Introduce cpuid_on_cpu() and cpuid_eax_on_cpu()
PATCH 1/3 Introduce cpuid_on_cpu() and cpuid_eax_on_cpu()
They will be used by cpuid driver and powernow-k8 cpufreq driver.
With these changes powernow-k8 driver could run correctly on OpenVZ kernels with virtual cpus enabled (SCHED_VCPU).
---
arch/i386/lib/Makefile | 2 - arch/i386/lib/cpuid-on-cpu.c | 67 +++++++++++++++++++++++++++++++++++++++++ arch/x86_64/lib/Makefile | 2 - arch/x86_64/lib/cpuid-on-cpu.c | 1 include/asm-i386/processor.h | 15 +++++++++ include/asm-x86_64/msr.h | 13 +++++++ 6 files changed, 98 insertions(+), 2 deletions(-)
arch/i386/lib/Makefile +1 -1
@@ -8,4 +8,4 @@ lib-y = checksum.o delay.o usercopy.o getuser.o putuser.o memcpy.o strstr.o \ lib-$(CONFIG_X86_USE_3DNOW) += mmx.o -obj-$(CONFIG_SMP) += msr-on-cpu.o +obj-$(CONFIG_SMP) += cpuid-on-cpu.o msr-on-cpu.o
arch/x86_64/lib/Makefile +1 -1
@@ -5,7 +5,7 @@ CFLAGS_csum-partial.o := -funroll-loops obj-y := io.o iomap_copy.o -obj-$(CONFIG_SMP) += msr-on-cpu.o +obj-$(CONFIG_SMP) += cpuid-on-cpu.o msr-on-cpu.o lib-y := csum-partial.o csum-copy.o csum-wrappers.o delay.o \ usercopy.o getuser.o putuser.o \
include/asm-i386/processor.h +15 -0
@@ -643,6 +643,21 @@ static inline unsigned int cpuid_edx(unsigned int op)return edx; } +#ifdef CONFIG_SMP +void cpuid_on_cpu (unsigned int cpu,u32 op,u32 *eax,u32 *ebx,u32 *ecx,u32 *edx); +u32 cpuid_eax_on_cpu (unsigned int cpu,u32 op); +#else +static inline void cpuid_on_cpu (unsigned int cpu,u32 op,u32 *eax,u32 *ebx,u32 *ecx,u32 *edx) +{ +cpuid (op, eax, ebx, ecx, edx); +} + +static inline u32 cpuid_eax_on_cpu (unsigned int cpu,u32 op) +{ +return cpuid_eax (op); +} +#endif +/* generic versions from gas */ #define GENERIC_NOP1 ".byte 0x90\n"#define GENERIC_NOP2 ".byte 0x89,0xf6\n"
include/asm-x86_64/msr.h +13 -0
@@ -163,6 +163,9 @@ static inline unsigned int cpuid_edx(unsigned int op)#ifdef CONFIG_SMP void rdmsr_on_cpu (unsigned int cpu,u32 msr_no,u32 *l,u32 *h);void wrmsr_on_cpu (unsigned int cpu,u32 msr_no,u32 l,u32 h); + +void cpuid_on_cpu (unsigned int cpu,u32 op,u32 *eax,u32 *ebx,u32 *ecx,u32 *edx); +u32 cpuid_eax_on_cpu (unsigned int cpu,u32 op);#else /* CONFIG_SMP */ static inline void rdmsr_on_cpu (unsigned int cpu,u32 msr_no,u32 *l,u32 *h) {
@@ -172,6 +175,16 @@ static inline void wrmsr_on_cpu(unsigned int cpu, u32 msr_no, u32 l, u32 h) {wrmsr (msr_no, l, h); } + +static inline void cpuid_on_cpu (unsigned int cpu,u32 op,u32 *eax,u32 *ebx,u32 *ecx,u32 *edx) +{ +cpuid (op, eax, ebx, ecx, edx); +} + +static inline u32 cpuid_eax_on_cpu (unsigned int cpu,u32 op) +{ +return cpuid_eax (op); +}#endif /* CONFIG_SMP */ #endif
arch/i386/lib/cpuid-on-cpu.c new file +67 -0
67 lines not on this page · raw message
arch/x86_64/lib/cpuid-on-cpu.c new file +1 -0
@@ -0,0 +1 @@ +#include "../../i386/lib/cpuid-on-cpu.c"
Quoted 4 lines · Alexey Dobriyan
They will be used by cpuid driver and powernow-k8 cpufreq driver.
With these changes powernow-k8 driver could run correctly on OpenVZ kernels with virtual cpus enabled (
SCHED_VCPU).
This means openvz has multiple virtual CPU levels? One for cpuid/rdmsr and one for the rest of the kernel? Both powernow-k8 and cpuid attempt to schedule to the target CPU so they should already run there. But it is some other CPU, but when they ask your _on_cpu() functions they suddenly get a "real" CPU? Where is the difference between these levels of virtualness?
That sounds quite fragile and will likely break often. I just rejected a similar concept -- virtual nodes and "physical nodes" for similar reasons.
Also it has weird semantics. For example if you have multiple virtual CPUs mapping to a single CPU then would the powernow-k8 driver try to set the frequency multiple times on the same physical CPU? That might go wrong actually because the CPU might not be happy to be poked again while it is already in a frequency change. Also there is no locking so in theory two vcpus might try to change frequency in parallel with probably quite bad effects.
I'm sure there are other scenarios with similar problems. e.g. what happens with microcode updates etc.?
Before adding any hacks like this I think your vcpu concept needs to be discussed properly on l-k. For me it doesn't look like it is something good right now though.
-Andi
Quoted 12 lines · Andi Kleen
Quoted 4 lines · Alexey Dobriyan
They will be used by cpuid driver and powernow-k8 cpufreq driver.
With these changes powernow-k8 driver could run correctly on OpenVZ kernels with virtual cpus enabled (
SCHED_VCPU).This means openvz has multiple virtual CPU levels? One for cpuid/rdmsr and one for the rest of the kernel? Both powernow-k8 and cpuid attempt to schedule to the target CPU so they should already run there. But it is some other CPU, but when they ask your
_on_cpu()functions they suddenly get a "real" CPU? Where is the difference between these levels of virtualness?
The CPUID and MSR drivers do not schedule to the target CPU; instead, on hardware, they rely on IPI'ing the target processor if it is not the one that's currently running.
There were a lot of discussion back when about which was the better solution. Alan Cox, in particular, really preferred the interrupt solution as being less likely to cause implicit deadlock.
I do want to add that it's been on my list for some time -- in fact, I keep implementing it half-way and then having other things to do -- to add MSR and CPUID ioctls() that allow the full register file to be set and read back, in order to support architecturally broken MSR and CPUID levels.
-hpa
Quoted 3 lines
The CPUID and MSR drivers do not schedule to the target CPU; instead, on hardware,
But powernow-k8 does.
-Andi
Quoted 7 lines · Andi Kleen
Quoted 4 lines · Alexey Dobriyan
They will be used by cpuid driver and powernow-k8 cpufreq driver.
With these changes powernow-k8 driver could run correctly on OpenVZ kernels with virtual cpus enabled (
SCHED_VCPU).This means openvz has multiple virtual CPU levels?
Not sure what do you mean.
Quoted 1 line
One for cpuid/rdmsr and one for the rest of the kernel?
Now I think I do. No, only one VCPU level:
+---+ +---+ +---+ PCPU | 0 | | 1 | | 2 | +---+ +---+ +---+ | \ | `---, V \ VCPU +---+ +---+ +---+ | 0 | | 1 | | 2 | +---+ +---+ +---+
PCPU chooses VCPU, chooses task from it's runqueue, or becomes idle.
Quoted 4 lines
Both powernow-k8 and cpuid attempt to schedule
to the target CPU so they should already run there. But it is some other CPU, but when they ask your_on_cpu()functions they suddenly get a "real" CPU? Where is the difference between these levels of virtualness?
*_on_cpu functions do some work on given physical CPU. set_cpus_allowed() in openvz operates on VCPU level, so process doing set_cpus_allowed() still could be scheduled anywhere. Which horribly breaks cpufreq drivers. To unbreak them, we need a way to do frequency changing work on given PCPU whose number comes from /sys/*/cpu/cpu* hierarchy.
Quoted 2 lines
That sounds quite fragile and will likely break often. I just rejected a similar concept -- virtual nodes and "physical nodes" for similar reasons.
Care to drop a link, so I could compare them? Don't recall something like that posted on l-k.
Quoted 3 lines
Also it has weird semantics. For example if you have multiple virtual CPUs mapping to a single CPU then would the powernow-k8 driver try to set the frequency multiple times on the same physical CPU?
If core cpufreq locking is OK, why would it?
Quoted 8 lines
That might
go wrong actually because the CPU might not be happy to be poked again while it is already in a frequency change. Also there is no locking so in theory two vcpus might try to change frequency in parallel with probably quite bad effects.I'm sure there are other scenarios with similar problems. e.g. what happens with microcode updates etc.?
apply_microcode() looks small enough to convert it to IPIs, but so far nobody asked for microcode updates in openvz.
Quoted 3 lines
Before adding any hacks like this I think your vcpu concept needs to be discussed properly on l-k. For me it doesn't look like it is something good right now though.
Andi, I think it all relies on correctness of core cpufreq locking. Core cpufreq locking is not changed with these patches so number of bugs should stay the same.
current way new way
----------- -------
set_cpus_allowed(cpu);
[frequency change part #1] [IPI part #1]
<=== nasty here ===>
[frequency change part #2] [IPI part #2]set_cpus_allowed(old)
Quoted 9 lines
Quoted 5 lines
Both powernow-k8 and cpuid attempt to schedule
to the target CPU so they should already run there. But it is some other CPU, but when they ask your_on_cpu()functions they suddenly get a "real" CPU? Where is the difference between these levels of virtualness?
*_on_cpufunctions do some work on given physical CPU.set_cpus_allowed()in openvz operates on VCPU level, so process doingset_cpus_allowed()still could be scheduled anywhere.
Ok so you have multple levels.
Quoted 5 lines
Quoted 3 lines
Also it has weird semantics. For example if you have multiple virtual CPUs mapping to a single CPU then would the powernow-k8 driver try to set the frequency multiple times on the same physical CPU?
If core cpufreq locking is OK, why would it?
It won't know about multiple CPUs mapping to a single CPU.
Quoted 2 lines
apply_microcode()looks small enough to convert it to IPIs, but so far nobody asked for microcode updates in openvz.
Well if they try it they will probably have problems.
Quoted 5 lines
Quoted 3 lines
Before adding any hacks like this I think your vcpu concept needs to be discussed properly on l-k. For me it doesn't look like it is something good right now though.
Andi, I think it all relies on correctness of core cpufreq locking.
I have my doubts it will cope with you changing all reasonable expected semantics under it.
-Andi
Quoted 24 lines · Andi Kleen
Quoted 8 lines
Quoted 4 lines
Both powernow-k8 and cpuid attempt to schedule
to the target CPU so they should already run there. But it is some other CPU, but when they ask your_on_cpu()functions they suddenly get a "real" CPU? Where is the difference between these levels of virtualness?
*_on_cpufunctions do some work on given physical CPU.set_cpus_allowed()in openvz operates on VCPU level, so process doingset_cpus_allowed()still could be scheduled anywhere.Ok so you have multple levels.
Quoted 5 lines
Quoted 3 lines
Also it has weird semantics. For example if you have multiple virtual CPUs mapping to a single CPU then would the powernow-k8 driver try to set the frequency multiple times on the same physical CPU?
If core cpufreq locking is OK, why would it?
It won't know about multiple CPUs mapping to a single CPU.
Quoted 2 lines
apply_microcode()looks small enough to convert it to IPIs, but so far nobody asked for microcode updates in openvz.Well if they try it they will probably have problems.
Quoted 8 lines
Quoted 5 lines
Quoted 3 lines
Before adding any hacks like this I think your vcpu concept needs to be discussed properly on l-k. For me it doesn't look like it is something good right now though.
Andi, I think it all relies on correctness of core cpufreq locking.
I have my doubts it will cope with you changing all reasonable expected semantics under it.
Synchronization promitives work as expected. Otherwise openvz'd be buried in bugs all over the map.
Core cpufreq has per-cpu array of rw-semaphores but the index of semaphore one want to down comes from userspace not from number of CPU process is executing virtual or physical.
Probably davej could say something.