1 /* SPDX-License-Identifier: BSD-3-Clause 2 * Copyright 2016 6WIND S.A. 3 * Copyright 2020 Mellanox Technologies, Ltd 4 */ 5 #include <rte_eal_memconfig.h> 6 #include <rte_errno.h> 7 #include <rte_mempool.h> 8 #include <rte_malloc.h> 9 #include <rte_rwlock.h> 10 11 #include "mlx5_glue.h" 12 #include "mlx5_common_mp.h" 13 #include "mlx5_common_mr.h" 14 #include "mlx5_common_utils.h" 15 16 struct mr_find_contig_memsegs_data { 17 uintptr_t addr; 18 uintptr_t start; 19 uintptr_t end; 20 const struct rte_memseg_list *msl; 21 }; 22 23 /** 24 * Expand B-tree table to a given size. Can't be called with holding 25 * memory_hotplug_lock or share_cache.rwlock due to rte_realloc(). 26 * 27 * @param bt 28 * Pointer to B-tree structure. 29 * @param n 30 * Number of entries for expansion. 31 * 32 * @return 33 * 0 on success, -1 on failure. 34 */ 35 static int 36 mr_btree_expand(struct mlx5_mr_btree *bt, int n) 37 { 38 void *mem; 39 int ret = 0; 40 41 if (n <= bt->size) 42 return ret; 43 /* 44 * Downside of directly using rte_realloc() is that SOCKET_ID_ANY is 45 * used inside if there's no room to expand. Because this is a quite 46 * rare case and a part of very slow path, it is very acceptable. 47 * Initially cache_bh[] will be given practically enough space and once 48 * it is expanded, expansion wouldn't be needed again ever. 49 */ 50 mem = rte_realloc(bt->table, n * sizeof(struct mr_cache_entry), 0); 51 if (mem == NULL) { 52 /* Not an error, B-tree search will be skipped. */ 53 DRV_LOG(WARNING, "failed to expand MR B-tree (%p) table", 54 (void *)bt); 55 ret = -1; 56 } else { 57 DRV_LOG(DEBUG, "expanded MR B-tree table (size=%u)", n); 58 bt->table = mem; 59 bt->size = n; 60 } 61 return ret; 62 } 63 64 /** 65 * Look up LKey from given B-tree lookup table, store the last index and return 66 * searched LKey. 67 * 68 * @param bt 69 * Pointer to B-tree structure. 70 * @param[out] idx 71 * Pointer to index. Even on search failure, returns index where it stops 72 * searching so that index can be used when inserting a new entry. 73 * @param addr 74 * Search key. 75 * 76 * @return 77 * Searched LKey on success, UINT32_MAX on no match. 78 */ 79 static uint32_t 80 mr_btree_lookup(struct mlx5_mr_btree *bt, uint16_t *idx, uintptr_t addr) 81 { 82 struct mr_cache_entry *lkp_tbl; 83 uint16_t n; 84 uint16_t base = 0; 85 86 MLX5_ASSERT(bt != NULL); 87 lkp_tbl = *bt->table; 88 n = bt->len; 89 /* First entry must be NULL for comparison. */ 90 MLX5_ASSERT(bt->len > 0 || (lkp_tbl[0].start == 0 && 91 lkp_tbl[0].lkey == UINT32_MAX)); 92 /* Binary search. */ 93 do { 94 register uint16_t delta = n >> 1; 95 96 if (addr < lkp_tbl[base + delta].start) { 97 n = delta; 98 } else { 99 base += delta; 100 n -= delta; 101 } 102 } while (n > 1); 103 MLX5_ASSERT(addr >= lkp_tbl[base].start); 104 *idx = base; 105 if (addr < lkp_tbl[base].end) 106 return lkp_tbl[base].lkey; 107 /* Not found. */ 108 return UINT32_MAX; 109 } 110 111 /** 112 * Insert an entry to B-tree lookup table. 113 * 114 * @param bt 115 * Pointer to B-tree structure. 116 * @param entry 117 * Pointer to new entry to insert. 118 * 119 * @return 120 * 0 on success, -1 on failure. 121 */ 122 static int 123 mr_btree_insert(struct mlx5_mr_btree *bt, struct mr_cache_entry *entry) 124 { 125 struct mr_cache_entry *lkp_tbl; 126 uint16_t idx = 0; 127 size_t shift; 128 129 MLX5_ASSERT(bt != NULL); 130 MLX5_ASSERT(bt->len <= bt->size); 131 MLX5_ASSERT(bt->len > 0); 132 lkp_tbl = *bt->table; 133 /* Find out the slot for insertion. */ 134 if (mr_btree_lookup(bt, &idx, entry->start) != UINT32_MAX) { 135 DRV_LOG(DEBUG, 136 "abort insertion to B-tree(%p): already exist at" 137 " idx=%u [0x%" PRIxPTR ", 0x%" PRIxPTR ") lkey=0x%x", 138 (void *)bt, idx, entry->start, entry->end, entry->lkey); 139 /* Already exist, return. */ 140 return 0; 141 } 142 /* If table is full, return error. */ 143 if (unlikely(bt->len == bt->size)) { 144 bt->overflow = 1; 145 return -1; 146 } 147 /* Insert entry. */ 148 ++idx; 149 shift = (bt->len - idx) * sizeof(struct mr_cache_entry); 150 if (shift) 151 memmove(&lkp_tbl[idx + 1], &lkp_tbl[idx], shift); 152 lkp_tbl[idx] = *entry; 153 bt->len++; 154 DRV_LOG(DEBUG, 155 "inserted B-tree(%p)[%u]," 156 " [0x%" PRIxPTR ", 0x%" PRIxPTR ") lkey=0x%x", 157 (void *)bt, idx, entry->start, entry->end, entry->lkey); 158 return 0; 159 } 160 161 /** 162 * Initialize B-tree and allocate memory for lookup table. 163 * 164 * @param bt 165 * Pointer to B-tree structure. 166 * @param n 167 * Number of entries to allocate. 168 * @param socket 169 * NUMA socket on which memory must be allocated. 170 * 171 * @return 172 * 0 on success, a negative errno value otherwise and rte_errno is set. 173 */ 174 int 175 mlx5_mr_btree_init(struct mlx5_mr_btree *bt, int n, int socket) 176 { 177 if (bt == NULL) { 178 rte_errno = EINVAL; 179 return -rte_errno; 180 } 181 MLX5_ASSERT(!bt->table && !bt->size); 182 memset(bt, 0, sizeof(*bt)); 183 bt->table = rte_calloc_socket("B-tree table", 184 n, sizeof(struct mr_cache_entry), 185 0, socket); 186 if (bt->table == NULL) { 187 rte_errno = ENOMEM; 188 DEBUG("failed to allocate memory for btree cache on socket %d", 189 socket); 190 return -rte_errno; 191 } 192 bt->size = n; 193 /* First entry must be NULL for binary search. */ 194 (*bt->table)[bt->len++] = (struct mr_cache_entry) { 195 .lkey = UINT32_MAX, 196 }; 197 DEBUG("initialized B-tree %p with table %p", 198 (void *)bt, (void *)bt->table); 199 return 0; 200 } 201 202 /** 203 * Free B-tree resources. 204 * 205 * @param bt 206 * Pointer to B-tree structure. 207 */ 208 void 209 mlx5_mr_btree_free(struct mlx5_mr_btree *bt) 210 { 211 if (bt == NULL) 212 return; 213 DEBUG("freeing B-tree %p with table %p", 214 (void *)bt, (void *)bt->table); 215 rte_free(bt->table); 216 memset(bt, 0, sizeof(*bt)); 217 } 218 219 /** 220 * Dump all the entries in a B-tree 221 * 222 * @param bt 223 * Pointer to B-tree structure. 224 */ 225 void 226 mlx5_mr_btree_dump(struct mlx5_mr_btree *bt __rte_unused) 227 { 228 #ifdef RTE_LIBRTE_MLX5_DEBUG 229 int idx; 230 struct mr_cache_entry *lkp_tbl; 231 232 if (bt == NULL) 233 return; 234 lkp_tbl = *bt->table; 235 for (idx = 0; idx < bt->len; ++idx) { 236 struct mr_cache_entry *entry = &lkp_tbl[idx]; 237 238 DEBUG("B-tree(%p)[%u]," 239 " [0x%" PRIxPTR ", 0x%" PRIxPTR ") lkey=0x%x", 240 (void *)bt, idx, entry->start, entry->end, entry->lkey); 241 } 242 #endif 243 } 244 245 /** 246 * Find virtually contiguous memory chunk in a given MR. 247 * 248 * @param dev 249 * Pointer to MR structure. 250 * @param[out] entry 251 * Pointer to returning MR cache entry. If not found, this will not be 252 * updated. 253 * @param start_idx 254 * Start index of the memseg bitmap. 255 * 256 * @return 257 * Next index to go on lookup. 258 */ 259 static int 260 mr_find_next_chunk(struct mlx5_mr *mr, struct mr_cache_entry *entry, 261 int base_idx) 262 { 263 uintptr_t start = 0; 264 uintptr_t end = 0; 265 uint32_t idx = 0; 266 267 /* MR for external memory doesn't have memseg list. */ 268 if (mr->msl == NULL) { 269 struct ibv_mr *ibv_mr = mr->ibv_mr; 270 271 MLX5_ASSERT(mr->ms_bmp_n == 1); 272 MLX5_ASSERT(mr->ms_n == 1); 273 MLX5_ASSERT(base_idx == 0); 274 /* 275 * Can't search it from memseg list but get it directly from 276 * verbs MR as there's only one chunk. 277 */ 278 entry->start = (uintptr_t)ibv_mr->addr; 279 entry->end = (uintptr_t)ibv_mr->addr + mr->ibv_mr->length; 280 entry->lkey = rte_cpu_to_be_32(mr->ibv_mr->lkey); 281 /* Returning 1 ends iteration. */ 282 return 1; 283 } 284 for (idx = base_idx; idx < mr->ms_bmp_n; ++idx) { 285 if (rte_bitmap_get(mr->ms_bmp, idx)) { 286 const struct rte_memseg_list *msl; 287 const struct rte_memseg *ms; 288 289 msl = mr->msl; 290 ms = rte_fbarray_get(&msl->memseg_arr, 291 mr->ms_base_idx + idx); 292 MLX5_ASSERT(msl->page_sz == ms->hugepage_sz); 293 if (!start) 294 start = ms->addr_64; 295 end = ms->addr_64 + ms->hugepage_sz; 296 } else if (start) { 297 /* Passed the end of a fragment. */ 298 break; 299 } 300 } 301 if (start) { 302 /* Found one chunk. */ 303 entry->start = start; 304 entry->end = end; 305 entry->lkey = rte_cpu_to_be_32(mr->ibv_mr->lkey); 306 } 307 return idx; 308 } 309 310 /** 311 * Insert a MR to the global B-tree cache. It may fail due to low-on-memory. 312 * Then, this entry will have to be searched by mr_lookup_list() in 313 * mlx5_mr_create() on miss. 314 * 315 * @param share_cache 316 * Pointer to a global shared MR cache. 317 * @param mr 318 * Pointer to MR to insert. 319 * 320 * @return 321 * 0 on success, -1 on failure. 322 */ 323 int 324 mlx5_mr_insert_cache(struct mlx5_mr_share_cache *share_cache, 325 struct mlx5_mr *mr) 326 { 327 unsigned int n; 328 329 DRV_LOG(DEBUG, "Inserting MR(%p) to global cache(%p)", 330 (void *)mr, (void *)share_cache); 331 for (n = 0; n < mr->ms_bmp_n; ) { 332 struct mr_cache_entry entry; 333 334 memset(&entry, 0, sizeof(entry)); 335 /* Find a contiguous chunk and advance the index. */ 336 n = mr_find_next_chunk(mr, &entry, n); 337 if (!entry.end) 338 break; 339 if (mr_btree_insert(&share_cache->cache, &entry) < 0) { 340 /* 341 * Overflowed, but the global table cannot be expanded 342 * because of deadlock. 343 */ 344 return -1; 345 } 346 } 347 return 0; 348 } 349 350 /** 351 * Look up address in the original global MR list. 352 * 353 * @param share_cache 354 * Pointer to a global shared MR cache. 355 * @param[out] entry 356 * Pointer to returning MR cache entry. If no match, this will not be updated. 357 * @param addr 358 * Search key. 359 * 360 * @return 361 * Found MR on match, NULL otherwise. 362 */ 363 struct mlx5_mr * 364 mlx5_mr_lookup_list(struct mlx5_mr_share_cache *share_cache, 365 struct mr_cache_entry *entry, uintptr_t addr) 366 { 367 struct mlx5_mr *mr; 368 369 /* Iterate all the existing MRs. */ 370 LIST_FOREACH(mr, &share_cache->mr_list, mr) { 371 unsigned int n; 372 373 if (mr->ms_n == 0) 374 continue; 375 for (n = 0; n < mr->ms_bmp_n; ) { 376 struct mr_cache_entry ret; 377 378 memset(&ret, 0, sizeof(ret)); 379 n = mr_find_next_chunk(mr, &ret, n); 380 if (addr >= ret.start && addr < ret.end) { 381 /* Found. */ 382 *entry = ret; 383 return mr; 384 } 385 } 386 } 387 return NULL; 388 } 389 390 /** 391 * Look up address on global MR cache. 392 * 393 * @param share_cache 394 * Pointer to a global shared MR cache. 395 * @param[out] entry 396 * Pointer to returning MR cache entry. If no match, this will not be updated. 397 * @param addr 398 * Search key. 399 * 400 * @return 401 * Searched LKey on success, UINT32_MAX on failure and rte_errno is set. 402 */ 403 uint32_t 404 mlx5_mr_lookup_cache(struct mlx5_mr_share_cache *share_cache, 405 struct mr_cache_entry *entry, uintptr_t addr) 406 { 407 uint16_t idx; 408 uint32_t lkey = UINT32_MAX; 409 struct mlx5_mr *mr; 410 411 /* 412 * If the global cache has overflowed since it failed to expand the 413 * B-tree table, it can't have all the existing MRs. Then, the address 414 * has to be searched by traversing the original MR list instead, which 415 * is very slow path. Otherwise, the global cache is all inclusive. 416 */ 417 if (!unlikely(share_cache->cache.overflow)) { 418 lkey = mr_btree_lookup(&share_cache->cache, &idx, addr); 419 if (lkey != UINT32_MAX) 420 *entry = (*share_cache->cache.table)[idx]; 421 } else { 422 /* Falling back to the slowest path. */ 423 mr = mlx5_mr_lookup_list(share_cache, entry, addr); 424 if (mr != NULL) 425 lkey = entry->lkey; 426 } 427 MLX5_ASSERT(lkey == UINT32_MAX || (addr >= entry->start && 428 addr < entry->end)); 429 return lkey; 430 } 431 432 /** 433 * Free MR resources. MR lock must not be held to avoid a deadlock. rte_free() 434 * can raise memory free event and the callback function will spin on the lock. 435 * 436 * @param mr 437 * Pointer to MR to free. 438 */ 439 static void 440 mr_free(struct mlx5_mr *mr) 441 { 442 if (mr == NULL) 443 return; 444 DRV_LOG(DEBUG, "freeing MR(%p):", (void *)mr); 445 if (mr->ibv_mr != NULL) 446 claim_zero(mlx5_glue->dereg_mr(mr->ibv_mr)); 447 if (mr->ms_bmp != NULL) 448 rte_bitmap_free(mr->ms_bmp); 449 rte_free(mr); 450 } 451 452 void 453 mlx5_mr_rebuild_cache(struct mlx5_mr_share_cache *share_cache) 454 { 455 struct mlx5_mr *mr; 456 457 DRV_LOG(DEBUG, "Rebuild dev cache[] %p", (void *)share_cache); 458 /* Flush cache to rebuild. */ 459 share_cache->cache.len = 1; 460 share_cache->cache.overflow = 0; 461 /* Iterate all the existing MRs. */ 462 LIST_FOREACH(mr, &share_cache->mr_list, mr) 463 if (mlx5_mr_insert_cache(share_cache, mr) < 0) 464 return; 465 } 466 467 /** 468 * Release resources of detached MR having no online entry. 469 * 470 * @param share_cache 471 * Pointer to a global shared MR cache. 472 */ 473 static void 474 mlx5_mr_garbage_collect(struct mlx5_mr_share_cache *share_cache) 475 { 476 struct mlx5_mr *mr_next; 477 struct mlx5_mr_list free_list = LIST_HEAD_INITIALIZER(free_list); 478 479 /* Must be called from the primary process. */ 480 MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY); 481 /* 482 * MR can't be freed with holding the lock because rte_free() could call 483 * memory free callback function. This will be a deadlock situation. 484 */ 485 rte_rwlock_write_lock(&share_cache->rwlock); 486 /* Detach the whole free list and release it after unlocking. */ 487 free_list = share_cache->mr_free_list; 488 LIST_INIT(&share_cache->mr_free_list); 489 rte_rwlock_write_unlock(&share_cache->rwlock); 490 /* Release resources. */ 491 mr_next = LIST_FIRST(&free_list); 492 while (mr_next != NULL) { 493 struct mlx5_mr *mr = mr_next; 494 495 mr_next = LIST_NEXT(mr, mr); 496 mr_free(mr); 497 } 498 } 499 500 /* Called during rte_memseg_contig_walk() by mlx5_mr_create(). */ 501 static int 502 mr_find_contig_memsegs_cb(const struct rte_memseg_list *msl, 503 const struct rte_memseg *ms, size_t len, void *arg) 504 { 505 struct mr_find_contig_memsegs_data *data = arg; 506 507 if (data->addr < ms->addr_64 || data->addr >= ms->addr_64 + len) 508 return 0; 509 /* Found, save it and stop walking. */ 510 data->start = ms->addr_64; 511 data->end = ms->addr_64 + len; 512 data->msl = msl; 513 return 1; 514 } 515 516 /** 517 * Create a new global Memory Region (MR) for a missing virtual address. 518 * This API should be called on a secondary process, then a request is sent to 519 * the primary process in order to create a MR for the address. As the global MR 520 * list is on the shared memory, following LKey lookup should succeed unless the 521 * request fails. 522 * 523 * @param pd 524 * Pointer to ibv_pd of a device (net, regex, vdpa,...). 525 * @param share_cache 526 * Pointer to a global shared MR cache. 527 * @param[out] entry 528 * Pointer to returning MR cache entry, found in the global cache or newly 529 * created. If failed to create one, this will not be updated. 530 * @param addr 531 * Target virtual address to register. 532 * @param mr_ext_memseg_en 533 * Configurable flag about external memory segment enable or not. 534 * 535 * @return 536 * Searched LKey on success, UINT32_MAX on failure and rte_errno is set. 537 */ 538 static uint32_t 539 mlx5_mr_create_secondary(struct ibv_pd *pd __rte_unused, 540 struct mlx5_mp_id *mp_id, 541 struct mlx5_mr_share_cache *share_cache, 542 struct mr_cache_entry *entry, uintptr_t addr, 543 unsigned int mr_ext_memseg_en __rte_unused) 544 { 545 int ret; 546 547 DEBUG("port %u requesting MR creation for address (%p)", 548 mp_id->port_id, (void *)addr); 549 ret = mlx5_mp_req_mr_create(mp_id, addr); 550 if (ret) { 551 DEBUG("Fail to request MR creation for address (%p)", 552 (void *)addr); 553 return UINT32_MAX; 554 } 555 rte_rwlock_read_lock(&share_cache->rwlock); 556 /* Fill in output data. */ 557 mlx5_mr_lookup_cache(share_cache, entry, addr); 558 /* Lookup can't fail. */ 559 MLX5_ASSERT(entry->lkey != UINT32_MAX); 560 rte_rwlock_read_unlock(&share_cache->rwlock); 561 DEBUG("MR CREATED by primary process for %p:\n" 562 " [0x%" PRIxPTR ", 0x%" PRIxPTR "), lkey=0x%x", 563 (void *)addr, entry->start, entry->end, entry->lkey); 564 return entry->lkey; 565 } 566 567 /** 568 * Create a new global Memory Region (MR) for a missing virtual address. 569 * Register entire virtually contiguous memory chunk around the address. 570 * 571 * @param pd 572 * Pointer to ibv_pd of a device (net, regex, vdpa,...). 573 * @param share_cache 574 * Pointer to a global shared MR cache. 575 * @param[out] entry 576 * Pointer to returning MR cache entry, found in the global cache or newly 577 * created. If failed to create one, this will not be updated. 578 * @param addr 579 * Target virtual address to register. 580 * @param mr_ext_memseg_en 581 * Configurable flag about external memory segment enable or not. 582 * 583 * @return 584 * Searched LKey on success, UINT32_MAX on failure and rte_errno is set. 585 */ 586 uint32_t 587 mlx5_mr_create_primary(struct ibv_pd *pd, 588 struct mlx5_mr_share_cache *share_cache, 589 struct mr_cache_entry *entry, uintptr_t addr, 590 unsigned int mr_ext_memseg_en) 591 { 592 struct mr_find_contig_memsegs_data data = {.addr = addr, }; 593 struct mr_find_contig_memsegs_data data_re; 594 const struct rte_memseg_list *msl; 595 const struct rte_memseg *ms; 596 struct mlx5_mr *mr = NULL; 597 int ms_idx_shift = -1; 598 uint32_t bmp_size; 599 void *bmp_mem; 600 uint32_t ms_n; 601 uint32_t n; 602 size_t len; 603 604 DRV_LOG(DEBUG, "Creating a MR using address (%p)", (void *)addr); 605 /* 606 * Release detached MRs if any. This can't be called with holding either 607 * memory_hotplug_lock or share_cache->rwlock. MRs on the free list have 608 * been detached by the memory free event but it couldn't be released 609 * inside the callback due to deadlock. As a result, releasing resources 610 * is quite opportunistic. 611 */ 612 mlx5_mr_garbage_collect(share_cache); 613 /* 614 * If enabled, find out a contiguous virtual address chunk in use, to 615 * which the given address belongs, in order to register maximum range. 616 * In the best case where mempools are not dynamically recreated and 617 * '--socket-mem' is specified as an EAL option, it is very likely to 618 * have only one MR(LKey) per a socket and per a hugepage-size even 619 * though the system memory is highly fragmented. As the whole memory 620 * chunk will be pinned by kernel, it can't be reused unless entire 621 * chunk is freed from EAL. 622 * 623 * If disabled, just register one memseg (page). Then, memory 624 * consumption will be minimized but it may drop performance if there 625 * are many MRs to lookup on the datapath. 626 */ 627 if (!mr_ext_memseg_en) { 628 data.msl = rte_mem_virt2memseg_list((void *)addr); 629 data.start = RTE_ALIGN_FLOOR(addr, data.msl->page_sz); 630 data.end = data.start + data.msl->page_sz; 631 } else if (!rte_memseg_contig_walk(mr_find_contig_memsegs_cb, &data)) { 632 DRV_LOG(WARNING, 633 "Unable to find virtually contiguous" 634 " chunk for address (%p)." 635 " rte_memseg_contig_walk() failed.", (void *)addr); 636 rte_errno = ENXIO; 637 goto err_nolock; 638 } 639 alloc_resources: 640 /* Addresses must be page-aligned. */ 641 MLX5_ASSERT(data.msl); 642 MLX5_ASSERT(rte_is_aligned((void *)data.start, data.msl->page_sz)); 643 MLX5_ASSERT(rte_is_aligned((void *)data.end, data.msl->page_sz)); 644 msl = data.msl; 645 ms = rte_mem_virt2memseg((void *)data.start, msl); 646 len = data.end - data.start; 647 MLX5_ASSERT(ms); 648 MLX5_ASSERT(msl->page_sz == ms->hugepage_sz); 649 /* Number of memsegs in the range. */ 650 ms_n = len / msl->page_sz; 651 DEBUG("Extending %p to [0x%" PRIxPTR ", 0x%" PRIxPTR ")," 652 " page_sz=0x%" PRIx64 ", ms_n=%u", 653 (void *)addr, data.start, data.end, msl->page_sz, ms_n); 654 /* Size of memory for bitmap. */ 655 bmp_size = rte_bitmap_get_memory_footprint(ms_n); 656 mr = rte_zmalloc_socket(NULL, 657 RTE_ALIGN_CEIL(sizeof(*mr), 658 RTE_CACHE_LINE_SIZE) + 659 bmp_size, 660 RTE_CACHE_LINE_SIZE, msl->socket_id); 661 if (mr == NULL) { 662 DEBUG("Unable to allocate memory for a new MR of" 663 " address (%p).", (void *)addr); 664 rte_errno = ENOMEM; 665 goto err_nolock; 666 } 667 mr->msl = msl; 668 /* 669 * Save the index of the first memseg and initialize memseg bitmap. To 670 * see if a memseg of ms_idx in the memseg-list is still valid, check: 671 * rte_bitmap_get(mr->bmp, ms_idx - mr->ms_base_idx) 672 */ 673 mr->ms_base_idx = rte_fbarray_find_idx(&msl->memseg_arr, ms); 674 bmp_mem = RTE_PTR_ALIGN_CEIL(mr + 1, RTE_CACHE_LINE_SIZE); 675 mr->ms_bmp = rte_bitmap_init(ms_n, bmp_mem, bmp_size); 676 if (mr->ms_bmp == NULL) { 677 DEBUG("Unable to initialize bitmap for a new MR of" 678 " address (%p).", (void *)addr); 679 rte_errno = EINVAL; 680 goto err_nolock; 681 } 682 /* 683 * Should recheck whether the extended contiguous chunk is still valid. 684 * Because memory_hotplug_lock can't be held if there's any memory 685 * related calls in a critical path, resource allocation above can't be 686 * locked. If the memory has been changed at this point, try again with 687 * just single page. If not, go on with the big chunk atomically from 688 * here. 689 */ 690 rte_mcfg_mem_read_lock(); 691 data_re = data; 692 if (len > msl->page_sz && 693 !rte_memseg_contig_walk(mr_find_contig_memsegs_cb, &data_re)) { 694 DEBUG("Unable to find virtually contiguous" 695 " chunk for address (%p)." 696 " rte_memseg_contig_walk() failed.", (void *)addr); 697 rte_errno = ENXIO; 698 goto err_memlock; 699 } 700 if (data.start != data_re.start || data.end != data_re.end) { 701 /* 702 * The extended contiguous chunk has been changed. Try again 703 * with single memseg instead. 704 */ 705 data.start = RTE_ALIGN_FLOOR(addr, msl->page_sz); 706 data.end = data.start + msl->page_sz; 707 rte_mcfg_mem_read_unlock(); 708 mr_free(mr); 709 goto alloc_resources; 710 } 711 MLX5_ASSERT(data.msl == data_re.msl); 712 rte_rwlock_write_lock(&share_cache->rwlock); 713 /* 714 * Check the address is really missing. If other thread already created 715 * one or it is not found due to overflow, abort and return. 716 */ 717 if (mlx5_mr_lookup_cache(share_cache, entry, addr) != UINT32_MAX) { 718 /* 719 * Insert to the global cache table. It may fail due to 720 * low-on-memory. Then, this entry will have to be searched 721 * here again. 722 */ 723 mr_btree_insert(&share_cache->cache, entry); 724 DEBUG("Found MR for %p on final lookup, abort", (void *)addr); 725 rte_rwlock_write_unlock(&share_cache->rwlock); 726 rte_mcfg_mem_read_unlock(); 727 /* 728 * Must be unlocked before calling rte_free() because 729 * mlx5_mr_mem_event_free_cb() can be called inside. 730 */ 731 mr_free(mr); 732 return entry->lkey; 733 } 734 /* 735 * Trim start and end addresses for verbs MR. Set bits for registering 736 * memsegs but exclude already registered ones. Bitmap can be 737 * fragmented. 738 */ 739 for (n = 0; n < ms_n; ++n) { 740 uintptr_t start; 741 struct mr_cache_entry ret; 742 743 memset(&ret, 0, sizeof(ret)); 744 start = data_re.start + n * msl->page_sz; 745 /* Exclude memsegs already registered by other MRs. */ 746 if (mlx5_mr_lookup_cache(share_cache, &ret, start) == 747 UINT32_MAX) { 748 /* 749 * Start from the first unregistered memseg in the 750 * extended range. 751 */ 752 if (ms_idx_shift == -1) { 753 mr->ms_base_idx += n; 754 data.start = start; 755 ms_idx_shift = n; 756 } 757 data.end = start + msl->page_sz; 758 rte_bitmap_set(mr->ms_bmp, n - ms_idx_shift); 759 ++mr->ms_n; 760 } 761 } 762 len = data.end - data.start; 763 mr->ms_bmp_n = len / msl->page_sz; 764 MLX5_ASSERT(ms_idx_shift + mr->ms_bmp_n <= ms_n); 765 /* 766 * Finally create a verbs MR for the memory chunk. ibv_reg_mr() can be 767 * called with holding the memory lock because it doesn't use 768 * mlx5_alloc_buf_extern() which eventually calls rte_malloc_socket() 769 * through mlx5_alloc_verbs_buf(). 770 */ 771 mr->ibv_mr = mlx5_glue->reg_mr(pd, (void *)data.start, len, 772 IBV_ACCESS_LOCAL_WRITE | 773 (haswell_broadwell_cpu ? 0 : 774 IBV_ACCESS_RELAXED_ORDERING)); 775 if (mr->ibv_mr == NULL) { 776 DEBUG("Fail to create a verbs MR for address (%p)", 777 (void *)addr); 778 rte_errno = EINVAL; 779 goto err_mrlock; 780 } 781 MLX5_ASSERT((uintptr_t)mr->ibv_mr->addr == data.start); 782 MLX5_ASSERT(mr->ibv_mr->length == len); 783 LIST_INSERT_HEAD(&share_cache->mr_list, mr, mr); 784 DEBUG("MR CREATED (%p) for %p:\n" 785 " [0x%" PRIxPTR ", 0x%" PRIxPTR ")," 786 " lkey=0x%x base_idx=%u ms_n=%u, ms_bmp_n=%u", 787 (void *)mr, (void *)addr, data.start, data.end, 788 rte_cpu_to_be_32(mr->ibv_mr->lkey), 789 mr->ms_base_idx, mr->ms_n, mr->ms_bmp_n); 790 /* Insert to the global cache table. */ 791 mlx5_mr_insert_cache(share_cache, mr); 792 /* Fill in output data. */ 793 mlx5_mr_lookup_cache(share_cache, entry, addr); 794 /* Lookup can't fail. */ 795 MLX5_ASSERT(entry->lkey != UINT32_MAX); 796 rte_rwlock_write_unlock(&share_cache->rwlock); 797 rte_mcfg_mem_read_unlock(); 798 return entry->lkey; 799 err_mrlock: 800 rte_rwlock_write_unlock(&share_cache->rwlock); 801 err_memlock: 802 rte_mcfg_mem_read_unlock(); 803 err_nolock: 804 /* 805 * In case of error, as this can be called in a datapath, a warning 806 * message per an error is preferable instead. Must be unlocked before 807 * calling rte_free() because mlx5_mr_mem_event_free_cb() can be called 808 * inside. 809 */ 810 mr_free(mr); 811 return UINT32_MAX; 812 } 813 814 /** 815 * Create a new global Memory Region (MR) for a missing virtual address. 816 * This can be called from primary and secondary process. 817 * 818 * @param pd 819 * Pointer to ibv_pd of a device (net, regex, vdpa,...). 820 * @param share_cache 821 * Pointer to a global shared MR cache. 822 * @param[out] entry 823 * Pointer to returning MR cache entry, found in the global cache or newly 824 * created. If failed to create one, this will not be updated. 825 * @param addr 826 * Target virtual address to register. 827 * 828 * @return 829 * Searched LKey on success, UINT32_MAX on failure and rte_errno is set. 830 */ 831 static uint32_t 832 mlx5_mr_create(struct ibv_pd *pd, struct mlx5_mp_id *mp_id, 833 struct mlx5_mr_share_cache *share_cache, 834 struct mr_cache_entry *entry, uintptr_t addr, 835 unsigned int mr_ext_memseg_en) 836 { 837 uint32_t ret = 0; 838 839 switch (rte_eal_process_type()) { 840 case RTE_PROC_PRIMARY: 841 ret = mlx5_mr_create_primary(pd, share_cache, entry, 842 addr, mr_ext_memseg_en); 843 break; 844 case RTE_PROC_SECONDARY: 845 ret = mlx5_mr_create_secondary(pd, mp_id, share_cache, entry, 846 addr, mr_ext_memseg_en); 847 break; 848 default: 849 break; 850 } 851 return ret; 852 } 853 854 /** 855 * Look up address in the global MR cache table. If not found, create a new MR. 856 * Insert the found/created entry to local bottom-half cache table. 857 * 858 * @param pd 859 * Pointer to ibv_pd of a device (net, regex, vdpa,...). 860 * @param share_cache 861 * Pointer to a global shared MR cache. 862 * @param mr_ctrl 863 * Pointer to per-queue MR control structure. 864 * @param[out] entry 865 * Pointer to returning MR cache entry, found in the global cache or newly 866 * created. If failed to create one, this is not written. 867 * @param addr 868 * Search key. 869 * 870 * @return 871 * Searched LKey on success, UINT32_MAX on no match. 872 */ 873 static uint32_t 874 mr_lookup_caches(struct ibv_pd *pd, struct mlx5_mp_id *mp_id, 875 struct mlx5_mr_share_cache *share_cache, 876 struct mlx5_mr_ctrl *mr_ctrl, 877 struct mr_cache_entry *entry, uintptr_t addr, 878 unsigned int mr_ext_memseg_en) 879 { 880 struct mlx5_mr_btree *bt = &mr_ctrl->cache_bh; 881 uint32_t lkey; 882 uint16_t idx; 883 884 /* If local cache table is full, try to double it. */ 885 if (unlikely(bt->len == bt->size)) 886 mr_btree_expand(bt, bt->size << 1); 887 /* Look up in the global cache. */ 888 rte_rwlock_read_lock(&share_cache->rwlock); 889 lkey = mr_btree_lookup(&share_cache->cache, &idx, addr); 890 if (lkey != UINT32_MAX) { 891 /* Found. */ 892 *entry = (*share_cache->cache.table)[idx]; 893 rte_rwlock_read_unlock(&share_cache->rwlock); 894 /* 895 * Update local cache. Even if it fails, return the found entry 896 * to update top-half cache. Next time, this entry will be found 897 * in the global cache. 898 */ 899 mr_btree_insert(bt, entry); 900 return lkey; 901 } 902 rte_rwlock_read_unlock(&share_cache->rwlock); 903 /* First time to see the address? Create a new MR. */ 904 lkey = mlx5_mr_create(pd, mp_id, share_cache, entry, addr, 905 mr_ext_memseg_en); 906 /* 907 * Update the local cache if successfully created a new global MR. Even 908 * if failed to create one, there's no action to take in this datapath 909 * code. As returning LKey is invalid, this will eventually make HW 910 * fail. 911 */ 912 if (lkey != UINT32_MAX) 913 mr_btree_insert(bt, entry); 914 return lkey; 915 } 916 917 /** 918 * Bottom-half of LKey search on datapath. First search in cache_bh[] and if 919 * misses, search in the global MR cache table and update the new entry to 920 * per-queue local caches. 921 * 922 * @param pd 923 * Pointer to ibv_pd of a device (net, regex, vdpa,...). 924 * @param share_cache 925 * Pointer to a global shared MR cache. 926 * @param mr_ctrl 927 * Pointer to per-queue MR control structure. 928 * @param addr 929 * Search key. 930 * 931 * @return 932 * Searched LKey on success, UINT32_MAX on no match. 933 */ 934 uint32_t mlx5_mr_addr2mr_bh(struct ibv_pd *pd, struct mlx5_mp_id *mp_id, 935 struct mlx5_mr_share_cache *share_cache, 936 struct mlx5_mr_ctrl *mr_ctrl, 937 uintptr_t addr, unsigned int mr_ext_memseg_en) 938 { 939 uint32_t lkey; 940 uint16_t bh_idx = 0; 941 /* Victim in top-half cache to replace with new entry. */ 942 struct mr_cache_entry *repl = &mr_ctrl->cache[mr_ctrl->head]; 943 944 /* Binary-search MR translation table. */ 945 lkey = mr_btree_lookup(&mr_ctrl->cache_bh, &bh_idx, addr); 946 /* Update top-half cache. */ 947 if (likely(lkey != UINT32_MAX)) { 948 *repl = (*mr_ctrl->cache_bh.table)[bh_idx]; 949 } else { 950 /* 951 * If missed in local lookup table, search in the global cache 952 * and local cache_bh[] will be updated inside if possible. 953 * Top-half cache entry will also be updated. 954 */ 955 lkey = mr_lookup_caches(pd, mp_id, share_cache, mr_ctrl, 956 repl, addr, mr_ext_memseg_en); 957 if (unlikely(lkey == UINT32_MAX)) 958 return UINT32_MAX; 959 } 960 /* Update the most recently used entry. */ 961 mr_ctrl->mru = mr_ctrl->head; 962 /* Point to the next victim, the oldest. */ 963 mr_ctrl->head = (mr_ctrl->head + 1) % MLX5_MR_CACHE_N; 964 return lkey; 965 } 966 967 /** 968 * Release all the created MRs and resources on global MR cache of a device. 969 * list. 970 * 971 * @param share_cache 972 * Pointer to a global shared MR cache. 973 */ 974 void 975 mlx5_mr_release_cache(struct mlx5_mr_share_cache *share_cache) 976 { 977 struct mlx5_mr *mr_next; 978 979 rte_rwlock_write_lock(&share_cache->rwlock); 980 /* Detach from MR list and move to free list. */ 981 mr_next = LIST_FIRST(&share_cache->mr_list); 982 while (mr_next != NULL) { 983 struct mlx5_mr *mr = mr_next; 984 985 mr_next = LIST_NEXT(mr, mr); 986 LIST_REMOVE(mr, mr); 987 LIST_INSERT_HEAD(&share_cache->mr_free_list, mr, mr); 988 } 989 LIST_INIT(&share_cache->mr_list); 990 /* Free global cache. */ 991 mlx5_mr_btree_free(&share_cache->cache); 992 rte_rwlock_write_unlock(&share_cache->rwlock); 993 /* Free all remaining MRs. */ 994 mlx5_mr_garbage_collect(share_cache); 995 } 996 997 /** 998 * Flush all of the local cache entries. 999 * 1000 * @param mr_ctrl 1001 * Pointer to per-queue MR local cache. 1002 */ 1003 void 1004 mlx5_mr_flush_local_cache(struct mlx5_mr_ctrl *mr_ctrl) 1005 { 1006 /* Reset the most-recently-used index. */ 1007 mr_ctrl->mru = 0; 1008 /* Reset the linear search array. */ 1009 mr_ctrl->head = 0; 1010 memset(mr_ctrl->cache, 0, sizeof(mr_ctrl->cache)); 1011 /* Reset the B-tree table. */ 1012 mr_ctrl->cache_bh.len = 1; 1013 mr_ctrl->cache_bh.overflow = 0; 1014 /* Update the generation number. */ 1015 mr_ctrl->cur_gen = *mr_ctrl->dev_gen_ptr; 1016 DRV_LOG(DEBUG, "mr_ctrl(%p): flushed, cur_gen=%d", 1017 (void *)mr_ctrl, mr_ctrl->cur_gen); 1018 } 1019 1020 /** 1021 * Creates a memory region for external memory, that is memory which is not 1022 * part of the DPDK memory segments. 1023 * 1024 * @param pd 1025 * Pointer to ibv_pd of a device (net, regex, vdpa,...). 1026 * @param addr 1027 * Starting virtual address of memory. 1028 * @param len 1029 * Length of memory segment being mapped. 1030 * @param socked_id 1031 * Socket to allocate heap memory for the control structures. 1032 * 1033 * @return 1034 * Pointer to MR structure on success, NULL otherwise. 1035 */ 1036 struct mlx5_mr * 1037 mlx5_create_mr_ext(struct ibv_pd *pd, uintptr_t addr, size_t len, int socket_id) 1038 { 1039 struct mlx5_mr *mr = NULL; 1040 1041 mr = rte_zmalloc_socket(NULL, 1042 RTE_ALIGN_CEIL(sizeof(*mr), 1043 RTE_CACHE_LINE_SIZE), 1044 RTE_CACHE_LINE_SIZE, socket_id); 1045 if (mr == NULL) 1046 return NULL; 1047 mr->ibv_mr = mlx5_glue->reg_mr(pd, (void *)addr, len, 1048 IBV_ACCESS_LOCAL_WRITE | 1049 (haswell_broadwell_cpu ? 0 : 1050 IBV_ACCESS_RELAXED_ORDERING)); 1051 if (mr->ibv_mr == NULL) { 1052 DRV_LOG(WARNING, 1053 "Fail to create a verbs MR for address (%p)", 1054 (void *)addr); 1055 rte_free(mr); 1056 return NULL; 1057 } 1058 mr->msl = NULL; /* Mark it is external memory. */ 1059 mr->ms_bmp = NULL; 1060 mr->ms_n = 1; 1061 mr->ms_bmp_n = 1; 1062 DRV_LOG(DEBUG, 1063 "MR CREATED (%p) for external memory %p:\n" 1064 " [0x%" PRIxPTR ", 0x%" PRIxPTR ")," 1065 " lkey=0x%x base_idx=%u ms_n=%u, ms_bmp_n=%u", 1066 (void *)mr, (void *)addr, 1067 addr, addr + len, rte_cpu_to_be_32(mr->ibv_mr->lkey), 1068 mr->ms_base_idx, mr->ms_n, mr->ms_bmp_n); 1069 return mr; 1070 } 1071 1072 /** 1073 * Dump all the created MRs and the global cache entries. 1074 * 1075 * @param sh 1076 * Pointer to Ethernet device shared context. 1077 */ 1078 void 1079 mlx5_mr_dump_cache(struct mlx5_mr_share_cache *share_cache __rte_unused) 1080 { 1081 #ifdef RTE_LIBRTE_MLX5_DEBUG 1082 struct mlx5_mr *mr; 1083 int mr_n = 0; 1084 int chunk_n = 0; 1085 1086 rte_rwlock_read_lock(&share_cache->rwlock); 1087 /* Iterate all the existing MRs. */ 1088 LIST_FOREACH(mr, &share_cache->mr_list, mr) { 1089 unsigned int n; 1090 1091 DEBUG("MR[%u], LKey = 0x%x, ms_n = %u, ms_bmp_n = %u", 1092 mr_n++, rte_cpu_to_be_32(mr->ibv_mr->lkey), 1093 mr->ms_n, mr->ms_bmp_n); 1094 if (mr->ms_n == 0) 1095 continue; 1096 for (n = 0; n < mr->ms_bmp_n; ) { 1097 struct mr_cache_entry ret = { 0, }; 1098 1099 n = mr_find_next_chunk(mr, &ret, n); 1100 if (!ret.end) 1101 break; 1102 DEBUG(" chunk[%u], [0x%" PRIxPTR ", 0x%" PRIxPTR ")", 1103 chunk_n++, ret.start, ret.end); 1104 } 1105 } 1106 DEBUG("Dumping global cache %p", (void *)share_cache); 1107 mlx5_mr_btree_dump(&share_cache->cache); 1108 rte_rwlock_read_unlock(&share_cache->rwlock); 1109 #endif 1110 } 1111