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 42 #include <linux/mlx4/qp.h> 43 44 #include "mlx4_ib.h" 45 #include "user.h" 46 47 enum { 48 MLX4_IB_ACK_REQ_FREQ = 8, 49 }; 50 51 enum { 52 MLX4_IB_DEFAULT_SCHED_QUEUE = 0x83, 53 MLX4_IB_DEFAULT_QP0_SCHED_QUEUE = 0x3f, 54 MLX4_IB_LINK_TYPE_IB = 0, 55 MLX4_IB_LINK_TYPE_ETH = 1 56 }; 57 58 enum { 59 /* 60 * Largest possible UD header: send with GRH and immediate 61 * data plus 18 bytes for an Ethernet header with VLAN/802.1Q 62 * tag. (LRH would only use 8 bytes, so Ethernet is the 63 * biggest case) 64 */ 65 MLX4_IB_UD_HEADER_SIZE = 82, 66 MLX4_IB_LSO_HEADER_SPARE = 128, 67 }; 68 69 enum { 70 MLX4_IB_IBOE_ETHERTYPE = 0x8915 71 }; 72 73 struct mlx4_ib_sqp { 74 struct mlx4_ib_qp qp; 75 int pkey_index; 76 u32 qkey; 77 u32 send_psn; 78 struct ib_ud_header ud_header; 79 u8 header_buf[MLX4_IB_UD_HEADER_SIZE]; 80 }; 81 82 enum { 83 MLX4_IB_MIN_SQ_STRIDE = 6, 84 MLX4_IB_CACHE_LINE_SIZE = 64, 85 }; 86 87 static const __be32 mlx4_ib_opcode[] = { 88 [IB_WR_SEND] = cpu_to_be32(MLX4_OPCODE_SEND), 89 [IB_WR_LSO] = cpu_to_be32(MLX4_OPCODE_LSO), 90 [IB_WR_SEND_WITH_IMM] = cpu_to_be32(MLX4_OPCODE_SEND_IMM), 91 [IB_WR_RDMA_WRITE] = cpu_to_be32(MLX4_OPCODE_RDMA_WRITE), 92 [IB_WR_RDMA_WRITE_WITH_IMM] = cpu_to_be32(MLX4_OPCODE_RDMA_WRITE_IMM), 93 [IB_WR_RDMA_READ] = cpu_to_be32(MLX4_OPCODE_RDMA_READ), 94 [IB_WR_ATOMIC_CMP_AND_SWP] = cpu_to_be32(MLX4_OPCODE_ATOMIC_CS), 95 [IB_WR_ATOMIC_FETCH_AND_ADD] = cpu_to_be32(MLX4_OPCODE_ATOMIC_FA), 96 [IB_WR_SEND_WITH_INV] = cpu_to_be32(MLX4_OPCODE_SEND_INVAL), 97 [IB_WR_LOCAL_INV] = cpu_to_be32(MLX4_OPCODE_LOCAL_INVAL), 98 [IB_WR_FAST_REG_MR] = cpu_to_be32(MLX4_OPCODE_FMR), 99 [IB_WR_MASKED_ATOMIC_CMP_AND_SWP] = cpu_to_be32(MLX4_OPCODE_MASKED_ATOMIC_CS), 100 [IB_WR_MASKED_ATOMIC_FETCH_AND_ADD] = cpu_to_be32(MLX4_OPCODE_MASKED_ATOMIC_FA), 101 }; 102 103 static struct mlx4_ib_sqp *to_msqp(struct mlx4_ib_qp *mqp) 104 { 105 return container_of(mqp, struct mlx4_ib_sqp, qp); 106 } 107 108 static int is_sqp(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp) 109 { 110 return qp->mqp.qpn >= dev->dev->caps.sqp_start && 111 qp->mqp.qpn <= dev->dev->caps.sqp_start + 3; 112 } 113 114 static int is_qp0(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp) 115 { 116 return qp->mqp.qpn >= dev->dev->caps.sqp_start && 117 qp->mqp.qpn <= dev->dev->caps.sqp_start + 1; 118 } 119 120 static void *get_wqe(struct mlx4_ib_qp *qp, int offset) 121 { 122 return mlx4_buf_offset(&qp->buf, offset); 123 } 124 125 static void *get_recv_wqe(struct mlx4_ib_qp *qp, int n) 126 { 127 return get_wqe(qp, qp->rq.offset + (n << qp->rq.wqe_shift)); 128 } 129 130 static void *get_send_wqe(struct mlx4_ib_qp *qp, int n) 131 { 132 return get_wqe(qp, qp->sq.offset + (n << qp->sq.wqe_shift)); 133 } 134 135 /* 136 * Stamp a SQ WQE so that it is invalid if prefetched by marking the 137 * first four bytes of every 64 byte chunk with 138 * 0x7FFFFFF | (invalid_ownership_value << 31). 139 * 140 * When the max work request size is less than or equal to the WQE 141 * basic block size, as an optimization, we can stamp all WQEs with 142 * 0xffffffff, and skip the very first chunk of each WQE. 143 */ 144 static void stamp_send_wqe(struct mlx4_ib_qp *qp, int n, int size) 145 { 146 __be32 *wqe; 147 int i; 148 int s; 149 int ind; 150 void *buf; 151 __be32 stamp; 152 struct mlx4_wqe_ctrl_seg *ctrl; 153 154 if (qp->sq_max_wqes_per_wr > 1) { 155 s = roundup(size, 1U << qp->sq.wqe_shift); 156 for (i = 0; i < s; i += 64) { 157 ind = (i >> qp->sq.wqe_shift) + n; 158 stamp = ind & qp->sq.wqe_cnt ? cpu_to_be32(0x7fffffff) : 159 cpu_to_be32(0xffffffff); 160 buf = get_send_wqe(qp, ind & (qp->sq.wqe_cnt - 1)); 161 wqe = buf + (i & ((1 << qp->sq.wqe_shift) - 1)); 162 *wqe = stamp; 163 } 164 } else { 165 ctrl = buf = get_send_wqe(qp, n & (qp->sq.wqe_cnt - 1)); 166 s = (ctrl->fence_size & 0x3f) << 4; 167 for (i = 64; i < s; i += 64) { 168 wqe = buf + i; 169 *wqe = cpu_to_be32(0xffffffff); 170 } 171 } 172 } 173 174 static void post_nop_wqe(struct mlx4_ib_qp *qp, int n, int size) 175 { 176 struct mlx4_wqe_ctrl_seg *ctrl; 177 struct mlx4_wqe_inline_seg *inl; 178 void *wqe; 179 int s; 180 181 ctrl = wqe = get_send_wqe(qp, n & (qp->sq.wqe_cnt - 1)); 182 s = sizeof(struct mlx4_wqe_ctrl_seg); 183 184 if (qp->ibqp.qp_type == IB_QPT_UD) { 185 struct mlx4_wqe_datagram_seg *dgram = wqe + sizeof *ctrl; 186 struct mlx4_av *av = (struct mlx4_av *)dgram->av; 187 memset(dgram, 0, sizeof *dgram); 188 av->port_pd = cpu_to_be32((qp->port << 24) | to_mpd(qp->ibqp.pd)->pdn); 189 s += sizeof(struct mlx4_wqe_datagram_seg); 190 } 191 192 /* Pad the remainder of the WQE with an inline data segment. */ 193 if (size > s) { 194 inl = wqe + s; 195 inl->byte_count = cpu_to_be32(1 << 31 | (size - s - sizeof *inl)); 196 } 197 ctrl->srcrb_flags = 0; 198 ctrl->fence_size = size / 16; 199 /* 200 * Make sure descriptor is fully written before setting ownership bit 201 * (because HW can start executing as soon as we do). 202 */ 203 wmb(); 204 205 ctrl->owner_opcode = cpu_to_be32(MLX4_OPCODE_NOP | MLX4_WQE_CTRL_NEC) | 206 (n & qp->sq.wqe_cnt ? cpu_to_be32(1 << 31) : 0); 207 208 stamp_send_wqe(qp, n + qp->sq_spare_wqes, size); 209 } 210 211 /* Post NOP WQE to prevent wrap-around in the middle of WR */ 212 static inline unsigned pad_wraparound(struct mlx4_ib_qp *qp, int ind) 213 { 214 unsigned s = qp->sq.wqe_cnt - (ind & (qp->sq.wqe_cnt - 1)); 215 if (unlikely(s < qp->sq_max_wqes_per_wr)) { 216 post_nop_wqe(qp, ind, s << qp->sq.wqe_shift); 217 ind += s; 218 } 219 return ind; 220 } 221 222 static void mlx4_ib_qp_event(struct mlx4_qp *qp, enum mlx4_event type) 223 { 224 struct ib_event event; 225 struct ib_qp *ibqp = &to_mibqp(qp)->ibqp; 226 227 if (type == MLX4_EVENT_TYPE_PATH_MIG) 228 to_mibqp(qp)->port = to_mibqp(qp)->alt_port; 229 230 if (ibqp->event_handler) { 231 event.device = ibqp->device; 232 event.element.qp = ibqp; 233 switch (type) { 234 case MLX4_EVENT_TYPE_PATH_MIG: 235 event.event = IB_EVENT_PATH_MIG; 236 break; 237 case MLX4_EVENT_TYPE_COMM_EST: 238 event.event = IB_EVENT_COMM_EST; 239 break; 240 case MLX4_EVENT_TYPE_SQ_DRAINED: 241 event.event = IB_EVENT_SQ_DRAINED; 242 break; 243 case MLX4_EVENT_TYPE_SRQ_QP_LAST_WQE: 244 event.event = IB_EVENT_QP_LAST_WQE_REACHED; 245 break; 246 case MLX4_EVENT_TYPE_WQ_CATAS_ERROR: 247 event.event = IB_EVENT_QP_FATAL; 248 break; 249 case MLX4_EVENT_TYPE_PATH_MIG_FAILED: 250 event.event = IB_EVENT_PATH_MIG_ERR; 251 break; 252 case MLX4_EVENT_TYPE_WQ_INVAL_REQ_ERROR: 253 event.event = IB_EVENT_QP_REQ_ERR; 254 break; 255 case MLX4_EVENT_TYPE_WQ_ACCESS_ERROR: 256 event.event = IB_EVENT_QP_ACCESS_ERR; 257 break; 258 default: 259 printk(KERN_WARNING "mlx4_ib: Unexpected event type %d " 260 "on QP %06x\n", type, qp->qpn); 261 return; 262 } 263 264 ibqp->event_handler(&event, ibqp->qp_context); 265 } 266 } 267 268 static int send_wqe_overhead(enum ib_qp_type type, u32 flags) 269 { 270 /* 271 * UD WQEs must have a datagram segment. 272 * RC and UC WQEs might have a remote address segment. 273 * MLX WQEs need two extra inline data segments (for the UD 274 * header and space for the ICRC). 275 */ 276 switch (type) { 277 case IB_QPT_UD: 278 return sizeof (struct mlx4_wqe_ctrl_seg) + 279 sizeof (struct mlx4_wqe_datagram_seg) + 280 ((flags & MLX4_IB_QP_LSO) ? MLX4_IB_LSO_HEADER_SPARE : 0); 281 case IB_QPT_UC: 282 return sizeof (struct mlx4_wqe_ctrl_seg) + 283 sizeof (struct mlx4_wqe_raddr_seg); 284 case IB_QPT_RC: 285 return sizeof (struct mlx4_wqe_ctrl_seg) + 286 sizeof (struct mlx4_wqe_atomic_seg) + 287 sizeof (struct mlx4_wqe_raddr_seg); 288 case IB_QPT_SMI: 289 case IB_QPT_GSI: 290 return sizeof (struct mlx4_wqe_ctrl_seg) + 291 ALIGN(MLX4_IB_UD_HEADER_SIZE + 292 DIV_ROUND_UP(MLX4_IB_UD_HEADER_SIZE, 293 MLX4_INLINE_ALIGN) * 294 sizeof (struct mlx4_wqe_inline_seg), 295 sizeof (struct mlx4_wqe_data_seg)) + 296 ALIGN(4 + 297 sizeof (struct mlx4_wqe_inline_seg), 298 sizeof (struct mlx4_wqe_data_seg)); 299 default: 300 return sizeof (struct mlx4_wqe_ctrl_seg); 301 } 302 } 303 304 static int set_rq_size(struct mlx4_ib_dev *dev, struct ib_qp_cap *cap, 305 int is_user, int has_rq, struct mlx4_ib_qp *qp) 306 { 307 /* Sanity check RQ size before proceeding */ 308 if (cap->max_recv_wr > dev->dev->caps.max_wqes || 309 cap->max_recv_sge > dev->dev->caps.max_rq_sg) 310 return -EINVAL; 311 312 if (!has_rq) { 313 if (cap->max_recv_wr) 314 return -EINVAL; 315 316 qp->rq.wqe_cnt = qp->rq.max_gs = 0; 317 } else { 318 /* HW requires >= 1 RQ entry with >= 1 gather entry */ 319 if (is_user && (!cap->max_recv_wr || !cap->max_recv_sge)) 320 return -EINVAL; 321 322 qp->rq.wqe_cnt = roundup_pow_of_two(max(1U, cap->max_recv_wr)); 323 qp->rq.max_gs = roundup_pow_of_two(max(1U, cap->max_recv_sge)); 324 qp->rq.wqe_shift = ilog2(qp->rq.max_gs * sizeof (struct mlx4_wqe_data_seg)); 325 } 326 327 cap->max_recv_wr = qp->rq.max_post = qp->rq.wqe_cnt; 328 cap->max_recv_sge = qp->rq.max_gs; 329 330 return 0; 331 } 332 333 static int set_kernel_sq_size(struct mlx4_ib_dev *dev, struct ib_qp_cap *cap, 334 enum ib_qp_type type, struct mlx4_ib_qp *qp) 335 { 336 int s; 337 338 /* Sanity check SQ size before proceeding */ 339 if (cap->max_send_wr > dev->dev->caps.max_wqes || 340 cap->max_send_sge > dev->dev->caps.max_sq_sg || 341 cap->max_inline_data + send_wqe_overhead(type, qp->flags) + 342 sizeof (struct mlx4_wqe_inline_seg) > dev->dev->caps.max_sq_desc_sz) 343 return -EINVAL; 344 345 /* 346 * For MLX transport we need 2 extra S/G entries: 347 * one for the header and one for the checksum at the end 348 */ 349 if ((type == IB_QPT_SMI || type == IB_QPT_GSI) && 350 cap->max_send_sge + 2 > dev->dev->caps.max_sq_sg) 351 return -EINVAL; 352 353 s = max(cap->max_send_sge * sizeof (struct mlx4_wqe_data_seg), 354 cap->max_inline_data + sizeof (struct mlx4_wqe_inline_seg)) + 355 send_wqe_overhead(type, qp->flags); 356 357 if (s > dev->dev->caps.max_sq_desc_sz) 358 return -EINVAL; 359 360 /* 361 * Hermon supports shrinking WQEs, such that a single work 362 * request can include multiple units of 1 << wqe_shift. This 363 * way, work requests can differ in size, and do not have to 364 * be a power of 2 in size, saving memory and speeding up send 365 * WR posting. Unfortunately, if we do this then the 366 * wqe_index field in CQEs can't be used to look up the WR ID 367 * anymore, so we do this only if selective signaling is off. 368 * 369 * Further, on 32-bit platforms, we can't use vmap() to make 370 * the QP buffer virtually contiguous. Thus we have to use 371 * constant-sized WRs to make sure a WR is always fully within 372 * a single page-sized chunk. 373 * 374 * Finally, we use NOP work requests to pad the end of the 375 * work queue, to avoid wrap-around in the middle of WR. We 376 * set NEC bit to avoid getting completions with error for 377 * these NOP WRs, but since NEC is only supported starting 378 * with firmware 2.2.232, we use constant-sized WRs for older 379 * firmware. 380 * 381 * And, since MLX QPs only support SEND, we use constant-sized 382 * WRs in this case. 383 * 384 * We look for the smallest value of wqe_shift such that the 385 * resulting number of wqes does not exceed device 386 * capabilities. 387 * 388 * We set WQE size to at least 64 bytes, this way stamping 389 * invalidates each WQE. 390 */ 391 if (dev->dev->caps.fw_ver >= MLX4_FW_VER_WQE_CTRL_NEC && 392 qp->sq_signal_bits && BITS_PER_LONG == 64 && 393 type != IB_QPT_SMI && type != IB_QPT_GSI) 394 qp->sq.wqe_shift = ilog2(64); 395 else 396 qp->sq.wqe_shift = ilog2(roundup_pow_of_two(s)); 397 398 for (;;) { 399 qp->sq_max_wqes_per_wr = DIV_ROUND_UP(s, 1U << qp->sq.wqe_shift); 400 401 /* 402 * We need to leave 2 KB + 1 WR of headroom in the SQ to 403 * allow HW to prefetch. 404 */ 405 qp->sq_spare_wqes = (2048 >> qp->sq.wqe_shift) + qp->sq_max_wqes_per_wr; 406 qp->sq.wqe_cnt = roundup_pow_of_two(cap->max_send_wr * 407 qp->sq_max_wqes_per_wr + 408 qp->sq_spare_wqes); 409 410 if (qp->sq.wqe_cnt <= dev->dev->caps.max_wqes) 411 break; 412 413 if (qp->sq_max_wqes_per_wr <= 1) 414 return -EINVAL; 415 416 ++qp->sq.wqe_shift; 417 } 418 419 qp->sq.max_gs = (min(dev->dev->caps.max_sq_desc_sz, 420 (qp->sq_max_wqes_per_wr << qp->sq.wqe_shift)) - 421 send_wqe_overhead(type, qp->flags)) / 422 sizeof (struct mlx4_wqe_data_seg); 423 424 qp->buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) + 425 (qp->sq.wqe_cnt << qp->sq.wqe_shift); 426 if (qp->rq.wqe_shift > qp->sq.wqe_shift) { 427 qp->rq.offset = 0; 428 qp->sq.offset = qp->rq.wqe_cnt << qp->rq.wqe_shift; 429 } else { 430 qp->rq.offset = qp->sq.wqe_cnt << qp->sq.wqe_shift; 431 qp->sq.offset = 0; 432 } 433 434 cap->max_send_wr = qp->sq.max_post = 435 (qp->sq.wqe_cnt - qp->sq_spare_wqes) / qp->sq_max_wqes_per_wr; 436 cap->max_send_sge = min(qp->sq.max_gs, 437 min(dev->dev->caps.max_sq_sg, 438 dev->dev->caps.max_rq_sg)); 439 /* We don't support inline sends for kernel QPs (yet) */ 440 cap->max_inline_data = 0; 441 442 return 0; 443 } 444 445 static int set_user_sq_size(struct mlx4_ib_dev *dev, 446 struct mlx4_ib_qp *qp, 447 struct mlx4_ib_create_qp *ucmd) 448 { 449 /* Sanity check SQ size before proceeding */ 450 if ((1 << ucmd->log_sq_bb_count) > dev->dev->caps.max_wqes || 451 ucmd->log_sq_stride > 452 ilog2(roundup_pow_of_two(dev->dev->caps.max_sq_desc_sz)) || 453 ucmd->log_sq_stride < MLX4_IB_MIN_SQ_STRIDE) 454 return -EINVAL; 455 456 qp->sq.wqe_cnt = 1 << ucmd->log_sq_bb_count; 457 qp->sq.wqe_shift = ucmd->log_sq_stride; 458 459 qp->buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) + 460 (qp->sq.wqe_cnt << qp->sq.wqe_shift); 461 462 return 0; 463 } 464 465 static int qp_has_rq(struct ib_qp_init_attr *attr) 466 { 467 if (attr->qp_type == IB_QPT_XRC_INI || attr->qp_type == IB_QPT_XRC_TGT) 468 return 0; 469 470 return !attr->srq; 471 } 472 473 static int create_qp_common(struct mlx4_ib_dev *dev, struct ib_pd *pd, 474 struct ib_qp_init_attr *init_attr, 475 struct ib_udata *udata, int sqpn, struct mlx4_ib_qp *qp) 476 { 477 int qpn; 478 int err; 479 480 mutex_init(&qp->mutex); 481 spin_lock_init(&qp->sq.lock); 482 spin_lock_init(&qp->rq.lock); 483 INIT_LIST_HEAD(&qp->gid_list); 484 485 qp->state = IB_QPS_RESET; 486 if (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR) 487 qp->sq_signal_bits = cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE); 488 489 err = set_rq_size(dev, &init_attr->cap, !!pd->uobject, qp_has_rq(init_attr), qp); 490 if (err) 491 goto err; 492 493 if (pd->uobject) { 494 struct mlx4_ib_create_qp ucmd; 495 496 if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd)) { 497 err = -EFAULT; 498 goto err; 499 } 500 501 qp->sq_no_prefetch = ucmd.sq_no_prefetch; 502 503 err = set_user_sq_size(dev, qp, &ucmd); 504 if (err) 505 goto err; 506 507 qp->umem = ib_umem_get(pd->uobject->context, ucmd.buf_addr, 508 qp->buf_size, 0, 0); 509 if (IS_ERR(qp->umem)) { 510 err = PTR_ERR(qp->umem); 511 goto err; 512 } 513 514 err = mlx4_mtt_init(dev->dev, ib_umem_page_count(qp->umem), 515 ilog2(qp->umem->page_size), &qp->mtt); 516 if (err) 517 goto err_buf; 518 519 err = mlx4_ib_umem_write_mtt(dev, &qp->mtt, qp->umem); 520 if (err) 521 goto err_mtt; 522 523 if (qp_has_rq(init_attr)) { 524 err = mlx4_ib_db_map_user(to_mucontext(pd->uobject->context), 525 ucmd.db_addr, &qp->db); 526 if (err) 527 goto err_mtt; 528 } 529 } else { 530 qp->sq_no_prefetch = 0; 531 532 if (init_attr->create_flags & IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK) 533 qp->flags |= MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK; 534 535 if (init_attr->create_flags & IB_QP_CREATE_IPOIB_UD_LSO) 536 qp->flags |= MLX4_IB_QP_LSO; 537 538 err = set_kernel_sq_size(dev, &init_attr->cap, init_attr->qp_type, qp); 539 if (err) 540 goto err; 541 542 if (qp_has_rq(init_attr)) { 543 err = mlx4_db_alloc(dev->dev, &qp->db, 0); 544 if (err) 545 goto err; 546 547 *qp->db.db = 0; 548 } 549 550 if (mlx4_buf_alloc(dev->dev, qp->buf_size, PAGE_SIZE * 2, &qp->buf)) { 551 err = -ENOMEM; 552 goto err_db; 553 } 554 555 err = mlx4_mtt_init(dev->dev, qp->buf.npages, qp->buf.page_shift, 556 &qp->mtt); 557 if (err) 558 goto err_buf; 559 560 err = mlx4_buf_write_mtt(dev->dev, &qp->mtt, &qp->buf); 561 if (err) 562 goto err_mtt; 563 564 qp->sq.wrid = kmalloc(qp->sq.wqe_cnt * sizeof (u64), GFP_KERNEL); 565 qp->rq.wrid = kmalloc(qp->rq.wqe_cnt * sizeof (u64), GFP_KERNEL); 566 567 if (!qp->sq.wrid || !qp->rq.wrid) { 568 err = -ENOMEM; 569 goto err_wrid; 570 } 571 } 572 573 if (sqpn) { 574 qpn = sqpn; 575 } else { 576 err = mlx4_qp_reserve_range(dev->dev, 1, 1, &qpn); 577 if (err) 578 goto err_wrid; 579 } 580 581 err = mlx4_qp_alloc(dev->dev, qpn, &qp->mqp); 582 if (err) 583 goto err_qpn; 584 585 if (init_attr->qp_type == IB_QPT_XRC_TGT) 586 qp->mqp.qpn |= (1 << 23); 587 588 /* 589 * Hardware wants QPN written in big-endian order (after 590 * shifting) for send doorbell. Precompute this value to save 591 * a little bit when posting sends. 592 */ 593 qp->doorbell_qpn = swab32(qp->mqp.qpn << 8); 594 595 qp->mqp.event = mlx4_ib_qp_event; 596 597 return 0; 598 599 err_qpn: 600 if (!sqpn) 601 mlx4_qp_release_range(dev->dev, qpn, 1); 602 603 err_wrid: 604 if (pd->uobject) { 605 if (qp_has_rq(init_attr)) 606 mlx4_ib_db_unmap_user(to_mucontext(pd->uobject->context), &qp->db); 607 } else { 608 kfree(qp->sq.wrid); 609 kfree(qp->rq.wrid); 610 } 611 612 err_mtt: 613 mlx4_mtt_cleanup(dev->dev, &qp->mtt); 614 615 err_buf: 616 if (pd->uobject) 617 ib_umem_release(qp->umem); 618 else 619 mlx4_buf_free(dev->dev, qp->buf_size, &qp->buf); 620 621 err_db: 622 if (!pd->uobject && qp_has_rq(init_attr)) 623 mlx4_db_free(dev->dev, &qp->db); 624 625 err: 626 return err; 627 } 628 629 static enum mlx4_qp_state to_mlx4_state(enum ib_qp_state state) 630 { 631 switch (state) { 632 case IB_QPS_RESET: return MLX4_QP_STATE_RST; 633 case IB_QPS_INIT: return MLX4_QP_STATE_INIT; 634 case IB_QPS_RTR: return MLX4_QP_STATE_RTR; 635 case IB_QPS_RTS: return MLX4_QP_STATE_RTS; 636 case IB_QPS_SQD: return MLX4_QP_STATE_SQD; 637 case IB_QPS_SQE: return MLX4_QP_STATE_SQER; 638 case IB_QPS_ERR: return MLX4_QP_STATE_ERR; 639 default: return -1; 640 } 641 } 642 643 static void mlx4_ib_lock_cqs(struct mlx4_ib_cq *send_cq, struct mlx4_ib_cq *recv_cq) 644 __acquires(&send_cq->lock) __acquires(&recv_cq->lock) 645 { 646 if (send_cq == recv_cq) { 647 spin_lock_irq(&send_cq->lock); 648 __acquire(&recv_cq->lock); 649 } else if (send_cq->mcq.cqn < recv_cq->mcq.cqn) { 650 spin_lock_irq(&send_cq->lock); 651 spin_lock_nested(&recv_cq->lock, SINGLE_DEPTH_NESTING); 652 } else { 653 spin_lock_irq(&recv_cq->lock); 654 spin_lock_nested(&send_cq->lock, SINGLE_DEPTH_NESTING); 655 } 656 } 657 658 static void mlx4_ib_unlock_cqs(struct mlx4_ib_cq *send_cq, struct mlx4_ib_cq *recv_cq) 659 __releases(&send_cq->lock) __releases(&recv_cq->lock) 660 { 661 if (send_cq == recv_cq) { 662 __release(&recv_cq->lock); 663 spin_unlock_irq(&send_cq->lock); 664 } else if (send_cq->mcq.cqn < recv_cq->mcq.cqn) { 665 spin_unlock(&recv_cq->lock); 666 spin_unlock_irq(&send_cq->lock); 667 } else { 668 spin_unlock(&send_cq->lock); 669 spin_unlock_irq(&recv_cq->lock); 670 } 671 } 672 673 static void del_gid_entries(struct mlx4_ib_qp *qp) 674 { 675 struct mlx4_ib_gid_entry *ge, *tmp; 676 677 list_for_each_entry_safe(ge, tmp, &qp->gid_list, list) { 678 list_del(&ge->list); 679 kfree(ge); 680 } 681 } 682 683 static struct mlx4_ib_pd *get_pd(struct mlx4_ib_qp *qp) 684 { 685 if (qp->ibqp.qp_type == IB_QPT_XRC_TGT) 686 return to_mpd(to_mxrcd(qp->ibqp.xrcd)->pd); 687 else 688 return to_mpd(qp->ibqp.pd); 689 } 690 691 static void get_cqs(struct mlx4_ib_qp *qp, 692 struct mlx4_ib_cq **send_cq, struct mlx4_ib_cq **recv_cq) 693 { 694 switch (qp->ibqp.qp_type) { 695 case IB_QPT_XRC_TGT: 696 *send_cq = to_mcq(to_mxrcd(qp->ibqp.xrcd)->cq); 697 *recv_cq = *send_cq; 698 break; 699 case IB_QPT_XRC_INI: 700 *send_cq = to_mcq(qp->ibqp.send_cq); 701 *recv_cq = *send_cq; 702 break; 703 default: 704 *send_cq = to_mcq(qp->ibqp.send_cq); 705 *recv_cq = to_mcq(qp->ibqp.recv_cq); 706 break; 707 } 708 } 709 710 static void destroy_qp_common(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp, 711 int is_user) 712 { 713 struct mlx4_ib_cq *send_cq, *recv_cq; 714 715 if (qp->state != IB_QPS_RESET) 716 if (mlx4_qp_modify(dev->dev, NULL, to_mlx4_state(qp->state), 717 MLX4_QP_STATE_RST, NULL, 0, 0, &qp->mqp)) 718 printk(KERN_WARNING "mlx4_ib: modify QP %06x to RESET failed.\n", 719 qp->mqp.qpn); 720 721 get_cqs(qp, &send_cq, &recv_cq); 722 723 mlx4_ib_lock_cqs(send_cq, recv_cq); 724 725 if (!is_user) { 726 __mlx4_ib_cq_clean(recv_cq, qp->mqp.qpn, 727 qp->ibqp.srq ? to_msrq(qp->ibqp.srq): NULL); 728 if (send_cq != recv_cq) 729 __mlx4_ib_cq_clean(send_cq, qp->mqp.qpn, NULL); 730 } 731 732 mlx4_qp_remove(dev->dev, &qp->mqp); 733 734 mlx4_ib_unlock_cqs(send_cq, recv_cq); 735 736 mlx4_qp_free(dev->dev, &qp->mqp); 737 738 if (!is_sqp(dev, qp)) 739 mlx4_qp_release_range(dev->dev, qp->mqp.qpn, 1); 740 741 mlx4_mtt_cleanup(dev->dev, &qp->mtt); 742 743 if (is_user) { 744 if (qp->rq.wqe_cnt) 745 mlx4_ib_db_unmap_user(to_mucontext(qp->ibqp.uobject->context), 746 &qp->db); 747 ib_umem_release(qp->umem); 748 } else { 749 kfree(qp->sq.wrid); 750 kfree(qp->rq.wrid); 751 mlx4_buf_free(dev->dev, qp->buf_size, &qp->buf); 752 if (qp->rq.wqe_cnt) 753 mlx4_db_free(dev->dev, &qp->db); 754 } 755 756 del_gid_entries(qp); 757 } 758 759 struct ib_qp *mlx4_ib_create_qp(struct ib_pd *pd, 760 struct ib_qp_init_attr *init_attr, 761 struct ib_udata *udata) 762 { 763 struct mlx4_ib_sqp *sqp; 764 struct mlx4_ib_qp *qp; 765 int err; 766 u16 xrcdn = 0; 767 768 /* 769 * We only support LSO and multicast loopback blocking, and 770 * only for kernel UD QPs. 771 */ 772 if (init_attr->create_flags & ~(IB_QP_CREATE_IPOIB_UD_LSO | 773 IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK)) 774 return ERR_PTR(-EINVAL); 775 776 if (init_attr->create_flags && 777 (udata || init_attr->qp_type != IB_QPT_UD)) 778 return ERR_PTR(-EINVAL); 779 780 switch (init_attr->qp_type) { 781 case IB_QPT_XRC_TGT: 782 pd = to_mxrcd(init_attr->xrcd)->pd; 783 xrcdn = to_mxrcd(init_attr->xrcd)->xrcdn; 784 init_attr->send_cq = to_mxrcd(init_attr->xrcd)->cq; 785 /* fall through */ 786 case IB_QPT_XRC_INI: 787 if (!(to_mdev(pd->device)->dev->caps.flags & MLX4_DEV_CAP_FLAG_XRC)) 788 return ERR_PTR(-ENOSYS); 789 init_attr->recv_cq = init_attr->send_cq; 790 /* fall through */ 791 case IB_QPT_RC: 792 case IB_QPT_UC: 793 case IB_QPT_UD: 794 { 795 qp = kzalloc(sizeof *qp, GFP_KERNEL); 796 if (!qp) 797 return ERR_PTR(-ENOMEM); 798 799 err = create_qp_common(to_mdev(pd->device), pd, init_attr, udata, 0, qp); 800 if (err) { 801 kfree(qp); 802 return ERR_PTR(err); 803 } 804 805 qp->ibqp.qp_num = qp->mqp.qpn; 806 qp->xrcdn = xrcdn; 807 808 break; 809 } 810 case IB_QPT_SMI: 811 case IB_QPT_GSI: 812 { 813 /* Userspace is not allowed to create special QPs: */ 814 if (udata) 815 return ERR_PTR(-EINVAL); 816 817 sqp = kzalloc(sizeof *sqp, GFP_KERNEL); 818 if (!sqp) 819 return ERR_PTR(-ENOMEM); 820 821 qp = &sqp->qp; 822 823 err = create_qp_common(to_mdev(pd->device), pd, init_attr, udata, 824 to_mdev(pd->device)->dev->caps.sqp_start + 825 (init_attr->qp_type == IB_QPT_SMI ? 0 : 2) + 826 init_attr->port_num - 1, 827 qp); 828 if (err) { 829 kfree(sqp); 830 return ERR_PTR(err); 831 } 832 833 qp->port = init_attr->port_num; 834 qp->ibqp.qp_num = init_attr->qp_type == IB_QPT_SMI ? 0 : 1; 835 836 break; 837 } 838 default: 839 /* Don't support raw QPs */ 840 return ERR_PTR(-EINVAL); 841 } 842 843 return &qp->ibqp; 844 } 845 846 int mlx4_ib_destroy_qp(struct ib_qp *qp) 847 { 848 struct mlx4_ib_dev *dev = to_mdev(qp->device); 849 struct mlx4_ib_qp *mqp = to_mqp(qp); 850 struct mlx4_ib_pd *pd; 851 852 if (is_qp0(dev, mqp)) 853 mlx4_CLOSE_PORT(dev->dev, mqp->port); 854 855 pd = get_pd(mqp); 856 destroy_qp_common(dev, mqp, !!pd->ibpd.uobject); 857 858 if (is_sqp(dev, mqp)) 859 kfree(to_msqp(mqp)); 860 else 861 kfree(mqp); 862 863 return 0; 864 } 865 866 static int to_mlx4_st(enum ib_qp_type type) 867 { 868 switch (type) { 869 case IB_QPT_RC: return MLX4_QP_ST_RC; 870 case IB_QPT_UC: return MLX4_QP_ST_UC; 871 case IB_QPT_UD: return MLX4_QP_ST_UD; 872 case IB_QPT_XRC_INI: 873 case IB_QPT_XRC_TGT: return MLX4_QP_ST_XRC; 874 case IB_QPT_SMI: 875 case IB_QPT_GSI: return MLX4_QP_ST_MLX; 876 default: return -1; 877 } 878 } 879 880 static __be32 to_mlx4_access_flags(struct mlx4_ib_qp *qp, const struct ib_qp_attr *attr, 881 int attr_mask) 882 { 883 u8 dest_rd_atomic; 884 u32 access_flags; 885 u32 hw_access_flags = 0; 886 887 if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) 888 dest_rd_atomic = attr->max_dest_rd_atomic; 889 else 890 dest_rd_atomic = qp->resp_depth; 891 892 if (attr_mask & IB_QP_ACCESS_FLAGS) 893 access_flags = attr->qp_access_flags; 894 else 895 access_flags = qp->atomic_rd_en; 896 897 if (!dest_rd_atomic) 898 access_flags &= IB_ACCESS_REMOTE_WRITE; 899 900 if (access_flags & IB_ACCESS_REMOTE_READ) 901 hw_access_flags |= MLX4_QP_BIT_RRE; 902 if (access_flags & IB_ACCESS_REMOTE_ATOMIC) 903 hw_access_flags |= MLX4_QP_BIT_RAE; 904 if (access_flags & IB_ACCESS_REMOTE_WRITE) 905 hw_access_flags |= MLX4_QP_BIT_RWE; 906 907 return cpu_to_be32(hw_access_flags); 908 } 909 910 static void store_sqp_attrs(struct mlx4_ib_sqp *sqp, const struct ib_qp_attr *attr, 911 int attr_mask) 912 { 913 if (attr_mask & IB_QP_PKEY_INDEX) 914 sqp->pkey_index = attr->pkey_index; 915 if (attr_mask & IB_QP_QKEY) 916 sqp->qkey = attr->qkey; 917 if (attr_mask & IB_QP_SQ_PSN) 918 sqp->send_psn = attr->sq_psn; 919 } 920 921 static void mlx4_set_sched(struct mlx4_qp_path *path, u8 port) 922 { 923 path->sched_queue = (path->sched_queue & 0xbf) | ((port - 1) << 6); 924 } 925 926 static int mlx4_set_path(struct mlx4_ib_dev *dev, const struct ib_ah_attr *ah, 927 struct mlx4_qp_path *path, u8 port) 928 { 929 int err; 930 int is_eth = rdma_port_get_link_layer(&dev->ib_dev, port) == 931 IB_LINK_LAYER_ETHERNET; 932 u8 mac[6]; 933 int is_mcast; 934 u16 vlan_tag; 935 int vidx; 936 937 path->grh_mylmc = ah->src_path_bits & 0x7f; 938 path->rlid = cpu_to_be16(ah->dlid); 939 if (ah->static_rate) { 940 path->static_rate = ah->static_rate + MLX4_STAT_RATE_OFFSET; 941 while (path->static_rate > IB_RATE_2_5_GBPS + MLX4_STAT_RATE_OFFSET && 942 !(1 << path->static_rate & dev->dev->caps.stat_rate_support)) 943 --path->static_rate; 944 } else 945 path->static_rate = 0; 946 947 if (ah->ah_flags & IB_AH_GRH) { 948 if (ah->grh.sgid_index >= dev->dev->caps.gid_table_len[port]) { 949 printk(KERN_ERR "sgid_index (%u) too large. max is %d\n", 950 ah->grh.sgid_index, dev->dev->caps.gid_table_len[port] - 1); 951 return -1; 952 } 953 954 path->grh_mylmc |= 1 << 7; 955 path->mgid_index = ah->grh.sgid_index; 956 path->hop_limit = ah->grh.hop_limit; 957 path->tclass_flowlabel = 958 cpu_to_be32((ah->grh.traffic_class << 20) | 959 (ah->grh.flow_label)); 960 memcpy(path->rgid, ah->grh.dgid.raw, 16); 961 } 962 963 if (is_eth) { 964 path->sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE | 965 ((port - 1) << 6) | ((ah->sl & 7) << 3) | ((ah->sl & 8) >> 1); 966 967 if (!(ah->ah_flags & IB_AH_GRH)) 968 return -1; 969 970 err = mlx4_ib_resolve_grh(dev, ah, mac, &is_mcast, port); 971 if (err) 972 return err; 973 974 memcpy(path->dmac, mac, 6); 975 path->ackto = MLX4_IB_LINK_TYPE_ETH; 976 /* use index 0 into MAC table for IBoE */ 977 path->grh_mylmc &= 0x80; 978 979 vlan_tag = rdma_get_vlan_id(&dev->iboe.gid_table[port - 1][ah->grh.sgid_index]); 980 if (vlan_tag < 0x1000) { 981 if (mlx4_find_cached_vlan(dev->dev, port, vlan_tag, &vidx)) 982 return -ENOENT; 983 984 path->vlan_index = vidx; 985 path->fl = 1 << 6; 986 } 987 } else 988 path->sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE | 989 ((port - 1) << 6) | ((ah->sl & 0xf) << 2); 990 991 return 0; 992 } 993 994 static void update_mcg_macs(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp) 995 { 996 struct mlx4_ib_gid_entry *ge, *tmp; 997 998 list_for_each_entry_safe(ge, tmp, &qp->gid_list, list) { 999 if (!ge->added && mlx4_ib_add_mc(dev, qp, &ge->gid)) { 1000 ge->added = 1; 1001 ge->port = qp->port; 1002 } 1003 } 1004 } 1005 1006 static int __mlx4_ib_modify_qp(struct ib_qp *ibqp, 1007 const struct ib_qp_attr *attr, int attr_mask, 1008 enum ib_qp_state cur_state, enum ib_qp_state new_state) 1009 { 1010 struct mlx4_ib_dev *dev = to_mdev(ibqp->device); 1011 struct mlx4_ib_qp *qp = to_mqp(ibqp); 1012 struct mlx4_ib_pd *pd; 1013 struct mlx4_ib_cq *send_cq, *recv_cq; 1014 struct mlx4_qp_context *context; 1015 enum mlx4_qp_optpar optpar = 0; 1016 int sqd_event; 1017 int err = -EINVAL; 1018 1019 context = kzalloc(sizeof *context, GFP_KERNEL); 1020 if (!context) 1021 return -ENOMEM; 1022 1023 context->flags = cpu_to_be32((to_mlx4_state(new_state) << 28) | 1024 (to_mlx4_st(ibqp->qp_type) << 16)); 1025 1026 if (!(attr_mask & IB_QP_PATH_MIG_STATE)) 1027 context->flags |= cpu_to_be32(MLX4_QP_PM_MIGRATED << 11); 1028 else { 1029 optpar |= MLX4_QP_OPTPAR_PM_STATE; 1030 switch (attr->path_mig_state) { 1031 case IB_MIG_MIGRATED: 1032 context->flags |= cpu_to_be32(MLX4_QP_PM_MIGRATED << 11); 1033 break; 1034 case IB_MIG_REARM: 1035 context->flags |= cpu_to_be32(MLX4_QP_PM_REARM << 11); 1036 break; 1037 case IB_MIG_ARMED: 1038 context->flags |= cpu_to_be32(MLX4_QP_PM_ARMED << 11); 1039 break; 1040 } 1041 } 1042 1043 if (ibqp->qp_type == IB_QPT_GSI || ibqp->qp_type == IB_QPT_SMI) 1044 context->mtu_msgmax = (IB_MTU_4096 << 5) | 11; 1045 else if (ibqp->qp_type == IB_QPT_UD) { 1046 if (qp->flags & MLX4_IB_QP_LSO) 1047 context->mtu_msgmax = (IB_MTU_4096 << 5) | 1048 ilog2(dev->dev->caps.max_gso_sz); 1049 else 1050 context->mtu_msgmax = (IB_MTU_4096 << 5) | 12; 1051 } else if (attr_mask & IB_QP_PATH_MTU) { 1052 if (attr->path_mtu < IB_MTU_256 || attr->path_mtu > IB_MTU_4096) { 1053 printk(KERN_ERR "path MTU (%u) is invalid\n", 1054 attr->path_mtu); 1055 goto out; 1056 } 1057 context->mtu_msgmax = (attr->path_mtu << 5) | 1058 ilog2(dev->dev->caps.max_msg_sz); 1059 } 1060 1061 if (qp->rq.wqe_cnt) 1062 context->rq_size_stride = ilog2(qp->rq.wqe_cnt) << 3; 1063 context->rq_size_stride |= qp->rq.wqe_shift - 4; 1064 1065 if (qp->sq.wqe_cnt) 1066 context->sq_size_stride = ilog2(qp->sq.wqe_cnt) << 3; 1067 context->sq_size_stride |= qp->sq.wqe_shift - 4; 1068 1069 if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) { 1070 context->sq_size_stride |= !!qp->sq_no_prefetch << 7; 1071 context->xrcd = cpu_to_be32((u32) qp->xrcdn); 1072 } 1073 1074 if (qp->ibqp.uobject) 1075 context->usr_page = cpu_to_be32(to_mucontext(ibqp->uobject->context)->uar.index); 1076 else 1077 context->usr_page = cpu_to_be32(dev->priv_uar.index); 1078 1079 if (attr_mask & IB_QP_DEST_QPN) 1080 context->remote_qpn = cpu_to_be32(attr->dest_qp_num); 1081 1082 if (attr_mask & IB_QP_PORT) { 1083 if (cur_state == IB_QPS_SQD && new_state == IB_QPS_SQD && 1084 !(attr_mask & IB_QP_AV)) { 1085 mlx4_set_sched(&context->pri_path, attr->port_num); 1086 optpar |= MLX4_QP_OPTPAR_SCHED_QUEUE; 1087 } 1088 } 1089 1090 if (cur_state == IB_QPS_INIT && new_state == IB_QPS_RTR) { 1091 if (dev->counters[qp->port - 1] != -1) { 1092 context->pri_path.counter_index = 1093 dev->counters[qp->port - 1]; 1094 optpar |= MLX4_QP_OPTPAR_COUNTER_INDEX; 1095 } else 1096 context->pri_path.counter_index = 0xff; 1097 } 1098 1099 if (attr_mask & IB_QP_PKEY_INDEX) { 1100 context->pri_path.pkey_index = attr->pkey_index; 1101 optpar |= MLX4_QP_OPTPAR_PKEY_INDEX; 1102 } 1103 1104 if (attr_mask & IB_QP_AV) { 1105 if (mlx4_set_path(dev, &attr->ah_attr, &context->pri_path, 1106 attr_mask & IB_QP_PORT ? attr->port_num : qp->port)) 1107 goto out; 1108 1109 optpar |= (MLX4_QP_OPTPAR_PRIMARY_ADDR_PATH | 1110 MLX4_QP_OPTPAR_SCHED_QUEUE); 1111 } 1112 1113 if (attr_mask & IB_QP_TIMEOUT) { 1114 context->pri_path.ackto |= attr->timeout << 3; 1115 optpar |= MLX4_QP_OPTPAR_ACK_TIMEOUT; 1116 } 1117 1118 if (attr_mask & IB_QP_ALT_PATH) { 1119 if (attr->alt_port_num == 0 || 1120 attr->alt_port_num > dev->dev->caps.num_ports) 1121 goto out; 1122 1123 if (attr->alt_pkey_index >= 1124 dev->dev->caps.pkey_table_len[attr->alt_port_num]) 1125 goto out; 1126 1127 if (mlx4_set_path(dev, &attr->alt_ah_attr, &context->alt_path, 1128 attr->alt_port_num)) 1129 goto out; 1130 1131 context->alt_path.pkey_index = attr->alt_pkey_index; 1132 context->alt_path.ackto = attr->alt_timeout << 3; 1133 optpar |= MLX4_QP_OPTPAR_ALT_ADDR_PATH; 1134 } 1135 1136 pd = get_pd(qp); 1137 get_cqs(qp, &send_cq, &recv_cq); 1138 context->pd = cpu_to_be32(pd->pdn); 1139 context->cqn_send = cpu_to_be32(send_cq->mcq.cqn); 1140 context->cqn_recv = cpu_to_be32(recv_cq->mcq.cqn); 1141 context->params1 = cpu_to_be32(MLX4_IB_ACK_REQ_FREQ << 28); 1142 1143 /* Set "fast registration enabled" for all kernel QPs */ 1144 if (!qp->ibqp.uobject) 1145 context->params1 |= cpu_to_be32(1 << 11); 1146 1147 if (attr_mask & IB_QP_RNR_RETRY) { 1148 context->params1 |= cpu_to_be32(attr->rnr_retry << 13); 1149 optpar |= MLX4_QP_OPTPAR_RNR_RETRY; 1150 } 1151 1152 if (attr_mask & IB_QP_RETRY_CNT) { 1153 context->params1 |= cpu_to_be32(attr->retry_cnt << 16); 1154 optpar |= MLX4_QP_OPTPAR_RETRY_COUNT; 1155 } 1156 1157 if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) { 1158 if (attr->max_rd_atomic) 1159 context->params1 |= 1160 cpu_to_be32(fls(attr->max_rd_atomic - 1) << 21); 1161 optpar |= MLX4_QP_OPTPAR_SRA_MAX; 1162 } 1163 1164 if (attr_mask & IB_QP_SQ_PSN) 1165 context->next_send_psn = cpu_to_be32(attr->sq_psn); 1166 1167 if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) { 1168 if (attr->max_dest_rd_atomic) 1169 context->params2 |= 1170 cpu_to_be32(fls(attr->max_dest_rd_atomic - 1) << 21); 1171 optpar |= MLX4_QP_OPTPAR_RRA_MAX; 1172 } 1173 1174 if (attr_mask & (IB_QP_ACCESS_FLAGS | IB_QP_MAX_DEST_RD_ATOMIC)) { 1175 context->params2 |= to_mlx4_access_flags(qp, attr, attr_mask); 1176 optpar |= MLX4_QP_OPTPAR_RWE | MLX4_QP_OPTPAR_RRE | MLX4_QP_OPTPAR_RAE; 1177 } 1178 1179 if (ibqp->srq) 1180 context->params2 |= cpu_to_be32(MLX4_QP_BIT_RIC); 1181 1182 if (attr_mask & IB_QP_MIN_RNR_TIMER) { 1183 context->rnr_nextrecvpsn |= cpu_to_be32(attr->min_rnr_timer << 24); 1184 optpar |= MLX4_QP_OPTPAR_RNR_TIMEOUT; 1185 } 1186 if (attr_mask & IB_QP_RQ_PSN) 1187 context->rnr_nextrecvpsn |= cpu_to_be32(attr->rq_psn); 1188 1189 if (attr_mask & IB_QP_QKEY) { 1190 context->qkey = cpu_to_be32(attr->qkey); 1191 optpar |= MLX4_QP_OPTPAR_Q_KEY; 1192 } 1193 1194 if (ibqp->srq) 1195 context->srqn = cpu_to_be32(1 << 24 | to_msrq(ibqp->srq)->msrq.srqn); 1196 1197 if (qp->rq.wqe_cnt && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) 1198 context->db_rec_addr = cpu_to_be64(qp->db.dma); 1199 1200 if (cur_state == IB_QPS_INIT && 1201 new_state == IB_QPS_RTR && 1202 (ibqp->qp_type == IB_QPT_GSI || ibqp->qp_type == IB_QPT_SMI || 1203 ibqp->qp_type == IB_QPT_UD)) { 1204 context->pri_path.sched_queue = (qp->port - 1) << 6; 1205 if (is_qp0(dev, qp)) 1206 context->pri_path.sched_queue |= MLX4_IB_DEFAULT_QP0_SCHED_QUEUE; 1207 else 1208 context->pri_path.sched_queue |= MLX4_IB_DEFAULT_SCHED_QUEUE; 1209 } 1210 1211 if (cur_state == IB_QPS_RTS && new_state == IB_QPS_SQD && 1212 attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY && attr->en_sqd_async_notify) 1213 sqd_event = 1; 1214 else 1215 sqd_event = 0; 1216 1217 if (!ibqp->uobject && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) 1218 context->rlkey |= (1 << 4); 1219 1220 /* 1221 * Before passing a kernel QP to the HW, make sure that the 1222 * ownership bits of the send queue are set and the SQ 1223 * headroom is stamped so that the hardware doesn't start 1224 * processing stale work requests. 1225 */ 1226 if (!ibqp->uobject && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) { 1227 struct mlx4_wqe_ctrl_seg *ctrl; 1228 int i; 1229 1230 for (i = 0; i < qp->sq.wqe_cnt; ++i) { 1231 ctrl = get_send_wqe(qp, i); 1232 ctrl->owner_opcode = cpu_to_be32(1 << 31); 1233 if (qp->sq_max_wqes_per_wr == 1) 1234 ctrl->fence_size = 1 << (qp->sq.wqe_shift - 4); 1235 1236 stamp_send_wqe(qp, i, 1 << qp->sq.wqe_shift); 1237 } 1238 } 1239 1240 err = mlx4_qp_modify(dev->dev, &qp->mtt, to_mlx4_state(cur_state), 1241 to_mlx4_state(new_state), context, optpar, 1242 sqd_event, &qp->mqp); 1243 if (err) 1244 goto out; 1245 1246 qp->state = new_state; 1247 1248 if (attr_mask & IB_QP_ACCESS_FLAGS) 1249 qp->atomic_rd_en = attr->qp_access_flags; 1250 if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) 1251 qp->resp_depth = attr->max_dest_rd_atomic; 1252 if (attr_mask & IB_QP_PORT) { 1253 qp->port = attr->port_num; 1254 update_mcg_macs(dev, qp); 1255 } 1256 if (attr_mask & IB_QP_ALT_PATH) 1257 qp->alt_port = attr->alt_port_num; 1258 1259 if (is_sqp(dev, qp)) 1260 store_sqp_attrs(to_msqp(qp), attr, attr_mask); 1261 1262 /* 1263 * If we moved QP0 to RTR, bring the IB link up; if we moved 1264 * QP0 to RESET or ERROR, bring the link back down. 1265 */ 1266 if (is_qp0(dev, qp)) { 1267 if (cur_state != IB_QPS_RTR && new_state == IB_QPS_RTR) 1268 if (mlx4_INIT_PORT(dev->dev, qp->port)) 1269 printk(KERN_WARNING "INIT_PORT failed for port %d\n", 1270 qp->port); 1271 1272 if (cur_state != IB_QPS_RESET && cur_state != IB_QPS_ERR && 1273 (new_state == IB_QPS_RESET || new_state == IB_QPS_ERR)) 1274 mlx4_CLOSE_PORT(dev->dev, qp->port); 1275 } 1276 1277 /* 1278 * If we moved a kernel QP to RESET, clean up all old CQ 1279 * entries and reinitialize the QP. 1280 */ 1281 if (new_state == IB_QPS_RESET && !ibqp->uobject) { 1282 mlx4_ib_cq_clean(recv_cq, qp->mqp.qpn, 1283 ibqp->srq ? to_msrq(ibqp->srq): NULL); 1284 if (send_cq != recv_cq) 1285 mlx4_ib_cq_clean(send_cq, qp->mqp.qpn, NULL); 1286 1287 qp->rq.head = 0; 1288 qp->rq.tail = 0; 1289 qp->sq.head = 0; 1290 qp->sq.tail = 0; 1291 qp->sq_next_wqe = 0; 1292 if (qp->rq.wqe_cnt) 1293 *qp->db.db = 0; 1294 } 1295 1296 out: 1297 kfree(context); 1298 return err; 1299 } 1300 1301 int mlx4_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, 1302 int attr_mask, struct ib_udata *udata) 1303 { 1304 struct mlx4_ib_dev *dev = to_mdev(ibqp->device); 1305 struct mlx4_ib_qp *qp = to_mqp(ibqp); 1306 enum ib_qp_state cur_state, new_state; 1307 int err = -EINVAL; 1308 1309 mutex_lock(&qp->mutex); 1310 1311 cur_state = attr_mask & IB_QP_CUR_STATE ? attr->cur_qp_state : qp->state; 1312 new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state; 1313 1314 if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type, attr_mask)) 1315 goto out; 1316 1317 if ((attr_mask & IB_QP_PORT) && 1318 (attr->port_num == 0 || attr->port_num > dev->dev->caps.num_ports)) { 1319 goto out; 1320 } 1321 1322 if (attr_mask & IB_QP_PKEY_INDEX) { 1323 int p = attr_mask & IB_QP_PORT ? attr->port_num : qp->port; 1324 if (attr->pkey_index >= dev->dev->caps.pkey_table_len[p]) 1325 goto out; 1326 } 1327 1328 if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC && 1329 attr->max_rd_atomic > dev->dev->caps.max_qp_init_rdma) { 1330 goto out; 1331 } 1332 1333 if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC && 1334 attr->max_dest_rd_atomic > dev->dev->caps.max_qp_dest_rdma) { 1335 goto out; 1336 } 1337 1338 if (cur_state == new_state && cur_state == IB_QPS_RESET) { 1339 err = 0; 1340 goto out; 1341 } 1342 1343 err = __mlx4_ib_modify_qp(ibqp, attr, attr_mask, cur_state, new_state); 1344 1345 out: 1346 mutex_unlock(&qp->mutex); 1347 return err; 1348 } 1349 1350 static int build_mlx_header(struct mlx4_ib_sqp *sqp, struct ib_send_wr *wr, 1351 void *wqe, unsigned *mlx_seg_len) 1352 { 1353 struct ib_device *ib_dev = sqp->qp.ibqp.device; 1354 struct mlx4_wqe_mlx_seg *mlx = wqe; 1355 struct mlx4_wqe_inline_seg *inl = wqe + sizeof *mlx; 1356 struct mlx4_ib_ah *ah = to_mah(wr->wr.ud.ah); 1357 union ib_gid sgid; 1358 u16 pkey; 1359 int send_size; 1360 int header_size; 1361 int spc; 1362 int i; 1363 int is_eth; 1364 int is_vlan = 0; 1365 int is_grh; 1366 u16 vlan; 1367 1368 send_size = 0; 1369 for (i = 0; i < wr->num_sge; ++i) 1370 send_size += wr->sg_list[i].length; 1371 1372 is_eth = rdma_port_get_link_layer(sqp->qp.ibqp.device, sqp->qp.port) == IB_LINK_LAYER_ETHERNET; 1373 is_grh = mlx4_ib_ah_grh_present(ah); 1374 if (is_eth) { 1375 ib_get_cached_gid(ib_dev, be32_to_cpu(ah->av.ib.port_pd) >> 24, 1376 ah->av.ib.gid_index, &sgid); 1377 vlan = rdma_get_vlan_id(&sgid); 1378 is_vlan = vlan < 0x1000; 1379 } 1380 ib_ud_header_init(send_size, !is_eth, is_eth, is_vlan, is_grh, 0, &sqp->ud_header); 1381 1382 if (!is_eth) { 1383 sqp->ud_header.lrh.service_level = 1384 be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 28; 1385 sqp->ud_header.lrh.destination_lid = ah->av.ib.dlid; 1386 sqp->ud_header.lrh.source_lid = cpu_to_be16(ah->av.ib.g_slid & 0x7f); 1387 } 1388 1389 if (is_grh) { 1390 sqp->ud_header.grh.traffic_class = 1391 (be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 20) & 0xff; 1392 sqp->ud_header.grh.flow_label = 1393 ah->av.ib.sl_tclass_flowlabel & cpu_to_be32(0xfffff); 1394 sqp->ud_header.grh.hop_limit = ah->av.ib.hop_limit; 1395 ib_get_cached_gid(ib_dev, be32_to_cpu(ah->av.ib.port_pd) >> 24, 1396 ah->av.ib.gid_index, &sqp->ud_header.grh.source_gid); 1397 memcpy(sqp->ud_header.grh.destination_gid.raw, 1398 ah->av.ib.dgid, 16); 1399 } 1400 1401 mlx->flags &= cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE); 1402 1403 if (!is_eth) { 1404 mlx->flags |= cpu_to_be32((!sqp->qp.ibqp.qp_num ? MLX4_WQE_MLX_VL15 : 0) | 1405 (sqp->ud_header.lrh.destination_lid == 1406 IB_LID_PERMISSIVE ? MLX4_WQE_MLX_SLR : 0) | 1407 (sqp->ud_header.lrh.service_level << 8)); 1408 mlx->rlid = sqp->ud_header.lrh.destination_lid; 1409 } 1410 1411 switch (wr->opcode) { 1412 case IB_WR_SEND: 1413 sqp->ud_header.bth.opcode = IB_OPCODE_UD_SEND_ONLY; 1414 sqp->ud_header.immediate_present = 0; 1415 break; 1416 case IB_WR_SEND_WITH_IMM: 1417 sqp->ud_header.bth.opcode = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE; 1418 sqp->ud_header.immediate_present = 1; 1419 sqp->ud_header.immediate_data = wr->ex.imm_data; 1420 break; 1421 default: 1422 return -EINVAL; 1423 } 1424 1425 if (is_eth) { 1426 u8 *smac; 1427 1428 memcpy(sqp->ud_header.eth.dmac_h, ah->av.eth.mac, 6); 1429 /* FIXME: cache smac value? */ 1430 smac = to_mdev(sqp->qp.ibqp.device)->iboe.netdevs[sqp->qp.port - 1]->dev_addr; 1431 memcpy(sqp->ud_header.eth.smac_h, smac, 6); 1432 if (!memcmp(sqp->ud_header.eth.smac_h, sqp->ud_header.eth.dmac_h, 6)) 1433 mlx->flags |= cpu_to_be32(MLX4_WQE_CTRL_FORCE_LOOPBACK); 1434 if (!is_vlan) { 1435 sqp->ud_header.eth.type = cpu_to_be16(MLX4_IB_IBOE_ETHERTYPE); 1436 } else { 1437 u16 pcp; 1438 1439 sqp->ud_header.vlan.type = cpu_to_be16(MLX4_IB_IBOE_ETHERTYPE); 1440 pcp = (be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 27 & 3) << 13; 1441 sqp->ud_header.vlan.tag = cpu_to_be16(vlan | pcp); 1442 } 1443 } else { 1444 sqp->ud_header.lrh.virtual_lane = !sqp->qp.ibqp.qp_num ? 15 : 0; 1445 if (sqp->ud_header.lrh.destination_lid == IB_LID_PERMISSIVE) 1446 sqp->ud_header.lrh.source_lid = IB_LID_PERMISSIVE; 1447 } 1448 sqp->ud_header.bth.solicited_event = !!(wr->send_flags & IB_SEND_SOLICITED); 1449 if (!sqp->qp.ibqp.qp_num) 1450 ib_get_cached_pkey(ib_dev, sqp->qp.port, sqp->pkey_index, &pkey); 1451 else 1452 ib_get_cached_pkey(ib_dev, sqp->qp.port, wr->wr.ud.pkey_index, &pkey); 1453 sqp->ud_header.bth.pkey = cpu_to_be16(pkey); 1454 sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->wr.ud.remote_qpn); 1455 sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1)); 1456 sqp->ud_header.deth.qkey = cpu_to_be32(wr->wr.ud.remote_qkey & 0x80000000 ? 1457 sqp->qkey : wr->wr.ud.remote_qkey); 1458 sqp->ud_header.deth.source_qpn = cpu_to_be32(sqp->qp.ibqp.qp_num); 1459 1460 header_size = ib_ud_header_pack(&sqp->ud_header, sqp->header_buf); 1461 1462 if (0) { 1463 printk(KERN_ERR "built UD header of size %d:\n", header_size); 1464 for (i = 0; i < header_size / 4; ++i) { 1465 if (i % 8 == 0) 1466 printk(" [%02x] ", i * 4); 1467 printk(" %08x", 1468 be32_to_cpu(((__be32 *) sqp->header_buf)[i])); 1469 if ((i + 1) % 8 == 0) 1470 printk("\n"); 1471 } 1472 printk("\n"); 1473 } 1474 1475 /* 1476 * Inline data segments may not cross a 64 byte boundary. If 1477 * our UD header is bigger than the space available up to the 1478 * next 64 byte boundary in the WQE, use two inline data 1479 * segments to hold the UD header. 1480 */ 1481 spc = MLX4_INLINE_ALIGN - 1482 ((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1)); 1483 if (header_size <= spc) { 1484 inl->byte_count = cpu_to_be32(1 << 31 | header_size); 1485 memcpy(inl + 1, sqp->header_buf, header_size); 1486 i = 1; 1487 } else { 1488 inl->byte_count = cpu_to_be32(1 << 31 | spc); 1489 memcpy(inl + 1, sqp->header_buf, spc); 1490 1491 inl = (void *) (inl + 1) + spc; 1492 memcpy(inl + 1, sqp->header_buf + spc, header_size - spc); 1493 /* 1494 * Need a barrier here to make sure all the data is 1495 * visible before the byte_count field is set. 1496 * Otherwise the HCA prefetcher could grab the 64-byte 1497 * chunk with this inline segment and get a valid (!= 1498 * 0xffffffff) byte count but stale data, and end up 1499 * generating a packet with bad headers. 1500 * 1501 * The first inline segment's byte_count field doesn't 1502 * need a barrier, because it comes after a 1503 * control/MLX segment and therefore is at an offset 1504 * of 16 mod 64. 1505 */ 1506 wmb(); 1507 inl->byte_count = cpu_to_be32(1 << 31 | (header_size - spc)); 1508 i = 2; 1509 } 1510 1511 *mlx_seg_len = 1512 ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + header_size, 16); 1513 return 0; 1514 } 1515 1516 static int mlx4_wq_overflow(struct mlx4_ib_wq *wq, int nreq, struct ib_cq *ib_cq) 1517 { 1518 unsigned cur; 1519 struct mlx4_ib_cq *cq; 1520 1521 cur = wq->head - wq->tail; 1522 if (likely(cur + nreq < wq->max_post)) 1523 return 0; 1524 1525 cq = to_mcq(ib_cq); 1526 spin_lock(&cq->lock); 1527 cur = wq->head - wq->tail; 1528 spin_unlock(&cq->lock); 1529 1530 return cur + nreq >= wq->max_post; 1531 } 1532 1533 static __be32 convert_access(int acc) 1534 { 1535 return (acc & IB_ACCESS_REMOTE_ATOMIC ? cpu_to_be32(MLX4_WQE_FMR_PERM_ATOMIC) : 0) | 1536 (acc & IB_ACCESS_REMOTE_WRITE ? cpu_to_be32(MLX4_WQE_FMR_PERM_REMOTE_WRITE) : 0) | 1537 (acc & IB_ACCESS_REMOTE_READ ? cpu_to_be32(MLX4_WQE_FMR_PERM_REMOTE_READ) : 0) | 1538 (acc & IB_ACCESS_LOCAL_WRITE ? cpu_to_be32(MLX4_WQE_FMR_PERM_LOCAL_WRITE) : 0) | 1539 cpu_to_be32(MLX4_WQE_FMR_PERM_LOCAL_READ); 1540 } 1541 1542 static void set_fmr_seg(struct mlx4_wqe_fmr_seg *fseg, struct ib_send_wr *wr) 1543 { 1544 struct mlx4_ib_fast_reg_page_list *mfrpl = to_mfrpl(wr->wr.fast_reg.page_list); 1545 int i; 1546 1547 for (i = 0; i < wr->wr.fast_reg.page_list_len; ++i) 1548 mfrpl->mapped_page_list[i] = 1549 cpu_to_be64(wr->wr.fast_reg.page_list->page_list[i] | 1550 MLX4_MTT_FLAG_PRESENT); 1551 1552 fseg->flags = convert_access(wr->wr.fast_reg.access_flags); 1553 fseg->mem_key = cpu_to_be32(wr->wr.fast_reg.rkey); 1554 fseg->buf_list = cpu_to_be64(mfrpl->map); 1555 fseg->start_addr = cpu_to_be64(wr->wr.fast_reg.iova_start); 1556 fseg->reg_len = cpu_to_be64(wr->wr.fast_reg.length); 1557 fseg->offset = 0; /* XXX -- is this just for ZBVA? */ 1558 fseg->page_size = cpu_to_be32(wr->wr.fast_reg.page_shift); 1559 fseg->reserved[0] = 0; 1560 fseg->reserved[1] = 0; 1561 } 1562 1563 static void set_local_inv_seg(struct mlx4_wqe_local_inval_seg *iseg, u32 rkey) 1564 { 1565 iseg->flags = 0; 1566 iseg->mem_key = cpu_to_be32(rkey); 1567 iseg->guest_id = 0; 1568 iseg->pa = 0; 1569 } 1570 1571 static __always_inline void set_raddr_seg(struct mlx4_wqe_raddr_seg *rseg, 1572 u64 remote_addr, u32 rkey) 1573 { 1574 rseg->raddr = cpu_to_be64(remote_addr); 1575 rseg->rkey = cpu_to_be32(rkey); 1576 rseg->reserved = 0; 1577 } 1578 1579 static void set_atomic_seg(struct mlx4_wqe_atomic_seg *aseg, struct ib_send_wr *wr) 1580 { 1581 if (wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP) { 1582 aseg->swap_add = cpu_to_be64(wr->wr.atomic.swap); 1583 aseg->compare = cpu_to_be64(wr->wr.atomic.compare_add); 1584 } else if (wr->opcode == IB_WR_MASKED_ATOMIC_FETCH_AND_ADD) { 1585 aseg->swap_add = cpu_to_be64(wr->wr.atomic.compare_add); 1586 aseg->compare = cpu_to_be64(wr->wr.atomic.compare_add_mask); 1587 } else { 1588 aseg->swap_add = cpu_to_be64(wr->wr.atomic.compare_add); 1589 aseg->compare = 0; 1590 } 1591 1592 } 1593 1594 static void set_masked_atomic_seg(struct mlx4_wqe_masked_atomic_seg *aseg, 1595 struct ib_send_wr *wr) 1596 { 1597 aseg->swap_add = cpu_to_be64(wr->wr.atomic.swap); 1598 aseg->swap_add_mask = cpu_to_be64(wr->wr.atomic.swap_mask); 1599 aseg->compare = cpu_to_be64(wr->wr.atomic.compare_add); 1600 aseg->compare_mask = cpu_to_be64(wr->wr.atomic.compare_add_mask); 1601 } 1602 1603 static void set_datagram_seg(struct mlx4_wqe_datagram_seg *dseg, 1604 struct ib_send_wr *wr) 1605 { 1606 memcpy(dseg->av, &to_mah(wr->wr.ud.ah)->av, sizeof (struct mlx4_av)); 1607 dseg->dqpn = cpu_to_be32(wr->wr.ud.remote_qpn); 1608 dseg->qkey = cpu_to_be32(wr->wr.ud.remote_qkey); 1609 dseg->vlan = to_mah(wr->wr.ud.ah)->av.eth.vlan; 1610 memcpy(dseg->mac, to_mah(wr->wr.ud.ah)->av.eth.mac, 6); 1611 } 1612 1613 static void set_mlx_icrc_seg(void *dseg) 1614 { 1615 u32 *t = dseg; 1616 struct mlx4_wqe_inline_seg *iseg = dseg; 1617 1618 t[1] = 0; 1619 1620 /* 1621 * Need a barrier here before writing the byte_count field to 1622 * make sure that all the data is visible before the 1623 * byte_count field is set. Otherwise, if the segment begins 1624 * a new cacheline, the HCA prefetcher could grab the 64-byte 1625 * chunk and get a valid (!= * 0xffffffff) byte count but 1626 * stale data, and end up sending the wrong data. 1627 */ 1628 wmb(); 1629 1630 iseg->byte_count = cpu_to_be32((1 << 31) | 4); 1631 } 1632 1633 static void set_data_seg(struct mlx4_wqe_data_seg *dseg, struct ib_sge *sg) 1634 { 1635 dseg->lkey = cpu_to_be32(sg->lkey); 1636 dseg->addr = cpu_to_be64(sg->addr); 1637 1638 /* 1639 * Need a barrier here before writing the byte_count field to 1640 * make sure that all the data is visible before the 1641 * byte_count field is set. Otherwise, if the segment begins 1642 * a new cacheline, the HCA prefetcher could grab the 64-byte 1643 * chunk and get a valid (!= * 0xffffffff) byte count but 1644 * stale data, and end up sending the wrong data. 1645 */ 1646 wmb(); 1647 1648 dseg->byte_count = cpu_to_be32(sg->length); 1649 } 1650 1651 static void __set_data_seg(struct mlx4_wqe_data_seg *dseg, struct ib_sge *sg) 1652 { 1653 dseg->byte_count = cpu_to_be32(sg->length); 1654 dseg->lkey = cpu_to_be32(sg->lkey); 1655 dseg->addr = cpu_to_be64(sg->addr); 1656 } 1657 1658 static int build_lso_seg(struct mlx4_wqe_lso_seg *wqe, struct ib_send_wr *wr, 1659 struct mlx4_ib_qp *qp, unsigned *lso_seg_len, 1660 __be32 *lso_hdr_sz, __be32 *blh) 1661 { 1662 unsigned halign = ALIGN(sizeof *wqe + wr->wr.ud.hlen, 16); 1663 1664 if (unlikely(halign > MLX4_IB_CACHE_LINE_SIZE)) 1665 *blh = cpu_to_be32(1 << 6); 1666 1667 if (unlikely(!(qp->flags & MLX4_IB_QP_LSO) && 1668 wr->num_sge > qp->sq.max_gs - (halign >> 4))) 1669 return -EINVAL; 1670 1671 memcpy(wqe->header, wr->wr.ud.header, wr->wr.ud.hlen); 1672 1673 *lso_hdr_sz = cpu_to_be32((wr->wr.ud.mss - wr->wr.ud.hlen) << 16 | 1674 wr->wr.ud.hlen); 1675 *lso_seg_len = halign; 1676 return 0; 1677 } 1678 1679 static __be32 send_ieth(struct ib_send_wr *wr) 1680 { 1681 switch (wr->opcode) { 1682 case IB_WR_SEND_WITH_IMM: 1683 case IB_WR_RDMA_WRITE_WITH_IMM: 1684 return wr->ex.imm_data; 1685 1686 case IB_WR_SEND_WITH_INV: 1687 return cpu_to_be32(wr->ex.invalidate_rkey); 1688 1689 default: 1690 return 0; 1691 } 1692 } 1693 1694 int mlx4_ib_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr, 1695 struct ib_send_wr **bad_wr) 1696 { 1697 struct mlx4_ib_qp *qp = to_mqp(ibqp); 1698 void *wqe; 1699 struct mlx4_wqe_ctrl_seg *ctrl; 1700 struct mlx4_wqe_data_seg *dseg; 1701 unsigned long flags; 1702 int nreq; 1703 int err = 0; 1704 unsigned ind; 1705 int uninitialized_var(stamp); 1706 int uninitialized_var(size); 1707 unsigned uninitialized_var(seglen); 1708 __be32 dummy; 1709 __be32 *lso_wqe; 1710 __be32 uninitialized_var(lso_hdr_sz); 1711 __be32 blh; 1712 int i; 1713 1714 spin_lock_irqsave(&qp->sq.lock, flags); 1715 1716 ind = qp->sq_next_wqe; 1717 1718 for (nreq = 0; wr; ++nreq, wr = wr->next) { 1719 lso_wqe = &dummy; 1720 blh = 0; 1721 1722 if (mlx4_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) { 1723 err = -ENOMEM; 1724 *bad_wr = wr; 1725 goto out; 1726 } 1727 1728 if (unlikely(wr->num_sge > qp->sq.max_gs)) { 1729 err = -EINVAL; 1730 *bad_wr = wr; 1731 goto out; 1732 } 1733 1734 ctrl = wqe = get_send_wqe(qp, ind & (qp->sq.wqe_cnt - 1)); 1735 qp->sq.wrid[(qp->sq.head + nreq) & (qp->sq.wqe_cnt - 1)] = wr->wr_id; 1736 1737 ctrl->srcrb_flags = 1738 (wr->send_flags & IB_SEND_SIGNALED ? 1739 cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE) : 0) | 1740 (wr->send_flags & IB_SEND_SOLICITED ? 1741 cpu_to_be32(MLX4_WQE_CTRL_SOLICITED) : 0) | 1742 ((wr->send_flags & IB_SEND_IP_CSUM) ? 1743 cpu_to_be32(MLX4_WQE_CTRL_IP_CSUM | 1744 MLX4_WQE_CTRL_TCP_UDP_CSUM) : 0) | 1745 qp->sq_signal_bits; 1746 1747 ctrl->imm = send_ieth(wr); 1748 1749 wqe += sizeof *ctrl; 1750 size = sizeof *ctrl / 16; 1751 1752 switch (ibqp->qp_type) { 1753 case IB_QPT_RC: 1754 case IB_QPT_UC: 1755 switch (wr->opcode) { 1756 case IB_WR_ATOMIC_CMP_AND_SWP: 1757 case IB_WR_ATOMIC_FETCH_AND_ADD: 1758 case IB_WR_MASKED_ATOMIC_FETCH_AND_ADD: 1759 set_raddr_seg(wqe, wr->wr.atomic.remote_addr, 1760 wr->wr.atomic.rkey); 1761 wqe += sizeof (struct mlx4_wqe_raddr_seg); 1762 1763 set_atomic_seg(wqe, wr); 1764 wqe += sizeof (struct mlx4_wqe_atomic_seg); 1765 1766 size += (sizeof (struct mlx4_wqe_raddr_seg) + 1767 sizeof (struct mlx4_wqe_atomic_seg)) / 16; 1768 1769 break; 1770 1771 case IB_WR_MASKED_ATOMIC_CMP_AND_SWP: 1772 set_raddr_seg(wqe, wr->wr.atomic.remote_addr, 1773 wr->wr.atomic.rkey); 1774 wqe += sizeof (struct mlx4_wqe_raddr_seg); 1775 1776 set_masked_atomic_seg(wqe, wr); 1777 wqe += sizeof (struct mlx4_wqe_masked_atomic_seg); 1778 1779 size += (sizeof (struct mlx4_wqe_raddr_seg) + 1780 sizeof (struct mlx4_wqe_masked_atomic_seg)) / 16; 1781 1782 break; 1783 1784 case IB_WR_RDMA_READ: 1785 case IB_WR_RDMA_WRITE: 1786 case IB_WR_RDMA_WRITE_WITH_IMM: 1787 set_raddr_seg(wqe, wr->wr.rdma.remote_addr, 1788 wr->wr.rdma.rkey); 1789 wqe += sizeof (struct mlx4_wqe_raddr_seg); 1790 size += sizeof (struct mlx4_wqe_raddr_seg) / 16; 1791 break; 1792 1793 case IB_WR_LOCAL_INV: 1794 ctrl->srcrb_flags |= 1795 cpu_to_be32(MLX4_WQE_CTRL_STRONG_ORDER); 1796 set_local_inv_seg(wqe, wr->ex.invalidate_rkey); 1797 wqe += sizeof (struct mlx4_wqe_local_inval_seg); 1798 size += sizeof (struct mlx4_wqe_local_inval_seg) / 16; 1799 break; 1800 1801 case IB_WR_FAST_REG_MR: 1802 ctrl->srcrb_flags |= 1803 cpu_to_be32(MLX4_WQE_CTRL_STRONG_ORDER); 1804 set_fmr_seg(wqe, wr); 1805 wqe += sizeof (struct mlx4_wqe_fmr_seg); 1806 size += sizeof (struct mlx4_wqe_fmr_seg) / 16; 1807 break; 1808 1809 default: 1810 /* No extra segments required for sends */ 1811 break; 1812 } 1813 break; 1814 1815 case IB_QPT_UD: 1816 set_datagram_seg(wqe, wr); 1817 wqe += sizeof (struct mlx4_wqe_datagram_seg); 1818 size += sizeof (struct mlx4_wqe_datagram_seg) / 16; 1819 1820 if (wr->opcode == IB_WR_LSO) { 1821 err = build_lso_seg(wqe, wr, qp, &seglen, &lso_hdr_sz, &blh); 1822 if (unlikely(err)) { 1823 *bad_wr = wr; 1824 goto out; 1825 } 1826 lso_wqe = (__be32 *) wqe; 1827 wqe += seglen; 1828 size += seglen / 16; 1829 } 1830 break; 1831 1832 case IB_QPT_SMI: 1833 case IB_QPT_GSI: 1834 err = build_mlx_header(to_msqp(qp), wr, ctrl, &seglen); 1835 if (unlikely(err)) { 1836 *bad_wr = wr; 1837 goto out; 1838 } 1839 wqe += seglen; 1840 size += seglen / 16; 1841 break; 1842 1843 default: 1844 break; 1845 } 1846 1847 /* 1848 * Write data segments in reverse order, so as to 1849 * overwrite cacheline stamp last within each 1850 * cacheline. This avoids issues with WQE 1851 * prefetching. 1852 */ 1853 1854 dseg = wqe; 1855 dseg += wr->num_sge - 1; 1856 size += wr->num_sge * (sizeof (struct mlx4_wqe_data_seg) / 16); 1857 1858 /* Add one more inline data segment for ICRC for MLX sends */ 1859 if (unlikely(qp->ibqp.qp_type == IB_QPT_SMI || 1860 qp->ibqp.qp_type == IB_QPT_GSI)) { 1861 set_mlx_icrc_seg(dseg + 1); 1862 size += sizeof (struct mlx4_wqe_data_seg) / 16; 1863 } 1864 1865 for (i = wr->num_sge - 1; i >= 0; --i, --dseg) 1866 set_data_seg(dseg, wr->sg_list + i); 1867 1868 /* 1869 * Possibly overwrite stamping in cacheline with LSO 1870 * segment only after making sure all data segments 1871 * are written. 1872 */ 1873 wmb(); 1874 *lso_wqe = lso_hdr_sz; 1875 1876 ctrl->fence_size = (wr->send_flags & IB_SEND_FENCE ? 1877 MLX4_WQE_CTRL_FENCE : 0) | size; 1878 1879 /* 1880 * Make sure descriptor is fully written before 1881 * setting ownership bit (because HW can start 1882 * executing as soon as we do). 1883 */ 1884 wmb(); 1885 1886 if (wr->opcode < 0 || wr->opcode >= ARRAY_SIZE(mlx4_ib_opcode)) { 1887 err = -EINVAL; 1888 goto out; 1889 } 1890 1891 ctrl->owner_opcode = mlx4_ib_opcode[wr->opcode] | 1892 (ind & qp->sq.wqe_cnt ? cpu_to_be32(1 << 31) : 0) | blh; 1893 1894 stamp = ind + qp->sq_spare_wqes; 1895 ind += DIV_ROUND_UP(size * 16, 1U << qp->sq.wqe_shift); 1896 1897 /* 1898 * We can improve latency by not stamping the last 1899 * send queue WQE until after ringing the doorbell, so 1900 * only stamp here if there are still more WQEs to post. 1901 * 1902 * Same optimization applies to padding with NOP wqe 1903 * in case of WQE shrinking (used to prevent wrap-around 1904 * in the middle of WR). 1905 */ 1906 if (wr->next) { 1907 stamp_send_wqe(qp, stamp, size * 16); 1908 ind = pad_wraparound(qp, ind); 1909 } 1910 } 1911 1912 out: 1913 if (likely(nreq)) { 1914 qp->sq.head += nreq; 1915 1916 /* 1917 * Make sure that descriptors are written before 1918 * doorbell record. 1919 */ 1920 wmb(); 1921 1922 writel(qp->doorbell_qpn, 1923 to_mdev(ibqp->device)->uar_map + MLX4_SEND_DOORBELL); 1924 1925 /* 1926 * Make sure doorbells don't leak out of SQ spinlock 1927 * and reach the HCA out of order. 1928 */ 1929 mmiowb(); 1930 1931 stamp_send_wqe(qp, stamp, size * 16); 1932 1933 ind = pad_wraparound(qp, ind); 1934 qp->sq_next_wqe = ind; 1935 } 1936 1937 spin_unlock_irqrestore(&qp->sq.lock, flags); 1938 1939 return err; 1940 } 1941 1942 int mlx4_ib_post_recv(struct ib_qp *ibqp, struct ib_recv_wr *wr, 1943 struct ib_recv_wr **bad_wr) 1944 { 1945 struct mlx4_ib_qp *qp = to_mqp(ibqp); 1946 struct mlx4_wqe_data_seg *scat; 1947 unsigned long flags; 1948 int err = 0; 1949 int nreq; 1950 int ind; 1951 int i; 1952 1953 spin_lock_irqsave(&qp->rq.lock, flags); 1954 1955 ind = qp->rq.head & (qp->rq.wqe_cnt - 1); 1956 1957 for (nreq = 0; wr; ++nreq, wr = wr->next) { 1958 if (mlx4_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) { 1959 err = -ENOMEM; 1960 *bad_wr = wr; 1961 goto out; 1962 } 1963 1964 if (unlikely(wr->num_sge > qp->rq.max_gs)) { 1965 err = -EINVAL; 1966 *bad_wr = wr; 1967 goto out; 1968 } 1969 1970 scat = get_recv_wqe(qp, ind); 1971 1972 for (i = 0; i < wr->num_sge; ++i) 1973 __set_data_seg(scat + i, wr->sg_list + i); 1974 1975 if (i < qp->rq.max_gs) { 1976 scat[i].byte_count = 0; 1977 scat[i].lkey = cpu_to_be32(MLX4_INVALID_LKEY); 1978 scat[i].addr = 0; 1979 } 1980 1981 qp->rq.wrid[ind] = wr->wr_id; 1982 1983 ind = (ind + 1) & (qp->rq.wqe_cnt - 1); 1984 } 1985 1986 out: 1987 if (likely(nreq)) { 1988 qp->rq.head += nreq; 1989 1990 /* 1991 * Make sure that descriptors are written before 1992 * doorbell record. 1993 */ 1994 wmb(); 1995 1996 *qp->db.db = cpu_to_be32(qp->rq.head & 0xffff); 1997 } 1998 1999 spin_unlock_irqrestore(&qp->rq.lock, flags); 2000 2001 return err; 2002 } 2003 2004 static inline enum ib_qp_state to_ib_qp_state(enum mlx4_qp_state mlx4_state) 2005 { 2006 switch (mlx4_state) { 2007 case MLX4_QP_STATE_RST: return IB_QPS_RESET; 2008 case MLX4_QP_STATE_INIT: return IB_QPS_INIT; 2009 case MLX4_QP_STATE_RTR: return IB_QPS_RTR; 2010 case MLX4_QP_STATE_RTS: return IB_QPS_RTS; 2011 case MLX4_QP_STATE_SQ_DRAINING: 2012 case MLX4_QP_STATE_SQD: return IB_QPS_SQD; 2013 case MLX4_QP_STATE_SQER: return IB_QPS_SQE; 2014 case MLX4_QP_STATE_ERR: return IB_QPS_ERR; 2015 default: return -1; 2016 } 2017 } 2018 2019 static inline enum ib_mig_state to_ib_mig_state(int mlx4_mig_state) 2020 { 2021 switch (mlx4_mig_state) { 2022 case MLX4_QP_PM_ARMED: return IB_MIG_ARMED; 2023 case MLX4_QP_PM_REARM: return IB_MIG_REARM; 2024 case MLX4_QP_PM_MIGRATED: return IB_MIG_MIGRATED; 2025 default: return -1; 2026 } 2027 } 2028 2029 static int to_ib_qp_access_flags(int mlx4_flags) 2030 { 2031 int ib_flags = 0; 2032 2033 if (mlx4_flags & MLX4_QP_BIT_RRE) 2034 ib_flags |= IB_ACCESS_REMOTE_READ; 2035 if (mlx4_flags & MLX4_QP_BIT_RWE) 2036 ib_flags |= IB_ACCESS_REMOTE_WRITE; 2037 if (mlx4_flags & MLX4_QP_BIT_RAE) 2038 ib_flags |= IB_ACCESS_REMOTE_ATOMIC; 2039 2040 return ib_flags; 2041 } 2042 2043 static void to_ib_ah_attr(struct mlx4_ib_dev *ibdev, struct ib_ah_attr *ib_ah_attr, 2044 struct mlx4_qp_path *path) 2045 { 2046 struct mlx4_dev *dev = ibdev->dev; 2047 int is_eth; 2048 2049 memset(ib_ah_attr, 0, sizeof *ib_ah_attr); 2050 ib_ah_attr->port_num = path->sched_queue & 0x40 ? 2 : 1; 2051 2052 if (ib_ah_attr->port_num == 0 || ib_ah_attr->port_num > dev->caps.num_ports) 2053 return; 2054 2055 is_eth = rdma_port_get_link_layer(&ibdev->ib_dev, ib_ah_attr->port_num) == 2056 IB_LINK_LAYER_ETHERNET; 2057 if (is_eth) 2058 ib_ah_attr->sl = ((path->sched_queue >> 3) & 0x7) | 2059 ((path->sched_queue & 4) << 1); 2060 else 2061 ib_ah_attr->sl = (path->sched_queue >> 2) & 0xf; 2062 2063 ib_ah_attr->dlid = be16_to_cpu(path->rlid); 2064 ib_ah_attr->src_path_bits = path->grh_mylmc & 0x7f; 2065 ib_ah_attr->static_rate = path->static_rate ? path->static_rate - 5 : 0; 2066 ib_ah_attr->ah_flags = (path->grh_mylmc & (1 << 7)) ? IB_AH_GRH : 0; 2067 if (ib_ah_attr->ah_flags) { 2068 ib_ah_attr->grh.sgid_index = path->mgid_index; 2069 ib_ah_attr->grh.hop_limit = path->hop_limit; 2070 ib_ah_attr->grh.traffic_class = 2071 (be32_to_cpu(path->tclass_flowlabel) >> 20) & 0xff; 2072 ib_ah_attr->grh.flow_label = 2073 be32_to_cpu(path->tclass_flowlabel) & 0xfffff; 2074 memcpy(ib_ah_attr->grh.dgid.raw, 2075 path->rgid, sizeof ib_ah_attr->grh.dgid.raw); 2076 } 2077 } 2078 2079 int mlx4_ib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr, int qp_attr_mask, 2080 struct ib_qp_init_attr *qp_init_attr) 2081 { 2082 struct mlx4_ib_dev *dev = to_mdev(ibqp->device); 2083 struct mlx4_ib_qp *qp = to_mqp(ibqp); 2084 struct mlx4_qp_context context; 2085 int mlx4_state; 2086 int err = 0; 2087 2088 mutex_lock(&qp->mutex); 2089 2090 if (qp->state == IB_QPS_RESET) { 2091 qp_attr->qp_state = IB_QPS_RESET; 2092 goto done; 2093 } 2094 2095 err = mlx4_qp_query(dev->dev, &qp->mqp, &context); 2096 if (err) { 2097 err = -EINVAL; 2098 goto out; 2099 } 2100 2101 mlx4_state = be32_to_cpu(context.flags) >> 28; 2102 2103 qp->state = to_ib_qp_state(mlx4_state); 2104 qp_attr->qp_state = qp->state; 2105 qp_attr->path_mtu = context.mtu_msgmax >> 5; 2106 qp_attr->path_mig_state = 2107 to_ib_mig_state((be32_to_cpu(context.flags) >> 11) & 0x3); 2108 qp_attr->qkey = be32_to_cpu(context.qkey); 2109 qp_attr->rq_psn = be32_to_cpu(context.rnr_nextrecvpsn) & 0xffffff; 2110 qp_attr->sq_psn = be32_to_cpu(context.next_send_psn) & 0xffffff; 2111 qp_attr->dest_qp_num = be32_to_cpu(context.remote_qpn) & 0xffffff; 2112 qp_attr->qp_access_flags = 2113 to_ib_qp_access_flags(be32_to_cpu(context.params2)); 2114 2115 if (qp->ibqp.qp_type == IB_QPT_RC || qp->ibqp.qp_type == IB_QPT_UC) { 2116 to_ib_ah_attr(dev, &qp_attr->ah_attr, &context.pri_path); 2117 to_ib_ah_attr(dev, &qp_attr->alt_ah_attr, &context.alt_path); 2118 qp_attr->alt_pkey_index = context.alt_path.pkey_index & 0x7f; 2119 qp_attr->alt_port_num = qp_attr->alt_ah_attr.port_num; 2120 } 2121 2122 qp_attr->pkey_index = context.pri_path.pkey_index & 0x7f; 2123 if (qp_attr->qp_state == IB_QPS_INIT) 2124 qp_attr->port_num = qp->port; 2125 else 2126 qp_attr->port_num = context.pri_path.sched_queue & 0x40 ? 2 : 1; 2127 2128 /* qp_attr->en_sqd_async_notify is only applicable in modify qp */ 2129 qp_attr->sq_draining = mlx4_state == MLX4_QP_STATE_SQ_DRAINING; 2130 2131 qp_attr->max_rd_atomic = 1 << ((be32_to_cpu(context.params1) >> 21) & 0x7); 2132 2133 qp_attr->max_dest_rd_atomic = 2134 1 << ((be32_to_cpu(context.params2) >> 21) & 0x7); 2135 qp_attr->min_rnr_timer = 2136 (be32_to_cpu(context.rnr_nextrecvpsn) >> 24) & 0x1f; 2137 qp_attr->timeout = context.pri_path.ackto >> 3; 2138 qp_attr->retry_cnt = (be32_to_cpu(context.params1) >> 16) & 0x7; 2139 qp_attr->rnr_retry = (be32_to_cpu(context.params1) >> 13) & 0x7; 2140 qp_attr->alt_timeout = context.alt_path.ackto >> 3; 2141 2142 done: 2143 qp_attr->cur_qp_state = qp_attr->qp_state; 2144 qp_attr->cap.max_recv_wr = qp->rq.wqe_cnt; 2145 qp_attr->cap.max_recv_sge = qp->rq.max_gs; 2146 2147 if (!ibqp->uobject) { 2148 qp_attr->cap.max_send_wr = qp->sq.wqe_cnt; 2149 qp_attr->cap.max_send_sge = qp->sq.max_gs; 2150 } else { 2151 qp_attr->cap.max_send_wr = 0; 2152 qp_attr->cap.max_send_sge = 0; 2153 } 2154 2155 /* 2156 * We don't support inline sends for kernel QPs (yet), and we 2157 * don't know what userspace's value should be. 2158 */ 2159 qp_attr->cap.max_inline_data = 0; 2160 2161 qp_init_attr->cap = qp_attr->cap; 2162 2163 qp_init_attr->create_flags = 0; 2164 if (qp->flags & MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK) 2165 qp_init_attr->create_flags |= IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK; 2166 2167 if (qp->flags & MLX4_IB_QP_LSO) 2168 qp_init_attr->create_flags |= IB_QP_CREATE_IPOIB_UD_LSO; 2169 2170 out: 2171 mutex_unlock(&qp->mutex); 2172 return err; 2173 } 2174 2175