Committer: Michael Beasley <mike@snafu.setup>
[mikesnafu-overlay.git] / drivers / infiniband / hw / ipath / ipath_rc.c
blob40f3e37d7adc3e883a8b3eff1a4c05cf7d7820f2
1 /*
2 * Copyright (c) 2006, 2007 QLogic Corporation. All rights reserved.
3 * Copyright (c) 2005, 2006 PathScale, 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
13 * conditions are met:
15 * - Redistributions of source code must retain the above
16 * copyright notice, this list of conditions and the following
17 * disclaimer.
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
31 * SOFTWARE.
34 #include "ipath_verbs.h"
35 #include "ipath_kernel.h"
37 /* cut down ridiculously long IB macro names */
38 #define OP(x) IB_OPCODE_RC_##x
40 static u32 restart_sge(struct ipath_sge_state *ss, struct ipath_swqe *wqe,
41 u32 psn, u32 pmtu)
43 u32 len;
45 len = ((psn - wqe->psn) & IPATH_PSN_MASK) * pmtu;
46 ss->sge = wqe->sg_list[0];
47 ss->sg_list = wqe->sg_list + 1;
48 ss->num_sge = wqe->wr.num_sge;
49 ipath_skip_sge(ss, len);
50 return wqe->length - len;
53 /**
54 * ipath_init_restart- initialize the qp->s_sge after a restart
55 * @qp: the QP who's SGE we're restarting
56 * @wqe: the work queue to initialize the QP's SGE from
58 * The QP s_lock should be held and interrupts disabled.
60 static void ipath_init_restart(struct ipath_qp *qp, struct ipath_swqe *wqe)
62 struct ipath_ibdev *dev;
64 qp->s_len = restart_sge(&qp->s_sge, wqe, qp->s_psn,
65 ib_mtu_enum_to_int(qp->path_mtu));
66 dev = to_idev(qp->ibqp.device);
67 spin_lock(&dev->pending_lock);
68 if (list_empty(&qp->timerwait))
69 list_add_tail(&qp->timerwait,
70 &dev->pending[dev->pending_index]);
71 spin_unlock(&dev->pending_lock);
74 /**
75 * ipath_make_rc_ack - construct a response packet (ACK, NAK, or RDMA read)
76 * @qp: a pointer to the QP
77 * @ohdr: a pointer to the IB header being constructed
78 * @pmtu: the path MTU
80 * Return 1 if constructed; otherwise, return 0.
81 * Note that we are in the responder's side of the QP context.
82 * Note the QP s_lock must be held.
84 static int ipath_make_rc_ack(struct ipath_ibdev *dev, struct ipath_qp *qp,
85 struct ipath_other_headers *ohdr, u32 pmtu)
87 struct ipath_ack_entry *e;
88 u32 hwords;
89 u32 len;
90 u32 bth0;
91 u32 bth2;
93 /* header size in 32-bit words LRH+BTH = (8+12)/4. */
94 hwords = 5;
96 switch (qp->s_ack_state) {
97 case OP(RDMA_READ_RESPONSE_LAST):
98 case OP(RDMA_READ_RESPONSE_ONLY):
99 case OP(ATOMIC_ACKNOWLEDGE):
101 * We can increment the tail pointer now that the last
102 * response has been sent instead of only being
103 * constructed.
105 if (++qp->s_tail_ack_queue > IPATH_MAX_RDMA_ATOMIC)
106 qp->s_tail_ack_queue = 0;
107 /* FALLTHROUGH */
108 case OP(SEND_ONLY):
109 case OP(ACKNOWLEDGE):
110 /* Check for no next entry in the queue. */
111 if (qp->r_head_ack_queue == qp->s_tail_ack_queue) {
112 if (qp->s_flags & IPATH_S_ACK_PENDING)
113 goto normal;
114 qp->s_ack_state = OP(ACKNOWLEDGE);
115 goto bail;
118 e = &qp->s_ack_queue[qp->s_tail_ack_queue];
119 if (e->opcode == OP(RDMA_READ_REQUEST)) {
120 /* Copy SGE state in case we need to resend */
121 qp->s_ack_rdma_sge = e->rdma_sge;
122 qp->s_cur_sge = &qp->s_ack_rdma_sge;
123 len = e->rdma_sge.sge.sge_length;
124 if (len > pmtu) {
125 len = pmtu;
126 qp->s_ack_state = OP(RDMA_READ_RESPONSE_FIRST);
127 } else {
128 qp->s_ack_state = OP(RDMA_READ_RESPONSE_ONLY);
129 e->sent = 1;
131 ohdr->u.aeth = ipath_compute_aeth(qp);
132 hwords++;
133 qp->s_ack_rdma_psn = e->psn;
134 bth2 = qp->s_ack_rdma_psn++ & IPATH_PSN_MASK;
135 } else {
136 /* COMPARE_SWAP or FETCH_ADD */
137 qp->s_cur_sge = NULL;
138 len = 0;
139 qp->s_ack_state = OP(ATOMIC_ACKNOWLEDGE);
140 ohdr->u.at.aeth = ipath_compute_aeth(qp);
141 ohdr->u.at.atomic_ack_eth[0] =
142 cpu_to_be32(e->atomic_data >> 32);
143 ohdr->u.at.atomic_ack_eth[1] =
144 cpu_to_be32(e->atomic_data);
145 hwords += sizeof(ohdr->u.at) / sizeof(u32);
146 bth2 = e->psn;
147 e->sent = 1;
149 bth0 = qp->s_ack_state << 24;
150 break;
152 case OP(RDMA_READ_RESPONSE_FIRST):
153 qp->s_ack_state = OP(RDMA_READ_RESPONSE_MIDDLE);
154 /* FALLTHROUGH */
155 case OP(RDMA_READ_RESPONSE_MIDDLE):
156 len = qp->s_ack_rdma_sge.sge.sge_length;
157 if (len > pmtu)
158 len = pmtu;
159 else {
160 ohdr->u.aeth = ipath_compute_aeth(qp);
161 hwords++;
162 qp->s_ack_state = OP(RDMA_READ_RESPONSE_LAST);
163 qp->s_ack_queue[qp->s_tail_ack_queue].sent = 1;
165 bth0 = qp->s_ack_state << 24;
166 bth2 = qp->s_ack_rdma_psn++ & IPATH_PSN_MASK;
167 break;
169 default:
170 normal:
172 * Send a regular ACK.
173 * Set the s_ack_state so we wait until after sending
174 * the ACK before setting s_ack_state to ACKNOWLEDGE
175 * (see above).
177 qp->s_ack_state = OP(SEND_ONLY);
178 qp->s_flags &= ~IPATH_S_ACK_PENDING;
179 qp->s_cur_sge = NULL;
180 if (qp->s_nak_state)
181 ohdr->u.aeth =
182 cpu_to_be32((qp->r_msn & IPATH_MSN_MASK) |
183 (qp->s_nak_state <<
184 IPATH_AETH_CREDIT_SHIFT));
185 else
186 ohdr->u.aeth = ipath_compute_aeth(qp);
187 hwords++;
188 len = 0;
189 bth0 = OP(ACKNOWLEDGE) << 24;
190 bth2 = qp->s_ack_psn & IPATH_PSN_MASK;
192 qp->s_hdrwords = hwords;
193 qp->s_cur_size = len;
194 ipath_make_ruc_header(dev, qp, ohdr, bth0, bth2);
195 return 1;
197 bail:
198 return 0;
202 * ipath_make_rc_req - construct a request packet (SEND, RDMA r/w, ATOMIC)
203 * @qp: a pointer to the QP
205 * Return 1 if constructed; otherwise, return 0.
207 int ipath_make_rc_req(struct ipath_qp *qp)
209 struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
210 struct ipath_other_headers *ohdr;
211 struct ipath_sge_state *ss;
212 struct ipath_swqe *wqe;
213 u32 hwords;
214 u32 len;
215 u32 bth0;
216 u32 bth2;
217 u32 pmtu = ib_mtu_enum_to_int(qp->path_mtu);
218 char newreq;
219 unsigned long flags;
220 int ret = 0;
222 ohdr = &qp->s_hdr.u.oth;
223 if (qp->remote_ah_attr.ah_flags & IB_AH_GRH)
224 ohdr = &qp->s_hdr.u.l.oth;
227 * The lock is needed to synchronize between the sending tasklet,
228 * the receive interrupt handler, and timeout resends.
230 spin_lock_irqsave(&qp->s_lock, flags);
232 /* Sending responses has higher priority over sending requests. */
233 if ((qp->r_head_ack_queue != qp->s_tail_ack_queue ||
234 (qp->s_flags & IPATH_S_ACK_PENDING) ||
235 qp->s_ack_state != OP(ACKNOWLEDGE)) &&
236 ipath_make_rc_ack(dev, qp, ohdr, pmtu))
237 goto done;
239 if (!(ib_ipath_state_ops[qp->state] & IPATH_PROCESS_SEND_OK) ||
240 qp->s_rnr_timeout || qp->s_wait_credit)
241 goto bail;
243 /* Limit the number of packets sent without an ACK. */
244 if (ipath_cmp24(qp->s_psn, qp->s_last_psn + IPATH_PSN_CREDIT) > 0) {
245 qp->s_wait_credit = 1;
246 dev->n_rc_stalls++;
247 goto bail;
250 /* header size in 32-bit words LRH+BTH = (8+12)/4. */
251 hwords = 5;
252 bth0 = 1 << 22; /* Set M bit */
254 /* Send a request. */
255 wqe = get_swqe_ptr(qp, qp->s_cur);
256 switch (qp->s_state) {
257 default:
259 * Resend an old request or start a new one.
261 * We keep track of the current SWQE so that
262 * we don't reset the "furthest progress" state
263 * if we need to back up.
265 newreq = 0;
266 if (qp->s_cur == qp->s_tail) {
267 /* Check if send work queue is empty. */
268 if (qp->s_tail == qp->s_head)
269 goto bail;
271 * If a fence is requested, wait for previous
272 * RDMA read and atomic operations to finish.
274 if ((wqe->wr.send_flags & IB_SEND_FENCE) &&
275 qp->s_num_rd_atomic) {
276 qp->s_flags |= IPATH_S_FENCE_PENDING;
277 goto bail;
279 wqe->psn = qp->s_next_psn;
280 newreq = 1;
283 * Note that we have to be careful not to modify the
284 * original work request since we may need to resend
285 * it.
287 len = wqe->length;
288 ss = &qp->s_sge;
289 bth2 = 0;
290 switch (wqe->wr.opcode) {
291 case IB_WR_SEND:
292 case IB_WR_SEND_WITH_IMM:
293 /* If no credit, return. */
294 if (qp->s_lsn != (u32) -1 &&
295 ipath_cmp24(wqe->ssn, qp->s_lsn + 1) > 0)
296 goto bail;
297 wqe->lpsn = wqe->psn;
298 if (len > pmtu) {
299 wqe->lpsn += (len - 1) / pmtu;
300 qp->s_state = OP(SEND_FIRST);
301 len = pmtu;
302 break;
304 if (wqe->wr.opcode == IB_WR_SEND)
305 qp->s_state = OP(SEND_ONLY);
306 else {
307 qp->s_state = OP(SEND_ONLY_WITH_IMMEDIATE);
308 /* Immediate data comes after the BTH */
309 ohdr->u.imm_data = wqe->wr.imm_data;
310 hwords += 1;
312 if (wqe->wr.send_flags & IB_SEND_SOLICITED)
313 bth0 |= 1 << 23;
314 bth2 = 1 << 31; /* Request ACK. */
315 if (++qp->s_cur == qp->s_size)
316 qp->s_cur = 0;
317 break;
319 case IB_WR_RDMA_WRITE:
320 if (newreq && qp->s_lsn != (u32) -1)
321 qp->s_lsn++;
322 /* FALLTHROUGH */
323 case IB_WR_RDMA_WRITE_WITH_IMM:
324 /* If no credit, return. */
325 if (qp->s_lsn != (u32) -1 &&
326 ipath_cmp24(wqe->ssn, qp->s_lsn + 1) > 0)
327 goto bail;
328 ohdr->u.rc.reth.vaddr =
329 cpu_to_be64(wqe->wr.wr.rdma.remote_addr);
330 ohdr->u.rc.reth.rkey =
331 cpu_to_be32(wqe->wr.wr.rdma.rkey);
332 ohdr->u.rc.reth.length = cpu_to_be32(len);
333 hwords += sizeof(struct ib_reth) / sizeof(u32);
334 wqe->lpsn = wqe->psn;
335 if (len > pmtu) {
336 wqe->lpsn += (len - 1) / pmtu;
337 qp->s_state = OP(RDMA_WRITE_FIRST);
338 len = pmtu;
339 break;
341 if (wqe->wr.opcode == IB_WR_RDMA_WRITE)
342 qp->s_state = OP(RDMA_WRITE_ONLY);
343 else {
344 qp->s_state =
345 OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE);
346 /* Immediate data comes after RETH */
347 ohdr->u.rc.imm_data = wqe->wr.imm_data;
348 hwords += 1;
349 if (wqe->wr.send_flags & IB_SEND_SOLICITED)
350 bth0 |= 1 << 23;
352 bth2 = 1 << 31; /* Request ACK. */
353 if (++qp->s_cur == qp->s_size)
354 qp->s_cur = 0;
355 break;
357 case IB_WR_RDMA_READ:
359 * Don't allow more operations to be started
360 * than the QP limits allow.
362 if (newreq) {
363 if (qp->s_num_rd_atomic >=
364 qp->s_max_rd_atomic) {
365 qp->s_flags |= IPATH_S_RDMAR_PENDING;
366 goto bail;
368 qp->s_num_rd_atomic++;
369 if (qp->s_lsn != (u32) -1)
370 qp->s_lsn++;
372 * Adjust s_next_psn to count the
373 * expected number of responses.
375 if (len > pmtu)
376 qp->s_next_psn += (len - 1) / pmtu;
377 wqe->lpsn = qp->s_next_psn++;
379 ohdr->u.rc.reth.vaddr =
380 cpu_to_be64(wqe->wr.wr.rdma.remote_addr);
381 ohdr->u.rc.reth.rkey =
382 cpu_to_be32(wqe->wr.wr.rdma.rkey);
383 ohdr->u.rc.reth.length = cpu_to_be32(len);
384 qp->s_state = OP(RDMA_READ_REQUEST);
385 hwords += sizeof(ohdr->u.rc.reth) / sizeof(u32);
386 ss = NULL;
387 len = 0;
388 if (++qp->s_cur == qp->s_size)
389 qp->s_cur = 0;
390 break;
392 case IB_WR_ATOMIC_CMP_AND_SWP:
393 case IB_WR_ATOMIC_FETCH_AND_ADD:
395 * Don't allow more operations to be started
396 * than the QP limits allow.
398 if (newreq) {
399 if (qp->s_num_rd_atomic >=
400 qp->s_max_rd_atomic) {
401 qp->s_flags |= IPATH_S_RDMAR_PENDING;
402 goto bail;
404 qp->s_num_rd_atomic++;
405 if (qp->s_lsn != (u32) -1)
406 qp->s_lsn++;
407 wqe->lpsn = wqe->psn;
409 if (wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
410 qp->s_state = OP(COMPARE_SWAP);
411 ohdr->u.atomic_eth.swap_data = cpu_to_be64(
412 wqe->wr.wr.atomic.swap);
413 ohdr->u.atomic_eth.compare_data = cpu_to_be64(
414 wqe->wr.wr.atomic.compare_add);
415 } else {
416 qp->s_state = OP(FETCH_ADD);
417 ohdr->u.atomic_eth.swap_data = cpu_to_be64(
418 wqe->wr.wr.atomic.compare_add);
419 ohdr->u.atomic_eth.compare_data = 0;
421 ohdr->u.atomic_eth.vaddr[0] = cpu_to_be32(
422 wqe->wr.wr.atomic.remote_addr >> 32);
423 ohdr->u.atomic_eth.vaddr[1] = cpu_to_be32(
424 wqe->wr.wr.atomic.remote_addr);
425 ohdr->u.atomic_eth.rkey = cpu_to_be32(
426 wqe->wr.wr.atomic.rkey);
427 hwords += sizeof(struct ib_atomic_eth) / sizeof(u32);
428 ss = NULL;
429 len = 0;
430 if (++qp->s_cur == qp->s_size)
431 qp->s_cur = 0;
432 break;
434 default:
435 goto bail;
437 qp->s_sge.sge = wqe->sg_list[0];
438 qp->s_sge.sg_list = wqe->sg_list + 1;
439 qp->s_sge.num_sge = wqe->wr.num_sge;
440 qp->s_len = wqe->length;
441 if (newreq) {
442 qp->s_tail++;
443 if (qp->s_tail >= qp->s_size)
444 qp->s_tail = 0;
446 bth2 |= qp->s_psn & IPATH_PSN_MASK;
447 if (wqe->wr.opcode == IB_WR_RDMA_READ)
448 qp->s_psn = wqe->lpsn + 1;
449 else {
450 qp->s_psn++;
451 if (ipath_cmp24(qp->s_psn, qp->s_next_psn) > 0)
452 qp->s_next_psn = qp->s_psn;
455 * Put the QP on the pending list so lost ACKs will cause
456 * a retry. More than one request can be pending so the
457 * QP may already be on the dev->pending list.
459 spin_lock(&dev->pending_lock);
460 if (list_empty(&qp->timerwait))
461 list_add_tail(&qp->timerwait,
462 &dev->pending[dev->pending_index]);
463 spin_unlock(&dev->pending_lock);
464 break;
466 case OP(RDMA_READ_RESPONSE_FIRST):
468 * This case can only happen if a send is restarted.
469 * See ipath_restart_rc().
471 ipath_init_restart(qp, wqe);
472 /* FALLTHROUGH */
473 case OP(SEND_FIRST):
474 qp->s_state = OP(SEND_MIDDLE);
475 /* FALLTHROUGH */
476 case OP(SEND_MIDDLE):
477 bth2 = qp->s_psn++ & IPATH_PSN_MASK;
478 if (ipath_cmp24(qp->s_psn, qp->s_next_psn) > 0)
479 qp->s_next_psn = qp->s_psn;
480 ss = &qp->s_sge;
481 len = qp->s_len;
482 if (len > pmtu) {
483 len = pmtu;
484 break;
486 if (wqe->wr.opcode == IB_WR_SEND)
487 qp->s_state = OP(SEND_LAST);
488 else {
489 qp->s_state = OP(SEND_LAST_WITH_IMMEDIATE);
490 /* Immediate data comes after the BTH */
491 ohdr->u.imm_data = wqe->wr.imm_data;
492 hwords += 1;
494 if (wqe->wr.send_flags & IB_SEND_SOLICITED)
495 bth0 |= 1 << 23;
496 bth2 |= 1 << 31; /* Request ACK. */
497 qp->s_cur++;
498 if (qp->s_cur >= qp->s_size)
499 qp->s_cur = 0;
500 break;
502 case OP(RDMA_READ_RESPONSE_LAST):
504 * This case can only happen if a RDMA write is restarted.
505 * See ipath_restart_rc().
507 ipath_init_restart(qp, wqe);
508 /* FALLTHROUGH */
509 case OP(RDMA_WRITE_FIRST):
510 qp->s_state = OP(RDMA_WRITE_MIDDLE);
511 /* FALLTHROUGH */
512 case OP(RDMA_WRITE_MIDDLE):
513 bth2 = qp->s_psn++ & IPATH_PSN_MASK;
514 if (ipath_cmp24(qp->s_psn, qp->s_next_psn) > 0)
515 qp->s_next_psn = qp->s_psn;
516 ss = &qp->s_sge;
517 len = qp->s_len;
518 if (len > pmtu) {
519 len = pmtu;
520 break;
522 if (wqe->wr.opcode == IB_WR_RDMA_WRITE)
523 qp->s_state = OP(RDMA_WRITE_LAST);
524 else {
525 qp->s_state = OP(RDMA_WRITE_LAST_WITH_IMMEDIATE);
526 /* Immediate data comes after the BTH */
527 ohdr->u.imm_data = wqe->wr.imm_data;
528 hwords += 1;
529 if (wqe->wr.send_flags & IB_SEND_SOLICITED)
530 bth0 |= 1 << 23;
532 bth2 |= 1 << 31; /* Request ACK. */
533 qp->s_cur++;
534 if (qp->s_cur >= qp->s_size)
535 qp->s_cur = 0;
536 break;
538 case OP(RDMA_READ_RESPONSE_MIDDLE):
540 * This case can only happen if a RDMA read is restarted.
541 * See ipath_restart_rc().
543 ipath_init_restart(qp, wqe);
544 len = ((qp->s_psn - wqe->psn) & IPATH_PSN_MASK) * pmtu;
545 ohdr->u.rc.reth.vaddr =
546 cpu_to_be64(wqe->wr.wr.rdma.remote_addr + len);
547 ohdr->u.rc.reth.rkey =
548 cpu_to_be32(wqe->wr.wr.rdma.rkey);
549 ohdr->u.rc.reth.length = cpu_to_be32(qp->s_len);
550 qp->s_state = OP(RDMA_READ_REQUEST);
551 hwords += sizeof(ohdr->u.rc.reth) / sizeof(u32);
552 bth2 = qp->s_psn++ & IPATH_PSN_MASK;
553 if (ipath_cmp24(qp->s_psn, qp->s_next_psn) > 0)
554 qp->s_next_psn = qp->s_psn;
555 ss = NULL;
556 len = 0;
557 qp->s_cur++;
558 if (qp->s_cur == qp->s_size)
559 qp->s_cur = 0;
560 break;
562 if (ipath_cmp24(qp->s_psn, qp->s_last_psn + IPATH_PSN_CREDIT - 1) >= 0)
563 bth2 |= 1 << 31; /* Request ACK. */
564 qp->s_len -= len;
565 qp->s_hdrwords = hwords;
566 qp->s_cur_sge = ss;
567 qp->s_cur_size = len;
568 ipath_make_ruc_header(dev, qp, ohdr, bth0 | (qp->s_state << 24), bth2);
569 done:
570 ret = 1;
571 bail:
572 spin_unlock_irqrestore(&qp->s_lock, flags);
573 return ret;
577 * send_rc_ack - Construct an ACK packet and send it
578 * @qp: a pointer to the QP
580 * This is called from ipath_rc_rcv() and only uses the receive
581 * side QP state.
582 * Note that RDMA reads and atomics are handled in the
583 * send side QP state and tasklet.
585 static void send_rc_ack(struct ipath_qp *qp)
587 struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
588 u16 lrh0;
589 u32 bth0;
590 u32 hwords;
591 struct ipath_ib_header hdr;
592 struct ipath_other_headers *ohdr;
593 unsigned long flags;
595 /* Don't send ACK or NAK if a RDMA read or atomic is pending. */
596 if (qp->r_head_ack_queue != qp->s_tail_ack_queue ||
597 (qp->s_flags & IPATH_S_ACK_PENDING) ||
598 qp->s_ack_state != OP(ACKNOWLEDGE))
599 goto queue_ack;
601 /* Construct the header. */
602 ohdr = &hdr.u.oth;
603 lrh0 = IPATH_LRH_BTH;
604 /* header size in 32-bit words LRH+BTH+AETH = (8+12+4)/4. */
605 hwords = 6;
606 if (unlikely(qp->remote_ah_attr.ah_flags & IB_AH_GRH)) {
607 hwords += ipath_make_grh(dev, &hdr.u.l.grh,
608 &qp->remote_ah_attr.grh,
609 hwords, 0);
610 ohdr = &hdr.u.l.oth;
611 lrh0 = IPATH_LRH_GRH;
613 /* read pkey_index w/o lock (its atomic) */
614 bth0 = ipath_get_pkey(dev->dd, qp->s_pkey_index) |
615 (OP(ACKNOWLEDGE) << 24) | (1 << 22);
616 if (qp->r_nak_state)
617 ohdr->u.aeth = cpu_to_be32((qp->r_msn & IPATH_MSN_MASK) |
618 (qp->r_nak_state <<
619 IPATH_AETH_CREDIT_SHIFT));
620 else
621 ohdr->u.aeth = ipath_compute_aeth(qp);
622 lrh0 |= qp->remote_ah_attr.sl << 4;
623 hdr.lrh[0] = cpu_to_be16(lrh0);
624 hdr.lrh[1] = cpu_to_be16(qp->remote_ah_attr.dlid);
625 hdr.lrh[2] = cpu_to_be16(hwords + SIZE_OF_CRC);
626 hdr.lrh[3] = cpu_to_be16(dev->dd->ipath_lid);
627 ohdr->bth[0] = cpu_to_be32(bth0);
628 ohdr->bth[1] = cpu_to_be32(qp->remote_qpn);
629 ohdr->bth[2] = cpu_to_be32(qp->r_ack_psn & IPATH_PSN_MASK);
632 * If we can send the ACK, clear the ACK state.
634 if (ipath_verbs_send(qp, &hdr, hwords, NULL, 0) == 0) {
635 dev->n_unicast_xmit++;
636 goto done;
640 * We are out of PIO buffers at the moment.
641 * Pass responsibility for sending the ACK to the
642 * send tasklet so that when a PIO buffer becomes
643 * available, the ACK is sent ahead of other outgoing
644 * packets.
646 dev->n_rc_qacks++;
648 queue_ack:
649 spin_lock_irqsave(&qp->s_lock, flags);
650 dev->n_rc_qacks++;
651 qp->s_flags |= IPATH_S_ACK_PENDING;
652 qp->s_nak_state = qp->r_nak_state;
653 qp->s_ack_psn = qp->r_ack_psn;
654 spin_unlock_irqrestore(&qp->s_lock, flags);
656 /* Call ipath_do_rc_send() in another thread. */
657 tasklet_hi_schedule(&qp->s_task);
659 done:
660 return;
664 * reset_psn - reset the QP state to send starting from PSN
665 * @qp: the QP
666 * @psn: the packet sequence number to restart at
668 * This is called from ipath_rc_rcv() to process an incoming RC ACK
669 * for the given QP.
670 * Called at interrupt level with the QP s_lock held.
672 static void reset_psn(struct ipath_qp *qp, u32 psn)
674 u32 n = qp->s_last;
675 struct ipath_swqe *wqe = get_swqe_ptr(qp, n);
676 u32 opcode;
678 qp->s_cur = n;
681 * If we are starting the request from the beginning,
682 * let the normal send code handle initialization.
684 if (ipath_cmp24(psn, wqe->psn) <= 0) {
685 qp->s_state = OP(SEND_LAST);
686 goto done;
689 /* Find the work request opcode corresponding to the given PSN. */
690 opcode = wqe->wr.opcode;
691 for (;;) {
692 int diff;
694 if (++n == qp->s_size)
695 n = 0;
696 if (n == qp->s_tail)
697 break;
698 wqe = get_swqe_ptr(qp, n);
699 diff = ipath_cmp24(psn, wqe->psn);
700 if (diff < 0)
701 break;
702 qp->s_cur = n;
704 * If we are starting the request from the beginning,
705 * let the normal send code handle initialization.
707 if (diff == 0) {
708 qp->s_state = OP(SEND_LAST);
709 goto done;
711 opcode = wqe->wr.opcode;
715 * Set the state to restart in the middle of a request.
716 * Don't change the s_sge, s_cur_sge, or s_cur_size.
717 * See ipath_do_rc_send().
719 switch (opcode) {
720 case IB_WR_SEND:
721 case IB_WR_SEND_WITH_IMM:
722 qp->s_state = OP(RDMA_READ_RESPONSE_FIRST);
723 break;
725 case IB_WR_RDMA_WRITE:
726 case IB_WR_RDMA_WRITE_WITH_IMM:
727 qp->s_state = OP(RDMA_READ_RESPONSE_LAST);
728 break;
730 case IB_WR_RDMA_READ:
731 qp->s_state = OP(RDMA_READ_RESPONSE_MIDDLE);
732 break;
734 default:
736 * This case shouldn't happen since its only
737 * one PSN per req.
739 qp->s_state = OP(SEND_LAST);
741 done:
742 qp->s_psn = psn;
746 * ipath_restart_rc - back up requester to resend the last un-ACKed request
747 * @qp: the QP to restart
748 * @psn: packet sequence number for the request
749 * @wc: the work completion request
751 * The QP s_lock should be held and interrupts disabled.
753 void ipath_restart_rc(struct ipath_qp *qp, u32 psn, struct ib_wc *wc)
755 struct ipath_swqe *wqe = get_swqe_ptr(qp, qp->s_last);
756 struct ipath_ibdev *dev;
758 if (qp->s_retry == 0) {
759 wc->wr_id = wqe->wr.wr_id;
760 wc->status = IB_WC_RETRY_EXC_ERR;
761 wc->opcode = ib_ipath_wc_opcode[wqe->wr.opcode];
762 wc->vendor_err = 0;
763 wc->byte_len = 0;
764 wc->qp = &qp->ibqp;
765 wc->imm_data = 0;
766 wc->src_qp = qp->remote_qpn;
767 wc->wc_flags = 0;
768 wc->pkey_index = 0;
769 wc->slid = qp->remote_ah_attr.dlid;
770 wc->sl = qp->remote_ah_attr.sl;
771 wc->dlid_path_bits = 0;
772 wc->port_num = 0;
773 ipath_sqerror_qp(qp, wc);
774 goto bail;
776 qp->s_retry--;
779 * Remove the QP from the timeout queue.
780 * Note: it may already have been removed by ipath_ib_timer().
782 dev = to_idev(qp->ibqp.device);
783 spin_lock(&dev->pending_lock);
784 if (!list_empty(&qp->timerwait))
785 list_del_init(&qp->timerwait);
786 spin_unlock(&dev->pending_lock);
788 if (wqe->wr.opcode == IB_WR_RDMA_READ)
789 dev->n_rc_resends++;
790 else
791 dev->n_rc_resends += (qp->s_psn - psn) & IPATH_PSN_MASK;
793 reset_psn(qp, psn);
794 tasklet_hi_schedule(&qp->s_task);
796 bail:
797 return;
800 static inline void update_last_psn(struct ipath_qp *qp, u32 psn)
802 if (qp->s_last_psn != psn) {
803 qp->s_last_psn = psn;
804 if (qp->s_wait_credit) {
805 qp->s_wait_credit = 0;
806 tasklet_hi_schedule(&qp->s_task);
812 * do_rc_ack - process an incoming RC ACK
813 * @qp: the QP the ACK came in on
814 * @psn: the packet sequence number of the ACK
815 * @opcode: the opcode of the request that resulted in the ACK
817 * This is called from ipath_rc_rcv_resp() to process an incoming RC ACK
818 * for the given QP.
819 * Called at interrupt level with the QP s_lock held and interrupts disabled.
820 * Returns 1 if OK, 0 if current operation should be aborted (NAK).
822 static int do_rc_ack(struct ipath_qp *qp, u32 aeth, u32 psn, int opcode,
823 u64 val)
825 struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
826 struct ib_wc wc;
827 struct ipath_swqe *wqe;
828 int ret = 0;
829 u32 ack_psn;
830 int diff;
833 * Remove the QP from the timeout queue (or RNR timeout queue).
834 * If ipath_ib_timer() has already removed it,
835 * it's OK since we hold the QP s_lock and ipath_restart_rc()
836 * just won't find anything to restart if we ACK everything.
838 spin_lock(&dev->pending_lock);
839 if (!list_empty(&qp->timerwait))
840 list_del_init(&qp->timerwait);
841 spin_unlock(&dev->pending_lock);
844 * Note that NAKs implicitly ACK outstanding SEND and RDMA write
845 * requests and implicitly NAK RDMA read and atomic requests issued
846 * before the NAK'ed request. The MSN won't include the NAK'ed
847 * request but will include an ACK'ed request(s).
849 ack_psn = psn;
850 if (aeth >> 29)
851 ack_psn--;
852 wqe = get_swqe_ptr(qp, qp->s_last);
855 * The MSN might be for a later WQE than the PSN indicates so
856 * only complete WQEs that the PSN finishes.
858 while ((diff = ipath_cmp24(ack_psn, wqe->lpsn)) >= 0) {
860 * RDMA_READ_RESPONSE_ONLY is a special case since
861 * we want to generate completion events for everything
862 * before the RDMA read, copy the data, then generate
863 * the completion for the read.
865 if (wqe->wr.opcode == IB_WR_RDMA_READ &&
866 opcode == OP(RDMA_READ_RESPONSE_ONLY) &&
867 diff == 0) {
868 ret = 1;
869 goto bail;
872 * If this request is a RDMA read or atomic, and the ACK is
873 * for a later operation, this ACK NAKs the RDMA read or
874 * atomic. In other words, only a RDMA_READ_LAST or ONLY
875 * can ACK a RDMA read and likewise for atomic ops. Note
876 * that the NAK case can only happen if relaxed ordering is
877 * used and requests are sent after an RDMA read or atomic
878 * is sent but before the response is received.
880 if ((wqe->wr.opcode == IB_WR_RDMA_READ &&
881 (opcode != OP(RDMA_READ_RESPONSE_LAST) || diff != 0)) ||
882 ((wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
883 wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD) &&
884 (opcode != OP(ATOMIC_ACKNOWLEDGE) || diff != 0))) {
886 * The last valid PSN seen is the previous
887 * request's.
889 update_last_psn(qp, wqe->psn - 1);
890 /* Retry this request. */
891 ipath_restart_rc(qp, wqe->psn, &wc);
893 * No need to process the ACK/NAK since we are
894 * restarting an earlier request.
896 goto bail;
898 if (wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
899 wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD)
900 *(u64 *) wqe->sg_list[0].vaddr = val;
901 if (qp->s_num_rd_atomic &&
902 (wqe->wr.opcode == IB_WR_RDMA_READ ||
903 wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
904 wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD)) {
905 qp->s_num_rd_atomic--;
906 /* Restart sending task if fence is complete */
907 if ((qp->s_flags & IPATH_S_FENCE_PENDING) &&
908 !qp->s_num_rd_atomic) {
909 qp->s_flags &= ~IPATH_S_FENCE_PENDING;
910 tasklet_hi_schedule(&qp->s_task);
911 } else if (qp->s_flags & IPATH_S_RDMAR_PENDING) {
912 qp->s_flags &= ~IPATH_S_RDMAR_PENDING;
913 tasklet_hi_schedule(&qp->s_task);
916 /* Post a send completion queue entry if requested. */
917 if (!(qp->s_flags & IPATH_S_SIGNAL_REQ_WR) ||
918 (wqe->wr.send_flags & IB_SEND_SIGNALED)) {
919 wc.wr_id = wqe->wr.wr_id;
920 wc.status = IB_WC_SUCCESS;
921 wc.opcode = ib_ipath_wc_opcode[wqe->wr.opcode];
922 wc.vendor_err = 0;
923 wc.byte_len = wqe->length;
924 wc.imm_data = 0;
925 wc.qp = &qp->ibqp;
926 wc.src_qp = qp->remote_qpn;
927 wc.wc_flags = 0;
928 wc.pkey_index = 0;
929 wc.slid = qp->remote_ah_attr.dlid;
930 wc.sl = qp->remote_ah_attr.sl;
931 wc.dlid_path_bits = 0;
932 wc.port_num = 0;
933 ipath_cq_enter(to_icq(qp->ibqp.send_cq), &wc, 0);
935 qp->s_retry = qp->s_retry_cnt;
937 * If we are completing a request which is in the process of
938 * being resent, we can stop resending it since we know the
939 * responder has already seen it.
941 if (qp->s_last == qp->s_cur) {
942 if (++qp->s_cur >= qp->s_size)
943 qp->s_cur = 0;
944 qp->s_last = qp->s_cur;
945 if (qp->s_last == qp->s_tail)
946 break;
947 wqe = get_swqe_ptr(qp, qp->s_cur);
948 qp->s_state = OP(SEND_LAST);
949 qp->s_psn = wqe->psn;
950 } else {
951 if (++qp->s_last >= qp->s_size)
952 qp->s_last = 0;
953 if (qp->s_last == qp->s_tail)
954 break;
955 wqe = get_swqe_ptr(qp, qp->s_last);
959 switch (aeth >> 29) {
960 case 0: /* ACK */
961 dev->n_rc_acks++;
962 /* If this is a partial ACK, reset the retransmit timer. */
963 if (qp->s_last != qp->s_tail) {
964 spin_lock(&dev->pending_lock);
965 if (list_empty(&qp->timerwait))
966 list_add_tail(&qp->timerwait,
967 &dev->pending[dev->pending_index]);
968 spin_unlock(&dev->pending_lock);
970 * If we get a partial ACK for a resent operation,
971 * we can stop resending the earlier packets and
972 * continue with the next packet the receiver wants.
974 if (ipath_cmp24(qp->s_psn, psn) <= 0) {
975 reset_psn(qp, psn + 1);
976 tasklet_hi_schedule(&qp->s_task);
978 } else if (ipath_cmp24(qp->s_psn, psn) <= 0) {
979 qp->s_state = OP(SEND_LAST);
980 qp->s_psn = psn + 1;
982 ipath_get_credit(qp, aeth);
983 qp->s_rnr_retry = qp->s_rnr_retry_cnt;
984 qp->s_retry = qp->s_retry_cnt;
985 update_last_psn(qp, psn);
986 ret = 1;
987 goto bail;
989 case 1: /* RNR NAK */
990 dev->n_rnr_naks++;
991 if (qp->s_last == qp->s_tail)
992 goto bail;
993 if (qp->s_rnr_retry == 0) {
994 wc.status = IB_WC_RNR_RETRY_EXC_ERR;
995 goto class_b;
997 if (qp->s_rnr_retry_cnt < 7)
998 qp->s_rnr_retry--;
1000 /* The last valid PSN is the previous PSN. */
1001 update_last_psn(qp, psn - 1);
1003 if (wqe->wr.opcode == IB_WR_RDMA_READ)
1004 dev->n_rc_resends++;
1005 else
1006 dev->n_rc_resends +=
1007 (qp->s_psn - psn) & IPATH_PSN_MASK;
1009 reset_psn(qp, psn);
1011 qp->s_rnr_timeout =
1012 ib_ipath_rnr_table[(aeth >> IPATH_AETH_CREDIT_SHIFT) &
1013 IPATH_AETH_CREDIT_MASK];
1014 ipath_insert_rnr_queue(qp);
1015 goto bail;
1017 case 3: /* NAK */
1018 if (qp->s_last == qp->s_tail)
1019 goto bail;
1020 /* The last valid PSN is the previous PSN. */
1021 update_last_psn(qp, psn - 1);
1022 switch ((aeth >> IPATH_AETH_CREDIT_SHIFT) &
1023 IPATH_AETH_CREDIT_MASK) {
1024 case 0: /* PSN sequence error */
1025 dev->n_seq_naks++;
1027 * Back up to the responder's expected PSN.
1028 * Note that we might get a NAK in the middle of an
1029 * RDMA READ response which terminates the RDMA
1030 * READ.
1032 ipath_restart_rc(qp, psn, &wc);
1033 break;
1035 case 1: /* Invalid Request */
1036 wc.status = IB_WC_REM_INV_REQ_ERR;
1037 dev->n_other_naks++;
1038 goto class_b;
1040 case 2: /* Remote Access Error */
1041 wc.status = IB_WC_REM_ACCESS_ERR;
1042 dev->n_other_naks++;
1043 goto class_b;
1045 case 3: /* Remote Operation Error */
1046 wc.status = IB_WC_REM_OP_ERR;
1047 dev->n_other_naks++;
1048 class_b:
1049 wc.wr_id = wqe->wr.wr_id;
1050 wc.opcode = ib_ipath_wc_opcode[wqe->wr.opcode];
1051 wc.vendor_err = 0;
1052 wc.byte_len = 0;
1053 wc.qp = &qp->ibqp;
1054 wc.imm_data = 0;
1055 wc.src_qp = qp->remote_qpn;
1056 wc.wc_flags = 0;
1057 wc.pkey_index = 0;
1058 wc.slid = qp->remote_ah_attr.dlid;
1059 wc.sl = qp->remote_ah_attr.sl;
1060 wc.dlid_path_bits = 0;
1061 wc.port_num = 0;
1062 ipath_sqerror_qp(qp, &wc);
1063 break;
1065 default:
1066 /* Ignore other reserved NAK error codes */
1067 goto reserved;
1069 qp->s_rnr_retry = qp->s_rnr_retry_cnt;
1070 goto bail;
1072 default: /* 2: reserved */
1073 reserved:
1074 /* Ignore reserved NAK codes. */
1075 goto bail;
1078 bail:
1079 return ret;
1083 * ipath_rc_rcv_resp - process an incoming RC response packet
1084 * @dev: the device this packet came in on
1085 * @ohdr: the other headers for this packet
1086 * @data: the packet data
1087 * @tlen: the packet length
1088 * @qp: the QP for this packet
1089 * @opcode: the opcode for this packet
1090 * @psn: the packet sequence number for this packet
1091 * @hdrsize: the header length
1092 * @pmtu: the path MTU
1093 * @header_in_data: true if part of the header data is in the data buffer
1095 * This is called from ipath_rc_rcv() to process an incoming RC response
1096 * packet for the given QP.
1097 * Called at interrupt level.
1099 static inline void ipath_rc_rcv_resp(struct ipath_ibdev *dev,
1100 struct ipath_other_headers *ohdr,
1101 void *data, u32 tlen,
1102 struct ipath_qp *qp,
1103 u32 opcode,
1104 u32 psn, u32 hdrsize, u32 pmtu,
1105 int header_in_data)
1107 struct ipath_swqe *wqe;
1108 unsigned long flags;
1109 struct ib_wc wc;
1110 int diff;
1111 u32 pad;
1112 u32 aeth;
1113 u64 val;
1115 spin_lock_irqsave(&qp->s_lock, flags);
1117 /* Ignore invalid responses. */
1118 if (ipath_cmp24(psn, qp->s_next_psn) >= 0)
1119 goto ack_done;
1121 /* Ignore duplicate responses. */
1122 diff = ipath_cmp24(psn, qp->s_last_psn);
1123 if (unlikely(diff <= 0)) {
1124 /* Update credits for "ghost" ACKs */
1125 if (diff == 0 && opcode == OP(ACKNOWLEDGE)) {
1126 if (!header_in_data)
1127 aeth = be32_to_cpu(ohdr->u.aeth);
1128 else {
1129 aeth = be32_to_cpu(((__be32 *) data)[0]);
1130 data += sizeof(__be32);
1132 if ((aeth >> 29) == 0)
1133 ipath_get_credit(qp, aeth);
1135 goto ack_done;
1138 if (unlikely(qp->s_last == qp->s_tail))
1139 goto ack_done;
1140 wqe = get_swqe_ptr(qp, qp->s_last);
1142 switch (opcode) {
1143 case OP(ACKNOWLEDGE):
1144 case OP(ATOMIC_ACKNOWLEDGE):
1145 case OP(RDMA_READ_RESPONSE_FIRST):
1146 if (!header_in_data)
1147 aeth = be32_to_cpu(ohdr->u.aeth);
1148 else {
1149 aeth = be32_to_cpu(((__be32 *) data)[0]);
1150 data += sizeof(__be32);
1152 if (opcode == OP(ATOMIC_ACKNOWLEDGE)) {
1153 if (!header_in_data) {
1154 __be32 *p = ohdr->u.at.atomic_ack_eth;
1156 val = ((u64) be32_to_cpu(p[0]) << 32) |
1157 be32_to_cpu(p[1]);
1158 } else
1159 val = be64_to_cpu(((__be64 *) data)[0]);
1160 } else
1161 val = 0;
1162 if (!do_rc_ack(qp, aeth, psn, opcode, val) ||
1163 opcode != OP(RDMA_READ_RESPONSE_FIRST))
1164 goto ack_done;
1165 hdrsize += 4;
1166 wqe = get_swqe_ptr(qp, qp->s_last);
1167 if (unlikely(wqe->wr.opcode != IB_WR_RDMA_READ))
1168 goto ack_op_err;
1170 * If this is a response to a resent RDMA read, we
1171 * have to be careful to copy the data to the right
1172 * location.
1174 qp->s_rdma_read_len = restart_sge(&qp->s_rdma_read_sge,
1175 wqe, psn, pmtu);
1176 goto read_middle;
1178 case OP(RDMA_READ_RESPONSE_MIDDLE):
1179 /* no AETH, no ACK */
1180 if (unlikely(ipath_cmp24(psn, qp->s_last_psn + 1))) {
1181 dev->n_rdma_seq++;
1182 ipath_restart_rc(qp, qp->s_last_psn + 1, &wc);
1183 goto ack_done;
1185 if (unlikely(wqe->wr.opcode != IB_WR_RDMA_READ))
1186 goto ack_op_err;
1187 read_middle:
1188 if (unlikely(tlen != (hdrsize + pmtu + 4)))
1189 goto ack_len_err;
1190 if (unlikely(pmtu >= qp->s_rdma_read_len))
1191 goto ack_len_err;
1193 /* We got a response so update the timeout. */
1194 spin_lock(&dev->pending_lock);
1195 if (qp->s_rnr_timeout == 0 && !list_empty(&qp->timerwait))
1196 list_move_tail(&qp->timerwait,
1197 &dev->pending[dev->pending_index]);
1198 spin_unlock(&dev->pending_lock);
1200 if (opcode == OP(RDMA_READ_RESPONSE_MIDDLE))
1201 qp->s_retry = qp->s_retry_cnt;
1204 * Update the RDMA receive state but do the copy w/o
1205 * holding the locks and blocking interrupts.
1207 qp->s_rdma_read_len -= pmtu;
1208 update_last_psn(qp, psn);
1209 spin_unlock_irqrestore(&qp->s_lock, flags);
1210 ipath_copy_sge(&qp->s_rdma_read_sge, data, pmtu);
1211 goto bail;
1213 case OP(RDMA_READ_RESPONSE_ONLY):
1214 if (!header_in_data)
1215 aeth = be32_to_cpu(ohdr->u.aeth);
1216 else
1217 aeth = be32_to_cpu(((__be32 *) data)[0]);
1218 if (!do_rc_ack(qp, aeth, psn, opcode, 0))
1219 goto ack_done;
1220 /* Get the number of bytes the message was padded by. */
1221 pad = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3;
1223 * Check that the data size is >= 0 && <= pmtu.
1224 * Remember to account for the AETH header (4) and
1225 * ICRC (4).
1227 if (unlikely(tlen < (hdrsize + pad + 8)))
1228 goto ack_len_err;
1230 * If this is a response to a resent RDMA read, we
1231 * have to be careful to copy the data to the right
1232 * location.
1234 wqe = get_swqe_ptr(qp, qp->s_last);
1235 qp->s_rdma_read_len = restart_sge(&qp->s_rdma_read_sge,
1236 wqe, psn, pmtu);
1237 goto read_last;
1239 case OP(RDMA_READ_RESPONSE_LAST):
1240 /* ACKs READ req. */
1241 if (unlikely(ipath_cmp24(psn, qp->s_last_psn + 1))) {
1242 dev->n_rdma_seq++;
1243 ipath_restart_rc(qp, qp->s_last_psn + 1, &wc);
1244 goto ack_done;
1246 if (unlikely(wqe->wr.opcode != IB_WR_RDMA_READ))
1247 goto ack_op_err;
1248 /* Get the number of bytes the message was padded by. */
1249 pad = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3;
1251 * Check that the data size is >= 1 && <= pmtu.
1252 * Remember to account for the AETH header (4) and
1253 * ICRC (4).
1255 if (unlikely(tlen <= (hdrsize + pad + 8)))
1256 goto ack_len_err;
1257 read_last:
1258 tlen -= hdrsize + pad + 8;
1259 if (unlikely(tlen != qp->s_rdma_read_len))
1260 goto ack_len_err;
1261 if (!header_in_data)
1262 aeth = be32_to_cpu(ohdr->u.aeth);
1263 else {
1264 aeth = be32_to_cpu(((__be32 *) data)[0]);
1265 data += sizeof(__be32);
1267 ipath_copy_sge(&qp->s_rdma_read_sge, data, tlen);
1268 (void) do_rc_ack(qp, aeth, psn,
1269 OP(RDMA_READ_RESPONSE_LAST), 0);
1270 goto ack_done;
1273 ack_done:
1274 spin_unlock_irqrestore(&qp->s_lock, flags);
1275 goto bail;
1277 ack_op_err:
1278 wc.status = IB_WC_LOC_QP_OP_ERR;
1279 goto ack_err;
1281 ack_len_err:
1282 wc.status = IB_WC_LOC_LEN_ERR;
1283 ack_err:
1284 wc.wr_id = wqe->wr.wr_id;
1285 wc.opcode = ib_ipath_wc_opcode[wqe->wr.opcode];
1286 wc.vendor_err = 0;
1287 wc.byte_len = 0;
1288 wc.imm_data = 0;
1289 wc.qp = &qp->ibqp;
1290 wc.src_qp = qp->remote_qpn;
1291 wc.wc_flags = 0;
1292 wc.pkey_index = 0;
1293 wc.slid = qp->remote_ah_attr.dlid;
1294 wc.sl = qp->remote_ah_attr.sl;
1295 wc.dlid_path_bits = 0;
1296 wc.port_num = 0;
1297 ipath_sqerror_qp(qp, &wc);
1298 spin_unlock_irqrestore(&qp->s_lock, flags);
1299 bail:
1300 return;
1304 * ipath_rc_rcv_error - process an incoming duplicate or error RC packet
1305 * @dev: the device this packet came in on
1306 * @ohdr: the other headers for this packet
1307 * @data: the packet data
1308 * @qp: the QP for this packet
1309 * @opcode: the opcode for this packet
1310 * @psn: the packet sequence number for this packet
1311 * @diff: the difference between the PSN and the expected PSN
1312 * @header_in_data: true if part of the header data is in the data buffer
1314 * This is called from ipath_rc_rcv() to process an unexpected
1315 * incoming RC packet for the given QP.
1316 * Called at interrupt level.
1317 * Return 1 if no more processing is needed; otherwise return 0 to
1318 * schedule a response to be sent.
1320 static inline int ipath_rc_rcv_error(struct ipath_ibdev *dev,
1321 struct ipath_other_headers *ohdr,
1322 void *data,
1323 struct ipath_qp *qp,
1324 u32 opcode,
1325 u32 psn,
1326 int diff,
1327 int header_in_data)
1329 struct ipath_ack_entry *e;
1330 u8 i, prev;
1331 int old_req;
1332 unsigned long flags;
1334 if (diff > 0) {
1336 * Packet sequence error.
1337 * A NAK will ACK earlier sends and RDMA writes.
1338 * Don't queue the NAK if we already sent one.
1340 if (!qp->r_nak_state) {
1341 qp->r_nak_state = IB_NAK_PSN_ERROR;
1342 /* Use the expected PSN. */
1343 qp->r_ack_psn = qp->r_psn;
1344 goto send_ack;
1346 goto done;
1350 * Handle a duplicate request. Don't re-execute SEND, RDMA
1351 * write or atomic op. Don't NAK errors, just silently drop
1352 * the duplicate request. Note that r_sge, r_len, and
1353 * r_rcv_len may be in use so don't modify them.
1355 * We are supposed to ACK the earliest duplicate PSN but we
1356 * can coalesce an outstanding duplicate ACK. We have to
1357 * send the earliest so that RDMA reads can be restarted at
1358 * the requester's expected PSN.
1360 * First, find where this duplicate PSN falls within the
1361 * ACKs previously sent.
1363 psn &= IPATH_PSN_MASK;
1364 e = NULL;
1365 old_req = 1;
1366 spin_lock_irqsave(&qp->s_lock, flags);
1367 for (i = qp->r_head_ack_queue; ; i = prev) {
1368 if (i == qp->s_tail_ack_queue)
1369 old_req = 0;
1370 if (i)
1371 prev = i - 1;
1372 else
1373 prev = IPATH_MAX_RDMA_ATOMIC;
1374 if (prev == qp->r_head_ack_queue) {
1375 e = NULL;
1376 break;
1378 e = &qp->s_ack_queue[prev];
1379 if (!e->opcode) {
1380 e = NULL;
1381 break;
1383 if (ipath_cmp24(psn, e->psn) >= 0) {
1384 if (prev == qp->s_tail_ack_queue)
1385 old_req = 0;
1386 break;
1389 switch (opcode) {
1390 case OP(RDMA_READ_REQUEST): {
1391 struct ib_reth *reth;
1392 u32 offset;
1393 u32 len;
1396 * If we didn't find the RDMA read request in the ack queue,
1397 * or the send tasklet is already backed up to send an
1398 * earlier entry, we can ignore this request.
1400 if (!e || e->opcode != OP(RDMA_READ_REQUEST) || old_req)
1401 goto unlock_done;
1402 /* RETH comes after BTH */
1403 if (!header_in_data)
1404 reth = &ohdr->u.rc.reth;
1405 else {
1406 reth = (struct ib_reth *)data;
1407 data += sizeof(*reth);
1410 * Address range must be a subset of the original
1411 * request and start on pmtu boundaries.
1412 * We reuse the old ack_queue slot since the requester
1413 * should not back up and request an earlier PSN for the
1414 * same request.
1416 offset = ((psn - e->psn) & IPATH_PSN_MASK) *
1417 ib_mtu_enum_to_int(qp->path_mtu);
1418 len = be32_to_cpu(reth->length);
1419 if (unlikely(offset + len > e->rdma_sge.sge.sge_length))
1420 goto unlock_done;
1421 if (len != 0) {
1422 u32 rkey = be32_to_cpu(reth->rkey);
1423 u64 vaddr = be64_to_cpu(reth->vaddr);
1424 int ok;
1426 ok = ipath_rkey_ok(qp, &e->rdma_sge,
1427 len, vaddr, rkey,
1428 IB_ACCESS_REMOTE_READ);
1429 if (unlikely(!ok))
1430 goto unlock_done;
1431 } else {
1432 e->rdma_sge.sg_list = NULL;
1433 e->rdma_sge.num_sge = 0;
1434 e->rdma_sge.sge.mr = NULL;
1435 e->rdma_sge.sge.vaddr = NULL;
1436 e->rdma_sge.sge.length = 0;
1437 e->rdma_sge.sge.sge_length = 0;
1439 e->psn = psn;
1440 qp->s_ack_state = OP(ACKNOWLEDGE);
1441 qp->s_tail_ack_queue = prev;
1442 break;
1445 case OP(COMPARE_SWAP):
1446 case OP(FETCH_ADD): {
1448 * If we didn't find the atomic request in the ack queue
1449 * or the send tasklet is already backed up to send an
1450 * earlier entry, we can ignore this request.
1452 if (!e || e->opcode != (u8) opcode || old_req)
1453 goto unlock_done;
1454 qp->s_ack_state = OP(ACKNOWLEDGE);
1455 qp->s_tail_ack_queue = prev;
1456 break;
1459 default:
1460 if (old_req)
1461 goto unlock_done;
1463 * Resend the most recent ACK if this request is
1464 * after all the previous RDMA reads and atomics.
1466 if (i == qp->r_head_ack_queue) {
1467 spin_unlock_irqrestore(&qp->s_lock, flags);
1468 qp->r_nak_state = 0;
1469 qp->r_ack_psn = qp->r_psn - 1;
1470 goto send_ack;
1473 * Try to send a simple ACK to work around a Mellanox bug
1474 * which doesn't accept a RDMA read response or atomic
1475 * response as an ACK for earlier SENDs or RDMA writes.
1477 if (qp->r_head_ack_queue == qp->s_tail_ack_queue &&
1478 !(qp->s_flags & IPATH_S_ACK_PENDING) &&
1479 qp->s_ack_state == OP(ACKNOWLEDGE)) {
1480 spin_unlock_irqrestore(&qp->s_lock, flags);
1481 qp->r_nak_state = 0;
1482 qp->r_ack_psn = qp->s_ack_queue[i].psn - 1;
1483 goto send_ack;
1486 * Resend the RDMA read or atomic op which
1487 * ACKs this duplicate request.
1489 qp->s_ack_state = OP(ACKNOWLEDGE);
1490 qp->s_tail_ack_queue = i;
1491 break;
1493 qp->r_nak_state = 0;
1494 tasklet_hi_schedule(&qp->s_task);
1496 unlock_done:
1497 spin_unlock_irqrestore(&qp->s_lock, flags);
1498 done:
1499 return 1;
1501 send_ack:
1502 return 0;
1505 static void ipath_rc_error(struct ipath_qp *qp, enum ib_wc_status err)
1507 unsigned long flags;
1508 int lastwqe;
1510 spin_lock_irqsave(&qp->s_lock, flags);
1511 qp->state = IB_QPS_ERR;
1512 lastwqe = ipath_error_qp(qp, err);
1513 spin_unlock_irqrestore(&qp->s_lock, flags);
1515 if (lastwqe) {
1516 struct ib_event ev;
1518 ev.device = qp->ibqp.device;
1519 ev.element.qp = &qp->ibqp;
1520 ev.event = IB_EVENT_QP_LAST_WQE_REACHED;
1521 qp->ibqp.event_handler(&ev, qp->ibqp.qp_context);
1525 static inline void ipath_update_ack_queue(struct ipath_qp *qp, unsigned n)
1527 unsigned long flags;
1528 unsigned next;
1530 next = n + 1;
1531 if (next > IPATH_MAX_RDMA_ATOMIC)
1532 next = 0;
1533 spin_lock_irqsave(&qp->s_lock, flags);
1534 if (n == qp->s_tail_ack_queue) {
1535 qp->s_tail_ack_queue = next;
1536 qp->s_ack_state = OP(ACKNOWLEDGE);
1538 spin_unlock_irqrestore(&qp->s_lock, flags);
1542 * ipath_rc_rcv - process an incoming RC packet
1543 * @dev: the device this packet came in on
1544 * @hdr: the header of this packet
1545 * @has_grh: true if the header has a GRH
1546 * @data: the packet data
1547 * @tlen: the packet length
1548 * @qp: the QP for this packet
1550 * This is called from ipath_qp_rcv() to process an incoming RC packet
1551 * for the given QP.
1552 * Called at interrupt level.
1554 void ipath_rc_rcv(struct ipath_ibdev *dev, struct ipath_ib_header *hdr,
1555 int has_grh, void *data, u32 tlen, struct ipath_qp *qp)
1557 struct ipath_other_headers *ohdr;
1558 u32 opcode;
1559 u32 hdrsize;
1560 u32 psn;
1561 u32 pad;
1562 struct ib_wc wc;
1563 u32 pmtu = ib_mtu_enum_to_int(qp->path_mtu);
1564 int diff;
1565 struct ib_reth *reth;
1566 int header_in_data;
1568 /* Validate the SLID. See Ch. 9.6.1.5 */
1569 if (unlikely(be16_to_cpu(hdr->lrh[3]) != qp->remote_ah_attr.dlid))
1570 goto done;
1572 /* Check for GRH */
1573 if (!has_grh) {
1574 ohdr = &hdr->u.oth;
1575 hdrsize = 8 + 12; /* LRH + BTH */
1576 psn = be32_to_cpu(ohdr->bth[2]);
1577 header_in_data = 0;
1578 } else {
1579 ohdr = &hdr->u.l.oth;
1580 hdrsize = 8 + 40 + 12; /* LRH + GRH + BTH */
1582 * The header with GRH is 60 bytes and the core driver sets
1583 * the eager header buffer size to 56 bytes so the last 4
1584 * bytes of the BTH header (PSN) is in the data buffer.
1586 header_in_data = dev->dd->ipath_rcvhdrentsize == 16;
1587 if (header_in_data) {
1588 psn = be32_to_cpu(((__be32 *) data)[0]);
1589 data += sizeof(__be32);
1590 } else
1591 psn = be32_to_cpu(ohdr->bth[2]);
1595 * Process responses (ACKs) before anything else. Note that the
1596 * packet sequence number will be for something in the send work
1597 * queue rather than the expected receive packet sequence number.
1598 * In other words, this QP is the requester.
1600 opcode = be32_to_cpu(ohdr->bth[0]) >> 24;
1601 if (opcode >= OP(RDMA_READ_RESPONSE_FIRST) &&
1602 opcode <= OP(ATOMIC_ACKNOWLEDGE)) {
1603 ipath_rc_rcv_resp(dev, ohdr, data, tlen, qp, opcode, psn,
1604 hdrsize, pmtu, header_in_data);
1605 goto done;
1608 /* Compute 24 bits worth of difference. */
1609 diff = ipath_cmp24(psn, qp->r_psn);
1610 if (unlikely(diff)) {
1611 if (ipath_rc_rcv_error(dev, ohdr, data, qp, opcode,
1612 psn, diff, header_in_data))
1613 goto done;
1614 goto send_ack;
1617 /* Check for opcode sequence errors. */
1618 switch (qp->r_state) {
1619 case OP(SEND_FIRST):
1620 case OP(SEND_MIDDLE):
1621 if (opcode == OP(SEND_MIDDLE) ||
1622 opcode == OP(SEND_LAST) ||
1623 opcode == OP(SEND_LAST_WITH_IMMEDIATE))
1624 break;
1625 nack_inv:
1626 ipath_rc_error(qp, IB_WC_REM_INV_REQ_ERR);
1627 qp->r_nak_state = IB_NAK_INVALID_REQUEST;
1628 qp->r_ack_psn = qp->r_psn;
1629 goto send_ack;
1631 case OP(RDMA_WRITE_FIRST):
1632 case OP(RDMA_WRITE_MIDDLE):
1633 if (opcode == OP(RDMA_WRITE_MIDDLE) ||
1634 opcode == OP(RDMA_WRITE_LAST) ||
1635 opcode == OP(RDMA_WRITE_LAST_WITH_IMMEDIATE))
1636 break;
1637 goto nack_inv;
1639 default:
1640 if (opcode == OP(SEND_MIDDLE) ||
1641 opcode == OP(SEND_LAST) ||
1642 opcode == OP(SEND_LAST_WITH_IMMEDIATE) ||
1643 opcode == OP(RDMA_WRITE_MIDDLE) ||
1644 opcode == OP(RDMA_WRITE_LAST) ||
1645 opcode == OP(RDMA_WRITE_LAST_WITH_IMMEDIATE))
1646 goto nack_inv;
1648 * Note that it is up to the requester to not send a new
1649 * RDMA read or atomic operation before receiving an ACK
1650 * for the previous operation.
1652 break;
1655 wc.imm_data = 0;
1656 wc.wc_flags = 0;
1658 /* OK, process the packet. */
1659 switch (opcode) {
1660 case OP(SEND_FIRST):
1661 if (!ipath_get_rwqe(qp, 0)) {
1662 rnr_nak:
1663 qp->r_nak_state = IB_RNR_NAK | qp->r_min_rnr_timer;
1664 qp->r_ack_psn = qp->r_psn;
1665 goto send_ack;
1667 qp->r_rcv_len = 0;
1668 /* FALLTHROUGH */
1669 case OP(SEND_MIDDLE):
1670 case OP(RDMA_WRITE_MIDDLE):
1671 send_middle:
1672 /* Check for invalid length PMTU or posted rwqe len. */
1673 if (unlikely(tlen != (hdrsize + pmtu + 4)))
1674 goto nack_inv;
1675 qp->r_rcv_len += pmtu;
1676 if (unlikely(qp->r_rcv_len > qp->r_len))
1677 goto nack_inv;
1678 ipath_copy_sge(&qp->r_sge, data, pmtu);
1679 break;
1681 case OP(RDMA_WRITE_LAST_WITH_IMMEDIATE):
1682 /* consume RWQE */
1683 if (!ipath_get_rwqe(qp, 1))
1684 goto rnr_nak;
1685 goto send_last_imm;
1687 case OP(SEND_ONLY):
1688 case OP(SEND_ONLY_WITH_IMMEDIATE):
1689 if (!ipath_get_rwqe(qp, 0))
1690 goto rnr_nak;
1691 qp->r_rcv_len = 0;
1692 if (opcode == OP(SEND_ONLY))
1693 goto send_last;
1694 /* FALLTHROUGH */
1695 case OP(SEND_LAST_WITH_IMMEDIATE):
1696 send_last_imm:
1697 if (header_in_data) {
1698 wc.imm_data = *(__be32 *) data;
1699 data += sizeof(__be32);
1700 } else {
1701 /* Immediate data comes after BTH */
1702 wc.imm_data = ohdr->u.imm_data;
1704 hdrsize += 4;
1705 wc.wc_flags = IB_WC_WITH_IMM;
1706 /* FALLTHROUGH */
1707 case OP(SEND_LAST):
1708 case OP(RDMA_WRITE_LAST):
1709 send_last:
1710 /* Get the number of bytes the message was padded by. */
1711 pad = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3;
1712 /* Check for invalid length. */
1713 /* XXX LAST len should be >= 1 */
1714 if (unlikely(tlen < (hdrsize + pad + 4)))
1715 goto nack_inv;
1716 /* Don't count the CRC. */
1717 tlen -= (hdrsize + pad + 4);
1718 wc.byte_len = tlen + qp->r_rcv_len;
1719 if (unlikely(wc.byte_len > qp->r_len))
1720 goto nack_inv;
1721 ipath_copy_sge(&qp->r_sge, data, tlen);
1722 qp->r_msn++;
1723 if (!qp->r_wrid_valid)
1724 break;
1725 qp->r_wrid_valid = 0;
1726 wc.wr_id = qp->r_wr_id;
1727 wc.status = IB_WC_SUCCESS;
1728 wc.opcode = IB_WC_RECV;
1729 wc.vendor_err = 0;
1730 wc.qp = &qp->ibqp;
1731 wc.src_qp = qp->remote_qpn;
1732 wc.pkey_index = 0;
1733 wc.slid = qp->remote_ah_attr.dlid;
1734 wc.sl = qp->remote_ah_attr.sl;
1735 wc.dlid_path_bits = 0;
1736 wc.port_num = 0;
1737 /* Signal completion event if the solicited bit is set. */
1738 ipath_cq_enter(to_icq(qp->ibqp.recv_cq), &wc,
1739 (ohdr->bth[0] &
1740 __constant_cpu_to_be32(1 << 23)) != 0);
1741 break;
1743 case OP(RDMA_WRITE_FIRST):
1744 case OP(RDMA_WRITE_ONLY):
1745 case OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE):
1746 if (unlikely(!(qp->qp_access_flags &
1747 IB_ACCESS_REMOTE_WRITE)))
1748 goto nack_inv;
1749 /* consume RWQE */
1750 /* RETH comes after BTH */
1751 if (!header_in_data)
1752 reth = &ohdr->u.rc.reth;
1753 else {
1754 reth = (struct ib_reth *)data;
1755 data += sizeof(*reth);
1757 hdrsize += sizeof(*reth);
1758 qp->r_len = be32_to_cpu(reth->length);
1759 qp->r_rcv_len = 0;
1760 if (qp->r_len != 0) {
1761 u32 rkey = be32_to_cpu(reth->rkey);
1762 u64 vaddr = be64_to_cpu(reth->vaddr);
1763 int ok;
1765 /* Check rkey & NAK */
1766 ok = ipath_rkey_ok(qp, &qp->r_sge,
1767 qp->r_len, vaddr, rkey,
1768 IB_ACCESS_REMOTE_WRITE);
1769 if (unlikely(!ok))
1770 goto nack_acc;
1771 } else {
1772 qp->r_sge.sg_list = NULL;
1773 qp->r_sge.sge.mr = NULL;
1774 qp->r_sge.sge.vaddr = NULL;
1775 qp->r_sge.sge.length = 0;
1776 qp->r_sge.sge.sge_length = 0;
1778 if (opcode == OP(RDMA_WRITE_FIRST))
1779 goto send_middle;
1780 else if (opcode == OP(RDMA_WRITE_ONLY))
1781 goto send_last;
1782 if (!ipath_get_rwqe(qp, 1))
1783 goto rnr_nak;
1784 goto send_last_imm;
1786 case OP(RDMA_READ_REQUEST): {
1787 struct ipath_ack_entry *e;
1788 u32 len;
1789 u8 next;
1791 if (unlikely(!(qp->qp_access_flags &
1792 IB_ACCESS_REMOTE_READ)))
1793 goto nack_inv;
1794 next = qp->r_head_ack_queue + 1;
1795 if (next > IPATH_MAX_RDMA_ATOMIC)
1796 next = 0;
1797 if (unlikely(next == qp->s_tail_ack_queue)) {
1798 if (!qp->s_ack_queue[next].sent)
1799 goto nack_inv;
1800 ipath_update_ack_queue(qp, next);
1802 e = &qp->s_ack_queue[qp->r_head_ack_queue];
1803 /* RETH comes after BTH */
1804 if (!header_in_data)
1805 reth = &ohdr->u.rc.reth;
1806 else {
1807 reth = (struct ib_reth *)data;
1808 data += sizeof(*reth);
1810 len = be32_to_cpu(reth->length);
1811 if (len) {
1812 u32 rkey = be32_to_cpu(reth->rkey);
1813 u64 vaddr = be64_to_cpu(reth->vaddr);
1814 int ok;
1816 /* Check rkey & NAK */
1817 ok = ipath_rkey_ok(qp, &e->rdma_sge, len, vaddr,
1818 rkey, IB_ACCESS_REMOTE_READ);
1819 if (unlikely(!ok))
1820 goto nack_acc;
1822 * Update the next expected PSN. We add 1 later
1823 * below, so only add the remainder here.
1825 if (len > pmtu)
1826 qp->r_psn += (len - 1) / pmtu;
1827 } else {
1828 e->rdma_sge.sg_list = NULL;
1829 e->rdma_sge.num_sge = 0;
1830 e->rdma_sge.sge.mr = NULL;
1831 e->rdma_sge.sge.vaddr = NULL;
1832 e->rdma_sge.sge.length = 0;
1833 e->rdma_sge.sge.sge_length = 0;
1835 e->opcode = opcode;
1836 e->sent = 0;
1837 e->psn = psn;
1839 * We need to increment the MSN here instead of when we
1840 * finish sending the result since a duplicate request would
1841 * increment it more than once.
1843 qp->r_msn++;
1844 qp->r_psn++;
1845 qp->r_state = opcode;
1846 qp->r_nak_state = 0;
1847 barrier();
1848 qp->r_head_ack_queue = next;
1850 /* Call ipath_do_rc_send() in another thread. */
1851 tasklet_hi_schedule(&qp->s_task);
1853 goto done;
1856 case OP(COMPARE_SWAP):
1857 case OP(FETCH_ADD): {
1858 struct ib_atomic_eth *ateth;
1859 struct ipath_ack_entry *e;
1860 u64 vaddr;
1861 atomic64_t *maddr;
1862 u64 sdata;
1863 u32 rkey;
1864 u8 next;
1866 if (unlikely(!(qp->qp_access_flags &
1867 IB_ACCESS_REMOTE_ATOMIC)))
1868 goto nack_inv;
1869 next = qp->r_head_ack_queue + 1;
1870 if (next > IPATH_MAX_RDMA_ATOMIC)
1871 next = 0;
1872 if (unlikely(next == qp->s_tail_ack_queue)) {
1873 if (!qp->s_ack_queue[next].sent)
1874 goto nack_inv;
1875 ipath_update_ack_queue(qp, next);
1877 if (!header_in_data)
1878 ateth = &ohdr->u.atomic_eth;
1879 else
1880 ateth = (struct ib_atomic_eth *)data;
1881 vaddr = ((u64) be32_to_cpu(ateth->vaddr[0]) << 32) |
1882 be32_to_cpu(ateth->vaddr[1]);
1883 if (unlikely(vaddr & (sizeof(u64) - 1)))
1884 goto nack_inv;
1885 rkey = be32_to_cpu(ateth->rkey);
1886 /* Check rkey & NAK */
1887 if (unlikely(!ipath_rkey_ok(qp, &qp->r_sge,
1888 sizeof(u64), vaddr, rkey,
1889 IB_ACCESS_REMOTE_ATOMIC)))
1890 goto nack_acc;
1891 /* Perform atomic OP and save result. */
1892 maddr = (atomic64_t *) qp->r_sge.sge.vaddr;
1893 sdata = be64_to_cpu(ateth->swap_data);
1894 e = &qp->s_ack_queue[qp->r_head_ack_queue];
1895 e->atomic_data = (opcode == OP(FETCH_ADD)) ?
1896 (u64) atomic64_add_return(sdata, maddr) - sdata :
1897 (u64) cmpxchg((u64 *) qp->r_sge.sge.vaddr,
1898 be64_to_cpu(ateth->compare_data),
1899 sdata);
1900 e->opcode = opcode;
1901 e->sent = 0;
1902 e->psn = psn & IPATH_PSN_MASK;
1903 qp->r_msn++;
1904 qp->r_psn++;
1905 qp->r_state = opcode;
1906 qp->r_nak_state = 0;
1907 barrier();
1908 qp->r_head_ack_queue = next;
1910 /* Call ipath_do_rc_send() in another thread. */
1911 tasklet_hi_schedule(&qp->s_task);
1913 goto done;
1916 default:
1917 /* NAK unknown opcodes. */
1918 goto nack_inv;
1920 qp->r_psn++;
1921 qp->r_state = opcode;
1922 qp->r_ack_psn = psn;
1923 qp->r_nak_state = 0;
1924 /* Send an ACK if requested or required. */
1925 if (psn & (1 << 31))
1926 goto send_ack;
1927 goto done;
1929 nack_acc:
1930 ipath_rc_error(qp, IB_WC_REM_ACCESS_ERR);
1931 qp->r_nak_state = IB_NAK_REMOTE_ACCESS_ERROR;
1932 qp->r_ack_psn = qp->r_psn;
1934 send_ack:
1935 send_rc_ack(qp);
1937 done:
1938 return;