1 /* 2 * Copyright (c) 2016 Hisilicon Limited. 3 * Copyright (c) 2007, 2008 Mellanox Technologies. All rights reserved. 4 * 5 * This software is available to you under a choice of one of two 6 * licenses. You may choose to be licensed under the terms of the GNU 7 * General Public License (GPL) Version 2, available from the file 8 * COPYING in the main directory of this source tree, or the 9 * OpenIB.org BSD license below: 10 * 11 * Redistribution and use in source and binary forms, with or 12 * without modification, are permitted provided that the following 13 * conditions are met: 14 * 15 * - Redistributions of source code must retain the above 16 * copyright notice, this list of conditions and the following 17 * disclaimer. 18 * 19 * - Redistributions in binary form must reproduce the above 20 * copyright notice, this list of conditions and the following 21 * disclaimer in the documentation and/or other materials 22 * provided with the distribution. 23 * 24 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, 25 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF 26 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND 27 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS 28 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN 29 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN 30 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE 31 * SOFTWARE. 32 */ 33 34 #include <linux/vmalloc.h> 35 #include <rdma/ib_umem.h> 36 #include "hns_roce_device.h" 37 #include "hns_roce_cmd.h" 38 #include "hns_roce_hem.h" 39 40 static u32 hw_index_to_key(int ind) 41 { 42 return ((u32)ind >> 24) | ((u32)ind << 8); 43 } 44 45 unsigned long key_to_hw_index(u32 key) 46 { 47 return (key << 24) | (key >> 8); 48 } 49 50 static int alloc_mr_key(struct hns_roce_dev *hr_dev, struct hns_roce_mr *mr) 51 { 52 struct hns_roce_ida *mtpt_ida = &hr_dev->mr_table.mtpt_ida; 53 struct ib_device *ibdev = &hr_dev->ib_dev; 54 int err; 55 int id; 56 57 /* Allocate a key for mr from mr_table */ 58 id = ida_alloc_range(&mtpt_ida->ida, mtpt_ida->min, mtpt_ida->max, 59 GFP_KERNEL); 60 if (id < 0) { 61 ibdev_err(ibdev, "failed to alloc id for MR key, id(%d)\n", id); 62 return -ENOMEM; 63 } 64 65 mr->key = hw_index_to_key(id); /* MR key */ 66 67 err = hns_roce_table_get(hr_dev, &hr_dev->mr_table.mtpt_table, 68 (unsigned long)id); 69 if (err) { 70 ibdev_err(ibdev, "failed to alloc mtpt, ret = %d.\n", err); 71 goto err_free_bitmap; 72 } 73 74 return 0; 75 err_free_bitmap: 76 ida_free(&mtpt_ida->ida, id); 77 return err; 78 } 79 80 static void free_mr_key(struct hns_roce_dev *hr_dev, struct hns_roce_mr *mr) 81 { 82 unsigned long obj = key_to_hw_index(mr->key); 83 84 hns_roce_table_put(hr_dev, &hr_dev->mr_table.mtpt_table, obj); 85 ida_free(&hr_dev->mr_table.mtpt_ida.ida, (int)obj); 86 } 87 88 static int alloc_mr_pbl(struct hns_roce_dev *hr_dev, struct hns_roce_mr *mr, 89 struct ib_udata *udata, u64 start) 90 { 91 struct ib_device *ibdev = &hr_dev->ib_dev; 92 bool is_fast = mr->type == MR_TYPE_FRMR; 93 struct hns_roce_buf_attr buf_attr = {}; 94 int err; 95 96 mr->pbl_hop_num = is_fast ? 1 : hr_dev->caps.pbl_hop_num; 97 buf_attr.page_shift = is_fast ? PAGE_SHIFT : 98 hr_dev->caps.pbl_buf_pg_sz + PAGE_SHIFT; 99 buf_attr.region[0].size = mr->size; 100 buf_attr.region[0].hopnum = mr->pbl_hop_num; 101 buf_attr.region_count = 1; 102 buf_attr.user_access = mr->access; 103 /* fast MR's buffer is alloced before mapping, not at creation */ 104 buf_attr.mtt_only = is_fast; 105 106 err = hns_roce_mtr_create(hr_dev, &mr->pbl_mtr, &buf_attr, 107 hr_dev->caps.pbl_ba_pg_sz + PAGE_SHIFT, 108 udata, start); 109 if (err) 110 ibdev_err(ibdev, "failed to alloc pbl mtr, ret = %d.\n", err); 111 else 112 mr->npages = mr->pbl_mtr.hem_cfg.buf_pg_count; 113 114 return err; 115 } 116 117 static void free_mr_pbl(struct hns_roce_dev *hr_dev, struct hns_roce_mr *mr) 118 { 119 hns_roce_mtr_destroy(hr_dev, &mr->pbl_mtr); 120 } 121 122 static void hns_roce_mr_free(struct hns_roce_dev *hr_dev, struct hns_roce_mr *mr) 123 { 124 struct ib_device *ibdev = &hr_dev->ib_dev; 125 int ret; 126 127 if (mr->enabled) { 128 ret = hns_roce_destroy_hw_ctx(hr_dev, HNS_ROCE_CMD_DESTROY_MPT, 129 key_to_hw_index(mr->key) & 130 (hr_dev->caps.num_mtpts - 1)); 131 if (ret) 132 ibdev_warn(ibdev, "failed to destroy mpt, ret = %d.\n", 133 ret); 134 } 135 136 free_mr_pbl(hr_dev, mr); 137 free_mr_key(hr_dev, mr); 138 } 139 140 static int hns_roce_mr_enable(struct hns_roce_dev *hr_dev, 141 struct hns_roce_mr *mr) 142 { 143 unsigned long mtpt_idx = key_to_hw_index(mr->key); 144 struct hns_roce_cmd_mailbox *mailbox; 145 struct device *dev = hr_dev->dev; 146 int ret; 147 148 /* Allocate mailbox memory */ 149 mailbox = hns_roce_alloc_cmd_mailbox(hr_dev); 150 if (IS_ERR(mailbox)) 151 return PTR_ERR(mailbox); 152 153 if (mr->type != MR_TYPE_FRMR) 154 ret = hr_dev->hw->write_mtpt(hr_dev, mailbox->buf, mr); 155 else 156 ret = hr_dev->hw->frmr_write_mtpt(hr_dev, mailbox->buf, mr); 157 if (ret) { 158 dev_err(dev, "failed to write mtpt, ret = %d.\n", ret); 159 goto err_page; 160 } 161 162 ret = hns_roce_create_hw_ctx(hr_dev, mailbox, HNS_ROCE_CMD_CREATE_MPT, 163 mtpt_idx & (hr_dev->caps.num_mtpts - 1)); 164 if (ret) { 165 dev_err(dev, "failed to create mpt, ret = %d.\n", ret); 166 goto err_page; 167 } 168 169 mr->enabled = 1; 170 171 err_page: 172 hns_roce_free_cmd_mailbox(hr_dev, mailbox); 173 174 return ret; 175 } 176 177 void hns_roce_init_mr_table(struct hns_roce_dev *hr_dev) 178 { 179 struct hns_roce_ida *mtpt_ida = &hr_dev->mr_table.mtpt_ida; 180 181 ida_init(&mtpt_ida->ida); 182 mtpt_ida->max = hr_dev->caps.num_mtpts - 1; 183 mtpt_ida->min = hr_dev->caps.reserved_mrws; 184 } 185 186 struct ib_mr *hns_roce_get_dma_mr(struct ib_pd *pd, int acc) 187 { 188 struct hns_roce_dev *hr_dev = to_hr_dev(pd->device); 189 struct hns_roce_mr *mr; 190 int ret; 191 192 mr = kzalloc(sizeof(*mr), GFP_KERNEL); 193 if (mr == NULL) 194 return ERR_PTR(-ENOMEM); 195 196 mr->type = MR_TYPE_DMA; 197 mr->pd = to_hr_pd(pd)->pdn; 198 mr->access = acc; 199 200 /* Allocate memory region key */ 201 hns_roce_hem_list_init(&mr->pbl_mtr.hem_list); 202 ret = alloc_mr_key(hr_dev, mr); 203 if (ret) 204 goto err_free; 205 206 ret = hns_roce_mr_enable(hr_dev, mr); 207 if (ret) 208 goto err_mr; 209 210 mr->ibmr.rkey = mr->ibmr.lkey = mr->key; 211 212 return &mr->ibmr; 213 err_mr: 214 free_mr_key(hr_dev, mr); 215 216 err_free: 217 kfree(mr); 218 return ERR_PTR(ret); 219 } 220 221 struct ib_mr *hns_roce_reg_user_mr(struct ib_pd *pd, u64 start, u64 length, 222 u64 virt_addr, int access_flags, 223 struct ib_udata *udata) 224 { 225 struct hns_roce_dev *hr_dev = to_hr_dev(pd->device); 226 struct hns_roce_mr *mr; 227 int ret; 228 229 mr = kzalloc(sizeof(*mr), GFP_KERNEL); 230 if (!mr) 231 return ERR_PTR(-ENOMEM); 232 233 mr->iova = virt_addr; 234 mr->size = length; 235 mr->pd = to_hr_pd(pd)->pdn; 236 mr->access = access_flags; 237 mr->type = MR_TYPE_MR; 238 239 ret = alloc_mr_key(hr_dev, mr); 240 if (ret) 241 goto err_alloc_mr; 242 243 ret = alloc_mr_pbl(hr_dev, mr, udata, start); 244 if (ret) 245 goto err_alloc_key; 246 247 ret = hns_roce_mr_enable(hr_dev, mr); 248 if (ret) 249 goto err_alloc_pbl; 250 251 mr->ibmr.rkey = mr->ibmr.lkey = mr->key; 252 mr->ibmr.length = length; 253 254 return &mr->ibmr; 255 256 err_alloc_pbl: 257 free_mr_pbl(hr_dev, mr); 258 err_alloc_key: 259 free_mr_key(hr_dev, mr); 260 err_alloc_mr: 261 kfree(mr); 262 return ERR_PTR(ret); 263 } 264 265 struct ib_mr *hns_roce_rereg_user_mr(struct ib_mr *ibmr, int flags, u64 start, 266 u64 length, u64 virt_addr, 267 int mr_access_flags, struct ib_pd *pd, 268 struct ib_udata *udata) 269 { 270 struct hns_roce_dev *hr_dev = to_hr_dev(ibmr->device); 271 struct ib_device *ib_dev = &hr_dev->ib_dev; 272 struct hns_roce_mr *mr = to_hr_mr(ibmr); 273 struct hns_roce_cmd_mailbox *mailbox; 274 unsigned long mtpt_idx; 275 int ret; 276 277 if (!mr->enabled) 278 return ERR_PTR(-EINVAL); 279 280 mailbox = hns_roce_alloc_cmd_mailbox(hr_dev); 281 if (IS_ERR(mailbox)) 282 return ERR_CAST(mailbox); 283 284 mtpt_idx = key_to_hw_index(mr->key) & (hr_dev->caps.num_mtpts - 1); 285 286 ret = hns_roce_cmd_mbox(hr_dev, 0, mailbox->dma, HNS_ROCE_CMD_QUERY_MPT, 287 mtpt_idx); 288 if (ret) 289 goto free_cmd_mbox; 290 291 ret = hns_roce_destroy_hw_ctx(hr_dev, HNS_ROCE_CMD_DESTROY_MPT, 292 mtpt_idx); 293 if (ret) 294 ibdev_warn(ib_dev, "failed to destroy MPT, ret = %d.\n", ret); 295 296 mr->enabled = 0; 297 mr->iova = virt_addr; 298 mr->size = length; 299 300 if (flags & IB_MR_REREG_PD) 301 mr->pd = to_hr_pd(pd)->pdn; 302 303 if (flags & IB_MR_REREG_ACCESS) 304 mr->access = mr_access_flags; 305 306 if (flags & IB_MR_REREG_TRANS) { 307 free_mr_pbl(hr_dev, mr); 308 ret = alloc_mr_pbl(hr_dev, mr, udata, start); 309 if (ret) { 310 ibdev_err(ib_dev, "failed to alloc mr PBL, ret = %d.\n", 311 ret); 312 goto free_cmd_mbox; 313 } 314 } 315 316 ret = hr_dev->hw->rereg_write_mtpt(hr_dev, mr, flags, mailbox->buf); 317 if (ret) { 318 ibdev_err(ib_dev, "failed to write mtpt, ret = %d.\n", ret); 319 goto free_cmd_mbox; 320 } 321 322 ret = hns_roce_create_hw_ctx(hr_dev, mailbox, HNS_ROCE_CMD_CREATE_MPT, 323 mtpt_idx); 324 if (ret) { 325 ibdev_err(ib_dev, "failed to create MPT, ret = %d.\n", ret); 326 goto free_cmd_mbox; 327 } 328 329 mr->enabled = 1; 330 331 free_cmd_mbox: 332 hns_roce_free_cmd_mailbox(hr_dev, mailbox); 333 334 if (ret) 335 return ERR_PTR(ret); 336 return NULL; 337 } 338 339 int hns_roce_dereg_mr(struct ib_mr *ibmr, struct ib_udata *udata) 340 { 341 struct hns_roce_dev *hr_dev = to_hr_dev(ibmr->device); 342 struct hns_roce_mr *mr = to_hr_mr(ibmr); 343 int ret = 0; 344 345 if (hr_dev->hw->dereg_mr) 346 hr_dev->hw->dereg_mr(hr_dev); 347 348 hns_roce_mr_free(hr_dev, mr); 349 kfree(mr); 350 351 return ret; 352 } 353 354 struct ib_mr *hns_roce_alloc_mr(struct ib_pd *pd, enum ib_mr_type mr_type, 355 u32 max_num_sg) 356 { 357 struct hns_roce_dev *hr_dev = to_hr_dev(pd->device); 358 struct device *dev = hr_dev->dev; 359 struct hns_roce_mr *mr; 360 int ret; 361 362 if (mr_type != IB_MR_TYPE_MEM_REG) 363 return ERR_PTR(-EINVAL); 364 365 if (max_num_sg > HNS_ROCE_FRMR_MAX_PA) { 366 dev_err(dev, "max_num_sg larger than %d\n", 367 HNS_ROCE_FRMR_MAX_PA); 368 return ERR_PTR(-EINVAL); 369 } 370 371 mr = kzalloc(sizeof(*mr), GFP_KERNEL); 372 if (!mr) 373 return ERR_PTR(-ENOMEM); 374 375 mr->type = MR_TYPE_FRMR; 376 mr->pd = to_hr_pd(pd)->pdn; 377 mr->size = max_num_sg * (1 << PAGE_SHIFT); 378 379 /* Allocate memory region key */ 380 ret = alloc_mr_key(hr_dev, mr); 381 if (ret) 382 goto err_free; 383 384 ret = alloc_mr_pbl(hr_dev, mr, NULL, 0); 385 if (ret) 386 goto err_key; 387 388 ret = hns_roce_mr_enable(hr_dev, mr); 389 if (ret) 390 goto err_pbl; 391 392 mr->ibmr.rkey = mr->ibmr.lkey = mr->key; 393 mr->ibmr.length = mr->size; 394 395 return &mr->ibmr; 396 397 err_key: 398 free_mr_key(hr_dev, mr); 399 err_pbl: 400 free_mr_pbl(hr_dev, mr); 401 err_free: 402 kfree(mr); 403 return ERR_PTR(ret); 404 } 405 406 static int hns_roce_set_page(struct ib_mr *ibmr, u64 addr) 407 { 408 struct hns_roce_mr *mr = to_hr_mr(ibmr); 409 410 if (likely(mr->npages < mr->pbl_mtr.hem_cfg.buf_pg_count)) { 411 mr->page_list[mr->npages++] = addr; 412 return 0; 413 } 414 415 return -ENOBUFS; 416 } 417 418 int hns_roce_map_mr_sg(struct ib_mr *ibmr, struct scatterlist *sg, int sg_nents, 419 unsigned int *sg_offset) 420 { 421 struct hns_roce_dev *hr_dev = to_hr_dev(ibmr->device); 422 struct ib_device *ibdev = &hr_dev->ib_dev; 423 struct hns_roce_mr *mr = to_hr_mr(ibmr); 424 struct hns_roce_mtr *mtr = &mr->pbl_mtr; 425 int ret = 0; 426 427 mr->npages = 0; 428 mr->page_list = kvcalloc(mr->pbl_mtr.hem_cfg.buf_pg_count, 429 sizeof(dma_addr_t), GFP_KERNEL); 430 if (!mr->page_list) 431 return ret; 432 433 ret = ib_sg_to_pages(ibmr, sg, sg_nents, sg_offset, hns_roce_set_page); 434 if (ret < 1) { 435 ibdev_err(ibdev, "failed to store sg pages %u %u, cnt = %d.\n", 436 mr->npages, mr->pbl_mtr.hem_cfg.buf_pg_count, ret); 437 goto err_page_list; 438 } 439 440 mtr->hem_cfg.region[0].offset = 0; 441 mtr->hem_cfg.region[0].count = mr->npages; 442 mtr->hem_cfg.region[0].hopnum = mr->pbl_hop_num; 443 mtr->hem_cfg.region_count = 1; 444 ret = hns_roce_mtr_map(hr_dev, mtr, mr->page_list, mr->npages); 445 if (ret) { 446 ibdev_err(ibdev, "failed to map sg mtr, ret = %d.\n", ret); 447 ret = 0; 448 } else { 449 mr->pbl_mtr.hem_cfg.buf_pg_shift = (u32)ilog2(ibmr->page_size); 450 ret = mr->npages; 451 } 452 453 err_page_list: 454 kvfree(mr->page_list); 455 mr->page_list = NULL; 456 457 return ret; 458 } 459 460 static void hns_roce_mw_free(struct hns_roce_dev *hr_dev, 461 struct hns_roce_mw *mw) 462 { 463 struct device *dev = hr_dev->dev; 464 int ret; 465 466 if (mw->enabled) { 467 ret = hns_roce_destroy_hw_ctx(hr_dev, HNS_ROCE_CMD_DESTROY_MPT, 468 key_to_hw_index(mw->rkey) & 469 (hr_dev->caps.num_mtpts - 1)); 470 if (ret) 471 dev_warn(dev, "MW DESTROY_MPT failed (%d)\n", ret); 472 473 hns_roce_table_put(hr_dev, &hr_dev->mr_table.mtpt_table, 474 key_to_hw_index(mw->rkey)); 475 } 476 477 ida_free(&hr_dev->mr_table.mtpt_ida.ida, 478 (int)key_to_hw_index(mw->rkey)); 479 } 480 481 static int hns_roce_mw_enable(struct hns_roce_dev *hr_dev, 482 struct hns_roce_mw *mw) 483 { 484 struct hns_roce_mr_table *mr_table = &hr_dev->mr_table; 485 struct hns_roce_cmd_mailbox *mailbox; 486 struct device *dev = hr_dev->dev; 487 unsigned long mtpt_idx = key_to_hw_index(mw->rkey); 488 int ret; 489 490 /* prepare HEM entry memory */ 491 ret = hns_roce_table_get(hr_dev, &mr_table->mtpt_table, mtpt_idx); 492 if (ret) 493 return ret; 494 495 mailbox = hns_roce_alloc_cmd_mailbox(hr_dev); 496 if (IS_ERR(mailbox)) { 497 ret = PTR_ERR(mailbox); 498 goto err_table; 499 } 500 501 ret = hr_dev->hw->mw_write_mtpt(mailbox->buf, mw); 502 if (ret) { 503 dev_err(dev, "MW write mtpt fail!\n"); 504 goto err_page; 505 } 506 507 ret = hns_roce_create_hw_ctx(hr_dev, mailbox, HNS_ROCE_CMD_CREATE_MPT, 508 mtpt_idx & (hr_dev->caps.num_mtpts - 1)); 509 if (ret) { 510 dev_err(dev, "MW CREATE_MPT failed (%d)\n", ret); 511 goto err_page; 512 } 513 514 mw->enabled = 1; 515 516 hns_roce_free_cmd_mailbox(hr_dev, mailbox); 517 518 return 0; 519 520 err_page: 521 hns_roce_free_cmd_mailbox(hr_dev, mailbox); 522 523 err_table: 524 hns_roce_table_put(hr_dev, &mr_table->mtpt_table, mtpt_idx); 525 526 return ret; 527 } 528 529 int hns_roce_alloc_mw(struct ib_mw *ibmw, struct ib_udata *udata) 530 { 531 struct hns_roce_dev *hr_dev = to_hr_dev(ibmw->device); 532 struct hns_roce_ida *mtpt_ida = &hr_dev->mr_table.mtpt_ida; 533 struct ib_device *ibdev = &hr_dev->ib_dev; 534 struct hns_roce_mw *mw = to_hr_mw(ibmw); 535 int ret; 536 int id; 537 538 /* Allocate a key for mw from mr_table */ 539 id = ida_alloc_range(&mtpt_ida->ida, mtpt_ida->min, mtpt_ida->max, 540 GFP_KERNEL); 541 if (id < 0) { 542 ibdev_err(ibdev, "failed to alloc id for MW key, id(%d)\n", id); 543 return -ENOMEM; 544 } 545 546 mw->rkey = hw_index_to_key(id); 547 548 ibmw->rkey = mw->rkey; 549 mw->pdn = to_hr_pd(ibmw->pd)->pdn; 550 mw->pbl_hop_num = hr_dev->caps.pbl_hop_num; 551 mw->pbl_ba_pg_sz = hr_dev->caps.pbl_ba_pg_sz; 552 mw->pbl_buf_pg_sz = hr_dev->caps.pbl_buf_pg_sz; 553 554 ret = hns_roce_mw_enable(hr_dev, mw); 555 if (ret) 556 goto err_mw; 557 558 return 0; 559 560 err_mw: 561 hns_roce_mw_free(hr_dev, mw); 562 return ret; 563 } 564 565 int hns_roce_dealloc_mw(struct ib_mw *ibmw) 566 { 567 struct hns_roce_dev *hr_dev = to_hr_dev(ibmw->device); 568 struct hns_roce_mw *mw = to_hr_mw(ibmw); 569 570 hns_roce_mw_free(hr_dev, mw); 571 return 0; 572 } 573 574 static int mtr_map_region(struct hns_roce_dev *hr_dev, struct hns_roce_mtr *mtr, 575 struct hns_roce_buf_region *region, dma_addr_t *pages, 576 int max_count) 577 { 578 int count, npage; 579 int offset, end; 580 __le64 *mtts; 581 u64 addr; 582 int i; 583 584 offset = region->offset; 585 end = offset + region->count; 586 npage = 0; 587 while (offset < end && npage < max_count) { 588 count = 0; 589 mtts = hns_roce_hem_list_find_mtt(hr_dev, &mtr->hem_list, 590 offset, &count, NULL); 591 if (!mtts) 592 return -ENOBUFS; 593 594 for (i = 0; i < count && npage < max_count; i++) { 595 addr = pages[npage]; 596 597 mtts[i] = cpu_to_le64(addr); 598 npage++; 599 } 600 offset += count; 601 } 602 603 return npage; 604 } 605 606 static inline bool mtr_has_mtt(struct hns_roce_buf_attr *attr) 607 { 608 int i; 609 610 for (i = 0; i < attr->region_count; i++) 611 if (attr->region[i].hopnum != HNS_ROCE_HOP_NUM_0 && 612 attr->region[i].hopnum > 0) 613 return true; 614 615 /* because the mtr only one root base address, when hopnum is 0 means 616 * root base address equals the first buffer address, thus all alloced 617 * memory must in a continuous space accessed by direct mode. 618 */ 619 return false; 620 } 621 622 static inline size_t mtr_bufs_size(struct hns_roce_buf_attr *attr) 623 { 624 size_t size = 0; 625 int i; 626 627 for (i = 0; i < attr->region_count; i++) 628 size += attr->region[i].size; 629 630 return size; 631 } 632 633 /* 634 * check the given pages in continuous address space 635 * Returns 0 on success, or the error page num. 636 */ 637 static inline int mtr_check_direct_pages(dma_addr_t *pages, int page_count, 638 unsigned int page_shift) 639 { 640 size_t page_size = 1 << page_shift; 641 int i; 642 643 for (i = 1; i < page_count; i++) 644 if (pages[i] - pages[i - 1] != page_size) 645 return i; 646 647 return 0; 648 } 649 650 static void mtr_free_bufs(struct hns_roce_dev *hr_dev, struct hns_roce_mtr *mtr) 651 { 652 /* release user buffers */ 653 if (mtr->umem) { 654 ib_umem_release(mtr->umem); 655 mtr->umem = NULL; 656 } 657 658 /* release kernel buffers */ 659 if (mtr->kmem) { 660 hns_roce_buf_free(hr_dev, mtr->kmem); 661 mtr->kmem = NULL; 662 } 663 } 664 665 static int mtr_alloc_bufs(struct hns_roce_dev *hr_dev, struct hns_roce_mtr *mtr, 666 struct hns_roce_buf_attr *buf_attr, 667 struct ib_udata *udata, unsigned long user_addr) 668 { 669 struct ib_device *ibdev = &hr_dev->ib_dev; 670 size_t total_size; 671 672 total_size = mtr_bufs_size(buf_attr); 673 674 if (udata) { 675 mtr->kmem = NULL; 676 mtr->umem = ib_umem_get(ibdev, user_addr, total_size, 677 buf_attr->user_access); 678 if (IS_ERR_OR_NULL(mtr->umem)) { 679 ibdev_err(ibdev, "failed to get umem, ret = %ld.\n", 680 PTR_ERR(mtr->umem)); 681 return -ENOMEM; 682 } 683 } else { 684 mtr->umem = NULL; 685 mtr->kmem = hns_roce_buf_alloc(hr_dev, total_size, 686 buf_attr->page_shift, 687 mtr->hem_cfg.is_direct ? 688 HNS_ROCE_BUF_DIRECT : 0); 689 if (IS_ERR(mtr->kmem)) { 690 ibdev_err(ibdev, "failed to alloc kmem, ret = %ld.\n", 691 PTR_ERR(mtr->kmem)); 692 return PTR_ERR(mtr->kmem); 693 } 694 } 695 696 return 0; 697 } 698 699 static int mtr_map_bufs(struct hns_roce_dev *hr_dev, struct hns_roce_mtr *mtr, 700 int page_count, unsigned int page_shift) 701 { 702 struct ib_device *ibdev = &hr_dev->ib_dev; 703 dma_addr_t *pages; 704 int npage; 705 int ret; 706 707 /* alloc a tmp array to store buffer's dma address */ 708 pages = kvcalloc(page_count, sizeof(dma_addr_t), GFP_KERNEL); 709 if (!pages) 710 return -ENOMEM; 711 712 if (mtr->umem) 713 npage = hns_roce_get_umem_bufs(hr_dev, pages, page_count, 714 mtr->umem, page_shift); 715 else 716 npage = hns_roce_get_kmem_bufs(hr_dev, pages, page_count, 717 mtr->kmem, page_shift); 718 719 if (npage != page_count) { 720 ibdev_err(ibdev, "failed to get mtr page %d != %d.\n", npage, 721 page_count); 722 ret = -ENOBUFS; 723 goto err_alloc_list; 724 } 725 726 if (mtr->hem_cfg.is_direct && npage > 1) { 727 ret = mtr_check_direct_pages(pages, npage, page_shift); 728 if (ret) { 729 ibdev_err(ibdev, "failed to check %s page: %d / %d.\n", 730 mtr->umem ? "umtr" : "kmtr", ret, npage); 731 ret = -ENOBUFS; 732 goto err_alloc_list; 733 } 734 } 735 736 ret = hns_roce_mtr_map(hr_dev, mtr, pages, page_count); 737 if (ret) 738 ibdev_err(ibdev, "failed to map mtr pages, ret = %d.\n", ret); 739 740 err_alloc_list: 741 kvfree(pages); 742 743 return ret; 744 } 745 746 int hns_roce_mtr_map(struct hns_roce_dev *hr_dev, struct hns_roce_mtr *mtr, 747 dma_addr_t *pages, unsigned int page_cnt) 748 { 749 struct ib_device *ibdev = &hr_dev->ib_dev; 750 struct hns_roce_buf_region *r; 751 unsigned int i, mapped_cnt; 752 int ret = 0; 753 754 /* 755 * Only use the first page address as root ba when hopnum is 0, this 756 * is because the addresses of all pages are consecutive in this case. 757 */ 758 if (mtr->hem_cfg.is_direct) { 759 mtr->hem_cfg.root_ba = pages[0]; 760 return 0; 761 } 762 763 for (i = 0, mapped_cnt = 0; i < mtr->hem_cfg.region_count && 764 mapped_cnt < page_cnt; i++) { 765 r = &mtr->hem_cfg.region[i]; 766 /* if hopnum is 0, no need to map pages in this region */ 767 if (!r->hopnum) { 768 mapped_cnt += r->count; 769 continue; 770 } 771 772 if (r->offset + r->count > page_cnt) { 773 ret = -EINVAL; 774 ibdev_err(ibdev, 775 "failed to check mtr%u count %u + %u > %u.\n", 776 i, r->offset, r->count, page_cnt); 777 return ret; 778 } 779 780 ret = mtr_map_region(hr_dev, mtr, r, &pages[r->offset], 781 page_cnt - mapped_cnt); 782 if (ret < 0) { 783 ibdev_err(ibdev, 784 "failed to map mtr%u offset %u, ret = %d.\n", 785 i, r->offset, ret); 786 return ret; 787 } 788 mapped_cnt += ret; 789 ret = 0; 790 } 791 792 if (mapped_cnt < page_cnt) { 793 ret = -ENOBUFS; 794 ibdev_err(ibdev, "failed to map mtr pages count: %u < %u.\n", 795 mapped_cnt, page_cnt); 796 } 797 798 return ret; 799 } 800 801 int hns_roce_mtr_find(struct hns_roce_dev *hr_dev, struct hns_roce_mtr *mtr, 802 u32 offset, u64 *mtt_buf, int mtt_max, u64 *base_addr) 803 { 804 struct hns_roce_hem_cfg *cfg = &mtr->hem_cfg; 805 int mtt_count, left; 806 u32 start_index; 807 int total = 0; 808 __le64 *mtts; 809 u32 npage; 810 u64 addr; 811 812 if (!mtt_buf || mtt_max < 1) 813 goto done; 814 815 /* no mtt memory in direct mode, so just return the buffer address */ 816 if (cfg->is_direct) { 817 start_index = offset >> HNS_HW_PAGE_SHIFT; 818 for (mtt_count = 0; mtt_count < cfg->region_count && 819 total < mtt_max; mtt_count++) { 820 npage = cfg->region[mtt_count].offset; 821 if (npage < start_index) 822 continue; 823 824 addr = cfg->root_ba + (npage << HNS_HW_PAGE_SHIFT); 825 mtt_buf[total] = addr; 826 827 total++; 828 } 829 830 goto done; 831 } 832 833 start_index = offset >> cfg->buf_pg_shift; 834 left = mtt_max; 835 while (left > 0) { 836 mtt_count = 0; 837 mtts = hns_roce_hem_list_find_mtt(hr_dev, &mtr->hem_list, 838 start_index + total, 839 &mtt_count, NULL); 840 if (!mtts || !mtt_count) 841 goto done; 842 843 npage = min(mtt_count, left); 844 left -= npage; 845 for (mtt_count = 0; mtt_count < npage; mtt_count++) 846 mtt_buf[total++] = le64_to_cpu(mtts[mtt_count]); 847 } 848 849 done: 850 if (base_addr) 851 *base_addr = cfg->root_ba; 852 853 return total; 854 } 855 856 static int mtr_init_buf_cfg(struct hns_roce_dev *hr_dev, 857 struct hns_roce_buf_attr *attr, 858 struct hns_roce_hem_cfg *cfg, 859 unsigned int *buf_page_shift, u64 unalinged_size) 860 { 861 struct hns_roce_buf_region *r; 862 u64 first_region_padding; 863 int page_cnt, region_cnt; 864 unsigned int page_shift; 865 size_t buf_size; 866 867 /* If mtt is disabled, all pages must be within a continuous range */ 868 cfg->is_direct = !mtr_has_mtt(attr); 869 buf_size = mtr_bufs_size(attr); 870 if (cfg->is_direct) { 871 /* When HEM buffer uses 0-level addressing, the page size is 872 * equal to the whole buffer size, and we split the buffer into 873 * small pages which is used to check whether the adjacent 874 * units are in the continuous space and its size is fixed to 875 * 4K based on hns ROCEE's requirement. 876 */ 877 page_shift = HNS_HW_PAGE_SHIFT; 878 879 /* The ROCEE requires the page size to be 4K * 2 ^ N. */ 880 cfg->buf_pg_count = 1; 881 cfg->buf_pg_shift = HNS_HW_PAGE_SHIFT + 882 order_base_2(DIV_ROUND_UP(buf_size, HNS_HW_PAGE_SIZE)); 883 first_region_padding = 0; 884 } else { 885 page_shift = attr->page_shift; 886 cfg->buf_pg_count = DIV_ROUND_UP(buf_size + unalinged_size, 887 1 << page_shift); 888 cfg->buf_pg_shift = page_shift; 889 first_region_padding = unalinged_size; 890 } 891 892 /* Convert buffer size to page index and page count for each region and 893 * the buffer's offset needs to be appended to the first region. 894 */ 895 for (page_cnt = 0, region_cnt = 0; region_cnt < attr->region_count && 896 region_cnt < ARRAY_SIZE(cfg->region); region_cnt++) { 897 r = &cfg->region[region_cnt]; 898 r->offset = page_cnt; 899 buf_size = hr_hw_page_align(attr->region[region_cnt].size + 900 first_region_padding); 901 r->count = DIV_ROUND_UP(buf_size, 1 << page_shift); 902 first_region_padding = 0; 903 page_cnt += r->count; 904 r->hopnum = to_hr_hem_hopnum(attr->region[region_cnt].hopnum, 905 r->count); 906 } 907 908 cfg->region_count = region_cnt; 909 *buf_page_shift = page_shift; 910 911 return page_cnt; 912 } 913 914 static int mtr_alloc_mtt(struct hns_roce_dev *hr_dev, struct hns_roce_mtr *mtr, 915 unsigned int ba_page_shift) 916 { 917 struct hns_roce_hem_cfg *cfg = &mtr->hem_cfg; 918 int ret; 919 920 hns_roce_hem_list_init(&mtr->hem_list); 921 if (!cfg->is_direct) { 922 ret = hns_roce_hem_list_request(hr_dev, &mtr->hem_list, 923 cfg->region, cfg->region_count, 924 ba_page_shift); 925 if (ret) 926 return ret; 927 cfg->root_ba = mtr->hem_list.root_ba; 928 cfg->ba_pg_shift = ba_page_shift; 929 } else { 930 cfg->ba_pg_shift = cfg->buf_pg_shift; 931 } 932 933 return 0; 934 } 935 936 static void mtr_free_mtt(struct hns_roce_dev *hr_dev, struct hns_roce_mtr *mtr) 937 { 938 hns_roce_hem_list_release(hr_dev, &mtr->hem_list); 939 } 940 941 /** 942 * hns_roce_mtr_create - Create hns memory translate region. 943 * 944 * @hr_dev: RoCE device struct pointer 945 * @mtr: memory translate region 946 * @buf_attr: buffer attribute for creating mtr 947 * @ba_page_shift: page shift for multi-hop base address table 948 * @udata: user space context, if it's NULL, means kernel space 949 * @user_addr: userspace virtual address to start at 950 */ 951 int hns_roce_mtr_create(struct hns_roce_dev *hr_dev, struct hns_roce_mtr *mtr, 952 struct hns_roce_buf_attr *buf_attr, 953 unsigned int ba_page_shift, struct ib_udata *udata, 954 unsigned long user_addr) 955 { 956 struct ib_device *ibdev = &hr_dev->ib_dev; 957 unsigned int buf_page_shift = 0; 958 int buf_page_cnt; 959 int ret; 960 961 buf_page_cnt = mtr_init_buf_cfg(hr_dev, buf_attr, &mtr->hem_cfg, 962 &buf_page_shift, 963 udata ? user_addr & ~PAGE_MASK : 0); 964 if (buf_page_cnt < 1 || buf_page_shift < HNS_HW_PAGE_SHIFT) { 965 ibdev_err(ibdev, "failed to init mtr cfg, count %d shift %u.\n", 966 buf_page_cnt, buf_page_shift); 967 return -EINVAL; 968 } 969 970 ret = mtr_alloc_mtt(hr_dev, mtr, ba_page_shift); 971 if (ret) { 972 ibdev_err(ibdev, "failed to alloc mtr mtt, ret = %d.\n", ret); 973 return ret; 974 } 975 976 /* The caller has its own buffer list and invokes the hns_roce_mtr_map() 977 * to finish the MTT configuration. 978 */ 979 if (buf_attr->mtt_only) { 980 mtr->umem = NULL; 981 mtr->kmem = NULL; 982 return 0; 983 } 984 985 ret = mtr_alloc_bufs(hr_dev, mtr, buf_attr, udata, user_addr); 986 if (ret) { 987 ibdev_err(ibdev, "failed to alloc mtr bufs, ret = %d.\n", ret); 988 goto err_alloc_mtt; 989 } 990 991 /* Write buffer's dma address to MTT */ 992 ret = mtr_map_bufs(hr_dev, mtr, buf_page_cnt, buf_page_shift); 993 if (ret) 994 ibdev_err(ibdev, "failed to map mtr bufs, ret = %d.\n", ret); 995 else 996 return 0; 997 998 mtr_free_bufs(hr_dev, mtr); 999 err_alloc_mtt: 1000 mtr_free_mtt(hr_dev, mtr); 1001 return ret; 1002 } 1003 1004 void hns_roce_mtr_destroy(struct hns_roce_dev *hr_dev, struct hns_roce_mtr *mtr) 1005 { 1006 /* release multi-hop addressing resource */ 1007 hns_roce_hem_list_release(hr_dev, &mtr->hem_list); 1008 1009 /* free buffers */ 1010 mtr_free_bufs(hr_dev, mtr); 1011 } 1012