/* * ita_irr.c - example of how to use code range restriction with the Itanium PMU * * Copyright (c) 2002-2006 Hewlett-Packard Development Company, L.P. * Contributed by Stephane Eranian * * 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/ia64. */ #include #include #include #include #include #include #include #include #include #include #define VECTOR_SIZE 1000000UL #define NUM_PMCS PFMLIB_MAX_PMCS #define NUM_PMDS PFMLIB_MAX_PMDS #define MAX_EVT_NAME_LEN 128 #define MAX_PMU_NAME_LEN 32 typedef struct { char *event_name; unsigned long expected_value; } event_desc_t; static event_desc_t event_list[]={ { "fp_ops_retired_hi", 0UL} , { "fp_ops_retired_lo", VECTOR_SIZE<<1 }, { NULL, 0UL } }; 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); } void saxpy(double *a, double *b, double *c, unsigned long size) { unsigned long i; for(i=0; i < size; i++) { c[i] = 2*a[i] + b[i]; } } void saxpy2(double *a, double *b, double *c, unsigned long size) { unsigned long i; for(i=0; i < size; i++) { c[i] = 2*a[i] + b[i]; } } static int do_test(void) { unsigned long size; double *a, *b, *c; size = VECTOR_SIZE; a = malloc(size*sizeof(double)); b = malloc(size*sizeof(double)); c = malloc(size*sizeof(double)); if (a == NULL || b == NULL || c == NULL) fatal_error("Cannot allocate vectors\n"); memset(a, 0, size*sizeof(double)); memset(b, 0, size*sizeof(double)); memset(c, 0, size*sizeof(double)); saxpy(a,b,c, size); saxpy2(a,b,c, size); return 0; } int main(int argc, char **argv) { event_desc_t *p; pfmlib_input_param_t inp; pfmlib_output_param_t outp; pfmlib_ita_input_param_t ita_inp; pfmlib_ita_output_param_t ita_outp; pfarg_pmr_t pd[NUM_PMDS]; pfarg_pmr_t pc[NUM_PMCS]; pfarg_pmr_t ibrs[8]; unsigned long range_start, range_end; pfmlib_options_t pfmlib_options; struct fd { /* function descriptor */ unsigned long addr; unsigned long gp; } *fd; int ret, type = 0; unsigned int i; int id; char name[MAX_EVT_NAME_LEN]; /* * 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)); /* * Let's make sure we run this on the right CPU family */ pfm_get_pmu_type(&type); if (type != PFMLIB_ITANIUM_PMU) { char model[MAX_PMU_NAME_LEN]; pfm_get_pmu_name(model, MAX_PMU_NAME_LEN); fatal_error("this program does not work with %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 = 1; /* set to 1 for verbose */ pfm_set_options(&pfmlib_options); /* * Compute the range we are interested in * * On IA-64, the function pointer does not point directly * to the function but to a descriptor which contains two * unsigned long: the first one is the actual start address * of the function, the second is the gp (global pointer) * to load into r1 before jumping into the function. Unlesss * we're jumping into a shared library the gp is the same as * the current gp. * * In the artificial example, we also rely on the compiler/linker * NOT reordering code layout. We depend on saxpy2() being just * after saxpy(). * */ fd = (struct fd *)saxpy; range_start = fd->addr; fd = (struct fd *)saxpy2; range_end = fd->addr; /* * linker may reorder saxpy() and saxpy2() */ if (range_end < range_start) { unsigned long tmp; tmp = range_start; range_start = range_end; range_end = tmp; } memset(pc, 0, sizeof(pc)); memset(pd, 0, sizeof(pd)); memset(ibrs,0, sizeof(ibrs)); memset(&inp,0, sizeof(inp)); memset(&outp,0, sizeof(outp)); memset(&ita_inp,0, sizeof(ita_inp)); memset(&ita_outp,0, sizeof(ita_outp)); /* * find requested event */ p = event_list; for (i=0; p->event_name ; i++, p++) { if (pfm_find_event(p->event_name, &inp.pfp_events[i].event) != PFMLIB_SUCCESS) fatal_error("Cannot find %s event\n", p->event_name); } /* * set the privilege mode: * PFM_PLM3 : user level only */ inp.pfp_dfl_plm = PFM_PLM3; /* * how many counters we use */ inp.pfp_event_count = i; /* * We use the library to figure out how to program the debug registers * to cover the data range we are interested in. The rr_end parameter * must point to the byte after the last element of the range (C-style range). * * Because of the masking mechanism and therefore alignment constraints used to implement * this feature, it may not be possible to exactly cover a given range. It may be that * the coverage exceeds the desired range. So it is possible to capture noise if * the surrounding addresses are also heavily used. You can figure out by how much the * actual range is off compared to the requested range by checking the rr_soff and rr_eoff * fields of rr_infos on return from the library call. * * Upon return, the rr_dbr array is programmed and the number of debug registers (not pairs) * used to cover the range is in rr_nbr_used. * */ ita_inp.pfp_ita_irange.rr_used = 1; /* indicate we use code range restriction */ ita_inp.pfp_ita_irange.rr_limits[0].rr_start = range_start; ita_inp.pfp_ita_irange.rr_limits[0].rr_end = range_end; /* * let the library figure out the values for the PMCS */ if ((ret=pfm_dispatch_events(&inp, &ita_inp, &outp, &ita_outp)) != PFMLIB_SUCCESS) fatal_error("cannot configure events: %s\n", pfm_strerror(ret)); /* * print offsets */ printf("code range : [0x%016lx-0x%016lx)\n" "start_offset:-0x%lx end_offset:+0x%lx\n" "%d pairs of debug registers used\n", range_start, range_end, ita_outp.pfp_ita_irange.rr_infos[0].rr_soff, ita_outp.pfp_ita_irange.rr_infos[0].rr_eoff, ita_outp.pfp_ita_irange.rr_nbr_used >> 1); /* * now create the session */ id = pfm_create(0, NULL); if (id == -1) { if (errno == ENOSYS) fatal_error("Your kernel does not have performance monitoring support!\n"); fatal_error("cannot create session %s\n", strerror(errno)); } /* * 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 cause 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; } for (i=0; i < outp.pfp_pmd_count; i++) { pd[i].reg_num = outp.pfp_pmds[i].reg_num; } /* * propagate the setup for the debug registers from the library to the arguments * to the syscall. The library does not know the type of the syscall * anymore. IBRs are mapped to PMC256-PMC263 */ for (i=0; i < ita_outp.pfp_ita_drange.rr_nbr_used; i++) { ibrs[i].reg_num = 256+ita_outp.pfp_ita_irange.rr_br[i].reg_num; ibrs[i].reg_value = ita_outp.pfp_ita_irange.rr_br[i].reg_value; } /* * Program the code debug registers. * * IMPORTANT: programming the debug register MUST always be done before the PMCs * otherwise the kernel will fail on PFM_WRITE_PMCS. This is for security reasons. */ if (pfm_write(id, 0, PFM_RW_PMC, ibrs, ita_outp.pfp_ita_irange.rr_nbr_used * sizeof(*ibrs)) == -1) fatal_error("pfm_write error errno %d\n",errno); /* * 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 coutning monitors. */ if (pfm_write(id, 0, PFM_RW_PMC, pc, outp.pfp_pmc_count * sizeof(*pc)) == -1) fatal_error("pfm_write error errno %d\n",errno); if (pfm_write(id, 0, PFM_RW_PMD, pd, outp.pfp_pmd_count * sizeof(*pd)) == -1) fatal_error("pfm_write(PMD) error errno %d\n",errno); /* * now attach session */ if (pfm_attach(id, 0, getpid()) == -1) fatal_error("pfm_attach error errno %d\n",errno); /* * Let's roll now. * * We run two distinct copies of the same function but we restrict measurement * to the first one (saxpy). Therefore the expected count is half what you would * get if code range restriction was not used. The core loop in both case uses * two floating point operation per iteration. */ if (pfm_set_state(id, 0, PFM_ST_START)) fatal_error("pfm_set_state error errno %d\n",errno); do_test(); if (pfm_set_state(id, 0, PFM_ST_STOP)) fatal_error("pfm_set_state error errno %d\n",errno); /* * now read the results */ if (pfm_read(id, 0, PFM_RW_PMD, pd, inp.pfp_event_count * sizeof(*pd)) == -1) fatal_error("pfm_read error errno %d\n",errno); /* * 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 < inp.pfp_event_count; i++) { pfm_get_full_event_name(&inp.pfp_events[i], name, MAX_EVT_NAME_LEN); printf("PMD%-3u %20lu %s (expected %lu)\n", pd[i].reg_num, pd[i].reg_value, name, event_list[i].expected_value); } /* * let's stop this now */ close(id); return 0; }