src/basic/bpf-program.c

   1 /* SPDX-License-Identifier: LGPL-2.1+ */
   2
   3 #include <fcntl.h>
   4 #include <sys/stat.h>
   5 #include <sys/types.h>
   6 #include <unistd.h>
   7
   8 #include "alloc-util.h"
   9 #include "bpf-program.h"
  10 #include "fd-util.h"
  11 #include "log.h"
  12 #include "missing.h"
  13 #include "path-util.h"
  14 #include "util.h"
  15
  16 int bpf_program_new(uint32_t prog_type, BPFProgram **ret) {
  17         _cleanup_(bpf_program_unrefp) BPFProgram *p = NULL;
  18
  19         p = new0(BPFProgram, 1);
  20         if (!p)
  21                 return log_oom();
  22
  23         p->n_ref = 1;
  24         p->prog_type = prog_type;
  25         p->kernel_fd = -1;
  26
  27         *ret = TAKE_PTR(p);
  28
  29         return 0;
  30 }
  31
  32 BPFProgram *bpf_program_ref(BPFProgram *p) {
  33         if (!p)
  34                 return NULL;
  35
  36         assert(p->n_ref > 0);
  37         p->n_ref++;
  38
  39         return p;
  40 }
  41
  42 BPFProgram *bpf_program_unref(BPFProgram *p) {
  43         if (!p)
  44                 return NULL;
  45
  46         assert(p->n_ref > 0);
  47         p->n_ref--;
  48
  49         if (p->n_ref > 0)
  50                 return NULL;
  51
  52         /* Unfortunately, the kernel currently doesn't implicitly detach BPF programs from their cgroups when the last
  53          * fd to the BPF program is closed. This has nasty side-effects since this means that abnormally terminated
  54          * programs that attached one of their BPF programs to a cgroup will leave this programs pinned for good with
  55          * zero chance of recovery, until the cgroup is removed. This is particularly problematic if the cgroup in
  56          * question is the root cgroup (or any other cgroup belonging to a service that cannot be restarted during
  57          * operation, such as dbus), as the memory for the BPF program can only be reclaimed through a reboot. To
  58          * counter this, we track closely to which cgroup a program was attached to and will detach it on our own
  59          * whenever we close the BPF fd. */
  60         (void) bpf_program_cgroup_detach(p);
  61
  62         safe_close(p->kernel_fd);
  63         free(p->instructions);
  64         free(p->attached_path);
  65
  66         return mfree(p);
  67 }
  68
  69 int bpf_program_add_instructions(BPFProgram *p, const struct bpf_insn *instructions, size_t count) {
  70
  71         assert(p);
  72
  73         if (p->kernel_fd >= 0) /* don't allow modification after we uploaded things to the kernel */
  74                 return -EBUSY;
  75
  76         if (!GREEDY_REALLOC(p->instructions, p->allocated, p->n_instructions + count))
  77                 return -ENOMEM;
  78
  79         memcpy(p->instructions + p->n_instructions, instructions, sizeof(struct bpf_insn) * count);
  80         p->n_instructions += count;
  81
  82         return 0;
  83 }
  84
  85 int bpf_program_load_kernel(BPFProgram *p, char *log_buf, size_t log_size) {
  86         union bpf_attr attr;
  87
  88         assert(p);
  89
  90         if (p->kernel_fd >= 0) { /* make this idempotent */
  91                 memzero(log_buf, log_size);
  92                 return 0;
  93         }
  94
  95         attr = (union bpf_attr) {
  96                 .prog_type = p->prog_type,
  97                 .insns = PTR_TO_UINT64(p->instructions),
  98                 .insn_cnt = p->n_instructions,
  99                 .license = PTR_TO_UINT64("GPL"),
 100                 .log_buf = PTR_TO_UINT64(log_buf),
 101                 .log_level = !!log_buf,
 102                 .log_size = log_size,
 103         };
 104
 105         p->kernel_fd = bpf(BPF_PROG_LOAD, &attr, sizeof(attr));
 106         if (p->kernel_fd < 0)
 107                 return -errno;
 108
 109         return 0;
 110 }
 111
 112 int bpf_program_cgroup_attach(BPFProgram *p, int type, const char *path, uint32_t flags) {
 113         _cleanup_free_ char *copy = NULL;
 114         _cleanup_close_ int fd = -1;
 115         union bpf_attr attr;
 116         int r;
 117
 118         assert(p);
 119         assert(type >= 0);
 120         assert(path);
 121
 122         if (!IN_SET(flags, 0, BPF_F_ALLOW_OVERRIDE, BPF_F_ALLOW_MULTI))
 123                 return -EINVAL;
 124
 125         /* We need to track which cgroup the program is attached to, and we can only track one attachment, hence let's
 126         * refuse this early. */
 127         if (p->attached_path) {
 128                 if (!path_equal(p->attached_path, path))
 129                         return -EBUSY;
 130                 if (p->attached_type != type)
 131                         return -EBUSY;
 132                 if (p->attached_flags != flags)
 133                         return -EBUSY;
 134
 135                 /* Here's a shortcut: if we previously attached this program already, then we don't have to do so
 136                  * again. Well, with one exception: if we are in BPF_F_ALLOW_OVERRIDE mode then someone else might have
 137                  * replaced our program since the last time, hence let's reattach it again, just to be safe. In flags
 138                  * == 0 mode this is not an issue since nobody else can replace our program in that case, and in flags
 139                  * == BPF_F_ALLOW_MULTI mode any other's program would be installed in addition to ours hence ours
 140                  * would remain in effect. */
 141                 if (flags != BPF_F_ALLOW_OVERRIDE)
 142                         return 0;
 143         }
 144
 145         /* Ensure we have a kernel object for this. */
 146         r = bpf_program_load_kernel(p, NULL, 0);
 147         if (r < 0)
 148                 return r;
 149
 150         copy = strdup(path);
 151         if (!copy)
 152                 return -ENOMEM;
 153
 154         fd = open(path, O_DIRECTORY|O_RDONLY|O_CLOEXEC);
 155         if (fd < 0)
 156                 return -errno;
 157
 158         attr = (union bpf_attr) {
 159                 .attach_type = type,
 160                 .target_fd = fd,
 161                 .attach_bpf_fd = p->kernel_fd,
 162                 .attach_flags = flags,
 163         };
 164
 165         if (bpf(BPF_PROG_ATTACH, &attr, sizeof(attr)) < 0)
 166                 return -errno;
 167
 168         free_and_replace(p->attached_path, copy);
 169         p->attached_type = type;
 170         p->attached_flags = flags;
 171
 172         return 0;
 173 }
 174
 175 int bpf_program_cgroup_detach(BPFProgram *p) {
 176         _cleanup_close_ int fd = -1;
 177
 178         assert(p);
 179
 180         if (!p->attached_path)
 181                 return -EUNATCH;
 182
 183         fd = open(p->attached_path, O_DIRECTORY|O_RDONLY|O_CLOEXEC);
 184         if (fd < 0) {
 185                 if (errno != ENOENT)
 186                         return -errno;
 187
 188                 /* If the cgroup does not exist anymore, then we don't have to explicitly detach, it got detached
 189                  * implicitly by the removal, hence don't complain */
 190
 191         } else {
 192                 union bpf_attr attr;
 193
 194                 attr = (union bpf_attr) {
 195                         .attach_type = p->attached_type,
 196                         .target_fd = fd,
 197                         .attach_bpf_fd = p->kernel_fd,
 198                 };
 199
 200                 if (bpf(BPF_PROG_DETACH, &attr, sizeof(attr)) < 0)
 201                         return -errno;
 202         }
 203
 204         p->attached_path = mfree(p->attached_path);
 205
 206         return 0;
 207 }
 208
 209 int bpf_map_new(enum bpf_map_type type, size_t key_size, size_t value_size, size_t max_entries, uint32_t flags) {
 210         union bpf_attr attr = {
 211                 .map_type = type,
 212                 .key_size = key_size,
 213                 .value_size = value_size,
 214                 .max_entries = max_entries,
 215                 .map_flags = flags,
 216         };
 217         int fd;
 218
 219         fd = bpf(BPF_MAP_CREATE, &attr, sizeof(attr));
 220         if (fd < 0)
 221                 return -errno;
 222
 223         return fd;
 224 }
 225
 226 int bpf_map_update_element(int fd, const void *key, void *value) {
 227
 228         union bpf_attr attr = {
 229                 .map_fd = fd,
 230                 .key = PTR_TO_UINT64(key),
 231                 .value = PTR_TO_UINT64(value),
 232         };
 233
 234         if (bpf(BPF_MAP_UPDATE_ELEM, &attr, sizeof(attr)) < 0)
 235                 return -errno;
 236
 237         return 0;
 238 }
 239
 240 int bpf_map_lookup_element(int fd, const void *key, void *value) {
 241
 242         union bpf_attr attr = {
 243                 .map_fd = fd,
 244                 .key = PTR_TO_UINT64(key),
 245                 .value = PTR_TO_UINT64(value),
 246         };
 247
 248         if (bpf(BPF_MAP_LOOKUP_ELEM, &attr, sizeof(attr)) < 0)
 249                 return -errno;
 250
 251         return 0;
 252 }