1 /* 2 * Copyright (c) 2006, 2007, 2008, 2009 QLogic Corporation. All rights reserved. 3 * Copyright (c) 2005, 2006 PathScale, Inc. 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/spinlock.h> 35 #include <rdma/ib_smi.h> 36 37 #include "qib.h" 38 #include "qib_mad.h" 39 40 /* 41 * Validate a RWQE and fill in the SGE state. 42 * Return 1 if OK. 43 */ 44 static int qib_init_sge(struct rvt_qp *qp, struct rvt_rwqe *wqe) 45 { 46 int i, j, ret; 47 struct ib_wc wc; 48 struct rvt_lkey_table *rkt; 49 struct rvt_pd *pd; 50 struct rvt_sge_state *ss; 51 52 rkt = &to_idev(qp->ibqp.device)->rdi.lkey_table; 53 pd = ibpd_to_rvtpd(qp->ibqp.srq ? qp->ibqp.srq->pd : qp->ibqp.pd); 54 ss = &qp->r_sge; 55 ss->sg_list = qp->r_sg_list; 56 qp->r_len = 0; 57 for (i = j = 0; i < wqe->num_sge; i++) { 58 if (wqe->sg_list[i].length == 0) 59 continue; 60 /* Check LKEY */ 61 if (!rvt_lkey_ok(rkt, pd, j ? &ss->sg_list[j - 1] : &ss->sge, 62 &wqe->sg_list[i], IB_ACCESS_LOCAL_WRITE)) 63 goto bad_lkey; 64 qp->r_len += wqe->sg_list[i].length; 65 j++; 66 } 67 ss->num_sge = j; 68 ss->total_len = qp->r_len; 69 ret = 1; 70 goto bail; 71 72 bad_lkey: 73 while (j) { 74 struct rvt_sge *sge = --j ? &ss->sg_list[j - 1] : &ss->sge; 75 76 rvt_put_mr(sge->mr); 77 } 78 ss->num_sge = 0; 79 memset(&wc, 0, sizeof(wc)); 80 wc.wr_id = wqe->wr_id; 81 wc.status = IB_WC_LOC_PROT_ERR; 82 wc.opcode = IB_WC_RECV; 83 wc.qp = &qp->ibqp; 84 /* Signal solicited completion event. */ 85 rvt_cq_enter(ibcq_to_rvtcq(qp->ibqp.recv_cq), &wc, 1); 86 ret = 0; 87 bail: 88 return ret; 89 } 90 91 /** 92 * qib_get_rwqe - copy the next RWQE into the QP's RWQE 93 * @qp: the QP 94 * @wr_id_only: update qp->r_wr_id only, not qp->r_sge 95 * 96 * Return -1 if there is a local error, 0 if no RWQE is available, 97 * otherwise return 1. 98 * 99 * Can be called from interrupt level. 100 */ 101 int qib_get_rwqe(struct rvt_qp *qp, int wr_id_only) 102 { 103 unsigned long flags; 104 struct rvt_rq *rq; 105 struct rvt_rwq *wq; 106 struct rvt_srq *srq; 107 struct rvt_rwqe *wqe; 108 void (*handler)(struct ib_event *, void *); 109 u32 tail; 110 int ret; 111 112 if (qp->ibqp.srq) { 113 srq = ibsrq_to_rvtsrq(qp->ibqp.srq); 114 handler = srq->ibsrq.event_handler; 115 rq = &srq->rq; 116 } else { 117 srq = NULL; 118 handler = NULL; 119 rq = &qp->r_rq; 120 } 121 122 spin_lock_irqsave(&rq->lock, flags); 123 if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK)) { 124 ret = 0; 125 goto unlock; 126 } 127 128 wq = rq->wq; 129 tail = wq->tail; 130 /* Validate tail before using it since it is user writable. */ 131 if (tail >= rq->size) 132 tail = 0; 133 if (unlikely(tail == wq->head)) { 134 ret = 0; 135 goto unlock; 136 } 137 /* Make sure entry is read after head index is read. */ 138 smp_rmb(); 139 wqe = rvt_get_rwqe_ptr(rq, tail); 140 /* 141 * Even though we update the tail index in memory, the verbs 142 * consumer is not supposed to post more entries until a 143 * completion is generated. 144 */ 145 if (++tail >= rq->size) 146 tail = 0; 147 wq->tail = tail; 148 if (!wr_id_only && !qib_init_sge(qp, wqe)) { 149 ret = -1; 150 goto unlock; 151 } 152 qp->r_wr_id = wqe->wr_id; 153 154 ret = 1; 155 set_bit(RVT_R_WRID_VALID, &qp->r_aflags); 156 if (handler) { 157 u32 n; 158 159 /* 160 * Validate head pointer value and compute 161 * the number of remaining WQEs. 162 */ 163 n = wq->head; 164 if (n >= rq->size) 165 n = 0; 166 if (n < tail) 167 n += rq->size - tail; 168 else 169 n -= tail; 170 if (n < srq->limit) { 171 struct ib_event ev; 172 173 srq->limit = 0; 174 spin_unlock_irqrestore(&rq->lock, flags); 175 ev.device = qp->ibqp.device; 176 ev.element.srq = qp->ibqp.srq; 177 ev.event = IB_EVENT_SRQ_LIMIT_REACHED; 178 handler(&ev, srq->ibsrq.srq_context); 179 goto bail; 180 } 181 } 182 unlock: 183 spin_unlock_irqrestore(&rq->lock, flags); 184 bail: 185 return ret; 186 } 187 188 /* 189 * Switch to alternate path. 190 * The QP s_lock should be held and interrupts disabled. 191 */ 192 void qib_migrate_qp(struct rvt_qp *qp) 193 { 194 struct ib_event ev; 195 196 qp->s_mig_state = IB_MIG_MIGRATED; 197 qp->remote_ah_attr = qp->alt_ah_attr; 198 qp->port_num = qp->alt_ah_attr.port_num; 199 qp->s_pkey_index = qp->s_alt_pkey_index; 200 201 ev.device = qp->ibqp.device; 202 ev.element.qp = &qp->ibqp; 203 ev.event = IB_EVENT_PATH_MIG; 204 qp->ibqp.event_handler(&ev, qp->ibqp.qp_context); 205 } 206 207 static __be64 get_sguid(struct qib_ibport *ibp, unsigned index) 208 { 209 if (!index) { 210 struct qib_pportdata *ppd = ppd_from_ibp(ibp); 211 212 return ppd->guid; 213 } 214 return ibp->guids[index - 1]; 215 } 216 217 static int gid_ok(union ib_gid *gid, __be64 gid_prefix, __be64 id) 218 { 219 return (gid->global.interface_id == id && 220 (gid->global.subnet_prefix == gid_prefix || 221 gid->global.subnet_prefix == IB_DEFAULT_GID_PREFIX)); 222 } 223 224 /* 225 * 226 * This should be called with the QP r_lock held. 227 * 228 * The s_lock will be acquired around the qib_migrate_qp() call. 229 */ 230 int qib_ruc_check_hdr(struct qib_ibport *ibp, struct ib_header *hdr, 231 int has_grh, struct rvt_qp *qp, u32 bth0) 232 { 233 __be64 guid; 234 unsigned long flags; 235 236 if (qp->s_mig_state == IB_MIG_ARMED && (bth0 & IB_BTH_MIG_REQ)) { 237 if (!has_grh) { 238 if (qp->alt_ah_attr.ah_flags & IB_AH_GRH) 239 goto err; 240 } else { 241 if (!(qp->alt_ah_attr.ah_flags & IB_AH_GRH)) 242 goto err; 243 guid = get_sguid(ibp, qp->alt_ah_attr.grh.sgid_index); 244 if (!gid_ok(&hdr->u.l.grh.dgid, 245 ibp->rvp.gid_prefix, guid)) 246 goto err; 247 if (!gid_ok(&hdr->u.l.grh.sgid, 248 qp->alt_ah_attr.grh.dgid.global.subnet_prefix, 249 qp->alt_ah_attr.grh.dgid.global.interface_id)) 250 goto err; 251 } 252 if (!qib_pkey_ok((u16)bth0, 253 qib_get_pkey(ibp, qp->s_alt_pkey_index))) { 254 qib_bad_pqkey(ibp, IB_NOTICE_TRAP_BAD_PKEY, 255 (u16)bth0, 256 (be16_to_cpu(hdr->lrh[0]) >> 4) & 0xF, 257 0, qp->ibqp.qp_num, 258 hdr->lrh[3], hdr->lrh[1]); 259 goto err; 260 } 261 /* Validate the SLID. See Ch. 9.6.1.5 and 17.2.8 */ 262 if (be16_to_cpu(hdr->lrh[3]) != qp->alt_ah_attr.dlid || 263 ppd_from_ibp(ibp)->port != qp->alt_ah_attr.port_num) 264 goto err; 265 spin_lock_irqsave(&qp->s_lock, flags); 266 qib_migrate_qp(qp); 267 spin_unlock_irqrestore(&qp->s_lock, flags); 268 } else { 269 if (!has_grh) { 270 if (qp->remote_ah_attr.ah_flags & IB_AH_GRH) 271 goto err; 272 } else { 273 if (!(qp->remote_ah_attr.ah_flags & IB_AH_GRH)) 274 goto err; 275 guid = get_sguid(ibp, 276 qp->remote_ah_attr.grh.sgid_index); 277 if (!gid_ok(&hdr->u.l.grh.dgid, 278 ibp->rvp.gid_prefix, guid)) 279 goto err; 280 if (!gid_ok(&hdr->u.l.grh.sgid, 281 qp->remote_ah_attr.grh.dgid.global.subnet_prefix, 282 qp->remote_ah_attr.grh.dgid.global.interface_id)) 283 goto err; 284 } 285 if (!qib_pkey_ok((u16)bth0, 286 qib_get_pkey(ibp, qp->s_pkey_index))) { 287 qib_bad_pqkey(ibp, IB_NOTICE_TRAP_BAD_PKEY, 288 (u16)bth0, 289 (be16_to_cpu(hdr->lrh[0]) >> 4) & 0xF, 290 0, qp->ibqp.qp_num, 291 hdr->lrh[3], hdr->lrh[1]); 292 goto err; 293 } 294 /* Validate the SLID. See Ch. 9.6.1.5 */ 295 if (be16_to_cpu(hdr->lrh[3]) != qp->remote_ah_attr.dlid || 296 ppd_from_ibp(ibp)->port != qp->port_num) 297 goto err; 298 if (qp->s_mig_state == IB_MIG_REARM && 299 !(bth0 & IB_BTH_MIG_REQ)) 300 qp->s_mig_state = IB_MIG_ARMED; 301 } 302 303 return 0; 304 305 err: 306 return 1; 307 } 308 309 /** 310 * qib_ruc_loopback - handle UC and RC lookback requests 311 * @sqp: the sending QP 312 * 313 * This is called from qib_do_send() to 314 * forward a WQE addressed to the same HCA. 315 * Note that although we are single threaded due to the tasklet, we still 316 * have to protect against post_send(). We don't have to worry about 317 * receive interrupts since this is a connected protocol and all packets 318 * will pass through here. 319 */ 320 static void qib_ruc_loopback(struct rvt_qp *sqp) 321 { 322 struct qib_ibport *ibp = to_iport(sqp->ibqp.device, sqp->port_num); 323 struct qib_pportdata *ppd = ppd_from_ibp(ibp); 324 struct qib_devdata *dd = ppd->dd; 325 struct rvt_dev_info *rdi = &dd->verbs_dev.rdi; 326 struct rvt_qp *qp; 327 struct rvt_swqe *wqe; 328 struct rvt_sge *sge; 329 unsigned long flags; 330 struct ib_wc wc; 331 u64 sdata; 332 atomic64_t *maddr; 333 enum ib_wc_status send_status; 334 int release; 335 int ret; 336 337 rcu_read_lock(); 338 /* 339 * Note that we check the responder QP state after 340 * checking the requester's state. 341 */ 342 qp = rvt_lookup_qpn(rdi, &ibp->rvp, sqp->remote_qpn); 343 if (!qp) 344 goto done; 345 346 spin_lock_irqsave(&sqp->s_lock, flags); 347 348 /* Return if we are already busy processing a work request. */ 349 if ((sqp->s_flags & (RVT_S_BUSY | RVT_S_ANY_WAIT)) || 350 !(ib_rvt_state_ops[sqp->state] & RVT_PROCESS_OR_FLUSH_SEND)) 351 goto unlock; 352 353 sqp->s_flags |= RVT_S_BUSY; 354 355 again: 356 smp_read_barrier_depends(); /* see post_one_send() */ 357 if (sqp->s_last == ACCESS_ONCE(sqp->s_head)) 358 goto clr_busy; 359 wqe = rvt_get_swqe_ptr(sqp, sqp->s_last); 360 361 /* Return if it is not OK to start a new work reqeust. */ 362 if (!(ib_rvt_state_ops[sqp->state] & RVT_PROCESS_NEXT_SEND_OK)) { 363 if (!(ib_rvt_state_ops[sqp->state] & RVT_FLUSH_SEND)) 364 goto clr_busy; 365 /* We are in the error state, flush the work request. */ 366 send_status = IB_WC_WR_FLUSH_ERR; 367 goto flush_send; 368 } 369 370 /* 371 * We can rely on the entry not changing without the s_lock 372 * being held until we update s_last. 373 * We increment s_cur to indicate s_last is in progress. 374 */ 375 if (sqp->s_last == sqp->s_cur) { 376 if (++sqp->s_cur >= sqp->s_size) 377 sqp->s_cur = 0; 378 } 379 spin_unlock_irqrestore(&sqp->s_lock, flags); 380 381 if (!qp || !(ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK) || 382 qp->ibqp.qp_type != sqp->ibqp.qp_type) { 383 ibp->rvp.n_pkt_drops++; 384 /* 385 * For RC, the requester would timeout and retry so 386 * shortcut the timeouts and just signal too many retries. 387 */ 388 if (sqp->ibqp.qp_type == IB_QPT_RC) 389 send_status = IB_WC_RETRY_EXC_ERR; 390 else 391 send_status = IB_WC_SUCCESS; 392 goto serr; 393 } 394 395 memset(&wc, 0, sizeof(wc)); 396 send_status = IB_WC_SUCCESS; 397 398 release = 1; 399 sqp->s_sge.sge = wqe->sg_list[0]; 400 sqp->s_sge.sg_list = wqe->sg_list + 1; 401 sqp->s_sge.num_sge = wqe->wr.num_sge; 402 sqp->s_len = wqe->length; 403 switch (wqe->wr.opcode) { 404 case IB_WR_SEND_WITH_IMM: 405 wc.wc_flags = IB_WC_WITH_IMM; 406 wc.ex.imm_data = wqe->wr.ex.imm_data; 407 /* FALLTHROUGH */ 408 case IB_WR_SEND: 409 ret = qib_get_rwqe(qp, 0); 410 if (ret < 0) 411 goto op_err; 412 if (!ret) 413 goto rnr_nak; 414 break; 415 416 case IB_WR_RDMA_WRITE_WITH_IMM: 417 if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_WRITE))) 418 goto inv_err; 419 wc.wc_flags = IB_WC_WITH_IMM; 420 wc.ex.imm_data = wqe->wr.ex.imm_data; 421 ret = qib_get_rwqe(qp, 1); 422 if (ret < 0) 423 goto op_err; 424 if (!ret) 425 goto rnr_nak; 426 /* FALLTHROUGH */ 427 case IB_WR_RDMA_WRITE: 428 if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_WRITE))) 429 goto inv_err; 430 if (wqe->length == 0) 431 break; 432 if (unlikely(!rvt_rkey_ok(qp, &qp->r_sge.sge, wqe->length, 433 wqe->rdma_wr.remote_addr, 434 wqe->rdma_wr.rkey, 435 IB_ACCESS_REMOTE_WRITE))) 436 goto acc_err; 437 qp->r_sge.sg_list = NULL; 438 qp->r_sge.num_sge = 1; 439 qp->r_sge.total_len = wqe->length; 440 break; 441 442 case IB_WR_RDMA_READ: 443 if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_READ))) 444 goto inv_err; 445 if (unlikely(!rvt_rkey_ok(qp, &sqp->s_sge.sge, wqe->length, 446 wqe->rdma_wr.remote_addr, 447 wqe->rdma_wr.rkey, 448 IB_ACCESS_REMOTE_READ))) 449 goto acc_err; 450 release = 0; 451 sqp->s_sge.sg_list = NULL; 452 sqp->s_sge.num_sge = 1; 453 qp->r_sge.sge = wqe->sg_list[0]; 454 qp->r_sge.sg_list = wqe->sg_list + 1; 455 qp->r_sge.num_sge = wqe->wr.num_sge; 456 qp->r_sge.total_len = wqe->length; 457 break; 458 459 case IB_WR_ATOMIC_CMP_AND_SWP: 460 case IB_WR_ATOMIC_FETCH_AND_ADD: 461 if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_ATOMIC))) 462 goto inv_err; 463 if (unlikely(!rvt_rkey_ok(qp, &qp->r_sge.sge, sizeof(u64), 464 wqe->atomic_wr.remote_addr, 465 wqe->atomic_wr.rkey, 466 IB_ACCESS_REMOTE_ATOMIC))) 467 goto acc_err; 468 /* Perform atomic OP and save result. */ 469 maddr = (atomic64_t *) qp->r_sge.sge.vaddr; 470 sdata = wqe->atomic_wr.compare_add; 471 *(u64 *) sqp->s_sge.sge.vaddr = 472 (wqe->atomic_wr.wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD) ? 473 (u64) atomic64_add_return(sdata, maddr) - sdata : 474 (u64) cmpxchg((u64 *) qp->r_sge.sge.vaddr, 475 sdata, wqe->atomic_wr.swap); 476 rvt_put_mr(qp->r_sge.sge.mr); 477 qp->r_sge.num_sge = 0; 478 goto send_comp; 479 480 default: 481 send_status = IB_WC_LOC_QP_OP_ERR; 482 goto serr; 483 } 484 485 sge = &sqp->s_sge.sge; 486 while (sqp->s_len) { 487 u32 len = sqp->s_len; 488 489 if (len > sge->length) 490 len = sge->length; 491 if (len > sge->sge_length) 492 len = sge->sge_length; 493 BUG_ON(len == 0); 494 qib_copy_sge(&qp->r_sge, sge->vaddr, len, release); 495 sge->vaddr += len; 496 sge->length -= len; 497 sge->sge_length -= len; 498 if (sge->sge_length == 0) { 499 if (!release) 500 rvt_put_mr(sge->mr); 501 if (--sqp->s_sge.num_sge) 502 *sge = *sqp->s_sge.sg_list++; 503 } else if (sge->length == 0 && sge->mr->lkey) { 504 if (++sge->n >= RVT_SEGSZ) { 505 if (++sge->m >= sge->mr->mapsz) 506 break; 507 sge->n = 0; 508 } 509 sge->vaddr = 510 sge->mr->map[sge->m]->segs[sge->n].vaddr; 511 sge->length = 512 sge->mr->map[sge->m]->segs[sge->n].length; 513 } 514 sqp->s_len -= len; 515 } 516 if (release) 517 rvt_put_ss(&qp->r_sge); 518 519 if (!test_and_clear_bit(RVT_R_WRID_VALID, &qp->r_aflags)) 520 goto send_comp; 521 522 if (wqe->wr.opcode == IB_WR_RDMA_WRITE_WITH_IMM) 523 wc.opcode = IB_WC_RECV_RDMA_WITH_IMM; 524 else 525 wc.opcode = IB_WC_RECV; 526 wc.wr_id = qp->r_wr_id; 527 wc.status = IB_WC_SUCCESS; 528 wc.byte_len = wqe->length; 529 wc.qp = &qp->ibqp; 530 wc.src_qp = qp->remote_qpn; 531 wc.slid = qp->remote_ah_attr.dlid; 532 wc.sl = qp->remote_ah_attr.sl; 533 wc.port_num = 1; 534 /* Signal completion event if the solicited bit is set. */ 535 rvt_cq_enter(ibcq_to_rvtcq(qp->ibqp.recv_cq), &wc, 536 wqe->wr.send_flags & IB_SEND_SOLICITED); 537 538 send_comp: 539 spin_lock_irqsave(&sqp->s_lock, flags); 540 ibp->rvp.n_loop_pkts++; 541 flush_send: 542 sqp->s_rnr_retry = sqp->s_rnr_retry_cnt; 543 qib_send_complete(sqp, wqe, send_status); 544 goto again; 545 546 rnr_nak: 547 /* Handle RNR NAK */ 548 if (qp->ibqp.qp_type == IB_QPT_UC) 549 goto send_comp; 550 ibp->rvp.n_rnr_naks++; 551 /* 552 * Note: we don't need the s_lock held since the BUSY flag 553 * makes this single threaded. 554 */ 555 if (sqp->s_rnr_retry == 0) { 556 send_status = IB_WC_RNR_RETRY_EXC_ERR; 557 goto serr; 558 } 559 if (sqp->s_rnr_retry_cnt < 7) 560 sqp->s_rnr_retry--; 561 spin_lock_irqsave(&sqp->s_lock, flags); 562 if (!(ib_rvt_state_ops[sqp->state] & RVT_PROCESS_RECV_OK)) 563 goto clr_busy; 564 rvt_add_rnr_timer(sqp, qp->r_min_rnr_timer << 565 IB_AETH_CREDIT_SHIFT); 566 goto clr_busy; 567 568 op_err: 569 send_status = IB_WC_REM_OP_ERR; 570 wc.status = IB_WC_LOC_QP_OP_ERR; 571 goto err; 572 573 inv_err: 574 send_status = IB_WC_REM_INV_REQ_ERR; 575 wc.status = IB_WC_LOC_QP_OP_ERR; 576 goto err; 577 578 acc_err: 579 send_status = IB_WC_REM_ACCESS_ERR; 580 wc.status = IB_WC_LOC_PROT_ERR; 581 err: 582 /* responder goes to error state */ 583 rvt_rc_error(qp, wc.status); 584 585 serr: 586 spin_lock_irqsave(&sqp->s_lock, flags); 587 qib_send_complete(sqp, wqe, send_status); 588 if (sqp->ibqp.qp_type == IB_QPT_RC) { 589 int lastwqe = rvt_error_qp(sqp, IB_WC_WR_FLUSH_ERR); 590 591 sqp->s_flags &= ~RVT_S_BUSY; 592 spin_unlock_irqrestore(&sqp->s_lock, flags); 593 if (lastwqe) { 594 struct ib_event ev; 595 596 ev.device = sqp->ibqp.device; 597 ev.element.qp = &sqp->ibqp; 598 ev.event = IB_EVENT_QP_LAST_WQE_REACHED; 599 sqp->ibqp.event_handler(&ev, sqp->ibqp.qp_context); 600 } 601 goto done; 602 } 603 clr_busy: 604 sqp->s_flags &= ~RVT_S_BUSY; 605 unlock: 606 spin_unlock_irqrestore(&sqp->s_lock, flags); 607 done: 608 rcu_read_unlock(); 609 } 610 611 /** 612 * qib_make_grh - construct a GRH header 613 * @ibp: a pointer to the IB port 614 * @hdr: a pointer to the GRH header being constructed 615 * @grh: the global route address to send to 616 * @hwords: the number of 32 bit words of header being sent 617 * @nwords: the number of 32 bit words of data being sent 618 * 619 * Return the size of the header in 32 bit words. 620 */ 621 u32 qib_make_grh(struct qib_ibport *ibp, struct ib_grh *hdr, 622 struct ib_global_route *grh, u32 hwords, u32 nwords) 623 { 624 hdr->version_tclass_flow = 625 cpu_to_be32((IB_GRH_VERSION << IB_GRH_VERSION_SHIFT) | 626 (grh->traffic_class << IB_GRH_TCLASS_SHIFT) | 627 (grh->flow_label << IB_GRH_FLOW_SHIFT)); 628 hdr->paylen = cpu_to_be16((hwords - 2 + nwords + SIZE_OF_CRC) << 2); 629 /* next_hdr is defined by C8-7 in ch. 8.4.1 */ 630 hdr->next_hdr = IB_GRH_NEXT_HDR; 631 hdr->hop_limit = grh->hop_limit; 632 /* The SGID is 32-bit aligned. */ 633 hdr->sgid.global.subnet_prefix = ibp->rvp.gid_prefix; 634 hdr->sgid.global.interface_id = grh->sgid_index ? 635 ibp->guids[grh->sgid_index - 1] : ppd_from_ibp(ibp)->guid; 636 hdr->dgid = grh->dgid; 637 638 /* GRH header size in 32-bit words. */ 639 return sizeof(struct ib_grh) / sizeof(u32); 640 } 641 642 void qib_make_ruc_header(struct rvt_qp *qp, struct ib_other_headers *ohdr, 643 u32 bth0, u32 bth2) 644 { 645 struct qib_qp_priv *priv = qp->priv; 646 struct qib_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num); 647 u16 lrh0; 648 u32 nwords; 649 u32 extra_bytes; 650 651 /* Construct the header. */ 652 extra_bytes = -qp->s_cur_size & 3; 653 nwords = (qp->s_cur_size + extra_bytes) >> 2; 654 lrh0 = QIB_LRH_BTH; 655 if (unlikely(qp->remote_ah_attr.ah_flags & IB_AH_GRH)) { 656 qp->s_hdrwords += qib_make_grh(ibp, &priv->s_hdr->u.l.grh, 657 &qp->remote_ah_attr.grh, 658 qp->s_hdrwords, nwords); 659 lrh0 = QIB_LRH_GRH; 660 } 661 lrh0 |= ibp->sl_to_vl[qp->remote_ah_attr.sl] << 12 | 662 qp->remote_ah_attr.sl << 4; 663 priv->s_hdr->lrh[0] = cpu_to_be16(lrh0); 664 priv->s_hdr->lrh[1] = cpu_to_be16(qp->remote_ah_attr.dlid); 665 priv->s_hdr->lrh[2] = 666 cpu_to_be16(qp->s_hdrwords + nwords + SIZE_OF_CRC); 667 priv->s_hdr->lrh[3] = cpu_to_be16(ppd_from_ibp(ibp)->lid | 668 qp->remote_ah_attr.src_path_bits); 669 bth0 |= qib_get_pkey(ibp, qp->s_pkey_index); 670 bth0 |= extra_bytes << 20; 671 if (qp->s_mig_state == IB_MIG_MIGRATED) 672 bth0 |= IB_BTH_MIG_REQ; 673 ohdr->bth[0] = cpu_to_be32(bth0); 674 ohdr->bth[1] = cpu_to_be32(qp->remote_qpn); 675 ohdr->bth[2] = cpu_to_be32(bth2); 676 this_cpu_inc(ibp->pmastats->n_unicast_xmit); 677 } 678 679 void _qib_do_send(struct work_struct *work) 680 { 681 struct qib_qp_priv *priv = container_of(work, struct qib_qp_priv, 682 s_work); 683 struct rvt_qp *qp = priv->owner; 684 685 qib_do_send(qp); 686 } 687 688 /** 689 * qib_do_send - perform a send on a QP 690 * @qp: pointer to the QP 691 * 692 * Process entries in the send work queue until credit or queue is 693 * exhausted. Only allow one CPU to send a packet per QP (tasklet). 694 * Otherwise, two threads could send packets out of order. 695 */ 696 void qib_do_send(struct rvt_qp *qp) 697 { 698 struct qib_qp_priv *priv = qp->priv; 699 struct qib_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num); 700 struct qib_pportdata *ppd = ppd_from_ibp(ibp); 701 int (*make_req)(struct rvt_qp *qp, unsigned long *flags); 702 unsigned long flags; 703 704 if ((qp->ibqp.qp_type == IB_QPT_RC || 705 qp->ibqp.qp_type == IB_QPT_UC) && 706 (qp->remote_ah_attr.dlid & ~((1 << ppd->lmc) - 1)) == ppd->lid) { 707 qib_ruc_loopback(qp); 708 return; 709 } 710 711 if (qp->ibqp.qp_type == IB_QPT_RC) 712 make_req = qib_make_rc_req; 713 else if (qp->ibqp.qp_type == IB_QPT_UC) 714 make_req = qib_make_uc_req; 715 else 716 make_req = qib_make_ud_req; 717 718 spin_lock_irqsave(&qp->s_lock, flags); 719 720 /* Return if we are already busy processing a work request. */ 721 if (!qib_send_ok(qp)) { 722 spin_unlock_irqrestore(&qp->s_lock, flags); 723 return; 724 } 725 726 qp->s_flags |= RVT_S_BUSY; 727 728 do { 729 /* Check for a constructed packet to be sent. */ 730 if (qp->s_hdrwords != 0) { 731 spin_unlock_irqrestore(&qp->s_lock, flags); 732 /* 733 * If the packet cannot be sent now, return and 734 * the send tasklet will be woken up later. 735 */ 736 if (qib_verbs_send(qp, priv->s_hdr, qp->s_hdrwords, 737 qp->s_cur_sge, qp->s_cur_size)) 738 return; 739 /* Record that s_hdr is empty. */ 740 qp->s_hdrwords = 0; 741 spin_lock_irqsave(&qp->s_lock, flags); 742 } 743 } while (make_req(qp, &flags)); 744 745 spin_unlock_irqrestore(&qp->s_lock, flags); 746 } 747 748 /* 749 * This should be called with s_lock held. 750 */ 751 void qib_send_complete(struct rvt_qp *qp, struct rvt_swqe *wqe, 752 enum ib_wc_status status) 753 { 754 u32 old_last, last; 755 756 if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_OR_FLUSH_SEND)) 757 return; 758 759 last = qp->s_last; 760 old_last = last; 761 if (++last >= qp->s_size) 762 last = 0; 763 qp->s_last = last; 764 /* See post_send() */ 765 barrier(); 766 rvt_put_swqe(wqe); 767 if (qp->ibqp.qp_type == IB_QPT_UD || 768 qp->ibqp.qp_type == IB_QPT_SMI || 769 qp->ibqp.qp_type == IB_QPT_GSI) 770 atomic_dec(&ibah_to_rvtah(wqe->ud_wr.ah)->refcount); 771 772 rvt_qp_swqe_complete(qp, wqe, status); 773 774 if (qp->s_acked == old_last) 775 qp->s_acked = last; 776 if (qp->s_cur == old_last) 777 qp->s_cur = last; 778 if (qp->s_tail == old_last) 779 qp->s_tail = last; 780 if (qp->state == IB_QPS_SQD && last == qp->s_cur) 781 qp->s_draining = 0; 782 } 783