Stephane Eranian wrote:
Will,
I got OProfile to work on perfmon v2.3 all IA-64 and X86. I need to
clean the user level code a little bit. I would appreciate if you could
take a look at it.
There is one thing I am not so sure about at this point: how does OProfile
handle unavailable counters when it assigns event -> counters,
take the configuration with the NMI watchdog, for instance.
Do you know anything about this?
Yes, I developed a patch to handle this for OProfile to work with some registers
being reserved for watchdog. However, I seemed to have forgotten to post the
patch. I have attached the patch to this email.
The way I have it setup right now does not use libpfm for event -> counter
assignment. I just use it for interface definitions and perfmon syscalls.
So I rely on OProfile to map events to OProfile counters, then I map
those onto perfmon PMC/PMD registers.
The other thing I noticed is that Montecito events are not available.
I guess no one has taken the time to enter the Montecito events into OProfile.
-Will
Index: libop/op_alloc_counter.c
===================================================================
RCS file: /cvsroot/oprofile/oprofile/libop/op_alloc_counter.c,v
retrieving revision 1.7
diff -u -r1.7 op_alloc_counter.c
--- libop/op_alloc_counter.c 17 Nov 2006 18:20:38 -0000 1.7
+++ libop/op_alloc_counter.c 18 Jan 2007 20:41:50 -0000
@@ -12,6 +12,8 @@
*/
#include <stdlib.h>
+#include <ctype.h>
+#include <dirent.h>
#include "op_events.h"
#include "op_libiberty.h"
@@ -143,6 +145,42 @@
return 0;
}
+/* determine which directories are counter directories
+ */
+static int perfcounterdir(const struct dirent * entry)
+{
+ return (isdigit(entry->d_name[0]));
+}
+
+
+/**
+ * @param mask pointer where to place bit mask of unavailable counters
+ *
+ * return >= 0 number of counters that are available
+ * < 0 could not determine number of counters
+ *
+ */
+static int op_get_counter_mask(u32 * mask)
+{
+ struct dirent **counterlist;
+ int count, i;
+ /* assume nothing is available */
+ u32 available=0;
+
+ count = scandir("/dev/oprofile", &counterlist, perfcounterdir,
+ alphasort);
+ if (count < 0)
+ /* unable to determine bit mask */
+ return -1;
+ /* convert to bit map (0 where counter exists) */
+ for (i=0; i<count; ++i) {
+ available |= 1 << atoi(counterlist[i]->d_name);
+ free(counterlist[i]);
+ }
+ *mask=~available;
+ free(counterlist);
+ return count;
+}
size_t * map_event_to_counter(struct op_event const * pev[], int nr_events,
op_cpu cpu_type)
@@ -150,8 +188,11 @@
counter_arc_head * ctr_arc;
size_t * counter_map;
int nr_counters;
+ u32 unavailable_counters = 0;
- nr_counters = op_get_nr_counters(cpu_type);
+ nr_counters = op_get_counter_mask(&unavailable_counters);
+ if (nr_counters < 0)
+ nr_counters = op_get_nr_counters(cpu_type);
if (nr_counters < nr_events)
return 0;
@@ -159,7 +200,8 @@
counter_map = xmalloc(nr_counters * sizeof(size_t));
- if (!allocate_counter(ctr_arc, nr_events, 0, 0, counter_map)) {
+ if (!allocate_counter(ctr_arc, nr_events, 0, unavailable_counters,
+ counter_map)) {
free(counter_map);
counter_map = 0;
}
_______________________________________________
perfmon mailing list
[email protected]
http://www.hpl.hp.com/hosted/linux/mail-archives/perfmon/
