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/net.h> 49 #include <rdma/ib_smi.h> 50 51 #include "hfi.h" 52 #include "mad.h" 53 #include "verbs_txreq.h" 54 #include "qp.h" 55 56 /** 57 * ud_loopback - handle send on loopback QPs 58 * @sqp: the sending QP 59 * @swqe: the send work request 60 * 61 * This is called from hfi1_make_ud_req() to forward a WQE addressed 62 * to the same HFI. 63 * Note that the receive interrupt handler may be calling hfi1_ud_rcv() 64 * while this is being called. 65 */ 66 static void ud_loopback(struct rvt_qp *sqp, struct rvt_swqe *swqe) 67 { 68 struct hfi1_ibport *ibp = to_iport(sqp->ibqp.device, sqp->port_num); 69 struct hfi1_pportdata *ppd; 70 struct rvt_qp *qp; 71 struct rdma_ah_attr *ah_attr; 72 unsigned long flags; 73 struct rvt_sge_state ssge; 74 struct rvt_sge *sge; 75 struct ib_wc wc; 76 u32 length; 77 enum ib_qp_type sqptype, dqptype; 78 79 rcu_read_lock(); 80 81 qp = rvt_lookup_qpn(ib_to_rvt(sqp->ibqp.device), &ibp->rvp, 82 swqe->ud_wr.remote_qpn); 83 if (!qp) { 84 ibp->rvp.n_pkt_drops++; 85 rcu_read_unlock(); 86 return; 87 } 88 89 sqptype = sqp->ibqp.qp_type == IB_QPT_GSI ? 90 IB_QPT_UD : sqp->ibqp.qp_type; 91 dqptype = qp->ibqp.qp_type == IB_QPT_GSI ? 92 IB_QPT_UD : qp->ibqp.qp_type; 93 94 if (dqptype != sqptype || 95 !(ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK)) { 96 ibp->rvp.n_pkt_drops++; 97 goto drop; 98 } 99 100 ah_attr = &ibah_to_rvtah(swqe->ud_wr.ah)->attr; 101 ppd = ppd_from_ibp(ibp); 102 103 if (qp->ibqp.qp_num > 1) { 104 u16 pkey; 105 u16 slid; 106 u8 sc5 = ibp->sl_to_sc[rdma_ah_get_sl(ah_attr)]; 107 108 pkey = hfi1_get_pkey(ibp, sqp->s_pkey_index); 109 slid = ppd->lid | (rdma_ah_get_path_bits(ah_attr) & 110 ((1 << ppd->lmc) - 1)); 111 if (unlikely(ingress_pkey_check(ppd, pkey, sc5, 112 qp->s_pkey_index, slid))) { 113 hfi1_bad_pqkey(ibp, OPA_TRAP_BAD_P_KEY, pkey, 114 rdma_ah_get_sl(ah_attr), 115 sqp->ibqp.qp_num, qp->ibqp.qp_num, 116 slid, rdma_ah_get_dlid(ah_attr)); 117 goto drop; 118 } 119 } 120 121 /* 122 * Check that the qkey matches (except for QP0, see 9.6.1.4.1). 123 * Qkeys with the high order bit set mean use the 124 * qkey from the QP context instead of the WR (see 10.2.5). 125 */ 126 if (qp->ibqp.qp_num) { 127 u32 qkey; 128 129 qkey = (int)swqe->ud_wr.remote_qkey < 0 ? 130 sqp->qkey : swqe->ud_wr.remote_qkey; 131 if (unlikely(qkey != qp->qkey)) { 132 u16 lid; 133 134 lid = ppd->lid | (rdma_ah_get_path_bits(ah_attr) & 135 ((1 << ppd->lmc) - 1)); 136 hfi1_bad_pqkey(ibp, OPA_TRAP_BAD_Q_KEY, qkey, 137 rdma_ah_get_sl(ah_attr), 138 sqp->ibqp.qp_num, qp->ibqp.qp_num, 139 lid, 140 rdma_ah_get_dlid(ah_attr)); 141 goto drop; 142 } 143 } 144 145 /* 146 * A GRH is expected to precede the data even if not 147 * present on the wire. 148 */ 149 length = swqe->length; 150 memset(&wc, 0, sizeof(wc)); 151 wc.byte_len = length + sizeof(struct ib_grh); 152 153 if (swqe->wr.opcode == IB_WR_SEND_WITH_IMM) { 154 wc.wc_flags = IB_WC_WITH_IMM; 155 wc.ex.imm_data = swqe->wr.ex.imm_data; 156 } 157 158 spin_lock_irqsave(&qp->r_lock, flags); 159 160 /* 161 * Get the next work request entry to find where to put the data. 162 */ 163 if (qp->r_flags & RVT_R_REUSE_SGE) { 164 qp->r_flags &= ~RVT_R_REUSE_SGE; 165 } else { 166 int ret; 167 168 ret = hfi1_rvt_get_rwqe(qp, 0); 169 if (ret < 0) { 170 rvt_rc_error(qp, IB_WC_LOC_QP_OP_ERR); 171 goto bail_unlock; 172 } 173 if (!ret) { 174 if (qp->ibqp.qp_num == 0) 175 ibp->rvp.n_vl15_dropped++; 176 goto bail_unlock; 177 } 178 } 179 /* Silently drop packets which are too big. */ 180 if (unlikely(wc.byte_len > qp->r_len)) { 181 qp->r_flags |= RVT_R_REUSE_SGE; 182 ibp->rvp.n_pkt_drops++; 183 goto bail_unlock; 184 } 185 186 if (rdma_ah_get_ah_flags(ah_attr) & IB_AH_GRH) { 187 struct ib_grh grh; 188 const struct ib_global_route *grd = rdma_ah_read_grh(ah_attr); 189 190 hfi1_make_grh(ibp, &grh, grd, 0, 0); 191 hfi1_copy_sge(&qp->r_sge, &grh, 192 sizeof(grh), true, false); 193 wc.wc_flags |= IB_WC_GRH; 194 } else { 195 rvt_skip_sge(&qp->r_sge, sizeof(struct ib_grh), true); 196 } 197 ssge.sg_list = swqe->sg_list + 1; 198 ssge.sge = *swqe->sg_list; 199 ssge.num_sge = swqe->wr.num_sge; 200 sge = &ssge.sge; 201 while (length) { 202 u32 len = sge->length; 203 204 if (len > length) 205 len = length; 206 if (len > sge->sge_length) 207 len = sge->sge_length; 208 WARN_ON_ONCE(len == 0); 209 hfi1_copy_sge(&qp->r_sge, sge->vaddr, len, true, false); 210 sge->vaddr += len; 211 sge->length -= len; 212 sge->sge_length -= len; 213 if (sge->sge_length == 0) { 214 if (--ssge.num_sge) 215 *sge = *ssge.sg_list++; 216 } else if (sge->length == 0 && sge->mr->lkey) { 217 if (++sge->n >= RVT_SEGSZ) { 218 if (++sge->m >= sge->mr->mapsz) 219 break; 220 sge->n = 0; 221 } 222 sge->vaddr = 223 sge->mr->map[sge->m]->segs[sge->n].vaddr; 224 sge->length = 225 sge->mr->map[sge->m]->segs[sge->n].length; 226 } 227 length -= len; 228 } 229 rvt_put_ss(&qp->r_sge); 230 if (!test_and_clear_bit(RVT_R_WRID_VALID, &qp->r_aflags)) 231 goto bail_unlock; 232 wc.wr_id = qp->r_wr_id; 233 wc.status = IB_WC_SUCCESS; 234 wc.opcode = IB_WC_RECV; 235 wc.qp = &qp->ibqp; 236 wc.src_qp = sqp->ibqp.qp_num; 237 if (qp->ibqp.qp_type == IB_QPT_GSI || qp->ibqp.qp_type == IB_QPT_SMI) { 238 if (sqp->ibqp.qp_type == IB_QPT_GSI || 239 sqp->ibqp.qp_type == IB_QPT_SMI) 240 wc.pkey_index = swqe->ud_wr.pkey_index; 241 else 242 wc.pkey_index = sqp->s_pkey_index; 243 } else { 244 wc.pkey_index = 0; 245 } 246 wc.slid = ppd->lid | (rdma_ah_get_path_bits(ah_attr) & 247 ((1 << ppd->lmc) - 1)); 248 /* Check for loopback when the port lid is not set */ 249 if (wc.slid == 0 && sqp->ibqp.qp_type == IB_QPT_GSI) 250 wc.slid = be16_to_cpu(IB_LID_PERMISSIVE); 251 wc.sl = rdma_ah_get_sl(ah_attr); 252 wc.dlid_path_bits = rdma_ah_get_dlid(ah_attr) & ((1 << ppd->lmc) - 1); 253 wc.port_num = qp->port_num; 254 /* Signal completion event if the solicited bit is set. */ 255 rvt_cq_enter(ibcq_to_rvtcq(qp->ibqp.recv_cq), &wc, 256 swqe->wr.send_flags & IB_SEND_SOLICITED); 257 ibp->rvp.n_loop_pkts++; 258 bail_unlock: 259 spin_unlock_irqrestore(&qp->r_lock, flags); 260 drop: 261 rcu_read_unlock(); 262 } 263 264 /** 265 * hfi1_make_ud_req - construct a UD request packet 266 * @qp: the QP 267 * 268 * Assume s_lock is held. 269 * 270 * Return 1 if constructed; otherwise, return 0. 271 */ 272 int hfi1_make_ud_req(struct rvt_qp *qp, struct hfi1_pkt_state *ps) 273 { 274 struct hfi1_qp_priv *priv = qp->priv; 275 struct ib_other_headers *ohdr; 276 struct rdma_ah_attr *ah_attr; 277 struct hfi1_pportdata *ppd; 278 struct hfi1_ibport *ibp; 279 struct rvt_swqe *wqe; 280 u32 nwords; 281 u32 extra_bytes; 282 u32 bth0; 283 u16 lrh0; 284 u16 lid; 285 int next_cur; 286 u8 sc5; 287 288 ps->s_txreq = get_txreq(ps->dev, qp); 289 if (IS_ERR(ps->s_txreq)) 290 goto bail_no_tx; 291 292 if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_NEXT_SEND_OK)) { 293 if (!(ib_rvt_state_ops[qp->state] & RVT_FLUSH_SEND)) 294 goto bail; 295 /* We are in the error state, flush the work request. */ 296 smp_read_barrier_depends(); /* see post_one_send */ 297 if (qp->s_last == ACCESS_ONCE(qp->s_head)) 298 goto bail; 299 /* If DMAs are in progress, we can't flush immediately. */ 300 if (iowait_sdma_pending(&priv->s_iowait)) { 301 qp->s_flags |= RVT_S_WAIT_DMA; 302 goto bail; 303 } 304 wqe = rvt_get_swqe_ptr(qp, qp->s_last); 305 hfi1_send_complete(qp, wqe, IB_WC_WR_FLUSH_ERR); 306 goto done_free_tx; 307 } 308 309 /* see post_one_send() */ 310 smp_read_barrier_depends(); 311 if (qp->s_cur == ACCESS_ONCE(qp->s_head)) 312 goto bail; 313 314 wqe = rvt_get_swqe_ptr(qp, qp->s_cur); 315 next_cur = qp->s_cur + 1; 316 if (next_cur >= qp->s_size) 317 next_cur = 0; 318 319 /* Construct the header. */ 320 ibp = to_iport(qp->ibqp.device, qp->port_num); 321 ppd = ppd_from_ibp(ibp); 322 ah_attr = &ibah_to_rvtah(wqe->ud_wr.ah)->attr; 323 if (rdma_ah_get_dlid(ah_attr) < be16_to_cpu(IB_MULTICAST_LID_BASE) || 324 rdma_ah_get_dlid(ah_attr) == be16_to_cpu(IB_LID_PERMISSIVE)) { 325 lid = rdma_ah_get_dlid(ah_attr) & ~((1 << ppd->lmc) - 1); 326 if (unlikely(!loopback && 327 (lid == ppd->lid || 328 (lid == be16_to_cpu(IB_LID_PERMISSIVE) && 329 qp->ibqp.qp_type == IB_QPT_GSI)))) { 330 unsigned long tflags = ps->flags; 331 /* 332 * If DMAs are in progress, we can't generate 333 * a completion for the loopback packet since 334 * it would be out of order. 335 * Instead of waiting, we could queue a 336 * zero length descriptor so we get a callback. 337 */ 338 if (iowait_sdma_pending(&priv->s_iowait)) { 339 qp->s_flags |= RVT_S_WAIT_DMA; 340 goto bail; 341 } 342 qp->s_cur = next_cur; 343 spin_unlock_irqrestore(&qp->s_lock, tflags); 344 ud_loopback(qp, wqe); 345 spin_lock_irqsave(&qp->s_lock, tflags); 346 ps->flags = tflags; 347 hfi1_send_complete(qp, wqe, IB_WC_SUCCESS); 348 goto done_free_tx; 349 } 350 } 351 352 qp->s_cur = next_cur; 353 extra_bytes = -wqe->length & 3; 354 nwords = (wqe->length + extra_bytes) >> 2; 355 356 /* header size in 32-bit words LRH+BTH+DETH = (8+12+8)/4. */ 357 qp->s_hdrwords = 7; 358 ps->s_txreq->s_cur_size = wqe->length; 359 ps->s_txreq->ss = &qp->s_sge; 360 qp->s_srate = rdma_ah_get_static_rate(ah_attr); 361 qp->srate_mbps = ib_rate_to_mbps(qp->s_srate); 362 qp->s_wqe = wqe; 363 qp->s_sge.sge = wqe->sg_list[0]; 364 qp->s_sge.sg_list = wqe->sg_list + 1; 365 qp->s_sge.num_sge = wqe->wr.num_sge; 366 qp->s_sge.total_len = wqe->length; 367 368 if (rdma_ah_get_ah_flags(ah_attr) & IB_AH_GRH) { 369 /* Header size in 32-bit words. */ 370 qp->s_hdrwords += hfi1_make_grh(ibp, 371 &ps->s_txreq->phdr.hdr.u.l.grh, 372 rdma_ah_read_grh(ah_attr), 373 qp->s_hdrwords, nwords); 374 lrh0 = HFI1_LRH_GRH; 375 ohdr = &ps->s_txreq->phdr.hdr.u.l.oth; 376 /* 377 * Don't worry about sending to locally attached multicast 378 * QPs. It is unspecified by the spec. what happens. 379 */ 380 } else { 381 /* Header size in 32-bit words. */ 382 lrh0 = HFI1_LRH_BTH; 383 ohdr = &ps->s_txreq->phdr.hdr.u.oth; 384 } 385 if (wqe->wr.opcode == IB_WR_SEND_WITH_IMM) { 386 qp->s_hdrwords++; 387 ohdr->u.ud.imm_data = wqe->wr.ex.imm_data; 388 bth0 = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE << 24; 389 } else { 390 bth0 = IB_OPCODE_UD_SEND_ONLY << 24; 391 } 392 sc5 = ibp->sl_to_sc[rdma_ah_get_sl(ah_attr)]; 393 lrh0 |= (rdma_ah_get_sl(ah_attr) & 0xf) << 4; 394 if (qp->ibqp.qp_type == IB_QPT_SMI) { 395 lrh0 |= 0xF000; /* Set VL (see ch. 13.5.3.1) */ 396 priv->s_sc = 0xf; 397 } else { 398 lrh0 |= (sc5 & 0xf) << 12; 399 priv->s_sc = sc5; 400 } 401 priv->s_sde = qp_to_sdma_engine(qp, priv->s_sc); 402 ps->s_txreq->sde = priv->s_sde; 403 priv->s_sendcontext = qp_to_send_context(qp, priv->s_sc); 404 ps->s_txreq->psc = priv->s_sendcontext; 405 ps->s_txreq->phdr.hdr.lrh[0] = cpu_to_be16(lrh0); 406 ps->s_txreq->phdr.hdr.lrh[1] = 407 cpu_to_be16(rdma_ah_get_dlid(ah_attr)); 408 ps->s_txreq->phdr.hdr.lrh[2] = 409 cpu_to_be16(qp->s_hdrwords + nwords + SIZE_OF_CRC); 410 if (rdma_ah_get_dlid(ah_attr) == be16_to_cpu(IB_LID_PERMISSIVE)) { 411 ps->s_txreq->phdr.hdr.lrh[3] = IB_LID_PERMISSIVE; 412 } else { 413 lid = ppd->lid; 414 if (lid) { 415 lid |= rdma_ah_get_path_bits(ah_attr) & 416 ((1 << ppd->lmc) - 1); 417 ps->s_txreq->phdr.hdr.lrh[3] = cpu_to_be16(lid); 418 } else { 419 ps->s_txreq->phdr.hdr.lrh[3] = IB_LID_PERMISSIVE; 420 } 421 } 422 if (wqe->wr.send_flags & IB_SEND_SOLICITED) 423 bth0 |= IB_BTH_SOLICITED; 424 bth0 |= extra_bytes << 20; 425 if (qp->ibqp.qp_type == IB_QPT_GSI || qp->ibqp.qp_type == IB_QPT_SMI) 426 bth0 |= hfi1_get_pkey(ibp, wqe->ud_wr.pkey_index); 427 else 428 bth0 |= hfi1_get_pkey(ibp, qp->s_pkey_index); 429 ohdr->bth[0] = cpu_to_be32(bth0); 430 ohdr->bth[1] = cpu_to_be32(wqe->ud_wr.remote_qpn); 431 ohdr->bth[2] = cpu_to_be32(mask_psn(wqe->psn)); 432 /* 433 * Qkeys with the high order bit set mean use the 434 * qkey from the QP context instead of the WR (see 10.2.5). 435 */ 436 ohdr->u.ud.deth[0] = cpu_to_be32((int)wqe->ud_wr.remote_qkey < 0 ? 437 qp->qkey : wqe->ud_wr.remote_qkey); 438 ohdr->u.ud.deth[1] = cpu_to_be32(qp->ibqp.qp_num); 439 /* disarm any ahg */ 440 priv->s_ahg->ahgcount = 0; 441 priv->s_ahg->ahgidx = 0; 442 priv->s_ahg->tx_flags = 0; 443 /* pbc */ 444 ps->s_txreq->hdr_dwords = qp->s_hdrwords + 2; 445 446 return 1; 447 448 done_free_tx: 449 hfi1_put_txreq(ps->s_txreq); 450 ps->s_txreq = NULL; 451 return 1; 452 453 bail: 454 hfi1_put_txreq(ps->s_txreq); 455 456 bail_no_tx: 457 ps->s_txreq = NULL; 458 qp->s_flags &= ~RVT_S_BUSY; 459 qp->s_hdrwords = 0; 460 return 0; 461 } 462 463 /* 464 * Hardware can't check this so we do it here. 465 * 466 * This is a slightly different algorithm than the standard pkey check. It 467 * special cases the management keys and allows for 0x7fff and 0xffff to be in 468 * the table at the same time. 469 * 470 * @returns the index found or -1 if not found 471 */ 472 int hfi1_lookup_pkey_idx(struct hfi1_ibport *ibp, u16 pkey) 473 { 474 struct hfi1_pportdata *ppd = ppd_from_ibp(ibp); 475 unsigned i; 476 477 if (pkey == FULL_MGMT_P_KEY || pkey == LIM_MGMT_P_KEY) { 478 unsigned lim_idx = -1; 479 480 for (i = 0; i < ARRAY_SIZE(ppd->pkeys); ++i) { 481 /* here we look for an exact match */ 482 if (ppd->pkeys[i] == pkey) 483 return i; 484 if (ppd->pkeys[i] == LIM_MGMT_P_KEY) 485 lim_idx = i; 486 } 487 488 /* did not find 0xffff return 0x7fff idx if found */ 489 if (pkey == FULL_MGMT_P_KEY) 490 return lim_idx; 491 492 /* no match... */ 493 return -1; 494 } 495 496 pkey &= 0x7fff; /* remove limited/full membership bit */ 497 498 for (i = 0; i < ARRAY_SIZE(ppd->pkeys); ++i) 499 if ((ppd->pkeys[i] & 0x7fff) == pkey) 500 return i; 501 502 /* 503 * Should not get here, this means hardware failed to validate pkeys. 504 */ 505 return -1; 506 } 507 508 void return_cnp(struct hfi1_ibport *ibp, struct rvt_qp *qp, u32 remote_qpn, 509 u32 pkey, u32 slid, u32 dlid, u8 sc5, 510 const struct ib_grh *old_grh) 511 { 512 u64 pbc, pbc_flags = 0; 513 u32 bth0, plen, vl, hwords = 5; 514 u16 lrh0; 515 u8 sl = ibp->sc_to_sl[sc5]; 516 struct ib_header hdr; 517 struct ib_other_headers *ohdr; 518 struct pio_buf *pbuf; 519 struct send_context *ctxt = qp_to_send_context(qp, sc5); 520 struct hfi1_pportdata *ppd = ppd_from_ibp(ibp); 521 522 if (old_grh) { 523 struct ib_grh *grh = &hdr.u.l.grh; 524 525 grh->version_tclass_flow = old_grh->version_tclass_flow; 526 grh->paylen = cpu_to_be16((hwords - 2 + SIZE_OF_CRC) << 2); 527 grh->hop_limit = 0xff; 528 grh->sgid = old_grh->dgid; 529 grh->dgid = old_grh->sgid; 530 ohdr = &hdr.u.l.oth; 531 lrh0 = HFI1_LRH_GRH; 532 hwords += sizeof(struct ib_grh) / sizeof(u32); 533 } else { 534 ohdr = &hdr.u.oth; 535 lrh0 = HFI1_LRH_BTH; 536 } 537 538 lrh0 |= (sc5 & 0xf) << 12 | sl << 4; 539 540 bth0 = pkey | (IB_OPCODE_CNP << 24); 541 ohdr->bth[0] = cpu_to_be32(bth0); 542 543 ohdr->bth[1] = cpu_to_be32(remote_qpn | (1 << IB_BECN_SHIFT)); 544 ohdr->bth[2] = 0; /* PSN 0 */ 545 546 hdr.lrh[0] = cpu_to_be16(lrh0); 547 hdr.lrh[1] = cpu_to_be16(dlid); 548 hdr.lrh[2] = cpu_to_be16(hwords + SIZE_OF_CRC); 549 hdr.lrh[3] = cpu_to_be16(slid); 550 551 plen = 2 /* PBC */ + hwords; 552 pbc_flags |= (!!(sc5 & 0x10)) << PBC_DC_INFO_SHIFT; 553 vl = sc_to_vlt(ppd->dd, sc5); 554 pbc = create_pbc(ppd, pbc_flags, qp->srate_mbps, vl, plen); 555 if (ctxt) { 556 pbuf = sc_buffer_alloc(ctxt, plen, NULL, NULL); 557 if (pbuf) 558 ppd->dd->pio_inline_send(ppd->dd, pbuf, pbc, 559 &hdr, hwords); 560 } 561 } 562 563 /* 564 * opa_smp_check() - Do the regular pkey checking, and the additional 565 * checks for SMPs specified in OPAv1 rev 1.0, 9/19/2016 update, section 566 * 9.10.25 ("SMA Packet Checks"). 567 * 568 * Note that: 569 * - Checks are done using the pkey directly from the packet's BTH, 570 * and specifically _not_ the pkey that we attach to the completion, 571 * which may be different. 572 * - These checks are specifically for "non-local" SMPs (i.e., SMPs 573 * which originated on another node). SMPs which are sent from, and 574 * destined to this node are checked in opa_local_smp_check(). 575 * 576 * At the point where opa_smp_check() is called, we know: 577 * - destination QP is QP0 578 * 579 * opa_smp_check() returns 0 if all checks succeed, 1 otherwise. 580 */ 581 static int opa_smp_check(struct hfi1_ibport *ibp, u16 pkey, u8 sc5, 582 struct rvt_qp *qp, u16 slid, struct opa_smp *smp) 583 { 584 struct hfi1_pportdata *ppd = ppd_from_ibp(ibp); 585 586 /* 587 * I don't think it's possible for us to get here with sc != 0xf, 588 * but check it to be certain. 589 */ 590 if (sc5 != 0xf) 591 return 1; 592 593 if (rcv_pkey_check(ppd, pkey, sc5, slid)) 594 return 1; 595 596 /* 597 * At this point we know (and so don't need to check again) that 598 * the pkey is either LIM_MGMT_P_KEY, or FULL_MGMT_P_KEY 599 * (see ingress_pkey_check). 600 */ 601 if (smp->mgmt_class != IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE && 602 smp->mgmt_class != IB_MGMT_CLASS_SUBN_LID_ROUTED) { 603 ingress_pkey_table_fail(ppd, pkey, slid); 604 return 1; 605 } 606 607 /* 608 * SMPs fall into one of four (disjoint) categories: 609 * SMA request, SMA response, SMA trap, or SMA trap repress. 610 * Our response depends, in part, on which type of SMP we're 611 * processing. 612 * 613 * If this is an SMA response, skip the check here. 614 * 615 * If this is an SMA request or SMA trap repress: 616 * - pkey != FULL_MGMT_P_KEY => 617 * increment port recv constraint errors, drop MAD 618 * 619 * Otherwise: 620 * - accept if the port is running an SM 621 * - drop MAD if it's an SMA trap 622 * - pkey == FULL_MGMT_P_KEY => 623 * reply with unsupported method 624 * - pkey != FULL_MGMT_P_KEY => 625 * increment port recv constraint errors, drop MAD 626 */ 627 switch (smp->method) { 628 case IB_MGMT_METHOD_GET_RESP: 629 case IB_MGMT_METHOD_REPORT_RESP: 630 break; 631 case IB_MGMT_METHOD_GET: 632 case IB_MGMT_METHOD_SET: 633 case IB_MGMT_METHOD_REPORT: 634 case IB_MGMT_METHOD_TRAP_REPRESS: 635 if (pkey != FULL_MGMT_P_KEY) { 636 ingress_pkey_table_fail(ppd, pkey, slid); 637 return 1; 638 } 639 break; 640 default: 641 if (ibp->rvp.port_cap_flags & IB_PORT_SM) 642 return 0; 643 if (smp->method == IB_MGMT_METHOD_TRAP) 644 return 1; 645 if (pkey == FULL_MGMT_P_KEY) { 646 smp->status |= IB_SMP_UNSUP_METHOD; 647 return 0; 648 } 649 ingress_pkey_table_fail(ppd, pkey, slid); 650 return 1; 651 } 652 return 0; 653 } 654 655 /** 656 * hfi1_ud_rcv - receive an incoming UD packet 657 * @ibp: the port the packet came in on 658 * @hdr: the packet header 659 * @rcv_flags: flags relevant to rcv processing 660 * @data: the packet data 661 * @tlen: the packet length 662 * @qp: the QP the packet came on 663 * 664 * This is called from qp_rcv() to process an incoming UD packet 665 * for the given QP. 666 * Called at interrupt level. 667 */ 668 void hfi1_ud_rcv(struct hfi1_packet *packet) 669 { 670 struct ib_other_headers *ohdr = packet->ohdr; 671 int opcode; 672 u32 hdrsize = packet->hlen; 673 struct ib_wc wc; 674 u32 qkey; 675 u32 src_qp; 676 u16 dlid, pkey; 677 int mgmt_pkey_idx = -1; 678 struct hfi1_ibport *ibp = rcd_to_iport(packet->rcd); 679 struct hfi1_pportdata *ppd = ppd_from_ibp(ibp); 680 struct ib_header *hdr = packet->hdr; 681 u32 rcv_flags = packet->rcv_flags; 682 void *data = packet->ebuf; 683 u32 tlen = packet->tlen; 684 struct rvt_qp *qp = packet->qp; 685 bool has_grh = rcv_flags & HFI1_HAS_GRH; 686 u8 sc5 = hfi1_9B_get_sc5(hdr, packet->rhf); 687 u32 bth1; 688 u8 sl_from_sc, sl; 689 u16 slid; 690 u8 extra_bytes; 691 692 qkey = be32_to_cpu(ohdr->u.ud.deth[0]); 693 src_qp = be32_to_cpu(ohdr->u.ud.deth[1]) & RVT_QPN_MASK; 694 dlid = ib_get_dlid(hdr); 695 bth1 = be32_to_cpu(ohdr->bth[1]); 696 slid = ib_get_slid(hdr); 697 pkey = ib_bth_get_pkey(ohdr); 698 opcode = ib_bth_get_opcode(ohdr); 699 sl = ib_get_sl(hdr); 700 extra_bytes = ib_bth_get_pad(ohdr); 701 extra_bytes += (SIZE_OF_CRC << 2); 702 sl_from_sc = ibp->sc_to_sl[sc5]; 703 704 process_ecn(qp, packet, (opcode != IB_OPCODE_CNP)); 705 /* 706 * Get the number of bytes the message was padded by 707 * and drop incomplete packets. 708 */ 709 if (unlikely(tlen < (hdrsize + extra_bytes))) 710 goto drop; 711 712 tlen -= hdrsize + extra_bytes; 713 714 /* 715 * Check that the permissive LID is only used on QP0 716 * and the QKEY matches (see 9.6.1.4.1 and 9.6.1.5.1). 717 */ 718 if (qp->ibqp.qp_num) { 719 if (unlikely(hdr->lrh[1] == IB_LID_PERMISSIVE || 720 hdr->lrh[3] == IB_LID_PERMISSIVE)) 721 goto drop; 722 if (qp->ibqp.qp_num > 1) { 723 if (unlikely(rcv_pkey_check(ppd, pkey, sc5, slid))) { 724 /* 725 * Traps will not be sent for packets dropped 726 * by the HW. This is fine, as sending trap 727 * for invalid pkeys is optional according to 728 * IB spec (release 1.3, section 10.9.4) 729 */ 730 hfi1_bad_pqkey(ibp, OPA_TRAP_BAD_P_KEY, 731 pkey, sl, 732 src_qp, qp->ibqp.qp_num, 733 slid, dlid); 734 return; 735 } 736 } else { 737 /* GSI packet */ 738 mgmt_pkey_idx = hfi1_lookup_pkey_idx(ibp, pkey); 739 if (mgmt_pkey_idx < 0) 740 goto drop; 741 } 742 if (unlikely(qkey != qp->qkey)) { 743 hfi1_bad_pqkey(ibp, OPA_TRAP_BAD_Q_KEY, qkey, sl, 744 src_qp, qp->ibqp.qp_num, 745 slid, dlid); 746 return; 747 } 748 /* Drop invalid MAD packets (see 13.5.3.1). */ 749 if (unlikely(qp->ibqp.qp_num == 1 && 750 (tlen > 2048 || (sc5 == 0xF)))) 751 goto drop; 752 } else { 753 /* Received on QP0, and so by definition, this is an SMP */ 754 struct opa_smp *smp = (struct opa_smp *)data; 755 756 if (opa_smp_check(ibp, pkey, sc5, qp, slid, smp)) 757 goto drop; 758 759 if (tlen > 2048) 760 goto drop; 761 if ((hdr->lrh[1] == IB_LID_PERMISSIVE || 762 hdr->lrh[3] == IB_LID_PERMISSIVE) && 763 smp->mgmt_class != IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) 764 goto drop; 765 766 /* look up SMI pkey */ 767 mgmt_pkey_idx = hfi1_lookup_pkey_idx(ibp, pkey); 768 if (mgmt_pkey_idx < 0) 769 goto drop; 770 } 771 772 if (qp->ibqp.qp_num > 1 && 773 opcode == IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE) { 774 wc.ex.imm_data = ohdr->u.ud.imm_data; 775 wc.wc_flags = IB_WC_WITH_IMM; 776 tlen -= sizeof(u32); 777 } else if (opcode == IB_OPCODE_UD_SEND_ONLY) { 778 wc.ex.imm_data = 0; 779 wc.wc_flags = 0; 780 } else { 781 goto drop; 782 } 783 784 /* 785 * A GRH is expected to precede the data even if not 786 * present on the wire. 787 */ 788 wc.byte_len = tlen + sizeof(struct ib_grh); 789 790 /* 791 * Get the next work request entry to find where to put the data. 792 */ 793 if (qp->r_flags & RVT_R_REUSE_SGE) { 794 qp->r_flags &= ~RVT_R_REUSE_SGE; 795 } else { 796 int ret; 797 798 ret = hfi1_rvt_get_rwqe(qp, 0); 799 if (ret < 0) { 800 rvt_rc_error(qp, IB_WC_LOC_QP_OP_ERR); 801 return; 802 } 803 if (!ret) { 804 if (qp->ibqp.qp_num == 0) 805 ibp->rvp.n_vl15_dropped++; 806 return; 807 } 808 } 809 /* Silently drop packets which are too big. */ 810 if (unlikely(wc.byte_len > qp->r_len)) { 811 qp->r_flags |= RVT_R_REUSE_SGE; 812 goto drop; 813 } 814 if (has_grh) { 815 hfi1_copy_sge(&qp->r_sge, &hdr->u.l.grh, 816 sizeof(struct ib_grh), true, false); 817 wc.wc_flags |= IB_WC_GRH; 818 } else { 819 rvt_skip_sge(&qp->r_sge, sizeof(struct ib_grh), true); 820 } 821 hfi1_copy_sge(&qp->r_sge, data, wc.byte_len - sizeof(struct ib_grh), 822 true, false); 823 rvt_put_ss(&qp->r_sge); 824 if (!test_and_clear_bit(RVT_R_WRID_VALID, &qp->r_aflags)) 825 return; 826 wc.wr_id = qp->r_wr_id; 827 wc.status = IB_WC_SUCCESS; 828 wc.opcode = IB_WC_RECV; 829 wc.vendor_err = 0; 830 wc.qp = &qp->ibqp; 831 wc.src_qp = src_qp; 832 833 if (qp->ibqp.qp_type == IB_QPT_GSI || 834 qp->ibqp.qp_type == IB_QPT_SMI) { 835 if (mgmt_pkey_idx < 0) { 836 if (net_ratelimit()) { 837 struct hfi1_devdata *dd = ppd->dd; 838 839 dd_dev_err(dd, "QP type %d mgmt_pkey_idx < 0 and packet not dropped???\n", 840 qp->ibqp.qp_type); 841 mgmt_pkey_idx = 0; 842 } 843 } 844 wc.pkey_index = (unsigned)mgmt_pkey_idx; 845 } else { 846 wc.pkey_index = 0; 847 } 848 849 wc.slid = slid; 850 wc.sl = sl_from_sc; 851 852 /* 853 * Save the LMC lower bits if the destination LID is a unicast LID. 854 */ 855 wc.dlid_path_bits = dlid >= be16_to_cpu(IB_MULTICAST_LID_BASE) ? 0 : 856 dlid & ((1 << ppd_from_ibp(ibp)->lmc) - 1); 857 wc.port_num = qp->port_num; 858 /* Signal completion event if the solicited bit is set. */ 859 rvt_cq_enter(ibcq_to_rvtcq(qp->ibqp.recv_cq), &wc, 860 (ohdr->bth[0] & 861 cpu_to_be32(IB_BTH_SOLICITED)) != 0); 862 return; 863 864 drop: 865 ibp->rvp.n_pkt_drops++; 866 } 867