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 41static u32 hw_index_to_key(unsigned long ind) 42{ 43 return (u32)(ind >> 24) | (ind << 8); 44} 45 46unsigned long key_to_hw_index(u32 key) 47{ 48 return (key << 24) | (key >> 8); 49} 50 51static 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 60int 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 69static 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; 99err_free_bitmap: 100 hns_roce_bitmap_free(&hr_dev->mr_table.mtpt_bitmap, obj, BITMAP_NO_RR); 101 return err; 102} 103 104static 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 112static 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 143static 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 148static 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 167static 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 204err_page: 205 hns_roce_free_cmd_mailbox(hr_dev, mailbox); 206 207 return ret; 208} 209 210int 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 222void 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 229struct 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; 254err_mr: 255 free_mr_key(hr_dev, mr); 256 257err_free: 258 kfree(mr); 259 return ERR_PTR(ret); 260} 261 262struct 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; 289 290 return &mr->ibmr; 291 292err_alloc_pbl: 293 free_mr_pbl(hr_dev, mr); 294err_alloc_key: 295 free_mr_key(hr_dev, mr); 296err_alloc_mr: 297 kfree(mr); 298 return ERR_PTR(ret); 299} 300 301static int rereg_mr_trans(struct ib_mr *ibmr, int flags, 302 u64 start, u64 length, 303 u64 virt_addr, int mr_access_flags, 304 struct hns_roce_cmd_mailbox *mailbox, 305 u32 pdn, struct ib_udata *udata) 306{ 307 struct hns_roce_dev *hr_dev = to_hr_dev(ibmr->device); 308 struct ib_device *ibdev = &hr_dev->ib_dev; 309 struct hns_roce_mr *mr = to_hr_mr(ibmr); 310 int ret; 311 312 free_mr_pbl(hr_dev, mr); 313 ret = alloc_mr_pbl(hr_dev, mr, length, udata, start, mr_access_flags); 314 if (ret) { 315 ibdev_err(ibdev, "failed to create mr PBL, ret = %d.\n", ret); 316 return ret; 317 } 318 319 ret = hr_dev->hw->rereg_write_mtpt(hr_dev, mr, flags, pdn, 320 mr_access_flags, virt_addr, 321 length, mailbox->buf); 322 if (ret) { 323 ibdev_err(ibdev, "failed to write mtpt, ret = %d.\n", ret); 324 free_mr_pbl(hr_dev, mr); 325 } 326 327 return ret; 328} 329 330int hns_roce_rereg_user_mr(struct ib_mr *ibmr, int flags, u64 start, u64 length, 331 u64 virt_addr, int mr_access_flags, struct ib_pd *pd, 332 struct ib_udata *udata) 333{ 334 struct hns_roce_dev *hr_dev = to_hr_dev(ibmr->device); 335 struct ib_device *ib_dev = &hr_dev->ib_dev; 336 struct hns_roce_mr *mr = to_hr_mr(ibmr); 337 struct hns_roce_cmd_mailbox *mailbox; 338 unsigned long mtpt_idx; 339 u32 pdn = 0; 340 int ret; 341 342 if (!mr->enabled) 343 return -EINVAL; 344 345 mailbox = hns_roce_alloc_cmd_mailbox(hr_dev); 346 if (IS_ERR(mailbox)) 347 return PTR_ERR(mailbox); 348 349 mtpt_idx = key_to_hw_index(mr->key) & (hr_dev->caps.num_mtpts - 1); 350 ret = hns_roce_cmd_mbox(hr_dev, 0, mailbox->dma, mtpt_idx, 0, 351 HNS_ROCE_CMD_QUERY_MPT, 352 HNS_ROCE_CMD_TIMEOUT_MSECS); 353 if (ret) 354 goto free_cmd_mbox; 355 356 ret = hns_roce_hw_destroy_mpt(hr_dev, NULL, mtpt_idx); 357 if (ret) 358 ibdev_warn(ib_dev, "failed to destroy MPT, ret = %d.\n", ret); 359 360 mr->enabled = 0; 361 362 if (flags & IB_MR_REREG_PD) 363 pdn = to_hr_pd(pd)->pdn; 364 365 if (flags & IB_MR_REREG_TRANS) { 366 ret = rereg_mr_trans(ibmr, flags, 367 start, length, 368 virt_addr, mr_access_flags, 369 mailbox, pdn, udata); 370 if (ret) 371 goto free_cmd_mbox; 372 } else { 373 ret = hr_dev->hw->rereg_write_mtpt(hr_dev, mr, flags, pdn, 374 mr_access_flags, virt_addr, 375 length, mailbox->buf); 376 if (ret) 377 goto free_cmd_mbox; 378 } 379 380 ret = hns_roce_hw_create_mpt(hr_dev, mailbox, mtpt_idx); 381 if (ret) { 382 ibdev_err(ib_dev, "failed to create MPT, ret = %d.\n", ret); 383 goto free_cmd_mbox; 384 } 385 386 mr->enabled = 1; 387 if (flags & IB_MR_REREG_ACCESS) 388 mr->access = mr_access_flags; 389 390 hns_roce_free_cmd_mailbox(hr_dev, mailbox); 391 392 return 0; 393 394free_cmd_mbox: 395 hns_roce_free_cmd_mailbox(hr_dev, mailbox); 396 397 return ret; 398} 399 400int hns_roce_dereg_mr(struct ib_mr *ibmr, struct ib_udata *udata) 401{ 402 struct hns_roce_dev *hr_dev = to_hr_dev(ibmr->device); 403 struct hns_roce_mr *mr = to_hr_mr(ibmr); 404 int ret = 0; 405 406 if (hr_dev->hw->dereg_mr) { 407 ret = hr_dev->hw->dereg_mr(hr_dev, mr, udata); 408 } else { 409 hns_roce_mr_free(hr_dev, mr); 410 kfree(mr); 411 } 412 413 return ret; 414} 415 416struct ib_mr *hns_roce_alloc_mr(struct ib_pd *pd, enum ib_mr_type mr_type, 417 u32 max_num_sg) 418{ 419 struct hns_roce_dev *hr_dev = to_hr_dev(pd->device); 420 struct device *dev = hr_dev->dev; 421 struct hns_roce_mr *mr; 422 u64 length; 423 int ret; 424 425 if (mr_type != IB_MR_TYPE_MEM_REG) 426 return ERR_PTR(-EINVAL); 427 428 if (max_num_sg > HNS_ROCE_FRMR_MAX_PA) { 429 dev_err(dev, "max_num_sg larger than %d\n", 430 HNS_ROCE_FRMR_MAX_PA); 431 return ERR_PTR(-EINVAL); 432 } 433 434 mr = kzalloc(sizeof(*mr), GFP_KERNEL); 435 if (!mr) 436 return ERR_PTR(-ENOMEM); 437 438 mr->type = MR_TYPE_FRMR; 439 440 /* Allocate memory region key */ 441 length = max_num_sg * (1 << PAGE_SHIFT); 442 ret = alloc_mr_key(hr_dev, mr, to_hr_pd(pd)->pdn, 0, length, 0); 443 if (ret) 444 goto err_free; 445 446 ret = alloc_mr_pbl(hr_dev, mr, length, NULL, 0, 0); 447 if (ret) 448 goto err_key; 449 450 ret = hns_roce_mr_enable(hr_dev, mr); 451 if (ret) 452 goto err_pbl; 453 454 mr->ibmr.rkey = mr->ibmr.lkey = mr->key; 455 mr->ibmr.length = length; 456 457 return &mr->ibmr; 458 459err_pbl: 460 free_mr_pbl(hr_dev, mr); 461err_key: 462 free_mr_key(hr_dev, mr); 463err_free: 464 kfree(mr); 465 return ERR_PTR(ret); 466} 467 468static int hns_roce_set_page(struct ib_mr *ibmr, u64 addr) 469{ 470 struct hns_roce_mr *mr = to_hr_mr(ibmr); 471 472 if (likely(mr->npages < mr->pbl_mtr.hem_cfg.buf_pg_count)) { 473 mr->page_list[mr->npages++] = addr; 474 return 0; 475 } 476 477 return -ENOBUFS; 478} 479 480int hns_roce_map_mr_sg(struct ib_mr *ibmr, struct scatterlist *sg, int sg_nents, 481 unsigned int *sg_offset) 482{ 483 struct hns_roce_dev *hr_dev = to_hr_dev(ibmr->device); 484 struct ib_device *ibdev = &hr_dev->ib_dev; 485 struct hns_roce_mr *mr = to_hr_mr(ibmr); 486 struct hns_roce_mtr *mtr = &mr->pbl_mtr; 487 int ret = 0; 488 489 mr->npages = 0; 490 mr->page_list = kvcalloc(mr->pbl_mtr.hem_cfg.buf_pg_count, 491 sizeof(dma_addr_t), GFP_KERNEL); 492 if (!mr->page_list) 493 return ret; 494 495 ret = ib_sg_to_pages(ibmr, sg, sg_nents, sg_offset, hns_roce_set_page); 496 if (ret < 1) { 497 ibdev_err(ibdev, "failed to store sg pages %u %u, cnt = %d.\n", 498 mr->npages, mr->pbl_mtr.hem_cfg.buf_pg_count, ret); 499 goto err_page_list; 500 } 501 502 mtr->hem_cfg.region[0].offset = 0; 503 mtr->hem_cfg.region[0].count = mr->npages; 504 mtr->hem_cfg.region[0].hopnum = mr->pbl_hop_num; 505 mtr->hem_cfg.region_count = 1; 506 ret = hns_roce_mtr_map(hr_dev, mtr, mr->page_list, mr->npages); 507 if (ret) { 508 ibdev_err(ibdev, "failed to map sg mtr, ret = %d.\n", ret); 509 ret = 0; 510 } else { 511 mr->pbl_mtr.hem_cfg.buf_pg_shift = ilog2(ibmr->page_size); 512 ret = mr->npages; 513 } 514 515err_page_list: 516 kvfree(mr->page_list); 517 mr->page_list = NULL; 518 519 return ret; 520} 521 522static void hns_roce_mw_free(struct hns_roce_dev *hr_dev, 523 struct hns_roce_mw *mw) 524{ 525 struct device *dev = hr_dev->dev; 526 int ret; 527 528 if (mw->enabled) { 529 ret = hns_roce_hw_destroy_mpt(hr_dev, NULL, 530 key_to_hw_index(mw->rkey) & 531 (hr_dev->caps.num_mtpts - 1)); 532 if (ret) 533 dev_warn(dev, "MW DESTROY_MPT failed (%d)\n", ret); 534 535 hns_roce_table_put(hr_dev, &hr_dev->mr_table.mtpt_table, 536 key_to_hw_index(mw->rkey)); 537 } 538 539 hns_roce_bitmap_free(&hr_dev->mr_table.mtpt_bitmap, 540 key_to_hw_index(mw->rkey), BITMAP_NO_RR); 541} 542 543static int hns_roce_mw_enable(struct hns_roce_dev *hr_dev, 544 struct hns_roce_mw *mw) 545{ 546 struct hns_roce_mr_table *mr_table = &hr_dev->mr_table; 547 struct hns_roce_cmd_mailbox *mailbox; 548 struct device *dev = hr_dev->dev; 549 unsigned long mtpt_idx = key_to_hw_index(mw->rkey); 550 int ret; 551 552 /* prepare HEM entry memory */ 553 ret = hns_roce_table_get(hr_dev, &mr_table->mtpt_table, mtpt_idx); 554 if (ret) 555 return ret; 556 557 mailbox = hns_roce_alloc_cmd_mailbox(hr_dev); 558 if (IS_ERR(mailbox)) { 559 ret = PTR_ERR(mailbox); 560 goto err_table; 561 } 562 563 ret = hr_dev->hw->mw_write_mtpt(mailbox->buf, mw); 564 if (ret) { 565 dev_err(dev, "MW write mtpt fail!\n"); 566 goto err_page; 567 } 568 569 ret = hns_roce_hw_create_mpt(hr_dev, mailbox, 570 mtpt_idx & (hr_dev->caps.num_mtpts - 1)); 571 if (ret) { 572 dev_err(dev, "MW CREATE_MPT failed (%d)\n", ret); 573 goto err_page; 574 } 575 576 mw->enabled = 1; 577 578 hns_roce_free_cmd_mailbox(hr_dev, mailbox); 579 580 return 0; 581 582err_page: 583 hns_roce_free_cmd_mailbox(hr_dev, mailbox); 584 585err_table: 586 hns_roce_table_put(hr_dev, &mr_table->mtpt_table, mtpt_idx); 587 588 return ret; 589} 590 591int hns_roce_alloc_mw(struct ib_mw *ibmw, struct ib_udata *udata) 592{ 593 struct hns_roce_dev *hr_dev = to_hr_dev(ibmw->device); 594 struct hns_roce_mw *mw = to_hr_mw(ibmw); 595 unsigned long index = 0; 596 int ret; 597 598 /* Allocate a key for mw from bitmap */ 599 ret = hns_roce_bitmap_alloc(&hr_dev->mr_table.mtpt_bitmap, &index); 600 if (ret) 601 return ret; 602 603 mw->rkey = hw_index_to_key(index); 604 605 ibmw->rkey = mw->rkey; 606 mw->pdn = to_hr_pd(ibmw->pd)->pdn; 607 mw->pbl_hop_num = hr_dev->caps.pbl_hop_num; 608 mw->pbl_ba_pg_sz = hr_dev->caps.pbl_ba_pg_sz; 609 mw->pbl_buf_pg_sz = hr_dev->caps.pbl_buf_pg_sz; 610 611 ret = hns_roce_mw_enable(hr_dev, mw); 612 if (ret) 613 goto err_mw; 614 615 return 0; 616 617err_mw: 618 hns_roce_mw_free(hr_dev, mw); 619 return ret; 620} 621 622int hns_roce_dealloc_mw(struct ib_mw *ibmw) 623{ 624 struct hns_roce_dev *hr_dev = to_hr_dev(ibmw->device); 625 struct hns_roce_mw *mw = to_hr_mw(ibmw); 626 627 hns_roce_mw_free(hr_dev, mw); 628 return 0; 629} 630 631static int mtr_map_region(struct hns_roce_dev *hr_dev, struct hns_roce_mtr *mtr, 632 dma_addr_t *pages, struct hns_roce_buf_region *region) 633{ 634 __le64 *mtts; 635 int offset; 636 int count; 637 int npage; 638 u64 addr; 639 int end; 640 int i; 641 642 /* if hopnum is 0, buffer cannot store BAs, so skip write mtt */ 643 if (!region->hopnum) 644 return 0; 645 646 offset = region->offset; 647 end = offset + region->count; 648 npage = 0; 649 while (offset < end) { 650 mtts = hns_roce_hem_list_find_mtt(hr_dev, &mtr->hem_list, 651 offset, &count, NULL); 652 if (!mtts) 653 return -ENOBUFS; 654 655 for (i = 0; i < count; i++) { 656 if (hr_dev->hw_rev == HNS_ROCE_HW_VER1) 657 addr = to_hr_hw_page_addr(pages[npage]); 658 else 659 addr = pages[npage]; 660 661 mtts[i] = cpu_to_le64(addr); 662 npage++; 663 } 664 offset += count; 665 } 666 667 return 0; 668} 669 670static inline bool mtr_has_mtt(struct hns_roce_buf_attr *attr) 671{ 672 int i; 673 674 for (i = 0; i < attr->region_count; i++) 675 if (attr->region[i].hopnum != HNS_ROCE_HOP_NUM_0 && 676 attr->region[i].hopnum > 0) 677 return true; 678 679 /* because the mtr only one root base address, when hopnum is 0 means 680 * root base address equals the first buffer address, thus all alloced 681 * memory must in a continuous space accessed by direct mode. 682 */ 683 return false; 684} 685 686static inline size_t mtr_bufs_size(struct hns_roce_buf_attr *attr) 687{ 688 size_t size = 0; 689 int i; 690 691 for (i = 0; i < attr->region_count; i++) 692 size += attr->region[i].size; 693 694 return size; 695} 696 697static inline size_t mtr_kmem_direct_size(bool is_direct, size_t alloc_size, 698 unsigned int page_shift) 699{ 700 if (is_direct) 701 return ALIGN(alloc_size, 1 << page_shift); 702 else 703 return HNS_HW_DIRECT_PAGE_COUNT << page_shift; 704} 705 706/* 707 * check the given pages in continuous address space 708 * Returns 0 on success, or the error page num. 709 */ 710static inline int mtr_check_direct_pages(dma_addr_t *pages, int page_count, 711 unsigned int page_shift) 712{ 713 size_t page_size = 1 << page_shift; 714 int i; 715 716 for (i = 1; i < page_count; i++) 717 if (pages[i] - pages[i - 1] != page_size) 718 return i; 719 720 return 0; 721} 722 723static void mtr_free_bufs(struct hns_roce_dev *hr_dev, struct hns_roce_mtr *mtr) 724{ 725 /* release user buffers */ 726 if (mtr->umem) { 727 ib_umem_release(mtr->umem); 728 mtr->umem = NULL; 729 } 730 731 /* release kernel buffers */ 732 if (mtr->kmem) { 733 hns_roce_buf_free(hr_dev, mtr->kmem); 734 kfree(mtr->kmem); 735 mtr->kmem = NULL; 736 } 737} 738 739static int mtr_alloc_bufs(struct hns_roce_dev *hr_dev, struct hns_roce_mtr *mtr, 740 struct hns_roce_buf_attr *buf_attr, bool is_direct, 741 struct ib_udata *udata, unsigned long user_addr) 742{ 743 struct ib_device *ibdev = &hr_dev->ib_dev; 744 unsigned int best_pg_shift; 745 int all_pg_count = 0; 746 size_t direct_size; 747 size_t total_size; 748 int ret; 749 750 total_size = mtr_bufs_size(buf_attr); 751 if (total_size < 1) { 752 ibdev_err(ibdev, "Failed to check mtr size\n"); 753 return -EINVAL; 754 } 755 756 if (udata) { 757 unsigned long pgsz_bitmap; 758 unsigned long page_size; 759 760 mtr->kmem = NULL; 761 mtr->umem = ib_umem_get(ibdev, user_addr, total_size, 762 buf_attr->user_access); 763 if (IS_ERR_OR_NULL(mtr->umem)) { 764 ibdev_err(ibdev, "Failed to get umem, ret %ld\n", 765 PTR_ERR(mtr->umem)); 766 return -ENOMEM; 767 } 768 if (buf_attr->fixed_page) 769 pgsz_bitmap = 1 << buf_attr->page_shift; 770 else 771 pgsz_bitmap = GENMASK(buf_attr->page_shift, PAGE_SHIFT); 772 773 page_size = ib_umem_find_best_pgsz(mtr->umem, pgsz_bitmap, 774 user_addr); 775 if (!page_size) 776 return -EINVAL; 777 best_pg_shift = order_base_2(page_size); 778 all_pg_count = ib_umem_num_dma_blocks(mtr->umem, page_size); 779 ret = 0; 780 } else { 781 mtr->umem = NULL; 782 mtr->kmem = kzalloc(sizeof(*mtr->kmem), GFP_KERNEL); 783 if (!mtr->kmem) { 784 ibdev_err(ibdev, "Failed to alloc kmem\n"); 785 return -ENOMEM; 786 } 787 direct_size = mtr_kmem_direct_size(is_direct, total_size, 788 buf_attr->page_shift); 789 ret = hns_roce_buf_alloc(hr_dev, total_size, direct_size, 790 mtr->kmem, buf_attr->page_shift); 791 if (ret) { 792 ibdev_err(ibdev, "Failed to alloc kmem, ret %d\n", ret); 793 goto err_alloc_mem; 794 } 795 best_pg_shift = buf_attr->page_shift; 796 all_pg_count = mtr->kmem->npages; 797 } 798 799 /* must bigger than minimum hardware page shift */ 800 if (best_pg_shift < HNS_HW_PAGE_SHIFT || all_pg_count < 1) { 801 ret = -EINVAL; 802 ibdev_err(ibdev, "Failed to check mtr page shift %d count %d\n", 803 best_pg_shift, all_pg_count); 804 goto err_alloc_mem; 805 } 806 807 mtr->hem_cfg.buf_pg_shift = best_pg_shift; 808 mtr->hem_cfg.buf_pg_count = all_pg_count; 809 810 return 0; 811err_alloc_mem: 812 mtr_free_bufs(hr_dev, mtr); 813 return ret; 814} 815 816static int mtr_get_pages(struct hns_roce_dev *hr_dev, struct hns_roce_mtr *mtr, 817 dma_addr_t *pages, int count, unsigned int page_shift) 818{ 819 struct ib_device *ibdev = &hr_dev->ib_dev; 820 int npage; 821 int err; 822 823 if (mtr->umem) 824 npage = hns_roce_get_umem_bufs(hr_dev, pages, count, 0, 825 mtr->umem, page_shift); 826 else 827 npage = hns_roce_get_kmem_bufs(hr_dev, pages, count, 0, 828 mtr->kmem); 829 830 if (mtr->hem_cfg.is_direct && npage > 1) { 831 err = mtr_check_direct_pages(pages, npage, page_shift); 832 if (err) { 833 ibdev_err(ibdev, "Failed to check %s direct page-%d\n", 834 mtr->umem ? "user" : "kernel", err); 835 npage = err; 836 } 837 } 838 839 return npage; 840} 841 842int hns_roce_mtr_map(struct hns_roce_dev *hr_dev, struct hns_roce_mtr *mtr, 843 dma_addr_t *pages, int page_cnt) 844{ 845 struct ib_device *ibdev = &hr_dev->ib_dev; 846 struct hns_roce_buf_region *r; 847 int err; 848 int i; 849 850 /* 851 * Only use the first page address as root ba when hopnum is 0, this 852 * is because the addresses of all pages are consecutive in this case. 853 */ 854 if (mtr->hem_cfg.is_direct) { 855 mtr->hem_cfg.root_ba = pages[0]; 856 return 0; 857 } 858 859 for (i = 0; i < mtr->hem_cfg.region_count; i++) { 860 r = &mtr->hem_cfg.region[i]; 861 if (r->offset + r->count > page_cnt) { 862 err = -EINVAL; 863 ibdev_err(ibdev, 864 "failed to check mtr%u end %u + %u, max %u.\n", 865 i, r->offset, r->count, page_cnt); 866 return err; 867 } 868 869 err = mtr_map_region(hr_dev, mtr, &pages[r->offset], r); 870 if (err) { 871 ibdev_err(ibdev, 872 "failed to map mtr%u offset %u, ret = %d.\n", 873 i, r->offset, err); 874 return err; 875 } 876 } 877 878 return 0; 879} 880 881int hns_roce_mtr_find(struct hns_roce_dev *hr_dev, struct hns_roce_mtr *mtr, 882 int offset, u64 *mtt_buf, int mtt_max, u64 *base_addr) 883{ 884 struct hns_roce_hem_cfg *cfg = &mtr->hem_cfg; 885 int start_index; 886 int mtt_count; 887 int total = 0; 888 __le64 *mtts; 889 int npage; 890 u64 addr; 891 int left; 892 893 if (!mtt_buf || mtt_max < 1) 894 goto done; 895 896 /* no mtt memory in direct mode, so just return the buffer address */ 897 if (cfg->is_direct) { 898 start_index = offset >> HNS_HW_PAGE_SHIFT; 899 for (mtt_count = 0; mtt_count < cfg->region_count && 900 total < mtt_max; mtt_count++) { 901 npage = cfg->region[mtt_count].offset; 902 if (npage < start_index) 903 continue; 904 905 addr = cfg->root_ba + (npage << HNS_HW_PAGE_SHIFT); 906 if (hr_dev->hw_rev == HNS_ROCE_HW_VER1) 907 mtt_buf[total] = to_hr_hw_page_addr(addr); 908 else 909 mtt_buf[total] = addr; 910 911 total++; 912 } 913 914 goto done; 915 } 916 917 start_index = offset >> cfg->buf_pg_shift; 918 left = mtt_max; 919 while (left > 0) { 920 mtt_count = 0; 921 mtts = hns_roce_hem_list_find_mtt(hr_dev, &mtr->hem_list, 922 start_index + total, 923 &mtt_count, NULL); 924 if (!mtts || !mtt_count) 925 goto done; 926 927 npage = min(mtt_count, left); 928 left -= npage; 929 for (mtt_count = 0; mtt_count < npage; mtt_count++) 930 mtt_buf[total++] = le64_to_cpu(mtts[mtt_count]); 931 } 932 933done: 934 if (base_addr) 935 *base_addr = cfg->root_ba; 936 937 return total; 938} 939 940static int mtr_init_buf_cfg(struct hns_roce_dev *hr_dev, 941 struct hns_roce_buf_attr *attr, 942 struct hns_roce_hem_cfg *cfg, 943 unsigned int *buf_page_shift) 944{ 945 struct hns_roce_buf_region *r; 946 unsigned int page_shift; 947 int page_cnt = 0; 948 size_t buf_size; 949 int region_cnt; 950 951 if (cfg->is_direct) { 952 buf_size = cfg->buf_pg_count << cfg->buf_pg_shift; 953 page_cnt = DIV_ROUND_UP(buf_size, HNS_HW_PAGE_SIZE); 954 /* 955 * When HEM buffer use level-0 addressing, the page size equals 956 * the buffer size, and the the page size = 4K * 2^N. 957 */ 958 cfg->buf_pg_shift = HNS_HW_PAGE_SHIFT + order_base_2(page_cnt); 959 if (attr->region_count > 1) { 960 cfg->buf_pg_count = page_cnt; 961 page_shift = HNS_HW_PAGE_SHIFT; 962 } else { 963 cfg->buf_pg_count = 1; 964 page_shift = cfg->buf_pg_shift; 965 if (buf_size != 1 << page_shift) { 966 ibdev_err(&hr_dev->ib_dev, 967 "failed to check direct size %zu shift %d.\n", 968 buf_size, page_shift); 969 return -EINVAL; 970 } 971 } 972 } else { 973 page_shift = cfg->buf_pg_shift; 974 } 975 976 /* convert buffer size to page index and page count */ 977 for (page_cnt = 0, region_cnt = 0; page_cnt < cfg->buf_pg_count && 978 region_cnt < attr->region_count && 979 region_cnt < ARRAY_SIZE(cfg->region); region_cnt++) { 980 r = &cfg->region[region_cnt]; 981 r->offset = page_cnt; 982 buf_size = hr_hw_page_align(attr->region[region_cnt].size); 983 r->count = DIV_ROUND_UP(buf_size, 1 << page_shift); 984 page_cnt += r->count; 985 r->hopnum = to_hr_hem_hopnum(attr->region[region_cnt].hopnum, 986 r->count); 987 } 988 989 if (region_cnt < 1) { 990 ibdev_err(&hr_dev->ib_dev, 991 "failed to check mtr region count, pages = %d.\n", 992 cfg->buf_pg_count); 993 return -ENOBUFS; 994 } 995 996 cfg->region_count = region_cnt; 997 *buf_page_shift = page_shift; 998 999 return page_cnt; 1000} 1001 1002/** 1003 * hns_roce_mtr_create - Create hns memory translate region. 1004 * 1005 * @mtr: memory translate region 1006 * @buf_attr: buffer attribute for creating mtr 1007 * @ba_page_shift: page shift for multi-hop base address table 1008 * @udata: user space context, if it's NULL, means kernel space 1009 * @user_addr: userspace virtual address to start at 1010 */ 1011int hns_roce_mtr_create(struct hns_roce_dev *hr_dev, struct hns_roce_mtr *mtr, 1012 struct hns_roce_buf_attr *buf_attr, 1013 unsigned int ba_page_shift, struct ib_udata *udata, 1014 unsigned long user_addr) 1015{ 1016 struct hns_roce_hem_cfg *cfg = &mtr->hem_cfg; 1017 struct ib_device *ibdev = &hr_dev->ib_dev; 1018 unsigned int buf_page_shift = 0; 1019 dma_addr_t *pages = NULL; 1020 int all_pg_cnt; 1021 int get_pg_cnt; 1022 int ret = 0; 1023 1024 /* if disable mtt, all pages must in a continuous address range */ 1025 cfg->is_direct = !mtr_has_mtt(buf_attr); 1026 1027 /* if buffer only need mtt, just init the hem cfg */ 1028 if (buf_attr->mtt_only) { 1029 cfg->buf_pg_shift = buf_attr->page_shift; 1030 cfg->buf_pg_count = mtr_bufs_size(buf_attr) >> 1031 buf_attr->page_shift; 1032 mtr->umem = NULL; 1033 mtr->kmem = NULL; 1034 } else { 1035 ret = mtr_alloc_bufs(hr_dev, mtr, buf_attr, cfg->is_direct, 1036 udata, user_addr); 1037 if (ret) { 1038 ibdev_err(ibdev, 1039 "failed to alloc mtr bufs, ret = %d.\n", ret); 1040 return ret; 1041 } 1042 } 1043 1044 all_pg_cnt = mtr_init_buf_cfg(hr_dev, buf_attr, cfg, &buf_page_shift); 1045 if (all_pg_cnt < 1) { 1046 ret = -ENOBUFS; 1047 ibdev_err(ibdev, "failed to init mtr buf cfg.\n"); 1048 goto err_alloc_bufs; 1049 } 1050 1051 hns_roce_hem_list_init(&mtr->hem_list); 1052 if (!cfg->is_direct) { 1053 ret = hns_roce_hem_list_request(hr_dev, &mtr->hem_list, 1054 cfg->region, cfg->region_count, 1055 ba_page_shift); 1056 if (ret) { 1057 ibdev_err(ibdev, "failed to request mtr hem, ret = %d.\n", 1058 ret); 1059 goto err_alloc_bufs; 1060 } 1061 cfg->root_ba = mtr->hem_list.root_ba; 1062 cfg->ba_pg_shift = ba_page_shift; 1063 } else { 1064 cfg->ba_pg_shift = cfg->buf_pg_shift; 1065 } 1066 1067 /* no buffer to map */ 1068 if (buf_attr->mtt_only) 1069 return 0; 1070 1071 /* alloc a tmp array to store buffer's dma address */ 1072 pages = kvcalloc(all_pg_cnt, sizeof(dma_addr_t), GFP_KERNEL); 1073 if (!pages) { 1074 ret = -ENOMEM; 1075 ibdev_err(ibdev, "failed to alloc mtr page list %d.\n", 1076 all_pg_cnt); 1077 goto err_alloc_hem_list; 1078 } 1079 1080 get_pg_cnt = mtr_get_pages(hr_dev, mtr, pages, all_pg_cnt, 1081 buf_page_shift); 1082 if (get_pg_cnt != all_pg_cnt) { 1083 ibdev_err(ibdev, "failed to get mtr page %d != %d.\n", 1084 get_pg_cnt, all_pg_cnt); 1085 ret = -ENOBUFS; 1086 goto err_alloc_page_list; 1087 } 1088 1089 /* write buffer's dma address to BA table */ 1090 ret = hns_roce_mtr_map(hr_dev, mtr, pages, all_pg_cnt); 1091 if (ret) { 1092 ibdev_err(ibdev, "failed to map mtr pages, ret = %d.\n", ret); 1093 goto err_alloc_page_list; 1094 } 1095 1096 /* drop tmp array */ 1097 kvfree(pages); 1098 return 0; 1099err_alloc_page_list: 1100 kvfree(pages); 1101err_alloc_hem_list: 1102 hns_roce_hem_list_release(hr_dev, &mtr->hem_list); 1103err_alloc_bufs: 1104 mtr_free_bufs(hr_dev, mtr); 1105 return ret; 1106} 1107 1108void hns_roce_mtr_destroy(struct hns_roce_dev *hr_dev, struct hns_roce_mtr *mtr) 1109{ 1110 /* release multi-hop addressing resource */ 1111 hns_roce_hem_list_release(hr_dev, &mtr->hem_list); 1112 1113 /* free buffers */ 1114 mtr_free_bufs(hr_dev, mtr); 1115} 1116