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