Re: [PATCH] Export per-thread resource usage via sysctl

On Saturday, July 16, 2011 8:04:58 am Alexander Best wrote:
> On Sat Jul 16 11, Pan Tsu wrote:
> > Alexander Best <arundel_at_freebsd.org> writes:
> > 
> > > On Fri Jul 15 11, John Baldwin wrote:
> > >> This change exports each individual thread's resource usage via sysctl when 
> > >> individual threads are requested via KERN_PROC_INC_THREAD.  This generally 
> > >> works correctly with 'top -m io' after the previous change to revert top(1) 
> > >> back to using KERN_PROC_PROC when threads are not enabled.  There is one issue 
> > >> in that top doesn't necessarily DTRT when disabling/enabling threads via 'H' 
> > >> at runtime while in io mode.  I may do some further work to clean that up.  
> > >> However, for just top run it will now show per-thread stats instead of 
> > >> duplicating the per-process stats for each thread.
> > >
> > > i'm not sure, if i understand what the patch is supposed to do. however after
> > > applying it, and recompiling/reinstalling the kernel, 'top -mio' displays the
> > > same stats for each thread of a process. if i understood you correctly, each
> > > thread should have individual stats.
> > >
> > > i'm running r224068 on amd64 and just reinstalled 'top'. anything i am missing?
> > 
> > FWIW, I see different numbers for a few threads of firefox-bin with top-3.8b1.
> > 
> >   http://img233.imageshack.us/img233/1570/81482202.png
> > 
> > Which is an improvement compared to how all threads showed same numbers
> > before applying the patch.
> 
> hmmm...not here. i did the following: 'top -mio -b -H -d2 999999' and had a
> look at the second output, where if found these lines:

Hmm, I never saw that.  However, I do need this patch to top to get consistently sane numbers
(otherwise top picks one thread as the old thread for all threads causing diffs):

Index: machine.c
===================================================================
--- machine.c	(revision 224062)
+++ machine.c	(working copy)
_at__at_ -235,6 +235,7 _at__at_ static int *pcpu_cpu_states;
 
 static int compare_jid(const void *a, const void *b);
 static int compare_pid(const void *a, const void *b);
+static int compare_tid(const void *a, const void *b);
 static const char *format_nice(const struct kinfo_proc *pp);
 static void getsysctl(const char *name, void *ptr, size_t len);
 static int swapmode(int *retavail, int *retfree);
_at__at_ -557,7 +558,7 _at__at_ get_old_proc(struct kinfo_proc *pp)
 	 * cache it.
 	 */
 	oldpp = bsearch(&pp, previous_pref, previous_proc_count,
-	    sizeof(*previous_pref), compare_pid);
+	    sizeof(*previous_pref), ps.thread ? compare_tid : compare_pid);
 	if (oldpp == NULL) {
 		pp->ki_udata = NOPROC;
 		return (NULL);
_at__at_ -652,7 +653,7 _at__at_ get_process_info(struct system_info *si, struct pr
 			previous_pref[i] = &previous_procs[i];
 		bcopy(pbase, previous_procs, nproc * sizeof(*previous_procs));
 		qsort(previous_pref, nproc, sizeof(*previous_pref),
-		    compare_pid);
+		    ps.thread ? compare_tid : compare_pid);
 	}
 	previous_proc_count = nproc;
 
_at__at_ -1059,6 +1060,18 _at__at_ compare_pid(const void *p1, const void *p2)
 	return ((*pp1)->ki_pid - (*pp2)->ki_pid);
 }
 
+static int
+compare_tid(const void *p1, const void *p2)
+{
+	const struct kinfo_proc * const *pp1 = p1;
+	const struct kinfo_proc * const *pp2 = p2;
+
+	if ((*pp2)->ki_tid < 0 || (*pp1)->ki_tid < 0)
+		abort();
+
+	return ((*pp1)->ki_tid - (*pp2)->ki_tid);
+}
+
 /*
  *  proc_compare - comparison function for "qsort"
  *	Compares the resource consumption of two processes using five

However, using this test app:

#include <sys/types.h>
#include <err.h>
#include <fcntl.h>
#include <pthread.h>
#include <unistd.h>
#include <stdio.h>

#define	BUFFER_LEN	16384
#define	FILE_LEN	(BUFFER_LEN * 128)

char buffer[BUFFER_LEN];
int fd;

static void *
work(void *arg)
{
	int mode = (intptr_t)arg;
	ssize_t result;
	off_t offset;
	static const char* names[] = { "idle", "hog", "read", "write" };

	pthread_set_name_np(pthread_self(), names[mode]);
	offset = 0;
	for (;;) {
		switch (mode) {
		case 0:
			sleep(60);
			break;
		case 1:
			pthread_yield();
			break;
		case 2:
		case 3:
			if (mode == 2)
				result = pread(fd, buffer, sizeof(buffer),
				    offset);
			else
				result = pwrite(fd, buffer, sizeof(buffer),
				    offset);
			if (result < 0)
				warn("%s", mode == 2 ? "pread" : "pwrite");
			offset += sizeof(buffer);
			if (offset >= FILE_LEN)
				offset = 0;
			usleep(1000 * 100);
			break;
		}
	}

	return (NULL);
}

int
main(int ac, char **av)
{
	pthread_t thread;
	char template[] = "/tmp/t.XXXXXXXX";
	int error, flags;

	fd = mkstemp(template);
	if (fd < 0)
		err(1, "mkstemp");
	if (unlink(template) < 0)
		err(1, "unlink(%s)", template);
	if (ftruncate(fd, FILE_LEN) < 0)
		err(1, "ftruncate");
	flags = fcntl(fd, F_GETFL);
	if (flags < 0)
		err(1, "fcntl(F_GETFL)");
	if (fcntl(fd, F_SETFL, flags | O_DIRECT) < 0)
		err(1, "fcntl(F_SETFL, O_DIRECT)");
	printf("PID: %ld\n", (long)getpid());
	error = pthread_create(&thread, NULL, work, (void *)0);
	if (error)
		errc(1, error, "pthread_create");
/*
	error = pthread_create(&thread, NULL, work, (void *)1);
	if (error)
		errc(1, error, "pthread_create");
*/
	error = pthread_create(&thread, NULL, work, (void *)2);
	if (error)
		errc(1, error, "pthread_create");
	(void)work((void *)3);
	return (0);
}

I get top output like so with this patch and the above patch applied to the
latest top:

  PID USERNAME     VCSW  IVCSW   READ  WRITE  FAULT  TOTAL PERCENT COMMAND
25003 jhb            20      0      0      6      0      6  75.00% {write}
25003 jhb            21      0      1      0      0      1  12.50% {read}
25003 jhb             0      0      0      0      0      0   0.00% {idle}

Without the top patch I get bogus output like so:

25003 jhb            20      0      0      0      0      0  -0.00% {write}
25003 jhb            41      5     34    -75      0    -41  33.33% {read}
25003 jhb         -2326     -1     -7    -75      0    -82  66.67% {idle}

-- 
John Baldwin