tools/bpftool: switch map event_pipe to libbpf's perf_buffer
authorAndrii Nakryiko <andriin@fb.com>
Sat, 6 Jul 2019 18:06:27 +0000 (11:06 -0700)
committerDaniel Borkmann <daniel@iogearbox.net>
Mon, 8 Jul 2019 13:35:43 +0000 (15:35 +0200)
Switch event_pipe implementation to rely on new libbpf perf buffer API
(it's raw low-level variant).

Signed-off-by: Andrii Nakryiko <andriin@fb.com>
Acked-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Acked-by: Yonghong Song <yhs@fb.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
tools/bpf/bpftool/map_perf_ring.c

index 0507dfaf7a8f93f5c0e2daf8de837e1163dff203..3f108ab177973761a03601e8d69da0719e3bd052 100644 (file)
@@ -28,7 +28,7 @@
 
 #define MMAP_PAGE_CNT  16
 
-static bool stop;
+static volatile bool stop;
 
 struct event_ring_info {
        int fd;
@@ -44,32 +44,44 @@ struct perf_event_sample {
        unsigned char data[];
 };
 
+struct perf_event_lost {
+       struct perf_event_header header;
+       __u64 id;
+       __u64 lost;
+};
+
 static void int_exit(int signo)
 {
        fprintf(stderr, "Stopping...\n");
        stop = true;
 }
 
+struct event_pipe_ctx {
+       bool all_cpus;
+       int cpu;
+       int idx;
+};
+
 static enum bpf_perf_event_ret
-print_bpf_output(struct perf_event_header *event, void *private_data)
+print_bpf_output(void *private_data, int cpu, struct perf_event_header *event)
 {
-       struct perf_event_sample *e = container_of(event, struct perf_event_sample,
+       struct perf_event_sample *e = container_of(event,
+                                                  struct perf_event_sample,
                                                   header);
-       struct event_ring_info *ring = private_data;
-       struct {
-               struct perf_event_header header;
-               __u64 id;
-               __u64 lost;
-       } *lost = (typeof(lost))event;
+       struct perf_event_lost *lost = container_of(event,
+                                                   struct perf_event_lost,
+                                                   header);
+       struct event_pipe_ctx *ctx = private_data;
+       int idx = ctx->all_cpus ? cpu : ctx->idx;
 
        if (json_output) {
                jsonw_start_object(json_wtr);
                jsonw_name(json_wtr, "type");
                jsonw_uint(json_wtr, e->header.type);
                jsonw_name(json_wtr, "cpu");
-               jsonw_uint(json_wtr, ring->cpu);
+               jsonw_uint(json_wtr, cpu);
                jsonw_name(json_wtr, "index");
-               jsonw_uint(json_wtr, ring->key);
+               jsonw_uint(json_wtr, idx);
                if (e->header.type == PERF_RECORD_SAMPLE) {
                        jsonw_name(json_wtr, "timestamp");
                        jsonw_uint(json_wtr, e->time);
@@ -89,7 +101,7 @@ print_bpf_output(struct perf_event_header *event, void *private_data)
                if (e->header.type == PERF_RECORD_SAMPLE) {
                        printf("== @%lld.%09lld CPU: %d index: %d =====\n",
                               e->time / 1000000000ULL, e->time % 1000000000ULL,
-                              ring->cpu, ring->key);
+                              cpu, idx);
                        fprint_hex(stdout, e->data, e->size, " ");
                        printf("\n");
                } else if (e->header.type == PERF_RECORD_LOST) {
@@ -103,87 +115,25 @@ print_bpf_output(struct perf_event_header *event, void *private_data)
        return LIBBPF_PERF_EVENT_CONT;
 }
 
-static void
-perf_event_read(struct event_ring_info *ring, void **buf, size_t *buf_len)
-{
-       enum bpf_perf_event_ret ret;
-
-       ret = bpf_perf_event_read_simple(ring->mem,
-                                        MMAP_PAGE_CNT * get_page_size(),
-                                        get_page_size(), buf, buf_len,
-                                        print_bpf_output, ring);
-       if (ret != LIBBPF_PERF_EVENT_CONT) {
-               fprintf(stderr, "perf read loop failed with %d\n", ret);
-               stop = true;
-       }
-}
-
-static int perf_mmap_size(void)
-{
-       return get_page_size() * (MMAP_PAGE_CNT + 1);
-}
-
-static void *perf_event_mmap(int fd)
-{
-       int mmap_size = perf_mmap_size();
-       void *base;
-
-       base = mmap(NULL, mmap_size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
-       if (base == MAP_FAILED) {
-               p_err("event mmap failed: %s\n", strerror(errno));
-               return NULL;
-       }
-
-       return base;
-}
-
-static void perf_event_unmap(void *mem)
-{
-       if (munmap(mem, perf_mmap_size()))
-               fprintf(stderr, "Can't unmap ring memory!\n");
-}
-
-static int bpf_perf_event_open(int map_fd, int key, int cpu)
+int do_event_pipe(int argc, char **argv)
 {
-       struct perf_event_attr attr = {
+       struct perf_event_attr perf_attr = {
                .sample_type = PERF_SAMPLE_RAW | PERF_SAMPLE_TIME,
                .type = PERF_TYPE_SOFTWARE,
                .config = PERF_COUNT_SW_BPF_OUTPUT,
+               .sample_period = 1,
+               .wakeup_events = 1,
        };
-       int pmu_fd;
-
-       pmu_fd = sys_perf_event_open(&attr, -1, cpu, -1, 0);
-       if (pmu_fd < 0) {
-               p_err("failed to open perf event %d for CPU %d", key, cpu);
-               return -1;
-       }
-
-       if (bpf_map_update_elem(map_fd, &key, &pmu_fd, BPF_ANY)) {
-               p_err("failed to update map for event %d for CPU %d", key, cpu);
-               goto err_close;
-       }
-       if (ioctl(pmu_fd, PERF_EVENT_IOC_ENABLE, 0)) {
-               p_err("failed to enable event %d for CPU %d", key, cpu);
-               goto err_close;
-       }
-
-       return pmu_fd;
-
-err_close:
-       close(pmu_fd);
-       return -1;
-}
-
-int do_event_pipe(int argc, char **argv)
-{
-       int i, nfds, map_fd, index = -1, cpu = -1;
        struct bpf_map_info map_info = {};
-       struct event_ring_info *rings;
-       size_t tmp_buf_sz = 0;
-       void *tmp_buf = NULL;
-       struct pollfd *pfds;
+       struct perf_buffer_raw_opts opts = {};
+       struct event_pipe_ctx ctx = {
+               .all_cpus = true,
+               .cpu = -1,
+               .idx = -1,
+       };
+       struct perf_buffer *pb;
        __u32 map_info_len;
-       bool do_all = true;
+       int err, map_fd;
 
        map_info_len = sizeof(map_info);
        map_fd = map_parse_fd_and_info(&argc, &argv, &map_info, &map_info_len);
@@ -205,7 +155,7 @@ int do_event_pipe(int argc, char **argv)
                        char *endptr;
 
                        NEXT_ARG();
-                       cpu = strtoul(*argv, &endptr, 0);
+                       ctx.cpu = strtoul(*argv, &endptr, 0);
                        if (*endptr) {
                                p_err("can't parse %s as CPU ID", **argv);
                                goto err_close_map;
@@ -216,7 +166,7 @@ int do_event_pipe(int argc, char **argv)
                        char *endptr;
 
                        NEXT_ARG();
-                       index = strtoul(*argv, &endptr, 0);
+                       ctx.idx = strtoul(*argv, &endptr, 0);
                        if (*endptr) {
                                p_err("can't parse %s as index", **argv);
                                goto err_close_map;
@@ -228,45 +178,32 @@ int do_event_pipe(int argc, char **argv)
                        goto err_close_map;
                }
 
-               do_all = false;
+               ctx.all_cpus = false;
        }
 
-       if (!do_all) {
-               if (index == -1 || cpu == -1) {
+       if (!ctx.all_cpus) {
+               if (ctx.idx == -1 || ctx.cpu == -1) {
                        p_err("cpu and index must be specified together");
                        goto err_close_map;
                }
-
-               nfds = 1;
        } else {
-               nfds = min(get_possible_cpus(), map_info.max_entries);
-               cpu = 0;
-               index = 0;
+               ctx.cpu = 0;
+               ctx.idx = 0;
        }
 
-       rings = calloc(nfds, sizeof(rings[0]));
-       if (!rings)
+       opts.attr = &perf_attr;
+       opts.event_cb = print_bpf_output;
+       opts.ctx = &ctx;
+       opts.cpu_cnt = ctx.all_cpus ? 0 : 1;
+       opts.cpus = &ctx.cpu;
+       opts.map_keys = &ctx.idx;
+
+       pb = perf_buffer__new_raw(map_fd, MMAP_PAGE_CNT, &opts);
+       err = libbpf_get_error(pb);
+       if (err) {
+               p_err("failed to create perf buffer: %s (%d)",
+                     strerror(err), err);
                goto err_close_map;
-
-       pfds = calloc(nfds, sizeof(pfds[0]));
-       if (!pfds)
-               goto err_free_rings;
-
-       for (i = 0; i < nfds; i++) {
-               rings[i].cpu = cpu + i;
-               rings[i].key = index + i;
-
-               rings[i].fd = bpf_perf_event_open(map_fd, rings[i].key,
-                                                 rings[i].cpu);
-               if (rings[i].fd < 0)
-                       goto err_close_fds_prev;
-
-               rings[i].mem = perf_event_mmap(rings[i].fd);
-               if (!rings[i].mem)
-                       goto err_close_fds_current;
-
-               pfds[i].fd = rings[i].fd;
-               pfds[i].events = POLLIN;
        }
 
        signal(SIGINT, int_exit);
@@ -277,34 +214,24 @@ int do_event_pipe(int argc, char **argv)
                jsonw_start_array(json_wtr);
 
        while (!stop) {
-               poll(pfds, nfds, 200);
-               for (i = 0; i < nfds; i++)
-                       perf_event_read(&rings[i], &tmp_buf, &tmp_buf_sz);
+               err = perf_buffer__poll(pb, 200);
+               if (err < 0 && err != -EINTR) {
+                       p_err("perf buffer polling failed: %s (%d)",
+                             strerror(err), err);
+                       goto err_close_pb;
+               }
        }
-       free(tmp_buf);
 
        if (json_output)
                jsonw_end_array(json_wtr);
 
-       for (i = 0; i < nfds; i++) {
-               perf_event_unmap(rings[i].mem);
-               close(rings[i].fd);
-       }
-       free(pfds);
-       free(rings);
+       perf_buffer__free(pb);
        close(map_fd);
 
        return 0;
 
-err_close_fds_prev:
-       while (i--) {
-               perf_event_unmap(rings[i].mem);
-err_close_fds_current:
-               close(rings[i].fd);
-       }
-       free(pfds);
-err_free_rings:
-       free(rings);
+err_close_pb:
+       perf_buffer__free(pb);
 err_close_map:
        close(map_fd);
        return -1;