[go: nahoru, domu]

1/*
2 * Copyright (c) 2006, 2007, 2008 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 "ipath_verbs.h"
37#include "ipath_kernel.h"
38
39/* cut down ridiculously long IB macro names */
40#define OP(x) IB_OPCODE_RC_##x
41
42static u32 restart_sge(struct ipath_sge_state *ss, struct ipath_swqe *wqe,
43		       u32 psn, u32 pmtu)
44{
45	u32 len;
46
47	len = ((psn - wqe->psn) & IPATH_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	ipath_skip_sge(ss, len);
52	return wqe->length - len;
53}
54
55/**
56 * ipath_init_restart- initialize the qp->s_sge after a restart
57 * @qp: the QP who's SGE we're restarting
58 * @wqe: the work queue to initialize the QP's SGE from
59 *
60 * The QP s_lock should be held and interrupts disabled.
61 */
62static void ipath_init_restart(struct ipath_qp *qp, struct ipath_swqe *wqe)
63{
64	struct ipath_ibdev *dev;
65
66	qp->s_len = restart_sge(&qp->s_sge, wqe, qp->s_psn,
67				ib_mtu_enum_to_int(qp->path_mtu));
68	dev = to_idev(qp->ibqp.device);
69	spin_lock(&dev->pending_lock);
70	if (list_empty(&qp->timerwait))
71		list_add_tail(&qp->timerwait,
72			      &dev->pending[dev->pending_index]);
73	spin_unlock(&dev->pending_lock);
74}
75
76/**
77 * ipath_make_rc_ack - construct a response packet (ACK, NAK, or RDMA read)
78 * @qp: a pointer to the QP
79 * @ohdr: a pointer to the IB header being constructed
80 * @pmtu: the path MTU
81 *
82 * Return 1 if constructed; otherwise, return 0.
83 * Note that we are in the responder's side of the QP context.
84 * Note the QP s_lock must be held.
85 */
86static int ipath_make_rc_ack(struct ipath_ibdev *dev, struct ipath_qp *qp,
87			     struct ipath_other_headers *ohdr, u32 pmtu)
88{
89	struct ipath_ack_entry *e;
90	u32 hwords;
91	u32 len;
92	u32 bth0;
93	u32 bth2;
94
95	/* Don't send an ACK if we aren't supposed to. */
96	if (!(ib_ipath_state_ops[qp->state] & IPATH_PROCESS_RECV_OK))
97		goto bail;
98
99	/* header size in 32-bit words LRH+BTH = (8+12)/4. */
100	hwords = 5;
101
102	switch (qp->s_ack_state) {
103	case OP(RDMA_READ_RESPONSE_LAST):
104	case OP(RDMA_READ_RESPONSE_ONLY):
105	case OP(ATOMIC_ACKNOWLEDGE):
106		/*
107		 * We can increment the tail pointer now that the last
108		 * response has been sent instead of only being
109		 * constructed.
110		 */
111		if (++qp->s_tail_ack_queue > IPATH_MAX_RDMA_ATOMIC)
112			qp->s_tail_ack_queue = 0;
113		/* FALLTHROUGH */
114	case OP(SEND_ONLY):
115	case OP(ACKNOWLEDGE):
116		/* Check for no next entry in the queue. */
117		if (qp->r_head_ack_queue == qp->s_tail_ack_queue) {
118			if (qp->s_flags & IPATH_S_ACK_PENDING)
119				goto normal;
120			qp->s_ack_state = OP(ACKNOWLEDGE);
121			goto bail;
122		}
123
124		e = &qp->s_ack_queue[qp->s_tail_ack_queue];
125		if (e->opcode == OP(RDMA_READ_REQUEST)) {
126			/* Copy SGE state in case we need to resend */
127			qp->s_ack_rdma_sge = e->rdma_sge;
128			qp->s_cur_sge = &qp->s_ack_rdma_sge;
129			len = e->rdma_sge.sge.sge_length;
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 = ipath_compute_aeth(qp);
138			hwords++;
139			qp->s_ack_rdma_psn = e->psn;
140			bth2 = qp->s_ack_rdma_psn++ & IPATH_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 = ipath_compute_aeth(qp);
147			ohdr->u.at.atomic_ack_eth[0] =
148				cpu_to_be32(e->atomic_data >> 32);
149			ohdr->u.at.atomic_ack_eth[1] =
150				cpu_to_be32(e->atomic_data);
151			hwords += sizeof(ohdr->u.at) / sizeof(u32);
152			bth2 = e->psn;
153			e->sent = 1;
154		}
155		bth0 = qp->s_ack_state << 24;
156		break;
157
158	case OP(RDMA_READ_RESPONSE_FIRST):
159		qp->s_ack_state = OP(RDMA_READ_RESPONSE_MIDDLE);
160		/* FALLTHROUGH */
161	case OP(RDMA_READ_RESPONSE_MIDDLE):
162		len = qp->s_ack_rdma_sge.sge.sge_length;
163		if (len > pmtu)
164			len = pmtu;
165		else {
166			ohdr->u.aeth = ipath_compute_aeth(qp);
167			hwords++;
168			qp->s_ack_state = OP(RDMA_READ_RESPONSE_LAST);
169			qp->s_ack_queue[qp->s_tail_ack_queue].sent = 1;
170		}
171		bth0 = qp->s_ack_state << 24;
172		bth2 = qp->s_ack_rdma_psn++ & IPATH_PSN_MASK;
173		break;
174
175	default:
176	normal:
177		/*
178		 * Send a regular ACK.
179		 * Set the s_ack_state so we wait until after sending
180		 * the ACK before setting s_ack_state to ACKNOWLEDGE
181		 * (see above).
182		 */
183		qp->s_ack_state = OP(SEND_ONLY);
184		qp->s_flags &= ~IPATH_S_ACK_PENDING;
185		qp->s_cur_sge = NULL;
186		if (qp->s_nak_state)
187			ohdr->u.aeth =
188				cpu_to_be32((qp->r_msn & IPATH_MSN_MASK) |
189					    (qp->s_nak_state <<
190					     IPATH_AETH_CREDIT_SHIFT));
191		else
192			ohdr->u.aeth = ipath_compute_aeth(qp);
193		hwords++;
194		len = 0;
195		bth0 = OP(ACKNOWLEDGE) << 24;
196		bth2 = qp->s_ack_psn & IPATH_PSN_MASK;
197	}
198	qp->s_hdrwords = hwords;
199	qp->s_cur_size = len;
200	ipath_make_ruc_header(dev, qp, ohdr, bth0, bth2);
201	return 1;
202
203bail:
204	return 0;
205}
206
207/**
208 * ipath_make_rc_req - construct a request packet (SEND, RDMA r/w, ATOMIC)
209 * @qp: a pointer to the QP
210 *
211 * Return 1 if constructed; otherwise, return 0.
212 */
213int ipath_make_rc_req(struct ipath_qp *qp)
214{
215	struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
216	struct ipath_other_headers *ohdr;
217	struct ipath_sge_state *ss;
218	struct ipath_swqe *wqe;
219	u32 hwords;
220	u32 len;
221	u32 bth0;
222	u32 bth2;
223	u32 pmtu = ib_mtu_enum_to_int(qp->path_mtu);
224	char newreq;
225	unsigned long flags;
226	int ret = 0;
227
228	ohdr = &qp->s_hdr.u.oth;
229	if (qp->remote_ah_attr.ah_flags & IB_AH_GRH)
230		ohdr = &qp->s_hdr.u.l.oth;
231
232	/*
233	 * The lock is needed to synchronize between the sending tasklet,
234	 * the receive interrupt handler, and timeout resends.
235	 */
236	spin_lock_irqsave(&qp->s_lock, flags);
237
238	/* Sending responses has higher priority over sending requests. */
239	if ((qp->r_head_ack_queue != qp->s_tail_ack_queue ||
240	     (qp->s_flags & IPATH_S_ACK_PENDING) ||
241	     qp->s_ack_state != OP(ACKNOWLEDGE)) &&
242	    ipath_make_rc_ack(dev, qp, ohdr, pmtu))
243		goto done;
244
245	if (!(ib_ipath_state_ops[qp->state] & IPATH_PROCESS_SEND_OK)) {
246		if (!(ib_ipath_state_ops[qp->state] & IPATH_FLUSH_SEND))
247			goto bail;
248		/* We are in the error state, flush the work request. */
249		if (qp->s_last == qp->s_head)
250			goto bail;
251		/* If DMAs are in progress, we can't flush immediately. */
252		if (atomic_read(&qp->s_dma_busy)) {
253			qp->s_flags |= IPATH_S_WAIT_DMA;
254			goto bail;
255		}
256		wqe = get_swqe_ptr(qp, qp->s_last);
257		ipath_send_complete(qp, wqe, IB_WC_WR_FLUSH_ERR);
258		goto done;
259	}
260
261	/* Leave BUSY set until RNR timeout. */
262	if (qp->s_rnr_timeout) {
263		qp->s_flags |= IPATH_S_WAITING;
264		goto bail;
265	}
266
267	/* header size in 32-bit words LRH+BTH = (8+12)/4. */
268	hwords = 5;
269	bth0 = 1 << 22; /* Set M bit */
270
271	/* Send a request. */
272	wqe = get_swqe_ptr(qp, qp->s_cur);
273	switch (qp->s_state) {
274	default:
275		if (!(ib_ipath_state_ops[qp->state] &
276		    IPATH_PROCESS_NEXT_SEND_OK))
277			goto bail;
278		/*
279		 * Resend an old request or start a new one.
280		 *
281		 * We keep track of the current SWQE so that
282		 * we don't reset the "furthest progress" state
283		 * if we need to back up.
284		 */
285		newreq = 0;
286		if (qp->s_cur == qp->s_tail) {
287			/* Check if send work queue is empty. */
288			if (qp->s_tail == qp->s_head)
289				goto bail;
290			/*
291			 * If a fence is requested, wait for previous
292			 * RDMA read and atomic operations to finish.
293			 */
294			if ((wqe->wr.send_flags & IB_SEND_FENCE) &&
295			    qp->s_num_rd_atomic) {
296				qp->s_flags |= IPATH_S_FENCE_PENDING;
297				goto bail;
298			}
299			wqe->psn = qp->s_next_psn;
300			newreq = 1;
301		}
302		/*
303		 * Note that we have to be careful not to modify the
304		 * original work request since we may need to resend
305		 * it.
306		 */
307		len = wqe->length;
308		ss = &qp->s_sge;
309		bth2 = 0;
310		switch (wqe->wr.opcode) {
311		case IB_WR_SEND:
312		case IB_WR_SEND_WITH_IMM:
313			/* If no credit, return. */
314			if (qp->s_lsn != (u32) -1 &&
315			    ipath_cmp24(wqe->ssn, qp->s_lsn + 1) > 0) {
316				qp->s_flags |= IPATH_S_WAIT_SSN_CREDIT;
317				goto bail;
318			}
319			wqe->lpsn = wqe->psn;
320			if (len > pmtu) {
321				wqe->lpsn += (len - 1) / pmtu;
322				qp->s_state = OP(SEND_FIRST);
323				len = pmtu;
324				break;
325			}
326			if (wqe->wr.opcode == IB_WR_SEND)
327				qp->s_state = OP(SEND_ONLY);
328			else {
329				qp->s_state = OP(SEND_ONLY_WITH_IMMEDIATE);
330				/* Immediate data comes after the BTH */
331				ohdr->u.imm_data = wqe->wr.ex.imm_data;
332				hwords += 1;
333			}
334			if (wqe->wr.send_flags & IB_SEND_SOLICITED)
335				bth0 |= 1 << 23;
336			bth2 = 1 << 31;	/* Request ACK. */
337			if (++qp->s_cur == qp->s_size)
338				qp->s_cur = 0;
339			break;
340
341		case IB_WR_RDMA_WRITE:
342			if (newreq && qp->s_lsn != (u32) -1)
343				qp->s_lsn++;
344			/* FALLTHROUGH */
345		case IB_WR_RDMA_WRITE_WITH_IMM:
346			/* If no credit, return. */
347			if (qp->s_lsn != (u32) -1 &&
348			    ipath_cmp24(wqe->ssn, qp->s_lsn + 1) > 0) {
349				qp->s_flags |= IPATH_S_WAIT_SSN_CREDIT;
350				goto bail;
351			}
352			ohdr->u.rc.reth.vaddr =
353				cpu_to_be64(wqe->wr.wr.rdma.remote_addr);
354			ohdr->u.rc.reth.rkey =
355				cpu_to_be32(wqe->wr.wr.rdma.rkey);
356			ohdr->u.rc.reth.length = cpu_to_be32(len);
357			hwords += sizeof(struct ib_reth) / sizeof(u32);
358			wqe->lpsn = wqe->psn;
359			if (len > pmtu) {
360				wqe->lpsn += (len - 1) / pmtu;
361				qp->s_state = OP(RDMA_WRITE_FIRST);
362				len = pmtu;
363				break;
364			}
365			if (wqe->wr.opcode == IB_WR_RDMA_WRITE)
366				qp->s_state = OP(RDMA_WRITE_ONLY);
367			else {
368				qp->s_state =
369					OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE);
370				/* Immediate data comes after RETH */
371				ohdr->u.rc.imm_data = wqe->wr.ex.imm_data;
372				hwords += 1;
373				if (wqe->wr.send_flags & IB_SEND_SOLICITED)
374					bth0 |= 1 << 23;
375			}
376			bth2 = 1 << 31;	/* Request ACK. */
377			if (++qp->s_cur == qp->s_size)
378				qp->s_cur = 0;
379			break;
380
381		case IB_WR_RDMA_READ:
382			/*
383			 * Don't allow more operations to be started
384			 * than the QP limits allow.
385			 */
386			if (newreq) {
387				if (qp->s_num_rd_atomic >=
388				    qp->s_max_rd_atomic) {
389					qp->s_flags |= IPATH_S_RDMAR_PENDING;
390					goto bail;
391				}
392				qp->s_num_rd_atomic++;
393				if (qp->s_lsn != (u32) -1)
394					qp->s_lsn++;
395				/*
396				 * Adjust s_next_psn to count the
397				 * expected number of responses.
398				 */
399				if (len > pmtu)
400					qp->s_next_psn += (len - 1) / pmtu;
401				wqe->lpsn = qp->s_next_psn++;
402			}
403			ohdr->u.rc.reth.vaddr =
404				cpu_to_be64(wqe->wr.wr.rdma.remote_addr);
405			ohdr->u.rc.reth.rkey =
406				cpu_to_be32(wqe->wr.wr.rdma.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			if (++qp->s_cur == qp->s_size)
413				qp->s_cur = 0;
414			break;
415
416		case IB_WR_ATOMIC_CMP_AND_SWP:
417		case IB_WR_ATOMIC_FETCH_AND_ADD:
418			/*
419			 * Don't allow more operations to be started
420			 * than the QP limits allow.
421			 */
422			if (newreq) {
423				if (qp->s_num_rd_atomic >=
424				    qp->s_max_rd_atomic) {
425					qp->s_flags |= IPATH_S_RDMAR_PENDING;
426					goto bail;
427				}
428				qp->s_num_rd_atomic++;
429				if (qp->s_lsn != (u32) -1)
430					qp->s_lsn++;
431				wqe->lpsn = wqe->psn;
432			}
433			if (wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
434				qp->s_state = OP(COMPARE_SWAP);
435				ohdr->u.atomic_eth.swap_data = cpu_to_be64(
436					wqe->wr.wr.atomic.swap);
437				ohdr->u.atomic_eth.compare_data = cpu_to_be64(
438					wqe->wr.wr.atomic.compare_add);
439			} else {
440				qp->s_state = OP(FETCH_ADD);
441				ohdr->u.atomic_eth.swap_data = cpu_to_be64(
442					wqe->wr.wr.atomic.compare_add);
443				ohdr->u.atomic_eth.compare_data = 0;
444			}
445			ohdr->u.atomic_eth.vaddr[0] = cpu_to_be32(
446				wqe->wr.wr.atomic.remote_addr >> 32);
447			ohdr->u.atomic_eth.vaddr[1] = cpu_to_be32(
448				wqe->wr.wr.atomic.remote_addr);
449			ohdr->u.atomic_eth.rkey = cpu_to_be32(
450				wqe->wr.wr.atomic.rkey);
451			hwords += sizeof(struct ib_atomic_eth) / sizeof(u32);
452			ss = NULL;
453			len = 0;
454			if (++qp->s_cur == qp->s_size)
455				qp->s_cur = 0;
456			break;
457
458		default:
459			goto bail;
460		}
461		qp->s_sge.sge = wqe->sg_list[0];
462		qp->s_sge.sg_list = wqe->sg_list + 1;
463		qp->s_sge.num_sge = wqe->wr.num_sge;
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		bth2 |= qp->s_psn & IPATH_PSN_MASK;
471		if (wqe->wr.opcode == IB_WR_RDMA_READ)
472			qp->s_psn = wqe->lpsn + 1;
473		else {
474			qp->s_psn++;
475			if (ipath_cmp24(qp->s_psn, qp->s_next_psn) > 0)
476				qp->s_next_psn = qp->s_psn;
477		}
478		/*
479		 * Put the QP on the pending list so lost ACKs will cause
480		 * a retry.  More than one request can be pending so the
481		 * QP may already be on the dev->pending list.
482		 */
483		spin_lock(&dev->pending_lock);
484		if (list_empty(&qp->timerwait))
485			list_add_tail(&qp->timerwait,
486				      &dev->pending[dev->pending_index]);
487		spin_unlock(&dev->pending_lock);
488		break;
489
490	case OP(RDMA_READ_RESPONSE_FIRST):
491		/*
492		 * This case can only happen if a send is restarted.
493		 * See ipath_restart_rc().
494		 */
495		ipath_init_restart(qp, wqe);
496		/* FALLTHROUGH */
497	case OP(SEND_FIRST):
498		qp->s_state = OP(SEND_MIDDLE);
499		/* FALLTHROUGH */
500	case OP(SEND_MIDDLE):
501		bth2 = qp->s_psn++ & IPATH_PSN_MASK;
502		if (ipath_cmp24(qp->s_psn, qp->s_next_psn) > 0)
503			qp->s_next_psn = qp->s_psn;
504		ss = &qp->s_sge;
505		len = qp->s_len;
506		if (len > pmtu) {
507			len = pmtu;
508			break;
509		}
510		if (wqe->wr.opcode == IB_WR_SEND)
511			qp->s_state = OP(SEND_LAST);
512		else {
513			qp->s_state = OP(SEND_LAST_WITH_IMMEDIATE);
514			/* Immediate data comes after the BTH */
515			ohdr->u.imm_data = wqe->wr.ex.imm_data;
516			hwords += 1;
517		}
518		if (wqe->wr.send_flags & IB_SEND_SOLICITED)
519			bth0 |= 1 << 23;
520		bth2 |= 1 << 31;	/* Request ACK. */
521		qp->s_cur++;
522		if (qp->s_cur >= qp->s_size)
523			qp->s_cur = 0;
524		break;
525
526	case OP(RDMA_READ_RESPONSE_LAST):
527		/*
528		 * This case can only happen if a RDMA write is restarted.
529		 * See ipath_restart_rc().
530		 */
531		ipath_init_restart(qp, wqe);
532		/* FALLTHROUGH */
533	case OP(RDMA_WRITE_FIRST):
534		qp->s_state = OP(RDMA_WRITE_MIDDLE);
535		/* FALLTHROUGH */
536	case OP(RDMA_WRITE_MIDDLE):
537		bth2 = qp->s_psn++ & IPATH_PSN_MASK;
538		if (ipath_cmp24(qp->s_psn, qp->s_next_psn) > 0)
539			qp->s_next_psn = qp->s_psn;
540		ss = &qp->s_sge;
541		len = qp->s_len;
542		if (len > pmtu) {
543			len = pmtu;
544			break;
545		}
546		if (wqe->wr.opcode == IB_WR_RDMA_WRITE)
547			qp->s_state = OP(RDMA_WRITE_LAST);
548		else {
549			qp->s_state = OP(RDMA_WRITE_LAST_WITH_IMMEDIATE);
550			/* Immediate data comes after the BTH */
551			ohdr->u.imm_data = wqe->wr.ex.imm_data;
552			hwords += 1;
553			if (wqe->wr.send_flags & IB_SEND_SOLICITED)
554				bth0 |= 1 << 23;
555		}
556		bth2 |= 1 << 31;	/* Request ACK. */
557		qp->s_cur++;
558		if (qp->s_cur >= qp->s_size)
559			qp->s_cur = 0;
560		break;
561
562	case OP(RDMA_READ_RESPONSE_MIDDLE):
563		/*
564		 * This case can only happen if a RDMA read is restarted.
565		 * See ipath_restart_rc().
566		 */
567		ipath_init_restart(qp, wqe);
568		len = ((qp->s_psn - wqe->psn) & IPATH_PSN_MASK) * pmtu;
569		ohdr->u.rc.reth.vaddr =
570			cpu_to_be64(wqe->wr.wr.rdma.remote_addr + len);
571		ohdr->u.rc.reth.rkey =
572			cpu_to_be32(wqe->wr.wr.rdma.rkey);
573		ohdr->u.rc.reth.length = cpu_to_be32(qp->s_len);
574		qp->s_state = OP(RDMA_READ_REQUEST);
575		hwords += sizeof(ohdr->u.rc.reth) / sizeof(u32);
576		bth2 = qp->s_psn & IPATH_PSN_MASK;
577		qp->s_psn = wqe->lpsn + 1;
578		ss = NULL;
579		len = 0;
580		qp->s_cur++;
581		if (qp->s_cur == qp->s_size)
582			qp->s_cur = 0;
583		break;
584	}
585	if (ipath_cmp24(qp->s_psn, qp->s_last_psn + IPATH_PSN_CREDIT - 1) >= 0)
586		bth2 |= 1 << 31;	/* Request ACK. */
587	qp->s_len -= len;
588	qp->s_hdrwords = hwords;
589	qp->s_cur_sge = ss;
590	qp->s_cur_size = len;
591	ipath_make_ruc_header(dev, qp, ohdr, bth0 | (qp->s_state << 24), bth2);
592done:
593	ret = 1;
594	goto unlock;
595
596bail:
597	qp->s_flags &= ~IPATH_S_BUSY;
598unlock:
599	spin_unlock_irqrestore(&qp->s_lock, flags);
600	return ret;
601}
602
603/**
604 * send_rc_ack - Construct an ACK packet and send it
605 * @qp: a pointer to the QP
606 *
607 * This is called from ipath_rc_rcv() and only uses the receive
608 * side QP state.
609 * Note that RDMA reads and atomics are handled in the
610 * send side QP state and tasklet.
611 */
612static void send_rc_ack(struct ipath_qp *qp)
613{
614	struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
615	struct ipath_devdata *dd;
616	u16 lrh0;
617	u32 bth0;
618	u32 hwords;
619	u32 __iomem *piobuf;
620	struct ipath_ib_header hdr;
621	struct ipath_other_headers *ohdr;
622	unsigned long flags;
623
624	spin_lock_irqsave(&qp->s_lock, flags);
625
626	/* Don't send ACK or NAK if a RDMA read or atomic is pending. */
627	if (qp->r_head_ack_queue != qp->s_tail_ack_queue ||
628	    (qp->s_flags & IPATH_S_ACK_PENDING) ||
629	    qp->s_ack_state != OP(ACKNOWLEDGE))
630		goto queue_ack;
631
632	spin_unlock_irqrestore(&qp->s_lock, flags);
633
634	/* Don't try to send ACKs if the link isn't ACTIVE */
635	dd = dev->dd;
636	if (!(dd->ipath_flags & IPATH_LINKACTIVE))
637		goto done;
638
639	piobuf = ipath_getpiobuf(dd, 0, NULL);
640	if (!piobuf) {
641		/*
642		 * We are out of PIO buffers at the moment.
643		 * Pass responsibility for sending the ACK to the
644		 * send tasklet so that when a PIO buffer becomes
645		 * available, the ACK is sent ahead of other outgoing
646		 * packets.
647		 */
648		spin_lock_irqsave(&qp->s_lock, flags);
649		goto queue_ack;
650	}
651
652	/* Construct the header. */
653	ohdr = &hdr.u.oth;
654	lrh0 = IPATH_LRH_BTH;
655	/* header size in 32-bit words LRH+BTH+AETH = (8+12+4)/4. */
656	hwords = 6;
657	if (unlikely(qp->remote_ah_attr.ah_flags & IB_AH_GRH)) {
658		hwords += ipath_make_grh(dev, &hdr.u.l.grh,
659					 &qp->remote_ah_attr.grh,
660					 hwords, 0);
661		ohdr = &hdr.u.l.oth;
662		lrh0 = IPATH_LRH_GRH;
663	}
664	/* read pkey_index w/o lock (its atomic) */
665	bth0 = ipath_get_pkey(dd, qp->s_pkey_index) |
666		(OP(ACKNOWLEDGE) << 24) | (1 << 22);
667	if (qp->r_nak_state)
668		ohdr->u.aeth = cpu_to_be32((qp->r_msn & IPATH_MSN_MASK) |
669					    (qp->r_nak_state <<
670					     IPATH_AETH_CREDIT_SHIFT));
671	else
672		ohdr->u.aeth = ipath_compute_aeth(qp);
673	lrh0 |= qp->remote_ah_attr.sl << 4;
674	hdr.lrh[0] = cpu_to_be16(lrh0);
675	hdr.lrh[1] = cpu_to_be16(qp->remote_ah_attr.dlid);
676	hdr.lrh[2] = cpu_to_be16(hwords + SIZE_OF_CRC);
677	hdr.lrh[3] = cpu_to_be16(dd->ipath_lid |
678				 qp->remote_ah_attr.src_path_bits);
679	ohdr->bth[0] = cpu_to_be32(bth0);
680	ohdr->bth[1] = cpu_to_be32(qp->remote_qpn);
681	ohdr->bth[2] = cpu_to_be32(qp->r_ack_psn & IPATH_PSN_MASK);
682
683	writeq(hwords + 1, piobuf);
684
685	if (dd->ipath_flags & IPATH_PIO_FLUSH_WC) {
686		u32 *hdrp = (u32 *) &hdr;
687
688		ipath_flush_wc();
689		__iowrite32_copy(piobuf + 2, hdrp, hwords - 1);
690		ipath_flush_wc();
691		__raw_writel(hdrp[hwords - 1], piobuf + hwords + 1);
692	} else
693		__iowrite32_copy(piobuf + 2, (u32 *) &hdr, hwords);
694
695	ipath_flush_wc();
696
697	dev->n_unicast_xmit++;
698	goto done;
699
700queue_ack:
701	if (ib_ipath_state_ops[qp->state] & IPATH_PROCESS_RECV_OK) {
702		dev->n_rc_qacks++;
703		qp->s_flags |= IPATH_S_ACK_PENDING;
704		qp->s_nak_state = qp->r_nak_state;
705		qp->s_ack_psn = qp->r_ack_psn;
706
707		/* Schedule the send tasklet. */
708		ipath_schedule_send(qp);
709	}
710	spin_unlock_irqrestore(&qp->s_lock, flags);
711done:
712	return;
713}
714
715/**
716 * reset_psn - reset the QP state to send starting from PSN
717 * @qp: the QP
718 * @psn: the packet sequence number to restart at
719 *
720 * This is called from ipath_rc_rcv() to process an incoming RC ACK
721 * for the given QP.
722 * Called at interrupt level with the QP s_lock held.
723 */
724static void reset_psn(struct ipath_qp *qp, u32 psn)
725{
726	u32 n = qp->s_last;
727	struct ipath_swqe *wqe = get_swqe_ptr(qp, n);
728	u32 opcode;
729
730	qp->s_cur = n;
731
732	/*
733	 * If we are starting the request from the beginning,
734	 * let the normal send code handle initialization.
735	 */
736	if (ipath_cmp24(psn, wqe->psn) <= 0) {
737		qp->s_state = OP(SEND_LAST);
738		goto done;
739	}
740
741	/* Find the work request opcode corresponding to the given PSN. */
742	opcode = wqe->wr.opcode;
743	for (;;) {
744		int diff;
745
746		if (++n == qp->s_size)
747			n = 0;
748		if (n == qp->s_tail)
749			break;
750		wqe = get_swqe_ptr(qp, n);
751		diff = ipath_cmp24(psn, wqe->psn);
752		if (diff < 0)
753			break;
754		qp->s_cur = n;
755		/*
756		 * If we are starting the request from the beginning,
757		 * let the normal send code handle initialization.
758		 */
759		if (diff == 0) {
760			qp->s_state = OP(SEND_LAST);
761			goto done;
762		}
763		opcode = wqe->wr.opcode;
764	}
765
766	/*
767	 * Set the state to restart in the middle of a request.
768	 * Don't change the s_sge, s_cur_sge, or s_cur_size.
769	 * See ipath_make_rc_req().
770	 */
771	switch (opcode) {
772	case IB_WR_SEND:
773	case IB_WR_SEND_WITH_IMM:
774		qp->s_state = OP(RDMA_READ_RESPONSE_FIRST);
775		break;
776
777	case IB_WR_RDMA_WRITE:
778	case IB_WR_RDMA_WRITE_WITH_IMM:
779		qp->s_state = OP(RDMA_READ_RESPONSE_LAST);
780		break;
781
782	case IB_WR_RDMA_READ:
783		qp->s_state = OP(RDMA_READ_RESPONSE_MIDDLE);
784		break;
785
786	default:
787		/*
788		 * This case shouldn't happen since its only
789		 * one PSN per req.
790		 */
791		qp->s_state = OP(SEND_LAST);
792	}
793done:
794	qp->s_psn = psn;
795}
796
797/**
798 * ipath_restart_rc - back up requester to resend the last un-ACKed request
799 * @qp: the QP to restart
800 * @psn: packet sequence number for the request
801 * @wc: the work completion request
802 *
803 * The QP s_lock should be held and interrupts disabled.
804 */
805void ipath_restart_rc(struct ipath_qp *qp, u32 psn)
806{
807	struct ipath_swqe *wqe = get_swqe_ptr(qp, qp->s_last);
808	struct ipath_ibdev *dev;
809
810	if (qp->s_retry == 0) {
811		ipath_send_complete(qp, wqe, IB_WC_RETRY_EXC_ERR);
812		ipath_error_qp(qp, IB_WC_WR_FLUSH_ERR);
813		goto bail;
814	}
815	qp->s_retry--;
816
817	/*
818	 * Remove the QP from the timeout queue.
819	 * Note: it may already have been removed by ipath_ib_timer().
820	 */
821	dev = to_idev(qp->ibqp.device);
822	spin_lock(&dev->pending_lock);
823	if (!list_empty(&qp->timerwait))
824		list_del_init(&qp->timerwait);
825	if (!list_empty(&qp->piowait))
826		list_del_init(&qp->piowait);
827	spin_unlock(&dev->pending_lock);
828
829	if (wqe->wr.opcode == IB_WR_RDMA_READ)
830		dev->n_rc_resends++;
831	else
832		dev->n_rc_resends += (qp->s_psn - psn) & IPATH_PSN_MASK;
833
834	reset_psn(qp, psn);
835	ipath_schedule_send(qp);
836
837bail:
838	return;
839}
840
841static inline void update_last_psn(struct ipath_qp *qp, u32 psn)
842{
843	qp->s_last_psn = psn;
844}
845
846/**
847 * do_rc_ack - process an incoming RC ACK
848 * @qp: the QP the ACK came in on
849 * @psn: the packet sequence number of the ACK
850 * @opcode: the opcode of the request that resulted in the ACK
851 *
852 * This is called from ipath_rc_rcv_resp() to process an incoming RC ACK
853 * for the given QP.
854 * Called at interrupt level with the QP s_lock held and interrupts disabled.
855 * Returns 1 if OK, 0 if current operation should be aborted (NAK).
856 */
857static int do_rc_ack(struct ipath_qp *qp, u32 aeth, u32 psn, int opcode,
858		     u64 val)
859{
860	struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
861	struct ib_wc wc;
862	enum ib_wc_status status;
863	struct ipath_swqe *wqe;
864	int ret = 0;
865	u32 ack_psn;
866	int diff;
867
868	/*
869	 * Remove the QP from the timeout queue (or RNR timeout queue).
870	 * If ipath_ib_timer() has already removed it,
871	 * it's OK since we hold the QP s_lock and ipath_restart_rc()
872	 * just won't find anything to restart if we ACK everything.
873	 */
874	spin_lock(&dev->pending_lock);
875	if (!list_empty(&qp->timerwait))
876		list_del_init(&qp->timerwait);
877	spin_unlock(&dev->pending_lock);
878
879	/*
880	 * Note that NAKs implicitly ACK outstanding SEND and RDMA write
881	 * requests and implicitly NAK RDMA read and atomic requests issued
882	 * before the NAK'ed request.  The MSN won't include the NAK'ed
883	 * request but will include an ACK'ed request(s).
884	 */
885	ack_psn = psn;
886	if (aeth >> 29)
887		ack_psn--;
888	wqe = get_swqe_ptr(qp, qp->s_last);
889
890	/*
891	 * The MSN might be for a later WQE than the PSN indicates so
892	 * only complete WQEs that the PSN finishes.
893	 */
894	while ((diff = ipath_cmp24(ack_psn, wqe->lpsn)) >= 0) {
895		/*
896		 * RDMA_READ_RESPONSE_ONLY is a special case since
897		 * we want to generate completion events for everything
898		 * before the RDMA read, copy the data, then generate
899		 * the completion for the read.
900		 */
901		if (wqe->wr.opcode == IB_WR_RDMA_READ &&
902		    opcode == OP(RDMA_READ_RESPONSE_ONLY) &&
903		    diff == 0) {
904			ret = 1;
905			goto bail;
906		}
907		/*
908		 * If this request is a RDMA read or atomic, and the ACK is
909		 * for a later operation, this ACK NAKs the RDMA read or
910		 * atomic.  In other words, only a RDMA_READ_LAST or ONLY
911		 * can ACK a RDMA read and likewise for atomic ops.  Note
912		 * that the NAK case can only happen if relaxed ordering is
913		 * used and requests are sent after an RDMA read or atomic
914		 * is sent but before the response is received.
915		 */
916		if ((wqe->wr.opcode == IB_WR_RDMA_READ &&
917		     (opcode != OP(RDMA_READ_RESPONSE_LAST) || diff != 0)) ||
918		    ((wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
919		      wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD) &&
920		     (opcode != OP(ATOMIC_ACKNOWLEDGE) || diff != 0))) {
921			/*
922			 * The last valid PSN seen is the previous
923			 * request's.
924			 */
925			update_last_psn(qp, wqe->psn - 1);
926			/* Retry this request. */
927			ipath_restart_rc(qp, wqe->psn);
928			/*
929			 * No need to process the ACK/NAK since we are
930			 * restarting an earlier request.
931			 */
932			goto bail;
933		}
934		if (wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
935		    wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD)
936			*(u64 *) wqe->sg_list[0].vaddr = val;
937		if (qp->s_num_rd_atomic &&
938		    (wqe->wr.opcode == IB_WR_RDMA_READ ||
939		     wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
940		     wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD)) {
941			qp->s_num_rd_atomic--;
942			/* Restart sending task if fence is complete */
943			if (((qp->s_flags & IPATH_S_FENCE_PENDING) &&
944			     !qp->s_num_rd_atomic) ||
945			    qp->s_flags & IPATH_S_RDMAR_PENDING)
946				ipath_schedule_send(qp);
947		}
948		/* Post a send completion queue entry if requested. */
949		if (!(qp->s_flags & IPATH_S_SIGNAL_REQ_WR) ||
950		    (wqe->wr.send_flags & IB_SEND_SIGNALED)) {
951			memset(&wc, 0, sizeof wc);
952			wc.wr_id = wqe->wr.wr_id;
953			wc.status = IB_WC_SUCCESS;
954			wc.opcode = ib_ipath_wc_opcode[wqe->wr.opcode];
955			wc.byte_len = wqe->length;
956			wc.qp = &qp->ibqp;
957			wc.src_qp = qp->remote_qpn;
958			wc.slid = qp->remote_ah_attr.dlid;
959			wc.sl = qp->remote_ah_attr.sl;
960			ipath_cq_enter(to_icq(qp->ibqp.send_cq), &wc, 0);
961		}
962		qp->s_retry = qp->s_retry_cnt;
963		/*
964		 * If we are completing a request which is in the process of
965		 * being resent, we can stop resending it since we know the
966		 * responder has already seen it.
967		 */
968		if (qp->s_last == qp->s_cur) {
969			if (++qp->s_cur >= qp->s_size)
970				qp->s_cur = 0;
971			qp->s_last = qp->s_cur;
972			if (qp->s_last == qp->s_tail)
973				break;
974			wqe = get_swqe_ptr(qp, qp->s_cur);
975			qp->s_state = OP(SEND_LAST);
976			qp->s_psn = wqe->psn;
977		} else {
978			if (++qp->s_last >= qp->s_size)
979				qp->s_last = 0;
980			if (qp->state == IB_QPS_SQD && qp->s_last == qp->s_cur)
981				qp->s_draining = 0;
982			if (qp->s_last == qp->s_tail)
983				break;
984			wqe = get_swqe_ptr(qp, qp->s_last);
985		}
986	}
987
988	switch (aeth >> 29) {
989	case 0:		/* ACK */
990		dev->n_rc_acks++;
991		/* If this is a partial ACK, reset the retransmit timer. */
992		if (qp->s_last != qp->s_tail) {
993			spin_lock(&dev->pending_lock);
994			if (list_empty(&qp->timerwait))
995				list_add_tail(&qp->timerwait,
996					&dev->pending[dev->pending_index]);
997			spin_unlock(&dev->pending_lock);
998			/*
999			 * If we get a partial ACK for a resent operation,
1000			 * we can stop resending the earlier packets and
1001			 * continue with the next packet the receiver wants.
1002			 */
1003			if (ipath_cmp24(qp->s_psn, psn) <= 0) {
1004				reset_psn(qp, psn + 1);
1005				ipath_schedule_send(qp);
1006			}
1007		} else if (ipath_cmp24(qp->s_psn, psn) <= 0) {
1008			qp->s_state = OP(SEND_LAST);
1009			qp->s_psn = psn + 1;
1010		}
1011		ipath_get_credit(qp, aeth);
1012		qp->s_rnr_retry = qp->s_rnr_retry_cnt;
1013		qp->s_retry = qp->s_retry_cnt;
1014		update_last_psn(qp, psn);
1015		ret = 1;
1016		goto bail;
1017
1018	case 1:		/* RNR NAK */
1019		dev->n_rnr_naks++;
1020		if (qp->s_last == qp->s_tail)
1021			goto bail;
1022		if (qp->s_rnr_retry == 0) {
1023			status = IB_WC_RNR_RETRY_EXC_ERR;
1024			goto class_b;
1025		}
1026		if (qp->s_rnr_retry_cnt < 7)
1027			qp->s_rnr_retry--;
1028
1029		/* The last valid PSN is the previous PSN. */
1030		update_last_psn(qp, psn - 1);
1031
1032		if (wqe->wr.opcode == IB_WR_RDMA_READ)
1033			dev->n_rc_resends++;
1034		else
1035			dev->n_rc_resends +=
1036				(qp->s_psn - psn) & IPATH_PSN_MASK;
1037
1038		reset_psn(qp, psn);
1039
1040		qp->s_rnr_timeout =
1041			ib_ipath_rnr_table[(aeth >> IPATH_AETH_CREDIT_SHIFT) &
1042					   IPATH_AETH_CREDIT_MASK];
1043		ipath_insert_rnr_queue(qp);
1044		ipath_schedule_send(qp);
1045		goto bail;
1046
1047	case 3:		/* NAK */
1048		if (qp->s_last == qp->s_tail)
1049			goto bail;
1050		/* The last valid PSN is the previous PSN. */
1051		update_last_psn(qp, psn - 1);
1052		switch ((aeth >> IPATH_AETH_CREDIT_SHIFT) &
1053			IPATH_AETH_CREDIT_MASK) {
1054		case 0:	/* PSN sequence error */
1055			dev->n_seq_naks++;
1056			/*
1057			 * Back up to the responder's expected PSN.
1058			 * Note that we might get a NAK in the middle of an
1059			 * RDMA READ response which terminates the RDMA
1060			 * READ.
1061			 */
1062			ipath_restart_rc(qp, psn);
1063			break;
1064
1065		case 1:	/* Invalid Request */
1066			status = IB_WC_REM_INV_REQ_ERR;
1067			dev->n_other_naks++;
1068			goto class_b;
1069
1070		case 2:	/* Remote Access Error */
1071			status = IB_WC_REM_ACCESS_ERR;
1072			dev->n_other_naks++;
1073			goto class_b;
1074
1075		case 3:	/* Remote Operation Error */
1076			status = IB_WC_REM_OP_ERR;
1077			dev->n_other_naks++;
1078		class_b:
1079			ipath_send_complete(qp, wqe, status);
1080			ipath_error_qp(qp, IB_WC_WR_FLUSH_ERR);
1081			break;
1082
1083		default:
1084			/* Ignore other reserved NAK error codes */
1085			goto reserved;
1086		}
1087		qp->s_rnr_retry = qp->s_rnr_retry_cnt;
1088		goto bail;
1089
1090	default:		/* 2: reserved */
1091	reserved:
1092		/* Ignore reserved NAK codes. */
1093		goto bail;
1094	}
1095
1096bail:
1097	return ret;
1098}
1099
1100/**
1101 * ipath_rc_rcv_resp - process an incoming RC response packet
1102 * @dev: the device this packet came in on
1103 * @ohdr: the other headers for this packet
1104 * @data: the packet data
1105 * @tlen: the packet length
1106 * @qp: the QP for this packet
1107 * @opcode: the opcode for this packet
1108 * @psn: the packet sequence number for this packet
1109 * @hdrsize: the header length
1110 * @pmtu: the path MTU
1111 * @header_in_data: true if part of the header data is in the data buffer
1112 *
1113 * This is called from ipath_rc_rcv() to process an incoming RC response
1114 * packet for the given QP.
1115 * Called at interrupt level.
1116 */
1117static inline void ipath_rc_rcv_resp(struct ipath_ibdev *dev,
1118				     struct ipath_other_headers *ohdr,
1119				     void *data, u32 tlen,
1120				     struct ipath_qp *qp,
1121				     u32 opcode,
1122				     u32 psn, u32 hdrsize, u32 pmtu,
1123				     int header_in_data)
1124{
1125	struct ipath_swqe *wqe;
1126	enum ib_wc_status status;
1127	unsigned long flags;
1128	int diff;
1129	u32 pad;
1130	u32 aeth;
1131	u64 val;
1132
1133	spin_lock_irqsave(&qp->s_lock, flags);
1134
1135	/* Double check we can process this now that we hold the s_lock. */
1136	if (!(ib_ipath_state_ops[qp->state] & IPATH_PROCESS_RECV_OK))
1137		goto ack_done;
1138
1139	/* Ignore invalid responses. */
1140	if (ipath_cmp24(psn, qp->s_next_psn) >= 0)
1141		goto ack_done;
1142
1143	/* Ignore duplicate responses. */
1144	diff = ipath_cmp24(psn, qp->s_last_psn);
1145	if (unlikely(diff <= 0)) {
1146		/* Update credits for "ghost" ACKs */
1147		if (diff == 0 && opcode == OP(ACKNOWLEDGE)) {
1148			if (!header_in_data)
1149				aeth = be32_to_cpu(ohdr->u.aeth);
1150			else {
1151				aeth = be32_to_cpu(((__be32 *) data)[0]);
1152				data += sizeof(__be32);
1153			}
1154			if ((aeth >> 29) == 0)
1155				ipath_get_credit(qp, aeth);
1156		}
1157		goto ack_done;
1158	}
1159
1160	if (unlikely(qp->s_last == qp->s_tail))
1161		goto ack_done;
1162	wqe = get_swqe_ptr(qp, qp->s_last);
1163	status = IB_WC_SUCCESS;
1164
1165	switch (opcode) {
1166	case OP(ACKNOWLEDGE):
1167	case OP(ATOMIC_ACKNOWLEDGE):
1168	case OP(RDMA_READ_RESPONSE_FIRST):
1169		if (!header_in_data)
1170			aeth = be32_to_cpu(ohdr->u.aeth);
1171		else {
1172			aeth = be32_to_cpu(((__be32 *) data)[0]);
1173			data += sizeof(__be32);
1174		}
1175		if (opcode == OP(ATOMIC_ACKNOWLEDGE)) {
1176			if (!header_in_data) {
1177				__be32 *p = ohdr->u.at.atomic_ack_eth;
1178
1179				val = ((u64) be32_to_cpu(p[0]) << 32) |
1180					be32_to_cpu(p[1]);
1181			} else
1182				val = be64_to_cpu(((__be64 *) data)[0]);
1183		} else
1184			val = 0;
1185		if (!do_rc_ack(qp, aeth, psn, opcode, val) ||
1186		    opcode != OP(RDMA_READ_RESPONSE_FIRST))
1187			goto ack_done;
1188		hdrsize += 4;
1189		wqe = get_swqe_ptr(qp, qp->s_last);
1190		if (unlikely(wqe->wr.opcode != IB_WR_RDMA_READ))
1191			goto ack_op_err;
1192		qp->r_flags &= ~IPATH_R_RDMAR_SEQ;
1193		/*
1194		 * If this is a response to a resent RDMA read, we
1195		 * have to be careful to copy the data to the right
1196		 * location.
1197		 */
1198		qp->s_rdma_read_len = restart_sge(&qp->s_rdma_read_sge,
1199						  wqe, psn, pmtu);
1200		goto read_middle;
1201
1202	case OP(RDMA_READ_RESPONSE_MIDDLE):
1203		/* no AETH, no ACK */
1204		if (unlikely(ipath_cmp24(psn, qp->s_last_psn + 1))) {
1205			dev->n_rdma_seq++;
1206			if (qp->r_flags & IPATH_R_RDMAR_SEQ)
1207				goto ack_done;
1208			qp->r_flags |= IPATH_R_RDMAR_SEQ;
1209			ipath_restart_rc(qp, qp->s_last_psn + 1);
1210			goto ack_done;
1211		}
1212		if (unlikely(wqe->wr.opcode != IB_WR_RDMA_READ))
1213			goto ack_op_err;
1214	read_middle:
1215		if (unlikely(tlen != (hdrsize + pmtu + 4)))
1216			goto ack_len_err;
1217		if (unlikely(pmtu >= qp->s_rdma_read_len))
1218			goto ack_len_err;
1219
1220		/* We got a response so update the timeout. */
1221		spin_lock(&dev->pending_lock);
1222		if (qp->s_rnr_timeout == 0 && !list_empty(&qp->timerwait))
1223			list_move_tail(&qp->timerwait,
1224				       &dev->pending[dev->pending_index]);
1225		spin_unlock(&dev->pending_lock);
1226
1227		if (opcode == OP(RDMA_READ_RESPONSE_MIDDLE))
1228			qp->s_retry = qp->s_retry_cnt;
1229
1230		/*
1231		 * Update the RDMA receive state but do the copy w/o
1232		 * holding the locks and blocking interrupts.
1233		 */
1234		qp->s_rdma_read_len -= pmtu;
1235		update_last_psn(qp, psn);
1236		spin_unlock_irqrestore(&qp->s_lock, flags);
1237		ipath_copy_sge(&qp->s_rdma_read_sge, data, pmtu);
1238		goto bail;
1239
1240	case OP(RDMA_READ_RESPONSE_ONLY):
1241		if (!header_in_data)
1242			aeth = be32_to_cpu(ohdr->u.aeth);
1243		else
1244			aeth = be32_to_cpu(((__be32 *) data)[0]);
1245		if (!do_rc_ack(qp, aeth, psn, opcode, 0))
1246			goto ack_done;
1247		/* Get the number of bytes the message was padded by. */
1248		pad = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3;
1249		/*
1250		 * Check that the data size is >= 0 && <= pmtu.
1251		 * Remember to account for the AETH header (4) and
1252		 * ICRC (4).
1253		 */
1254		if (unlikely(tlen < (hdrsize + pad + 8)))
1255			goto ack_len_err;
1256		/*
1257		 * If this is a response to a resent RDMA read, we
1258		 * have to be careful to copy the data to the right
1259		 * location.
1260		 */
1261		wqe = get_swqe_ptr(qp, qp->s_last);
1262		qp->s_rdma_read_len = restart_sge(&qp->s_rdma_read_sge,
1263						  wqe, psn, pmtu);
1264		goto read_last;
1265
1266	case OP(RDMA_READ_RESPONSE_LAST):
1267		/* ACKs READ req. */
1268		if (unlikely(ipath_cmp24(psn, qp->s_last_psn + 1))) {
1269			dev->n_rdma_seq++;
1270			if (qp->r_flags & IPATH_R_RDMAR_SEQ)
1271				goto ack_done;
1272			qp->r_flags |= IPATH_R_RDMAR_SEQ;
1273			ipath_restart_rc(qp, qp->s_last_psn + 1);
1274			goto ack_done;
1275		}
1276		if (unlikely(wqe->wr.opcode != IB_WR_RDMA_READ))
1277			goto ack_op_err;
1278		/* Get the number of bytes the message was padded by. */
1279		pad = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3;
1280		/*
1281		 * Check that the data size is >= 1 && <= pmtu.
1282		 * Remember to account for the AETH header (4) and
1283		 * ICRC (4).
1284		 */
1285		if (unlikely(tlen <= (hdrsize + pad + 8)))
1286			goto ack_len_err;
1287	read_last:
1288		tlen -= hdrsize + pad + 8;
1289		if (unlikely(tlen != qp->s_rdma_read_len))
1290			goto ack_len_err;
1291		if (!header_in_data)
1292			aeth = be32_to_cpu(ohdr->u.aeth);
1293		else {
1294			aeth = be32_to_cpu(((__be32 *) data)[0]);
1295			data += sizeof(__be32);
1296		}
1297		ipath_copy_sge(&qp->s_rdma_read_sge, data, tlen);
1298		(void) do_rc_ack(qp, aeth, psn,
1299				 OP(RDMA_READ_RESPONSE_LAST), 0);
1300		goto ack_done;
1301	}
1302
1303ack_op_err:
1304	status = IB_WC_LOC_QP_OP_ERR;
1305	goto ack_err;
1306
1307ack_len_err:
1308	status = IB_WC_LOC_LEN_ERR;
1309ack_err:
1310	ipath_send_complete(qp, wqe, status);
1311	ipath_error_qp(qp, IB_WC_WR_FLUSH_ERR);
1312ack_done:
1313	spin_unlock_irqrestore(&qp->s_lock, flags);
1314bail:
1315	return;
1316}
1317
1318/**
1319 * ipath_rc_rcv_error - process an incoming duplicate or error RC packet
1320 * @dev: the device this packet came in on
1321 * @ohdr: the other headers for this packet
1322 * @data: the packet data
1323 * @qp: the QP for this packet
1324 * @opcode: the opcode for this packet
1325 * @psn: the packet sequence number for this packet
1326 * @diff: the difference between the PSN and the expected PSN
1327 * @header_in_data: true if part of the header data is in the data buffer
1328 *
1329 * This is called from ipath_rc_rcv() to process an unexpected
1330 * incoming RC packet for the given QP.
1331 * Called at interrupt level.
1332 * Return 1 if no more processing is needed; otherwise return 0 to
1333 * schedule a response to be sent.
1334 */
1335static inline int ipath_rc_rcv_error(struct ipath_ibdev *dev,
1336				     struct ipath_other_headers *ohdr,
1337				     void *data,
1338				     struct ipath_qp *qp,
1339				     u32 opcode,
1340				     u32 psn,
1341				     int diff,
1342				     int header_in_data)
1343{
1344	struct ipath_ack_entry *e;
1345	u8 i, prev;
1346	int old_req;
1347	unsigned long flags;
1348
1349	if (diff > 0) {
1350		/*
1351		 * Packet sequence error.
1352		 * A NAK will ACK earlier sends and RDMA writes.
1353		 * Don't queue the NAK if we already sent one.
1354		 */
1355		if (!qp->r_nak_state) {
1356			qp->r_nak_state = IB_NAK_PSN_ERROR;
1357			/* Use the expected PSN. */
1358			qp->r_ack_psn = qp->r_psn;
1359			goto send_ack;
1360		}
1361		goto done;
1362	}
1363
1364	/*
1365	 * Handle a duplicate request.  Don't re-execute SEND, RDMA
1366	 * write or atomic op.  Don't NAK errors, just silently drop
1367	 * the duplicate request.  Note that r_sge, r_len, and
1368	 * r_rcv_len may be in use so don't modify them.
1369	 *
1370	 * We are supposed to ACK the earliest duplicate PSN but we
1371	 * can coalesce an outstanding duplicate ACK.  We have to
1372	 * send the earliest so that RDMA reads can be restarted at
1373	 * the requester's expected PSN.
1374	 *
1375	 * First, find where this duplicate PSN falls within the
1376	 * ACKs previously sent.
1377	 */
1378	psn &= IPATH_PSN_MASK;
1379	e = NULL;
1380	old_req = 1;
1381
1382	spin_lock_irqsave(&qp->s_lock, flags);
1383	/* Double check we can process this now that we hold the s_lock. */
1384	if (!(ib_ipath_state_ops[qp->state] & IPATH_PROCESS_RECV_OK))
1385		goto unlock_done;
1386
1387	for (i = qp->r_head_ack_queue; ; i = prev) {
1388		if (i == qp->s_tail_ack_queue)
1389			old_req = 0;
1390		if (i)
1391			prev = i - 1;
1392		else
1393			prev = IPATH_MAX_RDMA_ATOMIC;
1394		if (prev == qp->r_head_ack_queue) {
1395			e = NULL;
1396			break;
1397		}
1398		e = &qp->s_ack_queue[prev];
1399		if (!e->opcode) {
1400			e = NULL;
1401			break;
1402		}
1403		if (ipath_cmp24(psn, e->psn) >= 0) {
1404			if (prev == qp->s_tail_ack_queue)
1405				old_req = 0;
1406			break;
1407		}
1408	}
1409	switch (opcode) {
1410	case OP(RDMA_READ_REQUEST): {
1411		struct ib_reth *reth;
1412		u32 offset;
1413		u32 len;
1414
1415		/*
1416		 * If we didn't find the RDMA read request in the ack queue,
1417		 * or the send tasklet is already backed up to send an
1418		 * earlier entry, we can ignore this request.
1419		 */
1420		if (!e || e->opcode != OP(RDMA_READ_REQUEST) || old_req)
1421			goto unlock_done;
1422		/* RETH comes after BTH */
1423		if (!header_in_data)
1424			reth = &ohdr->u.rc.reth;
1425		else {
1426			reth = (struct ib_reth *)data;
1427			data += sizeof(*reth);
1428		}
1429		/*
1430		 * Address range must be a subset of the original
1431		 * request and start on pmtu boundaries.
1432		 * We reuse the old ack_queue slot since the requester
1433		 * should not back up and request an earlier PSN for the
1434		 * same request.
1435		 */
1436		offset = ((psn - e->psn) & IPATH_PSN_MASK) *
1437			ib_mtu_enum_to_int(qp->path_mtu);
1438		len = be32_to_cpu(reth->length);
1439		if (unlikely(offset + len > e->rdma_sge.sge.sge_length))
1440			goto unlock_done;
1441		if (len != 0) {
1442			u32 rkey = be32_to_cpu(reth->rkey);
1443			u64 vaddr = be64_to_cpu(reth->vaddr);
1444			int ok;
1445
1446			ok = ipath_rkey_ok(qp, &e->rdma_sge,
1447					   len, vaddr, rkey,
1448					   IB_ACCESS_REMOTE_READ);
1449			if (unlikely(!ok))
1450				goto unlock_done;
1451		} else {
1452			e->rdma_sge.sg_list = NULL;
1453			e->rdma_sge.num_sge = 0;
1454			e->rdma_sge.sge.mr = NULL;
1455			e->rdma_sge.sge.vaddr = NULL;
1456			e->rdma_sge.sge.length = 0;
1457			e->rdma_sge.sge.sge_length = 0;
1458		}
1459		e->psn = psn;
1460		qp->s_ack_state = OP(ACKNOWLEDGE);
1461		qp->s_tail_ack_queue = prev;
1462		break;
1463	}
1464
1465	case OP(COMPARE_SWAP):
1466	case OP(FETCH_ADD): {
1467		/*
1468		 * If we didn't find the atomic request in the ack queue
1469		 * or the send tasklet is already backed up to send an
1470		 * earlier entry, we can ignore this request.
1471		 */
1472		if (!e || e->opcode != (u8) opcode || old_req)
1473			goto unlock_done;
1474		qp->s_ack_state = OP(ACKNOWLEDGE);
1475		qp->s_tail_ack_queue = prev;
1476		break;
1477	}
1478
1479	default:
1480		if (old_req)
1481			goto unlock_done;
1482		/*
1483		 * Resend the most recent ACK if this request is
1484		 * after all the previous RDMA reads and atomics.
1485		 */
1486		if (i == qp->r_head_ack_queue) {
1487			spin_unlock_irqrestore(&qp->s_lock, flags);
1488			qp->r_nak_state = 0;
1489			qp->r_ack_psn = qp->r_psn - 1;
1490			goto send_ack;
1491		}
1492		/*
1493		 * Try to send a simple ACK to work around a Mellanox bug
1494		 * which doesn't accept a RDMA read response or atomic
1495		 * response as an ACK for earlier SENDs or RDMA writes.
1496		 */
1497		if (qp->r_head_ack_queue == qp->s_tail_ack_queue &&
1498		    !(qp->s_flags & IPATH_S_ACK_PENDING) &&
1499		    qp->s_ack_state == OP(ACKNOWLEDGE)) {
1500			spin_unlock_irqrestore(&qp->s_lock, flags);
1501			qp->r_nak_state = 0;
1502			qp->r_ack_psn = qp->s_ack_queue[i].psn - 1;
1503			goto send_ack;
1504		}
1505		/*
1506		 * Resend the RDMA read or atomic op which
1507		 * ACKs this duplicate request.
1508		 */
1509		qp->s_ack_state = OP(ACKNOWLEDGE);
1510		qp->s_tail_ack_queue = i;
1511		break;
1512	}
1513	qp->r_nak_state = 0;
1514	ipath_schedule_send(qp);
1515
1516unlock_done:
1517	spin_unlock_irqrestore(&qp->s_lock, flags);
1518done:
1519	return 1;
1520
1521send_ack:
1522	return 0;
1523}
1524
1525void ipath_rc_error(struct ipath_qp *qp, enum ib_wc_status err)
1526{
1527	unsigned long flags;
1528	int lastwqe;
1529
1530	spin_lock_irqsave(&qp->s_lock, flags);
1531	lastwqe = ipath_error_qp(qp, err);
1532	spin_unlock_irqrestore(&qp->s_lock, flags);
1533
1534	if (lastwqe) {
1535		struct ib_event ev;
1536
1537		ev.device = qp->ibqp.device;
1538		ev.element.qp = &qp->ibqp;
1539		ev.event = IB_EVENT_QP_LAST_WQE_REACHED;
1540		qp->ibqp.event_handler(&ev, qp->ibqp.qp_context);
1541	}
1542}
1543
1544static inline void ipath_update_ack_queue(struct ipath_qp *qp, unsigned n)
1545{
1546	unsigned next;
1547
1548	next = n + 1;
1549	if (next > IPATH_MAX_RDMA_ATOMIC)
1550		next = 0;
1551	if (n == qp->s_tail_ack_queue) {
1552		qp->s_tail_ack_queue = next;
1553		qp->s_ack_state = OP(ACKNOWLEDGE);
1554	}
1555}
1556
1557/**
1558 * ipath_rc_rcv - process an incoming RC packet
1559 * @dev: the device this packet came in on
1560 * @hdr: the header of this packet
1561 * @has_grh: true if the header has a GRH
1562 * @data: the packet data
1563 * @tlen: the packet length
1564 * @qp: the QP for this packet
1565 *
1566 * This is called from ipath_qp_rcv() to process an incoming RC packet
1567 * for the given QP.
1568 * Called at interrupt level.
1569 */
1570void ipath_rc_rcv(struct ipath_ibdev *dev, struct ipath_ib_header *hdr,
1571		  int has_grh, void *data, u32 tlen, struct ipath_qp *qp)
1572{
1573	struct ipath_other_headers *ohdr;
1574	u32 opcode;
1575	u32 hdrsize;
1576	u32 psn;
1577	u32 pad;
1578	struct ib_wc wc;
1579	u32 pmtu = ib_mtu_enum_to_int(qp->path_mtu);
1580	int diff;
1581	struct ib_reth *reth;
1582	int header_in_data;
1583	unsigned long flags;
1584
1585	/* Validate the SLID. See Ch. 9.6.1.5 */
1586	if (unlikely(be16_to_cpu(hdr->lrh[3]) != qp->remote_ah_attr.dlid))
1587		goto done;
1588
1589	/* Check for GRH */
1590	if (!has_grh) {
1591		ohdr = &hdr->u.oth;
1592		hdrsize = 8 + 12;	/* LRH + BTH */
1593		psn = be32_to_cpu(ohdr->bth[2]);
1594		header_in_data = 0;
1595	} else {
1596		ohdr = &hdr->u.l.oth;
1597		hdrsize = 8 + 40 + 12;	/* LRH + GRH + BTH */
1598		/*
1599		 * The header with GRH is 60 bytes and the core driver sets
1600		 * the eager header buffer size to 56 bytes so the last 4
1601		 * bytes of the BTH header (PSN) is in the data buffer.
1602		 */
1603		header_in_data = dev->dd->ipath_rcvhdrentsize == 16;
1604		if (header_in_data) {
1605			psn = be32_to_cpu(((__be32 *) data)[0]);
1606			data += sizeof(__be32);
1607		} else
1608			psn = be32_to_cpu(ohdr->bth[2]);
1609	}
1610
1611	/*
1612	 * Process responses (ACKs) before anything else.  Note that the
1613	 * packet sequence number will be for something in the send work
1614	 * queue rather than the expected receive packet sequence number.
1615	 * In other words, this QP is the requester.
1616	 */
1617	opcode = be32_to_cpu(ohdr->bth[0]) >> 24;
1618	if (opcode >= OP(RDMA_READ_RESPONSE_FIRST) &&
1619	    opcode <= OP(ATOMIC_ACKNOWLEDGE)) {
1620		ipath_rc_rcv_resp(dev, ohdr, data, tlen, qp, opcode, psn,
1621				  hdrsize, pmtu, header_in_data);
1622		goto done;
1623	}
1624
1625	/* Compute 24 bits worth of difference. */
1626	diff = ipath_cmp24(psn, qp->r_psn);
1627	if (unlikely(diff)) {
1628		if (ipath_rc_rcv_error(dev, ohdr, data, qp, opcode,
1629				       psn, diff, header_in_data))
1630			goto done;
1631		goto send_ack;
1632	}
1633
1634	/* Check for opcode sequence errors. */
1635	switch (qp->r_state) {
1636	case OP(SEND_FIRST):
1637	case OP(SEND_MIDDLE):
1638		if (opcode == OP(SEND_MIDDLE) ||
1639		    opcode == OP(SEND_LAST) ||
1640		    opcode == OP(SEND_LAST_WITH_IMMEDIATE))
1641			break;
1642		goto nack_inv;
1643
1644	case OP(RDMA_WRITE_FIRST):
1645	case OP(RDMA_WRITE_MIDDLE):
1646		if (opcode == OP(RDMA_WRITE_MIDDLE) ||
1647		    opcode == OP(RDMA_WRITE_LAST) ||
1648		    opcode == OP(RDMA_WRITE_LAST_WITH_IMMEDIATE))
1649			break;
1650		goto nack_inv;
1651
1652	default:
1653		if (opcode == OP(SEND_MIDDLE) ||
1654		    opcode == OP(SEND_LAST) ||
1655		    opcode == OP(SEND_LAST_WITH_IMMEDIATE) ||
1656		    opcode == OP(RDMA_WRITE_MIDDLE) ||
1657		    opcode == OP(RDMA_WRITE_LAST) ||
1658		    opcode == OP(RDMA_WRITE_LAST_WITH_IMMEDIATE))
1659			goto nack_inv;
1660		/*
1661		 * Note that it is up to the requester to not send a new
1662		 * RDMA read or atomic operation before receiving an ACK
1663		 * for the previous operation.
1664		 */
1665		break;
1666	}
1667
1668	memset(&wc, 0, sizeof wc);
1669
1670	/* OK, process the packet. */
1671	switch (opcode) {
1672	case OP(SEND_FIRST):
1673		if (!ipath_get_rwqe(qp, 0))
1674			goto rnr_nak;
1675		qp->r_rcv_len = 0;
1676		/* FALLTHROUGH */
1677	case OP(SEND_MIDDLE):
1678	case OP(RDMA_WRITE_MIDDLE):
1679	send_middle:
1680		/* Check for invalid length PMTU or posted rwqe len. */
1681		if (unlikely(tlen != (hdrsize + pmtu + 4)))
1682			goto nack_inv;
1683		qp->r_rcv_len += pmtu;
1684		if (unlikely(qp->r_rcv_len > qp->r_len))
1685			goto nack_inv;
1686		ipath_copy_sge(&qp->r_sge, data, pmtu);
1687		break;
1688
1689	case OP(RDMA_WRITE_LAST_WITH_IMMEDIATE):
1690		/* consume RWQE */
1691		if (!ipath_get_rwqe(qp, 1))
1692			goto rnr_nak;
1693		goto send_last_imm;
1694
1695	case OP(SEND_ONLY):
1696	case OP(SEND_ONLY_WITH_IMMEDIATE):
1697		if (!ipath_get_rwqe(qp, 0))
1698			goto rnr_nak;
1699		qp->r_rcv_len = 0;
1700		if (opcode == OP(SEND_ONLY))
1701			goto send_last;
1702		/* FALLTHROUGH */
1703	case OP(SEND_LAST_WITH_IMMEDIATE):
1704	send_last_imm:
1705		if (header_in_data) {
1706			wc.ex.imm_data = *(__be32 *) data;
1707			data += sizeof(__be32);
1708		} else {
1709			/* Immediate data comes after BTH */
1710			wc.ex.imm_data = ohdr->u.imm_data;
1711		}
1712		hdrsize += 4;
1713		wc.wc_flags = IB_WC_WITH_IMM;
1714		/* FALLTHROUGH */
1715	case OP(SEND_LAST):
1716	case OP(RDMA_WRITE_LAST):
1717	send_last:
1718		/* Get the number of bytes the message was padded by. */
1719		pad = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3;
1720		/* Check for invalid length. */
1721		/* XXX LAST len should be >= 1 */
1722		if (unlikely(tlen < (hdrsize + pad + 4)))
1723			goto nack_inv;
1724		/* Don't count the CRC. */
1725		tlen -= (hdrsize + pad + 4);
1726		wc.byte_len = tlen + qp->r_rcv_len;
1727		if (unlikely(wc.byte_len > qp->r_len))
1728			goto nack_inv;
1729		ipath_copy_sge(&qp->r_sge, data, tlen);
1730		qp->r_msn++;
1731		if (!test_and_clear_bit(IPATH_R_WRID_VALID, &qp->r_aflags))
1732			break;
1733		wc.wr_id = qp->r_wr_id;
1734		wc.status = IB_WC_SUCCESS;
1735		if (opcode == OP(RDMA_WRITE_LAST_WITH_IMMEDIATE) ||
1736		    opcode == OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE))
1737			wc.opcode = IB_WC_RECV_RDMA_WITH_IMM;
1738		else
1739			wc.opcode = IB_WC_RECV;
1740		wc.qp = &qp->ibqp;
1741		wc.src_qp = qp->remote_qpn;
1742		wc.slid = qp->remote_ah_attr.dlid;
1743		wc.sl = qp->remote_ah_attr.sl;
1744		/* Signal completion event if the solicited bit is set. */
1745		ipath_cq_enter(to_icq(qp->ibqp.recv_cq), &wc,
1746			       (ohdr->bth[0] &
1747				cpu_to_be32(1 << 23)) != 0);
1748		break;
1749
1750	case OP(RDMA_WRITE_FIRST):
1751	case OP(RDMA_WRITE_ONLY):
1752	case OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE):
1753		if (unlikely(!(qp->qp_access_flags &
1754			       IB_ACCESS_REMOTE_WRITE)))
1755			goto nack_inv;
1756		/* consume RWQE */
1757		/* RETH comes after BTH */
1758		if (!header_in_data)
1759			reth = &ohdr->u.rc.reth;
1760		else {
1761			reth = (struct ib_reth *)data;
1762			data += sizeof(*reth);
1763		}
1764		hdrsize += sizeof(*reth);
1765		qp->r_len = be32_to_cpu(reth->length);
1766		qp->r_rcv_len = 0;
1767		if (qp->r_len != 0) {
1768			u32 rkey = be32_to_cpu(reth->rkey);
1769			u64 vaddr = be64_to_cpu(reth->vaddr);
1770			int ok;
1771
1772			/* Check rkey & NAK */
1773			ok = ipath_rkey_ok(qp, &qp->r_sge,
1774					   qp->r_len, vaddr, rkey,
1775					   IB_ACCESS_REMOTE_WRITE);
1776			if (unlikely(!ok))
1777				goto nack_acc;
1778		} else {
1779			qp->r_sge.sg_list = NULL;
1780			qp->r_sge.sge.mr = NULL;
1781			qp->r_sge.sge.vaddr = NULL;
1782			qp->r_sge.sge.length = 0;
1783			qp->r_sge.sge.sge_length = 0;
1784		}
1785		if (opcode == OP(RDMA_WRITE_FIRST))
1786			goto send_middle;
1787		else if (opcode == OP(RDMA_WRITE_ONLY))
1788			goto send_last;
1789		if (!ipath_get_rwqe(qp, 1))
1790			goto rnr_nak;
1791		goto send_last_imm;
1792
1793	case OP(RDMA_READ_REQUEST): {
1794		struct ipath_ack_entry *e;
1795		u32 len;
1796		u8 next;
1797
1798		if (unlikely(!(qp->qp_access_flags &
1799			       IB_ACCESS_REMOTE_READ)))
1800			goto nack_inv;
1801		next = qp->r_head_ack_queue + 1;
1802		if (next > IPATH_MAX_RDMA_ATOMIC)
1803			next = 0;
1804		spin_lock_irqsave(&qp->s_lock, flags);
1805		/* Double check we can process this while holding the s_lock. */
1806		if (!(ib_ipath_state_ops[qp->state] & IPATH_PROCESS_RECV_OK))
1807			goto unlock;
1808		if (unlikely(next == qp->s_tail_ack_queue)) {
1809			if (!qp->s_ack_queue[next].sent)
1810				goto nack_inv_unlck;
1811			ipath_update_ack_queue(qp, next);
1812		}
1813		e = &qp->s_ack_queue[qp->r_head_ack_queue];
1814		/* RETH comes after BTH */
1815		if (!header_in_data)
1816			reth = &ohdr->u.rc.reth;
1817		else {
1818			reth = (struct ib_reth *)data;
1819			data += sizeof(*reth);
1820		}
1821		len = be32_to_cpu(reth->length);
1822		if (len) {
1823			u32 rkey = be32_to_cpu(reth->rkey);
1824			u64 vaddr = be64_to_cpu(reth->vaddr);
1825			int ok;
1826
1827			/* Check rkey & NAK */
1828			ok = ipath_rkey_ok(qp, &e->rdma_sge, len, vaddr,
1829					   rkey, IB_ACCESS_REMOTE_READ);
1830			if (unlikely(!ok))
1831				goto nack_acc_unlck;
1832			/*
1833			 * Update the next expected PSN.  We add 1 later
1834			 * below, so only add the remainder here.
1835			 */
1836			if (len > pmtu)
1837				qp->r_psn += (len - 1) / pmtu;
1838		} else {
1839			e->rdma_sge.sg_list = NULL;
1840			e->rdma_sge.num_sge = 0;
1841			e->rdma_sge.sge.mr = NULL;
1842			e->rdma_sge.sge.vaddr = NULL;
1843			e->rdma_sge.sge.length = 0;
1844			e->rdma_sge.sge.sge_length = 0;
1845		}
1846		e->opcode = opcode;
1847		e->sent = 0;
1848		e->psn = psn;
1849		/*
1850		 * We need to increment the MSN here instead of when we
1851		 * finish sending the result since a duplicate request would
1852		 * increment it more than once.
1853		 */
1854		qp->r_msn++;
1855		qp->r_psn++;
1856		qp->r_state = opcode;
1857		qp->r_nak_state = 0;
1858		qp->r_head_ack_queue = next;
1859
1860		/* Schedule the send tasklet. */
1861		ipath_schedule_send(qp);
1862
1863		goto unlock;
1864	}
1865
1866	case OP(COMPARE_SWAP):
1867	case OP(FETCH_ADD): {
1868		struct ib_atomic_eth *ateth;
1869		struct ipath_ack_entry *e;
1870		u64 vaddr;
1871		atomic64_t *maddr;
1872		u64 sdata;
1873		u32 rkey;
1874		u8 next;
1875
1876		if (unlikely(!(qp->qp_access_flags &
1877			       IB_ACCESS_REMOTE_ATOMIC)))
1878			goto nack_inv;
1879		next = qp->r_head_ack_queue + 1;
1880		if (next > IPATH_MAX_RDMA_ATOMIC)
1881			next = 0;
1882		spin_lock_irqsave(&qp->s_lock, flags);
1883		/* Double check we can process this while holding the s_lock. */
1884		if (!(ib_ipath_state_ops[qp->state] & IPATH_PROCESS_RECV_OK))
1885			goto unlock;
1886		if (unlikely(next == qp->s_tail_ack_queue)) {
1887			if (!qp->s_ack_queue[next].sent)
1888				goto nack_inv_unlck;
1889			ipath_update_ack_queue(qp, next);
1890		}
1891		if (!header_in_data)
1892			ateth = &ohdr->u.atomic_eth;
1893		else
1894			ateth = (struct ib_atomic_eth *)data;
1895		vaddr = ((u64) be32_to_cpu(ateth->vaddr[0]) << 32) |
1896			be32_to_cpu(ateth->vaddr[1]);
1897		if (unlikely(vaddr & (sizeof(u64) - 1)))
1898			goto nack_inv_unlck;
1899		rkey = be32_to_cpu(ateth->rkey);
1900		/* Check rkey & NAK */
1901		if (unlikely(!ipath_rkey_ok(qp, &qp->r_sge,
1902					    sizeof(u64), vaddr, rkey,
1903					    IB_ACCESS_REMOTE_ATOMIC)))
1904			goto nack_acc_unlck;
1905		/* Perform atomic OP and save result. */
1906		maddr = (atomic64_t *) qp->r_sge.sge.vaddr;
1907		sdata = be64_to_cpu(ateth->swap_data);
1908		e = &qp->s_ack_queue[qp->r_head_ack_queue];
1909		e->atomic_data = (opcode == OP(FETCH_ADD)) ?
1910			(u64) atomic64_add_return(sdata, maddr) - sdata :
1911			(u64) cmpxchg((u64 *) qp->r_sge.sge.vaddr,
1912				      be64_to_cpu(ateth->compare_data),
1913				      sdata);
1914		e->opcode = opcode;
1915		e->sent = 0;
1916		e->psn = psn & IPATH_PSN_MASK;
1917		qp->r_msn++;
1918		qp->r_psn++;
1919		qp->r_state = opcode;
1920		qp->r_nak_state = 0;
1921		qp->r_head_ack_queue = next;
1922
1923		/* Schedule the send tasklet. */
1924		ipath_schedule_send(qp);
1925
1926		goto unlock;
1927	}
1928
1929	default:
1930		/* NAK unknown opcodes. */
1931		goto nack_inv;
1932	}
1933	qp->r_psn++;
1934	qp->r_state = opcode;
1935	qp->r_ack_psn = psn;
1936	qp->r_nak_state = 0;
1937	/* Send an ACK if requested or required. */
1938	if (psn & (1 << 31))
1939		goto send_ack;
1940	goto done;
1941
1942rnr_nak:
1943	qp->r_nak_state = IB_RNR_NAK | qp->r_min_rnr_timer;
1944	qp->r_ack_psn = qp->r_psn;
1945	goto send_ack;
1946
1947nack_inv_unlck:
1948	spin_unlock_irqrestore(&qp->s_lock, flags);
1949nack_inv:
1950	ipath_rc_error(qp, IB_WC_LOC_QP_OP_ERR);
1951	qp->r_nak_state = IB_NAK_INVALID_REQUEST;
1952	qp->r_ack_psn = qp->r_psn;
1953	goto send_ack;
1954
1955nack_acc_unlck:
1956	spin_unlock_irqrestore(&qp->s_lock, flags);
1957nack_acc:
1958	ipath_rc_error(qp, IB_WC_LOC_PROT_ERR);
1959	qp->r_nak_state = IB_NAK_REMOTE_ACCESS_ERROR;
1960	qp->r_ack_psn = qp->r_psn;
1961send_ack:
1962	send_rc_ack(qp);
1963	goto done;
1964
1965unlock:
1966	spin_unlock_irqrestore(&qp->s_lock, flags);
1967done:
1968	return;
1969}
1970