IB/ipath: Limit # of packets sent without an ACK received
[linux-3.10.git] / 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->s_rnr_timeout = 0;
339         qp->s_head = 0;
340         qp->s_tail = 0;
341         qp->s_cur = 0;
342         qp->s_last = 0;
343         qp->s_ssn = 1;
344         qp->s_lsn = 0;
345         qp->s_wait_credit = 0;
346         if (qp->r_rq.wq) {
347                 qp->r_rq.wq->head = 0;
348                 qp->r_rq.wq->tail = 0;
349         }
350         qp->r_reuse_sge = 0;
351 }
352
353 /**
354  * ipath_error_qp - put a QP into an error state
355  * @qp: the QP to put into an error state
356  *
357  * Flushes both send and receive work queues.
358  * QP s_lock should be held and interrupts disabled.
359  */
360
361 void ipath_error_qp(struct ipath_qp *qp)
362 {
363         struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
364         struct ib_wc wc;
365
366         ipath_dbg(KERN_INFO "QP%d/%d in error state\n",
367                   qp->ibqp.qp_num, qp->remote_qpn);
368
369         spin_lock(&dev->pending_lock);
370         /* XXX What if its already removed by the timeout code? */
371         if (!list_empty(&qp->timerwait))
372                 list_del_init(&qp->timerwait);
373         if (!list_empty(&qp->piowait))
374                 list_del_init(&qp->piowait);
375         spin_unlock(&dev->pending_lock);
376
377         wc.status = IB_WC_WR_FLUSH_ERR;
378         wc.vendor_err = 0;
379         wc.byte_len = 0;
380         wc.imm_data = 0;
381         wc.qp_num = qp->ibqp.qp_num;
382         wc.src_qp = 0;
383         wc.wc_flags = 0;
384         wc.pkey_index = 0;
385         wc.slid = 0;
386         wc.sl = 0;
387         wc.dlid_path_bits = 0;
388         wc.port_num = 0;
389
390         while (qp->s_last != qp->s_head) {
391                 struct ipath_swqe *wqe = get_swqe_ptr(qp, qp->s_last);
392
393                 wc.wr_id = wqe->wr.wr_id;
394                 wc.opcode = ib_ipath_wc_opcode[wqe->wr.opcode];
395                 if (++qp->s_last >= qp->s_size)
396                         qp->s_last = 0;
397                 ipath_cq_enter(to_icq(qp->ibqp.send_cq), &wc, 1);
398         }
399         qp->s_cur = qp->s_tail = qp->s_head;
400         qp->s_hdrwords = 0;
401         qp->s_ack_state = IB_OPCODE_RC_ACKNOWLEDGE;
402
403         if (qp->r_rq.wq) {
404                 struct ipath_rwq *wq;
405                 u32 head;
406                 u32 tail;
407
408                 spin_lock(&qp->r_rq.lock);
409
410                 /* sanity check pointers before trusting them */
411                 wq = qp->r_rq.wq;
412                 head = wq->head;
413                 if (head >= qp->r_rq.size)
414                         head = 0;
415                 tail = wq->tail;
416                 if (tail >= qp->r_rq.size)
417                         tail = 0;
418                 wc.opcode = IB_WC_RECV;
419                 while (tail != head) {
420                         wc.wr_id = get_rwqe_ptr(&qp->r_rq, tail)->wr_id;
421                         if (++tail >= qp->r_rq.size)
422                                 tail = 0;
423                         ipath_cq_enter(to_icq(qp->ibqp.recv_cq), &wc, 1);
424                 }
425                 wq->tail = tail;
426
427                 spin_unlock(&qp->r_rq.lock);
428         }
429 }
430
431 /**
432  * ipath_modify_qp - modify the attributes of a queue pair
433  * @ibqp: the queue pair who's attributes we're modifying
434  * @attr: the new attributes
435  * @attr_mask: the mask of attributes to modify
436  * @udata: user data for ipathverbs.so
437  *
438  * Returns 0 on success, otherwise returns an errno.
439  */
440 int ipath_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
441                     int attr_mask, struct ib_udata *udata)
442 {
443         struct ipath_ibdev *dev = to_idev(ibqp->device);
444         struct ipath_qp *qp = to_iqp(ibqp);
445         enum ib_qp_state cur_state, new_state;
446         unsigned long flags;
447         int ret;
448
449         spin_lock_irqsave(&qp->s_lock, flags);
450
451         cur_state = attr_mask & IB_QP_CUR_STATE ?
452                 attr->cur_qp_state : qp->state;
453         new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
454
455         if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type,
456                                 attr_mask))
457                 goto inval;
458
459         if (attr_mask & IB_QP_AV) {
460                 if (attr->ah_attr.dlid == 0 ||
461                     attr->ah_attr.dlid >= IPATH_MULTICAST_LID_BASE)
462                         goto inval;
463
464                 if ((attr->ah_attr.ah_flags & IB_AH_GRH) &&
465                     (attr->ah_attr.grh.sgid_index > 1))
466                         goto inval;
467         }
468
469         if (attr_mask & IB_QP_PKEY_INDEX)
470                 if (attr->pkey_index >= ipath_get_npkeys(dev->dd))
471                         goto inval;
472
473         if (attr_mask & IB_QP_MIN_RNR_TIMER)
474                 if (attr->min_rnr_timer > 31)
475                         goto inval;
476
477         if (attr_mask & IB_QP_PORT)
478                 if (attr->port_num == 0 ||
479                     attr->port_num > ibqp->device->phys_port_cnt)
480                         goto inval;
481
482         if (attr_mask & IB_QP_PATH_MTU)
483                 if (attr->path_mtu > IB_MTU_4096)
484                         goto inval;
485
486         if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
487                 if (attr->max_dest_rd_atomic > 1)
488                         goto inval;
489
490         if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC)
491                 if (attr->max_rd_atomic > 1)
492                         goto inval;
493
494         if (attr_mask & IB_QP_PATH_MIG_STATE)
495                 if (attr->path_mig_state != IB_MIG_MIGRATED &&
496                     attr->path_mig_state != IB_MIG_REARM)
497                         goto inval;
498
499         switch (new_state) {
500         case IB_QPS_RESET:
501                 ipath_reset_qp(qp);
502                 break;
503
504         case IB_QPS_ERR:
505                 ipath_error_qp(qp);
506                 break;
507
508         default:
509                 break;
510
511         }
512
513         if (attr_mask & IB_QP_PKEY_INDEX)
514                 qp->s_pkey_index = attr->pkey_index;
515
516         if (attr_mask & IB_QP_DEST_QPN)
517                 qp->remote_qpn = attr->dest_qp_num;
518
519         if (attr_mask & IB_QP_SQ_PSN) {
520                 qp->s_psn = qp->s_next_psn = attr->sq_psn;
521                 qp->s_last_psn = qp->s_next_psn - 1;
522         }
523
524         if (attr_mask & IB_QP_RQ_PSN)
525                 qp->r_psn = attr->rq_psn;
526
527         if (attr_mask & IB_QP_ACCESS_FLAGS)
528                 qp->qp_access_flags = attr->qp_access_flags;
529
530         if (attr_mask & IB_QP_AV)
531                 qp->remote_ah_attr = attr->ah_attr;
532
533         if (attr_mask & IB_QP_PATH_MTU)
534                 qp->path_mtu = attr->path_mtu;
535
536         if (attr_mask & IB_QP_RETRY_CNT)
537                 qp->s_retry = qp->s_retry_cnt = attr->retry_cnt;
538
539         if (attr_mask & IB_QP_RNR_RETRY) {
540                 qp->s_rnr_retry = attr->rnr_retry;
541                 if (qp->s_rnr_retry > 7)
542                         qp->s_rnr_retry = 7;
543                 qp->s_rnr_retry_cnt = qp->s_rnr_retry;
544         }
545
546         if (attr_mask & IB_QP_MIN_RNR_TIMER)
547                 qp->r_min_rnr_timer = attr->min_rnr_timer;
548
549         if (attr_mask & IB_QP_TIMEOUT)
550                 qp->timeout = attr->timeout;
551
552         if (attr_mask & IB_QP_QKEY)
553                 qp->qkey = attr->qkey;
554
555         qp->state = new_state;
556         spin_unlock_irqrestore(&qp->s_lock, flags);
557
558         ret = 0;
559         goto bail;
560
561 inval:
562         spin_unlock_irqrestore(&qp->s_lock, flags);
563         ret = -EINVAL;
564
565 bail:
566         return ret;
567 }
568
569 int ipath_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
570                    int attr_mask, struct ib_qp_init_attr *init_attr)
571 {
572         struct ipath_qp *qp = to_iqp(ibqp);
573
574         attr->qp_state = qp->state;
575         attr->cur_qp_state = attr->qp_state;
576         attr->path_mtu = qp->path_mtu;
577         attr->path_mig_state = 0;
578         attr->qkey = qp->qkey;
579         attr->rq_psn = qp->r_psn;
580         attr->sq_psn = qp->s_next_psn;
581         attr->dest_qp_num = qp->remote_qpn;
582         attr->qp_access_flags = qp->qp_access_flags;
583         attr->cap.max_send_wr = qp->s_size - 1;
584         attr->cap.max_recv_wr = qp->ibqp.srq ? 0 : qp->r_rq.size - 1;
585         attr->cap.max_send_sge = qp->s_max_sge;
586         attr->cap.max_recv_sge = qp->r_rq.max_sge;
587         attr->cap.max_inline_data = 0;
588         attr->ah_attr = qp->remote_ah_attr;
589         memset(&attr->alt_ah_attr, 0, sizeof(attr->alt_ah_attr));
590         attr->pkey_index = qp->s_pkey_index;
591         attr->alt_pkey_index = 0;
592         attr->en_sqd_async_notify = 0;
593         attr->sq_draining = 0;
594         attr->max_rd_atomic = 1;
595         attr->max_dest_rd_atomic = 1;
596         attr->min_rnr_timer = qp->r_min_rnr_timer;
597         attr->port_num = 1;
598         attr->timeout = qp->timeout;
599         attr->retry_cnt = qp->s_retry_cnt;
600         attr->rnr_retry = qp->s_rnr_retry;
601         attr->alt_port_num = 0;
602         attr->alt_timeout = 0;
603
604         init_attr->event_handler = qp->ibqp.event_handler;
605         init_attr->qp_context = qp->ibqp.qp_context;
606         init_attr->send_cq = qp->ibqp.send_cq;
607         init_attr->recv_cq = qp->ibqp.recv_cq;
608         init_attr->srq = qp->ibqp.srq;
609         init_attr->cap = attr->cap;
610         if (qp->s_flags & (1 << IPATH_S_SIGNAL_REQ_WR))
611                 init_attr->sq_sig_type = IB_SIGNAL_REQ_WR;
612         else
613                 init_attr->sq_sig_type = IB_SIGNAL_ALL_WR;
614         init_attr->qp_type = qp->ibqp.qp_type;
615         init_attr->port_num = 1;
616         return 0;
617 }
618
619 /**
620  * ipath_compute_aeth - compute the AETH (syndrome + MSN)
621  * @qp: the queue pair to compute the AETH for
622  *
623  * Returns the AETH.
624  */
625 __be32 ipath_compute_aeth(struct ipath_qp *qp)
626 {
627         u32 aeth = qp->r_msn & IPATH_MSN_MASK;
628
629         if (qp->ibqp.srq) {
630                 /*
631                  * Shared receive queues don't generate credits.
632                  * Set the credit field to the invalid value.
633                  */
634                 aeth |= IPATH_AETH_CREDIT_INVAL << IPATH_AETH_CREDIT_SHIFT;
635         } else {
636                 u32 min, max, x;
637                 u32 credits;
638                 struct ipath_rwq *wq = qp->r_rq.wq;
639                 u32 head;
640                 u32 tail;
641
642                 /* sanity check pointers before trusting them */
643                 head = wq->head;
644                 if (head >= qp->r_rq.size)
645                         head = 0;
646                 tail = wq->tail;
647                 if (tail >= qp->r_rq.size)
648                         tail = 0;
649                 /*
650                  * Compute the number of credits available (RWQEs).
651                  * XXX Not holding the r_rq.lock here so there is a small
652                  * chance that the pair of reads are not atomic.
653                  */
654                 credits = head - tail;
655                 if ((int)credits < 0)
656                         credits += qp->r_rq.size;
657                 /*
658                  * Binary search the credit table to find the code to
659                  * use.
660                  */
661                 min = 0;
662                 max = 31;
663                 for (;;) {
664                         x = (min + max) / 2;
665                         if (credit_table[x] == credits)
666                                 break;
667                         if (credit_table[x] > credits)
668                                 max = x;
669                         else if (min == x)
670                                 break;
671                         else
672                                 min = x;
673                 }
674                 aeth |= x << IPATH_AETH_CREDIT_SHIFT;
675         }
676         return cpu_to_be32(aeth);
677 }
678
679 /**
680  * ipath_create_qp - create a queue pair for a device
681  * @ibpd: the protection domain who's device we create the queue pair for
682  * @init_attr: the attributes of the queue pair
683  * @udata: unused by InfiniPath
684  *
685  * Returns the queue pair on success, otherwise returns an errno.
686  *
687  * Called by the ib_create_qp() core verbs function.
688  */
689 struct ib_qp *ipath_create_qp(struct ib_pd *ibpd,
690                               struct ib_qp_init_attr *init_attr,
691                               struct ib_udata *udata)
692 {
693         struct ipath_qp *qp;
694         int err;
695         struct ipath_swqe *swq = NULL;
696         struct ipath_ibdev *dev;
697         size_t sz;
698         struct ib_qp *ret;
699
700         if (init_attr->cap.max_send_sge > ib_ipath_max_sges ||
701             init_attr->cap.max_recv_sge > ib_ipath_max_sges ||
702             init_attr->cap.max_send_wr > ib_ipath_max_qp_wrs ||
703             init_attr->cap.max_recv_wr > ib_ipath_max_qp_wrs) {
704                 ret = ERR_PTR(-ENOMEM);
705                 goto bail;
706         }
707
708         if (init_attr->cap.max_send_sge +
709             init_attr->cap.max_recv_sge +
710             init_attr->cap.max_send_wr +
711             init_attr->cap.max_recv_wr == 0) {
712                 ret = ERR_PTR(-EINVAL);
713                 goto bail;
714         }
715
716         switch (init_attr->qp_type) {
717         case IB_QPT_UC:
718         case IB_QPT_RC:
719                 sz = sizeof(struct ipath_sge) *
720                         init_attr->cap.max_send_sge +
721                         sizeof(struct ipath_swqe);
722                 swq = vmalloc((init_attr->cap.max_send_wr + 1) * sz);
723                 if (swq == NULL) {
724                         ret = ERR_PTR(-ENOMEM);
725                         goto bail;
726                 }
727                 /* FALLTHROUGH */
728         case IB_QPT_UD:
729         case IB_QPT_SMI:
730         case IB_QPT_GSI:
731                 sz = sizeof(*qp);
732                 if (init_attr->srq) {
733                         struct ipath_srq *srq = to_isrq(init_attr->srq);
734
735                         sz += sizeof(*qp->r_sg_list) *
736                                 srq->rq.max_sge;
737                 } else
738                         sz += sizeof(*qp->r_sg_list) *
739                                 init_attr->cap.max_recv_sge;
740                 qp = kmalloc(sz, GFP_KERNEL);
741                 if (!qp) {
742                         ret = ERR_PTR(-ENOMEM);
743                         goto bail_swq;
744                 }
745                 if (init_attr->srq) {
746                         sz = 0;
747                         qp->r_rq.size = 0;
748                         qp->r_rq.max_sge = 0;
749                         qp->r_rq.wq = NULL;
750                         init_attr->cap.max_recv_wr = 0;
751                         init_attr->cap.max_recv_sge = 0;
752                 } else {
753                         qp->r_rq.size = init_attr->cap.max_recv_wr + 1;
754                         qp->r_rq.max_sge = init_attr->cap.max_recv_sge;
755                         sz = (sizeof(struct ib_sge) * qp->r_rq.max_sge) +
756                                 sizeof(struct ipath_rwqe);
757                         qp->r_rq.wq = vmalloc_user(sizeof(struct ipath_rwq) +
758                                               qp->r_rq.size * sz);
759                         if (!qp->r_rq.wq) {
760                                 ret = ERR_PTR(-ENOMEM);
761                                 goto bail_qp;
762                         }
763                 }
764
765                 /*
766                  * ib_create_qp() will initialize qp->ibqp
767                  * except for qp->ibqp.qp_num.
768                  */
769                 spin_lock_init(&qp->s_lock);
770                 spin_lock_init(&qp->r_rq.lock);
771                 atomic_set(&qp->refcount, 0);
772                 init_waitqueue_head(&qp->wait);
773                 tasklet_init(&qp->s_task, ipath_do_ruc_send,
774                              (unsigned long)qp);
775                 INIT_LIST_HEAD(&qp->piowait);
776                 INIT_LIST_HEAD(&qp->timerwait);
777                 qp->state = IB_QPS_RESET;
778                 qp->s_wq = swq;
779                 qp->s_size = init_attr->cap.max_send_wr + 1;
780                 qp->s_max_sge = init_attr->cap.max_send_sge;
781                 if (init_attr->sq_sig_type == IB_SIGNAL_REQ_WR)
782                         qp->s_flags = 1 << IPATH_S_SIGNAL_REQ_WR;
783                 else
784                         qp->s_flags = 0;
785                 dev = to_idev(ibpd->device);
786                 err = ipath_alloc_qpn(&dev->qp_table, qp,
787                                       init_attr->qp_type);
788                 if (err) {
789                         ret = ERR_PTR(err);
790                         goto bail_rwq;
791                 }
792                 qp->ip = NULL;
793                 ipath_reset_qp(qp);
794                 break;
795
796         default:
797                 /* Don't support raw QPs */
798                 ret = ERR_PTR(-ENOSYS);
799                 goto bail;
800         }
801
802         init_attr->cap.max_inline_data = 0;
803
804         /*
805          * Return the address of the RWQ as the offset to mmap.
806          * See ipath_mmap() for details.
807          */
808         if (udata && udata->outlen >= sizeof(__u64)) {
809                 struct ipath_mmap_info *ip;
810                 __u64 offset = (__u64) qp->r_rq.wq;
811                 int err;
812
813                 err = ib_copy_to_udata(udata, &offset, sizeof(offset));
814                 if (err) {
815                         ret = ERR_PTR(err);
816                         goto bail_rwq;
817                 }
818
819                 if (qp->r_rq.wq) {
820                         /* Allocate info for ipath_mmap(). */
821                         ip = kmalloc(sizeof(*ip), GFP_KERNEL);
822                         if (!ip) {
823                                 ret = ERR_PTR(-ENOMEM);
824                                 goto bail_rwq;
825                         }
826                         qp->ip = ip;
827                         ip->context = ibpd->uobject->context;
828                         ip->obj = qp->r_rq.wq;
829                         kref_init(&ip->ref);
830                         ip->mmap_cnt = 0;
831                         ip->size = PAGE_ALIGN(sizeof(struct ipath_rwq) +
832                                               qp->r_rq.size * sz);
833                         spin_lock_irq(&dev->pending_lock);
834                         ip->next = dev->pending_mmaps;
835                         dev->pending_mmaps = ip;
836                         spin_unlock_irq(&dev->pending_lock);
837                 }
838         }
839
840         spin_lock(&dev->n_qps_lock);
841         if (dev->n_qps_allocated == ib_ipath_max_qps) {
842                 spin_unlock(&dev->n_qps_lock);
843                 ret = ERR_PTR(-ENOMEM);
844                 goto bail_ip;
845         }
846
847         dev->n_qps_allocated++;
848         spin_unlock(&dev->n_qps_lock);
849
850         ret = &qp->ibqp;
851         goto bail;
852
853 bail_ip:
854         kfree(qp->ip);
855 bail_rwq:
856         vfree(qp->r_rq.wq);
857 bail_qp:
858         kfree(qp);
859 bail_swq:
860         vfree(swq);
861 bail:
862         return ret;
863 }
864
865 /**
866  * ipath_destroy_qp - destroy a queue pair
867  * @ibqp: the queue pair to destroy
868  *
869  * Returns 0 on success.
870  *
871  * Note that this can be called while the QP is actively sending or
872  * receiving!
873  */
874 int ipath_destroy_qp(struct ib_qp *ibqp)
875 {
876         struct ipath_qp *qp = to_iqp(ibqp);
877         struct ipath_ibdev *dev = to_idev(ibqp->device);
878         unsigned long flags;
879
880         spin_lock_irqsave(&qp->s_lock, flags);
881         qp->state = IB_QPS_ERR;
882         spin_unlock_irqrestore(&qp->s_lock, flags);
883         spin_lock(&dev->n_qps_lock);
884         dev->n_qps_allocated--;
885         spin_unlock(&dev->n_qps_lock);
886
887         /* Stop the sending tasklet. */
888         tasklet_kill(&qp->s_task);
889
890         /* Make sure the QP isn't on the timeout list. */
891         spin_lock_irqsave(&dev->pending_lock, flags);
892         if (!list_empty(&qp->timerwait))
893                 list_del_init(&qp->timerwait);
894         if (!list_empty(&qp->piowait))
895                 list_del_init(&qp->piowait);
896         spin_unlock_irqrestore(&dev->pending_lock, flags);
897
898         /*
899          * Make sure that the QP is not in the QPN table so receive
900          * interrupts will discard packets for this QP.  XXX Also remove QP
901          * from multicast table.
902          */
903         if (atomic_read(&qp->refcount) != 0)
904                 ipath_free_qp(&dev->qp_table, qp);
905
906         if (qp->ip)
907                 kref_put(&qp->ip->ref, ipath_release_mmap_info);
908         else
909                 vfree(qp->r_rq.wq);
910         vfree(qp->s_wq);
911         kfree(qp);
912         return 0;
913 }
914
915 /**
916  * ipath_init_qp_table - initialize the QP table for a device
917  * @idev: the device who's QP table we're initializing
918  * @size: the size of the QP table
919  *
920  * Returns 0 on success, otherwise returns an errno.
921  */
922 int ipath_init_qp_table(struct ipath_ibdev *idev, int size)
923 {
924         int i;
925         int ret;
926
927         idev->qp_table.last = 1;        /* QPN 0 and 1 are special. */
928         idev->qp_table.max = size;
929         idev->qp_table.nmaps = 1;
930         idev->qp_table.table = kzalloc(size * sizeof(*idev->qp_table.table),
931                                        GFP_KERNEL);
932         if (idev->qp_table.table == NULL) {
933                 ret = -ENOMEM;
934                 goto bail;
935         }
936
937         for (i = 0; i < ARRAY_SIZE(idev->qp_table.map); i++) {
938                 atomic_set(&idev->qp_table.map[i].n_free, BITS_PER_PAGE);
939                 idev->qp_table.map[i].page = NULL;
940         }
941
942         ret = 0;
943
944 bail:
945         return ret;
946 }
947
948 /**
949  * ipath_sqerror_qp - put a QP's send queue into an error state
950  * @qp: QP who's send queue will be put into an error state
951  * @wc: the WC responsible for putting the QP in this state
952  *
953  * Flushes the send work queue.
954  * The QP s_lock should be held.
955  */
956
957 void ipath_sqerror_qp(struct ipath_qp *qp, struct ib_wc *wc)
958 {
959         struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
960         struct ipath_swqe *wqe = get_swqe_ptr(qp, qp->s_last);
961
962         ipath_dbg(KERN_INFO "Send queue error on QP%d/%d: err: %d\n",
963                   qp->ibqp.qp_num, qp->remote_qpn, wc->status);
964
965         spin_lock(&dev->pending_lock);
966         /* XXX What if its already removed by the timeout code? */
967         if (!list_empty(&qp->timerwait))
968                 list_del_init(&qp->timerwait);
969         if (!list_empty(&qp->piowait))
970                 list_del_init(&qp->piowait);
971         spin_unlock(&dev->pending_lock);
972
973         ipath_cq_enter(to_icq(qp->ibqp.send_cq), wc, 1);
974         if (++qp->s_last >= qp->s_size)
975                 qp->s_last = 0;
976
977         wc->status = IB_WC_WR_FLUSH_ERR;
978
979         while (qp->s_last != qp->s_head) {
980                 wc->wr_id = wqe->wr.wr_id;
981                 wc->opcode = ib_ipath_wc_opcode[wqe->wr.opcode];
982                 ipath_cq_enter(to_icq(qp->ibqp.send_cq), wc, 1);
983                 if (++qp->s_last >= qp->s_size)
984                         qp->s_last = 0;
985                 wqe = get_swqe_ptr(qp, qp->s_last);
986         }
987         qp->s_cur = qp->s_tail = qp->s_head;
988         qp->state = IB_QPS_SQE;
989 }
990
991 /**
992  * ipath_get_credit - flush the send work queue of a QP
993  * @qp: the qp who's send work queue to flush
994  * @aeth: the Acknowledge Extended Transport Header
995  *
996  * The QP s_lock should be held.
997  */
998 void ipath_get_credit(struct ipath_qp *qp, u32 aeth)
999 {
1000         u32 credit = (aeth >> IPATH_AETH_CREDIT_SHIFT) & IPATH_AETH_CREDIT_MASK;
1001
1002         /*
1003          * If the credit is invalid, we can send
1004          * as many packets as we like.  Otherwise, we have to
1005          * honor the credit field.
1006          */
1007         if (credit == IPATH_AETH_CREDIT_INVAL)
1008                 qp->s_lsn = (u32) -1;
1009         else if (qp->s_lsn != (u32) -1) {
1010                 /* Compute new LSN (i.e., MSN + credit) */
1011                 credit = (aeth + credit_table[credit]) & IPATH_MSN_MASK;
1012                 if (ipath_cmp24(credit, qp->s_lsn) > 0)
1013                         qp->s_lsn = credit;
1014         }
1015
1016         /* Restart sending if it was blocked due to lack of credits. */
1017         if (qp->s_cur != qp->s_head &&
1018             (qp->s_lsn == (u32) -1 ||
1019              ipath_cmp24(get_swqe_ptr(qp, qp->s_cur)->ssn,
1020                          qp->s_lsn + 1) <= 0))
1021                 tasklet_hi_schedule(&qp->s_task);
1022 }