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