Notifications for perf sideband events

kernel/events/core.c +2 -1

 	 */
-	if (event->attr.inherit || !is_sampling_event(event))
+	if (event->attr.inherit || event->attr.signal_on_wakeup ||
+			!is_sampling_event(event))
 		return -EINVAL;
 
 	atomic_add(refresh, &event->event_limit);

@@ -7362,12 +7362,15 @@ static int __perf_event_overflow(struct perf_event *event,
 	 * events
 	 */
 

kernel/events/core.c +0 -0

 	raw_spin_unlock_irq(&child_ctx->lock);
 
 	/*

kernel/events/ring_buffer.c +0 -0

index 2831480c63a2..4e7c728569a8 100644

kernel/events/ring_buffer.c +0 -0

 	handle->event->pending_wakeup = 1;
 	irq_work_queue(&handle->event->pending);
 }

Signature · 1 line

2.13.1

kernel/events/core.c static int _perf_event_refresh(struct perf_event *event, int refresh) · line 2680

Quoted 7 lines · Naveen N. Rao

kernel/events/core.c +0 -0

index 6c4e523dc1e2..812fcfc767f4 100644
@@ -2679,7 +2679,8 @@ static int _perf_event_refresh(struct perf_event *event, int refresh)
 	/*
 	 * not supported on inherited events

That comment wants updating to explain the signal crud..

Quoted 20 lines

kernel/events/core.c +8 -1

 	 */
-	if (event->attr.inherit || !is_sampling_event(event))
+	if (event->attr.inherit || event->attr.signal_on_wakeup ||
+			!is_sampling_event(event))
 		return -EINVAL;
 
 	atomic_add(refresh, &event->event_limit);

@@ -7362,12 +7362,15 @@ static int __perf_event_overflow(struct perf_event *event,
 	 * events
 	 */
 
+	if (!event->attr.signal_on_wakeup) {
+		int events = atomic_read(&event->event_limit);
+		event->pending_kill = POLL_IN;
+		if (events && atomic_dec_and_test(&event->event_limit)) {
+			ret = 1;
+			event->pending_kill = POLL_HUP;
 

kernel/events/core.c static int __perf_event_overflow(struct perf_event *event, · line 7374

Quoted 3 lines

kernel/events/core.c +2 -0

+			perf_event_disable_inatomic(event);
+		}
 	}

So even without event_limit (IOC_REFRESH) this would've generated SIGIO:POLL_IN.

kernel/events/core.c perf_event_exit_event(struct perf_event *child_event, · line 10414

Quoted 8 lines

kernel/events/core.c +1 -0

 	READ_ONCE(event->overflow_handler)(event, data, regs);

@@ -10408,6 +10411,7 @@ perf_event_exit_event(struct perf_event *child_event,
 		perf_group_detach(child_event);
 	list_del_event(child_event, child_ctx);
 	child_event->state = PERF_EVENT_STATE_EXIT; /* is_event_hup() */
+	child_event->pending_kill = POLL_HUP;

This looks like an undocumented change..

Quoted 12 lines

kernel/events/core.c +0 -0

 	raw_spin_unlock_irq(&child_ctx->lock);
 
 	/*

kernel/events/ring_buffer.c +0 -0

index 2831480c63a2..4e7c728569a8 100644
@@ -21,6 +21,9 @@ static void perf_output_wakeup(struct perf_output_handle *handle)
 {
 	atomic_set(&handle->rb->poll, POLLIN);
 

kernel/events/ring_buffer.c static void perf_output_wakeup(struct perf_output_handle *handle) · line 26

Quoted 3 lines

kernel/events/ring_buffer.c +3 -0

+	if (handle->event->attr.signal_on_wakeup)
+		handle->event->pending_kill = POLL_IN;
+

And this is the bit that generates SIGIO:POLL_IN on wakeup.

Quoted 7 lines

kernel/events/ring_buffer.c +0 -0

 	handle->event->pending_wakeup = 1;
 	irq_work_queue(&handle->event->pending);
 }

Signature · 1 line

2.13.1

kernel/events/core.c +1 -1

-	if (!event->attr.watermark) {
+	if (!event->attr.count_sb_events && !event->attr.watermark) {
 		int wakeup_events = event->attr.wakeup_events;
 
 		if (wakeup_events) {

kernel/events/ring_buffer.c +2 -0

index 4e7c728569a8..f43a6081141f 100644
@@ -197,6 +197,19 @@ __perf_output_begin(struct perf_output_handle *handle,
 	 * none of the data stores below can be lifted up by the compiler.
 	 */
 
+	if (event->attr.count_sb_events && !event->attr.watermark) {
+		int wakeup_events = event->attr.wakeup_events;

On Mon, Jun 19, 2017 at 08:01:08PM +0530, Naveen N. Rao wrote:

SNIP

Quoted 13 lines

kernel/events/core.c +1 -1

-	if (!event->attr.watermark) {
+	if (!event->attr.count_sb_events && !event->attr.watermark) {
 		int wakeup_events = event->attr.wakeup_events;
 
 		if (wakeup_events) {

kernel/events/ring_buffer.c +0 -0

index 4e7c728569a8..f43a6081141f 100644
@@ -197,6 +197,19 @@ __perf_output_begin(struct perf_output_handle *handle,
 	 * none of the data stores below can be lifted up by the compiler.
 	 */
 

kernel/events/ring_buffer.c __perf_output_begin(struct perf_output_handle *handle, · line 211

Quoted 12 lines

kernel/events/ring_buffer.c +12 -0

+	if (event->attr.count_sb_events && !event->attr.watermark) {
+		int wakeup_events = event->attr.wakeup_events;
+
+		if (wakeup_events) {
+			int events = local_inc_return(&rb->events);
+
+			if (events >= wakeup_events) {
+				local_sub(wakeup_events, &rb->events);
+				local_inc(&rb->wakeup);
+			}
+		}
+	}

hum, so there's the same wakeup code in perf_output_sample, but it's not called for non-sample (sideband) events

it'd be nice to have this wakeup only once in __perf_output_begin, but it'd change the behaviour for sideband events, which would start to follow the wakeup_events logic.. and possibly disturb Vince's tests?

kernel/events/ring_buffer.c +6 -0

index 4e7c728569a8..f43a6081141f 100644
@@ -197,6 +197,19 @@ __perf_output_begin(struct perf_output_handle *handle,
 	 * none of the data stores below can be lifted up by the compiler.
 	 */
 
+	if (event->attr.count_sb_events && !event->attr.watermark) {
+		int wakeup_events = event->attr.wakeup_events;
+
+		if (wakeup_events) {
+			int events = local_inc_return(&rb->events);
+

Quoted 19 lines · Naveen N. Rao

kernel/events/ring_buffer.c +11 -0

index 4e7c728569a8..f43a6081141f 100644
@@ -197,6 +197,19 @@ __perf_output_begin(struct perf_output_handle *handle,
 	 * none of the data stores below can be lifted up by the compiler.
 	 */
 
+	if (event->attr.count_sb_events && !event->attr.watermark) {
+		int wakeup_events = event->attr.wakeup_events;
+
+		if (wakeup_events) {
+			int events = local_inc_return(&rb->events);
+
+			if (events >= wakeup_events) {
+				local_sub(wakeup_events, &rb->events);
+				local_inc(&rb->wakeup);
+			}
+		}

kernel/events/ring_buffer.c __perf_output_begin(struct perf_output_handle *handle, · line 215

Quoted 5 lines

kernel/events/ring_buffer.c +2 -0

+	}
+
 	if (unlikely(head - local_read(&rb->wakeup) > rb->watermark))
 		local_add(rb->watermark, &rb->wakeup);
 

So this is a very performance sensitive function; not at all happy to add bits here ... :/

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.

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;

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/[email protected]/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(-)

Signature · 1 line

2.13.1

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;

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/[email protected]/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(-)

Signature · 1 line

2.13.1

Quoted 11 lines · Naveen N. Rao

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.

Quoted 7 lines

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

Quoted 5 lines

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.