[dpdk-dev] eal: check cpu flags at init

Message ID f7td1jqmzti.fsf@redhat.com (mailing list archive)
State Not Applicable, archived
Headers

Commit Message

Aaron Conole Sept. 26, 2016, 3:43 p.m. UTC
  Flavio Leitner <fbl@sysclose.org> writes:

> An application might be linked to DPDK but not really use it,
> so move the cpu flag check to the EAL initialization instead.
>
> Signed-off-by: Flavio Leitner <fbl@sysclose.org>
> ---
>  lib/librte_eal/bsdapp/eal/eal.c             | 3 +++
>  lib/librte_eal/common/eal_common_cpuflags.c | 6 ------
>  lib/librte_eal/linuxapp/eal/eal.c           | 3 +++
>  3 files changed, 6 insertions(+), 6 deletions(-)
>
> diff --git a/lib/librte_eal/bsdapp/eal/eal.c b/lib/librte_eal/bsdapp/eal/eal.c
> index a0c8f8c..c4b22af 100644
> --- a/lib/librte_eal/bsdapp/eal/eal.c
> +++ b/lib/librte_eal/bsdapp/eal/eal.c
> @@ -496,6 +496,9 @@ rte_eal_init(int argc, char **argv)
>  	char cpuset[RTE_CPU_AFFINITY_STR_LEN];
>  	char thread_name[RTE_MAX_THREAD_NAME_LEN];
>  
> +	/* checks if the machine is adequate */
> +	rte_cpu_check_supported();
> +

I think it makes sense to return a result here;  after all, since this
is no longer a *constructor*, we can actually handle a failure case.

So maybe the following diff:

--

and the change these hunks to:

if (!rte_cpu_check_supported()) {
	return -1;
}

My only concern is whether this change would be considered ABI
breaking.  I wouldn't think so, since it doesn't seem as though an
application would want to call this explicitly (and is spelled out as
such), but I can't be sure that it isn't already included in the
standard application API, and therefore needs to go through the change
process.

My $.02

-Aaron

>  	if (!rte_atomic32_test_and_set(&run_once))
>  		return -1;
>  
> diff --git a/lib/librte_eal/common/eal_common_cpuflags.c b/lib/librte_eal/common/eal_common_cpuflags.c
> index ecb1240..b5f76f7 100644
> --- a/lib/librte_eal/common/eal_common_cpuflags.c
> +++ b/lib/librte_eal/common/eal_common_cpuflags.c
> @@ -39,14 +39,8 @@
>  /**
>   * Checks if the machine is adequate for running the binary. If it is not, the
>   * program exits with status 1.
> - * The function attribute forces this function to be called before main(). But
> - * with ICC, the check is generated by the compiler.
>   */
> -#ifndef __INTEL_COMPILER
> -void __attribute__ ((__constructor__))
> -#else
>  void
> -#endif
>  rte_cpu_check_supported(void)
>  {
>  	/* This is generated at compile-time by the build system */
> diff --git a/lib/librte_eal/linuxapp/eal/eal.c b/lib/librte_eal/linuxapp/eal/eal.c
> index d5b81a3..4e88cfc 100644
> --- a/lib/librte_eal/linuxapp/eal/eal.c
> +++ b/lib/librte_eal/linuxapp/eal/eal.c
> @@ -740,6 +740,9 @@ rte_eal_init(int argc, char **argv)
>  	char cpuset[RTE_CPU_AFFINITY_STR_LEN];
>  	char thread_name[RTE_MAX_THREAD_NAME_LEN];
>  
> +	/* checks if the machine is adequate */
> +	rte_cpu_check_supported();
> +
>  	if (!rte_atomic32_test_and_set(&run_once))
>  		return -1;
  

Comments

Flavio Leitner Sept. 27, 2016, 6:32 p.m. UTC | #1
On Mon, Sep 26, 2016 at 11:43:37AM -0400, Aaron Conole wrote:
> My only concern is whether this change would be considered ABI
> breaking.  I wouldn't think so, since it doesn't seem as though an
> application would want to call this explicitly (and is spelled out as
> such), but I can't be sure that it isn't already included in the
> standard application API, and therefore needs to go through the change
> process.

I didn't want to change the original behavior more than needed.

I think another patch would be necessary to change the whole EAL
initialization because there's a bunch of rte_panic() there which
aren't friendly with callers either.
  
Aaron Conole Sept. 29, 2016, 8:42 p.m. UTC | #2
Flavio Leitner <fbl@sysclose.org> writes:

> On Mon, Sep 26, 2016 at 11:43:37AM -0400, Aaron Conole wrote:
>> My only concern is whether this change would be considered ABI
>> breaking.  I wouldn't think so, since it doesn't seem as though an
>> application would want to call this explicitly (and is spelled out as
>> such), but I can't be sure that it isn't already included in the
>> standard application API, and therefore needs to go through the change
>> process.
>
> I didn't want to change the original behavior more than needed.
>
> I think another patch would be necessary to change the whole EAL
> initialization because there's a bunch of rte_panic() there which
> aren't friendly with callers either.

Okay makes sense.

Acked-by: Aaron Conole <aconole@redhat.com>
  
Thomas Monjalon Oct. 3, 2016, 2:13 p.m. UTC | #3
2016-09-29 16:42, Aaron Conole:
> Flavio Leitner <fbl@sysclose.org> writes:
> 
> > On Mon, Sep 26, 2016 at 11:43:37AM -0400, Aaron Conole wrote:
> >> My only concern is whether this change would be considered ABI
> >> breaking.  I wouldn't think so, since it doesn't seem as though an
> >> application would want to call this explicitly (and is spelled out as
> >> such), but I can't be sure that it isn't already included in the
> >> standard application API, and therefore needs to go through the change
> >> process.
> >
> > I didn't want to change the original behavior more than needed.
> >
> > I think another patch would be necessary to change the whole EAL
> > initialization because there's a bunch of rte_panic() there which
> > aren't friendly with callers either.

Yes please, we need to remove all those panic/exit calls.

> Okay makes sense.
> 
> Acked-by: Aaron Conole <aconole@redhat.com>

Applied, thanks
  

Patch

diff --git a/lib/librte_eal/common/eal_common_cpuflags.c b/lib/librte_eal/common/eal_common_cpuflags.c
index ecb1240..eccf5f8 100644
--- a/lib/librte_eal/common/eal_common_cpuflags.c
+++ b/lib/librte_eal/common/eal_common_cpuflags.c
@@ -38,15 +38,9 @@ 
 
 /**
  * Checks if the machine is adequate for running the binary. If it is not, the
- * program exits with status 1.
- * The function attribute forces this function to be called before main(). But
- * with ICC, the check is generated by the compiler.
+ * function returns ENOTSUP.
  */
-#ifndef __INTEL_COMPILER
-void __attribute__ ((__constructor__))
-#else
-void
-#endif
+int
 rte_cpu_check_supported(void)
 {
 	/* This is generated at compile-time by the build system */
@@ -63,14 +57,15 @@  rte_cpu_check_supported(void)
 			fprintf(stderr,
 				"ERROR: CPU feature flag lookup failed with error %d\n",
 				ret);
-			exit(1);
+			return ENOTSUP;
 		}
 		if (!ret) {
 			fprintf(stderr,
 			        "ERROR: This system does not support \"%s\".\n"
 			        "Please check that RTE_MACHINE is set correctly.\n",
 			        rte_cpu_get_flag_name(compile_time_flags[i]));
-			exit(1);
+			return ENOTSUP;
 		}
 	}
+	return 0;
 }
diff --git a/lib/librte_eal/common/include/generic/rte_cpuflags.h b/lib/librte_eal/common/include/generic/rte_cpuflags.h
index 71321f3..6e4eb5a 100644
--- a/lib/librte_eal/common/include/generic/rte_cpuflags.h
+++ b/lib/librte_eal/common/include/generic/rte_cpuflags.h
@@ -79,7 +79,7 @@  rte_cpu_get_flag_enabled(enum rte_cpu_flag_t feature);
  * that were specified at compile time. It is called automatically within the
  * EAL, so does not need to be used by applications.
  */
-void
+int
 rte_cpu_check_supported(void);
 
 #endif /* _RTE_CPUFLAGS_H_ */