1 /* 2 * Copyright (c) 2016 Hisilicon Limited. 3 * 4 * This software is available to you under a choice of one of two 5 * licenses. You may choose to be licensed under the terms of the GNU 6 * General Public License (GPL) Version 2, available from the file 7 * COPYING in the main directory of this source tree, or the 8 * OpenIB.org BSD license below: 9 * 10 * Redistribution and use in source and binary forms, with or 11 * without modification, are permitted provided that the following 12 * conditions are met: 13 * 14 * - Redistributions of source code must retain the above 15 * copyright notice, this list of conditions and the following 16 * disclaimer. 17 * 18 * - Redistributions in binary form must reproduce the above 19 * copyright notice, this list of conditions and the following 20 * disclaimer in the documentation and/or other materials 21 * provided with the distribution. 22 * 23 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, 24 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF 25 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND 26 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS 27 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN 28 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN 29 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE 30 * SOFTWARE. 31 */ 32 33 #include <rdma/ib_umem.h> 34 #include <rdma/uverbs_ioctl.h> 35 #include "hns_roce_device.h" 36 #include "hns_roce_cmd.h" 37 #include "hns_roce_hem.h" 38 #include "hns_roce_common.h" 39 40 static u8 get_least_load_bankid_for_cq(struct hns_roce_bank *bank) 41 { 42 u32 least_load = bank[0].inuse; 43 u8 bankid = 0; 44 u32 bankcnt; 45 u8 i; 46 47 for (i = 1; i < HNS_ROCE_CQ_BANK_NUM; i++) { 48 bankcnt = bank[i].inuse; 49 if (bankcnt < least_load) { 50 least_load = bankcnt; 51 bankid = i; 52 } 53 } 54 55 return bankid; 56 } 57 58 static int alloc_cqn(struct hns_roce_dev *hr_dev, struct hns_roce_cq *hr_cq) 59 { 60 struct hns_roce_cq_table *cq_table = &hr_dev->cq_table; 61 struct hns_roce_bank *bank; 62 u8 bankid; 63 int id; 64 65 mutex_lock(&cq_table->bank_mutex); 66 bankid = get_least_load_bankid_for_cq(cq_table->bank); 67 bank = &cq_table->bank[bankid]; 68 69 id = ida_alloc_range(&bank->ida, bank->min, bank->max, GFP_KERNEL); 70 if (id < 0) { 71 mutex_unlock(&cq_table->bank_mutex); 72 return id; 73 } 74 75 /* the lower 2 bits is bankid */ 76 hr_cq->cqn = (id << CQ_BANKID_SHIFT) | bankid; 77 bank->inuse++; 78 mutex_unlock(&cq_table->bank_mutex); 79 80 return 0; 81 } 82 83 static inline u8 get_cq_bankid(unsigned long cqn) 84 { 85 /* The lower 2 bits of CQN are used to hash to different banks */ 86 return (u8)(cqn & GENMASK(1, 0)); 87 } 88 89 static void free_cqn(struct hns_roce_dev *hr_dev, unsigned long cqn) 90 { 91 struct hns_roce_cq_table *cq_table = &hr_dev->cq_table; 92 struct hns_roce_bank *bank; 93 94 bank = &cq_table->bank[get_cq_bankid(cqn)]; 95 96 ida_free(&bank->ida, cqn >> CQ_BANKID_SHIFT); 97 98 mutex_lock(&cq_table->bank_mutex); 99 bank->inuse--; 100 mutex_unlock(&cq_table->bank_mutex); 101 } 102 103 static int hns_roce_create_cqc(struct hns_roce_dev *hr_dev, 104 struct hns_roce_cq *hr_cq, 105 u64 *mtts, dma_addr_t dma_handle) 106 { 107 struct ib_device *ibdev = &hr_dev->ib_dev; 108 struct hns_roce_cmd_mailbox *mailbox; 109 int ret; 110 111 mailbox = hns_roce_alloc_cmd_mailbox(hr_dev); 112 if (IS_ERR(mailbox)) { 113 ibdev_err(ibdev, "failed to alloc mailbox for CQC.\n"); 114 return PTR_ERR(mailbox); 115 } 116 117 hr_dev->hw->write_cqc(hr_dev, hr_cq, mailbox->buf, mtts, dma_handle); 118 119 ret = hns_roce_create_hw_ctx(hr_dev, mailbox, HNS_ROCE_CMD_CREATE_CQC, 120 hr_cq->cqn); 121 if (ret) 122 ibdev_err(ibdev, 123 "failed to send create cmd for CQ(0x%lx), ret = %d.\n", 124 hr_cq->cqn, ret); 125 126 hns_roce_free_cmd_mailbox(hr_dev, mailbox); 127 128 return ret; 129 } 130 131 static int alloc_cqc(struct hns_roce_dev *hr_dev, struct hns_roce_cq *hr_cq) 132 { 133 struct hns_roce_cq_table *cq_table = &hr_dev->cq_table; 134 struct ib_device *ibdev = &hr_dev->ib_dev; 135 u64 mtts[MTT_MIN_COUNT] = {}; 136 dma_addr_t dma_handle; 137 int ret; 138 139 ret = hns_roce_mtr_find(hr_dev, &hr_cq->mtr, 0, mtts, ARRAY_SIZE(mtts), 140 &dma_handle); 141 if (!ret) { 142 ibdev_err(ibdev, "failed to find CQ mtr, ret = %d.\n", ret); 143 return -EINVAL; 144 } 145 146 /* Get CQC memory HEM(Hardware Entry Memory) table */ 147 ret = hns_roce_table_get(hr_dev, &cq_table->table, hr_cq->cqn); 148 if (ret) { 149 ibdev_err(ibdev, "failed to get CQ(0x%lx) context, ret = %d.\n", 150 hr_cq->cqn, ret); 151 return ret; 152 } 153 154 ret = xa_err(xa_store_irq(&cq_table->array, hr_cq->cqn, hr_cq, GFP_KERNEL)); 155 if (ret) { 156 ibdev_err(ibdev, "failed to xa_store CQ, ret = %d.\n", ret); 157 goto err_put; 158 } 159 160 ret = hns_roce_create_cqc(hr_dev, hr_cq, mtts, dma_handle); 161 if (ret) 162 goto err_xa; 163 164 return 0; 165 166 err_xa: 167 xa_erase_irq(&cq_table->array, hr_cq->cqn); 168 err_put: 169 hns_roce_table_put(hr_dev, &cq_table->table, hr_cq->cqn); 170 171 return ret; 172 } 173 174 static void free_cqc(struct hns_roce_dev *hr_dev, struct hns_roce_cq *hr_cq) 175 { 176 struct hns_roce_cq_table *cq_table = &hr_dev->cq_table; 177 struct device *dev = hr_dev->dev; 178 int ret; 179 180 ret = hns_roce_destroy_hw_ctx(hr_dev, HNS_ROCE_CMD_DESTROY_CQC, 181 hr_cq->cqn); 182 if (ret) 183 dev_err(dev, "DESTROY_CQ failed (%d) for CQN %06lx\n", ret, 184 hr_cq->cqn); 185 186 xa_erase_irq(&cq_table->array, hr_cq->cqn); 187 188 /* Waiting interrupt process procedure carried out */ 189 synchronize_irq(hr_dev->eq_table.eq[hr_cq->vector].irq); 190 191 /* wait for all interrupt processed */ 192 if (refcount_dec_and_test(&hr_cq->refcount)) 193 complete(&hr_cq->free); 194 wait_for_completion(&hr_cq->free); 195 196 hns_roce_table_put(hr_dev, &cq_table->table, hr_cq->cqn); 197 } 198 199 static int alloc_cq_buf(struct hns_roce_dev *hr_dev, struct hns_roce_cq *hr_cq, 200 struct ib_udata *udata, unsigned long addr) 201 { 202 struct ib_device *ibdev = &hr_dev->ib_dev; 203 struct hns_roce_buf_attr buf_attr = {}; 204 int ret; 205 206 buf_attr.page_shift = hr_dev->caps.cqe_buf_pg_sz + PAGE_SHIFT; 207 buf_attr.region[0].size = hr_cq->cq_depth * hr_cq->cqe_size; 208 buf_attr.region[0].hopnum = hr_dev->caps.cqe_hop_num; 209 buf_attr.region_count = 1; 210 211 ret = hns_roce_mtr_create(hr_dev, &hr_cq->mtr, &buf_attr, 212 hr_dev->caps.cqe_ba_pg_sz + PAGE_SHIFT, 213 udata, addr); 214 if (ret) 215 ibdev_err(ibdev, "failed to alloc CQ mtr, ret = %d.\n", ret); 216 217 return ret; 218 } 219 220 static void free_cq_buf(struct hns_roce_dev *hr_dev, struct hns_roce_cq *hr_cq) 221 { 222 hns_roce_mtr_destroy(hr_dev, &hr_cq->mtr); 223 } 224 225 static int alloc_cq_db(struct hns_roce_dev *hr_dev, struct hns_roce_cq *hr_cq, 226 struct ib_udata *udata, unsigned long addr, 227 struct hns_roce_ib_create_cq_resp *resp) 228 { 229 bool has_db = hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_CQ_RECORD_DB; 230 struct hns_roce_ucontext *uctx; 231 int err; 232 233 if (udata) { 234 if (has_db && 235 udata->outlen >= offsetofend(typeof(*resp), cap_flags)) { 236 uctx = rdma_udata_to_drv_context(udata, 237 struct hns_roce_ucontext, ibucontext); 238 err = hns_roce_db_map_user(uctx, addr, &hr_cq->db); 239 if (err) 240 return err; 241 hr_cq->flags |= HNS_ROCE_CQ_FLAG_RECORD_DB; 242 resp->cap_flags |= HNS_ROCE_CQ_FLAG_RECORD_DB; 243 } 244 } else { 245 if (has_db) { 246 err = hns_roce_alloc_db(hr_dev, &hr_cq->db, 1); 247 if (err) 248 return err; 249 hr_cq->set_ci_db = hr_cq->db.db_record; 250 *hr_cq->set_ci_db = 0; 251 hr_cq->flags |= HNS_ROCE_CQ_FLAG_RECORD_DB; 252 } 253 hr_cq->db_reg = hr_dev->reg_base + hr_dev->odb_offset + 254 DB_REG_OFFSET * hr_dev->priv_uar.index; 255 } 256 257 return 0; 258 } 259 260 static void free_cq_db(struct hns_roce_dev *hr_dev, struct hns_roce_cq *hr_cq, 261 struct ib_udata *udata) 262 { 263 struct hns_roce_ucontext *uctx; 264 265 if (!(hr_cq->flags & HNS_ROCE_CQ_FLAG_RECORD_DB)) 266 return; 267 268 hr_cq->flags &= ~HNS_ROCE_CQ_FLAG_RECORD_DB; 269 if (udata) { 270 uctx = rdma_udata_to_drv_context(udata, 271 struct hns_roce_ucontext, 272 ibucontext); 273 hns_roce_db_unmap_user(uctx, &hr_cq->db); 274 } else { 275 hns_roce_free_db(hr_dev, &hr_cq->db); 276 } 277 } 278 279 static int verify_cq_create_attr(struct hns_roce_dev *hr_dev, 280 const struct ib_cq_init_attr *attr) 281 { 282 struct ib_device *ibdev = &hr_dev->ib_dev; 283 284 if (!attr->cqe || attr->cqe > hr_dev->caps.max_cqes) { 285 ibdev_err(ibdev, "failed to check CQ count %u, max = %u.\n", 286 attr->cqe, hr_dev->caps.max_cqes); 287 return -EINVAL; 288 } 289 290 if (attr->comp_vector >= hr_dev->caps.num_comp_vectors) { 291 ibdev_err(ibdev, "failed to check CQ vector = %u, max = %d.\n", 292 attr->comp_vector, hr_dev->caps.num_comp_vectors); 293 return -EINVAL; 294 } 295 296 return 0; 297 } 298 299 static int get_cq_ucmd(struct hns_roce_cq *hr_cq, struct ib_udata *udata, 300 struct hns_roce_ib_create_cq *ucmd) 301 { 302 struct ib_device *ibdev = hr_cq->ib_cq.device; 303 int ret; 304 305 ret = ib_copy_from_udata(ucmd, udata, min(udata->inlen, sizeof(*ucmd))); 306 if (ret) { 307 ibdev_err(ibdev, "failed to copy CQ udata, ret = %d.\n", ret); 308 return ret; 309 } 310 311 return 0; 312 } 313 314 static void set_cq_param(struct hns_roce_cq *hr_cq, u32 cq_entries, int vector, 315 struct hns_roce_ib_create_cq *ucmd) 316 { 317 struct hns_roce_dev *hr_dev = to_hr_dev(hr_cq->ib_cq.device); 318 319 cq_entries = max(cq_entries, hr_dev->caps.min_cqes); 320 cq_entries = roundup_pow_of_two(cq_entries); 321 hr_cq->ib_cq.cqe = cq_entries - 1; /* used as cqe index */ 322 hr_cq->cq_depth = cq_entries; 323 hr_cq->vector = vector; 324 325 spin_lock_init(&hr_cq->lock); 326 INIT_LIST_HEAD(&hr_cq->sq_list); 327 INIT_LIST_HEAD(&hr_cq->rq_list); 328 } 329 330 static int set_cqe_size(struct hns_roce_cq *hr_cq, struct ib_udata *udata, 331 struct hns_roce_ib_create_cq *ucmd) 332 { 333 struct hns_roce_dev *hr_dev = to_hr_dev(hr_cq->ib_cq.device); 334 335 if (!udata) { 336 hr_cq->cqe_size = hr_dev->caps.cqe_sz; 337 return 0; 338 } 339 340 if (udata->inlen >= offsetofend(typeof(*ucmd), cqe_size)) { 341 if (ucmd->cqe_size != HNS_ROCE_V2_CQE_SIZE && 342 ucmd->cqe_size != HNS_ROCE_V3_CQE_SIZE) { 343 ibdev_err(&hr_dev->ib_dev, 344 "invalid cqe size %u.\n", ucmd->cqe_size); 345 return -EINVAL; 346 } 347 348 hr_cq->cqe_size = ucmd->cqe_size; 349 } else { 350 hr_cq->cqe_size = HNS_ROCE_V2_CQE_SIZE; 351 } 352 353 return 0; 354 } 355 356 int hns_roce_create_cq(struct ib_cq *ib_cq, const struct ib_cq_init_attr *attr, 357 struct ib_udata *udata) 358 { 359 struct hns_roce_dev *hr_dev = to_hr_dev(ib_cq->device); 360 struct hns_roce_ib_create_cq_resp resp = {}; 361 struct hns_roce_cq *hr_cq = to_hr_cq(ib_cq); 362 struct ib_device *ibdev = &hr_dev->ib_dev; 363 struct hns_roce_ib_create_cq ucmd = {}; 364 int ret; 365 366 if (attr->flags) 367 return -EOPNOTSUPP; 368 369 ret = verify_cq_create_attr(hr_dev, attr); 370 if (ret) 371 return ret; 372 373 if (udata) { 374 ret = get_cq_ucmd(hr_cq, udata, &ucmd); 375 if (ret) 376 return ret; 377 } 378 379 set_cq_param(hr_cq, attr->cqe, attr->comp_vector, &ucmd); 380 381 ret = set_cqe_size(hr_cq, udata, &ucmd); 382 if (ret) 383 return ret; 384 385 ret = alloc_cq_buf(hr_dev, hr_cq, udata, ucmd.buf_addr); 386 if (ret) { 387 ibdev_err(ibdev, "failed to alloc CQ buf, ret = %d.\n", ret); 388 return ret; 389 } 390 391 ret = alloc_cq_db(hr_dev, hr_cq, udata, ucmd.db_addr, &resp); 392 if (ret) { 393 ibdev_err(ibdev, "failed to alloc CQ db, ret = %d.\n", ret); 394 goto err_cq_buf; 395 } 396 397 ret = alloc_cqn(hr_dev, hr_cq); 398 if (ret) { 399 ibdev_err(ibdev, "failed to alloc CQN, ret = %d.\n", ret); 400 goto err_cq_db; 401 } 402 403 ret = alloc_cqc(hr_dev, hr_cq); 404 if (ret) { 405 ibdev_err(ibdev, 406 "failed to alloc CQ context, ret = %d.\n", ret); 407 goto err_cqn; 408 } 409 410 if (udata) { 411 resp.cqn = hr_cq->cqn; 412 ret = ib_copy_to_udata(udata, &resp, 413 min(udata->outlen, sizeof(resp))); 414 if (ret) 415 goto err_cqc; 416 } 417 418 hr_cq->cons_index = 0; 419 hr_cq->arm_sn = 1; 420 refcount_set(&hr_cq->refcount, 1); 421 init_completion(&hr_cq->free); 422 423 return 0; 424 425 err_cqc: 426 free_cqc(hr_dev, hr_cq); 427 err_cqn: 428 free_cqn(hr_dev, hr_cq->cqn); 429 err_cq_db: 430 free_cq_db(hr_dev, hr_cq, udata); 431 err_cq_buf: 432 free_cq_buf(hr_dev, hr_cq); 433 return ret; 434 } 435 436 int hns_roce_destroy_cq(struct ib_cq *ib_cq, struct ib_udata *udata) 437 { 438 struct hns_roce_dev *hr_dev = to_hr_dev(ib_cq->device); 439 struct hns_roce_cq *hr_cq = to_hr_cq(ib_cq); 440 441 free_cqc(hr_dev, hr_cq); 442 free_cqn(hr_dev, hr_cq->cqn); 443 free_cq_db(hr_dev, hr_cq, udata); 444 free_cq_buf(hr_dev, hr_cq); 445 446 return 0; 447 } 448 449 void hns_roce_cq_completion(struct hns_roce_dev *hr_dev, u32 cqn) 450 { 451 struct hns_roce_cq *hr_cq; 452 struct ib_cq *ibcq; 453 454 hr_cq = xa_load(&hr_dev->cq_table.array, 455 cqn & (hr_dev->caps.num_cqs - 1)); 456 if (!hr_cq) { 457 dev_warn(hr_dev->dev, "completion event for bogus CQ 0x%06x\n", 458 cqn); 459 return; 460 } 461 462 ++hr_cq->arm_sn; 463 ibcq = &hr_cq->ib_cq; 464 if (ibcq->comp_handler) 465 ibcq->comp_handler(ibcq, ibcq->cq_context); 466 } 467 468 void hns_roce_cq_event(struct hns_roce_dev *hr_dev, u32 cqn, int event_type) 469 { 470 struct device *dev = hr_dev->dev; 471 struct hns_roce_cq *hr_cq; 472 struct ib_event event; 473 struct ib_cq *ibcq; 474 475 if (event_type != HNS_ROCE_EVENT_TYPE_CQ_ID_INVALID && 476 event_type != HNS_ROCE_EVENT_TYPE_CQ_ACCESS_ERROR && 477 event_type != HNS_ROCE_EVENT_TYPE_CQ_OVERFLOW) { 478 dev_err(dev, "unexpected event type 0x%x on CQ 0x%06x\n", 479 event_type, cqn); 480 return; 481 } 482 483 xa_lock(&hr_dev->cq_table.array); 484 hr_cq = xa_load(&hr_dev->cq_table.array, 485 cqn & (hr_dev->caps.num_cqs - 1)); 486 if (hr_cq) 487 refcount_inc(&hr_cq->refcount); 488 xa_unlock(&hr_dev->cq_table.array); 489 if (!hr_cq) { 490 dev_warn(dev, "async event for bogus CQ 0x%06x\n", cqn); 491 return; 492 } 493 494 ibcq = &hr_cq->ib_cq; 495 if (ibcq->event_handler) { 496 event.device = ibcq->device; 497 event.element.cq = ibcq; 498 event.event = IB_EVENT_CQ_ERR; 499 ibcq->event_handler(&event, ibcq->cq_context); 500 } 501 502 if (refcount_dec_and_test(&hr_cq->refcount)) 503 complete(&hr_cq->free); 504 } 505 506 void hns_roce_init_cq_table(struct hns_roce_dev *hr_dev) 507 { 508 struct hns_roce_cq_table *cq_table = &hr_dev->cq_table; 509 unsigned int reserved_from_bot; 510 unsigned int i; 511 512 mutex_init(&cq_table->bank_mutex); 513 xa_init(&cq_table->array); 514 515 reserved_from_bot = hr_dev->caps.reserved_cqs; 516 517 for (i = 0; i < reserved_from_bot; i++) { 518 cq_table->bank[get_cq_bankid(i)].inuse++; 519 cq_table->bank[get_cq_bankid(i)].min++; 520 } 521 522 for (i = 0; i < HNS_ROCE_CQ_BANK_NUM; i++) { 523 ida_init(&cq_table->bank[i].ida); 524 cq_table->bank[i].max = hr_dev->caps.num_cqs / 525 HNS_ROCE_CQ_BANK_NUM - 1; 526 } 527 } 528 529 void hns_roce_cleanup_cq_table(struct hns_roce_dev *hr_dev) 530 { 531 int i; 532 533 for (i = 0; i < HNS_ROCE_CQ_BANK_NUM; i++) 534 ida_destroy(&hr_dev->cq_table.bank[i].ida); 535 } 536