Merge branch 'wimax-2.6.35.y' of git://git.kernel.org/pub/scm/linux/kernel/git/inaky...
[pandora-kernel.git] / drivers / infiniband / hw / ipath / ipath_verbs.c
1 /*
2  * Copyright (c) 2006, 2007, 2008 QLogic Corporation. All rights reserved.
3  * Copyright (c) 2005, 2006 PathScale, Inc. All rights reserved.
4  *
5  * This software is available to you under a choice of one of two
6  * licenses.  You may choose to be licensed under the terms of the GNU
7  * General Public License (GPL) Version 2, available from the file
8  * COPYING in the main directory of this source tree, or the
9  * OpenIB.org BSD license below:
10  *
11  *     Redistribution and use in source and binary forms, with or
12  *     without modification, are permitted provided that the following
13  *     conditions are met:
14  *
15  *      - Redistributions of source code must retain the above
16  *        copyright notice, this list of conditions and the following
17  *        disclaimer.
18  *
19  *      - Redistributions in binary form must reproduce the above
20  *        copyright notice, this list of conditions and the following
21  *        disclaimer in the documentation and/or other materials
22  *        provided with the distribution.
23  *
24  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
25  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
26  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
27  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
28  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
29  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
30  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
31  * SOFTWARE.
32  */
33
34 #include <rdma/ib_mad.h>
35 #include <rdma/ib_user_verbs.h>
36 #include <linux/io.h>
37 #include <linux/slab.h>
38 #include <linux/utsname.h>
39 #include <linux/rculist.h>
40
41 #include "ipath_kernel.h"
42 #include "ipath_verbs.h"
43 #include "ipath_common.h"
44
45 static unsigned int ib_ipath_qp_table_size = 251;
46 module_param_named(qp_table_size, ib_ipath_qp_table_size, uint, S_IRUGO);
47 MODULE_PARM_DESC(qp_table_size, "QP table size");
48
49 unsigned int ib_ipath_lkey_table_size = 12;
50 module_param_named(lkey_table_size, ib_ipath_lkey_table_size, uint,
51                    S_IRUGO);
52 MODULE_PARM_DESC(lkey_table_size,
53                  "LKEY table size in bits (2^n, 1 <= n <= 23)");
54
55 static unsigned int ib_ipath_max_pds = 0xFFFF;
56 module_param_named(max_pds, ib_ipath_max_pds, uint, S_IWUSR | S_IRUGO);
57 MODULE_PARM_DESC(max_pds,
58                  "Maximum number of protection domains to support");
59
60 static unsigned int ib_ipath_max_ahs = 0xFFFF;
61 module_param_named(max_ahs, ib_ipath_max_ahs, uint, S_IWUSR | S_IRUGO);
62 MODULE_PARM_DESC(max_ahs, "Maximum number of address handles to support");
63
64 unsigned int ib_ipath_max_cqes = 0x2FFFF;
65 module_param_named(max_cqes, ib_ipath_max_cqes, uint, S_IWUSR | S_IRUGO);
66 MODULE_PARM_DESC(max_cqes,
67                  "Maximum number of completion queue entries to support");
68
69 unsigned int ib_ipath_max_cqs = 0x1FFFF;
70 module_param_named(max_cqs, ib_ipath_max_cqs, uint, S_IWUSR | S_IRUGO);
71 MODULE_PARM_DESC(max_cqs, "Maximum number of completion queues to support");
72
73 unsigned int ib_ipath_max_qp_wrs = 0x3FFF;
74 module_param_named(max_qp_wrs, ib_ipath_max_qp_wrs, uint,
75                    S_IWUSR | S_IRUGO);
76 MODULE_PARM_DESC(max_qp_wrs, "Maximum number of QP WRs to support");
77
78 unsigned int ib_ipath_max_qps = 16384;
79 module_param_named(max_qps, ib_ipath_max_qps, uint, S_IWUSR | S_IRUGO);
80 MODULE_PARM_DESC(max_qps, "Maximum number of QPs to support");
81
82 unsigned int ib_ipath_max_sges = 0x60;
83 module_param_named(max_sges, ib_ipath_max_sges, uint, S_IWUSR | S_IRUGO);
84 MODULE_PARM_DESC(max_sges, "Maximum number of SGEs to support");
85
86 unsigned int ib_ipath_max_mcast_grps = 16384;
87 module_param_named(max_mcast_grps, ib_ipath_max_mcast_grps, uint,
88                    S_IWUSR | S_IRUGO);
89 MODULE_PARM_DESC(max_mcast_grps,
90                  "Maximum number of multicast groups to support");
91
92 unsigned int ib_ipath_max_mcast_qp_attached = 16;
93 module_param_named(max_mcast_qp_attached, ib_ipath_max_mcast_qp_attached,
94                    uint, S_IWUSR | S_IRUGO);
95 MODULE_PARM_DESC(max_mcast_qp_attached,
96                  "Maximum number of attached QPs to support");
97
98 unsigned int ib_ipath_max_srqs = 1024;
99 module_param_named(max_srqs, ib_ipath_max_srqs, uint, S_IWUSR | S_IRUGO);
100 MODULE_PARM_DESC(max_srqs, "Maximum number of SRQs to support");
101
102 unsigned int ib_ipath_max_srq_sges = 128;
103 module_param_named(max_srq_sges, ib_ipath_max_srq_sges,
104                    uint, S_IWUSR | S_IRUGO);
105 MODULE_PARM_DESC(max_srq_sges, "Maximum number of SRQ SGEs to support");
106
107 unsigned int ib_ipath_max_srq_wrs = 0x1FFFF;
108 module_param_named(max_srq_wrs, ib_ipath_max_srq_wrs,
109                    uint, S_IWUSR | S_IRUGO);
110 MODULE_PARM_DESC(max_srq_wrs, "Maximum number of SRQ WRs support");
111
112 static unsigned int ib_ipath_disable_sma;
113 module_param_named(disable_sma, ib_ipath_disable_sma, uint, S_IWUSR | S_IRUGO);
114 MODULE_PARM_DESC(disable_sma, "Disable the SMA");
115
116 /*
117  * Note that it is OK to post send work requests in the SQE and ERR
118  * states; ipath_do_send() will process them and generate error
119  * completions as per IB 1.2 C10-96.
120  */
121 const int ib_ipath_state_ops[IB_QPS_ERR + 1] = {
122         [IB_QPS_RESET] = 0,
123         [IB_QPS_INIT] = IPATH_POST_RECV_OK,
124         [IB_QPS_RTR] = IPATH_POST_RECV_OK | IPATH_PROCESS_RECV_OK,
125         [IB_QPS_RTS] = IPATH_POST_RECV_OK | IPATH_PROCESS_RECV_OK |
126             IPATH_POST_SEND_OK | IPATH_PROCESS_SEND_OK |
127             IPATH_PROCESS_NEXT_SEND_OK,
128         [IB_QPS_SQD] = IPATH_POST_RECV_OK | IPATH_PROCESS_RECV_OK |
129             IPATH_POST_SEND_OK | IPATH_PROCESS_SEND_OK,
130         [IB_QPS_SQE] = IPATH_POST_RECV_OK | IPATH_PROCESS_RECV_OK |
131             IPATH_POST_SEND_OK | IPATH_FLUSH_SEND,
132         [IB_QPS_ERR] = IPATH_POST_RECV_OK | IPATH_FLUSH_RECV |
133             IPATH_POST_SEND_OK | IPATH_FLUSH_SEND,
134 };
135
136 struct ipath_ucontext {
137         struct ib_ucontext ibucontext;
138 };
139
140 static inline struct ipath_ucontext *to_iucontext(struct ib_ucontext
141                                                   *ibucontext)
142 {
143         return container_of(ibucontext, struct ipath_ucontext, ibucontext);
144 }
145
146 /*
147  * Translate ib_wr_opcode into ib_wc_opcode.
148  */
149 const enum ib_wc_opcode ib_ipath_wc_opcode[] = {
150         [IB_WR_RDMA_WRITE] = IB_WC_RDMA_WRITE,
151         [IB_WR_RDMA_WRITE_WITH_IMM] = IB_WC_RDMA_WRITE,
152         [IB_WR_SEND] = IB_WC_SEND,
153         [IB_WR_SEND_WITH_IMM] = IB_WC_SEND,
154         [IB_WR_RDMA_READ] = IB_WC_RDMA_READ,
155         [IB_WR_ATOMIC_CMP_AND_SWP] = IB_WC_COMP_SWAP,
156         [IB_WR_ATOMIC_FETCH_AND_ADD] = IB_WC_FETCH_ADD
157 };
158
159 /*
160  * System image GUID.
161  */
162 static __be64 sys_image_guid;
163
164 /**
165  * ipath_copy_sge - copy data to SGE memory
166  * @ss: the SGE state
167  * @data: the data to copy
168  * @length: the length of the data
169  */
170 void ipath_copy_sge(struct ipath_sge_state *ss, void *data, u32 length)
171 {
172         struct ipath_sge *sge = &ss->sge;
173
174         while (length) {
175                 u32 len = sge->length;
176
177                 if (len > length)
178                         len = length;
179                 if (len > sge->sge_length)
180                         len = sge->sge_length;
181                 BUG_ON(len == 0);
182                 memcpy(sge->vaddr, data, len);
183                 sge->vaddr += len;
184                 sge->length -= len;
185                 sge->sge_length -= len;
186                 if (sge->sge_length == 0) {
187                         if (--ss->num_sge)
188                                 *sge = *ss->sg_list++;
189                 } else if (sge->length == 0 && sge->mr != NULL) {
190                         if (++sge->n >= IPATH_SEGSZ) {
191                                 if (++sge->m >= sge->mr->mapsz)
192                                         break;
193                                 sge->n = 0;
194                         }
195                         sge->vaddr =
196                                 sge->mr->map[sge->m]->segs[sge->n].vaddr;
197                         sge->length =
198                                 sge->mr->map[sge->m]->segs[sge->n].length;
199                 }
200                 data += len;
201                 length -= len;
202         }
203 }
204
205 /**
206  * ipath_skip_sge - skip over SGE memory - XXX almost dup of prev func
207  * @ss: the SGE state
208  * @length: the number of bytes to skip
209  */
210 void ipath_skip_sge(struct ipath_sge_state *ss, u32 length)
211 {
212         struct ipath_sge *sge = &ss->sge;
213
214         while (length) {
215                 u32 len = sge->length;
216
217                 if (len > length)
218                         len = length;
219                 if (len > sge->sge_length)
220                         len = sge->sge_length;
221                 BUG_ON(len == 0);
222                 sge->vaddr += len;
223                 sge->length -= len;
224                 sge->sge_length -= len;
225                 if (sge->sge_length == 0) {
226                         if (--ss->num_sge)
227                                 *sge = *ss->sg_list++;
228                 } else if (sge->length == 0 && sge->mr != NULL) {
229                         if (++sge->n >= IPATH_SEGSZ) {
230                                 if (++sge->m >= sge->mr->mapsz)
231                                         break;
232                                 sge->n = 0;
233                         }
234                         sge->vaddr =
235                                 sge->mr->map[sge->m]->segs[sge->n].vaddr;
236                         sge->length =
237                                 sge->mr->map[sge->m]->segs[sge->n].length;
238                 }
239                 length -= len;
240         }
241 }
242
243 /*
244  * Count the number of DMA descriptors needed to send length bytes of data.
245  * Don't modify the ipath_sge_state to get the count.
246  * Return zero if any of the segments is not aligned.
247  */
248 static u32 ipath_count_sge(struct ipath_sge_state *ss, u32 length)
249 {
250         struct ipath_sge *sg_list = ss->sg_list;
251         struct ipath_sge sge = ss->sge;
252         u8 num_sge = ss->num_sge;
253         u32 ndesc = 1;  /* count the header */
254
255         while (length) {
256                 u32 len = sge.length;
257
258                 if (len > length)
259                         len = length;
260                 if (len > sge.sge_length)
261                         len = sge.sge_length;
262                 BUG_ON(len == 0);
263                 if (((long) sge.vaddr & (sizeof(u32) - 1)) ||
264                     (len != length && (len & (sizeof(u32) - 1)))) {
265                         ndesc = 0;
266                         break;
267                 }
268                 ndesc++;
269                 sge.vaddr += len;
270                 sge.length -= len;
271                 sge.sge_length -= len;
272                 if (sge.sge_length == 0) {
273                         if (--num_sge)
274                                 sge = *sg_list++;
275                 } else if (sge.length == 0 && sge.mr != NULL) {
276                         if (++sge.n >= IPATH_SEGSZ) {
277                                 if (++sge.m >= sge.mr->mapsz)
278                                         break;
279                                 sge.n = 0;
280                         }
281                         sge.vaddr =
282                                 sge.mr->map[sge.m]->segs[sge.n].vaddr;
283                         sge.length =
284                                 sge.mr->map[sge.m]->segs[sge.n].length;
285                 }
286                 length -= len;
287         }
288         return ndesc;
289 }
290
291 /*
292  * Copy from the SGEs to the data buffer.
293  */
294 static void ipath_copy_from_sge(void *data, struct ipath_sge_state *ss,
295                                 u32 length)
296 {
297         struct ipath_sge *sge = &ss->sge;
298
299         while (length) {
300                 u32 len = sge->length;
301
302                 if (len > length)
303                         len = length;
304                 if (len > sge->sge_length)
305                         len = sge->sge_length;
306                 BUG_ON(len == 0);
307                 memcpy(data, sge->vaddr, len);
308                 sge->vaddr += len;
309                 sge->length -= len;
310                 sge->sge_length -= len;
311                 if (sge->sge_length == 0) {
312                         if (--ss->num_sge)
313                                 *sge = *ss->sg_list++;
314                 } else if (sge->length == 0 && sge->mr != NULL) {
315                         if (++sge->n >= IPATH_SEGSZ) {
316                                 if (++sge->m >= sge->mr->mapsz)
317                                         break;
318                                 sge->n = 0;
319                         }
320                         sge->vaddr =
321                                 sge->mr->map[sge->m]->segs[sge->n].vaddr;
322                         sge->length =
323                                 sge->mr->map[sge->m]->segs[sge->n].length;
324                 }
325                 data += len;
326                 length -= len;
327         }
328 }
329
330 /**
331  * ipath_post_one_send - post one RC, UC, or UD send work request
332  * @qp: the QP to post on
333  * @wr: the work request to send
334  */
335 static int ipath_post_one_send(struct ipath_qp *qp, struct ib_send_wr *wr)
336 {
337         struct ipath_swqe *wqe;
338         u32 next;
339         int i;
340         int j;
341         int acc;
342         int ret;
343         unsigned long flags;
344         struct ipath_devdata *dd = to_idev(qp->ibqp.device)->dd;
345
346         spin_lock_irqsave(&qp->s_lock, flags);
347
348         if (qp->ibqp.qp_type != IB_QPT_SMI &&
349             !(dd->ipath_flags & IPATH_LINKACTIVE)) {
350                 ret = -ENETDOWN;
351                 goto bail;
352         }
353
354         /* Check that state is OK to post send. */
355         if (unlikely(!(ib_ipath_state_ops[qp->state] & IPATH_POST_SEND_OK)))
356                 goto bail_inval;
357
358         /* IB spec says that num_sge == 0 is OK. */
359         if (wr->num_sge > qp->s_max_sge)
360                 goto bail_inval;
361
362         /*
363          * Don't allow RDMA reads or atomic operations on UC or
364          * undefined operations.
365          * Make sure buffer is large enough to hold the result for atomics.
366          */
367         if (qp->ibqp.qp_type == IB_QPT_UC) {
368                 if ((unsigned) wr->opcode >= IB_WR_RDMA_READ)
369                         goto bail_inval;
370         } else if (qp->ibqp.qp_type == IB_QPT_UD) {
371                 /* Check UD opcode */
372                 if (wr->opcode != IB_WR_SEND &&
373                     wr->opcode != IB_WR_SEND_WITH_IMM)
374                         goto bail_inval;
375                 /* Check UD destination address PD */
376                 if (qp->ibqp.pd != wr->wr.ud.ah->pd)
377                         goto bail_inval;
378         } else if ((unsigned) wr->opcode > IB_WR_ATOMIC_FETCH_AND_ADD)
379                 goto bail_inval;
380         else if (wr->opcode >= IB_WR_ATOMIC_CMP_AND_SWP &&
381                    (wr->num_sge == 0 ||
382                     wr->sg_list[0].length < sizeof(u64) ||
383                     wr->sg_list[0].addr & (sizeof(u64) - 1)))
384                 goto bail_inval;
385         else if (wr->opcode >= IB_WR_RDMA_READ && !qp->s_max_rd_atomic)
386                 goto bail_inval;
387
388         next = qp->s_head + 1;
389         if (next >= qp->s_size)
390                 next = 0;
391         if (next == qp->s_last) {
392                 ret = -ENOMEM;
393                 goto bail;
394         }
395
396         wqe = get_swqe_ptr(qp, qp->s_head);
397         wqe->wr = *wr;
398         wqe->length = 0;
399         if (wr->num_sge) {
400                 acc = wr->opcode >= IB_WR_RDMA_READ ?
401                         IB_ACCESS_LOCAL_WRITE : 0;
402                 for (i = 0, j = 0; i < wr->num_sge; i++) {
403                         u32 length = wr->sg_list[i].length;
404                         int ok;
405
406                         if (length == 0)
407                                 continue;
408                         ok = ipath_lkey_ok(qp, &wqe->sg_list[j],
409                                            &wr->sg_list[i], acc);
410                         if (!ok)
411                                 goto bail_inval;
412                         wqe->length += length;
413                         j++;
414                 }
415                 wqe->wr.num_sge = j;
416         }
417         if (qp->ibqp.qp_type == IB_QPT_UC ||
418             qp->ibqp.qp_type == IB_QPT_RC) {
419                 if (wqe->length > 0x80000000U)
420                         goto bail_inval;
421         } else if (wqe->length > to_idev(qp->ibqp.device)->dd->ipath_ibmtu)
422                 goto bail_inval;
423         wqe->ssn = qp->s_ssn++;
424         qp->s_head = next;
425
426         ret = 0;
427         goto bail;
428
429 bail_inval:
430         ret = -EINVAL;
431 bail:
432         spin_unlock_irqrestore(&qp->s_lock, flags);
433         return ret;
434 }
435
436 /**
437  * ipath_post_send - post a send on a QP
438  * @ibqp: the QP to post the send on
439  * @wr: the list of work requests to post
440  * @bad_wr: the first bad WR is put here
441  *
442  * This may be called from interrupt context.
443  */
444 static int ipath_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
445                            struct ib_send_wr **bad_wr)
446 {
447         struct ipath_qp *qp = to_iqp(ibqp);
448         int err = 0;
449
450         for (; wr; wr = wr->next) {
451                 err = ipath_post_one_send(qp, wr);
452                 if (err) {
453                         *bad_wr = wr;
454                         goto bail;
455                 }
456         }
457
458         /* Try to do the send work in the caller's context. */
459         ipath_do_send((unsigned long) qp);
460
461 bail:
462         return err;
463 }
464
465 /**
466  * ipath_post_receive - post a receive on a QP
467  * @ibqp: the QP to post the receive on
468  * @wr: the WR to post
469  * @bad_wr: the first bad WR is put here
470  *
471  * This may be called from interrupt context.
472  */
473 static int ipath_post_receive(struct ib_qp *ibqp, struct ib_recv_wr *wr,
474                               struct ib_recv_wr **bad_wr)
475 {
476         struct ipath_qp *qp = to_iqp(ibqp);
477         struct ipath_rwq *wq = qp->r_rq.wq;
478         unsigned long flags;
479         int ret;
480
481         /* Check that state is OK to post receive. */
482         if (!(ib_ipath_state_ops[qp->state] & IPATH_POST_RECV_OK) || !wq) {
483                 *bad_wr = wr;
484                 ret = -EINVAL;
485                 goto bail;
486         }
487
488         for (; wr; wr = wr->next) {
489                 struct ipath_rwqe *wqe;
490                 u32 next;
491                 int i;
492
493                 if ((unsigned) wr->num_sge > qp->r_rq.max_sge) {
494                         *bad_wr = wr;
495                         ret = -EINVAL;
496                         goto bail;
497                 }
498
499                 spin_lock_irqsave(&qp->r_rq.lock, flags);
500                 next = wq->head + 1;
501                 if (next >= qp->r_rq.size)
502                         next = 0;
503                 if (next == wq->tail) {
504                         spin_unlock_irqrestore(&qp->r_rq.lock, flags);
505                         *bad_wr = wr;
506                         ret = -ENOMEM;
507                         goto bail;
508                 }
509
510                 wqe = get_rwqe_ptr(&qp->r_rq, wq->head);
511                 wqe->wr_id = wr->wr_id;
512                 wqe->num_sge = wr->num_sge;
513                 for (i = 0; i < wr->num_sge; i++)
514                         wqe->sg_list[i] = wr->sg_list[i];
515                 /* Make sure queue entry is written before the head index. */
516                 smp_wmb();
517                 wq->head = next;
518                 spin_unlock_irqrestore(&qp->r_rq.lock, flags);
519         }
520         ret = 0;
521
522 bail:
523         return ret;
524 }
525
526 /**
527  * ipath_qp_rcv - processing an incoming packet on a QP
528  * @dev: the device the packet came on
529  * @hdr: the packet header
530  * @has_grh: true if the packet has a GRH
531  * @data: the packet data
532  * @tlen: the packet length
533  * @qp: the QP the packet came on
534  *
535  * This is called from ipath_ib_rcv() to process an incoming packet
536  * for the given QP.
537  * Called at interrupt level.
538  */
539 static void ipath_qp_rcv(struct ipath_ibdev *dev,
540                          struct ipath_ib_header *hdr, int has_grh,
541                          void *data, u32 tlen, struct ipath_qp *qp)
542 {
543         /* Check for valid receive state. */
544         if (!(ib_ipath_state_ops[qp->state] & IPATH_PROCESS_RECV_OK)) {
545                 dev->n_pkt_drops++;
546                 return;
547         }
548
549         switch (qp->ibqp.qp_type) {
550         case IB_QPT_SMI:
551         case IB_QPT_GSI:
552                 if (ib_ipath_disable_sma)
553                         break;
554                 /* FALLTHROUGH */
555         case IB_QPT_UD:
556                 ipath_ud_rcv(dev, hdr, has_grh, data, tlen, qp);
557                 break;
558
559         case IB_QPT_RC:
560                 ipath_rc_rcv(dev, hdr, has_grh, data, tlen, qp);
561                 break;
562
563         case IB_QPT_UC:
564                 ipath_uc_rcv(dev, hdr, has_grh, data, tlen, qp);
565                 break;
566
567         default:
568                 break;
569         }
570 }
571
572 /**
573  * ipath_ib_rcv - process an incoming packet
574  * @arg: the device pointer
575  * @rhdr: the header of the packet
576  * @data: the packet data
577  * @tlen: the packet length
578  *
579  * This is called from ipath_kreceive() to process an incoming packet at
580  * interrupt level. Tlen is the length of the header + data + CRC in bytes.
581  */
582 void ipath_ib_rcv(struct ipath_ibdev *dev, void *rhdr, void *data,
583                   u32 tlen)
584 {
585         struct ipath_ib_header *hdr = rhdr;
586         struct ipath_other_headers *ohdr;
587         struct ipath_qp *qp;
588         u32 qp_num;
589         int lnh;
590         u8 opcode;
591         u16 lid;
592
593         if (unlikely(dev == NULL))
594                 goto bail;
595
596         if (unlikely(tlen < 24)) {      /* LRH+BTH+CRC */
597                 dev->rcv_errors++;
598                 goto bail;
599         }
600
601         /* Check for a valid destination LID (see ch. 7.11.1). */
602         lid = be16_to_cpu(hdr->lrh[1]);
603         if (lid < IPATH_MULTICAST_LID_BASE) {
604                 lid &= ~((1 << dev->dd->ipath_lmc) - 1);
605                 if (unlikely(lid != dev->dd->ipath_lid)) {
606                         dev->rcv_errors++;
607                         goto bail;
608                 }
609         }
610
611         /* Check for GRH */
612         lnh = be16_to_cpu(hdr->lrh[0]) & 3;
613         if (lnh == IPATH_LRH_BTH)
614                 ohdr = &hdr->u.oth;
615         else if (lnh == IPATH_LRH_GRH)
616                 ohdr = &hdr->u.l.oth;
617         else {
618                 dev->rcv_errors++;
619                 goto bail;
620         }
621
622         opcode = be32_to_cpu(ohdr->bth[0]) >> 24;
623         dev->opstats[opcode].n_bytes += tlen;
624         dev->opstats[opcode].n_packets++;
625
626         /* Get the destination QP number. */
627         qp_num = be32_to_cpu(ohdr->bth[1]) & IPATH_QPN_MASK;
628         if (qp_num == IPATH_MULTICAST_QPN) {
629                 struct ipath_mcast *mcast;
630                 struct ipath_mcast_qp *p;
631
632                 if (lnh != IPATH_LRH_GRH) {
633                         dev->n_pkt_drops++;
634                         goto bail;
635                 }
636                 mcast = ipath_mcast_find(&hdr->u.l.grh.dgid);
637                 if (mcast == NULL) {
638                         dev->n_pkt_drops++;
639                         goto bail;
640                 }
641                 dev->n_multicast_rcv++;
642                 list_for_each_entry_rcu(p, &mcast->qp_list, list)
643                         ipath_qp_rcv(dev, hdr, 1, data, tlen, p->qp);
644                 /*
645                  * Notify ipath_multicast_detach() if it is waiting for us
646                  * to finish.
647                  */
648                 if (atomic_dec_return(&mcast->refcount) <= 1)
649                         wake_up(&mcast->wait);
650         } else {
651                 qp = ipath_lookup_qpn(&dev->qp_table, qp_num);
652                 if (qp) {
653                         dev->n_unicast_rcv++;
654                         ipath_qp_rcv(dev, hdr, lnh == IPATH_LRH_GRH, data,
655                                      tlen, qp);
656                         /*
657                          * Notify ipath_destroy_qp() if it is waiting
658                          * for us to finish.
659                          */
660                         if (atomic_dec_and_test(&qp->refcount))
661                                 wake_up(&qp->wait);
662                 } else
663                         dev->n_pkt_drops++;
664         }
665
666 bail:;
667 }
668
669 /**
670  * ipath_ib_timer - verbs timer
671  * @arg: the device pointer
672  *
673  * This is called from ipath_do_rcv_timer() at interrupt level to check for
674  * QPs which need retransmits and to collect performance numbers.
675  */
676 static void ipath_ib_timer(struct ipath_ibdev *dev)
677 {
678         struct ipath_qp *resend = NULL;
679         struct ipath_qp *rnr = NULL;
680         struct list_head *last;
681         struct ipath_qp *qp;
682         unsigned long flags;
683
684         if (dev == NULL)
685                 return;
686
687         spin_lock_irqsave(&dev->pending_lock, flags);
688         /* Start filling the next pending queue. */
689         if (++dev->pending_index >= ARRAY_SIZE(dev->pending))
690                 dev->pending_index = 0;
691         /* Save any requests still in the new queue, they have timed out. */
692         last = &dev->pending[dev->pending_index];
693         while (!list_empty(last)) {
694                 qp = list_entry(last->next, struct ipath_qp, timerwait);
695                 list_del_init(&qp->timerwait);
696                 qp->timer_next = resend;
697                 resend = qp;
698                 atomic_inc(&qp->refcount);
699         }
700         last = &dev->rnrwait;
701         if (!list_empty(last)) {
702                 qp = list_entry(last->next, struct ipath_qp, timerwait);
703                 if (--qp->s_rnr_timeout == 0) {
704                         do {
705                                 list_del_init(&qp->timerwait);
706                                 qp->timer_next = rnr;
707                                 rnr = qp;
708                                 atomic_inc(&qp->refcount);
709                                 if (list_empty(last))
710                                         break;
711                                 qp = list_entry(last->next, struct ipath_qp,
712                                                 timerwait);
713                         } while (qp->s_rnr_timeout == 0);
714                 }
715         }
716         /*
717          * We should only be in the started state if pma_sample_start != 0
718          */
719         if (dev->pma_sample_status == IB_PMA_SAMPLE_STATUS_STARTED &&
720             --dev->pma_sample_start == 0) {
721                 dev->pma_sample_status = IB_PMA_SAMPLE_STATUS_RUNNING;
722                 ipath_snapshot_counters(dev->dd, &dev->ipath_sword,
723                                         &dev->ipath_rword,
724                                         &dev->ipath_spkts,
725                                         &dev->ipath_rpkts,
726                                         &dev->ipath_xmit_wait);
727         }
728         if (dev->pma_sample_status == IB_PMA_SAMPLE_STATUS_RUNNING) {
729                 if (dev->pma_sample_interval == 0) {
730                         u64 ta, tb, tc, td, te;
731
732                         dev->pma_sample_status = IB_PMA_SAMPLE_STATUS_DONE;
733                         ipath_snapshot_counters(dev->dd, &ta, &tb,
734                                                 &tc, &td, &te);
735
736                         dev->ipath_sword = ta - dev->ipath_sword;
737                         dev->ipath_rword = tb - dev->ipath_rword;
738                         dev->ipath_spkts = tc - dev->ipath_spkts;
739                         dev->ipath_rpkts = td - dev->ipath_rpkts;
740                         dev->ipath_xmit_wait = te - dev->ipath_xmit_wait;
741                 }
742                 else
743                         dev->pma_sample_interval--;
744         }
745         spin_unlock_irqrestore(&dev->pending_lock, flags);
746
747         /* XXX What if timer fires again while this is running? */
748         while (resend != NULL) {
749                 qp = resend;
750                 resend = qp->timer_next;
751
752                 spin_lock_irqsave(&qp->s_lock, flags);
753                 if (qp->s_last != qp->s_tail &&
754                     ib_ipath_state_ops[qp->state] & IPATH_PROCESS_SEND_OK) {
755                         dev->n_timeouts++;
756                         ipath_restart_rc(qp, qp->s_last_psn + 1);
757                 }
758                 spin_unlock_irqrestore(&qp->s_lock, flags);
759
760                 /* Notify ipath_destroy_qp() if it is waiting. */
761                 if (atomic_dec_and_test(&qp->refcount))
762                         wake_up(&qp->wait);
763         }
764         while (rnr != NULL) {
765                 qp = rnr;
766                 rnr = qp->timer_next;
767
768                 spin_lock_irqsave(&qp->s_lock, flags);
769                 if (ib_ipath_state_ops[qp->state] & IPATH_PROCESS_SEND_OK)
770                         ipath_schedule_send(qp);
771                 spin_unlock_irqrestore(&qp->s_lock, flags);
772
773                 /* Notify ipath_destroy_qp() if it is waiting. */
774                 if (atomic_dec_and_test(&qp->refcount))
775                         wake_up(&qp->wait);
776         }
777 }
778
779 static void update_sge(struct ipath_sge_state *ss, u32 length)
780 {
781         struct ipath_sge *sge = &ss->sge;
782
783         sge->vaddr += length;
784         sge->length -= length;
785         sge->sge_length -= length;
786         if (sge->sge_length == 0) {
787                 if (--ss->num_sge)
788                         *sge = *ss->sg_list++;
789         } else if (sge->length == 0 && sge->mr != NULL) {
790                 if (++sge->n >= IPATH_SEGSZ) {
791                         if (++sge->m >= sge->mr->mapsz)
792                                 return;
793                         sge->n = 0;
794                 }
795                 sge->vaddr = sge->mr->map[sge->m]->segs[sge->n].vaddr;
796                 sge->length = sge->mr->map[sge->m]->segs[sge->n].length;
797         }
798 }
799
800 #ifdef __LITTLE_ENDIAN
801 static inline u32 get_upper_bits(u32 data, u32 shift)
802 {
803         return data >> shift;
804 }
805
806 static inline u32 set_upper_bits(u32 data, u32 shift)
807 {
808         return data << shift;
809 }
810
811 static inline u32 clear_upper_bytes(u32 data, u32 n, u32 off)
812 {
813         data <<= ((sizeof(u32) - n) * BITS_PER_BYTE);
814         data >>= ((sizeof(u32) - n - off) * BITS_PER_BYTE);
815         return data;
816 }
817 #else
818 static inline u32 get_upper_bits(u32 data, u32 shift)
819 {
820         return data << shift;
821 }
822
823 static inline u32 set_upper_bits(u32 data, u32 shift)
824 {
825         return data >> shift;
826 }
827
828 static inline u32 clear_upper_bytes(u32 data, u32 n, u32 off)
829 {
830         data >>= ((sizeof(u32) - n) * BITS_PER_BYTE);
831         data <<= ((sizeof(u32) - n - off) * BITS_PER_BYTE);
832         return data;
833 }
834 #endif
835
836 static void copy_io(u32 __iomem *piobuf, struct ipath_sge_state *ss,
837                     u32 length, unsigned flush_wc)
838 {
839         u32 extra = 0;
840         u32 data = 0;
841         u32 last;
842
843         while (1) {
844                 u32 len = ss->sge.length;
845                 u32 off;
846
847                 if (len > length)
848                         len = length;
849                 if (len > ss->sge.sge_length)
850                         len = ss->sge.sge_length;
851                 BUG_ON(len == 0);
852                 /* If the source address is not aligned, try to align it. */
853                 off = (unsigned long)ss->sge.vaddr & (sizeof(u32) - 1);
854                 if (off) {
855                         u32 *addr = (u32 *)((unsigned long)ss->sge.vaddr &
856                                             ~(sizeof(u32) - 1));
857                         u32 v = get_upper_bits(*addr, off * BITS_PER_BYTE);
858                         u32 y;
859
860                         y = sizeof(u32) - off;
861                         if (len > y)
862                                 len = y;
863                         if (len + extra >= sizeof(u32)) {
864                                 data |= set_upper_bits(v, extra *
865                                                        BITS_PER_BYTE);
866                                 len = sizeof(u32) - extra;
867                                 if (len == length) {
868                                         last = data;
869                                         break;
870                                 }
871                                 __raw_writel(data, piobuf);
872                                 piobuf++;
873                                 extra = 0;
874                                 data = 0;
875                         } else {
876                                 /* Clear unused upper bytes */
877                                 data |= clear_upper_bytes(v, len, extra);
878                                 if (len == length) {
879                                         last = data;
880                                         break;
881                                 }
882                                 extra += len;
883                         }
884                 } else if (extra) {
885                         /* Source address is aligned. */
886                         u32 *addr = (u32 *) ss->sge.vaddr;
887                         int shift = extra * BITS_PER_BYTE;
888                         int ushift = 32 - shift;
889                         u32 l = len;
890
891                         while (l >= sizeof(u32)) {
892                                 u32 v = *addr;
893
894                                 data |= set_upper_bits(v, shift);
895                                 __raw_writel(data, piobuf);
896                                 data = get_upper_bits(v, ushift);
897                                 piobuf++;
898                                 addr++;
899                                 l -= sizeof(u32);
900                         }
901                         /*
902                          * We still have 'extra' number of bytes leftover.
903                          */
904                         if (l) {
905                                 u32 v = *addr;
906
907                                 if (l + extra >= sizeof(u32)) {
908                                         data |= set_upper_bits(v, shift);
909                                         len -= l + extra - sizeof(u32);
910                                         if (len == length) {
911                                                 last = data;
912                                                 break;
913                                         }
914                                         __raw_writel(data, piobuf);
915                                         piobuf++;
916                                         extra = 0;
917                                         data = 0;
918                                 } else {
919                                         /* Clear unused upper bytes */
920                                         data |= clear_upper_bytes(v, l,
921                                                                   extra);
922                                         if (len == length) {
923                                                 last = data;
924                                                 break;
925                                         }
926                                         extra += l;
927                                 }
928                         } else if (len == length) {
929                                 last = data;
930                                 break;
931                         }
932                 } else if (len == length) {
933                         u32 w;
934
935                         /*
936                          * Need to round up for the last dword in the
937                          * packet.
938                          */
939                         w = (len + 3) >> 2;
940                         __iowrite32_copy(piobuf, ss->sge.vaddr, w - 1);
941                         piobuf += w - 1;
942                         last = ((u32 *) ss->sge.vaddr)[w - 1];
943                         break;
944                 } else {
945                         u32 w = len >> 2;
946
947                         __iowrite32_copy(piobuf, ss->sge.vaddr, w);
948                         piobuf += w;
949
950                         extra = len & (sizeof(u32) - 1);
951                         if (extra) {
952                                 u32 v = ((u32 *) ss->sge.vaddr)[w];
953
954                                 /* Clear unused upper bytes */
955                                 data = clear_upper_bytes(v, extra, 0);
956                         }
957                 }
958                 update_sge(ss, len);
959                 length -= len;
960         }
961         /* Update address before sending packet. */
962         update_sge(ss, length);
963         if (flush_wc) {
964                 /* must flush early everything before trigger word */
965                 ipath_flush_wc();
966                 __raw_writel(last, piobuf);
967                 /* be sure trigger word is written */
968                 ipath_flush_wc();
969         } else
970                 __raw_writel(last, piobuf);
971 }
972
973 /*
974  * Convert IB rate to delay multiplier.
975  */
976 unsigned ipath_ib_rate_to_mult(enum ib_rate rate)
977 {
978         switch (rate) {
979         case IB_RATE_2_5_GBPS: return 8;
980         case IB_RATE_5_GBPS:   return 4;
981         case IB_RATE_10_GBPS:  return 2;
982         case IB_RATE_20_GBPS:  return 1;
983         default:               return 0;
984         }
985 }
986
987 /*
988  * Convert delay multiplier to IB rate
989  */
990 static enum ib_rate ipath_mult_to_ib_rate(unsigned mult)
991 {
992         switch (mult) {
993         case 8:  return IB_RATE_2_5_GBPS;
994         case 4:  return IB_RATE_5_GBPS;
995         case 2:  return IB_RATE_10_GBPS;
996         case 1:  return IB_RATE_20_GBPS;
997         default: return IB_RATE_PORT_CURRENT;
998         }
999 }
1000
1001 static inline struct ipath_verbs_txreq *get_txreq(struct ipath_ibdev *dev)
1002 {
1003         struct ipath_verbs_txreq *tx = NULL;
1004         unsigned long flags;
1005
1006         spin_lock_irqsave(&dev->pending_lock, flags);
1007         if (!list_empty(&dev->txreq_free)) {
1008                 struct list_head *l = dev->txreq_free.next;
1009
1010                 list_del(l);
1011                 tx = list_entry(l, struct ipath_verbs_txreq, txreq.list);
1012         }
1013         spin_unlock_irqrestore(&dev->pending_lock, flags);
1014         return tx;
1015 }
1016
1017 static inline void put_txreq(struct ipath_ibdev *dev,
1018                              struct ipath_verbs_txreq *tx)
1019 {
1020         unsigned long flags;
1021
1022         spin_lock_irqsave(&dev->pending_lock, flags);
1023         list_add(&tx->txreq.list, &dev->txreq_free);
1024         spin_unlock_irqrestore(&dev->pending_lock, flags);
1025 }
1026
1027 static void sdma_complete(void *cookie, int status)
1028 {
1029         struct ipath_verbs_txreq *tx = cookie;
1030         struct ipath_qp *qp = tx->qp;
1031         struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
1032         unsigned long flags;
1033         enum ib_wc_status ibs = status == IPATH_SDMA_TXREQ_S_OK ?
1034                 IB_WC_SUCCESS : IB_WC_WR_FLUSH_ERR;
1035
1036         if (atomic_dec_and_test(&qp->s_dma_busy)) {
1037                 spin_lock_irqsave(&qp->s_lock, flags);
1038                 if (tx->wqe)
1039                         ipath_send_complete(qp, tx->wqe, ibs);
1040                 if ((ib_ipath_state_ops[qp->state] & IPATH_FLUSH_SEND &&
1041                      qp->s_last != qp->s_head) ||
1042                     (qp->s_flags & IPATH_S_WAIT_DMA))
1043                         ipath_schedule_send(qp);
1044                 spin_unlock_irqrestore(&qp->s_lock, flags);
1045                 wake_up(&qp->wait_dma);
1046         } else if (tx->wqe) {
1047                 spin_lock_irqsave(&qp->s_lock, flags);
1048                 ipath_send_complete(qp, tx->wqe, ibs);
1049                 spin_unlock_irqrestore(&qp->s_lock, flags);
1050         }
1051
1052         if (tx->txreq.flags & IPATH_SDMA_TXREQ_F_FREEBUF)
1053                 kfree(tx->txreq.map_addr);
1054         put_txreq(dev, tx);
1055
1056         if (atomic_dec_and_test(&qp->refcount))
1057                 wake_up(&qp->wait);
1058 }
1059
1060 static void decrement_dma_busy(struct ipath_qp *qp)
1061 {
1062         unsigned long flags;
1063
1064         if (atomic_dec_and_test(&qp->s_dma_busy)) {
1065                 spin_lock_irqsave(&qp->s_lock, flags);
1066                 if ((ib_ipath_state_ops[qp->state] & IPATH_FLUSH_SEND &&
1067                      qp->s_last != qp->s_head) ||
1068                     (qp->s_flags & IPATH_S_WAIT_DMA))
1069                         ipath_schedule_send(qp);
1070                 spin_unlock_irqrestore(&qp->s_lock, flags);
1071                 wake_up(&qp->wait_dma);
1072         }
1073 }
1074
1075 /*
1076  * Compute the number of clock cycles of delay before sending the next packet.
1077  * The multipliers reflect the number of clocks for the fastest rate so
1078  * one tick at 4xDDR is 8 ticks at 1xSDR.
1079  * If the destination port will take longer to receive a packet than
1080  * the outgoing link can send it, we need to delay sending the next packet
1081  * by the difference in time it takes the receiver to receive and the sender
1082  * to send this packet.
1083  * Note that this delay is always correct for UC and RC but not always
1084  * optimal for UD. For UD, the destination HCA can be different for each
1085  * packet, in which case, we could send packets to a different destination
1086  * while "waiting" for the delay. The overhead for doing this without
1087  * HW support is more than just paying the cost of delaying some packets
1088  * unnecessarily.
1089  */
1090 static inline unsigned ipath_pkt_delay(u32 plen, u8 snd_mult, u8 rcv_mult)
1091 {
1092         return (rcv_mult > snd_mult) ?
1093                 (plen * (rcv_mult - snd_mult) + 1) >> 1 : 0;
1094 }
1095
1096 static int ipath_verbs_send_dma(struct ipath_qp *qp,
1097                                 struct ipath_ib_header *hdr, u32 hdrwords,
1098                                 struct ipath_sge_state *ss, u32 len,
1099                                 u32 plen, u32 dwords)
1100 {
1101         struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
1102         struct ipath_devdata *dd = dev->dd;
1103         struct ipath_verbs_txreq *tx;
1104         u32 *piobuf;
1105         u32 control;
1106         u32 ndesc;
1107         int ret;
1108
1109         tx = qp->s_tx;
1110         if (tx) {
1111                 qp->s_tx = NULL;
1112                 /* resend previously constructed packet */
1113                 atomic_inc(&qp->s_dma_busy);
1114                 ret = ipath_sdma_verbs_send(dd, tx->ss, tx->len, tx);
1115                 if (ret) {
1116                         qp->s_tx = tx;
1117                         decrement_dma_busy(qp);
1118                 }
1119                 goto bail;
1120         }
1121
1122         tx = get_txreq(dev);
1123         if (!tx) {
1124                 ret = -EBUSY;
1125                 goto bail;
1126         }
1127
1128         /*
1129          * Get the saved delay count we computed for the previous packet
1130          * and save the delay count for this packet to be used next time
1131          * we get here.
1132          */
1133         control = qp->s_pkt_delay;
1134         qp->s_pkt_delay = ipath_pkt_delay(plen, dd->delay_mult, qp->s_dmult);
1135
1136         tx->qp = qp;
1137         atomic_inc(&qp->refcount);
1138         tx->wqe = qp->s_wqe;
1139         tx->txreq.callback = sdma_complete;
1140         tx->txreq.callback_cookie = tx;
1141         tx->txreq.flags = IPATH_SDMA_TXREQ_F_HEADTOHOST |
1142                 IPATH_SDMA_TXREQ_F_INTREQ | IPATH_SDMA_TXREQ_F_FREEDESC;
1143         if (plen + 1 >= IPATH_SMALLBUF_DWORDS)
1144                 tx->txreq.flags |= IPATH_SDMA_TXREQ_F_USELARGEBUF;
1145
1146         /* VL15 packets bypass credit check */
1147         if ((be16_to_cpu(hdr->lrh[0]) >> 12) == 15) {
1148                 control |= 1ULL << 31;
1149                 tx->txreq.flags |= IPATH_SDMA_TXREQ_F_VL15;
1150         }
1151
1152         if (len) {
1153                 /*
1154                  * Don't try to DMA if it takes more descriptors than
1155                  * the queue holds.
1156                  */
1157                 ndesc = ipath_count_sge(ss, len);
1158                 if (ndesc >= dd->ipath_sdma_descq_cnt)
1159                         ndesc = 0;
1160         } else
1161                 ndesc = 1;
1162         if (ndesc) {
1163                 tx->hdr.pbc[0] = cpu_to_le32(plen);
1164                 tx->hdr.pbc[1] = cpu_to_le32(control);
1165                 memcpy(&tx->hdr.hdr, hdr, hdrwords << 2);
1166                 tx->txreq.sg_count = ndesc;
1167                 tx->map_len = (hdrwords + 2) << 2;
1168                 tx->txreq.map_addr = &tx->hdr;
1169                 atomic_inc(&qp->s_dma_busy);
1170                 ret = ipath_sdma_verbs_send(dd, ss, dwords, tx);
1171                 if (ret) {
1172                         /* save ss and length in dwords */
1173                         tx->ss = ss;
1174                         tx->len = dwords;
1175                         qp->s_tx = tx;
1176                         decrement_dma_busy(qp);
1177                 }
1178                 goto bail;
1179         }
1180
1181         /* Allocate a buffer and copy the header and payload to it. */
1182         tx->map_len = (plen + 1) << 2;
1183         piobuf = kmalloc(tx->map_len, GFP_ATOMIC);
1184         if (unlikely(piobuf == NULL)) {
1185                 ret = -EBUSY;
1186                 goto err_tx;
1187         }
1188         tx->txreq.map_addr = piobuf;
1189         tx->txreq.flags |= IPATH_SDMA_TXREQ_F_FREEBUF;
1190         tx->txreq.sg_count = 1;
1191
1192         *piobuf++ = (__force u32) cpu_to_le32(plen);
1193         *piobuf++ = (__force u32) cpu_to_le32(control);
1194         memcpy(piobuf, hdr, hdrwords << 2);
1195         ipath_copy_from_sge(piobuf + hdrwords, ss, len);
1196
1197         atomic_inc(&qp->s_dma_busy);
1198         ret = ipath_sdma_verbs_send(dd, NULL, 0, tx);
1199         /*
1200          * If we couldn't queue the DMA request, save the info
1201          * and try again later rather than destroying the
1202          * buffer and undoing the side effects of the copy.
1203          */
1204         if (ret) {
1205                 tx->ss = NULL;
1206                 tx->len = 0;
1207                 qp->s_tx = tx;
1208                 decrement_dma_busy(qp);
1209         }
1210         dev->n_unaligned++;
1211         goto bail;
1212
1213 err_tx:
1214         if (atomic_dec_and_test(&qp->refcount))
1215                 wake_up(&qp->wait);
1216         put_txreq(dev, tx);
1217 bail:
1218         return ret;
1219 }
1220
1221 static int ipath_verbs_send_pio(struct ipath_qp *qp,
1222                                 struct ipath_ib_header *ibhdr, u32 hdrwords,
1223                                 struct ipath_sge_state *ss, u32 len,
1224                                 u32 plen, u32 dwords)
1225 {
1226         struct ipath_devdata *dd = to_idev(qp->ibqp.device)->dd;
1227         u32 *hdr = (u32 *) ibhdr;
1228         u32 __iomem *piobuf;
1229         unsigned flush_wc;
1230         u32 control;
1231         int ret;
1232         unsigned long flags;
1233
1234         piobuf = ipath_getpiobuf(dd, plen, NULL);
1235         if (unlikely(piobuf == NULL)) {
1236                 ret = -EBUSY;
1237                 goto bail;
1238         }
1239
1240         /*
1241          * Get the saved delay count we computed for the previous packet
1242          * and save the delay count for this packet to be used next time
1243          * we get here.
1244          */
1245         control = qp->s_pkt_delay;
1246         qp->s_pkt_delay = ipath_pkt_delay(plen, dd->delay_mult, qp->s_dmult);
1247
1248         /* VL15 packets bypass credit check */
1249         if ((be16_to_cpu(ibhdr->lrh[0]) >> 12) == 15)
1250                 control |= 1ULL << 31;
1251
1252         /*
1253          * Write the length to the control qword plus any needed flags.
1254          * We have to flush after the PBC for correctness on some cpus
1255          * or WC buffer can be written out of order.
1256          */
1257         writeq(((u64) control << 32) | plen, piobuf);
1258         piobuf += 2;
1259
1260         flush_wc = dd->ipath_flags & IPATH_PIO_FLUSH_WC;
1261         if (len == 0) {
1262                 /*
1263                  * If there is just the header portion, must flush before
1264                  * writing last word of header for correctness, and after
1265                  * the last header word (trigger word).
1266                  */
1267                 if (flush_wc) {
1268                         ipath_flush_wc();
1269                         __iowrite32_copy(piobuf, hdr, hdrwords - 1);
1270                         ipath_flush_wc();
1271                         __raw_writel(hdr[hdrwords - 1], piobuf + hdrwords - 1);
1272                         ipath_flush_wc();
1273                 } else
1274                         __iowrite32_copy(piobuf, hdr, hdrwords);
1275                 goto done;
1276         }
1277
1278         if (flush_wc)
1279                 ipath_flush_wc();
1280         __iowrite32_copy(piobuf, hdr, hdrwords);
1281         piobuf += hdrwords;
1282
1283         /* The common case is aligned and contained in one segment. */
1284         if (likely(ss->num_sge == 1 && len <= ss->sge.length &&
1285                    !((unsigned long)ss->sge.vaddr & (sizeof(u32) - 1)))) {
1286                 u32 *addr = (u32 *) ss->sge.vaddr;
1287
1288                 /* Update address before sending packet. */
1289                 update_sge(ss, len);
1290                 if (flush_wc) {
1291                         __iowrite32_copy(piobuf, addr, dwords - 1);
1292                         /* must flush early everything before trigger word */
1293                         ipath_flush_wc();
1294                         __raw_writel(addr[dwords - 1], piobuf + dwords - 1);
1295                         /* be sure trigger word is written */
1296                         ipath_flush_wc();
1297                 } else
1298                         __iowrite32_copy(piobuf, addr, dwords);
1299                 goto done;
1300         }
1301         copy_io(piobuf, ss, len, flush_wc);
1302 done:
1303         if (qp->s_wqe) {
1304                 spin_lock_irqsave(&qp->s_lock, flags);
1305                 ipath_send_complete(qp, qp->s_wqe, IB_WC_SUCCESS);
1306                 spin_unlock_irqrestore(&qp->s_lock, flags);
1307         }
1308         ret = 0;
1309 bail:
1310         return ret;
1311 }
1312
1313 /**
1314  * ipath_verbs_send - send a packet
1315  * @qp: the QP to send on
1316  * @hdr: the packet header
1317  * @hdrwords: the number of 32-bit words in the header
1318  * @ss: the SGE to send
1319  * @len: the length of the packet in bytes
1320  */
1321 int ipath_verbs_send(struct ipath_qp *qp, struct ipath_ib_header *hdr,
1322                      u32 hdrwords, struct ipath_sge_state *ss, u32 len)
1323 {
1324         struct ipath_devdata *dd = to_idev(qp->ibqp.device)->dd;
1325         u32 plen;
1326         int ret;
1327         u32 dwords = (len + 3) >> 2;
1328
1329         /*
1330          * Calculate the send buffer trigger address.
1331          * The +1 counts for the pbc control dword following the pbc length.
1332          */
1333         plen = hdrwords + dwords + 1;
1334
1335         /*
1336          * VL15 packets (IB_QPT_SMI) will always use PIO, so we
1337          * can defer SDMA restart until link goes ACTIVE without
1338          * worrying about just how we got there.
1339          */
1340         if (qp->ibqp.qp_type == IB_QPT_SMI ||
1341             !(dd->ipath_flags & IPATH_HAS_SEND_DMA))
1342                 ret = ipath_verbs_send_pio(qp, hdr, hdrwords, ss, len,
1343                                            plen, dwords);
1344         else
1345                 ret = ipath_verbs_send_dma(qp, hdr, hdrwords, ss, len,
1346                                            plen, dwords);
1347
1348         return ret;
1349 }
1350
1351 int ipath_snapshot_counters(struct ipath_devdata *dd, u64 *swords,
1352                             u64 *rwords, u64 *spkts, u64 *rpkts,
1353                             u64 *xmit_wait)
1354 {
1355         int ret;
1356
1357         if (!(dd->ipath_flags & IPATH_INITTED)) {
1358                 /* no hardware, freeze, etc. */
1359                 ret = -EINVAL;
1360                 goto bail;
1361         }
1362         *swords = ipath_snap_cntr(dd, dd->ipath_cregs->cr_wordsendcnt);
1363         *rwords = ipath_snap_cntr(dd, dd->ipath_cregs->cr_wordrcvcnt);
1364         *spkts = ipath_snap_cntr(dd, dd->ipath_cregs->cr_pktsendcnt);
1365         *rpkts = ipath_snap_cntr(dd, dd->ipath_cregs->cr_pktrcvcnt);
1366         *xmit_wait = ipath_snap_cntr(dd, dd->ipath_cregs->cr_sendstallcnt);
1367
1368         ret = 0;
1369
1370 bail:
1371         return ret;
1372 }
1373
1374 /**
1375  * ipath_get_counters - get various chip counters
1376  * @dd: the infinipath device
1377  * @cntrs: counters are placed here
1378  *
1379  * Return the counters needed by recv_pma_get_portcounters().
1380  */
1381 int ipath_get_counters(struct ipath_devdata *dd,
1382                        struct ipath_verbs_counters *cntrs)
1383 {
1384         struct ipath_cregs const *crp = dd->ipath_cregs;
1385         int ret;
1386
1387         if (!(dd->ipath_flags & IPATH_INITTED)) {
1388                 /* no hardware, freeze, etc. */
1389                 ret = -EINVAL;
1390                 goto bail;
1391         }
1392         cntrs->symbol_error_counter =
1393                 ipath_snap_cntr(dd, crp->cr_ibsymbolerrcnt);
1394         cntrs->link_error_recovery_counter =
1395                 ipath_snap_cntr(dd, crp->cr_iblinkerrrecovcnt);
1396         /*
1397          * The link downed counter counts when the other side downs the
1398          * connection.  We add in the number of times we downed the link
1399          * due to local link integrity errors to compensate.
1400          */
1401         cntrs->link_downed_counter =
1402                 ipath_snap_cntr(dd, crp->cr_iblinkdowncnt);
1403         cntrs->port_rcv_errors =
1404                 ipath_snap_cntr(dd, crp->cr_rxdroppktcnt) +
1405                 ipath_snap_cntr(dd, crp->cr_rcvovflcnt) +
1406                 ipath_snap_cntr(dd, crp->cr_portovflcnt) +
1407                 ipath_snap_cntr(dd, crp->cr_err_rlencnt) +
1408                 ipath_snap_cntr(dd, crp->cr_invalidrlencnt) +
1409                 ipath_snap_cntr(dd, crp->cr_errlinkcnt) +
1410                 ipath_snap_cntr(dd, crp->cr_erricrccnt) +
1411                 ipath_snap_cntr(dd, crp->cr_errvcrccnt) +
1412                 ipath_snap_cntr(dd, crp->cr_errlpcrccnt) +
1413                 ipath_snap_cntr(dd, crp->cr_badformatcnt) +
1414                 dd->ipath_rxfc_unsupvl_errs;
1415         if (crp->cr_rxotherlocalphyerrcnt)
1416                 cntrs->port_rcv_errors +=
1417                         ipath_snap_cntr(dd, crp->cr_rxotherlocalphyerrcnt);
1418         if (crp->cr_rxvlerrcnt)
1419                 cntrs->port_rcv_errors +=
1420                         ipath_snap_cntr(dd, crp->cr_rxvlerrcnt);
1421         cntrs->port_rcv_remphys_errors =
1422                 ipath_snap_cntr(dd, crp->cr_rcvebpcnt);
1423         cntrs->port_xmit_discards = ipath_snap_cntr(dd, crp->cr_unsupvlcnt);
1424         cntrs->port_xmit_data = ipath_snap_cntr(dd, crp->cr_wordsendcnt);
1425         cntrs->port_rcv_data = ipath_snap_cntr(dd, crp->cr_wordrcvcnt);
1426         cntrs->port_xmit_packets = ipath_snap_cntr(dd, crp->cr_pktsendcnt);
1427         cntrs->port_rcv_packets = ipath_snap_cntr(dd, crp->cr_pktrcvcnt);
1428         cntrs->local_link_integrity_errors =
1429                 crp->cr_locallinkintegrityerrcnt ?
1430                 ipath_snap_cntr(dd, crp->cr_locallinkintegrityerrcnt) :
1431                 ((dd->ipath_flags & IPATH_GPIO_ERRINTRS) ?
1432                  dd->ipath_lli_errs : dd->ipath_lli_errors);
1433         cntrs->excessive_buffer_overrun_errors =
1434                 crp->cr_excessbufferovflcnt ?
1435                 ipath_snap_cntr(dd, crp->cr_excessbufferovflcnt) :
1436                 dd->ipath_overrun_thresh_errs;
1437         cntrs->vl15_dropped = crp->cr_vl15droppedpktcnt ?
1438                 ipath_snap_cntr(dd, crp->cr_vl15droppedpktcnt) : 0;
1439
1440         ret = 0;
1441
1442 bail:
1443         return ret;
1444 }
1445
1446 /**
1447  * ipath_ib_piobufavail - callback when a PIO buffer is available
1448  * @arg: the device pointer
1449  *
1450  * This is called from ipath_intr() at interrupt level when a PIO buffer is
1451  * available after ipath_verbs_send() returned an error that no buffers were
1452  * available.  Return 1 if we consumed all the PIO buffers and we still have
1453  * QPs waiting for buffers (for now, just restart the send tasklet and
1454  * return zero).
1455  */
1456 int ipath_ib_piobufavail(struct ipath_ibdev *dev)
1457 {
1458         struct list_head *list;
1459         struct ipath_qp *qplist;
1460         struct ipath_qp *qp;
1461         unsigned long flags;
1462
1463         if (dev == NULL)
1464                 goto bail;
1465
1466         list = &dev->piowait;
1467         qplist = NULL;
1468
1469         spin_lock_irqsave(&dev->pending_lock, flags);
1470         while (!list_empty(list)) {
1471                 qp = list_entry(list->next, struct ipath_qp, piowait);
1472                 list_del_init(&qp->piowait);
1473                 qp->pio_next = qplist;
1474                 qplist = qp;
1475                 atomic_inc(&qp->refcount);
1476         }
1477         spin_unlock_irqrestore(&dev->pending_lock, flags);
1478
1479         while (qplist != NULL) {
1480                 qp = qplist;
1481                 qplist = qp->pio_next;
1482
1483                 spin_lock_irqsave(&qp->s_lock, flags);
1484                 if (ib_ipath_state_ops[qp->state] & IPATH_PROCESS_SEND_OK)
1485                         ipath_schedule_send(qp);
1486                 spin_unlock_irqrestore(&qp->s_lock, flags);
1487
1488                 /* Notify ipath_destroy_qp() if it is waiting. */
1489                 if (atomic_dec_and_test(&qp->refcount))
1490                         wake_up(&qp->wait);
1491         }
1492
1493 bail:
1494         return 0;
1495 }
1496
1497 static int ipath_query_device(struct ib_device *ibdev,
1498                               struct ib_device_attr *props)
1499 {
1500         struct ipath_ibdev *dev = to_idev(ibdev);
1501
1502         memset(props, 0, sizeof(*props));
1503
1504         props->device_cap_flags = IB_DEVICE_BAD_PKEY_CNTR |
1505                 IB_DEVICE_BAD_QKEY_CNTR | IB_DEVICE_SHUTDOWN_PORT |
1506                 IB_DEVICE_SYS_IMAGE_GUID | IB_DEVICE_RC_RNR_NAK_GEN |
1507                 IB_DEVICE_PORT_ACTIVE_EVENT | IB_DEVICE_SRQ_RESIZE;
1508         props->page_size_cap = PAGE_SIZE;
1509         props->vendor_id =
1510                 IPATH_SRC_OUI_1 << 16 | IPATH_SRC_OUI_2 << 8 | IPATH_SRC_OUI_3;
1511         props->vendor_part_id = dev->dd->ipath_deviceid;
1512         props->hw_ver = dev->dd->ipath_pcirev;
1513
1514         props->sys_image_guid = dev->sys_image_guid;
1515
1516         props->max_mr_size = ~0ull;
1517         props->max_qp = ib_ipath_max_qps;
1518         props->max_qp_wr = ib_ipath_max_qp_wrs;
1519         props->max_sge = ib_ipath_max_sges;
1520         props->max_cq = ib_ipath_max_cqs;
1521         props->max_ah = ib_ipath_max_ahs;
1522         props->max_cqe = ib_ipath_max_cqes;
1523         props->max_mr = dev->lk_table.max;
1524         props->max_fmr = dev->lk_table.max;
1525         props->max_map_per_fmr = 32767;
1526         props->max_pd = ib_ipath_max_pds;
1527         props->max_qp_rd_atom = IPATH_MAX_RDMA_ATOMIC;
1528         props->max_qp_init_rd_atom = 255;
1529         /* props->max_res_rd_atom */
1530         props->max_srq = ib_ipath_max_srqs;
1531         props->max_srq_wr = ib_ipath_max_srq_wrs;
1532         props->max_srq_sge = ib_ipath_max_srq_sges;
1533         /* props->local_ca_ack_delay */
1534         props->atomic_cap = IB_ATOMIC_GLOB;
1535         props->max_pkeys = ipath_get_npkeys(dev->dd);
1536         props->max_mcast_grp = ib_ipath_max_mcast_grps;
1537         props->max_mcast_qp_attach = ib_ipath_max_mcast_qp_attached;
1538         props->max_total_mcast_qp_attach = props->max_mcast_qp_attach *
1539                 props->max_mcast_grp;
1540
1541         return 0;
1542 }
1543
1544 const u8 ipath_cvt_physportstate[32] = {
1545         [INFINIPATH_IBCS_LT_STATE_DISABLED] = IB_PHYSPORTSTATE_DISABLED,
1546         [INFINIPATH_IBCS_LT_STATE_LINKUP] = IB_PHYSPORTSTATE_LINKUP,
1547         [INFINIPATH_IBCS_LT_STATE_POLLACTIVE] = IB_PHYSPORTSTATE_POLL,
1548         [INFINIPATH_IBCS_LT_STATE_POLLQUIET] = IB_PHYSPORTSTATE_POLL,
1549         [INFINIPATH_IBCS_LT_STATE_SLEEPDELAY] = IB_PHYSPORTSTATE_SLEEP,
1550         [INFINIPATH_IBCS_LT_STATE_SLEEPQUIET] = IB_PHYSPORTSTATE_SLEEP,
1551         [INFINIPATH_IBCS_LT_STATE_CFGDEBOUNCE] =
1552                 IB_PHYSPORTSTATE_CFG_TRAIN,
1553         [INFINIPATH_IBCS_LT_STATE_CFGRCVFCFG] =
1554                 IB_PHYSPORTSTATE_CFG_TRAIN,
1555         [INFINIPATH_IBCS_LT_STATE_CFGWAITRMT] =
1556                 IB_PHYSPORTSTATE_CFG_TRAIN,
1557         [INFINIPATH_IBCS_LT_STATE_CFGIDLE] = IB_PHYSPORTSTATE_CFG_TRAIN,
1558         [INFINIPATH_IBCS_LT_STATE_RECOVERRETRAIN] =
1559                 IB_PHYSPORTSTATE_LINK_ERR_RECOVER,
1560         [INFINIPATH_IBCS_LT_STATE_RECOVERWAITRMT] =
1561                 IB_PHYSPORTSTATE_LINK_ERR_RECOVER,
1562         [INFINIPATH_IBCS_LT_STATE_RECOVERIDLE] =
1563                 IB_PHYSPORTSTATE_LINK_ERR_RECOVER,
1564         [0x10] = IB_PHYSPORTSTATE_CFG_TRAIN,
1565         [0x11] = IB_PHYSPORTSTATE_CFG_TRAIN,
1566         [0x12] = IB_PHYSPORTSTATE_CFG_TRAIN,
1567         [0x13] = IB_PHYSPORTSTATE_CFG_TRAIN,
1568         [0x14] = IB_PHYSPORTSTATE_CFG_TRAIN,
1569         [0x15] = IB_PHYSPORTSTATE_CFG_TRAIN,
1570         [0x16] = IB_PHYSPORTSTATE_CFG_TRAIN,
1571         [0x17] = IB_PHYSPORTSTATE_CFG_TRAIN
1572 };
1573
1574 u32 ipath_get_cr_errpkey(struct ipath_devdata *dd)
1575 {
1576         return ipath_read_creg32(dd, dd->ipath_cregs->cr_errpkey);
1577 }
1578
1579 static int ipath_query_port(struct ib_device *ibdev,
1580                             u8 port, struct ib_port_attr *props)
1581 {
1582         struct ipath_ibdev *dev = to_idev(ibdev);
1583         struct ipath_devdata *dd = dev->dd;
1584         enum ib_mtu mtu;
1585         u16 lid = dd->ipath_lid;
1586         u64 ibcstat;
1587
1588         memset(props, 0, sizeof(*props));
1589         props->lid = lid ? lid : be16_to_cpu(IB_LID_PERMISSIVE);
1590         props->lmc = dd->ipath_lmc;
1591         props->sm_lid = dev->sm_lid;
1592         props->sm_sl = dev->sm_sl;
1593         ibcstat = dd->ipath_lastibcstat;
1594         /* map LinkState to IB portinfo values.  */
1595         props->state = ipath_ib_linkstate(dd, ibcstat) + 1;
1596
1597         /* See phys_state_show() */
1598         props->phys_state = /* MEA: assumes shift == 0 */
1599                 ipath_cvt_physportstate[dd->ipath_lastibcstat &
1600                 dd->ibcs_lts_mask];
1601         props->port_cap_flags = dev->port_cap_flags;
1602         props->gid_tbl_len = 1;
1603         props->max_msg_sz = 0x80000000;
1604         props->pkey_tbl_len = ipath_get_npkeys(dd);
1605         props->bad_pkey_cntr = ipath_get_cr_errpkey(dd) -
1606                 dev->z_pkey_violations;
1607         props->qkey_viol_cntr = dev->qkey_violations;
1608         props->active_width = dd->ipath_link_width_active;
1609         /* See rate_show() */
1610         props->active_speed = dd->ipath_link_speed_active;
1611         props->max_vl_num = 1;          /* VLCap = VL0 */
1612         props->init_type_reply = 0;
1613
1614         props->max_mtu = ipath_mtu4096 ? IB_MTU_4096 : IB_MTU_2048;
1615         switch (dd->ipath_ibmtu) {
1616         case 4096:
1617                 mtu = IB_MTU_4096;
1618                 break;
1619         case 2048:
1620                 mtu = IB_MTU_2048;
1621                 break;
1622         case 1024:
1623                 mtu = IB_MTU_1024;
1624                 break;
1625         case 512:
1626                 mtu = IB_MTU_512;
1627                 break;
1628         case 256:
1629                 mtu = IB_MTU_256;
1630                 break;
1631         default:
1632                 mtu = IB_MTU_2048;
1633         }
1634         props->active_mtu = mtu;
1635         props->subnet_timeout = dev->subnet_timeout;
1636
1637         return 0;
1638 }
1639
1640 static int ipath_modify_device(struct ib_device *device,
1641                                int device_modify_mask,
1642                                struct ib_device_modify *device_modify)
1643 {
1644         int ret;
1645
1646         if (device_modify_mask & ~(IB_DEVICE_MODIFY_SYS_IMAGE_GUID |
1647                                    IB_DEVICE_MODIFY_NODE_DESC)) {
1648                 ret = -EOPNOTSUPP;
1649                 goto bail;
1650         }
1651
1652         if (device_modify_mask & IB_DEVICE_MODIFY_NODE_DESC)
1653                 memcpy(device->node_desc, device_modify->node_desc, 64);
1654
1655         if (device_modify_mask & IB_DEVICE_MODIFY_SYS_IMAGE_GUID)
1656                 to_idev(device)->sys_image_guid =
1657                         cpu_to_be64(device_modify->sys_image_guid);
1658
1659         ret = 0;
1660
1661 bail:
1662         return ret;
1663 }
1664
1665 static int ipath_modify_port(struct ib_device *ibdev,
1666                              u8 port, int port_modify_mask,
1667                              struct ib_port_modify *props)
1668 {
1669         struct ipath_ibdev *dev = to_idev(ibdev);
1670
1671         dev->port_cap_flags |= props->set_port_cap_mask;
1672         dev->port_cap_flags &= ~props->clr_port_cap_mask;
1673         if (port_modify_mask & IB_PORT_SHUTDOWN)
1674                 ipath_set_linkstate(dev->dd, IPATH_IB_LINKDOWN);
1675         if (port_modify_mask & IB_PORT_RESET_QKEY_CNTR)
1676                 dev->qkey_violations = 0;
1677         return 0;
1678 }
1679
1680 static int ipath_query_gid(struct ib_device *ibdev, u8 port,
1681                            int index, union ib_gid *gid)
1682 {
1683         struct ipath_ibdev *dev = to_idev(ibdev);
1684         int ret;
1685
1686         if (index >= 1) {
1687                 ret = -EINVAL;
1688                 goto bail;
1689         }
1690         gid->global.subnet_prefix = dev->gid_prefix;
1691         gid->global.interface_id = dev->dd->ipath_guid;
1692
1693         ret = 0;
1694
1695 bail:
1696         return ret;
1697 }
1698
1699 static struct ib_pd *ipath_alloc_pd(struct ib_device *ibdev,
1700                                     struct ib_ucontext *context,
1701                                     struct ib_udata *udata)
1702 {
1703         struct ipath_ibdev *dev = to_idev(ibdev);
1704         struct ipath_pd *pd;
1705         struct ib_pd *ret;
1706
1707         /*
1708          * This is actually totally arbitrary.  Some correctness tests
1709          * assume there's a maximum number of PDs that can be allocated.
1710          * We don't actually have this limit, but we fail the test if
1711          * we allow allocations of more than we report for this value.
1712          */
1713
1714         pd = kmalloc(sizeof *pd, GFP_KERNEL);
1715         if (!pd) {
1716                 ret = ERR_PTR(-ENOMEM);
1717                 goto bail;
1718         }
1719
1720         spin_lock(&dev->n_pds_lock);
1721         if (dev->n_pds_allocated == ib_ipath_max_pds) {
1722                 spin_unlock(&dev->n_pds_lock);
1723                 kfree(pd);
1724                 ret = ERR_PTR(-ENOMEM);
1725                 goto bail;
1726         }
1727
1728         dev->n_pds_allocated++;
1729         spin_unlock(&dev->n_pds_lock);
1730
1731         /* ib_alloc_pd() will initialize pd->ibpd. */
1732         pd->user = udata != NULL;
1733
1734         ret = &pd->ibpd;
1735
1736 bail:
1737         return ret;
1738 }
1739
1740 static int ipath_dealloc_pd(struct ib_pd *ibpd)
1741 {
1742         struct ipath_pd *pd = to_ipd(ibpd);
1743         struct ipath_ibdev *dev = to_idev(ibpd->device);
1744
1745         spin_lock(&dev->n_pds_lock);
1746         dev->n_pds_allocated--;
1747         spin_unlock(&dev->n_pds_lock);
1748
1749         kfree(pd);
1750
1751         return 0;
1752 }
1753
1754 /**
1755  * ipath_create_ah - create an address handle
1756  * @pd: the protection domain
1757  * @ah_attr: the attributes of the AH
1758  *
1759  * This may be called from interrupt context.
1760  */
1761 static struct ib_ah *ipath_create_ah(struct ib_pd *pd,
1762                                      struct ib_ah_attr *ah_attr)
1763 {
1764         struct ipath_ah *ah;
1765         struct ib_ah *ret;
1766         struct ipath_ibdev *dev = to_idev(pd->device);
1767         unsigned long flags;
1768
1769         /* A multicast address requires a GRH (see ch. 8.4.1). */
1770         if (ah_attr->dlid >= IPATH_MULTICAST_LID_BASE &&
1771             ah_attr->dlid != IPATH_PERMISSIVE_LID &&
1772             !(ah_attr->ah_flags & IB_AH_GRH)) {
1773                 ret = ERR_PTR(-EINVAL);
1774                 goto bail;
1775         }
1776
1777         if (ah_attr->dlid == 0) {
1778                 ret = ERR_PTR(-EINVAL);
1779                 goto bail;
1780         }
1781
1782         if (ah_attr->port_num < 1 ||
1783             ah_attr->port_num > pd->device->phys_port_cnt) {
1784                 ret = ERR_PTR(-EINVAL);
1785                 goto bail;
1786         }
1787
1788         ah = kmalloc(sizeof *ah, GFP_ATOMIC);
1789         if (!ah) {
1790                 ret = ERR_PTR(-ENOMEM);
1791                 goto bail;
1792         }
1793
1794         spin_lock_irqsave(&dev->n_ahs_lock, flags);
1795         if (dev->n_ahs_allocated == ib_ipath_max_ahs) {
1796                 spin_unlock_irqrestore(&dev->n_ahs_lock, flags);
1797                 kfree(ah);
1798                 ret = ERR_PTR(-ENOMEM);
1799                 goto bail;
1800         }
1801
1802         dev->n_ahs_allocated++;
1803         spin_unlock_irqrestore(&dev->n_ahs_lock, flags);
1804
1805         /* ib_create_ah() will initialize ah->ibah. */
1806         ah->attr = *ah_attr;
1807         ah->attr.static_rate = ipath_ib_rate_to_mult(ah_attr->static_rate);
1808
1809         ret = &ah->ibah;
1810
1811 bail:
1812         return ret;
1813 }
1814
1815 /**
1816  * ipath_destroy_ah - destroy an address handle
1817  * @ibah: the AH to destroy
1818  *
1819  * This may be called from interrupt context.
1820  */
1821 static int ipath_destroy_ah(struct ib_ah *ibah)
1822 {
1823         struct ipath_ibdev *dev = to_idev(ibah->device);
1824         struct ipath_ah *ah = to_iah(ibah);
1825         unsigned long flags;
1826
1827         spin_lock_irqsave(&dev->n_ahs_lock, flags);
1828         dev->n_ahs_allocated--;
1829         spin_unlock_irqrestore(&dev->n_ahs_lock, flags);
1830
1831         kfree(ah);
1832
1833         return 0;
1834 }
1835
1836 static int ipath_query_ah(struct ib_ah *ibah, struct ib_ah_attr *ah_attr)
1837 {
1838         struct ipath_ah *ah = to_iah(ibah);
1839
1840         *ah_attr = ah->attr;
1841         ah_attr->static_rate = ipath_mult_to_ib_rate(ah->attr.static_rate);
1842
1843         return 0;
1844 }
1845
1846 /**
1847  * ipath_get_npkeys - return the size of the PKEY table for port 0
1848  * @dd: the infinipath device
1849  */
1850 unsigned ipath_get_npkeys(struct ipath_devdata *dd)
1851 {
1852         return ARRAY_SIZE(dd->ipath_pd[0]->port_pkeys);
1853 }
1854
1855 /**
1856  * ipath_get_pkey - return the indexed PKEY from the port PKEY table
1857  * @dd: the infinipath device
1858  * @index: the PKEY index
1859  */
1860 unsigned ipath_get_pkey(struct ipath_devdata *dd, unsigned index)
1861 {
1862         unsigned ret;
1863
1864         /* always a kernel port, no locking needed */
1865         if (index >= ARRAY_SIZE(dd->ipath_pd[0]->port_pkeys))
1866                 ret = 0;
1867         else
1868                 ret = dd->ipath_pd[0]->port_pkeys[index];
1869
1870         return ret;
1871 }
1872
1873 static int ipath_query_pkey(struct ib_device *ibdev, u8 port, u16 index,
1874                             u16 *pkey)
1875 {
1876         struct ipath_ibdev *dev = to_idev(ibdev);
1877         int ret;
1878
1879         if (index >= ipath_get_npkeys(dev->dd)) {
1880                 ret = -EINVAL;
1881                 goto bail;
1882         }
1883
1884         *pkey = ipath_get_pkey(dev->dd, index);
1885         ret = 0;
1886
1887 bail:
1888         return ret;
1889 }
1890
1891 /**
1892  * ipath_alloc_ucontext - allocate a ucontest
1893  * @ibdev: the infiniband device
1894  * @udata: not used by the InfiniPath driver
1895  */
1896
1897 static struct ib_ucontext *ipath_alloc_ucontext(struct ib_device *ibdev,
1898                                                 struct ib_udata *udata)
1899 {
1900         struct ipath_ucontext *context;
1901         struct ib_ucontext *ret;
1902
1903         context = kmalloc(sizeof *context, GFP_KERNEL);
1904         if (!context) {
1905                 ret = ERR_PTR(-ENOMEM);
1906                 goto bail;
1907         }
1908
1909         ret = &context->ibucontext;
1910
1911 bail:
1912         return ret;
1913 }
1914
1915 static int ipath_dealloc_ucontext(struct ib_ucontext *context)
1916 {
1917         kfree(to_iucontext(context));
1918         return 0;
1919 }
1920
1921 static int ipath_verbs_register_sysfs(struct ib_device *dev);
1922
1923 static void __verbs_timer(unsigned long arg)
1924 {
1925         struct ipath_devdata *dd = (struct ipath_devdata *) arg;
1926
1927         /* Handle verbs layer timeouts. */
1928         ipath_ib_timer(dd->verbs_dev);
1929
1930         mod_timer(&dd->verbs_timer, jiffies + 1);
1931 }
1932
1933 static int enable_timer(struct ipath_devdata *dd)
1934 {
1935         /*
1936          * Early chips had a design flaw where the chip and kernel idea
1937          * of the tail register don't always agree, and therefore we won't
1938          * get an interrupt on the next packet received.
1939          * If the board supports per packet receive interrupts, use it.
1940          * Otherwise, the timer function periodically checks for packets
1941          * to cover this case.
1942          * Either way, the timer is needed for verbs layer related
1943          * processing.
1944          */
1945         if (dd->ipath_flags & IPATH_GPIO_INTR) {
1946                 ipath_write_kreg(dd, dd->ipath_kregs->kr_debugportselect,
1947                                  0x2074076542310ULL);
1948                 /* Enable GPIO bit 2 interrupt */
1949                 dd->ipath_gpio_mask |= (u64) (1 << IPATH_GPIO_PORT0_BIT);
1950                 ipath_write_kreg(dd, dd->ipath_kregs->kr_gpio_mask,
1951                                  dd->ipath_gpio_mask);
1952         }
1953
1954         init_timer(&dd->verbs_timer);
1955         dd->verbs_timer.function = __verbs_timer;
1956         dd->verbs_timer.data = (unsigned long)dd;
1957         dd->verbs_timer.expires = jiffies + 1;
1958         add_timer(&dd->verbs_timer);
1959
1960         return 0;
1961 }
1962
1963 static int disable_timer(struct ipath_devdata *dd)
1964 {
1965         /* Disable GPIO bit 2 interrupt */
1966         if (dd->ipath_flags & IPATH_GPIO_INTR) {
1967                 /* Disable GPIO bit 2 interrupt */
1968                 dd->ipath_gpio_mask &= ~((u64) (1 << IPATH_GPIO_PORT0_BIT));
1969                 ipath_write_kreg(dd, dd->ipath_kregs->kr_gpio_mask,
1970                                  dd->ipath_gpio_mask);
1971                 /*
1972                  * We might want to undo changes to debugportselect,
1973                  * but how?
1974                  */
1975         }
1976
1977         del_timer_sync(&dd->verbs_timer);
1978
1979         return 0;
1980 }
1981
1982 /**
1983  * ipath_register_ib_device - register our device with the infiniband core
1984  * @dd: the device data structure
1985  * Return the allocated ipath_ibdev pointer or NULL on error.
1986  */
1987 int ipath_register_ib_device(struct ipath_devdata *dd)
1988 {
1989         struct ipath_verbs_counters cntrs;
1990         struct ipath_ibdev *idev;
1991         struct ib_device *dev;
1992         struct ipath_verbs_txreq *tx;
1993         unsigned i;
1994         int ret;
1995
1996         idev = (struct ipath_ibdev *)ib_alloc_device(sizeof *idev);
1997         if (idev == NULL) {
1998                 ret = -ENOMEM;
1999                 goto bail;
2000         }
2001
2002         dev = &idev->ibdev;
2003
2004         if (dd->ipath_sdma_descq_cnt) {
2005                 tx = kmalloc(dd->ipath_sdma_descq_cnt * sizeof *tx,
2006                              GFP_KERNEL);
2007                 if (tx == NULL) {
2008                         ret = -ENOMEM;
2009                         goto err_tx;
2010                 }
2011         } else
2012                 tx = NULL;
2013         idev->txreq_bufs = tx;
2014
2015         /* Only need to initialize non-zero fields. */
2016         spin_lock_init(&idev->n_pds_lock);
2017         spin_lock_init(&idev->n_ahs_lock);
2018         spin_lock_init(&idev->n_cqs_lock);
2019         spin_lock_init(&idev->n_qps_lock);
2020         spin_lock_init(&idev->n_srqs_lock);
2021         spin_lock_init(&idev->n_mcast_grps_lock);
2022
2023         spin_lock_init(&idev->qp_table.lock);
2024         spin_lock_init(&idev->lk_table.lock);
2025         idev->sm_lid = __constant_be16_to_cpu(IB_LID_PERMISSIVE);
2026         /* Set the prefix to the default value (see ch. 4.1.1) */
2027         idev->gid_prefix = __constant_cpu_to_be64(0xfe80000000000000ULL);
2028
2029         ret = ipath_init_qp_table(idev, ib_ipath_qp_table_size);
2030         if (ret)
2031                 goto err_qp;
2032
2033         /*
2034          * The top ib_ipath_lkey_table_size bits are used to index the
2035          * table.  The lower 8 bits can be owned by the user (copied from
2036          * the LKEY).  The remaining bits act as a generation number or tag.
2037          */
2038         idev->lk_table.max = 1 << ib_ipath_lkey_table_size;
2039         idev->lk_table.table = kzalloc(idev->lk_table.max *
2040                                        sizeof(*idev->lk_table.table),
2041                                        GFP_KERNEL);
2042         if (idev->lk_table.table == NULL) {
2043                 ret = -ENOMEM;
2044                 goto err_lk;
2045         }
2046         INIT_LIST_HEAD(&idev->pending_mmaps);
2047         spin_lock_init(&idev->pending_lock);
2048         idev->mmap_offset = PAGE_SIZE;
2049         spin_lock_init(&idev->mmap_offset_lock);
2050         INIT_LIST_HEAD(&idev->pending[0]);
2051         INIT_LIST_HEAD(&idev->pending[1]);
2052         INIT_LIST_HEAD(&idev->pending[2]);
2053         INIT_LIST_HEAD(&idev->piowait);
2054         INIT_LIST_HEAD(&idev->rnrwait);
2055         INIT_LIST_HEAD(&idev->txreq_free);
2056         idev->pending_index = 0;
2057         idev->port_cap_flags =
2058                 IB_PORT_SYS_IMAGE_GUID_SUP | IB_PORT_CLIENT_REG_SUP;
2059         if (dd->ipath_flags & IPATH_HAS_LINK_LATENCY)
2060                 idev->port_cap_flags |= IB_PORT_LINK_LATENCY_SUP;
2061         idev->pma_counter_select[0] = IB_PMA_PORT_XMIT_DATA;
2062         idev->pma_counter_select[1] = IB_PMA_PORT_RCV_DATA;
2063         idev->pma_counter_select[2] = IB_PMA_PORT_XMIT_PKTS;
2064         idev->pma_counter_select[3] = IB_PMA_PORT_RCV_PKTS;
2065         idev->pma_counter_select[4] = IB_PMA_PORT_XMIT_WAIT;
2066
2067         /* Snapshot current HW counters to "clear" them. */
2068         ipath_get_counters(dd, &cntrs);
2069         idev->z_symbol_error_counter = cntrs.symbol_error_counter;
2070         idev->z_link_error_recovery_counter =
2071                 cntrs.link_error_recovery_counter;
2072         idev->z_link_downed_counter = cntrs.link_downed_counter;
2073         idev->z_port_rcv_errors = cntrs.port_rcv_errors;
2074         idev->z_port_rcv_remphys_errors =
2075                 cntrs.port_rcv_remphys_errors;
2076         idev->z_port_xmit_discards = cntrs.port_xmit_discards;
2077         idev->z_port_xmit_data = cntrs.port_xmit_data;
2078         idev->z_port_rcv_data = cntrs.port_rcv_data;
2079         idev->z_port_xmit_packets = cntrs.port_xmit_packets;
2080         idev->z_port_rcv_packets = cntrs.port_rcv_packets;
2081         idev->z_local_link_integrity_errors =
2082                 cntrs.local_link_integrity_errors;
2083         idev->z_excessive_buffer_overrun_errors =
2084                 cntrs.excessive_buffer_overrun_errors;
2085         idev->z_vl15_dropped = cntrs.vl15_dropped;
2086
2087         for (i = 0; i < dd->ipath_sdma_descq_cnt; i++, tx++)
2088                 list_add(&tx->txreq.list, &idev->txreq_free);
2089
2090         /*
2091          * The system image GUID is supposed to be the same for all
2092          * IB HCAs in a single system but since there can be other
2093          * device types in the system, we can't be sure this is unique.
2094          */
2095         if (!sys_image_guid)
2096                 sys_image_guid = dd->ipath_guid;
2097         idev->sys_image_guid = sys_image_guid;
2098         idev->ib_unit = dd->ipath_unit;
2099         idev->dd = dd;
2100
2101         strlcpy(dev->name, "ipath%d", IB_DEVICE_NAME_MAX);
2102         dev->owner = THIS_MODULE;
2103         dev->node_guid = dd->ipath_guid;
2104         dev->uverbs_abi_ver = IPATH_UVERBS_ABI_VERSION;
2105         dev->uverbs_cmd_mask =
2106                 (1ull << IB_USER_VERBS_CMD_GET_CONTEXT)         |
2107                 (1ull << IB_USER_VERBS_CMD_QUERY_DEVICE)        |
2108                 (1ull << IB_USER_VERBS_CMD_QUERY_PORT)          |
2109                 (1ull << IB_USER_VERBS_CMD_ALLOC_PD)            |
2110                 (1ull << IB_USER_VERBS_CMD_DEALLOC_PD)          |
2111                 (1ull << IB_USER_VERBS_CMD_CREATE_AH)           |
2112                 (1ull << IB_USER_VERBS_CMD_DESTROY_AH)          |
2113                 (1ull << IB_USER_VERBS_CMD_QUERY_AH)            |
2114                 (1ull << IB_USER_VERBS_CMD_REG_MR)              |
2115                 (1ull << IB_USER_VERBS_CMD_DEREG_MR)            |
2116                 (1ull << IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL) |
2117                 (1ull << IB_USER_VERBS_CMD_CREATE_CQ)           |
2118                 (1ull << IB_USER_VERBS_CMD_RESIZE_CQ)           |
2119                 (1ull << IB_USER_VERBS_CMD_DESTROY_CQ)          |
2120                 (1ull << IB_USER_VERBS_CMD_POLL_CQ)             |
2121                 (1ull << IB_USER_VERBS_CMD_REQ_NOTIFY_CQ)       |
2122                 (1ull << IB_USER_VERBS_CMD_CREATE_QP)           |
2123                 (1ull << IB_USER_VERBS_CMD_QUERY_QP)            |
2124                 (1ull << IB_USER_VERBS_CMD_MODIFY_QP)           |
2125                 (1ull << IB_USER_VERBS_CMD_DESTROY_QP)          |
2126                 (1ull << IB_USER_VERBS_CMD_POST_SEND)           |
2127                 (1ull << IB_USER_VERBS_CMD_POST_RECV)           |
2128                 (1ull << IB_USER_VERBS_CMD_ATTACH_MCAST)        |
2129                 (1ull << IB_USER_VERBS_CMD_DETACH_MCAST)        |
2130                 (1ull << IB_USER_VERBS_CMD_CREATE_SRQ)          |
2131                 (1ull << IB_USER_VERBS_CMD_MODIFY_SRQ)          |
2132                 (1ull << IB_USER_VERBS_CMD_QUERY_SRQ)           |
2133                 (1ull << IB_USER_VERBS_CMD_DESTROY_SRQ)         |
2134                 (1ull << IB_USER_VERBS_CMD_POST_SRQ_RECV);
2135         dev->node_type = RDMA_NODE_IB_CA;
2136         dev->phys_port_cnt = 1;
2137         dev->num_comp_vectors = 1;
2138         dev->dma_device = &dd->pcidev->dev;
2139         dev->query_device = ipath_query_device;
2140         dev->modify_device = ipath_modify_device;
2141         dev->query_port = ipath_query_port;
2142         dev->modify_port = ipath_modify_port;
2143         dev->query_pkey = ipath_query_pkey;
2144         dev->query_gid = ipath_query_gid;
2145         dev->alloc_ucontext = ipath_alloc_ucontext;
2146         dev->dealloc_ucontext = ipath_dealloc_ucontext;
2147         dev->alloc_pd = ipath_alloc_pd;
2148         dev->dealloc_pd = ipath_dealloc_pd;
2149         dev->create_ah = ipath_create_ah;
2150         dev->destroy_ah = ipath_destroy_ah;
2151         dev->query_ah = ipath_query_ah;
2152         dev->create_srq = ipath_create_srq;
2153         dev->modify_srq = ipath_modify_srq;
2154         dev->query_srq = ipath_query_srq;
2155         dev->destroy_srq = ipath_destroy_srq;
2156         dev->create_qp = ipath_create_qp;
2157         dev->modify_qp = ipath_modify_qp;
2158         dev->query_qp = ipath_query_qp;
2159         dev->destroy_qp = ipath_destroy_qp;
2160         dev->post_send = ipath_post_send;
2161         dev->post_recv = ipath_post_receive;
2162         dev->post_srq_recv = ipath_post_srq_receive;
2163         dev->create_cq = ipath_create_cq;
2164         dev->destroy_cq = ipath_destroy_cq;
2165         dev->resize_cq = ipath_resize_cq;
2166         dev->poll_cq = ipath_poll_cq;
2167         dev->req_notify_cq = ipath_req_notify_cq;
2168         dev->get_dma_mr = ipath_get_dma_mr;
2169         dev->reg_phys_mr = ipath_reg_phys_mr;
2170         dev->reg_user_mr = ipath_reg_user_mr;
2171         dev->dereg_mr = ipath_dereg_mr;
2172         dev->alloc_fmr = ipath_alloc_fmr;
2173         dev->map_phys_fmr = ipath_map_phys_fmr;
2174         dev->unmap_fmr = ipath_unmap_fmr;
2175         dev->dealloc_fmr = ipath_dealloc_fmr;
2176         dev->attach_mcast = ipath_multicast_attach;
2177         dev->detach_mcast = ipath_multicast_detach;
2178         dev->process_mad = ipath_process_mad;
2179         dev->mmap = ipath_mmap;
2180         dev->dma_ops = &ipath_dma_mapping_ops;
2181
2182         snprintf(dev->node_desc, sizeof(dev->node_desc),
2183                  IPATH_IDSTR " %s", init_utsname()->nodename);
2184
2185         ret = ib_register_device(dev, NULL);
2186         if (ret)
2187                 goto err_reg;
2188
2189         if (ipath_verbs_register_sysfs(dev))
2190                 goto err_class;
2191
2192         enable_timer(dd);
2193
2194         goto bail;
2195
2196 err_class:
2197         ib_unregister_device(dev);
2198 err_reg:
2199         kfree(idev->lk_table.table);
2200 err_lk:
2201         kfree(idev->qp_table.table);
2202 err_qp:
2203         kfree(idev->txreq_bufs);
2204 err_tx:
2205         ib_dealloc_device(dev);
2206         ipath_dev_err(dd, "cannot register verbs: %d!\n", -ret);
2207         idev = NULL;
2208
2209 bail:
2210         dd->verbs_dev = idev;
2211         return ret;
2212 }
2213
2214 void ipath_unregister_ib_device(struct ipath_ibdev *dev)
2215 {
2216         struct ib_device *ibdev = &dev->ibdev;
2217         u32 qps_inuse;
2218
2219         ib_unregister_device(ibdev);
2220
2221         disable_timer(dev->dd);
2222
2223         if (!list_empty(&dev->pending[0]) ||
2224             !list_empty(&dev->pending[1]) ||
2225             !list_empty(&dev->pending[2]))
2226                 ipath_dev_err(dev->dd, "pending list not empty!\n");
2227         if (!list_empty(&dev->piowait))
2228                 ipath_dev_err(dev->dd, "piowait list not empty!\n");
2229         if (!list_empty(&dev->rnrwait))
2230                 ipath_dev_err(dev->dd, "rnrwait list not empty!\n");
2231         if (!ipath_mcast_tree_empty())
2232                 ipath_dev_err(dev->dd, "multicast table memory leak!\n");
2233         /*
2234          * Note that ipath_unregister_ib_device() can be called before all
2235          * the QPs are destroyed!
2236          */
2237         qps_inuse = ipath_free_all_qps(&dev->qp_table);
2238         if (qps_inuse)
2239                 ipath_dev_err(dev->dd, "QP memory leak! %u still in use\n",
2240                         qps_inuse);
2241         kfree(dev->qp_table.table);
2242         kfree(dev->lk_table.table);
2243         kfree(dev->txreq_bufs);
2244         ib_dealloc_device(ibdev);
2245 }
2246
2247 static ssize_t show_rev(struct device *device, struct device_attribute *attr,
2248                         char *buf)
2249 {
2250         struct ipath_ibdev *dev =
2251                 container_of(device, struct ipath_ibdev, ibdev.dev);
2252
2253         return sprintf(buf, "%x\n", dev->dd->ipath_pcirev);
2254 }
2255
2256 static ssize_t show_hca(struct device *device, struct device_attribute *attr,
2257                         char *buf)
2258 {
2259         struct ipath_ibdev *dev =
2260                 container_of(device, struct ipath_ibdev, ibdev.dev);
2261         int ret;
2262
2263         ret = dev->dd->ipath_f_get_boardname(dev->dd, buf, 128);
2264         if (ret < 0)
2265                 goto bail;
2266         strcat(buf, "\n");
2267         ret = strlen(buf);
2268
2269 bail:
2270         return ret;
2271 }
2272
2273 static ssize_t show_stats(struct device *device, struct device_attribute *attr,
2274                           char *buf)
2275 {
2276         struct ipath_ibdev *dev =
2277                 container_of(device, struct ipath_ibdev, ibdev.dev);
2278         int i;
2279         int len;
2280
2281         len = sprintf(buf,
2282                       "RC resends  %d\n"
2283                       "RC no QACK  %d\n"
2284                       "RC ACKs     %d\n"
2285                       "RC SEQ NAKs %d\n"
2286                       "RC RDMA seq %d\n"
2287                       "RC RNR NAKs %d\n"
2288                       "RC OTH NAKs %d\n"
2289                       "RC timeouts %d\n"
2290                       "RC RDMA dup %d\n"
2291                       "piobuf wait %d\n"
2292                       "unaligned   %d\n"
2293                       "PKT drops   %d\n"
2294                       "WQE errs    %d\n",
2295                       dev->n_rc_resends, dev->n_rc_qacks, dev->n_rc_acks,
2296                       dev->n_seq_naks, dev->n_rdma_seq, dev->n_rnr_naks,
2297                       dev->n_other_naks, dev->n_timeouts,
2298                       dev->n_rdma_dup_busy, dev->n_piowait, dev->n_unaligned,
2299                       dev->n_pkt_drops, dev->n_wqe_errs);
2300         for (i = 0; i < ARRAY_SIZE(dev->opstats); i++) {
2301                 const struct ipath_opcode_stats *si = &dev->opstats[i];
2302
2303                 if (!si->n_packets && !si->n_bytes)
2304                         continue;
2305                 len += sprintf(buf + len, "%02x %llu/%llu\n", i,
2306                                (unsigned long long) si->n_packets,
2307                                (unsigned long long) si->n_bytes);
2308         }
2309         return len;
2310 }
2311
2312 static DEVICE_ATTR(hw_rev, S_IRUGO, show_rev, NULL);
2313 static DEVICE_ATTR(hca_type, S_IRUGO, show_hca, NULL);
2314 static DEVICE_ATTR(board_id, S_IRUGO, show_hca, NULL);
2315 static DEVICE_ATTR(stats, S_IRUGO, show_stats, NULL);
2316
2317 static struct device_attribute *ipath_class_attributes[] = {
2318         &dev_attr_hw_rev,
2319         &dev_attr_hca_type,
2320         &dev_attr_board_id,
2321         &dev_attr_stats
2322 };
2323
2324 static int ipath_verbs_register_sysfs(struct ib_device *dev)
2325 {
2326         int i;
2327         int ret;
2328
2329         for (i = 0; i < ARRAY_SIZE(ipath_class_attributes); ++i)
2330                 if (device_create_file(&dev->dev,
2331                                        ipath_class_attributes[i])) {
2332                         ret = 1;
2333                         goto bail;
2334                 }
2335
2336         ret = 0;
2337
2338 bail:
2339         return ret;
2340 }