1 /* 2 * Copyright(c) 2015 - 2018 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 /* We support only two types - 9B and 16B for now */ 57 static const hfi1_make_req hfi1_make_ud_req_tbl[2] = { 58 [HFI1_PKT_TYPE_9B] = &hfi1_make_ud_req_9B, 59 [HFI1_PKT_TYPE_16B] = &hfi1_make_ud_req_16B 60 }; 61 62 /** 63 * ud_loopback - handle send on loopback QPs 64 * @sqp: the sending QP 65 * @swqe: the send work request 66 * 67 * This is called from hfi1_make_ud_req() to forward a WQE addressed 68 * to the same HFI. 69 * Note that the receive interrupt handler may be calling hfi1_ud_rcv() 70 * while this is being called. 71 */ 72 static void ud_loopback(struct rvt_qp *sqp, struct rvt_swqe *swqe) 73 { 74 struct hfi1_ibport *ibp = to_iport(sqp->ibqp.device, sqp->port_num); 75 struct hfi1_pportdata *ppd; 76 struct hfi1_qp_priv *priv = sqp->priv; 77 struct rvt_qp *qp; 78 struct rdma_ah_attr *ah_attr; 79 unsigned long flags; 80 struct rvt_sge_state ssge; 81 struct rvt_sge *sge; 82 struct ib_wc wc; 83 u32 length; 84 enum ib_qp_type sqptype, dqptype; 85 86 rcu_read_lock(); 87 88 qp = rvt_lookup_qpn(ib_to_rvt(sqp->ibqp.device), &ibp->rvp, 89 swqe->ud_wr.remote_qpn); 90 if (!qp) { 91 ibp->rvp.n_pkt_drops++; 92 rcu_read_unlock(); 93 return; 94 } 95 96 sqptype = sqp->ibqp.qp_type == IB_QPT_GSI ? 97 IB_QPT_UD : sqp->ibqp.qp_type; 98 dqptype = qp->ibqp.qp_type == IB_QPT_GSI ? 99 IB_QPT_UD : qp->ibqp.qp_type; 100 101 if (dqptype != sqptype || 102 !(ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK)) { 103 ibp->rvp.n_pkt_drops++; 104 goto drop; 105 } 106 107 ah_attr = &ibah_to_rvtah(swqe->ud_wr.ah)->attr; 108 ppd = ppd_from_ibp(ibp); 109 110 if (qp->ibqp.qp_num > 1) { 111 u16 pkey; 112 u32 slid; 113 u8 sc5 = ibp->sl_to_sc[rdma_ah_get_sl(ah_attr)]; 114 115 pkey = hfi1_get_pkey(ibp, sqp->s_pkey_index); 116 slid = ppd->lid | (rdma_ah_get_path_bits(ah_attr) & 117 ((1 << ppd->lmc) - 1)); 118 if (unlikely(ingress_pkey_check(ppd, pkey, sc5, 119 qp->s_pkey_index, 120 slid, false))) { 121 hfi1_bad_pkey(ibp, pkey, 122 rdma_ah_get_sl(ah_attr), 123 sqp->ibqp.qp_num, qp->ibqp.qp_num, 124 slid, rdma_ah_get_dlid(ah_attr)); 125 goto drop; 126 } 127 } 128 129 /* 130 * Check that the qkey matches (except for QP0, see 9.6.1.4.1). 131 * Qkeys with the high order bit set mean use the 132 * qkey from the QP context instead of the WR (see 10.2.5). 133 */ 134 if (qp->ibqp.qp_num) { 135 u32 qkey; 136 137 qkey = (int)swqe->ud_wr.remote_qkey < 0 ? 138 sqp->qkey : swqe->ud_wr.remote_qkey; 139 if (unlikely(qkey != qp->qkey)) 140 goto drop; /* silently drop per IBTA spec */ 141 } 142 143 /* 144 * A GRH is expected to precede the data even if not 145 * present on the wire. 146 */ 147 length = swqe->length; 148 memset(&wc, 0, sizeof(wc)); 149 wc.byte_len = length + sizeof(struct ib_grh); 150 151 if (swqe->wr.opcode == IB_WR_SEND_WITH_IMM) { 152 wc.wc_flags = IB_WC_WITH_IMM; 153 wc.ex.imm_data = swqe->wr.ex.imm_data; 154 } 155 156 spin_lock_irqsave(&qp->r_lock, flags); 157 158 /* 159 * Get the next work request entry to find where to put the data. 160 */ 161 if (qp->r_flags & RVT_R_REUSE_SGE) { 162 qp->r_flags &= ~RVT_R_REUSE_SGE; 163 } else { 164 int ret; 165 166 ret = rvt_get_rwqe(qp, false); 167 if (ret < 0) { 168 rvt_rc_error(qp, IB_WC_LOC_QP_OP_ERR); 169 goto bail_unlock; 170 } 171 if (!ret) { 172 if (qp->ibqp.qp_num == 0) 173 ibp->rvp.n_vl15_dropped++; 174 goto bail_unlock; 175 } 176 } 177 /* Silently drop packets which are too big. */ 178 if (unlikely(wc.byte_len > qp->r_len)) { 179 qp->r_flags |= RVT_R_REUSE_SGE; 180 ibp->rvp.n_pkt_drops++; 181 goto bail_unlock; 182 } 183 184 if (rdma_ah_get_ah_flags(ah_attr) & IB_AH_GRH) { 185 struct ib_grh grh; 186 struct ib_global_route grd = *(rdma_ah_read_grh(ah_attr)); 187 188 /* 189 * For loopback packets with extended LIDs, the 190 * sgid_index in the GRH is 0 and the dgid is 191 * OPA GID of the sender. While creating a response 192 * to the loopback packet, IB core creates the new 193 * sgid_index from the DGID and that will be the 194 * OPA_GID_INDEX. The new dgid is from the sgid 195 * index and that will be in the IB GID format. 196 * 197 * We now have a case where the sent packet had a 198 * different sgid_index and dgid compared to the 199 * one that was received in response. 200 * 201 * Fix this inconsistency. 202 */ 203 if (priv->hdr_type == HFI1_PKT_TYPE_16B) { 204 if (grd.sgid_index == 0) 205 grd.sgid_index = OPA_GID_INDEX; 206 207 if (ib_is_opa_gid(&grd.dgid)) 208 grd.dgid.global.interface_id = 209 cpu_to_be64(ppd->guids[HFI1_PORT_GUID_INDEX]); 210 } 211 212 hfi1_make_grh(ibp, &grh, &grd, 0, 0); 213 rvt_copy_sge(qp, &qp->r_sge, &grh, 214 sizeof(grh), true, false); 215 wc.wc_flags |= IB_WC_GRH; 216 } else { 217 rvt_skip_sge(&qp->r_sge, sizeof(struct ib_grh), true); 218 } 219 ssge.sg_list = swqe->sg_list + 1; 220 ssge.sge = *swqe->sg_list; 221 ssge.num_sge = swqe->wr.num_sge; 222 sge = &ssge.sge; 223 while (length) { 224 u32 len = sge->length; 225 226 if (len > length) 227 len = length; 228 if (len > sge->sge_length) 229 len = sge->sge_length; 230 WARN_ON_ONCE(len == 0); 231 rvt_copy_sge(qp, &qp->r_sge, sge->vaddr, len, true, false); 232 sge->vaddr += len; 233 sge->length -= len; 234 sge->sge_length -= len; 235 if (sge->sge_length == 0) { 236 if (--ssge.num_sge) 237 *sge = *ssge.sg_list++; 238 } else if (sge->length == 0 && sge->mr->lkey) { 239 if (++sge->n >= RVT_SEGSZ) { 240 if (++sge->m >= sge->mr->mapsz) 241 break; 242 sge->n = 0; 243 } 244 sge->vaddr = 245 sge->mr->map[sge->m]->segs[sge->n].vaddr; 246 sge->length = 247 sge->mr->map[sge->m]->segs[sge->n].length; 248 } 249 length -= len; 250 } 251 rvt_put_ss(&qp->r_sge); 252 if (!test_and_clear_bit(RVT_R_WRID_VALID, &qp->r_aflags)) 253 goto bail_unlock; 254 wc.wr_id = qp->r_wr_id; 255 wc.status = IB_WC_SUCCESS; 256 wc.opcode = IB_WC_RECV; 257 wc.qp = &qp->ibqp; 258 wc.src_qp = sqp->ibqp.qp_num; 259 if (qp->ibqp.qp_type == IB_QPT_GSI || qp->ibqp.qp_type == IB_QPT_SMI) { 260 if (sqp->ibqp.qp_type == IB_QPT_GSI || 261 sqp->ibqp.qp_type == IB_QPT_SMI) 262 wc.pkey_index = swqe->ud_wr.pkey_index; 263 else 264 wc.pkey_index = sqp->s_pkey_index; 265 } else { 266 wc.pkey_index = 0; 267 } 268 wc.slid = (ppd->lid | (rdma_ah_get_path_bits(ah_attr) & 269 ((1 << ppd->lmc) - 1))) & U16_MAX; 270 /* Check for loopback when the port lid is not set */ 271 if (wc.slid == 0 && sqp->ibqp.qp_type == IB_QPT_GSI) 272 wc.slid = be16_to_cpu(IB_LID_PERMISSIVE); 273 wc.sl = rdma_ah_get_sl(ah_attr); 274 wc.dlid_path_bits = rdma_ah_get_dlid(ah_attr) & ((1 << ppd->lmc) - 1); 275 wc.port_num = qp->port_num; 276 /* Signal completion event if the solicited bit is set. */ 277 rvt_cq_enter(ibcq_to_rvtcq(qp->ibqp.recv_cq), &wc, 278 swqe->wr.send_flags & IB_SEND_SOLICITED); 279 ibp->rvp.n_loop_pkts++; 280 bail_unlock: 281 spin_unlock_irqrestore(&qp->r_lock, flags); 282 drop: 283 rcu_read_unlock(); 284 } 285 286 static void hfi1_make_bth_deth(struct rvt_qp *qp, struct rvt_swqe *wqe, 287 struct ib_other_headers *ohdr, 288 u16 *pkey, u32 extra_bytes, bool bypass) 289 { 290 u32 bth0; 291 struct hfi1_ibport *ibp; 292 293 ibp = to_iport(qp->ibqp.device, qp->port_num); 294 if (wqe->wr.opcode == IB_WR_SEND_WITH_IMM) { 295 ohdr->u.ud.imm_data = wqe->wr.ex.imm_data; 296 bth0 = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE << 24; 297 } else { 298 bth0 = IB_OPCODE_UD_SEND_ONLY << 24; 299 } 300 301 if (wqe->wr.send_flags & IB_SEND_SOLICITED) 302 bth0 |= IB_BTH_SOLICITED; 303 bth0 |= extra_bytes << 20; 304 if (qp->ibqp.qp_type == IB_QPT_GSI || qp->ibqp.qp_type == IB_QPT_SMI) 305 *pkey = hfi1_get_pkey(ibp, wqe->ud_wr.pkey_index); 306 else 307 *pkey = hfi1_get_pkey(ibp, qp->s_pkey_index); 308 if (!bypass) 309 bth0 |= *pkey; 310 ohdr->bth[0] = cpu_to_be32(bth0); 311 ohdr->bth[1] = cpu_to_be32(wqe->ud_wr.remote_qpn); 312 ohdr->bth[2] = cpu_to_be32(mask_psn(wqe->psn)); 313 /* 314 * Qkeys with the high order bit set mean use the 315 * qkey from the QP context instead of the WR (see 10.2.5). 316 */ 317 ohdr->u.ud.deth[0] = cpu_to_be32((int)wqe->ud_wr.remote_qkey < 0 ? 318 qp->qkey : wqe->ud_wr.remote_qkey); 319 ohdr->u.ud.deth[1] = cpu_to_be32(qp->ibqp.qp_num); 320 } 321 322 void hfi1_make_ud_req_9B(struct rvt_qp *qp, struct hfi1_pkt_state *ps, 323 struct rvt_swqe *wqe) 324 { 325 u32 nwords, extra_bytes; 326 u16 len, slid, dlid, pkey; 327 u16 lrh0 = 0; 328 u8 sc5; 329 struct hfi1_qp_priv *priv = qp->priv; 330 struct ib_other_headers *ohdr; 331 struct rdma_ah_attr *ah_attr; 332 struct hfi1_pportdata *ppd; 333 struct hfi1_ibport *ibp; 334 struct ib_grh *grh; 335 336 ibp = to_iport(qp->ibqp.device, qp->port_num); 337 ppd = ppd_from_ibp(ibp); 338 ah_attr = &ibah_to_rvtah(wqe->ud_wr.ah)->attr; 339 340 extra_bytes = -wqe->length & 3; 341 nwords = ((wqe->length + extra_bytes) >> 2) + SIZE_OF_CRC; 342 /* header size in dwords LRH+BTH+DETH = (8+12+8)/4. */ 343 ps->s_txreq->hdr_dwords = 7; 344 if (wqe->wr.opcode == IB_WR_SEND_WITH_IMM) 345 ps->s_txreq->hdr_dwords++; 346 347 if (rdma_ah_get_ah_flags(ah_attr) & IB_AH_GRH) { 348 grh = &ps->s_txreq->phdr.hdr.ibh.u.l.grh; 349 ps->s_txreq->hdr_dwords += 350 hfi1_make_grh(ibp, grh, rdma_ah_read_grh(ah_attr), 351 ps->s_txreq->hdr_dwords - LRH_9B_DWORDS, 352 nwords); 353 lrh0 = HFI1_LRH_GRH; 354 ohdr = &ps->s_txreq->phdr.hdr.ibh.u.l.oth; 355 } else { 356 lrh0 = HFI1_LRH_BTH; 357 ohdr = &ps->s_txreq->phdr.hdr.ibh.u.oth; 358 } 359 360 sc5 = ibp->sl_to_sc[rdma_ah_get_sl(ah_attr)]; 361 lrh0 |= (rdma_ah_get_sl(ah_attr) & 0xf) << 4; 362 if (qp->ibqp.qp_type == IB_QPT_SMI) { 363 lrh0 |= 0xF000; /* Set VL (see ch. 13.5.3.1) */ 364 priv->s_sc = 0xf; 365 } else { 366 lrh0 |= (sc5 & 0xf) << 12; 367 priv->s_sc = sc5; 368 } 369 370 dlid = opa_get_lid(rdma_ah_get_dlid(ah_attr), 9B); 371 if (dlid == be16_to_cpu(IB_LID_PERMISSIVE)) { 372 slid = be16_to_cpu(IB_LID_PERMISSIVE); 373 } else { 374 u16 lid = (u16)ppd->lid; 375 376 if (lid) { 377 lid |= rdma_ah_get_path_bits(ah_attr) & 378 ((1 << ppd->lmc) - 1); 379 slid = lid; 380 } else { 381 slid = be16_to_cpu(IB_LID_PERMISSIVE); 382 } 383 } 384 hfi1_make_bth_deth(qp, wqe, ohdr, &pkey, extra_bytes, false); 385 len = ps->s_txreq->hdr_dwords + nwords; 386 387 /* Setup the packet */ 388 ps->s_txreq->phdr.hdr.hdr_type = HFI1_PKT_TYPE_9B; 389 hfi1_make_ib_hdr(&ps->s_txreq->phdr.hdr.ibh, 390 lrh0, len, dlid, slid); 391 } 392 393 void hfi1_make_ud_req_16B(struct rvt_qp *qp, struct hfi1_pkt_state *ps, 394 struct rvt_swqe *wqe) 395 { 396 struct hfi1_qp_priv *priv = qp->priv; 397 struct ib_other_headers *ohdr; 398 struct rdma_ah_attr *ah_attr; 399 struct hfi1_pportdata *ppd; 400 struct hfi1_ibport *ibp; 401 u32 dlid, slid, nwords, extra_bytes; 402 u32 dest_qp = wqe->ud_wr.remote_qpn; 403 u32 src_qp = qp->ibqp.qp_num; 404 u16 len, pkey; 405 u8 l4, sc5; 406 bool is_mgmt = false; 407 408 ibp = to_iport(qp->ibqp.device, qp->port_num); 409 ppd = ppd_from_ibp(ibp); 410 ah_attr = &ibah_to_rvtah(wqe->ud_wr.ah)->attr; 411 412 /* 413 * Build 16B Management Packet if either the destination 414 * or source queue pair number is 0 or 1. 415 */ 416 if (dest_qp == 0 || src_qp == 0 || dest_qp == 1 || src_qp == 1) { 417 /* header size in dwords 16B LRH+L4_FM = (16+8)/4. */ 418 ps->s_txreq->hdr_dwords = 6; 419 is_mgmt = true; 420 } else { 421 /* header size in dwords 16B LRH+BTH+DETH = (16+12+8)/4. */ 422 ps->s_txreq->hdr_dwords = 9; 423 if (wqe->wr.opcode == IB_WR_SEND_WITH_IMM) 424 ps->s_txreq->hdr_dwords++; 425 } 426 427 /* SW provides space for CRC and LT for bypass packets. */ 428 extra_bytes = hfi1_get_16b_padding((ps->s_txreq->hdr_dwords << 2), 429 wqe->length); 430 nwords = ((wqe->length + extra_bytes + SIZE_OF_LT) >> 2) + SIZE_OF_CRC; 431 432 if ((rdma_ah_get_ah_flags(ah_attr) & IB_AH_GRH) && 433 hfi1_check_mcast(rdma_ah_get_dlid(ah_attr))) { 434 struct ib_grh *grh; 435 struct ib_global_route *grd = rdma_ah_retrieve_grh(ah_attr); 436 /* 437 * Ensure OPA GIDs are transformed to IB gids 438 * before creating the GRH. 439 */ 440 if (grd->sgid_index == OPA_GID_INDEX) { 441 dd_dev_warn(ppd->dd, "Bad sgid_index. sgid_index: %d\n", 442 grd->sgid_index); 443 grd->sgid_index = 0; 444 } 445 grh = &ps->s_txreq->phdr.hdr.opah.u.l.grh; 446 ps->s_txreq->hdr_dwords += hfi1_make_grh( 447 ibp, grh, grd, 448 ps->s_txreq->hdr_dwords - LRH_16B_DWORDS, 449 nwords); 450 ohdr = &ps->s_txreq->phdr.hdr.opah.u.l.oth; 451 l4 = OPA_16B_L4_IB_GLOBAL; 452 } else { 453 ohdr = &ps->s_txreq->phdr.hdr.opah.u.oth; 454 l4 = OPA_16B_L4_IB_LOCAL; 455 } 456 457 sc5 = ibp->sl_to_sc[rdma_ah_get_sl(ah_attr)]; 458 if (qp->ibqp.qp_type == IB_QPT_SMI) 459 priv->s_sc = 0xf; 460 else 461 priv->s_sc = sc5; 462 463 dlid = opa_get_lid(rdma_ah_get_dlid(ah_attr), 16B); 464 if (!ppd->lid) 465 slid = be32_to_cpu(OPA_LID_PERMISSIVE); 466 else 467 slid = ppd->lid | (rdma_ah_get_path_bits(ah_attr) & 468 ((1 << ppd->lmc) - 1)); 469 470 if (is_mgmt) { 471 l4 = OPA_16B_L4_FM; 472 pkey = hfi1_get_pkey(ibp, wqe->ud_wr.pkey_index); 473 hfi1_16B_set_qpn(&ps->s_txreq->phdr.hdr.opah.u.mgmt, 474 dest_qp, src_qp); 475 } else { 476 hfi1_make_bth_deth(qp, wqe, ohdr, &pkey, extra_bytes, true); 477 } 478 /* Convert dwords to flits */ 479 len = (ps->s_txreq->hdr_dwords + nwords) >> 1; 480 481 /* Setup the packet */ 482 ps->s_txreq->phdr.hdr.hdr_type = HFI1_PKT_TYPE_16B; 483 hfi1_make_16b_hdr(&ps->s_txreq->phdr.hdr.opah, 484 slid, dlid, len, pkey, 0, 0, l4, priv->s_sc); 485 } 486 487 /** 488 * hfi1_make_ud_req - construct a UD request packet 489 * @qp: the QP 490 * 491 * Assume s_lock is held. 492 * 493 * Return 1 if constructed; otherwise, return 0. 494 */ 495 int hfi1_make_ud_req(struct rvt_qp *qp, struct hfi1_pkt_state *ps) 496 { 497 struct hfi1_qp_priv *priv = qp->priv; 498 struct rdma_ah_attr *ah_attr; 499 struct hfi1_pportdata *ppd; 500 struct hfi1_ibport *ibp; 501 struct rvt_swqe *wqe; 502 int next_cur; 503 u32 lid; 504 505 ps->s_txreq = get_txreq(ps->dev, qp); 506 if (!ps->s_txreq) 507 goto bail_no_tx; 508 509 if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_NEXT_SEND_OK)) { 510 if (!(ib_rvt_state_ops[qp->state] & RVT_FLUSH_SEND)) 511 goto bail; 512 /* We are in the error state, flush the work request. */ 513 if (qp->s_last == READ_ONCE(qp->s_head)) 514 goto bail; 515 /* If DMAs are in progress, we can't flush immediately. */ 516 if (iowait_sdma_pending(&priv->s_iowait)) { 517 qp->s_flags |= RVT_S_WAIT_DMA; 518 goto bail; 519 } 520 wqe = rvt_get_swqe_ptr(qp, qp->s_last); 521 rvt_send_complete(qp, wqe, IB_WC_WR_FLUSH_ERR); 522 goto done_free_tx; 523 } 524 525 /* see post_one_send() */ 526 if (qp->s_cur == READ_ONCE(qp->s_head)) 527 goto bail; 528 529 wqe = rvt_get_swqe_ptr(qp, qp->s_cur); 530 next_cur = qp->s_cur + 1; 531 if (next_cur >= qp->s_size) 532 next_cur = 0; 533 534 /* Construct the header. */ 535 ibp = to_iport(qp->ibqp.device, qp->port_num); 536 ppd = ppd_from_ibp(ibp); 537 ah_attr = &ibah_to_rvtah(wqe->ud_wr.ah)->attr; 538 priv->hdr_type = hfi1_get_hdr_type(ppd->lid, ah_attr); 539 if ((!hfi1_check_mcast(rdma_ah_get_dlid(ah_attr))) || 540 (rdma_ah_get_dlid(ah_attr) == be32_to_cpu(OPA_LID_PERMISSIVE))) { 541 lid = rdma_ah_get_dlid(ah_attr) & ~((1 << ppd->lmc) - 1); 542 if (unlikely(!loopback && 543 ((lid == ppd->lid) || 544 ((lid == be32_to_cpu(OPA_LID_PERMISSIVE)) && 545 (qp->ibqp.qp_type == IB_QPT_GSI))))) { 546 unsigned long tflags = ps->flags; 547 /* 548 * If DMAs are in progress, we can't generate 549 * a completion for the loopback packet since 550 * it would be out of order. 551 * Instead of waiting, we could queue a 552 * zero length descriptor so we get a callback. 553 */ 554 if (iowait_sdma_pending(&priv->s_iowait)) { 555 qp->s_flags |= RVT_S_WAIT_DMA; 556 goto bail; 557 } 558 qp->s_cur = next_cur; 559 spin_unlock_irqrestore(&qp->s_lock, tflags); 560 ud_loopback(qp, wqe); 561 spin_lock_irqsave(&qp->s_lock, tflags); 562 ps->flags = tflags; 563 rvt_send_complete(qp, wqe, IB_WC_SUCCESS); 564 goto done_free_tx; 565 } 566 } 567 568 qp->s_cur = next_cur; 569 ps->s_txreq->s_cur_size = wqe->length; 570 ps->s_txreq->ss = &qp->s_sge; 571 qp->s_srate = rdma_ah_get_static_rate(ah_attr); 572 qp->srate_mbps = ib_rate_to_mbps(qp->s_srate); 573 qp->s_wqe = wqe; 574 qp->s_sge.sge = wqe->sg_list[0]; 575 qp->s_sge.sg_list = wqe->sg_list + 1; 576 qp->s_sge.num_sge = wqe->wr.num_sge; 577 qp->s_sge.total_len = wqe->length; 578 579 /* Make the appropriate header */ 580 hfi1_make_ud_req_tbl[priv->hdr_type](qp, ps, qp->s_wqe); 581 priv->s_sde = qp_to_sdma_engine(qp, priv->s_sc); 582 ps->s_txreq->sde = priv->s_sde; 583 priv->s_sendcontext = qp_to_send_context(qp, priv->s_sc); 584 ps->s_txreq->psc = priv->s_sendcontext; 585 /* disarm any ahg */ 586 priv->s_ahg->ahgcount = 0; 587 priv->s_ahg->ahgidx = 0; 588 priv->s_ahg->tx_flags = 0; 589 590 return 1; 591 592 done_free_tx: 593 hfi1_put_txreq(ps->s_txreq); 594 ps->s_txreq = NULL; 595 return 1; 596 597 bail: 598 hfi1_put_txreq(ps->s_txreq); 599 600 bail_no_tx: 601 ps->s_txreq = NULL; 602 qp->s_flags &= ~RVT_S_BUSY; 603 return 0; 604 } 605 606 /* 607 * Hardware can't check this so we do it here. 608 * 609 * This is a slightly different algorithm than the standard pkey check. It 610 * special cases the management keys and allows for 0x7fff and 0xffff to be in 611 * the table at the same time. 612 * 613 * @returns the index found or -1 if not found 614 */ 615 int hfi1_lookup_pkey_idx(struct hfi1_ibport *ibp, u16 pkey) 616 { 617 struct hfi1_pportdata *ppd = ppd_from_ibp(ibp); 618 unsigned i; 619 620 if (pkey == FULL_MGMT_P_KEY || pkey == LIM_MGMT_P_KEY) { 621 unsigned lim_idx = -1; 622 623 for (i = 0; i < ARRAY_SIZE(ppd->pkeys); ++i) { 624 /* here we look for an exact match */ 625 if (ppd->pkeys[i] == pkey) 626 return i; 627 if (ppd->pkeys[i] == LIM_MGMT_P_KEY) 628 lim_idx = i; 629 } 630 631 /* did not find 0xffff return 0x7fff idx if found */ 632 if (pkey == FULL_MGMT_P_KEY) 633 return lim_idx; 634 635 /* no match... */ 636 return -1; 637 } 638 639 pkey &= 0x7fff; /* remove limited/full membership bit */ 640 641 for (i = 0; i < ARRAY_SIZE(ppd->pkeys); ++i) 642 if ((ppd->pkeys[i] & 0x7fff) == pkey) 643 return i; 644 645 /* 646 * Should not get here, this means hardware failed to validate pkeys. 647 */ 648 return -1; 649 } 650 651 void return_cnp_16B(struct hfi1_ibport *ibp, struct rvt_qp *qp, 652 u32 remote_qpn, u16 pkey, u32 slid, u32 dlid, 653 u8 sc5, const struct ib_grh *old_grh) 654 { 655 u64 pbc, pbc_flags = 0; 656 u32 bth0, plen, vl, hwords = 7; 657 u16 len; 658 u8 l4; 659 struct hfi1_16b_header hdr; 660 struct ib_other_headers *ohdr; 661 struct pio_buf *pbuf; 662 struct send_context *ctxt = qp_to_send_context(qp, sc5); 663 struct hfi1_pportdata *ppd = ppd_from_ibp(ibp); 664 u32 nwords; 665 666 /* Populate length */ 667 nwords = ((hfi1_get_16b_padding(hwords << 2, 0) + 668 SIZE_OF_LT) >> 2) + SIZE_OF_CRC; 669 if (old_grh) { 670 struct ib_grh *grh = &hdr.u.l.grh; 671 672 grh->version_tclass_flow = old_grh->version_tclass_flow; 673 grh->paylen = cpu_to_be16( 674 (hwords - LRH_16B_DWORDS + nwords) << 2); 675 grh->hop_limit = 0xff; 676 grh->sgid = old_grh->dgid; 677 grh->dgid = old_grh->sgid; 678 ohdr = &hdr.u.l.oth; 679 l4 = OPA_16B_L4_IB_GLOBAL; 680 hwords += sizeof(struct ib_grh) / sizeof(u32); 681 } else { 682 ohdr = &hdr.u.oth; 683 l4 = OPA_16B_L4_IB_LOCAL; 684 } 685 686 /* BIT 16 to 19 is TVER. Bit 20 to 22 is pad cnt */ 687 bth0 = (IB_OPCODE_CNP << 24) | (1 << 16) | 688 (hfi1_get_16b_padding(hwords << 2, 0) << 20); 689 ohdr->bth[0] = cpu_to_be32(bth0); 690 691 ohdr->bth[1] = cpu_to_be32(remote_qpn); 692 ohdr->bth[2] = 0; /* PSN 0 */ 693 694 /* Convert dwords to flits */ 695 len = (hwords + nwords) >> 1; 696 hfi1_make_16b_hdr(&hdr, slid, dlid, len, pkey, 1, 0, l4, sc5); 697 698 plen = 2 /* PBC */ + hwords + nwords; 699 pbc_flags |= PBC_PACKET_BYPASS | PBC_INSERT_BYPASS_ICRC; 700 vl = sc_to_vlt(ppd->dd, sc5); 701 pbc = create_pbc(ppd, pbc_flags, qp->srate_mbps, vl, plen); 702 if (ctxt) { 703 pbuf = sc_buffer_alloc(ctxt, plen, NULL, NULL); 704 if (pbuf) 705 ppd->dd->pio_inline_send(ppd->dd, pbuf, pbc, 706 &hdr, hwords); 707 } 708 } 709 710 void return_cnp(struct hfi1_ibport *ibp, struct rvt_qp *qp, u32 remote_qpn, 711 u16 pkey, u32 slid, u32 dlid, u8 sc5, 712 const struct ib_grh *old_grh) 713 { 714 u64 pbc, pbc_flags = 0; 715 u32 bth0, plen, vl, hwords = 5; 716 u16 lrh0; 717 u8 sl = ibp->sc_to_sl[sc5]; 718 struct ib_header hdr; 719 struct ib_other_headers *ohdr; 720 struct pio_buf *pbuf; 721 struct send_context *ctxt = qp_to_send_context(qp, sc5); 722 struct hfi1_pportdata *ppd = ppd_from_ibp(ibp); 723 724 if (old_grh) { 725 struct ib_grh *grh = &hdr.u.l.grh; 726 727 grh->version_tclass_flow = old_grh->version_tclass_flow; 728 grh->paylen = cpu_to_be16( 729 (hwords - LRH_9B_DWORDS + SIZE_OF_CRC) << 2); 730 grh->hop_limit = 0xff; 731 grh->sgid = old_grh->dgid; 732 grh->dgid = old_grh->sgid; 733 ohdr = &hdr.u.l.oth; 734 lrh0 = HFI1_LRH_GRH; 735 hwords += sizeof(struct ib_grh) / sizeof(u32); 736 } else { 737 ohdr = &hdr.u.oth; 738 lrh0 = HFI1_LRH_BTH; 739 } 740 741 lrh0 |= (sc5 & 0xf) << 12 | sl << 4; 742 743 bth0 = pkey | (IB_OPCODE_CNP << 24); 744 ohdr->bth[0] = cpu_to_be32(bth0); 745 746 ohdr->bth[1] = cpu_to_be32(remote_qpn | (1 << IB_BECN_SHIFT)); 747 ohdr->bth[2] = 0; /* PSN 0 */ 748 749 hfi1_make_ib_hdr(&hdr, lrh0, hwords + SIZE_OF_CRC, dlid, slid); 750 plen = 2 /* PBC */ + hwords; 751 pbc_flags |= (ib_is_sc5(sc5) << PBC_DC_INFO_SHIFT); 752 vl = sc_to_vlt(ppd->dd, sc5); 753 pbc = create_pbc(ppd, pbc_flags, qp->srate_mbps, vl, plen); 754 if (ctxt) { 755 pbuf = sc_buffer_alloc(ctxt, plen, NULL, NULL); 756 if (pbuf) 757 ppd->dd->pio_inline_send(ppd->dd, pbuf, pbc, 758 &hdr, hwords); 759 } 760 } 761 762 /* 763 * opa_smp_check() - Do the regular pkey checking, and the additional 764 * checks for SMPs specified in OPAv1 rev 1.0, 9/19/2016 update, section 765 * 9.10.25 ("SMA Packet Checks"). 766 * 767 * Note that: 768 * - Checks are done using the pkey directly from the packet's BTH, 769 * and specifically _not_ the pkey that we attach to the completion, 770 * which may be different. 771 * - These checks are specifically for "non-local" SMPs (i.e., SMPs 772 * which originated on another node). SMPs which are sent from, and 773 * destined to this node are checked in opa_local_smp_check(). 774 * 775 * At the point where opa_smp_check() is called, we know: 776 * - destination QP is QP0 777 * 778 * opa_smp_check() returns 0 if all checks succeed, 1 otherwise. 779 */ 780 static int opa_smp_check(struct hfi1_ibport *ibp, u16 pkey, u8 sc5, 781 struct rvt_qp *qp, u16 slid, struct opa_smp *smp) 782 { 783 struct hfi1_pportdata *ppd = ppd_from_ibp(ibp); 784 785 /* 786 * I don't think it's possible for us to get here with sc != 0xf, 787 * but check it to be certain. 788 */ 789 if (sc5 != 0xf) 790 return 1; 791 792 if (rcv_pkey_check(ppd, pkey, sc5, slid)) 793 return 1; 794 795 /* 796 * At this point we know (and so don't need to check again) that 797 * the pkey is either LIM_MGMT_P_KEY, or FULL_MGMT_P_KEY 798 * (see ingress_pkey_check). 799 */ 800 if (smp->mgmt_class != IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE && 801 smp->mgmt_class != IB_MGMT_CLASS_SUBN_LID_ROUTED) { 802 ingress_pkey_table_fail(ppd, pkey, slid); 803 return 1; 804 } 805 806 /* 807 * SMPs fall into one of four (disjoint) categories: 808 * SMA request, SMA response, SMA trap, or SMA trap repress. 809 * Our response depends, in part, on which type of SMP we're 810 * processing. 811 * 812 * If this is an SMA response, skip the check here. 813 * 814 * If this is an SMA request or SMA trap repress: 815 * - pkey != FULL_MGMT_P_KEY => 816 * increment port recv constraint errors, drop MAD 817 * 818 * Otherwise: 819 * - accept if the port is running an SM 820 * - drop MAD if it's an SMA trap 821 * - pkey == FULL_MGMT_P_KEY => 822 * reply with unsupported method 823 * - pkey != FULL_MGMT_P_KEY => 824 * increment port recv constraint errors, drop MAD 825 */ 826 switch (smp->method) { 827 case IB_MGMT_METHOD_GET_RESP: 828 case IB_MGMT_METHOD_REPORT_RESP: 829 break; 830 case IB_MGMT_METHOD_GET: 831 case IB_MGMT_METHOD_SET: 832 case IB_MGMT_METHOD_REPORT: 833 case IB_MGMT_METHOD_TRAP_REPRESS: 834 if (pkey != FULL_MGMT_P_KEY) { 835 ingress_pkey_table_fail(ppd, pkey, slid); 836 return 1; 837 } 838 break; 839 default: 840 if (ibp->rvp.port_cap_flags & IB_PORT_SM) 841 return 0; 842 if (smp->method == IB_MGMT_METHOD_TRAP) 843 return 1; 844 if (pkey == FULL_MGMT_P_KEY) { 845 smp->status |= IB_SMP_UNSUP_METHOD; 846 return 0; 847 } 848 ingress_pkey_table_fail(ppd, pkey, slid); 849 return 1; 850 } 851 return 0; 852 } 853 854 /** 855 * hfi1_ud_rcv - receive an incoming UD packet 856 * @ibp: the port the packet came in on 857 * @hdr: the packet header 858 * @rcv_flags: flags relevant to rcv processing 859 * @data: the packet data 860 * @tlen: the packet length 861 * @qp: the QP the packet came on 862 * 863 * This is called from qp_rcv() to process an incoming UD packet 864 * for the given QP. 865 * Called at interrupt level. 866 */ 867 void hfi1_ud_rcv(struct hfi1_packet *packet) 868 { 869 u32 hdrsize = packet->hlen; 870 struct ib_wc wc; 871 u32 src_qp; 872 u16 pkey; 873 int mgmt_pkey_idx = -1; 874 struct hfi1_ibport *ibp = rcd_to_iport(packet->rcd); 875 struct hfi1_pportdata *ppd = ppd_from_ibp(ibp); 876 void *data = packet->payload; 877 u32 tlen = packet->tlen; 878 struct rvt_qp *qp = packet->qp; 879 u8 sc5 = packet->sc; 880 u8 sl_from_sc; 881 u8 opcode = packet->opcode; 882 u8 sl = packet->sl; 883 u32 dlid = packet->dlid; 884 u32 slid = packet->slid; 885 u8 extra_bytes; 886 u8 l4 = 0; 887 bool dlid_is_permissive; 888 bool slid_is_permissive; 889 bool solicited = false; 890 891 extra_bytes = packet->pad + packet->extra_byte + (SIZE_OF_CRC << 2); 892 893 if (packet->etype == RHF_RCV_TYPE_BYPASS) { 894 u32 permissive_lid = 895 opa_get_lid(be32_to_cpu(OPA_LID_PERMISSIVE), 16B); 896 897 l4 = hfi1_16B_get_l4(packet->hdr); 898 pkey = hfi1_16B_get_pkey(packet->hdr); 899 dlid_is_permissive = (dlid == permissive_lid); 900 slid_is_permissive = (slid == permissive_lid); 901 } else { 902 pkey = ib_bth_get_pkey(packet->ohdr); 903 dlid_is_permissive = (dlid == be16_to_cpu(IB_LID_PERMISSIVE)); 904 slid_is_permissive = (slid == be16_to_cpu(IB_LID_PERMISSIVE)); 905 } 906 sl_from_sc = ibp->sc_to_sl[sc5]; 907 908 if (likely(l4 != OPA_16B_L4_FM)) { 909 src_qp = ib_get_sqpn(packet->ohdr); 910 solicited = ib_bth_is_solicited(packet->ohdr); 911 } else { 912 src_qp = hfi1_16B_get_src_qpn(packet->mgmt); 913 } 914 915 process_ecn(qp, packet, (opcode != IB_OPCODE_CNP)); 916 /* 917 * Get the number of bytes the message was padded by 918 * and drop incomplete packets. 919 */ 920 if (unlikely(tlen < (hdrsize + extra_bytes))) 921 goto drop; 922 923 tlen -= hdrsize + extra_bytes; 924 925 /* 926 * Check that the permissive LID is only used on QP0 927 * and the QKEY matches (see 9.6.1.4.1 and 9.6.1.5.1). 928 */ 929 if (qp->ibqp.qp_num) { 930 if (unlikely(dlid_is_permissive || slid_is_permissive)) 931 goto drop; 932 if (qp->ibqp.qp_num > 1) { 933 if (unlikely(rcv_pkey_check(ppd, pkey, sc5, slid))) { 934 /* 935 * Traps will not be sent for packets dropped 936 * by the HW. This is fine, as sending trap 937 * for invalid pkeys is optional according to 938 * IB spec (release 1.3, section 10.9.4) 939 */ 940 hfi1_bad_pkey(ibp, 941 pkey, sl, 942 src_qp, qp->ibqp.qp_num, 943 slid, dlid); 944 return; 945 } 946 } else { 947 /* GSI packet */ 948 mgmt_pkey_idx = hfi1_lookup_pkey_idx(ibp, pkey); 949 if (mgmt_pkey_idx < 0) 950 goto drop; 951 } 952 if (unlikely(l4 != OPA_16B_L4_FM && 953 ib_get_qkey(packet->ohdr) != qp->qkey)) 954 return; /* Silent drop */ 955 956 /* Drop invalid MAD packets (see 13.5.3.1). */ 957 if (unlikely(qp->ibqp.qp_num == 1 && 958 (tlen > 2048 || (sc5 == 0xF)))) 959 goto drop; 960 } else { 961 /* Received on QP0, and so by definition, this is an SMP */ 962 struct opa_smp *smp = (struct opa_smp *)data; 963 964 if (opa_smp_check(ibp, pkey, sc5, qp, slid, smp)) 965 goto drop; 966 967 if (tlen > 2048) 968 goto drop; 969 if ((dlid_is_permissive || slid_is_permissive) && 970 smp->mgmt_class != IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) 971 goto drop; 972 973 /* look up SMI pkey */ 974 mgmt_pkey_idx = hfi1_lookup_pkey_idx(ibp, pkey); 975 if (mgmt_pkey_idx < 0) 976 goto drop; 977 } 978 979 if (qp->ibqp.qp_num > 1 && 980 opcode == IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE) { 981 wc.ex.imm_data = packet->ohdr->u.ud.imm_data; 982 wc.wc_flags = IB_WC_WITH_IMM; 983 tlen -= sizeof(u32); 984 } else if (opcode == IB_OPCODE_UD_SEND_ONLY) { 985 wc.ex.imm_data = 0; 986 wc.wc_flags = 0; 987 } else { 988 goto drop; 989 } 990 991 /* 992 * A GRH is expected to precede the data even if not 993 * present on the wire. 994 */ 995 wc.byte_len = tlen + sizeof(struct ib_grh); 996 997 /* 998 * Get the next work request entry to find where to put the data. 999 */ 1000 if (qp->r_flags & RVT_R_REUSE_SGE) { 1001 qp->r_flags &= ~RVT_R_REUSE_SGE; 1002 } else { 1003 int ret; 1004 1005 ret = rvt_get_rwqe(qp, false); 1006 if (ret < 0) { 1007 rvt_rc_error(qp, IB_WC_LOC_QP_OP_ERR); 1008 return; 1009 } 1010 if (!ret) { 1011 if (qp->ibqp.qp_num == 0) 1012 ibp->rvp.n_vl15_dropped++; 1013 return; 1014 } 1015 } 1016 /* Silently drop packets which are too big. */ 1017 if (unlikely(wc.byte_len > qp->r_len)) { 1018 qp->r_flags |= RVT_R_REUSE_SGE; 1019 goto drop; 1020 } 1021 if (packet->grh) { 1022 rvt_copy_sge(qp, &qp->r_sge, packet->grh, 1023 sizeof(struct ib_grh), true, false); 1024 wc.wc_flags |= IB_WC_GRH; 1025 } else if (packet->etype == RHF_RCV_TYPE_BYPASS) { 1026 struct ib_grh grh; 1027 /* 1028 * Assuming we only created 16B on the send side 1029 * if we want to use large LIDs, since GRH was stripped 1030 * out when creating 16B, add back the GRH here. 1031 */ 1032 hfi1_make_ext_grh(packet, &grh, slid, dlid); 1033 rvt_copy_sge(qp, &qp->r_sge, &grh, 1034 sizeof(struct ib_grh), true, false); 1035 wc.wc_flags |= IB_WC_GRH; 1036 } else { 1037 rvt_skip_sge(&qp->r_sge, sizeof(struct ib_grh), true); 1038 } 1039 rvt_copy_sge(qp, &qp->r_sge, data, wc.byte_len - sizeof(struct ib_grh), 1040 true, false); 1041 rvt_put_ss(&qp->r_sge); 1042 if (!test_and_clear_bit(RVT_R_WRID_VALID, &qp->r_aflags)) 1043 return; 1044 wc.wr_id = qp->r_wr_id; 1045 wc.status = IB_WC_SUCCESS; 1046 wc.opcode = IB_WC_RECV; 1047 wc.vendor_err = 0; 1048 wc.qp = &qp->ibqp; 1049 wc.src_qp = src_qp; 1050 1051 if (qp->ibqp.qp_type == IB_QPT_GSI || 1052 qp->ibqp.qp_type == IB_QPT_SMI) { 1053 if (mgmt_pkey_idx < 0) { 1054 if (net_ratelimit()) { 1055 struct hfi1_devdata *dd = ppd->dd; 1056 1057 dd_dev_err(dd, "QP type %d mgmt_pkey_idx < 0 and packet not dropped???\n", 1058 qp->ibqp.qp_type); 1059 mgmt_pkey_idx = 0; 1060 } 1061 } 1062 wc.pkey_index = (unsigned)mgmt_pkey_idx; 1063 } else { 1064 wc.pkey_index = 0; 1065 } 1066 if (slid_is_permissive) 1067 slid = be32_to_cpu(OPA_LID_PERMISSIVE); 1068 wc.slid = slid & U16_MAX; 1069 wc.sl = sl_from_sc; 1070 1071 /* 1072 * Save the LMC lower bits if the destination LID is a unicast LID. 1073 */ 1074 wc.dlid_path_bits = hfi1_check_mcast(dlid) ? 0 : 1075 dlid & ((1 << ppd_from_ibp(ibp)->lmc) - 1); 1076 wc.port_num = qp->port_num; 1077 /* Signal completion event if the solicited bit is set. */ 1078 rvt_cq_enter(ibcq_to_rvtcq(qp->ibqp.recv_cq), &wc, solicited); 1079 return; 1080 1081 drop: 1082 ibp->rvp.n_pkt_drops++; 1083 } 1084