1 /* 2 * Copyright(c) 2015, 2016 Intel Corporation. 3 * 4 * This file is provided under a dual BSD/GPLv2 license. When using or 5 * redistributing this file, you may do so under either license. 6 * 7 * GPL LICENSE SUMMARY 8 * 9 * This program is free software; you can redistribute it and/or modify 10 * it under the terms of version 2 of the GNU General Public License as 11 * published by the Free Software Foundation. 12 * 13 * This program is distributed in the hope that it will be useful, but 14 * WITHOUT ANY WARRANTY; without even the implied warranty of 15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU 16 * General Public License for more details. 17 * 18 * BSD LICENSE 19 * 20 * Redistribution and use in source and binary forms, with or without 21 * modification, are permitted provided that the following conditions 22 * are met: 23 * 24 * - Redistributions of source code must retain the above copyright 25 * notice, this list of conditions and the following disclaimer. 26 * - Redistributions in binary form must reproduce the above copyright 27 * notice, this list of conditions and the following disclaimer in 28 * the documentation and/or other materials provided with the 29 * distribution. 30 * - Neither the name of Intel Corporation nor the names of its 31 * contributors may be used to endorse or promote products derived 32 * from this software without specific prior written permission. 33 * 34 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 35 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 36 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR 37 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT 38 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, 39 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT 40 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, 41 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY 42 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 43 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE 44 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 45 * 46 */ 47 48 #include <linux/spinlock.h> 49 50 #include "hfi.h" 51 #include "mad.h" 52 #include "qp.h" 53 #include "verbs_txreq.h" 54 #include "trace.h" 55 56 /* 57 * Convert the AETH RNR timeout code into the number of microseconds. 58 */ 59 const u32 ib_hfi1_rnr_table[32] = { 60 655360, /* 00: 655.36 */ 61 10, /* 01: .01 */ 62 20, /* 02 .02 */ 63 30, /* 03: .03 */ 64 40, /* 04: .04 */ 65 60, /* 05: .06 */ 66 80, /* 06: .08 */ 67 120, /* 07: .12 */ 68 160, /* 08: .16 */ 69 240, /* 09: .24 */ 70 320, /* 0A: .32 */ 71 480, /* 0B: .48 */ 72 640, /* 0C: .64 */ 73 960, /* 0D: .96 */ 74 1280, /* 0E: 1.28 */ 75 1920, /* 0F: 1.92 */ 76 2560, /* 10: 2.56 */ 77 3840, /* 11: 3.84 */ 78 5120, /* 12: 5.12 */ 79 7680, /* 13: 7.68 */ 80 10240, /* 14: 10.24 */ 81 15360, /* 15: 15.36 */ 82 20480, /* 16: 20.48 */ 83 30720, /* 17: 30.72 */ 84 40960, /* 18: 40.96 */ 85 61440, /* 19: 61.44 */ 86 81920, /* 1A: 81.92 */ 87 122880, /* 1B: 122.88 */ 88 163840, /* 1C: 163.84 */ 89 245760, /* 1D: 245.76 */ 90 327680, /* 1E: 327.68 */ 91 491520 /* 1F: 491.52 */ 92 }; 93 94 /* 95 * Validate a RWQE and fill in the SGE state. 96 * Return 1 if OK. 97 */ 98 static int init_sge(struct rvt_qp *qp, struct rvt_rwqe *wqe) 99 { 100 int i, j, ret; 101 struct ib_wc wc; 102 struct rvt_lkey_table *rkt; 103 struct rvt_pd *pd; 104 struct rvt_sge_state *ss; 105 106 rkt = &to_idev(qp->ibqp.device)->rdi.lkey_table; 107 pd = ibpd_to_rvtpd(qp->ibqp.srq ? qp->ibqp.srq->pd : qp->ibqp.pd); 108 ss = &qp->r_sge; 109 ss->sg_list = qp->r_sg_list; 110 qp->r_len = 0; 111 for (i = j = 0; i < wqe->num_sge; i++) { 112 if (wqe->sg_list[i].length == 0) 113 continue; 114 /* Check LKEY */ 115 if (!rvt_lkey_ok(rkt, pd, j ? &ss->sg_list[j - 1] : &ss->sge, 116 &wqe->sg_list[i], IB_ACCESS_LOCAL_WRITE)) 117 goto bad_lkey; 118 qp->r_len += wqe->sg_list[i].length; 119 j++; 120 } 121 ss->num_sge = j; 122 ss->total_len = qp->r_len; 123 ret = 1; 124 goto bail; 125 126 bad_lkey: 127 while (j) { 128 struct rvt_sge *sge = --j ? &ss->sg_list[j - 1] : &ss->sge; 129 130 rvt_put_mr(sge->mr); 131 } 132 ss->num_sge = 0; 133 memset(&wc, 0, sizeof(wc)); 134 wc.wr_id = wqe->wr_id; 135 wc.status = IB_WC_LOC_PROT_ERR; 136 wc.opcode = IB_WC_RECV; 137 wc.qp = &qp->ibqp; 138 /* Signal solicited completion event. */ 139 rvt_cq_enter(ibcq_to_rvtcq(qp->ibqp.recv_cq), &wc, 1); 140 ret = 0; 141 bail: 142 return ret; 143 } 144 145 /** 146 * hfi1_rvt_get_rwqe - copy the next RWQE into the QP's RWQE 147 * @qp: the QP 148 * @wr_id_only: update qp->r_wr_id only, not qp->r_sge 149 * 150 * Return -1 if there is a local error, 0 if no RWQE is available, 151 * otherwise return 1. 152 * 153 * Can be called from interrupt level. 154 */ 155 int hfi1_rvt_get_rwqe(struct rvt_qp *qp, int wr_id_only) 156 { 157 unsigned long flags; 158 struct rvt_rq *rq; 159 struct rvt_rwq *wq; 160 struct rvt_srq *srq; 161 struct rvt_rwqe *wqe; 162 void (*handler)(struct ib_event *, void *); 163 u32 tail; 164 int ret; 165 166 if (qp->ibqp.srq) { 167 srq = ibsrq_to_rvtsrq(qp->ibqp.srq); 168 handler = srq->ibsrq.event_handler; 169 rq = &srq->rq; 170 } else { 171 srq = NULL; 172 handler = NULL; 173 rq = &qp->r_rq; 174 } 175 176 spin_lock_irqsave(&rq->lock, flags); 177 if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK)) { 178 ret = 0; 179 goto unlock; 180 } 181 182 wq = rq->wq; 183 tail = wq->tail; 184 /* Validate tail before using it since it is user writable. */ 185 if (tail >= rq->size) 186 tail = 0; 187 if (unlikely(tail == wq->head)) { 188 ret = 0; 189 goto unlock; 190 } 191 /* Make sure entry is read after head index is read. */ 192 smp_rmb(); 193 wqe = rvt_get_rwqe_ptr(rq, tail); 194 /* 195 * Even though we update the tail index in memory, the verbs 196 * consumer is not supposed to post more entries until a 197 * completion is generated. 198 */ 199 if (++tail >= rq->size) 200 tail = 0; 201 wq->tail = tail; 202 if (!wr_id_only && !init_sge(qp, wqe)) { 203 ret = -1; 204 goto unlock; 205 } 206 qp->r_wr_id = wqe->wr_id; 207 208 ret = 1; 209 set_bit(RVT_R_WRID_VALID, &qp->r_aflags); 210 if (handler) { 211 u32 n; 212 213 /* 214 * Validate head pointer value and compute 215 * the number of remaining WQEs. 216 */ 217 n = wq->head; 218 if (n >= rq->size) 219 n = 0; 220 if (n < tail) 221 n += rq->size - tail; 222 else 223 n -= tail; 224 if (n < srq->limit) { 225 struct ib_event ev; 226 227 srq->limit = 0; 228 spin_unlock_irqrestore(&rq->lock, flags); 229 ev.device = qp->ibqp.device; 230 ev.element.srq = qp->ibqp.srq; 231 ev.event = IB_EVENT_SRQ_LIMIT_REACHED; 232 handler(&ev, srq->ibsrq.srq_context); 233 goto bail; 234 } 235 } 236 unlock: 237 spin_unlock_irqrestore(&rq->lock, flags); 238 bail: 239 return ret; 240 } 241 242 static __be64 get_sguid(struct hfi1_ibport *ibp, unsigned index) 243 { 244 if (!index) { 245 struct hfi1_pportdata *ppd = ppd_from_ibp(ibp); 246 247 return cpu_to_be64(ppd->guid); 248 } 249 return ibp->guids[index - 1]; 250 } 251 252 static int gid_ok(union ib_gid *gid, __be64 gid_prefix, __be64 id) 253 { 254 return (gid->global.interface_id == id && 255 (gid->global.subnet_prefix == gid_prefix || 256 gid->global.subnet_prefix == IB_DEFAULT_GID_PREFIX)); 257 } 258 259 /* 260 * 261 * This should be called with the QP r_lock held. 262 * 263 * The s_lock will be acquired around the hfi1_migrate_qp() call. 264 */ 265 int hfi1_ruc_check_hdr(struct hfi1_ibport *ibp, struct hfi1_ib_header *hdr, 266 int has_grh, struct rvt_qp *qp, u32 bth0) 267 { 268 __be64 guid; 269 unsigned long flags; 270 u8 sc5 = ibp->sl_to_sc[qp->remote_ah_attr.sl]; 271 272 if (qp->s_mig_state == IB_MIG_ARMED && (bth0 & IB_BTH_MIG_REQ)) { 273 if (!has_grh) { 274 if (qp->alt_ah_attr.ah_flags & IB_AH_GRH) 275 goto err; 276 } else { 277 if (!(qp->alt_ah_attr.ah_flags & IB_AH_GRH)) 278 goto err; 279 guid = get_sguid(ibp, qp->alt_ah_attr.grh.sgid_index); 280 if (!gid_ok(&hdr->u.l.grh.dgid, ibp->rvp.gid_prefix, 281 guid)) 282 goto err; 283 if (!gid_ok( 284 &hdr->u.l.grh.sgid, 285 qp->alt_ah_attr.grh.dgid.global.subnet_prefix, 286 qp->alt_ah_attr.grh.dgid.global.interface_id)) 287 goto err; 288 } 289 if (unlikely(rcv_pkey_check(ppd_from_ibp(ibp), (u16)bth0, 290 sc5, be16_to_cpu(hdr->lrh[3])))) { 291 hfi1_bad_pqkey(ibp, OPA_TRAP_BAD_P_KEY, 292 (u16)bth0, 293 (be16_to_cpu(hdr->lrh[0]) >> 4) & 0xF, 294 0, qp->ibqp.qp_num, 295 be16_to_cpu(hdr->lrh[3]), 296 be16_to_cpu(hdr->lrh[1])); 297 goto err; 298 } 299 /* Validate the SLID. See Ch. 9.6.1.5 and 17.2.8 */ 300 if (be16_to_cpu(hdr->lrh[3]) != qp->alt_ah_attr.dlid || 301 ppd_from_ibp(ibp)->port != qp->alt_ah_attr.port_num) 302 goto err; 303 spin_lock_irqsave(&qp->s_lock, flags); 304 hfi1_migrate_qp(qp); 305 spin_unlock_irqrestore(&qp->s_lock, flags); 306 } else { 307 if (!has_grh) { 308 if (qp->remote_ah_attr.ah_flags & IB_AH_GRH) 309 goto err; 310 } else { 311 if (!(qp->remote_ah_attr.ah_flags & IB_AH_GRH)) 312 goto err; 313 guid = get_sguid(ibp, 314 qp->remote_ah_attr.grh.sgid_index); 315 if (!gid_ok(&hdr->u.l.grh.dgid, ibp->rvp.gid_prefix, 316 guid)) 317 goto err; 318 if (!gid_ok( 319 &hdr->u.l.grh.sgid, 320 qp->remote_ah_attr.grh.dgid.global.subnet_prefix, 321 qp->remote_ah_attr.grh.dgid.global.interface_id)) 322 goto err; 323 } 324 if (unlikely(rcv_pkey_check(ppd_from_ibp(ibp), (u16)bth0, 325 sc5, be16_to_cpu(hdr->lrh[3])))) { 326 hfi1_bad_pqkey(ibp, OPA_TRAP_BAD_P_KEY, 327 (u16)bth0, 328 (be16_to_cpu(hdr->lrh[0]) >> 4) & 0xF, 329 0, qp->ibqp.qp_num, 330 be16_to_cpu(hdr->lrh[3]), 331 be16_to_cpu(hdr->lrh[1])); 332 goto err; 333 } 334 /* Validate the SLID. See Ch. 9.6.1.5 */ 335 if (be16_to_cpu(hdr->lrh[3]) != qp->remote_ah_attr.dlid || 336 ppd_from_ibp(ibp)->port != qp->port_num) 337 goto err; 338 if (qp->s_mig_state == IB_MIG_REARM && 339 !(bth0 & IB_BTH_MIG_REQ)) 340 qp->s_mig_state = IB_MIG_ARMED; 341 } 342 343 return 0; 344 345 err: 346 return 1; 347 } 348 349 /** 350 * ruc_loopback - handle UC and RC loopback requests 351 * @sqp: the sending QP 352 * 353 * This is called from hfi1_do_send() to 354 * forward a WQE addressed to the same HFI. 355 * Note that although we are single threaded due to the tasklet, we still 356 * have to protect against post_send(). We don't have to worry about 357 * receive interrupts since this is a connected protocol and all packets 358 * will pass through here. 359 */ 360 static void ruc_loopback(struct rvt_qp *sqp) 361 { 362 struct hfi1_ibport *ibp = to_iport(sqp->ibqp.device, sqp->port_num); 363 struct rvt_qp *qp; 364 struct rvt_swqe *wqe; 365 struct rvt_sge *sge; 366 unsigned long flags; 367 struct ib_wc wc; 368 u64 sdata; 369 atomic64_t *maddr; 370 enum ib_wc_status send_status; 371 int release; 372 int ret; 373 int copy_last = 0; 374 u32 to; 375 376 rcu_read_lock(); 377 378 /* 379 * Note that we check the responder QP state after 380 * checking the requester's state. 381 */ 382 qp = rvt_lookup_qpn(ib_to_rvt(sqp->ibqp.device), &ibp->rvp, 383 sqp->remote_qpn); 384 385 spin_lock_irqsave(&sqp->s_lock, flags); 386 387 /* Return if we are already busy processing a work request. */ 388 if ((sqp->s_flags & (RVT_S_BUSY | RVT_S_ANY_WAIT)) || 389 !(ib_rvt_state_ops[sqp->state] & RVT_PROCESS_OR_FLUSH_SEND)) 390 goto unlock; 391 392 sqp->s_flags |= RVT_S_BUSY; 393 394 again: 395 smp_read_barrier_depends(); /* see post_one_send() */ 396 if (sqp->s_last == ACCESS_ONCE(sqp->s_head)) 397 goto clr_busy; 398 wqe = rvt_get_swqe_ptr(sqp, sqp->s_last); 399 400 /* Return if it is not OK to start a new work request. */ 401 if (!(ib_rvt_state_ops[sqp->state] & RVT_PROCESS_NEXT_SEND_OK)) { 402 if (!(ib_rvt_state_ops[sqp->state] & RVT_FLUSH_SEND)) 403 goto clr_busy; 404 /* We are in the error state, flush the work request. */ 405 send_status = IB_WC_WR_FLUSH_ERR; 406 goto flush_send; 407 } 408 409 /* 410 * We can rely on the entry not changing without the s_lock 411 * being held until we update s_last. 412 * We increment s_cur to indicate s_last is in progress. 413 */ 414 if (sqp->s_last == sqp->s_cur) { 415 if (++sqp->s_cur >= sqp->s_size) 416 sqp->s_cur = 0; 417 } 418 spin_unlock_irqrestore(&sqp->s_lock, flags); 419 420 if (!qp || !(ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK) || 421 qp->ibqp.qp_type != sqp->ibqp.qp_type) { 422 ibp->rvp.n_pkt_drops++; 423 /* 424 * For RC, the requester would timeout and retry so 425 * shortcut the timeouts and just signal too many retries. 426 */ 427 if (sqp->ibqp.qp_type == IB_QPT_RC) 428 send_status = IB_WC_RETRY_EXC_ERR; 429 else 430 send_status = IB_WC_SUCCESS; 431 goto serr; 432 } 433 434 memset(&wc, 0, sizeof(wc)); 435 send_status = IB_WC_SUCCESS; 436 437 release = 1; 438 sqp->s_sge.sge = wqe->sg_list[0]; 439 sqp->s_sge.sg_list = wqe->sg_list + 1; 440 sqp->s_sge.num_sge = wqe->wr.num_sge; 441 sqp->s_len = wqe->length; 442 switch (wqe->wr.opcode) { 443 case IB_WR_SEND_WITH_IMM: 444 wc.wc_flags = IB_WC_WITH_IMM; 445 wc.ex.imm_data = wqe->wr.ex.imm_data; 446 /* FALLTHROUGH */ 447 case IB_WR_SEND: 448 ret = hfi1_rvt_get_rwqe(qp, 0); 449 if (ret < 0) 450 goto op_err; 451 if (!ret) 452 goto rnr_nak; 453 break; 454 455 case IB_WR_RDMA_WRITE_WITH_IMM: 456 if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_WRITE))) 457 goto inv_err; 458 wc.wc_flags = IB_WC_WITH_IMM; 459 wc.ex.imm_data = wqe->wr.ex.imm_data; 460 ret = hfi1_rvt_get_rwqe(qp, 1); 461 if (ret < 0) 462 goto op_err; 463 if (!ret) 464 goto rnr_nak; 465 /* skip copy_last set and qp_access_flags recheck */ 466 goto do_write; 467 case IB_WR_RDMA_WRITE: 468 copy_last = ibpd_to_rvtpd(qp->ibqp.pd)->user; 469 if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_WRITE))) 470 goto inv_err; 471 do_write: 472 if (wqe->length == 0) 473 break; 474 if (unlikely(!rvt_rkey_ok(qp, &qp->r_sge.sge, wqe->length, 475 wqe->rdma_wr.remote_addr, 476 wqe->rdma_wr.rkey, 477 IB_ACCESS_REMOTE_WRITE))) 478 goto acc_err; 479 qp->r_sge.sg_list = NULL; 480 qp->r_sge.num_sge = 1; 481 qp->r_sge.total_len = wqe->length; 482 break; 483 484 case IB_WR_RDMA_READ: 485 if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_READ))) 486 goto inv_err; 487 if (unlikely(!rvt_rkey_ok(qp, &sqp->s_sge.sge, wqe->length, 488 wqe->rdma_wr.remote_addr, 489 wqe->rdma_wr.rkey, 490 IB_ACCESS_REMOTE_READ))) 491 goto acc_err; 492 release = 0; 493 sqp->s_sge.sg_list = NULL; 494 sqp->s_sge.num_sge = 1; 495 qp->r_sge.sge = wqe->sg_list[0]; 496 qp->r_sge.sg_list = wqe->sg_list + 1; 497 qp->r_sge.num_sge = wqe->wr.num_sge; 498 qp->r_sge.total_len = wqe->length; 499 break; 500 501 case IB_WR_ATOMIC_CMP_AND_SWP: 502 case IB_WR_ATOMIC_FETCH_AND_ADD: 503 if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_ATOMIC))) 504 goto inv_err; 505 if (unlikely(!rvt_rkey_ok(qp, &qp->r_sge.sge, sizeof(u64), 506 wqe->atomic_wr.remote_addr, 507 wqe->atomic_wr.rkey, 508 IB_ACCESS_REMOTE_ATOMIC))) 509 goto acc_err; 510 /* Perform atomic OP and save result. */ 511 maddr = (atomic64_t *)qp->r_sge.sge.vaddr; 512 sdata = wqe->atomic_wr.compare_add; 513 *(u64 *)sqp->s_sge.sge.vaddr = 514 (wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD) ? 515 (u64)atomic64_add_return(sdata, maddr) - sdata : 516 (u64)cmpxchg((u64 *)qp->r_sge.sge.vaddr, 517 sdata, wqe->atomic_wr.swap); 518 rvt_put_mr(qp->r_sge.sge.mr); 519 qp->r_sge.num_sge = 0; 520 goto send_comp; 521 522 default: 523 send_status = IB_WC_LOC_QP_OP_ERR; 524 goto serr; 525 } 526 527 sge = &sqp->s_sge.sge; 528 while (sqp->s_len) { 529 u32 len = sqp->s_len; 530 531 if (len > sge->length) 532 len = sge->length; 533 if (len > sge->sge_length) 534 len = sge->sge_length; 535 WARN_ON_ONCE(len == 0); 536 hfi1_copy_sge(&qp->r_sge, sge->vaddr, len, release, copy_last); 537 sge->vaddr += len; 538 sge->length -= len; 539 sge->sge_length -= len; 540 if (sge->sge_length == 0) { 541 if (!release) 542 rvt_put_mr(sge->mr); 543 if (--sqp->s_sge.num_sge) 544 *sge = *sqp->s_sge.sg_list++; 545 } else if (sge->length == 0 && sge->mr->lkey) { 546 if (++sge->n >= RVT_SEGSZ) { 547 if (++sge->m >= sge->mr->mapsz) 548 break; 549 sge->n = 0; 550 } 551 sge->vaddr = 552 sge->mr->map[sge->m]->segs[sge->n].vaddr; 553 sge->length = 554 sge->mr->map[sge->m]->segs[sge->n].length; 555 } 556 sqp->s_len -= len; 557 } 558 if (release) 559 rvt_put_ss(&qp->r_sge); 560 561 if (!test_and_clear_bit(RVT_R_WRID_VALID, &qp->r_aflags)) 562 goto send_comp; 563 564 if (wqe->wr.opcode == IB_WR_RDMA_WRITE_WITH_IMM) 565 wc.opcode = IB_WC_RECV_RDMA_WITH_IMM; 566 else 567 wc.opcode = IB_WC_RECV; 568 wc.wr_id = qp->r_wr_id; 569 wc.status = IB_WC_SUCCESS; 570 wc.byte_len = wqe->length; 571 wc.qp = &qp->ibqp; 572 wc.src_qp = qp->remote_qpn; 573 wc.slid = qp->remote_ah_attr.dlid; 574 wc.sl = qp->remote_ah_attr.sl; 575 wc.port_num = 1; 576 /* Signal completion event if the solicited bit is set. */ 577 rvt_cq_enter(ibcq_to_rvtcq(qp->ibqp.recv_cq), &wc, 578 wqe->wr.send_flags & IB_SEND_SOLICITED); 579 580 send_comp: 581 spin_lock_irqsave(&sqp->s_lock, flags); 582 ibp->rvp.n_loop_pkts++; 583 flush_send: 584 sqp->s_rnr_retry = sqp->s_rnr_retry_cnt; 585 hfi1_send_complete(sqp, wqe, send_status); 586 goto again; 587 588 rnr_nak: 589 /* Handle RNR NAK */ 590 if (qp->ibqp.qp_type == IB_QPT_UC) 591 goto send_comp; 592 ibp->rvp.n_rnr_naks++; 593 /* 594 * Note: we don't need the s_lock held since the BUSY flag 595 * makes this single threaded. 596 */ 597 if (sqp->s_rnr_retry == 0) { 598 send_status = IB_WC_RNR_RETRY_EXC_ERR; 599 goto serr; 600 } 601 if (sqp->s_rnr_retry_cnt < 7) 602 sqp->s_rnr_retry--; 603 spin_lock_irqsave(&sqp->s_lock, flags); 604 if (!(ib_rvt_state_ops[sqp->state] & RVT_PROCESS_RECV_OK)) 605 goto clr_busy; 606 to = ib_hfi1_rnr_table[qp->r_min_rnr_timer]; 607 hfi1_add_rnr_timer(sqp, to); 608 goto clr_busy; 609 610 op_err: 611 send_status = IB_WC_REM_OP_ERR; 612 wc.status = IB_WC_LOC_QP_OP_ERR; 613 goto err; 614 615 inv_err: 616 send_status = IB_WC_REM_INV_REQ_ERR; 617 wc.status = IB_WC_LOC_QP_OP_ERR; 618 goto err; 619 620 acc_err: 621 send_status = IB_WC_REM_ACCESS_ERR; 622 wc.status = IB_WC_LOC_PROT_ERR; 623 err: 624 /* responder goes to error state */ 625 hfi1_rc_error(qp, wc.status); 626 627 serr: 628 spin_lock_irqsave(&sqp->s_lock, flags); 629 hfi1_send_complete(sqp, wqe, send_status); 630 if (sqp->ibqp.qp_type == IB_QPT_RC) { 631 int lastwqe = rvt_error_qp(sqp, IB_WC_WR_FLUSH_ERR); 632 633 sqp->s_flags &= ~RVT_S_BUSY; 634 spin_unlock_irqrestore(&sqp->s_lock, flags); 635 if (lastwqe) { 636 struct ib_event ev; 637 638 ev.device = sqp->ibqp.device; 639 ev.element.qp = &sqp->ibqp; 640 ev.event = IB_EVENT_QP_LAST_WQE_REACHED; 641 sqp->ibqp.event_handler(&ev, sqp->ibqp.qp_context); 642 } 643 goto done; 644 } 645 clr_busy: 646 sqp->s_flags &= ~RVT_S_BUSY; 647 unlock: 648 spin_unlock_irqrestore(&sqp->s_lock, flags); 649 done: 650 rcu_read_unlock(); 651 } 652 653 /** 654 * hfi1_make_grh - construct a GRH header 655 * @ibp: a pointer to the IB port 656 * @hdr: a pointer to the GRH header being constructed 657 * @grh: the global route address to send to 658 * @hwords: the number of 32 bit words of header being sent 659 * @nwords: the number of 32 bit words of data being sent 660 * 661 * Return the size of the header in 32 bit words. 662 */ 663 u32 hfi1_make_grh(struct hfi1_ibport *ibp, struct ib_grh *hdr, 664 struct ib_global_route *grh, u32 hwords, u32 nwords) 665 { 666 hdr->version_tclass_flow = 667 cpu_to_be32((IB_GRH_VERSION << IB_GRH_VERSION_SHIFT) | 668 (grh->traffic_class << IB_GRH_TCLASS_SHIFT) | 669 (grh->flow_label << IB_GRH_FLOW_SHIFT)); 670 hdr->paylen = cpu_to_be16((hwords - 2 + nwords + SIZE_OF_CRC) << 2); 671 /* next_hdr is defined by C8-7 in ch. 8.4.1 */ 672 hdr->next_hdr = IB_GRH_NEXT_HDR; 673 hdr->hop_limit = grh->hop_limit; 674 /* The SGID is 32-bit aligned. */ 675 hdr->sgid.global.subnet_prefix = ibp->rvp.gid_prefix; 676 hdr->sgid.global.interface_id = 677 grh->sgid_index && grh->sgid_index < ARRAY_SIZE(ibp->guids) ? 678 ibp->guids[grh->sgid_index - 1] : 679 cpu_to_be64(ppd_from_ibp(ibp)->guid); 680 hdr->dgid = grh->dgid; 681 682 /* GRH header size in 32-bit words. */ 683 return sizeof(struct ib_grh) / sizeof(u32); 684 } 685 686 #define BTH2_OFFSET (offsetof(struct hfi1_pio_header, hdr.u.oth.bth[2]) / 4) 687 688 /** 689 * build_ahg - create ahg in s_hdr 690 * @qp: a pointer to QP 691 * @npsn: the next PSN for the request/response 692 * 693 * This routine handles the AHG by allocating an ahg entry and causing the 694 * copy of the first middle. 695 * 696 * Subsequent middles use the copied entry, editing the 697 * PSN with 1 or 2 edits. 698 */ 699 static inline void build_ahg(struct rvt_qp *qp, u32 npsn) 700 { 701 struct hfi1_qp_priv *priv = qp->priv; 702 703 if (unlikely(qp->s_flags & RVT_S_AHG_CLEAR)) 704 clear_ahg(qp); 705 if (!(qp->s_flags & RVT_S_AHG_VALID)) { 706 /* first middle that needs copy */ 707 if (qp->s_ahgidx < 0) 708 qp->s_ahgidx = sdma_ahg_alloc(priv->s_sde); 709 if (qp->s_ahgidx >= 0) { 710 qp->s_ahgpsn = npsn; 711 priv->s_hdr->tx_flags |= SDMA_TXREQ_F_AHG_COPY; 712 /* save to protect a change in another thread */ 713 priv->s_hdr->sde = priv->s_sde; 714 priv->s_hdr->ahgidx = qp->s_ahgidx; 715 qp->s_flags |= RVT_S_AHG_VALID; 716 } 717 } else { 718 /* subsequent middle after valid */ 719 if (qp->s_ahgidx >= 0) { 720 priv->s_hdr->tx_flags |= SDMA_TXREQ_F_USE_AHG; 721 priv->s_hdr->ahgidx = qp->s_ahgidx; 722 priv->s_hdr->ahgcount++; 723 priv->s_hdr->ahgdesc[0] = 724 sdma_build_ahg_descriptor( 725 (__force u16)cpu_to_be16((u16)npsn), 726 BTH2_OFFSET, 727 16, 728 16); 729 if ((npsn & 0xffff0000) != 730 (qp->s_ahgpsn & 0xffff0000)) { 731 priv->s_hdr->ahgcount++; 732 priv->s_hdr->ahgdesc[1] = 733 sdma_build_ahg_descriptor( 734 (__force u16)cpu_to_be16( 735 (u16)(npsn >> 16)), 736 BTH2_OFFSET, 737 0, 738 16); 739 } 740 } 741 } 742 } 743 744 void hfi1_make_ruc_header(struct rvt_qp *qp, struct hfi1_other_headers *ohdr, 745 u32 bth0, u32 bth2, int middle, 746 struct hfi1_pkt_state *ps) 747 { 748 struct hfi1_qp_priv *priv = qp->priv; 749 struct hfi1_ibport *ibp = ps->ibp; 750 u16 lrh0; 751 u32 nwords; 752 u32 extra_bytes; 753 u32 bth1; 754 755 /* Construct the header. */ 756 extra_bytes = -qp->s_cur_size & 3; 757 nwords = (qp->s_cur_size + extra_bytes) >> 2; 758 lrh0 = HFI1_LRH_BTH; 759 if (unlikely(qp->remote_ah_attr.ah_flags & IB_AH_GRH)) { 760 qp->s_hdrwords += hfi1_make_grh(ibp, 761 &ps->s_txreq->phdr.hdr.u.l.grh, 762 &qp->remote_ah_attr.grh, 763 qp->s_hdrwords, nwords); 764 lrh0 = HFI1_LRH_GRH; 765 middle = 0; 766 } 767 lrh0 |= (priv->s_sc & 0xf) << 12 | (qp->remote_ah_attr.sl & 0xf) << 4; 768 /* 769 * reset s_hdr/AHG fields 770 * 771 * This insures that the ahgentry/ahgcount 772 * are at a non-AHG default to protect 773 * build_verbs_tx_desc() from using 774 * an include ahgidx. 775 * 776 * build_ahg() will modify as appropriate 777 * to use the AHG feature. 778 */ 779 priv->s_hdr->tx_flags = 0; 780 priv->s_hdr->ahgcount = 0; 781 priv->s_hdr->ahgidx = 0; 782 priv->s_hdr->sde = NULL; 783 if (qp->s_mig_state == IB_MIG_MIGRATED) 784 bth0 |= IB_BTH_MIG_REQ; 785 else 786 middle = 0; 787 if (middle) 788 build_ahg(qp, bth2); 789 else 790 qp->s_flags &= ~RVT_S_AHG_VALID; 791 ps->s_txreq->phdr.hdr.lrh[0] = cpu_to_be16(lrh0); 792 ps->s_txreq->phdr.hdr.lrh[1] = cpu_to_be16(qp->remote_ah_attr.dlid); 793 ps->s_txreq->phdr.hdr.lrh[2] = 794 cpu_to_be16(qp->s_hdrwords + nwords + SIZE_OF_CRC); 795 ps->s_txreq->phdr.hdr.lrh[3] = cpu_to_be16(ppd_from_ibp(ibp)->lid | 796 qp->remote_ah_attr.src_path_bits); 797 bth0 |= hfi1_get_pkey(ibp, qp->s_pkey_index); 798 bth0 |= extra_bytes << 20; 799 ohdr->bth[0] = cpu_to_be32(bth0); 800 bth1 = qp->remote_qpn; 801 if (qp->s_flags & RVT_S_ECN) { 802 qp->s_flags &= ~RVT_S_ECN; 803 /* we recently received a FECN, so return a BECN */ 804 bth1 |= (HFI1_BECN_MASK << HFI1_BECN_SHIFT); 805 } 806 ohdr->bth[1] = cpu_to_be32(bth1); 807 ohdr->bth[2] = cpu_to_be32(bth2); 808 } 809 810 /* when sending, force a reschedule every one of these periods */ 811 #define SEND_RESCHED_TIMEOUT (5 * HZ) /* 5s in jiffies */ 812 813 void _hfi1_do_send(struct work_struct *work) 814 { 815 struct iowait *wait = container_of(work, struct iowait, iowork); 816 struct rvt_qp *qp = iowait_to_qp(wait); 817 818 hfi1_do_send(qp); 819 } 820 821 /** 822 * hfi1_do_send - perform a send on a QP 823 * @work: contains a pointer to the QP 824 * 825 * Process entries in the send work queue until credit or queue is 826 * exhausted. Only allow one CPU to send a packet per QP (tasklet). 827 * Otherwise, two threads could send packets out of order. 828 */ 829 void hfi1_do_send(struct rvt_qp *qp) 830 { 831 struct hfi1_pkt_state ps; 832 struct hfi1_qp_priv *priv = qp->priv; 833 int (*make_req)(struct rvt_qp *qp, struct hfi1_pkt_state *ps); 834 unsigned long timeout; 835 unsigned long timeout_int; 836 int cpu; 837 838 ps.dev = to_idev(qp->ibqp.device); 839 ps.ibp = to_iport(qp->ibqp.device, qp->port_num); 840 ps.ppd = ppd_from_ibp(ps.ibp); 841 842 switch (qp->ibqp.qp_type) { 843 case IB_QPT_RC: 844 if (!loopback && ((qp->remote_ah_attr.dlid & ~((1 << ps.ppd->lmc 845 ) - 1)) == 846 ps.ppd->lid)) { 847 ruc_loopback(qp); 848 return; 849 } 850 make_req = hfi1_make_rc_req; 851 timeout_int = (qp->timeout_jiffies); 852 break; 853 case IB_QPT_UC: 854 if (!loopback && ((qp->remote_ah_attr.dlid & ~((1 << ps.ppd->lmc 855 ) - 1)) == 856 ps.ppd->lid)) { 857 ruc_loopback(qp); 858 return; 859 } 860 make_req = hfi1_make_uc_req; 861 timeout_int = SEND_RESCHED_TIMEOUT; 862 break; 863 default: 864 make_req = hfi1_make_ud_req; 865 timeout_int = SEND_RESCHED_TIMEOUT; 866 } 867 868 spin_lock_irqsave(&qp->s_lock, ps.flags); 869 870 /* Return if we are already busy processing a work request. */ 871 if (!hfi1_send_ok(qp)) { 872 spin_unlock_irqrestore(&qp->s_lock, ps.flags); 873 return; 874 } 875 876 qp->s_flags |= RVT_S_BUSY; 877 878 timeout = jiffies + (timeout_int) / 8; 879 cpu = priv->s_sde ? priv->s_sde->cpu : 880 cpumask_first(cpumask_of_node(ps.ppd->dd->node)); 881 /* insure a pre-built packet is handled */ 882 ps.s_txreq = get_waiting_verbs_txreq(qp); 883 do { 884 /* Check for a constructed packet to be sent. */ 885 if (qp->s_hdrwords != 0) { 886 spin_unlock_irqrestore(&qp->s_lock, ps.flags); 887 /* 888 * If the packet cannot be sent now, return and 889 * the send tasklet will be woken up later. 890 */ 891 if (hfi1_verbs_send(qp, &ps)) 892 return; 893 /* Record that s_hdr is empty. */ 894 qp->s_hdrwords = 0; 895 /* allow other tasks to run */ 896 if (unlikely(time_after(jiffies, timeout))) { 897 if (workqueue_congested(cpu, 898 ps.ppd->hfi1_wq)) { 899 spin_lock_irqsave( 900 &qp->s_lock, 901 ps.flags); 902 qp->s_flags &= ~RVT_S_BUSY; 903 hfi1_schedule_send(qp); 904 spin_unlock_irqrestore( 905 &qp->s_lock, 906 ps.flags); 907 this_cpu_inc( 908 *ps.ppd->dd->send_schedule); 909 return; 910 } 911 if (!irqs_disabled()) { 912 cond_resched(); 913 this_cpu_inc( 914 *ps.ppd->dd->send_schedule); 915 } 916 timeout = jiffies + (timeout_int) / 8; 917 } 918 spin_lock_irqsave(&qp->s_lock, ps.flags); 919 } 920 } while (make_req(qp, &ps)); 921 922 spin_unlock_irqrestore(&qp->s_lock, ps.flags); 923 } 924 925 /* 926 * This should be called with s_lock held. 927 */ 928 void hfi1_send_complete(struct rvt_qp *qp, struct rvt_swqe *wqe, 929 enum ib_wc_status status) 930 { 931 u32 old_last, last; 932 unsigned i; 933 934 if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_OR_FLUSH_SEND)) 935 return; 936 937 last = qp->s_last; 938 old_last = last; 939 if (++last >= qp->s_size) 940 last = 0; 941 qp->s_last = last; 942 /* See post_send() */ 943 barrier(); 944 for (i = 0; i < wqe->wr.num_sge; i++) { 945 struct rvt_sge *sge = &wqe->sg_list[i]; 946 947 rvt_put_mr(sge->mr); 948 } 949 if (qp->ibqp.qp_type == IB_QPT_UD || 950 qp->ibqp.qp_type == IB_QPT_SMI || 951 qp->ibqp.qp_type == IB_QPT_GSI) 952 atomic_dec(&ibah_to_rvtah(wqe->ud_wr.ah)->refcount); 953 954 /* See ch. 11.2.4.1 and 10.7.3.1 */ 955 if (!(qp->s_flags & RVT_S_SIGNAL_REQ_WR) || 956 (wqe->wr.send_flags & IB_SEND_SIGNALED) || 957 status != IB_WC_SUCCESS) { 958 struct ib_wc wc; 959 960 memset(&wc, 0, sizeof(wc)); 961 wc.wr_id = wqe->wr.wr_id; 962 wc.status = status; 963 wc.opcode = ib_hfi1_wc_opcode[wqe->wr.opcode]; 964 wc.qp = &qp->ibqp; 965 if (status == IB_WC_SUCCESS) 966 wc.byte_len = wqe->length; 967 rvt_cq_enter(ibcq_to_rvtcq(qp->ibqp.send_cq), &wc, 968 status != IB_WC_SUCCESS); 969 } 970 971 if (qp->s_acked == old_last) 972 qp->s_acked = last; 973 if (qp->s_cur == old_last) 974 qp->s_cur = last; 975 if (qp->s_tail == old_last) 976 qp->s_tail = last; 977 if (qp->state == IB_QPS_SQD && last == qp->s_cur) 978 qp->s_draining = 0; 979 } 980