2 * Copyright(c) 2015 - 2018 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>
55 #include "verbs_txreq.h"
58 struct rvt_ack_entry *find_prev_entry(struct rvt_qp *qp, u32 psn, u8 *prev,
59 u8 *prev_ack, bool *scheduled)
60 __must_hold(&qp->s_lock)
62 struct rvt_ack_entry *e = NULL;
66 for (i = qp->r_head_ack_queue; ; i = p) {
67 if (i == qp->s_tail_ack_queue)
72 p = rvt_size_atomic(ib_to_rvt(qp->ibqp.device));
73 if (p == qp->r_head_ack_queue) {
77 e = &qp->s_ack_queue[p];
82 if (cmp_psn(psn, e->psn) >= 0) {
83 if (p == qp->s_tail_ack_queue &&
84 cmp_psn(psn, e->lpsn) <= 0)
99 * make_rc_ack - construct a response packet (ACK, NAK, or RDMA read)
100 * @dev: the device for this QP
101 * @qp: a pointer to the QP
102 * @ohdr: a pointer to the IB header being constructed
103 * @ps: the xmit packet state
105 * Return 1 if constructed; otherwise, return 0.
106 * Note that we are in the responder's side of the QP context.
107 * Note the QP s_lock must be held.
109 static int make_rc_ack(struct hfi1_ibdev *dev, struct rvt_qp *qp,
110 struct ib_other_headers *ohdr,
111 struct hfi1_pkt_state *ps)
113 struct rvt_ack_entry *e;
116 u32 bth0 = 0, bth2 = 0;
117 u32 bth1 = qp->remote_qpn | (HFI1_CAP_IS_KSET(OPFN) << IB_BTHE_E_SHIFT);
120 struct hfi1_qp_priv *qpriv = qp->priv;
123 u8 next = qp->s_tail_ack_queue;
124 struct tid_rdma_request *req;
126 trace_hfi1_rsp_make_rc_ack(qp, 0);
127 lockdep_assert_held(&qp->s_lock);
128 /* Don't send an ACK if we aren't supposed to. */
129 if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK))
132 if (qpriv->hdr_type == HFI1_PKT_TYPE_9B)
133 /* header size in 32-bit words LRH+BTH = (8+12)/4. */
136 /* header size in 32-bit words 16B LRH+BTH = (16+12)/4. */
139 switch (qp->s_ack_state) {
140 case OP(RDMA_READ_RESPONSE_LAST):
141 case OP(RDMA_READ_RESPONSE_ONLY):
142 e = &qp->s_ack_queue[qp->s_tail_ack_queue];
143 release_rdma_sge_mr(e);
145 case OP(ATOMIC_ACKNOWLEDGE):
147 * We can increment the tail pointer now that the last
148 * response has been sent instead of only being
151 if (++next > rvt_size_atomic(&dev->rdi))
154 * Only advance the s_acked_ack_queue pointer if there
155 * have been no TID RDMA requests.
157 e = &qp->s_ack_queue[qp->s_tail_ack_queue];
158 if (e->opcode != TID_OP(WRITE_REQ) &&
159 qp->s_acked_ack_queue == qp->s_tail_ack_queue)
160 qp->s_acked_ack_queue = next;
161 qp->s_tail_ack_queue = next;
162 trace_hfi1_rsp_make_rc_ack(qp, e->psn);
165 case OP(ACKNOWLEDGE):
166 /* Check for no next entry in the queue. */
167 if (qp->r_head_ack_queue == qp->s_tail_ack_queue) {
168 if (qp->s_flags & RVT_S_ACK_PENDING)
173 e = &qp->s_ack_queue[qp->s_tail_ack_queue];
174 /* Check for tid write fence */
175 if ((qpriv->s_flags & HFI1_R_TID_WAIT_INTERLCK) ||
176 hfi1_tid_rdma_ack_interlock(qp, e)) {
177 iowait_set_flag(&qpriv->s_iowait, IOWAIT_PENDING_IB);
180 if (e->opcode == OP(RDMA_READ_REQUEST)) {
182 * If a RDMA read response is being resent and
183 * we haven't seen the duplicate request yet,
184 * then stop sending the remaining responses the
185 * responder has seen until the requester re-sends it.
187 len = e->rdma_sge.sge_length;
188 if (len && !e->rdma_sge.mr) {
189 if (qp->s_acked_ack_queue ==
190 qp->s_tail_ack_queue)
191 qp->s_acked_ack_queue =
192 qp->r_head_ack_queue;
193 qp->s_tail_ack_queue = qp->r_head_ack_queue;
196 /* Copy SGE state in case we need to resend */
197 ps->s_txreq->mr = e->rdma_sge.mr;
199 rvt_get_mr(ps->s_txreq->mr);
200 qp->s_ack_rdma_sge.sge = e->rdma_sge;
201 qp->s_ack_rdma_sge.num_sge = 1;
202 ps->s_txreq->ss = &qp->s_ack_rdma_sge;
205 qp->s_ack_state = OP(RDMA_READ_RESPONSE_FIRST);
207 qp->s_ack_state = OP(RDMA_READ_RESPONSE_ONLY);
210 ohdr->u.aeth = rvt_compute_aeth(qp);
212 qp->s_ack_rdma_psn = e->psn;
213 bth2 = mask_psn(qp->s_ack_rdma_psn++);
214 } else if (e->opcode == TID_OP(WRITE_REQ)) {
216 * If a TID RDMA WRITE RESP is being resent, we have to
217 * wait for the actual request. All requests that are to
218 * be resent will have their state set to
219 * TID_REQUEST_RESEND. When the new request arrives, the
220 * state will be changed to TID_REQUEST_RESEND_ACTIVE.
222 req = ack_to_tid_req(e);
223 if (req->state == TID_REQUEST_RESEND ||
224 req->state == TID_REQUEST_INIT_RESEND)
226 qp->s_ack_state = TID_OP(WRITE_RESP);
227 qp->s_ack_rdma_psn = mask_psn(e->psn + req->cur_seg);
229 } else if (e->opcode == TID_OP(READ_REQ)) {
231 * If a TID RDMA read response is being resent and
232 * we haven't seen the duplicate request yet,
233 * then stop sending the remaining responses the
234 * responder has seen until the requester re-sends it.
236 len = e->rdma_sge.sge_length;
237 if (len && !e->rdma_sge.mr) {
238 if (qp->s_acked_ack_queue ==
239 qp->s_tail_ack_queue)
240 qp->s_acked_ack_queue =
241 qp->r_head_ack_queue;
242 qp->s_tail_ack_queue = qp->r_head_ack_queue;
245 /* Copy SGE state in case we need to resend */
246 ps->s_txreq->mr = e->rdma_sge.mr;
248 rvt_get_mr(ps->s_txreq->mr);
249 qp->s_ack_rdma_sge.sge = e->rdma_sge;
250 qp->s_ack_rdma_sge.num_sge = 1;
251 qp->s_ack_state = TID_OP(READ_RESP);
254 /* COMPARE_SWAP or FETCH_ADD */
255 ps->s_txreq->ss = NULL;
257 qp->s_ack_state = OP(ATOMIC_ACKNOWLEDGE);
258 ohdr->u.at.aeth = rvt_compute_aeth(qp);
259 ib_u64_put(e->atomic_data, &ohdr->u.at.atomic_ack_eth);
260 hwords += sizeof(ohdr->u.at) / sizeof(u32);
261 bth2 = mask_psn(e->psn);
264 trace_hfi1_tid_write_rsp_make_rc_ack(qp);
265 bth0 = qp->s_ack_state << 24;
268 case OP(RDMA_READ_RESPONSE_FIRST):
269 qp->s_ack_state = OP(RDMA_READ_RESPONSE_MIDDLE);
271 case OP(RDMA_READ_RESPONSE_MIDDLE):
272 ps->s_txreq->ss = &qp->s_ack_rdma_sge;
273 ps->s_txreq->mr = qp->s_ack_rdma_sge.sge.mr;
275 rvt_get_mr(ps->s_txreq->mr);
276 len = qp->s_ack_rdma_sge.sge.sge_length;
279 middle = HFI1_CAP_IS_KSET(SDMA_AHG);
281 ohdr->u.aeth = rvt_compute_aeth(qp);
283 qp->s_ack_state = OP(RDMA_READ_RESPONSE_LAST);
284 e = &qp->s_ack_queue[qp->s_tail_ack_queue];
287 bth0 = qp->s_ack_state << 24;
288 bth2 = mask_psn(qp->s_ack_rdma_psn++);
291 case TID_OP(WRITE_RESP):
294 * 1. Check if RVT_S_ACK_PENDING is set. If yes,
296 * 2. Attempt to allocate TID resources.
297 * 3. Remove RVT_S_RESP_PENDING flags from s_flags
298 * 4. If resources not available:
299 * 4.1 Set RVT_S_WAIT_TID_SPACE
300 * 4.2 Queue QP on RCD TID queue
301 * 4.3 Put QP on iowait list.
302 * 4.4 Build IB RNR NAK with appropriate timeout value
303 * 4.5 Return indication progress made.
304 * 5. If resources are available:
305 * 5.1 Program HW flow CSRs
306 * 5.2 Build TID RDMA WRITE RESP packet
307 * 5.3 If more resources needed, do 2.1 - 2.3.
308 * 5.4 Wake up next QP on RCD TID queue.
309 * 5.5 Return indication progress made.
312 e = &qp->s_ack_queue[qp->s_tail_ack_queue];
313 req = ack_to_tid_req(e);
316 * Send scheduled RNR NAK's. RNR NAK's need to be sent at
317 * segment boundaries, not at request boundaries. Don't change
318 * s_ack_state because we are still in the middle of a request
320 if (qpriv->rnr_nak_state == TID_RNR_NAK_SEND &&
321 qp->s_tail_ack_queue == qpriv->r_tid_alloc &&
322 req->cur_seg == req->alloc_seg) {
323 qpriv->rnr_nak_state = TID_RNR_NAK_SENT;
324 goto normal_no_state;
327 bth2 = mask_psn(qp->s_ack_rdma_psn);
328 hdrlen = hfi1_build_tid_rdma_write_resp(qp, e, ohdr, &bth1,
335 bth0 = qp->s_ack_state << 24;
336 qp->s_ack_rdma_psn++;
337 trace_hfi1_tid_req_make_rc_ack_write(qp, 0, e->opcode, e->psn,
339 if (req->cur_seg != req->total_segs)
343 /* Do not free e->rdma_sge until all data are received */
344 qp->s_ack_state = OP(ATOMIC_ACKNOWLEDGE);
347 case TID_OP(READ_RESP):
349 e = &qp->s_ack_queue[qp->s_tail_ack_queue];
350 ps->s_txreq->ss = &qp->s_ack_rdma_sge;
351 delta = hfi1_build_tid_rdma_read_resp(qp, e, ohdr, &bth0,
360 * Increment qp->s_tail_ack_queue through s_ack_state
363 qp->s_ack_state = OP(RDMA_READ_RESPONSE_LAST);
366 case TID_OP(READ_REQ):
372 * Send a regular ACK.
373 * Set the s_ack_state so we wait until after sending
374 * the ACK before setting s_ack_state to ACKNOWLEDGE
377 qp->s_ack_state = OP(SEND_ONLY);
381 cpu_to_be32((qp->r_msn & IB_MSN_MASK) |
383 IB_AETH_CREDIT_SHIFT));
385 ohdr->u.aeth = rvt_compute_aeth(qp);
388 bth0 = OP(ACKNOWLEDGE) << 24;
389 bth2 = mask_psn(qp->s_ack_psn);
390 qp->s_flags &= ~RVT_S_ACK_PENDING;
391 ps->s_txreq->txreq.flags |= SDMA_TXREQ_F_VIP;
392 ps->s_txreq->ss = NULL;
394 qp->s_rdma_ack_cnt++;
395 ps->s_txreq->sde = qpriv->s_sde;
396 ps->s_txreq->s_cur_size = len;
397 ps->s_txreq->hdr_dwords = hwords;
398 hfi1_make_ruc_header(qp, ohdr, bth0, bth1, bth2, middle, ps);
401 spin_unlock_irqrestore(&qp->s_lock, ps->flags);
402 spin_lock_irqsave(&qp->r_lock, ps->flags);
403 spin_lock(&qp->s_lock);
404 rvt_error_qp(qp, IB_WC_WR_FLUSH_ERR);
405 spin_unlock(&qp->s_lock);
406 spin_unlock_irqrestore(&qp->r_lock, ps->flags);
407 spin_lock_irqsave(&qp->s_lock, ps->flags);
409 qp->s_ack_state = OP(ACKNOWLEDGE);
411 * Ensure s_rdma_ack_cnt changes are committed prior to resetting
415 qp->s_flags &= ~(RVT_S_RESP_PENDING
422 * hfi1_make_rc_req - construct a request packet (SEND, RDMA r/w, ATOMIC)
423 * @qp: a pointer to the QP
425 * Assumes s_lock is held.
427 * Return 1 if constructed; otherwise, return 0.
429 int hfi1_make_rc_req(struct rvt_qp *qp, struct hfi1_pkt_state *ps)
431 struct hfi1_qp_priv *priv = qp->priv;
432 struct hfi1_ibdev *dev = to_idev(qp->ibqp.device);
433 struct ib_other_headers *ohdr;
434 struct rvt_sge_state *ss = NULL;
435 struct rvt_swqe *wqe;
436 struct hfi1_swqe_priv *wpriv;
437 struct tid_rdma_request *req = NULL;
438 /* header size in 32-bit words LRH+BTH = (8+12)/4. */
441 u32 bth0 = 0, bth2 = 0;
442 u32 bth1 = qp->remote_qpn | (HFI1_CAP_IS_KSET(OPFN) << IB_BTHE_E_SHIFT);
447 struct tid_rdma_flow *flow = NULL;
448 struct tid_rdma_params *remote;
450 trace_hfi1_sender_make_rc_req(qp);
451 lockdep_assert_held(&qp->s_lock);
452 ps->s_txreq = get_txreq(ps->dev, qp);
456 if (priv->hdr_type == HFI1_PKT_TYPE_9B) {
457 /* header size in 32-bit words LRH+BTH = (8+12)/4. */
459 if (rdma_ah_get_ah_flags(&qp->remote_ah_attr) & IB_AH_GRH)
460 ohdr = &ps->s_txreq->phdr.hdr.ibh.u.l.oth;
462 ohdr = &ps->s_txreq->phdr.hdr.ibh.u.oth;
464 /* header size in 32-bit words 16B LRH+BTH = (16+12)/4. */
466 if ((rdma_ah_get_ah_flags(&qp->remote_ah_attr) & IB_AH_GRH) &&
467 (hfi1_check_mcast(rdma_ah_get_dlid(&qp->remote_ah_attr))))
468 ohdr = &ps->s_txreq->phdr.hdr.opah.u.l.oth;
470 ohdr = &ps->s_txreq->phdr.hdr.opah.u.oth;
473 /* Sending responses has higher priority over sending requests. */
474 if ((qp->s_flags & RVT_S_RESP_PENDING) &&
475 make_rc_ack(dev, qp, ohdr, ps))
478 if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_SEND_OK)) {
479 if (!(ib_rvt_state_ops[qp->state] & RVT_FLUSH_SEND))
481 /* We are in the error state, flush the work request. */
482 if (qp->s_last == READ_ONCE(qp->s_head))
484 /* If DMAs are in progress, we can't flush immediately. */
485 if (iowait_sdma_pending(&priv->s_iowait)) {
486 qp->s_flags |= RVT_S_WAIT_DMA;
490 wqe = rvt_get_swqe_ptr(qp, qp->s_last);
491 hfi1_trdma_send_complete(qp, wqe, qp->s_last != qp->s_acked ?
492 IB_WC_SUCCESS : IB_WC_WR_FLUSH_ERR);
493 /* will get called again */
497 if (qp->s_flags & (RVT_S_WAIT_RNR | RVT_S_WAIT_ACK | HFI1_S_WAIT_HALT))
500 if (cmp_psn(qp->s_psn, qp->s_sending_hpsn) <= 0) {
501 if (cmp_psn(qp->s_sending_psn, qp->s_sending_hpsn) <= 0) {
502 qp->s_flags |= RVT_S_WAIT_PSN;
505 qp->s_sending_psn = qp->s_psn;
506 qp->s_sending_hpsn = qp->s_psn - 1;
509 /* Send a request. */
510 wqe = rvt_get_swqe_ptr(qp, qp->s_cur);
512 switch (qp->s_state) {
514 if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_NEXT_SEND_OK))
517 * Resend an old request or start a new one.
519 * We keep track of the current SWQE so that
520 * we don't reset the "furthest progress" state
521 * if we need to back up.
524 if (qp->s_cur == qp->s_tail) {
525 /* Check if send work queue is empty. */
526 if (qp->s_tail == READ_ONCE(qp->s_head)) {
531 * If a fence is requested, wait for previous
532 * RDMA read and atomic operations to finish.
533 * However, there is no need to guard against
534 * TID RDMA READ after TID RDMA READ.
536 if ((wqe->wr.send_flags & IB_SEND_FENCE) &&
537 qp->s_num_rd_atomic &&
538 (wqe->wr.opcode != IB_WR_TID_RDMA_READ ||
539 priv->pending_tid_r_segs < qp->s_num_rd_atomic)) {
540 qp->s_flags |= RVT_S_WAIT_FENCE;
544 * Local operations are processed immediately
545 * after all prior requests have completed
547 if (wqe->wr.opcode == IB_WR_REG_MR ||
548 wqe->wr.opcode == IB_WR_LOCAL_INV) {
552 if (qp->s_last != qp->s_cur)
554 if (++qp->s_cur == qp->s_size)
556 if (++qp->s_tail == qp->s_size)
558 if (!(wqe->wr.send_flags &
559 RVT_SEND_COMPLETION_ONLY)) {
560 err = rvt_invalidate_rkey(
562 wqe->wr.ex.invalidate_rkey);
565 rvt_send_complete(qp, wqe,
566 err ? IB_WC_LOC_PROT_ERR
569 atomic_dec(&qp->local_ops_pending);
574 qp->s_psn = wqe->psn;
577 * Note that we have to be careful not to modify the
578 * original work request since we may need to resend
583 bth2 = mask_psn(qp->s_psn);
586 * Interlock between various IB requests and TID RDMA
589 if ((priv->s_flags & HFI1_S_TID_WAIT_INTERLCK) ||
590 hfi1_tid_rdma_wqe_interlock(qp, wqe))
593 switch (wqe->wr.opcode) {
595 case IB_WR_SEND_WITH_IMM:
596 case IB_WR_SEND_WITH_INV:
597 /* If no credit, return. */
598 if (!rvt_rc_credit_avail(qp, wqe))
601 qp->s_state = OP(SEND_FIRST);
605 if (wqe->wr.opcode == IB_WR_SEND) {
606 qp->s_state = OP(SEND_ONLY);
607 } else if (wqe->wr.opcode == IB_WR_SEND_WITH_IMM) {
608 qp->s_state = OP(SEND_ONLY_WITH_IMMEDIATE);
609 /* Immediate data comes after the BTH */
610 ohdr->u.imm_data = wqe->wr.ex.imm_data;
613 qp->s_state = OP(SEND_ONLY_WITH_INVALIDATE);
614 /* Invalidate rkey comes after the BTH */
615 ohdr->u.ieth = cpu_to_be32(
616 wqe->wr.ex.invalidate_rkey);
619 if (wqe->wr.send_flags & IB_SEND_SOLICITED)
620 bth0 |= IB_BTH_SOLICITED;
621 bth2 |= IB_BTH_REQ_ACK;
622 if (++qp->s_cur == qp->s_size)
626 case IB_WR_RDMA_WRITE:
627 if (newreq && !(qp->s_flags & RVT_S_UNLIMITED_CREDIT))
629 goto no_flow_control;
630 case IB_WR_RDMA_WRITE_WITH_IMM:
631 /* If no credit, return. */
632 if (!rvt_rc_credit_avail(qp, wqe))
636 wqe->rdma_wr.remote_addr,
638 ohdr->u.rc.reth.rkey =
639 cpu_to_be32(wqe->rdma_wr.rkey);
640 ohdr->u.rc.reth.length = cpu_to_be32(len);
641 hwords += sizeof(struct ib_reth) / sizeof(u32);
643 qp->s_state = OP(RDMA_WRITE_FIRST);
647 if (wqe->wr.opcode == IB_WR_RDMA_WRITE) {
648 qp->s_state = OP(RDMA_WRITE_ONLY);
651 OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE);
652 /* Immediate data comes after RETH */
653 ohdr->u.rc.imm_data = wqe->wr.ex.imm_data;
655 if (wqe->wr.send_flags & IB_SEND_SOLICITED)
656 bth0 |= IB_BTH_SOLICITED;
658 bth2 |= IB_BTH_REQ_ACK;
659 if (++qp->s_cur == qp->s_size)
663 case IB_WR_TID_RDMA_WRITE:
666 * Limit the number of TID RDMA WRITE requests.
668 if (atomic_read(&priv->n_tid_requests) >=
669 HFI1_TID_RDMA_WRITE_CNT)
672 if (!(qp->s_flags & RVT_S_UNLIMITED_CREDIT))
676 hwords += hfi1_build_tid_rdma_write_req(qp, wqe, ohdr,
680 if (priv->s_tid_cur == HFI1_QP_WQE_INVALID) {
681 priv->s_tid_cur = qp->s_cur;
682 if (priv->s_tid_tail == HFI1_QP_WQE_INVALID) {
683 priv->s_tid_tail = qp->s_cur;
684 priv->s_state = TID_OP(WRITE_RESP);
686 } else if (priv->s_tid_cur == priv->s_tid_head) {
687 struct rvt_swqe *__w;
688 struct tid_rdma_request *__r;
690 __w = rvt_get_swqe_ptr(qp, priv->s_tid_cur);
691 __r = wqe_to_tid_req(__w);
694 * The s_tid_cur pointer is advanced to s_cur if
695 * any of the following conditions about the WQE
696 * to which s_ti_cur currently points to are
698 * 1. The request is not a TID RDMA WRITE
700 * 2. The request is in the INACTIVE or
701 * COMPLETE states (TID RDMA READ requests
702 * stay at INACTIVE and TID RDMA WRITE
703 * transition to COMPLETE when done),
704 * 3. The request is in the ACTIVE or SYNC
705 * state and the number of completed
706 * segments is equal to the total segment
708 * (If ACTIVE, the request is waiting for
709 * ACKs. If SYNC, the request has not
710 * received any responses because it's
711 * waiting on a sync point.)
713 if (__w->wr.opcode != IB_WR_TID_RDMA_WRITE ||
714 __r->state == TID_REQUEST_INACTIVE ||
715 __r->state == TID_REQUEST_COMPLETE ||
716 ((__r->state == TID_REQUEST_ACTIVE ||
717 __r->state == TID_REQUEST_SYNC) &&
718 __r->comp_seg == __r->total_segs)) {
719 if (priv->s_tid_tail ==
722 TID_OP(WRITE_DATA_LAST)) {
723 priv->s_tid_tail = qp->s_cur;
727 priv->s_tid_cur = qp->s_cur;
730 * A corner case: when the last TID RDMA WRITE
731 * request was completed, s_tid_head,
732 * s_tid_cur, and s_tid_tail all point to the
733 * same location. Other requests are posted and
734 * s_cur wraps around to the same location,
735 * where a new TID RDMA WRITE is posted. In
736 * this case, none of the indices need to be
737 * updated. However, the priv->s_state should.
739 if (priv->s_tid_tail == qp->s_cur &&
740 priv->s_state == TID_OP(WRITE_DATA_LAST))
741 priv->s_state = TID_OP(WRITE_RESP);
743 req = wqe_to_tid_req(wqe);
745 priv->s_tid_head = qp->s_cur;
746 priv->pending_tid_w_resp += req->total_segs;
747 atomic_inc(&priv->n_tid_requests);
748 atomic_dec(&priv->n_requests);
750 req->state = TID_REQUEST_RESEND;
751 req->comp_seg = delta_psn(bth2, wqe->psn);
753 * Pull back any segments since we are going
754 * to re-receive them.
756 req->setup_head = req->clear_tail;
757 priv->pending_tid_w_resp +=
758 delta_psn(wqe->lpsn, bth2) + 1;
761 trace_hfi1_tid_write_sender_make_req(qp, newreq);
762 trace_hfi1_tid_req_make_req_write(qp, newreq,
766 if (++qp->s_cur == qp->s_size)
770 case IB_WR_RDMA_READ:
772 * Don't allow more operations to be started
773 * than the QP limits allow.
775 if (qp->s_num_rd_atomic >=
776 qp->s_max_rd_atomic) {
777 qp->s_flags |= RVT_S_WAIT_RDMAR;
780 qp->s_num_rd_atomic++;
781 if (newreq && !(qp->s_flags & RVT_S_UNLIMITED_CREDIT))
784 wqe->rdma_wr.remote_addr,
786 ohdr->u.rc.reth.rkey =
787 cpu_to_be32(wqe->rdma_wr.rkey);
788 ohdr->u.rc.reth.length = cpu_to_be32(len);
789 qp->s_state = OP(RDMA_READ_REQUEST);
790 hwords += sizeof(ohdr->u.rc.reth) / sizeof(u32);
793 bth2 |= IB_BTH_REQ_ACK;
794 if (++qp->s_cur == qp->s_size)
798 case IB_WR_TID_RDMA_READ:
799 trace_hfi1_tid_read_sender_make_req(qp, newreq);
801 req = wqe_to_tid_req(wqe);
802 trace_hfi1_tid_req_make_req_read(qp, newreq,
806 delta = cmp_psn(qp->s_psn, wqe->psn);
809 * Don't allow more operations to be started
810 * than the QP limits allow. We could get here under
811 * three conditions; (1) It's a new request; (2) We are
812 * sending the second or later segment of a request,
813 * but the qp->s_state is set to OP(RDMA_READ_REQUEST)
814 * when the last segment of a previous request is
815 * received just before this; (3) We are re-sending a
818 if (qp->s_num_rd_atomic >= qp->s_max_rd_atomic) {
819 qp->s_flags |= RVT_S_WAIT_RDMAR;
823 struct tid_rdma_flow *flow =
824 &req->flows[req->setup_head];
827 * Set up s_sge as it is needed for TID
828 * allocation. However, if the pages have been
829 * walked and mapped, skip it. An earlier try
830 * has failed to allocate the TID entries.
832 if (!flow->npagesets) {
833 qp->s_sge.sge = wqe->sg_list[0];
834 qp->s_sge.sg_list = wqe->sg_list + 1;
835 qp->s_sge.num_sge = wqe->wr.num_sge;
836 qp->s_sge.total_len = wqe->length;
837 qp->s_len = wqe->length;
839 req->clear_tail = req->setup_head;
840 req->flow_idx = req->setup_head;
841 req->state = TID_REQUEST_ACTIVE;
843 } else if (delta == 0) {
844 /* Re-send a request */
847 req->ack_pending = 0;
848 req->flow_idx = req->clear_tail;
849 req->state = TID_REQUEST_RESEND;
851 req->s_next_psn = qp->s_psn;
852 /* Read one segment at a time */
853 len = min_t(u32, req->seg_len,
854 wqe->length - req->seg_len * req->cur_seg);
855 delta = hfi1_build_tid_rdma_read_req(qp, wqe, ohdr,
859 /* Wait for TID space */
862 if (newreq && !(qp->s_flags & RVT_S_UNLIMITED_CREDIT))
866 /* Check if this is the last segment */
867 if (req->cur_seg >= req->total_segs &&
868 ++qp->s_cur == qp->s_size)
872 case IB_WR_ATOMIC_CMP_AND_SWP:
873 case IB_WR_ATOMIC_FETCH_AND_ADD:
875 * Don't allow more operations to be started
876 * than the QP limits allow.
878 if (qp->s_num_rd_atomic >=
879 qp->s_max_rd_atomic) {
880 qp->s_flags |= RVT_S_WAIT_RDMAR;
883 qp->s_num_rd_atomic++;
887 if (newreq && !(qp->s_flags & RVT_S_UNLIMITED_CREDIT))
889 if (wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
890 wqe->wr.opcode == IB_WR_OPFN) {
891 qp->s_state = OP(COMPARE_SWAP);
892 put_ib_ateth_swap(wqe->atomic_wr.swap,
893 &ohdr->u.atomic_eth);
894 put_ib_ateth_compare(wqe->atomic_wr.compare_add,
895 &ohdr->u.atomic_eth);
897 qp->s_state = OP(FETCH_ADD);
898 put_ib_ateth_swap(wqe->atomic_wr.compare_add,
899 &ohdr->u.atomic_eth);
900 put_ib_ateth_compare(0, &ohdr->u.atomic_eth);
902 put_ib_ateth_vaddr(wqe->atomic_wr.remote_addr,
903 &ohdr->u.atomic_eth);
904 ohdr->u.atomic_eth.rkey = cpu_to_be32(
905 wqe->atomic_wr.rkey);
906 hwords += sizeof(struct ib_atomic_eth) / sizeof(u32);
909 bth2 |= IB_BTH_REQ_ACK;
910 if (++qp->s_cur == qp->s_size)
917 if (wqe->wr.opcode != IB_WR_TID_RDMA_READ) {
918 qp->s_sge.sge = wqe->sg_list[0];
919 qp->s_sge.sg_list = wqe->sg_list + 1;
920 qp->s_sge.num_sge = wqe->wr.num_sge;
921 qp->s_sge.total_len = wqe->length;
922 qp->s_len = wqe->length;
926 if (qp->s_tail >= qp->s_size)
929 if (wqe->wr.opcode == IB_WR_RDMA_READ ||
930 wqe->wr.opcode == IB_WR_TID_RDMA_WRITE)
931 qp->s_psn = wqe->lpsn + 1;
932 else if (wqe->wr.opcode == IB_WR_TID_RDMA_READ)
933 qp->s_psn = req->s_next_psn;
938 case OP(RDMA_READ_RESPONSE_FIRST):
940 * qp->s_state is normally set to the opcode of the
941 * last packet constructed for new requests and therefore
942 * is never set to RDMA read response.
943 * RDMA_READ_RESPONSE_FIRST is used by the ACK processing
944 * thread to indicate a SEND needs to be restarted from an
945 * earlier PSN without interfering with the sending thread.
948 qp->s_len = restart_sge(&qp->s_sge, wqe, qp->s_psn, pmtu);
951 qp->s_state = OP(SEND_MIDDLE);
953 case OP(SEND_MIDDLE):
954 bth2 = mask_psn(qp->s_psn++);
959 middle = HFI1_CAP_IS_KSET(SDMA_AHG);
962 if (wqe->wr.opcode == IB_WR_SEND) {
963 qp->s_state = OP(SEND_LAST);
964 } else if (wqe->wr.opcode == IB_WR_SEND_WITH_IMM) {
965 qp->s_state = OP(SEND_LAST_WITH_IMMEDIATE);
966 /* Immediate data comes after the BTH */
967 ohdr->u.imm_data = wqe->wr.ex.imm_data;
970 qp->s_state = OP(SEND_LAST_WITH_INVALIDATE);
971 /* invalidate data comes after the BTH */
972 ohdr->u.ieth = cpu_to_be32(wqe->wr.ex.invalidate_rkey);
975 if (wqe->wr.send_flags & IB_SEND_SOLICITED)
976 bth0 |= IB_BTH_SOLICITED;
977 bth2 |= IB_BTH_REQ_ACK;
979 if (qp->s_cur >= qp->s_size)
983 case OP(RDMA_READ_RESPONSE_LAST):
985 * qp->s_state is normally set to the opcode of the
986 * last packet constructed for new requests and therefore
987 * is never set to RDMA read response.
988 * RDMA_READ_RESPONSE_LAST is used by the ACK processing
989 * thread to indicate a RDMA write needs to be restarted from
990 * an earlier PSN without interfering with the sending thread.
993 qp->s_len = restart_sge(&qp->s_sge, wqe, qp->s_psn, pmtu);
995 case OP(RDMA_WRITE_FIRST):
996 qp->s_state = OP(RDMA_WRITE_MIDDLE);
998 case OP(RDMA_WRITE_MIDDLE):
999 bth2 = mask_psn(qp->s_psn++);
1004 middle = HFI1_CAP_IS_KSET(SDMA_AHG);
1007 if (wqe->wr.opcode == IB_WR_RDMA_WRITE) {
1008 qp->s_state = OP(RDMA_WRITE_LAST);
1010 qp->s_state = OP(RDMA_WRITE_LAST_WITH_IMMEDIATE);
1011 /* Immediate data comes after the BTH */
1012 ohdr->u.imm_data = wqe->wr.ex.imm_data;
1014 if (wqe->wr.send_flags & IB_SEND_SOLICITED)
1015 bth0 |= IB_BTH_SOLICITED;
1017 bth2 |= IB_BTH_REQ_ACK;
1019 if (qp->s_cur >= qp->s_size)
1023 case OP(RDMA_READ_RESPONSE_MIDDLE):
1025 * qp->s_state is normally set to the opcode of the
1026 * last packet constructed for new requests and therefore
1027 * is never set to RDMA read response.
1028 * RDMA_READ_RESPONSE_MIDDLE is used by the ACK processing
1029 * thread to indicate a RDMA read needs to be restarted from
1030 * an earlier PSN without interfering with the sending thread.
1033 len = (delta_psn(qp->s_psn, wqe->psn)) * pmtu;
1035 wqe->rdma_wr.remote_addr + len,
1037 ohdr->u.rc.reth.rkey =
1038 cpu_to_be32(wqe->rdma_wr.rkey);
1039 ohdr->u.rc.reth.length = cpu_to_be32(wqe->length - len);
1040 qp->s_state = OP(RDMA_READ_REQUEST);
1041 hwords += sizeof(ohdr->u.rc.reth) / sizeof(u32);
1042 bth2 = mask_psn(qp->s_psn) | IB_BTH_REQ_ACK;
1043 qp->s_psn = wqe->lpsn + 1;
1047 if (qp->s_cur == qp->s_size)
1051 case TID_OP(WRITE_RESP):
1053 * This value for s_state is used for restarting a TID RDMA
1054 * WRITE request. See comment in OP(RDMA_READ_RESPONSE_MIDDLE
1057 req = wqe_to_tid_req(wqe);
1058 req->state = TID_REQUEST_RESEND;
1060 remote = rcu_dereference(priv->tid_rdma.remote);
1061 req->comp_seg = delta_psn(qp->s_psn, wqe->psn);
1062 len = wqe->length - (req->comp_seg * remote->max_len);
1065 bth2 = mask_psn(qp->s_psn);
1066 hwords += hfi1_build_tid_rdma_write_req(qp, wqe, ohdr, &bth1,
1068 qp->s_psn = wqe->lpsn + 1;
1070 qp->s_state = TID_OP(WRITE_REQ);
1071 priv->pending_tid_w_resp += delta_psn(wqe->lpsn, bth2) + 1;
1072 priv->s_tid_cur = qp->s_cur;
1073 if (++qp->s_cur == qp->s_size)
1075 trace_hfi1_tid_req_make_req_write(qp, 0, wqe->wr.opcode,
1076 wqe->psn, wqe->lpsn, req);
1079 case TID_OP(READ_RESP):
1080 if (wqe->wr.opcode != IB_WR_TID_RDMA_READ)
1082 /* This is used to restart a TID read request */
1083 req = wqe_to_tid_req(wqe);
1086 * Back down. The field qp->s_psn has been set to the psn with
1087 * which the request should be restart. It's OK to use division
1088 * as this is on the retry path.
1090 req->cur_seg = delta_psn(qp->s_psn, wqe->psn) / priv->pkts_ps;
1093 * The following function need to be redefined to return the
1094 * status to make sure that we find the flow. At the same
1095 * time, we can use the req->state change to check if the
1096 * call succeeds or not.
1098 req->state = TID_REQUEST_RESEND;
1099 hfi1_tid_rdma_restart_req(qp, wqe, &bth2);
1100 if (req->state != TID_REQUEST_ACTIVE) {
1102 * Failed to find the flow. Release all allocated tid
1105 hfi1_kern_exp_rcv_clear_all(req);
1106 hfi1_kern_clear_hw_flow(priv->rcd, qp);
1108 hfi1_trdma_send_complete(qp, wqe, IB_WC_LOC_QP_OP_ERR);
1111 req->state = TID_REQUEST_RESEND;
1112 len = min_t(u32, req->seg_len,
1113 wqe->length - req->seg_len * req->cur_seg);
1114 flow = &req->flows[req->flow_idx];
1116 req->s_next_psn = flow->flow_state.ib_lpsn + 1;
1117 delta = hfi1_build_tid_rdma_read_packet(wqe, ohdr, &bth1,
1120 /* Wait for TID space */
1125 /* Check if this is the last segment */
1126 if (req->cur_seg >= req->total_segs &&
1127 ++qp->s_cur == qp->s_size)
1129 qp->s_psn = req->s_next_psn;
1130 trace_hfi1_tid_req_make_req_read(qp, 0, wqe->wr.opcode,
1131 wqe->psn, wqe->lpsn, req);
1133 case TID_OP(READ_REQ):
1134 req = wqe_to_tid_req(wqe);
1135 delta = cmp_psn(qp->s_psn, wqe->psn);
1137 * If the current WR is not TID RDMA READ, or this is the start
1138 * of a new request, we need to change the qp->s_state so that
1139 * the request can be set up properly.
1141 if (wqe->wr.opcode != IB_WR_TID_RDMA_READ || delta == 0 ||
1142 qp->s_cur == qp->s_tail) {
1143 qp->s_state = OP(RDMA_READ_REQUEST);
1144 if (delta == 0 || qp->s_cur == qp->s_tail)
1151 if (qp->s_num_rd_atomic >= qp->s_max_rd_atomic) {
1152 qp->s_flags |= RVT_S_WAIT_RDMAR;
1157 /* Read one segment at a time */
1158 len = min_t(u32, req->seg_len,
1159 wqe->length - req->seg_len * req->cur_seg);
1160 delta = hfi1_build_tid_rdma_read_req(qp, wqe, ohdr, &bth1,
1163 /* Wait for TID space */
1168 /* Check if this is the last segment */
1169 if (req->cur_seg >= req->total_segs &&
1170 ++qp->s_cur == qp->s_size)
1172 qp->s_psn = req->s_next_psn;
1173 trace_hfi1_tid_req_make_req_read(qp, 0, wqe->wr.opcode,
1174 wqe->psn, wqe->lpsn, req);
1177 qp->s_sending_hpsn = bth2;
1178 delta = delta_psn(bth2, wqe->psn);
1179 if (delta && delta % HFI1_PSN_CREDIT == 0 &&
1180 wqe->wr.opcode != IB_WR_TID_RDMA_WRITE)
1181 bth2 |= IB_BTH_REQ_ACK;
1182 if (qp->s_flags & RVT_S_SEND_ONE) {
1183 qp->s_flags &= ~RVT_S_SEND_ONE;
1184 qp->s_flags |= RVT_S_WAIT_ACK;
1185 bth2 |= IB_BTH_REQ_ACK;
1188 ps->s_txreq->hdr_dwords = hwords;
1189 ps->s_txreq->sde = priv->s_sde;
1190 ps->s_txreq->ss = ss;
1191 ps->s_txreq->s_cur_size = len;
1192 hfi1_make_ruc_header(
1195 bth0 | (qp->s_state << 24),
1203 hfi1_put_txreq(ps->s_txreq);
1208 hfi1_put_txreq(ps->s_txreq);
1212 qp->s_flags &= ~RVT_S_BUSY;
1214 * If we didn't get a txreq, the QP will be woken up later to try
1215 * again. Set the flags to indicate which work item to wake
1218 iowait_set_flag(&priv->s_iowait, IOWAIT_PENDING_IB);
1222 static inline void hfi1_make_bth_aeth(struct rvt_qp *qp,
1223 struct ib_other_headers *ohdr,
1226 if (qp->r_nak_state)
1227 ohdr->u.aeth = cpu_to_be32((qp->r_msn & IB_MSN_MASK) |
1229 IB_AETH_CREDIT_SHIFT));
1231 ohdr->u.aeth = rvt_compute_aeth(qp);
1233 ohdr->bth[0] = cpu_to_be32(bth0);
1234 ohdr->bth[1] = cpu_to_be32(bth1 | qp->remote_qpn);
1235 ohdr->bth[2] = cpu_to_be32(mask_psn(qp->r_ack_psn));
1238 static inline void hfi1_queue_rc_ack(struct hfi1_packet *packet, bool is_fecn)
1240 struct rvt_qp *qp = packet->qp;
1241 struct hfi1_ibport *ibp;
1242 unsigned long flags;
1244 spin_lock_irqsave(&qp->s_lock, flags);
1245 if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK))
1247 ibp = rcd_to_iport(packet->rcd);
1248 this_cpu_inc(*ibp->rvp.rc_qacks);
1249 qp->s_flags |= RVT_S_ACK_PENDING | RVT_S_RESP_PENDING;
1250 qp->s_nak_state = qp->r_nak_state;
1251 qp->s_ack_psn = qp->r_ack_psn;
1253 qp->s_flags |= RVT_S_ECN;
1255 /* Schedule the send tasklet. */
1256 hfi1_schedule_send(qp);
1258 spin_unlock_irqrestore(&qp->s_lock, flags);
1261 static inline void hfi1_make_rc_ack_9B(struct hfi1_packet *packet,
1262 struct hfi1_opa_header *opa_hdr,
1263 u8 sc5, bool is_fecn,
1264 u64 *pbc_flags, u32 *hwords,
1267 struct rvt_qp *qp = packet->qp;
1268 struct hfi1_ibport *ibp = rcd_to_iport(packet->rcd);
1269 struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
1270 struct ib_header *hdr = &opa_hdr->ibh;
1271 struct ib_other_headers *ohdr;
1272 u16 lrh0 = HFI1_LRH_BTH;
1276 opa_hdr->hdr_type = HFI1_PKT_TYPE_9B;
1278 /* header size in 32-bit words LRH+BTH+AETH = (8+12+4)/4 */
1281 if (unlikely(rdma_ah_get_ah_flags(&qp->remote_ah_attr) & IB_AH_GRH)) {
1282 *hwords += hfi1_make_grh(ibp, &hdr->u.l.grh,
1283 rdma_ah_read_grh(&qp->remote_ah_attr),
1284 *hwords - 2, SIZE_OF_CRC);
1285 ohdr = &hdr->u.l.oth;
1286 lrh0 = HFI1_LRH_GRH;
1288 /* set PBC_DC_INFO bit (aka SC[4]) in pbc_flags */
1289 *pbc_flags |= ((!!(sc5 & 0x10)) << PBC_DC_INFO_SHIFT);
1291 /* read pkey_index w/o lock (its atomic) */
1292 pkey = hfi1_get_pkey(ibp, qp->s_pkey_index);
1294 lrh0 |= (sc5 & IB_SC_MASK) << IB_SC_SHIFT |
1295 (rdma_ah_get_sl(&qp->remote_ah_attr) & IB_SL_MASK) <<
1298 hfi1_make_ib_hdr(hdr, lrh0, *hwords + SIZE_OF_CRC,
1299 opa_get_lid(rdma_ah_get_dlid(&qp->remote_ah_attr), 9B),
1300 ppd->lid | rdma_ah_get_path_bits(&qp->remote_ah_attr));
1302 bth0 = pkey | (OP(ACKNOWLEDGE) << 24);
1303 if (qp->s_mig_state == IB_MIG_MIGRATED)
1304 bth0 |= IB_BTH_MIG_REQ;
1305 bth1 = (!!is_fecn) << IB_BECN_SHIFT;
1307 * Inline ACKs go out without the use of the Verbs send engine, so
1308 * we need to set the STL Verbs Extended bit here
1310 bth1 |= HFI1_CAP_IS_KSET(OPFN) << IB_BTHE_E_SHIFT;
1311 hfi1_make_bth_aeth(qp, ohdr, bth0, bth1);
1314 static inline void hfi1_make_rc_ack_16B(struct hfi1_packet *packet,
1315 struct hfi1_opa_header *opa_hdr,
1316 u8 sc5, bool is_fecn,
1317 u64 *pbc_flags, u32 *hwords,
1320 struct rvt_qp *qp = packet->qp;
1321 struct hfi1_ibport *ibp = rcd_to_iport(packet->rcd);
1322 struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
1323 struct hfi1_16b_header *hdr = &opa_hdr->opah;
1324 struct ib_other_headers *ohdr;
1327 bool becn = is_fecn;
1328 u8 l4 = OPA_16B_L4_IB_LOCAL;
1331 opa_hdr->hdr_type = HFI1_PKT_TYPE_16B;
1333 /* header size in 32-bit words 16B LRH+BTH+AETH = (16+12+4)/4 */
1335 extra_bytes = hfi1_get_16b_padding(*hwords << 2, 0);
1336 *nwords = SIZE_OF_CRC + ((extra_bytes + SIZE_OF_LT) >> 2);
1338 if (unlikely(rdma_ah_get_ah_flags(&qp->remote_ah_attr) & IB_AH_GRH) &&
1339 hfi1_check_mcast(rdma_ah_get_dlid(&qp->remote_ah_attr))) {
1340 *hwords += hfi1_make_grh(ibp, &hdr->u.l.grh,
1341 rdma_ah_read_grh(&qp->remote_ah_attr),
1342 *hwords - 4, *nwords);
1343 ohdr = &hdr->u.l.oth;
1344 l4 = OPA_16B_L4_IB_GLOBAL;
1346 *pbc_flags |= PBC_PACKET_BYPASS | PBC_INSERT_BYPASS_ICRC;
1348 /* read pkey_index w/o lock (its atomic) */
1349 pkey = hfi1_get_pkey(ibp, qp->s_pkey_index);
1351 /* Convert dwords to flits */
1352 len = (*hwords + *nwords) >> 1;
1354 hfi1_make_16b_hdr(hdr, ppd->lid |
1355 (rdma_ah_get_path_bits(&qp->remote_ah_attr) &
1356 ((1 << ppd->lmc) - 1)),
1357 opa_get_lid(rdma_ah_get_dlid(&qp->remote_ah_attr),
1358 16B), len, pkey, becn, 0, l4, sc5);
1360 bth0 = pkey | (OP(ACKNOWLEDGE) << 24);
1361 bth0 |= extra_bytes << 20;
1362 if (qp->s_mig_state == IB_MIG_MIGRATED)
1363 bth1 = OPA_BTH_MIG_REQ;
1364 hfi1_make_bth_aeth(qp, ohdr, bth0, bth1);
1367 typedef void (*hfi1_make_rc_ack)(struct hfi1_packet *packet,
1368 struct hfi1_opa_header *opa_hdr,
1369 u8 sc5, bool is_fecn,
1370 u64 *pbc_flags, u32 *hwords,
1373 /* We support only two types - 9B and 16B for now */
1374 static const hfi1_make_rc_ack hfi1_make_rc_ack_tbl[2] = {
1375 [HFI1_PKT_TYPE_9B] = &hfi1_make_rc_ack_9B,
1376 [HFI1_PKT_TYPE_16B] = &hfi1_make_rc_ack_16B
1380 * hfi1_send_rc_ack - Construct an ACK packet and send it
1381 * @qp: a pointer to the QP
1383 * This is called from hfi1_rc_rcv() and handle_receive_interrupt().
1384 * Note that RDMA reads and atomics are handled in the
1385 * send side QP state and send engine.
1387 void hfi1_send_rc_ack(struct hfi1_packet *packet, bool is_fecn)
1389 struct hfi1_ctxtdata *rcd = packet->rcd;
1390 struct rvt_qp *qp = packet->qp;
1391 struct hfi1_ibport *ibp = rcd_to_iport(rcd);
1392 struct hfi1_qp_priv *priv = qp->priv;
1393 struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
1394 u8 sc5 = ibp->sl_to_sc[rdma_ah_get_sl(&qp->remote_ah_attr)];
1395 u64 pbc, pbc_flags = 0;
1399 struct pio_buf *pbuf;
1400 struct hfi1_opa_header opa_hdr;
1402 /* clear the defer count */
1405 /* Don't send ACK or NAK if a RDMA read or atomic is pending. */
1406 if (qp->s_flags & RVT_S_RESP_PENDING) {
1407 hfi1_queue_rc_ack(packet, is_fecn);
1411 /* Ensure s_rdma_ack_cnt changes are committed */
1412 if (qp->s_rdma_ack_cnt) {
1413 hfi1_queue_rc_ack(packet, is_fecn);
1417 /* Don't try to send ACKs if the link isn't ACTIVE */
1418 if (driver_lstate(ppd) != IB_PORT_ACTIVE)
1421 /* Make the appropriate header */
1422 hfi1_make_rc_ack_tbl[priv->hdr_type](packet, &opa_hdr, sc5, is_fecn,
1423 &pbc_flags, &hwords, &nwords);
1425 plen = 2 /* PBC */ + hwords + nwords;
1426 pbc = create_pbc(ppd, pbc_flags, qp->srate_mbps,
1427 sc_to_vlt(ppd->dd, sc5), plen);
1428 pbuf = sc_buffer_alloc(rcd->sc, plen, NULL, NULL);
1429 if (IS_ERR_OR_NULL(pbuf)) {
1431 * We have no room to send at the moment. Pass
1432 * responsibility for sending the ACK to the send engine
1433 * so that when enough buffer space becomes available,
1434 * the ACK is sent ahead of other outgoing packets.
1436 hfi1_queue_rc_ack(packet, is_fecn);
1439 trace_ack_output_ibhdr(dd_from_ibdev(qp->ibqp.device),
1440 &opa_hdr, ib_is_sc5(sc5));
1442 /* write the pbc and data */
1443 ppd->dd->pio_inline_send(ppd->dd, pbuf, pbc,
1444 (priv->hdr_type == HFI1_PKT_TYPE_9B ?
1445 (void *)&opa_hdr.ibh :
1446 (void *)&opa_hdr.opah), hwords);
1451 * update_num_rd_atomic - update the qp->s_num_rd_atomic
1453 * @psn: the packet sequence number to restart at
1456 * This is called from reset_psn() to update qp->s_num_rd_atomic
1457 * for the current wqe.
1458 * Called at interrupt level with the QP s_lock held.
1460 static void update_num_rd_atomic(struct rvt_qp *qp, u32 psn,
1461 struct rvt_swqe *wqe)
1463 u32 opcode = wqe->wr.opcode;
1465 if (opcode == IB_WR_RDMA_READ ||
1466 opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
1467 opcode == IB_WR_ATOMIC_FETCH_AND_ADD) {
1468 qp->s_num_rd_atomic++;
1469 } else if (opcode == IB_WR_TID_RDMA_READ) {
1470 struct tid_rdma_request *req = wqe_to_tid_req(wqe);
1471 struct hfi1_qp_priv *priv = qp->priv;
1473 if (cmp_psn(psn, wqe->lpsn) <= 0) {
1476 cur_seg = (psn - wqe->psn) / priv->pkts_ps;
1477 req->ack_pending = cur_seg - req->comp_seg;
1478 priv->pending_tid_r_segs += req->ack_pending;
1479 qp->s_num_rd_atomic += req->ack_pending;
1480 trace_hfi1_tid_req_update_num_rd_atomic(qp, 0,
1486 priv->pending_tid_r_segs += req->total_segs;
1487 qp->s_num_rd_atomic += req->total_segs;
1493 * reset_psn - reset the QP state to send starting from PSN
1495 * @psn: the packet sequence number to restart at
1497 * This is called from hfi1_rc_rcv() to process an incoming RC ACK
1499 * Called at interrupt level with the QP s_lock held.
1501 static void reset_psn(struct rvt_qp *qp, u32 psn)
1503 u32 n = qp->s_acked;
1504 struct rvt_swqe *wqe = rvt_get_swqe_ptr(qp, n);
1506 struct hfi1_qp_priv *priv = qp->priv;
1508 lockdep_assert_held(&qp->s_lock);
1510 priv->pending_tid_r_segs = 0;
1511 priv->pending_tid_w_resp = 0;
1512 qp->s_num_rd_atomic = 0;
1515 * If we are starting the request from the beginning,
1516 * let the normal send code handle initialization.
1518 if (cmp_psn(psn, wqe->psn) <= 0) {
1519 qp->s_state = OP(SEND_LAST);
1522 update_num_rd_atomic(qp, psn, wqe);
1524 /* Find the work request opcode corresponding to the given PSN. */
1528 if (++n == qp->s_size)
1530 if (n == qp->s_tail)
1532 wqe = rvt_get_swqe_ptr(qp, n);
1533 diff = cmp_psn(psn, wqe->psn);
1535 /* Point wqe back to the previous one*/
1536 wqe = rvt_get_swqe_ptr(qp, qp->s_cur);
1541 * If we are starting the request from the beginning,
1542 * let the normal send code handle initialization.
1545 qp->s_state = OP(SEND_LAST);
1549 update_num_rd_atomic(qp, psn, wqe);
1551 opcode = wqe->wr.opcode;
1554 * Set the state to restart in the middle of a request.
1555 * Don't change the s_sge, s_cur_sge, or s_cur_size.
1556 * See hfi1_make_rc_req().
1560 case IB_WR_SEND_WITH_IMM:
1561 qp->s_state = OP(RDMA_READ_RESPONSE_FIRST);
1564 case IB_WR_RDMA_WRITE:
1565 case IB_WR_RDMA_WRITE_WITH_IMM:
1566 qp->s_state = OP(RDMA_READ_RESPONSE_LAST);
1569 case IB_WR_TID_RDMA_WRITE:
1570 qp->s_state = TID_OP(WRITE_RESP);
1573 case IB_WR_RDMA_READ:
1574 qp->s_state = OP(RDMA_READ_RESPONSE_MIDDLE);
1577 case IB_WR_TID_RDMA_READ:
1578 qp->s_state = TID_OP(READ_RESP);
1583 * This case shouldn't happen since its only
1586 qp->s_state = OP(SEND_LAST);
1589 priv->s_flags &= ~HFI1_S_TID_WAIT_INTERLCK;
1592 * Set RVT_S_WAIT_PSN as rc_complete() may start the timer
1593 * asynchronously before the send engine can get scheduled.
1594 * Doing it in hfi1_make_rc_req() is too late.
1596 if ((cmp_psn(qp->s_psn, qp->s_sending_hpsn) <= 0) &&
1597 (cmp_psn(qp->s_sending_psn, qp->s_sending_hpsn) <= 0))
1598 qp->s_flags |= RVT_S_WAIT_PSN;
1599 qp->s_flags &= ~HFI1_S_AHG_VALID;
1600 trace_hfi1_sender_reset_psn(qp);
1604 * Back up requester to resend the last un-ACKed request.
1605 * The QP r_lock and s_lock should be held and interrupts disabled.
1607 void hfi1_restart_rc(struct rvt_qp *qp, u32 psn, int wait)
1609 struct hfi1_qp_priv *priv = qp->priv;
1610 struct rvt_swqe *wqe = rvt_get_swqe_ptr(qp, qp->s_acked);
1611 struct hfi1_ibport *ibp;
1613 lockdep_assert_held(&qp->r_lock);
1614 lockdep_assert_held(&qp->s_lock);
1615 trace_hfi1_sender_restart_rc(qp);
1616 if (qp->s_retry == 0) {
1617 if (qp->s_mig_state == IB_MIG_ARMED) {
1618 hfi1_migrate_qp(qp);
1619 qp->s_retry = qp->s_retry_cnt;
1620 } else if (qp->s_last == qp->s_acked) {
1622 * We need special handling for the OPFN request WQEs as
1623 * they are not allowed to generate real user errors
1625 if (wqe->wr.opcode == IB_WR_OPFN) {
1626 struct hfi1_ibport *ibp =
1627 to_iport(qp->ibqp.device, qp->port_num);
1629 * Call opfn_conn_reply() with capcode and
1630 * remaining data as 0 to close out the
1633 opfn_conn_reply(qp, priv->opfn.curr);
1634 wqe = do_rc_completion(qp, wqe, ibp);
1635 qp->s_flags &= ~RVT_S_WAIT_ACK;
1637 trace_hfi1_tid_write_sender_restart_rc(qp, 0);
1638 if (wqe->wr.opcode == IB_WR_TID_RDMA_READ) {
1639 struct tid_rdma_request *req;
1641 req = wqe_to_tid_req(wqe);
1642 hfi1_kern_exp_rcv_clear_all(req);
1643 hfi1_kern_clear_hw_flow(priv->rcd, qp);
1646 hfi1_trdma_send_complete(qp, wqe,
1647 IB_WC_RETRY_EXC_ERR);
1648 rvt_error_qp(qp, IB_WC_WR_FLUSH_ERR);
1651 } else { /* need to handle delayed completion */
1658 ibp = to_iport(qp->ibqp.device, qp->port_num);
1659 if (wqe->wr.opcode == IB_WR_RDMA_READ ||
1660 wqe->wr.opcode == IB_WR_TID_RDMA_READ)
1661 ibp->rvp.n_rc_resends++;
1663 ibp->rvp.n_rc_resends += delta_psn(qp->s_psn, psn);
1665 qp->s_flags &= ~(RVT_S_WAIT_FENCE | RVT_S_WAIT_RDMAR |
1666 RVT_S_WAIT_SSN_CREDIT | RVT_S_WAIT_PSN |
1667 RVT_S_WAIT_ACK | HFI1_S_WAIT_TID_RESP);
1669 qp->s_flags |= RVT_S_SEND_ONE;
1674 * Set qp->s_sending_psn to the next PSN after the given one.
1675 * This would be psn+1 except when RDMA reads or TID RDMA ops
1678 static void reset_sending_psn(struct rvt_qp *qp, u32 psn)
1680 struct rvt_swqe *wqe;
1683 lockdep_assert_held(&qp->s_lock);
1684 /* Find the work request corresponding to the given PSN. */
1686 wqe = rvt_get_swqe_ptr(qp, n);
1687 if (cmp_psn(psn, wqe->lpsn) <= 0) {
1688 if (wqe->wr.opcode == IB_WR_RDMA_READ ||
1689 wqe->wr.opcode == IB_WR_TID_RDMA_READ ||
1690 wqe->wr.opcode == IB_WR_TID_RDMA_WRITE)
1691 qp->s_sending_psn = wqe->lpsn + 1;
1693 qp->s_sending_psn = psn + 1;
1696 if (++n == qp->s_size)
1698 if (n == qp->s_tail)
1704 * hfi1_rc_verbs_aborted - handle abort status
1706 * @opah: the opa header
1708 * This code modifies both ACK bit in BTH[2]
1709 * and the s_flags to go into send one mode.
1711 * This serves to throttle the send engine to only
1712 * send a single packet in the likely case the
1713 * a link has gone down.
1715 void hfi1_rc_verbs_aborted(struct rvt_qp *qp, struct hfi1_opa_header *opah)
1717 struct ib_other_headers *ohdr = hfi1_get_rc_ohdr(opah);
1718 u8 opcode = ib_bth_get_opcode(ohdr);
1721 /* ignore responses */
1722 if ((opcode >= OP(RDMA_READ_RESPONSE_FIRST) &&
1723 opcode <= OP(ATOMIC_ACKNOWLEDGE)) ||
1724 opcode == TID_OP(READ_RESP) ||
1725 opcode == TID_OP(WRITE_RESP))
1728 psn = ib_bth_get_psn(ohdr) | IB_BTH_REQ_ACK;
1729 ohdr->bth[2] = cpu_to_be32(psn);
1730 qp->s_flags |= RVT_S_SEND_ONE;
1734 * This should be called with the QP s_lock held and interrupts disabled.
1736 void hfi1_rc_send_complete(struct rvt_qp *qp, struct hfi1_opa_header *opah)
1738 struct ib_other_headers *ohdr;
1739 struct hfi1_qp_priv *priv = qp->priv;
1740 struct rvt_swqe *wqe;
1741 u32 opcode, head, tail;
1743 struct tid_rdma_request *req;
1745 lockdep_assert_held(&qp->s_lock);
1746 if (!(ib_rvt_state_ops[qp->state] & RVT_SEND_OR_FLUSH_OR_RECV_OK))
1749 ohdr = hfi1_get_rc_ohdr(opah);
1750 opcode = ib_bth_get_opcode(ohdr);
1751 if ((opcode >= OP(RDMA_READ_RESPONSE_FIRST) &&
1752 opcode <= OP(ATOMIC_ACKNOWLEDGE)) ||
1753 opcode == TID_OP(READ_RESP) ||
1754 opcode == TID_OP(WRITE_RESP)) {
1755 WARN_ON(!qp->s_rdma_ack_cnt);
1756 qp->s_rdma_ack_cnt--;
1760 psn = ib_bth_get_psn(ohdr);
1762 * Don't attempt to reset the sending PSN for packets in the
1763 * KDETH PSN space since the PSN does not match anything.
1765 if (opcode != TID_OP(WRITE_DATA) &&
1766 opcode != TID_OP(WRITE_DATA_LAST) &&
1767 opcode != TID_OP(ACK) && opcode != TID_OP(RESYNC))
1768 reset_sending_psn(qp, psn);
1770 /* Handle TID RDMA WRITE packets differently */
1771 if (opcode >= TID_OP(WRITE_REQ) &&
1772 opcode <= TID_OP(WRITE_DATA_LAST)) {
1773 head = priv->s_tid_head;
1774 tail = priv->s_tid_cur;
1776 * s_tid_cur is set to s_tid_head in the case, where
1777 * a new TID RDMA request is being started and all
1778 * previous ones have been completed.
1779 * Therefore, we need to do a secondary check in order
1780 * to properly determine whether we should start the
1783 wqe = rvt_get_swqe_ptr(qp, tail);
1784 req = wqe_to_tid_req(wqe);
1785 if (head == tail && req->comp_seg < req->total_segs) {
1787 tail = qp->s_size - 1;
1797 * Start timer after a packet requesting an ACK has been sent and
1798 * there are still requests that haven't been acked.
1800 if ((psn & IB_BTH_REQ_ACK) && tail != head &&
1801 opcode != TID_OP(WRITE_DATA) && opcode != TID_OP(WRITE_DATA_LAST) &&
1802 opcode != TID_OP(RESYNC) &&
1804 (RVT_S_TIMER | RVT_S_WAIT_RNR | RVT_S_WAIT_PSN)) &&
1805 (ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK)) {
1806 if (opcode == TID_OP(READ_REQ))
1807 rvt_add_retry_timer_ext(qp, priv->timeout_shift);
1809 rvt_add_retry_timer(qp);
1812 /* Start TID RDMA ACK timer */
1813 if ((opcode == TID_OP(WRITE_DATA) ||
1814 opcode == TID_OP(WRITE_DATA_LAST) ||
1815 opcode == TID_OP(RESYNC)) &&
1816 (psn & IB_BTH_REQ_ACK) &&
1817 !(priv->s_flags & HFI1_S_TID_RETRY_TIMER) &&
1818 (ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK)) {
1820 * The TID RDMA ACK packet could be received before this
1821 * function is called. Therefore, add the timer only if TID
1822 * RDMA ACK packets are actually pending.
1824 wqe = rvt_get_swqe_ptr(qp, qp->s_acked);
1825 req = wqe_to_tid_req(wqe);
1826 if (wqe->wr.opcode == IB_WR_TID_RDMA_WRITE &&
1827 req->ack_seg < req->cur_seg)
1828 hfi1_add_tid_retry_timer(qp);
1831 while (qp->s_last != qp->s_acked) {
1832 wqe = rvt_get_swqe_ptr(qp, qp->s_last);
1833 if (cmp_psn(wqe->lpsn, qp->s_sending_psn) >= 0 &&
1834 cmp_psn(qp->s_sending_psn, qp->s_sending_hpsn) <= 0)
1836 trdma_clean_swqe(qp, wqe);
1837 trace_hfi1_qp_send_completion(qp, wqe, qp->s_last);
1838 rvt_qp_complete_swqe(qp,
1840 ib_hfi1_wc_opcode[wqe->wr.opcode],
1844 * If we were waiting for sends to complete before re-sending,
1845 * and they are now complete, restart sending.
1847 trace_hfi1_sendcomplete(qp, psn);
1848 if (qp->s_flags & RVT_S_WAIT_PSN &&
1849 cmp_psn(qp->s_sending_psn, qp->s_sending_hpsn) > 0) {
1850 qp->s_flags &= ~RVT_S_WAIT_PSN;
1851 qp->s_sending_psn = qp->s_psn;
1852 qp->s_sending_hpsn = qp->s_psn - 1;
1853 hfi1_schedule_send(qp);
1857 static inline void update_last_psn(struct rvt_qp *qp, u32 psn)
1859 qp->s_last_psn = psn;
1863 * Generate a SWQE completion.
1864 * This is similar to hfi1_send_complete but has to check to be sure
1865 * that the SGEs are not being referenced if the SWQE is being resent.
1867 struct rvt_swqe *do_rc_completion(struct rvt_qp *qp,
1868 struct rvt_swqe *wqe,
1869 struct hfi1_ibport *ibp)
1871 struct hfi1_qp_priv *priv = qp->priv;
1873 lockdep_assert_held(&qp->s_lock);
1875 * Don't decrement refcount and don't generate a
1876 * completion if the SWQE is being resent until the send
1879 trace_hfi1_rc_completion(qp, wqe->lpsn);
1880 if (cmp_psn(wqe->lpsn, qp->s_sending_psn) < 0 ||
1881 cmp_psn(qp->s_sending_psn, qp->s_sending_hpsn) > 0) {
1882 trdma_clean_swqe(qp, wqe);
1883 trace_hfi1_qp_send_completion(qp, wqe, qp->s_last);
1884 rvt_qp_complete_swqe(qp,
1886 ib_hfi1_wc_opcode[wqe->wr.opcode],
1889 struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
1891 this_cpu_inc(*ibp->rvp.rc_delayed_comp);
1893 * If send progress not running attempt to progress
1896 if (ppd->dd->flags & HFI1_HAS_SEND_DMA) {
1897 struct sdma_engine *engine;
1898 u8 sl = rdma_ah_get_sl(&qp->remote_ah_attr);
1901 /* For now use sc to find engine */
1902 sc5 = ibp->sl_to_sc[sl];
1903 engine = qp_to_sdma_engine(qp, sc5);
1904 sdma_engine_progress_schedule(engine);
1908 qp->s_retry = qp->s_retry_cnt;
1910 * Don't update the last PSN if the request being completed is
1911 * a TID RDMA WRITE request.
1912 * Completion of the TID RDMA WRITE requests are done by the
1913 * TID RDMA ACKs and as such could be for a request that has
1914 * already been ACKed as far as the IB state machine is
1917 if (wqe->wr.opcode != IB_WR_TID_RDMA_WRITE)
1918 update_last_psn(qp, wqe->lpsn);
1921 * If we are completing a request which is in the process of
1922 * being resent, we can stop re-sending it since we know the
1923 * responder has already seen it.
1925 if (qp->s_acked == qp->s_cur) {
1926 if (++qp->s_cur >= qp->s_size)
1928 qp->s_acked = qp->s_cur;
1929 wqe = rvt_get_swqe_ptr(qp, qp->s_cur);
1930 if (qp->s_acked != qp->s_tail) {
1931 qp->s_state = OP(SEND_LAST);
1932 qp->s_psn = wqe->psn;
1935 if (++qp->s_acked >= qp->s_size)
1937 if (qp->state == IB_QPS_SQD && qp->s_acked == qp->s_cur)
1939 wqe = rvt_get_swqe_ptr(qp, qp->s_acked);
1941 if (priv->s_flags & HFI1_S_TID_WAIT_INTERLCK) {
1942 priv->s_flags &= ~HFI1_S_TID_WAIT_INTERLCK;
1943 hfi1_schedule_send(qp);
1948 static void set_restart_qp(struct rvt_qp *qp, struct hfi1_ctxtdata *rcd)
1950 /* Retry this request. */
1951 if (!(qp->r_flags & RVT_R_RDMAR_SEQ)) {
1952 qp->r_flags |= RVT_R_RDMAR_SEQ;
1953 hfi1_restart_rc(qp, qp->s_last_psn + 1, 0);
1954 if (list_empty(&qp->rspwait)) {
1955 qp->r_flags |= RVT_R_RSP_SEND;
1957 list_add_tail(&qp->rspwait, &rcd->qp_wait_list);
1963 * update_qp_retry_state - Update qp retry state.
1965 * @psn: the packet sequence number of the TID RDMA WRITE RESP.
1966 * @spsn: The start psn for the given TID RDMA WRITE swqe.
1967 * @lpsn: The last psn for the given TID RDMA WRITE swqe.
1969 * This function is called to update the qp retry state upon
1970 * receiving a TID WRITE RESP after the qp is scheduled to retry
1973 static void update_qp_retry_state(struct rvt_qp *qp, u32 psn, u32 spsn,
1976 struct hfi1_qp_priv *qpriv = qp->priv;
1978 qp->s_psn = psn + 1;
1980 * If this is the first TID RDMA WRITE RESP packet for the current
1981 * request, change the s_state so that the retry will be processed
1982 * correctly. Similarly, if this is the last TID RDMA WRITE RESP
1983 * packet, change the s_state and advance the s_cur.
1985 if (cmp_psn(psn, lpsn) >= 0) {
1986 qp->s_cur = qpriv->s_tid_cur + 1;
1987 if (qp->s_cur >= qp->s_size)
1989 qp->s_state = TID_OP(WRITE_REQ);
1990 } else if (!cmp_psn(psn, spsn)) {
1991 qp->s_cur = qpriv->s_tid_cur;
1992 qp->s_state = TID_OP(WRITE_RESP);
1997 * do_rc_ack - process an incoming RC ACK
1998 * @qp: the QP the ACK came in on
1999 * @psn: the packet sequence number of the ACK
2000 * @opcode: the opcode of the request that resulted in the ACK
2002 * This is called from rc_rcv_resp() to process an incoming RC ACK
2004 * May be called at interrupt level, with the QP s_lock held.
2005 * Returns 1 if OK, 0 if current operation should be aborted (NAK).
2007 int do_rc_ack(struct rvt_qp *qp, u32 aeth, u32 psn, int opcode,
2008 u64 val, struct hfi1_ctxtdata *rcd)
2010 struct hfi1_ibport *ibp;
2011 enum ib_wc_status status;
2012 struct hfi1_qp_priv *qpriv = qp->priv;
2013 struct rvt_swqe *wqe;
2017 struct rvt_dev_info *rdi;
2019 lockdep_assert_held(&qp->s_lock);
2021 * Note that NAKs implicitly ACK outstanding SEND and RDMA write
2022 * requests and implicitly NAK RDMA read and atomic requests issued
2023 * before the NAK'ed request. The MSN won't include the NAK'ed
2024 * request but will include an ACK'ed request(s).
2027 if (aeth >> IB_AETH_NAK_SHIFT)
2029 wqe = rvt_get_swqe_ptr(qp, qp->s_acked);
2030 ibp = rcd_to_iport(rcd);
2033 * The MSN might be for a later WQE than the PSN indicates so
2034 * only complete WQEs that the PSN finishes.
2036 while ((diff = delta_psn(ack_psn, wqe->lpsn)) >= 0) {
2038 * RDMA_READ_RESPONSE_ONLY is a special case since
2039 * we want to generate completion events for everything
2040 * before the RDMA read, copy the data, then generate
2041 * the completion for the read.
2043 if (wqe->wr.opcode == IB_WR_RDMA_READ &&
2044 opcode == OP(RDMA_READ_RESPONSE_ONLY) &&
2050 * If this request is a RDMA read or atomic, and the ACK is
2051 * for a later operation, this ACK NAKs the RDMA read or
2052 * atomic. In other words, only a RDMA_READ_LAST or ONLY
2053 * can ACK a RDMA read and likewise for atomic ops. Note
2054 * that the NAK case can only happen if relaxed ordering is
2055 * used and requests are sent after an RDMA read or atomic
2056 * is sent but before the response is received.
2058 if ((wqe->wr.opcode == IB_WR_RDMA_READ &&
2059 (opcode != OP(RDMA_READ_RESPONSE_LAST) || diff != 0)) ||
2060 (wqe->wr.opcode == IB_WR_TID_RDMA_READ &&
2061 (opcode != TID_OP(READ_RESP) || diff != 0)) ||
2062 ((wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
2063 wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD) &&
2064 (opcode != OP(ATOMIC_ACKNOWLEDGE) || diff != 0)) ||
2065 (wqe->wr.opcode == IB_WR_TID_RDMA_WRITE &&
2066 (delta_psn(psn, qp->s_last_psn) != 1))) {
2067 set_restart_qp(qp, rcd);
2069 * No need to process the ACK/NAK since we are
2070 * restarting an earlier request.
2074 if (wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
2075 wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD) {
2076 u64 *vaddr = wqe->sg_list[0].vaddr;
2079 if (wqe->wr.opcode == IB_WR_OPFN)
2080 opfn_conn_reply(qp, val);
2082 if (qp->s_num_rd_atomic &&
2083 (wqe->wr.opcode == IB_WR_RDMA_READ ||
2084 wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
2085 wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD)) {
2086 qp->s_num_rd_atomic--;
2087 /* Restart sending task if fence is complete */
2088 if ((qp->s_flags & RVT_S_WAIT_FENCE) &&
2089 !qp->s_num_rd_atomic) {
2090 qp->s_flags &= ~(RVT_S_WAIT_FENCE |
2092 hfi1_schedule_send(qp);
2093 } else if (qp->s_flags & RVT_S_WAIT_RDMAR) {
2094 qp->s_flags &= ~(RVT_S_WAIT_RDMAR |
2096 hfi1_schedule_send(qp);
2101 * TID RDMA WRITE requests will be completed by the TID RDMA
2102 * ACK packet handler (see tid_rdma.c).
2104 if (wqe->wr.opcode == IB_WR_TID_RDMA_WRITE)
2107 wqe = do_rc_completion(qp, wqe, ibp);
2108 if (qp->s_acked == qp->s_tail)
2112 trace_hfi1_rc_ack_do(qp, aeth, psn, wqe);
2113 trace_hfi1_sender_do_rc_ack(qp);
2114 switch (aeth >> IB_AETH_NAK_SHIFT) {
2116 this_cpu_inc(*ibp->rvp.rc_acks);
2117 if (wqe->wr.opcode == IB_WR_TID_RDMA_READ) {
2118 if (wqe_to_tid_req(wqe)->ack_pending)
2119 rvt_mod_retry_timer_ext(qp,
2120 qpriv->timeout_shift);
2122 rvt_stop_rc_timers(qp);
2123 } else if (qp->s_acked != qp->s_tail) {
2124 struct rvt_swqe *__w = NULL;
2126 if (qpriv->s_tid_cur != HFI1_QP_WQE_INVALID)
2127 __w = rvt_get_swqe_ptr(qp, qpriv->s_tid_cur);
2130 * Stop timers if we've received all of the TID RDMA
2131 * WRITE * responses.
2133 if (__w && __w->wr.opcode == IB_WR_TID_RDMA_WRITE &&
2134 opcode == TID_OP(WRITE_RESP)) {
2136 * Normally, the loop above would correctly
2137 * process all WQEs from s_acked onward and
2138 * either complete them or check for correct
2140 * However, for TID RDMA, due to pipelining,
2141 * the response may not be for the request at
2142 * s_acked so the above look would just be
2143 * skipped. This does not allow for checking
2144 * the PSN sequencing. It has to be done
2147 if (cmp_psn(psn, qp->s_last_psn + 1)) {
2148 set_restart_qp(qp, rcd);
2152 * If the psn is being resent, stop the
2155 if (qp->s_cur != qp->s_tail &&
2156 cmp_psn(qp->s_psn, psn) <= 0)
2157 update_qp_retry_state(qp, psn,
2160 else if (--qpriv->pending_tid_w_resp)
2161 rvt_mod_retry_timer(qp);
2163 rvt_stop_rc_timers(qp);
2166 * We are expecting more ACKs so
2167 * mod the retry timer.
2169 rvt_mod_retry_timer(qp);
2171 * We can stop re-sending the earlier packets
2172 * and continue with the next packet the
2175 if (cmp_psn(qp->s_psn, psn) <= 0)
2176 reset_psn(qp, psn + 1);
2179 /* No more acks - kill all timers */
2180 rvt_stop_rc_timers(qp);
2181 if (cmp_psn(qp->s_psn, psn) <= 0) {
2182 qp->s_state = OP(SEND_LAST);
2183 qp->s_psn = psn + 1;
2186 if (qp->s_flags & RVT_S_WAIT_ACK) {
2187 qp->s_flags &= ~RVT_S_WAIT_ACK;
2188 hfi1_schedule_send(qp);
2190 rvt_get_credit(qp, aeth);
2191 qp->s_rnr_retry = qp->s_rnr_retry_cnt;
2192 qp->s_retry = qp->s_retry_cnt;
2194 * If the current request is a TID RDMA WRITE request and the
2195 * response is not a TID RDMA WRITE RESP packet, s_last_psn
2196 * can't be advanced.
2198 if (wqe->wr.opcode == IB_WR_TID_RDMA_WRITE &&
2199 opcode != TID_OP(WRITE_RESP) &&
2200 cmp_psn(psn, wqe->psn) >= 0)
2202 update_last_psn(qp, psn);
2205 case 1: /* RNR NAK */
2206 ibp->rvp.n_rnr_naks++;
2207 if (qp->s_acked == qp->s_tail)
2209 if (qp->s_flags & RVT_S_WAIT_RNR)
2211 rdi = ib_to_rvt(qp->ibqp.device);
2212 if (!(rdi->post_parms[wqe->wr.opcode].flags &
2213 RVT_OPERATION_IGN_RNR_CNT)) {
2214 if (qp->s_rnr_retry == 0) {
2215 status = IB_WC_RNR_RETRY_EXC_ERR;
2218 if (qp->s_rnr_retry_cnt < 7 && qp->s_rnr_retry_cnt > 0)
2223 * The last valid PSN is the previous PSN. For TID RDMA WRITE
2224 * request, s_last_psn should be incremented only when a TID
2225 * RDMA WRITE RESP is received to avoid skipping lost TID RDMA
2226 * WRITE RESP packets.
2228 if (wqe->wr.opcode == IB_WR_TID_RDMA_WRITE) {
2229 reset_psn(qp, qp->s_last_psn + 1);
2231 update_last_psn(qp, psn - 1);
2235 ibp->rvp.n_rc_resends += delta_psn(qp->s_psn, psn);
2236 qp->s_flags &= ~(RVT_S_WAIT_SSN_CREDIT | RVT_S_WAIT_ACK);
2237 rvt_stop_rc_timers(qp);
2238 rvt_add_rnr_timer(qp, aeth);
2242 if (qp->s_acked == qp->s_tail)
2244 /* The last valid PSN is the previous PSN. */
2245 update_last_psn(qp, psn - 1);
2246 switch ((aeth >> IB_AETH_CREDIT_SHIFT) &
2247 IB_AETH_CREDIT_MASK) {
2248 case 0: /* PSN sequence error */
2249 ibp->rvp.n_seq_naks++;
2251 * Back up to the responder's expected PSN.
2252 * Note that we might get a NAK in the middle of an
2253 * RDMA READ response which terminates the RDMA
2256 hfi1_restart_rc(qp, psn, 0);
2257 hfi1_schedule_send(qp);
2260 case 1: /* Invalid Request */
2261 status = IB_WC_REM_INV_REQ_ERR;
2262 ibp->rvp.n_other_naks++;
2265 case 2: /* Remote Access Error */
2266 status = IB_WC_REM_ACCESS_ERR;
2267 ibp->rvp.n_other_naks++;
2270 case 3: /* Remote Operation Error */
2271 status = IB_WC_REM_OP_ERR;
2272 ibp->rvp.n_other_naks++;
2274 if (qp->s_last == qp->s_acked) {
2275 if (wqe->wr.opcode == IB_WR_TID_RDMA_READ)
2276 hfi1_kern_read_tid_flow_free(qp);
2278 hfi1_trdma_send_complete(qp, wqe, status);
2279 rvt_error_qp(qp, IB_WC_WR_FLUSH_ERR);
2284 /* Ignore other reserved NAK error codes */
2287 qp->s_retry = qp->s_retry_cnt;
2288 qp->s_rnr_retry = qp->s_rnr_retry_cnt;
2291 default: /* 2: reserved */
2293 /* Ignore reserved NAK codes. */
2296 /* cannot be reached */
2298 rvt_stop_rc_timers(qp);
2303 * We have seen an out of sequence RDMA read middle or last packet.
2304 * This ACKs SENDs and RDMA writes up to the first RDMA read or atomic SWQE.
2306 static void rdma_seq_err(struct rvt_qp *qp, struct hfi1_ibport *ibp, u32 psn,
2307 struct hfi1_ctxtdata *rcd)
2309 struct rvt_swqe *wqe;
2311 lockdep_assert_held(&qp->s_lock);
2312 /* Remove QP from retry timer */
2313 rvt_stop_rc_timers(qp);
2315 wqe = rvt_get_swqe_ptr(qp, qp->s_acked);
2317 while (cmp_psn(psn, wqe->lpsn) > 0) {
2318 if (wqe->wr.opcode == IB_WR_RDMA_READ ||
2319 wqe->wr.opcode == IB_WR_TID_RDMA_READ ||
2320 wqe->wr.opcode == IB_WR_TID_RDMA_WRITE ||
2321 wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
2322 wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD)
2324 wqe = do_rc_completion(qp, wqe, ibp);
2327 ibp->rvp.n_rdma_seq++;
2328 qp->r_flags |= RVT_R_RDMAR_SEQ;
2329 hfi1_restart_rc(qp, qp->s_last_psn + 1, 0);
2330 if (list_empty(&qp->rspwait)) {
2331 qp->r_flags |= RVT_R_RSP_SEND;
2333 list_add_tail(&qp->rspwait, &rcd->qp_wait_list);
2338 * rc_rcv_resp - process an incoming RC response packet
2339 * @packet: data packet information
2341 * This is called from hfi1_rc_rcv() to process an incoming RC response
2342 * packet for the given QP.
2343 * Called at interrupt level.
2345 static void rc_rcv_resp(struct hfi1_packet *packet)
2347 struct hfi1_ctxtdata *rcd = packet->rcd;
2348 void *data = packet->payload;
2349 u32 tlen = packet->tlen;
2350 struct rvt_qp *qp = packet->qp;
2351 struct hfi1_ibport *ibp;
2352 struct ib_other_headers *ohdr = packet->ohdr;
2353 struct rvt_swqe *wqe;
2354 enum ib_wc_status status;
2355 unsigned long flags;
2359 u32 psn = ib_bth_get_psn(packet->ohdr);
2360 u32 pmtu = qp->pmtu;
2361 u16 hdrsize = packet->hlen;
2362 u8 opcode = packet->opcode;
2363 u8 pad = packet->pad;
2364 u8 extra_bytes = pad + packet->extra_byte + (SIZE_OF_CRC << 2);
2366 spin_lock_irqsave(&qp->s_lock, flags);
2367 trace_hfi1_ack(qp, psn);
2369 /* Ignore invalid responses. */
2370 if (cmp_psn(psn, READ_ONCE(qp->s_next_psn)) >= 0)
2373 /* Ignore duplicate responses. */
2374 diff = cmp_psn(psn, qp->s_last_psn);
2375 if (unlikely(diff <= 0)) {
2376 /* Update credits for "ghost" ACKs */
2377 if (diff == 0 && opcode == OP(ACKNOWLEDGE)) {
2378 aeth = be32_to_cpu(ohdr->u.aeth);
2379 if ((aeth >> IB_AETH_NAK_SHIFT) == 0)
2380 rvt_get_credit(qp, aeth);
2386 * Skip everything other than the PSN we expect, if we are waiting
2387 * for a reply to a restarted RDMA read or atomic op.
2389 if (qp->r_flags & RVT_R_RDMAR_SEQ) {
2390 if (cmp_psn(psn, qp->s_last_psn + 1) != 0)
2392 qp->r_flags &= ~RVT_R_RDMAR_SEQ;
2395 if (unlikely(qp->s_acked == qp->s_tail))
2397 wqe = rvt_get_swqe_ptr(qp, qp->s_acked);
2398 status = IB_WC_SUCCESS;
2401 case OP(ACKNOWLEDGE):
2402 case OP(ATOMIC_ACKNOWLEDGE):
2403 case OP(RDMA_READ_RESPONSE_FIRST):
2404 aeth = be32_to_cpu(ohdr->u.aeth);
2405 if (opcode == OP(ATOMIC_ACKNOWLEDGE))
2406 val = ib_u64_get(&ohdr->u.at.atomic_ack_eth);
2409 if (!do_rc_ack(qp, aeth, psn, opcode, val, rcd) ||
2410 opcode != OP(RDMA_READ_RESPONSE_FIRST))
2412 wqe = rvt_get_swqe_ptr(qp, qp->s_acked);
2413 if (unlikely(wqe->wr.opcode != IB_WR_RDMA_READ))
2416 * If this is a response to a resent RDMA read, we
2417 * have to be careful to copy the data to the right
2420 qp->s_rdma_read_len = restart_sge(&qp->s_rdma_read_sge,
2424 case OP(RDMA_READ_RESPONSE_MIDDLE):
2425 /* no AETH, no ACK */
2426 if (unlikely(cmp_psn(psn, qp->s_last_psn + 1)))
2428 if (unlikely(wqe->wr.opcode != IB_WR_RDMA_READ))
2431 if (unlikely(tlen != (hdrsize + pmtu + extra_bytes)))
2433 if (unlikely(pmtu >= qp->s_rdma_read_len))
2437 * We got a response so update the timeout.
2438 * 4.096 usec. * (1 << qp->timeout)
2440 rvt_mod_retry_timer(qp);
2441 if (qp->s_flags & RVT_S_WAIT_ACK) {
2442 qp->s_flags &= ~RVT_S_WAIT_ACK;
2443 hfi1_schedule_send(qp);
2446 if (opcode == OP(RDMA_READ_RESPONSE_MIDDLE))
2447 qp->s_retry = qp->s_retry_cnt;
2450 * Update the RDMA receive state but do the copy w/o
2451 * holding the locks and blocking interrupts.
2453 qp->s_rdma_read_len -= pmtu;
2454 update_last_psn(qp, psn);
2455 spin_unlock_irqrestore(&qp->s_lock, flags);
2456 rvt_copy_sge(qp, &qp->s_rdma_read_sge,
2457 data, pmtu, false, false);
2460 case OP(RDMA_READ_RESPONSE_ONLY):
2461 aeth = be32_to_cpu(ohdr->u.aeth);
2462 if (!do_rc_ack(qp, aeth, psn, opcode, 0, rcd))
2465 * Check that the data size is >= 0 && <= pmtu.
2466 * Remember to account for ICRC (4).
2468 if (unlikely(tlen < (hdrsize + extra_bytes)))
2471 * If this is a response to a resent RDMA read, we
2472 * have to be careful to copy the data to the right
2475 wqe = rvt_get_swqe_ptr(qp, qp->s_acked);
2476 qp->s_rdma_read_len = restart_sge(&qp->s_rdma_read_sge,
2480 case OP(RDMA_READ_RESPONSE_LAST):
2481 /* ACKs READ req. */
2482 if (unlikely(cmp_psn(psn, qp->s_last_psn + 1)))
2484 if (unlikely(wqe->wr.opcode != IB_WR_RDMA_READ))
2487 * Check that the data size is >= 1 && <= pmtu.
2488 * Remember to account for ICRC (4).
2490 if (unlikely(tlen <= (hdrsize + extra_bytes)))
2493 tlen -= hdrsize + extra_bytes;
2494 if (unlikely(tlen != qp->s_rdma_read_len))
2496 aeth = be32_to_cpu(ohdr->u.aeth);
2497 rvt_copy_sge(qp, &qp->s_rdma_read_sge,
2498 data, tlen, false, false);
2499 WARN_ON(qp->s_rdma_read_sge.num_sge);
2500 (void)do_rc_ack(qp, aeth, psn,
2501 OP(RDMA_READ_RESPONSE_LAST), 0, rcd);
2506 status = IB_WC_LOC_QP_OP_ERR;
2510 ibp = rcd_to_iport(rcd);
2511 rdma_seq_err(qp, ibp, psn, rcd);
2515 status = IB_WC_LOC_LEN_ERR;
2517 if (qp->s_last == qp->s_acked) {
2518 rvt_send_complete(qp, wqe, status);
2519 rvt_error_qp(qp, IB_WC_WR_FLUSH_ERR);
2522 spin_unlock_irqrestore(&qp->s_lock, flags);
2527 static inline void rc_cancel_ack(struct rvt_qp *qp)
2530 if (list_empty(&qp->rspwait))
2532 list_del_init(&qp->rspwait);
2533 qp->r_flags &= ~RVT_R_RSP_NAK;
2538 * rc_rcv_error - process an incoming duplicate or error RC packet
2539 * @ohdr: the other headers for this packet
2540 * @data: the packet data
2541 * @qp: the QP for this packet
2542 * @opcode: the opcode for this packet
2543 * @psn: the packet sequence number for this packet
2544 * @diff: the difference between the PSN and the expected PSN
2546 * This is called from hfi1_rc_rcv() to process an unexpected
2547 * incoming RC packet for the given QP.
2548 * Called at interrupt level.
2549 * Return 1 if no more processing is needed; otherwise return 0 to
2550 * schedule a response to be sent.
2552 static noinline int rc_rcv_error(struct ib_other_headers *ohdr, void *data,
2553 struct rvt_qp *qp, u32 opcode, u32 psn,
2554 int diff, struct hfi1_ctxtdata *rcd)
2556 struct hfi1_ibport *ibp = rcd_to_iport(rcd);
2557 struct rvt_ack_entry *e;
2558 unsigned long flags;
2560 u8 mra; /* most recent ACK */
2563 trace_hfi1_rcv_error(qp, psn);
2566 * Packet sequence error.
2567 * A NAK will ACK earlier sends and RDMA writes.
2568 * Don't queue the NAK if we already sent one.
2570 if (!qp->r_nak_state) {
2571 ibp->rvp.n_rc_seqnak++;
2572 qp->r_nak_state = IB_NAK_PSN_ERROR;
2573 /* Use the expected PSN. */
2574 qp->r_ack_psn = qp->r_psn;
2576 * Wait to send the sequence NAK until all packets
2577 * in the receive queue have been processed.
2578 * Otherwise, we end up propagating congestion.
2580 rc_defered_ack(rcd, qp);
2586 * Handle a duplicate request. Don't re-execute SEND, RDMA
2587 * write or atomic op. Don't NAK errors, just silently drop
2588 * the duplicate request. Note that r_sge, r_len, and
2589 * r_rcv_len may be in use so don't modify them.
2591 * We are supposed to ACK the earliest duplicate PSN but we
2592 * can coalesce an outstanding duplicate ACK. We have to
2593 * send the earliest so that RDMA reads can be restarted at
2594 * the requester's expected PSN.
2596 * First, find where this duplicate PSN falls within the
2597 * ACKs previously sent.
2598 * old_req is true if there is an older response that is scheduled
2599 * to be sent before sending this one.
2603 ibp->rvp.n_rc_dupreq++;
2605 spin_lock_irqsave(&qp->s_lock, flags);
2607 e = find_prev_entry(qp, psn, &prev, &mra, &old_req);
2610 case OP(RDMA_READ_REQUEST): {
2611 struct ib_reth *reth;
2616 * If we didn't find the RDMA read request in the ack queue,
2617 * we can ignore this request.
2619 if (!e || e->opcode != OP(RDMA_READ_REQUEST))
2621 /* RETH comes after BTH */
2622 reth = &ohdr->u.rc.reth;
2624 * Address range must be a subset of the original
2625 * request and start on pmtu boundaries.
2626 * We reuse the old ack_queue slot since the requester
2627 * should not back up and request an earlier PSN for the
2630 offset = delta_psn(psn, e->psn) * qp->pmtu;
2631 len = be32_to_cpu(reth->length);
2632 if (unlikely(offset + len != e->rdma_sge.sge_length))
2634 release_rdma_sge_mr(e);
2636 u32 rkey = be32_to_cpu(reth->rkey);
2637 u64 vaddr = get_ib_reth_vaddr(reth);
2640 ok = rvt_rkey_ok(qp, &e->rdma_sge, len, vaddr, rkey,
2641 IB_ACCESS_REMOTE_READ);
2645 e->rdma_sge.vaddr = NULL;
2646 e->rdma_sge.length = 0;
2647 e->rdma_sge.sge_length = 0;
2652 if (qp->s_acked_ack_queue == qp->s_tail_ack_queue)
2653 qp->s_acked_ack_queue = prev;
2654 qp->s_tail_ack_queue = prev;
2658 case OP(COMPARE_SWAP):
2659 case OP(FETCH_ADD): {
2661 * If we didn't find the atomic request in the ack queue
2662 * or the send engine is already backed up to send an
2663 * earlier entry, we can ignore this request.
2665 if (!e || e->opcode != (u8)opcode || old_req)
2667 if (qp->s_tail_ack_queue == qp->s_acked_ack_queue)
2668 qp->s_acked_ack_queue = prev;
2669 qp->s_tail_ack_queue = prev;
2675 * Ignore this operation if it doesn't request an ACK
2676 * or an earlier RDMA read or atomic is going to be resent.
2678 if (!(psn & IB_BTH_REQ_ACK) || old_req)
2681 * Resend the most recent ACK if this request is
2682 * after all the previous RDMA reads and atomics.
2684 if (mra == qp->r_head_ack_queue) {
2685 spin_unlock_irqrestore(&qp->s_lock, flags);
2686 qp->r_nak_state = 0;
2687 qp->r_ack_psn = qp->r_psn - 1;
2692 * Resend the RDMA read or atomic op which
2693 * ACKs this duplicate request.
2695 if (qp->s_tail_ack_queue == qp->s_acked_ack_queue)
2696 qp->s_acked_ack_queue = mra;
2697 qp->s_tail_ack_queue = mra;
2700 qp->s_ack_state = OP(ACKNOWLEDGE);
2701 qp->s_flags |= RVT_S_RESP_PENDING;
2702 qp->r_nak_state = 0;
2703 hfi1_schedule_send(qp);
2706 spin_unlock_irqrestore(&qp->s_lock, flags);
2714 static void log_cca_event(struct hfi1_pportdata *ppd, u8 sl, u32 rlid,
2715 u32 lqpn, u32 rqpn, u8 svc_type)
2717 struct opa_hfi1_cong_log_event_internal *cc_event;
2718 unsigned long flags;
2720 if (sl >= OPA_MAX_SLS)
2723 spin_lock_irqsave(&ppd->cc_log_lock, flags);
2725 ppd->threshold_cong_event_map[sl / 8] |= 1 << (sl % 8);
2726 ppd->threshold_event_counter++;
2728 cc_event = &ppd->cc_events[ppd->cc_log_idx++];
2729 if (ppd->cc_log_idx == OPA_CONG_LOG_ELEMS)
2730 ppd->cc_log_idx = 0;
2731 cc_event->lqpn = lqpn & RVT_QPN_MASK;
2732 cc_event->rqpn = rqpn & RVT_QPN_MASK;
2734 cc_event->svc_type = svc_type;
2735 cc_event->rlid = rlid;
2736 /* keep timestamp in units of 1.024 usec */
2737 cc_event->timestamp = ktime_get_ns() / 1024;
2739 spin_unlock_irqrestore(&ppd->cc_log_lock, flags);
2742 void process_becn(struct hfi1_pportdata *ppd, u8 sl, u32 rlid, u32 lqpn,
2743 u32 rqpn, u8 svc_type)
2745 struct cca_timer *cca_timer;
2746 u16 ccti, ccti_incr, ccti_timer, ccti_limit;
2747 u8 trigger_threshold;
2748 struct cc_state *cc_state;
2749 unsigned long flags;
2751 if (sl >= OPA_MAX_SLS)
2754 cc_state = get_cc_state(ppd);
2760 * 1) increase CCTI (for this SL)
2761 * 2) select IPG (i.e., call set_link_ipg())
2764 ccti_limit = cc_state->cct.ccti_limit;
2765 ccti_incr = cc_state->cong_setting.entries[sl].ccti_increase;
2766 ccti_timer = cc_state->cong_setting.entries[sl].ccti_timer;
2768 cc_state->cong_setting.entries[sl].trigger_threshold;
2770 spin_lock_irqsave(&ppd->cca_timer_lock, flags);
2772 cca_timer = &ppd->cca_timer[sl];
2773 if (cca_timer->ccti < ccti_limit) {
2774 if (cca_timer->ccti + ccti_incr <= ccti_limit)
2775 cca_timer->ccti += ccti_incr;
2777 cca_timer->ccti = ccti_limit;
2781 ccti = cca_timer->ccti;
2783 if (!hrtimer_active(&cca_timer->hrtimer)) {
2784 /* ccti_timer is in units of 1.024 usec */
2785 unsigned long nsec = 1024 * ccti_timer;
2787 hrtimer_start(&cca_timer->hrtimer, ns_to_ktime(nsec),
2788 HRTIMER_MODE_REL_PINNED);
2791 spin_unlock_irqrestore(&ppd->cca_timer_lock, flags);
2793 if ((trigger_threshold != 0) && (ccti >= trigger_threshold))
2794 log_cca_event(ppd, sl, rlid, lqpn, rqpn, svc_type);
2798 * hfi1_rc_rcv - process an incoming RC packet
2799 * @packet: data packet information
2801 * This is called from qp_rcv() to process an incoming RC packet
2803 * May be called at interrupt level.
2805 void hfi1_rc_rcv(struct hfi1_packet *packet)
2807 struct hfi1_ctxtdata *rcd = packet->rcd;
2808 void *data = packet->payload;
2809 u32 tlen = packet->tlen;
2810 struct rvt_qp *qp = packet->qp;
2811 struct hfi1_qp_priv *qpriv = qp->priv;
2812 struct hfi1_ibport *ibp = rcd_to_iport(rcd);
2813 struct ib_other_headers *ohdr = packet->ohdr;
2814 u32 opcode = packet->opcode;
2815 u32 hdrsize = packet->hlen;
2816 u32 psn = ib_bth_get_psn(packet->ohdr);
2817 u32 pad = packet->pad;
2819 u32 pmtu = qp->pmtu;
2821 struct ib_reth *reth;
2822 unsigned long flags;
2824 bool copy_last = false, fecn;
2826 u8 extra_bytes = pad + packet->extra_byte + (SIZE_OF_CRC << 2);
2828 lockdep_assert_held(&qp->r_lock);
2830 if (hfi1_ruc_check_hdr(ibp, packet))
2833 fecn = process_ecn(qp, packet);
2834 opfn_trigger_conn_request(qp, be32_to_cpu(ohdr->bth[1]));
2837 * Process responses (ACKs) before anything else. Note that the
2838 * packet sequence number will be for something in the send work
2839 * queue rather than the expected receive packet sequence number.
2840 * In other words, this QP is the requester.
2842 if (opcode >= OP(RDMA_READ_RESPONSE_FIRST) &&
2843 opcode <= OP(ATOMIC_ACKNOWLEDGE)) {
2844 rc_rcv_resp(packet);
2848 /* Compute 24 bits worth of difference. */
2849 diff = delta_psn(psn, qp->r_psn);
2850 if (unlikely(diff)) {
2851 if (rc_rcv_error(ohdr, data, qp, opcode, psn, diff, rcd))
2856 /* Check for opcode sequence errors. */
2857 switch (qp->r_state) {
2858 case OP(SEND_FIRST):
2859 case OP(SEND_MIDDLE):
2860 if (opcode == OP(SEND_MIDDLE) ||
2861 opcode == OP(SEND_LAST) ||
2862 opcode == OP(SEND_LAST_WITH_IMMEDIATE) ||
2863 opcode == OP(SEND_LAST_WITH_INVALIDATE))
2867 case OP(RDMA_WRITE_FIRST):
2868 case OP(RDMA_WRITE_MIDDLE):
2869 if (opcode == OP(RDMA_WRITE_MIDDLE) ||
2870 opcode == OP(RDMA_WRITE_LAST) ||
2871 opcode == OP(RDMA_WRITE_LAST_WITH_IMMEDIATE))
2876 if (opcode == OP(SEND_MIDDLE) ||
2877 opcode == OP(SEND_LAST) ||
2878 opcode == OP(SEND_LAST_WITH_IMMEDIATE) ||
2879 opcode == OP(SEND_LAST_WITH_INVALIDATE) ||
2880 opcode == OP(RDMA_WRITE_MIDDLE) ||
2881 opcode == OP(RDMA_WRITE_LAST) ||
2882 opcode == OP(RDMA_WRITE_LAST_WITH_IMMEDIATE))
2885 * Note that it is up to the requester to not send a new
2886 * RDMA read or atomic operation before receiving an ACK
2887 * for the previous operation.
2892 if (qp->state == IB_QPS_RTR && !(qp->r_flags & RVT_R_COMM_EST))
2895 /* OK, process the packet. */
2897 case OP(SEND_FIRST):
2898 ret = rvt_get_rwqe(qp, false);
2905 case OP(SEND_MIDDLE):
2906 case OP(RDMA_WRITE_MIDDLE):
2908 /* Check for invalid length PMTU or posted rwqe len. */
2910 * There will be no padding for 9B packet but 16B packets
2911 * will come in with some padding since we always add
2912 * CRC and LT bytes which will need to be flit aligned
2914 if (unlikely(tlen != (hdrsize + pmtu + extra_bytes)))
2916 qp->r_rcv_len += pmtu;
2917 if (unlikely(qp->r_rcv_len > qp->r_len))
2919 rvt_copy_sge(qp, &qp->r_sge, data, pmtu, true, false);
2922 case OP(RDMA_WRITE_LAST_WITH_IMMEDIATE):
2924 ret = rvt_get_rwqe(qp, true);
2932 case OP(SEND_ONLY_WITH_IMMEDIATE):
2933 case OP(SEND_ONLY_WITH_INVALIDATE):
2934 ret = rvt_get_rwqe(qp, false);
2940 if (opcode == OP(SEND_ONLY))
2941 goto no_immediate_data;
2942 if (opcode == OP(SEND_ONLY_WITH_INVALIDATE))
2944 /* FALLTHROUGH -- for SEND_ONLY_WITH_IMMEDIATE */
2945 case OP(SEND_LAST_WITH_IMMEDIATE):
2947 wc.ex.imm_data = ohdr->u.imm_data;
2948 wc.wc_flags = IB_WC_WITH_IMM;
2950 case OP(SEND_LAST_WITH_INVALIDATE):
2952 rkey = be32_to_cpu(ohdr->u.ieth);
2953 if (rvt_invalidate_rkey(qp, rkey))
2954 goto no_immediate_data;
2955 wc.ex.invalidate_rkey = rkey;
2956 wc.wc_flags = IB_WC_WITH_INVALIDATE;
2958 case OP(RDMA_WRITE_LAST):
2959 copy_last = rvt_is_user_qp(qp);
2966 /* Check for invalid length. */
2967 /* LAST len should be >= 1 */
2968 if (unlikely(tlen < (hdrsize + extra_bytes)))
2970 /* Don't count the CRC(and padding and LT byte for 16B). */
2971 tlen -= (hdrsize + extra_bytes);
2972 wc.byte_len = tlen + qp->r_rcv_len;
2973 if (unlikely(wc.byte_len > qp->r_len))
2975 rvt_copy_sge(qp, &qp->r_sge, data, tlen, true, copy_last);
2976 rvt_put_ss(&qp->r_sge);
2978 if (!__test_and_clear_bit(RVT_R_WRID_VALID, &qp->r_aflags))
2980 wc.wr_id = qp->r_wr_id;
2981 wc.status = IB_WC_SUCCESS;
2982 if (opcode == OP(RDMA_WRITE_LAST_WITH_IMMEDIATE) ||
2983 opcode == OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE))
2984 wc.opcode = IB_WC_RECV_RDMA_WITH_IMM;
2986 wc.opcode = IB_WC_RECV;
2988 wc.src_qp = qp->remote_qpn;
2989 wc.slid = rdma_ah_get_dlid(&qp->remote_ah_attr) & U16_MAX;
2991 * It seems that IB mandates the presence of an SL in a
2992 * work completion only for the UD transport (see section
2993 * 11.4.2 of IBTA Vol. 1).
2995 * However, the way the SL is chosen below is consistent
2996 * with the way that IB/qib works and is trying avoid
2997 * introducing incompatibilities.
2999 * See also OPA Vol. 1, section 9.7.6, and table 9-17.
3001 wc.sl = rdma_ah_get_sl(&qp->remote_ah_attr);
3002 /* zero fields that are N/A */
3005 wc.dlid_path_bits = 0;
3007 /* Signal completion event if the solicited bit is set. */
3008 rvt_recv_cq(qp, &wc, ib_bth_is_solicited(ohdr));
3011 case OP(RDMA_WRITE_ONLY):
3012 copy_last = rvt_is_user_qp(qp);
3014 case OP(RDMA_WRITE_FIRST):
3015 case OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE):
3016 if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_WRITE)))
3019 reth = &ohdr->u.rc.reth;
3020 qp->r_len = be32_to_cpu(reth->length);
3022 qp->r_sge.sg_list = NULL;
3023 if (qp->r_len != 0) {
3024 u32 rkey = be32_to_cpu(reth->rkey);
3025 u64 vaddr = get_ib_reth_vaddr(reth);
3028 /* Check rkey & NAK */
3029 ok = rvt_rkey_ok(qp, &qp->r_sge.sge, qp->r_len, vaddr,
3030 rkey, IB_ACCESS_REMOTE_WRITE);
3033 qp->r_sge.num_sge = 1;
3035 qp->r_sge.num_sge = 0;
3036 qp->r_sge.sge.mr = NULL;
3037 qp->r_sge.sge.vaddr = NULL;
3038 qp->r_sge.sge.length = 0;
3039 qp->r_sge.sge.sge_length = 0;
3041 if (opcode == OP(RDMA_WRITE_FIRST))
3043 else if (opcode == OP(RDMA_WRITE_ONLY))
3044 goto no_immediate_data;
3045 ret = rvt_get_rwqe(qp, true);
3049 /* peer will send again */
3050 rvt_put_ss(&qp->r_sge);
3053 wc.ex.imm_data = ohdr->u.rc.imm_data;
3054 wc.wc_flags = IB_WC_WITH_IMM;
3057 case OP(RDMA_READ_REQUEST): {
3058 struct rvt_ack_entry *e;
3062 if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_READ)))
3064 next = qp->r_head_ack_queue + 1;
3065 /* s_ack_queue is size rvt_size_atomic()+1 so use > not >= */
3066 if (next > rvt_size_atomic(ib_to_rvt(qp->ibqp.device)))
3068 spin_lock_irqsave(&qp->s_lock, flags);
3069 if (unlikely(next == qp->s_acked_ack_queue)) {
3070 if (!qp->s_ack_queue[next].sent)
3071 goto nack_inv_unlck;
3072 update_ack_queue(qp, next);
3074 e = &qp->s_ack_queue[qp->r_head_ack_queue];
3075 release_rdma_sge_mr(e);
3076 reth = &ohdr->u.rc.reth;
3077 len = be32_to_cpu(reth->length);
3079 u32 rkey = be32_to_cpu(reth->rkey);
3080 u64 vaddr = get_ib_reth_vaddr(reth);
3083 /* Check rkey & NAK */
3084 ok = rvt_rkey_ok(qp, &e->rdma_sge, len, vaddr,
3085 rkey, IB_ACCESS_REMOTE_READ);
3087 goto nack_acc_unlck;
3089 * Update the next expected PSN. We add 1 later
3090 * below, so only add the remainder here.
3092 qp->r_psn += rvt_div_mtu(qp, len - 1);
3094 e->rdma_sge.mr = NULL;
3095 e->rdma_sge.vaddr = NULL;
3096 e->rdma_sge.length = 0;
3097 e->rdma_sge.sge_length = 0;
3102 e->lpsn = qp->r_psn;
3104 * We need to increment the MSN here instead of when we
3105 * finish sending the result since a duplicate request would
3106 * increment it more than once.
3110 qp->r_state = opcode;
3111 qp->r_nak_state = 0;
3112 qp->r_head_ack_queue = next;
3113 qpriv->r_tid_alloc = qp->r_head_ack_queue;
3115 /* Schedule the send engine. */
3116 qp->s_flags |= RVT_S_RESP_PENDING;
3118 qp->s_flags |= RVT_S_ECN;
3119 hfi1_schedule_send(qp);
3121 spin_unlock_irqrestore(&qp->s_lock, flags);
3125 case OP(COMPARE_SWAP):
3126 case OP(FETCH_ADD): {
3127 struct ib_atomic_eth *ateth = &ohdr->u.atomic_eth;
3128 u64 vaddr = get_ib_ateth_vaddr(ateth);
3129 bool opfn = opcode == OP(COMPARE_SWAP) &&
3130 vaddr == HFI1_VERBS_E_ATOMIC_VADDR;
3131 struct rvt_ack_entry *e;
3137 if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_ATOMIC) &&
3140 next = qp->r_head_ack_queue + 1;
3141 if (next > rvt_size_atomic(ib_to_rvt(qp->ibqp.device)))
3143 spin_lock_irqsave(&qp->s_lock, flags);
3144 if (unlikely(next == qp->s_acked_ack_queue)) {
3145 if (!qp->s_ack_queue[next].sent)
3146 goto nack_inv_unlck;
3147 update_ack_queue(qp, next);
3149 e = &qp->s_ack_queue[qp->r_head_ack_queue];
3150 release_rdma_sge_mr(e);
3151 /* Process OPFN special virtual address */
3153 opfn_conn_response(qp, e, ateth);
3156 if (unlikely(vaddr & (sizeof(u64) - 1)))
3157 goto nack_inv_unlck;
3158 rkey = be32_to_cpu(ateth->rkey);
3159 /* Check rkey & NAK */
3160 if (unlikely(!rvt_rkey_ok(qp, &qp->r_sge.sge, sizeof(u64),
3162 IB_ACCESS_REMOTE_ATOMIC)))
3163 goto nack_acc_unlck;
3164 /* Perform atomic OP and save result. */
3165 maddr = (atomic64_t *)qp->r_sge.sge.vaddr;
3166 sdata = get_ib_ateth_swap(ateth);
3167 e->atomic_data = (opcode == OP(FETCH_ADD)) ?
3168 (u64)atomic64_add_return(sdata, maddr) - sdata :
3169 (u64)cmpxchg((u64 *)qp->r_sge.sge.vaddr,
3170 get_ib_ateth_compare(ateth),
3172 rvt_put_mr(qp->r_sge.sge.mr);
3173 qp->r_sge.num_sge = 0;
3181 qp->r_state = opcode;
3182 qp->r_nak_state = 0;
3183 qp->r_head_ack_queue = next;
3184 qpriv->r_tid_alloc = qp->r_head_ack_queue;
3186 /* Schedule the send engine. */
3187 qp->s_flags |= RVT_S_RESP_PENDING;
3189 qp->s_flags |= RVT_S_ECN;
3190 hfi1_schedule_send(qp);
3192 spin_unlock_irqrestore(&qp->s_lock, flags);
3197 /* NAK unknown opcodes. */
3201 qp->r_state = opcode;
3202 qp->r_ack_psn = psn;
3203 qp->r_nak_state = 0;
3204 /* Send an ACK if requested or required. */
3205 if (psn & IB_BTH_REQ_ACK || fecn) {
3206 if (packet->numpkt == 0 || fecn ||
3207 qp->r_adefered >= HFI1_PSN_CREDIT) {
3212 rc_defered_ack(rcd, qp);
3217 qp->r_nak_state = qp->r_min_rnr_timer | IB_RNR_NAK;
3218 qp->r_ack_psn = qp->r_psn;
3219 /* Queue RNR NAK for later */
3220 rc_defered_ack(rcd, qp);
3224 rvt_rc_error(qp, IB_WC_LOC_QP_OP_ERR);
3225 qp->r_nak_state = IB_NAK_REMOTE_OPERATIONAL_ERROR;
3226 qp->r_ack_psn = qp->r_psn;
3227 /* Queue NAK for later */
3228 rc_defered_ack(rcd, qp);
3232 spin_unlock_irqrestore(&qp->s_lock, flags);
3234 rvt_rc_error(qp, IB_WC_LOC_QP_OP_ERR);
3235 qp->r_nak_state = IB_NAK_INVALID_REQUEST;
3236 qp->r_ack_psn = qp->r_psn;
3237 /* Queue NAK for later */
3238 rc_defered_ack(rcd, qp);
3242 spin_unlock_irqrestore(&qp->s_lock, flags);
3244 rvt_rc_error(qp, IB_WC_LOC_PROT_ERR);
3245 qp->r_nak_state = IB_NAK_REMOTE_ACCESS_ERROR;
3246 qp->r_ack_psn = qp->r_psn;
3248 hfi1_send_rc_ack(packet, fecn);
3251 void hfi1_rc_hdrerr(
3252 struct hfi1_ctxtdata *rcd,
3253 struct hfi1_packet *packet,
3256 struct hfi1_ibport *ibp = rcd_to_iport(rcd);
3261 if (hfi1_ruc_check_hdr(ibp, packet))
3264 psn = ib_bth_get_psn(packet->ohdr);
3265 opcode = ib_bth_get_opcode(packet->ohdr);
3267 /* Only deal with RDMA Writes for now */
3268 if (opcode < IB_OPCODE_RC_RDMA_READ_RESPONSE_FIRST) {
3269 diff = delta_psn(psn, qp->r_psn);
3270 if (!qp->r_nak_state && diff >= 0) {
3271 ibp->rvp.n_rc_seqnak++;
3272 qp->r_nak_state = IB_NAK_PSN_ERROR;
3273 /* Use the expected PSN. */
3274 qp->r_ack_psn = qp->r_psn;
3276 * Wait to send the sequence
3277 * NAK until all packets
3278 * in the receive queue have
3280 * Otherwise, we end up
3281 * propagating congestion.
3283 rc_defered_ack(rcd, qp);
3284 } /* Out of sequence NAK */
3285 } /* QP Request NAKs */