IB/ipath: Performance improvements via mmap of queues
[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_kernel.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 static int init_sge(struct ipath_qp *qp, struct ipath_rwqe *wqe)
110 {
111         struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
112         int user = to_ipd(qp->ibqp.pd)->user;
113         int i, j, ret;
114         struct ib_wc wc;
115
116         qp->r_len = 0;
117         for (i = j = 0; i < wqe->num_sge; i++) {
118                 if (wqe->sg_list[i].length == 0)
119                         continue;
120                 /* Check LKEY */
121                 if ((user && wqe->sg_list[i].lkey == 0) ||
122                     !ipath_lkey_ok(&dev->lk_table,
123                                    &qp->r_sg_list[j], &wqe->sg_list[i],
124                                    IB_ACCESS_LOCAL_WRITE))
125                         goto bad_lkey;
126                 qp->r_len += wqe->sg_list[i].length;
127                 j++;
128         }
129         qp->r_sge.sge = qp->r_sg_list[0];
130         qp->r_sge.sg_list = qp->r_sg_list + 1;
131         qp->r_sge.num_sge = j;
132         ret = 1;
133         goto bail;
134
135 bad_lkey:
136         wc.wr_id = wqe->wr_id;
137         wc.status = IB_WC_LOC_PROT_ERR;
138         wc.opcode = IB_WC_RECV;
139         wc.vendor_err = 0;
140         wc.byte_len = 0;
141         wc.imm_data = 0;
142         wc.qp_num = qp->ibqp.qp_num;
143         wc.src_qp = 0;
144         wc.wc_flags = 0;
145         wc.pkey_index = 0;
146         wc.slid = 0;
147         wc.sl = 0;
148         wc.dlid_path_bits = 0;
149         wc.port_num = 0;
150         /* Signal solicited completion event. */
151         ipath_cq_enter(to_icq(qp->ibqp.recv_cq), &wc, 1);
152         ret = 0;
153 bail:
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 wr_id only, not SGEs
161  *
162  * Return 0 if no RWQE is available, otherwise return 1.
163  *
164  * Can be called from interrupt level.
165  */
166 int 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         wq = rq->wq;
189         tail = wq->tail;
190         /* Validate tail before using it since it is user writable. */
191         if (tail >= rq->size)
192                 tail = 0;
193         do {
194                 if (unlikely(tail == wq->head)) {
195                         spin_unlock_irqrestore(&rq->lock, flags);
196                         ret = 0;
197                         goto bail;
198                 }
199                 wqe = get_rwqe_ptr(rq, tail);
200                 if (++tail >= rq->size)
201                         tail = 0;
202         } while (!wr_id_only && !init_sge(qp, wqe));
203         qp->r_wr_id = wqe->wr_id;
204         wq->tail = tail;
205
206         ret = 1;
207         if (handler) {
208                 u32 n;
209
210                 /*
211                  * validate head pointer value and compute
212                  * the number of remaining WQEs.
213                  */
214                 n = wq->head;
215                 if (n >= rq->size)
216                         n = 0;
217                 if (n < tail)
218                         n += rq->size - tail;
219                 else
220                         n -= tail;
221                 if (n < srq->limit) {
222                         struct ib_event ev;
223
224                         srq->limit = 0;
225                         spin_unlock_irqrestore(&rq->lock, flags);
226                         ev.device = qp->ibqp.device;
227                         ev.element.srq = qp->ibqp.srq;
228                         ev.event = IB_EVENT_SRQ_LIMIT_REACHED;
229                         handler(&ev, srq->ibsrq.srq_context);
230                         goto bail;
231                 }
232         }
233         spin_unlock_irqrestore(&rq->lock, flags);
234
235 bail:
236         return ret;
237 }
238
239 /**
240  * ipath_ruc_loopback - handle UC and RC lookback requests
241  * @sqp: the loopback QP
242  *
243  * This is called from ipath_do_uc_send() or ipath_do_rc_send() to
244  * forward a WQE addressed to the same HCA.
245  * Note that although we are single threaded due to the tasklet, we still
246  * have to protect against post_send().  We don't have to worry about
247  * receive interrupts since this is a connected protocol and all packets
248  * will pass through here.
249  */
250 static void ipath_ruc_loopback(struct ipath_qp *sqp)
251 {
252         struct ipath_ibdev *dev = to_idev(sqp->ibqp.device);
253         struct ipath_qp *qp;
254         struct ipath_swqe *wqe;
255         struct ipath_sge *sge;
256         unsigned long flags;
257         struct ib_wc wc;
258         u64 sdata;
259
260         qp = ipath_lookup_qpn(&dev->qp_table, sqp->remote_qpn);
261         if (!qp) {
262                 dev->n_pkt_drops++;
263                 return;
264         }
265
266 again:
267         spin_lock_irqsave(&sqp->s_lock, flags);
268
269         if (!(ib_ipath_state_ops[sqp->state] & IPATH_PROCESS_SEND_OK)) {
270                 spin_unlock_irqrestore(&sqp->s_lock, flags);
271                 goto done;
272         }
273
274         /* Get the next send request. */
275         if (sqp->s_last == sqp->s_head) {
276                 /* Send work queue is empty. */
277                 spin_unlock_irqrestore(&sqp->s_lock, flags);
278                 goto done;
279         }
280
281         /*
282          * We can rely on the entry not changing without the s_lock
283          * being held until we update s_last.
284          */
285         wqe = get_swqe_ptr(sqp, sqp->s_last);
286         spin_unlock_irqrestore(&sqp->s_lock, flags);
287
288         wc.wc_flags = 0;
289         wc.imm_data = 0;
290
291         sqp->s_sge.sge = wqe->sg_list[0];
292         sqp->s_sge.sg_list = wqe->sg_list + 1;
293         sqp->s_sge.num_sge = wqe->wr.num_sge;
294         sqp->s_len = wqe->length;
295         switch (wqe->wr.opcode) {
296         case IB_WR_SEND_WITH_IMM:
297                 wc.wc_flags = IB_WC_WITH_IMM;
298                 wc.imm_data = wqe->wr.imm_data;
299                 /* FALLTHROUGH */
300         case IB_WR_SEND:
301                 if (!ipath_get_rwqe(qp, 0)) {
302                 rnr_nak:
303                         /* Handle RNR NAK */
304                         if (qp->ibqp.qp_type == IB_QPT_UC)
305                                 goto send_comp;
306                         if (sqp->s_rnr_retry == 0) {
307                                 wc.status = IB_WC_RNR_RETRY_EXC_ERR;
308                                 goto err;
309                         }
310                         if (sqp->s_rnr_retry_cnt < 7)
311                                 sqp->s_rnr_retry--;
312                         dev->n_rnr_naks++;
313                         sqp->s_rnr_timeout =
314                                 ib_ipath_rnr_table[sqp->r_min_rnr_timer];
315                         ipath_insert_rnr_queue(sqp);
316                         goto done;
317                 }
318                 break;
319
320         case IB_WR_RDMA_WRITE_WITH_IMM:
321                 wc.wc_flags = IB_WC_WITH_IMM;
322                 wc.imm_data = wqe->wr.imm_data;
323                 if (!ipath_get_rwqe(qp, 1))
324                         goto rnr_nak;
325                 /* FALLTHROUGH */
326         case IB_WR_RDMA_WRITE:
327                 if (wqe->length == 0)
328                         break;
329                 if (unlikely(!ipath_rkey_ok(dev, &qp->r_sge, wqe->length,
330                                             wqe->wr.wr.rdma.remote_addr,
331                                             wqe->wr.wr.rdma.rkey,
332                                             IB_ACCESS_REMOTE_WRITE))) {
333                 acc_err:
334                         wc.status = IB_WC_REM_ACCESS_ERR;
335                 err:
336                         wc.wr_id = wqe->wr.wr_id;
337                         wc.opcode = ib_ipath_wc_opcode[wqe->wr.opcode];
338                         wc.vendor_err = 0;
339                         wc.byte_len = 0;
340                         wc.qp_num = sqp->ibqp.qp_num;
341                         wc.src_qp = sqp->remote_qpn;
342                         wc.pkey_index = 0;
343                         wc.slid = sqp->remote_ah_attr.dlid;
344                         wc.sl = sqp->remote_ah_attr.sl;
345                         wc.dlid_path_bits = 0;
346                         wc.port_num = 0;
347                         ipath_sqerror_qp(sqp, &wc);
348                         goto done;
349                 }
350                 break;
351
352         case IB_WR_RDMA_READ:
353                 if (unlikely(!ipath_rkey_ok(dev, &sqp->s_sge, wqe->length,
354                                             wqe->wr.wr.rdma.remote_addr,
355                                             wqe->wr.wr.rdma.rkey,
356                                             IB_ACCESS_REMOTE_READ)))
357                         goto acc_err;
358                 if (unlikely(!(qp->qp_access_flags &
359                                IB_ACCESS_REMOTE_READ)))
360                         goto acc_err;
361                 qp->r_sge.sge = wqe->sg_list[0];
362                 qp->r_sge.sg_list = wqe->sg_list + 1;
363                 qp->r_sge.num_sge = wqe->wr.num_sge;
364                 break;
365
366         case IB_WR_ATOMIC_CMP_AND_SWP:
367         case IB_WR_ATOMIC_FETCH_AND_ADD:
368                 if (unlikely(!ipath_rkey_ok(dev, &qp->r_sge, sizeof(u64),
369                                             wqe->wr.wr.rdma.remote_addr,
370                                             wqe->wr.wr.rdma.rkey,
371                                             IB_ACCESS_REMOTE_ATOMIC)))
372                         goto acc_err;
373                 /* Perform atomic OP and save result. */
374                 sdata = wqe->wr.wr.atomic.swap;
375                 spin_lock_irqsave(&dev->pending_lock, flags);
376                 qp->r_atomic_data = *(u64 *) qp->r_sge.sge.vaddr;
377                 if (wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD)
378                         *(u64 *) qp->r_sge.sge.vaddr =
379                                 qp->r_atomic_data + sdata;
380                 else if (qp->r_atomic_data == wqe->wr.wr.atomic.compare_add)
381                         *(u64 *) qp->r_sge.sge.vaddr = sdata;
382                 spin_unlock_irqrestore(&dev->pending_lock, flags);
383                 *(u64 *) sqp->s_sge.sge.vaddr = qp->r_atomic_data;
384                 goto send_comp;
385
386         default:
387                 goto done;
388         }
389
390         sge = &sqp->s_sge.sge;
391         while (sqp->s_len) {
392                 u32 len = sqp->s_len;
393
394                 if (len > sge->length)
395                         len = sge->length;
396                 BUG_ON(len == 0);
397                 ipath_copy_sge(&qp->r_sge, sge->vaddr, len);
398                 sge->vaddr += len;
399                 sge->length -= len;
400                 sge->sge_length -= len;
401                 if (sge->sge_length == 0) {
402                         if (--sqp->s_sge.num_sge)
403                                 *sge = *sqp->s_sge.sg_list++;
404                 } else if (sge->length == 0 && sge->mr != NULL) {
405                         if (++sge->n >= IPATH_SEGSZ) {
406                                 if (++sge->m >= sge->mr->mapsz)
407                                         break;
408                                 sge->n = 0;
409                         }
410                         sge->vaddr =
411                                 sge->mr->map[sge->m]->segs[sge->n].vaddr;
412                         sge->length =
413                                 sge->mr->map[sge->m]->segs[sge->n].length;
414                 }
415                 sqp->s_len -= len;
416         }
417
418         if (wqe->wr.opcode == IB_WR_RDMA_WRITE ||
419             wqe->wr.opcode == IB_WR_RDMA_READ)
420                 goto send_comp;
421
422         if (wqe->wr.opcode == IB_WR_RDMA_WRITE_WITH_IMM)
423                 wc.opcode = IB_WC_RECV_RDMA_WITH_IMM;
424         else
425                 wc.opcode = IB_WC_RECV;
426         wc.wr_id = qp->r_wr_id;
427         wc.status = IB_WC_SUCCESS;
428         wc.vendor_err = 0;
429         wc.byte_len = wqe->length;
430         wc.qp_num = qp->ibqp.qp_num;
431         wc.src_qp = qp->remote_qpn;
432         /* XXX do we know which pkey matched? Only needed for GSI. */
433         wc.pkey_index = 0;
434         wc.slid = qp->remote_ah_attr.dlid;
435         wc.sl = qp->remote_ah_attr.sl;
436         wc.dlid_path_bits = 0;
437         /* Signal completion event if the solicited bit is set. */
438         ipath_cq_enter(to_icq(qp->ibqp.recv_cq), &wc,
439                        wqe->wr.send_flags & IB_SEND_SOLICITED);
440
441 send_comp:
442         sqp->s_rnr_retry = sqp->s_rnr_retry_cnt;
443
444         if (!test_bit(IPATH_S_SIGNAL_REQ_WR, &sqp->s_flags) ||
445             (wqe->wr.send_flags & IB_SEND_SIGNALED)) {
446                 wc.wr_id = wqe->wr.wr_id;
447                 wc.status = IB_WC_SUCCESS;
448                 wc.opcode = ib_ipath_wc_opcode[wqe->wr.opcode];
449                 wc.vendor_err = 0;
450                 wc.byte_len = wqe->length;
451                 wc.qp_num = sqp->ibqp.qp_num;
452                 wc.src_qp = 0;
453                 wc.pkey_index = 0;
454                 wc.slid = 0;
455                 wc.sl = 0;
456                 wc.dlid_path_bits = 0;
457                 wc.port_num = 0;
458                 ipath_cq_enter(to_icq(sqp->ibqp.send_cq), &wc, 0);
459         }
460
461         /* Update s_last now that we are finished with the SWQE */
462         spin_lock_irqsave(&sqp->s_lock, flags);
463         if (++sqp->s_last >= sqp->s_size)
464                 sqp->s_last = 0;
465         spin_unlock_irqrestore(&sqp->s_lock, flags);
466         goto again;
467
468 done:
469         if (atomic_dec_and_test(&qp->refcount))
470                 wake_up(&qp->wait);
471 }
472
473 /**
474  * ipath_no_bufs_available - tell the layer driver we need buffers
475  * @qp: the QP that caused the problem
476  * @dev: the device we ran out of buffers on
477  *
478  * Called when we run out of PIO buffers.
479  */
480 void ipath_no_bufs_available(struct ipath_qp *qp, struct ipath_ibdev *dev)
481 {
482         unsigned long flags;
483
484         spin_lock_irqsave(&dev->pending_lock, flags);
485         if (list_empty(&qp->piowait))
486                 list_add_tail(&qp->piowait, &dev->piowait);
487         spin_unlock_irqrestore(&dev->pending_lock, flags);
488         /*
489          * Note that as soon as ipath_layer_want_buffer() is called and
490          * possibly before it returns, ipath_ib_piobufavail()
491          * could be called.  If we are still in the tasklet function,
492          * tasklet_hi_schedule() will not call us until the next time
493          * tasklet_hi_schedule() is called.
494          * We clear the tasklet flag now since we are committing to return
495          * from the tasklet function.
496          */
497         clear_bit(IPATH_S_BUSY, &qp->s_flags);
498         tasklet_unlock(&qp->s_task);
499         ipath_layer_want_buffer(dev->dd);
500         dev->n_piowait++;
501 }
502
503 /**
504  * ipath_post_ruc_send - post RC and UC sends
505  * @qp: the QP to post on
506  * @wr: the work request to send
507  */
508 int ipath_post_ruc_send(struct ipath_qp *qp, struct ib_send_wr *wr)
509 {
510         struct ipath_swqe *wqe;
511         unsigned long flags;
512         u32 next;
513         int i, j;
514         int acc;
515         int ret;
516
517         /*
518          * Don't allow RDMA reads or atomic operations on UC or
519          * undefined operations.
520          * Make sure buffer is large enough to hold the result for atomics.
521          */
522         if (qp->ibqp.qp_type == IB_QPT_UC) {
523                 if ((unsigned) wr->opcode >= IB_WR_RDMA_READ) {
524                         ret = -EINVAL;
525                         goto bail;
526                 }
527         } else if ((unsigned) wr->opcode > IB_WR_ATOMIC_FETCH_AND_ADD) {
528                 ret = -EINVAL;
529                 goto bail;
530         } else if (wr->opcode >= IB_WR_ATOMIC_CMP_AND_SWP &&
531                    (wr->num_sge == 0 ||
532                     wr->sg_list[0].length < sizeof(u64) ||
533                     wr->sg_list[0].addr & (sizeof(u64) - 1))) {
534                 ret = -EINVAL;
535                 goto bail;
536         }
537         /* IB spec says that num_sge == 0 is OK. */
538         if (wr->num_sge > qp->s_max_sge) {
539                 ret = -ENOMEM;
540                 goto bail;
541         }
542         spin_lock_irqsave(&qp->s_lock, flags);
543         next = qp->s_head + 1;
544         if (next >= qp->s_size)
545                 next = 0;
546         if (next == qp->s_last) {
547                 spin_unlock_irqrestore(&qp->s_lock, flags);
548                 ret = -EINVAL;
549                 goto bail;
550         }
551
552         wqe = get_swqe_ptr(qp, qp->s_head);
553         wqe->wr = *wr;
554         wqe->ssn = qp->s_ssn++;
555         wqe->sg_list[0].mr = NULL;
556         wqe->sg_list[0].vaddr = NULL;
557         wqe->sg_list[0].length = 0;
558         wqe->sg_list[0].sge_length = 0;
559         wqe->length = 0;
560         acc = wr->opcode >= IB_WR_RDMA_READ ? IB_ACCESS_LOCAL_WRITE : 0;
561         for (i = 0, j = 0; i < wr->num_sge; i++) {
562                 if (to_ipd(qp->ibqp.pd)->user && wr->sg_list[i].lkey == 0) {
563                         spin_unlock_irqrestore(&qp->s_lock, flags);
564                         ret = -EINVAL;
565                         goto bail;
566                 }
567                 if (wr->sg_list[i].length == 0)
568                         continue;
569                 if (!ipath_lkey_ok(&to_idev(qp->ibqp.device)->lk_table,
570                                    &wqe->sg_list[j], &wr->sg_list[i],
571                                    acc)) {
572                         spin_unlock_irqrestore(&qp->s_lock, flags);
573                         ret = -EINVAL;
574                         goto bail;
575                 }
576                 wqe->length += wr->sg_list[i].length;
577                 j++;
578         }
579         wqe->wr.num_sge = j;
580         qp->s_head = next;
581         spin_unlock_irqrestore(&qp->s_lock, flags);
582
583         ipath_do_ruc_send((unsigned long) qp);
584
585         ret = 0;
586
587 bail:
588         return ret;
589 }
590
591 /**
592  * ipath_make_grh - construct a GRH header
593  * @dev: a pointer to the ipath device
594  * @hdr: a pointer to the GRH header being constructed
595  * @grh: the global route address to send to
596  * @hwords: the number of 32 bit words of header being sent
597  * @nwords: the number of 32 bit words of data being sent
598  *
599  * Return the size of the header in 32 bit words.
600  */
601 u32 ipath_make_grh(struct ipath_ibdev *dev, struct ib_grh *hdr,
602                    struct ib_global_route *grh, u32 hwords, u32 nwords)
603 {
604         hdr->version_tclass_flow =
605                 cpu_to_be32((6 << 28) |
606                             (grh->traffic_class << 20) |
607                             grh->flow_label);
608         hdr->paylen = cpu_to_be16((hwords - 2 + nwords + SIZE_OF_CRC) << 2);
609         /* next_hdr is defined by C8-7 in ch. 8.4.1 */
610         hdr->next_hdr = 0x1B;
611         hdr->hop_limit = grh->hop_limit;
612         /* The SGID is 32-bit aligned. */
613         hdr->sgid.global.subnet_prefix = dev->gid_prefix;
614         hdr->sgid.global.interface_id = ipath_layer_get_guid(dev->dd);
615         hdr->dgid = grh->dgid;
616
617         /* GRH header size in 32-bit words. */
618         return sizeof(struct ib_grh) / sizeof(u32);
619 }
620
621 /**
622  * ipath_do_ruc_send - perform a send on an RC or UC QP
623  * @data: contains a pointer to the QP
624  *
625  * Process entries in the send work queue until credit or queue is
626  * exhausted.  Only allow one CPU to send a packet per QP (tasklet).
627  * Otherwise, after we drop the QP s_lock, two threads could send
628  * packets out of order.
629  */
630 void ipath_do_ruc_send(unsigned long data)
631 {
632         struct ipath_qp *qp = (struct ipath_qp *)data;
633         struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
634         unsigned long flags;
635         u16 lrh0;
636         u32 nwords;
637         u32 extra_bytes;
638         u32 bth0;
639         u32 bth2;
640         u32 pmtu = ib_mtu_enum_to_int(qp->path_mtu);
641         struct ipath_other_headers *ohdr;
642
643         if (test_and_set_bit(IPATH_S_BUSY, &qp->s_flags))
644                 goto bail;
645
646         if (unlikely(qp->remote_ah_attr.dlid ==
647                      ipath_layer_get_lid(dev->dd))) {
648                 ipath_ruc_loopback(qp);
649                 goto clear;
650         }
651
652         ohdr = &qp->s_hdr.u.oth;
653         if (qp->remote_ah_attr.ah_flags & IB_AH_GRH)
654                 ohdr = &qp->s_hdr.u.l.oth;
655
656 again:
657         /* Check for a constructed packet to be sent. */
658         if (qp->s_hdrwords != 0) {
659                 /*
660                  * If no PIO bufs are available, return.  An interrupt will
661                  * call ipath_ib_piobufavail() when one is available.
662                  */
663                 if (ipath_verbs_send(dev->dd, qp->s_hdrwords,
664                                      (u32 *) &qp->s_hdr, qp->s_cur_size,
665                                      qp->s_cur_sge)) {
666                         ipath_no_bufs_available(qp, dev);
667                         goto bail;
668                 }
669                 dev->n_unicast_xmit++;
670                 /* Record that we sent the packet and s_hdr is empty. */
671                 qp->s_hdrwords = 0;
672         }
673
674         /*
675          * The lock is needed to synchronize between setting
676          * qp->s_ack_state, resend timer, and post_send().
677          */
678         spin_lock_irqsave(&qp->s_lock, flags);
679
680         /* Sending responses has higher priority over sending requests. */
681         if (qp->s_ack_state != IB_OPCODE_RC_ACKNOWLEDGE &&
682             (bth0 = ipath_make_rc_ack(qp, ohdr, pmtu)) != 0)
683                 bth2 = qp->s_ack_psn++ & IPATH_PSN_MASK;
684         else if (!((qp->ibqp.qp_type == IB_QPT_RC) ?
685                    ipath_make_rc_req(qp, ohdr, pmtu, &bth0, &bth2) :
686                    ipath_make_uc_req(qp, ohdr, pmtu, &bth0, &bth2))) {
687                 /*
688                  * Clear the busy bit before unlocking to avoid races with
689                  * adding new work queue items and then failing to process
690                  * them.
691                  */
692                 clear_bit(IPATH_S_BUSY, &qp->s_flags);
693                 spin_unlock_irqrestore(&qp->s_lock, flags);
694                 goto bail;
695         }
696
697         spin_unlock_irqrestore(&qp->s_lock, flags);
698
699         /* Construct the header. */
700         extra_bytes = (4 - qp->s_cur_size) & 3;
701         nwords = (qp->s_cur_size + extra_bytes) >> 2;
702         lrh0 = IPATH_LRH_BTH;
703         if (unlikely(qp->remote_ah_attr.ah_flags & IB_AH_GRH)) {
704                 qp->s_hdrwords += ipath_make_grh(dev, &qp->s_hdr.u.l.grh,
705                                                  &qp->remote_ah_attr.grh,
706                                                  qp->s_hdrwords, nwords);
707                 lrh0 = IPATH_LRH_GRH;
708         }
709         lrh0 |= qp->remote_ah_attr.sl << 4;
710         qp->s_hdr.lrh[0] = cpu_to_be16(lrh0);
711         qp->s_hdr.lrh[1] = cpu_to_be16(qp->remote_ah_attr.dlid);
712         qp->s_hdr.lrh[2] = cpu_to_be16(qp->s_hdrwords + nwords +
713                                        SIZE_OF_CRC);
714         qp->s_hdr.lrh[3] = cpu_to_be16(ipath_layer_get_lid(dev->dd));
715         bth0 |= ipath_layer_get_pkey(dev->dd, qp->s_pkey_index);
716         bth0 |= extra_bytes << 20;
717         ohdr->bth[0] = cpu_to_be32(bth0);
718         ohdr->bth[1] = cpu_to_be32(qp->remote_qpn);
719         ohdr->bth[2] = cpu_to_be32(bth2);
720
721         /* Check for more work to do. */
722         goto again;
723
724 clear:
725         clear_bit(IPATH_S_BUSY, &qp->s_flags);
726 bail:
727         return;
728 }