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_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 kmemleak_not_leak(pages); 110 rb->pages = pages; 111 rb->nr_pages = nr_pages; 112 return rb; 113 } 114 115 err_free_pages: 116 for (i = 0; i < nr_pages; i++) 117 __free_page(pages[i]); 118 kvfree(pages); 119 return NULL; 120 } 121 122 static void bpf_ringbuf_notify(struct irq_work *work) 123 { 124 struct bpf_ringbuf *rb = container_of(work, struct bpf_ringbuf, work); 125 126 wake_up_all(&rb->waitq); 127 } 128 129 static struct bpf_ringbuf *bpf_ringbuf_alloc(size_t data_sz, int numa_node) 130 { 131 struct bpf_ringbuf *rb; 132 133 rb = bpf_ringbuf_area_alloc(data_sz, numa_node); 134 if (!rb) 135 return NULL; 136 137 spin_lock_init(&rb->spinlock); 138 init_waitqueue_head(&rb->waitq); 139 init_irq_work(&rb->work, bpf_ringbuf_notify); 140 141 rb->mask = data_sz - 1; 142 rb->consumer_pos = 0; 143 rb->producer_pos = 0; 144 145 return rb; 146 } 147 148 static struct bpf_map *ringbuf_map_alloc(union bpf_attr *attr) 149 { 150 struct bpf_ringbuf_map *rb_map; 151 152 if (attr->map_flags & ~RINGBUF_CREATE_FLAG_MASK) 153 return ERR_PTR(-EINVAL); 154 155 if (attr->key_size || attr->value_size || 156 !is_power_of_2(attr->max_entries) || 157 !PAGE_ALIGNED(attr->max_entries)) 158 return ERR_PTR(-EINVAL); 159 160 #ifdef CONFIG_64BIT 161 /* on 32-bit arch, it's impossible to overflow record's hdr->pgoff */ 162 if (attr->max_entries > RINGBUF_MAX_DATA_SZ) 163 return ERR_PTR(-E2BIG); 164 #endif 165 166 rb_map = kzalloc(sizeof(*rb_map), GFP_USER | __GFP_ACCOUNT); 167 if (!rb_map) 168 return ERR_PTR(-ENOMEM); 169 170 bpf_map_init_from_attr(&rb_map->map, attr); 171 172 rb_map->rb = bpf_ringbuf_alloc(attr->max_entries, rb_map->map.numa_node); 173 if (!rb_map->rb) { 174 kfree(rb_map); 175 return ERR_PTR(-ENOMEM); 176 } 177 178 return &rb_map->map; 179 } 180 181 static void bpf_ringbuf_free(struct bpf_ringbuf *rb) 182 { 183 /* copy pages pointer and nr_pages to local variable, as we are going 184 * to unmap rb itself with vunmap() below 185 */ 186 struct page **pages = rb->pages; 187 int i, nr_pages = rb->nr_pages; 188 189 vunmap(rb); 190 for (i = 0; i < nr_pages; i++) 191 __free_page(pages[i]); 192 kvfree(pages); 193 } 194 195 static void ringbuf_map_free(struct bpf_map *map) 196 { 197 struct bpf_ringbuf_map *rb_map; 198 199 rb_map = container_of(map, struct bpf_ringbuf_map, map); 200 bpf_ringbuf_free(rb_map->rb); 201 kfree(rb_map); 202 } 203 204 static void *ringbuf_map_lookup_elem(struct bpf_map *map, void *key) 205 { 206 return ERR_PTR(-ENOTSUPP); 207 } 208 209 static int ringbuf_map_update_elem(struct bpf_map *map, void *key, void *value, 210 u64 flags) 211 { 212 return -ENOTSUPP; 213 } 214 215 static int ringbuf_map_delete_elem(struct bpf_map *map, void *key) 216 { 217 return -ENOTSUPP; 218 } 219 220 static int ringbuf_map_get_next_key(struct bpf_map *map, void *key, 221 void *next_key) 222 { 223 return -ENOTSUPP; 224 } 225 226 static int ringbuf_map_mmap(struct bpf_map *map, struct vm_area_struct *vma) 227 { 228 struct bpf_ringbuf_map *rb_map; 229 230 rb_map = container_of(map, struct bpf_ringbuf_map, map); 231 232 if (vma->vm_flags & VM_WRITE) { 233 /* allow writable mapping for the consumer_pos only */ 234 if (vma->vm_pgoff != 0 || vma->vm_end - vma->vm_start != PAGE_SIZE) 235 return -EPERM; 236 } else { 237 vma->vm_flags &= ~VM_MAYWRITE; 238 } 239 /* remap_vmalloc_range() checks size and offset constraints */ 240 return remap_vmalloc_range(vma, rb_map->rb, 241 vma->vm_pgoff + RINGBUF_PGOFF); 242 } 243 244 static unsigned long ringbuf_avail_data_sz(struct bpf_ringbuf *rb) 245 { 246 unsigned long cons_pos, prod_pos; 247 248 cons_pos = smp_load_acquire(&rb->consumer_pos); 249 prod_pos = smp_load_acquire(&rb->producer_pos); 250 return prod_pos - cons_pos; 251 } 252 253 static __poll_t ringbuf_map_poll(struct bpf_map *map, struct file *filp, 254 struct poll_table_struct *pts) 255 { 256 struct bpf_ringbuf_map *rb_map; 257 258 rb_map = container_of(map, struct bpf_ringbuf_map, map); 259 poll_wait(filp, &rb_map->rb->waitq, pts); 260 261 if (ringbuf_avail_data_sz(rb_map->rb)) 262 return EPOLLIN | EPOLLRDNORM; 263 return 0; 264 } 265 266 static int ringbuf_map_btf_id; 267 const struct bpf_map_ops ringbuf_map_ops = { 268 .map_meta_equal = bpf_map_meta_equal, 269 .map_alloc = ringbuf_map_alloc, 270 .map_free = ringbuf_map_free, 271 .map_mmap = ringbuf_map_mmap, 272 .map_poll = ringbuf_map_poll, 273 .map_lookup_elem = ringbuf_map_lookup_elem, 274 .map_update_elem = ringbuf_map_update_elem, 275 .map_delete_elem = ringbuf_map_delete_elem, 276 .map_get_next_key = ringbuf_map_get_next_key, 277 .map_btf_name = "bpf_ringbuf_map", 278 .map_btf_id = &ringbuf_map_btf_id, 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, 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, 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, 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