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/platform_device.h> 35 #include <linux/vmalloc.h> 36 #include <rdma/ib_umem.h> 37 #include "hns_roce_device.h" 38 #include "hns_roce_cmd.h" 39 #include "hns_roce_hem.h" 40 41 static u32 hw_index_to_key(unsigned long ind) 42 { 43 return (u32)(ind >> 24) | (ind << 8); 44 } 45 46 unsigned long key_to_hw_index(u32 key) 47 { 48 return (key << 24) | (key >> 8); 49 } 50 51 static int hns_roce_hw_create_mpt(struct hns_roce_dev *hr_dev, 52 struct hns_roce_cmd_mailbox *mailbox, 53 unsigned long mpt_index) 54 { 55 return hns_roce_cmd_mbox(hr_dev, mailbox->dma, 0, mpt_index, 0, 56 HNS_ROCE_CMD_CREATE_MPT, 57 HNS_ROCE_CMD_TIMEOUT_MSECS); 58 } 59 60 int hns_roce_hw_destroy_mpt(struct hns_roce_dev *hr_dev, 61 struct hns_roce_cmd_mailbox *mailbox, 62 unsigned long mpt_index) 63 { 64 return hns_roce_cmd_mbox(hr_dev, 0, mailbox ? mailbox->dma : 0, 65 mpt_index, !mailbox, HNS_ROCE_CMD_DESTROY_MPT, 66 HNS_ROCE_CMD_TIMEOUT_MSECS); 67 } 68 69 static int alloc_mr_key(struct hns_roce_dev *hr_dev, struct hns_roce_mr *mr, 70 u32 pd, u64 iova, u64 size, u32 access) 71 { 72 struct ib_device *ibdev = &hr_dev->ib_dev; 73 unsigned long obj = 0; 74 int err; 75 76 /* Allocate a key for mr from mr_table */ 77 err = hns_roce_bitmap_alloc(&hr_dev->mr_table.mtpt_bitmap, &obj); 78 if (err) { 79 ibdev_err(ibdev, 80 "failed to alloc bitmap for MR key, ret = %d.\n", 81 err); 82 return -ENOMEM; 83 } 84 85 mr->iova = iova; /* MR va starting addr */ 86 mr->size = size; /* MR addr range */ 87 mr->pd = pd; /* MR num */ 88 mr->access = access; /* MR access permit */ 89 mr->enabled = 0; /* MR active status */ 90 mr->key = hw_index_to_key(obj); /* MR key */ 91 92 err = hns_roce_table_get(hr_dev, &hr_dev->mr_table.mtpt_table, obj); 93 if (err) { 94 ibdev_err(ibdev, "failed to alloc mtpt, ret = %d.\n", err); 95 goto err_free_bitmap; 96 } 97 98 return 0; 99 err_free_bitmap: 100 hns_roce_bitmap_free(&hr_dev->mr_table.mtpt_bitmap, obj, BITMAP_NO_RR); 101 return err; 102 } 103 104 static void free_mr_key(struct hns_roce_dev *hr_dev, struct hns_roce_mr *mr) 105 { 106 unsigned long obj = key_to_hw_index(mr->key); 107 108 hns_roce_table_put(hr_dev, &hr_dev->mr_table.mtpt_table, obj); 109 hns_roce_bitmap_free(&hr_dev->mr_table.mtpt_bitmap, obj, BITMAP_NO_RR); 110 } 111 112 static int alloc_mr_pbl(struct hns_roce_dev *hr_dev, struct hns_roce_mr *mr, 113 size_t length, struct ib_udata *udata, u64 start, 114 int access) 115 { 116 struct ib_device *ibdev = &hr_dev->ib_dev; 117 bool is_fast = mr->type == MR_TYPE_FRMR; 118 struct hns_roce_buf_attr buf_attr = {}; 119 int err; 120 121 mr->pbl_hop_num = is_fast ? 1 : hr_dev->caps.pbl_hop_num; 122 buf_attr.page_shift = is_fast ? PAGE_SHIFT : 123 hr_dev->caps.pbl_buf_pg_sz + PAGE_SHIFT; 124 buf_attr.region[0].size = length; 125 buf_attr.region[0].hopnum = mr->pbl_hop_num; 126 buf_attr.region_count = 1; 127 buf_attr.fixed_page = true; 128 buf_attr.user_access = access; 129 /* fast MR's buffer is alloced before mapping, not at creation */ 130 buf_attr.mtt_only = is_fast; 131 132 err = hns_roce_mtr_create(hr_dev, &mr->pbl_mtr, &buf_attr, 133 hr_dev->caps.pbl_ba_pg_sz + HNS_HW_PAGE_SHIFT, 134 udata, start); 135 if (err) 136 ibdev_err(ibdev, "failed to alloc pbl mtr, ret = %d.\n", err); 137 else 138 mr->npages = mr->pbl_mtr.hem_cfg.buf_pg_count; 139 140 return err; 141 } 142 143 static void free_mr_pbl(struct hns_roce_dev *hr_dev, struct hns_roce_mr *mr) 144 { 145 hns_roce_mtr_destroy(hr_dev, &mr->pbl_mtr); 146 } 147 148 static void hns_roce_mr_free(struct hns_roce_dev *hr_dev, 149 struct hns_roce_mr *mr) 150 { 151 struct ib_device *ibdev = &hr_dev->ib_dev; 152 int ret; 153 154 if (mr->enabled) { 155 ret = hns_roce_hw_destroy_mpt(hr_dev, NULL, 156 key_to_hw_index(mr->key) & 157 (hr_dev->caps.num_mtpts - 1)); 158 if (ret) 159 ibdev_warn(ibdev, "failed to destroy mpt, ret = %d.\n", 160 ret); 161 } 162 163 free_mr_pbl(hr_dev, mr); 164 free_mr_key(hr_dev, mr); 165 } 166 167 static int hns_roce_mr_enable(struct hns_roce_dev *hr_dev, 168 struct hns_roce_mr *mr) 169 { 170 unsigned long mtpt_idx = key_to_hw_index(mr->key); 171 struct hns_roce_cmd_mailbox *mailbox; 172 struct device *dev = hr_dev->dev; 173 int ret; 174 175 /* Allocate mailbox memory */ 176 mailbox = hns_roce_alloc_cmd_mailbox(hr_dev); 177 if (IS_ERR(mailbox)) { 178 ret = PTR_ERR(mailbox); 179 return ret; 180 } 181 182 if (mr->type != MR_TYPE_FRMR) 183 ret = hr_dev->hw->write_mtpt(hr_dev, mailbox->buf, mr, 184 mtpt_idx); 185 else 186 ret = hr_dev->hw->frmr_write_mtpt(hr_dev, mailbox->buf, mr); 187 if (ret) { 188 dev_err(dev, "failed to write mtpt, ret = %d.\n", ret); 189 goto err_page; 190 } 191 192 ret = hns_roce_hw_create_mpt(hr_dev, mailbox, 193 mtpt_idx & (hr_dev->caps.num_mtpts - 1)); 194 if (ret) { 195 dev_err(dev, "failed to create mpt, ret = %d.\n", ret); 196 goto err_page; 197 } 198 199 mr->enabled = 1; 200 hns_roce_free_cmd_mailbox(hr_dev, mailbox); 201 202 return 0; 203 204 err_page: 205 hns_roce_free_cmd_mailbox(hr_dev, mailbox); 206 207 return ret; 208 } 209 210 int hns_roce_init_mr_table(struct hns_roce_dev *hr_dev) 211 { 212 struct hns_roce_mr_table *mr_table = &hr_dev->mr_table; 213 int ret; 214 215 ret = hns_roce_bitmap_init(&mr_table->mtpt_bitmap, 216 hr_dev->caps.num_mtpts, 217 hr_dev->caps.num_mtpts - 1, 218 hr_dev->caps.reserved_mrws, 0); 219 return ret; 220 } 221 222 void hns_roce_cleanup_mr_table(struct hns_roce_dev *hr_dev) 223 { 224 struct hns_roce_mr_table *mr_table = &hr_dev->mr_table; 225 226 hns_roce_bitmap_cleanup(&mr_table->mtpt_bitmap); 227 } 228 229 struct ib_mr *hns_roce_get_dma_mr(struct ib_pd *pd, int acc) 230 { 231 struct hns_roce_dev *hr_dev = to_hr_dev(pd->device); 232 struct hns_roce_mr *mr; 233 int ret; 234 235 mr = kzalloc(sizeof(*mr), GFP_KERNEL); 236 if (mr == NULL) 237 return ERR_PTR(-ENOMEM); 238 239 mr->type = MR_TYPE_DMA; 240 241 /* Allocate memory region key */ 242 hns_roce_hem_list_init(&mr->pbl_mtr.hem_list); 243 ret = alloc_mr_key(hr_dev, mr, to_hr_pd(pd)->pdn, 0, 0, acc); 244 if (ret) 245 goto err_free; 246 247 ret = hns_roce_mr_enable(to_hr_dev(pd->device), mr); 248 if (ret) 249 goto err_mr; 250 251 mr->ibmr.rkey = mr->ibmr.lkey = mr->key; 252 253 return &mr->ibmr; 254 err_mr: 255 free_mr_key(hr_dev, mr); 256 257 err_free: 258 kfree(mr); 259 return ERR_PTR(ret); 260 } 261 262 struct ib_mr *hns_roce_reg_user_mr(struct ib_pd *pd, u64 start, u64 length, 263 u64 virt_addr, int access_flags, 264 struct ib_udata *udata) 265 { 266 struct hns_roce_dev *hr_dev = to_hr_dev(pd->device); 267 struct hns_roce_mr *mr; 268 int ret; 269 270 mr = kzalloc(sizeof(*mr), GFP_KERNEL); 271 if (!mr) 272 return ERR_PTR(-ENOMEM); 273 274 mr->type = MR_TYPE_MR; 275 ret = alloc_mr_key(hr_dev, mr, to_hr_pd(pd)->pdn, virt_addr, length, 276 access_flags); 277 if (ret) 278 goto err_alloc_mr; 279 280 ret = alloc_mr_pbl(hr_dev, mr, length, udata, start, access_flags); 281 if (ret) 282 goto err_alloc_key; 283 284 ret = hns_roce_mr_enable(hr_dev, mr); 285 if (ret) 286 goto err_alloc_pbl; 287 288 mr->ibmr.rkey = mr->ibmr.lkey = mr->key; 289 mr->ibmr.length = length; 290 291 return &mr->ibmr; 292 293 err_alloc_pbl: 294 free_mr_pbl(hr_dev, mr); 295 err_alloc_key: 296 free_mr_key(hr_dev, mr); 297 err_alloc_mr: 298 kfree(mr); 299 return ERR_PTR(ret); 300 } 301 302 static int rereg_mr_trans(struct ib_mr *ibmr, int flags, 303 u64 start, u64 length, 304 u64 virt_addr, int mr_access_flags, 305 struct hns_roce_cmd_mailbox *mailbox, 306 u32 pdn, struct ib_udata *udata) 307 { 308 struct hns_roce_dev *hr_dev = to_hr_dev(ibmr->device); 309 struct ib_device *ibdev = &hr_dev->ib_dev; 310 struct hns_roce_mr *mr = to_hr_mr(ibmr); 311 int ret; 312 313 free_mr_pbl(hr_dev, mr); 314 ret = alloc_mr_pbl(hr_dev, mr, length, udata, start, mr_access_flags); 315 if (ret) { 316 ibdev_err(ibdev, "failed to create mr PBL, ret = %d.\n", ret); 317 return ret; 318 } 319 320 ret = hr_dev->hw->rereg_write_mtpt(hr_dev, mr, flags, pdn, 321 mr_access_flags, virt_addr, 322 length, mailbox->buf); 323 if (ret) { 324 ibdev_err(ibdev, "failed to write mtpt, ret = %d.\n", ret); 325 free_mr_pbl(hr_dev, mr); 326 } 327 328 return ret; 329 } 330 331 struct ib_mr *hns_roce_rereg_user_mr(struct ib_mr *ibmr, int flags, u64 start, 332 u64 length, u64 virt_addr, 333 int mr_access_flags, struct ib_pd *pd, 334 struct ib_udata *udata) 335 { 336 struct hns_roce_dev *hr_dev = to_hr_dev(ibmr->device); 337 struct ib_device *ib_dev = &hr_dev->ib_dev; 338 struct hns_roce_mr *mr = to_hr_mr(ibmr); 339 struct hns_roce_cmd_mailbox *mailbox; 340 unsigned long mtpt_idx; 341 u32 pdn = 0; 342 int ret; 343 344 if (!mr->enabled) 345 return ERR_PTR(-EINVAL); 346 347 mailbox = hns_roce_alloc_cmd_mailbox(hr_dev); 348 if (IS_ERR(mailbox)) 349 return ERR_CAST(mailbox); 350 351 mtpt_idx = key_to_hw_index(mr->key) & (hr_dev->caps.num_mtpts - 1); 352 ret = hns_roce_cmd_mbox(hr_dev, 0, mailbox->dma, mtpt_idx, 0, 353 HNS_ROCE_CMD_QUERY_MPT, 354 HNS_ROCE_CMD_TIMEOUT_MSECS); 355 if (ret) 356 goto free_cmd_mbox; 357 358 ret = hns_roce_hw_destroy_mpt(hr_dev, NULL, mtpt_idx); 359 if (ret) 360 ibdev_warn(ib_dev, "failed to destroy MPT, ret = %d.\n", ret); 361 362 mr->enabled = 0; 363 364 if (flags & IB_MR_REREG_PD) 365 pdn = to_hr_pd(pd)->pdn; 366 367 if (flags & IB_MR_REREG_TRANS) { 368 ret = rereg_mr_trans(ibmr, flags, 369 start, length, 370 virt_addr, mr_access_flags, 371 mailbox, pdn, udata); 372 if (ret) 373 goto free_cmd_mbox; 374 } else { 375 ret = hr_dev->hw->rereg_write_mtpt(hr_dev, mr, flags, pdn, 376 mr_access_flags, virt_addr, 377 length, mailbox->buf); 378 if (ret) 379 goto free_cmd_mbox; 380 } 381 382 ret = hns_roce_hw_create_mpt(hr_dev, mailbox, mtpt_idx); 383 if (ret) { 384 ibdev_err(ib_dev, "failed to create MPT, ret = %d.\n", ret); 385 goto free_cmd_mbox; 386 } 387 388 mr->enabled = 1; 389 if (flags & IB_MR_REREG_ACCESS) 390 mr->access = mr_access_flags; 391 392 hns_roce_free_cmd_mailbox(hr_dev, mailbox); 393 394 return NULL; 395 396 free_cmd_mbox: 397 hns_roce_free_cmd_mailbox(hr_dev, mailbox); 398 399 return ERR_PTR(ret); 400 } 401 402 int hns_roce_dereg_mr(struct ib_mr *ibmr, struct ib_udata *udata) 403 { 404 struct hns_roce_dev *hr_dev = to_hr_dev(ibmr->device); 405 struct hns_roce_mr *mr = to_hr_mr(ibmr); 406 int ret = 0; 407 408 if (hr_dev->hw->dereg_mr) { 409 ret = hr_dev->hw->dereg_mr(hr_dev, mr, udata); 410 } else { 411 hns_roce_mr_free(hr_dev, mr); 412 kfree(mr); 413 } 414 415 return ret; 416 } 417 418 struct ib_mr *hns_roce_alloc_mr(struct ib_pd *pd, enum ib_mr_type mr_type, 419 u32 max_num_sg) 420 { 421 struct hns_roce_dev *hr_dev = to_hr_dev(pd->device); 422 struct device *dev = hr_dev->dev; 423 struct hns_roce_mr *mr; 424 u64 length; 425 int ret; 426 427 if (mr_type != IB_MR_TYPE_MEM_REG) 428 return ERR_PTR(-EINVAL); 429 430 if (max_num_sg > HNS_ROCE_FRMR_MAX_PA) { 431 dev_err(dev, "max_num_sg larger than %d\n", 432 HNS_ROCE_FRMR_MAX_PA); 433 return ERR_PTR(-EINVAL); 434 } 435 436 mr = kzalloc(sizeof(*mr), GFP_KERNEL); 437 if (!mr) 438 return ERR_PTR(-ENOMEM); 439 440 mr->type = MR_TYPE_FRMR; 441 442 /* Allocate memory region key */ 443 length = max_num_sg * (1 << PAGE_SHIFT); 444 ret = alloc_mr_key(hr_dev, mr, to_hr_pd(pd)->pdn, 0, length, 0); 445 if (ret) 446 goto err_free; 447 448 ret = alloc_mr_pbl(hr_dev, mr, length, NULL, 0, 0); 449 if (ret) 450 goto err_key; 451 452 ret = hns_roce_mr_enable(hr_dev, mr); 453 if (ret) 454 goto err_pbl; 455 456 mr->ibmr.rkey = mr->ibmr.lkey = mr->key; 457 mr->ibmr.length = length; 458 459 return &mr->ibmr; 460 461 err_key: 462 free_mr_key(hr_dev, mr); 463 err_pbl: 464 free_mr_pbl(hr_dev, mr); 465 err_free: 466 kfree(mr); 467 return ERR_PTR(ret); 468 } 469 470 static int hns_roce_set_page(struct ib_mr *ibmr, u64 addr) 471 { 472 struct hns_roce_mr *mr = to_hr_mr(ibmr); 473 474 if (likely(mr->npages < mr->pbl_mtr.hem_cfg.buf_pg_count)) { 475 mr->page_list[mr->npages++] = addr; 476 return 0; 477 } 478 479 return -ENOBUFS; 480 } 481 482 int hns_roce_map_mr_sg(struct ib_mr *ibmr, struct scatterlist *sg, int sg_nents, 483 unsigned int *sg_offset) 484 { 485 struct hns_roce_dev *hr_dev = to_hr_dev(ibmr->device); 486 struct ib_device *ibdev = &hr_dev->ib_dev; 487 struct hns_roce_mr *mr = to_hr_mr(ibmr); 488 struct hns_roce_mtr *mtr = &mr->pbl_mtr; 489 int ret = 0; 490 491 mr->npages = 0; 492 mr->page_list = kvcalloc(mr->pbl_mtr.hem_cfg.buf_pg_count, 493 sizeof(dma_addr_t), GFP_KERNEL); 494 if (!mr->page_list) 495 return ret; 496 497 ret = ib_sg_to_pages(ibmr, sg, sg_nents, sg_offset, hns_roce_set_page); 498 if (ret < 1) { 499 ibdev_err(ibdev, "failed to store sg pages %u %u, cnt = %d.\n", 500 mr->npages, mr->pbl_mtr.hem_cfg.buf_pg_count, ret); 501 goto err_page_list; 502 } 503 504 mtr->hem_cfg.region[0].offset = 0; 505 mtr->hem_cfg.region[0].count = mr->npages; 506 mtr->hem_cfg.region[0].hopnum = mr->pbl_hop_num; 507 mtr->hem_cfg.region_count = 1; 508 ret = hns_roce_mtr_map(hr_dev, mtr, mr->page_list, mr->npages); 509 if (ret) { 510 ibdev_err(ibdev, "failed to map sg mtr, ret = %d.\n", ret); 511 ret = 0; 512 } else { 513 mr->pbl_mtr.hem_cfg.buf_pg_shift = (u32)ilog2(ibmr->page_size); 514 ret = mr->npages; 515 } 516 517 err_page_list: 518 kvfree(mr->page_list); 519 mr->page_list = NULL; 520 521 return ret; 522 } 523 524 static void hns_roce_mw_free(struct hns_roce_dev *hr_dev, 525 struct hns_roce_mw *mw) 526 { 527 struct device *dev = hr_dev->dev; 528 int ret; 529 530 if (mw->enabled) { 531 ret = hns_roce_hw_destroy_mpt(hr_dev, NULL, 532 key_to_hw_index(mw->rkey) & 533 (hr_dev->caps.num_mtpts - 1)); 534 if (ret) 535 dev_warn(dev, "MW DESTROY_MPT failed (%d)\n", ret); 536 537 hns_roce_table_put(hr_dev, &hr_dev->mr_table.mtpt_table, 538 key_to_hw_index(mw->rkey)); 539 } 540 541 hns_roce_bitmap_free(&hr_dev->mr_table.mtpt_bitmap, 542 key_to_hw_index(mw->rkey), BITMAP_NO_RR); 543 } 544 545 static int hns_roce_mw_enable(struct hns_roce_dev *hr_dev, 546 struct hns_roce_mw *mw) 547 { 548 struct hns_roce_mr_table *mr_table = &hr_dev->mr_table; 549 struct hns_roce_cmd_mailbox *mailbox; 550 struct device *dev = hr_dev->dev; 551 unsigned long mtpt_idx = key_to_hw_index(mw->rkey); 552 int ret; 553 554 /* prepare HEM entry memory */ 555 ret = hns_roce_table_get(hr_dev, &mr_table->mtpt_table, mtpt_idx); 556 if (ret) 557 return ret; 558 559 mailbox = hns_roce_alloc_cmd_mailbox(hr_dev); 560 if (IS_ERR(mailbox)) { 561 ret = PTR_ERR(mailbox); 562 goto err_table; 563 } 564 565 ret = hr_dev->hw->mw_write_mtpt(mailbox->buf, mw); 566 if (ret) { 567 dev_err(dev, "MW write mtpt fail!\n"); 568 goto err_page; 569 } 570 571 ret = hns_roce_hw_create_mpt(hr_dev, mailbox, 572 mtpt_idx & (hr_dev->caps.num_mtpts - 1)); 573 if (ret) { 574 dev_err(dev, "MW CREATE_MPT failed (%d)\n", ret); 575 goto err_page; 576 } 577 578 mw->enabled = 1; 579 580 hns_roce_free_cmd_mailbox(hr_dev, mailbox); 581 582 return 0; 583 584 err_page: 585 hns_roce_free_cmd_mailbox(hr_dev, mailbox); 586 587 err_table: 588 hns_roce_table_put(hr_dev, &mr_table->mtpt_table, mtpt_idx); 589 590 return ret; 591 } 592 593 int hns_roce_alloc_mw(struct ib_mw *ibmw, struct ib_udata *udata) 594 { 595 struct hns_roce_dev *hr_dev = to_hr_dev(ibmw->device); 596 struct hns_roce_mw *mw = to_hr_mw(ibmw); 597 unsigned long index = 0; 598 int ret; 599 600 /* Allocate a key for mw from bitmap */ 601 ret = hns_roce_bitmap_alloc(&hr_dev->mr_table.mtpt_bitmap, &index); 602 if (ret) 603 return ret; 604 605 mw->rkey = hw_index_to_key(index); 606 607 ibmw->rkey = mw->rkey; 608 mw->pdn = to_hr_pd(ibmw->pd)->pdn; 609 mw->pbl_hop_num = hr_dev->caps.pbl_hop_num; 610 mw->pbl_ba_pg_sz = hr_dev->caps.pbl_ba_pg_sz; 611 mw->pbl_buf_pg_sz = hr_dev->caps.pbl_buf_pg_sz; 612 613 ret = hns_roce_mw_enable(hr_dev, mw); 614 if (ret) 615 goto err_mw; 616 617 return 0; 618 619 err_mw: 620 hns_roce_mw_free(hr_dev, mw); 621 return ret; 622 } 623 624 int hns_roce_dealloc_mw(struct ib_mw *ibmw) 625 { 626 struct hns_roce_dev *hr_dev = to_hr_dev(ibmw->device); 627 struct hns_roce_mw *mw = to_hr_mw(ibmw); 628 629 hns_roce_mw_free(hr_dev, mw); 630 return 0; 631 } 632 633 static int mtr_map_region(struct hns_roce_dev *hr_dev, struct hns_roce_mtr *mtr, 634 dma_addr_t *pages, struct hns_roce_buf_region *region) 635 { 636 __le64 *mtts; 637 int offset; 638 int count; 639 int npage; 640 u64 addr; 641 int end; 642 int i; 643 644 /* if hopnum is 0, buffer cannot store BAs, so skip write mtt */ 645 if (!region->hopnum) 646 return 0; 647 648 offset = region->offset; 649 end = offset + region->count; 650 npage = 0; 651 while (offset < end) { 652 mtts = hns_roce_hem_list_find_mtt(hr_dev, &mtr->hem_list, 653 offset, &count, NULL); 654 if (!mtts) 655 return -ENOBUFS; 656 657 for (i = 0; i < count; i++) { 658 if (hr_dev->hw_rev == HNS_ROCE_HW_VER1) 659 addr = to_hr_hw_page_addr(pages[npage]); 660 else 661 addr = pages[npage]; 662 663 mtts[i] = cpu_to_le64(addr); 664 npage++; 665 } 666 offset += count; 667 } 668 669 return 0; 670 } 671 672 static inline bool mtr_has_mtt(struct hns_roce_buf_attr *attr) 673 { 674 int i; 675 676 for (i = 0; i < attr->region_count; i++) 677 if (attr->region[i].hopnum != HNS_ROCE_HOP_NUM_0 && 678 attr->region[i].hopnum > 0) 679 return true; 680 681 /* because the mtr only one root base address, when hopnum is 0 means 682 * root base address equals the first buffer address, thus all alloced 683 * memory must in a continuous space accessed by direct mode. 684 */ 685 return false; 686 } 687 688 static inline size_t mtr_bufs_size(struct hns_roce_buf_attr *attr) 689 { 690 size_t size = 0; 691 int i; 692 693 for (i = 0; i < attr->region_count; i++) 694 size += attr->region[i].size; 695 696 return size; 697 } 698 699 /* 700 * check the given pages in continuous address space 701 * Returns 0 on success, or the error page num. 702 */ 703 static inline int mtr_check_direct_pages(dma_addr_t *pages, int page_count, 704 unsigned int page_shift) 705 { 706 size_t page_size = 1 << page_shift; 707 int i; 708 709 for (i = 1; i < page_count; i++) 710 if (pages[i] - pages[i - 1] != page_size) 711 return i; 712 713 return 0; 714 } 715 716 static void mtr_free_bufs(struct hns_roce_dev *hr_dev, struct hns_roce_mtr *mtr) 717 { 718 /* release user buffers */ 719 if (mtr->umem) { 720 ib_umem_release(mtr->umem); 721 mtr->umem = NULL; 722 } 723 724 /* release kernel buffers */ 725 if (mtr->kmem) { 726 hns_roce_buf_free(hr_dev, mtr->kmem); 727 mtr->kmem = NULL; 728 } 729 } 730 731 static int mtr_alloc_bufs(struct hns_roce_dev *hr_dev, struct hns_roce_mtr *mtr, 732 struct hns_roce_buf_attr *buf_attr, bool is_direct, 733 struct ib_udata *udata, unsigned long user_addr) 734 { 735 struct ib_device *ibdev = &hr_dev->ib_dev; 736 unsigned int best_pg_shift; 737 int all_pg_count = 0; 738 size_t total_size; 739 int ret; 740 741 total_size = mtr_bufs_size(buf_attr); 742 if (total_size < 1) { 743 ibdev_err(ibdev, "failed to check mtr size\n."); 744 return -EINVAL; 745 } 746 747 if (udata) { 748 unsigned long pgsz_bitmap; 749 unsigned long page_size; 750 751 mtr->kmem = NULL; 752 mtr->umem = ib_umem_get(ibdev, user_addr, total_size, 753 buf_attr->user_access); 754 if (IS_ERR_OR_NULL(mtr->umem)) { 755 ibdev_err(ibdev, "failed to get umem, ret = %ld.\n", 756 PTR_ERR(mtr->umem)); 757 return -ENOMEM; 758 } 759 if (buf_attr->fixed_page) 760 pgsz_bitmap = 1 << buf_attr->page_shift; 761 else 762 pgsz_bitmap = GENMASK(buf_attr->page_shift, PAGE_SHIFT); 763 764 page_size = ib_umem_find_best_pgsz(mtr->umem, pgsz_bitmap, 765 user_addr); 766 if (!page_size) 767 return -EINVAL; 768 best_pg_shift = order_base_2(page_size); 769 all_pg_count = ib_umem_num_dma_blocks(mtr->umem, page_size); 770 ret = 0; 771 } else { 772 mtr->umem = NULL; 773 mtr->kmem = 774 hns_roce_buf_alloc(hr_dev, total_size, 775 buf_attr->page_shift, 776 is_direct ? HNS_ROCE_BUF_DIRECT : 0); 777 if (IS_ERR(mtr->kmem)) { 778 ibdev_err(ibdev, "failed to alloc kmem, ret = %ld.\n", 779 PTR_ERR(mtr->kmem)); 780 return PTR_ERR(mtr->kmem); 781 } 782 783 best_pg_shift = buf_attr->page_shift; 784 all_pg_count = mtr->kmem->npages; 785 } 786 787 /* must bigger than minimum hardware page shift */ 788 if (best_pg_shift < HNS_HW_PAGE_SHIFT || all_pg_count < 1) { 789 ret = -EINVAL; 790 ibdev_err(ibdev, 791 "failed to check mtr, page shift = %u count = %d.\n", 792 best_pg_shift, all_pg_count); 793 goto err_alloc_mem; 794 } 795 796 mtr->hem_cfg.buf_pg_shift = best_pg_shift; 797 mtr->hem_cfg.buf_pg_count = all_pg_count; 798 799 return 0; 800 err_alloc_mem: 801 mtr_free_bufs(hr_dev, mtr); 802 return ret; 803 } 804 805 static int mtr_get_pages(struct hns_roce_dev *hr_dev, struct hns_roce_mtr *mtr, 806 dma_addr_t *pages, int count, unsigned int page_shift) 807 { 808 struct ib_device *ibdev = &hr_dev->ib_dev; 809 int npage; 810 int err; 811 812 if (mtr->umem) 813 npage = hns_roce_get_umem_bufs(hr_dev, pages, count, 0, 814 mtr->umem, page_shift); 815 else 816 npage = hns_roce_get_kmem_bufs(hr_dev, pages, count, 0, 817 mtr->kmem); 818 819 if (mtr->hem_cfg.is_direct && npage > 1) { 820 err = mtr_check_direct_pages(pages, npage, page_shift); 821 if (err) { 822 ibdev_err(ibdev, "Failed to check %s direct page-%d\n", 823 mtr->umem ? "user" : "kernel", err); 824 npage = err; 825 } 826 } 827 828 return npage; 829 } 830 831 int hns_roce_mtr_map(struct hns_roce_dev *hr_dev, struct hns_roce_mtr *mtr, 832 dma_addr_t *pages, unsigned int page_cnt) 833 { 834 struct ib_device *ibdev = &hr_dev->ib_dev; 835 struct hns_roce_buf_region *r; 836 unsigned int i; 837 int err; 838 839 /* 840 * Only use the first page address as root ba when hopnum is 0, this 841 * is because the addresses of all pages are consecutive in this case. 842 */ 843 if (mtr->hem_cfg.is_direct) { 844 mtr->hem_cfg.root_ba = pages[0]; 845 return 0; 846 } 847 848 for (i = 0; i < mtr->hem_cfg.region_count; i++) { 849 r = &mtr->hem_cfg.region[i]; 850 if (r->offset + r->count > page_cnt) { 851 err = -EINVAL; 852 ibdev_err(ibdev, 853 "failed to check mtr%u end %u + %u, max %u.\n", 854 i, r->offset, r->count, page_cnt); 855 return err; 856 } 857 858 err = mtr_map_region(hr_dev, mtr, &pages[r->offset], r); 859 if (err) { 860 ibdev_err(ibdev, 861 "failed to map mtr%u offset %u, ret = %d.\n", 862 i, r->offset, err); 863 return err; 864 } 865 } 866 867 return 0; 868 } 869 870 int hns_roce_mtr_find(struct hns_roce_dev *hr_dev, struct hns_roce_mtr *mtr, 871 int offset, u64 *mtt_buf, int mtt_max, u64 *base_addr) 872 { 873 struct hns_roce_hem_cfg *cfg = &mtr->hem_cfg; 874 int mtt_count, left; 875 int start_index; 876 int total = 0; 877 __le64 *mtts; 878 u32 npage; 879 u64 addr; 880 881 if (!mtt_buf || mtt_max < 1) 882 goto done; 883 884 /* no mtt memory in direct mode, so just return the buffer address */ 885 if (cfg->is_direct) { 886 start_index = offset >> HNS_HW_PAGE_SHIFT; 887 for (mtt_count = 0; mtt_count < cfg->region_count && 888 total < mtt_max; mtt_count++) { 889 npage = cfg->region[mtt_count].offset; 890 if (npage < start_index) 891 continue; 892 893 addr = cfg->root_ba + (npage << HNS_HW_PAGE_SHIFT); 894 if (hr_dev->hw_rev == HNS_ROCE_HW_VER1) 895 mtt_buf[total] = to_hr_hw_page_addr(addr); 896 else 897 mtt_buf[total] = addr; 898 899 total++; 900 } 901 902 goto done; 903 } 904 905 start_index = offset >> cfg->buf_pg_shift; 906 left = mtt_max; 907 while (left > 0) { 908 mtt_count = 0; 909 mtts = hns_roce_hem_list_find_mtt(hr_dev, &mtr->hem_list, 910 start_index + total, 911 &mtt_count, NULL); 912 if (!mtts || !mtt_count) 913 goto done; 914 915 npage = min(mtt_count, left); 916 left -= npage; 917 for (mtt_count = 0; mtt_count < npage; mtt_count++) 918 mtt_buf[total++] = le64_to_cpu(mtts[mtt_count]); 919 } 920 921 done: 922 if (base_addr) 923 *base_addr = cfg->root_ba; 924 925 return total; 926 } 927 928 static int mtr_init_buf_cfg(struct hns_roce_dev *hr_dev, 929 struct hns_roce_buf_attr *attr, 930 struct hns_roce_hem_cfg *cfg, 931 unsigned int *buf_page_shift) 932 { 933 struct hns_roce_buf_region *r; 934 unsigned int page_shift; 935 int page_cnt = 0; 936 size_t buf_size; 937 int region_cnt; 938 939 if (cfg->is_direct) { 940 buf_size = cfg->buf_pg_count << cfg->buf_pg_shift; 941 page_cnt = DIV_ROUND_UP(buf_size, HNS_HW_PAGE_SIZE); 942 /* 943 * When HEM buffer use level-0 addressing, the page size equals 944 * the buffer size, and the the page size = 4K * 2^N. 945 */ 946 cfg->buf_pg_shift = HNS_HW_PAGE_SHIFT + order_base_2(page_cnt); 947 if (attr->region_count > 1) { 948 cfg->buf_pg_count = page_cnt; 949 page_shift = HNS_HW_PAGE_SHIFT; 950 } else { 951 cfg->buf_pg_count = 1; 952 page_shift = cfg->buf_pg_shift; 953 if (buf_size != 1 << page_shift) { 954 ibdev_err(&hr_dev->ib_dev, 955 "failed to check direct size %zu shift %d.\n", 956 buf_size, page_shift); 957 return -EINVAL; 958 } 959 } 960 } else { 961 page_shift = cfg->buf_pg_shift; 962 } 963 964 /* convert buffer size to page index and page count */ 965 for (page_cnt = 0, region_cnt = 0; page_cnt < cfg->buf_pg_count && 966 region_cnt < attr->region_count && 967 region_cnt < ARRAY_SIZE(cfg->region); region_cnt++) { 968 r = &cfg->region[region_cnt]; 969 r->offset = page_cnt; 970 buf_size = hr_hw_page_align(attr->region[region_cnt].size); 971 r->count = DIV_ROUND_UP(buf_size, 1 << page_shift); 972 page_cnt += r->count; 973 r->hopnum = to_hr_hem_hopnum(attr->region[region_cnt].hopnum, 974 r->count); 975 } 976 977 if (region_cnt < 1) { 978 ibdev_err(&hr_dev->ib_dev, 979 "failed to check mtr region count, pages = %d.\n", 980 cfg->buf_pg_count); 981 return -ENOBUFS; 982 } 983 984 cfg->region_count = region_cnt; 985 *buf_page_shift = page_shift; 986 987 return page_cnt; 988 } 989 990 /** 991 * hns_roce_mtr_create - Create hns memory translate region. 992 * 993 * @mtr: memory translate region 994 * @buf_attr: buffer attribute for creating mtr 995 * @ba_page_shift: page shift for multi-hop base address table 996 * @udata: user space context, if it's NULL, means kernel space 997 * @user_addr: userspace virtual address to start at 998 */ 999 int hns_roce_mtr_create(struct hns_roce_dev *hr_dev, struct hns_roce_mtr *mtr, 1000 struct hns_roce_buf_attr *buf_attr, 1001 unsigned int ba_page_shift, struct ib_udata *udata, 1002 unsigned long user_addr) 1003 { 1004 struct hns_roce_hem_cfg *cfg = &mtr->hem_cfg; 1005 struct ib_device *ibdev = &hr_dev->ib_dev; 1006 unsigned int buf_page_shift = 0; 1007 dma_addr_t *pages = NULL; 1008 int all_pg_cnt; 1009 int get_pg_cnt; 1010 int ret = 0; 1011 1012 /* if disable mtt, all pages must in a continuous address range */ 1013 cfg->is_direct = !mtr_has_mtt(buf_attr); 1014 1015 /* if buffer only need mtt, just init the hem cfg */ 1016 if (buf_attr->mtt_only) { 1017 cfg->buf_pg_shift = buf_attr->page_shift; 1018 cfg->buf_pg_count = mtr_bufs_size(buf_attr) >> 1019 buf_attr->page_shift; 1020 mtr->umem = NULL; 1021 mtr->kmem = NULL; 1022 } else { 1023 ret = mtr_alloc_bufs(hr_dev, mtr, buf_attr, cfg->is_direct, 1024 udata, user_addr); 1025 if (ret) { 1026 ibdev_err(ibdev, 1027 "failed to alloc mtr bufs, ret = %d.\n", ret); 1028 return ret; 1029 } 1030 } 1031 1032 all_pg_cnt = mtr_init_buf_cfg(hr_dev, buf_attr, cfg, &buf_page_shift); 1033 if (all_pg_cnt < 1) { 1034 ret = -ENOBUFS; 1035 ibdev_err(ibdev, "failed to init mtr buf cfg.\n"); 1036 goto err_alloc_bufs; 1037 } 1038 1039 hns_roce_hem_list_init(&mtr->hem_list); 1040 if (!cfg->is_direct) { 1041 ret = hns_roce_hem_list_request(hr_dev, &mtr->hem_list, 1042 cfg->region, cfg->region_count, 1043 ba_page_shift); 1044 if (ret) { 1045 ibdev_err(ibdev, "failed to request mtr hem, ret = %d.\n", 1046 ret); 1047 goto err_alloc_bufs; 1048 } 1049 cfg->root_ba = mtr->hem_list.root_ba; 1050 cfg->ba_pg_shift = ba_page_shift; 1051 } else { 1052 cfg->ba_pg_shift = cfg->buf_pg_shift; 1053 } 1054 1055 /* no buffer to map */ 1056 if (buf_attr->mtt_only) 1057 return 0; 1058 1059 /* alloc a tmp array to store buffer's dma address */ 1060 pages = kvcalloc(all_pg_cnt, sizeof(dma_addr_t), GFP_KERNEL); 1061 if (!pages) { 1062 ret = -ENOMEM; 1063 ibdev_err(ibdev, "failed to alloc mtr page list %d.\n", 1064 all_pg_cnt); 1065 goto err_alloc_hem_list; 1066 } 1067 1068 get_pg_cnt = mtr_get_pages(hr_dev, mtr, pages, all_pg_cnt, 1069 buf_page_shift); 1070 if (get_pg_cnt != all_pg_cnt) { 1071 ibdev_err(ibdev, "failed to get mtr page %d != %d.\n", 1072 get_pg_cnt, all_pg_cnt); 1073 ret = -ENOBUFS; 1074 goto err_alloc_page_list; 1075 } 1076 1077 /* write buffer's dma address to BA table */ 1078 ret = hns_roce_mtr_map(hr_dev, mtr, pages, all_pg_cnt); 1079 if (ret) { 1080 ibdev_err(ibdev, "failed to map mtr pages, ret = %d.\n", ret); 1081 goto err_alloc_page_list; 1082 } 1083 1084 /* drop tmp array */ 1085 kvfree(pages); 1086 return 0; 1087 err_alloc_page_list: 1088 kvfree(pages); 1089 err_alloc_hem_list: 1090 hns_roce_hem_list_release(hr_dev, &mtr->hem_list); 1091 err_alloc_bufs: 1092 mtr_free_bufs(hr_dev, mtr); 1093 return ret; 1094 } 1095 1096 void hns_roce_mtr_destroy(struct hns_roce_dev *hr_dev, struct hns_roce_mtr *mtr) 1097 { 1098 /* release multi-hop addressing resource */ 1099 hns_roce_hem_list_release(hr_dev, &mtr->hem_list); 1100 1101 /* free buffers */ 1102 mtr_free_bufs(hr_dev, mtr); 1103 } 1104