Re: [Qemu-devel] [PATCH v2] target-arm: kvm: Differentiate registers based on write-back levels
On 22.07.2015 16:03, Peter Maydell wrote: On 22 July 2015 at 13:56, Claudio Fontana claudio.font...@huawei.com wrote: I can if you want check if this patch actually fixes the problem without the KVM workaround. Is this the version I am supposed to test, or should I wait for the next respin? Fixed version went into master earlier this week, so test that... thanks -- PMM Awesome, one week in vacation and it's done :-) Anyway, tested and works ok, no more time jumps. Thanks! CLaudio
Re: [Qemu-devel] [PATCH v2] target-arm: kvm: Differentiate registers based on write-back levels
On 17.07.2015 17:39, Christoffer Dall wrote: On Fri, Jul 17, 2015 at 03:29:56PM +0100, Peter Maydell wrote: On 16 July 2015 at 12:34, Christoffer Dall christoffer.d...@linaro.org wrote: Some registers like the CNTVCT register should only be written to the kernel as part of machine initialization or on vmload operations, but never during runtime, as this can potentially make time go backwards or create inconsistent time observations between VCPUs. Introduce a list of registers that should not be written back at runtime and check this list on syncing the register state to the KVM state. Thanks. I think this should go into QEMU 2.4, given that it fixes a bug with time misbehaving in guests. Are you happy that it's received enough testing? (I have given 32-bit KVM a spin but have no convenient 64-bit box to test with, and besides, I didn't notice the bug in the first place :-)) I tested this on both Juno and Mustang, with a simple loop kernel booting and doing hackbench test, but if we want to be on the extra careful side, perhaps Alex can run it through his migration test setup? I don't think that's necessary though. Signed-off-by: Christoffer Dall christoffer.d...@linaro.org --- Changes since RFC: - Move cpreg_level to kvm_arm_cpreg_level and into kvm32.c and kvm64.c - Changed struct name and declare as static const I have a couple of minor comments on the comments below, and you forgot to update the stub version of write_list_to_kvmstate(). I can just fix these up as I put it into target-arm.next, though. Fixed up version at: https://git.linaro.org/people/peter.maydell/qemu-arm.git target-arm.next If interested parties could test that by end-of-Monday that would be nice (since rc2 is scheduled for Tuesday). dtc | 2 +- target-arm/kvm.c | 6 +- target-arm/kvm32.c | 30 +- target-arm/kvm64.c | 30 +- target-arm/kvm_arm.h | 12 +++- target-arm/machine.c | 2 +- 6 files changed, 76 insertions(+), 6 deletions(-) diff --git a/dtc b/dtc index 65cc4d2..bc895d6 16 --- a/dtc +++ b/dtc @@ -1 +1 @@ -Subproject commit 65cc4d2748a2c2e6f27f1cf39e07a5dbabd80ebf +Subproject commit bc895d6d09695d05ceb8b52486ffe861d6cfbdde Stray submodule change :-) Damn, keeps happening to me. ok, I'll stop using git commit -a for qemu changes. diff --git a/target-arm/kvm32.c b/target-arm/kvm32.c index d7e7d68..6769815 100644 --- a/target-arm/kvm32.c +++ b/target-arm/kvm32.c @@ -153,6 +153,34 @@ bool kvm_arm_reg_syncs_via_cpreg_list(uint64_t regidx) } } +typedef struct CPRegStateLevel { +uint64_t regidx; +int level; +} CPRegStateLevel; + +/* All coprocessor registers not listed in the following table are assumed to + * be of the level KVM_PUT_RUNTIME_STATE, a register should be written less . If a register. + * often, you must add it to this table with a state of either + * KVM_PUT_RESET_STATE or KVM_PUT_FULL_STATE. + */ +static const CPRegStateLevel non_runtime_cpregs[] = { +{ KVM_REG_ARM_TIMER_CNT, KVM_PUT_FULL_STATE }, +}; diff --git a/target-arm/kvm64.c b/target-arm/kvm64.c index ac34f51..d59f41c 100644 --- a/target-arm/kvm64.c +++ b/target-arm/kvm64.c @@ -139,6 +139,34 @@ bool kvm_arm_reg_syncs_via_cpreg_list(uint64_t regidx) } } +typedef struct CPRegStateLevel { +uint64_t regidx; +int level; +} CPRegStateLevel; + +/* All system not listed in the following table are assumed to be of the level system registers + * KVM_PUT_RUNTIME_STATE, a register should be written less often, you must . If a register + * add it to this table with a state of either KVM_PUT_RESET_STATE or + * KVM_PUT_FULL_STATE. + */ +static const CPRegStateLevel non_runtime_cpregs[] = { +{ KVM_REG_ARM_TIMER_CNT, KVM_PUT_FULL_STATE }, +}; --- a/target-arm/kvm_arm.h +++ b/target-arm/kvm_arm.h @@ -83,7 +93,7 @@ bool kvm_arm_reg_syncs_via_cpreg_list(uint64_t regidx); * Note that we do not stop early on failure -- we will attempt * writing all registers in the list. */ -bool write_list_to_kvmstate(ARMCPU *cpu); +bool write_list_to_kvmstate(ARMCPU *cpu, int level); You forgot to update the stub function in target-arm/kvm-stub.c, so this breaks compilation on non-ARM hosts. whoops, who cares about non-ARM hosts anyway. Thanks for fixing these up, the fixed up version looks good. -Christoffer Hi, I can if you want check if this patch actually fixes the problem without the KVM workaround. Is this the version I am supposed to test, or should I wait for the next respin? Thanks, Claudio -- Claudio Fontana Server Virtualization Architect Huawei Technologies Duesseldorf GmbH Riesstraße 25 - 80992 München
Re: [Qemu-devel] [PATCH v2] target-arm: kvm: Differentiate registers based on write-back levels
On 22 July 2015 at 13:56, Claudio Fontana claudio.font...@huawei.com wrote: I can if you want check if this patch actually fixes the problem without the KVM workaround. Is this the version I am supposed to test, or should I wait for the next respin? Fixed version went into master earlier this week, so test that... thanks -- PMM
Re: [Qemu-devel] [PATCH v2] target-arm: kvm: Differentiate registers based on write-back levels
On 16 July 2015 at 12:34, Christoffer Dall christoffer.d...@linaro.org wrote: Some registers like the CNTVCT register should only be written to the kernel as part of machine initialization or on vmload operations, but never during runtime, as this can potentially make time go backwards or create inconsistent time observations between VCPUs. Introduce a list of registers that should not be written back at runtime and check this list on syncing the register state to the KVM state. Thanks. I think this should go into QEMU 2.4, given that it fixes a bug with time misbehaving in guests. Are you happy that it's received enough testing? (I have given 32-bit KVM a spin but have no convenient 64-bit box to test with, and besides, I didn't notice the bug in the first place :-)) Signed-off-by: Christoffer Dall christoffer.d...@linaro.org --- Changes since RFC: - Move cpreg_level to kvm_arm_cpreg_level and into kvm32.c and kvm64.c - Changed struct name and declare as static const I have a couple of minor comments on the comments below, and you forgot to update the stub version of write_list_to_kvmstate(). I can just fix these up as I put it into target-arm.next, though. Fixed up version at: https://git.linaro.org/people/peter.maydell/qemu-arm.git target-arm.next If interested parties could test that by end-of-Monday that would be nice (since rc2 is scheduled for Tuesday). dtc | 2 +- target-arm/kvm.c | 6 +- target-arm/kvm32.c | 30 +- target-arm/kvm64.c | 30 +- target-arm/kvm_arm.h | 12 +++- target-arm/machine.c | 2 +- 6 files changed, 76 insertions(+), 6 deletions(-) diff --git a/dtc b/dtc index 65cc4d2..bc895d6 16 --- a/dtc +++ b/dtc @@ -1 +1 @@ -Subproject commit 65cc4d2748a2c2e6f27f1cf39e07a5dbabd80ebf +Subproject commit bc895d6d09695d05ceb8b52486ffe861d6cfbdde Stray submodule change :-) diff --git a/target-arm/kvm32.c b/target-arm/kvm32.c index d7e7d68..6769815 100644 --- a/target-arm/kvm32.c +++ b/target-arm/kvm32.c @@ -153,6 +153,34 @@ bool kvm_arm_reg_syncs_via_cpreg_list(uint64_t regidx) } } +typedef struct CPRegStateLevel { +uint64_t regidx; +int level; +} CPRegStateLevel; + +/* All coprocessor registers not listed in the following table are assumed to + * be of the level KVM_PUT_RUNTIME_STATE, a register should be written less . If a register. + * often, you must add it to this table with a state of either + * KVM_PUT_RESET_STATE or KVM_PUT_FULL_STATE. + */ +static const CPRegStateLevel non_runtime_cpregs[] = { +{ KVM_REG_ARM_TIMER_CNT, KVM_PUT_FULL_STATE }, +}; diff --git a/target-arm/kvm64.c b/target-arm/kvm64.c index ac34f51..d59f41c 100644 --- a/target-arm/kvm64.c +++ b/target-arm/kvm64.c @@ -139,6 +139,34 @@ bool kvm_arm_reg_syncs_via_cpreg_list(uint64_t regidx) } } +typedef struct CPRegStateLevel { +uint64_t regidx; +int level; +} CPRegStateLevel; + +/* All system not listed in the following table are assumed to be of the level system registers + * KVM_PUT_RUNTIME_STATE, a register should be written less often, you must . If a register + * add it to this table with a state of either KVM_PUT_RESET_STATE or + * KVM_PUT_FULL_STATE. + */ +static const CPRegStateLevel non_runtime_cpregs[] = { +{ KVM_REG_ARM_TIMER_CNT, KVM_PUT_FULL_STATE }, +}; --- a/target-arm/kvm_arm.h +++ b/target-arm/kvm_arm.h @@ -83,7 +93,7 @@ bool kvm_arm_reg_syncs_via_cpreg_list(uint64_t regidx); * Note that we do not stop early on failure -- we will attempt * writing all registers in the list. */ -bool write_list_to_kvmstate(ARMCPU *cpu); +bool write_list_to_kvmstate(ARMCPU *cpu, int level); You forgot to update the stub function in target-arm/kvm-stub.c, so this breaks compilation on non-ARM hosts. thanks -- PMM
Re: [Qemu-devel] [PATCH v2] target-arm: kvm: Differentiate registers based on write-back levels
On Fri, Jul 17, 2015 at 03:29:56PM +0100, Peter Maydell wrote: On 16 July 2015 at 12:34, Christoffer Dall christoffer.d...@linaro.org wrote: Some registers like the CNTVCT register should only be written to the kernel as part of machine initialization or on vmload operations, but never during runtime, as this can potentially make time go backwards or create inconsistent time observations between VCPUs. Introduce a list of registers that should not be written back at runtime and check this list on syncing the register state to the KVM state. Thanks. I think this should go into QEMU 2.4, given that it fixes a bug with time misbehaving in guests. Are you happy that it's received enough testing? (I have given 32-bit KVM a spin but have no convenient 64-bit box to test with, and besides, I didn't notice the bug in the first place :-)) I tested this on both Juno and Mustang, with a simple loop kernel booting and doing hackbench test, but if we want to be on the extra careful side, perhaps Alex can run it through his migration test setup? I don't think that's necessary though. Signed-off-by: Christoffer Dall christoffer.d...@linaro.org --- Changes since RFC: - Move cpreg_level to kvm_arm_cpreg_level and into kvm32.c and kvm64.c - Changed struct name and declare as static const I have a couple of minor comments on the comments below, and you forgot to update the stub version of write_list_to_kvmstate(). I can just fix these up as I put it into target-arm.next, though. Fixed up version at: https://git.linaro.org/people/peter.maydell/qemu-arm.git target-arm.next If interested parties could test that by end-of-Monday that would be nice (since rc2 is scheduled for Tuesday). dtc | 2 +- target-arm/kvm.c | 6 +- target-arm/kvm32.c | 30 +- target-arm/kvm64.c | 30 +- target-arm/kvm_arm.h | 12 +++- target-arm/machine.c | 2 +- 6 files changed, 76 insertions(+), 6 deletions(-) diff --git a/dtc b/dtc index 65cc4d2..bc895d6 16 --- a/dtc +++ b/dtc @@ -1 +1 @@ -Subproject commit 65cc4d2748a2c2e6f27f1cf39e07a5dbabd80ebf +Subproject commit bc895d6d09695d05ceb8b52486ffe861d6cfbdde Stray submodule change :-) Damn, keeps happening to me. ok, I'll stop using git commit -a for qemu changes. diff --git a/target-arm/kvm32.c b/target-arm/kvm32.c index d7e7d68..6769815 100644 --- a/target-arm/kvm32.c +++ b/target-arm/kvm32.c @@ -153,6 +153,34 @@ bool kvm_arm_reg_syncs_via_cpreg_list(uint64_t regidx) } } +typedef struct CPRegStateLevel { +uint64_t regidx; +int level; +} CPRegStateLevel; + +/* All coprocessor registers not listed in the following table are assumed to + * be of the level KVM_PUT_RUNTIME_STATE, a register should be written less . If a register. + * often, you must add it to this table with a state of either + * KVM_PUT_RESET_STATE or KVM_PUT_FULL_STATE. + */ +static const CPRegStateLevel non_runtime_cpregs[] = { +{ KVM_REG_ARM_TIMER_CNT, KVM_PUT_FULL_STATE }, +}; diff --git a/target-arm/kvm64.c b/target-arm/kvm64.c index ac34f51..d59f41c 100644 --- a/target-arm/kvm64.c +++ b/target-arm/kvm64.c @@ -139,6 +139,34 @@ bool kvm_arm_reg_syncs_via_cpreg_list(uint64_t regidx) } } +typedef struct CPRegStateLevel { +uint64_t regidx; +int level; +} CPRegStateLevel; + +/* All system not listed in the following table are assumed to be of the level system registers + * KVM_PUT_RUNTIME_STATE, a register should be written less often, you must . If a register + * add it to this table with a state of either KVM_PUT_RESET_STATE or + * KVM_PUT_FULL_STATE. + */ +static const CPRegStateLevel non_runtime_cpregs[] = { +{ KVM_REG_ARM_TIMER_CNT, KVM_PUT_FULL_STATE }, +}; --- a/target-arm/kvm_arm.h +++ b/target-arm/kvm_arm.h @@ -83,7 +93,7 @@ bool kvm_arm_reg_syncs_via_cpreg_list(uint64_t regidx); * Note that we do not stop early on failure -- we will attempt * writing all registers in the list. */ -bool write_list_to_kvmstate(ARMCPU *cpu); +bool write_list_to_kvmstate(ARMCPU *cpu, int level); You forgot to update the stub function in target-arm/kvm-stub.c, so this breaks compilation on non-ARM hosts. whoops, who cares about non-ARM hosts anyway. Thanks for fixing these up, the fixed up version looks good. -Christoffer
[Qemu-devel] [PATCH v2] target-arm: kvm: Differentiate registers based on write-back levels
Some registers like the CNTVCT register should only be written to the kernel as part of machine initialization or on vmload operations, but never during runtime, as this can potentially make time go backwards or create inconsistent time observations between VCPUs. Introduce a list of registers that should not be written back at runtime and check this list on syncing the register state to the KVM state. Signed-off-by: Christoffer Dall christoffer.d...@linaro.org --- Changes since RFC: - Move cpreg_level to kvm_arm_cpreg_level and into kvm32.c and kvm64.c - Changed struct name and declare as static const dtc | 2 +- target-arm/kvm.c | 6 +- target-arm/kvm32.c | 30 +- target-arm/kvm64.c | 30 +- target-arm/kvm_arm.h | 12 +++- target-arm/machine.c | 2 +- 6 files changed, 76 insertions(+), 6 deletions(-) diff --git a/dtc b/dtc index 65cc4d2..bc895d6 16 --- a/dtc +++ b/dtc @@ -1 +1 @@ -Subproject commit 65cc4d2748a2c2e6f27f1cf39e07a5dbabd80ebf +Subproject commit bc895d6d09695d05ceb8b52486ffe861d6cfbdde diff --git a/target-arm/kvm.c b/target-arm/kvm.c index 548bfd7..b278542 100644 --- a/target-arm/kvm.c +++ b/target-arm/kvm.c @@ -409,7 +409,7 @@ bool write_kvmstate_to_list(ARMCPU *cpu) return ok; } -bool write_list_to_kvmstate(ARMCPU *cpu) +bool write_list_to_kvmstate(ARMCPU *cpu, int level) { CPUState *cs = CPU(cpu); int i; @@ -421,6 +421,10 @@ bool write_list_to_kvmstate(ARMCPU *cpu) uint32_t v32; int ret; +if (kvm_arm_cpreg_level(regidx) level) { +continue; +} + r.id = regidx; switch (regidx KVM_REG_SIZE_MASK) { case KVM_REG_SIZE_U32: diff --git a/target-arm/kvm32.c b/target-arm/kvm32.c index d7e7d68..6769815 100644 --- a/target-arm/kvm32.c +++ b/target-arm/kvm32.c @@ -153,6 +153,34 @@ bool kvm_arm_reg_syncs_via_cpreg_list(uint64_t regidx) } } +typedef struct CPRegStateLevel { +uint64_t regidx; +int level; +} CPRegStateLevel; + +/* All coprocessor registers not listed in the following table are assumed to + * be of the level KVM_PUT_RUNTIME_STATE, a register should be written less + * often, you must add it to this table with a state of either + * KVM_PUT_RESET_STATE or KVM_PUT_FULL_STATE. + */ +static const CPRegStateLevel non_runtime_cpregs[] = { +{ KVM_REG_ARM_TIMER_CNT, KVM_PUT_FULL_STATE }, +}; + +int kvm_arm_cpreg_level(uint64_t regidx) +{ +int i; + +for (i = 0; i ARRAY_SIZE(non_runtime_cpregs); i++) { +const CPRegStateLevel *l = non_runtime_cpregs[i]; +if (l-regidx == regidx) { +return l-level; +} +} + +return KVM_PUT_RUNTIME_STATE; +} + #define ARM_MPIDR_HWID_BITMASK 0xFF #define ARM_CPU_ID_MPIDR 0, 0, 0, 5 @@ -367,7 +395,7 @@ int kvm_arch_put_registers(CPUState *cs, int level) * managed to update the CPUARMState with, and only allowing those * to be written back up into the kernel). */ -if (!write_list_to_kvmstate(cpu)) { +if (!write_list_to_kvmstate(cpu, level)) { return EINVAL; } diff --git a/target-arm/kvm64.c b/target-arm/kvm64.c index ac34f51..d59f41c 100644 --- a/target-arm/kvm64.c +++ b/target-arm/kvm64.c @@ -139,6 +139,34 @@ bool kvm_arm_reg_syncs_via_cpreg_list(uint64_t regidx) } } +typedef struct CPRegStateLevel { +uint64_t regidx; +int level; +} CPRegStateLevel; + +/* All system not listed in the following table are assumed to be of the level + * KVM_PUT_RUNTIME_STATE, a register should be written less often, you must + * add it to this table with a state of either KVM_PUT_RESET_STATE or + * KVM_PUT_FULL_STATE. + */ +static const CPRegStateLevel non_runtime_cpregs[] = { +{ KVM_REG_ARM_TIMER_CNT, KVM_PUT_FULL_STATE }, +}; + +int kvm_arm_cpreg_level(uint64_t regidx) +{ +int i; + +for (i = 0; i ARRAY_SIZE(non_runtime_cpregs); i++) { +const CPRegStateLevel *l = non_runtime_cpregs[i]; +if (l-regidx == regidx) { +return l-level; +} +} + +return KVM_PUT_RUNTIME_STATE; +} + #define AARCH64_CORE_REG(x) (KVM_REG_ARM64 | KVM_REG_SIZE_U64 | \ KVM_REG_ARM_CORE | KVM_REG_ARM_CORE_REG(x)) @@ -280,7 +308,7 @@ int kvm_arch_put_registers(CPUState *cs, int level) return ret; } -if (!write_list_to_kvmstate(cpu)) { +if (!write_list_to_kvmstate(cpu, level)) { return EINVAL; } diff --git a/target-arm/kvm_arm.h b/target-arm/kvm_arm.h index 5abd591..7912d74 100644 --- a/target-arm/kvm_arm.h +++ b/target-arm/kvm_arm.h @@ -69,8 +69,18 @@ int kvm_arm_init_cpreg_list(ARMCPU *cpu); bool kvm_arm_reg_syncs_via_cpreg_list(uint64_t regidx); /** + * kvm_arm_cpreg_level + * regidx: KVM register index + * + * Return the level of this coprocessor/system register. Return value is + * either KVM_PUT_RUNTIME_STATE, KVM_PUT_RESET_STATE, or KVM_PUT_FULL_STATE. + */