tools/perf/arch/x86/tests/intel-cqm.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 #include "tests/tests.h"
   3 #include "cloexec.h"
   4 #include "debug.h"
   5 #include "evlist.h"
   6 #include "evsel.h"
   7 #include "arch-tests.h"
   8 #include <internal/lib.h> // page_size
   9
  10 #include <signal.h>
  11 #include <sys/mman.h>
  12 #include <sys/wait.h>
  13 #include <errno.h>
  14 #include <string.h>
  15
  16 static pid_t spawn(void)
  17 {
  18         pid_t pid;
  19
  20         pid = fork();
  21         if (pid)
  22                 return pid;
  23
  24         while(1)
  25                 sleep(5);
  26         return 0;
  27 }
  28
  29 /*
  30  * Create an event group that contains both a sampled hardware
  31  * (cpu-cycles) and software (intel_cqm/llc_occupancy/) event. We then
  32  * wait for the hardware perf counter to overflow and generate a PMI,
  33  * which triggers an event read for both of the events in the group.
  34  *
  35  * Since reading Intel CQM event counters requires sending SMP IPIs, the
  36  * CQM pmu needs to handle the above situation gracefully, and return
  37  * the last read counter value to avoid triggering a WARN_ON_ONCE() in
  38  * smp_call_function_many() caused by sending IPIs from NMI context.
  39  */
  40 int test__intel_cqm_count_nmi_context(struct test *test __maybe_unused, int subtest __maybe_unused)
  41 {
  42         struct evlist *evlist = NULL;
  43         struct evsel *evsel = NULL;
  44         struct perf_event_attr pe;
  45         int i, fd[2], flag, ret;
  46         size_t mmap_len;
  47         void *event;
  48         pid_t pid;
  49         int err = TEST_FAIL;
  50
  51         flag = perf_event_open_cloexec_flag();
  52
  53         evlist = evlist__new();
  54         if (!evlist) {
  55                 pr_debug("evlist__new failed\n");
  56                 return TEST_FAIL;
  57         }
  58
  59         ret = parse_events(evlist, "intel_cqm/llc_occupancy/", NULL);
  60         if (ret) {
  61                 pr_debug("parse_events failed, is \"intel_cqm/llc_occupancy/\" available?\n");
  62                 err = TEST_SKIP;
  63                 goto out;
  64         }
  65
  66         evsel = evlist__first(evlist);
  67         if (!evsel) {
  68                 pr_debug("evlist__first failed\n");
  69                 goto out;
  70         }
  71
  72         memset(&pe, 0, sizeof(pe));
  73         pe.size = sizeof(pe);
  74
  75         pe.type = PERF_TYPE_HARDWARE;
  76         pe.config = PERF_COUNT_HW_CPU_CYCLES;
  77         pe.read_format = PERF_FORMAT_GROUP;
  78
  79         pe.sample_period = 128;
  80         pe.sample_type = PERF_SAMPLE_IP | PERF_SAMPLE_READ;
  81
  82         pid = spawn();
  83
  84         fd[0] = sys_perf_event_open(&pe, pid, -1, -1, flag);
  85         if (fd[0] < 0) {
  86                 pr_debug("failed to open event\n");
  87                 goto out;
  88         }
  89
  90         memset(&pe, 0, sizeof(pe));
  91         pe.size = sizeof(pe);
  92
  93         pe.type = evsel->attr.type;
  94         pe.config = evsel->attr.config;
  95
  96         fd[1] = sys_perf_event_open(&pe, pid, -1, fd[0], flag);
  97         if (fd[1] < 0) {
  98                 pr_debug("failed to open event\n");
  99                 goto out;
 100         }
 101
 102         /*
 103          * Pick a power-of-two number of pages + 1 for the meta-data
 104          * page (struct perf_event_mmap_page). See tools/perf/design.txt.
 105          */
 106         mmap_len = page_size * 65;
 107
 108         event = mmap(NULL, mmap_len, PROT_READ, MAP_SHARED, fd[0], 0);
 109         if (event == (void *)(-1)) {
 110                 pr_debug("failed to mmap %d\n", errno);
 111                 goto out;
 112         }
 113
 114         sleep(1);
 115
 116         err = TEST_OK;
 117
 118         munmap(event, mmap_len);
 119
 120         for (i = 0; i < 2; i++)
 121                 close(fd[i]);
 122
 123         kill(pid, SIGKILL);
 124         wait(NULL);
 125 out:
 126         evlist__delete(evlist);
 127         return err;
 128 }