eal/x86: cache queried CPU flags

Message ID 20241007110725.377550-1-bruce.richardson@intel.com (mailing list archive)
State Superseded
Delegated to: Thomas Monjalon
Headers
Series eal/x86: cache queried CPU flags |

Checks

Context Check Description
ci/checkpatch success coding style OK
ci/loongarch-compilation success Compilation OK
ci/loongarch-unit-testing success Unit Testing PASS
ci/github-robot: build success github build: passed
ci/iol-broadcom-Performance success Performance Testing PASS
ci/iol-intel-Performance success Performance Testing PASS
ci/iol-mellanox-Performance success Performance Testing PASS
ci/iol-marvell-Functional success Functional Testing PASS
ci/iol-intel-Functional success Functional Testing PASS
ci/iol-compile-amd64-testing success Testing PASS
ci/iol-unit-arm64-testing success Testing PASS
ci/iol-unit-amd64-testing success Testing PASS
ci/iol-compile-arm64-testing success Testing PASS
ci/iol-sample-apps-testing warning Testing issues

Commit Message

Bruce Richardson Oct. 7, 2024, 11:06 a.m. UTC
Rather than re-querying the HW each time a CPU flag is requested, we can
just save the return value in the flags array. This should speed up
repeated querying of CPU flags, and provides a workaround for a reported
issue where errors are seen with constant querying of the AVX-512 CPU
flag from a non-AVX VM.

Bugzilla Id: 1501

Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
---
 lib/eal/x86/rte_cpuflags.c | 20 +++++++++++++++-----
 1 file changed, 15 insertions(+), 5 deletions(-)
  

Comments

David Marchand Oct. 7, 2024, 11:23 a.m. UTC | #1
On Mon, Oct 7, 2024 at 1:07 PM Bruce Richardson
<bruce.richardson@intel.com> wrote:
>
> Rather than re-querying the HW each time a CPU flag is requested, we can
> just save the return value in the flags array. This should speed up
> repeated querying of CPU flags, and provides a workaround for a reported
> issue where errors are seen with constant querying of the AVX-512 CPU
> flag from a non-AVX VM.

This could be applied to other arches cpuflag implementation.
Copying other arches maintainers as a fyi.

>
> Bugzilla Id: 1501

Bugzilla ID*

>
> Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>

I did a quick pass, it lgtm.
  
Wathsala Wathawana Vithanage Oct. 7, 2024, 2:47 p.m. UTC | #2
> On Mon, Oct 7, 2024 at 1:07 PM Bruce Richardson
> <bruce.richardson@intel.com> wrote:
> >
> > Rather than re-querying the HW each time a CPU flag is requested, we
> > can just save the return value in the flags array. This should speed
> > up repeated querying of CPU flags, and provides a workaround for a
> > reported issue where errors are seen with constant querying of the
> > AVX-512 CPU flag from a non-AVX VM.
> 
> This could be applied to other arches cpuflag implementation.
> Copying other arches maintainers as a fyi.
> 
+1
  
Wathsala Wathawana Vithanage Oct. 7, 2024, 4:29 p.m. UTC | #3
> Rather than re-querying the HW each time a CPU flag is requested, we can just
> save the return value in the flags array. This should speed up repeated
> querying of CPU flags, and provides a workaround for a reported issue where
> errors are seen with constant querying of the AVX-512 CPU flag from a non-
> AVX VM.
> 
> Bugzilla Id: 1501
> 
> Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
> ---

Hi Bruce,
Why x86/rte_cpuflags.c is not using rte_getauxval() ?
  
Bruce Richardson Oct. 7, 2024, 4:43 p.m. UTC | #4
On Mon, Oct 07, 2024 at 04:29:31PM +0000, Wathsala Wathawana Vithanage wrote:
> > Rather than re-querying the HW each time a CPU flag is requested, we can just
> > save the return value in the flags array. This should speed up repeated
> > querying of CPU flags, and provides a workaround for a reported issue where
> > errors are seen with constant querying of the AVX-512 CPU flag from a non-
> > AVX VM.
> > 
> > Bugzilla Id: 1501
> > 
> > Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
> > ---
> 
> Hi Bruce,
> Why x86/rte_cpuflags.c is not using rte_getauxval() ?

I don't see any such DPDK function, unless I'm missing something? I do see
a linux function which includes hw capability flags. I suspect the reasons
we don't use that are:

1. Lack of awareness of it (at least on my end! :-))
2. Originally when we added flags for new instruction sets we may have
   needed to support their detection on platforms where the kernel/libc was
   too old to be aware of them.
3. It's linux only, so we'd still need these cpuid calls for other OS's
   anyway.

Therefore, I don't see any benefit in using that function over what we do
now. Is there some benefit that you see that makes you think it would be
worthwhile switching?

Thanks,
/Bruce
  
Wathsala Wathawana Vithanage Oct. 7, 2024, 5:30 p.m. UTC | #5
> > Hi Bruce,
> > Why x86/rte_cpuflags.c is not using rte_getauxval() ?
> 
> I don't see any such DPDK function, unless I'm missing something? I do see a
> linux function which includes hw capability flags. I suspect the reasons we
> don't use that are:
> 
> 1. Lack of awareness of it (at least on my end! :-)) 2. Originally when we added
> flags for new instruction sets we may have
>    needed to support their detection on platforms where the kernel/libc was
>    too old to be aware of them.
> 3. It's linux only, so we'd still need these cpuid calls for other OS's
>    anyway.
> 
> Therefore, I don't see any benefit in using that function over what we do now.
> Is there some benefit that you see that makes you think it would be
> worthwhile switching?
> 

Agree, all these are good reasons to avoid it. 
getauxval() reads these values from process stack (ELF loader places it there), so it's
a relatively fast way of finding cpu features. I'm wondering if a similar mechanism is
available in FreeBSD and Windows. 
If so, it might be beneficial to have a common API for such functions.
But as you mentioned, if the kernel isn't aware of the feature it may fail.
  
Konstantin Ananyev Oct. 11, 2024, 12:42 p.m. UTC | #6
> Rather than re-querying the HW each time a CPU flag is requested, we can
> just save the return value in the flags array. This should speed up
> repeated querying of CPU flags, and provides a workaround for a reported
> issue where errors are seen with constant querying of the AVX-512 CPU
> flag from a non-AVX VM.
> 
> Bugzilla Id: 1501
> 
> Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
> ---
>  lib/eal/x86/rte_cpuflags.c | 20 +++++++++++++++-----
>  1 file changed, 15 insertions(+), 5 deletions(-)
> 
> diff --git a/lib/eal/x86/rte_cpuflags.c b/lib/eal/x86/rte_cpuflags.c
> index 26163ab746..62e782fb4b 100644
> --- a/lib/eal/x86/rte_cpuflags.c
> +++ b/lib/eal/x86/rte_cpuflags.c
> @@ -8,6 +8,7 @@
>  #include <errno.h>
>  #include <stdint.h>
>  #include <string.h>
> +#include <stdbool.h>
> 
>  #include "rte_cpuid.h"
> 
> @@ -21,12 +22,14 @@ struct feature_entry {
>  	uint32_t bit;				/**< cpuid register bit */
>  #define CPU_FLAG_NAME_MAX_LEN 64
>  	char name[CPU_FLAG_NAME_MAX_LEN];       /**< String for printing */
> +	bool has_value;
> +	bool value;
>  };
> 
>  #define FEAT_DEF(name, leaf, subleaf, reg, bit) \
>  	[RTE_CPUFLAG_##name] = {leaf, subleaf, reg, bit, #name },
> 
> -const struct feature_entry rte_cpu_feature_table[] = {
> +struct feature_entry rte_cpu_feature_table[] = {
>  	FEAT_DEF(SSE3, 0x00000001, 0, RTE_REG_ECX,  0)
>  	FEAT_DEF(PCLMULQDQ, 0x00000001, 0, RTE_REG_ECX,  1)
>  	FEAT_DEF(DTES64, 0x00000001, 0, RTE_REG_ECX,  2)
> @@ -147,7 +150,7 @@ const struct feature_entry rte_cpu_feature_table[] = {
>  int
>  rte_cpu_get_flag_enabled(enum rte_cpu_flag_t feature)
>  {
> -	const struct feature_entry *feat;
> +	struct feature_entry *feat;
>  	cpuid_registers_t regs;
>  	unsigned int maxleaf;
> 
> @@ -156,6 +159,8 @@ rte_cpu_get_flag_enabled(enum rte_cpu_flag_t feature)
>  		return -ENOENT;
> 
>  	feat = &rte_cpu_feature_table[feature];
> +	if (feat->has_value)
> +		return feat->value;
> 
>  	if (!feat->leaf)
>  		/* This entry in the table wasn't filled out! */
> @@ -163,8 +168,10 @@ rte_cpu_get_flag_enabled(enum rte_cpu_flag_t feature)
> 
>  	maxleaf = __get_cpuid_max(feat->leaf & 0x80000000, NULL);
> 
> -	if (maxleaf < feat->leaf)
> -		return 0;
> +	if (maxleaf < feat->leaf) {
> +		feat->value = 0;
> +		goto out;
> +	}
> 
>  #ifdef RTE_TOOLCHAIN_MSVC
>  	__cpuidex(regs, feat->leaf, feat->subleaf);
> @@ -175,7 +182,10 @@ rte_cpu_get_flag_enabled(enum rte_cpu_flag_t feature)
>  #endif
> 
>  	/* check if the feature is enabled */
> -	return (regs[feat->reg] >> feat->bit) & 1;
> +	feat->value = (regs[feat->reg] >> feat->bit) & 1;
> +out:
> +	feat->has_value = true;
> +	return feat->value;

If that function can be called by 2 (or more) threads simultaneously,
then In theory  'feat->has_value = true;' can be reordered with 
' feat->value = (regs[feat->reg] >> feat->bit) & 1;'  (by cpu or complier)
and some thread(s) can get wrong feat->value.
The probability of such collision is really low, but still seems not impossible. 

>  }
> 
>  const char *
> --
> 2.43.0
  
Bruce Richardson Oct. 11, 2024, 12:48 p.m. UTC | #7
On Fri, Oct 11, 2024 at 12:42:01PM +0000, Konstantin Ananyev wrote:
> 
> 
> > Rather than re-querying the HW each time a CPU flag is requested, we can
> > just save the return value in the flags array. This should speed up
> > repeated querying of CPU flags, and provides a workaround for a reported
> > issue where errors are seen with constant querying of the AVX-512 CPU
> > flag from a non-AVX VM.
> > 
> > Bugzilla Id: 1501
> > 
> > Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
> > ---
> >  lib/eal/x86/rte_cpuflags.c | 20 +++++++++++++++-----
> >  1 file changed, 15 insertions(+), 5 deletions(-)
> > 
> > diff --git a/lib/eal/x86/rte_cpuflags.c b/lib/eal/x86/rte_cpuflags.c
> > index 26163ab746..62e782fb4b 100644
> > --- a/lib/eal/x86/rte_cpuflags.c
> > +++ b/lib/eal/x86/rte_cpuflags.c
> > @@ -8,6 +8,7 @@
> >  #include <errno.h>
> >  #include <stdint.h>
> >  #include <string.h>
> > +#include <stdbool.h>
> > 
> >  #include "rte_cpuid.h"
> > 
> > @@ -21,12 +22,14 @@ struct feature_entry {
> >  	uint32_t bit;				/**< cpuid register bit */
> >  #define CPU_FLAG_NAME_MAX_LEN 64
> >  	char name[CPU_FLAG_NAME_MAX_LEN];       /**< String for printing */
> > +	bool has_value;
> > +	bool value;
> >  };
> > 
> >  #define FEAT_DEF(name, leaf, subleaf, reg, bit) \
> >  	[RTE_CPUFLAG_##name] = {leaf, subleaf, reg, bit, #name },
> > 
> > -const struct feature_entry rte_cpu_feature_table[] = {
> > +struct feature_entry rte_cpu_feature_table[] = {
> >  	FEAT_DEF(SSE3, 0x00000001, 0, RTE_REG_ECX,  0)
> >  	FEAT_DEF(PCLMULQDQ, 0x00000001, 0, RTE_REG_ECX,  1)
> >  	FEAT_DEF(DTES64, 0x00000001, 0, RTE_REG_ECX,  2)
> > @@ -147,7 +150,7 @@ const struct feature_entry rte_cpu_feature_table[] = {
> >  int
> >  rte_cpu_get_flag_enabled(enum rte_cpu_flag_t feature)
> >  {
> > -	const struct feature_entry *feat;
> > +	struct feature_entry *feat;
> >  	cpuid_registers_t regs;
> >  	unsigned int maxleaf;
> > 
> > @@ -156,6 +159,8 @@ rte_cpu_get_flag_enabled(enum rte_cpu_flag_t feature)
> >  		return -ENOENT;
> > 
> >  	feat = &rte_cpu_feature_table[feature];
> > +	if (feat->has_value)
> > +		return feat->value;
> > 
> >  	if (!feat->leaf)
> >  		/* This entry in the table wasn't filled out! */
> > @@ -163,8 +168,10 @@ rte_cpu_get_flag_enabled(enum rte_cpu_flag_t feature)
> > 
> >  	maxleaf = __get_cpuid_max(feat->leaf & 0x80000000, NULL);
> > 
> > -	if (maxleaf < feat->leaf)
> > -		return 0;
> > +	if (maxleaf < feat->leaf) {
> > +		feat->value = 0;
> > +		goto out;
> > +	}
> > 
> >  #ifdef RTE_TOOLCHAIN_MSVC
> >  	__cpuidex(regs, feat->leaf, feat->subleaf);
> > @@ -175,7 +182,10 @@ rte_cpu_get_flag_enabled(enum rte_cpu_flag_t feature)
> >  #endif
> > 
> >  	/* check if the feature is enabled */
> > -	return (regs[feat->reg] >> feat->bit) & 1;
> > +	feat->value = (regs[feat->reg] >> feat->bit) & 1;
> > +out:
> > +	feat->has_value = true;
> > +	return feat->value;
> 
> If that function can be called by 2 (or more) threads simultaneously,
> then In theory  'feat->has_value = true;' can be reordered with 
> ' feat->value = (regs[feat->reg] >> feat->bit) & 1;'  (by cpu or complier)
> and some thread(s) can get wrong feat->value.
> The probability of such collision is really low, but still seems not impossible. 
> 

Well since this code is x86-specific the externally visible store ordering
will match the instruction store ordering. Therefore, I think a compiler
barrier is all that is necessary before feat->has_value assignment,
correct?

/Bruce
  
Konstantin Ananyev Oct. 11, 2024, 1 p.m. UTC | #8
> -----Original Message-----
> From: Bruce Richardson <bruce.richardson@intel.com>
> Sent: Friday, October 11, 2024 1:48 PM
> To: Konstantin Ananyev <konstantin.ananyev@huawei.com>
> Cc: dev@dpdk.org; david.marchand@redhat.com
> Subject: Re: [PATCH] eal/x86: cache queried CPU flags
> 
> On Fri, Oct 11, 2024 at 12:42:01PM +0000, Konstantin Ananyev wrote:
> >
> >
> > > Rather than re-querying the HW each time a CPU flag is requested, we can
> > > just save the return value in the flags array. This should speed up
> > > repeated querying of CPU flags, and provides a workaround for a reported
> > > issue where errors are seen with constant querying of the AVX-512 CPU
> > > flag from a non-AVX VM.
> > >
> > > Bugzilla Id: 1501
> > >
> > > Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
> > > ---
> > >  lib/eal/x86/rte_cpuflags.c | 20 +++++++++++++++-----
> > >  1 file changed, 15 insertions(+), 5 deletions(-)
> > >
> > > diff --git a/lib/eal/x86/rte_cpuflags.c b/lib/eal/x86/rte_cpuflags.c
> > > index 26163ab746..62e782fb4b 100644
> > > --- a/lib/eal/x86/rte_cpuflags.c
> > > +++ b/lib/eal/x86/rte_cpuflags.c
> > > @@ -8,6 +8,7 @@
> > >  #include <errno.h>
> > >  #include <stdint.h>
> > >  #include <string.h>
> > > +#include <stdbool.h>
> > >
> > >  #include "rte_cpuid.h"
> > >
> > > @@ -21,12 +22,14 @@ struct feature_entry {
> > >  	uint32_t bit;				/**< cpuid register bit */
> > >  #define CPU_FLAG_NAME_MAX_LEN 64
> > >  	char name[CPU_FLAG_NAME_MAX_LEN];       /**< String for printing */
> > > +	bool has_value;
> > > +	bool value;
> > >  };
> > >
> > >  #define FEAT_DEF(name, leaf, subleaf, reg, bit) \
> > >  	[RTE_CPUFLAG_##name] = {leaf, subleaf, reg, bit, #name },
> > >
> > > -const struct feature_entry rte_cpu_feature_table[] = {
> > > +struct feature_entry rte_cpu_feature_table[] = {
> > >  	FEAT_DEF(SSE3, 0x00000001, 0, RTE_REG_ECX,  0)
> > >  	FEAT_DEF(PCLMULQDQ, 0x00000001, 0, RTE_REG_ECX,  1)
> > >  	FEAT_DEF(DTES64, 0x00000001, 0, RTE_REG_ECX,  2)
> > > @@ -147,7 +150,7 @@ const struct feature_entry rte_cpu_feature_table[] = {
> > >  int
> > >  rte_cpu_get_flag_enabled(enum rte_cpu_flag_t feature)
> > >  {
> > > -	const struct feature_entry *feat;
> > > +	struct feature_entry *feat;
> > >  	cpuid_registers_t regs;
> > >  	unsigned int maxleaf;
> > >
> > > @@ -156,6 +159,8 @@ rte_cpu_get_flag_enabled(enum rte_cpu_flag_t feature)
> > >  		return -ENOENT;
> > >
> > >  	feat = &rte_cpu_feature_table[feature];
> > > +	if (feat->has_value)
> > > +		return feat->value;
> > >
> > >  	if (!feat->leaf)
> > >  		/* This entry in the table wasn't filled out! */
> > > @@ -163,8 +168,10 @@ rte_cpu_get_flag_enabled(enum rte_cpu_flag_t feature)
> > >
> > >  	maxleaf = __get_cpuid_max(feat->leaf & 0x80000000, NULL);
> > >
> > > -	if (maxleaf < feat->leaf)
> > > -		return 0;
> > > +	if (maxleaf < feat->leaf) {
> > > +		feat->value = 0;
> > > +		goto out;
> > > +	}
> > >
> > >  #ifdef RTE_TOOLCHAIN_MSVC
> > >  	__cpuidex(regs, feat->leaf, feat->subleaf);
> > > @@ -175,7 +182,10 @@ rte_cpu_get_flag_enabled(enum rte_cpu_flag_t feature)
> > >  #endif
> > >
> > >  	/* check if the feature is enabled */
> > > -	return (regs[feat->reg] >> feat->bit) & 1;
> > > +	feat->value = (regs[feat->reg] >> feat->bit) & 1;
> > > +out:
> > > +	feat->has_value = true;
> > > +	return feat->value;
> >
> > If that function can be called by 2 (or more) threads simultaneously,
> > then In theory  'feat->has_value = true;' can be reordered with
> > ' feat->value = (regs[feat->reg] >> feat->bit) & 1;'  (by cpu or complier)
> > and some thread(s) can get wrong feat->value.
> > The probability of such collision is really low, but still seems not impossible.
> >
> 
> Well since this code is x86-specific the externally visible store ordering
> will match the instruction store ordering. Therefore, I think a compiler
> barrier is all that is necessary before feat->has_value assignment,
> correct?

Yep, seems so.
  

Patch

diff --git a/lib/eal/x86/rte_cpuflags.c b/lib/eal/x86/rte_cpuflags.c
index 26163ab746..62e782fb4b 100644
--- a/lib/eal/x86/rte_cpuflags.c
+++ b/lib/eal/x86/rte_cpuflags.c
@@ -8,6 +8,7 @@ 
 #include <errno.h>
 #include <stdint.h>
 #include <string.h>
+#include <stdbool.h>
 
 #include "rte_cpuid.h"
 
@@ -21,12 +22,14 @@  struct feature_entry {
 	uint32_t bit;				/**< cpuid register bit */
 #define CPU_FLAG_NAME_MAX_LEN 64
 	char name[CPU_FLAG_NAME_MAX_LEN];       /**< String for printing */
+	bool has_value;
+	bool value;
 };
 
 #define FEAT_DEF(name, leaf, subleaf, reg, bit) \
 	[RTE_CPUFLAG_##name] = {leaf, subleaf, reg, bit, #name },
 
-const struct feature_entry rte_cpu_feature_table[] = {
+struct feature_entry rte_cpu_feature_table[] = {
 	FEAT_DEF(SSE3, 0x00000001, 0, RTE_REG_ECX,  0)
 	FEAT_DEF(PCLMULQDQ, 0x00000001, 0, RTE_REG_ECX,  1)
 	FEAT_DEF(DTES64, 0x00000001, 0, RTE_REG_ECX,  2)
@@ -147,7 +150,7 @@  const struct feature_entry rte_cpu_feature_table[] = {
 int
 rte_cpu_get_flag_enabled(enum rte_cpu_flag_t feature)
 {
-	const struct feature_entry *feat;
+	struct feature_entry *feat;
 	cpuid_registers_t regs;
 	unsigned int maxleaf;
 
@@ -156,6 +159,8 @@  rte_cpu_get_flag_enabled(enum rte_cpu_flag_t feature)
 		return -ENOENT;
 
 	feat = &rte_cpu_feature_table[feature];
+	if (feat->has_value)
+		return feat->value;
 
 	if (!feat->leaf)
 		/* This entry in the table wasn't filled out! */
@@ -163,8 +168,10 @@  rte_cpu_get_flag_enabled(enum rte_cpu_flag_t feature)
 
 	maxleaf = __get_cpuid_max(feat->leaf & 0x80000000, NULL);
 
-	if (maxleaf < feat->leaf)
-		return 0;
+	if (maxleaf < feat->leaf) {
+		feat->value = 0;
+		goto out;
+	}
 
 #ifdef RTE_TOOLCHAIN_MSVC
 	__cpuidex(regs, feat->leaf, feat->subleaf);
@@ -175,7 +182,10 @@  rte_cpu_get_flag_enabled(enum rte_cpu_flag_t feature)
 #endif
 
 	/* check if the feature is enabled */
-	return (regs[feat->reg] >> feat->bit) & 1;
+	feat->value = (regs[feat->reg] >> feat->bit) & 1;
+out:
+	feat->has_value = true;
+	return feat->value;
 }
 
 const char *