1/* 2 * Copyright (c) 2004 Topspin Communications. All rights reserved. 3 * Copyright (c) 2005 Cisco Systems. All rights reserved. 4 * Copyright (c) 2005 Mellanox Technologies. All rights reserved. 5 * Copyright (c) 2004 Voltaire, Inc. All rights reserved. 6 * 7 * This software is available to you under a choice of one of two 8 * licenses. You may choose to be licensed under the terms of the GNU 9 * General Public License (GPL) Version 2, available from the file 10 * COPYING in the main directory of this source tree, or the 11 * OpenIB.org BSD license below: 12 * 13 * Redistribution and use in source and binary forms, with or 14 * without modification, are permitted provided that the following 15 * conditions are met: 16 * 17 * - Redistributions of source code must retain the above 18 * copyright notice, this list of conditions and the following 19 * disclaimer. 20 * 21 * - Redistributions in binary form must reproduce the above 22 * copyright notice, this list of conditions and the following 23 * disclaimer in the documentation and/or other materials 24 * provided with the distribution. 25 * 26 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, 27 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF 28 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND 29 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS 30 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN 31 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN 32 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE 33 * SOFTWARE. 34 */ 35 36#include <linux/string.h> 37#include <linux/slab.h> 38#include <linux/sched.h> 39 40#include <asm/io.h> 41 42#include <rdma/ib_verbs.h> 43#include <rdma/ib_cache.h> 44#include <rdma/ib_pack.h> 45 46#include "mthca_dev.h" 47#include "mthca_cmd.h" 48#include "mthca_memfree.h" 49#include "mthca_wqe.h" 50 51enum { 52 MTHCA_MAX_DIRECT_QP_SIZE = 4 * PAGE_SIZE, 53 MTHCA_ACK_REQ_FREQ = 10, 54 MTHCA_FLIGHT_LIMIT = 9, 55 MTHCA_UD_HEADER_SIZE = 72, /* largest UD header possible */ 56 MTHCA_INLINE_HEADER_SIZE = 4, /* data segment overhead for inline */ 57 MTHCA_INLINE_CHUNK_SIZE = 16 /* inline data segment chunk */ 58}; 59 60enum { 61 MTHCA_QP_STATE_RST = 0, 62 MTHCA_QP_STATE_INIT = 1, 63 MTHCA_QP_STATE_RTR = 2, 64 MTHCA_QP_STATE_RTS = 3, 65 MTHCA_QP_STATE_SQE = 4, 66 MTHCA_QP_STATE_SQD = 5, 67 MTHCA_QP_STATE_ERR = 6, 68 MTHCA_QP_STATE_DRAINING = 7 69}; 70 71enum { 72 MTHCA_QP_ST_RC = 0x0, 73 MTHCA_QP_ST_UC = 0x1, 74 MTHCA_QP_ST_RD = 0x2, 75 MTHCA_QP_ST_UD = 0x3, 76 MTHCA_QP_ST_MLX = 0x7 77}; 78 79enum { 80 MTHCA_QP_PM_MIGRATED = 0x3, 81 MTHCA_QP_PM_ARMED = 0x0, 82 MTHCA_QP_PM_REARM = 0x1 83}; 84 85enum { 86 /* qp_context flags */ 87 MTHCA_QP_BIT_DE = 1 << 8, 88 /* params1 */ 89 MTHCA_QP_BIT_SRE = 1 << 15, 90 MTHCA_QP_BIT_SWE = 1 << 14, 91 MTHCA_QP_BIT_SAE = 1 << 13, 92 MTHCA_QP_BIT_SIC = 1 << 4, 93 MTHCA_QP_BIT_SSC = 1 << 3, 94 /* params2 */ 95 MTHCA_QP_BIT_RRE = 1 << 15, 96 MTHCA_QP_BIT_RWE = 1 << 14, 97 MTHCA_QP_BIT_RAE = 1 << 13, 98 MTHCA_QP_BIT_RIC = 1 << 4, 99 MTHCA_QP_BIT_RSC = 1 << 3 100}; 101 102enum { 103 MTHCA_SEND_DOORBELL_FENCE = 1 << 5 104}; 105 106struct mthca_qp_path { 107 __be32 port_pkey; 108 u8 rnr_retry; 109 u8 g_mylmc; 110 __be16 rlid; 111 u8 ackto; 112 u8 mgid_index; 113 u8 static_rate; 114 u8 hop_limit; 115 __be32 sl_tclass_flowlabel; 116 u8 rgid[16]; 117} __attribute__((packed)); 118 119struct mthca_qp_context { 120 __be32 flags; 121 __be32 tavor_sched_queue; /* Reserved on Arbel */ 122 u8 mtu_msgmax; 123 u8 rq_size_stride; /* Reserved on Tavor */ 124 u8 sq_size_stride; /* Reserved on Tavor */ 125 u8 rlkey_arbel_sched_queue; /* Reserved on Tavor */ 126 __be32 usr_page; 127 __be32 local_qpn; 128 __be32 remote_qpn; 129 u32 reserved1[2]; 130 struct mthca_qp_path pri_path; 131 struct mthca_qp_path alt_path; 132 __be32 rdd; 133 __be32 pd; 134 __be32 wqe_base; 135 __be32 wqe_lkey; 136 __be32 params1; 137 __be32 reserved2; 138 __be32 next_send_psn; 139 __be32 cqn_snd; 140 __be32 snd_wqe_base_l; /* Next send WQE on Tavor */ 141 __be32 snd_db_index; /* (debugging only entries) */ 142 __be32 last_acked_psn; 143 __be32 ssn; 144 __be32 params2; 145 __be32 rnr_nextrecvpsn; 146 __be32 ra_buff_indx; 147 __be32 cqn_rcv; 148 __be32 rcv_wqe_base_l; /* Next recv WQE on Tavor */ 149 __be32 rcv_db_index; /* (debugging only entries) */ 150 __be32 qkey; 151 __be32 srqn; 152 __be32 rmsn; 153 __be16 rq_wqe_counter; /* reserved on Tavor */ 154 __be16 sq_wqe_counter; /* reserved on Tavor */ 155 u32 reserved3[18]; 156} __attribute__((packed)); 157 158struct mthca_qp_param { 159 __be32 opt_param_mask; 160 u32 reserved1; 161 struct mthca_qp_context context; 162 u32 reserved2[62]; 163} __attribute__((packed)); 164 165enum { 166 MTHCA_QP_OPTPAR_ALT_ADDR_PATH = 1 << 0, 167 MTHCA_QP_OPTPAR_RRE = 1 << 1, 168 MTHCA_QP_OPTPAR_RAE = 1 << 2, 169 MTHCA_QP_OPTPAR_RWE = 1 << 3, 170 MTHCA_QP_OPTPAR_PKEY_INDEX = 1 << 4, 171 MTHCA_QP_OPTPAR_Q_KEY = 1 << 5, 172 MTHCA_QP_OPTPAR_RNR_TIMEOUT = 1 << 6, 173 MTHCA_QP_OPTPAR_PRIMARY_ADDR_PATH = 1 << 7, 174 MTHCA_QP_OPTPAR_SRA_MAX = 1 << 8, 175 MTHCA_QP_OPTPAR_RRA_MAX = 1 << 9, 176 MTHCA_QP_OPTPAR_PM_STATE = 1 << 10, 177 MTHCA_QP_OPTPAR_PORT_NUM = 1 << 11, 178 MTHCA_QP_OPTPAR_RETRY_COUNT = 1 << 12, 179 MTHCA_QP_OPTPAR_ALT_RNR_RETRY = 1 << 13, 180 MTHCA_QP_OPTPAR_ACK_TIMEOUT = 1 << 14, 181 MTHCA_QP_OPTPAR_RNR_RETRY = 1 << 15, 182 MTHCA_QP_OPTPAR_SCHED_QUEUE = 1 << 16 183}; 184 185static const u8 mthca_opcode[] = { 186 [IB_WR_SEND] = MTHCA_OPCODE_SEND, 187 [IB_WR_SEND_WITH_IMM] = MTHCA_OPCODE_SEND_IMM, 188 [IB_WR_RDMA_WRITE] = MTHCA_OPCODE_RDMA_WRITE, 189 [IB_WR_RDMA_WRITE_WITH_IMM] = MTHCA_OPCODE_RDMA_WRITE_IMM, 190 [IB_WR_RDMA_READ] = MTHCA_OPCODE_RDMA_READ, 191 [IB_WR_ATOMIC_CMP_AND_SWP] = MTHCA_OPCODE_ATOMIC_CS, 192 [IB_WR_ATOMIC_FETCH_AND_ADD] = MTHCA_OPCODE_ATOMIC_FA, 193}; 194 195static int is_sqp(struct mthca_dev *dev, struct mthca_qp *qp) 196{ 197 return qp->qpn >= dev->qp_table.sqp_start && 198 qp->qpn <= dev->qp_table.sqp_start + 3; 199} 200 201static int is_qp0(struct mthca_dev *dev, struct mthca_qp *qp) 202{ 203 return qp->qpn >= dev->qp_table.sqp_start && 204 qp->qpn <= dev->qp_table.sqp_start + 1; 205} 206 207static void *get_recv_wqe(struct mthca_qp *qp, int n) 208{ 209 if (qp->is_direct) 210 return qp->queue.direct.buf + (n << qp->rq.wqe_shift); 211 else 212 return qp->queue.page_list[(n << qp->rq.wqe_shift) >> PAGE_SHIFT].buf + 213 ((n << qp->rq.wqe_shift) & (PAGE_SIZE - 1)); 214} 215 216static void *get_send_wqe(struct mthca_qp *qp, int n) 217{ 218 if (qp->is_direct) 219 return qp->queue.direct.buf + qp->send_wqe_offset + 220 (n << qp->sq.wqe_shift); 221 else 222 return qp->queue.page_list[(qp->send_wqe_offset + 223 (n << qp->sq.wqe_shift)) >> 224 PAGE_SHIFT].buf + 225 ((qp->send_wqe_offset + (n << qp->sq.wqe_shift)) & 226 (PAGE_SIZE - 1)); 227} 228 229static void mthca_wq_reset(struct mthca_wq *wq) 230{ 231 wq->next_ind = 0; 232 wq->last_comp = wq->max - 1; 233 wq->head = 0; 234 wq->tail = 0; 235} 236 237void mthca_qp_event(struct mthca_dev *dev, u32 qpn, 238 enum ib_event_type event_type) 239{ 240 struct mthca_qp *qp; 241 struct ib_event event; 242 243 spin_lock(&dev->qp_table.lock); 244 qp = mthca_array_get(&dev->qp_table.qp, qpn & (dev->limits.num_qps - 1)); 245 if (qp) 246 ++qp->refcount; 247 spin_unlock(&dev->qp_table.lock); 248 249 if (!qp) { 250 mthca_warn(dev, "Async event for bogus QP %08x\n", qpn); 251 return; 252 } 253 254 if (event_type == IB_EVENT_PATH_MIG) 255 qp->port = qp->alt_port; 256 257 event.device = &dev->ib_dev; 258 event.event = event_type; 259 event.element.qp = &qp->ibqp; 260 if (qp->ibqp.event_handler) 261 qp->ibqp.event_handler(&event, qp->ibqp.qp_context); 262 263 spin_lock(&dev->qp_table.lock); 264 if (!--qp->refcount) 265 wake_up(&qp->wait); 266 spin_unlock(&dev->qp_table.lock); 267} 268 269static int to_mthca_state(enum ib_qp_state ib_state) 270{ 271 switch (ib_state) { 272 case IB_QPS_RESET: return MTHCA_QP_STATE_RST; 273 case IB_QPS_INIT: return MTHCA_QP_STATE_INIT; 274 case IB_QPS_RTR: return MTHCA_QP_STATE_RTR; 275 case IB_QPS_RTS: return MTHCA_QP_STATE_RTS; 276 case IB_QPS_SQD: return MTHCA_QP_STATE_SQD; 277 case IB_QPS_SQE: return MTHCA_QP_STATE_SQE; 278 case IB_QPS_ERR: return MTHCA_QP_STATE_ERR; 279 default: return -1; 280 } 281} 282 283enum { RC, UC, UD, RD, RDEE, MLX, NUM_TRANS }; 284 285static int to_mthca_st(int transport) 286{ 287 switch (transport) { 288 case RC: return MTHCA_QP_ST_RC; 289 case UC: return MTHCA_QP_ST_UC; 290 case UD: return MTHCA_QP_ST_UD; 291 case RD: return MTHCA_QP_ST_RD; 292 case MLX: return MTHCA_QP_ST_MLX; 293 default: return -1; 294 } 295} 296 297static void store_attrs(struct mthca_sqp *sqp, const struct ib_qp_attr *attr, 298 int attr_mask) 299{ 300 if (attr_mask & IB_QP_PKEY_INDEX) 301 sqp->pkey_index = attr->pkey_index; 302 if (attr_mask & IB_QP_QKEY) 303 sqp->qkey = attr->qkey; 304 if (attr_mask & IB_QP_SQ_PSN) 305 sqp->send_psn = attr->sq_psn; 306} 307 308static void init_port(struct mthca_dev *dev, int port) 309{ 310 int err; 311 u8 status; 312 struct mthca_init_ib_param param; 313 314 memset(¶m, 0, sizeof param); 315 316 param.port_width = dev->limits.port_width_cap; 317 param.vl_cap = dev->limits.vl_cap; 318 param.mtu_cap = dev->limits.mtu_cap; 319 param.gid_cap = dev->limits.gid_table_len; 320 param.pkey_cap = dev->limits.pkey_table_len; 321 322 err = mthca_INIT_IB(dev, ¶m, port, &status); 323 if (err) 324 mthca_warn(dev, "INIT_IB failed, return code %d.\n", err); 325 if (status) 326 mthca_warn(dev, "INIT_IB returned status %02x.\n", status); 327} 328 329static __be32 get_hw_access_flags(struct mthca_qp *qp, const struct ib_qp_attr *attr, 330 int attr_mask) 331{ 332 u8 dest_rd_atomic; 333 u32 access_flags; 334 u32 hw_access_flags = 0; 335 336 if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) 337 dest_rd_atomic = attr->max_dest_rd_atomic; 338 else 339 dest_rd_atomic = qp->resp_depth; 340 341 if (attr_mask & IB_QP_ACCESS_FLAGS) 342 access_flags = attr->qp_access_flags; 343 else 344 access_flags = qp->atomic_rd_en; 345 346 if (!dest_rd_atomic) 347 access_flags &= IB_ACCESS_REMOTE_WRITE; 348 349 if (access_flags & IB_ACCESS_REMOTE_READ) 350 hw_access_flags |= MTHCA_QP_BIT_RRE; 351 if (access_flags & IB_ACCESS_REMOTE_ATOMIC) 352 hw_access_flags |= MTHCA_QP_BIT_RAE; 353 if (access_flags & IB_ACCESS_REMOTE_WRITE) 354 hw_access_flags |= MTHCA_QP_BIT_RWE; 355 356 return cpu_to_be32(hw_access_flags); 357} 358 359static inline enum ib_qp_state to_ib_qp_state(int mthca_state) 360{ 361 switch (mthca_state) { 362 case MTHCA_QP_STATE_RST: return IB_QPS_RESET; 363 case MTHCA_QP_STATE_INIT: return IB_QPS_INIT; 364 case MTHCA_QP_STATE_RTR: return IB_QPS_RTR; 365 case MTHCA_QP_STATE_RTS: return IB_QPS_RTS; 366 case MTHCA_QP_STATE_DRAINING: 367 case MTHCA_QP_STATE_SQD: return IB_QPS_SQD; 368 case MTHCA_QP_STATE_SQE: return IB_QPS_SQE; 369 case MTHCA_QP_STATE_ERR: return IB_QPS_ERR; 370 default: return -1; 371 } 372} 373 374static inline enum ib_mig_state to_ib_mig_state(int mthca_mig_state) 375{ 376 switch (mthca_mig_state) { 377 case 0: return IB_MIG_ARMED; 378 case 1: return IB_MIG_REARM; 379 case 3: return IB_MIG_MIGRATED; 380 default: return -1; 381 } 382} 383 384static int to_ib_qp_access_flags(int mthca_flags) 385{ 386 int ib_flags = 0; 387 388 if (mthca_flags & MTHCA_QP_BIT_RRE) 389 ib_flags |= IB_ACCESS_REMOTE_READ; 390 if (mthca_flags & MTHCA_QP_BIT_RWE) 391 ib_flags |= IB_ACCESS_REMOTE_WRITE; 392 if (mthca_flags & MTHCA_QP_BIT_RAE) 393 ib_flags |= IB_ACCESS_REMOTE_ATOMIC; 394 395 return ib_flags; 396} 397 398static void to_ib_ah_attr(struct mthca_dev *dev, struct ib_ah_attr *ib_ah_attr, 399 struct mthca_qp_path *path) 400{ 401 memset(ib_ah_attr, 0, sizeof *ib_ah_attr); 402 ib_ah_attr->port_num = (be32_to_cpu(path->port_pkey) >> 24) & 0x3; 403 404 if (ib_ah_attr->port_num == 0 || ib_ah_attr->port_num > dev->limits.num_ports) 405 return; 406 407 ib_ah_attr->dlid = be16_to_cpu(path->rlid); 408 ib_ah_attr->sl = be32_to_cpu(path->sl_tclass_flowlabel) >> 28; 409 ib_ah_attr->src_path_bits = path->g_mylmc & 0x7f; 410 ib_ah_attr->static_rate = mthca_rate_to_ib(dev, 411 path->static_rate & 0xf, 412 ib_ah_attr->port_num); 413 ib_ah_attr->ah_flags = (path->g_mylmc & (1 << 7)) ? IB_AH_GRH : 0; 414 if (ib_ah_attr->ah_flags) { 415 ib_ah_attr->grh.sgid_index = path->mgid_index & (dev->limits.gid_table_len - 1); 416 ib_ah_attr->grh.hop_limit = path->hop_limit; 417 ib_ah_attr->grh.traffic_class = 418 (be32_to_cpu(path->sl_tclass_flowlabel) >> 20) & 0xff; 419 ib_ah_attr->grh.flow_label = 420 be32_to_cpu(path->sl_tclass_flowlabel) & 0xfffff; 421 memcpy(ib_ah_attr->grh.dgid.raw, 422 path->rgid, sizeof ib_ah_attr->grh.dgid.raw); 423 } 424} 425 426int mthca_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr, int qp_attr_mask, 427 struct ib_qp_init_attr *qp_init_attr) 428{ 429 struct mthca_dev *dev = to_mdev(ibqp->device); 430 struct mthca_qp *qp = to_mqp(ibqp); 431 int err = 0; 432 struct mthca_mailbox *mailbox = NULL; 433 struct mthca_qp_param *qp_param; 434 struct mthca_qp_context *context; 435 int mthca_state; 436 u8 status; 437 438 mutex_lock(&qp->mutex); 439 440 if (qp->state == IB_QPS_RESET) { 441 qp_attr->qp_state = IB_QPS_RESET; 442 goto done; 443 } 444 445 mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL); 446 if (IS_ERR(mailbox)) { 447 err = PTR_ERR(mailbox); 448 goto out; 449 } 450 451 err = mthca_QUERY_QP(dev, qp->qpn, 0, mailbox, &status); 452 if (err) 453 goto out_mailbox; 454 if (status) { 455 mthca_warn(dev, "QUERY_QP returned status %02x\n", status); 456 err = -EINVAL; 457 goto out_mailbox; 458 } 459 460 qp_param = mailbox->buf; 461 context = &qp_param->context; 462 mthca_state = be32_to_cpu(context->flags) >> 28; 463 464 qp->state = to_ib_qp_state(mthca_state); 465 qp_attr->qp_state = qp->state; 466 qp_attr->path_mtu = context->mtu_msgmax >> 5; 467 qp_attr->path_mig_state = 468 to_ib_mig_state((be32_to_cpu(context->flags) >> 11) & 0x3); 469 qp_attr->qkey = be32_to_cpu(context->qkey); 470 qp_attr->rq_psn = be32_to_cpu(context->rnr_nextrecvpsn) & 0xffffff; 471 qp_attr->sq_psn = be32_to_cpu(context->next_send_psn) & 0xffffff; 472 qp_attr->dest_qp_num = be32_to_cpu(context->remote_qpn) & 0xffffff; 473 qp_attr->qp_access_flags = 474 to_ib_qp_access_flags(be32_to_cpu(context->params2)); 475 476 if (qp->transport == RC || qp->transport == UC) { 477 to_ib_ah_attr(dev, &qp_attr->ah_attr, &context->pri_path); 478 to_ib_ah_attr(dev, &qp_attr->alt_ah_attr, &context->alt_path); 479 qp_attr->alt_pkey_index = 480 be32_to_cpu(context->alt_path.port_pkey) & 0x7f; 481 qp_attr->alt_port_num = qp_attr->alt_ah_attr.port_num; 482 } 483 484 qp_attr->pkey_index = be32_to_cpu(context->pri_path.port_pkey) & 0x7f; 485 qp_attr->port_num = 486 (be32_to_cpu(context->pri_path.port_pkey) >> 24) & 0x3; 487 488 /* qp_attr->en_sqd_async_notify is only applicable in modify qp */ 489 qp_attr->sq_draining = mthca_state == MTHCA_QP_STATE_DRAINING; 490 491 qp_attr->max_rd_atomic = 1 << ((be32_to_cpu(context->params1) >> 21) & 0x7); 492 493 qp_attr->max_dest_rd_atomic = 494 1 << ((be32_to_cpu(context->params2) >> 21) & 0x7); 495 qp_attr->min_rnr_timer = 496 (be32_to_cpu(context->rnr_nextrecvpsn) >> 24) & 0x1f; 497 qp_attr->timeout = context->pri_path.ackto >> 3; 498 qp_attr->retry_cnt = (be32_to_cpu(context->params1) >> 16) & 0x7; 499 qp_attr->rnr_retry = context->pri_path.rnr_retry >> 5; 500 qp_attr->alt_timeout = context->alt_path.ackto >> 3; 501 502done: 503 qp_attr->cur_qp_state = qp_attr->qp_state; 504 qp_attr->cap.max_send_wr = qp->sq.max; 505 qp_attr->cap.max_recv_wr = qp->rq.max; 506 qp_attr->cap.max_send_sge = qp->sq.max_gs; 507 qp_attr->cap.max_recv_sge = qp->rq.max_gs; 508 qp_attr->cap.max_inline_data = qp->max_inline_data; 509 510 qp_init_attr->cap = qp_attr->cap; 511 512out_mailbox: 513 mthca_free_mailbox(dev, mailbox); 514 515out: 516 mutex_unlock(&qp->mutex); 517 return err; 518} 519 520static int mthca_path_set(struct mthca_dev *dev, const struct ib_ah_attr *ah, 521 struct mthca_qp_path *path, u8 port) 522{ 523 path->g_mylmc = ah->src_path_bits & 0x7f; 524 path->rlid = cpu_to_be16(ah->dlid); 525 path->static_rate = mthca_get_rate(dev, ah->static_rate, port); 526 527 if (ah->ah_flags & IB_AH_GRH) { 528 if (ah->grh.sgid_index >= dev->limits.gid_table_len) { 529 mthca_dbg(dev, "sgid_index (%u) too large. max is %d\n", 530 ah->grh.sgid_index, dev->limits.gid_table_len-1); 531 return -1; 532 } 533 534 path->g_mylmc |= 1 << 7; 535 path->mgid_index = ah->grh.sgid_index; 536 path->hop_limit = ah->grh.hop_limit; 537 path->sl_tclass_flowlabel = 538 cpu_to_be32((ah->sl << 28) | 539 (ah->grh.traffic_class << 20) | 540 (ah->grh.flow_label)); 541 memcpy(path->rgid, ah->grh.dgid.raw, 16); 542 } else 543 path->sl_tclass_flowlabel = cpu_to_be32(ah->sl << 28); 544 545 return 0; 546} 547 548static int __mthca_modify_qp(struct ib_qp *ibqp, 549 const struct ib_qp_attr *attr, int attr_mask, 550 enum ib_qp_state cur_state, enum ib_qp_state new_state) 551{ 552 struct mthca_dev *dev = to_mdev(ibqp->device); 553 struct mthca_qp *qp = to_mqp(ibqp); 554 struct mthca_mailbox *mailbox; 555 struct mthca_qp_param *qp_param; 556 struct mthca_qp_context *qp_context; 557 u32 sqd_event = 0; 558 u8 status; 559 int err = -EINVAL; 560 561 mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL); 562 if (IS_ERR(mailbox)) { 563 err = PTR_ERR(mailbox); 564 goto out; 565 } 566 qp_param = mailbox->buf; 567 qp_context = &qp_param->context; 568 memset(qp_param, 0, sizeof *qp_param); 569 570 qp_context->flags = cpu_to_be32((to_mthca_state(new_state) << 28) | 571 (to_mthca_st(qp->transport) << 16)); 572 qp_context->flags |= cpu_to_be32(MTHCA_QP_BIT_DE); 573 if (!(attr_mask & IB_QP_PATH_MIG_STATE)) 574 qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_MIGRATED << 11); 575 else { 576 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PM_STATE); 577 switch (attr->path_mig_state) { 578 case IB_MIG_MIGRATED: 579 qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_MIGRATED << 11); 580 break; 581 case IB_MIG_REARM: 582 qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_REARM << 11); 583 break; 584 case IB_MIG_ARMED: 585 qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_ARMED << 11); 586 break; 587 } 588 } 589 590 /* leave tavor_sched_queue as 0 */ 591 592 if (qp->transport == MLX || qp->transport == UD) 593 qp_context->mtu_msgmax = (IB_MTU_2048 << 5) | 11; 594 else if (attr_mask & IB_QP_PATH_MTU) { 595 if (attr->path_mtu < IB_MTU_256 || attr->path_mtu > IB_MTU_2048) { 596 mthca_dbg(dev, "path MTU (%u) is invalid\n", 597 attr->path_mtu); 598 goto out_mailbox; 599 } 600 qp_context->mtu_msgmax = (attr->path_mtu << 5) | 31; 601 } 602 603 if (mthca_is_memfree(dev)) { 604 if (qp->rq.max) 605 qp_context->rq_size_stride = ilog2(qp->rq.max) << 3; 606 qp_context->rq_size_stride |= qp->rq.wqe_shift - 4; 607 608 if (qp->sq.max) 609 qp_context->sq_size_stride = ilog2(qp->sq.max) << 3; 610 qp_context->sq_size_stride |= qp->sq.wqe_shift - 4; 611 } 612 613 /* leave arbel_sched_queue as 0 */ 614 615 if (qp->ibqp.uobject) 616 qp_context->usr_page = 617 cpu_to_be32(to_mucontext(qp->ibqp.uobject->context)->uar.index); 618 else 619 qp_context->usr_page = cpu_to_be32(dev->driver_uar.index); 620 qp_context->local_qpn = cpu_to_be32(qp->qpn); 621 if (attr_mask & IB_QP_DEST_QPN) { 622 qp_context->remote_qpn = cpu_to_be32(attr->dest_qp_num); 623 } 624 625 if (qp->transport == MLX) 626 qp_context->pri_path.port_pkey |= 627 cpu_to_be32(qp->port << 24); 628 else { 629 if (attr_mask & IB_QP_PORT) { 630 qp_context->pri_path.port_pkey |= 631 cpu_to_be32(attr->port_num << 24); 632 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PORT_NUM); 633 } 634 } 635 636 if (attr_mask & IB_QP_PKEY_INDEX) { 637 qp_context->pri_path.port_pkey |= 638 cpu_to_be32(attr->pkey_index); 639 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PKEY_INDEX); 640 } 641 642 if (attr_mask & IB_QP_RNR_RETRY) { 643 qp_context->alt_path.rnr_retry = qp_context->pri_path.rnr_retry = 644 attr->rnr_retry << 5; 645 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RNR_RETRY | 646 MTHCA_QP_OPTPAR_ALT_RNR_RETRY); 647 } 648 649 if (attr_mask & IB_QP_AV) { 650 if (mthca_path_set(dev, &attr->ah_attr, &qp_context->pri_path, 651 attr_mask & IB_QP_PORT ? attr->port_num : qp->port)) 652 goto out_mailbox; 653 654 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PRIMARY_ADDR_PATH); 655 } 656 657 if (ibqp->qp_type == IB_QPT_RC && 658 cur_state == IB_QPS_INIT && new_state == IB_QPS_RTR) { 659 u8 sched_queue = ibqp->uobject ? 0x2 : 0x1; 660 661 if (mthca_is_memfree(dev)) 662 qp_context->rlkey_arbel_sched_queue |= sched_queue; 663 else 664 qp_context->tavor_sched_queue |= cpu_to_be32(sched_queue); 665 666 qp_param->opt_param_mask |= 667 cpu_to_be32(MTHCA_QP_OPTPAR_SCHED_QUEUE); 668 } 669 670 if (attr_mask & IB_QP_TIMEOUT) { 671 qp_context->pri_path.ackto = attr->timeout << 3; 672 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_ACK_TIMEOUT); 673 } 674 675 if (attr_mask & IB_QP_ALT_PATH) { 676 if (attr->alt_pkey_index >= dev->limits.pkey_table_len) { 677 mthca_dbg(dev, "Alternate P_Key index (%u) too large. max is %d\n", 678 attr->alt_pkey_index, dev->limits.pkey_table_len-1); 679 goto out_mailbox; 680 } 681 682 if (attr->alt_port_num == 0 || attr->alt_port_num > dev->limits.num_ports) { 683 mthca_dbg(dev, "Alternate port number (%u) is invalid\n", 684 attr->alt_port_num); 685 goto out_mailbox; 686 } 687 688 if (mthca_path_set(dev, &attr->alt_ah_attr, &qp_context->alt_path, 689 attr->alt_ah_attr.port_num)) 690 goto out_mailbox; 691 692 qp_context->alt_path.port_pkey |= cpu_to_be32(attr->alt_pkey_index | 693 attr->alt_port_num << 24); 694 qp_context->alt_path.ackto = attr->alt_timeout << 3; 695 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_ALT_ADDR_PATH); 696 } 697 698 /* leave rdd as 0 */ 699 qp_context->pd = cpu_to_be32(to_mpd(ibqp->pd)->pd_num); 700 /* leave wqe_base as 0 (we always create an MR based at 0 for WQs) */ 701 qp_context->wqe_lkey = cpu_to_be32(qp->mr.ibmr.lkey); 702 qp_context->params1 = cpu_to_be32((MTHCA_ACK_REQ_FREQ << 28) | 703 (MTHCA_FLIGHT_LIMIT << 24) | 704 MTHCA_QP_BIT_SWE); 705 if (qp->sq_policy == IB_SIGNAL_ALL_WR) 706 qp_context->params1 |= cpu_to_be32(MTHCA_QP_BIT_SSC); 707 if (attr_mask & IB_QP_RETRY_CNT) { 708 qp_context->params1 |= cpu_to_be32(attr->retry_cnt << 16); 709 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RETRY_COUNT); 710 } 711 712 if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) { 713 if (attr->max_rd_atomic) { 714 qp_context->params1 |= 715 cpu_to_be32(MTHCA_QP_BIT_SRE | 716 MTHCA_QP_BIT_SAE); 717 qp_context->params1 |= 718 cpu_to_be32(fls(attr->max_rd_atomic - 1) << 21); 719 } 720 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_SRA_MAX); 721 } 722 723 if (attr_mask & IB_QP_SQ_PSN) 724 qp_context->next_send_psn = cpu_to_be32(attr->sq_psn); 725 qp_context->cqn_snd = cpu_to_be32(to_mcq(ibqp->send_cq)->cqn); 726 727 if (mthca_is_memfree(dev)) { 728 qp_context->snd_wqe_base_l = cpu_to_be32(qp->send_wqe_offset); 729 qp_context->snd_db_index = cpu_to_be32(qp->sq.db_index); 730 } 731 732 if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) { 733 if (attr->max_dest_rd_atomic) 734 qp_context->params2 |= 735 cpu_to_be32(fls(attr->max_dest_rd_atomic - 1) << 21); 736 737 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RRA_MAX); 738 } 739 740 if (attr_mask & (IB_QP_ACCESS_FLAGS | IB_QP_MAX_DEST_RD_ATOMIC)) { 741 qp_context->params2 |= get_hw_access_flags(qp, attr, attr_mask); 742 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RWE | 743 MTHCA_QP_OPTPAR_RRE | 744 MTHCA_QP_OPTPAR_RAE); 745 } 746 747 qp_context->params2 |= cpu_to_be32(MTHCA_QP_BIT_RSC); 748 749 if (ibqp->srq) 750 qp_context->params2 |= cpu_to_be32(MTHCA_QP_BIT_RIC); 751 752 if (attr_mask & IB_QP_MIN_RNR_TIMER) { 753 qp_context->rnr_nextrecvpsn |= cpu_to_be32(attr->min_rnr_timer << 24); 754 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RNR_TIMEOUT); 755 } 756 if (attr_mask & IB_QP_RQ_PSN) 757 qp_context->rnr_nextrecvpsn |= cpu_to_be32(attr->rq_psn); 758 759 qp_context->ra_buff_indx = 760 cpu_to_be32(dev->qp_table.rdb_base + 761 ((qp->qpn & (dev->limits.num_qps - 1)) * MTHCA_RDB_ENTRY_SIZE << 762 dev->qp_table.rdb_shift)); 763 764 qp_context->cqn_rcv = cpu_to_be32(to_mcq(ibqp->recv_cq)->cqn); 765 766 if (mthca_is_memfree(dev)) 767 qp_context->rcv_db_index = cpu_to_be32(qp->rq.db_index); 768 769 if (attr_mask & IB_QP_QKEY) { 770 qp_context->qkey = cpu_to_be32(attr->qkey); 771 qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_Q_KEY); 772 } 773 774 if (ibqp->srq) 775 qp_context->srqn = cpu_to_be32(1 << 24 | 776 to_msrq(ibqp->srq)->srqn); 777 778 if (cur_state == IB_QPS_RTS && new_state == IB_QPS_SQD && 779 attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY && 780 attr->en_sqd_async_notify) 781 sqd_event = 1 << 31; 782 783 err = mthca_MODIFY_QP(dev, cur_state, new_state, qp->qpn, 0, 784 mailbox, sqd_event, &status); 785 if (err) 786 goto out_mailbox; 787 if (status) { 788 mthca_warn(dev, "modify QP %d->%d returned status %02x.\n", 789 cur_state, new_state, status); 790 err = -EINVAL; 791 goto out_mailbox; 792 } 793 794 qp->state = new_state; 795 if (attr_mask & IB_QP_ACCESS_FLAGS) 796 qp->atomic_rd_en = attr->qp_access_flags; 797 if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) 798 qp->resp_depth = attr->max_dest_rd_atomic; 799 if (attr_mask & IB_QP_PORT) 800 qp->port = attr->port_num; 801 if (attr_mask & IB_QP_ALT_PATH) 802 qp->alt_port = attr->alt_port_num; 803 804 if (is_sqp(dev, qp)) 805 store_attrs(to_msqp(qp), attr, attr_mask); 806 807 /* 808 * If we moved QP0 to RTR, bring the IB link up; if we moved 809 * QP0 to RESET or ERROR, bring the link back down. 810 */ 811 if (is_qp0(dev, qp)) { 812 if (cur_state != IB_QPS_RTR && 813 new_state == IB_QPS_RTR) 814 init_port(dev, qp->port); 815 816 if (cur_state != IB_QPS_RESET && 817 cur_state != IB_QPS_ERR && 818 (new_state == IB_QPS_RESET || 819 new_state == IB_QPS_ERR)) 820 mthca_CLOSE_IB(dev, qp->port, &status); 821 } 822 823 /* 824 * If we moved a kernel QP to RESET, clean up all old CQ 825 * entries and reinitialize the QP. 826 */ 827 if (new_state == IB_QPS_RESET && !qp->ibqp.uobject) { 828 mthca_cq_clean(dev, to_mcq(qp->ibqp.recv_cq), qp->qpn, 829 qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL); 830 if (qp->ibqp.send_cq != qp->ibqp.recv_cq) 831 mthca_cq_clean(dev, to_mcq(qp->ibqp.send_cq), qp->qpn, NULL); 832 833 mthca_wq_reset(&qp->sq); 834 qp->sq.last = get_send_wqe(qp, qp->sq.max - 1); 835 836 mthca_wq_reset(&qp->rq); 837 qp->rq.last = get_recv_wqe(qp, qp->rq.max - 1); 838 839 if (mthca_is_memfree(dev)) { 840 *qp->sq.db = 0; 841 *qp->rq.db = 0; 842 } 843 } 844 845out_mailbox: 846 mthca_free_mailbox(dev, mailbox); 847out: 848 return err; 849} 850 851int mthca_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, int attr_mask, 852 struct ib_udata *udata) 853{ 854 struct mthca_dev *dev = to_mdev(ibqp->device); 855 struct mthca_qp *qp = to_mqp(ibqp); 856 enum ib_qp_state cur_state, new_state; 857 int err = -EINVAL; 858 859 mutex_lock(&qp->mutex); 860 if (attr_mask & IB_QP_CUR_STATE) { 861 cur_state = attr->cur_qp_state; 862 } else { 863 spin_lock_irq(&qp->sq.lock); 864 spin_lock(&qp->rq.lock); 865 cur_state = qp->state; 866 spin_unlock(&qp->rq.lock); 867 spin_unlock_irq(&qp->sq.lock); 868 } 869 870 new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state; 871 872 if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type, attr_mask)) { 873 mthca_dbg(dev, "Bad QP transition (transport %d) " 874 "%d->%d with attr 0x%08x\n", 875 qp->transport, cur_state, new_state, 876 attr_mask); 877 goto out; 878 } 879 880 if ((attr_mask & IB_QP_PKEY_INDEX) && 881 attr->pkey_index >= dev->limits.pkey_table_len) { 882 mthca_dbg(dev, "P_Key index (%u) too large. max is %d\n", 883 attr->pkey_index, dev->limits.pkey_table_len-1); 884 goto out; 885 } 886 887 if ((attr_mask & IB_QP_PORT) && 888 (attr->port_num == 0 || attr->port_num > dev->limits.num_ports)) { 889 mthca_dbg(dev, "Port number (%u) is invalid\n", attr->port_num); 890 goto out; 891 } 892 893 if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC && 894 attr->max_rd_atomic > dev->limits.max_qp_init_rdma) { 895 mthca_dbg(dev, "Max rdma_atomic as initiator %u too large (max is %d)\n", 896 attr->max_rd_atomic, dev->limits.max_qp_init_rdma); 897 goto out; 898 } 899 900 if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC && 901 attr->max_dest_rd_atomic > 1 << dev->qp_table.rdb_shift) { 902 mthca_dbg(dev, "Max rdma_atomic as responder %u too large (max %d)\n", 903 attr->max_dest_rd_atomic, 1 << dev->qp_table.rdb_shift); 904 goto out; 905 } 906 907 if (cur_state == new_state && cur_state == IB_QPS_RESET) { 908 err = 0; 909 goto out; 910 } 911 912 err = __mthca_modify_qp(ibqp, attr, attr_mask, cur_state, new_state); 913 914out: 915 mutex_unlock(&qp->mutex); 916 return err; 917} 918 919static int mthca_max_data_size(struct mthca_dev *dev, struct mthca_qp *qp, int desc_sz) 920{ 921 /* 922 * Calculate the maximum size of WQE s/g segments, excluding 923 * the next segment and other non-data segments. 924 */ 925 int max_data_size = desc_sz - sizeof (struct mthca_next_seg); 926 927 switch (qp->transport) { 928 case MLX: 929 max_data_size -= 2 * sizeof (struct mthca_data_seg); 930 break; 931 932 case UD: 933 if (mthca_is_memfree(dev)) 934 max_data_size -= sizeof (struct mthca_arbel_ud_seg); 935 else 936 max_data_size -= sizeof (struct mthca_tavor_ud_seg); 937 break; 938 939 default: 940 max_data_size -= sizeof (struct mthca_raddr_seg); 941 break; 942 } 943 944 return max_data_size; 945} 946 947static inline int mthca_max_inline_data(struct mthca_pd *pd, int max_data_size) 948{ 949 /* We don't support inline data for kernel QPs (yet). */ 950 return pd->ibpd.uobject ? max_data_size - MTHCA_INLINE_HEADER_SIZE : 0; 951} 952 953static void mthca_adjust_qp_caps(struct mthca_dev *dev, 954 struct mthca_pd *pd, 955 struct mthca_qp *qp) 956{ 957 int max_data_size = mthca_max_data_size(dev, qp, 958 min(dev->limits.max_desc_sz, 959 1 << qp->sq.wqe_shift)); 960 961 qp->max_inline_data = mthca_max_inline_data(pd, max_data_size); 962 963 qp->sq.max_gs = min_t(int, dev->limits.max_sg, 964 max_data_size / sizeof (struct mthca_data_seg)); 965 qp->rq.max_gs = min_t(int, dev->limits.max_sg, 966 (min(dev->limits.max_desc_sz, 1 << qp->rq.wqe_shift) - 967 sizeof (struct mthca_next_seg)) / 968 sizeof (struct mthca_data_seg)); 969} 970 971/* 972 * Allocate and register buffer for WQEs. qp->rq.max, sq.max, 973 * rq.max_gs and sq.max_gs must all be assigned. 974 * mthca_alloc_wqe_buf will calculate rq.wqe_shift and 975 * sq.wqe_shift (as well as send_wqe_offset, is_direct, and 976 * queue) 977 */ 978static int mthca_alloc_wqe_buf(struct mthca_dev *dev, 979 struct mthca_pd *pd, 980 struct mthca_qp *qp) 981{ 982 int size; 983 int err = -ENOMEM; 984 985 size = sizeof (struct mthca_next_seg) + 986 qp->rq.max_gs * sizeof (struct mthca_data_seg); 987 988 if (size > dev->limits.max_desc_sz) 989 return -EINVAL; 990 991 for (qp->rq.wqe_shift = 6; 1 << qp->rq.wqe_shift < size; 992 qp->rq.wqe_shift++) 993 ; /* nothing */ 994 995 size = qp->sq.max_gs * sizeof (struct mthca_data_seg); 996 switch (qp->transport) { 997 case MLX: 998 size += 2 * sizeof (struct mthca_data_seg); 999 break; 1000 1001 case UD: 1002 size += mthca_is_memfree(dev) ? 1003 sizeof (struct mthca_arbel_ud_seg) : 1004 sizeof (struct mthca_tavor_ud_seg); 1005 break; 1006 1007 case UC: 1008 size += sizeof (struct mthca_raddr_seg); 1009 break; 1010 1011 case RC: 1012 size += sizeof (struct mthca_raddr_seg); 1013 /* 1014 * An atomic op will require an atomic segment, a 1015 * remote address segment and one scatter entry. 1016 */ 1017 size = max_t(int, size, 1018 sizeof (struct mthca_atomic_seg) + 1019 sizeof (struct mthca_raddr_seg) + 1020 sizeof (struct mthca_data_seg)); 1021 break; 1022 1023 default: 1024 break; 1025 } 1026 1027 /* Make sure that we have enough space for a bind request */ 1028 size = max_t(int, size, sizeof (struct mthca_bind_seg)); 1029 1030 size += sizeof (struct mthca_next_seg); 1031 1032 if (size > dev->limits.max_desc_sz) 1033 return -EINVAL; 1034 1035 for (qp->sq.wqe_shift = 6; 1 << qp->sq.wqe_shift < size; 1036 qp->sq.wqe_shift++) 1037 ; /* nothing */ 1038 1039 qp->send_wqe_offset = ALIGN(qp->rq.max << qp->rq.wqe_shift, 1040 1 << qp->sq.wqe_shift); 1041 1042 /* 1043 * If this is a userspace QP, we don't actually have to 1044 * allocate anything. All we need is to calculate the WQE 1045 * sizes and the send_wqe_offset, so we're done now. 1046 */ 1047 if (pd->ibpd.uobject) 1048 return 0; 1049 1050 size = PAGE_ALIGN(qp->send_wqe_offset + 1051 (qp->sq.max << qp->sq.wqe_shift)); 1052 1053 qp->wrid = kmalloc((qp->rq.max + qp->sq.max) * sizeof (u64), 1054 GFP_KERNEL); 1055 if (!qp->wrid) 1056 goto err_out; 1057 1058 err = mthca_buf_alloc(dev, size, MTHCA_MAX_DIRECT_QP_SIZE, 1059 &qp->queue, &qp->is_direct, pd, 0, &qp->mr); 1060 if (err) 1061 goto err_out; 1062 1063 return 0; 1064 1065err_out: 1066 kfree(qp->wrid); 1067 return err; 1068} 1069 1070static void mthca_free_wqe_buf(struct mthca_dev *dev, 1071 struct mthca_qp *qp) 1072{ 1073 mthca_buf_free(dev, PAGE_ALIGN(qp->send_wqe_offset + 1074 (qp->sq.max << qp->sq.wqe_shift)), 1075 &qp->queue, qp->is_direct, &qp->mr); 1076 kfree(qp->wrid); 1077} 1078 1079static int mthca_map_memfree(struct mthca_dev *dev, 1080 struct mthca_qp *qp) 1081{ 1082 int ret; 1083 1084 if (mthca_is_memfree(dev)) { 1085 ret = mthca_table_get(dev, dev->qp_table.qp_table, qp->qpn); 1086 if (ret) 1087 return ret; 1088 1089 ret = mthca_table_get(dev, dev->qp_table.eqp_table, qp->qpn); 1090 if (ret) 1091 goto err_qpc; 1092 1093 ret = mthca_table_get(dev, dev->qp_table.rdb_table, 1094 qp->qpn << dev->qp_table.rdb_shift); 1095 if (ret) 1096 goto err_eqpc; 1097 1098 } 1099 1100 return 0; 1101 1102err_eqpc: 1103 mthca_table_put(dev, dev->qp_table.eqp_table, qp->qpn); 1104 1105err_qpc: 1106 mthca_table_put(dev, dev->qp_table.qp_table, qp->qpn); 1107 1108 return ret; 1109} 1110 1111static void mthca_unmap_memfree(struct mthca_dev *dev, 1112 struct mthca_qp *qp) 1113{ 1114 mthca_table_put(dev, dev->qp_table.rdb_table, 1115 qp->qpn << dev->qp_table.rdb_shift); 1116 mthca_table_put(dev, dev->qp_table.eqp_table, qp->qpn); 1117 mthca_table_put(dev, dev->qp_table.qp_table, qp->qpn); 1118} 1119 1120static int mthca_alloc_memfree(struct mthca_dev *dev, 1121 struct mthca_qp *qp) 1122{ 1123 if (mthca_is_memfree(dev)) { 1124 qp->rq.db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_RQ, 1125 qp->qpn, &qp->rq.db); 1126 if (qp->rq.db_index < 0) 1127 return -ENOMEM; 1128 1129 qp->sq.db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_SQ, 1130 qp->qpn, &qp->sq.db); 1131 if (qp->sq.db_index < 0) { 1132 mthca_free_db(dev, MTHCA_DB_TYPE_RQ, qp->rq.db_index); 1133 return -ENOMEM; 1134 } 1135 } 1136 1137 return 0; 1138} 1139 1140static void mthca_free_memfree(struct mthca_dev *dev, 1141 struct mthca_qp *qp) 1142{ 1143 if (mthca_is_memfree(dev)) { 1144 mthca_free_db(dev, MTHCA_DB_TYPE_SQ, qp->sq.db_index); 1145 mthca_free_db(dev, MTHCA_DB_TYPE_RQ, qp->rq.db_index); 1146 } 1147} 1148 1149static int mthca_alloc_qp_common(struct mthca_dev *dev, 1150 struct mthca_pd *pd, 1151 struct mthca_cq *send_cq, 1152 struct mthca_cq *recv_cq, 1153 enum ib_sig_type send_policy, 1154 struct mthca_qp *qp) 1155{ 1156 int ret; 1157 int i; 1158 struct mthca_next_seg *next; 1159 1160 qp->refcount = 1; 1161 init_waitqueue_head(&qp->wait); 1162 mutex_init(&qp->mutex); 1163 qp->state = IB_QPS_RESET; 1164 qp->atomic_rd_en = 0; 1165 qp->resp_depth = 0; 1166 qp->sq_policy = send_policy; 1167 mthca_wq_reset(&qp->sq); 1168 mthca_wq_reset(&qp->rq); 1169 1170 spin_lock_init(&qp->sq.lock); 1171 spin_lock_init(&qp->rq.lock); 1172 1173 ret = mthca_map_memfree(dev, qp); 1174 if (ret) 1175 return ret; 1176 1177 ret = mthca_alloc_wqe_buf(dev, pd, qp); 1178 if (ret) { 1179 mthca_unmap_memfree(dev, qp); 1180 return ret; 1181 } 1182 1183 mthca_adjust_qp_caps(dev, pd, qp); 1184 1185 /* 1186 * If this is a userspace QP, we're done now. The doorbells 1187 * will be allocated and buffers will be initialized in 1188 * userspace. 1189 */ 1190 if (pd->ibpd.uobject) 1191 return 0; 1192 1193 ret = mthca_alloc_memfree(dev, qp); 1194 if (ret) { 1195 mthca_free_wqe_buf(dev, qp); 1196 mthca_unmap_memfree(dev, qp); 1197 return ret; 1198 } 1199 1200 if (mthca_is_memfree(dev)) { 1201 struct mthca_data_seg *scatter; 1202 int size = (sizeof (struct mthca_next_seg) + 1203 qp->rq.max_gs * sizeof (struct mthca_data_seg)) / 16; 1204 1205 for (i = 0; i < qp->rq.max; ++i) { 1206 next = get_recv_wqe(qp, i); 1207 next->nda_op = cpu_to_be32(((i + 1) & (qp->rq.max - 1)) << 1208 qp->rq.wqe_shift); 1209 next->ee_nds = cpu_to_be32(size); 1210 1211 for (scatter = (void *) (next + 1); 1212 (void *) scatter < (void *) next + (1 << qp->rq.wqe_shift); 1213 ++scatter) 1214 scatter->lkey = cpu_to_be32(MTHCA_INVAL_LKEY); 1215 } 1216 1217 for (i = 0; i < qp->sq.max; ++i) { 1218 next = get_send_wqe(qp, i); 1219 next->nda_op = cpu_to_be32((((i + 1) & (qp->sq.max - 1)) << 1220 qp->sq.wqe_shift) + 1221 qp->send_wqe_offset); 1222 } 1223 } else { 1224 for (i = 0; i < qp->rq.max; ++i) { 1225 next = get_recv_wqe(qp, i); 1226 next->nda_op = htonl((((i + 1) % qp->rq.max) << 1227 qp->rq.wqe_shift) | 1); 1228 } 1229 1230 } 1231 1232 qp->sq.last = get_send_wqe(qp, qp->sq.max - 1); 1233 qp->rq.last = get_recv_wqe(qp, qp->rq.max - 1); 1234 1235 return 0; 1236} 1237 1238static int mthca_set_qp_size(struct mthca_dev *dev, struct ib_qp_cap *cap, 1239 struct mthca_pd *pd, struct mthca_qp *qp) 1240{ 1241 int max_data_size = mthca_max_data_size(dev, qp, dev->limits.max_desc_sz); 1242 1243 /* Sanity check QP size before proceeding */ 1244 if (cap->max_send_wr > dev->limits.max_wqes || 1245 cap->max_recv_wr > dev->limits.max_wqes || 1246 cap->max_send_sge > dev->limits.max_sg || 1247 cap->max_recv_sge > dev->limits.max_sg || 1248 cap->max_inline_data > mthca_max_inline_data(pd, max_data_size)) 1249 return -EINVAL; 1250 1251 /* 1252 * For MLX transport we need 2 extra send gather entries: 1253 * one for the header and one for the checksum at the end 1254 */ 1255 if (qp->transport == MLX && cap->max_send_sge + 2 > dev->limits.max_sg) 1256 return -EINVAL; 1257 1258 if (mthca_is_memfree(dev)) { 1259 qp->rq.max = cap->max_recv_wr ? 1260 roundup_pow_of_two(cap->max_recv_wr) : 0; 1261 qp->sq.max = cap->max_send_wr ? 1262 roundup_pow_of_two(cap->max_send_wr) : 0; 1263 } else { 1264 qp->rq.max = cap->max_recv_wr; 1265 qp->sq.max = cap->max_send_wr; 1266 } 1267 1268 qp->rq.max_gs = cap->max_recv_sge; 1269 qp->sq.max_gs = max_t(int, cap->max_send_sge, 1270 ALIGN(cap->max_inline_data + MTHCA_INLINE_HEADER_SIZE, 1271 MTHCA_INLINE_CHUNK_SIZE) / 1272 sizeof (struct mthca_data_seg)); 1273 1274 return 0; 1275} 1276 1277int mthca_alloc_qp(struct mthca_dev *dev, 1278 struct mthca_pd *pd, 1279 struct mthca_cq *send_cq, 1280 struct mthca_cq *recv_cq, 1281 enum ib_qp_type type, 1282 enum ib_sig_type send_policy, 1283 struct ib_qp_cap *cap, 1284 struct mthca_qp *qp) 1285{ 1286 int err; 1287 1288 switch (type) { 1289 case IB_QPT_RC: qp->transport = RC; break; 1290 case IB_QPT_UC: qp->transport = UC; break; 1291 case IB_QPT_UD: qp->transport = UD; break; 1292 default: return -EINVAL; 1293 } 1294 1295 err = mthca_set_qp_size(dev, cap, pd, qp); 1296 if (err) 1297 return err; 1298 1299 qp->qpn = mthca_alloc(&dev->qp_table.alloc); 1300 if (qp->qpn == -1) 1301 return -ENOMEM; 1302 1303 /* initialize port to zero for error-catching. */ 1304 qp->port = 0; 1305 1306 err = mthca_alloc_qp_common(dev, pd, send_cq, recv_cq, 1307 send_policy, qp); 1308 if (err) { 1309 mthca_free(&dev->qp_table.alloc, qp->qpn); 1310 return err; 1311 } 1312 1313 spin_lock_irq(&dev->qp_table.lock); 1314 mthca_array_set(&dev->qp_table.qp, 1315 qp->qpn & (dev->limits.num_qps - 1), qp); 1316 spin_unlock_irq(&dev->qp_table.lock); 1317 1318 return 0; 1319} 1320 1321static void mthca_lock_cqs(struct mthca_cq *send_cq, struct mthca_cq *recv_cq) 1322 __acquires(&send_cq->lock) __acquires(&recv_cq->lock) 1323{ 1324 if (send_cq == recv_cq) { 1325 spin_lock_irq(&send_cq->lock); 1326 __acquire(&recv_cq->lock); 1327 } else if (send_cq->cqn < recv_cq->cqn) { 1328 spin_lock_irq(&send_cq->lock); 1329 spin_lock_nested(&recv_cq->lock, SINGLE_DEPTH_NESTING); 1330 } else { 1331 spin_lock_irq(&recv_cq->lock); 1332 spin_lock_nested(&send_cq->lock, SINGLE_DEPTH_NESTING); 1333 } 1334} 1335 1336static void mthca_unlock_cqs(struct mthca_cq *send_cq, struct mthca_cq *recv_cq) 1337 __releases(&send_cq->lock) __releases(&recv_cq->lock) 1338{ 1339 if (send_cq == recv_cq) { 1340 __release(&recv_cq->lock); 1341 spin_unlock_irq(&send_cq->lock); 1342 } else if (send_cq->cqn < recv_cq->cqn) { 1343 spin_unlock(&recv_cq->lock); 1344 spin_unlock_irq(&send_cq->lock); 1345 } else { 1346 spin_unlock(&send_cq->lock); 1347 spin_unlock_irq(&recv_cq->lock); 1348 } 1349} 1350 1351int mthca_alloc_sqp(struct mthca_dev *dev, 1352 struct mthca_pd *pd, 1353 struct mthca_cq *send_cq, 1354 struct mthca_cq *recv_cq, 1355 enum ib_sig_type send_policy, 1356 struct ib_qp_cap *cap, 1357 int qpn, 1358 int port, 1359 struct mthca_sqp *sqp) 1360{ 1361 u32 mqpn = qpn * 2 + dev->qp_table.sqp_start + port - 1; 1362 int err; 1363 1364 sqp->qp.transport = MLX; 1365 err = mthca_set_qp_size(dev, cap, pd, &sqp->qp); 1366 if (err) 1367 return err; 1368 1369 sqp->header_buf_size = sqp->qp.sq.max * MTHCA_UD_HEADER_SIZE; 1370 sqp->header_buf = dma_alloc_coherent(&dev->pdev->dev, sqp->header_buf_size, 1371 &sqp->header_dma, GFP_KERNEL); 1372 if (!sqp->header_buf) 1373 return -ENOMEM; 1374 1375 spin_lock_irq(&dev->qp_table.lock); 1376 if (mthca_array_get(&dev->qp_table.qp, mqpn)) 1377 err = -EBUSY; 1378 else 1379 mthca_array_set(&dev->qp_table.qp, mqpn, sqp); 1380 spin_unlock_irq(&dev->qp_table.lock); 1381 1382 if (err) 1383 goto err_out; 1384 1385 sqp->qp.port = port; 1386 sqp->qp.qpn = mqpn; 1387 sqp->qp.transport = MLX; 1388 1389 err = mthca_alloc_qp_common(dev, pd, send_cq, recv_cq, 1390 send_policy, &sqp->qp); 1391 if (err) 1392 goto err_out_free; 1393 1394 atomic_inc(&pd->sqp_count); 1395 1396 return 0; 1397 1398 err_out_free: 1399 /* 1400 * Lock CQs here, so that CQ polling code can do QP lookup 1401 * without taking a lock. 1402 */ 1403 mthca_lock_cqs(send_cq, recv_cq); 1404 1405 spin_lock(&dev->qp_table.lock); 1406 mthca_array_clear(&dev->qp_table.qp, mqpn); 1407 spin_unlock(&dev->qp_table.lock); 1408 1409 mthca_unlock_cqs(send_cq, recv_cq); 1410 1411 err_out: 1412 dma_free_coherent(&dev->pdev->dev, sqp->header_buf_size, 1413 sqp->header_buf, sqp->header_dma); 1414 1415 return err; 1416} 1417 1418static inline int get_qp_refcount(struct mthca_dev *dev, struct mthca_qp *qp) 1419{ 1420 int c; 1421 1422 spin_lock_irq(&dev->qp_table.lock); 1423 c = qp->refcount; 1424 spin_unlock_irq(&dev->qp_table.lock); 1425 1426 return c; 1427} 1428 1429void mthca_free_qp(struct mthca_dev *dev, 1430 struct mthca_qp *qp) 1431{ 1432 u8 status; 1433 struct mthca_cq *send_cq; 1434 struct mthca_cq *recv_cq; 1435 1436 send_cq = to_mcq(qp->ibqp.send_cq); 1437 recv_cq = to_mcq(qp->ibqp.recv_cq); 1438 1439 /* 1440 * Lock CQs here, so that CQ polling code can do QP lookup 1441 * without taking a lock. 1442 */ 1443 mthca_lock_cqs(send_cq, recv_cq); 1444 1445 spin_lock(&dev->qp_table.lock); 1446 mthca_array_clear(&dev->qp_table.qp, 1447 qp->qpn & (dev->limits.num_qps - 1)); 1448 --qp->refcount; 1449 spin_unlock(&dev->qp_table.lock); 1450 1451 mthca_unlock_cqs(send_cq, recv_cq); 1452 1453 wait_event(qp->wait, !get_qp_refcount(dev, qp)); 1454 1455 if (qp->state != IB_QPS_RESET) 1456 mthca_MODIFY_QP(dev, qp->state, IB_QPS_RESET, qp->qpn, 0, 1457 NULL, 0, &status); 1458 1459 /* 1460 * If this is a userspace QP, the buffers, MR, CQs and so on 1461 * will be cleaned up in userspace, so all we have to do is 1462 * unref the mem-free tables and free the QPN in our table. 1463 */ 1464 if (!qp->ibqp.uobject) { 1465 mthca_cq_clean(dev, recv_cq, qp->qpn, 1466 qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL); 1467 if (send_cq != recv_cq) 1468 mthca_cq_clean(dev, send_cq, qp->qpn, NULL); 1469 1470 mthca_free_memfree(dev, qp); 1471 mthca_free_wqe_buf(dev, qp); 1472 } 1473 1474 mthca_unmap_memfree(dev, qp); 1475 1476 if (is_sqp(dev, qp)) { 1477 atomic_dec(&(to_mpd(qp->ibqp.pd)->sqp_count)); 1478 dma_free_coherent(&dev->pdev->dev, 1479 to_msqp(qp)->header_buf_size, 1480 to_msqp(qp)->header_buf, 1481 to_msqp(qp)->header_dma); 1482 } else 1483 mthca_free(&dev->qp_table.alloc, qp->qpn); 1484} 1485 1486/* Create UD header for an MLX send and build a data segment for it */ 1487static int build_mlx_header(struct mthca_dev *dev, struct mthca_sqp *sqp, 1488 int ind, struct ib_send_wr *wr, 1489 struct mthca_mlx_seg *mlx, 1490 struct mthca_data_seg *data) 1491{ 1492 int header_size; 1493 int err; 1494 u16 pkey; 1495 1496 ib_ud_header_init(256, /* assume a MAD */ 1497 mthca_ah_grh_present(to_mah(wr->wr.ud.ah)), 0, 1498 &sqp->ud_header); 1499 1500 err = mthca_read_ah(dev, to_mah(wr->wr.ud.ah), &sqp->ud_header); 1501 if (err) 1502 return err; 1503 mlx->flags &= ~cpu_to_be32(MTHCA_NEXT_SOLICIT | 1); 1504 mlx->flags |= cpu_to_be32((!sqp->qp.ibqp.qp_num ? MTHCA_MLX_VL15 : 0) | 1505 (sqp->ud_header.lrh.destination_lid == 1506 IB_LID_PERMISSIVE ? MTHCA_MLX_SLR : 0) | 1507 (sqp->ud_header.lrh.service_level << 8)); 1508 mlx->rlid = sqp->ud_header.lrh.destination_lid; 1509 mlx->vcrc = 0; 1510 1511 switch (wr->opcode) { 1512 case IB_WR_SEND: 1513 sqp->ud_header.bth.opcode = IB_OPCODE_UD_SEND_ONLY; 1514 sqp->ud_header.immediate_present = 0; 1515 break; 1516 case IB_WR_SEND_WITH_IMM: 1517 sqp->ud_header.bth.opcode = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE; 1518 sqp->ud_header.immediate_present = 1; 1519 sqp->ud_header.immediate_data = wr->ex.imm_data; 1520 break; 1521 default: 1522 return -EINVAL; 1523 } 1524 1525 sqp->ud_header.lrh.virtual_lane = !sqp->qp.ibqp.qp_num ? 15 : 0; 1526 if (sqp->ud_header.lrh.destination_lid == IB_LID_PERMISSIVE) 1527 sqp->ud_header.lrh.source_lid = IB_LID_PERMISSIVE; 1528 sqp->ud_header.bth.solicited_event = !!(wr->send_flags & IB_SEND_SOLICITED); 1529 if (!sqp->qp.ibqp.qp_num) 1530 ib_get_cached_pkey(&dev->ib_dev, sqp->qp.port, 1531 sqp->pkey_index, &pkey); 1532 else 1533 ib_get_cached_pkey(&dev->ib_dev, sqp->qp.port, 1534 wr->wr.ud.pkey_index, &pkey); 1535 sqp->ud_header.bth.pkey = cpu_to_be16(pkey); 1536 sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->wr.ud.remote_qpn); 1537 sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1)); 1538 sqp->ud_header.deth.qkey = cpu_to_be32(wr->wr.ud.remote_qkey & 0x80000000 ? 1539 sqp->qkey : wr->wr.ud.remote_qkey); 1540 sqp->ud_header.deth.source_qpn = cpu_to_be32(sqp->qp.ibqp.qp_num); 1541 1542 header_size = ib_ud_header_pack(&sqp->ud_header, 1543 sqp->header_buf + 1544 ind * MTHCA_UD_HEADER_SIZE); 1545 1546 data->byte_count = cpu_to_be32(header_size); 1547 data->lkey = cpu_to_be32(to_mpd(sqp->qp.ibqp.pd)->ntmr.ibmr.lkey); 1548 data->addr = cpu_to_be64(sqp->header_dma + 1549 ind * MTHCA_UD_HEADER_SIZE); 1550 1551 return 0; 1552} 1553 1554static inline int mthca_wq_overflow(struct mthca_wq *wq, int nreq, 1555 struct ib_cq *ib_cq) 1556{ 1557 unsigned cur; 1558 struct mthca_cq *cq; 1559 1560 cur = wq->head - wq->tail; 1561 if (likely(cur + nreq < wq->max)) 1562 return 0; 1563 1564 cq = to_mcq(ib_cq); 1565 spin_lock(&cq->lock); 1566 cur = wq->head - wq->tail; 1567 spin_unlock(&cq->lock); 1568 1569 return cur + nreq >= wq->max; 1570} 1571 1572static __always_inline void set_raddr_seg(struct mthca_raddr_seg *rseg, 1573 u64 remote_addr, u32 rkey) 1574{ 1575 rseg->raddr = cpu_to_be64(remote_addr); 1576 rseg->rkey = cpu_to_be32(rkey); 1577 rseg->reserved = 0; 1578} 1579 1580static __always_inline void set_atomic_seg(struct mthca_atomic_seg *aseg, 1581 struct ib_send_wr *wr) 1582{ 1583 if (wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP) { 1584 aseg->swap_add = cpu_to_be64(wr->wr.atomic.swap); 1585 aseg->compare = cpu_to_be64(wr->wr.atomic.compare_add); 1586 } else { 1587 aseg->swap_add = cpu_to_be64(wr->wr.atomic.compare_add); 1588 aseg->compare = 0; 1589 } 1590 1591} 1592 1593static void set_tavor_ud_seg(struct mthca_tavor_ud_seg *useg, 1594 struct ib_send_wr *wr) 1595{ 1596 useg->lkey = cpu_to_be32(to_mah(wr->wr.ud.ah)->key); 1597 useg->av_addr = cpu_to_be64(to_mah(wr->wr.ud.ah)->avdma); 1598 useg->dqpn = cpu_to_be32(wr->wr.ud.remote_qpn); 1599 useg->qkey = cpu_to_be32(wr->wr.ud.remote_qkey); 1600 1601} 1602 1603static void set_arbel_ud_seg(struct mthca_arbel_ud_seg *useg, 1604 struct ib_send_wr *wr) 1605{ 1606 memcpy(useg->av, to_mah(wr->wr.ud.ah)->av, MTHCA_AV_SIZE); 1607 useg->dqpn = cpu_to_be32(wr->wr.ud.remote_qpn); 1608 useg->qkey = cpu_to_be32(wr->wr.ud.remote_qkey); 1609} 1610 1611int mthca_tavor_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr, 1612 struct ib_send_wr **bad_wr) 1613{ 1614 struct mthca_dev *dev = to_mdev(ibqp->device); 1615 struct mthca_qp *qp = to_mqp(ibqp); 1616 void *wqe; 1617 void *prev_wqe; 1618 unsigned long flags; 1619 int err = 0; 1620 int nreq; 1621 int i; 1622 int size; 1623 /* 1624 * f0 and size0 are only used if nreq != 0, and they will 1625 * always be initialized the first time through the main loop 1626 * before nreq is incremented. So nreq cannot become non-zero 1627 * without initializing f0 and size0, and they are in fact 1628 * never used uninitialized. 1629 */ 1630 int uninitialized_var(size0); 1631 u32 uninitialized_var(f0); 1632 int ind; 1633 u8 op0 = 0; 1634 1635 spin_lock_irqsave(&qp->sq.lock, flags); 1636 1637 1638 ind = qp->sq.next_ind; 1639 1640 for (nreq = 0; wr; ++nreq, wr = wr->next) { 1641 if (mthca_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) { 1642 mthca_err(dev, "SQ %06x full (%u head, %u tail," 1643 " %d max, %d nreq)\n", qp->qpn, 1644 qp->sq.head, qp->sq.tail, 1645 qp->sq.max, nreq); 1646 err = -ENOMEM; 1647 *bad_wr = wr; 1648 goto out; 1649 } 1650 1651 wqe = get_send_wqe(qp, ind); 1652 prev_wqe = qp->sq.last; 1653 qp->sq.last = wqe; 1654 1655 ((struct mthca_next_seg *) wqe)->nda_op = 0; 1656 ((struct mthca_next_seg *) wqe)->ee_nds = 0; 1657 ((struct mthca_next_seg *) wqe)->flags = 1658 ((wr->send_flags & IB_SEND_SIGNALED) ? 1659 cpu_to_be32(MTHCA_NEXT_CQ_UPDATE) : 0) | 1660 ((wr->send_flags & IB_SEND_SOLICITED) ? 1661 cpu_to_be32(MTHCA_NEXT_SOLICIT) : 0) | 1662 cpu_to_be32(1); 1663 if (wr->opcode == IB_WR_SEND_WITH_IMM || 1664 wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM) 1665 ((struct mthca_next_seg *) wqe)->imm = wr->ex.imm_data; 1666 1667 wqe += sizeof (struct mthca_next_seg); 1668 size = sizeof (struct mthca_next_seg) / 16; 1669 1670 switch (qp->transport) { 1671 case RC: 1672 switch (wr->opcode) { 1673 case IB_WR_ATOMIC_CMP_AND_SWP: 1674 case IB_WR_ATOMIC_FETCH_AND_ADD: 1675 set_raddr_seg(wqe, wr->wr.atomic.remote_addr, 1676 wr->wr.atomic.rkey); 1677 wqe += sizeof (struct mthca_raddr_seg); 1678 1679 set_atomic_seg(wqe, wr); 1680 wqe += sizeof (struct mthca_atomic_seg); 1681 size += (sizeof (struct mthca_raddr_seg) + 1682 sizeof (struct mthca_atomic_seg)) / 16; 1683 break; 1684 1685 case IB_WR_RDMA_WRITE: 1686 case IB_WR_RDMA_WRITE_WITH_IMM: 1687 case IB_WR_RDMA_READ: 1688 set_raddr_seg(wqe, wr->wr.rdma.remote_addr, 1689 wr->wr.rdma.rkey); 1690 wqe += sizeof (struct mthca_raddr_seg); 1691 size += sizeof (struct mthca_raddr_seg) / 16; 1692 break; 1693 1694 default: 1695 /* No extra segments required for sends */ 1696 break; 1697 } 1698 1699 break; 1700 1701 case UC: 1702 switch (wr->opcode) { 1703 case IB_WR_RDMA_WRITE: 1704 case IB_WR_RDMA_WRITE_WITH_IMM: 1705 set_raddr_seg(wqe, wr->wr.rdma.remote_addr, 1706 wr->wr.rdma.rkey); 1707 wqe += sizeof (struct mthca_raddr_seg); 1708 size += sizeof (struct mthca_raddr_seg) / 16; 1709 break; 1710 1711 default: 1712 /* No extra segments required for sends */ 1713 break; 1714 } 1715 1716 break; 1717 1718 case UD: 1719 set_tavor_ud_seg(wqe, wr); 1720 wqe += sizeof (struct mthca_tavor_ud_seg); 1721 size += sizeof (struct mthca_tavor_ud_seg) / 16; 1722 break; 1723 1724 case MLX: 1725 err = build_mlx_header(dev, to_msqp(qp), ind, wr, 1726 wqe - sizeof (struct mthca_next_seg), 1727 wqe); 1728 if (err) { 1729 *bad_wr = wr; 1730 goto out; 1731 } 1732 wqe += sizeof (struct mthca_data_seg); 1733 size += sizeof (struct mthca_data_seg) / 16; 1734 break; 1735 } 1736 1737 if (wr->num_sge > qp->sq.max_gs) { 1738 mthca_err(dev, "too many gathers\n"); 1739 err = -EINVAL; 1740 *bad_wr = wr; 1741 goto out; 1742 } 1743 1744 for (i = 0; i < wr->num_sge; ++i) { 1745 mthca_set_data_seg(wqe, wr->sg_list + i); 1746 wqe += sizeof (struct mthca_data_seg); 1747 size += sizeof (struct mthca_data_seg) / 16; 1748 } 1749 1750 /* Add one more inline data segment for ICRC */ 1751 if (qp->transport == MLX) { 1752 ((struct mthca_data_seg *) wqe)->byte_count = 1753 cpu_to_be32((1 << 31) | 4); 1754 ((u32 *) wqe)[1] = 0; 1755 wqe += sizeof (struct mthca_data_seg); 1756 size += sizeof (struct mthca_data_seg) / 16; 1757 } 1758 1759 qp->wrid[ind + qp->rq.max] = wr->wr_id; 1760 1761 if (wr->opcode >= ARRAY_SIZE(mthca_opcode)) { 1762 mthca_err(dev, "opcode invalid\n"); 1763 err = -EINVAL; 1764 *bad_wr = wr; 1765 goto out; 1766 } 1767 1768 ((struct mthca_next_seg *) prev_wqe)->nda_op = 1769 cpu_to_be32(((ind << qp->sq.wqe_shift) + 1770 qp->send_wqe_offset) | 1771 mthca_opcode[wr->opcode]); 1772 wmb(); 1773 ((struct mthca_next_seg *) prev_wqe)->ee_nds = 1774 cpu_to_be32((nreq ? 0 : MTHCA_NEXT_DBD) | size | 1775 ((wr->send_flags & IB_SEND_FENCE) ? 1776 MTHCA_NEXT_FENCE : 0)); 1777 1778 if (!nreq) { 1779 size0 = size; 1780 op0 = mthca_opcode[wr->opcode]; 1781 f0 = wr->send_flags & IB_SEND_FENCE ? 1782 MTHCA_SEND_DOORBELL_FENCE : 0; 1783 } 1784 1785 ++ind; 1786 if (unlikely(ind >= qp->sq.max)) 1787 ind -= qp->sq.max; 1788 } 1789 1790out: 1791 if (likely(nreq)) { 1792 wmb(); 1793 1794 mthca_write64(((qp->sq.next_ind << qp->sq.wqe_shift) + 1795 qp->send_wqe_offset) | f0 | op0, 1796 (qp->qpn << 8) | size0, 1797 dev->kar + MTHCA_SEND_DOORBELL, 1798 MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock)); 1799 /* 1800 * Make sure doorbells don't leak out of SQ spinlock 1801 * and reach the HCA out of order: 1802 */ 1803 mmiowb(); 1804 } 1805 1806 qp->sq.next_ind = ind; 1807 qp->sq.head += nreq; 1808 1809 spin_unlock_irqrestore(&qp->sq.lock, flags); 1810 return err; 1811} 1812 1813int mthca_tavor_post_receive(struct ib_qp *ibqp, struct ib_recv_wr *wr, 1814 struct ib_recv_wr **bad_wr) 1815{ 1816 struct mthca_dev *dev = to_mdev(ibqp->device); 1817 struct mthca_qp *qp = to_mqp(ibqp); 1818 unsigned long flags; 1819 int err = 0; 1820 int nreq; 1821 int i; 1822 int size; 1823 /* 1824 * size0 is only used if nreq != 0, and it will always be 1825 * initialized the first time through the main loop before 1826 * nreq is incremented. So nreq cannot become non-zero 1827 * without initializing size0, and it is in fact never used 1828 * uninitialized. 1829 */ 1830 int uninitialized_var(size0); 1831 int ind; 1832 void *wqe; 1833 void *prev_wqe; 1834 1835 spin_lock_irqsave(&qp->rq.lock, flags); 1836 1837 1838 ind = qp->rq.next_ind; 1839 1840 for (nreq = 0; wr; wr = wr->next) { 1841 if (mthca_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) { 1842 mthca_err(dev, "RQ %06x full (%u head, %u tail," 1843 " %d max, %d nreq)\n", qp->qpn, 1844 qp->rq.head, qp->rq.tail, 1845 qp->rq.max, nreq); 1846 err = -ENOMEM; 1847 *bad_wr = wr; 1848 goto out; 1849 } 1850 1851 wqe = get_recv_wqe(qp, ind); 1852 prev_wqe = qp->rq.last; 1853 qp->rq.last = wqe; 1854 1855 ((struct mthca_next_seg *) wqe)->ee_nds = 1856 cpu_to_be32(MTHCA_NEXT_DBD); 1857 ((struct mthca_next_seg *) wqe)->flags = 0; 1858 1859 wqe += sizeof (struct mthca_next_seg); 1860 size = sizeof (struct mthca_next_seg) / 16; 1861 1862 if (unlikely(wr->num_sge > qp->rq.max_gs)) { 1863 err = -EINVAL; 1864 *bad_wr = wr; 1865 goto out; 1866 } 1867 1868 for (i = 0; i < wr->num_sge; ++i) { 1869 mthca_set_data_seg(wqe, wr->sg_list + i); 1870 wqe += sizeof (struct mthca_data_seg); 1871 size += sizeof (struct mthca_data_seg) / 16; 1872 } 1873 1874 qp->wrid[ind] = wr->wr_id; 1875 1876 ((struct mthca_next_seg *) prev_wqe)->ee_nds = 1877 cpu_to_be32(MTHCA_NEXT_DBD | size); 1878 1879 if (!nreq) 1880 size0 = size; 1881 1882 ++ind; 1883 if (unlikely(ind >= qp->rq.max)) 1884 ind -= qp->rq.max; 1885 1886 ++nreq; 1887 if (unlikely(nreq == MTHCA_TAVOR_MAX_WQES_PER_RECV_DB)) { 1888 nreq = 0; 1889 1890 wmb(); 1891 1892 mthca_write64((qp->rq.next_ind << qp->rq.wqe_shift) | size0, 1893 qp->qpn << 8, dev->kar + MTHCA_RECEIVE_DOORBELL, 1894 MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock)); 1895 1896 qp->rq.next_ind = ind; 1897 qp->rq.head += MTHCA_TAVOR_MAX_WQES_PER_RECV_DB; 1898 } 1899 } 1900 1901out: 1902 if (likely(nreq)) { 1903 wmb(); 1904 1905 mthca_write64((qp->rq.next_ind << qp->rq.wqe_shift) | size0, 1906 qp->qpn << 8 | nreq, dev->kar + MTHCA_RECEIVE_DOORBELL, 1907 MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock)); 1908 } 1909 1910 qp->rq.next_ind = ind; 1911 qp->rq.head += nreq; 1912 1913 /* 1914 * Make sure doorbells don't leak out of RQ spinlock and reach 1915 * the HCA out of order: 1916 */ 1917 mmiowb(); 1918 1919 spin_unlock_irqrestore(&qp->rq.lock, flags); 1920 return err; 1921} 1922 1923int mthca_arbel_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr, 1924 struct ib_send_wr **bad_wr) 1925{ 1926 struct mthca_dev *dev = to_mdev(ibqp->device); 1927 struct mthca_qp *qp = to_mqp(ibqp); 1928 u32 dbhi; 1929 void *wqe; 1930 void *prev_wqe; 1931 unsigned long flags; 1932 int err = 0; 1933 int nreq; 1934 int i; 1935 int size; 1936 /* 1937 * f0 and size0 are only used if nreq != 0, and they will 1938 * always be initialized the first time through the main loop 1939 * before nreq is incremented. So nreq cannot become non-zero 1940 * without initializing f0 and size0, and they are in fact 1941 * never used uninitialized. 1942 */ 1943 int uninitialized_var(size0); 1944 u32 uninitialized_var(f0); 1945 int ind; 1946 u8 op0 = 0; 1947 1948 spin_lock_irqsave(&qp->sq.lock, flags); 1949 1950 1951 ind = qp->sq.head & (qp->sq.max - 1); 1952 1953 for (nreq = 0; wr; ++nreq, wr = wr->next) { 1954 if (unlikely(nreq == MTHCA_ARBEL_MAX_WQES_PER_SEND_DB)) { 1955 nreq = 0; 1956 1957 dbhi = (MTHCA_ARBEL_MAX_WQES_PER_SEND_DB << 24) | 1958 ((qp->sq.head & 0xffff) << 8) | f0 | op0; 1959 1960 qp->sq.head += MTHCA_ARBEL_MAX_WQES_PER_SEND_DB; 1961 1962 /* 1963 * Make sure that descriptors are written before 1964 * doorbell record. 1965 */ 1966 wmb(); 1967 *qp->sq.db = cpu_to_be32(qp->sq.head & 0xffff); 1968 1969 /* 1970 * Make sure doorbell record is written before we 1971 * write MMIO send doorbell. 1972 */ 1973 wmb(); 1974 1975 mthca_write64(dbhi, (qp->qpn << 8) | size0, 1976 dev->kar + MTHCA_SEND_DOORBELL, 1977 MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock)); 1978 } 1979 1980 if (mthca_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) { 1981 mthca_err(dev, "SQ %06x full (%u head, %u tail," 1982 " %d max, %d nreq)\n", qp->qpn, 1983 qp->sq.head, qp->sq.tail, 1984 qp->sq.max, nreq); 1985 err = -ENOMEM; 1986 *bad_wr = wr; 1987 goto out; 1988 } 1989 1990 wqe = get_send_wqe(qp, ind); 1991 prev_wqe = qp->sq.last; 1992 qp->sq.last = wqe; 1993 1994 ((struct mthca_next_seg *) wqe)->flags = 1995 ((wr->send_flags & IB_SEND_SIGNALED) ? 1996 cpu_to_be32(MTHCA_NEXT_CQ_UPDATE) : 0) | 1997 ((wr->send_flags & IB_SEND_SOLICITED) ? 1998 cpu_to_be32(MTHCA_NEXT_SOLICIT) : 0) | 1999 ((wr->send_flags & IB_SEND_IP_CSUM) ? 2000 cpu_to_be32(MTHCA_NEXT_IP_CSUM | MTHCA_NEXT_TCP_UDP_CSUM) : 0) | 2001 cpu_to_be32(1); 2002 if (wr->opcode == IB_WR_SEND_WITH_IMM || 2003 wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM) 2004 ((struct mthca_next_seg *) wqe)->imm = wr->ex.imm_data; 2005 2006 wqe += sizeof (struct mthca_next_seg); 2007 size = sizeof (struct mthca_next_seg) / 16; 2008 2009 switch (qp->transport) { 2010 case RC: 2011 switch (wr->opcode) { 2012 case IB_WR_ATOMIC_CMP_AND_SWP: 2013 case IB_WR_ATOMIC_FETCH_AND_ADD: 2014 set_raddr_seg(wqe, wr->wr.atomic.remote_addr, 2015 wr->wr.atomic.rkey); 2016 wqe += sizeof (struct mthca_raddr_seg); 2017 2018 set_atomic_seg(wqe, wr); 2019 wqe += sizeof (struct mthca_atomic_seg); 2020 size += (sizeof (struct mthca_raddr_seg) + 2021 sizeof (struct mthca_atomic_seg)) / 16; 2022 break; 2023 2024 case IB_WR_RDMA_READ: 2025 case IB_WR_RDMA_WRITE: 2026 case IB_WR_RDMA_WRITE_WITH_IMM: 2027 set_raddr_seg(wqe, wr->wr.rdma.remote_addr, 2028 wr->wr.rdma.rkey); 2029 wqe += sizeof (struct mthca_raddr_seg); 2030 size += sizeof (struct mthca_raddr_seg) / 16; 2031 break; 2032 2033 default: 2034 /* No extra segments required for sends */ 2035 break; 2036 } 2037 2038 break; 2039 2040 case UC: 2041 switch (wr->opcode) { 2042 case IB_WR_RDMA_WRITE: 2043 case IB_WR_RDMA_WRITE_WITH_IMM: 2044 set_raddr_seg(wqe, wr->wr.rdma.remote_addr, 2045 wr->wr.rdma.rkey); 2046 wqe += sizeof (struct mthca_raddr_seg); 2047 size += sizeof (struct mthca_raddr_seg) / 16; 2048 break; 2049 2050 default: 2051 /* No extra segments required for sends */ 2052 break; 2053 } 2054 2055 break; 2056 2057 case UD: 2058 set_arbel_ud_seg(wqe, wr); 2059 wqe += sizeof (struct mthca_arbel_ud_seg); 2060 size += sizeof (struct mthca_arbel_ud_seg) / 16; 2061 break; 2062 2063 case MLX: 2064 err = build_mlx_header(dev, to_msqp(qp), ind, wr, 2065 wqe - sizeof (struct mthca_next_seg), 2066 wqe); 2067 if (err) { 2068 *bad_wr = wr; 2069 goto out; 2070 } 2071 wqe += sizeof (struct mthca_data_seg); 2072 size += sizeof (struct mthca_data_seg) / 16; 2073 break; 2074 } 2075 2076 if (wr->num_sge > qp->sq.max_gs) { 2077 mthca_err(dev, "too many gathers\n"); 2078 err = -EINVAL; 2079 *bad_wr = wr; 2080 goto out; 2081 } 2082 2083 for (i = 0; i < wr->num_sge; ++i) { 2084 mthca_set_data_seg(wqe, wr->sg_list + i); 2085 wqe += sizeof (struct mthca_data_seg); 2086 size += sizeof (struct mthca_data_seg) / 16; 2087 } 2088 2089 /* Add one more inline data segment for ICRC */ 2090 if (qp->transport == MLX) { 2091 ((struct mthca_data_seg *) wqe)->byte_count = 2092 cpu_to_be32((1 << 31) | 4); 2093 ((u32 *) wqe)[1] = 0; 2094 wqe += sizeof (struct mthca_data_seg); 2095 size += sizeof (struct mthca_data_seg) / 16; 2096 } 2097 2098 qp->wrid[ind + qp->rq.max] = wr->wr_id; 2099 2100 if (wr->opcode >= ARRAY_SIZE(mthca_opcode)) { 2101 mthca_err(dev, "opcode invalid\n"); 2102 err = -EINVAL; 2103 *bad_wr = wr; 2104 goto out; 2105 } 2106 2107 ((struct mthca_next_seg *) prev_wqe)->nda_op = 2108 cpu_to_be32(((ind << qp->sq.wqe_shift) + 2109 qp->send_wqe_offset) | 2110 mthca_opcode[wr->opcode]); 2111 wmb(); 2112 ((struct mthca_next_seg *) prev_wqe)->ee_nds = 2113 cpu_to_be32(MTHCA_NEXT_DBD | size | 2114 ((wr->send_flags & IB_SEND_FENCE) ? 2115 MTHCA_NEXT_FENCE : 0)); 2116 2117 if (!nreq) { 2118 size0 = size; 2119 op0 = mthca_opcode[wr->opcode]; 2120 f0 = wr->send_flags & IB_SEND_FENCE ? 2121 MTHCA_SEND_DOORBELL_FENCE : 0; 2122 } 2123 2124 ++ind; 2125 if (unlikely(ind >= qp->sq.max)) 2126 ind -= qp->sq.max; 2127 } 2128 2129out: 2130 if (likely(nreq)) { 2131 dbhi = (nreq << 24) | ((qp->sq.head & 0xffff) << 8) | f0 | op0; 2132 2133 qp->sq.head += nreq; 2134 2135 /* 2136 * Make sure that descriptors are written before 2137 * doorbell record. 2138 */ 2139 wmb(); 2140 *qp->sq.db = cpu_to_be32(qp->sq.head & 0xffff); 2141 2142 /* 2143 * Make sure doorbell record is written before we 2144 * write MMIO send doorbell. 2145 */ 2146 wmb(); 2147 2148 mthca_write64(dbhi, (qp->qpn << 8) | size0, dev->kar + MTHCA_SEND_DOORBELL, 2149 MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock)); 2150 } 2151 2152 /* 2153 * Make sure doorbells don't leak out of SQ spinlock and reach 2154 * the HCA out of order: 2155 */ 2156 mmiowb(); 2157 2158 spin_unlock_irqrestore(&qp->sq.lock, flags); 2159 return err; 2160} 2161 2162int mthca_arbel_post_receive(struct ib_qp *ibqp, struct ib_recv_wr *wr, 2163 struct ib_recv_wr **bad_wr) 2164{ 2165 struct mthca_dev *dev = to_mdev(ibqp->device); 2166 struct mthca_qp *qp = to_mqp(ibqp); 2167 unsigned long flags; 2168 int err = 0; 2169 int nreq; 2170 int ind; 2171 int i; 2172 void *wqe; 2173 2174 spin_lock_irqsave(&qp->rq.lock, flags); 2175 2176 2177 ind = qp->rq.head & (qp->rq.max - 1); 2178 2179 for (nreq = 0; wr; ++nreq, wr = wr->next) { 2180 if (mthca_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) { 2181 mthca_err(dev, "RQ %06x full (%u head, %u tail," 2182 " %d max, %d nreq)\n", qp->qpn, 2183 qp->rq.head, qp->rq.tail, 2184 qp->rq.max, nreq); 2185 err = -ENOMEM; 2186 *bad_wr = wr; 2187 goto out; 2188 } 2189 2190 wqe = get_recv_wqe(qp, ind); 2191 2192 ((struct mthca_next_seg *) wqe)->flags = 0; 2193 2194 wqe += sizeof (struct mthca_next_seg); 2195 2196 if (unlikely(wr->num_sge > qp->rq.max_gs)) { 2197 err = -EINVAL; 2198 *bad_wr = wr; 2199 goto out; 2200 } 2201 2202 for (i = 0; i < wr->num_sge; ++i) { 2203 mthca_set_data_seg(wqe, wr->sg_list + i); 2204 wqe += sizeof (struct mthca_data_seg); 2205 } 2206 2207 if (i < qp->rq.max_gs) 2208 mthca_set_data_seg_inval(wqe); 2209 2210 qp->wrid[ind] = wr->wr_id; 2211 2212 ++ind; 2213 if (unlikely(ind >= qp->rq.max)) 2214 ind -= qp->rq.max; 2215 } 2216out: 2217 if (likely(nreq)) { 2218 qp->rq.head += nreq; 2219 2220 /* 2221 * Make sure that descriptors are written before 2222 * doorbell record. 2223 */ 2224 wmb(); 2225 *qp->rq.db = cpu_to_be32(qp->rq.head & 0xffff); 2226 } 2227 2228 spin_unlock_irqrestore(&qp->rq.lock, flags); 2229 return err; 2230} 2231 2232void mthca_free_err_wqe(struct mthca_dev *dev, struct mthca_qp *qp, int is_send, 2233 int index, int *dbd, __be32 *new_wqe) 2234{ 2235 struct mthca_next_seg *next; 2236 2237 /* 2238 * For SRQs, all receive WQEs generate a CQE, so we're always 2239 * at the end of the doorbell chain. 2240 */ 2241 if (qp->ibqp.srq && !is_send) { 2242 *new_wqe = 0; 2243 return; 2244 } 2245 2246 if (is_send) 2247 next = get_send_wqe(qp, index); 2248 else 2249 next = get_recv_wqe(qp, index); 2250 2251 *dbd = !!(next->ee_nds & cpu_to_be32(MTHCA_NEXT_DBD)); 2252 if (next->ee_nds & cpu_to_be32(0x3f)) 2253 *new_wqe = (next->nda_op & cpu_to_be32(~0x3f)) | 2254 (next->ee_nds & cpu_to_be32(0x3f)); 2255 else 2256 *new_wqe = 0; 2257} 2258 2259int mthca_init_qp_table(struct mthca_dev *dev) 2260{ 2261 int err; 2262 u8 status; 2263 int i; 2264 2265 spin_lock_init(&dev->qp_table.lock); 2266 2267 /* 2268 * We reserve 2 extra QPs per port for the special QPs. The 2269 * special QP for port 1 has to be even, so round up. 2270 */ 2271 dev->qp_table.sqp_start = (dev->limits.reserved_qps + 1) & ~1UL; 2272 err = mthca_alloc_init(&dev->qp_table.alloc, 2273 dev->limits.num_qps, 2274 (1 << 24) - 1, 2275 dev->qp_table.sqp_start + 2276 MTHCA_MAX_PORTS * 2); 2277 if (err) 2278 return err; 2279 2280 err = mthca_array_init(&dev->qp_table.qp, 2281 dev->limits.num_qps); 2282 if (err) { 2283 mthca_alloc_cleanup(&dev->qp_table.alloc); 2284 return err; 2285 } 2286 2287 for (i = 0; i < 2; ++i) { 2288 err = mthca_CONF_SPECIAL_QP(dev, i ? IB_QPT_GSI : IB_QPT_SMI, 2289 dev->qp_table.sqp_start + i * 2, 2290 &status); 2291 if (err) 2292 goto err_out; 2293 if (status) { 2294 mthca_warn(dev, "CONF_SPECIAL_QP returned " 2295 "status %02x, aborting.\n", 2296 status); 2297 err = -EINVAL; 2298 goto err_out; 2299 } 2300 } 2301 return 0; 2302 2303 err_out: 2304 for (i = 0; i < 2; ++i) 2305 mthca_CONF_SPECIAL_QP(dev, i, 0, &status); 2306 2307 mthca_array_cleanup(&dev->qp_table.qp, dev->limits.num_qps); 2308 mthca_alloc_cleanup(&dev->qp_table.alloc); 2309 2310 return err; 2311} 2312 2313void mthca_cleanup_qp_table(struct mthca_dev *dev) 2314{ 2315 int i; 2316 u8 status; 2317 2318 for (i = 0; i < 2; ++i) 2319 mthca_CONF_SPECIAL_QP(dev, i, 0, &status); 2320 2321 mthca_array_cleanup(&dev->qp_table.qp, dev->limits.num_qps); 2322 mthca_alloc_cleanup(&dev->qp_table.alloc); 2323} 2324