1/* 2 * Copyright (c) 2006, 2007, 2008, 2009 QLogic Corporation. All rights reserved. 3 * Copyright (c) 2005, 2006 PathScale, Inc. All rights reserved. 4 * 5 * This software is available to you under a choice of one of two 6 * licenses. You may choose to be licensed under the terms of the GNU 7 * General Public License (GPL) Version 2, available from the file 8 * COPYING in the main directory of this source tree, or the 9 * OpenIB.org BSD license below: 10 * 11 * Redistribution and use in source and binary forms, with or 12 * without modification, are permitted provided that the following 13 * conditions are met: 14 * 15 * - Redistributions of source code must retain the above 16 * copyright notice, this list of conditions and the following 17 * disclaimer. 18 * 19 * - Redistributions in binary form must reproduce the above 20 * copyright notice, this list of conditions and the following 21 * disclaimer in the documentation and/or other materials 22 * provided with the distribution. 23 * 24 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, 25 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF 26 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND 27 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS 28 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN 29 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN 30 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE 31 * SOFTWARE. 32 */ 33 34#include <linux/io.h> 35 36#include "qib.h" 37 38/* cut down ridiculously long IB macro names */ 39#define OP(x) IB_OPCODE_RC_##x 40 41 42static u32 restart_sge(struct rvt_sge_state *ss, struct rvt_swqe *wqe, 43 u32 psn, u32 pmtu) 44{ 45 u32 len; 46 47 len = ((psn - wqe->psn) & QIB_PSN_MASK) * pmtu; 48 return rvt_restart_sge(ss, wqe, len); 49} 50 51/** 52 * qib_make_rc_ack - construct a response packet (ACK, NAK, or RDMA read) 53 * @dev: the device for this QP 54 * @qp: a pointer to the QP 55 * @ohdr: a pointer to the IB header being constructed 56 * @pmtu: the path MTU 57 * 58 * Return 1 if constructed; otherwise, return 0. 59 * Note that we are in the responder's side of the QP context. 60 * Note the QP s_lock must be held. 61 */ 62static int qib_make_rc_ack(struct qib_ibdev *dev, struct rvt_qp *qp, 63 struct ib_other_headers *ohdr, u32 pmtu) 64{ 65 struct rvt_ack_entry *e; 66 u32 hwords; 67 u32 len; 68 u32 bth0; 69 u32 bth2; 70 71 /* Don't send an ACK if we aren't supposed to. */ 72 if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK)) 73 goto bail; 74 75 /* header size in 32-bit words LRH+BTH = (8+12)/4. */ 76 hwords = 5; 77 78 switch (qp->s_ack_state) { 79 case OP(RDMA_READ_RESPONSE_LAST): 80 case OP(RDMA_READ_RESPONSE_ONLY): 81 e = &qp->s_ack_queue[qp->s_tail_ack_queue]; 82 if (e->rdma_sge.mr) { 83 rvt_put_mr(e->rdma_sge.mr); 84 e->rdma_sge.mr = NULL; 85 } 86 fallthrough; 87 case OP(ATOMIC_ACKNOWLEDGE): 88 /* 89 * We can increment the tail pointer now that the last 90 * response has been sent instead of only being 91 * constructed. 92 */ 93 if (++qp->s_tail_ack_queue > QIB_MAX_RDMA_ATOMIC) 94 qp->s_tail_ack_queue = 0; 95 fallthrough; 96 case OP(SEND_ONLY): 97 case OP(ACKNOWLEDGE): 98 /* Check for no next entry in the queue. */ 99 if (qp->r_head_ack_queue == qp->s_tail_ack_queue) { 100 if (qp->s_flags & RVT_S_ACK_PENDING) 101 goto normal; 102 goto bail; 103 } 104 105 e = &qp->s_ack_queue[qp->s_tail_ack_queue]; 106 if (e->opcode == OP(RDMA_READ_REQUEST)) { 107 /* 108 * If a RDMA read response is being resent and 109 * we haven't seen the duplicate request yet, 110 * then stop sending the remaining responses the 111 * responder has seen until the requester resends it. 112 */ 113 len = e->rdma_sge.sge_length; 114 if (len && !e->rdma_sge.mr) { 115 qp->s_tail_ack_queue = qp->r_head_ack_queue; 116 goto bail; 117 } 118 /* Copy SGE state in case we need to resend */ 119 qp->s_rdma_mr = e->rdma_sge.mr; 120 if (qp->s_rdma_mr) 121 rvt_get_mr(qp->s_rdma_mr); 122 qp->s_ack_rdma_sge.sge = e->rdma_sge; 123 qp->s_ack_rdma_sge.num_sge = 1; 124 qp->s_cur_sge = &qp->s_ack_rdma_sge; 125 if (len > pmtu) { 126 len = pmtu; 127 qp->s_ack_state = OP(RDMA_READ_RESPONSE_FIRST); 128 } else { 129 qp->s_ack_state = OP(RDMA_READ_RESPONSE_ONLY); 130 e->sent = 1; 131 } 132 ohdr->u.aeth = rvt_compute_aeth(qp); 133 hwords++; 134 qp->s_ack_rdma_psn = e->psn; 135 bth2 = qp->s_ack_rdma_psn++ & QIB_PSN_MASK; 136 } else { 137 /* COMPARE_SWAP or FETCH_ADD */ 138 qp->s_cur_sge = NULL; 139 len = 0; 140 qp->s_ack_state = OP(ATOMIC_ACKNOWLEDGE); 141 ohdr->u.at.aeth = rvt_compute_aeth(qp); 142 ib_u64_put(e->atomic_data, &ohdr->u.at.atomic_ack_eth); 143 hwords += sizeof(ohdr->u.at) / sizeof(u32); 144 bth2 = e->psn & QIB_PSN_MASK; 145 e->sent = 1; 146 } 147 bth0 = qp->s_ack_state << 24; 148 break; 149 150 case OP(RDMA_READ_RESPONSE_FIRST): 151 qp->s_ack_state = OP(RDMA_READ_RESPONSE_MIDDLE); 152 fallthrough; 153 case OP(RDMA_READ_RESPONSE_MIDDLE): 154 qp->s_cur_sge = &qp->s_ack_rdma_sge; 155 qp->s_rdma_mr = qp->s_ack_rdma_sge.sge.mr; 156 if (qp->s_rdma_mr) 157 rvt_get_mr(qp->s_rdma_mr); 158 len = qp->s_ack_rdma_sge.sge.sge_length; 159 if (len > pmtu) 160 len = pmtu; 161 else { 162 ohdr->u.aeth = rvt_compute_aeth(qp); 163 hwords++; 164 qp->s_ack_state = OP(RDMA_READ_RESPONSE_LAST); 165 e = &qp->s_ack_queue[qp->s_tail_ack_queue]; 166 e->sent = 1; 167 } 168 bth0 = qp->s_ack_state << 24; 169 bth2 = qp->s_ack_rdma_psn++ & QIB_PSN_MASK; 170 break; 171 172 default: 173normal: 174 /* 175 * Send a regular ACK. 176 * Set the s_ack_state so we wait until after sending 177 * the ACK before setting s_ack_state to ACKNOWLEDGE 178 * (see above). 179 */ 180 qp->s_ack_state = OP(SEND_ONLY); 181 qp->s_flags &= ~RVT_S_ACK_PENDING; 182 qp->s_cur_sge = NULL; 183 if (qp->s_nak_state) 184 ohdr->u.aeth = 185 cpu_to_be32((qp->r_msn & IB_MSN_MASK) | 186 (qp->s_nak_state << 187 IB_AETH_CREDIT_SHIFT)); 188 else 189 ohdr->u.aeth = rvt_compute_aeth(qp); 190 hwords++; 191 len = 0; 192 bth0 = OP(ACKNOWLEDGE) << 24; 193 bth2 = qp->s_ack_psn & QIB_PSN_MASK; 194 } 195 qp->s_rdma_ack_cnt++; 196 qp->s_hdrwords = hwords; 197 qp->s_cur_size = len; 198 qib_make_ruc_header(qp, ohdr, bth0, bth2); 199 return 1; 200 201bail: 202 qp->s_ack_state = OP(ACKNOWLEDGE); 203 qp->s_flags &= ~(RVT_S_RESP_PENDING | RVT_S_ACK_PENDING); 204 return 0; 205} 206 207/** 208 * qib_make_rc_req - construct a request packet (SEND, RDMA r/w, ATOMIC) 209 * @qp: a pointer to the QP 210 * 211 * Assumes the s_lock is held. 212 * 213 * Return 1 if constructed; otherwise, return 0. 214 */ 215int qib_make_rc_req(struct rvt_qp *qp, unsigned long *flags) 216{ 217 struct qib_qp_priv *priv = qp->priv; 218 struct qib_ibdev *dev = to_idev(qp->ibqp.device); 219 struct ib_other_headers *ohdr; 220 struct rvt_sge_state *ss; 221 struct rvt_swqe *wqe; 222 u32 hwords; 223 u32 len; 224 u32 bth0; 225 u32 bth2; 226 u32 pmtu = qp->pmtu; 227 char newreq; 228 int ret = 0; 229 int delta; 230 231 ohdr = &priv->s_hdr->u.oth; 232 if (rdma_ah_get_ah_flags(&qp->remote_ah_attr) & IB_AH_GRH) 233 ohdr = &priv->s_hdr->u.l.oth; 234 235 /* Sending responses has higher priority over sending requests. */ 236 if ((qp->s_flags & RVT_S_RESP_PENDING) && 237 qib_make_rc_ack(dev, qp, ohdr, pmtu)) 238 goto done; 239 240 if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_SEND_OK)) { 241 if (!(ib_rvt_state_ops[qp->state] & RVT_FLUSH_SEND)) 242 goto bail; 243 /* We are in the error state, flush the work request. */ 244 if (qp->s_last == READ_ONCE(qp->s_head)) 245 goto bail; 246 /* If DMAs are in progress, we can't flush immediately. */ 247 if (atomic_read(&priv->s_dma_busy)) { 248 qp->s_flags |= RVT_S_WAIT_DMA; 249 goto bail; 250 } 251 wqe = rvt_get_swqe_ptr(qp, qp->s_last); 252 rvt_send_complete(qp, wqe, qp->s_last != qp->s_acked ? 253 IB_WC_SUCCESS : IB_WC_WR_FLUSH_ERR); 254 /* will get called again */ 255 goto done; 256 } 257 258 if (qp->s_flags & (RVT_S_WAIT_RNR | RVT_S_WAIT_ACK)) 259 goto bail; 260 261 if (qib_cmp24(qp->s_psn, qp->s_sending_hpsn) <= 0) { 262 if (qib_cmp24(qp->s_sending_psn, qp->s_sending_hpsn) <= 0) { 263 qp->s_flags |= RVT_S_WAIT_PSN; 264 goto bail; 265 } 266 qp->s_sending_psn = qp->s_psn; 267 qp->s_sending_hpsn = qp->s_psn - 1; 268 } 269 270 /* header size in 32-bit words LRH+BTH = (8+12)/4. */ 271 hwords = 5; 272 bth0 = 0; 273 274 /* Send a request. */ 275 wqe = rvt_get_swqe_ptr(qp, qp->s_cur); 276 switch (qp->s_state) { 277 default: 278 if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_NEXT_SEND_OK)) 279 goto bail; 280 /* 281 * Resend an old request or start a new one. 282 * 283 * We keep track of the current SWQE so that 284 * we don't reset the "furthest progress" state 285 * if we need to back up. 286 */ 287 newreq = 0; 288 if (qp->s_cur == qp->s_tail) { 289 /* Check if send work queue is empty. */ 290 if (qp->s_tail == READ_ONCE(qp->s_head)) 291 goto bail; 292 /* 293 * If a fence is requested, wait for previous 294 * RDMA read and atomic operations to finish. 295 */ 296 if ((wqe->wr.send_flags & IB_SEND_FENCE) && 297 qp->s_num_rd_atomic) { 298 qp->s_flags |= RVT_S_WAIT_FENCE; 299 goto bail; 300 } 301 newreq = 1; 302 qp->s_psn = wqe->psn; 303 } 304 /* 305 * Note that we have to be careful not to modify the 306 * original work request since we may need to resend 307 * it. 308 */ 309 len = wqe->length; 310 ss = &qp->s_sge; 311 bth2 = qp->s_psn & QIB_PSN_MASK; 312 switch (wqe->wr.opcode) { 313 case IB_WR_SEND: 314 case IB_WR_SEND_WITH_IMM: 315 /* If no credit, return. */ 316 if (!rvt_rc_credit_avail(qp, wqe)) 317 goto bail; 318 if (len > pmtu) { 319 qp->s_state = OP(SEND_FIRST); 320 len = pmtu; 321 break; 322 } 323 if (wqe->wr.opcode == IB_WR_SEND) 324 qp->s_state = OP(SEND_ONLY); 325 else { 326 qp->s_state = OP(SEND_ONLY_WITH_IMMEDIATE); 327 /* Immediate data comes after the BTH */ 328 ohdr->u.imm_data = wqe->wr.ex.imm_data; 329 hwords += 1; 330 } 331 if (wqe->wr.send_flags & IB_SEND_SOLICITED) 332 bth0 |= IB_BTH_SOLICITED; 333 bth2 |= IB_BTH_REQ_ACK; 334 if (++qp->s_cur == qp->s_size) 335 qp->s_cur = 0; 336 break; 337 338 case IB_WR_RDMA_WRITE: 339 if (newreq && !(qp->s_flags & RVT_S_UNLIMITED_CREDIT)) 340 qp->s_lsn++; 341 goto no_flow_control; 342 case IB_WR_RDMA_WRITE_WITH_IMM: 343 /* If no credit, return. */ 344 if (!rvt_rc_credit_avail(qp, wqe)) 345 goto bail; 346no_flow_control: 347 ohdr->u.rc.reth.vaddr = 348 cpu_to_be64(wqe->rdma_wr.remote_addr); 349 ohdr->u.rc.reth.rkey = 350 cpu_to_be32(wqe->rdma_wr.rkey); 351 ohdr->u.rc.reth.length = cpu_to_be32(len); 352 hwords += sizeof(struct ib_reth) / sizeof(u32); 353 if (len > pmtu) { 354 qp->s_state = OP(RDMA_WRITE_FIRST); 355 len = pmtu; 356 break; 357 } 358 if (wqe->rdma_wr.wr.opcode == IB_WR_RDMA_WRITE) 359 qp->s_state = OP(RDMA_WRITE_ONLY); 360 else { 361 qp->s_state = OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE); 362 /* Immediate data comes after RETH */ 363 ohdr->u.rc.imm_data = 364 wqe->rdma_wr.wr.ex.imm_data; 365 hwords += 1; 366 if (wqe->rdma_wr.wr.send_flags & IB_SEND_SOLICITED) 367 bth0 |= IB_BTH_SOLICITED; 368 } 369 bth2 |= IB_BTH_REQ_ACK; 370 if (++qp->s_cur == qp->s_size) 371 qp->s_cur = 0; 372 break; 373 374 case IB_WR_RDMA_READ: 375 /* 376 * Don't allow more operations to be started 377 * than the QP limits allow. 378 */ 379 if (newreq) { 380 if (qp->s_num_rd_atomic >= 381 qp->s_max_rd_atomic) { 382 qp->s_flags |= RVT_S_WAIT_RDMAR; 383 goto bail; 384 } 385 qp->s_num_rd_atomic++; 386 if (!(qp->s_flags & RVT_S_UNLIMITED_CREDIT)) 387 qp->s_lsn++; 388 } 389 390 ohdr->u.rc.reth.vaddr = 391 cpu_to_be64(wqe->rdma_wr.remote_addr); 392 ohdr->u.rc.reth.rkey = 393 cpu_to_be32(wqe->rdma_wr.rkey); 394 ohdr->u.rc.reth.length = cpu_to_be32(len); 395 qp->s_state = OP(RDMA_READ_REQUEST); 396 hwords += sizeof(ohdr->u.rc.reth) / sizeof(u32); 397 ss = NULL; 398 len = 0; 399 bth2 |= IB_BTH_REQ_ACK; 400 if (++qp->s_cur == qp->s_size) 401 qp->s_cur = 0; 402 break; 403 404 case IB_WR_ATOMIC_CMP_AND_SWP: 405 case IB_WR_ATOMIC_FETCH_AND_ADD: 406 /* 407 * Don't allow more operations to be started 408 * than the QP limits allow. 409 */ 410 if (newreq) { 411 if (qp->s_num_rd_atomic >= 412 qp->s_max_rd_atomic) { 413 qp->s_flags |= RVT_S_WAIT_RDMAR; 414 goto bail; 415 } 416 qp->s_num_rd_atomic++; 417 if (!(qp->s_flags & RVT_S_UNLIMITED_CREDIT)) 418 qp->s_lsn++; 419 } 420 if (wqe->atomic_wr.wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP) { 421 qp->s_state = OP(COMPARE_SWAP); 422 put_ib_ateth_swap(wqe->atomic_wr.swap, 423 &ohdr->u.atomic_eth); 424 put_ib_ateth_compare(wqe->atomic_wr.compare_add, 425 &ohdr->u.atomic_eth); 426 } else { 427 qp->s_state = OP(FETCH_ADD); 428 put_ib_ateth_swap(wqe->atomic_wr.compare_add, 429 &ohdr->u.atomic_eth); 430 put_ib_ateth_compare(0, &ohdr->u.atomic_eth); 431 } 432 put_ib_ateth_vaddr(wqe->atomic_wr.remote_addr, 433 &ohdr->u.atomic_eth); 434 ohdr->u.atomic_eth.rkey = cpu_to_be32( 435 wqe->atomic_wr.rkey); 436 hwords += sizeof(struct ib_atomic_eth) / sizeof(u32); 437 ss = NULL; 438 len = 0; 439 bth2 |= IB_BTH_REQ_ACK; 440 if (++qp->s_cur == qp->s_size) 441 qp->s_cur = 0; 442 break; 443 444 default: 445 goto bail; 446 } 447 qp->s_sge.sge = wqe->sg_list[0]; 448 qp->s_sge.sg_list = wqe->sg_list + 1; 449 qp->s_sge.num_sge = wqe->wr.num_sge; 450 qp->s_sge.total_len = wqe->length; 451 qp->s_len = wqe->length; 452 if (newreq) { 453 qp->s_tail++; 454 if (qp->s_tail >= qp->s_size) 455 qp->s_tail = 0; 456 } 457 if (wqe->wr.opcode == IB_WR_RDMA_READ) 458 qp->s_psn = wqe->lpsn + 1; 459 else 460 qp->s_psn++; 461 break; 462 463 case OP(RDMA_READ_RESPONSE_FIRST): 464 /* 465 * qp->s_state is normally set to the opcode of the 466 * last packet constructed for new requests and therefore 467 * is never set to RDMA read response. 468 * RDMA_READ_RESPONSE_FIRST is used by the ACK processing 469 * thread to indicate a SEND needs to be restarted from an 470 * earlier PSN without interferring with the sending thread. 471 * See qib_restart_rc(). 472 */ 473 qp->s_len = restart_sge(&qp->s_sge, wqe, qp->s_psn, pmtu); 474 fallthrough; 475 case OP(SEND_FIRST): 476 qp->s_state = OP(SEND_MIDDLE); 477 fallthrough; 478 case OP(SEND_MIDDLE): 479 bth2 = qp->s_psn++ & QIB_PSN_MASK; 480 ss = &qp->s_sge; 481 len = qp->s_len; 482 if (len > pmtu) { 483 len = pmtu; 484 break; 485 } 486 if (wqe->wr.opcode == IB_WR_SEND) 487 qp->s_state = OP(SEND_LAST); 488 else { 489 qp->s_state = OP(SEND_LAST_WITH_IMMEDIATE); 490 /* Immediate data comes after the BTH */ 491 ohdr->u.imm_data = wqe->wr.ex.imm_data; 492 hwords += 1; 493 } 494 if (wqe->wr.send_flags & IB_SEND_SOLICITED) 495 bth0 |= IB_BTH_SOLICITED; 496 bth2 |= IB_BTH_REQ_ACK; 497 qp->s_cur++; 498 if (qp->s_cur >= qp->s_size) 499 qp->s_cur = 0; 500 break; 501 502 case OP(RDMA_READ_RESPONSE_LAST): 503 /* 504 * qp->s_state is normally set to the opcode of the 505 * last packet constructed for new requests and therefore 506 * is never set to RDMA read response. 507 * RDMA_READ_RESPONSE_LAST is used by the ACK processing 508 * thread to indicate a RDMA write needs to be restarted from 509 * an earlier PSN without interferring with the sending thread. 510 * See qib_restart_rc(). 511 */ 512 qp->s_len = restart_sge(&qp->s_sge, wqe, qp->s_psn, pmtu); 513 fallthrough; 514 case OP(RDMA_WRITE_FIRST): 515 qp->s_state = OP(RDMA_WRITE_MIDDLE); 516 fallthrough; 517 case OP(RDMA_WRITE_MIDDLE): 518 bth2 = qp->s_psn++ & QIB_PSN_MASK; 519 ss = &qp->s_sge; 520 len = qp->s_len; 521 if (len > pmtu) { 522 len = pmtu; 523 break; 524 } 525 if (wqe->wr.opcode == IB_WR_RDMA_WRITE) 526 qp->s_state = OP(RDMA_WRITE_LAST); 527 else { 528 qp->s_state = OP(RDMA_WRITE_LAST_WITH_IMMEDIATE); 529 /* Immediate data comes after the BTH */ 530 ohdr->u.imm_data = wqe->wr.ex.imm_data; 531 hwords += 1; 532 if (wqe->wr.send_flags & IB_SEND_SOLICITED) 533 bth0 |= IB_BTH_SOLICITED; 534 } 535 bth2 |= IB_BTH_REQ_ACK; 536 qp->s_cur++; 537 if (qp->s_cur >= qp->s_size) 538 qp->s_cur = 0; 539 break; 540 541 case OP(RDMA_READ_RESPONSE_MIDDLE): 542 /* 543 * qp->s_state is normally set to the opcode of the 544 * last packet constructed for new requests and therefore 545 * is never set to RDMA read response. 546 * RDMA_READ_RESPONSE_MIDDLE is used by the ACK processing 547 * thread to indicate a RDMA read needs to be restarted from 548 * an earlier PSN without interferring with the sending thread. 549 * See qib_restart_rc(). 550 */ 551 len = ((qp->s_psn - wqe->psn) & QIB_PSN_MASK) * pmtu; 552 ohdr->u.rc.reth.vaddr = 553 cpu_to_be64(wqe->rdma_wr.remote_addr + len); 554 ohdr->u.rc.reth.rkey = 555 cpu_to_be32(wqe->rdma_wr.rkey); 556 ohdr->u.rc.reth.length = cpu_to_be32(wqe->length - len); 557 qp->s_state = OP(RDMA_READ_REQUEST); 558 hwords += sizeof(ohdr->u.rc.reth) / sizeof(u32); 559 bth2 = (qp->s_psn & QIB_PSN_MASK) | IB_BTH_REQ_ACK; 560 qp->s_psn = wqe->lpsn + 1; 561 ss = NULL; 562 len = 0; 563 qp->s_cur++; 564 if (qp->s_cur == qp->s_size) 565 qp->s_cur = 0; 566 break; 567 } 568 qp->s_sending_hpsn = bth2; 569 delta = (((int) bth2 - (int) wqe->psn) << 8) >> 8; 570 if (delta && delta % QIB_PSN_CREDIT == 0) 571 bth2 |= IB_BTH_REQ_ACK; 572 if (qp->s_flags & RVT_S_SEND_ONE) { 573 qp->s_flags &= ~RVT_S_SEND_ONE; 574 qp->s_flags |= RVT_S_WAIT_ACK; 575 bth2 |= IB_BTH_REQ_ACK; 576 } 577 qp->s_len -= len; 578 qp->s_hdrwords = hwords; 579 qp->s_cur_sge = ss; 580 qp->s_cur_size = len; 581 qib_make_ruc_header(qp, ohdr, bth0 | (qp->s_state << 24), bth2); 582done: 583 return 1; 584bail: 585 qp->s_flags &= ~RVT_S_BUSY; 586 return ret; 587} 588 589/** 590 * qib_send_rc_ack - Construct an ACK packet and send it 591 * @qp: a pointer to the QP 592 * 593 * This is called from qib_rc_rcv() and qib_kreceive(). 594 * Note that RDMA reads and atomics are handled in the 595 * send side QP state and tasklet. 596 */ 597void qib_send_rc_ack(struct rvt_qp *qp) 598{ 599 struct qib_devdata *dd = dd_from_ibdev(qp->ibqp.device); 600 struct qib_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num); 601 struct qib_pportdata *ppd = ppd_from_ibp(ibp); 602 u64 pbc; 603 u16 lrh0; 604 u32 bth0; 605 u32 hwords; 606 u32 pbufn; 607 u32 __iomem *piobuf; 608 struct ib_header hdr; 609 struct ib_other_headers *ohdr; 610 u32 control; 611 unsigned long flags; 612 613 spin_lock_irqsave(&qp->s_lock, flags); 614 615 if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK)) 616 goto unlock; 617 618 /* Don't send ACK or NAK if a RDMA read or atomic is pending. */ 619 if ((qp->s_flags & RVT_S_RESP_PENDING) || qp->s_rdma_ack_cnt) 620 goto queue_ack; 621 622 /* Construct the header with s_lock held so APM doesn't change it. */ 623 ohdr = &hdr.u.oth; 624 lrh0 = QIB_LRH_BTH; 625 /* header size in 32-bit words LRH+BTH+AETH = (8+12+4)/4. */ 626 hwords = 6; 627 if (unlikely(rdma_ah_get_ah_flags(&qp->remote_ah_attr) & 628 IB_AH_GRH)) { 629 hwords += qib_make_grh(ibp, &hdr.u.l.grh, 630 rdma_ah_read_grh(&qp->remote_ah_attr), 631 hwords, 0); 632 ohdr = &hdr.u.l.oth; 633 lrh0 = QIB_LRH_GRH; 634 } 635 /* read pkey_index w/o lock (its atomic) */ 636 bth0 = qib_get_pkey(ibp, qp->s_pkey_index) | (OP(ACKNOWLEDGE) << 24); 637 if (qp->s_mig_state == IB_MIG_MIGRATED) 638 bth0 |= IB_BTH_MIG_REQ; 639 if (qp->r_nak_state) 640 ohdr->u.aeth = cpu_to_be32((qp->r_msn & IB_MSN_MASK) | 641 (qp->r_nak_state << 642 IB_AETH_CREDIT_SHIFT)); 643 else 644 ohdr->u.aeth = rvt_compute_aeth(qp); 645 lrh0 |= ibp->sl_to_vl[rdma_ah_get_sl(&qp->remote_ah_attr)] << 12 | 646 rdma_ah_get_sl(&qp->remote_ah_attr) << 4; 647 hdr.lrh[0] = cpu_to_be16(lrh0); 648 hdr.lrh[1] = cpu_to_be16(rdma_ah_get_dlid(&qp->remote_ah_attr)); 649 hdr.lrh[2] = cpu_to_be16(hwords + SIZE_OF_CRC); 650 hdr.lrh[3] = cpu_to_be16(ppd->lid | 651 rdma_ah_get_path_bits(&qp->remote_ah_attr)); 652 ohdr->bth[0] = cpu_to_be32(bth0); 653 ohdr->bth[1] = cpu_to_be32(qp->remote_qpn); 654 ohdr->bth[2] = cpu_to_be32(qp->r_ack_psn & QIB_PSN_MASK); 655 656 spin_unlock_irqrestore(&qp->s_lock, flags); 657 658 /* Don't try to send ACKs if the link isn't ACTIVE */ 659 if (!(ppd->lflags & QIBL_LINKACTIVE)) 660 goto done; 661 662 control = dd->f_setpbc_control(ppd, hwords + SIZE_OF_CRC, 663 qp->s_srate, lrh0 >> 12); 664 /* length is + 1 for the control dword */ 665 pbc = ((u64) control << 32) | (hwords + 1); 666 667 piobuf = dd->f_getsendbuf(ppd, pbc, &pbufn); 668 if (!piobuf) { 669 /* 670 * We are out of PIO buffers at the moment. 671 * Pass responsibility for sending the ACK to the 672 * send tasklet so that when a PIO buffer becomes 673 * available, the ACK is sent ahead of other outgoing 674 * packets. 675 */ 676 spin_lock_irqsave(&qp->s_lock, flags); 677 goto queue_ack; 678 } 679 680 /* 681 * Write the pbc. 682 * We have to flush after the PBC for correctness 683 * on some cpus or WC buffer can be written out of order. 684 */ 685 writeq(pbc, piobuf); 686 687 if (dd->flags & QIB_PIO_FLUSH_WC) { 688 u32 *hdrp = (u32 *) &hdr; 689 690 qib_flush_wc(); 691 qib_pio_copy(piobuf + 2, hdrp, hwords - 1); 692 qib_flush_wc(); 693 __raw_writel(hdrp[hwords - 1], piobuf + hwords + 1); 694 } else 695 qib_pio_copy(piobuf + 2, (u32 *) &hdr, hwords); 696 697 if (dd->flags & QIB_USE_SPCL_TRIG) { 698 u32 spcl_off = (pbufn >= dd->piobcnt2k) ? 2047 : 1023; 699 700 qib_flush_wc(); 701 __raw_writel(0xaebecede, piobuf + spcl_off); 702 } 703 704 qib_flush_wc(); 705 qib_sendbuf_done(dd, pbufn); 706 707 this_cpu_inc(ibp->pmastats->n_unicast_xmit); 708 goto done; 709 710queue_ack: 711 if (ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK) { 712 this_cpu_inc(*ibp->rvp.rc_qacks); 713 qp->s_flags |= RVT_S_ACK_PENDING | RVT_S_RESP_PENDING; 714 qp->s_nak_state = qp->r_nak_state; 715 qp->s_ack_psn = qp->r_ack_psn; 716 717 /* Schedule the send tasklet. */ 718 qib_schedule_send(qp); 719 } 720unlock: 721 spin_unlock_irqrestore(&qp->s_lock, flags); 722done: 723 return; 724} 725 726/** 727 * reset_psn - reset the QP state to send starting from PSN 728 * @qp: the QP 729 * @psn: the packet sequence number to restart at 730 * 731 * This is called from qib_rc_rcv() to process an incoming RC ACK 732 * for the given QP. 733 * Called at interrupt level with the QP s_lock held. 734 */ 735static void reset_psn(struct rvt_qp *qp, u32 psn) 736{ 737 u32 n = qp->s_acked; 738 struct rvt_swqe *wqe = rvt_get_swqe_ptr(qp, n); 739 u32 opcode; 740 741 qp->s_cur = n; 742 743 /* 744 * If we are starting the request from the beginning, 745 * let the normal send code handle initialization. 746 */ 747 if (qib_cmp24(psn, wqe->psn) <= 0) { 748 qp->s_state = OP(SEND_LAST); 749 goto done; 750 } 751 752 /* Find the work request opcode corresponding to the given PSN. */ 753 opcode = wqe->wr.opcode; 754 for (;;) { 755 int diff; 756 757 if (++n == qp->s_size) 758 n = 0; 759 if (n == qp->s_tail) 760 break; 761 wqe = rvt_get_swqe_ptr(qp, n); 762 diff = qib_cmp24(psn, wqe->psn); 763 if (diff < 0) 764 break; 765 qp->s_cur = n; 766 /* 767 * If we are starting the request from the beginning, 768 * let the normal send code handle initialization. 769 */ 770 if (diff == 0) { 771 qp->s_state = OP(SEND_LAST); 772 goto done; 773 } 774 opcode = wqe->wr.opcode; 775 } 776 777 /* 778 * Set the state to restart in the middle of a request. 779 * Don't change the s_sge, s_cur_sge, or s_cur_size. 780 * See qib_make_rc_req(). 781 */ 782 switch (opcode) { 783 case IB_WR_SEND: 784 case IB_WR_SEND_WITH_IMM: 785 qp->s_state = OP(RDMA_READ_RESPONSE_FIRST); 786 break; 787 788 case IB_WR_RDMA_WRITE: 789 case IB_WR_RDMA_WRITE_WITH_IMM: 790 qp->s_state = OP(RDMA_READ_RESPONSE_LAST); 791 break; 792 793 case IB_WR_RDMA_READ: 794 qp->s_state = OP(RDMA_READ_RESPONSE_MIDDLE); 795 break; 796 797 default: 798 /* 799 * This case shouldn't happen since its only 800 * one PSN per req. 801 */ 802 qp->s_state = OP(SEND_LAST); 803 } 804done: 805 qp->s_psn = psn; 806 /* 807 * Set RVT_S_WAIT_PSN as qib_rc_complete() may start the timer 808 * asynchronously before the send tasklet can get scheduled. 809 * Doing it in qib_make_rc_req() is too late. 810 */ 811 if ((qib_cmp24(qp->s_psn, qp->s_sending_hpsn) <= 0) && 812 (qib_cmp24(qp->s_sending_psn, qp->s_sending_hpsn) <= 0)) 813 qp->s_flags |= RVT_S_WAIT_PSN; 814} 815 816/* 817 * Back up requester to resend the last un-ACKed request. 818 * The QP r_lock and s_lock should be held and interrupts disabled. 819 */ 820void qib_restart_rc(struct rvt_qp *qp, u32 psn, int wait) 821{ 822 struct rvt_swqe *wqe = rvt_get_swqe_ptr(qp, qp->s_acked); 823 struct qib_ibport *ibp; 824 825 if (qp->s_retry == 0) { 826 if (qp->s_mig_state == IB_MIG_ARMED) { 827 qib_migrate_qp(qp); 828 qp->s_retry = qp->s_retry_cnt; 829 } else if (qp->s_last == qp->s_acked) { 830 rvt_send_complete(qp, wqe, IB_WC_RETRY_EXC_ERR); 831 rvt_error_qp(qp, IB_WC_WR_FLUSH_ERR); 832 return; 833 } else /* XXX need to handle delayed completion */ 834 return; 835 } else 836 qp->s_retry--; 837 838 ibp = to_iport(qp->ibqp.device, qp->port_num); 839 if (wqe->wr.opcode == IB_WR_RDMA_READ) 840 ibp->rvp.n_rc_resends++; 841 else 842 ibp->rvp.n_rc_resends += (qp->s_psn - psn) & QIB_PSN_MASK; 843 844 qp->s_flags &= ~(RVT_S_WAIT_FENCE | RVT_S_WAIT_RDMAR | 845 RVT_S_WAIT_SSN_CREDIT | RVT_S_WAIT_PSN | 846 RVT_S_WAIT_ACK); 847 if (wait) 848 qp->s_flags |= RVT_S_SEND_ONE; 849 reset_psn(qp, psn); 850} 851 852/* 853 * Set qp->s_sending_psn to the next PSN after the given one. 854 * This would be psn+1 except when RDMA reads are present. 855 */ 856static void reset_sending_psn(struct rvt_qp *qp, u32 psn) 857{ 858 struct rvt_swqe *wqe; 859 u32 n = qp->s_last; 860 861 /* Find the work request corresponding to the given PSN. */ 862 for (;;) { 863 wqe = rvt_get_swqe_ptr(qp, n); 864 if (qib_cmp24(psn, wqe->lpsn) <= 0) { 865 if (wqe->wr.opcode == IB_WR_RDMA_READ) 866 qp->s_sending_psn = wqe->lpsn + 1; 867 else 868 qp->s_sending_psn = psn + 1; 869 break; 870 } 871 if (++n == qp->s_size) 872 n = 0; 873 if (n == qp->s_tail) 874 break; 875 } 876} 877 878/* 879 * This should be called with the QP s_lock held and interrupts disabled. 880 */ 881void qib_rc_send_complete(struct rvt_qp *qp, struct ib_header *hdr) 882{ 883 struct ib_other_headers *ohdr; 884 struct rvt_swqe *wqe; 885 u32 opcode; 886 u32 psn; 887 888 if (!(ib_rvt_state_ops[qp->state] & RVT_SEND_OR_FLUSH_OR_RECV_OK)) 889 return; 890 891 /* Find out where the BTH is */ 892 if ((be16_to_cpu(hdr->lrh[0]) & 3) == QIB_LRH_BTH) 893 ohdr = &hdr->u.oth; 894 else 895 ohdr = &hdr->u.l.oth; 896 897 opcode = be32_to_cpu(ohdr->bth[0]) >> 24; 898 if (opcode >= OP(RDMA_READ_RESPONSE_FIRST) && 899 opcode <= OP(ATOMIC_ACKNOWLEDGE)) { 900 WARN_ON(!qp->s_rdma_ack_cnt); 901 qp->s_rdma_ack_cnt--; 902 return; 903 } 904 905 psn = be32_to_cpu(ohdr->bth[2]); 906 reset_sending_psn(qp, psn); 907 908 /* 909 * Start timer after a packet requesting an ACK has been sent and 910 * there are still requests that haven't been acked. 911 */ 912 if ((psn & IB_BTH_REQ_ACK) && qp->s_acked != qp->s_tail && 913 !(qp->s_flags & (RVT_S_TIMER | RVT_S_WAIT_RNR | RVT_S_WAIT_PSN)) && 914 (ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK)) 915 rvt_add_retry_timer(qp); 916 917 while (qp->s_last != qp->s_acked) { 918 wqe = rvt_get_swqe_ptr(qp, qp->s_last); 919 if (qib_cmp24(wqe->lpsn, qp->s_sending_psn) >= 0 && 920 qib_cmp24(qp->s_sending_psn, qp->s_sending_hpsn) <= 0) 921 break; 922 rvt_qp_complete_swqe(qp, 923 wqe, 924 ib_qib_wc_opcode[wqe->wr.opcode], 925 IB_WC_SUCCESS); 926 } 927 /* 928 * If we were waiting for sends to complete before resending, 929 * and they are now complete, restart sending. 930 */ 931 if (qp->s_flags & RVT_S_WAIT_PSN && 932 qib_cmp24(qp->s_sending_psn, qp->s_sending_hpsn) > 0) { 933 qp->s_flags &= ~RVT_S_WAIT_PSN; 934 qp->s_sending_psn = qp->s_psn; 935 qp->s_sending_hpsn = qp->s_psn - 1; 936 qib_schedule_send(qp); 937 } 938} 939 940static inline void update_last_psn(struct rvt_qp *qp, u32 psn) 941{ 942 qp->s_last_psn = psn; 943} 944 945/* 946 * Generate a SWQE completion. 947 * This is similar to qib_send_complete but has to check to be sure 948 * that the SGEs are not being referenced if the SWQE is being resent. 949 */ 950static struct rvt_swqe *do_rc_completion(struct rvt_qp *qp, 951 struct rvt_swqe *wqe, 952 struct qib_ibport *ibp) 953{ 954 /* 955 * Don't decrement refcount and don't generate a 956 * completion if the SWQE is being resent until the send 957 * is finished. 958 */ 959 if (qib_cmp24(wqe->lpsn, qp->s_sending_psn) < 0 || 960 qib_cmp24(qp->s_sending_psn, qp->s_sending_hpsn) > 0) 961 rvt_qp_complete_swqe(qp, 962 wqe, 963 ib_qib_wc_opcode[wqe->wr.opcode], 964 IB_WC_SUCCESS); 965 else 966 this_cpu_inc(*ibp->rvp.rc_delayed_comp); 967 968 qp->s_retry = qp->s_retry_cnt; 969 update_last_psn(qp, wqe->lpsn); 970 971 /* 972 * If we are completing a request which is in the process of 973 * being resent, we can stop resending it since we know the 974 * responder has already seen it. 975 */ 976 if (qp->s_acked == qp->s_cur) { 977 if (++qp->s_cur >= qp->s_size) 978 qp->s_cur = 0; 979 qp->s_acked = qp->s_cur; 980 wqe = rvt_get_swqe_ptr(qp, qp->s_cur); 981 if (qp->s_acked != qp->s_tail) { 982 qp->s_state = OP(SEND_LAST); 983 qp->s_psn = wqe->psn; 984 } 985 } else { 986 if (++qp->s_acked >= qp->s_size) 987 qp->s_acked = 0; 988 if (qp->state == IB_QPS_SQD && qp->s_acked == qp->s_cur) 989 qp->s_draining = 0; 990 wqe = rvt_get_swqe_ptr(qp, qp->s_acked); 991 } 992 return wqe; 993} 994 995/** 996 * do_rc_ack - process an incoming RC ACK 997 * @qp: the QP the ACK came in on 998 * @psn: the packet sequence number of the ACK 999 * @opcode: the opcode of the request that resulted in the ACK 1000 * 1001 * This is called from qib_rc_rcv_resp() to process an incoming RC ACK 1002 * for the given QP. 1003 * Called at interrupt level with the QP s_lock held. 1004 * Returns 1 if OK, 0 if current operation should be aborted (NAK). 1005 */ 1006static int do_rc_ack(struct rvt_qp *qp, u32 aeth, u32 psn, int opcode, 1007 u64 val, struct qib_ctxtdata *rcd) 1008{ 1009 struct qib_ibport *ibp; 1010 enum ib_wc_status status; 1011 struct rvt_swqe *wqe; 1012 int ret = 0; 1013 u32 ack_psn; 1014 int diff; 1015 1016 /* 1017 * Note that NAKs implicitly ACK outstanding SEND and RDMA write 1018 * requests and implicitly NAK RDMA read and atomic requests issued 1019 * before the NAK'ed request. The MSN won't include the NAK'ed 1020 * request but will include an ACK'ed request(s). 1021 */ 1022 ack_psn = psn; 1023 if (aeth >> IB_AETH_NAK_SHIFT) 1024 ack_psn--; 1025 wqe = rvt_get_swqe_ptr(qp, qp->s_acked); 1026 ibp = to_iport(qp->ibqp.device, qp->port_num); 1027 1028 /* 1029 * The MSN might be for a later WQE than the PSN indicates so 1030 * only complete WQEs that the PSN finishes. 1031 */ 1032 while ((diff = qib_cmp24(ack_psn, wqe->lpsn)) >= 0) { 1033 /* 1034 * RDMA_READ_RESPONSE_ONLY is a special case since 1035 * we want to generate completion events for everything 1036 * before the RDMA read, copy the data, then generate 1037 * the completion for the read. 1038 */ 1039 if (wqe->wr.opcode == IB_WR_RDMA_READ && 1040 opcode == OP(RDMA_READ_RESPONSE_ONLY) && 1041 diff == 0) { 1042 ret = 1; 1043 goto bail; 1044 } 1045 /* 1046 * If this request is a RDMA read or atomic, and the ACK is 1047 * for a later operation, this ACK NAKs the RDMA read or 1048 * atomic. In other words, only a RDMA_READ_LAST or ONLY 1049 * can ACK a RDMA read and likewise for atomic ops. Note 1050 * that the NAK case can only happen if relaxed ordering is 1051 * used and requests are sent after an RDMA read or atomic 1052 * is sent but before the response is received. 1053 */ 1054 if ((wqe->wr.opcode == IB_WR_RDMA_READ && 1055 (opcode != OP(RDMA_READ_RESPONSE_LAST) || diff != 0)) || 1056 ((wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP || 1057 wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD) && 1058 (opcode != OP(ATOMIC_ACKNOWLEDGE) || diff != 0))) { 1059 /* Retry this request. */ 1060 if (!(qp->r_flags & RVT_R_RDMAR_SEQ)) { 1061 qp->r_flags |= RVT_R_RDMAR_SEQ; 1062 qib_restart_rc(qp, qp->s_last_psn + 1, 0); 1063 if (list_empty(&qp->rspwait)) { 1064 qp->r_flags |= RVT_R_RSP_SEND; 1065 rvt_get_qp(qp); 1066 list_add_tail(&qp->rspwait, 1067 &rcd->qp_wait_list); 1068 } 1069 } 1070 /* 1071 * No need to process the ACK/NAK since we are 1072 * restarting an earlier request. 1073 */ 1074 goto bail; 1075 } 1076 if (wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP || 1077 wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD) { 1078 u64 *vaddr = wqe->sg_list[0].vaddr; 1079 *vaddr = val; 1080 } 1081 if (qp->s_num_rd_atomic && 1082 (wqe->wr.opcode == IB_WR_RDMA_READ || 1083 wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP || 1084 wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD)) { 1085 qp->s_num_rd_atomic--; 1086 /* Restart sending task if fence is complete */ 1087 if ((qp->s_flags & RVT_S_WAIT_FENCE) && 1088 !qp->s_num_rd_atomic) { 1089 qp->s_flags &= ~(RVT_S_WAIT_FENCE | 1090 RVT_S_WAIT_ACK); 1091 qib_schedule_send(qp); 1092 } else if (qp->s_flags & RVT_S_WAIT_RDMAR) { 1093 qp->s_flags &= ~(RVT_S_WAIT_RDMAR | 1094 RVT_S_WAIT_ACK); 1095 qib_schedule_send(qp); 1096 } 1097 } 1098 wqe = do_rc_completion(qp, wqe, ibp); 1099 if (qp->s_acked == qp->s_tail) 1100 break; 1101 } 1102 1103 switch (aeth >> IB_AETH_NAK_SHIFT) { 1104 case 0: /* ACK */ 1105 this_cpu_inc(*ibp->rvp.rc_acks); 1106 if (qp->s_acked != qp->s_tail) { 1107 /* 1108 * We are expecting more ACKs so 1109 * reset the retransmit timer. 1110 */ 1111 rvt_mod_retry_timer(qp); 1112 /* 1113 * We can stop resending the earlier packets and 1114 * continue with the next packet the receiver wants. 1115 */ 1116 if (qib_cmp24(qp->s_psn, psn) <= 0) 1117 reset_psn(qp, psn + 1); 1118 } else { 1119 /* No more acks - kill all timers */ 1120 rvt_stop_rc_timers(qp); 1121 if (qib_cmp24(qp->s_psn, psn) <= 0) { 1122 qp->s_state = OP(SEND_LAST); 1123 qp->s_psn = psn + 1; 1124 } 1125 } 1126 if (qp->s_flags & RVT_S_WAIT_ACK) { 1127 qp->s_flags &= ~RVT_S_WAIT_ACK; 1128 qib_schedule_send(qp); 1129 } 1130 rvt_get_credit(qp, aeth); 1131 qp->s_rnr_retry = qp->s_rnr_retry_cnt; 1132 qp->s_retry = qp->s_retry_cnt; 1133 update_last_psn(qp, psn); 1134 return 1; 1135 1136 case 1: /* RNR NAK */ 1137 ibp->rvp.n_rnr_naks++; 1138 if (qp->s_acked == qp->s_tail) 1139 goto bail; 1140 if (qp->s_flags & RVT_S_WAIT_RNR) 1141 goto bail; 1142 if (qp->s_rnr_retry == 0) { 1143 status = IB_WC_RNR_RETRY_EXC_ERR; 1144 goto class_b; 1145 } 1146 if (qp->s_rnr_retry_cnt < 7) 1147 qp->s_rnr_retry--; 1148 1149 /* The last valid PSN is the previous PSN. */ 1150 update_last_psn(qp, psn - 1); 1151 1152 ibp->rvp.n_rc_resends += (qp->s_psn - psn) & QIB_PSN_MASK; 1153 1154 reset_psn(qp, psn); 1155 1156 qp->s_flags &= ~(RVT_S_WAIT_SSN_CREDIT | RVT_S_WAIT_ACK); 1157 rvt_stop_rc_timers(qp); 1158 rvt_add_rnr_timer(qp, aeth); 1159 return 0; 1160 1161 case 3: /* NAK */ 1162 if (qp->s_acked == qp->s_tail) 1163 goto bail; 1164 /* The last valid PSN is the previous PSN. */ 1165 update_last_psn(qp, psn - 1); 1166 switch ((aeth >> IB_AETH_CREDIT_SHIFT) & 1167 IB_AETH_CREDIT_MASK) { 1168 case 0: /* PSN sequence error */ 1169 ibp->rvp.n_seq_naks++; 1170 /* 1171 * Back up to the responder's expected PSN. 1172 * Note that we might get a NAK in the middle of an 1173 * RDMA READ response which terminates the RDMA 1174 * READ. 1175 */ 1176 qib_restart_rc(qp, psn, 0); 1177 qib_schedule_send(qp); 1178 break; 1179 1180 case 1: /* Invalid Request */ 1181 status = IB_WC_REM_INV_REQ_ERR; 1182 ibp->rvp.n_other_naks++; 1183 goto class_b; 1184 1185 case 2: /* Remote Access Error */ 1186 status = IB_WC_REM_ACCESS_ERR; 1187 ibp->rvp.n_other_naks++; 1188 goto class_b; 1189 1190 case 3: /* Remote Operation Error */ 1191 status = IB_WC_REM_OP_ERR; 1192 ibp->rvp.n_other_naks++; 1193class_b: 1194 if (qp->s_last == qp->s_acked) { 1195 rvt_send_complete(qp, wqe, status); 1196 rvt_error_qp(qp, IB_WC_WR_FLUSH_ERR); 1197 } 1198 break; 1199 1200 default: 1201 /* Ignore other reserved NAK error codes */ 1202 goto reserved; 1203 } 1204 qp->s_retry = qp->s_retry_cnt; 1205 qp->s_rnr_retry = qp->s_rnr_retry_cnt; 1206 goto bail; 1207 1208 default: /* 2: reserved */ 1209reserved: 1210 /* Ignore reserved NAK codes. */ 1211 goto bail; 1212 } 1213 1214bail: 1215 rvt_stop_rc_timers(qp); 1216 return ret; 1217} 1218 1219/* 1220 * We have seen an out of sequence RDMA read middle or last packet. 1221 * This ACKs SENDs and RDMA writes up to the first RDMA read or atomic SWQE. 1222 */ 1223static void rdma_seq_err(struct rvt_qp *qp, struct qib_ibport *ibp, u32 psn, 1224 struct qib_ctxtdata *rcd) 1225{ 1226 struct rvt_swqe *wqe; 1227 1228 /* Remove QP from retry timer */ 1229 rvt_stop_rc_timers(qp); 1230 1231 wqe = rvt_get_swqe_ptr(qp, qp->s_acked); 1232 1233 while (qib_cmp24(psn, wqe->lpsn) > 0) { 1234 if (wqe->wr.opcode == IB_WR_RDMA_READ || 1235 wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP || 1236 wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD) 1237 break; 1238 wqe = do_rc_completion(qp, wqe, ibp); 1239 } 1240 1241 ibp->rvp.n_rdma_seq++; 1242 qp->r_flags |= RVT_R_RDMAR_SEQ; 1243 qib_restart_rc(qp, qp->s_last_psn + 1, 0); 1244 if (list_empty(&qp->rspwait)) { 1245 qp->r_flags |= RVT_R_RSP_SEND; 1246 rvt_get_qp(qp); 1247 list_add_tail(&qp->rspwait, &rcd->qp_wait_list); 1248 } 1249} 1250 1251/** 1252 * qib_rc_rcv_resp - process an incoming RC response packet 1253 * @ibp: the port this packet came in on 1254 * @ohdr: the other headers for this packet 1255 * @data: the packet data 1256 * @tlen: the packet length 1257 * @qp: the QP for this packet 1258 * @opcode: the opcode for this packet 1259 * @psn: the packet sequence number for this packet 1260 * @hdrsize: the header length 1261 * @pmtu: the path MTU 1262 * 1263 * This is called from qib_rc_rcv() to process an incoming RC response 1264 * packet for the given QP. 1265 * Called at interrupt level. 1266 */ 1267static void qib_rc_rcv_resp(struct qib_ibport *ibp, 1268 struct ib_other_headers *ohdr, 1269 void *data, u32 tlen, 1270 struct rvt_qp *qp, 1271 u32 opcode, 1272 u32 psn, u32 hdrsize, u32 pmtu, 1273 struct qib_ctxtdata *rcd) 1274{ 1275 struct rvt_swqe *wqe; 1276 struct qib_pportdata *ppd = ppd_from_ibp(ibp); 1277 enum ib_wc_status status; 1278 unsigned long flags; 1279 int diff; 1280 u32 pad; 1281 u32 aeth; 1282 u64 val; 1283 1284 if (opcode != OP(RDMA_READ_RESPONSE_MIDDLE)) { 1285 /* 1286 * If ACK'd PSN on SDMA busy list try to make progress to 1287 * reclaim SDMA credits. 1288 */ 1289 if ((qib_cmp24(psn, qp->s_sending_psn) >= 0) && 1290 (qib_cmp24(qp->s_sending_psn, qp->s_sending_hpsn) <= 0)) { 1291 1292 /* 1293 * If send tasklet not running attempt to progress 1294 * SDMA queue. 1295 */ 1296 if (!(qp->s_flags & RVT_S_BUSY)) { 1297 /* Acquire SDMA Lock */ 1298 spin_lock_irqsave(&ppd->sdma_lock, flags); 1299 /* Invoke sdma make progress */ 1300 qib_sdma_make_progress(ppd); 1301 /* Release SDMA Lock */ 1302 spin_unlock_irqrestore(&ppd->sdma_lock, flags); 1303 } 1304 } 1305 } 1306 1307 spin_lock_irqsave(&qp->s_lock, flags); 1308 if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK)) 1309 goto ack_done; 1310 1311 /* Ignore invalid responses. */ 1312 if (qib_cmp24(psn, READ_ONCE(qp->s_next_psn)) >= 0) 1313 goto ack_done; 1314 1315 /* Ignore duplicate responses. */ 1316 diff = qib_cmp24(psn, qp->s_last_psn); 1317 if (unlikely(diff <= 0)) { 1318 /* Update credits for "ghost" ACKs */ 1319 if (diff == 0 && opcode == OP(ACKNOWLEDGE)) { 1320 aeth = be32_to_cpu(ohdr->u.aeth); 1321 if ((aeth >> IB_AETH_NAK_SHIFT) == 0) 1322 rvt_get_credit(qp, aeth); 1323 } 1324 goto ack_done; 1325 } 1326 1327 /* 1328 * Skip everything other than the PSN we expect, if we are waiting 1329 * for a reply to a restarted RDMA read or atomic op. 1330 */ 1331 if (qp->r_flags & RVT_R_RDMAR_SEQ) { 1332 if (qib_cmp24(psn, qp->s_last_psn + 1) != 0) 1333 goto ack_done; 1334 qp->r_flags &= ~RVT_R_RDMAR_SEQ; 1335 } 1336 1337 if (unlikely(qp->s_acked == qp->s_tail)) 1338 goto ack_done; 1339 wqe = rvt_get_swqe_ptr(qp, qp->s_acked); 1340 status = IB_WC_SUCCESS; 1341 1342 switch (opcode) { 1343 case OP(ACKNOWLEDGE): 1344 case OP(ATOMIC_ACKNOWLEDGE): 1345 case OP(RDMA_READ_RESPONSE_FIRST): 1346 aeth = be32_to_cpu(ohdr->u.aeth); 1347 if (opcode == OP(ATOMIC_ACKNOWLEDGE)) 1348 val = ib_u64_get(&ohdr->u.at.atomic_ack_eth); 1349 else 1350 val = 0; 1351 if (!do_rc_ack(qp, aeth, psn, opcode, val, rcd) || 1352 opcode != OP(RDMA_READ_RESPONSE_FIRST)) 1353 goto ack_done; 1354 hdrsize += 4; 1355 wqe = rvt_get_swqe_ptr(qp, qp->s_acked); 1356 if (unlikely(wqe->wr.opcode != IB_WR_RDMA_READ)) 1357 goto ack_op_err; 1358 /* 1359 * If this is a response to a resent RDMA read, we 1360 * have to be careful to copy the data to the right 1361 * location. 1362 */ 1363 qp->s_rdma_read_len = restart_sge(&qp->s_rdma_read_sge, 1364 wqe, psn, pmtu); 1365 goto read_middle; 1366 1367 case OP(RDMA_READ_RESPONSE_MIDDLE): 1368 /* no AETH, no ACK */ 1369 if (unlikely(qib_cmp24(psn, qp->s_last_psn + 1))) 1370 goto ack_seq_err; 1371 if (unlikely(wqe->wr.opcode != IB_WR_RDMA_READ)) 1372 goto ack_op_err; 1373read_middle: 1374 if (unlikely(tlen != (hdrsize + pmtu + 4))) 1375 goto ack_len_err; 1376 if (unlikely(pmtu >= qp->s_rdma_read_len)) 1377 goto ack_len_err; 1378 1379 /* 1380 * We got a response so update the timeout. 1381 * 4.096 usec. * (1 << qp->timeout) 1382 */ 1383 rvt_mod_retry_timer(qp); 1384 if (qp->s_flags & RVT_S_WAIT_ACK) { 1385 qp->s_flags &= ~RVT_S_WAIT_ACK; 1386 qib_schedule_send(qp); 1387 } 1388 1389 if (opcode == OP(RDMA_READ_RESPONSE_MIDDLE)) 1390 qp->s_retry = qp->s_retry_cnt; 1391 1392 /* 1393 * Update the RDMA receive state but do the copy w/o 1394 * holding the locks and blocking interrupts. 1395 */ 1396 qp->s_rdma_read_len -= pmtu; 1397 update_last_psn(qp, psn); 1398 spin_unlock_irqrestore(&qp->s_lock, flags); 1399 rvt_copy_sge(qp, &qp->s_rdma_read_sge, 1400 data, pmtu, false, false); 1401 goto bail; 1402 1403 case OP(RDMA_READ_RESPONSE_ONLY): 1404 aeth = be32_to_cpu(ohdr->u.aeth); 1405 if (!do_rc_ack(qp, aeth, psn, opcode, 0, rcd)) 1406 goto ack_done; 1407 /* Get the number of bytes the message was padded by. */ 1408 pad = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3; 1409 /* 1410 * Check that the data size is >= 0 && <= pmtu. 1411 * Remember to account for the AETH header (4) and 1412 * ICRC (4). 1413 */ 1414 if (unlikely(tlen < (hdrsize + pad + 8))) 1415 goto ack_len_err; 1416 /* 1417 * If this is a response to a resent RDMA read, we 1418 * have to be careful to copy the data to the right 1419 * location. 1420 */ 1421 wqe = rvt_get_swqe_ptr(qp, qp->s_acked); 1422 qp->s_rdma_read_len = restart_sge(&qp->s_rdma_read_sge, 1423 wqe, psn, pmtu); 1424 goto read_last; 1425 1426 case OP(RDMA_READ_RESPONSE_LAST): 1427 /* ACKs READ req. */ 1428 if (unlikely(qib_cmp24(psn, qp->s_last_psn + 1))) 1429 goto ack_seq_err; 1430 if (unlikely(wqe->wr.opcode != IB_WR_RDMA_READ)) 1431 goto ack_op_err; 1432 /* Get the number of bytes the message was padded by. */ 1433 pad = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3; 1434 /* 1435 * Check that the data size is >= 1 && <= pmtu. 1436 * Remember to account for the AETH header (4) and 1437 * ICRC (4). 1438 */ 1439 if (unlikely(tlen <= (hdrsize + pad + 8))) 1440 goto ack_len_err; 1441read_last: 1442 tlen -= hdrsize + pad + 8; 1443 if (unlikely(tlen != qp->s_rdma_read_len)) 1444 goto ack_len_err; 1445 aeth = be32_to_cpu(ohdr->u.aeth); 1446 rvt_copy_sge(qp, &qp->s_rdma_read_sge, 1447 data, tlen, false, false); 1448 WARN_ON(qp->s_rdma_read_sge.num_sge); 1449 (void) do_rc_ack(qp, aeth, psn, 1450 OP(RDMA_READ_RESPONSE_LAST), 0, rcd); 1451 goto ack_done; 1452 } 1453 1454ack_op_err: 1455 status = IB_WC_LOC_QP_OP_ERR; 1456 goto ack_err; 1457 1458ack_seq_err: 1459 rdma_seq_err(qp, ibp, psn, rcd); 1460 goto ack_done; 1461 1462ack_len_err: 1463 status = IB_WC_LOC_LEN_ERR; 1464ack_err: 1465 if (qp->s_last == qp->s_acked) { 1466 rvt_send_complete(qp, wqe, status); 1467 rvt_error_qp(qp, IB_WC_WR_FLUSH_ERR); 1468 } 1469ack_done: 1470 spin_unlock_irqrestore(&qp->s_lock, flags); 1471bail: 1472 return; 1473} 1474 1475/** 1476 * qib_rc_rcv_error - process an incoming duplicate or error RC packet 1477 * @ohdr: the other headers for this packet 1478 * @data: the packet data 1479 * @qp: the QP for this packet 1480 * @opcode: the opcode for this packet 1481 * @psn: the packet sequence number for this packet 1482 * @diff: the difference between the PSN and the expected PSN 1483 * 1484 * This is called from qib_rc_rcv() to process an unexpected 1485 * incoming RC packet for the given QP. 1486 * Called at interrupt level. 1487 * Return 1 if no more processing is needed; otherwise return 0 to 1488 * schedule a response to be sent. 1489 */ 1490static int qib_rc_rcv_error(struct ib_other_headers *ohdr, 1491 void *data, 1492 struct rvt_qp *qp, 1493 u32 opcode, 1494 u32 psn, 1495 int diff, 1496 struct qib_ctxtdata *rcd) 1497{ 1498 struct qib_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num); 1499 struct rvt_ack_entry *e; 1500 unsigned long flags; 1501 u8 i, prev; 1502 int old_req; 1503 1504 if (diff > 0) { 1505 /* 1506 * Packet sequence error. 1507 * A NAK will ACK earlier sends and RDMA writes. 1508 * Don't queue the NAK if we already sent one. 1509 */ 1510 if (!qp->r_nak_state) { 1511 ibp->rvp.n_rc_seqnak++; 1512 qp->r_nak_state = IB_NAK_PSN_ERROR; 1513 /* Use the expected PSN. */ 1514 qp->r_ack_psn = qp->r_psn; 1515 /* 1516 * Wait to send the sequence NAK until all packets 1517 * in the receive queue have been processed. 1518 * Otherwise, we end up propagating congestion. 1519 */ 1520 if (list_empty(&qp->rspwait)) { 1521 qp->r_flags |= RVT_R_RSP_NAK; 1522 rvt_get_qp(qp); 1523 list_add_tail(&qp->rspwait, &rcd->qp_wait_list); 1524 } 1525 } 1526 goto done; 1527 } 1528 1529 /* 1530 * Handle a duplicate request. Don't re-execute SEND, RDMA 1531 * write or atomic op. Don't NAK errors, just silently drop 1532 * the duplicate request. Note that r_sge, r_len, and 1533 * r_rcv_len may be in use so don't modify them. 1534 * 1535 * We are supposed to ACK the earliest duplicate PSN but we 1536 * can coalesce an outstanding duplicate ACK. We have to 1537 * send the earliest so that RDMA reads can be restarted at 1538 * the requester's expected PSN. 1539 * 1540 * First, find where this duplicate PSN falls within the 1541 * ACKs previously sent. 1542 * old_req is true if there is an older response that is scheduled 1543 * to be sent before sending this one. 1544 */ 1545 e = NULL; 1546 old_req = 1; 1547 ibp->rvp.n_rc_dupreq++; 1548 1549 spin_lock_irqsave(&qp->s_lock, flags); 1550 1551 for (i = qp->r_head_ack_queue; ; i = prev) { 1552 if (i == qp->s_tail_ack_queue) 1553 old_req = 0; 1554 if (i) 1555 prev = i - 1; 1556 else 1557 prev = QIB_MAX_RDMA_ATOMIC; 1558 if (prev == qp->r_head_ack_queue) { 1559 e = NULL; 1560 break; 1561 } 1562 e = &qp->s_ack_queue[prev]; 1563 if (!e->opcode) { 1564 e = NULL; 1565 break; 1566 } 1567 if (qib_cmp24(psn, e->psn) >= 0) { 1568 if (prev == qp->s_tail_ack_queue && 1569 qib_cmp24(psn, e->lpsn) <= 0) 1570 old_req = 0; 1571 break; 1572 } 1573 } 1574 switch (opcode) { 1575 case OP(RDMA_READ_REQUEST): { 1576 struct ib_reth *reth; 1577 u32 offset; 1578 u32 len; 1579 1580 /* 1581 * If we didn't find the RDMA read request in the ack queue, 1582 * we can ignore this request. 1583 */ 1584 if (!e || e->opcode != OP(RDMA_READ_REQUEST)) 1585 goto unlock_done; 1586 /* RETH comes after BTH */ 1587 reth = &ohdr->u.rc.reth; 1588 /* 1589 * Address range must be a subset of the original 1590 * request and start on pmtu boundaries. 1591 * We reuse the old ack_queue slot since the requester 1592 * should not back up and request an earlier PSN for the 1593 * same request. 1594 */ 1595 offset = ((psn - e->psn) & QIB_PSN_MASK) * 1596 qp->pmtu; 1597 len = be32_to_cpu(reth->length); 1598 if (unlikely(offset + len != e->rdma_sge.sge_length)) 1599 goto unlock_done; 1600 if (e->rdma_sge.mr) { 1601 rvt_put_mr(e->rdma_sge.mr); 1602 e->rdma_sge.mr = NULL; 1603 } 1604 if (len != 0) { 1605 u32 rkey = be32_to_cpu(reth->rkey); 1606 u64 vaddr = be64_to_cpu(reth->vaddr); 1607 int ok; 1608 1609 ok = rvt_rkey_ok(qp, &e->rdma_sge, len, vaddr, rkey, 1610 IB_ACCESS_REMOTE_READ); 1611 if (unlikely(!ok)) 1612 goto unlock_done; 1613 } else { 1614 e->rdma_sge.vaddr = NULL; 1615 e->rdma_sge.length = 0; 1616 e->rdma_sge.sge_length = 0; 1617 } 1618 e->psn = psn; 1619 if (old_req) 1620 goto unlock_done; 1621 qp->s_tail_ack_queue = prev; 1622 break; 1623 } 1624 1625 case OP(COMPARE_SWAP): 1626 case OP(FETCH_ADD): { 1627 /* 1628 * If we didn't find the atomic request in the ack queue 1629 * or the send tasklet is already backed up to send an 1630 * earlier entry, we can ignore this request. 1631 */ 1632 if (!e || e->opcode != (u8) opcode || old_req) 1633 goto unlock_done; 1634 qp->s_tail_ack_queue = prev; 1635 break; 1636 } 1637 1638 default: 1639 /* 1640 * Ignore this operation if it doesn't request an ACK 1641 * or an earlier RDMA read or atomic is going to be resent. 1642 */ 1643 if (!(psn & IB_BTH_REQ_ACK) || old_req) 1644 goto unlock_done; 1645 /* 1646 * Resend the most recent ACK if this request is 1647 * after all the previous RDMA reads and atomics. 1648 */ 1649 if (i == qp->r_head_ack_queue) { 1650 spin_unlock_irqrestore(&qp->s_lock, flags); 1651 qp->r_nak_state = 0; 1652 qp->r_ack_psn = qp->r_psn - 1; 1653 goto send_ack; 1654 } 1655 /* 1656 * Try to send a simple ACK to work around a Mellanox bug 1657 * which doesn't accept a RDMA read response or atomic 1658 * response as an ACK for earlier SENDs or RDMA writes. 1659 */ 1660 if (!(qp->s_flags & RVT_S_RESP_PENDING)) { 1661 spin_unlock_irqrestore(&qp->s_lock, flags); 1662 qp->r_nak_state = 0; 1663 qp->r_ack_psn = qp->s_ack_queue[i].psn - 1; 1664 goto send_ack; 1665 } 1666 /* 1667 * Resend the RDMA read or atomic op which 1668 * ACKs this duplicate request. 1669 */ 1670 qp->s_tail_ack_queue = i; 1671 break; 1672 } 1673 qp->s_ack_state = OP(ACKNOWLEDGE); 1674 qp->s_flags |= RVT_S_RESP_PENDING; 1675 qp->r_nak_state = 0; 1676 qib_schedule_send(qp); 1677 1678unlock_done: 1679 spin_unlock_irqrestore(&qp->s_lock, flags); 1680done: 1681 return 1; 1682 1683send_ack: 1684 return 0; 1685} 1686 1687static inline void qib_update_ack_queue(struct rvt_qp *qp, unsigned n) 1688{ 1689 unsigned next; 1690 1691 next = n + 1; 1692 if (next > QIB_MAX_RDMA_ATOMIC) 1693 next = 0; 1694 qp->s_tail_ack_queue = next; 1695 qp->s_ack_state = OP(ACKNOWLEDGE); 1696} 1697 1698/** 1699 * qib_rc_rcv - process an incoming RC packet 1700 * @rcd: the context pointer 1701 * @hdr: the header of this packet 1702 * @has_grh: true if the header has a GRH 1703 * @data: the packet data 1704 * @tlen: the packet length 1705 * @qp: the QP for this packet 1706 * 1707 * This is called from qib_qp_rcv() to process an incoming RC packet 1708 * for the given QP. 1709 * Called at interrupt level. 1710 */ 1711void qib_rc_rcv(struct qib_ctxtdata *rcd, struct ib_header *hdr, 1712 int has_grh, void *data, u32 tlen, struct rvt_qp *qp) 1713{ 1714 struct qib_ibport *ibp = &rcd->ppd->ibport_data; 1715 struct ib_other_headers *ohdr; 1716 u32 opcode; 1717 u32 hdrsize; 1718 u32 psn; 1719 u32 pad; 1720 struct ib_wc wc; 1721 u32 pmtu = qp->pmtu; 1722 int diff; 1723 struct ib_reth *reth; 1724 unsigned long flags; 1725 int ret; 1726 1727 /* Check for GRH */ 1728 if (!has_grh) { 1729 ohdr = &hdr->u.oth; 1730 hdrsize = 8 + 12; /* LRH + BTH */ 1731 } else { 1732 ohdr = &hdr->u.l.oth; 1733 hdrsize = 8 + 40 + 12; /* LRH + GRH + BTH */ 1734 } 1735 1736 opcode = be32_to_cpu(ohdr->bth[0]); 1737 if (qib_ruc_check_hdr(ibp, hdr, has_grh, qp, opcode)) 1738 return; 1739 1740 psn = be32_to_cpu(ohdr->bth[2]); 1741 opcode >>= 24; 1742 1743 /* 1744 * Process responses (ACKs) before anything else. Note that the 1745 * packet sequence number will be for something in the send work 1746 * queue rather than the expected receive packet sequence number. 1747 * In other words, this QP is the requester. 1748 */ 1749 if (opcode >= OP(RDMA_READ_RESPONSE_FIRST) && 1750 opcode <= OP(ATOMIC_ACKNOWLEDGE)) { 1751 qib_rc_rcv_resp(ibp, ohdr, data, tlen, qp, opcode, psn, 1752 hdrsize, pmtu, rcd); 1753 return; 1754 } 1755 1756 /* Compute 24 bits worth of difference. */ 1757 diff = qib_cmp24(psn, qp->r_psn); 1758 if (unlikely(diff)) { 1759 if (qib_rc_rcv_error(ohdr, data, qp, opcode, psn, diff, rcd)) 1760 return; 1761 goto send_ack; 1762 } 1763 1764 /* Check for opcode sequence errors. */ 1765 switch (qp->r_state) { 1766 case OP(SEND_FIRST): 1767 case OP(SEND_MIDDLE): 1768 if (opcode == OP(SEND_MIDDLE) || 1769 opcode == OP(SEND_LAST) || 1770 opcode == OP(SEND_LAST_WITH_IMMEDIATE)) 1771 break; 1772 goto nack_inv; 1773 1774 case OP(RDMA_WRITE_FIRST): 1775 case OP(RDMA_WRITE_MIDDLE): 1776 if (opcode == OP(RDMA_WRITE_MIDDLE) || 1777 opcode == OP(RDMA_WRITE_LAST) || 1778 opcode == OP(RDMA_WRITE_LAST_WITH_IMMEDIATE)) 1779 break; 1780 goto nack_inv; 1781 1782 default: 1783 if (opcode == OP(SEND_MIDDLE) || 1784 opcode == OP(SEND_LAST) || 1785 opcode == OP(SEND_LAST_WITH_IMMEDIATE) || 1786 opcode == OP(RDMA_WRITE_MIDDLE) || 1787 opcode == OP(RDMA_WRITE_LAST) || 1788 opcode == OP(RDMA_WRITE_LAST_WITH_IMMEDIATE)) 1789 goto nack_inv; 1790 /* 1791 * Note that it is up to the requester to not send a new 1792 * RDMA read or atomic operation before receiving an ACK 1793 * for the previous operation. 1794 */ 1795 break; 1796 } 1797 1798 if (qp->state == IB_QPS_RTR && !(qp->r_flags & RVT_R_COMM_EST)) 1799 rvt_comm_est(qp); 1800 1801 /* OK, process the packet. */ 1802 switch (opcode) { 1803 case OP(SEND_FIRST): 1804 ret = rvt_get_rwqe(qp, false); 1805 if (ret < 0) 1806 goto nack_op_err; 1807 if (!ret) 1808 goto rnr_nak; 1809 qp->r_rcv_len = 0; 1810 fallthrough; 1811 case OP(SEND_MIDDLE): 1812 case OP(RDMA_WRITE_MIDDLE): 1813send_middle: 1814 /* Check for invalid length PMTU or posted rwqe len. */ 1815 if (unlikely(tlen != (hdrsize + pmtu + 4))) 1816 goto nack_inv; 1817 qp->r_rcv_len += pmtu; 1818 if (unlikely(qp->r_rcv_len > qp->r_len)) 1819 goto nack_inv; 1820 rvt_copy_sge(qp, &qp->r_sge, data, pmtu, true, false); 1821 break; 1822 1823 case OP(RDMA_WRITE_LAST_WITH_IMMEDIATE): 1824 /* consume RWQE */ 1825 ret = rvt_get_rwqe(qp, true); 1826 if (ret < 0) 1827 goto nack_op_err; 1828 if (!ret) 1829 goto rnr_nak; 1830 goto send_last_imm; 1831 1832 case OP(SEND_ONLY): 1833 case OP(SEND_ONLY_WITH_IMMEDIATE): 1834 ret = rvt_get_rwqe(qp, false); 1835 if (ret < 0) 1836 goto nack_op_err; 1837 if (!ret) 1838 goto rnr_nak; 1839 qp->r_rcv_len = 0; 1840 if (opcode == OP(SEND_ONLY)) 1841 goto no_immediate_data; 1842 fallthrough; /* for SEND_ONLY_WITH_IMMEDIATE */ 1843 case OP(SEND_LAST_WITH_IMMEDIATE): 1844send_last_imm: 1845 wc.ex.imm_data = ohdr->u.imm_data; 1846 hdrsize += 4; 1847 wc.wc_flags = IB_WC_WITH_IMM; 1848 goto send_last; 1849 case OP(SEND_LAST): 1850 case OP(RDMA_WRITE_LAST): 1851no_immediate_data: 1852 wc.wc_flags = 0; 1853 wc.ex.imm_data = 0; 1854send_last: 1855 /* Get the number of bytes the message was padded by. */ 1856 pad = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3; 1857 /* Check for invalid length. */ 1858 /* XXX LAST len should be >= 1 */ 1859 if (unlikely(tlen < (hdrsize + pad + 4))) 1860 goto nack_inv; 1861 /* Don't count the CRC. */ 1862 tlen -= (hdrsize + pad + 4); 1863 wc.byte_len = tlen + qp->r_rcv_len; 1864 if (unlikely(wc.byte_len > qp->r_len)) 1865 goto nack_inv; 1866 rvt_copy_sge(qp, &qp->r_sge, data, tlen, true, false); 1867 rvt_put_ss(&qp->r_sge); 1868 qp->r_msn++; 1869 if (!test_and_clear_bit(RVT_R_WRID_VALID, &qp->r_aflags)) 1870 break; 1871 wc.wr_id = qp->r_wr_id; 1872 wc.status = IB_WC_SUCCESS; 1873 if (opcode == OP(RDMA_WRITE_LAST_WITH_IMMEDIATE) || 1874 opcode == OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE)) 1875 wc.opcode = IB_WC_RECV_RDMA_WITH_IMM; 1876 else 1877 wc.opcode = IB_WC_RECV; 1878 wc.qp = &qp->ibqp; 1879 wc.src_qp = qp->remote_qpn; 1880 wc.slid = rdma_ah_get_dlid(&qp->remote_ah_attr); 1881 wc.sl = rdma_ah_get_sl(&qp->remote_ah_attr); 1882 /* zero fields that are N/A */ 1883 wc.vendor_err = 0; 1884 wc.pkey_index = 0; 1885 wc.dlid_path_bits = 0; 1886 wc.port_num = 0; 1887 /* Signal completion event if the solicited bit is set. */ 1888 rvt_recv_cq(qp, &wc, ib_bth_is_solicited(ohdr)); 1889 break; 1890 1891 case OP(RDMA_WRITE_FIRST): 1892 case OP(RDMA_WRITE_ONLY): 1893 case OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE): 1894 if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_WRITE))) 1895 goto nack_inv; 1896 /* consume RWQE */ 1897 reth = &ohdr->u.rc.reth; 1898 hdrsize += sizeof(*reth); 1899 qp->r_len = be32_to_cpu(reth->length); 1900 qp->r_rcv_len = 0; 1901 qp->r_sge.sg_list = NULL; 1902 if (qp->r_len != 0) { 1903 u32 rkey = be32_to_cpu(reth->rkey); 1904 u64 vaddr = be64_to_cpu(reth->vaddr); 1905 int ok; 1906 1907 /* Check rkey & NAK */ 1908 ok = rvt_rkey_ok(qp, &qp->r_sge.sge, qp->r_len, vaddr, 1909 rkey, IB_ACCESS_REMOTE_WRITE); 1910 if (unlikely(!ok)) 1911 goto nack_acc; 1912 qp->r_sge.num_sge = 1; 1913 } else { 1914 qp->r_sge.num_sge = 0; 1915 qp->r_sge.sge.mr = NULL; 1916 qp->r_sge.sge.vaddr = NULL; 1917 qp->r_sge.sge.length = 0; 1918 qp->r_sge.sge.sge_length = 0; 1919 } 1920 if (opcode == OP(RDMA_WRITE_FIRST)) 1921 goto send_middle; 1922 else if (opcode == OP(RDMA_WRITE_ONLY)) 1923 goto no_immediate_data; 1924 ret = rvt_get_rwqe(qp, true); 1925 if (ret < 0) 1926 goto nack_op_err; 1927 if (!ret) { 1928 rvt_put_ss(&qp->r_sge); 1929 goto rnr_nak; 1930 } 1931 wc.ex.imm_data = ohdr->u.rc.imm_data; 1932 hdrsize += 4; 1933 wc.wc_flags = IB_WC_WITH_IMM; 1934 goto send_last; 1935 1936 case OP(RDMA_READ_REQUEST): { 1937 struct rvt_ack_entry *e; 1938 u32 len; 1939 u8 next; 1940 1941 if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_READ))) 1942 goto nack_inv; 1943 next = qp->r_head_ack_queue + 1; 1944 /* s_ack_queue is size QIB_MAX_RDMA_ATOMIC+1 so use > not >= */ 1945 if (next > QIB_MAX_RDMA_ATOMIC) 1946 next = 0; 1947 spin_lock_irqsave(&qp->s_lock, flags); 1948 if (unlikely(next == qp->s_tail_ack_queue)) { 1949 if (!qp->s_ack_queue[next].sent) 1950 goto nack_inv_unlck; 1951 qib_update_ack_queue(qp, next); 1952 } 1953 e = &qp->s_ack_queue[qp->r_head_ack_queue]; 1954 if (e->opcode == OP(RDMA_READ_REQUEST) && e->rdma_sge.mr) { 1955 rvt_put_mr(e->rdma_sge.mr); 1956 e->rdma_sge.mr = NULL; 1957 } 1958 reth = &ohdr->u.rc.reth; 1959 len = be32_to_cpu(reth->length); 1960 if (len) { 1961 u32 rkey = be32_to_cpu(reth->rkey); 1962 u64 vaddr = be64_to_cpu(reth->vaddr); 1963 int ok; 1964 1965 /* Check rkey & NAK */ 1966 ok = rvt_rkey_ok(qp, &e->rdma_sge, len, vaddr, 1967 rkey, IB_ACCESS_REMOTE_READ); 1968 if (unlikely(!ok)) 1969 goto nack_acc_unlck; 1970 /* 1971 * Update the next expected PSN. We add 1 later 1972 * below, so only add the remainder here. 1973 */ 1974 qp->r_psn += rvt_div_mtu(qp, len - 1); 1975 } else { 1976 e->rdma_sge.mr = NULL; 1977 e->rdma_sge.vaddr = NULL; 1978 e->rdma_sge.length = 0; 1979 e->rdma_sge.sge_length = 0; 1980 } 1981 e->opcode = opcode; 1982 e->sent = 0; 1983 e->psn = psn; 1984 e->lpsn = qp->r_psn; 1985 /* 1986 * We need to increment the MSN here instead of when we 1987 * finish sending the result since a duplicate request would 1988 * increment it more than once. 1989 */ 1990 qp->r_msn++; 1991 qp->r_psn++; 1992 qp->r_state = opcode; 1993 qp->r_nak_state = 0; 1994 qp->r_head_ack_queue = next; 1995 1996 /* Schedule the send tasklet. */ 1997 qp->s_flags |= RVT_S_RESP_PENDING; 1998 qib_schedule_send(qp); 1999 2000 goto sunlock; 2001 } 2002 2003 case OP(COMPARE_SWAP): 2004 case OP(FETCH_ADD): { 2005 struct ib_atomic_eth *ateth; 2006 struct rvt_ack_entry *e; 2007 u64 vaddr; 2008 atomic64_t *maddr; 2009 u64 sdata; 2010 u32 rkey; 2011 u8 next; 2012 2013 if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_ATOMIC))) 2014 goto nack_inv; 2015 next = qp->r_head_ack_queue + 1; 2016 if (next > QIB_MAX_RDMA_ATOMIC) 2017 next = 0; 2018 spin_lock_irqsave(&qp->s_lock, flags); 2019 if (unlikely(next == qp->s_tail_ack_queue)) { 2020 if (!qp->s_ack_queue[next].sent) 2021 goto nack_inv_unlck; 2022 qib_update_ack_queue(qp, next); 2023 } 2024 e = &qp->s_ack_queue[qp->r_head_ack_queue]; 2025 if (e->opcode == OP(RDMA_READ_REQUEST) && e->rdma_sge.mr) { 2026 rvt_put_mr(e->rdma_sge.mr); 2027 e->rdma_sge.mr = NULL; 2028 } 2029 ateth = &ohdr->u.atomic_eth; 2030 vaddr = get_ib_ateth_vaddr(ateth); 2031 if (unlikely(vaddr & (sizeof(u64) - 1))) 2032 goto nack_inv_unlck; 2033 rkey = be32_to_cpu(ateth->rkey); 2034 /* Check rkey & NAK */ 2035 if (unlikely(!rvt_rkey_ok(qp, &qp->r_sge.sge, sizeof(u64), 2036 vaddr, rkey, 2037 IB_ACCESS_REMOTE_ATOMIC))) 2038 goto nack_acc_unlck; 2039 /* Perform atomic OP and save result. */ 2040 maddr = (atomic64_t *) qp->r_sge.sge.vaddr; 2041 sdata = get_ib_ateth_swap(ateth); 2042 e->atomic_data = (opcode == OP(FETCH_ADD)) ? 2043 (u64) atomic64_add_return(sdata, maddr) - sdata : 2044 (u64) cmpxchg((u64 *) qp->r_sge.sge.vaddr, 2045 get_ib_ateth_compare(ateth), 2046 sdata); 2047 rvt_put_mr(qp->r_sge.sge.mr); 2048 qp->r_sge.num_sge = 0; 2049 e->opcode = opcode; 2050 e->sent = 0; 2051 e->psn = psn; 2052 e->lpsn = psn; 2053 qp->r_msn++; 2054 qp->r_psn++; 2055 qp->r_state = opcode; 2056 qp->r_nak_state = 0; 2057 qp->r_head_ack_queue = next; 2058 2059 /* Schedule the send tasklet. */ 2060 qp->s_flags |= RVT_S_RESP_PENDING; 2061 qib_schedule_send(qp); 2062 2063 goto sunlock; 2064 } 2065 2066 default: 2067 /* NAK unknown opcodes. */ 2068 goto nack_inv; 2069 } 2070 qp->r_psn++; 2071 qp->r_state = opcode; 2072 qp->r_ack_psn = psn; 2073 qp->r_nak_state = 0; 2074 /* Send an ACK if requested or required. */ 2075 if (psn & (1 << 31)) 2076 goto send_ack; 2077 return; 2078 2079rnr_nak: 2080 qp->r_nak_state = IB_RNR_NAK | qp->r_min_rnr_timer; 2081 qp->r_ack_psn = qp->r_psn; 2082 /* Queue RNR NAK for later */ 2083 if (list_empty(&qp->rspwait)) { 2084 qp->r_flags |= RVT_R_RSP_NAK; 2085 rvt_get_qp(qp); 2086 list_add_tail(&qp->rspwait, &rcd->qp_wait_list); 2087 } 2088 return; 2089 2090nack_op_err: 2091 rvt_rc_error(qp, IB_WC_LOC_QP_OP_ERR); 2092 qp->r_nak_state = IB_NAK_REMOTE_OPERATIONAL_ERROR; 2093 qp->r_ack_psn = qp->r_psn; 2094 /* Queue NAK for later */ 2095 if (list_empty(&qp->rspwait)) { 2096 qp->r_flags |= RVT_R_RSP_NAK; 2097 rvt_get_qp(qp); 2098 list_add_tail(&qp->rspwait, &rcd->qp_wait_list); 2099 } 2100 return; 2101 2102nack_inv_unlck: 2103 spin_unlock_irqrestore(&qp->s_lock, flags); 2104nack_inv: 2105 rvt_rc_error(qp, IB_WC_LOC_QP_OP_ERR); 2106 qp->r_nak_state = IB_NAK_INVALID_REQUEST; 2107 qp->r_ack_psn = qp->r_psn; 2108 /* Queue NAK for later */ 2109 if (list_empty(&qp->rspwait)) { 2110 qp->r_flags |= RVT_R_RSP_NAK; 2111 rvt_get_qp(qp); 2112 list_add_tail(&qp->rspwait, &rcd->qp_wait_list); 2113 } 2114 return; 2115 2116nack_acc_unlck: 2117 spin_unlock_irqrestore(&qp->s_lock, flags); 2118nack_acc: 2119 rvt_rc_error(qp, IB_WC_LOC_PROT_ERR); 2120 qp->r_nak_state = IB_NAK_REMOTE_ACCESS_ERROR; 2121 qp->r_ack_psn = qp->r_psn; 2122send_ack: 2123 qib_send_rc_ack(qp); 2124 return; 2125 2126sunlock: 2127 spin_unlock_irqrestore(&qp->s_lock, flags); 2128} 2129