Juan,
Found the problem. It is in the example.
The reason you ceased to get notification is because the notification
queue is full.
For each notification a message is enqueued. When you get the notification, you
need to extract that message via a read(). So I think all that is
missing in your
program is.
pfarg_msg_t msg;
read(fd, &msg, sizeof(msg));
in the signal handler before pfm_restart();
In the case of trivial sampling, it is possible to request no
notification messages be
sent by setting a context flag.
ctx.ctx_flags = PFM_FL_OVFL_NO_MSG
pfm_create_context(&ctx, "default", .....)
In either case, you will getting infinite number of notifications.
On Thu, May 14, 2009 at 2:18 PM, stephane eranian
<eran...@googlemail.com> wrote:
> Juan,
>
> I am able to reproduce the problem with a modified mont_dear.c. There
> it stops after the
> 8th notification. This is quite strange given this is all self
> monitoring. Let me investigate.
> I also want to verify if this is IA-64 specific or not.
>
>
> On Thu, May 14, 2009 at 11:49 AM, Juan Angel Lorenzo
> <jalmailingli...@gmail.com> wrote:
>> Hi,
>>
>> El mié, 13-05-2009 a las 14:11 +0200, stephane eranian escribió:
>>
>>>
>>> > The number of notifications is the number of times the sampling buffer
>>> > overflowed, isn't it? So why does it look limited to 31? I mean, even if
>>>
>>> Yes, assuming your resume monitoring() via pfm_restart() at the end of
>>> each notification.
>>
>> Yes, as far as I saw, this is what the "mont_dear.c" code does. It
>> resumes monitoring every time a notification is handled. But it looks
>> there is a limit of 31 times the monitoring can be resumed. And then it
>> stops. I don't understand why. I thought this process could be done
>> infinite times. Could it be perhaps because of any limit in the
>> operating system settings?
>>
>>
>>> Did you read the pmds in the handler via pfm_read_pmds()?
>>>
>> Yes. Briefly, this is what I did (mixing code from mont_dear.c,
>> mont_rr.c and notify_self.c):
>>
>> #include <perfmon/perfmon.h>
>> #include <perfmon/perfmon_default_smpl.h>
>> #include <perfmon/pfmlib_montecito.h>
>>
>> #define SMPL_PERIOD (40)
>> #define M_PMD(x) (1UL<<(x))
>> #define DEAR_REGS_MASK (M_PMD(32)|M_PMD(33)|M_PMD(36))
>>
>> typedef pfm_default_smpl_hdr_t dear_hdr_t;
>> typedef pfm_default_smpl_entry_t dear_entry_t;
>> typedef pfm_default_smpl_ctx_arg_t dear_ctx_t;
>>
>> static int ctx_fd;
>> static char *event1_name = "data_ear_cache_lat4";
>> static pfarg_reg_t pd[NUM_PMDS];
>>
>>
>> static void
>> sigio_handler(int n, struct siginfo *info, struct sigcontext *sc)
>> {
>> pfm_msg_t msg;
>> pfm_mont_pmd_reg_t reg;
>> int fd = ctx_fd;
>> int r;
>>
>> if (fd != ctx_fd) {
>> fatal_error("handler does not get valid file descriptor\n");
>> }
>>
>> if (event1_name && perfmonctl(fd, PFM_READ_PMDS, pd+1, 1) == -1) {
>> fatal_error("PFM_READ_PMDS: %s", strerror(errno));
>> }
>>
>> r = read(fd, &msg, sizeof(msg));
>> if (r != sizeof(msg)) {
>> fatal_error("cannot read overflow message: %s\n", strerror(errno));
>> }
>>
>> printf("PMD pfarg_reg_t.reg_num: %u\n",pd[1].reg_num);
>> printf("PMD pfarg_reg_t.reg_value: %lu\n",pd[1].reg_value);
>> reg = (pfm_mont_pmd_reg_t)pd[1].reg_value;
>> printf("PMD32: 0x%016lx\n", reg.pmd32_mont_reg.dear_daddr);
>>
>> ........
>>
>> /*
>> * And resume monitoring
>> */
>> if (perfmonctl(fd, PFM_RESTART,NULL, 0) == -1) {
>> fatal_error("PFM_RESTART: %s", strerror(errno));
>> }
>> }
>>
>>
>>
>> int
>> main(int argc, char **argv)
>> {
>>
>> pfarg_reg_t pc[NUM_PMCS];
>> pfmlib_input_param_t inp;
>> pfmlib_output_param_t outp;
>> pfmlib_event_t ev;
>> dear_ctx_t ctx;
>> pfarg_load_t load_args;
>> pfmlib_options_t pfmlib_options;
>> struct sigaction act;
>>
>> ......
>> unsigned long range_start, range_end;
>> pfmlib_mont_input_param_t mont_inp;
>> pfmlib_mont_output_param_t mont_outp;
>> pfarg_dbreg_t dbrs[8];
>>
>> ......
>>
>> if (pfm_find_full_event(event1_name, &ev) != PFMLIB_SUCCESS)
>> fatal_error("cannot find event %s\n", event1_name);
>>
>> inp.pfp_dfl_plm = PFM_PLM3|PFM_PLM0;
>>
>> /*
>> * how many counters we use
>> */
>> inp.pfp_event_count = 1;
>>
>> /*
>> * propagate the event descriptor
>> */
>> inp.pfp_events[0] = ev;
>>
>>
>> mont_inp.pfp_mont_drange.rr_used = 1;
>> mont_inp.pfp_mont_drange.rr_limits[0].rr_start = range_start;
>> mont_inp.pfp_mont_drange.rr_limits[0].rr_end = range_end;
>>
>> if ((ret=pfm_dispatch_events(&inp, &mont_inp,
>> &outp,&mont_outp)) != PFMLIB_SUCCESS)
>> fatal_error("cannot configure events: %s\n",
>> pfm_strerror(ret));
>>
>> /*
>> * now create the context for self monitoring/per-task
>> */
>> if (perfmonctl(0, PFM_CREATE_CONTEXT, &ctx, 1) == -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));
>> }
>>
>> ctx_fd = ctx.ctx_arg.ctx_fd;
>>
>>
>> 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;
>> }
>>
>>
>> /*
>> * figure out pmd mapping from output pmc
>> */
>> for (i=0; i < outp.pfp_pmd_count; i++)
>> pd[i].reg_num = outp.pfp_pmds[i].reg_num;
>>
>> /*
>> * We want to get notified when the counter used for our first
>> * event overflows
>> */
>> pc[0].reg_flags |= PFM_REGFL_OVFL_NOTIFY;
>> pc[0].reg_reset_pmds[0] |= 1UL << outp.pfp_pmds[1].reg_num;
>>
>>
>> /*
>> * initialize the PMD and the sampling period
>> */
>> pd[0].reg_value = - SMPL_PERIOD;
>> pd[0].reg_long_reset = - SMPL_PERIOD;
>> pd[0].reg_short_reset = - SMPL_PERIOD;
>>
>>
>> if (perfmonctl(ctx_fd, PFM_WRITE_DBRS, dbrs,
>> mont_outp.pfp_mont_drange.rr_nbr_used) == -1) {
>> fatal_error( "child: perfmonctl error PFM_WRITE_DBRS
>> errno %d\n",errno);
>> }
>>
>>
>> if (perfmonctl(ctx_fd, PFM_WRITE_PMCS, pc, outp.pfp_pmc_count))
>> fatal_error("pfm_write_pmcs error errno %d\n",errno);
>>
>> if (perfmonctl(ctx_fd, PFM_WRITE_PMDS, pd, outp.pfp_pmd_count))
>> fatal_error("pfm_write_pmds error errno %d\n",errno);
>>
>> /*
>> * attach context to stopped task
>> */
>> load_args.load_pid = getpid();
>> if (perfmonctl(ctx_fd, PFM_LOAD_CONTEXT, &load_args, 1))
>> fatal_error("pfm_load_context error errno %d\n",errno);
>>
>>
>> /*
>> * setup asynchronous notification on the file descriptor
>> */
>> ret = fcntl(ctx_fd, F_SETFL, fcntl(ctx_fd, F_GETFL, 0) |
>> O_ASYNC);
>>
>> /*
>> * get ownership of the descriptor
>> */
>> ret = fcntl(ctx_fd, F_SETOWN, getpid());
>>
>> /*
>> * Let's roll now.
>> */
>> pfm_self_start(ctx_fd);
>>
>> .......
>>
>> pfm_self_stop(ctx_fd);
>>
>> close(ctx_fd);
>> return 0;
>> }
>>
>>
>> There's one thing I'm not sure is right. There was a line in
>> notify_self.c:
>>
>> pc[0].reg_reset_pmds[0] |= 1UL << outp.pfp_pmcs[1].reg_num;
>>
>> which gives me a "pfm_write_pmcs error errno 22" error in this code. I
>> think it intends to set the byte 32th in the reg_reset_pmds[0] field, so
>> I changed it to:
>>
>> pc[0].reg_reset_pmds[0] |= 1UL << outp.pfp_pmds[1].reg_num;
>>
>> So the code runs, but when I get a notification, the value of
>> pd[1].reg_value, in the sigio_handler, is always 0.
>>
>> Again, thanks for the help and apologies for the long email. If this
>> code is not clear enough, I can attach the whole file.
>>
>> Juan
>>
>>
>
------------------------------------------------------------------------------
The NEW KODAK i700 Series Scanners deliver under ANY circumstances! Your
production scanning environment may not be a perfect world - but thanks to
Kodak, there's a perfect scanner to get the job done! With the NEW KODAK i700
Series Scanner you'll get full speed at 300 dpi even with all image
processing features enabled. http://p.sf.net/sfu/kodak-com
_______________________________________________
perfmon2-devel mailing list
perfmon2-devel@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/perfmon2-devel
