1 /* 2 * Copyright (c) 2007 Cisco Systems, Inc. All rights reserved. 3 * Copyright (c) 2007, 2008 Mellanox Technologies. All rights reserved. 4 * 5 * This software is available to you under a choice of one of two 6 * licenses. You may choose to be licensed under the terms of the GNU 7 * General Public License (GPL) Version 2, available from the file 8 * COPYING in the main directory of this source tree, or the 9 * OpenIB.org BSD license below: 10 * 11 * Redistribution and use in source and binary forms, with or 12 * without modification, are permitted provided that the following 13 * conditions are met: 14 * 15 * - Redistributions of source code must retain the above 16 * copyright notice, this list of conditions and the following 17 * disclaimer. 18 * 19 * - Redistributions in binary form must reproduce the above 20 * copyright notice, this list of conditions and the following 21 * disclaimer in the documentation and/or other materials 22 * provided with the distribution. 23 * 24 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, 25 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF 26 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND 27 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS 28 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN 29 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN 30 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE 31 * SOFTWARE. 32 */ 33 34 #include <linux/log2.h> 35 #include <linux/etherdevice.h> 36 #include <net/ip.h> 37 #include <linux/slab.h> 38 #include <linux/netdevice.h> 39 40 #include <rdma/ib_cache.h> 41 #include <rdma/ib_pack.h> 42 #include <rdma/ib_addr.h> 43 #include <rdma/ib_mad.h> 44 #include <rdma/uverbs_ioctl.h> 45 46 #include <linux/mlx4/driver.h> 47 #include <linux/mlx4/qp.h> 48 49 #include "mlx4_ib.h" 50 #include <rdma/mlx4-abi.h> 51 52 static void mlx4_ib_lock_cqs(struct mlx4_ib_cq *send_cq, 53 struct mlx4_ib_cq *recv_cq); 54 static void mlx4_ib_unlock_cqs(struct mlx4_ib_cq *send_cq, 55 struct mlx4_ib_cq *recv_cq); 56 static int _mlx4_ib_modify_wq(struct ib_wq *ibwq, enum ib_wq_state new_state, 57 struct ib_udata *udata); 58 59 enum { 60 MLX4_IB_ACK_REQ_FREQ = 8, 61 }; 62 63 enum { 64 MLX4_IB_DEFAULT_SCHED_QUEUE = 0x83, 65 MLX4_IB_DEFAULT_QP0_SCHED_QUEUE = 0x3f, 66 MLX4_IB_LINK_TYPE_IB = 0, 67 MLX4_IB_LINK_TYPE_ETH = 1 68 }; 69 70 enum { 71 MLX4_IB_MIN_SQ_STRIDE = 6, 72 MLX4_IB_CACHE_LINE_SIZE = 64, 73 }; 74 75 enum { 76 MLX4_RAW_QP_MTU = 7, 77 MLX4_RAW_QP_MSGMAX = 31, 78 }; 79 80 #ifndef ETH_ALEN 81 #define ETH_ALEN 6 82 #endif 83 84 static const __be32 mlx4_ib_opcode[] = { 85 [IB_WR_SEND] = cpu_to_be32(MLX4_OPCODE_SEND), 86 [IB_WR_LSO] = cpu_to_be32(MLX4_OPCODE_LSO), 87 [IB_WR_SEND_WITH_IMM] = cpu_to_be32(MLX4_OPCODE_SEND_IMM), 88 [IB_WR_RDMA_WRITE] = cpu_to_be32(MLX4_OPCODE_RDMA_WRITE), 89 [IB_WR_RDMA_WRITE_WITH_IMM] = cpu_to_be32(MLX4_OPCODE_RDMA_WRITE_IMM), 90 [IB_WR_RDMA_READ] = cpu_to_be32(MLX4_OPCODE_RDMA_READ), 91 [IB_WR_ATOMIC_CMP_AND_SWP] = cpu_to_be32(MLX4_OPCODE_ATOMIC_CS), 92 [IB_WR_ATOMIC_FETCH_AND_ADD] = cpu_to_be32(MLX4_OPCODE_ATOMIC_FA), 93 [IB_WR_SEND_WITH_INV] = cpu_to_be32(MLX4_OPCODE_SEND_INVAL), 94 [IB_WR_LOCAL_INV] = cpu_to_be32(MLX4_OPCODE_LOCAL_INVAL), 95 [IB_WR_REG_MR] = cpu_to_be32(MLX4_OPCODE_FMR), 96 [IB_WR_MASKED_ATOMIC_CMP_AND_SWP] = cpu_to_be32(MLX4_OPCODE_MASKED_ATOMIC_CS), 97 [IB_WR_MASKED_ATOMIC_FETCH_AND_ADD] = cpu_to_be32(MLX4_OPCODE_MASKED_ATOMIC_FA), 98 }; 99 100 enum mlx4_ib_source_type { 101 MLX4_IB_QP_SRC = 0, 102 MLX4_IB_RWQ_SRC = 1, 103 }; 104 105 static int is_tunnel_qp(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp) 106 { 107 if (!mlx4_is_master(dev->dev)) 108 return 0; 109 110 return qp->mqp.qpn >= dev->dev->phys_caps.base_tunnel_sqpn && 111 qp->mqp.qpn < dev->dev->phys_caps.base_tunnel_sqpn + 112 8 * MLX4_MFUNC_MAX; 113 } 114 115 static int is_sqp(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp) 116 { 117 int proxy_sqp = 0; 118 int real_sqp = 0; 119 int i; 120 /* PPF or Native -- real SQP */ 121 real_sqp = ((mlx4_is_master(dev->dev) || !mlx4_is_mfunc(dev->dev)) && 122 qp->mqp.qpn >= dev->dev->phys_caps.base_sqpn && 123 qp->mqp.qpn <= dev->dev->phys_caps.base_sqpn + 3); 124 if (real_sqp) 125 return 1; 126 /* VF or PF -- proxy SQP */ 127 if (mlx4_is_mfunc(dev->dev)) { 128 for (i = 0; i < dev->dev->caps.num_ports; i++) { 129 if (qp->mqp.qpn == dev->dev->caps.spec_qps[i].qp0_proxy || 130 qp->mqp.qpn == dev->dev->caps.spec_qps[i].qp1_proxy) { 131 proxy_sqp = 1; 132 break; 133 } 134 } 135 } 136 if (proxy_sqp) 137 return 1; 138 139 return !!(qp->flags & MLX4_IB_ROCE_V2_GSI_QP); 140 } 141 142 /* used for INIT/CLOSE port logic */ 143 static int is_qp0(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp) 144 { 145 int proxy_qp0 = 0; 146 int real_qp0 = 0; 147 int i; 148 /* PPF or Native -- real QP0 */ 149 real_qp0 = ((mlx4_is_master(dev->dev) || !mlx4_is_mfunc(dev->dev)) && 150 qp->mqp.qpn >= dev->dev->phys_caps.base_sqpn && 151 qp->mqp.qpn <= dev->dev->phys_caps.base_sqpn + 1); 152 if (real_qp0) 153 return 1; 154 /* VF or PF -- proxy QP0 */ 155 if (mlx4_is_mfunc(dev->dev)) { 156 for (i = 0; i < dev->dev->caps.num_ports; i++) { 157 if (qp->mqp.qpn == dev->dev->caps.spec_qps[i].qp0_proxy) { 158 proxy_qp0 = 1; 159 break; 160 } 161 } 162 } 163 return proxy_qp0; 164 } 165 166 static void *get_wqe(struct mlx4_ib_qp *qp, int offset) 167 { 168 return mlx4_buf_offset(&qp->buf, offset); 169 } 170 171 static void *get_recv_wqe(struct mlx4_ib_qp *qp, int n) 172 { 173 return get_wqe(qp, qp->rq.offset + (n << qp->rq.wqe_shift)); 174 } 175 176 static void *get_send_wqe(struct mlx4_ib_qp *qp, int n) 177 { 178 return get_wqe(qp, qp->sq.offset + (n << qp->sq.wqe_shift)); 179 } 180 181 /* 182 * Stamp a SQ WQE so that it is invalid if prefetched by marking the 183 * first four bytes of every 64 byte chunk with 0xffffffff, except for 184 * the very first chunk of the WQE. 185 */ 186 static void stamp_send_wqe(struct mlx4_ib_qp *qp, int n) 187 { 188 __be32 *wqe; 189 int i; 190 int s; 191 void *buf; 192 struct mlx4_wqe_ctrl_seg *ctrl; 193 194 buf = get_send_wqe(qp, n & (qp->sq.wqe_cnt - 1)); 195 ctrl = (struct mlx4_wqe_ctrl_seg *)buf; 196 s = (ctrl->qpn_vlan.fence_size & 0x3f) << 4; 197 for (i = 64; i < s; i += 64) { 198 wqe = buf + i; 199 *wqe = cpu_to_be32(0xffffffff); 200 } 201 } 202 203 static void mlx4_ib_qp_event(struct mlx4_qp *qp, enum mlx4_event type) 204 { 205 struct ib_event event; 206 struct ib_qp *ibqp = &to_mibqp(qp)->ibqp; 207 208 if (type == MLX4_EVENT_TYPE_PATH_MIG) 209 to_mibqp(qp)->port = to_mibqp(qp)->alt_port; 210 211 if (ibqp->event_handler) { 212 event.device = ibqp->device; 213 event.element.qp = ibqp; 214 switch (type) { 215 case MLX4_EVENT_TYPE_PATH_MIG: 216 event.event = IB_EVENT_PATH_MIG; 217 break; 218 case MLX4_EVENT_TYPE_COMM_EST: 219 event.event = IB_EVENT_COMM_EST; 220 break; 221 case MLX4_EVENT_TYPE_SQ_DRAINED: 222 event.event = IB_EVENT_SQ_DRAINED; 223 break; 224 case MLX4_EVENT_TYPE_SRQ_QP_LAST_WQE: 225 event.event = IB_EVENT_QP_LAST_WQE_REACHED; 226 break; 227 case MLX4_EVENT_TYPE_WQ_CATAS_ERROR: 228 event.event = IB_EVENT_QP_FATAL; 229 break; 230 case MLX4_EVENT_TYPE_PATH_MIG_FAILED: 231 event.event = IB_EVENT_PATH_MIG_ERR; 232 break; 233 case MLX4_EVENT_TYPE_WQ_INVAL_REQ_ERROR: 234 event.event = IB_EVENT_QP_REQ_ERR; 235 break; 236 case MLX4_EVENT_TYPE_WQ_ACCESS_ERROR: 237 event.event = IB_EVENT_QP_ACCESS_ERR; 238 break; 239 default: 240 pr_warn("Unexpected event type %d " 241 "on QP %06x\n", type, qp->qpn); 242 return; 243 } 244 245 ibqp->event_handler(&event, ibqp->qp_context); 246 } 247 } 248 249 static void mlx4_ib_wq_event(struct mlx4_qp *qp, enum mlx4_event type) 250 { 251 pr_warn_ratelimited("Unexpected event type %d on WQ 0x%06x. Events are not supported for WQs\n", 252 type, qp->qpn); 253 } 254 255 static int send_wqe_overhead(enum mlx4_ib_qp_type type, u32 flags) 256 { 257 /* 258 * UD WQEs must have a datagram segment. 259 * RC and UC WQEs might have a remote address segment. 260 * MLX WQEs need two extra inline data segments (for the UD 261 * header and space for the ICRC). 262 */ 263 switch (type) { 264 case MLX4_IB_QPT_UD: 265 return sizeof (struct mlx4_wqe_ctrl_seg) + 266 sizeof (struct mlx4_wqe_datagram_seg) + 267 ((flags & MLX4_IB_QP_LSO) ? MLX4_IB_LSO_HEADER_SPARE : 0); 268 case MLX4_IB_QPT_PROXY_SMI_OWNER: 269 case MLX4_IB_QPT_PROXY_SMI: 270 case MLX4_IB_QPT_PROXY_GSI: 271 return sizeof (struct mlx4_wqe_ctrl_seg) + 272 sizeof (struct mlx4_wqe_datagram_seg) + 64; 273 case MLX4_IB_QPT_TUN_SMI_OWNER: 274 case MLX4_IB_QPT_TUN_GSI: 275 return sizeof (struct mlx4_wqe_ctrl_seg) + 276 sizeof (struct mlx4_wqe_datagram_seg); 277 278 case MLX4_IB_QPT_UC: 279 return sizeof (struct mlx4_wqe_ctrl_seg) + 280 sizeof (struct mlx4_wqe_raddr_seg); 281 case MLX4_IB_QPT_RC: 282 return sizeof (struct mlx4_wqe_ctrl_seg) + 283 sizeof (struct mlx4_wqe_masked_atomic_seg) + 284 sizeof (struct mlx4_wqe_raddr_seg); 285 case MLX4_IB_QPT_SMI: 286 case MLX4_IB_QPT_GSI: 287 return sizeof (struct mlx4_wqe_ctrl_seg) + 288 ALIGN(MLX4_IB_UD_HEADER_SIZE + 289 DIV_ROUND_UP(MLX4_IB_UD_HEADER_SIZE, 290 MLX4_INLINE_ALIGN) * 291 sizeof (struct mlx4_wqe_inline_seg), 292 sizeof (struct mlx4_wqe_data_seg)) + 293 ALIGN(4 + 294 sizeof (struct mlx4_wqe_inline_seg), 295 sizeof (struct mlx4_wqe_data_seg)); 296 default: 297 return sizeof (struct mlx4_wqe_ctrl_seg); 298 } 299 } 300 301 static int set_rq_size(struct mlx4_ib_dev *dev, struct ib_qp_cap *cap, 302 bool is_user, bool has_rq, struct mlx4_ib_qp *qp, 303 u32 inl_recv_sz) 304 { 305 /* Sanity check RQ size before proceeding */ 306 if (cap->max_recv_wr > dev->dev->caps.max_wqes - MLX4_IB_SQ_MAX_SPARE || 307 cap->max_recv_sge > min(dev->dev->caps.max_sq_sg, dev->dev->caps.max_rq_sg)) 308 return -EINVAL; 309 310 if (!has_rq) { 311 if (cap->max_recv_wr || inl_recv_sz) 312 return -EINVAL; 313 314 qp->rq.wqe_cnt = qp->rq.max_gs = 0; 315 } else { 316 u32 max_inl_recv_sz = dev->dev->caps.max_rq_sg * 317 sizeof(struct mlx4_wqe_data_seg); 318 u32 wqe_size; 319 320 /* HW requires >= 1 RQ entry with >= 1 gather entry */ 321 if (is_user && (!cap->max_recv_wr || !cap->max_recv_sge || 322 inl_recv_sz > max_inl_recv_sz)) 323 return -EINVAL; 324 325 qp->rq.wqe_cnt = roundup_pow_of_two(max(1U, cap->max_recv_wr)); 326 qp->rq.max_gs = roundup_pow_of_two(max(1U, cap->max_recv_sge)); 327 wqe_size = qp->rq.max_gs * sizeof(struct mlx4_wqe_data_seg); 328 qp->rq.wqe_shift = ilog2(max_t(u32, wqe_size, inl_recv_sz)); 329 } 330 331 /* leave userspace return values as they were, so as not to break ABI */ 332 if (is_user) { 333 cap->max_recv_wr = qp->rq.max_post = qp->rq.wqe_cnt; 334 cap->max_recv_sge = qp->rq.max_gs; 335 } else { 336 cap->max_recv_wr = qp->rq.max_post = 337 min(dev->dev->caps.max_wqes - MLX4_IB_SQ_MAX_SPARE, qp->rq.wqe_cnt); 338 cap->max_recv_sge = min(qp->rq.max_gs, 339 min(dev->dev->caps.max_sq_sg, 340 dev->dev->caps.max_rq_sg)); 341 } 342 343 return 0; 344 } 345 346 static int set_kernel_sq_size(struct mlx4_ib_dev *dev, struct ib_qp_cap *cap, 347 enum mlx4_ib_qp_type type, struct mlx4_ib_qp *qp) 348 { 349 int s; 350 351 /* Sanity check SQ size before proceeding */ 352 if (cap->max_send_wr > (dev->dev->caps.max_wqes - MLX4_IB_SQ_MAX_SPARE) || 353 cap->max_send_sge > min(dev->dev->caps.max_sq_sg, dev->dev->caps.max_rq_sg) || 354 cap->max_inline_data + send_wqe_overhead(type, qp->flags) + 355 sizeof (struct mlx4_wqe_inline_seg) > dev->dev->caps.max_sq_desc_sz) 356 return -EINVAL; 357 358 /* 359 * For MLX transport we need 2 extra S/G entries: 360 * one for the header and one for the checksum at the end 361 */ 362 if ((type == MLX4_IB_QPT_SMI || type == MLX4_IB_QPT_GSI || 363 type & (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER)) && 364 cap->max_send_sge + 2 > dev->dev->caps.max_sq_sg) 365 return -EINVAL; 366 367 s = max(cap->max_send_sge * sizeof (struct mlx4_wqe_data_seg), 368 cap->max_inline_data + sizeof (struct mlx4_wqe_inline_seg)) + 369 send_wqe_overhead(type, qp->flags); 370 371 if (s > dev->dev->caps.max_sq_desc_sz) 372 return -EINVAL; 373 374 qp->sq.wqe_shift = ilog2(roundup_pow_of_two(s)); 375 376 /* 377 * We need to leave 2 KB + 1 WR of headroom in the SQ to 378 * allow HW to prefetch. 379 */ 380 qp->sq_spare_wqes = MLX4_IB_SQ_HEADROOM(qp->sq.wqe_shift); 381 qp->sq.wqe_cnt = roundup_pow_of_two(cap->max_send_wr + 382 qp->sq_spare_wqes); 383 384 qp->sq.max_gs = 385 (min(dev->dev->caps.max_sq_desc_sz, 386 (1 << qp->sq.wqe_shift)) - 387 send_wqe_overhead(type, qp->flags)) / 388 sizeof (struct mlx4_wqe_data_seg); 389 390 qp->buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) + 391 (qp->sq.wqe_cnt << qp->sq.wqe_shift); 392 if (qp->rq.wqe_shift > qp->sq.wqe_shift) { 393 qp->rq.offset = 0; 394 qp->sq.offset = qp->rq.wqe_cnt << qp->rq.wqe_shift; 395 } else { 396 qp->rq.offset = qp->sq.wqe_cnt << qp->sq.wqe_shift; 397 qp->sq.offset = 0; 398 } 399 400 cap->max_send_wr = qp->sq.max_post = 401 qp->sq.wqe_cnt - qp->sq_spare_wqes; 402 cap->max_send_sge = min(qp->sq.max_gs, 403 min(dev->dev->caps.max_sq_sg, 404 dev->dev->caps.max_rq_sg)); 405 /* We don't support inline sends for kernel QPs (yet) */ 406 cap->max_inline_data = 0; 407 408 return 0; 409 } 410 411 static int set_user_sq_size(struct mlx4_ib_dev *dev, 412 struct mlx4_ib_qp *qp, 413 struct mlx4_ib_create_qp *ucmd) 414 { 415 /* Sanity check SQ size before proceeding */ 416 if ((1 << ucmd->log_sq_bb_count) > dev->dev->caps.max_wqes || 417 ucmd->log_sq_stride > 418 ilog2(roundup_pow_of_two(dev->dev->caps.max_sq_desc_sz)) || 419 ucmd->log_sq_stride < MLX4_IB_MIN_SQ_STRIDE) 420 return -EINVAL; 421 422 qp->sq.wqe_cnt = 1 << ucmd->log_sq_bb_count; 423 qp->sq.wqe_shift = ucmd->log_sq_stride; 424 425 qp->buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) + 426 (qp->sq.wqe_cnt << qp->sq.wqe_shift); 427 428 return 0; 429 } 430 431 static int alloc_proxy_bufs(struct ib_device *dev, struct mlx4_ib_qp *qp) 432 { 433 int i; 434 435 qp->sqp_proxy_rcv = 436 kmalloc_array(qp->rq.wqe_cnt, sizeof(struct mlx4_ib_buf), 437 GFP_KERNEL); 438 if (!qp->sqp_proxy_rcv) 439 return -ENOMEM; 440 for (i = 0; i < qp->rq.wqe_cnt; i++) { 441 qp->sqp_proxy_rcv[i].addr = 442 kmalloc(sizeof (struct mlx4_ib_proxy_sqp_hdr), 443 GFP_KERNEL); 444 if (!qp->sqp_proxy_rcv[i].addr) 445 goto err; 446 qp->sqp_proxy_rcv[i].map = 447 ib_dma_map_single(dev, qp->sqp_proxy_rcv[i].addr, 448 sizeof (struct mlx4_ib_proxy_sqp_hdr), 449 DMA_FROM_DEVICE); 450 if (ib_dma_mapping_error(dev, qp->sqp_proxy_rcv[i].map)) { 451 kfree(qp->sqp_proxy_rcv[i].addr); 452 goto err; 453 } 454 } 455 return 0; 456 457 err: 458 while (i > 0) { 459 --i; 460 ib_dma_unmap_single(dev, qp->sqp_proxy_rcv[i].map, 461 sizeof (struct mlx4_ib_proxy_sqp_hdr), 462 DMA_FROM_DEVICE); 463 kfree(qp->sqp_proxy_rcv[i].addr); 464 } 465 kfree(qp->sqp_proxy_rcv); 466 qp->sqp_proxy_rcv = NULL; 467 return -ENOMEM; 468 } 469 470 static void free_proxy_bufs(struct ib_device *dev, struct mlx4_ib_qp *qp) 471 { 472 int i; 473 474 for (i = 0; i < qp->rq.wqe_cnt; i++) { 475 ib_dma_unmap_single(dev, qp->sqp_proxy_rcv[i].map, 476 sizeof (struct mlx4_ib_proxy_sqp_hdr), 477 DMA_FROM_DEVICE); 478 kfree(qp->sqp_proxy_rcv[i].addr); 479 } 480 kfree(qp->sqp_proxy_rcv); 481 } 482 483 static bool qp_has_rq(struct ib_qp_init_attr *attr) 484 { 485 if (attr->qp_type == IB_QPT_XRC_INI || attr->qp_type == IB_QPT_XRC_TGT) 486 return false; 487 488 return !attr->srq; 489 } 490 491 static int qp0_enabled_vf(struct mlx4_dev *dev, int qpn) 492 { 493 int i; 494 for (i = 0; i < dev->caps.num_ports; i++) { 495 if (qpn == dev->caps.spec_qps[i].qp0_proxy) 496 return !!dev->caps.spec_qps[i].qp0_qkey; 497 } 498 return 0; 499 } 500 501 static void mlx4_ib_free_qp_counter(struct mlx4_ib_dev *dev, 502 struct mlx4_ib_qp *qp) 503 { 504 mutex_lock(&dev->counters_table[qp->port - 1].mutex); 505 mlx4_counter_free(dev->dev, qp->counter_index->index); 506 list_del(&qp->counter_index->list); 507 mutex_unlock(&dev->counters_table[qp->port - 1].mutex); 508 509 kfree(qp->counter_index); 510 qp->counter_index = NULL; 511 } 512 513 static int set_qp_rss(struct mlx4_ib_dev *dev, struct mlx4_ib_rss *rss_ctx, 514 struct ib_qp_init_attr *init_attr, 515 struct mlx4_ib_create_qp_rss *ucmd) 516 { 517 rss_ctx->base_qpn_tbl_sz = init_attr->rwq_ind_tbl->ind_tbl[0]->wq_num | 518 (init_attr->rwq_ind_tbl->log_ind_tbl_size << 24); 519 520 if ((ucmd->rx_hash_function == MLX4_IB_RX_HASH_FUNC_TOEPLITZ) && 521 (dev->dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_RSS_TOP)) { 522 memcpy(rss_ctx->rss_key, ucmd->rx_hash_key, 523 MLX4_EN_RSS_KEY_SIZE); 524 } else { 525 pr_debug("RX Hash function is not supported\n"); 526 return (-EOPNOTSUPP); 527 } 528 529 if (ucmd->rx_hash_fields_mask & ~(MLX4_IB_RX_HASH_SRC_IPV4 | 530 MLX4_IB_RX_HASH_DST_IPV4 | 531 MLX4_IB_RX_HASH_SRC_IPV6 | 532 MLX4_IB_RX_HASH_DST_IPV6 | 533 MLX4_IB_RX_HASH_SRC_PORT_TCP | 534 MLX4_IB_RX_HASH_DST_PORT_TCP | 535 MLX4_IB_RX_HASH_SRC_PORT_UDP | 536 MLX4_IB_RX_HASH_DST_PORT_UDP | 537 MLX4_IB_RX_HASH_INNER)) { 538 pr_debug("RX Hash fields_mask has unsupported mask (0x%llx)\n", 539 ucmd->rx_hash_fields_mask); 540 return (-EOPNOTSUPP); 541 } 542 543 if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_IPV4) && 544 (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_IPV4)) { 545 rss_ctx->flags = MLX4_RSS_IPV4; 546 } else if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_IPV4) || 547 (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_IPV4)) { 548 pr_debug("RX Hash fields_mask is not supported - both IPv4 SRC and DST must be set\n"); 549 return (-EOPNOTSUPP); 550 } 551 552 if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_IPV6) && 553 (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_IPV6)) { 554 rss_ctx->flags |= MLX4_RSS_IPV6; 555 } else if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_IPV6) || 556 (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_IPV6)) { 557 pr_debug("RX Hash fields_mask is not supported - both IPv6 SRC and DST must be set\n"); 558 return (-EOPNOTSUPP); 559 } 560 561 if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_PORT_UDP) && 562 (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_PORT_UDP)) { 563 if (!(dev->dev->caps.flags & MLX4_DEV_CAP_FLAG_UDP_RSS)) { 564 pr_debug("RX Hash fields_mask for UDP is not supported\n"); 565 return (-EOPNOTSUPP); 566 } 567 568 if (rss_ctx->flags & MLX4_RSS_IPV4) 569 rss_ctx->flags |= MLX4_RSS_UDP_IPV4; 570 if (rss_ctx->flags & MLX4_RSS_IPV6) 571 rss_ctx->flags |= MLX4_RSS_UDP_IPV6; 572 if (!(rss_ctx->flags & (MLX4_RSS_IPV6 | MLX4_RSS_IPV4))) { 573 pr_debug("RX Hash fields_mask is not supported - UDP must be set with IPv4 or IPv6\n"); 574 return (-EOPNOTSUPP); 575 } 576 } else if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_PORT_UDP) || 577 (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_PORT_UDP)) { 578 pr_debug("RX Hash fields_mask is not supported - both UDP SRC and DST must be set\n"); 579 return (-EOPNOTSUPP); 580 } 581 582 if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_PORT_TCP) && 583 (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_PORT_TCP)) { 584 if (rss_ctx->flags & MLX4_RSS_IPV4) 585 rss_ctx->flags |= MLX4_RSS_TCP_IPV4; 586 if (rss_ctx->flags & MLX4_RSS_IPV6) 587 rss_ctx->flags |= MLX4_RSS_TCP_IPV6; 588 if (!(rss_ctx->flags & (MLX4_RSS_IPV6 | MLX4_RSS_IPV4))) { 589 pr_debug("RX Hash fields_mask is not supported - TCP must be set with IPv4 or IPv6\n"); 590 return (-EOPNOTSUPP); 591 } 592 } else if ((ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_SRC_PORT_TCP) || 593 (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_DST_PORT_TCP)) { 594 pr_debug("RX Hash fields_mask is not supported - both TCP SRC and DST must be set\n"); 595 return (-EOPNOTSUPP); 596 } 597 598 if (ucmd->rx_hash_fields_mask & MLX4_IB_RX_HASH_INNER) { 599 if (dev->dev->caps.tunnel_offload_mode == 600 MLX4_TUNNEL_OFFLOAD_MODE_VXLAN) { 601 /* 602 * Hash according to inner headers if exist, otherwise 603 * according to outer headers. 604 */ 605 rss_ctx->flags |= MLX4_RSS_BY_INNER_HEADERS_IPONLY; 606 } else { 607 pr_debug("RSS Hash for inner headers isn't supported\n"); 608 return (-EOPNOTSUPP); 609 } 610 } 611 612 return 0; 613 } 614 615 static int create_qp_rss(struct mlx4_ib_dev *dev, 616 struct ib_qp_init_attr *init_attr, 617 struct mlx4_ib_create_qp_rss *ucmd, 618 struct mlx4_ib_qp *qp) 619 { 620 int qpn; 621 int err; 622 623 qp->mqp.usage = MLX4_RES_USAGE_USER_VERBS; 624 625 err = mlx4_qp_reserve_range(dev->dev, 1, 1, &qpn, 0, qp->mqp.usage); 626 if (err) 627 return err; 628 629 err = mlx4_qp_alloc(dev->dev, qpn, &qp->mqp); 630 if (err) 631 goto err_qpn; 632 633 INIT_LIST_HEAD(&qp->gid_list); 634 INIT_LIST_HEAD(&qp->steering_rules); 635 636 qp->mlx4_ib_qp_type = MLX4_IB_QPT_RAW_PACKET; 637 qp->state = IB_QPS_RESET; 638 639 /* Set dummy send resources to be compatible with HV and PRM */ 640 qp->sq_no_prefetch = 1; 641 qp->sq.wqe_cnt = 1; 642 qp->sq.wqe_shift = MLX4_IB_MIN_SQ_STRIDE; 643 qp->buf_size = qp->sq.wqe_cnt << MLX4_IB_MIN_SQ_STRIDE; 644 qp->mtt = (to_mqp( 645 (struct ib_qp *)init_attr->rwq_ind_tbl->ind_tbl[0]))->mtt; 646 647 qp->rss_ctx = kzalloc(sizeof(*qp->rss_ctx), GFP_KERNEL); 648 if (!qp->rss_ctx) { 649 err = -ENOMEM; 650 goto err_qp_alloc; 651 } 652 653 err = set_qp_rss(dev, qp->rss_ctx, init_attr, ucmd); 654 if (err) 655 goto err; 656 657 return 0; 658 659 err: 660 kfree(qp->rss_ctx); 661 662 err_qp_alloc: 663 mlx4_qp_remove(dev->dev, &qp->mqp); 664 mlx4_qp_free(dev->dev, &qp->mqp); 665 666 err_qpn: 667 mlx4_qp_release_range(dev->dev, qpn, 1); 668 return err; 669 } 670 671 static int _mlx4_ib_create_qp_rss(struct ib_pd *pd, struct mlx4_ib_qp *qp, 672 struct ib_qp_init_attr *init_attr, 673 struct ib_udata *udata) 674 { 675 struct mlx4_ib_create_qp_rss ucmd = {}; 676 size_t required_cmd_sz; 677 int err; 678 679 if (!udata) { 680 pr_debug("RSS QP with NULL udata\n"); 681 return -EINVAL; 682 } 683 684 if (udata->outlen) 685 return -EOPNOTSUPP; 686 687 required_cmd_sz = offsetof(typeof(ucmd), reserved1) + 688 sizeof(ucmd.reserved1); 689 if (udata->inlen < required_cmd_sz) { 690 pr_debug("invalid inlen\n"); 691 return -EINVAL; 692 } 693 694 if (ib_copy_from_udata(&ucmd, udata, min(sizeof(ucmd), udata->inlen))) { 695 pr_debug("copy failed\n"); 696 return -EFAULT; 697 } 698 699 if (memchr_inv(ucmd.reserved, 0, sizeof(ucmd.reserved))) 700 return -EOPNOTSUPP; 701 702 if (ucmd.comp_mask || ucmd.reserved1) 703 return -EOPNOTSUPP; 704 705 if (udata->inlen > sizeof(ucmd) && 706 !ib_is_udata_cleared(udata, sizeof(ucmd), 707 udata->inlen - sizeof(ucmd))) { 708 pr_debug("inlen is not supported\n"); 709 return -EOPNOTSUPP; 710 } 711 712 if (init_attr->qp_type != IB_QPT_RAW_PACKET) { 713 pr_debug("RSS QP with unsupported QP type %d\n", 714 init_attr->qp_type); 715 return -EOPNOTSUPP; 716 } 717 718 if (init_attr->create_flags) { 719 pr_debug("RSS QP doesn't support create flags\n"); 720 return -EOPNOTSUPP; 721 } 722 723 if (init_attr->send_cq || init_attr->cap.max_send_wr) { 724 pr_debug("RSS QP with unsupported send attributes\n"); 725 return -EOPNOTSUPP; 726 } 727 728 qp->pri.vid = 0xFFFF; 729 qp->alt.vid = 0xFFFF; 730 731 err = create_qp_rss(to_mdev(pd->device), init_attr, &ucmd, qp); 732 if (err) 733 return err; 734 735 qp->ibqp.qp_num = qp->mqp.qpn; 736 return 0; 737 } 738 739 /* 740 * This function allocates a WQN from a range which is consecutive and aligned 741 * to its size. In case the range is full, then it creates a new range and 742 * allocates WQN from it. The new range will be used for following allocations. 743 */ 744 static int mlx4_ib_alloc_wqn(struct mlx4_ib_ucontext *context, 745 struct mlx4_ib_qp *qp, int range_size, int *wqn) 746 { 747 struct mlx4_ib_dev *dev = to_mdev(context->ibucontext.device); 748 struct mlx4_wqn_range *range; 749 int err = 0; 750 751 mutex_lock(&context->wqn_ranges_mutex); 752 753 range = list_first_entry_or_null(&context->wqn_ranges_list, 754 struct mlx4_wqn_range, list); 755 756 if (!range || (range->refcount == range->size) || range->dirty) { 757 range = kzalloc(sizeof(*range), GFP_KERNEL); 758 if (!range) { 759 err = -ENOMEM; 760 goto out; 761 } 762 763 err = mlx4_qp_reserve_range(dev->dev, range_size, 764 range_size, &range->base_wqn, 0, 765 qp->mqp.usage); 766 if (err) { 767 kfree(range); 768 goto out; 769 } 770 771 range->size = range_size; 772 list_add(&range->list, &context->wqn_ranges_list); 773 } else if (range_size != 1) { 774 /* 775 * Requesting a new range (>1) when last range is still open, is 776 * not valid. 777 */ 778 err = -EINVAL; 779 goto out; 780 } 781 782 qp->wqn_range = range; 783 784 *wqn = range->base_wqn + range->refcount; 785 786 range->refcount++; 787 788 out: 789 mutex_unlock(&context->wqn_ranges_mutex); 790 791 return err; 792 } 793 794 static void mlx4_ib_release_wqn(struct mlx4_ib_ucontext *context, 795 struct mlx4_ib_qp *qp, bool dirty_release) 796 { 797 struct mlx4_ib_dev *dev = to_mdev(context->ibucontext.device); 798 struct mlx4_wqn_range *range; 799 800 mutex_lock(&context->wqn_ranges_mutex); 801 802 range = qp->wqn_range; 803 804 range->refcount--; 805 if (!range->refcount) { 806 mlx4_qp_release_range(dev->dev, range->base_wqn, 807 range->size); 808 list_del(&range->list); 809 kfree(range); 810 } else if (dirty_release) { 811 /* 812 * A range which one of its WQNs is destroyed, won't be able to be 813 * reused for further WQN allocations. 814 * The next created WQ will allocate a new range. 815 */ 816 range->dirty = true; 817 } 818 819 mutex_unlock(&context->wqn_ranges_mutex); 820 } 821 822 static int create_rq(struct ib_pd *pd, struct ib_qp_init_attr *init_attr, 823 struct ib_udata *udata, struct mlx4_ib_qp *qp) 824 { 825 struct mlx4_ib_dev *dev = to_mdev(pd->device); 826 int qpn; 827 int err; 828 struct mlx4_ib_ucontext *context = rdma_udata_to_drv_context( 829 udata, struct mlx4_ib_ucontext, ibucontext); 830 struct mlx4_ib_cq *mcq; 831 unsigned long flags; 832 int range_size; 833 struct mlx4_ib_create_wq wq; 834 size_t copy_len; 835 int shift; 836 int n; 837 838 qp->mlx4_ib_qp_type = MLX4_IB_QPT_RAW_PACKET; 839 840 spin_lock_init(&qp->sq.lock); 841 spin_lock_init(&qp->rq.lock); 842 INIT_LIST_HEAD(&qp->gid_list); 843 INIT_LIST_HEAD(&qp->steering_rules); 844 845 qp->state = IB_QPS_RESET; 846 847 copy_len = min(sizeof(struct mlx4_ib_create_wq), udata->inlen); 848 849 if (ib_copy_from_udata(&wq, udata, copy_len)) { 850 err = -EFAULT; 851 goto err; 852 } 853 854 if (wq.comp_mask || wq.reserved[0] || wq.reserved[1] || 855 wq.reserved[2]) { 856 pr_debug("user command isn't supported\n"); 857 err = -EOPNOTSUPP; 858 goto err; 859 } 860 861 if (wq.log_range_size > ilog2(dev->dev->caps.max_rss_tbl_sz)) { 862 pr_debug("WQN range size must be equal or smaller than %d\n", 863 dev->dev->caps.max_rss_tbl_sz); 864 err = -EOPNOTSUPP; 865 goto err; 866 } 867 range_size = 1 << wq.log_range_size; 868 869 if (init_attr->create_flags & IB_QP_CREATE_SCATTER_FCS) 870 qp->flags |= MLX4_IB_QP_SCATTER_FCS; 871 872 err = set_rq_size(dev, &init_attr->cap, true, true, qp, qp->inl_recv_sz); 873 if (err) 874 goto err; 875 876 qp->sq_no_prefetch = 1; 877 qp->sq.wqe_cnt = 1; 878 qp->sq.wqe_shift = MLX4_IB_MIN_SQ_STRIDE; 879 qp->buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) + 880 (qp->sq.wqe_cnt << qp->sq.wqe_shift); 881 882 qp->umem = ib_umem_get(pd->device, wq.buf_addr, qp->buf_size, 0); 883 if (IS_ERR(qp->umem)) { 884 err = PTR_ERR(qp->umem); 885 goto err; 886 } 887 888 shift = mlx4_ib_umem_calc_optimal_mtt_size(qp->umem, 0, &n); 889 err = mlx4_mtt_init(dev->dev, n, shift, &qp->mtt); 890 891 if (err) 892 goto err_buf; 893 894 err = mlx4_ib_umem_write_mtt(dev, &qp->mtt, qp->umem); 895 if (err) 896 goto err_mtt; 897 898 err = mlx4_ib_db_map_user(udata, wq.db_addr, &qp->db); 899 if (err) 900 goto err_mtt; 901 qp->mqp.usage = MLX4_RES_USAGE_USER_VERBS; 902 903 err = mlx4_ib_alloc_wqn(context, qp, range_size, &qpn); 904 if (err) 905 goto err_wrid; 906 907 err = mlx4_qp_alloc(dev->dev, qpn, &qp->mqp); 908 if (err) 909 goto err_qpn; 910 911 /* 912 * Hardware wants QPN written in big-endian order (after 913 * shifting) for send doorbell. Precompute this value to save 914 * a little bit when posting sends. 915 */ 916 qp->doorbell_qpn = swab32(qp->mqp.qpn << 8); 917 918 qp->mqp.event = mlx4_ib_wq_event; 919 920 spin_lock_irqsave(&dev->reset_flow_resource_lock, flags); 921 mlx4_ib_lock_cqs(to_mcq(init_attr->send_cq), 922 to_mcq(init_attr->recv_cq)); 923 /* Maintain device to QPs access, needed for further handling 924 * via reset flow 925 */ 926 list_add_tail(&qp->qps_list, &dev->qp_list); 927 /* Maintain CQ to QPs access, needed for further handling 928 * via reset flow 929 */ 930 mcq = to_mcq(init_attr->send_cq); 931 list_add_tail(&qp->cq_send_list, &mcq->send_qp_list); 932 mcq = to_mcq(init_attr->recv_cq); 933 list_add_tail(&qp->cq_recv_list, &mcq->recv_qp_list); 934 mlx4_ib_unlock_cqs(to_mcq(init_attr->send_cq), 935 to_mcq(init_attr->recv_cq)); 936 spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags); 937 return 0; 938 939 err_qpn: 940 mlx4_ib_release_wqn(context, qp, 0); 941 err_wrid: 942 mlx4_ib_db_unmap_user(context, &qp->db); 943 944 err_mtt: 945 mlx4_mtt_cleanup(dev->dev, &qp->mtt); 946 err_buf: 947 ib_umem_release(qp->umem); 948 err: 949 return err; 950 } 951 952 static int create_qp_common(struct ib_pd *pd, struct ib_qp_init_attr *init_attr, 953 struct ib_udata *udata, int sqpn, 954 struct mlx4_ib_qp *qp) 955 { 956 struct mlx4_ib_dev *dev = to_mdev(pd->device); 957 int qpn; 958 int err; 959 struct mlx4_ib_ucontext *context = rdma_udata_to_drv_context( 960 udata, struct mlx4_ib_ucontext, ibucontext); 961 enum mlx4_ib_qp_type qp_type = (enum mlx4_ib_qp_type) init_attr->qp_type; 962 struct mlx4_ib_cq *mcq; 963 unsigned long flags; 964 965 /* When tunneling special qps, we use a plain UD qp */ 966 if (sqpn) { 967 if (mlx4_is_mfunc(dev->dev) && 968 (!mlx4_is_master(dev->dev) || 969 !(init_attr->create_flags & MLX4_IB_SRIOV_SQP))) { 970 if (init_attr->qp_type == IB_QPT_GSI) 971 qp_type = MLX4_IB_QPT_PROXY_GSI; 972 else { 973 if (mlx4_is_master(dev->dev) || 974 qp0_enabled_vf(dev->dev, sqpn)) 975 qp_type = MLX4_IB_QPT_PROXY_SMI_OWNER; 976 else 977 qp_type = MLX4_IB_QPT_PROXY_SMI; 978 } 979 } 980 qpn = sqpn; 981 /* add extra sg entry for tunneling */ 982 init_attr->cap.max_recv_sge++; 983 } else if (init_attr->create_flags & MLX4_IB_SRIOV_TUNNEL_QP) { 984 struct mlx4_ib_qp_tunnel_init_attr *tnl_init = 985 container_of(init_attr, 986 struct mlx4_ib_qp_tunnel_init_attr, init_attr); 987 if ((tnl_init->proxy_qp_type != IB_QPT_SMI && 988 tnl_init->proxy_qp_type != IB_QPT_GSI) || 989 !mlx4_is_master(dev->dev)) 990 return -EINVAL; 991 if (tnl_init->proxy_qp_type == IB_QPT_GSI) 992 qp_type = MLX4_IB_QPT_TUN_GSI; 993 else if (tnl_init->slave == mlx4_master_func_num(dev->dev) || 994 mlx4_vf_smi_enabled(dev->dev, tnl_init->slave, 995 tnl_init->port)) 996 qp_type = MLX4_IB_QPT_TUN_SMI_OWNER; 997 else 998 qp_type = MLX4_IB_QPT_TUN_SMI; 999 /* we are definitely in the PPF here, since we are creating 1000 * tunnel QPs. base_tunnel_sqpn is therefore valid. */ 1001 qpn = dev->dev->phys_caps.base_tunnel_sqpn + 8 * tnl_init->slave 1002 + tnl_init->proxy_qp_type * 2 + tnl_init->port - 1; 1003 sqpn = qpn; 1004 } 1005 1006 if (init_attr->qp_type == IB_QPT_SMI || 1007 init_attr->qp_type == IB_QPT_GSI || qp_type == MLX4_IB_QPT_SMI || 1008 qp_type == MLX4_IB_QPT_GSI || 1009 (qp_type & (MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_SMI_OWNER | 1010 MLX4_IB_QPT_PROXY_GSI | MLX4_IB_QPT_TUN_SMI_OWNER))) { 1011 qp->sqp = kzalloc(sizeof(struct mlx4_ib_sqp), GFP_KERNEL); 1012 if (!qp->sqp) 1013 return -ENOMEM; 1014 } 1015 1016 qp->mlx4_ib_qp_type = qp_type; 1017 1018 spin_lock_init(&qp->sq.lock); 1019 spin_lock_init(&qp->rq.lock); 1020 INIT_LIST_HEAD(&qp->gid_list); 1021 INIT_LIST_HEAD(&qp->steering_rules); 1022 1023 qp->state = IB_QPS_RESET; 1024 if (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR) 1025 qp->sq_signal_bits = cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE); 1026 1027 if (udata) { 1028 struct mlx4_ib_create_qp ucmd; 1029 size_t copy_len; 1030 int shift; 1031 int n; 1032 1033 copy_len = sizeof(struct mlx4_ib_create_qp); 1034 1035 if (ib_copy_from_udata(&ucmd, udata, copy_len)) { 1036 err = -EFAULT; 1037 goto err; 1038 } 1039 1040 qp->inl_recv_sz = ucmd.inl_recv_sz; 1041 1042 if (init_attr->create_flags & IB_QP_CREATE_SCATTER_FCS) { 1043 if (!(dev->dev->caps.flags & 1044 MLX4_DEV_CAP_FLAG_FCS_KEEP)) { 1045 pr_debug("scatter FCS is unsupported\n"); 1046 err = -EOPNOTSUPP; 1047 goto err; 1048 } 1049 1050 qp->flags |= MLX4_IB_QP_SCATTER_FCS; 1051 } 1052 1053 err = set_rq_size(dev, &init_attr->cap, udata, 1054 qp_has_rq(init_attr), qp, qp->inl_recv_sz); 1055 if (err) 1056 goto err; 1057 1058 qp->sq_no_prefetch = ucmd.sq_no_prefetch; 1059 1060 err = set_user_sq_size(dev, qp, &ucmd); 1061 if (err) 1062 goto err; 1063 1064 qp->umem = 1065 ib_umem_get(pd->device, ucmd.buf_addr, qp->buf_size, 0); 1066 if (IS_ERR(qp->umem)) { 1067 err = PTR_ERR(qp->umem); 1068 goto err; 1069 } 1070 1071 shift = mlx4_ib_umem_calc_optimal_mtt_size(qp->umem, 0, &n); 1072 err = mlx4_mtt_init(dev->dev, n, shift, &qp->mtt); 1073 1074 if (err) 1075 goto err_buf; 1076 1077 err = mlx4_ib_umem_write_mtt(dev, &qp->mtt, qp->umem); 1078 if (err) 1079 goto err_mtt; 1080 1081 if (qp_has_rq(init_attr)) { 1082 err = mlx4_ib_db_map_user(udata, ucmd.db_addr, &qp->db); 1083 if (err) 1084 goto err_mtt; 1085 } 1086 qp->mqp.usage = MLX4_RES_USAGE_USER_VERBS; 1087 } else { 1088 err = set_rq_size(dev, &init_attr->cap, udata, 1089 qp_has_rq(init_attr), qp, 0); 1090 if (err) 1091 goto err; 1092 1093 qp->sq_no_prefetch = 0; 1094 1095 if (init_attr->create_flags & IB_QP_CREATE_IPOIB_UD_LSO) 1096 qp->flags |= MLX4_IB_QP_LSO; 1097 1098 if (init_attr->create_flags & IB_QP_CREATE_NETIF_QP) { 1099 if (dev->steering_support == 1100 MLX4_STEERING_MODE_DEVICE_MANAGED) 1101 qp->flags |= MLX4_IB_QP_NETIF; 1102 else 1103 goto err; 1104 } 1105 1106 err = set_kernel_sq_size(dev, &init_attr->cap, qp_type, qp); 1107 if (err) 1108 goto err; 1109 1110 if (qp_has_rq(init_attr)) { 1111 err = mlx4_db_alloc(dev->dev, &qp->db, 0); 1112 if (err) 1113 goto err; 1114 1115 *qp->db.db = 0; 1116 } 1117 1118 if (mlx4_buf_alloc(dev->dev, qp->buf_size, PAGE_SIZE * 2, 1119 &qp->buf)) { 1120 err = -ENOMEM; 1121 goto err_db; 1122 } 1123 1124 err = mlx4_mtt_init(dev->dev, qp->buf.npages, qp->buf.page_shift, 1125 &qp->mtt); 1126 if (err) 1127 goto err_buf; 1128 1129 err = mlx4_buf_write_mtt(dev->dev, &qp->mtt, &qp->buf); 1130 if (err) 1131 goto err_mtt; 1132 1133 qp->sq.wrid = kvmalloc_array(qp->sq.wqe_cnt, 1134 sizeof(u64), GFP_KERNEL); 1135 qp->rq.wrid = kvmalloc_array(qp->rq.wqe_cnt, 1136 sizeof(u64), GFP_KERNEL); 1137 if (!qp->sq.wrid || !qp->rq.wrid) { 1138 err = -ENOMEM; 1139 goto err_wrid; 1140 } 1141 qp->mqp.usage = MLX4_RES_USAGE_DRIVER; 1142 } 1143 1144 if (sqpn) { 1145 if (qp->mlx4_ib_qp_type & (MLX4_IB_QPT_PROXY_SMI_OWNER | 1146 MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI)) { 1147 if (alloc_proxy_bufs(pd->device, qp)) { 1148 err = -ENOMEM; 1149 goto err_wrid; 1150 } 1151 } 1152 } else { 1153 /* Raw packet QPNs may not have bits 6,7 set in their qp_num; 1154 * otherwise, the WQE BlueFlame setup flow wrongly causes 1155 * VLAN insertion. */ 1156 if (init_attr->qp_type == IB_QPT_RAW_PACKET) 1157 err = mlx4_qp_reserve_range(dev->dev, 1, 1, &qpn, 1158 (init_attr->cap.max_send_wr ? 1159 MLX4_RESERVE_ETH_BF_QP : 0) | 1160 (init_attr->cap.max_recv_wr ? 1161 MLX4_RESERVE_A0_QP : 0), 1162 qp->mqp.usage); 1163 else 1164 if (qp->flags & MLX4_IB_QP_NETIF) 1165 err = mlx4_ib_steer_qp_alloc(dev, 1, &qpn); 1166 else 1167 err = mlx4_qp_reserve_range(dev->dev, 1, 1, 1168 &qpn, 0, qp->mqp.usage); 1169 if (err) 1170 goto err_proxy; 1171 } 1172 1173 if (init_attr->create_flags & IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK) 1174 qp->flags |= MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK; 1175 1176 err = mlx4_qp_alloc(dev->dev, qpn, &qp->mqp); 1177 if (err) 1178 goto err_qpn; 1179 1180 if (init_attr->qp_type == IB_QPT_XRC_TGT) 1181 qp->mqp.qpn |= (1 << 23); 1182 1183 /* 1184 * Hardware wants QPN written in big-endian order (after 1185 * shifting) for send doorbell. Precompute this value to save 1186 * a little bit when posting sends. 1187 */ 1188 qp->doorbell_qpn = swab32(qp->mqp.qpn << 8); 1189 1190 qp->mqp.event = mlx4_ib_qp_event; 1191 1192 spin_lock_irqsave(&dev->reset_flow_resource_lock, flags); 1193 mlx4_ib_lock_cqs(to_mcq(init_attr->send_cq), 1194 to_mcq(init_attr->recv_cq)); 1195 /* Maintain device to QPs access, needed for further handling 1196 * via reset flow 1197 */ 1198 list_add_tail(&qp->qps_list, &dev->qp_list); 1199 /* Maintain CQ to QPs access, needed for further handling 1200 * via reset flow 1201 */ 1202 mcq = to_mcq(init_attr->send_cq); 1203 list_add_tail(&qp->cq_send_list, &mcq->send_qp_list); 1204 mcq = to_mcq(init_attr->recv_cq); 1205 list_add_tail(&qp->cq_recv_list, &mcq->recv_qp_list); 1206 mlx4_ib_unlock_cqs(to_mcq(init_attr->send_cq), 1207 to_mcq(init_attr->recv_cq)); 1208 spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags); 1209 return 0; 1210 1211 err_qpn: 1212 if (!sqpn) { 1213 if (qp->flags & MLX4_IB_QP_NETIF) 1214 mlx4_ib_steer_qp_free(dev, qpn, 1); 1215 else 1216 mlx4_qp_release_range(dev->dev, qpn, 1); 1217 } 1218 err_proxy: 1219 if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI) 1220 free_proxy_bufs(pd->device, qp); 1221 err_wrid: 1222 if (udata) { 1223 if (qp_has_rq(init_attr)) 1224 mlx4_ib_db_unmap_user(context, &qp->db); 1225 } else { 1226 kvfree(qp->sq.wrid); 1227 kvfree(qp->rq.wrid); 1228 } 1229 1230 err_mtt: 1231 mlx4_mtt_cleanup(dev->dev, &qp->mtt); 1232 1233 err_buf: 1234 if (!qp->umem) 1235 mlx4_buf_free(dev->dev, qp->buf_size, &qp->buf); 1236 ib_umem_release(qp->umem); 1237 1238 err_db: 1239 if (!udata && qp_has_rq(init_attr)) 1240 mlx4_db_free(dev->dev, &qp->db); 1241 1242 err: 1243 kfree(qp->sqp); 1244 return err; 1245 } 1246 1247 static enum mlx4_qp_state to_mlx4_state(enum ib_qp_state state) 1248 { 1249 switch (state) { 1250 case IB_QPS_RESET: return MLX4_QP_STATE_RST; 1251 case IB_QPS_INIT: return MLX4_QP_STATE_INIT; 1252 case IB_QPS_RTR: return MLX4_QP_STATE_RTR; 1253 case IB_QPS_RTS: return MLX4_QP_STATE_RTS; 1254 case IB_QPS_SQD: return MLX4_QP_STATE_SQD; 1255 case IB_QPS_SQE: return MLX4_QP_STATE_SQER; 1256 case IB_QPS_ERR: return MLX4_QP_STATE_ERR; 1257 default: return -1; 1258 } 1259 } 1260 1261 static void mlx4_ib_lock_cqs(struct mlx4_ib_cq *send_cq, struct mlx4_ib_cq *recv_cq) 1262 __acquires(&send_cq->lock) __acquires(&recv_cq->lock) 1263 { 1264 if (send_cq == recv_cq) { 1265 spin_lock(&send_cq->lock); 1266 __acquire(&recv_cq->lock); 1267 } else if (send_cq->mcq.cqn < recv_cq->mcq.cqn) { 1268 spin_lock(&send_cq->lock); 1269 spin_lock_nested(&recv_cq->lock, SINGLE_DEPTH_NESTING); 1270 } else { 1271 spin_lock(&recv_cq->lock); 1272 spin_lock_nested(&send_cq->lock, SINGLE_DEPTH_NESTING); 1273 } 1274 } 1275 1276 static void mlx4_ib_unlock_cqs(struct mlx4_ib_cq *send_cq, struct mlx4_ib_cq *recv_cq) 1277 __releases(&send_cq->lock) __releases(&recv_cq->lock) 1278 { 1279 if (send_cq == recv_cq) { 1280 __release(&recv_cq->lock); 1281 spin_unlock(&send_cq->lock); 1282 } else if (send_cq->mcq.cqn < recv_cq->mcq.cqn) { 1283 spin_unlock(&recv_cq->lock); 1284 spin_unlock(&send_cq->lock); 1285 } else { 1286 spin_unlock(&send_cq->lock); 1287 spin_unlock(&recv_cq->lock); 1288 } 1289 } 1290 1291 static void del_gid_entries(struct mlx4_ib_qp *qp) 1292 { 1293 struct mlx4_ib_gid_entry *ge, *tmp; 1294 1295 list_for_each_entry_safe(ge, tmp, &qp->gid_list, list) { 1296 list_del(&ge->list); 1297 kfree(ge); 1298 } 1299 } 1300 1301 static struct mlx4_ib_pd *get_pd(struct mlx4_ib_qp *qp) 1302 { 1303 if (qp->ibqp.qp_type == IB_QPT_XRC_TGT) 1304 return to_mpd(to_mxrcd(qp->ibqp.xrcd)->pd); 1305 else 1306 return to_mpd(qp->ibqp.pd); 1307 } 1308 1309 static void get_cqs(struct mlx4_ib_qp *qp, enum mlx4_ib_source_type src, 1310 struct mlx4_ib_cq **send_cq, struct mlx4_ib_cq **recv_cq) 1311 { 1312 switch (qp->ibqp.qp_type) { 1313 case IB_QPT_XRC_TGT: 1314 *send_cq = to_mcq(to_mxrcd(qp->ibqp.xrcd)->cq); 1315 *recv_cq = *send_cq; 1316 break; 1317 case IB_QPT_XRC_INI: 1318 *send_cq = to_mcq(qp->ibqp.send_cq); 1319 *recv_cq = *send_cq; 1320 break; 1321 default: 1322 *recv_cq = (src == MLX4_IB_QP_SRC) ? to_mcq(qp->ibqp.recv_cq) : 1323 to_mcq(qp->ibwq.cq); 1324 *send_cq = (src == MLX4_IB_QP_SRC) ? to_mcq(qp->ibqp.send_cq) : 1325 *recv_cq; 1326 break; 1327 } 1328 } 1329 1330 static void destroy_qp_rss(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp) 1331 { 1332 if (qp->state != IB_QPS_RESET) { 1333 int i; 1334 1335 for (i = 0; i < (1 << qp->ibqp.rwq_ind_tbl->log_ind_tbl_size); 1336 i++) { 1337 struct ib_wq *ibwq = qp->ibqp.rwq_ind_tbl->ind_tbl[i]; 1338 struct mlx4_ib_qp *wq = to_mqp((struct ib_qp *)ibwq); 1339 1340 mutex_lock(&wq->mutex); 1341 1342 wq->rss_usecnt--; 1343 1344 mutex_unlock(&wq->mutex); 1345 } 1346 1347 if (mlx4_qp_modify(dev->dev, NULL, to_mlx4_state(qp->state), 1348 MLX4_QP_STATE_RST, NULL, 0, 0, &qp->mqp)) 1349 pr_warn("modify QP %06x to RESET failed.\n", 1350 qp->mqp.qpn); 1351 } 1352 1353 mlx4_qp_remove(dev->dev, &qp->mqp); 1354 mlx4_qp_free(dev->dev, &qp->mqp); 1355 mlx4_qp_release_range(dev->dev, qp->mqp.qpn, 1); 1356 del_gid_entries(qp); 1357 } 1358 1359 static void destroy_qp_common(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp, 1360 enum mlx4_ib_source_type src, 1361 struct ib_udata *udata) 1362 { 1363 struct mlx4_ib_cq *send_cq, *recv_cq; 1364 unsigned long flags; 1365 1366 if (qp->state != IB_QPS_RESET) { 1367 if (mlx4_qp_modify(dev->dev, NULL, to_mlx4_state(qp->state), 1368 MLX4_QP_STATE_RST, NULL, 0, 0, &qp->mqp)) 1369 pr_warn("modify QP %06x to RESET failed.\n", 1370 qp->mqp.qpn); 1371 if (qp->pri.smac || (!qp->pri.smac && qp->pri.smac_port)) { 1372 mlx4_unregister_mac(dev->dev, qp->pri.smac_port, qp->pri.smac); 1373 qp->pri.smac = 0; 1374 qp->pri.smac_port = 0; 1375 } 1376 if (qp->alt.smac) { 1377 mlx4_unregister_mac(dev->dev, qp->alt.smac_port, qp->alt.smac); 1378 qp->alt.smac = 0; 1379 } 1380 if (qp->pri.vid < 0x1000) { 1381 mlx4_unregister_vlan(dev->dev, qp->pri.vlan_port, qp->pri.vid); 1382 qp->pri.vid = 0xFFFF; 1383 qp->pri.candidate_vid = 0xFFFF; 1384 qp->pri.update_vid = 0; 1385 } 1386 if (qp->alt.vid < 0x1000) { 1387 mlx4_unregister_vlan(dev->dev, qp->alt.vlan_port, qp->alt.vid); 1388 qp->alt.vid = 0xFFFF; 1389 qp->alt.candidate_vid = 0xFFFF; 1390 qp->alt.update_vid = 0; 1391 } 1392 } 1393 1394 get_cqs(qp, src, &send_cq, &recv_cq); 1395 1396 spin_lock_irqsave(&dev->reset_flow_resource_lock, flags); 1397 mlx4_ib_lock_cqs(send_cq, recv_cq); 1398 1399 /* del from lists under both locks above to protect reset flow paths */ 1400 list_del(&qp->qps_list); 1401 list_del(&qp->cq_send_list); 1402 list_del(&qp->cq_recv_list); 1403 if (!udata) { 1404 __mlx4_ib_cq_clean(recv_cq, qp->mqp.qpn, 1405 qp->ibqp.srq ? to_msrq(qp->ibqp.srq): NULL); 1406 if (send_cq != recv_cq) 1407 __mlx4_ib_cq_clean(send_cq, qp->mqp.qpn, NULL); 1408 } 1409 1410 mlx4_qp_remove(dev->dev, &qp->mqp); 1411 1412 mlx4_ib_unlock_cqs(send_cq, recv_cq); 1413 spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags); 1414 1415 mlx4_qp_free(dev->dev, &qp->mqp); 1416 1417 if (!is_sqp(dev, qp) && !is_tunnel_qp(dev, qp)) { 1418 if (qp->flags & MLX4_IB_QP_NETIF) 1419 mlx4_ib_steer_qp_free(dev, qp->mqp.qpn, 1); 1420 else if (src == MLX4_IB_RWQ_SRC) 1421 mlx4_ib_release_wqn( 1422 rdma_udata_to_drv_context( 1423 udata, 1424 struct mlx4_ib_ucontext, 1425 ibucontext), 1426 qp, 1); 1427 else 1428 mlx4_qp_release_range(dev->dev, qp->mqp.qpn, 1); 1429 } 1430 1431 mlx4_mtt_cleanup(dev->dev, &qp->mtt); 1432 1433 if (udata) { 1434 if (qp->rq.wqe_cnt) { 1435 struct mlx4_ib_ucontext *mcontext = 1436 rdma_udata_to_drv_context( 1437 udata, 1438 struct mlx4_ib_ucontext, 1439 ibucontext); 1440 1441 mlx4_ib_db_unmap_user(mcontext, &qp->db); 1442 } 1443 } else { 1444 kvfree(qp->sq.wrid); 1445 kvfree(qp->rq.wrid); 1446 if (qp->mlx4_ib_qp_type & (MLX4_IB_QPT_PROXY_SMI_OWNER | 1447 MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI)) 1448 free_proxy_bufs(&dev->ib_dev, qp); 1449 mlx4_buf_free(dev->dev, qp->buf_size, &qp->buf); 1450 if (qp->rq.wqe_cnt) 1451 mlx4_db_free(dev->dev, &qp->db); 1452 } 1453 ib_umem_release(qp->umem); 1454 1455 del_gid_entries(qp); 1456 } 1457 1458 static u32 get_sqp_num(struct mlx4_ib_dev *dev, struct ib_qp_init_attr *attr) 1459 { 1460 /* Native or PPF */ 1461 if (!mlx4_is_mfunc(dev->dev) || 1462 (mlx4_is_master(dev->dev) && 1463 attr->create_flags & MLX4_IB_SRIOV_SQP)) { 1464 return dev->dev->phys_caps.base_sqpn + 1465 (attr->qp_type == IB_QPT_SMI ? 0 : 2) + 1466 attr->port_num - 1; 1467 } 1468 /* PF or VF -- creating proxies */ 1469 if (attr->qp_type == IB_QPT_SMI) 1470 return dev->dev->caps.spec_qps[attr->port_num - 1].qp0_proxy; 1471 else 1472 return dev->dev->caps.spec_qps[attr->port_num - 1].qp1_proxy; 1473 } 1474 1475 static int _mlx4_ib_create_qp(struct ib_pd *pd, struct mlx4_ib_qp *qp, 1476 struct ib_qp_init_attr *init_attr, 1477 struct ib_udata *udata) 1478 { 1479 int err; 1480 int sup_u_create_flags = MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK; 1481 u16 xrcdn = 0; 1482 1483 if (init_attr->rwq_ind_tbl) 1484 return _mlx4_ib_create_qp_rss(pd, qp, init_attr, udata); 1485 1486 /* 1487 * We only support LSO, vendor flag1, and multicast loopback blocking, 1488 * and only for kernel UD QPs. 1489 */ 1490 if (init_attr->create_flags & ~(MLX4_IB_QP_LSO | 1491 MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK | 1492 MLX4_IB_SRIOV_TUNNEL_QP | 1493 MLX4_IB_SRIOV_SQP | 1494 MLX4_IB_QP_NETIF | 1495 MLX4_IB_QP_CREATE_ROCE_V2_GSI)) 1496 return -EOPNOTSUPP; 1497 1498 if (init_attr->create_flags & IB_QP_CREATE_NETIF_QP) { 1499 if (init_attr->qp_type != IB_QPT_UD) 1500 return -EINVAL; 1501 } 1502 1503 if (init_attr->create_flags) { 1504 if (udata && init_attr->create_flags & ~(sup_u_create_flags)) 1505 return -EINVAL; 1506 1507 if ((init_attr->create_flags & ~(MLX4_IB_SRIOV_SQP | 1508 MLX4_IB_QP_CREATE_ROCE_V2_GSI | 1509 MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK) && 1510 init_attr->qp_type != IB_QPT_UD) || 1511 (init_attr->create_flags & MLX4_IB_SRIOV_SQP && 1512 init_attr->qp_type > IB_QPT_GSI) || 1513 (init_attr->create_flags & MLX4_IB_QP_CREATE_ROCE_V2_GSI && 1514 init_attr->qp_type != IB_QPT_GSI)) 1515 return -EINVAL; 1516 } 1517 1518 switch (init_attr->qp_type) { 1519 case IB_QPT_XRC_TGT: 1520 pd = to_mxrcd(init_attr->xrcd)->pd; 1521 xrcdn = to_mxrcd(init_attr->xrcd)->xrcdn; 1522 init_attr->send_cq = to_mxrcd(init_attr->xrcd)->cq; 1523 fallthrough; 1524 case IB_QPT_XRC_INI: 1525 if (!(to_mdev(pd->device)->dev->caps.flags & MLX4_DEV_CAP_FLAG_XRC)) 1526 return -ENOSYS; 1527 init_attr->recv_cq = init_attr->send_cq; 1528 fallthrough; 1529 case IB_QPT_RC: 1530 case IB_QPT_UC: 1531 case IB_QPT_RAW_PACKET: 1532 case IB_QPT_UD: 1533 qp->pri.vid = 0xFFFF; 1534 qp->alt.vid = 0xFFFF; 1535 err = create_qp_common(pd, init_attr, udata, 0, qp); 1536 if (err) 1537 return err; 1538 1539 qp->ibqp.qp_num = qp->mqp.qpn; 1540 qp->xrcdn = xrcdn; 1541 break; 1542 case IB_QPT_SMI: 1543 case IB_QPT_GSI: 1544 { 1545 int sqpn; 1546 1547 if (init_attr->create_flags & MLX4_IB_QP_CREATE_ROCE_V2_GSI) { 1548 int res = mlx4_qp_reserve_range(to_mdev(pd->device)->dev, 1549 1, 1, &sqpn, 0, 1550 MLX4_RES_USAGE_DRIVER); 1551 1552 if (res) 1553 return res; 1554 } else { 1555 sqpn = get_sqp_num(to_mdev(pd->device), init_attr); 1556 } 1557 1558 qp->pri.vid = 0xFFFF; 1559 qp->alt.vid = 0xFFFF; 1560 err = create_qp_common(pd, init_attr, udata, sqpn, qp); 1561 if (err) 1562 return err; 1563 1564 if (init_attr->create_flags & 1565 (MLX4_IB_SRIOV_SQP | MLX4_IB_SRIOV_TUNNEL_QP)) 1566 /* Internal QP created with ib_create_qp */ 1567 rdma_restrack_no_track(&qp->ibqp.res); 1568 1569 qp->port = init_attr->port_num; 1570 qp->ibqp.qp_num = init_attr->qp_type == IB_QPT_SMI ? 0 : 1571 init_attr->create_flags & MLX4_IB_QP_CREATE_ROCE_V2_GSI ? sqpn : 1; 1572 break; 1573 } 1574 default: 1575 /* Don't support raw QPs */ 1576 return -EOPNOTSUPP; 1577 } 1578 return 0; 1579 } 1580 1581 int mlx4_ib_create_qp(struct ib_qp *ibqp, struct ib_qp_init_attr *init_attr, 1582 struct ib_udata *udata) 1583 { 1584 struct ib_device *device = ibqp->device; 1585 struct mlx4_ib_dev *dev = to_mdev(device); 1586 struct mlx4_ib_qp *qp = to_mqp(ibqp); 1587 struct ib_pd *pd = ibqp->pd; 1588 int ret; 1589 1590 mutex_init(&qp->mutex); 1591 ret = _mlx4_ib_create_qp(pd, qp, init_attr, udata); 1592 if (ret) 1593 return ret; 1594 1595 if (init_attr->qp_type == IB_QPT_GSI && 1596 !(init_attr->create_flags & MLX4_IB_QP_CREATE_ROCE_V2_GSI)) { 1597 struct mlx4_ib_sqp *sqp = qp->sqp; 1598 int is_eth = rdma_cap_eth_ah(&dev->ib_dev, init_attr->port_num); 1599 1600 if (is_eth && 1601 dev->dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_ROCE_V1_V2) { 1602 init_attr->create_flags |= MLX4_IB_QP_CREATE_ROCE_V2_GSI; 1603 sqp->roce_v2_gsi = ib_create_qp(pd, init_attr); 1604 1605 if (IS_ERR(sqp->roce_v2_gsi)) { 1606 pr_err("Failed to create GSI QP for RoCEv2 (%ld)\n", PTR_ERR(sqp->roce_v2_gsi)); 1607 sqp->roce_v2_gsi = NULL; 1608 } else { 1609 to_mqp(sqp->roce_v2_gsi)->flags |= 1610 MLX4_IB_ROCE_V2_GSI_QP; 1611 } 1612 1613 init_attr->create_flags &= ~MLX4_IB_QP_CREATE_ROCE_V2_GSI; 1614 } 1615 } 1616 return 0; 1617 } 1618 1619 static int _mlx4_ib_destroy_qp(struct ib_qp *qp, struct ib_udata *udata) 1620 { 1621 struct mlx4_ib_dev *dev = to_mdev(qp->device); 1622 struct mlx4_ib_qp *mqp = to_mqp(qp); 1623 1624 if (is_qp0(dev, mqp)) 1625 mlx4_CLOSE_PORT(dev->dev, mqp->port); 1626 1627 if (mqp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI && 1628 dev->qp1_proxy[mqp->port - 1] == mqp) { 1629 mutex_lock(&dev->qp1_proxy_lock[mqp->port - 1]); 1630 dev->qp1_proxy[mqp->port - 1] = NULL; 1631 mutex_unlock(&dev->qp1_proxy_lock[mqp->port - 1]); 1632 } 1633 1634 if (mqp->counter_index) 1635 mlx4_ib_free_qp_counter(dev, mqp); 1636 1637 if (qp->rwq_ind_tbl) { 1638 destroy_qp_rss(dev, mqp); 1639 } else { 1640 destroy_qp_common(dev, mqp, MLX4_IB_QP_SRC, udata); 1641 } 1642 1643 kfree(mqp->sqp); 1644 return 0; 1645 } 1646 1647 int mlx4_ib_destroy_qp(struct ib_qp *qp, struct ib_udata *udata) 1648 { 1649 struct mlx4_ib_qp *mqp = to_mqp(qp); 1650 1651 if (mqp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI) { 1652 struct mlx4_ib_sqp *sqp = mqp->sqp; 1653 1654 if (sqp->roce_v2_gsi) 1655 ib_destroy_qp(sqp->roce_v2_gsi); 1656 } 1657 1658 return _mlx4_ib_destroy_qp(qp, udata); 1659 } 1660 1661 static int to_mlx4_st(struct mlx4_ib_dev *dev, enum mlx4_ib_qp_type type) 1662 { 1663 switch (type) { 1664 case MLX4_IB_QPT_RC: return MLX4_QP_ST_RC; 1665 case MLX4_IB_QPT_UC: return MLX4_QP_ST_UC; 1666 case MLX4_IB_QPT_UD: return MLX4_QP_ST_UD; 1667 case MLX4_IB_QPT_XRC_INI: 1668 case MLX4_IB_QPT_XRC_TGT: return MLX4_QP_ST_XRC; 1669 case MLX4_IB_QPT_SMI: 1670 case MLX4_IB_QPT_GSI: 1671 case MLX4_IB_QPT_RAW_PACKET: return MLX4_QP_ST_MLX; 1672 1673 case MLX4_IB_QPT_PROXY_SMI_OWNER: 1674 case MLX4_IB_QPT_TUN_SMI_OWNER: return (mlx4_is_mfunc(dev->dev) ? 1675 MLX4_QP_ST_MLX : -1); 1676 case MLX4_IB_QPT_PROXY_SMI: 1677 case MLX4_IB_QPT_TUN_SMI: 1678 case MLX4_IB_QPT_PROXY_GSI: 1679 case MLX4_IB_QPT_TUN_GSI: return (mlx4_is_mfunc(dev->dev) ? 1680 MLX4_QP_ST_UD : -1); 1681 default: return -1; 1682 } 1683 } 1684 1685 static __be32 to_mlx4_access_flags(struct mlx4_ib_qp *qp, const struct ib_qp_attr *attr, 1686 int attr_mask) 1687 { 1688 u8 dest_rd_atomic; 1689 u32 access_flags; 1690 u32 hw_access_flags = 0; 1691 1692 if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) 1693 dest_rd_atomic = attr->max_dest_rd_atomic; 1694 else 1695 dest_rd_atomic = qp->resp_depth; 1696 1697 if (attr_mask & IB_QP_ACCESS_FLAGS) 1698 access_flags = attr->qp_access_flags; 1699 else 1700 access_flags = qp->atomic_rd_en; 1701 1702 if (!dest_rd_atomic) 1703 access_flags &= IB_ACCESS_REMOTE_WRITE; 1704 1705 if (access_flags & IB_ACCESS_REMOTE_READ) 1706 hw_access_flags |= MLX4_QP_BIT_RRE; 1707 if (access_flags & IB_ACCESS_REMOTE_ATOMIC) 1708 hw_access_flags |= MLX4_QP_BIT_RAE; 1709 if (access_flags & IB_ACCESS_REMOTE_WRITE) 1710 hw_access_flags |= MLX4_QP_BIT_RWE; 1711 1712 return cpu_to_be32(hw_access_flags); 1713 } 1714 1715 static void store_sqp_attrs(struct mlx4_ib_sqp *sqp, const struct ib_qp_attr *attr, 1716 int attr_mask) 1717 { 1718 if (attr_mask & IB_QP_PKEY_INDEX) 1719 sqp->pkey_index = attr->pkey_index; 1720 if (attr_mask & IB_QP_QKEY) 1721 sqp->qkey = attr->qkey; 1722 if (attr_mask & IB_QP_SQ_PSN) 1723 sqp->send_psn = attr->sq_psn; 1724 } 1725 1726 static void mlx4_set_sched(struct mlx4_qp_path *path, u8 port) 1727 { 1728 path->sched_queue = (path->sched_queue & 0xbf) | ((port - 1) << 6); 1729 } 1730 1731 static int _mlx4_set_path(struct mlx4_ib_dev *dev, 1732 const struct rdma_ah_attr *ah, 1733 u64 smac, u16 vlan_tag, struct mlx4_qp_path *path, 1734 struct mlx4_roce_smac_vlan_info *smac_info, u8 port) 1735 { 1736 int vidx; 1737 int smac_index; 1738 int err; 1739 1740 path->grh_mylmc = rdma_ah_get_path_bits(ah) & 0x7f; 1741 path->rlid = cpu_to_be16(rdma_ah_get_dlid(ah)); 1742 if (rdma_ah_get_static_rate(ah)) { 1743 path->static_rate = rdma_ah_get_static_rate(ah) + 1744 MLX4_STAT_RATE_OFFSET; 1745 while (path->static_rate > IB_RATE_2_5_GBPS + MLX4_STAT_RATE_OFFSET && 1746 !(1 << path->static_rate & dev->dev->caps.stat_rate_support)) 1747 --path->static_rate; 1748 } else 1749 path->static_rate = 0; 1750 1751 if (rdma_ah_get_ah_flags(ah) & IB_AH_GRH) { 1752 const struct ib_global_route *grh = rdma_ah_read_grh(ah); 1753 int real_sgid_index = 1754 mlx4_ib_gid_index_to_real_index(dev, grh->sgid_attr); 1755 1756 if (real_sgid_index < 0) 1757 return real_sgid_index; 1758 if (real_sgid_index >= dev->dev->caps.gid_table_len[port]) { 1759 pr_err("sgid_index (%u) too large. max is %d\n", 1760 real_sgid_index, dev->dev->caps.gid_table_len[port] - 1); 1761 return -1; 1762 } 1763 1764 path->grh_mylmc |= 1 << 7; 1765 path->mgid_index = real_sgid_index; 1766 path->hop_limit = grh->hop_limit; 1767 path->tclass_flowlabel = 1768 cpu_to_be32((grh->traffic_class << 20) | 1769 (grh->flow_label)); 1770 memcpy(path->rgid, grh->dgid.raw, 16); 1771 } 1772 1773 if (ah->type == RDMA_AH_ATTR_TYPE_ROCE) { 1774 if (!(rdma_ah_get_ah_flags(ah) & IB_AH_GRH)) 1775 return -1; 1776 1777 path->sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE | 1778 ((port - 1) << 6) | ((rdma_ah_get_sl(ah) & 7) << 3); 1779 1780 path->feup |= MLX4_FEUP_FORCE_ETH_UP; 1781 if (vlan_tag < 0x1000) { 1782 if (smac_info->vid < 0x1000) { 1783 /* both valid vlan ids */ 1784 if (smac_info->vid != vlan_tag) { 1785 /* different VIDs. unreg old and reg new */ 1786 err = mlx4_register_vlan(dev->dev, port, vlan_tag, &vidx); 1787 if (err) 1788 return err; 1789 smac_info->candidate_vid = vlan_tag; 1790 smac_info->candidate_vlan_index = vidx; 1791 smac_info->candidate_vlan_port = port; 1792 smac_info->update_vid = 1; 1793 path->vlan_index = vidx; 1794 } else { 1795 path->vlan_index = smac_info->vlan_index; 1796 } 1797 } else { 1798 /* no current vlan tag in qp */ 1799 err = mlx4_register_vlan(dev->dev, port, vlan_tag, &vidx); 1800 if (err) 1801 return err; 1802 smac_info->candidate_vid = vlan_tag; 1803 smac_info->candidate_vlan_index = vidx; 1804 smac_info->candidate_vlan_port = port; 1805 smac_info->update_vid = 1; 1806 path->vlan_index = vidx; 1807 } 1808 path->feup |= MLX4_FVL_FORCE_ETH_VLAN; 1809 path->fl = 1 << 6; 1810 } else { 1811 /* have current vlan tag. unregister it at modify-qp success */ 1812 if (smac_info->vid < 0x1000) { 1813 smac_info->candidate_vid = 0xFFFF; 1814 smac_info->update_vid = 1; 1815 } 1816 } 1817 1818 /* get smac_index for RoCE use. 1819 * If no smac was yet assigned, register one. 1820 * If one was already assigned, but the new mac differs, 1821 * unregister the old one and register the new one. 1822 */ 1823 if ((!smac_info->smac && !smac_info->smac_port) || 1824 smac_info->smac != smac) { 1825 /* register candidate now, unreg if needed, after success */ 1826 smac_index = mlx4_register_mac(dev->dev, port, smac); 1827 if (smac_index >= 0) { 1828 smac_info->candidate_smac_index = smac_index; 1829 smac_info->candidate_smac = smac; 1830 smac_info->candidate_smac_port = port; 1831 } else { 1832 return -EINVAL; 1833 } 1834 } else { 1835 smac_index = smac_info->smac_index; 1836 } 1837 memcpy(path->dmac, ah->roce.dmac, 6); 1838 path->ackto = MLX4_IB_LINK_TYPE_ETH; 1839 /* put MAC table smac index for IBoE */ 1840 path->grh_mylmc = (u8) (smac_index) | 0x80; 1841 } else { 1842 path->sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE | 1843 ((port - 1) << 6) | ((rdma_ah_get_sl(ah) & 0xf) << 2); 1844 } 1845 1846 return 0; 1847 } 1848 1849 static int mlx4_set_path(struct mlx4_ib_dev *dev, const struct ib_qp_attr *qp, 1850 enum ib_qp_attr_mask qp_attr_mask, 1851 struct mlx4_ib_qp *mqp, 1852 struct mlx4_qp_path *path, u8 port, 1853 u16 vlan_id, u8 *smac) 1854 { 1855 return _mlx4_set_path(dev, &qp->ah_attr, 1856 mlx4_mac_to_u64(smac), 1857 vlan_id, 1858 path, &mqp->pri, port); 1859 } 1860 1861 static int mlx4_set_alt_path(struct mlx4_ib_dev *dev, 1862 const struct ib_qp_attr *qp, 1863 enum ib_qp_attr_mask qp_attr_mask, 1864 struct mlx4_ib_qp *mqp, 1865 struct mlx4_qp_path *path, u8 port) 1866 { 1867 return _mlx4_set_path(dev, &qp->alt_ah_attr, 1868 0, 1869 0xffff, 1870 path, &mqp->alt, port); 1871 } 1872 1873 static void update_mcg_macs(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp) 1874 { 1875 struct mlx4_ib_gid_entry *ge, *tmp; 1876 1877 list_for_each_entry_safe(ge, tmp, &qp->gid_list, list) { 1878 if (!ge->added && mlx4_ib_add_mc(dev, qp, &ge->gid)) { 1879 ge->added = 1; 1880 ge->port = qp->port; 1881 } 1882 } 1883 } 1884 1885 static int handle_eth_ud_smac_index(struct mlx4_ib_dev *dev, 1886 struct mlx4_ib_qp *qp, 1887 struct mlx4_qp_context *context) 1888 { 1889 u64 u64_mac; 1890 int smac_index; 1891 1892 u64_mac = atomic64_read(&dev->iboe.mac[qp->port - 1]); 1893 1894 context->pri_path.sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE | ((qp->port - 1) << 6); 1895 if (!qp->pri.smac && !qp->pri.smac_port) { 1896 smac_index = mlx4_register_mac(dev->dev, qp->port, u64_mac); 1897 if (smac_index >= 0) { 1898 qp->pri.candidate_smac_index = smac_index; 1899 qp->pri.candidate_smac = u64_mac; 1900 qp->pri.candidate_smac_port = qp->port; 1901 context->pri_path.grh_mylmc = 0x80 | (u8) smac_index; 1902 } else { 1903 return -ENOENT; 1904 } 1905 } 1906 return 0; 1907 } 1908 1909 static int create_qp_lb_counter(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp) 1910 { 1911 struct counter_index *new_counter_index; 1912 int err; 1913 u32 tmp_idx; 1914 1915 if (rdma_port_get_link_layer(&dev->ib_dev, qp->port) != 1916 IB_LINK_LAYER_ETHERNET || 1917 !(qp->flags & MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK) || 1918 !(dev->dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_LB_SRC_CHK)) 1919 return 0; 1920 1921 err = mlx4_counter_alloc(dev->dev, &tmp_idx, MLX4_RES_USAGE_DRIVER); 1922 if (err) 1923 return err; 1924 1925 new_counter_index = kmalloc(sizeof(*new_counter_index), GFP_KERNEL); 1926 if (!new_counter_index) { 1927 mlx4_counter_free(dev->dev, tmp_idx); 1928 return -ENOMEM; 1929 } 1930 1931 new_counter_index->index = tmp_idx; 1932 new_counter_index->allocated = 1; 1933 qp->counter_index = new_counter_index; 1934 1935 mutex_lock(&dev->counters_table[qp->port - 1].mutex); 1936 list_add_tail(&new_counter_index->list, 1937 &dev->counters_table[qp->port - 1].counters_list); 1938 mutex_unlock(&dev->counters_table[qp->port - 1].mutex); 1939 1940 return 0; 1941 } 1942 1943 enum { 1944 MLX4_QPC_ROCE_MODE_1 = 0, 1945 MLX4_QPC_ROCE_MODE_2 = 2, 1946 MLX4_QPC_ROCE_MODE_UNDEFINED = 0xff 1947 }; 1948 1949 static u8 gid_type_to_qpc(enum ib_gid_type gid_type) 1950 { 1951 switch (gid_type) { 1952 case IB_GID_TYPE_ROCE: 1953 return MLX4_QPC_ROCE_MODE_1; 1954 case IB_GID_TYPE_ROCE_UDP_ENCAP: 1955 return MLX4_QPC_ROCE_MODE_2; 1956 default: 1957 return MLX4_QPC_ROCE_MODE_UNDEFINED; 1958 } 1959 } 1960 1961 /* 1962 * Go over all RSS QP's childes (WQs) and apply their HW state according to 1963 * their logic state if the RSS QP is the first RSS QP associated for the WQ. 1964 */ 1965 static int bringup_rss_rwqs(struct ib_rwq_ind_table *ind_tbl, u8 port_num, 1966 struct ib_udata *udata) 1967 { 1968 int err = 0; 1969 int i; 1970 1971 for (i = 0; i < (1 << ind_tbl->log_ind_tbl_size); i++) { 1972 struct ib_wq *ibwq = ind_tbl->ind_tbl[i]; 1973 struct mlx4_ib_qp *wq = to_mqp((struct ib_qp *)ibwq); 1974 1975 mutex_lock(&wq->mutex); 1976 1977 /* Mlx4_ib restrictions: 1978 * WQ's is associated to a port according to the RSS QP it is 1979 * associates to. 1980 * In case the WQ is associated to a different port by another 1981 * RSS QP, return a failure. 1982 */ 1983 if ((wq->rss_usecnt > 0) && (wq->port != port_num)) { 1984 err = -EINVAL; 1985 mutex_unlock(&wq->mutex); 1986 break; 1987 } 1988 wq->port = port_num; 1989 if ((wq->rss_usecnt == 0) && (ibwq->state == IB_WQS_RDY)) { 1990 err = _mlx4_ib_modify_wq(ibwq, IB_WQS_RDY, udata); 1991 if (err) { 1992 mutex_unlock(&wq->mutex); 1993 break; 1994 } 1995 } 1996 wq->rss_usecnt++; 1997 1998 mutex_unlock(&wq->mutex); 1999 } 2000 2001 if (i && err) { 2002 int j; 2003 2004 for (j = (i - 1); j >= 0; j--) { 2005 struct ib_wq *ibwq = ind_tbl->ind_tbl[j]; 2006 struct mlx4_ib_qp *wq = to_mqp((struct ib_qp *)ibwq); 2007 2008 mutex_lock(&wq->mutex); 2009 2010 if ((wq->rss_usecnt == 1) && 2011 (ibwq->state == IB_WQS_RDY)) 2012 if (_mlx4_ib_modify_wq(ibwq, IB_WQS_RESET, 2013 udata)) 2014 pr_warn("failed to reverse WQN=0x%06x\n", 2015 ibwq->wq_num); 2016 wq->rss_usecnt--; 2017 2018 mutex_unlock(&wq->mutex); 2019 } 2020 } 2021 2022 return err; 2023 } 2024 2025 static void bring_down_rss_rwqs(struct ib_rwq_ind_table *ind_tbl, 2026 struct ib_udata *udata) 2027 { 2028 int i; 2029 2030 for (i = 0; i < (1 << ind_tbl->log_ind_tbl_size); i++) { 2031 struct ib_wq *ibwq = ind_tbl->ind_tbl[i]; 2032 struct mlx4_ib_qp *wq = to_mqp((struct ib_qp *)ibwq); 2033 2034 mutex_lock(&wq->mutex); 2035 2036 if ((wq->rss_usecnt == 1) && (ibwq->state == IB_WQS_RDY)) 2037 if (_mlx4_ib_modify_wq(ibwq, IB_WQS_RESET, udata)) 2038 pr_warn("failed to reverse WQN=%x\n", 2039 ibwq->wq_num); 2040 wq->rss_usecnt--; 2041 2042 mutex_unlock(&wq->mutex); 2043 } 2044 } 2045 2046 static void fill_qp_rss_context(struct mlx4_qp_context *context, 2047 struct mlx4_ib_qp *qp) 2048 { 2049 struct mlx4_rss_context *rss_context; 2050 2051 rss_context = (void *)context + offsetof(struct mlx4_qp_context, 2052 pri_path) + MLX4_RSS_OFFSET_IN_QPC_PRI_PATH; 2053 2054 rss_context->base_qpn = cpu_to_be32(qp->rss_ctx->base_qpn_tbl_sz); 2055 rss_context->default_qpn = 2056 cpu_to_be32(qp->rss_ctx->base_qpn_tbl_sz & 0xffffff); 2057 if (qp->rss_ctx->flags & (MLX4_RSS_UDP_IPV4 | MLX4_RSS_UDP_IPV6)) 2058 rss_context->base_qpn_udp = rss_context->default_qpn; 2059 rss_context->flags = qp->rss_ctx->flags; 2060 /* Currently support just toeplitz */ 2061 rss_context->hash_fn = MLX4_RSS_HASH_TOP; 2062 2063 memcpy(rss_context->rss_key, qp->rss_ctx->rss_key, 2064 MLX4_EN_RSS_KEY_SIZE); 2065 } 2066 2067 static int __mlx4_ib_modify_qp(void *src, enum mlx4_ib_source_type src_type, 2068 const struct ib_qp_attr *attr, int attr_mask, 2069 enum ib_qp_state cur_state, 2070 enum ib_qp_state new_state, 2071 struct ib_udata *udata) 2072 { 2073 struct ib_srq *ibsrq; 2074 const struct ib_gid_attr *gid_attr = NULL; 2075 struct ib_rwq_ind_table *rwq_ind_tbl; 2076 enum ib_qp_type qp_type; 2077 struct mlx4_ib_dev *dev; 2078 struct mlx4_ib_qp *qp; 2079 struct mlx4_ib_pd *pd; 2080 struct mlx4_ib_cq *send_cq, *recv_cq; 2081 struct mlx4_ib_ucontext *ucontext = rdma_udata_to_drv_context( 2082 udata, struct mlx4_ib_ucontext, ibucontext); 2083 struct mlx4_qp_context *context; 2084 enum mlx4_qp_optpar optpar = 0; 2085 int sqd_event; 2086 int steer_qp = 0; 2087 int err = -EINVAL; 2088 int counter_index; 2089 2090 if (src_type == MLX4_IB_RWQ_SRC) { 2091 struct ib_wq *ibwq; 2092 2093 ibwq = (struct ib_wq *)src; 2094 ibsrq = NULL; 2095 rwq_ind_tbl = NULL; 2096 qp_type = IB_QPT_RAW_PACKET; 2097 qp = to_mqp((struct ib_qp *)ibwq); 2098 dev = to_mdev(ibwq->device); 2099 pd = to_mpd(ibwq->pd); 2100 } else { 2101 struct ib_qp *ibqp; 2102 2103 ibqp = (struct ib_qp *)src; 2104 ibsrq = ibqp->srq; 2105 rwq_ind_tbl = ibqp->rwq_ind_tbl; 2106 qp_type = ibqp->qp_type; 2107 qp = to_mqp(ibqp); 2108 dev = to_mdev(ibqp->device); 2109 pd = get_pd(qp); 2110 } 2111 2112 /* APM is not supported under RoCE */ 2113 if (attr_mask & IB_QP_ALT_PATH && 2114 rdma_port_get_link_layer(&dev->ib_dev, qp->port) == 2115 IB_LINK_LAYER_ETHERNET) 2116 return -ENOTSUPP; 2117 2118 context = kzalloc(sizeof *context, GFP_KERNEL); 2119 if (!context) 2120 return -ENOMEM; 2121 2122 context->flags = cpu_to_be32((to_mlx4_state(new_state) << 28) | 2123 (to_mlx4_st(dev, qp->mlx4_ib_qp_type) << 16)); 2124 2125 if (!(attr_mask & IB_QP_PATH_MIG_STATE)) 2126 context->flags |= cpu_to_be32(MLX4_QP_PM_MIGRATED << 11); 2127 else { 2128 optpar |= MLX4_QP_OPTPAR_PM_STATE; 2129 switch (attr->path_mig_state) { 2130 case IB_MIG_MIGRATED: 2131 context->flags |= cpu_to_be32(MLX4_QP_PM_MIGRATED << 11); 2132 break; 2133 case IB_MIG_REARM: 2134 context->flags |= cpu_to_be32(MLX4_QP_PM_REARM << 11); 2135 break; 2136 case IB_MIG_ARMED: 2137 context->flags |= cpu_to_be32(MLX4_QP_PM_ARMED << 11); 2138 break; 2139 } 2140 } 2141 2142 if (qp->inl_recv_sz) 2143 context->param3 |= cpu_to_be32(1 << 25); 2144 2145 if (qp->flags & MLX4_IB_QP_SCATTER_FCS) 2146 context->param3 |= cpu_to_be32(1 << 29); 2147 2148 if (qp_type == IB_QPT_GSI || qp_type == IB_QPT_SMI) 2149 context->mtu_msgmax = (IB_MTU_4096 << 5) | 11; 2150 else if (qp_type == IB_QPT_RAW_PACKET) 2151 context->mtu_msgmax = (MLX4_RAW_QP_MTU << 5) | MLX4_RAW_QP_MSGMAX; 2152 else if (qp_type == IB_QPT_UD) { 2153 if (qp->flags & MLX4_IB_QP_LSO) 2154 context->mtu_msgmax = (IB_MTU_4096 << 5) | 2155 ilog2(dev->dev->caps.max_gso_sz); 2156 else 2157 context->mtu_msgmax = (IB_MTU_4096 << 5) | 13; 2158 } else if (attr_mask & IB_QP_PATH_MTU) { 2159 if (attr->path_mtu < IB_MTU_256 || attr->path_mtu > IB_MTU_4096) { 2160 pr_err("path MTU (%u) is invalid\n", 2161 attr->path_mtu); 2162 goto out; 2163 } 2164 context->mtu_msgmax = (attr->path_mtu << 5) | 2165 ilog2(dev->dev->caps.max_msg_sz); 2166 } 2167 2168 if (!rwq_ind_tbl) { /* PRM RSS receive side should be left zeros */ 2169 if (qp->rq.wqe_cnt) 2170 context->rq_size_stride = ilog2(qp->rq.wqe_cnt) << 3; 2171 context->rq_size_stride |= qp->rq.wqe_shift - 4; 2172 } 2173 2174 if (qp->sq.wqe_cnt) 2175 context->sq_size_stride = ilog2(qp->sq.wqe_cnt) << 3; 2176 context->sq_size_stride |= qp->sq.wqe_shift - 4; 2177 2178 if (new_state == IB_QPS_RESET && qp->counter_index) 2179 mlx4_ib_free_qp_counter(dev, qp); 2180 2181 if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) { 2182 context->sq_size_stride |= !!qp->sq_no_prefetch << 7; 2183 context->xrcd = cpu_to_be32((u32) qp->xrcdn); 2184 if (qp_type == IB_QPT_RAW_PACKET) 2185 context->param3 |= cpu_to_be32(1 << 30); 2186 } 2187 2188 if (ucontext) 2189 context->usr_page = cpu_to_be32( 2190 mlx4_to_hw_uar_index(dev->dev, ucontext->uar.index)); 2191 else 2192 context->usr_page = cpu_to_be32( 2193 mlx4_to_hw_uar_index(dev->dev, dev->priv_uar.index)); 2194 2195 if (attr_mask & IB_QP_DEST_QPN) 2196 context->remote_qpn = cpu_to_be32(attr->dest_qp_num); 2197 2198 if (attr_mask & IB_QP_PORT) { 2199 if (cur_state == IB_QPS_SQD && new_state == IB_QPS_SQD && 2200 !(attr_mask & IB_QP_AV)) { 2201 mlx4_set_sched(&context->pri_path, attr->port_num); 2202 optpar |= MLX4_QP_OPTPAR_SCHED_QUEUE; 2203 } 2204 } 2205 2206 if (cur_state == IB_QPS_INIT && new_state == IB_QPS_RTR) { 2207 err = create_qp_lb_counter(dev, qp); 2208 if (err) 2209 goto out; 2210 2211 counter_index = 2212 dev->counters_table[qp->port - 1].default_counter; 2213 if (qp->counter_index) 2214 counter_index = qp->counter_index->index; 2215 2216 if (counter_index != -1) { 2217 context->pri_path.counter_index = counter_index; 2218 optpar |= MLX4_QP_OPTPAR_COUNTER_INDEX; 2219 if (qp->counter_index) { 2220 context->pri_path.fl |= 2221 MLX4_FL_ETH_SRC_CHECK_MC_LB; 2222 context->pri_path.vlan_control |= 2223 MLX4_CTRL_ETH_SRC_CHECK_IF_COUNTER; 2224 } 2225 } else 2226 context->pri_path.counter_index = 2227 MLX4_SINK_COUNTER_INDEX(dev->dev); 2228 2229 if (qp->flags & MLX4_IB_QP_NETIF) { 2230 mlx4_ib_steer_qp_reg(dev, qp, 1); 2231 steer_qp = 1; 2232 } 2233 2234 if (qp_type == IB_QPT_GSI) { 2235 enum ib_gid_type gid_type = qp->flags & MLX4_IB_ROCE_V2_GSI_QP ? 2236 IB_GID_TYPE_ROCE_UDP_ENCAP : IB_GID_TYPE_ROCE; 2237 u8 qpc_roce_mode = gid_type_to_qpc(gid_type); 2238 2239 context->rlkey_roce_mode |= (qpc_roce_mode << 6); 2240 } 2241 } 2242 2243 if (attr_mask & IB_QP_PKEY_INDEX) { 2244 if (qp->mlx4_ib_qp_type & MLX4_IB_QPT_ANY_SRIOV) 2245 context->pri_path.disable_pkey_check = 0x40; 2246 context->pri_path.pkey_index = attr->pkey_index; 2247 optpar |= MLX4_QP_OPTPAR_PKEY_INDEX; 2248 } 2249 2250 if (attr_mask & IB_QP_AV) { 2251 u8 port_num = mlx4_is_bonded(dev->dev) ? 1 : 2252 attr_mask & IB_QP_PORT ? attr->port_num : qp->port; 2253 u16 vlan = 0xffff; 2254 u8 smac[ETH_ALEN]; 2255 int is_eth = 2256 rdma_cap_eth_ah(&dev->ib_dev, port_num) && 2257 rdma_ah_get_ah_flags(&attr->ah_attr) & IB_AH_GRH; 2258 2259 if (is_eth) { 2260 gid_attr = attr->ah_attr.grh.sgid_attr; 2261 err = rdma_read_gid_l2_fields(gid_attr, &vlan, 2262 &smac[0]); 2263 if (err) 2264 goto out; 2265 } 2266 2267 if (mlx4_set_path(dev, attr, attr_mask, qp, &context->pri_path, 2268 port_num, vlan, smac)) 2269 goto out; 2270 2271 optpar |= (MLX4_QP_OPTPAR_PRIMARY_ADDR_PATH | 2272 MLX4_QP_OPTPAR_SCHED_QUEUE); 2273 2274 if (is_eth && 2275 (cur_state == IB_QPS_INIT && new_state == IB_QPS_RTR)) { 2276 u8 qpc_roce_mode = gid_type_to_qpc(gid_attr->gid_type); 2277 2278 if (qpc_roce_mode == MLX4_QPC_ROCE_MODE_UNDEFINED) { 2279 err = -EINVAL; 2280 goto out; 2281 } 2282 context->rlkey_roce_mode |= (qpc_roce_mode << 6); 2283 } 2284 2285 } 2286 2287 if (attr_mask & IB_QP_TIMEOUT) { 2288 context->pri_path.ackto |= attr->timeout << 3; 2289 optpar |= MLX4_QP_OPTPAR_ACK_TIMEOUT; 2290 } 2291 2292 if (attr_mask & IB_QP_ALT_PATH) { 2293 if (attr->alt_port_num == 0 || 2294 attr->alt_port_num > dev->dev->caps.num_ports) 2295 goto out; 2296 2297 if (attr->alt_pkey_index >= 2298 dev->dev->caps.pkey_table_len[attr->alt_port_num]) 2299 goto out; 2300 2301 if (mlx4_set_alt_path(dev, attr, attr_mask, qp, 2302 &context->alt_path, 2303 attr->alt_port_num)) 2304 goto out; 2305 2306 context->alt_path.pkey_index = attr->alt_pkey_index; 2307 context->alt_path.ackto = attr->alt_timeout << 3; 2308 optpar |= MLX4_QP_OPTPAR_ALT_ADDR_PATH; 2309 } 2310 2311 context->pd = cpu_to_be32(pd->pdn); 2312 2313 if (!rwq_ind_tbl) { 2314 context->params1 = cpu_to_be32(MLX4_IB_ACK_REQ_FREQ << 28); 2315 get_cqs(qp, src_type, &send_cq, &recv_cq); 2316 } else { /* Set dummy CQs to be compatible with HV and PRM */ 2317 send_cq = to_mcq(rwq_ind_tbl->ind_tbl[0]->cq); 2318 recv_cq = send_cq; 2319 } 2320 context->cqn_send = cpu_to_be32(send_cq->mcq.cqn); 2321 context->cqn_recv = cpu_to_be32(recv_cq->mcq.cqn); 2322 2323 /* Set "fast registration enabled" for all kernel QPs */ 2324 if (!ucontext) 2325 context->params1 |= cpu_to_be32(1 << 11); 2326 2327 if (attr_mask & IB_QP_RNR_RETRY) { 2328 context->params1 |= cpu_to_be32(attr->rnr_retry << 13); 2329 optpar |= MLX4_QP_OPTPAR_RNR_RETRY; 2330 } 2331 2332 if (attr_mask & IB_QP_RETRY_CNT) { 2333 context->params1 |= cpu_to_be32(attr->retry_cnt << 16); 2334 optpar |= MLX4_QP_OPTPAR_RETRY_COUNT; 2335 } 2336 2337 if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) { 2338 if (attr->max_rd_atomic) 2339 context->params1 |= 2340 cpu_to_be32(fls(attr->max_rd_atomic - 1) << 21); 2341 optpar |= MLX4_QP_OPTPAR_SRA_MAX; 2342 } 2343 2344 if (attr_mask & IB_QP_SQ_PSN) 2345 context->next_send_psn = cpu_to_be32(attr->sq_psn); 2346 2347 if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) { 2348 if (attr->max_dest_rd_atomic) 2349 context->params2 |= 2350 cpu_to_be32(fls(attr->max_dest_rd_atomic - 1) << 21); 2351 optpar |= MLX4_QP_OPTPAR_RRA_MAX; 2352 } 2353 2354 if (attr_mask & (IB_QP_ACCESS_FLAGS | IB_QP_MAX_DEST_RD_ATOMIC)) { 2355 context->params2 |= to_mlx4_access_flags(qp, attr, attr_mask); 2356 optpar |= MLX4_QP_OPTPAR_RWE | MLX4_QP_OPTPAR_RRE | MLX4_QP_OPTPAR_RAE; 2357 } 2358 2359 if (ibsrq) 2360 context->params2 |= cpu_to_be32(MLX4_QP_BIT_RIC); 2361 2362 if (attr_mask & IB_QP_MIN_RNR_TIMER) { 2363 context->rnr_nextrecvpsn |= cpu_to_be32(attr->min_rnr_timer << 24); 2364 optpar |= MLX4_QP_OPTPAR_RNR_TIMEOUT; 2365 } 2366 if (attr_mask & IB_QP_RQ_PSN) 2367 context->rnr_nextrecvpsn |= cpu_to_be32(attr->rq_psn); 2368 2369 /* proxy and tunnel qp qkeys will be changed in modify-qp wrappers */ 2370 if (attr_mask & IB_QP_QKEY) { 2371 if (qp->mlx4_ib_qp_type & 2372 (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER)) 2373 context->qkey = cpu_to_be32(IB_QP_SET_QKEY); 2374 else { 2375 if (mlx4_is_mfunc(dev->dev) && 2376 !(qp->mlx4_ib_qp_type & MLX4_IB_QPT_ANY_SRIOV) && 2377 (attr->qkey & MLX4_RESERVED_QKEY_MASK) == 2378 MLX4_RESERVED_QKEY_BASE) { 2379 pr_err("Cannot use reserved QKEY" 2380 " 0x%x (range 0xffff0000..0xffffffff" 2381 " is reserved)\n", attr->qkey); 2382 err = -EINVAL; 2383 goto out; 2384 } 2385 context->qkey = cpu_to_be32(attr->qkey); 2386 } 2387 optpar |= MLX4_QP_OPTPAR_Q_KEY; 2388 } 2389 2390 if (ibsrq) 2391 context->srqn = cpu_to_be32(1 << 24 | 2392 to_msrq(ibsrq)->msrq.srqn); 2393 2394 if (qp->rq.wqe_cnt && 2395 cur_state == IB_QPS_RESET && 2396 new_state == IB_QPS_INIT) 2397 context->db_rec_addr = cpu_to_be64(qp->db.dma); 2398 2399 if (cur_state == IB_QPS_INIT && 2400 new_state == IB_QPS_RTR && 2401 (qp_type == IB_QPT_GSI || qp_type == IB_QPT_SMI || 2402 qp_type == IB_QPT_UD || qp_type == IB_QPT_RAW_PACKET)) { 2403 context->pri_path.sched_queue = (qp->port - 1) << 6; 2404 if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_SMI || 2405 qp->mlx4_ib_qp_type & 2406 (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER)) { 2407 context->pri_path.sched_queue |= MLX4_IB_DEFAULT_QP0_SCHED_QUEUE; 2408 if (qp->mlx4_ib_qp_type != MLX4_IB_QPT_SMI) 2409 context->pri_path.fl = 0x80; 2410 } else { 2411 if (qp->mlx4_ib_qp_type & MLX4_IB_QPT_ANY_SRIOV) 2412 context->pri_path.fl = 0x80; 2413 context->pri_path.sched_queue |= MLX4_IB_DEFAULT_SCHED_QUEUE; 2414 } 2415 if (rdma_port_get_link_layer(&dev->ib_dev, qp->port) == 2416 IB_LINK_LAYER_ETHERNET) { 2417 if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_TUN_GSI || 2418 qp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI) 2419 context->pri_path.feup = 1 << 7; /* don't fsm */ 2420 /* handle smac_index */ 2421 if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_UD || 2422 qp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI || 2423 qp->mlx4_ib_qp_type == MLX4_IB_QPT_TUN_GSI) { 2424 err = handle_eth_ud_smac_index(dev, qp, context); 2425 if (err) { 2426 err = -EINVAL; 2427 goto out; 2428 } 2429 if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI) 2430 dev->qp1_proxy[qp->port - 1] = qp; 2431 } 2432 } 2433 } 2434 2435 if (qp_type == IB_QPT_RAW_PACKET) { 2436 context->pri_path.ackto = (context->pri_path.ackto & 0xf8) | 2437 MLX4_IB_LINK_TYPE_ETH; 2438 if (dev->dev->caps.tunnel_offload_mode == MLX4_TUNNEL_OFFLOAD_MODE_VXLAN) { 2439 /* set QP to receive both tunneled & non-tunneled packets */ 2440 if (!rwq_ind_tbl) 2441 context->srqn = cpu_to_be32(7 << 28); 2442 } 2443 } 2444 2445 if (qp_type == IB_QPT_UD && (new_state == IB_QPS_RTR)) { 2446 int is_eth = rdma_port_get_link_layer( 2447 &dev->ib_dev, qp->port) == 2448 IB_LINK_LAYER_ETHERNET; 2449 if (is_eth) { 2450 context->pri_path.ackto = MLX4_IB_LINK_TYPE_ETH; 2451 optpar |= MLX4_QP_OPTPAR_PRIMARY_ADDR_PATH; 2452 } 2453 } 2454 2455 if (cur_state == IB_QPS_RTS && new_state == IB_QPS_SQD && 2456 attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY && attr->en_sqd_async_notify) 2457 sqd_event = 1; 2458 else 2459 sqd_event = 0; 2460 2461 if (!ucontext && 2462 cur_state == IB_QPS_RESET && 2463 new_state == IB_QPS_INIT) 2464 context->rlkey_roce_mode |= (1 << 4); 2465 2466 /* 2467 * Before passing a kernel QP to the HW, make sure that the 2468 * ownership bits of the send queue are set and the SQ 2469 * headroom is stamped so that the hardware doesn't start 2470 * processing stale work requests. 2471 */ 2472 if (!ucontext && 2473 cur_state == IB_QPS_RESET && 2474 new_state == IB_QPS_INIT) { 2475 struct mlx4_wqe_ctrl_seg *ctrl; 2476 int i; 2477 2478 for (i = 0; i < qp->sq.wqe_cnt; ++i) { 2479 ctrl = get_send_wqe(qp, i); 2480 ctrl->owner_opcode = cpu_to_be32(1 << 31); 2481 ctrl->qpn_vlan.fence_size = 2482 1 << (qp->sq.wqe_shift - 4); 2483 stamp_send_wqe(qp, i); 2484 } 2485 } 2486 2487 if (rwq_ind_tbl && 2488 cur_state == IB_QPS_RESET && 2489 new_state == IB_QPS_INIT) { 2490 fill_qp_rss_context(context, qp); 2491 context->flags |= cpu_to_be32(1 << MLX4_RSS_QPC_FLAG_OFFSET); 2492 } 2493 2494 err = mlx4_qp_modify(dev->dev, &qp->mtt, to_mlx4_state(cur_state), 2495 to_mlx4_state(new_state), context, optpar, 2496 sqd_event, &qp->mqp); 2497 if (err) 2498 goto out; 2499 2500 qp->state = new_state; 2501 2502 if (attr_mask & IB_QP_ACCESS_FLAGS) 2503 qp->atomic_rd_en = attr->qp_access_flags; 2504 if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) 2505 qp->resp_depth = attr->max_dest_rd_atomic; 2506 if (attr_mask & IB_QP_PORT) { 2507 qp->port = attr->port_num; 2508 update_mcg_macs(dev, qp); 2509 } 2510 if (attr_mask & IB_QP_ALT_PATH) 2511 qp->alt_port = attr->alt_port_num; 2512 2513 if (is_sqp(dev, qp)) 2514 store_sqp_attrs(qp->sqp, attr, attr_mask); 2515 2516 /* 2517 * If we moved QP0 to RTR, bring the IB link up; if we moved 2518 * QP0 to RESET or ERROR, bring the link back down. 2519 */ 2520 if (is_qp0(dev, qp)) { 2521 if (cur_state != IB_QPS_RTR && new_state == IB_QPS_RTR) 2522 if (mlx4_INIT_PORT(dev->dev, qp->port)) 2523 pr_warn("INIT_PORT failed for port %d\n", 2524 qp->port); 2525 2526 if (cur_state != IB_QPS_RESET && cur_state != IB_QPS_ERR && 2527 (new_state == IB_QPS_RESET || new_state == IB_QPS_ERR)) 2528 mlx4_CLOSE_PORT(dev->dev, qp->port); 2529 } 2530 2531 /* 2532 * If we moved a kernel QP to RESET, clean up all old CQ 2533 * entries and reinitialize the QP. 2534 */ 2535 if (new_state == IB_QPS_RESET) { 2536 if (!ucontext) { 2537 mlx4_ib_cq_clean(recv_cq, qp->mqp.qpn, 2538 ibsrq ? to_msrq(ibsrq) : NULL); 2539 if (send_cq != recv_cq) 2540 mlx4_ib_cq_clean(send_cq, qp->mqp.qpn, NULL); 2541 2542 qp->rq.head = 0; 2543 qp->rq.tail = 0; 2544 qp->sq.head = 0; 2545 qp->sq.tail = 0; 2546 qp->sq_next_wqe = 0; 2547 if (qp->rq.wqe_cnt) 2548 *qp->db.db = 0; 2549 2550 if (qp->flags & MLX4_IB_QP_NETIF) 2551 mlx4_ib_steer_qp_reg(dev, qp, 0); 2552 } 2553 if (qp->pri.smac || (!qp->pri.smac && qp->pri.smac_port)) { 2554 mlx4_unregister_mac(dev->dev, qp->pri.smac_port, qp->pri.smac); 2555 qp->pri.smac = 0; 2556 qp->pri.smac_port = 0; 2557 } 2558 if (qp->alt.smac) { 2559 mlx4_unregister_mac(dev->dev, qp->alt.smac_port, qp->alt.smac); 2560 qp->alt.smac = 0; 2561 } 2562 if (qp->pri.vid < 0x1000) { 2563 mlx4_unregister_vlan(dev->dev, qp->pri.vlan_port, qp->pri.vid); 2564 qp->pri.vid = 0xFFFF; 2565 qp->pri.candidate_vid = 0xFFFF; 2566 qp->pri.update_vid = 0; 2567 } 2568 2569 if (qp->alt.vid < 0x1000) { 2570 mlx4_unregister_vlan(dev->dev, qp->alt.vlan_port, qp->alt.vid); 2571 qp->alt.vid = 0xFFFF; 2572 qp->alt.candidate_vid = 0xFFFF; 2573 qp->alt.update_vid = 0; 2574 } 2575 } 2576 out: 2577 if (err && qp->counter_index) 2578 mlx4_ib_free_qp_counter(dev, qp); 2579 if (err && steer_qp) 2580 mlx4_ib_steer_qp_reg(dev, qp, 0); 2581 kfree(context); 2582 if (qp->pri.candidate_smac || 2583 (!qp->pri.candidate_smac && qp->pri.candidate_smac_port)) { 2584 if (err) { 2585 mlx4_unregister_mac(dev->dev, qp->pri.candidate_smac_port, qp->pri.candidate_smac); 2586 } else { 2587 if (qp->pri.smac || (!qp->pri.smac && qp->pri.smac_port)) 2588 mlx4_unregister_mac(dev->dev, qp->pri.smac_port, qp->pri.smac); 2589 qp->pri.smac = qp->pri.candidate_smac; 2590 qp->pri.smac_index = qp->pri.candidate_smac_index; 2591 qp->pri.smac_port = qp->pri.candidate_smac_port; 2592 } 2593 qp->pri.candidate_smac = 0; 2594 qp->pri.candidate_smac_index = 0; 2595 qp->pri.candidate_smac_port = 0; 2596 } 2597 if (qp->alt.candidate_smac) { 2598 if (err) { 2599 mlx4_unregister_mac(dev->dev, qp->alt.candidate_smac_port, qp->alt.candidate_smac); 2600 } else { 2601 if (qp->alt.smac) 2602 mlx4_unregister_mac(dev->dev, qp->alt.smac_port, qp->alt.smac); 2603 qp->alt.smac = qp->alt.candidate_smac; 2604 qp->alt.smac_index = qp->alt.candidate_smac_index; 2605 qp->alt.smac_port = qp->alt.candidate_smac_port; 2606 } 2607 qp->alt.candidate_smac = 0; 2608 qp->alt.candidate_smac_index = 0; 2609 qp->alt.candidate_smac_port = 0; 2610 } 2611 2612 if (qp->pri.update_vid) { 2613 if (err) { 2614 if (qp->pri.candidate_vid < 0x1000) 2615 mlx4_unregister_vlan(dev->dev, qp->pri.candidate_vlan_port, 2616 qp->pri.candidate_vid); 2617 } else { 2618 if (qp->pri.vid < 0x1000) 2619 mlx4_unregister_vlan(dev->dev, qp->pri.vlan_port, 2620 qp->pri.vid); 2621 qp->pri.vid = qp->pri.candidate_vid; 2622 qp->pri.vlan_port = qp->pri.candidate_vlan_port; 2623 qp->pri.vlan_index = qp->pri.candidate_vlan_index; 2624 } 2625 qp->pri.candidate_vid = 0xFFFF; 2626 qp->pri.update_vid = 0; 2627 } 2628 2629 if (qp->alt.update_vid) { 2630 if (err) { 2631 if (qp->alt.candidate_vid < 0x1000) 2632 mlx4_unregister_vlan(dev->dev, qp->alt.candidate_vlan_port, 2633 qp->alt.candidate_vid); 2634 } else { 2635 if (qp->alt.vid < 0x1000) 2636 mlx4_unregister_vlan(dev->dev, qp->alt.vlan_port, 2637 qp->alt.vid); 2638 qp->alt.vid = qp->alt.candidate_vid; 2639 qp->alt.vlan_port = qp->alt.candidate_vlan_port; 2640 qp->alt.vlan_index = qp->alt.candidate_vlan_index; 2641 } 2642 qp->alt.candidate_vid = 0xFFFF; 2643 qp->alt.update_vid = 0; 2644 } 2645 2646 return err; 2647 } 2648 2649 enum { 2650 MLX4_IB_MODIFY_QP_RSS_SUP_ATTR_MSK = (IB_QP_STATE | 2651 IB_QP_PORT), 2652 }; 2653 2654 static int _mlx4_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, 2655 int attr_mask, struct ib_udata *udata) 2656 { 2657 struct mlx4_ib_dev *dev = to_mdev(ibqp->device); 2658 struct mlx4_ib_qp *qp = to_mqp(ibqp); 2659 enum ib_qp_state cur_state, new_state; 2660 int err = -EINVAL; 2661 mutex_lock(&qp->mutex); 2662 2663 cur_state = attr_mask & IB_QP_CUR_STATE ? attr->cur_qp_state : qp->state; 2664 new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state; 2665 2666 if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type, 2667 attr_mask)) { 2668 pr_debug("qpn 0x%x: invalid attribute mask specified " 2669 "for transition %d to %d. qp_type %d," 2670 " attr_mask 0x%x\n", 2671 ibqp->qp_num, cur_state, new_state, 2672 ibqp->qp_type, attr_mask); 2673 goto out; 2674 } 2675 2676 if (ibqp->rwq_ind_tbl) { 2677 if (!(((cur_state == IB_QPS_RESET) && 2678 (new_state == IB_QPS_INIT)) || 2679 ((cur_state == IB_QPS_INIT) && 2680 (new_state == IB_QPS_RTR)))) { 2681 pr_debug("qpn 0x%x: RSS QP unsupported transition %d to %d\n", 2682 ibqp->qp_num, cur_state, new_state); 2683 2684 err = -EOPNOTSUPP; 2685 goto out; 2686 } 2687 2688 if (attr_mask & ~MLX4_IB_MODIFY_QP_RSS_SUP_ATTR_MSK) { 2689 pr_debug("qpn 0x%x: RSS QP unsupported attribute mask 0x%x for transition %d to %d\n", 2690 ibqp->qp_num, attr_mask, cur_state, new_state); 2691 2692 err = -EOPNOTSUPP; 2693 goto out; 2694 } 2695 } 2696 2697 if (mlx4_is_bonded(dev->dev) && (attr_mask & IB_QP_PORT)) { 2698 if ((cur_state == IB_QPS_RESET) && (new_state == IB_QPS_INIT)) { 2699 if ((ibqp->qp_type == IB_QPT_RC) || 2700 (ibqp->qp_type == IB_QPT_UD) || 2701 (ibqp->qp_type == IB_QPT_UC) || 2702 (ibqp->qp_type == IB_QPT_RAW_PACKET) || 2703 (ibqp->qp_type == IB_QPT_XRC_INI)) { 2704 attr->port_num = mlx4_ib_bond_next_port(dev); 2705 } 2706 } else { 2707 /* no sense in changing port_num 2708 * when ports are bonded */ 2709 attr_mask &= ~IB_QP_PORT; 2710 } 2711 } 2712 2713 if ((attr_mask & IB_QP_PORT) && 2714 (attr->port_num == 0 || attr->port_num > dev->num_ports)) { 2715 pr_debug("qpn 0x%x: invalid port number (%d) specified " 2716 "for transition %d to %d. qp_type %d\n", 2717 ibqp->qp_num, attr->port_num, cur_state, 2718 new_state, ibqp->qp_type); 2719 goto out; 2720 } 2721 2722 if ((attr_mask & IB_QP_PORT) && (ibqp->qp_type == IB_QPT_RAW_PACKET) && 2723 (rdma_port_get_link_layer(&dev->ib_dev, attr->port_num) != 2724 IB_LINK_LAYER_ETHERNET)) 2725 goto out; 2726 2727 if (attr_mask & IB_QP_PKEY_INDEX) { 2728 int p = attr_mask & IB_QP_PORT ? attr->port_num : qp->port; 2729 if (attr->pkey_index >= dev->dev->caps.pkey_table_len[p]) { 2730 pr_debug("qpn 0x%x: invalid pkey index (%d) specified " 2731 "for transition %d to %d. qp_type %d\n", 2732 ibqp->qp_num, attr->pkey_index, cur_state, 2733 new_state, ibqp->qp_type); 2734 goto out; 2735 } 2736 } 2737 2738 if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC && 2739 attr->max_rd_atomic > dev->dev->caps.max_qp_init_rdma) { 2740 pr_debug("qpn 0x%x: max_rd_atomic (%d) too large. " 2741 "Transition %d to %d. qp_type %d\n", 2742 ibqp->qp_num, attr->max_rd_atomic, cur_state, 2743 new_state, ibqp->qp_type); 2744 goto out; 2745 } 2746 2747 if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC && 2748 attr->max_dest_rd_atomic > dev->dev->caps.max_qp_dest_rdma) { 2749 pr_debug("qpn 0x%x: max_dest_rd_atomic (%d) too large. " 2750 "Transition %d to %d. qp_type %d\n", 2751 ibqp->qp_num, attr->max_dest_rd_atomic, cur_state, 2752 new_state, ibqp->qp_type); 2753 goto out; 2754 } 2755 2756 if (cur_state == new_state && cur_state == IB_QPS_RESET) { 2757 err = 0; 2758 goto out; 2759 } 2760 2761 if (ibqp->rwq_ind_tbl && (new_state == IB_QPS_INIT)) { 2762 err = bringup_rss_rwqs(ibqp->rwq_ind_tbl, attr->port_num, 2763 udata); 2764 if (err) 2765 goto out; 2766 } 2767 2768 err = __mlx4_ib_modify_qp(ibqp, MLX4_IB_QP_SRC, attr, attr_mask, 2769 cur_state, new_state, udata); 2770 2771 if (ibqp->rwq_ind_tbl && err) 2772 bring_down_rss_rwqs(ibqp->rwq_ind_tbl, udata); 2773 2774 if (mlx4_is_bonded(dev->dev) && (attr_mask & IB_QP_PORT)) 2775 attr->port_num = 1; 2776 2777 out: 2778 mutex_unlock(&qp->mutex); 2779 return err; 2780 } 2781 2782 int mlx4_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, 2783 int attr_mask, struct ib_udata *udata) 2784 { 2785 struct mlx4_ib_qp *mqp = to_mqp(ibqp); 2786 int ret; 2787 2788 if (attr_mask & ~IB_QP_ATTR_STANDARD_BITS) 2789 return -EOPNOTSUPP; 2790 2791 ret = _mlx4_ib_modify_qp(ibqp, attr, attr_mask, udata); 2792 2793 if (mqp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI) { 2794 struct mlx4_ib_sqp *sqp = mqp->sqp; 2795 int err = 0; 2796 2797 if (sqp->roce_v2_gsi) 2798 err = ib_modify_qp(sqp->roce_v2_gsi, attr, attr_mask); 2799 if (err) 2800 pr_err("Failed to modify GSI QP for RoCEv2 (%d)\n", 2801 err); 2802 } 2803 return ret; 2804 } 2805 2806 static int vf_get_qp0_qkey(struct mlx4_dev *dev, int qpn, u32 *qkey) 2807 { 2808 int i; 2809 for (i = 0; i < dev->caps.num_ports; i++) { 2810 if (qpn == dev->caps.spec_qps[i].qp0_proxy || 2811 qpn == dev->caps.spec_qps[i].qp0_tunnel) { 2812 *qkey = dev->caps.spec_qps[i].qp0_qkey; 2813 return 0; 2814 } 2815 } 2816 return -EINVAL; 2817 } 2818 2819 static int build_sriov_qp0_header(struct mlx4_ib_qp *qp, 2820 const struct ib_ud_wr *wr, 2821 void *wqe, unsigned *mlx_seg_len) 2822 { 2823 struct mlx4_ib_dev *mdev = to_mdev(qp->ibqp.device); 2824 struct mlx4_ib_sqp *sqp = qp->sqp; 2825 struct ib_device *ib_dev = qp->ibqp.device; 2826 struct mlx4_wqe_mlx_seg *mlx = wqe; 2827 struct mlx4_wqe_inline_seg *inl = wqe + sizeof *mlx; 2828 struct mlx4_ib_ah *ah = to_mah(wr->ah); 2829 u16 pkey; 2830 u32 qkey; 2831 int send_size; 2832 int header_size; 2833 int spc; 2834 int err; 2835 int i; 2836 2837 if (wr->wr.opcode != IB_WR_SEND) 2838 return -EINVAL; 2839 2840 send_size = 0; 2841 2842 for (i = 0; i < wr->wr.num_sge; ++i) 2843 send_size += wr->wr.sg_list[i].length; 2844 2845 /* for proxy-qp0 sends, need to add in size of tunnel header */ 2846 /* for tunnel-qp0 sends, tunnel header is already in s/g list */ 2847 if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_SMI_OWNER) 2848 send_size += sizeof (struct mlx4_ib_tunnel_header); 2849 2850 ib_ud_header_init(send_size, 1, 0, 0, 0, 0, 0, 0, &sqp->ud_header); 2851 2852 if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_SMI_OWNER) { 2853 sqp->ud_header.lrh.service_level = 2854 be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 28; 2855 sqp->ud_header.lrh.destination_lid = 2856 cpu_to_be16(ah->av.ib.g_slid & 0x7f); 2857 sqp->ud_header.lrh.source_lid = 2858 cpu_to_be16(ah->av.ib.g_slid & 0x7f); 2859 } 2860 2861 mlx->flags &= cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE); 2862 2863 /* force loopback */ 2864 mlx->flags |= cpu_to_be32(MLX4_WQE_MLX_VL15 | 0x1 | MLX4_WQE_MLX_SLR); 2865 mlx->rlid = sqp->ud_header.lrh.destination_lid; 2866 2867 sqp->ud_header.lrh.virtual_lane = 0; 2868 sqp->ud_header.bth.solicited_event = !!(wr->wr.send_flags & IB_SEND_SOLICITED); 2869 err = ib_get_cached_pkey(ib_dev, qp->port, 0, &pkey); 2870 if (err) 2871 return err; 2872 sqp->ud_header.bth.pkey = cpu_to_be16(pkey); 2873 if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_TUN_SMI_OWNER) 2874 sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->remote_qpn); 2875 else 2876 sqp->ud_header.bth.destination_qpn = 2877 cpu_to_be32(mdev->dev->caps.spec_qps[qp->port - 1].qp0_tunnel); 2878 2879 sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1)); 2880 if (mlx4_is_master(mdev->dev)) { 2881 if (mlx4_get_parav_qkey(mdev->dev, qp->mqp.qpn, &qkey)) 2882 return -EINVAL; 2883 } else { 2884 if (vf_get_qp0_qkey(mdev->dev, qp->mqp.qpn, &qkey)) 2885 return -EINVAL; 2886 } 2887 sqp->ud_header.deth.qkey = cpu_to_be32(qkey); 2888 sqp->ud_header.deth.source_qpn = cpu_to_be32(qp->mqp.qpn); 2889 2890 sqp->ud_header.bth.opcode = IB_OPCODE_UD_SEND_ONLY; 2891 sqp->ud_header.immediate_present = 0; 2892 2893 header_size = ib_ud_header_pack(&sqp->ud_header, sqp->header_buf); 2894 2895 /* 2896 * Inline data segments may not cross a 64 byte boundary. If 2897 * our UD header is bigger than the space available up to the 2898 * next 64 byte boundary in the WQE, use two inline data 2899 * segments to hold the UD header. 2900 */ 2901 spc = MLX4_INLINE_ALIGN - 2902 ((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1)); 2903 if (header_size <= spc) { 2904 inl->byte_count = cpu_to_be32(1 << 31 | header_size); 2905 memcpy(inl + 1, sqp->header_buf, header_size); 2906 i = 1; 2907 } else { 2908 inl->byte_count = cpu_to_be32(1 << 31 | spc); 2909 memcpy(inl + 1, sqp->header_buf, spc); 2910 2911 inl = (void *) (inl + 1) + spc; 2912 memcpy(inl + 1, sqp->header_buf + spc, header_size - spc); 2913 /* 2914 * Need a barrier here to make sure all the data is 2915 * visible before the byte_count field is set. 2916 * Otherwise the HCA prefetcher could grab the 64-byte 2917 * chunk with this inline segment and get a valid (!= 2918 * 0xffffffff) byte count but stale data, and end up 2919 * generating a packet with bad headers. 2920 * 2921 * The first inline segment's byte_count field doesn't 2922 * need a barrier, because it comes after a 2923 * control/MLX segment and therefore is at an offset 2924 * of 16 mod 64. 2925 */ 2926 wmb(); 2927 inl->byte_count = cpu_to_be32(1 << 31 | (header_size - spc)); 2928 i = 2; 2929 } 2930 2931 *mlx_seg_len = 2932 ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + header_size, 16); 2933 return 0; 2934 } 2935 2936 static u8 sl_to_vl(struct mlx4_ib_dev *dev, u8 sl, int port_num) 2937 { 2938 union sl2vl_tbl_to_u64 tmp_vltab; 2939 u8 vl; 2940 2941 if (sl > 15) 2942 return 0xf; 2943 tmp_vltab.sl64 = atomic64_read(&dev->sl2vl[port_num - 1]); 2944 vl = tmp_vltab.sl8[sl >> 1]; 2945 if (sl & 1) 2946 vl &= 0x0f; 2947 else 2948 vl >>= 4; 2949 return vl; 2950 } 2951 2952 static int fill_gid_by_hw_index(struct mlx4_ib_dev *ibdev, u8 port_num, 2953 int index, union ib_gid *gid, 2954 enum ib_gid_type *gid_type) 2955 { 2956 struct mlx4_ib_iboe *iboe = &ibdev->iboe; 2957 struct mlx4_port_gid_table *port_gid_table; 2958 unsigned long flags; 2959 2960 port_gid_table = &iboe->gids[port_num - 1]; 2961 spin_lock_irqsave(&iboe->lock, flags); 2962 memcpy(gid, &port_gid_table->gids[index].gid, sizeof(*gid)); 2963 *gid_type = port_gid_table->gids[index].gid_type; 2964 spin_unlock_irqrestore(&iboe->lock, flags); 2965 if (rdma_is_zero_gid(gid)) 2966 return -ENOENT; 2967 2968 return 0; 2969 } 2970 2971 #define MLX4_ROCEV2_QP1_SPORT 0xC000 2972 static int build_mlx_header(struct mlx4_ib_qp *qp, const struct ib_ud_wr *wr, 2973 void *wqe, unsigned *mlx_seg_len) 2974 { 2975 struct mlx4_ib_sqp *sqp = qp->sqp; 2976 struct ib_device *ib_dev = qp->ibqp.device; 2977 struct mlx4_ib_dev *ibdev = to_mdev(ib_dev); 2978 struct mlx4_wqe_mlx_seg *mlx = wqe; 2979 struct mlx4_wqe_ctrl_seg *ctrl = wqe; 2980 struct mlx4_wqe_inline_seg *inl = wqe + sizeof *mlx; 2981 struct mlx4_ib_ah *ah = to_mah(wr->ah); 2982 union ib_gid sgid; 2983 u16 pkey; 2984 int send_size; 2985 int header_size; 2986 int spc; 2987 int i; 2988 int err = 0; 2989 u16 vlan = 0xffff; 2990 bool is_eth; 2991 bool is_vlan = false; 2992 bool is_grh; 2993 bool is_udp = false; 2994 int ip_version = 0; 2995 2996 send_size = 0; 2997 for (i = 0; i < wr->wr.num_sge; ++i) 2998 send_size += wr->wr.sg_list[i].length; 2999 3000 is_eth = rdma_port_get_link_layer(qp->ibqp.device, qp->port) == IB_LINK_LAYER_ETHERNET; 3001 is_grh = mlx4_ib_ah_grh_present(ah); 3002 if (is_eth) { 3003 enum ib_gid_type gid_type; 3004 if (mlx4_is_mfunc(to_mdev(ib_dev)->dev)) { 3005 /* When multi-function is enabled, the ib_core gid 3006 * indexes don't necessarily match the hw ones, so 3007 * we must use our own cache */ 3008 err = mlx4_get_roce_gid_from_slave(to_mdev(ib_dev)->dev, 3009 be32_to_cpu(ah->av.ib.port_pd) >> 24, 3010 ah->av.ib.gid_index, &sgid.raw[0]); 3011 if (err) 3012 return err; 3013 } else { 3014 err = fill_gid_by_hw_index(ibdev, qp->port, 3015 ah->av.ib.gid_index, &sgid, 3016 &gid_type); 3017 if (!err) { 3018 is_udp = gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP; 3019 if (is_udp) { 3020 if (ipv6_addr_v4mapped((struct in6_addr *)&sgid)) 3021 ip_version = 4; 3022 else 3023 ip_version = 6; 3024 is_grh = false; 3025 } 3026 } else { 3027 return err; 3028 } 3029 } 3030 if (ah->av.eth.vlan != cpu_to_be16(0xffff)) { 3031 vlan = be16_to_cpu(ah->av.eth.vlan) & 0x0fff; 3032 is_vlan = true; 3033 } 3034 } 3035 err = ib_ud_header_init(send_size, !is_eth, is_eth, is_vlan, is_grh, 3036 ip_version, is_udp, 0, &sqp->ud_header); 3037 if (err) 3038 return err; 3039 3040 if (!is_eth) { 3041 sqp->ud_header.lrh.service_level = 3042 be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 28; 3043 sqp->ud_header.lrh.destination_lid = ah->av.ib.dlid; 3044 sqp->ud_header.lrh.source_lid = cpu_to_be16(ah->av.ib.g_slid & 0x7f); 3045 } 3046 3047 if (is_grh || (ip_version == 6)) { 3048 sqp->ud_header.grh.traffic_class = 3049 (be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 20) & 0xff; 3050 sqp->ud_header.grh.flow_label = 3051 ah->av.ib.sl_tclass_flowlabel & cpu_to_be32(0xfffff); 3052 sqp->ud_header.grh.hop_limit = ah->av.ib.hop_limit; 3053 if (is_eth) { 3054 memcpy(sqp->ud_header.grh.source_gid.raw, sgid.raw, 16); 3055 } else { 3056 if (mlx4_is_mfunc(to_mdev(ib_dev)->dev)) { 3057 /* When multi-function is enabled, the ib_core gid 3058 * indexes don't necessarily match the hw ones, so 3059 * we must use our own cache 3060 */ 3061 sqp->ud_header.grh.source_gid.global 3062 .subnet_prefix = 3063 cpu_to_be64(atomic64_read( 3064 &(to_mdev(ib_dev) 3065 ->sriov 3066 .demux[qp->port - 1] 3067 .subnet_prefix))); 3068 sqp->ud_header.grh.source_gid.global 3069 .interface_id = 3070 to_mdev(ib_dev) 3071 ->sriov.demux[qp->port - 1] 3072 .guid_cache[ah->av.ib.gid_index]; 3073 } else { 3074 sqp->ud_header.grh.source_gid = 3075 ah->ibah.sgid_attr->gid; 3076 } 3077 } 3078 memcpy(sqp->ud_header.grh.destination_gid.raw, 3079 ah->av.ib.dgid, 16); 3080 } 3081 3082 if (ip_version == 4) { 3083 sqp->ud_header.ip4.tos = 3084 (be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 20) & 0xff; 3085 sqp->ud_header.ip4.id = 0; 3086 sqp->ud_header.ip4.frag_off = htons(IP_DF); 3087 sqp->ud_header.ip4.ttl = ah->av.eth.hop_limit; 3088 3089 memcpy(&sqp->ud_header.ip4.saddr, 3090 sgid.raw + 12, 4); 3091 memcpy(&sqp->ud_header.ip4.daddr, ah->av.ib.dgid + 12, 4); 3092 sqp->ud_header.ip4.check = ib_ud_ip4_csum(&sqp->ud_header); 3093 } 3094 3095 if (is_udp) { 3096 sqp->ud_header.udp.dport = htons(ROCE_V2_UDP_DPORT); 3097 sqp->ud_header.udp.sport = htons(MLX4_ROCEV2_QP1_SPORT); 3098 sqp->ud_header.udp.csum = 0; 3099 } 3100 3101 mlx->flags &= cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE); 3102 3103 if (!is_eth) { 3104 mlx->flags |= 3105 cpu_to_be32((!qp->ibqp.qp_num ? MLX4_WQE_MLX_VL15 : 0) | 3106 (sqp->ud_header.lrh.destination_lid == 3107 IB_LID_PERMISSIVE ? 3108 MLX4_WQE_MLX_SLR : 3109 0) | 3110 (sqp->ud_header.lrh.service_level << 8)); 3111 if (ah->av.ib.port_pd & cpu_to_be32(0x80000000)) 3112 mlx->flags |= cpu_to_be32(0x1); /* force loopback */ 3113 mlx->rlid = sqp->ud_header.lrh.destination_lid; 3114 } 3115 3116 switch (wr->wr.opcode) { 3117 case IB_WR_SEND: 3118 sqp->ud_header.bth.opcode = IB_OPCODE_UD_SEND_ONLY; 3119 sqp->ud_header.immediate_present = 0; 3120 break; 3121 case IB_WR_SEND_WITH_IMM: 3122 sqp->ud_header.bth.opcode = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE; 3123 sqp->ud_header.immediate_present = 1; 3124 sqp->ud_header.immediate_data = wr->wr.ex.imm_data; 3125 break; 3126 default: 3127 return -EINVAL; 3128 } 3129 3130 if (is_eth) { 3131 u16 ether_type; 3132 u16 pcp = (be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 29) << 13; 3133 3134 ether_type = (!is_udp) ? ETH_P_IBOE: 3135 (ip_version == 4 ? ETH_P_IP : ETH_P_IPV6); 3136 3137 mlx->sched_prio = cpu_to_be16(pcp); 3138 3139 ether_addr_copy(sqp->ud_header.eth.smac_h, ah->av.eth.s_mac); 3140 ether_addr_copy(sqp->ud_header.eth.dmac_h, ah->av.eth.mac); 3141 memcpy(&ctrl->srcrb_flags16[0], ah->av.eth.mac, 2); 3142 memcpy(&ctrl->imm, ah->av.eth.mac + 2, 4); 3143 3144 if (!memcmp(sqp->ud_header.eth.smac_h, sqp->ud_header.eth.dmac_h, 6)) 3145 mlx->flags |= cpu_to_be32(MLX4_WQE_CTRL_FORCE_LOOPBACK); 3146 if (!is_vlan) { 3147 sqp->ud_header.eth.type = cpu_to_be16(ether_type); 3148 } else { 3149 sqp->ud_header.vlan.type = cpu_to_be16(ether_type); 3150 sqp->ud_header.vlan.tag = cpu_to_be16(vlan | pcp); 3151 } 3152 } else { 3153 sqp->ud_header.lrh.virtual_lane = 3154 !qp->ibqp.qp_num ? 3155 15 : 3156 sl_to_vl(to_mdev(ib_dev), 3157 sqp->ud_header.lrh.service_level, 3158 qp->port); 3159 if (qp->ibqp.qp_num && sqp->ud_header.lrh.virtual_lane == 15) 3160 return -EINVAL; 3161 if (sqp->ud_header.lrh.destination_lid == IB_LID_PERMISSIVE) 3162 sqp->ud_header.lrh.source_lid = IB_LID_PERMISSIVE; 3163 } 3164 sqp->ud_header.bth.solicited_event = !!(wr->wr.send_flags & IB_SEND_SOLICITED); 3165 if (!qp->ibqp.qp_num) 3166 err = ib_get_cached_pkey(ib_dev, qp->port, sqp->pkey_index, 3167 &pkey); 3168 else 3169 err = ib_get_cached_pkey(ib_dev, qp->port, wr->pkey_index, 3170 &pkey); 3171 if (err) 3172 return err; 3173 3174 sqp->ud_header.bth.pkey = cpu_to_be16(pkey); 3175 sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->remote_qpn); 3176 sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1)); 3177 sqp->ud_header.deth.qkey = cpu_to_be32(wr->remote_qkey & 0x80000000 ? 3178 sqp->qkey : wr->remote_qkey); 3179 sqp->ud_header.deth.source_qpn = cpu_to_be32(qp->ibqp.qp_num); 3180 3181 header_size = ib_ud_header_pack(&sqp->ud_header, sqp->header_buf); 3182 3183 if (0) { 3184 pr_err("built UD header of size %d:\n", header_size); 3185 for (i = 0; i < header_size / 4; ++i) { 3186 if (i % 8 == 0) 3187 pr_err(" [%02x] ", i * 4); 3188 pr_cont(" %08x", 3189 be32_to_cpu(((__be32 *) sqp->header_buf)[i])); 3190 if ((i + 1) % 8 == 0) 3191 pr_cont("\n"); 3192 } 3193 pr_err("\n"); 3194 } 3195 3196 /* 3197 * Inline data segments may not cross a 64 byte boundary. If 3198 * our UD header is bigger than the space available up to the 3199 * next 64 byte boundary in the WQE, use two inline data 3200 * segments to hold the UD header. 3201 */ 3202 spc = MLX4_INLINE_ALIGN - 3203 ((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1)); 3204 if (header_size <= spc) { 3205 inl->byte_count = cpu_to_be32(1 << 31 | header_size); 3206 memcpy(inl + 1, sqp->header_buf, header_size); 3207 i = 1; 3208 } else { 3209 inl->byte_count = cpu_to_be32(1 << 31 | spc); 3210 memcpy(inl + 1, sqp->header_buf, spc); 3211 3212 inl = (void *) (inl + 1) + spc; 3213 memcpy(inl + 1, sqp->header_buf + spc, header_size - spc); 3214 /* 3215 * Need a barrier here to make sure all the data is 3216 * visible before the byte_count field is set. 3217 * Otherwise the HCA prefetcher could grab the 64-byte 3218 * chunk with this inline segment and get a valid (!= 3219 * 0xffffffff) byte count but stale data, and end up 3220 * generating a packet with bad headers. 3221 * 3222 * The first inline segment's byte_count field doesn't 3223 * need a barrier, because it comes after a 3224 * control/MLX segment and therefore is at an offset 3225 * of 16 mod 64. 3226 */ 3227 wmb(); 3228 inl->byte_count = cpu_to_be32(1 << 31 | (header_size - spc)); 3229 i = 2; 3230 } 3231 3232 *mlx_seg_len = 3233 ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + header_size, 16); 3234 return 0; 3235 } 3236 3237 static int mlx4_wq_overflow(struct mlx4_ib_wq *wq, int nreq, struct ib_cq *ib_cq) 3238 { 3239 unsigned cur; 3240 struct mlx4_ib_cq *cq; 3241 3242 cur = wq->head - wq->tail; 3243 if (likely(cur + nreq < wq->max_post)) 3244 return 0; 3245 3246 cq = to_mcq(ib_cq); 3247 spin_lock(&cq->lock); 3248 cur = wq->head - wq->tail; 3249 spin_unlock(&cq->lock); 3250 3251 return cur + nreq >= wq->max_post; 3252 } 3253 3254 static __be32 convert_access(int acc) 3255 { 3256 return (acc & IB_ACCESS_REMOTE_ATOMIC ? 3257 cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_ATOMIC) : 0) | 3258 (acc & IB_ACCESS_REMOTE_WRITE ? 3259 cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_REMOTE_WRITE) : 0) | 3260 (acc & IB_ACCESS_REMOTE_READ ? 3261 cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_REMOTE_READ) : 0) | 3262 (acc & IB_ACCESS_LOCAL_WRITE ? cpu_to_be32(MLX4_WQE_FMR_PERM_LOCAL_WRITE) : 0) | 3263 cpu_to_be32(MLX4_WQE_FMR_PERM_LOCAL_READ); 3264 } 3265 3266 static void set_reg_seg(struct mlx4_wqe_fmr_seg *fseg, 3267 const struct ib_reg_wr *wr) 3268 { 3269 struct mlx4_ib_mr *mr = to_mmr(wr->mr); 3270 3271 fseg->flags = convert_access(wr->access); 3272 fseg->mem_key = cpu_to_be32(wr->key); 3273 fseg->buf_list = cpu_to_be64(mr->page_map); 3274 fseg->start_addr = cpu_to_be64(mr->ibmr.iova); 3275 fseg->reg_len = cpu_to_be64(mr->ibmr.length); 3276 fseg->offset = 0; /* XXX -- is this just for ZBVA? */ 3277 fseg->page_size = cpu_to_be32(ilog2(mr->ibmr.page_size)); 3278 fseg->reserved[0] = 0; 3279 fseg->reserved[1] = 0; 3280 } 3281 3282 static void set_local_inv_seg(struct mlx4_wqe_local_inval_seg *iseg, u32 rkey) 3283 { 3284 memset(iseg, 0, sizeof(*iseg)); 3285 iseg->mem_key = cpu_to_be32(rkey); 3286 } 3287 3288 static __always_inline void set_raddr_seg(struct mlx4_wqe_raddr_seg *rseg, 3289 u64 remote_addr, u32 rkey) 3290 { 3291 rseg->raddr = cpu_to_be64(remote_addr); 3292 rseg->rkey = cpu_to_be32(rkey); 3293 rseg->reserved = 0; 3294 } 3295 3296 static void set_atomic_seg(struct mlx4_wqe_atomic_seg *aseg, 3297 const struct ib_atomic_wr *wr) 3298 { 3299 if (wr->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP) { 3300 aseg->swap_add = cpu_to_be64(wr->swap); 3301 aseg->compare = cpu_to_be64(wr->compare_add); 3302 } else if (wr->wr.opcode == IB_WR_MASKED_ATOMIC_FETCH_AND_ADD) { 3303 aseg->swap_add = cpu_to_be64(wr->compare_add); 3304 aseg->compare = cpu_to_be64(wr->compare_add_mask); 3305 } else { 3306 aseg->swap_add = cpu_to_be64(wr->compare_add); 3307 aseg->compare = 0; 3308 } 3309 3310 } 3311 3312 static void set_masked_atomic_seg(struct mlx4_wqe_masked_atomic_seg *aseg, 3313 const struct ib_atomic_wr *wr) 3314 { 3315 aseg->swap_add = cpu_to_be64(wr->swap); 3316 aseg->swap_add_mask = cpu_to_be64(wr->swap_mask); 3317 aseg->compare = cpu_to_be64(wr->compare_add); 3318 aseg->compare_mask = cpu_to_be64(wr->compare_add_mask); 3319 } 3320 3321 static void set_datagram_seg(struct mlx4_wqe_datagram_seg *dseg, 3322 const struct ib_ud_wr *wr) 3323 { 3324 memcpy(dseg->av, &to_mah(wr->ah)->av, sizeof (struct mlx4_av)); 3325 dseg->dqpn = cpu_to_be32(wr->remote_qpn); 3326 dseg->qkey = cpu_to_be32(wr->remote_qkey); 3327 dseg->vlan = to_mah(wr->ah)->av.eth.vlan; 3328 memcpy(dseg->mac, to_mah(wr->ah)->av.eth.mac, 6); 3329 } 3330 3331 static void set_tunnel_datagram_seg(struct mlx4_ib_dev *dev, 3332 struct mlx4_wqe_datagram_seg *dseg, 3333 const struct ib_ud_wr *wr, 3334 enum mlx4_ib_qp_type qpt) 3335 { 3336 union mlx4_ext_av *av = &to_mah(wr->ah)->av; 3337 struct mlx4_av sqp_av = {0}; 3338 int port = *((u8 *) &av->ib.port_pd) & 0x3; 3339 3340 /* force loopback */ 3341 sqp_av.port_pd = av->ib.port_pd | cpu_to_be32(0x80000000); 3342 sqp_av.g_slid = av->ib.g_slid & 0x7f; /* no GRH */ 3343 sqp_av.sl_tclass_flowlabel = av->ib.sl_tclass_flowlabel & 3344 cpu_to_be32(0xf0000000); 3345 3346 memcpy(dseg->av, &sqp_av, sizeof (struct mlx4_av)); 3347 if (qpt == MLX4_IB_QPT_PROXY_GSI) 3348 dseg->dqpn = cpu_to_be32(dev->dev->caps.spec_qps[port - 1].qp1_tunnel); 3349 else 3350 dseg->dqpn = cpu_to_be32(dev->dev->caps.spec_qps[port - 1].qp0_tunnel); 3351 /* Use QKEY from the QP context, which is set by master */ 3352 dseg->qkey = cpu_to_be32(IB_QP_SET_QKEY); 3353 } 3354 3355 static void build_tunnel_header(const struct ib_ud_wr *wr, void *wqe, 3356 unsigned *mlx_seg_len) 3357 { 3358 struct mlx4_wqe_inline_seg *inl = wqe; 3359 struct mlx4_ib_tunnel_header hdr; 3360 struct mlx4_ib_ah *ah = to_mah(wr->ah); 3361 int spc; 3362 int i; 3363 3364 memcpy(&hdr.av, &ah->av, sizeof hdr.av); 3365 hdr.remote_qpn = cpu_to_be32(wr->remote_qpn); 3366 hdr.pkey_index = cpu_to_be16(wr->pkey_index); 3367 hdr.qkey = cpu_to_be32(wr->remote_qkey); 3368 memcpy(hdr.mac, ah->av.eth.mac, 6); 3369 hdr.vlan = ah->av.eth.vlan; 3370 3371 spc = MLX4_INLINE_ALIGN - 3372 ((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1)); 3373 if (sizeof (hdr) <= spc) { 3374 memcpy(inl + 1, &hdr, sizeof (hdr)); 3375 wmb(); 3376 inl->byte_count = cpu_to_be32(1 << 31 | sizeof (hdr)); 3377 i = 1; 3378 } else { 3379 memcpy(inl + 1, &hdr, spc); 3380 wmb(); 3381 inl->byte_count = cpu_to_be32(1 << 31 | spc); 3382 3383 inl = (void *) (inl + 1) + spc; 3384 memcpy(inl + 1, (void *) &hdr + spc, sizeof (hdr) - spc); 3385 wmb(); 3386 inl->byte_count = cpu_to_be32(1 << 31 | (sizeof (hdr) - spc)); 3387 i = 2; 3388 } 3389 3390 *mlx_seg_len = 3391 ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + sizeof (hdr), 16); 3392 } 3393 3394 static void set_mlx_icrc_seg(void *dseg) 3395 { 3396 u32 *t = dseg; 3397 struct mlx4_wqe_inline_seg *iseg = dseg; 3398 3399 t[1] = 0; 3400 3401 /* 3402 * Need a barrier here before writing the byte_count field to 3403 * make sure that all the data is visible before the 3404 * byte_count field is set. Otherwise, if the segment begins 3405 * a new cacheline, the HCA prefetcher could grab the 64-byte 3406 * chunk and get a valid (!= * 0xffffffff) byte count but 3407 * stale data, and end up sending the wrong data. 3408 */ 3409 wmb(); 3410 3411 iseg->byte_count = cpu_to_be32((1 << 31) | 4); 3412 } 3413 3414 static void set_data_seg(struct mlx4_wqe_data_seg *dseg, struct ib_sge *sg) 3415 { 3416 dseg->lkey = cpu_to_be32(sg->lkey); 3417 dseg->addr = cpu_to_be64(sg->addr); 3418 3419 /* 3420 * Need a barrier here before writing the byte_count field to 3421 * make sure that all the data is visible before the 3422 * byte_count field is set. Otherwise, if the segment begins 3423 * a new cacheline, the HCA prefetcher could grab the 64-byte 3424 * chunk and get a valid (!= * 0xffffffff) byte count but 3425 * stale data, and end up sending the wrong data. 3426 */ 3427 wmb(); 3428 3429 dseg->byte_count = cpu_to_be32(sg->length); 3430 } 3431 3432 static void __set_data_seg(struct mlx4_wqe_data_seg *dseg, struct ib_sge *sg) 3433 { 3434 dseg->byte_count = cpu_to_be32(sg->length); 3435 dseg->lkey = cpu_to_be32(sg->lkey); 3436 dseg->addr = cpu_to_be64(sg->addr); 3437 } 3438 3439 static int build_lso_seg(struct mlx4_wqe_lso_seg *wqe, 3440 const struct ib_ud_wr *wr, struct mlx4_ib_qp *qp, 3441 unsigned *lso_seg_len, __be32 *lso_hdr_sz, __be32 *blh) 3442 { 3443 unsigned halign = ALIGN(sizeof *wqe + wr->hlen, 16); 3444 3445 if (unlikely(halign > MLX4_IB_CACHE_LINE_SIZE)) 3446 *blh = cpu_to_be32(1 << 6); 3447 3448 if (unlikely(!(qp->flags & MLX4_IB_QP_LSO) && 3449 wr->wr.num_sge > qp->sq.max_gs - (halign >> 4))) 3450 return -EINVAL; 3451 3452 memcpy(wqe->header, wr->header, wr->hlen); 3453 3454 *lso_hdr_sz = cpu_to_be32(wr->mss << 16 | wr->hlen); 3455 *lso_seg_len = halign; 3456 return 0; 3457 } 3458 3459 static __be32 send_ieth(const struct ib_send_wr *wr) 3460 { 3461 switch (wr->opcode) { 3462 case IB_WR_SEND_WITH_IMM: 3463 case IB_WR_RDMA_WRITE_WITH_IMM: 3464 return wr->ex.imm_data; 3465 3466 case IB_WR_SEND_WITH_INV: 3467 return cpu_to_be32(wr->ex.invalidate_rkey); 3468 3469 default: 3470 return 0; 3471 } 3472 } 3473 3474 static void add_zero_len_inline(void *wqe) 3475 { 3476 struct mlx4_wqe_inline_seg *inl = wqe; 3477 memset(wqe, 0, 16); 3478 inl->byte_count = cpu_to_be32(1 << 31); 3479 } 3480 3481 static int _mlx4_ib_post_send(struct ib_qp *ibqp, const struct ib_send_wr *wr, 3482 const struct ib_send_wr **bad_wr, bool drain) 3483 { 3484 struct mlx4_ib_qp *qp = to_mqp(ibqp); 3485 void *wqe; 3486 struct mlx4_wqe_ctrl_seg *ctrl; 3487 struct mlx4_wqe_data_seg *dseg; 3488 unsigned long flags; 3489 int nreq; 3490 int err = 0; 3491 unsigned ind; 3492 int size; 3493 unsigned seglen; 3494 __be32 dummy; 3495 __be32 *lso_wqe; 3496 __be32 lso_hdr_sz; 3497 __be32 blh; 3498 int i; 3499 struct mlx4_ib_dev *mdev = to_mdev(ibqp->device); 3500 3501 if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI) { 3502 struct mlx4_ib_sqp *sqp = qp->sqp; 3503 3504 if (sqp->roce_v2_gsi) { 3505 struct mlx4_ib_ah *ah = to_mah(ud_wr(wr)->ah); 3506 enum ib_gid_type gid_type; 3507 union ib_gid gid; 3508 3509 if (!fill_gid_by_hw_index(mdev, qp->port, 3510 ah->av.ib.gid_index, 3511 &gid, &gid_type)) 3512 qp = (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) ? 3513 to_mqp(sqp->roce_v2_gsi) : qp; 3514 else 3515 pr_err("Failed to get gid at index %d. RoCEv2 will not work properly\n", 3516 ah->av.ib.gid_index); 3517 } 3518 } 3519 3520 spin_lock_irqsave(&qp->sq.lock, flags); 3521 if (mdev->dev->persist->state & MLX4_DEVICE_STATE_INTERNAL_ERROR && 3522 !drain) { 3523 err = -EIO; 3524 *bad_wr = wr; 3525 nreq = 0; 3526 goto out; 3527 } 3528 3529 ind = qp->sq_next_wqe; 3530 3531 for (nreq = 0; wr; ++nreq, wr = wr->next) { 3532 lso_wqe = &dummy; 3533 blh = 0; 3534 3535 if (mlx4_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) { 3536 err = -ENOMEM; 3537 *bad_wr = wr; 3538 goto out; 3539 } 3540 3541 if (unlikely(wr->num_sge > qp->sq.max_gs)) { 3542 err = -EINVAL; 3543 *bad_wr = wr; 3544 goto out; 3545 } 3546 3547 ctrl = wqe = get_send_wqe(qp, ind & (qp->sq.wqe_cnt - 1)); 3548 qp->sq.wrid[(qp->sq.head + nreq) & (qp->sq.wqe_cnt - 1)] = wr->wr_id; 3549 3550 ctrl->srcrb_flags = 3551 (wr->send_flags & IB_SEND_SIGNALED ? 3552 cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE) : 0) | 3553 (wr->send_flags & IB_SEND_SOLICITED ? 3554 cpu_to_be32(MLX4_WQE_CTRL_SOLICITED) : 0) | 3555 ((wr->send_flags & IB_SEND_IP_CSUM) ? 3556 cpu_to_be32(MLX4_WQE_CTRL_IP_CSUM | 3557 MLX4_WQE_CTRL_TCP_UDP_CSUM) : 0) | 3558 qp->sq_signal_bits; 3559 3560 ctrl->imm = send_ieth(wr); 3561 3562 wqe += sizeof *ctrl; 3563 size = sizeof *ctrl / 16; 3564 3565 switch (qp->mlx4_ib_qp_type) { 3566 case MLX4_IB_QPT_RC: 3567 case MLX4_IB_QPT_UC: 3568 switch (wr->opcode) { 3569 case IB_WR_ATOMIC_CMP_AND_SWP: 3570 case IB_WR_ATOMIC_FETCH_AND_ADD: 3571 case IB_WR_MASKED_ATOMIC_FETCH_AND_ADD: 3572 set_raddr_seg(wqe, atomic_wr(wr)->remote_addr, 3573 atomic_wr(wr)->rkey); 3574 wqe += sizeof (struct mlx4_wqe_raddr_seg); 3575 3576 set_atomic_seg(wqe, atomic_wr(wr)); 3577 wqe += sizeof (struct mlx4_wqe_atomic_seg); 3578 3579 size += (sizeof (struct mlx4_wqe_raddr_seg) + 3580 sizeof (struct mlx4_wqe_atomic_seg)) / 16; 3581 3582 break; 3583 3584 case IB_WR_MASKED_ATOMIC_CMP_AND_SWP: 3585 set_raddr_seg(wqe, atomic_wr(wr)->remote_addr, 3586 atomic_wr(wr)->rkey); 3587 wqe += sizeof (struct mlx4_wqe_raddr_seg); 3588 3589 set_masked_atomic_seg(wqe, atomic_wr(wr)); 3590 wqe += sizeof (struct mlx4_wqe_masked_atomic_seg); 3591 3592 size += (sizeof (struct mlx4_wqe_raddr_seg) + 3593 sizeof (struct mlx4_wqe_masked_atomic_seg)) / 16; 3594 3595 break; 3596 3597 case IB_WR_RDMA_READ: 3598 case IB_WR_RDMA_WRITE: 3599 case IB_WR_RDMA_WRITE_WITH_IMM: 3600 set_raddr_seg(wqe, rdma_wr(wr)->remote_addr, 3601 rdma_wr(wr)->rkey); 3602 wqe += sizeof (struct mlx4_wqe_raddr_seg); 3603 size += sizeof (struct mlx4_wqe_raddr_seg) / 16; 3604 break; 3605 3606 case IB_WR_LOCAL_INV: 3607 ctrl->srcrb_flags |= 3608 cpu_to_be32(MLX4_WQE_CTRL_STRONG_ORDER); 3609 set_local_inv_seg(wqe, wr->ex.invalidate_rkey); 3610 wqe += sizeof (struct mlx4_wqe_local_inval_seg); 3611 size += sizeof (struct mlx4_wqe_local_inval_seg) / 16; 3612 break; 3613 3614 case IB_WR_REG_MR: 3615 ctrl->srcrb_flags |= 3616 cpu_to_be32(MLX4_WQE_CTRL_STRONG_ORDER); 3617 set_reg_seg(wqe, reg_wr(wr)); 3618 wqe += sizeof(struct mlx4_wqe_fmr_seg); 3619 size += sizeof(struct mlx4_wqe_fmr_seg) / 16; 3620 break; 3621 3622 default: 3623 /* No extra segments required for sends */ 3624 break; 3625 } 3626 break; 3627 3628 case MLX4_IB_QPT_TUN_SMI_OWNER: 3629 err = build_sriov_qp0_header(qp, ud_wr(wr), ctrl, 3630 &seglen); 3631 if (unlikely(err)) { 3632 *bad_wr = wr; 3633 goto out; 3634 } 3635 wqe += seglen; 3636 size += seglen / 16; 3637 break; 3638 case MLX4_IB_QPT_TUN_SMI: 3639 case MLX4_IB_QPT_TUN_GSI: 3640 /* this is a UD qp used in MAD responses to slaves. */ 3641 set_datagram_seg(wqe, ud_wr(wr)); 3642 /* set the forced-loopback bit in the data seg av */ 3643 *(__be32 *) wqe |= cpu_to_be32(0x80000000); 3644 wqe += sizeof (struct mlx4_wqe_datagram_seg); 3645 size += sizeof (struct mlx4_wqe_datagram_seg) / 16; 3646 break; 3647 case MLX4_IB_QPT_UD: 3648 set_datagram_seg(wqe, ud_wr(wr)); 3649 wqe += sizeof (struct mlx4_wqe_datagram_seg); 3650 size += sizeof (struct mlx4_wqe_datagram_seg) / 16; 3651 3652 if (wr->opcode == IB_WR_LSO) { 3653 err = build_lso_seg(wqe, ud_wr(wr), qp, &seglen, 3654 &lso_hdr_sz, &blh); 3655 if (unlikely(err)) { 3656 *bad_wr = wr; 3657 goto out; 3658 } 3659 lso_wqe = (__be32 *) wqe; 3660 wqe += seglen; 3661 size += seglen / 16; 3662 } 3663 break; 3664 3665 case MLX4_IB_QPT_PROXY_SMI_OWNER: 3666 err = build_sriov_qp0_header(qp, ud_wr(wr), ctrl, 3667 &seglen); 3668 if (unlikely(err)) { 3669 *bad_wr = wr; 3670 goto out; 3671 } 3672 wqe += seglen; 3673 size += seglen / 16; 3674 /* to start tunnel header on a cache-line boundary */ 3675 add_zero_len_inline(wqe); 3676 wqe += 16; 3677 size++; 3678 build_tunnel_header(ud_wr(wr), wqe, &seglen); 3679 wqe += seglen; 3680 size += seglen / 16; 3681 break; 3682 case MLX4_IB_QPT_PROXY_SMI: 3683 case MLX4_IB_QPT_PROXY_GSI: 3684 /* If we are tunneling special qps, this is a UD qp. 3685 * In this case we first add a UD segment targeting 3686 * the tunnel qp, and then add a header with address 3687 * information */ 3688 set_tunnel_datagram_seg(to_mdev(ibqp->device), wqe, 3689 ud_wr(wr), 3690 qp->mlx4_ib_qp_type); 3691 wqe += sizeof (struct mlx4_wqe_datagram_seg); 3692 size += sizeof (struct mlx4_wqe_datagram_seg) / 16; 3693 build_tunnel_header(ud_wr(wr), wqe, &seglen); 3694 wqe += seglen; 3695 size += seglen / 16; 3696 break; 3697 3698 case MLX4_IB_QPT_SMI: 3699 case MLX4_IB_QPT_GSI: 3700 err = build_mlx_header(qp, ud_wr(wr), ctrl, &seglen); 3701 if (unlikely(err)) { 3702 *bad_wr = wr; 3703 goto out; 3704 } 3705 wqe += seglen; 3706 size += seglen / 16; 3707 break; 3708 3709 default: 3710 break; 3711 } 3712 3713 /* 3714 * Write data segments in reverse order, so as to 3715 * overwrite cacheline stamp last within each 3716 * cacheline. This avoids issues with WQE 3717 * prefetching. 3718 */ 3719 3720 dseg = wqe; 3721 dseg += wr->num_sge - 1; 3722 size += wr->num_sge * (sizeof (struct mlx4_wqe_data_seg) / 16); 3723 3724 /* Add one more inline data segment for ICRC for MLX sends */ 3725 if (unlikely(qp->mlx4_ib_qp_type == MLX4_IB_QPT_SMI || 3726 qp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI || 3727 qp->mlx4_ib_qp_type & 3728 (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER))) { 3729 set_mlx_icrc_seg(dseg + 1); 3730 size += sizeof (struct mlx4_wqe_data_seg) / 16; 3731 } 3732 3733 for (i = wr->num_sge - 1; i >= 0; --i, --dseg) 3734 set_data_seg(dseg, wr->sg_list + i); 3735 3736 /* 3737 * Possibly overwrite stamping in cacheline with LSO 3738 * segment only after making sure all data segments 3739 * are written. 3740 */ 3741 wmb(); 3742 *lso_wqe = lso_hdr_sz; 3743 3744 ctrl->qpn_vlan.fence_size = (wr->send_flags & IB_SEND_FENCE ? 3745 MLX4_WQE_CTRL_FENCE : 0) | size; 3746 3747 /* 3748 * Make sure descriptor is fully written before 3749 * setting ownership bit (because HW can start 3750 * executing as soon as we do). 3751 */ 3752 wmb(); 3753 3754 if (wr->opcode < 0 || wr->opcode >= ARRAY_SIZE(mlx4_ib_opcode)) { 3755 *bad_wr = wr; 3756 err = -EINVAL; 3757 goto out; 3758 } 3759 3760 ctrl->owner_opcode = mlx4_ib_opcode[wr->opcode] | 3761 (ind & qp->sq.wqe_cnt ? cpu_to_be32(1 << 31) : 0) | blh; 3762 3763 /* 3764 * We can improve latency by not stamping the last 3765 * send queue WQE until after ringing the doorbell, so 3766 * only stamp here if there are still more WQEs to post. 3767 */ 3768 if (wr->next) 3769 stamp_send_wqe(qp, ind + qp->sq_spare_wqes); 3770 ind++; 3771 } 3772 3773 out: 3774 if (likely(nreq)) { 3775 qp->sq.head += nreq; 3776 3777 /* 3778 * Make sure that descriptors are written before 3779 * doorbell record. 3780 */ 3781 wmb(); 3782 3783 writel_relaxed(qp->doorbell_qpn, 3784 to_mdev(ibqp->device)->uar_map + MLX4_SEND_DOORBELL); 3785 3786 stamp_send_wqe(qp, ind + qp->sq_spare_wqes - 1); 3787 3788 qp->sq_next_wqe = ind; 3789 } 3790 3791 spin_unlock_irqrestore(&qp->sq.lock, flags); 3792 3793 return err; 3794 } 3795 3796 int mlx4_ib_post_send(struct ib_qp *ibqp, const struct ib_send_wr *wr, 3797 const struct ib_send_wr **bad_wr) 3798 { 3799 return _mlx4_ib_post_send(ibqp, wr, bad_wr, false); 3800 } 3801 3802 static int _mlx4_ib_post_recv(struct ib_qp *ibqp, const struct ib_recv_wr *wr, 3803 const struct ib_recv_wr **bad_wr, bool drain) 3804 { 3805 struct mlx4_ib_qp *qp = to_mqp(ibqp); 3806 struct mlx4_wqe_data_seg *scat; 3807 unsigned long flags; 3808 int err = 0; 3809 int nreq; 3810 int ind; 3811 int max_gs; 3812 int i; 3813 struct mlx4_ib_dev *mdev = to_mdev(ibqp->device); 3814 3815 max_gs = qp->rq.max_gs; 3816 spin_lock_irqsave(&qp->rq.lock, flags); 3817 3818 if (mdev->dev->persist->state & MLX4_DEVICE_STATE_INTERNAL_ERROR && 3819 !drain) { 3820 err = -EIO; 3821 *bad_wr = wr; 3822 nreq = 0; 3823 goto out; 3824 } 3825 3826 ind = qp->rq.head & (qp->rq.wqe_cnt - 1); 3827 3828 for (nreq = 0; wr; ++nreq, wr = wr->next) { 3829 if (mlx4_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) { 3830 err = -ENOMEM; 3831 *bad_wr = wr; 3832 goto out; 3833 } 3834 3835 if (unlikely(wr->num_sge > qp->rq.max_gs)) { 3836 err = -EINVAL; 3837 *bad_wr = wr; 3838 goto out; 3839 } 3840 3841 scat = get_recv_wqe(qp, ind); 3842 3843 if (qp->mlx4_ib_qp_type & (MLX4_IB_QPT_PROXY_SMI_OWNER | 3844 MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI)) { 3845 ib_dma_sync_single_for_device(ibqp->device, 3846 qp->sqp_proxy_rcv[ind].map, 3847 sizeof (struct mlx4_ib_proxy_sqp_hdr), 3848 DMA_FROM_DEVICE); 3849 scat->byte_count = 3850 cpu_to_be32(sizeof (struct mlx4_ib_proxy_sqp_hdr)); 3851 /* use dma lkey from upper layer entry */ 3852 scat->lkey = cpu_to_be32(wr->sg_list->lkey); 3853 scat->addr = cpu_to_be64(qp->sqp_proxy_rcv[ind].map); 3854 scat++; 3855 max_gs--; 3856 } 3857 3858 for (i = 0; i < wr->num_sge; ++i) 3859 __set_data_seg(scat + i, wr->sg_list + i); 3860 3861 if (i < max_gs) { 3862 scat[i].byte_count = 0; 3863 scat[i].lkey = cpu_to_be32(MLX4_INVALID_LKEY); 3864 scat[i].addr = 0; 3865 } 3866 3867 qp->rq.wrid[ind] = wr->wr_id; 3868 3869 ind = (ind + 1) & (qp->rq.wqe_cnt - 1); 3870 } 3871 3872 out: 3873 if (likely(nreq)) { 3874 qp->rq.head += nreq; 3875 3876 /* 3877 * Make sure that descriptors are written before 3878 * doorbell record. 3879 */ 3880 wmb(); 3881 3882 *qp->db.db = cpu_to_be32(qp->rq.head & 0xffff); 3883 } 3884 3885 spin_unlock_irqrestore(&qp->rq.lock, flags); 3886 3887 return err; 3888 } 3889 3890 int mlx4_ib_post_recv(struct ib_qp *ibqp, const struct ib_recv_wr *wr, 3891 const struct ib_recv_wr **bad_wr) 3892 { 3893 return _mlx4_ib_post_recv(ibqp, wr, bad_wr, false); 3894 } 3895 3896 static inline enum ib_qp_state to_ib_qp_state(enum mlx4_qp_state mlx4_state) 3897 { 3898 switch (mlx4_state) { 3899 case MLX4_QP_STATE_RST: return IB_QPS_RESET; 3900 case MLX4_QP_STATE_INIT: return IB_QPS_INIT; 3901 case MLX4_QP_STATE_RTR: return IB_QPS_RTR; 3902 case MLX4_QP_STATE_RTS: return IB_QPS_RTS; 3903 case MLX4_QP_STATE_SQ_DRAINING: 3904 case MLX4_QP_STATE_SQD: return IB_QPS_SQD; 3905 case MLX4_QP_STATE_SQER: return IB_QPS_SQE; 3906 case MLX4_QP_STATE_ERR: return IB_QPS_ERR; 3907 default: return -1; 3908 } 3909 } 3910 3911 static inline enum ib_mig_state to_ib_mig_state(int mlx4_mig_state) 3912 { 3913 switch (mlx4_mig_state) { 3914 case MLX4_QP_PM_ARMED: return IB_MIG_ARMED; 3915 case MLX4_QP_PM_REARM: return IB_MIG_REARM; 3916 case MLX4_QP_PM_MIGRATED: return IB_MIG_MIGRATED; 3917 default: return -1; 3918 } 3919 } 3920 3921 static int to_ib_qp_access_flags(int mlx4_flags) 3922 { 3923 int ib_flags = 0; 3924 3925 if (mlx4_flags & MLX4_QP_BIT_RRE) 3926 ib_flags |= IB_ACCESS_REMOTE_READ; 3927 if (mlx4_flags & MLX4_QP_BIT_RWE) 3928 ib_flags |= IB_ACCESS_REMOTE_WRITE; 3929 if (mlx4_flags & MLX4_QP_BIT_RAE) 3930 ib_flags |= IB_ACCESS_REMOTE_ATOMIC; 3931 3932 return ib_flags; 3933 } 3934 3935 static void to_rdma_ah_attr(struct mlx4_ib_dev *ibdev, 3936 struct rdma_ah_attr *ah_attr, 3937 struct mlx4_qp_path *path) 3938 { 3939 struct mlx4_dev *dev = ibdev->dev; 3940 u8 port_num = path->sched_queue & 0x40 ? 2 : 1; 3941 3942 memset(ah_attr, 0, sizeof(*ah_attr)); 3943 if (port_num == 0 || port_num > dev->caps.num_ports) 3944 return; 3945 ah_attr->type = rdma_ah_find_type(&ibdev->ib_dev, port_num); 3946 3947 if (ah_attr->type == RDMA_AH_ATTR_TYPE_ROCE) 3948 rdma_ah_set_sl(ah_attr, ((path->sched_queue >> 3) & 0x7) | 3949 ((path->sched_queue & 4) << 1)); 3950 else 3951 rdma_ah_set_sl(ah_attr, (path->sched_queue >> 2) & 0xf); 3952 rdma_ah_set_port_num(ah_attr, port_num); 3953 3954 rdma_ah_set_dlid(ah_attr, be16_to_cpu(path->rlid)); 3955 rdma_ah_set_path_bits(ah_attr, path->grh_mylmc & 0x7f); 3956 rdma_ah_set_static_rate(ah_attr, 3957 path->static_rate ? path->static_rate - 5 : 0); 3958 if (path->grh_mylmc & (1 << 7)) { 3959 rdma_ah_set_grh(ah_attr, NULL, 3960 be32_to_cpu(path->tclass_flowlabel) & 0xfffff, 3961 path->mgid_index, 3962 path->hop_limit, 3963 (be32_to_cpu(path->tclass_flowlabel) 3964 >> 20) & 0xff); 3965 rdma_ah_set_dgid_raw(ah_attr, path->rgid); 3966 } 3967 } 3968 3969 int mlx4_ib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr, int qp_attr_mask, 3970 struct ib_qp_init_attr *qp_init_attr) 3971 { 3972 struct mlx4_ib_dev *dev = to_mdev(ibqp->device); 3973 struct mlx4_ib_qp *qp = to_mqp(ibqp); 3974 struct mlx4_qp_context context; 3975 int mlx4_state; 3976 int err = 0; 3977 3978 if (ibqp->rwq_ind_tbl) 3979 return -EOPNOTSUPP; 3980 3981 mutex_lock(&qp->mutex); 3982 3983 if (qp->state == IB_QPS_RESET) { 3984 qp_attr->qp_state = IB_QPS_RESET; 3985 goto done; 3986 } 3987 3988 err = mlx4_qp_query(dev->dev, &qp->mqp, &context); 3989 if (err) { 3990 err = -EINVAL; 3991 goto out; 3992 } 3993 3994 mlx4_state = be32_to_cpu(context.flags) >> 28; 3995 3996 qp->state = to_ib_qp_state(mlx4_state); 3997 qp_attr->qp_state = qp->state; 3998 qp_attr->path_mtu = context.mtu_msgmax >> 5; 3999 qp_attr->path_mig_state = 4000 to_ib_mig_state((be32_to_cpu(context.flags) >> 11) & 0x3); 4001 qp_attr->qkey = be32_to_cpu(context.qkey); 4002 qp_attr->rq_psn = be32_to_cpu(context.rnr_nextrecvpsn) & 0xffffff; 4003 qp_attr->sq_psn = be32_to_cpu(context.next_send_psn) & 0xffffff; 4004 qp_attr->dest_qp_num = be32_to_cpu(context.remote_qpn) & 0xffffff; 4005 qp_attr->qp_access_flags = 4006 to_ib_qp_access_flags(be32_to_cpu(context.params2)); 4007 4008 if (qp->ibqp.qp_type == IB_QPT_RC || qp->ibqp.qp_type == IB_QPT_UC || 4009 qp->ibqp.qp_type == IB_QPT_XRC_INI || 4010 qp->ibqp.qp_type == IB_QPT_XRC_TGT) { 4011 to_rdma_ah_attr(dev, &qp_attr->ah_attr, &context.pri_path); 4012 to_rdma_ah_attr(dev, &qp_attr->alt_ah_attr, &context.alt_path); 4013 qp_attr->alt_pkey_index = context.alt_path.pkey_index & 0x7f; 4014 qp_attr->alt_port_num = 4015 rdma_ah_get_port_num(&qp_attr->alt_ah_attr); 4016 } 4017 4018 qp_attr->pkey_index = context.pri_path.pkey_index & 0x7f; 4019 if (qp_attr->qp_state == IB_QPS_INIT) 4020 qp_attr->port_num = qp->port; 4021 else 4022 qp_attr->port_num = context.pri_path.sched_queue & 0x40 ? 2 : 1; 4023 4024 /* qp_attr->en_sqd_async_notify is only applicable in modify qp */ 4025 qp_attr->sq_draining = mlx4_state == MLX4_QP_STATE_SQ_DRAINING; 4026 4027 qp_attr->max_rd_atomic = 1 << ((be32_to_cpu(context.params1) >> 21) & 0x7); 4028 4029 qp_attr->max_dest_rd_atomic = 4030 1 << ((be32_to_cpu(context.params2) >> 21) & 0x7); 4031 qp_attr->min_rnr_timer = 4032 (be32_to_cpu(context.rnr_nextrecvpsn) >> 24) & 0x1f; 4033 qp_attr->timeout = context.pri_path.ackto >> 3; 4034 qp_attr->retry_cnt = (be32_to_cpu(context.params1) >> 16) & 0x7; 4035 qp_attr->rnr_retry = (be32_to_cpu(context.params1) >> 13) & 0x7; 4036 qp_attr->alt_timeout = context.alt_path.ackto >> 3; 4037 4038 done: 4039 qp_attr->cur_qp_state = qp_attr->qp_state; 4040 qp_attr->cap.max_recv_wr = qp->rq.wqe_cnt; 4041 qp_attr->cap.max_recv_sge = qp->rq.max_gs; 4042 4043 if (!ibqp->uobject) { 4044 qp_attr->cap.max_send_wr = qp->sq.wqe_cnt; 4045 qp_attr->cap.max_send_sge = qp->sq.max_gs; 4046 } else { 4047 qp_attr->cap.max_send_wr = 0; 4048 qp_attr->cap.max_send_sge = 0; 4049 } 4050 4051 /* 4052 * We don't support inline sends for kernel QPs (yet), and we 4053 * don't know what userspace's value should be. 4054 */ 4055 qp_attr->cap.max_inline_data = 0; 4056 4057 qp_init_attr->cap = qp_attr->cap; 4058 4059 qp_init_attr->create_flags = 0; 4060 if (qp->flags & MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK) 4061 qp_init_attr->create_flags |= IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK; 4062 4063 if (qp->flags & MLX4_IB_QP_LSO) 4064 qp_init_attr->create_flags |= IB_QP_CREATE_IPOIB_UD_LSO; 4065 4066 if (qp->flags & MLX4_IB_QP_NETIF) 4067 qp_init_attr->create_flags |= IB_QP_CREATE_NETIF_QP; 4068 4069 qp_init_attr->sq_sig_type = 4070 qp->sq_signal_bits == cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE) ? 4071 IB_SIGNAL_ALL_WR : IB_SIGNAL_REQ_WR; 4072 4073 out: 4074 mutex_unlock(&qp->mutex); 4075 return err; 4076 } 4077 4078 struct ib_wq *mlx4_ib_create_wq(struct ib_pd *pd, 4079 struct ib_wq_init_attr *init_attr, 4080 struct ib_udata *udata) 4081 { 4082 struct mlx4_dev *dev = to_mdev(pd->device)->dev; 4083 struct ib_qp_init_attr ib_qp_init_attr = {}; 4084 struct mlx4_ib_qp *qp; 4085 struct mlx4_ib_create_wq ucmd; 4086 int err, required_cmd_sz; 4087 4088 if (!udata) 4089 return ERR_PTR(-EINVAL); 4090 4091 required_cmd_sz = offsetof(typeof(ucmd), comp_mask) + 4092 sizeof(ucmd.comp_mask); 4093 if (udata->inlen < required_cmd_sz) { 4094 pr_debug("invalid inlen\n"); 4095 return ERR_PTR(-EINVAL); 4096 } 4097 4098 if (udata->inlen > sizeof(ucmd) && 4099 !ib_is_udata_cleared(udata, sizeof(ucmd), 4100 udata->inlen - sizeof(ucmd))) { 4101 pr_debug("inlen is not supported\n"); 4102 return ERR_PTR(-EOPNOTSUPP); 4103 } 4104 4105 if (udata->outlen) 4106 return ERR_PTR(-EOPNOTSUPP); 4107 4108 if (init_attr->wq_type != IB_WQT_RQ) { 4109 pr_debug("unsupported wq type %d\n", init_attr->wq_type); 4110 return ERR_PTR(-EOPNOTSUPP); 4111 } 4112 4113 if (init_attr->create_flags & ~IB_WQ_FLAGS_SCATTER_FCS || 4114 !(dev->caps.flags & MLX4_DEV_CAP_FLAG_FCS_KEEP)) { 4115 pr_debug("unsupported create_flags %u\n", 4116 init_attr->create_flags); 4117 return ERR_PTR(-EOPNOTSUPP); 4118 } 4119 4120 qp = kzalloc(sizeof(*qp), GFP_KERNEL); 4121 if (!qp) 4122 return ERR_PTR(-ENOMEM); 4123 4124 mutex_init(&qp->mutex); 4125 qp->pri.vid = 0xFFFF; 4126 qp->alt.vid = 0xFFFF; 4127 4128 ib_qp_init_attr.qp_context = init_attr->wq_context; 4129 ib_qp_init_attr.qp_type = IB_QPT_RAW_PACKET; 4130 ib_qp_init_attr.cap.max_recv_wr = init_attr->max_wr; 4131 ib_qp_init_attr.cap.max_recv_sge = init_attr->max_sge; 4132 ib_qp_init_attr.recv_cq = init_attr->cq; 4133 ib_qp_init_attr.send_cq = ib_qp_init_attr.recv_cq; /* Dummy CQ */ 4134 4135 if (init_attr->create_flags & IB_WQ_FLAGS_SCATTER_FCS) 4136 ib_qp_init_attr.create_flags |= IB_QP_CREATE_SCATTER_FCS; 4137 4138 err = create_rq(pd, &ib_qp_init_attr, udata, qp); 4139 if (err) { 4140 kfree(qp); 4141 return ERR_PTR(err); 4142 } 4143 4144 qp->ibwq.event_handler = init_attr->event_handler; 4145 qp->ibwq.wq_num = qp->mqp.qpn; 4146 qp->ibwq.state = IB_WQS_RESET; 4147 4148 return &qp->ibwq; 4149 } 4150 4151 static int ib_wq2qp_state(enum ib_wq_state state) 4152 { 4153 switch (state) { 4154 case IB_WQS_RESET: 4155 return IB_QPS_RESET; 4156 case IB_WQS_RDY: 4157 return IB_QPS_RTR; 4158 default: 4159 return IB_QPS_ERR; 4160 } 4161 } 4162 4163 static int _mlx4_ib_modify_wq(struct ib_wq *ibwq, enum ib_wq_state new_state, 4164 struct ib_udata *udata) 4165 { 4166 struct mlx4_ib_qp *qp = to_mqp((struct ib_qp *)ibwq); 4167 enum ib_qp_state qp_cur_state; 4168 enum ib_qp_state qp_new_state; 4169 int attr_mask; 4170 int err; 4171 4172 /* ib_qp.state represents the WQ HW state while ib_wq.state represents 4173 * the WQ logic state. 4174 */ 4175 qp_cur_state = qp->state; 4176 qp_new_state = ib_wq2qp_state(new_state); 4177 4178 if (ib_wq2qp_state(new_state) == qp_cur_state) 4179 return 0; 4180 4181 if (new_state == IB_WQS_RDY) { 4182 struct ib_qp_attr attr = {}; 4183 4184 attr.port_num = qp->port; 4185 attr_mask = IB_QP_PORT; 4186 4187 err = __mlx4_ib_modify_qp(ibwq, MLX4_IB_RWQ_SRC, &attr, 4188 attr_mask, IB_QPS_RESET, IB_QPS_INIT, 4189 udata); 4190 if (err) { 4191 pr_debug("WQN=0x%06x failed to apply RST->INIT on the HW QP\n", 4192 ibwq->wq_num); 4193 return err; 4194 } 4195 4196 qp_cur_state = IB_QPS_INIT; 4197 } 4198 4199 attr_mask = 0; 4200 err = __mlx4_ib_modify_qp(ibwq, MLX4_IB_RWQ_SRC, NULL, attr_mask, 4201 qp_cur_state, qp_new_state, udata); 4202 4203 if (err && (qp_cur_state == IB_QPS_INIT)) { 4204 qp_new_state = IB_QPS_RESET; 4205 if (__mlx4_ib_modify_qp(ibwq, MLX4_IB_RWQ_SRC, NULL, 4206 attr_mask, IB_QPS_INIT, IB_QPS_RESET, 4207 udata)) { 4208 pr_warn("WQN=0x%06x failed with reverting HW's resources failure\n", 4209 ibwq->wq_num); 4210 qp_new_state = IB_QPS_INIT; 4211 } 4212 } 4213 4214 qp->state = qp_new_state; 4215 4216 return err; 4217 } 4218 4219 int mlx4_ib_modify_wq(struct ib_wq *ibwq, struct ib_wq_attr *wq_attr, 4220 u32 wq_attr_mask, struct ib_udata *udata) 4221 { 4222 struct mlx4_ib_qp *qp = to_mqp((struct ib_qp *)ibwq); 4223 struct mlx4_ib_modify_wq ucmd = {}; 4224 size_t required_cmd_sz; 4225 enum ib_wq_state cur_state, new_state; 4226 int err = 0; 4227 4228 required_cmd_sz = offsetof(typeof(ucmd), reserved) + 4229 sizeof(ucmd.reserved); 4230 if (udata->inlen < required_cmd_sz) 4231 return -EINVAL; 4232 4233 if (udata->inlen > sizeof(ucmd) && 4234 !ib_is_udata_cleared(udata, sizeof(ucmd), 4235 udata->inlen - sizeof(ucmd))) 4236 return -EOPNOTSUPP; 4237 4238 if (ib_copy_from_udata(&ucmd, udata, min(sizeof(ucmd), udata->inlen))) 4239 return -EFAULT; 4240 4241 if (ucmd.comp_mask || ucmd.reserved) 4242 return -EOPNOTSUPP; 4243 4244 if (wq_attr_mask & IB_WQ_FLAGS) 4245 return -EOPNOTSUPP; 4246 4247 cur_state = wq_attr->curr_wq_state; 4248 new_state = wq_attr->wq_state; 4249 4250 if ((new_state == IB_WQS_RDY) && (cur_state == IB_WQS_ERR)) 4251 return -EINVAL; 4252 4253 if ((new_state == IB_WQS_ERR) && (cur_state == IB_WQS_RESET)) 4254 return -EINVAL; 4255 4256 /* Need to protect against the parent RSS which also may modify WQ 4257 * state. 4258 */ 4259 mutex_lock(&qp->mutex); 4260 4261 /* Can update HW state only if a RSS QP has already associated to this 4262 * WQ, so we can apply its port on the WQ. 4263 */ 4264 if (qp->rss_usecnt) 4265 err = _mlx4_ib_modify_wq(ibwq, new_state, udata); 4266 4267 if (!err) 4268 ibwq->state = new_state; 4269 4270 mutex_unlock(&qp->mutex); 4271 4272 return err; 4273 } 4274 4275 int mlx4_ib_destroy_wq(struct ib_wq *ibwq, struct ib_udata *udata) 4276 { 4277 struct mlx4_ib_dev *dev = to_mdev(ibwq->device); 4278 struct mlx4_ib_qp *qp = to_mqp((struct ib_qp *)ibwq); 4279 4280 if (qp->counter_index) 4281 mlx4_ib_free_qp_counter(dev, qp); 4282 4283 destroy_qp_common(dev, qp, MLX4_IB_RWQ_SRC, udata); 4284 4285 kfree(qp); 4286 return 0; 4287 } 4288 4289 int mlx4_ib_create_rwq_ind_table(struct ib_rwq_ind_table *rwq_ind_table, 4290 struct ib_rwq_ind_table_init_attr *init_attr, 4291 struct ib_udata *udata) 4292 { 4293 struct mlx4_ib_create_rwq_ind_tbl_resp resp = {}; 4294 unsigned int ind_tbl_size = 1 << init_attr->log_ind_tbl_size; 4295 struct ib_device *device = rwq_ind_table->device; 4296 unsigned int base_wqn; 4297 size_t min_resp_len; 4298 int i, err = 0; 4299 4300 if (udata->inlen > 0 && 4301 !ib_is_udata_cleared(udata, 0, 4302 udata->inlen)) 4303 return -EOPNOTSUPP; 4304 4305 min_resp_len = offsetof(typeof(resp), reserved) + sizeof(resp.reserved); 4306 if (udata->outlen && udata->outlen < min_resp_len) 4307 return -EINVAL; 4308 4309 if (ind_tbl_size > 4310 device->attrs.rss_caps.max_rwq_indirection_table_size) { 4311 pr_debug("log_ind_tbl_size = %d is bigger than supported = %d\n", 4312 ind_tbl_size, 4313 device->attrs.rss_caps.max_rwq_indirection_table_size); 4314 return -EINVAL; 4315 } 4316 4317 base_wqn = init_attr->ind_tbl[0]->wq_num; 4318 4319 if (base_wqn % ind_tbl_size) { 4320 pr_debug("WQN=0x%x isn't aligned with indirection table size\n", 4321 base_wqn); 4322 return -EINVAL; 4323 } 4324 4325 for (i = 1; i < ind_tbl_size; i++) { 4326 if (++base_wqn != init_attr->ind_tbl[i]->wq_num) { 4327 pr_debug("indirection table's WQNs aren't consecutive\n"); 4328 return -EINVAL; 4329 } 4330 } 4331 4332 if (udata->outlen) { 4333 resp.response_length = offsetof(typeof(resp), response_length) + 4334 sizeof(resp.response_length); 4335 err = ib_copy_to_udata(udata, &resp, resp.response_length); 4336 } 4337 4338 return err; 4339 } 4340 4341 struct mlx4_ib_drain_cqe { 4342 struct ib_cqe cqe; 4343 struct completion done; 4344 }; 4345 4346 static void mlx4_ib_drain_qp_done(struct ib_cq *cq, struct ib_wc *wc) 4347 { 4348 struct mlx4_ib_drain_cqe *cqe = container_of(wc->wr_cqe, 4349 struct mlx4_ib_drain_cqe, 4350 cqe); 4351 4352 complete(&cqe->done); 4353 } 4354 4355 /* This function returns only once the drained WR was completed */ 4356 static void handle_drain_completion(struct ib_cq *cq, 4357 struct mlx4_ib_drain_cqe *sdrain, 4358 struct mlx4_ib_dev *dev) 4359 { 4360 struct mlx4_dev *mdev = dev->dev; 4361 4362 if (cq->poll_ctx == IB_POLL_DIRECT) { 4363 while (wait_for_completion_timeout(&sdrain->done, HZ / 10) <= 0) 4364 ib_process_cq_direct(cq, -1); 4365 return; 4366 } 4367 4368 if (mdev->persist->state == MLX4_DEVICE_STATE_INTERNAL_ERROR) { 4369 struct mlx4_ib_cq *mcq = to_mcq(cq); 4370 bool triggered = false; 4371 unsigned long flags; 4372 4373 spin_lock_irqsave(&dev->reset_flow_resource_lock, flags); 4374 /* Make sure that the CQ handler won't run if wasn't run yet */ 4375 if (!mcq->mcq.reset_notify_added) 4376 mcq->mcq.reset_notify_added = 1; 4377 else 4378 triggered = true; 4379 spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags); 4380 4381 if (triggered) { 4382 /* Wait for any scheduled/running task to be ended */ 4383 switch (cq->poll_ctx) { 4384 case IB_POLL_SOFTIRQ: 4385 irq_poll_disable(&cq->iop); 4386 irq_poll_enable(&cq->iop); 4387 break; 4388 case IB_POLL_WORKQUEUE: 4389 cancel_work_sync(&cq->work); 4390 break; 4391 default: 4392 WARN_ON_ONCE(1); 4393 } 4394 } 4395 4396 /* Run the CQ handler - this makes sure that the drain WR will 4397 * be processed if wasn't processed yet. 4398 */ 4399 mcq->mcq.comp(&mcq->mcq); 4400 } 4401 4402 wait_for_completion(&sdrain->done); 4403 } 4404 4405 void mlx4_ib_drain_sq(struct ib_qp *qp) 4406 { 4407 struct ib_cq *cq = qp->send_cq; 4408 struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR }; 4409 struct mlx4_ib_drain_cqe sdrain; 4410 const struct ib_send_wr *bad_swr; 4411 struct ib_rdma_wr swr = { 4412 .wr = { 4413 .next = NULL, 4414 { .wr_cqe = &sdrain.cqe, }, 4415 .opcode = IB_WR_RDMA_WRITE, 4416 }, 4417 }; 4418 int ret; 4419 struct mlx4_ib_dev *dev = to_mdev(qp->device); 4420 struct mlx4_dev *mdev = dev->dev; 4421 4422 ret = ib_modify_qp(qp, &attr, IB_QP_STATE); 4423 if (ret && mdev->persist->state != MLX4_DEVICE_STATE_INTERNAL_ERROR) { 4424 WARN_ONCE(ret, "failed to drain send queue: %d\n", ret); 4425 return; 4426 } 4427 4428 sdrain.cqe.done = mlx4_ib_drain_qp_done; 4429 init_completion(&sdrain.done); 4430 4431 ret = _mlx4_ib_post_send(qp, &swr.wr, &bad_swr, true); 4432 if (ret) { 4433 WARN_ONCE(ret, "failed to drain send queue: %d\n", ret); 4434 return; 4435 } 4436 4437 handle_drain_completion(cq, &sdrain, dev); 4438 } 4439 4440 void mlx4_ib_drain_rq(struct ib_qp *qp) 4441 { 4442 struct ib_cq *cq = qp->recv_cq; 4443 struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR }; 4444 struct mlx4_ib_drain_cqe rdrain; 4445 struct ib_recv_wr rwr = {}; 4446 const struct ib_recv_wr *bad_rwr; 4447 int ret; 4448 struct mlx4_ib_dev *dev = to_mdev(qp->device); 4449 struct mlx4_dev *mdev = dev->dev; 4450 4451 ret = ib_modify_qp(qp, &attr, IB_QP_STATE); 4452 if (ret && mdev->persist->state != MLX4_DEVICE_STATE_INTERNAL_ERROR) { 4453 WARN_ONCE(ret, "failed to drain recv queue: %d\n", ret); 4454 return; 4455 } 4456 4457 rwr.wr_cqe = &rdrain.cqe; 4458 rdrain.cqe.done = mlx4_ib_drain_qp_done; 4459 init_completion(&rdrain.done); 4460 4461 ret = _mlx4_ib_post_recv(qp, &rwr, &bad_rwr, true); 4462 if (ret) { 4463 WARN_ONCE(ret, "failed to drain recv queue: %d\n", ret); 4464 return; 4465 } 4466 4467 handle_drain_completion(cq, &rdrain, dev); 4468 } 4469