[rfc][patch] perf,tools: Remove test_attr_* muck from perf.h

[rfc][patch] perf,tools: Remove test_attr_* muck from perf.h

[Peter Zijlstra]

Hi,

I occasionally build the odd program against raw perf and use perf.h for
this.

Now I find that no longer works because of:

52502bf201a85 ("perf tests: Add framework for automated perf_event_attr tests")

So revert the hunk touching on perf.h.

Also, it seems to be for snake shit only, and snakes stink, except on a
plane :-)

---
 tools/perf/perf.h | 14 +-------------
 1 file changed, 1 insertion(+), 13 deletions(-)

diff --git a/tools/perf/perf.h b/tools/perf/perf.h
index ebdad3376c67..4a089338b60c 100644
--- a/tools/perf/perf.h
+++ b/tools/perf/perf.h
@@ -212,25 +212,13 @@ static inline unsigned long long rdclock(void)
 	(void) (&_min1 == &_min2);		\
 	_min1 < _min2 ? _min1 : _min2; })
 
-extern bool test_attr__enabled;
-void test_attr__init(void);
-void test_attr__open(struct perf_event_attr *attr, pid_t pid, int cpu,
-		     int fd, int group_fd, unsigned long flags);
-
 static inline int
 sys_perf_event_open(struct perf_event_attr *attr,
 		      pid_t pid, int cpu, int group_fd,
 		      unsigned long flags)
 {
-	int fd;
-
-	fd = syscall(__NR_perf_event_open, attr, pid, cpu,
+	return syscall(__NR_perf_event_open, attr, pid, cpu,
 		     group_fd, flags);
-
-	if (unlikely(test_attr__enabled))
-		test_attr__open(attr, pid, cpu, fd, group_fd, flags);
-
-	return fd;
 }
 
 #define MAX_COUNTERS			256

Re: [rfc][patch] perf,tools: Remove test_attr_* muck from perf.h

[Jiri Olsa]

On Thu, May 01, 2014 at 10:27:30AM +0200, Peter Zijlstra wrote:
> Hi,
> 
> I occasionally build the odd program against raw perf and use perf.h for
> this.
> 
> Now I find that no longer works because of:
> 
> 52502bf201a85 ("perf tests: Add framework for automated perf_event_attr tests")
> 
> So revert the hunk touching on perf.h.

But but but, then you'll brake attr tests

> 
> Also, it seems to be for snake shit only, and snakes stink, except on a
> plane :-)

wanna get to lwn quotes page, huh? ;)

any details about the breakage? maybe we could find some
way to make you happy and keep the snake shit in place

jirka

Re: [rfc][patch] perf,tools: Remove test_attr_* muck from perf.h

[Peter Zijlstra]

[-- Attachment #1: Type: text/plain, Size: 2548 bytes --]

On Thu, May 01, 2014 at 08:35:05PM +0200, Jiri Olsa wrote:
> On Thu, May 01, 2014 at 10:27:30AM +0200, Peter Zijlstra wrote:
> > Hi,
> > 
> > I occasionally build the odd program against raw perf and use perf.h for
> > this.
> > 
> > Now I find that no longer works because of:
> > 
> > 52502bf201a85 ("perf tests: Add framework for automated perf_event_attr tests")
> > 
> > So revert the hunk touching on perf.h.
> 
> But but but, then you'll brake attr tests
> 
> > 
> > Also, it seems to be for snake shit only, and snakes stink, except on a
> > plane :-)
> 
> wanna get to lwn quotes page, huh? ;)
> 
> any details about the breakage? maybe we could find some
> way to make you happy and keep the snake shit in place

linux-2.6/tools/perf# gcc -o perf_bug perf_bug.c
/tmp/cclObJsz.o: In function `sys_perf_event_open':
perf_bug.c:(.text+0x48): undefined reference to `test_attr__enabled'
perf_bug.c:(.text+0x72): undefined reference to `test_attr__open'
collect2: error: ld returned 1 exit status

Initially perf.h was meant to have just the syscall wrapper and few
system level defines (like the membars etc..)

But now its grown all kinds of gunk outside of that.

So ideally we'd split the thing into two files, one which is a
stand-alone syscall wrapper + system utils things and one which includes
the former and adds all the perf (tool, not kernel) gunk on.

I see all sorts of extern perf_{host,guest} things and util/target.h
includes.. that all should definitely move elsewhere.

---
#include <sys/types.h>
#include <sys/wait.h>
#include <sys/mman.h>
#include <signal.h>
#include <unistd.h>
#include <stdio.h>
#include "perf.h"

struct perf_event_attr perf_attr = {
	.type = PERF_TYPE_SOFTWARE,
	.config = PERF_COUNT_SW_EMULATION_FAULTS,
	.sample_period = 0,
	.inherit = 1,
};

int main(void)
{
	void *page;
	pid_t pid;
	int fd[16];

	fd[0] = sys_perf_event_open(&perf_attr, 0, -1, -1, 0);
	if (fd[0] < 0) {
		perror("sys_perf_event_open 0");
		return 0;
	}

	perf_attr.inherit = 0;

	fd[1] = sys_perf_event_open(&perf_attr, 0, -1, fd[0], 0);
	if (fd[1] < 0) {
		perror("sys_perf_event_open 1");
		return 0;
	}

	page = mmap(NULL, 3*4096, PROT_READ|PROT_WRITE, MAP_SHARED, fd[1], 0);
	if (page == ((void *)-1)) {
		perror("mmap");
		return 0;
	}

	pid = fork();
	if (!pid) {
		/* child */
		while (1) ;
		return 0;
	}
	usleep(500000);

	munmap(page, 3*4096);

	close(fd[1]);
	kill(pid, SIGTERM);
	waitpid(pid, NULL, 0);

	return 0;
}

[-- Attachment #2: Type: application/pgp-signature, Size: 836 bytes --]

Re: [rfc][patch] perf,tools: Remove test_attr_* muck from perf.h

[Jiri Olsa]

On Fri, May 02, 2014 at 01:52:01PM +0200, Peter Zijlstra wrote:
> On Thu, May 01, 2014 at 08:35:05PM +0200, Jiri Olsa wrote:
> > On Thu, May 01, 2014 at 10:27:30AM +0200, Peter Zijlstra wrote:
> > > Hi,
> > > 
> > > I occasionally build the odd program against raw perf and use perf.h for
> > > this.
> > > 
> > > Now I find that no longer works because of:
> > > 
> > > 52502bf201a85 ("perf tests: Add framework for automated perf_event_attr tests")
> > > 
> > > So revert the hunk touching on perf.h.
> > 
> > But but but, then you'll brake attr tests
> > 
> > > 
> > > Also, it seems to be for snake shit only, and snakes stink, except on a
> > > plane :-)
> > 
> > wanna get to lwn quotes page, huh? ;)
> > 
> > any details about the breakage? maybe we could find some
> > way to make you happy and keep the snake shit in place
> 
> linux-2.6/tools/perf# gcc -o perf_bug perf_bug.c
> /tmp/cclObJsz.o: In function `sys_perf_event_open':
> perf_bug.c:(.text+0x48): undefined reference to `test_attr__enabled'
> perf_bug.c:(.text+0x72): undefined reference to `test_attr__open'
> collect2: error: ld returned 1 exit status
> 
> Initially perf.h was meant to have just the syscall wrapper and few
> system level defines (like the membars etc..)
> 
> But now its grown all kinds of gunk outside of that.
> 
> So ideally we'd split the thing into two files, one which is a
> stand-alone syscall wrapper + system utils things and one which includes
> the former and adds all the perf (tool, not kernel) gunk on.
> 
> I see all sorts of extern perf_{host,guest} things and util/target.h
> includes.. that all should definitely move elsewhere.

sounds good, I'll prepare something along those lines

jirka