| Message ID | 1470933366-1364-3-git-send-email-mark.rutland@arm.com |
|---|---|
| State | Superseded |
| Headers | show |
Hi, Apologies for the delay in replying. On Fri, Aug 12, 2016 at 12:01:23PM +0200, Jiri Olsa wrote: > On Thu, Aug 11, 2016 at 05:36:06PM +0100, Mark Rutland wrote: > > The perf tools can read a cpumask file for a PMU, describing a subset of > > CPUs which that PMU covers. So far this has only been used to cater for > > uncore PMUs, which in practice happen to only have a single CPU > > described in the mask. > > > > Until recently, the perf tools only correctly handled cpumask containing > > a single CPU, and only when monitoring in system-wide mode. For example, > > prior to commit 00e727bb389359c8 ("perf stat: Balance opening and > > reading events"), a mask with more than a single CPU could cause > > perf stat to hang. When a CPU PMU covers a subset of CPUs, but lacks a > > cpumask, perf record will fail to open events (on the cores the PMU does > > not support), and gives up. > > > > For systems with heterogeneous CPUs such as ARM big.LITTLE systems, this > > presents a problem. We have a PMU for each microarchitecture (e.g. a big > > PMU and a little PMU), and would like to expose a cpumask for each (so > > as to allow perf record and other tools to do the right thing). However, > > doing so kernel-side will cause old perf binaries to not function (e.g. > > hitting the issue solved by 00e727bb389359c8), and thus commits the > > cardinal sin of breaking (existing) userspace. > > > > To address this chicken-and-egg problem, this patch adds support got a > > new file, supported_cpumask, which is largely identical to the existing > > cpumask file. A kernel can expose this file, knowing that new perf > > binaries will correctly support it, while old perf binaries will not > > look for it (and thus will not be broken). > > I might have asked before, but what's the kernel side state of this? Kernel-side, we do not currently expose a cpumask, and I do not have a current patch series to do so. I wanted to figure out if this was the right direction or whether I was going off into the weeds. Clearly that's jsut confusing, so I guess I should respin this long with the kernel-side patches? Implementation wise, it's fairly trivial to add (e.g. [1]). Thanks, Mark. [1] http://lkml.kernel.org/r/1466529109-21715-9-git-send-email-jeremy.linton@arm.com
diff --git a/tools/perf/util/pmu.c b/tools/perf/util/pmu.c index ddb0261..ded0cb2 100644 --- a/tools/perf/util/pmu.c +++ b/tools/perf/util/pmu.c @@ -445,14 +445,23 @@ static struct cpu_map *pmu_cpumask(const char *name) FILE *file; struct cpu_map *cpus; const char *sysfs = sysfs__mountpoint(); + const char *templates[] = { + "%s/bus/event_source/devices/%s/cpumask", + "%s/bus/event_source/devices/%s/supported_cpumask", + NULL + }; + const char **template; if (!sysfs) return NULL; - snprintf(path, PATH_MAX, - "%s/bus/event_source/devices/%s/cpumask", sysfs, name); + for (template = templates; *template; template++) { + snprintf(path, PATH_MAX, *template, sysfs, name); + if (stat(path, &st) == 0) + break; + } - if (stat(path, &st) < 0) + if (!*template) return NULL; file = fopen(path, "r");
The perf tools can read a cpumask file for a PMU, describing a subset of CPUs which that PMU covers. So far this has only been used to cater for uncore PMUs, which in practice happen to only have a single CPU described in the mask. Until recently, the perf tools only correctly handled cpumask containing a single CPU, and only when monitoring in system-wide mode. For example, prior to commit 00e727bb389359c8 ("perf stat: Balance opening and reading events"), a mask with more than a single CPU could cause perf stat to hang. When a CPU PMU covers a subset of CPUs, but lacks a cpumask, perf record will fail to open events (on the cores the PMU does not support), and gives up. For systems with heterogeneous CPUs such as ARM big.LITTLE systems, this presents a problem. We have a PMU for each microarchitecture (e.g. a big PMU and a little PMU), and would like to expose a cpumask for each (so as to allow perf record and other tools to do the right thing). However, doing so kernel-side will cause old perf binaries to not function (e.g. hitting the issue solved by 00e727bb389359c8), and thus commits the cardinal sin of breaking (existing) userspace. To address this chicken-and-egg problem, this patch adds support got a new file, supported_cpumask, which is largely identical to the existing cpumask file. A kernel can expose this file, knowing that new perf binaries will correctly support it, while old perf binaries will not look for it (and thus will not be broken). Signed-off-by: Mark Rutland <mark.rutland@arm.com> --- tools/perf/util/pmu.c | 15 ++++++++++++--- 1 file changed, 12 insertions(+), 3 deletions(-) -- 1.9.1