Skip to contentSkip to conversation

perf: support enable and disable commands in stat and record modes

PATCH Superseded Alexey Budankov 2 participants 23 messages Jun 1, 2020 – Jun 3, 2020 perf events

v6 of 12 Superseded Latest is v12 →

lore
Share & tools
lore
#0 lore raw

v6 cover perf: support enable and disable commands in stat and record modes

Changes in v6:
- split re-factoring of events handling loops for stat mode into smaller incremental parts
- added parts missing at v5
- corrected v5 runtime issues

v5: https://lore.kernel.org/lkml/[email protected]/

Changes in v5:
- split re-factoring of events handling loops for stat mode into smaller incremental parts

v4: https://lore.kernel.org/lkml/[email protected]/

Changes in v4:
- made checking of ctlfd state unconditional in record trace streaming loop
- introduced static poll fds to keep evlist__filter_pollfd() unaffected
- handled ret code of evlist__initialize_ctlfd() where need
- renamed and structured handle_events() function
- applied anonymous structs where needed

v3: https://lore.kernel.org/lkml/[email protected]/

Changes in v3:
- renamed functions and types from perf_evlist_ to evlist_ to avoid clash with libperf code;
- extended commands to be strings of variable length consisting of command name and also possibly including command specific data;
- merged docs update with the code changes;
- updated docs for -D,--delay=-1 option for stat and record modes;

v2: https://lore.kernel.org/lkml/[email protected]/

Changes in v2:
- renamed resume and pause commands to enable and disable ones, renamed CTL_CMD_RESUME and CTL_CMD_PAUSE to CTL_CMD_ENABLE and CTL_CMD_DISABLE to fit to the appropriate ioctls and avoid mixing up with PAUSE_OUTPUT ioctl;
- factored out event handling loop into a handle_events() for stat mode;
- separated -D,--delay=-1 into separate patches for stat and record modes;

v1: https://lore.kernel.org/lkml/[email protected]/

repo: tip of git://git.kernel.org/pub/scm/linux/kernel/git/acme/linux.git perf/core

The patch set implements handling of 'start disabled', 'enable' and 'disable' external control commands which can be provided for stat and record modes of the tool from an external controlling process. 'start disabled' command can be used to postpone enabling of events in the beginning of a monitoring session. 'enable' and 'disable' commands can be used to enable and disable events correspondingly any time after the start of the session.

The 'start disabled', 'enable' and 'disable' external control commands can be used to focus measurement on specially selected time intervals of workload execution. Focused measurement reduces tool intrusion and influence on workload behavior, reduces distortion and amount of collected and stored data, mitigates data accuracy loss because measurement and data capturing happen only during intervals of interest.

A controlling process can be a bash shell script [1], native executable or any other language program that can directly work with file descriptors, e.g. pipes [2], and spawn a process, specially the tool one.

-D,--delay <val> option is extended with -1 value to skip events enabling in the beginning of a monitoring session ('start disabled' command). --ctl-fd and --ctl-fd-ack command line options are introduced to provide the tool with a pair of file descriptors to listen to control commands and reply to the controlling process on the completion of received commands.

The tool reads control command message from ctl-fd descriptor, handles the command and optionally replies acknowledgement message to fd-ack descriptor, if it is specified on the command line. 'enable' command is recognized as 'enable' string message and 'disable' command is recognized as 'disable' string message both received from ctl-fd descriptor. Completion message is 'ack\n' and sent to fd-ack descriptor.

Example bash script demonstrating simple use case follows:

#!/bin/bash

ctl_dir=/tmp/

ctl_fifo=${ctl_dir}perf_ctl.fifo
test -p ${ctl_fifo} && unlink ${ctl_fifo}
mkfifo ${ctl_fifo} && exec {ctl_fd}<>${ctl_fifo}

ctl_ack_fifo=${ctl_dir}perf_ctl_ack.fifo
test -p ${ctl_ack_fifo} && unlink ${ctl_ack_fifo} mkfifo ${ctl_ack_fifo} && exec {ctl_fd_ack}<>${ctl_ack_fifo}

perf stat -D -1 -e cpu-cycles -a -I 1000                \
          --ctl-fd ${ctl_fd} --ctl-fd-ack ${ctl_fd_ack} \
          -- sleep 40 &

perf_pid=$!

sleep 5 && echo 'enable' >&${ctl_fd} && read -u ${ctl_fd_ack} e1 && echo "enabled(${e1})" sleep 10 && echo 'disable' >&${ctl_fd} && read -u ${ctl_fd_ack} d1 && echo "disabled(${d1})" sleep 5 && echo 'enable' >&${ctl_fd} && read -u ${ctl_fd_ack} e2 && echo "enabled(${e2})" sleep 10 && echo 'disable' >&${ctl_fd} && read -u ${ctl_fd_ack} d2 && echo "disabled(${d2})"

exec {ctl_fd_ack}>&- && unlink ${ctl_ack_fifo}
exec {ctl_fd}>&- && unlink ${ctl_fifo}

wait -n ${perf_pid}
exit $?

Script output:

[root@host dir] example
Events disabled
# time counts unit events

     1.001101062      <not counted>      cpu-cycles                                                  
     2.002994944      <not counted>      cpu-cycles                                                  
     3.004864340      <not counted>      cpu-cycles                                                  
     4.006727177      <not counted>      cpu-cycles                                                  
Events enabled
enabled(ack)
     4.993808464          3,124,246      cpu-cycles                                                  
     5.008597004          3,325,624      cpu-cycles                                                  
     6.010387483         83,472,992      cpu-cycles                                                  
     7.012266598         55,877,621      cpu-cycles                                                  
     8.014175695         97,892,729      cpu-cycles                                                  
     9.016056093         68,461,242      cpu-cycles                                                  
    10.017937507         55,449,643      cpu-cycles                                                  
    11.019830154         68,938,167      cpu-cycles                                                  
    12.021719952         55,164,101      cpu-cycles                                                  
    13.023627550         70,535,720      cpu-cycles                                                  
    14.025580995         53,240,125      cpu-cycles                                                  
disabled(ack)
    14.997518260         53,558,068      cpu-cycles                                                  
Events disabled
    15.027216416      <not counted>      cpu-cycles                                                  
    16.029052729      <not counted>      cpu-cycles                                                  
    17.030904762      <not counted>      cpu-cycles                                                  
    18.032073424      <not counted>      cpu-cycles                                                  
    19.033805074      <not counted>      cpu-cycles                                                  
Events enabled
enabled(ack)
    20.001279097          3,021,022      cpu-cycles                                                  
    20.035044381          6,434,367      cpu-cycles                                                  
    21.036923813         89,358,251      cpu-cycles                                                  
    22.038825169         72,516,351      cpu-cycles                                                  
#           time             counts unit events
    23.040715596         55,046,157      cpu-cycles                                                  
    24.042643757         78,128,649      cpu-cycles                                                  
    25.044558535         61,052,428      cpu-cycles                                                  
    26.046452785         62,142,806      cpu-cycles                                                  
    27.048353021         74,477,971      cpu-cycles                                                  
    28.050241286         61,001,623      cpu-cycles                                                  
    29.052149961         61,653,502      cpu-cycles                                                  
disabled(ack)
    30.004980264         82,729,640      cpu-cycles                                                  
Events disabled
    30.053516176      <not counted>      cpu-cycles                                                  
    31.055348366      <not counted>      cpu-cycles                                                  
    32.057202097      <not counted>      cpu-cycles                                                  
    33.059040702      <not counted>      cpu-cycles                                                  
    34.060843288      <not counted>      cpu-cycles                                                  
    35.000888624      <not counted>      cpu-cycles                                                  

[root@host dir]#

[1] http://man7.org/linux/man-pages/man1/bash.1.html
[2] http://man7.org/linux/man-pages/man2/pipe.2.html

---
Alexey Budankov (13):
tools/libperf: introduce notion of static polled file descriptors perf evlist: introduce control file descriptors
perf evlist: implement control command handling functions perf stat: factor out body of event handling loop for system wide perf stat: move target check to loop control statement perf stat: factor out body of event handling loop for fork case perf stat: factor out event handling loop into dispatch_events() perf stat: extend -D,--delay option with -1 value
perf stat: implement control commands handling
perf stat: introduce --ctl-fd[-ack] options
perf record: extend -D,--delay option with -1 value perf record: implement control commands handling
perf record: introduce --ctl-fd[-ack] options

 tools/lib/api/fd/array.c                 |  42 ++++++-
 tools/lib/api/fd/array.h                 |   7 ++
 tools/lib/perf/evlist.c                  |  11 ++
 tools/lib/perf/include/internal/evlist.h |   2 +
 tools/perf/Documentation/perf-record.txt |   5 +-
 tools/perf/Documentation/perf-stat.txt   |  45 ++++++-
 tools/perf/builtin-record.c              |  38 +++++-
 tools/perf/builtin-stat.c                | 149 +++++++++++++++++------
 tools/perf/builtin-trace.c               |   2 +-
 tools/perf/util/evlist.c                 | 131 ++++++++++++++++++++
 tools/perf/util/evlist.h                 |  25 ++++
 tools/perf/util/record.h                 |   4 +-
 tools/perf/util/stat.h                   |   4 +-
 13 files changed, 415 insertions(+), 50 deletions(-)
Signature · 1 line
2.24.1
#1 lore raw

PATCH v6 1/13 tools/libperf: introduce notion of static polled file descriptors

Implement adding of file descriptors by fdarray__add_stat() to fix-sized (currently 1) stat_entries array located at struct fdarray. Append added file descriptors to the array used by poll() syscall during fdarray__poll() call. Copy poll() result of the added descriptors from the array back to the storage for separate analysis.

---
 tools/lib/api/fd/array.c                 | 42 +++++++++++++++++++++++-
 tools/lib/api/fd/array.h                 |  7 ++++
 tools/lib/perf/evlist.c                  | 11 +++++++
 tools/lib/perf/include/internal/evlist.h |  2 ++
 4 files changed, 61 insertions(+), 1 deletion(-)
tools/lib/api/fd/array.c +41 -1
index 58d44d5eee31..b0027f2169c7 100644
@@ -11,10 +11,16 @@
 
 void fdarray__init(struct fdarray *fda, int nr_autogrow)
 {
+	int i;
+
 	fda->entries	 = NULL;
 	fda->priv	 = NULL;
 	fda->nr		 = fda->nr_alloc = 0;
 	fda->nr_autogrow = nr_autogrow;
+
+	fda->nr_stat = 0;
+	for (i = 0; i < FDARRAY__STAT_ENTRIES_MAX; i++)
+		fda->stat_entries[i].fd = -1;
 }
 
 int fdarray__grow(struct fdarray *fda, int nr)
@@ -83,6 +89,20 @@ int fdarray__add(struct fdarray *fda, int fd, short revents)
 	return pos;
 }
 
+int fdarray__add_stat(struct fdarray *fda, int fd, short revents)
+{
+	int pos = fda->nr_stat;
+
+	if (pos >= FDARRAY__STAT_ENTRIES_MAX)
+		return -1;
+
+	fda->stat_entries[pos].fd = fd;
+	fda->stat_entries[pos].events = revents;
+	fda->nr_stat++;
+
+	return pos;
+}
+
 int fdarray__filter(struct fdarray *fda, short revents,
 		    void (*entry_destructor)(struct fdarray *fda, int fd, void *arg),
 		    void *arg)
@@ -113,7 +133,27 @@ int fdarray__filter(struct fdarray *fda, short revents,
 
 int fdarray__poll(struct fdarray *fda, int timeout)
 {
-	return poll(fda->entries, fda->nr, timeout);
+	int nr, i, pos, res;
+
+	nr = fda->nr;
+
+	for (i = 0; i < fda->nr_stat; i++) {
+		if (fda->stat_entries[i].fd != -1) {
+			pos = fdarray__add(fda, fda->stat_entries[i].fd,
+					   fda->stat_entries[i].events);
+			if (pos >= 0)
+				fda->priv[pos].idx = i;
+		}
+	}
+
+	res = poll(fda->entries, fda->nr, timeout);
+
+	for (i = nr; i < fda->nr; i++)
+		fda->stat_entries[fda->priv[i].idx] = fda->entries[i];
+
+	fda->nr = nr;
+
+	return res;
 }
 
 int fdarray__fprintf(struct fdarray *fda, FILE *fp)
tools/lib/api/fd/array.h +7 -0
index b39557d1a88f..9bca72e80b09 100644
@@ -3,6 +3,7 @@
 #define __API_FD_ARRAY__
 
 #include <stdio.h>
+#include <poll.h>
 
 struct pollfd;
 
@@ -16,6 +17,9 @@ struct pollfd;
  *	  I.e. using 'fda->priv[N].idx = * value' where N < fda->nr is ok,
  *	  but doing 'fda->priv = malloc(M)' is not allowed.
  */
+
+#define FDARRAY__STAT_ENTRIES_MAX	1
+
 struct fdarray {
 	int	       nr;
 	int	       nr_alloc;
@@ -25,6 +29,8 @@ struct fdarray {
 		int    idx;
 		void   *ptr;
 	} *priv;
+	int	       nr_stat;
+	struct pollfd  stat_entries[FDARRAY__STAT_ENTRIES_MAX];
 };
 
 void fdarray__init(struct fdarray *fda, int nr_autogrow);
@@ -34,6 +40,7 @@ struct fdarray *fdarray__new(int nr_alloc, int nr_autogrow);
 void fdarray__delete(struct fdarray *fda);
 
 int fdarray__add(struct fdarray *fda, int fd, short revents);
+int fdarray__add_stat(struct fdarray *fda, int fd, short revents);
 int fdarray__poll(struct fdarray *fda, int timeout);
 int fdarray__filter(struct fdarray *fda, short revents,
 		    void (*entry_destructor)(struct fdarray *fda, int fd, void *arg),
tools/lib/perf/evlist.c +11 -0
index 6a875a0f01bb..e68e4c08e7c2 100644
@@ -317,6 +317,17 @@ int perf_evlist__add_pollfd(struct perf_evlist *evlist, int fd,
 	return pos;
 }
 
+int perf_evlist__add_pollfd_stat(struct perf_evlist *evlist, int fd,
+			         short revent)
+{
+	int pos = fdarray__add_stat(&evlist->pollfd, fd, revent | POLLERR | POLLHUP);
+
+	if (pos >= 0)
+		fcntl(fd, F_SETFL, O_NONBLOCK);
+
+	return pos;
+}
+
 static void perf_evlist__munmap_filtered(struct fdarray *fda, int fd,
 					 void *arg __maybe_unused)
 {
tools/lib/perf/include/internal/evlist.h +2 -0
index 74dc8c3f0b66..2b3b4518c05e 100644
@@ -46,6 +46,8 @@ struct perf_evlist_mmap_ops {
 int perf_evlist__alloc_pollfd(struct perf_evlist *evlist);
 int perf_evlist__add_pollfd(struct perf_evlist *evlist, int fd,
 			    void *ptr, short revent);
+int perf_evlist__add_pollfd_stat(struct perf_evlist *evlist, int fd,
+			         short revent);
 
 int perf_evlist__mmap_ops(struct perf_evlist *evlist,
 			  struct perf_evlist_mmap_ops *ops,
Signature · 1 line
2.24.1
reviewer
#2 lore raw

Re: [PATCH v6 01/13] tools/libperf: introduce notion of static polled file descriptors

Quoted 6 lines · Alexey Budankov

Implement adding of file descriptors by fdarray__add_stat() to fix-sized (currently 1) stat_entries array located at struct fdarray. Append added file descriptors to the array used by poll() syscall during fdarray__poll() call. Copy poll() result of the added descriptors from the array back to the storage for separate analysis.

Why not instead call evlist__add_pollfd() before other fds are added, so the fda->entries[] position is always fixed. Then this patch is not needed.

Quoted 155 lines
---
 tools/lib/api/fd/array.c                 | 42 +++++++++++++++++++++++-
 tools/lib/api/fd/array.h                 |  7 ++++
 tools/lib/perf/evlist.c                  | 11 +++++++
 tools/lib/perf/include/internal/evlist.h |  2 ++
 4 files changed, 61 insertions(+), 1 deletion(-)
tools/lib/api/fd/array.c +41 -1
index 58d44d5eee31..b0027f2169c7 100644
@@ -11,10 +11,16 @@
 
 void fdarray__init(struct fdarray *fda, int nr_autogrow)
 {
+	int i;
+
 	fda->entries	 = NULL;
 	fda->priv	 = NULL;
 	fda->nr		 = fda->nr_alloc = 0;
 	fda->nr_autogrow = nr_autogrow;
+
+	fda->nr_stat = 0;
+	for (i = 0; i < FDARRAY__STAT_ENTRIES_MAX; i++)
+		fda->stat_entries[i].fd = -1;
 }
 
 int fdarray__grow(struct fdarray *fda, int nr)
@@ -83,6 +89,20 @@ int fdarray__add(struct fdarray *fda, int fd, short revents)
 	return pos;
 }
 
+int fdarray__add_stat(struct fdarray *fda, int fd, short revents)
+{
+	int pos = fda->nr_stat;
+
+	if (pos >= FDARRAY__STAT_ENTRIES_MAX)
+		return -1;
+
+	fda->stat_entries[pos].fd = fd;
+	fda->stat_entries[pos].events = revents;
+	fda->nr_stat++;
+
+	return pos;
+}
+
 int fdarray__filter(struct fdarray *fda, short revents,
 		    void (*entry_destructor)(struct fdarray *fda, int fd, void *arg),
 		    void *arg)
@@ -113,7 +133,27 @@ int fdarray__filter(struct fdarray *fda, short revents,
 
 int fdarray__poll(struct fdarray *fda, int timeout)
 {
-	return poll(fda->entries, fda->nr, timeout);
+	int nr, i, pos, res;
+
+	nr = fda->nr;
+
+	for (i = 0; i < fda->nr_stat; i++) {
+		if (fda->stat_entries[i].fd != -1) {
+			pos = fdarray__add(fda, fda->stat_entries[i].fd,
+					   fda->stat_entries[i].events);
+			if (pos >= 0)
+				fda->priv[pos].idx = i;
+		}
+	}
+
+	res = poll(fda->entries, fda->nr, timeout);
+
+	for (i = nr; i < fda->nr; i++)
+		fda->stat_entries[fda->priv[i].idx] = fda->entries[i];
+
+	fda->nr = nr;
+
+	return res;
 }
 
 int fdarray__fprintf(struct fdarray *fda, FILE *fp)
tools/lib/api/fd/array.h +7 -0
index b39557d1a88f..9bca72e80b09 100644
@@ -3,6 +3,7 @@
 #define __API_FD_ARRAY__
 
 #include <stdio.h>
+#include <poll.h>
 
 struct pollfd;
 
@@ -16,6 +17,9 @@ struct pollfd;
  *	  I.e. using 'fda->priv[N].idx = * value' where N < fda->nr is ok,
  *	  but doing 'fda->priv = malloc(M)' is not allowed.
  */
+
+#define FDARRAY__STAT_ENTRIES_MAX	1
+
 struct fdarray {
 	int	       nr;
 	int	       nr_alloc;
@@ -25,6 +29,8 @@ struct fdarray {
 		int    idx;
 		void   *ptr;
 	} *priv;
+	int	       nr_stat;
+	struct pollfd  stat_entries[FDARRAY__STAT_ENTRIES_MAX];
 };
 
 void fdarray__init(struct fdarray *fda, int nr_autogrow);
@@ -34,6 +40,7 @@ struct fdarray *fdarray__new(int nr_alloc, int nr_autogrow);
 void fdarray__delete(struct fdarray *fda);
 
 int fdarray__add(struct fdarray *fda, int fd, short revents);
+int fdarray__add_stat(struct fdarray *fda, int fd, short revents);
 int fdarray__poll(struct fdarray *fda, int timeout);
 int fdarray__filter(struct fdarray *fda, short revents,
 		    void (*entry_destructor)(struct fdarray *fda, int fd, void *arg),
tools/lib/perf/evlist.c +11 -0
index 6a875a0f01bb..e68e4c08e7c2 100644
@@ -317,6 +317,17 @@ int perf_evlist__add_pollfd(struct perf_evlist *evlist, int fd,
 	return pos;
 }
 
+int perf_evlist__add_pollfd_stat(struct perf_evlist *evlist, int fd,
+			         short revent)
+{
+	int pos = fdarray__add_stat(&evlist->pollfd, fd, revent | POLLERR | POLLHUP);
+
+	if (pos >= 0)
+		fcntl(fd, F_SETFL, O_NONBLOCK);
+
+	return pos;
+}
+
 static void perf_evlist__munmap_filtered(struct fdarray *fda, int fd,
 					 void *arg __maybe_unused)
 {
tools/lib/perf/include/internal/evlist.h +2 -0
index 74dc8c3f0b66..2b3b4518c05e 100644
@@ -46,6 +46,8 @@ struct perf_evlist_mmap_ops {
 int perf_evlist__alloc_pollfd(struct perf_evlist *evlist);
 int perf_evlist__add_pollfd(struct perf_evlist *evlist, int fd,
 			    void *ptr, short revent);
+int perf_evlist__add_pollfd_stat(struct perf_evlist *evlist, int fd,
+			         short revent);
 
 int perf_evlist__mmap_ops(struct perf_evlist *evlist,
 			  struct perf_evlist_mmap_ops *ops,
#3 lore raw

Re: [PATCH v6 01/13] tools/libperf: introduce notion of static polled file descriptors

Hi,

Quoted 10 lines · Adrian Hunter
Quoted 6 lines · Alexey Budankov

Implement adding of file descriptors by fdarray__add_stat() to fix-sized (currently 1) stat_entries array located at struct fdarray. Append added file descriptors to the array used by poll() syscall during fdarray__poll() call. Copy poll() result of the added descriptors from the array back to the storage for separate analysis.

Why not instead call evlist__add_pollfd() before other fds are added, so the fda->entries[] position is always fixed. Then this patch is not needed.

It then will block event consumption loop, at least in record mode, due to change sin initial assumptions behind fdarray__filter(). So extension of the API with 'static' fds looks safer w.r.t. possible functional regressions at the same time extending the API with ability to atomically wait for (poll()) not only event fds but also any other fds during monitoring.

Feel free to ask more.

~Alexey

reviewer
#4 lore raw

Re: [PATCH v6 01/13] tools/libperf: introduce notion of static polled file descriptors

Quoted 11 lines · Alexey Budankov

Hi,

Quoted 8 lines · Adrian Hunter
Quoted 6 lines · Alexey Budankov

Implement adding of file descriptors by fdarray__add_stat() to fix-sized (currently 1) stat_entries array located at struct fdarray. Append added file descriptors to the array used by poll() syscall during fdarray__poll() call. Copy poll() result of the added descriptors from the array back to the storage for separate analysis.

Quoted 8 lines
Quoted 2 lines

Why not instead call evlist__add_pollfd() before other fds are added, so the fda->entries[] position is always fixed. Then this patch is not needed.

It then will block event consumption loop, at least in record mode, due to change sin initial assumptions behind fdarray__filter(). So extension of the API with 'static' fds looks safer w.r.t. possible functional regressions at the same time extending the API with ability to atomically wait for (poll()) not only event fds but also any other fds during monitoring.

So make fdarray__filter() return the number of filterable fds remaining.

#5 lore raw

RE: [PATCH v6 01/13] tools/libperf: introduce notion of static polled file descriptors

Quoted 30 lines

-----Original Message-----
From: Hunter, Adrian <[email protected]>
Sent: Wednesday, June 3, 2020 3:24 PM
To: Alexey Budankov <[email protected]>
Cc: Arnaldo Carvalho de Melo <[email protected]>; Jiri Olsa <[email protected]>; Namhyung Kim <[email protected]>; Alexander Shishkin <[email protected]>; Peter Zijlstra <[email protected]>; Ingo Molnar <[email protected]>; Andi Kleen <[email protected]>; linux-kernel <[email protected]>
Subject: Re: [PATCH v6 01/13] tools/libperf: introduce notion of static polled file descriptors

Quoted 13 lines · Alexey Budankov

Hi,

Quoted 10 lines · Adrian Hunter
Quoted 6 lines · Alexey Budankov

Implement adding of file descriptors by fdarray__add_stat() to fix-sized (currently 1) stat_entries array located at struct fdarray. Append added file descriptors to the array used by poll() syscall during fdarray__poll() call. Copy poll() result of the added descriptors from the array back to the storage for separate analysis.

Why not instead call evlist__add_pollfd() before other fds are added, so the fda->entries[] position is always fixed. Then this patch is not

needed.

Quoted 1 line
Quoted 9 lines
Quoted 5 lines

It then will block event consumption loop, at least in record mode, due to change sin initial assumptions behind fdarray__filter(). So extension of the API with 'static' fds looks safer w.r.t. possible functional regressions at the same time extending the API with ability to atomically wait for (poll()) not only event fds but also any other fds

during monitoring.

So make fdarray__filter() return the number of filterable fds remaining.

Or perhaps simpler, compare the return value to the number of fds that are known not to be filterable

#6 lore raw

Re: [PATCH v6 01/13] tools/libperf: introduce notion of static polled file descriptors

Quoted 35 lines · Adrian
Quoted 35 lines

-----Original Message-----
From: Hunter, Adrian <[email protected]>
Sent: Wednesday, June 3, 2020 3:24 PM
To: Alexey Budankov <[email protected]>
Cc: Arnaldo Carvalho de Melo <[email protected]>; Jiri Olsa <[email protected]>; Namhyung Kim <[email protected]>; Alexander Shishkin <[email protected]>; Peter Zijlstra <[email protected]>; Ingo Molnar <[email protected]>; Andi Kleen <[email protected]>; linux-kernel <[email protected]>
Subject: Re: [PATCH v6 01/13] tools/libperf: introduce notion of static polled file descriptors

Quoted 13 lines · Alexey Budankov

Hi,

Quoted 10 lines · Adrian Hunter
Quoted 6 lines · Alexey Budankov

Implement adding of file descriptors by fdarray__add_stat() to fix-sized (currently 1) stat_entries array located at struct fdarray. Append added file descriptors to the array used by poll() syscall during fdarray__poll() call. Copy poll() result of the added descriptors from the array back to the storage for separate analysis.

Why not instead call evlist__add_pollfd() before other fds are added, so the fda->entries[] position is always fixed. Then this patch is not

needed.

Quoted 6 lines

It then will block event consumption loop, at least in record mode, due to change sin initial assumptions behind fdarray__filter(). So extension of the API with 'static' fds looks safer w.r.t. possible functional regressions at the same time extending the API with ability to atomically wait for (poll()) not only event fds but also any other fds

Quoted 7 lines
Quoted 4 lines

during monitoring.

So make fdarray__filter() return the number of filterable fds remaining.

Or perhaps simpler, compare the return value to the number of fds that are known not to be filterable

Well, various implementations are for sure possible but the proposed design avoids implicit code assumptions and dependency on API calls order as in API implementation as in client code and that is why it's been proposed this way.

reviewer
#7 lore raw

Re: [PATCH v6 01/13] tools/libperf: introduce notion of static polled file descriptors

Quoted 43 lines · Alexey Budankov
Quoted 41 lines · Adrian
Quoted 39 lines

-----Original Message-----
From: Hunter, Adrian <[email protected]>
Sent: Wednesday, June 3, 2020 3:24 PM
To: Alexey Budankov <[email protected]>
Cc: Arnaldo Carvalho de Melo <[email protected]>; Jiri Olsa <[email protected]>; Namhyung Kim <[email protected]>; Alexander Shishkin <[email protected]>; Peter Zijlstra <[email protected]>; Ingo Molnar <[email protected]>; Andi Kleen <[email protected]>; linux-kernel <[email protected]>
Subject: Re: [PATCH v6 01/13] tools/libperf: introduce notion of static polled file descriptors

Quoted 13 lines · Alexey Budankov

Hi,

Quoted 10 lines · Adrian Hunter
Quoted 6 lines · Alexey Budankov

Implement adding of file descriptors by fdarray__add_stat() to fix-sized (currently 1) stat_entries array located at struct fdarray. Append added file descriptors to the array used by poll() syscall during fdarray__poll() call. Copy poll() result of the added descriptors from the array back to the storage for separate analysis.

Why not instead call evlist__add_pollfd() before other fds are added, so the fda->entries[] position is always fixed. Then this patch is not

needed.

Quoted 6 lines

It then will block event consumption loop, at least in record mode, due to change sin initial assumptions behind fdarray__filter(). So extension of the API with 'static' fds looks safer w.r.t. possible functional regressions at the same time extending the API with ability to atomically wait for (poll()) not only event fds but also any other fds

during monitoring.

So make fdarray__filter() return the number of filterable fds remaining.

Quoted 6 lines
Quoted 1 line

Or perhaps simpler, compare the return value to the number of fds that are known not to be filterable

Well, various implementations are for sure possible but the proposed design avoids implicit code assumptions and dependency on API calls order as in API implementation as in client code and that is why it's been proposed this way.

The dependencies are confined one function, namely __cmd_record(), it is simpler, and avoids the inelegant changes to fdarray__poll().

#8 lore raw

Re: [PATCH v6 01/13] tools/libperf: introduce notion of static polled file descriptors

Quoted 46 lines · Adrian Hunter
Quoted 45 lines · Alexey Budankov
Quoted 42 lines · Adrian
Quoted 30 lines

-----Original Message-----
From: Hunter, Adrian <[email protected]>
Sent: Wednesday, June 3, 2020 3:24 PM
To: Alexey Budankov <[email protected]>
Cc: Arnaldo Carvalho de Melo <[email protected]>; Jiri Olsa <[email protected]>; Namhyung Kim <[email protected]>; Alexander Shishkin <[email protected]>; Peter Zijlstra <[email protected]>; Ingo Molnar <[email protected]>; Andi Kleen <[email protected]>; linux-kernel <[email protected]>
Subject: Re: [PATCH v6 01/13] tools/libperf: introduce notion of static polled file descriptors

Quoted 13 lines · Alexey Budankov

Hi,

Quoted 10 lines · Adrian Hunter
Quoted 6 lines · Alexey Budankov

Implement adding of file descriptors by fdarray__add_stat() to fix-sized (currently 1) stat_entries array located at struct fdarray. Append added file descriptors to the array used by poll() syscall during fdarray__poll() call. Copy poll() result of the added descriptors from the array back to the storage for separate analysis.

Why not instead call evlist__add_pollfd() before other fds are added, so the fda->entries[] position is always fixed. Then this patch is not

needed.

Quoted 1 line
Quoted 9 lines
Quoted 5 lines

It then will block event consumption loop, at least in record mode, due to change sin initial assumptions behind fdarray__filter(). So extension of the API with 'static' fds looks safer w.r.t. possible functional regressions at the same time extending the API with ability to atomically wait for (poll()) not only event fds but also any other fds

during monitoring.

So make fdarray__filter() return the number of filterable fds remaining.

Or perhaps simpler, compare the return value to the number of fds that are known not to be filterable

Quoted 7 lines
Quoted 4 lines

Well, various implementations are for sure possible but the proposed design avoids implicit code assumptions and dependency on API calls order as in API implementation as in client code and that is why it's been proposed this way.

The dependencies are confined one function, namely __cmd_record(), it is simpler, and avoids the inelegant changes to fdarray__poll().

As far the changes don't cause regressions and serve their purpose well they are acceptable.

#9 lore raw

PATCH v6 2/13 perf evlist: introduce control file descriptors

In reply to Alexey Budankov

Define and initialize control file descriptors.

---
 tools/perf/util/evlist.c | 3 +++
 tools/perf/util/evlist.h | 5 +++++
 2 files changed, 8 insertions(+)
tools/perf/util/evlist.c +3 -0
index 173b4f0e0e6e..47541b5cab46 100644
@@ -63,6 +63,9 @@ void evlist__init(struct evlist *evlist, struct perf_cpu_map *cpus,
 	perf_evlist__set_maps(&evlist->core, cpus, threads);
 	evlist->workload.pid = -1;
 	evlist->bkw_mmap_state = BKW_MMAP_NOTREADY;
+	evlist->ctl_fd.fd = -1;
+	evlist->ctl_fd.ack = -1;
+	evlist->ctl_fd.pos = -1;
 }
 
 struct evlist *evlist__new(void)
tools/perf/util/evlist.h +5 -0
index b6f325dfb4d2..0d8b361f1c8e 100644
@@ -74,6 +74,11 @@ struct evlist {
 		pthread_t		th;
 		volatile int		done;
 	} thread;
+	struct {
+		int	fd;
+		int	ack;
+		int	pos;
+	} ctl_fd;
 };
 
 struct evsel_str_handler {
Signature · 1 line
2.24.1
#10 lore raw

PATCH v6 3/13 perf evlist: implement control command handling functions

In reply to Alexey Budankov

Implement functions of initialization, finalization and processing of control command messages coming from control file descriptors. Allocate control file descriptor as a static descriptor in struct fdarray of struct perf_evlist object by perf_evlist__add_pollfd_stat().

---
 tools/perf/util/evlist.c | 128 +++++++++++++++++++++++++++++++++++++++
 tools/perf/util/evlist.h |  17 ++++++
 2 files changed, 145 insertions(+)
tools/perf/util/evlist.c +128 -0
index 47541b5cab46..fbd98f741af9 100644
@@ -1718,3 +1718,131 @@ struct evsel *perf_evlist__reset_weak_group(struct evlist *evsel_list,
 	}
 	return leader;
 }
+
+int evlist__initialize_ctlfd(struct evlist *evlist, int fd, int ack)
+{
+	if (fd == -1) {
+		pr_debug("Control descriptor is not initialized\n");
+		return 0;
+	}
+
+	evlist->ctl_fd.pos = perf_evlist__add_pollfd_stat(&evlist->core, fd, POLLIN);
+	if (evlist->ctl_fd.pos < 0) {
+		evlist->ctl_fd.pos = -1;
+		pr_err("Failed to add ctl fd entry: %m\n");
+		return -1;
+	}
+
+	evlist->ctl_fd.fd = fd;
+	evlist->ctl_fd.ack = ack;
+
+	return 0;
+}
+
+int evlist__finalize_ctlfd(struct evlist *evlist)
+{
+	if (evlist->ctl_fd.pos == -1)
+		return 0;
+
+	evlist->core.pollfd.stat_entries[evlist->ctl_fd.pos].fd = -1;
+	evlist->ctl_fd.pos = -1;
+	evlist->ctl_fd.ack = -1;
+	evlist->ctl_fd.fd = -1;
+
+	return 0;
+}
+
+static int evlist__ctlfd_recv(struct evlist *evlist, enum evlist_ctl_cmd *cmd,
+			      char *cmd_data, size_t data_size)
+{
+	int err;
+	char c;
+	size_t bytes_read = 0;
+
+	memset(cmd_data, 0, data_size--);
+
+	do {
+		err = read(evlist->ctl_fd.fd, &c, 1);
+		if (err > 0) {
+			if (c == '\n' || c == '\0')
+				break;
+			cmd_data[bytes_read++] = c;
+			if (bytes_read == data_size)
+				break;
+		} else {
+			if (err == -1)
+				pr_err("Failed to read from ctlfd %d: %m\n", evlist->ctl_fd.fd);
+			break;
+		}
+	} while (1);
+
+	pr_debug("Message from ctl_fd: \"%s%s\"\n", cmd_data,
+		 bytes_read == data_size ? "" : c == '\n' ? "\\n" : "\\0");
+
+	if (err > 0) {
+		if (!strncmp(cmd_data, EVLIST_CTL_CMD_ENABLE_TAG,
+			     strlen(EVLIST_CTL_CMD_ENABLE_TAG))) {
+			*cmd = EVLIST_CTL_CMD_ENABLE;
+		} else if (!strncmp(cmd_data, EVLIST_CTL_CMD_DISABLE_TAG,
+				    strlen(EVLIST_CTL_CMD_DISABLE_TAG))) {
+			*cmd = EVLIST_CTL_CMD_DISABLE;
+		}
+	}
+
+	return err;
+}
+
+static int evlist__ctlfd_ack(struct evlist *evlist)
+{
+	int err;
+
+	if (evlist->ctl_fd.ack == -1)
+		return 0;
+
+	err = write(evlist->ctl_fd.ack, EVLIST_CTL_CMD_ACK_TAG,
+		    sizeof(EVLIST_CTL_CMD_ACK_TAG));
+	if (err == -1)
+		pr_err("failed to write to ctl_ack_fd %d: %m\n", evlist->ctl_fd.ack);
+
+	return err;
+}
+
+int evlist__ctlfd_process(struct evlist *evlist, enum evlist_ctl_cmd *cmd)
+{
+	int err = 0;
+	char cmd_data[EVLIST_CTL_CMD_MAX_LEN];
+	int ctlfd_pos = evlist->ctl_fd.pos;
+	struct pollfd *stat_entries = evlist->core.pollfd.stat_entries;
+
+	if (ctlfd_pos == -1 || !stat_entries[ctlfd_pos].revents)
+		return 0;
+
+	if (stat_entries[ctlfd_pos].revents & POLLIN) {
+		err = evlist__ctlfd_recv(evlist, cmd, cmd_data,
+					 EVLIST_CTL_CMD_MAX_LEN);
+		if (err > 0) {
+			switch (*cmd) {
+			case EVLIST_CTL_CMD_ENABLE:
+				evlist__enable(evlist);
+				break;
+			case EVLIST_CTL_CMD_DISABLE:
+				evlist__disable(evlist);
+				break;
+			case EVLIST_CTL_CMD_ACK:
+			case EVLIST_CTL_CMD_UNSUPPORTED:
+			default:
+				pr_debug("ctlfd: unsupported %d\n", *cmd);
+				break;
+			}
+			if (!(*cmd == EVLIST_CTL_CMD_ACK || *cmd == EVLIST_CTL_CMD_UNSUPPORTED))
+				evlist__ctlfd_ack(evlist);
+		}
+	}
+
+	if (stat_entries[ctlfd_pos].revents & (POLLHUP | POLLERR))
+		evlist__finalize_ctlfd(evlist);
+	else
+		stat_entries[ctlfd_pos].revents = 0;
+
+	return err;
+}
tools/perf/util/evlist.h +17 -0
index 0d8b361f1c8e..bccf0a970371 100644
@@ -360,4 +360,21 @@ void perf_evlist__force_leader(struct evlist *evlist);
 struct evsel *perf_evlist__reset_weak_group(struct evlist *evlist,
 						 struct evsel *evsel,
 						bool close);
+#define EVLIST_CTL_CMD_ENABLE_TAG  "enable"
+#define EVLIST_CTL_CMD_DISABLE_TAG "disable"
+#define EVLIST_CTL_CMD_ACK_TAG     "ack\n"
+
+#define EVLIST_CTL_CMD_MAX_LEN 64
+
+enum evlist_ctl_cmd {
+	EVLIST_CTL_CMD_UNSUPPORTED = 0,
+	EVLIST_CTL_CMD_ENABLE,
+	EVLIST_CTL_CMD_DISABLE,
+	EVLIST_CTL_CMD_ACK
+};
+
+int evlist__initialize_ctlfd(struct evlist *evlist, int ctl_fd, int ctl_fd_ack);
+int evlist__finalize_ctlfd(struct evlist *evlist);
+int evlist__ctlfd_process(struct evlist *evlist, enum evlist_ctl_cmd *cmd);
+
 #endif /* __PERF_EVLIST_H */
Signature · 1 line
2.24.1
#11 lore raw

PATCH v6 4/13 perf stat: factor out body of event handling loop for system wide

In reply to Alexey Budankov

Introduce process_timeout() and process_interval() functions that factor out body of event handling loop for attach and system wide monitoring use cases.

---
 tools/perf/builtin-stat.c | 28 ++++++++++++++++++++--------
 1 file changed, 20 insertions(+), 8 deletions(-)
tools/perf/builtin-stat.c +20 -8
index b2b79aa161dd..6dd28c714f4a 100644
@@ -422,6 +422,23 @@ static void process_interval(void)
 	print_counters(&rs, 0, NULL);
 }
 
+static bool print_interval(unsigned int interval, int *times)
+{
+	if (interval) {
+		process_interval();
+		if (interval_count && !(--(*times)))
+			return true;
+	}
+	return false;
+}
+
+static bool process_timeout(int timeout, unsigned int interval, int *times)
+{
+	if (timeout)
+		return true;
+	return print_interval(interval, times);
+}
+
 static void enable_counters(void)
 {
 	if (stat_config.initial_delay)
@@ -558,6 +575,7 @@ static int __run_perf_stat(int argc, const char **argv, int run_idx)
 	struct affinity affinity;
 	int i, cpu;
 	bool second_pass = false;
+	bool stop = false;
 
 	if (interval) {
 		ts.tv_sec  = interval / USEC_PER_MSEC;
@@ -752,17 +770,11 @@ static int __run_perf_stat(int argc, const char **argv, int run_idx)
 			psignal(WTERMSIG(status), argv[0]);
 	} else {
 		enable_counters();
-		while (!done) {
+		while (!done && !stop) {
 			nanosleep(&ts, NULL);
 			if (!is_target_alive(&target, evsel_list->core.threads))
 				break;
-			if (timeout)
-				break;
-			if (interval) {
-				process_interval();
-				if (interval_count && !(--times))
-					break;
-			}
+			stop = process_timeout(timeout, interval, &times);
 		}
 	}
 
Signature · 1 line
2.24.1
#12 lore raw

PATCH v6 5/13 perf stat: move target check to loop control statement

In reply to Alexey Budankov

Check for target existence in loop control statement jointly with 'stop' indicator based on command line values and external asynchronous 'done' signal.

---
 tools/perf/builtin-stat.c | 4 +---
 1 file changed, 1 insertion(+), 3 deletions(-)
tools/perf/builtin-stat.c +1 -3
index 6dd28c714f4a..9c090eaae1e1 100644
@@ -770,10 +770,8 @@ static int __run_perf_stat(int argc, const char **argv, int run_idx)
 			psignal(WTERMSIG(status), argv[0]);
 	} else {
 		enable_counters();
-		while (!done && !stop) {
+		while (!done && !stop && is_target_alive(&target, evsel_list->core.threads)) {
 			nanosleep(&ts, NULL);
-			if (!is_target_alive(&target, evsel_list->core.threads))
-				break;
 			stop = process_timeout(timeout, interval, &times);
 		}
 	}
Signature · 1 line
2.24.1
#13 lore raw

PATCH v6 6/13 perf stat: factor out body of event handling loop for fork case

In reply to Alexey Budankov

Factor out body of event handling loop for fork case reusing process_timeout() and process_interval() functions.

---
 tools/perf/builtin-stat.c | 8 ++------
 1 file changed, 2 insertions(+), 6 deletions(-)
tools/perf/builtin-stat.c +2 -6
index 9c090eaae1e1..edd6fdcd1c72 100644
@@ -745,13 +745,9 @@ static int __run_perf_stat(int argc, const char **argv, int run_idx)
 		enable_counters();
 
 		if (interval || timeout) {
-			while (!waitpid(child_pid, &status, WNOHANG)) {
+			while (!stop && !waitpid(child_pid, &status, WNOHANG)) {
 				nanosleep(&ts, NULL);
-				if (timeout)
-					break;
-				process_interval();
-				if (interval_count && !(--times))
-					break;
+				stop = process_timeout(timeout, interval, &times);
 			}
 		}
 		if (child_pid != -1) {
Signature · 1 line
2.24.1
#14 lore raw

PATCH v6 7/13 perf stat: factor out event handling loop into dispatch_events()

In reply to Alexey Budankov

Consolidate event dispatching loops for fork, attach and system wide monitoring use cases into common dispatch_events() function.

---
 tools/perf/builtin-stat.c | 35 ++++++++++++++++++++++++-----------
 1 file changed, 24 insertions(+), 11 deletions(-)
tools/perf/builtin-stat.c +24 -11
index edd6fdcd1c72..f2d1b8c1481d 100644
@@ -504,6 +504,27 @@ static bool is_target_alive(struct target *_target,
 	return false;
 }
 
+static int dispatch_events(bool forks, int timeout, int interval, int *times, struct timespec *ts)
+{
+	bool stop = false;
+	int child = 0, status = 0;
+
+	while (1) {
+		if (forks)
+			child = waitpid(child_pid, &status, WNOHANG);
+		else
+			child = !is_target_alive(&target, evsel_list->core.threads) ? 1 : 0;
+
+		if (done || stop || child)
+			break;
+
+		nanosleep(ts, NULL);
+		stop = process_timeout(timeout, interval, times);
+	}
+
+	return status;
+}
+
 enum counter_recovery {
 	COUNTER_SKIP,
 	COUNTER_RETRY,
@@ -575,7 +596,6 @@ static int __run_perf_stat(int argc, const char **argv, int run_idx)
 	struct affinity affinity;
 	int i, cpu;
 	bool second_pass = false;
-	bool stop = false;
 
 	if (interval) {
 		ts.tv_sec  = interval / USEC_PER_MSEC;
@@ -744,12 +764,8 @@ static int __run_perf_stat(int argc, const char **argv, int run_idx)
 		perf_evlist__start_workload(evsel_list);
 		enable_counters();
 
-		if (interval || timeout) {
-			while (!stop && !waitpid(child_pid, &status, WNOHANG)) {
-				nanosleep(&ts, NULL);
-				stop = process_timeout(timeout, interval, &times);
-			}
-		}
+		if (interval || timeout)
+			status = dispatch_events(forks, timeout, interval, &times, &ts);
 		if (child_pid != -1) {
 			if (timeout)
 				kill(child_pid, SIGTERM);
@@ -766,10 +782,7 @@ static int __run_perf_stat(int argc, const char **argv, int run_idx)
 			psignal(WTERMSIG(status), argv[0]);
 	} else {
 		enable_counters();
-		while (!done && !stop && is_target_alive(&target, evsel_list->core.threads)) {
-			nanosleep(&ts, NULL);
-			stop = process_timeout(timeout, interval, &times);
-		}
+		dispatch_events(forks, timeout, interval, &times, &ts);
 	}
 
 	disable_counters();
Signature · 1 line
2.24.1
#15 lore raw

PATCH v6 8/13 perf stat: extend -D,--delay option with -1 value

In reply to Alexey Budankov

Extend -D,--delay option with -1 value to start monitoring with events disabled to be enabled later by enable command provided via control file descriptor.

---
 tools/perf/Documentation/perf-stat.txt |  5 +++--
 tools/perf/builtin-stat.c              | 18 ++++++++++++++----
 tools/perf/util/evlist.h               |  3 +++
 tools/perf/util/stat.h                 |  2 +-
 4 files changed, 21 insertions(+), 7 deletions(-)
tools/perf/Documentation/perf-stat.txt +3 -2
index b029ee728a0b..9f32f6cd558d 100644
@@ -238,8 +238,9 @@ mode, use --per-node in addition to -a. (system-wide).
 
 -D msecs::
 --delay msecs::
-After starting the program, wait msecs before measuring. This is useful to
-filter out the startup phase of the program, which is often very different.
+After starting the program, wait msecs before measuring (-1: start with events
+disabled). This is useful to filter out the startup phase of the program,
+which is often very different.
 
 -T::
 --transaction::
tools/perf/builtin-stat.c +14 -4
index f2d1b8c1481d..8cf248bb9e78 100644
@@ -441,16 +441,26 @@ static bool process_timeout(int timeout, unsigned int interval, int *times)
 
 static void enable_counters(void)
 {
-	if (stat_config.initial_delay)
+	if (stat_config.initial_delay < 0) {
+		pr_info(EVLIST_DISABLED_MSG);
+		return;
+	}
+
+	if (stat_config.initial_delay > 0) {
+		pr_info(EVLIST_DISABLED_MSG);
 		usleep(stat_config.initial_delay * USEC_PER_MSEC);
+	}
 
 	/*
 	 * We need to enable counters only if:
 	 * - we don't have tracee (attaching to task or cpu)
 	 * - we have initial delay configured
 	 */
-	if (!target__none(&target) || stat_config.initial_delay)
+	if (!target__none(&target) || stat_config.initial_delay) {
 		evlist__enable(evsel_list);
+		if (stat_config.initial_delay > 0)
+			pr_info(EVLIST_ENABLED_MSG);
+	}
 }
 
 static void disable_counters(void)
@@ -1007,8 +1017,8 @@ static struct option stat_options[] = {
 		     "aggregate counts per thread", AGGR_THREAD),
 	OPT_SET_UINT(0, "per-node", &stat_config.aggr_mode,
 		     "aggregate counts per numa node", AGGR_NODE),
-	OPT_UINTEGER('D', "delay", &stat_config.initial_delay,
-		     "ms to wait before starting measurement after program start"),
+	OPT_INTEGER('D', "delay", &stat_config.initial_delay,
+		    "ms to wait before starting measurement after program start (-1: start with events disabled)"),
 	OPT_CALLBACK_NOOPT(0, "metric-only", &stat_config.metric_only, NULL,
 			"Only print computed metrics. No raw values", enable_metric_only),
 	OPT_BOOLEAN(0, "metric-no-group", &stat_config.metric_no_group,
tools/perf/util/evlist.h +3 -0
index bccf0a970371..7c3726a685f5 100644
@@ -377,4 +377,7 @@ int evlist__initialize_ctlfd(struct evlist *evlist, int ctl_fd, int ctl_fd_ack);
 int evlist__finalize_ctlfd(struct evlist *evlist);
 int evlist__ctlfd_process(struct evlist *evlist, enum evlist_ctl_cmd *cmd);
 
+#define EVLIST_ENABLED_MSG "Events enabled\n"
+#define EVLIST_DISABLED_MSG "Events disabled\n"
+
 #endif /* __PERF_EVLIST_H */
tools/perf/util/stat.h +1 -1
index f75ae679eb28..626421ef35c2 100644
@@ -116,7 +116,7 @@ struct perf_stat_config {
 	FILE			*output;
 	unsigned int		 interval;
 	unsigned int		 timeout;
-	unsigned int		 initial_delay;
+	int			 initial_delay;
 	unsigned int		 unit_width;
 	unsigned int		 metric_only_len;
 	int			 times;
Signature · 1 line
2.24.1
#16 lore raw

PATCH v6 9/13 perf stat: implement control commands handling

In reply to Alexey Budankov

Implement handling of 'enable' and 'disable' control commands coming from control file descriptor. process_evlist() function checks for events on static fd and makes required operations. If poll event splits initiated timeout interval then the reminder is calculated and still waited in the following poll() syscall.

---
 tools/perf/builtin-stat.c | 67 +++++++++++++++++++++++++++++----------
 1 file changed, 50 insertions(+), 17 deletions(-)
tools/perf/builtin-stat.c +50 -17
index 8cf248bb9e78..4a197b5eee23 100644
@@ -439,6 +439,31 @@ static bool process_timeout(int timeout, unsigned int interval, int *times)
 	return print_interval(interval, times);
 }
 
+static bool process_evlist(struct evlist *evlist, unsigned int interval, int *times)
+{
+	bool stop = false;
+	enum evlist_ctl_cmd cmd = EVLIST_CTL_CMD_UNSUPPORTED;
+
+	if (evlist__ctlfd_process(evlist, &cmd) > 0) {
+		switch (cmd) {
+		case EVLIST_CTL_CMD_ENABLE:
+			pr_info(EVLIST_ENABLED_MSG);
+			stop = print_interval(interval, times);
+			break;
+		case EVLIST_CTL_CMD_DISABLE:
+			stop = print_interval(interval, times);
+			pr_info(EVLIST_DISABLED_MSG);
+			break;
+		case EVLIST_CTL_CMD_ACK:
+		case EVLIST_CTL_CMD_UNSUPPORTED:
+		default:
+			break;
+		}
+	}
+
+	return stop;
+}
+
 static void enable_counters(void)
 {
 	if (stat_config.initial_delay < 0) {
@@ -514,10 +539,21 @@ static bool is_target_alive(struct target *_target,
 	return false;
 }
 
-static int dispatch_events(bool forks, int timeout, int interval, int *times, struct timespec *ts)
+static int dispatch_events(bool forks, int timeout, int interval, int *times)
 {
 	bool stop = false;
 	int child = 0, status = 0;
+	int time_to_sleep, sleep_time;
+	struct timespec time_start, time_stop, time_diff;
+
+	if (interval)
+		sleep_time = interval;
+	else if (timeout)
+		sleep_time = timeout;
+	else
+		sleep_time = 1000;
+
+	time_to_sleep = sleep_time;
 
 	while (1) {
 		if (forks)
@@ -528,8 +564,17 @@ static int dispatch_events(bool forks, int timeout, int interval, int *times, st
 		if (done || stop || child)
 			break;
 
-		nanosleep(ts, NULL);
-		stop = process_timeout(timeout, interval, times);
+		clock_gettime(CLOCK_MONOTONIC, &time_start);
+		if (!(evlist__poll(evsel_list, time_to_sleep) > 0)) { /* poll timeout or EINTR */
+			stop = process_timeout(timeout, interval, times);
+			time_to_sleep = sleep_time;
+		} else { /* fd revent */
+			stop = process_evlist(evsel_list, interval, times);
+			clock_gettime(CLOCK_MONOTONIC, &time_stop);
+			diff_timespec(&time_diff, &time_stop, &time_start);
+			time_to_sleep -= time_diff.tv_sec * MSEC_PER_SEC +
+					time_diff.tv_nsec / NSEC_PER_MSEC;
+		}
 	}
 
 	return status;
@@ -598,7 +643,6 @@ static int __run_perf_stat(int argc, const char **argv, int run_idx)
 	char msg[BUFSIZ];
 	unsigned long long t0, t1;
 	struct evsel *counter;
-	struct timespec ts;
 	size_t l;
 	int status = 0;
 	const bool forks = (argc > 0);
@@ -607,17 +651,6 @@ static int __run_perf_stat(int argc, const char **argv, int run_idx)
 	int i, cpu;
 	bool second_pass = false;
 
-	if (interval) {
-		ts.tv_sec  = interval / USEC_PER_MSEC;
-		ts.tv_nsec = (interval % USEC_PER_MSEC) * NSEC_PER_MSEC;
-	} else if (timeout) {
-		ts.tv_sec  = timeout / USEC_PER_MSEC;
-		ts.tv_nsec = (timeout % USEC_PER_MSEC) * NSEC_PER_MSEC;
-	} else {
-		ts.tv_sec  = 1;
-		ts.tv_nsec = 0;
-	}
-
 	if (forks) {
 		if (perf_evlist__prepare_workload(evsel_list, &target, argv, is_pipe,
 						  workload_exec_failed_signal) < 0) {
@@ -775,7 +808,7 @@ static int __run_perf_stat(int argc, const char **argv, int run_idx)
 		enable_counters();
 
 		if (interval || timeout)
-			status = dispatch_events(forks, timeout, interval, &times, &ts);
+			status = dispatch_events(forks, timeout, interval, &times);
 		if (child_pid != -1) {
 			if (timeout)
 				kill(child_pid, SIGTERM);
@@ -792,7 +825,7 @@ static int __run_perf_stat(int argc, const char **argv, int run_idx)
 			psignal(WTERMSIG(status), argv[0]);
 	} else {
 		enable_counters();
-		dispatch_events(forks, timeout, interval, &times, &ts);
+		dispatch_events(forks, timeout, interval, &times);
 	}
 
 	disable_counters();
Signature · 1 line
2.24.1
#17 lore raw

PATCH v6 10/13 perf stat: introduce --ctl-fd[-ack] options

In reply to Alexey Budankov

Introduce --ctl-fd[-ack] options to pass open file descriptors numbers from command line. Extend perf-stat.txt file with --ctl-fd[-ack] options description. Document possible usage model introduced by --ctl-fd[-ack] options by providing example bash shell script.

---
 tools/perf/Documentation/perf-stat.txt | 40 ++++++++++++++++++++++++++
 tools/perf/builtin-stat.c              | 11 +++++++
 tools/perf/util/stat.h                 |  2 ++
 3 files changed, 53 insertions(+)
tools/perf/Documentation/perf-stat.txt +40 -0
index 9f32f6cd558d..c4a5d31062ef 100644
@@ -176,6 +176,46 @@ with it.  --append may be used here.  Examples:
      3>results  perf stat --log-fd 3          -- $cmd
      3>>results perf stat --log-fd 3 --append -- $cmd
 
+--ctl-fd::
+--ctl-fd-ack::
+
+Listen on ctl-fd descriptor for command to control measurement ('enable': enable events,
+'disable': disable events). Optionally send control command completion ('ack') to fd-ack
+descriptor to synchronize with the controlling process. Example of bash shell script
+to enable and disable events during measurements:
+
+#!/bin/bash
+
+ctl_dir=/tmp/
+
+ctl_fifo=${ctl_dir}perf_ctl.fifo
+test -p ${ctl_fifo} && unlink ${ctl_fifo}
+mkfifo ${ctl_fifo}
+exec {ctl_fd}<>${ctl_fifo}
+
+ctl_ack_fifo=${ctl_dir}perf_ctl_ack.fifo
+test -p ${ctl_ack_fifo} && unlink ${ctl_ack_fifo}
+mkfifo ${ctl_ack_fifo}
+exec {ctl_fd_ack}<>${ctl_ack_fifo}
+
+perf stat -D -1 -e cpu-cycles -a -I 1000                \
+          --ctl-fd ${ctl_fd} --ctl-fd-ack ${ctl_fd_ack} \
+          -- sleep 30 &
+perf_pid=$!
+
+sleep 5  && echo 'enable' >&${ctl_fd} && read -u ${ctl_fd_ack} e1 && echo "enabled(${e1})"
+sleep 10 && echo 'disable' >&${ctl_fd} && read -u ${ctl_fd_ack} d1 && echo "disabled(${d1})"
+
+exec {ctl_fd_ack}>&-
+unlink ${ctl_ack_fifo}
+
+exec {ctl_fd}>&-
+unlink ${ctl_fifo}
+
+wait -n ${perf_pid}
+exit $?
+
+
 --pre::
 --post::
 	Pre and post measurement hooks, e.g.:
tools/perf/builtin-stat.c +11 -0
index 4a197b5eee23..a878e106ead0 100644
@@ -188,6 +188,8 @@ static struct perf_stat_config stat_config = {
 	.metric_only_len	= METRIC_ONLY_LEN,
 	.walltime_nsecs_stats	= &walltime_nsecs_stats,
 	.big_num		= true,
+	.ctl_fd			= -1,
+	.ctl_fd_ack		= -1
 };
 
 static inline void diff_timespec(struct timespec *r, struct timespec *a,
@@ -1080,6 +1082,10 @@ static struct option stat_options[] = {
 		"libpfm4 event selector. use 'perf list' to list available events",
 		parse_libpfm_events_option),
 #endif
+	OPT_INTEGER(0, "ctl-fd", &stat_config.ctl_fd,
+		    "Listen on fd descriptor for command to control measurement ('enable': enable events, 'disable': disable events)"),
+	OPT_INTEGER(0, "ctl-fd-ack", &stat_config.ctl_fd_ack,
+		    "Send control command completion ('ack') to fd ack descriptor"),
 	OPT_END()
 };
 
@@ -2249,6 +2255,9 @@ int cmd_stat(int argc, const char **argv)
 	signal(SIGALRM, skip_signal);
 	signal(SIGABRT, skip_signal);
 
+	if (evlist__initialize_ctlfd(evsel_list, stat_config.ctl_fd, stat_config.ctl_fd_ack))
+		goto out;
+
 	status = 0;
 	for (run_idx = 0; forever || run_idx < stat_config.run_count; run_idx++) {
 		if (stat_config.run_count != 1 && verbose > 0)
@@ -2268,6 +2277,8 @@ int cmd_stat(int argc, const char **argv)
 	if (!forever && status != -1 && (!interval || stat_config.summary))
 		print_counters(NULL, argc, argv);
 
+	evlist__finalize_ctlfd(evsel_list);
+
 	if (STAT_RECORD) {
 		/*
 		 * We synthesize the kernel mmap record just so that older tools
tools/perf/util/stat.h +2 -0
index 626421ef35c2..06f0baabe775 100644
@@ -133,6 +133,8 @@ struct perf_stat_config {
 	struct perf_cpu_map		*cpus_aggr_map;
 	u64			*walltime_run;
 	struct rblist		 metric_events;
+	int			 ctl_fd;
+	int			 ctl_fd_ack;
 };
 
 void perf_stat__set_big_num(int set);
Signature · 1 line
2.24.1
#18 lore raw

PATCH v6 11/13 perf record: extend -D,--delay option with -1 value

In reply to Alexey Budankov

Extend -D,--delay option with -1 to start collection with events disabled to be enabled later by enable command provided via control file descriptor.

---
 tools/perf/Documentation/perf-record.txt |  5 +++--
 tools/perf/builtin-record.c              | 12 ++++++++----
 tools/perf/builtin-trace.c               |  2 +-
 tools/perf/util/record.h                 |  2 +-
 4 files changed, 13 insertions(+), 8 deletions(-)
tools/perf/Documentation/perf-record.txt +3 -2
index fa8a5fcd27ab..a84376605805 100644
@@ -407,8 +407,9 @@ if combined with -a or -C options.
 
 -D::
 --delay=::
-After starting the program, wait msecs before measuring. This is useful to
-filter out the startup phase of the program, which is often very different.
+After starting the program, wait msecs before measuring (-1: start with events
+disabled). This is useful to filter out the startup phase of the program, which
+is often very different.
 
 -I::
 --intr-regs::
tools/perf/builtin-record.c +8 -4
index e108d90ae2ed..d0b29a1070a0 100644
@@ -1749,8 +1749,12 @@ static int __cmd_record(struct record *rec, int argc, const char **argv)
 	}
 
 	if (opts->initial_delay) {
-		usleep(opts->initial_delay * USEC_PER_MSEC);
-		evlist__enable(rec->evlist);
+		pr_info(EVLIST_DISABLED_MSG);
+		if (opts->initial_delay > 0) {
+			usleep(opts->initial_delay * USEC_PER_MSEC);
+			evlist__enable(rec->evlist);
+			pr_info(EVLIST_ENABLED_MSG);
+		}
 	}
 
 	trigger_ready(&auxtrace_snapshot_trigger);
@@ -2462,8 +2466,8 @@ static struct option __record_options[] = {
 	OPT_CALLBACK('G', "cgroup", &record.evlist, "name",
 		     "monitor event in cgroup name only",
 		     parse_cgroups),
-	OPT_UINTEGER('D', "delay", &record.opts.initial_delay,
-		  "ms to wait before starting measurement after program start"),
+	OPT_INTEGER('D', "delay", &record.opts.initial_delay,
+		  "ms to wait before starting measurement after program start (-1: start with events disabled)"),
 	OPT_BOOLEAN(0, "kcore", &record.opts.kcore, "copy /proc/kcore"),
 	OPT_STRING('u', "uid", &record.opts.target.uid_str, "user",
 		   "user to profile"),
tools/perf/builtin-trace.c +1 -1
index 4cbb64edc998..290149b1b3b5 100644
@@ -4813,7 +4813,7 @@ int cmd_trace(int argc, const char **argv)
 			"per thread proc mmap processing timeout in ms"),
 	OPT_CALLBACK('G', "cgroup", &trace, "name", "monitor event in cgroup name only",
 		     trace__parse_cgroups),
-	OPT_UINTEGER('D', "delay", &trace.opts.initial_delay,
+	OPT_INTEGER('D', "delay", &trace.opts.initial_delay,
 		     "ms to wait before starting measurement after program "
 		     "start"),
 	OPTS_EVSWITCH(&trace.evswitch),
tools/perf/util/record.h +1 -1
index 39d1de4b2a36..da138dcb4d34 100644
@@ -61,7 +61,7 @@ struct record_opts {
 	const char    *auxtrace_snapshot_opts;
 	const char    *auxtrace_sample_opts;
 	bool	      sample_transaction;
-	unsigned      initial_delay;
+	int	      initial_delay;
 	bool	      use_clockid;
 	clockid_t     clockid;
 	u64	      clockid_res_ns;
Signature · 1 line
2.24.1
#19 lore raw

PATCH v6 12/13 perf record: implement control commands handling

In reply to Alexey Budankov

Implement handling of 'enable' and 'disable' control commands coming from control file descriptor.

---
 tools/perf/builtin-record.c | 16 ++++++++++++++++
 1 file changed, 16 insertions(+)
tools/perf/builtin-record.c +16 -0
index d0b29a1070a0..0394e068dde8 100644
@@ -1527,6 +1527,7 @@ static int __cmd_record(struct record *rec, int argc, const char **argv)
 	bool disabled = false, draining = false;
 	int fd;
 	float ratio = 0;
+	enum evlist_ctl_cmd cmd = EVLIST_CTL_CMD_UNSUPPORTED;
 
 	atexit(record__sig_exit);
 	signal(SIGCHLD, sig_handler);
@@ -1830,6 +1831,21 @@ static int __cmd_record(struct record *rec, int argc, const char **argv)
 				alarm(rec->switch_output.time);
 		}
 
+		if (evlist__ctlfd_process(rec->evlist, &cmd) > 0) {
+			switch (cmd) {
+			case EVLIST_CTL_CMD_ENABLE:
+				pr_info(EVLIST_ENABLED_MSG);
+				break;
+			case EVLIST_CTL_CMD_DISABLE:
+				pr_info(EVLIST_DISABLED_MSG);
+				break;
+			case EVLIST_CTL_CMD_ACK:
+			case EVLIST_CTL_CMD_UNSUPPORTED:
+			default:
+				break;
+			}
+		}
+
 		if (hits == rec->samples) {
 			if (done || draining)
 				break;
Signature · 1 line
2.24.1
#20 lore raw

PATCH v6 13/13 perf record: introduce --ctl-fd[-ack] options

In reply to Alexey Budankov

Introduce --ctl-fd[-ack] options to pass open file descriptors numbers from command line. Extend perf-record.txt file with --ctl-fd[-ack] options description. Document possible usage model introduced by --ctl-fd[-ack] options by providing example bash shell script.

---
 tools/perf/builtin-record.c | 10 ++++++++++
 tools/perf/util/record.h    |  2 ++
 2 files changed, 12 insertions(+)
tools/perf/builtin-record.c +10 -0
index 0394e068dde8..8494ce964738 100644
@@ -1749,6 +1749,9 @@ static int __cmd_record(struct record *rec, int argc, const char **argv)
 		perf_evlist__start_workload(rec->evlist);
 	}
 
+	if (evlist__initialize_ctlfd(rec->evlist, opts->ctl_fd, opts->ctl_fd_ack))
+		goto out_child;
+
 	if (opts->initial_delay) {
 		pr_info(EVLIST_DISABLED_MSG);
 		if (opts->initial_delay > 0) {
@@ -1895,6 +1898,7 @@ static int __cmd_record(struct record *rec, int argc, const char **argv)
 		record__synthesize_workload(rec, true);
 
 out_child:
+	evlist__finalize_ctlfd(rec->evlist);
 	record__mmap_read_all(rec, true);
 	record__aio_mmap_read_sync(rec);
 
@@ -2380,6 +2384,8 @@ static struct record record = {
 		},
 		.mmap_flush          = MMAP_FLUSH_DEFAULT,
 		.nr_threads_synthesize = 1,
+		.ctl_fd              = -1,
+		.ctl_fd_ack          = -1,
 	},
 	.tool = {
 		.sample		= process_sample_event,
@@ -2581,6 +2587,10 @@ static struct option __record_options[] = {
 		"libpfm4 event selector. use 'perf list' to list available events",
 		parse_libpfm_events_option),
 #endif
+	OPT_INTEGER(0, "ctl-fd", &record.opts.ctl_fd,
+		    "Listen on fd descriptor for command to control measurement ('enable': enable events, 'disable': disable events)"),
+	OPT_INTEGER(0, "ctl-fd-ack", &record.opts.ctl_fd_ack,
+		   "Send control command completion ('ack') to fd ack descriptor"),
 	OPT_END()
 };
 
tools/perf/util/record.h +2 -0
index da138dcb4d34..4cb72a478af1 100644
@@ -70,6 +70,8 @@ struct record_opts {
 	int	      mmap_flush;
 	unsigned int  comp_level;
 	unsigned int  nr_threads_synthesize;
+	int	      ctl_fd;
+	int	      ctl_fd_ack;
 };
 
 extern const char * const *record_usage;
Signature · 1 line
2.24.1
reviewer
#21 lore raw

Re: [PATCH v6 13/13] perf record: introduce --ctl-fd[-ack] options

Quoted 3 lines · Alexey Budankov

Introduce --ctl-fd[-ack] options to pass open file descriptors numbers from command line. Extend perf-record.txt file with --ctl-fd[-ack]

perf-record.txt change seems to be missing.

Quoted 65 lines

options description. Document possible usage model introduced by --ctl-fd[-ack] options by providing example bash shell script.

---
 tools/perf/builtin-record.c | 10 ++++++++++
 tools/perf/util/record.h    |  2 ++
 2 files changed, 12 insertions(+)
tools/perf/builtin-record.c +10 -0
index 0394e068dde8..8494ce964738 100644
@@ -1749,6 +1749,9 @@ static int __cmd_record(struct record *rec, int argc, const char **argv)
 		perf_evlist__start_workload(rec->evlist);
 	}
 
+	if (evlist__initialize_ctlfd(rec->evlist, opts->ctl_fd, opts->ctl_fd_ack))
+		goto out_child;
+
 	if (opts->initial_delay) {
 		pr_info(EVLIST_DISABLED_MSG);
 		if (opts->initial_delay > 0) {
@@ -1895,6 +1898,7 @@ static int __cmd_record(struct record *rec, int argc, const char **argv)
 		record__synthesize_workload(rec, true);
 
 out_child:
+	evlist__finalize_ctlfd(rec->evlist);
 	record__mmap_read_all(rec, true);
 	record__aio_mmap_read_sync(rec);
 
@@ -2380,6 +2384,8 @@ static struct record record = {
 		},
 		.mmap_flush          = MMAP_FLUSH_DEFAULT,
 		.nr_threads_synthesize = 1,
+		.ctl_fd              = -1,
+		.ctl_fd_ack          = -1,
 	},
 	.tool = {
 		.sample		= process_sample_event,
@@ -2581,6 +2587,10 @@ static struct option __record_options[] = {
 		"libpfm4 event selector. use 'perf list' to list available events",
 		parse_libpfm_events_option),
 #endif
+	OPT_INTEGER(0, "ctl-fd", &record.opts.ctl_fd,
+		    "Listen on fd descriptor for command to control measurement ('enable': enable events, 'disable': disable events)"),
+	OPT_INTEGER(0, "ctl-fd-ack", &record.opts.ctl_fd_ack,
+		   "Send control command completion ('ack') to fd ack descriptor"),
 	OPT_END()
 };
 
tools/perf/util/record.h +2 -0
index da138dcb4d34..4cb72a478af1 100644
@@ -70,6 +70,8 @@ struct record_opts {
 	int	      mmap_flush;
 	unsigned int  comp_level;
 	unsigned int  nr_threads_synthesize;
+	int	      ctl_fd;
+	int	      ctl_fd_ack;
 };
 
 extern const char * const *record_usage;
#22 lore raw

Re: [PATCH v6 13/13] perf record: introduce --ctl-fd[-ack] options

Quoted 6 lines · Adrian Hunter
Quoted 3 lines · Alexey Budankov

Introduce --ctl-fd[-ack] options to pass open file descriptors numbers from command line. Extend perf-record.txt file with --ctl-fd[-ack]

perf-record.txt change seems to be missing.

Good catch, thanks!
Will add it to v7.

~Alexey