IB/qib: Decode path MTU optimization
[pandora-kernel.git] / drivers / infiniband / hw / qib / qib_uc.c
1 /*
2  * Copyright (c) 2006, 2007, 2008, 2009, 2010 QLogic Corporation.
3  * All rights reserved.
4  * Copyright (c) 2005, 2006 PathScale, Inc. All rights reserved.
5  *
6  * This software is available to you under a choice of one of two
7  * licenses.  You may choose to be licensed under the terms of the GNU
8  * General Public License (GPL) Version 2, available from the file
9  * COPYING in the main directory of this source tree, or the
10  * OpenIB.org BSD license below:
11  *
12  *     Redistribution and use in source and binary forms, with or
13  *     without modification, are permitted provided that the following
14  *     conditions are met:
15  *
16  *      - Redistributions of source code must retain the above
17  *        copyright notice, this list of conditions and the following
18  *        disclaimer.
19  *
20  *      - Redistributions in binary form must reproduce the above
21  *        copyright notice, this list of conditions and the following
22  *        disclaimer in the documentation and/or other materials
23  *        provided with the distribution.
24  *
25  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
26  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
27  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
28  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
29  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
30  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
31  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
32  * SOFTWARE.
33  */
34
35 #include "qib.h"
36
37 /* cut down ridiculously long IB macro names */
38 #define OP(x) IB_OPCODE_UC_##x
39
40 /**
41  * qib_make_uc_req - construct a request packet (SEND, RDMA write)
42  * @qp: a pointer to the QP
43  *
44  * Return 1 if constructed; otherwise, return 0.
45  */
46 int qib_make_uc_req(struct qib_qp *qp)
47 {
48         struct qib_other_headers *ohdr;
49         struct qib_swqe *wqe;
50         unsigned long flags;
51         u32 hwords;
52         u32 bth0;
53         u32 len;
54         u32 pmtu = qp->pmtu;
55         int ret = 0;
56
57         spin_lock_irqsave(&qp->s_lock, flags);
58
59         if (!(ib_qib_state_ops[qp->state] & QIB_PROCESS_SEND_OK)) {
60                 if (!(ib_qib_state_ops[qp->state] & QIB_FLUSH_SEND))
61                         goto bail;
62                 /* We are in the error state, flush the work request. */
63                 if (qp->s_last == qp->s_head)
64                         goto bail;
65                 /* If DMAs are in progress, we can't flush immediately. */
66                 if (atomic_read(&qp->s_dma_busy)) {
67                         qp->s_flags |= QIB_S_WAIT_DMA;
68                         goto bail;
69                 }
70                 wqe = get_swqe_ptr(qp, qp->s_last);
71                 qib_send_complete(qp, wqe, IB_WC_WR_FLUSH_ERR);
72                 goto done;
73         }
74
75         ohdr = &qp->s_hdr.u.oth;
76         if (qp->remote_ah_attr.ah_flags & IB_AH_GRH)
77                 ohdr = &qp->s_hdr.u.l.oth;
78
79         /* header size in 32-bit words LRH+BTH = (8+12)/4. */
80         hwords = 5;
81         bth0 = 0;
82
83         /* Get the next send request. */
84         wqe = get_swqe_ptr(qp, qp->s_cur);
85         qp->s_wqe = NULL;
86         switch (qp->s_state) {
87         default:
88                 if (!(ib_qib_state_ops[qp->state] &
89                     QIB_PROCESS_NEXT_SEND_OK))
90                         goto bail;
91                 /* Check if send work queue is empty. */
92                 if (qp->s_cur == qp->s_head)
93                         goto bail;
94                 /*
95                  * Start a new request.
96                  */
97                 wqe->psn = qp->s_next_psn;
98                 qp->s_psn = qp->s_next_psn;
99                 qp->s_sge.sge = wqe->sg_list[0];
100                 qp->s_sge.sg_list = wqe->sg_list + 1;
101                 qp->s_sge.num_sge = wqe->wr.num_sge;
102                 qp->s_sge.total_len = wqe->length;
103                 len = wqe->length;
104                 qp->s_len = len;
105                 switch (wqe->wr.opcode) {
106                 case IB_WR_SEND:
107                 case IB_WR_SEND_WITH_IMM:
108                         if (len > pmtu) {
109                                 qp->s_state = OP(SEND_FIRST);
110                                 len = pmtu;
111                                 break;
112                         }
113                         if (wqe->wr.opcode == IB_WR_SEND)
114                                 qp->s_state = OP(SEND_ONLY);
115                         else {
116                                 qp->s_state =
117                                         OP(SEND_ONLY_WITH_IMMEDIATE);
118                                 /* Immediate data comes after the BTH */
119                                 ohdr->u.imm_data = wqe->wr.ex.imm_data;
120                                 hwords += 1;
121                         }
122                         if (wqe->wr.send_flags & IB_SEND_SOLICITED)
123                                 bth0 |= IB_BTH_SOLICITED;
124                         qp->s_wqe = wqe;
125                         if (++qp->s_cur >= qp->s_size)
126                                 qp->s_cur = 0;
127                         break;
128
129                 case IB_WR_RDMA_WRITE:
130                 case IB_WR_RDMA_WRITE_WITH_IMM:
131                         ohdr->u.rc.reth.vaddr =
132                                 cpu_to_be64(wqe->wr.wr.rdma.remote_addr);
133                         ohdr->u.rc.reth.rkey =
134                                 cpu_to_be32(wqe->wr.wr.rdma.rkey);
135                         ohdr->u.rc.reth.length = cpu_to_be32(len);
136                         hwords += sizeof(struct ib_reth) / 4;
137                         if (len > pmtu) {
138                                 qp->s_state = OP(RDMA_WRITE_FIRST);
139                                 len = pmtu;
140                                 break;
141                         }
142                         if (wqe->wr.opcode == IB_WR_RDMA_WRITE)
143                                 qp->s_state = OP(RDMA_WRITE_ONLY);
144                         else {
145                                 qp->s_state =
146                                         OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE);
147                                 /* Immediate data comes after the RETH */
148                                 ohdr->u.rc.imm_data = wqe->wr.ex.imm_data;
149                                 hwords += 1;
150                                 if (wqe->wr.send_flags & IB_SEND_SOLICITED)
151                                         bth0 |= IB_BTH_SOLICITED;
152                         }
153                         qp->s_wqe = wqe;
154                         if (++qp->s_cur >= qp->s_size)
155                                 qp->s_cur = 0;
156                         break;
157
158                 default:
159                         goto bail;
160                 }
161                 break;
162
163         case OP(SEND_FIRST):
164                 qp->s_state = OP(SEND_MIDDLE);
165                 /* FALLTHROUGH */
166         case OP(SEND_MIDDLE):
167                 len = qp->s_len;
168                 if (len > pmtu) {
169                         len = pmtu;
170                         break;
171                 }
172                 if (wqe->wr.opcode == IB_WR_SEND)
173                         qp->s_state = OP(SEND_LAST);
174                 else {
175                         qp->s_state = OP(SEND_LAST_WITH_IMMEDIATE);
176                         /* Immediate data comes after the BTH */
177                         ohdr->u.imm_data = wqe->wr.ex.imm_data;
178                         hwords += 1;
179                 }
180                 if (wqe->wr.send_flags & IB_SEND_SOLICITED)
181                         bth0 |= IB_BTH_SOLICITED;
182                 qp->s_wqe = wqe;
183                 if (++qp->s_cur >= qp->s_size)
184                         qp->s_cur = 0;
185                 break;
186
187         case OP(RDMA_WRITE_FIRST):
188                 qp->s_state = OP(RDMA_WRITE_MIDDLE);
189                 /* FALLTHROUGH */
190         case OP(RDMA_WRITE_MIDDLE):
191                 len = qp->s_len;
192                 if (len > pmtu) {
193                         len = pmtu;
194                         break;
195                 }
196                 if (wqe->wr.opcode == IB_WR_RDMA_WRITE)
197                         qp->s_state = OP(RDMA_WRITE_LAST);
198                 else {
199                         qp->s_state =
200                                 OP(RDMA_WRITE_LAST_WITH_IMMEDIATE);
201                         /* Immediate data comes after the BTH */
202                         ohdr->u.imm_data = wqe->wr.ex.imm_data;
203                         hwords += 1;
204                         if (wqe->wr.send_flags & IB_SEND_SOLICITED)
205                                 bth0 |= IB_BTH_SOLICITED;
206                 }
207                 qp->s_wqe = wqe;
208                 if (++qp->s_cur >= qp->s_size)
209                         qp->s_cur = 0;
210                 break;
211         }
212         qp->s_len -= len;
213         qp->s_hdrwords = hwords;
214         qp->s_cur_sge = &qp->s_sge;
215         qp->s_cur_size = len;
216         qib_make_ruc_header(qp, ohdr, bth0 | (qp->s_state << 24),
217                             qp->s_next_psn++ & QIB_PSN_MASK);
218 done:
219         ret = 1;
220         goto unlock;
221
222 bail:
223         qp->s_flags &= ~QIB_S_BUSY;
224 unlock:
225         spin_unlock_irqrestore(&qp->s_lock, flags);
226         return ret;
227 }
228
229 /**
230  * qib_uc_rcv - handle an incoming UC packet
231  * @ibp: the port the packet came in on
232  * @hdr: the header of the packet
233  * @has_grh: true if the packet has a GRH
234  * @data: the packet data
235  * @tlen: the length of the packet
236  * @qp: the QP for this packet.
237  *
238  * This is called from qib_qp_rcv() to process an incoming UC packet
239  * for the given QP.
240  * Called at interrupt level.
241  */
242 void qib_uc_rcv(struct qib_ibport *ibp, struct qib_ib_header *hdr,
243                 int has_grh, void *data, u32 tlen, struct qib_qp *qp)
244 {
245         struct qib_other_headers *ohdr;
246         unsigned long flags;
247         u32 opcode;
248         u32 hdrsize;
249         u32 psn;
250         u32 pad;
251         struct ib_wc wc;
252         u32 pmtu = qp->pmtu;
253         struct ib_reth *reth;
254         int ret;
255
256         /* Check for GRH */
257         if (!has_grh) {
258                 ohdr = &hdr->u.oth;
259                 hdrsize = 8 + 12;       /* LRH + BTH */
260         } else {
261                 ohdr = &hdr->u.l.oth;
262                 hdrsize = 8 + 40 + 12;  /* LRH + GRH + BTH */
263         }
264
265         opcode = be32_to_cpu(ohdr->bth[0]);
266         spin_lock_irqsave(&qp->s_lock, flags);
267         if (qib_ruc_check_hdr(ibp, hdr, has_grh, qp, opcode))
268                 goto sunlock;
269         spin_unlock_irqrestore(&qp->s_lock, flags);
270
271         psn = be32_to_cpu(ohdr->bth[2]);
272         opcode >>= 24;
273
274         /* Compare the PSN verses the expected PSN. */
275         if (unlikely(qib_cmp24(psn, qp->r_psn) != 0)) {
276                 /*
277                  * Handle a sequence error.
278                  * Silently drop any current message.
279                  */
280                 qp->r_psn = psn;
281 inv:
282                 if (qp->r_state == OP(SEND_FIRST) ||
283                     qp->r_state == OP(SEND_MIDDLE)) {
284                         set_bit(QIB_R_REWIND_SGE, &qp->r_aflags);
285                         qp->r_sge.num_sge = 0;
286                 } else
287                         while (qp->r_sge.num_sge) {
288                                 atomic_dec(&qp->r_sge.sge.mr->refcount);
289                                 if (--qp->r_sge.num_sge)
290                                         qp->r_sge.sge = *qp->r_sge.sg_list++;
291                         }
292                 qp->r_state = OP(SEND_LAST);
293                 switch (opcode) {
294                 case OP(SEND_FIRST):
295                 case OP(SEND_ONLY):
296                 case OP(SEND_ONLY_WITH_IMMEDIATE):
297                         goto send_first;
298
299                 case OP(RDMA_WRITE_FIRST):
300                 case OP(RDMA_WRITE_ONLY):
301                 case OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE):
302                         goto rdma_first;
303
304                 default:
305                         goto drop;
306                 }
307         }
308
309         /* Check for opcode sequence errors. */
310         switch (qp->r_state) {
311         case OP(SEND_FIRST):
312         case OP(SEND_MIDDLE):
313                 if (opcode == OP(SEND_MIDDLE) ||
314                     opcode == OP(SEND_LAST) ||
315                     opcode == OP(SEND_LAST_WITH_IMMEDIATE))
316                         break;
317                 goto inv;
318
319         case OP(RDMA_WRITE_FIRST):
320         case OP(RDMA_WRITE_MIDDLE):
321                 if (opcode == OP(RDMA_WRITE_MIDDLE) ||
322                     opcode == OP(RDMA_WRITE_LAST) ||
323                     opcode == OP(RDMA_WRITE_LAST_WITH_IMMEDIATE))
324                         break;
325                 goto inv;
326
327         default:
328                 if (opcode == OP(SEND_FIRST) ||
329                     opcode == OP(SEND_ONLY) ||
330                     opcode == OP(SEND_ONLY_WITH_IMMEDIATE) ||
331                     opcode == OP(RDMA_WRITE_FIRST) ||
332                     opcode == OP(RDMA_WRITE_ONLY) ||
333                     opcode == OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE))
334                         break;
335                 goto inv;
336         }
337
338         if (qp->state == IB_QPS_RTR && !(qp->r_flags & QIB_R_COMM_EST)) {
339                 qp->r_flags |= QIB_R_COMM_EST;
340                 if (qp->ibqp.event_handler) {
341                         struct ib_event ev;
342
343                         ev.device = qp->ibqp.device;
344                         ev.element.qp = &qp->ibqp;
345                         ev.event = IB_EVENT_COMM_EST;
346                         qp->ibqp.event_handler(&ev, qp->ibqp.qp_context);
347                 }
348         }
349
350         /* OK, process the packet. */
351         switch (opcode) {
352         case OP(SEND_FIRST):
353         case OP(SEND_ONLY):
354         case OP(SEND_ONLY_WITH_IMMEDIATE):
355 send_first:
356                 if (test_and_clear_bit(QIB_R_REWIND_SGE, &qp->r_aflags))
357                         qp->r_sge = qp->s_rdma_read_sge;
358                 else {
359                         ret = qib_get_rwqe(qp, 0);
360                         if (ret < 0)
361                                 goto op_err;
362                         if (!ret)
363                                 goto drop;
364                         /*
365                          * qp->s_rdma_read_sge will be the owner
366                          * of the mr references.
367                          */
368                         qp->s_rdma_read_sge = qp->r_sge;
369                 }
370                 qp->r_rcv_len = 0;
371                 if (opcode == OP(SEND_ONLY))
372                         goto no_immediate_data;
373                 else if (opcode == OP(SEND_ONLY_WITH_IMMEDIATE))
374                         goto send_last_imm;
375                 /* FALLTHROUGH */
376         case OP(SEND_MIDDLE):
377                 /* Check for invalid length PMTU or posted rwqe len. */
378                 if (unlikely(tlen != (hdrsize + pmtu + 4)))
379                         goto rewind;
380                 qp->r_rcv_len += pmtu;
381                 if (unlikely(qp->r_rcv_len > qp->r_len))
382                         goto rewind;
383                 qib_copy_sge(&qp->r_sge, data, pmtu, 0);
384                 break;
385
386         case OP(SEND_LAST_WITH_IMMEDIATE):
387 send_last_imm:
388                 wc.ex.imm_data = ohdr->u.imm_data;
389                 hdrsize += 4;
390                 wc.wc_flags = IB_WC_WITH_IMM;
391                 goto send_last;
392         case OP(SEND_LAST):
393 no_immediate_data:
394                 wc.ex.imm_data = 0;
395                 wc.wc_flags = 0;
396 send_last:
397                 /* Get the number of bytes the message was padded by. */
398                 pad = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3;
399                 /* Check for invalid length. */
400                 /* XXX LAST len should be >= 1 */
401                 if (unlikely(tlen < (hdrsize + pad + 4)))
402                         goto rewind;
403                 /* Don't count the CRC. */
404                 tlen -= (hdrsize + pad + 4);
405                 wc.byte_len = tlen + qp->r_rcv_len;
406                 if (unlikely(wc.byte_len > qp->r_len))
407                         goto rewind;
408                 wc.opcode = IB_WC_RECV;
409 last_imm:
410                 qib_copy_sge(&qp->r_sge, data, tlen, 0);
411                 while (qp->s_rdma_read_sge.num_sge) {
412                         atomic_dec(&qp->s_rdma_read_sge.sge.mr->refcount);
413                         if (--qp->s_rdma_read_sge.num_sge)
414                                 qp->s_rdma_read_sge.sge =
415                                         *qp->s_rdma_read_sge.sg_list++;
416                 }
417                 wc.wr_id = qp->r_wr_id;
418                 wc.status = IB_WC_SUCCESS;
419                 wc.qp = &qp->ibqp;
420                 wc.src_qp = qp->remote_qpn;
421                 wc.slid = qp->remote_ah_attr.dlid;
422                 wc.sl = qp->remote_ah_attr.sl;
423                 /* zero fields that are N/A */
424                 wc.vendor_err = 0;
425                 wc.pkey_index = 0;
426                 wc.dlid_path_bits = 0;
427                 wc.port_num = 0;
428                 wc.csum_ok = 0;
429                 /* Signal completion event if the solicited bit is set. */
430                 qib_cq_enter(to_icq(qp->ibqp.recv_cq), &wc,
431                              (ohdr->bth[0] &
432                                 cpu_to_be32(IB_BTH_SOLICITED)) != 0);
433                 break;
434
435         case OP(RDMA_WRITE_FIRST):
436         case OP(RDMA_WRITE_ONLY):
437         case OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE): /* consume RWQE */
438 rdma_first:
439                 if (unlikely(!(qp->qp_access_flags &
440                                IB_ACCESS_REMOTE_WRITE))) {
441                         goto drop;
442                 }
443                 reth = &ohdr->u.rc.reth;
444                 hdrsize += sizeof(*reth);
445                 qp->r_len = be32_to_cpu(reth->length);
446                 qp->r_rcv_len = 0;
447                 qp->r_sge.sg_list = NULL;
448                 if (qp->r_len != 0) {
449                         u32 rkey = be32_to_cpu(reth->rkey);
450                         u64 vaddr = be64_to_cpu(reth->vaddr);
451                         int ok;
452
453                         /* Check rkey */
454                         ok = qib_rkey_ok(qp, &qp->r_sge.sge, qp->r_len,
455                                          vaddr, rkey, IB_ACCESS_REMOTE_WRITE);
456                         if (unlikely(!ok))
457                                 goto drop;
458                         qp->r_sge.num_sge = 1;
459                 } else {
460                         qp->r_sge.num_sge = 0;
461                         qp->r_sge.sge.mr = NULL;
462                         qp->r_sge.sge.vaddr = NULL;
463                         qp->r_sge.sge.length = 0;
464                         qp->r_sge.sge.sge_length = 0;
465                 }
466                 if (opcode == OP(RDMA_WRITE_ONLY))
467                         goto rdma_last;
468                 else if (opcode == OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE)) {
469                         wc.ex.imm_data = ohdr->u.rc.imm_data;
470                         goto rdma_last_imm;
471                 }
472                 /* FALLTHROUGH */
473         case OP(RDMA_WRITE_MIDDLE):
474                 /* Check for invalid length PMTU or posted rwqe len. */
475                 if (unlikely(tlen != (hdrsize + pmtu + 4)))
476                         goto drop;
477                 qp->r_rcv_len += pmtu;
478                 if (unlikely(qp->r_rcv_len > qp->r_len))
479                         goto drop;
480                 qib_copy_sge(&qp->r_sge, data, pmtu, 1);
481                 break;
482
483         case OP(RDMA_WRITE_LAST_WITH_IMMEDIATE):
484                 wc.ex.imm_data = ohdr->u.imm_data;
485 rdma_last_imm:
486                 hdrsize += 4;
487                 wc.wc_flags = IB_WC_WITH_IMM;
488
489                 /* Get the number of bytes the message was padded by. */
490                 pad = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3;
491                 /* Check for invalid length. */
492                 /* XXX LAST len should be >= 1 */
493                 if (unlikely(tlen < (hdrsize + pad + 4)))
494                         goto drop;
495                 /* Don't count the CRC. */
496                 tlen -= (hdrsize + pad + 4);
497                 if (unlikely(tlen + qp->r_rcv_len != qp->r_len))
498                         goto drop;
499                 if (test_and_clear_bit(QIB_R_REWIND_SGE, &qp->r_aflags))
500                         while (qp->s_rdma_read_sge.num_sge) {
501                                 atomic_dec(&qp->s_rdma_read_sge.sge.mr->
502                                            refcount);
503                                 if (--qp->s_rdma_read_sge.num_sge)
504                                         qp->s_rdma_read_sge.sge =
505                                                 *qp->s_rdma_read_sge.sg_list++;
506                         }
507                 else {
508                         ret = qib_get_rwqe(qp, 1);
509                         if (ret < 0)
510                                 goto op_err;
511                         if (!ret)
512                                 goto drop;
513                 }
514                 wc.byte_len = qp->r_len;
515                 wc.opcode = IB_WC_RECV_RDMA_WITH_IMM;
516                 goto last_imm;
517
518         case OP(RDMA_WRITE_LAST):
519 rdma_last:
520                 /* Get the number of bytes the message was padded by. */
521                 pad = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3;
522                 /* Check for invalid length. */
523                 /* XXX LAST len should be >= 1 */
524                 if (unlikely(tlen < (hdrsize + pad + 4)))
525                         goto drop;
526                 /* Don't count the CRC. */
527                 tlen -= (hdrsize + pad + 4);
528                 if (unlikely(tlen + qp->r_rcv_len != qp->r_len))
529                         goto drop;
530                 qib_copy_sge(&qp->r_sge, data, tlen, 1);
531                 while (qp->r_sge.num_sge) {
532                         atomic_dec(&qp->r_sge.sge.mr->refcount);
533                         if (--qp->r_sge.num_sge)
534                                 qp->r_sge.sge = *qp->r_sge.sg_list++;
535                 }
536                 break;
537
538         default:
539                 /* Drop packet for unknown opcodes. */
540                 goto drop;
541         }
542         qp->r_psn++;
543         qp->r_state = opcode;
544         return;
545
546 rewind:
547         set_bit(QIB_R_REWIND_SGE, &qp->r_aflags);
548         qp->r_sge.num_sge = 0;
549 drop:
550         ibp->n_pkt_drops++;
551         return;
552
553 op_err:
554         qib_rc_error(qp, IB_WC_LOC_QP_OP_ERR);
555         return;
556
557 sunlock:
558         spin_unlock_irqrestore(&qp->s_lock, flags);
559 }