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