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 #include "srq.h" 39 #include "qp.h" 40 41 static void mlx5_ib_cq_comp(struct mlx5_core_cq *cq, struct mlx5_eqe *eqe) 42 { 43 struct ib_cq *ibcq = &to_mibcq(cq)->ibcq; 44 45 ibcq->comp_handler(ibcq, ibcq->cq_context); 46 } 47 48 static void mlx5_ib_cq_event(struct mlx5_core_cq *mcq, enum mlx5_event type) 49 { 50 struct mlx5_ib_cq *cq = container_of(mcq, struct mlx5_ib_cq, mcq); 51 struct mlx5_ib_dev *dev = to_mdev(cq->ibcq.device); 52 struct ib_cq *ibcq = &cq->ibcq; 53 struct ib_event event; 54 55 if (type != MLX5_EVENT_TYPE_CQ_ERROR) { 56 mlx5_ib_warn(dev, "Unexpected event type %d on CQ %06x\n", 57 type, mcq->cqn); 58 return; 59 } 60 61 if (ibcq->event_handler) { 62 event.device = &dev->ib_dev; 63 event.event = IB_EVENT_CQ_ERR; 64 event.element.cq = ibcq; 65 ibcq->event_handler(&event, ibcq->cq_context); 66 } 67 } 68 69 static void *get_cqe(struct mlx5_ib_cq *cq, int n) 70 { 71 return mlx5_frag_buf_get_wqe(&cq->buf.fbc, n); 72 } 73 74 static u8 sw_ownership_bit(int n, int nent) 75 { 76 return (n & nent) ? 1 : 0; 77 } 78 79 static void *get_sw_cqe(struct mlx5_ib_cq *cq, int n) 80 { 81 void *cqe = get_cqe(cq, n & cq->ibcq.cqe); 82 struct mlx5_cqe64 *cqe64; 83 84 cqe64 = (cq->mcq.cqe_sz == 64) ? cqe : cqe + 64; 85 86 if (likely(get_cqe_opcode(cqe64) != MLX5_CQE_INVALID) && 87 !((cqe64->op_own & MLX5_CQE_OWNER_MASK) ^ !!(n & (cq->ibcq.cqe + 1)))) { 88 return cqe; 89 } else { 90 return NULL; 91 } 92 } 93 94 static void *next_cqe_sw(struct mlx5_ib_cq *cq) 95 { 96 return get_sw_cqe(cq, cq->mcq.cons_index); 97 } 98 99 static enum ib_wc_opcode get_umr_comp(struct mlx5_ib_wq *wq, int idx) 100 { 101 switch (wq->wr_data[idx]) { 102 case MLX5_IB_WR_UMR: 103 return 0; 104 105 case IB_WR_LOCAL_INV: 106 return IB_WC_LOCAL_INV; 107 108 case IB_WR_REG_MR: 109 return IB_WC_REG_MR; 110 111 default: 112 pr_warn("unknown completion status\n"); 113 return 0; 114 } 115 } 116 117 static void handle_good_req(struct ib_wc *wc, struct mlx5_cqe64 *cqe, 118 struct mlx5_ib_wq *wq, int idx) 119 { 120 wc->wc_flags = 0; 121 switch (be32_to_cpu(cqe->sop_drop_qpn) >> 24) { 122 case MLX5_OPCODE_RDMA_WRITE_IMM: 123 wc->wc_flags |= IB_WC_WITH_IMM; 124 fallthrough; 125 case MLX5_OPCODE_RDMA_WRITE: 126 wc->opcode = IB_WC_RDMA_WRITE; 127 break; 128 case MLX5_OPCODE_SEND_IMM: 129 wc->wc_flags |= IB_WC_WITH_IMM; 130 fallthrough; 131 case MLX5_OPCODE_SEND: 132 case MLX5_OPCODE_SEND_INVAL: 133 wc->opcode = IB_WC_SEND; 134 break; 135 case MLX5_OPCODE_RDMA_READ: 136 wc->opcode = IB_WC_RDMA_READ; 137 wc->byte_len = be32_to_cpu(cqe->byte_cnt); 138 break; 139 case MLX5_OPCODE_ATOMIC_CS: 140 wc->opcode = IB_WC_COMP_SWAP; 141 wc->byte_len = 8; 142 break; 143 case MLX5_OPCODE_ATOMIC_FA: 144 wc->opcode = IB_WC_FETCH_ADD; 145 wc->byte_len = 8; 146 break; 147 case MLX5_OPCODE_ATOMIC_MASKED_CS: 148 wc->opcode = IB_WC_MASKED_COMP_SWAP; 149 wc->byte_len = 8; 150 break; 151 case MLX5_OPCODE_ATOMIC_MASKED_FA: 152 wc->opcode = IB_WC_MASKED_FETCH_ADD; 153 wc->byte_len = 8; 154 break; 155 case MLX5_OPCODE_UMR: 156 wc->opcode = get_umr_comp(wq, idx); 157 break; 158 } 159 } 160 161 enum { 162 MLX5_GRH_IN_BUFFER = 1, 163 MLX5_GRH_IN_CQE = 2, 164 }; 165 166 static void handle_responder(struct ib_wc *wc, struct mlx5_cqe64 *cqe, 167 struct mlx5_ib_qp *qp) 168 { 169 enum rdma_link_layer ll = rdma_port_get_link_layer(qp->ibqp.device, 1); 170 struct mlx5_ib_dev *dev = to_mdev(qp->ibqp.device); 171 struct mlx5_ib_srq *srq = NULL; 172 struct mlx5_ib_wq *wq; 173 u16 wqe_ctr; 174 u8 roce_packet_type; 175 bool vlan_present; 176 u8 g; 177 178 if (qp->ibqp.srq || qp->ibqp.xrcd) { 179 struct mlx5_core_srq *msrq = NULL; 180 181 if (qp->ibqp.xrcd) { 182 msrq = mlx5_cmd_get_srq(dev, be32_to_cpu(cqe->srqn)); 183 if (msrq) 184 srq = to_mibsrq(msrq); 185 } else { 186 srq = to_msrq(qp->ibqp.srq); 187 } 188 if (srq) { 189 wqe_ctr = be16_to_cpu(cqe->wqe_counter); 190 wc->wr_id = srq->wrid[wqe_ctr]; 191 mlx5_ib_free_srq_wqe(srq, wqe_ctr); 192 if (msrq) 193 mlx5_core_res_put(&msrq->common); 194 } 195 } else { 196 wq = &qp->rq; 197 wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)]; 198 ++wq->tail; 199 } 200 wc->byte_len = be32_to_cpu(cqe->byte_cnt); 201 202 switch (get_cqe_opcode(cqe)) { 203 case MLX5_CQE_RESP_WR_IMM: 204 wc->opcode = IB_WC_RECV_RDMA_WITH_IMM; 205 wc->wc_flags = IB_WC_WITH_IMM; 206 wc->ex.imm_data = cqe->immediate; 207 break; 208 case MLX5_CQE_RESP_SEND: 209 wc->opcode = IB_WC_RECV; 210 wc->wc_flags = IB_WC_IP_CSUM_OK; 211 if (unlikely(!((cqe->hds_ip_ext & CQE_L3_OK) && 212 (cqe->hds_ip_ext & CQE_L4_OK)))) 213 wc->wc_flags = 0; 214 break; 215 case MLX5_CQE_RESP_SEND_IMM: 216 wc->opcode = IB_WC_RECV; 217 wc->wc_flags = IB_WC_WITH_IMM; 218 wc->ex.imm_data = cqe->immediate; 219 break; 220 case MLX5_CQE_RESP_SEND_INV: 221 wc->opcode = IB_WC_RECV; 222 wc->wc_flags = IB_WC_WITH_INVALIDATE; 223 wc->ex.invalidate_rkey = be32_to_cpu(cqe->inval_rkey); 224 break; 225 } 226 wc->src_qp = be32_to_cpu(cqe->flags_rqpn) & 0xffffff; 227 wc->dlid_path_bits = cqe->ml_path; 228 g = (be32_to_cpu(cqe->flags_rqpn) >> 28) & 3; 229 wc->wc_flags |= g ? IB_WC_GRH : 0; 230 if (is_qp1(qp->type)) { 231 u16 pkey = be32_to_cpu(cqe->pkey) & 0xffff; 232 233 ib_find_cached_pkey(&dev->ib_dev, qp->port, pkey, 234 &wc->pkey_index); 235 } else { 236 wc->pkey_index = 0; 237 } 238 239 if (ll != IB_LINK_LAYER_ETHERNET) { 240 wc->slid = be16_to_cpu(cqe->slid); 241 wc->sl = (be32_to_cpu(cqe->flags_rqpn) >> 24) & 0xf; 242 return; 243 } 244 245 wc->slid = 0; 246 vlan_present = cqe->l4_l3_hdr_type & 0x1; 247 roce_packet_type = (be32_to_cpu(cqe->flags_rqpn) >> 24) & 0x3; 248 if (vlan_present) { 249 wc->vlan_id = (be16_to_cpu(cqe->vlan_info)) & 0xfff; 250 wc->sl = (be16_to_cpu(cqe->vlan_info) >> 13) & 0x7; 251 wc->wc_flags |= IB_WC_WITH_VLAN; 252 } else { 253 wc->sl = 0; 254 } 255 256 switch (roce_packet_type) { 257 case MLX5_CQE_ROCE_L3_HEADER_TYPE_GRH: 258 wc->network_hdr_type = RDMA_NETWORK_ROCE_V1; 259 break; 260 case MLX5_CQE_ROCE_L3_HEADER_TYPE_IPV6: 261 wc->network_hdr_type = RDMA_NETWORK_IPV6; 262 break; 263 case MLX5_CQE_ROCE_L3_HEADER_TYPE_IPV4: 264 wc->network_hdr_type = RDMA_NETWORK_IPV4; 265 break; 266 } 267 wc->wc_flags |= IB_WC_WITH_NETWORK_HDR_TYPE; 268 } 269 270 static void dump_cqe(struct mlx5_ib_dev *dev, struct mlx5_err_cqe *cqe) 271 { 272 mlx5_ib_warn(dev, "dump error cqe\n"); 273 mlx5_dump_err_cqe(dev->mdev, cqe); 274 } 275 276 static void mlx5_handle_error_cqe(struct mlx5_ib_dev *dev, 277 struct mlx5_err_cqe *cqe, 278 struct ib_wc *wc) 279 { 280 int dump = 1; 281 282 switch (cqe->syndrome) { 283 case MLX5_CQE_SYNDROME_LOCAL_LENGTH_ERR: 284 wc->status = IB_WC_LOC_LEN_ERR; 285 break; 286 case MLX5_CQE_SYNDROME_LOCAL_QP_OP_ERR: 287 wc->status = IB_WC_LOC_QP_OP_ERR; 288 break; 289 case MLX5_CQE_SYNDROME_LOCAL_PROT_ERR: 290 wc->status = IB_WC_LOC_PROT_ERR; 291 break; 292 case MLX5_CQE_SYNDROME_WR_FLUSH_ERR: 293 dump = 0; 294 wc->status = IB_WC_WR_FLUSH_ERR; 295 break; 296 case MLX5_CQE_SYNDROME_MW_BIND_ERR: 297 wc->status = IB_WC_MW_BIND_ERR; 298 break; 299 case MLX5_CQE_SYNDROME_BAD_RESP_ERR: 300 wc->status = IB_WC_BAD_RESP_ERR; 301 break; 302 case MLX5_CQE_SYNDROME_LOCAL_ACCESS_ERR: 303 wc->status = IB_WC_LOC_ACCESS_ERR; 304 break; 305 case MLX5_CQE_SYNDROME_REMOTE_INVAL_REQ_ERR: 306 wc->status = IB_WC_REM_INV_REQ_ERR; 307 break; 308 case MLX5_CQE_SYNDROME_REMOTE_ACCESS_ERR: 309 wc->status = IB_WC_REM_ACCESS_ERR; 310 break; 311 case MLX5_CQE_SYNDROME_REMOTE_OP_ERR: 312 wc->status = IB_WC_REM_OP_ERR; 313 break; 314 case MLX5_CQE_SYNDROME_TRANSPORT_RETRY_EXC_ERR: 315 wc->status = IB_WC_RETRY_EXC_ERR; 316 dump = 0; 317 break; 318 case MLX5_CQE_SYNDROME_RNR_RETRY_EXC_ERR: 319 wc->status = IB_WC_RNR_RETRY_EXC_ERR; 320 dump = 0; 321 break; 322 case MLX5_CQE_SYNDROME_REMOTE_ABORTED_ERR: 323 wc->status = IB_WC_REM_ABORT_ERR; 324 break; 325 default: 326 wc->status = IB_WC_GENERAL_ERR; 327 break; 328 } 329 330 wc->vendor_err = cqe->vendor_err_synd; 331 if (dump) 332 dump_cqe(dev, cqe); 333 } 334 335 static void handle_atomics(struct mlx5_ib_qp *qp, struct mlx5_cqe64 *cqe64, 336 u16 tail, u16 head) 337 { 338 u16 idx; 339 340 do { 341 idx = tail & (qp->sq.wqe_cnt - 1); 342 if (idx == head) 343 break; 344 345 tail = qp->sq.w_list[idx].next; 346 } while (1); 347 tail = qp->sq.w_list[idx].next; 348 qp->sq.last_poll = tail; 349 } 350 351 static void free_cq_buf(struct mlx5_ib_dev *dev, struct mlx5_ib_cq_buf *buf) 352 { 353 mlx5_frag_buf_free(dev->mdev, &buf->frag_buf); 354 } 355 356 static void get_sig_err_item(struct mlx5_sig_err_cqe *cqe, 357 struct ib_sig_err *item) 358 { 359 u16 syndrome = be16_to_cpu(cqe->syndrome); 360 361 #define GUARD_ERR (1 << 13) 362 #define APPTAG_ERR (1 << 12) 363 #define REFTAG_ERR (1 << 11) 364 365 if (syndrome & GUARD_ERR) { 366 item->err_type = IB_SIG_BAD_GUARD; 367 item->expected = be32_to_cpu(cqe->expected_trans_sig) >> 16; 368 item->actual = be32_to_cpu(cqe->actual_trans_sig) >> 16; 369 } else 370 if (syndrome & REFTAG_ERR) { 371 item->err_type = IB_SIG_BAD_REFTAG; 372 item->expected = be32_to_cpu(cqe->expected_reftag); 373 item->actual = be32_to_cpu(cqe->actual_reftag); 374 } else 375 if (syndrome & APPTAG_ERR) { 376 item->err_type = IB_SIG_BAD_APPTAG; 377 item->expected = be32_to_cpu(cqe->expected_trans_sig) & 0xffff; 378 item->actual = be32_to_cpu(cqe->actual_trans_sig) & 0xffff; 379 } else { 380 pr_err("Got signature completion error with bad syndrome %04x\n", 381 syndrome); 382 } 383 384 item->sig_err_offset = be64_to_cpu(cqe->err_offset); 385 item->key = be32_to_cpu(cqe->mkey); 386 } 387 388 static void sw_comp(struct mlx5_ib_qp *qp, int num_entries, struct ib_wc *wc, 389 int *npolled, bool is_send) 390 { 391 struct mlx5_ib_wq *wq; 392 unsigned int cur; 393 int np; 394 int i; 395 396 wq = (is_send) ? &qp->sq : &qp->rq; 397 cur = wq->head - wq->tail; 398 np = *npolled; 399 400 if (cur == 0) 401 return; 402 403 for (i = 0; i < cur && np < num_entries; i++) { 404 unsigned int idx; 405 406 idx = (is_send) ? wq->last_poll : wq->tail; 407 idx &= (wq->wqe_cnt - 1); 408 wc->wr_id = wq->wrid[idx]; 409 wc->status = IB_WC_WR_FLUSH_ERR; 410 wc->vendor_err = MLX5_CQE_SYNDROME_WR_FLUSH_ERR; 411 wq->tail++; 412 if (is_send) 413 wq->last_poll = wq->w_list[idx].next; 414 np++; 415 wc->qp = &qp->ibqp; 416 wc++; 417 } 418 *npolled = np; 419 } 420 421 static void mlx5_ib_poll_sw_comp(struct mlx5_ib_cq *cq, int num_entries, 422 struct ib_wc *wc, int *npolled) 423 { 424 struct mlx5_ib_qp *qp; 425 426 *npolled = 0; 427 /* Find uncompleted WQEs belonging to that cq and return mmics ones */ 428 list_for_each_entry(qp, &cq->list_send_qp, cq_send_list) { 429 sw_comp(qp, num_entries, wc + *npolled, npolled, true); 430 if (*npolled >= num_entries) 431 return; 432 } 433 434 list_for_each_entry(qp, &cq->list_recv_qp, cq_recv_list) { 435 sw_comp(qp, num_entries, wc + *npolled, npolled, false); 436 if (*npolled >= num_entries) 437 return; 438 } 439 } 440 441 static int mlx5_poll_one(struct mlx5_ib_cq *cq, 442 struct mlx5_ib_qp **cur_qp, 443 struct ib_wc *wc) 444 { 445 struct mlx5_ib_dev *dev = to_mdev(cq->ibcq.device); 446 struct mlx5_err_cqe *err_cqe; 447 struct mlx5_cqe64 *cqe64; 448 struct mlx5_core_qp *mqp; 449 struct mlx5_ib_wq *wq; 450 uint8_t opcode; 451 uint32_t qpn; 452 u16 wqe_ctr; 453 void *cqe; 454 int idx; 455 456 repoll: 457 cqe = next_cqe_sw(cq); 458 if (!cqe) 459 return -EAGAIN; 460 461 cqe64 = (cq->mcq.cqe_sz == 64) ? cqe : cqe + 64; 462 463 ++cq->mcq.cons_index; 464 465 /* Make sure we read CQ entry contents after we've checked the 466 * ownership bit. 467 */ 468 rmb(); 469 470 opcode = get_cqe_opcode(cqe64); 471 if (unlikely(opcode == MLX5_CQE_RESIZE_CQ)) { 472 if (likely(cq->resize_buf)) { 473 free_cq_buf(dev, &cq->buf); 474 cq->buf = *cq->resize_buf; 475 kfree(cq->resize_buf); 476 cq->resize_buf = NULL; 477 goto repoll; 478 } else { 479 mlx5_ib_warn(dev, "unexpected resize cqe\n"); 480 } 481 } 482 483 qpn = ntohl(cqe64->sop_drop_qpn) & 0xffffff; 484 if (!*cur_qp || (qpn != (*cur_qp)->ibqp.qp_num)) { 485 /* We do not have to take the QP table lock here, 486 * because CQs will be locked while QPs are removed 487 * from the table. 488 */ 489 mqp = radix_tree_lookup(&dev->qp_table.tree, qpn); 490 *cur_qp = to_mibqp(mqp); 491 } 492 493 wc->qp = &(*cur_qp)->ibqp; 494 switch (opcode) { 495 case MLX5_CQE_REQ: 496 wq = &(*cur_qp)->sq; 497 wqe_ctr = be16_to_cpu(cqe64->wqe_counter); 498 idx = wqe_ctr & (wq->wqe_cnt - 1); 499 handle_good_req(wc, cqe64, wq, idx); 500 handle_atomics(*cur_qp, cqe64, wq->last_poll, idx); 501 wc->wr_id = wq->wrid[idx]; 502 wq->tail = wq->wqe_head[idx] + 1; 503 wc->status = IB_WC_SUCCESS; 504 break; 505 case MLX5_CQE_RESP_WR_IMM: 506 case MLX5_CQE_RESP_SEND: 507 case MLX5_CQE_RESP_SEND_IMM: 508 case MLX5_CQE_RESP_SEND_INV: 509 handle_responder(wc, cqe64, *cur_qp); 510 wc->status = IB_WC_SUCCESS; 511 break; 512 case MLX5_CQE_RESIZE_CQ: 513 break; 514 case MLX5_CQE_REQ_ERR: 515 case MLX5_CQE_RESP_ERR: 516 err_cqe = (struct mlx5_err_cqe *)cqe64; 517 mlx5_handle_error_cqe(dev, err_cqe, wc); 518 mlx5_ib_dbg(dev, "%s error cqe on cqn 0x%x:\n", 519 opcode == MLX5_CQE_REQ_ERR ? 520 "Requestor" : "Responder", cq->mcq.cqn); 521 mlx5_ib_dbg(dev, "syndrome 0x%x, vendor syndrome 0x%x\n", 522 err_cqe->syndrome, err_cqe->vendor_err_synd); 523 if (opcode == MLX5_CQE_REQ_ERR) { 524 wq = &(*cur_qp)->sq; 525 wqe_ctr = be16_to_cpu(cqe64->wqe_counter); 526 idx = wqe_ctr & (wq->wqe_cnt - 1); 527 wc->wr_id = wq->wrid[idx]; 528 wq->tail = wq->wqe_head[idx] + 1; 529 } else { 530 struct mlx5_ib_srq *srq; 531 532 if ((*cur_qp)->ibqp.srq) { 533 srq = to_msrq((*cur_qp)->ibqp.srq); 534 wqe_ctr = be16_to_cpu(cqe64->wqe_counter); 535 wc->wr_id = srq->wrid[wqe_ctr]; 536 mlx5_ib_free_srq_wqe(srq, wqe_ctr); 537 } else { 538 wq = &(*cur_qp)->rq; 539 wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)]; 540 ++wq->tail; 541 } 542 } 543 break; 544 case MLX5_CQE_SIG_ERR: { 545 struct mlx5_sig_err_cqe *sig_err_cqe = 546 (struct mlx5_sig_err_cqe *)cqe64; 547 struct mlx5_core_sig_ctx *sig; 548 549 xa_lock(&dev->sig_mrs); 550 sig = xa_load(&dev->sig_mrs, 551 mlx5_base_mkey(be32_to_cpu(sig_err_cqe->mkey))); 552 get_sig_err_item(sig_err_cqe, &sig->err_item); 553 sig->sig_err_exists = true; 554 sig->sigerr_count++; 555 556 mlx5_ib_warn(dev, "CQN: 0x%x Got SIGERR on key: 0x%x err_type %x err_offset %llx expected %x actual %x\n", 557 cq->mcq.cqn, sig->err_item.key, 558 sig->err_item.err_type, 559 sig->err_item.sig_err_offset, 560 sig->err_item.expected, 561 sig->err_item.actual); 562 563 xa_unlock(&dev->sig_mrs); 564 goto repoll; 565 } 566 } 567 568 return 0; 569 } 570 571 static int poll_soft_wc(struct mlx5_ib_cq *cq, int num_entries, 572 struct ib_wc *wc, bool is_fatal_err) 573 { 574 struct mlx5_ib_dev *dev = to_mdev(cq->ibcq.device); 575 struct mlx5_ib_wc *soft_wc, *next; 576 int npolled = 0; 577 578 list_for_each_entry_safe(soft_wc, next, &cq->wc_list, list) { 579 if (npolled >= num_entries) 580 break; 581 582 mlx5_ib_dbg(dev, "polled software generated completion on CQ 0x%x\n", 583 cq->mcq.cqn); 584 585 if (unlikely(is_fatal_err)) { 586 soft_wc->wc.status = IB_WC_WR_FLUSH_ERR; 587 soft_wc->wc.vendor_err = MLX5_CQE_SYNDROME_WR_FLUSH_ERR; 588 } 589 wc[npolled++] = soft_wc->wc; 590 list_del(&soft_wc->list); 591 kfree(soft_wc); 592 } 593 594 return npolled; 595 } 596 597 int mlx5_ib_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc) 598 { 599 struct mlx5_ib_cq *cq = to_mcq(ibcq); 600 struct mlx5_ib_qp *cur_qp = NULL; 601 struct mlx5_ib_dev *dev = to_mdev(cq->ibcq.device); 602 struct mlx5_core_dev *mdev = dev->mdev; 603 unsigned long flags; 604 int soft_polled = 0; 605 int npolled; 606 607 spin_lock_irqsave(&cq->lock, flags); 608 if (mdev->state == MLX5_DEVICE_STATE_INTERNAL_ERROR) { 609 /* make sure no soft wqe's are waiting */ 610 if (unlikely(!list_empty(&cq->wc_list))) 611 soft_polled = poll_soft_wc(cq, num_entries, wc, true); 612 613 mlx5_ib_poll_sw_comp(cq, num_entries - soft_polled, 614 wc + soft_polled, &npolled); 615 goto out; 616 } 617 618 if (unlikely(!list_empty(&cq->wc_list))) 619 soft_polled = poll_soft_wc(cq, num_entries, wc, false); 620 621 for (npolled = 0; npolled < num_entries - soft_polled; npolled++) { 622 if (mlx5_poll_one(cq, &cur_qp, wc + soft_polled + npolled)) 623 break; 624 } 625 626 if (npolled) 627 mlx5_cq_set_ci(&cq->mcq); 628 out: 629 spin_unlock_irqrestore(&cq->lock, flags); 630 631 return soft_polled + npolled; 632 } 633 634 int mlx5_ib_arm_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags flags) 635 { 636 struct mlx5_core_dev *mdev = to_mdev(ibcq->device)->mdev; 637 struct mlx5_ib_cq *cq = to_mcq(ibcq); 638 void __iomem *uar_page = mdev->priv.uar->map; 639 unsigned long irq_flags; 640 int ret = 0; 641 642 spin_lock_irqsave(&cq->lock, irq_flags); 643 if (cq->notify_flags != IB_CQ_NEXT_COMP) 644 cq->notify_flags = flags & IB_CQ_SOLICITED_MASK; 645 646 if ((flags & IB_CQ_REPORT_MISSED_EVENTS) && !list_empty(&cq->wc_list)) 647 ret = 1; 648 spin_unlock_irqrestore(&cq->lock, irq_flags); 649 650 mlx5_cq_arm(&cq->mcq, 651 (flags & IB_CQ_SOLICITED_MASK) == IB_CQ_SOLICITED ? 652 MLX5_CQ_DB_REQ_NOT_SOL : MLX5_CQ_DB_REQ_NOT, 653 uar_page, to_mcq(ibcq)->mcq.cons_index); 654 655 return ret; 656 } 657 658 static int alloc_cq_frag_buf(struct mlx5_ib_dev *dev, 659 struct mlx5_ib_cq_buf *buf, 660 int nent, 661 int cqe_size) 662 { 663 struct mlx5_frag_buf *frag_buf = &buf->frag_buf; 664 u8 log_wq_stride = 6 + (cqe_size == 128 ? 1 : 0); 665 u8 log_wq_sz = ilog2(cqe_size); 666 int err; 667 668 err = mlx5_frag_buf_alloc_node(dev->mdev, 669 nent * cqe_size, 670 frag_buf, 671 dev->mdev->priv.numa_node); 672 if (err) 673 return err; 674 675 mlx5_init_fbc(frag_buf->frags, log_wq_stride, log_wq_sz, &buf->fbc); 676 677 buf->cqe_size = cqe_size; 678 buf->nent = nent; 679 680 return 0; 681 } 682 683 enum { 684 MLX5_CQE_RES_FORMAT_HASH = 0, 685 MLX5_CQE_RES_FORMAT_CSUM = 1, 686 MLX5_CQE_RES_FORMAT_CSUM_STRIDX = 3, 687 }; 688 689 static int mini_cqe_res_format_to_hw(struct mlx5_ib_dev *dev, u8 format) 690 { 691 switch (format) { 692 case MLX5_IB_CQE_RES_FORMAT_HASH: 693 return MLX5_CQE_RES_FORMAT_HASH; 694 case MLX5_IB_CQE_RES_FORMAT_CSUM: 695 return MLX5_CQE_RES_FORMAT_CSUM; 696 case MLX5_IB_CQE_RES_FORMAT_CSUM_STRIDX: 697 if (MLX5_CAP_GEN(dev->mdev, mini_cqe_resp_stride_index)) 698 return MLX5_CQE_RES_FORMAT_CSUM_STRIDX; 699 return -EOPNOTSUPP; 700 default: 701 return -EINVAL; 702 } 703 } 704 705 static int create_cq_user(struct mlx5_ib_dev *dev, struct ib_udata *udata, 706 struct mlx5_ib_cq *cq, int entries, u32 **cqb, 707 int *cqe_size, int *index, int *inlen) 708 { 709 struct mlx5_ib_create_cq ucmd = {}; 710 unsigned long page_size; 711 unsigned int page_offset_quantized; 712 size_t ucmdlen; 713 __be64 *pas; 714 int ncont; 715 void *cqc; 716 int err; 717 struct mlx5_ib_ucontext *context = rdma_udata_to_drv_context( 718 udata, struct mlx5_ib_ucontext, ibucontext); 719 720 ucmdlen = min(udata->inlen, sizeof(ucmd)); 721 if (ucmdlen < offsetof(struct mlx5_ib_create_cq, flags)) 722 return -EINVAL; 723 724 if (ib_copy_from_udata(&ucmd, udata, ucmdlen)) 725 return -EFAULT; 726 727 if ((ucmd.flags & ~(MLX5_IB_CREATE_CQ_FLAGS_CQE_128B_PAD | 728 MLX5_IB_CREATE_CQ_FLAGS_UAR_PAGE_INDEX | 729 MLX5_IB_CREATE_CQ_FLAGS_REAL_TIME_TS))) 730 return -EINVAL; 731 732 if ((ucmd.cqe_size != 64 && ucmd.cqe_size != 128) || 733 ucmd.reserved0 || ucmd.reserved1) 734 return -EINVAL; 735 736 *cqe_size = ucmd.cqe_size; 737 738 cq->buf.umem = 739 ib_umem_get(&dev->ib_dev, ucmd.buf_addr, 740 entries * ucmd.cqe_size, IB_ACCESS_LOCAL_WRITE); 741 if (IS_ERR(cq->buf.umem)) { 742 err = PTR_ERR(cq->buf.umem); 743 return err; 744 } 745 746 page_size = mlx5_umem_find_best_cq_quantized_pgoff( 747 cq->buf.umem, cqc, log_page_size, MLX5_ADAPTER_PAGE_SHIFT, 748 page_offset, 64, &page_offset_quantized); 749 if (!page_size) { 750 err = -EINVAL; 751 goto err_umem; 752 } 753 754 err = mlx5_ib_db_map_user(context, ucmd.db_addr, &cq->db); 755 if (err) 756 goto err_umem; 757 758 ncont = ib_umem_num_dma_blocks(cq->buf.umem, page_size); 759 mlx5_ib_dbg( 760 dev, 761 "addr 0x%llx, size %u, npages %zu, page_size %lu, ncont %d\n", 762 ucmd.buf_addr, entries * ucmd.cqe_size, 763 ib_umem_num_pages(cq->buf.umem), page_size, ncont); 764 765 *inlen = MLX5_ST_SZ_BYTES(create_cq_in) + 766 MLX5_FLD_SZ_BYTES(create_cq_in, pas[0]) * ncont; 767 *cqb = kvzalloc(*inlen, GFP_KERNEL); 768 if (!*cqb) { 769 err = -ENOMEM; 770 goto err_db; 771 } 772 773 pas = (__be64 *)MLX5_ADDR_OF(create_cq_in, *cqb, pas); 774 mlx5_ib_populate_pas(cq->buf.umem, page_size, pas, 0); 775 776 cqc = MLX5_ADDR_OF(create_cq_in, *cqb, cq_context); 777 MLX5_SET(cqc, cqc, log_page_size, 778 order_base_2(page_size) - MLX5_ADAPTER_PAGE_SHIFT); 779 MLX5_SET(cqc, cqc, page_offset, page_offset_quantized); 780 781 if (ucmd.flags & MLX5_IB_CREATE_CQ_FLAGS_UAR_PAGE_INDEX) { 782 *index = ucmd.uar_page_index; 783 } else if (context->bfregi.lib_uar_dyn) { 784 err = -EINVAL; 785 goto err_cqb; 786 } else { 787 *index = context->bfregi.sys_pages[0]; 788 } 789 790 if (ucmd.cqe_comp_en == 1) { 791 int mini_cqe_format; 792 793 if (!((*cqe_size == 128 && 794 MLX5_CAP_GEN(dev->mdev, cqe_compression_128)) || 795 (*cqe_size == 64 && 796 MLX5_CAP_GEN(dev->mdev, cqe_compression)))) { 797 err = -EOPNOTSUPP; 798 mlx5_ib_warn(dev, "CQE compression is not supported for size %d!\n", 799 *cqe_size); 800 goto err_cqb; 801 } 802 803 mini_cqe_format = 804 mini_cqe_res_format_to_hw(dev, 805 ucmd.cqe_comp_res_format); 806 if (mini_cqe_format < 0) { 807 err = mini_cqe_format; 808 mlx5_ib_dbg(dev, "CQE compression res format %d error: %d\n", 809 ucmd.cqe_comp_res_format, err); 810 goto err_cqb; 811 } 812 813 MLX5_SET(cqc, cqc, cqe_comp_en, 1); 814 MLX5_SET(cqc, cqc, mini_cqe_res_format, mini_cqe_format); 815 } 816 817 if (ucmd.flags & MLX5_IB_CREATE_CQ_FLAGS_CQE_128B_PAD) { 818 if (*cqe_size != 128 || 819 !MLX5_CAP_GEN(dev->mdev, cqe_128_always)) { 820 err = -EOPNOTSUPP; 821 mlx5_ib_warn(dev, 822 "CQE padding is not supported for CQE size of %dB!\n", 823 *cqe_size); 824 goto err_cqb; 825 } 826 827 cq->private_flags |= MLX5_IB_CQ_PR_FLAGS_CQE_128_PAD; 828 } 829 830 if (ucmd.flags & MLX5_IB_CREATE_CQ_FLAGS_REAL_TIME_TS) 831 cq->private_flags |= MLX5_IB_CQ_PR_FLAGS_REAL_TIME_TS; 832 833 MLX5_SET(create_cq_in, *cqb, uid, context->devx_uid); 834 return 0; 835 836 err_cqb: 837 kvfree(*cqb); 838 839 err_db: 840 mlx5_ib_db_unmap_user(context, &cq->db); 841 842 err_umem: 843 ib_umem_release(cq->buf.umem); 844 return err; 845 } 846 847 static void destroy_cq_user(struct mlx5_ib_cq *cq, struct ib_udata *udata) 848 { 849 struct mlx5_ib_ucontext *context = rdma_udata_to_drv_context( 850 udata, struct mlx5_ib_ucontext, ibucontext); 851 852 mlx5_ib_db_unmap_user(context, &cq->db); 853 ib_umem_release(cq->buf.umem); 854 } 855 856 static void init_cq_frag_buf(struct mlx5_ib_cq_buf *buf) 857 { 858 int i; 859 void *cqe; 860 struct mlx5_cqe64 *cqe64; 861 862 for (i = 0; i < buf->nent; i++) { 863 cqe = mlx5_frag_buf_get_wqe(&buf->fbc, i); 864 cqe64 = buf->cqe_size == 64 ? cqe : cqe + 64; 865 cqe64->op_own = MLX5_CQE_INVALID << 4; 866 } 867 } 868 869 static int create_cq_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *cq, 870 int entries, int cqe_size, 871 u32 **cqb, int *index, int *inlen) 872 { 873 __be64 *pas; 874 void *cqc; 875 int err; 876 877 err = mlx5_db_alloc(dev->mdev, &cq->db); 878 if (err) 879 return err; 880 881 cq->mcq.set_ci_db = cq->db.db; 882 cq->mcq.arm_db = cq->db.db + 1; 883 cq->mcq.cqe_sz = cqe_size; 884 885 err = alloc_cq_frag_buf(dev, &cq->buf, entries, cqe_size); 886 if (err) 887 goto err_db; 888 889 init_cq_frag_buf(&cq->buf); 890 891 *inlen = MLX5_ST_SZ_BYTES(create_cq_in) + 892 MLX5_FLD_SZ_BYTES(create_cq_in, pas[0]) * 893 cq->buf.frag_buf.npages; 894 *cqb = kvzalloc(*inlen, GFP_KERNEL); 895 if (!*cqb) { 896 err = -ENOMEM; 897 goto err_buf; 898 } 899 900 pas = (__be64 *)MLX5_ADDR_OF(create_cq_in, *cqb, pas); 901 mlx5_fill_page_frag_array(&cq->buf.frag_buf, pas); 902 903 cqc = MLX5_ADDR_OF(create_cq_in, *cqb, cq_context); 904 MLX5_SET(cqc, cqc, log_page_size, 905 cq->buf.frag_buf.page_shift - 906 MLX5_ADAPTER_PAGE_SHIFT); 907 908 *index = dev->mdev->priv.uar->index; 909 910 return 0; 911 912 err_buf: 913 free_cq_buf(dev, &cq->buf); 914 915 err_db: 916 mlx5_db_free(dev->mdev, &cq->db); 917 return err; 918 } 919 920 static void destroy_cq_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *cq) 921 { 922 free_cq_buf(dev, &cq->buf); 923 mlx5_db_free(dev->mdev, &cq->db); 924 } 925 926 static void notify_soft_wc_handler(struct work_struct *work) 927 { 928 struct mlx5_ib_cq *cq = container_of(work, struct mlx5_ib_cq, 929 notify_work); 930 931 cq->ibcq.comp_handler(&cq->ibcq, cq->ibcq.cq_context); 932 } 933 934 int mlx5_ib_create_cq(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr, 935 struct ib_udata *udata) 936 { 937 struct ib_device *ibdev = ibcq->device; 938 int entries = attr->cqe; 939 int vector = attr->comp_vector; 940 struct mlx5_ib_dev *dev = to_mdev(ibdev); 941 struct mlx5_ib_cq *cq = to_mcq(ibcq); 942 u32 out[MLX5_ST_SZ_DW(create_cq_out)]; 943 int index; 944 int inlen; 945 u32 *cqb = NULL; 946 void *cqc; 947 int cqe_size; 948 unsigned int irqn; 949 int eqn; 950 int err; 951 952 if (entries < 0 || 953 (entries > (1 << MLX5_CAP_GEN(dev->mdev, log_max_cq_sz)))) 954 return -EINVAL; 955 956 if (check_cq_create_flags(attr->flags)) 957 return -EOPNOTSUPP; 958 959 entries = roundup_pow_of_two(entries + 1); 960 if (entries > (1 << MLX5_CAP_GEN(dev->mdev, log_max_cq_sz))) 961 return -EINVAL; 962 963 cq->ibcq.cqe = entries - 1; 964 mutex_init(&cq->resize_mutex); 965 spin_lock_init(&cq->lock); 966 cq->resize_buf = NULL; 967 cq->resize_umem = NULL; 968 cq->create_flags = attr->flags; 969 INIT_LIST_HEAD(&cq->list_send_qp); 970 INIT_LIST_HEAD(&cq->list_recv_qp); 971 972 if (udata) { 973 err = create_cq_user(dev, udata, cq, entries, &cqb, &cqe_size, 974 &index, &inlen); 975 if (err) 976 return err; 977 } else { 978 cqe_size = cache_line_size() == 128 ? 128 : 64; 979 err = create_cq_kernel(dev, cq, entries, cqe_size, &cqb, 980 &index, &inlen); 981 if (err) 982 return err; 983 984 INIT_WORK(&cq->notify_work, notify_soft_wc_handler); 985 } 986 987 err = mlx5_vector2eqn(dev->mdev, vector, &eqn, &irqn); 988 if (err) 989 goto err_cqb; 990 991 cq->cqe_size = cqe_size; 992 993 cqc = MLX5_ADDR_OF(create_cq_in, cqb, cq_context); 994 MLX5_SET(cqc, cqc, cqe_sz, 995 cqe_sz_to_mlx_sz(cqe_size, 996 cq->private_flags & 997 MLX5_IB_CQ_PR_FLAGS_CQE_128_PAD)); 998 MLX5_SET(cqc, cqc, log_cq_size, ilog2(entries)); 999 MLX5_SET(cqc, cqc, uar_page, index); 1000 MLX5_SET(cqc, cqc, c_eqn, eqn); 1001 MLX5_SET64(cqc, cqc, dbr_addr, cq->db.dma); 1002 if (cq->create_flags & IB_UVERBS_CQ_FLAGS_IGNORE_OVERRUN) 1003 MLX5_SET(cqc, cqc, oi, 1); 1004 1005 err = mlx5_core_create_cq(dev->mdev, &cq->mcq, cqb, inlen, out, sizeof(out)); 1006 if (err) 1007 goto err_cqb; 1008 1009 mlx5_ib_dbg(dev, "cqn 0x%x\n", cq->mcq.cqn); 1010 cq->mcq.irqn = irqn; 1011 if (udata) 1012 cq->mcq.tasklet_ctx.comp = mlx5_ib_cq_comp; 1013 else 1014 cq->mcq.comp = mlx5_ib_cq_comp; 1015 cq->mcq.event = mlx5_ib_cq_event; 1016 1017 INIT_LIST_HEAD(&cq->wc_list); 1018 1019 if (udata) 1020 if (ib_copy_to_udata(udata, &cq->mcq.cqn, sizeof(__u32))) { 1021 err = -EFAULT; 1022 goto err_cmd; 1023 } 1024 1025 1026 kvfree(cqb); 1027 return 0; 1028 1029 err_cmd: 1030 mlx5_core_destroy_cq(dev->mdev, &cq->mcq); 1031 1032 err_cqb: 1033 kvfree(cqb); 1034 if (udata) 1035 destroy_cq_user(cq, udata); 1036 else 1037 destroy_cq_kernel(dev, cq); 1038 return err; 1039 } 1040 1041 int mlx5_ib_destroy_cq(struct ib_cq *cq, struct ib_udata *udata) 1042 { 1043 struct mlx5_ib_dev *dev = to_mdev(cq->device); 1044 struct mlx5_ib_cq *mcq = to_mcq(cq); 1045 int ret; 1046 1047 ret = mlx5_core_destroy_cq(dev->mdev, &mcq->mcq); 1048 if (ret) 1049 return ret; 1050 1051 if (udata) 1052 destroy_cq_user(mcq, udata); 1053 else 1054 destroy_cq_kernel(dev, mcq); 1055 return 0; 1056 } 1057 1058 static int is_equal_rsn(struct mlx5_cqe64 *cqe64, u32 rsn) 1059 { 1060 return rsn == (ntohl(cqe64->sop_drop_qpn) & 0xffffff); 1061 } 1062 1063 void __mlx5_ib_cq_clean(struct mlx5_ib_cq *cq, u32 rsn, struct mlx5_ib_srq *srq) 1064 { 1065 struct mlx5_cqe64 *cqe64, *dest64; 1066 void *cqe, *dest; 1067 u32 prod_index; 1068 int nfreed = 0; 1069 u8 owner_bit; 1070 1071 if (!cq) 1072 return; 1073 1074 /* First we need to find the current producer index, so we 1075 * know where to start cleaning from. It doesn't matter if HW 1076 * adds new entries after this loop -- the QP we're worried 1077 * about is already in RESET, so the new entries won't come 1078 * from our QP and therefore don't need to be checked. 1079 */ 1080 for (prod_index = cq->mcq.cons_index; get_sw_cqe(cq, prod_index); prod_index++) 1081 if (prod_index == cq->mcq.cons_index + cq->ibcq.cqe) 1082 break; 1083 1084 /* Now sweep backwards through the CQ, removing CQ entries 1085 * that match our QP by copying older entries on top of them. 1086 */ 1087 while ((int) --prod_index - (int) cq->mcq.cons_index >= 0) { 1088 cqe = get_cqe(cq, prod_index & cq->ibcq.cqe); 1089 cqe64 = (cq->mcq.cqe_sz == 64) ? cqe : cqe + 64; 1090 if (is_equal_rsn(cqe64, rsn)) { 1091 if (srq && (ntohl(cqe64->srqn) & 0xffffff)) 1092 mlx5_ib_free_srq_wqe(srq, be16_to_cpu(cqe64->wqe_counter)); 1093 ++nfreed; 1094 } else if (nfreed) { 1095 dest = get_cqe(cq, (prod_index + nfreed) & cq->ibcq.cqe); 1096 dest64 = (cq->mcq.cqe_sz == 64) ? dest : dest + 64; 1097 owner_bit = dest64->op_own & MLX5_CQE_OWNER_MASK; 1098 memcpy(dest, cqe, cq->mcq.cqe_sz); 1099 dest64->op_own = owner_bit | 1100 (dest64->op_own & ~MLX5_CQE_OWNER_MASK); 1101 } 1102 } 1103 1104 if (nfreed) { 1105 cq->mcq.cons_index += nfreed; 1106 /* Make sure update of buffer contents is done before 1107 * updating consumer index. 1108 */ 1109 wmb(); 1110 mlx5_cq_set_ci(&cq->mcq); 1111 } 1112 } 1113 1114 void mlx5_ib_cq_clean(struct mlx5_ib_cq *cq, u32 qpn, struct mlx5_ib_srq *srq) 1115 { 1116 if (!cq) 1117 return; 1118 1119 spin_lock_irq(&cq->lock); 1120 __mlx5_ib_cq_clean(cq, qpn, srq); 1121 spin_unlock_irq(&cq->lock); 1122 } 1123 1124 int mlx5_ib_modify_cq(struct ib_cq *cq, u16 cq_count, u16 cq_period) 1125 { 1126 struct mlx5_ib_dev *dev = to_mdev(cq->device); 1127 struct mlx5_ib_cq *mcq = to_mcq(cq); 1128 int err; 1129 1130 if (!MLX5_CAP_GEN(dev->mdev, cq_moderation)) 1131 return -EOPNOTSUPP; 1132 1133 if (cq_period > MLX5_MAX_CQ_PERIOD) 1134 return -EINVAL; 1135 1136 err = mlx5_core_modify_cq_moderation(dev->mdev, &mcq->mcq, 1137 cq_period, cq_count); 1138 if (err) 1139 mlx5_ib_warn(dev, "modify cq 0x%x failed\n", mcq->mcq.cqn); 1140 1141 return err; 1142 } 1143 1144 static int resize_user(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *cq, 1145 int entries, struct ib_udata *udata, 1146 int *cqe_size) 1147 { 1148 struct mlx5_ib_resize_cq ucmd; 1149 struct ib_umem *umem; 1150 int err; 1151 1152 err = ib_copy_from_udata(&ucmd, udata, sizeof(ucmd)); 1153 if (err) 1154 return err; 1155 1156 if (ucmd.reserved0 || ucmd.reserved1) 1157 return -EINVAL; 1158 1159 /* check multiplication overflow */ 1160 if (ucmd.cqe_size && SIZE_MAX / ucmd.cqe_size <= entries - 1) 1161 return -EINVAL; 1162 1163 umem = ib_umem_get(&dev->ib_dev, ucmd.buf_addr, 1164 (size_t)ucmd.cqe_size * entries, 1165 IB_ACCESS_LOCAL_WRITE); 1166 if (IS_ERR(umem)) { 1167 err = PTR_ERR(umem); 1168 return err; 1169 } 1170 1171 cq->resize_umem = umem; 1172 *cqe_size = ucmd.cqe_size; 1173 1174 return 0; 1175 } 1176 1177 static int resize_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *cq, 1178 int entries, int cqe_size) 1179 { 1180 int err; 1181 1182 cq->resize_buf = kzalloc(sizeof(*cq->resize_buf), GFP_KERNEL); 1183 if (!cq->resize_buf) 1184 return -ENOMEM; 1185 1186 err = alloc_cq_frag_buf(dev, cq->resize_buf, entries, cqe_size); 1187 if (err) 1188 goto ex; 1189 1190 init_cq_frag_buf(cq->resize_buf); 1191 1192 return 0; 1193 1194 ex: 1195 kfree(cq->resize_buf); 1196 return err; 1197 } 1198 1199 static int copy_resize_cqes(struct mlx5_ib_cq *cq) 1200 { 1201 struct mlx5_ib_dev *dev = to_mdev(cq->ibcq.device); 1202 struct mlx5_cqe64 *scqe64; 1203 struct mlx5_cqe64 *dcqe64; 1204 void *start_cqe; 1205 void *scqe; 1206 void *dcqe; 1207 int ssize; 1208 int dsize; 1209 int i; 1210 u8 sw_own; 1211 1212 ssize = cq->buf.cqe_size; 1213 dsize = cq->resize_buf->cqe_size; 1214 if (ssize != dsize) { 1215 mlx5_ib_warn(dev, "resize from different cqe size is not supported\n"); 1216 return -EINVAL; 1217 } 1218 1219 i = cq->mcq.cons_index; 1220 scqe = get_sw_cqe(cq, i); 1221 scqe64 = ssize == 64 ? scqe : scqe + 64; 1222 start_cqe = scqe; 1223 if (!scqe) { 1224 mlx5_ib_warn(dev, "expected cqe in sw ownership\n"); 1225 return -EINVAL; 1226 } 1227 1228 while (get_cqe_opcode(scqe64) != MLX5_CQE_RESIZE_CQ) { 1229 dcqe = mlx5_frag_buf_get_wqe(&cq->resize_buf->fbc, 1230 (i + 1) & cq->resize_buf->nent); 1231 dcqe64 = dsize == 64 ? dcqe : dcqe + 64; 1232 sw_own = sw_ownership_bit(i + 1, cq->resize_buf->nent); 1233 memcpy(dcqe, scqe, dsize); 1234 dcqe64->op_own = (dcqe64->op_own & ~MLX5_CQE_OWNER_MASK) | sw_own; 1235 1236 ++i; 1237 scqe = get_sw_cqe(cq, i); 1238 scqe64 = ssize == 64 ? scqe : scqe + 64; 1239 if (!scqe) { 1240 mlx5_ib_warn(dev, "expected cqe in sw ownership\n"); 1241 return -EINVAL; 1242 } 1243 1244 if (scqe == start_cqe) { 1245 pr_warn("resize CQ failed to get resize CQE, CQN 0x%x\n", 1246 cq->mcq.cqn); 1247 return -ENOMEM; 1248 } 1249 } 1250 ++cq->mcq.cons_index; 1251 return 0; 1252 } 1253 1254 int mlx5_ib_resize_cq(struct ib_cq *ibcq, int entries, struct ib_udata *udata) 1255 { 1256 struct mlx5_ib_dev *dev = to_mdev(ibcq->device); 1257 struct mlx5_ib_cq *cq = to_mcq(ibcq); 1258 void *cqc; 1259 u32 *in; 1260 int err; 1261 int npas; 1262 __be64 *pas; 1263 unsigned int page_offset_quantized = 0; 1264 unsigned int page_shift; 1265 int inlen; 1266 int cqe_size; 1267 unsigned long flags; 1268 1269 if (!MLX5_CAP_GEN(dev->mdev, cq_resize)) { 1270 pr_info("Firmware does not support resize CQ\n"); 1271 return -ENOSYS; 1272 } 1273 1274 if (entries < 1 || 1275 entries > (1 << MLX5_CAP_GEN(dev->mdev, log_max_cq_sz))) { 1276 mlx5_ib_warn(dev, "wrong entries number %d, max %d\n", 1277 entries, 1278 1 << MLX5_CAP_GEN(dev->mdev, log_max_cq_sz)); 1279 return -EINVAL; 1280 } 1281 1282 entries = roundup_pow_of_two(entries + 1); 1283 if (entries > (1 << MLX5_CAP_GEN(dev->mdev, log_max_cq_sz)) + 1) 1284 return -EINVAL; 1285 1286 if (entries == ibcq->cqe + 1) 1287 return 0; 1288 1289 mutex_lock(&cq->resize_mutex); 1290 if (udata) { 1291 unsigned long page_size; 1292 1293 err = resize_user(dev, cq, entries, udata, &cqe_size); 1294 if (err) 1295 goto ex; 1296 1297 page_size = mlx5_umem_find_best_cq_quantized_pgoff( 1298 cq->resize_umem, cqc, log_page_size, 1299 MLX5_ADAPTER_PAGE_SHIFT, page_offset, 64, 1300 &page_offset_quantized); 1301 if (!page_size) { 1302 err = -EINVAL; 1303 goto ex_resize; 1304 } 1305 npas = ib_umem_num_dma_blocks(cq->resize_umem, page_size); 1306 page_shift = order_base_2(page_size); 1307 } else { 1308 struct mlx5_frag_buf *frag_buf; 1309 1310 cqe_size = 64; 1311 err = resize_kernel(dev, cq, entries, cqe_size); 1312 if (err) 1313 goto ex; 1314 frag_buf = &cq->resize_buf->frag_buf; 1315 npas = frag_buf->npages; 1316 page_shift = frag_buf->page_shift; 1317 } 1318 1319 inlen = MLX5_ST_SZ_BYTES(modify_cq_in) + 1320 MLX5_FLD_SZ_BYTES(modify_cq_in, pas[0]) * npas; 1321 1322 in = kvzalloc(inlen, GFP_KERNEL); 1323 if (!in) { 1324 err = -ENOMEM; 1325 goto ex_resize; 1326 } 1327 1328 pas = (__be64 *)MLX5_ADDR_OF(modify_cq_in, in, pas); 1329 if (udata) 1330 mlx5_ib_populate_pas(cq->resize_umem, 1UL << page_shift, pas, 1331 0); 1332 else 1333 mlx5_fill_page_frag_array(&cq->resize_buf->frag_buf, pas); 1334 1335 MLX5_SET(modify_cq_in, in, 1336 modify_field_select_resize_field_select.resize_field_select.resize_field_select, 1337 MLX5_MODIFY_CQ_MASK_LOG_SIZE | 1338 MLX5_MODIFY_CQ_MASK_PG_OFFSET | 1339 MLX5_MODIFY_CQ_MASK_PG_SIZE); 1340 1341 cqc = MLX5_ADDR_OF(modify_cq_in, in, cq_context); 1342 1343 MLX5_SET(cqc, cqc, log_page_size, 1344 page_shift - MLX5_ADAPTER_PAGE_SHIFT); 1345 MLX5_SET(cqc, cqc, page_offset, page_offset_quantized); 1346 MLX5_SET(cqc, cqc, cqe_sz, 1347 cqe_sz_to_mlx_sz(cqe_size, 1348 cq->private_flags & 1349 MLX5_IB_CQ_PR_FLAGS_CQE_128_PAD)); 1350 MLX5_SET(cqc, cqc, log_cq_size, ilog2(entries)); 1351 1352 MLX5_SET(modify_cq_in, in, op_mod, MLX5_CQ_OPMOD_RESIZE); 1353 MLX5_SET(modify_cq_in, in, cqn, cq->mcq.cqn); 1354 1355 err = mlx5_core_modify_cq(dev->mdev, &cq->mcq, in, inlen); 1356 if (err) 1357 goto ex_alloc; 1358 1359 if (udata) { 1360 cq->ibcq.cqe = entries - 1; 1361 ib_umem_release(cq->buf.umem); 1362 cq->buf.umem = cq->resize_umem; 1363 cq->resize_umem = NULL; 1364 } else { 1365 struct mlx5_ib_cq_buf tbuf; 1366 int resized = 0; 1367 1368 spin_lock_irqsave(&cq->lock, flags); 1369 if (cq->resize_buf) { 1370 err = copy_resize_cqes(cq); 1371 if (!err) { 1372 tbuf = cq->buf; 1373 cq->buf = *cq->resize_buf; 1374 kfree(cq->resize_buf); 1375 cq->resize_buf = NULL; 1376 resized = 1; 1377 } 1378 } 1379 cq->ibcq.cqe = entries - 1; 1380 spin_unlock_irqrestore(&cq->lock, flags); 1381 if (resized) 1382 free_cq_buf(dev, &tbuf); 1383 } 1384 mutex_unlock(&cq->resize_mutex); 1385 1386 kvfree(in); 1387 return 0; 1388 1389 ex_alloc: 1390 kvfree(in); 1391 1392 ex_resize: 1393 ib_umem_release(cq->resize_umem); 1394 if (!udata) { 1395 free_cq_buf(dev, cq->resize_buf); 1396 cq->resize_buf = NULL; 1397 } 1398 ex: 1399 mutex_unlock(&cq->resize_mutex); 1400 return err; 1401 } 1402 1403 int mlx5_ib_get_cqe_size(struct ib_cq *ibcq) 1404 { 1405 struct mlx5_ib_cq *cq; 1406 1407 if (!ibcq) 1408 return 128; 1409 1410 cq = to_mcq(ibcq); 1411 return cq->cqe_size; 1412 } 1413 1414 /* Called from atomic context */ 1415 int mlx5_ib_generate_wc(struct ib_cq *ibcq, struct ib_wc *wc) 1416 { 1417 struct mlx5_ib_wc *soft_wc; 1418 struct mlx5_ib_cq *cq = to_mcq(ibcq); 1419 unsigned long flags; 1420 1421 soft_wc = kmalloc(sizeof(*soft_wc), GFP_ATOMIC); 1422 if (!soft_wc) 1423 return -ENOMEM; 1424 1425 soft_wc->wc = *wc; 1426 spin_lock_irqsave(&cq->lock, flags); 1427 list_add_tail(&soft_wc->list, &cq->wc_list); 1428 if (cq->notify_flags == IB_CQ_NEXT_COMP || 1429 wc->status != IB_WC_SUCCESS) { 1430 cq->notify_flags = 0; 1431 schedule_work(&cq->notify_work); 1432 } 1433 spin_unlock_irqrestore(&cq->lock, flags); 1434 1435 return 0; 1436 } 1437