Merge branches 'armv7', 'at91', 'misc' and 'omap' into devel
[linux-2.6] / 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                 /*
102                  * We can increment the tail pointer now that the last
103                  * response has been sent instead of only being
104                  * constructed.
105                  */
106                 if (++qp->s_tail_ack_queue > IPATH_MAX_RDMA_ATOMIC)
107                         qp->s_tail_ack_queue = 0;
108                 /* FALLTHROUGH */
109         case OP(SEND_ONLY):
110         case OP(ACKNOWLEDGE):
111                 /* Check for no next entry in the queue. */
112                 if (qp->r_head_ack_queue == qp->s_tail_ack_queue) {
113                         if (qp->s_flags & IPATH_S_ACK_PENDING)
114                                 goto normal;
115                         qp->s_ack_state = OP(ACKNOWLEDGE);
116                         goto bail;
117                 }
118
119                 e = &qp->s_ack_queue[qp->s_tail_ack_queue];
120                 if (e->opcode == OP(RDMA_READ_REQUEST)) {
121                         /* Copy SGE state in case we need to resend */
122                         qp->s_ack_rdma_sge = e->rdma_sge;
123                         qp->s_cur_sge = &qp->s_ack_rdma_sge;
124                         len = e->rdma_sge.sge.sge_length;
125                         if (len > pmtu) {
126                                 len = pmtu;
127                                 qp->s_ack_state = OP(RDMA_READ_RESPONSE_FIRST);
128                         } else
129                                 qp->s_ack_state = OP(RDMA_READ_RESPONSE_ONLY);
130                         ohdr->u.aeth = ipath_compute_aeth(qp);
131                         hwords++;
132                         qp->s_ack_rdma_psn = e->psn;
133                         bth2 = qp->s_ack_rdma_psn++ & IPATH_PSN_MASK;
134                 } else {
135                         /* COMPARE_SWAP or FETCH_ADD */
136                         qp->s_cur_sge = NULL;
137                         len = 0;
138                         qp->s_ack_state = OP(ATOMIC_ACKNOWLEDGE);
139                         ohdr->u.at.aeth = ipath_compute_aeth(qp);
140                         ohdr->u.at.atomic_ack_eth[0] =
141                                 cpu_to_be32(e->atomic_data >> 32);
142                         ohdr->u.at.atomic_ack_eth[1] =
143                                 cpu_to_be32(e->atomic_data);
144                         hwords += sizeof(ohdr->u.at) / sizeof(u32);
145                         bth2 = e->psn;
146                 }
147                 bth0 = qp->s_ack_state << 24;
148                 break;
149
150         case OP(RDMA_READ_RESPONSE_FIRST):
151                 qp->s_ack_state = OP(RDMA_READ_RESPONSE_MIDDLE);
152                 /* FALLTHROUGH */
153         case OP(RDMA_READ_RESPONSE_MIDDLE):
154                 len = qp->s_ack_rdma_sge.sge.sge_length;
155                 if (len > pmtu)
156                         len = pmtu;
157                 else {
158                         ohdr->u.aeth = ipath_compute_aeth(qp);
159                         hwords++;
160                         qp->s_ack_state = OP(RDMA_READ_RESPONSE_LAST);
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(SEND_ONLY);
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 != OP(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             (qp->s_flags & IPATH_S_ACK_PENDING) ||
590             qp->s_ack_state != OP(ACKNOWLEDGE))
591                 goto queue_ack;
592
593         /* Construct the header. */
594         ohdr = &hdr.u.oth;
595         lrh0 = IPATH_LRH_BTH;
596         /* header size in 32-bit words LRH+BTH+AETH = (8+12+4)/4. */
597         hwords = 6;
598         if (unlikely(qp->remote_ah_attr.ah_flags & IB_AH_GRH)) {
599                 hwords += ipath_make_grh(dev, &hdr.u.l.grh,
600                                          &qp->remote_ah_attr.grh,
601                                          hwords, 0);
602                 ohdr = &hdr.u.l.oth;
603                 lrh0 = IPATH_LRH_GRH;
604         }
605         /* read pkey_index w/o lock (its atomic) */
606         bth0 = ipath_get_pkey(dev->dd, qp->s_pkey_index) |
607                 OP(ACKNOWLEDGE) << 24;
608         if (qp->r_nak_state)
609                 ohdr->u.aeth = cpu_to_be32((qp->r_msn & IPATH_MSN_MASK) |
610                                             (qp->r_nak_state <<
611                                              IPATH_AETH_CREDIT_SHIFT));
612         else
613                 ohdr->u.aeth = ipath_compute_aeth(qp);
614         lrh0 |= qp->remote_ah_attr.sl << 4;
615         hdr.lrh[0] = cpu_to_be16(lrh0);
616         hdr.lrh[1] = cpu_to_be16(qp->remote_ah_attr.dlid);
617         hdr.lrh[2] = cpu_to_be16(hwords + SIZE_OF_CRC);
618         hdr.lrh[3] = cpu_to_be16(dev->dd->ipath_lid);
619         ohdr->bth[0] = cpu_to_be32(bth0);
620         ohdr->bth[1] = cpu_to_be32(qp->remote_qpn);
621         ohdr->bth[2] = cpu_to_be32(qp->r_ack_psn & IPATH_PSN_MASK);
622
623         /*
624          * If we can send the ACK, clear the ACK state.
625          */
626         if (ipath_verbs_send(dev->dd, hwords, (u32 *) &hdr, 0, NULL) == 0) {
627                 dev->n_unicast_xmit++;
628                 goto done;
629         }
630
631         /*
632          * We are out of PIO buffers at the moment.
633          * Pass responsibility for sending the ACK to the
634          * send tasklet so that when a PIO buffer becomes
635          * available, the ACK is sent ahead of other outgoing
636          * packets.
637          */
638         dev->n_rc_qacks++;
639
640 queue_ack:
641         spin_lock_irqsave(&qp->s_lock, flags);
642         qp->s_flags |= IPATH_S_ACK_PENDING;
643         qp->s_nak_state = qp->r_nak_state;
644         qp->s_ack_psn = qp->r_ack_psn;
645         spin_unlock_irqrestore(&qp->s_lock, flags);
646
647         /* Call ipath_do_rc_send() in another thread. */
648         tasklet_hi_schedule(&qp->s_task);
649
650 done:
651         return;
652 }
653
654 /**
655  * reset_psn - reset the QP state to send starting from PSN
656  * @qp: the QP
657  * @psn: the packet sequence number to restart at
658  *
659  * This is called from ipath_rc_rcv() to process an incoming RC ACK
660  * for the given QP.
661  * Called at interrupt level with the QP s_lock held.
662  */
663 static void reset_psn(struct ipath_qp *qp, u32 psn)
664 {
665         u32 n = qp->s_last;
666         struct ipath_swqe *wqe = get_swqe_ptr(qp, n);
667         u32 opcode;
668
669         qp->s_cur = n;
670
671         /*
672          * If we are starting the request from the beginning,
673          * let the normal send code handle initialization.
674          */
675         if (ipath_cmp24(psn, wqe->psn) <= 0) {
676                 qp->s_state = OP(SEND_LAST);
677                 goto done;
678         }
679
680         /* Find the work request opcode corresponding to the given PSN. */
681         opcode = wqe->wr.opcode;
682         for (;;) {
683                 int diff;
684
685                 if (++n == qp->s_size)
686                         n = 0;
687                 if (n == qp->s_tail)
688                         break;
689                 wqe = get_swqe_ptr(qp, n);
690                 diff = ipath_cmp24(psn, wqe->psn);
691                 if (diff < 0)
692                         break;
693                 qp->s_cur = n;
694                 /*
695                  * If we are starting the request from the beginning,
696                  * let the normal send code handle initialization.
697                  */
698                 if (diff == 0) {
699                         qp->s_state = OP(SEND_LAST);
700                         goto done;
701                 }
702                 opcode = wqe->wr.opcode;
703         }
704
705         /*
706          * Set the state to restart in the middle of a request.
707          * Don't change the s_sge, s_cur_sge, or s_cur_size.
708          * See ipath_do_rc_send().
709          */
710         switch (opcode) {
711         case IB_WR_SEND:
712         case IB_WR_SEND_WITH_IMM:
713                 qp->s_state = OP(RDMA_READ_RESPONSE_FIRST);
714                 break;
715
716         case IB_WR_RDMA_WRITE:
717         case IB_WR_RDMA_WRITE_WITH_IMM:
718                 qp->s_state = OP(RDMA_READ_RESPONSE_LAST);
719                 break;
720
721         case IB_WR_RDMA_READ:
722                 qp->s_state = OP(RDMA_READ_RESPONSE_MIDDLE);
723                 break;
724
725         default:
726                 /*
727                  * This case shouldn't happen since its only
728                  * one PSN per req.
729                  */
730                 qp->s_state = OP(SEND_LAST);
731         }
732 done:
733         qp->s_psn = psn;
734 }
735
736 /**
737  * ipath_restart_rc - back up requester to resend the last un-ACKed request
738  * @qp: the QP to restart
739  * @psn: packet sequence number for the request
740  * @wc: the work completion request
741  *
742  * The QP s_lock should be held and interrupts disabled.
743  */
744 void ipath_restart_rc(struct ipath_qp *qp, u32 psn, struct ib_wc *wc)
745 {
746         struct ipath_swqe *wqe = get_swqe_ptr(qp, qp->s_last);
747         struct ipath_ibdev *dev;
748
749         if (qp->s_retry == 0) {
750                 wc->wr_id = wqe->wr.wr_id;
751                 wc->status = IB_WC_RETRY_EXC_ERR;
752                 wc->opcode = ib_ipath_wc_opcode[wqe->wr.opcode];
753                 wc->vendor_err = 0;
754                 wc->byte_len = 0;
755                 wc->qp = &qp->ibqp;
756                 wc->src_qp = qp->remote_qpn;
757                 wc->pkey_index = 0;
758                 wc->slid = qp->remote_ah_attr.dlid;
759                 wc->sl = qp->remote_ah_attr.sl;
760                 wc->dlid_path_bits = 0;
761                 wc->port_num = 0;
762                 ipath_sqerror_qp(qp, wc);
763                 goto bail;
764         }
765         qp->s_retry--;
766
767         /*
768          * Remove the QP from the timeout queue.
769          * Note: it may already have been removed by ipath_ib_timer().
770          */
771         dev = to_idev(qp->ibqp.device);
772         spin_lock(&dev->pending_lock);
773         if (!list_empty(&qp->timerwait))
774                 list_del_init(&qp->timerwait);
775         spin_unlock(&dev->pending_lock);
776
777         if (wqe->wr.opcode == IB_WR_RDMA_READ)
778                 dev->n_rc_resends++;
779         else
780                 dev->n_rc_resends += (qp->s_psn - psn) & IPATH_PSN_MASK;
781
782         reset_psn(qp, psn);
783         tasklet_hi_schedule(&qp->s_task);
784
785 bail:
786         return;
787 }
788
789 static inline void update_last_psn(struct ipath_qp *qp, u32 psn)
790 {
791         if (qp->s_wait_credit) {
792                 qp->s_wait_credit = 0;
793                 tasklet_hi_schedule(&qp->s_task);
794         }
795         qp->s_last_psn = psn;
796 }
797
798 /**
799  * do_rc_ack - process an incoming RC ACK
800  * @qp: the QP the ACK came in on
801  * @psn: the packet sequence number of the ACK
802  * @opcode: the opcode of the request that resulted in the ACK
803  *
804  * This is called from ipath_rc_rcv_resp() to process an incoming RC ACK
805  * for the given QP.
806  * Called at interrupt level with the QP s_lock held and interrupts disabled.
807  * Returns 1 if OK, 0 if current operation should be aborted (NAK).
808  */
809 static int do_rc_ack(struct ipath_qp *qp, u32 aeth, u32 psn, int opcode)
810 {
811         struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
812         struct ib_wc wc;
813         struct ipath_swqe *wqe;
814         int ret = 0;
815         u32 ack_psn;
816
817         /*
818          * Remove the QP from the timeout queue (or RNR timeout queue).
819          * If ipath_ib_timer() has already removed it,
820          * it's OK since we hold the QP s_lock and ipath_restart_rc()
821          * just won't find anything to restart if we ACK everything.
822          */
823         spin_lock(&dev->pending_lock);
824         if (!list_empty(&qp->timerwait))
825                 list_del_init(&qp->timerwait);
826         spin_unlock(&dev->pending_lock);
827
828         /*
829          * Note that NAKs implicitly ACK outstanding SEND and RDMA write
830          * requests and implicitly NAK RDMA read and atomic requests issued
831          * before the NAK'ed request.  The MSN won't include the NAK'ed
832          * request but will include an ACK'ed request(s).
833          */
834         ack_psn = psn;
835         if (aeth >> 29)
836                 ack_psn--;
837         wqe = get_swqe_ptr(qp, qp->s_last);
838
839         /*
840          * The MSN might be for a later WQE than the PSN indicates so
841          * only complete WQEs that the PSN finishes.
842          */
843         while (ipath_cmp24(ack_psn, wqe->lpsn) >= 0) {
844                 /*
845                  * If this request is a RDMA read or atomic, and the ACK is
846                  * for a later operation, this ACK NAKs the RDMA read or
847                  * atomic.  In other words, only a RDMA_READ_LAST or ONLY
848                  * can ACK a RDMA read and likewise for atomic ops.  Note
849                  * that the NAK case can only happen if relaxed ordering is
850                  * used and requests are sent after an RDMA read or atomic
851                  * is sent but before the response is received.
852                  */
853                 if ((wqe->wr.opcode == IB_WR_RDMA_READ &&
854                      (opcode != OP(RDMA_READ_RESPONSE_LAST) ||
855                       ipath_cmp24(ack_psn, wqe->lpsn) != 0)) ||
856                     ((wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
857                       wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD) &&
858                      (opcode != OP(ATOMIC_ACKNOWLEDGE) ||
859                       ipath_cmp24(wqe->psn, psn) != 0))) {
860                         /*
861                          * The last valid PSN seen is the previous
862                          * request's.
863                          */
864                         update_last_psn(qp, wqe->psn - 1);
865                         /* Retry this request. */
866                         ipath_restart_rc(qp, wqe->psn, &wc);
867                         /*
868                          * No need to process the ACK/NAK since we are
869                          * restarting an earlier request.
870                          */
871                         goto bail;
872                 }
873                 if (qp->s_num_rd_atomic &&
874                     (wqe->wr.opcode == IB_WR_RDMA_READ ||
875                      wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
876                      wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD)) {
877                         qp->s_num_rd_atomic--;
878                         /* Restart sending task if fence is complete */
879                         if ((qp->s_flags & IPATH_S_FENCE_PENDING) &&
880                             !qp->s_num_rd_atomic) {
881                                 qp->s_flags &= ~IPATH_S_FENCE_PENDING;
882                                 tasklet_hi_schedule(&qp->s_task);
883                         } else if (qp->s_flags & IPATH_S_RDMAR_PENDING) {
884                                 qp->s_flags &= ~IPATH_S_RDMAR_PENDING;
885                                 tasklet_hi_schedule(&qp->s_task);
886                         }
887                 }
888                 /* Post a send completion queue entry if requested. */
889                 if (!(qp->s_flags & IPATH_S_SIGNAL_REQ_WR) ||
890                     (wqe->wr.send_flags & IB_SEND_SIGNALED)) {
891                         wc.wr_id = wqe->wr.wr_id;
892                         wc.status = IB_WC_SUCCESS;
893                         wc.opcode = ib_ipath_wc_opcode[wqe->wr.opcode];
894                         wc.vendor_err = 0;
895                         wc.byte_len = wqe->length;
896                         wc.imm_data = 0;
897                         wc.qp = &qp->ibqp;
898                         wc.src_qp = qp->remote_qpn;
899                         wc.wc_flags = 0;
900                         wc.pkey_index = 0;
901                         wc.slid = qp->remote_ah_attr.dlid;
902                         wc.sl = qp->remote_ah_attr.sl;
903                         wc.dlid_path_bits = 0;
904                         wc.port_num = 0;
905                         ipath_cq_enter(to_icq(qp->ibqp.send_cq), &wc, 0);
906                 }
907                 qp->s_retry = qp->s_retry_cnt;
908                 /*
909                  * If we are completing a request which is in the process of
910                  * being resent, we can stop resending it since we know the
911                  * responder has already seen it.
912                  */
913                 if (qp->s_last == qp->s_cur) {
914                         if (++qp->s_cur >= qp->s_size)
915                                 qp->s_cur = 0;
916                         qp->s_last = qp->s_cur;
917                         if (qp->s_last == qp->s_tail)
918                                 break;
919                         wqe = get_swqe_ptr(qp, qp->s_cur);
920                         qp->s_state = OP(SEND_LAST);
921                         qp->s_psn = wqe->psn;
922                 } else {
923                         if (++qp->s_last >= qp->s_size)
924                                 qp->s_last = 0;
925                         if (qp->s_last == qp->s_tail)
926                                 break;
927                         wqe = get_swqe_ptr(qp, qp->s_last);
928                 }
929         }
930
931         switch (aeth >> 29) {
932         case 0:         /* ACK */
933                 dev->n_rc_acks++;
934                 /* If this is a partial ACK, reset the retransmit timer. */
935                 if (qp->s_last != qp->s_tail) {
936                         spin_lock(&dev->pending_lock);
937                         list_add_tail(&qp->timerwait,
938                                       &dev->pending[dev->pending_index]);
939                         spin_unlock(&dev->pending_lock);
940                         /*
941                          * If we get a partial ACK for a resent operation,
942                          * we can stop resending the earlier packets and
943                          * continue with the next packet the receiver wants.
944                          */
945                         if (ipath_cmp24(qp->s_psn, psn) <= 0) {
946                                 reset_psn(qp, psn + 1);
947                                 tasklet_hi_schedule(&qp->s_task);
948                         }
949                 } else if (ipath_cmp24(qp->s_psn, psn) <= 0) {
950                         qp->s_state = OP(SEND_LAST);
951                         qp->s_psn = psn + 1;
952                 }
953                 ipath_get_credit(qp, aeth);
954                 qp->s_rnr_retry = qp->s_rnr_retry_cnt;
955                 qp->s_retry = qp->s_retry_cnt;
956                 update_last_psn(qp, psn);
957                 ret = 1;
958                 goto bail;
959
960         case 1:         /* RNR NAK */
961                 dev->n_rnr_naks++;
962                 if (qp->s_last == qp->s_tail)
963                         goto bail;
964                 if (qp->s_rnr_retry == 0) {
965                         wc.status = IB_WC_RNR_RETRY_EXC_ERR;
966                         goto class_b;
967                 }
968                 if (qp->s_rnr_retry_cnt < 7)
969                         qp->s_rnr_retry--;
970
971                 /* The last valid PSN is the previous PSN. */
972                 update_last_psn(qp, psn - 1);
973
974                 if (wqe->wr.opcode == IB_WR_RDMA_READ)
975                         dev->n_rc_resends++;
976                 else
977                         dev->n_rc_resends +=
978                                 (qp->s_psn - psn) & IPATH_PSN_MASK;
979
980                 reset_psn(qp, psn);
981
982                 qp->s_rnr_timeout =
983                         ib_ipath_rnr_table[(aeth >> IPATH_AETH_CREDIT_SHIFT) &
984                                            IPATH_AETH_CREDIT_MASK];
985                 ipath_insert_rnr_queue(qp);
986                 goto bail;
987
988         case 3:         /* NAK */
989                 if (qp->s_last == qp->s_tail)
990                         goto bail;
991                 /* The last valid PSN is the previous PSN. */
992                 update_last_psn(qp, psn - 1);
993                 switch ((aeth >> IPATH_AETH_CREDIT_SHIFT) &
994                         IPATH_AETH_CREDIT_MASK) {
995                 case 0: /* PSN sequence error */
996                         dev->n_seq_naks++;
997                         /*
998                          * Back up to the responder's expected PSN.
999                          * Note that we might get a NAK in the middle of an
1000                          * RDMA READ response which terminates the RDMA
1001                          * READ.
1002                          */
1003                         ipath_restart_rc(qp, psn, &wc);
1004                         break;
1005
1006                 case 1: /* Invalid Request */
1007                         wc.status = IB_WC_REM_INV_REQ_ERR;
1008                         dev->n_other_naks++;
1009                         goto class_b;
1010
1011                 case 2: /* Remote Access Error */
1012                         wc.status = IB_WC_REM_ACCESS_ERR;
1013                         dev->n_other_naks++;
1014                         goto class_b;
1015
1016                 case 3: /* Remote Operation Error */
1017                         wc.status = IB_WC_REM_OP_ERR;
1018                         dev->n_other_naks++;
1019                 class_b:
1020                         wc.wr_id = wqe->wr.wr_id;
1021                         wc.opcode = ib_ipath_wc_opcode[wqe->wr.opcode];
1022                         wc.vendor_err = 0;
1023                         wc.byte_len = 0;
1024                         wc.qp = &qp->ibqp;
1025                         wc.src_qp = qp->remote_qpn;
1026                         wc.pkey_index = 0;
1027                         wc.slid = qp->remote_ah_attr.dlid;
1028                         wc.sl = qp->remote_ah_attr.sl;
1029                         wc.dlid_path_bits = 0;
1030                         wc.port_num = 0;
1031                         ipath_sqerror_qp(qp, &wc);
1032                         break;
1033
1034                 default:
1035                         /* Ignore other reserved NAK error codes */
1036                         goto reserved;
1037                 }
1038                 qp->s_rnr_retry = qp->s_rnr_retry_cnt;
1039                 goto bail;
1040
1041         default:                /* 2: reserved */
1042         reserved:
1043                 /* Ignore reserved NAK codes. */
1044                 goto bail;
1045         }
1046
1047 bail:
1048         return ret;
1049 }
1050
1051 /**
1052  * ipath_rc_rcv_resp - process an incoming RC response packet
1053  * @dev: the device this packet came in on
1054  * @ohdr: the other headers for this packet
1055  * @data: the packet data
1056  * @tlen: the packet length
1057  * @qp: the QP for this packet
1058  * @opcode: the opcode for this packet
1059  * @psn: the packet sequence number for this packet
1060  * @hdrsize: the header length
1061  * @pmtu: the path MTU
1062  * @header_in_data: true if part of the header data is in the data buffer
1063  *
1064  * This is called from ipath_rc_rcv() to process an incoming RC response
1065  * packet for the given QP.
1066  * Called at interrupt level.
1067  */
1068 static inline void ipath_rc_rcv_resp(struct ipath_ibdev *dev,
1069                                      struct ipath_other_headers *ohdr,
1070                                      void *data, u32 tlen,
1071                                      struct ipath_qp *qp,
1072                                      u32 opcode,
1073                                      u32 psn, u32 hdrsize, u32 pmtu,
1074                                      int header_in_data)
1075 {
1076         struct ipath_swqe *wqe;
1077         unsigned long flags;
1078         struct ib_wc wc;
1079         int diff;
1080         u32 pad;
1081         u32 aeth;
1082
1083         spin_lock_irqsave(&qp->s_lock, flags);
1084
1085         /* Ignore invalid responses. */
1086         if (ipath_cmp24(psn, qp->s_next_psn) >= 0)
1087                 goto ack_done;
1088
1089         /* Ignore duplicate responses. */
1090         diff = ipath_cmp24(psn, qp->s_last_psn);
1091         if (unlikely(diff <= 0)) {
1092                 /* Update credits for "ghost" ACKs */
1093                 if (diff == 0 && opcode == OP(ACKNOWLEDGE)) {
1094                         if (!header_in_data)
1095                                 aeth = be32_to_cpu(ohdr->u.aeth);
1096                         else {
1097                                 aeth = be32_to_cpu(((__be32 *) data)[0]);
1098                                 data += sizeof(__be32);
1099                         }
1100                         if ((aeth >> 29) == 0)
1101                                 ipath_get_credit(qp, aeth);
1102                 }
1103                 goto ack_done;
1104         }
1105
1106         if (unlikely(qp->s_last == qp->s_tail))
1107                 goto ack_done;
1108         wqe = get_swqe_ptr(qp, qp->s_last);
1109
1110         switch (opcode) {
1111         case OP(ACKNOWLEDGE):
1112         case OP(ATOMIC_ACKNOWLEDGE):
1113         case OP(RDMA_READ_RESPONSE_FIRST):
1114                 if (!header_in_data)
1115                         aeth = be32_to_cpu(ohdr->u.aeth);
1116                 else {
1117                         aeth = be32_to_cpu(((__be32 *) data)[0]);
1118                         data += sizeof(__be32);
1119                 }
1120                 if (opcode == OP(ATOMIC_ACKNOWLEDGE)) {
1121                         u64 val;
1122
1123                         if (!header_in_data) {
1124                                 __be32 *p = ohdr->u.at.atomic_ack_eth;
1125
1126                                 val = ((u64) be32_to_cpu(p[0]) << 32) |
1127                                         be32_to_cpu(p[1]);
1128                         } else
1129                                 val = be64_to_cpu(((__be64 *) data)[0]);
1130                         *(u64 *) wqe->sg_list[0].vaddr = val;
1131                 }
1132                 if (!do_rc_ack(qp, aeth, psn, opcode) ||
1133                     opcode != OP(RDMA_READ_RESPONSE_FIRST))
1134                         goto ack_done;
1135                 hdrsize += 4;
1136                 if (unlikely(wqe->wr.opcode != IB_WR_RDMA_READ))
1137                         goto ack_op_err;
1138                 /*
1139                  * If this is a response to a resent RDMA read, we
1140                  * have to be careful to copy the data to the right
1141                  * location.
1142                  */
1143                 qp->s_rdma_read_len = restart_sge(&qp->s_rdma_read_sge,
1144                                                   wqe, psn, pmtu);
1145                 goto read_middle;
1146
1147         case OP(RDMA_READ_RESPONSE_MIDDLE):
1148                 /* no AETH, no ACK */
1149                 if (unlikely(ipath_cmp24(psn, qp->s_last_psn + 1))) {
1150                         dev->n_rdma_seq++;
1151                         ipath_restart_rc(qp, qp->s_last_psn + 1, &wc);
1152                         goto ack_done;
1153                 }
1154                 if (unlikely(wqe->wr.opcode != IB_WR_RDMA_READ))
1155                         goto ack_op_err;
1156         read_middle:
1157                 if (unlikely(tlen != (hdrsize + pmtu + 4)))
1158                         goto ack_len_err;
1159                 if (unlikely(pmtu >= qp->s_rdma_read_len))
1160                         goto ack_len_err;
1161
1162                 /* We got a response so update the timeout. */
1163                 spin_lock(&dev->pending_lock);
1164                 if (qp->s_rnr_timeout == 0 && !list_empty(&qp->timerwait))
1165                         list_move_tail(&qp->timerwait,
1166                                        &dev->pending[dev->pending_index]);
1167                 spin_unlock(&dev->pending_lock);
1168                 /*
1169                  * Update the RDMA receive state but do the copy w/o
1170                  * holding the locks and blocking interrupts.
1171                  */
1172                 qp->s_rdma_read_len -= pmtu;
1173                 update_last_psn(qp, psn);
1174                 spin_unlock_irqrestore(&qp->s_lock, flags);
1175                 ipath_copy_sge(&qp->s_rdma_read_sge, data, pmtu);
1176                 goto bail;
1177
1178         case OP(RDMA_READ_RESPONSE_ONLY):
1179                 if (unlikely(ipath_cmp24(psn, qp->s_last_psn + 1))) {
1180                         dev->n_rdma_seq++;
1181                         ipath_restart_rc(qp, qp->s_last_psn + 1, &wc);
1182                         goto ack_done;
1183                 }
1184                 if (unlikely(wqe->wr.opcode != IB_WR_RDMA_READ))
1185                         goto ack_op_err;
1186                 /* Get the number of bytes the message was padded by. */
1187                 pad = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3;
1188                 /*
1189                  * Check that the data size is >= 0 && <= pmtu.
1190                  * Remember to account for the AETH header (4) and
1191                  * ICRC (4).
1192                  */
1193                 if (unlikely(tlen < (hdrsize + pad + 8)))
1194                         goto ack_len_err;
1195                 /*
1196                  * If this is a response to a resent RDMA read, we
1197                  * have to be careful to copy the data to the right
1198                  * location.
1199                  */
1200                 qp->s_rdma_read_len = restart_sge(&qp->s_rdma_read_sge,
1201                                                   wqe, psn, pmtu);
1202                 goto read_last;
1203
1204         case OP(RDMA_READ_RESPONSE_LAST):
1205                 /* ACKs READ req. */
1206                 if (unlikely(ipath_cmp24(psn, qp->s_last_psn + 1))) {
1207                         dev->n_rdma_seq++;
1208                         ipath_restart_rc(qp, qp->s_last_psn + 1, &wc);
1209                         goto ack_done;
1210                 }
1211                 if (unlikely(wqe->wr.opcode != IB_WR_RDMA_READ))
1212                         goto ack_op_err;
1213                 /* Get the number of bytes the message was padded by. */
1214                 pad = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3;
1215                 /*
1216                  * Check that the data size is >= 1 && <= pmtu.
1217                  * Remember to account for the AETH header (4) and
1218                  * ICRC (4).
1219                  */
1220                 if (unlikely(tlen <= (hdrsize + pad + 8)))
1221                         goto ack_len_err;
1222         read_last:
1223                 tlen -= hdrsize + pad + 8;
1224                 if (unlikely(tlen != qp->s_rdma_read_len))
1225                         goto ack_len_err;
1226                 if (!header_in_data)
1227                         aeth = be32_to_cpu(ohdr->u.aeth);
1228                 else {
1229                         aeth = be32_to_cpu(((__be32 *) data)[0]);
1230                         data += sizeof(__be32);
1231                 }
1232                 ipath_copy_sge(&qp->s_rdma_read_sge, data, tlen);
1233                 (void) do_rc_ack(qp, aeth, psn, OP(RDMA_READ_RESPONSE_LAST));
1234                 goto ack_done;
1235         }
1236
1237 ack_done:
1238         spin_unlock_irqrestore(&qp->s_lock, flags);
1239         goto bail;
1240
1241 ack_op_err:
1242         wc.status = IB_WC_LOC_QP_OP_ERR;
1243         goto ack_err;
1244
1245 ack_len_err:
1246         wc.status = IB_WC_LOC_LEN_ERR;
1247 ack_err:
1248         wc.wr_id = wqe->wr.wr_id;
1249         wc.opcode = ib_ipath_wc_opcode[wqe->wr.opcode];
1250         wc.vendor_err = 0;
1251         wc.byte_len = 0;
1252         wc.imm_data = 0;
1253         wc.qp = &qp->ibqp;
1254         wc.src_qp = qp->remote_qpn;
1255         wc.wc_flags = 0;
1256         wc.pkey_index = 0;
1257         wc.slid = qp->remote_ah_attr.dlid;
1258         wc.sl = qp->remote_ah_attr.sl;
1259         wc.dlid_path_bits = 0;
1260         wc.port_num = 0;
1261         ipath_sqerror_qp(qp, &wc);
1262         spin_unlock_irqrestore(&qp->s_lock, flags);
1263 bail:
1264         return;
1265 }
1266
1267 /**
1268  * ipath_rc_rcv_error - process an incoming duplicate or error RC packet
1269  * @dev: the device this packet came in on
1270  * @ohdr: the other headers for this packet
1271  * @data: the packet data
1272  * @qp: the QP for this packet
1273  * @opcode: the opcode for this packet
1274  * @psn: the packet sequence number for this packet
1275  * @diff: the difference between the PSN and the expected PSN
1276  * @header_in_data: true if part of the header data is in the data buffer
1277  *
1278  * This is called from ipath_rc_rcv() to process an unexpected
1279  * incoming RC packet for the given QP.
1280  * Called at interrupt level.
1281  * Return 1 if no more processing is needed; otherwise return 0 to
1282  * schedule a response to be sent.
1283  */
1284 static inline int ipath_rc_rcv_error(struct ipath_ibdev *dev,
1285                                      struct ipath_other_headers *ohdr,
1286                                      void *data,
1287                                      struct ipath_qp *qp,
1288                                      u32 opcode,
1289                                      u32 psn,
1290                                      int diff,
1291                                      int header_in_data)
1292 {
1293         struct ipath_ack_entry *e;
1294         u8 i, prev;
1295         int old_req;
1296         unsigned long flags;
1297
1298         if (diff > 0) {
1299                 /*
1300                  * Packet sequence error.
1301                  * A NAK will ACK earlier sends and RDMA writes.
1302                  * Don't queue the NAK if we already sent one.
1303                  */
1304                 if (!qp->r_nak_state) {
1305                         qp->r_nak_state = IB_NAK_PSN_ERROR;
1306                         /* Use the expected PSN. */
1307                         qp->r_ack_psn = qp->r_psn;
1308                         goto send_ack;
1309                 }
1310                 goto done;
1311         }
1312
1313         /*
1314          * Handle a duplicate request.  Don't re-execute SEND, RDMA
1315          * write or atomic op.  Don't NAK errors, just silently drop
1316          * the duplicate request.  Note that r_sge, r_len, and
1317          * r_rcv_len may be in use so don't modify them.
1318          *
1319          * We are supposed to ACK the earliest duplicate PSN but we
1320          * can coalesce an outstanding duplicate ACK.  We have to
1321          * send the earliest so that RDMA reads can be restarted at
1322          * the requester's expected PSN.
1323          *
1324          * First, find where this duplicate PSN falls within the
1325          * ACKs previously sent.
1326          */
1327         psn &= IPATH_PSN_MASK;
1328         e = NULL;
1329         old_req = 1;
1330         spin_lock_irqsave(&qp->s_lock, flags);
1331         for (i = qp->r_head_ack_queue; ; i = prev) {
1332                 if (i == qp->s_tail_ack_queue)
1333                         old_req = 0;
1334                 if (i)
1335                         prev = i - 1;
1336                 else
1337                         prev = IPATH_MAX_RDMA_ATOMIC;
1338                 if (prev == qp->r_head_ack_queue) {
1339                         e = NULL;
1340                         break;
1341                 }
1342                 e = &qp->s_ack_queue[prev];
1343                 if (!e->opcode) {
1344                         e = NULL;
1345                         break;
1346                 }
1347                 if (ipath_cmp24(psn, e->psn) >= 0)
1348                         break;
1349         }
1350         switch (opcode) {
1351         case OP(RDMA_READ_REQUEST): {
1352                 struct ib_reth *reth;
1353                 u32 offset;
1354                 u32 len;
1355
1356                 /*
1357                  * If we didn't find the RDMA read request in the ack queue,
1358                  * or the send tasklet is already backed up to send an
1359                  * earlier entry, we can ignore this request.
1360                  */
1361                 if (!e || e->opcode != OP(RDMA_READ_REQUEST) || old_req)
1362                         goto unlock_done;
1363                 /* RETH comes after BTH */
1364                 if (!header_in_data)
1365                         reth = &ohdr->u.rc.reth;
1366                 else {
1367                         reth = (struct ib_reth *)data;
1368                         data += sizeof(*reth);
1369                 }
1370                 /*
1371                  * Address range must be a subset of the original
1372                  * request and start on pmtu boundaries.
1373                  * We reuse the old ack_queue slot since the requester
1374                  * should not back up and request an earlier PSN for the
1375                  * same request.
1376                  */
1377                 offset = ((psn - e->psn) & IPATH_PSN_MASK) *
1378                         ib_mtu_enum_to_int(qp->path_mtu);
1379                 len = be32_to_cpu(reth->length);
1380                 if (unlikely(offset + len > e->rdma_sge.sge.sge_length))
1381                         goto unlock_done;
1382                 if (len != 0) {
1383                         u32 rkey = be32_to_cpu(reth->rkey);
1384                         u64 vaddr = be64_to_cpu(reth->vaddr);
1385                         int ok;
1386
1387                         ok = ipath_rkey_ok(qp, &e->rdma_sge,
1388                                            len, vaddr, rkey,
1389                                            IB_ACCESS_REMOTE_READ);
1390                         if (unlikely(!ok))
1391                                 goto unlock_done;
1392                 } else {
1393                         e->rdma_sge.sg_list = NULL;
1394                         e->rdma_sge.num_sge = 0;
1395                         e->rdma_sge.sge.mr = NULL;
1396                         e->rdma_sge.sge.vaddr = NULL;
1397                         e->rdma_sge.sge.length = 0;
1398                         e->rdma_sge.sge.sge_length = 0;
1399                 }
1400                 e->psn = psn;
1401                 qp->s_ack_state = OP(ACKNOWLEDGE);
1402                 qp->s_tail_ack_queue = prev;
1403                 break;
1404         }
1405
1406         case OP(COMPARE_SWAP):
1407         case OP(FETCH_ADD): {
1408                 /*
1409                  * If we didn't find the atomic request in the ack queue
1410                  * or the send tasklet is already backed up to send an
1411                  * earlier entry, we can ignore this request.
1412                  */
1413                 if (!e || e->opcode != (u8) opcode || old_req)
1414                         goto unlock_done;
1415                 qp->s_ack_state = OP(ACKNOWLEDGE);
1416                 qp->s_tail_ack_queue = prev;
1417                 break;
1418         }
1419
1420         default:
1421                 if (old_req)
1422                         goto unlock_done;
1423                 /*
1424                  * Resend the most recent ACK if this request is
1425                  * after all the previous RDMA reads and atomics.
1426                  */
1427                 if (i == qp->r_head_ack_queue) {
1428                         spin_unlock_irqrestore(&qp->s_lock, flags);
1429                         qp->r_nak_state = 0;
1430                         qp->r_ack_psn = qp->r_psn - 1;
1431                         goto send_ack;
1432                 }
1433                 /*
1434                  * Resend the RDMA read or atomic op which
1435                  * ACKs this duplicate request.
1436                  */
1437                 qp->s_ack_state = OP(ACKNOWLEDGE);
1438                 qp->s_tail_ack_queue = i;
1439                 break;
1440         }
1441         qp->r_nak_state = 0;
1442         tasklet_hi_schedule(&qp->s_task);
1443
1444 unlock_done:
1445         spin_unlock_irqrestore(&qp->s_lock, flags);
1446 done:
1447         return 1;
1448
1449 send_ack:
1450         return 0;
1451 }
1452
1453 static void ipath_rc_error(struct ipath_qp *qp, enum ib_wc_status err)
1454 {
1455         unsigned long flags;
1456
1457         spin_lock_irqsave(&qp->s_lock, flags);
1458         qp->state = IB_QPS_ERR;
1459         ipath_error_qp(qp, err);
1460         spin_unlock_irqrestore(&qp->s_lock, flags);
1461 }
1462
1463 /**
1464  * ipath_rc_rcv - process an incoming RC packet
1465  * @dev: the device this packet came in on
1466  * @hdr: the header of this packet
1467  * @has_grh: true if the header has a GRH
1468  * @data: the packet data
1469  * @tlen: the packet length
1470  * @qp: the QP for this packet
1471  *
1472  * This is called from ipath_qp_rcv() to process an incoming RC packet
1473  * for the given QP.
1474  * Called at interrupt level.
1475  */
1476 void ipath_rc_rcv(struct ipath_ibdev *dev, struct ipath_ib_header *hdr,
1477                   int has_grh, void *data, u32 tlen, struct ipath_qp *qp)
1478 {
1479         struct ipath_other_headers *ohdr;
1480         u32 opcode;
1481         u32 hdrsize;
1482         u32 psn;
1483         u32 pad;
1484         struct ib_wc wc;
1485         u32 pmtu = ib_mtu_enum_to_int(qp->path_mtu);
1486         int diff;
1487         struct ib_reth *reth;
1488         int header_in_data;
1489
1490         /* Validate the SLID. See Ch. 9.6.1.5 */
1491         if (unlikely(be16_to_cpu(hdr->lrh[3]) != qp->remote_ah_attr.dlid))
1492                 goto done;
1493
1494         /* Check for GRH */
1495         if (!has_grh) {
1496                 ohdr = &hdr->u.oth;
1497                 hdrsize = 8 + 12;       /* LRH + BTH */
1498                 psn = be32_to_cpu(ohdr->bth[2]);
1499                 header_in_data = 0;
1500         } else {
1501                 ohdr = &hdr->u.l.oth;
1502                 hdrsize = 8 + 40 + 12;  /* LRH + GRH + BTH */
1503                 /*
1504                  * The header with GRH is 60 bytes and the core driver sets
1505                  * the eager header buffer size to 56 bytes so the last 4
1506                  * bytes of the BTH header (PSN) is in the data buffer.
1507                  */
1508                 header_in_data = dev->dd->ipath_rcvhdrentsize == 16;
1509                 if (header_in_data) {
1510                         psn = be32_to_cpu(((__be32 *) data)[0]);
1511                         data += sizeof(__be32);
1512                 } else
1513                         psn = be32_to_cpu(ohdr->bth[2]);
1514         }
1515
1516         /*
1517          * Process responses (ACKs) before anything else.  Note that the
1518          * packet sequence number will be for something in the send work
1519          * queue rather than the expected receive packet sequence number.
1520          * In other words, this QP is the requester.
1521          */
1522         opcode = be32_to_cpu(ohdr->bth[0]) >> 24;
1523         if (opcode >= OP(RDMA_READ_RESPONSE_FIRST) &&
1524             opcode <= OP(ATOMIC_ACKNOWLEDGE)) {
1525                 ipath_rc_rcv_resp(dev, ohdr, data, tlen, qp, opcode, psn,
1526                                   hdrsize, pmtu, header_in_data);
1527                 goto done;
1528         }
1529
1530         /* Compute 24 bits worth of difference. */
1531         diff = ipath_cmp24(psn, qp->r_psn);
1532         if (unlikely(diff)) {
1533                 if (ipath_rc_rcv_error(dev, ohdr, data, qp, opcode,
1534                                        psn, diff, header_in_data))
1535                         goto done;
1536                 goto send_ack;
1537         }
1538
1539         /* Check for opcode sequence errors. */
1540         switch (qp->r_state) {
1541         case OP(SEND_FIRST):
1542         case OP(SEND_MIDDLE):
1543                 if (opcode == OP(SEND_MIDDLE) ||
1544                     opcode == OP(SEND_LAST) ||
1545                     opcode == OP(SEND_LAST_WITH_IMMEDIATE))
1546                         break;
1547         nack_inv:
1548                 ipath_rc_error(qp, IB_WC_REM_INV_REQ_ERR);
1549                 qp->r_nak_state = IB_NAK_INVALID_REQUEST;
1550                 qp->r_ack_psn = qp->r_psn;
1551                 goto send_ack;
1552
1553         case OP(RDMA_WRITE_FIRST):
1554         case OP(RDMA_WRITE_MIDDLE):
1555                 if (opcode == OP(RDMA_WRITE_MIDDLE) ||
1556                     opcode == OP(RDMA_WRITE_LAST) ||
1557                     opcode == OP(RDMA_WRITE_LAST_WITH_IMMEDIATE))
1558                         break;
1559                 goto nack_inv;
1560
1561         default:
1562                 if (opcode == OP(SEND_MIDDLE) ||
1563                     opcode == OP(SEND_LAST) ||
1564                     opcode == OP(SEND_LAST_WITH_IMMEDIATE) ||
1565                     opcode == OP(RDMA_WRITE_MIDDLE) ||
1566                     opcode == OP(RDMA_WRITE_LAST) ||
1567                     opcode == OP(RDMA_WRITE_LAST_WITH_IMMEDIATE))
1568                         goto nack_inv;
1569                 /*
1570                  * Note that it is up to the requester to not send a new
1571                  * RDMA read or atomic operation before receiving an ACK
1572                  * for the previous operation.
1573                  */
1574                 break;
1575         }
1576
1577         wc.imm_data = 0;
1578         wc.wc_flags = 0;
1579
1580         /* OK, process the packet. */
1581         switch (opcode) {
1582         case OP(SEND_FIRST):
1583                 if (!ipath_get_rwqe(qp, 0)) {
1584                 rnr_nak:
1585                         /*
1586                          * A RNR NAK will ACK earlier sends and RDMA writes.
1587                          * Don't queue the NAK if a RDMA read or atomic
1588                          * is pending though.
1589                          */
1590                         if (qp->r_nak_state)
1591                                 goto done;
1592                         qp->r_nak_state = IB_RNR_NAK | qp->r_min_rnr_timer;
1593                         qp->r_ack_psn = qp->r_psn;
1594                         goto send_ack;
1595                 }
1596                 qp->r_rcv_len = 0;
1597                 /* FALLTHROUGH */
1598         case OP(SEND_MIDDLE):
1599         case OP(RDMA_WRITE_MIDDLE):
1600         send_middle:
1601                 /* Check for invalid length PMTU or posted rwqe len. */
1602                 if (unlikely(tlen != (hdrsize + pmtu + 4)))
1603                         goto nack_inv;
1604                 qp->r_rcv_len += pmtu;
1605                 if (unlikely(qp->r_rcv_len > qp->r_len))
1606                         goto nack_inv;
1607                 ipath_copy_sge(&qp->r_sge, data, pmtu);
1608                 break;
1609
1610         case OP(RDMA_WRITE_LAST_WITH_IMMEDIATE):
1611                 /* consume RWQE */
1612                 if (!ipath_get_rwqe(qp, 1))
1613                         goto rnr_nak;
1614                 goto send_last_imm;
1615
1616         case OP(SEND_ONLY):
1617         case OP(SEND_ONLY_WITH_IMMEDIATE):
1618                 if (!ipath_get_rwqe(qp, 0))
1619                         goto rnr_nak;
1620                 qp->r_rcv_len = 0;
1621                 if (opcode == OP(SEND_ONLY))
1622                         goto send_last;
1623                 /* FALLTHROUGH */
1624         case OP(SEND_LAST_WITH_IMMEDIATE):
1625         send_last_imm:
1626                 if (header_in_data) {
1627                         wc.imm_data = *(__be32 *) data;
1628                         data += sizeof(__be32);
1629                 } else {
1630                         /* Immediate data comes after BTH */
1631                         wc.imm_data = ohdr->u.imm_data;
1632                 }
1633                 hdrsize += 4;
1634                 wc.wc_flags = IB_WC_WITH_IMM;
1635                 /* FALLTHROUGH */
1636         case OP(SEND_LAST):
1637         case OP(RDMA_WRITE_LAST):
1638         send_last:
1639                 /* Get the number of bytes the message was padded by. */
1640                 pad = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3;
1641                 /* Check for invalid length. */
1642                 /* XXX LAST len should be >= 1 */
1643                 if (unlikely(tlen < (hdrsize + pad + 4)))
1644                         goto nack_inv;
1645                 /* Don't count the CRC. */
1646                 tlen -= (hdrsize + pad + 4);
1647                 wc.byte_len = tlen + qp->r_rcv_len;
1648                 if (unlikely(wc.byte_len > qp->r_len))
1649                         goto nack_inv;
1650                 ipath_copy_sge(&qp->r_sge, data, tlen);
1651                 qp->r_msn++;
1652                 if (!qp->r_wrid_valid)
1653                         break;
1654                 qp->r_wrid_valid = 0;
1655                 wc.wr_id = qp->r_wr_id;
1656                 wc.status = IB_WC_SUCCESS;
1657                 wc.opcode = IB_WC_RECV;
1658                 wc.vendor_err = 0;
1659                 wc.qp = &qp->ibqp;
1660                 wc.src_qp = qp->remote_qpn;
1661                 wc.pkey_index = 0;
1662                 wc.slid = qp->remote_ah_attr.dlid;
1663                 wc.sl = qp->remote_ah_attr.sl;
1664                 wc.dlid_path_bits = 0;
1665                 wc.port_num = 0;
1666                 /* Signal completion event if the solicited bit is set. */
1667                 ipath_cq_enter(to_icq(qp->ibqp.recv_cq), &wc,
1668                                (ohdr->bth[0] &
1669                                 __constant_cpu_to_be32(1 << 23)) != 0);
1670                 break;
1671
1672         case OP(RDMA_WRITE_FIRST):
1673         case OP(RDMA_WRITE_ONLY):
1674         case OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE):
1675                 /* consume RWQE */
1676                 /* RETH comes after BTH */
1677                 if (!header_in_data)
1678                         reth = &ohdr->u.rc.reth;
1679                 else {
1680                         reth = (struct ib_reth *)data;
1681                         data += sizeof(*reth);
1682                 }
1683                 hdrsize += sizeof(*reth);
1684                 qp->r_len = be32_to_cpu(reth->length);
1685                 qp->r_rcv_len = 0;
1686                 if (qp->r_len != 0) {
1687                         u32 rkey = be32_to_cpu(reth->rkey);
1688                         u64 vaddr = be64_to_cpu(reth->vaddr);
1689                         int ok;
1690
1691                         /* Check rkey & NAK */
1692                         ok = ipath_rkey_ok(qp, &qp->r_sge,
1693                                            qp->r_len, vaddr, rkey,
1694                                            IB_ACCESS_REMOTE_WRITE);
1695                         if (unlikely(!ok))
1696                                 goto nack_acc;
1697                 } else {
1698                         qp->r_sge.sg_list = NULL;
1699                         qp->r_sge.sge.mr = NULL;
1700                         qp->r_sge.sge.vaddr = NULL;
1701                         qp->r_sge.sge.length = 0;
1702                         qp->r_sge.sge.sge_length = 0;
1703                 }
1704                 if (unlikely(!(qp->qp_access_flags &
1705                                IB_ACCESS_REMOTE_WRITE)))
1706                         goto nack_acc;
1707                 if (opcode == OP(RDMA_WRITE_FIRST))
1708                         goto send_middle;
1709                 else if (opcode == OP(RDMA_WRITE_ONLY))
1710                         goto send_last;
1711                 if (!ipath_get_rwqe(qp, 1))
1712                         goto rnr_nak;
1713                 goto send_last_imm;
1714
1715         case OP(RDMA_READ_REQUEST): {
1716                 struct ipath_ack_entry *e;
1717                 u32 len;
1718                 u8 next;
1719
1720                 if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_READ)))
1721                         goto nack_acc;
1722                 next = qp->r_head_ack_queue + 1;
1723                 if (next > IPATH_MAX_RDMA_ATOMIC)
1724                         next = 0;
1725                 if (unlikely(next == qp->s_tail_ack_queue))
1726                         goto nack_inv;
1727                 e = &qp->s_ack_queue[qp->r_head_ack_queue];
1728                 /* RETH comes after BTH */
1729                 if (!header_in_data)
1730                         reth = &ohdr->u.rc.reth;
1731                 else {
1732                         reth = (struct ib_reth *)data;
1733                         data += sizeof(*reth);
1734                 }
1735                 len = be32_to_cpu(reth->length);
1736                 if (len) {
1737                         u32 rkey = be32_to_cpu(reth->rkey);
1738                         u64 vaddr = be64_to_cpu(reth->vaddr);
1739                         int ok;
1740
1741                         /* Check rkey & NAK */
1742                         ok = ipath_rkey_ok(qp, &e->rdma_sge, len, vaddr,
1743                                            rkey, IB_ACCESS_REMOTE_READ);
1744                         if (unlikely(!ok))
1745                                 goto nack_acc;
1746                         /*
1747                          * Update the next expected PSN.  We add 1 later
1748                          * below, so only add the remainder here.
1749                          */
1750                         if (len > pmtu)
1751                                 qp->r_psn += (len - 1) / pmtu;
1752                 } else {
1753                         e->rdma_sge.sg_list = NULL;
1754                         e->rdma_sge.num_sge = 0;
1755                         e->rdma_sge.sge.mr = NULL;
1756                         e->rdma_sge.sge.vaddr = NULL;
1757                         e->rdma_sge.sge.length = 0;
1758                         e->rdma_sge.sge.sge_length = 0;
1759                 }
1760                 e->opcode = opcode;
1761                 e->psn = psn;
1762                 /*
1763                  * We need to increment the MSN here instead of when we
1764                  * finish sending the result since a duplicate request would
1765                  * increment it more than once.
1766                  */
1767                 qp->r_msn++;
1768                 qp->r_psn++;
1769                 qp->r_state = opcode;
1770                 qp->r_nak_state = 0;
1771                 barrier();
1772                 qp->r_head_ack_queue = next;
1773
1774                 /* Call ipath_do_rc_send() in another thread. */
1775                 tasklet_hi_schedule(&qp->s_task);
1776
1777                 goto done;
1778         }
1779
1780         case OP(COMPARE_SWAP):
1781         case OP(FETCH_ADD): {
1782                 struct ib_atomic_eth *ateth;
1783                 struct ipath_ack_entry *e;
1784                 u64 vaddr;
1785                 atomic64_t *maddr;
1786                 u64 sdata;
1787                 u32 rkey;
1788                 u8 next;
1789
1790                 if (unlikely(!(qp->qp_access_flags &
1791                                IB_ACCESS_REMOTE_ATOMIC)))
1792                         goto nack_acc;
1793                 next = qp->r_head_ack_queue + 1;
1794                 if (next > IPATH_MAX_RDMA_ATOMIC)
1795                         next = 0;
1796                 if (unlikely(next == qp->s_tail_ack_queue))
1797                         goto nack_inv;
1798                 if (!header_in_data)
1799                         ateth = &ohdr->u.atomic_eth;
1800                 else
1801                         ateth = (struct ib_atomic_eth *)data;
1802                 vaddr = ((u64) be32_to_cpu(ateth->vaddr[0]) << 32) |
1803                         be32_to_cpu(ateth->vaddr[1]);
1804                 if (unlikely(vaddr & (sizeof(u64) - 1)))
1805                         goto nack_inv;
1806                 rkey = be32_to_cpu(ateth->rkey);
1807                 /* Check rkey & NAK */
1808                 if (unlikely(!ipath_rkey_ok(qp, &qp->r_sge,
1809                                             sizeof(u64), vaddr, rkey,
1810                                             IB_ACCESS_REMOTE_ATOMIC)))
1811                         goto nack_acc;
1812                 /* Perform atomic OP and save result. */
1813                 maddr = (atomic64_t *) qp->r_sge.sge.vaddr;
1814                 sdata = be64_to_cpu(ateth->swap_data);
1815                 e = &qp->s_ack_queue[qp->r_head_ack_queue];
1816                 e->atomic_data = (opcode == OP(FETCH_ADD)) ?
1817                         (u64) atomic64_add_return(sdata, maddr) - sdata :
1818                         (u64) cmpxchg((u64 *) qp->r_sge.sge.vaddr,
1819                                       be64_to_cpu(ateth->compare_data),
1820                                       sdata);
1821                 e->opcode = opcode;
1822                 e->psn = psn & IPATH_PSN_MASK;
1823                 qp->r_msn++;
1824                 qp->r_psn++;
1825                 qp->r_state = opcode;
1826                 qp->r_nak_state = 0;
1827                 barrier();
1828                 qp->r_head_ack_queue = next;
1829
1830                 /* Call ipath_do_rc_send() in another thread. */
1831                 tasklet_hi_schedule(&qp->s_task);
1832
1833                 goto done;
1834         }
1835
1836         default:
1837                 /* NAK unknown opcodes. */
1838                 goto nack_inv;
1839         }
1840         qp->r_psn++;
1841         qp->r_state = opcode;
1842         qp->r_ack_psn = psn;
1843         qp->r_nak_state = 0;
1844         /* Send an ACK if requested or required. */
1845         if (psn & (1 << 31))
1846                 goto send_ack;
1847         goto done;
1848
1849 nack_acc:
1850         ipath_rc_error(qp, IB_WC_REM_ACCESS_ERR);
1851         qp->r_nak_state = IB_NAK_REMOTE_ACCESS_ERROR;
1852         qp->r_ack_psn = qp->r_psn;
1853
1854 send_ack:
1855         send_rc_ack(qp);
1856
1857 done:
1858         return;
1859 }