations, and it should always have valid entries if any accesses are made to usespace regions. However other operating systems which may modify SLB entry 0 or do more fancy things with segments might be affected.
Quoted 5 lines
When POWER9 slbia support is added in the next patch, this becomes a real problem because some new slbia variants don't invalidate all non-zero entries.
Signed-off-by Nicholas Piggin <[email protected]>
Applied to ppc-for-5.0, thanks.
Quoted 44 lines
--- target/ppc/mmu-hash64.c | 21 +++++++++++++++------ 1 file changed, 15 insertions(+), 6 deletions(-)target/ppc/mmu-hash64.c +15
-6@@ -100,20 +100,29 @@ void helper_slbia(CPUPPCState *env)PowerPCCPU *cpu =env_archcpu (env);int n; +/* +* slbia must always flush all TLB (which is equivalent to ERAT in ppc +* architecture). Matching on SLB_ESID_V is not good enough, because slbmte +* can overwrite a valid SLB without flushing its lookaside information. +* +* It would be possible to keep the TLB in synch with the SLB by flushing +* when a valid entry is overwritten by slbmte, and therefore slbia would +* not have to flush unless it evicts a valid SLB entry. However it is +* expected that slbmte is more common than slbia, and slbia is usually +* going to evict valid SLB entries, so that tradeoff is unlikely to be a +* good one. +*/ +/* XXX: Warning: slbia never invalidates the first segment */ for (n =1 ; n < cpu->hash64_opts->slb_size; n++) {ppc_slb_t *slb = &env->slb[n];if (slb->esid &SLB_ESID_V ) { slb->esid &= ~SLB_ESID_V ; -/* -* XXX: given the fact that segment size is 256 MB or 1TB, -* and we still don't have a tlb_flush_mask(env, n, mask) -* in QEMU, we just invalidate all TLBs -*/ - env->tlb_need_flush |=TLB_NEED_LOCAL_FLUSH ; } } + + env->tlb_need_flush |=TLB_NEED_LOCAL_FLUSH ; } static void __helper_slbie(CPUPPCState *env, target_ulong addr,
Signature · 4 lines
David Gibson | I'll have my music baroque, and my code david AT gibson.dropbear.id.au | minimalist, thank you. NOT _the_ _other_ | _way_ _around_! http://www.ozlabs.org/~dgibson