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