[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/sched.h>
35#include <linux/spinlock.h>
36
37#include "ipath_verbs.h"
38#include "ipath_kernel.h"
39
40/*
41 * Convert the AETH RNR timeout code into the number of milliseconds.
42 */
43const u32 ib_ipath_rnr_table[32] = {
44	656,			/* 0 */
45	1,			/* 1 */
46	1,			/* 2 */
47	1,			/* 3 */
48	1,			/* 4 */
49	1,			/* 5 */
50	1,			/* 6 */
51	1,			/* 7 */
52	1,			/* 8 */
53	1,			/* 9 */
54	1,			/* A */
55	1,			/* B */
56	1,			/* C */
57	1,			/* D */
58	2,			/* E */
59	2,			/* F */
60	3,			/* 10 */
61	4,			/* 11 */
62	6,			/* 12 */
63	8,			/* 13 */
64	11,			/* 14 */
65	16,			/* 15 */
66	21,			/* 16 */
67	31,			/* 17 */
68	41,			/* 18 */
69	62,			/* 19 */
70	82,			/* 1A */
71	123,			/* 1B */
72	164,			/* 1C */
73	246,			/* 1D */
74	328,			/* 1E */
75	492			/* 1F */
76};
77
78/**
79 * ipath_insert_rnr_queue - put QP on the RNR timeout list for the device
80 * @qp: the QP
81 *
82 * Called with the QP s_lock held and interrupts disabled.
83 * XXX Use a simple list for now.  We might need a priority
84 * queue if we have lots of QPs waiting for RNR timeouts
85 * but that should be rare.
86 */
87void ipath_insert_rnr_queue(struct ipath_qp *qp)
88{
89	struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
90
91	/* We already did a spin_lock_irqsave(), so just use spin_lock */
92	spin_lock(&dev->pending_lock);
93	if (list_empty(&dev->rnrwait))
94		list_add(&qp->timerwait, &dev->rnrwait);
95	else {
96		struct list_head *l = &dev->rnrwait;
97		struct ipath_qp *nqp = list_entry(l->next, struct ipath_qp,
98						  timerwait);
99
100		while (qp->s_rnr_timeout >= nqp->s_rnr_timeout) {
101			qp->s_rnr_timeout -= nqp->s_rnr_timeout;
102			l = l->next;
103			if (l->next == &dev->rnrwait) {
104				nqp = NULL;
105				break;
106			}
107			nqp = list_entry(l->next, struct ipath_qp,
108					 timerwait);
109		}
110		if (nqp)
111			nqp->s_rnr_timeout -= qp->s_rnr_timeout;
112		list_add(&qp->timerwait, l);
113	}
114	spin_unlock(&dev->pending_lock);
115}
116
117/**
118 * ipath_init_sge - Validate a RWQE and fill in the SGE state
119 * @qp: the QP
120 *
121 * Return 1 if OK.
122 */
123int ipath_init_sge(struct ipath_qp *qp, struct ipath_rwqe *wqe,
124		   u32 *lengthp, struct ipath_sge_state *ss)
125{
126	int i, j, ret;
127	struct ib_wc wc;
128
129	*lengthp = 0;
130	for (i = j = 0; i < wqe->num_sge; i++) {
131		if (wqe->sg_list[i].length == 0)
132			continue;
133		/* Check LKEY */
134		if (!ipath_lkey_ok(qp, j ? &ss->sg_list[j - 1] : &ss->sge,
135				   &wqe->sg_list[i], IB_ACCESS_LOCAL_WRITE))
136			goto bad_lkey;
137		*lengthp += wqe->sg_list[i].length;
138		j++;
139	}
140	ss->num_sge = j;
141	ret = 1;
142	goto bail;
143
144bad_lkey:
145	memset(&wc, 0, sizeof(wc));
146	wc.wr_id = wqe->wr_id;
147	wc.status = IB_WC_LOC_PROT_ERR;
148	wc.opcode = IB_WC_RECV;
149	wc.qp = &qp->ibqp;
150	/* Signal solicited completion event. */
151	ipath_cq_enter(to_icq(qp->ibqp.recv_cq), &wc, 1);
152	ret = 0;
153bail:
154	return ret;
155}
156
157/**
158 * ipath_get_rwqe - copy the next RWQE into the QP's RWQE
159 * @qp: the QP
160 * @wr_id_only: update qp->r_wr_id only, not qp->r_sge
161 *
162 * Return 0 if no RWQE is available, otherwise return 1.
163 *
164 * Can be called from interrupt level.
165 */
166int ipath_get_rwqe(struct ipath_qp *qp, int wr_id_only)
167{
168	unsigned long flags;
169	struct ipath_rq *rq;
170	struct ipath_rwq *wq;
171	struct ipath_srq *srq;
172	struct ipath_rwqe *wqe;
173	void (*handler)(struct ib_event *, void *);
174	u32 tail;
175	int ret;
176
177	if (qp->ibqp.srq) {
178		srq = to_isrq(qp->ibqp.srq);
179		handler = srq->ibsrq.event_handler;
180		rq = &srq->rq;
181	} else {
182		srq = NULL;
183		handler = NULL;
184		rq = &qp->r_rq;
185	}
186
187	spin_lock_irqsave(&rq->lock, flags);
188	if (!(ib_ipath_state_ops[qp->state] & IPATH_PROCESS_RECV_OK)) {
189		ret = 0;
190		goto unlock;
191	}
192
193	wq = rq->wq;
194	tail = wq->tail;
195	/* Validate tail before using it since it is user writable. */
196	if (tail >= rq->size)
197		tail = 0;
198	do {
199		if (unlikely(tail == wq->head)) {
200			ret = 0;
201			goto unlock;
202		}
203		/* Make sure entry is read after head index is read. */
204		smp_rmb();
205		wqe = get_rwqe_ptr(rq, tail);
206		if (++tail >= rq->size)
207			tail = 0;
208		if (wr_id_only)
209			break;
210		qp->r_sge.sg_list = qp->r_sg_list;
211	} while (!ipath_init_sge(qp, wqe, &qp->r_len, &qp->r_sge));
212	qp->r_wr_id = wqe->wr_id;
213	wq->tail = tail;
214
215	ret = 1;
216	set_bit(IPATH_R_WRID_VALID, &qp->r_aflags);
217	if (handler) {
218		u32 n;
219
220		/*
221		 * validate head pointer value and compute
222		 * the number of remaining WQEs.
223		 */
224		n = wq->head;
225		if (n >= rq->size)
226			n = 0;
227		if (n < tail)
228			n += rq->size - tail;
229		else
230			n -= tail;
231		if (n < srq->limit) {
232			struct ib_event ev;
233
234			srq->limit = 0;
235			spin_unlock_irqrestore(&rq->lock, flags);
236			ev.device = qp->ibqp.device;
237			ev.element.srq = qp->ibqp.srq;
238			ev.event = IB_EVENT_SRQ_LIMIT_REACHED;
239			handler(&ev, srq->ibsrq.srq_context);
240			goto bail;
241		}
242	}
243unlock:
244	spin_unlock_irqrestore(&rq->lock, flags);
245bail:
246	return ret;
247}
248
249/**
250 * ipath_ruc_loopback - handle UC and RC lookback requests
251 * @sqp: the sending QP
252 *
253 * This is called from ipath_do_send() to
254 * forward a WQE addressed to the same HCA.
255 * Note that although we are single threaded due to the tasklet, we still
256 * have to protect against post_send().  We don't have to worry about
257 * receive interrupts since this is a connected protocol and all packets
258 * will pass through here.
259 */
260static void ipath_ruc_loopback(struct ipath_qp *sqp)
261{
262	struct ipath_ibdev *dev = to_idev(sqp->ibqp.device);
263	struct ipath_qp *qp;
264	struct ipath_swqe *wqe;
265	struct ipath_sge *sge;
266	unsigned long flags;
267	struct ib_wc wc;
268	u64 sdata;
269	atomic64_t *maddr;
270	enum ib_wc_status send_status;
271
272	/*
273	 * Note that we check the responder QP state after
274	 * checking the requester's state.
275	 */
276	qp = ipath_lookup_qpn(&dev->qp_table, sqp->remote_qpn);
277
278	spin_lock_irqsave(&sqp->s_lock, flags);
279
280	/* Return if we are already busy processing a work request. */
281	if ((sqp->s_flags & (IPATH_S_BUSY | IPATH_S_ANY_WAIT)) ||
282	    !(ib_ipath_state_ops[sqp->state] & IPATH_PROCESS_OR_FLUSH_SEND))
283		goto unlock;
284
285	sqp->s_flags |= IPATH_S_BUSY;
286
287again:
288	if (sqp->s_last == sqp->s_head)
289		goto clr_busy;
290	wqe = get_swqe_ptr(sqp, sqp->s_last);
291
292	/* Return if it is not OK to start a new work reqeust. */
293	if (!(ib_ipath_state_ops[sqp->state] & IPATH_PROCESS_NEXT_SEND_OK)) {
294		if (!(ib_ipath_state_ops[sqp->state] & IPATH_FLUSH_SEND))
295			goto clr_busy;
296		/* We are in the error state, flush the work request. */
297		send_status = IB_WC_WR_FLUSH_ERR;
298		goto flush_send;
299	}
300
301	/*
302	 * We can rely on the entry not changing without the s_lock
303	 * being held until we update s_last.
304	 * We increment s_cur to indicate s_last is in progress.
305	 */
306	if (sqp->s_last == sqp->s_cur) {
307		if (++sqp->s_cur >= sqp->s_size)
308			sqp->s_cur = 0;
309	}
310	spin_unlock_irqrestore(&sqp->s_lock, flags);
311
312	if (!qp || !(ib_ipath_state_ops[qp->state] & IPATH_PROCESS_RECV_OK)) {
313		dev->n_pkt_drops++;
314		/*
315		 * For RC, the requester would timeout and retry so
316		 * shortcut the timeouts and just signal too many retries.
317		 */
318		if (sqp->ibqp.qp_type == IB_QPT_RC)
319			send_status = IB_WC_RETRY_EXC_ERR;
320		else
321			send_status = IB_WC_SUCCESS;
322		goto serr;
323	}
324
325	memset(&wc, 0, sizeof wc);
326	send_status = IB_WC_SUCCESS;
327
328	sqp->s_sge.sge = wqe->sg_list[0];
329	sqp->s_sge.sg_list = wqe->sg_list + 1;
330	sqp->s_sge.num_sge = wqe->wr.num_sge;
331	sqp->s_len = wqe->length;
332	switch (wqe->wr.opcode) {
333	case IB_WR_SEND_WITH_IMM:
334		wc.wc_flags = IB_WC_WITH_IMM;
335		wc.ex.imm_data = wqe->wr.ex.imm_data;
336		/* FALLTHROUGH */
337	case IB_WR_SEND:
338		if (!ipath_get_rwqe(qp, 0))
339			goto rnr_nak;
340		break;
341
342	case IB_WR_RDMA_WRITE_WITH_IMM:
343		if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_WRITE)))
344			goto inv_err;
345		wc.wc_flags = IB_WC_WITH_IMM;
346		wc.ex.imm_data = wqe->wr.ex.imm_data;
347		if (!ipath_get_rwqe(qp, 1))
348			goto rnr_nak;
349		/* FALLTHROUGH */
350	case IB_WR_RDMA_WRITE:
351		if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_WRITE)))
352			goto inv_err;
353		if (wqe->length == 0)
354			break;
355		if (unlikely(!ipath_rkey_ok(qp, &qp->r_sge, wqe->length,
356					    wqe->wr.wr.rdma.remote_addr,
357					    wqe->wr.wr.rdma.rkey,
358					    IB_ACCESS_REMOTE_WRITE)))
359			goto acc_err;
360		break;
361
362	case IB_WR_RDMA_READ:
363		if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_READ)))
364			goto inv_err;
365		if (unlikely(!ipath_rkey_ok(qp, &sqp->s_sge, wqe->length,
366					    wqe->wr.wr.rdma.remote_addr,
367					    wqe->wr.wr.rdma.rkey,
368					    IB_ACCESS_REMOTE_READ)))
369			goto acc_err;
370		qp->r_sge.sge = wqe->sg_list[0];
371		qp->r_sge.sg_list = wqe->sg_list + 1;
372		qp->r_sge.num_sge = wqe->wr.num_sge;
373		break;
374
375	case IB_WR_ATOMIC_CMP_AND_SWP:
376	case IB_WR_ATOMIC_FETCH_AND_ADD:
377		if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_ATOMIC)))
378			goto inv_err;
379		if (unlikely(!ipath_rkey_ok(qp, &qp->r_sge, sizeof(u64),
380					    wqe->wr.wr.atomic.remote_addr,
381					    wqe->wr.wr.atomic.rkey,
382					    IB_ACCESS_REMOTE_ATOMIC)))
383			goto acc_err;
384		/* Perform atomic OP and save result. */
385		maddr = (atomic64_t *) qp->r_sge.sge.vaddr;
386		sdata = wqe->wr.wr.atomic.compare_add;
387		*(u64 *) sqp->s_sge.sge.vaddr =
388			(wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD) ?
389			(u64) atomic64_add_return(sdata, maddr) - sdata :
390			(u64) cmpxchg((u64 *) qp->r_sge.sge.vaddr,
391				      sdata, wqe->wr.wr.atomic.swap);
392		goto send_comp;
393
394	default:
395		send_status = IB_WC_LOC_QP_OP_ERR;
396		goto serr;
397	}
398
399	sge = &sqp->s_sge.sge;
400	while (sqp->s_len) {
401		u32 len = sqp->s_len;
402
403		if (len > sge->length)
404			len = sge->length;
405		if (len > sge->sge_length)
406			len = sge->sge_length;
407		BUG_ON(len == 0);
408		ipath_copy_sge(&qp->r_sge, sge->vaddr, len);
409		sge->vaddr += len;
410		sge->length -= len;
411		sge->sge_length -= len;
412		if (sge->sge_length == 0) {
413			if (--sqp->s_sge.num_sge)
414				*sge = *sqp->s_sge.sg_list++;
415		} else if (sge->length == 0 && sge->mr != NULL) {
416			if (++sge->n >= IPATH_SEGSZ) {
417				if (++sge->m >= sge->mr->mapsz)
418					break;
419				sge->n = 0;
420			}
421			sge->vaddr =
422				sge->mr->map[sge->m]->segs[sge->n].vaddr;
423			sge->length =
424				sge->mr->map[sge->m]->segs[sge->n].length;
425		}
426		sqp->s_len -= len;
427	}
428
429	if (!test_and_clear_bit(IPATH_R_WRID_VALID, &qp->r_aflags))
430		goto send_comp;
431
432	if (wqe->wr.opcode == IB_WR_RDMA_WRITE_WITH_IMM)
433		wc.opcode = IB_WC_RECV_RDMA_WITH_IMM;
434	else
435		wc.opcode = IB_WC_RECV;
436	wc.wr_id = qp->r_wr_id;
437	wc.status = IB_WC_SUCCESS;
438	wc.byte_len = wqe->length;
439	wc.qp = &qp->ibqp;
440	wc.src_qp = qp->remote_qpn;
441	wc.slid = qp->remote_ah_attr.dlid;
442	wc.sl = qp->remote_ah_attr.sl;
443	wc.port_num = 1;
444	/* Signal completion event if the solicited bit is set. */
445	ipath_cq_enter(to_icq(qp->ibqp.recv_cq), &wc,
446		       wqe->wr.send_flags & IB_SEND_SOLICITED);
447
448send_comp:
449	spin_lock_irqsave(&sqp->s_lock, flags);
450flush_send:
451	sqp->s_rnr_retry = sqp->s_rnr_retry_cnt;
452	ipath_send_complete(sqp, wqe, send_status);
453	goto again;
454
455rnr_nak:
456	/* Handle RNR NAK */
457	if (qp->ibqp.qp_type == IB_QPT_UC)
458		goto send_comp;
459	/*
460	 * Note: we don't need the s_lock held since the BUSY flag
461	 * makes this single threaded.
462	 */
463	if (sqp->s_rnr_retry == 0) {
464		send_status = IB_WC_RNR_RETRY_EXC_ERR;
465		goto serr;
466	}
467	if (sqp->s_rnr_retry_cnt < 7)
468		sqp->s_rnr_retry--;
469	spin_lock_irqsave(&sqp->s_lock, flags);
470	if (!(ib_ipath_state_ops[sqp->state] & IPATH_PROCESS_RECV_OK))
471		goto clr_busy;
472	sqp->s_flags |= IPATH_S_WAITING;
473	dev->n_rnr_naks++;
474	sqp->s_rnr_timeout = ib_ipath_rnr_table[qp->r_min_rnr_timer];
475	ipath_insert_rnr_queue(sqp);
476	goto clr_busy;
477
478inv_err:
479	send_status = IB_WC_REM_INV_REQ_ERR;
480	wc.status = IB_WC_LOC_QP_OP_ERR;
481	goto err;
482
483acc_err:
484	send_status = IB_WC_REM_ACCESS_ERR;
485	wc.status = IB_WC_LOC_PROT_ERR;
486err:
487	/* responder goes to error state */
488	ipath_rc_error(qp, wc.status);
489
490serr:
491	spin_lock_irqsave(&sqp->s_lock, flags);
492	ipath_send_complete(sqp, wqe, send_status);
493	if (sqp->ibqp.qp_type == IB_QPT_RC) {
494		int lastwqe = ipath_error_qp(sqp, IB_WC_WR_FLUSH_ERR);
495
496		sqp->s_flags &= ~IPATH_S_BUSY;
497		spin_unlock_irqrestore(&sqp->s_lock, flags);
498		if (lastwqe) {
499			struct ib_event ev;
500
501			ev.device = sqp->ibqp.device;
502			ev.element.qp = &sqp->ibqp;
503			ev.event = IB_EVENT_QP_LAST_WQE_REACHED;
504			sqp->ibqp.event_handler(&ev, sqp->ibqp.qp_context);
505		}
506		goto done;
507	}
508clr_busy:
509	sqp->s_flags &= ~IPATH_S_BUSY;
510unlock:
511	spin_unlock_irqrestore(&sqp->s_lock, flags);
512done:
513	if (qp && atomic_dec_and_test(&qp->refcount))
514		wake_up(&qp->wait);
515}
516
517static void want_buffer(struct ipath_devdata *dd, struct ipath_qp *qp)
518{
519	if (!(dd->ipath_flags & IPATH_HAS_SEND_DMA) ||
520	    qp->ibqp.qp_type == IB_QPT_SMI) {
521		unsigned long flags;
522
523		spin_lock_irqsave(&dd->ipath_sendctrl_lock, flags);
524		dd->ipath_sendctrl |= INFINIPATH_S_PIOINTBUFAVAIL;
525		ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl,
526				 dd->ipath_sendctrl);
527		ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
528		spin_unlock_irqrestore(&dd->ipath_sendctrl_lock, flags);
529	}
530}
531
532/**
533 * ipath_no_bufs_available - tell the layer driver we need buffers
534 * @qp: the QP that caused the problem
535 * @dev: the device we ran out of buffers on
536 *
537 * Called when we run out of PIO buffers.
538 * If we are now in the error state, return zero to flush the
539 * send work request.
540 */
541static int ipath_no_bufs_available(struct ipath_qp *qp,
542				    struct ipath_ibdev *dev)
543{
544	unsigned long flags;
545	int ret = 1;
546
547	/*
548	 * Note that as soon as want_buffer() is called and
549	 * possibly before it returns, ipath_ib_piobufavail()
550	 * could be called. Therefore, put QP on the piowait list before
551	 * enabling the PIO avail interrupt.
552	 */
553	spin_lock_irqsave(&qp->s_lock, flags);
554	if (ib_ipath_state_ops[qp->state] & IPATH_PROCESS_SEND_OK) {
555		dev->n_piowait++;
556		qp->s_flags |= IPATH_S_WAITING;
557		qp->s_flags &= ~IPATH_S_BUSY;
558		spin_lock(&dev->pending_lock);
559		if (list_empty(&qp->piowait))
560			list_add_tail(&qp->piowait, &dev->piowait);
561		spin_unlock(&dev->pending_lock);
562	} else
563		ret = 0;
564	spin_unlock_irqrestore(&qp->s_lock, flags);
565	if (ret)
566		want_buffer(dev->dd, qp);
567	return ret;
568}
569
570/**
571 * ipath_make_grh - construct a GRH header
572 * @dev: a pointer to the ipath device
573 * @hdr: a pointer to the GRH header being constructed
574 * @grh: the global route address to send to
575 * @hwords: the number of 32 bit words of header being sent
576 * @nwords: the number of 32 bit words of data being sent
577 *
578 * Return the size of the header in 32 bit words.
579 */
580u32 ipath_make_grh(struct ipath_ibdev *dev, struct ib_grh *hdr,
581		   struct ib_global_route *grh, u32 hwords, u32 nwords)
582{
583	hdr->version_tclass_flow =
584		cpu_to_be32((6 << 28) |
585			    (grh->traffic_class << 20) |
586			    grh->flow_label);
587	hdr->paylen = cpu_to_be16((hwords - 2 + nwords + SIZE_OF_CRC) << 2);
588	/* next_hdr is defined by C8-7 in ch. 8.4.1 */
589	hdr->next_hdr = 0x1B;
590	hdr->hop_limit = grh->hop_limit;
591	/* The SGID is 32-bit aligned. */
592	hdr->sgid.global.subnet_prefix = dev->gid_prefix;
593	hdr->sgid.global.interface_id = dev->dd->ipath_guid;
594	hdr->dgid = grh->dgid;
595
596	/* GRH header size in 32-bit words. */
597	return sizeof(struct ib_grh) / sizeof(u32);
598}
599
600void ipath_make_ruc_header(struct ipath_ibdev *dev, struct ipath_qp *qp,
601			   struct ipath_other_headers *ohdr,
602			   u32 bth0, u32 bth2)
603{
604	u16 lrh0;
605	u32 nwords;
606	u32 extra_bytes;
607
608	/* Construct the header. */
609	extra_bytes = -qp->s_cur_size & 3;
610	nwords = (qp->s_cur_size + extra_bytes) >> 2;
611	lrh0 = IPATH_LRH_BTH;
612	if (unlikely(qp->remote_ah_attr.ah_flags & IB_AH_GRH)) {
613		qp->s_hdrwords += ipath_make_grh(dev, &qp->s_hdr.u.l.grh,
614						 &qp->remote_ah_attr.grh,
615						 qp->s_hdrwords, nwords);
616		lrh0 = IPATH_LRH_GRH;
617	}
618	lrh0 |= qp->remote_ah_attr.sl << 4;
619	qp->s_hdr.lrh[0] = cpu_to_be16(lrh0);
620	qp->s_hdr.lrh[1] = cpu_to_be16(qp->remote_ah_attr.dlid);
621	qp->s_hdr.lrh[2] = cpu_to_be16(qp->s_hdrwords + nwords + SIZE_OF_CRC);
622	qp->s_hdr.lrh[3] = cpu_to_be16(dev->dd->ipath_lid |
623				       qp->remote_ah_attr.src_path_bits);
624	bth0 |= ipath_get_pkey(dev->dd, qp->s_pkey_index);
625	bth0 |= extra_bytes << 20;
626	ohdr->bth[0] = cpu_to_be32(bth0 | (1 << 22));
627	ohdr->bth[1] = cpu_to_be32(qp->remote_qpn);
628	ohdr->bth[2] = cpu_to_be32(bth2);
629}
630
631/**
632 * ipath_do_send - perform a send on a QP
633 * @data: contains a pointer to the QP
634 *
635 * Process entries in the send work queue until credit or queue is
636 * exhausted.  Only allow one CPU to send a packet per QP (tasklet).
637 * Otherwise, two threads could send packets out of order.
638 */
639void ipath_do_send(unsigned long data)
640{
641	struct ipath_qp *qp = (struct ipath_qp *)data;
642	struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
643	int (*make_req)(struct ipath_qp *qp);
644	unsigned long flags;
645
646	if ((qp->ibqp.qp_type == IB_QPT_RC ||
647	     qp->ibqp.qp_type == IB_QPT_UC) &&
648	    qp->remote_ah_attr.dlid == dev->dd->ipath_lid) {
649		ipath_ruc_loopback(qp);
650		goto bail;
651	}
652
653	if (qp->ibqp.qp_type == IB_QPT_RC)
654	       make_req = ipath_make_rc_req;
655	else if (qp->ibqp.qp_type == IB_QPT_UC)
656	       make_req = ipath_make_uc_req;
657	else
658	       make_req = ipath_make_ud_req;
659
660	spin_lock_irqsave(&qp->s_lock, flags);
661
662	/* Return if we are already busy processing a work request. */
663	if ((qp->s_flags & (IPATH_S_BUSY | IPATH_S_ANY_WAIT)) ||
664	    !(ib_ipath_state_ops[qp->state] & IPATH_PROCESS_OR_FLUSH_SEND)) {
665		spin_unlock_irqrestore(&qp->s_lock, flags);
666		goto bail;
667	}
668
669	qp->s_flags |= IPATH_S_BUSY;
670
671	spin_unlock_irqrestore(&qp->s_lock, flags);
672
673again:
674	/* Check for a constructed packet to be sent. */
675	if (qp->s_hdrwords != 0) {
676		/*
677		 * If no PIO bufs are available, return.  An interrupt will
678		 * call ipath_ib_piobufavail() when one is available.
679		 */
680		if (ipath_verbs_send(qp, &qp->s_hdr, qp->s_hdrwords,
681				     qp->s_cur_sge, qp->s_cur_size)) {
682			if (ipath_no_bufs_available(qp, dev))
683				goto bail;
684		}
685		dev->n_unicast_xmit++;
686		/* Record that we sent the packet and s_hdr is empty. */
687		qp->s_hdrwords = 0;
688	}
689
690	if (make_req(qp))
691		goto again;
692
693bail:;
694}
695
696/*
697 * This should be called with s_lock held.
698 */
699void ipath_send_complete(struct ipath_qp *qp, struct ipath_swqe *wqe,
700			 enum ib_wc_status status)
701{
702	u32 old_last, last;
703
704	if (!(ib_ipath_state_ops[qp->state] & IPATH_PROCESS_OR_FLUSH_SEND))
705		return;
706
707	/* See ch. 11.2.4.1 and 10.7.3.1 */
708	if (!(qp->s_flags & IPATH_S_SIGNAL_REQ_WR) ||
709	    (wqe->wr.send_flags & IB_SEND_SIGNALED) ||
710	    status != IB_WC_SUCCESS) {
711		struct ib_wc wc;
712
713		memset(&wc, 0, sizeof wc);
714		wc.wr_id = wqe->wr.wr_id;
715		wc.status = status;
716		wc.opcode = ib_ipath_wc_opcode[wqe->wr.opcode];
717		wc.qp = &qp->ibqp;
718		if (status == IB_WC_SUCCESS)
719			wc.byte_len = wqe->length;
720		ipath_cq_enter(to_icq(qp->ibqp.send_cq), &wc,
721			       status != IB_WC_SUCCESS);
722	}
723
724	old_last = last = qp->s_last;
725	if (++last >= qp->s_size)
726		last = 0;
727	qp->s_last = last;
728	if (qp->s_cur == old_last)
729		qp->s_cur = last;
730	if (qp->s_tail == old_last)
731		qp->s_tail = last;
732	if (qp->state == IB_QPS_SQD && last == qp->s_cur)
733		qp->s_draining = 0;
734}
735