Hi guys,
It seems that an older problem has crept back in to the kernel patch. I
seem to remember this being fixed once before...but at the moment, it is
still there.
The normal test cases inside of perfmon don't catch this case, so I have
included one here. Basically it's self that does multiple start and
stops. Manoj, Mark, can you verify this? Test case has been attached.
One good run looks like this:
PMD0 67073650 CYCLES
PMD1 53000183 INSNS_COMPLETE
PMD0 66098650 CYCLES
PMD1 53000181 INSNS_COMPLETE
PMD0 66094923 CYCLES
PMD1 53000181 INSNS_COMPLETE
Occasionally it looks like this. Sometimes it's easier to see if you run
multiple copies.
PMD0 66066007 CYCLES
PMD1 53000182 INSNS_COMPLETE
PMD0 19935908 CYCLES
PMD1 15730260 INSNS_COMPLETE
PMD0 58149463 CYCLES
PMD1 45930517 INSNS_COMPLETE
I have verified this on both 64 and 32 bit builds on a MIPS 25K with an
unmodified (except for syscalls) patch.
Stefane, do you have any obvious ideas about where I should look? The
code currently saves and restores all counters (unlike my previous
mails).
---
BTW, the MIPS LIBPFM code I have needed some fixes for 64 bit builds as
well as some other minor things.
1) Change uint32_t to unsigned int in lib/pfmlib_gen_mips64_priv.h
2) Add perfmon/pfmlib_gen_mips64.h to include/Makefile at line 92
3) Code for get_event_counters should look like this:
> *code = 0xff & (gen_mips64_pe[i].pme_code >> (cnt*8));
Instead of:
< *code = 0xf & (gen_mips64_pe[i].pme_code >> (cnt*4));
4) Fix up domain bits to be 'standardized'
diff -r1.5 pfmlib_gen_mips64.h
35,37c35,37
< * PFM_PLM0 = SUPERVISOR
< * PFM_PLM1 = INTERRUPT
< * PFM_PLM2 = KERNEL
---
> * PFM_PLM0 = KERNEL
> * PFM_PLM2 = INTERRUPT
> * PFM_PLM1 = SUPERVISOR
In stuff_regs:
< reg.sel_os = plm & PFM_PLM2 ? 1 : 0;
---
> reg.sel_os = plm & PFM_PLM0 ? 1 : 0;
> reg.sel_exl = plm & PFM_PLM2 ? 1 : 0;
202d227
< reg.sel_exl = plm & PFM_PLM0 ? 1 : 0;
/*
* self.c - example of a simple self monitoring task
*
* Copyright (c) 2002-2006 Hewlett-Packard Development Company, L.P.
* Contributed by Stephane Eranian <[EMAIL PROTECTED]>
*
* Permission is hereby granted, free of charge, to any person obtaining a copy
* of this software and associated documentation files (the "Software"), to deal
* in the Software without restriction, including without limitation the rights
* to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies
* of the Software, and to permit persons to whom the Software is furnished to do so,
* subject to the following conditions:
*
* The above copyright notice and this permission notice shall be included in all
* copies or substantial portions of the Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED,
* INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A
* PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT
* HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF
* CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE
* OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
*
* This file is part of libpfm, a performance monitoring support library for
* applications on Linux.
*/
#include <sys/types.h>
#include <inttypes.h>
#include <stdio.h>
#include <stdlib.h>
#include <stdarg.h>
#include <errno.h>
#include <unistd.h>
#include <string.h>
#include <signal.h>
#include <perfmon/pfmlib.h>
#include <perfmon/perfmon.h>
#include "detect_pmcs.h"
#define NUM_PMCS PFMLIB_MAX_PMCS
#define NUM_PMDS PFMLIB_MAX_PMDS
#define ITERS 1000000ULL
#define TABSIZE 1024
static int a[TABSIZE], b[TABSIZE];
/*
* our test code (function cannot be made static otherwise it is optimized away)
*/
uint64_t
noploop(uint64_t loop)
{
unsigned int j = 0;
while (loop--) { a[j] += a[j]*loop + b[j];j = (j+1)%TABSIZE; }
return a[0];
}
static void fatal_error(char *fmt,...) __attribute__((noreturn));
static void
fatal_error(char *fmt, ...)
{
va_list ap;
va_start(ap, fmt);
vfprintf(stderr, fmt, ap);
va_end(ap);
exit(1);
}
int
main(int argc, char **argv)
{
char **p;
unsigned int i;
int ret, ctx_fd;
pfmlib_input_param_t inp;
pfmlib_output_param_t outp;
pfarg_pmd_t pd[NUM_PMDS];
pfarg_pmc_t pc[NUM_PMCS];
pfarg_ctx_t ctx;
pfarg_load_t load_args;
pfmlib_options_t pfmlib_options;
unsigned int num_counters;
size_t len;
char *name;
/*
* Initialize pfm library (required before we can use it)
*/
ret = pfm_initialize();
if (ret != PFMLIB_SUCCESS)
fatal_error("Cannot initialize library: %s\n", pfm_strerror(ret));
pfm_get_max_event_name_len(&len);
name = malloc(len+1);
if (!name)
fatal_error("cannot allocate event name buffer\n");
pfm_get_num_counters(&num_counters);
/*
* pass options to library (optional)
*/
memset(&pfmlib_options, 0, sizeof(pfmlib_options));
pfmlib_options.pfm_debug = 1; /* set to 1 for debug */
pfmlib_options.pfm_verbose = 1; /* set to 1 for verbose */
pfm_set_options(&pfmlib_options);
memset(pd, 0, sizeof(pd));
memset(pc, 0, sizeof(pc));
memset(&ctx, 0, sizeof(ctx));
memset(&load_args, 0, sizeof(load_args));
/*
* prepare parameters to library.
*/
memset(&inp,0, sizeof(inp));
memset(&outp,0, sizeof(outp));
/*
* be nice to user!
*/
if (argc > 1) {
p = argv+1;
for (i=0; *p ; i++, p++) {
ret = pfm_find_full_event(*p, &inp.pfp_events[i]);
if (ret != PFMLIB_SUCCESS)
fatal_error("event %s: %s\n", *p, pfm_strerror(ret));
}
} else {
if (pfm_get_cycle_event(&inp.pfp_events[0]) != PFMLIB_SUCCESS)
fatal_error("cannot find cycle event\n");
if (pfm_get_inst_retired_event(&inp.pfp_events[1]) != PFMLIB_SUCCESS)
fatal_error("cannot find inst retired event\n");
i = 2;
}
/*
* set the default privilege mode for all counters:
* PFM_PLM3 : user level only
*/
inp.pfp_dfl_plm = PFM_PLM3;
if (i > num_counters) {
i = num_counters;
printf("too many events provided (max=%d events), using first %d event(s)\n", num_counters, i);
}
/*
* how many counters we use
*/
inp.pfp_event_count = i;
/*
* now create a new context, per process context.
* This just creates a new context with some initial state, it is not
* active nor attached to any process.
*/
ctx_fd = pfm_create_context(&ctx, NULL, NULL, 0);
if (ctx_fd == -1) {
if (errno == ENOSYS)
fatal_error("Your kernel does not have performance monitoring support!\n");
fatal_error("Can't create PFM context %s\n", strerror(errno));
}
/*
* build the pfp_unavail_pmcs bitmask by looking
* at what perfmon has available. It is not always
* the case that all PMU registers are actually available
* to applications. For instance, on IA-32 platforms, some
* registers may be reserved for the NMI watchdog timer.
*
* With this bitmap, the library knows which registers NOT to
* use. Of source, it is possible that no valid assignement may
* be possible if certina PMU registers are not available.
*/
detect_unavail_pmcs(ctx_fd, &inp.pfp_unavail_pmcs);
/*
* let the library figure out the values for the PMCS
*/
if ((ret=pfm_dispatch_events(&inp, NULL, &outp, NULL)) != PFMLIB_SUCCESS)
fatal_error("cannot configure events: %s\n", pfm_strerror(ret));
/*
* Now prepare the argument to initialize the PMDs and PMCS.
* We use pfp_pmc_count to determine the number of PMC to intialize.
* We use pfp_pmd_count to determine the number of PMD to initialize.
* Some events/features may cause extra PMCs to be used, leading to:
* - pfp_pmc_count may be >= pfp_event_count
* - pfp_pmd_count may be >= pfp_event_count
*/
for (i=0; i < outp.pfp_pmc_count; i++) {
pc[i].reg_num = outp.pfp_pmcs[i].reg_num;
pc[i].reg_value = outp.pfp_pmcs[i].reg_value;
}
for (i=0; i < outp.pfp_pmd_count; i++) {
pd[i].reg_num = outp.pfp_pmds[i].reg_num;
}
/*
* Now program the registers
*/
if (pfm_write_pmcs(ctx_fd, pc, outp.pfp_pmc_count))
fatal_error("pfm_write_pmcs error errno %d\n",errno);
if (pfm_write_pmds(ctx_fd, pd, outp.pfp_pmd_count))
fatal_error("pfm_write_pmds error errno %d\n",errno);
/*
* now we load (i.e., attach) the context to ourself
*/
load_args.load_pid = getpid();
if (pfm_load_context(ctx_fd, &load_args))
fatal_error("pfm_load_context error errno %d\n",errno);
noploop(ITERS);
/*
* Let's roll now
*/
if (pfm_start(ctx_fd, NULL))
fatal_error("pfm_start error errno %d\n",errno);
noploop(ITERS);
if (pfm_stop(ctx_fd))
fatal_error("pfm_stop error errno %d\n",errno);
/*
* now read the results. We use pfp_event_count because
* libpfm guarantees that counters for the events always
* come first.
*/
if (pfm_read_pmds(ctx_fd, pd, inp.pfp_event_count))
fatal_error( "pfm_read_pmds error errno %d\n",errno);
/*
* print the results
*/
for (i=0; i < inp.pfp_event_count; i++) {
pfm_get_full_event_name(&inp.pfp_events[i], name, len+1);
printf("PMD%-3u %20"PRIu64" %s\n",
pd[i].reg_num,
pd[i].reg_value,
name);
pd[i].reg_value = 0ULL;
}
if (pfm_write_pmds(ctx_fd, pd, outp.pfp_pmd_count))
fatal_error("pfm_write_pmds error errno %d\n",errno);
/*
* Let's roll now
*/
if (pfm_start(ctx_fd, NULL))
fatal_error("pfm_start error errno %d\n",errno);
noploop(ITERS);
if (pfm_stop(ctx_fd))
fatal_error("pfm_stop error errno %d\n",errno);
/*
* now read the results. We use pfp_event_count because
* libpfm guarantees that counters for the events always
* come first.
*/
if (pfm_read_pmds(ctx_fd, pd, inp.pfp_event_count))
fatal_error( "pfm_read_pmds error errno %d\n",errno);
/*
* print the results
*/
for (i=0; i < inp.pfp_event_count; i++) {
pfm_get_full_event_name(&inp.pfp_events[i], name, len+1);
printf("PMD%-3u %20"PRIu64" %s\n",
pd[i].reg_num,
pd[i].reg_value,
name);
pd[i].reg_value = 0ULL;
}
if (pfm_write_pmds(ctx_fd, pd, outp.pfp_pmd_count))
fatal_error("pfm_write_pmds error errno %d\n",errno);
/*
* Let's roll now
*/
if (pfm_start(ctx_fd, NULL))
fatal_error("pfm_start error errno %d\n",errno);
noploop(ITERS);
if (pfm_stop(ctx_fd))
fatal_error("pfm_stop error errno %d\n",errno);
/*
* now read the results. We use pfp_event_count because
* libpfm guarantees that counters for the events always
* come first.
*/
if (pfm_read_pmds(ctx_fd, pd, inp.pfp_event_count))
fatal_error( "pfm_read_pmds error errno %d\n",errno);
/*
* print the results
*/
for (i=0; i < inp.pfp_event_count; i++) {
pfm_get_full_event_name(&inp.pfp_events[i], name, len+1);
printf("PMD%-3u %20"PRIu64" %s\n",
pd[i].reg_num,
pd[i].reg_value,
name);
pd[i].reg_value = 0ULL;
}
free(name);
/*
* and destroy our context
*/
close(ctx_fd);
return 0;
}
_______________________________________________
perfmon mailing list
[email protected]
http://www.hpl.hp.com/hosted/linux/mail-archives/perfmon/
