772bc59fb85c314fc4e1084c1656039a10c1fa6f
[linux-3.10.git] / drivers / infiniband / hw / ipath / ipath_ruc.c
1 /*
2  * Copyright (c) 2006 QLogic, Inc. 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 "ipath_verbs.h"
35 #include "ipath_common.h"
36
37 /*
38  * Convert the AETH RNR timeout code into the number of milliseconds.
39  */
40 const u32 ib_ipath_rnr_table[32] = {
41         656,                    /* 0 */
42         1,                      /* 1 */
43         1,                      /* 2 */
44         1,                      /* 3 */
45         1,                      /* 4 */
46         1,                      /* 5 */
47         1,                      /* 6 */
48         1,                      /* 7 */
49         1,                      /* 8 */
50         1,                      /* 9 */
51         1,                      /* A */
52         1,                      /* B */
53         1,                      /* C */
54         1,                      /* D */
55         2,                      /* E */
56         2,                      /* F */
57         3,                      /* 10 */
58         4,                      /* 11 */
59         6,                      /* 12 */
60         8,                      /* 13 */
61         11,                     /* 14 */
62         16,                     /* 15 */
63         21,                     /* 16 */
64         31,                     /* 17 */
65         41,                     /* 18 */
66         62,                     /* 19 */
67         82,                     /* 1A */
68         123,                    /* 1B */
69         164,                    /* 1C */
70         246,                    /* 1D */
71         328,                    /* 1E */
72         492                     /* 1F */
73 };
74
75 /**
76  * ipath_insert_rnr_queue - put QP on the RNR timeout list for the device
77  * @qp: the QP
78  *
79  * XXX Use a simple list for now.  We might need a priority
80  * queue if we have lots of QPs waiting for RNR timeouts
81  * but that should be rare.
82  */
83 void ipath_insert_rnr_queue(struct ipath_qp *qp)
84 {
85         struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
86         unsigned long flags;
87
88         spin_lock_irqsave(&dev->pending_lock, flags);
89         if (list_empty(&dev->rnrwait))
90                 list_add(&qp->timerwait, &dev->rnrwait);
91         else {
92                 struct list_head *l = &dev->rnrwait;
93                 struct ipath_qp *nqp = list_entry(l->next, struct ipath_qp,
94                                                   timerwait);
95
96                 while (qp->s_rnr_timeout >= nqp->s_rnr_timeout) {
97                         qp->s_rnr_timeout -= nqp->s_rnr_timeout;
98                         l = l->next;
99                         if (l->next == &dev->rnrwait)
100                                 break;
101                         nqp = list_entry(l->next, struct ipath_qp,
102                                          timerwait);
103                 }
104                 list_add(&qp->timerwait, l);
105         }
106         spin_unlock_irqrestore(&dev->pending_lock, flags);
107 }
108
109 /**
110  * ipath_get_rwqe - copy the next RWQE into the QP's RWQE
111  * @qp: the QP
112  * @wr_id_only: update wr_id only, not SGEs
113  *
114  * Return 0 if no RWQE is available, otherwise return 1.
115  *
116  * Can be called from interrupt level.
117  */
118 int ipath_get_rwqe(struct ipath_qp *qp, int wr_id_only)
119 {
120         unsigned long flags;
121         struct ipath_rq *rq;
122         struct ipath_srq *srq;
123         struct ipath_rwqe *wqe;
124         int ret = 1;
125
126         if (!qp->ibqp.srq) {
127                 rq = &qp->r_rq;
128                 spin_lock_irqsave(&rq->lock, flags);
129
130                 if (unlikely(rq->tail == rq->head)) {
131                         ret = 0;
132                         goto done;
133                 }
134                 wqe = get_rwqe_ptr(rq, rq->tail);
135                 qp->r_wr_id = wqe->wr_id;
136                 if (!wr_id_only) {
137                         qp->r_sge.sge = wqe->sg_list[0];
138                         qp->r_sge.sg_list = wqe->sg_list + 1;
139                         qp->r_sge.num_sge = wqe->num_sge;
140                         qp->r_len = wqe->length;
141                 }
142                 if (++rq->tail >= rq->size)
143                         rq->tail = 0;
144                 goto done;
145         }
146
147         srq = to_isrq(qp->ibqp.srq);
148         rq = &srq->rq;
149         spin_lock_irqsave(&rq->lock, flags);
150
151         if (unlikely(rq->tail == rq->head)) {
152                 ret = 0;
153                 goto done;
154         }
155         wqe = get_rwqe_ptr(rq, rq->tail);
156         qp->r_wr_id = wqe->wr_id;
157         if (!wr_id_only) {
158                 qp->r_sge.sge = wqe->sg_list[0];
159                 qp->r_sge.sg_list = wqe->sg_list + 1;
160                 qp->r_sge.num_sge = wqe->num_sge;
161                 qp->r_len = wqe->length;
162         }
163         if (++rq->tail >= rq->size)
164                 rq->tail = 0;
165         if (srq->ibsrq.event_handler) {
166                 struct ib_event ev;
167                 u32 n;
168
169                 if (rq->head < rq->tail)
170                         n = rq->size + rq->head - rq->tail;
171                 else
172                         n = rq->head - rq->tail;
173                 if (n < srq->limit) {
174                         srq->limit = 0;
175                         spin_unlock_irqrestore(&rq->lock, flags);
176                         ev.device = qp->ibqp.device;
177                         ev.element.srq = qp->ibqp.srq;
178                         ev.event = IB_EVENT_SRQ_LIMIT_REACHED;
179                         srq->ibsrq.event_handler(&ev,
180                                                  srq->ibsrq.srq_context);
181                         goto bail;
182                 }
183         }
184
185 done:
186         spin_unlock_irqrestore(&rq->lock, flags);
187 bail:
188         return ret;
189 }
190
191 /**
192  * ipath_ruc_loopback - handle UC and RC lookback requests
193  * @sqp: the loopback QP
194  *
195  * This is called from ipath_do_uc_send() or ipath_do_rc_send() to
196  * forward a WQE addressed to the same HCA.
197  * Note that although we are single threaded due to the tasklet, we still
198  * have to protect against post_send().  We don't have to worry about
199  * receive interrupts since this is a connected protocol and all packets
200  * will pass through here.
201  */
202 static void ipath_ruc_loopback(struct ipath_qp *sqp)
203 {
204         struct ipath_ibdev *dev = to_idev(sqp->ibqp.device);
205         struct ipath_qp *qp;
206         struct ipath_swqe *wqe;
207         struct ipath_sge *sge;
208         unsigned long flags;
209         struct ib_wc wc;
210         u64 sdata;
211
212         qp = ipath_lookup_qpn(&dev->qp_table, sqp->remote_qpn);
213         if (!qp) {
214                 dev->n_pkt_drops++;
215                 return;
216         }
217
218 again:
219         spin_lock_irqsave(&sqp->s_lock, flags);
220
221         if (!(ib_ipath_state_ops[sqp->state] & IPATH_PROCESS_SEND_OK)) {
222                 spin_unlock_irqrestore(&sqp->s_lock, flags);
223                 goto done;
224         }
225
226         /* Get the next send request. */
227         if (sqp->s_last == sqp->s_head) {
228                 /* Send work queue is empty. */
229                 spin_unlock_irqrestore(&sqp->s_lock, flags);
230                 goto done;
231         }
232
233         /*
234          * We can rely on the entry not changing without the s_lock
235          * being held until we update s_last.
236          */
237         wqe = get_swqe_ptr(sqp, sqp->s_last);
238         spin_unlock_irqrestore(&sqp->s_lock, flags);
239
240         wc.wc_flags = 0;
241         wc.imm_data = 0;
242
243         sqp->s_sge.sge = wqe->sg_list[0];
244         sqp->s_sge.sg_list = wqe->sg_list + 1;
245         sqp->s_sge.num_sge = wqe->wr.num_sge;
246         sqp->s_len = wqe->length;
247         switch (wqe->wr.opcode) {
248         case IB_WR_SEND_WITH_IMM:
249                 wc.wc_flags = IB_WC_WITH_IMM;
250                 wc.imm_data = wqe->wr.imm_data;
251                 /* FALLTHROUGH */
252         case IB_WR_SEND:
253                 if (!ipath_get_rwqe(qp, 0)) {
254                 rnr_nak:
255                         /* Handle RNR NAK */
256                         if (qp->ibqp.qp_type == IB_QPT_UC)
257                                 goto send_comp;
258                         if (sqp->s_rnr_retry == 0) {
259                                 wc.status = IB_WC_RNR_RETRY_EXC_ERR;
260                                 goto err;
261                         }
262                         if (sqp->s_rnr_retry_cnt < 7)
263                                 sqp->s_rnr_retry--;
264                         dev->n_rnr_naks++;
265                         sqp->s_rnr_timeout =
266                                 ib_ipath_rnr_table[sqp->r_min_rnr_timer];
267                         ipath_insert_rnr_queue(sqp);
268                         goto done;
269                 }
270                 break;
271
272         case IB_WR_RDMA_WRITE_WITH_IMM:
273                 wc.wc_flags = IB_WC_WITH_IMM;
274                 wc.imm_data = wqe->wr.imm_data;
275                 if (!ipath_get_rwqe(qp, 1))
276                         goto rnr_nak;
277                 /* FALLTHROUGH */
278         case IB_WR_RDMA_WRITE:
279                 if (wqe->length == 0)
280                         break;
281                 if (unlikely(!ipath_rkey_ok(dev, &qp->r_sge, wqe->length,
282                                             wqe->wr.wr.rdma.remote_addr,
283                                             wqe->wr.wr.rdma.rkey,
284                                             IB_ACCESS_REMOTE_WRITE))) {
285                 acc_err:
286                         wc.status = IB_WC_REM_ACCESS_ERR;
287                 err:
288                         wc.wr_id = wqe->wr.wr_id;
289                         wc.opcode = ib_ipath_wc_opcode[wqe->wr.opcode];
290                         wc.vendor_err = 0;
291                         wc.byte_len = 0;
292                         wc.qp_num = sqp->ibqp.qp_num;
293                         wc.src_qp = sqp->remote_qpn;
294                         wc.pkey_index = 0;
295                         wc.slid = sqp->remote_ah_attr.dlid;
296                         wc.sl = sqp->remote_ah_attr.sl;
297                         wc.dlid_path_bits = 0;
298                         wc.port_num = 0;
299                         ipath_sqerror_qp(sqp, &wc);
300                         goto done;
301                 }
302                 break;
303
304         case IB_WR_RDMA_READ:
305                 if (unlikely(!ipath_rkey_ok(dev, &sqp->s_sge, wqe->length,
306                                             wqe->wr.wr.rdma.remote_addr,
307                                             wqe->wr.wr.rdma.rkey,
308                                             IB_ACCESS_REMOTE_READ)))
309                         goto acc_err;
310                 if (unlikely(!(qp->qp_access_flags &
311                                IB_ACCESS_REMOTE_READ)))
312                         goto acc_err;
313                 qp->r_sge.sge = wqe->sg_list[0];
314                 qp->r_sge.sg_list = wqe->sg_list + 1;
315                 qp->r_sge.num_sge = wqe->wr.num_sge;
316                 break;
317
318         case IB_WR_ATOMIC_CMP_AND_SWP:
319         case IB_WR_ATOMIC_FETCH_AND_ADD:
320                 if (unlikely(!ipath_rkey_ok(dev, &qp->r_sge, sizeof(u64),
321                                             wqe->wr.wr.rdma.remote_addr,
322                                             wqe->wr.wr.rdma.rkey,
323                                             IB_ACCESS_REMOTE_ATOMIC)))
324                         goto acc_err;
325                 /* Perform atomic OP and save result. */
326                 sdata = wqe->wr.wr.atomic.swap;
327                 spin_lock_irqsave(&dev->pending_lock, flags);
328                 qp->r_atomic_data = *(u64 *) qp->r_sge.sge.vaddr;
329                 if (wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD)
330                         *(u64 *) qp->r_sge.sge.vaddr =
331                                 qp->r_atomic_data + sdata;
332                 else if (qp->r_atomic_data == wqe->wr.wr.atomic.compare_add)
333                         *(u64 *) qp->r_sge.sge.vaddr = sdata;
334                 spin_unlock_irqrestore(&dev->pending_lock, flags);
335                 *(u64 *) sqp->s_sge.sge.vaddr = qp->r_atomic_data;
336                 goto send_comp;
337
338         default:
339                 goto done;
340         }
341
342         sge = &sqp->s_sge.sge;
343         while (sqp->s_len) {
344                 u32 len = sqp->s_len;
345
346                 if (len > sge->length)
347                         len = sge->length;
348                 BUG_ON(len == 0);
349                 ipath_copy_sge(&qp->r_sge, sge->vaddr, len);
350                 sge->vaddr += len;
351                 sge->length -= len;
352                 sge->sge_length -= len;
353                 if (sge->sge_length == 0) {
354                         if (--sqp->s_sge.num_sge)
355                                 *sge = *sqp->s_sge.sg_list++;
356                 } else if (sge->length == 0 && sge->mr != NULL) {
357                         if (++sge->n >= IPATH_SEGSZ) {
358                                 if (++sge->m >= sge->mr->mapsz)
359                                         break;
360                                 sge->n = 0;
361                         }
362                         sge->vaddr =
363                                 sge->mr->map[sge->m]->segs[sge->n].vaddr;
364                         sge->length =
365                                 sge->mr->map[sge->m]->segs[sge->n].length;
366                 }
367                 sqp->s_len -= len;
368         }
369
370         if (wqe->wr.opcode == IB_WR_RDMA_WRITE ||
371             wqe->wr.opcode == IB_WR_RDMA_READ)
372                 goto send_comp;
373
374         if (wqe->wr.opcode == IB_WR_RDMA_WRITE_WITH_IMM)
375                 wc.opcode = IB_WC_RECV_RDMA_WITH_IMM;
376         else
377                 wc.opcode = IB_WC_RECV;
378         wc.wr_id = qp->r_wr_id;
379         wc.status = IB_WC_SUCCESS;
380         wc.vendor_err = 0;
381         wc.byte_len = wqe->length;
382         wc.qp_num = qp->ibqp.qp_num;
383         wc.src_qp = qp->remote_qpn;
384         /* XXX do we know which pkey matched? Only needed for GSI. */
385         wc.pkey_index = 0;
386         wc.slid = qp->remote_ah_attr.dlid;
387         wc.sl = qp->remote_ah_attr.sl;
388         wc.dlid_path_bits = 0;
389         /* Signal completion event if the solicited bit is set. */
390         ipath_cq_enter(to_icq(qp->ibqp.recv_cq), &wc,
391                        wqe->wr.send_flags & IB_SEND_SOLICITED);
392
393 send_comp:
394         sqp->s_rnr_retry = sqp->s_rnr_retry_cnt;
395
396         if (!test_bit(IPATH_S_SIGNAL_REQ_WR, &sqp->s_flags) ||
397             (wqe->wr.send_flags & IB_SEND_SIGNALED)) {
398                 wc.wr_id = wqe->wr.wr_id;
399                 wc.status = IB_WC_SUCCESS;
400                 wc.opcode = ib_ipath_wc_opcode[wqe->wr.opcode];
401                 wc.vendor_err = 0;
402                 wc.byte_len = wqe->length;
403                 wc.qp_num = sqp->ibqp.qp_num;
404                 wc.src_qp = 0;
405                 wc.pkey_index = 0;
406                 wc.slid = 0;
407                 wc.sl = 0;
408                 wc.dlid_path_bits = 0;
409                 wc.port_num = 0;
410                 ipath_cq_enter(to_icq(sqp->ibqp.send_cq), &wc, 0);
411         }
412
413         /* Update s_last now that we are finished with the SWQE */
414         spin_lock_irqsave(&sqp->s_lock, flags);
415         if (++sqp->s_last >= sqp->s_size)
416                 sqp->s_last = 0;
417         spin_unlock_irqrestore(&sqp->s_lock, flags);
418         goto again;
419
420 done:
421         if (atomic_dec_and_test(&qp->refcount))
422                 wake_up(&qp->wait);
423 }
424
425 /**
426  * ipath_no_bufs_available - tell the layer driver we need buffers
427  * @qp: the QP that caused the problem
428  * @dev: the device we ran out of buffers on
429  *
430  * Called when we run out of PIO buffers.
431  */
432 void ipath_no_bufs_available(struct ipath_qp *qp, struct ipath_ibdev *dev)
433 {
434         unsigned long flags;
435
436         spin_lock_irqsave(&dev->pending_lock, flags);
437         if (list_empty(&qp->piowait))
438                 list_add_tail(&qp->piowait, &dev->piowait);
439         spin_unlock_irqrestore(&dev->pending_lock, flags);
440         /*
441          * Note that as soon as ipath_layer_want_buffer() is called and
442          * possibly before it returns, ipath_ib_piobufavail()
443          * could be called.  If we are still in the tasklet function,
444          * tasklet_hi_schedule() will not call us until the next time
445          * tasklet_hi_schedule() is called.
446          * We clear the tasklet flag now since we are committing to return
447          * from the tasklet function.
448          */
449         clear_bit(IPATH_S_BUSY, &qp->s_flags);
450         tasklet_unlock(&qp->s_task);
451         ipath_layer_want_buffer(dev->dd);
452         dev->n_piowait++;
453 }
454
455 /**
456  * ipath_post_ruc_send - post RC and UC sends
457  * @qp: the QP to post on
458  * @wr: the work request to send
459  */
460 int ipath_post_ruc_send(struct ipath_qp *qp, struct ib_send_wr *wr)
461 {
462         struct ipath_swqe *wqe;
463         unsigned long flags;
464         u32 next;
465         int i, j;
466         int acc;
467         int ret;
468
469         /*
470          * Don't allow RDMA reads or atomic operations on UC or
471          * undefined operations.
472          * Make sure buffer is large enough to hold the result for atomics.
473          */
474         if (qp->ibqp.qp_type == IB_QPT_UC) {
475                 if ((unsigned) wr->opcode >= IB_WR_RDMA_READ) {
476                         ret = -EINVAL;
477                         goto bail;
478                 }
479         } else if ((unsigned) wr->opcode > IB_WR_ATOMIC_FETCH_AND_ADD) {
480                 ret = -EINVAL;
481                 goto bail;
482         } else if (wr->opcode >= IB_WR_ATOMIC_CMP_AND_SWP &&
483                    (wr->num_sge == 0 ||
484                     wr->sg_list[0].length < sizeof(u64) ||
485                     wr->sg_list[0].addr & (sizeof(u64) - 1))) {
486                 ret = -EINVAL;
487                 goto bail;
488         }
489         /* IB spec says that num_sge == 0 is OK. */
490         if (wr->num_sge > qp->s_max_sge) {
491                 ret = -ENOMEM;
492                 goto bail;
493         }
494         spin_lock_irqsave(&qp->s_lock, flags);
495         next = qp->s_head + 1;
496         if (next >= qp->s_size)
497                 next = 0;
498         if (next == qp->s_last) {
499                 spin_unlock_irqrestore(&qp->s_lock, flags);
500                 ret = -EINVAL;
501                 goto bail;
502         }
503
504         wqe = get_swqe_ptr(qp, qp->s_head);
505         wqe->wr = *wr;
506         wqe->ssn = qp->s_ssn++;
507         wqe->sg_list[0].mr = NULL;
508         wqe->sg_list[0].vaddr = NULL;
509         wqe->sg_list[0].length = 0;
510         wqe->sg_list[0].sge_length = 0;
511         wqe->length = 0;
512         acc = wr->opcode >= IB_WR_RDMA_READ ? IB_ACCESS_LOCAL_WRITE : 0;
513         for (i = 0, j = 0; i < wr->num_sge; i++) {
514                 if (to_ipd(qp->ibqp.pd)->user && wr->sg_list[i].lkey == 0) {
515                         spin_unlock_irqrestore(&qp->s_lock, flags);
516                         ret = -EINVAL;
517                         goto bail;
518                 }
519                 if (wr->sg_list[i].length == 0)
520                         continue;
521                 if (!ipath_lkey_ok(&to_idev(qp->ibqp.device)->lk_table,
522                                    &wqe->sg_list[j], &wr->sg_list[i],
523                                    acc)) {
524                         spin_unlock_irqrestore(&qp->s_lock, flags);
525                         ret = -EINVAL;
526                         goto bail;
527                 }
528                 wqe->length += wr->sg_list[i].length;
529                 j++;
530         }
531         wqe->wr.num_sge = j;
532         qp->s_head = next;
533         spin_unlock_irqrestore(&qp->s_lock, flags);
534
535         ipath_do_ruc_send((unsigned long) qp);
536
537         ret = 0;
538
539 bail:
540         return ret;
541 }
542
543 /**
544  * ipath_make_grh - construct a GRH header
545  * @dev: a pointer to the ipath device
546  * @hdr: a pointer to the GRH header being constructed
547  * @grh: the global route address to send to
548  * @hwords: the number of 32 bit words of header being sent
549  * @nwords: the number of 32 bit words of data being sent
550  *
551  * Return the size of the header in 32 bit words.
552  */
553 u32 ipath_make_grh(struct ipath_ibdev *dev, struct ib_grh *hdr,
554                    struct ib_global_route *grh, u32 hwords, u32 nwords)
555 {
556         hdr->version_tclass_flow =
557                 cpu_to_be32((6 << 28) |
558                             (grh->traffic_class << 20) |
559                             grh->flow_label);
560         hdr->paylen = cpu_to_be16((hwords - 2 + nwords + SIZE_OF_CRC) << 2);
561         /* next_hdr is defined by C8-7 in ch. 8.4.1 */
562         hdr->next_hdr = 0x1B;
563         hdr->hop_limit = grh->hop_limit;
564         /* The SGID is 32-bit aligned. */
565         hdr->sgid.global.subnet_prefix = dev->gid_prefix;
566         hdr->sgid.global.interface_id = ipath_layer_get_guid(dev->dd);
567         hdr->dgid = grh->dgid;
568
569         /* GRH header size in 32-bit words. */
570         return sizeof(struct ib_grh) / sizeof(u32);
571 }
572
573 /**
574  * ipath_do_ruc_send - perform a send on an RC or UC QP
575  * @data: contains a pointer to the QP
576  *
577  * Process entries in the send work queue until credit or queue is
578  * exhausted.  Only allow one CPU to send a packet per QP (tasklet).
579  * Otherwise, after we drop the QP s_lock, two threads could send
580  * packets out of order.
581  */
582 void ipath_do_ruc_send(unsigned long data)
583 {
584         struct ipath_qp *qp = (struct ipath_qp *)data;
585         struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
586         unsigned long flags;
587         u16 lrh0;
588         u32 nwords;
589         u32 extra_bytes;
590         u32 bth0;
591         u32 bth2;
592         u32 pmtu = ib_mtu_enum_to_int(qp->path_mtu);
593         struct ipath_other_headers *ohdr;
594
595         if (test_and_set_bit(IPATH_S_BUSY, &qp->s_flags))
596                 goto bail;
597
598         if (unlikely(qp->remote_ah_attr.dlid ==
599                      ipath_layer_get_lid(dev->dd))) {
600                 ipath_ruc_loopback(qp);
601                 goto clear;
602         }
603
604         ohdr = &qp->s_hdr.u.oth;
605         if (qp->remote_ah_attr.ah_flags & IB_AH_GRH)
606                 ohdr = &qp->s_hdr.u.l.oth;
607
608 again:
609         /* Check for a constructed packet to be sent. */
610         if (qp->s_hdrwords != 0) {
611                 /*
612                  * If no PIO bufs are available, return.  An interrupt will
613                  * call ipath_ib_piobufavail() when one is available.
614                  */
615                 if (ipath_verbs_send(dev->dd, qp->s_hdrwords,
616                                      (u32 *) &qp->s_hdr, qp->s_cur_size,
617                                      qp->s_cur_sge)) {
618                         ipath_no_bufs_available(qp, dev);
619                         goto bail;
620                 }
621                 dev->n_unicast_xmit++;
622                 /* Record that we sent the packet and s_hdr is empty. */
623                 qp->s_hdrwords = 0;
624         }
625
626         /*
627          * The lock is needed to synchronize between setting
628          * qp->s_ack_state, resend timer, and post_send().
629          */
630         spin_lock_irqsave(&qp->s_lock, flags);
631
632         /* Sending responses has higher priority over sending requests. */
633         if (qp->s_ack_state != IB_OPCODE_RC_ACKNOWLEDGE &&
634             (bth0 = ipath_make_rc_ack(qp, ohdr, pmtu)) != 0)
635                 bth2 = qp->s_ack_psn++ & IPATH_PSN_MASK;
636         else if (!((qp->ibqp.qp_type == IB_QPT_RC) ?
637                    ipath_make_rc_req(qp, ohdr, pmtu, &bth0, &bth2) :
638                    ipath_make_uc_req(qp, ohdr, pmtu, &bth0, &bth2))) {
639                 /*
640                  * Clear the busy bit before unlocking to avoid races with
641                  * adding new work queue items and then failing to process
642                  * them.
643                  */
644                 clear_bit(IPATH_S_BUSY, &qp->s_flags);
645                 spin_unlock_irqrestore(&qp->s_lock, flags);
646                 goto bail;
647         }
648
649         spin_unlock_irqrestore(&qp->s_lock, flags);
650
651         /* Construct the header. */
652         extra_bytes = (4 - qp->s_cur_size) & 3;
653         nwords = (qp->s_cur_size + extra_bytes) >> 2;
654         lrh0 = IPATH_LRH_BTH;
655         if (unlikely(qp->remote_ah_attr.ah_flags & IB_AH_GRH)) {
656                 qp->s_hdrwords += ipath_make_grh(dev, &qp->s_hdr.u.l.grh,
657                                                  &qp->remote_ah_attr.grh,
658                                                  qp->s_hdrwords, nwords);
659                 lrh0 = IPATH_LRH_GRH;
660         }
661         lrh0 |= qp->remote_ah_attr.sl << 4;
662         qp->s_hdr.lrh[0] = cpu_to_be16(lrh0);
663         qp->s_hdr.lrh[1] = cpu_to_be16(qp->remote_ah_attr.dlid);
664         qp->s_hdr.lrh[2] = cpu_to_be16(qp->s_hdrwords + nwords +
665                                        SIZE_OF_CRC);
666         qp->s_hdr.lrh[3] = cpu_to_be16(ipath_layer_get_lid(dev->dd));
667         bth0 |= ipath_layer_get_pkey(dev->dd, qp->s_pkey_index);
668         bth0 |= extra_bytes << 20;
669         ohdr->bth[0] = cpu_to_be32(bth0);
670         ohdr->bth[1] = cpu_to_be32(qp->remote_qpn);
671         ohdr->bth[2] = cpu_to_be32(bth2);
672
673         /* Check for more work to do. */
674         goto again;
675
676 clear:
677         clear_bit(IPATH_S_BUSY, &qp->s_flags);
678 bail:
679         return;
680 }