Merge branch 'master' into gfs2
[linux-2.6] / drivers / infiniband / hw / ipath / ipath_qp.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 <linux/err.h>
35 #include <linux/vmalloc.h>
36
37 #include "ipath_verbs.h"
38 #include "ipath_kernel.h"
39
40 #define BITS_PER_PAGE           (PAGE_SIZE*BITS_PER_BYTE)
41 #define BITS_PER_PAGE_MASK      (BITS_PER_PAGE-1)
42 #define mk_qpn(qpt, map, off)   (((map) - (qpt)->map) * BITS_PER_PAGE + \
43                                  (off))
44 #define find_next_offset(map, off) find_next_zero_bit((map)->page, \
45                                                       BITS_PER_PAGE, off)
46
47 /*
48  * Convert the AETH credit code into the number of credits.
49  */
50 static u32 credit_table[31] = {
51         0,                      /* 0 */
52         1,                      /* 1 */
53         2,                      /* 2 */
54         3,                      /* 3 */
55         4,                      /* 4 */
56         6,                      /* 5 */
57         8,                      /* 6 */
58         12,                     /* 7 */
59         16,                     /* 8 */
60         24,                     /* 9 */
61         32,                     /* A */
62         48,                     /* B */
63         64,                     /* C */
64         96,                     /* D */
65         128,                    /* E */
66         192,                    /* F */
67         256,                    /* 10 */
68         384,                    /* 11 */
69         512,                    /* 12 */
70         768,                    /* 13 */
71         1024,                   /* 14 */
72         1536,                   /* 15 */
73         2048,                   /* 16 */
74         3072,                   /* 17 */
75         4096,                   /* 18 */
76         6144,                   /* 19 */
77         8192,                   /* 1A */
78         12288,                  /* 1B */
79         16384,                  /* 1C */
80         24576,                  /* 1D */
81         32768                   /* 1E */
82 };
83
84 static u32 alloc_qpn(struct ipath_qp_table *qpt)
85 {
86         u32 i, offset, max_scan, qpn;
87         struct qpn_map *map;
88         u32 ret;
89
90         qpn = qpt->last + 1;
91         if (qpn >= QPN_MAX)
92                 qpn = 2;
93         offset = qpn & BITS_PER_PAGE_MASK;
94         map = &qpt->map[qpn / BITS_PER_PAGE];
95         max_scan = qpt->nmaps - !offset;
96         for (i = 0;;) {
97                 if (unlikely(!map->page)) {
98                         unsigned long page = get_zeroed_page(GFP_KERNEL);
99                         unsigned long flags;
100
101                         /*
102                          * Free the page if someone raced with us
103                          * installing it:
104                          */
105                         spin_lock_irqsave(&qpt->lock, flags);
106                         if (map->page)
107                                 free_page(page);
108                         else
109                                 map->page = (void *)page;
110                         spin_unlock_irqrestore(&qpt->lock, flags);
111                         if (unlikely(!map->page))
112                                 break;
113                 }
114                 if (likely(atomic_read(&map->n_free))) {
115                         do {
116                                 if (!test_and_set_bit(offset, map->page)) {
117                                         atomic_dec(&map->n_free);
118                                         qpt->last = qpn;
119                                         ret = qpn;
120                                         goto bail;
121                                 }
122                                 offset = find_next_offset(map, offset);
123                                 qpn = mk_qpn(qpt, map, offset);
124                                 /*
125                                  * This test differs from alloc_pidmap().
126                                  * If find_next_offset() does find a zero
127                                  * bit, we don't need to check for QPN
128                                  * wrapping around past our starting QPN.
129                                  * We just need to be sure we don't loop
130                                  * forever.
131                                  */
132                         } while (offset < BITS_PER_PAGE && qpn < QPN_MAX);
133                 }
134                 /*
135                  * In order to keep the number of pages allocated to a
136                  * minimum, we scan the all existing pages before increasing
137                  * the size of the bitmap table.
138                  */
139                 if (++i > max_scan) {
140                         if (qpt->nmaps == QPNMAP_ENTRIES)
141                                 break;
142                         map = &qpt->map[qpt->nmaps++];
143                         offset = 0;
144                 } else if (map < &qpt->map[qpt->nmaps]) {
145                         ++map;
146                         offset = 0;
147                 } else {
148                         map = &qpt->map[0];
149                         offset = 2;
150                 }
151                 qpn = mk_qpn(qpt, map, offset);
152         }
153
154         ret = 0;
155
156 bail:
157         return ret;
158 }
159
160 static void free_qpn(struct ipath_qp_table *qpt, u32 qpn)
161 {
162         struct qpn_map *map;
163
164         map = qpt->map + qpn / BITS_PER_PAGE;
165         if (map->page)
166                 clear_bit(qpn & BITS_PER_PAGE_MASK, map->page);
167         atomic_inc(&map->n_free);
168 }
169
170 /**
171  * ipath_alloc_qpn - allocate a QP number
172  * @qpt: the QP table
173  * @qp: the QP
174  * @type: the QP type (IB_QPT_SMI and IB_QPT_GSI are special)
175  *
176  * Allocate the next available QPN and put the QP into the hash table.
177  * The hash table holds a reference to the QP.
178  */
179 static int ipath_alloc_qpn(struct ipath_qp_table *qpt, struct ipath_qp *qp,
180                            enum ib_qp_type type)
181 {
182         unsigned long flags;
183         u32 qpn;
184         int ret;
185
186         if (type == IB_QPT_SMI)
187                 qpn = 0;
188         else if (type == IB_QPT_GSI)
189                 qpn = 1;
190         else {
191                 /* Allocate the next available QPN */
192                 qpn = alloc_qpn(qpt);
193                 if (qpn == 0) {
194                         ret = -ENOMEM;
195                         goto bail;
196                 }
197         }
198         qp->ibqp.qp_num = qpn;
199
200         /* Add the QP to the hash table. */
201         spin_lock_irqsave(&qpt->lock, flags);
202
203         qpn %= qpt->max;
204         qp->next = qpt->table[qpn];
205         qpt->table[qpn] = qp;
206         atomic_inc(&qp->refcount);
207
208         spin_unlock_irqrestore(&qpt->lock, flags);
209         ret = 0;
210
211 bail:
212         return ret;
213 }
214
215 /**
216  * ipath_free_qp - remove a QP from the QP table
217  * @qpt: the QP table
218  * @qp: the QP to remove
219  *
220  * Remove the QP from the table so it can't be found asynchronously by
221  * the receive interrupt routine.
222  */
223 static void ipath_free_qp(struct ipath_qp_table *qpt, struct ipath_qp *qp)
224 {
225         struct ipath_qp *q, **qpp;
226         unsigned long flags;
227         int fnd = 0;
228
229         spin_lock_irqsave(&qpt->lock, flags);
230
231         /* Remove QP from the hash table. */
232         qpp = &qpt->table[qp->ibqp.qp_num % qpt->max];
233         for (; (q = *qpp) != NULL; qpp = &q->next) {
234                 if (q == qp) {
235                         *qpp = qp->next;
236                         qp->next = NULL;
237                         atomic_dec(&qp->refcount);
238                         fnd = 1;
239                         break;
240                 }
241         }
242
243         spin_unlock_irqrestore(&qpt->lock, flags);
244
245         if (!fnd)
246                 return;
247
248         /* If QPN is not reserved, mark QPN free in the bitmap. */
249         if (qp->ibqp.qp_num > 1)
250                 free_qpn(qpt, qp->ibqp.qp_num);
251
252         wait_event(qp->wait, !atomic_read(&qp->refcount));
253 }
254
255 /**
256  * ipath_free_all_qps - remove all QPs from the table
257  * @qpt: the QP table to empty
258  */
259 void ipath_free_all_qps(struct ipath_qp_table *qpt)
260 {
261         unsigned long flags;
262         struct ipath_qp *qp, *nqp;
263         u32 n;
264
265         for (n = 0; n < qpt->max; n++) {
266                 spin_lock_irqsave(&qpt->lock, flags);
267                 qp = qpt->table[n];
268                 qpt->table[n] = NULL;
269                 spin_unlock_irqrestore(&qpt->lock, flags);
270
271                 while (qp) {
272                         nqp = qp->next;
273                         if (qp->ibqp.qp_num > 1)
274                                 free_qpn(qpt, qp->ibqp.qp_num);
275                         if (!atomic_dec_and_test(&qp->refcount) ||
276                             !ipath_destroy_qp(&qp->ibqp))
277                                 ipath_dbg(KERN_INFO "QP memory leak!\n");
278                         qp = nqp;
279                 }
280         }
281
282         for (n = 0; n < ARRAY_SIZE(qpt->map); n++) {
283                 if (qpt->map[n].page)
284                         free_page((unsigned long)qpt->map[n].page);
285         }
286 }
287
288 /**
289  * ipath_lookup_qpn - return the QP with the given QPN
290  * @qpt: the QP table
291  * @qpn: the QP number to look up
292  *
293  * The caller is responsible for decrementing the QP reference count
294  * when done.
295  */
296 struct ipath_qp *ipath_lookup_qpn(struct ipath_qp_table *qpt, u32 qpn)
297 {
298         unsigned long flags;
299         struct ipath_qp *qp;
300
301         spin_lock_irqsave(&qpt->lock, flags);
302
303         for (qp = qpt->table[qpn % qpt->max]; qp; qp = qp->next) {
304                 if (qp->ibqp.qp_num == qpn) {
305                         atomic_inc(&qp->refcount);
306                         break;
307                 }
308         }
309
310         spin_unlock_irqrestore(&qpt->lock, flags);
311         return qp;
312 }
313
314 /**
315  * ipath_reset_qp - initialize the QP state to the reset state
316  * @qp: the QP to reset
317  */
318 static void ipath_reset_qp(struct ipath_qp *qp)
319 {
320         qp->remote_qpn = 0;
321         qp->qkey = 0;
322         qp->qp_access_flags = 0;
323         clear_bit(IPATH_S_BUSY, &qp->s_flags);
324         qp->s_hdrwords = 0;
325         qp->s_psn = 0;
326         qp->r_psn = 0;
327         qp->r_msn = 0;
328         if (qp->ibqp.qp_type == IB_QPT_RC) {
329                 qp->s_state = IB_OPCODE_RC_SEND_LAST;
330                 qp->r_state = IB_OPCODE_RC_SEND_LAST;
331         } else {
332                 qp->s_state = IB_OPCODE_UC_SEND_LAST;
333                 qp->r_state = IB_OPCODE_UC_SEND_LAST;
334         }
335         qp->s_ack_state = IB_OPCODE_RC_ACKNOWLEDGE;
336         qp->r_ack_state = IB_OPCODE_RC_ACKNOWLEDGE;
337         qp->r_nak_state = 0;
338         qp->r_wrid_valid = 0;
339         qp->s_rnr_timeout = 0;
340         qp->s_head = 0;
341         qp->s_tail = 0;
342         qp->s_cur = 0;
343         qp->s_last = 0;
344         qp->s_ssn = 1;
345         qp->s_lsn = 0;
346         qp->s_wait_credit = 0;
347         if (qp->r_rq.wq) {
348                 qp->r_rq.wq->head = 0;
349                 qp->r_rq.wq->tail = 0;
350         }
351         qp->r_reuse_sge = 0;
352 }
353
354 /**
355  * ipath_error_qp - put a QP into an error state
356  * @qp: the QP to put into an error state
357  * @err: the receive completion error to signal if a RWQE is active
358  *
359  * Flushes both send and receive work queues.
360  * QP s_lock should be held and interrupts disabled.
361  */
362
363 void ipath_error_qp(struct ipath_qp *qp, enum ib_wc_status err)
364 {
365         struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
366         struct ib_wc wc;
367
368         ipath_dbg(KERN_INFO "QP%d/%d in error state\n",
369                   qp->ibqp.qp_num, qp->remote_qpn);
370
371         spin_lock(&dev->pending_lock);
372         /* XXX What if its already removed by the timeout code? */
373         if (!list_empty(&qp->timerwait))
374                 list_del_init(&qp->timerwait);
375         if (!list_empty(&qp->piowait))
376                 list_del_init(&qp->piowait);
377         spin_unlock(&dev->pending_lock);
378
379         wc.vendor_err = 0;
380         wc.byte_len = 0;
381         wc.imm_data = 0;
382         wc.qp_num = qp->ibqp.qp_num;
383         wc.src_qp = 0;
384         wc.wc_flags = 0;
385         wc.pkey_index = 0;
386         wc.slid = 0;
387         wc.sl = 0;
388         wc.dlid_path_bits = 0;
389         wc.port_num = 0;
390         if (qp->r_wrid_valid) {
391                 qp->r_wrid_valid = 0;
392                 wc.status = err;
393                 ipath_cq_enter(to_icq(qp->ibqp.send_cq), &wc, 1);
394         }
395         wc.status = IB_WC_WR_FLUSH_ERR;
396
397         while (qp->s_last != qp->s_head) {
398                 struct ipath_swqe *wqe = get_swqe_ptr(qp, qp->s_last);
399
400                 wc.wr_id = wqe->wr.wr_id;
401                 wc.opcode = ib_ipath_wc_opcode[wqe->wr.opcode];
402                 if (++qp->s_last >= qp->s_size)
403                         qp->s_last = 0;
404                 ipath_cq_enter(to_icq(qp->ibqp.send_cq), &wc, 1);
405         }
406         qp->s_cur = qp->s_tail = qp->s_head;
407         qp->s_hdrwords = 0;
408         qp->s_ack_state = IB_OPCODE_RC_ACKNOWLEDGE;
409
410         if (qp->r_rq.wq) {
411                 struct ipath_rwq *wq;
412                 u32 head;
413                 u32 tail;
414
415                 spin_lock(&qp->r_rq.lock);
416
417                 /* sanity check pointers before trusting them */
418                 wq = qp->r_rq.wq;
419                 head = wq->head;
420                 if (head >= qp->r_rq.size)
421                         head = 0;
422                 tail = wq->tail;
423                 if (tail >= qp->r_rq.size)
424                         tail = 0;
425                 wc.opcode = IB_WC_RECV;
426                 while (tail != head) {
427                         wc.wr_id = get_rwqe_ptr(&qp->r_rq, tail)->wr_id;
428                         if (++tail >= qp->r_rq.size)
429                                 tail = 0;
430                         ipath_cq_enter(to_icq(qp->ibqp.recv_cq), &wc, 1);
431                 }
432                 wq->tail = tail;
433
434                 spin_unlock(&qp->r_rq.lock);
435         }
436 }
437
438 /**
439  * ipath_modify_qp - modify the attributes of a queue pair
440  * @ibqp: the queue pair who's attributes we're modifying
441  * @attr: the new attributes
442  * @attr_mask: the mask of attributes to modify
443  * @udata: user data for ipathverbs.so
444  *
445  * Returns 0 on success, otherwise returns an errno.
446  */
447 int ipath_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
448                     int attr_mask, struct ib_udata *udata)
449 {
450         struct ipath_ibdev *dev = to_idev(ibqp->device);
451         struct ipath_qp *qp = to_iqp(ibqp);
452         enum ib_qp_state cur_state, new_state;
453         unsigned long flags;
454         int ret;
455
456         spin_lock_irqsave(&qp->s_lock, flags);
457
458         cur_state = attr_mask & IB_QP_CUR_STATE ?
459                 attr->cur_qp_state : qp->state;
460         new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
461
462         if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type,
463                                 attr_mask))
464                 goto inval;
465
466         if (attr_mask & IB_QP_AV) {
467                 if (attr->ah_attr.dlid == 0 ||
468                     attr->ah_attr.dlid >= IPATH_MULTICAST_LID_BASE)
469                         goto inval;
470
471                 if ((attr->ah_attr.ah_flags & IB_AH_GRH) &&
472                     (attr->ah_attr.grh.sgid_index > 1))
473                         goto inval;
474         }
475
476         if (attr_mask & IB_QP_PKEY_INDEX)
477                 if (attr->pkey_index >= ipath_get_npkeys(dev->dd))
478                         goto inval;
479
480         if (attr_mask & IB_QP_MIN_RNR_TIMER)
481                 if (attr->min_rnr_timer > 31)
482                         goto inval;
483
484         if (attr_mask & IB_QP_PORT)
485                 if (attr->port_num == 0 ||
486                     attr->port_num > ibqp->device->phys_port_cnt)
487                         goto inval;
488
489         if (attr_mask & IB_QP_PATH_MTU)
490                 if (attr->path_mtu > IB_MTU_4096)
491                         goto inval;
492
493         if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
494                 if (attr->max_dest_rd_atomic > 1)
495                         goto inval;
496
497         if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC)
498                 if (attr->max_rd_atomic > 1)
499                         goto inval;
500
501         if (attr_mask & IB_QP_PATH_MIG_STATE)
502                 if (attr->path_mig_state != IB_MIG_MIGRATED &&
503                     attr->path_mig_state != IB_MIG_REARM)
504                         goto inval;
505
506         switch (new_state) {
507         case IB_QPS_RESET:
508                 ipath_reset_qp(qp);
509                 break;
510
511         case IB_QPS_ERR:
512                 ipath_error_qp(qp, IB_WC_GENERAL_ERR);
513                 break;
514
515         default:
516                 break;
517
518         }
519
520         if (attr_mask & IB_QP_PKEY_INDEX)
521                 qp->s_pkey_index = attr->pkey_index;
522
523         if (attr_mask & IB_QP_DEST_QPN)
524                 qp->remote_qpn = attr->dest_qp_num;
525
526         if (attr_mask & IB_QP_SQ_PSN) {
527                 qp->s_psn = qp->s_next_psn = attr->sq_psn;
528                 qp->s_last_psn = qp->s_next_psn - 1;
529         }
530
531         if (attr_mask & IB_QP_RQ_PSN)
532                 qp->r_psn = attr->rq_psn;
533
534         if (attr_mask & IB_QP_ACCESS_FLAGS)
535                 qp->qp_access_flags = attr->qp_access_flags;
536
537         if (attr_mask & IB_QP_AV)
538                 qp->remote_ah_attr = attr->ah_attr;
539
540         if (attr_mask & IB_QP_PATH_MTU)
541                 qp->path_mtu = attr->path_mtu;
542
543         if (attr_mask & IB_QP_RETRY_CNT)
544                 qp->s_retry = qp->s_retry_cnt = attr->retry_cnt;
545
546         if (attr_mask & IB_QP_RNR_RETRY) {
547                 qp->s_rnr_retry = attr->rnr_retry;
548                 if (qp->s_rnr_retry > 7)
549                         qp->s_rnr_retry = 7;
550                 qp->s_rnr_retry_cnt = qp->s_rnr_retry;
551         }
552
553         if (attr_mask & IB_QP_MIN_RNR_TIMER)
554                 qp->r_min_rnr_timer = attr->min_rnr_timer;
555
556         if (attr_mask & IB_QP_TIMEOUT)
557                 qp->timeout = attr->timeout;
558
559         if (attr_mask & IB_QP_QKEY)
560                 qp->qkey = attr->qkey;
561
562         qp->state = new_state;
563         spin_unlock_irqrestore(&qp->s_lock, flags);
564
565         ret = 0;
566         goto bail;
567
568 inval:
569         spin_unlock_irqrestore(&qp->s_lock, flags);
570         ret = -EINVAL;
571
572 bail:
573         return ret;
574 }
575
576 int ipath_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
577                    int attr_mask, struct ib_qp_init_attr *init_attr)
578 {
579         struct ipath_qp *qp = to_iqp(ibqp);
580
581         attr->qp_state = qp->state;
582         attr->cur_qp_state = attr->qp_state;
583         attr->path_mtu = qp->path_mtu;
584         attr->path_mig_state = 0;
585         attr->qkey = qp->qkey;
586         attr->rq_psn = qp->r_psn;
587         attr->sq_psn = qp->s_next_psn;
588         attr->dest_qp_num = qp->remote_qpn;
589         attr->qp_access_flags = qp->qp_access_flags;
590         attr->cap.max_send_wr = qp->s_size - 1;
591         attr->cap.max_recv_wr = qp->ibqp.srq ? 0 : qp->r_rq.size - 1;
592         attr->cap.max_send_sge = qp->s_max_sge;
593         attr->cap.max_recv_sge = qp->r_rq.max_sge;
594         attr->cap.max_inline_data = 0;
595         attr->ah_attr = qp->remote_ah_attr;
596         memset(&attr->alt_ah_attr, 0, sizeof(attr->alt_ah_attr));
597         attr->pkey_index = qp->s_pkey_index;
598         attr->alt_pkey_index = 0;
599         attr->en_sqd_async_notify = 0;
600         attr->sq_draining = 0;
601         attr->max_rd_atomic = 1;
602         attr->max_dest_rd_atomic = 1;
603         attr->min_rnr_timer = qp->r_min_rnr_timer;
604         attr->port_num = 1;
605         attr->timeout = qp->timeout;
606         attr->retry_cnt = qp->s_retry_cnt;
607         attr->rnr_retry = qp->s_rnr_retry;
608         attr->alt_port_num = 0;
609         attr->alt_timeout = 0;
610
611         init_attr->event_handler = qp->ibqp.event_handler;
612         init_attr->qp_context = qp->ibqp.qp_context;
613         init_attr->send_cq = qp->ibqp.send_cq;
614         init_attr->recv_cq = qp->ibqp.recv_cq;
615         init_attr->srq = qp->ibqp.srq;
616         init_attr->cap = attr->cap;
617         if (qp->s_flags & (1 << IPATH_S_SIGNAL_REQ_WR))
618                 init_attr->sq_sig_type = IB_SIGNAL_REQ_WR;
619         else
620                 init_attr->sq_sig_type = IB_SIGNAL_ALL_WR;
621         init_attr->qp_type = qp->ibqp.qp_type;
622         init_attr->port_num = 1;
623         return 0;
624 }
625
626 /**
627  * ipath_compute_aeth - compute the AETH (syndrome + MSN)
628  * @qp: the queue pair to compute the AETH for
629  *
630  * Returns the AETH.
631  */
632 __be32 ipath_compute_aeth(struct ipath_qp *qp)
633 {
634         u32 aeth = qp->r_msn & IPATH_MSN_MASK;
635
636         if (qp->ibqp.srq) {
637                 /*
638                  * Shared receive queues don't generate credits.
639                  * Set the credit field to the invalid value.
640                  */
641                 aeth |= IPATH_AETH_CREDIT_INVAL << IPATH_AETH_CREDIT_SHIFT;
642         } else {
643                 u32 min, max, x;
644                 u32 credits;
645                 struct ipath_rwq *wq = qp->r_rq.wq;
646                 u32 head;
647                 u32 tail;
648
649                 /* sanity check pointers before trusting them */
650                 head = wq->head;
651                 if (head >= qp->r_rq.size)
652                         head = 0;
653                 tail = wq->tail;
654                 if (tail >= qp->r_rq.size)
655                         tail = 0;
656                 /*
657                  * Compute the number of credits available (RWQEs).
658                  * XXX Not holding the r_rq.lock here so there is a small
659                  * chance that the pair of reads are not atomic.
660                  */
661                 credits = head - tail;
662                 if ((int)credits < 0)
663                         credits += qp->r_rq.size;
664                 /*
665                  * Binary search the credit table to find the code to
666                  * use.
667                  */
668                 min = 0;
669                 max = 31;
670                 for (;;) {
671                         x = (min + max) / 2;
672                         if (credit_table[x] == credits)
673                                 break;
674                         if (credit_table[x] > credits)
675                                 max = x;
676                         else if (min == x)
677                                 break;
678                         else
679                                 min = x;
680                 }
681                 aeth |= x << IPATH_AETH_CREDIT_SHIFT;
682         }
683         return cpu_to_be32(aeth);
684 }
685
686 /**
687  * ipath_create_qp - create a queue pair for a device
688  * @ibpd: the protection domain who's device we create the queue pair for
689  * @init_attr: the attributes of the queue pair
690  * @udata: unused by InfiniPath
691  *
692  * Returns the queue pair on success, otherwise returns an errno.
693  *
694  * Called by the ib_create_qp() core verbs function.
695  */
696 struct ib_qp *ipath_create_qp(struct ib_pd *ibpd,
697                               struct ib_qp_init_attr *init_attr,
698                               struct ib_udata *udata)
699 {
700         struct ipath_qp *qp;
701         int err;
702         struct ipath_swqe *swq = NULL;
703         struct ipath_ibdev *dev;
704         size_t sz;
705         struct ib_qp *ret;
706
707         if (init_attr->cap.max_send_sge > ib_ipath_max_sges ||
708             init_attr->cap.max_recv_sge > ib_ipath_max_sges ||
709             init_attr->cap.max_send_wr > ib_ipath_max_qp_wrs ||
710             init_attr->cap.max_recv_wr > ib_ipath_max_qp_wrs) {
711                 ret = ERR_PTR(-ENOMEM);
712                 goto bail;
713         }
714
715         if (init_attr->cap.max_send_sge +
716             init_attr->cap.max_recv_sge +
717             init_attr->cap.max_send_wr +
718             init_attr->cap.max_recv_wr == 0) {
719                 ret = ERR_PTR(-EINVAL);
720                 goto bail;
721         }
722
723         switch (init_attr->qp_type) {
724         case IB_QPT_UC:
725         case IB_QPT_RC:
726                 sz = sizeof(struct ipath_sge) *
727                         init_attr->cap.max_send_sge +
728                         sizeof(struct ipath_swqe);
729                 swq = vmalloc((init_attr->cap.max_send_wr + 1) * sz);
730                 if (swq == NULL) {
731                         ret = ERR_PTR(-ENOMEM);
732                         goto bail;
733                 }
734                 /* FALLTHROUGH */
735         case IB_QPT_UD:
736         case IB_QPT_SMI:
737         case IB_QPT_GSI:
738                 sz = sizeof(*qp);
739                 if (init_attr->srq) {
740                         struct ipath_srq *srq = to_isrq(init_attr->srq);
741
742                         sz += sizeof(*qp->r_sg_list) *
743                                 srq->rq.max_sge;
744                 } else
745                         sz += sizeof(*qp->r_sg_list) *
746                                 init_attr->cap.max_recv_sge;
747                 qp = kmalloc(sz, GFP_KERNEL);
748                 if (!qp) {
749                         ret = ERR_PTR(-ENOMEM);
750                         goto bail_swq;
751                 }
752                 if (init_attr->srq) {
753                         sz = 0;
754                         qp->r_rq.size = 0;
755                         qp->r_rq.max_sge = 0;
756                         qp->r_rq.wq = NULL;
757                         init_attr->cap.max_recv_wr = 0;
758                         init_attr->cap.max_recv_sge = 0;
759                 } else {
760                         qp->r_rq.size = init_attr->cap.max_recv_wr + 1;
761                         qp->r_rq.max_sge = init_attr->cap.max_recv_sge;
762                         sz = (sizeof(struct ib_sge) * qp->r_rq.max_sge) +
763                                 sizeof(struct ipath_rwqe);
764                         qp->r_rq.wq = vmalloc_user(sizeof(struct ipath_rwq) +
765                                               qp->r_rq.size * sz);
766                         if (!qp->r_rq.wq) {
767                                 ret = ERR_PTR(-ENOMEM);
768                                 goto bail_qp;
769                         }
770                 }
771
772                 /*
773                  * ib_create_qp() will initialize qp->ibqp
774                  * except for qp->ibqp.qp_num.
775                  */
776                 spin_lock_init(&qp->s_lock);
777                 spin_lock_init(&qp->r_rq.lock);
778                 atomic_set(&qp->refcount, 0);
779                 init_waitqueue_head(&qp->wait);
780                 tasklet_init(&qp->s_task, ipath_do_ruc_send,
781                              (unsigned long)qp);
782                 INIT_LIST_HEAD(&qp->piowait);
783                 INIT_LIST_HEAD(&qp->timerwait);
784                 qp->state = IB_QPS_RESET;
785                 qp->s_wq = swq;
786                 qp->s_size = init_attr->cap.max_send_wr + 1;
787                 qp->s_max_sge = init_attr->cap.max_send_sge;
788                 if (init_attr->sq_sig_type == IB_SIGNAL_REQ_WR)
789                         qp->s_flags = 1 << IPATH_S_SIGNAL_REQ_WR;
790                 else
791                         qp->s_flags = 0;
792                 dev = to_idev(ibpd->device);
793                 err = ipath_alloc_qpn(&dev->qp_table, qp,
794                                       init_attr->qp_type);
795                 if (err) {
796                         ret = ERR_PTR(err);
797                         goto bail_rwq;
798                 }
799                 qp->ip = NULL;
800                 ipath_reset_qp(qp);
801                 break;
802
803         default:
804                 /* Don't support raw QPs */
805                 ret = ERR_PTR(-ENOSYS);
806                 goto bail;
807         }
808
809         init_attr->cap.max_inline_data = 0;
810
811         /*
812          * Return the address of the RWQ as the offset to mmap.
813          * See ipath_mmap() for details.
814          */
815         if (udata && udata->outlen >= sizeof(__u64)) {
816                 struct ipath_mmap_info *ip;
817                 __u64 offset = (__u64) qp->r_rq.wq;
818                 int err;
819
820                 err = ib_copy_to_udata(udata, &offset, sizeof(offset));
821                 if (err) {
822                         ret = ERR_PTR(err);
823                         goto bail_rwq;
824                 }
825
826                 if (qp->r_rq.wq) {
827                         /* Allocate info for ipath_mmap(). */
828                         ip = kmalloc(sizeof(*ip), GFP_KERNEL);
829                         if (!ip) {
830                                 ret = ERR_PTR(-ENOMEM);
831                                 goto bail_rwq;
832                         }
833                         qp->ip = ip;
834                         ip->context = ibpd->uobject->context;
835                         ip->obj = qp->r_rq.wq;
836                         kref_init(&ip->ref);
837                         ip->mmap_cnt = 0;
838                         ip->size = PAGE_ALIGN(sizeof(struct ipath_rwq) +
839                                               qp->r_rq.size * sz);
840                         spin_lock_irq(&dev->pending_lock);
841                         ip->next = dev->pending_mmaps;
842                         dev->pending_mmaps = ip;
843                         spin_unlock_irq(&dev->pending_lock);
844                 }
845         }
846
847         spin_lock(&dev->n_qps_lock);
848         if (dev->n_qps_allocated == ib_ipath_max_qps) {
849                 spin_unlock(&dev->n_qps_lock);
850                 ret = ERR_PTR(-ENOMEM);
851                 goto bail_ip;
852         }
853
854         dev->n_qps_allocated++;
855         spin_unlock(&dev->n_qps_lock);
856
857         ret = &qp->ibqp;
858         goto bail;
859
860 bail_ip:
861         kfree(qp->ip);
862 bail_rwq:
863         vfree(qp->r_rq.wq);
864 bail_qp:
865         kfree(qp);
866 bail_swq:
867         vfree(swq);
868 bail:
869         return ret;
870 }
871
872 /**
873  * ipath_destroy_qp - destroy a queue pair
874  * @ibqp: the queue pair to destroy
875  *
876  * Returns 0 on success.
877  *
878  * Note that this can be called while the QP is actively sending or
879  * receiving!
880  */
881 int ipath_destroy_qp(struct ib_qp *ibqp)
882 {
883         struct ipath_qp *qp = to_iqp(ibqp);
884         struct ipath_ibdev *dev = to_idev(ibqp->device);
885         unsigned long flags;
886
887         spin_lock_irqsave(&qp->s_lock, flags);
888         qp->state = IB_QPS_ERR;
889         spin_unlock_irqrestore(&qp->s_lock, flags);
890         spin_lock(&dev->n_qps_lock);
891         dev->n_qps_allocated--;
892         spin_unlock(&dev->n_qps_lock);
893
894         /* Stop the sending tasklet. */
895         tasklet_kill(&qp->s_task);
896
897         /* Make sure the QP isn't on the timeout list. */
898         spin_lock_irqsave(&dev->pending_lock, flags);
899         if (!list_empty(&qp->timerwait))
900                 list_del_init(&qp->timerwait);
901         if (!list_empty(&qp->piowait))
902                 list_del_init(&qp->piowait);
903         spin_unlock_irqrestore(&dev->pending_lock, flags);
904
905         /*
906          * Make sure that the QP is not in the QPN table so receive
907          * interrupts will discard packets for this QP.  XXX Also remove QP
908          * from multicast table.
909          */
910         if (atomic_read(&qp->refcount) != 0)
911                 ipath_free_qp(&dev->qp_table, qp);
912
913         if (qp->ip)
914                 kref_put(&qp->ip->ref, ipath_release_mmap_info);
915         else
916                 vfree(qp->r_rq.wq);
917         vfree(qp->s_wq);
918         kfree(qp);
919         return 0;
920 }
921
922 /**
923  * ipath_init_qp_table - initialize the QP table for a device
924  * @idev: the device who's QP table we're initializing
925  * @size: the size of the QP table
926  *
927  * Returns 0 on success, otherwise returns an errno.
928  */
929 int ipath_init_qp_table(struct ipath_ibdev *idev, int size)
930 {
931         int i;
932         int ret;
933
934         idev->qp_table.last = 1;        /* QPN 0 and 1 are special. */
935         idev->qp_table.max = size;
936         idev->qp_table.nmaps = 1;
937         idev->qp_table.table = kzalloc(size * sizeof(*idev->qp_table.table),
938                                        GFP_KERNEL);
939         if (idev->qp_table.table == NULL) {
940                 ret = -ENOMEM;
941                 goto bail;
942         }
943
944         for (i = 0; i < ARRAY_SIZE(idev->qp_table.map); i++) {
945                 atomic_set(&idev->qp_table.map[i].n_free, BITS_PER_PAGE);
946                 idev->qp_table.map[i].page = NULL;
947         }
948
949         ret = 0;
950
951 bail:
952         return ret;
953 }
954
955 /**
956  * ipath_sqerror_qp - put a QP's send queue into an error state
957  * @qp: QP who's send queue will be put into an error state
958  * @wc: the WC responsible for putting the QP in this state
959  *
960  * Flushes the send work queue.
961  * The QP s_lock should be held.
962  */
963
964 void ipath_sqerror_qp(struct ipath_qp *qp, struct ib_wc *wc)
965 {
966         struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
967         struct ipath_swqe *wqe = get_swqe_ptr(qp, qp->s_last);
968
969         ipath_dbg(KERN_INFO "Send queue error on QP%d/%d: err: %d\n",
970                   qp->ibqp.qp_num, qp->remote_qpn, wc->status);
971
972         spin_lock(&dev->pending_lock);
973         /* XXX What if its already removed by the timeout code? */
974         if (!list_empty(&qp->timerwait))
975                 list_del_init(&qp->timerwait);
976         if (!list_empty(&qp->piowait))
977                 list_del_init(&qp->piowait);
978         spin_unlock(&dev->pending_lock);
979
980         ipath_cq_enter(to_icq(qp->ibqp.send_cq), wc, 1);
981         if (++qp->s_last >= qp->s_size)
982                 qp->s_last = 0;
983
984         wc->status = IB_WC_WR_FLUSH_ERR;
985
986         while (qp->s_last != qp->s_head) {
987                 wc->wr_id = wqe->wr.wr_id;
988                 wc->opcode = ib_ipath_wc_opcode[wqe->wr.opcode];
989                 ipath_cq_enter(to_icq(qp->ibqp.send_cq), wc, 1);
990                 if (++qp->s_last >= qp->s_size)
991                         qp->s_last = 0;
992                 wqe = get_swqe_ptr(qp, qp->s_last);
993         }
994         qp->s_cur = qp->s_tail = qp->s_head;
995         qp->state = IB_QPS_SQE;
996 }
997
998 /**
999  * ipath_get_credit - flush the send work queue of a QP
1000  * @qp: the qp who's send work queue to flush
1001  * @aeth: the Acknowledge Extended Transport Header
1002  *
1003  * The QP s_lock should be held.
1004  */
1005 void ipath_get_credit(struct ipath_qp *qp, u32 aeth)
1006 {
1007         u32 credit = (aeth >> IPATH_AETH_CREDIT_SHIFT) & IPATH_AETH_CREDIT_MASK;
1008
1009         /*
1010          * If the credit is invalid, we can send
1011          * as many packets as we like.  Otherwise, we have to
1012          * honor the credit field.
1013          */
1014         if (credit == IPATH_AETH_CREDIT_INVAL)
1015                 qp->s_lsn = (u32) -1;
1016         else if (qp->s_lsn != (u32) -1) {
1017                 /* Compute new LSN (i.e., MSN + credit) */
1018                 credit = (aeth + credit_table[credit]) & IPATH_MSN_MASK;
1019                 if (ipath_cmp24(credit, qp->s_lsn) > 0)
1020                         qp->s_lsn = credit;
1021         }
1022
1023         /* Restart sending if it was blocked due to lack of credits. */
1024         if (qp->s_cur != qp->s_head &&
1025             (qp->s_lsn == (u32) -1 ||
1026              ipath_cmp24(get_swqe_ptr(qp, qp->s_cur)->ssn,
1027                          qp->s_lsn + 1) <= 0))
1028                 tasklet_hi_schedule(&qp->s_task);
1029 }