kernel/bpf/ringbuf.c

   1 #include <linux/bpf.h>
   2 #include <linux/btf.h>
   3 #include <linux/err.h>
   4 #include <linux/irq_work.h>
   5 #include <linux/slab.h>
   6 #include <linux/filter.h>
   7 #include <linux/mm.h>
   8 #include <linux/vmalloc.h>
   9 #include <linux/wait.h>
  10 #include <linux/poll.h>
  11 #include <linux/kmemleak.h>
  12 #include <uapi/linux/btf.h>
  13 #include <linux/btf_ids.h>
  14
  15 #define RINGBUF_CREATE_FLAG_MASK (BPF_F_NUMA_NODE)
  16
  17 /* non-mmap()'able part of bpf_ringbuf (everything up to consumer page) */
  18 #define RINGBUF_PGOFF \
  19         (offsetof(struct bpf_ringbuf, consumer_pos) >> PAGE_SHIFT)
  20 /* consumer page and producer page */
  21 #define RINGBUF_POS_PAGES 2
  22
  23 #define RINGBUF_MAX_RECORD_SZ (UINT_MAX/4)
  24
  25 /* Maximum size of ring buffer area is limited by 32-bit page offset within
  26  * record header, counted in pages. Reserve 8 bits for extensibility, and take
  27  * into account few extra pages for consumer/producer pages and
  28  * non-mmap()'able parts. This gives 64GB limit, which seems plenty for single
  29  * ring buffer.
  30  */
  31 #define RINGBUF_MAX_DATA_SZ \
  32         (((1ULL << 24) - RINGBUF_POS_PAGES - RINGBUF_PGOFF) * PAGE_SIZE)
  33
  34 struct bpf_ringbuf {
  35         wait_queue_head_t waitq;
  36         struct irq_work work;
  37         u64 mask;
  38         struct page **pages;
  39         int nr_pages;
  40         spinlock_t spinlock ____cacheline_aligned_in_smp;
  41         /* Consumer and producer counters are put into separate pages to allow
  42          * mapping consumer page as r/w, but restrict producer page to r/o.
  43          * This protects producer position from being modified by user-space
  44          * application and ruining in-kernel position tracking.
  45          */
  46         unsigned long consumer_pos __aligned(PAGE_SIZE);
  47         unsigned long producer_pos __aligned(PAGE_SIZE);
  48         char data[] __aligned(PAGE_SIZE);
  49 };
  50
  51 struct bpf_ringbuf_map {
  52         struct bpf_map map;
  53         struct bpf_ringbuf *rb;
  54 };
  55
  56 /* 8-byte ring buffer record header structure */
  57 struct bpf_ringbuf_hdr {
  58         u32 len;
  59         u32 pg_off;
  60 };
  61
  62 static struct bpf_ringbuf *bpf_ringbuf_area_alloc(size_t data_sz, int numa_node)
  63 {
  64         const gfp_t flags = GFP_KERNEL_ACCOUNT | __GFP_RETRY_MAYFAIL |
  65                             __GFP_NOWARN | __GFP_ZERO;
  66         int nr_meta_pages = RINGBUF_PGOFF + RINGBUF_POS_PAGES;
  67         int nr_data_pages = data_sz >> PAGE_SHIFT;
  68         int nr_pages = nr_meta_pages + nr_data_pages;
  69         struct page **pages, *page;
  70         struct bpf_ringbuf *rb;
  71         size_t array_size;
  72         int i;
  73
  74         /* Each data page is mapped twice to allow "virtual"
  75          * continuous read of samples wrapping around the end of ring
  76          * buffer area:
  77          * ------------------------------------------------------
  78          * | meta pages |  real data pages  |  same data pages  |
  79          * ------------------------------------------------------
  80          * |            | 1 2 3 4 5 6 7 8 9 | 1 2 3 4 5 6 7 8 9 |
  81          * ------------------------------------------------------
  82          * |            | TA             DA | TA             DA |
  83          * ------------------------------------------------------
  84          *                               ^^^^^^^
  85          *                                  |
  86          * Here, no need to worry about special handling of wrapped-around
  87          * data due to double-mapped data pages. This works both in kernel and
  88          * when mmap()'ed in user-space, simplifying both kernel and
  89          * user-space implementations significantly.
  90          */
  91         array_size = (nr_meta_pages + 2 * nr_data_pages) * sizeof(*pages);
  92         pages = bpf_map_area_alloc(array_size, numa_node);
  93         if (!pages)
  94                 return NULL;
  95
  96         for (i = 0; i < nr_pages; i++) {
  97                 page = alloc_pages_node(numa_node, flags, 0);
  98                 if (!page) {
  99                         nr_pages = i;
 100                         goto err_free_pages;
 101                 }
 102                 pages[i] = page;
 103                 if (i >= nr_meta_pages)
 104                         pages[nr_data_pages + i] = page;
 105         }
 106
 107         rb = vmap(pages, nr_meta_pages + 2 * nr_data_pages,
 108                   VM_MAP | VM_USERMAP, PAGE_KERNEL);
 109         if (rb) {
 110                 kmemleak_not_leak(pages);
 111                 rb->pages = pages;
 112                 rb->nr_pages = nr_pages;
 113                 return rb;
 114         }
 115
 116 err_free_pages:
 117         for (i = 0; i < nr_pages; i++)
 118                 __free_page(pages[i]);
 119         kvfree(pages);
 120         return NULL;
 121 }
 122
 123 static void bpf_ringbuf_notify(struct irq_work *work)
 124 {
 125         struct bpf_ringbuf *rb = container_of(work, struct bpf_ringbuf, work);
 126
 127         wake_up_all(&rb->waitq);
 128 }
 129
 130 static struct bpf_ringbuf *bpf_ringbuf_alloc(size_t data_sz, int numa_node)
 131 {
 132         struct bpf_ringbuf *rb;
 133
 134         rb = bpf_ringbuf_area_alloc(data_sz, numa_node);
 135         if (!rb)
 136                 return NULL;
 137
 138         spin_lock_init(&rb->spinlock);
 139         init_waitqueue_head(&rb->waitq);
 140         init_irq_work(&rb->work, bpf_ringbuf_notify);
 141
 142         rb->mask = data_sz - 1;
 143         rb->consumer_pos = 0;
 144         rb->producer_pos = 0;
 145
 146         return rb;
 147 }
 148
 149 static struct bpf_map *ringbuf_map_alloc(union bpf_attr *attr)
 150 {
 151         struct bpf_ringbuf_map *rb_map;
 152
 153         if (attr->map_flags & ~RINGBUF_CREATE_FLAG_MASK)
 154                 return ERR_PTR(-EINVAL);
 155
 156         if (attr->key_size || attr->value_size ||
 157             !is_power_of_2(attr->max_entries) ||
 158             !PAGE_ALIGNED(attr->max_entries))
 159                 return ERR_PTR(-EINVAL);
 160
 161 #ifdef CONFIG_64BIT
 162         /* on 32-bit arch, it's impossible to overflow record's hdr->pgoff */
 163         if (attr->max_entries > RINGBUF_MAX_DATA_SZ)
 164                 return ERR_PTR(-E2BIG);
 165 #endif
 166
 167         rb_map = kzalloc(sizeof(*rb_map), GFP_USER | __GFP_ACCOUNT);
 168         if (!rb_map)
 169                 return ERR_PTR(-ENOMEM);
 170
 171         bpf_map_init_from_attr(&rb_map->map, attr);
 172
 173         rb_map->rb = bpf_ringbuf_alloc(attr->max_entries, rb_map->map.numa_node);
 174         if (!rb_map->rb) {
 175                 kfree(rb_map);
 176                 return ERR_PTR(-ENOMEM);
 177         }
 178
 179         return &rb_map->map;
 180 }
 181
 182 static void bpf_ringbuf_free(struct bpf_ringbuf *rb)
 183 {
 184         /* copy pages pointer and nr_pages to local variable, as we are going
 185          * to unmap rb itself with vunmap() below
 186          */
 187         struct page **pages = rb->pages;
 188         int i, nr_pages = rb->nr_pages;
 189
 190         vunmap(rb);
 191         for (i = 0; i < nr_pages; i++)
 192                 __free_page(pages[i]);
 193         kvfree(pages);
 194 }
 195
 196 static void ringbuf_map_free(struct bpf_map *map)
 197 {
 198         struct bpf_ringbuf_map *rb_map;
 199
 200         rb_map = container_of(map, struct bpf_ringbuf_map, map);
 201         bpf_ringbuf_free(rb_map->rb);
 202         kfree(rb_map);
 203 }
 204
 205 static void *ringbuf_map_lookup_elem(struct bpf_map *map, void *key)
 206 {
 207         return ERR_PTR(-ENOTSUPP);
 208 }
 209
 210 static int ringbuf_map_update_elem(struct bpf_map *map, void *key, void *value,
 211                                    u64 flags)
 212 {
 213         return -ENOTSUPP;
 214 }
 215
 216 static int ringbuf_map_delete_elem(struct bpf_map *map, void *key)
 217 {
 218         return -ENOTSUPP;
 219 }
 220
 221 static int ringbuf_map_get_next_key(struct bpf_map *map, void *key,
 222                                     void *next_key)
 223 {
 224         return -ENOTSUPP;
 225 }
 226
 227 static int ringbuf_map_mmap(struct bpf_map *map, struct vm_area_struct *vma)
 228 {
 229         struct bpf_ringbuf_map *rb_map;
 230
 231         rb_map = container_of(map, struct bpf_ringbuf_map, map);
 232
 233         if (vma->vm_flags & VM_WRITE) {
 234                 /* allow writable mapping for the consumer_pos only */
 235                 if (vma->vm_pgoff != 0 || vma->vm_end - vma->vm_start != PAGE_SIZE)
 236                         return -EPERM;
 237         } else {
 238                 vma->vm_flags &= ~VM_MAYWRITE;
 239         }
 240         /* remap_vmalloc_range() checks size and offset constraints */
 241         return remap_vmalloc_range(vma, rb_map->rb,
 242                                    vma->vm_pgoff + RINGBUF_PGOFF);
 243 }
 244
 245 static unsigned long ringbuf_avail_data_sz(struct bpf_ringbuf *rb)
 246 {
 247         unsigned long cons_pos, prod_pos;
 248
 249         cons_pos = smp_load_acquire(&rb->consumer_pos);
 250         prod_pos = smp_load_acquire(&rb->producer_pos);
 251         return prod_pos - cons_pos;
 252 }
 253
 254 static __poll_t ringbuf_map_poll(struct bpf_map *map, struct file *filp,
 255                                  struct poll_table_struct *pts)
 256 {
 257         struct bpf_ringbuf_map *rb_map;
 258
 259         rb_map = container_of(map, struct bpf_ringbuf_map, map);
 260         poll_wait(filp, &rb_map->rb->waitq, pts);
 261
 262         if (ringbuf_avail_data_sz(rb_map->rb))
 263                 return EPOLLIN | EPOLLRDNORM;
 264         return 0;
 265 }
 266
 267 BTF_ID_LIST_SINGLE(ringbuf_map_btf_ids, struct, bpf_ringbuf_map)
 268 const struct bpf_map_ops ringbuf_map_ops = {
 269         .map_meta_equal = bpf_map_meta_equal,
 270         .map_alloc = ringbuf_map_alloc,
 271         .map_free = ringbuf_map_free,
 272         .map_mmap = ringbuf_map_mmap,
 273         .map_poll = ringbuf_map_poll,
 274         .map_lookup_elem = ringbuf_map_lookup_elem,
 275         .map_update_elem = ringbuf_map_update_elem,
 276         .map_delete_elem = ringbuf_map_delete_elem,
 277         .map_get_next_key = ringbuf_map_get_next_key,
 278         .map_btf_id = &ringbuf_map_btf_ids[0],
 279 };
 280
 281 /* Given pointer to ring buffer record metadata and struct bpf_ringbuf itself,
 282  * calculate offset from record metadata to ring buffer in pages, rounded
 283  * down. This page offset is stored as part of record metadata and allows to
 284  * restore struct bpf_ringbuf * from record pointer. This page offset is
 285  * stored at offset 4 of record metadata header.
 286  */
 287 static size_t bpf_ringbuf_rec_pg_off(struct bpf_ringbuf *rb,
 288                                      struct bpf_ringbuf_hdr *hdr)
 289 {
 290         return ((void *)hdr - (void *)rb) >> PAGE_SHIFT;
 291 }
 292
 293 /* Given pointer to ring buffer record header, restore pointer to struct
 294  * bpf_ringbuf itself by using page offset stored at offset 4
 295  */
 296 static struct bpf_ringbuf *
 297 bpf_ringbuf_restore_from_rec(struct bpf_ringbuf_hdr *hdr)
 298 {
 299         unsigned long addr = (unsigned long)(void *)hdr;
 300         unsigned long off = (unsigned long)hdr->pg_off << PAGE_SHIFT;
 301
 302         return (void*)((addr & PAGE_MASK) - off);
 303 }
 304
 305 static void *__bpf_ringbuf_reserve(struct bpf_ringbuf *rb, u64 size)
 306 {
 307         unsigned long cons_pos, prod_pos, new_prod_pos, flags;
 308         u32 len, pg_off;
 309         struct bpf_ringbuf_hdr *hdr;
 310
 311         if (unlikely(size > RINGBUF_MAX_RECORD_SZ))
 312                 return NULL;
 313
 314         len = round_up(size + BPF_RINGBUF_HDR_SZ, 8);
 315         if (len > rb->mask + 1)
 316                 return NULL;
 317
 318         cons_pos = smp_load_acquire(&rb->consumer_pos);
 319
 320         if (in_nmi()) {
 321                 if (!spin_trylock_irqsave(&rb->spinlock, flags))
 322                         return NULL;
 323         } else {
 324                 spin_lock_irqsave(&rb->spinlock, flags);
 325         }
 326
 327         prod_pos = rb->producer_pos;
 328         new_prod_pos = prod_pos + len;
 329
 330         /* check for out of ringbuf space by ensuring producer position
 331          * doesn't advance more than (ringbuf_size - 1) ahead
 332          */
 333         if (new_prod_pos - cons_pos > rb->mask) {
 334                 spin_unlock_irqrestore(&rb->spinlock, flags);
 335                 return NULL;
 336         }
 337
 338         hdr = (void *)rb->data + (prod_pos & rb->mask);
 339         pg_off = bpf_ringbuf_rec_pg_off(rb, hdr);
 340         hdr->len = size | BPF_RINGBUF_BUSY_BIT;
 341         hdr->pg_off = pg_off;
 342
 343         /* pairs with consumer's smp_load_acquire() */
 344         smp_store_release(&rb->producer_pos, new_prod_pos);
 345
 346         spin_unlock_irqrestore(&rb->spinlock, flags);
 347
 348         return (void *)hdr + BPF_RINGBUF_HDR_SZ;
 349 }
 350
 351 BPF_CALL_3(bpf_ringbuf_reserve, struct bpf_map *, map, u64, size, u64, flags)
 352 {
 353         struct bpf_ringbuf_map *rb_map;
 354
 355         if (unlikely(flags))
 356                 return 0;
 357
 358         rb_map = container_of(map, struct bpf_ringbuf_map, map);
 359         return (unsigned long)__bpf_ringbuf_reserve(rb_map->rb, size);
 360 }
 361
 362 const struct bpf_func_proto bpf_ringbuf_reserve_proto = {
 363         .func           = bpf_ringbuf_reserve,
 364         .ret_type       = RET_PTR_TO_ALLOC_MEM_OR_NULL,
 365         .arg1_type      = ARG_CONST_MAP_PTR,
 366         .arg2_type      = ARG_CONST_ALLOC_SIZE_OR_ZERO,
 367         .arg3_type      = ARG_ANYTHING,
 368 };
 369
 370 static void bpf_ringbuf_commit(void *sample, u64 flags, bool discard)
 371 {
 372         unsigned long rec_pos, cons_pos;
 373         struct bpf_ringbuf_hdr *hdr;
 374         struct bpf_ringbuf *rb;
 375         u32 new_len;
 376
 377         hdr = sample - BPF_RINGBUF_HDR_SZ;
 378         rb = bpf_ringbuf_restore_from_rec(hdr);
 379         new_len = hdr->len ^ BPF_RINGBUF_BUSY_BIT;
 380         if (discard)
 381                 new_len |= BPF_RINGBUF_DISCARD_BIT;
 382
 383         /* update record header with correct final size prefix */
 384         xchg(&hdr->len, new_len);
 385
 386         /* if consumer caught up and is waiting for our record, notify about
 387          * new data availability
 388          */
 389         rec_pos = (void *)hdr - (void *)rb->data;
 390         cons_pos = smp_load_acquire(&rb->consumer_pos) & rb->mask;
 391
 392         if (flags & BPF_RB_FORCE_WAKEUP)
 393                 irq_work_queue(&rb->work);
 394         else if (cons_pos == rec_pos && !(flags & BPF_RB_NO_WAKEUP))
 395                 irq_work_queue(&rb->work);
 396 }
 397
 398 BPF_CALL_2(bpf_ringbuf_submit, void *, sample, u64, flags)
 399 {
 400         bpf_ringbuf_commit(sample, flags, false /* discard */);
 401         return 0;
 402 }
 403
 404 const struct bpf_func_proto bpf_ringbuf_submit_proto = {
 405         .func           = bpf_ringbuf_submit,
 406         .ret_type       = RET_VOID,
 407         .arg1_type      = ARG_PTR_TO_ALLOC_MEM | OBJ_RELEASE,
 408         .arg2_type      = ARG_ANYTHING,
 409 };
 410
 411 BPF_CALL_2(bpf_ringbuf_discard, void *, sample, u64, flags)
 412 {
 413         bpf_ringbuf_commit(sample, flags, true /* discard */);
 414         return 0;
 415 }
 416
 417 const struct bpf_func_proto bpf_ringbuf_discard_proto = {
 418         .func           = bpf_ringbuf_discard,
 419         .ret_type       = RET_VOID,
 420         .arg1_type      = ARG_PTR_TO_ALLOC_MEM | OBJ_RELEASE,
 421         .arg2_type      = ARG_ANYTHING,
 422 };
 423
 424 BPF_CALL_4(bpf_ringbuf_output, struct bpf_map *, map, void *, data, u64, size,
 425            u64, flags)
 426 {
 427         struct bpf_ringbuf_map *rb_map;
 428         void *rec;
 429
 430         if (unlikely(flags & ~(BPF_RB_NO_WAKEUP | BPF_RB_FORCE_WAKEUP)))
 431                 return -EINVAL;
 432
 433         rb_map = container_of(map, struct bpf_ringbuf_map, map);
 434         rec = __bpf_ringbuf_reserve(rb_map->rb, size);
 435         if (!rec)
 436                 return -EAGAIN;
 437
 438         memcpy(rec, data, size);
 439         bpf_ringbuf_commit(rec, flags, false /* discard */);
 440         return 0;
 441 }
 442
 443 const struct bpf_func_proto bpf_ringbuf_output_proto = {
 444         .func           = bpf_ringbuf_output,
 445         .ret_type       = RET_INTEGER,
 446         .arg1_type      = ARG_CONST_MAP_PTR,
 447         .arg2_type      = ARG_PTR_TO_MEM | MEM_RDONLY,
 448         .arg3_type      = ARG_CONST_SIZE_OR_ZERO,
 449         .arg4_type      = ARG_ANYTHING,
 450 };
 451
 452 BPF_CALL_2(bpf_ringbuf_query, struct bpf_map *, map, u64, flags)
 453 {
 454         struct bpf_ringbuf *rb;
 455
 456         rb = container_of(map, struct bpf_ringbuf_map, map)->rb;
 457
 458         switch (flags) {
 459         case BPF_RB_AVAIL_DATA:
 460                 return ringbuf_avail_data_sz(rb);
 461         case BPF_RB_RING_SIZE:
 462                 return rb->mask + 1;
 463         case BPF_RB_CONS_POS:
 464                 return smp_load_acquire(&rb->consumer_pos);
 465         case BPF_RB_PROD_POS:
 466                 return smp_load_acquire(&rb->producer_pos);
 467         default:
 468                 return 0;
 469         }
 470 }
 471
 472 const struct bpf_func_proto bpf_ringbuf_query_proto = {
 473         .func           = bpf_ringbuf_query,
 474         .ret_type       = RET_INTEGER,
 475         .arg1_type      = ARG_CONST_MAP_PTR,
 476         .arg2_type      = ARG_ANYTHING,
 477 };
 478
 479 BPF_CALL_4(bpf_ringbuf_reserve_dynptr, struct bpf_map *, map, u32, size, u64, flags,
 480            struct bpf_dynptr_kern *, ptr)
 481 {
 482         struct bpf_ringbuf_map *rb_map;
 483         void *sample;
 484         int err;
 485
 486         if (unlikely(flags)) {
 487                 bpf_dynptr_set_null(ptr);
 488                 return -EINVAL;
 489         }
 490
 491         err = bpf_dynptr_check_size(size);
 492         if (err) {
 493                 bpf_dynptr_set_null(ptr);
 494                 return err;
 495         }
 496
 497         rb_map = container_of(map, struct bpf_ringbuf_map, map);
 498
 499         sample = __bpf_ringbuf_reserve(rb_map->rb, size);
 500         if (!sample) {
 501                 bpf_dynptr_set_null(ptr);
 502                 return -EINVAL;
 503         }
 504
 505         bpf_dynptr_init(ptr, sample, BPF_DYNPTR_TYPE_RINGBUF, 0, size);
 506
 507         return 0;
 508 }
 509
 510 const struct bpf_func_proto bpf_ringbuf_reserve_dynptr_proto = {
 511         .func           = bpf_ringbuf_reserve_dynptr,
 512         .ret_type       = RET_INTEGER,
 513         .arg1_type      = ARG_CONST_MAP_PTR,
 514         .arg2_type      = ARG_ANYTHING,
 515         .arg3_type      = ARG_ANYTHING,
 516         .arg4_type      = ARG_PTR_TO_DYNPTR | DYNPTR_TYPE_RINGBUF | MEM_UNINIT,
 517 };
 518
 519 BPF_CALL_2(bpf_ringbuf_submit_dynptr, struct bpf_dynptr_kern *, ptr, u64, flags)
 520 {
 521         if (!ptr->data)
 522                 return 0;
 523
 524         bpf_ringbuf_commit(ptr->data, flags, false /* discard */);
 525
 526         bpf_dynptr_set_null(ptr);
 527
 528         return 0;
 529 }
 530
 531 const struct bpf_func_proto bpf_ringbuf_submit_dynptr_proto = {
 532         .func           = bpf_ringbuf_submit_dynptr,
 533         .ret_type       = RET_VOID,
 534         .arg1_type      = ARG_PTR_TO_DYNPTR | DYNPTR_TYPE_RINGBUF | OBJ_RELEASE,
 535         .arg2_type      = ARG_ANYTHING,
 536 };
 537
 538 BPF_CALL_2(bpf_ringbuf_discard_dynptr, struct bpf_dynptr_kern *, ptr, u64, flags)
 539 {
 540         if (!ptr->data)
 541                 return 0;
 542
 543         bpf_ringbuf_commit(ptr->data, flags, true /* discard */);
 544
 545         bpf_dynptr_set_null(ptr);
 546
 547         return 0;
 548 }
 549
 550 const struct bpf_func_proto bpf_ringbuf_discard_dynptr_proto = {
 551         .func           = bpf_ringbuf_discard_dynptr,
 552         .ret_type       = RET_VOID,
 553         .arg1_type      = ARG_PTR_TO_DYNPTR | DYNPTR_TYPE_RINGBUF | OBJ_RELEASE,
 554         .arg2_type      = ARG_ANYTHING,
 555 };