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 "hfi.h" 49 #include "verbs_txreq.h" 50 #include "qp.h" 51 52 /* cut down ridiculously long IB macro names */ 53 #define OP(x) UC_OP(x) 54 55 /** 56 * hfi1_make_uc_req - construct a request packet (SEND, RDMA write) 57 * @qp: a pointer to the QP 58 * 59 * Assume s_lock is held. 60 * 61 * Return 1 if constructed; otherwise, return 0. 62 */ 63 int hfi1_make_uc_req(struct rvt_qp *qp, struct hfi1_pkt_state *ps) 64 { 65 struct hfi1_qp_priv *priv = qp->priv; 66 struct ib_other_headers *ohdr; 67 struct rvt_swqe *wqe; 68 u32 hwords; 69 u32 bth0 = 0; 70 u32 len; 71 u32 pmtu = qp->pmtu; 72 int middle = 0; 73 74 ps->s_txreq = get_txreq(ps->dev, qp); 75 if (!ps->s_txreq) 76 goto bail_no_tx; 77 78 if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_SEND_OK)) { 79 if (!(ib_rvt_state_ops[qp->state] & RVT_FLUSH_SEND)) 80 goto bail; 81 /* We are in the error state, flush the work request. */ 82 if (qp->s_last == READ_ONCE(qp->s_head)) 83 goto bail; 84 /* If DMAs are in progress, we can't flush immediately. */ 85 if (iowait_sdma_pending(&priv->s_iowait)) { 86 qp->s_flags |= RVT_S_WAIT_DMA; 87 goto bail; 88 } 89 clear_ahg(qp); 90 wqe = rvt_get_swqe_ptr(qp, qp->s_last); 91 rvt_send_complete(qp, wqe, IB_WC_WR_FLUSH_ERR); 92 goto done_free_tx; 93 } 94 95 if (priv->hdr_type == HFI1_PKT_TYPE_9B) { 96 /* header size in 32-bit words LRH+BTH = (8+12)/4. */ 97 hwords = 5; 98 if (rdma_ah_get_ah_flags(&qp->remote_ah_attr) & IB_AH_GRH) 99 ohdr = &ps->s_txreq->phdr.hdr.ibh.u.l.oth; 100 else 101 ohdr = &ps->s_txreq->phdr.hdr.ibh.u.oth; 102 } else { 103 /* header size in 32-bit words 16B LRH+BTH = (16+12)/4. */ 104 hwords = 7; 105 if ((rdma_ah_get_ah_flags(&qp->remote_ah_attr) & IB_AH_GRH) && 106 (hfi1_check_mcast(rdma_ah_get_dlid(&qp->remote_ah_attr)))) 107 ohdr = &ps->s_txreq->phdr.hdr.opah.u.l.oth; 108 else 109 ohdr = &ps->s_txreq->phdr.hdr.opah.u.oth; 110 } 111 112 /* Get the next send request. */ 113 wqe = rvt_get_swqe_ptr(qp, qp->s_cur); 114 qp->s_wqe = NULL; 115 switch (qp->s_state) { 116 default: 117 if (!(ib_rvt_state_ops[qp->state] & 118 RVT_PROCESS_NEXT_SEND_OK)) 119 goto bail; 120 /* Check if send work queue is empty. */ 121 if (qp->s_cur == READ_ONCE(qp->s_head)) { 122 clear_ahg(qp); 123 goto bail; 124 } 125 /* 126 * Local operations are processed immediately 127 * after all prior requests have completed. 128 */ 129 if (wqe->wr.opcode == IB_WR_REG_MR || 130 wqe->wr.opcode == IB_WR_LOCAL_INV) { 131 int local_ops = 0; 132 int err = 0; 133 134 if (qp->s_last != qp->s_cur) 135 goto bail; 136 if (++qp->s_cur == qp->s_size) 137 qp->s_cur = 0; 138 if (!(wqe->wr.send_flags & RVT_SEND_COMPLETION_ONLY)) { 139 err = rvt_invalidate_rkey( 140 qp, wqe->wr.ex.invalidate_rkey); 141 local_ops = 1; 142 } 143 rvt_send_complete(qp, wqe, err ? IB_WC_LOC_PROT_ERR 144 : IB_WC_SUCCESS); 145 if (local_ops) 146 atomic_dec(&qp->local_ops_pending); 147 goto done_free_tx; 148 } 149 /* 150 * Start a new request. 151 */ 152 qp->s_psn = wqe->psn; 153 qp->s_sge.sge = wqe->sg_list[0]; 154 qp->s_sge.sg_list = wqe->sg_list + 1; 155 qp->s_sge.num_sge = wqe->wr.num_sge; 156 qp->s_sge.total_len = wqe->length; 157 len = wqe->length; 158 qp->s_len = len; 159 switch (wqe->wr.opcode) { 160 case IB_WR_SEND: 161 case IB_WR_SEND_WITH_IMM: 162 if (len > pmtu) { 163 qp->s_state = OP(SEND_FIRST); 164 len = pmtu; 165 break; 166 } 167 if (wqe->wr.opcode == IB_WR_SEND) { 168 qp->s_state = OP(SEND_ONLY); 169 } else { 170 qp->s_state = 171 OP(SEND_ONLY_WITH_IMMEDIATE); 172 /* Immediate data comes after the BTH */ 173 ohdr->u.imm_data = wqe->wr.ex.imm_data; 174 hwords += 1; 175 } 176 if (wqe->wr.send_flags & IB_SEND_SOLICITED) 177 bth0 |= IB_BTH_SOLICITED; 178 qp->s_wqe = wqe; 179 if (++qp->s_cur >= qp->s_size) 180 qp->s_cur = 0; 181 break; 182 183 case IB_WR_RDMA_WRITE: 184 case IB_WR_RDMA_WRITE_WITH_IMM: 185 ohdr->u.rc.reth.vaddr = 186 cpu_to_be64(wqe->rdma_wr.remote_addr); 187 ohdr->u.rc.reth.rkey = 188 cpu_to_be32(wqe->rdma_wr.rkey); 189 ohdr->u.rc.reth.length = cpu_to_be32(len); 190 hwords += sizeof(struct ib_reth) / 4; 191 if (len > pmtu) { 192 qp->s_state = OP(RDMA_WRITE_FIRST); 193 len = pmtu; 194 break; 195 } 196 if (wqe->wr.opcode == IB_WR_RDMA_WRITE) { 197 qp->s_state = OP(RDMA_WRITE_ONLY); 198 } else { 199 qp->s_state = 200 OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE); 201 /* Immediate data comes after the RETH */ 202 ohdr->u.rc.imm_data = wqe->wr.ex.imm_data; 203 hwords += 1; 204 if (wqe->wr.send_flags & IB_SEND_SOLICITED) 205 bth0 |= IB_BTH_SOLICITED; 206 } 207 qp->s_wqe = wqe; 208 if (++qp->s_cur >= qp->s_size) 209 qp->s_cur = 0; 210 break; 211 212 default: 213 goto bail; 214 } 215 break; 216 217 case OP(SEND_FIRST): 218 qp->s_state = OP(SEND_MIDDLE); 219 /* FALLTHROUGH */ 220 case OP(SEND_MIDDLE): 221 len = qp->s_len; 222 if (len > pmtu) { 223 len = pmtu; 224 middle = HFI1_CAP_IS_KSET(SDMA_AHG); 225 break; 226 } 227 if (wqe->wr.opcode == IB_WR_SEND) { 228 qp->s_state = OP(SEND_LAST); 229 } else { 230 qp->s_state = OP(SEND_LAST_WITH_IMMEDIATE); 231 /* Immediate data comes after the BTH */ 232 ohdr->u.imm_data = wqe->wr.ex.imm_data; 233 hwords += 1; 234 } 235 if (wqe->wr.send_flags & IB_SEND_SOLICITED) 236 bth0 |= IB_BTH_SOLICITED; 237 qp->s_wqe = wqe; 238 if (++qp->s_cur >= qp->s_size) 239 qp->s_cur = 0; 240 break; 241 242 case OP(RDMA_WRITE_FIRST): 243 qp->s_state = OP(RDMA_WRITE_MIDDLE); 244 /* FALLTHROUGH */ 245 case OP(RDMA_WRITE_MIDDLE): 246 len = qp->s_len; 247 if (len > pmtu) { 248 len = pmtu; 249 middle = HFI1_CAP_IS_KSET(SDMA_AHG); 250 break; 251 } 252 if (wqe->wr.opcode == IB_WR_RDMA_WRITE) { 253 qp->s_state = OP(RDMA_WRITE_LAST); 254 } else { 255 qp->s_state = 256 OP(RDMA_WRITE_LAST_WITH_IMMEDIATE); 257 /* Immediate data comes after the BTH */ 258 ohdr->u.imm_data = wqe->wr.ex.imm_data; 259 hwords += 1; 260 if (wqe->wr.send_flags & IB_SEND_SOLICITED) 261 bth0 |= IB_BTH_SOLICITED; 262 } 263 qp->s_wqe = wqe; 264 if (++qp->s_cur >= qp->s_size) 265 qp->s_cur = 0; 266 break; 267 } 268 qp->s_len -= len; 269 ps->s_txreq->hdr_dwords = hwords; 270 ps->s_txreq->sde = priv->s_sde; 271 ps->s_txreq->ss = &qp->s_sge; 272 ps->s_txreq->s_cur_size = len; 273 hfi1_make_ruc_header(qp, ohdr, bth0 | (qp->s_state << 24), 274 qp->remote_qpn, mask_psn(qp->s_psn++), 275 middle, ps); 276 return 1; 277 278 done_free_tx: 279 hfi1_put_txreq(ps->s_txreq); 280 ps->s_txreq = NULL; 281 return 1; 282 283 bail: 284 hfi1_put_txreq(ps->s_txreq); 285 286 bail_no_tx: 287 ps->s_txreq = NULL; 288 qp->s_flags &= ~RVT_S_BUSY; 289 return 0; 290 } 291 292 /** 293 * hfi1_uc_rcv - handle an incoming UC packet 294 * @ibp: the port the packet came in on 295 * @hdr: the header of the packet 296 * @rcv_flags: flags relevant to rcv processing 297 * @data: the packet data 298 * @tlen: the length of the packet 299 * @qp: the QP for this packet. 300 * 301 * This is called from qp_rcv() to process an incoming UC packet 302 * for the given QP. 303 * Called at interrupt level. 304 */ 305 void hfi1_uc_rcv(struct hfi1_packet *packet) 306 { 307 struct hfi1_ibport *ibp = rcd_to_iport(packet->rcd); 308 void *data = packet->payload; 309 u32 tlen = packet->tlen; 310 struct rvt_qp *qp = packet->qp; 311 struct ib_other_headers *ohdr = packet->ohdr; 312 u32 opcode = packet->opcode; 313 u32 hdrsize = packet->hlen; 314 u32 psn; 315 u32 pad = packet->pad; 316 struct ib_wc wc; 317 u32 pmtu = qp->pmtu; 318 struct ib_reth *reth; 319 int ret; 320 u8 extra_bytes = pad + packet->extra_byte + (SIZE_OF_CRC << 2); 321 322 if (hfi1_ruc_check_hdr(ibp, packet)) 323 return; 324 325 process_ecn(qp, packet); 326 327 psn = ib_bth_get_psn(ohdr); 328 /* Compare the PSN verses the expected PSN. */ 329 if (unlikely(cmp_psn(psn, qp->r_psn) != 0)) { 330 /* 331 * Handle a sequence error. 332 * Silently drop any current message. 333 */ 334 qp->r_psn = psn; 335 inv: 336 if (qp->r_state == OP(SEND_FIRST) || 337 qp->r_state == OP(SEND_MIDDLE)) { 338 set_bit(RVT_R_REWIND_SGE, &qp->r_aflags); 339 qp->r_sge.num_sge = 0; 340 } else { 341 rvt_put_ss(&qp->r_sge); 342 } 343 qp->r_state = OP(SEND_LAST); 344 switch (opcode) { 345 case OP(SEND_FIRST): 346 case OP(SEND_ONLY): 347 case OP(SEND_ONLY_WITH_IMMEDIATE): 348 goto send_first; 349 350 case OP(RDMA_WRITE_FIRST): 351 case OP(RDMA_WRITE_ONLY): 352 case OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE): 353 goto rdma_first; 354 355 default: 356 goto drop; 357 } 358 } 359 360 /* Check for opcode sequence errors. */ 361 switch (qp->r_state) { 362 case OP(SEND_FIRST): 363 case OP(SEND_MIDDLE): 364 if (opcode == OP(SEND_MIDDLE) || 365 opcode == OP(SEND_LAST) || 366 opcode == OP(SEND_LAST_WITH_IMMEDIATE)) 367 break; 368 goto inv; 369 370 case OP(RDMA_WRITE_FIRST): 371 case OP(RDMA_WRITE_MIDDLE): 372 if (opcode == OP(RDMA_WRITE_MIDDLE) || 373 opcode == OP(RDMA_WRITE_LAST) || 374 opcode == OP(RDMA_WRITE_LAST_WITH_IMMEDIATE)) 375 break; 376 goto inv; 377 378 default: 379 if (opcode == OP(SEND_FIRST) || 380 opcode == OP(SEND_ONLY) || 381 opcode == OP(SEND_ONLY_WITH_IMMEDIATE) || 382 opcode == OP(RDMA_WRITE_FIRST) || 383 opcode == OP(RDMA_WRITE_ONLY) || 384 opcode == OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE)) 385 break; 386 goto inv; 387 } 388 389 if (qp->state == IB_QPS_RTR && !(qp->r_flags & RVT_R_COMM_EST)) 390 rvt_comm_est(qp); 391 392 /* OK, process the packet. */ 393 switch (opcode) { 394 case OP(SEND_FIRST): 395 case OP(SEND_ONLY): 396 case OP(SEND_ONLY_WITH_IMMEDIATE): 397 send_first: 398 if (test_and_clear_bit(RVT_R_REWIND_SGE, &qp->r_aflags)) { 399 qp->r_sge = qp->s_rdma_read_sge; 400 } else { 401 ret = rvt_get_rwqe(qp, false); 402 if (ret < 0) 403 goto op_err; 404 if (!ret) 405 goto drop; 406 /* 407 * qp->s_rdma_read_sge will be the owner 408 * of the mr references. 409 */ 410 qp->s_rdma_read_sge = qp->r_sge; 411 } 412 qp->r_rcv_len = 0; 413 if (opcode == OP(SEND_ONLY)) 414 goto no_immediate_data; 415 else if (opcode == OP(SEND_ONLY_WITH_IMMEDIATE)) 416 goto send_last_imm; 417 /* FALLTHROUGH */ 418 case OP(SEND_MIDDLE): 419 /* Check for invalid length PMTU or posted rwqe len. */ 420 /* 421 * There will be no padding for 9B packet but 16B packets 422 * will come in with some padding since we always add 423 * CRC and LT bytes which will need to be flit aligned 424 */ 425 if (unlikely(tlen != (hdrsize + pmtu + extra_bytes))) 426 goto rewind; 427 qp->r_rcv_len += pmtu; 428 if (unlikely(qp->r_rcv_len > qp->r_len)) 429 goto rewind; 430 rvt_copy_sge(qp, &qp->r_sge, data, pmtu, false, false); 431 break; 432 433 case OP(SEND_LAST_WITH_IMMEDIATE): 434 send_last_imm: 435 wc.ex.imm_data = ohdr->u.imm_data; 436 wc.wc_flags = IB_WC_WITH_IMM; 437 goto send_last; 438 case OP(SEND_LAST): 439 no_immediate_data: 440 wc.ex.imm_data = 0; 441 wc.wc_flags = 0; 442 send_last: 443 /* Check for invalid length. */ 444 /* LAST len should be >= 1 */ 445 if (unlikely(tlen < (hdrsize + extra_bytes))) 446 goto rewind; 447 /* Don't count the CRC. */ 448 tlen -= (hdrsize + extra_bytes); 449 wc.byte_len = tlen + qp->r_rcv_len; 450 if (unlikely(wc.byte_len > qp->r_len)) 451 goto rewind; 452 wc.opcode = IB_WC_RECV; 453 rvt_copy_sge(qp, &qp->r_sge, data, tlen, false, false); 454 rvt_put_ss(&qp->s_rdma_read_sge); 455 last_imm: 456 wc.wr_id = qp->r_wr_id; 457 wc.status = IB_WC_SUCCESS; 458 wc.qp = &qp->ibqp; 459 wc.src_qp = qp->remote_qpn; 460 wc.slid = rdma_ah_get_dlid(&qp->remote_ah_attr) & U16_MAX; 461 /* 462 * It seems that IB mandates the presence of an SL in a 463 * work completion only for the UD transport (see section 464 * 11.4.2 of IBTA Vol. 1). 465 * 466 * However, the way the SL is chosen below is consistent 467 * with the way that IB/qib works and is trying avoid 468 * introducing incompatibilities. 469 * 470 * See also OPA Vol. 1, section 9.7.6, and table 9-17. 471 */ 472 wc.sl = rdma_ah_get_sl(&qp->remote_ah_attr); 473 /* zero fields that are N/A */ 474 wc.vendor_err = 0; 475 wc.pkey_index = 0; 476 wc.dlid_path_bits = 0; 477 wc.port_num = 0; 478 /* Signal completion event if the solicited bit is set. */ 479 rvt_cq_enter(ibcq_to_rvtcq(qp->ibqp.recv_cq), &wc, 480 ib_bth_is_solicited(ohdr)); 481 break; 482 483 case OP(RDMA_WRITE_FIRST): 484 case OP(RDMA_WRITE_ONLY): 485 case OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE): /* consume RWQE */ 486 rdma_first: 487 if (unlikely(!(qp->qp_access_flags & 488 IB_ACCESS_REMOTE_WRITE))) { 489 goto drop; 490 } 491 reth = &ohdr->u.rc.reth; 492 qp->r_len = be32_to_cpu(reth->length); 493 qp->r_rcv_len = 0; 494 qp->r_sge.sg_list = NULL; 495 if (qp->r_len != 0) { 496 u32 rkey = be32_to_cpu(reth->rkey); 497 u64 vaddr = be64_to_cpu(reth->vaddr); 498 int ok; 499 500 /* Check rkey */ 501 ok = rvt_rkey_ok(qp, &qp->r_sge.sge, qp->r_len, 502 vaddr, rkey, IB_ACCESS_REMOTE_WRITE); 503 if (unlikely(!ok)) 504 goto drop; 505 qp->r_sge.num_sge = 1; 506 } else { 507 qp->r_sge.num_sge = 0; 508 qp->r_sge.sge.mr = NULL; 509 qp->r_sge.sge.vaddr = NULL; 510 qp->r_sge.sge.length = 0; 511 qp->r_sge.sge.sge_length = 0; 512 } 513 if (opcode == OP(RDMA_WRITE_ONLY)) { 514 goto rdma_last; 515 } else if (opcode == OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE)) { 516 wc.ex.imm_data = ohdr->u.rc.imm_data; 517 goto rdma_last_imm; 518 } 519 /* FALLTHROUGH */ 520 case OP(RDMA_WRITE_MIDDLE): 521 /* Check for invalid length PMTU or posted rwqe len. */ 522 if (unlikely(tlen != (hdrsize + pmtu + 4))) 523 goto drop; 524 qp->r_rcv_len += pmtu; 525 if (unlikely(qp->r_rcv_len > qp->r_len)) 526 goto drop; 527 rvt_copy_sge(qp, &qp->r_sge, data, pmtu, true, false); 528 break; 529 530 case OP(RDMA_WRITE_LAST_WITH_IMMEDIATE): 531 wc.ex.imm_data = ohdr->u.imm_data; 532 rdma_last_imm: 533 wc.wc_flags = IB_WC_WITH_IMM; 534 535 /* Check for invalid length. */ 536 /* LAST len should be >= 1 */ 537 if (unlikely(tlen < (hdrsize + pad + 4))) 538 goto drop; 539 /* Don't count the CRC. */ 540 tlen -= (hdrsize + extra_bytes); 541 if (unlikely(tlen + qp->r_rcv_len != qp->r_len)) 542 goto drop; 543 if (test_and_clear_bit(RVT_R_REWIND_SGE, &qp->r_aflags)) { 544 rvt_put_ss(&qp->s_rdma_read_sge); 545 } else { 546 ret = rvt_get_rwqe(qp, true); 547 if (ret < 0) 548 goto op_err; 549 if (!ret) 550 goto drop; 551 } 552 wc.byte_len = qp->r_len; 553 wc.opcode = IB_WC_RECV_RDMA_WITH_IMM; 554 rvt_copy_sge(qp, &qp->r_sge, data, tlen, true, false); 555 rvt_put_ss(&qp->r_sge); 556 goto last_imm; 557 558 case OP(RDMA_WRITE_LAST): 559 rdma_last: 560 /* Check for invalid length. */ 561 /* LAST len should be >= 1 */ 562 if (unlikely(tlen < (hdrsize + pad + 4))) 563 goto drop; 564 /* Don't count the CRC. */ 565 tlen -= (hdrsize + extra_bytes); 566 if (unlikely(tlen + qp->r_rcv_len != qp->r_len)) 567 goto drop; 568 rvt_copy_sge(qp, &qp->r_sge, data, tlen, true, false); 569 rvt_put_ss(&qp->r_sge); 570 break; 571 572 default: 573 /* Drop packet for unknown opcodes. */ 574 goto drop; 575 } 576 qp->r_psn++; 577 qp->r_state = opcode; 578 return; 579 580 rewind: 581 set_bit(RVT_R_REWIND_SGE, &qp->r_aflags); 582 qp->r_sge.num_sge = 0; 583 drop: 584 ibp->rvp.n_pkt_drops++; 585 return; 586 587 op_err: 588 rvt_rc_error(qp, IB_WC_LOC_QP_OP_ERR); 589 } 590