libbpf: Add legacy uprobe attaching support
authorAndrii Nakryiko <andrii@kernel.org>
Tue, 21 Sep 2021 21:00:36 +0000 (14:00 -0700)
committerAlexei Starovoitov <ast@kernel.org>
Wed, 22 Sep 2021 02:40:09 +0000 (19:40 -0700)
Similarly to recently added legacy kprobe attach interface support
through tracefs, support attaching uprobes using the legacy interface if
host kernel doesn't support newer FD-based interface.

For uprobes event name consists of "libbpf_" prefix, PID, sanitized
binary path and offset within that binary. Structuraly the code is
aligned with kprobe logic refactoring in previous patch. struct
bpf_link_perf is re-used and all the same legacy_probe_name and
legacy_is_retprobe fields are used to ensure proper cleanup on
bpf_link__destroy().

Users should be aware, though, that on old kernels which don't support
FD-based interface for kprobe/uprobe attachment, if the application
crashes before bpf_link__destroy() is called, uprobe legacy
events will be left in tracefs. This is the same limitation as with
legacy kprobe interfaces.

Signed-off-by: Andrii Nakryiko <andrii@kernel.org>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/20210921210036.1545557-5-andrii@kernel.org
tools/lib/bpf/libbpf.c

index aa842f0..ef5db34 100644 (file)
@@ -9021,6 +9021,7 @@ struct bpf_link_perf {
 };
 
 static int remove_kprobe_event_legacy(const char *probe_name, bool retprobe);
+static int remove_uprobe_event_legacy(const char *probe_name, bool retprobe);
 
 static int bpf_link_perf_detach(struct bpf_link *link)
 {
@@ -9034,11 +9035,14 @@ static int bpf_link_perf_detach(struct bpf_link *link)
                close(perf_link->perf_event_fd);
        close(link->fd);
 
-       /* legacy kprobe needs to be removed after perf event fd closure */
+       /* legacy uprobe/kprobe needs to be removed after perf event fd closure */
        if (perf_link->legacy_probe_name) {
                if (perf_link->legacy_is_kprobe) {
                        err = remove_kprobe_event_legacy(perf_link->legacy_probe_name,
                                                         perf_link->legacy_is_retprobe);
+               } else {
+                       err = remove_uprobe_event_legacy(perf_link->legacy_probe_name,
+                                                        perf_link->legacy_is_retprobe);
                }
        }
 
@@ -9450,17 +9454,96 @@ static struct bpf_link *attach_kprobe(const struct bpf_program *prog)
        return link;
 }
 
+static void gen_uprobe_legacy_event_name(char *buf, size_t buf_sz,
+                                        const char *binary_path, uint64_t offset)
+{
+       int i;
+
+       snprintf(buf, buf_sz, "libbpf_%u_%s_0x%zx", getpid(), binary_path, (size_t)offset);
+
+       /* sanitize binary_path in the probe name */
+       for (i = 0; buf[i]; i++) {
+               if (!isalnum(buf[i]))
+                       buf[i] = '_';
+       }
+}
+
+static inline int add_uprobe_event_legacy(const char *probe_name, bool retprobe,
+                                         const char *binary_path, size_t offset)
+{
+       const char *file = "/sys/kernel/debug/tracing/uprobe_events";
+
+       return append_to_file(file, "%c:%s/%s %s:0x%zx",
+                             retprobe ? 'r' : 'p',
+                             retprobe ? "uretprobes" : "uprobes",
+                             probe_name, binary_path, offset);
+}
+
+static inline int remove_uprobe_event_legacy(const char *probe_name, bool retprobe)
+{
+       const char *file = "/sys/kernel/debug/tracing/uprobe_events";
+
+       return append_to_file(file, "-:%s/%s", retprobe ? "uretprobes" : "uprobes", probe_name);
+}
+
+static int determine_uprobe_perf_type_legacy(const char *probe_name, bool retprobe)
+{
+       char file[512];
+
+       snprintf(file, sizeof(file),
+                "/sys/kernel/debug/tracing/events/%s/%s/id",
+                retprobe ? "uretprobes" : "uprobes", probe_name);
+
+       return parse_uint_from_file(file, "%d\n");
+}
+
+static int perf_event_uprobe_open_legacy(const char *probe_name, bool retprobe,
+                                        const char *binary_path, size_t offset, int pid)
+{
+       struct perf_event_attr attr;
+       int type, pfd, err;
+
+       err = add_uprobe_event_legacy(probe_name, retprobe, binary_path, offset);
+       if (err < 0) {
+               pr_warn("failed to add legacy uprobe event for %s:0x%zx: %d\n",
+                       binary_path, (size_t)offset, err);
+               return err;
+       }
+       type = determine_uprobe_perf_type_legacy(probe_name, retprobe);
+       if (type < 0) {
+               pr_warn("failed to determine legacy uprobe event id for %s:0x%zx: %d\n",
+                       binary_path, offset, err);
+               return type;
+       }
+
+       memset(&attr, 0, sizeof(attr));
+       attr.size = sizeof(attr);
+       attr.config = type;
+       attr.type = PERF_TYPE_TRACEPOINT;
+
+       pfd = syscall(__NR_perf_event_open, &attr,
+                     pid < 0 ? -1 : pid, /* pid */
+                     pid == -1 ? 0 : -1, /* cpu */
+                     -1 /* group_fd */,  PERF_FLAG_FD_CLOEXEC);
+       if (pfd < 0) {
+               err = -errno;
+               pr_warn("legacy uprobe perf_event_open() failed: %d\n", err);
+               return err;
+       }
+       return pfd;
+}
+
 LIBBPF_API struct bpf_link *
 bpf_program__attach_uprobe_opts(const struct bpf_program *prog, pid_t pid,
                                const char *binary_path, size_t func_offset,
                                const struct bpf_uprobe_opts *opts)
 {
        DECLARE_LIBBPF_OPTS(bpf_perf_event_opts, pe_opts);
-       char errmsg[STRERR_BUFSIZE];
+       char errmsg[STRERR_BUFSIZE], *legacy_probe = NULL;
        struct bpf_link *link;
        size_t ref_ctr_off;
        int pfd, err;
-       bool retprobe;
+       bool retprobe, legacy;
 
        if (!OPTS_VALID(opts, bpf_uprobe_opts))
                return libbpf_err_ptr(-EINVAL);
@@ -9469,15 +9552,35 @@ bpf_program__attach_uprobe_opts(const struct bpf_program *prog, pid_t pid,
        ref_ctr_off = OPTS_GET(opts, ref_ctr_offset, 0);
        pe_opts.bpf_cookie = OPTS_GET(opts, bpf_cookie, 0);
 
-       pfd = perf_event_open_probe(true /* uprobe */, retprobe, binary_path,
-                                   func_offset, pid, ref_ctr_off);
+       legacy = determine_uprobe_perf_type() < 0;
+       if (!legacy) {
+               pfd = perf_event_open_probe(true /* uprobe */, retprobe, binary_path,
+                                           func_offset, pid, ref_ctr_off);
+       } else {
+               char probe_name[512];
+
+               if (ref_ctr_off)
+                       return libbpf_err_ptr(-EINVAL);
+
+               gen_uprobe_legacy_event_name(probe_name, sizeof(probe_name),
+                                            binary_path, func_offset);
+
+               legacy_probe = strdup(probe_name);
+               if (!legacy_probe)
+                       return libbpf_err_ptr(-ENOMEM);
+
+               pfd = perf_event_uprobe_open_legacy(legacy_probe, retprobe,
+                                                   binary_path, func_offset, pid);
+       }
        if (pfd < 0) {
+               err = -errno;
                pr_warn("prog '%s': failed to create %s '%s:0x%zx' perf event: %s\n",
                        prog->name, retprobe ? "uretprobe" : "uprobe",
                        binary_path, func_offset,
-                       libbpf_strerror_r(pfd, errmsg, sizeof(errmsg)));
-               return libbpf_err_ptr(pfd);
+                       libbpf_strerror_r(err, errmsg, sizeof(errmsg)));
+               goto err_out;
        }
+
        link = bpf_program__attach_perf_event_opts(prog, pfd, &pe_opts);
        err = libbpf_get_error(link);
        if (err) {
@@ -9486,9 +9589,20 @@ bpf_program__attach_uprobe_opts(const struct bpf_program *prog, pid_t pid,
                        prog->name, retprobe ? "uretprobe" : "uprobe",
                        binary_path, func_offset,
                        libbpf_strerror_r(err, errmsg, sizeof(errmsg)));
-               return libbpf_err_ptr(err);
+               goto err_out;
+       }
+       if (legacy) {
+               struct bpf_link_perf *perf_link = container_of(link, struct bpf_link_perf, link);
+
+               perf_link->legacy_probe_name = legacy_probe;
+               perf_link->legacy_is_kprobe = false;
+               perf_link->legacy_is_retprobe = retprobe;
        }
        return link;
+err_out:
+       free(legacy_probe);
+       return libbpf_err_ptr(err);
+
 }
 
 struct bpf_link *bpf_program__attach_uprobe(const struct bpf_program *prog,