xref: /linux/drivers/infiniband/hw/qib/qib_rc.c (revision e5c86679d5e864947a52fb31e45a425dea3e7fa9)
1 /*
2  * Copyright (c) 2006, 2007, 2008, 2009 QLogic Corporation. All rights reserved.
3  * Copyright (c) 2005, 2006 PathScale, Inc. All rights reserved.
4  *
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:
10  *
11  *     Redistribution and use in source and binary forms, with or
12  *     without modification, are permitted provided that the following
13  *     conditions are met:
14  *
15  *      - Redistributions of source code must retain the above
16  *        copyright notice, this list of conditions and the following
17  *        disclaimer.
18  *
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.
23  *
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.
32  */
33 
34 #include <linux/io.h>
35 
36 #include "qib.h"
37 
38 /* cut down ridiculously long IB macro names */
39 #define OP(x) IB_OPCODE_RC_##x
40 
41 
42 static u32 restart_sge(struct rvt_sge_state *ss, struct rvt_swqe *wqe,
43 		       u32 psn, u32 pmtu)
44 {
45 	u32 len;
46 
47 	len = ((psn - wqe->psn) & QIB_PSN_MASK) * pmtu;
48 	ss->sge = wqe->sg_list[0];
49 	ss->sg_list = wqe->sg_list + 1;
50 	ss->num_sge = wqe->wr.num_sge;
51 	ss->total_len = wqe->length;
52 	rvt_skip_sge(ss, len, false);
53 	return wqe->length - len;
54 }
55 
56 /**
57  * qib_make_rc_ack - construct a response packet (ACK, NAK, or RDMA read)
58  * @dev: the device for this QP
59  * @qp: a pointer to the QP
60  * @ohdr: a pointer to the IB header being constructed
61  * @pmtu: the path MTU
62  *
63  * Return 1 if constructed; otherwise, return 0.
64  * Note that we are in the responder's side of the QP context.
65  * Note the QP s_lock must be held.
66  */
67 static int qib_make_rc_ack(struct qib_ibdev *dev, struct rvt_qp *qp,
68 			   struct ib_other_headers *ohdr, u32 pmtu)
69 {
70 	struct rvt_ack_entry *e;
71 	u32 hwords;
72 	u32 len;
73 	u32 bth0;
74 	u32 bth2;
75 
76 	/* Don't send an ACK if we aren't supposed to. */
77 	if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK))
78 		goto bail;
79 
80 	/* header size in 32-bit words LRH+BTH = (8+12)/4. */
81 	hwords = 5;
82 
83 	switch (qp->s_ack_state) {
84 	case OP(RDMA_READ_RESPONSE_LAST):
85 	case OP(RDMA_READ_RESPONSE_ONLY):
86 		e = &qp->s_ack_queue[qp->s_tail_ack_queue];
87 		if (e->rdma_sge.mr) {
88 			rvt_put_mr(e->rdma_sge.mr);
89 			e->rdma_sge.mr = NULL;
90 		}
91 		/* FALLTHROUGH */
92 	case OP(ATOMIC_ACKNOWLEDGE):
93 		/*
94 		 * We can increment the tail pointer now that the last
95 		 * response has been sent instead of only being
96 		 * constructed.
97 		 */
98 		if (++qp->s_tail_ack_queue > QIB_MAX_RDMA_ATOMIC)
99 			qp->s_tail_ack_queue = 0;
100 		/* FALLTHROUGH */
101 	case OP(SEND_ONLY):
102 	case OP(ACKNOWLEDGE):
103 		/* Check for no next entry in the queue. */
104 		if (qp->r_head_ack_queue == qp->s_tail_ack_queue) {
105 			if (qp->s_flags & RVT_S_ACK_PENDING)
106 				goto normal;
107 			goto bail;
108 		}
109 
110 		e = &qp->s_ack_queue[qp->s_tail_ack_queue];
111 		if (e->opcode == OP(RDMA_READ_REQUEST)) {
112 			/*
113 			 * If a RDMA read response is being resent and
114 			 * we haven't seen the duplicate request yet,
115 			 * then stop sending the remaining responses the
116 			 * responder has seen until the requester resends it.
117 			 */
118 			len = e->rdma_sge.sge_length;
119 			if (len && !e->rdma_sge.mr) {
120 				qp->s_tail_ack_queue = qp->r_head_ack_queue;
121 				goto bail;
122 			}
123 			/* Copy SGE state in case we need to resend */
124 			qp->s_rdma_mr = e->rdma_sge.mr;
125 			if (qp->s_rdma_mr)
126 				rvt_get_mr(qp->s_rdma_mr);
127 			qp->s_ack_rdma_sge.sge = e->rdma_sge;
128 			qp->s_ack_rdma_sge.num_sge = 1;
129 			qp->s_cur_sge = &qp->s_ack_rdma_sge;
130 			if (len > pmtu) {
131 				len = pmtu;
132 				qp->s_ack_state = OP(RDMA_READ_RESPONSE_FIRST);
133 			} else {
134 				qp->s_ack_state = OP(RDMA_READ_RESPONSE_ONLY);
135 				e->sent = 1;
136 			}
137 			ohdr->u.aeth = rvt_compute_aeth(qp);
138 			hwords++;
139 			qp->s_ack_rdma_psn = e->psn;
140 			bth2 = qp->s_ack_rdma_psn++ & QIB_PSN_MASK;
141 		} else {
142 			/* COMPARE_SWAP or FETCH_ADD */
143 			qp->s_cur_sge = NULL;
144 			len = 0;
145 			qp->s_ack_state = OP(ATOMIC_ACKNOWLEDGE);
146 			ohdr->u.at.aeth = rvt_compute_aeth(qp);
147 			ib_u64_put(e->atomic_data, &ohdr->u.at.atomic_ack_eth);
148 			hwords += sizeof(ohdr->u.at) / sizeof(u32);
149 			bth2 = e->psn & QIB_PSN_MASK;
150 			e->sent = 1;
151 		}
152 		bth0 = qp->s_ack_state << 24;
153 		break;
154 
155 	case OP(RDMA_READ_RESPONSE_FIRST):
156 		qp->s_ack_state = OP(RDMA_READ_RESPONSE_MIDDLE);
157 		/* FALLTHROUGH */
158 	case OP(RDMA_READ_RESPONSE_MIDDLE):
159 		qp->s_cur_sge = &qp->s_ack_rdma_sge;
160 		qp->s_rdma_mr = qp->s_ack_rdma_sge.sge.mr;
161 		if (qp->s_rdma_mr)
162 			rvt_get_mr(qp->s_rdma_mr);
163 		len = qp->s_ack_rdma_sge.sge.sge_length;
164 		if (len > pmtu)
165 			len = pmtu;
166 		else {
167 			ohdr->u.aeth = rvt_compute_aeth(qp);
168 			hwords++;
169 			qp->s_ack_state = OP(RDMA_READ_RESPONSE_LAST);
170 			e = &qp->s_ack_queue[qp->s_tail_ack_queue];
171 			e->sent = 1;
172 		}
173 		bth0 = qp->s_ack_state << 24;
174 		bth2 = qp->s_ack_rdma_psn++ & QIB_PSN_MASK;
175 		break;
176 
177 	default:
178 normal:
179 		/*
180 		 * Send a regular ACK.
181 		 * Set the s_ack_state so we wait until after sending
182 		 * the ACK before setting s_ack_state to ACKNOWLEDGE
183 		 * (see above).
184 		 */
185 		qp->s_ack_state = OP(SEND_ONLY);
186 		qp->s_flags &= ~RVT_S_ACK_PENDING;
187 		qp->s_cur_sge = NULL;
188 		if (qp->s_nak_state)
189 			ohdr->u.aeth =
190 				cpu_to_be32((qp->r_msn & IB_MSN_MASK) |
191 					    (qp->s_nak_state <<
192 					     IB_AETH_CREDIT_SHIFT));
193 		else
194 			ohdr->u.aeth = rvt_compute_aeth(qp);
195 		hwords++;
196 		len = 0;
197 		bth0 = OP(ACKNOWLEDGE) << 24;
198 		bth2 = qp->s_ack_psn & QIB_PSN_MASK;
199 	}
200 	qp->s_rdma_ack_cnt++;
201 	qp->s_hdrwords = hwords;
202 	qp->s_cur_size = len;
203 	qib_make_ruc_header(qp, ohdr, bth0, bth2);
204 	return 1;
205 
206 bail:
207 	qp->s_ack_state = OP(ACKNOWLEDGE);
208 	qp->s_flags &= ~(RVT_S_RESP_PENDING | RVT_S_ACK_PENDING);
209 	return 0;
210 }
211 
212 /**
213  * qib_make_rc_req - construct a request packet (SEND, RDMA r/w, ATOMIC)
214  * @qp: a pointer to the QP
215  *
216  * Assumes the s_lock is held.
217  *
218  * Return 1 if constructed; otherwise, return 0.
219  */
220 int qib_make_rc_req(struct rvt_qp *qp, unsigned long *flags)
221 {
222 	struct qib_qp_priv *priv = qp->priv;
223 	struct qib_ibdev *dev = to_idev(qp->ibqp.device);
224 	struct ib_other_headers *ohdr;
225 	struct rvt_sge_state *ss;
226 	struct rvt_swqe *wqe;
227 	u32 hwords;
228 	u32 len;
229 	u32 bth0;
230 	u32 bth2;
231 	u32 pmtu = qp->pmtu;
232 	char newreq;
233 	int ret = 0;
234 	int delta;
235 
236 	ohdr = &priv->s_hdr->u.oth;
237 	if (qp->remote_ah_attr.ah_flags & IB_AH_GRH)
238 		ohdr = &priv->s_hdr->u.l.oth;
239 
240 	/* Sending responses has higher priority over sending requests. */
241 	if ((qp->s_flags & RVT_S_RESP_PENDING) &&
242 	    qib_make_rc_ack(dev, qp, ohdr, pmtu))
243 		goto done;
244 
245 	if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_SEND_OK)) {
246 		if (!(ib_rvt_state_ops[qp->state] & RVT_FLUSH_SEND))
247 			goto bail;
248 		/* We are in the error state, flush the work request. */
249 		smp_read_barrier_depends(); /* see post_one_send() */
250 		if (qp->s_last == READ_ONCE(qp->s_head))
251 			goto bail;
252 		/* If DMAs are in progress, we can't flush immediately. */
253 		if (atomic_read(&priv->s_dma_busy)) {
254 			qp->s_flags |= RVT_S_WAIT_DMA;
255 			goto bail;
256 		}
257 		wqe = rvt_get_swqe_ptr(qp, qp->s_last);
258 		qib_send_complete(qp, wqe, qp->s_last != qp->s_acked ?
259 			IB_WC_SUCCESS : IB_WC_WR_FLUSH_ERR);
260 		/* will get called again */
261 		goto done;
262 	}
263 
264 	if (qp->s_flags & (RVT_S_WAIT_RNR | RVT_S_WAIT_ACK))
265 		goto bail;
266 
267 	if (qib_cmp24(qp->s_psn, qp->s_sending_hpsn) <= 0) {
268 		if (qib_cmp24(qp->s_sending_psn, qp->s_sending_hpsn) <= 0) {
269 			qp->s_flags |= RVT_S_WAIT_PSN;
270 			goto bail;
271 		}
272 		qp->s_sending_psn = qp->s_psn;
273 		qp->s_sending_hpsn = qp->s_psn - 1;
274 	}
275 
276 	/* header size in 32-bit words LRH+BTH = (8+12)/4. */
277 	hwords = 5;
278 	bth0 = 0;
279 
280 	/* Send a request. */
281 	wqe = rvt_get_swqe_ptr(qp, qp->s_cur);
282 	switch (qp->s_state) {
283 	default:
284 		if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_NEXT_SEND_OK))
285 			goto bail;
286 		/*
287 		 * Resend an old request or start a new one.
288 		 *
289 		 * We keep track of the current SWQE so that
290 		 * we don't reset the "furthest progress" state
291 		 * if we need to back up.
292 		 */
293 		newreq = 0;
294 		if (qp->s_cur == qp->s_tail) {
295 			/* Check if send work queue is empty. */
296 			smp_read_barrier_depends(); /* see post_one_send() */
297 			if (qp->s_tail == READ_ONCE(qp->s_head))
298 				goto bail;
299 			/*
300 			 * If a fence is requested, wait for previous
301 			 * RDMA read and atomic operations to finish.
302 			 */
303 			if ((wqe->wr.send_flags & IB_SEND_FENCE) &&
304 			    qp->s_num_rd_atomic) {
305 				qp->s_flags |= RVT_S_WAIT_FENCE;
306 				goto bail;
307 			}
308 			newreq = 1;
309 			qp->s_psn = wqe->psn;
310 		}
311 		/*
312 		 * Note that we have to be careful not to modify the
313 		 * original work request since we may need to resend
314 		 * it.
315 		 */
316 		len = wqe->length;
317 		ss = &qp->s_sge;
318 		bth2 = qp->s_psn & QIB_PSN_MASK;
319 		switch (wqe->wr.opcode) {
320 		case IB_WR_SEND:
321 		case IB_WR_SEND_WITH_IMM:
322 			/* If no credit, return. */
323 			if (!(qp->s_flags & RVT_S_UNLIMITED_CREDIT) &&
324 			    rvt_cmp_msn(wqe->ssn, qp->s_lsn + 1) > 0) {
325 				qp->s_flags |= RVT_S_WAIT_SSN_CREDIT;
326 				goto bail;
327 			}
328 			if (len > pmtu) {
329 				qp->s_state = OP(SEND_FIRST);
330 				len = pmtu;
331 				break;
332 			}
333 			if (wqe->wr.opcode == IB_WR_SEND)
334 				qp->s_state = OP(SEND_ONLY);
335 			else {
336 				qp->s_state = OP(SEND_ONLY_WITH_IMMEDIATE);
337 				/* Immediate data comes after the BTH */
338 				ohdr->u.imm_data = wqe->wr.ex.imm_data;
339 				hwords += 1;
340 			}
341 			if (wqe->wr.send_flags & IB_SEND_SOLICITED)
342 				bth0 |= IB_BTH_SOLICITED;
343 			bth2 |= IB_BTH_REQ_ACK;
344 			if (++qp->s_cur == qp->s_size)
345 				qp->s_cur = 0;
346 			break;
347 
348 		case IB_WR_RDMA_WRITE:
349 			if (newreq && !(qp->s_flags & RVT_S_UNLIMITED_CREDIT))
350 				qp->s_lsn++;
351 			/* FALLTHROUGH */
352 		case IB_WR_RDMA_WRITE_WITH_IMM:
353 			/* If no credit, return. */
354 			if (!(qp->s_flags & RVT_S_UNLIMITED_CREDIT) &&
355 			    rvt_cmp_msn(wqe->ssn, qp->s_lsn + 1) > 0) {
356 				qp->s_flags |= RVT_S_WAIT_SSN_CREDIT;
357 				goto bail;
358 			}
359 
360 			ohdr->u.rc.reth.vaddr =
361 				cpu_to_be64(wqe->rdma_wr.remote_addr);
362 			ohdr->u.rc.reth.rkey =
363 				cpu_to_be32(wqe->rdma_wr.rkey);
364 			ohdr->u.rc.reth.length = cpu_to_be32(len);
365 			hwords += sizeof(struct ib_reth) / sizeof(u32);
366 			if (len > pmtu) {
367 				qp->s_state = OP(RDMA_WRITE_FIRST);
368 				len = pmtu;
369 				break;
370 			}
371 			if (wqe->rdma_wr.wr.opcode == IB_WR_RDMA_WRITE)
372 				qp->s_state = OP(RDMA_WRITE_ONLY);
373 			else {
374 				qp->s_state = OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE);
375 				/* Immediate data comes after RETH */
376 				ohdr->u.rc.imm_data =
377 					wqe->rdma_wr.wr.ex.imm_data;
378 				hwords += 1;
379 				if (wqe->rdma_wr.wr.send_flags & IB_SEND_SOLICITED)
380 					bth0 |= IB_BTH_SOLICITED;
381 			}
382 			bth2 |= IB_BTH_REQ_ACK;
383 			if (++qp->s_cur == qp->s_size)
384 				qp->s_cur = 0;
385 			break;
386 
387 		case IB_WR_RDMA_READ:
388 			/*
389 			 * Don't allow more operations to be started
390 			 * than the QP limits allow.
391 			 */
392 			if (newreq) {
393 				if (qp->s_num_rd_atomic >=
394 				    qp->s_max_rd_atomic) {
395 					qp->s_flags |= RVT_S_WAIT_RDMAR;
396 					goto bail;
397 				}
398 				qp->s_num_rd_atomic++;
399 				if (!(qp->s_flags & RVT_S_UNLIMITED_CREDIT))
400 					qp->s_lsn++;
401 			}
402 
403 			ohdr->u.rc.reth.vaddr =
404 				cpu_to_be64(wqe->rdma_wr.remote_addr);
405 			ohdr->u.rc.reth.rkey =
406 				cpu_to_be32(wqe->rdma_wr.rkey);
407 			ohdr->u.rc.reth.length = cpu_to_be32(len);
408 			qp->s_state = OP(RDMA_READ_REQUEST);
409 			hwords += sizeof(ohdr->u.rc.reth) / sizeof(u32);
410 			ss = NULL;
411 			len = 0;
412 			bth2 |= IB_BTH_REQ_ACK;
413 			if (++qp->s_cur == qp->s_size)
414 				qp->s_cur = 0;
415 			break;
416 
417 		case IB_WR_ATOMIC_CMP_AND_SWP:
418 		case IB_WR_ATOMIC_FETCH_AND_ADD:
419 			/*
420 			 * Don't allow more operations to be started
421 			 * than the QP limits allow.
422 			 */
423 			if (newreq) {
424 				if (qp->s_num_rd_atomic >=
425 				    qp->s_max_rd_atomic) {
426 					qp->s_flags |= RVT_S_WAIT_RDMAR;
427 					goto bail;
428 				}
429 				qp->s_num_rd_atomic++;
430 				if (!(qp->s_flags & RVT_S_UNLIMITED_CREDIT))
431 					qp->s_lsn++;
432 			}
433 			if (wqe->atomic_wr.wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
434 				qp->s_state = OP(COMPARE_SWAP);
435 				put_ib_ateth_swap(wqe->atomic_wr.swap,
436 						  &ohdr->u.atomic_eth);
437 				put_ib_ateth_swap(wqe->atomic_wr.compare_add,
438 						  &ohdr->u.atomic_eth);
439 			} else {
440 				qp->s_state = OP(FETCH_ADD);
441 				put_ib_ateth_swap(wqe->atomic_wr.compare_add,
442 						  &ohdr->u.atomic_eth);
443 				put_ib_ateth_swap(0, &ohdr->u.atomic_eth);
444 			}
445 			put_ib_ateth_vaddr(wqe->atomic_wr.remote_addr,
446 					   &ohdr->u.atomic_eth);
447 			ohdr->u.atomic_eth.rkey = cpu_to_be32(
448 				wqe->atomic_wr.rkey);
449 			hwords += sizeof(struct ib_atomic_eth) / sizeof(u32);
450 			ss = NULL;
451 			len = 0;
452 			bth2 |= IB_BTH_REQ_ACK;
453 			if (++qp->s_cur == qp->s_size)
454 				qp->s_cur = 0;
455 			break;
456 
457 		default:
458 			goto bail;
459 		}
460 		qp->s_sge.sge = wqe->sg_list[0];
461 		qp->s_sge.sg_list = wqe->sg_list + 1;
462 		qp->s_sge.num_sge = wqe->wr.num_sge;
463 		qp->s_sge.total_len = wqe->length;
464 		qp->s_len = wqe->length;
465 		if (newreq) {
466 			qp->s_tail++;
467 			if (qp->s_tail >= qp->s_size)
468 				qp->s_tail = 0;
469 		}
470 		if (wqe->wr.opcode == IB_WR_RDMA_READ)
471 			qp->s_psn = wqe->lpsn + 1;
472 		else
473 			qp->s_psn++;
474 		break;
475 
476 	case OP(RDMA_READ_RESPONSE_FIRST):
477 		/*
478 		 * qp->s_state is normally set to the opcode of the
479 		 * last packet constructed for new requests and therefore
480 		 * is never set to RDMA read response.
481 		 * RDMA_READ_RESPONSE_FIRST is used by the ACK processing
482 		 * thread to indicate a SEND needs to be restarted from an
483 		 * earlier PSN without interferring with the sending thread.
484 		 * See qib_restart_rc().
485 		 */
486 		qp->s_len = restart_sge(&qp->s_sge, wqe, qp->s_psn, pmtu);
487 		/* FALLTHROUGH */
488 	case OP(SEND_FIRST):
489 		qp->s_state = OP(SEND_MIDDLE);
490 		/* FALLTHROUGH */
491 	case OP(SEND_MIDDLE):
492 		bth2 = qp->s_psn++ & QIB_PSN_MASK;
493 		ss = &qp->s_sge;
494 		len = qp->s_len;
495 		if (len > pmtu) {
496 			len = pmtu;
497 			break;
498 		}
499 		if (wqe->wr.opcode == IB_WR_SEND)
500 			qp->s_state = OP(SEND_LAST);
501 		else {
502 			qp->s_state = OP(SEND_LAST_WITH_IMMEDIATE);
503 			/* Immediate data comes after the BTH */
504 			ohdr->u.imm_data = wqe->wr.ex.imm_data;
505 			hwords += 1;
506 		}
507 		if (wqe->wr.send_flags & IB_SEND_SOLICITED)
508 			bth0 |= IB_BTH_SOLICITED;
509 		bth2 |= IB_BTH_REQ_ACK;
510 		qp->s_cur++;
511 		if (qp->s_cur >= qp->s_size)
512 			qp->s_cur = 0;
513 		break;
514 
515 	case OP(RDMA_READ_RESPONSE_LAST):
516 		/*
517 		 * qp->s_state is normally set to the opcode of the
518 		 * last packet constructed for new requests and therefore
519 		 * is never set to RDMA read response.
520 		 * RDMA_READ_RESPONSE_LAST is used by the ACK processing
521 		 * thread to indicate a RDMA write needs to be restarted from
522 		 * an earlier PSN without interferring with the sending thread.
523 		 * See qib_restart_rc().
524 		 */
525 		qp->s_len = restart_sge(&qp->s_sge, wqe, qp->s_psn, pmtu);
526 		/* FALLTHROUGH */
527 	case OP(RDMA_WRITE_FIRST):
528 		qp->s_state = OP(RDMA_WRITE_MIDDLE);
529 		/* FALLTHROUGH */
530 	case OP(RDMA_WRITE_MIDDLE):
531 		bth2 = qp->s_psn++ & QIB_PSN_MASK;
532 		ss = &qp->s_sge;
533 		len = qp->s_len;
534 		if (len > pmtu) {
535 			len = pmtu;
536 			break;
537 		}
538 		if (wqe->wr.opcode == IB_WR_RDMA_WRITE)
539 			qp->s_state = OP(RDMA_WRITE_LAST);
540 		else {
541 			qp->s_state = OP(RDMA_WRITE_LAST_WITH_IMMEDIATE);
542 			/* Immediate data comes after the BTH */
543 			ohdr->u.imm_data = wqe->wr.ex.imm_data;
544 			hwords += 1;
545 			if (wqe->wr.send_flags & IB_SEND_SOLICITED)
546 				bth0 |= IB_BTH_SOLICITED;
547 		}
548 		bth2 |= IB_BTH_REQ_ACK;
549 		qp->s_cur++;
550 		if (qp->s_cur >= qp->s_size)
551 			qp->s_cur = 0;
552 		break;
553 
554 	case OP(RDMA_READ_RESPONSE_MIDDLE):
555 		/*
556 		 * qp->s_state is normally set to the opcode of the
557 		 * last packet constructed for new requests and therefore
558 		 * is never set to RDMA read response.
559 		 * RDMA_READ_RESPONSE_MIDDLE is used by the ACK processing
560 		 * thread to indicate a RDMA read needs to be restarted from
561 		 * an earlier PSN without interferring with the sending thread.
562 		 * See qib_restart_rc().
563 		 */
564 		len = ((qp->s_psn - wqe->psn) & QIB_PSN_MASK) * pmtu;
565 		ohdr->u.rc.reth.vaddr =
566 			cpu_to_be64(wqe->rdma_wr.remote_addr + len);
567 		ohdr->u.rc.reth.rkey =
568 			cpu_to_be32(wqe->rdma_wr.rkey);
569 		ohdr->u.rc.reth.length = cpu_to_be32(wqe->length - len);
570 		qp->s_state = OP(RDMA_READ_REQUEST);
571 		hwords += sizeof(ohdr->u.rc.reth) / sizeof(u32);
572 		bth2 = (qp->s_psn & QIB_PSN_MASK) | IB_BTH_REQ_ACK;
573 		qp->s_psn = wqe->lpsn + 1;
574 		ss = NULL;
575 		len = 0;
576 		qp->s_cur++;
577 		if (qp->s_cur == qp->s_size)
578 			qp->s_cur = 0;
579 		break;
580 	}
581 	qp->s_sending_hpsn = bth2;
582 	delta = (((int) bth2 - (int) wqe->psn) << 8) >> 8;
583 	if (delta && delta % QIB_PSN_CREDIT == 0)
584 		bth2 |= IB_BTH_REQ_ACK;
585 	if (qp->s_flags & RVT_S_SEND_ONE) {
586 		qp->s_flags &= ~RVT_S_SEND_ONE;
587 		qp->s_flags |= RVT_S_WAIT_ACK;
588 		bth2 |= IB_BTH_REQ_ACK;
589 	}
590 	qp->s_len -= len;
591 	qp->s_hdrwords = hwords;
592 	qp->s_cur_sge = ss;
593 	qp->s_cur_size = len;
594 	qib_make_ruc_header(qp, ohdr, bth0 | (qp->s_state << 24), bth2);
595 done:
596 	return 1;
597 bail:
598 	qp->s_flags &= ~RVT_S_BUSY;
599 	return ret;
600 }
601 
602 /**
603  * qib_send_rc_ack - Construct an ACK packet and send it
604  * @qp: a pointer to the QP
605  *
606  * This is called from qib_rc_rcv() and qib_kreceive().
607  * Note that RDMA reads and atomics are handled in the
608  * send side QP state and tasklet.
609  */
610 void qib_send_rc_ack(struct rvt_qp *qp)
611 {
612 	struct qib_devdata *dd = dd_from_ibdev(qp->ibqp.device);
613 	struct qib_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num);
614 	struct qib_pportdata *ppd = ppd_from_ibp(ibp);
615 	u64 pbc;
616 	u16 lrh0;
617 	u32 bth0;
618 	u32 hwords;
619 	u32 pbufn;
620 	u32 __iomem *piobuf;
621 	struct ib_header hdr;
622 	struct ib_other_headers *ohdr;
623 	u32 control;
624 	unsigned long flags;
625 
626 	spin_lock_irqsave(&qp->s_lock, flags);
627 
628 	if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK))
629 		goto unlock;
630 
631 	/* Don't send ACK or NAK if a RDMA read or atomic is pending. */
632 	if ((qp->s_flags & RVT_S_RESP_PENDING) || qp->s_rdma_ack_cnt)
633 		goto queue_ack;
634 
635 	/* Construct the header with s_lock held so APM doesn't change it. */
636 	ohdr = &hdr.u.oth;
637 	lrh0 = QIB_LRH_BTH;
638 	/* header size in 32-bit words LRH+BTH+AETH = (8+12+4)/4. */
639 	hwords = 6;
640 	if (unlikely(qp->remote_ah_attr.ah_flags & IB_AH_GRH)) {
641 		hwords += qib_make_grh(ibp, &hdr.u.l.grh,
642 				       &qp->remote_ah_attr.grh, hwords, 0);
643 		ohdr = &hdr.u.l.oth;
644 		lrh0 = QIB_LRH_GRH;
645 	}
646 	/* read pkey_index w/o lock (its atomic) */
647 	bth0 = qib_get_pkey(ibp, qp->s_pkey_index) | (OP(ACKNOWLEDGE) << 24);
648 	if (qp->s_mig_state == IB_MIG_MIGRATED)
649 		bth0 |= IB_BTH_MIG_REQ;
650 	if (qp->r_nak_state)
651 		ohdr->u.aeth = cpu_to_be32((qp->r_msn & IB_MSN_MASK) |
652 					    (qp->r_nak_state <<
653 					     IB_AETH_CREDIT_SHIFT));
654 	else
655 		ohdr->u.aeth = rvt_compute_aeth(qp);
656 	lrh0 |= ibp->sl_to_vl[qp->remote_ah_attr.sl] << 12 |
657 		qp->remote_ah_attr.sl << 4;
658 	hdr.lrh[0] = cpu_to_be16(lrh0);
659 	hdr.lrh[1] = cpu_to_be16(qp->remote_ah_attr.dlid);
660 	hdr.lrh[2] = cpu_to_be16(hwords + SIZE_OF_CRC);
661 	hdr.lrh[3] = cpu_to_be16(ppd->lid | qp->remote_ah_attr.src_path_bits);
662 	ohdr->bth[0] = cpu_to_be32(bth0);
663 	ohdr->bth[1] = cpu_to_be32(qp->remote_qpn);
664 	ohdr->bth[2] = cpu_to_be32(qp->r_ack_psn & QIB_PSN_MASK);
665 
666 	spin_unlock_irqrestore(&qp->s_lock, flags);
667 
668 	/* Don't try to send ACKs if the link isn't ACTIVE */
669 	if (!(ppd->lflags & QIBL_LINKACTIVE))
670 		goto done;
671 
672 	control = dd->f_setpbc_control(ppd, hwords + SIZE_OF_CRC,
673 				       qp->s_srate, lrh0 >> 12);
674 	/* length is + 1 for the control dword */
675 	pbc = ((u64) control << 32) | (hwords + 1);
676 
677 	piobuf = dd->f_getsendbuf(ppd, pbc, &pbufn);
678 	if (!piobuf) {
679 		/*
680 		 * We are out of PIO buffers at the moment.
681 		 * Pass responsibility for sending the ACK to the
682 		 * send tasklet so that when a PIO buffer becomes
683 		 * available, the ACK is sent ahead of other outgoing
684 		 * packets.
685 		 */
686 		spin_lock_irqsave(&qp->s_lock, flags);
687 		goto queue_ack;
688 	}
689 
690 	/*
691 	 * Write the pbc.
692 	 * We have to flush after the PBC for correctness
693 	 * on some cpus or WC buffer can be written out of order.
694 	 */
695 	writeq(pbc, piobuf);
696 
697 	if (dd->flags & QIB_PIO_FLUSH_WC) {
698 		u32 *hdrp = (u32 *) &hdr;
699 
700 		qib_flush_wc();
701 		qib_pio_copy(piobuf + 2, hdrp, hwords - 1);
702 		qib_flush_wc();
703 		__raw_writel(hdrp[hwords - 1], piobuf + hwords + 1);
704 	} else
705 		qib_pio_copy(piobuf + 2, (u32 *) &hdr, hwords);
706 
707 	if (dd->flags & QIB_USE_SPCL_TRIG) {
708 		u32 spcl_off = (pbufn >= dd->piobcnt2k) ? 2047 : 1023;
709 
710 		qib_flush_wc();
711 		__raw_writel(0xaebecede, piobuf + spcl_off);
712 	}
713 
714 	qib_flush_wc();
715 	qib_sendbuf_done(dd, pbufn);
716 
717 	this_cpu_inc(ibp->pmastats->n_unicast_xmit);
718 	goto done;
719 
720 queue_ack:
721 	if (ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK) {
722 		this_cpu_inc(*ibp->rvp.rc_qacks);
723 		qp->s_flags |= RVT_S_ACK_PENDING | RVT_S_RESP_PENDING;
724 		qp->s_nak_state = qp->r_nak_state;
725 		qp->s_ack_psn = qp->r_ack_psn;
726 
727 		/* Schedule the send tasklet. */
728 		qib_schedule_send(qp);
729 	}
730 unlock:
731 	spin_unlock_irqrestore(&qp->s_lock, flags);
732 done:
733 	return;
734 }
735 
736 /**
737  * reset_psn - reset the QP state to send starting from PSN
738  * @qp: the QP
739  * @psn: the packet sequence number to restart at
740  *
741  * This is called from qib_rc_rcv() to process an incoming RC ACK
742  * for the given QP.
743  * Called at interrupt level with the QP s_lock held.
744  */
745 static void reset_psn(struct rvt_qp *qp, u32 psn)
746 {
747 	u32 n = qp->s_acked;
748 	struct rvt_swqe *wqe = rvt_get_swqe_ptr(qp, n);
749 	u32 opcode;
750 
751 	qp->s_cur = n;
752 
753 	/*
754 	 * If we are starting the request from the beginning,
755 	 * let the normal send code handle initialization.
756 	 */
757 	if (qib_cmp24(psn, wqe->psn) <= 0) {
758 		qp->s_state = OP(SEND_LAST);
759 		goto done;
760 	}
761 
762 	/* Find the work request opcode corresponding to the given PSN. */
763 	opcode = wqe->wr.opcode;
764 	for (;;) {
765 		int diff;
766 
767 		if (++n == qp->s_size)
768 			n = 0;
769 		if (n == qp->s_tail)
770 			break;
771 		wqe = rvt_get_swqe_ptr(qp, n);
772 		diff = qib_cmp24(psn, wqe->psn);
773 		if (diff < 0)
774 			break;
775 		qp->s_cur = n;
776 		/*
777 		 * If we are starting the request from the beginning,
778 		 * let the normal send code handle initialization.
779 		 */
780 		if (diff == 0) {
781 			qp->s_state = OP(SEND_LAST);
782 			goto done;
783 		}
784 		opcode = wqe->wr.opcode;
785 	}
786 
787 	/*
788 	 * Set the state to restart in the middle of a request.
789 	 * Don't change the s_sge, s_cur_sge, or s_cur_size.
790 	 * See qib_make_rc_req().
791 	 */
792 	switch (opcode) {
793 	case IB_WR_SEND:
794 	case IB_WR_SEND_WITH_IMM:
795 		qp->s_state = OP(RDMA_READ_RESPONSE_FIRST);
796 		break;
797 
798 	case IB_WR_RDMA_WRITE:
799 	case IB_WR_RDMA_WRITE_WITH_IMM:
800 		qp->s_state = OP(RDMA_READ_RESPONSE_LAST);
801 		break;
802 
803 	case IB_WR_RDMA_READ:
804 		qp->s_state = OP(RDMA_READ_RESPONSE_MIDDLE);
805 		break;
806 
807 	default:
808 		/*
809 		 * This case shouldn't happen since its only
810 		 * one PSN per req.
811 		 */
812 		qp->s_state = OP(SEND_LAST);
813 	}
814 done:
815 	qp->s_psn = psn;
816 	/*
817 	 * Set RVT_S_WAIT_PSN as qib_rc_complete() may start the timer
818 	 * asynchronously before the send tasklet can get scheduled.
819 	 * Doing it in qib_make_rc_req() is too late.
820 	 */
821 	if ((qib_cmp24(qp->s_psn, qp->s_sending_hpsn) <= 0) &&
822 	    (qib_cmp24(qp->s_sending_psn, qp->s_sending_hpsn) <= 0))
823 		qp->s_flags |= RVT_S_WAIT_PSN;
824 }
825 
826 /*
827  * Back up requester to resend the last un-ACKed request.
828  * The QP r_lock and s_lock should be held and interrupts disabled.
829  */
830 void qib_restart_rc(struct rvt_qp *qp, u32 psn, int wait)
831 {
832 	struct rvt_swqe *wqe = rvt_get_swqe_ptr(qp, qp->s_acked);
833 	struct qib_ibport *ibp;
834 
835 	if (qp->s_retry == 0) {
836 		if (qp->s_mig_state == IB_MIG_ARMED) {
837 			qib_migrate_qp(qp);
838 			qp->s_retry = qp->s_retry_cnt;
839 		} else if (qp->s_last == qp->s_acked) {
840 			qib_send_complete(qp, wqe, IB_WC_RETRY_EXC_ERR);
841 			rvt_error_qp(qp, IB_WC_WR_FLUSH_ERR);
842 			return;
843 		} else /* XXX need to handle delayed completion */
844 			return;
845 	} else
846 		qp->s_retry--;
847 
848 	ibp = to_iport(qp->ibqp.device, qp->port_num);
849 	if (wqe->wr.opcode == IB_WR_RDMA_READ)
850 		ibp->rvp.n_rc_resends++;
851 	else
852 		ibp->rvp.n_rc_resends += (qp->s_psn - psn) & QIB_PSN_MASK;
853 
854 	qp->s_flags &= ~(RVT_S_WAIT_FENCE | RVT_S_WAIT_RDMAR |
855 			 RVT_S_WAIT_SSN_CREDIT | RVT_S_WAIT_PSN |
856 			 RVT_S_WAIT_ACK);
857 	if (wait)
858 		qp->s_flags |= RVT_S_SEND_ONE;
859 	reset_psn(qp, psn);
860 }
861 
862 /*
863  * Set qp->s_sending_psn to the next PSN after the given one.
864  * This would be psn+1 except when RDMA reads are present.
865  */
866 static void reset_sending_psn(struct rvt_qp *qp, u32 psn)
867 {
868 	struct rvt_swqe *wqe;
869 	u32 n = qp->s_last;
870 
871 	/* Find the work request corresponding to the given PSN. */
872 	for (;;) {
873 		wqe = rvt_get_swqe_ptr(qp, n);
874 		if (qib_cmp24(psn, wqe->lpsn) <= 0) {
875 			if (wqe->wr.opcode == IB_WR_RDMA_READ)
876 				qp->s_sending_psn = wqe->lpsn + 1;
877 			else
878 				qp->s_sending_psn = psn + 1;
879 			break;
880 		}
881 		if (++n == qp->s_size)
882 			n = 0;
883 		if (n == qp->s_tail)
884 			break;
885 	}
886 }
887 
888 /*
889  * This should be called with the QP s_lock held and interrupts disabled.
890  */
891 void qib_rc_send_complete(struct rvt_qp *qp, struct ib_header *hdr)
892 {
893 	struct ib_other_headers *ohdr;
894 	struct rvt_swqe *wqe;
895 	u32 opcode;
896 	u32 psn;
897 
898 	if (!(ib_rvt_state_ops[qp->state] & RVT_SEND_OR_FLUSH_OR_RECV_OK))
899 		return;
900 
901 	/* Find out where the BTH is */
902 	if ((be16_to_cpu(hdr->lrh[0]) & 3) == QIB_LRH_BTH)
903 		ohdr = &hdr->u.oth;
904 	else
905 		ohdr = &hdr->u.l.oth;
906 
907 	opcode = be32_to_cpu(ohdr->bth[0]) >> 24;
908 	if (opcode >= OP(RDMA_READ_RESPONSE_FIRST) &&
909 	    opcode <= OP(ATOMIC_ACKNOWLEDGE)) {
910 		WARN_ON(!qp->s_rdma_ack_cnt);
911 		qp->s_rdma_ack_cnt--;
912 		return;
913 	}
914 
915 	psn = be32_to_cpu(ohdr->bth[2]);
916 	reset_sending_psn(qp, psn);
917 
918 	/*
919 	 * Start timer after a packet requesting an ACK has been sent and
920 	 * there are still requests that haven't been acked.
921 	 */
922 	if ((psn & IB_BTH_REQ_ACK) && qp->s_acked != qp->s_tail &&
923 	    !(qp->s_flags & (RVT_S_TIMER | RVT_S_WAIT_RNR | RVT_S_WAIT_PSN)) &&
924 	    (ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK))
925 		rvt_add_retry_timer(qp);
926 
927 	while (qp->s_last != qp->s_acked) {
928 		u32 s_last;
929 
930 		wqe = rvt_get_swqe_ptr(qp, qp->s_last);
931 		if (qib_cmp24(wqe->lpsn, qp->s_sending_psn) >= 0 &&
932 		    qib_cmp24(qp->s_sending_psn, qp->s_sending_hpsn) <= 0)
933 			break;
934 		s_last = qp->s_last;
935 		if (++s_last >= qp->s_size)
936 			s_last = 0;
937 		qp->s_last = s_last;
938 		/* see post_send() */
939 		barrier();
940 		rvt_put_swqe(wqe);
941 		rvt_qp_swqe_complete(qp, wqe, IB_WC_SUCCESS);
942 	}
943 	/*
944 	 * If we were waiting for sends to complete before resending,
945 	 * and they are now complete, restart sending.
946 	 */
947 	if (qp->s_flags & RVT_S_WAIT_PSN &&
948 	    qib_cmp24(qp->s_sending_psn, qp->s_sending_hpsn) > 0) {
949 		qp->s_flags &= ~RVT_S_WAIT_PSN;
950 		qp->s_sending_psn = qp->s_psn;
951 		qp->s_sending_hpsn = qp->s_psn - 1;
952 		qib_schedule_send(qp);
953 	}
954 }
955 
956 static inline void update_last_psn(struct rvt_qp *qp, u32 psn)
957 {
958 	qp->s_last_psn = psn;
959 }
960 
961 /*
962  * Generate a SWQE completion.
963  * This is similar to qib_send_complete but has to check to be sure
964  * that the SGEs are not being referenced if the SWQE is being resent.
965  */
966 static struct rvt_swqe *do_rc_completion(struct rvt_qp *qp,
967 					 struct rvt_swqe *wqe,
968 					 struct qib_ibport *ibp)
969 {
970 	/*
971 	 * Don't decrement refcount and don't generate a
972 	 * completion if the SWQE is being resent until the send
973 	 * is finished.
974 	 */
975 	if (qib_cmp24(wqe->lpsn, qp->s_sending_psn) < 0 ||
976 	    qib_cmp24(qp->s_sending_psn, qp->s_sending_hpsn) > 0) {
977 		u32 s_last;
978 
979 		rvt_put_swqe(wqe);
980 		s_last = qp->s_last;
981 		if (++s_last >= qp->s_size)
982 			s_last = 0;
983 		qp->s_last = s_last;
984 		/* see post_send() */
985 		barrier();
986 		rvt_qp_swqe_complete(qp, wqe, IB_WC_SUCCESS);
987 	} else
988 		this_cpu_inc(*ibp->rvp.rc_delayed_comp);
989 
990 	qp->s_retry = qp->s_retry_cnt;
991 	update_last_psn(qp, wqe->lpsn);
992 
993 	/*
994 	 * If we are completing a request which is in the process of
995 	 * being resent, we can stop resending it since we know the
996 	 * responder has already seen it.
997 	 */
998 	if (qp->s_acked == qp->s_cur) {
999 		if (++qp->s_cur >= qp->s_size)
1000 			qp->s_cur = 0;
1001 		qp->s_acked = qp->s_cur;
1002 		wqe = rvt_get_swqe_ptr(qp, qp->s_cur);
1003 		if (qp->s_acked != qp->s_tail) {
1004 			qp->s_state = OP(SEND_LAST);
1005 			qp->s_psn = wqe->psn;
1006 		}
1007 	} else {
1008 		if (++qp->s_acked >= qp->s_size)
1009 			qp->s_acked = 0;
1010 		if (qp->state == IB_QPS_SQD && qp->s_acked == qp->s_cur)
1011 			qp->s_draining = 0;
1012 		wqe = rvt_get_swqe_ptr(qp, qp->s_acked);
1013 	}
1014 	return wqe;
1015 }
1016 
1017 /**
1018  * do_rc_ack - process an incoming RC ACK
1019  * @qp: the QP the ACK came in on
1020  * @psn: the packet sequence number of the ACK
1021  * @opcode: the opcode of the request that resulted in the ACK
1022  *
1023  * This is called from qib_rc_rcv_resp() to process an incoming RC ACK
1024  * for the given QP.
1025  * Called at interrupt level with the QP s_lock held.
1026  * Returns 1 if OK, 0 if current operation should be aborted (NAK).
1027  */
1028 static int do_rc_ack(struct rvt_qp *qp, u32 aeth, u32 psn, int opcode,
1029 		     u64 val, struct qib_ctxtdata *rcd)
1030 {
1031 	struct qib_ibport *ibp;
1032 	enum ib_wc_status status;
1033 	struct rvt_swqe *wqe;
1034 	int ret = 0;
1035 	u32 ack_psn;
1036 	int diff;
1037 
1038 	/*
1039 	 * Note that NAKs implicitly ACK outstanding SEND and RDMA write
1040 	 * requests and implicitly NAK RDMA read and atomic requests issued
1041 	 * before the NAK'ed request.  The MSN won't include the NAK'ed
1042 	 * request but will include an ACK'ed request(s).
1043 	 */
1044 	ack_psn = psn;
1045 	if (aeth >> IB_AETH_NAK_SHIFT)
1046 		ack_psn--;
1047 	wqe = rvt_get_swqe_ptr(qp, qp->s_acked);
1048 	ibp = to_iport(qp->ibqp.device, qp->port_num);
1049 
1050 	/*
1051 	 * The MSN might be for a later WQE than the PSN indicates so
1052 	 * only complete WQEs that the PSN finishes.
1053 	 */
1054 	while ((diff = qib_cmp24(ack_psn, wqe->lpsn)) >= 0) {
1055 		/*
1056 		 * RDMA_READ_RESPONSE_ONLY is a special case since
1057 		 * we want to generate completion events for everything
1058 		 * before the RDMA read, copy the data, then generate
1059 		 * the completion for the read.
1060 		 */
1061 		if (wqe->wr.opcode == IB_WR_RDMA_READ &&
1062 		    opcode == OP(RDMA_READ_RESPONSE_ONLY) &&
1063 		    diff == 0) {
1064 			ret = 1;
1065 			goto bail;
1066 		}
1067 		/*
1068 		 * If this request is a RDMA read or atomic, and the ACK is
1069 		 * for a later operation, this ACK NAKs the RDMA read or
1070 		 * atomic.  In other words, only a RDMA_READ_LAST or ONLY
1071 		 * can ACK a RDMA read and likewise for atomic ops.  Note
1072 		 * that the NAK case can only happen if relaxed ordering is
1073 		 * used and requests are sent after an RDMA read or atomic
1074 		 * is sent but before the response is received.
1075 		 */
1076 		if ((wqe->wr.opcode == IB_WR_RDMA_READ &&
1077 		     (opcode != OP(RDMA_READ_RESPONSE_LAST) || diff != 0)) ||
1078 		    ((wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
1079 		      wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD) &&
1080 		     (opcode != OP(ATOMIC_ACKNOWLEDGE) || diff != 0))) {
1081 			/* Retry this request. */
1082 			if (!(qp->r_flags & RVT_R_RDMAR_SEQ)) {
1083 				qp->r_flags |= RVT_R_RDMAR_SEQ;
1084 				qib_restart_rc(qp, qp->s_last_psn + 1, 0);
1085 				if (list_empty(&qp->rspwait)) {
1086 					qp->r_flags |= RVT_R_RSP_SEND;
1087 					rvt_get_qp(qp);
1088 					list_add_tail(&qp->rspwait,
1089 						      &rcd->qp_wait_list);
1090 				}
1091 			}
1092 			/*
1093 			 * No need to process the ACK/NAK since we are
1094 			 * restarting an earlier request.
1095 			 */
1096 			goto bail;
1097 		}
1098 		if (wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
1099 		    wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD) {
1100 			u64 *vaddr = wqe->sg_list[0].vaddr;
1101 			*vaddr = val;
1102 		}
1103 		if (qp->s_num_rd_atomic &&
1104 		    (wqe->wr.opcode == IB_WR_RDMA_READ ||
1105 		     wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
1106 		     wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD)) {
1107 			qp->s_num_rd_atomic--;
1108 			/* Restart sending task if fence is complete */
1109 			if ((qp->s_flags & RVT_S_WAIT_FENCE) &&
1110 			    !qp->s_num_rd_atomic) {
1111 				qp->s_flags &= ~(RVT_S_WAIT_FENCE |
1112 						 RVT_S_WAIT_ACK);
1113 				qib_schedule_send(qp);
1114 			} else if (qp->s_flags & RVT_S_WAIT_RDMAR) {
1115 				qp->s_flags &= ~(RVT_S_WAIT_RDMAR |
1116 						 RVT_S_WAIT_ACK);
1117 				qib_schedule_send(qp);
1118 			}
1119 		}
1120 		wqe = do_rc_completion(qp, wqe, ibp);
1121 		if (qp->s_acked == qp->s_tail)
1122 			break;
1123 	}
1124 
1125 	switch (aeth >> IB_AETH_NAK_SHIFT) {
1126 	case 0:         /* ACK */
1127 		this_cpu_inc(*ibp->rvp.rc_acks);
1128 		if (qp->s_acked != qp->s_tail) {
1129 			/*
1130 			 * We are expecting more ACKs so
1131 			 * reset the retransmit timer.
1132 			 */
1133 			rvt_mod_retry_timer(qp);
1134 			/*
1135 			 * We can stop resending the earlier packets and
1136 			 * continue with the next packet the receiver wants.
1137 			 */
1138 			if (qib_cmp24(qp->s_psn, psn) <= 0)
1139 				reset_psn(qp, psn + 1);
1140 		} else {
1141 			/* No more acks - kill all timers */
1142 			rvt_stop_rc_timers(qp);
1143 			if (qib_cmp24(qp->s_psn, psn) <= 0) {
1144 				qp->s_state = OP(SEND_LAST);
1145 				qp->s_psn = psn + 1;
1146 			}
1147 		}
1148 		if (qp->s_flags & RVT_S_WAIT_ACK) {
1149 			qp->s_flags &= ~RVT_S_WAIT_ACK;
1150 			qib_schedule_send(qp);
1151 		}
1152 		rvt_get_credit(qp, aeth);
1153 		qp->s_rnr_retry = qp->s_rnr_retry_cnt;
1154 		qp->s_retry = qp->s_retry_cnt;
1155 		update_last_psn(qp, psn);
1156 		return 1;
1157 
1158 	case 1:         /* RNR NAK */
1159 		ibp->rvp.n_rnr_naks++;
1160 		if (qp->s_acked == qp->s_tail)
1161 			goto bail;
1162 		if (qp->s_flags & RVT_S_WAIT_RNR)
1163 			goto bail;
1164 		if (qp->s_rnr_retry == 0) {
1165 			status = IB_WC_RNR_RETRY_EXC_ERR;
1166 			goto class_b;
1167 		}
1168 		if (qp->s_rnr_retry_cnt < 7)
1169 			qp->s_rnr_retry--;
1170 
1171 		/* The last valid PSN is the previous PSN. */
1172 		update_last_psn(qp, psn - 1);
1173 
1174 		ibp->rvp.n_rc_resends += (qp->s_psn - psn) & QIB_PSN_MASK;
1175 
1176 		reset_psn(qp, psn);
1177 
1178 		qp->s_flags &= ~(RVT_S_WAIT_SSN_CREDIT | RVT_S_WAIT_ACK);
1179 		rvt_stop_rc_timers(qp);
1180 		rvt_add_rnr_timer(qp, aeth);
1181 		return 0;
1182 
1183 	case 3:         /* NAK */
1184 		if (qp->s_acked == qp->s_tail)
1185 			goto bail;
1186 		/* The last valid PSN is the previous PSN. */
1187 		update_last_psn(qp, psn - 1);
1188 		switch ((aeth >> IB_AETH_CREDIT_SHIFT) &
1189 			IB_AETH_CREDIT_MASK) {
1190 		case 0: /* PSN sequence error */
1191 			ibp->rvp.n_seq_naks++;
1192 			/*
1193 			 * Back up to the responder's expected PSN.
1194 			 * Note that we might get a NAK in the middle of an
1195 			 * RDMA READ response which terminates the RDMA
1196 			 * READ.
1197 			 */
1198 			qib_restart_rc(qp, psn, 0);
1199 			qib_schedule_send(qp);
1200 			break;
1201 
1202 		case 1: /* Invalid Request */
1203 			status = IB_WC_REM_INV_REQ_ERR;
1204 			ibp->rvp.n_other_naks++;
1205 			goto class_b;
1206 
1207 		case 2: /* Remote Access Error */
1208 			status = IB_WC_REM_ACCESS_ERR;
1209 			ibp->rvp.n_other_naks++;
1210 			goto class_b;
1211 
1212 		case 3: /* Remote Operation Error */
1213 			status = IB_WC_REM_OP_ERR;
1214 			ibp->rvp.n_other_naks++;
1215 class_b:
1216 			if (qp->s_last == qp->s_acked) {
1217 				qib_send_complete(qp, wqe, status);
1218 				rvt_error_qp(qp, IB_WC_WR_FLUSH_ERR);
1219 			}
1220 			break;
1221 
1222 		default:
1223 			/* Ignore other reserved NAK error codes */
1224 			goto reserved;
1225 		}
1226 		qp->s_retry = qp->s_retry_cnt;
1227 		qp->s_rnr_retry = qp->s_rnr_retry_cnt;
1228 		goto bail;
1229 
1230 	default:                /* 2: reserved */
1231 reserved:
1232 		/* Ignore reserved NAK codes. */
1233 		goto bail;
1234 	}
1235 
1236 bail:
1237 	rvt_stop_rc_timers(qp);
1238 	return ret;
1239 }
1240 
1241 /*
1242  * We have seen an out of sequence RDMA read middle or last packet.
1243  * This ACKs SENDs and RDMA writes up to the first RDMA read or atomic SWQE.
1244  */
1245 static void rdma_seq_err(struct rvt_qp *qp, struct qib_ibport *ibp, u32 psn,
1246 			 struct qib_ctxtdata *rcd)
1247 {
1248 	struct rvt_swqe *wqe;
1249 
1250 	/* Remove QP from retry timer */
1251 	rvt_stop_rc_timers(qp);
1252 
1253 	wqe = rvt_get_swqe_ptr(qp, qp->s_acked);
1254 
1255 	while (qib_cmp24(psn, wqe->lpsn) > 0) {
1256 		if (wqe->wr.opcode == IB_WR_RDMA_READ ||
1257 		    wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
1258 		    wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD)
1259 			break;
1260 		wqe = do_rc_completion(qp, wqe, ibp);
1261 	}
1262 
1263 	ibp->rvp.n_rdma_seq++;
1264 	qp->r_flags |= RVT_R_RDMAR_SEQ;
1265 	qib_restart_rc(qp, qp->s_last_psn + 1, 0);
1266 	if (list_empty(&qp->rspwait)) {
1267 		qp->r_flags |= RVT_R_RSP_SEND;
1268 		rvt_get_qp(qp);
1269 		list_add_tail(&qp->rspwait, &rcd->qp_wait_list);
1270 	}
1271 }
1272 
1273 /**
1274  * qib_rc_rcv_resp - process an incoming RC response packet
1275  * @ibp: the port this packet came in on
1276  * @ohdr: the other headers for this packet
1277  * @data: the packet data
1278  * @tlen: the packet length
1279  * @qp: the QP for this packet
1280  * @opcode: the opcode for this packet
1281  * @psn: the packet sequence number for this packet
1282  * @hdrsize: the header length
1283  * @pmtu: the path MTU
1284  *
1285  * This is called from qib_rc_rcv() to process an incoming RC response
1286  * packet for the given QP.
1287  * Called at interrupt level.
1288  */
1289 static void qib_rc_rcv_resp(struct qib_ibport *ibp,
1290 			    struct ib_other_headers *ohdr,
1291 			    void *data, u32 tlen,
1292 			    struct rvt_qp *qp,
1293 			    u32 opcode,
1294 			    u32 psn, u32 hdrsize, u32 pmtu,
1295 			    struct qib_ctxtdata *rcd)
1296 {
1297 	struct rvt_swqe *wqe;
1298 	struct qib_pportdata *ppd = ppd_from_ibp(ibp);
1299 	enum ib_wc_status status;
1300 	unsigned long flags;
1301 	int diff;
1302 	u32 pad;
1303 	u32 aeth;
1304 	u64 val;
1305 
1306 	if (opcode != OP(RDMA_READ_RESPONSE_MIDDLE)) {
1307 		/*
1308 		 * If ACK'd PSN on SDMA busy list try to make progress to
1309 		 * reclaim SDMA credits.
1310 		 */
1311 		if ((qib_cmp24(psn, qp->s_sending_psn) >= 0) &&
1312 		    (qib_cmp24(qp->s_sending_psn, qp->s_sending_hpsn) <= 0)) {
1313 
1314 			/*
1315 			 * If send tasklet not running attempt to progress
1316 			 * SDMA queue.
1317 			 */
1318 			if (!(qp->s_flags & RVT_S_BUSY)) {
1319 				/* Acquire SDMA Lock */
1320 				spin_lock_irqsave(&ppd->sdma_lock, flags);
1321 				/* Invoke sdma make progress */
1322 				qib_sdma_make_progress(ppd);
1323 				/* Release SDMA Lock */
1324 				spin_unlock_irqrestore(&ppd->sdma_lock, flags);
1325 			}
1326 		}
1327 	}
1328 
1329 	spin_lock_irqsave(&qp->s_lock, flags);
1330 	if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK))
1331 		goto ack_done;
1332 
1333 	/* Ignore invalid responses. */
1334 	smp_read_barrier_depends(); /* see post_one_send */
1335 	if (qib_cmp24(psn, READ_ONCE(qp->s_next_psn)) >= 0)
1336 		goto ack_done;
1337 
1338 	/* Ignore duplicate responses. */
1339 	diff = qib_cmp24(psn, qp->s_last_psn);
1340 	if (unlikely(diff <= 0)) {
1341 		/* Update credits for "ghost" ACKs */
1342 		if (diff == 0 && opcode == OP(ACKNOWLEDGE)) {
1343 			aeth = be32_to_cpu(ohdr->u.aeth);
1344 			if ((aeth >> IB_AETH_NAK_SHIFT) == 0)
1345 				rvt_get_credit(qp, aeth);
1346 		}
1347 		goto ack_done;
1348 	}
1349 
1350 	/*
1351 	 * Skip everything other than the PSN we expect, if we are waiting
1352 	 * for a reply to a restarted RDMA read or atomic op.
1353 	 */
1354 	if (qp->r_flags & RVT_R_RDMAR_SEQ) {
1355 		if (qib_cmp24(psn, qp->s_last_psn + 1) != 0)
1356 			goto ack_done;
1357 		qp->r_flags &= ~RVT_R_RDMAR_SEQ;
1358 	}
1359 
1360 	if (unlikely(qp->s_acked == qp->s_tail))
1361 		goto ack_done;
1362 	wqe = rvt_get_swqe_ptr(qp, qp->s_acked);
1363 	status = IB_WC_SUCCESS;
1364 
1365 	switch (opcode) {
1366 	case OP(ACKNOWLEDGE):
1367 	case OP(ATOMIC_ACKNOWLEDGE):
1368 	case OP(RDMA_READ_RESPONSE_FIRST):
1369 		aeth = be32_to_cpu(ohdr->u.aeth);
1370 		if (opcode == OP(ATOMIC_ACKNOWLEDGE))
1371 			val = ib_u64_get(&ohdr->u.at.atomic_ack_eth);
1372 		else
1373 			val = 0;
1374 		if (!do_rc_ack(qp, aeth, psn, opcode, val, rcd) ||
1375 		    opcode != OP(RDMA_READ_RESPONSE_FIRST))
1376 			goto ack_done;
1377 		hdrsize += 4;
1378 		wqe = rvt_get_swqe_ptr(qp, qp->s_acked);
1379 		if (unlikely(wqe->wr.opcode != IB_WR_RDMA_READ))
1380 			goto ack_op_err;
1381 		/*
1382 		 * If this is a response to a resent RDMA read, we
1383 		 * have to be careful to copy the data to the right
1384 		 * location.
1385 		 */
1386 		qp->s_rdma_read_len = restart_sge(&qp->s_rdma_read_sge,
1387 						  wqe, psn, pmtu);
1388 		goto read_middle;
1389 
1390 	case OP(RDMA_READ_RESPONSE_MIDDLE):
1391 		/* no AETH, no ACK */
1392 		if (unlikely(qib_cmp24(psn, qp->s_last_psn + 1)))
1393 			goto ack_seq_err;
1394 		if (unlikely(wqe->wr.opcode != IB_WR_RDMA_READ))
1395 			goto ack_op_err;
1396 read_middle:
1397 		if (unlikely(tlen != (hdrsize + pmtu + 4)))
1398 			goto ack_len_err;
1399 		if (unlikely(pmtu >= qp->s_rdma_read_len))
1400 			goto ack_len_err;
1401 
1402 		/*
1403 		 * We got a response so update the timeout.
1404 		 * 4.096 usec. * (1 << qp->timeout)
1405 		 */
1406 		rvt_mod_retry_timer(qp);
1407 		if (qp->s_flags & RVT_S_WAIT_ACK) {
1408 			qp->s_flags &= ~RVT_S_WAIT_ACK;
1409 			qib_schedule_send(qp);
1410 		}
1411 
1412 		if (opcode == OP(RDMA_READ_RESPONSE_MIDDLE))
1413 			qp->s_retry = qp->s_retry_cnt;
1414 
1415 		/*
1416 		 * Update the RDMA receive state but do the copy w/o
1417 		 * holding the locks and blocking interrupts.
1418 		 */
1419 		qp->s_rdma_read_len -= pmtu;
1420 		update_last_psn(qp, psn);
1421 		spin_unlock_irqrestore(&qp->s_lock, flags);
1422 		qib_copy_sge(&qp->s_rdma_read_sge, data, pmtu, 0);
1423 		goto bail;
1424 
1425 	case OP(RDMA_READ_RESPONSE_ONLY):
1426 		aeth = be32_to_cpu(ohdr->u.aeth);
1427 		if (!do_rc_ack(qp, aeth, psn, opcode, 0, rcd))
1428 			goto ack_done;
1429 		/* Get the number of bytes the message was padded by. */
1430 		pad = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3;
1431 		/*
1432 		 * Check that the data size is >= 0 && <= pmtu.
1433 		 * Remember to account for the AETH header (4) and
1434 		 * ICRC (4).
1435 		 */
1436 		if (unlikely(tlen < (hdrsize + pad + 8)))
1437 			goto ack_len_err;
1438 		/*
1439 		 * If this is a response to a resent RDMA read, we
1440 		 * have to be careful to copy the data to the right
1441 		 * location.
1442 		 */
1443 		wqe = rvt_get_swqe_ptr(qp, qp->s_acked);
1444 		qp->s_rdma_read_len = restart_sge(&qp->s_rdma_read_sge,
1445 						  wqe, psn, pmtu);
1446 		goto read_last;
1447 
1448 	case OP(RDMA_READ_RESPONSE_LAST):
1449 		/* ACKs READ req. */
1450 		if (unlikely(qib_cmp24(psn, qp->s_last_psn + 1)))
1451 			goto ack_seq_err;
1452 		if (unlikely(wqe->wr.opcode != IB_WR_RDMA_READ))
1453 			goto ack_op_err;
1454 		/* Get the number of bytes the message was padded by. */
1455 		pad = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3;
1456 		/*
1457 		 * Check that the data size is >= 1 && <= pmtu.
1458 		 * Remember to account for the AETH header (4) and
1459 		 * ICRC (4).
1460 		 */
1461 		if (unlikely(tlen <= (hdrsize + pad + 8)))
1462 			goto ack_len_err;
1463 read_last:
1464 		tlen -= hdrsize + pad + 8;
1465 		if (unlikely(tlen != qp->s_rdma_read_len))
1466 			goto ack_len_err;
1467 		aeth = be32_to_cpu(ohdr->u.aeth);
1468 		qib_copy_sge(&qp->s_rdma_read_sge, data, tlen, 0);
1469 		WARN_ON(qp->s_rdma_read_sge.num_sge);
1470 		(void) do_rc_ack(qp, aeth, psn,
1471 				 OP(RDMA_READ_RESPONSE_LAST), 0, rcd);
1472 		goto ack_done;
1473 	}
1474 
1475 ack_op_err:
1476 	status = IB_WC_LOC_QP_OP_ERR;
1477 	goto ack_err;
1478 
1479 ack_seq_err:
1480 	rdma_seq_err(qp, ibp, psn, rcd);
1481 	goto ack_done;
1482 
1483 ack_len_err:
1484 	status = IB_WC_LOC_LEN_ERR;
1485 ack_err:
1486 	if (qp->s_last == qp->s_acked) {
1487 		qib_send_complete(qp, wqe, status);
1488 		rvt_error_qp(qp, IB_WC_WR_FLUSH_ERR);
1489 	}
1490 ack_done:
1491 	spin_unlock_irqrestore(&qp->s_lock, flags);
1492 bail:
1493 	return;
1494 }
1495 
1496 /**
1497  * qib_rc_rcv_error - process an incoming duplicate or error RC packet
1498  * @ohdr: the other headers for this packet
1499  * @data: the packet data
1500  * @qp: the QP for this packet
1501  * @opcode: the opcode for this packet
1502  * @psn: the packet sequence number for this packet
1503  * @diff: the difference between the PSN and the expected PSN
1504  *
1505  * This is called from qib_rc_rcv() to process an unexpected
1506  * incoming RC packet for the given QP.
1507  * Called at interrupt level.
1508  * Return 1 if no more processing is needed; otherwise return 0 to
1509  * schedule a response to be sent.
1510  */
1511 static int qib_rc_rcv_error(struct ib_other_headers *ohdr,
1512 			    void *data,
1513 			    struct rvt_qp *qp,
1514 			    u32 opcode,
1515 			    u32 psn,
1516 			    int diff,
1517 			    struct qib_ctxtdata *rcd)
1518 {
1519 	struct qib_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num);
1520 	struct rvt_ack_entry *e;
1521 	unsigned long flags;
1522 	u8 i, prev;
1523 	int old_req;
1524 
1525 	if (diff > 0) {
1526 		/*
1527 		 * Packet sequence error.
1528 		 * A NAK will ACK earlier sends and RDMA writes.
1529 		 * Don't queue the NAK if we already sent one.
1530 		 */
1531 		if (!qp->r_nak_state) {
1532 			ibp->rvp.n_rc_seqnak++;
1533 			qp->r_nak_state = IB_NAK_PSN_ERROR;
1534 			/* Use the expected PSN. */
1535 			qp->r_ack_psn = qp->r_psn;
1536 			/*
1537 			 * Wait to send the sequence NAK until all packets
1538 			 * in the receive queue have been processed.
1539 			 * Otherwise, we end up propagating congestion.
1540 			 */
1541 			if (list_empty(&qp->rspwait)) {
1542 				qp->r_flags |= RVT_R_RSP_NAK;
1543 				rvt_get_qp(qp);
1544 				list_add_tail(&qp->rspwait, &rcd->qp_wait_list);
1545 			}
1546 		}
1547 		goto done;
1548 	}
1549 
1550 	/*
1551 	 * Handle a duplicate request.  Don't re-execute SEND, RDMA
1552 	 * write or atomic op.  Don't NAK errors, just silently drop
1553 	 * the duplicate request.  Note that r_sge, r_len, and
1554 	 * r_rcv_len may be in use so don't modify them.
1555 	 *
1556 	 * We are supposed to ACK the earliest duplicate PSN but we
1557 	 * can coalesce an outstanding duplicate ACK.  We have to
1558 	 * send the earliest so that RDMA reads can be restarted at
1559 	 * the requester's expected PSN.
1560 	 *
1561 	 * First, find where this duplicate PSN falls within the
1562 	 * ACKs previously sent.
1563 	 * old_req is true if there is an older response that is scheduled
1564 	 * to be sent before sending this one.
1565 	 */
1566 	e = NULL;
1567 	old_req = 1;
1568 	ibp->rvp.n_rc_dupreq++;
1569 
1570 	spin_lock_irqsave(&qp->s_lock, flags);
1571 
1572 	for (i = qp->r_head_ack_queue; ; i = prev) {
1573 		if (i == qp->s_tail_ack_queue)
1574 			old_req = 0;
1575 		if (i)
1576 			prev = i - 1;
1577 		else
1578 			prev = QIB_MAX_RDMA_ATOMIC;
1579 		if (prev == qp->r_head_ack_queue) {
1580 			e = NULL;
1581 			break;
1582 		}
1583 		e = &qp->s_ack_queue[prev];
1584 		if (!e->opcode) {
1585 			e = NULL;
1586 			break;
1587 		}
1588 		if (qib_cmp24(psn, e->psn) >= 0) {
1589 			if (prev == qp->s_tail_ack_queue &&
1590 			    qib_cmp24(psn, e->lpsn) <= 0)
1591 				old_req = 0;
1592 			break;
1593 		}
1594 	}
1595 	switch (opcode) {
1596 	case OP(RDMA_READ_REQUEST): {
1597 		struct ib_reth *reth;
1598 		u32 offset;
1599 		u32 len;
1600 
1601 		/*
1602 		 * If we didn't find the RDMA read request in the ack queue,
1603 		 * we can ignore this request.
1604 		 */
1605 		if (!e || e->opcode != OP(RDMA_READ_REQUEST))
1606 			goto unlock_done;
1607 		/* RETH comes after BTH */
1608 		reth = &ohdr->u.rc.reth;
1609 		/*
1610 		 * Address range must be a subset of the original
1611 		 * request and start on pmtu boundaries.
1612 		 * We reuse the old ack_queue slot since the requester
1613 		 * should not back up and request an earlier PSN for the
1614 		 * same request.
1615 		 */
1616 		offset = ((psn - e->psn) & QIB_PSN_MASK) *
1617 			qp->pmtu;
1618 		len = be32_to_cpu(reth->length);
1619 		if (unlikely(offset + len != e->rdma_sge.sge_length))
1620 			goto unlock_done;
1621 		if (e->rdma_sge.mr) {
1622 			rvt_put_mr(e->rdma_sge.mr);
1623 			e->rdma_sge.mr = NULL;
1624 		}
1625 		if (len != 0) {
1626 			u32 rkey = be32_to_cpu(reth->rkey);
1627 			u64 vaddr = be64_to_cpu(reth->vaddr);
1628 			int ok;
1629 
1630 			ok = rvt_rkey_ok(qp, &e->rdma_sge, len, vaddr, rkey,
1631 					 IB_ACCESS_REMOTE_READ);
1632 			if (unlikely(!ok))
1633 				goto unlock_done;
1634 		} else {
1635 			e->rdma_sge.vaddr = NULL;
1636 			e->rdma_sge.length = 0;
1637 			e->rdma_sge.sge_length = 0;
1638 		}
1639 		e->psn = psn;
1640 		if (old_req)
1641 			goto unlock_done;
1642 		qp->s_tail_ack_queue = prev;
1643 		break;
1644 	}
1645 
1646 	case OP(COMPARE_SWAP):
1647 	case OP(FETCH_ADD): {
1648 		/*
1649 		 * If we didn't find the atomic request in the ack queue
1650 		 * or the send tasklet is already backed up to send an
1651 		 * earlier entry, we can ignore this request.
1652 		 */
1653 		if (!e || e->opcode != (u8) opcode || old_req)
1654 			goto unlock_done;
1655 		qp->s_tail_ack_queue = prev;
1656 		break;
1657 	}
1658 
1659 	default:
1660 		/*
1661 		 * Ignore this operation if it doesn't request an ACK
1662 		 * or an earlier RDMA read or atomic is going to be resent.
1663 		 */
1664 		if (!(psn & IB_BTH_REQ_ACK) || old_req)
1665 			goto unlock_done;
1666 		/*
1667 		 * Resend the most recent ACK if this request is
1668 		 * after all the previous RDMA reads and atomics.
1669 		 */
1670 		if (i == qp->r_head_ack_queue) {
1671 			spin_unlock_irqrestore(&qp->s_lock, flags);
1672 			qp->r_nak_state = 0;
1673 			qp->r_ack_psn = qp->r_psn - 1;
1674 			goto send_ack;
1675 		}
1676 		/*
1677 		 * Try to send a simple ACK to work around a Mellanox bug
1678 		 * which doesn't accept a RDMA read response or atomic
1679 		 * response as an ACK for earlier SENDs or RDMA writes.
1680 		 */
1681 		if (!(qp->s_flags & RVT_S_RESP_PENDING)) {
1682 			spin_unlock_irqrestore(&qp->s_lock, flags);
1683 			qp->r_nak_state = 0;
1684 			qp->r_ack_psn = qp->s_ack_queue[i].psn - 1;
1685 			goto send_ack;
1686 		}
1687 		/*
1688 		 * Resend the RDMA read or atomic op which
1689 		 * ACKs this duplicate request.
1690 		 */
1691 		qp->s_tail_ack_queue = i;
1692 		break;
1693 	}
1694 	qp->s_ack_state = OP(ACKNOWLEDGE);
1695 	qp->s_flags |= RVT_S_RESP_PENDING;
1696 	qp->r_nak_state = 0;
1697 	qib_schedule_send(qp);
1698 
1699 unlock_done:
1700 	spin_unlock_irqrestore(&qp->s_lock, flags);
1701 done:
1702 	return 1;
1703 
1704 send_ack:
1705 	return 0;
1706 }
1707 
1708 static inline void qib_update_ack_queue(struct rvt_qp *qp, unsigned n)
1709 {
1710 	unsigned next;
1711 
1712 	next = n + 1;
1713 	if (next > QIB_MAX_RDMA_ATOMIC)
1714 		next = 0;
1715 	qp->s_tail_ack_queue = next;
1716 	qp->s_ack_state = OP(ACKNOWLEDGE);
1717 }
1718 
1719 /**
1720  * qib_rc_rcv - process an incoming RC packet
1721  * @rcd: the context pointer
1722  * @hdr: the header of this packet
1723  * @has_grh: true if the header has a GRH
1724  * @data: the packet data
1725  * @tlen: the packet length
1726  * @qp: the QP for this packet
1727  *
1728  * This is called from qib_qp_rcv() to process an incoming RC packet
1729  * for the given QP.
1730  * Called at interrupt level.
1731  */
1732 void qib_rc_rcv(struct qib_ctxtdata *rcd, struct ib_header *hdr,
1733 		int has_grh, void *data, u32 tlen, struct rvt_qp *qp)
1734 {
1735 	struct qib_ibport *ibp = &rcd->ppd->ibport_data;
1736 	struct ib_other_headers *ohdr;
1737 	u32 opcode;
1738 	u32 hdrsize;
1739 	u32 psn;
1740 	u32 pad;
1741 	struct ib_wc wc;
1742 	u32 pmtu = qp->pmtu;
1743 	int diff;
1744 	struct ib_reth *reth;
1745 	unsigned long flags;
1746 	int ret;
1747 
1748 	/* Check for GRH */
1749 	if (!has_grh) {
1750 		ohdr = &hdr->u.oth;
1751 		hdrsize = 8 + 12;       /* LRH + BTH */
1752 	} else {
1753 		ohdr = &hdr->u.l.oth;
1754 		hdrsize = 8 + 40 + 12;  /* LRH + GRH + BTH */
1755 	}
1756 
1757 	opcode = be32_to_cpu(ohdr->bth[0]);
1758 	if (qib_ruc_check_hdr(ibp, hdr, has_grh, qp, opcode))
1759 		return;
1760 
1761 	psn = be32_to_cpu(ohdr->bth[2]);
1762 	opcode >>= 24;
1763 
1764 	/*
1765 	 * Process responses (ACKs) before anything else.  Note that the
1766 	 * packet sequence number will be for something in the send work
1767 	 * queue rather than the expected receive packet sequence number.
1768 	 * In other words, this QP is the requester.
1769 	 */
1770 	if (opcode >= OP(RDMA_READ_RESPONSE_FIRST) &&
1771 	    opcode <= OP(ATOMIC_ACKNOWLEDGE)) {
1772 		qib_rc_rcv_resp(ibp, ohdr, data, tlen, qp, opcode, psn,
1773 				hdrsize, pmtu, rcd);
1774 		return;
1775 	}
1776 
1777 	/* Compute 24 bits worth of difference. */
1778 	diff = qib_cmp24(psn, qp->r_psn);
1779 	if (unlikely(diff)) {
1780 		if (qib_rc_rcv_error(ohdr, data, qp, opcode, psn, diff, rcd))
1781 			return;
1782 		goto send_ack;
1783 	}
1784 
1785 	/* Check for opcode sequence errors. */
1786 	switch (qp->r_state) {
1787 	case OP(SEND_FIRST):
1788 	case OP(SEND_MIDDLE):
1789 		if (opcode == OP(SEND_MIDDLE) ||
1790 		    opcode == OP(SEND_LAST) ||
1791 		    opcode == OP(SEND_LAST_WITH_IMMEDIATE))
1792 			break;
1793 		goto nack_inv;
1794 
1795 	case OP(RDMA_WRITE_FIRST):
1796 	case OP(RDMA_WRITE_MIDDLE):
1797 		if (opcode == OP(RDMA_WRITE_MIDDLE) ||
1798 		    opcode == OP(RDMA_WRITE_LAST) ||
1799 		    opcode == OP(RDMA_WRITE_LAST_WITH_IMMEDIATE))
1800 			break;
1801 		goto nack_inv;
1802 
1803 	default:
1804 		if (opcode == OP(SEND_MIDDLE) ||
1805 		    opcode == OP(SEND_LAST) ||
1806 		    opcode == OP(SEND_LAST_WITH_IMMEDIATE) ||
1807 		    opcode == OP(RDMA_WRITE_MIDDLE) ||
1808 		    opcode == OP(RDMA_WRITE_LAST) ||
1809 		    opcode == OP(RDMA_WRITE_LAST_WITH_IMMEDIATE))
1810 			goto nack_inv;
1811 		/*
1812 		 * Note that it is up to the requester to not send a new
1813 		 * RDMA read or atomic operation before receiving an ACK
1814 		 * for the previous operation.
1815 		 */
1816 		break;
1817 	}
1818 
1819 	if (qp->state == IB_QPS_RTR && !(qp->r_flags & RVT_R_COMM_EST))
1820 		rvt_comm_est(qp);
1821 
1822 	/* OK, process the packet. */
1823 	switch (opcode) {
1824 	case OP(SEND_FIRST):
1825 		ret = qib_get_rwqe(qp, 0);
1826 		if (ret < 0)
1827 			goto nack_op_err;
1828 		if (!ret)
1829 			goto rnr_nak;
1830 		qp->r_rcv_len = 0;
1831 		/* FALLTHROUGH */
1832 	case OP(SEND_MIDDLE):
1833 	case OP(RDMA_WRITE_MIDDLE):
1834 send_middle:
1835 		/* Check for invalid length PMTU or posted rwqe len. */
1836 		if (unlikely(tlen != (hdrsize + pmtu + 4)))
1837 			goto nack_inv;
1838 		qp->r_rcv_len += pmtu;
1839 		if (unlikely(qp->r_rcv_len > qp->r_len))
1840 			goto nack_inv;
1841 		qib_copy_sge(&qp->r_sge, data, pmtu, 1);
1842 		break;
1843 
1844 	case OP(RDMA_WRITE_LAST_WITH_IMMEDIATE):
1845 		/* consume RWQE */
1846 		ret = qib_get_rwqe(qp, 1);
1847 		if (ret < 0)
1848 			goto nack_op_err;
1849 		if (!ret)
1850 			goto rnr_nak;
1851 		goto send_last_imm;
1852 
1853 	case OP(SEND_ONLY):
1854 	case OP(SEND_ONLY_WITH_IMMEDIATE):
1855 		ret = qib_get_rwqe(qp, 0);
1856 		if (ret < 0)
1857 			goto nack_op_err;
1858 		if (!ret)
1859 			goto rnr_nak;
1860 		qp->r_rcv_len = 0;
1861 		if (opcode == OP(SEND_ONLY))
1862 			goto no_immediate_data;
1863 		/* FALLTHROUGH for SEND_ONLY_WITH_IMMEDIATE */
1864 	case OP(SEND_LAST_WITH_IMMEDIATE):
1865 send_last_imm:
1866 		wc.ex.imm_data = ohdr->u.imm_data;
1867 		hdrsize += 4;
1868 		wc.wc_flags = IB_WC_WITH_IMM;
1869 		goto send_last;
1870 	case OP(SEND_LAST):
1871 	case OP(RDMA_WRITE_LAST):
1872 no_immediate_data:
1873 		wc.wc_flags = 0;
1874 		wc.ex.imm_data = 0;
1875 send_last:
1876 		/* Get the number of bytes the message was padded by. */
1877 		pad = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3;
1878 		/* Check for invalid length. */
1879 		/* XXX LAST len should be >= 1 */
1880 		if (unlikely(tlen < (hdrsize + pad + 4)))
1881 			goto nack_inv;
1882 		/* Don't count the CRC. */
1883 		tlen -= (hdrsize + pad + 4);
1884 		wc.byte_len = tlen + qp->r_rcv_len;
1885 		if (unlikely(wc.byte_len > qp->r_len))
1886 			goto nack_inv;
1887 		qib_copy_sge(&qp->r_sge, data, tlen, 1);
1888 		rvt_put_ss(&qp->r_sge);
1889 		qp->r_msn++;
1890 		if (!test_and_clear_bit(RVT_R_WRID_VALID, &qp->r_aflags))
1891 			break;
1892 		wc.wr_id = qp->r_wr_id;
1893 		wc.status = IB_WC_SUCCESS;
1894 		if (opcode == OP(RDMA_WRITE_LAST_WITH_IMMEDIATE) ||
1895 		    opcode == OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE))
1896 			wc.opcode = IB_WC_RECV_RDMA_WITH_IMM;
1897 		else
1898 			wc.opcode = IB_WC_RECV;
1899 		wc.qp = &qp->ibqp;
1900 		wc.src_qp = qp->remote_qpn;
1901 		wc.slid = qp->remote_ah_attr.dlid;
1902 		wc.sl = qp->remote_ah_attr.sl;
1903 		/* zero fields that are N/A */
1904 		wc.vendor_err = 0;
1905 		wc.pkey_index = 0;
1906 		wc.dlid_path_bits = 0;
1907 		wc.port_num = 0;
1908 		/* Signal completion event if the solicited bit is set. */
1909 		rvt_cq_enter(ibcq_to_rvtcq(qp->ibqp.recv_cq), &wc,
1910 			     (ohdr->bth[0] &
1911 			      cpu_to_be32(IB_BTH_SOLICITED)) != 0);
1912 		break;
1913 
1914 	case OP(RDMA_WRITE_FIRST):
1915 	case OP(RDMA_WRITE_ONLY):
1916 	case OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE):
1917 		if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_WRITE)))
1918 			goto nack_inv;
1919 		/* consume RWQE */
1920 		reth = &ohdr->u.rc.reth;
1921 		hdrsize += sizeof(*reth);
1922 		qp->r_len = be32_to_cpu(reth->length);
1923 		qp->r_rcv_len = 0;
1924 		qp->r_sge.sg_list = NULL;
1925 		if (qp->r_len != 0) {
1926 			u32 rkey = be32_to_cpu(reth->rkey);
1927 			u64 vaddr = be64_to_cpu(reth->vaddr);
1928 			int ok;
1929 
1930 			/* Check rkey & NAK */
1931 			ok = rvt_rkey_ok(qp, &qp->r_sge.sge, qp->r_len, vaddr,
1932 					 rkey, IB_ACCESS_REMOTE_WRITE);
1933 			if (unlikely(!ok))
1934 				goto nack_acc;
1935 			qp->r_sge.num_sge = 1;
1936 		} else {
1937 			qp->r_sge.num_sge = 0;
1938 			qp->r_sge.sge.mr = NULL;
1939 			qp->r_sge.sge.vaddr = NULL;
1940 			qp->r_sge.sge.length = 0;
1941 			qp->r_sge.sge.sge_length = 0;
1942 		}
1943 		if (opcode == OP(RDMA_WRITE_FIRST))
1944 			goto send_middle;
1945 		else if (opcode == OP(RDMA_WRITE_ONLY))
1946 			goto no_immediate_data;
1947 		ret = qib_get_rwqe(qp, 1);
1948 		if (ret < 0)
1949 			goto nack_op_err;
1950 		if (!ret)
1951 			goto rnr_nak;
1952 		wc.ex.imm_data = ohdr->u.rc.imm_data;
1953 		hdrsize += 4;
1954 		wc.wc_flags = IB_WC_WITH_IMM;
1955 		goto send_last;
1956 
1957 	case OP(RDMA_READ_REQUEST): {
1958 		struct rvt_ack_entry *e;
1959 		u32 len;
1960 		u8 next;
1961 
1962 		if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_READ)))
1963 			goto nack_inv;
1964 		next = qp->r_head_ack_queue + 1;
1965 		/* s_ack_queue is size QIB_MAX_RDMA_ATOMIC+1 so use > not >= */
1966 		if (next > QIB_MAX_RDMA_ATOMIC)
1967 			next = 0;
1968 		spin_lock_irqsave(&qp->s_lock, flags);
1969 		if (unlikely(next == qp->s_tail_ack_queue)) {
1970 			if (!qp->s_ack_queue[next].sent)
1971 				goto nack_inv_unlck;
1972 			qib_update_ack_queue(qp, next);
1973 		}
1974 		e = &qp->s_ack_queue[qp->r_head_ack_queue];
1975 		if (e->opcode == OP(RDMA_READ_REQUEST) && e->rdma_sge.mr) {
1976 			rvt_put_mr(e->rdma_sge.mr);
1977 			e->rdma_sge.mr = NULL;
1978 		}
1979 		reth = &ohdr->u.rc.reth;
1980 		len = be32_to_cpu(reth->length);
1981 		if (len) {
1982 			u32 rkey = be32_to_cpu(reth->rkey);
1983 			u64 vaddr = be64_to_cpu(reth->vaddr);
1984 			int ok;
1985 
1986 			/* Check rkey & NAK */
1987 			ok = rvt_rkey_ok(qp, &e->rdma_sge, len, vaddr,
1988 					 rkey, IB_ACCESS_REMOTE_READ);
1989 			if (unlikely(!ok))
1990 				goto nack_acc_unlck;
1991 			/*
1992 			 * Update the next expected PSN.  We add 1 later
1993 			 * below, so only add the remainder here.
1994 			 */
1995 			qp->r_psn += rvt_div_mtu(qp, len - 1);
1996 		} else {
1997 			e->rdma_sge.mr = NULL;
1998 			e->rdma_sge.vaddr = NULL;
1999 			e->rdma_sge.length = 0;
2000 			e->rdma_sge.sge_length = 0;
2001 		}
2002 		e->opcode = opcode;
2003 		e->sent = 0;
2004 		e->psn = psn;
2005 		e->lpsn = qp->r_psn;
2006 		/*
2007 		 * We need to increment the MSN here instead of when we
2008 		 * finish sending the result since a duplicate request would
2009 		 * increment it more than once.
2010 		 */
2011 		qp->r_msn++;
2012 		qp->r_psn++;
2013 		qp->r_state = opcode;
2014 		qp->r_nak_state = 0;
2015 		qp->r_head_ack_queue = next;
2016 
2017 		/* Schedule the send tasklet. */
2018 		qp->s_flags |= RVT_S_RESP_PENDING;
2019 		qib_schedule_send(qp);
2020 
2021 		goto sunlock;
2022 	}
2023 
2024 	case OP(COMPARE_SWAP):
2025 	case OP(FETCH_ADD): {
2026 		struct ib_atomic_eth *ateth;
2027 		struct rvt_ack_entry *e;
2028 		u64 vaddr;
2029 		atomic64_t *maddr;
2030 		u64 sdata;
2031 		u32 rkey;
2032 		u8 next;
2033 
2034 		if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_ATOMIC)))
2035 			goto nack_inv;
2036 		next = qp->r_head_ack_queue + 1;
2037 		if (next > QIB_MAX_RDMA_ATOMIC)
2038 			next = 0;
2039 		spin_lock_irqsave(&qp->s_lock, flags);
2040 		if (unlikely(next == qp->s_tail_ack_queue)) {
2041 			if (!qp->s_ack_queue[next].sent)
2042 				goto nack_inv_unlck;
2043 			qib_update_ack_queue(qp, next);
2044 		}
2045 		e = &qp->s_ack_queue[qp->r_head_ack_queue];
2046 		if (e->opcode == OP(RDMA_READ_REQUEST) && e->rdma_sge.mr) {
2047 			rvt_put_mr(e->rdma_sge.mr);
2048 			e->rdma_sge.mr = NULL;
2049 		}
2050 		ateth = &ohdr->u.atomic_eth;
2051 		vaddr = get_ib_ateth_vaddr(ateth);
2052 		if (unlikely(vaddr & (sizeof(u64) - 1)))
2053 			goto nack_inv_unlck;
2054 		rkey = be32_to_cpu(ateth->rkey);
2055 		/* Check rkey & NAK */
2056 		if (unlikely(!rvt_rkey_ok(qp, &qp->r_sge.sge, sizeof(u64),
2057 					  vaddr, rkey,
2058 					  IB_ACCESS_REMOTE_ATOMIC)))
2059 			goto nack_acc_unlck;
2060 		/* Perform atomic OP and save result. */
2061 		maddr = (atomic64_t *) qp->r_sge.sge.vaddr;
2062 		sdata = get_ib_ateth_swap(ateth);
2063 		e->atomic_data = (opcode == OP(FETCH_ADD)) ?
2064 			(u64) atomic64_add_return(sdata, maddr) - sdata :
2065 			(u64) cmpxchg((u64 *) qp->r_sge.sge.vaddr,
2066 				      get_ib_ateth_compare(ateth),
2067 				      sdata);
2068 		rvt_put_mr(qp->r_sge.sge.mr);
2069 		qp->r_sge.num_sge = 0;
2070 		e->opcode = opcode;
2071 		e->sent = 0;
2072 		e->psn = psn;
2073 		e->lpsn = psn;
2074 		qp->r_msn++;
2075 		qp->r_psn++;
2076 		qp->r_state = opcode;
2077 		qp->r_nak_state = 0;
2078 		qp->r_head_ack_queue = next;
2079 
2080 		/* Schedule the send tasklet. */
2081 		qp->s_flags |= RVT_S_RESP_PENDING;
2082 		qib_schedule_send(qp);
2083 
2084 		goto sunlock;
2085 	}
2086 
2087 	default:
2088 		/* NAK unknown opcodes. */
2089 		goto nack_inv;
2090 	}
2091 	qp->r_psn++;
2092 	qp->r_state = opcode;
2093 	qp->r_ack_psn = psn;
2094 	qp->r_nak_state = 0;
2095 	/* Send an ACK if requested or required. */
2096 	if (psn & (1 << 31))
2097 		goto send_ack;
2098 	return;
2099 
2100 rnr_nak:
2101 	qp->r_nak_state = IB_RNR_NAK | qp->r_min_rnr_timer;
2102 	qp->r_ack_psn = qp->r_psn;
2103 	/* Queue RNR NAK for later */
2104 	if (list_empty(&qp->rspwait)) {
2105 		qp->r_flags |= RVT_R_RSP_NAK;
2106 		rvt_get_qp(qp);
2107 		list_add_tail(&qp->rspwait, &rcd->qp_wait_list);
2108 	}
2109 	return;
2110 
2111 nack_op_err:
2112 	rvt_rc_error(qp, IB_WC_LOC_QP_OP_ERR);
2113 	qp->r_nak_state = IB_NAK_REMOTE_OPERATIONAL_ERROR;
2114 	qp->r_ack_psn = qp->r_psn;
2115 	/* Queue NAK for later */
2116 	if (list_empty(&qp->rspwait)) {
2117 		qp->r_flags |= RVT_R_RSP_NAK;
2118 		rvt_get_qp(qp);
2119 		list_add_tail(&qp->rspwait, &rcd->qp_wait_list);
2120 	}
2121 	return;
2122 
2123 nack_inv_unlck:
2124 	spin_unlock_irqrestore(&qp->s_lock, flags);
2125 nack_inv:
2126 	rvt_rc_error(qp, IB_WC_LOC_QP_OP_ERR);
2127 	qp->r_nak_state = IB_NAK_INVALID_REQUEST;
2128 	qp->r_ack_psn = qp->r_psn;
2129 	/* Queue NAK for later */
2130 	if (list_empty(&qp->rspwait)) {
2131 		qp->r_flags |= RVT_R_RSP_NAK;
2132 		rvt_get_qp(qp);
2133 		list_add_tail(&qp->rspwait, &rcd->qp_wait_list);
2134 	}
2135 	return;
2136 
2137 nack_acc_unlck:
2138 	spin_unlock_irqrestore(&qp->s_lock, flags);
2139 nack_acc:
2140 	rvt_rc_error(qp, IB_WC_LOC_PROT_ERR);
2141 	qp->r_nak_state = IB_NAK_REMOTE_ACCESS_ERROR;
2142 	qp->r_ack_psn = qp->r_psn;
2143 send_ack:
2144 	qib_send_rc_ack(qp);
2145 	return;
2146 
2147 sunlock:
2148 	spin_unlock_irqrestore(&qp->s_lock, flags);
2149 }
2150