IB/ipath: Update copyright dates
[linux-2.6] / drivers / infiniband / hw / ipath / ipath_verbs.c
1 /*
2  * Copyright (c) 2006, 2007 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/utsname.h>
38
39 #include "ipath_kernel.h"
40 #include "ipath_verbs.h"
41 #include "ipath_common.h"
42
43 static unsigned int ib_ipath_qp_table_size = 251;
44 module_param_named(qp_table_size, ib_ipath_qp_table_size, uint, S_IRUGO);
45 MODULE_PARM_DESC(qp_table_size, "QP table size");
46
47 unsigned int ib_ipath_lkey_table_size = 12;
48 module_param_named(lkey_table_size, ib_ipath_lkey_table_size, uint,
49                    S_IRUGO);
50 MODULE_PARM_DESC(lkey_table_size,
51                  "LKEY table size in bits (2^n, 1 <= n <= 23)");
52
53 static unsigned int ib_ipath_max_pds = 0xFFFF;
54 module_param_named(max_pds, ib_ipath_max_pds, uint, S_IWUSR | S_IRUGO);
55 MODULE_PARM_DESC(max_pds,
56                  "Maximum number of protection domains to support");
57
58 static unsigned int ib_ipath_max_ahs = 0xFFFF;
59 module_param_named(max_ahs, ib_ipath_max_ahs, uint, S_IWUSR | S_IRUGO);
60 MODULE_PARM_DESC(max_ahs, "Maximum number of address handles to support");
61
62 unsigned int ib_ipath_max_cqes = 0x2FFFF;
63 module_param_named(max_cqes, ib_ipath_max_cqes, uint, S_IWUSR | S_IRUGO);
64 MODULE_PARM_DESC(max_cqes,
65                  "Maximum number of completion queue entries to support");
66
67 unsigned int ib_ipath_max_cqs = 0x1FFFF;
68 module_param_named(max_cqs, ib_ipath_max_cqs, uint, S_IWUSR | S_IRUGO);
69 MODULE_PARM_DESC(max_cqs, "Maximum number of completion queues to support");
70
71 unsigned int ib_ipath_max_qp_wrs = 0x3FFF;
72 module_param_named(max_qp_wrs, ib_ipath_max_qp_wrs, uint,
73                    S_IWUSR | S_IRUGO);
74 MODULE_PARM_DESC(max_qp_wrs, "Maximum number of QP WRs to support");
75
76 unsigned int ib_ipath_max_qps = 16384;
77 module_param_named(max_qps, ib_ipath_max_qps, uint, S_IWUSR | S_IRUGO);
78 MODULE_PARM_DESC(max_qps, "Maximum number of QPs to support");
79
80 unsigned int ib_ipath_max_sges = 0x60;
81 module_param_named(max_sges, ib_ipath_max_sges, uint, S_IWUSR | S_IRUGO);
82 MODULE_PARM_DESC(max_sges, "Maximum number of SGEs to support");
83
84 unsigned int ib_ipath_max_mcast_grps = 16384;
85 module_param_named(max_mcast_grps, ib_ipath_max_mcast_grps, uint,
86                    S_IWUSR | S_IRUGO);
87 MODULE_PARM_DESC(max_mcast_grps,
88                  "Maximum number of multicast groups to support");
89
90 unsigned int ib_ipath_max_mcast_qp_attached = 16;
91 module_param_named(max_mcast_qp_attached, ib_ipath_max_mcast_qp_attached,
92                    uint, S_IWUSR | S_IRUGO);
93 MODULE_PARM_DESC(max_mcast_qp_attached,
94                  "Maximum number of attached QPs to support");
95
96 unsigned int ib_ipath_max_srqs = 1024;
97 module_param_named(max_srqs, ib_ipath_max_srqs, uint, S_IWUSR | S_IRUGO);
98 MODULE_PARM_DESC(max_srqs, "Maximum number of SRQs to support");
99
100 unsigned int ib_ipath_max_srq_sges = 128;
101 module_param_named(max_srq_sges, ib_ipath_max_srq_sges,
102                    uint, S_IWUSR | S_IRUGO);
103 MODULE_PARM_DESC(max_srq_sges, "Maximum number of SRQ SGEs to support");
104
105 unsigned int ib_ipath_max_srq_wrs = 0x1FFFF;
106 module_param_named(max_srq_wrs, ib_ipath_max_srq_wrs,
107                    uint, S_IWUSR | S_IRUGO);
108 MODULE_PARM_DESC(max_srq_wrs, "Maximum number of SRQ WRs support");
109
110 static unsigned int ib_ipath_disable_sma;
111 module_param_named(disable_sma, ib_ipath_disable_sma, uint, S_IWUSR | S_IRUGO);
112 MODULE_PARM_DESC(ib_ipath_disable_sma, "Disable the SMA");
113
114 const int ib_ipath_state_ops[IB_QPS_ERR + 1] = {
115         [IB_QPS_RESET] = 0,
116         [IB_QPS_INIT] = IPATH_POST_RECV_OK,
117         [IB_QPS_RTR] = IPATH_POST_RECV_OK | IPATH_PROCESS_RECV_OK,
118         [IB_QPS_RTS] = IPATH_POST_RECV_OK | IPATH_PROCESS_RECV_OK |
119             IPATH_POST_SEND_OK | IPATH_PROCESS_SEND_OK,
120         [IB_QPS_SQD] = IPATH_POST_RECV_OK | IPATH_PROCESS_RECV_OK |
121             IPATH_POST_SEND_OK,
122         [IB_QPS_SQE] = IPATH_POST_RECV_OK | IPATH_PROCESS_RECV_OK,
123         [IB_QPS_ERR] = 0,
124 };
125
126 struct ipath_ucontext {
127         struct ib_ucontext ibucontext;
128 };
129
130 static inline struct ipath_ucontext *to_iucontext(struct ib_ucontext
131                                                   *ibucontext)
132 {
133         return container_of(ibucontext, struct ipath_ucontext, ibucontext);
134 }
135
136 /*
137  * Translate ib_wr_opcode into ib_wc_opcode.
138  */
139 const enum ib_wc_opcode ib_ipath_wc_opcode[] = {
140         [IB_WR_RDMA_WRITE] = IB_WC_RDMA_WRITE,
141         [IB_WR_RDMA_WRITE_WITH_IMM] = IB_WC_RDMA_WRITE,
142         [IB_WR_SEND] = IB_WC_SEND,
143         [IB_WR_SEND_WITH_IMM] = IB_WC_SEND,
144         [IB_WR_RDMA_READ] = IB_WC_RDMA_READ,
145         [IB_WR_ATOMIC_CMP_AND_SWP] = IB_WC_COMP_SWAP,
146         [IB_WR_ATOMIC_FETCH_AND_ADD] = IB_WC_FETCH_ADD
147 };
148
149 /*
150  * System image GUID.
151  */
152 static __be64 sys_image_guid;
153
154 /**
155  * ipath_copy_sge - copy data to SGE memory
156  * @ss: the SGE state
157  * @data: the data to copy
158  * @length: the length of the data
159  */
160 void ipath_copy_sge(struct ipath_sge_state *ss, void *data, u32 length)
161 {
162         struct ipath_sge *sge = &ss->sge;
163
164         while (length) {
165                 u32 len = sge->length;
166
167                 if (len > length)
168                         len = length;
169                 if (len > sge->sge_length)
170                         len = sge->sge_length;
171                 BUG_ON(len == 0);
172                 memcpy(sge->vaddr, data, len);
173                 sge->vaddr += len;
174                 sge->length -= len;
175                 sge->sge_length -= len;
176                 if (sge->sge_length == 0) {
177                         if (--ss->num_sge)
178                                 *sge = *ss->sg_list++;
179                 } else if (sge->length == 0 && sge->mr != NULL) {
180                         if (++sge->n >= IPATH_SEGSZ) {
181                                 if (++sge->m >= sge->mr->mapsz)
182                                         break;
183                                 sge->n = 0;
184                         }
185                         sge->vaddr =
186                                 sge->mr->map[sge->m]->segs[sge->n].vaddr;
187                         sge->length =
188                                 sge->mr->map[sge->m]->segs[sge->n].length;
189                 }
190                 data += len;
191                 length -= len;
192         }
193 }
194
195 /**
196  * ipath_skip_sge - skip over SGE memory - XXX almost dup of prev func
197  * @ss: the SGE state
198  * @length: the number of bytes to skip
199  */
200 void ipath_skip_sge(struct ipath_sge_state *ss, u32 length)
201 {
202         struct ipath_sge *sge = &ss->sge;
203
204         while (length) {
205                 u32 len = sge->length;
206
207                 if (len > length)
208                         len = length;
209                 if (len > sge->sge_length)
210                         len = sge->sge_length;
211                 BUG_ON(len == 0);
212                 sge->vaddr += len;
213                 sge->length -= len;
214                 sge->sge_length -= len;
215                 if (sge->sge_length == 0) {
216                         if (--ss->num_sge)
217                                 *sge = *ss->sg_list++;
218                 } else if (sge->length == 0 && sge->mr != NULL) {
219                         if (++sge->n >= IPATH_SEGSZ) {
220                                 if (++sge->m >= sge->mr->mapsz)
221                                         break;
222                                 sge->n = 0;
223                         }
224                         sge->vaddr =
225                                 sge->mr->map[sge->m]->segs[sge->n].vaddr;
226                         sge->length =
227                                 sge->mr->map[sge->m]->segs[sge->n].length;
228                 }
229                 length -= len;
230         }
231 }
232
233 /**
234  * ipath_post_send - post a send on a QP
235  * @ibqp: the QP to post the send on
236  * @wr: the list of work requests to post
237  * @bad_wr: the first bad WR is put here
238  *
239  * This may be called from interrupt context.
240  */
241 static int ipath_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
242                            struct ib_send_wr **bad_wr)
243 {
244         struct ipath_qp *qp = to_iqp(ibqp);
245         int err = 0;
246
247         /* Check that state is OK to post send. */
248         if (!(ib_ipath_state_ops[qp->state] & IPATH_POST_SEND_OK)) {
249                 *bad_wr = wr;
250                 err = -EINVAL;
251                 goto bail;
252         }
253
254         for (; wr; wr = wr->next) {
255                 switch (qp->ibqp.qp_type) {
256                 case IB_QPT_UC:
257                 case IB_QPT_RC:
258                         err = ipath_post_ruc_send(qp, wr);
259                         break;
260
261                 case IB_QPT_SMI:
262                 case IB_QPT_GSI:
263                 case IB_QPT_UD:
264                         err = ipath_post_ud_send(qp, wr);
265                         break;
266
267                 default:
268                         err = -EINVAL;
269                 }
270                 if (err) {
271                         *bad_wr = wr;
272                         break;
273                 }
274         }
275
276 bail:
277         return err;
278 }
279
280 /**
281  * ipath_post_receive - post a receive on a QP
282  * @ibqp: the QP to post the receive on
283  * @wr: the WR to post
284  * @bad_wr: the first bad WR is put here
285  *
286  * This may be called from interrupt context.
287  */
288 static int ipath_post_receive(struct ib_qp *ibqp, struct ib_recv_wr *wr,
289                               struct ib_recv_wr **bad_wr)
290 {
291         struct ipath_qp *qp = to_iqp(ibqp);
292         struct ipath_rwq *wq = qp->r_rq.wq;
293         unsigned long flags;
294         int ret;
295
296         /* Check that state is OK to post receive. */
297         if (!(ib_ipath_state_ops[qp->state] & IPATH_POST_RECV_OK) || !wq) {
298                 *bad_wr = wr;
299                 ret = -EINVAL;
300                 goto bail;
301         }
302
303         for (; wr; wr = wr->next) {
304                 struct ipath_rwqe *wqe;
305                 u32 next;
306                 int i;
307
308                 if ((unsigned) wr->num_sge > qp->r_rq.max_sge) {
309                         *bad_wr = wr;
310                         ret = -ENOMEM;
311                         goto bail;
312                 }
313
314                 spin_lock_irqsave(&qp->r_rq.lock, flags);
315                 next = wq->head + 1;
316                 if (next >= qp->r_rq.size)
317                         next = 0;
318                 if (next == wq->tail) {
319                         spin_unlock_irqrestore(&qp->r_rq.lock, flags);
320                         *bad_wr = wr;
321                         ret = -ENOMEM;
322                         goto bail;
323                 }
324
325                 wqe = get_rwqe_ptr(&qp->r_rq, wq->head);
326                 wqe->wr_id = wr->wr_id;
327                 wqe->num_sge = wr->num_sge;
328                 for (i = 0; i < wr->num_sge; i++)
329                         wqe->sg_list[i] = wr->sg_list[i];
330                 wq->head = next;
331                 spin_unlock_irqrestore(&qp->r_rq.lock, flags);
332         }
333         ret = 0;
334
335 bail:
336         return ret;
337 }
338
339 /**
340  * ipath_qp_rcv - processing an incoming packet on a QP
341  * @dev: the device the packet came on
342  * @hdr: the packet header
343  * @has_grh: true if the packet has a GRH
344  * @data: the packet data
345  * @tlen: the packet length
346  * @qp: the QP the packet came on
347  *
348  * This is called from ipath_ib_rcv() to process an incoming packet
349  * for the given QP.
350  * Called at interrupt level.
351  */
352 static void ipath_qp_rcv(struct ipath_ibdev *dev,
353                          struct ipath_ib_header *hdr, int has_grh,
354                          void *data, u32 tlen, struct ipath_qp *qp)
355 {
356         /* Check for valid receive state. */
357         if (!(ib_ipath_state_ops[qp->state] & IPATH_PROCESS_RECV_OK)) {
358                 dev->n_pkt_drops++;
359                 return;
360         }
361
362         switch (qp->ibqp.qp_type) {
363         case IB_QPT_SMI:
364         case IB_QPT_GSI:
365                 if (ib_ipath_disable_sma)
366                         break;
367                 /* FALLTHROUGH */
368         case IB_QPT_UD:
369                 ipath_ud_rcv(dev, hdr, has_grh, data, tlen, qp);
370                 break;
371
372         case IB_QPT_RC:
373                 ipath_rc_rcv(dev, hdr, has_grh, data, tlen, qp);
374                 break;
375
376         case IB_QPT_UC:
377                 ipath_uc_rcv(dev, hdr, has_grh, data, tlen, qp);
378                 break;
379
380         default:
381                 break;
382         }
383 }
384
385 /**
386  * ipath_ib_rcv - process an incoming packet
387  * @arg: the device pointer
388  * @rhdr: the header of the packet
389  * @data: the packet data
390  * @tlen: the packet length
391  *
392  * This is called from ipath_kreceive() to process an incoming packet at
393  * interrupt level. Tlen is the length of the header + data + CRC in bytes.
394  */
395 void ipath_ib_rcv(struct ipath_ibdev *dev, void *rhdr, void *data,
396                   u32 tlen)
397 {
398         struct ipath_ib_header *hdr = rhdr;
399         struct ipath_other_headers *ohdr;
400         struct ipath_qp *qp;
401         u32 qp_num;
402         int lnh;
403         u8 opcode;
404         u16 lid;
405
406         if (unlikely(dev == NULL))
407                 goto bail;
408
409         if (unlikely(tlen < 24)) {      /* LRH+BTH+CRC */
410                 dev->rcv_errors++;
411                 goto bail;
412         }
413
414         /* Check for a valid destination LID (see ch. 7.11.1). */
415         lid = be16_to_cpu(hdr->lrh[1]);
416         if (lid < IPATH_MULTICAST_LID_BASE) {
417                 lid &= ~((1 << (dev->mkeyprot_resv_lmc & 7)) - 1);
418                 if (unlikely(lid != dev->dd->ipath_lid)) {
419                         dev->rcv_errors++;
420                         goto bail;
421                 }
422         }
423
424         /* Check for GRH */
425         lnh = be16_to_cpu(hdr->lrh[0]) & 3;
426         if (lnh == IPATH_LRH_BTH)
427                 ohdr = &hdr->u.oth;
428         else if (lnh == IPATH_LRH_GRH)
429                 ohdr = &hdr->u.l.oth;
430         else {
431                 dev->rcv_errors++;
432                 goto bail;
433         }
434
435         opcode = be32_to_cpu(ohdr->bth[0]) >> 24;
436         dev->opstats[opcode].n_bytes += tlen;
437         dev->opstats[opcode].n_packets++;
438
439         /* Get the destination QP number. */
440         qp_num = be32_to_cpu(ohdr->bth[1]) & IPATH_QPN_MASK;
441         if (qp_num == IPATH_MULTICAST_QPN) {
442                 struct ipath_mcast *mcast;
443                 struct ipath_mcast_qp *p;
444
445                 if (lnh != IPATH_LRH_GRH) {
446                         dev->n_pkt_drops++;
447                         goto bail;
448                 }
449                 mcast = ipath_mcast_find(&hdr->u.l.grh.dgid);
450                 if (mcast == NULL) {
451                         dev->n_pkt_drops++;
452                         goto bail;
453                 }
454                 dev->n_multicast_rcv++;
455                 list_for_each_entry_rcu(p, &mcast->qp_list, list)
456                         ipath_qp_rcv(dev, hdr, 1, data, tlen, p->qp);
457                 /*
458                  * Notify ipath_multicast_detach() if it is waiting for us
459                  * to finish.
460                  */
461                 if (atomic_dec_return(&mcast->refcount) <= 1)
462                         wake_up(&mcast->wait);
463         } else {
464                 qp = ipath_lookup_qpn(&dev->qp_table, qp_num);
465                 if (qp) {
466                         dev->n_unicast_rcv++;
467                         ipath_qp_rcv(dev, hdr, lnh == IPATH_LRH_GRH, data,
468                                      tlen, qp);
469                         /*
470                          * Notify ipath_destroy_qp() if it is waiting
471                          * for us to finish.
472                          */
473                         if (atomic_dec_and_test(&qp->refcount))
474                                 wake_up(&qp->wait);
475                 } else
476                         dev->n_pkt_drops++;
477         }
478
479 bail:;
480 }
481
482 /**
483  * ipath_ib_timer - verbs timer
484  * @arg: the device pointer
485  *
486  * This is called from ipath_do_rcv_timer() at interrupt level to check for
487  * QPs which need retransmits and to collect performance numbers.
488  */
489 void ipath_ib_timer(struct ipath_ibdev *dev)
490 {
491         struct ipath_qp *resend = NULL;
492         struct list_head *last;
493         struct ipath_qp *qp;
494         unsigned long flags;
495
496         if (dev == NULL)
497                 return;
498
499         spin_lock_irqsave(&dev->pending_lock, flags);
500         /* Start filling the next pending queue. */
501         if (++dev->pending_index >= ARRAY_SIZE(dev->pending))
502                 dev->pending_index = 0;
503         /* Save any requests still in the new queue, they have timed out. */
504         last = &dev->pending[dev->pending_index];
505         while (!list_empty(last)) {
506                 qp = list_entry(last->next, struct ipath_qp, timerwait);
507                 list_del_init(&qp->timerwait);
508                 qp->timer_next = resend;
509                 resend = qp;
510                 atomic_inc(&qp->refcount);
511         }
512         last = &dev->rnrwait;
513         if (!list_empty(last)) {
514                 qp = list_entry(last->next, struct ipath_qp, timerwait);
515                 if (--qp->s_rnr_timeout == 0) {
516                         do {
517                                 list_del_init(&qp->timerwait);
518                                 tasklet_hi_schedule(&qp->s_task);
519                                 if (list_empty(last))
520                                         break;
521                                 qp = list_entry(last->next, struct ipath_qp,
522                                                 timerwait);
523                         } while (qp->s_rnr_timeout == 0);
524                 }
525         }
526         /*
527          * We should only be in the started state if pma_sample_start != 0
528          */
529         if (dev->pma_sample_status == IB_PMA_SAMPLE_STATUS_STARTED &&
530             --dev->pma_sample_start == 0) {
531                 dev->pma_sample_status = IB_PMA_SAMPLE_STATUS_RUNNING;
532                 ipath_snapshot_counters(dev->dd, &dev->ipath_sword,
533                                         &dev->ipath_rword,
534                                         &dev->ipath_spkts,
535                                         &dev->ipath_rpkts,
536                                         &dev->ipath_xmit_wait);
537         }
538         if (dev->pma_sample_status == IB_PMA_SAMPLE_STATUS_RUNNING) {
539                 if (dev->pma_sample_interval == 0) {
540                         u64 ta, tb, tc, td, te;
541
542                         dev->pma_sample_status = IB_PMA_SAMPLE_STATUS_DONE;
543                         ipath_snapshot_counters(dev->dd, &ta, &tb,
544                                                 &tc, &td, &te);
545
546                         dev->ipath_sword = ta - dev->ipath_sword;
547                         dev->ipath_rword = tb - dev->ipath_rword;
548                         dev->ipath_spkts = tc - dev->ipath_spkts;
549                         dev->ipath_rpkts = td - dev->ipath_rpkts;
550                         dev->ipath_xmit_wait = te - dev->ipath_xmit_wait;
551                 }
552                 else
553                         dev->pma_sample_interval--;
554         }
555         spin_unlock_irqrestore(&dev->pending_lock, flags);
556
557         /* XXX What if timer fires again while this is running? */
558         for (qp = resend; qp != NULL; qp = qp->timer_next) {
559                 struct ib_wc wc;
560
561                 spin_lock_irqsave(&qp->s_lock, flags);
562                 if (qp->s_last != qp->s_tail && qp->state == IB_QPS_RTS) {
563                         dev->n_timeouts++;
564                         ipath_restart_rc(qp, qp->s_last_psn + 1, &wc);
565                 }
566                 spin_unlock_irqrestore(&qp->s_lock, flags);
567
568                 /* Notify ipath_destroy_qp() if it is waiting. */
569                 if (atomic_dec_and_test(&qp->refcount))
570                         wake_up(&qp->wait);
571         }
572 }
573
574 static void update_sge(struct ipath_sge_state *ss, u32 length)
575 {
576         struct ipath_sge *sge = &ss->sge;
577
578         sge->vaddr += length;
579         sge->length -= length;
580         sge->sge_length -= length;
581         if (sge->sge_length == 0) {
582                 if (--ss->num_sge)
583                         *sge = *ss->sg_list++;
584         } else if (sge->length == 0 && sge->mr != NULL) {
585                 if (++sge->n >= IPATH_SEGSZ) {
586                         if (++sge->m >= sge->mr->mapsz)
587                                 return;
588                         sge->n = 0;
589                 }
590                 sge->vaddr = sge->mr->map[sge->m]->segs[sge->n].vaddr;
591                 sge->length = sge->mr->map[sge->m]->segs[sge->n].length;
592         }
593 }
594
595 #ifdef __LITTLE_ENDIAN
596 static inline u32 get_upper_bits(u32 data, u32 shift)
597 {
598         return data >> shift;
599 }
600
601 static inline u32 set_upper_bits(u32 data, u32 shift)
602 {
603         return data << shift;
604 }
605
606 static inline u32 clear_upper_bytes(u32 data, u32 n, u32 off)
607 {
608         data <<= ((sizeof(u32) - n) * BITS_PER_BYTE);
609         data >>= ((sizeof(u32) - n - off) * BITS_PER_BYTE);
610         return data;
611 }
612 #else
613 static inline u32 get_upper_bits(u32 data, u32 shift)
614 {
615         return data << shift;
616 }
617
618 static inline u32 set_upper_bits(u32 data, u32 shift)
619 {
620         return data >> shift;
621 }
622
623 static inline u32 clear_upper_bytes(u32 data, u32 n, u32 off)
624 {
625         data >>= ((sizeof(u32) - n) * BITS_PER_BYTE);
626         data <<= ((sizeof(u32) - n - off) * BITS_PER_BYTE);
627         return data;
628 }
629 #endif
630
631 static void copy_io(u32 __iomem *piobuf, struct ipath_sge_state *ss,
632                     u32 length)
633 {
634         u32 extra = 0;
635         u32 data = 0;
636         u32 last;
637
638         while (1) {
639                 u32 len = ss->sge.length;
640                 u32 off;
641
642                 BUG_ON(len == 0);
643                 if (len > length)
644                         len = length;
645                 if (len > ss->sge.sge_length)
646                         len = ss->sge.sge_length;
647                 /* If the source address is not aligned, try to align it. */
648                 off = (unsigned long)ss->sge.vaddr & (sizeof(u32) - 1);
649                 if (off) {
650                         u32 *addr = (u32 *)((unsigned long)ss->sge.vaddr &
651                                             ~(sizeof(u32) - 1));
652                         u32 v = get_upper_bits(*addr, off * BITS_PER_BYTE);
653                         u32 y;
654
655                         y = sizeof(u32) - off;
656                         if (len > y)
657                                 len = y;
658                         if (len + extra >= sizeof(u32)) {
659                                 data |= set_upper_bits(v, extra *
660                                                        BITS_PER_BYTE);
661                                 len = sizeof(u32) - extra;
662                                 if (len == length) {
663                                         last = data;
664                                         break;
665                                 }
666                                 __raw_writel(data, piobuf);
667                                 piobuf++;
668                                 extra = 0;
669                                 data = 0;
670                         } else {
671                                 /* Clear unused upper bytes */
672                                 data |= clear_upper_bytes(v, len, extra);
673                                 if (len == length) {
674                                         last = data;
675                                         break;
676                                 }
677                                 extra += len;
678                         }
679                 } else if (extra) {
680                         /* Source address is aligned. */
681                         u32 *addr = (u32 *) ss->sge.vaddr;
682                         int shift = extra * BITS_PER_BYTE;
683                         int ushift = 32 - shift;
684                         u32 l = len;
685
686                         while (l >= sizeof(u32)) {
687                                 u32 v = *addr;
688
689                                 data |= set_upper_bits(v, shift);
690                                 __raw_writel(data, piobuf);
691                                 data = get_upper_bits(v, ushift);
692                                 piobuf++;
693                                 addr++;
694                                 l -= sizeof(u32);
695                         }
696                         /*
697                          * We still have 'extra' number of bytes leftover.
698                          */
699                         if (l) {
700                                 u32 v = *addr;
701
702                                 if (l + extra >= sizeof(u32)) {
703                                         data |= set_upper_bits(v, shift);
704                                         len -= l + extra - sizeof(u32);
705                                         if (len == length) {
706                                                 last = data;
707                                                 break;
708                                         }
709                                         __raw_writel(data, piobuf);
710                                         piobuf++;
711                                         extra = 0;
712                                         data = 0;
713                                 } else {
714                                         /* Clear unused upper bytes */
715                                         data |= clear_upper_bytes(v, l,
716                                                                   extra);
717                                         if (len == length) {
718                                                 last = data;
719                                                 break;
720                                         }
721                                         extra += l;
722                                 }
723                         } else if (len == length) {
724                                 last = data;
725                                 break;
726                         }
727                 } else if (len == length) {
728                         u32 w;
729
730                         /*
731                          * Need to round up for the last dword in the
732                          * packet.
733                          */
734                         w = (len + 3) >> 2;
735                         __iowrite32_copy(piobuf, ss->sge.vaddr, w - 1);
736                         piobuf += w - 1;
737                         last = ((u32 *) ss->sge.vaddr)[w - 1];
738                         break;
739                 } else {
740                         u32 w = len >> 2;
741
742                         __iowrite32_copy(piobuf, ss->sge.vaddr, w);
743                         piobuf += w;
744
745                         extra = len & (sizeof(u32) - 1);
746                         if (extra) {
747                                 u32 v = ((u32 *) ss->sge.vaddr)[w];
748
749                                 /* Clear unused upper bytes */
750                                 data = clear_upper_bytes(v, extra, 0);
751                         }
752                 }
753                 update_sge(ss, len);
754                 length -= len;
755         }
756         /* Update address before sending packet. */
757         update_sge(ss, length);
758         /* must flush early everything before trigger word */
759         ipath_flush_wc();
760         __raw_writel(last, piobuf);
761         /* be sure trigger word is written */
762         ipath_flush_wc();
763 }
764
765 /**
766  * ipath_verbs_send - send a packet
767  * @dd: the infinipath device
768  * @hdrwords: the number of words in the header
769  * @hdr: the packet header
770  * @len: the length of the packet in bytes
771  * @ss: the SGE to send
772  */
773 int ipath_verbs_send(struct ipath_devdata *dd, u32 hdrwords,
774                      u32 *hdr, u32 len, struct ipath_sge_state *ss)
775 {
776         u32 __iomem *piobuf;
777         u32 plen;
778         int ret;
779
780         /* +1 is for the qword padding of pbc */
781         plen = hdrwords + ((len + 3) >> 2) + 1;
782         if (unlikely((plen << 2) > dd->ipath_ibmaxlen)) {
783                 ret = -EINVAL;
784                 goto bail;
785         }
786
787         /* Get a PIO buffer to use. */
788         piobuf = ipath_getpiobuf(dd, NULL);
789         if (unlikely(piobuf == NULL)) {
790                 ret = -EBUSY;
791                 goto bail;
792         }
793
794         /*
795          * Write len to control qword, no flags.
796          * We have to flush after the PBC for correctness on some cpus
797          * or WC buffer can be written out of order.
798          */
799         writeq(plen, piobuf);
800         ipath_flush_wc();
801         piobuf += 2;
802         if (len == 0) {
803                 /*
804                  * If there is just the header portion, must flush before
805                  * writing last word of header for correctness, and after
806                  * the last header word (trigger word).
807                  */
808                 __iowrite32_copy(piobuf, hdr, hdrwords - 1);
809                 ipath_flush_wc();
810                 __raw_writel(hdr[hdrwords - 1], piobuf + hdrwords - 1);
811                 ipath_flush_wc();
812                 ret = 0;
813                 goto bail;
814         }
815
816         __iowrite32_copy(piobuf, hdr, hdrwords);
817         piobuf += hdrwords;
818
819         /* The common case is aligned and contained in one segment. */
820         if (likely(ss->num_sge == 1 && len <= ss->sge.length &&
821                    !((unsigned long)ss->sge.vaddr & (sizeof(u32) - 1)))) {
822                 u32 w;
823                 u32 *addr = (u32 *) ss->sge.vaddr;
824
825                 /* Update address before sending packet. */
826                 update_sge(ss, len);
827                 /* Need to round up for the last dword in the packet. */
828                 w = (len + 3) >> 2;
829                 __iowrite32_copy(piobuf, addr, w - 1);
830                 /* must flush early everything before trigger word */
831                 ipath_flush_wc();
832                 __raw_writel(addr[w - 1], piobuf + w - 1);
833                 /* be sure trigger word is written */
834                 ipath_flush_wc();
835                 ret = 0;
836                 goto bail;
837         }
838         copy_io(piobuf, ss, len);
839         ret = 0;
840
841 bail:
842         return ret;
843 }
844
845 int ipath_snapshot_counters(struct ipath_devdata *dd, u64 *swords,
846                             u64 *rwords, u64 *spkts, u64 *rpkts,
847                             u64 *xmit_wait)
848 {
849         int ret;
850
851         if (!(dd->ipath_flags & IPATH_INITTED)) {
852                 /* no hardware, freeze, etc. */
853                 ipath_dbg("unit %u not usable\n", dd->ipath_unit);
854                 ret = -EINVAL;
855                 goto bail;
856         }
857         *swords = ipath_snap_cntr(dd, dd->ipath_cregs->cr_wordsendcnt);
858         *rwords = ipath_snap_cntr(dd, dd->ipath_cregs->cr_wordrcvcnt);
859         *spkts = ipath_snap_cntr(dd, dd->ipath_cregs->cr_pktsendcnt);
860         *rpkts = ipath_snap_cntr(dd, dd->ipath_cregs->cr_pktrcvcnt);
861         *xmit_wait = ipath_snap_cntr(dd, dd->ipath_cregs->cr_sendstallcnt);
862
863         ret = 0;
864
865 bail:
866         return ret;
867 }
868
869 /**
870  * ipath_get_counters - get various chip counters
871  * @dd: the infinipath device
872  * @cntrs: counters are placed here
873  *
874  * Return the counters needed by recv_pma_get_portcounters().
875  */
876 int ipath_get_counters(struct ipath_devdata *dd,
877                        struct ipath_verbs_counters *cntrs)
878 {
879         int ret;
880
881         if (!(dd->ipath_flags & IPATH_INITTED)) {
882                 /* no hardware, freeze, etc. */
883                 ipath_dbg("unit %u not usable\n", dd->ipath_unit);
884                 ret = -EINVAL;
885                 goto bail;
886         }
887         cntrs->symbol_error_counter =
888                 ipath_snap_cntr(dd, dd->ipath_cregs->cr_ibsymbolerrcnt);
889         cntrs->link_error_recovery_counter =
890                 ipath_snap_cntr(dd, dd->ipath_cregs->cr_iblinkerrrecovcnt);
891         /*
892          * The link downed counter counts when the other side downs the
893          * connection.  We add in the number of times we downed the link
894          * due to local link integrity errors to compensate.
895          */
896         cntrs->link_downed_counter =
897                 ipath_snap_cntr(dd, dd->ipath_cregs->cr_iblinkdowncnt);
898         cntrs->port_rcv_errors =
899                 ipath_snap_cntr(dd, dd->ipath_cregs->cr_rxdroppktcnt) +
900                 ipath_snap_cntr(dd, dd->ipath_cregs->cr_rcvovflcnt) +
901                 ipath_snap_cntr(dd, dd->ipath_cregs->cr_portovflcnt) +
902                 ipath_snap_cntr(dd, dd->ipath_cregs->cr_err_rlencnt) +
903                 ipath_snap_cntr(dd, dd->ipath_cregs->cr_invalidrlencnt) +
904                 ipath_snap_cntr(dd, dd->ipath_cregs->cr_erricrccnt) +
905                 ipath_snap_cntr(dd, dd->ipath_cregs->cr_errvcrccnt) +
906                 ipath_snap_cntr(dd, dd->ipath_cregs->cr_errlpcrccnt) +
907                 ipath_snap_cntr(dd, dd->ipath_cregs->cr_badformatcnt) +
908                 dd->ipath_rxfc_unsupvl_errs;
909         cntrs->port_rcv_remphys_errors =
910                 ipath_snap_cntr(dd, dd->ipath_cregs->cr_rcvebpcnt);
911         cntrs->port_xmit_discards =
912                 ipath_snap_cntr(dd, dd->ipath_cregs->cr_unsupvlcnt);
913         cntrs->port_xmit_data =
914                 ipath_snap_cntr(dd, dd->ipath_cregs->cr_wordsendcnt);
915         cntrs->port_rcv_data =
916                 ipath_snap_cntr(dd, dd->ipath_cregs->cr_wordrcvcnt);
917         cntrs->port_xmit_packets =
918                 ipath_snap_cntr(dd, dd->ipath_cregs->cr_pktsendcnt);
919         cntrs->port_rcv_packets =
920                 ipath_snap_cntr(dd, dd->ipath_cregs->cr_pktrcvcnt);
921         cntrs->local_link_integrity_errors =
922                 (dd->ipath_flags & IPATH_GPIO_ERRINTRS) ?
923                 dd->ipath_lli_errs : dd->ipath_lli_errors;
924         cntrs->excessive_buffer_overrun_errors = dd->ipath_overrun_thresh_errs;
925
926         ret = 0;
927
928 bail:
929         return ret;
930 }
931
932 /**
933  * ipath_ib_piobufavail - callback when a PIO buffer is available
934  * @arg: the device pointer
935  *
936  * This is called from ipath_intr() at interrupt level when a PIO buffer is
937  * available after ipath_verbs_send() returned an error that no buffers were
938  * available.  Return 1 if we consumed all the PIO buffers and we still have
939  * QPs waiting for buffers (for now, just do a tasklet_hi_schedule and
940  * return zero).
941  */
942 int ipath_ib_piobufavail(struct ipath_ibdev *dev)
943 {
944         struct ipath_qp *qp;
945         unsigned long flags;
946
947         if (dev == NULL)
948                 goto bail;
949
950         spin_lock_irqsave(&dev->pending_lock, flags);
951         while (!list_empty(&dev->piowait)) {
952                 qp = list_entry(dev->piowait.next, struct ipath_qp,
953                                 piowait);
954                 list_del_init(&qp->piowait);
955                 clear_bit(IPATH_S_BUSY, &qp->s_busy);
956                 tasklet_hi_schedule(&qp->s_task);
957         }
958         spin_unlock_irqrestore(&dev->pending_lock, flags);
959
960 bail:
961         return 0;
962 }
963
964 static int ipath_query_device(struct ib_device *ibdev,
965                               struct ib_device_attr *props)
966 {
967         struct ipath_ibdev *dev = to_idev(ibdev);
968
969         memset(props, 0, sizeof(*props));
970
971         props->device_cap_flags = IB_DEVICE_BAD_PKEY_CNTR |
972                 IB_DEVICE_BAD_QKEY_CNTR | IB_DEVICE_SHUTDOWN_PORT |
973                 IB_DEVICE_SYS_IMAGE_GUID;
974         props->page_size_cap = PAGE_SIZE;
975         props->vendor_id = dev->dd->ipath_vendorid;
976         props->vendor_part_id = dev->dd->ipath_deviceid;
977         props->hw_ver = dev->dd->ipath_pcirev;
978
979         props->sys_image_guid = dev->sys_image_guid;
980
981         props->max_mr_size = ~0ull;
982         props->max_qp = ib_ipath_max_qps;
983         props->max_qp_wr = ib_ipath_max_qp_wrs;
984         props->max_sge = ib_ipath_max_sges;
985         props->max_cq = ib_ipath_max_cqs;
986         props->max_ah = ib_ipath_max_ahs;
987         props->max_cqe = ib_ipath_max_cqes;
988         props->max_mr = dev->lk_table.max;
989         props->max_fmr = dev->lk_table.max;
990         props->max_map_per_fmr = 32767;
991         props->max_pd = ib_ipath_max_pds;
992         props->max_qp_rd_atom = IPATH_MAX_RDMA_ATOMIC;
993         props->max_qp_init_rd_atom = 255;
994         /* props->max_res_rd_atom */
995         props->max_srq = ib_ipath_max_srqs;
996         props->max_srq_wr = ib_ipath_max_srq_wrs;
997         props->max_srq_sge = ib_ipath_max_srq_sges;
998         /* props->local_ca_ack_delay */
999         props->atomic_cap = IB_ATOMIC_GLOB;
1000         props->max_pkeys = ipath_get_npkeys(dev->dd);
1001         props->max_mcast_grp = ib_ipath_max_mcast_grps;
1002         props->max_mcast_qp_attach = ib_ipath_max_mcast_qp_attached;
1003         props->max_total_mcast_qp_attach = props->max_mcast_qp_attach *
1004                 props->max_mcast_grp;
1005
1006         return 0;
1007 }
1008
1009 const u8 ipath_cvt_physportstate[16] = {
1010         [INFINIPATH_IBCS_LT_STATE_DISABLED] = 3,
1011         [INFINIPATH_IBCS_LT_STATE_LINKUP] = 5,
1012         [INFINIPATH_IBCS_LT_STATE_POLLACTIVE] = 2,
1013         [INFINIPATH_IBCS_LT_STATE_POLLQUIET] = 2,
1014         [INFINIPATH_IBCS_LT_STATE_SLEEPDELAY] = 1,
1015         [INFINIPATH_IBCS_LT_STATE_SLEEPQUIET] = 1,
1016         [INFINIPATH_IBCS_LT_STATE_CFGDEBOUNCE] = 4,
1017         [INFINIPATH_IBCS_LT_STATE_CFGRCVFCFG] = 4,
1018         [INFINIPATH_IBCS_LT_STATE_CFGWAITRMT] = 4,
1019         [INFINIPATH_IBCS_LT_STATE_CFGIDLE] = 4,
1020         [INFINIPATH_IBCS_LT_STATE_RECOVERRETRAIN] = 6,
1021         [INFINIPATH_IBCS_LT_STATE_RECOVERWAITRMT] = 6,
1022         [INFINIPATH_IBCS_LT_STATE_RECOVERIDLE] = 6,
1023 };
1024
1025 u32 ipath_get_cr_errpkey(struct ipath_devdata *dd)
1026 {
1027         return ipath_read_creg32(dd, dd->ipath_cregs->cr_errpkey);
1028 }
1029
1030 static int ipath_query_port(struct ib_device *ibdev,
1031                             u8 port, struct ib_port_attr *props)
1032 {
1033         struct ipath_ibdev *dev = to_idev(ibdev);
1034         enum ib_mtu mtu;
1035         u16 lid = dev->dd->ipath_lid;
1036         u64 ibcstat;
1037
1038         memset(props, 0, sizeof(*props));
1039         props->lid = lid ? lid : __constant_be16_to_cpu(IB_LID_PERMISSIVE);
1040         props->lmc = dev->mkeyprot_resv_lmc & 7;
1041         props->sm_lid = dev->sm_lid;
1042         props->sm_sl = dev->sm_sl;
1043         ibcstat = dev->dd->ipath_lastibcstat;
1044         props->state = ((ibcstat >> 4) & 0x3) + 1;
1045         /* See phys_state_show() */
1046         props->phys_state = ipath_cvt_physportstate[
1047                 dev->dd->ipath_lastibcstat & 0xf];
1048         props->port_cap_flags = dev->port_cap_flags;
1049         props->gid_tbl_len = 1;
1050         props->max_msg_sz = 0x80000000;
1051         props->pkey_tbl_len = ipath_get_npkeys(dev->dd);
1052         props->bad_pkey_cntr = ipath_get_cr_errpkey(dev->dd) -
1053                 dev->z_pkey_violations;
1054         props->qkey_viol_cntr = dev->qkey_violations;
1055         props->active_width = IB_WIDTH_4X;
1056         /* See rate_show() */
1057         props->active_speed = 1;        /* Regular 10Mbs speed. */
1058         props->max_vl_num = 1;          /* VLCap = VL0 */
1059         props->init_type_reply = 0;
1060
1061         /*
1062          * Note: the chips support a maximum MTU of 4096, but the driver
1063          * hasn't implemented this feature yet, so set the maximum value
1064          * to 2048.
1065          */
1066         props->max_mtu = IB_MTU_2048;
1067         switch (dev->dd->ipath_ibmtu) {
1068         case 4096:
1069                 mtu = IB_MTU_4096;
1070                 break;
1071         case 2048:
1072                 mtu = IB_MTU_2048;
1073                 break;
1074         case 1024:
1075                 mtu = IB_MTU_1024;
1076                 break;
1077         case 512:
1078                 mtu = IB_MTU_512;
1079                 break;
1080         case 256:
1081                 mtu = IB_MTU_256;
1082                 break;
1083         default:
1084                 mtu = IB_MTU_2048;
1085         }
1086         props->active_mtu = mtu;
1087         props->subnet_timeout = dev->subnet_timeout;
1088
1089         return 0;
1090 }
1091
1092 static int ipath_modify_device(struct ib_device *device,
1093                                int device_modify_mask,
1094                                struct ib_device_modify *device_modify)
1095 {
1096         int ret;
1097
1098         if (device_modify_mask & ~(IB_DEVICE_MODIFY_SYS_IMAGE_GUID |
1099                                    IB_DEVICE_MODIFY_NODE_DESC)) {
1100                 ret = -EOPNOTSUPP;
1101                 goto bail;
1102         }
1103
1104         if (device_modify_mask & IB_DEVICE_MODIFY_NODE_DESC)
1105                 memcpy(device->node_desc, device_modify->node_desc, 64);
1106
1107         if (device_modify_mask & IB_DEVICE_MODIFY_SYS_IMAGE_GUID)
1108                 to_idev(device)->sys_image_guid =
1109                         cpu_to_be64(device_modify->sys_image_guid);
1110
1111         ret = 0;
1112
1113 bail:
1114         return ret;
1115 }
1116
1117 static int ipath_modify_port(struct ib_device *ibdev,
1118                              u8 port, int port_modify_mask,
1119                              struct ib_port_modify *props)
1120 {
1121         struct ipath_ibdev *dev = to_idev(ibdev);
1122
1123         dev->port_cap_flags |= props->set_port_cap_mask;
1124         dev->port_cap_flags &= ~props->clr_port_cap_mask;
1125         if (port_modify_mask & IB_PORT_SHUTDOWN)
1126                 ipath_set_linkstate(dev->dd, IPATH_IB_LINKDOWN);
1127         if (port_modify_mask & IB_PORT_RESET_QKEY_CNTR)
1128                 dev->qkey_violations = 0;
1129         return 0;
1130 }
1131
1132 static int ipath_query_gid(struct ib_device *ibdev, u8 port,
1133                            int index, union ib_gid *gid)
1134 {
1135         struct ipath_ibdev *dev = to_idev(ibdev);
1136         int ret;
1137
1138         if (index >= 1) {
1139                 ret = -EINVAL;
1140                 goto bail;
1141         }
1142         gid->global.subnet_prefix = dev->gid_prefix;
1143         gid->global.interface_id = dev->dd->ipath_guid;
1144
1145         ret = 0;
1146
1147 bail:
1148         return ret;
1149 }
1150
1151 static struct ib_pd *ipath_alloc_pd(struct ib_device *ibdev,
1152                                     struct ib_ucontext *context,
1153                                     struct ib_udata *udata)
1154 {
1155         struct ipath_ibdev *dev = to_idev(ibdev);
1156         struct ipath_pd *pd;
1157         struct ib_pd *ret;
1158
1159         /*
1160          * This is actually totally arbitrary.  Some correctness tests
1161          * assume there's a maximum number of PDs that can be allocated.
1162          * We don't actually have this limit, but we fail the test if
1163          * we allow allocations of more than we report for this value.
1164          */
1165
1166         pd = kmalloc(sizeof *pd, GFP_KERNEL);
1167         if (!pd) {
1168                 ret = ERR_PTR(-ENOMEM);
1169                 goto bail;
1170         }
1171
1172         spin_lock(&dev->n_pds_lock);
1173         if (dev->n_pds_allocated == ib_ipath_max_pds) {
1174                 spin_unlock(&dev->n_pds_lock);
1175                 kfree(pd);
1176                 ret = ERR_PTR(-ENOMEM);
1177                 goto bail;
1178         }
1179
1180         dev->n_pds_allocated++;
1181         spin_unlock(&dev->n_pds_lock);
1182
1183         /* ib_alloc_pd() will initialize pd->ibpd. */
1184         pd->user = udata != NULL;
1185
1186         ret = &pd->ibpd;
1187
1188 bail:
1189         return ret;
1190 }
1191
1192 static int ipath_dealloc_pd(struct ib_pd *ibpd)
1193 {
1194         struct ipath_pd *pd = to_ipd(ibpd);
1195         struct ipath_ibdev *dev = to_idev(ibpd->device);
1196
1197         spin_lock(&dev->n_pds_lock);
1198         dev->n_pds_allocated--;
1199         spin_unlock(&dev->n_pds_lock);
1200
1201         kfree(pd);
1202
1203         return 0;
1204 }
1205
1206 /**
1207  * ipath_create_ah - create an address handle
1208  * @pd: the protection domain
1209  * @ah_attr: the attributes of the AH
1210  *
1211  * This may be called from interrupt context.
1212  */
1213 static struct ib_ah *ipath_create_ah(struct ib_pd *pd,
1214                                      struct ib_ah_attr *ah_attr)
1215 {
1216         struct ipath_ah *ah;
1217         struct ib_ah *ret;
1218         struct ipath_ibdev *dev = to_idev(pd->device);
1219         unsigned long flags;
1220
1221         /* A multicast address requires a GRH (see ch. 8.4.1). */
1222         if (ah_attr->dlid >= IPATH_MULTICAST_LID_BASE &&
1223             ah_attr->dlid != IPATH_PERMISSIVE_LID &&
1224             !(ah_attr->ah_flags & IB_AH_GRH)) {
1225                 ret = ERR_PTR(-EINVAL);
1226                 goto bail;
1227         }
1228
1229         if (ah_attr->dlid == 0) {
1230                 ret = ERR_PTR(-EINVAL);
1231                 goto bail;
1232         }
1233
1234         if (ah_attr->port_num < 1 ||
1235             ah_attr->port_num > pd->device->phys_port_cnt) {
1236                 ret = ERR_PTR(-EINVAL);
1237                 goto bail;
1238         }
1239
1240         ah = kmalloc(sizeof *ah, GFP_ATOMIC);
1241         if (!ah) {
1242                 ret = ERR_PTR(-ENOMEM);
1243                 goto bail;
1244         }
1245
1246         spin_lock_irqsave(&dev->n_ahs_lock, flags);
1247         if (dev->n_ahs_allocated == ib_ipath_max_ahs) {
1248                 spin_unlock_irqrestore(&dev->n_ahs_lock, flags);
1249                 kfree(ah);
1250                 ret = ERR_PTR(-ENOMEM);
1251                 goto bail;
1252         }
1253
1254         dev->n_ahs_allocated++;
1255         spin_unlock_irqrestore(&dev->n_ahs_lock, flags);
1256
1257         /* ib_create_ah() will initialize ah->ibah. */
1258         ah->attr = *ah_attr;
1259
1260         ret = &ah->ibah;
1261
1262 bail:
1263         return ret;
1264 }
1265
1266 /**
1267  * ipath_destroy_ah - destroy an address handle
1268  * @ibah: the AH to destroy
1269  *
1270  * This may be called from interrupt context.
1271  */
1272 static int ipath_destroy_ah(struct ib_ah *ibah)
1273 {
1274         struct ipath_ibdev *dev = to_idev(ibah->device);
1275         struct ipath_ah *ah = to_iah(ibah);
1276         unsigned long flags;
1277
1278         spin_lock_irqsave(&dev->n_ahs_lock, flags);
1279         dev->n_ahs_allocated--;
1280         spin_unlock_irqrestore(&dev->n_ahs_lock, flags);
1281
1282         kfree(ah);
1283
1284         return 0;
1285 }
1286
1287 static int ipath_query_ah(struct ib_ah *ibah, struct ib_ah_attr *ah_attr)
1288 {
1289         struct ipath_ah *ah = to_iah(ibah);
1290
1291         *ah_attr = ah->attr;
1292
1293         return 0;
1294 }
1295
1296 /**
1297  * ipath_get_npkeys - return the size of the PKEY table for port 0
1298  * @dd: the infinipath device
1299  */
1300 unsigned ipath_get_npkeys(struct ipath_devdata *dd)
1301 {
1302         return ARRAY_SIZE(dd->ipath_pd[0]->port_pkeys);
1303 }
1304
1305 /**
1306  * ipath_get_pkey - return the indexed PKEY from the port 0 PKEY table
1307  * @dd: the infinipath device
1308  * @index: the PKEY index
1309  */
1310 unsigned ipath_get_pkey(struct ipath_devdata *dd, unsigned index)
1311 {
1312         unsigned ret;
1313
1314         if (index >= ARRAY_SIZE(dd->ipath_pd[0]->port_pkeys))
1315                 ret = 0;
1316         else
1317                 ret = dd->ipath_pd[0]->port_pkeys[index];
1318
1319         return ret;
1320 }
1321
1322 static int ipath_query_pkey(struct ib_device *ibdev, u8 port, u16 index,
1323                             u16 *pkey)
1324 {
1325         struct ipath_ibdev *dev = to_idev(ibdev);
1326         int ret;
1327
1328         if (index >= ipath_get_npkeys(dev->dd)) {
1329                 ret = -EINVAL;
1330                 goto bail;
1331         }
1332
1333         *pkey = ipath_get_pkey(dev->dd, index);
1334         ret = 0;
1335
1336 bail:
1337         return ret;
1338 }
1339
1340 /**
1341  * ipath_alloc_ucontext - allocate a ucontest
1342  * @ibdev: the infiniband device
1343  * @udata: not used by the InfiniPath driver
1344  */
1345
1346 static struct ib_ucontext *ipath_alloc_ucontext(struct ib_device *ibdev,
1347                                                 struct ib_udata *udata)
1348 {
1349         struct ipath_ucontext *context;
1350         struct ib_ucontext *ret;
1351
1352         context = kmalloc(sizeof *context, GFP_KERNEL);
1353         if (!context) {
1354                 ret = ERR_PTR(-ENOMEM);
1355                 goto bail;
1356         }
1357
1358         ret = &context->ibucontext;
1359
1360 bail:
1361         return ret;
1362 }
1363
1364 static int ipath_dealloc_ucontext(struct ib_ucontext *context)
1365 {
1366         kfree(to_iucontext(context));
1367         return 0;
1368 }
1369
1370 static int ipath_verbs_register_sysfs(struct ib_device *dev);
1371
1372 static void __verbs_timer(unsigned long arg)
1373 {
1374         struct ipath_devdata *dd = (struct ipath_devdata *) arg;
1375
1376         /*
1377          * If port 0 receive packet interrupts are not available, or
1378          * can be missed, poll the receive queue
1379          */
1380         if (dd->ipath_flags & IPATH_POLL_RX_INTR)
1381                 ipath_kreceive(dd);
1382
1383         /* Handle verbs layer timeouts. */
1384         ipath_ib_timer(dd->verbs_dev);
1385
1386         mod_timer(&dd->verbs_timer, jiffies + 1);
1387 }
1388
1389 static int enable_timer(struct ipath_devdata *dd)
1390 {
1391         /*
1392          * Early chips had a design flaw where the chip and kernel idea
1393          * of the tail register don't always agree, and therefore we won't
1394          * get an interrupt on the next packet received.
1395          * If the board supports per packet receive interrupts, use it.
1396          * Otherwise, the timer function periodically checks for packets
1397          * to cover this case.
1398          * Either way, the timer is needed for verbs layer related
1399          * processing.
1400          */
1401         if (dd->ipath_flags & IPATH_GPIO_INTR) {
1402                 ipath_write_kreg(dd, dd->ipath_kregs->kr_debugportselect,
1403                                  0x2074076542310ULL);
1404                 /* Enable GPIO bit 2 interrupt */
1405                 dd->ipath_gpio_mask |= (u64) (1 << IPATH_GPIO_PORT0_BIT);
1406                 ipath_write_kreg(dd, dd->ipath_kregs->kr_gpio_mask,
1407                                  dd->ipath_gpio_mask);
1408         }
1409
1410         init_timer(&dd->verbs_timer);
1411         dd->verbs_timer.function = __verbs_timer;
1412         dd->verbs_timer.data = (unsigned long)dd;
1413         dd->verbs_timer.expires = jiffies + 1;
1414         add_timer(&dd->verbs_timer);
1415
1416         return 0;
1417 }
1418
1419 static int disable_timer(struct ipath_devdata *dd)
1420 {
1421         /* Disable GPIO bit 2 interrupt */
1422         if (dd->ipath_flags & IPATH_GPIO_INTR) {
1423                 u64 val;
1424                 /* Disable GPIO bit 2 interrupt */
1425                 val = ipath_read_kreg64(dd, dd->ipath_kregs->kr_gpio_mask);
1426                 dd->ipath_gpio_mask &= ~((u64) (1 << IPATH_GPIO_PORT0_BIT));
1427                 ipath_write_kreg(dd, dd->ipath_kregs->kr_gpio_mask,
1428                                  dd->ipath_gpio_mask);
1429                 /*
1430                  * We might want to undo changes to debugportselect,
1431                  * but how?
1432                  */
1433         }
1434
1435         del_timer_sync(&dd->verbs_timer);
1436
1437         return 0;
1438 }
1439
1440 /**
1441  * ipath_register_ib_device - register our device with the infiniband core
1442  * @dd: the device data structure
1443  * Return the allocated ipath_ibdev pointer or NULL on error.
1444  */
1445 int ipath_register_ib_device(struct ipath_devdata *dd)
1446 {
1447         struct ipath_verbs_counters cntrs;
1448         struct ipath_ibdev *idev;
1449         struct ib_device *dev;
1450         int ret;
1451
1452         idev = (struct ipath_ibdev *)ib_alloc_device(sizeof *idev);
1453         if (idev == NULL) {
1454                 ret = -ENOMEM;
1455                 goto bail;
1456         }
1457
1458         dev = &idev->ibdev;
1459
1460         /* Only need to initialize non-zero fields. */
1461         spin_lock_init(&idev->n_pds_lock);
1462         spin_lock_init(&idev->n_ahs_lock);
1463         spin_lock_init(&idev->n_cqs_lock);
1464         spin_lock_init(&idev->n_qps_lock);
1465         spin_lock_init(&idev->n_srqs_lock);
1466         spin_lock_init(&idev->n_mcast_grps_lock);
1467
1468         spin_lock_init(&idev->qp_table.lock);
1469         spin_lock_init(&idev->lk_table.lock);
1470         idev->sm_lid = __constant_be16_to_cpu(IB_LID_PERMISSIVE);
1471         /* Set the prefix to the default value (see ch. 4.1.1) */
1472         idev->gid_prefix = __constant_cpu_to_be64(0xfe80000000000000ULL);
1473
1474         ret = ipath_init_qp_table(idev, ib_ipath_qp_table_size);
1475         if (ret)
1476                 goto err_qp;
1477
1478         /*
1479          * The top ib_ipath_lkey_table_size bits are used to index the
1480          * table.  The lower 8 bits can be owned by the user (copied from
1481          * the LKEY).  The remaining bits act as a generation number or tag.
1482          */
1483         idev->lk_table.max = 1 << ib_ipath_lkey_table_size;
1484         idev->lk_table.table = kzalloc(idev->lk_table.max *
1485                                        sizeof(*idev->lk_table.table),
1486                                        GFP_KERNEL);
1487         if (idev->lk_table.table == NULL) {
1488                 ret = -ENOMEM;
1489                 goto err_lk;
1490         }
1491         INIT_LIST_HEAD(&idev->pending_mmaps);
1492         spin_lock_init(&idev->pending_lock);
1493         idev->mmap_offset = PAGE_SIZE;
1494         spin_lock_init(&idev->mmap_offset_lock);
1495         INIT_LIST_HEAD(&idev->pending[0]);
1496         INIT_LIST_HEAD(&idev->pending[1]);
1497         INIT_LIST_HEAD(&idev->pending[2]);
1498         INIT_LIST_HEAD(&idev->piowait);
1499         INIT_LIST_HEAD(&idev->rnrwait);
1500         idev->pending_index = 0;
1501         idev->port_cap_flags =
1502                 IB_PORT_SYS_IMAGE_GUID_SUP | IB_PORT_CLIENT_REG_SUP;
1503         idev->pma_counter_select[0] = IB_PMA_PORT_XMIT_DATA;
1504         idev->pma_counter_select[1] = IB_PMA_PORT_RCV_DATA;
1505         idev->pma_counter_select[2] = IB_PMA_PORT_XMIT_PKTS;
1506         idev->pma_counter_select[3] = IB_PMA_PORT_RCV_PKTS;
1507         idev->pma_counter_select[4] = IB_PMA_PORT_XMIT_WAIT;
1508         idev->link_width_enabled = 3;   /* 1x or 4x */
1509
1510         /* Snapshot current HW counters to "clear" them. */
1511         ipath_get_counters(dd, &cntrs);
1512         idev->z_symbol_error_counter = cntrs.symbol_error_counter;
1513         idev->z_link_error_recovery_counter =
1514                 cntrs.link_error_recovery_counter;
1515         idev->z_link_downed_counter = cntrs.link_downed_counter;
1516         idev->z_port_rcv_errors = cntrs.port_rcv_errors;
1517         idev->z_port_rcv_remphys_errors =
1518                 cntrs.port_rcv_remphys_errors;
1519         idev->z_port_xmit_discards = cntrs.port_xmit_discards;
1520         idev->z_port_xmit_data = cntrs.port_xmit_data;
1521         idev->z_port_rcv_data = cntrs.port_rcv_data;
1522         idev->z_port_xmit_packets = cntrs.port_xmit_packets;
1523         idev->z_port_rcv_packets = cntrs.port_rcv_packets;
1524         idev->z_local_link_integrity_errors =
1525                 cntrs.local_link_integrity_errors;
1526         idev->z_excessive_buffer_overrun_errors =
1527                 cntrs.excessive_buffer_overrun_errors;
1528
1529         /*
1530          * The system image GUID is supposed to be the same for all
1531          * IB HCAs in a single system but since there can be other
1532          * device types in the system, we can't be sure this is unique.
1533          */
1534         if (!sys_image_guid)
1535                 sys_image_guid = dd->ipath_guid;
1536         idev->sys_image_guid = sys_image_guid;
1537         idev->ib_unit = dd->ipath_unit;
1538         idev->dd = dd;
1539
1540         strlcpy(dev->name, "ipath%d", IB_DEVICE_NAME_MAX);
1541         dev->owner = THIS_MODULE;
1542         dev->node_guid = dd->ipath_guid;
1543         dev->uverbs_abi_ver = IPATH_UVERBS_ABI_VERSION;
1544         dev->uverbs_cmd_mask =
1545                 (1ull << IB_USER_VERBS_CMD_GET_CONTEXT)         |
1546                 (1ull << IB_USER_VERBS_CMD_QUERY_DEVICE)        |
1547                 (1ull << IB_USER_VERBS_CMD_QUERY_PORT)          |
1548                 (1ull << IB_USER_VERBS_CMD_ALLOC_PD)            |
1549                 (1ull << IB_USER_VERBS_CMD_DEALLOC_PD)          |
1550                 (1ull << IB_USER_VERBS_CMD_CREATE_AH)           |
1551                 (1ull << IB_USER_VERBS_CMD_DESTROY_AH)          |
1552                 (1ull << IB_USER_VERBS_CMD_QUERY_AH)            |
1553                 (1ull << IB_USER_VERBS_CMD_REG_MR)              |
1554                 (1ull << IB_USER_VERBS_CMD_DEREG_MR)            |
1555                 (1ull << IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL) |
1556                 (1ull << IB_USER_VERBS_CMD_CREATE_CQ)           |
1557                 (1ull << IB_USER_VERBS_CMD_RESIZE_CQ)           |
1558                 (1ull << IB_USER_VERBS_CMD_DESTROY_CQ)          |
1559                 (1ull << IB_USER_VERBS_CMD_POLL_CQ)             |
1560                 (1ull << IB_USER_VERBS_CMD_REQ_NOTIFY_CQ)       |
1561                 (1ull << IB_USER_VERBS_CMD_CREATE_QP)           |
1562                 (1ull << IB_USER_VERBS_CMD_QUERY_QP)            |
1563                 (1ull << IB_USER_VERBS_CMD_MODIFY_QP)           |
1564                 (1ull << IB_USER_VERBS_CMD_DESTROY_QP)          |
1565                 (1ull << IB_USER_VERBS_CMD_POST_SEND)           |
1566                 (1ull << IB_USER_VERBS_CMD_POST_RECV)           |
1567                 (1ull << IB_USER_VERBS_CMD_ATTACH_MCAST)        |
1568                 (1ull << IB_USER_VERBS_CMD_DETACH_MCAST)        |
1569                 (1ull << IB_USER_VERBS_CMD_CREATE_SRQ)          |
1570                 (1ull << IB_USER_VERBS_CMD_MODIFY_SRQ)          |
1571                 (1ull << IB_USER_VERBS_CMD_QUERY_SRQ)           |
1572                 (1ull << IB_USER_VERBS_CMD_DESTROY_SRQ)         |
1573                 (1ull << IB_USER_VERBS_CMD_POST_SRQ_RECV);
1574         dev->node_type = RDMA_NODE_IB_CA;
1575         dev->phys_port_cnt = 1;
1576         dev->num_comp_vectors = 1;
1577         dev->dma_device = &dd->pcidev->dev;
1578         dev->query_device = ipath_query_device;
1579         dev->modify_device = ipath_modify_device;
1580         dev->query_port = ipath_query_port;
1581         dev->modify_port = ipath_modify_port;
1582         dev->query_pkey = ipath_query_pkey;
1583         dev->query_gid = ipath_query_gid;
1584         dev->alloc_ucontext = ipath_alloc_ucontext;
1585         dev->dealloc_ucontext = ipath_dealloc_ucontext;
1586         dev->alloc_pd = ipath_alloc_pd;
1587         dev->dealloc_pd = ipath_dealloc_pd;
1588         dev->create_ah = ipath_create_ah;
1589         dev->destroy_ah = ipath_destroy_ah;
1590         dev->query_ah = ipath_query_ah;
1591         dev->create_srq = ipath_create_srq;
1592         dev->modify_srq = ipath_modify_srq;
1593         dev->query_srq = ipath_query_srq;
1594         dev->destroy_srq = ipath_destroy_srq;
1595         dev->create_qp = ipath_create_qp;
1596         dev->modify_qp = ipath_modify_qp;
1597         dev->query_qp = ipath_query_qp;
1598         dev->destroy_qp = ipath_destroy_qp;
1599         dev->post_send = ipath_post_send;
1600         dev->post_recv = ipath_post_receive;
1601         dev->post_srq_recv = ipath_post_srq_receive;
1602         dev->create_cq = ipath_create_cq;
1603         dev->destroy_cq = ipath_destroy_cq;
1604         dev->resize_cq = ipath_resize_cq;
1605         dev->poll_cq = ipath_poll_cq;
1606         dev->req_notify_cq = ipath_req_notify_cq;
1607         dev->get_dma_mr = ipath_get_dma_mr;
1608         dev->reg_phys_mr = ipath_reg_phys_mr;
1609         dev->reg_user_mr = ipath_reg_user_mr;
1610         dev->dereg_mr = ipath_dereg_mr;
1611         dev->alloc_fmr = ipath_alloc_fmr;
1612         dev->map_phys_fmr = ipath_map_phys_fmr;
1613         dev->unmap_fmr = ipath_unmap_fmr;
1614         dev->dealloc_fmr = ipath_dealloc_fmr;
1615         dev->attach_mcast = ipath_multicast_attach;
1616         dev->detach_mcast = ipath_multicast_detach;
1617         dev->process_mad = ipath_process_mad;
1618         dev->mmap = ipath_mmap;
1619         dev->dma_ops = &ipath_dma_mapping_ops;
1620
1621         snprintf(dev->node_desc, sizeof(dev->node_desc),
1622                  IPATH_IDSTR " %s", init_utsname()->nodename);
1623
1624         ret = ib_register_device(dev);
1625         if (ret)
1626                 goto err_reg;
1627
1628         if (ipath_verbs_register_sysfs(dev))
1629                 goto err_class;
1630
1631         enable_timer(dd);
1632
1633         goto bail;
1634
1635 err_class:
1636         ib_unregister_device(dev);
1637 err_reg:
1638         kfree(idev->lk_table.table);
1639 err_lk:
1640         kfree(idev->qp_table.table);
1641 err_qp:
1642         ib_dealloc_device(dev);
1643         ipath_dev_err(dd, "cannot register verbs: %d!\n", -ret);
1644         idev = NULL;
1645
1646 bail:
1647         dd->verbs_dev = idev;
1648         return ret;
1649 }
1650
1651 void ipath_unregister_ib_device(struct ipath_ibdev *dev)
1652 {
1653         struct ib_device *ibdev = &dev->ibdev;
1654
1655         disable_timer(dev->dd);
1656
1657         ib_unregister_device(ibdev);
1658
1659         if (!list_empty(&dev->pending[0]) ||
1660             !list_empty(&dev->pending[1]) ||
1661             !list_empty(&dev->pending[2]))
1662                 ipath_dev_err(dev->dd, "pending list not empty!\n");
1663         if (!list_empty(&dev->piowait))
1664                 ipath_dev_err(dev->dd, "piowait list not empty!\n");
1665         if (!list_empty(&dev->rnrwait))
1666                 ipath_dev_err(dev->dd, "rnrwait list not empty!\n");
1667         if (!ipath_mcast_tree_empty())
1668                 ipath_dev_err(dev->dd, "multicast table memory leak!\n");
1669         /*
1670          * Note that ipath_unregister_ib_device() can be called before all
1671          * the QPs are destroyed!
1672          */
1673         ipath_free_all_qps(&dev->qp_table);
1674         kfree(dev->qp_table.table);
1675         kfree(dev->lk_table.table);
1676         ib_dealloc_device(ibdev);
1677 }
1678
1679 static ssize_t show_rev(struct class_device *cdev, char *buf)
1680 {
1681         struct ipath_ibdev *dev =
1682                 container_of(cdev, struct ipath_ibdev, ibdev.class_dev);
1683
1684         return sprintf(buf, "%x\n", dev->dd->ipath_pcirev);
1685 }
1686
1687 static ssize_t show_hca(struct class_device *cdev, char *buf)
1688 {
1689         struct ipath_ibdev *dev =
1690                 container_of(cdev, struct ipath_ibdev, ibdev.class_dev);
1691         int ret;
1692
1693         ret = dev->dd->ipath_f_get_boardname(dev->dd, buf, 128);
1694         if (ret < 0)
1695                 goto bail;
1696         strcat(buf, "\n");
1697         ret = strlen(buf);
1698
1699 bail:
1700         return ret;
1701 }
1702
1703 static ssize_t show_stats(struct class_device *cdev, char *buf)
1704 {
1705         struct ipath_ibdev *dev =
1706                 container_of(cdev, struct ipath_ibdev, ibdev.class_dev);
1707         int i;
1708         int len;
1709
1710         len = sprintf(buf,
1711                       "RC resends  %d\n"
1712                       "RC no QACK  %d\n"
1713                       "RC ACKs     %d\n"
1714                       "RC SEQ NAKs %d\n"
1715                       "RC RDMA seq %d\n"
1716                       "RC RNR NAKs %d\n"
1717                       "RC OTH NAKs %d\n"
1718                       "RC timeouts %d\n"
1719                       "RC RDMA dup %d\n"
1720                       "RC stalls   %d\n"
1721                       "piobuf wait %d\n"
1722                       "no piobuf   %d\n"
1723                       "PKT drops   %d\n"
1724                       "WQE errs    %d\n",
1725                       dev->n_rc_resends, dev->n_rc_qacks, dev->n_rc_acks,
1726                       dev->n_seq_naks, dev->n_rdma_seq, dev->n_rnr_naks,
1727                       dev->n_other_naks, dev->n_timeouts,
1728                       dev->n_rdma_dup_busy, dev->n_rc_stalls, dev->n_piowait,
1729                       dev->n_no_piobuf, dev->n_pkt_drops, dev->n_wqe_errs);
1730         for (i = 0; i < ARRAY_SIZE(dev->opstats); i++) {
1731                 const struct ipath_opcode_stats *si = &dev->opstats[i];
1732
1733                 if (!si->n_packets && !si->n_bytes)
1734                         continue;
1735                 len += sprintf(buf + len, "%02x %llu/%llu\n", i,
1736                                (unsigned long long) si->n_packets,
1737                                (unsigned long long) si->n_bytes);
1738         }
1739         return len;
1740 }
1741
1742 static CLASS_DEVICE_ATTR(hw_rev, S_IRUGO, show_rev, NULL);
1743 static CLASS_DEVICE_ATTR(hca_type, S_IRUGO, show_hca, NULL);
1744 static CLASS_DEVICE_ATTR(board_id, S_IRUGO, show_hca, NULL);
1745 static CLASS_DEVICE_ATTR(stats, S_IRUGO, show_stats, NULL);
1746
1747 static struct class_device_attribute *ipath_class_attributes[] = {
1748         &class_device_attr_hw_rev,
1749         &class_device_attr_hca_type,
1750         &class_device_attr_board_id,
1751         &class_device_attr_stats
1752 };
1753
1754 static int ipath_verbs_register_sysfs(struct ib_device *dev)
1755 {
1756         int i;
1757         int ret;
1758
1759         for (i = 0; i < ARRAY_SIZE(ipath_class_attributes); ++i)
1760                 if (class_device_create_file(&dev->class_dev,
1761                                              ipath_class_attributes[i])) {
1762                         ret = 1;
1763                         goto bail;
1764                 }
1765
1766         ret = 0;
1767
1768 bail:
1769         return ret;
1770 }