On Fri, Apr 05, 2019 at 12:16:39PM +0200, Florian Weimer wrote: > * Peter Zijlstra: > > > On Wed, Apr 03, 2019 at 11:08:09PM +0300, Alexey Dobriyan wrote: > >> Currently there is no easy way to get the number of CPUs on the system. > > The size of the affinity mask is only related to the number of CPUs in > the system in such a way that the number of CPUs cannot be larger than > the number of bits in the affinity mask. > > >> Glibc in particular shipped with 1024 CPUs support maximum at some point > >> which is quite surprising as glibc maitainers should know better. > > This dates back to a time when the kernel was never going to support > more than 1024 CPUs. > > A lot of distribution kernels still enforce a hard limit, which papers > over firmware bugs which tell the kernel that the system can be > hot-plugged to a ridiculous number of sockets/CPUs. > > >> Another group dynamically grow buffer until cpumask fits. This is > >> inefficient as multiple system calls are done. > >> > >> Nobody seems to parse "/sys/devices/system/cpu/possible". > >> Even if someone does, parsing sysfs is much slower than necessary. > > > > True; but I suppose glibc already does lots of that anyway, right? It > > does contain the right information. > > If I recall correctly my last investigation, > /sys/devices/system/cpu/possible does not reflect the size of the > affinity mask, either. > > >> Patch overloads sched_getaffinity(len=0) to simply return "nr_cpu_ids". > >> This will make gettting CPU mask require at most 2 system calls > >> and will eliminate unnecessary code. > >> > >> len=0 is chosen so that > >> * passing zeroes is the simplest thing > >> > >> syscall(__NR_sched_getaffinity, 0, 0, NULL) > >> > >> will simply do the right thing, > >> > >> * old kernels returned -EINVAL unconditionally. > >> > >> Note: glibc segfaults upon exiting from system call because it tries to > >> clear the rest of the buffer if return value is positive, so > >> applications will have to use syscall(3). > >> Good news is that it proves noone uses sched_getaffinity(pid, 0, NULL). > > Given that old kernels fail with EINVAL, that evidence is fairly > restricted. > > I'm not sure if it's a good idea to overload this interface. I expect > that users will want to call sched_getaffinity (the system call wrapper) > with cpusetsize == 0 to query the value, so there will be pressure on > glibc to remove the memset. At that point we have an API that obscurely > fails with old glibc versions, but suceeds with newer ones, which isn't > great. I can do "if (len == 536870912)" so that bit count overflows on old kernels into EINVAL and is unlikely to be used ever.
