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 <linux/count_zeros.h> 36 #include <rdma/ib_umem.h> 37 #include <linux/math.h> 38 #include "hns_roce_device.h" 39 #include "hns_roce_cmd.h" 40 #include "hns_roce_hem.h" 41 #include "hns_roce_trace.h" 42 43 static u32 hw_index_to_key(int ind) 44 { 45 return ((u32)ind >> 24) | ((u32)ind << 8); 46 } 47 48 unsigned long key_to_hw_index(u32 key) 49 { 50 return (key << 24) | (key >> 8); 51 } 52 53 static int alloc_mr_key(struct hns_roce_dev *hr_dev, struct hns_roce_mr *mr) 54 { 55 struct hns_roce_ida *mtpt_ida = &hr_dev->mr_table.mtpt_ida; 56 struct ib_device *ibdev = &hr_dev->ib_dev; 57 int err; 58 int id; 59 60 /* Allocate a key for mr from mr_table */ 61 id = ida_alloc_range(&mtpt_ida->ida, mtpt_ida->min, mtpt_ida->max, 62 GFP_KERNEL); 63 if (id < 0) { 64 ibdev_err(ibdev, "failed to alloc id for MR key, id(%d)\n", id); 65 return -ENOMEM; 66 } 67 68 mr->key = hw_index_to_key(id); /* MR key */ 69 70 err = hns_roce_table_get(hr_dev, &hr_dev->mr_table.mtpt_table, 71 (unsigned long)id); 72 if (err) { 73 ibdev_err(ibdev, "failed to alloc mtpt, ret = %d.\n", err); 74 goto err_free_bitmap; 75 } 76 77 return 0; 78 err_free_bitmap: 79 ida_free(&mtpt_ida->ida, id); 80 return err; 81 } 82 83 static void free_mr_key(struct hns_roce_dev *hr_dev, struct hns_roce_mr *mr) 84 { 85 unsigned long obj = key_to_hw_index(mr->key); 86 87 hns_roce_table_put(hr_dev, &hr_dev->mr_table.mtpt_table, obj); 88 ida_free(&hr_dev->mr_table.mtpt_ida.ida, (int)obj); 89 } 90 91 static int alloc_mr_pbl(struct hns_roce_dev *hr_dev, struct hns_roce_mr *mr, 92 struct ib_udata *udata, u64 start) 93 { 94 struct ib_device *ibdev = &hr_dev->ib_dev; 95 bool is_fast = mr->type == MR_TYPE_FRMR; 96 struct hns_roce_buf_attr buf_attr = {}; 97 int err; 98 99 mr->pbl_hop_num = is_fast ? 1 : hr_dev->caps.pbl_hop_num; 100 buf_attr.page_shift = is_fast ? PAGE_SHIFT : 101 hr_dev->caps.pbl_buf_pg_sz + PAGE_SHIFT; 102 buf_attr.region[0].size = mr->size; 103 buf_attr.region[0].hopnum = mr->pbl_hop_num; 104 buf_attr.region_count = 1; 105 buf_attr.user_access = mr->access; 106 /* fast MR's buffer is alloced before mapping, not at creation */ 107 buf_attr.mtt_only = is_fast; 108 buf_attr.iova = mr->iova; 109 /* pagesize and hopnum is fixed for fast MR */ 110 buf_attr.adaptive = !is_fast; 111 buf_attr.type = MTR_PBL; 112 113 err = hns_roce_mtr_create(hr_dev, &mr->pbl_mtr, &buf_attr, 114 hr_dev->caps.pbl_ba_pg_sz + PAGE_SHIFT, 115 udata, start); 116 if (err) { 117 ibdev_err(ibdev, "failed to alloc pbl mtr, ret = %d.\n", err); 118 return err; 119 } 120 121 mr->npages = mr->pbl_mtr.hem_cfg.buf_pg_count; 122 mr->pbl_hop_num = buf_attr.region[0].hopnum; 123 124 return err; 125 } 126 127 static void free_mr_pbl(struct hns_roce_dev *hr_dev, struct hns_roce_mr *mr) 128 { 129 hns_roce_mtr_destroy(hr_dev, &mr->pbl_mtr); 130 } 131 132 static void hns_roce_mr_free(struct hns_roce_dev *hr_dev, struct hns_roce_mr *mr) 133 { 134 struct ib_device *ibdev = &hr_dev->ib_dev; 135 int ret; 136 137 if (mr->enabled) { 138 ret = hns_roce_destroy_hw_ctx(hr_dev, HNS_ROCE_CMD_DESTROY_MPT, 139 key_to_hw_index(mr->key) & 140 (hr_dev->caps.num_mtpts - 1)); 141 if (ret) 142 ibdev_warn_ratelimited(ibdev, "failed to destroy mpt, ret = %d.\n", 143 ret); 144 } 145 146 free_mr_pbl(hr_dev, mr); 147 free_mr_key(hr_dev, mr); 148 } 149 150 static int hns_roce_mr_enable(struct hns_roce_dev *hr_dev, 151 struct hns_roce_mr *mr) 152 { 153 unsigned long mtpt_idx = key_to_hw_index(mr->key); 154 struct hns_roce_cmd_mailbox *mailbox; 155 struct device *dev = hr_dev->dev; 156 int ret; 157 158 /* Allocate mailbox memory */ 159 mailbox = hns_roce_alloc_cmd_mailbox(hr_dev); 160 if (IS_ERR(mailbox)) 161 return PTR_ERR(mailbox); 162 163 trace_hns_mr(mr); 164 if (mr->type != MR_TYPE_FRMR) 165 ret = hr_dev->hw->write_mtpt(hr_dev, mailbox->buf, mr); 166 else 167 ret = hr_dev->hw->frmr_write_mtpt(mailbox->buf, mr); 168 if (ret) { 169 dev_err(dev, "failed to write mtpt, ret = %d.\n", ret); 170 goto err_page; 171 } 172 173 ret = hns_roce_create_hw_ctx(hr_dev, mailbox, HNS_ROCE_CMD_CREATE_MPT, 174 mtpt_idx & (hr_dev->caps.num_mtpts - 1)); 175 if (ret) { 176 dev_err(dev, "failed to create mpt, ret = %d.\n", ret); 177 goto err_page; 178 } 179 180 mr->enabled = 1; 181 182 err_page: 183 hns_roce_free_cmd_mailbox(hr_dev, mailbox); 184 185 return ret; 186 } 187 188 void hns_roce_init_mr_table(struct hns_roce_dev *hr_dev) 189 { 190 struct hns_roce_ida *mtpt_ida = &hr_dev->mr_table.mtpt_ida; 191 192 ida_init(&mtpt_ida->ida); 193 mtpt_ida->max = hr_dev->caps.num_mtpts - 1; 194 mtpt_ida->min = hr_dev->caps.reserved_mrws; 195 } 196 197 struct ib_mr *hns_roce_get_dma_mr(struct ib_pd *pd, int acc) 198 { 199 struct hns_roce_dev *hr_dev = to_hr_dev(pd->device); 200 struct hns_roce_mr *mr; 201 int ret; 202 203 mr = kzalloc(sizeof(*mr), GFP_KERNEL); 204 if (!mr) 205 return ERR_PTR(-ENOMEM); 206 207 mr->type = MR_TYPE_DMA; 208 mr->pd = to_hr_pd(pd)->pdn; 209 mr->access = acc; 210 211 /* Allocate memory region key */ 212 hns_roce_hem_list_init(&mr->pbl_mtr.hem_list); 213 ret = alloc_mr_key(hr_dev, mr); 214 if (ret) 215 goto err_free; 216 217 ret = hns_roce_mr_enable(hr_dev, mr); 218 if (ret) 219 goto err_mr; 220 221 mr->ibmr.rkey = mr->ibmr.lkey = mr->key; 222 223 return &mr->ibmr; 224 err_mr: 225 free_mr_key(hr_dev, mr); 226 227 err_free: 228 kfree(mr); 229 return ERR_PTR(ret); 230 } 231 232 struct ib_mr *hns_roce_reg_user_mr(struct ib_pd *pd, u64 start, u64 length, 233 u64 virt_addr, int access_flags, 234 struct ib_udata *udata) 235 { 236 struct hns_roce_dev *hr_dev = to_hr_dev(pd->device); 237 struct hns_roce_mr *mr; 238 int ret; 239 240 mr = kzalloc(sizeof(*mr), GFP_KERNEL); 241 if (!mr) { 242 ret = -ENOMEM; 243 goto err_out; 244 } 245 246 mr->iova = virt_addr; 247 mr->size = length; 248 mr->pd = to_hr_pd(pd)->pdn; 249 mr->access = access_flags; 250 mr->type = MR_TYPE_MR; 251 252 ret = alloc_mr_key(hr_dev, mr); 253 if (ret) 254 goto err_alloc_mr; 255 256 ret = alloc_mr_pbl(hr_dev, mr, udata, start); 257 if (ret) 258 goto err_alloc_key; 259 260 ret = hns_roce_mr_enable(hr_dev, mr); 261 if (ret) 262 goto err_alloc_pbl; 263 264 mr->ibmr.rkey = mr->ibmr.lkey = mr->key; 265 266 return &mr->ibmr; 267 268 err_alloc_pbl: 269 free_mr_pbl(hr_dev, mr); 270 err_alloc_key: 271 free_mr_key(hr_dev, mr); 272 err_alloc_mr: 273 kfree(mr); 274 err_out: 275 atomic64_inc(&hr_dev->dfx_cnt[HNS_ROCE_DFX_MR_REG_ERR_CNT]); 276 277 return ERR_PTR(ret); 278 } 279 280 struct ib_mr *hns_roce_rereg_user_mr(struct ib_mr *ibmr, int flags, u64 start, 281 u64 length, u64 virt_addr, 282 int mr_access_flags, struct ib_pd *pd, 283 struct ib_udata *udata) 284 { 285 struct hns_roce_dev *hr_dev = to_hr_dev(ibmr->device); 286 struct ib_device *ib_dev = &hr_dev->ib_dev; 287 struct hns_roce_mr *mr = to_hr_mr(ibmr); 288 struct hns_roce_cmd_mailbox *mailbox; 289 unsigned long mtpt_idx; 290 int ret; 291 292 if (!mr->enabled) { 293 ret = -EINVAL; 294 goto err_out; 295 } 296 297 mailbox = hns_roce_alloc_cmd_mailbox(hr_dev); 298 ret = PTR_ERR_OR_ZERO(mailbox); 299 if (ret) 300 goto err_out; 301 302 mtpt_idx = key_to_hw_index(mr->key) & (hr_dev->caps.num_mtpts - 1); 303 304 ret = hns_roce_cmd_mbox(hr_dev, 0, mailbox->dma, HNS_ROCE_CMD_QUERY_MPT, 305 mtpt_idx); 306 if (ret) 307 goto free_cmd_mbox; 308 309 ret = hns_roce_destroy_hw_ctx(hr_dev, HNS_ROCE_CMD_DESTROY_MPT, 310 mtpt_idx); 311 if (ret) 312 ibdev_warn(ib_dev, "failed to destroy MPT, ret = %d.\n", ret); 313 314 mr->enabled = 0; 315 mr->iova = virt_addr; 316 mr->size = length; 317 318 if (flags & IB_MR_REREG_PD) 319 mr->pd = to_hr_pd(pd)->pdn; 320 321 if (flags & IB_MR_REREG_ACCESS) 322 mr->access = mr_access_flags; 323 324 if (flags & IB_MR_REREG_TRANS) { 325 free_mr_pbl(hr_dev, mr); 326 ret = alloc_mr_pbl(hr_dev, mr, udata, start); 327 if (ret) { 328 ibdev_err(ib_dev, "failed to alloc mr PBL, ret = %d.\n", 329 ret); 330 goto free_cmd_mbox; 331 } 332 } 333 334 ret = hr_dev->hw->rereg_write_mtpt(hr_dev, mr, flags, mailbox->buf); 335 if (ret) { 336 ibdev_err(ib_dev, "failed to write mtpt, ret = %d.\n", ret); 337 goto free_cmd_mbox; 338 } 339 340 ret = hns_roce_create_hw_ctx(hr_dev, mailbox, HNS_ROCE_CMD_CREATE_MPT, 341 mtpt_idx); 342 if (ret) { 343 ibdev_err(ib_dev, "failed to create MPT, ret = %d.\n", ret); 344 goto free_cmd_mbox; 345 } 346 347 mr->enabled = 1; 348 349 free_cmd_mbox: 350 hns_roce_free_cmd_mailbox(hr_dev, mailbox); 351 352 err_out: 353 if (ret) { 354 atomic64_inc(&hr_dev->dfx_cnt[HNS_ROCE_DFX_MR_REREG_ERR_CNT]); 355 return ERR_PTR(ret); 356 } 357 358 return NULL; 359 } 360 361 int hns_roce_dereg_mr(struct ib_mr *ibmr, struct ib_udata *udata) 362 { 363 struct hns_roce_dev *hr_dev = to_hr_dev(ibmr->device); 364 struct hns_roce_mr *mr = to_hr_mr(ibmr); 365 366 if (hr_dev->hw->dereg_mr) 367 hr_dev->hw->dereg_mr(hr_dev); 368 369 hns_roce_mr_free(hr_dev, mr); 370 kfree(mr); 371 372 return 0; 373 } 374 375 struct ib_mr *hns_roce_alloc_mr(struct ib_pd *pd, enum ib_mr_type mr_type, 376 u32 max_num_sg) 377 { 378 struct hns_roce_dev *hr_dev = to_hr_dev(pd->device); 379 struct device *dev = hr_dev->dev; 380 struct hns_roce_mr *mr; 381 int ret; 382 383 if (mr_type != IB_MR_TYPE_MEM_REG) 384 return ERR_PTR(-EINVAL); 385 386 if (max_num_sg > HNS_ROCE_FRMR_MAX_PA) { 387 dev_err(dev, "max_num_sg larger than %d\n", 388 HNS_ROCE_FRMR_MAX_PA); 389 return ERR_PTR(-EINVAL); 390 } 391 392 mr = kzalloc(sizeof(*mr), GFP_KERNEL); 393 if (!mr) 394 return ERR_PTR(-ENOMEM); 395 396 mr->type = MR_TYPE_FRMR; 397 mr->pd = to_hr_pd(pd)->pdn; 398 mr->size = max_num_sg * (1 << PAGE_SHIFT); 399 400 /* Allocate memory region key */ 401 ret = alloc_mr_key(hr_dev, mr); 402 if (ret) 403 goto err_free; 404 405 ret = alloc_mr_pbl(hr_dev, mr, NULL, 0); 406 if (ret) 407 goto err_key; 408 409 ret = hns_roce_mr_enable(hr_dev, mr); 410 if (ret) 411 goto err_pbl; 412 413 mr->ibmr.rkey = mr->ibmr.lkey = mr->key; 414 mr->ibmr.length = mr->size; 415 416 return &mr->ibmr; 417 418 err_pbl: 419 free_mr_pbl(hr_dev, mr); 420 err_key: 421 free_mr_key(hr_dev, mr); 422 err_free: 423 kfree(mr); 424 return ERR_PTR(ret); 425 } 426 427 static int hns_roce_set_page(struct ib_mr *ibmr, u64 addr) 428 { 429 struct hns_roce_mr *mr = to_hr_mr(ibmr); 430 431 if (likely(mr->npages < mr->pbl_mtr.hem_cfg.buf_pg_count)) { 432 mr->page_list[mr->npages++] = addr; 433 return 0; 434 } 435 436 return -ENOBUFS; 437 } 438 439 int hns_roce_map_mr_sg(struct ib_mr *ibmr, struct scatterlist *sg, int sg_nents, 440 unsigned int *sg_offset_p) 441 { 442 unsigned int sg_offset = sg_offset_p ? *sg_offset_p : 0; 443 struct hns_roce_dev *hr_dev = to_hr_dev(ibmr->device); 444 struct ib_device *ibdev = &hr_dev->ib_dev; 445 struct hns_roce_mr *mr = to_hr_mr(ibmr); 446 struct hns_roce_mtr *mtr = &mr->pbl_mtr; 447 int ret, sg_num = 0; 448 449 if (!IS_ALIGNED(sg_offset, HNS_ROCE_FRMR_ALIGN_SIZE) || 450 ibmr->page_size < HNS_HW_PAGE_SIZE || 451 ibmr->page_size > HNS_HW_MAX_PAGE_SIZE) 452 return sg_num; 453 454 mr->npages = 0; 455 mr->page_list = kvcalloc(mr->pbl_mtr.hem_cfg.buf_pg_count, 456 sizeof(dma_addr_t), GFP_KERNEL); 457 if (!mr->page_list) 458 return sg_num; 459 460 sg_num = ib_sg_to_pages(ibmr, sg, sg_nents, sg_offset_p, hns_roce_set_page); 461 if (sg_num < 1) { 462 ibdev_err(ibdev, "failed to store sg pages %u %u, cnt = %d.\n", 463 mr->npages, mr->pbl_mtr.hem_cfg.buf_pg_count, sg_num); 464 goto err_page_list; 465 } 466 467 mtr->hem_cfg.region[0].offset = 0; 468 mtr->hem_cfg.region[0].count = mr->npages; 469 mtr->hem_cfg.region[0].hopnum = mr->pbl_hop_num; 470 mtr->hem_cfg.region_count = 1; 471 ret = hns_roce_mtr_map(hr_dev, mtr, mr->page_list, mr->npages); 472 if (ret) { 473 ibdev_err(ibdev, "failed to map sg mtr, ret = %d.\n", ret); 474 sg_num = 0; 475 } else { 476 mr->pbl_mtr.hem_cfg.buf_pg_shift = (u32)ilog2(ibmr->page_size); 477 } 478 479 err_page_list: 480 kvfree(mr->page_list); 481 mr->page_list = NULL; 482 483 return sg_num; 484 } 485 486 static void hns_roce_mw_free(struct hns_roce_dev *hr_dev, 487 struct hns_roce_mw *mw) 488 { 489 struct device *dev = hr_dev->dev; 490 int ret; 491 492 if (mw->enabled) { 493 ret = hns_roce_destroy_hw_ctx(hr_dev, HNS_ROCE_CMD_DESTROY_MPT, 494 key_to_hw_index(mw->rkey) & 495 (hr_dev->caps.num_mtpts - 1)); 496 if (ret) 497 dev_warn(dev, "MW DESTROY_MPT failed (%d)\n", ret); 498 499 hns_roce_table_put(hr_dev, &hr_dev->mr_table.mtpt_table, 500 key_to_hw_index(mw->rkey)); 501 } 502 503 ida_free(&hr_dev->mr_table.mtpt_ida.ida, 504 (int)key_to_hw_index(mw->rkey)); 505 } 506 507 static int hns_roce_mw_enable(struct hns_roce_dev *hr_dev, 508 struct hns_roce_mw *mw) 509 { 510 struct hns_roce_mr_table *mr_table = &hr_dev->mr_table; 511 struct hns_roce_cmd_mailbox *mailbox; 512 struct device *dev = hr_dev->dev; 513 unsigned long mtpt_idx = key_to_hw_index(mw->rkey); 514 int ret; 515 516 /* prepare HEM entry memory */ 517 ret = hns_roce_table_get(hr_dev, &mr_table->mtpt_table, mtpt_idx); 518 if (ret) 519 return ret; 520 521 mailbox = hns_roce_alloc_cmd_mailbox(hr_dev); 522 if (IS_ERR(mailbox)) { 523 ret = PTR_ERR(mailbox); 524 goto err_table; 525 } 526 527 ret = hr_dev->hw->mw_write_mtpt(mailbox->buf, mw); 528 if (ret) { 529 dev_err(dev, "MW write mtpt fail!\n"); 530 goto err_page; 531 } 532 533 ret = hns_roce_create_hw_ctx(hr_dev, mailbox, HNS_ROCE_CMD_CREATE_MPT, 534 mtpt_idx & (hr_dev->caps.num_mtpts - 1)); 535 if (ret) { 536 dev_err(dev, "MW CREATE_MPT failed (%d)\n", ret); 537 goto err_page; 538 } 539 540 mw->enabled = 1; 541 542 hns_roce_free_cmd_mailbox(hr_dev, mailbox); 543 544 return 0; 545 546 err_page: 547 hns_roce_free_cmd_mailbox(hr_dev, mailbox); 548 549 err_table: 550 hns_roce_table_put(hr_dev, &mr_table->mtpt_table, mtpt_idx); 551 552 return ret; 553 } 554 555 int hns_roce_alloc_mw(struct ib_mw *ibmw, struct ib_udata *udata) 556 { 557 struct hns_roce_dev *hr_dev = to_hr_dev(ibmw->device); 558 struct hns_roce_ida *mtpt_ida = &hr_dev->mr_table.mtpt_ida; 559 struct ib_device *ibdev = &hr_dev->ib_dev; 560 struct hns_roce_mw *mw = to_hr_mw(ibmw); 561 int ret; 562 int id; 563 564 /* Allocate a key for mw from mr_table */ 565 id = ida_alloc_range(&mtpt_ida->ida, mtpt_ida->min, mtpt_ida->max, 566 GFP_KERNEL); 567 if (id < 0) { 568 ibdev_err(ibdev, "failed to alloc id for MW key, id(%d)\n", id); 569 return -ENOMEM; 570 } 571 572 mw->rkey = hw_index_to_key(id); 573 574 ibmw->rkey = mw->rkey; 575 mw->pdn = to_hr_pd(ibmw->pd)->pdn; 576 mw->pbl_hop_num = hr_dev->caps.pbl_hop_num; 577 mw->pbl_ba_pg_sz = hr_dev->caps.pbl_ba_pg_sz; 578 mw->pbl_buf_pg_sz = hr_dev->caps.pbl_buf_pg_sz; 579 580 ret = hns_roce_mw_enable(hr_dev, mw); 581 if (ret) 582 goto err_mw; 583 584 return 0; 585 586 err_mw: 587 hns_roce_mw_free(hr_dev, mw); 588 return ret; 589 } 590 591 int hns_roce_dealloc_mw(struct ib_mw *ibmw) 592 { 593 struct hns_roce_dev *hr_dev = to_hr_dev(ibmw->device); 594 struct hns_roce_mw *mw = to_hr_mw(ibmw); 595 596 hns_roce_mw_free(hr_dev, mw); 597 return 0; 598 } 599 600 static int mtr_map_region(struct hns_roce_dev *hr_dev, struct hns_roce_mtr *mtr, 601 struct hns_roce_buf_region *region, dma_addr_t *pages, 602 int max_count) 603 { 604 int count, npage; 605 int offset, end; 606 __le64 *mtts; 607 u64 addr; 608 int i; 609 610 offset = region->offset; 611 end = offset + region->count; 612 npage = 0; 613 while (offset < end && npage < max_count) { 614 count = 0; 615 mtts = hns_roce_hem_list_find_mtt(hr_dev, &mtr->hem_list, 616 offset, &count); 617 if (!mtts) 618 return -ENOBUFS; 619 620 for (i = 0; i < count && npage < max_count; i++) { 621 addr = pages[npage]; 622 623 mtts[i] = cpu_to_le64(addr); 624 npage++; 625 } 626 offset += count; 627 } 628 629 return npage; 630 } 631 632 static inline bool mtr_has_mtt(struct hns_roce_buf_attr *attr) 633 { 634 int i; 635 636 for (i = 0; i < attr->region_count; i++) 637 if (attr->region[i].hopnum != HNS_ROCE_HOP_NUM_0 && 638 attr->region[i].hopnum > 0) 639 return true; 640 641 /* because the mtr only one root base address, when hopnum is 0 means 642 * root base address equals the first buffer address, thus all alloced 643 * memory must in a continuous space accessed by direct mode. 644 */ 645 return false; 646 } 647 648 static inline size_t mtr_bufs_size(struct hns_roce_buf_attr *attr) 649 { 650 size_t size = 0; 651 int i; 652 653 for (i = 0; i < attr->region_count; i++) 654 size += attr->region[i].size; 655 656 return size; 657 } 658 659 /* 660 * check the given pages in continuous address space 661 * Returns 0 on success, or the error page num. 662 */ 663 static inline int mtr_check_direct_pages(dma_addr_t *pages, int page_count, 664 unsigned int page_shift) 665 { 666 size_t page_size = 1 << page_shift; 667 int i; 668 669 for (i = 1; i < page_count; i++) 670 if (pages[i] - pages[i - 1] != page_size) 671 return i; 672 673 return 0; 674 } 675 676 static void mtr_free_bufs(struct hns_roce_dev *hr_dev, struct hns_roce_mtr *mtr) 677 { 678 /* release user buffers */ 679 if (mtr->umem) { 680 ib_umem_release(mtr->umem); 681 mtr->umem = NULL; 682 } 683 684 /* release kernel buffers */ 685 if (mtr->kmem) { 686 hns_roce_buf_free(hr_dev, mtr->kmem); 687 mtr->kmem = NULL; 688 } 689 } 690 691 static int mtr_alloc_bufs(struct hns_roce_dev *hr_dev, struct hns_roce_mtr *mtr, 692 struct hns_roce_buf_attr *buf_attr, 693 struct ib_udata *udata, unsigned long user_addr) 694 { 695 struct ib_device *ibdev = &hr_dev->ib_dev; 696 size_t total_size; 697 698 total_size = mtr_bufs_size(buf_attr); 699 700 if (udata) { 701 mtr->kmem = NULL; 702 mtr->umem = ib_umem_get(ibdev, user_addr, total_size, 703 buf_attr->user_access); 704 if (IS_ERR(mtr->umem)) { 705 ibdev_err(ibdev, "failed to get umem, ret = %ld.\n", 706 PTR_ERR(mtr->umem)); 707 return -ENOMEM; 708 } 709 } else { 710 mtr->umem = NULL; 711 mtr->kmem = hns_roce_buf_alloc(hr_dev, total_size, 712 buf_attr->page_shift, 713 !mtr_has_mtt(buf_attr) ? 714 HNS_ROCE_BUF_DIRECT : 0); 715 if (IS_ERR(mtr->kmem)) { 716 ibdev_err(ibdev, "failed to alloc kmem, ret = %ld.\n", 717 PTR_ERR(mtr->kmem)); 718 return PTR_ERR(mtr->kmem); 719 } 720 } 721 722 return 0; 723 } 724 725 static int cal_mtr_pg_cnt(struct hns_roce_mtr *mtr) 726 { 727 struct hns_roce_buf_region *region; 728 int page_cnt = 0; 729 int i; 730 731 for (i = 0; i < mtr->hem_cfg.region_count; i++) { 732 region = &mtr->hem_cfg.region[i]; 733 page_cnt += region->count; 734 } 735 736 return page_cnt; 737 } 738 739 static bool need_split_huge_page(struct hns_roce_mtr *mtr) 740 { 741 /* When HEM buffer uses 0-level addressing, the page size is 742 * equal to the whole buffer size. If the current MTR has multiple 743 * regions, we split the buffer into small pages(4k, required by hns 744 * ROCEE). These pages will be used in multiple regions. 745 */ 746 return mtr->hem_cfg.is_direct && mtr->hem_cfg.region_count > 1; 747 } 748 749 static int mtr_map_bufs(struct hns_roce_dev *hr_dev, struct hns_roce_mtr *mtr) 750 { 751 struct ib_device *ibdev = &hr_dev->ib_dev; 752 int page_count = cal_mtr_pg_cnt(mtr); 753 unsigned int page_shift; 754 dma_addr_t *pages; 755 int npage; 756 int ret; 757 758 page_shift = need_split_huge_page(mtr) ? HNS_HW_PAGE_SHIFT : 759 mtr->hem_cfg.buf_pg_shift; 760 /* alloc a tmp array to store buffer's dma address */ 761 pages = kvcalloc(page_count, sizeof(dma_addr_t), GFP_KERNEL); 762 if (!pages) 763 return -ENOMEM; 764 765 if (mtr->umem) 766 npage = hns_roce_get_umem_bufs(pages, page_count, 767 mtr->umem, page_shift); 768 else 769 npage = hns_roce_get_kmem_bufs(hr_dev, pages, page_count, 770 mtr->kmem, page_shift); 771 772 if (npage != page_count) { 773 ibdev_err(ibdev, "failed to get mtr page %d != %d.\n", npage, 774 page_count); 775 ret = -ENOBUFS; 776 goto err_alloc_list; 777 } 778 779 if (need_split_huge_page(mtr) && npage > 1) { 780 ret = mtr_check_direct_pages(pages, npage, page_shift); 781 if (ret) { 782 ibdev_err(ibdev, "failed to check %s page: %d / %d.\n", 783 mtr->umem ? "umtr" : "kmtr", ret, npage); 784 ret = -ENOBUFS; 785 goto err_alloc_list; 786 } 787 } 788 789 ret = hns_roce_mtr_map(hr_dev, mtr, pages, page_count); 790 if (ret) 791 ibdev_err(ibdev, "failed to map mtr pages, ret = %d.\n", ret); 792 793 err_alloc_list: 794 kvfree(pages); 795 796 return ret; 797 } 798 799 int hns_roce_mtr_map(struct hns_roce_dev *hr_dev, struct hns_roce_mtr *mtr, 800 dma_addr_t *pages, unsigned int page_cnt) 801 { 802 struct ib_device *ibdev = &hr_dev->ib_dev; 803 struct hns_roce_buf_region *r; 804 unsigned int i, mapped_cnt; 805 int ret = 0; 806 807 /* 808 * Only use the first page address as root ba when hopnum is 0, this 809 * is because the addresses of all pages are consecutive in this case. 810 */ 811 if (mtr->hem_cfg.is_direct) { 812 mtr->hem_cfg.root_ba = pages[0]; 813 return 0; 814 } 815 816 for (i = 0, mapped_cnt = 0; i < mtr->hem_cfg.region_count && 817 mapped_cnt < page_cnt; i++) { 818 r = &mtr->hem_cfg.region[i]; 819 820 if (r->offset + r->count > page_cnt) { 821 ret = -EINVAL; 822 ibdev_err(ibdev, 823 "failed to check mtr%u count %u + %u > %u.\n", 824 i, r->offset, r->count, page_cnt); 825 return ret; 826 } 827 828 ret = mtr_map_region(hr_dev, mtr, r, &pages[r->offset], 829 page_cnt - mapped_cnt); 830 if (ret < 0) { 831 ibdev_err(ibdev, 832 "failed to map mtr%u offset %u, ret = %d.\n", 833 i, r->offset, ret); 834 return ret; 835 } 836 mapped_cnt += ret; 837 ret = 0; 838 } 839 840 if (mapped_cnt < page_cnt) { 841 ret = -ENOBUFS; 842 ibdev_err(ibdev, "failed to map mtr pages count: %u < %u.\n", 843 mapped_cnt, page_cnt); 844 } 845 846 return ret; 847 } 848 849 static int hns_roce_get_direct_addr_mtt(struct hns_roce_hem_cfg *cfg, 850 u32 start_index, u64 *mtt_buf, 851 int mtt_cnt) 852 { 853 int mtt_count; 854 int total = 0; 855 u32 npage; 856 u64 addr; 857 858 if (mtt_cnt > cfg->region_count) 859 return -EINVAL; 860 861 for (mtt_count = 0; mtt_count < cfg->region_count && total < mtt_cnt; 862 mtt_count++) { 863 npage = cfg->region[mtt_count].offset; 864 if (npage < start_index) 865 continue; 866 867 addr = cfg->root_ba + (npage << HNS_HW_PAGE_SHIFT); 868 mtt_buf[total] = addr; 869 870 total++; 871 } 872 873 if (!total) 874 return -ENOENT; 875 876 return 0; 877 } 878 879 static int hns_roce_get_mhop_mtt(struct hns_roce_dev *hr_dev, 880 struct hns_roce_mtr *mtr, u32 start_index, 881 u64 *mtt_buf, int mtt_cnt) 882 { 883 int left = mtt_cnt; 884 int total = 0; 885 int mtt_count; 886 __le64 *mtts; 887 u32 npage; 888 889 while (left > 0) { 890 mtt_count = 0; 891 mtts = hns_roce_hem_list_find_mtt(hr_dev, &mtr->hem_list, 892 start_index + total, 893 &mtt_count); 894 if (!mtts || !mtt_count) 895 break; 896 897 npage = min(mtt_count, left); 898 left -= npage; 899 for (mtt_count = 0; mtt_count < npage; mtt_count++) 900 mtt_buf[total++] = le64_to_cpu(mtts[mtt_count]); 901 } 902 903 if (!total) 904 return -ENOENT; 905 906 return 0; 907 } 908 909 int hns_roce_mtr_find(struct hns_roce_dev *hr_dev, struct hns_roce_mtr *mtr, 910 u32 offset, u64 *mtt_buf, int mtt_max) 911 { 912 struct hns_roce_hem_cfg *cfg = &mtr->hem_cfg; 913 u32 start_index; 914 int ret; 915 916 if (!mtt_buf || mtt_max < 1) 917 return -EINVAL; 918 919 /* no mtt memory in direct mode, so just return the buffer address */ 920 if (cfg->is_direct) { 921 start_index = offset >> HNS_HW_PAGE_SHIFT; 922 ret = hns_roce_get_direct_addr_mtt(cfg, start_index, 923 mtt_buf, mtt_max); 924 } else { 925 start_index = offset >> cfg->buf_pg_shift; 926 ret = hns_roce_get_mhop_mtt(hr_dev, mtr, start_index, 927 mtt_buf, mtt_max); 928 } 929 return ret; 930 } 931 932 static int get_best_page_shift(struct hns_roce_dev *hr_dev, 933 struct hns_roce_mtr *mtr, 934 struct hns_roce_buf_attr *buf_attr) 935 { 936 unsigned int page_sz; 937 938 if (!buf_attr->adaptive || buf_attr->type != MTR_PBL || !mtr->umem) 939 return 0; 940 941 page_sz = ib_umem_find_best_pgsz(mtr->umem, 942 hr_dev->caps.page_size_cap, 943 buf_attr->iova); 944 if (!page_sz) 945 return -EINVAL; 946 947 buf_attr->page_shift = order_base_2(page_sz); 948 return 0; 949 } 950 951 static int get_best_hop_num(struct hns_roce_dev *hr_dev, 952 struct hns_roce_mtr *mtr, 953 struct hns_roce_buf_attr *buf_attr, 954 unsigned int ba_pg_shift) 955 { 956 #define INVALID_HOPNUM -1 957 #define MIN_BA_CNT 1 958 size_t buf_pg_sz = 1 << buf_attr->page_shift; 959 struct ib_device *ibdev = &hr_dev->ib_dev; 960 size_t ba_pg_sz = 1 << ba_pg_shift; 961 int hop_num = INVALID_HOPNUM; 962 size_t unit = MIN_BA_CNT; 963 size_t ba_cnt; 964 int j; 965 966 if (!buf_attr->adaptive || buf_attr->type != MTR_PBL) 967 return 0; 968 969 /* Caculating the number of buf pages, each buf page need a BA */ 970 if (mtr->umem) 971 ba_cnt = ib_umem_num_dma_blocks(mtr->umem, buf_pg_sz); 972 else 973 ba_cnt = DIV_ROUND_UP(buf_attr->region[0].size, buf_pg_sz); 974 975 for (j = 0; j <= HNS_ROCE_MAX_HOP_NUM; j++) { 976 if (ba_cnt <= unit) { 977 hop_num = j; 978 break; 979 } 980 /* Number of BAs can be represented at per hop */ 981 unit *= ba_pg_sz / BA_BYTE_LEN; 982 } 983 984 if (hop_num < 0) { 985 ibdev_err(ibdev, 986 "failed to calculate a valid hopnum.\n"); 987 return -EINVAL; 988 } 989 990 buf_attr->region[0].hopnum = hop_num; 991 992 return 0; 993 } 994 995 static bool is_buf_attr_valid(struct hns_roce_dev *hr_dev, 996 struct hns_roce_buf_attr *attr) 997 { 998 struct ib_device *ibdev = &hr_dev->ib_dev; 999 1000 if (attr->region_count > ARRAY_SIZE(attr->region) || 1001 attr->region_count < 1 || attr->page_shift < HNS_HW_PAGE_SHIFT) { 1002 ibdev_err(ibdev, 1003 "invalid buf attr, region count %u, page shift %u.\n", 1004 attr->region_count, attr->page_shift); 1005 return false; 1006 } 1007 1008 return true; 1009 } 1010 1011 static int mtr_init_buf_cfg(struct hns_roce_dev *hr_dev, 1012 struct hns_roce_mtr *mtr, 1013 struct hns_roce_buf_attr *attr) 1014 { 1015 struct hns_roce_hem_cfg *cfg = &mtr->hem_cfg; 1016 struct hns_roce_buf_region *r; 1017 size_t buf_pg_sz; 1018 size_t buf_size; 1019 int page_cnt, i; 1020 u64 pgoff = 0; 1021 1022 if (!is_buf_attr_valid(hr_dev, attr)) 1023 return -EINVAL; 1024 1025 /* If mtt is disabled, all pages must be within a continuous range */ 1026 cfg->is_direct = !mtr_has_mtt(attr); 1027 cfg->region_count = attr->region_count; 1028 buf_size = mtr_bufs_size(attr); 1029 if (need_split_huge_page(mtr)) { 1030 buf_pg_sz = HNS_HW_PAGE_SIZE; 1031 cfg->buf_pg_count = 1; 1032 /* The ROCEE requires the page size to be 4K * 2 ^ N. */ 1033 cfg->buf_pg_shift = HNS_HW_PAGE_SHIFT + 1034 order_base_2(DIV_ROUND_UP(buf_size, HNS_HW_PAGE_SIZE)); 1035 } else { 1036 buf_pg_sz = 1 << attr->page_shift; 1037 cfg->buf_pg_count = mtr->umem ? 1038 ib_umem_num_dma_blocks(mtr->umem, buf_pg_sz) : 1039 DIV_ROUND_UP(buf_size, buf_pg_sz); 1040 cfg->buf_pg_shift = attr->page_shift; 1041 pgoff = mtr->umem ? mtr->umem->address & ~PAGE_MASK : 0; 1042 } 1043 1044 /* Convert buffer size to page index and page count for each region and 1045 * the buffer's offset needs to be appended to the first region. 1046 */ 1047 for (page_cnt = 0, i = 0; i < attr->region_count; i++) { 1048 r = &cfg->region[i]; 1049 r->offset = page_cnt; 1050 buf_size = hr_hw_page_align(attr->region[i].size + pgoff); 1051 if (attr->type == MTR_PBL && mtr->umem) 1052 r->count = ib_umem_num_dma_blocks(mtr->umem, buf_pg_sz); 1053 else 1054 r->count = DIV_ROUND_UP(buf_size, buf_pg_sz); 1055 1056 pgoff = 0; 1057 page_cnt += r->count; 1058 r->hopnum = to_hr_hem_hopnum(attr->region[i].hopnum, r->count); 1059 } 1060 1061 return 0; 1062 } 1063 1064 static u64 cal_pages_per_l1ba(unsigned int ba_per_bt, unsigned int hopnum) 1065 { 1066 return int_pow(ba_per_bt, hopnum - 1); 1067 } 1068 1069 static unsigned int cal_best_bt_pg_sz(struct hns_roce_dev *hr_dev, 1070 struct hns_roce_mtr *mtr, 1071 unsigned int pg_shift) 1072 { 1073 unsigned long cap = hr_dev->caps.page_size_cap; 1074 struct hns_roce_buf_region *re; 1075 unsigned int pgs_per_l1ba; 1076 unsigned int ba_per_bt; 1077 unsigned int ba_num; 1078 int i; 1079 1080 for_each_set_bit_from(pg_shift, &cap, sizeof(cap) * BITS_PER_BYTE) { 1081 if (!(BIT(pg_shift) & cap)) 1082 continue; 1083 1084 ba_per_bt = BIT(pg_shift) / BA_BYTE_LEN; 1085 ba_num = 0; 1086 for (i = 0; i < mtr->hem_cfg.region_count; i++) { 1087 re = &mtr->hem_cfg.region[i]; 1088 if (re->hopnum == 0) 1089 continue; 1090 1091 pgs_per_l1ba = cal_pages_per_l1ba(ba_per_bt, re->hopnum); 1092 ba_num += DIV_ROUND_UP(re->count, pgs_per_l1ba); 1093 } 1094 1095 if (ba_num <= ba_per_bt) 1096 return pg_shift; 1097 } 1098 1099 return 0; 1100 } 1101 1102 static int mtr_alloc_mtt(struct hns_roce_dev *hr_dev, struct hns_roce_mtr *mtr, 1103 unsigned int ba_page_shift) 1104 { 1105 struct hns_roce_hem_cfg *cfg = &mtr->hem_cfg; 1106 int ret; 1107 1108 hns_roce_hem_list_init(&mtr->hem_list); 1109 if (!cfg->is_direct) { 1110 ba_page_shift = cal_best_bt_pg_sz(hr_dev, mtr, ba_page_shift); 1111 if (!ba_page_shift) 1112 return -ERANGE; 1113 1114 ret = hns_roce_hem_list_request(hr_dev, &mtr->hem_list, 1115 cfg->region, cfg->region_count, 1116 ba_page_shift); 1117 if (ret) 1118 return ret; 1119 cfg->root_ba = mtr->hem_list.root_ba; 1120 cfg->ba_pg_shift = ba_page_shift; 1121 } else { 1122 cfg->ba_pg_shift = cfg->buf_pg_shift; 1123 } 1124 1125 return 0; 1126 } 1127 1128 static void mtr_free_mtt(struct hns_roce_dev *hr_dev, struct hns_roce_mtr *mtr) 1129 { 1130 hns_roce_hem_list_release(hr_dev, &mtr->hem_list); 1131 } 1132 1133 /** 1134 * hns_roce_mtr_create - Create hns memory translate region. 1135 * 1136 * @hr_dev: RoCE device struct pointer 1137 * @mtr: memory translate region 1138 * @buf_attr: buffer attribute for creating mtr 1139 * @ba_page_shift: page shift for multi-hop base address table 1140 * @udata: user space context, if it's NULL, means kernel space 1141 * @user_addr: userspace virtual address to start at 1142 */ 1143 int hns_roce_mtr_create(struct hns_roce_dev *hr_dev, struct hns_roce_mtr *mtr, 1144 struct hns_roce_buf_attr *buf_attr, 1145 unsigned int ba_page_shift, struct ib_udata *udata, 1146 unsigned long user_addr) 1147 { 1148 struct ib_device *ibdev = &hr_dev->ib_dev; 1149 int ret; 1150 1151 trace_hns_buf_attr(buf_attr); 1152 /* The caller has its own buffer list and invokes the hns_roce_mtr_map() 1153 * to finish the MTT configuration. 1154 */ 1155 if (buf_attr->mtt_only) { 1156 mtr->umem = NULL; 1157 mtr->kmem = NULL; 1158 } else { 1159 ret = mtr_alloc_bufs(hr_dev, mtr, buf_attr, udata, user_addr); 1160 if (ret) { 1161 ibdev_err(ibdev, 1162 "failed to alloc mtr bufs, ret = %d.\n", ret); 1163 return ret; 1164 } 1165 1166 ret = get_best_page_shift(hr_dev, mtr, buf_attr); 1167 if (ret) 1168 goto err_init_buf; 1169 1170 ret = get_best_hop_num(hr_dev, mtr, buf_attr, ba_page_shift); 1171 if (ret) 1172 goto err_init_buf; 1173 } 1174 1175 ret = mtr_init_buf_cfg(hr_dev, mtr, buf_attr); 1176 if (ret) 1177 goto err_init_buf; 1178 1179 ret = mtr_alloc_mtt(hr_dev, mtr, ba_page_shift); 1180 if (ret) { 1181 ibdev_err(ibdev, "failed to alloc mtr mtt, ret = %d.\n", ret); 1182 goto err_init_buf; 1183 } 1184 1185 if (buf_attr->mtt_only) 1186 return 0; 1187 1188 /* Write buffer's dma address to MTT */ 1189 ret = mtr_map_bufs(hr_dev, mtr); 1190 if (ret) { 1191 ibdev_err(ibdev, "failed to map mtr bufs, ret = %d.\n", ret); 1192 goto err_alloc_mtt; 1193 } 1194 1195 return 0; 1196 1197 err_alloc_mtt: 1198 mtr_free_mtt(hr_dev, mtr); 1199 err_init_buf: 1200 mtr_free_bufs(hr_dev, mtr); 1201 1202 return ret; 1203 } 1204 1205 void hns_roce_mtr_destroy(struct hns_roce_dev *hr_dev, struct hns_roce_mtr *mtr) 1206 { 1207 /* release multi-hop addressing resource */ 1208 hns_roce_hem_list_release(hr_dev, &mtr->hem_list); 1209 1210 /* free buffers */ 1211 mtr_free_bufs(hr_dev, mtr); 1212 } 1213