Pull battery into release branch
[linux-2.6] / drivers / infiniband / hw / ipath / ipath_qp.c
1 /*
2  * Copyright (c) 2006, 2007 QLogic Corporation. All rights reserved.
3  * Copyright (c) 2005, 2006 PathScale, Inc. All rights reserved.
4  *
5  * This software is available to you under a choice of one of two
6  * licenses.  You may choose to be licensed under the terms of the GNU
7  * General Public License (GPL) Version 2, available from the file
8  * COPYING in the main directory of this source tree, or the
9  * OpenIB.org BSD license below:
10  *
11  *     Redistribution and use in source and binary forms, with or
12  *     without modification, are permitted provided that the following
13  *     conditions are met:
14  *
15  *      - Redistributions of source code must retain the above
16  *        copyright notice, this list of conditions and the following
17  *        disclaimer.
18  *
19  *      - Redistributions in binary form must reproduce the above
20  *        copyright notice, this list of conditions and the following
21  *        disclaimer in the documentation and/or other materials
22  *        provided with the distribution.
23  *
24  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
25  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
26  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
27  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
28  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
29  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
30  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
31  * SOFTWARE.
32  */
33
34 #include <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
85 static void get_map_page(struct ipath_qp_table *qpt, struct qpn_map *map)
86 {
87         unsigned long page = get_zeroed_page(GFP_KERNEL);
88         unsigned long flags;
89
90         /*
91          * Free the page if someone raced with us installing it.
92          */
93
94         spin_lock_irqsave(&qpt->lock, flags);
95         if (map->page)
96                 free_page(page);
97         else
98                 map->page = (void *)page;
99         spin_unlock_irqrestore(&qpt->lock, flags);
100 }
101
102
103 static int alloc_qpn(struct ipath_qp_table *qpt, enum ib_qp_type type)
104 {
105         u32 i, offset, max_scan, qpn;
106         struct qpn_map *map;
107         u32 ret = -1;
108
109         if (type == IB_QPT_SMI)
110                 ret = 0;
111         else if (type == IB_QPT_GSI)
112                 ret = 1;
113
114         if (ret != -1) {
115                 map = &qpt->map[0];
116                 if (unlikely(!map->page)) {
117                         get_map_page(qpt, map);
118                         if (unlikely(!map->page)) {
119                                 ret = -ENOMEM;
120                                 goto bail;
121                         }
122                 }
123                 if (!test_and_set_bit(ret, map->page))
124                         atomic_dec(&map->n_free);
125                 else
126                         ret = -EBUSY;
127                 goto bail;
128         }
129
130         qpn = qpt->last + 1;
131         if (qpn >= QPN_MAX)
132                 qpn = 2;
133         offset = qpn & BITS_PER_PAGE_MASK;
134         map = &qpt->map[qpn / BITS_PER_PAGE];
135         max_scan = qpt->nmaps - !offset;
136         for (i = 0;;) {
137                 if (unlikely(!map->page)) {
138                         get_map_page(qpt, map);
139                         if (unlikely(!map->page))
140                                 break;
141                 }
142                 if (likely(atomic_read(&map->n_free))) {
143                         do {
144                                 if (!test_and_set_bit(offset, map->page)) {
145                                         atomic_dec(&map->n_free);
146                                         qpt->last = qpn;
147                                         ret = qpn;
148                                         goto bail;
149                                 }
150                                 offset = find_next_offset(map, offset);
151                                 qpn = mk_qpn(qpt, map, offset);
152                                 /*
153                                  * This test differs from alloc_pidmap().
154                                  * If find_next_offset() does find a zero
155                                  * bit, we don't need to check for QPN
156                                  * wrapping around past our starting QPN.
157                                  * We just need to be sure we don't loop
158                                  * forever.
159                                  */
160                         } while (offset < BITS_PER_PAGE && qpn < QPN_MAX);
161                 }
162                 /*
163                  * In order to keep the number of pages allocated to a
164                  * minimum, we scan the all existing pages before increasing
165                  * the size of the bitmap table.
166                  */
167                 if (++i > max_scan) {
168                         if (qpt->nmaps == QPNMAP_ENTRIES)
169                                 break;
170                         map = &qpt->map[qpt->nmaps++];
171                         offset = 0;
172                 } else if (map < &qpt->map[qpt->nmaps]) {
173                         ++map;
174                         offset = 0;
175                 } else {
176                         map = &qpt->map[0];
177                         offset = 2;
178                 }
179                 qpn = mk_qpn(qpt, map, offset);
180         }
181
182         ret = -ENOMEM;
183
184 bail:
185         return ret;
186 }
187
188 static void free_qpn(struct ipath_qp_table *qpt, u32 qpn)
189 {
190         struct qpn_map *map;
191
192         map = qpt->map + qpn / BITS_PER_PAGE;
193         if (map->page)
194                 clear_bit(qpn & BITS_PER_PAGE_MASK, map->page);
195         atomic_inc(&map->n_free);
196 }
197
198 /**
199  * ipath_alloc_qpn - allocate a QP number
200  * @qpt: the QP table
201  * @qp: the QP
202  * @type: the QP type (IB_QPT_SMI and IB_QPT_GSI are special)
203  *
204  * Allocate the next available QPN and put the QP into the hash table.
205  * The hash table holds a reference to the QP.
206  */
207 static int ipath_alloc_qpn(struct ipath_qp_table *qpt, struct ipath_qp *qp,
208                            enum ib_qp_type type)
209 {
210         unsigned long flags;
211         int ret;
212
213         ret = alloc_qpn(qpt, type);
214         if (ret < 0)
215                 goto bail;
216         qp->ibqp.qp_num = ret;
217
218         /* Add the QP to the hash table. */
219         spin_lock_irqsave(&qpt->lock, flags);
220
221         ret %= qpt->max;
222         qp->next = qpt->table[ret];
223         qpt->table[ret] = qp;
224         atomic_inc(&qp->refcount);
225
226         spin_unlock_irqrestore(&qpt->lock, flags);
227         ret = 0;
228
229 bail:
230         return ret;
231 }
232
233 /**
234  * ipath_free_qp - remove a QP from the QP table
235  * @qpt: the QP table
236  * @qp: the QP to remove
237  *
238  * Remove the QP from the table so it can't be found asynchronously by
239  * the receive interrupt routine.
240  */
241 static void ipath_free_qp(struct ipath_qp_table *qpt, struct ipath_qp *qp)
242 {
243         struct ipath_qp *q, **qpp;
244         unsigned long flags;
245         int fnd = 0;
246
247         spin_lock_irqsave(&qpt->lock, flags);
248
249         /* Remove QP from the hash table. */
250         qpp = &qpt->table[qp->ibqp.qp_num % qpt->max];
251         for (; (q = *qpp) != NULL; qpp = &q->next) {
252                 if (q == qp) {
253                         *qpp = qp->next;
254                         qp->next = NULL;
255                         atomic_dec(&qp->refcount);
256                         fnd = 1;
257                         break;
258                 }
259         }
260
261         spin_unlock_irqrestore(&qpt->lock, flags);
262
263         if (!fnd)
264                 return;
265
266         free_qpn(qpt, qp->ibqp.qp_num);
267
268         wait_event(qp->wait, !atomic_read(&qp->refcount));
269 }
270
271 /**
272  * ipath_free_all_qps - remove all QPs from the table
273  * @qpt: the QP table to empty
274  */
275 void ipath_free_all_qps(struct ipath_qp_table *qpt)
276 {
277         unsigned long flags;
278         struct ipath_qp *qp, *nqp;
279         u32 n;
280
281         for (n = 0; n < qpt->max; n++) {
282                 spin_lock_irqsave(&qpt->lock, flags);
283                 qp = qpt->table[n];
284                 qpt->table[n] = NULL;
285                 spin_unlock_irqrestore(&qpt->lock, flags);
286
287                 while (qp) {
288                         nqp = qp->next;
289                         free_qpn(qpt, qp->ibqp.qp_num);
290                         if (!atomic_dec_and_test(&qp->refcount) ||
291                             !ipath_destroy_qp(&qp->ibqp))
292                                 ipath_dbg("QP memory leak!\n");
293                         qp = nqp;
294                 }
295         }
296
297         for (n = 0; n < ARRAY_SIZE(qpt->map); n++) {
298                 if (qpt->map[n].page)
299                         free_page((unsigned long)qpt->map[n].page);
300         }
301 }
302
303 /**
304  * ipath_lookup_qpn - return the QP with the given QPN
305  * @qpt: the QP table
306  * @qpn: the QP number to look up
307  *
308  * The caller is responsible for decrementing the QP reference count
309  * when done.
310  */
311 struct ipath_qp *ipath_lookup_qpn(struct ipath_qp_table *qpt, u32 qpn)
312 {
313         unsigned long flags;
314         struct ipath_qp *qp;
315
316         spin_lock_irqsave(&qpt->lock, flags);
317
318         for (qp = qpt->table[qpn % qpt->max]; qp; qp = qp->next) {
319                 if (qp->ibqp.qp_num == qpn) {
320                         atomic_inc(&qp->refcount);
321                         break;
322                 }
323         }
324
325         spin_unlock_irqrestore(&qpt->lock, flags);
326         return qp;
327 }
328
329 /**
330  * ipath_reset_qp - initialize the QP state to the reset state
331  * @qp: the QP to reset
332  */
333 static void ipath_reset_qp(struct ipath_qp *qp)
334 {
335         qp->remote_qpn = 0;
336         qp->qkey = 0;
337         qp->qp_access_flags = 0;
338         qp->s_busy = 0;
339         qp->s_flags &= IPATH_S_SIGNAL_REQ_WR;
340         qp->s_hdrwords = 0;
341         qp->s_psn = 0;
342         qp->r_psn = 0;
343         qp->r_msn = 0;
344         if (qp->ibqp.qp_type == IB_QPT_RC) {
345                 qp->s_state = IB_OPCODE_RC_SEND_LAST;
346                 qp->r_state = IB_OPCODE_RC_SEND_LAST;
347         } else {
348                 qp->s_state = IB_OPCODE_UC_SEND_LAST;
349                 qp->r_state = IB_OPCODE_UC_SEND_LAST;
350         }
351         qp->s_ack_state = IB_OPCODE_RC_ACKNOWLEDGE;
352         qp->r_nak_state = 0;
353         qp->r_wrid_valid = 0;
354         qp->s_rnr_timeout = 0;
355         qp->s_head = 0;
356         qp->s_tail = 0;
357         qp->s_cur = 0;
358         qp->s_last = 0;
359         qp->s_ssn = 1;
360         qp->s_lsn = 0;
361         qp->s_wait_credit = 0;
362         memset(qp->s_ack_queue, 0, sizeof(qp->s_ack_queue));
363         qp->r_head_ack_queue = 0;
364         qp->s_tail_ack_queue = 0;
365         qp->s_num_rd_atomic = 0;
366         if (qp->r_rq.wq) {
367                 qp->r_rq.wq->head = 0;
368                 qp->r_rq.wq->tail = 0;
369         }
370         qp->r_reuse_sge = 0;
371 }
372
373 /**
374  * ipath_error_qp - put a QP into an error state
375  * @qp: the QP to put into an error state
376  * @err: the receive completion error to signal if a RWQE is active
377  *
378  * Flushes both send and receive work queues.
379  * The QP s_lock should be held and interrupts disabled.
380  */
381
382 void ipath_error_qp(struct ipath_qp *qp, enum ib_wc_status err)
383 {
384         struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
385         struct ib_wc wc;
386
387         ipath_dbg("QP%d/%d in error state\n",
388                   qp->ibqp.qp_num, qp->remote_qpn);
389
390         spin_lock(&dev->pending_lock);
391         /* XXX What if its already removed by the timeout code? */
392         if (!list_empty(&qp->timerwait))
393                 list_del_init(&qp->timerwait);
394         if (!list_empty(&qp->piowait))
395                 list_del_init(&qp->piowait);
396         spin_unlock(&dev->pending_lock);
397
398         wc.vendor_err = 0;
399         wc.byte_len = 0;
400         wc.imm_data = 0;
401         wc.qp = &qp->ibqp;
402         wc.src_qp = 0;
403         wc.wc_flags = 0;
404         wc.pkey_index = 0;
405         wc.slid = 0;
406         wc.sl = 0;
407         wc.dlid_path_bits = 0;
408         wc.port_num = 0;
409         if (qp->r_wrid_valid) {
410                 qp->r_wrid_valid = 0;
411                 wc.wr_id = qp->r_wr_id;
412                 wc.opcode = IB_WC_RECV;
413                 wc.status = err;
414                 ipath_cq_enter(to_icq(qp->ibqp.send_cq), &wc, 1);
415         }
416         wc.status = IB_WC_WR_FLUSH_ERR;
417
418         while (qp->s_last != qp->s_head) {
419                 struct ipath_swqe *wqe = get_swqe_ptr(qp, qp->s_last);
420
421                 wc.wr_id = wqe->wr.wr_id;
422                 wc.opcode = ib_ipath_wc_opcode[wqe->wr.opcode];
423                 if (++qp->s_last >= qp->s_size)
424                         qp->s_last = 0;
425                 ipath_cq_enter(to_icq(qp->ibqp.send_cq), &wc, 1);
426         }
427         qp->s_cur = qp->s_tail = qp->s_head;
428         qp->s_hdrwords = 0;
429         qp->s_ack_state = IB_OPCODE_RC_ACKNOWLEDGE;
430
431         if (qp->r_rq.wq) {
432                 struct ipath_rwq *wq;
433                 u32 head;
434                 u32 tail;
435
436                 spin_lock(&qp->r_rq.lock);
437
438                 /* sanity check pointers before trusting them */
439                 wq = qp->r_rq.wq;
440                 head = wq->head;
441                 if (head >= qp->r_rq.size)
442                         head = 0;
443                 tail = wq->tail;
444                 if (tail >= qp->r_rq.size)
445                         tail = 0;
446                 wc.opcode = IB_WC_RECV;
447                 while (tail != head) {
448                         wc.wr_id = get_rwqe_ptr(&qp->r_rq, tail)->wr_id;
449                         if (++tail >= qp->r_rq.size)
450                                 tail = 0;
451                         ipath_cq_enter(to_icq(qp->ibqp.recv_cq), &wc, 1);
452                 }
453                 wq->tail = tail;
454
455                 spin_unlock(&qp->r_rq.lock);
456         }
457 }
458
459 /**
460  * ipath_modify_qp - modify the attributes of a queue pair
461  * @ibqp: the queue pair who's attributes we're modifying
462  * @attr: the new attributes
463  * @attr_mask: the mask of attributes to modify
464  * @udata: user data for ipathverbs.so
465  *
466  * Returns 0 on success, otherwise returns an errno.
467  */
468 int ipath_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
469                     int attr_mask, struct ib_udata *udata)
470 {
471         struct ipath_ibdev *dev = to_idev(ibqp->device);
472         struct ipath_qp *qp = to_iqp(ibqp);
473         enum ib_qp_state cur_state, new_state;
474         unsigned long flags;
475         int ret;
476
477         spin_lock_irqsave(&qp->s_lock, flags);
478
479         cur_state = attr_mask & IB_QP_CUR_STATE ?
480                 attr->cur_qp_state : qp->state;
481         new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
482
483         if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type,
484                                 attr_mask))
485                 goto inval;
486
487         if (attr_mask & IB_QP_AV) {
488                 if (attr->ah_attr.dlid == 0 ||
489                     attr->ah_attr.dlid >= IPATH_MULTICAST_LID_BASE)
490                         goto inval;
491
492                 if ((attr->ah_attr.ah_flags & IB_AH_GRH) &&
493                     (attr->ah_attr.grh.sgid_index > 1))
494                         goto inval;
495         }
496
497         if (attr_mask & IB_QP_PKEY_INDEX)
498                 if (attr->pkey_index >= ipath_get_npkeys(dev->dd))
499                         goto inval;
500
501         if (attr_mask & IB_QP_MIN_RNR_TIMER)
502                 if (attr->min_rnr_timer > 31)
503                         goto inval;
504
505         if (attr_mask & IB_QP_PORT)
506                 if (attr->port_num == 0 ||
507                     attr->port_num > ibqp->device->phys_port_cnt)
508                         goto inval;
509
510         /*
511          * Note: the chips support a maximum MTU of 4096, but the driver
512          * hasn't implemented this feature yet, so don't allow Path MTU
513          * values greater than 2048.
514          */
515         if (attr_mask & IB_QP_PATH_MTU)
516                 if (attr->path_mtu > IB_MTU_2048)
517                         goto inval;
518
519         if (attr_mask & IB_QP_PATH_MIG_STATE)
520                 if (attr->path_mig_state != IB_MIG_MIGRATED &&
521                     attr->path_mig_state != IB_MIG_REARM)
522                         goto inval;
523
524         if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
525                 if (attr->max_dest_rd_atomic > IPATH_MAX_RDMA_ATOMIC)
526                         goto inval;
527
528         switch (new_state) {
529         case IB_QPS_RESET:
530                 ipath_reset_qp(qp);
531                 break;
532
533         case IB_QPS_ERR:
534                 ipath_error_qp(qp, IB_WC_WR_FLUSH_ERR);
535                 break;
536
537         default:
538                 break;
539
540         }
541
542         if (attr_mask & IB_QP_PKEY_INDEX)
543                 qp->s_pkey_index = attr->pkey_index;
544
545         if (attr_mask & IB_QP_DEST_QPN)
546                 qp->remote_qpn = attr->dest_qp_num;
547
548         if (attr_mask & IB_QP_SQ_PSN) {
549                 qp->s_psn = qp->s_next_psn = attr->sq_psn;
550                 qp->s_last_psn = qp->s_next_psn - 1;
551         }
552
553         if (attr_mask & IB_QP_RQ_PSN)
554                 qp->r_psn = attr->rq_psn;
555
556         if (attr_mask & IB_QP_ACCESS_FLAGS)
557                 qp->qp_access_flags = attr->qp_access_flags;
558
559         if (attr_mask & IB_QP_AV)
560                 qp->remote_ah_attr = attr->ah_attr;
561
562         if (attr_mask & IB_QP_PATH_MTU)
563                 qp->path_mtu = attr->path_mtu;
564
565         if (attr_mask & IB_QP_RETRY_CNT)
566                 qp->s_retry = qp->s_retry_cnt = attr->retry_cnt;
567
568         if (attr_mask & IB_QP_RNR_RETRY) {
569                 qp->s_rnr_retry = attr->rnr_retry;
570                 if (qp->s_rnr_retry > 7)
571                         qp->s_rnr_retry = 7;
572                 qp->s_rnr_retry_cnt = qp->s_rnr_retry;
573         }
574
575         if (attr_mask & IB_QP_MIN_RNR_TIMER)
576                 qp->r_min_rnr_timer = attr->min_rnr_timer;
577
578         if (attr_mask & IB_QP_TIMEOUT)
579                 qp->timeout = attr->timeout;
580
581         if (attr_mask & IB_QP_QKEY)
582                 qp->qkey = attr->qkey;
583
584         if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
585                 qp->r_max_rd_atomic = attr->max_dest_rd_atomic;
586
587         if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC)
588                 qp->s_max_rd_atomic = attr->max_rd_atomic;
589
590         qp->state = new_state;
591         spin_unlock_irqrestore(&qp->s_lock, flags);
592
593         ret = 0;
594         goto bail;
595
596 inval:
597         spin_unlock_irqrestore(&qp->s_lock, flags);
598         ret = -EINVAL;
599
600 bail:
601         return ret;
602 }
603
604 int ipath_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
605                    int attr_mask, struct ib_qp_init_attr *init_attr)
606 {
607         struct ipath_qp *qp = to_iqp(ibqp);
608
609         attr->qp_state = qp->state;
610         attr->cur_qp_state = attr->qp_state;
611         attr->path_mtu = qp->path_mtu;
612         attr->path_mig_state = 0;
613         attr->qkey = qp->qkey;
614         attr->rq_psn = qp->r_psn;
615         attr->sq_psn = qp->s_next_psn;
616         attr->dest_qp_num = qp->remote_qpn;
617         attr->qp_access_flags = qp->qp_access_flags;
618         attr->cap.max_send_wr = qp->s_size - 1;
619         attr->cap.max_recv_wr = qp->ibqp.srq ? 0 : qp->r_rq.size - 1;
620         attr->cap.max_send_sge = qp->s_max_sge;
621         attr->cap.max_recv_sge = qp->r_rq.max_sge;
622         attr->cap.max_inline_data = 0;
623         attr->ah_attr = qp->remote_ah_attr;
624         memset(&attr->alt_ah_attr, 0, sizeof(attr->alt_ah_attr));
625         attr->pkey_index = qp->s_pkey_index;
626         attr->alt_pkey_index = 0;
627         attr->en_sqd_async_notify = 0;
628         attr->sq_draining = 0;
629         attr->max_rd_atomic = qp->s_max_rd_atomic;
630         attr->max_dest_rd_atomic = qp->r_max_rd_atomic;
631         attr->min_rnr_timer = qp->r_min_rnr_timer;
632         attr->port_num = 1;
633         attr->timeout = qp->timeout;
634         attr->retry_cnt = qp->s_retry_cnt;
635         attr->rnr_retry = qp->s_rnr_retry;
636         attr->alt_port_num = 0;
637         attr->alt_timeout = 0;
638
639         init_attr->event_handler = qp->ibqp.event_handler;
640         init_attr->qp_context = qp->ibqp.qp_context;
641         init_attr->send_cq = qp->ibqp.send_cq;
642         init_attr->recv_cq = qp->ibqp.recv_cq;
643         init_attr->srq = qp->ibqp.srq;
644         init_attr->cap = attr->cap;
645         if (qp->s_flags & IPATH_S_SIGNAL_REQ_WR)
646                 init_attr->sq_sig_type = IB_SIGNAL_REQ_WR;
647         else
648                 init_attr->sq_sig_type = IB_SIGNAL_ALL_WR;
649         init_attr->qp_type = qp->ibqp.qp_type;
650         init_attr->port_num = 1;
651         return 0;
652 }
653
654 /**
655  * ipath_compute_aeth - compute the AETH (syndrome + MSN)
656  * @qp: the queue pair to compute the AETH for
657  *
658  * Returns the AETH.
659  */
660 __be32 ipath_compute_aeth(struct ipath_qp *qp)
661 {
662         u32 aeth = qp->r_msn & IPATH_MSN_MASK;
663
664         if (qp->ibqp.srq) {
665                 /*
666                  * Shared receive queues don't generate credits.
667                  * Set the credit field to the invalid value.
668                  */
669                 aeth |= IPATH_AETH_CREDIT_INVAL << IPATH_AETH_CREDIT_SHIFT;
670         } else {
671                 u32 min, max, x;
672                 u32 credits;
673                 struct ipath_rwq *wq = qp->r_rq.wq;
674                 u32 head;
675                 u32 tail;
676
677                 /* sanity check pointers before trusting them */
678                 head = wq->head;
679                 if (head >= qp->r_rq.size)
680                         head = 0;
681                 tail = wq->tail;
682                 if (tail >= qp->r_rq.size)
683                         tail = 0;
684                 /*
685                  * Compute the number of credits available (RWQEs).
686                  * XXX Not holding the r_rq.lock here so there is a small
687                  * chance that the pair of reads are not atomic.
688                  */
689                 credits = head - tail;
690                 if ((int)credits < 0)
691                         credits += qp->r_rq.size;
692                 /*
693                  * Binary search the credit table to find the code to
694                  * use.
695                  */
696                 min = 0;
697                 max = 31;
698                 for (;;) {
699                         x = (min + max) / 2;
700                         if (credit_table[x] == credits)
701                                 break;
702                         if (credit_table[x] > credits)
703                                 max = x;
704                         else if (min == x)
705                                 break;
706                         else
707                                 min = x;
708                 }
709                 aeth |= x << IPATH_AETH_CREDIT_SHIFT;
710         }
711         return cpu_to_be32(aeth);
712 }
713
714 /**
715  * ipath_create_qp - create a queue pair for a device
716  * @ibpd: the protection domain who's device we create the queue pair for
717  * @init_attr: the attributes of the queue pair
718  * @udata: unused by InfiniPath
719  *
720  * Returns the queue pair on success, otherwise returns an errno.
721  *
722  * Called by the ib_create_qp() core verbs function.
723  */
724 struct ib_qp *ipath_create_qp(struct ib_pd *ibpd,
725                               struct ib_qp_init_attr *init_attr,
726                               struct ib_udata *udata)
727 {
728         struct ipath_qp *qp;
729         int err;
730         struct ipath_swqe *swq = NULL;
731         struct ipath_ibdev *dev;
732         size_t sz;
733         struct ib_qp *ret;
734
735         if (init_attr->cap.max_send_sge > ib_ipath_max_sges ||
736             init_attr->cap.max_recv_sge > ib_ipath_max_sges ||
737             init_attr->cap.max_send_wr > ib_ipath_max_qp_wrs ||
738             init_attr->cap.max_recv_wr > ib_ipath_max_qp_wrs) {
739                 ret = ERR_PTR(-ENOMEM);
740                 goto bail;
741         }
742
743         if (init_attr->cap.max_send_sge +
744             init_attr->cap.max_recv_sge +
745             init_attr->cap.max_send_wr +
746             init_attr->cap.max_recv_wr == 0) {
747                 ret = ERR_PTR(-EINVAL);
748                 goto bail;
749         }
750
751         switch (init_attr->qp_type) {
752         case IB_QPT_UC:
753         case IB_QPT_RC:
754                 sz = sizeof(struct ipath_sge) *
755                         init_attr->cap.max_send_sge +
756                         sizeof(struct ipath_swqe);
757                 swq = vmalloc((init_attr->cap.max_send_wr + 1) * sz);
758                 if (swq == NULL) {
759                         ret = ERR_PTR(-ENOMEM);
760                         goto bail;
761                 }
762                 /* FALLTHROUGH */
763         case IB_QPT_UD:
764         case IB_QPT_SMI:
765         case IB_QPT_GSI:
766                 sz = sizeof(*qp);
767                 if (init_attr->srq) {
768                         struct ipath_srq *srq = to_isrq(init_attr->srq);
769
770                         sz += sizeof(*qp->r_sg_list) *
771                                 srq->rq.max_sge;
772                 } else
773                         sz += sizeof(*qp->r_sg_list) *
774                                 init_attr->cap.max_recv_sge;
775                 qp = kmalloc(sz, GFP_KERNEL);
776                 if (!qp) {
777                         ret = ERR_PTR(-ENOMEM);
778                         goto bail_swq;
779                 }
780                 if (init_attr->srq) {
781                         sz = 0;
782                         qp->r_rq.size = 0;
783                         qp->r_rq.max_sge = 0;
784                         qp->r_rq.wq = NULL;
785                         init_attr->cap.max_recv_wr = 0;
786                         init_attr->cap.max_recv_sge = 0;
787                 } else {
788                         qp->r_rq.size = init_attr->cap.max_recv_wr + 1;
789                         qp->r_rq.max_sge = init_attr->cap.max_recv_sge;
790                         sz = (sizeof(struct ib_sge) * qp->r_rq.max_sge) +
791                                 sizeof(struct ipath_rwqe);
792                         qp->r_rq.wq = vmalloc_user(sizeof(struct ipath_rwq) +
793                                               qp->r_rq.size * sz);
794                         if (!qp->r_rq.wq) {
795                                 ret = ERR_PTR(-ENOMEM);
796                                 goto bail_qp;
797                         }
798                 }
799
800                 /*
801                  * ib_create_qp() will initialize qp->ibqp
802                  * except for qp->ibqp.qp_num.
803                  */
804                 spin_lock_init(&qp->s_lock);
805                 spin_lock_init(&qp->r_rq.lock);
806                 atomic_set(&qp->refcount, 0);
807                 init_waitqueue_head(&qp->wait);
808                 tasklet_init(&qp->s_task, ipath_do_ruc_send,
809                              (unsigned long)qp);
810                 INIT_LIST_HEAD(&qp->piowait);
811                 INIT_LIST_HEAD(&qp->timerwait);
812                 qp->state = IB_QPS_RESET;
813                 qp->s_wq = swq;
814                 qp->s_size = init_attr->cap.max_send_wr + 1;
815                 qp->s_max_sge = init_attr->cap.max_send_sge;
816                 if (init_attr->sq_sig_type == IB_SIGNAL_REQ_WR)
817                         qp->s_flags = IPATH_S_SIGNAL_REQ_WR;
818                 else
819                         qp->s_flags = 0;
820                 dev = to_idev(ibpd->device);
821                 err = ipath_alloc_qpn(&dev->qp_table, qp,
822                                       init_attr->qp_type);
823                 if (err) {
824                         ret = ERR_PTR(err);
825                         goto bail_rwq;
826                 }
827                 qp->ip = NULL;
828                 ipath_reset_qp(qp);
829                 break;
830
831         default:
832                 /* Don't support raw QPs */
833                 ret = ERR_PTR(-ENOSYS);
834                 goto bail;
835         }
836
837         init_attr->cap.max_inline_data = 0;
838
839         /*
840          * Return the address of the RWQ as the offset to mmap.
841          * See ipath_mmap() for details.
842          */
843         if (udata && udata->outlen >= sizeof(__u64)) {
844                 int err;
845
846                 if (!qp->r_rq.wq) {
847                         __u64 offset = 0;
848
849                         err = ib_copy_to_udata(udata, &offset,
850                                                sizeof(offset));
851                         if (err) {
852                                 ret = ERR_PTR(err);
853                                 goto bail_rwq;
854                         }
855                 } else {
856                         u32 s = sizeof(struct ipath_rwq) +
857                                 qp->r_rq.size * sz;
858
859                         qp->ip =
860                             ipath_create_mmap_info(dev, s,
861                                                    ibpd->uobject->context,
862                                                    qp->r_rq.wq);
863                         if (!qp->ip) {
864                                 ret = ERR_PTR(-ENOMEM);
865                                 goto bail_rwq;
866                         }
867
868                         err = ib_copy_to_udata(udata, &(qp->ip->offset),
869                                                sizeof(qp->ip->offset));
870                         if (err) {
871                                 ret = ERR_PTR(err);
872                                 goto bail_ip;
873                         }
874                 }
875         }
876
877         spin_lock(&dev->n_qps_lock);
878         if (dev->n_qps_allocated == ib_ipath_max_qps) {
879                 spin_unlock(&dev->n_qps_lock);
880                 ret = ERR_PTR(-ENOMEM);
881                 goto bail_ip;
882         }
883
884         dev->n_qps_allocated++;
885         spin_unlock(&dev->n_qps_lock);
886
887         if (qp->ip) {
888                 spin_lock_irq(&dev->pending_lock);
889                 list_add(&qp->ip->pending_mmaps, &dev->pending_mmaps);
890                 spin_unlock_irq(&dev->pending_lock);
891         }
892
893         ret = &qp->ibqp;
894         goto bail;
895
896 bail_ip:
897         kfree(qp->ip);
898 bail_rwq:
899         vfree(qp->r_rq.wq);
900 bail_qp:
901         kfree(qp);
902 bail_swq:
903         vfree(swq);
904 bail:
905         return ret;
906 }
907
908 /**
909  * ipath_destroy_qp - destroy a queue pair
910  * @ibqp: the queue pair to destroy
911  *
912  * Returns 0 on success.
913  *
914  * Note that this can be called while the QP is actively sending or
915  * receiving!
916  */
917 int ipath_destroy_qp(struct ib_qp *ibqp)
918 {
919         struct ipath_qp *qp = to_iqp(ibqp);
920         struct ipath_ibdev *dev = to_idev(ibqp->device);
921         unsigned long flags;
922
923         spin_lock_irqsave(&qp->s_lock, flags);
924         qp->state = IB_QPS_ERR;
925         spin_unlock_irqrestore(&qp->s_lock, flags);
926         spin_lock(&dev->n_qps_lock);
927         dev->n_qps_allocated--;
928         spin_unlock(&dev->n_qps_lock);
929
930         /* Stop the sending tasklet. */
931         tasklet_kill(&qp->s_task);
932
933         /* Make sure the QP isn't on the timeout list. */
934         spin_lock_irqsave(&dev->pending_lock, flags);
935         if (!list_empty(&qp->timerwait))
936                 list_del_init(&qp->timerwait);
937         if (!list_empty(&qp->piowait))
938                 list_del_init(&qp->piowait);
939         spin_unlock_irqrestore(&dev->pending_lock, flags);
940
941         /*
942          * Make sure that the QP is not in the QPN table so receive
943          * interrupts will discard packets for this QP.  XXX Also remove QP
944          * from multicast table.
945          */
946         if (atomic_read(&qp->refcount) != 0)
947                 ipath_free_qp(&dev->qp_table, qp);
948
949         if (qp->ip)
950                 kref_put(&qp->ip->ref, ipath_release_mmap_info);
951         else
952                 vfree(qp->r_rq.wq);
953         vfree(qp->s_wq);
954         kfree(qp);
955         return 0;
956 }
957
958 /**
959  * ipath_init_qp_table - initialize the QP table for a device
960  * @idev: the device who's QP table we're initializing
961  * @size: the size of the QP table
962  *
963  * Returns 0 on success, otherwise returns an errno.
964  */
965 int ipath_init_qp_table(struct ipath_ibdev *idev, int size)
966 {
967         int i;
968         int ret;
969
970         idev->qp_table.last = 1;        /* QPN 0 and 1 are special. */
971         idev->qp_table.max = size;
972         idev->qp_table.nmaps = 1;
973         idev->qp_table.table = kzalloc(size * sizeof(*idev->qp_table.table),
974                                        GFP_KERNEL);
975         if (idev->qp_table.table == NULL) {
976                 ret = -ENOMEM;
977                 goto bail;
978         }
979
980         for (i = 0; i < ARRAY_SIZE(idev->qp_table.map); i++) {
981                 atomic_set(&idev->qp_table.map[i].n_free, BITS_PER_PAGE);
982                 idev->qp_table.map[i].page = NULL;
983         }
984
985         ret = 0;
986
987 bail:
988         return ret;
989 }
990
991 /**
992  * ipath_sqerror_qp - put a QP's send queue into an error state
993  * @qp: QP who's send queue will be put into an error state
994  * @wc: the WC responsible for putting the QP in this state
995  *
996  * Flushes the send work queue.
997  * The QP s_lock should be held and interrupts disabled.
998  */
999
1000 void ipath_sqerror_qp(struct ipath_qp *qp, struct ib_wc *wc)
1001 {
1002         struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
1003         struct ipath_swqe *wqe = get_swqe_ptr(qp, qp->s_last);
1004
1005         ipath_dbg("Send queue error on QP%d/%d: err: %d\n",
1006                   qp->ibqp.qp_num, qp->remote_qpn, wc->status);
1007
1008         spin_lock(&dev->pending_lock);
1009         /* XXX What if its already removed by the timeout code? */
1010         if (!list_empty(&qp->timerwait))
1011                 list_del_init(&qp->timerwait);
1012         if (!list_empty(&qp->piowait))
1013                 list_del_init(&qp->piowait);
1014         spin_unlock(&dev->pending_lock);
1015
1016         ipath_cq_enter(to_icq(qp->ibqp.send_cq), wc, 1);
1017         if (++qp->s_last >= qp->s_size)
1018                 qp->s_last = 0;
1019
1020         wc->status = IB_WC_WR_FLUSH_ERR;
1021
1022         while (qp->s_last != qp->s_head) {
1023                 wqe = get_swqe_ptr(qp, qp->s_last);
1024                 wc->wr_id = wqe->wr.wr_id;
1025                 wc->opcode = ib_ipath_wc_opcode[wqe->wr.opcode];
1026                 ipath_cq_enter(to_icq(qp->ibqp.send_cq), wc, 1);
1027                 if (++qp->s_last >= qp->s_size)
1028                         qp->s_last = 0;
1029         }
1030         qp->s_cur = qp->s_tail = qp->s_head;
1031         qp->state = IB_QPS_SQE;
1032 }
1033
1034 /**
1035  * ipath_get_credit - flush the send work queue of a QP
1036  * @qp: the qp who's send work queue to flush
1037  * @aeth: the Acknowledge Extended Transport Header
1038  *
1039  * The QP s_lock should be held.
1040  */
1041 void ipath_get_credit(struct ipath_qp *qp, u32 aeth)
1042 {
1043         u32 credit = (aeth >> IPATH_AETH_CREDIT_SHIFT) & IPATH_AETH_CREDIT_MASK;
1044
1045         /*
1046          * If the credit is invalid, we can send
1047          * as many packets as we like.  Otherwise, we have to
1048          * honor the credit field.
1049          */
1050         if (credit == IPATH_AETH_CREDIT_INVAL)
1051                 qp->s_lsn = (u32) -1;
1052         else if (qp->s_lsn != (u32) -1) {
1053                 /* Compute new LSN (i.e., MSN + credit) */
1054                 credit = (aeth + credit_table[credit]) & IPATH_MSN_MASK;
1055                 if (ipath_cmp24(credit, qp->s_lsn) > 0)
1056                         qp->s_lsn = credit;
1057         }
1058
1059         /* Restart sending if it was blocked due to lack of credits. */
1060         if (qp->s_cur != qp->s_head &&
1061             (qp->s_lsn == (u32) -1 ||
1062              ipath_cmp24(get_swqe_ptr(qp, qp->s_cur)->ssn,
1063                          qp->s_lsn + 1) <= 0))
1064                 tasklet_hi_schedule(&qp->s_task);
1065 }