kernel/bpf/bpf_struct_ops.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /* Copyright (c) 2019 Facebook */
   3
   4 #include <linux/bpf.h>
   5 #include <linux/bpf_verifier.h>
   6 #include <linux/btf.h>
   7 #include <linux/filter.h>
   8 #include <linux/slab.h>
   9 #include <linux/numa.h>
  10 #include <linux/seq_file.h>
  11 #include <linux/refcount.h>
  12 #include <linux/mutex.h>
  13
  14 enum bpf_struct_ops_state {
  15         BPF_STRUCT_OPS_STATE_INIT,
  16         BPF_STRUCT_OPS_STATE_INUSE,
  17         BPF_STRUCT_OPS_STATE_TOBEFREE,
  18 };
  19
  20 #define BPF_STRUCT_OPS_COMMON_VALUE                     \
  21         refcount_t refcnt;                              \
  22         enum bpf_struct_ops_state state
  23
  24 struct bpf_struct_ops_value {
  25         BPF_STRUCT_OPS_COMMON_VALUE;
  26         char data[] ____cacheline_aligned_in_smp;
  27 };
  28
  29 struct bpf_struct_ops_map {
  30         struct bpf_map map;
  31         struct rcu_head rcu;
  32         const struct bpf_struct_ops *st_ops;
  33         /* protect map_update */
  34         struct mutex lock;
  35         /* progs has all the bpf_prog that is populated
  36          * to the func ptr of the kernel's struct
  37          * (in kvalue.data).
  38          */
  39         struct bpf_prog **progs;
  40         /* image is a page that has all the trampolines
  41          * that stores the func args before calling the bpf_prog.
  42          * A PAGE_SIZE "image" is enough to store all trampoline for
  43          * "progs[]".
  44          */
  45         void *image;
  46         /* uvalue->data stores the kernel struct
  47          * (e.g. tcp_congestion_ops) that is more useful
  48          * to userspace than the kvalue.  For example,
  49          * the bpf_prog's id is stored instead of the kernel
  50          * address of a func ptr.
  51          */
  52         struct bpf_struct_ops_value *uvalue;
  53         /* kvalue.data stores the actual kernel's struct
  54          * (e.g. tcp_congestion_ops) that will be
  55          * registered to the kernel subsystem.
  56          */
  57         struct bpf_struct_ops_value kvalue;
  58 };
  59
  60 #define VALUE_PREFIX "bpf_struct_ops_"
  61 #define VALUE_PREFIX_LEN (sizeof(VALUE_PREFIX) - 1)
  62
  63 /* bpf_struct_ops_##_name (e.g. bpf_struct_ops_tcp_congestion_ops) is
  64  * the map's value exposed to the userspace and its btf-type-id is
  65  * stored at the map->btf_vmlinux_value_type_id.
  66  *
  67  */
  68 #define BPF_STRUCT_OPS_TYPE(_name)                              \
  69 extern struct bpf_struct_ops bpf_##_name;                       \
  70                                                                 \
  71 struct bpf_struct_ops_##_name {                                         \
  72         BPF_STRUCT_OPS_COMMON_VALUE;                            \
  73         struct _name data ____cacheline_aligned_in_smp;         \
  74 };
  75 #include "bpf_struct_ops_types.h"
  76 #undef BPF_STRUCT_OPS_TYPE
  77
  78 enum {
  79 #define BPF_STRUCT_OPS_TYPE(_name) BPF_STRUCT_OPS_TYPE_##_name,
  80 #include "bpf_struct_ops_types.h"
  81 #undef BPF_STRUCT_OPS_TYPE
  82         __NR_BPF_STRUCT_OPS_TYPE,
  83 };
  84
  85 static struct bpf_struct_ops * const bpf_struct_ops[] = {
  86 #define BPF_STRUCT_OPS_TYPE(_name)                              \
  87         [BPF_STRUCT_OPS_TYPE_##_name] = &bpf_##_name,
  88 #include "bpf_struct_ops_types.h"
  89 #undef BPF_STRUCT_OPS_TYPE
  90 };
  91
  92 const struct bpf_verifier_ops bpf_struct_ops_verifier_ops = {
  93 };
  94
  95 const struct bpf_prog_ops bpf_struct_ops_prog_ops = {
  96 };
  97
  98 static const struct btf_type *module_type;
  99
 100 void bpf_struct_ops_init(struct btf *btf, struct bpf_verifier_log *log)
 101 {
 102         s32 type_id, value_id, module_id;
 103         const struct btf_member *member;
 104         struct bpf_struct_ops *st_ops;
 105         const struct btf_type *t;
 106         char value_name[128];
 107         const char *mname;
 108         u32 i, j;
 109
 110         /* Ensure BTF type is emitted for "struct bpf_struct_ops_##_name" */
 111 #define BPF_STRUCT_OPS_TYPE(_name) BTF_TYPE_EMIT(struct bpf_struct_ops_##_name);
 112 #include "bpf_struct_ops_types.h"
 113 #undef BPF_STRUCT_OPS_TYPE
 114
 115         module_id = btf_find_by_name_kind(btf, "module", BTF_KIND_STRUCT);
 116         if (module_id < 0) {
 117                 pr_warn("Cannot find struct module in btf_vmlinux\n");
 118                 return;
 119         }
 120         module_type = btf_type_by_id(btf, module_id);
 121
 122         for (i = 0; i < ARRAY_SIZE(bpf_struct_ops); i++) {
 123                 st_ops = bpf_struct_ops[i];
 124
 125                 if (strlen(st_ops->name) + VALUE_PREFIX_LEN >=
 126                     sizeof(value_name)) {
 127                         pr_warn("struct_ops name %s is too long\n",
 128                                 st_ops->name);
 129                         continue;
 130                 }
 131                 sprintf(value_name, "%s%s", VALUE_PREFIX, st_ops->name);
 132
 133                 value_id = btf_find_by_name_kind(btf, value_name,
 134                                                  BTF_KIND_STRUCT);
 135                 if (value_id < 0) {
 136                         pr_warn("Cannot find struct %s in btf_vmlinux\n",
 137                                 value_name);
 138                         continue;
 139                 }
 140
 141                 type_id = btf_find_by_name_kind(btf, st_ops->name,
 142                                                 BTF_KIND_STRUCT);
 143                 if (type_id < 0) {
 144                         pr_warn("Cannot find struct %s in btf_vmlinux\n",
 145                                 st_ops->name);
 146                         continue;
 147                 }
 148                 t = btf_type_by_id(btf, type_id);
 149                 if (btf_type_vlen(t) > BPF_STRUCT_OPS_MAX_NR_MEMBERS) {
 150                         pr_warn("Cannot support #%u members in struct %s\n",
 151                                 btf_type_vlen(t), st_ops->name);
 152                         continue;
 153                 }
 154
 155                 for_each_member(j, t, member) {
 156                         const struct btf_type *func_proto;
 157
 158                         mname = btf_name_by_offset(btf, member->name_off);
 159                         if (!*mname) {
 160                                 pr_warn("anon member in struct %s is not supported\n",
 161                                         st_ops->name);
 162                                 break;
 163                         }
 164
 165                         if (btf_member_bitfield_size(t, member)) {
 166                                 pr_warn("bit field member %s in struct %s is not supported\n",
 167                                         mname, st_ops->name);
 168                                 break;
 169                         }
 170
 171                         func_proto = btf_type_resolve_func_ptr(btf,
 172                                                                member->type,
 173                                                                NULL);
 174                         if (func_proto &&
 175                             btf_distill_func_proto(log, btf,
 176                                                    func_proto, mname,
 177                                                    &st_ops->func_models[j])) {
 178                                 pr_warn("Error in parsing func ptr %s in struct %s\n",
 179                                         mname, st_ops->name);
 180                                 break;
 181                         }
 182                 }
 183
 184                 if (j == btf_type_vlen(t)) {
 185                         if (st_ops->init(btf)) {
 186                                 pr_warn("Error in init bpf_struct_ops %s\n",
 187                                         st_ops->name);
 188                         } else {
 189                                 st_ops->type_id = type_id;
 190                                 st_ops->type = t;
 191                                 st_ops->value_id = value_id;
 192                                 st_ops->value_type = btf_type_by_id(btf,
 193                                                                     value_id);
 194                         }
 195                 }
 196         }
 197 }
 198
 199 extern struct btf *btf_vmlinux;
 200
 201 static const struct bpf_struct_ops *
 202 bpf_struct_ops_find_value(u32 value_id)
 203 {
 204         unsigned int i;
 205
 206         if (!value_id || !btf_vmlinux)
 207                 return NULL;
 208
 209         for (i = 0; i < ARRAY_SIZE(bpf_struct_ops); i++) {
 210                 if (bpf_struct_ops[i]->value_id == value_id)
 211                         return bpf_struct_ops[i];
 212         }
 213
 214         return NULL;
 215 }
 216
 217 const struct bpf_struct_ops *bpf_struct_ops_find(u32 type_id)
 218 {
 219         unsigned int i;
 220
 221         if (!type_id || !btf_vmlinux)
 222                 return NULL;
 223
 224         for (i = 0; i < ARRAY_SIZE(bpf_struct_ops); i++) {
 225                 if (bpf_struct_ops[i]->type_id == type_id)
 226                         return bpf_struct_ops[i];
 227         }
 228
 229         return NULL;
 230 }
 231
 232 static int bpf_struct_ops_map_get_next_key(struct bpf_map *map, void *key,
 233                                            void *next_key)
 234 {
 235         if (key && *(u32 *)key == 0)
 236                 return -ENOENT;
 237
 238         *(u32 *)next_key = 0;
 239         return 0;
 240 }
 241
 242 int bpf_struct_ops_map_sys_lookup_elem(struct bpf_map *map, void *key,
 243                                        void *value)
 244 {
 245         struct bpf_struct_ops_map *st_map = (struct bpf_struct_ops_map *)map;
 246         struct bpf_struct_ops_value *uvalue, *kvalue;
 247         enum bpf_struct_ops_state state;
 248
 249         if (unlikely(*(u32 *)key != 0))
 250                 return -ENOENT;
 251
 252         kvalue = &st_map->kvalue;
 253         /* Pair with smp_store_release() during map_update */
 254         state = smp_load_acquire(&kvalue->state);
 255         if (state == BPF_STRUCT_OPS_STATE_INIT) {
 256                 memset(value, 0, map->value_size);
 257                 return 0;
 258         }
 259
 260         /* No lock is needed.  state and refcnt do not need
 261          * to be updated together under atomic context.
 262          */
 263         uvalue = (struct bpf_struct_ops_value *)value;
 264         memcpy(uvalue, st_map->uvalue, map->value_size);
 265         uvalue->state = state;
 266         refcount_set(&uvalue->refcnt, refcount_read(&kvalue->refcnt));
 267
 268         return 0;
 269 }
 270
 271 static void *bpf_struct_ops_map_lookup_elem(struct bpf_map *map, void *key)
 272 {
 273         return ERR_PTR(-EINVAL);
 274 }
 275
 276 static void bpf_struct_ops_map_put_progs(struct bpf_struct_ops_map *st_map)
 277 {
 278         const struct btf_type *t = st_map->st_ops->type;
 279         u32 i;
 280
 281         for (i = 0; i < btf_type_vlen(t); i++) {
 282                 if (st_map->progs[i]) {
 283                         bpf_prog_put(st_map->progs[i]);
 284                         st_map->progs[i] = NULL;
 285                 }
 286         }
 287 }
 288
 289 static int check_zero_holes(const struct btf_type *t, void *data)
 290 {
 291         const struct btf_member *member;
 292         u32 i, moff, msize, prev_mend = 0;
 293         const struct btf_type *mtype;
 294
 295         for_each_member(i, t, member) {
 296                 moff = btf_member_bit_offset(t, member) / 8;
 297                 if (moff > prev_mend &&
 298                     memchr_inv(data + prev_mend, 0, moff - prev_mend))
 299                         return -EINVAL;
 300
 301                 mtype = btf_type_by_id(btf_vmlinux, member->type);
 302                 mtype = btf_resolve_size(btf_vmlinux, mtype, &msize);
 303                 if (IS_ERR(mtype))
 304                         return PTR_ERR(mtype);
 305                 prev_mend = moff + msize;
 306         }
 307
 308         if (t->size > prev_mend &&
 309             memchr_inv(data + prev_mend, 0, t->size - prev_mend))
 310                 return -EINVAL;
 311
 312         return 0;
 313 }
 314
 315 static int bpf_struct_ops_map_update_elem(struct bpf_map *map, void *key,
 316                                           void *value, u64 flags)
 317 {
 318         struct bpf_struct_ops_map *st_map = (struct bpf_struct_ops_map *)map;
 319         const struct bpf_struct_ops *st_ops = st_map->st_ops;
 320         struct bpf_struct_ops_value *uvalue, *kvalue;
 321         const struct btf_member *member;
 322         const struct btf_type *t = st_ops->type;
 323         struct bpf_tramp_progs *tprogs = NULL;
 324         void *udata, *kdata;
 325         int prog_fd, err = 0;
 326         void *image;
 327         u32 i;
 328
 329         if (flags)
 330                 return -EINVAL;
 331
 332         if (*(u32 *)key != 0)
 333                 return -E2BIG;
 334
 335         err = check_zero_holes(st_ops->value_type, value);
 336         if (err)
 337                 return err;
 338
 339         uvalue = (struct bpf_struct_ops_value *)value;
 340         err = check_zero_holes(t, uvalue->data);
 341         if (err)
 342                 return err;
 343
 344         if (uvalue->state || refcount_read(&uvalue->refcnt))
 345                 return -EINVAL;
 346
 347         tprogs = kcalloc(BPF_TRAMP_MAX, sizeof(*tprogs), GFP_KERNEL);
 348         if (!tprogs)
 349                 return -ENOMEM;
 350
 351         uvalue = (struct bpf_struct_ops_value *)st_map->uvalue;
 352         kvalue = (struct bpf_struct_ops_value *)&st_map->kvalue;
 353
 354         mutex_lock(&st_map->lock);
 355
 356         if (kvalue->state != BPF_STRUCT_OPS_STATE_INIT) {
 357                 err = -EBUSY;
 358                 goto unlock;
 359         }
 360
 361         memcpy(uvalue, value, map->value_size);
 362
 363         udata = &uvalue->data;
 364         kdata = &kvalue->data;
 365         image = st_map->image;
 366
 367         for_each_member(i, t, member) {
 368                 const struct btf_type *mtype, *ptype;
 369                 struct bpf_prog *prog;
 370                 u32 moff;
 371
 372                 moff = btf_member_bit_offset(t, member) / 8;
 373                 ptype = btf_type_resolve_ptr(btf_vmlinux, member->type, NULL);
 374                 if (ptype == module_type) {
 375                         if (*(void **)(udata + moff))
 376                                 goto reset_unlock;
 377                         *(void **)(kdata + moff) = BPF_MODULE_OWNER;
 378                         continue;
 379                 }
 380
 381                 err = st_ops->init_member(t, member, kdata, udata);
 382                 if (err < 0)
 383                         goto reset_unlock;
 384
 385                 /* The ->init_member() has handled this member */
 386                 if (err > 0)
 387                         continue;
 388
 389                 /* If st_ops->init_member does not handle it,
 390                  * we will only handle func ptrs and zero-ed members
 391                  * here.  Reject everything else.
 392                  */
 393
 394                 /* All non func ptr member must be 0 */
 395                 if (!ptype || !btf_type_is_func_proto(ptype)) {
 396                         u32 msize;
 397
 398                         mtype = btf_type_by_id(btf_vmlinux, member->type);
 399                         mtype = btf_resolve_size(btf_vmlinux, mtype, &msize);
 400                         if (IS_ERR(mtype)) {
 401                                 err = PTR_ERR(mtype);
 402                                 goto reset_unlock;
 403                         }
 404
 405                         if (memchr_inv(udata + moff, 0, msize)) {
 406                                 err = -EINVAL;
 407                                 goto reset_unlock;
 408                         }
 409
 410                         continue;
 411                 }
 412
 413                 prog_fd = (int)(*(unsigned long *)(udata + moff));
 414                 /* Similar check as the attr->attach_prog_fd */
 415                 if (!prog_fd)
 416                         continue;
 417
 418                 prog = bpf_prog_get(prog_fd);
 419                 if (IS_ERR(prog)) {
 420                         err = PTR_ERR(prog);
 421                         goto reset_unlock;
 422                 }
 423                 st_map->progs[i] = prog;
 424
 425                 if (prog->type != BPF_PROG_TYPE_STRUCT_OPS ||
 426                     prog->aux->attach_btf_id != st_ops->type_id ||
 427                     prog->expected_attach_type != i) {
 428                         err = -EINVAL;
 429                         goto reset_unlock;
 430                 }
 431
 432                 tprogs[BPF_TRAMP_FENTRY].progs[0] = prog;
 433                 tprogs[BPF_TRAMP_FENTRY].nr_progs = 1;
 434                 err = arch_prepare_bpf_trampoline(NULL, image,
 435                                                   st_map->image + PAGE_SIZE,
 436                                                   &st_ops->func_models[i], 0,
 437                                                   tprogs, NULL);
 438                 if (err < 0)
 439                         goto reset_unlock;
 440
 441                 *(void **)(kdata + moff) = image;
 442                 image += err;
 443
 444                 /* put prog_id to udata */
 445                 *(unsigned long *)(udata + moff) = prog->aux->id;
 446         }
 447
 448         refcount_set(&kvalue->refcnt, 1);
 449         bpf_map_inc(map);
 450
 451         set_memory_ro((long)st_map->image, 1);
 452         set_memory_x((long)st_map->image, 1);
 453         err = st_ops->reg(kdata);
 454         if (likely(!err)) {
 455                 /* Pair with smp_load_acquire() during lookup_elem().
 456                  * It ensures the above udata updates (e.g. prog->aux->id)
 457                  * can be seen once BPF_STRUCT_OPS_STATE_INUSE is set.
 458                  */
 459                 smp_store_release(&kvalue->state, BPF_STRUCT_OPS_STATE_INUSE);
 460                 goto unlock;
 461         }
 462
 463         /* Error during st_ops->reg().  It is very unlikely since
 464          * the above init_member() should have caught it earlier
 465          * before reg().  The only possibility is if there was a race
 466          * in registering the struct_ops (under the same name) to
 467          * a sub-system through different struct_ops's maps.
 468          */
 469         set_memory_nx((long)st_map->image, 1);
 470         set_memory_rw((long)st_map->image, 1);
 471         bpf_map_put(map);
 472
 473 reset_unlock:
 474         bpf_struct_ops_map_put_progs(st_map);
 475         memset(uvalue, 0, map->value_size);
 476         memset(kvalue, 0, map->value_size);
 477 unlock:
 478         kfree(tprogs);
 479         mutex_unlock(&st_map->lock);
 480         return err;
 481 }
 482
 483 static int bpf_struct_ops_map_delete_elem(struct bpf_map *map, void *key)
 484 {
 485         enum bpf_struct_ops_state prev_state;
 486         struct bpf_struct_ops_map *st_map;
 487
 488         st_map = (struct bpf_struct_ops_map *)map;
 489         prev_state = cmpxchg(&st_map->kvalue.state,
 490                              BPF_STRUCT_OPS_STATE_INUSE,
 491                              BPF_STRUCT_OPS_STATE_TOBEFREE);
 492         switch (prev_state) {
 493         case BPF_STRUCT_OPS_STATE_INUSE:
 494                 st_map->st_ops->unreg(&st_map->kvalue.data);
 495                 if (refcount_dec_and_test(&st_map->kvalue.refcnt))
 496                         bpf_map_put(map);
 497                 return 0;
 498         case BPF_STRUCT_OPS_STATE_TOBEFREE:
 499                 return -EINPROGRESS;
 500         case BPF_STRUCT_OPS_STATE_INIT:
 501                 return -ENOENT;
 502         default:
 503                 WARN_ON_ONCE(1);
 504                 /* Should never happen.  Treat it as not found. */
 505                 return -ENOENT;
 506         }
 507 }
 508
 509 static void bpf_struct_ops_map_seq_show_elem(struct bpf_map *map, void *key,
 510                                              struct seq_file *m)
 511 {
 512         void *value;
 513         int err;
 514
 515         value = kmalloc(map->value_size, GFP_USER | __GFP_NOWARN);
 516         if (!value)
 517                 return;
 518
 519         err = bpf_struct_ops_map_sys_lookup_elem(map, key, value);
 520         if (!err) {
 521                 btf_type_seq_show(btf_vmlinux, map->btf_vmlinux_value_type_id,
 522                                   value, m);
 523                 seq_puts(m, "\n");
 524         }
 525
 526         kfree(value);
 527 }
 528
 529 static void bpf_struct_ops_map_free(struct bpf_map *map)
 530 {
 531         struct bpf_struct_ops_map *st_map = (struct bpf_struct_ops_map *)map;
 532
 533         if (st_map->progs)
 534                 bpf_struct_ops_map_put_progs(st_map);
 535         bpf_map_area_free(st_map->progs);
 536         bpf_jit_free_exec(st_map->image);
 537         bpf_map_area_free(st_map->uvalue);
 538         bpf_map_area_free(st_map);
 539 }
 540
 541 static int bpf_struct_ops_map_alloc_check(union bpf_attr *attr)
 542 {
 543         if (attr->key_size != sizeof(unsigned int) || attr->max_entries != 1 ||
 544             attr->map_flags || !attr->btf_vmlinux_value_type_id)
 545                 return -EINVAL;
 546         return 0;
 547 }
 548
 549 static struct bpf_map *bpf_struct_ops_map_alloc(union bpf_attr *attr)
 550 {
 551         const struct bpf_struct_ops *st_ops;
 552         size_t st_map_size;
 553         struct bpf_struct_ops_map *st_map;
 554         const struct btf_type *t, *vt;
 555         struct bpf_map *map;
 556
 557         if (!bpf_capable())
 558                 return ERR_PTR(-EPERM);
 559
 560         st_ops = bpf_struct_ops_find_value(attr->btf_vmlinux_value_type_id);
 561         if (!st_ops)
 562                 return ERR_PTR(-ENOTSUPP);
 563
 564         vt = st_ops->value_type;
 565         if (attr->value_size != vt->size)
 566                 return ERR_PTR(-EINVAL);
 567
 568         t = st_ops->type;
 569
 570         st_map_size = sizeof(*st_map) +
 571                 /* kvalue stores the
 572                  * struct bpf_struct_ops_tcp_congestions_ops
 573                  */
 574                 (vt->size - sizeof(struct bpf_struct_ops_value));
 575
 576         st_map = bpf_map_area_alloc(st_map_size, NUMA_NO_NODE);
 577         if (!st_map)
 578                 return ERR_PTR(-ENOMEM);
 579
 580         st_map->st_ops = st_ops;
 581         map = &st_map->map;
 582
 583         st_map->uvalue = bpf_map_area_alloc(vt->size, NUMA_NO_NODE);
 584         st_map->progs =
 585                 bpf_map_area_alloc(btf_type_vlen(t) * sizeof(struct bpf_prog *),
 586                                    NUMA_NO_NODE);
 587         st_map->image = bpf_jit_alloc_exec(PAGE_SIZE);
 588         if (!st_map->uvalue || !st_map->progs || !st_map->image) {
 589                 bpf_struct_ops_map_free(map);
 590                 return ERR_PTR(-ENOMEM);
 591         }
 592
 593         mutex_init(&st_map->lock);
 594         set_vm_flush_reset_perms(st_map->image);
 595         bpf_map_init_from_attr(map, attr);
 596
 597         return map;
 598 }
 599
 600 static int bpf_struct_ops_map_btf_id;
 601 const struct bpf_map_ops bpf_struct_ops_map_ops = {
 602         .map_alloc_check = bpf_struct_ops_map_alloc_check,
 603         .map_alloc = bpf_struct_ops_map_alloc,
 604         .map_free = bpf_struct_ops_map_free,
 605         .map_get_next_key = bpf_struct_ops_map_get_next_key,
 606         .map_lookup_elem = bpf_struct_ops_map_lookup_elem,
 607         .map_delete_elem = bpf_struct_ops_map_delete_elem,
 608         .map_update_elem = bpf_struct_ops_map_update_elem,
 609         .map_seq_show_elem = bpf_struct_ops_map_seq_show_elem,
 610         .map_btf_name = "bpf_struct_ops_map",
 611         .map_btf_id = &bpf_struct_ops_map_btf_id,
 612 };
 613
 614 /* "const void *" because some subsystem is
 615  * passing a const (e.g. const struct tcp_congestion_ops *)
 616  */
 617 bool bpf_struct_ops_get(const void *kdata)
 618 {
 619         struct bpf_struct_ops_value *kvalue;
 620
 621         kvalue = container_of(kdata, struct bpf_struct_ops_value, data);
 622
 623         return refcount_inc_not_zero(&kvalue->refcnt);
 624 }
 625
 626 static void bpf_struct_ops_put_rcu(struct rcu_head *head)
 627 {
 628         struct bpf_struct_ops_map *st_map;
 629
 630         st_map = container_of(head, struct bpf_struct_ops_map, rcu);
 631         bpf_map_put(&st_map->map);
 632 }
 633
 634 void bpf_struct_ops_put(const void *kdata)
 635 {
 636         struct bpf_struct_ops_value *kvalue;
 637
 638         kvalue = container_of(kdata, struct bpf_struct_ops_value, data);
 639         if (refcount_dec_and_test(&kvalue->refcnt)) {
 640                 struct bpf_struct_ops_map *st_map;
 641
 642                 st_map = container_of(kvalue, struct bpf_struct_ops_map,
 643                                       kvalue);
 644                 /* The struct_ops's function may switch to another struct_ops.
 645                  *
 646                  * For example, bpf_tcp_cc_x->init() may switch to
 647                  * another tcp_cc_y by calling
 648                  * setsockopt(TCP_CONGESTION, "tcp_cc_y").
 649                  * During the switch,  bpf_struct_ops_put(tcp_cc_x) is called
 650                  * and its map->refcnt may reach 0 which then free its
 651                  * trampoline image while tcp_cc_x is still running.
 652                  *
 653                  * Thus, a rcu grace period is needed here.
 654                  */
 655                 call_rcu(&st_map->rcu, bpf_struct_ops_put_rcu);
 656         }
 657 }