1 // SPDX-License-Identifier: (GPL-2.0-only OR BSD-2-Clause) 2 /* Copyright (C) 2016-2018 Netronome Systems, Inc. */ 3 4 /* 5 * nfp_net_offload.c 6 * Netronome network device driver: TC offload functions for PF and VF 7 */ 8 9 #define pr_fmt(fmt) "NFP net bpf: " fmt 10 11 #include <linux/bpf.h> 12 #include <linux/kernel.h> 13 #include <linux/netdevice.h> 14 #include <linux/pci.h> 15 #include <linux/jiffies.h> 16 #include <linux/timer.h> 17 #include <linux/list.h> 18 #include <linux/mm.h> 19 20 #include <net/pkt_cls.h> 21 #include <net/tc_act/tc_gact.h> 22 #include <net/tc_act/tc_mirred.h> 23 24 #include "main.h" 25 #include "../ccm.h" 26 #include "../nfp_app.h" 27 #include "../nfp_net_ctrl.h" 28 #include "../nfp_net.h" 29 30 static int 31 nfp_map_ptr_record(struct nfp_app_bpf *bpf, struct nfp_prog *nfp_prog, 32 struct bpf_map *map) 33 { 34 struct nfp_bpf_neutral_map *record; 35 int err; 36 37 /* Reuse path - other offloaded program is already tracking this map. */ 38 record = rhashtable_lookup_fast(&bpf->maps_neutral, &map->id, 39 nfp_bpf_maps_neutral_params); 40 if (record) { 41 nfp_prog->map_records[nfp_prog->map_records_cnt++] = record; 42 record->count++; 43 return 0; 44 } 45 46 /* Grab a single ref to the map for our record. The prog destroy ndo 47 * happens after free_used_maps(). 48 */ 49 bpf_map_inc(map); 50 51 record = kmalloc(sizeof(*record), GFP_KERNEL); 52 if (!record) { 53 err = -ENOMEM; 54 goto err_map_put; 55 } 56 57 record->ptr = map; 58 record->map_id = map->id; 59 record->count = 1; 60 61 err = rhashtable_insert_fast(&bpf->maps_neutral, &record->l, 62 nfp_bpf_maps_neutral_params); 63 if (err) 64 goto err_free_rec; 65 66 nfp_prog->map_records[nfp_prog->map_records_cnt++] = record; 67 68 return 0; 69 70 err_free_rec: 71 kfree(record); 72 err_map_put: 73 bpf_map_put(map); 74 return err; 75 } 76 77 static void 78 nfp_map_ptrs_forget(struct nfp_app_bpf *bpf, struct nfp_prog *nfp_prog) 79 { 80 bool freed = false; 81 int i; 82 83 for (i = 0; i < nfp_prog->map_records_cnt; i++) { 84 if (--nfp_prog->map_records[i]->count) { 85 nfp_prog->map_records[i] = NULL; 86 continue; 87 } 88 89 WARN_ON(rhashtable_remove_fast(&bpf->maps_neutral, 90 &nfp_prog->map_records[i]->l, 91 nfp_bpf_maps_neutral_params)); 92 freed = true; 93 } 94 95 if (freed) { 96 synchronize_rcu(); 97 98 for (i = 0; i < nfp_prog->map_records_cnt; i++) 99 if (nfp_prog->map_records[i]) { 100 bpf_map_put(nfp_prog->map_records[i]->ptr); 101 kfree(nfp_prog->map_records[i]); 102 } 103 } 104 105 kfree(nfp_prog->map_records); 106 nfp_prog->map_records = NULL; 107 nfp_prog->map_records_cnt = 0; 108 } 109 110 static int 111 nfp_map_ptrs_record(struct nfp_app_bpf *bpf, struct nfp_prog *nfp_prog, 112 struct bpf_prog *prog) 113 { 114 int i, cnt, err; 115 116 /* Quickly count the maps we will have to remember */ 117 cnt = 0; 118 for (i = 0; i < prog->aux->used_map_cnt; i++) 119 if (bpf_map_offload_neutral(prog->aux->used_maps[i])) 120 cnt++; 121 if (!cnt) 122 return 0; 123 124 nfp_prog->map_records = kmalloc_array(cnt, 125 sizeof(nfp_prog->map_records[0]), 126 GFP_KERNEL); 127 if (!nfp_prog->map_records) 128 return -ENOMEM; 129 130 for (i = 0; i < prog->aux->used_map_cnt; i++) 131 if (bpf_map_offload_neutral(prog->aux->used_maps[i])) { 132 err = nfp_map_ptr_record(bpf, nfp_prog, 133 prog->aux->used_maps[i]); 134 if (err) { 135 nfp_map_ptrs_forget(bpf, nfp_prog); 136 return err; 137 } 138 } 139 WARN_ON(cnt != nfp_prog->map_records_cnt); 140 141 return 0; 142 } 143 144 static int 145 nfp_prog_prepare(struct nfp_prog *nfp_prog, const struct bpf_insn *prog, 146 unsigned int cnt) 147 { 148 struct nfp_insn_meta *meta; 149 unsigned int i; 150 151 for (i = 0; i < cnt; i++) { 152 meta = kzalloc(sizeof(*meta), GFP_KERNEL); 153 if (!meta) 154 return -ENOMEM; 155 156 meta->insn = prog[i]; 157 meta->n = i; 158 if (is_mbpf_alu(meta)) { 159 meta->umin_src = U64_MAX; 160 meta->umin_dst = U64_MAX; 161 } 162 163 list_add_tail(&meta->l, &nfp_prog->insns); 164 } 165 nfp_prog->n_insns = cnt; 166 167 nfp_bpf_jit_prepare(nfp_prog); 168 169 return 0; 170 } 171 172 static void nfp_prog_free(struct nfp_prog *nfp_prog) 173 { 174 struct nfp_insn_meta *meta, *tmp; 175 176 kfree(nfp_prog->subprog); 177 178 list_for_each_entry_safe(meta, tmp, &nfp_prog->insns, l) { 179 list_del(&meta->l); 180 kfree(meta); 181 } 182 kfree(nfp_prog); 183 } 184 185 static int nfp_bpf_verifier_prep(struct bpf_prog *prog) 186 { 187 struct nfp_prog *nfp_prog; 188 int ret; 189 190 nfp_prog = kzalloc(sizeof(*nfp_prog), GFP_KERNEL); 191 if (!nfp_prog) 192 return -ENOMEM; 193 prog->aux->offload->dev_priv = nfp_prog; 194 195 INIT_LIST_HEAD(&nfp_prog->insns); 196 nfp_prog->type = prog->type; 197 nfp_prog->bpf = bpf_offload_dev_priv(prog->aux->offload->offdev); 198 199 ret = nfp_prog_prepare(nfp_prog, prog->insnsi, prog->len); 200 if (ret) 201 goto err_free; 202 203 nfp_prog->verifier_meta = nfp_prog_first_meta(nfp_prog); 204 205 return 0; 206 207 err_free: 208 nfp_prog_free(nfp_prog); 209 210 return ret; 211 } 212 213 static int nfp_bpf_translate(struct bpf_prog *prog) 214 { 215 struct nfp_net *nn = netdev_priv(prog->aux->offload->netdev); 216 struct nfp_prog *nfp_prog = prog->aux->offload->dev_priv; 217 unsigned int max_instr; 218 int err; 219 220 /* We depend on dead code elimination succeeding */ 221 if (prog->aux->offload->opt_failed) 222 return -EINVAL; 223 224 max_instr = nn_readw(nn, NFP_NET_CFG_BPF_MAX_LEN); 225 nfp_prog->__prog_alloc_len = max_instr * sizeof(u64); 226 227 nfp_prog->prog = kvmalloc(nfp_prog->__prog_alloc_len, GFP_KERNEL); 228 if (!nfp_prog->prog) 229 return -ENOMEM; 230 231 err = nfp_bpf_jit(nfp_prog); 232 if (err) 233 return err; 234 235 prog->aux->offload->jited_len = nfp_prog->prog_len * sizeof(u64); 236 prog->aux->offload->jited_image = nfp_prog->prog; 237 238 return nfp_map_ptrs_record(nfp_prog->bpf, nfp_prog, prog); 239 } 240 241 static void nfp_bpf_destroy(struct bpf_prog *prog) 242 { 243 struct nfp_prog *nfp_prog = prog->aux->offload->dev_priv; 244 245 kvfree(nfp_prog->prog); 246 nfp_map_ptrs_forget(nfp_prog->bpf, nfp_prog); 247 nfp_prog_free(nfp_prog); 248 } 249 250 /* Atomic engine requires values to be in big endian, we need to byte swap 251 * the value words used with xadd. 252 */ 253 static void nfp_map_bpf_byte_swap(struct nfp_bpf_map *nfp_map, void *value) 254 { 255 u32 *word = value; 256 unsigned int i; 257 258 for (i = 0; i < DIV_ROUND_UP(nfp_map->offmap->map.value_size, 4); i++) 259 if (nfp_map->use_map[i].type == NFP_MAP_USE_ATOMIC_CNT) 260 word[i] = (__force u32)cpu_to_be32(word[i]); 261 } 262 263 /* Mark value as unsafely initialized in case it becomes atomic later 264 * and we didn't byte swap something non-byte swap neutral. 265 */ 266 static void 267 nfp_map_bpf_byte_swap_record(struct nfp_bpf_map *nfp_map, void *value) 268 { 269 u32 *word = value; 270 unsigned int i; 271 272 for (i = 0; i < DIV_ROUND_UP(nfp_map->offmap->map.value_size, 4); i++) 273 if (nfp_map->use_map[i].type == NFP_MAP_UNUSED && 274 word[i] != (__force u32)cpu_to_be32(word[i])) 275 nfp_map->use_map[i].non_zero_update = 1; 276 } 277 278 static int 279 nfp_bpf_map_lookup_entry(struct bpf_offloaded_map *offmap, 280 void *key, void *value) 281 { 282 int err; 283 284 err = nfp_bpf_ctrl_lookup_entry(offmap, key, value); 285 if (err) 286 return err; 287 288 nfp_map_bpf_byte_swap(offmap->dev_priv, value); 289 return 0; 290 } 291 292 static int 293 nfp_bpf_map_update_entry(struct bpf_offloaded_map *offmap, 294 void *key, void *value, u64 flags) 295 { 296 nfp_map_bpf_byte_swap(offmap->dev_priv, value); 297 nfp_map_bpf_byte_swap_record(offmap->dev_priv, value); 298 return nfp_bpf_ctrl_update_entry(offmap, key, value, flags); 299 } 300 301 static int 302 nfp_bpf_map_get_next_key(struct bpf_offloaded_map *offmap, 303 void *key, void *next_key) 304 { 305 if (!key) 306 return nfp_bpf_ctrl_getfirst_entry(offmap, next_key); 307 return nfp_bpf_ctrl_getnext_entry(offmap, key, next_key); 308 } 309 310 static int 311 nfp_bpf_map_delete_elem(struct bpf_offloaded_map *offmap, void *key) 312 { 313 if (offmap->map.map_type == BPF_MAP_TYPE_ARRAY) 314 return -EINVAL; 315 return nfp_bpf_ctrl_del_entry(offmap, key); 316 } 317 318 static const struct bpf_map_dev_ops nfp_bpf_map_ops = { 319 .map_get_next_key = nfp_bpf_map_get_next_key, 320 .map_lookup_elem = nfp_bpf_map_lookup_entry, 321 .map_update_elem = nfp_bpf_map_update_entry, 322 .map_delete_elem = nfp_bpf_map_delete_elem, 323 }; 324 325 static int 326 nfp_bpf_map_alloc(struct nfp_app_bpf *bpf, struct bpf_offloaded_map *offmap) 327 { 328 struct nfp_bpf_map *nfp_map; 329 unsigned int use_map_size; 330 long long int res; 331 332 if (!bpf->maps.types) 333 return -EOPNOTSUPP; 334 335 if (offmap->map.map_flags || 336 offmap->map.numa_node != NUMA_NO_NODE) { 337 pr_info("map flags are not supported\n"); 338 return -EINVAL; 339 } 340 341 if (!(bpf->maps.types & 1 << offmap->map.map_type)) { 342 pr_info("map type not supported\n"); 343 return -EOPNOTSUPP; 344 } 345 if (bpf->maps.max_maps == bpf->maps_in_use) { 346 pr_info("too many maps for a device\n"); 347 return -ENOMEM; 348 } 349 if (bpf->maps.max_elems - bpf->map_elems_in_use < 350 offmap->map.max_entries) { 351 pr_info("map with too many elements: %u, left: %u\n", 352 offmap->map.max_entries, 353 bpf->maps.max_elems - bpf->map_elems_in_use); 354 return -ENOMEM; 355 } 356 357 if (round_up(offmap->map.key_size, 8) + 358 round_up(offmap->map.value_size, 8) > bpf->maps.max_elem_sz) { 359 pr_info("map elements too large: %u, FW max element size (key+value): %u\n", 360 round_up(offmap->map.key_size, 8) + 361 round_up(offmap->map.value_size, 8), 362 bpf->maps.max_elem_sz); 363 return -ENOMEM; 364 } 365 if (offmap->map.key_size > bpf->maps.max_key_sz) { 366 pr_info("map key size %u, FW max is %u\n", 367 offmap->map.key_size, bpf->maps.max_key_sz); 368 return -ENOMEM; 369 } 370 if (offmap->map.value_size > bpf->maps.max_val_sz) { 371 pr_info("map value size %u, FW max is %u\n", 372 offmap->map.value_size, bpf->maps.max_val_sz); 373 return -ENOMEM; 374 } 375 376 use_map_size = DIV_ROUND_UP(offmap->map.value_size, 4) * 377 sizeof_field(struct nfp_bpf_map, use_map[0]); 378 379 nfp_map = kzalloc(sizeof(*nfp_map) + use_map_size, GFP_USER); 380 if (!nfp_map) 381 return -ENOMEM; 382 383 offmap->dev_priv = nfp_map; 384 nfp_map->offmap = offmap; 385 nfp_map->bpf = bpf; 386 spin_lock_init(&nfp_map->cache_lock); 387 388 res = nfp_bpf_ctrl_alloc_map(bpf, &offmap->map); 389 if (res < 0) { 390 kfree(nfp_map); 391 return res; 392 } 393 394 nfp_map->tid = res; 395 offmap->dev_ops = &nfp_bpf_map_ops; 396 bpf->maps_in_use++; 397 bpf->map_elems_in_use += offmap->map.max_entries; 398 list_add_tail(&nfp_map->l, &bpf->map_list); 399 400 return 0; 401 } 402 403 static int 404 nfp_bpf_map_free(struct nfp_app_bpf *bpf, struct bpf_offloaded_map *offmap) 405 { 406 struct nfp_bpf_map *nfp_map = offmap->dev_priv; 407 408 nfp_bpf_ctrl_free_map(bpf, nfp_map); 409 dev_consume_skb_any(nfp_map->cache); 410 WARN_ON_ONCE(nfp_map->cache_blockers); 411 list_del_init(&nfp_map->l); 412 bpf->map_elems_in_use -= offmap->map.max_entries; 413 bpf->maps_in_use--; 414 kfree(nfp_map); 415 416 return 0; 417 } 418 419 int nfp_ndo_bpf(struct nfp_app *app, struct nfp_net *nn, struct netdev_bpf *bpf) 420 { 421 switch (bpf->command) { 422 case BPF_OFFLOAD_MAP_ALLOC: 423 return nfp_bpf_map_alloc(app->priv, bpf->offmap); 424 case BPF_OFFLOAD_MAP_FREE: 425 return nfp_bpf_map_free(app->priv, bpf->offmap); 426 default: 427 return -EINVAL; 428 } 429 } 430 431 static unsigned long 432 nfp_bpf_perf_event_copy(void *dst, const void *src, 433 unsigned long off, unsigned long len) 434 { 435 memcpy(dst, src + off, len); 436 return 0; 437 } 438 439 int nfp_bpf_event_output(struct nfp_app_bpf *bpf, const void *data, 440 unsigned int len) 441 { 442 struct cmsg_bpf_event *cbe = (void *)data; 443 struct nfp_bpf_neutral_map *record; 444 u32 pkt_size, data_size, map_id; 445 u64 map_id_full; 446 447 if (len < sizeof(struct cmsg_bpf_event)) 448 return -EINVAL; 449 450 pkt_size = be32_to_cpu(cbe->pkt_size); 451 data_size = be32_to_cpu(cbe->data_size); 452 map_id_full = be64_to_cpu(cbe->map_ptr); 453 map_id = map_id_full; 454 455 if (len < sizeof(struct cmsg_bpf_event) + pkt_size + data_size) 456 return -EINVAL; 457 if (cbe->hdr.ver != NFP_CCM_ABI_VERSION) 458 return -EINVAL; 459 460 rcu_read_lock(); 461 record = rhashtable_lookup(&bpf->maps_neutral, &map_id, 462 nfp_bpf_maps_neutral_params); 463 if (!record || map_id_full > U32_MAX) { 464 rcu_read_unlock(); 465 cmsg_warn(bpf, "perf event: map id %lld (0x%llx) not recognized, dropping event\n", 466 map_id_full, map_id_full); 467 return -EINVAL; 468 } 469 470 bpf_event_output(record->ptr, be32_to_cpu(cbe->cpu_id), 471 &cbe->data[round_up(pkt_size, 4)], data_size, 472 cbe->data, pkt_size, nfp_bpf_perf_event_copy); 473 rcu_read_unlock(); 474 475 return 0; 476 } 477 478 static int 479 nfp_net_bpf_load(struct nfp_net *nn, struct bpf_prog *prog, 480 struct netlink_ext_ack *extack) 481 { 482 struct nfp_prog *nfp_prog = prog->aux->offload->dev_priv; 483 unsigned int fw_mtu, pkt_off, max_stack, max_prog_len; 484 dma_addr_t dma_addr; 485 void *img; 486 int err; 487 488 fw_mtu = nn_readb(nn, NFP_NET_CFG_BPF_INL_MTU) * 64 - 32; 489 pkt_off = min(prog->aux->max_pkt_offset, nn->dp.netdev->mtu); 490 if (fw_mtu < pkt_off) { 491 NL_SET_ERR_MSG_MOD(extack, "BPF offload not supported with potential packet access beyond HW packet split boundary"); 492 return -EOPNOTSUPP; 493 } 494 495 max_stack = nn_readb(nn, NFP_NET_CFG_BPF_STACK_SZ) * 64; 496 if (nfp_prog->stack_size > max_stack) { 497 NL_SET_ERR_MSG_MOD(extack, "stack too large"); 498 return -EOPNOTSUPP; 499 } 500 501 max_prog_len = nn_readw(nn, NFP_NET_CFG_BPF_MAX_LEN); 502 if (nfp_prog->prog_len > max_prog_len) { 503 NL_SET_ERR_MSG_MOD(extack, "program too long"); 504 return -EOPNOTSUPP; 505 } 506 507 img = nfp_bpf_relo_for_vnic(nfp_prog, nn->app_priv); 508 if (IS_ERR(img)) 509 return PTR_ERR(img); 510 511 dma_addr = dma_map_single(nn->dp.dev, img, 512 nfp_prog->prog_len * sizeof(u64), 513 DMA_TO_DEVICE); 514 if (dma_mapping_error(nn->dp.dev, dma_addr)) { 515 kfree(img); 516 return -ENOMEM; 517 } 518 519 nn_writew(nn, NFP_NET_CFG_BPF_SIZE, nfp_prog->prog_len); 520 nn_writeq(nn, NFP_NET_CFG_BPF_ADDR, dma_addr); 521 522 /* Load up the JITed code */ 523 err = nfp_net_reconfig(nn, NFP_NET_CFG_UPDATE_BPF); 524 if (err) 525 NL_SET_ERR_MSG_MOD(extack, 526 "FW command error while loading BPF"); 527 528 dma_unmap_single(nn->dp.dev, dma_addr, nfp_prog->prog_len * sizeof(u64), 529 DMA_TO_DEVICE); 530 kfree(img); 531 532 return err; 533 } 534 535 static void 536 nfp_net_bpf_start(struct nfp_net *nn, struct netlink_ext_ack *extack) 537 { 538 int err; 539 540 /* Enable passing packets through BPF function */ 541 nn->dp.ctrl |= NFP_NET_CFG_CTRL_BPF; 542 nn_writel(nn, NFP_NET_CFG_CTRL, nn->dp.ctrl); 543 err = nfp_net_reconfig(nn, NFP_NET_CFG_UPDATE_GEN); 544 if (err) 545 NL_SET_ERR_MSG_MOD(extack, 546 "FW command error while enabling BPF"); 547 } 548 549 static int nfp_net_bpf_stop(struct nfp_net *nn) 550 { 551 if (!(nn->dp.ctrl & NFP_NET_CFG_CTRL_BPF)) 552 return 0; 553 554 nn->dp.ctrl &= ~NFP_NET_CFG_CTRL_BPF; 555 nn_writel(nn, NFP_NET_CFG_CTRL, nn->dp.ctrl); 556 557 return nfp_net_reconfig(nn, NFP_NET_CFG_UPDATE_GEN); 558 } 559 560 int nfp_net_bpf_offload(struct nfp_net *nn, struct bpf_prog *prog, 561 bool old_prog, struct netlink_ext_ack *extack) 562 { 563 int err; 564 565 if (prog && !bpf_offload_dev_match(prog, nn->dp.netdev)) 566 return -EINVAL; 567 568 if (prog && old_prog) { 569 u8 cap; 570 571 cap = nn_readb(nn, NFP_NET_CFG_BPF_CAP); 572 if (!(cap & NFP_NET_BPF_CAP_RELO)) { 573 NL_SET_ERR_MSG_MOD(extack, 574 "FW does not support live reload"); 575 return -EBUSY; 576 } 577 } 578 579 /* Something else is loaded, different program type? */ 580 if (!old_prog && nn->dp.ctrl & NFP_NET_CFG_CTRL_BPF) 581 return -EBUSY; 582 583 if (old_prog && !prog) 584 return nfp_net_bpf_stop(nn); 585 586 err = nfp_net_bpf_load(nn, prog, extack); 587 if (err) 588 return err; 589 590 if (!old_prog) 591 nfp_net_bpf_start(nn, extack); 592 593 return 0; 594 } 595 596 const struct bpf_prog_offload_ops nfp_bpf_dev_ops = { 597 .insn_hook = nfp_verify_insn, 598 .finalize = nfp_bpf_finalize, 599 .replace_insn = nfp_bpf_opt_replace_insn, 600 .remove_insns = nfp_bpf_opt_remove_insns, 601 .prepare = nfp_bpf_verifier_prep, 602 .translate = nfp_bpf_translate, 603 .destroy = nfp_bpf_destroy, 604 }; 605