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