Groups | Search | Server Info | Keyboard shortcuts | Login | Register [http] [https] [nntp] [nntps]

Groups > linux.kernel > #1685689 > unrolled thread

Re: [PATCH 0/2] Notifications for perf sideband events

Back to article view | Back to linux.kernel

This discussion starts older than the indexed window; earlier articles aren't shown. The article labeled Started by below is the oldest one visible, not the original post.

Contents

  Re: [PATCH 0/2] Notifications for perf sideband events Jiri Olsa <jolsa@redhat.com> - 2017-07-12 12:50 +0200
    Re: [PATCH 0/2] Notifications for perf sideband events "Naveen N. Rao" <naveen.n.rao@linux.vnet.ibm.com> - 2017-07-13 16:30 +0200
      Re: [PATCH 0/2] Notifications for perf sideband events Vince Weaver <vincent.weaver@maine.edu> - 2017-07-13 17:40 +0200

#1685689 — Re: [PATCH 0/2] Notifications for perf sideband events

On Mon, Jun 19, 2017 at 08:01:06PM +0530, Naveen N. Rao wrote:
> Currently, there is no way to ask for signals to be delivered when a
> certain number of sideband events have been logged into the ring buffer.
> This is problematic if we are only interested in, say, context switch
> events. Furthermore, signals are more useful (rather than polling) for
> self-profiling. This series provides for a way to achieve this.
> 
> We ride on top of the existing support for ring buffer wakeup to
> generate signals as desired. Counting sideband events still requires
> some changes in the output path, but in normal cases, it ends up being
> just a comparison.
> 
> The test program below demonstrates how a process can profile itself for
> context switch events and how it can control notification through
> signals. The key changes include the below perf_event_attr settings as
> well as use of IOC_ENABLE:
> 	pe.signal_on_wakeup = 1;
> 	pe.count_sb_events = 1;
> 	pe.wakeup_events = 2;

Vince,
could you please check on this? thanks

Naveen,
have you run Vince's test suite on this?
  http://github.com/deater/perf_event_tests.git

jirka

> 
> To keep things simple, PERF_EVENT_IOC_REFRESH cannot be used if any of
> the new attributes are set.
> 
> RFC v2:
> https://www.mail-archive.com/linux-kernel@vger.kernel.org/msg1420363.html
> Changes:
>   - Send HUP on perf_event_exit_event() (suggested by Jiri)
>   - Disable use of IOC_REFRESH if signal_on_wakeup/count_sb_events is
>     set.
> 
> 
> - Naveen
> 
> ---
> Here is a sample program demonstrating the same:
> 
>     #define _GNU_SOURCE
> 
>     #include <stdlib.h>
>     #include <stdio.h>
>     #include <unistd.h>
>     #include <fcntl.h>
>     #include <string.h>
>     #include <signal.h>
>     #include <sys/ioctl.h>
>     #include <sys/mman.h>
>     #include <linux/perf_event.h>
>     #include <asm/unistd.h>
> 
>     static long
>     perf_event_open(struct perf_event_attr *hw_event, pid_t pid,
> 		   int cpu, int group_fd, unsigned long flags)
>     {
> 	return syscall(__NR_perf_event_open, hw_event, pid, cpu,
> 		      group_fd, flags);
>     }
> 
>     static void sigio_handler(int n, siginfo_t *info, void *uc)
>     {
> 	fprintf (stderr, "Caught %s\n", info->si_code == POLL_HUP ? "POLL_HUP" :
> 		       (info->si_code == POLL_IN ? "POLL_IN" : "other signal"));
>     }
> 
>     int main(int argc, char **argv)
>     {
> 	struct perf_event_attr pe;
> 	struct sigaction act;
> 	int fd;
> 	void *buf;
> 
> 	memset(&act, 0, sizeof(act));
> 	act.sa_sigaction = sigio_handler;
> 	act.sa_flags = SA_SIGINFO;
> 	sigaction(SIGIO, &act, 0);
> 
> 	memset(&pe, 0, sizeof(struct perf_event_attr));
> 	pe.size = sizeof(struct perf_event_attr);
> 	pe.type = PERF_TYPE_SOFTWARE;
> 	pe.config = PERF_COUNT_SW_DUMMY;
> 	pe.disabled = 1;
> 	pe.sample_period = 1;
> 	pe.context_switch = 1;
> 	pe.signal_on_wakeup = 1;
> 	pe.count_sb_events = 1;
> 	pe.wakeup_events = 2;
> 
> 	fd = perf_event_open(&pe, 0, -1, -1, 0);
> 	if (fd == -1) {
> 	    fprintf(stderr, "Error opening leader %lx\n", (unsigned long)pe.config);
> 	    exit(EXIT_FAILURE);
> 	}
> 
> 	buf = mmap(NULL, sysconf(_SC_PAGESIZE) * 2, PROT_READ|PROT_WRITE, MAP_SHARED, fd, 0);
> 	if (buf == MAP_FAILED) {
> 	    fprintf(stderr, "Can't mmap buffer\n");
> 	    return -1;
> 	}
> 
> 	if (fcntl(fd, F_SETFL, fcntl(fd, F_GETFL, 0) | O_ASYNC) == -1)
> 	    return -2;
> 
> 	if (fcntl(fd, F_SETSIG, SIGIO) == -1)
> 	    return -3;
> 
> 	if (fcntl(fd, F_SETOWN, getpid()) == -1)
> 	    return -4;
> 
> 	if (ioctl(fd, PERF_EVENT_IOC_ENABLE, 0) == -1)
> 	    return -5;
> 
> 	fprintf (stderr, "Sleep 1\n");
> 	sleep(1);
> 	fprintf (stderr, "Sleep 2\n");
> 	sleep(1);
> 	fprintf (stderr, "Sleep 3\n");
> 	sleep(1);
> 
> 	/* Disable the event counter */
> 	ioctl(fd, PERF_EVENT_IOC_DISABLE, 1);
> 
> 	close(fd);
> 
> 	return 0;
>     }
> 
> 
> A sample output:
>     $ time ./cs
>     Sleep 1
>     Caught POLL_IN
>     Sleep 2
>     Caught POLL_IN
>     Sleep 3
>     Caught POLL_IN
> 
>     real	0m3.040s
>     user	0m0.001s
>     sys	0m0.003s
> 
> 
> Naveen N. Rao (2):
>   kernel/events: Add option to notify through signals on wakeup
>   kernel/events: Add option to enable counting sideband events in
>     wakeup_events
> 
>  include/uapi/linux/perf_event.h |  4 +++-
>  kernel/events/core.c            | 20 ++++++++++++--------
>  kernel/events/ring_buffer.c     | 16 ++++++++++++++++
>  3 files changed, 31 insertions(+), 9 deletions(-)
> 
> -- 
> 2.13.1
> 

[toc] | [next] | [standalone]

#1686580

[ Adding Vince...]


On 2017/07/12 12:48PM, Jiri Olsa wrote:
> On Mon, Jun 19, 2017 at 08:01:06PM +0530, Naveen N. Rao wrote:
> > Currently, there is no way to ask for signals to be delivered when a
> > certain number of sideband events have been logged into the ring buffer.
> > This is problematic if we are only interested in, say, context switch
> > events. Furthermore, signals are more useful (rather than polling) for
> > self-profiling. This series provides for a way to achieve this.
> > 
> > We ride on top of the existing support for ring buffer wakeup to
> > generate signals as desired. Counting sideband events still requires
> > some changes in the output path, but in normal cases, it ends up being
> > just a comparison.
> > 
> > The test program below demonstrates how a process can profile itself for
> > context switch events and how it can control notification through
> > signals. The key changes include the below perf_event_attr settings as
> > well as use of IOC_ENABLE:
> > 	pe.signal_on_wakeup = 1;
> > 	pe.count_sb_events = 1;
> > 	pe.wakeup_events = 2;
> 
> Vince,
> could you please check on this? thanks
> 
> Naveen,
> have you run Vince's test suite on this?
>   http://github.com/deater/perf_event_tests.git

I just tried this and I see quite a few failures even without these 
patches.

The behavior is similar with/without these patches and all the ioctl 
tests pass, but I see some failures with the overflow tests. I'll look 
into those tests in detail tomorrow. I may be missing something.


Thanks,
Naveen

[toc] | [prev] | [next] | [standalone]

#1686630

On Thu, 13 Jul 2017, Naveen N. Rao wrote:

> > could you please check on this? thanks
> > 
> > Naveen,
> > have you run Vince's test suite on this?
> >   http://github.com/deater/perf_event_tests.git
> 
> I just tried this and I see quite a few failures even without these 
> patches.
> 
> The behavior is similar with/without these patches and all the ioctl 
> tests pass, but I see some failures with the overflow tests. I'll look 
> into those tests in detail tomorrow. I may be missing something.

the signal/overflow interface has always been a troublesome one :(
It's part of why I haven't had much to say about the whole kernel 
addresses leaking into samples mess.

I had noticed some of the related overflow tests have been failing 
recently (both in perf_event_tests and in PAPI), but I haven't had
a chance to see if it was an actual regression or just due to the way the 
tests are designed.  I'll see if I can figure out what's going on.

Vince

[toc] | [prev] | [standalone]

Back to top | Article view | linux.kernel

csiph-web