1 // SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB 2 /* Copyright (c) 2020 Mellanox Technologies Ltd. */ 3 4 #include <linux/vhost_types.h> 5 #include <linux/vdpa.h> 6 #include <linux/gcd.h> 7 #include <linux/string.h> 8 #include <linux/mlx5/qp.h> 9 #include "mlx5_vdpa.h" 10 11 /* DIV_ROUND_UP where the divider is a power of 2 give by its log base 2 value */ 12 #define MLX5_DIV_ROUND_UP_POW2(_n, _s) \ 13 ({ \ 14 u64 __s = _s; \ 15 u64 _res; \ 16 _res = (((_n) + (1 << (__s)) - 1) >> (__s)); \ 17 _res; \ 18 }) 19 20 static int get_octo_len(u64 len, int page_shift) 21 { 22 u64 page_size = 1ULL << page_shift; 23 int npages; 24 25 npages = ALIGN(len, page_size) >> page_shift; 26 return (npages + 1) / 2; 27 } 28 29 static void mlx5_set_access_mode(void *mkc, int mode) 30 { 31 MLX5_SET(mkc, mkc, access_mode_1_0, mode & 0x3); 32 MLX5_SET(mkc, mkc, access_mode_4_2, mode >> 2); 33 } 34 35 static void populate_mtts(struct mlx5_vdpa_direct_mr *mr, __be64 *mtt) 36 { 37 struct scatterlist *sg; 38 int nsg = mr->nsg; 39 u64 dma_addr; 40 u64 dma_len; 41 int j = 0; 42 int i; 43 44 for_each_sg(mr->sg_head.sgl, sg, mr->nent, i) { 45 for (dma_addr = sg_dma_address(sg), dma_len = sg_dma_len(sg); 46 nsg && dma_len; 47 nsg--, dma_addr += BIT(mr->log_size), dma_len -= BIT(mr->log_size)) 48 mtt[j++] = cpu_to_be64(dma_addr); 49 } 50 } 51 52 static int create_direct_mr(struct mlx5_vdpa_dev *mvdev, struct mlx5_vdpa_direct_mr *mr) 53 { 54 int inlen; 55 void *mkc; 56 void *in; 57 int err; 58 59 inlen = MLX5_ST_SZ_BYTES(create_mkey_in) + roundup(MLX5_ST_SZ_BYTES(mtt) * mr->nsg, 16); 60 in = kvzalloc(inlen, GFP_KERNEL); 61 if (!in) 62 return -ENOMEM; 63 64 MLX5_SET(create_mkey_in, in, uid, mvdev->res.uid); 65 mkc = MLX5_ADDR_OF(create_mkey_in, in, memory_key_mkey_entry); 66 MLX5_SET(mkc, mkc, lw, !!(mr->perm & VHOST_MAP_WO)); 67 MLX5_SET(mkc, mkc, lr, !!(mr->perm & VHOST_MAP_RO)); 68 mlx5_set_access_mode(mkc, MLX5_MKC_ACCESS_MODE_MTT); 69 MLX5_SET(mkc, mkc, qpn, 0xffffff); 70 MLX5_SET(mkc, mkc, pd, mvdev->res.pdn); 71 MLX5_SET64(mkc, mkc, start_addr, mr->offset); 72 MLX5_SET64(mkc, mkc, len, mr->end - mr->start); 73 MLX5_SET(mkc, mkc, log_page_size, mr->log_size); 74 MLX5_SET(mkc, mkc, translations_octword_size, 75 get_octo_len(mr->end - mr->start, mr->log_size)); 76 MLX5_SET(create_mkey_in, in, translations_octword_actual_size, 77 get_octo_len(mr->end - mr->start, mr->log_size)); 78 populate_mtts(mr, MLX5_ADDR_OF(create_mkey_in, in, klm_pas_mtt)); 79 err = mlx5_vdpa_create_mkey(mvdev, &mr->mr, in, inlen); 80 kvfree(in); 81 if (err) { 82 mlx5_vdpa_warn(mvdev, "Failed to create direct MR\n"); 83 return err; 84 } 85 86 return 0; 87 } 88 89 static void destroy_direct_mr(struct mlx5_vdpa_dev *mvdev, struct mlx5_vdpa_direct_mr *mr) 90 { 91 mlx5_vdpa_destroy_mkey(mvdev, mr->mr); 92 } 93 94 static u64 map_start(struct vhost_iotlb_map *map, struct mlx5_vdpa_direct_mr *mr) 95 { 96 return max_t(u64, map->start, mr->start); 97 } 98 99 static u64 map_end(struct vhost_iotlb_map *map, struct mlx5_vdpa_direct_mr *mr) 100 { 101 return min_t(u64, map->last + 1, mr->end); 102 } 103 104 static u64 maplen(struct vhost_iotlb_map *map, struct mlx5_vdpa_direct_mr *mr) 105 { 106 return map_end(map, mr) - map_start(map, mr); 107 } 108 109 #define MLX5_VDPA_INVALID_START_ADDR ((u64)-1) 110 #define MLX5_VDPA_INVALID_LEN ((u64)-1) 111 112 static u64 indir_start_addr(struct mlx5_vdpa_mr *mkey) 113 { 114 struct mlx5_vdpa_direct_mr *s; 115 116 s = list_first_entry_or_null(&mkey->head, struct mlx5_vdpa_direct_mr, list); 117 if (!s) 118 return MLX5_VDPA_INVALID_START_ADDR; 119 120 return s->start; 121 } 122 123 static u64 indir_len(struct mlx5_vdpa_mr *mkey) 124 { 125 struct mlx5_vdpa_direct_mr *s; 126 struct mlx5_vdpa_direct_mr *e; 127 128 s = list_first_entry_or_null(&mkey->head, struct mlx5_vdpa_direct_mr, list); 129 if (!s) 130 return MLX5_VDPA_INVALID_LEN; 131 132 e = list_last_entry(&mkey->head, struct mlx5_vdpa_direct_mr, list); 133 134 return e->end - s->start; 135 } 136 137 #define LOG_MAX_KLM_SIZE 30 138 #define MAX_KLM_SIZE BIT(LOG_MAX_KLM_SIZE) 139 140 static u32 klm_bcount(u64 size) 141 { 142 return (u32)size; 143 } 144 145 static void fill_indir(struct mlx5_vdpa_dev *mvdev, struct mlx5_vdpa_mr *mkey, void *in) 146 { 147 struct mlx5_vdpa_direct_mr *dmr; 148 struct mlx5_klm *klmarr; 149 struct mlx5_klm *klm; 150 bool first = true; 151 u64 preve; 152 int i; 153 154 klmarr = MLX5_ADDR_OF(create_mkey_in, in, klm_pas_mtt); 155 i = 0; 156 list_for_each_entry(dmr, &mkey->head, list) { 157 again: 158 klm = &klmarr[i++]; 159 if (first) { 160 preve = dmr->start; 161 first = false; 162 } 163 164 if (preve == dmr->start) { 165 klm->key = cpu_to_be32(dmr->mr); 166 klm->bcount = cpu_to_be32(klm_bcount(dmr->end - dmr->start)); 167 preve = dmr->end; 168 } else { 169 klm->key = cpu_to_be32(mvdev->res.null_mkey); 170 klm->bcount = cpu_to_be32(klm_bcount(dmr->start - preve)); 171 preve = dmr->start; 172 goto again; 173 } 174 } 175 } 176 177 static int klm_byte_size(int nklms) 178 { 179 return 16 * ALIGN(nklms, 4); 180 } 181 182 static int create_indirect_key(struct mlx5_vdpa_dev *mvdev, struct mlx5_vdpa_mr *mr) 183 { 184 int inlen; 185 void *mkc; 186 void *in; 187 int err; 188 u64 start; 189 u64 len; 190 191 start = indir_start_addr(mr); 192 len = indir_len(mr); 193 if (start == MLX5_VDPA_INVALID_START_ADDR || len == MLX5_VDPA_INVALID_LEN) 194 return -EINVAL; 195 196 inlen = MLX5_ST_SZ_BYTES(create_mkey_in) + klm_byte_size(mr->num_klms); 197 in = kzalloc(inlen, GFP_KERNEL); 198 if (!in) 199 return -ENOMEM; 200 201 MLX5_SET(create_mkey_in, in, uid, mvdev->res.uid); 202 mkc = MLX5_ADDR_OF(create_mkey_in, in, memory_key_mkey_entry); 203 MLX5_SET(mkc, mkc, lw, 1); 204 MLX5_SET(mkc, mkc, lr, 1); 205 mlx5_set_access_mode(mkc, MLX5_MKC_ACCESS_MODE_KLMS); 206 MLX5_SET(mkc, mkc, qpn, 0xffffff); 207 MLX5_SET(mkc, mkc, pd, mvdev->res.pdn); 208 MLX5_SET64(mkc, mkc, start_addr, start); 209 MLX5_SET64(mkc, mkc, len, len); 210 MLX5_SET(mkc, mkc, translations_octword_size, klm_byte_size(mr->num_klms) / 16); 211 MLX5_SET(create_mkey_in, in, translations_octword_actual_size, mr->num_klms); 212 fill_indir(mvdev, mr, in); 213 err = mlx5_vdpa_create_mkey(mvdev, &mr->mkey, in, inlen); 214 kfree(in); 215 return err; 216 } 217 218 static void destroy_indirect_key(struct mlx5_vdpa_dev *mvdev, struct mlx5_vdpa_mr *mkey) 219 { 220 mlx5_vdpa_destroy_mkey(mvdev, mkey->mkey); 221 } 222 223 static int map_direct_mr(struct mlx5_vdpa_dev *mvdev, struct mlx5_vdpa_direct_mr *mr, 224 struct vhost_iotlb *iotlb) 225 { 226 struct vhost_iotlb_map *map; 227 unsigned long lgcd = 0; 228 int log_entity_size; 229 unsigned long size; 230 u64 start = 0; 231 int err; 232 struct page *pg; 233 unsigned int nsg; 234 int sglen; 235 u64 pa; 236 u64 paend; 237 struct scatterlist *sg; 238 struct device *dma = mvdev->vdev.dma_dev; 239 240 for (map = vhost_iotlb_itree_first(iotlb, mr->start, mr->end - 1); 241 map; map = vhost_iotlb_itree_next(map, start, mr->end - 1)) { 242 size = maplen(map, mr); 243 lgcd = gcd(lgcd, size); 244 start += size; 245 } 246 log_entity_size = ilog2(lgcd); 247 248 sglen = 1 << log_entity_size; 249 nsg = MLX5_DIV_ROUND_UP_POW2(mr->end - mr->start, log_entity_size); 250 251 err = sg_alloc_table(&mr->sg_head, nsg, GFP_KERNEL); 252 if (err) 253 return err; 254 255 sg = mr->sg_head.sgl; 256 for (map = vhost_iotlb_itree_first(iotlb, mr->start, mr->end - 1); 257 map; map = vhost_iotlb_itree_next(map, mr->start, mr->end - 1)) { 258 paend = map->addr + maplen(map, mr); 259 for (pa = map->addr; pa < paend; pa += sglen) { 260 pg = pfn_to_page(__phys_to_pfn(pa)); 261 if (!sg) { 262 mlx5_vdpa_warn(mvdev, "sg null. start 0x%llx, end 0x%llx\n", 263 map->start, map->last + 1); 264 err = -ENOMEM; 265 goto err_map; 266 } 267 sg_set_page(sg, pg, sglen, 0); 268 sg = sg_next(sg); 269 if (!sg) 270 goto done; 271 } 272 } 273 done: 274 mr->log_size = log_entity_size; 275 mr->nsg = nsg; 276 mr->nent = dma_map_sg_attrs(dma, mr->sg_head.sgl, mr->nsg, DMA_BIDIRECTIONAL, 0); 277 if (!mr->nent) { 278 err = -ENOMEM; 279 goto err_map; 280 } 281 282 err = create_direct_mr(mvdev, mr); 283 if (err) 284 goto err_direct; 285 286 return 0; 287 288 err_direct: 289 dma_unmap_sg_attrs(dma, mr->sg_head.sgl, mr->nsg, DMA_BIDIRECTIONAL, 0); 290 err_map: 291 sg_free_table(&mr->sg_head); 292 return err; 293 } 294 295 static void unmap_direct_mr(struct mlx5_vdpa_dev *mvdev, struct mlx5_vdpa_direct_mr *mr) 296 { 297 struct device *dma = mvdev->vdev.dma_dev; 298 299 destroy_direct_mr(mvdev, mr); 300 dma_unmap_sg_attrs(dma, mr->sg_head.sgl, mr->nsg, DMA_BIDIRECTIONAL, 0); 301 sg_free_table(&mr->sg_head); 302 } 303 304 static int add_direct_chain(struct mlx5_vdpa_dev *mvdev, u64 start, u64 size, u8 perm, 305 struct vhost_iotlb *iotlb) 306 { 307 struct mlx5_vdpa_mr *mr = &mvdev->mr; 308 struct mlx5_vdpa_direct_mr *dmr; 309 struct mlx5_vdpa_direct_mr *n; 310 LIST_HEAD(tmp); 311 u64 st; 312 u64 sz; 313 int err; 314 315 st = start; 316 while (size) { 317 sz = (u32)min_t(u64, MAX_KLM_SIZE, size); 318 dmr = kzalloc(sizeof(*dmr), GFP_KERNEL); 319 if (!dmr) { 320 err = -ENOMEM; 321 goto err_alloc; 322 } 323 324 dmr->start = st; 325 dmr->end = st + sz; 326 dmr->perm = perm; 327 err = map_direct_mr(mvdev, dmr, iotlb); 328 if (err) { 329 kfree(dmr); 330 goto err_alloc; 331 } 332 333 list_add_tail(&dmr->list, &tmp); 334 size -= sz; 335 mr->num_directs++; 336 mr->num_klms++; 337 st += sz; 338 } 339 list_splice_tail(&tmp, &mr->head); 340 return 0; 341 342 err_alloc: 343 list_for_each_entry_safe(dmr, n, &mr->head, list) { 344 list_del_init(&dmr->list); 345 unmap_direct_mr(mvdev, dmr); 346 kfree(dmr); 347 } 348 return err; 349 } 350 351 /* The iotlb pointer contains a list of maps. Go over the maps, possibly 352 * merging mergeable maps, and create direct memory keys that provide the 353 * device access to memory. The direct mkeys are then referred to by the 354 * indirect memory key that provides access to the enitre address space given 355 * by iotlb. 356 */ 357 static int create_user_mr(struct mlx5_vdpa_dev *mvdev, struct vhost_iotlb *iotlb) 358 { 359 struct mlx5_vdpa_mr *mr = &mvdev->mr; 360 struct mlx5_vdpa_direct_mr *dmr; 361 struct mlx5_vdpa_direct_mr *n; 362 struct vhost_iotlb_map *map; 363 u32 pperm = U16_MAX; 364 u64 last = U64_MAX; 365 u64 ps = U64_MAX; 366 u64 pe = U64_MAX; 367 u64 start = 0; 368 int err = 0; 369 int nnuls; 370 371 INIT_LIST_HEAD(&mr->head); 372 for (map = vhost_iotlb_itree_first(iotlb, start, last); map; 373 map = vhost_iotlb_itree_next(map, start, last)) { 374 start = map->start; 375 if (pe == map->start && pperm == map->perm) { 376 pe = map->last + 1; 377 } else { 378 if (ps != U64_MAX) { 379 if (pe < map->start) { 380 /* We have a hole in the map. Check how 381 * many null keys are required to fill it. 382 */ 383 nnuls = MLX5_DIV_ROUND_UP_POW2(map->start - pe, 384 LOG_MAX_KLM_SIZE); 385 mr->num_klms += nnuls; 386 } 387 err = add_direct_chain(mvdev, ps, pe - ps, pperm, iotlb); 388 if (err) 389 goto err_chain; 390 } 391 ps = map->start; 392 pe = map->last + 1; 393 pperm = map->perm; 394 } 395 } 396 err = add_direct_chain(mvdev, ps, pe - ps, pperm, iotlb); 397 if (err) 398 goto err_chain; 399 400 /* Create the memory key that defines the guests's address space. This 401 * memory key refers to the direct keys that contain the MTT 402 * translations 403 */ 404 err = create_indirect_key(mvdev, mr); 405 if (err) 406 goto err_chain; 407 408 mr->user_mr = true; 409 return 0; 410 411 err_chain: 412 list_for_each_entry_safe_reverse(dmr, n, &mr->head, list) { 413 list_del_init(&dmr->list); 414 unmap_direct_mr(mvdev, dmr); 415 kfree(dmr); 416 } 417 return err; 418 } 419 420 static int create_dma_mr(struct mlx5_vdpa_dev *mvdev, struct mlx5_vdpa_mr *mr) 421 { 422 int inlen = MLX5_ST_SZ_BYTES(create_mkey_in); 423 void *mkc; 424 u32 *in; 425 int err; 426 427 in = kzalloc(inlen, GFP_KERNEL); 428 if (!in) 429 return -ENOMEM; 430 431 mkc = MLX5_ADDR_OF(create_mkey_in, in, memory_key_mkey_entry); 432 433 MLX5_SET(mkc, mkc, access_mode_1_0, MLX5_MKC_ACCESS_MODE_PA); 434 MLX5_SET(mkc, mkc, length64, 1); 435 MLX5_SET(mkc, mkc, lw, 1); 436 MLX5_SET(mkc, mkc, lr, 1); 437 MLX5_SET(mkc, mkc, pd, mvdev->res.pdn); 438 MLX5_SET(mkc, mkc, qpn, 0xffffff); 439 440 err = mlx5_vdpa_create_mkey(mvdev, &mr->mkey, in, inlen); 441 if (!err) 442 mr->user_mr = false; 443 444 kfree(in); 445 return err; 446 } 447 448 static void destroy_dma_mr(struct mlx5_vdpa_dev *mvdev, struct mlx5_vdpa_mr *mr) 449 { 450 mlx5_vdpa_destroy_mkey(mvdev, mr->mkey); 451 } 452 453 static int dup_iotlb(struct mlx5_vdpa_dev *mvdev, struct vhost_iotlb *src) 454 { 455 struct vhost_iotlb_map *map; 456 u64 start = 0, last = ULLONG_MAX; 457 int err; 458 459 if (!src) { 460 err = vhost_iotlb_add_range(mvdev->cvq.iotlb, start, last, start, VHOST_ACCESS_RW); 461 return err; 462 } 463 464 for (map = vhost_iotlb_itree_first(src, start, last); map; 465 map = vhost_iotlb_itree_next(map, start, last)) { 466 err = vhost_iotlb_add_range(mvdev->cvq.iotlb, map->start, map->last, 467 map->addr, map->perm); 468 if (err) 469 return err; 470 } 471 return 0; 472 } 473 474 static void prune_iotlb(struct mlx5_vdpa_dev *mvdev) 475 { 476 vhost_iotlb_del_range(mvdev->cvq.iotlb, 0, ULLONG_MAX); 477 } 478 479 static void destroy_user_mr(struct mlx5_vdpa_dev *mvdev, struct mlx5_vdpa_mr *mr) 480 { 481 struct mlx5_vdpa_direct_mr *dmr; 482 struct mlx5_vdpa_direct_mr *n; 483 484 destroy_indirect_key(mvdev, mr); 485 list_for_each_entry_safe_reverse(dmr, n, &mr->head, list) { 486 list_del_init(&dmr->list); 487 unmap_direct_mr(mvdev, dmr); 488 kfree(dmr); 489 } 490 } 491 492 static void _mlx5_vdpa_destroy_cvq_mr(struct mlx5_vdpa_dev *mvdev, unsigned int asid) 493 { 494 if (mvdev->group2asid[MLX5_VDPA_CVQ_GROUP] != asid) 495 return; 496 497 prune_iotlb(mvdev); 498 } 499 500 static void _mlx5_vdpa_destroy_dvq_mr(struct mlx5_vdpa_dev *mvdev, unsigned int asid) 501 { 502 struct mlx5_vdpa_mr *mr = &mvdev->mr; 503 504 if (mvdev->group2asid[MLX5_VDPA_DATAVQ_GROUP] != asid) 505 return; 506 507 if (!mr->initialized) 508 return; 509 510 if (mr->user_mr) 511 destroy_user_mr(mvdev, mr); 512 else 513 destroy_dma_mr(mvdev, mr); 514 515 mr->initialized = false; 516 } 517 518 void mlx5_vdpa_destroy_mr_asid(struct mlx5_vdpa_dev *mvdev, unsigned int asid) 519 { 520 struct mlx5_vdpa_mr *mr = &mvdev->mr; 521 522 mutex_lock(&mr->mkey_mtx); 523 524 _mlx5_vdpa_destroy_dvq_mr(mvdev, asid); 525 _mlx5_vdpa_destroy_cvq_mr(mvdev, asid); 526 527 mutex_unlock(&mr->mkey_mtx); 528 } 529 530 void mlx5_vdpa_destroy_mr(struct mlx5_vdpa_dev *mvdev) 531 { 532 mlx5_vdpa_destroy_mr_asid(mvdev, mvdev->group2asid[MLX5_VDPA_CVQ_GROUP]); 533 mlx5_vdpa_destroy_mr_asid(mvdev, mvdev->group2asid[MLX5_VDPA_DATAVQ_GROUP]); 534 } 535 536 static int _mlx5_vdpa_create_cvq_mr(struct mlx5_vdpa_dev *mvdev, 537 struct vhost_iotlb *iotlb, 538 unsigned int asid) 539 { 540 if (mvdev->group2asid[MLX5_VDPA_CVQ_GROUP] != asid) 541 return 0; 542 543 return dup_iotlb(mvdev, iotlb); 544 } 545 546 static int _mlx5_vdpa_create_dvq_mr(struct mlx5_vdpa_dev *mvdev, 547 struct vhost_iotlb *iotlb, 548 unsigned int asid) 549 { 550 struct mlx5_vdpa_mr *mr = &mvdev->mr; 551 int err; 552 553 if (mvdev->group2asid[MLX5_VDPA_DATAVQ_GROUP] != asid) 554 return 0; 555 556 if (mr->initialized) 557 return 0; 558 559 if (iotlb) 560 err = create_user_mr(mvdev, iotlb); 561 else 562 err = create_dma_mr(mvdev, mr); 563 564 if (err) 565 return err; 566 567 mr->initialized = true; 568 569 return 0; 570 } 571 572 static int _mlx5_vdpa_create_mr(struct mlx5_vdpa_dev *mvdev, 573 struct vhost_iotlb *iotlb, unsigned int asid) 574 { 575 int err; 576 577 err = _mlx5_vdpa_create_dvq_mr(mvdev, iotlb, asid); 578 if (err) 579 return err; 580 581 err = _mlx5_vdpa_create_cvq_mr(mvdev, iotlb, asid); 582 if (err) 583 goto out_err; 584 585 return 0; 586 587 out_err: 588 _mlx5_vdpa_destroy_dvq_mr(mvdev, asid); 589 590 return err; 591 } 592 593 int mlx5_vdpa_create_mr(struct mlx5_vdpa_dev *mvdev, struct vhost_iotlb *iotlb, 594 unsigned int asid) 595 { 596 int err; 597 598 mutex_lock(&mvdev->mr.mkey_mtx); 599 err = _mlx5_vdpa_create_mr(mvdev, iotlb, asid); 600 mutex_unlock(&mvdev->mr.mkey_mtx); 601 return err; 602 } 603 604 int mlx5_vdpa_handle_set_map(struct mlx5_vdpa_dev *mvdev, struct vhost_iotlb *iotlb, 605 bool *change_map, unsigned int asid) 606 { 607 struct mlx5_vdpa_mr *mr = &mvdev->mr; 608 int err = 0; 609 610 *change_map = false; 611 mutex_lock(&mr->mkey_mtx); 612 if (mr->initialized) { 613 mlx5_vdpa_info(mvdev, "memory map update\n"); 614 *change_map = true; 615 } 616 if (!*change_map) 617 err = _mlx5_vdpa_create_mr(mvdev, iotlb, asid); 618 mutex_unlock(&mr->mkey_mtx); 619 620 return err; 621 } 622