Merge master.kernel.org:/pub/scm/linux/kernel/git/gregkh/driver-2.6
[linux-2.6] / drivers / infiniband / hw / ipath / ipath_verbs.c
1 /*
2  * Copyright (c) 2006 QLogic, Inc. All rights reserved.
3  * Copyright (c) 2005, 2006 PathScale, Inc. All rights reserved.
4  *
5  * This software is available to you under a choice of one of two
6  * licenses.  You may choose to be licensed under the terms of the GNU
7  * General Public License (GPL) Version 2, available from the file
8  * COPYING in the main directory of this source tree, or the
9  * OpenIB.org BSD license below:
10  *
11  *     Redistribution and use in source and binary forms, with or
12  *     without modification, are permitted provided that the following
13  *     conditions are met:
14  *
15  *      - Redistributions of source code must retain the above
16  *        copyright notice, this list of conditions and the following
17  *        disclaimer.
18  *
19  *      - Redistributions in binary form must reproduce the above
20  *        copyright notice, this list of conditions and the following
21  *        disclaimer in the documentation and/or other materials
22  *        provided with the distribution.
23  *
24  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
25  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
26  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
27  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
28  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
29  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
30  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
31  * SOFTWARE.
32  */
33
34 #include <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                 BUG_ON(len == 0);
168                 if (len > length)
169                         len = length;
170                 memcpy(sge->vaddr, data, len);
171                 sge->vaddr += len;
172                 sge->length -= len;
173                 sge->sge_length -= len;
174                 if (sge->sge_length == 0) {
175                         if (--ss->num_sge)
176                                 *sge = *ss->sg_list++;
177                 } else if (sge->length == 0 && sge->mr != NULL) {
178                         if (++sge->n >= IPATH_SEGSZ) {
179                                 if (++sge->m >= sge->mr->mapsz)
180                                         break;
181                                 sge->n = 0;
182                         }
183                         sge->vaddr =
184                                 sge->mr->map[sge->m]->segs[sge->n].vaddr;
185                         sge->length =
186                                 sge->mr->map[sge->m]->segs[sge->n].length;
187                 }
188                 data += len;
189                 length -= len;
190         }
191 }
192
193 /**
194  * ipath_skip_sge - skip over SGE memory - XXX almost dup of prev func
195  * @ss: the SGE state
196  * @length: the number of bytes to skip
197  */
198 void ipath_skip_sge(struct ipath_sge_state *ss, u32 length)
199 {
200         struct ipath_sge *sge = &ss->sge;
201
202         while (length) {
203                 u32 len = sge->length;
204
205                 BUG_ON(len == 0);
206                 if (len > length)
207                         len = length;
208                 sge->vaddr += len;
209                 sge->length -= len;
210                 sge->sge_length -= len;
211                 if (sge->sge_length == 0) {
212                         if (--ss->num_sge)
213                                 *sge = *ss->sg_list++;
214                 } else if (sge->length == 0 && sge->mr != NULL) {
215                         if (++sge->n >= IPATH_SEGSZ) {
216                                 if (++sge->m >= sge->mr->mapsz)
217                                         break;
218                                 sge->n = 0;
219                         }
220                         sge->vaddr =
221                                 sge->mr->map[sge->m]->segs[sge->n].vaddr;
222                         sge->length =
223                                 sge->mr->map[sge->m]->segs[sge->n].length;
224                 }
225                 length -= len;
226         }
227 }
228
229 /**
230  * ipath_post_send - post a send on a QP
231  * @ibqp: the QP to post the send on
232  * @wr: the list of work requests to post
233  * @bad_wr: the first bad WR is put here
234  *
235  * This may be called from interrupt context.
236  */
237 static int ipath_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
238                            struct ib_send_wr **bad_wr)
239 {
240         struct ipath_qp *qp = to_iqp(ibqp);
241         int err = 0;
242
243         /* Check that state is OK to post send. */
244         if (!(ib_ipath_state_ops[qp->state] & IPATH_POST_SEND_OK)) {
245                 *bad_wr = wr;
246                 err = -EINVAL;
247                 goto bail;
248         }
249
250         for (; wr; wr = wr->next) {
251                 switch (qp->ibqp.qp_type) {
252                 case IB_QPT_UC:
253                 case IB_QPT_RC:
254                         err = ipath_post_ruc_send(qp, wr);
255                         break;
256
257                 case IB_QPT_SMI:
258                 case IB_QPT_GSI:
259                 case IB_QPT_UD:
260                         err = ipath_post_ud_send(qp, wr);
261                         break;
262
263                 default:
264                         err = -EINVAL;
265                 }
266                 if (err) {
267                         *bad_wr = wr;
268                         break;
269                 }
270         }
271
272 bail:
273         return err;
274 }
275
276 /**
277  * ipath_post_receive - post a receive on a QP
278  * @ibqp: the QP to post the receive on
279  * @wr: the WR to post
280  * @bad_wr: the first bad WR is put here
281  *
282  * This may be called from interrupt context.
283  */
284 static int ipath_post_receive(struct ib_qp *ibqp, struct ib_recv_wr *wr,
285                               struct ib_recv_wr **bad_wr)
286 {
287         struct ipath_qp *qp = to_iqp(ibqp);
288         struct ipath_rwq *wq = qp->r_rq.wq;
289         unsigned long flags;
290         int ret;
291
292         /* Check that state is OK to post receive. */
293         if (!(ib_ipath_state_ops[qp->state] & IPATH_POST_RECV_OK) || !wq) {
294                 *bad_wr = wr;
295                 ret = -EINVAL;
296                 goto bail;
297         }
298
299         for (; wr; wr = wr->next) {
300                 struct ipath_rwqe *wqe;
301                 u32 next;
302                 int i;
303
304                 if ((unsigned) wr->num_sge > qp->r_rq.max_sge) {
305                         *bad_wr = wr;
306                         ret = -ENOMEM;
307                         goto bail;
308                 }
309
310                 spin_lock_irqsave(&qp->r_rq.lock, flags);
311                 next = wq->head + 1;
312                 if (next >= qp->r_rq.size)
313                         next = 0;
314                 if (next == wq->tail) {
315                         spin_unlock_irqrestore(&qp->r_rq.lock, flags);
316                         *bad_wr = wr;
317                         ret = -ENOMEM;
318                         goto bail;
319                 }
320
321                 wqe = get_rwqe_ptr(&qp->r_rq, wq->head);
322                 wqe->wr_id = wr->wr_id;
323                 wqe->num_sge = wr->num_sge;
324                 for (i = 0; i < wr->num_sge; i++)
325                         wqe->sg_list[i] = wr->sg_list[i];
326                 wq->head = next;
327                 spin_unlock_irqrestore(&qp->r_rq.lock, flags);
328         }
329         ret = 0;
330
331 bail:
332         return ret;
333 }
334
335 /**
336  * ipath_qp_rcv - processing an incoming packet on a QP
337  * @dev: the device the packet came on
338  * @hdr: the packet header
339  * @has_grh: true if the packet has a GRH
340  * @data: the packet data
341  * @tlen: the packet length
342  * @qp: the QP the packet came on
343  *
344  * This is called from ipath_ib_rcv() to process an incoming packet
345  * for the given QP.
346  * Called at interrupt level.
347  */
348 static void ipath_qp_rcv(struct ipath_ibdev *dev,
349                          struct ipath_ib_header *hdr, int has_grh,
350                          void *data, u32 tlen, struct ipath_qp *qp)
351 {
352         /* Check for valid receive state. */
353         if (!(ib_ipath_state_ops[qp->state] & IPATH_PROCESS_RECV_OK)) {
354                 dev->n_pkt_drops++;
355                 return;
356         }
357
358         switch (qp->ibqp.qp_type) {
359         case IB_QPT_SMI:
360         case IB_QPT_GSI:
361                 if (ib_ipath_disable_sma)
362                         break;
363                 /* FALLTHROUGH */
364         case IB_QPT_UD:
365                 ipath_ud_rcv(dev, hdr, has_grh, data, tlen, qp);
366                 break;
367
368         case IB_QPT_RC:
369                 ipath_rc_rcv(dev, hdr, has_grh, data, tlen, qp);
370                 break;
371
372         case IB_QPT_UC:
373                 ipath_uc_rcv(dev, hdr, has_grh, data, tlen, qp);
374                 break;
375
376         default:
377                 break;
378         }
379 }
380
381 /**
382  * ipath_ib_rcv - process an incoming packet
383  * @arg: the device pointer
384  * @rhdr: the header of the packet
385  * @data: the packet data
386  * @tlen: the packet length
387  *
388  * This is called from ipath_kreceive() to process an incoming packet at
389  * interrupt level. Tlen is the length of the header + data + CRC in bytes.
390  */
391 void ipath_ib_rcv(struct ipath_ibdev *dev, void *rhdr, void *data,
392                   u32 tlen)
393 {
394         struct ipath_ib_header *hdr = rhdr;
395         struct ipath_other_headers *ohdr;
396         struct ipath_qp *qp;
397         u32 qp_num;
398         int lnh;
399         u8 opcode;
400         u16 lid;
401
402         if (unlikely(dev == NULL))
403                 goto bail;
404
405         if (unlikely(tlen < 24)) {      /* LRH+BTH+CRC */
406                 dev->rcv_errors++;
407                 goto bail;
408         }
409
410         /* Check for a valid destination LID (see ch. 7.11.1). */
411         lid = be16_to_cpu(hdr->lrh[1]);
412         if (lid < IPATH_MULTICAST_LID_BASE) {
413                 lid &= ~((1 << (dev->mkeyprot_resv_lmc & 7)) - 1);
414                 if (unlikely(lid != dev->dd->ipath_lid)) {
415                         dev->rcv_errors++;
416                         goto bail;
417                 }
418         }
419
420         /* Check for GRH */
421         lnh = be16_to_cpu(hdr->lrh[0]) & 3;
422         if (lnh == IPATH_LRH_BTH)
423                 ohdr = &hdr->u.oth;
424         else if (lnh == IPATH_LRH_GRH)
425                 ohdr = &hdr->u.l.oth;
426         else {
427                 dev->rcv_errors++;
428                 goto bail;
429         }
430
431         opcode = be32_to_cpu(ohdr->bth[0]) >> 24;
432         dev->opstats[opcode].n_bytes += tlen;
433         dev->opstats[opcode].n_packets++;
434
435         /* Get the destination QP number. */
436         qp_num = be32_to_cpu(ohdr->bth[1]) & IPATH_QPN_MASK;
437         if (qp_num == IPATH_MULTICAST_QPN) {
438                 struct ipath_mcast *mcast;
439                 struct ipath_mcast_qp *p;
440
441                 if (lnh != IPATH_LRH_GRH) {
442                         dev->n_pkt_drops++;
443                         goto bail;
444                 }
445                 mcast = ipath_mcast_find(&hdr->u.l.grh.dgid);
446                 if (mcast == NULL) {
447                         dev->n_pkt_drops++;
448                         goto bail;
449                 }
450                 dev->n_multicast_rcv++;
451                 list_for_each_entry_rcu(p, &mcast->qp_list, list)
452                         ipath_qp_rcv(dev, hdr, 1, data, tlen, p->qp);
453                 /*
454                  * Notify ipath_multicast_detach() if it is waiting for us
455                  * to finish.
456                  */
457                 if (atomic_dec_return(&mcast->refcount) <= 1)
458                         wake_up(&mcast->wait);
459         } else {
460                 qp = ipath_lookup_qpn(&dev->qp_table, qp_num);
461                 if (qp) {
462                         dev->n_unicast_rcv++;
463                         ipath_qp_rcv(dev, hdr, lnh == IPATH_LRH_GRH, data,
464                                      tlen, qp);
465                         /*
466                          * Notify ipath_destroy_qp() if it is waiting
467                          * for us to finish.
468                          */
469                         if (atomic_dec_and_test(&qp->refcount))
470                                 wake_up(&qp->wait);
471                 } else
472                         dev->n_pkt_drops++;
473         }
474
475 bail:;
476 }
477
478 /**
479  * ipath_ib_timer - verbs timer
480  * @arg: the device pointer
481  *
482  * This is called from ipath_do_rcv_timer() at interrupt level to check for
483  * QPs which need retransmits and to collect performance numbers.
484  */
485 void ipath_ib_timer(struct ipath_ibdev *dev)
486 {
487         struct ipath_qp *resend = NULL;
488         struct list_head *last;
489         struct ipath_qp *qp;
490         unsigned long flags;
491
492         if (dev == NULL)
493                 return;
494
495         spin_lock_irqsave(&dev->pending_lock, flags);
496         /* Start filling the next pending queue. */
497         if (++dev->pending_index >= ARRAY_SIZE(dev->pending))
498                 dev->pending_index = 0;
499         /* Save any requests still in the new queue, they have timed out. */
500         last = &dev->pending[dev->pending_index];
501         while (!list_empty(last)) {
502                 qp = list_entry(last->next, struct ipath_qp, timerwait);
503                 list_del_init(&qp->timerwait);
504                 qp->timer_next = resend;
505                 resend = qp;
506                 atomic_inc(&qp->refcount);
507         }
508         last = &dev->rnrwait;
509         if (!list_empty(last)) {
510                 qp = list_entry(last->next, struct ipath_qp, timerwait);
511                 if (--qp->s_rnr_timeout == 0) {
512                         do {
513                                 list_del_init(&qp->timerwait);
514                                 tasklet_hi_schedule(&qp->s_task);
515                                 if (list_empty(last))
516                                         break;
517                                 qp = list_entry(last->next, struct ipath_qp,
518                                                 timerwait);
519                         } while (qp->s_rnr_timeout == 0);
520                 }
521         }
522         /*
523          * We should only be in the started state if pma_sample_start != 0
524          */
525         if (dev->pma_sample_status == IB_PMA_SAMPLE_STATUS_STARTED &&
526             --dev->pma_sample_start == 0) {
527                 dev->pma_sample_status = IB_PMA_SAMPLE_STATUS_RUNNING;
528                 ipath_snapshot_counters(dev->dd, &dev->ipath_sword,
529                                         &dev->ipath_rword,
530                                         &dev->ipath_spkts,
531                                         &dev->ipath_rpkts,
532                                         &dev->ipath_xmit_wait);
533         }
534         if (dev->pma_sample_status == IB_PMA_SAMPLE_STATUS_RUNNING) {
535                 if (dev->pma_sample_interval == 0) {
536                         u64 ta, tb, tc, td, te;
537
538                         dev->pma_sample_status = IB_PMA_SAMPLE_STATUS_DONE;
539                         ipath_snapshot_counters(dev->dd, &ta, &tb,
540                                                 &tc, &td, &te);
541
542                         dev->ipath_sword = ta - dev->ipath_sword;
543                         dev->ipath_rword = tb - dev->ipath_rword;
544                         dev->ipath_spkts = tc - dev->ipath_spkts;
545                         dev->ipath_rpkts = td - dev->ipath_rpkts;
546                         dev->ipath_xmit_wait = te - dev->ipath_xmit_wait;
547                 }
548                 else
549                         dev->pma_sample_interval--;
550         }
551         spin_unlock_irqrestore(&dev->pending_lock, flags);
552
553         /* XXX What if timer fires again while this is running? */
554         for (qp = resend; qp != NULL; qp = qp->timer_next) {
555                 struct ib_wc wc;
556
557                 spin_lock_irqsave(&qp->s_lock, flags);
558                 if (qp->s_last != qp->s_tail && qp->state == IB_QPS_RTS) {
559                         dev->n_timeouts++;
560                         ipath_restart_rc(qp, qp->s_last_psn + 1, &wc);
561                 }
562                 spin_unlock_irqrestore(&qp->s_lock, flags);
563
564                 /* Notify ipath_destroy_qp() if it is waiting. */
565                 if (atomic_dec_and_test(&qp->refcount))
566                         wake_up(&qp->wait);
567         }
568 }
569
570 static void update_sge(struct ipath_sge_state *ss, u32 length)
571 {
572         struct ipath_sge *sge = &ss->sge;
573
574         sge->vaddr += length;
575         sge->length -= length;
576         sge->sge_length -= length;
577         if (sge->sge_length == 0) {
578                 if (--ss->num_sge)
579                         *sge = *ss->sg_list++;
580         } else if (sge->length == 0 && sge->mr != NULL) {
581                 if (++sge->n >= IPATH_SEGSZ) {
582                         if (++sge->m >= sge->mr->mapsz)
583                                 return;
584                         sge->n = 0;
585                 }
586                 sge->vaddr = sge->mr->map[sge->m]->segs[sge->n].vaddr;
587                 sge->length = sge->mr->map[sge->m]->segs[sge->n].length;
588         }
589 }
590
591 #ifdef __LITTLE_ENDIAN
592 static inline u32 get_upper_bits(u32 data, u32 shift)
593 {
594         return data >> shift;
595 }
596
597 static inline u32 set_upper_bits(u32 data, u32 shift)
598 {
599         return data << shift;
600 }
601
602 static inline u32 clear_upper_bytes(u32 data, u32 n, u32 off)
603 {
604         data <<= ((sizeof(u32) - n) * BITS_PER_BYTE);
605         data >>= ((sizeof(u32) - n - off) * BITS_PER_BYTE);
606         return data;
607 }
608 #else
609 static inline u32 get_upper_bits(u32 data, u32 shift)
610 {
611         return data << shift;
612 }
613
614 static inline u32 set_upper_bits(u32 data, u32 shift)
615 {
616         return data >> shift;
617 }
618
619 static inline u32 clear_upper_bytes(u32 data, u32 n, u32 off)
620 {
621         data >>= ((sizeof(u32) - n) * BITS_PER_BYTE);
622         data <<= ((sizeof(u32) - n - off) * BITS_PER_BYTE);
623         return data;
624 }
625 #endif
626
627 static void copy_io(u32 __iomem *piobuf, struct ipath_sge_state *ss,
628                     u32 length)
629 {
630         u32 extra = 0;
631         u32 data = 0;
632         u32 last;
633
634         while (1) {
635                 u32 len = ss->sge.length;
636                 u32 off;
637
638                 BUG_ON(len == 0);
639                 if (len > length)
640                         len = length;
641                 if (len > ss->sge.sge_length)
642                         len = ss->sge.sge_length;
643                 /* If the source address is not aligned, try to align it. */
644                 off = (unsigned long)ss->sge.vaddr & (sizeof(u32) - 1);
645                 if (off) {
646                         u32 *addr = (u32 *)((unsigned long)ss->sge.vaddr &
647                                             ~(sizeof(u32) - 1));
648                         u32 v = get_upper_bits(*addr, off * BITS_PER_BYTE);
649                         u32 y;
650
651                         y = sizeof(u32) - off;
652                         if (len > y)
653                                 len = y;
654                         if (len + extra >= sizeof(u32)) {
655                                 data |= set_upper_bits(v, extra *
656                                                        BITS_PER_BYTE);
657                                 len = sizeof(u32) - extra;
658                                 if (len == length) {
659                                         last = data;
660                                         break;
661                                 }
662                                 __raw_writel(data, piobuf);
663                                 piobuf++;
664                                 extra = 0;
665                                 data = 0;
666                         } else {
667                                 /* Clear unused upper bytes */
668                                 data |= clear_upper_bytes(v, len, extra);
669                                 if (len == length) {
670                                         last = data;
671                                         break;
672                                 }
673                                 extra += len;
674                         }
675                 } else if (extra) {
676                         /* Source address is aligned. */
677                         u32 *addr = (u32 *) ss->sge.vaddr;
678                         int shift = extra * BITS_PER_BYTE;
679                         int ushift = 32 - shift;
680                         u32 l = len;
681
682                         while (l >= sizeof(u32)) {
683                                 u32 v = *addr;
684
685                                 data |= set_upper_bits(v, shift);
686                                 __raw_writel(data, piobuf);
687                                 data = get_upper_bits(v, ushift);
688                                 piobuf++;
689                                 addr++;
690                                 l -= sizeof(u32);
691                         }
692                         /*
693                          * We still have 'extra' number of bytes leftover.
694                          */
695                         if (l) {
696                                 u32 v = *addr;
697
698                                 if (l + extra >= sizeof(u32)) {
699                                         data |= set_upper_bits(v, shift);
700                                         len -= l + extra - sizeof(u32);
701                                         if (len == length) {
702                                                 last = data;
703                                                 break;
704                                         }
705                                         __raw_writel(data, piobuf);
706                                         piobuf++;
707                                         extra = 0;
708                                         data = 0;
709                                 } else {
710                                         /* Clear unused upper bytes */
711                                         data |= clear_upper_bytes(v, l,
712                                                                   extra);
713                                         if (len == length) {
714                                                 last = data;
715                                                 break;
716                                         }
717                                         extra += l;
718                                 }
719                         } else if (len == length) {
720                                 last = data;
721                                 break;
722                         }
723                 } else if (len == length) {
724                         u32 w;
725
726                         /*
727                          * Need to round up for the last dword in the
728                          * packet.
729                          */
730                         w = (len + 3) >> 2;
731                         __iowrite32_copy(piobuf, ss->sge.vaddr, w - 1);
732                         piobuf += w - 1;
733                         last = ((u32 *) ss->sge.vaddr)[w - 1];
734                         break;
735                 } else {
736                         u32 w = len >> 2;
737
738                         __iowrite32_copy(piobuf, ss->sge.vaddr, w);
739                         piobuf += w;
740
741                         extra = len & (sizeof(u32) - 1);
742                         if (extra) {
743                                 u32 v = ((u32 *) ss->sge.vaddr)[w];
744
745                                 /* Clear unused upper bytes */
746                                 data = clear_upper_bytes(v, extra, 0);
747                         }
748                 }
749                 update_sge(ss, len);
750                 length -= len;
751         }
752         /* Update address before sending packet. */
753         update_sge(ss, length);
754         /* must flush early everything before trigger word */
755         ipath_flush_wc();
756         __raw_writel(last, piobuf);
757         /* be sure trigger word is written */
758         ipath_flush_wc();
759 }
760
761 /**
762  * ipath_verbs_send - send a packet
763  * @dd: the infinipath device
764  * @hdrwords: the number of words in the header
765  * @hdr: the packet header
766  * @len: the length of the packet in bytes
767  * @ss: the SGE to send
768  */
769 int ipath_verbs_send(struct ipath_devdata *dd, u32 hdrwords,
770                      u32 *hdr, u32 len, struct ipath_sge_state *ss)
771 {
772         u32 __iomem *piobuf;
773         u32 plen;
774         int ret;
775
776         /* +1 is for the qword padding of pbc */
777         plen = hdrwords + ((len + 3) >> 2) + 1;
778         if (unlikely((plen << 2) > dd->ipath_ibmaxlen)) {
779                 ret = -EINVAL;
780                 goto bail;
781         }
782
783         /* Get a PIO buffer to use. */
784         piobuf = ipath_getpiobuf(dd, NULL);
785         if (unlikely(piobuf == NULL)) {
786                 ret = -EBUSY;
787                 goto bail;
788         }
789
790         /*
791          * Write len to control qword, no flags.
792          * We have to flush after the PBC for correctness on some cpus
793          * or WC buffer can be written out of order.
794          */
795         writeq(plen, piobuf);
796         ipath_flush_wc();
797         piobuf += 2;
798         if (len == 0) {
799                 /*
800                  * If there is just the header portion, must flush before
801                  * writing last word of header for correctness, and after
802                  * the last header word (trigger word).
803                  */
804                 __iowrite32_copy(piobuf, hdr, hdrwords - 1);
805                 ipath_flush_wc();
806                 __raw_writel(hdr[hdrwords - 1], piobuf + hdrwords - 1);
807                 ipath_flush_wc();
808                 ret = 0;
809                 goto bail;
810         }
811
812         __iowrite32_copy(piobuf, hdr, hdrwords);
813         piobuf += hdrwords;
814
815         /* The common case is aligned and contained in one segment. */
816         if (likely(ss->num_sge == 1 && len <= ss->sge.length &&
817                    !((unsigned long)ss->sge.vaddr & (sizeof(u32) - 1)))) {
818                 u32 w;
819                 u32 *addr = (u32 *) ss->sge.vaddr;
820
821                 /* Update address before sending packet. */
822                 update_sge(ss, len);
823                 /* Need to round up for the last dword in the packet. */
824                 w = (len + 3) >> 2;
825                 __iowrite32_copy(piobuf, addr, w - 1);
826                 /* must flush early everything before trigger word */
827                 ipath_flush_wc();
828                 __raw_writel(addr[w - 1], piobuf + w - 1);
829                 /* be sure trigger word is written */
830                 ipath_flush_wc();
831                 ret = 0;
832                 goto bail;
833         }
834         copy_io(piobuf, ss, len);
835         ret = 0;
836
837 bail:
838         return ret;
839 }
840
841 int ipath_snapshot_counters(struct ipath_devdata *dd, u64 *swords,
842                             u64 *rwords, u64 *spkts, u64 *rpkts,
843                             u64 *xmit_wait)
844 {
845         int ret;
846
847         if (!(dd->ipath_flags & IPATH_INITTED)) {
848                 /* no hardware, freeze, etc. */
849                 ipath_dbg("unit %u not usable\n", dd->ipath_unit);
850                 ret = -EINVAL;
851                 goto bail;
852         }
853         *swords = ipath_snap_cntr(dd, dd->ipath_cregs->cr_wordsendcnt);
854         *rwords = ipath_snap_cntr(dd, dd->ipath_cregs->cr_wordrcvcnt);
855         *spkts = ipath_snap_cntr(dd, dd->ipath_cregs->cr_pktsendcnt);
856         *rpkts = ipath_snap_cntr(dd, dd->ipath_cregs->cr_pktrcvcnt);
857         *xmit_wait = ipath_snap_cntr(dd, dd->ipath_cregs->cr_sendstallcnt);
858
859         ret = 0;
860
861 bail:
862         return ret;
863 }
864
865 /**
866  * ipath_get_counters - get various chip counters
867  * @dd: the infinipath device
868  * @cntrs: counters are placed here
869  *
870  * Return the counters needed by recv_pma_get_portcounters().
871  */
872 int ipath_get_counters(struct ipath_devdata *dd,
873                        struct ipath_verbs_counters *cntrs)
874 {
875         int ret;
876
877         if (!(dd->ipath_flags & IPATH_INITTED)) {
878                 /* no hardware, freeze, etc. */
879                 ipath_dbg("unit %u not usable\n", dd->ipath_unit);
880                 ret = -EINVAL;
881                 goto bail;
882         }
883         cntrs->symbol_error_counter =
884                 ipath_snap_cntr(dd, dd->ipath_cregs->cr_ibsymbolerrcnt);
885         cntrs->link_error_recovery_counter =
886                 ipath_snap_cntr(dd, dd->ipath_cregs->cr_iblinkerrrecovcnt);
887         /*
888          * The link downed counter counts when the other side downs the
889          * connection.  We add in the number of times we downed the link
890          * due to local link integrity errors to compensate.
891          */
892         cntrs->link_downed_counter =
893                 ipath_snap_cntr(dd, dd->ipath_cregs->cr_iblinkdowncnt);
894         cntrs->port_rcv_errors =
895                 ipath_snap_cntr(dd, dd->ipath_cregs->cr_rxdroppktcnt) +
896                 ipath_snap_cntr(dd, dd->ipath_cregs->cr_rcvovflcnt) +
897                 ipath_snap_cntr(dd, dd->ipath_cregs->cr_portovflcnt) +
898                 ipath_snap_cntr(dd, dd->ipath_cregs->cr_err_rlencnt) +
899                 ipath_snap_cntr(dd, dd->ipath_cregs->cr_invalidrlencnt) +
900                 ipath_snap_cntr(dd, dd->ipath_cregs->cr_erricrccnt) +
901                 ipath_snap_cntr(dd, dd->ipath_cregs->cr_errvcrccnt) +
902                 ipath_snap_cntr(dd, dd->ipath_cregs->cr_errlpcrccnt) +
903                 ipath_snap_cntr(dd, dd->ipath_cregs->cr_badformatcnt) +
904                 dd->ipath_rxfc_unsupvl_errs;
905         cntrs->port_rcv_remphys_errors =
906                 ipath_snap_cntr(dd, dd->ipath_cregs->cr_rcvebpcnt);
907         cntrs->port_xmit_discards =
908                 ipath_snap_cntr(dd, dd->ipath_cregs->cr_unsupvlcnt);
909         cntrs->port_xmit_data =
910                 ipath_snap_cntr(dd, dd->ipath_cregs->cr_wordsendcnt);
911         cntrs->port_rcv_data =
912                 ipath_snap_cntr(dd, dd->ipath_cregs->cr_wordrcvcnt);
913         cntrs->port_xmit_packets =
914                 ipath_snap_cntr(dd, dd->ipath_cregs->cr_pktsendcnt);
915         cntrs->port_rcv_packets =
916                 ipath_snap_cntr(dd, dd->ipath_cregs->cr_pktrcvcnt);
917         cntrs->local_link_integrity_errors =
918                 (dd->ipath_flags & IPATH_GPIO_ERRINTRS) ?
919                 dd->ipath_lli_errs : dd->ipath_lli_errors;
920         cntrs->excessive_buffer_overrun_errors = dd->ipath_overrun_thresh_errs;
921
922         ret = 0;
923
924 bail:
925         return ret;
926 }
927
928 /**
929  * ipath_ib_piobufavail - callback when a PIO buffer is available
930  * @arg: the device pointer
931  *
932  * This is called from ipath_intr() at interrupt level when a PIO buffer is
933  * available after ipath_verbs_send() returned an error that no buffers were
934  * available.  Return 1 if we consumed all the PIO buffers and we still have
935  * QPs waiting for buffers (for now, just do a tasklet_hi_schedule and
936  * return zero).
937  */
938 int ipath_ib_piobufavail(struct ipath_ibdev *dev)
939 {
940         struct ipath_qp *qp;
941         unsigned long flags;
942
943         if (dev == NULL)
944                 goto bail;
945
946         spin_lock_irqsave(&dev->pending_lock, flags);
947         while (!list_empty(&dev->piowait)) {
948                 qp = list_entry(dev->piowait.next, struct ipath_qp,
949                                 piowait);
950                 list_del_init(&qp->piowait);
951                 tasklet_hi_schedule(&qp->s_task);
952         }
953         spin_unlock_irqrestore(&dev->pending_lock, flags);
954
955 bail:
956         return 0;
957 }
958
959 static int ipath_query_device(struct ib_device *ibdev,
960                               struct ib_device_attr *props)
961 {
962         struct ipath_ibdev *dev = to_idev(ibdev);
963
964         memset(props, 0, sizeof(*props));
965
966         props->device_cap_flags = IB_DEVICE_BAD_PKEY_CNTR |
967                 IB_DEVICE_BAD_QKEY_CNTR | IB_DEVICE_SHUTDOWN_PORT |
968                 IB_DEVICE_SYS_IMAGE_GUID;
969         props->page_size_cap = PAGE_SIZE;
970         props->vendor_id = dev->dd->ipath_vendorid;
971         props->vendor_part_id = dev->dd->ipath_deviceid;
972         props->hw_ver = dev->dd->ipath_pcirev;
973
974         props->sys_image_guid = dev->sys_image_guid;
975
976         props->max_mr_size = ~0ull;
977         props->max_qp = ib_ipath_max_qps;
978         props->max_qp_wr = ib_ipath_max_qp_wrs;
979         props->max_sge = ib_ipath_max_sges;
980         props->max_cq = ib_ipath_max_cqs;
981         props->max_ah = ib_ipath_max_ahs;
982         props->max_cqe = ib_ipath_max_cqes;
983         props->max_mr = dev->lk_table.max;
984         props->max_pd = ib_ipath_max_pds;
985         props->max_qp_rd_atom = IPATH_MAX_RDMA_ATOMIC;
986         props->max_qp_init_rd_atom = 255;
987         /* props->max_res_rd_atom */
988         props->max_srq = ib_ipath_max_srqs;
989         props->max_srq_wr = ib_ipath_max_srq_wrs;
990         props->max_srq_sge = ib_ipath_max_srq_sges;
991         /* props->local_ca_ack_delay */
992         props->atomic_cap = IB_ATOMIC_GLOB;
993         props->max_pkeys = ipath_get_npkeys(dev->dd);
994         props->max_mcast_grp = ib_ipath_max_mcast_grps;
995         props->max_mcast_qp_attach = ib_ipath_max_mcast_qp_attached;
996         props->max_total_mcast_qp_attach = props->max_mcast_qp_attach *
997                 props->max_mcast_grp;
998
999         return 0;
1000 }
1001
1002 const u8 ipath_cvt_physportstate[16] = {
1003         [INFINIPATH_IBCS_LT_STATE_DISABLED] = 3,
1004         [INFINIPATH_IBCS_LT_STATE_LINKUP] = 5,
1005         [INFINIPATH_IBCS_LT_STATE_POLLACTIVE] = 2,
1006         [INFINIPATH_IBCS_LT_STATE_POLLQUIET] = 2,
1007         [INFINIPATH_IBCS_LT_STATE_SLEEPDELAY] = 1,
1008         [INFINIPATH_IBCS_LT_STATE_SLEEPQUIET] = 1,
1009         [INFINIPATH_IBCS_LT_STATE_CFGDEBOUNCE] = 4,
1010         [INFINIPATH_IBCS_LT_STATE_CFGRCVFCFG] = 4,
1011         [INFINIPATH_IBCS_LT_STATE_CFGWAITRMT] = 4,
1012         [INFINIPATH_IBCS_LT_STATE_CFGIDLE] = 4,
1013         [INFINIPATH_IBCS_LT_STATE_RECOVERRETRAIN] = 6,
1014         [INFINIPATH_IBCS_LT_STATE_RECOVERWAITRMT] = 6,
1015         [INFINIPATH_IBCS_LT_STATE_RECOVERIDLE] = 6,
1016 };
1017
1018 u32 ipath_get_cr_errpkey(struct ipath_devdata *dd)
1019 {
1020         return ipath_read_creg32(dd, dd->ipath_cregs->cr_errpkey);
1021 }
1022
1023 static int ipath_query_port(struct ib_device *ibdev,
1024                             u8 port, struct ib_port_attr *props)
1025 {
1026         struct ipath_ibdev *dev = to_idev(ibdev);
1027         enum ib_mtu mtu;
1028         u16 lid = dev->dd->ipath_lid;
1029         u64 ibcstat;
1030
1031         memset(props, 0, sizeof(*props));
1032         props->lid = lid ? lid : __constant_be16_to_cpu(IB_LID_PERMISSIVE);
1033         props->lmc = dev->mkeyprot_resv_lmc & 7;
1034         props->sm_lid = dev->sm_lid;
1035         props->sm_sl = dev->sm_sl;
1036         ibcstat = dev->dd->ipath_lastibcstat;
1037         props->state = ((ibcstat >> 4) & 0x3) + 1;
1038         /* See phys_state_show() */
1039         props->phys_state = ipath_cvt_physportstate[
1040                 dev->dd->ipath_lastibcstat & 0xf];
1041         props->port_cap_flags = dev->port_cap_flags;
1042         props->gid_tbl_len = 1;
1043         props->max_msg_sz = 0x80000000;
1044         props->pkey_tbl_len = ipath_get_npkeys(dev->dd);
1045         props->bad_pkey_cntr = ipath_get_cr_errpkey(dev->dd) -
1046                 dev->z_pkey_violations;
1047         props->qkey_viol_cntr = dev->qkey_violations;
1048         props->active_width = IB_WIDTH_4X;
1049         /* See rate_show() */
1050         props->active_speed = 1;        /* Regular 10Mbs speed. */
1051         props->max_vl_num = 1;          /* VLCap = VL0 */
1052         props->init_type_reply = 0;
1053
1054         props->max_mtu = IB_MTU_4096;
1055         switch (dev->dd->ipath_ibmtu) {
1056         case 4096:
1057                 mtu = IB_MTU_4096;
1058                 break;
1059         case 2048:
1060                 mtu = IB_MTU_2048;
1061                 break;
1062         case 1024:
1063                 mtu = IB_MTU_1024;
1064                 break;
1065         case 512:
1066                 mtu = IB_MTU_512;
1067                 break;
1068         case 256:
1069                 mtu = IB_MTU_256;
1070                 break;
1071         default:
1072                 mtu = IB_MTU_2048;
1073         }
1074         props->active_mtu = mtu;
1075         props->subnet_timeout = dev->subnet_timeout;
1076
1077         return 0;
1078 }
1079
1080 static int ipath_modify_device(struct ib_device *device,
1081                                int device_modify_mask,
1082                                struct ib_device_modify *device_modify)
1083 {
1084         int ret;
1085
1086         if (device_modify_mask & ~(IB_DEVICE_MODIFY_SYS_IMAGE_GUID |
1087                                    IB_DEVICE_MODIFY_NODE_DESC)) {
1088                 ret = -EOPNOTSUPP;
1089                 goto bail;
1090         }
1091
1092         if (device_modify_mask & IB_DEVICE_MODIFY_NODE_DESC)
1093                 memcpy(device->node_desc, device_modify->node_desc, 64);
1094
1095         if (device_modify_mask & IB_DEVICE_MODIFY_SYS_IMAGE_GUID)
1096                 to_idev(device)->sys_image_guid =
1097                         cpu_to_be64(device_modify->sys_image_guid);
1098
1099         ret = 0;
1100
1101 bail:
1102         return ret;
1103 }
1104
1105 static int ipath_modify_port(struct ib_device *ibdev,
1106                              u8 port, int port_modify_mask,
1107                              struct ib_port_modify *props)
1108 {
1109         struct ipath_ibdev *dev = to_idev(ibdev);
1110
1111         dev->port_cap_flags |= props->set_port_cap_mask;
1112         dev->port_cap_flags &= ~props->clr_port_cap_mask;
1113         if (port_modify_mask & IB_PORT_SHUTDOWN)
1114                 ipath_set_linkstate(dev->dd, IPATH_IB_LINKDOWN);
1115         if (port_modify_mask & IB_PORT_RESET_QKEY_CNTR)
1116                 dev->qkey_violations = 0;
1117         return 0;
1118 }
1119
1120 static int ipath_query_gid(struct ib_device *ibdev, u8 port,
1121                            int index, union ib_gid *gid)
1122 {
1123         struct ipath_ibdev *dev = to_idev(ibdev);
1124         int ret;
1125
1126         if (index >= 1) {
1127                 ret = -EINVAL;
1128                 goto bail;
1129         }
1130         gid->global.subnet_prefix = dev->gid_prefix;
1131         gid->global.interface_id = dev->dd->ipath_guid;
1132
1133         ret = 0;
1134
1135 bail:
1136         return ret;
1137 }
1138
1139 static struct ib_pd *ipath_alloc_pd(struct ib_device *ibdev,
1140                                     struct ib_ucontext *context,
1141                                     struct ib_udata *udata)
1142 {
1143         struct ipath_ibdev *dev = to_idev(ibdev);
1144         struct ipath_pd *pd;
1145         struct ib_pd *ret;
1146
1147         /*
1148          * This is actually totally arbitrary.  Some correctness tests
1149          * assume there's a maximum number of PDs that can be allocated.
1150          * We don't actually have this limit, but we fail the test if
1151          * we allow allocations of more than we report for this value.
1152          */
1153
1154         pd = kmalloc(sizeof *pd, GFP_KERNEL);
1155         if (!pd) {
1156                 ret = ERR_PTR(-ENOMEM);
1157                 goto bail;
1158         }
1159
1160         spin_lock(&dev->n_pds_lock);
1161         if (dev->n_pds_allocated == ib_ipath_max_pds) {
1162                 spin_unlock(&dev->n_pds_lock);
1163                 kfree(pd);
1164                 ret = ERR_PTR(-ENOMEM);
1165                 goto bail;
1166         }
1167
1168         dev->n_pds_allocated++;
1169         spin_unlock(&dev->n_pds_lock);
1170
1171         /* ib_alloc_pd() will initialize pd->ibpd. */
1172         pd->user = udata != NULL;
1173
1174         ret = &pd->ibpd;
1175
1176 bail:
1177         return ret;
1178 }
1179
1180 static int ipath_dealloc_pd(struct ib_pd *ibpd)
1181 {
1182         struct ipath_pd *pd = to_ipd(ibpd);
1183         struct ipath_ibdev *dev = to_idev(ibpd->device);
1184
1185         spin_lock(&dev->n_pds_lock);
1186         dev->n_pds_allocated--;
1187         spin_unlock(&dev->n_pds_lock);
1188
1189         kfree(pd);
1190
1191         return 0;
1192 }
1193
1194 /**
1195  * ipath_create_ah - create an address handle
1196  * @pd: the protection domain
1197  * @ah_attr: the attributes of the AH
1198  *
1199  * This may be called from interrupt context.
1200  */
1201 static struct ib_ah *ipath_create_ah(struct ib_pd *pd,
1202                                      struct ib_ah_attr *ah_attr)
1203 {
1204         struct ipath_ah *ah;
1205         struct ib_ah *ret;
1206         struct ipath_ibdev *dev = to_idev(pd->device);
1207         unsigned long flags;
1208
1209         /* A multicast address requires a GRH (see ch. 8.4.1). */
1210         if (ah_attr->dlid >= IPATH_MULTICAST_LID_BASE &&
1211             ah_attr->dlid != IPATH_PERMISSIVE_LID &&
1212             !(ah_attr->ah_flags & IB_AH_GRH)) {
1213                 ret = ERR_PTR(-EINVAL);
1214                 goto bail;
1215         }
1216
1217         if (ah_attr->dlid == 0) {
1218                 ret = ERR_PTR(-EINVAL);
1219                 goto bail;
1220         }
1221
1222         if (ah_attr->port_num < 1 ||
1223             ah_attr->port_num > pd->device->phys_port_cnt) {
1224                 ret = ERR_PTR(-EINVAL);
1225                 goto bail;
1226         }
1227
1228         ah = kmalloc(sizeof *ah, GFP_ATOMIC);
1229         if (!ah) {
1230                 ret = ERR_PTR(-ENOMEM);
1231                 goto bail;
1232         }
1233
1234         spin_lock_irqsave(&dev->n_ahs_lock, flags);
1235         if (dev->n_ahs_allocated == ib_ipath_max_ahs) {
1236                 spin_unlock_irqrestore(&dev->n_ahs_lock, flags);
1237                 kfree(ah);
1238                 ret = ERR_PTR(-ENOMEM);
1239                 goto bail;
1240         }
1241
1242         dev->n_ahs_allocated++;
1243         spin_unlock_irqrestore(&dev->n_ahs_lock, flags);
1244
1245         /* ib_create_ah() will initialize ah->ibah. */
1246         ah->attr = *ah_attr;
1247
1248         ret = &ah->ibah;
1249
1250 bail:
1251         return ret;
1252 }
1253
1254 /**
1255  * ipath_destroy_ah - destroy an address handle
1256  * @ibah: the AH to destroy
1257  *
1258  * This may be called from interrupt context.
1259  */
1260 static int ipath_destroy_ah(struct ib_ah *ibah)
1261 {
1262         struct ipath_ibdev *dev = to_idev(ibah->device);
1263         struct ipath_ah *ah = to_iah(ibah);
1264         unsigned long flags;
1265
1266         spin_lock_irqsave(&dev->n_ahs_lock, flags);
1267         dev->n_ahs_allocated--;
1268         spin_unlock_irqrestore(&dev->n_ahs_lock, flags);
1269
1270         kfree(ah);
1271
1272         return 0;
1273 }
1274
1275 static int ipath_query_ah(struct ib_ah *ibah, struct ib_ah_attr *ah_attr)
1276 {
1277         struct ipath_ah *ah = to_iah(ibah);
1278
1279         *ah_attr = ah->attr;
1280
1281         return 0;
1282 }
1283
1284 /**
1285  * ipath_get_npkeys - return the size of the PKEY table for port 0
1286  * @dd: the infinipath device
1287  */
1288 unsigned ipath_get_npkeys(struct ipath_devdata *dd)
1289 {
1290         return ARRAY_SIZE(dd->ipath_pd[0]->port_pkeys);
1291 }
1292
1293 /**
1294  * ipath_get_pkey - return the indexed PKEY from the port 0 PKEY table
1295  * @dd: the infinipath device
1296  * @index: the PKEY index
1297  */
1298 unsigned ipath_get_pkey(struct ipath_devdata *dd, unsigned index)
1299 {
1300         unsigned ret;
1301
1302         if (index >= ARRAY_SIZE(dd->ipath_pd[0]->port_pkeys))
1303                 ret = 0;
1304         else
1305                 ret = dd->ipath_pd[0]->port_pkeys[index];
1306
1307         return ret;
1308 }
1309
1310 static int ipath_query_pkey(struct ib_device *ibdev, u8 port, u16 index,
1311                             u16 *pkey)
1312 {
1313         struct ipath_ibdev *dev = to_idev(ibdev);
1314         int ret;
1315
1316         if (index >= ipath_get_npkeys(dev->dd)) {
1317                 ret = -EINVAL;
1318                 goto bail;
1319         }
1320
1321         *pkey = ipath_get_pkey(dev->dd, index);
1322         ret = 0;
1323
1324 bail:
1325         return ret;
1326 }
1327
1328 /**
1329  * ipath_alloc_ucontext - allocate a ucontest
1330  * @ibdev: the infiniband device
1331  * @udata: not used by the InfiniPath driver
1332  */
1333
1334 static struct ib_ucontext *ipath_alloc_ucontext(struct ib_device *ibdev,
1335                                                 struct ib_udata *udata)
1336 {
1337         struct ipath_ucontext *context;
1338         struct ib_ucontext *ret;
1339
1340         context = kmalloc(sizeof *context, GFP_KERNEL);
1341         if (!context) {
1342                 ret = ERR_PTR(-ENOMEM);
1343                 goto bail;
1344         }
1345
1346         ret = &context->ibucontext;
1347
1348 bail:
1349         return ret;
1350 }
1351
1352 static int ipath_dealloc_ucontext(struct ib_ucontext *context)
1353 {
1354         kfree(to_iucontext(context));
1355         return 0;
1356 }
1357
1358 static int ipath_verbs_register_sysfs(struct ib_device *dev);
1359
1360 static void __verbs_timer(unsigned long arg)
1361 {
1362         struct ipath_devdata *dd = (struct ipath_devdata *) arg;
1363
1364         /*
1365          * If port 0 receive packet interrupts are not available, or
1366          * can be missed, poll the receive queue
1367          */
1368         if (dd->ipath_flags & IPATH_POLL_RX_INTR)
1369                 ipath_kreceive(dd);
1370
1371         /* Handle verbs layer timeouts. */
1372         ipath_ib_timer(dd->verbs_dev);
1373
1374         mod_timer(&dd->verbs_timer, jiffies + 1);
1375 }
1376
1377 static int enable_timer(struct ipath_devdata *dd)
1378 {
1379         /*
1380          * Early chips had a design flaw where the chip and kernel idea
1381          * of the tail register don't always agree, and therefore we won't
1382          * get an interrupt on the next packet received.
1383          * If the board supports per packet receive interrupts, use it.
1384          * Otherwise, the timer function periodically checks for packets
1385          * to cover this case.
1386          * Either way, the timer is needed for verbs layer related
1387          * processing.
1388          */
1389         if (dd->ipath_flags & IPATH_GPIO_INTR) {
1390                 u64 val;
1391                 ipath_write_kreg(dd, dd->ipath_kregs->kr_debugportselect,
1392                                  0x2074076542310ULL);
1393                 /* Enable GPIO bit 2 interrupt */
1394                 val = ipath_read_kreg64(dd, dd->ipath_kregs->kr_gpio_mask);
1395                 val |= (u64) (1 << IPATH_GPIO_PORT0_BIT);
1396                 ipath_write_kreg( dd, dd->ipath_kregs->kr_gpio_mask, val);
1397         }
1398
1399         init_timer(&dd->verbs_timer);
1400         dd->verbs_timer.function = __verbs_timer;
1401         dd->verbs_timer.data = (unsigned long)dd;
1402         dd->verbs_timer.expires = jiffies + 1;
1403         add_timer(&dd->verbs_timer);
1404
1405         return 0;
1406 }
1407
1408 static int disable_timer(struct ipath_devdata *dd)
1409 {
1410         /* Disable GPIO bit 2 interrupt */
1411         if (dd->ipath_flags & IPATH_GPIO_INTR) {
1412                 u64 val;
1413                 /* Disable GPIO bit 2 interrupt */
1414                 val = ipath_read_kreg64(dd, dd->ipath_kregs->kr_gpio_mask);
1415                 val &= ~((u64) (1 << IPATH_GPIO_PORT0_BIT));
1416                 ipath_write_kreg( dd, dd->ipath_kregs->kr_gpio_mask, val);
1417                 /*
1418                  * We might want to undo changes to debugportselect,
1419                  * but how?
1420                  */
1421         }
1422
1423         del_timer_sync(&dd->verbs_timer);
1424
1425         return 0;
1426 }
1427
1428 /**
1429  * ipath_register_ib_device - register our device with the infiniband core
1430  * @dd: the device data structure
1431  * Return the allocated ipath_ibdev pointer or NULL on error.
1432  */
1433 int ipath_register_ib_device(struct ipath_devdata *dd)
1434 {
1435         struct ipath_verbs_counters cntrs;
1436         struct ipath_ibdev *idev;
1437         struct ib_device *dev;
1438         int ret;
1439
1440         idev = (struct ipath_ibdev *)ib_alloc_device(sizeof *idev);
1441         if (idev == NULL) {
1442                 ret = -ENOMEM;
1443                 goto bail;
1444         }
1445
1446         dev = &idev->ibdev;
1447
1448         /* Only need to initialize non-zero fields. */
1449         spin_lock_init(&idev->n_pds_lock);
1450         spin_lock_init(&idev->n_ahs_lock);
1451         spin_lock_init(&idev->n_cqs_lock);
1452         spin_lock_init(&idev->n_qps_lock);
1453         spin_lock_init(&idev->n_srqs_lock);
1454         spin_lock_init(&idev->n_mcast_grps_lock);
1455
1456         spin_lock_init(&idev->qp_table.lock);
1457         spin_lock_init(&idev->lk_table.lock);
1458         idev->sm_lid = __constant_be16_to_cpu(IB_LID_PERMISSIVE);
1459         /* Set the prefix to the default value (see ch. 4.1.1) */
1460         idev->gid_prefix = __constant_cpu_to_be64(0xfe80000000000000ULL);
1461
1462         ret = ipath_init_qp_table(idev, ib_ipath_qp_table_size);
1463         if (ret)
1464                 goto err_qp;
1465
1466         /*
1467          * The top ib_ipath_lkey_table_size bits are used to index the
1468          * table.  The lower 8 bits can be owned by the user (copied from
1469          * the LKEY).  The remaining bits act as a generation number or tag.
1470          */
1471         idev->lk_table.max = 1 << ib_ipath_lkey_table_size;
1472         idev->lk_table.table = kzalloc(idev->lk_table.max *
1473                                        sizeof(*idev->lk_table.table),
1474                                        GFP_KERNEL);
1475         if (idev->lk_table.table == NULL) {
1476                 ret = -ENOMEM;
1477                 goto err_lk;
1478         }
1479         spin_lock_init(&idev->pending_lock);
1480         INIT_LIST_HEAD(&idev->pending[0]);
1481         INIT_LIST_HEAD(&idev->pending[1]);
1482         INIT_LIST_HEAD(&idev->pending[2]);
1483         INIT_LIST_HEAD(&idev->piowait);
1484         INIT_LIST_HEAD(&idev->rnrwait);
1485         idev->pending_index = 0;
1486         idev->port_cap_flags =
1487                 IB_PORT_SYS_IMAGE_GUID_SUP | IB_PORT_CLIENT_REG_SUP;
1488         idev->pma_counter_select[0] = IB_PMA_PORT_XMIT_DATA;
1489         idev->pma_counter_select[1] = IB_PMA_PORT_RCV_DATA;
1490         idev->pma_counter_select[2] = IB_PMA_PORT_XMIT_PKTS;
1491         idev->pma_counter_select[3] = IB_PMA_PORT_RCV_PKTS;
1492         idev->pma_counter_select[4] = IB_PMA_PORT_XMIT_WAIT;
1493         idev->link_width_enabled = 3;   /* 1x or 4x */
1494
1495         /* Snapshot current HW counters to "clear" them. */
1496         ipath_get_counters(dd, &cntrs);
1497         idev->z_symbol_error_counter = cntrs.symbol_error_counter;
1498         idev->z_link_error_recovery_counter =
1499                 cntrs.link_error_recovery_counter;
1500         idev->z_link_downed_counter = cntrs.link_downed_counter;
1501         idev->z_port_rcv_errors = cntrs.port_rcv_errors;
1502         idev->z_port_rcv_remphys_errors =
1503                 cntrs.port_rcv_remphys_errors;
1504         idev->z_port_xmit_discards = cntrs.port_xmit_discards;
1505         idev->z_port_xmit_data = cntrs.port_xmit_data;
1506         idev->z_port_rcv_data = cntrs.port_rcv_data;
1507         idev->z_port_xmit_packets = cntrs.port_xmit_packets;
1508         idev->z_port_rcv_packets = cntrs.port_rcv_packets;
1509         idev->z_local_link_integrity_errors =
1510                 cntrs.local_link_integrity_errors;
1511         idev->z_excessive_buffer_overrun_errors =
1512                 cntrs.excessive_buffer_overrun_errors;
1513
1514         /*
1515          * The system image GUID is supposed to be the same for all
1516          * IB HCAs in a single system but since there can be other
1517          * device types in the system, we can't be sure this is unique.
1518          */
1519         if (!sys_image_guid)
1520                 sys_image_guid = dd->ipath_guid;
1521         idev->sys_image_guid = sys_image_guid;
1522         idev->ib_unit = dd->ipath_unit;
1523         idev->dd = dd;
1524
1525         strlcpy(dev->name, "ipath%d", IB_DEVICE_NAME_MAX);
1526         dev->owner = THIS_MODULE;
1527         dev->node_guid = dd->ipath_guid;
1528         dev->uverbs_abi_ver = IPATH_UVERBS_ABI_VERSION;
1529         dev->uverbs_cmd_mask =
1530                 (1ull << IB_USER_VERBS_CMD_GET_CONTEXT)         |
1531                 (1ull << IB_USER_VERBS_CMD_QUERY_DEVICE)        |
1532                 (1ull << IB_USER_VERBS_CMD_QUERY_PORT)          |
1533                 (1ull << IB_USER_VERBS_CMD_ALLOC_PD)            |
1534                 (1ull << IB_USER_VERBS_CMD_DEALLOC_PD)          |
1535                 (1ull << IB_USER_VERBS_CMD_CREATE_AH)           |
1536                 (1ull << IB_USER_VERBS_CMD_DESTROY_AH)          |
1537                 (1ull << IB_USER_VERBS_CMD_QUERY_AH)            |
1538                 (1ull << IB_USER_VERBS_CMD_REG_MR)              |
1539                 (1ull << IB_USER_VERBS_CMD_DEREG_MR)            |
1540                 (1ull << IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL) |
1541                 (1ull << IB_USER_VERBS_CMD_CREATE_CQ)           |
1542                 (1ull << IB_USER_VERBS_CMD_RESIZE_CQ)           |
1543                 (1ull << IB_USER_VERBS_CMD_DESTROY_CQ)          |
1544                 (1ull << IB_USER_VERBS_CMD_POLL_CQ)             |
1545                 (1ull << IB_USER_VERBS_CMD_REQ_NOTIFY_CQ)       |
1546                 (1ull << IB_USER_VERBS_CMD_CREATE_QP)           |
1547                 (1ull << IB_USER_VERBS_CMD_QUERY_QP)            |
1548                 (1ull << IB_USER_VERBS_CMD_MODIFY_QP)           |
1549                 (1ull << IB_USER_VERBS_CMD_DESTROY_QP)          |
1550                 (1ull << IB_USER_VERBS_CMD_POST_SEND)           |
1551                 (1ull << IB_USER_VERBS_CMD_POST_RECV)           |
1552                 (1ull << IB_USER_VERBS_CMD_ATTACH_MCAST)        |
1553                 (1ull << IB_USER_VERBS_CMD_DETACH_MCAST)        |
1554                 (1ull << IB_USER_VERBS_CMD_CREATE_SRQ)          |
1555                 (1ull << IB_USER_VERBS_CMD_MODIFY_SRQ)          |
1556                 (1ull << IB_USER_VERBS_CMD_QUERY_SRQ)           |
1557                 (1ull << IB_USER_VERBS_CMD_DESTROY_SRQ)         |
1558                 (1ull << IB_USER_VERBS_CMD_POST_SRQ_RECV);
1559         dev->node_type = RDMA_NODE_IB_CA;
1560         dev->phys_port_cnt = 1;
1561         dev->dma_device = &dd->pcidev->dev;
1562         dev->query_device = ipath_query_device;
1563         dev->modify_device = ipath_modify_device;
1564         dev->query_port = ipath_query_port;
1565         dev->modify_port = ipath_modify_port;
1566         dev->query_pkey = ipath_query_pkey;
1567         dev->query_gid = ipath_query_gid;
1568         dev->alloc_ucontext = ipath_alloc_ucontext;
1569         dev->dealloc_ucontext = ipath_dealloc_ucontext;
1570         dev->alloc_pd = ipath_alloc_pd;
1571         dev->dealloc_pd = ipath_dealloc_pd;
1572         dev->create_ah = ipath_create_ah;
1573         dev->destroy_ah = ipath_destroy_ah;
1574         dev->query_ah = ipath_query_ah;
1575         dev->create_srq = ipath_create_srq;
1576         dev->modify_srq = ipath_modify_srq;
1577         dev->query_srq = ipath_query_srq;
1578         dev->destroy_srq = ipath_destroy_srq;
1579         dev->create_qp = ipath_create_qp;
1580         dev->modify_qp = ipath_modify_qp;
1581         dev->query_qp = ipath_query_qp;
1582         dev->destroy_qp = ipath_destroy_qp;
1583         dev->post_send = ipath_post_send;
1584         dev->post_recv = ipath_post_receive;
1585         dev->post_srq_recv = ipath_post_srq_receive;
1586         dev->create_cq = ipath_create_cq;
1587         dev->destroy_cq = ipath_destroy_cq;
1588         dev->resize_cq = ipath_resize_cq;
1589         dev->poll_cq = ipath_poll_cq;
1590         dev->req_notify_cq = ipath_req_notify_cq;
1591         dev->get_dma_mr = ipath_get_dma_mr;
1592         dev->reg_phys_mr = ipath_reg_phys_mr;
1593         dev->reg_user_mr = ipath_reg_user_mr;
1594         dev->dereg_mr = ipath_dereg_mr;
1595         dev->alloc_fmr = ipath_alloc_fmr;
1596         dev->map_phys_fmr = ipath_map_phys_fmr;
1597         dev->unmap_fmr = ipath_unmap_fmr;
1598         dev->dealloc_fmr = ipath_dealloc_fmr;
1599         dev->attach_mcast = ipath_multicast_attach;
1600         dev->detach_mcast = ipath_multicast_detach;
1601         dev->process_mad = ipath_process_mad;
1602         dev->mmap = ipath_mmap;
1603         dev->dma_ops = &ipath_dma_mapping_ops;
1604
1605         snprintf(dev->node_desc, sizeof(dev->node_desc),
1606                  IPATH_IDSTR " %s", init_utsname()->nodename);
1607
1608         ret = ib_register_device(dev);
1609         if (ret)
1610                 goto err_reg;
1611
1612         if (ipath_verbs_register_sysfs(dev))
1613                 goto err_class;
1614
1615         enable_timer(dd);
1616
1617         goto bail;
1618
1619 err_class:
1620         ib_unregister_device(dev);
1621 err_reg:
1622         kfree(idev->lk_table.table);
1623 err_lk:
1624         kfree(idev->qp_table.table);
1625 err_qp:
1626         ib_dealloc_device(dev);
1627         ipath_dev_err(dd, "cannot register verbs: %d!\n", -ret);
1628         idev = NULL;
1629
1630 bail:
1631         dd->verbs_dev = idev;
1632         return ret;
1633 }
1634
1635 void ipath_unregister_ib_device(struct ipath_ibdev *dev)
1636 {
1637         struct ib_device *ibdev = &dev->ibdev;
1638
1639         disable_timer(dev->dd);
1640
1641         ib_unregister_device(ibdev);
1642
1643         if (!list_empty(&dev->pending[0]) ||
1644             !list_empty(&dev->pending[1]) ||
1645             !list_empty(&dev->pending[2]))
1646                 ipath_dev_err(dev->dd, "pending list not empty!\n");
1647         if (!list_empty(&dev->piowait))
1648                 ipath_dev_err(dev->dd, "piowait list not empty!\n");
1649         if (!list_empty(&dev->rnrwait))
1650                 ipath_dev_err(dev->dd, "rnrwait list not empty!\n");
1651         if (!ipath_mcast_tree_empty())
1652                 ipath_dev_err(dev->dd, "multicast table memory leak!\n");
1653         /*
1654          * Note that ipath_unregister_ib_device() can be called before all
1655          * the QPs are destroyed!
1656          */
1657         ipath_free_all_qps(&dev->qp_table);
1658         kfree(dev->qp_table.table);
1659         kfree(dev->lk_table.table);
1660         ib_dealloc_device(ibdev);
1661 }
1662
1663 static ssize_t show_rev(struct class_device *cdev, char *buf)
1664 {
1665         struct ipath_ibdev *dev =
1666                 container_of(cdev, struct ipath_ibdev, ibdev.class_dev);
1667
1668         return sprintf(buf, "%x\n", dev->dd->ipath_pcirev);
1669 }
1670
1671 static ssize_t show_hca(struct class_device *cdev, char *buf)
1672 {
1673         struct ipath_ibdev *dev =
1674                 container_of(cdev, struct ipath_ibdev, ibdev.class_dev);
1675         int ret;
1676
1677         ret = dev->dd->ipath_f_get_boardname(dev->dd, buf, 128);
1678         if (ret < 0)
1679                 goto bail;
1680         strcat(buf, "\n");
1681         ret = strlen(buf);
1682
1683 bail:
1684         return ret;
1685 }
1686
1687 static ssize_t show_stats(struct class_device *cdev, char *buf)
1688 {
1689         struct ipath_ibdev *dev =
1690                 container_of(cdev, struct ipath_ibdev, ibdev.class_dev);
1691         int i;
1692         int len;
1693
1694         len = sprintf(buf,
1695                       "RC resends  %d\n"
1696                       "RC no QACK  %d\n"
1697                       "RC ACKs     %d\n"
1698                       "RC SEQ NAKs %d\n"
1699                       "RC RDMA seq %d\n"
1700                       "RC RNR NAKs %d\n"
1701                       "RC OTH NAKs %d\n"
1702                       "RC timeouts %d\n"
1703                       "RC RDMA dup %d\n"
1704                       "RC stalls   %d\n"
1705                       "piobuf wait %d\n"
1706                       "no piobuf   %d\n"
1707                       "PKT drops   %d\n"
1708                       "WQE errs    %d\n",
1709                       dev->n_rc_resends, dev->n_rc_qacks, dev->n_rc_acks,
1710                       dev->n_seq_naks, dev->n_rdma_seq, dev->n_rnr_naks,
1711                       dev->n_other_naks, dev->n_timeouts,
1712                       dev->n_rdma_dup_busy, dev->n_rc_stalls, dev->n_piowait,
1713                       dev->n_no_piobuf, dev->n_pkt_drops, dev->n_wqe_errs);
1714         for (i = 0; i < ARRAY_SIZE(dev->opstats); i++) {
1715                 const struct ipath_opcode_stats *si = &dev->opstats[i];
1716
1717                 if (!si->n_packets && !si->n_bytes)
1718                         continue;
1719                 len += sprintf(buf + len, "%02x %llu/%llu\n", i,
1720                                (unsigned long long) si->n_packets,
1721                                (unsigned long long) si->n_bytes);
1722         }
1723         return len;
1724 }
1725
1726 static CLASS_DEVICE_ATTR(hw_rev, S_IRUGO, show_rev, NULL);
1727 static CLASS_DEVICE_ATTR(hca_type, S_IRUGO, show_hca, NULL);
1728 static CLASS_DEVICE_ATTR(board_id, S_IRUGO, show_hca, NULL);
1729 static CLASS_DEVICE_ATTR(stats, S_IRUGO, show_stats, NULL);
1730
1731 static struct class_device_attribute *ipath_class_attributes[] = {
1732         &class_device_attr_hw_rev,
1733         &class_device_attr_hca_type,
1734         &class_device_attr_board_id,
1735         &class_device_attr_stats
1736 };
1737
1738 static int ipath_verbs_register_sysfs(struct ib_device *dev)
1739 {
1740         int i;
1741         int ret;
1742
1743         for (i = 0; i < ARRAY_SIZE(ipath_class_attributes); ++i)
1744                 if (class_device_create_file(&dev->class_dev,
1745                                              ipath_class_attributes[i])) {
1746                         ret = 1;
1747                         goto bail;
1748                 }
1749
1750         ret = 0;
1751
1752 bail:
1753         return ret;
1754 }