RDPID is a new instruction that reads MSR_TSC_AUX quickly. This
should be considerably faster than reading the GDT. Add a
cpufeature for it and use it from __vdso_getcpu when available.
Signed-off-by: Andy Lutomirski <[email protected]>
---
I don't have a Cannonlake CPU (or whatever CPU I'd need for this).
Could someone who has such a beast give this a try?
Boris, could you double-check me? You're a lot more familiar with
CPUID stuff and the instruction tables than I am, and I can't fall
back to a real disassembler because the instruction is too new.
Also, it's time for someone to do UMIP. I'll see if I can convince
someone in KVM land to emulate it to make it easier to test.
Changes from v1:
- Remove rdpid() from special_instructions.h. (Was a leftover -- sorry.)
arch/x86/include/asm/cpufeatures.h | 1 +
arch/x86/include/asm/vgtod.h | 7 ++++++-
2 files changed, 7 insertions(+), 1 deletion(-)
diff --git a/arch/x86/include/asm/cpufeatures.h b/arch/x86/include/asm/cpufeatures.h
index 7bfb6b70c745..beaf2fb601ee 100644
--- a/arch/x86/include/asm/cpufeatures.h
+++ b/arch/x86/include/asm/cpufeatures.h
@@ -279,6 +279,7 @@
/* Intel-defined CPU features, CPUID level 0x00000007:0 (ecx), word 16 */
#define X86_FEATURE_PKU (16*32+ 3) /* Protection Keys for Userspace */
#define X86_FEATURE_OSPKE (16*32+ 4) /* OS Protection Keys Enable */
+#define X86_FEATURE_RDPID (16*32+ 22) /* RDPID instruction */
/*
* BUG word(s)
diff --git a/arch/x86/include/asm/vgtod.h b/arch/x86/include/asm/vgtod.h
index e728699db774..3a01996db58f 100644
--- a/arch/x86/include/asm/vgtod.h
+++ b/arch/x86/include/asm/vgtod.h
@@ -89,8 +89,13 @@ static inline unsigned int __getcpu(void)
* works on all CPUs. This is volatile so that it orders
* correctly wrt barrier() and to keep gcc from cleverly
* hoisting it out of the calling function.
+ *
+ * If RDPID is available, use it.
*/
- asm volatile ("lsl %1,%0" : "=r" (p) : "r" (__PER_CPU_SEG));
+ alternative_io ("lsl %[p],%[seg]",
+ ".byte 0xf3,0x0f,0xc7,0xf8", /* RDPID %eax/rax */
+ X86_FEATURE_RDPID,
+ [p] "=a" (p), [seg] "r" (__PER_CPU_SEG));
return p;
}
--
2.5.5
On Wed, Apr 20, 2016 at 06:16:01PM -0700, Andy Lutomirski wrote:
> Also, it's time for someone to do UMIP. I'll see if I can convince
> someone in KVM land to emulate it to make it easier to test.
That'll be fun - we can simply set that bit in CR4 and see who screams
:-P
> Changes from v1:
> - Remove rdpid() from special_instructions.h. (Was a leftover -- sorry.)
>
> arch/x86/include/asm/cpufeatures.h | 1 +
> arch/x86/include/asm/vgtod.h | 7 ++++++-
> 2 files changed, 7 insertions(+), 1 deletion(-)
>
> diff --git a/arch/x86/include/asm/cpufeatures.h b/arch/x86/include/asm/cpufeatures.h
> index 7bfb6b70c745..beaf2fb601ee 100644
> --- a/arch/x86/include/asm/cpufeatures.h
> +++ b/arch/x86/include/asm/cpufeatures.h
> @@ -279,6 +279,7 @@
> /* Intel-defined CPU features, CPUID level 0x00000007:0 (ecx), word 16 */
> #define X86_FEATURE_PKU (16*32+ 3) /* Protection Keys for Userspace */
> #define X86_FEATURE_OSPKE (16*32+ 4) /* OS Protection Keys Enable */
> +#define X86_FEATURE_RDPID (16*32+ 22) /* RDPID instruction */
>
> /*
> * BUG word(s)
> diff --git a/arch/x86/include/asm/vgtod.h b/arch/x86/include/asm/vgtod.h
> index e728699db774..3a01996db58f 100644
> --- a/arch/x86/include/asm/vgtod.h
> +++ b/arch/x86/include/asm/vgtod.h
> @@ -89,8 +89,13 @@ static inline unsigned int __getcpu(void)
> * works on all CPUs. This is volatile so that it orders
> * correctly wrt barrier() and to keep gcc from cleverly
> * hoisting it out of the calling function.
> + *
> + * If RDPID is available, use it.
> */
> - asm volatile ("lsl %1,%0" : "=r" (p) : "r" (__PER_CPU_SEG));
> + alternative_io ("lsl %[p],%[seg]",
> + ".byte 0xf3,0x0f,0xc7,0xf8", /* RDPID %eax/rax */
AFAICT, 0xf8 is correct, if I'm reading the SDM right:
bits [7:6] must be 11b for opcode group 9 and RDPID is in the 11b row,
bits [5:3] are ModRM.reg and they need to be 111b for RDPID (0x7 column)
and the last three [2:0] select the register and they must be 000b for
rAX.
HOWEVER, you need to make the asm output register constraint "=a"
because you're specifying rAX as a destination register for RDPID.
Also, I'm wondering: should we supply that alternative in a separate
inline function in special_instructions.h for wider use? I.e., something
like read_cpu_num() or so...
--
Regards/Gruss,
Boris.
ECO tip #101: Trim your mails when you reply.
On Thu, Apr 21, 2016 at 5:16 AM, Borislav Petkov <[email protected]> wrote:
> On Wed, Apr 20, 2016 at 06:16:01PM -0700, Andy Lutomirski wrote:
>> Also, it's time for someone to do UMIP. I'll see if I can convince
>> someone in KVM land to emulate it to make it easier to test.
>
> That'll be fun - we can simply set that bit in CR4 and see who screams
> :-P
>
>> Changes from v1:
>> - Remove rdpid() from special_instructions.h. (Was a leftover -- sorry.)
>>
>> arch/x86/include/asm/cpufeatures.h | 1 +
>> arch/x86/include/asm/vgtod.h | 7 ++++++-
>> 2 files changed, 7 insertions(+), 1 deletion(-)
>>
>> diff --git a/arch/x86/include/asm/cpufeatures.h b/arch/x86/include/asm/cpufeatures.h
>> index 7bfb6b70c745..beaf2fb601ee 100644
>> --- a/arch/x86/include/asm/cpufeatures.h
>> +++ b/arch/x86/include/asm/cpufeatures.h
>> @@ -279,6 +279,7 @@
>> /* Intel-defined CPU features, CPUID level 0x00000007:0 (ecx), word 16 */
>> #define X86_FEATURE_PKU (16*32+ 3) /* Protection Keys for Userspace */
>> #define X86_FEATURE_OSPKE (16*32+ 4) /* OS Protection Keys Enable */
>> +#define X86_FEATURE_RDPID (16*32+ 22) /* RDPID instruction */
>>
>> /*
>> * BUG word(s)
>> diff --git a/arch/x86/include/asm/vgtod.h b/arch/x86/include/asm/vgtod.h
>> index e728699db774..3a01996db58f 100644
>> --- a/arch/x86/include/asm/vgtod.h
>> +++ b/arch/x86/include/asm/vgtod.h
>> @@ -89,8 +89,13 @@ static inline unsigned int __getcpu(void)
>> * works on all CPUs. This is volatile so that it orders
>> * correctly wrt barrier() and to keep gcc from cleverly
>> * hoisting it out of the calling function.
>> + *
>> + * If RDPID is available, use it.
>> */
>> - asm volatile ("lsl %1,%0" : "=r" (p) : "r" (__PER_CPU_SEG));
>> + alternative_io ("lsl %[p],%[seg]",
>> + ".byte 0xf3,0x0f,0xc7,0xf8", /* RDPID %eax/rax */
>
> AFAICT, 0xf8 is correct, if I'm reading the SDM right:
>
> bits [7:6] must be 11b for opcode group 9 and RDPID is in the 11b row,
> bits [5:3] are ModRM.reg and they need to be 111b for RDPID (0x7 column)
> and the last three [2:0] select the register and they must be 000b for
> rAX.
>
> HOWEVER, you need to make the asm output register constraint "=a"
> because you're specifying rAX as a destination register for RDPID.
Didn't I?
>
> Also, I'm wondering: should we supply that alternative in a separate
> inline function in special_instructions.h for wider use? I.e., something
> like read_cpu_num() or so...
>
I thought about it, and there were two reasons:
1. I don't think we want to use __getcpu in the kernel. LSL is fairly
slow, and we'd still need to mask off the node number.
raw_smp_processor_id(), in contrast, is a single load.
2. I have no way to benchmark this thing. I'm assuming the RDPID will
be faster than LSL, but that doesn't mean it's faster than a load.
(It could be -- it will save a cache line.)
So we might actually want something that does an alternative where the
two choices are the percpu load and RDPID ; AND, but that wouldn't end
up sharing code. But I'll leave that to someone with an actual
RDPID-supporting CPU :)
--Andy
On Thu, Apr 21, 2016 at 08:25:45AM -0700, Andy Lutomirski wrote:
> Didn't I?
Bah, I cut off the line which has the "=a" and then did the commenting.
Sorry about the noise.
> I thought about it, and there were two reasons:
>
> 1. I don't think we want to use __getcpu in the kernel. LSL is fairly
> slow, and we'd still need to mask off the node number.
> raw_smp_processor_id(), in contrast, is a single load.
Right.
> 2. I have no way to benchmark this thing. I'm assuming the RDPID will
> be faster than LSL, but that doesn't mean it's faster than a load.
> (It could be -- it will save a cache line.)
But the RDPID reads an MSR. So it probably is microcode and thus slower
than a load... I guess one of the reasons for the RDPID is to avoid the
serialization cost of RDTSCP.
> So we might actually want something that does an alternative where the
> two choices are the percpu load and RDPID ; AND, but that wouldn't end
> up sharing code. But I'll leave that to someone with an actual
> RDPID-supporting CPU :)
Right.
--
Regards/Gruss,
Boris.
ECO tip #101: Trim your mails when you reply.