]> pilppa.org Git - linux-2.6-omap-h63xx.git/blob - drivers/infiniband/hw/ipath/ipath_rc.c
IB/ipath: Don't put QP in timeout queue if waiting to send
[linux-2.6-omap-h63xx.git] / drivers / infiniband / hw / ipath / ipath_rc.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 /* cut down ridiculously long IB macro names */
38 #define OP(x) IB_OPCODE_RC_##x
39
40 static u32 restart_sge(struct ipath_sge_state *ss, struct ipath_swqe *wqe,
41                        u32 psn, u32 pmtu)
42 {
43         u32 len;
44
45         len = ((psn - wqe->psn) & IPATH_PSN_MASK) * pmtu;
46         ss->sge = wqe->sg_list[0];
47         ss->sg_list = wqe->sg_list + 1;
48         ss->num_sge = wqe->wr.num_sge;
49         ipath_skip_sge(ss, len);
50         return wqe->length - len;
51 }
52
53 /**
54  * ipath_init_restart- initialize the qp->s_sge after a restart
55  * @qp: the QP who's SGE we're restarting
56  * @wqe: the work queue to initialize the QP's SGE from
57  *
58  * The QP s_lock should be held and interrupts disabled.
59  */
60 static void ipath_init_restart(struct ipath_qp *qp, struct ipath_swqe *wqe)
61 {
62         struct ipath_ibdev *dev;
63
64         qp->s_len = restart_sge(&qp->s_sge, wqe, qp->s_psn,
65                                 ib_mtu_enum_to_int(qp->path_mtu));
66         dev = to_idev(qp->ibqp.device);
67         spin_lock(&dev->pending_lock);
68         if (list_empty(&qp->timerwait))
69                 list_add_tail(&qp->timerwait,
70                               &dev->pending[dev->pending_index]);
71         spin_unlock(&dev->pending_lock);
72 }
73
74 /**
75  * ipath_make_rc_ack - construct a response packet (ACK, NAK, or RDMA read)
76  * @qp: a pointer to the QP
77  * @ohdr: a pointer to the IB header being constructed
78  * @pmtu: the path MTU
79  *
80  * Return 1 if constructed; otherwise, return 0.
81  * Note that we are in the responder's side of the QP context.
82  * Note the QP s_lock must be held.
83  */
84 static int ipath_make_rc_ack(struct ipath_qp *qp,
85                              struct ipath_other_headers *ohdr,
86                              u32 pmtu, u32 *bth0p, u32 *bth2p)
87 {
88         struct ipath_ack_entry *e;
89         u32 hwords;
90         u32 len;
91         u32 bth0;
92         u32 bth2;
93
94         /* header size in 32-bit words LRH+BTH = (8+12)/4. */
95         hwords = 5;
96
97         switch (qp->s_ack_state) {
98         case OP(RDMA_READ_RESPONSE_LAST):
99         case OP(RDMA_READ_RESPONSE_ONLY):
100         case OP(ATOMIC_ACKNOWLEDGE):
101                 qp->s_ack_state = OP(ACKNOWLEDGE);
102                 /* FALLTHROUGH */
103         case OP(ACKNOWLEDGE):
104                 /* Check for no next entry in the queue. */
105                 if (qp->r_head_ack_queue == qp->s_tail_ack_queue) {
106                         if (qp->s_flags & IPATH_S_ACK_PENDING)
107                                 goto normal;
108                         goto bail;
109                 }
110
111                 e = &qp->s_ack_queue[qp->s_tail_ack_queue];
112                 if (e->opcode == OP(RDMA_READ_REQUEST)) {
113                         /* Copy SGE state in case we need to resend */
114                         qp->s_ack_rdma_sge = e->rdma_sge;
115                         qp->s_cur_sge = &qp->s_ack_rdma_sge;
116                         len = e->rdma_sge.sge.sge_length;
117                         if (len > pmtu) {
118                                 len = pmtu;
119                                 qp->s_ack_state = OP(RDMA_READ_RESPONSE_FIRST);
120                         } else {
121                                 qp->s_ack_state = OP(RDMA_READ_RESPONSE_ONLY);
122                                 if (++qp->s_tail_ack_queue >
123                                     IPATH_MAX_RDMA_ATOMIC)
124                                         qp->s_tail_ack_queue = 0;
125                         }
126                         ohdr->u.aeth = ipath_compute_aeth(qp);
127                         hwords++;
128                         qp->s_ack_rdma_psn = e->psn;
129                         bth2 = qp->s_ack_rdma_psn++ & IPATH_PSN_MASK;
130                 } else {
131                         /* COMPARE_SWAP or FETCH_ADD */
132                         qp->s_cur_sge = NULL;
133                         len = 0;
134                         qp->s_ack_state = OP(ATOMIC_ACKNOWLEDGE);
135                         ohdr->u.at.aeth = ipath_compute_aeth(qp);
136                         ohdr->u.at.atomic_ack_eth[0] =
137                                 cpu_to_be32(e->atomic_data >> 32);
138                         ohdr->u.at.atomic_ack_eth[1] =
139                                 cpu_to_be32(e->atomic_data);
140                         hwords += sizeof(ohdr->u.at) / sizeof(u32);
141                         bth2 = e->psn;
142                         if (++qp->s_tail_ack_queue > IPATH_MAX_RDMA_ATOMIC)
143                                 qp->s_tail_ack_queue = 0;
144                 }
145                 bth0 = qp->s_ack_state << 24;
146                 break;
147
148         case OP(RDMA_READ_RESPONSE_FIRST):
149                 qp->s_ack_state = OP(RDMA_READ_RESPONSE_MIDDLE);
150                 /* FALLTHROUGH */
151         case OP(RDMA_READ_RESPONSE_MIDDLE):
152                 len = qp->s_ack_rdma_sge.sge.sge_length;
153                 if (len > pmtu)
154                         len = pmtu;
155                 else {
156                         ohdr->u.aeth = ipath_compute_aeth(qp);
157                         hwords++;
158                         qp->s_ack_state = OP(RDMA_READ_RESPONSE_LAST);
159                         if (++qp->s_tail_ack_queue > IPATH_MAX_RDMA_ATOMIC)
160                                 qp->s_tail_ack_queue = 0;
161                 }
162                 bth0 = qp->s_ack_state << 24;
163                 bth2 = qp->s_ack_rdma_psn++ & IPATH_PSN_MASK;
164                 break;
165
166         default:
167         normal:
168                 /*
169                  * Send a regular ACK.
170                  * Set the s_ack_state so we wait until after sending
171                  * the ACK before setting s_ack_state to ACKNOWLEDGE
172                  * (see above).
173                  */
174                 qp->s_ack_state = OP(ATOMIC_ACKNOWLEDGE);
175                 qp->s_flags &= ~IPATH_S_ACK_PENDING;
176                 qp->s_cur_sge = NULL;
177                 if (qp->s_nak_state)
178                         ohdr->u.aeth =
179                                 cpu_to_be32((qp->r_msn & IPATH_MSN_MASK) |
180                                             (qp->s_nak_state <<
181                                              IPATH_AETH_CREDIT_SHIFT));
182                 else
183                         ohdr->u.aeth = ipath_compute_aeth(qp);
184                 hwords++;
185                 len = 0;
186                 bth0 = OP(ACKNOWLEDGE) << 24;
187                 bth2 = qp->s_ack_psn & IPATH_PSN_MASK;
188         }
189         qp->s_hdrwords = hwords;
190         qp->s_cur_size = len;
191         *bth0p = bth0;
192         *bth2p = bth2;
193         return 1;
194
195 bail:
196         return 0;
197 }
198
199 /**
200  * ipath_make_rc_req - construct a request packet (SEND, RDMA r/w, ATOMIC)
201  * @qp: a pointer to the QP
202  * @ohdr: a pointer to the IB header being constructed
203  * @pmtu: the path MTU
204  * @bth0p: pointer to the BTH opcode word
205  * @bth2p: pointer to the BTH PSN word
206  *
207  * Return 1 if constructed; otherwise, return 0.
208  * Note the QP s_lock must be held and interrupts disabled.
209  */
210 int ipath_make_rc_req(struct ipath_qp *qp,
211                       struct ipath_other_headers *ohdr,
212                       u32 pmtu, u32 *bth0p, u32 *bth2p)
213 {
214         struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
215         struct ipath_sge_state *ss;
216         struct ipath_swqe *wqe;
217         u32 hwords;
218         u32 len;
219         u32 bth0;
220         u32 bth2;
221         char newreq;
222
223         /* Sending responses has higher priority over sending requests. */
224         if ((qp->r_head_ack_queue != qp->s_tail_ack_queue ||
225              (qp->s_flags & IPATH_S_ACK_PENDING) ||
226              qp->s_ack_state != IB_OPCODE_RC_ACKNOWLEDGE) &&
227             ipath_make_rc_ack(qp, ohdr, pmtu, bth0p, bth2p))
228                 goto done;
229
230         if (!(ib_ipath_state_ops[qp->state] & IPATH_PROCESS_SEND_OK) ||
231             qp->s_rnr_timeout || qp->s_wait_credit)
232                 goto bail;
233
234         /* Limit the number of packets sent without an ACK. */
235         if (ipath_cmp24(qp->s_psn, qp->s_last_psn + IPATH_PSN_CREDIT) > 0) {
236                 qp->s_wait_credit = 1;
237                 dev->n_rc_stalls++;
238                 goto bail;
239         }
240
241         /* header size in 32-bit words LRH+BTH = (8+12)/4. */
242         hwords = 5;
243         bth0 = 0;
244
245         /* Send a request. */
246         wqe = get_swqe_ptr(qp, qp->s_cur);
247         switch (qp->s_state) {
248         default:
249                 /*
250                  * Resend an old request or start a new one.
251                  *
252                  * We keep track of the current SWQE so that
253                  * we don't reset the "furthest progress" state
254                  * if we need to back up.
255                  */
256                 newreq = 0;
257                 if (qp->s_cur == qp->s_tail) {
258                         /* Check if send work queue is empty. */
259                         if (qp->s_tail == qp->s_head)
260                                 goto bail;
261                         /*
262                          * If a fence is requested, wait for previous
263                          * RDMA read and atomic operations to finish.
264                          */
265                         if ((wqe->wr.send_flags & IB_SEND_FENCE) &&
266                             qp->s_num_rd_atomic) {
267                                 qp->s_flags |= IPATH_S_FENCE_PENDING;
268                                 goto bail;
269                         }
270                         wqe->psn = qp->s_next_psn;
271                         newreq = 1;
272                 }
273                 /*
274                  * Note that we have to be careful not to modify the
275                  * original work request since we may need to resend
276                  * it.
277                  */
278                 len = wqe->length;
279                 ss = &qp->s_sge;
280                 bth2 = 0;
281                 switch (wqe->wr.opcode) {
282                 case IB_WR_SEND:
283                 case IB_WR_SEND_WITH_IMM:
284                         /* If no credit, return. */
285                         if (qp->s_lsn != (u32) -1 &&
286                             ipath_cmp24(wqe->ssn, qp->s_lsn + 1) > 0)
287                                 goto bail;
288                         wqe->lpsn = wqe->psn;
289                         if (len > pmtu) {
290                                 wqe->lpsn += (len - 1) / pmtu;
291                                 qp->s_state = OP(SEND_FIRST);
292                                 len = pmtu;
293                                 break;
294                         }
295                         if (wqe->wr.opcode == IB_WR_SEND)
296                                 qp->s_state = OP(SEND_ONLY);
297                         else {
298                                 qp->s_state = OP(SEND_ONLY_WITH_IMMEDIATE);
299                                 /* Immediate data comes after the BTH */
300                                 ohdr->u.imm_data = wqe->wr.imm_data;
301                                 hwords += 1;
302                         }
303                         if (wqe->wr.send_flags & IB_SEND_SOLICITED)
304                                 bth0 |= 1 << 23;
305                         bth2 = 1 << 31; /* Request ACK. */
306                         if (++qp->s_cur == qp->s_size)
307                                 qp->s_cur = 0;
308                         break;
309
310                 case IB_WR_RDMA_WRITE:
311                         if (newreq && qp->s_lsn != (u32) -1)
312                                 qp->s_lsn++;
313                         /* FALLTHROUGH */
314                 case IB_WR_RDMA_WRITE_WITH_IMM:
315                         /* If no credit, return. */
316                         if (qp->s_lsn != (u32) -1 &&
317                             ipath_cmp24(wqe->ssn, qp->s_lsn + 1) > 0)
318                                 goto bail;
319                         ohdr->u.rc.reth.vaddr =
320                                 cpu_to_be64(wqe->wr.wr.rdma.remote_addr);
321                         ohdr->u.rc.reth.rkey =
322                                 cpu_to_be32(wqe->wr.wr.rdma.rkey);
323                         ohdr->u.rc.reth.length = cpu_to_be32(len);
324                         hwords += sizeof(struct ib_reth) / sizeof(u32);
325                         wqe->lpsn = wqe->psn;
326                         if (len > pmtu) {
327                                 wqe->lpsn += (len - 1) / pmtu;
328                                 qp->s_state = OP(RDMA_WRITE_FIRST);
329                                 len = pmtu;
330                                 break;
331                         }
332                         if (wqe->wr.opcode == IB_WR_RDMA_WRITE)
333                                 qp->s_state = OP(RDMA_WRITE_ONLY);
334                         else {
335                                 qp->s_state =
336                                         OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE);
337                                 /* Immediate data comes after RETH */
338                                 ohdr->u.rc.imm_data = wqe->wr.imm_data;
339                                 hwords += 1;
340                                 if (wqe->wr.send_flags & IB_SEND_SOLICITED)
341                                         bth0 |= 1 << 23;
342                         }
343                         bth2 = 1 << 31; /* Request ACK. */
344                         if (++qp->s_cur == qp->s_size)
345                                 qp->s_cur = 0;
346                         break;
347
348                 case IB_WR_RDMA_READ:
349                         /*
350                          * Don't allow more operations to be started
351                          * than the QP limits allow.
352                          */
353                         if (newreq) {
354                                 if (qp->s_num_rd_atomic >=
355                                     qp->s_max_rd_atomic) {
356                                         qp->s_flags |= IPATH_S_RDMAR_PENDING;
357                                         goto bail;
358                                 }
359                                 qp->s_num_rd_atomic++;
360                                 if (qp->s_lsn != (u32) -1)
361                                         qp->s_lsn++;
362                                 /*
363                                  * Adjust s_next_psn to count the
364                                  * expected number of responses.
365                                  */
366                                 if (len > pmtu)
367                                         qp->s_next_psn += (len - 1) / pmtu;
368                                 wqe->lpsn = qp->s_next_psn++;
369                         }
370                         ohdr->u.rc.reth.vaddr =
371                                 cpu_to_be64(wqe->wr.wr.rdma.remote_addr);
372                         ohdr->u.rc.reth.rkey =
373                                 cpu_to_be32(wqe->wr.wr.rdma.rkey);
374                         ohdr->u.rc.reth.length = cpu_to_be32(len);
375                         qp->s_state = OP(RDMA_READ_REQUEST);
376                         hwords += sizeof(ohdr->u.rc.reth) / sizeof(u32);
377                         ss = NULL;
378                         len = 0;
379                         if (++qp->s_cur == qp->s_size)
380                                 qp->s_cur = 0;
381                         break;
382
383                 case IB_WR_ATOMIC_CMP_AND_SWP:
384                 case IB_WR_ATOMIC_FETCH_AND_ADD:
385                         /*
386                          * Don't allow more operations to be started
387                          * than the QP limits allow.
388                          */
389                         if (newreq) {
390                                 if (qp->s_num_rd_atomic >=
391                                     qp->s_max_rd_atomic) {
392                                         qp->s_flags |= IPATH_S_RDMAR_PENDING;
393                                         goto bail;
394                                 }
395                                 qp->s_num_rd_atomic++;
396                                 if (qp->s_lsn != (u32) -1)
397                                         qp->s_lsn++;
398                                 wqe->lpsn = wqe->psn;
399                         }
400                         if (wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
401                                 qp->s_state = OP(COMPARE_SWAP);
402                                 ohdr->u.atomic_eth.swap_data = cpu_to_be64(
403                                         wqe->wr.wr.atomic.swap);
404                                 ohdr->u.atomic_eth.compare_data = cpu_to_be64(
405                                         wqe->wr.wr.atomic.compare_add);
406                         } else {
407                                 qp->s_state = OP(FETCH_ADD);
408                                 ohdr->u.atomic_eth.swap_data = cpu_to_be64(
409                                         wqe->wr.wr.atomic.compare_add);
410                                 ohdr->u.atomic_eth.compare_data = 0;
411                         }
412                         ohdr->u.atomic_eth.vaddr[0] = cpu_to_be32(
413                                 wqe->wr.wr.atomic.remote_addr >> 32);
414                         ohdr->u.atomic_eth.vaddr[1] = cpu_to_be32(
415                                 wqe->wr.wr.atomic.remote_addr);
416                         ohdr->u.atomic_eth.rkey = cpu_to_be32(
417                                 wqe->wr.wr.atomic.rkey);
418                         hwords += sizeof(struct ib_atomic_eth) / sizeof(u32);
419                         ss = NULL;
420                         len = 0;
421                         if (++qp->s_cur == qp->s_size)
422                                 qp->s_cur = 0;
423                         break;
424
425                 default:
426                         goto bail;
427                 }
428                 qp->s_sge.sge = wqe->sg_list[0];
429                 qp->s_sge.sg_list = wqe->sg_list + 1;
430                 qp->s_sge.num_sge = wqe->wr.num_sge;
431                 qp->s_len = wqe->length;
432                 if (newreq) {
433                         qp->s_tail++;
434                         if (qp->s_tail >= qp->s_size)
435                                 qp->s_tail = 0;
436                 }
437                 bth2 |= qp->s_psn & IPATH_PSN_MASK;
438                 if (wqe->wr.opcode == IB_WR_RDMA_READ)
439                         qp->s_psn = wqe->lpsn + 1;
440                 else {
441                         qp->s_psn++;
442                         if (ipath_cmp24(qp->s_psn, qp->s_next_psn) > 0)
443                                 qp->s_next_psn = qp->s_psn;
444                 }
445                 /*
446                  * Put the QP on the pending list so lost ACKs will cause
447                  * a retry.  More than one request can be pending so the
448                  * QP may already be on the dev->pending list.
449                  */
450                 spin_lock(&dev->pending_lock);
451                 if (list_empty(&qp->timerwait))
452                         list_add_tail(&qp->timerwait,
453                                       &dev->pending[dev->pending_index]);
454                 spin_unlock(&dev->pending_lock);
455                 break;
456
457         case OP(RDMA_READ_RESPONSE_FIRST):
458                 /*
459                  * This case can only happen if a send is restarted.
460                  * See ipath_restart_rc().
461                  */
462                 ipath_init_restart(qp, wqe);
463                 /* FALLTHROUGH */
464         case OP(SEND_FIRST):
465                 qp->s_state = OP(SEND_MIDDLE);
466                 /* FALLTHROUGH */
467         case OP(SEND_MIDDLE):
468                 bth2 = qp->s_psn++ & IPATH_PSN_MASK;
469                 if (ipath_cmp24(qp->s_psn, qp->s_next_psn) > 0)
470                         qp->s_next_psn = qp->s_psn;
471                 ss = &qp->s_sge;
472                 len = qp->s_len;
473                 if (len > pmtu) {
474                         len = pmtu;
475                         break;
476                 }
477                 if (wqe->wr.opcode == IB_WR_SEND)
478                         qp->s_state = OP(SEND_LAST);
479                 else {
480                         qp->s_state = OP(SEND_LAST_WITH_IMMEDIATE);
481                         /* Immediate data comes after the BTH */
482                         ohdr->u.imm_data = wqe->wr.imm_data;
483                         hwords += 1;
484                 }
485                 if (wqe->wr.send_flags & IB_SEND_SOLICITED)
486                         bth0 |= 1 << 23;
487                 bth2 |= 1 << 31;        /* Request ACK. */
488                 qp->s_cur++;
489                 if (qp->s_cur >= qp->s_size)
490                         qp->s_cur = 0;
491                 break;
492
493         case OP(RDMA_READ_RESPONSE_LAST):
494                 /*
495                  * This case can only happen if a RDMA write is restarted.
496                  * See ipath_restart_rc().
497                  */
498                 ipath_init_restart(qp, wqe);
499                 /* FALLTHROUGH */
500         case OP(RDMA_WRITE_FIRST):
501                 qp->s_state = OP(RDMA_WRITE_MIDDLE);
502                 /* FALLTHROUGH */
503         case OP(RDMA_WRITE_MIDDLE):
504                 bth2 = qp->s_psn++ & IPATH_PSN_MASK;
505                 if (ipath_cmp24(qp->s_psn, qp->s_next_psn) > 0)
506                         qp->s_next_psn = qp->s_psn;
507                 ss = &qp->s_sge;
508                 len = qp->s_len;
509                 if (len > pmtu) {
510                         len = pmtu;
511                         break;
512                 }
513                 if (wqe->wr.opcode == IB_WR_RDMA_WRITE)
514                         qp->s_state = OP(RDMA_WRITE_LAST);
515                 else {
516                         qp->s_state = OP(RDMA_WRITE_LAST_WITH_IMMEDIATE);
517                         /* Immediate data comes after the BTH */
518                         ohdr->u.imm_data = wqe->wr.imm_data;
519                         hwords += 1;
520                         if (wqe->wr.send_flags & IB_SEND_SOLICITED)
521                                 bth0 |= 1 << 23;
522                 }
523                 bth2 |= 1 << 31;        /* Request ACK. */
524                 qp->s_cur++;
525                 if (qp->s_cur >= qp->s_size)
526                         qp->s_cur = 0;
527                 break;
528
529         case OP(RDMA_READ_RESPONSE_MIDDLE):
530                 /*
531                  * This case can only happen if a RDMA read is restarted.
532                  * See ipath_restart_rc().
533                  */
534                 ipath_init_restart(qp, wqe);
535                 len = ((qp->s_psn - wqe->psn) & IPATH_PSN_MASK) * pmtu;
536                 ohdr->u.rc.reth.vaddr =
537                         cpu_to_be64(wqe->wr.wr.rdma.remote_addr + len);
538                 ohdr->u.rc.reth.rkey =
539                         cpu_to_be32(wqe->wr.wr.rdma.rkey);
540                 ohdr->u.rc.reth.length = cpu_to_be32(qp->s_len);
541                 qp->s_state = OP(RDMA_READ_REQUEST);
542                 hwords += sizeof(ohdr->u.rc.reth) / sizeof(u32);
543                 bth2 = qp->s_psn++ & IPATH_PSN_MASK;
544                 if (ipath_cmp24(qp->s_psn, qp->s_next_psn) > 0)
545                         qp->s_next_psn = qp->s_psn;
546                 ss = NULL;
547                 len = 0;
548                 qp->s_cur++;
549                 if (qp->s_cur == qp->s_size)
550                         qp->s_cur = 0;
551                 break;
552         }
553         if (ipath_cmp24(qp->s_psn, qp->s_last_psn + IPATH_PSN_CREDIT - 1) >= 0)
554                 bth2 |= 1 << 31;        /* Request ACK. */
555         qp->s_len -= len;
556         qp->s_hdrwords = hwords;
557         qp->s_cur_sge = ss;
558         qp->s_cur_size = len;
559         *bth0p = bth0 | (qp->s_state << 24);
560         *bth2p = bth2;
561 done:
562         return 1;
563
564 bail:
565         return 0;
566 }
567
568 /**
569  * send_rc_ack - Construct an ACK packet and send it
570  * @qp: a pointer to the QP
571  *
572  * This is called from ipath_rc_rcv() and only uses the receive
573  * side QP state.
574  * Note that RDMA reads and atomics are handled in the
575  * send side QP state and tasklet.
576  */
577 static void send_rc_ack(struct ipath_qp *qp)
578 {
579         struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
580         u16 lrh0;
581         u32 bth0;
582         u32 hwords;
583         struct ipath_ib_header hdr;
584         struct ipath_other_headers *ohdr;
585         unsigned long flags;
586
587         /* Don't send ACK or NAK if a RDMA read or atomic is pending. */
588         if (qp->r_head_ack_queue != qp->s_tail_ack_queue)
589                 goto queue_ack;
590
591         /* Construct the header. */
592         ohdr = &hdr.u.oth;
593         lrh0 = IPATH_LRH_BTH;
594         /* header size in 32-bit words LRH+BTH+AETH = (8+12+4)/4. */
595         hwords = 6;
596         if (unlikely(qp->remote_ah_attr.ah_flags & IB_AH_GRH)) {
597                 hwords += ipath_make_grh(dev, &hdr.u.l.grh,
598                                          &qp->remote_ah_attr.grh,
599                                          hwords, 0);
600                 ohdr = &hdr.u.l.oth;
601                 lrh0 = IPATH_LRH_GRH;
602         }
603         /* read pkey_index w/o lock (its atomic) */
604         bth0 = ipath_get_pkey(dev->dd, qp->s_pkey_index) |
605                 OP(ACKNOWLEDGE) << 24;
606         if (qp->r_nak_state)
607                 ohdr->u.aeth = cpu_to_be32((qp->r_msn & IPATH_MSN_MASK) |
608                                             (qp->r_nak_state <<
609                                              IPATH_AETH_CREDIT_SHIFT));
610         else
611                 ohdr->u.aeth = ipath_compute_aeth(qp);
612         lrh0 |= qp->remote_ah_attr.sl << 4;
613         hdr.lrh[0] = cpu_to_be16(lrh0);
614         hdr.lrh[1] = cpu_to_be16(qp->remote_ah_attr.dlid);
615         hdr.lrh[2] = cpu_to_be16(hwords + SIZE_OF_CRC);
616         hdr.lrh[3] = cpu_to_be16(dev->dd->ipath_lid);
617         ohdr->bth[0] = cpu_to_be32(bth0);
618         ohdr->bth[1] = cpu_to_be32(qp->remote_qpn);
619         ohdr->bth[2] = cpu_to_be32(qp->r_ack_psn & IPATH_PSN_MASK);
620
621         /*
622          * If we can send the ACK, clear the ACK state.
623          */
624         if (ipath_verbs_send(dev->dd, hwords, (u32 *) &hdr, 0, NULL) == 0) {
625                 dev->n_unicast_xmit++;
626                 goto done;
627         }
628
629         /*
630          * We are out of PIO buffers at the moment.
631          * Pass responsibility for sending the ACK to the
632          * send tasklet so that when a PIO buffer becomes
633          * available, the ACK is sent ahead of other outgoing
634          * packets.
635          */
636         dev->n_rc_qacks++;
637
638 queue_ack:
639         spin_lock_irqsave(&qp->s_lock, flags);
640         qp->s_flags |= IPATH_S_ACK_PENDING;
641         qp->s_nak_state = qp->r_nak_state;
642         qp->s_ack_psn = qp->r_ack_psn;
643         spin_unlock_irqrestore(&qp->s_lock, flags);
644
645         /* Call ipath_do_rc_send() in another thread. */
646         tasklet_hi_schedule(&qp->s_task);
647
648 done:
649         return;
650 }
651
652 /**
653  * reset_psn - reset the QP state to send starting from PSN
654  * @qp: the QP
655  * @psn: the packet sequence number to restart at
656  *
657  * This is called from ipath_rc_rcv() to process an incoming RC ACK
658  * for the given QP.
659  * Called at interrupt level with the QP s_lock held.
660  */
661 static void reset_psn(struct ipath_qp *qp, u32 psn)
662 {
663         u32 n = qp->s_last;
664         struct ipath_swqe *wqe = get_swqe_ptr(qp, n);
665         u32 opcode;
666
667         qp->s_cur = n;
668
669         /*
670          * If we are starting the request from the beginning,
671          * let the normal send code handle initialization.
672          */
673         if (ipath_cmp24(psn, wqe->psn) <= 0) {
674                 qp->s_state = OP(SEND_LAST);
675                 goto done;
676         }
677
678         /* Find the work request opcode corresponding to the given PSN. */
679         opcode = wqe->wr.opcode;
680         for (;;) {
681                 int diff;
682
683                 if (++n == qp->s_size)
684                         n = 0;
685                 if (n == qp->s_tail)
686                         break;
687                 wqe = get_swqe_ptr(qp, n);
688                 diff = ipath_cmp24(psn, wqe->psn);
689                 if (diff < 0)
690                         break;
691                 qp->s_cur = n;
692                 /*
693                  * If we are starting the request from the beginning,
694                  * let the normal send code handle initialization.
695                  */
696                 if (diff == 0) {
697                         qp->s_state = OP(SEND_LAST);
698                         goto done;
699                 }
700                 opcode = wqe->wr.opcode;
701         }
702
703         /*
704          * Set the state to restart in the middle of a request.
705          * Don't change the s_sge, s_cur_sge, or s_cur_size.
706          * See ipath_do_rc_send().
707          */
708         switch (opcode) {
709         case IB_WR_SEND:
710         case IB_WR_SEND_WITH_IMM:
711                 qp->s_state = OP(RDMA_READ_RESPONSE_FIRST);
712                 break;
713
714         case IB_WR_RDMA_WRITE:
715         case IB_WR_RDMA_WRITE_WITH_IMM:
716                 qp->s_state = OP(RDMA_READ_RESPONSE_LAST);
717                 break;
718
719         case IB_WR_RDMA_READ:
720                 qp->s_state = OP(RDMA_READ_RESPONSE_MIDDLE);
721                 break;
722
723         default:
724                 /*
725                  * This case shouldn't happen since its only
726                  * one PSN per req.
727                  */
728                 qp->s_state = OP(SEND_LAST);
729         }
730 done:
731         qp->s_psn = psn;
732 }
733
734 /**
735  * ipath_restart_rc - back up requester to resend the last un-ACKed request
736  * @qp: the QP to restart
737  * @psn: packet sequence number for the request
738  * @wc: the work completion request
739  *
740  * The QP s_lock should be held and interrupts disabled.
741  */
742 void ipath_restart_rc(struct ipath_qp *qp, u32 psn, struct ib_wc *wc)
743 {
744         struct ipath_swqe *wqe = get_swqe_ptr(qp, qp->s_last);
745         struct ipath_ibdev *dev;
746
747         if (qp->s_retry == 0) {
748                 wc->wr_id = wqe->wr.wr_id;
749                 wc->status = IB_WC_RETRY_EXC_ERR;
750                 wc->opcode = ib_ipath_wc_opcode[wqe->wr.opcode];
751                 wc->vendor_err = 0;
752                 wc->byte_len = 0;
753                 wc->qp = &qp->ibqp;
754                 wc->src_qp = qp->remote_qpn;
755                 wc->pkey_index = 0;
756                 wc->slid = qp->remote_ah_attr.dlid;
757                 wc->sl = qp->remote_ah_attr.sl;
758                 wc->dlid_path_bits = 0;
759                 wc->port_num = 0;
760                 ipath_sqerror_qp(qp, wc);
761                 goto bail;
762         }
763         qp->s_retry--;
764
765         /*
766          * Remove the QP from the timeout queue.
767          * Note: it may already have been removed by ipath_ib_timer().
768          */
769         dev = to_idev(qp->ibqp.device);
770         spin_lock(&dev->pending_lock);
771         if (!list_empty(&qp->timerwait))
772                 list_del_init(&qp->timerwait);
773         spin_unlock(&dev->pending_lock);
774
775         if (wqe->wr.opcode == IB_WR_RDMA_READ)
776                 dev->n_rc_resends++;
777         else
778                 dev->n_rc_resends += (qp->s_psn - psn) & IPATH_PSN_MASK;
779
780         reset_psn(qp, psn);
781         tasklet_hi_schedule(&qp->s_task);
782
783 bail:
784         return;
785 }
786
787 static inline void update_last_psn(struct ipath_qp *qp, u32 psn)
788 {
789         if (qp->s_wait_credit) {
790                 qp->s_wait_credit = 0;
791                 tasklet_hi_schedule(&qp->s_task);
792         }
793         qp->s_last_psn = psn;
794 }
795
796 /**
797  * do_rc_ack - process an incoming RC ACK
798  * @qp: the QP the ACK came in on
799  * @psn: the packet sequence number of the ACK
800  * @opcode: the opcode of the request that resulted in the ACK
801  *
802  * This is called from ipath_rc_rcv_resp() to process an incoming RC ACK
803  * for the given QP.
804  * Called at interrupt level with the QP s_lock held and interrupts disabled.
805  * Returns 1 if OK, 0 if current operation should be aborted (NAK).
806  */
807 static int do_rc_ack(struct ipath_qp *qp, u32 aeth, u32 psn, int opcode)
808 {
809         struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
810         struct ib_wc wc;
811         struct ipath_swqe *wqe;
812         int ret = 0;
813         u32 ack_psn;
814
815         /*
816          * Remove the QP from the timeout queue (or RNR timeout queue).
817          * If ipath_ib_timer() has already removed it,
818          * it's OK since we hold the QP s_lock and ipath_restart_rc()
819          * just won't find anything to restart if we ACK everything.
820          */
821         spin_lock(&dev->pending_lock);
822         if (!list_empty(&qp->timerwait))
823                 list_del_init(&qp->timerwait);
824         spin_unlock(&dev->pending_lock);
825
826         /*
827          * Note that NAKs implicitly ACK outstanding SEND and RDMA write
828          * requests and implicitly NAK RDMA read and atomic requests issued
829          * before the NAK'ed request.  The MSN won't include the NAK'ed
830          * request but will include an ACK'ed request(s).
831          */
832         ack_psn = psn;
833         if (aeth >> 29)
834                 ack_psn--;
835         wqe = get_swqe_ptr(qp, qp->s_last);
836
837         /*
838          * The MSN might be for a later WQE than the PSN indicates so
839          * only complete WQEs that the PSN finishes.
840          */
841         while (ipath_cmp24(ack_psn, wqe->lpsn) >= 0) {
842                 /*
843                  * If this request is a RDMA read or atomic, and the ACK is
844                  * for a later operation, this ACK NAKs the RDMA read or
845                  * atomic.  In other words, only a RDMA_READ_LAST or ONLY
846                  * can ACK a RDMA read and likewise for atomic ops.  Note
847                  * that the NAK case can only happen if relaxed ordering is
848                  * used and requests are sent after an RDMA read or atomic
849                  * is sent but before the response is received.
850                  */
851                 if ((wqe->wr.opcode == IB_WR_RDMA_READ &&
852                      (opcode != OP(RDMA_READ_RESPONSE_LAST) ||
853                       ipath_cmp24(ack_psn, wqe->lpsn) != 0)) ||
854                     ((wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
855                       wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD) &&
856                      (opcode != OP(ATOMIC_ACKNOWLEDGE) ||
857                       ipath_cmp24(wqe->psn, psn) != 0))) {
858                         /*
859                          * The last valid PSN seen is the previous
860                          * request's.
861                          */
862                         update_last_psn(qp, wqe->psn - 1);
863                         /* Retry this request. */
864                         ipath_restart_rc(qp, wqe->psn, &wc);
865                         /*
866                          * No need to process the ACK/NAK since we are
867                          * restarting an earlier request.
868                          */
869                         goto bail;
870                 }
871                 if (qp->s_num_rd_atomic &&
872                     (wqe->wr.opcode == IB_WR_RDMA_READ ||
873                      wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
874                      wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD)) {
875                         qp->s_num_rd_atomic--;
876                         /* Restart sending task if fence is complete */
877                         if ((qp->s_flags & IPATH_S_FENCE_PENDING) &&
878                             !qp->s_num_rd_atomic) {
879                                 qp->s_flags &= ~IPATH_S_FENCE_PENDING;
880                                 tasklet_hi_schedule(&qp->s_task);
881                         } else if (qp->s_flags & IPATH_S_RDMAR_PENDING) {
882                                 qp->s_flags &= ~IPATH_S_RDMAR_PENDING;
883                                 tasklet_hi_schedule(&qp->s_task);
884                         }
885                 }
886                 /* Post a send completion queue entry if requested. */
887                 if (!(qp->s_flags & IPATH_S_SIGNAL_REQ_WR) ||
888                     (wqe->wr.send_flags & IB_SEND_SIGNALED)) {
889                         wc.wr_id = wqe->wr.wr_id;
890                         wc.status = IB_WC_SUCCESS;
891                         wc.opcode = ib_ipath_wc_opcode[wqe->wr.opcode];
892                         wc.vendor_err = 0;
893                         wc.byte_len = wqe->length;
894                         wc.imm_data = 0;
895                         wc.qp = &qp->ibqp;
896                         wc.src_qp = qp->remote_qpn;
897                         wc.wc_flags = 0;
898                         wc.pkey_index = 0;
899                         wc.slid = qp->remote_ah_attr.dlid;
900                         wc.sl = qp->remote_ah_attr.sl;
901                         wc.dlid_path_bits = 0;
902                         wc.port_num = 0;
903                         ipath_cq_enter(to_icq(qp->ibqp.send_cq), &wc, 0);
904                 }
905                 qp->s_retry = qp->s_retry_cnt;
906                 /*
907                  * If we are completing a request which is in the process of
908                  * being resent, we can stop resending it since we know the
909                  * responder has already seen it.
910                  */
911                 if (qp->s_last == qp->s_cur) {
912                         if (++qp->s_cur >= qp->s_size)
913                                 qp->s_cur = 0;
914                         qp->s_last = qp->s_cur;
915                         if (qp->s_last == qp->s_tail)
916                                 break;
917                         wqe = get_swqe_ptr(qp, qp->s_cur);
918                         qp->s_state = OP(SEND_LAST);
919                         qp->s_psn = wqe->psn;
920                 } else {
921                         if (++qp->s_last >= qp->s_size)
922                                 qp->s_last = 0;
923                         if (qp->s_last == qp->s_tail)
924                                 break;
925                         wqe = get_swqe_ptr(qp, qp->s_last);
926                 }
927         }
928
929         switch (aeth >> 29) {
930         case 0:         /* ACK */
931                 dev->n_rc_acks++;
932                 /* If this is a partial ACK, reset the retransmit timer. */
933                 if (qp->s_last != qp->s_tail) {
934                         spin_lock(&dev->pending_lock);
935                         list_add_tail(&qp->timerwait,
936                                       &dev->pending[dev->pending_index]);
937                         spin_unlock(&dev->pending_lock);
938                         /*
939                          * If we get a partial ACK for a resent operation,
940                          * we can stop resending the earlier packets and
941                          * continue with the next packet the receiver wants.
942                          */
943                         if (ipath_cmp24(qp->s_psn, psn) <= 0) {
944                                 reset_psn(qp, psn + 1);
945                                 tasklet_hi_schedule(&qp->s_task);
946                         }
947                 } else if (ipath_cmp24(qp->s_psn, psn) <= 0) {
948                         qp->s_state = OP(SEND_LAST);
949                         qp->s_psn = psn + 1;
950                 }
951                 ipath_get_credit(qp, aeth);
952                 qp->s_rnr_retry = qp->s_rnr_retry_cnt;
953                 qp->s_retry = qp->s_retry_cnt;
954                 update_last_psn(qp, psn);
955                 ret = 1;
956                 goto bail;
957
958         case 1:         /* RNR NAK */
959                 dev->n_rnr_naks++;
960                 if (qp->s_last == qp->s_tail)
961                         goto bail;
962                 if (qp->s_rnr_retry == 0) {
963                         wc.status = IB_WC_RNR_RETRY_EXC_ERR;
964                         goto class_b;
965                 }
966                 if (qp->s_rnr_retry_cnt < 7)
967                         qp->s_rnr_retry--;
968
969                 /* The last valid PSN is the previous PSN. */
970                 update_last_psn(qp, psn - 1);
971
972                 if (wqe->wr.opcode == IB_WR_RDMA_READ)
973                         dev->n_rc_resends++;
974                 else
975                         dev->n_rc_resends +=
976                                 (qp->s_psn - psn) & IPATH_PSN_MASK;
977
978                 reset_psn(qp, psn);
979
980                 qp->s_rnr_timeout =
981                         ib_ipath_rnr_table[(aeth >> IPATH_AETH_CREDIT_SHIFT) &
982                                            IPATH_AETH_CREDIT_MASK];
983                 ipath_insert_rnr_queue(qp);
984                 goto bail;
985
986         case 3:         /* NAK */
987                 if (qp->s_last == qp->s_tail)
988                         goto bail;
989                 /* The last valid PSN is the previous PSN. */
990                 update_last_psn(qp, psn - 1);
991                 switch ((aeth >> IPATH_AETH_CREDIT_SHIFT) &
992                         IPATH_AETH_CREDIT_MASK) {
993                 case 0: /* PSN sequence error */
994                         dev->n_seq_naks++;
995                         /*
996                          * Back up to the responder's expected PSN.
997                          * Note that we might get a NAK in the middle of an
998                          * RDMA READ response which terminates the RDMA
999                          * READ.
1000                          */
1001                         ipath_restart_rc(qp, psn, &wc);
1002                         break;
1003
1004                 case 1: /* Invalid Request */
1005                         wc.status = IB_WC_REM_INV_REQ_ERR;
1006                         dev->n_other_naks++;
1007                         goto class_b;
1008
1009                 case 2: /* Remote Access Error */
1010                         wc.status = IB_WC_REM_ACCESS_ERR;
1011                         dev->n_other_naks++;
1012                         goto class_b;
1013
1014                 case 3: /* Remote Operation Error */
1015                         wc.status = IB_WC_REM_OP_ERR;
1016                         dev->n_other_naks++;
1017                 class_b:
1018                         wc.wr_id = wqe->wr.wr_id;
1019                         wc.opcode = ib_ipath_wc_opcode[wqe->wr.opcode];
1020                         wc.vendor_err = 0;
1021                         wc.byte_len = 0;
1022                         wc.qp = &qp->ibqp;
1023                         wc.src_qp = qp->remote_qpn;
1024                         wc.pkey_index = 0;
1025                         wc.slid = qp->remote_ah_attr.dlid;
1026                         wc.sl = qp->remote_ah_attr.sl;
1027                         wc.dlid_path_bits = 0;
1028                         wc.port_num = 0;
1029                         ipath_sqerror_qp(qp, &wc);
1030                         break;
1031
1032                 default:
1033                         /* Ignore other reserved NAK error codes */
1034                         goto reserved;
1035                 }
1036                 qp->s_rnr_retry = qp->s_rnr_retry_cnt;
1037                 goto bail;
1038
1039         default:                /* 2: reserved */
1040         reserved:
1041                 /* Ignore reserved NAK codes. */
1042                 goto bail;
1043         }
1044
1045 bail:
1046         return ret;
1047 }
1048
1049 /**
1050  * ipath_rc_rcv_resp - process an incoming RC response packet
1051  * @dev: the device this packet came in on
1052  * @ohdr: the other headers for this packet
1053  * @data: the packet data
1054  * @tlen: the packet length
1055  * @qp: the QP for this packet
1056  * @opcode: the opcode for this packet
1057  * @psn: the packet sequence number for this packet
1058  * @hdrsize: the header length
1059  * @pmtu: the path MTU
1060  * @header_in_data: true if part of the header data is in the data buffer
1061  *
1062  * This is called from ipath_rc_rcv() to process an incoming RC response
1063  * packet for the given QP.
1064  * Called at interrupt level.
1065  */
1066 static inline void ipath_rc_rcv_resp(struct ipath_ibdev *dev,
1067                                      struct ipath_other_headers *ohdr,
1068                                      void *data, u32 tlen,
1069                                      struct ipath_qp *qp,
1070                                      u32 opcode,
1071                                      u32 psn, u32 hdrsize, u32 pmtu,
1072                                      int header_in_data)
1073 {
1074         struct ipath_swqe *wqe;
1075         unsigned long flags;
1076         struct ib_wc wc;
1077         int diff;
1078         u32 pad;
1079         u32 aeth;
1080
1081         spin_lock_irqsave(&qp->s_lock, flags);
1082
1083         /* Ignore invalid responses. */
1084         if (ipath_cmp24(psn, qp->s_next_psn) >= 0)
1085                 goto ack_done;
1086
1087         /* Ignore duplicate responses. */
1088         diff = ipath_cmp24(psn, qp->s_last_psn);
1089         if (unlikely(diff <= 0)) {
1090                 /* Update credits for "ghost" ACKs */
1091                 if (diff == 0 && opcode == OP(ACKNOWLEDGE)) {
1092                         if (!header_in_data)
1093                                 aeth = be32_to_cpu(ohdr->u.aeth);
1094                         else {
1095                                 aeth = be32_to_cpu(((__be32 *) data)[0]);
1096                                 data += sizeof(__be32);
1097                         }
1098                         if ((aeth >> 29) == 0)
1099                                 ipath_get_credit(qp, aeth);
1100                 }
1101                 goto ack_done;
1102         }
1103
1104         if (unlikely(qp->s_last == qp->s_tail))
1105                 goto ack_done;
1106         wqe = get_swqe_ptr(qp, qp->s_last);
1107
1108         switch (opcode) {
1109         case OP(ACKNOWLEDGE):
1110         case OP(ATOMIC_ACKNOWLEDGE):
1111         case OP(RDMA_READ_RESPONSE_FIRST):
1112                 if (!header_in_data)
1113                         aeth = be32_to_cpu(ohdr->u.aeth);
1114                 else {
1115                         aeth = be32_to_cpu(((__be32 *) data)[0]);
1116                         data += sizeof(__be32);
1117                 }
1118                 if (opcode == OP(ATOMIC_ACKNOWLEDGE)) {
1119                         u64 val;
1120
1121                         if (!header_in_data) {
1122                                 __be32 *p = ohdr->u.at.atomic_ack_eth;
1123
1124                                 val = ((u64) be32_to_cpu(p[0]) << 32) |
1125                                         be32_to_cpu(p[1]);
1126                         } else
1127                                 val = be64_to_cpu(((__be64 *) data)[0]);
1128                         *(u64 *) wqe->sg_list[0].vaddr = val;
1129                 }
1130                 if (!do_rc_ack(qp, aeth, psn, opcode) ||
1131                     opcode != OP(RDMA_READ_RESPONSE_FIRST))
1132                         goto ack_done;
1133                 hdrsize += 4;
1134                 if (unlikely(wqe->wr.opcode != IB_WR_RDMA_READ))
1135                         goto ack_op_err;
1136                 /*
1137                  * If this is a response to a resent RDMA read, we
1138                  * have to be careful to copy the data to the right
1139                  * location.
1140                  */
1141                 qp->s_rdma_read_len = restart_sge(&qp->s_rdma_read_sge,
1142                                                   wqe, psn, pmtu);
1143                 goto read_middle;
1144
1145         case OP(RDMA_READ_RESPONSE_MIDDLE):
1146                 /* no AETH, no ACK */
1147                 if (unlikely(ipath_cmp24(psn, qp->s_last_psn + 1))) {
1148                         dev->n_rdma_seq++;
1149                         ipath_restart_rc(qp, qp->s_last_psn + 1, &wc);
1150                         goto ack_done;
1151                 }
1152                 if (unlikely(wqe->wr.opcode != IB_WR_RDMA_READ))
1153                         goto ack_op_err;
1154         read_middle:
1155                 if (unlikely(tlen != (hdrsize + pmtu + 4)))
1156                         goto ack_len_err;
1157                 if (unlikely(pmtu >= qp->s_rdma_read_len))
1158                         goto ack_len_err;
1159
1160                 /* We got a response so update the timeout. */
1161                 spin_lock(&dev->pending_lock);
1162                 if (qp->s_rnr_timeout == 0 && !list_empty(&qp->timerwait))
1163                         list_move_tail(&qp->timerwait,
1164                                        &dev->pending[dev->pending_index]);
1165                 spin_unlock(&dev->pending_lock);
1166                 /*
1167                  * Update the RDMA receive state but do the copy w/o
1168                  * holding the locks and blocking interrupts.
1169                  */
1170                 qp->s_rdma_read_len -= pmtu;
1171                 update_last_psn(qp, psn);
1172                 spin_unlock_irqrestore(&qp->s_lock, flags);
1173                 ipath_copy_sge(&qp->s_rdma_read_sge, data, pmtu);
1174                 goto bail;
1175
1176         case OP(RDMA_READ_RESPONSE_ONLY):
1177                 if (unlikely(ipath_cmp24(psn, qp->s_last_psn + 1))) {
1178                         dev->n_rdma_seq++;
1179                         ipath_restart_rc(qp, qp->s_last_psn + 1, &wc);
1180                         goto ack_done;
1181                 }
1182                 if (unlikely(wqe->wr.opcode != IB_WR_RDMA_READ))
1183                         goto ack_op_err;
1184                 /* Get the number of bytes the message was padded by. */
1185                 pad = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3;
1186                 /*
1187                  * Check that the data size is >= 0 && <= pmtu.
1188                  * Remember to account for the AETH header (4) and
1189                  * ICRC (4).
1190                  */
1191                 if (unlikely(tlen < (hdrsize + pad + 8)))
1192                         goto ack_len_err;
1193                 /*
1194                  * If this is a response to a resent RDMA read, we
1195                  * have to be careful to copy the data to the right
1196                  * location.
1197                  */
1198                 qp->s_rdma_read_len = restart_sge(&qp->s_rdma_read_sge,
1199                                                   wqe, psn, pmtu);
1200                 goto read_last;
1201
1202         case OP(RDMA_READ_RESPONSE_LAST):
1203                 /* ACKs READ req. */
1204                 if (unlikely(ipath_cmp24(psn, qp->s_last_psn + 1))) {
1205                         dev->n_rdma_seq++;
1206                         ipath_restart_rc(qp, qp->s_last_psn + 1, &wc);
1207                         goto ack_done;
1208                 }
1209                 if (unlikely(wqe->wr.opcode != IB_WR_RDMA_READ))
1210                         goto ack_op_err;
1211                 /* Get the number of bytes the message was padded by. */
1212                 pad = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3;
1213                 /*
1214                  * Check that the data size is >= 1 && <= pmtu.
1215                  * Remember to account for the AETH header (4) and
1216                  * ICRC (4).
1217                  */
1218                 if (unlikely(tlen <= (hdrsize + pad + 8)))
1219                         goto ack_len_err;
1220         read_last:
1221                 tlen -= hdrsize + pad + 8;
1222                 if (unlikely(tlen != qp->s_rdma_read_len))
1223                         goto ack_len_err;
1224                 if (!header_in_data)
1225                         aeth = be32_to_cpu(ohdr->u.aeth);
1226                 else {
1227                         aeth = be32_to_cpu(((__be32 *) data)[0]);
1228                         data += sizeof(__be32);
1229                 }
1230                 ipath_copy_sge(&qp->s_rdma_read_sge, data, tlen);
1231                 (void) do_rc_ack(qp, aeth, psn, OP(RDMA_READ_RESPONSE_LAST));
1232                 goto ack_done;
1233         }
1234
1235 ack_done:
1236         spin_unlock_irqrestore(&qp->s_lock, flags);
1237         goto bail;
1238
1239 ack_op_err:
1240         wc.status = IB_WC_LOC_QP_OP_ERR;
1241         goto ack_err;
1242
1243 ack_len_err:
1244         wc.status = IB_WC_LOC_LEN_ERR;
1245 ack_err:
1246         wc.wr_id = wqe->wr.wr_id;
1247         wc.opcode = ib_ipath_wc_opcode[wqe->wr.opcode];
1248         wc.vendor_err = 0;
1249         wc.byte_len = 0;
1250         wc.imm_data = 0;
1251         wc.qp = &qp->ibqp;
1252         wc.src_qp = qp->remote_qpn;
1253         wc.wc_flags = 0;
1254         wc.pkey_index = 0;
1255         wc.slid = qp->remote_ah_attr.dlid;
1256         wc.sl = qp->remote_ah_attr.sl;
1257         wc.dlid_path_bits = 0;
1258         wc.port_num = 0;
1259         ipath_sqerror_qp(qp, &wc);
1260 bail:
1261         return;
1262 }
1263
1264 /**
1265  * ipath_rc_rcv_error - process an incoming duplicate or error RC packet
1266  * @dev: the device this packet came in on
1267  * @ohdr: the other headers for this packet
1268  * @data: the packet data
1269  * @qp: the QP for this packet
1270  * @opcode: the opcode for this packet
1271  * @psn: the packet sequence number for this packet
1272  * @diff: the difference between the PSN and the expected PSN
1273  * @header_in_data: true if part of the header data is in the data buffer
1274  *
1275  * This is called from ipath_rc_rcv() to process an unexpected
1276  * incoming RC packet for the given QP.
1277  * Called at interrupt level.
1278  * Return 1 if no more processing is needed; otherwise return 0 to
1279  * schedule a response to be sent.
1280  */
1281 static inline int ipath_rc_rcv_error(struct ipath_ibdev *dev,
1282                                      struct ipath_other_headers *ohdr,
1283                                      void *data,
1284                                      struct ipath_qp *qp,
1285                                      u32 opcode,
1286                                      u32 psn,
1287                                      int diff,
1288                                      int header_in_data)
1289 {
1290         struct ipath_ack_entry *e;
1291         u8 i, prev;
1292         int old_req;
1293         unsigned long flags;
1294
1295         if (diff > 0) {
1296                 /*
1297                  * Packet sequence error.
1298                  * A NAK will ACK earlier sends and RDMA writes.
1299                  * Don't queue the NAK if we already sent one.
1300                  */
1301                 if (!qp->r_nak_state) {
1302                         qp->r_nak_state = IB_NAK_PSN_ERROR;
1303                         /* Use the expected PSN. */
1304                         qp->r_ack_psn = qp->r_psn;
1305                         goto send_ack;
1306                 }
1307                 goto done;
1308         }
1309
1310         /*
1311          * Handle a duplicate request.  Don't re-execute SEND, RDMA
1312          * write or atomic op.  Don't NAK errors, just silently drop
1313          * the duplicate request.  Note that r_sge, r_len, and
1314          * r_rcv_len may be in use so don't modify them.
1315          *
1316          * We are supposed to ACK the earliest duplicate PSN but we
1317          * can coalesce an outstanding duplicate ACK.  We have to
1318          * send the earliest so that RDMA reads can be restarted at
1319          * the requester's expected PSN.
1320          *
1321          * First, find where this duplicate PSN falls within the
1322          * ACKs previously sent.
1323          */
1324         psn &= IPATH_PSN_MASK;
1325         e = NULL;
1326         old_req = 1;
1327         spin_lock_irqsave(&qp->s_lock, flags);
1328         for (i = qp->r_head_ack_queue; ; i = prev) {
1329                 if (i == qp->s_tail_ack_queue)
1330                         old_req = 0;
1331                 if (i)
1332                         prev = i - 1;
1333                 else
1334                         prev = IPATH_MAX_RDMA_ATOMIC;
1335                 if (prev == qp->r_head_ack_queue) {
1336                         e = NULL;
1337                         break;
1338                 }
1339                 e = &qp->s_ack_queue[prev];
1340                 if (!e->opcode) {
1341                         e = NULL;
1342                         break;
1343                 }
1344                 if (ipath_cmp24(psn, e->psn) >= 0)
1345                         break;
1346         }
1347         switch (opcode) {
1348         case OP(RDMA_READ_REQUEST): {
1349                 struct ib_reth *reth;
1350                 u32 offset;
1351                 u32 len;
1352
1353                 /*
1354                  * If we didn't find the RDMA read request in the ack queue,
1355                  * or the send tasklet is already backed up to send an
1356                  * earlier entry, we can ignore this request.
1357                  */
1358                 if (!e || e->opcode != OP(RDMA_READ_REQUEST) || old_req)
1359                         goto unlock_done;
1360                 /* RETH comes after BTH */
1361                 if (!header_in_data)
1362                         reth = &ohdr->u.rc.reth;
1363                 else {
1364                         reth = (struct ib_reth *)data;
1365                         data += sizeof(*reth);
1366                 }
1367                 /*
1368                  * Address range must be a subset of the original
1369                  * request and start on pmtu boundaries.
1370                  * We reuse the old ack_queue slot since the requester
1371                  * should not back up and request an earlier PSN for the
1372                  * same request.
1373                  */
1374                 offset = ((psn - e->psn) & IPATH_PSN_MASK) *
1375                         ib_mtu_enum_to_int(qp->path_mtu);
1376                 len = be32_to_cpu(reth->length);
1377                 if (unlikely(offset + len > e->rdma_sge.sge.sge_length))
1378                         goto unlock_done;
1379                 if (len != 0) {
1380                         u32 rkey = be32_to_cpu(reth->rkey);
1381                         u64 vaddr = be64_to_cpu(reth->vaddr);
1382                         int ok;
1383
1384                         ok = ipath_rkey_ok(qp, &e->rdma_sge,
1385                                            len, vaddr, rkey,
1386                                            IB_ACCESS_REMOTE_READ);
1387                         if (unlikely(!ok))
1388                                 goto unlock_done;
1389                 } else {
1390                         e->rdma_sge.sg_list = NULL;
1391                         e->rdma_sge.num_sge = 0;
1392                         e->rdma_sge.sge.mr = NULL;
1393                         e->rdma_sge.sge.vaddr = NULL;
1394                         e->rdma_sge.sge.length = 0;
1395                         e->rdma_sge.sge.sge_length = 0;
1396                 }
1397                 e->psn = psn;
1398                 qp->s_ack_state = OP(ACKNOWLEDGE);
1399                 qp->s_tail_ack_queue = prev;
1400                 break;
1401         }
1402
1403         case OP(COMPARE_SWAP):
1404         case OP(FETCH_ADD): {
1405                 /*
1406                  * If we didn't find the atomic request in the ack queue
1407                  * or the send tasklet is already backed up to send an
1408                  * earlier entry, we can ignore this request.
1409                  */
1410                 if (!e || e->opcode != (u8) opcode || old_req)
1411                         goto unlock_done;
1412                 qp->s_ack_state = OP(ACKNOWLEDGE);
1413                 qp->s_tail_ack_queue = prev;
1414                 break;
1415         }
1416
1417         default:
1418                 if (old_req)
1419                         goto unlock_done;
1420                 /*
1421                  * Resend the most recent ACK if this request is
1422                  * after all the previous RDMA reads and atomics.
1423                  */
1424                 if (i == qp->r_head_ack_queue) {
1425                         spin_unlock_irqrestore(&qp->s_lock, flags);
1426                         qp->r_nak_state = 0;
1427                         qp->r_ack_psn = qp->r_psn - 1;
1428                         goto send_ack;
1429                 }
1430                 /*
1431                  * Resend the RDMA read or atomic op which
1432                  * ACKs this duplicate request.
1433                  */
1434                 qp->s_ack_state = OP(ACKNOWLEDGE);
1435                 qp->s_tail_ack_queue = i;
1436                 break;
1437         }
1438         qp->r_nak_state = 0;
1439         spin_unlock_irq(&qp->s_lock);
1440         tasklet_hi_schedule(&qp->s_task);
1441
1442 unlock_done:
1443         spin_unlock_irqrestore(&qp->s_lock, flags);
1444 done:
1445         return 1;
1446
1447 send_ack:
1448         return 0;
1449 }
1450
1451 static void ipath_rc_error(struct ipath_qp *qp, enum ib_wc_status err)
1452 {
1453         unsigned long flags;
1454
1455         spin_lock_irqsave(&qp->s_lock, flags);
1456         qp->state = IB_QPS_ERR;
1457         ipath_error_qp(qp, err);
1458         spin_unlock_irqrestore(&qp->s_lock, flags);
1459 }
1460
1461 /**
1462  * ipath_rc_rcv - process an incoming RC packet
1463  * @dev: the device this packet came in on
1464  * @hdr: the header of this packet
1465  * @has_grh: true if the header has a GRH
1466  * @data: the packet data
1467  * @tlen: the packet length
1468  * @qp: the QP for this packet
1469  *
1470  * This is called from ipath_qp_rcv() to process an incoming RC packet
1471  * for the given QP.
1472  * Called at interrupt level.
1473  */
1474 void ipath_rc_rcv(struct ipath_ibdev *dev, struct ipath_ib_header *hdr,
1475                   int has_grh, void *data, u32 tlen, struct ipath_qp *qp)
1476 {
1477         struct ipath_other_headers *ohdr;
1478         u32 opcode;
1479         u32 hdrsize;
1480         u32 psn;
1481         u32 pad;
1482         struct ib_wc wc;
1483         u32 pmtu = ib_mtu_enum_to_int(qp->path_mtu);
1484         int diff;
1485         struct ib_reth *reth;
1486         int header_in_data;
1487
1488         /* Validate the SLID. See Ch. 9.6.1.5 */
1489         if (unlikely(be16_to_cpu(hdr->lrh[3]) != qp->remote_ah_attr.dlid))
1490                 goto done;
1491
1492         /* Check for GRH */
1493         if (!has_grh) {
1494                 ohdr = &hdr->u.oth;
1495                 hdrsize = 8 + 12;       /* LRH + BTH */
1496                 psn = be32_to_cpu(ohdr->bth[2]);
1497                 header_in_data = 0;
1498         } else {
1499                 ohdr = &hdr->u.l.oth;
1500                 hdrsize = 8 + 40 + 12;  /* LRH + GRH + BTH */
1501                 /*
1502                  * The header with GRH is 60 bytes and the core driver sets
1503                  * the eager header buffer size to 56 bytes so the last 4
1504                  * bytes of the BTH header (PSN) is in the data buffer.
1505                  */
1506                 header_in_data = dev->dd->ipath_rcvhdrentsize == 16;
1507                 if (header_in_data) {
1508                         psn = be32_to_cpu(((__be32 *) data)[0]);
1509                         data += sizeof(__be32);
1510                 } else
1511                         psn = be32_to_cpu(ohdr->bth[2]);
1512         }
1513
1514         /*
1515          * Process responses (ACKs) before anything else.  Note that the
1516          * packet sequence number will be for something in the send work
1517          * queue rather than the expected receive packet sequence number.
1518          * In other words, this QP is the requester.
1519          */
1520         opcode = be32_to_cpu(ohdr->bth[0]) >> 24;
1521         if (opcode >= OP(RDMA_READ_RESPONSE_FIRST) &&
1522             opcode <= OP(ATOMIC_ACKNOWLEDGE)) {
1523                 ipath_rc_rcv_resp(dev, ohdr, data, tlen, qp, opcode, psn,
1524                                   hdrsize, pmtu, header_in_data);
1525                 goto done;
1526         }
1527
1528         /* Compute 24 bits worth of difference. */
1529         diff = ipath_cmp24(psn, qp->r_psn);
1530         if (unlikely(diff)) {
1531                 if (ipath_rc_rcv_error(dev, ohdr, data, qp, opcode,
1532                                        psn, diff, header_in_data))
1533                         goto done;
1534                 goto send_ack;
1535         }
1536
1537         /* Check for opcode sequence errors. */
1538         switch (qp->r_state) {
1539         case OP(SEND_FIRST):
1540         case OP(SEND_MIDDLE):
1541                 if (opcode == OP(SEND_MIDDLE) ||
1542                     opcode == OP(SEND_LAST) ||
1543                     opcode == OP(SEND_LAST_WITH_IMMEDIATE))
1544                         break;
1545         nack_inv:
1546                 ipath_rc_error(qp, IB_WC_REM_INV_REQ_ERR);
1547                 qp->r_nak_state = IB_NAK_INVALID_REQUEST;
1548                 qp->r_ack_psn = qp->r_psn;
1549                 goto send_ack;
1550
1551         case OP(RDMA_WRITE_FIRST):
1552         case OP(RDMA_WRITE_MIDDLE):
1553                 if (opcode == OP(RDMA_WRITE_MIDDLE) ||
1554                     opcode == OP(RDMA_WRITE_LAST) ||
1555                     opcode == OP(RDMA_WRITE_LAST_WITH_IMMEDIATE))
1556                         break;
1557                 goto nack_inv;
1558
1559         default:
1560                 if (opcode == OP(SEND_MIDDLE) ||
1561                     opcode == OP(SEND_LAST) ||
1562                     opcode == OP(SEND_LAST_WITH_IMMEDIATE) ||
1563                     opcode == OP(RDMA_WRITE_MIDDLE) ||
1564                     opcode == OP(RDMA_WRITE_LAST) ||
1565                     opcode == OP(RDMA_WRITE_LAST_WITH_IMMEDIATE))
1566                         goto nack_inv;
1567                 /*
1568                  * Note that it is up to the requester to not send a new
1569                  * RDMA read or atomic operation before receiving an ACK
1570                  * for the previous operation.
1571                  */
1572                 break;
1573         }
1574
1575         wc.imm_data = 0;
1576         wc.wc_flags = 0;
1577
1578         /* OK, process the packet. */
1579         switch (opcode) {
1580         case OP(SEND_FIRST):
1581                 if (!ipath_get_rwqe(qp, 0)) {
1582                 rnr_nak:
1583                         /*
1584                          * A RNR NAK will ACK earlier sends and RDMA writes.
1585                          * Don't queue the NAK if a RDMA read or atomic
1586                          * is pending though.
1587                          */
1588                         if (qp->r_nak_state)
1589                                 goto done;
1590                         qp->r_nak_state = IB_RNR_NAK | qp->r_min_rnr_timer;
1591                         qp->r_ack_psn = qp->r_psn;
1592                         goto send_ack;
1593                 }
1594                 qp->r_rcv_len = 0;
1595                 /* FALLTHROUGH */
1596         case OP(SEND_MIDDLE):
1597         case OP(RDMA_WRITE_MIDDLE):
1598         send_middle:
1599                 /* Check for invalid length PMTU or posted rwqe len. */
1600                 if (unlikely(tlen != (hdrsize + pmtu + 4)))
1601                         goto nack_inv;
1602                 qp->r_rcv_len += pmtu;
1603                 if (unlikely(qp->r_rcv_len > qp->r_len))
1604                         goto nack_inv;
1605                 ipath_copy_sge(&qp->r_sge, data, pmtu);
1606                 break;
1607
1608         case OP(RDMA_WRITE_LAST_WITH_IMMEDIATE):
1609                 /* consume RWQE */
1610                 if (!ipath_get_rwqe(qp, 1))
1611                         goto rnr_nak;
1612                 goto send_last_imm;
1613
1614         case OP(SEND_ONLY):
1615         case OP(SEND_ONLY_WITH_IMMEDIATE):
1616                 if (!ipath_get_rwqe(qp, 0))
1617                         goto rnr_nak;
1618                 qp->r_rcv_len = 0;
1619                 if (opcode == OP(SEND_ONLY))
1620                         goto send_last;
1621                 /* FALLTHROUGH */
1622         case OP(SEND_LAST_WITH_IMMEDIATE):
1623         send_last_imm:
1624                 if (header_in_data) {
1625                         wc.imm_data = *(__be32 *) data;
1626                         data += sizeof(__be32);
1627                 } else {
1628                         /* Immediate data comes after BTH */
1629                         wc.imm_data = ohdr->u.imm_data;
1630                 }
1631                 hdrsize += 4;
1632                 wc.wc_flags = IB_WC_WITH_IMM;
1633                 /* FALLTHROUGH */
1634         case OP(SEND_LAST):
1635         case OP(RDMA_WRITE_LAST):
1636         send_last:
1637                 /* Get the number of bytes the message was padded by. */
1638                 pad = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3;
1639                 /* Check for invalid length. */
1640                 /* XXX LAST len should be >= 1 */
1641                 if (unlikely(tlen < (hdrsize + pad + 4)))
1642                         goto nack_inv;
1643                 /* Don't count the CRC. */
1644                 tlen -= (hdrsize + pad + 4);
1645                 wc.byte_len = tlen + qp->r_rcv_len;
1646                 if (unlikely(wc.byte_len > qp->r_len))
1647                         goto nack_inv;
1648                 ipath_copy_sge(&qp->r_sge, data, tlen);
1649                 qp->r_msn++;
1650                 if (!qp->r_wrid_valid)
1651                         break;
1652                 qp->r_wrid_valid = 0;
1653                 wc.wr_id = qp->r_wr_id;
1654                 wc.status = IB_WC_SUCCESS;
1655                 wc.opcode = IB_WC_RECV;
1656                 wc.vendor_err = 0;
1657                 wc.qp = &qp->ibqp;
1658                 wc.src_qp = qp->remote_qpn;
1659                 wc.pkey_index = 0;
1660                 wc.slid = qp->remote_ah_attr.dlid;
1661                 wc.sl = qp->remote_ah_attr.sl;
1662                 wc.dlid_path_bits = 0;
1663                 wc.port_num = 0;
1664                 /* Signal completion event if the solicited bit is set. */
1665                 ipath_cq_enter(to_icq(qp->ibqp.recv_cq), &wc,
1666                                (ohdr->bth[0] &
1667                                 __constant_cpu_to_be32(1 << 23)) != 0);
1668                 break;
1669
1670         case OP(RDMA_WRITE_FIRST):
1671         case OP(RDMA_WRITE_ONLY):
1672         case OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE):
1673                 /* consume RWQE */
1674                 /* RETH comes after BTH */
1675                 if (!header_in_data)
1676                         reth = &ohdr->u.rc.reth;
1677                 else {
1678                         reth = (struct ib_reth *)data;
1679                         data += sizeof(*reth);
1680                 }
1681                 hdrsize += sizeof(*reth);
1682                 qp->r_len = be32_to_cpu(reth->length);
1683                 qp->r_rcv_len = 0;
1684                 if (qp->r_len != 0) {
1685                         u32 rkey = be32_to_cpu(reth->rkey);
1686                         u64 vaddr = be64_to_cpu(reth->vaddr);
1687                         int ok;
1688
1689                         /* Check rkey & NAK */
1690                         ok = ipath_rkey_ok(qp, &qp->r_sge,
1691                                            qp->r_len, vaddr, rkey,
1692                                            IB_ACCESS_REMOTE_WRITE);
1693                         if (unlikely(!ok))
1694                                 goto nack_acc;
1695                 } else {
1696                         qp->r_sge.sg_list = NULL;
1697                         qp->r_sge.sge.mr = NULL;
1698                         qp->r_sge.sge.vaddr = NULL;
1699                         qp->r_sge.sge.length = 0;
1700                         qp->r_sge.sge.sge_length = 0;
1701                 }
1702                 if (unlikely(!(qp->qp_access_flags &
1703                                IB_ACCESS_REMOTE_WRITE)))
1704                         goto nack_acc;
1705                 if (opcode == OP(RDMA_WRITE_FIRST))
1706                         goto send_middle;
1707                 else if (opcode == OP(RDMA_WRITE_ONLY))
1708                         goto send_last;
1709                 if (!ipath_get_rwqe(qp, 1))
1710                         goto rnr_nak;
1711                 goto send_last_imm;
1712
1713         case OP(RDMA_READ_REQUEST): {
1714                 struct ipath_ack_entry *e;
1715                 u32 len;
1716                 u8 next;
1717
1718                 if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_READ)))
1719                         goto nack_acc;
1720                 next = qp->r_head_ack_queue + 1;
1721                 if (next > IPATH_MAX_RDMA_ATOMIC)
1722                         next = 0;
1723                 if (unlikely(next == qp->s_tail_ack_queue))
1724                         goto nack_inv;
1725                 e = &qp->s_ack_queue[qp->r_head_ack_queue];
1726                 /* RETH comes after BTH */
1727                 if (!header_in_data)
1728                         reth = &ohdr->u.rc.reth;
1729                 else {
1730                         reth = (struct ib_reth *)data;
1731                         data += sizeof(*reth);
1732                 }
1733                 len = be32_to_cpu(reth->length);
1734                 if (len) {
1735                         u32 rkey = be32_to_cpu(reth->rkey);
1736                         u64 vaddr = be64_to_cpu(reth->vaddr);
1737                         int ok;
1738
1739                         /* Check rkey & NAK */
1740                         ok = ipath_rkey_ok(qp, &e->rdma_sge, len, vaddr,
1741                                            rkey, IB_ACCESS_REMOTE_READ);
1742                         if (unlikely(!ok))
1743                                 goto nack_acc;
1744                         /*
1745                          * Update the next expected PSN.  We add 1 later
1746                          * below, so only add the remainder here.
1747                          */
1748                         if (len > pmtu)
1749                                 qp->r_psn += (len - 1) / pmtu;
1750                 } else {
1751                         e->rdma_sge.sg_list = NULL;
1752                         e->rdma_sge.num_sge = 0;
1753                         e->rdma_sge.sge.mr = NULL;
1754                         e->rdma_sge.sge.vaddr = NULL;
1755                         e->rdma_sge.sge.length = 0;
1756                         e->rdma_sge.sge.sge_length = 0;
1757                 }
1758                 e->opcode = opcode;
1759                 e->psn = psn;
1760                 /*
1761                  * We need to increment the MSN here instead of when we
1762                  * finish sending the result since a duplicate request would
1763                  * increment it more than once.
1764                  */
1765                 qp->r_msn++;
1766                 qp->r_psn++;
1767                 qp->r_state = opcode;
1768                 qp->r_nak_state = 0;
1769                 barrier();
1770                 qp->r_head_ack_queue = next;
1771
1772                 /* Call ipath_do_rc_send() in another thread. */
1773                 tasklet_hi_schedule(&qp->s_task);
1774
1775                 goto done;
1776         }
1777
1778         case OP(COMPARE_SWAP):
1779         case OP(FETCH_ADD): {
1780                 struct ib_atomic_eth *ateth;
1781                 struct ipath_ack_entry *e;
1782                 u64 vaddr;
1783                 atomic64_t *maddr;
1784                 u64 sdata;
1785                 u32 rkey;
1786                 u8 next;
1787
1788                 if (unlikely(!(qp->qp_access_flags &
1789                                IB_ACCESS_REMOTE_ATOMIC)))
1790                         goto nack_acc;
1791                 next = qp->r_head_ack_queue + 1;
1792                 if (next > IPATH_MAX_RDMA_ATOMIC)
1793                         next = 0;
1794                 if (unlikely(next == qp->s_tail_ack_queue))
1795                         goto nack_inv;
1796                 if (!header_in_data)
1797                         ateth = &ohdr->u.atomic_eth;
1798                 else
1799                         ateth = (struct ib_atomic_eth *)data;
1800                 vaddr = ((u64) be32_to_cpu(ateth->vaddr[0]) << 32) |
1801                         be32_to_cpu(ateth->vaddr[1]);
1802                 if (unlikely(vaddr & (sizeof(u64) - 1)))
1803                         goto nack_inv;
1804                 rkey = be32_to_cpu(ateth->rkey);
1805                 /* Check rkey & NAK */
1806                 if (unlikely(!ipath_rkey_ok(qp, &qp->r_sge,
1807                                             sizeof(u64), vaddr, rkey,
1808                                             IB_ACCESS_REMOTE_ATOMIC)))
1809                         goto nack_acc;
1810                 /* Perform atomic OP and save result. */
1811                 maddr = (atomic64_t *) qp->r_sge.sge.vaddr;
1812                 sdata = be64_to_cpu(ateth->swap_data);
1813                 e = &qp->s_ack_queue[qp->r_head_ack_queue];
1814                 e->atomic_data = (opcode == OP(FETCH_ADD)) ?
1815                         (u64) atomic64_add_return(sdata, maddr) - sdata :
1816                         (u64) cmpxchg((u64 *) qp->r_sge.sge.vaddr,
1817                                       be64_to_cpu(ateth->compare_data),
1818                                       sdata);
1819                 e->opcode = opcode;
1820                 e->psn = psn & IPATH_PSN_MASK;
1821                 qp->r_msn++;
1822                 qp->r_psn++;
1823                 qp->r_state = opcode;
1824                 qp->r_nak_state = 0;
1825                 barrier();
1826                 qp->r_head_ack_queue = next;
1827
1828                 /* Call ipath_do_rc_send() in another thread. */
1829                 tasklet_hi_schedule(&qp->s_task);
1830
1831                 goto done;
1832         }
1833
1834         default:
1835                 /* NAK unknown opcodes. */
1836                 goto nack_inv;
1837         }
1838         qp->r_psn++;
1839         qp->r_state = opcode;
1840         qp->r_ack_psn = psn;
1841         qp->r_nak_state = 0;
1842         /* Send an ACK if requested or required. */
1843         if (psn & (1 << 31))
1844                 goto send_ack;
1845         goto done;
1846
1847 nack_acc:
1848         ipath_rc_error(qp, IB_WC_REM_ACCESS_ERR);
1849         qp->r_nak_state = IB_NAK_REMOTE_ACCESS_ERROR;
1850         qp->r_ack_psn = qp->r_psn;
1851
1852 send_ack:
1853         send_rc_ack(qp);
1854
1855 done:
1856         return;
1857 }