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
  14 #define RINGBUF_CREATE_FLAG_MASK (BPF_F_NUMA_NODE)
  15
  16 /* non-mmap()'able part of bpf_ringbuf (everything up to consumer page) */
  17 #define RINGBUF_PGOFF \
  18         (offsetof(struct bpf_ringbuf, consumer_pos) >> PAGE_SHIFT)
  19 /* consumer page and producer page */
  20 #define RINGBUF_POS_PAGES 2
  21
  22 #define RINGBUF_MAX_RECORD_SZ (UINT_MAX/4)
  23
  24 /* Maximum size of ring buffer area is limited by 32-bit page offset within
  25  * record header, counted in pages. Reserve 8 bits for extensibility, and take
  26  * into account few extra pages for consumer/producer pages and
  27  * non-mmap()'able parts. This gives 64GB limit, which seems plenty for single
  28  * ring buffer.
  29  */
  30 #define RINGBUF_MAX_DATA_SZ \
  31         (((1ULL << 24) - RINGBUF_POS_PAGES - RINGBUF_PGOFF) * PAGE_SIZE)
  32
  33 struct bpf_ringbuf {
  34         wait_queue_head_t waitq;
  35         struct irq_work work;
  36         u64 mask;
  37         struct page **pages;
  38         int nr_pages;
  39         spinlock_t spinlock ____cacheline_aligned_in_smp;
  40         /* Consumer and producer counters are put into separate pages to allow
  41          * mapping consumer page as r/w, but restrict producer page to r/o.
  42          * This protects producer position from being modified by user-space
  43          * application and ruining in-kernel position tracking.
  44          */
  45         unsigned long consumer_pos __aligned(PAGE_SIZE);
  46         unsigned long producer_pos __aligned(PAGE_SIZE);
  47         char data[] __aligned(PAGE_SIZE);
  48 };
  49
  50 struct bpf_ringbuf_map {
  51         struct bpf_map map;
  52         struct bpf_map_memory memory;
  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 | __GFP_RETRY_MAYFAIL | __GFP_NOWARN |
  65                             __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         if (array_size > PAGE_SIZE)
  93                 pages = vmalloc_node(array_size, numa_node);
  94         else
  95                 pages = kmalloc_node(array_size, flags, numa_node);
  96         if (!pages)
  97                 return NULL;
  98
  99         for (i = 0; i < nr_pages; i++) {
 100                 page = alloc_pages_node(numa_node, flags, 0);
 101                 if (!page) {
 102                         nr_pages = i;
 103                         goto err_free_pages;
 104                 }
 105                 pages[i] = page;
 106                 if (i >= nr_meta_pages)
 107                         pages[nr_data_pages + i] = page;
 108         }
 109
 110         rb = vmap(pages, nr_meta_pages + 2 * nr_data_pages,
 111                   VM_MAP | VM_USERMAP, PAGE_KERNEL);
 112         if (rb) {
 113                 kmemleak_not_leak(pages);
 114                 rb->pages = pages;
 115                 rb->nr_pages = nr_pages;
 116                 return rb;
 117         }
 118
 119 err_free_pages:
 120         for (i = 0; i < nr_pages; i++)
 121                 __free_page(pages[i]);
 122         kvfree(pages);
 123         return NULL;
 124 }
 125
 126 static void bpf_ringbuf_notify(struct irq_work *work)
 127 {
 128         struct bpf_ringbuf *rb = container_of(work, struct bpf_ringbuf, work);
 129
 130         wake_up_all(&rb->waitq);
 131 }
 132
 133 static struct bpf_ringbuf *bpf_ringbuf_alloc(size_t data_sz, int numa_node)
 134 {
 135         struct bpf_ringbuf *rb;
 136
 137         rb = bpf_ringbuf_area_alloc(data_sz, numa_node);
 138         if (!rb)
 139                 return ERR_PTR(-ENOMEM);
 140
 141         spin_lock_init(&rb->spinlock);
 142         init_waitqueue_head(&rb->waitq);
 143         init_irq_work(&rb->work, bpf_ringbuf_notify);
 144
 145         rb->mask = data_sz - 1;
 146         rb->consumer_pos = 0;
 147         rb->producer_pos = 0;
 148
 149         return rb;
 150 }
 151
 152 static struct bpf_map *ringbuf_map_alloc(union bpf_attr *attr)
 153 {
 154         struct bpf_ringbuf_map *rb_map;
 155         u64 cost;
 156         int err;
 157
 158         if (attr->map_flags & ~RINGBUF_CREATE_FLAG_MASK)
 159                 return ERR_PTR(-EINVAL);
 160
 161         if (attr->key_size || attr->value_size ||
 162             !is_power_of_2(attr->max_entries) ||
 163             !PAGE_ALIGNED(attr->max_entries))
 164                 return ERR_PTR(-EINVAL);
 165
 166 #ifdef CONFIG_64BIT
 167         /* on 32-bit arch, it's impossible to overflow record's hdr->pgoff */
 168         if (attr->max_entries > RINGBUF_MAX_DATA_SZ)
 169                 return ERR_PTR(-E2BIG);
 170 #endif
 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 int ringbuf_map_mmap(struct bpf_map *map, struct vm_area_struct *vma)
 246 {
 247         struct bpf_ringbuf_map *rb_map;
 248
 249         rb_map = container_of(map, struct bpf_ringbuf_map, map);
 250
 251         if (vma->vm_flags & VM_WRITE) {
 252                 /* allow writable mapping for the consumer_pos only */
 253                 if (vma->vm_pgoff != 0 || vma->vm_end - vma->vm_start != PAGE_SIZE)
 254                         return -EPERM;
 255         } else {
 256                 vma->vm_flags &= ~VM_MAYWRITE;
 257         }
 258         /* remap_vmalloc_range() checks size and offset constraints */
 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         if (len > rb->mask + 1)
 335                 return NULL;
 336
 337         cons_pos = smp_load_acquire(&rb->consumer_pos);
 338
 339         if (in_nmi()) {
 340                 if (!spin_trylock_irqsave(&rb->spinlock, flags))
 341                         return NULL;
 342         } else {
 343                 spin_lock_irqsave(&rb->spinlock, flags);
 344         }
 345
 346         prod_pos = rb->producer_pos;
 347         new_prod_pos = prod_pos + len;
 348
 349         /* check for out of ringbuf space by ensuring producer position
 350          * doesn't advance more than (ringbuf_size - 1) ahead
 351          */
 352         if (new_prod_pos - cons_pos > rb->mask) {
 353                 spin_unlock_irqrestore(&rb->spinlock, flags);
 354                 return NULL;
 355         }
 356
 357         hdr = (void *)rb->data + (prod_pos & rb->mask);
 358         pg_off = bpf_ringbuf_rec_pg_off(rb, hdr);
 359         hdr->len = size | BPF_RINGBUF_BUSY_BIT;
 360         hdr->pg_off = pg_off;
 361
 362         /* pairs with consumer's smp_load_acquire() */
 363         smp_store_release(&rb->producer_pos, new_prod_pos);
 364
 365         spin_unlock_irqrestore(&rb->spinlock, flags);
 366
 367         return (void *)hdr + BPF_RINGBUF_HDR_SZ;
 368 }
 369
 370 BPF_CALL_3(bpf_ringbuf_reserve, struct bpf_map *, map, u64, size, u64, flags)
 371 {
 372         struct bpf_ringbuf_map *rb_map;
 373
 374         if (unlikely(flags))
 375                 return 0;
 376
 377         rb_map = container_of(map, struct bpf_ringbuf_map, map);
 378         return (unsigned long)__bpf_ringbuf_reserve(rb_map->rb, size);
 379 }
 380
 381 const struct bpf_func_proto bpf_ringbuf_reserve_proto = {
 382         .func           = bpf_ringbuf_reserve,
 383         .ret_type       = RET_PTR_TO_ALLOC_MEM_OR_NULL,
 384         .arg1_type      = ARG_CONST_MAP_PTR,
 385         .arg2_type      = ARG_CONST_ALLOC_SIZE_OR_ZERO,
 386         .arg3_type      = ARG_ANYTHING,
 387 };
 388
 389 static void bpf_ringbuf_commit(void *sample, u64 flags, bool discard)
 390 {
 391         unsigned long rec_pos, cons_pos;
 392         struct bpf_ringbuf_hdr *hdr;
 393         struct bpf_ringbuf *rb;
 394         u32 new_len;
 395
 396         hdr = sample - BPF_RINGBUF_HDR_SZ;
 397         rb = bpf_ringbuf_restore_from_rec(hdr);
 398         new_len = hdr->len ^ BPF_RINGBUF_BUSY_BIT;
 399         if (discard)
 400                 new_len |= BPF_RINGBUF_DISCARD_BIT;
 401
 402         /* update record header with correct final size prefix */
 403         xchg(&hdr->len, new_len);
 404
 405         /* if consumer caught up and is waiting for our record, notify about
 406          * new data availability
 407          */
 408         rec_pos = (void *)hdr - (void *)rb->data;
 409         cons_pos = smp_load_acquire(&rb->consumer_pos) & rb->mask;
 410
 411         if (flags & BPF_RB_FORCE_WAKEUP)
 412                 irq_work_queue(&rb->work);
 413         else if (cons_pos == rec_pos && !(flags & BPF_RB_NO_WAKEUP))
 414                 irq_work_queue(&rb->work);
 415 }
 416
 417 BPF_CALL_2(bpf_ringbuf_submit, void *, sample, u64, flags)
 418 {
 419         bpf_ringbuf_commit(sample, flags, false /* discard */);
 420         return 0;
 421 }
 422
 423 const struct bpf_func_proto bpf_ringbuf_submit_proto = {
 424         .func           = bpf_ringbuf_submit,
 425         .ret_type       = RET_VOID,
 426         .arg1_type      = ARG_PTR_TO_ALLOC_MEM,
 427         .arg2_type      = ARG_ANYTHING,
 428 };
 429
 430 BPF_CALL_2(bpf_ringbuf_discard, void *, sample, u64, flags)
 431 {
 432         bpf_ringbuf_commit(sample, flags, true /* discard */);
 433         return 0;
 434 }
 435
 436 const struct bpf_func_proto bpf_ringbuf_discard_proto = {
 437         .func           = bpf_ringbuf_discard,
 438         .ret_type       = RET_VOID,
 439         .arg1_type      = ARG_PTR_TO_ALLOC_MEM,
 440         .arg2_type      = ARG_ANYTHING,
 441 };
 442
 443 BPF_CALL_4(bpf_ringbuf_output, struct bpf_map *, map, void *, data, u64, size,
 444            u64, flags)
 445 {
 446         struct bpf_ringbuf_map *rb_map;
 447         void *rec;
 448
 449         if (unlikely(flags & ~(BPF_RB_NO_WAKEUP | BPF_RB_FORCE_WAKEUP)))
 450                 return -EINVAL;
 451
 452         rb_map = container_of(map, struct bpf_ringbuf_map, map);
 453         rec = __bpf_ringbuf_reserve(rb_map->rb, size);
 454         if (!rec)
 455                 return -EAGAIN;
 456
 457         memcpy(rec, data, size);
 458         bpf_ringbuf_commit(rec, flags, false /* discard */);
 459         return 0;
 460 }
 461
 462 const struct bpf_func_proto bpf_ringbuf_output_proto = {
 463         .func           = bpf_ringbuf_output,
 464         .ret_type       = RET_INTEGER,
 465         .arg1_type      = ARG_CONST_MAP_PTR,
 466         .arg2_type      = ARG_PTR_TO_MEM,
 467         .arg3_type      = ARG_CONST_SIZE_OR_ZERO,
 468         .arg4_type      = ARG_ANYTHING,
 469 };
 470
 471 BPF_CALL_2(bpf_ringbuf_query, struct bpf_map *, map, u64, flags)
 472 {
 473         struct bpf_ringbuf *rb;
 474
 475         rb = container_of(map, struct bpf_ringbuf_map, map)->rb;
 476
 477         switch (flags) {
 478         case BPF_RB_AVAIL_DATA:
 479                 return ringbuf_avail_data_sz(rb);
 480         case BPF_RB_RING_SIZE:
 481                 return rb->mask + 1;
 482         case BPF_RB_CONS_POS:
 483                 return smp_load_acquire(&rb->consumer_pos);
 484         case BPF_RB_PROD_POS:
 485                 return smp_load_acquire(&rb->producer_pos);
 486         default:
 487                 return 0;
 488         }
 489 }
 490
 491 const struct bpf_func_proto bpf_ringbuf_query_proto = {
 492         .func           = bpf_ringbuf_query,
 493         .ret_type       = RET_INTEGER,
 494         .arg1_type      = ARG_CONST_MAP_PTR,
 495         .arg2_type      = ARG_ANYTHING,
 496 };