[regression] sched_getaffinity fails with EINVAL in 2.6.31-rc4

[regression] sched_getaffinity fails with EINVAL in 2.6.31-rc4

[Gregory Haskins]

[-- Attachment #1: Type: text/plain, Size: 485 bytes --]

I am not sure when this started, but noticed that sched_getaffinity is
not working in -rc4.

[pid  6254] sched_getaffinity(0, 128, 0x7f80586b1fe0) = -1 EINVAL
(Invalid argument)

Here is a simple program for reproduction:

#include <sched.h>

int main(void)
{
	cpu_set_t mask;
	int ret;

	ret = sched_getaffinity(0, sizeof(mask), &mask);
	if (ret < 0)
		perror("getaffinity");

	return ret;
}

If I get time, I will bisect this later today.

Regards,
-Greg


[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 267 bytes --]

Re: [regression] sched_getaffinity fails with EINVAL in 2.6.31-rc4

[Steven Rostedt]

On Mon, 27 Jul 2009, Gregory Haskins wrote:

> I am not sure when this started, but noticed that sched_getaffinity is
> not working in -rc4.
> 
> [pid  6254] sched_getaffinity(0, 128, 0x7f80586b1fe0) = -1 EINVAL
> (Invalid argument)
> 
> Here is a simple program for reproduction:

Thanks,

I have a box available in which I can look at this further.

-- Steve

> 
> #include <sched.h>
> 
> int main(void)
> {
> 	cpu_set_t mask;
> 	int ret;
> 
> 	ret = sched_getaffinity(0, sizeof(mask), &mask);
> 	if (ret < 0)
> 		perror("getaffinity");
> 
> 	return ret;
> }
> 
> If I get time, I will bisect this later today.
> 
> Regards,
> -Greg
> 
> 

Re: [regression] sched_getaffinity fails with EINVAL in 2.6.31-rc4

[David Rientjes]

On Mon, 27 Jul 2009, Gregory Haskins wrote:

> I am not sure when this started, but noticed that sched_getaffinity is
> not working in -rc4.
> 
> [pid  6254] sched_getaffinity(0, 128, 0x7f80586b1fe0) = -1 EINVAL
> (Invalid argument)
> 
> Here is a simple program for reproduction:
> 
> #include <sched.h>
> 
> int main(void)
> {
> 	cpu_set_t mask;
> 	int ret;
> 
> 	ret = sched_getaffinity(0, sizeof(mask), &mask);
> 	if (ret < 0)
> 		perror("getaffinity");
> 
> 	return ret;
> }
> 
> If I get time, I will bisect this later today.
> 

It's most likely due to the cpumask changes.  I don't know what cpu_set_t 
is, but it's probably smaller than CONFIG_NR_CPUS rounded up to a multiple 
of sizeof(long).

Re: [regression] sched_getaffinity fails with EINVAL in 2.6.31-rc4

[Rusty Russell]

On Tue, 28 Jul 2009 03:42:15 am David Rientjes wrote:
> On Mon, 27 Jul 2009, Gregory Haskins wrote:
> 
> > I am not sure when this started, but noticed that sched_getaffinity is
> > not working in -rc4.
> > 
> > [pid  6254] sched_getaffinity(0, 128, 0x7f80586b1fe0) = -1 EINVAL
> > (Invalid argument)
> > 
> > Here is a simple program for reproduction:
> > 
> > #include <sched.h>
> > 
> > int main(void)
> > {
> > 	cpu_set_t mask;
> > 	int ret;
> > 
> > 	ret = sched_getaffinity(0, sizeof(mask), &mask);
> > 	if (ret < 0)
> > 		perror("getaffinity");
> > 
> > 	return ret;
> > }
> > 
> > If I get time, I will bisect this later today.
> > 
> 
> It's most likely due to the cpumask changes.  I don't know what cpu_set_t 
> is, but it's probably smaller than CONFIG_NR_CPUS rounded up to a multiple 
> of sizeof(long).

It's simply that you finally ran this on a kernel which had
CONFIG_NR_CPUS > 1024.

I argued strongly against the set and getaffinity interfaces when they were
introduced; they can't be used in general without a loop (enlarge cpu_set,
retry).

Cheers,
Rusty.

Re: [regression] sched_getaffinity fails with EINVAL in 2.6.31-rc4

[David Rientjes]

On Wed, 29 Jul 2009, Rusty Russell wrote:

> I argued strongly against the set and getaffinity interfaces when they were
> introduced; they can't be used in general without a loop (enlarge cpu_set,
> retry).
> 

Right, so the functional regression here is glibc 2.3.2 -> 2.3.3.

Re: [regression] sched_getaffinity fails with EINVAL in 2.6.31-rc4

[Gregory Haskins]

[-- Attachment #1: Type: text/plain, Size: 1257 bytes --]

Rusty Russell wrote:
> On Tue, 28 Jul 2009 03:42:15 am David Rientjes wrote:
>> On Mon, 27 Jul 2009, Gregory Haskins wrote:
>>
>>> I am not sure when this started, but noticed that sched_getaffinity is
>>> not working in -rc4.
>>>
>>> [pid  6254] sched_getaffinity(0, 128, 0x7f80586b1fe0) = -1 EINVAL
>>> (Invalid argument)
>>>
>>> Here is a simple program for reproduction:
>>>
>>> #include <sched.h>
>>>
>>> int main(void)
>>> {
>>> 	cpu_set_t mask;
>>> 	int ret;
>>>
>>> 	ret = sched_getaffinity(0, sizeof(mask), &mask);
>>> 	if (ret < 0)
>>> 		perror("getaffinity");
>>>
>>> 	return ret;
>>> }
>>>
>>> If I get time, I will bisect this later today.
>>>
>> It's most likely due to the cpumask changes.  I don't know what cpu_set_t 
>> is, but it's probably smaller than CONFIG_NR_CPUS rounded up to a multiple 
>> of sizeof(long).
> 
> It's simply that you finally ran this on a kernel which had
> CONFIG_NR_CPUS > 1024.

Indeed.  I thought I had replied to the list that I figured that out,
but I think it might have been a trimmed list from someone elses reply.

Long story short, MAXSMP=y crept in, which set NR_CPUS=4096.  Backing
this off < 1024 indeed fixes the issue.

Thanks to all who replied!
-Greg


[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 267 bytes --]