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/io.h> 49 #include <rdma/rdma_vt.h> 50 #include <rdma/rdmavt_qp.h> 51 52 #include "hfi.h" 53 #include "qp.h" 54 #include "verbs_txreq.h" 55 #include "trace.h" 56 57 /* cut down ridiculously long IB macro names */ 58 #define OP(x) RC_OP(x) 59 60 static u32 restart_sge(struct rvt_sge_state *ss, struct rvt_swqe *wqe, 61 u32 psn, u32 pmtu) 62 { 63 u32 len; 64 65 len = delta_psn(psn, wqe->psn) * pmtu; 66 ss->sge = wqe->sg_list[0]; 67 ss->sg_list = wqe->sg_list + 1; 68 ss->num_sge = wqe->wr.num_sge; 69 ss->total_len = wqe->length; 70 rvt_skip_sge(ss, len, false); 71 return wqe->length - len; 72 } 73 74 /** 75 * make_rc_ack - construct a response packet (ACK, NAK, or RDMA read) 76 * @dev: the device for this QP 77 * @qp: a pointer to the QP 78 * @ohdr: a pointer to the IB header being constructed 79 * @ps: the xmit packet state 80 * 81 * Return 1 if constructed; otherwise, return 0. 82 * Note that we are in the responder's side of the QP context. 83 * Note the QP s_lock must be held. 84 */ 85 static int make_rc_ack(struct hfi1_ibdev *dev, struct rvt_qp *qp, 86 struct ib_other_headers *ohdr, 87 struct hfi1_pkt_state *ps) 88 { 89 struct rvt_ack_entry *e; 90 u32 hwords; 91 u32 len; 92 u32 bth0; 93 u32 bth2; 94 int middle = 0; 95 u32 pmtu = qp->pmtu; 96 struct hfi1_qp_priv *priv = qp->priv; 97 98 lockdep_assert_held(&qp->s_lock); 99 /* Don't send an ACK if we aren't supposed to. */ 100 if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK)) 101 goto bail; 102 103 if (priv->hdr_type == HFI1_PKT_TYPE_9B) 104 /* header size in 32-bit words LRH+BTH = (8+12)/4. */ 105 hwords = 5; 106 else 107 /* header size in 32-bit words 16B LRH+BTH = (16+12)/4. */ 108 hwords = 7; 109 110 switch (qp->s_ack_state) { 111 case OP(RDMA_READ_RESPONSE_LAST): 112 case OP(RDMA_READ_RESPONSE_ONLY): 113 e = &qp->s_ack_queue[qp->s_tail_ack_queue]; 114 if (e->rdma_sge.mr) { 115 rvt_put_mr(e->rdma_sge.mr); 116 e->rdma_sge.mr = NULL; 117 } 118 /* FALLTHROUGH */ 119 case OP(ATOMIC_ACKNOWLEDGE): 120 /* 121 * We can increment the tail pointer now that the last 122 * response has been sent instead of only being 123 * constructed. 124 */ 125 if (++qp->s_tail_ack_queue > HFI1_MAX_RDMA_ATOMIC) 126 qp->s_tail_ack_queue = 0; 127 /* FALLTHROUGH */ 128 case OP(SEND_ONLY): 129 case OP(ACKNOWLEDGE): 130 /* Check for no next entry in the queue. */ 131 if (qp->r_head_ack_queue == qp->s_tail_ack_queue) { 132 if (qp->s_flags & RVT_S_ACK_PENDING) 133 goto normal; 134 goto bail; 135 } 136 137 e = &qp->s_ack_queue[qp->s_tail_ack_queue]; 138 if (e->opcode == OP(RDMA_READ_REQUEST)) { 139 /* 140 * If a RDMA read response is being resent and 141 * we haven't seen the duplicate request yet, 142 * then stop sending the remaining responses the 143 * responder has seen until the requester re-sends it. 144 */ 145 len = e->rdma_sge.sge_length; 146 if (len && !e->rdma_sge.mr) { 147 qp->s_tail_ack_queue = qp->r_head_ack_queue; 148 goto bail; 149 } 150 /* Copy SGE state in case we need to resend */ 151 ps->s_txreq->mr = e->rdma_sge.mr; 152 if (ps->s_txreq->mr) 153 rvt_get_mr(ps->s_txreq->mr); 154 qp->s_ack_rdma_sge.sge = e->rdma_sge; 155 qp->s_ack_rdma_sge.num_sge = 1; 156 ps->s_txreq->ss = &qp->s_ack_rdma_sge; 157 if (len > pmtu) { 158 len = pmtu; 159 qp->s_ack_state = OP(RDMA_READ_RESPONSE_FIRST); 160 } else { 161 qp->s_ack_state = OP(RDMA_READ_RESPONSE_ONLY); 162 e->sent = 1; 163 } 164 ohdr->u.aeth = rvt_compute_aeth(qp); 165 hwords++; 166 qp->s_ack_rdma_psn = e->psn; 167 bth2 = mask_psn(qp->s_ack_rdma_psn++); 168 } else { 169 /* COMPARE_SWAP or FETCH_ADD */ 170 ps->s_txreq->ss = NULL; 171 len = 0; 172 qp->s_ack_state = OP(ATOMIC_ACKNOWLEDGE); 173 ohdr->u.at.aeth = rvt_compute_aeth(qp); 174 ib_u64_put(e->atomic_data, &ohdr->u.at.atomic_ack_eth); 175 hwords += sizeof(ohdr->u.at) / sizeof(u32); 176 bth2 = mask_psn(e->psn); 177 e->sent = 1; 178 } 179 bth0 = qp->s_ack_state << 24; 180 break; 181 182 case OP(RDMA_READ_RESPONSE_FIRST): 183 qp->s_ack_state = OP(RDMA_READ_RESPONSE_MIDDLE); 184 /* FALLTHROUGH */ 185 case OP(RDMA_READ_RESPONSE_MIDDLE): 186 ps->s_txreq->ss = &qp->s_ack_rdma_sge; 187 ps->s_txreq->mr = qp->s_ack_rdma_sge.sge.mr; 188 if (ps->s_txreq->mr) 189 rvt_get_mr(ps->s_txreq->mr); 190 len = qp->s_ack_rdma_sge.sge.sge_length; 191 if (len > pmtu) { 192 len = pmtu; 193 middle = HFI1_CAP_IS_KSET(SDMA_AHG); 194 } else { 195 ohdr->u.aeth = rvt_compute_aeth(qp); 196 hwords++; 197 qp->s_ack_state = OP(RDMA_READ_RESPONSE_LAST); 198 e = &qp->s_ack_queue[qp->s_tail_ack_queue]; 199 e->sent = 1; 200 } 201 bth0 = qp->s_ack_state << 24; 202 bth2 = mask_psn(qp->s_ack_rdma_psn++); 203 break; 204 205 default: 206 normal: 207 /* 208 * Send a regular ACK. 209 * Set the s_ack_state so we wait until after sending 210 * the ACK before setting s_ack_state to ACKNOWLEDGE 211 * (see above). 212 */ 213 qp->s_ack_state = OP(SEND_ONLY); 214 qp->s_flags &= ~RVT_S_ACK_PENDING; 215 ps->s_txreq->ss = NULL; 216 if (qp->s_nak_state) 217 ohdr->u.aeth = 218 cpu_to_be32((qp->r_msn & IB_MSN_MASK) | 219 (qp->s_nak_state << 220 IB_AETH_CREDIT_SHIFT)); 221 else 222 ohdr->u.aeth = rvt_compute_aeth(qp); 223 hwords++; 224 len = 0; 225 bth0 = OP(ACKNOWLEDGE) << 24; 226 bth2 = mask_psn(qp->s_ack_psn); 227 } 228 qp->s_rdma_ack_cnt++; 229 ps->s_txreq->sde = priv->s_sde; 230 ps->s_txreq->s_cur_size = len; 231 ps->s_txreq->hdr_dwords = hwords; 232 hfi1_make_ruc_header(qp, ohdr, bth0, bth2, middle, ps); 233 return 1; 234 235 bail: 236 qp->s_ack_state = OP(ACKNOWLEDGE); 237 /* 238 * Ensure s_rdma_ack_cnt changes are committed prior to resetting 239 * RVT_S_RESP_PENDING 240 */ 241 smp_wmb(); 242 qp->s_flags &= ~(RVT_S_RESP_PENDING 243 | RVT_S_ACK_PENDING 244 | HFI1_S_AHG_VALID); 245 return 0; 246 } 247 248 /** 249 * hfi1_make_rc_req - construct a request packet (SEND, RDMA r/w, ATOMIC) 250 * @qp: a pointer to the QP 251 * 252 * Assumes s_lock is held. 253 * 254 * Return 1 if constructed; otherwise, return 0. 255 */ 256 int hfi1_make_rc_req(struct rvt_qp *qp, struct hfi1_pkt_state *ps) 257 { 258 struct hfi1_qp_priv *priv = qp->priv; 259 struct hfi1_ibdev *dev = to_idev(qp->ibqp.device); 260 struct ib_other_headers *ohdr; 261 struct rvt_sge_state *ss; 262 struct rvt_swqe *wqe; 263 u32 hwords; 264 u32 len; 265 u32 bth0 = 0; 266 u32 bth2; 267 u32 pmtu = qp->pmtu; 268 char newreq; 269 int middle = 0; 270 int delta; 271 272 lockdep_assert_held(&qp->s_lock); 273 ps->s_txreq = get_txreq(ps->dev, qp); 274 if (!ps->s_txreq) 275 goto bail_no_tx; 276 277 if (priv->hdr_type == HFI1_PKT_TYPE_9B) { 278 /* header size in 32-bit words LRH+BTH = (8+12)/4. */ 279 hwords = 5; 280 if (rdma_ah_get_ah_flags(&qp->remote_ah_attr) & IB_AH_GRH) 281 ohdr = &ps->s_txreq->phdr.hdr.ibh.u.l.oth; 282 else 283 ohdr = &ps->s_txreq->phdr.hdr.ibh.u.oth; 284 } else { 285 /* header size in 32-bit words 16B LRH+BTH = (16+12)/4. */ 286 hwords = 7; 287 if ((rdma_ah_get_ah_flags(&qp->remote_ah_attr) & IB_AH_GRH) && 288 (hfi1_check_mcast(rdma_ah_get_dlid(&qp->remote_ah_attr)))) 289 ohdr = &ps->s_txreq->phdr.hdr.opah.u.l.oth; 290 else 291 ohdr = &ps->s_txreq->phdr.hdr.opah.u.oth; 292 } 293 294 /* Sending responses has higher priority over sending requests. */ 295 if ((qp->s_flags & RVT_S_RESP_PENDING) && 296 make_rc_ack(dev, qp, ohdr, ps)) 297 return 1; 298 299 if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_SEND_OK)) { 300 if (!(ib_rvt_state_ops[qp->state] & RVT_FLUSH_SEND)) 301 goto bail; 302 /* We are in the error state, flush the work request. */ 303 if (qp->s_last == READ_ONCE(qp->s_head)) 304 goto bail; 305 /* If DMAs are in progress, we can't flush immediately. */ 306 if (iowait_sdma_pending(&priv->s_iowait)) { 307 qp->s_flags |= RVT_S_WAIT_DMA; 308 goto bail; 309 } 310 clear_ahg(qp); 311 wqe = rvt_get_swqe_ptr(qp, qp->s_last); 312 rvt_send_complete(qp, wqe, qp->s_last != qp->s_acked ? 313 IB_WC_SUCCESS : IB_WC_WR_FLUSH_ERR); 314 /* will get called again */ 315 goto done_free_tx; 316 } 317 318 if (qp->s_flags & (RVT_S_WAIT_RNR | RVT_S_WAIT_ACK)) 319 goto bail; 320 321 if (cmp_psn(qp->s_psn, qp->s_sending_hpsn) <= 0) { 322 if (cmp_psn(qp->s_sending_psn, qp->s_sending_hpsn) <= 0) { 323 qp->s_flags |= RVT_S_WAIT_PSN; 324 goto bail; 325 } 326 qp->s_sending_psn = qp->s_psn; 327 qp->s_sending_hpsn = qp->s_psn - 1; 328 } 329 330 /* Send a request. */ 331 wqe = rvt_get_swqe_ptr(qp, qp->s_cur); 332 switch (qp->s_state) { 333 default: 334 if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_NEXT_SEND_OK)) 335 goto bail; 336 /* 337 * Resend an old request or start a new one. 338 * 339 * We keep track of the current SWQE so that 340 * we don't reset the "furthest progress" state 341 * if we need to back up. 342 */ 343 newreq = 0; 344 if (qp->s_cur == qp->s_tail) { 345 /* Check if send work queue is empty. */ 346 if (qp->s_tail == READ_ONCE(qp->s_head)) { 347 clear_ahg(qp); 348 goto bail; 349 } 350 /* 351 * If a fence is requested, wait for previous 352 * RDMA read and atomic operations to finish. 353 */ 354 if ((wqe->wr.send_flags & IB_SEND_FENCE) && 355 qp->s_num_rd_atomic) { 356 qp->s_flags |= RVT_S_WAIT_FENCE; 357 goto bail; 358 } 359 /* 360 * Local operations are processed immediately 361 * after all prior requests have completed 362 */ 363 if (wqe->wr.opcode == IB_WR_REG_MR || 364 wqe->wr.opcode == IB_WR_LOCAL_INV) { 365 int local_ops = 0; 366 int err = 0; 367 368 if (qp->s_last != qp->s_cur) 369 goto bail; 370 if (++qp->s_cur == qp->s_size) 371 qp->s_cur = 0; 372 if (++qp->s_tail == qp->s_size) 373 qp->s_tail = 0; 374 if (!(wqe->wr.send_flags & 375 RVT_SEND_COMPLETION_ONLY)) { 376 err = rvt_invalidate_rkey( 377 qp, 378 wqe->wr.ex.invalidate_rkey); 379 local_ops = 1; 380 } 381 rvt_send_complete(qp, wqe, 382 err ? IB_WC_LOC_PROT_ERR 383 : IB_WC_SUCCESS); 384 if (local_ops) 385 atomic_dec(&qp->local_ops_pending); 386 goto done_free_tx; 387 } 388 389 newreq = 1; 390 qp->s_psn = wqe->psn; 391 } 392 /* 393 * Note that we have to be careful not to modify the 394 * original work request since we may need to resend 395 * it. 396 */ 397 len = wqe->length; 398 ss = &qp->s_sge; 399 bth2 = mask_psn(qp->s_psn); 400 switch (wqe->wr.opcode) { 401 case IB_WR_SEND: 402 case IB_WR_SEND_WITH_IMM: 403 case IB_WR_SEND_WITH_INV: 404 /* If no credit, return. */ 405 if (!(qp->s_flags & RVT_S_UNLIMITED_CREDIT) && 406 rvt_cmp_msn(wqe->ssn, qp->s_lsn + 1) > 0) { 407 qp->s_flags |= RVT_S_WAIT_SSN_CREDIT; 408 goto bail; 409 } 410 if (len > pmtu) { 411 qp->s_state = OP(SEND_FIRST); 412 len = pmtu; 413 break; 414 } 415 if (wqe->wr.opcode == IB_WR_SEND) { 416 qp->s_state = OP(SEND_ONLY); 417 } else if (wqe->wr.opcode == IB_WR_SEND_WITH_IMM) { 418 qp->s_state = OP(SEND_ONLY_WITH_IMMEDIATE); 419 /* Immediate data comes after the BTH */ 420 ohdr->u.imm_data = wqe->wr.ex.imm_data; 421 hwords += 1; 422 } else { 423 qp->s_state = OP(SEND_ONLY_WITH_INVALIDATE); 424 /* Invalidate rkey comes after the BTH */ 425 ohdr->u.ieth = cpu_to_be32( 426 wqe->wr.ex.invalidate_rkey); 427 hwords += 1; 428 } 429 if (wqe->wr.send_flags & IB_SEND_SOLICITED) 430 bth0 |= IB_BTH_SOLICITED; 431 bth2 |= IB_BTH_REQ_ACK; 432 if (++qp->s_cur == qp->s_size) 433 qp->s_cur = 0; 434 break; 435 436 case IB_WR_RDMA_WRITE: 437 if (newreq && !(qp->s_flags & RVT_S_UNLIMITED_CREDIT)) 438 qp->s_lsn++; 439 goto no_flow_control; 440 case IB_WR_RDMA_WRITE_WITH_IMM: 441 /* If no credit, return. */ 442 if (!(qp->s_flags & RVT_S_UNLIMITED_CREDIT) && 443 rvt_cmp_msn(wqe->ssn, qp->s_lsn + 1) > 0) { 444 qp->s_flags |= RVT_S_WAIT_SSN_CREDIT; 445 goto bail; 446 } 447 no_flow_control: 448 put_ib_reth_vaddr( 449 wqe->rdma_wr.remote_addr, 450 &ohdr->u.rc.reth); 451 ohdr->u.rc.reth.rkey = 452 cpu_to_be32(wqe->rdma_wr.rkey); 453 ohdr->u.rc.reth.length = cpu_to_be32(len); 454 hwords += sizeof(struct ib_reth) / sizeof(u32); 455 if (len > pmtu) { 456 qp->s_state = OP(RDMA_WRITE_FIRST); 457 len = pmtu; 458 break; 459 } 460 if (wqe->wr.opcode == IB_WR_RDMA_WRITE) { 461 qp->s_state = OP(RDMA_WRITE_ONLY); 462 } else { 463 qp->s_state = 464 OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE); 465 /* Immediate data comes after RETH */ 466 ohdr->u.rc.imm_data = wqe->wr.ex.imm_data; 467 hwords += 1; 468 if (wqe->wr.send_flags & IB_SEND_SOLICITED) 469 bth0 |= IB_BTH_SOLICITED; 470 } 471 bth2 |= IB_BTH_REQ_ACK; 472 if (++qp->s_cur == qp->s_size) 473 qp->s_cur = 0; 474 break; 475 476 case IB_WR_RDMA_READ: 477 /* 478 * Don't allow more operations to be started 479 * than the QP limits allow. 480 */ 481 if (newreq) { 482 if (qp->s_num_rd_atomic >= 483 qp->s_max_rd_atomic) { 484 qp->s_flags |= RVT_S_WAIT_RDMAR; 485 goto bail; 486 } 487 qp->s_num_rd_atomic++; 488 if (!(qp->s_flags & RVT_S_UNLIMITED_CREDIT)) 489 qp->s_lsn++; 490 } 491 put_ib_reth_vaddr( 492 wqe->rdma_wr.remote_addr, 493 &ohdr->u.rc.reth); 494 ohdr->u.rc.reth.rkey = 495 cpu_to_be32(wqe->rdma_wr.rkey); 496 ohdr->u.rc.reth.length = cpu_to_be32(len); 497 qp->s_state = OP(RDMA_READ_REQUEST); 498 hwords += sizeof(ohdr->u.rc.reth) / sizeof(u32); 499 ss = NULL; 500 len = 0; 501 bth2 |= IB_BTH_REQ_ACK; 502 if (++qp->s_cur == qp->s_size) 503 qp->s_cur = 0; 504 break; 505 506 case IB_WR_ATOMIC_CMP_AND_SWP: 507 case IB_WR_ATOMIC_FETCH_AND_ADD: 508 /* 509 * Don't allow more operations to be started 510 * than the QP limits allow. 511 */ 512 if (newreq) { 513 if (qp->s_num_rd_atomic >= 514 qp->s_max_rd_atomic) { 515 qp->s_flags |= RVT_S_WAIT_RDMAR; 516 goto bail; 517 } 518 qp->s_num_rd_atomic++; 519 if (!(qp->s_flags & RVT_S_UNLIMITED_CREDIT)) 520 qp->s_lsn++; 521 } 522 if (wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP) { 523 qp->s_state = OP(COMPARE_SWAP); 524 put_ib_ateth_swap(wqe->atomic_wr.swap, 525 &ohdr->u.atomic_eth); 526 put_ib_ateth_compare(wqe->atomic_wr.compare_add, 527 &ohdr->u.atomic_eth); 528 } else { 529 qp->s_state = OP(FETCH_ADD); 530 put_ib_ateth_swap(wqe->atomic_wr.compare_add, 531 &ohdr->u.atomic_eth); 532 put_ib_ateth_compare(0, &ohdr->u.atomic_eth); 533 } 534 put_ib_ateth_vaddr(wqe->atomic_wr.remote_addr, 535 &ohdr->u.atomic_eth); 536 ohdr->u.atomic_eth.rkey = cpu_to_be32( 537 wqe->atomic_wr.rkey); 538 hwords += sizeof(struct ib_atomic_eth) / sizeof(u32); 539 ss = NULL; 540 len = 0; 541 bth2 |= IB_BTH_REQ_ACK; 542 if (++qp->s_cur == qp->s_size) 543 qp->s_cur = 0; 544 break; 545 546 default: 547 goto bail; 548 } 549 qp->s_sge.sge = wqe->sg_list[0]; 550 qp->s_sge.sg_list = wqe->sg_list + 1; 551 qp->s_sge.num_sge = wqe->wr.num_sge; 552 qp->s_sge.total_len = wqe->length; 553 qp->s_len = wqe->length; 554 if (newreq) { 555 qp->s_tail++; 556 if (qp->s_tail >= qp->s_size) 557 qp->s_tail = 0; 558 } 559 if (wqe->wr.opcode == IB_WR_RDMA_READ) 560 qp->s_psn = wqe->lpsn + 1; 561 else 562 qp->s_psn++; 563 break; 564 565 case OP(RDMA_READ_RESPONSE_FIRST): 566 /* 567 * qp->s_state is normally set to the opcode of the 568 * last packet constructed for new requests and therefore 569 * is never set to RDMA read response. 570 * RDMA_READ_RESPONSE_FIRST is used by the ACK processing 571 * thread to indicate a SEND needs to be restarted from an 572 * earlier PSN without interfering with the sending thread. 573 * See restart_rc(). 574 */ 575 qp->s_len = restart_sge(&qp->s_sge, wqe, qp->s_psn, pmtu); 576 /* FALLTHROUGH */ 577 case OP(SEND_FIRST): 578 qp->s_state = OP(SEND_MIDDLE); 579 /* FALLTHROUGH */ 580 case OP(SEND_MIDDLE): 581 bth2 = mask_psn(qp->s_psn++); 582 ss = &qp->s_sge; 583 len = qp->s_len; 584 if (len > pmtu) { 585 len = pmtu; 586 middle = HFI1_CAP_IS_KSET(SDMA_AHG); 587 break; 588 } 589 if (wqe->wr.opcode == IB_WR_SEND) { 590 qp->s_state = OP(SEND_LAST); 591 } else if (wqe->wr.opcode == IB_WR_SEND_WITH_IMM) { 592 qp->s_state = OP(SEND_LAST_WITH_IMMEDIATE); 593 /* Immediate data comes after the BTH */ 594 ohdr->u.imm_data = wqe->wr.ex.imm_data; 595 hwords += 1; 596 } else { 597 qp->s_state = OP(SEND_LAST_WITH_INVALIDATE); 598 /* invalidate data comes after the BTH */ 599 ohdr->u.ieth = cpu_to_be32(wqe->wr.ex.invalidate_rkey); 600 hwords += 1; 601 } 602 if (wqe->wr.send_flags & IB_SEND_SOLICITED) 603 bth0 |= IB_BTH_SOLICITED; 604 bth2 |= IB_BTH_REQ_ACK; 605 qp->s_cur++; 606 if (qp->s_cur >= qp->s_size) 607 qp->s_cur = 0; 608 break; 609 610 case OP(RDMA_READ_RESPONSE_LAST): 611 /* 612 * qp->s_state is normally set to the opcode of the 613 * last packet constructed for new requests and therefore 614 * is never set to RDMA read response. 615 * RDMA_READ_RESPONSE_LAST is used by the ACK processing 616 * thread to indicate a RDMA write needs to be restarted from 617 * an earlier PSN without interfering with the sending thread. 618 * See restart_rc(). 619 */ 620 qp->s_len = restart_sge(&qp->s_sge, wqe, qp->s_psn, pmtu); 621 /* FALLTHROUGH */ 622 case OP(RDMA_WRITE_FIRST): 623 qp->s_state = OP(RDMA_WRITE_MIDDLE); 624 /* FALLTHROUGH */ 625 case OP(RDMA_WRITE_MIDDLE): 626 bth2 = mask_psn(qp->s_psn++); 627 ss = &qp->s_sge; 628 len = qp->s_len; 629 if (len > pmtu) { 630 len = pmtu; 631 middle = HFI1_CAP_IS_KSET(SDMA_AHG); 632 break; 633 } 634 if (wqe->wr.opcode == IB_WR_RDMA_WRITE) { 635 qp->s_state = OP(RDMA_WRITE_LAST); 636 } else { 637 qp->s_state = OP(RDMA_WRITE_LAST_WITH_IMMEDIATE); 638 /* Immediate data comes after the BTH */ 639 ohdr->u.imm_data = wqe->wr.ex.imm_data; 640 hwords += 1; 641 if (wqe->wr.send_flags & IB_SEND_SOLICITED) 642 bth0 |= IB_BTH_SOLICITED; 643 } 644 bth2 |= IB_BTH_REQ_ACK; 645 qp->s_cur++; 646 if (qp->s_cur >= qp->s_size) 647 qp->s_cur = 0; 648 break; 649 650 case OP(RDMA_READ_RESPONSE_MIDDLE): 651 /* 652 * qp->s_state is normally set to the opcode of the 653 * last packet constructed for new requests and therefore 654 * is never set to RDMA read response. 655 * RDMA_READ_RESPONSE_MIDDLE is used by the ACK processing 656 * thread to indicate a RDMA read needs to be restarted from 657 * an earlier PSN without interfering with the sending thread. 658 * See restart_rc(). 659 */ 660 len = (delta_psn(qp->s_psn, wqe->psn)) * pmtu; 661 put_ib_reth_vaddr( 662 wqe->rdma_wr.remote_addr + len, 663 &ohdr->u.rc.reth); 664 ohdr->u.rc.reth.rkey = 665 cpu_to_be32(wqe->rdma_wr.rkey); 666 ohdr->u.rc.reth.length = cpu_to_be32(wqe->length - len); 667 qp->s_state = OP(RDMA_READ_REQUEST); 668 hwords += sizeof(ohdr->u.rc.reth) / sizeof(u32); 669 bth2 = mask_psn(qp->s_psn) | IB_BTH_REQ_ACK; 670 qp->s_psn = wqe->lpsn + 1; 671 ss = NULL; 672 len = 0; 673 qp->s_cur++; 674 if (qp->s_cur == qp->s_size) 675 qp->s_cur = 0; 676 break; 677 } 678 qp->s_sending_hpsn = bth2; 679 delta = delta_psn(bth2, wqe->psn); 680 if (delta && delta % HFI1_PSN_CREDIT == 0) 681 bth2 |= IB_BTH_REQ_ACK; 682 if (qp->s_flags & RVT_S_SEND_ONE) { 683 qp->s_flags &= ~RVT_S_SEND_ONE; 684 qp->s_flags |= RVT_S_WAIT_ACK; 685 bth2 |= IB_BTH_REQ_ACK; 686 } 687 qp->s_len -= len; 688 ps->s_txreq->hdr_dwords = hwords; 689 ps->s_txreq->sde = priv->s_sde; 690 ps->s_txreq->ss = ss; 691 ps->s_txreq->s_cur_size = len; 692 hfi1_make_ruc_header( 693 qp, 694 ohdr, 695 bth0 | (qp->s_state << 24), 696 bth2, 697 middle, 698 ps); 699 return 1; 700 701 done_free_tx: 702 hfi1_put_txreq(ps->s_txreq); 703 ps->s_txreq = NULL; 704 return 1; 705 706 bail: 707 hfi1_put_txreq(ps->s_txreq); 708 709 bail_no_tx: 710 ps->s_txreq = NULL; 711 qp->s_flags &= ~RVT_S_BUSY; 712 return 0; 713 } 714 715 static inline void hfi1_make_bth_aeth(struct rvt_qp *qp, 716 struct ib_other_headers *ohdr, 717 u32 bth0, u32 bth1) 718 { 719 if (qp->r_nak_state) 720 ohdr->u.aeth = cpu_to_be32((qp->r_msn & IB_MSN_MASK) | 721 (qp->r_nak_state << 722 IB_AETH_CREDIT_SHIFT)); 723 else 724 ohdr->u.aeth = rvt_compute_aeth(qp); 725 726 ohdr->bth[0] = cpu_to_be32(bth0); 727 ohdr->bth[1] = cpu_to_be32(bth1 | qp->remote_qpn); 728 ohdr->bth[2] = cpu_to_be32(mask_psn(qp->r_ack_psn)); 729 } 730 731 static inline void hfi1_queue_rc_ack(struct hfi1_packet *packet, bool is_fecn) 732 { 733 struct rvt_qp *qp = packet->qp; 734 struct hfi1_ibport *ibp; 735 unsigned long flags; 736 737 spin_lock_irqsave(&qp->s_lock, flags); 738 if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK)) 739 goto unlock; 740 ibp = rcd_to_iport(packet->rcd); 741 this_cpu_inc(*ibp->rvp.rc_qacks); 742 qp->s_flags |= RVT_S_ACK_PENDING | RVT_S_RESP_PENDING; 743 qp->s_nak_state = qp->r_nak_state; 744 qp->s_ack_psn = qp->r_ack_psn; 745 if (is_fecn) 746 qp->s_flags |= RVT_S_ECN; 747 748 /* Schedule the send tasklet. */ 749 hfi1_schedule_send(qp); 750 unlock: 751 spin_unlock_irqrestore(&qp->s_lock, flags); 752 } 753 754 static inline void hfi1_make_rc_ack_9B(struct hfi1_packet *packet, 755 struct hfi1_opa_header *opa_hdr, 756 u8 sc5, bool is_fecn, 757 u64 *pbc_flags, u32 *hwords, 758 u32 *nwords) 759 { 760 struct rvt_qp *qp = packet->qp; 761 struct hfi1_ibport *ibp = rcd_to_iport(packet->rcd); 762 struct hfi1_pportdata *ppd = ppd_from_ibp(ibp); 763 struct ib_header *hdr = &opa_hdr->ibh; 764 struct ib_other_headers *ohdr; 765 u16 lrh0 = HFI1_LRH_BTH; 766 u16 pkey; 767 u32 bth0, bth1; 768 769 opa_hdr->hdr_type = HFI1_PKT_TYPE_9B; 770 ohdr = &hdr->u.oth; 771 /* header size in 32-bit words LRH+BTH+AETH = (8+12+4)/4 */ 772 *hwords = 6; 773 774 if (unlikely(rdma_ah_get_ah_flags(&qp->remote_ah_attr) & IB_AH_GRH)) { 775 *hwords += hfi1_make_grh(ibp, &hdr->u.l.grh, 776 rdma_ah_read_grh(&qp->remote_ah_attr), 777 *hwords - 2, SIZE_OF_CRC); 778 ohdr = &hdr->u.l.oth; 779 lrh0 = HFI1_LRH_GRH; 780 } 781 /* set PBC_DC_INFO bit (aka SC[4]) in pbc_flags */ 782 *pbc_flags |= ((!!(sc5 & 0x10)) << PBC_DC_INFO_SHIFT); 783 784 /* read pkey_index w/o lock (its atomic) */ 785 pkey = hfi1_get_pkey(ibp, qp->s_pkey_index); 786 787 lrh0 |= (sc5 & IB_SC_MASK) << IB_SC_SHIFT | 788 (rdma_ah_get_sl(&qp->remote_ah_attr) & IB_SL_MASK) << 789 IB_SL_SHIFT; 790 791 hfi1_make_ib_hdr(hdr, lrh0, *hwords + SIZE_OF_CRC, 792 opa_get_lid(rdma_ah_get_dlid(&qp->remote_ah_attr), 9B), 793 ppd->lid | rdma_ah_get_path_bits(&qp->remote_ah_attr)); 794 795 bth0 = pkey | (OP(ACKNOWLEDGE) << 24); 796 if (qp->s_mig_state == IB_MIG_MIGRATED) 797 bth0 |= IB_BTH_MIG_REQ; 798 bth1 = (!!is_fecn) << IB_BECN_SHIFT; 799 hfi1_make_bth_aeth(qp, ohdr, bth0, bth1); 800 } 801 802 static inline void hfi1_make_rc_ack_16B(struct hfi1_packet *packet, 803 struct hfi1_opa_header *opa_hdr, 804 u8 sc5, bool is_fecn, 805 u64 *pbc_flags, u32 *hwords, 806 u32 *nwords) 807 { 808 struct rvt_qp *qp = packet->qp; 809 struct hfi1_ibport *ibp = rcd_to_iport(packet->rcd); 810 struct hfi1_pportdata *ppd = ppd_from_ibp(ibp); 811 struct hfi1_16b_header *hdr = &opa_hdr->opah; 812 struct ib_other_headers *ohdr; 813 u32 bth0, bth1 = 0; 814 u16 len, pkey; 815 bool becn = is_fecn; 816 u8 l4 = OPA_16B_L4_IB_LOCAL; 817 u8 extra_bytes; 818 819 opa_hdr->hdr_type = HFI1_PKT_TYPE_16B; 820 ohdr = &hdr->u.oth; 821 /* header size in 32-bit words 16B LRH+BTH+AETH = (16+12+4)/4 */ 822 *hwords = 8; 823 extra_bytes = hfi1_get_16b_padding(*hwords << 2, 0); 824 *nwords = SIZE_OF_CRC + ((extra_bytes + SIZE_OF_LT) >> 2); 825 826 if (unlikely(rdma_ah_get_ah_flags(&qp->remote_ah_attr) & IB_AH_GRH) && 827 hfi1_check_mcast(rdma_ah_get_dlid(&qp->remote_ah_attr))) { 828 *hwords += hfi1_make_grh(ibp, &hdr->u.l.grh, 829 rdma_ah_read_grh(&qp->remote_ah_attr), 830 *hwords - 4, *nwords); 831 ohdr = &hdr->u.l.oth; 832 l4 = OPA_16B_L4_IB_GLOBAL; 833 } 834 *pbc_flags |= PBC_PACKET_BYPASS | PBC_INSERT_BYPASS_ICRC; 835 836 /* read pkey_index w/o lock (its atomic) */ 837 pkey = hfi1_get_pkey(ibp, qp->s_pkey_index); 838 839 /* Convert dwords to flits */ 840 len = (*hwords + *nwords) >> 1; 841 842 hfi1_make_16b_hdr(hdr, ppd->lid | 843 (rdma_ah_get_path_bits(&qp->remote_ah_attr) & 844 ((1 << ppd->lmc) - 1)), 845 opa_get_lid(rdma_ah_get_dlid(&qp->remote_ah_attr), 846 16B), len, pkey, becn, 0, l4, sc5); 847 848 bth0 = pkey | (OP(ACKNOWLEDGE) << 24); 849 bth0 |= extra_bytes << 20; 850 if (qp->s_mig_state == IB_MIG_MIGRATED) 851 bth1 = OPA_BTH_MIG_REQ; 852 hfi1_make_bth_aeth(qp, ohdr, bth0, bth1); 853 } 854 855 typedef void (*hfi1_make_rc_ack)(struct hfi1_packet *packet, 856 struct hfi1_opa_header *opa_hdr, 857 u8 sc5, bool is_fecn, 858 u64 *pbc_flags, u32 *hwords, 859 u32 *nwords); 860 861 /* We support only two types - 9B and 16B for now */ 862 static const hfi1_make_rc_ack hfi1_make_rc_ack_tbl[2] = { 863 [HFI1_PKT_TYPE_9B] = &hfi1_make_rc_ack_9B, 864 [HFI1_PKT_TYPE_16B] = &hfi1_make_rc_ack_16B 865 }; 866 867 /** 868 * hfi1_send_rc_ack - Construct an ACK packet and send it 869 * @qp: a pointer to the QP 870 * 871 * This is called from hfi1_rc_rcv() and handle_receive_interrupt(). 872 * Note that RDMA reads and atomics are handled in the 873 * send side QP state and send engine. 874 */ 875 void hfi1_send_rc_ack(struct hfi1_packet *packet, bool is_fecn) 876 { 877 struct hfi1_ctxtdata *rcd = packet->rcd; 878 struct rvt_qp *qp = packet->qp; 879 struct hfi1_ibport *ibp = rcd_to_iport(rcd); 880 struct hfi1_qp_priv *priv = qp->priv; 881 struct hfi1_pportdata *ppd = ppd_from_ibp(ibp); 882 u8 sc5 = ibp->sl_to_sc[rdma_ah_get_sl(&qp->remote_ah_attr)]; 883 u64 pbc, pbc_flags = 0; 884 u32 hwords = 0; 885 u32 nwords = 0; 886 u32 plen; 887 struct pio_buf *pbuf; 888 struct hfi1_opa_header opa_hdr; 889 890 /* clear the defer count */ 891 qp->r_adefered = 0; 892 893 /* Don't send ACK or NAK if a RDMA read or atomic is pending. */ 894 if (qp->s_flags & RVT_S_RESP_PENDING) { 895 hfi1_queue_rc_ack(packet, is_fecn); 896 return; 897 } 898 899 /* Ensure s_rdma_ack_cnt changes are committed */ 900 if (qp->s_rdma_ack_cnt) { 901 hfi1_queue_rc_ack(packet, is_fecn); 902 return; 903 } 904 905 /* Don't try to send ACKs if the link isn't ACTIVE */ 906 if (driver_lstate(ppd) != IB_PORT_ACTIVE) 907 return; 908 909 /* Make the appropriate header */ 910 hfi1_make_rc_ack_tbl[priv->hdr_type](packet, &opa_hdr, sc5, is_fecn, 911 &pbc_flags, &hwords, &nwords); 912 913 plen = 2 /* PBC */ + hwords + nwords; 914 pbc = create_pbc(ppd, pbc_flags, qp->srate_mbps, 915 sc_to_vlt(ppd->dd, sc5), plen); 916 pbuf = sc_buffer_alloc(rcd->sc, plen, NULL, NULL); 917 if (!pbuf) { 918 /* 919 * We have no room to send at the moment. Pass 920 * responsibility for sending the ACK to the send engine 921 * so that when enough buffer space becomes available, 922 * the ACK is sent ahead of other outgoing packets. 923 */ 924 hfi1_queue_rc_ack(packet, is_fecn); 925 return; 926 } 927 trace_ack_output_ibhdr(dd_from_ibdev(qp->ibqp.device), 928 &opa_hdr, ib_is_sc5(sc5)); 929 930 /* write the pbc and data */ 931 ppd->dd->pio_inline_send(ppd->dd, pbuf, pbc, 932 (priv->hdr_type == HFI1_PKT_TYPE_9B ? 933 (void *)&opa_hdr.ibh : 934 (void *)&opa_hdr.opah), hwords); 935 return; 936 } 937 938 /** 939 * reset_psn - reset the QP state to send starting from PSN 940 * @qp: the QP 941 * @psn: the packet sequence number to restart at 942 * 943 * This is called from hfi1_rc_rcv() to process an incoming RC ACK 944 * for the given QP. 945 * Called at interrupt level with the QP s_lock held. 946 */ 947 static void reset_psn(struct rvt_qp *qp, u32 psn) 948 { 949 u32 n = qp->s_acked; 950 struct rvt_swqe *wqe = rvt_get_swqe_ptr(qp, n); 951 u32 opcode; 952 953 lockdep_assert_held(&qp->s_lock); 954 qp->s_cur = n; 955 956 /* 957 * If we are starting the request from the beginning, 958 * let the normal send code handle initialization. 959 */ 960 if (cmp_psn(psn, wqe->psn) <= 0) { 961 qp->s_state = OP(SEND_LAST); 962 goto done; 963 } 964 965 /* Find the work request opcode corresponding to the given PSN. */ 966 opcode = wqe->wr.opcode; 967 for (;;) { 968 int diff; 969 970 if (++n == qp->s_size) 971 n = 0; 972 if (n == qp->s_tail) 973 break; 974 wqe = rvt_get_swqe_ptr(qp, n); 975 diff = cmp_psn(psn, wqe->psn); 976 if (diff < 0) 977 break; 978 qp->s_cur = n; 979 /* 980 * If we are starting the request from the beginning, 981 * let the normal send code handle initialization. 982 */ 983 if (diff == 0) { 984 qp->s_state = OP(SEND_LAST); 985 goto done; 986 } 987 opcode = wqe->wr.opcode; 988 } 989 990 /* 991 * Set the state to restart in the middle of a request. 992 * Don't change the s_sge, s_cur_sge, or s_cur_size. 993 * See hfi1_make_rc_req(). 994 */ 995 switch (opcode) { 996 case IB_WR_SEND: 997 case IB_WR_SEND_WITH_IMM: 998 qp->s_state = OP(RDMA_READ_RESPONSE_FIRST); 999 break; 1000 1001 case IB_WR_RDMA_WRITE: 1002 case IB_WR_RDMA_WRITE_WITH_IMM: 1003 qp->s_state = OP(RDMA_READ_RESPONSE_LAST); 1004 break; 1005 1006 case IB_WR_RDMA_READ: 1007 qp->s_state = OP(RDMA_READ_RESPONSE_MIDDLE); 1008 break; 1009 1010 default: 1011 /* 1012 * This case shouldn't happen since its only 1013 * one PSN per req. 1014 */ 1015 qp->s_state = OP(SEND_LAST); 1016 } 1017 done: 1018 qp->s_psn = psn; 1019 /* 1020 * Set RVT_S_WAIT_PSN as rc_complete() may start the timer 1021 * asynchronously before the send engine can get scheduled. 1022 * Doing it in hfi1_make_rc_req() is too late. 1023 */ 1024 if ((cmp_psn(qp->s_psn, qp->s_sending_hpsn) <= 0) && 1025 (cmp_psn(qp->s_sending_psn, qp->s_sending_hpsn) <= 0)) 1026 qp->s_flags |= RVT_S_WAIT_PSN; 1027 qp->s_flags &= ~HFI1_S_AHG_VALID; 1028 } 1029 1030 /* 1031 * Back up requester to resend the last un-ACKed request. 1032 * The QP r_lock and s_lock should be held and interrupts disabled. 1033 */ 1034 void hfi1_restart_rc(struct rvt_qp *qp, u32 psn, int wait) 1035 { 1036 struct rvt_swqe *wqe = rvt_get_swqe_ptr(qp, qp->s_acked); 1037 struct hfi1_ibport *ibp; 1038 1039 lockdep_assert_held(&qp->r_lock); 1040 lockdep_assert_held(&qp->s_lock); 1041 if (qp->s_retry == 0) { 1042 if (qp->s_mig_state == IB_MIG_ARMED) { 1043 hfi1_migrate_qp(qp); 1044 qp->s_retry = qp->s_retry_cnt; 1045 } else if (qp->s_last == qp->s_acked) { 1046 rvt_send_complete(qp, wqe, IB_WC_RETRY_EXC_ERR); 1047 rvt_error_qp(qp, IB_WC_WR_FLUSH_ERR); 1048 return; 1049 } else { /* need to handle delayed completion */ 1050 return; 1051 } 1052 } else { 1053 qp->s_retry--; 1054 } 1055 1056 ibp = to_iport(qp->ibqp.device, qp->port_num); 1057 if (wqe->wr.opcode == IB_WR_RDMA_READ) 1058 ibp->rvp.n_rc_resends++; 1059 else 1060 ibp->rvp.n_rc_resends += delta_psn(qp->s_psn, psn); 1061 1062 qp->s_flags &= ~(RVT_S_WAIT_FENCE | RVT_S_WAIT_RDMAR | 1063 RVT_S_WAIT_SSN_CREDIT | RVT_S_WAIT_PSN | 1064 RVT_S_WAIT_ACK); 1065 if (wait) 1066 qp->s_flags |= RVT_S_SEND_ONE; 1067 reset_psn(qp, psn); 1068 } 1069 1070 /* 1071 * Set qp->s_sending_psn to the next PSN after the given one. 1072 * This would be psn+1 except when RDMA reads are present. 1073 */ 1074 static void reset_sending_psn(struct rvt_qp *qp, u32 psn) 1075 { 1076 struct rvt_swqe *wqe; 1077 u32 n = qp->s_last; 1078 1079 lockdep_assert_held(&qp->s_lock); 1080 /* Find the work request corresponding to the given PSN. */ 1081 for (;;) { 1082 wqe = rvt_get_swqe_ptr(qp, n); 1083 if (cmp_psn(psn, wqe->lpsn) <= 0) { 1084 if (wqe->wr.opcode == IB_WR_RDMA_READ) 1085 qp->s_sending_psn = wqe->lpsn + 1; 1086 else 1087 qp->s_sending_psn = psn + 1; 1088 break; 1089 } 1090 if (++n == qp->s_size) 1091 n = 0; 1092 if (n == qp->s_tail) 1093 break; 1094 } 1095 } 1096 1097 /* 1098 * This should be called with the QP s_lock held and interrupts disabled. 1099 */ 1100 void hfi1_rc_send_complete(struct rvt_qp *qp, struct hfi1_opa_header *opah) 1101 { 1102 struct ib_other_headers *ohdr; 1103 struct hfi1_qp_priv *priv = qp->priv; 1104 struct rvt_swqe *wqe; 1105 struct ib_header *hdr = NULL; 1106 struct hfi1_16b_header *hdr_16b = NULL; 1107 u32 opcode; 1108 u32 psn; 1109 1110 lockdep_assert_held(&qp->s_lock); 1111 if (!(ib_rvt_state_ops[qp->state] & RVT_SEND_OR_FLUSH_OR_RECV_OK)) 1112 return; 1113 1114 /* Find out where the BTH is */ 1115 if (priv->hdr_type == HFI1_PKT_TYPE_9B) { 1116 hdr = &opah->ibh; 1117 if (ib_get_lnh(hdr) == HFI1_LRH_BTH) 1118 ohdr = &hdr->u.oth; 1119 else 1120 ohdr = &hdr->u.l.oth; 1121 } else { 1122 u8 l4; 1123 1124 hdr_16b = &opah->opah; 1125 l4 = hfi1_16B_get_l4(hdr_16b); 1126 if (l4 == OPA_16B_L4_IB_LOCAL) 1127 ohdr = &hdr_16b->u.oth; 1128 else 1129 ohdr = &hdr_16b->u.l.oth; 1130 } 1131 1132 opcode = ib_bth_get_opcode(ohdr); 1133 if (opcode >= OP(RDMA_READ_RESPONSE_FIRST) && 1134 opcode <= OP(ATOMIC_ACKNOWLEDGE)) { 1135 WARN_ON(!qp->s_rdma_ack_cnt); 1136 qp->s_rdma_ack_cnt--; 1137 return; 1138 } 1139 1140 psn = ib_bth_get_psn(ohdr); 1141 reset_sending_psn(qp, psn); 1142 1143 /* 1144 * Start timer after a packet requesting an ACK has been sent and 1145 * there are still requests that haven't been acked. 1146 */ 1147 if ((psn & IB_BTH_REQ_ACK) && qp->s_acked != qp->s_tail && 1148 !(qp->s_flags & 1149 (RVT_S_TIMER | RVT_S_WAIT_RNR | RVT_S_WAIT_PSN)) && 1150 (ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK)) 1151 rvt_add_retry_timer(qp); 1152 1153 while (qp->s_last != qp->s_acked) { 1154 u32 s_last; 1155 1156 wqe = rvt_get_swqe_ptr(qp, qp->s_last); 1157 if (cmp_psn(wqe->lpsn, qp->s_sending_psn) >= 0 && 1158 cmp_psn(qp->s_sending_psn, qp->s_sending_hpsn) <= 0) 1159 break; 1160 rvt_qp_wqe_unreserve(qp, wqe); 1161 s_last = qp->s_last; 1162 trace_hfi1_qp_send_completion(qp, wqe, s_last); 1163 if (++s_last >= qp->s_size) 1164 s_last = 0; 1165 qp->s_last = s_last; 1166 /* see post_send() */ 1167 barrier(); 1168 rvt_put_swqe(wqe); 1169 rvt_qp_swqe_complete(qp, 1170 wqe, 1171 ib_hfi1_wc_opcode[wqe->wr.opcode], 1172 IB_WC_SUCCESS); 1173 } 1174 /* 1175 * If we were waiting for sends to complete before re-sending, 1176 * and they are now complete, restart sending. 1177 */ 1178 trace_hfi1_sendcomplete(qp, psn); 1179 if (qp->s_flags & RVT_S_WAIT_PSN && 1180 cmp_psn(qp->s_sending_psn, qp->s_sending_hpsn) > 0) { 1181 qp->s_flags &= ~RVT_S_WAIT_PSN; 1182 qp->s_sending_psn = qp->s_psn; 1183 qp->s_sending_hpsn = qp->s_psn - 1; 1184 hfi1_schedule_send(qp); 1185 } 1186 } 1187 1188 static inline void update_last_psn(struct rvt_qp *qp, u32 psn) 1189 { 1190 qp->s_last_psn = psn; 1191 } 1192 1193 /* 1194 * Generate a SWQE completion. 1195 * This is similar to hfi1_send_complete but has to check to be sure 1196 * that the SGEs are not being referenced if the SWQE is being resent. 1197 */ 1198 static struct rvt_swqe *do_rc_completion(struct rvt_qp *qp, 1199 struct rvt_swqe *wqe, 1200 struct hfi1_ibport *ibp) 1201 { 1202 lockdep_assert_held(&qp->s_lock); 1203 /* 1204 * Don't decrement refcount and don't generate a 1205 * completion if the SWQE is being resent until the send 1206 * is finished. 1207 */ 1208 if (cmp_psn(wqe->lpsn, qp->s_sending_psn) < 0 || 1209 cmp_psn(qp->s_sending_psn, qp->s_sending_hpsn) > 0) { 1210 u32 s_last; 1211 1212 rvt_put_swqe(wqe); 1213 rvt_qp_wqe_unreserve(qp, wqe); 1214 s_last = qp->s_last; 1215 trace_hfi1_qp_send_completion(qp, wqe, s_last); 1216 if (++s_last >= qp->s_size) 1217 s_last = 0; 1218 qp->s_last = s_last; 1219 /* see post_send() */ 1220 barrier(); 1221 rvt_qp_swqe_complete(qp, 1222 wqe, 1223 ib_hfi1_wc_opcode[wqe->wr.opcode], 1224 IB_WC_SUCCESS); 1225 } else { 1226 struct hfi1_pportdata *ppd = ppd_from_ibp(ibp); 1227 1228 this_cpu_inc(*ibp->rvp.rc_delayed_comp); 1229 /* 1230 * If send progress not running attempt to progress 1231 * SDMA queue. 1232 */ 1233 if (ppd->dd->flags & HFI1_HAS_SEND_DMA) { 1234 struct sdma_engine *engine; 1235 u8 sl = rdma_ah_get_sl(&qp->remote_ah_attr); 1236 u8 sc5; 1237 1238 /* For now use sc to find engine */ 1239 sc5 = ibp->sl_to_sc[sl]; 1240 engine = qp_to_sdma_engine(qp, sc5); 1241 sdma_engine_progress_schedule(engine); 1242 } 1243 } 1244 1245 qp->s_retry = qp->s_retry_cnt; 1246 update_last_psn(qp, wqe->lpsn); 1247 1248 /* 1249 * If we are completing a request which is in the process of 1250 * being resent, we can stop re-sending it since we know the 1251 * responder has already seen it. 1252 */ 1253 if (qp->s_acked == qp->s_cur) { 1254 if (++qp->s_cur >= qp->s_size) 1255 qp->s_cur = 0; 1256 qp->s_acked = qp->s_cur; 1257 wqe = rvt_get_swqe_ptr(qp, qp->s_cur); 1258 if (qp->s_acked != qp->s_tail) { 1259 qp->s_state = OP(SEND_LAST); 1260 qp->s_psn = wqe->psn; 1261 } 1262 } else { 1263 if (++qp->s_acked >= qp->s_size) 1264 qp->s_acked = 0; 1265 if (qp->state == IB_QPS_SQD && qp->s_acked == qp->s_cur) 1266 qp->s_draining = 0; 1267 wqe = rvt_get_swqe_ptr(qp, qp->s_acked); 1268 } 1269 return wqe; 1270 } 1271 1272 /** 1273 * do_rc_ack - process an incoming RC ACK 1274 * @qp: the QP the ACK came in on 1275 * @psn: the packet sequence number of the ACK 1276 * @opcode: the opcode of the request that resulted in the ACK 1277 * 1278 * This is called from rc_rcv_resp() to process an incoming RC ACK 1279 * for the given QP. 1280 * May be called at interrupt level, with the QP s_lock held. 1281 * Returns 1 if OK, 0 if current operation should be aborted (NAK). 1282 */ 1283 static int do_rc_ack(struct rvt_qp *qp, u32 aeth, u32 psn, int opcode, 1284 u64 val, struct hfi1_ctxtdata *rcd) 1285 { 1286 struct hfi1_ibport *ibp; 1287 enum ib_wc_status status; 1288 struct rvt_swqe *wqe; 1289 int ret = 0; 1290 u32 ack_psn; 1291 int diff; 1292 1293 lockdep_assert_held(&qp->s_lock); 1294 /* 1295 * Note that NAKs implicitly ACK outstanding SEND and RDMA write 1296 * requests and implicitly NAK RDMA read and atomic requests issued 1297 * before the NAK'ed request. The MSN won't include the NAK'ed 1298 * request but will include an ACK'ed request(s). 1299 */ 1300 ack_psn = psn; 1301 if (aeth >> IB_AETH_NAK_SHIFT) 1302 ack_psn--; 1303 wqe = rvt_get_swqe_ptr(qp, qp->s_acked); 1304 ibp = rcd_to_iport(rcd); 1305 1306 /* 1307 * The MSN might be for a later WQE than the PSN indicates so 1308 * only complete WQEs that the PSN finishes. 1309 */ 1310 while ((diff = delta_psn(ack_psn, wqe->lpsn)) >= 0) { 1311 /* 1312 * RDMA_READ_RESPONSE_ONLY is a special case since 1313 * we want to generate completion events for everything 1314 * before the RDMA read, copy the data, then generate 1315 * the completion for the read. 1316 */ 1317 if (wqe->wr.opcode == IB_WR_RDMA_READ && 1318 opcode == OP(RDMA_READ_RESPONSE_ONLY) && 1319 diff == 0) { 1320 ret = 1; 1321 goto bail_stop; 1322 } 1323 /* 1324 * If this request is a RDMA read or atomic, and the ACK is 1325 * for a later operation, this ACK NAKs the RDMA read or 1326 * atomic. In other words, only a RDMA_READ_LAST or ONLY 1327 * can ACK a RDMA read and likewise for atomic ops. Note 1328 * that the NAK case can only happen if relaxed ordering is 1329 * used and requests are sent after an RDMA read or atomic 1330 * is sent but before the response is received. 1331 */ 1332 if ((wqe->wr.opcode == IB_WR_RDMA_READ && 1333 (opcode != OP(RDMA_READ_RESPONSE_LAST) || diff != 0)) || 1334 ((wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP || 1335 wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD) && 1336 (opcode != OP(ATOMIC_ACKNOWLEDGE) || diff != 0))) { 1337 /* Retry this request. */ 1338 if (!(qp->r_flags & RVT_R_RDMAR_SEQ)) { 1339 qp->r_flags |= RVT_R_RDMAR_SEQ; 1340 hfi1_restart_rc(qp, qp->s_last_psn + 1, 0); 1341 if (list_empty(&qp->rspwait)) { 1342 qp->r_flags |= RVT_R_RSP_SEND; 1343 rvt_get_qp(qp); 1344 list_add_tail(&qp->rspwait, 1345 &rcd->qp_wait_list); 1346 } 1347 } 1348 /* 1349 * No need to process the ACK/NAK since we are 1350 * restarting an earlier request. 1351 */ 1352 goto bail_stop; 1353 } 1354 if (wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP || 1355 wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD) { 1356 u64 *vaddr = wqe->sg_list[0].vaddr; 1357 *vaddr = val; 1358 } 1359 if (qp->s_num_rd_atomic && 1360 (wqe->wr.opcode == IB_WR_RDMA_READ || 1361 wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP || 1362 wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD)) { 1363 qp->s_num_rd_atomic--; 1364 /* Restart sending task if fence is complete */ 1365 if ((qp->s_flags & RVT_S_WAIT_FENCE) && 1366 !qp->s_num_rd_atomic) { 1367 qp->s_flags &= ~(RVT_S_WAIT_FENCE | 1368 RVT_S_WAIT_ACK); 1369 hfi1_schedule_send(qp); 1370 } else if (qp->s_flags & RVT_S_WAIT_RDMAR) { 1371 qp->s_flags &= ~(RVT_S_WAIT_RDMAR | 1372 RVT_S_WAIT_ACK); 1373 hfi1_schedule_send(qp); 1374 } 1375 } 1376 wqe = do_rc_completion(qp, wqe, ibp); 1377 if (qp->s_acked == qp->s_tail) 1378 break; 1379 } 1380 1381 switch (aeth >> IB_AETH_NAK_SHIFT) { 1382 case 0: /* ACK */ 1383 this_cpu_inc(*ibp->rvp.rc_acks); 1384 if (qp->s_acked != qp->s_tail) { 1385 /* 1386 * We are expecting more ACKs so 1387 * mod the retry timer. 1388 */ 1389 rvt_mod_retry_timer(qp); 1390 /* 1391 * We can stop re-sending the earlier packets and 1392 * continue with the next packet the receiver wants. 1393 */ 1394 if (cmp_psn(qp->s_psn, psn) <= 0) 1395 reset_psn(qp, psn + 1); 1396 } else { 1397 /* No more acks - kill all timers */ 1398 rvt_stop_rc_timers(qp); 1399 if (cmp_psn(qp->s_psn, psn) <= 0) { 1400 qp->s_state = OP(SEND_LAST); 1401 qp->s_psn = psn + 1; 1402 } 1403 } 1404 if (qp->s_flags & RVT_S_WAIT_ACK) { 1405 qp->s_flags &= ~RVT_S_WAIT_ACK; 1406 hfi1_schedule_send(qp); 1407 } 1408 rvt_get_credit(qp, aeth); 1409 qp->s_rnr_retry = qp->s_rnr_retry_cnt; 1410 qp->s_retry = qp->s_retry_cnt; 1411 update_last_psn(qp, psn); 1412 return 1; 1413 1414 case 1: /* RNR NAK */ 1415 ibp->rvp.n_rnr_naks++; 1416 if (qp->s_acked == qp->s_tail) 1417 goto bail_stop; 1418 if (qp->s_flags & RVT_S_WAIT_RNR) 1419 goto bail_stop; 1420 if (qp->s_rnr_retry == 0) { 1421 status = IB_WC_RNR_RETRY_EXC_ERR; 1422 goto class_b; 1423 } 1424 if (qp->s_rnr_retry_cnt < 7) 1425 qp->s_rnr_retry--; 1426 1427 /* The last valid PSN is the previous PSN. */ 1428 update_last_psn(qp, psn - 1); 1429 1430 ibp->rvp.n_rc_resends += delta_psn(qp->s_psn, psn); 1431 1432 reset_psn(qp, psn); 1433 1434 qp->s_flags &= ~(RVT_S_WAIT_SSN_CREDIT | RVT_S_WAIT_ACK); 1435 rvt_stop_rc_timers(qp); 1436 rvt_add_rnr_timer(qp, aeth); 1437 return 0; 1438 1439 case 3: /* NAK */ 1440 if (qp->s_acked == qp->s_tail) 1441 goto bail_stop; 1442 /* The last valid PSN is the previous PSN. */ 1443 update_last_psn(qp, psn - 1); 1444 switch ((aeth >> IB_AETH_CREDIT_SHIFT) & 1445 IB_AETH_CREDIT_MASK) { 1446 case 0: /* PSN sequence error */ 1447 ibp->rvp.n_seq_naks++; 1448 /* 1449 * Back up to the responder's expected PSN. 1450 * Note that we might get a NAK in the middle of an 1451 * RDMA READ response which terminates the RDMA 1452 * READ. 1453 */ 1454 hfi1_restart_rc(qp, psn, 0); 1455 hfi1_schedule_send(qp); 1456 break; 1457 1458 case 1: /* Invalid Request */ 1459 status = IB_WC_REM_INV_REQ_ERR; 1460 ibp->rvp.n_other_naks++; 1461 goto class_b; 1462 1463 case 2: /* Remote Access Error */ 1464 status = IB_WC_REM_ACCESS_ERR; 1465 ibp->rvp.n_other_naks++; 1466 goto class_b; 1467 1468 case 3: /* Remote Operation Error */ 1469 status = IB_WC_REM_OP_ERR; 1470 ibp->rvp.n_other_naks++; 1471 class_b: 1472 if (qp->s_last == qp->s_acked) { 1473 rvt_send_complete(qp, wqe, status); 1474 rvt_error_qp(qp, IB_WC_WR_FLUSH_ERR); 1475 } 1476 break; 1477 1478 default: 1479 /* Ignore other reserved NAK error codes */ 1480 goto reserved; 1481 } 1482 qp->s_retry = qp->s_retry_cnt; 1483 qp->s_rnr_retry = qp->s_rnr_retry_cnt; 1484 goto bail_stop; 1485 1486 default: /* 2: reserved */ 1487 reserved: 1488 /* Ignore reserved NAK codes. */ 1489 goto bail_stop; 1490 } 1491 /* cannot be reached */ 1492 bail_stop: 1493 rvt_stop_rc_timers(qp); 1494 return ret; 1495 } 1496 1497 /* 1498 * We have seen an out of sequence RDMA read middle or last packet. 1499 * This ACKs SENDs and RDMA writes up to the first RDMA read or atomic SWQE. 1500 */ 1501 static void rdma_seq_err(struct rvt_qp *qp, struct hfi1_ibport *ibp, u32 psn, 1502 struct hfi1_ctxtdata *rcd) 1503 { 1504 struct rvt_swqe *wqe; 1505 1506 lockdep_assert_held(&qp->s_lock); 1507 /* Remove QP from retry timer */ 1508 rvt_stop_rc_timers(qp); 1509 1510 wqe = rvt_get_swqe_ptr(qp, qp->s_acked); 1511 1512 while (cmp_psn(psn, wqe->lpsn) > 0) { 1513 if (wqe->wr.opcode == IB_WR_RDMA_READ || 1514 wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP || 1515 wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD) 1516 break; 1517 wqe = do_rc_completion(qp, wqe, ibp); 1518 } 1519 1520 ibp->rvp.n_rdma_seq++; 1521 qp->r_flags |= RVT_R_RDMAR_SEQ; 1522 hfi1_restart_rc(qp, qp->s_last_psn + 1, 0); 1523 if (list_empty(&qp->rspwait)) { 1524 qp->r_flags |= RVT_R_RSP_SEND; 1525 rvt_get_qp(qp); 1526 list_add_tail(&qp->rspwait, &rcd->qp_wait_list); 1527 } 1528 } 1529 1530 /** 1531 * rc_rcv_resp - process an incoming RC response packet 1532 * @packet: data packet information 1533 * 1534 * This is called from hfi1_rc_rcv() to process an incoming RC response 1535 * packet for the given QP. 1536 * Called at interrupt level. 1537 */ 1538 static void rc_rcv_resp(struct hfi1_packet *packet) 1539 { 1540 struct hfi1_ctxtdata *rcd = packet->rcd; 1541 void *data = packet->payload; 1542 u32 tlen = packet->tlen; 1543 struct rvt_qp *qp = packet->qp; 1544 struct hfi1_ibport *ibp; 1545 struct ib_other_headers *ohdr = packet->ohdr; 1546 struct rvt_swqe *wqe; 1547 enum ib_wc_status status; 1548 unsigned long flags; 1549 int diff; 1550 u64 val; 1551 u32 aeth; 1552 u32 psn = ib_bth_get_psn(packet->ohdr); 1553 u32 pmtu = qp->pmtu; 1554 u16 hdrsize = packet->hlen; 1555 u8 opcode = packet->opcode; 1556 u8 pad = packet->pad; 1557 u8 extra_bytes = pad + packet->extra_byte + (SIZE_OF_CRC << 2); 1558 1559 spin_lock_irqsave(&qp->s_lock, flags); 1560 trace_hfi1_ack(qp, psn); 1561 1562 /* Ignore invalid responses. */ 1563 if (cmp_psn(psn, READ_ONCE(qp->s_next_psn)) >= 0) 1564 goto ack_done; 1565 1566 /* Ignore duplicate responses. */ 1567 diff = cmp_psn(psn, qp->s_last_psn); 1568 if (unlikely(diff <= 0)) { 1569 /* Update credits for "ghost" ACKs */ 1570 if (diff == 0 && opcode == OP(ACKNOWLEDGE)) { 1571 aeth = be32_to_cpu(ohdr->u.aeth); 1572 if ((aeth >> IB_AETH_NAK_SHIFT) == 0) 1573 rvt_get_credit(qp, aeth); 1574 } 1575 goto ack_done; 1576 } 1577 1578 /* 1579 * Skip everything other than the PSN we expect, if we are waiting 1580 * for a reply to a restarted RDMA read or atomic op. 1581 */ 1582 if (qp->r_flags & RVT_R_RDMAR_SEQ) { 1583 if (cmp_psn(psn, qp->s_last_psn + 1) != 0) 1584 goto ack_done; 1585 qp->r_flags &= ~RVT_R_RDMAR_SEQ; 1586 } 1587 1588 if (unlikely(qp->s_acked == qp->s_tail)) 1589 goto ack_done; 1590 wqe = rvt_get_swqe_ptr(qp, qp->s_acked); 1591 status = IB_WC_SUCCESS; 1592 1593 switch (opcode) { 1594 case OP(ACKNOWLEDGE): 1595 case OP(ATOMIC_ACKNOWLEDGE): 1596 case OP(RDMA_READ_RESPONSE_FIRST): 1597 aeth = be32_to_cpu(ohdr->u.aeth); 1598 if (opcode == OP(ATOMIC_ACKNOWLEDGE)) 1599 val = ib_u64_get(&ohdr->u.at.atomic_ack_eth); 1600 else 1601 val = 0; 1602 if (!do_rc_ack(qp, aeth, psn, opcode, val, rcd) || 1603 opcode != OP(RDMA_READ_RESPONSE_FIRST)) 1604 goto ack_done; 1605 wqe = rvt_get_swqe_ptr(qp, qp->s_acked); 1606 if (unlikely(wqe->wr.opcode != IB_WR_RDMA_READ)) 1607 goto ack_op_err; 1608 /* 1609 * If this is a response to a resent RDMA read, we 1610 * have to be careful to copy the data to the right 1611 * location. 1612 */ 1613 qp->s_rdma_read_len = restart_sge(&qp->s_rdma_read_sge, 1614 wqe, psn, pmtu); 1615 goto read_middle; 1616 1617 case OP(RDMA_READ_RESPONSE_MIDDLE): 1618 /* no AETH, no ACK */ 1619 if (unlikely(cmp_psn(psn, qp->s_last_psn + 1))) 1620 goto ack_seq_err; 1621 if (unlikely(wqe->wr.opcode != IB_WR_RDMA_READ)) 1622 goto ack_op_err; 1623 read_middle: 1624 if (unlikely(tlen != (hdrsize + pmtu + extra_bytes))) 1625 goto ack_len_err; 1626 if (unlikely(pmtu >= qp->s_rdma_read_len)) 1627 goto ack_len_err; 1628 1629 /* 1630 * We got a response so update the timeout. 1631 * 4.096 usec. * (1 << qp->timeout) 1632 */ 1633 rvt_mod_retry_timer(qp); 1634 if (qp->s_flags & RVT_S_WAIT_ACK) { 1635 qp->s_flags &= ~RVT_S_WAIT_ACK; 1636 hfi1_schedule_send(qp); 1637 } 1638 1639 if (opcode == OP(RDMA_READ_RESPONSE_MIDDLE)) 1640 qp->s_retry = qp->s_retry_cnt; 1641 1642 /* 1643 * Update the RDMA receive state but do the copy w/o 1644 * holding the locks and blocking interrupts. 1645 */ 1646 qp->s_rdma_read_len -= pmtu; 1647 update_last_psn(qp, psn); 1648 spin_unlock_irqrestore(&qp->s_lock, flags); 1649 rvt_copy_sge(qp, &qp->s_rdma_read_sge, 1650 data, pmtu, false, false); 1651 goto bail; 1652 1653 case OP(RDMA_READ_RESPONSE_ONLY): 1654 aeth = be32_to_cpu(ohdr->u.aeth); 1655 if (!do_rc_ack(qp, aeth, psn, opcode, 0, rcd)) 1656 goto ack_done; 1657 /* 1658 * Check that the data size is >= 0 && <= pmtu. 1659 * Remember to account for ICRC (4). 1660 */ 1661 if (unlikely(tlen < (hdrsize + extra_bytes))) 1662 goto ack_len_err; 1663 /* 1664 * If this is a response to a resent RDMA read, we 1665 * have to be careful to copy the data to the right 1666 * location. 1667 */ 1668 wqe = rvt_get_swqe_ptr(qp, qp->s_acked); 1669 qp->s_rdma_read_len = restart_sge(&qp->s_rdma_read_sge, 1670 wqe, psn, pmtu); 1671 goto read_last; 1672 1673 case OP(RDMA_READ_RESPONSE_LAST): 1674 /* ACKs READ req. */ 1675 if (unlikely(cmp_psn(psn, qp->s_last_psn + 1))) 1676 goto ack_seq_err; 1677 if (unlikely(wqe->wr.opcode != IB_WR_RDMA_READ)) 1678 goto ack_op_err; 1679 /* 1680 * Check that the data size is >= 1 && <= pmtu. 1681 * Remember to account for ICRC (4). 1682 */ 1683 if (unlikely(tlen <= (hdrsize + extra_bytes))) 1684 goto ack_len_err; 1685 read_last: 1686 tlen -= hdrsize + extra_bytes; 1687 if (unlikely(tlen != qp->s_rdma_read_len)) 1688 goto ack_len_err; 1689 aeth = be32_to_cpu(ohdr->u.aeth); 1690 rvt_copy_sge(qp, &qp->s_rdma_read_sge, 1691 data, tlen, false, false); 1692 WARN_ON(qp->s_rdma_read_sge.num_sge); 1693 (void)do_rc_ack(qp, aeth, psn, 1694 OP(RDMA_READ_RESPONSE_LAST), 0, rcd); 1695 goto ack_done; 1696 } 1697 1698 ack_op_err: 1699 status = IB_WC_LOC_QP_OP_ERR; 1700 goto ack_err; 1701 1702 ack_seq_err: 1703 ibp = rcd_to_iport(rcd); 1704 rdma_seq_err(qp, ibp, psn, rcd); 1705 goto ack_done; 1706 1707 ack_len_err: 1708 status = IB_WC_LOC_LEN_ERR; 1709 ack_err: 1710 if (qp->s_last == qp->s_acked) { 1711 rvt_send_complete(qp, wqe, status); 1712 rvt_error_qp(qp, IB_WC_WR_FLUSH_ERR); 1713 } 1714 ack_done: 1715 spin_unlock_irqrestore(&qp->s_lock, flags); 1716 bail: 1717 return; 1718 } 1719 1720 static inline void rc_defered_ack(struct hfi1_ctxtdata *rcd, 1721 struct rvt_qp *qp) 1722 { 1723 if (list_empty(&qp->rspwait)) { 1724 qp->r_flags |= RVT_R_RSP_NAK; 1725 rvt_get_qp(qp); 1726 list_add_tail(&qp->rspwait, &rcd->qp_wait_list); 1727 } 1728 } 1729 1730 static inline void rc_cancel_ack(struct rvt_qp *qp) 1731 { 1732 qp->r_adefered = 0; 1733 if (list_empty(&qp->rspwait)) 1734 return; 1735 list_del_init(&qp->rspwait); 1736 qp->r_flags &= ~RVT_R_RSP_NAK; 1737 rvt_put_qp(qp); 1738 } 1739 1740 /** 1741 * rc_rcv_error - process an incoming duplicate or error RC packet 1742 * @ohdr: the other headers for this packet 1743 * @data: the packet data 1744 * @qp: the QP for this packet 1745 * @opcode: the opcode for this packet 1746 * @psn: the packet sequence number for this packet 1747 * @diff: the difference between the PSN and the expected PSN 1748 * 1749 * This is called from hfi1_rc_rcv() to process an unexpected 1750 * incoming RC packet for the given QP. 1751 * Called at interrupt level. 1752 * Return 1 if no more processing is needed; otherwise return 0 to 1753 * schedule a response to be sent. 1754 */ 1755 static noinline int rc_rcv_error(struct ib_other_headers *ohdr, void *data, 1756 struct rvt_qp *qp, u32 opcode, u32 psn, 1757 int diff, struct hfi1_ctxtdata *rcd) 1758 { 1759 struct hfi1_ibport *ibp = rcd_to_iport(rcd); 1760 struct rvt_ack_entry *e; 1761 unsigned long flags; 1762 u8 i, prev; 1763 int old_req; 1764 1765 trace_hfi1_rcv_error(qp, psn); 1766 if (diff > 0) { 1767 /* 1768 * Packet sequence error. 1769 * A NAK will ACK earlier sends and RDMA writes. 1770 * Don't queue the NAK if we already sent one. 1771 */ 1772 if (!qp->r_nak_state) { 1773 ibp->rvp.n_rc_seqnak++; 1774 qp->r_nak_state = IB_NAK_PSN_ERROR; 1775 /* Use the expected PSN. */ 1776 qp->r_ack_psn = qp->r_psn; 1777 /* 1778 * Wait to send the sequence NAK until all packets 1779 * in the receive queue have been processed. 1780 * Otherwise, we end up propagating congestion. 1781 */ 1782 rc_defered_ack(rcd, qp); 1783 } 1784 goto done; 1785 } 1786 1787 /* 1788 * Handle a duplicate request. Don't re-execute SEND, RDMA 1789 * write or atomic op. Don't NAK errors, just silently drop 1790 * the duplicate request. Note that r_sge, r_len, and 1791 * r_rcv_len may be in use so don't modify them. 1792 * 1793 * We are supposed to ACK the earliest duplicate PSN but we 1794 * can coalesce an outstanding duplicate ACK. We have to 1795 * send the earliest so that RDMA reads can be restarted at 1796 * the requester's expected PSN. 1797 * 1798 * First, find where this duplicate PSN falls within the 1799 * ACKs previously sent. 1800 * old_req is true if there is an older response that is scheduled 1801 * to be sent before sending this one. 1802 */ 1803 e = NULL; 1804 old_req = 1; 1805 ibp->rvp.n_rc_dupreq++; 1806 1807 spin_lock_irqsave(&qp->s_lock, flags); 1808 1809 for (i = qp->r_head_ack_queue; ; i = prev) { 1810 if (i == qp->s_tail_ack_queue) 1811 old_req = 0; 1812 if (i) 1813 prev = i - 1; 1814 else 1815 prev = HFI1_MAX_RDMA_ATOMIC; 1816 if (prev == qp->r_head_ack_queue) { 1817 e = NULL; 1818 break; 1819 } 1820 e = &qp->s_ack_queue[prev]; 1821 if (!e->opcode) { 1822 e = NULL; 1823 break; 1824 } 1825 if (cmp_psn(psn, e->psn) >= 0) { 1826 if (prev == qp->s_tail_ack_queue && 1827 cmp_psn(psn, e->lpsn) <= 0) 1828 old_req = 0; 1829 break; 1830 } 1831 } 1832 switch (opcode) { 1833 case OP(RDMA_READ_REQUEST): { 1834 struct ib_reth *reth; 1835 u32 offset; 1836 u32 len; 1837 1838 /* 1839 * If we didn't find the RDMA read request in the ack queue, 1840 * we can ignore this request. 1841 */ 1842 if (!e || e->opcode != OP(RDMA_READ_REQUEST)) 1843 goto unlock_done; 1844 /* RETH comes after BTH */ 1845 reth = &ohdr->u.rc.reth; 1846 /* 1847 * Address range must be a subset of the original 1848 * request and start on pmtu boundaries. 1849 * We reuse the old ack_queue slot since the requester 1850 * should not back up and request an earlier PSN for the 1851 * same request. 1852 */ 1853 offset = delta_psn(psn, e->psn) * qp->pmtu; 1854 len = be32_to_cpu(reth->length); 1855 if (unlikely(offset + len != e->rdma_sge.sge_length)) 1856 goto unlock_done; 1857 if (e->rdma_sge.mr) { 1858 rvt_put_mr(e->rdma_sge.mr); 1859 e->rdma_sge.mr = NULL; 1860 } 1861 if (len != 0) { 1862 u32 rkey = be32_to_cpu(reth->rkey); 1863 u64 vaddr = get_ib_reth_vaddr(reth); 1864 int ok; 1865 1866 ok = rvt_rkey_ok(qp, &e->rdma_sge, len, vaddr, rkey, 1867 IB_ACCESS_REMOTE_READ); 1868 if (unlikely(!ok)) 1869 goto unlock_done; 1870 } else { 1871 e->rdma_sge.vaddr = NULL; 1872 e->rdma_sge.length = 0; 1873 e->rdma_sge.sge_length = 0; 1874 } 1875 e->psn = psn; 1876 if (old_req) 1877 goto unlock_done; 1878 qp->s_tail_ack_queue = prev; 1879 break; 1880 } 1881 1882 case OP(COMPARE_SWAP): 1883 case OP(FETCH_ADD): { 1884 /* 1885 * If we didn't find the atomic request in the ack queue 1886 * or the send engine is already backed up to send an 1887 * earlier entry, we can ignore this request. 1888 */ 1889 if (!e || e->opcode != (u8)opcode || old_req) 1890 goto unlock_done; 1891 qp->s_tail_ack_queue = prev; 1892 break; 1893 } 1894 1895 default: 1896 /* 1897 * Ignore this operation if it doesn't request an ACK 1898 * or an earlier RDMA read or atomic is going to be resent. 1899 */ 1900 if (!(psn & IB_BTH_REQ_ACK) || old_req) 1901 goto unlock_done; 1902 /* 1903 * Resend the most recent ACK if this request is 1904 * after all the previous RDMA reads and atomics. 1905 */ 1906 if (i == qp->r_head_ack_queue) { 1907 spin_unlock_irqrestore(&qp->s_lock, flags); 1908 qp->r_nak_state = 0; 1909 qp->r_ack_psn = qp->r_psn - 1; 1910 goto send_ack; 1911 } 1912 1913 /* 1914 * Resend the RDMA read or atomic op which 1915 * ACKs this duplicate request. 1916 */ 1917 qp->s_tail_ack_queue = i; 1918 break; 1919 } 1920 qp->s_ack_state = OP(ACKNOWLEDGE); 1921 qp->s_flags |= RVT_S_RESP_PENDING; 1922 qp->r_nak_state = 0; 1923 hfi1_schedule_send(qp); 1924 1925 unlock_done: 1926 spin_unlock_irqrestore(&qp->s_lock, flags); 1927 done: 1928 return 1; 1929 1930 send_ack: 1931 return 0; 1932 } 1933 1934 static inline void update_ack_queue(struct rvt_qp *qp, unsigned n) 1935 { 1936 unsigned next; 1937 1938 next = n + 1; 1939 if (next > HFI1_MAX_RDMA_ATOMIC) 1940 next = 0; 1941 qp->s_tail_ack_queue = next; 1942 qp->s_ack_state = OP(ACKNOWLEDGE); 1943 } 1944 1945 static void log_cca_event(struct hfi1_pportdata *ppd, u8 sl, u32 rlid, 1946 u32 lqpn, u32 rqpn, u8 svc_type) 1947 { 1948 struct opa_hfi1_cong_log_event_internal *cc_event; 1949 unsigned long flags; 1950 1951 if (sl >= OPA_MAX_SLS) 1952 return; 1953 1954 spin_lock_irqsave(&ppd->cc_log_lock, flags); 1955 1956 ppd->threshold_cong_event_map[sl / 8] |= 1 << (sl % 8); 1957 ppd->threshold_event_counter++; 1958 1959 cc_event = &ppd->cc_events[ppd->cc_log_idx++]; 1960 if (ppd->cc_log_idx == OPA_CONG_LOG_ELEMS) 1961 ppd->cc_log_idx = 0; 1962 cc_event->lqpn = lqpn & RVT_QPN_MASK; 1963 cc_event->rqpn = rqpn & RVT_QPN_MASK; 1964 cc_event->sl = sl; 1965 cc_event->svc_type = svc_type; 1966 cc_event->rlid = rlid; 1967 /* keep timestamp in units of 1.024 usec */ 1968 cc_event->timestamp = ktime_get_ns() / 1024; 1969 1970 spin_unlock_irqrestore(&ppd->cc_log_lock, flags); 1971 } 1972 1973 void process_becn(struct hfi1_pportdata *ppd, u8 sl, u32 rlid, u32 lqpn, 1974 u32 rqpn, u8 svc_type) 1975 { 1976 struct cca_timer *cca_timer; 1977 u16 ccti, ccti_incr, ccti_timer, ccti_limit; 1978 u8 trigger_threshold; 1979 struct cc_state *cc_state; 1980 unsigned long flags; 1981 1982 if (sl >= OPA_MAX_SLS) 1983 return; 1984 1985 cc_state = get_cc_state(ppd); 1986 1987 if (!cc_state) 1988 return; 1989 1990 /* 1991 * 1) increase CCTI (for this SL) 1992 * 2) select IPG (i.e., call set_link_ipg()) 1993 * 3) start timer 1994 */ 1995 ccti_limit = cc_state->cct.ccti_limit; 1996 ccti_incr = cc_state->cong_setting.entries[sl].ccti_increase; 1997 ccti_timer = cc_state->cong_setting.entries[sl].ccti_timer; 1998 trigger_threshold = 1999 cc_state->cong_setting.entries[sl].trigger_threshold; 2000 2001 spin_lock_irqsave(&ppd->cca_timer_lock, flags); 2002 2003 cca_timer = &ppd->cca_timer[sl]; 2004 if (cca_timer->ccti < ccti_limit) { 2005 if (cca_timer->ccti + ccti_incr <= ccti_limit) 2006 cca_timer->ccti += ccti_incr; 2007 else 2008 cca_timer->ccti = ccti_limit; 2009 set_link_ipg(ppd); 2010 } 2011 2012 ccti = cca_timer->ccti; 2013 2014 if (!hrtimer_active(&cca_timer->hrtimer)) { 2015 /* ccti_timer is in units of 1.024 usec */ 2016 unsigned long nsec = 1024 * ccti_timer; 2017 2018 hrtimer_start(&cca_timer->hrtimer, ns_to_ktime(nsec), 2019 HRTIMER_MODE_REL_PINNED); 2020 } 2021 2022 spin_unlock_irqrestore(&ppd->cca_timer_lock, flags); 2023 2024 if ((trigger_threshold != 0) && (ccti >= trigger_threshold)) 2025 log_cca_event(ppd, sl, rlid, lqpn, rqpn, svc_type); 2026 } 2027 2028 /** 2029 * hfi1_rc_rcv - process an incoming RC packet 2030 * @packet: data packet information 2031 * 2032 * This is called from qp_rcv() to process an incoming RC packet 2033 * for the given QP. 2034 * May be called at interrupt level. 2035 */ 2036 void hfi1_rc_rcv(struct hfi1_packet *packet) 2037 { 2038 struct hfi1_ctxtdata *rcd = packet->rcd; 2039 void *data = packet->payload; 2040 u32 tlen = packet->tlen; 2041 struct rvt_qp *qp = packet->qp; 2042 struct hfi1_ibport *ibp = rcd_to_iport(rcd); 2043 struct ib_other_headers *ohdr = packet->ohdr; 2044 u32 opcode = packet->opcode; 2045 u32 hdrsize = packet->hlen; 2046 u32 psn = ib_bth_get_psn(packet->ohdr); 2047 u32 pad = packet->pad; 2048 struct ib_wc wc; 2049 u32 pmtu = qp->pmtu; 2050 int diff; 2051 struct ib_reth *reth; 2052 unsigned long flags; 2053 int ret; 2054 bool copy_last = false, fecn; 2055 u32 rkey; 2056 u8 extra_bytes = pad + packet->extra_byte + (SIZE_OF_CRC << 2); 2057 2058 lockdep_assert_held(&qp->r_lock); 2059 2060 if (hfi1_ruc_check_hdr(ibp, packet)) 2061 return; 2062 2063 fecn = process_ecn(qp, packet); 2064 2065 /* 2066 * Process responses (ACKs) before anything else. Note that the 2067 * packet sequence number will be for something in the send work 2068 * queue rather than the expected receive packet sequence number. 2069 * In other words, this QP is the requester. 2070 */ 2071 if (opcode >= OP(RDMA_READ_RESPONSE_FIRST) && 2072 opcode <= OP(ATOMIC_ACKNOWLEDGE)) { 2073 rc_rcv_resp(packet); 2074 return; 2075 } 2076 2077 /* Compute 24 bits worth of difference. */ 2078 diff = delta_psn(psn, qp->r_psn); 2079 if (unlikely(diff)) { 2080 if (rc_rcv_error(ohdr, data, qp, opcode, psn, diff, rcd)) 2081 return; 2082 goto send_ack; 2083 } 2084 2085 /* Check for opcode sequence errors. */ 2086 switch (qp->r_state) { 2087 case OP(SEND_FIRST): 2088 case OP(SEND_MIDDLE): 2089 if (opcode == OP(SEND_MIDDLE) || 2090 opcode == OP(SEND_LAST) || 2091 opcode == OP(SEND_LAST_WITH_IMMEDIATE) || 2092 opcode == OP(SEND_LAST_WITH_INVALIDATE)) 2093 break; 2094 goto nack_inv; 2095 2096 case OP(RDMA_WRITE_FIRST): 2097 case OP(RDMA_WRITE_MIDDLE): 2098 if (opcode == OP(RDMA_WRITE_MIDDLE) || 2099 opcode == OP(RDMA_WRITE_LAST) || 2100 opcode == OP(RDMA_WRITE_LAST_WITH_IMMEDIATE)) 2101 break; 2102 goto nack_inv; 2103 2104 default: 2105 if (opcode == OP(SEND_MIDDLE) || 2106 opcode == OP(SEND_LAST) || 2107 opcode == OP(SEND_LAST_WITH_IMMEDIATE) || 2108 opcode == OP(SEND_LAST_WITH_INVALIDATE) || 2109 opcode == OP(RDMA_WRITE_MIDDLE) || 2110 opcode == OP(RDMA_WRITE_LAST) || 2111 opcode == OP(RDMA_WRITE_LAST_WITH_IMMEDIATE)) 2112 goto nack_inv; 2113 /* 2114 * Note that it is up to the requester to not send a new 2115 * RDMA read or atomic operation before receiving an ACK 2116 * for the previous operation. 2117 */ 2118 break; 2119 } 2120 2121 if (qp->state == IB_QPS_RTR && !(qp->r_flags & RVT_R_COMM_EST)) 2122 rvt_comm_est(qp); 2123 2124 /* OK, process the packet. */ 2125 switch (opcode) { 2126 case OP(SEND_FIRST): 2127 ret = rvt_get_rwqe(qp, false); 2128 if (ret < 0) 2129 goto nack_op_err; 2130 if (!ret) 2131 goto rnr_nak; 2132 qp->r_rcv_len = 0; 2133 /* FALLTHROUGH */ 2134 case OP(SEND_MIDDLE): 2135 case OP(RDMA_WRITE_MIDDLE): 2136 send_middle: 2137 /* Check for invalid length PMTU or posted rwqe len. */ 2138 /* 2139 * There will be no padding for 9B packet but 16B packets 2140 * will come in with some padding since we always add 2141 * CRC and LT bytes which will need to be flit aligned 2142 */ 2143 if (unlikely(tlen != (hdrsize + pmtu + extra_bytes))) 2144 goto nack_inv; 2145 qp->r_rcv_len += pmtu; 2146 if (unlikely(qp->r_rcv_len > qp->r_len)) 2147 goto nack_inv; 2148 rvt_copy_sge(qp, &qp->r_sge, data, pmtu, true, false); 2149 break; 2150 2151 case OP(RDMA_WRITE_LAST_WITH_IMMEDIATE): 2152 /* consume RWQE */ 2153 ret = rvt_get_rwqe(qp, true); 2154 if (ret < 0) 2155 goto nack_op_err; 2156 if (!ret) 2157 goto rnr_nak; 2158 goto send_last_imm; 2159 2160 case OP(SEND_ONLY): 2161 case OP(SEND_ONLY_WITH_IMMEDIATE): 2162 case OP(SEND_ONLY_WITH_INVALIDATE): 2163 ret = rvt_get_rwqe(qp, false); 2164 if (ret < 0) 2165 goto nack_op_err; 2166 if (!ret) 2167 goto rnr_nak; 2168 qp->r_rcv_len = 0; 2169 if (opcode == OP(SEND_ONLY)) 2170 goto no_immediate_data; 2171 if (opcode == OP(SEND_ONLY_WITH_INVALIDATE)) 2172 goto send_last_inv; 2173 /* FALLTHROUGH -- for SEND_ONLY_WITH_IMMEDIATE */ 2174 case OP(SEND_LAST_WITH_IMMEDIATE): 2175 send_last_imm: 2176 wc.ex.imm_data = ohdr->u.imm_data; 2177 wc.wc_flags = IB_WC_WITH_IMM; 2178 goto send_last; 2179 case OP(SEND_LAST_WITH_INVALIDATE): 2180 send_last_inv: 2181 rkey = be32_to_cpu(ohdr->u.ieth); 2182 if (rvt_invalidate_rkey(qp, rkey)) 2183 goto no_immediate_data; 2184 wc.ex.invalidate_rkey = rkey; 2185 wc.wc_flags = IB_WC_WITH_INVALIDATE; 2186 goto send_last; 2187 case OP(RDMA_WRITE_LAST): 2188 copy_last = rvt_is_user_qp(qp); 2189 /* fall through */ 2190 case OP(SEND_LAST): 2191 no_immediate_data: 2192 wc.wc_flags = 0; 2193 wc.ex.imm_data = 0; 2194 send_last: 2195 /* Check for invalid length. */ 2196 /* LAST len should be >= 1 */ 2197 if (unlikely(tlen < (hdrsize + extra_bytes))) 2198 goto nack_inv; 2199 /* Don't count the CRC(and padding and LT byte for 16B). */ 2200 tlen -= (hdrsize + extra_bytes); 2201 wc.byte_len = tlen + qp->r_rcv_len; 2202 if (unlikely(wc.byte_len > qp->r_len)) 2203 goto nack_inv; 2204 rvt_copy_sge(qp, &qp->r_sge, data, tlen, true, copy_last); 2205 rvt_put_ss(&qp->r_sge); 2206 qp->r_msn++; 2207 if (!__test_and_clear_bit(RVT_R_WRID_VALID, &qp->r_aflags)) 2208 break; 2209 wc.wr_id = qp->r_wr_id; 2210 wc.status = IB_WC_SUCCESS; 2211 if (opcode == OP(RDMA_WRITE_LAST_WITH_IMMEDIATE) || 2212 opcode == OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE)) 2213 wc.opcode = IB_WC_RECV_RDMA_WITH_IMM; 2214 else 2215 wc.opcode = IB_WC_RECV; 2216 wc.qp = &qp->ibqp; 2217 wc.src_qp = qp->remote_qpn; 2218 wc.slid = rdma_ah_get_dlid(&qp->remote_ah_attr) & U16_MAX; 2219 /* 2220 * It seems that IB mandates the presence of an SL in a 2221 * work completion only for the UD transport (see section 2222 * 11.4.2 of IBTA Vol. 1). 2223 * 2224 * However, the way the SL is chosen below is consistent 2225 * with the way that IB/qib works and is trying avoid 2226 * introducing incompatibilities. 2227 * 2228 * See also OPA Vol. 1, section 9.7.6, and table 9-17. 2229 */ 2230 wc.sl = rdma_ah_get_sl(&qp->remote_ah_attr); 2231 /* zero fields that are N/A */ 2232 wc.vendor_err = 0; 2233 wc.pkey_index = 0; 2234 wc.dlid_path_bits = 0; 2235 wc.port_num = 0; 2236 /* Signal completion event if the solicited bit is set. */ 2237 rvt_cq_enter(ibcq_to_rvtcq(qp->ibqp.recv_cq), &wc, 2238 ib_bth_is_solicited(ohdr)); 2239 break; 2240 2241 case OP(RDMA_WRITE_ONLY): 2242 copy_last = rvt_is_user_qp(qp); 2243 /* fall through */ 2244 case OP(RDMA_WRITE_FIRST): 2245 case OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE): 2246 if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_WRITE))) 2247 goto nack_inv; 2248 /* consume RWQE */ 2249 reth = &ohdr->u.rc.reth; 2250 qp->r_len = be32_to_cpu(reth->length); 2251 qp->r_rcv_len = 0; 2252 qp->r_sge.sg_list = NULL; 2253 if (qp->r_len != 0) { 2254 u32 rkey = be32_to_cpu(reth->rkey); 2255 u64 vaddr = get_ib_reth_vaddr(reth); 2256 int ok; 2257 2258 /* Check rkey & NAK */ 2259 ok = rvt_rkey_ok(qp, &qp->r_sge.sge, qp->r_len, vaddr, 2260 rkey, IB_ACCESS_REMOTE_WRITE); 2261 if (unlikely(!ok)) 2262 goto nack_acc; 2263 qp->r_sge.num_sge = 1; 2264 } else { 2265 qp->r_sge.num_sge = 0; 2266 qp->r_sge.sge.mr = NULL; 2267 qp->r_sge.sge.vaddr = NULL; 2268 qp->r_sge.sge.length = 0; 2269 qp->r_sge.sge.sge_length = 0; 2270 } 2271 if (opcode == OP(RDMA_WRITE_FIRST)) 2272 goto send_middle; 2273 else if (opcode == OP(RDMA_WRITE_ONLY)) 2274 goto no_immediate_data; 2275 ret = rvt_get_rwqe(qp, true); 2276 if (ret < 0) 2277 goto nack_op_err; 2278 if (!ret) { 2279 /* peer will send again */ 2280 rvt_put_ss(&qp->r_sge); 2281 goto rnr_nak; 2282 } 2283 wc.ex.imm_data = ohdr->u.rc.imm_data; 2284 wc.wc_flags = IB_WC_WITH_IMM; 2285 goto send_last; 2286 2287 case OP(RDMA_READ_REQUEST): { 2288 struct rvt_ack_entry *e; 2289 u32 len; 2290 u8 next; 2291 2292 if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_READ))) 2293 goto nack_inv; 2294 next = qp->r_head_ack_queue + 1; 2295 /* s_ack_queue is size HFI1_MAX_RDMA_ATOMIC+1 so use > not >= */ 2296 if (next > HFI1_MAX_RDMA_ATOMIC) 2297 next = 0; 2298 spin_lock_irqsave(&qp->s_lock, flags); 2299 if (unlikely(next == qp->s_tail_ack_queue)) { 2300 if (!qp->s_ack_queue[next].sent) 2301 goto nack_inv_unlck; 2302 update_ack_queue(qp, next); 2303 } 2304 e = &qp->s_ack_queue[qp->r_head_ack_queue]; 2305 if (e->opcode == OP(RDMA_READ_REQUEST) && e->rdma_sge.mr) { 2306 rvt_put_mr(e->rdma_sge.mr); 2307 e->rdma_sge.mr = NULL; 2308 } 2309 reth = &ohdr->u.rc.reth; 2310 len = be32_to_cpu(reth->length); 2311 if (len) { 2312 u32 rkey = be32_to_cpu(reth->rkey); 2313 u64 vaddr = get_ib_reth_vaddr(reth); 2314 int ok; 2315 2316 /* Check rkey & NAK */ 2317 ok = rvt_rkey_ok(qp, &e->rdma_sge, len, vaddr, 2318 rkey, IB_ACCESS_REMOTE_READ); 2319 if (unlikely(!ok)) 2320 goto nack_acc_unlck; 2321 /* 2322 * Update the next expected PSN. We add 1 later 2323 * below, so only add the remainder here. 2324 */ 2325 qp->r_psn += rvt_div_mtu(qp, len - 1); 2326 } else { 2327 e->rdma_sge.mr = NULL; 2328 e->rdma_sge.vaddr = NULL; 2329 e->rdma_sge.length = 0; 2330 e->rdma_sge.sge_length = 0; 2331 } 2332 e->opcode = opcode; 2333 e->sent = 0; 2334 e->psn = psn; 2335 e->lpsn = qp->r_psn; 2336 /* 2337 * We need to increment the MSN here instead of when we 2338 * finish sending the result since a duplicate request would 2339 * increment it more than once. 2340 */ 2341 qp->r_msn++; 2342 qp->r_psn++; 2343 qp->r_state = opcode; 2344 qp->r_nak_state = 0; 2345 qp->r_head_ack_queue = next; 2346 2347 /* Schedule the send engine. */ 2348 qp->s_flags |= RVT_S_RESP_PENDING; 2349 if (fecn) 2350 qp->s_flags |= RVT_S_ECN; 2351 hfi1_schedule_send(qp); 2352 2353 spin_unlock_irqrestore(&qp->s_lock, flags); 2354 return; 2355 } 2356 2357 case OP(COMPARE_SWAP): 2358 case OP(FETCH_ADD): { 2359 struct ib_atomic_eth *ateth; 2360 struct rvt_ack_entry *e; 2361 u64 vaddr; 2362 atomic64_t *maddr; 2363 u64 sdata; 2364 u32 rkey; 2365 u8 next; 2366 2367 if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_ATOMIC))) 2368 goto nack_inv; 2369 next = qp->r_head_ack_queue + 1; 2370 if (next > HFI1_MAX_RDMA_ATOMIC) 2371 next = 0; 2372 spin_lock_irqsave(&qp->s_lock, flags); 2373 if (unlikely(next == qp->s_tail_ack_queue)) { 2374 if (!qp->s_ack_queue[next].sent) 2375 goto nack_inv_unlck; 2376 update_ack_queue(qp, next); 2377 } 2378 e = &qp->s_ack_queue[qp->r_head_ack_queue]; 2379 if (e->opcode == OP(RDMA_READ_REQUEST) && e->rdma_sge.mr) { 2380 rvt_put_mr(e->rdma_sge.mr); 2381 e->rdma_sge.mr = NULL; 2382 } 2383 ateth = &ohdr->u.atomic_eth; 2384 vaddr = get_ib_ateth_vaddr(ateth); 2385 if (unlikely(vaddr & (sizeof(u64) - 1))) 2386 goto nack_inv_unlck; 2387 rkey = be32_to_cpu(ateth->rkey); 2388 /* Check rkey & NAK */ 2389 if (unlikely(!rvt_rkey_ok(qp, &qp->r_sge.sge, sizeof(u64), 2390 vaddr, rkey, 2391 IB_ACCESS_REMOTE_ATOMIC))) 2392 goto nack_acc_unlck; 2393 /* Perform atomic OP and save result. */ 2394 maddr = (atomic64_t *)qp->r_sge.sge.vaddr; 2395 sdata = get_ib_ateth_swap(ateth); 2396 e->atomic_data = (opcode == OP(FETCH_ADD)) ? 2397 (u64)atomic64_add_return(sdata, maddr) - sdata : 2398 (u64)cmpxchg((u64 *)qp->r_sge.sge.vaddr, 2399 get_ib_ateth_compare(ateth), 2400 sdata); 2401 rvt_put_mr(qp->r_sge.sge.mr); 2402 qp->r_sge.num_sge = 0; 2403 e->opcode = opcode; 2404 e->sent = 0; 2405 e->psn = psn; 2406 e->lpsn = psn; 2407 qp->r_msn++; 2408 qp->r_psn++; 2409 qp->r_state = opcode; 2410 qp->r_nak_state = 0; 2411 qp->r_head_ack_queue = next; 2412 2413 /* Schedule the send engine. */ 2414 qp->s_flags |= RVT_S_RESP_PENDING; 2415 if (fecn) 2416 qp->s_flags |= RVT_S_ECN; 2417 hfi1_schedule_send(qp); 2418 2419 spin_unlock_irqrestore(&qp->s_lock, flags); 2420 return; 2421 } 2422 2423 default: 2424 /* NAK unknown opcodes. */ 2425 goto nack_inv; 2426 } 2427 qp->r_psn++; 2428 qp->r_state = opcode; 2429 qp->r_ack_psn = psn; 2430 qp->r_nak_state = 0; 2431 /* Send an ACK if requested or required. */ 2432 if (psn & IB_BTH_REQ_ACK || fecn) { 2433 if (packet->numpkt == 0 || fecn || 2434 qp->r_adefered >= HFI1_PSN_CREDIT) { 2435 rc_cancel_ack(qp); 2436 goto send_ack; 2437 } 2438 qp->r_adefered++; 2439 rc_defered_ack(rcd, qp); 2440 } 2441 return; 2442 2443 rnr_nak: 2444 qp->r_nak_state = qp->r_min_rnr_timer | IB_RNR_NAK; 2445 qp->r_ack_psn = qp->r_psn; 2446 /* Queue RNR NAK for later */ 2447 rc_defered_ack(rcd, qp); 2448 return; 2449 2450 nack_op_err: 2451 rvt_rc_error(qp, IB_WC_LOC_QP_OP_ERR); 2452 qp->r_nak_state = IB_NAK_REMOTE_OPERATIONAL_ERROR; 2453 qp->r_ack_psn = qp->r_psn; 2454 /* Queue NAK for later */ 2455 rc_defered_ack(rcd, qp); 2456 return; 2457 2458 nack_inv_unlck: 2459 spin_unlock_irqrestore(&qp->s_lock, flags); 2460 nack_inv: 2461 rvt_rc_error(qp, IB_WC_LOC_QP_OP_ERR); 2462 qp->r_nak_state = IB_NAK_INVALID_REQUEST; 2463 qp->r_ack_psn = qp->r_psn; 2464 /* Queue NAK for later */ 2465 rc_defered_ack(rcd, qp); 2466 return; 2467 2468 nack_acc_unlck: 2469 spin_unlock_irqrestore(&qp->s_lock, flags); 2470 nack_acc: 2471 rvt_rc_error(qp, IB_WC_LOC_PROT_ERR); 2472 qp->r_nak_state = IB_NAK_REMOTE_ACCESS_ERROR; 2473 qp->r_ack_psn = qp->r_psn; 2474 send_ack: 2475 hfi1_send_rc_ack(packet, fecn); 2476 } 2477 2478 void hfi1_rc_hdrerr( 2479 struct hfi1_ctxtdata *rcd, 2480 struct hfi1_packet *packet, 2481 struct rvt_qp *qp) 2482 { 2483 struct hfi1_ibport *ibp = rcd_to_iport(rcd); 2484 int diff; 2485 u32 opcode; 2486 u32 psn; 2487 2488 if (hfi1_ruc_check_hdr(ibp, packet)) 2489 return; 2490 2491 psn = ib_bth_get_psn(packet->ohdr); 2492 opcode = ib_bth_get_opcode(packet->ohdr); 2493 2494 /* Only deal with RDMA Writes for now */ 2495 if (opcode < IB_OPCODE_RC_RDMA_READ_RESPONSE_FIRST) { 2496 diff = delta_psn(psn, qp->r_psn); 2497 if (!qp->r_nak_state && diff >= 0) { 2498 ibp->rvp.n_rc_seqnak++; 2499 qp->r_nak_state = IB_NAK_PSN_ERROR; 2500 /* Use the expected PSN. */ 2501 qp->r_ack_psn = qp->r_psn; 2502 /* 2503 * Wait to send the sequence 2504 * NAK until all packets 2505 * in the receive queue have 2506 * been processed. 2507 * Otherwise, we end up 2508 * propagating congestion. 2509 */ 2510 rc_defered_ack(rcd, qp); 2511 } /* Out of sequence NAK */ 2512 } /* QP Request NAKs */ 2513 } 2514