Merge commit 'v2.6.28-rc9' into x86/apic
[linux-2.6] / drivers / infiniband / hw / ipath / ipath_ruc.c
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/spinlock.h>
35
36 #include "ipath_verbs.h"
37 #include "ipath_kernel.h"
38
39 /*
40  * Convert the AETH RNR timeout code into the number of milliseconds.
41  */
42 const u32 ib_ipath_rnr_table[32] = {
43         656,                    /* 0 */
44         1,                      /* 1 */
45         1,                      /* 2 */
46         1,                      /* 3 */
47         1,                      /* 4 */
48         1,                      /* 5 */
49         1,                      /* 6 */
50         1,                      /* 7 */
51         1,                      /* 8 */
52         1,                      /* 9 */
53         1,                      /* A */
54         1,                      /* B */
55         1,                      /* C */
56         1,                      /* D */
57         2,                      /* E */
58         2,                      /* F */
59         3,                      /* 10 */
60         4,                      /* 11 */
61         6,                      /* 12 */
62         8,                      /* 13 */
63         11,                     /* 14 */
64         16,                     /* 15 */
65         21,                     /* 16 */
66         31,                     /* 17 */
67         41,                     /* 18 */
68         62,                     /* 19 */
69         82,                     /* 1A */
70         123,                    /* 1B */
71         164,                    /* 1C */
72         246,                    /* 1D */
73         328,                    /* 1E */
74         492                     /* 1F */
75 };
76
77 /**
78  * ipath_insert_rnr_queue - put QP on the RNR timeout list for the device
79  * @qp: the QP
80  *
81  * Called with the QP s_lock held and interrupts disabled.
82  * XXX Use a simple list for now.  We might need a priority
83  * queue if we have lots of QPs waiting for RNR timeouts
84  * but that should be rare.
85  */
86 void ipath_insert_rnr_queue(struct ipath_qp *qp)
87 {
88         struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
89
90         /* We already did a spin_lock_irqsave(), so just use spin_lock */
91         spin_lock(&dev->pending_lock);
92         if (list_empty(&dev->rnrwait))
93                 list_add(&qp->timerwait, &dev->rnrwait);
94         else {
95                 struct list_head *l = &dev->rnrwait;
96                 struct ipath_qp *nqp = list_entry(l->next, struct ipath_qp,
97                                                   timerwait);
98
99                 while (qp->s_rnr_timeout >= nqp->s_rnr_timeout) {
100                         qp->s_rnr_timeout -= nqp->s_rnr_timeout;
101                         l = l->next;
102                         if (l->next == &dev->rnrwait) {
103                                 nqp = NULL;
104                                 break;
105                         }
106                         nqp = list_entry(l->next, struct ipath_qp,
107                                          timerwait);
108                 }
109                 if (nqp)
110                         nqp->s_rnr_timeout -= qp->s_rnr_timeout;
111                 list_add(&qp->timerwait, l);
112         }
113         spin_unlock(&dev->pending_lock);
114 }
115
116 /**
117  * ipath_init_sge - Validate a RWQE and fill in the SGE state
118  * @qp: the QP
119  *
120  * Return 1 if OK.
121  */
122 int ipath_init_sge(struct ipath_qp *qp, struct ipath_rwqe *wqe,
123                    u32 *lengthp, struct ipath_sge_state *ss)
124 {
125         int i, j, ret;
126         struct ib_wc wc;
127
128         *lengthp = 0;
129         for (i = j = 0; i < wqe->num_sge; i++) {
130                 if (wqe->sg_list[i].length == 0)
131                         continue;
132                 /* Check LKEY */
133                 if (!ipath_lkey_ok(qp, j ? &ss->sg_list[j - 1] : &ss->sge,
134                                    &wqe->sg_list[i], IB_ACCESS_LOCAL_WRITE))
135                         goto bad_lkey;
136                 *lengthp += wqe->sg_list[i].length;
137                 j++;
138         }
139         ss->num_sge = j;
140         ret = 1;
141         goto bail;
142
143 bad_lkey:
144         memset(&wc, 0, sizeof(wc));
145         wc.wr_id = wqe->wr_id;
146         wc.status = IB_WC_LOC_PROT_ERR;
147         wc.opcode = IB_WC_RECV;
148         wc.qp = &qp->ibqp;
149         /* Signal solicited completion event. */
150         ipath_cq_enter(to_icq(qp->ibqp.recv_cq), &wc, 1);
151         ret = 0;
152 bail:
153         return ret;
154 }
155
156 /**
157  * ipath_get_rwqe - copy the next RWQE into the QP's RWQE
158  * @qp: the QP
159  * @wr_id_only: update qp->r_wr_id only, not qp->r_sge
160  *
161  * Return 0 if no RWQE is available, otherwise return 1.
162  *
163  * Can be called from interrupt level.
164  */
165 int ipath_get_rwqe(struct ipath_qp *qp, int wr_id_only)
166 {
167         unsigned long flags;
168         struct ipath_rq *rq;
169         struct ipath_rwq *wq;
170         struct ipath_srq *srq;
171         struct ipath_rwqe *wqe;
172         void (*handler)(struct ib_event *, void *);
173         u32 tail;
174         int ret;
175
176         if (qp->ibqp.srq) {
177                 srq = to_isrq(qp->ibqp.srq);
178                 handler = srq->ibsrq.event_handler;
179                 rq = &srq->rq;
180         } else {
181                 srq = NULL;
182                 handler = NULL;
183                 rq = &qp->r_rq;
184         }
185
186         spin_lock_irqsave(&rq->lock, flags);
187         if (!(ib_ipath_state_ops[qp->state] & IPATH_PROCESS_RECV_OK)) {
188                 ret = 0;
189                 goto unlock;
190         }
191
192         wq = rq->wq;
193         tail = wq->tail;
194         /* Validate tail before using it since it is user writable. */
195         if (tail >= rq->size)
196                 tail = 0;
197         do {
198                 if (unlikely(tail == wq->head)) {
199                         ret = 0;
200                         goto unlock;
201                 }
202                 /* Make sure entry is read after head index is read. */
203                 smp_rmb();
204                 wqe = get_rwqe_ptr(rq, tail);
205                 if (++tail >= rq->size)
206                         tail = 0;
207                 if (wr_id_only)
208                         break;
209                 qp->r_sge.sg_list = qp->r_sg_list;
210         } while (!ipath_init_sge(qp, wqe, &qp->r_len, &qp->r_sge));
211         qp->r_wr_id = wqe->wr_id;
212         wq->tail = tail;
213
214         ret = 1;
215         set_bit(IPATH_R_WRID_VALID, &qp->r_aflags);
216         if (handler) {
217                 u32 n;
218
219                 /*
220                  * validate head pointer value and compute
221                  * the number of remaining WQEs.
222                  */
223                 n = wq->head;
224                 if (n >= rq->size)
225                         n = 0;
226                 if (n < tail)
227                         n += rq->size - tail;
228                 else
229                         n -= tail;
230                 if (n < srq->limit) {
231                         struct ib_event ev;
232
233                         srq->limit = 0;
234                         spin_unlock_irqrestore(&rq->lock, flags);
235                         ev.device = qp->ibqp.device;
236                         ev.element.srq = qp->ibqp.srq;
237                         ev.event = IB_EVENT_SRQ_LIMIT_REACHED;
238                         handler(&ev, srq->ibsrq.srq_context);
239                         goto bail;
240                 }
241         }
242 unlock:
243         spin_unlock_irqrestore(&rq->lock, flags);
244 bail:
245         return ret;
246 }
247
248 /**
249  * ipath_ruc_loopback - handle UC and RC lookback requests
250  * @sqp: the sending QP
251  *
252  * This is called from ipath_do_send() to
253  * forward a WQE addressed to the same HCA.
254  * Note that although we are single threaded due to the tasklet, we still
255  * have to protect against post_send().  We don't have to worry about
256  * receive interrupts since this is a connected protocol and all packets
257  * will pass through here.
258  */
259 static void ipath_ruc_loopback(struct ipath_qp *sqp)
260 {
261         struct ipath_ibdev *dev = to_idev(sqp->ibqp.device);
262         struct ipath_qp *qp;
263         struct ipath_swqe *wqe;
264         struct ipath_sge *sge;
265         unsigned long flags;
266         struct ib_wc wc;
267         u64 sdata;
268         atomic64_t *maddr;
269         enum ib_wc_status send_status;
270
271         /*
272          * Note that we check the responder QP state after
273          * checking the requester's state.
274          */
275         qp = ipath_lookup_qpn(&dev->qp_table, sqp->remote_qpn);
276
277         spin_lock_irqsave(&sqp->s_lock, flags);
278
279         /* Return if we are already busy processing a work request. */
280         if ((sqp->s_flags & (IPATH_S_BUSY | IPATH_S_ANY_WAIT)) ||
281             !(ib_ipath_state_ops[sqp->state] & IPATH_PROCESS_OR_FLUSH_SEND))
282                 goto unlock;
283
284         sqp->s_flags |= IPATH_S_BUSY;
285
286 again:
287         if (sqp->s_last == sqp->s_head)
288                 goto clr_busy;
289         wqe = get_swqe_ptr(sqp, sqp->s_last);
290
291         /* Return if it is not OK to start a new work reqeust. */
292         if (!(ib_ipath_state_ops[sqp->state] & IPATH_PROCESS_NEXT_SEND_OK)) {
293                 if (!(ib_ipath_state_ops[sqp->state] & IPATH_FLUSH_SEND))
294                         goto clr_busy;
295                 /* We are in the error state, flush the work request. */
296                 send_status = IB_WC_WR_FLUSH_ERR;
297                 goto flush_send;
298         }
299
300         /*
301          * We can rely on the entry not changing without the s_lock
302          * being held until we update s_last.
303          * We increment s_cur to indicate s_last is in progress.
304          */
305         if (sqp->s_last == sqp->s_cur) {
306                 if (++sqp->s_cur >= sqp->s_size)
307                         sqp->s_cur = 0;
308         }
309         spin_unlock_irqrestore(&sqp->s_lock, flags);
310
311         if (!qp || !(ib_ipath_state_ops[qp->state] & IPATH_PROCESS_RECV_OK)) {
312                 dev->n_pkt_drops++;
313                 /*
314                  * For RC, the requester would timeout and retry so
315                  * shortcut the timeouts and just signal too many retries.
316                  */
317                 if (sqp->ibqp.qp_type == IB_QPT_RC)
318                         send_status = IB_WC_RETRY_EXC_ERR;
319                 else
320                         send_status = IB_WC_SUCCESS;
321                 goto serr;
322         }
323
324         memset(&wc, 0, sizeof wc);
325         send_status = IB_WC_SUCCESS;
326
327         sqp->s_sge.sge = wqe->sg_list[0];
328         sqp->s_sge.sg_list = wqe->sg_list + 1;
329         sqp->s_sge.num_sge = wqe->wr.num_sge;
330         sqp->s_len = wqe->length;
331         switch (wqe->wr.opcode) {
332         case IB_WR_SEND_WITH_IMM:
333                 wc.wc_flags = IB_WC_WITH_IMM;
334                 wc.ex.imm_data = wqe->wr.ex.imm_data;
335                 /* FALLTHROUGH */
336         case IB_WR_SEND:
337                 if (!ipath_get_rwqe(qp, 0))
338                         goto rnr_nak;
339                 break;
340
341         case IB_WR_RDMA_WRITE_WITH_IMM:
342                 if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_WRITE)))
343                         goto inv_err;
344                 wc.wc_flags = IB_WC_WITH_IMM;
345                 wc.ex.imm_data = wqe->wr.ex.imm_data;
346                 if (!ipath_get_rwqe(qp, 1))
347                         goto rnr_nak;
348                 /* FALLTHROUGH */
349         case IB_WR_RDMA_WRITE:
350                 if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_WRITE)))
351                         goto inv_err;
352                 if (wqe->length == 0)
353                         break;
354                 if (unlikely(!ipath_rkey_ok(qp, &qp->r_sge, wqe->length,
355                                             wqe->wr.wr.rdma.remote_addr,
356                                             wqe->wr.wr.rdma.rkey,
357                                             IB_ACCESS_REMOTE_WRITE)))
358                         goto acc_err;
359                 break;
360
361         case IB_WR_RDMA_READ:
362                 if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_READ)))
363                         goto inv_err;
364                 if (unlikely(!ipath_rkey_ok(qp, &sqp->s_sge, wqe->length,
365                                             wqe->wr.wr.rdma.remote_addr,
366                                             wqe->wr.wr.rdma.rkey,
367                                             IB_ACCESS_REMOTE_READ)))
368                         goto acc_err;
369                 qp->r_sge.sge = wqe->sg_list[0];
370                 qp->r_sge.sg_list = wqe->sg_list + 1;
371                 qp->r_sge.num_sge = wqe->wr.num_sge;
372                 break;
373
374         case IB_WR_ATOMIC_CMP_AND_SWP:
375         case IB_WR_ATOMIC_FETCH_AND_ADD:
376                 if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_ATOMIC)))
377                         goto inv_err;
378                 if (unlikely(!ipath_rkey_ok(qp, &qp->r_sge, sizeof(u64),
379                                             wqe->wr.wr.atomic.remote_addr,
380                                             wqe->wr.wr.atomic.rkey,
381                                             IB_ACCESS_REMOTE_ATOMIC)))
382                         goto acc_err;
383                 /* Perform atomic OP and save result. */
384                 maddr = (atomic64_t *) qp->r_sge.sge.vaddr;
385                 sdata = wqe->wr.wr.atomic.compare_add;
386                 *(u64 *) sqp->s_sge.sge.vaddr =
387                         (wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD) ?
388                         (u64) atomic64_add_return(sdata, maddr) - sdata :
389                         (u64) cmpxchg((u64 *) qp->r_sge.sge.vaddr,
390                                       sdata, wqe->wr.wr.atomic.swap);
391                 goto send_comp;
392
393         default:
394                 send_status = IB_WC_LOC_QP_OP_ERR;
395                 goto serr;
396         }
397
398         sge = &sqp->s_sge.sge;
399         while (sqp->s_len) {
400                 u32 len = sqp->s_len;
401
402                 if (len > sge->length)
403                         len = sge->length;
404                 if (len > sge->sge_length)
405                         len = sge->sge_length;
406                 BUG_ON(len == 0);
407                 ipath_copy_sge(&qp->r_sge, sge->vaddr, len);
408                 sge->vaddr += len;
409                 sge->length -= len;
410                 sge->sge_length -= len;
411                 if (sge->sge_length == 0) {
412                         if (--sqp->s_sge.num_sge)
413                                 *sge = *sqp->s_sge.sg_list++;
414                 } else if (sge->length == 0 && sge->mr != NULL) {
415                         if (++sge->n >= IPATH_SEGSZ) {
416                                 if (++sge->m >= sge->mr->mapsz)
417                                         break;
418                                 sge->n = 0;
419                         }
420                         sge->vaddr =
421                                 sge->mr->map[sge->m]->segs[sge->n].vaddr;
422                         sge->length =
423                                 sge->mr->map[sge->m]->segs[sge->n].length;
424                 }
425                 sqp->s_len -= len;
426         }
427
428         if (!test_and_clear_bit(IPATH_R_WRID_VALID, &qp->r_aflags))
429                 goto send_comp;
430
431         if (wqe->wr.opcode == IB_WR_RDMA_WRITE_WITH_IMM)
432                 wc.opcode = IB_WC_RECV_RDMA_WITH_IMM;
433         else
434                 wc.opcode = IB_WC_RECV;
435         wc.wr_id = qp->r_wr_id;
436         wc.status = IB_WC_SUCCESS;
437         wc.byte_len = wqe->length;
438         wc.qp = &qp->ibqp;
439         wc.src_qp = qp->remote_qpn;
440         wc.slid = qp->remote_ah_attr.dlid;
441         wc.sl = qp->remote_ah_attr.sl;
442         wc.port_num = 1;
443         /* Signal completion event if the solicited bit is set. */
444         ipath_cq_enter(to_icq(qp->ibqp.recv_cq), &wc,
445                        wqe->wr.send_flags & IB_SEND_SOLICITED);
446
447 send_comp:
448         spin_lock_irqsave(&sqp->s_lock, flags);
449 flush_send:
450         sqp->s_rnr_retry = sqp->s_rnr_retry_cnt;
451         ipath_send_complete(sqp, wqe, send_status);
452         goto again;
453
454 rnr_nak:
455         /* Handle RNR NAK */
456         if (qp->ibqp.qp_type == IB_QPT_UC)
457                 goto send_comp;
458         /*
459          * Note: we don't need the s_lock held since the BUSY flag
460          * makes this single threaded.
461          */
462         if (sqp->s_rnr_retry == 0) {
463                 send_status = IB_WC_RNR_RETRY_EXC_ERR;
464                 goto serr;
465         }
466         if (sqp->s_rnr_retry_cnt < 7)
467                 sqp->s_rnr_retry--;
468         spin_lock_irqsave(&sqp->s_lock, flags);
469         if (!(ib_ipath_state_ops[sqp->state] & IPATH_PROCESS_RECV_OK))
470                 goto clr_busy;
471         sqp->s_flags |= IPATH_S_WAITING;
472         dev->n_rnr_naks++;
473         sqp->s_rnr_timeout = ib_ipath_rnr_table[qp->r_min_rnr_timer];
474         ipath_insert_rnr_queue(sqp);
475         goto clr_busy;
476
477 inv_err:
478         send_status = IB_WC_REM_INV_REQ_ERR;
479         wc.status = IB_WC_LOC_QP_OP_ERR;
480         goto err;
481
482 acc_err:
483         send_status = IB_WC_REM_ACCESS_ERR;
484         wc.status = IB_WC_LOC_PROT_ERR;
485 err:
486         /* responder goes to error state */
487         ipath_rc_error(qp, wc.status);
488
489 serr:
490         spin_lock_irqsave(&sqp->s_lock, flags);
491         ipath_send_complete(sqp, wqe, send_status);
492         if (sqp->ibqp.qp_type == IB_QPT_RC) {
493                 int lastwqe = ipath_error_qp(sqp, IB_WC_WR_FLUSH_ERR);
494
495                 sqp->s_flags &= ~IPATH_S_BUSY;
496                 spin_unlock_irqrestore(&sqp->s_lock, flags);
497                 if (lastwqe) {
498                         struct ib_event ev;
499
500                         ev.device = sqp->ibqp.device;
501                         ev.element.qp = &sqp->ibqp;
502                         ev.event = IB_EVENT_QP_LAST_WQE_REACHED;
503                         sqp->ibqp.event_handler(&ev, sqp->ibqp.qp_context);
504                 }
505                 goto done;
506         }
507 clr_busy:
508         sqp->s_flags &= ~IPATH_S_BUSY;
509 unlock:
510         spin_unlock_irqrestore(&sqp->s_lock, flags);
511 done:
512         if (qp && atomic_dec_and_test(&qp->refcount))
513                 wake_up(&qp->wait);
514 }
515
516 static void want_buffer(struct ipath_devdata *dd, struct ipath_qp *qp)
517 {
518         if (!(dd->ipath_flags & IPATH_HAS_SEND_DMA) ||
519             qp->ibqp.qp_type == IB_QPT_SMI) {
520                 unsigned long flags;
521
522                 spin_lock_irqsave(&dd->ipath_sendctrl_lock, flags);
523                 dd->ipath_sendctrl |= INFINIPATH_S_PIOINTBUFAVAIL;
524                 ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl,
525                                  dd->ipath_sendctrl);
526                 ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
527                 spin_unlock_irqrestore(&dd->ipath_sendctrl_lock, flags);
528         }
529 }
530
531 /**
532  * ipath_no_bufs_available - tell the layer driver we need buffers
533  * @qp: the QP that caused the problem
534  * @dev: the device we ran out of buffers on
535  *
536  * Called when we run out of PIO buffers.
537  * If we are now in the error state, return zero to flush the
538  * send work request.
539  */
540 static int ipath_no_bufs_available(struct ipath_qp *qp,
541                                     struct ipath_ibdev *dev)
542 {
543         unsigned long flags;
544         int ret = 1;
545
546         /*
547          * Note that as soon as want_buffer() is called and
548          * possibly before it returns, ipath_ib_piobufavail()
549          * could be called. Therefore, put QP on the piowait list before
550          * enabling the PIO avail interrupt.
551          */
552         spin_lock_irqsave(&qp->s_lock, flags);
553         if (ib_ipath_state_ops[qp->state] & IPATH_PROCESS_SEND_OK) {
554                 dev->n_piowait++;
555                 qp->s_flags |= IPATH_S_WAITING;
556                 qp->s_flags &= ~IPATH_S_BUSY;
557                 spin_lock(&dev->pending_lock);
558                 if (list_empty(&qp->piowait))
559                         list_add_tail(&qp->piowait, &dev->piowait);
560                 spin_unlock(&dev->pending_lock);
561         } else
562                 ret = 0;
563         spin_unlock_irqrestore(&qp->s_lock, flags);
564         if (ret)
565                 want_buffer(dev->dd, qp);
566         return ret;
567 }
568
569 /**
570  * ipath_make_grh - construct a GRH header
571  * @dev: a pointer to the ipath device
572  * @hdr: a pointer to the GRH header being constructed
573  * @grh: the global route address to send to
574  * @hwords: the number of 32 bit words of header being sent
575  * @nwords: the number of 32 bit words of data being sent
576  *
577  * Return the size of the header in 32 bit words.
578  */
579 u32 ipath_make_grh(struct ipath_ibdev *dev, struct ib_grh *hdr,
580                    struct ib_global_route *grh, u32 hwords, u32 nwords)
581 {
582         hdr->version_tclass_flow =
583                 cpu_to_be32((6 << 28) |
584                             (grh->traffic_class << 20) |
585                             grh->flow_label);
586         hdr->paylen = cpu_to_be16((hwords - 2 + nwords + SIZE_OF_CRC) << 2);
587         /* next_hdr is defined by C8-7 in ch. 8.4.1 */
588         hdr->next_hdr = 0x1B;
589         hdr->hop_limit = grh->hop_limit;
590         /* The SGID is 32-bit aligned. */
591         hdr->sgid.global.subnet_prefix = dev->gid_prefix;
592         hdr->sgid.global.interface_id = dev->dd->ipath_guid;
593         hdr->dgid = grh->dgid;
594
595         /* GRH header size in 32-bit words. */
596         return sizeof(struct ib_grh) / sizeof(u32);
597 }
598
599 void ipath_make_ruc_header(struct ipath_ibdev *dev, struct ipath_qp *qp,
600                            struct ipath_other_headers *ohdr,
601                            u32 bth0, u32 bth2)
602 {
603         u16 lrh0;
604         u32 nwords;
605         u32 extra_bytes;
606
607         /* Construct the header. */
608         extra_bytes = -qp->s_cur_size & 3;
609         nwords = (qp->s_cur_size + extra_bytes) >> 2;
610         lrh0 = IPATH_LRH_BTH;
611         if (unlikely(qp->remote_ah_attr.ah_flags & IB_AH_GRH)) {
612                 qp->s_hdrwords += ipath_make_grh(dev, &qp->s_hdr.u.l.grh,
613                                                  &qp->remote_ah_attr.grh,
614                                                  qp->s_hdrwords, nwords);
615                 lrh0 = IPATH_LRH_GRH;
616         }
617         lrh0 |= qp->remote_ah_attr.sl << 4;
618         qp->s_hdr.lrh[0] = cpu_to_be16(lrh0);
619         qp->s_hdr.lrh[1] = cpu_to_be16(qp->remote_ah_attr.dlid);
620         qp->s_hdr.lrh[2] = cpu_to_be16(qp->s_hdrwords + nwords + SIZE_OF_CRC);
621         qp->s_hdr.lrh[3] = cpu_to_be16(dev->dd->ipath_lid |
622                                        qp->remote_ah_attr.src_path_bits);
623         bth0 |= ipath_get_pkey(dev->dd, qp->s_pkey_index);
624         bth0 |= extra_bytes << 20;
625         ohdr->bth[0] = cpu_to_be32(bth0 | (1 << 22));
626         ohdr->bth[1] = cpu_to_be32(qp->remote_qpn);
627         ohdr->bth[2] = cpu_to_be32(bth2);
628 }
629
630 /**
631  * ipath_do_send - perform a send on a QP
632  * @data: contains a pointer to the QP
633  *
634  * Process entries in the send work queue until credit or queue is
635  * exhausted.  Only allow one CPU to send a packet per QP (tasklet).
636  * Otherwise, two threads could send packets out of order.
637  */
638 void ipath_do_send(unsigned long data)
639 {
640         struct ipath_qp *qp = (struct ipath_qp *)data;
641         struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
642         int (*make_req)(struct ipath_qp *qp);
643         unsigned long flags;
644
645         if ((qp->ibqp.qp_type == IB_QPT_RC ||
646              qp->ibqp.qp_type == IB_QPT_UC) &&
647             qp->remote_ah_attr.dlid == dev->dd->ipath_lid) {
648                 ipath_ruc_loopback(qp);
649                 goto bail;
650         }
651
652         if (qp->ibqp.qp_type == IB_QPT_RC)
653                make_req = ipath_make_rc_req;
654         else if (qp->ibqp.qp_type == IB_QPT_UC)
655                make_req = ipath_make_uc_req;
656         else
657                make_req = ipath_make_ud_req;
658
659         spin_lock_irqsave(&qp->s_lock, flags);
660
661         /* Return if we are already busy processing a work request. */
662         if ((qp->s_flags & (IPATH_S_BUSY | IPATH_S_ANY_WAIT)) ||
663             !(ib_ipath_state_ops[qp->state] & IPATH_PROCESS_OR_FLUSH_SEND)) {
664                 spin_unlock_irqrestore(&qp->s_lock, flags);
665                 goto bail;
666         }
667
668         qp->s_flags |= IPATH_S_BUSY;
669
670         spin_unlock_irqrestore(&qp->s_lock, flags);
671
672 again:
673         /* Check for a constructed packet to be sent. */
674         if (qp->s_hdrwords != 0) {
675                 /*
676                  * If no PIO bufs are available, return.  An interrupt will
677                  * call ipath_ib_piobufavail() when one is available.
678                  */
679                 if (ipath_verbs_send(qp, &qp->s_hdr, qp->s_hdrwords,
680                                      qp->s_cur_sge, qp->s_cur_size)) {
681                         if (ipath_no_bufs_available(qp, dev))
682                                 goto bail;
683                 }
684                 dev->n_unicast_xmit++;
685                 /* Record that we sent the packet and s_hdr is empty. */
686                 qp->s_hdrwords = 0;
687         }
688
689         if (make_req(qp))
690                 goto again;
691
692 bail:;
693 }
694
695 /*
696  * This should be called with s_lock held.
697  */
698 void ipath_send_complete(struct ipath_qp *qp, struct ipath_swqe *wqe,
699                          enum ib_wc_status status)
700 {
701         u32 old_last, last;
702
703         if (!(ib_ipath_state_ops[qp->state] & IPATH_PROCESS_OR_FLUSH_SEND))
704                 return;
705
706         /* See ch. 11.2.4.1 and 10.7.3.1 */
707         if (!(qp->s_flags & IPATH_S_SIGNAL_REQ_WR) ||
708             (wqe->wr.send_flags & IB_SEND_SIGNALED) ||
709             status != IB_WC_SUCCESS) {
710                 struct ib_wc wc;
711
712                 memset(&wc, 0, sizeof wc);
713                 wc.wr_id = wqe->wr.wr_id;
714                 wc.status = status;
715                 wc.opcode = ib_ipath_wc_opcode[wqe->wr.opcode];
716                 wc.qp = &qp->ibqp;
717                 if (status == IB_WC_SUCCESS)
718                         wc.byte_len = wqe->length;
719                 ipath_cq_enter(to_icq(qp->ibqp.send_cq), &wc,
720                                status != IB_WC_SUCCESS);
721         }
722
723         old_last = last = qp->s_last;
724         if (++last >= qp->s_size)
725                 last = 0;
726         qp->s_last = last;
727         if (qp->s_cur == old_last)
728                 qp->s_cur = last;
729         if (qp->s_tail == old_last)
730                 qp->s_tail = last;
731         if (qp->state == IB_QPS_SQD && last == qp->s_cur)
732                 qp->s_draining = 0;
733 }