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