]> git.proxmox.com Git - mirror_ubuntu-kernels.git/commitdiff
bpftool: Add struct_ops support
authorMartin KaFai Lau <kafai@fb.com>
Wed, 18 Mar 2020 17:16:56 +0000 (10:16 -0700)
committerDaniel Borkmann <daniel@iogearbox.net>
Fri, 20 Mar 2020 14:51:35 +0000 (15:51 +0100)
This patch adds struct_ops support to the bpftool.

To recap a bit on the recent bpf_struct_ops feature on the kernel side:
It currently supports "struct tcp_congestion_ops" to be implemented
in bpf.  At a high level, bpf_struct_ops is struct_ops map populated
with a number of bpf progs.  bpf_struct_ops currently supports the
"struct tcp_congestion_ops".  However, the bpf_struct_ops design is
generic enough that other kernel struct ops can be supported in
the future.

Although struct_ops is map+progs at a high lever, there are differences
in details.  For example,
1) After registering a struct_ops, the struct_ops is held by the kernel
   subsystem (e.g. tcp-cc).  Thus, there is no need to pin a
   struct_ops map or its progs in order to keep them around.
2) To iterate all struct_ops in a system, it iterates all maps
   in type BPF_MAP_TYPE_STRUCT_OPS.  BPF_MAP_TYPE_STRUCT_OPS is
   the current usual filter.  In the future, it may need to
   filter by other struct_ops specific properties.  e.g. filter by
   tcp_congestion_ops or other kernel subsystem ops in the future.
3) struct_ops requires the running kernel having BTF info.  That allows
   more flexibility in handling other kernel structs.  e.g. it can
   always dump the latest bpf_map_info.
4) Also, "struct_ops" command is not intended to repeat all features
   already provided by "map" or "prog".  For example, if there really
   is a need to pin the struct_ops map, the user can use the "map" cmd
   to do that.

While the first attempt was to reuse parts from map/prog.c,  it ended up
not a lot to share.  The only obvious item is the map_parse_fds() but
that still requires modifications to accommodate struct_ops map specific
filtering (for the immediate and the future needs).  Together with the
earlier mentioned differences, it is better to part away from map/prog.c.

The initial set of subcmds are, register, unregister, show, and dump.

For register, it registers all struct_ops maps that can be found in an
obj file.  Option can be added in the future to specify a particular
struct_ops map.  Also, the common bpf_tcp_cc is stateless (e.g.
bpf_cubic.c and bpf_dctcp.c).  The "reuse map" feature is not
implemented in this patch and it can be considered later also.

For other subcmds, please see the man doc for details.

A sample output of dump:
[root@arch-fb-vm1 bpf]# bpftool struct_ops dump name cubic
[{
        "bpf_map_info": {
            "type": 26,
            "id": 64,
            "key_size": 4,
            "value_size": 256,
            "max_entries": 1,
            "map_flags": 0,
            "name": "cubic",
            "ifindex": 0,
            "btf_vmlinux_value_type_id": 18452,
            "netns_dev": 0,
            "netns_ino": 0,
            "btf_id": 52,
            "btf_key_type_id": 0,
            "btf_value_type_id": 0
        }
    },{
        "bpf_struct_ops_tcp_congestion_ops": {
            "refcnt": {
                "refs": {
                    "counter": 1
                }
            },
            "state": "BPF_STRUCT_OPS_STATE_INUSE",
            "data": {
                "list": {
                    "next": 0,
                    "prev": 0
                },
                "key": 0,
                "flags": 0,
                "init": "void (struct sock *) bictcp_init/prog_id:138",
                "release": "void (struct sock *) 0",
                "ssthresh": "u32 (struct sock *) bictcp_recalc_ssthresh/prog_id:141",
                "cong_avoid": "void (struct sock *, u32, u32) bictcp_cong_avoid/prog_id:140",
                "set_state": "void (struct sock *, u8) bictcp_state/prog_id:142",
                "cwnd_event": "void (struct sock *, enum tcp_ca_event) bictcp_cwnd_event/prog_id:139",
                "in_ack_event": "void (struct sock *, u32) 0",
                "undo_cwnd": "u32 (struct sock *) tcp_reno_undo_cwnd/prog_id:144",
                "pkts_acked": "void (struct sock *, const struct ack_sample *) bictcp_acked/prog_id:143",
                "min_tso_segs": "u32 (struct sock *) 0",
                "sndbuf_expand": "u32 (struct sock *) 0",
                "cong_control": "void (struct sock *, const struct rate_sample *) 0",
                "get_info": "size_t (struct sock *, u32, int *, union tcp_cc_info *) 0",
                "name": "bpf_cubic",
                "owner": 0
            }
        }
    }
]

Signed-off-by: Martin KaFai Lau <kafai@fb.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Quentin Monnet <quentin@isovalent.com>
Link: https://lore.kernel.org/bpf/20200318171656.129650-1-kafai@fb.com
tools/bpf/bpftool/Documentation/bpftool-struct_ops.rst [new file with mode: 0644]
tools/bpf/bpftool/bash-completion/bpftool
tools/bpf/bpftool/main.c
tools/bpf/bpftool/main.h
tools/bpf/bpftool/struct_ops.c [new file with mode: 0644]

diff --git a/tools/bpf/bpftool/Documentation/bpftool-struct_ops.rst b/tools/bpf/bpftool/Documentation/bpftool-struct_ops.rst
new file mode 100644 (file)
index 0000000..f045cc8
--- /dev/null
@@ -0,0 +1,116 @@
+==================
+bpftool-struct_ops
+==================
+-------------------------------------------------------------------------------
+tool to register/unregister/introspect BPF struct_ops
+-------------------------------------------------------------------------------
+
+:Manual section: 8
+
+SYNOPSIS
+========
+
+       **bpftool** [*OPTIONS*] **struct_ops** *COMMAND*
+
+       *OPTIONS* := { { **-j** | **--json** } [{ **-p** | **--pretty** }] }
+
+       *COMMANDS* :=
+       { **show** | **list** | **dump** | **register** | **unregister** | **help** }
+
+STRUCT_OPS COMMANDS
+===================
+
+|      **bpftool** **struct_ops { show | list }** [*STRUCT_OPS_MAP*]
+|      **bpftool** **struct_ops dump** [*STRUCT_OPS_MAP*]
+|      **bpftool** **struct_ops register** *OBJ*
+|      **bpftool** **struct_ops unregister** *STRUCT_OPS_MAP*
+|      **bpftool** **struct_ops help**
+|
+|      *STRUCT_OPS_MAP* := { **id** *STRUCT_OPS_MAP_ID* | **name** *STRUCT_OPS_MAP_NAME* }
+|      *OBJ* := /a/file/of/bpf_struct_ops.o
+
+
+DESCRIPTION
+===========
+       **bpftool struct_ops { show | list }** [*STRUCT_OPS_MAP*]
+                 Show brief information about the struct_ops in the system.
+                 If *STRUCT_OPS_MAP* is specified, it shows information only
+                 for the given struct_ops.  Otherwise, it lists all struct_ops
+                 currently existing in the system.
+
+                 Output will start with struct_ops map ID, followed by its map
+                 name and its struct_ops's kernel type.
+
+       **bpftool struct_ops dump** [*STRUCT_OPS_MAP*]
+                 Dump details information about the struct_ops in the system.
+                 If *STRUCT_OPS_MAP* is specified, it dumps information only
+                 for the given struct_ops.  Otherwise, it dumps all struct_ops
+                 currently existing in the system.
+
+       **bpftool struct_ops register** *OBJ*
+                 Register bpf struct_ops from *OBJ*.  All struct_ops under
+                 the ELF section ".struct_ops" will be registered to
+                 its kernel subsystem.
+
+       **bpftool struct_ops unregister**  *STRUCT_OPS_MAP*
+                 Unregister the *STRUCT_OPS_MAP* from the kernel subsystem.
+
+       **bpftool struct_ops help**
+                 Print short help message.
+
+OPTIONS
+=======
+       -h, --help
+                 Print short generic help message (similar to **bpftool help**).
+
+       -V, --version
+                 Print version number (similar to **bpftool version**).
+
+       -j, --json
+                 Generate JSON output. For commands that cannot produce JSON, this
+                 option has no effect.
+
+       -p, --pretty
+                 Generate human-readable JSON output. Implies **-j**.
+
+       -d, --debug
+                 Print all logs available, even debug-level information. This
+                 includes logs from libbpf as well as from the verifier, when
+                 attempting to load programs.
+
+EXAMPLES
+========
+**# bpftool struct_ops show**
+
+::
+
+    100: dctcp           tcp_congestion_ops
+    105: cubic           tcp_congestion_ops
+
+**# bpftool struct_ops unregister id 105**
+
+::
+
+   Unregistered tcp_congestion_ops cubic id 105
+
+**# bpftool struct_ops register bpf_cubic.o**
+
+::
+
+   Registered tcp_congestion_ops cubic id 110
+
+
+SEE ALSO
+========
+       **bpf**\ (2),
+       **bpf-helpers**\ (7),
+       **bpftool**\ (8),
+       **bpftool-prog**\ (8),
+       **bpftool-map**\ (8),
+       **bpftool-cgroup**\ (8),
+       **bpftool-feature**\ (8),
+       **bpftool-net**\ (8),
+       **bpftool-perf**\ (8),
+       **bpftool-btf**\ (8)
+       **bpftool-gen**\ (8)
+       
index 9b0534f558f16188d49510f2fd94f976ff1e1cde..45ee99b159e222e38a01efd97110ce181cf337f0 100644 (file)
@@ -576,6 +576,34 @@ _bpftool()
                     ;;
             esac
             ;;
+        struct_ops)
+            local STRUCT_OPS_TYPE='id name'
+            case $command in
+                show|list|dump|unregister)
+                    case $prev in
+                        $command)
+                            COMPREPLY=( $( compgen -W "$STRUCT_OPS_TYPE" -- "$cur" ) )
+                            ;;
+                        id)
+                            _bpftool_get_map_ids_for_type struct_ops
+                            ;;
+                        name)
+                            _bpftool_get_map_names_for_type struct_ops
+                            ;;
+                    esac
+                    return 0
+                    ;;
+                register)
+                    _filedir
+                    return 0
+                    ;;
+                *)
+                    [[ $prev == $object ]] && \
+                        COMPREPLY=( $( compgen -W 'register unregister show list dump help' \
+                            -- "$cur" ) )
+                    ;;
+            esac
+            ;;
         map)
             local MAP_TYPE='id pinned name'
             case $command in
index 06449e846e4b8ba3b55fae7d8750337f8f050046..466c269eabdd8a28564788402bbefa0ffac7a020 100644 (file)
@@ -58,7 +58,7 @@ static int do_help(int argc, char **argv)
                "       %s batch file FILE\n"
                "       %s version\n"
                "\n"
-               "       OBJECT := { prog | map | cgroup | perf | net | feature | btf | gen }\n"
+               "       OBJECT := { prog | map | cgroup | perf | net | feature | btf | gen | struct_ops }\n"
                "       " HELP_SPEC_OPTIONS "\n"
                "",
                bin_name, bin_name, bin_name);
@@ -221,6 +221,7 @@ static const struct cmd cmds[] = {
        { "feature",    do_feature },
        { "btf",        do_btf },
        { "gen",        do_gen },
+       { "struct_ops", do_struct_ops },
        { "version",    do_version },
        { 0 }
 };
index 6db2398ae7e9adb258ec3f326f745dbf6a026537..86f14ce26fd7013aa0d931dfac9a66973cf7abc3 100644 (file)
@@ -161,6 +161,7 @@ int do_tracelog(int argc, char **arg);
 int do_feature(int argc, char **argv);
 int do_btf(int argc, char **argv);
 int do_gen(int argc, char **argv);
+int do_struct_ops(int argc, char **argv);
 
 int parse_u32_arg(int *argc, char ***argv, __u32 *val, const char *what);
 int prog_parse_fd(int *argc, char ***argv);
diff --git a/tools/bpf/bpftool/struct_ops.c b/tools/bpf/bpftool/struct_ops.c
new file mode 100644 (file)
index 0000000..2a7befb
--- /dev/null
@@ -0,0 +1,596 @@
+// SPDX-License-Identifier: (GPL-2.0-only OR BSD-2-Clause)
+/* Copyright (C) 2020 Facebook */
+
+#include <errno.h>
+#include <stdio.h>
+#include <unistd.h>
+
+#include <linux/err.h>
+
+#include <bpf/bpf.h>
+#include <bpf/btf.h>
+#include <bpf/libbpf.h>
+
+#include "json_writer.h"
+#include "main.h"
+
+#define STRUCT_OPS_VALUE_PREFIX "bpf_struct_ops_"
+
+static const struct btf_type *map_info_type;
+static __u32 map_info_alloc_len;
+static struct btf *btf_vmlinux;
+static __s32 map_info_type_id;
+
+struct res {
+       unsigned int nr_maps;
+       unsigned int nr_errs;
+};
+
+static const struct btf *get_btf_vmlinux(void)
+{
+       if (btf_vmlinux)
+               return btf_vmlinux;
+
+       btf_vmlinux = libbpf_find_kernel_btf();
+       if (IS_ERR(btf_vmlinux))
+               p_err("struct_ops requires kernel CONFIG_DEBUG_INFO_BTF=y");
+
+       return btf_vmlinux;
+}
+
+static const char *get_kern_struct_ops_name(const struct bpf_map_info *info)
+{
+       const struct btf *kern_btf;
+       const struct btf_type *t;
+       const char *st_ops_name;
+
+       kern_btf = get_btf_vmlinux();
+       if (IS_ERR(kern_btf))
+               return "<btf_vmlinux_not_found>";
+
+       t = btf__type_by_id(kern_btf, info->btf_vmlinux_value_type_id);
+       st_ops_name = btf__name_by_offset(kern_btf, t->name_off);
+       st_ops_name += strlen(STRUCT_OPS_VALUE_PREFIX);
+
+       return st_ops_name;
+}
+
+static __s32 get_map_info_type_id(void)
+{
+       const struct btf *kern_btf;
+
+       if (map_info_type_id)
+               return map_info_type_id;
+
+       kern_btf = get_btf_vmlinux();
+       if (IS_ERR(kern_btf)) {
+               map_info_type_id = PTR_ERR(kern_btf);
+               return map_info_type_id;
+       }
+
+       map_info_type_id = btf__find_by_name_kind(kern_btf, "bpf_map_info",
+                                                 BTF_KIND_STRUCT);
+       if (map_info_type_id < 0) {
+               p_err("can't find bpf_map_info from btf_vmlinux");
+               return map_info_type_id;
+       }
+       map_info_type = btf__type_by_id(kern_btf, map_info_type_id);
+
+       /* Ensure map_info_alloc() has at least what the bpftool needs */
+       map_info_alloc_len = map_info_type->size;
+       if (map_info_alloc_len < sizeof(struct bpf_map_info))
+               map_info_alloc_len = sizeof(struct bpf_map_info);
+
+       return map_info_type_id;
+}
+
+/* If the subcmd needs to print out the bpf_map_info,
+ * it should always call map_info_alloc to allocate
+ * a bpf_map_info object instead of allocating it
+ * on the stack.
+ *
+ * map_info_alloc() will take the running kernel's btf
+ * into account.  i.e. it will consider the
+ * sizeof(struct bpf_map_info) of the running kernel.
+ *
+ * It will enable the "struct_ops" cmd to print the latest
+ * "struct bpf_map_info".
+ *
+ * [ Recall that "struct_ops" requires the kernel's btf to
+ *   be available ]
+ */
+static struct bpf_map_info *map_info_alloc(__u32 *alloc_len)
+{
+       struct bpf_map_info *info;
+
+       if (get_map_info_type_id() < 0)
+               return NULL;
+
+       info = calloc(1, map_info_alloc_len);
+       if (!info)
+               p_err("mem alloc failed");
+       else
+               *alloc_len = map_info_alloc_len;
+
+       return info;
+}
+
+/* It iterates all struct_ops maps of the system.
+ * It returns the fd in "*res_fd" and map_info in "*info".
+ * In the very first iteration, info->id should be 0.
+ * An optional map "*name" filter can be specified.
+ * The filter can be made more flexible in the future.
+ * e.g. filter by kernel-struct-ops-name, regex-name, glob-name, ...etc.
+ *
+ * Return value:
+ *     1: A struct_ops map found.  It is returned in "*res_fd" and "*info".
+ *       The caller can continue to call get_next in the future.
+ *     0: No struct_ops map is returned.
+ *        All struct_ops map has been found.
+ *    -1: Error and the caller should abort the iteration.
+ */
+static int get_next_struct_ops_map(const char *name, int *res_fd,
+                                  struct bpf_map_info *info, __u32 info_len)
+{
+       __u32 id = info->id;
+       int err, fd;
+
+       while (true) {
+               err = bpf_map_get_next_id(id, &id);
+               if (err) {
+                       if (errno == ENOENT)
+                               return 0;
+                       p_err("can't get next map: %s", strerror(errno));
+                       return -1;
+               }
+
+               fd = bpf_map_get_fd_by_id(id);
+               if (fd < 0) {
+                       if (errno == ENOENT)
+                               continue;
+                       p_err("can't get map by id (%u): %s",
+                             id, strerror(errno));
+                       return -1;
+               }
+
+               err = bpf_obj_get_info_by_fd(fd, info, &info_len);
+               if (err) {
+                       p_err("can't get map info: %s", strerror(errno));
+                       close(fd);
+                       return -1;
+               }
+
+               if (info->type == BPF_MAP_TYPE_STRUCT_OPS &&
+                   (!name || !strcmp(name, info->name))) {
+                       *res_fd = fd;
+                       return 1;
+               }
+               close(fd);
+       }
+}
+
+static int cmd_retval(const struct res *res, bool must_have_one_map)
+{
+       if (res->nr_errs || (!res->nr_maps && must_have_one_map))
+               return -1;
+
+       return 0;
+}
+
+/* "data" is the work_func private storage */
+typedef int (*work_func)(int fd, const struct bpf_map_info *info, void *data,
+                        struct json_writer *wtr);
+
+/* Find all struct_ops map in the system.
+ * Filter out by "name" (if specified).
+ * Then call "func(fd, info, data, wtr)" on each struct_ops map found.
+ */
+static struct res do_search(const char *name, work_func func, void *data,
+                           struct json_writer *wtr)
+{
+       struct bpf_map_info *info;
+       struct res res = {};
+       __u32 info_len;
+       int fd, err;
+
+       info = map_info_alloc(&info_len);
+       if (!info) {
+               res.nr_errs++;
+               return res;
+       }
+
+       if (wtr)
+               jsonw_start_array(wtr);
+       while ((err = get_next_struct_ops_map(name, &fd, info, info_len)) == 1) {
+               res.nr_maps++;
+               err = func(fd, info, data, wtr);
+               if (err)
+                       res.nr_errs++;
+               close(fd);
+       }
+       if (wtr)
+               jsonw_end_array(wtr);
+
+       if (err)
+               res.nr_errs++;
+
+       if (!wtr && name && !res.nr_errs && !res.nr_maps)
+               /* It is not printing empty [].
+                * Thus, needs to specifically say nothing found
+                * for "name" here.
+                */
+               p_err("no struct_ops found for %s", name);
+       else if (!wtr && json_output && !res.nr_errs)
+               /* The "func()" above is not writing any json (i.e. !wtr
+                * test here).
+                *
+                * However, "-j" is enabled and there is no errs here,
+                * so call json_null() as the current convention of
+                * other cmds.
+                */
+               jsonw_null(json_wtr);
+
+       free(info);
+       return res;
+}
+
+static struct res do_one_id(const char *id_str, work_func func, void *data,
+                           struct json_writer *wtr)
+{
+       struct bpf_map_info *info;
+       struct res res = {};
+       unsigned long id;
+       __u32 info_len;
+       char *endptr;
+       int fd;
+
+       id = strtoul(id_str, &endptr, 0);
+       if (*endptr || !id || id > UINT32_MAX) {
+               p_err("invalid id %s", id_str);
+               res.nr_errs++;
+               return res;
+       }
+
+       fd = bpf_map_get_fd_by_id(id);
+       if (fd == -1) {
+               p_err("can't get map by id (%lu): %s", id, strerror(errno));
+               res.nr_errs++;
+               return res;
+       }
+
+       info = map_info_alloc(&info_len);
+       if (!info) {
+               res.nr_errs++;
+               goto done;
+       }
+
+       if (bpf_obj_get_info_by_fd(fd, info, &info_len)) {
+               p_err("can't get map info: %s", strerror(errno));
+               res.nr_errs++;
+               goto done;
+       }
+
+       if (info->type != BPF_MAP_TYPE_STRUCT_OPS) {
+               p_err("%s id %u is not a struct_ops map", info->name, info->id);
+               res.nr_errs++;
+               goto done;
+       }
+
+       res.nr_maps++;
+
+       if (func(fd, info, data, wtr))
+               res.nr_errs++;
+       else if (!wtr && json_output)
+               /* The "func()" above is not writing any json (i.e. !wtr
+                * test here).
+                *
+                * However, "-j" is enabled and there is no errs here,
+                * so call json_null() as the current convention of
+                * other cmds.
+                */
+               jsonw_null(json_wtr);
+
+done:
+       free(info);
+       close(fd);
+
+       return res;
+}
+
+static struct res do_work_on_struct_ops(const char *search_type,
+                                       const char *search_term,
+                                       work_func func, void *data,
+                                       struct json_writer *wtr)
+{
+       if (search_type) {
+               if (is_prefix(search_type, "id"))
+                       return do_one_id(search_term, func, data, wtr);
+               else if (!is_prefix(search_type, "name"))
+                       usage();
+       }
+
+       return do_search(search_term, func, data, wtr);
+}
+
+static int __do_show(int fd, const struct bpf_map_info *info, void *data,
+                    struct json_writer *wtr)
+{
+       if (wtr) {
+               jsonw_start_object(wtr);
+               jsonw_uint_field(wtr, "id", info->id);
+               jsonw_string_field(wtr, "name", info->name);
+               jsonw_string_field(wtr, "kernel_struct_ops",
+                                  get_kern_struct_ops_name(info));
+               jsonw_end_object(wtr);
+       } else {
+               printf("%u: %-15s %-32s\n", info->id, info->name,
+                      get_kern_struct_ops_name(info));
+       }
+
+       return 0;
+}
+
+static int do_show(int argc, char **argv)
+{
+       const char *search_type = NULL, *search_term = NULL;
+       struct res res;
+
+       if (argc && argc != 2)
+               usage();
+
+       if (argc == 2) {
+               search_type = GET_ARG();
+               search_term = GET_ARG();
+       }
+
+       res = do_work_on_struct_ops(search_type, search_term, __do_show,
+                                   NULL, json_wtr);
+
+       return cmd_retval(&res, !!search_term);
+}
+
+static int __do_dump(int fd, const struct bpf_map_info *info, void *data,
+                    struct json_writer *wtr)
+{
+       struct btf_dumper *d = (struct btf_dumper *)data;
+       const struct btf_type *struct_ops_type;
+       const struct btf *kern_btf = d->btf;
+       const char *struct_ops_name;
+       int zero = 0;
+       void *value;
+
+       /* note: d->jw == wtr */
+
+       kern_btf = d->btf;
+
+       /* The kernel supporting BPF_MAP_TYPE_STRUCT_OPS must have
+        * btf_vmlinux_value_type_id.
+        */
+       struct_ops_type = btf__type_by_id(kern_btf,
+                                         info->btf_vmlinux_value_type_id);
+       struct_ops_name = btf__name_by_offset(kern_btf,
+                                             struct_ops_type->name_off);
+       value = calloc(1, info->value_size);
+       if (!value) {
+               p_err("mem alloc failed");
+               return -1;
+       }
+
+       if (bpf_map_lookup_elem(fd, &zero, value)) {
+               p_err("can't lookup struct_ops map %s id %u",
+                     info->name, info->id);
+               free(value);
+               return -1;
+       }
+
+       jsonw_start_object(wtr);
+       jsonw_name(wtr, "bpf_map_info");
+       btf_dumper_type(d, map_info_type_id, (void *)info);
+       jsonw_end_object(wtr);
+
+       jsonw_start_object(wtr);
+       jsonw_name(wtr, struct_ops_name);
+       btf_dumper_type(d, info->btf_vmlinux_value_type_id, value);
+       jsonw_end_object(wtr);
+
+       free(value);
+
+       return 0;
+}
+
+static int do_dump(int argc, char **argv)
+{
+       const char *search_type = NULL, *search_term = NULL;
+       json_writer_t *wtr = json_wtr;
+       const struct btf *kern_btf;
+       struct btf_dumper d = {};
+       struct res res;
+
+       if (argc && argc != 2)
+               usage();
+
+       if (argc == 2) {
+               search_type = GET_ARG();
+               search_term = GET_ARG();
+       }
+
+       kern_btf = get_btf_vmlinux();
+       if (IS_ERR(kern_btf))
+               return -1;
+
+       if (!json_output) {
+               wtr = jsonw_new(stdout);
+               if (!wtr) {
+                       p_err("can't create json writer");
+                       return -1;
+               }
+               jsonw_pretty(wtr, true);
+       }
+
+       d.btf = kern_btf;
+       d.jw = wtr;
+       d.is_plain_text = !json_output;
+       d.prog_id_as_func_ptr = true;
+
+       res = do_work_on_struct_ops(search_type, search_term, __do_dump, &d,
+                                   wtr);
+
+       if (!json_output)
+               jsonw_destroy(&wtr);
+
+       return cmd_retval(&res, !!search_term);
+}
+
+static int __do_unregister(int fd, const struct bpf_map_info *info, void *data,
+                          struct json_writer *wtr)
+{
+       int zero = 0;
+
+       if (bpf_map_delete_elem(fd, &zero)) {
+               p_err("can't unload %s %s id %u: %s",
+                     get_kern_struct_ops_name(info), info->name,
+                     info->id, strerror(errno));
+               return -1;
+       }
+
+       p_info("Unregistered %s %s id %u",
+              get_kern_struct_ops_name(info), info->name,
+              info->id);
+
+       return 0;
+}
+
+static int do_unregister(int argc, char **argv)
+{
+       const char *search_type, *search_term;
+       struct res res;
+
+       if (argc != 2)
+               usage();
+
+       search_type = GET_ARG();
+       search_term = GET_ARG();
+
+       res = do_work_on_struct_ops(search_type, search_term,
+                                   __do_unregister, NULL, NULL);
+
+       return cmd_retval(&res, true);
+}
+
+static int do_register(int argc, char **argv)
+{
+       const struct bpf_map_def *def;
+       struct bpf_map_info info = {};
+       __u32 info_len = sizeof(info);
+       int nr_errs = 0, nr_maps = 0;
+       struct bpf_object *obj;
+       struct bpf_link *link;
+       struct bpf_map *map;
+       const char *file;
+
+       if (argc != 1)
+               usage();
+
+       file = GET_ARG();
+
+       obj = bpf_object__open(file);
+       if (IS_ERR_OR_NULL(obj))
+               return -1;
+
+       set_max_rlimit();
+
+       if (bpf_object__load(obj)) {
+               bpf_object__close(obj);
+               return -1;
+       }
+
+       bpf_object__for_each_map(map, obj) {
+               def = bpf_map__def(map);
+               if (def->type != BPF_MAP_TYPE_STRUCT_OPS)
+                       continue;
+
+               link = bpf_map__attach_struct_ops(map);
+               if (IS_ERR(link)) {
+                       p_err("can't register struct_ops %s: %s",
+                             bpf_map__name(map),
+                             strerror(-PTR_ERR(link)));
+                       nr_errs++;
+                       continue;
+               }
+               nr_maps++;
+
+               bpf_link__disconnect(link);
+               bpf_link__destroy(link);
+
+               if (!bpf_obj_get_info_by_fd(bpf_map__fd(map), &info,
+                                           &info_len))
+                       p_info("Registered %s %s id %u",
+                              get_kern_struct_ops_name(&info),
+                              bpf_map__name(map),
+                              info.id);
+               else
+                       /* Not p_err.  The struct_ops was attached
+                        * successfully.
+                        */
+                       p_info("Registered %s but can't find id: %s",
+                              bpf_map__name(map), strerror(errno));
+       }
+
+       bpf_object__close(obj);
+
+       if (nr_errs)
+               return -1;
+
+       if (!nr_maps) {
+               p_err("no struct_ops found in %s", file);
+               return -1;
+       }
+
+       if (json_output)
+               jsonw_null(json_wtr);
+
+       return 0;
+}
+
+static int do_help(int argc, char **argv)
+{
+       if (json_output) {
+               jsonw_null(json_wtr);
+               return 0;
+       }
+
+       fprintf(stderr,
+               "Usage: %s %s { show | list } [STRUCT_OPS_MAP]\n"
+               "       %s %s dump [STRUCT_OPS_MAP]\n"
+               "       %s %s register OBJ\n"
+               "       %s %s unregister STRUCT_OPS_MAP\n"
+               "       %s %s help\n"
+               "\n"
+               "       OPTIONS := { {-j|--json} [{-p|--pretty}] }\n"
+               "       STRUCT_OPS_MAP := [ id STRUCT_OPS_MAP_ID | name STRUCT_OPS_MAP_NAME ]\n",
+               bin_name, argv[-2], bin_name, argv[-2],
+               bin_name, argv[-2], bin_name, argv[-2],
+               bin_name, argv[-2]);
+
+       return 0;
+}
+
+static const struct cmd cmds[] = {
+       { "show",       do_show },
+       { "list",       do_show },
+       { "register",   do_register },
+       { "unregister", do_unregister },
+       { "dump",       do_dump },
+       { "help",       do_help },
+       { 0 }
+};
+
+int do_struct_ops(int argc, char **argv)
+{
+       int err;
+
+       err = cmd_select(cmds, argc, argv, do_help);
+
+       btf__free(btf_vmlinux);
+       return err;
+}