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 | __GFP_RETRY_MAYFAIL | __GFP_NOWARN |
  64                             __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         if (array_size > PAGE_SIZE)
  92                 pages = vmalloc_node(array_size, numa_node);
  93         else
  94                 pages = kmalloc_node(array_size, flags, numa_node);
  95         if (!pages)
  96                 return NULL;
  97
  98         for (i = 0; i < nr_pages; i++) {
  99                 page = alloc_pages_node(numa_node, flags, 0);
 100                 if (!page) {
 101                         nr_pages = i;
 102                         goto err_free_pages;
 103                 }
 104                 pages[i] = page;
 105                 if (i >= nr_meta_pages)
 106                         pages[nr_data_pages + i] = page;
 107         }
 108
 109         rb = vmap(pages, nr_meta_pages + 2 * nr_data_pages,
 110                   VM_ALLOC | VM_USERMAP, PAGE_KERNEL);
 111         if (rb) {
 112                 rb->pages = pages;
 113                 rb->nr_pages = nr_pages;
 114                 return rb;
 115         }
 116
 117 err_free_pages:
 118         for (i = 0; i < nr_pages; i++)
 119                 __free_page(pages[i]);
 120         kvfree(pages);
 121         return NULL;
 122 }
 123
 124 static void bpf_ringbuf_notify(struct irq_work *work)
 125 {
 126         struct bpf_ringbuf *rb = container_of(work, struct bpf_ringbuf, work);
 127
 128         wake_up_all(&rb->waitq);
 129 }
 130
 131 static struct bpf_ringbuf *bpf_ringbuf_alloc(size_t data_sz, int numa_node)
 132 {
 133         struct bpf_ringbuf *rb;
 134
 135         if (!data_sz || !PAGE_ALIGNED(data_sz))
 136                 return ERR_PTR(-EINVAL);
 137
 138 #ifdef CONFIG_64BIT
 139         /* on 32-bit arch, it's impossible to overflow record's hdr->pgoff */
 140         if (data_sz > RINGBUF_MAX_DATA_SZ)
 141                 return ERR_PTR(-E2BIG);
 142 #endif
 143
 144         rb = bpf_ringbuf_area_alloc(data_sz, numa_node);
 145         if (!rb)
 146                 return ERR_PTR(-ENOMEM);
 147
 148         spin_lock_init(&rb->spinlock);
 149         init_waitqueue_head(&rb->waitq);
 150         init_irq_work(&rb->work, bpf_ringbuf_notify);
 151
 152         rb->mask = data_sz - 1;
 153         rb->consumer_pos = 0;
 154         rb->producer_pos = 0;
 155
 156         return rb;
 157 }
 158
 159 static struct bpf_map *ringbuf_map_alloc(union bpf_attr *attr)
 160 {
 161         struct bpf_ringbuf_map *rb_map;
 162         u64 cost;
 163         int err;
 164
 165         if (attr->map_flags & ~RINGBUF_CREATE_FLAG_MASK)
 166                 return ERR_PTR(-EINVAL);
 167
 168         if (attr->key_size || attr->value_size ||
 169             attr->max_entries == 0 || !PAGE_ALIGNED(attr->max_entries))
 170                 return ERR_PTR(-EINVAL);
 171
 172         rb_map = kzalloc(sizeof(*rb_map), GFP_USER);
 173         if (!rb_map)
 174                 return ERR_PTR(-ENOMEM);
 175
 176         bpf_map_init_from_attr(&rb_map->map, attr);
 177
 178         cost = sizeof(struct bpf_ringbuf_map) +
 179                sizeof(struct bpf_ringbuf) +
 180                attr->max_entries;
 181         err = bpf_map_charge_init(&rb_map->map.memory, cost);
 182         if (err)
 183                 goto err_free_map;
 184
 185         rb_map->rb = bpf_ringbuf_alloc(attr->max_entries, rb_map->map.numa_node);
 186         if (IS_ERR(rb_map->rb)) {
 187                 err = PTR_ERR(rb_map->rb);
 188                 goto err_uncharge;
 189         }
 190
 191         return &rb_map->map;
 192
 193 err_uncharge:
 194         bpf_map_charge_finish(&rb_map->map.memory);
 195 err_free_map:
 196         kfree(rb_map);
 197         return ERR_PTR(err);
 198 }
 199
 200 static void bpf_ringbuf_free(struct bpf_ringbuf *rb)
 201 {
 202         /* copy pages pointer and nr_pages to local variable, as we are going
 203          * to unmap rb itself with vunmap() below
 204          */
 205         struct page **pages = rb->pages;
 206         int i, nr_pages = rb->nr_pages;
 207
 208         vunmap(rb);
 209         for (i = 0; i < nr_pages; i++)
 210                 __free_page(pages[i]);
 211         kvfree(pages);
 212 }
 213
 214 static void ringbuf_map_free(struct bpf_map *map)
 215 {
 216         struct bpf_ringbuf_map *rb_map;
 217
 218         rb_map = container_of(map, struct bpf_ringbuf_map, map);
 219         bpf_ringbuf_free(rb_map->rb);
 220         kfree(rb_map);
 221 }
 222
 223 static void *ringbuf_map_lookup_elem(struct bpf_map *map, void *key)
 224 {
 225         return ERR_PTR(-ENOTSUPP);
 226 }
 227
 228 static int ringbuf_map_update_elem(struct bpf_map *map, void *key, void *value,
 229                                    u64 flags)
 230 {
 231         return -ENOTSUPP;
 232 }
 233
 234 static int ringbuf_map_delete_elem(struct bpf_map *map, void *key)
 235 {
 236         return -ENOTSUPP;
 237 }
 238
 239 static int ringbuf_map_get_next_key(struct bpf_map *map, void *key,
 240                                     void *next_key)
 241 {
 242         return -ENOTSUPP;
 243 }
 244
 245 static size_t bpf_ringbuf_mmap_page_cnt(const struct bpf_ringbuf *rb)
 246 {
 247         size_t data_pages = (rb->mask + 1) >> PAGE_SHIFT;
 248
 249         /* consumer page + producer page + 2 x data pages */
 250         return RINGBUF_POS_PAGES + 2 * data_pages;
 251 }
 252
 253 static int ringbuf_map_mmap(struct bpf_map *map, struct vm_area_struct *vma)
 254 {
 255         struct bpf_ringbuf_map *rb_map;
 256         size_t mmap_sz;
 257
 258         rb_map = container_of(map, struct bpf_ringbuf_map, map);
 259         mmap_sz = bpf_ringbuf_mmap_page_cnt(rb_map->rb) << PAGE_SHIFT;
 260
 261         if (vma->vm_pgoff * PAGE_SIZE + (vma->vm_end - vma->vm_start) > mmap_sz)
 262                 return -EINVAL;
 263
 264         return remap_vmalloc_range(vma, rb_map->rb,
 265                                    vma->vm_pgoff + RINGBUF_PGOFF);
 266 }
 267
 268 static unsigned long ringbuf_avail_data_sz(struct bpf_ringbuf *rb)
 269 {
 270         unsigned long cons_pos, prod_pos;
 271
 272         cons_pos = smp_load_acquire(&rb->consumer_pos);
 273         prod_pos = smp_load_acquire(&rb->producer_pos);
 274         return prod_pos - cons_pos;
 275 }
 276
 277 static __poll_t ringbuf_map_poll(struct bpf_map *map, struct file *filp,
 278                                  struct poll_table_struct *pts)
 279 {
 280         struct bpf_ringbuf_map *rb_map;
 281
 282         rb_map = container_of(map, struct bpf_ringbuf_map, map);
 283         poll_wait(filp, &rb_map->rb->waitq, pts);
 284
 285         if (ringbuf_avail_data_sz(rb_map->rb))
 286                 return EPOLLIN | EPOLLRDNORM;
 287         return 0;
 288 }
 289
 290 static int ringbuf_map_btf_id;
 291 const struct bpf_map_ops ringbuf_map_ops = {
 292         .map_alloc = ringbuf_map_alloc,
 293         .map_free = ringbuf_map_free,
 294         .map_mmap = ringbuf_map_mmap,
 295         .map_poll = ringbuf_map_poll,
 296         .map_lookup_elem = ringbuf_map_lookup_elem,
 297         .map_update_elem = ringbuf_map_update_elem,
 298         .map_delete_elem = ringbuf_map_delete_elem,
 299         .map_get_next_key = ringbuf_map_get_next_key,
 300         .map_btf_name = "bpf_ringbuf_map",
 301         .map_btf_id = &ringbuf_map_btf_id,
 302 };
 303
 304 /* Given pointer to ring buffer record metadata and struct bpf_ringbuf itself,
 305  * calculate offset from record metadata to ring buffer in pages, rounded
 306  * down. This page offset is stored as part of record metadata and allows to
 307  * restore struct bpf_ringbuf * from record pointer. This page offset is
 308  * stored at offset 4 of record metadata header.
 309  */
 310 static size_t bpf_ringbuf_rec_pg_off(struct bpf_ringbuf *rb,
 311                                      struct bpf_ringbuf_hdr *hdr)
 312 {
 313         return ((void *)hdr - (void *)rb) >> PAGE_SHIFT;
 314 }
 315
 316 /* Given pointer to ring buffer record header, restore pointer to struct
 317  * bpf_ringbuf itself by using page offset stored at offset 4
 318  */
 319 static struct bpf_ringbuf *
 320 bpf_ringbuf_restore_from_rec(struct bpf_ringbuf_hdr *hdr)
 321 {
 322         unsigned long addr = (unsigned long)(void *)hdr;
 323         unsigned long off = (unsigned long)hdr->pg_off << PAGE_SHIFT;
 324
 325         return (void*)((addr & PAGE_MASK) - off);
 326 }
 327
 328 static void *__bpf_ringbuf_reserve(struct bpf_ringbuf *rb, u64 size)
 329 {
 330         unsigned long cons_pos, prod_pos, new_prod_pos, flags;
 331         u32 len, pg_off;
 332         struct bpf_ringbuf_hdr *hdr;
 333
 334         if (unlikely(size > RINGBUF_MAX_RECORD_SZ))
 335                 return NULL;
 336
 337         len = round_up(size + BPF_RINGBUF_HDR_SZ, 8);
 338         cons_pos = smp_load_acquire(&rb->consumer_pos);
 339
 340         if (in_nmi()) {
 341                 if (!spin_trylock_irqsave(&rb->spinlock, flags))
 342                         return NULL;
 343         } else {
 344                 spin_lock_irqsave(&rb->spinlock, flags);
 345         }
 346
 347         prod_pos = rb->producer_pos;
 348         new_prod_pos = prod_pos + len;
 349
 350         /* check for out of ringbuf space by ensuring producer position
 351          * doesn't advance more than (ringbuf_size - 1) ahead
 352          */
 353         if (new_prod_pos - cons_pos > rb->mask) {
 354                 spin_unlock_irqrestore(&rb->spinlock, flags);
 355                 return NULL;
 356         }
 357
 358         hdr = (void *)rb->data + (prod_pos & rb->mask);
 359         pg_off = bpf_ringbuf_rec_pg_off(rb, hdr);
 360         hdr->len = size | BPF_RINGBUF_BUSY_BIT;
 361         hdr->pg_off = pg_off;
 362
 363         /* pairs with consumer's smp_load_acquire() */
 364         smp_store_release(&rb->producer_pos, new_prod_pos);
 365
 366         spin_unlock_irqrestore(&rb->spinlock, flags);
 367
 368         return (void *)hdr + BPF_RINGBUF_HDR_SZ;
 369 }
 370
 371 BPF_CALL_3(bpf_ringbuf_reserve, struct bpf_map *, map, u64, size, u64, flags)
 372 {
 373         struct bpf_ringbuf_map *rb_map;
 374
 375         if (unlikely(flags))
 376                 return 0;
 377
 378         rb_map = container_of(map, struct bpf_ringbuf_map, map);
 379         return (unsigned long)__bpf_ringbuf_reserve(rb_map->rb, size);
 380 }
 381
 382 const struct bpf_func_proto bpf_ringbuf_reserve_proto = {
 383         .func           = bpf_ringbuf_reserve,
 384         .ret_type       = RET_PTR_TO_ALLOC_MEM_OR_NULL,
 385         .arg1_type      = ARG_CONST_MAP_PTR,
 386         .arg2_type      = ARG_CONST_ALLOC_SIZE_OR_ZERO,
 387         .arg3_type      = ARG_ANYTHING,
 388 };
 389
 390 static void bpf_ringbuf_commit(void *sample, u64 flags, bool discard)
 391 {
 392         unsigned long rec_pos, cons_pos;
 393         struct bpf_ringbuf_hdr *hdr;
 394         struct bpf_ringbuf *rb;
 395         u32 new_len;
 396
 397         hdr = sample - BPF_RINGBUF_HDR_SZ;
 398         rb = bpf_ringbuf_restore_from_rec(hdr);
 399         new_len = hdr->len ^ BPF_RINGBUF_BUSY_BIT;
 400         if (discard)
 401                 new_len |= BPF_RINGBUF_DISCARD_BIT;
 402
 403         /* update record header with correct final size prefix */
 404         xchg(&hdr->len, new_len);
 405
 406         /* if consumer caught up and is waiting for our record, notify about
 407          * new data availability
 408          */
 409         rec_pos = (void *)hdr - (void *)rb->data;
 410         cons_pos = smp_load_acquire(&rb->consumer_pos) & rb->mask;
 411
 412         if (flags & BPF_RB_FORCE_WAKEUP)
 413                 irq_work_queue(&rb->work);
 414         else if (cons_pos == rec_pos && !(flags & BPF_RB_NO_WAKEUP))
 415                 irq_work_queue(&rb->work);
 416 }
 417
 418 BPF_CALL_2(bpf_ringbuf_submit, void *, sample, u64, flags)
 419 {
 420         bpf_ringbuf_commit(sample, flags, false /* discard */);
 421         return 0;
 422 }
 423
 424 const struct bpf_func_proto bpf_ringbuf_submit_proto = {
 425         .func           = bpf_ringbuf_submit,
 426         .ret_type       = RET_VOID,
 427         .arg1_type      = ARG_PTR_TO_ALLOC_MEM,
 428         .arg2_type      = ARG_ANYTHING,
 429 };
 430
 431 BPF_CALL_2(bpf_ringbuf_discard, void *, sample, u64, flags)
 432 {
 433         bpf_ringbuf_commit(sample, flags, true /* discard */);
 434         return 0;
 435 }
 436
 437 const struct bpf_func_proto bpf_ringbuf_discard_proto = {
 438         .func           = bpf_ringbuf_discard,
 439         .ret_type       = RET_VOID,
 440         .arg1_type      = ARG_PTR_TO_ALLOC_MEM,
 441         .arg2_type      = ARG_ANYTHING,
 442 };
 443
 444 BPF_CALL_4(bpf_ringbuf_output, struct bpf_map *, map, void *, data, u64, size,
 445            u64, flags)
 446 {
 447         struct bpf_ringbuf_map *rb_map;
 448         void *rec;
 449
 450         if (unlikely(flags & ~(BPF_RB_NO_WAKEUP | BPF_RB_FORCE_WAKEUP)))
 451                 return -EINVAL;
 452
 453         rb_map = container_of(map, struct bpf_ringbuf_map, map);
 454         rec = __bpf_ringbuf_reserve(rb_map->rb, size);
 455         if (!rec)
 456                 return -EAGAIN;
 457
 458         memcpy(rec, data, size);
 459         bpf_ringbuf_commit(rec, flags, false /* discard */);
 460         return 0;
 461 }
 462
 463 const struct bpf_func_proto bpf_ringbuf_output_proto = {
 464         .func           = bpf_ringbuf_output,
 465         .ret_type       = RET_INTEGER,
 466         .arg1_type      = ARG_CONST_MAP_PTR,
 467         .arg2_type      = ARG_PTR_TO_MEM,
 468         .arg3_type      = ARG_CONST_SIZE_OR_ZERO,
 469         .arg4_type      = ARG_ANYTHING,
 470 };
 471
 472 BPF_CALL_2(bpf_ringbuf_query, struct bpf_map *, map, u64, flags)
 473 {
 474         struct bpf_ringbuf *rb;
 475
 476         rb = container_of(map, struct bpf_ringbuf_map, map)->rb;
 477
 478         switch (flags) {
 479         case BPF_RB_AVAIL_DATA:
 480                 return ringbuf_avail_data_sz(rb);
 481         case BPF_RB_RING_SIZE:
 482                 return rb->mask + 1;
 483         case BPF_RB_CONS_POS:
 484                 return smp_load_acquire(&rb->consumer_pos);
 485         case BPF_RB_PROD_POS:
 486                 return smp_load_acquire(&rb->producer_pos);
 487         default:
 488                 return 0;
 489         }
 490 }
 491
 492 const struct bpf_func_proto bpf_ringbuf_query_proto = {
 493         .func           = bpf_ringbuf_query,
 494         .ret_type       = RET_INTEGER,
 495         .arg1_type      = ARG_CONST_MAP_PTR,
 496         .arg2_type      = ARG_ANYTHING,
 497 };