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 #define RINGBUF_NR_META_PAGES (RINGBUF_PGOFF + RINGBUF_POS_PAGES)
  23
  24 #define RINGBUF_MAX_RECORD_SZ (UINT_MAX/4)
  25
  26 struct bpf_ringbuf {
  27         wait_queue_head_t waitq;
  28         struct irq_work work;
  29         u64 mask;
  30         struct page **pages;
  31         int nr_pages;
  32         spinlock_t spinlock ____cacheline_aligned_in_smp;
  33         /* For user-space producer ring buffers, an atomic_t busy bit is used
  34          * to synchronize access to the ring buffers in the kernel, rather than
  35          * the spinlock that is used for kernel-producer ring buffers. This is
  36          * done because the ring buffer must hold a lock across a BPF program's
  37          * callback:
  38          *
  39          *    __bpf_user_ringbuf_peek() // lock acquired
  40          * -> program callback_fn()
  41          * -> __bpf_user_ringbuf_sample_release() // lock released
  42          *
  43          * It is unsafe and incorrect to hold an IRQ spinlock across what could
  44          * be a long execution window, so we instead simply disallow concurrent
  45          * access to the ring buffer by kernel consumers, and return -EBUSY from
  46          * __bpf_user_ringbuf_peek() if the busy bit is held by another task.
  47          */
  48         atomic_t busy ____cacheline_aligned_in_smp;
  49         /* Consumer and producer counters are put into separate pages to
  50          * allow each position to be mapped with different permissions.
  51          * This prevents a user-space application from modifying the
  52          * position and ruining in-kernel tracking. The permissions of the
  53          * pages depend on who is producing samples: user-space or the
  54          * kernel.
  55          *
  56          * Kernel-producer
  57          * ---------------
  58          * The producer position and data pages are mapped as r/o in
  59          * userspace. For this approach, bits in the header of samples are
  60          * used to signal to user-space, and to other producers, whether a
  61          * sample is currently being written.
  62          *
  63          * User-space producer
  64          * -------------------
  65          * Only the page containing the consumer position is mapped r/o in
  66          * user-space. User-space producers also use bits of the header to
  67          * communicate to the kernel, but the kernel must carefully check and
  68          * validate each sample to ensure that they're correctly formatted, and
  69          * fully contained within the ring buffer.
  70          */
  71         unsigned long consumer_pos __aligned(PAGE_SIZE);
  72         unsigned long producer_pos __aligned(PAGE_SIZE);
  73         char data[] __aligned(PAGE_SIZE);
  74 };
  75
  76 struct bpf_ringbuf_map {
  77         struct bpf_map map;
  78         struct bpf_ringbuf *rb;
  79 };
  80
  81 /* 8-byte ring buffer record header structure */
  82 struct bpf_ringbuf_hdr {
  83         u32 len;
  84         u32 pg_off;
  85 };
  86
  87 static struct bpf_ringbuf *bpf_ringbuf_area_alloc(size_t data_sz, int numa_node)
  88 {
  89         const gfp_t flags = GFP_KERNEL_ACCOUNT | __GFP_RETRY_MAYFAIL |
  90                             __GFP_NOWARN | __GFP_ZERO;
  91         int nr_meta_pages = RINGBUF_NR_META_PAGES;
  92         int nr_data_pages = data_sz >> PAGE_SHIFT;
  93         int nr_pages = nr_meta_pages + nr_data_pages;
  94         struct page **pages, *page;
  95         struct bpf_ringbuf *rb;
  96         size_t array_size;
  97         int i;
  98
  99         /* Each data page is mapped twice to allow "virtual"
 100          * continuous read of samples wrapping around the end of ring
 101          * buffer area:
 102          * ------------------------------------------------------
 103          * | meta pages |  real data pages  |  same data pages  |
 104          * ------------------------------------------------------
 105          * |            | 1 2 3 4 5 6 7 8 9 | 1 2 3 4 5 6 7 8 9 |
 106          * ------------------------------------------------------
 107          * |            | TA             DA | TA             DA |
 108          * ------------------------------------------------------
 109          *                               ^^^^^^^
 110          *                                  |
 111          * Here, no need to worry about special handling of wrapped-around
 112          * data due to double-mapped data pages. This works both in kernel and
 113          * when mmap()'ed in user-space, simplifying both kernel and
 114          * user-space implementations significantly.
 115          */
 116         array_size = (nr_meta_pages + 2 * nr_data_pages) * sizeof(*pages);
 117         pages = bpf_map_area_alloc(array_size, numa_node);
 118         if (!pages)
 119                 return NULL;
 120
 121         for (i = 0; i < nr_pages; i++) {
 122                 page = alloc_pages_node(numa_node, flags, 0);
 123                 if (!page) {
 124                         nr_pages = i;
 125                         goto err_free_pages;
 126                 }
 127                 pages[i] = page;
 128                 if (i >= nr_meta_pages)
 129                         pages[nr_data_pages + i] = page;
 130         }
 131
 132         rb = vmap(pages, nr_meta_pages + 2 * nr_data_pages,
 133                   VM_MAP | VM_USERMAP, PAGE_KERNEL);
 134         if (rb) {
 135                 kmemleak_not_leak(pages);
 136                 rb->pages = pages;
 137                 rb->nr_pages = nr_pages;
 138                 return rb;
 139         }
 140
 141 err_free_pages:
 142         for (i = 0; i < nr_pages; i++)
 143                 __free_page(pages[i]);
 144         bpf_map_area_free(pages);
 145         return NULL;
 146 }
 147
 148 static void bpf_ringbuf_notify(struct irq_work *work)
 149 {
 150         struct bpf_ringbuf *rb = container_of(work, struct bpf_ringbuf, work);
 151
 152         wake_up_all(&rb->waitq);
 153 }
 154
 155 /* Maximum size of ring buffer area is limited by 32-bit page offset within
 156  * record header, counted in pages. Reserve 8 bits for extensibility, and
 157  * take into account few extra pages for consumer/producer pages and
 158  * non-mmap()'able parts, the current maximum size would be:
 159  *
 160  *     (((1ULL << 24) - RINGBUF_POS_PAGES - RINGBUF_PGOFF) * PAGE_SIZE)
 161  *
 162  * This gives 64GB limit, which seems plenty for single ring buffer. Now
 163  * considering that the maximum value of data_sz is (4GB - 1), there
 164  * will be no overflow, so just note the size limit in the comments.
 165  */
 166 static struct bpf_ringbuf *bpf_ringbuf_alloc(size_t data_sz, int numa_node)
 167 {
 168         struct bpf_ringbuf *rb;
 169
 170         rb = bpf_ringbuf_area_alloc(data_sz, numa_node);
 171         if (!rb)
 172                 return NULL;
 173
 174         spin_lock_init(&rb->spinlock);
 175         atomic_set(&rb->busy, 0);
 176         init_waitqueue_head(&rb->waitq);
 177         init_irq_work(&rb->work, bpf_ringbuf_notify);
 178
 179         rb->mask = data_sz - 1;
 180         rb->consumer_pos = 0;
 181         rb->producer_pos = 0;
 182
 183         return rb;
 184 }
 185
 186 static struct bpf_map *ringbuf_map_alloc(union bpf_attr *attr)
 187 {
 188         struct bpf_ringbuf_map *rb_map;
 189
 190         if (attr->map_flags & ~RINGBUF_CREATE_FLAG_MASK)
 191                 return ERR_PTR(-EINVAL);
 192
 193         if (attr->key_size || attr->value_size ||
 194             !is_power_of_2(attr->max_entries) ||
 195             !PAGE_ALIGNED(attr->max_entries))
 196                 return ERR_PTR(-EINVAL);
 197
 198         rb_map = bpf_map_area_alloc(sizeof(*rb_map), NUMA_NO_NODE);
 199         if (!rb_map)
 200                 return ERR_PTR(-ENOMEM);
 201
 202         bpf_map_init_from_attr(&rb_map->map, attr);
 203
 204         rb_map->rb = bpf_ringbuf_alloc(attr->max_entries, rb_map->map.numa_node);
 205         if (!rb_map->rb) {
 206                 bpf_map_area_free(rb_map);
 207                 return ERR_PTR(-ENOMEM);
 208         }
 209
 210         return &rb_map->map;
 211 }
 212
 213 static void bpf_ringbuf_free(struct bpf_ringbuf *rb)
 214 {
 215         /* copy pages pointer and nr_pages to local variable, as we are going
 216          * to unmap rb itself with vunmap() below
 217          */
 218         struct page **pages = rb->pages;
 219         int i, nr_pages = rb->nr_pages;
 220
 221         vunmap(rb);
 222         for (i = 0; i < nr_pages; i++)
 223                 __free_page(pages[i]);
 224         bpf_map_area_free(pages);
 225 }
 226
 227 static void ringbuf_map_free(struct bpf_map *map)
 228 {
 229         struct bpf_ringbuf_map *rb_map;
 230
 231         rb_map = container_of(map, struct bpf_ringbuf_map, map);
 232         bpf_ringbuf_free(rb_map->rb);
 233         bpf_map_area_free(rb_map);
 234 }
 235
 236 static void *ringbuf_map_lookup_elem(struct bpf_map *map, void *key)
 237 {
 238         return ERR_PTR(-ENOTSUPP);
 239 }
 240
 241 static long ringbuf_map_update_elem(struct bpf_map *map, void *key, void *value,
 242                                     u64 flags)
 243 {
 244         return -ENOTSUPP;
 245 }
 246
 247 static long ringbuf_map_delete_elem(struct bpf_map *map, void *key)
 248 {
 249         return -ENOTSUPP;
 250 }
 251
 252 static int ringbuf_map_get_next_key(struct bpf_map *map, void *key,
 253                                     void *next_key)
 254 {
 255         return -ENOTSUPP;
 256 }
 257
 258 static int ringbuf_map_mmap_kern(struct bpf_map *map, struct vm_area_struct *vma)
 259 {
 260         struct bpf_ringbuf_map *rb_map;
 261
 262         rb_map = container_of(map, struct bpf_ringbuf_map, map);
 263
 264         if (vma->vm_flags & VM_WRITE) {
 265                 /* allow writable mapping for the consumer_pos only */
 266                 if (vma->vm_pgoff != 0 || vma->vm_end - vma->vm_start != PAGE_SIZE)
 267                         return -EPERM;
 268         } else {
 269                 vm_flags_clear(vma, VM_MAYWRITE);
 270         }
 271         /* remap_vmalloc_range() checks size and offset constraints */
 272         return remap_vmalloc_range(vma, rb_map->rb,
 273                                    vma->vm_pgoff + RINGBUF_PGOFF);
 274 }
 275
 276 static int ringbuf_map_mmap_user(struct bpf_map *map, struct vm_area_struct *vma)
 277 {
 278         struct bpf_ringbuf_map *rb_map;
 279
 280         rb_map = container_of(map, struct bpf_ringbuf_map, map);
 281
 282         if (vma->vm_flags & VM_WRITE) {
 283                 if (vma->vm_pgoff == 0)
 284                         /* Disallow writable mappings to the consumer pointer,
 285                          * and allow writable mappings to both the producer
 286                          * position, and the ring buffer data itself.
 287                          */
 288                         return -EPERM;
 289         } else {
 290                 vm_flags_clear(vma, VM_MAYWRITE);
 291         }
 292         /* remap_vmalloc_range() checks size and offset constraints */
 293         return remap_vmalloc_range(vma, rb_map->rb, vma->vm_pgoff + RINGBUF_PGOFF);
 294 }
 295
 296 static unsigned long ringbuf_avail_data_sz(struct bpf_ringbuf *rb)
 297 {
 298         unsigned long cons_pos, prod_pos;
 299
 300         cons_pos = smp_load_acquire(&rb->consumer_pos);
 301         prod_pos = smp_load_acquire(&rb->producer_pos);
 302         return prod_pos - cons_pos;
 303 }
 304
 305 static u32 ringbuf_total_data_sz(const struct bpf_ringbuf *rb)
 306 {
 307         return rb->mask + 1;
 308 }
 309
 310 static __poll_t ringbuf_map_poll_kern(struct bpf_map *map, struct file *filp,
 311                                       struct poll_table_struct *pts)
 312 {
 313         struct bpf_ringbuf_map *rb_map;
 314
 315         rb_map = container_of(map, struct bpf_ringbuf_map, map);
 316         poll_wait(filp, &rb_map->rb->waitq, pts);
 317
 318         if (ringbuf_avail_data_sz(rb_map->rb))
 319                 return EPOLLIN | EPOLLRDNORM;
 320         return 0;
 321 }
 322
 323 static __poll_t ringbuf_map_poll_user(struct bpf_map *map, struct file *filp,
 324                                       struct poll_table_struct *pts)
 325 {
 326         struct bpf_ringbuf_map *rb_map;
 327
 328         rb_map = container_of(map, struct bpf_ringbuf_map, map);
 329         poll_wait(filp, &rb_map->rb->waitq, pts);
 330
 331         if (ringbuf_avail_data_sz(rb_map->rb) < ringbuf_total_data_sz(rb_map->rb))
 332                 return EPOLLOUT | EPOLLWRNORM;
 333         return 0;
 334 }
 335
 336 static u64 ringbuf_map_mem_usage(const struct bpf_map *map)
 337 {
 338         struct bpf_ringbuf *rb;
 339         int nr_data_pages;
 340         int nr_meta_pages;
 341         u64 usage = sizeof(struct bpf_ringbuf_map);
 342
 343         rb = container_of(map, struct bpf_ringbuf_map, map)->rb;
 344         usage += (u64)rb->nr_pages << PAGE_SHIFT;
 345         nr_meta_pages = RINGBUF_NR_META_PAGES;
 346         nr_data_pages = map->max_entries >> PAGE_SHIFT;
 347         usage += (nr_meta_pages + 2 * nr_data_pages) * sizeof(struct page *);
 348         return usage;
 349 }
 350
 351 BTF_ID_LIST_SINGLE(ringbuf_map_btf_ids, struct, bpf_ringbuf_map)
 352 const struct bpf_map_ops ringbuf_map_ops = {
 353         .map_meta_equal = bpf_map_meta_equal,
 354         .map_alloc = ringbuf_map_alloc,
 355         .map_free = ringbuf_map_free,
 356         .map_mmap = ringbuf_map_mmap_kern,
 357         .map_poll = ringbuf_map_poll_kern,
 358         .map_lookup_elem = ringbuf_map_lookup_elem,
 359         .map_update_elem = ringbuf_map_update_elem,
 360         .map_delete_elem = ringbuf_map_delete_elem,
 361         .map_get_next_key = ringbuf_map_get_next_key,
 362         .map_mem_usage = ringbuf_map_mem_usage,
 363         .map_btf_id = &ringbuf_map_btf_ids[0],
 364 };
 365
 366 BTF_ID_LIST_SINGLE(user_ringbuf_map_btf_ids, struct, bpf_ringbuf_map)
 367 const struct bpf_map_ops user_ringbuf_map_ops = {
 368         .map_meta_equal = bpf_map_meta_equal,
 369         .map_alloc = ringbuf_map_alloc,
 370         .map_free = ringbuf_map_free,
 371         .map_mmap = ringbuf_map_mmap_user,
 372         .map_poll = ringbuf_map_poll_user,
 373         .map_lookup_elem = ringbuf_map_lookup_elem,
 374         .map_update_elem = ringbuf_map_update_elem,
 375         .map_delete_elem = ringbuf_map_delete_elem,
 376         .map_get_next_key = ringbuf_map_get_next_key,
 377         .map_mem_usage = ringbuf_map_mem_usage,
 378         .map_btf_id = &user_ringbuf_map_btf_ids[0],
 379 };
 380
 381 /* Given pointer to ring buffer record metadata and struct bpf_ringbuf itself,
 382  * calculate offset from record metadata to ring buffer in pages, rounded
 383  * down. This page offset is stored as part of record metadata and allows to
 384  * restore struct bpf_ringbuf * from record pointer. This page offset is
 385  * stored at offset 4 of record metadata header.
 386  */
 387 static size_t bpf_ringbuf_rec_pg_off(struct bpf_ringbuf *rb,
 388                                      struct bpf_ringbuf_hdr *hdr)
 389 {
 390         return ((void *)hdr - (void *)rb) >> PAGE_SHIFT;
 391 }
 392
 393 /* Given pointer to ring buffer record header, restore pointer to struct
 394  * bpf_ringbuf itself by using page offset stored at offset 4
 395  */
 396 static struct bpf_ringbuf *
 397 bpf_ringbuf_restore_from_rec(struct bpf_ringbuf_hdr *hdr)
 398 {
 399         unsigned long addr = (unsigned long)(void *)hdr;
 400         unsigned long off = (unsigned long)hdr->pg_off << PAGE_SHIFT;
 401
 402         return (void*)((addr & PAGE_MASK) - off);
 403 }
 404
 405 static void *__bpf_ringbuf_reserve(struct bpf_ringbuf *rb, u64 size)
 406 {
 407         unsigned long cons_pos, prod_pos, new_prod_pos, flags;
 408         u32 len, pg_off;
 409         struct bpf_ringbuf_hdr *hdr;
 410
 411         if (unlikely(size > RINGBUF_MAX_RECORD_SZ))
 412                 return NULL;
 413
 414         len = round_up(size + BPF_RINGBUF_HDR_SZ, 8);
 415         if (len > ringbuf_total_data_sz(rb))
 416                 return NULL;
 417
 418         cons_pos = smp_load_acquire(&rb->consumer_pos);
 419
 420         if (in_nmi()) {
 421                 if (!spin_trylock_irqsave(&rb->spinlock, flags))
 422                         return NULL;
 423         } else {
 424                 spin_lock_irqsave(&rb->spinlock, flags);
 425         }
 426
 427         prod_pos = rb->producer_pos;
 428         new_prod_pos = prod_pos + len;
 429
 430         /* check for out of ringbuf space by ensuring producer position
 431          * doesn't advance more than (ringbuf_size - 1) ahead
 432          */
 433         if (new_prod_pos - cons_pos > rb->mask) {
 434                 spin_unlock_irqrestore(&rb->spinlock, flags);
 435                 return NULL;
 436         }
 437
 438         hdr = (void *)rb->data + (prod_pos & rb->mask);
 439         pg_off = bpf_ringbuf_rec_pg_off(rb, hdr);
 440         hdr->len = size | BPF_RINGBUF_BUSY_BIT;
 441         hdr->pg_off = pg_off;
 442
 443         /* pairs with consumer's smp_load_acquire() */
 444         smp_store_release(&rb->producer_pos, new_prod_pos);
 445
 446         spin_unlock_irqrestore(&rb->spinlock, flags);
 447
 448         return (void *)hdr + BPF_RINGBUF_HDR_SZ;
 449 }
 450
 451 BPF_CALL_3(bpf_ringbuf_reserve, struct bpf_map *, map, u64, size, u64, flags)
 452 {
 453         struct bpf_ringbuf_map *rb_map;
 454
 455         if (unlikely(flags))
 456                 return 0;
 457
 458         rb_map = container_of(map, struct bpf_ringbuf_map, map);
 459         return (unsigned long)__bpf_ringbuf_reserve(rb_map->rb, size);
 460 }
 461
 462 const struct bpf_func_proto bpf_ringbuf_reserve_proto = {
 463         .func           = bpf_ringbuf_reserve,
 464         .ret_type       = RET_PTR_TO_RINGBUF_MEM_OR_NULL,
 465         .arg1_type      = ARG_CONST_MAP_PTR,
 466         .arg2_type      = ARG_CONST_ALLOC_SIZE_OR_ZERO,
 467         .arg3_type      = ARG_ANYTHING,
 468 };
 469
 470 static void bpf_ringbuf_commit(void *sample, u64 flags, bool discard)
 471 {
 472         unsigned long rec_pos, cons_pos;
 473         struct bpf_ringbuf_hdr *hdr;
 474         struct bpf_ringbuf *rb;
 475         u32 new_len;
 476
 477         hdr = sample - BPF_RINGBUF_HDR_SZ;
 478         rb = bpf_ringbuf_restore_from_rec(hdr);
 479         new_len = hdr->len ^ BPF_RINGBUF_BUSY_BIT;
 480         if (discard)
 481                 new_len |= BPF_RINGBUF_DISCARD_BIT;
 482
 483         /* update record header with correct final size prefix */
 484         xchg(&hdr->len, new_len);
 485
 486         /* if consumer caught up and is waiting for our record, notify about
 487          * new data availability
 488          */
 489         rec_pos = (void *)hdr - (void *)rb->data;
 490         cons_pos = smp_load_acquire(&rb->consumer_pos) & rb->mask;
 491
 492         if (flags & BPF_RB_FORCE_WAKEUP)
 493                 irq_work_queue(&rb->work);
 494         else if (cons_pos == rec_pos && !(flags & BPF_RB_NO_WAKEUP))
 495                 irq_work_queue(&rb->work);
 496 }
 497
 498 BPF_CALL_2(bpf_ringbuf_submit, void *, sample, u64, flags)
 499 {
 500         bpf_ringbuf_commit(sample, flags, false /* discard */);
 501         return 0;
 502 }
 503
 504 const struct bpf_func_proto bpf_ringbuf_submit_proto = {
 505         .func           = bpf_ringbuf_submit,
 506         .ret_type       = RET_VOID,
 507         .arg1_type      = ARG_PTR_TO_RINGBUF_MEM | OBJ_RELEASE,
 508         .arg2_type      = ARG_ANYTHING,
 509 };
 510
 511 BPF_CALL_2(bpf_ringbuf_discard, void *, sample, u64, flags)
 512 {
 513         bpf_ringbuf_commit(sample, flags, true /* discard */);
 514         return 0;
 515 }
 516
 517 const struct bpf_func_proto bpf_ringbuf_discard_proto = {
 518         .func           = bpf_ringbuf_discard,
 519         .ret_type       = RET_VOID,
 520         .arg1_type      = ARG_PTR_TO_RINGBUF_MEM | OBJ_RELEASE,
 521         .arg2_type      = ARG_ANYTHING,
 522 };
 523
 524 BPF_CALL_4(bpf_ringbuf_output, struct bpf_map *, map, void *, data, u64, size,
 525            u64, flags)
 526 {
 527         struct bpf_ringbuf_map *rb_map;
 528         void *rec;
 529
 530         if (unlikely(flags & ~(BPF_RB_NO_WAKEUP | BPF_RB_FORCE_WAKEUP)))
 531                 return -EINVAL;
 532
 533         rb_map = container_of(map, struct bpf_ringbuf_map, map);
 534         rec = __bpf_ringbuf_reserve(rb_map->rb, size);
 535         if (!rec)
 536                 return -EAGAIN;
 537
 538         memcpy(rec, data, size);
 539         bpf_ringbuf_commit(rec, flags, false /* discard */);
 540         return 0;
 541 }
 542
 543 const struct bpf_func_proto bpf_ringbuf_output_proto = {
 544         .func           = bpf_ringbuf_output,
 545         .ret_type       = RET_INTEGER,
 546         .arg1_type      = ARG_CONST_MAP_PTR,
 547         .arg2_type      = ARG_PTR_TO_MEM | MEM_RDONLY,
 548         .arg3_type      = ARG_CONST_SIZE_OR_ZERO,
 549         .arg4_type      = ARG_ANYTHING,
 550 };
 551
 552 BPF_CALL_2(bpf_ringbuf_query, struct bpf_map *, map, u64, flags)
 553 {
 554         struct bpf_ringbuf *rb;
 555
 556         rb = container_of(map, struct bpf_ringbuf_map, map)->rb;
 557
 558         switch (flags) {
 559         case BPF_RB_AVAIL_DATA:
 560                 return ringbuf_avail_data_sz(rb);
 561         case BPF_RB_RING_SIZE:
 562                 return ringbuf_total_data_sz(rb);
 563         case BPF_RB_CONS_POS:
 564                 return smp_load_acquire(&rb->consumer_pos);
 565         case BPF_RB_PROD_POS:
 566                 return smp_load_acquire(&rb->producer_pos);
 567         default:
 568                 return 0;
 569         }
 570 }
 571
 572 const struct bpf_func_proto bpf_ringbuf_query_proto = {
 573         .func           = bpf_ringbuf_query,
 574         .ret_type       = RET_INTEGER,
 575         .arg1_type      = ARG_CONST_MAP_PTR,
 576         .arg2_type      = ARG_ANYTHING,
 577 };
 578
 579 BPF_CALL_4(bpf_ringbuf_reserve_dynptr, struct bpf_map *, map, u32, size, u64, flags,
 580            struct bpf_dynptr_kern *, ptr)
 581 {
 582         struct bpf_ringbuf_map *rb_map;
 583         void *sample;
 584         int err;
 585
 586         if (unlikely(flags)) {
 587                 bpf_dynptr_set_null(ptr);
 588                 return -EINVAL;
 589         }
 590
 591         err = bpf_dynptr_check_size(size);
 592         if (err) {
 593                 bpf_dynptr_set_null(ptr);
 594                 return err;
 595         }
 596
 597         rb_map = container_of(map, struct bpf_ringbuf_map, map);
 598
 599         sample = __bpf_ringbuf_reserve(rb_map->rb, size);
 600         if (!sample) {
 601                 bpf_dynptr_set_null(ptr);
 602                 return -EINVAL;
 603         }
 604
 605         bpf_dynptr_init(ptr, sample, BPF_DYNPTR_TYPE_RINGBUF, 0, size);
 606
 607         return 0;
 608 }
 609
 610 const struct bpf_func_proto bpf_ringbuf_reserve_dynptr_proto = {
 611         .func           = bpf_ringbuf_reserve_dynptr,
 612         .ret_type       = RET_INTEGER,
 613         .arg1_type      = ARG_CONST_MAP_PTR,
 614         .arg2_type      = ARG_ANYTHING,
 615         .arg3_type      = ARG_ANYTHING,
 616         .arg4_type      = ARG_PTR_TO_DYNPTR | DYNPTR_TYPE_RINGBUF | MEM_UNINIT,
 617 };
 618
 619 BPF_CALL_2(bpf_ringbuf_submit_dynptr, struct bpf_dynptr_kern *, ptr, u64, flags)
 620 {
 621         if (!ptr->data)
 622                 return 0;
 623
 624         bpf_ringbuf_commit(ptr->data, flags, false /* discard */);
 625
 626         bpf_dynptr_set_null(ptr);
 627
 628         return 0;
 629 }
 630
 631 const struct bpf_func_proto bpf_ringbuf_submit_dynptr_proto = {
 632         .func           = bpf_ringbuf_submit_dynptr,
 633         .ret_type       = RET_VOID,
 634         .arg1_type      = ARG_PTR_TO_DYNPTR | DYNPTR_TYPE_RINGBUF | OBJ_RELEASE,
 635         .arg2_type      = ARG_ANYTHING,
 636 };
 637
 638 BPF_CALL_2(bpf_ringbuf_discard_dynptr, struct bpf_dynptr_kern *, ptr, u64, flags)
 639 {
 640         if (!ptr->data)
 641                 return 0;
 642
 643         bpf_ringbuf_commit(ptr->data, flags, true /* discard */);
 644
 645         bpf_dynptr_set_null(ptr);
 646
 647         return 0;
 648 }
 649
 650 const struct bpf_func_proto bpf_ringbuf_discard_dynptr_proto = {
 651         .func           = bpf_ringbuf_discard_dynptr,
 652         .ret_type       = RET_VOID,
 653         .arg1_type      = ARG_PTR_TO_DYNPTR | DYNPTR_TYPE_RINGBUF | OBJ_RELEASE,
 654         .arg2_type      = ARG_ANYTHING,
 655 };
 656
 657 static int __bpf_user_ringbuf_peek(struct bpf_ringbuf *rb, void **sample, u32 *size)
 658 {
 659         int err;
 660         u32 hdr_len, sample_len, total_len, flags, *hdr;
 661         u64 cons_pos, prod_pos;
 662
 663         /* Synchronizes with smp_store_release() in user-space producer. */
 664         prod_pos = smp_load_acquire(&rb->producer_pos);
 665         if (prod_pos % 8)
 666                 return -EINVAL;
 667
 668         /* Synchronizes with smp_store_release() in __bpf_user_ringbuf_sample_release() */
 669         cons_pos = smp_load_acquire(&rb->consumer_pos);
 670         if (cons_pos >= prod_pos)
 671                 return -ENODATA;
 672
 673         hdr = (u32 *)((uintptr_t)rb->data + (uintptr_t)(cons_pos & rb->mask));
 674         /* Synchronizes with smp_store_release() in user-space producer. */
 675         hdr_len = smp_load_acquire(hdr);
 676         flags = hdr_len & (BPF_RINGBUF_BUSY_BIT | BPF_RINGBUF_DISCARD_BIT);
 677         sample_len = hdr_len & ~flags;
 678         total_len = round_up(sample_len + BPF_RINGBUF_HDR_SZ, 8);
 679
 680         /* The sample must fit within the region advertised by the producer position. */
 681         if (total_len > prod_pos - cons_pos)
 682                 return -EINVAL;
 683
 684         /* The sample must fit within the data region of the ring buffer. */
 685         if (total_len > ringbuf_total_data_sz(rb))
 686                 return -E2BIG;
 687
 688         /* The sample must fit into a struct bpf_dynptr. */
 689         err = bpf_dynptr_check_size(sample_len);
 690         if (err)
 691                 return -E2BIG;
 692
 693         if (flags & BPF_RINGBUF_DISCARD_BIT) {
 694                 /* If the discard bit is set, the sample should be skipped.
 695                  *
 696                  * Update the consumer pos, and return -EAGAIN so the caller
 697                  * knows to skip this sample and try to read the next one.
 698                  */
 699                 smp_store_release(&rb->consumer_pos, cons_pos + total_len);
 700                 return -EAGAIN;
 701         }
 702
 703         if (flags & BPF_RINGBUF_BUSY_BIT)
 704                 return -ENODATA;
 705
 706         *sample = (void *)((uintptr_t)rb->data +
 707                            (uintptr_t)((cons_pos + BPF_RINGBUF_HDR_SZ) & rb->mask));
 708         *size = sample_len;
 709         return 0;
 710 }
 711
 712 static void __bpf_user_ringbuf_sample_release(struct bpf_ringbuf *rb, size_t size, u64 flags)
 713 {
 714         u64 consumer_pos;
 715         u32 rounded_size = round_up(size + BPF_RINGBUF_HDR_SZ, 8);
 716
 717         /* Using smp_load_acquire() is unnecessary here, as the busy-bit
 718          * prevents another task from writing to consumer_pos after it was read
 719          * by this task with smp_load_acquire() in __bpf_user_ringbuf_peek().
 720          */
 721         consumer_pos = rb->consumer_pos;
 722          /* Synchronizes with smp_load_acquire() in user-space producer. */
 723         smp_store_release(&rb->consumer_pos, consumer_pos + rounded_size);
 724 }
 725
 726 BPF_CALL_4(bpf_user_ringbuf_drain, struct bpf_map *, map,
 727            void *, callback_fn, void *, callback_ctx, u64, flags)
 728 {
 729         struct bpf_ringbuf *rb;
 730         long samples, discarded_samples = 0, ret = 0;
 731         bpf_callback_t callback = (bpf_callback_t)callback_fn;
 732         u64 wakeup_flags = BPF_RB_NO_WAKEUP | BPF_RB_FORCE_WAKEUP;
 733         int busy = 0;
 734
 735         if (unlikely(flags & ~wakeup_flags))
 736                 return -EINVAL;
 737
 738         rb = container_of(map, struct bpf_ringbuf_map, map)->rb;
 739
 740         /* If another consumer is already consuming a sample, wait for them to finish. */
 741         if (!atomic_try_cmpxchg(&rb->busy, &busy, 1))
 742                 return -EBUSY;
 743
 744         for (samples = 0; samples < BPF_MAX_USER_RINGBUF_SAMPLES && ret == 0; samples++) {
 745                 int err;
 746                 u32 size;
 747                 void *sample;
 748                 struct bpf_dynptr_kern dynptr;
 749
 750                 err = __bpf_user_ringbuf_peek(rb, &sample, &size);
 751                 if (err) {
 752                         if (err == -ENODATA) {
 753                                 break;
 754                         } else if (err == -EAGAIN) {
 755                                 discarded_samples++;
 756                                 continue;
 757                         } else {
 758                                 ret = err;
 759                                 goto schedule_work_return;
 760                         }
 761                 }
 762
 763                 bpf_dynptr_init(&dynptr, sample, BPF_DYNPTR_TYPE_LOCAL, 0, size);
 764                 ret = callback((uintptr_t)&dynptr, (uintptr_t)callback_ctx, 0, 0, 0);
 765                 __bpf_user_ringbuf_sample_release(rb, size, flags);
 766         }
 767         ret = samples - discarded_samples;
 768
 769 schedule_work_return:
 770         /* Prevent the clearing of the busy-bit from being reordered before the
 771          * storing of any rb consumer or producer positions.
 772          */
 773         smp_mb__before_atomic();
 774         atomic_set(&rb->busy, 0);
 775
 776         if (flags & BPF_RB_FORCE_WAKEUP)
 777                 irq_work_queue(&rb->work);
 778         else if (!(flags & BPF_RB_NO_WAKEUP) && samples > 0)
 779                 irq_work_queue(&rb->work);
 780         return ret;
 781 }
 782
 783 const struct bpf_func_proto bpf_user_ringbuf_drain_proto = {
 784         .func           = bpf_user_ringbuf_drain,
 785         .ret_type       = RET_INTEGER,
 786         .arg1_type      = ARG_CONST_MAP_PTR,
 787         .arg2_type      = ARG_PTR_TO_FUNC,
 788         .arg3_type      = ARG_PTR_TO_STACK_OR_NULL,
 789         .arg4_type      = ARG_ANYTHING,
 790 };