2 * Copyright (c) 2016 Mellanox Technologies Ltd. All rights reserved.
3 * Copyright (c) 2015 System Fabric Works, Inc. All rights reserved.
5 * This software is available to you under a choice of one of two
6 * licenses. You may choose to be licensed under the terms of the GNU
7 * General Public License (GPL) Version 2, available from the file
8 * COPYING in the main directory of this source tree, or the
9 * OpenIB.org BSD license below:
11 * Redistribution and use in source and binary forms, with or
12 * without modification, are permitted provided that the following
15 * - Redistributions of source code must retain the above
16 * copyright notice, this list of conditions and the following
19 * - Redistributions in binary form must reproduce the above
20 * copyright notice, this list of conditions and the following
21 * disclaimer in the documentation and/or other materials
22 * provided with the distribution.
24 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
25 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
26 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
27 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
28 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
29 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
30 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
34 #include <linux/skbuff.h>
38 #include "rxe_queue.h"
40 static int next_opcode(struct rxe_qp *qp, struct rxe_send_wqe *wqe,
43 static inline void retry_first_write_send(struct rxe_qp *qp,
44 struct rxe_send_wqe *wqe,
45 unsigned mask, int npsn)
49 for (i = 0; i < npsn; i++) {
50 int to_send = (wqe->dma.resid > qp->mtu) ?
51 qp->mtu : wqe->dma.resid;
53 qp->req.opcode = next_opcode(qp, wqe,
56 if (wqe->wr.send_flags & IB_SEND_INLINE) {
57 wqe->dma.resid -= to_send;
58 wqe->dma.sge_offset += to_send;
60 advance_dma_data(&wqe->dma, to_send);
62 if (mask & WR_WRITE_MASK)
67 static void req_retry(struct rxe_qp *qp)
69 struct rxe_send_wqe *wqe;
70 unsigned int wqe_index;
75 wqe = queue_head(qp->sq.queue);
76 npsn = (qp->comp.psn - wqe->first_psn) & BTH_PSN_MASK;
78 qp->req.wqe_index = consumer_index(qp->sq.queue);
79 qp->req.psn = qp->comp.psn;
82 for (wqe_index = consumer_index(qp->sq.queue);
83 wqe_index != producer_index(qp->sq.queue);
84 wqe_index = next_index(qp->sq.queue, wqe_index)) {
85 wqe = addr_from_index(qp->sq.queue, wqe_index);
86 mask = wr_opcode_mask(wqe->wr.opcode, qp);
88 if (wqe->state == wqe_state_posted)
91 if (wqe->state == wqe_state_done)
94 wqe->iova = (mask & WR_ATOMIC_MASK) ?
95 wqe->wr.wr.atomic.remote_addr :
96 (mask & WR_READ_OR_WRITE_MASK) ?
97 wqe->wr.wr.rdma.remote_addr :
100 if (!first || (mask & WR_READ_MASK) == 0) {
101 wqe->dma.resid = wqe->dma.length;
102 wqe->dma.cur_sge = 0;
103 wqe->dma.sge_offset = 0;
109 if (mask & WR_WRITE_OR_SEND_MASK)
110 retry_first_write_send(qp, wqe, mask, npsn);
112 if (mask & WR_READ_MASK)
113 wqe->iova += npsn * qp->mtu;
116 wqe->state = wqe_state_posted;
120 void rnr_nak_timer(unsigned long data)
122 struct rxe_qp *qp = (struct rxe_qp *)data;
124 pr_debug("rnr nak timer fired\n");
125 rxe_run_task(&qp->req.task, 1);
128 static struct rxe_send_wqe *req_next_wqe(struct rxe_qp *qp)
130 struct rxe_send_wqe *wqe = queue_head(qp->sq.queue);
133 if (unlikely(qp->req.state == QP_STATE_DRAIN)) {
134 /* check to see if we are drained;
135 * state_lock used by requester and completer
137 spin_lock_irqsave(&qp->state_lock, flags);
139 if (qp->req.state != QP_STATE_DRAIN) {
140 /* comp just finished */
141 spin_unlock_irqrestore(&qp->state_lock,
146 if (wqe && ((qp->req.wqe_index !=
147 consumer_index(qp->sq.queue)) ||
148 (wqe->state != wqe_state_posted))) {
149 /* comp not done yet */
150 spin_unlock_irqrestore(&qp->state_lock,
155 qp->req.state = QP_STATE_DRAINED;
156 spin_unlock_irqrestore(&qp->state_lock, flags);
158 if (qp->ibqp.event_handler) {
161 ev.device = qp->ibqp.device;
162 ev.element.qp = &qp->ibqp;
163 ev.event = IB_EVENT_SQ_DRAINED;
164 qp->ibqp.event_handler(&ev,
165 qp->ibqp.qp_context);
170 if (qp->req.wqe_index == producer_index(qp->sq.queue))
173 wqe = addr_from_index(qp->sq.queue, qp->req.wqe_index);
175 if (unlikely((qp->req.state == QP_STATE_DRAIN ||
176 qp->req.state == QP_STATE_DRAINED) &&
177 (wqe->state != wqe_state_processing)))
180 if (unlikely((wqe->wr.send_flags & IB_SEND_FENCE) &&
181 (qp->req.wqe_index != consumer_index(qp->sq.queue)))) {
182 qp->req.wait_fence = 1;
186 wqe->mask = wr_opcode_mask(wqe->wr.opcode, qp);
190 static int next_opcode_rc(struct rxe_qp *qp, unsigned opcode, int fits)
193 case IB_WR_RDMA_WRITE:
194 if (qp->req.opcode == IB_OPCODE_RC_RDMA_WRITE_FIRST ||
195 qp->req.opcode == IB_OPCODE_RC_RDMA_WRITE_MIDDLE)
197 IB_OPCODE_RC_RDMA_WRITE_LAST :
198 IB_OPCODE_RC_RDMA_WRITE_MIDDLE;
201 IB_OPCODE_RC_RDMA_WRITE_ONLY :
202 IB_OPCODE_RC_RDMA_WRITE_FIRST;
204 case IB_WR_RDMA_WRITE_WITH_IMM:
205 if (qp->req.opcode == IB_OPCODE_RC_RDMA_WRITE_FIRST ||
206 qp->req.opcode == IB_OPCODE_RC_RDMA_WRITE_MIDDLE)
208 IB_OPCODE_RC_RDMA_WRITE_LAST_WITH_IMMEDIATE :
209 IB_OPCODE_RC_RDMA_WRITE_MIDDLE;
212 IB_OPCODE_RC_RDMA_WRITE_ONLY_WITH_IMMEDIATE :
213 IB_OPCODE_RC_RDMA_WRITE_FIRST;
216 if (qp->req.opcode == IB_OPCODE_RC_SEND_FIRST ||
217 qp->req.opcode == IB_OPCODE_RC_SEND_MIDDLE)
219 IB_OPCODE_RC_SEND_LAST :
220 IB_OPCODE_RC_SEND_MIDDLE;
223 IB_OPCODE_RC_SEND_ONLY :
224 IB_OPCODE_RC_SEND_FIRST;
226 case IB_WR_SEND_WITH_IMM:
227 if (qp->req.opcode == IB_OPCODE_RC_SEND_FIRST ||
228 qp->req.opcode == IB_OPCODE_RC_SEND_MIDDLE)
230 IB_OPCODE_RC_SEND_LAST_WITH_IMMEDIATE :
231 IB_OPCODE_RC_SEND_MIDDLE;
234 IB_OPCODE_RC_SEND_ONLY_WITH_IMMEDIATE :
235 IB_OPCODE_RC_SEND_FIRST;
237 case IB_WR_RDMA_READ:
238 return IB_OPCODE_RC_RDMA_READ_REQUEST;
240 case IB_WR_ATOMIC_CMP_AND_SWP:
241 return IB_OPCODE_RC_COMPARE_SWAP;
243 case IB_WR_ATOMIC_FETCH_AND_ADD:
244 return IB_OPCODE_RC_FETCH_ADD;
246 case IB_WR_SEND_WITH_INV:
247 if (qp->req.opcode == IB_OPCODE_RC_SEND_FIRST ||
248 qp->req.opcode == IB_OPCODE_RC_SEND_MIDDLE)
249 return fits ? IB_OPCODE_RC_SEND_LAST_WITH_INVALIDATE :
250 IB_OPCODE_RC_SEND_MIDDLE;
252 return fits ? IB_OPCODE_RC_SEND_ONLY_WITH_INVALIDATE :
253 IB_OPCODE_RC_SEND_FIRST;
255 case IB_WR_LOCAL_INV:
262 static int next_opcode_uc(struct rxe_qp *qp, unsigned opcode, int fits)
265 case IB_WR_RDMA_WRITE:
266 if (qp->req.opcode == IB_OPCODE_UC_RDMA_WRITE_FIRST ||
267 qp->req.opcode == IB_OPCODE_UC_RDMA_WRITE_MIDDLE)
269 IB_OPCODE_UC_RDMA_WRITE_LAST :
270 IB_OPCODE_UC_RDMA_WRITE_MIDDLE;
273 IB_OPCODE_UC_RDMA_WRITE_ONLY :
274 IB_OPCODE_UC_RDMA_WRITE_FIRST;
276 case IB_WR_RDMA_WRITE_WITH_IMM:
277 if (qp->req.opcode == IB_OPCODE_UC_RDMA_WRITE_FIRST ||
278 qp->req.opcode == IB_OPCODE_UC_RDMA_WRITE_MIDDLE)
280 IB_OPCODE_UC_RDMA_WRITE_LAST_WITH_IMMEDIATE :
281 IB_OPCODE_UC_RDMA_WRITE_MIDDLE;
284 IB_OPCODE_UC_RDMA_WRITE_ONLY_WITH_IMMEDIATE :
285 IB_OPCODE_UC_RDMA_WRITE_FIRST;
288 if (qp->req.opcode == IB_OPCODE_UC_SEND_FIRST ||
289 qp->req.opcode == IB_OPCODE_UC_SEND_MIDDLE)
291 IB_OPCODE_UC_SEND_LAST :
292 IB_OPCODE_UC_SEND_MIDDLE;
295 IB_OPCODE_UC_SEND_ONLY :
296 IB_OPCODE_UC_SEND_FIRST;
298 case IB_WR_SEND_WITH_IMM:
299 if (qp->req.opcode == IB_OPCODE_UC_SEND_FIRST ||
300 qp->req.opcode == IB_OPCODE_UC_SEND_MIDDLE)
302 IB_OPCODE_UC_SEND_LAST_WITH_IMMEDIATE :
303 IB_OPCODE_UC_SEND_MIDDLE;
306 IB_OPCODE_UC_SEND_ONLY_WITH_IMMEDIATE :
307 IB_OPCODE_UC_SEND_FIRST;
313 static int next_opcode(struct rxe_qp *qp, struct rxe_send_wqe *wqe,
316 int fits = (wqe->dma.resid <= qp->mtu);
318 switch (qp_type(qp)) {
320 return next_opcode_rc(qp, opcode, fits);
323 return next_opcode_uc(qp, opcode, fits);
330 return IB_OPCODE_UD_SEND_ONLY;
332 case IB_WR_SEND_WITH_IMM:
333 return IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE;
344 static inline int check_init_depth(struct rxe_qp *qp, struct rxe_send_wqe *wqe)
348 if (wqe->has_rd_atomic)
351 qp->req.need_rd_atomic = 1;
352 depth = atomic_dec_return(&qp->req.rd_atomic);
355 qp->req.need_rd_atomic = 0;
356 wqe->has_rd_atomic = 1;
360 atomic_inc(&qp->req.rd_atomic);
364 static inline int get_mtu(struct rxe_qp *qp, struct rxe_send_wqe *wqe)
366 struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
367 struct rxe_port *port;
370 if ((qp_type(qp) == IB_QPT_RC) || (qp_type(qp) == IB_QPT_UC))
376 return port->mtu_cap;
379 static struct sk_buff *init_req_packet(struct rxe_qp *qp,
380 struct rxe_send_wqe *wqe,
381 int opcode, int payload,
382 struct rxe_pkt_info *pkt)
384 struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
385 struct rxe_port *port = &rxe->port;
387 struct rxe_send_wr *ibwr = &wqe->wr;
389 int pad = (-payload) & 0x3;
396 /* length from start of bth to end of icrc */
397 paylen = rxe_opcode[opcode].length + payload + pad + RXE_ICRC_SIZE;
399 /* pkt->hdr, rxe, port_num and mask are initialized in ifc
402 pkt->opcode = opcode;
404 pkt->psn = qp->req.psn;
405 pkt->mask = rxe_opcode[opcode].mask;
406 pkt->paylen = paylen;
411 av = rxe_get_av(pkt);
412 skb = rxe->ifc_ops->init_packet(rxe, av, paylen, pkt);
417 solicited = (ibwr->send_flags & IB_SEND_SOLICITED) &&
418 (pkt->mask & RXE_END_MASK) &&
419 ((pkt->mask & (RXE_SEND_MASK)) ||
420 (pkt->mask & (RXE_WRITE_MASK | RXE_IMMDT_MASK)) ==
421 (RXE_WRITE_MASK | RXE_IMMDT_MASK));
423 pkey = (qp_type(qp) == IB_QPT_GSI) ?
424 port->pkey_tbl[ibwr->wr.ud.pkey_index] :
425 port->pkey_tbl[qp->attr.pkey_index];
427 qp_num = (pkt->mask & RXE_DETH_MASK) ? ibwr->wr.ud.remote_qpn :
428 qp->attr.dest_qp_num;
430 ack_req = ((pkt->mask & RXE_END_MASK) ||
431 (qp->req.noack_pkts++ > RXE_MAX_PKT_PER_ACK));
433 qp->req.noack_pkts = 0;
435 bth_init(pkt, pkt->opcode, solicited, 0, pad, pkey, qp_num,
438 /* init optional headers */
439 if (pkt->mask & RXE_RETH_MASK) {
440 reth_set_rkey(pkt, ibwr->wr.rdma.rkey);
441 reth_set_va(pkt, wqe->iova);
442 reth_set_len(pkt, wqe->dma.length);
445 if (pkt->mask & RXE_IMMDT_MASK)
446 immdt_set_imm(pkt, ibwr->ex.imm_data);
448 if (pkt->mask & RXE_IETH_MASK)
449 ieth_set_rkey(pkt, ibwr->ex.invalidate_rkey);
451 if (pkt->mask & RXE_ATMETH_MASK) {
452 atmeth_set_va(pkt, wqe->iova);
453 if (opcode == IB_OPCODE_RC_COMPARE_SWAP ||
454 opcode == IB_OPCODE_RD_COMPARE_SWAP) {
455 atmeth_set_swap_add(pkt, ibwr->wr.atomic.swap);
456 atmeth_set_comp(pkt, ibwr->wr.atomic.compare_add);
458 atmeth_set_swap_add(pkt, ibwr->wr.atomic.compare_add);
460 atmeth_set_rkey(pkt, ibwr->wr.atomic.rkey);
463 if (pkt->mask & RXE_DETH_MASK) {
464 if (qp->ibqp.qp_num == 1)
465 deth_set_qkey(pkt, GSI_QKEY);
467 deth_set_qkey(pkt, ibwr->wr.ud.remote_qkey);
468 deth_set_sqp(pkt, qp->ibqp.qp_num);
474 static int fill_packet(struct rxe_qp *qp, struct rxe_send_wqe *wqe,
475 struct rxe_pkt_info *pkt, struct sk_buff *skb,
478 struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
483 err = rxe->ifc_ops->prepare(rxe, pkt, skb, &crc);
487 if (pkt->mask & RXE_WRITE_OR_SEND) {
488 if (wqe->wr.send_flags & IB_SEND_INLINE) {
489 u8 *tmp = &wqe->dma.inline_data[wqe->dma.sge_offset];
491 crc = crc32_le(crc, tmp, paylen);
493 memcpy(payload_addr(pkt), tmp, paylen);
495 wqe->dma.resid -= paylen;
496 wqe->dma.sge_offset += paylen;
498 err = copy_data(rxe, qp->pd, 0, &wqe->dma,
499 payload_addr(pkt), paylen,
506 p = payload_addr(pkt) + paylen + bth_pad(pkt);
513 static void update_wqe_state(struct rxe_qp *qp,
514 struct rxe_send_wqe *wqe,
515 struct rxe_pkt_info *pkt)
517 if (pkt->mask & RXE_END_MASK) {
518 if (qp_type(qp) == IB_QPT_RC)
519 wqe->state = wqe_state_pending;
521 wqe->state = wqe_state_processing;
525 static void update_wqe_psn(struct rxe_qp *qp,
526 struct rxe_send_wqe *wqe,
527 struct rxe_pkt_info *pkt,
530 /* number of packets left to send including current one */
531 int num_pkt = (wqe->dma.resid + payload + qp->mtu - 1) / qp->mtu;
533 /* handle zero length packet case */
537 if (pkt->mask & RXE_START_MASK) {
538 wqe->first_psn = qp->req.psn;
539 wqe->last_psn = (qp->req.psn + num_pkt - 1) & BTH_PSN_MASK;
542 if (pkt->mask & RXE_READ_MASK)
543 qp->req.psn = (wqe->first_psn + num_pkt) & BTH_PSN_MASK;
545 qp->req.psn = (qp->req.psn + 1) & BTH_PSN_MASK;
548 static void save_state(struct rxe_send_wqe *wqe,
550 struct rxe_send_wqe *rollback_wqe,
551 struct rxe_qp *rollback_qp)
553 rollback_wqe->state = wqe->state;
554 rollback_wqe->first_psn = wqe->first_psn;
555 rollback_wqe->last_psn = wqe->last_psn;
556 rollback_qp->req.psn = qp->req.psn;
559 static void rollback_state(struct rxe_send_wqe *wqe,
561 struct rxe_send_wqe *rollback_wqe,
562 struct rxe_qp *rollback_qp)
564 wqe->state = rollback_wqe->state;
565 wqe->first_psn = rollback_wqe->first_psn;
566 wqe->last_psn = rollback_wqe->last_psn;
567 qp->req.psn = rollback_qp->req.psn;
570 static void update_state(struct rxe_qp *qp, struct rxe_send_wqe *wqe,
571 struct rxe_pkt_info *pkt, int payload)
573 qp->req.opcode = pkt->opcode;
575 if (pkt->mask & RXE_END_MASK)
576 qp->req.wqe_index = next_index(qp->sq.queue, qp->req.wqe_index);
578 qp->need_req_skb = 0;
580 if (qp->qp_timeout_jiffies && !timer_pending(&qp->retrans_timer))
581 mod_timer(&qp->retrans_timer,
582 jiffies + qp->qp_timeout_jiffies);
585 int rxe_requester(void *arg)
587 struct rxe_qp *qp = (struct rxe_qp *)arg;
588 struct rxe_pkt_info pkt;
590 struct rxe_send_wqe *wqe;
596 struct rxe_qp rollback_qp;
597 struct rxe_send_wqe rollback_wqe;
600 if (unlikely(!qp->valid || qp->req.state == QP_STATE_ERROR))
603 if (unlikely(qp->req.state == QP_STATE_RESET)) {
604 qp->req.wqe_index = consumer_index(qp->sq.queue);
606 qp->req.need_rd_atomic = 0;
607 qp->req.wait_psn = 0;
608 qp->req.need_retry = 0;
612 if (unlikely(qp->req.need_retry)) {
614 qp->req.need_retry = 0;
617 wqe = req_next_wqe(qp);
621 if (wqe->mask & WR_REG_MASK) {
622 if (wqe->wr.opcode == IB_WR_LOCAL_INV) {
623 struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
626 rmr = rxe_pool_get_index(&rxe->mr_pool,
627 wqe->wr.ex.invalidate_rkey >> 8);
629 pr_err("No mr for key %#x\n", wqe->wr.ex.invalidate_rkey);
630 wqe->state = wqe_state_error;
631 wqe->status = IB_WC_MW_BIND_ERR;
634 rmr->state = RXE_MEM_STATE_FREE;
635 wqe->state = wqe_state_done;
636 wqe->status = IB_WC_SUCCESS;
637 } else if (wqe->wr.opcode == IB_WR_REG_MR) {
638 struct rxe_mem *rmr = to_rmr(wqe->wr.wr.reg.mr);
640 rmr->state = RXE_MEM_STATE_VALID;
641 rmr->access = wqe->wr.wr.reg.access;
642 rmr->lkey = wqe->wr.wr.reg.key;
643 rmr->rkey = wqe->wr.wr.reg.key;
644 wqe->state = wqe_state_done;
645 wqe->status = IB_WC_SUCCESS;
649 qp->req.wqe_index = next_index(qp->sq.queue,
654 if (unlikely(qp_type(qp) == IB_QPT_RC &&
655 qp->req.psn > (qp->comp.psn + RXE_MAX_UNACKED_PSNS))) {
656 qp->req.wait_psn = 1;
660 /* Limit the number of inflight SKBs per QP */
661 if (unlikely(atomic_read(&qp->skb_out) >
662 RXE_INFLIGHT_SKBS_PER_QP_HIGH)) {
663 qp->need_req_skb = 1;
667 opcode = next_opcode(qp, wqe, wqe->wr.opcode);
668 if (unlikely(opcode < 0)) {
669 wqe->status = IB_WC_LOC_QP_OP_ERR;
673 mask = rxe_opcode[opcode].mask;
674 if (unlikely(mask & RXE_READ_OR_ATOMIC)) {
675 if (check_init_depth(qp, wqe))
679 mtu = get_mtu(qp, wqe);
680 payload = (mask & RXE_WRITE_OR_SEND) ? wqe->dma.resid : 0;
682 if (qp_type(qp) == IB_QPT_UD) {
683 /* C10-93.1.1: If the total sum of all the buffer lengths specified for a
684 * UD message exceeds the MTU of the port as returned by QueryHCA, the CI
685 * shall not emit any packets for this message. Further, the CI shall not
686 * generate an error due to this condition.
689 /* fake a successful UD send */
690 wqe->first_psn = qp->req.psn;
691 wqe->last_psn = qp->req.psn;
692 qp->req.psn = (qp->req.psn + 1) & BTH_PSN_MASK;
693 qp->req.opcode = IB_OPCODE_UD_SEND_ONLY;
694 qp->req.wqe_index = next_index(qp->sq.queue,
696 wqe->state = wqe_state_done;
697 wqe->status = IB_WC_SUCCESS;
703 skb = init_req_packet(qp, wqe, opcode, payload, &pkt);
704 if (unlikely(!skb)) {
705 pr_err("Failed allocating skb\n");
709 if (fill_packet(qp, wqe, &pkt, skb, payload)) {
710 pr_debug("Error during fill packet\n");
715 * To prevent a race on wqe access between requester and completer,
716 * wqe members state and psn need to be set before calling
718 * Otherwise, completer might initiate an unjustified retry flow.
720 save_state(wqe, qp, &rollback_wqe, &rollback_qp);
721 update_wqe_state(qp, wqe, &pkt);
722 update_wqe_psn(qp, wqe, &pkt, payload);
723 ret = rxe_xmit_packet(to_rdev(qp->ibqp.device), qp, &pkt, skb);
725 qp->need_req_skb = 1;
728 rollback_state(wqe, qp, &rollback_wqe, &rollback_qp);
730 if (ret == -EAGAIN) {
731 rxe_run_task(&qp->req.task, 1);
738 update_state(qp, wqe, &pkt, payload);
744 wqe->status = IB_WC_LOC_PROT_ERR;
745 wqe->state = wqe_state_error;
748 if (qp_type(qp) != IB_QPT_RC) {
749 while (rxe_completer(qp) == 0)