]> asedeno.scripts.mit.edu Git - linux.git/commitdiff
libbpf: add perf buffer API
authorAndrii Nakryiko <andriin@fb.com>
Sat, 6 Jul 2019 18:06:24 +0000 (11:06 -0700)
committerDaniel Borkmann <daniel@iogearbox.net>
Mon, 8 Jul 2019 13:35:43 +0000 (15:35 +0200)
BPF_MAP_TYPE_PERF_EVENT_ARRAY map is often used to send data from BPF program
to user space for additional processing. libbpf already has very low-level API
to read single CPU perf buffer, bpf_perf_event_read_simple(), but it's hard to
use and requires a lot of code to set everything up. This patch adds
perf_buffer abstraction on top of it, abstracting setting up and polling
per-CPU logic into simple and convenient API, similar to what BCC provides.

perf_buffer__new() sets up per-CPU ring buffers and updates corresponding BPF
map entries. It accepts two user-provided callbacks: one for handling raw
samples and one for get notifications of lost samples due to buffer overflow.

perf_buffer__new_raw() is similar, but provides more control over how
perf events are set up (by accepting user-provided perf_event_attr), how
they are handled (perf_event_header pointer is passed directly to
user-provided callback), and on which CPUs ring buffers are created
(it's possible to provide a list of CPUs and corresponding map keys to
update). This API allows advanced users fuller control.

perf_buffer__poll() is used to fetch ring buffer data across all CPUs,
utilizing epoll instance.

perf_buffer__free() does corresponding clean up and unsets FDs from BPF map.

All APIs are not thread-safe. User should ensure proper locking/coordination if
used in multi-threaded set up.

Signed-off-by: Andrii Nakryiko <andriin@fb.com>
Acked-by: Yonghong Song <yhs@fb.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
tools/lib/bpf/libbpf.c
tools/lib/bpf/libbpf.h
tools/lib/bpf/libbpf.map

index 2a08eb106221251a752cbcc7cbd247d85b8d6a04..ae569b50e2e08301f233205eedeec5a456c97425 100644 (file)
@@ -32,7 +32,9 @@
 #include <linux/limits.h>
 #include <linux/perf_event.h>
 #include <linux/ring_buffer.h>
+#include <sys/epoll.h>
 #include <sys/ioctl.h>
+#include <sys/mman.h>
 #include <sys/stat.h>
 #include <sys/types.h>
 #include <sys/vfs.h>
@@ -4354,6 +4356,370 @@ bpf_perf_event_read_simple(void *mmap_mem, size_t mmap_size, size_t page_size,
        return ret;
 }
 
+struct perf_buffer;
+
+struct perf_buffer_params {
+       struct perf_event_attr *attr;
+       /* if event_cb is specified, it takes precendence */
+       perf_buffer_event_fn event_cb;
+       /* sample_cb and lost_cb are higher-level common-case callbacks */
+       perf_buffer_sample_fn sample_cb;
+       perf_buffer_lost_fn lost_cb;
+       void *ctx;
+       int cpu_cnt;
+       int *cpus;
+       int *map_keys;
+};
+
+struct perf_cpu_buf {
+       struct perf_buffer *pb;
+       void *base; /* mmap()'ed memory */
+       void *buf; /* for reconstructing segmented data */
+       size_t buf_size;
+       int fd;
+       int cpu;
+       int map_key;
+};
+
+struct perf_buffer {
+       perf_buffer_event_fn event_cb;
+       perf_buffer_sample_fn sample_cb;
+       perf_buffer_lost_fn lost_cb;
+       void *ctx; /* passed into callbacks */
+
+       size_t page_size;
+       size_t mmap_size;
+       struct perf_cpu_buf **cpu_bufs;
+       struct epoll_event *events;
+       int cpu_cnt;
+       int epoll_fd; /* perf event FD */
+       int map_fd; /* BPF_MAP_TYPE_PERF_EVENT_ARRAY BPF map FD */
+};
+
+static void perf_buffer__free_cpu_buf(struct perf_buffer *pb,
+                                     struct perf_cpu_buf *cpu_buf)
+{
+       if (!cpu_buf)
+               return;
+       if (cpu_buf->base &&
+           munmap(cpu_buf->base, pb->mmap_size + pb->page_size))
+               pr_warning("failed to munmap cpu_buf #%d\n", cpu_buf->cpu);
+       if (cpu_buf->fd >= 0) {
+               ioctl(cpu_buf->fd, PERF_EVENT_IOC_DISABLE, 0);
+               close(cpu_buf->fd);
+       }
+       free(cpu_buf->buf);
+       free(cpu_buf);
+}
+
+void perf_buffer__free(struct perf_buffer *pb)
+{
+       int i;
+
+       if (!pb)
+               return;
+       if (pb->cpu_bufs) {
+               for (i = 0; i < pb->cpu_cnt && pb->cpu_bufs[i]; i++) {
+                       struct perf_cpu_buf *cpu_buf = pb->cpu_bufs[i];
+
+                       bpf_map_delete_elem(pb->map_fd, &cpu_buf->map_key);
+                       perf_buffer__free_cpu_buf(pb, cpu_buf);
+               }
+               free(pb->cpu_bufs);
+       }
+       if (pb->epoll_fd >= 0)
+               close(pb->epoll_fd);
+       free(pb->events);
+       free(pb);
+}
+
+static struct perf_cpu_buf *
+perf_buffer__open_cpu_buf(struct perf_buffer *pb, struct perf_event_attr *attr,
+                         int cpu, int map_key)
+{
+       struct perf_cpu_buf *cpu_buf;
+       char msg[STRERR_BUFSIZE];
+       int err;
+
+       cpu_buf = calloc(1, sizeof(*cpu_buf));
+       if (!cpu_buf)
+               return ERR_PTR(-ENOMEM);
+
+       cpu_buf->pb = pb;
+       cpu_buf->cpu = cpu;
+       cpu_buf->map_key = map_key;
+
+       cpu_buf->fd = syscall(__NR_perf_event_open, attr, -1 /* pid */, cpu,
+                             -1, PERF_FLAG_FD_CLOEXEC);
+       if (cpu_buf->fd < 0) {
+               err = -errno;
+               pr_warning("failed to open perf buffer event on cpu #%d: %s\n",
+                          cpu, libbpf_strerror_r(err, msg, sizeof(msg)));
+               goto error;
+       }
+
+       cpu_buf->base = mmap(NULL, pb->mmap_size + pb->page_size,
+                            PROT_READ | PROT_WRITE, MAP_SHARED,
+                            cpu_buf->fd, 0);
+       if (cpu_buf->base == MAP_FAILED) {
+               cpu_buf->base = NULL;
+               err = -errno;
+               pr_warning("failed to mmap perf buffer on cpu #%d: %s\n",
+                          cpu, libbpf_strerror_r(err, msg, sizeof(msg)));
+               goto error;
+       }
+
+       if (ioctl(cpu_buf->fd, PERF_EVENT_IOC_ENABLE, 0) < 0) {
+               err = -errno;
+               pr_warning("failed to enable perf buffer event on cpu #%d: %s\n",
+                          cpu, libbpf_strerror_r(err, msg, sizeof(msg)));
+               goto error;
+       }
+
+       return cpu_buf;
+
+error:
+       perf_buffer__free_cpu_buf(pb, cpu_buf);
+       return (struct perf_cpu_buf *)ERR_PTR(err);
+}
+
+static struct perf_buffer *__perf_buffer__new(int map_fd, size_t page_cnt,
+                                             struct perf_buffer_params *p);
+
+struct perf_buffer *perf_buffer__new(int map_fd, size_t page_cnt,
+                                    const struct perf_buffer_opts *opts)
+{
+       struct perf_buffer_params p = {};
+       struct perf_event_attr attr = {
+               .config = PERF_COUNT_SW_BPF_OUTPUT,
+               .type = PERF_TYPE_SOFTWARE,
+               .sample_type = PERF_SAMPLE_RAW,
+               .sample_period = 1,
+               .wakeup_events = 1,
+       };
+
+       p.attr = &attr;
+       p.sample_cb = opts ? opts->sample_cb : NULL;
+       p.lost_cb = opts ? opts->lost_cb : NULL;
+       p.ctx = opts ? opts->ctx : NULL;
+
+       return __perf_buffer__new(map_fd, page_cnt, &p);
+}
+
+struct perf_buffer *
+perf_buffer__new_raw(int map_fd, size_t page_cnt,
+                    const struct perf_buffer_raw_opts *opts)
+{
+       struct perf_buffer_params p = {};
+
+       p.attr = opts->attr;
+       p.event_cb = opts->event_cb;
+       p.ctx = opts->ctx;
+       p.cpu_cnt = opts->cpu_cnt;
+       p.cpus = opts->cpus;
+       p.map_keys = opts->map_keys;
+
+       return __perf_buffer__new(map_fd, page_cnt, &p);
+}
+
+static struct perf_buffer *__perf_buffer__new(int map_fd, size_t page_cnt,
+                                             struct perf_buffer_params *p)
+{
+       struct bpf_map_info map = {};
+       char msg[STRERR_BUFSIZE];
+       struct perf_buffer *pb;
+       __u32 map_info_len;
+       int err, i;
+
+       if (page_cnt & (page_cnt - 1)) {
+               pr_warning("page count should be power of two, but is %zu\n",
+                          page_cnt);
+               return ERR_PTR(-EINVAL);
+       }
+
+       map_info_len = sizeof(map);
+       err = bpf_obj_get_info_by_fd(map_fd, &map, &map_info_len);
+       if (err) {
+               err = -errno;
+               pr_warning("failed to get map info for map FD %d: %s\n",
+                          map_fd, libbpf_strerror_r(err, msg, sizeof(msg)));
+               return ERR_PTR(err);
+       }
+
+       if (map.type != BPF_MAP_TYPE_PERF_EVENT_ARRAY) {
+               pr_warning("map '%s' should be BPF_MAP_TYPE_PERF_EVENT_ARRAY\n",
+                          map.name);
+               return ERR_PTR(-EINVAL);
+       }
+
+       pb = calloc(1, sizeof(*pb));
+       if (!pb)
+               return ERR_PTR(-ENOMEM);
+
+       pb->event_cb = p->event_cb;
+       pb->sample_cb = p->sample_cb;
+       pb->lost_cb = p->lost_cb;
+       pb->ctx = p->ctx;
+
+       pb->page_size = getpagesize();
+       pb->mmap_size = pb->page_size * page_cnt;
+       pb->map_fd = map_fd;
+
+       pb->epoll_fd = epoll_create1(EPOLL_CLOEXEC);
+       if (pb->epoll_fd < 0) {
+               err = -errno;
+               pr_warning("failed to create epoll instance: %s\n",
+                          libbpf_strerror_r(err, msg, sizeof(msg)));
+               goto error;
+       }
+
+       if (p->cpu_cnt > 0) {
+               pb->cpu_cnt = p->cpu_cnt;
+       } else {
+               pb->cpu_cnt = libbpf_num_possible_cpus();
+               if (pb->cpu_cnt < 0) {
+                       err = pb->cpu_cnt;
+                       goto error;
+               }
+               if (map.max_entries < pb->cpu_cnt)
+                       pb->cpu_cnt = map.max_entries;
+       }
+
+       pb->events = calloc(pb->cpu_cnt, sizeof(*pb->events));
+       if (!pb->events) {
+               err = -ENOMEM;
+               pr_warning("failed to allocate events: out of memory\n");
+               goto error;
+       }
+       pb->cpu_bufs = calloc(pb->cpu_cnt, sizeof(*pb->cpu_bufs));
+       if (!pb->cpu_bufs) {
+               err = -ENOMEM;
+               pr_warning("failed to allocate buffers: out of memory\n");
+               goto error;
+       }
+
+       for (i = 0; i < pb->cpu_cnt; i++) {
+               struct perf_cpu_buf *cpu_buf;
+               int cpu, map_key;
+
+               cpu = p->cpu_cnt > 0 ? p->cpus[i] : i;
+               map_key = p->cpu_cnt > 0 ? p->map_keys[i] : i;
+
+               cpu_buf = perf_buffer__open_cpu_buf(pb, p->attr, cpu, map_key);
+               if (IS_ERR(cpu_buf)) {
+                       err = PTR_ERR(cpu_buf);
+                       goto error;
+               }
+
+               pb->cpu_bufs[i] = cpu_buf;
+
+               err = bpf_map_update_elem(pb->map_fd, &map_key,
+                                         &cpu_buf->fd, 0);
+               if (err) {
+                       err = -errno;
+                       pr_warning("failed to set cpu #%d, key %d -> perf FD %d: %s\n",
+                                  cpu, map_key, cpu_buf->fd,
+                                  libbpf_strerror_r(err, msg, sizeof(msg)));
+                       goto error;
+               }
+
+               pb->events[i].events = EPOLLIN;
+               pb->events[i].data.ptr = cpu_buf;
+               if (epoll_ctl(pb->epoll_fd, EPOLL_CTL_ADD, cpu_buf->fd,
+                             &pb->events[i]) < 0) {
+                       err = -errno;
+                       pr_warning("failed to epoll_ctl cpu #%d perf FD %d: %s\n",
+                                  cpu, cpu_buf->fd,
+                                  libbpf_strerror_r(err, msg, sizeof(msg)));
+                       goto error;
+               }
+       }
+
+       return pb;
+
+error:
+       if (pb)
+               perf_buffer__free(pb);
+       return ERR_PTR(err);
+}
+
+struct perf_sample_raw {
+       struct perf_event_header header;
+       uint32_t size;
+       char data[0];
+};
+
+struct perf_sample_lost {
+       struct perf_event_header header;
+       uint64_t id;
+       uint64_t lost;
+       uint64_t sample_id;
+};
+
+static enum bpf_perf_event_ret
+perf_buffer__process_record(struct perf_event_header *e, void *ctx)
+{
+       struct perf_cpu_buf *cpu_buf = ctx;
+       struct perf_buffer *pb = cpu_buf->pb;
+       void *data = e;
+
+       /* user wants full control over parsing perf event */
+       if (pb->event_cb)
+               return pb->event_cb(pb->ctx, cpu_buf->cpu, e);
+
+       switch (e->type) {
+       case PERF_RECORD_SAMPLE: {
+               struct perf_sample_raw *s = data;
+
+               if (pb->sample_cb)
+                       pb->sample_cb(pb->ctx, cpu_buf->cpu, s->data, s->size);
+               break;
+       }
+       case PERF_RECORD_LOST: {
+               struct perf_sample_lost *s = data;
+
+               if (pb->lost_cb)
+                       pb->lost_cb(pb->ctx, cpu_buf->cpu, s->lost);
+               break;
+       }
+       default:
+               pr_warning("unknown perf sample type %d\n", e->type);
+               return LIBBPF_PERF_EVENT_ERROR;
+       }
+       return LIBBPF_PERF_EVENT_CONT;
+}
+
+static int perf_buffer__process_records(struct perf_buffer *pb,
+                                       struct perf_cpu_buf *cpu_buf)
+{
+       enum bpf_perf_event_ret ret;
+
+       ret = bpf_perf_event_read_simple(cpu_buf->base, pb->mmap_size,
+                                        pb->page_size, &cpu_buf->buf,
+                                        &cpu_buf->buf_size,
+                                        perf_buffer__process_record, cpu_buf);
+       if (ret != LIBBPF_PERF_EVENT_CONT)
+               return ret;
+       return 0;
+}
+
+int perf_buffer__poll(struct perf_buffer *pb, int timeout_ms)
+{
+       int i, cnt, err;
+
+       cnt = epoll_wait(pb->epoll_fd, pb->events, pb->cpu_cnt, timeout_ms);
+       for (i = 0; i < cnt; i++) {
+               struct perf_cpu_buf *cpu_buf = pb->events[i].data.ptr;
+
+               err = perf_buffer__process_records(pb, cpu_buf);
+               if (err) {
+                       pr_warning("error while processing records: %d\n", err);
+                       return err;
+               }
+       }
+       return cnt < 0 ? -errno : cnt;
+}
+
 struct bpf_prog_info_array_desc {
        int     array_offset;   /* e.g. offset of jited_prog_insns */
        int     count_offset;   /* e.g. offset of jited_prog_len */
index f55933784f95208b230b283355aad53bfc92708f..5cbf459ece0b29eb3e36981d0b3cfe5a951ae562 100644 (file)
@@ -358,6 +358,26 @@ LIBBPF_API int bpf_prog_load(const char *file, enum bpf_prog_type type,
 LIBBPF_API int bpf_set_link_xdp_fd(int ifindex, int fd, __u32 flags);
 LIBBPF_API int bpf_get_link_xdp_id(int ifindex, __u32 *prog_id, __u32 flags);
 
+struct perf_buffer;
+
+typedef void (*perf_buffer_sample_fn)(void *ctx, int cpu,
+                                     void *data, __u32 size);
+typedef void (*perf_buffer_lost_fn)(void *ctx, int cpu, __u64 cnt);
+
+/* common use perf buffer options */
+struct perf_buffer_opts {
+       /* if specified, sample_cb is called for each sample */
+       perf_buffer_sample_fn sample_cb;
+       /* if specified, lost_cb is called for each batch of lost samples */
+       perf_buffer_lost_fn lost_cb;
+       /* ctx is provided to sample_cb and lost_cb */
+       void *ctx;
+};
+
+LIBBPF_API struct perf_buffer *
+perf_buffer__new(int map_fd, size_t page_cnt,
+                const struct perf_buffer_opts *opts);
+
 enum bpf_perf_event_ret {
        LIBBPF_PERF_EVENT_DONE  = 0,
        LIBBPF_PERF_EVENT_ERROR = -1,
@@ -365,6 +385,35 @@ enum bpf_perf_event_ret {
 };
 
 struct perf_event_header;
+
+typedef enum bpf_perf_event_ret
+(*perf_buffer_event_fn)(void *ctx, int cpu, struct perf_event_header *event);
+
+/* raw perf buffer options, giving most power and control */
+struct perf_buffer_raw_opts {
+       /* perf event attrs passed directly into perf_event_open() */
+       struct perf_event_attr *attr;
+       /* raw event callback */
+       perf_buffer_event_fn event_cb;
+       /* ctx is provided to event_cb */
+       void *ctx;
+       /* if cpu_cnt == 0, open all on all possible CPUs (up to the number of
+        * max_entries of given PERF_EVENT_ARRAY map)
+        */
+       int cpu_cnt;
+       /* if cpu_cnt > 0, cpus is an array of CPUs to open ring buffers on */
+       int *cpus;
+       /* if cpu_cnt > 0, map_keys specify map keys to set per-CPU FDs for */
+       int *map_keys;
+};
+
+LIBBPF_API struct perf_buffer *
+perf_buffer__new_raw(int map_fd, size_t page_cnt,
+                    const struct perf_buffer_raw_opts *opts);
+
+LIBBPF_API void perf_buffer__free(struct perf_buffer *pb);
+LIBBPF_API int perf_buffer__poll(struct perf_buffer *pb, int timeout_ms);
+
 typedef enum bpf_perf_event_ret
        (*bpf_perf_event_print_t)(struct perf_event_header *hdr,
                                  void *private_data);
index e6b7d4edbc936e461cb960155047dab429951c5e..f9d316e873d8d2d7ff37d138fdef0cbbbd425bcd 100644 (file)
@@ -179,4 +179,8 @@ LIBBPF_0.0.4 {
                btf_dump__new;
                btf__parse_elf;
                libbpf_num_possible_cpus;
+               perf_buffer__free;
+               perf_buffer__new;
+               perf_buffer__new_raw;
+               perf_buffer__poll;
 } LIBBPF_0.0.3;