Re: What *is* the API for sched_getaffinity? Should sched_getaffinity always succeed when using cpu_set_t?

[KOSAKI Motohiro <kosaki dot motohiro at gmail dot com>]

> Hi!
> > > But I'm not 100% sure what the change did but it looks to me like this
> > > changed kernel to create cpu sys entires for all sockets on the bus
> > > rather than only for present ones.
> >
> > Please look at /sys/devices/system/cpu/possible on your machine. It may be
> > different from /sys/devices/system/cpu/cpuX.
> >
> > Example, my x86 machine show:
> >
> > $ cat /sys/devices/system/cpu/possible
> > 0-31
> >
> > $ find  /sys/devices/system/cpu -maxdepth  1 -name 'cpu[0-9]*'
> > /sys/devices/system/cpu/cpu0
> > /sys/devices/system/cpu/cpu1
> > /sys/devices/system/cpu/cpu2
> > /sys/devices/system/cpu/cpu3
> > /sys/devices/system/cpu/cpu4
> > /sys/devices/system/cpu/cpu5
> > /sys/devices/system/cpu/cpu6
> > /sys/devices/system/cpu/cpu7
>
> No it's not different on any of my machines, but that is likely because
> I have boring hardware that does not support cpu hotplugging.

too bad.


> > > It also looks like some time ago both calls were aliases and the only
> > > source of information about available processors was /proc/cpuinfo.
> > > (see git log on glibc sysdeps/unix/sysv/linux/getsysstats.c)
> >
> > If running such old system, Linux doesn't have a hotplug feature. So, it is
> > correct IMHO.
> >
> > Ho hm, do you mean you are worry about chroot and/or namespace feature? If so,
> > you are correct. If processes can't access /sys, /proc/cpuinfo might tell us
> > wrong number of cpus information. But, is this serious? When /sys is no accessible,
> > a lot of functions don't work correctly. It is not __get_nprocs_conf() specific.
> > If it is unacceptable, we need to implement sysconf syscall, instead of using /sys.
>
> Well, you need to mount at least three filesystems to make linux work in
> chroot proc, dev and sys. But I'm not sure if sysconf syscall would be
> acceptable, you know the mantra kernel people keeps repeating "It could
> be done in userspace".
>
> Perhaps we could introduce syscall to query sysfs values. It could take
> sysfs path as identificator and userspace buffer and would return the
> value, defacto shortcut for open(), read(), close() as it is done now.
> That would simplify some things, but still I'm not sure whether this
> would be acceptable.

Oh, I didn't suggest you. I just asked your point. I don't want to enforce
any changes if it doesn't your point.


> > > Anyway this interface needs better description in manual pages. I can
> > > prepare a patch once we are sure what information it returns.
> >
> > Great!
>
> Now we only need to figure out what exactly it does and make that into
> understandable wording.

btw, I have an another question. the manual says _SC_NPROCESSORS_CONF could
be used maximum parallelism performance. But I don't understand why number of
configured cpus help it.

Example, if an administrator has found a cpu problem in firmware log and he
disable one of cpus, and current status is folloing:

present cpus: 0-7
online cpus: 0-6

Why application should think the cpus are 8? I mean cpu offlining is typically
not short term. And as I already said it doesn't work correct if cpu is hot added
dynamically. I'm not sure who and when should use this.