2 * Copyright (c) 2005, 2006 PathScale, Inc. All rights reserved.
4 * This software is available to you under a choice of one of two
5 * licenses. You may choose to be licensed under the terms of the GNU
6 * General Public License (GPL) Version 2, available from the file
7 * COPYING in the main directory of this source tree, or the
8 * OpenIB.org BSD license below:
10 * Redistribution and use in source and binary forms, with or
11 * without modification, are permitted provided that the following
14 * - Redistributions of source code must retain the above
15 * copyright notice, this list of conditions and the following
18 * - Redistributions in binary form must reproduce the above
19 * copyright notice, this list of conditions and the following
20 * disclaimer in the documentation and/or other materials
21 * provided with the distribution.
23 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
33 #include "ipath_verbs.h"
34 #include "ips_common.h"
36 /* cut down ridiculously long IB macro names */
37 #define OP(x) IB_OPCODE_RC_##x
40 * ipath_init_restart- initialize the qp->s_sge after a restart
41 * @qp: the QP who's SGE we're restarting
42 * @wqe: the work queue to initialize the QP's SGE from
44 * The QP s_lock should be held.
46 static void ipath_init_restart(struct ipath_qp *qp, struct ipath_swqe *wqe)
48 struct ipath_ibdev *dev;
51 len = ((qp->s_psn - wqe->psn) & IPS_PSN_MASK) *
52 ib_mtu_enum_to_int(qp->path_mtu);
53 qp->s_sge.sge = wqe->sg_list[0];
54 qp->s_sge.sg_list = wqe->sg_list + 1;
55 qp->s_sge.num_sge = wqe->wr.num_sge;
56 ipath_skip_sge(&qp->s_sge, len);
57 qp->s_len = wqe->length - len;
58 dev = to_idev(qp->ibqp.device);
59 spin_lock(&dev->pending_lock);
60 if (qp->timerwait.next == LIST_POISON1)
61 list_add_tail(&qp->timerwait,
62 &dev->pending[dev->pending_index]);
63 spin_unlock(&dev->pending_lock);
67 * ipath_make_rc_ack - construct a response packet (ACK, NAK, or RDMA read)
68 * @qp: a pointer to the QP
69 * @ohdr: a pointer to the IB header being constructed
72 * Return bth0 if constructed; otherwise, return 0.
73 * Note the QP s_lock must be held.
75 static inline u32 ipath_make_rc_ack(struct ipath_qp *qp,
76 struct ipath_other_headers *ohdr,
79 struct ipath_sge_state *ss;
84 /* header size in 32-bit words LRH+BTH = (8+12)/4. */
88 * Send a response. Note that we are in the responder's
89 * side of the QP context.
91 switch (qp->s_ack_state) {
92 case OP(RDMA_READ_REQUEST):
97 qp->s_ack_state = OP(RDMA_READ_RESPONSE_FIRST);
100 qp->s_ack_state = OP(RDMA_READ_RESPONSE_ONLY);
101 qp->s_rdma_len -= len;
102 bth0 = qp->s_ack_state << 24;
103 ohdr->u.aeth = ipath_compute_aeth(qp);
107 case OP(RDMA_READ_RESPONSE_FIRST):
108 qp->s_ack_state = OP(RDMA_READ_RESPONSE_MIDDLE);
110 case OP(RDMA_READ_RESPONSE_MIDDLE):
111 ss = &qp->s_rdma_sge;
112 len = qp->s_rdma_len;
116 ohdr->u.aeth = ipath_compute_aeth(qp);
118 qp->s_ack_state = OP(RDMA_READ_RESPONSE_LAST);
120 qp->s_rdma_len -= len;
121 bth0 = qp->s_ack_state << 24;
124 case OP(RDMA_READ_RESPONSE_LAST):
125 case OP(RDMA_READ_RESPONSE_ONLY):
127 * We have to prevent new requests from changing
128 * the r_sge state while a ipath_verbs_send()
130 * Changing r_state allows the receiver
131 * to continue processing new packets.
132 * We do it here now instead of above so
133 * that we are sure the packet was sent before
134 * changing the state.
136 qp->r_state = OP(RDMA_READ_RESPONSE_LAST);
137 qp->s_ack_state = OP(ACKNOWLEDGE);
140 case OP(COMPARE_SWAP):
144 qp->r_state = OP(SEND_LAST);
145 qp->s_ack_state = OP(ACKNOWLEDGE);
146 bth0 = IB_OPCODE_ATOMIC_ACKNOWLEDGE << 24;
147 ohdr->u.at.aeth = ipath_compute_aeth(qp);
148 ohdr->u.at.atomic_ack_eth = cpu_to_be64(qp->s_ack_atomic);
149 hwords += sizeof(ohdr->u.at) / 4;
153 /* Send a regular ACK. */
156 qp->s_ack_state = OP(ACKNOWLEDGE);
157 bth0 = qp->s_ack_state << 24;
158 ohdr->u.aeth = ipath_compute_aeth(qp);
161 qp->s_hdrwords = hwords;
163 qp->s_cur_size = len;
169 * ipath_make_rc_req - construct a request packet (SEND, RDMA r/w, ATOMIC)
170 * @qp: a pointer to the QP
171 * @ohdr: a pointer to the IB header being constructed
172 * @pmtu: the path MTU
173 * @bth0p: pointer to the BTH opcode word
174 * @bth2p: pointer to the BTH PSN word
176 * Return 1 if constructed; otherwise, return 0.
177 * Note the QP s_lock must be held.
179 static inline int ipath_make_rc_req(struct ipath_qp *qp,
180 struct ipath_other_headers *ohdr,
181 u32 pmtu, u32 *bth0p, u32 *bth2p)
183 struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
184 struct ipath_sge_state *ss;
185 struct ipath_swqe *wqe;
192 if (!(ib_ipath_state_ops[qp->state] & IPATH_PROCESS_SEND_OK) ||
196 /* header size in 32-bit words LRH+BTH = (8+12)/4. */
200 /* Send a request. */
201 wqe = get_swqe_ptr(qp, qp->s_cur);
202 switch (qp->s_state) {
205 * Resend an old request or start a new one.
207 * We keep track of the current SWQE so that
208 * we don't reset the "furthest progress" state
209 * if we need to back up.
212 if (qp->s_cur == qp->s_tail) {
213 /* Check if send work queue is empty. */
214 if (qp->s_tail == qp->s_head)
216 qp->s_psn = wqe->psn = qp->s_next_psn;
220 * Note that we have to be careful not to modify the
221 * original work request since we may need to resend
224 qp->s_sge.sge = wqe->sg_list[0];
225 qp->s_sge.sg_list = wqe->sg_list + 1;
226 qp->s_sge.num_sge = wqe->wr.num_sge;
227 qp->s_len = len = wqe->length;
230 switch (wqe->wr.opcode) {
232 case IB_WR_SEND_WITH_IMM:
233 /* If no credit, return. */
234 if (qp->s_lsn != (u32) -1 &&
235 ipath_cmp24(wqe->ssn, qp->s_lsn + 1) > 0)
237 wqe->lpsn = wqe->psn;
239 wqe->lpsn += (len - 1) / pmtu;
240 qp->s_state = OP(SEND_FIRST);
244 if (wqe->wr.opcode == IB_WR_SEND)
245 qp->s_state = OP(SEND_ONLY);
247 qp->s_state = OP(SEND_ONLY_WITH_IMMEDIATE);
248 /* Immediate data comes after the BTH */
249 ohdr->u.imm_data = wqe->wr.imm_data;
252 if (wqe->wr.send_flags & IB_SEND_SOLICITED)
254 bth2 = 1 << 31; /* Request ACK. */
255 if (++qp->s_cur == qp->s_size)
259 case IB_WR_RDMA_WRITE:
263 case IB_WR_RDMA_WRITE_WITH_IMM:
264 /* If no credit, return. */
265 if (qp->s_lsn != (u32) -1 &&
266 ipath_cmp24(wqe->ssn, qp->s_lsn + 1) > 0)
268 ohdr->u.rc.reth.vaddr =
269 cpu_to_be64(wqe->wr.wr.rdma.remote_addr);
270 ohdr->u.rc.reth.rkey =
271 cpu_to_be32(wqe->wr.wr.rdma.rkey);
272 ohdr->u.rc.reth.length = cpu_to_be32(len);
273 hwords += sizeof(struct ib_reth) / 4;
274 wqe->lpsn = wqe->psn;
276 wqe->lpsn += (len - 1) / pmtu;
277 qp->s_state = OP(RDMA_WRITE_FIRST);
281 if (wqe->wr.opcode == IB_WR_RDMA_WRITE)
282 qp->s_state = OP(RDMA_WRITE_ONLY);
285 OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE);
286 /* Immediate data comes
288 ohdr->u.rc.imm_data = wqe->wr.imm_data;
290 if (wqe->wr.send_flags & IB_SEND_SOLICITED)
293 bth2 = 1 << 31; /* Request ACK. */
294 if (++qp->s_cur == qp->s_size)
298 case IB_WR_RDMA_READ:
299 ohdr->u.rc.reth.vaddr =
300 cpu_to_be64(wqe->wr.wr.rdma.remote_addr);
301 ohdr->u.rc.reth.rkey =
302 cpu_to_be32(wqe->wr.wr.rdma.rkey);
303 ohdr->u.rc.reth.length = cpu_to_be32(len);
304 qp->s_state = OP(RDMA_READ_REQUEST);
305 hwords += sizeof(ohdr->u.rc.reth) / 4;
309 * Adjust s_next_psn to count the
310 * expected number of responses.
313 qp->s_next_psn += (len - 1) / pmtu;
314 wqe->lpsn = qp->s_next_psn++;
318 if (++qp->s_cur == qp->s_size)
322 case IB_WR_ATOMIC_CMP_AND_SWP:
323 case IB_WR_ATOMIC_FETCH_AND_ADD:
324 if (wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP)
325 qp->s_state = OP(COMPARE_SWAP);
327 qp->s_state = OP(FETCH_ADD);
328 ohdr->u.atomic_eth.vaddr = cpu_to_be64(
329 wqe->wr.wr.atomic.remote_addr);
330 ohdr->u.atomic_eth.rkey = cpu_to_be32(
331 wqe->wr.wr.atomic.rkey);
332 ohdr->u.atomic_eth.swap_data = cpu_to_be64(
333 wqe->wr.wr.atomic.swap);
334 ohdr->u.atomic_eth.compare_data = cpu_to_be64(
335 wqe->wr.wr.atomic.compare_add);
336 hwords += sizeof(struct ib_atomic_eth) / 4;
339 wqe->lpsn = wqe->psn;
341 if (++qp->s_cur == qp->s_size)
352 if (qp->s_tail >= qp->s_size)
355 bth2 |= qp->s_psn++ & IPS_PSN_MASK;
356 if ((int)(qp->s_psn - qp->s_next_psn) > 0)
357 qp->s_next_psn = qp->s_psn;
358 spin_lock(&dev->pending_lock);
359 if (qp->timerwait.next == LIST_POISON1)
360 list_add_tail(&qp->timerwait,
361 &dev->pending[dev->pending_index]);
362 spin_unlock(&dev->pending_lock);
365 case OP(RDMA_READ_RESPONSE_FIRST):
367 * This case can only happen if a send is restarted. See
368 * ipath_restart_rc().
370 ipath_init_restart(qp, wqe);
373 qp->s_state = OP(SEND_MIDDLE);
375 case OP(SEND_MIDDLE):
376 bth2 = qp->s_psn++ & IPS_PSN_MASK;
377 if ((int)(qp->s_psn - qp->s_next_psn) > 0)
378 qp->s_next_psn = qp->s_psn;
383 * Request an ACK every 1/2 MB to avoid retransmit
386 if (((wqe->length - len) % (512 * 1024)) == 0)
391 if (wqe->wr.opcode == IB_WR_SEND)
392 qp->s_state = OP(SEND_LAST);
394 qp->s_state = OP(SEND_LAST_WITH_IMMEDIATE);
395 /* Immediate data comes after the BTH */
396 ohdr->u.imm_data = wqe->wr.imm_data;
399 if (wqe->wr.send_flags & IB_SEND_SOLICITED)
401 bth2 |= 1 << 31; /* Request ACK. */
403 if (qp->s_cur >= qp->s_size)
407 case OP(RDMA_READ_RESPONSE_LAST):
409 * This case can only happen if a RDMA write is restarted.
410 * See ipath_restart_rc().
412 ipath_init_restart(qp, wqe);
414 case OP(RDMA_WRITE_FIRST):
415 qp->s_state = OP(RDMA_WRITE_MIDDLE);
417 case OP(RDMA_WRITE_MIDDLE):
418 bth2 = qp->s_psn++ & IPS_PSN_MASK;
419 if ((int)(qp->s_psn - qp->s_next_psn) > 0)
420 qp->s_next_psn = qp->s_psn;
425 * Request an ACK every 1/2 MB to avoid retransmit
428 if (((wqe->length - len) % (512 * 1024)) == 0)
433 if (wqe->wr.opcode == IB_WR_RDMA_WRITE)
434 qp->s_state = OP(RDMA_WRITE_LAST);
436 qp->s_state = OP(RDMA_WRITE_LAST_WITH_IMMEDIATE);
437 /* Immediate data comes after the BTH */
438 ohdr->u.imm_data = wqe->wr.imm_data;
440 if (wqe->wr.send_flags & IB_SEND_SOLICITED)
443 bth2 |= 1 << 31; /* Request ACK. */
445 if (qp->s_cur >= qp->s_size)
449 case OP(RDMA_READ_RESPONSE_MIDDLE):
451 * This case can only happen if a RDMA read is restarted.
452 * See ipath_restart_rc().
454 ipath_init_restart(qp, wqe);
455 len = ((qp->s_psn - wqe->psn) & IPS_PSN_MASK) * pmtu;
456 ohdr->u.rc.reth.vaddr =
457 cpu_to_be64(wqe->wr.wr.rdma.remote_addr + len);
458 ohdr->u.rc.reth.rkey =
459 cpu_to_be32(wqe->wr.wr.rdma.rkey);
460 ohdr->u.rc.reth.length = cpu_to_be32(qp->s_len);
461 qp->s_state = OP(RDMA_READ_REQUEST);
462 hwords += sizeof(ohdr->u.rc.reth) / 4;
463 bth2 = qp->s_psn++ & IPS_PSN_MASK;
464 if ((int)(qp->s_psn - qp->s_next_psn) > 0)
465 qp->s_next_psn = qp->s_psn;
469 if (qp->s_cur == qp->s_size)
473 case OP(RDMA_READ_REQUEST):
474 case OP(COMPARE_SWAP):
477 * We shouldn't start anything new until this request is
478 * finished. The ACK will handle rescheduling us. XXX The
479 * number of outstanding ones is negotiated at connection
480 * setup time (see pg. 258,289)? XXX Also, if we support
481 * multiple outstanding requests, we need to check the WQE
482 * IB_SEND_FENCE flag and not send a new request if a RDMA
483 * read or atomic is pending.
488 qp->s_hdrwords = hwords;
490 qp->s_cur_size = len;
491 *bth0p = bth0 | (qp->s_state << 24);
499 static inline void ipath_make_rc_grh(struct ipath_qp *qp,
500 struct ib_global_route *grh,
503 struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
505 /* GRH header size in 32-bit words. */
506 qp->s_hdrwords += 10;
507 qp->s_hdr.u.l.grh.version_tclass_flow =
508 cpu_to_be32((6 << 28) |
509 (grh->traffic_class << 20) |
511 qp->s_hdr.u.l.grh.paylen =
512 cpu_to_be16(((qp->s_hdrwords - 12) + nwords +
514 /* next_hdr is defined by C8-7 in ch. 8.4.1 */
515 qp->s_hdr.u.l.grh.next_hdr = 0x1B;
516 qp->s_hdr.u.l.grh.hop_limit = grh->hop_limit;
517 /* The SGID is 32-bit aligned. */
518 qp->s_hdr.u.l.grh.sgid.global.subnet_prefix = dev->gid_prefix;
519 qp->s_hdr.u.l.grh.sgid.global.interface_id =
520 ipath_layer_get_guid(dev->dd);
521 qp->s_hdr.u.l.grh.dgid = grh->dgid;
525 * ipath_do_rc_send - perform a send on an RC QP
526 * @data: contains a pointer to the QP
528 * Process entries in the send work queue until credit or queue is
529 * exhausted. Only allow one CPU to send a packet per QP (tasklet).
530 * Otherwise, after we drop the QP s_lock, two threads could send
531 * packets out of order.
533 void ipath_do_rc_send(unsigned long data)
535 struct ipath_qp *qp = (struct ipath_qp *)data;
536 struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
543 u32 pmtu = ib_mtu_enum_to_int(qp->path_mtu);
544 struct ipath_other_headers *ohdr;
546 if (test_and_set_bit(IPATH_S_BUSY, &qp->s_flags))
549 if (unlikely(qp->remote_ah_attr.dlid ==
550 ipath_layer_get_lid(dev->dd))) {
554 * Pass in an uninitialized ib_wc to be consistent with
555 * other places where ipath_ruc_loopback() is called.
557 ipath_ruc_loopback(qp, &wc);
561 ohdr = &qp->s_hdr.u.oth;
562 if (qp->remote_ah_attr.ah_flags & IB_AH_GRH)
563 ohdr = &qp->s_hdr.u.l.oth;
566 /* Check for a constructed packet to be sent. */
567 if (qp->s_hdrwords != 0) {
569 * If no PIO bufs are available, return. An interrupt will
570 * call ipath_ib_piobufavail() when one is available.
572 _VERBS_INFO("h %u %p\n", qp->s_hdrwords, &qp->s_hdr);
573 _VERBS_INFO("d %u %p %u %p %u %u %u %u\n", qp->s_cur_size,
574 qp->s_cur_sge->sg_list,
575 qp->s_cur_sge->num_sge,
576 qp->s_cur_sge->sge.vaddr,
577 qp->s_cur_sge->sge.sge_length,
578 qp->s_cur_sge->sge.length,
579 qp->s_cur_sge->sge.m,
580 qp->s_cur_sge->sge.n);
581 if (ipath_verbs_send(dev->dd, qp->s_hdrwords,
582 (u32 *) &qp->s_hdr, qp->s_cur_size,
584 ipath_no_bufs_available(qp, dev);
587 dev->n_unicast_xmit++;
588 /* Record that we sent the packet and s_hdr is empty. */
593 * The lock is needed to synchronize between setting
594 * qp->s_ack_state, resend timer, and post_send().
596 spin_lock_irqsave(&qp->s_lock, flags);
598 /* Sending responses has higher priority over sending requests. */
599 if (qp->s_ack_state != OP(ACKNOWLEDGE) &&
600 (bth0 = ipath_make_rc_ack(qp, ohdr, pmtu)) != 0)
601 bth2 = qp->s_ack_psn++ & IPS_PSN_MASK;
602 else if (!ipath_make_rc_req(qp, ohdr, pmtu, &bth0, &bth2))
605 spin_unlock_irqrestore(&qp->s_lock, flags);
607 /* Construct the header. */
608 extra_bytes = (4 - qp->s_cur_size) & 3;
609 nwords = (qp->s_cur_size + extra_bytes) >> 2;
611 if (unlikely(qp->remote_ah_attr.ah_flags & IB_AH_GRH)) {
612 ipath_make_rc_grh(qp, &qp->remote_ah_attr.grh, nwords);
615 lrh0 |= qp->remote_ah_attr.sl << 4;
616 qp->s_hdr.lrh[0] = cpu_to_be16(lrh0);
617 qp->s_hdr.lrh[1] = cpu_to_be16(qp->remote_ah_attr.dlid);
618 qp->s_hdr.lrh[2] = cpu_to_be16(qp->s_hdrwords + nwords +
620 qp->s_hdr.lrh[3] = cpu_to_be16(ipath_layer_get_lid(dev->dd));
621 bth0 |= ipath_layer_get_pkey(dev->dd, qp->s_pkey_index);
622 bth0 |= extra_bytes << 20;
623 ohdr->bth[0] = cpu_to_be32(bth0);
624 ohdr->bth[1] = cpu_to_be32(qp->remote_qpn);
625 ohdr->bth[2] = cpu_to_be32(bth2);
627 /* Check for more work to do. */
631 spin_unlock_irqrestore(&qp->s_lock, flags);
633 clear_bit(IPATH_S_BUSY, &qp->s_flags);
638 static void send_rc_ack(struct ipath_qp *qp)
640 struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
643 struct ipath_other_headers *ohdr;
645 /* Construct the header. */
646 ohdr = &qp->s_hdr.u.oth;
648 /* header size in 32-bit words LRH+BTH+AETH = (8+12+4)/4. */
650 if (unlikely(qp->remote_ah_attr.ah_flags & IB_AH_GRH)) {
651 ipath_make_rc_grh(qp, &qp->remote_ah_attr.grh, 0);
652 ohdr = &qp->s_hdr.u.l.oth;
655 bth0 = ipath_layer_get_pkey(dev->dd, qp->s_pkey_index);
656 ohdr->u.aeth = ipath_compute_aeth(qp);
657 if (qp->s_ack_state >= OP(COMPARE_SWAP)) {
658 bth0 |= IB_OPCODE_ATOMIC_ACKNOWLEDGE << 24;
659 ohdr->u.at.atomic_ack_eth = cpu_to_be64(qp->s_ack_atomic);
660 qp->s_hdrwords += sizeof(ohdr->u.at.atomic_ack_eth) / 4;
663 bth0 |= OP(ACKNOWLEDGE) << 24;
664 lrh0 |= qp->remote_ah_attr.sl << 4;
665 qp->s_hdr.lrh[0] = cpu_to_be16(lrh0);
666 qp->s_hdr.lrh[1] = cpu_to_be16(qp->remote_ah_attr.dlid);
667 qp->s_hdr.lrh[2] = cpu_to_be16(qp->s_hdrwords + SIZE_OF_CRC);
668 qp->s_hdr.lrh[3] = cpu_to_be16(ipath_layer_get_lid(dev->dd));
669 ohdr->bth[0] = cpu_to_be32(bth0);
670 ohdr->bth[1] = cpu_to_be32(qp->remote_qpn);
671 ohdr->bth[2] = cpu_to_be32(qp->s_ack_psn & IPS_PSN_MASK);
674 * If we can send the ACK, clear the ACK state.
676 if (ipath_verbs_send(dev->dd, qp->s_hdrwords, (u32 *) &qp->s_hdr,
678 qp->s_ack_state = OP(ACKNOWLEDGE);
680 dev->n_unicast_xmit++;
685 * ipath_restart_rc - back up requester to resend the last un-ACKed request
686 * @qp: the QP to restart
687 * @psn: packet sequence number for the request
688 * @wc: the work completion request
690 * The QP s_lock should be held.
692 void ipath_restart_rc(struct ipath_qp *qp, u32 psn, struct ib_wc *wc)
694 struct ipath_swqe *wqe = get_swqe_ptr(qp, qp->s_last);
695 struct ipath_ibdev *dev;
699 * If there are no requests pending, we are done.
701 if (ipath_cmp24(psn, qp->s_next_psn) >= 0 ||
702 qp->s_last == qp->s_tail)
705 if (qp->s_retry == 0) {
706 wc->wr_id = wqe->wr.wr_id;
707 wc->status = IB_WC_RETRY_EXC_ERR;
708 wc->opcode = ib_ipath_wc_opcode[wqe->wr.opcode];
711 wc->qp_num = qp->ibqp.qp_num;
712 wc->src_qp = qp->remote_qpn;
714 wc->slid = qp->remote_ah_attr.dlid;
715 wc->sl = qp->remote_ah_attr.sl;
716 wc->dlid_path_bits = 0;
718 ipath_sqerror_qp(qp, wc);
724 * Remove the QP from the timeout queue.
725 * Note: it may already have been removed by ipath_ib_timer().
727 dev = to_idev(qp->ibqp.device);
728 spin_lock(&dev->pending_lock);
729 if (qp->timerwait.next != LIST_POISON1)
730 list_del(&qp->timerwait);
731 spin_unlock(&dev->pending_lock);
733 if (wqe->wr.opcode == IB_WR_RDMA_READ)
736 dev->n_rc_resends += (int)qp->s_psn - (int)psn;
739 * If we are starting the request from the beginning, let the normal
740 * send code handle initialization.
742 qp->s_cur = qp->s_last;
743 if (ipath_cmp24(psn, wqe->psn) <= 0) {
744 qp->s_state = OP(SEND_LAST);
745 qp->s_psn = wqe->psn;
749 if (++n == qp->s_size)
751 if (n == qp->s_tail) {
752 if (ipath_cmp24(psn, qp->s_next_psn) >= 0) {
754 wqe = get_swqe_ptr(qp, n);
758 wqe = get_swqe_ptr(qp, n);
759 if (ipath_cmp24(psn, wqe->psn) < 0)
766 * Reset the state to restart in the middle of a request.
767 * Don't change the s_sge, s_cur_sge, or s_cur_size.
768 * See ipath_do_rc_send().
770 switch (wqe->wr.opcode) {
772 case IB_WR_SEND_WITH_IMM:
773 qp->s_state = OP(RDMA_READ_RESPONSE_FIRST);
776 case IB_WR_RDMA_WRITE:
777 case IB_WR_RDMA_WRITE_WITH_IMM:
778 qp->s_state = OP(RDMA_READ_RESPONSE_LAST);
781 case IB_WR_RDMA_READ:
783 OP(RDMA_READ_RESPONSE_MIDDLE);
788 * This case shouldn't happen since its only
791 qp->s_state = OP(SEND_LAST);
796 tasklet_hi_schedule(&qp->s_task);
803 * reset_psn - reset the QP state to send starting from PSN
805 * @psn: the packet sequence number to restart at
807 * This is called from ipath_rc_rcv() to process an incoming RC ACK
809 * Called at interrupt level with the QP s_lock held.
811 static void reset_psn(struct ipath_qp *qp, u32 psn)
813 struct ipath_swqe *wqe;
817 wqe = get_swqe_ptr(qp, n);
819 if (++n == qp->s_size)
821 if (n == qp->s_tail) {
822 if (ipath_cmp24(psn, qp->s_next_psn) >= 0) {
824 wqe = get_swqe_ptr(qp, n);
828 wqe = get_swqe_ptr(qp, n);
829 if (ipath_cmp24(psn, wqe->psn) < 0)
836 * Set the state to restart in the middle of a
837 * request. Don't change the s_sge, s_cur_sge, or
838 * s_cur_size. See ipath_do_rc_send().
840 switch (wqe->wr.opcode) {
842 case IB_WR_SEND_WITH_IMM:
843 qp->s_state = OP(RDMA_READ_RESPONSE_FIRST);
846 case IB_WR_RDMA_WRITE:
847 case IB_WR_RDMA_WRITE_WITH_IMM:
848 qp->s_state = OP(RDMA_READ_RESPONSE_LAST);
851 case IB_WR_RDMA_READ:
852 qp->s_state = OP(RDMA_READ_RESPONSE_MIDDLE);
857 * This case shouldn't happen since its only
860 qp->s_state = OP(SEND_LAST);
865 * do_rc_ack - process an incoming RC ACK
866 * @qp: the QP the ACK came in on
867 * @psn: the packet sequence number of the ACK
868 * @opcode: the opcode of the request that resulted in the ACK
870 * This is called from ipath_rc_rcv() to process an incoming RC ACK
872 * Called at interrupt level with the QP s_lock held.
873 * Returns 1 if OK, 0 if current operation should be aborted (NAK).
875 static int do_rc_ack(struct ipath_qp *qp, u32 aeth, u32 psn, int opcode)
877 struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
879 struct ipath_swqe *wqe;
883 * Remove the QP from the timeout queue (or RNR timeout queue).
884 * If ipath_ib_timer() has already removed it,
885 * it's OK since we hold the QP s_lock and ipath_restart_rc()
886 * just won't find anything to restart if we ACK everything.
888 spin_lock(&dev->pending_lock);
889 if (qp->timerwait.next != LIST_POISON1)
890 list_del(&qp->timerwait);
891 spin_unlock(&dev->pending_lock);
894 * Note that NAKs implicitly ACK outstanding SEND and RDMA write
895 * requests and implicitly NAK RDMA read and atomic requests issued
896 * before the NAK'ed request. The MSN won't include the NAK'ed
897 * request but will include an ACK'ed request(s).
899 wqe = get_swqe_ptr(qp, qp->s_last);
901 /* Nothing is pending to ACK/NAK. */
902 if (qp->s_last == qp->s_tail)
906 * The MSN might be for a later WQE than the PSN indicates so
907 * only complete WQEs that the PSN finishes.
909 while (ipath_cmp24(psn, wqe->lpsn) >= 0) {
910 /* If we are ACKing a WQE, the MSN should be >= the SSN. */
911 if (ipath_cmp24(aeth, wqe->ssn) < 0)
914 * If this request is a RDMA read or atomic, and the ACK is
915 * for a later operation, this ACK NAKs the RDMA read or
916 * atomic. In other words, only a RDMA_READ_LAST or ONLY
917 * can ACK a RDMA read and likewise for atomic ops. Note
918 * that the NAK case can only happen if relaxed ordering is
919 * used and requests are sent after an RDMA read or atomic
920 * is sent but before the response is received.
922 if ((wqe->wr.opcode == IB_WR_RDMA_READ &&
923 opcode != OP(RDMA_READ_RESPONSE_LAST)) ||
924 ((wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
925 wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD) &&
926 (opcode != OP(ATOMIC_ACKNOWLEDGE) ||
927 ipath_cmp24(wqe->psn, psn) != 0))) {
929 * The last valid PSN seen is the previous
932 qp->s_last_psn = wqe->psn - 1;
933 /* Retry this request. */
934 ipath_restart_rc(qp, wqe->psn, &wc);
936 * No need to process the ACK/NAK since we are
937 * restarting an earlier request.
941 /* Post a send completion queue entry if requested. */
942 if (!test_bit(IPATH_S_SIGNAL_REQ_WR, &qp->s_flags) ||
943 (wqe->wr.send_flags & IB_SEND_SIGNALED)) {
944 wc.wr_id = wqe->wr.wr_id;
945 wc.status = IB_WC_SUCCESS;
946 wc.opcode = ib_ipath_wc_opcode[wqe->wr.opcode];
948 wc.byte_len = wqe->length;
949 wc.qp_num = qp->ibqp.qp_num;
950 wc.src_qp = qp->remote_qpn;
952 wc.slid = qp->remote_ah_attr.dlid;
953 wc.sl = qp->remote_ah_attr.sl;
954 wc.dlid_path_bits = 0;
956 ipath_cq_enter(to_icq(qp->ibqp.send_cq), &wc, 0);
958 qp->s_retry = qp->s_retry_cnt;
960 * If we are completing a request which is in the process of
961 * being resent, we can stop resending it since we know the
962 * responder has already seen it.
964 if (qp->s_last == qp->s_cur) {
965 if (++qp->s_cur >= qp->s_size)
967 wqe = get_swqe_ptr(qp, qp->s_cur);
968 qp->s_state = OP(SEND_LAST);
969 qp->s_psn = wqe->psn;
971 if (++qp->s_last >= qp->s_size)
973 wqe = get_swqe_ptr(qp, qp->s_last);
974 if (qp->s_last == qp->s_tail)
978 switch (aeth >> 29) {
981 /* If this is a partial ACK, reset the retransmit timer. */
982 if (qp->s_last != qp->s_tail) {
983 spin_lock(&dev->pending_lock);
984 list_add_tail(&qp->timerwait,
985 &dev->pending[dev->pending_index]);
986 spin_unlock(&dev->pending_lock);
988 ipath_get_credit(qp, aeth);
989 qp->s_rnr_retry = qp->s_rnr_retry_cnt;
990 qp->s_retry = qp->s_retry_cnt;
991 qp->s_last_psn = psn;
995 case 1: /* RNR NAK */
997 if (qp->s_rnr_retry == 0) {
998 if (qp->s_last == qp->s_tail)
1001 wc.status = IB_WC_RNR_RETRY_EXC_ERR;
1004 if (qp->s_rnr_retry_cnt < 7)
1006 if (qp->s_last == qp->s_tail)
1009 /* The last valid PSN seen is the previous request's. */
1010 qp->s_last_psn = wqe->psn - 1;
1012 dev->n_rc_resends += (int)qp->s_psn - (int)psn;
1015 * If we are starting the request from the beginning, let
1016 * the normal send code handle initialization.
1018 qp->s_cur = qp->s_last;
1019 wqe = get_swqe_ptr(qp, qp->s_cur);
1020 if (ipath_cmp24(psn, wqe->psn) <= 0) {
1021 qp->s_state = OP(SEND_LAST);
1022 qp->s_psn = wqe->psn;
1027 ib_ipath_rnr_table[(aeth >> IPS_AETH_CREDIT_SHIFT) &
1028 IPS_AETH_CREDIT_MASK];
1029 ipath_insert_rnr_queue(qp);
1033 /* The last valid PSN seen is the previous request's. */
1034 if (qp->s_last != qp->s_tail)
1035 qp->s_last_psn = wqe->psn - 1;
1036 switch ((aeth >> IPS_AETH_CREDIT_SHIFT) &
1037 IPS_AETH_CREDIT_MASK) {
1038 case 0: /* PSN sequence error */
1041 * Back up to the responder's expected PSN. XXX
1042 * Note that we might get a NAK in the middle of an
1043 * RDMA READ response which terminates the RDMA
1046 if (qp->s_last == qp->s_tail)
1049 if (ipath_cmp24(psn, wqe->psn) < 0)
1052 /* Retry the request. */
1053 ipath_restart_rc(qp, psn, &wc);
1056 case 1: /* Invalid Request */
1057 wc.status = IB_WC_REM_INV_REQ_ERR;
1058 dev->n_other_naks++;
1061 case 2: /* Remote Access Error */
1062 wc.status = IB_WC_REM_ACCESS_ERR;
1063 dev->n_other_naks++;
1066 case 3: /* Remote Operation Error */
1067 wc.status = IB_WC_REM_OP_ERR;
1068 dev->n_other_naks++;
1070 wc.wr_id = wqe->wr.wr_id;
1071 wc.opcode = ib_ipath_wc_opcode[wqe->wr.opcode];
1074 wc.qp_num = qp->ibqp.qp_num;
1075 wc.src_qp = qp->remote_qpn;
1077 wc.slid = qp->remote_ah_attr.dlid;
1078 wc.sl = qp->remote_ah_attr.sl;
1079 wc.dlid_path_bits = 0;
1081 ipath_sqerror_qp(qp, &wc);
1085 /* Ignore other reserved NAK error codes */
1088 qp->s_rnr_retry = qp->s_rnr_retry_cnt;
1091 default: /* 2: reserved */
1093 /* Ignore reserved NAK codes. */
1102 * ipath_rc_rcv_resp - process an incoming RC response packet
1103 * @dev: the device this packet came in on
1104 * @ohdr: the other headers for this packet
1105 * @data: the packet data
1106 * @tlen: the packet length
1107 * @qp: the QP for this packet
1108 * @opcode: the opcode for this packet
1109 * @psn: the packet sequence number for this packet
1110 * @hdrsize: the header length
1111 * @pmtu: the path MTU
1112 * @header_in_data: true if part of the header data is in the data buffer
1114 * This is called from ipath_rc_rcv() to process an incoming RC response
1115 * packet for the given QP.
1116 * Called at interrupt level.
1118 static inline void ipath_rc_rcv_resp(struct ipath_ibdev *dev,
1119 struct ipath_other_headers *ohdr,
1120 void *data, u32 tlen,
1121 struct ipath_qp *qp,
1123 u32 psn, u32 hdrsize, u32 pmtu,
1126 unsigned long flags;
1132 spin_lock_irqsave(&qp->s_lock, flags);
1134 /* Ignore invalid responses. */
1135 if (ipath_cmp24(psn, qp->s_next_psn) >= 0)
1138 /* Ignore duplicate responses. */
1139 diff = ipath_cmp24(psn, qp->s_last_psn);
1140 if (unlikely(diff <= 0)) {
1141 /* Update credits for "ghost" ACKs */
1142 if (diff == 0 && opcode == OP(ACKNOWLEDGE)) {
1143 if (!header_in_data)
1144 aeth = be32_to_cpu(ohdr->u.aeth);
1146 aeth = be32_to_cpu(((__be32 *) data)[0]);
1147 data += sizeof(__be32);
1149 if ((aeth >> 29) == 0)
1150 ipath_get_credit(qp, aeth);
1156 case OP(ACKNOWLEDGE):
1157 case OP(ATOMIC_ACKNOWLEDGE):
1158 case OP(RDMA_READ_RESPONSE_FIRST):
1159 if (!header_in_data)
1160 aeth = be32_to_cpu(ohdr->u.aeth);
1162 aeth = be32_to_cpu(((__be32 *) data)[0]);
1163 data += sizeof(__be32);
1165 if (opcode == OP(ATOMIC_ACKNOWLEDGE))
1166 *(u64 *) qp->s_sge.sge.vaddr = *(u64 *) data;
1167 if (!do_rc_ack(qp, aeth, psn, opcode) ||
1168 opcode != OP(RDMA_READ_RESPONSE_FIRST))
1172 * do_rc_ack() has already checked the PSN so skip
1173 * the sequence check.
1177 case OP(RDMA_READ_RESPONSE_MIDDLE):
1178 /* no AETH, no ACK */
1179 if (unlikely(ipath_cmp24(psn, qp->s_last_psn + 1))) {
1181 ipath_restart_rc(qp, qp->s_last_psn + 1, &wc);
1185 if (unlikely(qp->s_state != OP(RDMA_READ_REQUEST)))
1187 if (unlikely(tlen != (hdrsize + pmtu + 4)))
1189 if (unlikely(pmtu >= qp->s_len))
1191 /* We got a response so update the timeout. */
1192 if (unlikely(qp->s_last == qp->s_tail ||
1193 get_swqe_ptr(qp, qp->s_last)->wr.opcode !=
1196 spin_lock(&dev->pending_lock);
1197 if (qp->s_rnr_timeout == 0 &&
1198 qp->timerwait.next != LIST_POISON1)
1199 list_move_tail(&qp->timerwait,
1200 &dev->pending[dev->pending_index]);
1201 spin_unlock(&dev->pending_lock);
1203 * Update the RDMA receive state but do the copy w/o holding the
1204 * locks and blocking interrupts. XXX Yet another place that
1205 * affects relaxed RDMA order since we don't want s_sge modified.
1208 qp->s_last_psn = psn;
1209 spin_unlock_irqrestore(&qp->s_lock, flags);
1210 ipath_copy_sge(&qp->s_sge, data, pmtu);
1213 case OP(RDMA_READ_RESPONSE_LAST):
1214 /* ACKs READ req. */
1215 if (unlikely(ipath_cmp24(psn, qp->s_last_psn + 1))) {
1217 ipath_restart_rc(qp, qp->s_last_psn + 1, &wc);
1221 case OP(RDMA_READ_RESPONSE_ONLY):
1222 if (unlikely(qp->s_state != OP(RDMA_READ_REQUEST)))
1225 * Get the number of bytes the message was padded by.
1227 pad = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3;
1229 * Check that the data size is >= 1 && <= pmtu.
1230 * Remember to account for the AETH header (4) and
1233 if (unlikely(tlen <= (hdrsize + pad + 8))) {
1235 * XXX Need to generate an error CQ
1240 tlen -= hdrsize + pad + 8;
1241 if (unlikely(tlen != qp->s_len)) {
1243 * XXX Need to generate an error CQ
1248 if (!header_in_data)
1249 aeth = be32_to_cpu(ohdr->u.aeth);
1251 aeth = be32_to_cpu(((__be32 *) data)[0]);
1252 data += sizeof(__be32);
1254 ipath_copy_sge(&qp->s_sge, data, tlen);
1255 if (do_rc_ack(qp, aeth, psn, OP(RDMA_READ_RESPONSE_LAST))) {
1257 * Change the state so we contimue
1258 * processing new requests.
1260 qp->s_state = OP(SEND_LAST);
1266 spin_unlock_irqrestore(&qp->s_lock, flags);
1272 * ipath_rc_rcv_error - process an incoming duplicate or error RC packet
1273 * @dev: the device this packet came in on
1274 * @ohdr: the other headers for this packet
1275 * @data: the packet data
1276 * @qp: the QP for this packet
1277 * @opcode: the opcode for this packet
1278 * @psn: the packet sequence number for this packet
1279 * @diff: the difference between the PSN and the expected PSN
1280 * @header_in_data: true if part of the header data is in the data buffer
1282 * This is called from ipath_rc_rcv() to process an unexpected
1283 * incoming RC packet for the given QP.
1284 * Called at interrupt level.
1285 * Return 1 if no more processing is needed; otherwise return 0 to
1286 * schedule a response to be sent and the s_lock unlocked.
1288 static inline int ipath_rc_rcv_error(struct ipath_ibdev *dev,
1289 struct ipath_other_headers *ohdr,
1291 struct ipath_qp *qp,
1297 struct ib_reth *reth;
1301 * Packet sequence error.
1302 * A NAK will ACK earlier sends and RDMA writes.
1303 * Don't queue the NAK if a RDMA read, atomic, or
1304 * NAK is pending though.
1306 spin_lock(&qp->s_lock);
1307 if ((qp->s_ack_state >= OP(RDMA_READ_REQUEST) &&
1308 qp->s_ack_state != IB_OPCODE_ACKNOWLEDGE) ||
1309 qp->s_nak_state != 0) {
1310 spin_unlock(&qp->s_lock);
1313 qp->s_ack_state = OP(SEND_ONLY);
1314 qp->s_nak_state = IB_NAK_PSN_ERROR;
1315 /* Use the expected PSN. */
1316 qp->s_ack_psn = qp->r_psn;
1321 * Handle a duplicate request. Don't re-execute SEND, RDMA
1322 * write or atomic op. Don't NAK errors, just silently drop
1323 * the duplicate request. Note that r_sge, r_len, and
1324 * r_rcv_len may be in use so don't modify them.
1326 * We are supposed to ACK the earliest duplicate PSN but we
1327 * can coalesce an outstanding duplicate ACK. We have to
1328 * send the earliest so that RDMA reads can be restarted at
1329 * the requester's expected PSN.
1331 spin_lock(&qp->s_lock);
1332 if (qp->s_ack_state != IB_OPCODE_ACKNOWLEDGE &&
1333 ipath_cmp24(psn, qp->s_ack_psn) >= 0) {
1334 if (qp->s_ack_state < IB_OPCODE_RDMA_READ_REQUEST)
1335 qp->s_ack_psn = psn;
1336 spin_unlock(&qp->s_lock);
1340 case OP(RDMA_READ_REQUEST):
1342 * We have to be careful to not change s_rdma_sge
1343 * while ipath_do_rc_send() is using it and not
1344 * holding the s_lock.
1346 if (qp->s_ack_state != OP(ACKNOWLEDGE) &&
1347 qp->s_ack_state >= IB_OPCODE_RDMA_READ_REQUEST) {
1348 spin_unlock(&qp->s_lock);
1349 dev->n_rdma_dup_busy++;
1352 /* RETH comes after BTH */
1353 if (!header_in_data)
1354 reth = &ohdr->u.rc.reth;
1356 reth = (struct ib_reth *)data;
1357 data += sizeof(*reth);
1359 qp->s_rdma_len = be32_to_cpu(reth->length);
1360 if (qp->s_rdma_len != 0) {
1361 u32 rkey = be32_to_cpu(reth->rkey);
1362 u64 vaddr = be64_to_cpu(reth->vaddr);
1366 * Address range must be a subset of the original
1367 * request and start on pmtu boundaries.
1369 ok = ipath_rkey_ok(dev, &qp->s_rdma_sge,
1370 qp->s_rdma_len, vaddr, rkey,
1371 IB_ACCESS_REMOTE_READ);
1375 qp->s_rdma_sge.sg_list = NULL;
1376 qp->s_rdma_sge.num_sge = 0;
1377 qp->s_rdma_sge.sge.mr = NULL;
1378 qp->s_rdma_sge.sge.vaddr = NULL;
1379 qp->s_rdma_sge.sge.length = 0;
1380 qp->s_rdma_sge.sge.sge_length = 0;
1384 case OP(COMPARE_SWAP):
1387 * Check for the PSN of the last atomic operations
1388 * performed and resend the result if found.
1390 if ((psn & IPS_PSN_MASK) != qp->r_atomic_psn) {
1391 spin_unlock(&qp->s_lock);
1394 qp->s_ack_atomic = qp->r_atomic_data;
1397 qp->s_ack_state = opcode;
1398 qp->s_nak_state = 0;
1399 qp->s_ack_psn = psn;
1408 * ipath_rc_rcv - process an incoming RC packet
1409 * @dev: the device this packet came in on
1410 * @hdr: the header of this packet
1411 * @has_grh: true if the header has a GRH
1412 * @data: the packet data
1413 * @tlen: the packet length
1414 * @qp: the QP for this packet
1416 * This is called from ipath_qp_rcv() to process an incoming RC packet
1418 * Called at interrupt level.
1420 void ipath_rc_rcv(struct ipath_ibdev *dev, struct ipath_ib_header *hdr,
1421 int has_grh, void *data, u32 tlen, struct ipath_qp *qp)
1423 struct ipath_other_headers *ohdr;
1428 unsigned long flags;
1430 u32 pmtu = ib_mtu_enum_to_int(qp->path_mtu);
1432 struct ib_reth *reth;
1438 hdrsize = 8 + 12; /* LRH + BTH */
1439 psn = be32_to_cpu(ohdr->bth[2]);
1442 ohdr = &hdr->u.l.oth;
1443 hdrsize = 8 + 40 + 12; /* LRH + GRH + BTH */
1445 * The header with GRH is 60 bytes and the core driver sets
1446 * the eager header buffer size to 56 bytes so the last 4
1447 * bytes of the BTH header (PSN) is in the data buffer.
1450 ipath_layer_get_rcvhdrentsize(dev->dd) == 16;
1451 if (header_in_data) {
1452 psn = be32_to_cpu(((__be32 *) data)[0]);
1453 data += sizeof(__be32);
1455 psn = be32_to_cpu(ohdr->bth[2]);
1458 * The opcode is in the low byte when its in network order
1459 * (top byte when in host order).
1461 opcode = be32_to_cpu(ohdr->bth[0]) >> 24;
1464 * Process responses (ACKs) before anything else. Note that the
1465 * packet sequence number will be for something in the send work
1466 * queue rather than the expected receive packet sequence number.
1467 * In other words, this QP is the requester.
1469 if (opcode >= OP(RDMA_READ_RESPONSE_FIRST) &&
1470 opcode <= OP(ATOMIC_ACKNOWLEDGE)) {
1471 ipath_rc_rcv_resp(dev, ohdr, data, tlen, qp, opcode, psn,
1472 hdrsize, pmtu, header_in_data);
1476 spin_lock_irqsave(&qp->r_rq.lock, flags);
1478 /* Compute 24 bits worth of difference. */
1479 diff = ipath_cmp24(psn, qp->r_psn);
1480 if (unlikely(diff)) {
1481 if (ipath_rc_rcv_error(dev, ohdr, data, qp, opcode,
1482 psn, diff, header_in_data))
1487 /* Check for opcode sequence errors. */
1488 switch (qp->r_state) {
1489 case OP(SEND_FIRST):
1490 case OP(SEND_MIDDLE):
1491 if (opcode == OP(SEND_MIDDLE) ||
1492 opcode == OP(SEND_LAST) ||
1493 opcode == OP(SEND_LAST_WITH_IMMEDIATE))
1497 * A NAK will ACK earlier sends and RDMA writes. Don't queue the
1498 * NAK if a RDMA read, atomic, or NAK is pending though.
1500 spin_lock(&qp->s_lock);
1501 if (qp->s_ack_state >= OP(RDMA_READ_REQUEST) &&
1502 qp->s_ack_state != IB_OPCODE_ACKNOWLEDGE) {
1503 spin_unlock(&qp->s_lock);
1506 /* XXX Flush WQEs */
1507 qp->state = IB_QPS_ERR;
1508 qp->s_ack_state = OP(SEND_ONLY);
1509 qp->s_nak_state = IB_NAK_INVALID_REQUEST;
1510 qp->s_ack_psn = qp->r_psn;
1513 case OP(RDMA_WRITE_FIRST):
1514 case OP(RDMA_WRITE_MIDDLE):
1515 if (opcode == OP(RDMA_WRITE_MIDDLE) ||
1516 opcode == OP(RDMA_WRITE_LAST) ||
1517 opcode == OP(RDMA_WRITE_LAST_WITH_IMMEDIATE))
1521 case OP(RDMA_READ_REQUEST):
1522 case OP(COMPARE_SWAP):
1525 * Drop all new requests until a response has been sent. A
1526 * new request then ACKs the RDMA response we sent. Relaxed
1527 * ordering would allow new requests to be processed but we
1528 * would need to keep a queue of rwqe's for all that are in
1529 * progress. Note that we can't RNR NAK this request since
1530 * the RDMA READ or atomic response is already queued to be
1531 * sent (unless we implement a response send queue).
1536 if (opcode == OP(SEND_MIDDLE) ||
1537 opcode == OP(SEND_LAST) ||
1538 opcode == OP(SEND_LAST_WITH_IMMEDIATE) ||
1539 opcode == OP(RDMA_WRITE_MIDDLE) ||
1540 opcode == OP(RDMA_WRITE_LAST) ||
1541 opcode == OP(RDMA_WRITE_LAST_WITH_IMMEDIATE))
1549 /* OK, process the packet. */
1551 case OP(SEND_FIRST):
1552 if (!ipath_get_rwqe(qp, 0)) {
1555 * A RNR NAK will ACK earlier sends and RDMA writes.
1556 * Don't queue the NAK if a RDMA read or atomic
1557 * is pending though.
1559 spin_lock(&qp->s_lock);
1560 if (qp->s_ack_state >=
1561 OP(RDMA_READ_REQUEST) &&
1562 qp->s_ack_state != IB_OPCODE_ACKNOWLEDGE) {
1563 spin_unlock(&qp->s_lock);
1566 qp->s_ack_state = OP(SEND_ONLY);
1567 qp->s_nak_state = IB_RNR_NAK | qp->s_min_rnr_timer;
1568 qp->s_ack_psn = qp->r_psn;
1573 case OP(SEND_MIDDLE):
1574 case OP(RDMA_WRITE_MIDDLE):
1576 /* Check for invalid length PMTU or posted rwqe len. */
1577 if (unlikely(tlen != (hdrsize + pmtu + 4)))
1579 qp->r_rcv_len += pmtu;
1580 if (unlikely(qp->r_rcv_len > qp->r_len))
1582 ipath_copy_sge(&qp->r_sge, data, pmtu);
1585 case OP(RDMA_WRITE_LAST_WITH_IMMEDIATE):
1587 if (!ipath_get_rwqe(qp, 1))
1592 case OP(SEND_ONLY_WITH_IMMEDIATE):
1593 if (!ipath_get_rwqe(qp, 0))
1596 if (opcode == OP(SEND_ONLY))
1599 case OP(SEND_LAST_WITH_IMMEDIATE):
1601 if (header_in_data) {
1602 wc.imm_data = *(__be32 *) data;
1603 data += sizeof(__be32);
1605 /* Immediate data comes after BTH */
1606 wc.imm_data = ohdr->u.imm_data;
1609 wc.wc_flags = IB_WC_WITH_IMM;
1612 case OP(RDMA_WRITE_LAST):
1614 /* Get the number of bytes the message was padded by. */
1615 pad = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3;
1616 /* Check for invalid length. */
1617 /* XXX LAST len should be >= 1 */
1618 if (unlikely(tlen < (hdrsize + pad + 4)))
1620 /* Don't count the CRC. */
1621 tlen -= (hdrsize + pad + 4);
1622 wc.byte_len = tlen + qp->r_rcv_len;
1623 if (unlikely(wc.byte_len > qp->r_len))
1625 ipath_copy_sge(&qp->r_sge, data, tlen);
1626 atomic_inc(&qp->msn);
1627 if (opcode == OP(RDMA_WRITE_LAST) ||
1628 opcode == OP(RDMA_WRITE_ONLY))
1630 wc.wr_id = qp->r_wr_id;
1631 wc.status = IB_WC_SUCCESS;
1632 wc.opcode = IB_WC_RECV;
1634 wc.qp_num = qp->ibqp.qp_num;
1635 wc.src_qp = qp->remote_qpn;
1637 wc.slid = qp->remote_ah_attr.dlid;
1638 wc.sl = qp->remote_ah_attr.sl;
1639 wc.dlid_path_bits = 0;
1641 /* Signal completion event if the solicited bit is set. */
1642 ipath_cq_enter(to_icq(qp->ibqp.recv_cq), &wc,
1644 __constant_cpu_to_be32(1 << 23)) != 0);
1647 case OP(RDMA_WRITE_FIRST):
1648 case OP(RDMA_WRITE_ONLY):
1649 case OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE):
1651 /* RETH comes after BTH */
1652 if (!header_in_data)
1653 reth = &ohdr->u.rc.reth;
1655 reth = (struct ib_reth *)data;
1656 data += sizeof(*reth);
1658 hdrsize += sizeof(*reth);
1659 qp->r_len = be32_to_cpu(reth->length);
1661 if (qp->r_len != 0) {
1662 u32 rkey = be32_to_cpu(reth->rkey);
1663 u64 vaddr = be64_to_cpu(reth->vaddr);
1666 /* Check rkey & NAK */
1667 ok = ipath_rkey_ok(dev, &qp->r_sge,
1668 qp->r_len, vaddr, rkey,
1669 IB_ACCESS_REMOTE_WRITE);
1670 if (unlikely(!ok)) {
1673 * A NAK will ACK earlier sends and RDMA
1674 * writes. Don't queue the NAK if a RDMA
1675 * read, atomic, or NAK is pending though.
1677 spin_lock(&qp->s_lock);
1678 if (qp->s_ack_state >=
1679 OP(RDMA_READ_REQUEST) &&
1681 IB_OPCODE_ACKNOWLEDGE) {
1682 spin_unlock(&qp->s_lock);
1685 /* XXX Flush WQEs */
1686 qp->state = IB_QPS_ERR;
1687 qp->s_ack_state = OP(RDMA_WRITE_ONLY);
1689 IB_NAK_REMOTE_ACCESS_ERROR;
1690 qp->s_ack_psn = qp->r_psn;
1694 qp->r_sge.sg_list = NULL;
1695 qp->r_sge.sge.mr = NULL;
1696 qp->r_sge.sge.vaddr = NULL;
1697 qp->r_sge.sge.length = 0;
1698 qp->r_sge.sge.sge_length = 0;
1700 if (unlikely(!(qp->qp_access_flags &
1701 IB_ACCESS_REMOTE_WRITE)))
1703 if (opcode == OP(RDMA_WRITE_FIRST))
1705 else if (opcode == OP(RDMA_WRITE_ONLY))
1707 if (!ipath_get_rwqe(qp, 1))
1711 case OP(RDMA_READ_REQUEST):
1712 /* RETH comes after BTH */
1713 if (!header_in_data)
1714 reth = &ohdr->u.rc.reth;
1716 reth = (struct ib_reth *)data;
1717 data += sizeof(*reth);
1719 spin_lock(&qp->s_lock);
1720 if (qp->s_ack_state != OP(ACKNOWLEDGE) &&
1721 qp->s_ack_state >= IB_OPCODE_RDMA_READ_REQUEST) {
1722 spin_unlock(&qp->s_lock);
1725 qp->s_rdma_len = be32_to_cpu(reth->length);
1726 if (qp->s_rdma_len != 0) {
1727 u32 rkey = be32_to_cpu(reth->rkey);
1728 u64 vaddr = be64_to_cpu(reth->vaddr);
1731 /* Check rkey & NAK */
1732 ok = ipath_rkey_ok(dev, &qp->s_rdma_sge,
1733 qp->s_rdma_len, vaddr, rkey,
1734 IB_ACCESS_REMOTE_READ);
1735 if (unlikely(!ok)) {
1736 spin_unlock(&qp->s_lock);
1740 * Update the next expected PSN. We add 1 later
1741 * below, so only add the remainder here.
1743 if (qp->s_rdma_len > pmtu)
1744 qp->r_psn += (qp->s_rdma_len - 1) / pmtu;
1746 qp->s_rdma_sge.sg_list = NULL;
1747 qp->s_rdma_sge.num_sge = 0;
1748 qp->s_rdma_sge.sge.mr = NULL;
1749 qp->s_rdma_sge.sge.vaddr = NULL;
1750 qp->s_rdma_sge.sge.length = 0;
1751 qp->s_rdma_sge.sge.sge_length = 0;
1753 if (unlikely(!(qp->qp_access_flags &
1754 IB_ACCESS_REMOTE_READ)))
1757 * We need to increment the MSN here instead of when we
1758 * finish sending the result since a duplicate request would
1759 * increment it more than once.
1761 atomic_inc(&qp->msn);
1762 qp->s_ack_state = opcode;
1763 qp->s_nak_state = 0;
1764 qp->s_ack_psn = psn;
1766 qp->r_state = opcode;
1769 case OP(COMPARE_SWAP):
1770 case OP(FETCH_ADD): {
1771 struct ib_atomic_eth *ateth;
1776 if (!header_in_data)
1777 ateth = &ohdr->u.atomic_eth;
1779 ateth = (struct ib_atomic_eth *)data;
1780 data += sizeof(*ateth);
1782 vaddr = be64_to_cpu(ateth->vaddr);
1783 if (unlikely(vaddr & (sizeof(u64) - 1)))
1785 rkey = be32_to_cpu(ateth->rkey);
1786 /* Check rkey & NAK */
1787 if (unlikely(!ipath_rkey_ok(dev, &qp->r_sge,
1788 sizeof(u64), vaddr, rkey,
1789 IB_ACCESS_REMOTE_ATOMIC)))
1791 if (unlikely(!(qp->qp_access_flags &
1792 IB_ACCESS_REMOTE_ATOMIC)))
1794 /* Perform atomic OP and save result. */
1795 sdata = be64_to_cpu(ateth->swap_data);
1796 spin_lock(&dev->pending_lock);
1797 qp->r_atomic_data = *(u64 *) qp->r_sge.sge.vaddr;
1798 if (opcode == OP(FETCH_ADD))
1799 *(u64 *) qp->r_sge.sge.vaddr =
1800 qp->r_atomic_data + sdata;
1801 else if (qp->r_atomic_data ==
1802 be64_to_cpu(ateth->compare_data))
1803 *(u64 *) qp->r_sge.sge.vaddr = sdata;
1804 spin_unlock(&dev->pending_lock);
1805 atomic_inc(&qp->msn);
1806 qp->r_atomic_psn = psn & IPS_PSN_MASK;
1812 /* Drop packet for unknown opcodes. */
1816 qp->r_state = opcode;
1817 /* Send an ACK if requested or required. */
1818 if (psn & (1 << 31)) {
1820 * Coalesce ACKs unless there is a RDMA READ or
1823 spin_lock(&qp->s_lock);
1824 if (qp->s_ack_state == OP(ACKNOWLEDGE) ||
1825 qp->s_ack_state < IB_OPCODE_RDMA_READ_REQUEST) {
1826 qp->s_ack_state = opcode;
1827 qp->s_nak_state = 0;
1828 qp->s_ack_psn = psn;
1829 qp->s_ack_atomic = qp->r_atomic_data;
1832 spin_unlock(&qp->s_lock);
1835 spin_unlock_irqrestore(&qp->r_rq.lock, flags);
1840 * Try to send ACK right away but not if ipath_do_rc_send() is
1843 if (qp->s_hdrwords == 0 &&
1844 (qp->s_ack_state < IB_OPCODE_RDMA_READ_REQUEST ||
1845 qp->s_ack_state >= IB_OPCODE_COMPARE_SWAP))
1849 spin_unlock(&qp->s_lock);
1850 spin_unlock_irqrestore(&qp->r_rq.lock, flags);
1852 /* Call ipath_do_rc_send() in another thread. */
1853 tasklet_hi_schedule(&qp->s_task);