int
main(int argc, char **argv)
{
pfmlib_options_t pfmlib_options;
if (argc < 2)
fatal_error("You must specify a command to execute\n");
/*
* Initialize pfm library (required before we can use it)
*/
if (pfm_initialize() != PFMLIB_SUCCESS) {
fatal_error("Can't initialize library\n");
}
/*
* pass options to library (optional)
*/
memset(&pfmlib_options, 0, sizeof(pfmlib_options));
pfmlib_options.pfm_debug = 0; /* set to 1 for debug */
pfmlib_options.pfm_verbose = 0; /* set to 1 for verbose */
pfm_set_options(&pfmlib_options);
return mainloop(argv+1);
}
int
main(int argc, char **argv)
{
pfmlib_options_t pfmlib_options;
unsigned long delay;
pid_t pid;
int ret;
if (argc < 2)
fatal_error("usage: %s pid [timeout]\n", argv[0]);
pid = atoi(argv[1]);
delay = argc > 2 ? strtoul(argv[2], NULL, 10) : 10;
/*
* pass options to library (optional)
*/
memset(&pfmlib_options, 0, sizeof(pfmlib_options));
pfmlib_options.pfm_debug = 0; /* set to 1 for debug */
pfm_set_options(&pfmlib_options);
/*
* 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));
return parent(pid, delay);
}
int
main(int argc, char **argv)
{
char **p;
int i, ret;
pid_t pid = getpid();
pfmlib_param_t evt;
pfarg_reg_t pd[NUM_PMDS];
pfarg_context_t ctx[1];
pfmlib_options_t pfmlib_options;
/*
* Initialize pfm library (required before we can use it)
*/
if (pfm_initialize() != PFMLIB_SUCCESS) {
printf("Can't initialize library\n");
exit(1);
}
/*
* check that the user did not specify too many events
*/
if (argc-1 > pfm_get_num_counters()) {
printf("Too many events specified\n");
exit(1);
}
/*
* pass options to library (optional)
*/
memset(&pfmlib_options, 0, sizeof(pfmlib_options));
pfmlib_options.pfm_debug = 0; /* set to 1 for debug */
pfm_set_options(&pfmlib_options);
memset(pd, 0, sizeof(pd));
memset(ctx, 0, sizeof(ctx));
/*
* prepare parameters to library. we don't use any Itanium
* specific features here. so the pfp_model is NULL.
*/
memset(&evt,0, sizeof(evt));
/*
* be nice to user!
*/
p = argc > 1 ? argv+1 : event_list;
for (i=0; *p ; i++, p++) {
if (pfm_find_event(*p, &evt.pfp_events[i].event) != PFMLIB_SUCCESS) {
fatal_error("Cannot find %s event\n", *p);
}
}
/*
* set the default privilege mode for all counters:
* PFM_PLM3 : user level only
*/
evt.pfp_dfl_plm = PFM_PLM3;
/*
* how many counters we use
*/
evt.pfp_event_count = i;
/*
* let the library figure out the values for the PMCS
*/
if ((ret=pfm_dispatch_events(&evt)) != PFMLIB_SUCCESS) {
fatal_error("cannot configure events: %s\n", pfm_strerror(ret));
}
/*
* for this example, we have decided not to get notified
* on counter overflows and the monitoring is not to be inherited
* in derived tasks.
*/
ctx[0].ctx_flags = PFM_FL_INHERIT_NONE;
/*
* now create the context for self monitoring/per-task
*/
if (perfmonctl(pid, 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));
}
/*
* Must be done before any PMD/PMD calls (unfreeze PMU). Initialize
* PMC/PMD to safe values. psr.up is cleared.
*/
if (perfmonctl(pid, PFM_ENABLE, NULL, 0) == -1) {
fatal_error("perfmonctl error PFM_ENABLE errno %d\n",errno);
}
/*
* Now prepare the argument to initialize the PMDs.
* the memset(pd) initialized the entire array to zero already, so
* we just have to fill in the register numbers from the pc[] array.
*/
for (i=0; i < evt.pfp_event_count; i++) {
pd[i].reg_num = evt.pfp_pc[i].reg_num;
}
/*
* Now program the registers
*
* We don't use the save variable to indicate the number of elements passed to
* the kernel because, as we said earlier, pc may contain more elements than
* the number of events we specified, i.e., contains more thann coutning monitors.
*/
if (perfmonctl(pid, PFM_WRITE_PMCS, evt.pfp_pc, evt.pfp_pc_count) == -1) {
fatal_error("perfmonctl error PFM_WRITE_PMCS errno %d\n",errno);
}
if (perfmonctl(pid, PFM_WRITE_PMDS, pd, evt.pfp_event_count) == -1) {
{int i; for(i=0; i < evt.pfp_event_count; i++) printf("pmd%d: 0x%x\n", i, pd[i].reg_flags);}
fatal_error("perfmonctl error PFM_WRITE_PMDS errno %d\n",errno);
}
/*
* Let's roll now
*/
pfm_start();
noploop(10000000);
pfm_stop();
/*
* now read the results
*/
if (perfmonctl(pid, PFM_READ_PMDS, pd, evt.pfp_event_count) == -1) {
fatal_error( "perfmonctl error READ_PMDS errno %d\n",errno);
return -1;
}
/*
* print the results
*
* It is important to realize, that the first event we specified may not
* be in PMD4. Not all events can be measured by any monitor. That's why
* we need to use the pc[] array to figure out where event i was allocated.
*
*/
for (i=0; i < evt.pfp_event_count; i++) {
char *name;
pfm_get_event_name(evt.pfp_events[i].event, &name);
printf("PMD%u %20lu %s\n",
pd[i].reg_num,
pd[i].reg_value,
name);
}
/*
* let's stop this now
*/
if (perfmonctl(pid, PFM_DESTROY_CONTEXT, NULL, 0) == -1) {
fatal_error( "child: perfmonctl error PFM_DESTROY errno %d\n",errno);
}
return 0;
}
int
main(void)
{
int ret;
int type = 0;
pid_t pid = getpid();
pfmlib_ita2_param_t ita_param;
pfarg_reg_t pd[NUM_PMDS];
pfarg_context_t ctx[1];
pfmlib_options_t pfmlib_options;
struct sigaction act;
/*
* Initialize pfm library (required before we can use it)
*/
if (pfm_initialize() != PFMLIB_SUCCESS) {
fatal_error("Can't initialize library\n");
}
/*
* Let's make sure we run this on the right CPU
*/
pfm_get_pmu_type(&type);
if (type != PFMLIB_ITANIUM2_PMU) {
char *model;
pfm_get_pmu_name(&model);
fatal_error("this program does not work with %s PMU\n", model);
}
/*
* Install the overflow handler (SIGPROF)
*/
memset(&act, 0, sizeof(act));
act.sa_handler = (sig_t)overflow_handler;
sigaction (SIGPROF, &act, 0);
/*
* pass options to library (optional)
*/
memset(&pfmlib_options, 0, sizeof(pfmlib_options));
pfmlib_options.pfm_debug = 0; /* set to 1 for debug */
pfmlib_options.pfm_verbose = 0; /* set to 1 for debug */
pfm_set_options(&pfmlib_options);
memset(pd, 0, sizeof(pd));
memset(ctx, 0, sizeof(ctx));
/*
* prepare parameters to library. we don't use any Itanium
* specific features here. so the pfp_model is NULL.
*/
memset(&evt,0, sizeof(evt));
memset(&ita_param,0, sizeof(ita_param));
/*
* because we use a model specific feature, we must initialize the
* model specific pfmlib parameter structure and link it to the
* common structure.
* The magic number is a simple mechanism used by the library to check
* that the model specific data structure is decent. You must set it manually
* otherwise the model specific feature won't work.
*/
ita_param.pfp_magic = PFMLIB_ITA2_PARAM_MAGIC;
evt.pfp_model = &ita_param;
/*
* Before calling pfm_find_dispatch(), we must specify what kind
* of branches we want to capture. We are interesteed in all the mispredicted branches,
* therefore we program we set the various fields of the BTB config to:
*/
ita_param.pfp_ita2_btb.btb_used = 1;
ita_param.pfp_ita2_btb.btb_ds = 0;
ita_param.pfp_ita2_btb.btb_tm = 0x3;
ita_param.pfp_ita2_btb.btb_ptm = 0x3;
ita_param.pfp_ita2_btb.btb_ppm = 0x3;
ita_param.pfp_ita2_btb.btb_brt = 0x0;
ita_param.pfp_ita2_btb.btb_plm = PFM_PLM3;
/*
* To count the number of occurence of this instruction, we must
* program a counting monitor with the IA64_TAGGED_INST_RETIRED_PMC8
* event.
*/
if (pfm_find_event_byname("BRANCH_EVENT", &evt.pfp_events[0].event) != PFMLIB_SUCCESS) {
fatal_error("cannot find event BRANCH_EVENT\n");
}
/*
* set the (global) privilege mode:
* PFM_PLM3 : user level only
*/
evt.pfp_dfl_plm = PFM_PLM3;
/*
* how many counters we use
*/
evt.pfp_event_count = 1;
/*
* let the library figure out the values for the PMCS
*/
if ((ret=pfm_dispatch_events(&evt)) != PFMLIB_SUCCESS) {
fatal_error("cannot configure events: %s\n", pfm_strerror(ret));
}
/*
* for this example, we will get notified ONLY when the sampling
* buffer is full. The monitoring is not to be inherited
* in derived tasks
*/
ctx[0].ctx_flags = PFM_FL_INHERIT_NONE;
ctx[0].ctx_notify_pid = getpid();
ctx[0].ctx_smpl_entries = SMPL_BUF_NENTRIES;
ctx[0].ctx_smpl_regs[0] = smpl_regs = BTB_REGS_MASK;
/*
* now create the context for self monitoring/per-task
*/
if (perfmonctl(pid, 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));
}
printf("Sampling buffer mapped at %p\n", ctx[0].ctx_smpl_vaddr);
smpl_vaddr = ctx[0].ctx_smpl_vaddr;
/*
* Must be done before any PMD/PMD calls (unfreeze PMU). Initialize
* PMC/PMD to safe values. psr.up is cleared.
*/
if (perfmonctl(pid, PFM_ENABLE, NULL, 0) == -1) {
fatal_error("perfmonctl error PFM_ENABLE errno %d\n",errno);
}
/*
* indicate we want notification when buffer is full
*/
evt.pfp_pc[0].reg_flags |= PFM_REGFL_OVFL_NOTIFY;
/*
* Now prepare the argument to initialize the PMD and the sampling period
*/
pd[0].reg_num = evt.pfp_pc[0].reg_num;
pd[0].reg_value = (~0UL) - SMPL_PERIOD +1;
pd[0].reg_long_reset = (~0UL) - SMPL_PERIOD +1;
pd[0].reg_short_reset = (~0UL) - SMPL_PERIOD +1;
/*
* When our counter overflows, we want to BTB index to be reset, so that we keep
* in sync. This is required to make it possible to interpret pmd16 on overflow
* to avoid repeating the same branch several times.
*/
evt.pfp_pc[0].reg_reset_pmds[0] = M_PMD(16);
/*
* reset pmd16, short and long reset value are set to zero as well
*/
pd[1].reg_num = 16;
pd[1].reg_value = 0UL;
/*
* Now program the registers
*
* We don't use the save variable to indicate the number of elements passed to
* the kernel because, as we said earlier, pc may contain more elements than
* the number of events we specified, i.e., contains more thann coutning monitors.
*/
if (perfmonctl(pid, PFM_WRITE_PMCS, evt.pfp_pc, evt.pfp_pc_count) == -1) {
fatal_error("perfmonctl error PFM_WRITE_PMCS errno %d\n",errno);
}
if (perfmonctl(pid, PFM_WRITE_PMDS, pd, 2) == -1) {
fatal_error("perfmonctl error PFM_WRITE_PMDS errno %d\n",errno);
}
/*
* Let's roll now.
*/
do_test(100000);
/*
* We must call the processing routine to cover the last entries recorded
* in the sampling buffer, i.e. which may not be full
*/
process_smpl_buffer();
/*
* let's stop this now
*/
if (perfmonctl(pid, PFM_DESTROY_CONTEXT, NULL, 0) == -1) {
fatal_error("perfmonctl error PFM_DESTROY errno %d\n",errno);
}
return 0;
}
int
main(void)
{
pfmlib_input_param_t inp;
pfmlib_output_param_t outp;
pfmlib_ita2_input_param_t ita2_inp;
pfarg_reg_t pd[NUM_PMDS];
pfarg_reg_t pc[NUM_PMCS];
pfarg_context_t ctx[1];
pfarg_load_t load_args;
pfmlib_options_t pfmlib_options;
int ret;
int type = 0;
int id;
unsigned int i;
char name[MAX_EVT_NAME_LEN];
/*
* Initialize pfm library (required before we can use it)
*/
if (pfm_initialize() != PFMLIB_SUCCESS) {
fatal_error("Can't initialize library\n");
}
/*
* Let's make sure we run this on the right CPU
*/
pfm_get_pmu_type(&type);
if (type != PFMLIB_ITANIUM2_PMU) {
char model[MAX_PMU_NAME_LEN];
pfm_get_pmu_name(model, MAX_PMU_NAME_LEN);
fatal_error("this program does not work with the %s PMU\n", model);
}
/*
* pass options to library (optional)
*/
memset(&pfmlib_options, 0, sizeof(pfmlib_options));
pfmlib_options.pfm_debug = 0; /* set to 1 for debug */
pfmlib_options.pfm_verbose = 0; /* 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));
memset(&inp,0, sizeof(inp));
memset(&outp,0, sizeof(outp));
memset(&ita2_inp,0, sizeof(ita2_inp));
/*
* We indicate that we are using the PMC8 opcode matcher. This is required
* otherwise the library add PMC8 to the list of PMC to pogram during
* pfm_dispatch_events().
*/
ita2_inp.pfp_ita2_pmc8.opcm_used = 1;
/*
* We want to match all the br.cloop in our test function.
* This branch is an IP-relative branch for which the major
* opcode (bits [40-37]=4) and the btype field is 5 (which represents
* bits[6-8]) so it is included in the match/mask fields of PMC8.
* It is necessarily in a B slot.
*
* We don't care which operands are used with br.cloop therefore
* the mask field of pmc8 is set such that only the 4 bits of the
* opcode and 3 bits of btype must match exactly. This is accomplished by
* clearing the top 4 bits and bits [6-8] of the mask field and setting the
* remaining bits. Similarly, the match field only has the opcode value and btype
* set according to the encoding of br.cloop, the
* remaining bits are zero. Bit 60 of PMC8 is set to indicate
* that we look only in B slots (this is the only possibility for
* this instruction anyway).
*
* So the binary representation of the value for PMC8 is as follows:
*
* 6666555555555544444444443333333333222222222211111111110000000000
* 3210987654321098765432109876543210987654321098765432109876543210
* ----------------------------------------------------------------
* 0001010000000000000000101000000000000011111111111111000111111000
*
* which yields a value of 0x1400028003fff1f8.
*
* Depending on the level of optimization to compile this code, it may
* be that the count reported could be zero, if the compiler uses a br.cond
* instead of br.cloop.
*
*
* The 0x1 sets the ig_ad field to make sure we ignore any range restriction.
* Also bit 2 must always be set
*/
ita2_inp.pfp_ita2_pmc8.pmc_val = 0x1400028003fff1fa;
/*
* To count the number of occurence of this instruction, we must
* program a counting monitor with the IA64_TAGGED_INST_RETIRED_PMC8
* event.
*/
if (pfm_find_full_event("IA64_TAGGED_INST_RETIRED_IBRP0_PMC8", &inp.pfp_events[0]) != PFMLIB_SUCCESS) {
fatal_error("cannot find event IA64_TAGGED_INST_RETIRED_IBRP0_PMC8\n");
}
/*
* set the privilege mode:
* PFM_PLM3 : user level only
*/
inp.pfp_dfl_plm = PFM_PLM3;
/*
* how many counters we use
*/
inp.pfp_event_count = 1;
/*
* let the library figure out the values for the PMCS
*/
if ((ret=pfm_dispatch_events(&inp, &ita2_inp, &outp, NULL)) != 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));
}
/*
* extract the unique identifier for our context, a regular file descriptor
*/
id = ctx[0].ctx_fd;
/*
* Now prepare the argument to initialize the PMDs and PMCS.
* We must pfp_pmc_count to determine the number of PMC to intialize.
* We must use pfp_event_count to determine the number of PMD to initialize.
* Some events causes extra PMCs to be used, so pfp_pmc_count may be >= pfp_event_count.
*
* This step is new compared to libpfm-2.x. It is necessary because the library no
* longer knows about the kernel data structures.
*/
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;
}
/*
* the PMC controlling the event ALWAYS come first, that's why this loop
* is safe even when extra PMC are needed to support a particular event.
*/
for (i=0; i < inp.pfp_event_count; i++) {
pd[i].reg_num = pc[i].reg_num;
}
printf("event_count=%d id=%d\n", inp.pfp_event_count, id);
/*
* Now program the registers
*
* We don't use the save variable to indicate the number of elements passed to
* the kernel because, as we said earlier, pc may contain more elements than
* the number of events we specified, i.e., contains more thann coutning monitors.
*/
if (perfmonctl(id, PFM_WRITE_PMCS, pc, outp.pfp_pmc_count) == -1) {
fatal_error("perfmonctl error PFM_WRITE_PMCS errno %d\n",errno);
}
if (perfmonctl(id, PFM_WRITE_PMDS, pd, inp.pfp_event_count) == -1) {
fatal_error("perfmonctl error PFM_WRITE_PMDS errno %d\n",errno);
}
/*
* now we load (i.e., attach) the context to ourself
*/
load_args.load_pid = getpid();
if (perfmonctl(id, PFM_LOAD_CONTEXT, &load_args, 1) == -1) {
fatal_error("perfmonctl error PFM_LOAD_CONTEXT errno %d\n",errno);
}
/*
* Let's roll now.
*/
pfm_self_start(id);
do_test(100UL);
pfm_self_stop(id);
/*
* now read the results
*/
if (perfmonctl(id, PFM_READ_PMDS, pd, inp.pfp_event_count) == -1) {
fatal_error("perfmonctl error READ_PMDS errno %d\n",errno);
}
/*
* print the results
*/
pfm_get_full_event_name(&inp.pfp_events[0], name, MAX_EVT_NAME_LEN);
printf("PMD%u %20lu %s\n",
pd[0].reg_num,
pd[0].reg_value,
name);
if (pd[0].reg_value != 0)
printf("compiler used br.cloop\n");
else
printf("compiler did not use br.cloop\n");
/*
* let's stop this now
*/
close(id);
return 0;
}
int
main(void)
{
int ret;
int type = 0;
char *name;
pid_t pid = getpid();
pfmlib_param_t evt;
pfmlib_ita2_param_t ita2_param;
pfarg_reg_t pd[NUM_PMDS];
pfarg_context_t ctx[1];
pfmlib_options_t pfmlib_options;
/*
* Initialize pfm library (required before we can use it)
*/
if (pfm_initialize() != PFMLIB_SUCCESS) {
fatal_error("Can't initialize library\n");
}
/*
* Let's make sure we run this on the right CPU
*/
pfm_get_pmu_type(&type);
if (type != PFMLIB_ITANIUM2_PMU) {
char *model;
pfm_get_pmu_name(&model);
fatal_error("this program does not work with the %s PMU\n", model);
}
/*
* pass options to library (optional)
*/
memset(&pfmlib_options, 0, sizeof(pfmlib_options));
pfmlib_options.pfm_debug = 0; /* set to 1 for debug */
pfmlib_options.pfm_verbose = 0; /* set to 1 for verbose */
pfm_set_options(&pfmlib_options);
memset(pd, 0, sizeof(pd));
memset(ctx, 0, sizeof(ctx));
memset(&evt,0, sizeof(evt));
memset(&ita2_param,0, sizeof(ita2_param));
/*
* because we use a model specific feature, we must initialize the
* model specific pfmlib parameter structure and link it to the
* common structure.
* The magic number is a simple mechanism used by the library to check
* that the model specific data structure is decent. You must set it manually
* otherwise the model specific feature won't work.
*/
ita2_param.pfp_magic = PFMLIB_ITA2_PARAM_MAGIC;
evt.pfp_model = &ita2_param;
/*
* We indicate that we are using the PMC8 opcode matcher. This is required
* otherwise the library add PMC8 to the list of PMC to pogram during
* pfm_dispatch_events().
*/
ita2_param.pfp_ita2_pmc8.opcm_used = 1;
/*
* We want to match all the br.cloop in our test function.
* This branch is an IP-relative branch for which the major
* opcode (bits [40-37]=4) and the btype field is 5 (which represents
* bits[6-8]) so it is included in the match/mask fields of PMC8.
* It is necessarily in a B slot.
*
* We don't care which operands are used with br.cloop therefore
* the mask field of pmc8 is set such that only the 4 bits of the
* opcode and 3 bits of btype must match exactly. This is accomplished by
* clearing the top 4 bits and bits [6-8] of the mask field and setting the
* remaining bits. Similarly, the match field only has the opcode value and btype
* set according to the encoding of br.cloop, the
* remaining bits are zero. Bit 60 of PMC8 is set to indicate
* that we look only in B slots (this is the only possibility for
* this instruction anyway).
*
* So the binary representation of the value for PMC8 is as follows:
*
* 6666555555555544444444443333333333222222222211111111110000000000
* 3210987654321098765432109876543210987654321098765432109876543210
* ----------------------------------------------------------------
* 0001010000000000000000101000000000000011111111111111000111111000
*
* which yields a value of 0x1400028003fff1f8.
*
* Depending on the level of optimization to compile this code, it may
* be that the count reported could be zero, if the compiler uses a br.cond
* instead of br.cloop.
*
*
* The 0x1 sets the ig_ad field to make sure we ignore any range restriction.
* Also bit 2 must always be set
*/
ita2_param.pfp_ita2_pmc8.pmc_val = 0x1400028003fff1fa;
/*
* To count the number of occurence of this instruction, we must
* program a counting monitor with the IA64_TAGGED_INST_RETIRED_PMC8
* event.
*/
if (pfm_find_event_byname("IA64_TAGGED_INST_RETIRED_IBRP0_PMC8", &evt.pfp_events[0].event) != PFMLIB_SUCCESS) {
fatal_error("cannot find event IA64_TAGGED_INST_RETIRED_IBRP0_PMC8\n");
}
/*
* set the privilege mode:
* PFM_PLM3 : user level only
*/
evt.pfp_dfl_plm = PFM_PLM3;
/*
* how many counters we use
*/
evt.pfp_event_count = 1;
/*
* let the library figure out the values for the PMCS
*/
if ((ret=pfm_dispatch_events(&evt)) != PFMLIB_SUCCESS) {
fatal_error("cannot configure events: %s\n", pfm_strerror(ret));
}
/*
* for this example, we have decided not to get notified
* on counter overflows and the monitoring is not to be inherited
* in derived tasks
*/
ctx[0].ctx_flags = PFM_FL_INHERIT_NONE;
/*
* now create the context for self monitoring/per-task
*/
if (perfmonctl(pid, 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));
}
/*
* Must be done before any PMD/PMD calls (unfreeze PMU). Initialize
* PMC/PMD to safe values. psr.up is cleared.
*/
if (perfmonctl(pid, PFM_ENABLE, NULL, 0) == -1) {
fatal_error("perfmonctl error PFM_ENABLE errno %d\n",errno);
}
/*
* Now prepare the argument to initialize the PMD.
*/
pd[0].reg_num = evt.pfp_pc[0].reg_num;
/*
* Now program the registers
*
* We don't use the save variable to indicate the number of elements passed to
* the kernel because, as we said earlier, pc may contain more elements than
* the number of events we specified, i.e., contains more thann coutning monitors.
*/
if (perfmonctl(pid, PFM_WRITE_PMCS, evt.pfp_pc, evt.pfp_pc_count) == -1) {
fatal_error("perfmonctl error PFM_WRITE_PMCS errno %d\n",errno);
}
if (perfmonctl(pid, PFM_WRITE_PMDS, pd, evt.pfp_event_count) == -1) {
fatal_error("perfmonctl error PFM_WRITE_PMDS errno %d\n",errno);
}
/*
* Let's roll now.
*/
pfm_start();
do_test(100UL);
pfm_stop();
/*
* now read the results
*/
if (perfmonctl(pid, PFM_READ_PMDS, pd, evt.pfp_event_count) == -1) {
fatal_error("perfmonctl error READ_PMDS errno %d\n",errno);
}
/*
* print the results
*/
pfm_get_event_name(evt.pfp_events[0].event, &name);
printf("PMD%u %20lu %s\n",
pd[0].reg_num,
pd[0].reg_value,
name);
if (pd[0].reg_value != 0)
printf("compiler used br.cloop\n");
else
printf("compiler did not use br.cloop\n");
/*
* let's stop this now
*/
if (perfmonctl(pid, PFM_DESTROY_CONTEXT, NULL, 0) == -1) {
fatal_error("perfmonctl error PFM_DESTROY errno %d\n",errno);
}
return 0;
}
int
main(int argc, char **argv)
{
pfarg_context_t ctx[1];
pfmlib_input_param_t inp;
pfmlib_output_param_t outp;
pfarg_reg_t pc[NUM_PMCS];
pfarg_load_t load_args;
pfmlib_options_t pfmlib_options;
struct sigaction act;
size_t len;
unsigned int i, num_counters;
int ret;
/*
* Initialize pfm library (required before we can use it)
*/
if (pfm_initialize() != PFMLIB_SUCCESS) {
printf("Can't initialize library\n");
exit(1);
}
/*
* Install the signal handler (SIGIO)
*/
memset(&act, 0, sizeof(act));
act.sa_handler = (sig_t)sigio_handler;
sigaction (SIGIO, &act, 0);
/*
* pass options to library (optional)
*/
memset(&pfmlib_options, 0, sizeof(pfmlib_options));
pfmlib_options.pfm_debug = 0; /* set to 1 for debug */
pfm_set_options(&pfmlib_options);
memset(pc, 0, sizeof(pc));
memset(ctx, 0, sizeof(ctx));
memset(&load_args, 0, sizeof(load_args));
memset(&inp,0, sizeof(inp));
memset(&outp,0, sizeof(outp));
pfm_get_num_counters(&num_counters);
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;
if (i > num_counters) {
i = num_counters;
printf("too many events provided (max=%d events), using first %d event(s)\n", num_counters, i);
}
/*
* set the default privilege mode for all counters:
* PFM_PLM3 : user level only
*/
inp.pfp_dfl_plm = PFM_PLM3;
/*
* how many counters we use
*/
inp.pfp_event_count = i;
/*
* how many counters we use
*/
if (i > 1) {
inp.pfp_event_count = i;
pfm_get_max_event_name_len(&len);
event1_name = malloc(len+1);
if (event1_name == NULL)
fatal_error("cannot allocate event name\n");
pfm_get_full_event_name(&inp.pfp_events[1], event1_name, len+1);
}
/*
* 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));
}
/*
* when we know we are self-monitoring and we have only one context, then
* when we get an overflow we know where it is coming from. Therefore we can
* save the call to the kernel to extract the notification message. By default,
* a message is generated. The queue of messages has a limited size, therefore
* it is important to clear the queue by reading the message on overflow. Failure
* to do so may result in a queue full and you will lose notification messages.
*
* With the PFM_FL_OVFL_NO_MSG, no message will be queue, but you will still get
* the signal. Similarly, the PFM_MSG_END will be generated.
*/
ctx[0].ctx_flags = PFM_FL_OVFL_NO_MSG;
/*
* 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_fd;
/*
* Now prepare the argument to initialize the PMDs and PMCS.
* We use pfp_pmc_count to determine the number of registers to
* setup. Note that this field can 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 < inp.pfp_event_count; i++) {
pd[i].reg_num = pc[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_pmcs[1].reg_num;
/*
* we arm the first counter, such that it will overflow
* after SMPL_PERIOD events have been observed
*/
pd[0].reg_value = (~0UL) - SMPL_PERIOD + 1;
pd[0].reg_long_reset = (~0UL) - SMPL_PERIOD + 1;
pd[0].reg_short_reset = (~0UL) - SMPL_PERIOD + 1;
/*
* Now program the registers
*
* We don't use the save variable to indicate the number of elements passed to
* the kernel because, as we said earlier, pc may contain more elements than
* the number of events we specified, i.e., contains more than counting monitors.
*/
if (perfmonctl(ctx_fd, PFM_WRITE_PMCS, pc, outp.pfp_pmc_count) == -1) {
fatal_error("perfmonctl error PFM_WRITE_PMCS errno %d\n",errno);
}
if (perfmonctl(ctx_fd, PFM_WRITE_PMDS, pd, inp.pfp_event_count) == -1) {
fatal_error("perfmonctl error PFM_WRITE_PMDS errno %d\n",errno);
}
/*
* we want to monitor ourself
*/
load_args.load_pid = getpid();
if (perfmonctl(ctx_fd, PFM_LOAD_CONTEXT, &load_args, 1) == -1) {
fatal_error("perfmonctl error PFM_WRITE_PMDS 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);
if (ret == -1) {
fatal_error("cannot set ASYNC: %s\n", strerror(errno));
}
/*
* get ownership of the descriptor
*/
ret = fcntl(ctx_fd, F_SETOWN, getpid());
if (ret == -1) {
fatal_error("cannot setown: %s\n", strerror(errno));
}
/*
* Let's roll now
*/
pfm_self_start(ctx_fd);
busyloop();
pfm_self_stop(ctx_fd);
/*
* free our context
*/
close(ctx_fd);
return 0;
}
int
main(int argc, char **argv)
{
pfmlib_input_param_t inp;
pfmlib_output_param_t outp;
pfmlib_core_input_param_t mod_inp;
pfmlib_options_t pfmlib_options;
pfarg_pmr_t pc[NUM_PMCS];
pfarg_pmd_attr_t pd[NUM_PMDS];
pfarg_sinfo_t sif;
struct pollfd fds;
smpl_arg_t buf_arg;
pfarg_msg_t msg;
smpl_hdr_t *hdr;
void *buf_addr;
uint64_t pebs_size;
pid_t pid;
int ret, fd, type;
unsigned int i;
uint32_t ctx_flags;
if (argc < 2)
fatal_error("you need to pass a program to sample\n");
if (pfm_initialize() != PFMLIB_SUCCESS)
fatal_error("libpfm intialization failed\n");
/*
* check we are on an Intel Core PMU
*/
pfm_get_pmu_type(&type);
if (type != PFMLIB_INTEL_CORE_PMU && type != PFMLIB_INTEL_ATOM_PMU)
fatal_error("This program only works with an Intel Core processor\n");
/*
* pass options to library (optional)
*/
memset(&pfmlib_options, 0, sizeof(pfmlib_options));
pfmlib_options.pfm_debug = 0; /* 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(&inp, 0, sizeof(inp));
memset(&outp, 0, sizeof(outp));
memset(&mod_inp, 0, sizeof(mod_inp));
memset(&sif, 0, sizeof(sif));
memset(&buf_arg, 0, sizeof(buf_arg));
memset(&fds, 0, sizeof(fds));
/*
* search for our sampling event
*/
if (pfm_find_full_event(SMPL_EVENT, &inp.pfp_events[0]) != PFMLIB_SUCCESS)
fatal_error("cannot find sampling event %s\n", SMPL_EVENT);
inp.pfp_event_count = 1;
inp.pfp_dfl_plm = PFM_PLM3;
/*
* important: inform libpfm we do use PEBS
*/
mod_inp.pfp_core_pebs.pebs_used = 1;
/*
* sampling buffer parameters
*/
pebs_size = 3 * getpagesize();
buf_arg.buf_size = pebs_size;
/*
* sampling period cannot use more bits than HW counter can supoprt
*/
buf_arg.cnt_reset = -SMPL_PERIOD;
/*
* We want a system-wide context for sampling
*/
ctx_flags = PFM_FL_SYSTEM_WIDE | PFM_FL_SMPL_FMT;
/*
* trigger notification (interrupt) when reaching the very end of
* the buffer
*/
buf_arg.intr_thres = (pebs_size/sizeof(smpl_entry_t))*90/100;
/*
* we want to measure CPU0, thus we pin ourself to the CPU before invoking
* perfmon. This ensures that the sampling buffer will be allocated on the
* same NUMA node.
*/
ret = pin_cpu(getpid(), 0);
if (ret)
fatal_error("cannot pin on CPU0");
/*
* create session and sampling buffer
*/
fd = pfm_create(ctx_flags, &sif, FMT_NAME, &buf_arg, sizeof(buf_arg));
if (fd == -1) {
if (errno == ENOSYS) {
fatal_error("Your kernel does not have performance monitoring support!\n");
}
fatal_error("cannot create session %s, maybe you do not have the PEBS sampling format in the kernel.\nCheck /sys/kernel/perfmon/formats\n", strerror(errno));
}
/*
* map buffer into our address space
*/
buf_addr = mmap(NULL, (size_t)buf_arg.buf_size, PROT_READ, MAP_PRIVATE, fd, 0);
printf("session [%d] buffer mapped @%p\n", fd, buf_addr);
if (buf_addr == MAP_FAILED)
fatal_error("cannot mmap sampling buffer errno %d\n", errno);
hdr = (smpl_hdr_t *)buf_addr;
printf("pebs_base=0x%llx pebs_end=0x%llx index=0x%llx\n"
"intr=0x%llx version=%u.%u\n"
"entry_size=%zu ds_size=%zu\n",
(unsigned long long)hdr->ds.pebs_buf_base,
(unsigned long long)hdr->ds.pebs_abs_max,
(unsigned long long)hdr->ds.pebs_index,
(unsigned long long)hdr->ds.pebs_intr_thres,
PFM_VERSION_MAJOR(hdr->version),
PFM_VERSION_MINOR(hdr->version),
sizeof(smpl_entry_t),
sizeof(hdr->ds));
if (PFM_VERSION_MAJOR(hdr->version) < 1)
fatal_error("invalid buffer format version\n");
/*
* get which PMC registers are available
*/
detect_unavail_pmu_regs(&sif, &inp.pfp_unavail_pmcs, NULL);
/*
* let libpfm figure out how to assign event onto PMU registers
*/
if (pfm_dispatch_events(&inp, &mod_inp, &outp, NULL) != PFMLIB_SUCCESS)
fatal_error("cannot assign event %s\n", SMPL_EVENT);
/*
* propagate PMC setup from libpfm to perfmon
*/
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;
/*
* must disable 64-bit emulation on the PMC0 counter.
* PMC0 is the only counter useable with PEBS. We must disable
* 64-bit emulation to avoid getting interrupts for each
* sampling period, PEBS takes care of this part.
*/
if (pc[i].reg_num == 0)
pc[i].reg_flags = PFM_REGFL_NO_EMUL64;
}
/*
* propagate PMD set from libpfm to perfmon
*/
for (i=0; i < outp.pfp_pmd_count; i++)
pd[i].reg_num = outp.pfp_pmds[i].reg_num;
/*
* setup sampling period for first counter
* we want notification on overflow, i.e., when buffer is full
*/
pd[0].reg_flags = PFM_REGFL_OVFL_NOTIFY;
pd[0].reg_value = -SMPL_PERIOD;
pd[0].reg_long_reset = -SMPL_PERIOD;
pd[0].reg_short_reset = -SMPL_PERIOD;
/*
* Now program the registers
*/
if (pfm_write(fd, 0, PFM_RW_PMC, pc, outp.pfp_pmc_count * sizeof(*pc)) == -1)
fatal_error("pfm_write error errno %d\n",errno);
if (pfm_write(fd, 0, PFM_RW_PMD_ATTR, pd, outp.pfp_pmd_count * sizeof(*pd)) == -1)
fatal_error("pfm_write(PMD) error errno %d\n",errno);
/*
* attach the session to CPU0
*/
if (pfm_attach(fd, 0, 0) == -1)
fatal_error("pfm_attach error errno %d\n",errno);
/*
* Create the child task
*/
signal(SIGCHLD, handler);
if ((pid=fork()) == -1)
fatal_error("Cannot fork process\n");
if (pid == 0) {
/* child does not inherit context file descriptor */
close(fd);
/* if child is too short-lived we may not measure it */
child(argv+1);
}
/*
* start monitoring
*/
if (pfm_set_state(fd, 0, PFM_ST_START) == -1)
fatal_error("pfm_set_state(start) error errno %d\n",errno);
fds.fd = fd;
fds.events = POLLIN;
/*
* core loop
*/
for(;done == 0;) {
/*
* Must use a timeout to avoid a race condition
* with the SIGCHLD signal
*/
ret = poll(&fds, 1, 500);
/*
* if timeout expired, then check done
*/
if (ret == 0)
continue;
if (ret == -1) {
if(ret == -1 && errno == EINTR) {
warning("read interrupted, retrying\n");
continue;
}
fatal_error("poll failed: %s\n", strerror(errno));
}
ret = read(fd, &msg, sizeof(msg));
if (ret == -1)
fatal_error("cannot read perfmon msg: %s\n", strerror(errno));
switch(msg.type) {
case PFM_MSG_OVFL: /* the sampling buffer is full */
process_smpl_buf(hdr);
/*
* reactivate monitoring once we are done with the samples
* in syste-wide, interface guarantees monitoring is active
* upon return from the pfm_restart() syscall
*/
if (pfm_set_state(fd, 0, PFM_ST_RESTART) == -1)
fatal_error("pfm_set_state(restart) error errno %d\n",errno);
break;
default: fatal_error("unknown message type %d\n", msg.type);
}
}
/*
* cleanup child
*/
waitpid(pid, NULL, 0);
/*
* stop monitoring, this is required in order to guarantee that the PEBS buffer
* header is updated with the latest position, such that we see see the final
* samples
*/
if (pfm_set_state(fd, 0, PFM_ST_STOP) == -1)
fatal_error("pfm_set_state(stop) error errno %d\n",errno);
/*
* check for any leftover samples. Must have monitoring stopped
* for this operation to have guarantee it is up to date
*/
process_smpl_buf(hdr);
/*
* close session
*/
close(fd);
/*
* unmap sampling buffer and actually free the perfmon session
*/
munmap(buf_addr, (size_t)buf_arg.buf_size);
return 0;
}
