2 * Copyright(c) 2015, 2016 Intel Corporation.
4 * This file is provided under a dual BSD/GPLv2 license. When using or
5 * redistributing this file, you may do so under either license.
9 * This program is free software; you can redistribute it and/or modify
10 * it under the terms of version 2 of the GNU General Public License as
11 * published by the Free Software Foundation.
13 * This program is distributed in the hope that it will be useful, but
14 * WITHOUT ANY WARRANTY; without even the implied warranty of
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
16 * General Public License for more details.
20 * Redistribution and use in source and binary forms, with or without
21 * modification, are permitted provided that the following conditions
24 * - Redistributions of source code must retain the above copyright
25 * notice, this list of conditions and the following disclaimer.
26 * - Redistributions in binary form must reproduce the above copyright
27 * notice, this list of conditions and the following disclaimer in
28 * the documentation and/or other materials provided with the
30 * - Neither the name of Intel Corporation nor the names of its
31 * contributors may be used to endorse or promote products derived
32 * from this software without specific prior written permission.
34 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
35 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
36 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
37 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
38 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
39 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
40 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
41 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
42 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
43 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
44 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
49 #include <rdma/rdma_vt.h>
50 #include <rdma/rdmavt_qp.h>
54 #include "verbs_txreq.h"
57 /* cut down ridiculously long IB macro names */
58 #define OP(x) RC_OP(x)
60 static u32 restart_sge(struct rvt_sge_state *ss, struct rvt_swqe *wqe,
65 len = delta_psn(psn, wqe->psn) * pmtu;
66 ss->sge = wqe->sg_list[0];
67 ss->sg_list = wqe->sg_list + 1;
68 ss->num_sge = wqe->wr.num_sge;
69 ss->total_len = wqe->length;
70 rvt_skip_sge(ss, len, false);
71 return wqe->length - len;
75 * make_rc_ack - construct a response packet (ACK, NAK, or RDMA read)
76 * @dev: the device for this QP
77 * @qp: a pointer to the QP
78 * @ohdr: a pointer to the IB header being constructed
79 * @ps: the xmit packet state
81 * Return 1 if constructed; otherwise, return 0.
82 * Note that we are in the responder's side of the QP context.
83 * Note the QP s_lock must be held.
85 static int make_rc_ack(struct hfi1_ibdev *dev, struct rvt_qp *qp,
86 struct ib_other_headers *ohdr,
87 struct hfi1_pkt_state *ps)
89 struct rvt_ack_entry *e;
96 struct hfi1_qp_priv *priv = qp->priv;
98 lockdep_assert_held(&qp->s_lock);
99 /* Don't send an ACK if we aren't supposed to. */
100 if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK))
103 /* header size in 32-bit words LRH+BTH = (8+12)/4. */
106 switch (qp->s_ack_state) {
107 case OP(RDMA_READ_RESPONSE_LAST):
108 case OP(RDMA_READ_RESPONSE_ONLY):
109 e = &qp->s_ack_queue[qp->s_tail_ack_queue];
110 if (e->rdma_sge.mr) {
111 rvt_put_mr(e->rdma_sge.mr);
112 e->rdma_sge.mr = NULL;
115 case OP(ATOMIC_ACKNOWLEDGE):
117 * We can increment the tail pointer now that the last
118 * response has been sent instead of only being
121 if (++qp->s_tail_ack_queue > HFI1_MAX_RDMA_ATOMIC)
122 qp->s_tail_ack_queue = 0;
125 case OP(ACKNOWLEDGE):
126 /* Check for no next entry in the queue. */
127 if (qp->r_head_ack_queue == qp->s_tail_ack_queue) {
128 if (qp->s_flags & RVT_S_ACK_PENDING)
133 e = &qp->s_ack_queue[qp->s_tail_ack_queue];
134 if (e->opcode == OP(RDMA_READ_REQUEST)) {
136 * If a RDMA read response is being resent and
137 * we haven't seen the duplicate request yet,
138 * then stop sending the remaining responses the
139 * responder has seen until the requester re-sends it.
141 len = e->rdma_sge.sge_length;
142 if (len && !e->rdma_sge.mr) {
143 qp->s_tail_ack_queue = qp->r_head_ack_queue;
146 /* Copy SGE state in case we need to resend */
147 ps->s_txreq->mr = e->rdma_sge.mr;
149 rvt_get_mr(ps->s_txreq->mr);
150 qp->s_ack_rdma_sge.sge = e->rdma_sge;
151 qp->s_ack_rdma_sge.num_sge = 1;
152 ps->s_txreq->ss = &qp->s_ack_rdma_sge;
155 qp->s_ack_state = OP(RDMA_READ_RESPONSE_FIRST);
157 qp->s_ack_state = OP(RDMA_READ_RESPONSE_ONLY);
160 ohdr->u.aeth = rvt_compute_aeth(qp);
162 qp->s_ack_rdma_psn = e->psn;
163 bth2 = mask_psn(qp->s_ack_rdma_psn++);
165 /* COMPARE_SWAP or FETCH_ADD */
166 ps->s_txreq->ss = NULL;
168 qp->s_ack_state = OP(ATOMIC_ACKNOWLEDGE);
169 ohdr->u.at.aeth = rvt_compute_aeth(qp);
170 ib_u64_put(e->atomic_data, &ohdr->u.at.atomic_ack_eth);
171 hwords += sizeof(ohdr->u.at) / sizeof(u32);
172 bth2 = mask_psn(e->psn);
175 bth0 = qp->s_ack_state << 24;
178 case OP(RDMA_READ_RESPONSE_FIRST):
179 qp->s_ack_state = OP(RDMA_READ_RESPONSE_MIDDLE);
181 case OP(RDMA_READ_RESPONSE_MIDDLE):
182 ps->s_txreq->ss = &qp->s_ack_rdma_sge;
183 ps->s_txreq->mr = qp->s_ack_rdma_sge.sge.mr;
185 rvt_get_mr(ps->s_txreq->mr);
186 len = qp->s_ack_rdma_sge.sge.sge_length;
189 middle = HFI1_CAP_IS_KSET(SDMA_AHG);
191 ohdr->u.aeth = rvt_compute_aeth(qp);
193 qp->s_ack_state = OP(RDMA_READ_RESPONSE_LAST);
194 e = &qp->s_ack_queue[qp->s_tail_ack_queue];
197 bth0 = qp->s_ack_state << 24;
198 bth2 = mask_psn(qp->s_ack_rdma_psn++);
204 * Send a regular ACK.
205 * Set the s_ack_state so we wait until after sending
206 * the ACK before setting s_ack_state to ACKNOWLEDGE
209 qp->s_ack_state = OP(SEND_ONLY);
210 qp->s_flags &= ~RVT_S_ACK_PENDING;
211 ps->s_txreq->ss = NULL;
214 cpu_to_be32((qp->r_msn & IB_MSN_MASK) |
216 IB_AETH_CREDIT_SHIFT));
218 ohdr->u.aeth = rvt_compute_aeth(qp);
221 bth0 = OP(ACKNOWLEDGE) << 24;
222 bth2 = mask_psn(qp->s_ack_psn);
224 qp->s_rdma_ack_cnt++;
225 qp->s_hdrwords = hwords;
226 ps->s_txreq->sde = priv->s_sde;
227 ps->s_txreq->s_cur_size = len;
228 hfi1_make_ruc_header(qp, ohdr, bth0, bth2, middle, ps);
230 ps->s_txreq->hdr_dwords = qp->s_hdrwords + 2;
234 qp->s_ack_state = OP(ACKNOWLEDGE);
236 * Ensure s_rdma_ack_cnt changes are committed prior to resetting
240 qp->s_flags &= ~(RVT_S_RESP_PENDING
247 * hfi1_make_rc_req - construct a request packet (SEND, RDMA r/w, ATOMIC)
248 * @qp: a pointer to the QP
250 * Assumes s_lock is held.
252 * Return 1 if constructed; otherwise, return 0.
254 int hfi1_make_rc_req(struct rvt_qp *qp, struct hfi1_pkt_state *ps)
256 struct hfi1_qp_priv *priv = qp->priv;
257 struct hfi1_ibdev *dev = to_idev(qp->ibqp.device);
258 struct ib_other_headers *ohdr;
259 struct rvt_sge_state *ss;
260 struct rvt_swqe *wqe;
261 /* header size in 32-bit words LRH+BTH = (8+12)/4. */
271 lockdep_assert_held(&qp->s_lock);
272 ps->s_txreq = get_txreq(ps->dev, qp);
273 if (IS_ERR(ps->s_txreq))
276 ohdr = &ps->s_txreq->phdr.hdr.u.oth;
277 if (qp->remote_ah_attr.ah_flags & IB_AH_GRH)
278 ohdr = &ps->s_txreq->phdr.hdr.u.l.oth;
280 /* Sending responses has higher priority over sending requests. */
281 if ((qp->s_flags & RVT_S_RESP_PENDING) &&
282 make_rc_ack(dev, qp, ohdr, ps))
285 if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_SEND_OK)) {
286 if (!(ib_rvt_state_ops[qp->state] & RVT_FLUSH_SEND))
288 /* We are in the error state, flush the work request. */
289 smp_read_barrier_depends(); /* see post_one_send() */
290 if (qp->s_last == READ_ONCE(qp->s_head))
292 /* If DMAs are in progress, we can't flush immediately. */
293 if (iowait_sdma_pending(&priv->s_iowait)) {
294 qp->s_flags |= RVT_S_WAIT_DMA;
298 wqe = rvt_get_swqe_ptr(qp, qp->s_last);
299 hfi1_send_complete(qp, wqe, qp->s_last != qp->s_acked ?
300 IB_WC_SUCCESS : IB_WC_WR_FLUSH_ERR);
301 /* will get called again */
305 if (qp->s_flags & (RVT_S_WAIT_RNR | RVT_S_WAIT_ACK))
308 if (cmp_psn(qp->s_psn, qp->s_sending_hpsn) <= 0) {
309 if (cmp_psn(qp->s_sending_psn, qp->s_sending_hpsn) <= 0) {
310 qp->s_flags |= RVT_S_WAIT_PSN;
313 qp->s_sending_psn = qp->s_psn;
314 qp->s_sending_hpsn = qp->s_psn - 1;
317 /* Send a request. */
318 wqe = rvt_get_swqe_ptr(qp, qp->s_cur);
319 switch (qp->s_state) {
321 if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_NEXT_SEND_OK))
324 * Resend an old request or start a new one.
326 * We keep track of the current SWQE so that
327 * we don't reset the "furthest progress" state
328 * if we need to back up.
331 if (qp->s_cur == qp->s_tail) {
332 /* Check if send work queue is empty. */
333 smp_read_barrier_depends(); /* see post_one_send() */
334 if (qp->s_tail == READ_ONCE(qp->s_head)) {
339 * If a fence is requested, wait for previous
340 * RDMA read and atomic operations to finish.
342 if ((wqe->wr.send_flags & IB_SEND_FENCE) &&
343 qp->s_num_rd_atomic) {
344 qp->s_flags |= RVT_S_WAIT_FENCE;
348 * Local operations are processed immediately
349 * after all prior requests have completed
351 if (wqe->wr.opcode == IB_WR_REG_MR ||
352 wqe->wr.opcode == IB_WR_LOCAL_INV) {
356 if (qp->s_last != qp->s_cur)
358 if (++qp->s_cur == qp->s_size)
360 if (++qp->s_tail == qp->s_size)
362 if (!(wqe->wr.send_flags &
363 RVT_SEND_COMPLETION_ONLY)) {
364 err = rvt_invalidate_rkey(
366 wqe->wr.ex.invalidate_rkey);
369 hfi1_send_complete(qp, wqe,
370 err ? IB_WC_LOC_PROT_ERR
373 atomic_dec(&qp->local_ops_pending);
379 qp->s_psn = wqe->psn;
382 * Note that we have to be careful not to modify the
383 * original work request since we may need to resend
388 bth2 = mask_psn(qp->s_psn);
389 switch (wqe->wr.opcode) {
391 case IB_WR_SEND_WITH_IMM:
392 case IB_WR_SEND_WITH_INV:
393 /* If no credit, return. */
394 if (!(qp->s_flags & RVT_S_UNLIMITED_CREDIT) &&
395 rvt_cmp_msn(wqe->ssn, qp->s_lsn + 1) > 0) {
396 qp->s_flags |= RVT_S_WAIT_SSN_CREDIT;
400 qp->s_state = OP(SEND_FIRST);
404 if (wqe->wr.opcode == IB_WR_SEND) {
405 qp->s_state = OP(SEND_ONLY);
406 } else if (wqe->wr.opcode == IB_WR_SEND_WITH_IMM) {
407 qp->s_state = OP(SEND_ONLY_WITH_IMMEDIATE);
408 /* Immediate data comes after the BTH */
409 ohdr->u.imm_data = wqe->wr.ex.imm_data;
412 qp->s_state = OP(SEND_ONLY_WITH_INVALIDATE);
413 /* Invalidate rkey comes after the BTH */
414 ohdr->u.ieth = cpu_to_be32(
415 wqe->wr.ex.invalidate_rkey);
418 if (wqe->wr.send_flags & IB_SEND_SOLICITED)
419 bth0 |= IB_BTH_SOLICITED;
420 bth2 |= IB_BTH_REQ_ACK;
421 if (++qp->s_cur == qp->s_size)
425 case IB_WR_RDMA_WRITE:
426 if (newreq && !(qp->s_flags & RVT_S_UNLIMITED_CREDIT))
429 case IB_WR_RDMA_WRITE_WITH_IMM:
430 /* If no credit, return. */
431 if (!(qp->s_flags & RVT_S_UNLIMITED_CREDIT) &&
432 rvt_cmp_msn(wqe->ssn, qp->s_lsn + 1) > 0) {
433 qp->s_flags |= RVT_S_WAIT_SSN_CREDIT;
437 wqe->rdma_wr.remote_addr,
439 ohdr->u.rc.reth.rkey =
440 cpu_to_be32(wqe->rdma_wr.rkey);
441 ohdr->u.rc.reth.length = cpu_to_be32(len);
442 hwords += sizeof(struct ib_reth) / sizeof(u32);
444 qp->s_state = OP(RDMA_WRITE_FIRST);
448 if (wqe->wr.opcode == IB_WR_RDMA_WRITE) {
449 qp->s_state = OP(RDMA_WRITE_ONLY);
452 OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE);
453 /* Immediate data comes after RETH */
454 ohdr->u.rc.imm_data = wqe->wr.ex.imm_data;
456 if (wqe->wr.send_flags & IB_SEND_SOLICITED)
457 bth0 |= IB_BTH_SOLICITED;
459 bth2 |= IB_BTH_REQ_ACK;
460 if (++qp->s_cur == qp->s_size)
464 case IB_WR_RDMA_READ:
466 * Don't allow more operations to be started
467 * than the QP limits allow.
470 if (qp->s_num_rd_atomic >=
471 qp->s_max_rd_atomic) {
472 qp->s_flags |= RVT_S_WAIT_RDMAR;
475 qp->s_num_rd_atomic++;
476 if (!(qp->s_flags & RVT_S_UNLIMITED_CREDIT))
480 wqe->rdma_wr.remote_addr,
482 ohdr->u.rc.reth.rkey =
483 cpu_to_be32(wqe->rdma_wr.rkey);
484 ohdr->u.rc.reth.length = cpu_to_be32(len);
485 qp->s_state = OP(RDMA_READ_REQUEST);
486 hwords += sizeof(ohdr->u.rc.reth) / sizeof(u32);
489 bth2 |= IB_BTH_REQ_ACK;
490 if (++qp->s_cur == qp->s_size)
494 case IB_WR_ATOMIC_CMP_AND_SWP:
495 case IB_WR_ATOMIC_FETCH_AND_ADD:
497 * Don't allow more operations to be started
498 * than the QP limits allow.
501 if (qp->s_num_rd_atomic >=
502 qp->s_max_rd_atomic) {
503 qp->s_flags |= RVT_S_WAIT_RDMAR;
506 qp->s_num_rd_atomic++;
507 if (!(qp->s_flags & RVT_S_UNLIMITED_CREDIT))
510 if (wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
511 qp->s_state = OP(COMPARE_SWAP);
512 put_ib_ateth_swap(wqe->atomic_wr.swap,
513 &ohdr->u.atomic_eth);
514 put_ib_ateth_compare(wqe->atomic_wr.compare_add,
515 &ohdr->u.atomic_eth);
517 qp->s_state = OP(FETCH_ADD);
518 put_ib_ateth_swap(wqe->atomic_wr.compare_add,
519 &ohdr->u.atomic_eth);
520 put_ib_ateth_compare(0, &ohdr->u.atomic_eth);
522 put_ib_ateth_vaddr(wqe->atomic_wr.remote_addr,
523 &ohdr->u.atomic_eth);
524 ohdr->u.atomic_eth.rkey = cpu_to_be32(
525 wqe->atomic_wr.rkey);
526 hwords += sizeof(struct ib_atomic_eth) / sizeof(u32);
529 bth2 |= IB_BTH_REQ_ACK;
530 if (++qp->s_cur == qp->s_size)
537 qp->s_sge.sge = wqe->sg_list[0];
538 qp->s_sge.sg_list = wqe->sg_list + 1;
539 qp->s_sge.num_sge = wqe->wr.num_sge;
540 qp->s_sge.total_len = wqe->length;
541 qp->s_len = wqe->length;
544 if (qp->s_tail >= qp->s_size)
547 if (wqe->wr.opcode == IB_WR_RDMA_READ)
548 qp->s_psn = wqe->lpsn + 1;
553 case OP(RDMA_READ_RESPONSE_FIRST):
555 * qp->s_state is normally set to the opcode of the
556 * last packet constructed for new requests and therefore
557 * is never set to RDMA read response.
558 * RDMA_READ_RESPONSE_FIRST is used by the ACK processing
559 * thread to indicate a SEND needs to be restarted from an
560 * earlier PSN without interfering with the sending thread.
563 qp->s_len = restart_sge(&qp->s_sge, wqe, qp->s_psn, pmtu);
566 qp->s_state = OP(SEND_MIDDLE);
568 case OP(SEND_MIDDLE):
569 bth2 = mask_psn(qp->s_psn++);
574 middle = HFI1_CAP_IS_KSET(SDMA_AHG);
577 if (wqe->wr.opcode == IB_WR_SEND) {
578 qp->s_state = OP(SEND_LAST);
579 } else if (wqe->wr.opcode == IB_WR_SEND_WITH_IMM) {
580 qp->s_state = OP(SEND_LAST_WITH_IMMEDIATE);
581 /* Immediate data comes after the BTH */
582 ohdr->u.imm_data = wqe->wr.ex.imm_data;
585 qp->s_state = OP(SEND_LAST_WITH_INVALIDATE);
586 /* invalidate data comes after the BTH */
587 ohdr->u.ieth = cpu_to_be32(wqe->wr.ex.invalidate_rkey);
590 if (wqe->wr.send_flags & IB_SEND_SOLICITED)
591 bth0 |= IB_BTH_SOLICITED;
592 bth2 |= IB_BTH_REQ_ACK;
594 if (qp->s_cur >= qp->s_size)
598 case OP(RDMA_READ_RESPONSE_LAST):
600 * qp->s_state is normally set to the opcode of the
601 * last packet constructed for new requests and therefore
602 * is never set to RDMA read response.
603 * RDMA_READ_RESPONSE_LAST is used by the ACK processing
604 * thread to indicate a RDMA write needs to be restarted from
605 * an earlier PSN without interfering with the sending thread.
608 qp->s_len = restart_sge(&qp->s_sge, wqe, qp->s_psn, pmtu);
610 case OP(RDMA_WRITE_FIRST):
611 qp->s_state = OP(RDMA_WRITE_MIDDLE);
613 case OP(RDMA_WRITE_MIDDLE):
614 bth2 = mask_psn(qp->s_psn++);
619 middle = HFI1_CAP_IS_KSET(SDMA_AHG);
622 if (wqe->wr.opcode == IB_WR_RDMA_WRITE) {
623 qp->s_state = OP(RDMA_WRITE_LAST);
625 qp->s_state = OP(RDMA_WRITE_LAST_WITH_IMMEDIATE);
626 /* Immediate data comes after the BTH */
627 ohdr->u.imm_data = wqe->wr.ex.imm_data;
629 if (wqe->wr.send_flags & IB_SEND_SOLICITED)
630 bth0 |= IB_BTH_SOLICITED;
632 bth2 |= IB_BTH_REQ_ACK;
634 if (qp->s_cur >= qp->s_size)
638 case OP(RDMA_READ_RESPONSE_MIDDLE):
640 * qp->s_state is normally set to the opcode of the
641 * last packet constructed for new requests and therefore
642 * is never set to RDMA read response.
643 * RDMA_READ_RESPONSE_MIDDLE is used by the ACK processing
644 * thread to indicate a RDMA read needs to be restarted from
645 * an earlier PSN without interfering with the sending thread.
648 len = (delta_psn(qp->s_psn, wqe->psn)) * pmtu;
650 wqe->rdma_wr.remote_addr + len,
652 ohdr->u.rc.reth.rkey =
653 cpu_to_be32(wqe->rdma_wr.rkey);
654 ohdr->u.rc.reth.length = cpu_to_be32(wqe->length - len);
655 qp->s_state = OP(RDMA_READ_REQUEST);
656 hwords += sizeof(ohdr->u.rc.reth) / sizeof(u32);
657 bth2 = mask_psn(qp->s_psn) | IB_BTH_REQ_ACK;
658 qp->s_psn = wqe->lpsn + 1;
662 if (qp->s_cur == qp->s_size)
666 qp->s_sending_hpsn = bth2;
667 delta = delta_psn(bth2, wqe->psn);
668 if (delta && delta % HFI1_PSN_CREDIT == 0)
669 bth2 |= IB_BTH_REQ_ACK;
670 if (qp->s_flags & RVT_S_SEND_ONE) {
671 qp->s_flags &= ~RVT_S_SEND_ONE;
672 qp->s_flags |= RVT_S_WAIT_ACK;
673 bth2 |= IB_BTH_REQ_ACK;
676 qp->s_hdrwords = hwords;
677 ps->s_txreq->sde = priv->s_sde;
678 ps->s_txreq->ss = ss;
679 ps->s_txreq->s_cur_size = len;
680 hfi1_make_ruc_header(
683 bth0 | (qp->s_state << 24),
688 ps->s_txreq->hdr_dwords = qp->s_hdrwords + 2;
692 hfi1_put_txreq(ps->s_txreq);
697 hfi1_put_txreq(ps->s_txreq);
701 qp->s_flags &= ~RVT_S_BUSY;
707 * hfi1_send_rc_ack - Construct an ACK packet and send it
708 * @qp: a pointer to the QP
710 * This is called from hfi1_rc_rcv() and handle_receive_interrupt().
711 * Note that RDMA reads and atomics are handled in the
712 * send side QP state and send engine.
714 void hfi1_send_rc_ack(struct hfi1_ctxtdata *rcd, struct rvt_qp *qp,
717 struct hfi1_ibport *ibp = rcd_to_iport(rcd);
718 struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
719 u64 pbc, pbc_flags = 0;
725 struct send_context *sc;
726 struct pio_buf *pbuf;
727 struct ib_header hdr;
728 struct ib_other_headers *ohdr;
730 struct hfi1_qp_priv *priv = qp->priv;
732 /* clear the defer count */
733 priv->r_adefered = 0;
735 /* Don't send ACK or NAK if a RDMA read or atomic is pending. */
736 if (qp->s_flags & RVT_S_RESP_PENDING)
739 /* Ensure s_rdma_ack_cnt changes are committed */
740 smp_read_barrier_depends();
741 if (qp->s_rdma_ack_cnt)
744 /* Construct the header */
745 /* header size in 32-bit words LRH+BTH+AETH = (8+12+4)/4 */
747 if (unlikely(qp->remote_ah_attr.ah_flags & IB_AH_GRH)) {
748 hwords += hfi1_make_grh(ibp, &hdr.u.l.grh,
749 &qp->remote_ah_attr.grh, hwords, 0);
756 /* read pkey_index w/o lock (its atomic) */
757 bth0 = hfi1_get_pkey(ibp, qp->s_pkey_index) | (OP(ACKNOWLEDGE) << 24);
758 if (qp->s_mig_state == IB_MIG_MIGRATED)
759 bth0 |= IB_BTH_MIG_REQ;
761 ohdr->u.aeth = cpu_to_be32((qp->r_msn & IB_MSN_MASK) |
763 IB_AETH_CREDIT_SHIFT));
765 ohdr->u.aeth = rvt_compute_aeth(qp);
766 sc5 = ibp->sl_to_sc[qp->remote_ah_attr.sl];
767 /* set PBC_DC_INFO bit (aka SC[4]) in pbc_flags */
768 pbc_flags |= ((!!(sc5 & 0x10)) << PBC_DC_INFO_SHIFT);
769 lrh0 |= (sc5 & 0xf) << 12 | (qp->remote_ah_attr.sl & 0xf) << 4;
770 hdr.lrh[0] = cpu_to_be16(lrh0);
771 hdr.lrh[1] = cpu_to_be16(qp->remote_ah_attr.dlid);
772 hdr.lrh[2] = cpu_to_be16(hwords + SIZE_OF_CRC);
773 hdr.lrh[3] = cpu_to_be16(ppd->lid | qp->remote_ah_attr.src_path_bits);
774 ohdr->bth[0] = cpu_to_be32(bth0);
775 ohdr->bth[1] = cpu_to_be32(qp->remote_qpn);
776 ohdr->bth[1] |= cpu_to_be32((!!is_fecn) << HFI1_BECN_SHIFT);
777 ohdr->bth[2] = cpu_to_be32(mask_psn(qp->r_ack_psn));
779 /* Don't try to send ACKs if the link isn't ACTIVE */
780 if (driver_lstate(ppd) != IB_PORT_ACTIVE)
784 plen = 2 /* PBC */ + hwords;
785 vl = sc_to_vlt(ppd->dd, sc5);
786 pbc = create_pbc(ppd, pbc_flags, qp->srate_mbps, vl, plen);
788 pbuf = sc_buffer_alloc(sc, plen, NULL, NULL);
791 * We have no room to send at the moment. Pass
792 * responsibility for sending the ACK to the send engine
793 * so that when enough buffer space becomes available,
794 * the ACK is sent ahead of other outgoing packets.
799 trace_ack_output_ibhdr(dd_from_ibdev(qp->ibqp.device), &hdr);
801 /* write the pbc and data */
802 ppd->dd->pio_inline_send(ppd->dd, pbuf, pbc, &hdr, hwords);
807 spin_lock_irqsave(&qp->s_lock, flags);
808 if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK))
810 this_cpu_inc(*ibp->rvp.rc_qacks);
811 qp->s_flags |= RVT_S_ACK_PENDING | RVT_S_RESP_PENDING;
812 qp->s_nak_state = qp->r_nak_state;
813 qp->s_ack_psn = qp->r_ack_psn;
815 qp->s_flags |= RVT_S_ECN;
817 /* Schedule the send engine. */
818 hfi1_schedule_send(qp);
820 spin_unlock_irqrestore(&qp->s_lock, flags);
824 * reset_psn - reset the QP state to send starting from PSN
826 * @psn: the packet sequence number to restart at
828 * This is called from hfi1_rc_rcv() to process an incoming RC ACK
830 * Called at interrupt level with the QP s_lock held.
832 static void reset_psn(struct rvt_qp *qp, u32 psn)
835 struct rvt_swqe *wqe = rvt_get_swqe_ptr(qp, n);
838 lockdep_assert_held(&qp->s_lock);
842 * If we are starting the request from the beginning,
843 * let the normal send code handle initialization.
845 if (cmp_psn(psn, wqe->psn) <= 0) {
846 qp->s_state = OP(SEND_LAST);
850 /* Find the work request opcode corresponding to the given PSN. */
851 opcode = wqe->wr.opcode;
855 if (++n == qp->s_size)
859 wqe = rvt_get_swqe_ptr(qp, n);
860 diff = cmp_psn(psn, wqe->psn);
865 * If we are starting the request from the beginning,
866 * let the normal send code handle initialization.
869 qp->s_state = OP(SEND_LAST);
872 opcode = wqe->wr.opcode;
876 * Set the state to restart in the middle of a request.
877 * Don't change the s_sge, s_cur_sge, or s_cur_size.
878 * See hfi1_make_rc_req().
882 case IB_WR_SEND_WITH_IMM:
883 qp->s_state = OP(RDMA_READ_RESPONSE_FIRST);
886 case IB_WR_RDMA_WRITE:
887 case IB_WR_RDMA_WRITE_WITH_IMM:
888 qp->s_state = OP(RDMA_READ_RESPONSE_LAST);
891 case IB_WR_RDMA_READ:
892 qp->s_state = OP(RDMA_READ_RESPONSE_MIDDLE);
897 * This case shouldn't happen since its only
900 qp->s_state = OP(SEND_LAST);
905 * Set RVT_S_WAIT_PSN as rc_complete() may start the timer
906 * asynchronously before the send engine can get scheduled.
907 * Doing it in hfi1_make_rc_req() is too late.
909 if ((cmp_psn(qp->s_psn, qp->s_sending_hpsn) <= 0) &&
910 (cmp_psn(qp->s_sending_psn, qp->s_sending_hpsn) <= 0))
911 qp->s_flags |= RVT_S_WAIT_PSN;
912 qp->s_flags &= ~RVT_S_AHG_VALID;
916 * Back up requester to resend the last un-ACKed request.
917 * The QP r_lock and s_lock should be held and interrupts disabled.
919 void hfi1_restart_rc(struct rvt_qp *qp, u32 psn, int wait)
921 struct rvt_swqe *wqe = rvt_get_swqe_ptr(qp, qp->s_acked);
922 struct hfi1_ibport *ibp;
924 lockdep_assert_held(&qp->r_lock);
925 lockdep_assert_held(&qp->s_lock);
926 if (qp->s_retry == 0) {
927 if (qp->s_mig_state == IB_MIG_ARMED) {
929 qp->s_retry = qp->s_retry_cnt;
930 } else if (qp->s_last == qp->s_acked) {
931 hfi1_send_complete(qp, wqe, IB_WC_RETRY_EXC_ERR);
932 rvt_error_qp(qp, IB_WC_WR_FLUSH_ERR);
934 } else { /* need to handle delayed completion */
941 ibp = to_iport(qp->ibqp.device, qp->port_num);
942 if (wqe->wr.opcode == IB_WR_RDMA_READ)
943 ibp->rvp.n_rc_resends++;
945 ibp->rvp.n_rc_resends += delta_psn(qp->s_psn, psn);
947 qp->s_flags &= ~(RVT_S_WAIT_FENCE | RVT_S_WAIT_RDMAR |
948 RVT_S_WAIT_SSN_CREDIT | RVT_S_WAIT_PSN |
951 qp->s_flags |= RVT_S_SEND_ONE;
956 * Set qp->s_sending_psn to the next PSN after the given one.
957 * This would be psn+1 except when RDMA reads are present.
959 static void reset_sending_psn(struct rvt_qp *qp, u32 psn)
961 struct rvt_swqe *wqe;
964 lockdep_assert_held(&qp->s_lock);
965 /* Find the work request corresponding to the given PSN. */
967 wqe = rvt_get_swqe_ptr(qp, n);
968 if (cmp_psn(psn, wqe->lpsn) <= 0) {
969 if (wqe->wr.opcode == IB_WR_RDMA_READ)
970 qp->s_sending_psn = wqe->lpsn + 1;
972 qp->s_sending_psn = psn + 1;
975 if (++n == qp->s_size)
983 * This should be called with the QP s_lock held and interrupts disabled.
985 void hfi1_rc_send_complete(struct rvt_qp *qp, struct ib_header *hdr)
987 struct ib_other_headers *ohdr;
988 struct rvt_swqe *wqe;
992 lockdep_assert_held(&qp->s_lock);
993 if (!(ib_rvt_state_ops[qp->state] & RVT_SEND_OR_FLUSH_OR_RECV_OK))
996 /* Find out where the BTH is */
997 if ((be16_to_cpu(hdr->lrh[0]) & 3) == HFI1_LRH_BTH)
1000 ohdr = &hdr->u.l.oth;
1002 opcode = be32_to_cpu(ohdr->bth[0]) >> 24;
1003 if (opcode >= OP(RDMA_READ_RESPONSE_FIRST) &&
1004 opcode <= OP(ATOMIC_ACKNOWLEDGE)) {
1005 WARN_ON(!qp->s_rdma_ack_cnt);
1006 qp->s_rdma_ack_cnt--;
1010 psn = be32_to_cpu(ohdr->bth[2]);
1011 reset_sending_psn(qp, psn);
1014 * Start timer after a packet requesting an ACK has been sent and
1015 * there are still requests that haven't been acked.
1017 if ((psn & IB_BTH_REQ_ACK) && qp->s_acked != qp->s_tail &&
1019 (RVT_S_TIMER | RVT_S_WAIT_RNR | RVT_S_WAIT_PSN)) &&
1020 (ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK))
1021 rvt_add_retry_timer(qp);
1023 while (qp->s_last != qp->s_acked) {
1026 wqe = rvt_get_swqe_ptr(qp, qp->s_last);
1027 if (cmp_psn(wqe->lpsn, qp->s_sending_psn) >= 0 &&
1028 cmp_psn(qp->s_sending_psn, qp->s_sending_hpsn) <= 0)
1030 s_last = qp->s_last;
1031 if (++s_last >= qp->s_size)
1033 qp->s_last = s_last;
1034 /* see post_send() */
1037 rvt_qp_swqe_complete(qp, wqe, IB_WC_SUCCESS);
1040 * If we were waiting for sends to complete before re-sending,
1041 * and they are now complete, restart sending.
1043 trace_hfi1_sendcomplete(qp, psn);
1044 if (qp->s_flags & RVT_S_WAIT_PSN &&
1045 cmp_psn(qp->s_sending_psn, qp->s_sending_hpsn) > 0) {
1046 qp->s_flags &= ~RVT_S_WAIT_PSN;
1047 qp->s_sending_psn = qp->s_psn;
1048 qp->s_sending_hpsn = qp->s_psn - 1;
1049 hfi1_schedule_send(qp);
1053 static inline void update_last_psn(struct rvt_qp *qp, u32 psn)
1055 qp->s_last_psn = psn;
1059 * Generate a SWQE completion.
1060 * This is similar to hfi1_send_complete but has to check to be sure
1061 * that the SGEs are not being referenced if the SWQE is being resent.
1063 static struct rvt_swqe *do_rc_completion(struct rvt_qp *qp,
1064 struct rvt_swqe *wqe,
1065 struct hfi1_ibport *ibp)
1067 lockdep_assert_held(&qp->s_lock);
1069 * Don't decrement refcount and don't generate a
1070 * completion if the SWQE is being resent until the send
1073 if (cmp_psn(wqe->lpsn, qp->s_sending_psn) < 0 ||
1074 cmp_psn(qp->s_sending_psn, qp->s_sending_hpsn) > 0) {
1078 s_last = qp->s_last;
1079 if (++s_last >= qp->s_size)
1081 qp->s_last = s_last;
1082 /* see post_send() */
1084 rvt_qp_swqe_complete(qp, wqe, IB_WC_SUCCESS);
1086 struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
1088 this_cpu_inc(*ibp->rvp.rc_delayed_comp);
1090 * If send progress not running attempt to progress
1093 if (ppd->dd->flags & HFI1_HAS_SEND_DMA) {
1094 struct sdma_engine *engine;
1097 /* For now use sc to find engine */
1098 sc5 = ibp->sl_to_sc[qp->remote_ah_attr.sl];
1099 engine = qp_to_sdma_engine(qp, sc5);
1100 sdma_engine_progress_schedule(engine);
1104 qp->s_retry = qp->s_retry_cnt;
1105 update_last_psn(qp, wqe->lpsn);
1108 * If we are completing a request which is in the process of
1109 * being resent, we can stop re-sending it since we know the
1110 * responder has already seen it.
1112 if (qp->s_acked == qp->s_cur) {
1113 if (++qp->s_cur >= qp->s_size)
1115 qp->s_acked = qp->s_cur;
1116 wqe = rvt_get_swqe_ptr(qp, qp->s_cur);
1117 if (qp->s_acked != qp->s_tail) {
1118 qp->s_state = OP(SEND_LAST);
1119 qp->s_psn = wqe->psn;
1122 if (++qp->s_acked >= qp->s_size)
1124 if (qp->state == IB_QPS_SQD && qp->s_acked == qp->s_cur)
1126 wqe = rvt_get_swqe_ptr(qp, qp->s_acked);
1132 * do_rc_ack - process an incoming RC ACK
1133 * @qp: the QP the ACK came in on
1134 * @psn: the packet sequence number of the ACK
1135 * @opcode: the opcode of the request that resulted in the ACK
1137 * This is called from rc_rcv_resp() to process an incoming RC ACK
1139 * May be called at interrupt level, with the QP s_lock held.
1140 * Returns 1 if OK, 0 if current operation should be aborted (NAK).
1142 static int do_rc_ack(struct rvt_qp *qp, u32 aeth, u32 psn, int opcode,
1143 u64 val, struct hfi1_ctxtdata *rcd)
1145 struct hfi1_ibport *ibp;
1146 enum ib_wc_status status;
1147 struct rvt_swqe *wqe;
1152 lockdep_assert_held(&qp->s_lock);
1154 * Note that NAKs implicitly ACK outstanding SEND and RDMA write
1155 * requests and implicitly NAK RDMA read and atomic requests issued
1156 * before the NAK'ed request. The MSN won't include the NAK'ed
1157 * request but will include an ACK'ed request(s).
1160 if (aeth >> IB_AETH_NAK_SHIFT)
1162 wqe = rvt_get_swqe_ptr(qp, qp->s_acked);
1163 ibp = rcd_to_iport(rcd);
1166 * The MSN might be for a later WQE than the PSN indicates so
1167 * only complete WQEs that the PSN finishes.
1169 while ((diff = delta_psn(ack_psn, wqe->lpsn)) >= 0) {
1171 * RDMA_READ_RESPONSE_ONLY is a special case since
1172 * we want to generate completion events for everything
1173 * before the RDMA read, copy the data, then generate
1174 * the completion for the read.
1176 if (wqe->wr.opcode == IB_WR_RDMA_READ &&
1177 opcode == OP(RDMA_READ_RESPONSE_ONLY) &&
1183 * If this request is a RDMA read or atomic, and the ACK is
1184 * for a later operation, this ACK NAKs the RDMA read or
1185 * atomic. In other words, only a RDMA_READ_LAST or ONLY
1186 * can ACK a RDMA read and likewise for atomic ops. Note
1187 * that the NAK case can only happen if relaxed ordering is
1188 * used and requests are sent after an RDMA read or atomic
1189 * is sent but before the response is received.
1191 if ((wqe->wr.opcode == IB_WR_RDMA_READ &&
1192 (opcode != OP(RDMA_READ_RESPONSE_LAST) || diff != 0)) ||
1193 ((wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
1194 wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD) &&
1195 (opcode != OP(ATOMIC_ACKNOWLEDGE) || diff != 0))) {
1196 /* Retry this request. */
1197 if (!(qp->r_flags & RVT_R_RDMAR_SEQ)) {
1198 qp->r_flags |= RVT_R_RDMAR_SEQ;
1199 hfi1_restart_rc(qp, qp->s_last_psn + 1, 0);
1200 if (list_empty(&qp->rspwait)) {
1201 qp->r_flags |= RVT_R_RSP_SEND;
1203 list_add_tail(&qp->rspwait,
1204 &rcd->qp_wait_list);
1208 * No need to process the ACK/NAK since we are
1209 * restarting an earlier request.
1213 if (wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
1214 wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD) {
1215 u64 *vaddr = wqe->sg_list[0].vaddr;
1218 if (qp->s_num_rd_atomic &&
1219 (wqe->wr.opcode == IB_WR_RDMA_READ ||
1220 wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
1221 wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD)) {
1222 qp->s_num_rd_atomic--;
1223 /* Restart sending task if fence is complete */
1224 if ((qp->s_flags & RVT_S_WAIT_FENCE) &&
1225 !qp->s_num_rd_atomic) {
1226 qp->s_flags &= ~(RVT_S_WAIT_FENCE |
1228 hfi1_schedule_send(qp);
1229 } else if (qp->s_flags & RVT_S_WAIT_RDMAR) {
1230 qp->s_flags &= ~(RVT_S_WAIT_RDMAR |
1232 hfi1_schedule_send(qp);
1235 wqe = do_rc_completion(qp, wqe, ibp);
1236 if (qp->s_acked == qp->s_tail)
1240 switch (aeth >> IB_AETH_NAK_SHIFT) {
1242 this_cpu_inc(*ibp->rvp.rc_acks);
1243 if (qp->s_acked != qp->s_tail) {
1245 * We are expecting more ACKs so
1246 * mod the retry timer.
1248 rvt_mod_retry_timer(qp);
1250 * We can stop re-sending the earlier packets and
1251 * continue with the next packet the receiver wants.
1253 if (cmp_psn(qp->s_psn, psn) <= 0)
1254 reset_psn(qp, psn + 1);
1256 /* No more acks - kill all timers */
1257 rvt_stop_rc_timers(qp);
1258 if (cmp_psn(qp->s_psn, psn) <= 0) {
1259 qp->s_state = OP(SEND_LAST);
1260 qp->s_psn = psn + 1;
1263 if (qp->s_flags & RVT_S_WAIT_ACK) {
1264 qp->s_flags &= ~RVT_S_WAIT_ACK;
1265 hfi1_schedule_send(qp);
1267 rvt_get_credit(qp, aeth);
1268 qp->s_rnr_retry = qp->s_rnr_retry_cnt;
1269 qp->s_retry = qp->s_retry_cnt;
1270 update_last_psn(qp, psn);
1273 case 1: /* RNR NAK */
1274 ibp->rvp.n_rnr_naks++;
1275 if (qp->s_acked == qp->s_tail)
1277 if (qp->s_flags & RVT_S_WAIT_RNR)
1279 if (qp->s_rnr_retry == 0) {
1280 status = IB_WC_RNR_RETRY_EXC_ERR;
1283 if (qp->s_rnr_retry_cnt < 7)
1286 /* The last valid PSN is the previous PSN. */
1287 update_last_psn(qp, psn - 1);
1289 ibp->rvp.n_rc_resends += delta_psn(qp->s_psn, psn);
1293 qp->s_flags &= ~(RVT_S_WAIT_SSN_CREDIT | RVT_S_WAIT_ACK);
1294 rvt_stop_rc_timers(qp);
1295 rvt_add_rnr_timer(qp, aeth);
1299 if (qp->s_acked == qp->s_tail)
1301 /* The last valid PSN is the previous PSN. */
1302 update_last_psn(qp, psn - 1);
1303 switch ((aeth >> IB_AETH_CREDIT_SHIFT) &
1304 IB_AETH_CREDIT_MASK) {
1305 case 0: /* PSN sequence error */
1306 ibp->rvp.n_seq_naks++;
1308 * Back up to the responder's expected PSN.
1309 * Note that we might get a NAK in the middle of an
1310 * RDMA READ response which terminates the RDMA
1313 hfi1_restart_rc(qp, psn, 0);
1314 hfi1_schedule_send(qp);
1317 case 1: /* Invalid Request */
1318 status = IB_WC_REM_INV_REQ_ERR;
1319 ibp->rvp.n_other_naks++;
1322 case 2: /* Remote Access Error */
1323 status = IB_WC_REM_ACCESS_ERR;
1324 ibp->rvp.n_other_naks++;
1327 case 3: /* Remote Operation Error */
1328 status = IB_WC_REM_OP_ERR;
1329 ibp->rvp.n_other_naks++;
1331 if (qp->s_last == qp->s_acked) {
1332 hfi1_send_complete(qp, wqe, status);
1333 rvt_error_qp(qp, IB_WC_WR_FLUSH_ERR);
1338 /* Ignore other reserved NAK error codes */
1341 qp->s_retry = qp->s_retry_cnt;
1342 qp->s_rnr_retry = qp->s_rnr_retry_cnt;
1345 default: /* 2: reserved */
1347 /* Ignore reserved NAK codes. */
1350 /* cannot be reached */
1352 rvt_stop_rc_timers(qp);
1357 * We have seen an out of sequence RDMA read middle or last packet.
1358 * This ACKs SENDs and RDMA writes up to the first RDMA read or atomic SWQE.
1360 static void rdma_seq_err(struct rvt_qp *qp, struct hfi1_ibport *ibp, u32 psn,
1361 struct hfi1_ctxtdata *rcd)
1363 struct rvt_swqe *wqe;
1365 lockdep_assert_held(&qp->s_lock);
1366 /* Remove QP from retry timer */
1367 rvt_stop_rc_timers(qp);
1369 wqe = rvt_get_swqe_ptr(qp, qp->s_acked);
1371 while (cmp_psn(psn, wqe->lpsn) > 0) {
1372 if (wqe->wr.opcode == IB_WR_RDMA_READ ||
1373 wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
1374 wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD)
1376 wqe = do_rc_completion(qp, wqe, ibp);
1379 ibp->rvp.n_rdma_seq++;
1380 qp->r_flags |= RVT_R_RDMAR_SEQ;
1381 hfi1_restart_rc(qp, qp->s_last_psn + 1, 0);
1382 if (list_empty(&qp->rspwait)) {
1383 qp->r_flags |= RVT_R_RSP_SEND;
1385 list_add_tail(&qp->rspwait, &rcd->qp_wait_list);
1390 * rc_rcv_resp - process an incoming RC response packet
1391 * @ibp: the port this packet came in on
1392 * @ohdr: the other headers for this packet
1393 * @data: the packet data
1394 * @tlen: the packet length
1395 * @qp: the QP for this packet
1396 * @opcode: the opcode for this packet
1397 * @psn: the packet sequence number for this packet
1398 * @hdrsize: the header length
1399 * @pmtu: the path MTU
1401 * This is called from hfi1_rc_rcv() to process an incoming RC response
1402 * packet for the given QP.
1403 * Called at interrupt level.
1405 static void rc_rcv_resp(struct hfi1_ibport *ibp,
1406 struct ib_other_headers *ohdr,
1407 void *data, u32 tlen, struct rvt_qp *qp,
1408 u32 opcode, u32 psn, u32 hdrsize, u32 pmtu,
1409 struct hfi1_ctxtdata *rcd)
1411 struct rvt_swqe *wqe;
1412 enum ib_wc_status status;
1413 unsigned long flags;
1419 spin_lock_irqsave(&qp->s_lock, flags);
1421 trace_hfi1_ack(qp, psn);
1423 /* Ignore invalid responses. */
1424 smp_read_barrier_depends(); /* see post_one_send */
1425 if (cmp_psn(psn, READ_ONCE(qp->s_next_psn)) >= 0)
1428 /* Ignore duplicate responses. */
1429 diff = cmp_psn(psn, qp->s_last_psn);
1430 if (unlikely(diff <= 0)) {
1431 /* Update credits for "ghost" ACKs */
1432 if (diff == 0 && opcode == OP(ACKNOWLEDGE)) {
1433 aeth = be32_to_cpu(ohdr->u.aeth);
1434 if ((aeth >> IB_AETH_NAK_SHIFT) == 0)
1435 rvt_get_credit(qp, aeth);
1441 * Skip everything other than the PSN we expect, if we are waiting
1442 * for a reply to a restarted RDMA read or atomic op.
1444 if (qp->r_flags & RVT_R_RDMAR_SEQ) {
1445 if (cmp_psn(psn, qp->s_last_psn + 1) != 0)
1447 qp->r_flags &= ~RVT_R_RDMAR_SEQ;
1450 if (unlikely(qp->s_acked == qp->s_tail))
1452 wqe = rvt_get_swqe_ptr(qp, qp->s_acked);
1453 status = IB_WC_SUCCESS;
1456 case OP(ACKNOWLEDGE):
1457 case OP(ATOMIC_ACKNOWLEDGE):
1458 case OP(RDMA_READ_RESPONSE_FIRST):
1459 aeth = be32_to_cpu(ohdr->u.aeth);
1460 if (opcode == OP(ATOMIC_ACKNOWLEDGE))
1461 val = ib_u64_get(&ohdr->u.at.atomic_ack_eth);
1464 if (!do_rc_ack(qp, aeth, psn, opcode, val, rcd) ||
1465 opcode != OP(RDMA_READ_RESPONSE_FIRST))
1467 wqe = rvt_get_swqe_ptr(qp, qp->s_acked);
1468 if (unlikely(wqe->wr.opcode != IB_WR_RDMA_READ))
1471 * If this is a response to a resent RDMA read, we
1472 * have to be careful to copy the data to the right
1475 qp->s_rdma_read_len = restart_sge(&qp->s_rdma_read_sge,
1479 case OP(RDMA_READ_RESPONSE_MIDDLE):
1480 /* no AETH, no ACK */
1481 if (unlikely(cmp_psn(psn, qp->s_last_psn + 1)))
1483 if (unlikely(wqe->wr.opcode != IB_WR_RDMA_READ))
1486 if (unlikely(tlen != (hdrsize + pmtu + 4)))
1488 if (unlikely(pmtu >= qp->s_rdma_read_len))
1492 * We got a response so update the timeout.
1493 * 4.096 usec. * (1 << qp->timeout)
1495 rvt_mod_retry_timer(qp);
1496 if (qp->s_flags & RVT_S_WAIT_ACK) {
1497 qp->s_flags &= ~RVT_S_WAIT_ACK;
1498 hfi1_schedule_send(qp);
1501 if (opcode == OP(RDMA_READ_RESPONSE_MIDDLE))
1502 qp->s_retry = qp->s_retry_cnt;
1505 * Update the RDMA receive state but do the copy w/o
1506 * holding the locks and blocking interrupts.
1508 qp->s_rdma_read_len -= pmtu;
1509 update_last_psn(qp, psn);
1510 spin_unlock_irqrestore(&qp->s_lock, flags);
1511 hfi1_copy_sge(&qp->s_rdma_read_sge, data, pmtu, false, false);
1514 case OP(RDMA_READ_RESPONSE_ONLY):
1515 aeth = be32_to_cpu(ohdr->u.aeth);
1516 if (!do_rc_ack(qp, aeth, psn, opcode, 0, rcd))
1518 /* Get the number of bytes the message was padded by. */
1519 pad = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3;
1521 * Check that the data size is >= 0 && <= pmtu.
1522 * Remember to account for ICRC (4).
1524 if (unlikely(tlen < (hdrsize + pad + 4)))
1527 * If this is a response to a resent RDMA read, we
1528 * have to be careful to copy the data to the right
1531 wqe = rvt_get_swqe_ptr(qp, qp->s_acked);
1532 qp->s_rdma_read_len = restart_sge(&qp->s_rdma_read_sge,
1536 case OP(RDMA_READ_RESPONSE_LAST):
1537 /* ACKs READ req. */
1538 if (unlikely(cmp_psn(psn, qp->s_last_psn + 1)))
1540 if (unlikely(wqe->wr.opcode != IB_WR_RDMA_READ))
1542 /* Get the number of bytes the message was padded by. */
1543 pad = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3;
1545 * Check that the data size is >= 1 && <= pmtu.
1546 * Remember to account for ICRC (4).
1548 if (unlikely(tlen <= (hdrsize + pad + 4)))
1551 tlen -= hdrsize + pad + 4;
1552 if (unlikely(tlen != qp->s_rdma_read_len))
1554 aeth = be32_to_cpu(ohdr->u.aeth);
1555 hfi1_copy_sge(&qp->s_rdma_read_sge, data, tlen, false, false);
1556 WARN_ON(qp->s_rdma_read_sge.num_sge);
1557 (void)do_rc_ack(qp, aeth, psn,
1558 OP(RDMA_READ_RESPONSE_LAST), 0, rcd);
1563 status = IB_WC_LOC_QP_OP_ERR;
1567 rdma_seq_err(qp, ibp, psn, rcd);
1571 status = IB_WC_LOC_LEN_ERR;
1573 if (qp->s_last == qp->s_acked) {
1574 hfi1_send_complete(qp, wqe, status);
1575 rvt_error_qp(qp, IB_WC_WR_FLUSH_ERR);
1578 spin_unlock_irqrestore(&qp->s_lock, flags);
1583 static inline void rc_defered_ack(struct hfi1_ctxtdata *rcd,
1586 if (list_empty(&qp->rspwait)) {
1587 qp->r_flags |= RVT_R_RSP_NAK;
1589 list_add_tail(&qp->rspwait, &rcd->qp_wait_list);
1593 static inline void rc_cancel_ack(struct rvt_qp *qp)
1595 struct hfi1_qp_priv *priv = qp->priv;
1597 priv->r_adefered = 0;
1598 if (list_empty(&qp->rspwait))
1600 list_del_init(&qp->rspwait);
1601 qp->r_flags &= ~RVT_R_RSP_NAK;
1606 * rc_rcv_error - process an incoming duplicate or error RC packet
1607 * @ohdr: the other headers for this packet
1608 * @data: the packet data
1609 * @qp: the QP for this packet
1610 * @opcode: the opcode for this packet
1611 * @psn: the packet sequence number for this packet
1612 * @diff: the difference between the PSN and the expected PSN
1614 * This is called from hfi1_rc_rcv() to process an unexpected
1615 * incoming RC packet for the given QP.
1616 * Called at interrupt level.
1617 * Return 1 if no more processing is needed; otherwise return 0 to
1618 * schedule a response to be sent.
1620 static noinline int rc_rcv_error(struct ib_other_headers *ohdr, void *data,
1621 struct rvt_qp *qp, u32 opcode, u32 psn,
1622 int diff, struct hfi1_ctxtdata *rcd)
1624 struct hfi1_ibport *ibp = rcd_to_iport(rcd);
1625 struct rvt_ack_entry *e;
1626 unsigned long flags;
1630 trace_hfi1_rcv_error(qp, psn);
1633 * Packet sequence error.
1634 * A NAK will ACK earlier sends and RDMA writes.
1635 * Don't queue the NAK if we already sent one.
1637 if (!qp->r_nak_state) {
1638 ibp->rvp.n_rc_seqnak++;
1639 qp->r_nak_state = IB_NAK_PSN_ERROR;
1640 /* Use the expected PSN. */
1641 qp->r_ack_psn = qp->r_psn;
1643 * Wait to send the sequence NAK until all packets
1644 * in the receive queue have been processed.
1645 * Otherwise, we end up propagating congestion.
1647 rc_defered_ack(rcd, qp);
1653 * Handle a duplicate request. Don't re-execute SEND, RDMA
1654 * write or atomic op. Don't NAK errors, just silently drop
1655 * the duplicate request. Note that r_sge, r_len, and
1656 * r_rcv_len may be in use so don't modify them.
1658 * We are supposed to ACK the earliest duplicate PSN but we
1659 * can coalesce an outstanding duplicate ACK. We have to
1660 * send the earliest so that RDMA reads can be restarted at
1661 * the requester's expected PSN.
1663 * First, find where this duplicate PSN falls within the
1664 * ACKs previously sent.
1665 * old_req is true if there is an older response that is scheduled
1666 * to be sent before sending this one.
1670 ibp->rvp.n_rc_dupreq++;
1672 spin_lock_irqsave(&qp->s_lock, flags);
1674 for (i = qp->r_head_ack_queue; ; i = prev) {
1675 if (i == qp->s_tail_ack_queue)
1680 prev = HFI1_MAX_RDMA_ATOMIC;
1681 if (prev == qp->r_head_ack_queue) {
1685 e = &qp->s_ack_queue[prev];
1690 if (cmp_psn(psn, e->psn) >= 0) {
1691 if (prev == qp->s_tail_ack_queue &&
1692 cmp_psn(psn, e->lpsn) <= 0)
1698 case OP(RDMA_READ_REQUEST): {
1699 struct ib_reth *reth;
1704 * If we didn't find the RDMA read request in the ack queue,
1705 * we can ignore this request.
1707 if (!e || e->opcode != OP(RDMA_READ_REQUEST))
1709 /* RETH comes after BTH */
1710 reth = &ohdr->u.rc.reth;
1712 * Address range must be a subset of the original
1713 * request and start on pmtu boundaries.
1714 * We reuse the old ack_queue slot since the requester
1715 * should not back up and request an earlier PSN for the
1718 offset = delta_psn(psn, e->psn) * qp->pmtu;
1719 len = be32_to_cpu(reth->length);
1720 if (unlikely(offset + len != e->rdma_sge.sge_length))
1722 if (e->rdma_sge.mr) {
1723 rvt_put_mr(e->rdma_sge.mr);
1724 e->rdma_sge.mr = NULL;
1727 u32 rkey = be32_to_cpu(reth->rkey);
1728 u64 vaddr = get_ib_reth_vaddr(reth);
1731 ok = rvt_rkey_ok(qp, &e->rdma_sge, len, vaddr, rkey,
1732 IB_ACCESS_REMOTE_READ);
1736 e->rdma_sge.vaddr = NULL;
1737 e->rdma_sge.length = 0;
1738 e->rdma_sge.sge_length = 0;
1743 qp->s_tail_ack_queue = prev;
1747 case OP(COMPARE_SWAP):
1748 case OP(FETCH_ADD): {
1750 * If we didn't find the atomic request in the ack queue
1751 * or the send engine is already backed up to send an
1752 * earlier entry, we can ignore this request.
1754 if (!e || e->opcode != (u8)opcode || old_req)
1756 qp->s_tail_ack_queue = prev;
1762 * Ignore this operation if it doesn't request an ACK
1763 * or an earlier RDMA read or atomic is going to be resent.
1765 if (!(psn & IB_BTH_REQ_ACK) || old_req)
1768 * Resend the most recent ACK if this request is
1769 * after all the previous RDMA reads and atomics.
1771 if (i == qp->r_head_ack_queue) {
1772 spin_unlock_irqrestore(&qp->s_lock, flags);
1773 qp->r_nak_state = 0;
1774 qp->r_ack_psn = qp->r_psn - 1;
1779 * Resend the RDMA read or atomic op which
1780 * ACKs this duplicate request.
1782 qp->s_tail_ack_queue = i;
1785 qp->s_ack_state = OP(ACKNOWLEDGE);
1786 qp->s_flags |= RVT_S_RESP_PENDING;
1787 qp->r_nak_state = 0;
1788 hfi1_schedule_send(qp);
1791 spin_unlock_irqrestore(&qp->s_lock, flags);
1799 static inline void update_ack_queue(struct rvt_qp *qp, unsigned n)
1804 if (next > HFI1_MAX_RDMA_ATOMIC)
1806 qp->s_tail_ack_queue = next;
1807 qp->s_ack_state = OP(ACKNOWLEDGE);
1810 static void log_cca_event(struct hfi1_pportdata *ppd, u8 sl, u32 rlid,
1811 u32 lqpn, u32 rqpn, u8 svc_type)
1813 struct opa_hfi1_cong_log_event_internal *cc_event;
1814 unsigned long flags;
1816 if (sl >= OPA_MAX_SLS)
1819 spin_lock_irqsave(&ppd->cc_log_lock, flags);
1821 ppd->threshold_cong_event_map[sl / 8] |= 1 << (sl % 8);
1822 ppd->threshold_event_counter++;
1824 cc_event = &ppd->cc_events[ppd->cc_log_idx++];
1825 if (ppd->cc_log_idx == OPA_CONG_LOG_ELEMS)
1826 ppd->cc_log_idx = 0;
1827 cc_event->lqpn = lqpn & RVT_QPN_MASK;
1828 cc_event->rqpn = rqpn & RVT_QPN_MASK;
1830 cc_event->svc_type = svc_type;
1831 cc_event->rlid = rlid;
1832 /* keep timestamp in units of 1.024 usec */
1833 cc_event->timestamp = ktime_to_ns(ktime_get()) / 1024;
1835 spin_unlock_irqrestore(&ppd->cc_log_lock, flags);
1838 void process_becn(struct hfi1_pportdata *ppd, u8 sl, u16 rlid, u32 lqpn,
1839 u32 rqpn, u8 svc_type)
1841 struct cca_timer *cca_timer;
1842 u16 ccti, ccti_incr, ccti_timer, ccti_limit;
1843 u8 trigger_threshold;
1844 struct cc_state *cc_state;
1845 unsigned long flags;
1847 if (sl >= OPA_MAX_SLS)
1850 cc_state = get_cc_state(ppd);
1856 * 1) increase CCTI (for this SL)
1857 * 2) select IPG (i.e., call set_link_ipg())
1860 ccti_limit = cc_state->cct.ccti_limit;
1861 ccti_incr = cc_state->cong_setting.entries[sl].ccti_increase;
1862 ccti_timer = cc_state->cong_setting.entries[sl].ccti_timer;
1864 cc_state->cong_setting.entries[sl].trigger_threshold;
1866 spin_lock_irqsave(&ppd->cca_timer_lock, flags);
1868 cca_timer = &ppd->cca_timer[sl];
1869 if (cca_timer->ccti < ccti_limit) {
1870 if (cca_timer->ccti + ccti_incr <= ccti_limit)
1871 cca_timer->ccti += ccti_incr;
1873 cca_timer->ccti = ccti_limit;
1877 ccti = cca_timer->ccti;
1879 if (!hrtimer_active(&cca_timer->hrtimer)) {
1880 /* ccti_timer is in units of 1.024 usec */
1881 unsigned long nsec = 1024 * ccti_timer;
1883 hrtimer_start(&cca_timer->hrtimer, ns_to_ktime(nsec),
1887 spin_unlock_irqrestore(&ppd->cca_timer_lock, flags);
1889 if ((trigger_threshold != 0) && (ccti >= trigger_threshold))
1890 log_cca_event(ppd, sl, rlid, lqpn, rqpn, svc_type);
1894 * hfi1_rc_rcv - process an incoming RC packet
1895 * @rcd: the context pointer
1896 * @hdr: the header of this packet
1897 * @rcv_flags: flags relevant to rcv processing
1898 * @data: the packet data
1899 * @tlen: the packet length
1900 * @qp: the QP for this packet
1902 * This is called from qp_rcv() to process an incoming RC packet
1904 * May be called at interrupt level.
1906 void hfi1_rc_rcv(struct hfi1_packet *packet)
1908 struct hfi1_ctxtdata *rcd = packet->rcd;
1909 struct ib_header *hdr = packet->hdr;
1910 u32 rcv_flags = packet->rcv_flags;
1911 void *data = packet->ebuf;
1912 u32 tlen = packet->tlen;
1913 struct rvt_qp *qp = packet->qp;
1914 struct hfi1_ibport *ibp = rcd_to_iport(rcd);
1915 struct ib_other_headers *ohdr = packet->ohdr;
1917 u32 hdrsize = packet->hlen;
1921 u32 pmtu = qp->pmtu;
1923 struct ib_reth *reth;
1924 unsigned long flags;
1925 int ret, is_fecn = 0;
1926 bool copy_last = false;
1929 lockdep_assert_held(&qp->r_lock);
1930 bth0 = be32_to_cpu(ohdr->bth[0]);
1931 if (hfi1_ruc_check_hdr(ibp, hdr, rcv_flags & HFI1_HAS_GRH, qp, bth0))
1934 is_fecn = process_ecn(qp, packet, false);
1936 psn = be32_to_cpu(ohdr->bth[2]);
1937 opcode = (bth0 >> 24) & 0xff;
1940 * Process responses (ACKs) before anything else. Note that the
1941 * packet sequence number will be for something in the send work
1942 * queue rather than the expected receive packet sequence number.
1943 * In other words, this QP is the requester.
1945 if (opcode >= OP(RDMA_READ_RESPONSE_FIRST) &&
1946 opcode <= OP(ATOMIC_ACKNOWLEDGE)) {
1947 rc_rcv_resp(ibp, ohdr, data, tlen, qp, opcode, psn,
1948 hdrsize, pmtu, rcd);
1954 /* Compute 24 bits worth of difference. */
1955 diff = delta_psn(psn, qp->r_psn);
1956 if (unlikely(diff)) {
1957 if (rc_rcv_error(ohdr, data, qp, opcode, psn, diff, rcd))
1962 /* Check for opcode sequence errors. */
1963 switch (qp->r_state) {
1964 case OP(SEND_FIRST):
1965 case OP(SEND_MIDDLE):
1966 if (opcode == OP(SEND_MIDDLE) ||
1967 opcode == OP(SEND_LAST) ||
1968 opcode == OP(SEND_LAST_WITH_IMMEDIATE) ||
1969 opcode == OP(SEND_LAST_WITH_INVALIDATE))
1973 case OP(RDMA_WRITE_FIRST):
1974 case OP(RDMA_WRITE_MIDDLE):
1975 if (opcode == OP(RDMA_WRITE_MIDDLE) ||
1976 opcode == OP(RDMA_WRITE_LAST) ||
1977 opcode == OP(RDMA_WRITE_LAST_WITH_IMMEDIATE))
1982 if (opcode == OP(SEND_MIDDLE) ||
1983 opcode == OP(SEND_LAST) ||
1984 opcode == OP(SEND_LAST_WITH_IMMEDIATE) ||
1985 opcode == OP(SEND_LAST_WITH_INVALIDATE) ||
1986 opcode == OP(RDMA_WRITE_MIDDLE) ||
1987 opcode == OP(RDMA_WRITE_LAST) ||
1988 opcode == OP(RDMA_WRITE_LAST_WITH_IMMEDIATE))
1991 * Note that it is up to the requester to not send a new
1992 * RDMA read or atomic operation before receiving an ACK
1993 * for the previous operation.
1998 if (qp->state == IB_QPS_RTR && !(qp->r_flags & RVT_R_COMM_EST))
2001 /* OK, process the packet. */
2003 case OP(SEND_FIRST):
2004 ret = hfi1_rvt_get_rwqe(qp, 0);
2011 case OP(SEND_MIDDLE):
2012 case OP(RDMA_WRITE_MIDDLE):
2014 /* Check for invalid length PMTU or posted rwqe len. */
2015 if (unlikely(tlen != (hdrsize + pmtu + 4)))
2017 qp->r_rcv_len += pmtu;
2018 if (unlikely(qp->r_rcv_len > qp->r_len))
2020 hfi1_copy_sge(&qp->r_sge, data, pmtu, true, false);
2023 case OP(RDMA_WRITE_LAST_WITH_IMMEDIATE):
2025 ret = hfi1_rvt_get_rwqe(qp, 1);
2033 case OP(SEND_ONLY_WITH_IMMEDIATE):
2034 case OP(SEND_ONLY_WITH_INVALIDATE):
2035 ret = hfi1_rvt_get_rwqe(qp, 0);
2041 if (opcode == OP(SEND_ONLY))
2042 goto no_immediate_data;
2043 if (opcode == OP(SEND_ONLY_WITH_INVALIDATE))
2045 /* FALLTHROUGH for SEND_ONLY_WITH_IMMEDIATE */
2046 case OP(SEND_LAST_WITH_IMMEDIATE):
2048 wc.ex.imm_data = ohdr->u.imm_data;
2049 wc.wc_flags = IB_WC_WITH_IMM;
2051 case OP(SEND_LAST_WITH_INVALIDATE):
2053 rkey = be32_to_cpu(ohdr->u.ieth);
2054 if (rvt_invalidate_rkey(qp, rkey))
2055 goto no_immediate_data;
2056 wc.ex.invalidate_rkey = rkey;
2057 wc.wc_flags = IB_WC_WITH_INVALIDATE;
2059 case OP(RDMA_WRITE_LAST):
2060 copy_last = rvt_is_user_qp(qp);
2067 /* Get the number of bytes the message was padded by. */
2068 pad = (bth0 >> 20) & 3;
2069 /* Check for invalid length. */
2070 /* LAST len should be >= 1 */
2071 if (unlikely(tlen < (hdrsize + pad + 4)))
2073 /* Don't count the CRC. */
2074 tlen -= (hdrsize + pad + 4);
2075 wc.byte_len = tlen + qp->r_rcv_len;
2076 if (unlikely(wc.byte_len > qp->r_len))
2078 hfi1_copy_sge(&qp->r_sge, data, tlen, true, copy_last);
2079 rvt_put_ss(&qp->r_sge);
2081 if (!__test_and_clear_bit(RVT_R_WRID_VALID, &qp->r_aflags))
2083 wc.wr_id = qp->r_wr_id;
2084 wc.status = IB_WC_SUCCESS;
2085 if (opcode == OP(RDMA_WRITE_LAST_WITH_IMMEDIATE) ||
2086 opcode == OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE))
2087 wc.opcode = IB_WC_RECV_RDMA_WITH_IMM;
2089 wc.opcode = IB_WC_RECV;
2091 wc.src_qp = qp->remote_qpn;
2092 wc.slid = qp->remote_ah_attr.dlid;
2094 * It seems that IB mandates the presence of an SL in a
2095 * work completion only for the UD transport (see section
2096 * 11.4.2 of IBTA Vol. 1).
2098 * However, the way the SL is chosen below is consistent
2099 * with the way that IB/qib works and is trying avoid
2100 * introducing incompatibilities.
2102 * See also OPA Vol. 1, section 9.7.6, and table 9-17.
2104 wc.sl = qp->remote_ah_attr.sl;
2105 /* zero fields that are N/A */
2108 wc.dlid_path_bits = 0;
2110 /* Signal completion event if the solicited bit is set. */
2111 rvt_cq_enter(ibcq_to_rvtcq(qp->ibqp.recv_cq), &wc,
2112 (bth0 & IB_BTH_SOLICITED) != 0);
2115 case OP(RDMA_WRITE_ONLY):
2116 copy_last = rvt_is_user_qp(qp);
2118 case OP(RDMA_WRITE_FIRST):
2119 case OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE):
2120 if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_WRITE)))
2123 reth = &ohdr->u.rc.reth;
2124 qp->r_len = be32_to_cpu(reth->length);
2126 qp->r_sge.sg_list = NULL;
2127 if (qp->r_len != 0) {
2128 u32 rkey = be32_to_cpu(reth->rkey);
2129 u64 vaddr = get_ib_reth_vaddr(reth);
2132 /* Check rkey & NAK */
2133 ok = rvt_rkey_ok(qp, &qp->r_sge.sge, qp->r_len, vaddr,
2134 rkey, IB_ACCESS_REMOTE_WRITE);
2137 qp->r_sge.num_sge = 1;
2139 qp->r_sge.num_sge = 0;
2140 qp->r_sge.sge.mr = NULL;
2141 qp->r_sge.sge.vaddr = NULL;
2142 qp->r_sge.sge.length = 0;
2143 qp->r_sge.sge.sge_length = 0;
2145 if (opcode == OP(RDMA_WRITE_FIRST))
2147 else if (opcode == OP(RDMA_WRITE_ONLY))
2148 goto no_immediate_data;
2149 ret = hfi1_rvt_get_rwqe(qp, 1);
2154 wc.ex.imm_data = ohdr->u.rc.imm_data;
2155 wc.wc_flags = IB_WC_WITH_IMM;
2158 case OP(RDMA_READ_REQUEST): {
2159 struct rvt_ack_entry *e;
2163 if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_READ)))
2165 next = qp->r_head_ack_queue + 1;
2166 /* s_ack_queue is size HFI1_MAX_RDMA_ATOMIC+1 so use > not >= */
2167 if (next > HFI1_MAX_RDMA_ATOMIC)
2169 spin_lock_irqsave(&qp->s_lock, flags);
2170 if (unlikely(next == qp->s_tail_ack_queue)) {
2171 if (!qp->s_ack_queue[next].sent)
2172 goto nack_inv_unlck;
2173 update_ack_queue(qp, next);
2175 e = &qp->s_ack_queue[qp->r_head_ack_queue];
2176 if (e->opcode == OP(RDMA_READ_REQUEST) && e->rdma_sge.mr) {
2177 rvt_put_mr(e->rdma_sge.mr);
2178 e->rdma_sge.mr = NULL;
2180 reth = &ohdr->u.rc.reth;
2181 len = be32_to_cpu(reth->length);
2183 u32 rkey = be32_to_cpu(reth->rkey);
2184 u64 vaddr = get_ib_reth_vaddr(reth);
2187 /* Check rkey & NAK */
2188 ok = rvt_rkey_ok(qp, &e->rdma_sge, len, vaddr,
2189 rkey, IB_ACCESS_REMOTE_READ);
2191 goto nack_acc_unlck;
2193 * Update the next expected PSN. We add 1 later
2194 * below, so only add the remainder here.
2196 qp->r_psn += rvt_div_mtu(qp, len - 1);
2198 e->rdma_sge.mr = NULL;
2199 e->rdma_sge.vaddr = NULL;
2200 e->rdma_sge.length = 0;
2201 e->rdma_sge.sge_length = 0;
2206 e->lpsn = qp->r_psn;
2208 * We need to increment the MSN here instead of when we
2209 * finish sending the result since a duplicate request would
2210 * increment it more than once.
2214 qp->r_state = opcode;
2215 qp->r_nak_state = 0;
2216 qp->r_head_ack_queue = next;
2218 /* Schedule the send engine. */
2219 qp->s_flags |= RVT_S_RESP_PENDING;
2220 hfi1_schedule_send(qp);
2222 spin_unlock_irqrestore(&qp->s_lock, flags);
2228 case OP(COMPARE_SWAP):
2229 case OP(FETCH_ADD): {
2230 struct ib_atomic_eth *ateth;
2231 struct rvt_ack_entry *e;
2238 if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_ATOMIC)))
2240 next = qp->r_head_ack_queue + 1;
2241 if (next > HFI1_MAX_RDMA_ATOMIC)
2243 spin_lock_irqsave(&qp->s_lock, flags);
2244 if (unlikely(next == qp->s_tail_ack_queue)) {
2245 if (!qp->s_ack_queue[next].sent)
2246 goto nack_inv_unlck;
2247 update_ack_queue(qp, next);
2249 e = &qp->s_ack_queue[qp->r_head_ack_queue];
2250 if (e->opcode == OP(RDMA_READ_REQUEST) && e->rdma_sge.mr) {
2251 rvt_put_mr(e->rdma_sge.mr);
2252 e->rdma_sge.mr = NULL;
2254 ateth = &ohdr->u.atomic_eth;
2255 vaddr = get_ib_ateth_vaddr(ateth);
2256 if (unlikely(vaddr & (sizeof(u64) - 1)))
2257 goto nack_inv_unlck;
2258 rkey = be32_to_cpu(ateth->rkey);
2259 /* Check rkey & NAK */
2260 if (unlikely(!rvt_rkey_ok(qp, &qp->r_sge.sge, sizeof(u64),
2262 IB_ACCESS_REMOTE_ATOMIC)))
2263 goto nack_acc_unlck;
2264 /* Perform atomic OP and save result. */
2265 maddr = (atomic64_t *)qp->r_sge.sge.vaddr;
2266 sdata = get_ib_ateth_swap(ateth);
2267 e->atomic_data = (opcode == OP(FETCH_ADD)) ?
2268 (u64)atomic64_add_return(sdata, maddr) - sdata :
2269 (u64)cmpxchg((u64 *)qp->r_sge.sge.vaddr,
2270 get_ib_ateth_compare(ateth),
2272 rvt_put_mr(qp->r_sge.sge.mr);
2273 qp->r_sge.num_sge = 0;
2280 qp->r_state = opcode;
2281 qp->r_nak_state = 0;
2282 qp->r_head_ack_queue = next;
2284 /* Schedule the send engine. */
2285 qp->s_flags |= RVT_S_RESP_PENDING;
2286 hfi1_schedule_send(qp);
2288 spin_unlock_irqrestore(&qp->s_lock, flags);
2295 /* NAK unknown opcodes. */
2299 qp->r_state = opcode;
2300 qp->r_ack_psn = psn;
2301 qp->r_nak_state = 0;
2302 /* Send an ACK if requested or required. */
2303 if (psn & IB_BTH_REQ_ACK) {
2304 struct hfi1_qp_priv *priv = qp->priv;
2306 if (packet->numpkt == 0) {
2310 if (priv->r_adefered >= HFI1_PSN_CREDIT) {
2314 if (unlikely(is_fecn)) {
2319 rc_defered_ack(rcd, qp);
2324 qp->r_nak_state = qp->r_min_rnr_timer | IB_RNR_NAK;
2325 qp->r_ack_psn = qp->r_psn;
2326 /* Queue RNR NAK for later */
2327 rc_defered_ack(rcd, qp);
2331 rvt_rc_error(qp, IB_WC_LOC_QP_OP_ERR);
2332 qp->r_nak_state = IB_NAK_REMOTE_OPERATIONAL_ERROR;
2333 qp->r_ack_psn = qp->r_psn;
2334 /* Queue NAK for later */
2335 rc_defered_ack(rcd, qp);
2339 spin_unlock_irqrestore(&qp->s_lock, flags);
2341 rvt_rc_error(qp, IB_WC_LOC_QP_OP_ERR);
2342 qp->r_nak_state = IB_NAK_INVALID_REQUEST;
2343 qp->r_ack_psn = qp->r_psn;
2344 /* Queue NAK for later */
2345 rc_defered_ack(rcd, qp);
2349 spin_unlock_irqrestore(&qp->s_lock, flags);
2351 rvt_rc_error(qp, IB_WC_LOC_PROT_ERR);
2352 qp->r_nak_state = IB_NAK_REMOTE_ACCESS_ERROR;
2353 qp->r_ack_psn = qp->r_psn;
2355 hfi1_send_rc_ack(rcd, qp, is_fecn);
2358 void hfi1_rc_hdrerr(
2359 struct hfi1_ctxtdata *rcd,
2360 struct ib_header *hdr,
2364 int has_grh = rcv_flags & HFI1_HAS_GRH;
2365 struct ib_other_headers *ohdr;
2366 struct hfi1_ibport *ibp = rcd_to_iport(rcd);
2374 ohdr = &hdr->u.l.oth;
2376 bth0 = be32_to_cpu(ohdr->bth[0]);
2377 if (hfi1_ruc_check_hdr(ibp, hdr, has_grh, qp, bth0))
2380 psn = be32_to_cpu(ohdr->bth[2]);
2381 opcode = (bth0 >> 24) & 0xff;
2383 /* Only deal with RDMA Writes for now */
2384 if (opcode < IB_OPCODE_RC_RDMA_READ_RESPONSE_FIRST) {
2385 diff = delta_psn(psn, qp->r_psn);
2386 if (!qp->r_nak_state && diff >= 0) {
2387 ibp->rvp.n_rc_seqnak++;
2388 qp->r_nak_state = IB_NAK_PSN_ERROR;
2389 /* Use the expected PSN. */
2390 qp->r_ack_psn = qp->r_psn;
2392 * Wait to send the sequence
2393 * NAK until all packets
2394 * in the receive queue have
2396 * Otherwise, we end up
2397 * propagating congestion.
2399 rc_defered_ack(rcd, qp);
2400 } /* Out of sequence NAK */
2401 } /* QP Request NAKs */