1 /* 2 * Copyright (c) 2013-2015, Mellanox Technologies. All rights reserved. 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 <linux/kref.h> 34 #include <rdma/ib_umem.h> 35 #include <rdma/ib_user_verbs.h> 36 #include <rdma/ib_cache.h> 37 #include "mlx5_ib.h" 38 39 static void mlx5_ib_cq_comp(struct mlx5_core_cq *cq) 40 { 41 struct ib_cq *ibcq = &to_mibcq(cq)->ibcq; 42 43 ibcq->comp_handler(ibcq, ibcq->cq_context); 44 } 45 46 static void mlx5_ib_cq_event(struct mlx5_core_cq *mcq, enum mlx5_event type) 47 { 48 struct mlx5_ib_cq *cq = container_of(mcq, struct mlx5_ib_cq, mcq); 49 struct mlx5_ib_dev *dev = to_mdev(cq->ibcq.device); 50 struct ib_cq *ibcq = &cq->ibcq; 51 struct ib_event event; 52 53 if (type != MLX5_EVENT_TYPE_CQ_ERROR) { 54 mlx5_ib_warn(dev, "Unexpected event type %d on CQ %06x\n", 55 type, mcq->cqn); 56 return; 57 } 58 59 if (ibcq->event_handler) { 60 event.device = &dev->ib_dev; 61 event.event = IB_EVENT_CQ_ERR; 62 event.element.cq = ibcq; 63 ibcq->event_handler(&event, ibcq->cq_context); 64 } 65 } 66 67 static void *get_cqe_from_buf(struct mlx5_ib_cq_buf *buf, int n, int size) 68 { 69 return mlx5_buf_offset(&buf->buf, n * size); 70 } 71 72 static void *get_cqe(struct mlx5_ib_cq *cq, int n) 73 { 74 return get_cqe_from_buf(&cq->buf, n, cq->mcq.cqe_sz); 75 } 76 77 static u8 sw_ownership_bit(int n, int nent) 78 { 79 return (n & nent) ? 1 : 0; 80 } 81 82 static void *get_sw_cqe(struct mlx5_ib_cq *cq, int n) 83 { 84 void *cqe = get_cqe(cq, n & cq->ibcq.cqe); 85 struct mlx5_cqe64 *cqe64; 86 87 cqe64 = (cq->mcq.cqe_sz == 64) ? cqe : cqe + 64; 88 89 if (likely((cqe64->op_own) >> 4 != MLX5_CQE_INVALID) && 90 !((cqe64->op_own & MLX5_CQE_OWNER_MASK) ^ !!(n & (cq->ibcq.cqe + 1)))) { 91 return cqe; 92 } else { 93 return NULL; 94 } 95 } 96 97 static void *next_cqe_sw(struct mlx5_ib_cq *cq) 98 { 99 return get_sw_cqe(cq, cq->mcq.cons_index); 100 } 101 102 static enum ib_wc_opcode get_umr_comp(struct mlx5_ib_wq *wq, int idx) 103 { 104 switch (wq->wr_data[idx]) { 105 case MLX5_IB_WR_UMR: 106 return 0; 107 108 case IB_WR_LOCAL_INV: 109 return IB_WC_LOCAL_INV; 110 111 case IB_WR_REG_MR: 112 return IB_WC_REG_MR; 113 114 default: 115 pr_warn("unknown completion status\n"); 116 return 0; 117 } 118 } 119 120 static void handle_good_req(struct ib_wc *wc, struct mlx5_cqe64 *cqe, 121 struct mlx5_ib_wq *wq, int idx) 122 { 123 wc->wc_flags = 0; 124 switch (be32_to_cpu(cqe->sop_drop_qpn) >> 24) { 125 case MLX5_OPCODE_RDMA_WRITE_IMM: 126 wc->wc_flags |= IB_WC_WITH_IMM; 127 case MLX5_OPCODE_RDMA_WRITE: 128 wc->opcode = IB_WC_RDMA_WRITE; 129 break; 130 case MLX5_OPCODE_SEND_IMM: 131 wc->wc_flags |= IB_WC_WITH_IMM; 132 case MLX5_OPCODE_SEND: 133 case MLX5_OPCODE_SEND_INVAL: 134 wc->opcode = IB_WC_SEND; 135 break; 136 case MLX5_OPCODE_RDMA_READ: 137 wc->opcode = IB_WC_RDMA_READ; 138 wc->byte_len = be32_to_cpu(cqe->byte_cnt); 139 break; 140 case MLX5_OPCODE_ATOMIC_CS: 141 wc->opcode = IB_WC_COMP_SWAP; 142 wc->byte_len = 8; 143 break; 144 case MLX5_OPCODE_ATOMIC_FA: 145 wc->opcode = IB_WC_FETCH_ADD; 146 wc->byte_len = 8; 147 break; 148 case MLX5_OPCODE_ATOMIC_MASKED_CS: 149 wc->opcode = IB_WC_MASKED_COMP_SWAP; 150 wc->byte_len = 8; 151 break; 152 case MLX5_OPCODE_ATOMIC_MASKED_FA: 153 wc->opcode = IB_WC_MASKED_FETCH_ADD; 154 wc->byte_len = 8; 155 break; 156 case MLX5_OPCODE_UMR: 157 wc->opcode = get_umr_comp(wq, idx); 158 break; 159 } 160 } 161 162 enum { 163 MLX5_GRH_IN_BUFFER = 1, 164 MLX5_GRH_IN_CQE = 2, 165 }; 166 167 static void handle_responder(struct ib_wc *wc, struct mlx5_cqe64 *cqe, 168 struct mlx5_ib_qp *qp) 169 { 170 enum rdma_link_layer ll = rdma_port_get_link_layer(qp->ibqp.device, 1); 171 struct mlx5_ib_dev *dev = to_mdev(qp->ibqp.device); 172 struct mlx5_ib_srq *srq; 173 struct mlx5_ib_wq *wq; 174 u16 wqe_ctr; 175 u8 g; 176 177 if (qp->ibqp.srq || qp->ibqp.xrcd) { 178 struct mlx5_core_srq *msrq = NULL; 179 180 if (qp->ibqp.xrcd) { 181 msrq = mlx5_core_get_srq(dev->mdev, 182 be32_to_cpu(cqe->srqn)); 183 srq = to_mibsrq(msrq); 184 } else { 185 srq = to_msrq(qp->ibqp.srq); 186 } 187 if (srq) { 188 wqe_ctr = be16_to_cpu(cqe->wqe_counter); 189 wc->wr_id = srq->wrid[wqe_ctr]; 190 mlx5_ib_free_srq_wqe(srq, wqe_ctr); 191 if (msrq && atomic_dec_and_test(&msrq->refcount)) 192 complete(&msrq->free); 193 } 194 } else { 195 wq = &qp->rq; 196 wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)]; 197 ++wq->tail; 198 } 199 wc->byte_len = be32_to_cpu(cqe->byte_cnt); 200 201 switch (cqe->op_own >> 4) { 202 case MLX5_CQE_RESP_WR_IMM: 203 wc->opcode = IB_WC_RECV_RDMA_WITH_IMM; 204 wc->wc_flags = IB_WC_WITH_IMM; 205 wc->ex.imm_data = cqe->imm_inval_pkey; 206 break; 207 case MLX5_CQE_RESP_SEND: 208 wc->opcode = IB_WC_RECV; 209 wc->wc_flags = IB_WC_IP_CSUM_OK; 210 if (unlikely(!((cqe->hds_ip_ext & CQE_L3_OK) && 211 (cqe->hds_ip_ext & CQE_L4_OK)))) 212 wc->wc_flags = 0; 213 break; 214 case MLX5_CQE_RESP_SEND_IMM: 215 wc->opcode = IB_WC_RECV; 216 wc->wc_flags = IB_WC_WITH_IMM; 217 wc->ex.imm_data = cqe->imm_inval_pkey; 218 break; 219 case MLX5_CQE_RESP_SEND_INV: 220 wc->opcode = IB_WC_RECV; 221 wc->wc_flags = IB_WC_WITH_INVALIDATE; 222 wc->ex.invalidate_rkey = be32_to_cpu(cqe->imm_inval_pkey); 223 break; 224 } 225 wc->slid = be16_to_cpu(cqe->slid); 226 wc->sl = (be32_to_cpu(cqe->flags_rqpn) >> 24) & 0xf; 227 wc->src_qp = be32_to_cpu(cqe->flags_rqpn) & 0xffffff; 228 wc->dlid_path_bits = cqe->ml_path; 229 g = (be32_to_cpu(cqe->flags_rqpn) >> 28) & 3; 230 wc->wc_flags |= g ? IB_WC_GRH : 0; 231 if (unlikely(is_qp1(qp->ibqp.qp_type))) { 232 u16 pkey = be32_to_cpu(cqe->imm_inval_pkey) & 0xffff; 233 234 ib_find_cached_pkey(&dev->ib_dev, qp->port, pkey, 235 &wc->pkey_index); 236 } else { 237 wc->pkey_index = 0; 238 } 239 240 if (ll != IB_LINK_LAYER_ETHERNET) 241 return; 242 243 switch (wc->sl & 0x3) { 244 case MLX5_CQE_ROCE_L3_HEADER_TYPE_GRH: 245 wc->network_hdr_type = RDMA_NETWORK_IB; 246 break; 247 case MLX5_CQE_ROCE_L3_HEADER_TYPE_IPV6: 248 wc->network_hdr_type = RDMA_NETWORK_IPV6; 249 break; 250 case MLX5_CQE_ROCE_L3_HEADER_TYPE_IPV4: 251 wc->network_hdr_type = RDMA_NETWORK_IPV4; 252 break; 253 } 254 wc->wc_flags |= IB_WC_WITH_NETWORK_HDR_TYPE; 255 } 256 257 static void dump_cqe(struct mlx5_ib_dev *dev, struct mlx5_err_cqe *cqe) 258 { 259 __be32 *p = (__be32 *)cqe; 260 int i; 261 262 mlx5_ib_warn(dev, "dump error cqe\n"); 263 for (i = 0; i < sizeof(*cqe) / 16; i++, p += 4) 264 pr_info("%08x %08x %08x %08x\n", be32_to_cpu(p[0]), 265 be32_to_cpu(p[1]), be32_to_cpu(p[2]), 266 be32_to_cpu(p[3])); 267 } 268 269 static void mlx5_handle_error_cqe(struct mlx5_ib_dev *dev, 270 struct mlx5_err_cqe *cqe, 271 struct ib_wc *wc) 272 { 273 int dump = 1; 274 275 switch (cqe->syndrome) { 276 case MLX5_CQE_SYNDROME_LOCAL_LENGTH_ERR: 277 wc->status = IB_WC_LOC_LEN_ERR; 278 break; 279 case MLX5_CQE_SYNDROME_LOCAL_QP_OP_ERR: 280 wc->status = IB_WC_LOC_QP_OP_ERR; 281 break; 282 case MLX5_CQE_SYNDROME_LOCAL_PROT_ERR: 283 wc->status = IB_WC_LOC_PROT_ERR; 284 break; 285 case MLX5_CQE_SYNDROME_WR_FLUSH_ERR: 286 dump = 0; 287 wc->status = IB_WC_WR_FLUSH_ERR; 288 break; 289 case MLX5_CQE_SYNDROME_MW_BIND_ERR: 290 wc->status = IB_WC_MW_BIND_ERR; 291 break; 292 case MLX5_CQE_SYNDROME_BAD_RESP_ERR: 293 wc->status = IB_WC_BAD_RESP_ERR; 294 break; 295 case MLX5_CQE_SYNDROME_LOCAL_ACCESS_ERR: 296 wc->status = IB_WC_LOC_ACCESS_ERR; 297 break; 298 case MLX5_CQE_SYNDROME_REMOTE_INVAL_REQ_ERR: 299 wc->status = IB_WC_REM_INV_REQ_ERR; 300 break; 301 case MLX5_CQE_SYNDROME_REMOTE_ACCESS_ERR: 302 wc->status = IB_WC_REM_ACCESS_ERR; 303 break; 304 case MLX5_CQE_SYNDROME_REMOTE_OP_ERR: 305 wc->status = IB_WC_REM_OP_ERR; 306 break; 307 case MLX5_CQE_SYNDROME_TRANSPORT_RETRY_EXC_ERR: 308 wc->status = IB_WC_RETRY_EXC_ERR; 309 dump = 0; 310 break; 311 case MLX5_CQE_SYNDROME_RNR_RETRY_EXC_ERR: 312 wc->status = IB_WC_RNR_RETRY_EXC_ERR; 313 dump = 0; 314 break; 315 case MLX5_CQE_SYNDROME_REMOTE_ABORTED_ERR: 316 wc->status = IB_WC_REM_ABORT_ERR; 317 break; 318 default: 319 wc->status = IB_WC_GENERAL_ERR; 320 break; 321 } 322 323 wc->vendor_err = cqe->vendor_err_synd; 324 if (dump) 325 dump_cqe(dev, cqe); 326 } 327 328 static int is_atomic_response(struct mlx5_ib_qp *qp, uint16_t idx) 329 { 330 /* TBD: waiting decision 331 */ 332 return 0; 333 } 334 335 static void *mlx5_get_atomic_laddr(struct mlx5_ib_qp *qp, uint16_t idx) 336 { 337 struct mlx5_wqe_data_seg *dpseg; 338 void *addr; 339 340 dpseg = mlx5_get_send_wqe(qp, idx) + sizeof(struct mlx5_wqe_ctrl_seg) + 341 sizeof(struct mlx5_wqe_raddr_seg) + 342 sizeof(struct mlx5_wqe_atomic_seg); 343 addr = (void *)(unsigned long)be64_to_cpu(dpseg->addr); 344 return addr; 345 } 346 347 static void handle_atomic(struct mlx5_ib_qp *qp, struct mlx5_cqe64 *cqe64, 348 uint16_t idx) 349 { 350 void *addr; 351 int byte_count; 352 int i; 353 354 if (!is_atomic_response(qp, idx)) 355 return; 356 357 byte_count = be32_to_cpu(cqe64->byte_cnt); 358 addr = mlx5_get_atomic_laddr(qp, idx); 359 360 if (byte_count == 4) { 361 *(uint32_t *)addr = be32_to_cpu(*((__be32 *)addr)); 362 } else { 363 for (i = 0; i < byte_count; i += 8) { 364 *(uint64_t *)addr = be64_to_cpu(*((__be64 *)addr)); 365 addr += 8; 366 } 367 } 368 369 return; 370 } 371 372 static void handle_atomics(struct mlx5_ib_qp *qp, struct mlx5_cqe64 *cqe64, 373 u16 tail, u16 head) 374 { 375 u16 idx; 376 377 do { 378 idx = tail & (qp->sq.wqe_cnt - 1); 379 handle_atomic(qp, cqe64, idx); 380 if (idx == head) 381 break; 382 383 tail = qp->sq.w_list[idx].next; 384 } while (1); 385 tail = qp->sq.w_list[idx].next; 386 qp->sq.last_poll = tail; 387 } 388 389 static void free_cq_buf(struct mlx5_ib_dev *dev, struct mlx5_ib_cq_buf *buf) 390 { 391 mlx5_buf_free(dev->mdev, &buf->buf); 392 } 393 394 static void get_sig_err_item(struct mlx5_sig_err_cqe *cqe, 395 struct ib_sig_err *item) 396 { 397 u16 syndrome = be16_to_cpu(cqe->syndrome); 398 399 #define GUARD_ERR (1 << 13) 400 #define APPTAG_ERR (1 << 12) 401 #define REFTAG_ERR (1 << 11) 402 403 if (syndrome & GUARD_ERR) { 404 item->err_type = IB_SIG_BAD_GUARD; 405 item->expected = be32_to_cpu(cqe->expected_trans_sig) >> 16; 406 item->actual = be32_to_cpu(cqe->actual_trans_sig) >> 16; 407 } else 408 if (syndrome & REFTAG_ERR) { 409 item->err_type = IB_SIG_BAD_REFTAG; 410 item->expected = be32_to_cpu(cqe->expected_reftag); 411 item->actual = be32_to_cpu(cqe->actual_reftag); 412 } else 413 if (syndrome & APPTAG_ERR) { 414 item->err_type = IB_SIG_BAD_APPTAG; 415 item->expected = be32_to_cpu(cqe->expected_trans_sig) & 0xffff; 416 item->actual = be32_to_cpu(cqe->actual_trans_sig) & 0xffff; 417 } else { 418 pr_err("Got signature completion error with bad syndrome %04x\n", 419 syndrome); 420 } 421 422 item->sig_err_offset = be64_to_cpu(cqe->err_offset); 423 item->key = be32_to_cpu(cqe->mkey); 424 } 425 426 static void sw_send_comp(struct mlx5_ib_qp *qp, int num_entries, 427 struct ib_wc *wc, int *npolled) 428 { 429 struct mlx5_ib_wq *wq; 430 unsigned int cur; 431 unsigned int idx; 432 int np; 433 int i; 434 435 wq = &qp->sq; 436 cur = wq->head - wq->tail; 437 np = *npolled; 438 439 if (cur == 0) 440 return; 441 442 for (i = 0; i < cur && np < num_entries; i++) { 443 idx = wq->last_poll & (wq->wqe_cnt - 1); 444 wc->wr_id = wq->wrid[idx]; 445 wc->status = IB_WC_WR_FLUSH_ERR; 446 wc->vendor_err = MLX5_CQE_SYNDROME_WR_FLUSH_ERR; 447 wq->tail++; 448 np++; 449 wc->qp = &qp->ibqp; 450 wc++; 451 wq->last_poll = wq->w_list[idx].next; 452 } 453 *npolled = np; 454 } 455 456 static void sw_recv_comp(struct mlx5_ib_qp *qp, int num_entries, 457 struct ib_wc *wc, int *npolled) 458 { 459 struct mlx5_ib_wq *wq; 460 unsigned int cur; 461 int np; 462 int i; 463 464 wq = &qp->rq; 465 cur = wq->head - wq->tail; 466 np = *npolled; 467 468 if (cur == 0) 469 return; 470 471 for (i = 0; i < cur && np < num_entries; i++) { 472 wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)]; 473 wc->status = IB_WC_WR_FLUSH_ERR; 474 wc->vendor_err = MLX5_CQE_SYNDROME_WR_FLUSH_ERR; 475 wq->tail++; 476 np++; 477 wc->qp = &qp->ibqp; 478 wc++; 479 } 480 *npolled = np; 481 } 482 483 static void mlx5_ib_poll_sw_comp(struct mlx5_ib_cq *cq, int num_entries, 484 struct ib_wc *wc, int *npolled) 485 { 486 struct mlx5_ib_qp *qp; 487 488 *npolled = 0; 489 /* Find uncompleted WQEs belonging to that cq and retrun mmics ones */ 490 list_for_each_entry(qp, &cq->list_send_qp, cq_send_list) { 491 sw_send_comp(qp, num_entries, wc + *npolled, npolled); 492 if (*npolled >= num_entries) 493 return; 494 } 495 496 list_for_each_entry(qp, &cq->list_recv_qp, cq_recv_list) { 497 sw_recv_comp(qp, num_entries, wc + *npolled, npolled); 498 if (*npolled >= num_entries) 499 return; 500 } 501 } 502 503 static int mlx5_poll_one(struct mlx5_ib_cq *cq, 504 struct mlx5_ib_qp **cur_qp, 505 struct ib_wc *wc) 506 { 507 struct mlx5_ib_dev *dev = to_mdev(cq->ibcq.device); 508 struct mlx5_err_cqe *err_cqe; 509 struct mlx5_cqe64 *cqe64; 510 struct mlx5_core_qp *mqp; 511 struct mlx5_ib_wq *wq; 512 struct mlx5_sig_err_cqe *sig_err_cqe; 513 struct mlx5_core_mkey *mmkey; 514 struct mlx5_ib_mr *mr; 515 uint8_t opcode; 516 uint32_t qpn; 517 u16 wqe_ctr; 518 void *cqe; 519 int idx; 520 521 repoll: 522 cqe = next_cqe_sw(cq); 523 if (!cqe) 524 return -EAGAIN; 525 526 cqe64 = (cq->mcq.cqe_sz == 64) ? cqe : cqe + 64; 527 528 ++cq->mcq.cons_index; 529 530 /* Make sure we read CQ entry contents after we've checked the 531 * ownership bit. 532 */ 533 rmb(); 534 535 opcode = cqe64->op_own >> 4; 536 if (unlikely(opcode == MLX5_CQE_RESIZE_CQ)) { 537 if (likely(cq->resize_buf)) { 538 free_cq_buf(dev, &cq->buf); 539 cq->buf = *cq->resize_buf; 540 kfree(cq->resize_buf); 541 cq->resize_buf = NULL; 542 goto repoll; 543 } else { 544 mlx5_ib_warn(dev, "unexpected resize cqe\n"); 545 } 546 } 547 548 qpn = ntohl(cqe64->sop_drop_qpn) & 0xffffff; 549 if (!*cur_qp || (qpn != (*cur_qp)->ibqp.qp_num)) { 550 /* We do not have to take the QP table lock here, 551 * because CQs will be locked while QPs are removed 552 * from the table. 553 */ 554 mqp = __mlx5_qp_lookup(dev->mdev, qpn); 555 *cur_qp = to_mibqp(mqp); 556 } 557 558 wc->qp = &(*cur_qp)->ibqp; 559 switch (opcode) { 560 case MLX5_CQE_REQ: 561 wq = &(*cur_qp)->sq; 562 wqe_ctr = be16_to_cpu(cqe64->wqe_counter); 563 idx = wqe_ctr & (wq->wqe_cnt - 1); 564 handle_good_req(wc, cqe64, wq, idx); 565 handle_atomics(*cur_qp, cqe64, wq->last_poll, idx); 566 wc->wr_id = wq->wrid[idx]; 567 wq->tail = wq->wqe_head[idx] + 1; 568 wc->status = IB_WC_SUCCESS; 569 break; 570 case MLX5_CQE_RESP_WR_IMM: 571 case MLX5_CQE_RESP_SEND: 572 case MLX5_CQE_RESP_SEND_IMM: 573 case MLX5_CQE_RESP_SEND_INV: 574 handle_responder(wc, cqe64, *cur_qp); 575 wc->status = IB_WC_SUCCESS; 576 break; 577 case MLX5_CQE_RESIZE_CQ: 578 break; 579 case MLX5_CQE_REQ_ERR: 580 case MLX5_CQE_RESP_ERR: 581 err_cqe = (struct mlx5_err_cqe *)cqe64; 582 mlx5_handle_error_cqe(dev, err_cqe, wc); 583 mlx5_ib_dbg(dev, "%s error cqe on cqn 0x%x:\n", 584 opcode == MLX5_CQE_REQ_ERR ? 585 "Requestor" : "Responder", cq->mcq.cqn); 586 mlx5_ib_dbg(dev, "syndrome 0x%x, vendor syndrome 0x%x\n", 587 err_cqe->syndrome, err_cqe->vendor_err_synd); 588 if (opcode == MLX5_CQE_REQ_ERR) { 589 wq = &(*cur_qp)->sq; 590 wqe_ctr = be16_to_cpu(cqe64->wqe_counter); 591 idx = wqe_ctr & (wq->wqe_cnt - 1); 592 wc->wr_id = wq->wrid[idx]; 593 wq->tail = wq->wqe_head[idx] + 1; 594 } else { 595 struct mlx5_ib_srq *srq; 596 597 if ((*cur_qp)->ibqp.srq) { 598 srq = to_msrq((*cur_qp)->ibqp.srq); 599 wqe_ctr = be16_to_cpu(cqe64->wqe_counter); 600 wc->wr_id = srq->wrid[wqe_ctr]; 601 mlx5_ib_free_srq_wqe(srq, wqe_ctr); 602 } else { 603 wq = &(*cur_qp)->rq; 604 wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)]; 605 ++wq->tail; 606 } 607 } 608 break; 609 case MLX5_CQE_SIG_ERR: 610 sig_err_cqe = (struct mlx5_sig_err_cqe *)cqe64; 611 612 read_lock(&dev->mdev->priv.mkey_table.lock); 613 mmkey = __mlx5_mr_lookup(dev->mdev, 614 mlx5_base_mkey(be32_to_cpu(sig_err_cqe->mkey))); 615 mr = to_mibmr(mmkey); 616 get_sig_err_item(sig_err_cqe, &mr->sig->err_item); 617 mr->sig->sig_err_exists = true; 618 mr->sig->sigerr_count++; 619 620 mlx5_ib_warn(dev, "CQN: 0x%x Got SIGERR on key: 0x%x err_type %x err_offset %llx expected %x actual %x\n", 621 cq->mcq.cqn, mr->sig->err_item.key, 622 mr->sig->err_item.err_type, 623 mr->sig->err_item.sig_err_offset, 624 mr->sig->err_item.expected, 625 mr->sig->err_item.actual); 626 627 read_unlock(&dev->mdev->priv.mkey_table.lock); 628 goto repoll; 629 } 630 631 return 0; 632 } 633 634 static int poll_soft_wc(struct mlx5_ib_cq *cq, int num_entries, 635 struct ib_wc *wc) 636 { 637 struct mlx5_ib_dev *dev = to_mdev(cq->ibcq.device); 638 struct mlx5_ib_wc *soft_wc, *next; 639 int npolled = 0; 640 641 list_for_each_entry_safe(soft_wc, next, &cq->wc_list, list) { 642 if (npolled >= num_entries) 643 break; 644 645 mlx5_ib_dbg(dev, "polled software generated completion on CQ 0x%x\n", 646 cq->mcq.cqn); 647 648 wc[npolled++] = soft_wc->wc; 649 list_del(&soft_wc->list); 650 kfree(soft_wc); 651 } 652 653 return npolled; 654 } 655 656 int mlx5_ib_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc) 657 { 658 struct mlx5_ib_cq *cq = to_mcq(ibcq); 659 struct mlx5_ib_qp *cur_qp = NULL; 660 struct mlx5_ib_dev *dev = to_mdev(cq->ibcq.device); 661 struct mlx5_core_dev *mdev = dev->mdev; 662 unsigned long flags; 663 int soft_polled = 0; 664 int npolled; 665 666 spin_lock_irqsave(&cq->lock, flags); 667 if (mdev->state == MLX5_DEVICE_STATE_INTERNAL_ERROR) { 668 mlx5_ib_poll_sw_comp(cq, num_entries, wc, &npolled); 669 goto out; 670 } 671 672 if (unlikely(!list_empty(&cq->wc_list))) 673 soft_polled = poll_soft_wc(cq, num_entries, wc); 674 675 for (npolled = 0; npolled < num_entries - soft_polled; npolled++) { 676 if (mlx5_poll_one(cq, &cur_qp, wc + soft_polled + npolled)) 677 break; 678 } 679 680 if (npolled) 681 mlx5_cq_set_ci(&cq->mcq); 682 out: 683 spin_unlock_irqrestore(&cq->lock, flags); 684 685 return soft_polled + npolled; 686 } 687 688 int mlx5_ib_arm_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags flags) 689 { 690 struct mlx5_core_dev *mdev = to_mdev(ibcq->device)->mdev; 691 struct mlx5_ib_cq *cq = to_mcq(ibcq); 692 void __iomem *uar_page = mdev->priv.uuari.uars[0].map; 693 unsigned long irq_flags; 694 int ret = 0; 695 696 spin_lock_irqsave(&cq->lock, irq_flags); 697 if (cq->notify_flags != IB_CQ_NEXT_COMP) 698 cq->notify_flags = flags & IB_CQ_SOLICITED_MASK; 699 700 if ((flags & IB_CQ_REPORT_MISSED_EVENTS) && !list_empty(&cq->wc_list)) 701 ret = 1; 702 spin_unlock_irqrestore(&cq->lock, irq_flags); 703 704 mlx5_cq_arm(&cq->mcq, 705 (flags & IB_CQ_SOLICITED_MASK) == IB_CQ_SOLICITED ? 706 MLX5_CQ_DB_REQ_NOT_SOL : MLX5_CQ_DB_REQ_NOT, 707 uar_page, 708 MLX5_GET_DOORBELL_LOCK(&mdev->priv.cq_uar_lock), 709 to_mcq(ibcq)->mcq.cons_index); 710 711 return ret; 712 } 713 714 static int alloc_cq_buf(struct mlx5_ib_dev *dev, struct mlx5_ib_cq_buf *buf, 715 int nent, int cqe_size) 716 { 717 int err; 718 719 err = mlx5_buf_alloc(dev->mdev, nent * cqe_size, &buf->buf); 720 if (err) 721 return err; 722 723 buf->cqe_size = cqe_size; 724 buf->nent = nent; 725 726 return 0; 727 } 728 729 static int create_cq_user(struct mlx5_ib_dev *dev, struct ib_udata *udata, 730 struct ib_ucontext *context, struct mlx5_ib_cq *cq, 731 int entries, u32 **cqb, 732 int *cqe_size, int *index, int *inlen) 733 { 734 struct mlx5_ib_create_cq ucmd = {}; 735 size_t ucmdlen; 736 int page_shift; 737 __be64 *pas; 738 int npages; 739 int ncont; 740 void *cqc; 741 int err; 742 743 ucmdlen = 744 (udata->inlen - sizeof(struct ib_uverbs_cmd_hdr) < 745 sizeof(ucmd)) ? (sizeof(ucmd) - 746 sizeof(ucmd.reserved)) : sizeof(ucmd); 747 748 if (ib_copy_from_udata(&ucmd, udata, ucmdlen)) 749 return -EFAULT; 750 751 if (ucmdlen == sizeof(ucmd) && 752 ucmd.reserved != 0) 753 return -EINVAL; 754 755 if (ucmd.cqe_size != 64 && ucmd.cqe_size != 128) 756 return -EINVAL; 757 758 *cqe_size = ucmd.cqe_size; 759 760 cq->buf.umem = ib_umem_get(context, ucmd.buf_addr, 761 entries * ucmd.cqe_size, 762 IB_ACCESS_LOCAL_WRITE, 1); 763 if (IS_ERR(cq->buf.umem)) { 764 err = PTR_ERR(cq->buf.umem); 765 return err; 766 } 767 768 err = mlx5_ib_db_map_user(to_mucontext(context), ucmd.db_addr, 769 &cq->db); 770 if (err) 771 goto err_umem; 772 773 mlx5_ib_cont_pages(cq->buf.umem, ucmd.buf_addr, 0, &npages, &page_shift, 774 &ncont, NULL); 775 mlx5_ib_dbg(dev, "addr 0x%llx, size %u, npages %d, page_shift %d, ncont %d\n", 776 ucmd.buf_addr, entries * ucmd.cqe_size, npages, page_shift, ncont); 777 778 *inlen = MLX5_ST_SZ_BYTES(create_cq_in) + 779 MLX5_FLD_SZ_BYTES(create_cq_in, pas[0]) * ncont; 780 *cqb = mlx5_vzalloc(*inlen); 781 if (!*cqb) { 782 err = -ENOMEM; 783 goto err_db; 784 } 785 786 pas = (__be64 *)MLX5_ADDR_OF(create_cq_in, *cqb, pas); 787 mlx5_ib_populate_pas(dev, cq->buf.umem, page_shift, pas, 0); 788 789 cqc = MLX5_ADDR_OF(create_cq_in, *cqb, cq_context); 790 MLX5_SET(cqc, cqc, log_page_size, 791 page_shift - MLX5_ADAPTER_PAGE_SHIFT); 792 793 *index = to_mucontext(context)->uuari.uars[0].index; 794 795 if (ucmd.cqe_comp_en == 1) { 796 if (unlikely((*cqe_size != 64) || 797 !MLX5_CAP_GEN(dev->mdev, cqe_compression))) { 798 err = -EOPNOTSUPP; 799 mlx5_ib_warn(dev, "CQE compression is not supported for size %d!\n", 800 *cqe_size); 801 goto err_cqb; 802 } 803 804 if (unlikely(!ucmd.cqe_comp_res_format || 805 !(ucmd.cqe_comp_res_format < 806 MLX5_IB_CQE_RES_RESERVED) || 807 (ucmd.cqe_comp_res_format & 808 (ucmd.cqe_comp_res_format - 1)))) { 809 err = -EOPNOTSUPP; 810 mlx5_ib_warn(dev, "CQE compression res format %d is not supported!\n", 811 ucmd.cqe_comp_res_format); 812 goto err_cqb; 813 } 814 815 MLX5_SET(cqc, cqc, cqe_comp_en, 1); 816 MLX5_SET(cqc, cqc, mini_cqe_res_format, 817 ilog2(ucmd.cqe_comp_res_format)); 818 } 819 820 return 0; 821 822 err_cqb: 823 kfree(cqb); 824 825 err_db: 826 mlx5_ib_db_unmap_user(to_mucontext(context), &cq->db); 827 828 err_umem: 829 ib_umem_release(cq->buf.umem); 830 return err; 831 } 832 833 static void destroy_cq_user(struct mlx5_ib_cq *cq, struct ib_ucontext *context) 834 { 835 mlx5_ib_db_unmap_user(to_mucontext(context), &cq->db); 836 ib_umem_release(cq->buf.umem); 837 } 838 839 static void init_cq_buf(struct mlx5_ib_cq *cq, struct mlx5_ib_cq_buf *buf) 840 { 841 int i; 842 void *cqe; 843 struct mlx5_cqe64 *cqe64; 844 845 for (i = 0; i < buf->nent; i++) { 846 cqe = get_cqe_from_buf(buf, i, buf->cqe_size); 847 cqe64 = buf->cqe_size == 64 ? cqe : cqe + 64; 848 cqe64->op_own = MLX5_CQE_INVALID << 4; 849 } 850 } 851 852 static int create_cq_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *cq, 853 int entries, int cqe_size, 854 u32 **cqb, int *index, int *inlen) 855 { 856 __be64 *pas; 857 void *cqc; 858 int err; 859 860 err = mlx5_db_alloc(dev->mdev, &cq->db); 861 if (err) 862 return err; 863 864 cq->mcq.set_ci_db = cq->db.db; 865 cq->mcq.arm_db = cq->db.db + 1; 866 cq->mcq.cqe_sz = cqe_size; 867 868 err = alloc_cq_buf(dev, &cq->buf, entries, cqe_size); 869 if (err) 870 goto err_db; 871 872 init_cq_buf(cq, &cq->buf); 873 874 *inlen = MLX5_ST_SZ_BYTES(create_cq_in) + 875 MLX5_FLD_SZ_BYTES(create_cq_in, pas[0]) * cq->buf.buf.npages; 876 *cqb = mlx5_vzalloc(*inlen); 877 if (!*cqb) { 878 err = -ENOMEM; 879 goto err_buf; 880 } 881 882 pas = (__be64 *)MLX5_ADDR_OF(create_cq_in, *cqb, pas); 883 mlx5_fill_page_array(&cq->buf.buf, pas); 884 885 cqc = MLX5_ADDR_OF(create_cq_in, *cqb, cq_context); 886 MLX5_SET(cqc, cqc, log_page_size, 887 cq->buf.buf.page_shift - MLX5_ADAPTER_PAGE_SHIFT); 888 889 *index = dev->mdev->priv.uuari.uars[0].index; 890 891 return 0; 892 893 err_buf: 894 free_cq_buf(dev, &cq->buf); 895 896 err_db: 897 mlx5_db_free(dev->mdev, &cq->db); 898 return err; 899 } 900 901 static void destroy_cq_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *cq) 902 { 903 free_cq_buf(dev, &cq->buf); 904 mlx5_db_free(dev->mdev, &cq->db); 905 } 906 907 static void notify_soft_wc_handler(struct work_struct *work) 908 { 909 struct mlx5_ib_cq *cq = container_of(work, struct mlx5_ib_cq, 910 notify_work); 911 912 cq->ibcq.comp_handler(&cq->ibcq, cq->ibcq.cq_context); 913 } 914 915 struct ib_cq *mlx5_ib_create_cq(struct ib_device *ibdev, 916 const struct ib_cq_init_attr *attr, 917 struct ib_ucontext *context, 918 struct ib_udata *udata) 919 { 920 int entries = attr->cqe; 921 int vector = attr->comp_vector; 922 struct mlx5_ib_dev *dev = to_mdev(ibdev); 923 struct mlx5_ib_cq *cq; 924 int uninitialized_var(index); 925 int uninitialized_var(inlen); 926 u32 *cqb = NULL; 927 void *cqc; 928 int cqe_size; 929 unsigned int irqn; 930 int eqn; 931 int err; 932 933 if (entries < 0 || 934 (entries > (1 << MLX5_CAP_GEN(dev->mdev, log_max_cq_sz)))) 935 return ERR_PTR(-EINVAL); 936 937 if (check_cq_create_flags(attr->flags)) 938 return ERR_PTR(-EOPNOTSUPP); 939 940 entries = roundup_pow_of_two(entries + 1); 941 if (entries > (1 << MLX5_CAP_GEN(dev->mdev, log_max_cq_sz))) 942 return ERR_PTR(-EINVAL); 943 944 cq = kzalloc(sizeof(*cq), GFP_KERNEL); 945 if (!cq) 946 return ERR_PTR(-ENOMEM); 947 948 cq->ibcq.cqe = entries - 1; 949 mutex_init(&cq->resize_mutex); 950 spin_lock_init(&cq->lock); 951 cq->resize_buf = NULL; 952 cq->resize_umem = NULL; 953 cq->create_flags = attr->flags; 954 INIT_LIST_HEAD(&cq->list_send_qp); 955 INIT_LIST_HEAD(&cq->list_recv_qp); 956 957 if (context) { 958 err = create_cq_user(dev, udata, context, cq, entries, 959 &cqb, &cqe_size, &index, &inlen); 960 if (err) 961 goto err_create; 962 } else { 963 cqe_size = cache_line_size() == 128 ? 128 : 64; 964 err = create_cq_kernel(dev, cq, entries, cqe_size, &cqb, 965 &index, &inlen); 966 if (err) 967 goto err_create; 968 969 INIT_WORK(&cq->notify_work, notify_soft_wc_handler); 970 } 971 972 err = mlx5_vector2eqn(dev->mdev, vector, &eqn, &irqn); 973 if (err) 974 goto err_cqb; 975 976 cq->cqe_size = cqe_size; 977 978 cqc = MLX5_ADDR_OF(create_cq_in, cqb, cq_context); 979 MLX5_SET(cqc, cqc, cqe_sz, cqe_sz_to_mlx_sz(cqe_size)); 980 MLX5_SET(cqc, cqc, log_cq_size, ilog2(entries)); 981 MLX5_SET(cqc, cqc, uar_page, index); 982 MLX5_SET(cqc, cqc, c_eqn, eqn); 983 MLX5_SET64(cqc, cqc, dbr_addr, cq->db.dma); 984 if (cq->create_flags & IB_CQ_FLAGS_IGNORE_OVERRUN) 985 MLX5_SET(cqc, cqc, oi, 1); 986 987 err = mlx5_core_create_cq(dev->mdev, &cq->mcq, cqb, inlen); 988 if (err) 989 goto err_cqb; 990 991 mlx5_ib_dbg(dev, "cqn 0x%x\n", cq->mcq.cqn); 992 cq->mcq.irqn = irqn; 993 if (context) 994 cq->mcq.tasklet_ctx.comp = mlx5_ib_cq_comp; 995 else 996 cq->mcq.comp = mlx5_ib_cq_comp; 997 cq->mcq.event = mlx5_ib_cq_event; 998 999 INIT_LIST_HEAD(&cq->wc_list); 1000 1001 if (context) 1002 if (ib_copy_to_udata(udata, &cq->mcq.cqn, sizeof(__u32))) { 1003 err = -EFAULT; 1004 goto err_cmd; 1005 } 1006 1007 1008 kvfree(cqb); 1009 return &cq->ibcq; 1010 1011 err_cmd: 1012 mlx5_core_destroy_cq(dev->mdev, &cq->mcq); 1013 1014 err_cqb: 1015 kvfree(cqb); 1016 if (context) 1017 destroy_cq_user(cq, context); 1018 else 1019 destroy_cq_kernel(dev, cq); 1020 1021 err_create: 1022 kfree(cq); 1023 1024 return ERR_PTR(err); 1025 } 1026 1027 1028 int mlx5_ib_destroy_cq(struct ib_cq *cq) 1029 { 1030 struct mlx5_ib_dev *dev = to_mdev(cq->device); 1031 struct mlx5_ib_cq *mcq = to_mcq(cq); 1032 struct ib_ucontext *context = NULL; 1033 1034 if (cq->uobject) 1035 context = cq->uobject->context; 1036 1037 mlx5_core_destroy_cq(dev->mdev, &mcq->mcq); 1038 if (context) 1039 destroy_cq_user(mcq, context); 1040 else 1041 destroy_cq_kernel(dev, mcq); 1042 1043 kfree(mcq); 1044 1045 return 0; 1046 } 1047 1048 static int is_equal_rsn(struct mlx5_cqe64 *cqe64, u32 rsn) 1049 { 1050 return rsn == (ntohl(cqe64->sop_drop_qpn) & 0xffffff); 1051 } 1052 1053 void __mlx5_ib_cq_clean(struct mlx5_ib_cq *cq, u32 rsn, struct mlx5_ib_srq *srq) 1054 { 1055 struct mlx5_cqe64 *cqe64, *dest64; 1056 void *cqe, *dest; 1057 u32 prod_index; 1058 int nfreed = 0; 1059 u8 owner_bit; 1060 1061 if (!cq) 1062 return; 1063 1064 /* First we need to find the current producer index, so we 1065 * know where to start cleaning from. It doesn't matter if HW 1066 * adds new entries after this loop -- the QP we're worried 1067 * about is already in RESET, so the new entries won't come 1068 * from our QP and therefore don't need to be checked. 1069 */ 1070 for (prod_index = cq->mcq.cons_index; get_sw_cqe(cq, prod_index); prod_index++) 1071 if (prod_index == cq->mcq.cons_index + cq->ibcq.cqe) 1072 break; 1073 1074 /* Now sweep backwards through the CQ, removing CQ entries 1075 * that match our QP by copying older entries on top of them. 1076 */ 1077 while ((int) --prod_index - (int) cq->mcq.cons_index >= 0) { 1078 cqe = get_cqe(cq, prod_index & cq->ibcq.cqe); 1079 cqe64 = (cq->mcq.cqe_sz == 64) ? cqe : cqe + 64; 1080 if (is_equal_rsn(cqe64, rsn)) { 1081 if (srq && (ntohl(cqe64->srqn) & 0xffffff)) 1082 mlx5_ib_free_srq_wqe(srq, be16_to_cpu(cqe64->wqe_counter)); 1083 ++nfreed; 1084 } else if (nfreed) { 1085 dest = get_cqe(cq, (prod_index + nfreed) & cq->ibcq.cqe); 1086 dest64 = (cq->mcq.cqe_sz == 64) ? dest : dest + 64; 1087 owner_bit = dest64->op_own & MLX5_CQE_OWNER_MASK; 1088 memcpy(dest, cqe, cq->mcq.cqe_sz); 1089 dest64->op_own = owner_bit | 1090 (dest64->op_own & ~MLX5_CQE_OWNER_MASK); 1091 } 1092 } 1093 1094 if (nfreed) { 1095 cq->mcq.cons_index += nfreed; 1096 /* Make sure update of buffer contents is done before 1097 * updating consumer index. 1098 */ 1099 wmb(); 1100 mlx5_cq_set_ci(&cq->mcq); 1101 } 1102 } 1103 1104 void mlx5_ib_cq_clean(struct mlx5_ib_cq *cq, u32 qpn, struct mlx5_ib_srq *srq) 1105 { 1106 if (!cq) 1107 return; 1108 1109 spin_lock_irq(&cq->lock); 1110 __mlx5_ib_cq_clean(cq, qpn, srq); 1111 spin_unlock_irq(&cq->lock); 1112 } 1113 1114 int mlx5_ib_modify_cq(struct ib_cq *cq, u16 cq_count, u16 cq_period) 1115 { 1116 struct mlx5_ib_dev *dev = to_mdev(cq->device); 1117 struct mlx5_ib_cq *mcq = to_mcq(cq); 1118 int err; 1119 1120 if (!MLX5_CAP_GEN(dev->mdev, cq_moderation)) 1121 return -ENOSYS; 1122 1123 err = mlx5_core_modify_cq_moderation(dev->mdev, &mcq->mcq, 1124 cq_period, cq_count); 1125 if (err) 1126 mlx5_ib_warn(dev, "modify cq 0x%x failed\n", mcq->mcq.cqn); 1127 1128 return err; 1129 } 1130 1131 static int resize_user(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *cq, 1132 int entries, struct ib_udata *udata, int *npas, 1133 int *page_shift, int *cqe_size) 1134 { 1135 struct mlx5_ib_resize_cq ucmd; 1136 struct ib_umem *umem; 1137 int err; 1138 int npages; 1139 struct ib_ucontext *context = cq->buf.umem->context; 1140 1141 err = ib_copy_from_udata(&ucmd, udata, sizeof(ucmd)); 1142 if (err) 1143 return err; 1144 1145 if (ucmd.reserved0 || ucmd.reserved1) 1146 return -EINVAL; 1147 1148 umem = ib_umem_get(context, ucmd.buf_addr, entries * ucmd.cqe_size, 1149 IB_ACCESS_LOCAL_WRITE, 1); 1150 if (IS_ERR(umem)) { 1151 err = PTR_ERR(umem); 1152 return err; 1153 } 1154 1155 mlx5_ib_cont_pages(umem, ucmd.buf_addr, 0, &npages, page_shift, 1156 npas, NULL); 1157 1158 cq->resize_umem = umem; 1159 *cqe_size = ucmd.cqe_size; 1160 1161 return 0; 1162 } 1163 1164 static void un_resize_user(struct mlx5_ib_cq *cq) 1165 { 1166 ib_umem_release(cq->resize_umem); 1167 } 1168 1169 static int resize_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *cq, 1170 int entries, int cqe_size) 1171 { 1172 int err; 1173 1174 cq->resize_buf = kzalloc(sizeof(*cq->resize_buf), GFP_KERNEL); 1175 if (!cq->resize_buf) 1176 return -ENOMEM; 1177 1178 err = alloc_cq_buf(dev, cq->resize_buf, entries, cqe_size); 1179 if (err) 1180 goto ex; 1181 1182 init_cq_buf(cq, cq->resize_buf); 1183 1184 return 0; 1185 1186 ex: 1187 kfree(cq->resize_buf); 1188 return err; 1189 } 1190 1191 static void un_resize_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *cq) 1192 { 1193 free_cq_buf(dev, cq->resize_buf); 1194 cq->resize_buf = NULL; 1195 } 1196 1197 static int copy_resize_cqes(struct mlx5_ib_cq *cq) 1198 { 1199 struct mlx5_ib_dev *dev = to_mdev(cq->ibcq.device); 1200 struct mlx5_cqe64 *scqe64; 1201 struct mlx5_cqe64 *dcqe64; 1202 void *start_cqe; 1203 void *scqe; 1204 void *dcqe; 1205 int ssize; 1206 int dsize; 1207 int i; 1208 u8 sw_own; 1209 1210 ssize = cq->buf.cqe_size; 1211 dsize = cq->resize_buf->cqe_size; 1212 if (ssize != dsize) { 1213 mlx5_ib_warn(dev, "resize from different cqe size is not supported\n"); 1214 return -EINVAL; 1215 } 1216 1217 i = cq->mcq.cons_index; 1218 scqe = get_sw_cqe(cq, i); 1219 scqe64 = ssize == 64 ? scqe : scqe + 64; 1220 start_cqe = scqe; 1221 if (!scqe) { 1222 mlx5_ib_warn(dev, "expected cqe in sw ownership\n"); 1223 return -EINVAL; 1224 } 1225 1226 while ((scqe64->op_own >> 4) != MLX5_CQE_RESIZE_CQ) { 1227 dcqe = get_cqe_from_buf(cq->resize_buf, 1228 (i + 1) & (cq->resize_buf->nent), 1229 dsize); 1230 dcqe64 = dsize == 64 ? dcqe : dcqe + 64; 1231 sw_own = sw_ownership_bit(i + 1, cq->resize_buf->nent); 1232 memcpy(dcqe, scqe, dsize); 1233 dcqe64->op_own = (dcqe64->op_own & ~MLX5_CQE_OWNER_MASK) | sw_own; 1234 1235 ++i; 1236 scqe = get_sw_cqe(cq, i); 1237 scqe64 = ssize == 64 ? scqe : scqe + 64; 1238 if (!scqe) { 1239 mlx5_ib_warn(dev, "expected cqe in sw ownership\n"); 1240 return -EINVAL; 1241 } 1242 1243 if (scqe == start_cqe) { 1244 pr_warn("resize CQ failed to get resize CQE, CQN 0x%x\n", 1245 cq->mcq.cqn); 1246 return -ENOMEM; 1247 } 1248 } 1249 ++cq->mcq.cons_index; 1250 return 0; 1251 } 1252 1253 int mlx5_ib_resize_cq(struct ib_cq *ibcq, int entries, struct ib_udata *udata) 1254 { 1255 struct mlx5_ib_dev *dev = to_mdev(ibcq->device); 1256 struct mlx5_ib_cq *cq = to_mcq(ibcq); 1257 void *cqc; 1258 u32 *in; 1259 int err; 1260 int npas; 1261 __be64 *pas; 1262 int page_shift; 1263 int inlen; 1264 int uninitialized_var(cqe_size); 1265 unsigned long flags; 1266 1267 if (!MLX5_CAP_GEN(dev->mdev, cq_resize)) { 1268 pr_info("Firmware does not support resize CQ\n"); 1269 return -ENOSYS; 1270 } 1271 1272 if (entries < 1 || 1273 entries > (1 << MLX5_CAP_GEN(dev->mdev, log_max_cq_sz))) { 1274 mlx5_ib_warn(dev, "wrong entries number %d, max %d\n", 1275 entries, 1276 1 << MLX5_CAP_GEN(dev->mdev, log_max_cq_sz)); 1277 return -EINVAL; 1278 } 1279 1280 entries = roundup_pow_of_two(entries + 1); 1281 if (entries > (1 << MLX5_CAP_GEN(dev->mdev, log_max_cq_sz)) + 1) 1282 return -EINVAL; 1283 1284 if (entries == ibcq->cqe + 1) 1285 return 0; 1286 1287 mutex_lock(&cq->resize_mutex); 1288 if (udata) { 1289 err = resize_user(dev, cq, entries, udata, &npas, &page_shift, 1290 &cqe_size); 1291 } else { 1292 cqe_size = 64; 1293 err = resize_kernel(dev, cq, entries, cqe_size); 1294 if (!err) { 1295 npas = cq->resize_buf->buf.npages; 1296 page_shift = cq->resize_buf->buf.page_shift; 1297 } 1298 } 1299 1300 if (err) 1301 goto ex; 1302 1303 inlen = MLX5_ST_SZ_BYTES(modify_cq_in) + 1304 MLX5_FLD_SZ_BYTES(modify_cq_in, pas[0]) * npas; 1305 1306 in = mlx5_vzalloc(inlen); 1307 if (!in) { 1308 err = -ENOMEM; 1309 goto ex_resize; 1310 } 1311 1312 pas = (__be64 *)MLX5_ADDR_OF(modify_cq_in, in, pas); 1313 if (udata) 1314 mlx5_ib_populate_pas(dev, cq->resize_umem, page_shift, 1315 pas, 0); 1316 else 1317 mlx5_fill_page_array(&cq->resize_buf->buf, pas); 1318 1319 MLX5_SET(modify_cq_in, in, 1320 modify_field_select_resize_field_select.resize_field_select.resize_field_select, 1321 MLX5_MODIFY_CQ_MASK_LOG_SIZE | 1322 MLX5_MODIFY_CQ_MASK_PG_OFFSET | 1323 MLX5_MODIFY_CQ_MASK_PG_SIZE); 1324 1325 cqc = MLX5_ADDR_OF(modify_cq_in, in, cq_context); 1326 1327 MLX5_SET(cqc, cqc, log_page_size, 1328 page_shift - MLX5_ADAPTER_PAGE_SHIFT); 1329 MLX5_SET(cqc, cqc, cqe_sz, cqe_sz_to_mlx_sz(cqe_size)); 1330 MLX5_SET(cqc, cqc, log_cq_size, ilog2(entries)); 1331 1332 MLX5_SET(modify_cq_in, in, op_mod, MLX5_CQ_OPMOD_RESIZE); 1333 MLX5_SET(modify_cq_in, in, cqn, cq->mcq.cqn); 1334 1335 err = mlx5_core_modify_cq(dev->mdev, &cq->mcq, in, inlen); 1336 if (err) 1337 goto ex_alloc; 1338 1339 if (udata) { 1340 cq->ibcq.cqe = entries - 1; 1341 ib_umem_release(cq->buf.umem); 1342 cq->buf.umem = cq->resize_umem; 1343 cq->resize_umem = NULL; 1344 } else { 1345 struct mlx5_ib_cq_buf tbuf; 1346 int resized = 0; 1347 1348 spin_lock_irqsave(&cq->lock, flags); 1349 if (cq->resize_buf) { 1350 err = copy_resize_cqes(cq); 1351 if (!err) { 1352 tbuf = cq->buf; 1353 cq->buf = *cq->resize_buf; 1354 kfree(cq->resize_buf); 1355 cq->resize_buf = NULL; 1356 resized = 1; 1357 } 1358 } 1359 cq->ibcq.cqe = entries - 1; 1360 spin_unlock_irqrestore(&cq->lock, flags); 1361 if (resized) 1362 free_cq_buf(dev, &tbuf); 1363 } 1364 mutex_unlock(&cq->resize_mutex); 1365 1366 kvfree(in); 1367 return 0; 1368 1369 ex_alloc: 1370 kvfree(in); 1371 1372 ex_resize: 1373 if (udata) 1374 un_resize_user(cq); 1375 else 1376 un_resize_kernel(dev, cq); 1377 ex: 1378 mutex_unlock(&cq->resize_mutex); 1379 return err; 1380 } 1381 1382 int mlx5_ib_get_cqe_size(struct mlx5_ib_dev *dev, struct ib_cq *ibcq) 1383 { 1384 struct mlx5_ib_cq *cq; 1385 1386 if (!ibcq) 1387 return 128; 1388 1389 cq = to_mcq(ibcq); 1390 return cq->cqe_size; 1391 } 1392 1393 /* Called from atomic context */ 1394 int mlx5_ib_generate_wc(struct ib_cq *ibcq, struct ib_wc *wc) 1395 { 1396 struct mlx5_ib_wc *soft_wc; 1397 struct mlx5_ib_cq *cq = to_mcq(ibcq); 1398 unsigned long flags; 1399 1400 soft_wc = kmalloc(sizeof(*soft_wc), GFP_ATOMIC); 1401 if (!soft_wc) 1402 return -ENOMEM; 1403 1404 soft_wc->wc = *wc; 1405 spin_lock_irqsave(&cq->lock, flags); 1406 list_add_tail(&soft_wc->list, &cq->wc_list); 1407 if (cq->notify_flags == IB_CQ_NEXT_COMP || 1408 wc->status != IB_WC_SUCCESS) { 1409 cq->notify_flags = 0; 1410 schedule_work(&cq->notify_work); 1411 } 1412 spin_unlock_irqrestore(&cq->lock, flags); 1413 1414 return 0; 1415 } 1416