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/slab.h> 36 #include <linux/netdevice.h> 37 38 #include <rdma/ib_cache.h> 39 #include <rdma/ib_pack.h> 40 #include <rdma/ib_addr.h> 41 #include <rdma/ib_mad.h> 42 43 #include <linux/mlx4/qp.h> 44 45 #include "mlx4_ib.h" 46 #include "user.h" 47 48 enum { 49 MLX4_IB_ACK_REQ_FREQ = 8, 50 }; 51 52 enum { 53 MLX4_IB_DEFAULT_SCHED_QUEUE = 0x83, 54 MLX4_IB_DEFAULT_QP0_SCHED_QUEUE = 0x3f, 55 MLX4_IB_LINK_TYPE_IB = 0, 56 MLX4_IB_LINK_TYPE_ETH = 1 57 }; 58 59 enum { 60 /* 61 * Largest possible UD header: send with GRH and immediate 62 * data plus 18 bytes for an Ethernet header with VLAN/802.1Q 63 * tag. (LRH would only use 8 bytes, so Ethernet is the 64 * biggest case) 65 */ 66 MLX4_IB_UD_HEADER_SIZE = 82, 67 MLX4_IB_LSO_HEADER_SPARE = 128, 68 }; 69 70 enum { 71 MLX4_IB_IBOE_ETHERTYPE = 0x8915 72 }; 73 74 struct mlx4_ib_sqp { 75 struct mlx4_ib_qp qp; 76 int pkey_index; 77 u32 qkey; 78 u32 send_psn; 79 struct ib_ud_header ud_header; 80 u8 header_buf[MLX4_IB_UD_HEADER_SIZE]; 81 }; 82 83 enum { 84 MLX4_IB_MIN_SQ_STRIDE = 6, 85 MLX4_IB_CACHE_LINE_SIZE = 64, 86 }; 87 88 enum { 89 MLX4_RAW_QP_MTU = 7, 90 MLX4_RAW_QP_MSGMAX = 31, 91 }; 92 93 #ifndef ETH_ALEN 94 #define ETH_ALEN 6 95 #endif 96 static inline u64 mlx4_mac_to_u64(u8 *addr) 97 { 98 u64 mac = 0; 99 int i; 100 101 for (i = 0; i < ETH_ALEN; i++) { 102 mac <<= 8; 103 mac |= addr[i]; 104 } 105 return mac; 106 } 107 108 static const __be32 mlx4_ib_opcode[] = { 109 [IB_WR_SEND] = cpu_to_be32(MLX4_OPCODE_SEND), 110 [IB_WR_LSO] = cpu_to_be32(MLX4_OPCODE_LSO), 111 [IB_WR_SEND_WITH_IMM] = cpu_to_be32(MLX4_OPCODE_SEND_IMM), 112 [IB_WR_RDMA_WRITE] = cpu_to_be32(MLX4_OPCODE_RDMA_WRITE), 113 [IB_WR_RDMA_WRITE_WITH_IMM] = cpu_to_be32(MLX4_OPCODE_RDMA_WRITE_IMM), 114 [IB_WR_RDMA_READ] = cpu_to_be32(MLX4_OPCODE_RDMA_READ), 115 [IB_WR_ATOMIC_CMP_AND_SWP] = cpu_to_be32(MLX4_OPCODE_ATOMIC_CS), 116 [IB_WR_ATOMIC_FETCH_AND_ADD] = cpu_to_be32(MLX4_OPCODE_ATOMIC_FA), 117 [IB_WR_SEND_WITH_INV] = cpu_to_be32(MLX4_OPCODE_SEND_INVAL), 118 [IB_WR_LOCAL_INV] = cpu_to_be32(MLX4_OPCODE_LOCAL_INVAL), 119 [IB_WR_FAST_REG_MR] = cpu_to_be32(MLX4_OPCODE_FMR), 120 [IB_WR_MASKED_ATOMIC_CMP_AND_SWP] = cpu_to_be32(MLX4_OPCODE_MASKED_ATOMIC_CS), 121 [IB_WR_MASKED_ATOMIC_FETCH_AND_ADD] = cpu_to_be32(MLX4_OPCODE_MASKED_ATOMIC_FA), 122 [IB_WR_BIND_MW] = cpu_to_be32(MLX4_OPCODE_BIND_MW), 123 }; 124 125 static struct mlx4_ib_sqp *to_msqp(struct mlx4_ib_qp *mqp) 126 { 127 return container_of(mqp, struct mlx4_ib_sqp, qp); 128 } 129 130 static int is_tunnel_qp(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp) 131 { 132 if (!mlx4_is_master(dev->dev)) 133 return 0; 134 135 return qp->mqp.qpn >= dev->dev->phys_caps.base_tunnel_sqpn && 136 qp->mqp.qpn < dev->dev->phys_caps.base_tunnel_sqpn + 137 8 * MLX4_MFUNC_MAX; 138 } 139 140 static int is_sqp(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp) 141 { 142 int proxy_sqp = 0; 143 int real_sqp = 0; 144 int i; 145 /* PPF or Native -- real SQP */ 146 real_sqp = ((mlx4_is_master(dev->dev) || !mlx4_is_mfunc(dev->dev)) && 147 qp->mqp.qpn >= dev->dev->phys_caps.base_sqpn && 148 qp->mqp.qpn <= dev->dev->phys_caps.base_sqpn + 3); 149 if (real_sqp) 150 return 1; 151 /* VF or PF -- proxy SQP */ 152 if (mlx4_is_mfunc(dev->dev)) { 153 for (i = 0; i < dev->dev->caps.num_ports; i++) { 154 if (qp->mqp.qpn == dev->dev->caps.qp0_proxy[i] || 155 qp->mqp.qpn == dev->dev->caps.qp1_proxy[i]) { 156 proxy_sqp = 1; 157 break; 158 } 159 } 160 } 161 return proxy_sqp; 162 } 163 164 /* used for INIT/CLOSE port logic */ 165 static int is_qp0(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp) 166 { 167 int proxy_qp0 = 0; 168 int real_qp0 = 0; 169 int i; 170 /* PPF or Native -- real QP0 */ 171 real_qp0 = ((mlx4_is_master(dev->dev) || !mlx4_is_mfunc(dev->dev)) && 172 qp->mqp.qpn >= dev->dev->phys_caps.base_sqpn && 173 qp->mqp.qpn <= dev->dev->phys_caps.base_sqpn + 1); 174 if (real_qp0) 175 return 1; 176 /* VF or PF -- proxy QP0 */ 177 if (mlx4_is_mfunc(dev->dev)) { 178 for (i = 0; i < dev->dev->caps.num_ports; i++) { 179 if (qp->mqp.qpn == dev->dev->caps.qp0_proxy[i]) { 180 proxy_qp0 = 1; 181 break; 182 } 183 } 184 } 185 return proxy_qp0; 186 } 187 188 static void *get_wqe(struct mlx4_ib_qp *qp, int offset) 189 { 190 return mlx4_buf_offset(&qp->buf, offset); 191 } 192 193 static void *get_recv_wqe(struct mlx4_ib_qp *qp, int n) 194 { 195 return get_wqe(qp, qp->rq.offset + (n << qp->rq.wqe_shift)); 196 } 197 198 static void *get_send_wqe(struct mlx4_ib_qp *qp, int n) 199 { 200 return get_wqe(qp, qp->sq.offset + (n << qp->sq.wqe_shift)); 201 } 202 203 /* 204 * Stamp a SQ WQE so that it is invalid if prefetched by marking the 205 * first four bytes of every 64 byte chunk with 206 * 0x7FFFFFF | (invalid_ownership_value << 31). 207 * 208 * When the max work request size is less than or equal to the WQE 209 * basic block size, as an optimization, we can stamp all WQEs with 210 * 0xffffffff, and skip the very first chunk of each WQE. 211 */ 212 static void stamp_send_wqe(struct mlx4_ib_qp *qp, int n, int size) 213 { 214 __be32 *wqe; 215 int i; 216 int s; 217 int ind; 218 void *buf; 219 __be32 stamp; 220 struct mlx4_wqe_ctrl_seg *ctrl; 221 222 if (qp->sq_max_wqes_per_wr > 1) { 223 s = roundup(size, 1U << qp->sq.wqe_shift); 224 for (i = 0; i < s; i += 64) { 225 ind = (i >> qp->sq.wqe_shift) + n; 226 stamp = ind & qp->sq.wqe_cnt ? cpu_to_be32(0x7fffffff) : 227 cpu_to_be32(0xffffffff); 228 buf = get_send_wqe(qp, ind & (qp->sq.wqe_cnt - 1)); 229 wqe = buf + (i & ((1 << qp->sq.wqe_shift) - 1)); 230 *wqe = stamp; 231 } 232 } else { 233 ctrl = buf = get_send_wqe(qp, n & (qp->sq.wqe_cnt - 1)); 234 s = (ctrl->fence_size & 0x3f) << 4; 235 for (i = 64; i < s; i += 64) { 236 wqe = buf + i; 237 *wqe = cpu_to_be32(0xffffffff); 238 } 239 } 240 } 241 242 static void post_nop_wqe(struct mlx4_ib_qp *qp, int n, int size) 243 { 244 struct mlx4_wqe_ctrl_seg *ctrl; 245 struct mlx4_wqe_inline_seg *inl; 246 void *wqe; 247 int s; 248 249 ctrl = wqe = get_send_wqe(qp, n & (qp->sq.wqe_cnt - 1)); 250 s = sizeof(struct mlx4_wqe_ctrl_seg); 251 252 if (qp->ibqp.qp_type == IB_QPT_UD) { 253 struct mlx4_wqe_datagram_seg *dgram = wqe + sizeof *ctrl; 254 struct mlx4_av *av = (struct mlx4_av *)dgram->av; 255 memset(dgram, 0, sizeof *dgram); 256 av->port_pd = cpu_to_be32((qp->port << 24) | to_mpd(qp->ibqp.pd)->pdn); 257 s += sizeof(struct mlx4_wqe_datagram_seg); 258 } 259 260 /* Pad the remainder of the WQE with an inline data segment. */ 261 if (size > s) { 262 inl = wqe + s; 263 inl->byte_count = cpu_to_be32(1 << 31 | (size - s - sizeof *inl)); 264 } 265 ctrl->srcrb_flags = 0; 266 ctrl->fence_size = size / 16; 267 /* 268 * Make sure descriptor is fully written before setting ownership bit 269 * (because HW can start executing as soon as we do). 270 */ 271 wmb(); 272 273 ctrl->owner_opcode = cpu_to_be32(MLX4_OPCODE_NOP | MLX4_WQE_CTRL_NEC) | 274 (n & qp->sq.wqe_cnt ? cpu_to_be32(1 << 31) : 0); 275 276 stamp_send_wqe(qp, n + qp->sq_spare_wqes, size); 277 } 278 279 /* Post NOP WQE to prevent wrap-around in the middle of WR */ 280 static inline unsigned pad_wraparound(struct mlx4_ib_qp *qp, int ind) 281 { 282 unsigned s = qp->sq.wqe_cnt - (ind & (qp->sq.wqe_cnt - 1)); 283 if (unlikely(s < qp->sq_max_wqes_per_wr)) { 284 post_nop_wqe(qp, ind, s << qp->sq.wqe_shift); 285 ind += s; 286 } 287 return ind; 288 } 289 290 static void mlx4_ib_qp_event(struct mlx4_qp *qp, enum mlx4_event type) 291 { 292 struct ib_event event; 293 struct ib_qp *ibqp = &to_mibqp(qp)->ibqp; 294 295 if (type == MLX4_EVENT_TYPE_PATH_MIG) 296 to_mibqp(qp)->port = to_mibqp(qp)->alt_port; 297 298 if (ibqp->event_handler) { 299 event.device = ibqp->device; 300 event.element.qp = ibqp; 301 switch (type) { 302 case MLX4_EVENT_TYPE_PATH_MIG: 303 event.event = IB_EVENT_PATH_MIG; 304 break; 305 case MLX4_EVENT_TYPE_COMM_EST: 306 event.event = IB_EVENT_COMM_EST; 307 break; 308 case MLX4_EVENT_TYPE_SQ_DRAINED: 309 event.event = IB_EVENT_SQ_DRAINED; 310 break; 311 case MLX4_EVENT_TYPE_SRQ_QP_LAST_WQE: 312 event.event = IB_EVENT_QP_LAST_WQE_REACHED; 313 break; 314 case MLX4_EVENT_TYPE_WQ_CATAS_ERROR: 315 event.event = IB_EVENT_QP_FATAL; 316 break; 317 case MLX4_EVENT_TYPE_PATH_MIG_FAILED: 318 event.event = IB_EVENT_PATH_MIG_ERR; 319 break; 320 case MLX4_EVENT_TYPE_WQ_INVAL_REQ_ERROR: 321 event.event = IB_EVENT_QP_REQ_ERR; 322 break; 323 case MLX4_EVENT_TYPE_WQ_ACCESS_ERROR: 324 event.event = IB_EVENT_QP_ACCESS_ERR; 325 break; 326 default: 327 pr_warn("Unexpected event type %d " 328 "on QP %06x\n", type, qp->qpn); 329 return; 330 } 331 332 ibqp->event_handler(&event, ibqp->qp_context); 333 } 334 } 335 336 static int send_wqe_overhead(enum mlx4_ib_qp_type type, u32 flags) 337 { 338 /* 339 * UD WQEs must have a datagram segment. 340 * RC and UC WQEs might have a remote address segment. 341 * MLX WQEs need two extra inline data segments (for the UD 342 * header and space for the ICRC). 343 */ 344 switch (type) { 345 case MLX4_IB_QPT_UD: 346 return sizeof (struct mlx4_wqe_ctrl_seg) + 347 sizeof (struct mlx4_wqe_datagram_seg) + 348 ((flags & MLX4_IB_QP_LSO) ? MLX4_IB_LSO_HEADER_SPARE : 0); 349 case MLX4_IB_QPT_PROXY_SMI_OWNER: 350 case MLX4_IB_QPT_PROXY_SMI: 351 case MLX4_IB_QPT_PROXY_GSI: 352 return sizeof (struct mlx4_wqe_ctrl_seg) + 353 sizeof (struct mlx4_wqe_datagram_seg) + 64; 354 case MLX4_IB_QPT_TUN_SMI_OWNER: 355 case MLX4_IB_QPT_TUN_GSI: 356 return sizeof (struct mlx4_wqe_ctrl_seg) + 357 sizeof (struct mlx4_wqe_datagram_seg); 358 359 case MLX4_IB_QPT_UC: 360 return sizeof (struct mlx4_wqe_ctrl_seg) + 361 sizeof (struct mlx4_wqe_raddr_seg); 362 case MLX4_IB_QPT_RC: 363 return sizeof (struct mlx4_wqe_ctrl_seg) + 364 sizeof (struct mlx4_wqe_atomic_seg) + 365 sizeof (struct mlx4_wqe_raddr_seg); 366 case MLX4_IB_QPT_SMI: 367 case MLX4_IB_QPT_GSI: 368 return sizeof (struct mlx4_wqe_ctrl_seg) + 369 ALIGN(MLX4_IB_UD_HEADER_SIZE + 370 DIV_ROUND_UP(MLX4_IB_UD_HEADER_SIZE, 371 MLX4_INLINE_ALIGN) * 372 sizeof (struct mlx4_wqe_inline_seg), 373 sizeof (struct mlx4_wqe_data_seg)) + 374 ALIGN(4 + 375 sizeof (struct mlx4_wqe_inline_seg), 376 sizeof (struct mlx4_wqe_data_seg)); 377 default: 378 return sizeof (struct mlx4_wqe_ctrl_seg); 379 } 380 } 381 382 static int set_rq_size(struct mlx4_ib_dev *dev, struct ib_qp_cap *cap, 383 int is_user, int has_rq, struct mlx4_ib_qp *qp) 384 { 385 /* Sanity check RQ size before proceeding */ 386 if (cap->max_recv_wr > dev->dev->caps.max_wqes - MLX4_IB_SQ_MAX_SPARE || 387 cap->max_recv_sge > min(dev->dev->caps.max_sq_sg, dev->dev->caps.max_rq_sg)) 388 return -EINVAL; 389 390 if (!has_rq) { 391 if (cap->max_recv_wr) 392 return -EINVAL; 393 394 qp->rq.wqe_cnt = qp->rq.max_gs = 0; 395 } else { 396 /* HW requires >= 1 RQ entry with >= 1 gather entry */ 397 if (is_user && (!cap->max_recv_wr || !cap->max_recv_sge)) 398 return -EINVAL; 399 400 qp->rq.wqe_cnt = roundup_pow_of_two(max(1U, cap->max_recv_wr)); 401 qp->rq.max_gs = roundup_pow_of_two(max(1U, cap->max_recv_sge)); 402 qp->rq.wqe_shift = ilog2(qp->rq.max_gs * sizeof (struct mlx4_wqe_data_seg)); 403 } 404 405 /* leave userspace return values as they were, so as not to break ABI */ 406 if (is_user) { 407 cap->max_recv_wr = qp->rq.max_post = qp->rq.wqe_cnt; 408 cap->max_recv_sge = qp->rq.max_gs; 409 } else { 410 cap->max_recv_wr = qp->rq.max_post = 411 min(dev->dev->caps.max_wqes - MLX4_IB_SQ_MAX_SPARE, qp->rq.wqe_cnt); 412 cap->max_recv_sge = min(qp->rq.max_gs, 413 min(dev->dev->caps.max_sq_sg, 414 dev->dev->caps.max_rq_sg)); 415 } 416 417 return 0; 418 } 419 420 static int set_kernel_sq_size(struct mlx4_ib_dev *dev, struct ib_qp_cap *cap, 421 enum mlx4_ib_qp_type type, struct mlx4_ib_qp *qp) 422 { 423 int s; 424 425 /* Sanity check SQ size before proceeding */ 426 if (cap->max_send_wr > (dev->dev->caps.max_wqes - MLX4_IB_SQ_MAX_SPARE) || 427 cap->max_send_sge > min(dev->dev->caps.max_sq_sg, dev->dev->caps.max_rq_sg) || 428 cap->max_inline_data + send_wqe_overhead(type, qp->flags) + 429 sizeof (struct mlx4_wqe_inline_seg) > dev->dev->caps.max_sq_desc_sz) 430 return -EINVAL; 431 432 /* 433 * For MLX transport we need 2 extra S/G entries: 434 * one for the header and one for the checksum at the end 435 */ 436 if ((type == MLX4_IB_QPT_SMI || type == MLX4_IB_QPT_GSI || 437 type & (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER)) && 438 cap->max_send_sge + 2 > dev->dev->caps.max_sq_sg) 439 return -EINVAL; 440 441 s = max(cap->max_send_sge * sizeof (struct mlx4_wqe_data_seg), 442 cap->max_inline_data + sizeof (struct mlx4_wqe_inline_seg)) + 443 send_wqe_overhead(type, qp->flags); 444 445 if (s > dev->dev->caps.max_sq_desc_sz) 446 return -EINVAL; 447 448 /* 449 * Hermon supports shrinking WQEs, such that a single work 450 * request can include multiple units of 1 << wqe_shift. This 451 * way, work requests can differ in size, and do not have to 452 * be a power of 2 in size, saving memory and speeding up send 453 * WR posting. Unfortunately, if we do this then the 454 * wqe_index field in CQEs can't be used to look up the WR ID 455 * anymore, so we do this only if selective signaling is off. 456 * 457 * Further, on 32-bit platforms, we can't use vmap() to make 458 * the QP buffer virtually contiguous. Thus we have to use 459 * constant-sized WRs to make sure a WR is always fully within 460 * a single page-sized chunk. 461 * 462 * Finally, we use NOP work requests to pad the end of the 463 * work queue, to avoid wrap-around in the middle of WR. We 464 * set NEC bit to avoid getting completions with error for 465 * these NOP WRs, but since NEC is only supported starting 466 * with firmware 2.2.232, we use constant-sized WRs for older 467 * firmware. 468 * 469 * And, since MLX QPs only support SEND, we use constant-sized 470 * WRs in this case. 471 * 472 * We look for the smallest value of wqe_shift such that the 473 * resulting number of wqes does not exceed device 474 * capabilities. 475 * 476 * We set WQE size to at least 64 bytes, this way stamping 477 * invalidates each WQE. 478 */ 479 if (dev->dev->caps.fw_ver >= MLX4_FW_VER_WQE_CTRL_NEC && 480 qp->sq_signal_bits && BITS_PER_LONG == 64 && 481 type != MLX4_IB_QPT_SMI && type != MLX4_IB_QPT_GSI && 482 !(type & (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_PROXY_SMI | 483 MLX4_IB_QPT_PROXY_GSI | MLX4_IB_QPT_TUN_SMI_OWNER))) 484 qp->sq.wqe_shift = ilog2(64); 485 else 486 qp->sq.wqe_shift = ilog2(roundup_pow_of_two(s)); 487 488 for (;;) { 489 qp->sq_max_wqes_per_wr = DIV_ROUND_UP(s, 1U << qp->sq.wqe_shift); 490 491 /* 492 * We need to leave 2 KB + 1 WR of headroom in the SQ to 493 * allow HW to prefetch. 494 */ 495 qp->sq_spare_wqes = (2048 >> qp->sq.wqe_shift) + qp->sq_max_wqes_per_wr; 496 qp->sq.wqe_cnt = roundup_pow_of_two(cap->max_send_wr * 497 qp->sq_max_wqes_per_wr + 498 qp->sq_spare_wqes); 499 500 if (qp->sq.wqe_cnt <= dev->dev->caps.max_wqes) 501 break; 502 503 if (qp->sq_max_wqes_per_wr <= 1) 504 return -EINVAL; 505 506 ++qp->sq.wqe_shift; 507 } 508 509 qp->sq.max_gs = (min(dev->dev->caps.max_sq_desc_sz, 510 (qp->sq_max_wqes_per_wr << qp->sq.wqe_shift)) - 511 send_wqe_overhead(type, qp->flags)) / 512 sizeof (struct mlx4_wqe_data_seg); 513 514 qp->buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) + 515 (qp->sq.wqe_cnt << qp->sq.wqe_shift); 516 if (qp->rq.wqe_shift > qp->sq.wqe_shift) { 517 qp->rq.offset = 0; 518 qp->sq.offset = qp->rq.wqe_cnt << qp->rq.wqe_shift; 519 } else { 520 qp->rq.offset = qp->sq.wqe_cnt << qp->sq.wqe_shift; 521 qp->sq.offset = 0; 522 } 523 524 cap->max_send_wr = qp->sq.max_post = 525 (qp->sq.wqe_cnt - qp->sq_spare_wqes) / qp->sq_max_wqes_per_wr; 526 cap->max_send_sge = min(qp->sq.max_gs, 527 min(dev->dev->caps.max_sq_sg, 528 dev->dev->caps.max_rq_sg)); 529 /* We don't support inline sends for kernel QPs (yet) */ 530 cap->max_inline_data = 0; 531 532 return 0; 533 } 534 535 static int set_user_sq_size(struct mlx4_ib_dev *dev, 536 struct mlx4_ib_qp *qp, 537 struct mlx4_ib_create_qp *ucmd) 538 { 539 /* Sanity check SQ size before proceeding */ 540 if ((1 << ucmd->log_sq_bb_count) > dev->dev->caps.max_wqes || 541 ucmd->log_sq_stride > 542 ilog2(roundup_pow_of_two(dev->dev->caps.max_sq_desc_sz)) || 543 ucmd->log_sq_stride < MLX4_IB_MIN_SQ_STRIDE) 544 return -EINVAL; 545 546 qp->sq.wqe_cnt = 1 << ucmd->log_sq_bb_count; 547 qp->sq.wqe_shift = ucmd->log_sq_stride; 548 549 qp->buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) + 550 (qp->sq.wqe_cnt << qp->sq.wqe_shift); 551 552 return 0; 553 } 554 555 static int alloc_proxy_bufs(struct ib_device *dev, struct mlx4_ib_qp *qp) 556 { 557 int i; 558 559 qp->sqp_proxy_rcv = 560 kmalloc(sizeof (struct mlx4_ib_buf) * qp->rq.wqe_cnt, 561 GFP_KERNEL); 562 if (!qp->sqp_proxy_rcv) 563 return -ENOMEM; 564 for (i = 0; i < qp->rq.wqe_cnt; i++) { 565 qp->sqp_proxy_rcv[i].addr = 566 kmalloc(sizeof (struct mlx4_ib_proxy_sqp_hdr), 567 GFP_KERNEL); 568 if (!qp->sqp_proxy_rcv[i].addr) 569 goto err; 570 qp->sqp_proxy_rcv[i].map = 571 ib_dma_map_single(dev, qp->sqp_proxy_rcv[i].addr, 572 sizeof (struct mlx4_ib_proxy_sqp_hdr), 573 DMA_FROM_DEVICE); 574 } 575 return 0; 576 577 err: 578 while (i > 0) { 579 --i; 580 ib_dma_unmap_single(dev, qp->sqp_proxy_rcv[i].map, 581 sizeof (struct mlx4_ib_proxy_sqp_hdr), 582 DMA_FROM_DEVICE); 583 kfree(qp->sqp_proxy_rcv[i].addr); 584 } 585 kfree(qp->sqp_proxy_rcv); 586 qp->sqp_proxy_rcv = NULL; 587 return -ENOMEM; 588 } 589 590 static void free_proxy_bufs(struct ib_device *dev, struct mlx4_ib_qp *qp) 591 { 592 int i; 593 594 for (i = 0; i < qp->rq.wqe_cnt; i++) { 595 ib_dma_unmap_single(dev, qp->sqp_proxy_rcv[i].map, 596 sizeof (struct mlx4_ib_proxy_sqp_hdr), 597 DMA_FROM_DEVICE); 598 kfree(qp->sqp_proxy_rcv[i].addr); 599 } 600 kfree(qp->sqp_proxy_rcv); 601 } 602 603 static int qp_has_rq(struct ib_qp_init_attr *attr) 604 { 605 if (attr->qp_type == IB_QPT_XRC_INI || attr->qp_type == IB_QPT_XRC_TGT) 606 return 0; 607 608 return !attr->srq; 609 } 610 611 static int create_qp_common(struct mlx4_ib_dev *dev, struct ib_pd *pd, 612 struct ib_qp_init_attr *init_attr, 613 struct ib_udata *udata, int sqpn, struct mlx4_ib_qp **caller_qp) 614 { 615 int qpn; 616 int err; 617 struct mlx4_ib_sqp *sqp; 618 struct mlx4_ib_qp *qp; 619 enum mlx4_ib_qp_type qp_type = (enum mlx4_ib_qp_type) init_attr->qp_type; 620 621 /* When tunneling special qps, we use a plain UD qp */ 622 if (sqpn) { 623 if (mlx4_is_mfunc(dev->dev) && 624 (!mlx4_is_master(dev->dev) || 625 !(init_attr->create_flags & MLX4_IB_SRIOV_SQP))) { 626 if (init_attr->qp_type == IB_QPT_GSI) 627 qp_type = MLX4_IB_QPT_PROXY_GSI; 628 else if (mlx4_is_master(dev->dev)) 629 qp_type = MLX4_IB_QPT_PROXY_SMI_OWNER; 630 else 631 qp_type = MLX4_IB_QPT_PROXY_SMI; 632 } 633 qpn = sqpn; 634 /* add extra sg entry for tunneling */ 635 init_attr->cap.max_recv_sge++; 636 } else if (init_attr->create_flags & MLX4_IB_SRIOV_TUNNEL_QP) { 637 struct mlx4_ib_qp_tunnel_init_attr *tnl_init = 638 container_of(init_attr, 639 struct mlx4_ib_qp_tunnel_init_attr, init_attr); 640 if ((tnl_init->proxy_qp_type != IB_QPT_SMI && 641 tnl_init->proxy_qp_type != IB_QPT_GSI) || 642 !mlx4_is_master(dev->dev)) 643 return -EINVAL; 644 if (tnl_init->proxy_qp_type == IB_QPT_GSI) 645 qp_type = MLX4_IB_QPT_TUN_GSI; 646 else if (tnl_init->slave == mlx4_master_func_num(dev->dev)) 647 qp_type = MLX4_IB_QPT_TUN_SMI_OWNER; 648 else 649 qp_type = MLX4_IB_QPT_TUN_SMI; 650 /* we are definitely in the PPF here, since we are creating 651 * tunnel QPs. base_tunnel_sqpn is therefore valid. */ 652 qpn = dev->dev->phys_caps.base_tunnel_sqpn + 8 * tnl_init->slave 653 + tnl_init->proxy_qp_type * 2 + tnl_init->port - 1; 654 sqpn = qpn; 655 } 656 657 if (!*caller_qp) { 658 if (qp_type == MLX4_IB_QPT_SMI || qp_type == MLX4_IB_QPT_GSI || 659 (qp_type & (MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_SMI_OWNER | 660 MLX4_IB_QPT_PROXY_GSI | MLX4_IB_QPT_TUN_SMI_OWNER))) { 661 sqp = kzalloc(sizeof (struct mlx4_ib_sqp), GFP_KERNEL); 662 if (!sqp) 663 return -ENOMEM; 664 qp = &sqp->qp; 665 qp->pri.vid = 0xFFFF; 666 qp->alt.vid = 0xFFFF; 667 } else { 668 qp = kzalloc(sizeof (struct mlx4_ib_qp), GFP_KERNEL); 669 if (!qp) 670 return -ENOMEM; 671 qp->pri.vid = 0xFFFF; 672 qp->alt.vid = 0xFFFF; 673 } 674 } else 675 qp = *caller_qp; 676 677 qp->mlx4_ib_qp_type = qp_type; 678 679 mutex_init(&qp->mutex); 680 spin_lock_init(&qp->sq.lock); 681 spin_lock_init(&qp->rq.lock); 682 INIT_LIST_HEAD(&qp->gid_list); 683 INIT_LIST_HEAD(&qp->steering_rules); 684 685 qp->state = IB_QPS_RESET; 686 if (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR) 687 qp->sq_signal_bits = cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE); 688 689 err = set_rq_size(dev, &init_attr->cap, !!pd->uobject, qp_has_rq(init_attr), qp); 690 if (err) 691 goto err; 692 693 if (pd->uobject) { 694 struct mlx4_ib_create_qp ucmd; 695 696 if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd)) { 697 err = -EFAULT; 698 goto err; 699 } 700 701 qp->sq_no_prefetch = ucmd.sq_no_prefetch; 702 703 err = set_user_sq_size(dev, qp, &ucmd); 704 if (err) 705 goto err; 706 707 qp->umem = ib_umem_get(pd->uobject->context, ucmd.buf_addr, 708 qp->buf_size, 0, 0); 709 if (IS_ERR(qp->umem)) { 710 err = PTR_ERR(qp->umem); 711 goto err; 712 } 713 714 err = mlx4_mtt_init(dev->dev, ib_umem_page_count(qp->umem), 715 ilog2(qp->umem->page_size), &qp->mtt); 716 if (err) 717 goto err_buf; 718 719 err = mlx4_ib_umem_write_mtt(dev, &qp->mtt, qp->umem); 720 if (err) 721 goto err_mtt; 722 723 if (qp_has_rq(init_attr)) { 724 err = mlx4_ib_db_map_user(to_mucontext(pd->uobject->context), 725 ucmd.db_addr, &qp->db); 726 if (err) 727 goto err_mtt; 728 } 729 } else { 730 qp->sq_no_prefetch = 0; 731 732 if (init_attr->create_flags & IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK) 733 qp->flags |= MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK; 734 735 if (init_attr->create_flags & IB_QP_CREATE_IPOIB_UD_LSO) 736 qp->flags |= MLX4_IB_QP_LSO; 737 738 if (init_attr->create_flags & IB_QP_CREATE_NETIF_QP) { 739 if (dev->steering_support == 740 MLX4_STEERING_MODE_DEVICE_MANAGED) 741 qp->flags |= MLX4_IB_QP_NETIF; 742 else 743 goto err; 744 } 745 746 err = set_kernel_sq_size(dev, &init_attr->cap, qp_type, qp); 747 if (err) 748 goto err; 749 750 if (qp_has_rq(init_attr)) { 751 err = mlx4_db_alloc(dev->dev, &qp->db, 0); 752 if (err) 753 goto err; 754 755 *qp->db.db = 0; 756 } 757 758 if (mlx4_buf_alloc(dev->dev, qp->buf_size, PAGE_SIZE * 2, &qp->buf)) { 759 err = -ENOMEM; 760 goto err_db; 761 } 762 763 err = mlx4_mtt_init(dev->dev, qp->buf.npages, qp->buf.page_shift, 764 &qp->mtt); 765 if (err) 766 goto err_buf; 767 768 err = mlx4_buf_write_mtt(dev->dev, &qp->mtt, &qp->buf); 769 if (err) 770 goto err_mtt; 771 772 qp->sq.wrid = kmalloc(qp->sq.wqe_cnt * sizeof (u64), GFP_KERNEL); 773 qp->rq.wrid = kmalloc(qp->rq.wqe_cnt * sizeof (u64), GFP_KERNEL); 774 775 if (!qp->sq.wrid || !qp->rq.wrid) { 776 err = -ENOMEM; 777 goto err_wrid; 778 } 779 } 780 781 if (sqpn) { 782 if (qp->mlx4_ib_qp_type & (MLX4_IB_QPT_PROXY_SMI_OWNER | 783 MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI)) { 784 if (alloc_proxy_bufs(pd->device, qp)) { 785 err = -ENOMEM; 786 goto err_wrid; 787 } 788 } 789 } else { 790 /* Raw packet QPNs must be aligned to 8 bits. If not, the WQE 791 * BlueFlame setup flow wrongly causes VLAN insertion. */ 792 if (init_attr->qp_type == IB_QPT_RAW_PACKET) 793 err = mlx4_qp_reserve_range(dev->dev, 1, 1 << 8, &qpn); 794 else 795 if (qp->flags & MLX4_IB_QP_NETIF) 796 err = mlx4_ib_steer_qp_alloc(dev, 1, &qpn); 797 else 798 err = mlx4_qp_reserve_range(dev->dev, 1, 1, 799 &qpn); 800 if (err) 801 goto err_proxy; 802 } 803 804 err = mlx4_qp_alloc(dev->dev, qpn, &qp->mqp); 805 if (err) 806 goto err_qpn; 807 808 if (init_attr->qp_type == IB_QPT_XRC_TGT) 809 qp->mqp.qpn |= (1 << 23); 810 811 /* 812 * Hardware wants QPN written in big-endian order (after 813 * shifting) for send doorbell. Precompute this value to save 814 * a little bit when posting sends. 815 */ 816 qp->doorbell_qpn = swab32(qp->mqp.qpn << 8); 817 818 qp->mqp.event = mlx4_ib_qp_event; 819 if (!*caller_qp) 820 *caller_qp = qp; 821 return 0; 822 823 err_qpn: 824 if (!sqpn) { 825 if (qp->flags & MLX4_IB_QP_NETIF) 826 mlx4_ib_steer_qp_free(dev, qpn, 1); 827 else 828 mlx4_qp_release_range(dev->dev, qpn, 1); 829 } 830 err_proxy: 831 if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI) 832 free_proxy_bufs(pd->device, qp); 833 err_wrid: 834 if (pd->uobject) { 835 if (qp_has_rq(init_attr)) 836 mlx4_ib_db_unmap_user(to_mucontext(pd->uobject->context), &qp->db); 837 } else { 838 kfree(qp->sq.wrid); 839 kfree(qp->rq.wrid); 840 } 841 842 err_mtt: 843 mlx4_mtt_cleanup(dev->dev, &qp->mtt); 844 845 err_buf: 846 if (pd->uobject) 847 ib_umem_release(qp->umem); 848 else 849 mlx4_buf_free(dev->dev, qp->buf_size, &qp->buf); 850 851 err_db: 852 if (!pd->uobject && qp_has_rq(init_attr)) 853 mlx4_db_free(dev->dev, &qp->db); 854 855 err: 856 if (!*caller_qp) 857 kfree(qp); 858 return err; 859 } 860 861 static enum mlx4_qp_state to_mlx4_state(enum ib_qp_state state) 862 { 863 switch (state) { 864 case IB_QPS_RESET: return MLX4_QP_STATE_RST; 865 case IB_QPS_INIT: return MLX4_QP_STATE_INIT; 866 case IB_QPS_RTR: return MLX4_QP_STATE_RTR; 867 case IB_QPS_RTS: return MLX4_QP_STATE_RTS; 868 case IB_QPS_SQD: return MLX4_QP_STATE_SQD; 869 case IB_QPS_SQE: return MLX4_QP_STATE_SQER; 870 case IB_QPS_ERR: return MLX4_QP_STATE_ERR; 871 default: return -1; 872 } 873 } 874 875 static void mlx4_ib_lock_cqs(struct mlx4_ib_cq *send_cq, struct mlx4_ib_cq *recv_cq) 876 __acquires(&send_cq->lock) __acquires(&recv_cq->lock) 877 { 878 if (send_cq == recv_cq) { 879 spin_lock_irq(&send_cq->lock); 880 __acquire(&recv_cq->lock); 881 } else if (send_cq->mcq.cqn < recv_cq->mcq.cqn) { 882 spin_lock_irq(&send_cq->lock); 883 spin_lock_nested(&recv_cq->lock, SINGLE_DEPTH_NESTING); 884 } else { 885 spin_lock_irq(&recv_cq->lock); 886 spin_lock_nested(&send_cq->lock, SINGLE_DEPTH_NESTING); 887 } 888 } 889 890 static void mlx4_ib_unlock_cqs(struct mlx4_ib_cq *send_cq, struct mlx4_ib_cq *recv_cq) 891 __releases(&send_cq->lock) __releases(&recv_cq->lock) 892 { 893 if (send_cq == recv_cq) { 894 __release(&recv_cq->lock); 895 spin_unlock_irq(&send_cq->lock); 896 } else if (send_cq->mcq.cqn < recv_cq->mcq.cqn) { 897 spin_unlock(&recv_cq->lock); 898 spin_unlock_irq(&send_cq->lock); 899 } else { 900 spin_unlock(&send_cq->lock); 901 spin_unlock_irq(&recv_cq->lock); 902 } 903 } 904 905 static void del_gid_entries(struct mlx4_ib_qp *qp) 906 { 907 struct mlx4_ib_gid_entry *ge, *tmp; 908 909 list_for_each_entry_safe(ge, tmp, &qp->gid_list, list) { 910 list_del(&ge->list); 911 kfree(ge); 912 } 913 } 914 915 static struct mlx4_ib_pd *get_pd(struct mlx4_ib_qp *qp) 916 { 917 if (qp->ibqp.qp_type == IB_QPT_XRC_TGT) 918 return to_mpd(to_mxrcd(qp->ibqp.xrcd)->pd); 919 else 920 return to_mpd(qp->ibqp.pd); 921 } 922 923 static void get_cqs(struct mlx4_ib_qp *qp, 924 struct mlx4_ib_cq **send_cq, struct mlx4_ib_cq **recv_cq) 925 { 926 switch (qp->ibqp.qp_type) { 927 case IB_QPT_XRC_TGT: 928 *send_cq = to_mcq(to_mxrcd(qp->ibqp.xrcd)->cq); 929 *recv_cq = *send_cq; 930 break; 931 case IB_QPT_XRC_INI: 932 *send_cq = to_mcq(qp->ibqp.send_cq); 933 *recv_cq = *send_cq; 934 break; 935 default: 936 *send_cq = to_mcq(qp->ibqp.send_cq); 937 *recv_cq = to_mcq(qp->ibqp.recv_cq); 938 break; 939 } 940 } 941 942 static void destroy_qp_common(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp, 943 int is_user) 944 { 945 struct mlx4_ib_cq *send_cq, *recv_cq; 946 947 if (qp->state != IB_QPS_RESET) { 948 if (mlx4_qp_modify(dev->dev, NULL, to_mlx4_state(qp->state), 949 MLX4_QP_STATE_RST, NULL, 0, 0, &qp->mqp)) 950 pr_warn("modify QP %06x to RESET failed.\n", 951 qp->mqp.qpn); 952 if (qp->pri.smac) { 953 mlx4_unregister_mac(dev->dev, qp->pri.smac_port, qp->pri.smac); 954 qp->pri.smac = 0; 955 } 956 if (qp->alt.smac) { 957 mlx4_unregister_mac(dev->dev, qp->alt.smac_port, qp->alt.smac); 958 qp->alt.smac = 0; 959 } 960 if (qp->pri.vid < 0x1000) { 961 mlx4_unregister_vlan(dev->dev, qp->pri.vlan_port, qp->pri.vid); 962 qp->pri.vid = 0xFFFF; 963 qp->pri.candidate_vid = 0xFFFF; 964 qp->pri.update_vid = 0; 965 } 966 if (qp->alt.vid < 0x1000) { 967 mlx4_unregister_vlan(dev->dev, qp->alt.vlan_port, qp->alt.vid); 968 qp->alt.vid = 0xFFFF; 969 qp->alt.candidate_vid = 0xFFFF; 970 qp->alt.update_vid = 0; 971 } 972 } 973 974 get_cqs(qp, &send_cq, &recv_cq); 975 976 mlx4_ib_lock_cqs(send_cq, recv_cq); 977 978 if (!is_user) { 979 __mlx4_ib_cq_clean(recv_cq, qp->mqp.qpn, 980 qp->ibqp.srq ? to_msrq(qp->ibqp.srq): NULL); 981 if (send_cq != recv_cq) 982 __mlx4_ib_cq_clean(send_cq, qp->mqp.qpn, NULL); 983 } 984 985 mlx4_qp_remove(dev->dev, &qp->mqp); 986 987 mlx4_ib_unlock_cqs(send_cq, recv_cq); 988 989 mlx4_qp_free(dev->dev, &qp->mqp); 990 991 if (!is_sqp(dev, qp) && !is_tunnel_qp(dev, qp)) { 992 if (qp->flags & MLX4_IB_QP_NETIF) 993 mlx4_ib_steer_qp_free(dev, qp->mqp.qpn, 1); 994 else 995 mlx4_qp_release_range(dev->dev, qp->mqp.qpn, 1); 996 } 997 998 mlx4_mtt_cleanup(dev->dev, &qp->mtt); 999 1000 if (is_user) { 1001 if (qp->rq.wqe_cnt) 1002 mlx4_ib_db_unmap_user(to_mucontext(qp->ibqp.uobject->context), 1003 &qp->db); 1004 ib_umem_release(qp->umem); 1005 } else { 1006 kfree(qp->sq.wrid); 1007 kfree(qp->rq.wrid); 1008 if (qp->mlx4_ib_qp_type & (MLX4_IB_QPT_PROXY_SMI_OWNER | 1009 MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI)) 1010 free_proxy_bufs(&dev->ib_dev, qp); 1011 mlx4_buf_free(dev->dev, qp->buf_size, &qp->buf); 1012 if (qp->rq.wqe_cnt) 1013 mlx4_db_free(dev->dev, &qp->db); 1014 } 1015 1016 del_gid_entries(qp); 1017 } 1018 1019 static u32 get_sqp_num(struct mlx4_ib_dev *dev, struct ib_qp_init_attr *attr) 1020 { 1021 /* Native or PPF */ 1022 if (!mlx4_is_mfunc(dev->dev) || 1023 (mlx4_is_master(dev->dev) && 1024 attr->create_flags & MLX4_IB_SRIOV_SQP)) { 1025 return dev->dev->phys_caps.base_sqpn + 1026 (attr->qp_type == IB_QPT_SMI ? 0 : 2) + 1027 attr->port_num - 1; 1028 } 1029 /* PF or VF -- creating proxies */ 1030 if (attr->qp_type == IB_QPT_SMI) 1031 return dev->dev->caps.qp0_proxy[attr->port_num - 1]; 1032 else 1033 return dev->dev->caps.qp1_proxy[attr->port_num - 1]; 1034 } 1035 1036 struct ib_qp *mlx4_ib_create_qp(struct ib_pd *pd, 1037 struct ib_qp_init_attr *init_attr, 1038 struct ib_udata *udata) 1039 { 1040 struct mlx4_ib_qp *qp = NULL; 1041 int err; 1042 u16 xrcdn = 0; 1043 1044 /* 1045 * We only support LSO, vendor flag1, and multicast loopback blocking, 1046 * and only for kernel UD QPs. 1047 */ 1048 if (init_attr->create_flags & ~(MLX4_IB_QP_LSO | 1049 MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK | 1050 MLX4_IB_SRIOV_TUNNEL_QP | 1051 MLX4_IB_SRIOV_SQP | 1052 MLX4_IB_QP_NETIF)) 1053 return ERR_PTR(-EINVAL); 1054 1055 if (init_attr->create_flags & IB_QP_CREATE_NETIF_QP) { 1056 if (init_attr->qp_type != IB_QPT_UD) 1057 return ERR_PTR(-EINVAL); 1058 } 1059 1060 if (init_attr->create_flags && 1061 (udata || 1062 ((init_attr->create_flags & ~MLX4_IB_SRIOV_SQP) && 1063 init_attr->qp_type != IB_QPT_UD) || 1064 ((init_attr->create_flags & MLX4_IB_SRIOV_SQP) && 1065 init_attr->qp_type > IB_QPT_GSI))) 1066 return ERR_PTR(-EINVAL); 1067 1068 switch (init_attr->qp_type) { 1069 case IB_QPT_XRC_TGT: 1070 pd = to_mxrcd(init_attr->xrcd)->pd; 1071 xrcdn = to_mxrcd(init_attr->xrcd)->xrcdn; 1072 init_attr->send_cq = to_mxrcd(init_attr->xrcd)->cq; 1073 /* fall through */ 1074 case IB_QPT_XRC_INI: 1075 if (!(to_mdev(pd->device)->dev->caps.flags & MLX4_DEV_CAP_FLAG_XRC)) 1076 return ERR_PTR(-ENOSYS); 1077 init_attr->recv_cq = init_attr->send_cq; 1078 /* fall through */ 1079 case IB_QPT_RC: 1080 case IB_QPT_UC: 1081 case IB_QPT_RAW_PACKET: 1082 qp = kzalloc(sizeof *qp, GFP_KERNEL); 1083 if (!qp) 1084 return ERR_PTR(-ENOMEM); 1085 qp->pri.vid = 0xFFFF; 1086 qp->alt.vid = 0xFFFF; 1087 /* fall through */ 1088 case IB_QPT_UD: 1089 { 1090 err = create_qp_common(to_mdev(pd->device), pd, init_attr, 1091 udata, 0, &qp); 1092 if (err) 1093 return ERR_PTR(err); 1094 1095 qp->ibqp.qp_num = qp->mqp.qpn; 1096 qp->xrcdn = xrcdn; 1097 1098 break; 1099 } 1100 case IB_QPT_SMI: 1101 case IB_QPT_GSI: 1102 { 1103 /* Userspace is not allowed to create special QPs: */ 1104 if (udata) 1105 return ERR_PTR(-EINVAL); 1106 1107 err = create_qp_common(to_mdev(pd->device), pd, init_attr, udata, 1108 get_sqp_num(to_mdev(pd->device), init_attr), 1109 &qp); 1110 if (err) 1111 return ERR_PTR(err); 1112 1113 qp->port = init_attr->port_num; 1114 qp->ibqp.qp_num = init_attr->qp_type == IB_QPT_SMI ? 0 : 1; 1115 1116 break; 1117 } 1118 default: 1119 /* Don't support raw QPs */ 1120 return ERR_PTR(-EINVAL); 1121 } 1122 1123 return &qp->ibqp; 1124 } 1125 1126 int mlx4_ib_destroy_qp(struct ib_qp *qp) 1127 { 1128 struct mlx4_ib_dev *dev = to_mdev(qp->device); 1129 struct mlx4_ib_qp *mqp = to_mqp(qp); 1130 struct mlx4_ib_pd *pd; 1131 1132 if (is_qp0(dev, mqp)) 1133 mlx4_CLOSE_PORT(dev->dev, mqp->port); 1134 1135 if (dev->qp1_proxy[mqp->port - 1] == mqp) { 1136 mutex_lock(&dev->qp1_proxy_lock[mqp->port - 1]); 1137 dev->qp1_proxy[mqp->port - 1] = NULL; 1138 mutex_unlock(&dev->qp1_proxy_lock[mqp->port - 1]); 1139 } 1140 1141 pd = get_pd(mqp); 1142 destroy_qp_common(dev, mqp, !!pd->ibpd.uobject); 1143 1144 if (is_sqp(dev, mqp)) 1145 kfree(to_msqp(mqp)); 1146 else 1147 kfree(mqp); 1148 1149 return 0; 1150 } 1151 1152 static int to_mlx4_st(struct mlx4_ib_dev *dev, enum mlx4_ib_qp_type type) 1153 { 1154 switch (type) { 1155 case MLX4_IB_QPT_RC: return MLX4_QP_ST_RC; 1156 case MLX4_IB_QPT_UC: return MLX4_QP_ST_UC; 1157 case MLX4_IB_QPT_UD: return MLX4_QP_ST_UD; 1158 case MLX4_IB_QPT_XRC_INI: 1159 case MLX4_IB_QPT_XRC_TGT: return MLX4_QP_ST_XRC; 1160 case MLX4_IB_QPT_SMI: 1161 case MLX4_IB_QPT_GSI: 1162 case MLX4_IB_QPT_RAW_PACKET: return MLX4_QP_ST_MLX; 1163 1164 case MLX4_IB_QPT_PROXY_SMI_OWNER: 1165 case MLX4_IB_QPT_TUN_SMI_OWNER: return (mlx4_is_mfunc(dev->dev) ? 1166 MLX4_QP_ST_MLX : -1); 1167 case MLX4_IB_QPT_PROXY_SMI: 1168 case MLX4_IB_QPT_TUN_SMI: 1169 case MLX4_IB_QPT_PROXY_GSI: 1170 case MLX4_IB_QPT_TUN_GSI: return (mlx4_is_mfunc(dev->dev) ? 1171 MLX4_QP_ST_UD : -1); 1172 default: return -1; 1173 } 1174 } 1175 1176 static __be32 to_mlx4_access_flags(struct mlx4_ib_qp *qp, const struct ib_qp_attr *attr, 1177 int attr_mask) 1178 { 1179 u8 dest_rd_atomic; 1180 u32 access_flags; 1181 u32 hw_access_flags = 0; 1182 1183 if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) 1184 dest_rd_atomic = attr->max_dest_rd_atomic; 1185 else 1186 dest_rd_atomic = qp->resp_depth; 1187 1188 if (attr_mask & IB_QP_ACCESS_FLAGS) 1189 access_flags = attr->qp_access_flags; 1190 else 1191 access_flags = qp->atomic_rd_en; 1192 1193 if (!dest_rd_atomic) 1194 access_flags &= IB_ACCESS_REMOTE_WRITE; 1195 1196 if (access_flags & IB_ACCESS_REMOTE_READ) 1197 hw_access_flags |= MLX4_QP_BIT_RRE; 1198 if (access_flags & IB_ACCESS_REMOTE_ATOMIC) 1199 hw_access_flags |= MLX4_QP_BIT_RAE; 1200 if (access_flags & IB_ACCESS_REMOTE_WRITE) 1201 hw_access_flags |= MLX4_QP_BIT_RWE; 1202 1203 return cpu_to_be32(hw_access_flags); 1204 } 1205 1206 static void store_sqp_attrs(struct mlx4_ib_sqp *sqp, const struct ib_qp_attr *attr, 1207 int attr_mask) 1208 { 1209 if (attr_mask & IB_QP_PKEY_INDEX) 1210 sqp->pkey_index = attr->pkey_index; 1211 if (attr_mask & IB_QP_QKEY) 1212 sqp->qkey = attr->qkey; 1213 if (attr_mask & IB_QP_SQ_PSN) 1214 sqp->send_psn = attr->sq_psn; 1215 } 1216 1217 static void mlx4_set_sched(struct mlx4_qp_path *path, u8 port) 1218 { 1219 path->sched_queue = (path->sched_queue & 0xbf) | ((port - 1) << 6); 1220 } 1221 1222 static int _mlx4_set_path(struct mlx4_ib_dev *dev, const struct ib_ah_attr *ah, 1223 u64 smac, u16 vlan_tag, struct mlx4_qp_path *path, 1224 struct mlx4_roce_smac_vlan_info *smac_info, u8 port) 1225 { 1226 int is_eth = rdma_port_get_link_layer(&dev->ib_dev, port) == 1227 IB_LINK_LAYER_ETHERNET; 1228 int vidx; 1229 int smac_index; 1230 int err; 1231 1232 1233 path->grh_mylmc = ah->src_path_bits & 0x7f; 1234 path->rlid = cpu_to_be16(ah->dlid); 1235 if (ah->static_rate) { 1236 path->static_rate = ah->static_rate + MLX4_STAT_RATE_OFFSET; 1237 while (path->static_rate > IB_RATE_2_5_GBPS + MLX4_STAT_RATE_OFFSET && 1238 !(1 << path->static_rate & dev->dev->caps.stat_rate_support)) 1239 --path->static_rate; 1240 } else 1241 path->static_rate = 0; 1242 1243 if (ah->ah_flags & IB_AH_GRH) { 1244 if (ah->grh.sgid_index >= dev->dev->caps.gid_table_len[port]) { 1245 pr_err("sgid_index (%u) too large. max is %d\n", 1246 ah->grh.sgid_index, dev->dev->caps.gid_table_len[port] - 1); 1247 return -1; 1248 } 1249 1250 path->grh_mylmc |= 1 << 7; 1251 path->mgid_index = ah->grh.sgid_index; 1252 path->hop_limit = ah->grh.hop_limit; 1253 path->tclass_flowlabel = 1254 cpu_to_be32((ah->grh.traffic_class << 20) | 1255 (ah->grh.flow_label)); 1256 memcpy(path->rgid, ah->grh.dgid.raw, 16); 1257 } 1258 1259 if (is_eth) { 1260 if (!(ah->ah_flags & IB_AH_GRH)) 1261 return -1; 1262 1263 path->sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE | 1264 ((port - 1) << 6) | ((ah->sl & 7) << 3); 1265 1266 path->feup |= MLX4_FEUP_FORCE_ETH_UP; 1267 if (vlan_tag < 0x1000) { 1268 if (smac_info->vid < 0x1000) { 1269 /* both valid vlan ids */ 1270 if (smac_info->vid != vlan_tag) { 1271 /* different VIDs. unreg old and reg new */ 1272 err = mlx4_register_vlan(dev->dev, port, vlan_tag, &vidx); 1273 if (err) 1274 return err; 1275 smac_info->candidate_vid = vlan_tag; 1276 smac_info->candidate_vlan_index = vidx; 1277 smac_info->candidate_vlan_port = port; 1278 smac_info->update_vid = 1; 1279 path->vlan_index = vidx; 1280 } else { 1281 path->vlan_index = smac_info->vlan_index; 1282 } 1283 } else { 1284 /* no current vlan tag in qp */ 1285 err = mlx4_register_vlan(dev->dev, port, vlan_tag, &vidx); 1286 if (err) 1287 return err; 1288 smac_info->candidate_vid = vlan_tag; 1289 smac_info->candidate_vlan_index = vidx; 1290 smac_info->candidate_vlan_port = port; 1291 smac_info->update_vid = 1; 1292 path->vlan_index = vidx; 1293 } 1294 path->feup |= MLX4_FVL_FORCE_ETH_VLAN; 1295 path->fl = 1 << 6; 1296 } else { 1297 /* have current vlan tag. unregister it at modify-qp success */ 1298 if (smac_info->vid < 0x1000) { 1299 smac_info->candidate_vid = 0xFFFF; 1300 smac_info->update_vid = 1; 1301 } 1302 } 1303 1304 /* get smac_index for RoCE use. 1305 * If no smac was yet assigned, register one. 1306 * If one was already assigned, but the new mac differs, 1307 * unregister the old one and register the new one. 1308 */ 1309 if (!smac_info->smac || smac_info->smac != smac) { 1310 /* register candidate now, unreg if needed, after success */ 1311 smac_index = mlx4_register_mac(dev->dev, port, smac); 1312 if (smac_index >= 0) { 1313 smac_info->candidate_smac_index = smac_index; 1314 smac_info->candidate_smac = smac; 1315 smac_info->candidate_smac_port = port; 1316 } else { 1317 return -EINVAL; 1318 } 1319 } else { 1320 smac_index = smac_info->smac_index; 1321 } 1322 1323 memcpy(path->dmac, ah->dmac, 6); 1324 path->ackto = MLX4_IB_LINK_TYPE_ETH; 1325 /* put MAC table smac index for IBoE */ 1326 path->grh_mylmc = (u8) (smac_index) | 0x80; 1327 } else { 1328 path->sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE | 1329 ((port - 1) << 6) | ((ah->sl & 0xf) << 2); 1330 } 1331 1332 return 0; 1333 } 1334 1335 static int mlx4_set_path(struct mlx4_ib_dev *dev, const struct ib_qp_attr *qp, 1336 enum ib_qp_attr_mask qp_attr_mask, 1337 struct mlx4_ib_qp *mqp, 1338 struct mlx4_qp_path *path, u8 port) 1339 { 1340 return _mlx4_set_path(dev, &qp->ah_attr, 1341 mlx4_mac_to_u64((u8 *)qp->smac), 1342 (qp_attr_mask & IB_QP_VID) ? qp->vlan_id : 0xffff, 1343 path, &mqp->pri, port); 1344 } 1345 1346 static int mlx4_set_alt_path(struct mlx4_ib_dev *dev, 1347 const struct ib_qp_attr *qp, 1348 enum ib_qp_attr_mask qp_attr_mask, 1349 struct mlx4_ib_qp *mqp, 1350 struct mlx4_qp_path *path, u8 port) 1351 { 1352 return _mlx4_set_path(dev, &qp->alt_ah_attr, 1353 mlx4_mac_to_u64((u8 *)qp->alt_smac), 1354 (qp_attr_mask & IB_QP_ALT_VID) ? 1355 qp->alt_vlan_id : 0xffff, 1356 path, &mqp->alt, port); 1357 } 1358 1359 static void update_mcg_macs(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp) 1360 { 1361 struct mlx4_ib_gid_entry *ge, *tmp; 1362 1363 list_for_each_entry_safe(ge, tmp, &qp->gid_list, list) { 1364 if (!ge->added && mlx4_ib_add_mc(dev, qp, &ge->gid)) { 1365 ge->added = 1; 1366 ge->port = qp->port; 1367 } 1368 } 1369 } 1370 1371 static int handle_eth_ud_smac_index(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp, u8 *smac, 1372 struct mlx4_qp_context *context) 1373 { 1374 struct net_device *ndev; 1375 u64 u64_mac; 1376 int smac_index; 1377 1378 1379 ndev = dev->iboe.netdevs[qp->port - 1]; 1380 if (ndev) { 1381 smac = ndev->dev_addr; 1382 u64_mac = mlx4_mac_to_u64(smac); 1383 } else { 1384 u64_mac = dev->dev->caps.def_mac[qp->port]; 1385 } 1386 1387 context->pri_path.sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE | ((qp->port - 1) << 6); 1388 if (!qp->pri.smac) { 1389 smac_index = mlx4_register_mac(dev->dev, qp->port, u64_mac); 1390 if (smac_index >= 0) { 1391 qp->pri.candidate_smac_index = smac_index; 1392 qp->pri.candidate_smac = u64_mac; 1393 qp->pri.candidate_smac_port = qp->port; 1394 context->pri_path.grh_mylmc = 0x80 | (u8) smac_index; 1395 } else { 1396 return -ENOENT; 1397 } 1398 } 1399 return 0; 1400 } 1401 1402 static int __mlx4_ib_modify_qp(struct ib_qp *ibqp, 1403 const struct ib_qp_attr *attr, int attr_mask, 1404 enum ib_qp_state cur_state, enum ib_qp_state new_state) 1405 { 1406 struct mlx4_ib_dev *dev = to_mdev(ibqp->device); 1407 struct mlx4_ib_qp *qp = to_mqp(ibqp); 1408 struct mlx4_ib_pd *pd; 1409 struct mlx4_ib_cq *send_cq, *recv_cq; 1410 struct mlx4_qp_context *context; 1411 enum mlx4_qp_optpar optpar = 0; 1412 int sqd_event; 1413 int steer_qp = 0; 1414 int err = -EINVAL; 1415 1416 context = kzalloc(sizeof *context, GFP_KERNEL); 1417 if (!context) 1418 return -ENOMEM; 1419 1420 context->flags = cpu_to_be32((to_mlx4_state(new_state) << 28) | 1421 (to_mlx4_st(dev, qp->mlx4_ib_qp_type) << 16)); 1422 1423 if (!(attr_mask & IB_QP_PATH_MIG_STATE)) 1424 context->flags |= cpu_to_be32(MLX4_QP_PM_MIGRATED << 11); 1425 else { 1426 optpar |= MLX4_QP_OPTPAR_PM_STATE; 1427 switch (attr->path_mig_state) { 1428 case IB_MIG_MIGRATED: 1429 context->flags |= cpu_to_be32(MLX4_QP_PM_MIGRATED << 11); 1430 break; 1431 case IB_MIG_REARM: 1432 context->flags |= cpu_to_be32(MLX4_QP_PM_REARM << 11); 1433 break; 1434 case IB_MIG_ARMED: 1435 context->flags |= cpu_to_be32(MLX4_QP_PM_ARMED << 11); 1436 break; 1437 } 1438 } 1439 1440 if (ibqp->qp_type == IB_QPT_GSI || ibqp->qp_type == IB_QPT_SMI) 1441 context->mtu_msgmax = (IB_MTU_4096 << 5) | 11; 1442 else if (ibqp->qp_type == IB_QPT_RAW_PACKET) 1443 context->mtu_msgmax = (MLX4_RAW_QP_MTU << 5) | MLX4_RAW_QP_MSGMAX; 1444 else if (ibqp->qp_type == IB_QPT_UD) { 1445 if (qp->flags & MLX4_IB_QP_LSO) 1446 context->mtu_msgmax = (IB_MTU_4096 << 5) | 1447 ilog2(dev->dev->caps.max_gso_sz); 1448 else 1449 context->mtu_msgmax = (IB_MTU_4096 << 5) | 12; 1450 } else if (attr_mask & IB_QP_PATH_MTU) { 1451 if (attr->path_mtu < IB_MTU_256 || attr->path_mtu > IB_MTU_4096) { 1452 pr_err("path MTU (%u) is invalid\n", 1453 attr->path_mtu); 1454 goto out; 1455 } 1456 context->mtu_msgmax = (attr->path_mtu << 5) | 1457 ilog2(dev->dev->caps.max_msg_sz); 1458 } 1459 1460 if (qp->rq.wqe_cnt) 1461 context->rq_size_stride = ilog2(qp->rq.wqe_cnt) << 3; 1462 context->rq_size_stride |= qp->rq.wqe_shift - 4; 1463 1464 if (qp->sq.wqe_cnt) 1465 context->sq_size_stride = ilog2(qp->sq.wqe_cnt) << 3; 1466 context->sq_size_stride |= qp->sq.wqe_shift - 4; 1467 1468 if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) { 1469 context->sq_size_stride |= !!qp->sq_no_prefetch << 7; 1470 context->xrcd = cpu_to_be32((u32) qp->xrcdn); 1471 if (ibqp->qp_type == IB_QPT_RAW_PACKET) 1472 context->param3 |= cpu_to_be32(1 << 30); 1473 } 1474 1475 if (qp->ibqp.uobject) 1476 context->usr_page = cpu_to_be32(to_mucontext(ibqp->uobject->context)->uar.index); 1477 else 1478 context->usr_page = cpu_to_be32(dev->priv_uar.index); 1479 1480 if (attr_mask & IB_QP_DEST_QPN) 1481 context->remote_qpn = cpu_to_be32(attr->dest_qp_num); 1482 1483 if (attr_mask & IB_QP_PORT) { 1484 if (cur_state == IB_QPS_SQD && new_state == IB_QPS_SQD && 1485 !(attr_mask & IB_QP_AV)) { 1486 mlx4_set_sched(&context->pri_path, attr->port_num); 1487 optpar |= MLX4_QP_OPTPAR_SCHED_QUEUE; 1488 } 1489 } 1490 1491 if (cur_state == IB_QPS_INIT && new_state == IB_QPS_RTR) { 1492 if (dev->counters[qp->port - 1] != -1) { 1493 context->pri_path.counter_index = 1494 dev->counters[qp->port - 1]; 1495 optpar |= MLX4_QP_OPTPAR_COUNTER_INDEX; 1496 } else 1497 context->pri_path.counter_index = 0xff; 1498 1499 if (qp->flags & MLX4_IB_QP_NETIF) { 1500 mlx4_ib_steer_qp_reg(dev, qp, 1); 1501 steer_qp = 1; 1502 } 1503 } 1504 1505 if (attr_mask & IB_QP_PKEY_INDEX) { 1506 if (qp->mlx4_ib_qp_type & MLX4_IB_QPT_ANY_SRIOV) 1507 context->pri_path.disable_pkey_check = 0x40; 1508 context->pri_path.pkey_index = attr->pkey_index; 1509 optpar |= MLX4_QP_OPTPAR_PKEY_INDEX; 1510 } 1511 1512 if (attr_mask & IB_QP_AV) { 1513 if (mlx4_set_path(dev, attr, attr_mask, qp, &context->pri_path, 1514 attr_mask & IB_QP_PORT ? 1515 attr->port_num : qp->port)) 1516 goto out; 1517 1518 optpar |= (MLX4_QP_OPTPAR_PRIMARY_ADDR_PATH | 1519 MLX4_QP_OPTPAR_SCHED_QUEUE); 1520 } 1521 1522 if (attr_mask & IB_QP_TIMEOUT) { 1523 context->pri_path.ackto |= attr->timeout << 3; 1524 optpar |= MLX4_QP_OPTPAR_ACK_TIMEOUT; 1525 } 1526 1527 if (attr_mask & IB_QP_ALT_PATH) { 1528 if (attr->alt_port_num == 0 || 1529 attr->alt_port_num > dev->dev->caps.num_ports) 1530 goto out; 1531 1532 if (attr->alt_pkey_index >= 1533 dev->dev->caps.pkey_table_len[attr->alt_port_num]) 1534 goto out; 1535 1536 if (mlx4_set_alt_path(dev, attr, attr_mask, qp, 1537 &context->alt_path, 1538 attr->alt_port_num)) 1539 goto out; 1540 1541 context->alt_path.pkey_index = attr->alt_pkey_index; 1542 context->alt_path.ackto = attr->alt_timeout << 3; 1543 optpar |= MLX4_QP_OPTPAR_ALT_ADDR_PATH; 1544 } 1545 1546 pd = get_pd(qp); 1547 get_cqs(qp, &send_cq, &recv_cq); 1548 context->pd = cpu_to_be32(pd->pdn); 1549 context->cqn_send = cpu_to_be32(send_cq->mcq.cqn); 1550 context->cqn_recv = cpu_to_be32(recv_cq->mcq.cqn); 1551 context->params1 = cpu_to_be32(MLX4_IB_ACK_REQ_FREQ << 28); 1552 1553 /* Set "fast registration enabled" for all kernel QPs */ 1554 if (!qp->ibqp.uobject) 1555 context->params1 |= cpu_to_be32(1 << 11); 1556 1557 if (attr_mask & IB_QP_RNR_RETRY) { 1558 context->params1 |= cpu_to_be32(attr->rnr_retry << 13); 1559 optpar |= MLX4_QP_OPTPAR_RNR_RETRY; 1560 } 1561 1562 if (attr_mask & IB_QP_RETRY_CNT) { 1563 context->params1 |= cpu_to_be32(attr->retry_cnt << 16); 1564 optpar |= MLX4_QP_OPTPAR_RETRY_COUNT; 1565 } 1566 1567 if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) { 1568 if (attr->max_rd_atomic) 1569 context->params1 |= 1570 cpu_to_be32(fls(attr->max_rd_atomic - 1) << 21); 1571 optpar |= MLX4_QP_OPTPAR_SRA_MAX; 1572 } 1573 1574 if (attr_mask & IB_QP_SQ_PSN) 1575 context->next_send_psn = cpu_to_be32(attr->sq_psn); 1576 1577 if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) { 1578 if (attr->max_dest_rd_atomic) 1579 context->params2 |= 1580 cpu_to_be32(fls(attr->max_dest_rd_atomic - 1) << 21); 1581 optpar |= MLX4_QP_OPTPAR_RRA_MAX; 1582 } 1583 1584 if (attr_mask & (IB_QP_ACCESS_FLAGS | IB_QP_MAX_DEST_RD_ATOMIC)) { 1585 context->params2 |= to_mlx4_access_flags(qp, attr, attr_mask); 1586 optpar |= MLX4_QP_OPTPAR_RWE | MLX4_QP_OPTPAR_RRE | MLX4_QP_OPTPAR_RAE; 1587 } 1588 1589 if (ibqp->srq) 1590 context->params2 |= cpu_to_be32(MLX4_QP_BIT_RIC); 1591 1592 if (attr_mask & IB_QP_MIN_RNR_TIMER) { 1593 context->rnr_nextrecvpsn |= cpu_to_be32(attr->min_rnr_timer << 24); 1594 optpar |= MLX4_QP_OPTPAR_RNR_TIMEOUT; 1595 } 1596 if (attr_mask & IB_QP_RQ_PSN) 1597 context->rnr_nextrecvpsn |= cpu_to_be32(attr->rq_psn); 1598 1599 /* proxy and tunnel qp qkeys will be changed in modify-qp wrappers */ 1600 if (attr_mask & IB_QP_QKEY) { 1601 if (qp->mlx4_ib_qp_type & 1602 (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER)) 1603 context->qkey = cpu_to_be32(IB_QP_SET_QKEY); 1604 else { 1605 if (mlx4_is_mfunc(dev->dev) && 1606 !(qp->mlx4_ib_qp_type & MLX4_IB_QPT_ANY_SRIOV) && 1607 (attr->qkey & MLX4_RESERVED_QKEY_MASK) == 1608 MLX4_RESERVED_QKEY_BASE) { 1609 pr_err("Cannot use reserved QKEY" 1610 " 0x%x (range 0xffff0000..0xffffffff" 1611 " is reserved)\n", attr->qkey); 1612 err = -EINVAL; 1613 goto out; 1614 } 1615 context->qkey = cpu_to_be32(attr->qkey); 1616 } 1617 optpar |= MLX4_QP_OPTPAR_Q_KEY; 1618 } 1619 1620 if (ibqp->srq) 1621 context->srqn = cpu_to_be32(1 << 24 | to_msrq(ibqp->srq)->msrq.srqn); 1622 1623 if (qp->rq.wqe_cnt && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) 1624 context->db_rec_addr = cpu_to_be64(qp->db.dma); 1625 1626 if (cur_state == IB_QPS_INIT && 1627 new_state == IB_QPS_RTR && 1628 (ibqp->qp_type == IB_QPT_GSI || ibqp->qp_type == IB_QPT_SMI || 1629 ibqp->qp_type == IB_QPT_UD || 1630 ibqp->qp_type == IB_QPT_RAW_PACKET)) { 1631 context->pri_path.sched_queue = (qp->port - 1) << 6; 1632 if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_SMI || 1633 qp->mlx4_ib_qp_type & 1634 (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER)) { 1635 context->pri_path.sched_queue |= MLX4_IB_DEFAULT_QP0_SCHED_QUEUE; 1636 if (qp->mlx4_ib_qp_type != MLX4_IB_QPT_SMI) 1637 context->pri_path.fl = 0x80; 1638 } else { 1639 if (qp->mlx4_ib_qp_type & MLX4_IB_QPT_ANY_SRIOV) 1640 context->pri_path.fl = 0x80; 1641 context->pri_path.sched_queue |= MLX4_IB_DEFAULT_SCHED_QUEUE; 1642 } 1643 if (rdma_port_get_link_layer(&dev->ib_dev, qp->port) == 1644 IB_LINK_LAYER_ETHERNET) { 1645 if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_TUN_GSI || 1646 qp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI) 1647 context->pri_path.feup = 1 << 7; /* don't fsm */ 1648 /* handle smac_index */ 1649 if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_UD || 1650 qp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI || 1651 qp->mlx4_ib_qp_type == MLX4_IB_QPT_TUN_GSI) { 1652 err = handle_eth_ud_smac_index(dev, qp, (u8 *)attr->smac, context); 1653 if (err) 1654 return -EINVAL; 1655 if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI) 1656 dev->qp1_proxy[qp->port - 1] = qp; 1657 } 1658 } 1659 } 1660 1661 if (qp->ibqp.qp_type == IB_QPT_RAW_PACKET) 1662 context->pri_path.ackto = (context->pri_path.ackto & 0xf8) | 1663 MLX4_IB_LINK_TYPE_ETH; 1664 1665 if (ibqp->qp_type == IB_QPT_UD && (new_state == IB_QPS_RTR)) { 1666 int is_eth = rdma_port_get_link_layer( 1667 &dev->ib_dev, qp->port) == 1668 IB_LINK_LAYER_ETHERNET; 1669 if (is_eth) { 1670 context->pri_path.ackto = MLX4_IB_LINK_TYPE_ETH; 1671 optpar |= MLX4_QP_OPTPAR_PRIMARY_ADDR_PATH; 1672 } 1673 } 1674 1675 1676 if (cur_state == IB_QPS_RTS && new_state == IB_QPS_SQD && 1677 attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY && attr->en_sqd_async_notify) 1678 sqd_event = 1; 1679 else 1680 sqd_event = 0; 1681 1682 if (!ibqp->uobject && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) 1683 context->rlkey |= (1 << 4); 1684 1685 /* 1686 * Before passing a kernel QP to the HW, make sure that the 1687 * ownership bits of the send queue are set and the SQ 1688 * headroom is stamped so that the hardware doesn't start 1689 * processing stale work requests. 1690 */ 1691 if (!ibqp->uobject && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) { 1692 struct mlx4_wqe_ctrl_seg *ctrl; 1693 int i; 1694 1695 for (i = 0; i < qp->sq.wqe_cnt; ++i) { 1696 ctrl = get_send_wqe(qp, i); 1697 ctrl->owner_opcode = cpu_to_be32(1 << 31); 1698 if (qp->sq_max_wqes_per_wr == 1) 1699 ctrl->fence_size = 1 << (qp->sq.wqe_shift - 4); 1700 1701 stamp_send_wqe(qp, i, 1 << qp->sq.wqe_shift); 1702 } 1703 } 1704 1705 err = mlx4_qp_modify(dev->dev, &qp->mtt, to_mlx4_state(cur_state), 1706 to_mlx4_state(new_state), context, optpar, 1707 sqd_event, &qp->mqp); 1708 if (err) 1709 goto out; 1710 1711 qp->state = new_state; 1712 1713 if (attr_mask & IB_QP_ACCESS_FLAGS) 1714 qp->atomic_rd_en = attr->qp_access_flags; 1715 if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) 1716 qp->resp_depth = attr->max_dest_rd_atomic; 1717 if (attr_mask & IB_QP_PORT) { 1718 qp->port = attr->port_num; 1719 update_mcg_macs(dev, qp); 1720 } 1721 if (attr_mask & IB_QP_ALT_PATH) 1722 qp->alt_port = attr->alt_port_num; 1723 1724 if (is_sqp(dev, qp)) 1725 store_sqp_attrs(to_msqp(qp), attr, attr_mask); 1726 1727 /* 1728 * If we moved QP0 to RTR, bring the IB link up; if we moved 1729 * QP0 to RESET or ERROR, bring the link back down. 1730 */ 1731 if (is_qp0(dev, qp)) { 1732 if (cur_state != IB_QPS_RTR && new_state == IB_QPS_RTR) 1733 if (mlx4_INIT_PORT(dev->dev, qp->port)) 1734 pr_warn("INIT_PORT failed for port %d\n", 1735 qp->port); 1736 1737 if (cur_state != IB_QPS_RESET && cur_state != IB_QPS_ERR && 1738 (new_state == IB_QPS_RESET || new_state == IB_QPS_ERR)) 1739 mlx4_CLOSE_PORT(dev->dev, qp->port); 1740 } 1741 1742 /* 1743 * If we moved a kernel QP to RESET, clean up all old CQ 1744 * entries and reinitialize the QP. 1745 */ 1746 if (new_state == IB_QPS_RESET) { 1747 if (!ibqp->uobject) { 1748 mlx4_ib_cq_clean(recv_cq, qp->mqp.qpn, 1749 ibqp->srq ? to_msrq(ibqp->srq) : NULL); 1750 if (send_cq != recv_cq) 1751 mlx4_ib_cq_clean(send_cq, qp->mqp.qpn, NULL); 1752 1753 qp->rq.head = 0; 1754 qp->rq.tail = 0; 1755 qp->sq.head = 0; 1756 qp->sq.tail = 0; 1757 qp->sq_next_wqe = 0; 1758 if (qp->rq.wqe_cnt) 1759 *qp->db.db = 0; 1760 1761 if (qp->flags & MLX4_IB_QP_NETIF) 1762 mlx4_ib_steer_qp_reg(dev, qp, 0); 1763 } 1764 if (qp->pri.smac) { 1765 mlx4_unregister_mac(dev->dev, qp->pri.smac_port, qp->pri.smac); 1766 qp->pri.smac = 0; 1767 } 1768 if (qp->alt.smac) { 1769 mlx4_unregister_mac(dev->dev, qp->alt.smac_port, qp->alt.smac); 1770 qp->alt.smac = 0; 1771 } 1772 if (qp->pri.vid < 0x1000) { 1773 mlx4_unregister_vlan(dev->dev, qp->pri.vlan_port, qp->pri.vid); 1774 qp->pri.vid = 0xFFFF; 1775 qp->pri.candidate_vid = 0xFFFF; 1776 qp->pri.update_vid = 0; 1777 } 1778 1779 if (qp->alt.vid < 0x1000) { 1780 mlx4_unregister_vlan(dev->dev, qp->alt.vlan_port, qp->alt.vid); 1781 qp->alt.vid = 0xFFFF; 1782 qp->alt.candidate_vid = 0xFFFF; 1783 qp->alt.update_vid = 0; 1784 } 1785 } 1786 out: 1787 if (err && steer_qp) 1788 mlx4_ib_steer_qp_reg(dev, qp, 0); 1789 kfree(context); 1790 if (qp->pri.candidate_smac) { 1791 if (err) { 1792 mlx4_unregister_mac(dev->dev, qp->pri.candidate_smac_port, qp->pri.candidate_smac); 1793 } else { 1794 if (qp->pri.smac) 1795 mlx4_unregister_mac(dev->dev, qp->pri.smac_port, qp->pri.smac); 1796 qp->pri.smac = qp->pri.candidate_smac; 1797 qp->pri.smac_index = qp->pri.candidate_smac_index; 1798 qp->pri.smac_port = qp->pri.candidate_smac_port; 1799 } 1800 qp->pri.candidate_smac = 0; 1801 qp->pri.candidate_smac_index = 0; 1802 qp->pri.candidate_smac_port = 0; 1803 } 1804 if (qp->alt.candidate_smac) { 1805 if (err) { 1806 mlx4_unregister_mac(dev->dev, qp->alt.candidate_smac_port, qp->alt.candidate_smac); 1807 } else { 1808 if (qp->alt.smac) 1809 mlx4_unregister_mac(dev->dev, qp->alt.smac_port, qp->alt.smac); 1810 qp->alt.smac = qp->alt.candidate_smac; 1811 qp->alt.smac_index = qp->alt.candidate_smac_index; 1812 qp->alt.smac_port = qp->alt.candidate_smac_port; 1813 } 1814 qp->alt.candidate_smac = 0; 1815 qp->alt.candidate_smac_index = 0; 1816 qp->alt.candidate_smac_port = 0; 1817 } 1818 1819 if (qp->pri.update_vid) { 1820 if (err) { 1821 if (qp->pri.candidate_vid < 0x1000) 1822 mlx4_unregister_vlan(dev->dev, qp->pri.candidate_vlan_port, 1823 qp->pri.candidate_vid); 1824 } else { 1825 if (qp->pri.vid < 0x1000) 1826 mlx4_unregister_vlan(dev->dev, qp->pri.vlan_port, 1827 qp->pri.vid); 1828 qp->pri.vid = qp->pri.candidate_vid; 1829 qp->pri.vlan_port = qp->pri.candidate_vlan_port; 1830 qp->pri.vlan_index = qp->pri.candidate_vlan_index; 1831 } 1832 qp->pri.candidate_vid = 0xFFFF; 1833 qp->pri.update_vid = 0; 1834 } 1835 1836 if (qp->alt.update_vid) { 1837 if (err) { 1838 if (qp->alt.candidate_vid < 0x1000) 1839 mlx4_unregister_vlan(dev->dev, qp->alt.candidate_vlan_port, 1840 qp->alt.candidate_vid); 1841 } else { 1842 if (qp->alt.vid < 0x1000) 1843 mlx4_unregister_vlan(dev->dev, qp->alt.vlan_port, 1844 qp->alt.vid); 1845 qp->alt.vid = qp->alt.candidate_vid; 1846 qp->alt.vlan_port = qp->alt.candidate_vlan_port; 1847 qp->alt.vlan_index = qp->alt.candidate_vlan_index; 1848 } 1849 qp->alt.candidate_vid = 0xFFFF; 1850 qp->alt.update_vid = 0; 1851 } 1852 1853 return err; 1854 } 1855 1856 int mlx4_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, 1857 int attr_mask, struct ib_udata *udata) 1858 { 1859 struct mlx4_ib_dev *dev = to_mdev(ibqp->device); 1860 struct mlx4_ib_qp *qp = to_mqp(ibqp); 1861 enum ib_qp_state cur_state, new_state; 1862 int err = -EINVAL; 1863 int ll; 1864 mutex_lock(&qp->mutex); 1865 1866 cur_state = attr_mask & IB_QP_CUR_STATE ? attr->cur_qp_state : qp->state; 1867 new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state; 1868 1869 if (cur_state == new_state && cur_state == IB_QPS_RESET) { 1870 ll = IB_LINK_LAYER_UNSPECIFIED; 1871 } else { 1872 int port = attr_mask & IB_QP_PORT ? attr->port_num : qp->port; 1873 ll = rdma_port_get_link_layer(&dev->ib_dev, port); 1874 } 1875 1876 if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type, 1877 attr_mask, ll)) { 1878 pr_debug("qpn 0x%x: invalid attribute mask specified " 1879 "for transition %d to %d. qp_type %d," 1880 " attr_mask 0x%x\n", 1881 ibqp->qp_num, cur_state, new_state, 1882 ibqp->qp_type, attr_mask); 1883 goto out; 1884 } 1885 1886 if ((attr_mask & IB_QP_PORT) && 1887 (attr->port_num == 0 || attr->port_num > dev->num_ports)) { 1888 pr_debug("qpn 0x%x: invalid port number (%d) specified " 1889 "for transition %d to %d. qp_type %d\n", 1890 ibqp->qp_num, attr->port_num, cur_state, 1891 new_state, ibqp->qp_type); 1892 goto out; 1893 } 1894 1895 if ((attr_mask & IB_QP_PORT) && (ibqp->qp_type == IB_QPT_RAW_PACKET) && 1896 (rdma_port_get_link_layer(&dev->ib_dev, attr->port_num) != 1897 IB_LINK_LAYER_ETHERNET)) 1898 goto out; 1899 1900 if (attr_mask & IB_QP_PKEY_INDEX) { 1901 int p = attr_mask & IB_QP_PORT ? attr->port_num : qp->port; 1902 if (attr->pkey_index >= dev->dev->caps.pkey_table_len[p]) { 1903 pr_debug("qpn 0x%x: invalid pkey index (%d) specified " 1904 "for transition %d to %d. qp_type %d\n", 1905 ibqp->qp_num, attr->pkey_index, cur_state, 1906 new_state, ibqp->qp_type); 1907 goto out; 1908 } 1909 } 1910 1911 if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC && 1912 attr->max_rd_atomic > dev->dev->caps.max_qp_init_rdma) { 1913 pr_debug("qpn 0x%x: max_rd_atomic (%d) too large. " 1914 "Transition %d to %d. qp_type %d\n", 1915 ibqp->qp_num, attr->max_rd_atomic, cur_state, 1916 new_state, ibqp->qp_type); 1917 goto out; 1918 } 1919 1920 if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC && 1921 attr->max_dest_rd_atomic > dev->dev->caps.max_qp_dest_rdma) { 1922 pr_debug("qpn 0x%x: max_dest_rd_atomic (%d) too large. " 1923 "Transition %d to %d. qp_type %d\n", 1924 ibqp->qp_num, attr->max_dest_rd_atomic, cur_state, 1925 new_state, ibqp->qp_type); 1926 goto out; 1927 } 1928 1929 if (cur_state == new_state && cur_state == IB_QPS_RESET) { 1930 err = 0; 1931 goto out; 1932 } 1933 1934 err = __mlx4_ib_modify_qp(ibqp, attr, attr_mask, cur_state, new_state); 1935 1936 out: 1937 mutex_unlock(&qp->mutex); 1938 return err; 1939 } 1940 1941 static int build_sriov_qp0_header(struct mlx4_ib_sqp *sqp, 1942 struct ib_send_wr *wr, 1943 void *wqe, unsigned *mlx_seg_len) 1944 { 1945 struct mlx4_ib_dev *mdev = to_mdev(sqp->qp.ibqp.device); 1946 struct ib_device *ib_dev = &mdev->ib_dev; 1947 struct mlx4_wqe_mlx_seg *mlx = wqe; 1948 struct mlx4_wqe_inline_seg *inl = wqe + sizeof *mlx; 1949 struct mlx4_ib_ah *ah = to_mah(wr->wr.ud.ah); 1950 u16 pkey; 1951 u32 qkey; 1952 int send_size; 1953 int header_size; 1954 int spc; 1955 int i; 1956 1957 if (wr->opcode != IB_WR_SEND) 1958 return -EINVAL; 1959 1960 send_size = 0; 1961 1962 for (i = 0; i < wr->num_sge; ++i) 1963 send_size += wr->sg_list[i].length; 1964 1965 /* for proxy-qp0 sends, need to add in size of tunnel header */ 1966 /* for tunnel-qp0 sends, tunnel header is already in s/g list */ 1967 if (sqp->qp.mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_SMI_OWNER) 1968 send_size += sizeof (struct mlx4_ib_tunnel_header); 1969 1970 ib_ud_header_init(send_size, 1, 0, 0, 0, 0, &sqp->ud_header); 1971 1972 if (sqp->qp.mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_SMI_OWNER) { 1973 sqp->ud_header.lrh.service_level = 1974 be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 28; 1975 sqp->ud_header.lrh.destination_lid = 1976 cpu_to_be16(ah->av.ib.g_slid & 0x7f); 1977 sqp->ud_header.lrh.source_lid = 1978 cpu_to_be16(ah->av.ib.g_slid & 0x7f); 1979 } 1980 1981 mlx->flags &= cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE); 1982 1983 /* force loopback */ 1984 mlx->flags |= cpu_to_be32(MLX4_WQE_MLX_VL15 | 0x1 | MLX4_WQE_MLX_SLR); 1985 mlx->rlid = sqp->ud_header.lrh.destination_lid; 1986 1987 sqp->ud_header.lrh.virtual_lane = 0; 1988 sqp->ud_header.bth.solicited_event = !!(wr->send_flags & IB_SEND_SOLICITED); 1989 ib_get_cached_pkey(ib_dev, sqp->qp.port, 0, &pkey); 1990 sqp->ud_header.bth.pkey = cpu_to_be16(pkey); 1991 if (sqp->qp.mlx4_ib_qp_type == MLX4_IB_QPT_TUN_SMI_OWNER) 1992 sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->wr.ud.remote_qpn); 1993 else 1994 sqp->ud_header.bth.destination_qpn = 1995 cpu_to_be32(mdev->dev->caps.qp0_tunnel[sqp->qp.port - 1]); 1996 1997 sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1)); 1998 if (mlx4_get_parav_qkey(mdev->dev, sqp->qp.mqp.qpn, &qkey)) 1999 return -EINVAL; 2000 sqp->ud_header.deth.qkey = cpu_to_be32(qkey); 2001 sqp->ud_header.deth.source_qpn = cpu_to_be32(sqp->qp.mqp.qpn); 2002 2003 sqp->ud_header.bth.opcode = IB_OPCODE_UD_SEND_ONLY; 2004 sqp->ud_header.immediate_present = 0; 2005 2006 header_size = ib_ud_header_pack(&sqp->ud_header, sqp->header_buf); 2007 2008 /* 2009 * Inline data segments may not cross a 64 byte boundary. If 2010 * our UD header is bigger than the space available up to the 2011 * next 64 byte boundary in the WQE, use two inline data 2012 * segments to hold the UD header. 2013 */ 2014 spc = MLX4_INLINE_ALIGN - 2015 ((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1)); 2016 if (header_size <= spc) { 2017 inl->byte_count = cpu_to_be32(1 << 31 | header_size); 2018 memcpy(inl + 1, sqp->header_buf, header_size); 2019 i = 1; 2020 } else { 2021 inl->byte_count = cpu_to_be32(1 << 31 | spc); 2022 memcpy(inl + 1, sqp->header_buf, spc); 2023 2024 inl = (void *) (inl + 1) + spc; 2025 memcpy(inl + 1, sqp->header_buf + spc, header_size - spc); 2026 /* 2027 * Need a barrier here to make sure all the data is 2028 * visible before the byte_count field is set. 2029 * Otherwise the HCA prefetcher could grab the 64-byte 2030 * chunk with this inline segment and get a valid (!= 2031 * 0xffffffff) byte count but stale data, and end up 2032 * generating a packet with bad headers. 2033 * 2034 * The first inline segment's byte_count field doesn't 2035 * need a barrier, because it comes after a 2036 * control/MLX segment and therefore is at an offset 2037 * of 16 mod 64. 2038 */ 2039 wmb(); 2040 inl->byte_count = cpu_to_be32(1 << 31 | (header_size - spc)); 2041 i = 2; 2042 } 2043 2044 *mlx_seg_len = 2045 ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + header_size, 16); 2046 return 0; 2047 } 2048 2049 static int build_mlx_header(struct mlx4_ib_sqp *sqp, struct ib_send_wr *wr, 2050 void *wqe, unsigned *mlx_seg_len) 2051 { 2052 struct ib_device *ib_dev = sqp->qp.ibqp.device; 2053 struct mlx4_wqe_mlx_seg *mlx = wqe; 2054 struct mlx4_wqe_ctrl_seg *ctrl = wqe; 2055 struct mlx4_wqe_inline_seg *inl = wqe + sizeof *mlx; 2056 struct mlx4_ib_ah *ah = to_mah(wr->wr.ud.ah); 2057 union ib_gid sgid; 2058 u16 pkey; 2059 int send_size; 2060 int header_size; 2061 int spc; 2062 int i; 2063 int err = 0; 2064 u16 vlan = 0xffff; 2065 bool is_eth; 2066 bool is_vlan = false; 2067 bool is_grh; 2068 2069 send_size = 0; 2070 for (i = 0; i < wr->num_sge; ++i) 2071 send_size += wr->sg_list[i].length; 2072 2073 is_eth = rdma_port_get_link_layer(sqp->qp.ibqp.device, sqp->qp.port) == IB_LINK_LAYER_ETHERNET; 2074 is_grh = mlx4_ib_ah_grh_present(ah); 2075 if (is_eth) { 2076 if (mlx4_is_mfunc(to_mdev(ib_dev)->dev)) { 2077 /* When multi-function is enabled, the ib_core gid 2078 * indexes don't necessarily match the hw ones, so 2079 * we must use our own cache */ 2080 err = mlx4_get_roce_gid_from_slave(to_mdev(ib_dev)->dev, 2081 be32_to_cpu(ah->av.ib.port_pd) >> 24, 2082 ah->av.ib.gid_index, &sgid.raw[0]); 2083 if (err) 2084 return err; 2085 } else { 2086 err = ib_get_cached_gid(ib_dev, 2087 be32_to_cpu(ah->av.ib.port_pd) >> 24, 2088 ah->av.ib.gid_index, &sgid); 2089 if (err) 2090 return err; 2091 } 2092 2093 if (ah->av.eth.vlan != cpu_to_be16(0xffff)) { 2094 vlan = be16_to_cpu(ah->av.eth.vlan) & 0x0fff; 2095 is_vlan = 1; 2096 } 2097 } 2098 ib_ud_header_init(send_size, !is_eth, is_eth, is_vlan, is_grh, 0, &sqp->ud_header); 2099 2100 if (!is_eth) { 2101 sqp->ud_header.lrh.service_level = 2102 be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 28; 2103 sqp->ud_header.lrh.destination_lid = ah->av.ib.dlid; 2104 sqp->ud_header.lrh.source_lid = cpu_to_be16(ah->av.ib.g_slid & 0x7f); 2105 } 2106 2107 if (is_grh) { 2108 sqp->ud_header.grh.traffic_class = 2109 (be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 20) & 0xff; 2110 sqp->ud_header.grh.flow_label = 2111 ah->av.ib.sl_tclass_flowlabel & cpu_to_be32(0xfffff); 2112 sqp->ud_header.grh.hop_limit = ah->av.ib.hop_limit; 2113 if (is_eth) 2114 memcpy(sqp->ud_header.grh.source_gid.raw, sgid.raw, 16); 2115 else { 2116 if (mlx4_is_mfunc(to_mdev(ib_dev)->dev)) { 2117 /* When multi-function is enabled, the ib_core gid 2118 * indexes don't necessarily match the hw ones, so 2119 * we must use our own cache */ 2120 sqp->ud_header.grh.source_gid.global.subnet_prefix = 2121 to_mdev(ib_dev)->sriov.demux[sqp->qp.port - 1]. 2122 subnet_prefix; 2123 sqp->ud_header.grh.source_gid.global.interface_id = 2124 to_mdev(ib_dev)->sriov.demux[sqp->qp.port - 1]. 2125 guid_cache[ah->av.ib.gid_index]; 2126 } else 2127 ib_get_cached_gid(ib_dev, 2128 be32_to_cpu(ah->av.ib.port_pd) >> 24, 2129 ah->av.ib.gid_index, 2130 &sqp->ud_header.grh.source_gid); 2131 } 2132 memcpy(sqp->ud_header.grh.destination_gid.raw, 2133 ah->av.ib.dgid, 16); 2134 } 2135 2136 mlx->flags &= cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE); 2137 2138 if (!is_eth) { 2139 mlx->flags |= cpu_to_be32((!sqp->qp.ibqp.qp_num ? MLX4_WQE_MLX_VL15 : 0) | 2140 (sqp->ud_header.lrh.destination_lid == 2141 IB_LID_PERMISSIVE ? MLX4_WQE_MLX_SLR : 0) | 2142 (sqp->ud_header.lrh.service_level << 8)); 2143 if (ah->av.ib.port_pd & cpu_to_be32(0x80000000)) 2144 mlx->flags |= cpu_to_be32(0x1); /* force loopback */ 2145 mlx->rlid = sqp->ud_header.lrh.destination_lid; 2146 } 2147 2148 switch (wr->opcode) { 2149 case IB_WR_SEND: 2150 sqp->ud_header.bth.opcode = IB_OPCODE_UD_SEND_ONLY; 2151 sqp->ud_header.immediate_present = 0; 2152 break; 2153 case IB_WR_SEND_WITH_IMM: 2154 sqp->ud_header.bth.opcode = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE; 2155 sqp->ud_header.immediate_present = 1; 2156 sqp->ud_header.immediate_data = wr->ex.imm_data; 2157 break; 2158 default: 2159 return -EINVAL; 2160 } 2161 2162 if (is_eth) { 2163 u8 *smac; 2164 struct in6_addr in6; 2165 2166 u16 pcp = (be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 29) << 13; 2167 2168 mlx->sched_prio = cpu_to_be16(pcp); 2169 2170 memcpy(sqp->ud_header.eth.dmac_h, ah->av.eth.mac, 6); 2171 /* FIXME: cache smac value? */ 2172 memcpy(&ctrl->srcrb_flags16[0], ah->av.eth.mac, 2); 2173 memcpy(&ctrl->imm, ah->av.eth.mac + 2, 4); 2174 memcpy(&in6, sgid.raw, sizeof(in6)); 2175 2176 if (!mlx4_is_mfunc(to_mdev(ib_dev)->dev)) 2177 smac = to_mdev(sqp->qp.ibqp.device)-> 2178 iboe.netdevs[sqp->qp.port - 1]->dev_addr; 2179 else /* use the src mac of the tunnel */ 2180 smac = ah->av.eth.s_mac; 2181 memcpy(sqp->ud_header.eth.smac_h, smac, 6); 2182 if (!memcmp(sqp->ud_header.eth.smac_h, sqp->ud_header.eth.dmac_h, 6)) 2183 mlx->flags |= cpu_to_be32(MLX4_WQE_CTRL_FORCE_LOOPBACK); 2184 if (!is_vlan) { 2185 sqp->ud_header.eth.type = cpu_to_be16(MLX4_IB_IBOE_ETHERTYPE); 2186 } else { 2187 sqp->ud_header.vlan.type = cpu_to_be16(MLX4_IB_IBOE_ETHERTYPE); 2188 sqp->ud_header.vlan.tag = cpu_to_be16(vlan | pcp); 2189 } 2190 } else { 2191 sqp->ud_header.lrh.virtual_lane = !sqp->qp.ibqp.qp_num ? 15 : 0; 2192 if (sqp->ud_header.lrh.destination_lid == IB_LID_PERMISSIVE) 2193 sqp->ud_header.lrh.source_lid = IB_LID_PERMISSIVE; 2194 } 2195 sqp->ud_header.bth.solicited_event = !!(wr->send_flags & IB_SEND_SOLICITED); 2196 if (!sqp->qp.ibqp.qp_num) 2197 ib_get_cached_pkey(ib_dev, sqp->qp.port, sqp->pkey_index, &pkey); 2198 else 2199 ib_get_cached_pkey(ib_dev, sqp->qp.port, wr->wr.ud.pkey_index, &pkey); 2200 sqp->ud_header.bth.pkey = cpu_to_be16(pkey); 2201 sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->wr.ud.remote_qpn); 2202 sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1)); 2203 sqp->ud_header.deth.qkey = cpu_to_be32(wr->wr.ud.remote_qkey & 0x80000000 ? 2204 sqp->qkey : wr->wr.ud.remote_qkey); 2205 sqp->ud_header.deth.source_qpn = cpu_to_be32(sqp->qp.ibqp.qp_num); 2206 2207 header_size = ib_ud_header_pack(&sqp->ud_header, sqp->header_buf); 2208 2209 if (0) { 2210 pr_err("built UD header of size %d:\n", header_size); 2211 for (i = 0; i < header_size / 4; ++i) { 2212 if (i % 8 == 0) 2213 pr_err(" [%02x] ", i * 4); 2214 pr_cont(" %08x", 2215 be32_to_cpu(((__be32 *) sqp->header_buf)[i])); 2216 if ((i + 1) % 8 == 0) 2217 pr_cont("\n"); 2218 } 2219 pr_err("\n"); 2220 } 2221 2222 /* 2223 * Inline data segments may not cross a 64 byte boundary. If 2224 * our UD header is bigger than the space available up to the 2225 * next 64 byte boundary in the WQE, use two inline data 2226 * segments to hold the UD header. 2227 */ 2228 spc = MLX4_INLINE_ALIGN - 2229 ((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1)); 2230 if (header_size <= spc) { 2231 inl->byte_count = cpu_to_be32(1 << 31 | header_size); 2232 memcpy(inl + 1, sqp->header_buf, header_size); 2233 i = 1; 2234 } else { 2235 inl->byte_count = cpu_to_be32(1 << 31 | spc); 2236 memcpy(inl + 1, sqp->header_buf, spc); 2237 2238 inl = (void *) (inl + 1) + spc; 2239 memcpy(inl + 1, sqp->header_buf + spc, header_size - spc); 2240 /* 2241 * Need a barrier here to make sure all the data is 2242 * visible before the byte_count field is set. 2243 * Otherwise the HCA prefetcher could grab the 64-byte 2244 * chunk with this inline segment and get a valid (!= 2245 * 0xffffffff) byte count but stale data, and end up 2246 * generating a packet with bad headers. 2247 * 2248 * The first inline segment's byte_count field doesn't 2249 * need a barrier, because it comes after a 2250 * control/MLX segment and therefore is at an offset 2251 * of 16 mod 64. 2252 */ 2253 wmb(); 2254 inl->byte_count = cpu_to_be32(1 << 31 | (header_size - spc)); 2255 i = 2; 2256 } 2257 2258 *mlx_seg_len = 2259 ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + header_size, 16); 2260 return 0; 2261 } 2262 2263 static int mlx4_wq_overflow(struct mlx4_ib_wq *wq, int nreq, struct ib_cq *ib_cq) 2264 { 2265 unsigned cur; 2266 struct mlx4_ib_cq *cq; 2267 2268 cur = wq->head - wq->tail; 2269 if (likely(cur + nreq < wq->max_post)) 2270 return 0; 2271 2272 cq = to_mcq(ib_cq); 2273 spin_lock(&cq->lock); 2274 cur = wq->head - wq->tail; 2275 spin_unlock(&cq->lock); 2276 2277 return cur + nreq >= wq->max_post; 2278 } 2279 2280 static __be32 convert_access(int acc) 2281 { 2282 return (acc & IB_ACCESS_REMOTE_ATOMIC ? 2283 cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_ATOMIC) : 0) | 2284 (acc & IB_ACCESS_REMOTE_WRITE ? 2285 cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_REMOTE_WRITE) : 0) | 2286 (acc & IB_ACCESS_REMOTE_READ ? 2287 cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_REMOTE_READ) : 0) | 2288 (acc & IB_ACCESS_LOCAL_WRITE ? cpu_to_be32(MLX4_WQE_FMR_PERM_LOCAL_WRITE) : 0) | 2289 cpu_to_be32(MLX4_WQE_FMR_PERM_LOCAL_READ); 2290 } 2291 2292 static void set_fmr_seg(struct mlx4_wqe_fmr_seg *fseg, struct ib_send_wr *wr) 2293 { 2294 struct mlx4_ib_fast_reg_page_list *mfrpl = to_mfrpl(wr->wr.fast_reg.page_list); 2295 int i; 2296 2297 for (i = 0; i < wr->wr.fast_reg.page_list_len; ++i) 2298 mfrpl->mapped_page_list[i] = 2299 cpu_to_be64(wr->wr.fast_reg.page_list->page_list[i] | 2300 MLX4_MTT_FLAG_PRESENT); 2301 2302 fseg->flags = convert_access(wr->wr.fast_reg.access_flags); 2303 fseg->mem_key = cpu_to_be32(wr->wr.fast_reg.rkey); 2304 fseg->buf_list = cpu_to_be64(mfrpl->map); 2305 fseg->start_addr = cpu_to_be64(wr->wr.fast_reg.iova_start); 2306 fseg->reg_len = cpu_to_be64(wr->wr.fast_reg.length); 2307 fseg->offset = 0; /* XXX -- is this just for ZBVA? */ 2308 fseg->page_size = cpu_to_be32(wr->wr.fast_reg.page_shift); 2309 fseg->reserved[0] = 0; 2310 fseg->reserved[1] = 0; 2311 } 2312 2313 static void set_bind_seg(struct mlx4_wqe_bind_seg *bseg, struct ib_send_wr *wr) 2314 { 2315 bseg->flags1 = 2316 convert_access(wr->wr.bind_mw.bind_info.mw_access_flags) & 2317 cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_REMOTE_READ | 2318 MLX4_WQE_FMR_AND_BIND_PERM_REMOTE_WRITE | 2319 MLX4_WQE_FMR_AND_BIND_PERM_ATOMIC); 2320 bseg->flags2 = 0; 2321 if (wr->wr.bind_mw.mw->type == IB_MW_TYPE_2) 2322 bseg->flags2 |= cpu_to_be32(MLX4_WQE_BIND_TYPE_2); 2323 if (wr->wr.bind_mw.bind_info.mw_access_flags & IB_ZERO_BASED) 2324 bseg->flags2 |= cpu_to_be32(MLX4_WQE_BIND_ZERO_BASED); 2325 bseg->new_rkey = cpu_to_be32(wr->wr.bind_mw.rkey); 2326 bseg->lkey = cpu_to_be32(wr->wr.bind_mw.bind_info.mr->lkey); 2327 bseg->addr = cpu_to_be64(wr->wr.bind_mw.bind_info.addr); 2328 bseg->length = cpu_to_be64(wr->wr.bind_mw.bind_info.length); 2329 } 2330 2331 static void set_local_inv_seg(struct mlx4_wqe_local_inval_seg *iseg, u32 rkey) 2332 { 2333 memset(iseg, 0, sizeof(*iseg)); 2334 iseg->mem_key = cpu_to_be32(rkey); 2335 } 2336 2337 static __always_inline void set_raddr_seg(struct mlx4_wqe_raddr_seg *rseg, 2338 u64 remote_addr, u32 rkey) 2339 { 2340 rseg->raddr = cpu_to_be64(remote_addr); 2341 rseg->rkey = cpu_to_be32(rkey); 2342 rseg->reserved = 0; 2343 } 2344 2345 static void set_atomic_seg(struct mlx4_wqe_atomic_seg *aseg, struct ib_send_wr *wr) 2346 { 2347 if (wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP) { 2348 aseg->swap_add = cpu_to_be64(wr->wr.atomic.swap); 2349 aseg->compare = cpu_to_be64(wr->wr.atomic.compare_add); 2350 } else if (wr->opcode == IB_WR_MASKED_ATOMIC_FETCH_AND_ADD) { 2351 aseg->swap_add = cpu_to_be64(wr->wr.atomic.compare_add); 2352 aseg->compare = cpu_to_be64(wr->wr.atomic.compare_add_mask); 2353 } else { 2354 aseg->swap_add = cpu_to_be64(wr->wr.atomic.compare_add); 2355 aseg->compare = 0; 2356 } 2357 2358 } 2359 2360 static void set_masked_atomic_seg(struct mlx4_wqe_masked_atomic_seg *aseg, 2361 struct ib_send_wr *wr) 2362 { 2363 aseg->swap_add = cpu_to_be64(wr->wr.atomic.swap); 2364 aseg->swap_add_mask = cpu_to_be64(wr->wr.atomic.swap_mask); 2365 aseg->compare = cpu_to_be64(wr->wr.atomic.compare_add); 2366 aseg->compare_mask = cpu_to_be64(wr->wr.atomic.compare_add_mask); 2367 } 2368 2369 static void set_datagram_seg(struct mlx4_wqe_datagram_seg *dseg, 2370 struct ib_send_wr *wr) 2371 { 2372 memcpy(dseg->av, &to_mah(wr->wr.ud.ah)->av, sizeof (struct mlx4_av)); 2373 dseg->dqpn = cpu_to_be32(wr->wr.ud.remote_qpn); 2374 dseg->qkey = cpu_to_be32(wr->wr.ud.remote_qkey); 2375 dseg->vlan = to_mah(wr->wr.ud.ah)->av.eth.vlan; 2376 memcpy(dseg->mac, to_mah(wr->wr.ud.ah)->av.eth.mac, 6); 2377 } 2378 2379 static void set_tunnel_datagram_seg(struct mlx4_ib_dev *dev, 2380 struct mlx4_wqe_datagram_seg *dseg, 2381 struct ib_send_wr *wr, enum ib_qp_type qpt) 2382 { 2383 union mlx4_ext_av *av = &to_mah(wr->wr.ud.ah)->av; 2384 struct mlx4_av sqp_av = {0}; 2385 int port = *((u8 *) &av->ib.port_pd) & 0x3; 2386 2387 /* force loopback */ 2388 sqp_av.port_pd = av->ib.port_pd | cpu_to_be32(0x80000000); 2389 sqp_av.g_slid = av->ib.g_slid & 0x7f; /* no GRH */ 2390 sqp_av.sl_tclass_flowlabel = av->ib.sl_tclass_flowlabel & 2391 cpu_to_be32(0xf0000000); 2392 2393 memcpy(dseg->av, &sqp_av, sizeof (struct mlx4_av)); 2394 /* This function used only for sending on QP1 proxies */ 2395 dseg->dqpn = cpu_to_be32(dev->dev->caps.qp1_tunnel[port - 1]); 2396 /* Use QKEY from the QP context, which is set by master */ 2397 dseg->qkey = cpu_to_be32(IB_QP_SET_QKEY); 2398 } 2399 2400 static void build_tunnel_header(struct ib_send_wr *wr, void *wqe, unsigned *mlx_seg_len) 2401 { 2402 struct mlx4_wqe_inline_seg *inl = wqe; 2403 struct mlx4_ib_tunnel_header hdr; 2404 struct mlx4_ib_ah *ah = to_mah(wr->wr.ud.ah); 2405 int spc; 2406 int i; 2407 2408 memcpy(&hdr.av, &ah->av, sizeof hdr.av); 2409 hdr.remote_qpn = cpu_to_be32(wr->wr.ud.remote_qpn); 2410 hdr.pkey_index = cpu_to_be16(wr->wr.ud.pkey_index); 2411 hdr.qkey = cpu_to_be32(wr->wr.ud.remote_qkey); 2412 memcpy(hdr.mac, ah->av.eth.mac, 6); 2413 hdr.vlan = ah->av.eth.vlan; 2414 2415 spc = MLX4_INLINE_ALIGN - 2416 ((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1)); 2417 if (sizeof (hdr) <= spc) { 2418 memcpy(inl + 1, &hdr, sizeof (hdr)); 2419 wmb(); 2420 inl->byte_count = cpu_to_be32(1 << 31 | sizeof (hdr)); 2421 i = 1; 2422 } else { 2423 memcpy(inl + 1, &hdr, spc); 2424 wmb(); 2425 inl->byte_count = cpu_to_be32(1 << 31 | spc); 2426 2427 inl = (void *) (inl + 1) + spc; 2428 memcpy(inl + 1, (void *) &hdr + spc, sizeof (hdr) - spc); 2429 wmb(); 2430 inl->byte_count = cpu_to_be32(1 << 31 | (sizeof (hdr) - spc)); 2431 i = 2; 2432 } 2433 2434 *mlx_seg_len = 2435 ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + sizeof (hdr), 16); 2436 } 2437 2438 static void set_mlx_icrc_seg(void *dseg) 2439 { 2440 u32 *t = dseg; 2441 struct mlx4_wqe_inline_seg *iseg = dseg; 2442 2443 t[1] = 0; 2444 2445 /* 2446 * Need a barrier here before writing the byte_count field to 2447 * make sure that all the data is visible before the 2448 * byte_count field is set. Otherwise, if the segment begins 2449 * a new cacheline, the HCA prefetcher could grab the 64-byte 2450 * chunk and get a valid (!= * 0xffffffff) byte count but 2451 * stale data, and end up sending the wrong data. 2452 */ 2453 wmb(); 2454 2455 iseg->byte_count = cpu_to_be32((1 << 31) | 4); 2456 } 2457 2458 static void set_data_seg(struct mlx4_wqe_data_seg *dseg, struct ib_sge *sg) 2459 { 2460 dseg->lkey = cpu_to_be32(sg->lkey); 2461 dseg->addr = cpu_to_be64(sg->addr); 2462 2463 /* 2464 * Need a barrier here before writing the byte_count field to 2465 * make sure that all the data is visible before the 2466 * byte_count field is set. Otherwise, if the segment begins 2467 * a new cacheline, the HCA prefetcher could grab the 64-byte 2468 * chunk and get a valid (!= * 0xffffffff) byte count but 2469 * stale data, and end up sending the wrong data. 2470 */ 2471 wmb(); 2472 2473 dseg->byte_count = cpu_to_be32(sg->length); 2474 } 2475 2476 static void __set_data_seg(struct mlx4_wqe_data_seg *dseg, struct ib_sge *sg) 2477 { 2478 dseg->byte_count = cpu_to_be32(sg->length); 2479 dseg->lkey = cpu_to_be32(sg->lkey); 2480 dseg->addr = cpu_to_be64(sg->addr); 2481 } 2482 2483 static int build_lso_seg(struct mlx4_wqe_lso_seg *wqe, struct ib_send_wr *wr, 2484 struct mlx4_ib_qp *qp, unsigned *lso_seg_len, 2485 __be32 *lso_hdr_sz, __be32 *blh) 2486 { 2487 unsigned halign = ALIGN(sizeof *wqe + wr->wr.ud.hlen, 16); 2488 2489 if (unlikely(halign > MLX4_IB_CACHE_LINE_SIZE)) 2490 *blh = cpu_to_be32(1 << 6); 2491 2492 if (unlikely(!(qp->flags & MLX4_IB_QP_LSO) && 2493 wr->num_sge > qp->sq.max_gs - (halign >> 4))) 2494 return -EINVAL; 2495 2496 memcpy(wqe->header, wr->wr.ud.header, wr->wr.ud.hlen); 2497 2498 *lso_hdr_sz = cpu_to_be32((wr->wr.ud.mss - wr->wr.ud.hlen) << 16 | 2499 wr->wr.ud.hlen); 2500 *lso_seg_len = halign; 2501 return 0; 2502 } 2503 2504 static __be32 send_ieth(struct ib_send_wr *wr) 2505 { 2506 switch (wr->opcode) { 2507 case IB_WR_SEND_WITH_IMM: 2508 case IB_WR_RDMA_WRITE_WITH_IMM: 2509 return wr->ex.imm_data; 2510 2511 case IB_WR_SEND_WITH_INV: 2512 return cpu_to_be32(wr->ex.invalidate_rkey); 2513 2514 default: 2515 return 0; 2516 } 2517 } 2518 2519 static void add_zero_len_inline(void *wqe) 2520 { 2521 struct mlx4_wqe_inline_seg *inl = wqe; 2522 memset(wqe, 0, 16); 2523 inl->byte_count = cpu_to_be32(1 << 31); 2524 } 2525 2526 int mlx4_ib_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr, 2527 struct ib_send_wr **bad_wr) 2528 { 2529 struct mlx4_ib_qp *qp = to_mqp(ibqp); 2530 void *wqe; 2531 struct mlx4_wqe_ctrl_seg *ctrl; 2532 struct mlx4_wqe_data_seg *dseg; 2533 unsigned long flags; 2534 int nreq; 2535 int err = 0; 2536 unsigned ind; 2537 int uninitialized_var(stamp); 2538 int uninitialized_var(size); 2539 unsigned uninitialized_var(seglen); 2540 __be32 dummy; 2541 __be32 *lso_wqe; 2542 __be32 uninitialized_var(lso_hdr_sz); 2543 __be32 blh; 2544 int i; 2545 2546 spin_lock_irqsave(&qp->sq.lock, flags); 2547 2548 ind = qp->sq_next_wqe; 2549 2550 for (nreq = 0; wr; ++nreq, wr = wr->next) { 2551 lso_wqe = &dummy; 2552 blh = 0; 2553 2554 if (mlx4_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) { 2555 err = -ENOMEM; 2556 *bad_wr = wr; 2557 goto out; 2558 } 2559 2560 if (unlikely(wr->num_sge > qp->sq.max_gs)) { 2561 err = -EINVAL; 2562 *bad_wr = wr; 2563 goto out; 2564 } 2565 2566 ctrl = wqe = get_send_wqe(qp, ind & (qp->sq.wqe_cnt - 1)); 2567 qp->sq.wrid[(qp->sq.head + nreq) & (qp->sq.wqe_cnt - 1)] = wr->wr_id; 2568 2569 ctrl->srcrb_flags = 2570 (wr->send_flags & IB_SEND_SIGNALED ? 2571 cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE) : 0) | 2572 (wr->send_flags & IB_SEND_SOLICITED ? 2573 cpu_to_be32(MLX4_WQE_CTRL_SOLICITED) : 0) | 2574 ((wr->send_flags & IB_SEND_IP_CSUM) ? 2575 cpu_to_be32(MLX4_WQE_CTRL_IP_CSUM | 2576 MLX4_WQE_CTRL_TCP_UDP_CSUM) : 0) | 2577 qp->sq_signal_bits; 2578 2579 ctrl->imm = send_ieth(wr); 2580 2581 wqe += sizeof *ctrl; 2582 size = sizeof *ctrl / 16; 2583 2584 switch (qp->mlx4_ib_qp_type) { 2585 case MLX4_IB_QPT_RC: 2586 case MLX4_IB_QPT_UC: 2587 switch (wr->opcode) { 2588 case IB_WR_ATOMIC_CMP_AND_SWP: 2589 case IB_WR_ATOMIC_FETCH_AND_ADD: 2590 case IB_WR_MASKED_ATOMIC_FETCH_AND_ADD: 2591 set_raddr_seg(wqe, wr->wr.atomic.remote_addr, 2592 wr->wr.atomic.rkey); 2593 wqe += sizeof (struct mlx4_wqe_raddr_seg); 2594 2595 set_atomic_seg(wqe, wr); 2596 wqe += sizeof (struct mlx4_wqe_atomic_seg); 2597 2598 size += (sizeof (struct mlx4_wqe_raddr_seg) + 2599 sizeof (struct mlx4_wqe_atomic_seg)) / 16; 2600 2601 break; 2602 2603 case IB_WR_MASKED_ATOMIC_CMP_AND_SWP: 2604 set_raddr_seg(wqe, wr->wr.atomic.remote_addr, 2605 wr->wr.atomic.rkey); 2606 wqe += sizeof (struct mlx4_wqe_raddr_seg); 2607 2608 set_masked_atomic_seg(wqe, wr); 2609 wqe += sizeof (struct mlx4_wqe_masked_atomic_seg); 2610 2611 size += (sizeof (struct mlx4_wqe_raddr_seg) + 2612 sizeof (struct mlx4_wqe_masked_atomic_seg)) / 16; 2613 2614 break; 2615 2616 case IB_WR_RDMA_READ: 2617 case IB_WR_RDMA_WRITE: 2618 case IB_WR_RDMA_WRITE_WITH_IMM: 2619 set_raddr_seg(wqe, wr->wr.rdma.remote_addr, 2620 wr->wr.rdma.rkey); 2621 wqe += sizeof (struct mlx4_wqe_raddr_seg); 2622 size += sizeof (struct mlx4_wqe_raddr_seg) / 16; 2623 break; 2624 2625 case IB_WR_LOCAL_INV: 2626 ctrl->srcrb_flags |= 2627 cpu_to_be32(MLX4_WQE_CTRL_STRONG_ORDER); 2628 set_local_inv_seg(wqe, wr->ex.invalidate_rkey); 2629 wqe += sizeof (struct mlx4_wqe_local_inval_seg); 2630 size += sizeof (struct mlx4_wqe_local_inval_seg) / 16; 2631 break; 2632 2633 case IB_WR_FAST_REG_MR: 2634 ctrl->srcrb_flags |= 2635 cpu_to_be32(MLX4_WQE_CTRL_STRONG_ORDER); 2636 set_fmr_seg(wqe, wr); 2637 wqe += sizeof (struct mlx4_wqe_fmr_seg); 2638 size += sizeof (struct mlx4_wqe_fmr_seg) / 16; 2639 break; 2640 2641 case IB_WR_BIND_MW: 2642 ctrl->srcrb_flags |= 2643 cpu_to_be32(MLX4_WQE_CTRL_STRONG_ORDER); 2644 set_bind_seg(wqe, wr); 2645 wqe += sizeof(struct mlx4_wqe_bind_seg); 2646 size += sizeof(struct mlx4_wqe_bind_seg) / 16; 2647 break; 2648 default: 2649 /* No extra segments required for sends */ 2650 break; 2651 } 2652 break; 2653 2654 case MLX4_IB_QPT_TUN_SMI_OWNER: 2655 err = build_sriov_qp0_header(to_msqp(qp), wr, ctrl, &seglen); 2656 if (unlikely(err)) { 2657 *bad_wr = wr; 2658 goto out; 2659 } 2660 wqe += seglen; 2661 size += seglen / 16; 2662 break; 2663 case MLX4_IB_QPT_TUN_SMI: 2664 case MLX4_IB_QPT_TUN_GSI: 2665 /* this is a UD qp used in MAD responses to slaves. */ 2666 set_datagram_seg(wqe, wr); 2667 /* set the forced-loopback bit in the data seg av */ 2668 *(__be32 *) wqe |= cpu_to_be32(0x80000000); 2669 wqe += sizeof (struct mlx4_wqe_datagram_seg); 2670 size += sizeof (struct mlx4_wqe_datagram_seg) / 16; 2671 break; 2672 case MLX4_IB_QPT_UD: 2673 set_datagram_seg(wqe, wr); 2674 wqe += sizeof (struct mlx4_wqe_datagram_seg); 2675 size += sizeof (struct mlx4_wqe_datagram_seg) / 16; 2676 2677 if (wr->opcode == IB_WR_LSO) { 2678 err = build_lso_seg(wqe, wr, qp, &seglen, &lso_hdr_sz, &blh); 2679 if (unlikely(err)) { 2680 *bad_wr = wr; 2681 goto out; 2682 } 2683 lso_wqe = (__be32 *) wqe; 2684 wqe += seglen; 2685 size += seglen / 16; 2686 } 2687 break; 2688 2689 case MLX4_IB_QPT_PROXY_SMI_OWNER: 2690 if (unlikely(!mlx4_is_master(to_mdev(ibqp->device)->dev))) { 2691 err = -ENOSYS; 2692 *bad_wr = wr; 2693 goto out; 2694 } 2695 err = build_sriov_qp0_header(to_msqp(qp), wr, ctrl, &seglen); 2696 if (unlikely(err)) { 2697 *bad_wr = wr; 2698 goto out; 2699 } 2700 wqe += seglen; 2701 size += seglen / 16; 2702 /* to start tunnel header on a cache-line boundary */ 2703 add_zero_len_inline(wqe); 2704 wqe += 16; 2705 size++; 2706 build_tunnel_header(wr, wqe, &seglen); 2707 wqe += seglen; 2708 size += seglen / 16; 2709 break; 2710 case MLX4_IB_QPT_PROXY_SMI: 2711 /* don't allow QP0 sends on guests */ 2712 err = -ENOSYS; 2713 *bad_wr = wr; 2714 goto out; 2715 case MLX4_IB_QPT_PROXY_GSI: 2716 /* If we are tunneling special qps, this is a UD qp. 2717 * In this case we first add a UD segment targeting 2718 * the tunnel qp, and then add a header with address 2719 * information */ 2720 set_tunnel_datagram_seg(to_mdev(ibqp->device), wqe, wr, ibqp->qp_type); 2721 wqe += sizeof (struct mlx4_wqe_datagram_seg); 2722 size += sizeof (struct mlx4_wqe_datagram_seg) / 16; 2723 build_tunnel_header(wr, wqe, &seglen); 2724 wqe += seglen; 2725 size += seglen / 16; 2726 break; 2727 2728 case MLX4_IB_QPT_SMI: 2729 case MLX4_IB_QPT_GSI: 2730 err = build_mlx_header(to_msqp(qp), wr, ctrl, &seglen); 2731 if (unlikely(err)) { 2732 *bad_wr = wr; 2733 goto out; 2734 } 2735 wqe += seglen; 2736 size += seglen / 16; 2737 break; 2738 2739 default: 2740 break; 2741 } 2742 2743 /* 2744 * Write data segments in reverse order, so as to 2745 * overwrite cacheline stamp last within each 2746 * cacheline. This avoids issues with WQE 2747 * prefetching. 2748 */ 2749 2750 dseg = wqe; 2751 dseg += wr->num_sge - 1; 2752 size += wr->num_sge * (sizeof (struct mlx4_wqe_data_seg) / 16); 2753 2754 /* Add one more inline data segment for ICRC for MLX sends */ 2755 if (unlikely(qp->mlx4_ib_qp_type == MLX4_IB_QPT_SMI || 2756 qp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI || 2757 qp->mlx4_ib_qp_type & 2758 (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER))) { 2759 set_mlx_icrc_seg(dseg + 1); 2760 size += sizeof (struct mlx4_wqe_data_seg) / 16; 2761 } 2762 2763 for (i = wr->num_sge - 1; i >= 0; --i, --dseg) 2764 set_data_seg(dseg, wr->sg_list + i); 2765 2766 /* 2767 * Possibly overwrite stamping in cacheline with LSO 2768 * segment only after making sure all data segments 2769 * are written. 2770 */ 2771 wmb(); 2772 *lso_wqe = lso_hdr_sz; 2773 2774 ctrl->fence_size = (wr->send_flags & IB_SEND_FENCE ? 2775 MLX4_WQE_CTRL_FENCE : 0) | size; 2776 2777 /* 2778 * Make sure descriptor is fully written before 2779 * setting ownership bit (because HW can start 2780 * executing as soon as we do). 2781 */ 2782 wmb(); 2783 2784 if (wr->opcode < 0 || wr->opcode >= ARRAY_SIZE(mlx4_ib_opcode)) { 2785 *bad_wr = wr; 2786 err = -EINVAL; 2787 goto out; 2788 } 2789 2790 ctrl->owner_opcode = mlx4_ib_opcode[wr->opcode] | 2791 (ind & qp->sq.wqe_cnt ? cpu_to_be32(1 << 31) : 0) | blh; 2792 2793 stamp = ind + qp->sq_spare_wqes; 2794 ind += DIV_ROUND_UP(size * 16, 1U << qp->sq.wqe_shift); 2795 2796 /* 2797 * We can improve latency by not stamping the last 2798 * send queue WQE until after ringing the doorbell, so 2799 * only stamp here if there are still more WQEs to post. 2800 * 2801 * Same optimization applies to padding with NOP wqe 2802 * in case of WQE shrinking (used to prevent wrap-around 2803 * in the middle of WR). 2804 */ 2805 if (wr->next) { 2806 stamp_send_wqe(qp, stamp, size * 16); 2807 ind = pad_wraparound(qp, ind); 2808 } 2809 } 2810 2811 out: 2812 if (likely(nreq)) { 2813 qp->sq.head += nreq; 2814 2815 /* 2816 * Make sure that descriptors are written before 2817 * doorbell record. 2818 */ 2819 wmb(); 2820 2821 writel(qp->doorbell_qpn, 2822 to_mdev(ibqp->device)->uar_map + MLX4_SEND_DOORBELL); 2823 2824 /* 2825 * Make sure doorbells don't leak out of SQ spinlock 2826 * and reach the HCA out of order. 2827 */ 2828 mmiowb(); 2829 2830 stamp_send_wqe(qp, stamp, size * 16); 2831 2832 ind = pad_wraparound(qp, ind); 2833 qp->sq_next_wqe = ind; 2834 } 2835 2836 spin_unlock_irqrestore(&qp->sq.lock, flags); 2837 2838 return err; 2839 } 2840 2841 int mlx4_ib_post_recv(struct ib_qp *ibqp, struct ib_recv_wr *wr, 2842 struct ib_recv_wr **bad_wr) 2843 { 2844 struct mlx4_ib_qp *qp = to_mqp(ibqp); 2845 struct mlx4_wqe_data_seg *scat; 2846 unsigned long flags; 2847 int err = 0; 2848 int nreq; 2849 int ind; 2850 int max_gs; 2851 int i; 2852 2853 max_gs = qp->rq.max_gs; 2854 spin_lock_irqsave(&qp->rq.lock, flags); 2855 2856 ind = qp->rq.head & (qp->rq.wqe_cnt - 1); 2857 2858 for (nreq = 0; wr; ++nreq, wr = wr->next) { 2859 if (mlx4_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) { 2860 err = -ENOMEM; 2861 *bad_wr = wr; 2862 goto out; 2863 } 2864 2865 if (unlikely(wr->num_sge > qp->rq.max_gs)) { 2866 err = -EINVAL; 2867 *bad_wr = wr; 2868 goto out; 2869 } 2870 2871 scat = get_recv_wqe(qp, ind); 2872 2873 if (qp->mlx4_ib_qp_type & (MLX4_IB_QPT_PROXY_SMI_OWNER | 2874 MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI)) { 2875 ib_dma_sync_single_for_device(ibqp->device, 2876 qp->sqp_proxy_rcv[ind].map, 2877 sizeof (struct mlx4_ib_proxy_sqp_hdr), 2878 DMA_FROM_DEVICE); 2879 scat->byte_count = 2880 cpu_to_be32(sizeof (struct mlx4_ib_proxy_sqp_hdr)); 2881 /* use dma lkey from upper layer entry */ 2882 scat->lkey = cpu_to_be32(wr->sg_list->lkey); 2883 scat->addr = cpu_to_be64(qp->sqp_proxy_rcv[ind].map); 2884 scat++; 2885 max_gs--; 2886 } 2887 2888 for (i = 0; i < wr->num_sge; ++i) 2889 __set_data_seg(scat + i, wr->sg_list + i); 2890 2891 if (i < max_gs) { 2892 scat[i].byte_count = 0; 2893 scat[i].lkey = cpu_to_be32(MLX4_INVALID_LKEY); 2894 scat[i].addr = 0; 2895 } 2896 2897 qp->rq.wrid[ind] = wr->wr_id; 2898 2899 ind = (ind + 1) & (qp->rq.wqe_cnt - 1); 2900 } 2901 2902 out: 2903 if (likely(nreq)) { 2904 qp->rq.head += nreq; 2905 2906 /* 2907 * Make sure that descriptors are written before 2908 * doorbell record. 2909 */ 2910 wmb(); 2911 2912 *qp->db.db = cpu_to_be32(qp->rq.head & 0xffff); 2913 } 2914 2915 spin_unlock_irqrestore(&qp->rq.lock, flags); 2916 2917 return err; 2918 } 2919 2920 static inline enum ib_qp_state to_ib_qp_state(enum mlx4_qp_state mlx4_state) 2921 { 2922 switch (mlx4_state) { 2923 case MLX4_QP_STATE_RST: return IB_QPS_RESET; 2924 case MLX4_QP_STATE_INIT: return IB_QPS_INIT; 2925 case MLX4_QP_STATE_RTR: return IB_QPS_RTR; 2926 case MLX4_QP_STATE_RTS: return IB_QPS_RTS; 2927 case MLX4_QP_STATE_SQ_DRAINING: 2928 case MLX4_QP_STATE_SQD: return IB_QPS_SQD; 2929 case MLX4_QP_STATE_SQER: return IB_QPS_SQE; 2930 case MLX4_QP_STATE_ERR: return IB_QPS_ERR; 2931 default: return -1; 2932 } 2933 } 2934 2935 static inline enum ib_mig_state to_ib_mig_state(int mlx4_mig_state) 2936 { 2937 switch (mlx4_mig_state) { 2938 case MLX4_QP_PM_ARMED: return IB_MIG_ARMED; 2939 case MLX4_QP_PM_REARM: return IB_MIG_REARM; 2940 case MLX4_QP_PM_MIGRATED: return IB_MIG_MIGRATED; 2941 default: return -1; 2942 } 2943 } 2944 2945 static int to_ib_qp_access_flags(int mlx4_flags) 2946 { 2947 int ib_flags = 0; 2948 2949 if (mlx4_flags & MLX4_QP_BIT_RRE) 2950 ib_flags |= IB_ACCESS_REMOTE_READ; 2951 if (mlx4_flags & MLX4_QP_BIT_RWE) 2952 ib_flags |= IB_ACCESS_REMOTE_WRITE; 2953 if (mlx4_flags & MLX4_QP_BIT_RAE) 2954 ib_flags |= IB_ACCESS_REMOTE_ATOMIC; 2955 2956 return ib_flags; 2957 } 2958 2959 static void to_ib_ah_attr(struct mlx4_ib_dev *ibdev, struct ib_ah_attr *ib_ah_attr, 2960 struct mlx4_qp_path *path) 2961 { 2962 struct mlx4_dev *dev = ibdev->dev; 2963 int is_eth; 2964 2965 memset(ib_ah_attr, 0, sizeof *ib_ah_attr); 2966 ib_ah_attr->port_num = path->sched_queue & 0x40 ? 2 : 1; 2967 2968 if (ib_ah_attr->port_num == 0 || ib_ah_attr->port_num > dev->caps.num_ports) 2969 return; 2970 2971 is_eth = rdma_port_get_link_layer(&ibdev->ib_dev, ib_ah_attr->port_num) == 2972 IB_LINK_LAYER_ETHERNET; 2973 if (is_eth) 2974 ib_ah_attr->sl = ((path->sched_queue >> 3) & 0x7) | 2975 ((path->sched_queue & 4) << 1); 2976 else 2977 ib_ah_attr->sl = (path->sched_queue >> 2) & 0xf; 2978 2979 ib_ah_attr->dlid = be16_to_cpu(path->rlid); 2980 ib_ah_attr->src_path_bits = path->grh_mylmc & 0x7f; 2981 ib_ah_attr->static_rate = path->static_rate ? path->static_rate - 5 : 0; 2982 ib_ah_attr->ah_flags = (path->grh_mylmc & (1 << 7)) ? IB_AH_GRH : 0; 2983 if (ib_ah_attr->ah_flags) { 2984 ib_ah_attr->grh.sgid_index = path->mgid_index; 2985 ib_ah_attr->grh.hop_limit = path->hop_limit; 2986 ib_ah_attr->grh.traffic_class = 2987 (be32_to_cpu(path->tclass_flowlabel) >> 20) & 0xff; 2988 ib_ah_attr->grh.flow_label = 2989 be32_to_cpu(path->tclass_flowlabel) & 0xfffff; 2990 memcpy(ib_ah_attr->grh.dgid.raw, 2991 path->rgid, sizeof ib_ah_attr->grh.dgid.raw); 2992 } 2993 } 2994 2995 int mlx4_ib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr, int qp_attr_mask, 2996 struct ib_qp_init_attr *qp_init_attr) 2997 { 2998 struct mlx4_ib_dev *dev = to_mdev(ibqp->device); 2999 struct mlx4_ib_qp *qp = to_mqp(ibqp); 3000 struct mlx4_qp_context context; 3001 int mlx4_state; 3002 int err = 0; 3003 3004 mutex_lock(&qp->mutex); 3005 3006 if (qp->state == IB_QPS_RESET) { 3007 qp_attr->qp_state = IB_QPS_RESET; 3008 goto done; 3009 } 3010 3011 err = mlx4_qp_query(dev->dev, &qp->mqp, &context); 3012 if (err) { 3013 err = -EINVAL; 3014 goto out; 3015 } 3016 3017 mlx4_state = be32_to_cpu(context.flags) >> 28; 3018 3019 qp->state = to_ib_qp_state(mlx4_state); 3020 qp_attr->qp_state = qp->state; 3021 qp_attr->path_mtu = context.mtu_msgmax >> 5; 3022 qp_attr->path_mig_state = 3023 to_ib_mig_state((be32_to_cpu(context.flags) >> 11) & 0x3); 3024 qp_attr->qkey = be32_to_cpu(context.qkey); 3025 qp_attr->rq_psn = be32_to_cpu(context.rnr_nextrecvpsn) & 0xffffff; 3026 qp_attr->sq_psn = be32_to_cpu(context.next_send_psn) & 0xffffff; 3027 qp_attr->dest_qp_num = be32_to_cpu(context.remote_qpn) & 0xffffff; 3028 qp_attr->qp_access_flags = 3029 to_ib_qp_access_flags(be32_to_cpu(context.params2)); 3030 3031 if (qp->ibqp.qp_type == IB_QPT_RC || qp->ibqp.qp_type == IB_QPT_UC) { 3032 to_ib_ah_attr(dev, &qp_attr->ah_attr, &context.pri_path); 3033 to_ib_ah_attr(dev, &qp_attr->alt_ah_attr, &context.alt_path); 3034 qp_attr->alt_pkey_index = context.alt_path.pkey_index & 0x7f; 3035 qp_attr->alt_port_num = qp_attr->alt_ah_attr.port_num; 3036 } 3037 3038 qp_attr->pkey_index = context.pri_path.pkey_index & 0x7f; 3039 if (qp_attr->qp_state == IB_QPS_INIT) 3040 qp_attr->port_num = qp->port; 3041 else 3042 qp_attr->port_num = context.pri_path.sched_queue & 0x40 ? 2 : 1; 3043 3044 /* qp_attr->en_sqd_async_notify is only applicable in modify qp */ 3045 qp_attr->sq_draining = mlx4_state == MLX4_QP_STATE_SQ_DRAINING; 3046 3047 qp_attr->max_rd_atomic = 1 << ((be32_to_cpu(context.params1) >> 21) & 0x7); 3048 3049 qp_attr->max_dest_rd_atomic = 3050 1 << ((be32_to_cpu(context.params2) >> 21) & 0x7); 3051 qp_attr->min_rnr_timer = 3052 (be32_to_cpu(context.rnr_nextrecvpsn) >> 24) & 0x1f; 3053 qp_attr->timeout = context.pri_path.ackto >> 3; 3054 qp_attr->retry_cnt = (be32_to_cpu(context.params1) >> 16) & 0x7; 3055 qp_attr->rnr_retry = (be32_to_cpu(context.params1) >> 13) & 0x7; 3056 qp_attr->alt_timeout = context.alt_path.ackto >> 3; 3057 3058 done: 3059 qp_attr->cur_qp_state = qp_attr->qp_state; 3060 qp_attr->cap.max_recv_wr = qp->rq.wqe_cnt; 3061 qp_attr->cap.max_recv_sge = qp->rq.max_gs; 3062 3063 if (!ibqp->uobject) { 3064 qp_attr->cap.max_send_wr = qp->sq.wqe_cnt; 3065 qp_attr->cap.max_send_sge = qp->sq.max_gs; 3066 } else { 3067 qp_attr->cap.max_send_wr = 0; 3068 qp_attr->cap.max_send_sge = 0; 3069 } 3070 3071 /* 3072 * We don't support inline sends for kernel QPs (yet), and we 3073 * don't know what userspace's value should be. 3074 */ 3075 qp_attr->cap.max_inline_data = 0; 3076 3077 qp_init_attr->cap = qp_attr->cap; 3078 3079 qp_init_attr->create_flags = 0; 3080 if (qp->flags & MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK) 3081 qp_init_attr->create_flags |= IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK; 3082 3083 if (qp->flags & MLX4_IB_QP_LSO) 3084 qp_init_attr->create_flags |= IB_QP_CREATE_IPOIB_UD_LSO; 3085 3086 if (qp->flags & MLX4_IB_QP_NETIF) 3087 qp_init_attr->create_flags |= IB_QP_CREATE_NETIF_QP; 3088 3089 qp_init_attr->sq_sig_type = 3090 qp->sq_signal_bits == cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE) ? 3091 IB_SIGNAL_ALL_WR : IB_SIGNAL_REQ_WR; 3092 3093 out: 3094 mutex_unlock(&qp->mutex); 3095 return err; 3096 } 3097 3098