1 #include <stdio.h> 2 #include <fcntl.h> 3 #include <stdint.h> 4 #include <stdlib.h> 5 6 #include <linux/err.h> 7 8 #include "util/ftrace.h" 9 #include "util/cpumap.h" 10 #include "util/thread_map.h" 11 #include "util/debug.h" 12 #include "util/evlist.h" 13 #include "util/bpf_counter.h" 14 15 #include "util/bpf_skel/func_latency.skel.h" 16 17 static struct func_latency_bpf *skel; 18 19 int perf_ftrace__latency_prepare_bpf(struct perf_ftrace *ftrace) 20 { 21 int fd, err; 22 int i, ncpus = 1, ntasks = 1; 23 struct filter_entry *func; 24 25 if (!list_is_singular(&ftrace->filters)) { 26 pr_err("ERROR: %s target function(s).\n", 27 list_empty(&ftrace->filters) ? "No" : "Too many"); 28 return -1; 29 } 30 31 func = list_first_entry(&ftrace->filters, struct filter_entry, list); 32 33 skel = func_latency_bpf__open(); 34 if (!skel) { 35 pr_err("Failed to open func latency skeleton\n"); 36 return -1; 37 } 38 39 /* don't need to set cpu filter for system-wide mode */ 40 if (ftrace->target.cpu_list) { 41 ncpus = perf_cpu_map__nr(ftrace->evlist->core.cpus); 42 bpf_map__set_max_entries(skel->maps.cpu_filter, ncpus); 43 } 44 45 if (target__has_task(&ftrace->target) || target__none(&ftrace->target)) { 46 ntasks = perf_thread_map__nr(ftrace->evlist->core.threads); 47 bpf_map__set_max_entries(skel->maps.task_filter, ntasks); 48 } 49 50 set_max_rlimit(); 51 52 err = func_latency_bpf__load(skel); 53 if (err) { 54 pr_err("Failed to load func latency skeleton\n"); 55 goto out; 56 } 57 58 if (ftrace->target.cpu_list) { 59 u32 cpu; 60 u8 val = 1; 61 62 skel->bss->has_cpu = 1; 63 fd = bpf_map__fd(skel->maps.cpu_filter); 64 65 for (i = 0; i < ncpus; i++) { 66 cpu = perf_cpu_map__cpu(ftrace->evlist->core.cpus, i).cpu; 67 bpf_map_update_elem(fd, &cpu, &val, BPF_ANY); 68 } 69 } 70 71 if (target__has_task(&ftrace->target) || target__none(&ftrace->target)) { 72 u32 pid; 73 u8 val = 1; 74 75 skel->bss->has_task = 1; 76 fd = bpf_map__fd(skel->maps.task_filter); 77 78 for (i = 0; i < ntasks; i++) { 79 pid = perf_thread_map__pid(ftrace->evlist->core.threads, i); 80 bpf_map_update_elem(fd, &pid, &val, BPF_ANY); 81 } 82 } 83 84 skel->bss->use_nsec = ftrace->use_nsec; 85 86 skel->links.func_begin = bpf_program__attach_kprobe(skel->progs.func_begin, 87 false, func->name); 88 if (IS_ERR(skel->links.func_begin)) { 89 pr_err("Failed to attach fentry program\n"); 90 err = PTR_ERR(skel->links.func_begin); 91 goto out; 92 } 93 94 skel->links.func_end = bpf_program__attach_kprobe(skel->progs.func_end, 95 true, func->name); 96 if (IS_ERR(skel->links.func_end)) { 97 pr_err("Failed to attach fexit program\n"); 98 err = PTR_ERR(skel->links.func_end); 99 goto out; 100 } 101 102 /* XXX: we don't actually use this fd - just for poll() */ 103 return open("/dev/null", O_RDONLY); 104 105 out: 106 return err; 107 } 108 109 int perf_ftrace__latency_start_bpf(struct perf_ftrace *ftrace __maybe_unused) 110 { 111 skel->bss->enabled = 1; 112 return 0; 113 } 114 115 int perf_ftrace__latency_stop_bpf(struct perf_ftrace *ftrace __maybe_unused) 116 { 117 skel->bss->enabled = 0; 118 return 0; 119 } 120 121 int perf_ftrace__latency_read_bpf(struct perf_ftrace *ftrace __maybe_unused, 122 int buckets[]) 123 { 124 int i, fd, err; 125 u32 idx; 126 u64 *hist; 127 int ncpus = cpu__max_cpu().cpu; 128 129 fd = bpf_map__fd(skel->maps.latency); 130 131 hist = calloc(ncpus, sizeof(*hist)); 132 if (hist == NULL) 133 return -ENOMEM; 134 135 for (idx = 0; idx < NUM_BUCKET; idx++) { 136 err = bpf_map_lookup_elem(fd, &idx, hist); 137 if (err) { 138 buckets[idx] = 0; 139 continue; 140 } 141 142 for (i = 0; i < ncpus; i++) 143 buckets[idx] += hist[i]; 144 } 145 146 free(hist); 147 return 0; 148 } 149 150 int perf_ftrace__latency_cleanup_bpf(struct perf_ftrace *ftrace __maybe_unused) 151 { 152 func_latency_bpf__destroy(skel); 153 return 0; 154 } 155