/* * pfmlib_perf_events.c: encode events for perf_event API * * Copyright (c) 2009 Google, Inc * 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. */ #include #include #include #include #include #include #include "pfmlib_priv.h" #include "pfmlib_perf_event_priv.h" #define PERF_PROC_FILE "/proc/sys/kernel/perf_event_paranoid" #ifdef min #undef min #endif #define min(a, b) ((a) < (b) ? (a) : (b)) /* * contains ONLY attributes related to PMU features */ static const pfmlib_attr_desc_t perf_event_mods[]={ PFM_ATTR_B("u", "monitor at user level"), /* monitor user level */ PFM_ATTR_B("k", "monitor at kernel level"), /* monitor kernel level */ PFM_ATTR_B("h", "monitor at hypervisor level"), /* monitor hypervisor level */ PFM_ATTR_SKIP, PFM_ATTR_SKIP, PFM_ATTR_SKIP, PFM_ATTR_SKIP, PFM_ATTR_B("mg", "monitor guest execution"), /* monitor guest level */ PFM_ATTR_B("mh", "monitor host execution"), /* monitor host level */ PFM_ATTR_NULL /* end-marker to avoid exporting number of entries */ }; /* * contains all attributes controlled by perf_events. That includes PMU attributes * and pure software attributes such as sampling periods */ static const pfmlib_attr_desc_t perf_event_ext_mods[]={ PFM_ATTR_B("u", "monitor at user level"), /* monitor user level */ PFM_ATTR_B("k", "monitor at kernel level"), /* monitor kernel level */ PFM_ATTR_B("h", "monitor at hypervisor level"), /* monitor hypervisor level */ PFM_ATTR_I("period", "sampling period"), /* sampling period */ PFM_ATTR_I("freq", "sampling frequency (Hz)"), /* sampling frequency */ PFM_ATTR_I("precise", "precise ip"), /* anti-skid mechanism */ PFM_ATTR_B("excl", "exclusive access"), /* exclusive PMU access */ PFM_ATTR_B("mg", "monitor guest execution"), /* monitor guest level */ PFM_ATTR_B("mh", "monitor host execution"), /* monitor host level */ PFM_ATTR_I("cpu", "CPU to program"), /* CPU to program */ PFM_ATTR_B("pinned", "pin event to counters"), /* pin event to PMU */ PFM_ATTR_NULL /* end-marker to avoid exporting number of entries */ }; static int pfmlib_perf_event_encode(void *this, const char *str, int dfl_plm, void *data) { pfm_perf_encode_arg_t arg; pfm_perf_encode_arg_t *uarg = data; pfmlib_os_t *os = this; struct perf_event_attr my_attr, *attr; pfmlib_pmu_t *pmu; pfmlib_event_desc_t e; pfmlib_event_attr_info_t *a; size_t orig_sz, asz, sz = sizeof(arg); uint64_t ival; int has_plm = 0, has_vmx_plm = 0; int i, plm = 0, ret, vmx_plm = 0; int cpu = -1, pinned = 0; sz = pfmlib_check_struct(uarg, uarg->size, PFM_PERF_ENCODE_ABI0, sz); if (!sz) return PFM_ERR_INVAL; /* copy input */ memcpy(&arg, uarg, sz); /* pointer to our internal attr struct */ memset(&my_attr, 0, sizeof(my_attr)); attr = &my_attr; /* * copy user attr to our internal version * size == 0 is interpreted minimal possible * size (ABI_VER0) */ /* size of attr struct passed by user */ orig_sz = uarg->attr->size; if (orig_sz == 0) asz = PERF_ATTR_SIZE_VER0; else asz = min(sizeof(*attr), orig_sz); /* * we copy the user struct to preserve whatever may * have been initialized but that we do not use */ memcpy(attr, uarg->attr, asz); /* restore internal size (just in case we need it) */ attr->size = sizeof(my_attr); /* useful for debugging */ if (asz != sizeof(*attr)) __pfm_vbprintf("warning: mismatch attr struct size " "user=%d libpfm=%zu\n", asz, sizeof(*attr)); memset(&e, 0, sizeof(e)); e.osid = os->id; e.os_data = attr; e.dfl_plm = dfl_plm; /* after this call, need to call pfmlib_release_event() */ ret = pfmlib_parse_event(str, &e); if (ret != PFM_SUCCESS) return ret; pmu = e.pmu; ret = PFM_ERR_NOTSUPP; if (!pmu->get_event_encoding[e.osid]) { DPRINT("PMU %s does not support PFM_OS_NONE\n", pmu->name); goto done; } ret = pmu->get_event_encoding[e.osid](pmu, &e); if (ret != PFM_SUCCESS) goto done; /* * process perf_event attributes */ for (i = 0; i < e.nattrs; i++) { a = attr(&e, i); if (a->ctrl != PFM_ATTR_CTRL_PERF_EVENT) continue; ival = e.attrs[i].ival; switch(a->idx) { case PERF_ATTR_U: if (ival) plm |= PFM_PLM3; has_plm = 1; break; case PERF_ATTR_K: if (ival) plm |= PFM_PLM0; has_plm = 1; break; case PERF_ATTR_H: if (ival) plm |= PFM_PLMH; has_plm = 1; break; case PERF_ATTR_PE: if (!ival || attr->freq) { ret = PFM_ERR_ATTR_VAL; goto done; } attr->sample_period = ival; break; case PERF_ATTR_FR: if (!ival || attr->sample_period) { ret = PFM_ERR_ATTR_VAL; goto done; } attr->sample_freq = ival; attr->freq = 1; break; case PERF_ATTR_PR: if (ival > 3) { ret = PFM_ERR_ATTR_VAL; goto done; } attr->precise_ip = ival; break; case PERF_ATTR_EX: if (ival && !attr->exclusive) attr->exclusive = 1; break; case PERF_ATTR_MG: vmx_plm |= PFM_PLM3; has_vmx_plm = 1; break; case PERF_ATTR_MH: vmx_plm |= PFM_PLM0; has_vmx_plm = 1; break; case PERF_ATTR_CPU: if (ival >= INT_MAX) { ret = PFM_ERR_ATTR_VAL; goto done; } cpu = (int)ival; break; case PERF_ATTR_PIN: pinned = (int)!!ival; break; } } /* * if no priv level mask was provided * with the event, then use dfl_plm */ if (!has_plm) plm = dfl_plm; /* exclude_guest by default */ if (!has_vmx_plm) vmx_plm = PFM_PLM0; /* * perf_event plm work by exclusion, so use logical or * goal here is to set to zero any exclude_* not supported * by underlying PMU */ plm |= (~pmu->supported_plm) & PFM_PLM_ALL; vmx_plm |= (~pmu->supported_plm) & PFM_PLM_ALL; attr->exclude_user = !(plm & PFM_PLM3); attr->exclude_kernel = !(plm & PFM_PLM0); attr->exclude_hv = !(plm & PFM_PLMH); attr->exclude_guest = !(vmx_plm & PFM_PLM3); attr->exclude_host = !(vmx_plm & PFM_PLM0); attr->pinned = pinned; __pfm_vbprintf("PERF[type=%x config=0x%"PRIx64" config1=0x%"PRIx64 " excl=%d e_u=%d e_k=%d e_hv=%d e_host=%d e_gu=%d period=%"PRIu64" freq=%d" " precise=%d pinned=%d] %s\n", attr->type, attr->config, attr->config1, attr->exclusive, attr->exclude_user, attr->exclude_kernel, attr->exclude_hv, attr->exclude_host, attr->exclude_guest, attr->sample_period, attr->freq, attr->precise_ip, attr->pinned, str); /* * propagate event index if necessary */ arg.idx = pfmlib_pidx2idx(e.pmu, e.event); /* propagate cpu */ arg.cpu = cpu; /* propagate our changes, that overwrites attr->size */ memcpy(uarg->attr, attr, asz); /* restore user size */ uarg->attr->size = orig_sz; /* * fstr not requested, stop here */ ret = PFM_SUCCESS; if (!arg.fstr) { memcpy(uarg, &arg, sz); goto done; } for (i=0; i < e.npattrs; i++) { int idx; if (e.pattrs[i].ctrl != PFM_ATTR_CTRL_PERF_EVENT) continue; idx = e.pattrs[i].idx; switch (idx) { case PERF_ATTR_K: evt_strcat(e.fstr, ":%s=%lu", perf_event_ext_mods[idx].name, !!(plm & PFM_PLM0)); break; case PERF_ATTR_U: evt_strcat(e.fstr, ":%s=%lu", perf_event_ext_mods[idx].name, !!(plm & PFM_PLM3)); break; case PERF_ATTR_H: evt_strcat(e.fstr, ":%s=%lu", perf_event_ext_mods[idx].name, !!(plm & PFM_PLMH)); break; case PERF_ATTR_PR: evt_strcat(e.fstr, ":%s=%d", perf_event_ext_mods[idx].name, attr->precise_ip); break; case PERF_ATTR_PE: case PERF_ATTR_FR: if (attr->freq && attr->sample_period) evt_strcat(e.fstr, ":%s=%"PRIu64, perf_event_ext_mods[idx].name, attr->sample_period); else if (attr->sample_period) evt_strcat(e.fstr, ":%s=%"PRIu64, perf_event_ext_mods[idx].name, attr->sample_period); break; case PERF_ATTR_MG: evt_strcat(e.fstr, ":%s=%lu", perf_event_ext_mods[idx].name, !attr->exclude_guest); break; case PERF_ATTR_MH: evt_strcat(e.fstr, ":%s=%lu", perf_event_ext_mods[idx].name, !attr->exclude_host); break; case PERF_ATTR_EX: evt_strcat(e.fstr, ":%s=%lu", perf_event_ext_mods[idx].name, attr->exclusive); break; } } ret = pfmlib_build_fstr(&e, arg.fstr); if (ret == PFM_SUCCESS) memcpy(uarg, &arg, sz); done: pfmlib_release_event(&e); return ret; } /* * get OS-specific event attributes */ static int perf_get_os_nattrs(void *this, pfmlib_event_desc_t *e) { pfmlib_os_t *os = this; int i, n = 0; for (i = 0; os->atdesc[i].name; i++) if (!is_empty_attr(os->atdesc+i)) n++; return n; } static int perf_get_os_attr_info(void *this, pfmlib_event_desc_t *e) { pfmlib_os_t *os = this; pfmlib_event_attr_info_t *info; int i, k, j = e->npattrs; for (i = k = 0; os->atdesc[i].name; i++) { /* skip padding entries */ if (is_empty_attr(os->atdesc+i)) continue; info = e->pattrs + j + k; info->name = os->atdesc[i].name; info->desc = os->atdesc[i].desc; info->equiv= NULL; info->code = i; info->idx = i; /* namespace-specific index */ info->type = os->atdesc[i].type; info->is_dfl = 0; info->ctrl = PFM_ATTR_CTRL_PERF_EVENT; k++; } e->npattrs += k; return PFM_SUCCESS; } /* * old interface, maintained for backward compatibility with earlier versions of the library */ int pfm_get_perf_event_encoding(const char *str, int dfl_plm, struct perf_event_attr *attr, char **fstr, int *idx) { pfm_perf_encode_arg_t arg; int ret; if (PFMLIB_INITIALIZED() == 0) return PFM_ERR_NOINIT; /* idx and fstr can be NULL */ if (!(attr && str)) return PFM_ERR_INVAL; if (dfl_plm & ~(PFM_PLM_ALL)) return PFM_ERR_INVAL; memset(&arg, 0, sizeof(arg)); /* do not clear attr, some fields may be initialized by caller already, e.g., size */ arg.attr = attr; arg.fstr = fstr; ret = pfm_get_os_event_encoding(str, dfl_plm, PFM_OS_PERF_EVENT_EXT, &arg); if (ret != PFM_SUCCESS) return ret; if (idx) *idx = arg.idx; return PFM_SUCCESS; } static int pfm_perf_event_os_detect(void *this) { int ret = access(PERF_PROC_FILE, F_OK); return ret ? PFM_ERR_NOTSUPP : PFM_SUCCESS; } pfmlib_os_t pfmlib_os_perf={ .name = "perf_event", .id = PFM_OS_PERF_EVENT, .atdesc = perf_event_mods, .detect = pfm_perf_event_os_detect, .get_os_attr_info = perf_get_os_attr_info, .get_os_nattrs = perf_get_os_nattrs, .encode = pfmlib_perf_event_encode, }; pfmlib_os_t pfmlib_os_perf_ext={ .name = "perf_event extended", .id = PFM_OS_PERF_EVENT_EXT, .atdesc = perf_event_ext_mods, .detect = pfm_perf_event_os_detect, .get_os_attr_info = perf_get_os_attr_info, .get_os_nattrs = perf_get_os_nattrs, .encode = pfmlib_perf_event_encode, };