kernel/bpf/arraymap.c

   1 /* Copyright (c) 2011-2014 PLUMgrid, http://plumgrid.com
   2  * Copyright (c) 2016,2017 Facebook
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of version 2 of the GNU General Public
   6  * License as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful, but
   9  * WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
  11  * General Public License for more details.
  12  */
  13 #include <linux/bpf.h>
  14 #include <linux/err.h>
  15 #include <linux/slab.h>
  16 #include <linux/mm.h>
  17 #include <linux/filter.h>
  18 #include <linux/perf_event.h>
  19
  20 #include "map_in_map.h"
  21
  22 #define ARRAY_CREATE_FLAG_MASK \
  23         (BPF_F_NUMA_NODE | BPF_F_RDONLY | BPF_F_WRONLY)
  24
  25 static void bpf_array_free_percpu(struct bpf_array *array)
  26 {
  27         int i;
  28
  29         for (i = 0; i < array->map.max_entries; i++) {
  30                 free_percpu(array->pptrs[i]);
  31                 cond_resched();
  32         }
  33 }
  34
  35 static int bpf_array_alloc_percpu(struct bpf_array *array)
  36 {
  37         void __percpu *ptr;
  38         int i;
  39
  40         for (i = 0; i < array->map.max_entries; i++) {
  41                 ptr = __alloc_percpu_gfp(array->elem_size, 8,
  42                                          GFP_USER | __GFP_NOWARN);
  43                 if (!ptr) {
  44                         bpf_array_free_percpu(array);
  45                         return -ENOMEM;
  46                 }
  47                 array->pptrs[i] = ptr;
  48                 cond_resched();
  49         }
  50
  51         return 0;
  52 }
  53
  54 /* Called from syscall */
  55 static int array_map_alloc_check(union bpf_attr *attr)
  56 {
  57         bool percpu = attr->map_type == BPF_MAP_TYPE_PERCPU_ARRAY;
  58         int numa_node = bpf_map_attr_numa_node(attr);
  59
  60         /* check sanity of attributes */
  61         if (attr->max_entries == 0 || attr->key_size != 4 ||
  62             attr->value_size == 0 ||
  63             attr->map_flags & ~ARRAY_CREATE_FLAG_MASK ||
  64             (percpu && numa_node != NUMA_NO_NODE))
  65                 return -EINVAL;
  66
  67         if (attr->value_size > KMALLOC_MAX_SIZE)
  68                 /* if value_size is bigger, the user space won't be able to
  69                  * access the elements.
  70                  */
  71                 return -E2BIG;
  72
  73         return 0;
  74 }
  75
  76 static struct bpf_map *array_map_alloc(union bpf_attr *attr)
  77 {
  78         bool percpu = attr->map_type == BPF_MAP_TYPE_PERCPU_ARRAY;
  79         int ret, numa_node = bpf_map_attr_numa_node(attr);
  80         u32 elem_size, index_mask, max_entries;
  81         bool unpriv = !capable(CAP_SYS_ADMIN);
  82         u64 cost, array_size, mask64;
  83         struct bpf_array *array;
  84
  85         elem_size = round_up(attr->value_size, 8);
  86
  87         max_entries = attr->max_entries;
  88
  89         /* On 32 bit archs roundup_pow_of_two() with max_entries that has
  90          * upper most bit set in u32 space is undefined behavior due to
  91          * resulting 1U << 32, so do it manually here in u64 space.
  92          */
  93         mask64 = fls_long(max_entries - 1);
  94         mask64 = 1ULL << mask64;
  95         mask64 -= 1;
  96
  97         index_mask = mask64;
  98         if (unpriv) {
  99                 /* round up array size to nearest power of 2,
 100                  * since cpu will speculate within index_mask limits
 101                  */
 102                 max_entries = index_mask + 1;
 103                 /* Check for overflows. */
 104                 if (max_entries < attr->max_entries)
 105                         return ERR_PTR(-E2BIG);
 106         }
 107
 108         array_size = sizeof(*array);
 109         if (percpu)
 110                 array_size += (u64) max_entries * sizeof(void *);
 111         else
 112                 array_size += (u64) max_entries * elem_size;
 113
 114         /* make sure there is no u32 overflow later in round_up() */
 115         cost = array_size;
 116         if (cost >= U32_MAX - PAGE_SIZE)
 117                 return ERR_PTR(-ENOMEM);
 118         if (percpu) {
 119                 cost += (u64)attr->max_entries * elem_size * num_possible_cpus();
 120                 if (cost >= U32_MAX - PAGE_SIZE)
 121                         return ERR_PTR(-ENOMEM);
 122         }
 123         cost = round_up(cost, PAGE_SIZE) >> PAGE_SHIFT;
 124
 125         ret = bpf_map_precharge_memlock(cost);
 126         if (ret < 0)
 127                 return ERR_PTR(ret);
 128
 129         /* allocate all map elements and zero-initialize them */
 130         array = bpf_map_area_alloc(array_size, numa_node);
 131         if (!array)
 132                 return ERR_PTR(-ENOMEM);
 133         array->index_mask = index_mask;
 134         array->map.unpriv_array = unpriv;
 135
 136         /* copy mandatory map attributes */
 137         bpf_map_init_from_attr(&array->map, attr);
 138         array->map.pages = cost;
 139         array->elem_size = elem_size;
 140
 141         if (percpu && bpf_array_alloc_percpu(array)) {
 142                 bpf_map_area_free(array);
 143                 return ERR_PTR(-ENOMEM);
 144         }
 145
 146         return &array->map;
 147 }
 148
 149 /* Called from syscall or from eBPF program */
 150 static void *array_map_lookup_elem(struct bpf_map *map, void *key)
 151 {
 152         struct bpf_array *array = container_of(map, struct bpf_array, map);
 153         u32 index = *(u32 *)key;
 154
 155         if (unlikely(index >= array->map.max_entries))
 156                 return NULL;
 157
 158         return array->value + array->elem_size * (index & array->index_mask);
 159 }
 160
 161 /* emit BPF instructions equivalent to C code of array_map_lookup_elem() */
 162 static u32 array_map_gen_lookup(struct bpf_map *map, struct bpf_insn *insn_buf)
 163 {
 164         struct bpf_array *array = container_of(map, struct bpf_array, map);
 165         struct bpf_insn *insn = insn_buf;
 166         u32 elem_size = round_up(map->value_size, 8);
 167         const int ret = BPF_REG_0;
 168         const int map_ptr = BPF_REG_1;
 169         const int index = BPF_REG_2;
 170
 171         *insn++ = BPF_ALU64_IMM(BPF_ADD, map_ptr, offsetof(struct bpf_array, value));
 172         *insn++ = BPF_LDX_MEM(BPF_W, ret, index, 0);
 173         if (map->unpriv_array) {
 174                 *insn++ = BPF_JMP_IMM(BPF_JGE, ret, map->max_entries, 4);
 175                 *insn++ = BPF_ALU32_IMM(BPF_AND, ret, array->index_mask);
 176         } else {
 177                 *insn++ = BPF_JMP_IMM(BPF_JGE, ret, map->max_entries, 3);
 178         }
 179
 180         if (is_power_of_2(elem_size)) {
 181                 *insn++ = BPF_ALU64_IMM(BPF_LSH, ret, ilog2(elem_size));
 182         } else {
 183                 *insn++ = BPF_ALU64_IMM(BPF_MUL, ret, elem_size);
 184         }
 185         *insn++ = BPF_ALU64_REG(BPF_ADD, ret, map_ptr);
 186         *insn++ = BPF_JMP_IMM(BPF_JA, 0, 0, 1);
 187         *insn++ = BPF_MOV64_IMM(ret, 0);
 188         return insn - insn_buf;
 189 }
 190
 191 /* Called from eBPF program */
 192 static void *percpu_array_map_lookup_elem(struct bpf_map *map, void *key)
 193 {
 194         struct bpf_array *array = container_of(map, struct bpf_array, map);
 195         u32 index = *(u32 *)key;
 196
 197         if (unlikely(index >= array->map.max_entries))
 198                 return NULL;
 199
 200         return this_cpu_ptr(array->pptrs[index & array->index_mask]);
 201 }
 202
 203 int bpf_percpu_array_copy(struct bpf_map *map, void *key, void *value)
 204 {
 205         struct bpf_array *array = container_of(map, struct bpf_array, map);
 206         u32 index = *(u32 *)key;
 207         void __percpu *pptr;
 208         int cpu, off = 0;
 209         u32 size;
 210
 211         if (unlikely(index >= array->map.max_entries))
 212                 return -ENOENT;
 213
 214         /* per_cpu areas are zero-filled and bpf programs can only
 215          * access 'value_size' of them, so copying rounded areas
 216          * will not leak any kernel data
 217          */
 218         size = round_up(map->value_size, 8);
 219         rcu_read_lock();
 220         pptr = array->pptrs[index & array->index_mask];
 221         for_each_possible_cpu(cpu) {
 222                 bpf_long_memcpy(value + off, per_cpu_ptr(pptr, cpu), size);
 223                 off += size;
 224         }
 225         rcu_read_unlock();
 226         return 0;
 227 }
 228
 229 /* Called from syscall */
 230 static int array_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
 231 {
 232         struct bpf_array *array = container_of(map, struct bpf_array, map);
 233         u32 index = key ? *(u32 *)key : U32_MAX;
 234         u32 *next = (u32 *)next_key;
 235
 236         if (index >= array->map.max_entries) {
 237                 *next = 0;
 238                 return 0;
 239         }
 240
 241         if (index == array->map.max_entries - 1)
 242                 return -ENOENT;
 243
 244         *next = index + 1;
 245         return 0;
 246 }
 247
 248 /* Called from syscall or from eBPF program */
 249 static int array_map_update_elem(struct bpf_map *map, void *key, void *value,
 250                                  u64 map_flags)
 251 {
 252         struct bpf_array *array = container_of(map, struct bpf_array, map);
 253         u32 index = *(u32 *)key;
 254
 255         if (unlikely(map_flags > BPF_EXIST))
 256                 /* unknown flags */
 257                 return -EINVAL;
 258
 259         if (unlikely(index >= array->map.max_entries))
 260                 /* all elements were pre-allocated, cannot insert a new one */
 261                 return -E2BIG;
 262
 263         if (unlikely(map_flags == BPF_NOEXIST))
 264                 /* all elements already exist */
 265                 return -EEXIST;
 266
 267         if (array->map.map_type == BPF_MAP_TYPE_PERCPU_ARRAY)
 268                 memcpy(this_cpu_ptr(array->pptrs[index & array->index_mask]),
 269                        value, map->value_size);
 270         else
 271                 memcpy(array->value +
 272                        array->elem_size * (index & array->index_mask),
 273                        value, map->value_size);
 274         return 0;
 275 }
 276
 277 int bpf_percpu_array_update(struct bpf_map *map, void *key, void *value,
 278                             u64 map_flags)
 279 {
 280         struct bpf_array *array = container_of(map, struct bpf_array, map);
 281         u32 index = *(u32 *)key;
 282         void __percpu *pptr;
 283         int cpu, off = 0;
 284         u32 size;
 285
 286         if (unlikely(map_flags > BPF_EXIST))
 287                 /* unknown flags */
 288                 return -EINVAL;
 289
 290         if (unlikely(index >= array->map.max_entries))
 291                 /* all elements were pre-allocated, cannot insert a new one */
 292                 return -E2BIG;
 293
 294         if (unlikely(map_flags == BPF_NOEXIST))
 295                 /* all elements already exist */
 296                 return -EEXIST;
 297
 298         /* the user space will provide round_up(value_size, 8) bytes that
 299          * will be copied into per-cpu area. bpf programs can only access
 300          * value_size of it. During lookup the same extra bytes will be
 301          * returned or zeros which were zero-filled by percpu_alloc,
 302          * so no kernel data leaks possible
 303          */
 304         size = round_up(map->value_size, 8);
 305         rcu_read_lock();
 306         pptr = array->pptrs[index & array->index_mask];
 307         for_each_possible_cpu(cpu) {
 308                 bpf_long_memcpy(per_cpu_ptr(pptr, cpu), value + off, size);
 309                 off += size;
 310         }
 311         rcu_read_unlock();
 312         return 0;
 313 }
 314
 315 /* Called from syscall or from eBPF program */
 316 static int array_map_delete_elem(struct bpf_map *map, void *key)
 317 {
 318         return -EINVAL;
 319 }
 320
 321 /* Called when map->refcnt goes to zero, either from workqueue or from syscall */
 322 static void array_map_free(struct bpf_map *map)
 323 {
 324         struct bpf_array *array = container_of(map, struct bpf_array, map);
 325
 326         /* at this point bpf_prog->aux->refcnt == 0 and this map->refcnt == 0,
 327          * so the programs (can be more than one that used this map) were
 328          * disconnected from events. Wait for outstanding programs to complete
 329          * and free the array
 330          */
 331         synchronize_rcu();
 332
 333         if (array->map.map_type == BPF_MAP_TYPE_PERCPU_ARRAY)
 334                 bpf_array_free_percpu(array);
 335
 336         bpf_map_area_free(array);
 337 }
 338
 339 const struct bpf_map_ops array_map_ops = {
 340         .map_alloc_check = array_map_alloc_check,
 341         .map_alloc = array_map_alloc,
 342         .map_free = array_map_free,
 343         .map_get_next_key = array_map_get_next_key,
 344         .map_lookup_elem = array_map_lookup_elem,
 345         .map_update_elem = array_map_update_elem,
 346         .map_delete_elem = array_map_delete_elem,
 347         .map_gen_lookup = array_map_gen_lookup,
 348 };
 349
 350 const struct bpf_map_ops percpu_array_map_ops = {
 351         .map_alloc_check = array_map_alloc_check,
 352         .map_alloc = array_map_alloc,
 353         .map_free = array_map_free,
 354         .map_get_next_key = array_map_get_next_key,
 355         .map_lookup_elem = percpu_array_map_lookup_elem,
 356         .map_update_elem = array_map_update_elem,
 357         .map_delete_elem = array_map_delete_elem,
 358 };
 359
 360 static int fd_array_map_alloc_check(union bpf_attr *attr)
 361 {
 362         /* only file descriptors can be stored in this type of map */
 363         if (attr->value_size != sizeof(u32))
 364                 return -EINVAL;
 365         return array_map_alloc_check(attr);
 366 }
 367
 368 static void fd_array_map_free(struct bpf_map *map)
 369 {
 370         struct bpf_array *array = container_of(map, struct bpf_array, map);
 371         int i;
 372
 373         synchronize_rcu();
 374
 375         /* make sure it's empty */
 376         for (i = 0; i < array->map.max_entries; i++)
 377                 BUG_ON(array->ptrs[i] != NULL);
 378
 379         bpf_map_area_free(array);
 380 }
 381
 382 static void *fd_array_map_lookup_elem(struct bpf_map *map, void *key)
 383 {
 384         return NULL;
 385 }
 386
 387 /* only called from syscall */
 388 int bpf_fd_array_map_lookup_elem(struct bpf_map *map, void *key, u32 *value)
 389 {
 390         void **elem, *ptr;
 391         int ret =  0;
 392
 393         if (!map->ops->map_fd_sys_lookup_elem)
 394                 return -ENOTSUPP;
 395
 396         rcu_read_lock();
 397         elem = array_map_lookup_elem(map, key);
 398         if (elem && (ptr = READ_ONCE(*elem)))
 399                 *value = map->ops->map_fd_sys_lookup_elem(ptr);
 400         else
 401                 ret = -ENOENT;
 402         rcu_read_unlock();
 403
 404         return ret;
 405 }
 406
 407 /* only called from syscall */
 408 int bpf_fd_array_map_update_elem(struct bpf_map *map, struct file *map_file,
 409                                  void *key, void *value, u64 map_flags)
 410 {
 411         struct bpf_array *array = container_of(map, struct bpf_array, map);
 412         void *new_ptr, *old_ptr;
 413         u32 index = *(u32 *)key, ufd;
 414
 415         if (map_flags != BPF_ANY)
 416                 return -EINVAL;
 417
 418         if (index >= array->map.max_entries)
 419                 return -E2BIG;
 420
 421         ufd = *(u32 *)value;
 422         new_ptr = map->ops->map_fd_get_ptr(map, map_file, ufd);
 423         if (IS_ERR(new_ptr))
 424                 return PTR_ERR(new_ptr);
 425
 426         old_ptr = xchg(array->ptrs + index, new_ptr);
 427         if (old_ptr)
 428                 map->ops->map_fd_put_ptr(old_ptr);
 429
 430         return 0;
 431 }
 432
 433 static int fd_array_map_delete_elem(struct bpf_map *map, void *key)
 434 {
 435         struct bpf_array *array = container_of(map, struct bpf_array, map);
 436         void *old_ptr;
 437         u32 index = *(u32 *)key;
 438
 439         if (index >= array->map.max_entries)
 440                 return -E2BIG;
 441
 442         old_ptr = xchg(array->ptrs + index, NULL);
 443         if (old_ptr) {
 444                 map->ops->map_fd_put_ptr(old_ptr);
 445                 return 0;
 446         } else {
 447                 return -ENOENT;
 448         }
 449 }
 450
 451 static void *prog_fd_array_get_ptr(struct bpf_map *map,
 452                                    struct file *map_file, int fd)
 453 {
 454         struct bpf_array *array = container_of(map, struct bpf_array, map);
 455         struct bpf_prog *prog = bpf_prog_get(fd);
 456
 457         if (IS_ERR(prog))
 458                 return prog;
 459
 460         if (!bpf_prog_array_compatible(array, prog)) {
 461                 bpf_prog_put(prog);
 462                 return ERR_PTR(-EINVAL);
 463         }
 464
 465         return prog;
 466 }
 467
 468 static void prog_fd_array_put_ptr(void *ptr)
 469 {
 470         bpf_prog_put(ptr);
 471 }
 472
 473 static u32 prog_fd_array_sys_lookup_elem(void *ptr)
 474 {
 475         return ((struct bpf_prog *)ptr)->aux->id;
 476 }
 477
 478 /* decrement refcnt of all bpf_progs that are stored in this map */
 479 void bpf_fd_array_map_clear(struct bpf_map *map)
 480 {
 481         struct bpf_array *array = container_of(map, struct bpf_array, map);
 482         int i;
 483
 484         for (i = 0; i < array->map.max_entries; i++)
 485                 fd_array_map_delete_elem(map, &i);
 486 }
 487
 488 const struct bpf_map_ops prog_array_map_ops = {
 489         .map_alloc_check = fd_array_map_alloc_check,
 490         .map_alloc = array_map_alloc,
 491         .map_free = fd_array_map_free,
 492         .map_get_next_key = array_map_get_next_key,
 493         .map_lookup_elem = fd_array_map_lookup_elem,
 494         .map_delete_elem = fd_array_map_delete_elem,
 495         .map_fd_get_ptr = prog_fd_array_get_ptr,
 496         .map_fd_put_ptr = prog_fd_array_put_ptr,
 497         .map_fd_sys_lookup_elem = prog_fd_array_sys_lookup_elem,
 498 };
 499
 500 static struct bpf_event_entry *bpf_event_entry_gen(struct file *perf_file,
 501                                                    struct file *map_file)
 502 {
 503         struct bpf_event_entry *ee;
 504
 505         ee = kzalloc(sizeof(*ee), GFP_ATOMIC);
 506         if (ee) {
 507                 ee->event = perf_file->private_data;
 508                 ee->perf_file = perf_file;
 509                 ee->map_file = map_file;
 510         }
 511
 512         return ee;
 513 }
 514
 515 static void __bpf_event_entry_free(struct rcu_head *rcu)
 516 {
 517         struct bpf_event_entry *ee;
 518
 519         ee = container_of(rcu, struct bpf_event_entry, rcu);
 520         fput(ee->perf_file);
 521         kfree(ee);
 522 }
 523
 524 static void bpf_event_entry_free_rcu(struct bpf_event_entry *ee)
 525 {
 526         call_rcu(&ee->rcu, __bpf_event_entry_free);
 527 }
 528
 529 static void *perf_event_fd_array_get_ptr(struct bpf_map *map,
 530                                          struct file *map_file, int fd)
 531 {
 532         struct bpf_event_entry *ee;
 533         struct perf_event *event;
 534         struct file *perf_file;
 535         u64 value;
 536
 537         perf_file = perf_event_get(fd);
 538         if (IS_ERR(perf_file))
 539                 return perf_file;
 540
 541         ee = ERR_PTR(-EOPNOTSUPP);
 542         event = perf_file->private_data;
 543         if (perf_event_read_local(event, &value, NULL, NULL) == -EOPNOTSUPP)
 544                 goto err_out;
 545
 546         ee = bpf_event_entry_gen(perf_file, map_file);
 547         if (ee)
 548                 return ee;
 549         ee = ERR_PTR(-ENOMEM);
 550 err_out:
 551         fput(perf_file);
 552         return ee;
 553 }
 554
 555 static void perf_event_fd_array_put_ptr(void *ptr)
 556 {
 557         bpf_event_entry_free_rcu(ptr);
 558 }
 559
 560 static void perf_event_fd_array_release(struct bpf_map *map,
 561                                         struct file *map_file)
 562 {
 563         struct bpf_array *array = container_of(map, struct bpf_array, map);
 564         struct bpf_event_entry *ee;
 565         int i;
 566
 567         rcu_read_lock();
 568         for (i = 0; i < array->map.max_entries; i++) {
 569                 ee = READ_ONCE(array->ptrs[i]);
 570                 if (ee && ee->map_file == map_file)
 571                         fd_array_map_delete_elem(map, &i);
 572         }
 573         rcu_read_unlock();
 574 }
 575
 576 const struct bpf_map_ops perf_event_array_map_ops = {
 577         .map_alloc_check = fd_array_map_alloc_check,
 578         .map_alloc = array_map_alloc,
 579         .map_free = fd_array_map_free,
 580         .map_get_next_key = array_map_get_next_key,
 581         .map_lookup_elem = fd_array_map_lookup_elem,
 582         .map_delete_elem = fd_array_map_delete_elem,
 583         .map_fd_get_ptr = perf_event_fd_array_get_ptr,
 584         .map_fd_put_ptr = perf_event_fd_array_put_ptr,
 585         .map_release = perf_event_fd_array_release,
 586 };
 587
 588 #ifdef CONFIG_CGROUPS
 589 static void *cgroup_fd_array_get_ptr(struct bpf_map *map,
 590                                      struct file *map_file /* not used */,
 591                                      int fd)
 592 {
 593         return cgroup_get_from_fd(fd);
 594 }
 595
 596 static void cgroup_fd_array_put_ptr(void *ptr)
 597 {
 598         /* cgroup_put free cgrp after a rcu grace period */
 599         cgroup_put(ptr);
 600 }
 601
 602 static void cgroup_fd_array_free(struct bpf_map *map)
 603 {
 604         bpf_fd_array_map_clear(map);
 605         fd_array_map_free(map);
 606 }
 607
 608 const struct bpf_map_ops cgroup_array_map_ops = {
 609         .map_alloc_check = fd_array_map_alloc_check,
 610         .map_alloc = array_map_alloc,
 611         .map_free = cgroup_fd_array_free,
 612         .map_get_next_key = array_map_get_next_key,
 613         .map_lookup_elem = fd_array_map_lookup_elem,
 614         .map_delete_elem = fd_array_map_delete_elem,
 615         .map_fd_get_ptr = cgroup_fd_array_get_ptr,
 616         .map_fd_put_ptr = cgroup_fd_array_put_ptr,
 617 };
 618 #endif
 619
 620 static struct bpf_map *array_of_map_alloc(union bpf_attr *attr)
 621 {
 622         struct bpf_map *map, *inner_map_meta;
 623
 624         inner_map_meta = bpf_map_meta_alloc(attr->inner_map_fd);
 625         if (IS_ERR(inner_map_meta))
 626                 return inner_map_meta;
 627
 628         map = array_map_alloc(attr);
 629         if (IS_ERR(map)) {
 630                 bpf_map_meta_free(inner_map_meta);
 631                 return map;
 632         }
 633
 634         map->inner_map_meta = inner_map_meta;
 635
 636         return map;
 637 }
 638
 639 static void array_of_map_free(struct bpf_map *map)
 640 {
 641         /* map->inner_map_meta is only accessed by syscall which
 642          * is protected by fdget/fdput.
 643          */
 644         bpf_map_meta_free(map->inner_map_meta);
 645         bpf_fd_array_map_clear(map);
 646         fd_array_map_free(map);
 647 }
 648
 649 static void *array_of_map_lookup_elem(struct bpf_map *map, void *key)
 650 {
 651         struct bpf_map **inner_map = array_map_lookup_elem(map, key);
 652
 653         if (!inner_map)
 654                 return NULL;
 655
 656         return READ_ONCE(*inner_map);
 657 }
 658
 659 static u32 array_of_map_gen_lookup(struct bpf_map *map,
 660                                    struct bpf_insn *insn_buf)
 661 {
 662         struct bpf_array *array = container_of(map, struct bpf_array, map);
 663         u32 elem_size = round_up(map->value_size, 8);
 664         struct bpf_insn *insn = insn_buf;
 665         const int ret = BPF_REG_0;
 666         const int map_ptr = BPF_REG_1;
 667         const int index = BPF_REG_2;
 668
 669         *insn++ = BPF_ALU64_IMM(BPF_ADD, map_ptr, offsetof(struct bpf_array, value));
 670         *insn++ = BPF_LDX_MEM(BPF_W, ret, index, 0);
 671         if (map->unpriv_array) {
 672                 *insn++ = BPF_JMP_IMM(BPF_JGE, ret, map->max_entries, 6);
 673                 *insn++ = BPF_ALU32_IMM(BPF_AND, ret, array->index_mask);
 674         } else {
 675                 *insn++ = BPF_JMP_IMM(BPF_JGE, ret, map->max_entries, 5);
 676         }
 677         if (is_power_of_2(elem_size))
 678                 *insn++ = BPF_ALU64_IMM(BPF_LSH, ret, ilog2(elem_size));
 679         else
 680                 *insn++ = BPF_ALU64_IMM(BPF_MUL, ret, elem_size);
 681         *insn++ = BPF_ALU64_REG(BPF_ADD, ret, map_ptr);
 682         *insn++ = BPF_LDX_MEM(BPF_DW, ret, ret, 0);
 683         *insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 1);
 684         *insn++ = BPF_JMP_IMM(BPF_JA, 0, 0, 1);
 685         *insn++ = BPF_MOV64_IMM(ret, 0);
 686
 687         return insn - insn_buf;
 688 }
 689
 690 const struct bpf_map_ops array_of_maps_map_ops = {
 691         .map_alloc_check = fd_array_map_alloc_check,
 692         .map_alloc = array_of_map_alloc,
 693         .map_free = array_of_map_free,
 694         .map_get_next_key = array_map_get_next_key,
 695         .map_lookup_elem = array_of_map_lookup_elem,
 696         .map_delete_elem = fd_array_map_delete_elem,
 697         .map_fd_get_ptr = bpf_map_fd_get_ptr,
 698         .map_fd_put_ptr = bpf_map_fd_put_ptr,
 699         .map_fd_sys_lookup_elem = bpf_map_fd_sys_lookup_elem,
 700         .map_gen_lookup = array_of_map_gen_lookup,
 701 };