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 static void rc_timeout(unsigned long arg); 42 43 static u32 restart_sge(struct rvt_sge_state *ss, struct rvt_swqe *wqe, 44 u32 psn, u32 pmtu) 45 { 46 u32 len; 47 48 len = ((psn - wqe->psn) & QIB_PSN_MASK) * pmtu; 49 ss->sge = wqe->sg_list[0]; 50 ss->sg_list = wqe->sg_list + 1; 51 ss->num_sge = wqe->wr.num_sge; 52 ss->total_len = wqe->length; 53 qib_skip_sge(ss, len, 0); 54 return wqe->length - len; 55 } 56 57 static void start_timer(struct rvt_qp *qp) 58 { 59 qp->s_flags |= RVT_S_TIMER; 60 qp->s_timer.function = rc_timeout; 61 /* 4.096 usec. * (1 << qp->timeout) */ 62 qp->s_timer.expires = jiffies + qp->timeout_jiffies; 63 add_timer(&qp->s_timer); 64 } 65 66 /** 67 * qib_make_rc_ack - construct a response packet (ACK, NAK, or RDMA read) 68 * @dev: the device for this QP 69 * @qp: a pointer to the QP 70 * @ohdr: a pointer to the IB header being constructed 71 * @pmtu: the path MTU 72 * 73 * Return 1 if constructed; otherwise, return 0. 74 * Note that we are in the responder's side of the QP context. 75 * Note the QP s_lock must be held. 76 */ 77 static int qib_make_rc_ack(struct qib_ibdev *dev, struct rvt_qp *qp, 78 struct ib_other_headers *ohdr, u32 pmtu) 79 { 80 struct rvt_ack_entry *e; 81 u32 hwords; 82 u32 len; 83 u32 bth0; 84 u32 bth2; 85 86 /* Don't send an ACK if we aren't supposed to. */ 87 if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK)) 88 goto bail; 89 90 /* header size in 32-bit words LRH+BTH = (8+12)/4. */ 91 hwords = 5; 92 93 switch (qp->s_ack_state) { 94 case OP(RDMA_READ_RESPONSE_LAST): 95 case OP(RDMA_READ_RESPONSE_ONLY): 96 e = &qp->s_ack_queue[qp->s_tail_ack_queue]; 97 if (e->rdma_sge.mr) { 98 rvt_put_mr(e->rdma_sge.mr); 99 e->rdma_sge.mr = NULL; 100 } 101 /* FALLTHROUGH */ 102 case OP(ATOMIC_ACKNOWLEDGE): 103 /* 104 * We can increment the tail pointer now that the last 105 * response has been sent instead of only being 106 * constructed. 107 */ 108 if (++qp->s_tail_ack_queue > QIB_MAX_RDMA_ATOMIC) 109 qp->s_tail_ack_queue = 0; 110 /* FALLTHROUGH */ 111 case OP(SEND_ONLY): 112 case OP(ACKNOWLEDGE): 113 /* Check for no next entry in the queue. */ 114 if (qp->r_head_ack_queue == qp->s_tail_ack_queue) { 115 if (qp->s_flags & RVT_S_ACK_PENDING) 116 goto normal; 117 goto bail; 118 } 119 120 e = &qp->s_ack_queue[qp->s_tail_ack_queue]; 121 if (e->opcode == OP(RDMA_READ_REQUEST)) { 122 /* 123 * If a RDMA read response is being resent and 124 * we haven't seen the duplicate request yet, 125 * then stop sending the remaining responses the 126 * responder has seen until the requester resends it. 127 */ 128 len = e->rdma_sge.sge_length; 129 if (len && !e->rdma_sge.mr) { 130 qp->s_tail_ack_queue = qp->r_head_ack_queue; 131 goto bail; 132 } 133 /* Copy SGE state in case we need to resend */ 134 qp->s_rdma_mr = e->rdma_sge.mr; 135 if (qp->s_rdma_mr) 136 rvt_get_mr(qp->s_rdma_mr); 137 qp->s_ack_rdma_sge.sge = e->rdma_sge; 138 qp->s_ack_rdma_sge.num_sge = 1; 139 qp->s_cur_sge = &qp->s_ack_rdma_sge; 140 if (len > pmtu) { 141 len = pmtu; 142 qp->s_ack_state = OP(RDMA_READ_RESPONSE_FIRST); 143 } else { 144 qp->s_ack_state = OP(RDMA_READ_RESPONSE_ONLY); 145 e->sent = 1; 146 } 147 ohdr->u.aeth = qib_compute_aeth(qp); 148 hwords++; 149 qp->s_ack_rdma_psn = e->psn; 150 bth2 = qp->s_ack_rdma_psn++ & QIB_PSN_MASK; 151 } else { 152 /* COMPARE_SWAP or FETCH_ADD */ 153 qp->s_cur_sge = NULL; 154 len = 0; 155 qp->s_ack_state = OP(ATOMIC_ACKNOWLEDGE); 156 ohdr->u.at.aeth = qib_compute_aeth(qp); 157 ib_u64_put(e->atomic_data, &ohdr->u.at.atomic_ack_eth); 158 hwords += sizeof(ohdr->u.at) / sizeof(u32); 159 bth2 = e->psn & QIB_PSN_MASK; 160 e->sent = 1; 161 } 162 bth0 = qp->s_ack_state << 24; 163 break; 164 165 case OP(RDMA_READ_RESPONSE_FIRST): 166 qp->s_ack_state = OP(RDMA_READ_RESPONSE_MIDDLE); 167 /* FALLTHROUGH */ 168 case OP(RDMA_READ_RESPONSE_MIDDLE): 169 qp->s_cur_sge = &qp->s_ack_rdma_sge; 170 qp->s_rdma_mr = qp->s_ack_rdma_sge.sge.mr; 171 if (qp->s_rdma_mr) 172 rvt_get_mr(qp->s_rdma_mr); 173 len = qp->s_ack_rdma_sge.sge.sge_length; 174 if (len > pmtu) 175 len = pmtu; 176 else { 177 ohdr->u.aeth = qib_compute_aeth(qp); 178 hwords++; 179 qp->s_ack_state = OP(RDMA_READ_RESPONSE_LAST); 180 e = &qp->s_ack_queue[qp->s_tail_ack_queue]; 181 e->sent = 1; 182 } 183 bth0 = qp->s_ack_state << 24; 184 bth2 = qp->s_ack_rdma_psn++ & QIB_PSN_MASK; 185 break; 186 187 default: 188 normal: 189 /* 190 * Send a regular ACK. 191 * Set the s_ack_state so we wait until after sending 192 * the ACK before setting s_ack_state to ACKNOWLEDGE 193 * (see above). 194 */ 195 qp->s_ack_state = OP(SEND_ONLY); 196 qp->s_flags &= ~RVT_S_ACK_PENDING; 197 qp->s_cur_sge = NULL; 198 if (qp->s_nak_state) 199 ohdr->u.aeth = 200 cpu_to_be32((qp->r_msn & QIB_MSN_MASK) | 201 (qp->s_nak_state << 202 QIB_AETH_CREDIT_SHIFT)); 203 else 204 ohdr->u.aeth = qib_compute_aeth(qp); 205 hwords++; 206 len = 0; 207 bth0 = OP(ACKNOWLEDGE) << 24; 208 bth2 = qp->s_ack_psn & QIB_PSN_MASK; 209 } 210 qp->s_rdma_ack_cnt++; 211 qp->s_hdrwords = hwords; 212 qp->s_cur_size = len; 213 qib_make_ruc_header(qp, ohdr, bth0, bth2); 214 return 1; 215 216 bail: 217 qp->s_ack_state = OP(ACKNOWLEDGE); 218 qp->s_flags &= ~(RVT_S_RESP_PENDING | RVT_S_ACK_PENDING); 219 return 0; 220 } 221 222 /** 223 * qib_make_rc_req - construct a request packet (SEND, RDMA r/w, ATOMIC) 224 * @qp: a pointer to the QP 225 * 226 * Assumes the s_lock is held. 227 * 228 * Return 1 if constructed; otherwise, return 0. 229 */ 230 int qib_make_rc_req(struct rvt_qp *qp, unsigned long *flags) 231 { 232 struct qib_qp_priv *priv = qp->priv; 233 struct qib_ibdev *dev = to_idev(qp->ibqp.device); 234 struct ib_other_headers *ohdr; 235 struct rvt_sge_state *ss; 236 struct rvt_swqe *wqe; 237 u32 hwords; 238 u32 len; 239 u32 bth0; 240 u32 bth2; 241 u32 pmtu = qp->pmtu; 242 char newreq; 243 int ret = 0; 244 int delta; 245 246 ohdr = &priv->s_hdr->u.oth; 247 if (qp->remote_ah_attr.ah_flags & IB_AH_GRH) 248 ohdr = &priv->s_hdr->u.l.oth; 249 250 /* Sending responses has higher priority over sending requests. */ 251 if ((qp->s_flags & RVT_S_RESP_PENDING) && 252 qib_make_rc_ack(dev, qp, ohdr, pmtu)) 253 goto done; 254 255 if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_SEND_OK)) { 256 if (!(ib_rvt_state_ops[qp->state] & RVT_FLUSH_SEND)) 257 goto bail; 258 /* We are in the error state, flush the work request. */ 259 smp_read_barrier_depends(); /* see post_one_send() */ 260 if (qp->s_last == ACCESS_ONCE(qp->s_head)) 261 goto bail; 262 /* If DMAs are in progress, we can't flush immediately. */ 263 if (atomic_read(&priv->s_dma_busy)) { 264 qp->s_flags |= RVT_S_WAIT_DMA; 265 goto bail; 266 } 267 wqe = rvt_get_swqe_ptr(qp, qp->s_last); 268 qib_send_complete(qp, wqe, qp->s_last != qp->s_acked ? 269 IB_WC_SUCCESS : IB_WC_WR_FLUSH_ERR); 270 /* will get called again */ 271 goto done; 272 } 273 274 if (qp->s_flags & (RVT_S_WAIT_RNR | RVT_S_WAIT_ACK)) 275 goto bail; 276 277 if (qib_cmp24(qp->s_psn, qp->s_sending_hpsn) <= 0) { 278 if (qib_cmp24(qp->s_sending_psn, qp->s_sending_hpsn) <= 0) { 279 qp->s_flags |= RVT_S_WAIT_PSN; 280 goto bail; 281 } 282 qp->s_sending_psn = qp->s_psn; 283 qp->s_sending_hpsn = qp->s_psn - 1; 284 } 285 286 /* header size in 32-bit words LRH+BTH = (8+12)/4. */ 287 hwords = 5; 288 bth0 = 0; 289 290 /* Send a request. */ 291 wqe = rvt_get_swqe_ptr(qp, qp->s_cur); 292 switch (qp->s_state) { 293 default: 294 if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_NEXT_SEND_OK)) 295 goto bail; 296 /* 297 * Resend an old request or start a new one. 298 * 299 * We keep track of the current SWQE so that 300 * we don't reset the "furthest progress" state 301 * if we need to back up. 302 */ 303 newreq = 0; 304 if (qp->s_cur == qp->s_tail) { 305 /* Check if send work queue is empty. */ 306 if (qp->s_tail == qp->s_head) 307 goto bail; 308 /* 309 * If a fence is requested, wait for previous 310 * RDMA read and atomic operations to finish. 311 */ 312 if ((wqe->wr.send_flags & IB_SEND_FENCE) && 313 qp->s_num_rd_atomic) { 314 qp->s_flags |= RVT_S_WAIT_FENCE; 315 goto bail; 316 } 317 newreq = 1; 318 qp->s_psn = wqe->psn; 319 } 320 /* 321 * Note that we have to be careful not to modify the 322 * original work request since we may need to resend 323 * it. 324 */ 325 len = wqe->length; 326 ss = &qp->s_sge; 327 bth2 = qp->s_psn & QIB_PSN_MASK; 328 switch (wqe->wr.opcode) { 329 case IB_WR_SEND: 330 case IB_WR_SEND_WITH_IMM: 331 /* If no credit, return. */ 332 if (!(qp->s_flags & RVT_S_UNLIMITED_CREDIT) && 333 qib_cmp24(wqe->ssn, qp->s_lsn + 1) > 0) { 334 qp->s_flags |= RVT_S_WAIT_SSN_CREDIT; 335 goto bail; 336 } 337 if (len > pmtu) { 338 qp->s_state = OP(SEND_FIRST); 339 len = pmtu; 340 break; 341 } 342 if (wqe->wr.opcode == IB_WR_SEND) 343 qp->s_state = OP(SEND_ONLY); 344 else { 345 qp->s_state = OP(SEND_ONLY_WITH_IMMEDIATE); 346 /* Immediate data comes after the BTH */ 347 ohdr->u.imm_data = wqe->wr.ex.imm_data; 348 hwords += 1; 349 } 350 if (wqe->wr.send_flags & IB_SEND_SOLICITED) 351 bth0 |= IB_BTH_SOLICITED; 352 bth2 |= IB_BTH_REQ_ACK; 353 if (++qp->s_cur == qp->s_size) 354 qp->s_cur = 0; 355 break; 356 357 case IB_WR_RDMA_WRITE: 358 if (newreq && !(qp->s_flags & RVT_S_UNLIMITED_CREDIT)) 359 qp->s_lsn++; 360 /* FALLTHROUGH */ 361 case IB_WR_RDMA_WRITE_WITH_IMM: 362 /* If no credit, return. */ 363 if (!(qp->s_flags & RVT_S_UNLIMITED_CREDIT) && 364 qib_cmp24(wqe->ssn, qp->s_lsn + 1) > 0) { 365 qp->s_flags |= RVT_S_WAIT_SSN_CREDIT; 366 goto bail; 367 } 368 369 ohdr->u.rc.reth.vaddr = 370 cpu_to_be64(wqe->rdma_wr.remote_addr); 371 ohdr->u.rc.reth.rkey = 372 cpu_to_be32(wqe->rdma_wr.rkey); 373 ohdr->u.rc.reth.length = cpu_to_be32(len); 374 hwords += sizeof(struct ib_reth) / sizeof(u32); 375 if (len > pmtu) { 376 qp->s_state = OP(RDMA_WRITE_FIRST); 377 len = pmtu; 378 break; 379 } 380 if (wqe->rdma_wr.wr.opcode == IB_WR_RDMA_WRITE) 381 qp->s_state = OP(RDMA_WRITE_ONLY); 382 else { 383 qp->s_state = OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE); 384 /* Immediate data comes after RETH */ 385 ohdr->u.rc.imm_data = 386 wqe->rdma_wr.wr.ex.imm_data; 387 hwords += 1; 388 if (wqe->rdma_wr.wr.send_flags & IB_SEND_SOLICITED) 389 bth0 |= IB_BTH_SOLICITED; 390 } 391 bth2 |= IB_BTH_REQ_ACK; 392 if (++qp->s_cur == qp->s_size) 393 qp->s_cur = 0; 394 break; 395 396 case IB_WR_RDMA_READ: 397 /* 398 * Don't allow more operations to be started 399 * than the QP limits allow. 400 */ 401 if (newreq) { 402 if (qp->s_num_rd_atomic >= 403 qp->s_max_rd_atomic) { 404 qp->s_flags |= RVT_S_WAIT_RDMAR; 405 goto bail; 406 } 407 qp->s_num_rd_atomic++; 408 if (!(qp->s_flags & RVT_S_UNLIMITED_CREDIT)) 409 qp->s_lsn++; 410 } 411 412 ohdr->u.rc.reth.vaddr = 413 cpu_to_be64(wqe->rdma_wr.remote_addr); 414 ohdr->u.rc.reth.rkey = 415 cpu_to_be32(wqe->rdma_wr.rkey); 416 ohdr->u.rc.reth.length = cpu_to_be32(len); 417 qp->s_state = OP(RDMA_READ_REQUEST); 418 hwords += sizeof(ohdr->u.rc.reth) / sizeof(u32); 419 ss = NULL; 420 len = 0; 421 bth2 |= IB_BTH_REQ_ACK; 422 if (++qp->s_cur == qp->s_size) 423 qp->s_cur = 0; 424 break; 425 426 case IB_WR_ATOMIC_CMP_AND_SWP: 427 case IB_WR_ATOMIC_FETCH_AND_ADD: 428 /* 429 * Don't allow more operations to be started 430 * than the QP limits allow. 431 */ 432 if (newreq) { 433 if (qp->s_num_rd_atomic >= 434 qp->s_max_rd_atomic) { 435 qp->s_flags |= RVT_S_WAIT_RDMAR; 436 goto bail; 437 } 438 qp->s_num_rd_atomic++; 439 if (!(qp->s_flags & RVT_S_UNLIMITED_CREDIT)) 440 qp->s_lsn++; 441 } 442 if (wqe->atomic_wr.wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP) { 443 qp->s_state = OP(COMPARE_SWAP); 444 put_ib_ateth_swap(wqe->atomic_wr.swap, 445 &ohdr->u.atomic_eth); 446 put_ib_ateth_swap(wqe->atomic_wr.compare_add, 447 &ohdr->u.atomic_eth); 448 } else { 449 qp->s_state = OP(FETCH_ADD); 450 put_ib_ateth_swap(wqe->atomic_wr.compare_add, 451 &ohdr->u.atomic_eth); 452 put_ib_ateth_swap(0, &ohdr->u.atomic_eth); 453 } 454 put_ib_ateth_vaddr(wqe->atomic_wr.remote_addr, 455 &ohdr->u.atomic_eth); 456 ohdr->u.atomic_eth.rkey = cpu_to_be32( 457 wqe->atomic_wr.rkey); 458 hwords += sizeof(struct ib_atomic_eth) / sizeof(u32); 459 ss = NULL; 460 len = 0; 461 bth2 |= IB_BTH_REQ_ACK; 462 if (++qp->s_cur == qp->s_size) 463 qp->s_cur = 0; 464 break; 465 466 default: 467 goto bail; 468 } 469 qp->s_sge.sge = wqe->sg_list[0]; 470 qp->s_sge.sg_list = wqe->sg_list + 1; 471 qp->s_sge.num_sge = wqe->wr.num_sge; 472 qp->s_sge.total_len = wqe->length; 473 qp->s_len = wqe->length; 474 if (newreq) { 475 qp->s_tail++; 476 if (qp->s_tail >= qp->s_size) 477 qp->s_tail = 0; 478 } 479 if (wqe->wr.opcode == IB_WR_RDMA_READ) 480 qp->s_psn = wqe->lpsn + 1; 481 else 482 qp->s_psn++; 483 break; 484 485 case OP(RDMA_READ_RESPONSE_FIRST): 486 /* 487 * qp->s_state is normally set to the opcode of the 488 * last packet constructed for new requests and therefore 489 * is never set to RDMA read response. 490 * RDMA_READ_RESPONSE_FIRST is used by the ACK processing 491 * thread to indicate a SEND needs to be restarted from an 492 * earlier PSN without interferring with the sending thread. 493 * See qib_restart_rc(). 494 */ 495 qp->s_len = restart_sge(&qp->s_sge, wqe, qp->s_psn, pmtu); 496 /* FALLTHROUGH */ 497 case OP(SEND_FIRST): 498 qp->s_state = OP(SEND_MIDDLE); 499 /* FALLTHROUGH */ 500 case OP(SEND_MIDDLE): 501 bth2 = qp->s_psn++ & QIB_PSN_MASK; 502 ss = &qp->s_sge; 503 len = qp->s_len; 504 if (len > pmtu) { 505 len = pmtu; 506 break; 507 } 508 if (wqe->wr.opcode == IB_WR_SEND) 509 qp->s_state = OP(SEND_LAST); 510 else { 511 qp->s_state = OP(SEND_LAST_WITH_IMMEDIATE); 512 /* Immediate data comes after the BTH */ 513 ohdr->u.imm_data = wqe->wr.ex.imm_data; 514 hwords += 1; 515 } 516 if (wqe->wr.send_flags & IB_SEND_SOLICITED) 517 bth0 |= IB_BTH_SOLICITED; 518 bth2 |= IB_BTH_REQ_ACK; 519 qp->s_cur++; 520 if (qp->s_cur >= qp->s_size) 521 qp->s_cur = 0; 522 break; 523 524 case OP(RDMA_READ_RESPONSE_LAST): 525 /* 526 * qp->s_state is normally set to the opcode of the 527 * last packet constructed for new requests and therefore 528 * is never set to RDMA read response. 529 * RDMA_READ_RESPONSE_LAST is used by the ACK processing 530 * thread to indicate a RDMA write needs to be restarted from 531 * an earlier PSN without interferring with the sending thread. 532 * See qib_restart_rc(). 533 */ 534 qp->s_len = restart_sge(&qp->s_sge, wqe, qp->s_psn, pmtu); 535 /* FALLTHROUGH */ 536 case OP(RDMA_WRITE_FIRST): 537 qp->s_state = OP(RDMA_WRITE_MIDDLE); 538 /* FALLTHROUGH */ 539 case OP(RDMA_WRITE_MIDDLE): 540 bth2 = qp->s_psn++ & QIB_PSN_MASK; 541 ss = &qp->s_sge; 542 len = qp->s_len; 543 if (len > pmtu) { 544 len = pmtu; 545 break; 546 } 547 if (wqe->wr.opcode == IB_WR_RDMA_WRITE) 548 qp->s_state = OP(RDMA_WRITE_LAST); 549 else { 550 qp->s_state = OP(RDMA_WRITE_LAST_WITH_IMMEDIATE); 551 /* Immediate data comes after the BTH */ 552 ohdr->u.imm_data = wqe->wr.ex.imm_data; 553 hwords += 1; 554 if (wqe->wr.send_flags & IB_SEND_SOLICITED) 555 bth0 |= IB_BTH_SOLICITED; 556 } 557 bth2 |= IB_BTH_REQ_ACK; 558 qp->s_cur++; 559 if (qp->s_cur >= qp->s_size) 560 qp->s_cur = 0; 561 break; 562 563 case OP(RDMA_READ_RESPONSE_MIDDLE): 564 /* 565 * qp->s_state is normally set to the opcode of the 566 * last packet constructed for new requests and therefore 567 * is never set to RDMA read response. 568 * RDMA_READ_RESPONSE_MIDDLE is used by the ACK processing 569 * thread to indicate a RDMA read needs to be restarted from 570 * an earlier PSN without interferring with the sending thread. 571 * See qib_restart_rc(). 572 */ 573 len = ((qp->s_psn - wqe->psn) & QIB_PSN_MASK) * pmtu; 574 ohdr->u.rc.reth.vaddr = 575 cpu_to_be64(wqe->rdma_wr.remote_addr + len); 576 ohdr->u.rc.reth.rkey = 577 cpu_to_be32(wqe->rdma_wr.rkey); 578 ohdr->u.rc.reth.length = cpu_to_be32(wqe->length - len); 579 qp->s_state = OP(RDMA_READ_REQUEST); 580 hwords += sizeof(ohdr->u.rc.reth) / sizeof(u32); 581 bth2 = (qp->s_psn & QIB_PSN_MASK) | IB_BTH_REQ_ACK; 582 qp->s_psn = wqe->lpsn + 1; 583 ss = NULL; 584 len = 0; 585 qp->s_cur++; 586 if (qp->s_cur == qp->s_size) 587 qp->s_cur = 0; 588 break; 589 } 590 qp->s_sending_hpsn = bth2; 591 delta = (((int) bth2 - (int) wqe->psn) << 8) >> 8; 592 if (delta && delta % QIB_PSN_CREDIT == 0) 593 bth2 |= IB_BTH_REQ_ACK; 594 if (qp->s_flags & RVT_S_SEND_ONE) { 595 qp->s_flags &= ~RVT_S_SEND_ONE; 596 qp->s_flags |= RVT_S_WAIT_ACK; 597 bth2 |= IB_BTH_REQ_ACK; 598 } 599 qp->s_len -= len; 600 qp->s_hdrwords = hwords; 601 qp->s_cur_sge = ss; 602 qp->s_cur_size = len; 603 qib_make_ruc_header(qp, ohdr, bth0 | (qp->s_state << 24), bth2); 604 done: 605 return 1; 606 bail: 607 qp->s_flags &= ~RVT_S_BUSY; 608 return ret; 609 } 610 611 /** 612 * qib_send_rc_ack - Construct an ACK packet and send it 613 * @qp: a pointer to the QP 614 * 615 * This is called from qib_rc_rcv() and qib_kreceive(). 616 * Note that RDMA reads and atomics are handled in the 617 * send side QP state and tasklet. 618 */ 619 void qib_send_rc_ack(struct rvt_qp *qp) 620 { 621 struct qib_devdata *dd = dd_from_ibdev(qp->ibqp.device); 622 struct qib_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num); 623 struct qib_pportdata *ppd = ppd_from_ibp(ibp); 624 u64 pbc; 625 u16 lrh0; 626 u32 bth0; 627 u32 hwords; 628 u32 pbufn; 629 u32 __iomem *piobuf; 630 struct ib_header hdr; 631 struct ib_other_headers *ohdr; 632 u32 control; 633 unsigned long flags; 634 635 spin_lock_irqsave(&qp->s_lock, flags); 636 637 if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK)) 638 goto unlock; 639 640 /* Don't send ACK or NAK if a RDMA read or atomic is pending. */ 641 if ((qp->s_flags & RVT_S_RESP_PENDING) || qp->s_rdma_ack_cnt) 642 goto queue_ack; 643 644 /* Construct the header with s_lock held so APM doesn't change it. */ 645 ohdr = &hdr.u.oth; 646 lrh0 = QIB_LRH_BTH; 647 /* header size in 32-bit words LRH+BTH+AETH = (8+12+4)/4. */ 648 hwords = 6; 649 if (unlikely(qp->remote_ah_attr.ah_flags & IB_AH_GRH)) { 650 hwords += qib_make_grh(ibp, &hdr.u.l.grh, 651 &qp->remote_ah_attr.grh, hwords, 0); 652 ohdr = &hdr.u.l.oth; 653 lrh0 = QIB_LRH_GRH; 654 } 655 /* read pkey_index w/o lock (its atomic) */ 656 bth0 = qib_get_pkey(ibp, qp->s_pkey_index) | (OP(ACKNOWLEDGE) << 24); 657 if (qp->s_mig_state == IB_MIG_MIGRATED) 658 bth0 |= IB_BTH_MIG_REQ; 659 if (qp->r_nak_state) 660 ohdr->u.aeth = cpu_to_be32((qp->r_msn & QIB_MSN_MASK) | 661 (qp->r_nak_state << 662 QIB_AETH_CREDIT_SHIFT)); 663 else 664 ohdr->u.aeth = qib_compute_aeth(qp); 665 lrh0 |= ibp->sl_to_vl[qp->remote_ah_attr.sl] << 12 | 666 qp->remote_ah_attr.sl << 4; 667 hdr.lrh[0] = cpu_to_be16(lrh0); 668 hdr.lrh[1] = cpu_to_be16(qp->remote_ah_attr.dlid); 669 hdr.lrh[2] = cpu_to_be16(hwords + SIZE_OF_CRC); 670 hdr.lrh[3] = cpu_to_be16(ppd->lid | qp->remote_ah_attr.src_path_bits); 671 ohdr->bth[0] = cpu_to_be32(bth0); 672 ohdr->bth[1] = cpu_to_be32(qp->remote_qpn); 673 ohdr->bth[2] = cpu_to_be32(qp->r_ack_psn & QIB_PSN_MASK); 674 675 spin_unlock_irqrestore(&qp->s_lock, flags); 676 677 /* Don't try to send ACKs if the link isn't ACTIVE */ 678 if (!(ppd->lflags & QIBL_LINKACTIVE)) 679 goto done; 680 681 control = dd->f_setpbc_control(ppd, hwords + SIZE_OF_CRC, 682 qp->s_srate, lrh0 >> 12); 683 /* length is + 1 for the control dword */ 684 pbc = ((u64) control << 32) | (hwords + 1); 685 686 piobuf = dd->f_getsendbuf(ppd, pbc, &pbufn); 687 if (!piobuf) { 688 /* 689 * We are out of PIO buffers at the moment. 690 * Pass responsibility for sending the ACK to the 691 * send tasklet so that when a PIO buffer becomes 692 * available, the ACK is sent ahead of other outgoing 693 * packets. 694 */ 695 spin_lock_irqsave(&qp->s_lock, flags); 696 goto queue_ack; 697 } 698 699 /* 700 * Write the pbc. 701 * We have to flush after the PBC for correctness 702 * on some cpus or WC buffer can be written out of order. 703 */ 704 writeq(pbc, piobuf); 705 706 if (dd->flags & QIB_PIO_FLUSH_WC) { 707 u32 *hdrp = (u32 *) &hdr; 708 709 qib_flush_wc(); 710 qib_pio_copy(piobuf + 2, hdrp, hwords - 1); 711 qib_flush_wc(); 712 __raw_writel(hdrp[hwords - 1], piobuf + hwords + 1); 713 } else 714 qib_pio_copy(piobuf + 2, (u32 *) &hdr, hwords); 715 716 if (dd->flags & QIB_USE_SPCL_TRIG) { 717 u32 spcl_off = (pbufn >= dd->piobcnt2k) ? 2047 : 1023; 718 719 qib_flush_wc(); 720 __raw_writel(0xaebecede, piobuf + spcl_off); 721 } 722 723 qib_flush_wc(); 724 qib_sendbuf_done(dd, pbufn); 725 726 this_cpu_inc(ibp->pmastats->n_unicast_xmit); 727 goto done; 728 729 queue_ack: 730 if (ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK) { 731 this_cpu_inc(*ibp->rvp.rc_qacks); 732 qp->s_flags |= RVT_S_ACK_PENDING | RVT_S_RESP_PENDING; 733 qp->s_nak_state = qp->r_nak_state; 734 qp->s_ack_psn = qp->r_ack_psn; 735 736 /* Schedule the send tasklet. */ 737 qib_schedule_send(qp); 738 } 739 unlock: 740 spin_unlock_irqrestore(&qp->s_lock, flags); 741 done: 742 return; 743 } 744 745 /** 746 * reset_psn - reset the QP state to send starting from PSN 747 * @qp: the QP 748 * @psn: the packet sequence number to restart at 749 * 750 * This is called from qib_rc_rcv() to process an incoming RC ACK 751 * for the given QP. 752 * Called at interrupt level with the QP s_lock held. 753 */ 754 static void reset_psn(struct rvt_qp *qp, u32 psn) 755 { 756 u32 n = qp->s_acked; 757 struct rvt_swqe *wqe = rvt_get_swqe_ptr(qp, n); 758 u32 opcode; 759 760 qp->s_cur = n; 761 762 /* 763 * If we are starting the request from the beginning, 764 * let the normal send code handle initialization. 765 */ 766 if (qib_cmp24(psn, wqe->psn) <= 0) { 767 qp->s_state = OP(SEND_LAST); 768 goto done; 769 } 770 771 /* Find the work request opcode corresponding to the given PSN. */ 772 opcode = wqe->wr.opcode; 773 for (;;) { 774 int diff; 775 776 if (++n == qp->s_size) 777 n = 0; 778 if (n == qp->s_tail) 779 break; 780 wqe = rvt_get_swqe_ptr(qp, n); 781 diff = qib_cmp24(psn, wqe->psn); 782 if (diff < 0) 783 break; 784 qp->s_cur = n; 785 /* 786 * If we are starting the request from the beginning, 787 * let the normal send code handle initialization. 788 */ 789 if (diff == 0) { 790 qp->s_state = OP(SEND_LAST); 791 goto done; 792 } 793 opcode = wqe->wr.opcode; 794 } 795 796 /* 797 * Set the state to restart in the middle of a request. 798 * Don't change the s_sge, s_cur_sge, or s_cur_size. 799 * See qib_make_rc_req(). 800 */ 801 switch (opcode) { 802 case IB_WR_SEND: 803 case IB_WR_SEND_WITH_IMM: 804 qp->s_state = OP(RDMA_READ_RESPONSE_FIRST); 805 break; 806 807 case IB_WR_RDMA_WRITE: 808 case IB_WR_RDMA_WRITE_WITH_IMM: 809 qp->s_state = OP(RDMA_READ_RESPONSE_LAST); 810 break; 811 812 case IB_WR_RDMA_READ: 813 qp->s_state = OP(RDMA_READ_RESPONSE_MIDDLE); 814 break; 815 816 default: 817 /* 818 * This case shouldn't happen since its only 819 * one PSN per req. 820 */ 821 qp->s_state = OP(SEND_LAST); 822 } 823 done: 824 qp->s_psn = psn; 825 /* 826 * Set RVT_S_WAIT_PSN as qib_rc_complete() may start the timer 827 * asynchronously before the send tasklet can get scheduled. 828 * Doing it in qib_make_rc_req() is too late. 829 */ 830 if ((qib_cmp24(qp->s_psn, qp->s_sending_hpsn) <= 0) && 831 (qib_cmp24(qp->s_sending_psn, qp->s_sending_hpsn) <= 0)) 832 qp->s_flags |= RVT_S_WAIT_PSN; 833 } 834 835 /* 836 * Back up requester to resend the last un-ACKed request. 837 * The QP r_lock and s_lock should be held and interrupts disabled. 838 */ 839 static void qib_restart_rc(struct rvt_qp *qp, u32 psn, int wait) 840 { 841 struct rvt_swqe *wqe = rvt_get_swqe_ptr(qp, qp->s_acked); 842 struct qib_ibport *ibp; 843 844 if (qp->s_retry == 0) { 845 if (qp->s_mig_state == IB_MIG_ARMED) { 846 qib_migrate_qp(qp); 847 qp->s_retry = qp->s_retry_cnt; 848 } else if (qp->s_last == qp->s_acked) { 849 qib_send_complete(qp, wqe, IB_WC_RETRY_EXC_ERR); 850 rvt_error_qp(qp, IB_WC_WR_FLUSH_ERR); 851 return; 852 } else /* XXX need to handle delayed completion */ 853 return; 854 } else 855 qp->s_retry--; 856 857 ibp = to_iport(qp->ibqp.device, qp->port_num); 858 if (wqe->wr.opcode == IB_WR_RDMA_READ) 859 ibp->rvp.n_rc_resends++; 860 else 861 ibp->rvp.n_rc_resends += (qp->s_psn - psn) & QIB_PSN_MASK; 862 863 qp->s_flags &= ~(RVT_S_WAIT_FENCE | RVT_S_WAIT_RDMAR | 864 RVT_S_WAIT_SSN_CREDIT | RVT_S_WAIT_PSN | 865 RVT_S_WAIT_ACK); 866 if (wait) 867 qp->s_flags |= RVT_S_SEND_ONE; 868 reset_psn(qp, psn); 869 } 870 871 /* 872 * This is called from s_timer for missing responses. 873 */ 874 static void rc_timeout(unsigned long arg) 875 { 876 struct rvt_qp *qp = (struct rvt_qp *)arg; 877 struct qib_ibport *ibp; 878 unsigned long flags; 879 880 spin_lock_irqsave(&qp->r_lock, flags); 881 spin_lock(&qp->s_lock); 882 if (qp->s_flags & RVT_S_TIMER) { 883 ibp = to_iport(qp->ibqp.device, qp->port_num); 884 ibp->rvp.n_rc_timeouts++; 885 qp->s_flags &= ~RVT_S_TIMER; 886 del_timer(&qp->s_timer); 887 qib_restart_rc(qp, qp->s_last_psn + 1, 1); 888 qib_schedule_send(qp); 889 } 890 spin_unlock(&qp->s_lock); 891 spin_unlock_irqrestore(&qp->r_lock, flags); 892 } 893 894 /* 895 * This is called from s_timer for RNR timeouts. 896 */ 897 void qib_rc_rnr_retry(unsigned long arg) 898 { 899 struct rvt_qp *qp = (struct rvt_qp *)arg; 900 unsigned long flags; 901 902 spin_lock_irqsave(&qp->s_lock, flags); 903 if (qp->s_flags & RVT_S_WAIT_RNR) { 904 qp->s_flags &= ~RVT_S_WAIT_RNR; 905 del_timer(&qp->s_timer); 906 qib_schedule_send(qp); 907 } 908 spin_unlock_irqrestore(&qp->s_lock, flags); 909 } 910 911 /* 912 * Set qp->s_sending_psn to the next PSN after the given one. 913 * This would be psn+1 except when RDMA reads are present. 914 */ 915 static void reset_sending_psn(struct rvt_qp *qp, u32 psn) 916 { 917 struct rvt_swqe *wqe; 918 u32 n = qp->s_last; 919 920 /* Find the work request corresponding to the given PSN. */ 921 for (;;) { 922 wqe = rvt_get_swqe_ptr(qp, n); 923 if (qib_cmp24(psn, wqe->lpsn) <= 0) { 924 if (wqe->wr.opcode == IB_WR_RDMA_READ) 925 qp->s_sending_psn = wqe->lpsn + 1; 926 else 927 qp->s_sending_psn = psn + 1; 928 break; 929 } 930 if (++n == qp->s_size) 931 n = 0; 932 if (n == qp->s_tail) 933 break; 934 } 935 } 936 937 /* 938 * This should be called with the QP s_lock held and interrupts disabled. 939 */ 940 void qib_rc_send_complete(struct rvt_qp *qp, struct ib_header *hdr) 941 { 942 struct ib_other_headers *ohdr; 943 struct rvt_swqe *wqe; 944 u32 opcode; 945 u32 psn; 946 947 if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_OR_FLUSH_SEND)) 948 return; 949 950 /* Find out where the BTH is */ 951 if ((be16_to_cpu(hdr->lrh[0]) & 3) == QIB_LRH_BTH) 952 ohdr = &hdr->u.oth; 953 else 954 ohdr = &hdr->u.l.oth; 955 956 opcode = be32_to_cpu(ohdr->bth[0]) >> 24; 957 if (opcode >= OP(RDMA_READ_RESPONSE_FIRST) && 958 opcode <= OP(ATOMIC_ACKNOWLEDGE)) { 959 WARN_ON(!qp->s_rdma_ack_cnt); 960 qp->s_rdma_ack_cnt--; 961 return; 962 } 963 964 psn = be32_to_cpu(ohdr->bth[2]); 965 reset_sending_psn(qp, psn); 966 967 /* 968 * Start timer after a packet requesting an ACK has been sent and 969 * there are still requests that haven't been acked. 970 */ 971 if ((psn & IB_BTH_REQ_ACK) && qp->s_acked != qp->s_tail && 972 !(qp->s_flags & (RVT_S_TIMER | RVT_S_WAIT_RNR | RVT_S_WAIT_PSN)) && 973 (ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK)) 974 start_timer(qp); 975 976 while (qp->s_last != qp->s_acked) { 977 u32 s_last; 978 979 wqe = rvt_get_swqe_ptr(qp, qp->s_last); 980 if (qib_cmp24(wqe->lpsn, qp->s_sending_psn) >= 0 && 981 qib_cmp24(qp->s_sending_psn, qp->s_sending_hpsn) <= 0) 982 break; 983 s_last = qp->s_last; 984 if (++s_last >= qp->s_size) 985 s_last = 0; 986 qp->s_last = s_last; 987 /* see post_send() */ 988 barrier(); 989 rvt_put_swqe(wqe); 990 rvt_qp_swqe_complete(qp, wqe, IB_WC_SUCCESS); 991 } 992 /* 993 * If we were waiting for sends to complete before resending, 994 * and they are now complete, restart sending. 995 */ 996 if (qp->s_flags & RVT_S_WAIT_PSN && 997 qib_cmp24(qp->s_sending_psn, qp->s_sending_hpsn) > 0) { 998 qp->s_flags &= ~RVT_S_WAIT_PSN; 999 qp->s_sending_psn = qp->s_psn; 1000 qp->s_sending_hpsn = qp->s_psn - 1; 1001 qib_schedule_send(qp); 1002 } 1003 } 1004 1005 static inline void update_last_psn(struct rvt_qp *qp, u32 psn) 1006 { 1007 qp->s_last_psn = psn; 1008 } 1009 1010 /* 1011 * Generate a SWQE completion. 1012 * This is similar to qib_send_complete but has to check to be sure 1013 * that the SGEs are not being referenced if the SWQE is being resent. 1014 */ 1015 static struct rvt_swqe *do_rc_completion(struct rvt_qp *qp, 1016 struct rvt_swqe *wqe, 1017 struct qib_ibport *ibp) 1018 { 1019 /* 1020 * Don't decrement refcount and don't generate a 1021 * completion if the SWQE is being resent until the send 1022 * is finished. 1023 */ 1024 if (qib_cmp24(wqe->lpsn, qp->s_sending_psn) < 0 || 1025 qib_cmp24(qp->s_sending_psn, qp->s_sending_hpsn) > 0) { 1026 u32 s_last; 1027 1028 rvt_put_swqe(wqe); 1029 s_last = qp->s_last; 1030 if (++s_last >= qp->s_size) 1031 s_last = 0; 1032 qp->s_last = s_last; 1033 /* see post_send() */ 1034 barrier(); 1035 rvt_qp_swqe_complete(qp, wqe, IB_WC_SUCCESS); 1036 } else 1037 this_cpu_inc(*ibp->rvp.rc_delayed_comp); 1038 1039 qp->s_retry = qp->s_retry_cnt; 1040 update_last_psn(qp, wqe->lpsn); 1041 1042 /* 1043 * If we are completing a request which is in the process of 1044 * being resent, we can stop resending it since we know the 1045 * responder has already seen it. 1046 */ 1047 if (qp->s_acked == qp->s_cur) { 1048 if (++qp->s_cur >= qp->s_size) 1049 qp->s_cur = 0; 1050 qp->s_acked = qp->s_cur; 1051 wqe = rvt_get_swqe_ptr(qp, qp->s_cur); 1052 if (qp->s_acked != qp->s_tail) { 1053 qp->s_state = OP(SEND_LAST); 1054 qp->s_psn = wqe->psn; 1055 } 1056 } else { 1057 if (++qp->s_acked >= qp->s_size) 1058 qp->s_acked = 0; 1059 if (qp->state == IB_QPS_SQD && qp->s_acked == qp->s_cur) 1060 qp->s_draining = 0; 1061 wqe = rvt_get_swqe_ptr(qp, qp->s_acked); 1062 } 1063 return wqe; 1064 } 1065 1066 /** 1067 * do_rc_ack - process an incoming RC ACK 1068 * @qp: the QP the ACK came in on 1069 * @psn: the packet sequence number of the ACK 1070 * @opcode: the opcode of the request that resulted in the ACK 1071 * 1072 * This is called from qib_rc_rcv_resp() to process an incoming RC ACK 1073 * for the given QP. 1074 * Called at interrupt level with the QP s_lock held. 1075 * Returns 1 if OK, 0 if current operation should be aborted (NAK). 1076 */ 1077 static int do_rc_ack(struct rvt_qp *qp, u32 aeth, u32 psn, int opcode, 1078 u64 val, struct qib_ctxtdata *rcd) 1079 { 1080 struct qib_ibport *ibp; 1081 enum ib_wc_status status; 1082 struct rvt_swqe *wqe; 1083 int ret = 0; 1084 u32 ack_psn; 1085 int diff; 1086 1087 /* Remove QP from retry timer */ 1088 if (qp->s_flags & (RVT_S_TIMER | RVT_S_WAIT_RNR)) { 1089 qp->s_flags &= ~(RVT_S_TIMER | RVT_S_WAIT_RNR); 1090 del_timer(&qp->s_timer); 1091 } 1092 1093 /* 1094 * Note that NAKs implicitly ACK outstanding SEND and RDMA write 1095 * requests and implicitly NAK RDMA read and atomic requests issued 1096 * before the NAK'ed request. The MSN won't include the NAK'ed 1097 * request but will include an ACK'ed request(s). 1098 */ 1099 ack_psn = psn; 1100 if (aeth >> 29) 1101 ack_psn--; 1102 wqe = rvt_get_swqe_ptr(qp, qp->s_acked); 1103 ibp = to_iport(qp->ibqp.device, qp->port_num); 1104 1105 /* 1106 * The MSN might be for a later WQE than the PSN indicates so 1107 * only complete WQEs that the PSN finishes. 1108 */ 1109 while ((diff = qib_cmp24(ack_psn, wqe->lpsn)) >= 0) { 1110 /* 1111 * RDMA_READ_RESPONSE_ONLY is a special case since 1112 * we want to generate completion events for everything 1113 * before the RDMA read, copy the data, then generate 1114 * the completion for the read. 1115 */ 1116 if (wqe->wr.opcode == IB_WR_RDMA_READ && 1117 opcode == OP(RDMA_READ_RESPONSE_ONLY) && 1118 diff == 0) { 1119 ret = 1; 1120 goto bail; 1121 } 1122 /* 1123 * If this request is a RDMA read or atomic, and the ACK is 1124 * for a later operation, this ACK NAKs the RDMA read or 1125 * atomic. In other words, only a RDMA_READ_LAST or ONLY 1126 * can ACK a RDMA read and likewise for atomic ops. Note 1127 * that the NAK case can only happen if relaxed ordering is 1128 * used and requests are sent after an RDMA read or atomic 1129 * is sent but before the response is received. 1130 */ 1131 if ((wqe->wr.opcode == IB_WR_RDMA_READ && 1132 (opcode != OP(RDMA_READ_RESPONSE_LAST) || diff != 0)) || 1133 ((wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP || 1134 wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD) && 1135 (opcode != OP(ATOMIC_ACKNOWLEDGE) || diff != 0))) { 1136 /* Retry this request. */ 1137 if (!(qp->r_flags & RVT_R_RDMAR_SEQ)) { 1138 qp->r_flags |= RVT_R_RDMAR_SEQ; 1139 qib_restart_rc(qp, qp->s_last_psn + 1, 0); 1140 if (list_empty(&qp->rspwait)) { 1141 qp->r_flags |= RVT_R_RSP_SEND; 1142 rvt_get_qp(qp); 1143 list_add_tail(&qp->rspwait, 1144 &rcd->qp_wait_list); 1145 } 1146 } 1147 /* 1148 * No need to process the ACK/NAK since we are 1149 * restarting an earlier request. 1150 */ 1151 goto bail; 1152 } 1153 if (wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP || 1154 wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD) { 1155 u64 *vaddr = wqe->sg_list[0].vaddr; 1156 *vaddr = val; 1157 } 1158 if (qp->s_num_rd_atomic && 1159 (wqe->wr.opcode == IB_WR_RDMA_READ || 1160 wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP || 1161 wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD)) { 1162 qp->s_num_rd_atomic--; 1163 /* Restart sending task if fence is complete */ 1164 if ((qp->s_flags & RVT_S_WAIT_FENCE) && 1165 !qp->s_num_rd_atomic) { 1166 qp->s_flags &= ~(RVT_S_WAIT_FENCE | 1167 RVT_S_WAIT_ACK); 1168 qib_schedule_send(qp); 1169 } else if (qp->s_flags & RVT_S_WAIT_RDMAR) { 1170 qp->s_flags &= ~(RVT_S_WAIT_RDMAR | 1171 RVT_S_WAIT_ACK); 1172 qib_schedule_send(qp); 1173 } 1174 } 1175 wqe = do_rc_completion(qp, wqe, ibp); 1176 if (qp->s_acked == qp->s_tail) 1177 break; 1178 } 1179 1180 switch (aeth >> 29) { 1181 case 0: /* ACK */ 1182 this_cpu_inc(*ibp->rvp.rc_acks); 1183 if (qp->s_acked != qp->s_tail) { 1184 /* 1185 * We are expecting more ACKs so 1186 * reset the retransmit timer. 1187 */ 1188 start_timer(qp); 1189 /* 1190 * We can stop resending the earlier packets and 1191 * continue with the next packet the receiver wants. 1192 */ 1193 if (qib_cmp24(qp->s_psn, psn) <= 0) 1194 reset_psn(qp, psn + 1); 1195 } else if (qib_cmp24(qp->s_psn, psn) <= 0) { 1196 qp->s_state = OP(SEND_LAST); 1197 qp->s_psn = psn + 1; 1198 } 1199 if (qp->s_flags & RVT_S_WAIT_ACK) { 1200 qp->s_flags &= ~RVT_S_WAIT_ACK; 1201 qib_schedule_send(qp); 1202 } 1203 qib_get_credit(qp, aeth); 1204 qp->s_rnr_retry = qp->s_rnr_retry_cnt; 1205 qp->s_retry = qp->s_retry_cnt; 1206 update_last_psn(qp, psn); 1207 ret = 1; 1208 goto bail; 1209 1210 case 1: /* RNR NAK */ 1211 ibp->rvp.n_rnr_naks++; 1212 if (qp->s_acked == qp->s_tail) 1213 goto bail; 1214 if (qp->s_flags & RVT_S_WAIT_RNR) 1215 goto bail; 1216 if (qp->s_rnr_retry == 0) { 1217 status = IB_WC_RNR_RETRY_EXC_ERR; 1218 goto class_b; 1219 } 1220 if (qp->s_rnr_retry_cnt < 7) 1221 qp->s_rnr_retry--; 1222 1223 /* The last valid PSN is the previous PSN. */ 1224 update_last_psn(qp, psn - 1); 1225 1226 ibp->rvp.n_rc_resends += (qp->s_psn - psn) & QIB_PSN_MASK; 1227 1228 reset_psn(qp, psn); 1229 1230 qp->s_flags &= ~(RVT_S_WAIT_SSN_CREDIT | RVT_S_WAIT_ACK); 1231 qp->s_flags |= RVT_S_WAIT_RNR; 1232 qp->s_timer.function = qib_rc_rnr_retry; 1233 qp->s_timer.expires = jiffies + usecs_to_jiffies( 1234 ib_qib_rnr_table[(aeth >> QIB_AETH_CREDIT_SHIFT) & 1235 QIB_AETH_CREDIT_MASK]); 1236 add_timer(&qp->s_timer); 1237 goto bail; 1238 1239 case 3: /* NAK */ 1240 if (qp->s_acked == qp->s_tail) 1241 goto bail; 1242 /* The last valid PSN is the previous PSN. */ 1243 update_last_psn(qp, psn - 1); 1244 switch ((aeth >> QIB_AETH_CREDIT_SHIFT) & 1245 QIB_AETH_CREDIT_MASK) { 1246 case 0: /* PSN sequence error */ 1247 ibp->rvp.n_seq_naks++; 1248 /* 1249 * Back up to the responder's expected PSN. 1250 * Note that we might get a NAK in the middle of an 1251 * RDMA READ response which terminates the RDMA 1252 * READ. 1253 */ 1254 qib_restart_rc(qp, psn, 0); 1255 qib_schedule_send(qp); 1256 break; 1257 1258 case 1: /* Invalid Request */ 1259 status = IB_WC_REM_INV_REQ_ERR; 1260 ibp->rvp.n_other_naks++; 1261 goto class_b; 1262 1263 case 2: /* Remote Access Error */ 1264 status = IB_WC_REM_ACCESS_ERR; 1265 ibp->rvp.n_other_naks++; 1266 goto class_b; 1267 1268 case 3: /* Remote Operation Error */ 1269 status = IB_WC_REM_OP_ERR; 1270 ibp->rvp.n_other_naks++; 1271 class_b: 1272 if (qp->s_last == qp->s_acked) { 1273 qib_send_complete(qp, wqe, status); 1274 rvt_error_qp(qp, IB_WC_WR_FLUSH_ERR); 1275 } 1276 break; 1277 1278 default: 1279 /* Ignore other reserved NAK error codes */ 1280 goto reserved; 1281 } 1282 qp->s_retry = qp->s_retry_cnt; 1283 qp->s_rnr_retry = qp->s_rnr_retry_cnt; 1284 goto bail; 1285 1286 default: /* 2: reserved */ 1287 reserved: 1288 /* Ignore reserved NAK codes. */ 1289 goto bail; 1290 } 1291 1292 bail: 1293 return ret; 1294 } 1295 1296 /* 1297 * We have seen an out of sequence RDMA read middle or last packet. 1298 * This ACKs SENDs and RDMA writes up to the first RDMA read or atomic SWQE. 1299 */ 1300 static void rdma_seq_err(struct rvt_qp *qp, struct qib_ibport *ibp, u32 psn, 1301 struct qib_ctxtdata *rcd) 1302 { 1303 struct rvt_swqe *wqe; 1304 1305 /* Remove QP from retry timer */ 1306 if (qp->s_flags & (RVT_S_TIMER | RVT_S_WAIT_RNR)) { 1307 qp->s_flags &= ~(RVT_S_TIMER | RVT_S_WAIT_RNR); 1308 del_timer(&qp->s_timer); 1309 } 1310 1311 wqe = rvt_get_swqe_ptr(qp, qp->s_acked); 1312 1313 while (qib_cmp24(psn, wqe->lpsn) > 0) { 1314 if (wqe->wr.opcode == IB_WR_RDMA_READ || 1315 wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP || 1316 wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD) 1317 break; 1318 wqe = do_rc_completion(qp, wqe, ibp); 1319 } 1320 1321 ibp->rvp.n_rdma_seq++; 1322 qp->r_flags |= RVT_R_RDMAR_SEQ; 1323 qib_restart_rc(qp, qp->s_last_psn + 1, 0); 1324 if (list_empty(&qp->rspwait)) { 1325 qp->r_flags |= RVT_R_RSP_SEND; 1326 rvt_get_qp(qp); 1327 list_add_tail(&qp->rspwait, &rcd->qp_wait_list); 1328 } 1329 } 1330 1331 /** 1332 * qib_rc_rcv_resp - process an incoming RC response packet 1333 * @ibp: the port this packet came in on 1334 * @ohdr: the other headers for this packet 1335 * @data: the packet data 1336 * @tlen: the packet length 1337 * @qp: the QP for this packet 1338 * @opcode: the opcode for this packet 1339 * @psn: the packet sequence number for this packet 1340 * @hdrsize: the header length 1341 * @pmtu: the path MTU 1342 * 1343 * This is called from qib_rc_rcv() to process an incoming RC response 1344 * packet for the given QP. 1345 * Called at interrupt level. 1346 */ 1347 static void qib_rc_rcv_resp(struct qib_ibport *ibp, 1348 struct ib_other_headers *ohdr, 1349 void *data, u32 tlen, 1350 struct rvt_qp *qp, 1351 u32 opcode, 1352 u32 psn, u32 hdrsize, u32 pmtu, 1353 struct qib_ctxtdata *rcd) 1354 { 1355 struct rvt_swqe *wqe; 1356 struct qib_pportdata *ppd = ppd_from_ibp(ibp); 1357 enum ib_wc_status status; 1358 unsigned long flags; 1359 int diff; 1360 u32 pad; 1361 u32 aeth; 1362 u64 val; 1363 1364 if (opcode != OP(RDMA_READ_RESPONSE_MIDDLE)) { 1365 /* 1366 * If ACK'd PSN on SDMA busy list try to make progress to 1367 * reclaim SDMA credits. 1368 */ 1369 if ((qib_cmp24(psn, qp->s_sending_psn) >= 0) && 1370 (qib_cmp24(qp->s_sending_psn, qp->s_sending_hpsn) <= 0)) { 1371 1372 /* 1373 * If send tasklet not running attempt to progress 1374 * SDMA queue. 1375 */ 1376 if (!(qp->s_flags & RVT_S_BUSY)) { 1377 /* Acquire SDMA Lock */ 1378 spin_lock_irqsave(&ppd->sdma_lock, flags); 1379 /* Invoke sdma make progress */ 1380 qib_sdma_make_progress(ppd); 1381 /* Release SDMA Lock */ 1382 spin_unlock_irqrestore(&ppd->sdma_lock, flags); 1383 } 1384 } 1385 } 1386 1387 spin_lock_irqsave(&qp->s_lock, flags); 1388 if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK)) 1389 goto ack_done; 1390 1391 /* Ignore invalid responses. */ 1392 smp_read_barrier_depends(); /* see post_one_send */ 1393 if (qib_cmp24(psn, ACCESS_ONCE(qp->s_next_psn)) >= 0) 1394 goto ack_done; 1395 1396 /* Ignore duplicate responses. */ 1397 diff = qib_cmp24(psn, qp->s_last_psn); 1398 if (unlikely(diff <= 0)) { 1399 /* Update credits for "ghost" ACKs */ 1400 if (diff == 0 && opcode == OP(ACKNOWLEDGE)) { 1401 aeth = be32_to_cpu(ohdr->u.aeth); 1402 if ((aeth >> 29) == 0) 1403 qib_get_credit(qp, aeth); 1404 } 1405 goto ack_done; 1406 } 1407 1408 /* 1409 * Skip everything other than the PSN we expect, if we are waiting 1410 * for a reply to a restarted RDMA read or atomic op. 1411 */ 1412 if (qp->r_flags & RVT_R_RDMAR_SEQ) { 1413 if (qib_cmp24(psn, qp->s_last_psn + 1) != 0) 1414 goto ack_done; 1415 qp->r_flags &= ~RVT_R_RDMAR_SEQ; 1416 } 1417 1418 if (unlikely(qp->s_acked == qp->s_tail)) 1419 goto ack_done; 1420 wqe = rvt_get_swqe_ptr(qp, qp->s_acked); 1421 status = IB_WC_SUCCESS; 1422 1423 switch (opcode) { 1424 case OP(ACKNOWLEDGE): 1425 case OP(ATOMIC_ACKNOWLEDGE): 1426 case OP(RDMA_READ_RESPONSE_FIRST): 1427 aeth = be32_to_cpu(ohdr->u.aeth); 1428 if (opcode == OP(ATOMIC_ACKNOWLEDGE)) 1429 val = ib_u64_get(&ohdr->u.at.atomic_ack_eth); 1430 else 1431 val = 0; 1432 if (!do_rc_ack(qp, aeth, psn, opcode, val, rcd) || 1433 opcode != OP(RDMA_READ_RESPONSE_FIRST)) 1434 goto ack_done; 1435 hdrsize += 4; 1436 wqe = rvt_get_swqe_ptr(qp, qp->s_acked); 1437 if (unlikely(wqe->wr.opcode != IB_WR_RDMA_READ)) 1438 goto ack_op_err; 1439 /* 1440 * If this is a response to a resent RDMA read, we 1441 * have to be careful to copy the data to the right 1442 * location. 1443 */ 1444 qp->s_rdma_read_len = restart_sge(&qp->s_rdma_read_sge, 1445 wqe, psn, pmtu); 1446 goto read_middle; 1447 1448 case OP(RDMA_READ_RESPONSE_MIDDLE): 1449 /* no AETH, no ACK */ 1450 if (unlikely(qib_cmp24(psn, qp->s_last_psn + 1))) 1451 goto ack_seq_err; 1452 if (unlikely(wqe->wr.opcode != IB_WR_RDMA_READ)) 1453 goto ack_op_err; 1454 read_middle: 1455 if (unlikely(tlen != (hdrsize + pmtu + 4))) 1456 goto ack_len_err; 1457 if (unlikely(pmtu >= qp->s_rdma_read_len)) 1458 goto ack_len_err; 1459 1460 /* 1461 * We got a response so update the timeout. 1462 * 4.096 usec. * (1 << qp->timeout) 1463 */ 1464 qp->s_flags |= RVT_S_TIMER; 1465 mod_timer(&qp->s_timer, jiffies + qp->timeout_jiffies); 1466 if (qp->s_flags & RVT_S_WAIT_ACK) { 1467 qp->s_flags &= ~RVT_S_WAIT_ACK; 1468 qib_schedule_send(qp); 1469 } 1470 1471 if (opcode == OP(RDMA_READ_RESPONSE_MIDDLE)) 1472 qp->s_retry = qp->s_retry_cnt; 1473 1474 /* 1475 * Update the RDMA receive state but do the copy w/o 1476 * holding the locks and blocking interrupts. 1477 */ 1478 qp->s_rdma_read_len -= pmtu; 1479 update_last_psn(qp, psn); 1480 spin_unlock_irqrestore(&qp->s_lock, flags); 1481 qib_copy_sge(&qp->s_rdma_read_sge, data, pmtu, 0); 1482 goto bail; 1483 1484 case OP(RDMA_READ_RESPONSE_ONLY): 1485 aeth = be32_to_cpu(ohdr->u.aeth); 1486 if (!do_rc_ack(qp, aeth, psn, opcode, 0, rcd)) 1487 goto ack_done; 1488 /* Get the number of bytes the message was padded by. */ 1489 pad = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3; 1490 /* 1491 * Check that the data size is >= 0 && <= pmtu. 1492 * Remember to account for the AETH header (4) and 1493 * ICRC (4). 1494 */ 1495 if (unlikely(tlen < (hdrsize + pad + 8))) 1496 goto ack_len_err; 1497 /* 1498 * If this is a response to a resent RDMA read, we 1499 * have to be careful to copy the data to the right 1500 * location. 1501 */ 1502 wqe = rvt_get_swqe_ptr(qp, qp->s_acked); 1503 qp->s_rdma_read_len = restart_sge(&qp->s_rdma_read_sge, 1504 wqe, psn, pmtu); 1505 goto read_last; 1506 1507 case OP(RDMA_READ_RESPONSE_LAST): 1508 /* ACKs READ req. */ 1509 if (unlikely(qib_cmp24(psn, qp->s_last_psn + 1))) 1510 goto ack_seq_err; 1511 if (unlikely(wqe->wr.opcode != IB_WR_RDMA_READ)) 1512 goto ack_op_err; 1513 /* Get the number of bytes the message was padded by. */ 1514 pad = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3; 1515 /* 1516 * Check that the data size is >= 1 && <= pmtu. 1517 * Remember to account for the AETH header (4) and 1518 * ICRC (4). 1519 */ 1520 if (unlikely(tlen <= (hdrsize + pad + 8))) 1521 goto ack_len_err; 1522 read_last: 1523 tlen -= hdrsize + pad + 8; 1524 if (unlikely(tlen != qp->s_rdma_read_len)) 1525 goto ack_len_err; 1526 aeth = be32_to_cpu(ohdr->u.aeth); 1527 qib_copy_sge(&qp->s_rdma_read_sge, data, tlen, 0); 1528 WARN_ON(qp->s_rdma_read_sge.num_sge); 1529 (void) do_rc_ack(qp, aeth, psn, 1530 OP(RDMA_READ_RESPONSE_LAST), 0, rcd); 1531 goto ack_done; 1532 } 1533 1534 ack_op_err: 1535 status = IB_WC_LOC_QP_OP_ERR; 1536 goto ack_err; 1537 1538 ack_seq_err: 1539 rdma_seq_err(qp, ibp, psn, rcd); 1540 goto ack_done; 1541 1542 ack_len_err: 1543 status = IB_WC_LOC_LEN_ERR; 1544 ack_err: 1545 if (qp->s_last == qp->s_acked) { 1546 qib_send_complete(qp, wqe, status); 1547 rvt_error_qp(qp, IB_WC_WR_FLUSH_ERR); 1548 } 1549 ack_done: 1550 spin_unlock_irqrestore(&qp->s_lock, flags); 1551 bail: 1552 return; 1553 } 1554 1555 /** 1556 * qib_rc_rcv_error - process an incoming duplicate or error RC packet 1557 * @ohdr: the other headers for this packet 1558 * @data: the packet data 1559 * @qp: the QP for this packet 1560 * @opcode: the opcode for this packet 1561 * @psn: the packet sequence number for this packet 1562 * @diff: the difference between the PSN and the expected PSN 1563 * 1564 * This is called from qib_rc_rcv() to process an unexpected 1565 * incoming RC packet for the given QP. 1566 * Called at interrupt level. 1567 * Return 1 if no more processing is needed; otherwise return 0 to 1568 * schedule a response to be sent. 1569 */ 1570 static int qib_rc_rcv_error(struct ib_other_headers *ohdr, 1571 void *data, 1572 struct rvt_qp *qp, 1573 u32 opcode, 1574 u32 psn, 1575 int diff, 1576 struct qib_ctxtdata *rcd) 1577 { 1578 struct qib_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num); 1579 struct rvt_ack_entry *e; 1580 unsigned long flags; 1581 u8 i, prev; 1582 int old_req; 1583 1584 if (diff > 0) { 1585 /* 1586 * Packet sequence error. 1587 * A NAK will ACK earlier sends and RDMA writes. 1588 * Don't queue the NAK if we already sent one. 1589 */ 1590 if (!qp->r_nak_state) { 1591 ibp->rvp.n_rc_seqnak++; 1592 qp->r_nak_state = IB_NAK_PSN_ERROR; 1593 /* Use the expected PSN. */ 1594 qp->r_ack_psn = qp->r_psn; 1595 /* 1596 * Wait to send the sequence NAK until all packets 1597 * in the receive queue have been processed. 1598 * Otherwise, we end up propagating congestion. 1599 */ 1600 if (list_empty(&qp->rspwait)) { 1601 qp->r_flags |= RVT_R_RSP_NAK; 1602 rvt_get_qp(qp); 1603 list_add_tail(&qp->rspwait, &rcd->qp_wait_list); 1604 } 1605 } 1606 goto done; 1607 } 1608 1609 /* 1610 * Handle a duplicate request. Don't re-execute SEND, RDMA 1611 * write or atomic op. Don't NAK errors, just silently drop 1612 * the duplicate request. Note that r_sge, r_len, and 1613 * r_rcv_len may be in use so don't modify them. 1614 * 1615 * We are supposed to ACK the earliest duplicate PSN but we 1616 * can coalesce an outstanding duplicate ACK. We have to 1617 * send the earliest so that RDMA reads can be restarted at 1618 * the requester's expected PSN. 1619 * 1620 * First, find where this duplicate PSN falls within the 1621 * ACKs previously sent. 1622 * old_req is true if there is an older response that is scheduled 1623 * to be sent before sending this one. 1624 */ 1625 e = NULL; 1626 old_req = 1; 1627 ibp->rvp.n_rc_dupreq++; 1628 1629 spin_lock_irqsave(&qp->s_lock, flags); 1630 1631 for (i = qp->r_head_ack_queue; ; i = prev) { 1632 if (i == qp->s_tail_ack_queue) 1633 old_req = 0; 1634 if (i) 1635 prev = i - 1; 1636 else 1637 prev = QIB_MAX_RDMA_ATOMIC; 1638 if (prev == qp->r_head_ack_queue) { 1639 e = NULL; 1640 break; 1641 } 1642 e = &qp->s_ack_queue[prev]; 1643 if (!e->opcode) { 1644 e = NULL; 1645 break; 1646 } 1647 if (qib_cmp24(psn, e->psn) >= 0) { 1648 if (prev == qp->s_tail_ack_queue && 1649 qib_cmp24(psn, e->lpsn) <= 0) 1650 old_req = 0; 1651 break; 1652 } 1653 } 1654 switch (opcode) { 1655 case OP(RDMA_READ_REQUEST): { 1656 struct ib_reth *reth; 1657 u32 offset; 1658 u32 len; 1659 1660 /* 1661 * If we didn't find the RDMA read request in the ack queue, 1662 * we can ignore this request. 1663 */ 1664 if (!e || e->opcode != OP(RDMA_READ_REQUEST)) 1665 goto unlock_done; 1666 /* RETH comes after BTH */ 1667 reth = &ohdr->u.rc.reth; 1668 /* 1669 * Address range must be a subset of the original 1670 * request and start on pmtu boundaries. 1671 * We reuse the old ack_queue slot since the requester 1672 * should not back up and request an earlier PSN for the 1673 * same request. 1674 */ 1675 offset = ((psn - e->psn) & QIB_PSN_MASK) * 1676 qp->pmtu; 1677 len = be32_to_cpu(reth->length); 1678 if (unlikely(offset + len != e->rdma_sge.sge_length)) 1679 goto unlock_done; 1680 if (e->rdma_sge.mr) { 1681 rvt_put_mr(e->rdma_sge.mr); 1682 e->rdma_sge.mr = NULL; 1683 } 1684 if (len != 0) { 1685 u32 rkey = be32_to_cpu(reth->rkey); 1686 u64 vaddr = be64_to_cpu(reth->vaddr); 1687 int ok; 1688 1689 ok = rvt_rkey_ok(qp, &e->rdma_sge, len, vaddr, rkey, 1690 IB_ACCESS_REMOTE_READ); 1691 if (unlikely(!ok)) 1692 goto unlock_done; 1693 } else { 1694 e->rdma_sge.vaddr = NULL; 1695 e->rdma_sge.length = 0; 1696 e->rdma_sge.sge_length = 0; 1697 } 1698 e->psn = psn; 1699 if (old_req) 1700 goto unlock_done; 1701 qp->s_tail_ack_queue = prev; 1702 break; 1703 } 1704 1705 case OP(COMPARE_SWAP): 1706 case OP(FETCH_ADD): { 1707 /* 1708 * If we didn't find the atomic request in the ack queue 1709 * or the send tasklet is already backed up to send an 1710 * earlier entry, we can ignore this request. 1711 */ 1712 if (!e || e->opcode != (u8) opcode || old_req) 1713 goto unlock_done; 1714 qp->s_tail_ack_queue = prev; 1715 break; 1716 } 1717 1718 default: 1719 /* 1720 * Ignore this operation if it doesn't request an ACK 1721 * or an earlier RDMA read or atomic is going to be resent. 1722 */ 1723 if (!(psn & IB_BTH_REQ_ACK) || old_req) 1724 goto unlock_done; 1725 /* 1726 * Resend the most recent ACK if this request is 1727 * after all the previous RDMA reads and atomics. 1728 */ 1729 if (i == qp->r_head_ack_queue) { 1730 spin_unlock_irqrestore(&qp->s_lock, flags); 1731 qp->r_nak_state = 0; 1732 qp->r_ack_psn = qp->r_psn - 1; 1733 goto send_ack; 1734 } 1735 /* 1736 * Try to send a simple ACK to work around a Mellanox bug 1737 * which doesn't accept a RDMA read response or atomic 1738 * response as an ACK for earlier SENDs or RDMA writes. 1739 */ 1740 if (!(qp->s_flags & RVT_S_RESP_PENDING)) { 1741 spin_unlock_irqrestore(&qp->s_lock, flags); 1742 qp->r_nak_state = 0; 1743 qp->r_ack_psn = qp->s_ack_queue[i].psn - 1; 1744 goto send_ack; 1745 } 1746 /* 1747 * Resend the RDMA read or atomic op which 1748 * ACKs this duplicate request. 1749 */ 1750 qp->s_tail_ack_queue = i; 1751 break; 1752 } 1753 qp->s_ack_state = OP(ACKNOWLEDGE); 1754 qp->s_flags |= RVT_S_RESP_PENDING; 1755 qp->r_nak_state = 0; 1756 qib_schedule_send(qp); 1757 1758 unlock_done: 1759 spin_unlock_irqrestore(&qp->s_lock, flags); 1760 done: 1761 return 1; 1762 1763 send_ack: 1764 return 0; 1765 } 1766 1767 void qib_rc_error(struct rvt_qp *qp, enum ib_wc_status err) 1768 { 1769 unsigned long flags; 1770 int lastwqe; 1771 1772 spin_lock_irqsave(&qp->s_lock, flags); 1773 lastwqe = rvt_error_qp(qp, err); 1774 spin_unlock_irqrestore(&qp->s_lock, flags); 1775 1776 if (lastwqe) { 1777 struct ib_event ev; 1778 1779 ev.device = qp->ibqp.device; 1780 ev.element.qp = &qp->ibqp; 1781 ev.event = IB_EVENT_QP_LAST_WQE_REACHED; 1782 qp->ibqp.event_handler(&ev, qp->ibqp.qp_context); 1783 } 1784 } 1785 1786 static inline void qib_update_ack_queue(struct rvt_qp *qp, unsigned n) 1787 { 1788 unsigned next; 1789 1790 next = n + 1; 1791 if (next > QIB_MAX_RDMA_ATOMIC) 1792 next = 0; 1793 qp->s_tail_ack_queue = next; 1794 qp->s_ack_state = OP(ACKNOWLEDGE); 1795 } 1796 1797 /** 1798 * qib_rc_rcv - process an incoming RC packet 1799 * @rcd: the context pointer 1800 * @hdr: the header of this packet 1801 * @has_grh: true if the header has a GRH 1802 * @data: the packet data 1803 * @tlen: the packet length 1804 * @qp: the QP for this packet 1805 * 1806 * This is called from qib_qp_rcv() to process an incoming RC packet 1807 * for the given QP. 1808 * Called at interrupt level. 1809 */ 1810 void qib_rc_rcv(struct qib_ctxtdata *rcd, struct ib_header *hdr, 1811 int has_grh, void *data, u32 tlen, struct rvt_qp *qp) 1812 { 1813 struct qib_ibport *ibp = &rcd->ppd->ibport_data; 1814 struct ib_other_headers *ohdr; 1815 u32 opcode; 1816 u32 hdrsize; 1817 u32 psn; 1818 u32 pad; 1819 struct ib_wc wc; 1820 u32 pmtu = qp->pmtu; 1821 int diff; 1822 struct ib_reth *reth; 1823 unsigned long flags; 1824 int ret; 1825 1826 /* Check for GRH */ 1827 if (!has_grh) { 1828 ohdr = &hdr->u.oth; 1829 hdrsize = 8 + 12; /* LRH + BTH */ 1830 } else { 1831 ohdr = &hdr->u.l.oth; 1832 hdrsize = 8 + 40 + 12; /* LRH + GRH + BTH */ 1833 } 1834 1835 opcode = be32_to_cpu(ohdr->bth[0]); 1836 if (qib_ruc_check_hdr(ibp, hdr, has_grh, qp, opcode)) 1837 return; 1838 1839 psn = be32_to_cpu(ohdr->bth[2]); 1840 opcode >>= 24; 1841 1842 /* 1843 * Process responses (ACKs) before anything else. Note that the 1844 * packet sequence number will be for something in the send work 1845 * queue rather than the expected receive packet sequence number. 1846 * In other words, this QP is the requester. 1847 */ 1848 if (opcode >= OP(RDMA_READ_RESPONSE_FIRST) && 1849 opcode <= OP(ATOMIC_ACKNOWLEDGE)) { 1850 qib_rc_rcv_resp(ibp, ohdr, data, tlen, qp, opcode, psn, 1851 hdrsize, pmtu, rcd); 1852 return; 1853 } 1854 1855 /* Compute 24 bits worth of difference. */ 1856 diff = qib_cmp24(psn, qp->r_psn); 1857 if (unlikely(diff)) { 1858 if (qib_rc_rcv_error(ohdr, data, qp, opcode, psn, diff, rcd)) 1859 return; 1860 goto send_ack; 1861 } 1862 1863 /* Check for opcode sequence errors. */ 1864 switch (qp->r_state) { 1865 case OP(SEND_FIRST): 1866 case OP(SEND_MIDDLE): 1867 if (opcode == OP(SEND_MIDDLE) || 1868 opcode == OP(SEND_LAST) || 1869 opcode == OP(SEND_LAST_WITH_IMMEDIATE)) 1870 break; 1871 goto nack_inv; 1872 1873 case OP(RDMA_WRITE_FIRST): 1874 case OP(RDMA_WRITE_MIDDLE): 1875 if (opcode == OP(RDMA_WRITE_MIDDLE) || 1876 opcode == OP(RDMA_WRITE_LAST) || 1877 opcode == OP(RDMA_WRITE_LAST_WITH_IMMEDIATE)) 1878 break; 1879 goto nack_inv; 1880 1881 default: 1882 if (opcode == OP(SEND_MIDDLE) || 1883 opcode == OP(SEND_LAST) || 1884 opcode == OP(SEND_LAST_WITH_IMMEDIATE) || 1885 opcode == OP(RDMA_WRITE_MIDDLE) || 1886 opcode == OP(RDMA_WRITE_LAST) || 1887 opcode == OP(RDMA_WRITE_LAST_WITH_IMMEDIATE)) 1888 goto nack_inv; 1889 /* 1890 * Note that it is up to the requester to not send a new 1891 * RDMA read or atomic operation before receiving an ACK 1892 * for the previous operation. 1893 */ 1894 break; 1895 } 1896 1897 if (qp->state == IB_QPS_RTR && !(qp->r_flags & RVT_R_COMM_EST)) { 1898 qp->r_flags |= RVT_R_COMM_EST; 1899 if (qp->ibqp.event_handler) { 1900 struct ib_event ev; 1901 1902 ev.device = qp->ibqp.device; 1903 ev.element.qp = &qp->ibqp; 1904 ev.event = IB_EVENT_COMM_EST; 1905 qp->ibqp.event_handler(&ev, qp->ibqp.qp_context); 1906 } 1907 } 1908 1909 /* OK, process the packet. */ 1910 switch (opcode) { 1911 case OP(SEND_FIRST): 1912 ret = qib_get_rwqe(qp, 0); 1913 if (ret < 0) 1914 goto nack_op_err; 1915 if (!ret) 1916 goto rnr_nak; 1917 qp->r_rcv_len = 0; 1918 /* FALLTHROUGH */ 1919 case OP(SEND_MIDDLE): 1920 case OP(RDMA_WRITE_MIDDLE): 1921 send_middle: 1922 /* Check for invalid length PMTU or posted rwqe len. */ 1923 if (unlikely(tlen != (hdrsize + pmtu + 4))) 1924 goto nack_inv; 1925 qp->r_rcv_len += pmtu; 1926 if (unlikely(qp->r_rcv_len > qp->r_len)) 1927 goto nack_inv; 1928 qib_copy_sge(&qp->r_sge, data, pmtu, 1); 1929 break; 1930 1931 case OP(RDMA_WRITE_LAST_WITH_IMMEDIATE): 1932 /* consume RWQE */ 1933 ret = qib_get_rwqe(qp, 1); 1934 if (ret < 0) 1935 goto nack_op_err; 1936 if (!ret) 1937 goto rnr_nak; 1938 goto send_last_imm; 1939 1940 case OP(SEND_ONLY): 1941 case OP(SEND_ONLY_WITH_IMMEDIATE): 1942 ret = qib_get_rwqe(qp, 0); 1943 if (ret < 0) 1944 goto nack_op_err; 1945 if (!ret) 1946 goto rnr_nak; 1947 qp->r_rcv_len = 0; 1948 if (opcode == OP(SEND_ONLY)) 1949 goto no_immediate_data; 1950 /* FALLTHROUGH for SEND_ONLY_WITH_IMMEDIATE */ 1951 case OP(SEND_LAST_WITH_IMMEDIATE): 1952 send_last_imm: 1953 wc.ex.imm_data = ohdr->u.imm_data; 1954 hdrsize += 4; 1955 wc.wc_flags = IB_WC_WITH_IMM; 1956 goto send_last; 1957 case OP(SEND_LAST): 1958 case OP(RDMA_WRITE_LAST): 1959 no_immediate_data: 1960 wc.wc_flags = 0; 1961 wc.ex.imm_data = 0; 1962 send_last: 1963 /* Get the number of bytes the message was padded by. */ 1964 pad = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3; 1965 /* Check for invalid length. */ 1966 /* XXX LAST len should be >= 1 */ 1967 if (unlikely(tlen < (hdrsize + pad + 4))) 1968 goto nack_inv; 1969 /* Don't count the CRC. */ 1970 tlen -= (hdrsize + pad + 4); 1971 wc.byte_len = tlen + qp->r_rcv_len; 1972 if (unlikely(wc.byte_len > qp->r_len)) 1973 goto nack_inv; 1974 qib_copy_sge(&qp->r_sge, data, tlen, 1); 1975 rvt_put_ss(&qp->r_sge); 1976 qp->r_msn++; 1977 if (!test_and_clear_bit(RVT_R_WRID_VALID, &qp->r_aflags)) 1978 break; 1979 wc.wr_id = qp->r_wr_id; 1980 wc.status = IB_WC_SUCCESS; 1981 if (opcode == OP(RDMA_WRITE_LAST_WITH_IMMEDIATE) || 1982 opcode == OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE)) 1983 wc.opcode = IB_WC_RECV_RDMA_WITH_IMM; 1984 else 1985 wc.opcode = IB_WC_RECV; 1986 wc.qp = &qp->ibqp; 1987 wc.src_qp = qp->remote_qpn; 1988 wc.slid = qp->remote_ah_attr.dlid; 1989 wc.sl = qp->remote_ah_attr.sl; 1990 /* zero fields that are N/A */ 1991 wc.vendor_err = 0; 1992 wc.pkey_index = 0; 1993 wc.dlid_path_bits = 0; 1994 wc.port_num = 0; 1995 /* Signal completion event if the solicited bit is set. */ 1996 rvt_cq_enter(ibcq_to_rvtcq(qp->ibqp.recv_cq), &wc, 1997 (ohdr->bth[0] & 1998 cpu_to_be32(IB_BTH_SOLICITED)) != 0); 1999 break; 2000 2001 case OP(RDMA_WRITE_FIRST): 2002 case OP(RDMA_WRITE_ONLY): 2003 case OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE): 2004 if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_WRITE))) 2005 goto nack_inv; 2006 /* consume RWQE */ 2007 reth = &ohdr->u.rc.reth; 2008 hdrsize += sizeof(*reth); 2009 qp->r_len = be32_to_cpu(reth->length); 2010 qp->r_rcv_len = 0; 2011 qp->r_sge.sg_list = NULL; 2012 if (qp->r_len != 0) { 2013 u32 rkey = be32_to_cpu(reth->rkey); 2014 u64 vaddr = be64_to_cpu(reth->vaddr); 2015 int ok; 2016 2017 /* Check rkey & NAK */ 2018 ok = rvt_rkey_ok(qp, &qp->r_sge.sge, qp->r_len, vaddr, 2019 rkey, IB_ACCESS_REMOTE_WRITE); 2020 if (unlikely(!ok)) 2021 goto nack_acc; 2022 qp->r_sge.num_sge = 1; 2023 } else { 2024 qp->r_sge.num_sge = 0; 2025 qp->r_sge.sge.mr = NULL; 2026 qp->r_sge.sge.vaddr = NULL; 2027 qp->r_sge.sge.length = 0; 2028 qp->r_sge.sge.sge_length = 0; 2029 } 2030 if (opcode == OP(RDMA_WRITE_FIRST)) 2031 goto send_middle; 2032 else if (opcode == OP(RDMA_WRITE_ONLY)) 2033 goto no_immediate_data; 2034 ret = qib_get_rwqe(qp, 1); 2035 if (ret < 0) 2036 goto nack_op_err; 2037 if (!ret) 2038 goto rnr_nak; 2039 wc.ex.imm_data = ohdr->u.rc.imm_data; 2040 hdrsize += 4; 2041 wc.wc_flags = IB_WC_WITH_IMM; 2042 goto send_last; 2043 2044 case OP(RDMA_READ_REQUEST): { 2045 struct rvt_ack_entry *e; 2046 u32 len; 2047 u8 next; 2048 2049 if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_READ))) 2050 goto nack_inv; 2051 next = qp->r_head_ack_queue + 1; 2052 /* s_ack_queue is size QIB_MAX_RDMA_ATOMIC+1 so use > not >= */ 2053 if (next > QIB_MAX_RDMA_ATOMIC) 2054 next = 0; 2055 spin_lock_irqsave(&qp->s_lock, flags); 2056 if (unlikely(next == qp->s_tail_ack_queue)) { 2057 if (!qp->s_ack_queue[next].sent) 2058 goto nack_inv_unlck; 2059 qib_update_ack_queue(qp, next); 2060 } 2061 e = &qp->s_ack_queue[qp->r_head_ack_queue]; 2062 if (e->opcode == OP(RDMA_READ_REQUEST) && e->rdma_sge.mr) { 2063 rvt_put_mr(e->rdma_sge.mr); 2064 e->rdma_sge.mr = NULL; 2065 } 2066 reth = &ohdr->u.rc.reth; 2067 len = be32_to_cpu(reth->length); 2068 if (len) { 2069 u32 rkey = be32_to_cpu(reth->rkey); 2070 u64 vaddr = be64_to_cpu(reth->vaddr); 2071 int ok; 2072 2073 /* Check rkey & NAK */ 2074 ok = rvt_rkey_ok(qp, &e->rdma_sge, len, vaddr, 2075 rkey, IB_ACCESS_REMOTE_READ); 2076 if (unlikely(!ok)) 2077 goto nack_acc_unlck; 2078 /* 2079 * Update the next expected PSN. We add 1 later 2080 * below, so only add the remainder here. 2081 */ 2082 qp->r_psn += rvt_div_mtu(qp, len - 1); 2083 } else { 2084 e->rdma_sge.mr = NULL; 2085 e->rdma_sge.vaddr = NULL; 2086 e->rdma_sge.length = 0; 2087 e->rdma_sge.sge_length = 0; 2088 } 2089 e->opcode = opcode; 2090 e->sent = 0; 2091 e->psn = psn; 2092 e->lpsn = qp->r_psn; 2093 /* 2094 * We need to increment the MSN here instead of when we 2095 * finish sending the result since a duplicate request would 2096 * increment it more than once. 2097 */ 2098 qp->r_msn++; 2099 qp->r_psn++; 2100 qp->r_state = opcode; 2101 qp->r_nak_state = 0; 2102 qp->r_head_ack_queue = next; 2103 2104 /* Schedule the send tasklet. */ 2105 qp->s_flags |= RVT_S_RESP_PENDING; 2106 qib_schedule_send(qp); 2107 2108 goto sunlock; 2109 } 2110 2111 case OP(COMPARE_SWAP): 2112 case OP(FETCH_ADD): { 2113 struct ib_atomic_eth *ateth; 2114 struct rvt_ack_entry *e; 2115 u64 vaddr; 2116 atomic64_t *maddr; 2117 u64 sdata; 2118 u32 rkey; 2119 u8 next; 2120 2121 if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_ATOMIC))) 2122 goto nack_inv; 2123 next = qp->r_head_ack_queue + 1; 2124 if (next > QIB_MAX_RDMA_ATOMIC) 2125 next = 0; 2126 spin_lock_irqsave(&qp->s_lock, flags); 2127 if (unlikely(next == qp->s_tail_ack_queue)) { 2128 if (!qp->s_ack_queue[next].sent) 2129 goto nack_inv_unlck; 2130 qib_update_ack_queue(qp, next); 2131 } 2132 e = &qp->s_ack_queue[qp->r_head_ack_queue]; 2133 if (e->opcode == OP(RDMA_READ_REQUEST) && e->rdma_sge.mr) { 2134 rvt_put_mr(e->rdma_sge.mr); 2135 e->rdma_sge.mr = NULL; 2136 } 2137 ateth = &ohdr->u.atomic_eth; 2138 vaddr = get_ib_ateth_vaddr(ateth); 2139 if (unlikely(vaddr & (sizeof(u64) - 1))) 2140 goto nack_inv_unlck; 2141 rkey = be32_to_cpu(ateth->rkey); 2142 /* Check rkey & NAK */ 2143 if (unlikely(!rvt_rkey_ok(qp, &qp->r_sge.sge, sizeof(u64), 2144 vaddr, rkey, 2145 IB_ACCESS_REMOTE_ATOMIC))) 2146 goto nack_acc_unlck; 2147 /* Perform atomic OP and save result. */ 2148 maddr = (atomic64_t *) qp->r_sge.sge.vaddr; 2149 sdata = get_ib_ateth_swap(ateth); 2150 e->atomic_data = (opcode == OP(FETCH_ADD)) ? 2151 (u64) atomic64_add_return(sdata, maddr) - sdata : 2152 (u64) cmpxchg((u64 *) qp->r_sge.sge.vaddr, 2153 get_ib_ateth_compare(ateth), 2154 sdata); 2155 rvt_put_mr(qp->r_sge.sge.mr); 2156 qp->r_sge.num_sge = 0; 2157 e->opcode = opcode; 2158 e->sent = 0; 2159 e->psn = psn; 2160 e->lpsn = psn; 2161 qp->r_msn++; 2162 qp->r_psn++; 2163 qp->r_state = opcode; 2164 qp->r_nak_state = 0; 2165 qp->r_head_ack_queue = next; 2166 2167 /* Schedule the send tasklet. */ 2168 qp->s_flags |= RVT_S_RESP_PENDING; 2169 qib_schedule_send(qp); 2170 2171 goto sunlock; 2172 } 2173 2174 default: 2175 /* NAK unknown opcodes. */ 2176 goto nack_inv; 2177 } 2178 qp->r_psn++; 2179 qp->r_state = opcode; 2180 qp->r_ack_psn = psn; 2181 qp->r_nak_state = 0; 2182 /* Send an ACK if requested or required. */ 2183 if (psn & (1 << 31)) 2184 goto send_ack; 2185 return; 2186 2187 rnr_nak: 2188 qp->r_nak_state = IB_RNR_NAK | qp->r_min_rnr_timer; 2189 qp->r_ack_psn = qp->r_psn; 2190 /* Queue RNR NAK for later */ 2191 if (list_empty(&qp->rspwait)) { 2192 qp->r_flags |= RVT_R_RSP_NAK; 2193 rvt_get_qp(qp); 2194 list_add_tail(&qp->rspwait, &rcd->qp_wait_list); 2195 } 2196 return; 2197 2198 nack_op_err: 2199 qib_rc_error(qp, IB_WC_LOC_QP_OP_ERR); 2200 qp->r_nak_state = IB_NAK_REMOTE_OPERATIONAL_ERROR; 2201 qp->r_ack_psn = qp->r_psn; 2202 /* Queue NAK for later */ 2203 if (list_empty(&qp->rspwait)) { 2204 qp->r_flags |= RVT_R_RSP_NAK; 2205 rvt_get_qp(qp); 2206 list_add_tail(&qp->rspwait, &rcd->qp_wait_list); 2207 } 2208 return; 2209 2210 nack_inv_unlck: 2211 spin_unlock_irqrestore(&qp->s_lock, flags); 2212 nack_inv: 2213 qib_rc_error(qp, IB_WC_LOC_QP_OP_ERR); 2214 qp->r_nak_state = IB_NAK_INVALID_REQUEST; 2215 qp->r_ack_psn = qp->r_psn; 2216 /* Queue NAK for later */ 2217 if (list_empty(&qp->rspwait)) { 2218 qp->r_flags |= RVT_R_RSP_NAK; 2219 rvt_get_qp(qp); 2220 list_add_tail(&qp->rspwait, &rcd->qp_wait_list); 2221 } 2222 return; 2223 2224 nack_acc_unlck: 2225 spin_unlock_irqrestore(&qp->s_lock, flags); 2226 nack_acc: 2227 qib_rc_error(qp, IB_WC_LOC_PROT_ERR); 2228 qp->r_nak_state = IB_NAK_REMOTE_ACCESS_ERROR; 2229 qp->r_ack_psn = qp->r_psn; 2230 send_ack: 2231 qib_send_rc_ack(qp); 2232 return; 2233 2234 sunlock: 2235 spin_unlock_irqrestore(&qp->s_lock, flags); 2236 } 2237