IB/ipath: Fix QP error completion queue entries
[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("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         qp->s_busy = 0;
324         qp->s_flags &= ~IPATH_S_SIGNAL_REQ_WR;
325         qp->s_hdrwords = 0;
326         qp->s_psn = 0;
327         qp->r_psn = 0;
328         qp->r_msn = 0;
329         if (qp->ibqp.qp_type == IB_QPT_RC) {
330                 qp->s_state = IB_OPCODE_RC_SEND_LAST;
331                 qp->r_state = IB_OPCODE_RC_SEND_LAST;
332         } else {
333                 qp->s_state = IB_OPCODE_UC_SEND_LAST;
334                 qp->r_state = IB_OPCODE_UC_SEND_LAST;
335         }
336         qp->s_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         memset(qp->s_ack_queue, 0, sizeof(qp->s_ack_queue));
348         qp->r_head_ack_queue = 0;
349         qp->s_tail_ack_queue = 0;
350         qp->s_num_rd_atomic = 0;
351         if (qp->r_rq.wq) {
352                 qp->r_rq.wq->head = 0;
353                 qp->r_rq.wq->tail = 0;
354         }
355         qp->r_reuse_sge = 0;
356 }
357
358 /**
359  * ipath_error_qp - put a QP into an error state
360  * @qp: the QP to put into an error state
361  * @err: the receive completion error to signal if a RWQE is active
362  *
363  * Flushes both send and receive work queues.
364  * The QP s_lock should be held and interrupts disabled.
365  */
366
367 void ipath_error_qp(struct ipath_qp *qp, enum ib_wc_status err)
368 {
369         struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
370         struct ib_wc wc;
371
372         ipath_dbg("QP%d/%d in error state\n",
373                   qp->ibqp.qp_num, qp->remote_qpn);
374
375         spin_lock(&dev->pending_lock);
376         /* XXX What if its already removed by the timeout code? */
377         if (!list_empty(&qp->timerwait))
378                 list_del_init(&qp->timerwait);
379         if (!list_empty(&qp->piowait))
380                 list_del_init(&qp->piowait);
381         spin_unlock(&dev->pending_lock);
382
383         wc.vendor_err = 0;
384         wc.byte_len = 0;
385         wc.imm_data = 0;
386         wc.qp = &qp->ibqp;
387         wc.src_qp = 0;
388         wc.wc_flags = 0;
389         wc.pkey_index = 0;
390         wc.slid = 0;
391         wc.sl = 0;
392         wc.dlid_path_bits = 0;
393         wc.port_num = 0;
394         if (qp->r_wrid_valid) {
395                 qp->r_wrid_valid = 0;
396                 wc.wr_id = qp->r_wr_id;
397                 wc.opcode = IB_WC_RECV;
398                 wc.status = err;
399                 ipath_cq_enter(to_icq(qp->ibqp.send_cq), &wc, 1);
400         }
401         wc.status = IB_WC_WR_FLUSH_ERR;
402
403         while (qp->s_last != qp->s_head) {
404                 struct ipath_swqe *wqe = get_swqe_ptr(qp, qp->s_last);
405
406                 wc.wr_id = wqe->wr.wr_id;
407                 wc.opcode = ib_ipath_wc_opcode[wqe->wr.opcode];
408                 if (++qp->s_last >= qp->s_size)
409                         qp->s_last = 0;
410                 ipath_cq_enter(to_icq(qp->ibqp.send_cq), &wc, 1);
411         }
412         qp->s_cur = qp->s_tail = qp->s_head;
413         qp->s_hdrwords = 0;
414         qp->s_ack_state = IB_OPCODE_RC_ACKNOWLEDGE;
415
416         if (qp->r_rq.wq) {
417                 struct ipath_rwq *wq;
418                 u32 head;
419                 u32 tail;
420
421                 spin_lock(&qp->r_rq.lock);
422
423                 /* sanity check pointers before trusting them */
424                 wq = qp->r_rq.wq;
425                 head = wq->head;
426                 if (head >= qp->r_rq.size)
427                         head = 0;
428                 tail = wq->tail;
429                 if (tail >= qp->r_rq.size)
430                         tail = 0;
431                 wc.opcode = IB_WC_RECV;
432                 while (tail != head) {
433                         wc.wr_id = get_rwqe_ptr(&qp->r_rq, tail)->wr_id;
434                         if (++tail >= qp->r_rq.size)
435                                 tail = 0;
436                         ipath_cq_enter(to_icq(qp->ibqp.recv_cq), &wc, 1);
437                 }
438                 wq->tail = tail;
439
440                 spin_unlock(&qp->r_rq.lock);
441         }
442 }
443
444 /**
445  * ipath_modify_qp - modify the attributes of a queue pair
446  * @ibqp: the queue pair who's attributes we're modifying
447  * @attr: the new attributes
448  * @attr_mask: the mask of attributes to modify
449  * @udata: user data for ipathverbs.so
450  *
451  * Returns 0 on success, otherwise returns an errno.
452  */
453 int ipath_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
454                     int attr_mask, struct ib_udata *udata)
455 {
456         struct ipath_ibdev *dev = to_idev(ibqp->device);
457         struct ipath_qp *qp = to_iqp(ibqp);
458         enum ib_qp_state cur_state, new_state;
459         unsigned long flags;
460         int ret;
461
462         spin_lock_irqsave(&qp->s_lock, flags);
463
464         cur_state = attr_mask & IB_QP_CUR_STATE ?
465                 attr->cur_qp_state : qp->state;
466         new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
467
468         if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type,
469                                 attr_mask))
470                 goto inval;
471
472         if (attr_mask & IB_QP_AV) {
473                 if (attr->ah_attr.dlid == 0 ||
474                     attr->ah_attr.dlid >= IPATH_MULTICAST_LID_BASE)
475                         goto inval;
476
477                 if ((attr->ah_attr.ah_flags & IB_AH_GRH) &&
478                     (attr->ah_attr.grh.sgid_index > 1))
479                         goto inval;
480         }
481
482         if (attr_mask & IB_QP_PKEY_INDEX)
483                 if (attr->pkey_index >= ipath_get_npkeys(dev->dd))
484                         goto inval;
485
486         if (attr_mask & IB_QP_MIN_RNR_TIMER)
487                 if (attr->min_rnr_timer > 31)
488                         goto inval;
489
490         if (attr_mask & IB_QP_PORT)
491                 if (attr->port_num == 0 ||
492                     attr->port_num > ibqp->device->phys_port_cnt)
493                         goto inval;
494
495         if (attr_mask & IB_QP_PATH_MTU)
496                 if (attr->path_mtu > IB_MTU_4096)
497                         goto inval;
498
499         if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
500                 if (attr->max_dest_rd_atomic > 1)
501                         goto inval;
502
503         if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC)
504                 if (attr->max_rd_atomic > 1)
505                         goto inval;
506
507         if (attr_mask & IB_QP_PATH_MIG_STATE)
508                 if (attr->path_mig_state != IB_MIG_MIGRATED &&
509                     attr->path_mig_state != IB_MIG_REARM)
510                         goto inval;
511
512         if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
513                 if (attr->max_dest_rd_atomic > IPATH_MAX_RDMA_ATOMIC)
514                         goto inval;
515
516         switch (new_state) {
517         case IB_QPS_RESET:
518                 ipath_reset_qp(qp);
519                 break;
520
521         case IB_QPS_ERR:
522                 ipath_error_qp(qp, IB_WC_GENERAL_ERR);
523                 break;
524
525         default:
526                 break;
527
528         }
529
530         if (attr_mask & IB_QP_PKEY_INDEX)
531                 qp->s_pkey_index = attr->pkey_index;
532
533         if (attr_mask & IB_QP_DEST_QPN)
534                 qp->remote_qpn = attr->dest_qp_num;
535
536         if (attr_mask & IB_QP_SQ_PSN) {
537                 qp->s_psn = qp->s_next_psn = attr->sq_psn;
538                 qp->s_last_psn = qp->s_next_psn - 1;
539         }
540
541         if (attr_mask & IB_QP_RQ_PSN)
542                 qp->r_psn = attr->rq_psn;
543
544         if (attr_mask & IB_QP_ACCESS_FLAGS)
545                 qp->qp_access_flags = attr->qp_access_flags;
546
547         if (attr_mask & IB_QP_AV)
548                 qp->remote_ah_attr = attr->ah_attr;
549
550         if (attr_mask & IB_QP_PATH_MTU)
551                 qp->path_mtu = attr->path_mtu;
552
553         if (attr_mask & IB_QP_RETRY_CNT)
554                 qp->s_retry = qp->s_retry_cnt = attr->retry_cnt;
555
556         if (attr_mask & IB_QP_RNR_RETRY) {
557                 qp->s_rnr_retry = attr->rnr_retry;
558                 if (qp->s_rnr_retry > 7)
559                         qp->s_rnr_retry = 7;
560                 qp->s_rnr_retry_cnt = qp->s_rnr_retry;
561         }
562
563         if (attr_mask & IB_QP_MIN_RNR_TIMER)
564                 qp->r_min_rnr_timer = attr->min_rnr_timer;
565
566         if (attr_mask & IB_QP_TIMEOUT)
567                 qp->timeout = attr->timeout;
568
569         if (attr_mask & IB_QP_QKEY)
570                 qp->qkey = attr->qkey;
571
572         if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
573                 qp->r_max_rd_atomic = attr->max_dest_rd_atomic;
574
575         if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC)
576                 qp->s_max_rd_atomic = attr->max_rd_atomic;
577
578         qp->state = new_state;
579         spin_unlock_irqrestore(&qp->s_lock, flags);
580
581         ret = 0;
582         goto bail;
583
584 inval:
585         spin_unlock_irqrestore(&qp->s_lock, flags);
586         ret = -EINVAL;
587
588 bail:
589         return ret;
590 }
591
592 int ipath_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
593                    int attr_mask, struct ib_qp_init_attr *init_attr)
594 {
595         struct ipath_qp *qp = to_iqp(ibqp);
596
597         attr->qp_state = qp->state;
598         attr->cur_qp_state = attr->qp_state;
599         attr->path_mtu = qp->path_mtu;
600         attr->path_mig_state = 0;
601         attr->qkey = qp->qkey;
602         attr->rq_psn = qp->r_psn;
603         attr->sq_psn = qp->s_next_psn;
604         attr->dest_qp_num = qp->remote_qpn;
605         attr->qp_access_flags = qp->qp_access_flags;
606         attr->cap.max_send_wr = qp->s_size - 1;
607         attr->cap.max_recv_wr = qp->ibqp.srq ? 0 : qp->r_rq.size - 1;
608         attr->cap.max_send_sge = qp->s_max_sge;
609         attr->cap.max_recv_sge = qp->r_rq.max_sge;
610         attr->cap.max_inline_data = 0;
611         attr->ah_attr = qp->remote_ah_attr;
612         memset(&attr->alt_ah_attr, 0, sizeof(attr->alt_ah_attr));
613         attr->pkey_index = qp->s_pkey_index;
614         attr->alt_pkey_index = 0;
615         attr->en_sqd_async_notify = 0;
616         attr->sq_draining = 0;
617         attr->max_rd_atomic = qp->s_max_rd_atomic;
618         attr->max_dest_rd_atomic = qp->r_max_rd_atomic;
619         attr->min_rnr_timer = qp->r_min_rnr_timer;
620         attr->port_num = 1;
621         attr->timeout = qp->timeout;
622         attr->retry_cnt = qp->s_retry_cnt;
623         attr->rnr_retry = qp->s_rnr_retry;
624         attr->alt_port_num = 0;
625         attr->alt_timeout = 0;
626
627         init_attr->event_handler = qp->ibqp.event_handler;
628         init_attr->qp_context = qp->ibqp.qp_context;
629         init_attr->send_cq = qp->ibqp.send_cq;
630         init_attr->recv_cq = qp->ibqp.recv_cq;
631         init_attr->srq = qp->ibqp.srq;
632         init_attr->cap = attr->cap;
633         if (qp->s_flags & IPATH_S_SIGNAL_REQ_WR)
634                 init_attr->sq_sig_type = IB_SIGNAL_REQ_WR;
635         else
636                 init_attr->sq_sig_type = IB_SIGNAL_ALL_WR;
637         init_attr->qp_type = qp->ibqp.qp_type;
638         init_attr->port_num = 1;
639         return 0;
640 }
641
642 /**
643  * ipath_compute_aeth - compute the AETH (syndrome + MSN)
644  * @qp: the queue pair to compute the AETH for
645  *
646  * Returns the AETH.
647  */
648 __be32 ipath_compute_aeth(struct ipath_qp *qp)
649 {
650         u32 aeth = qp->r_msn & IPATH_MSN_MASK;
651
652         if (qp->ibqp.srq) {
653                 /*
654                  * Shared receive queues don't generate credits.
655                  * Set the credit field to the invalid value.
656                  */
657                 aeth |= IPATH_AETH_CREDIT_INVAL << IPATH_AETH_CREDIT_SHIFT;
658         } else {
659                 u32 min, max, x;
660                 u32 credits;
661                 struct ipath_rwq *wq = qp->r_rq.wq;
662                 u32 head;
663                 u32 tail;
664
665                 /* sanity check pointers before trusting them */
666                 head = wq->head;
667                 if (head >= qp->r_rq.size)
668                         head = 0;
669                 tail = wq->tail;
670                 if (tail >= qp->r_rq.size)
671                         tail = 0;
672                 /*
673                  * Compute the number of credits available (RWQEs).
674                  * XXX Not holding the r_rq.lock here so there is a small
675                  * chance that the pair of reads are not atomic.
676                  */
677                 credits = head - tail;
678                 if ((int)credits < 0)
679                         credits += qp->r_rq.size;
680                 /*
681                  * Binary search the credit table to find the code to
682                  * use.
683                  */
684                 min = 0;
685                 max = 31;
686                 for (;;) {
687                         x = (min + max) / 2;
688                         if (credit_table[x] == credits)
689                                 break;
690                         if (credit_table[x] > credits)
691                                 max = x;
692                         else if (min == x)
693                                 break;
694                         else
695                                 min = x;
696                 }
697                 aeth |= x << IPATH_AETH_CREDIT_SHIFT;
698         }
699         return cpu_to_be32(aeth);
700 }
701
702 /**
703  * ipath_create_qp - create a queue pair for a device
704  * @ibpd: the protection domain who's device we create the queue pair for
705  * @init_attr: the attributes of the queue pair
706  * @udata: unused by InfiniPath
707  *
708  * Returns the queue pair on success, otherwise returns an errno.
709  *
710  * Called by the ib_create_qp() core verbs function.
711  */
712 struct ib_qp *ipath_create_qp(struct ib_pd *ibpd,
713                               struct ib_qp_init_attr *init_attr,
714                               struct ib_udata *udata)
715 {
716         struct ipath_qp *qp;
717         int err;
718         struct ipath_swqe *swq = NULL;
719         struct ipath_ibdev *dev;
720         size_t sz;
721         struct ib_qp *ret;
722
723         if (init_attr->cap.max_send_sge > ib_ipath_max_sges ||
724             init_attr->cap.max_recv_sge > ib_ipath_max_sges ||
725             init_attr->cap.max_send_wr > ib_ipath_max_qp_wrs ||
726             init_attr->cap.max_recv_wr > ib_ipath_max_qp_wrs) {
727                 ret = ERR_PTR(-ENOMEM);
728                 goto bail;
729         }
730
731         if (init_attr->cap.max_send_sge +
732             init_attr->cap.max_recv_sge +
733             init_attr->cap.max_send_wr +
734             init_attr->cap.max_recv_wr == 0) {
735                 ret = ERR_PTR(-EINVAL);
736                 goto bail;
737         }
738
739         switch (init_attr->qp_type) {
740         case IB_QPT_UC:
741         case IB_QPT_RC:
742                 sz = sizeof(struct ipath_sge) *
743                         init_attr->cap.max_send_sge +
744                         sizeof(struct ipath_swqe);
745                 swq = vmalloc((init_attr->cap.max_send_wr + 1) * sz);
746                 if (swq == NULL) {
747                         ret = ERR_PTR(-ENOMEM);
748                         goto bail;
749                 }
750                 /* FALLTHROUGH */
751         case IB_QPT_UD:
752         case IB_QPT_SMI:
753         case IB_QPT_GSI:
754                 sz = sizeof(*qp);
755                 if (init_attr->srq) {
756                         struct ipath_srq *srq = to_isrq(init_attr->srq);
757
758                         sz += sizeof(*qp->r_sg_list) *
759                                 srq->rq.max_sge;
760                 } else
761                         sz += sizeof(*qp->r_sg_list) *
762                                 init_attr->cap.max_recv_sge;
763                 qp = kmalloc(sz, GFP_KERNEL);
764                 if (!qp) {
765                         ret = ERR_PTR(-ENOMEM);
766                         goto bail_swq;
767                 }
768                 if (init_attr->srq) {
769                         sz = 0;
770                         qp->r_rq.size = 0;
771                         qp->r_rq.max_sge = 0;
772                         qp->r_rq.wq = NULL;
773                         init_attr->cap.max_recv_wr = 0;
774                         init_attr->cap.max_recv_sge = 0;
775                 } else {
776                         qp->r_rq.size = init_attr->cap.max_recv_wr + 1;
777                         qp->r_rq.max_sge = init_attr->cap.max_recv_sge;
778                         sz = (sizeof(struct ib_sge) * qp->r_rq.max_sge) +
779                                 sizeof(struct ipath_rwqe);
780                         qp->r_rq.wq = vmalloc_user(sizeof(struct ipath_rwq) +
781                                               qp->r_rq.size * sz);
782                         if (!qp->r_rq.wq) {
783                                 ret = ERR_PTR(-ENOMEM);
784                                 goto bail_qp;
785                         }
786                 }
787
788                 /*
789                  * ib_create_qp() will initialize qp->ibqp
790                  * except for qp->ibqp.qp_num.
791                  */
792                 spin_lock_init(&qp->s_lock);
793                 spin_lock_init(&qp->r_rq.lock);
794                 atomic_set(&qp->refcount, 0);
795                 init_waitqueue_head(&qp->wait);
796                 tasklet_init(&qp->s_task, ipath_do_ruc_send,
797                              (unsigned long)qp);
798                 INIT_LIST_HEAD(&qp->piowait);
799                 INIT_LIST_HEAD(&qp->timerwait);
800                 qp->state = IB_QPS_RESET;
801                 qp->s_wq = swq;
802                 qp->s_size = init_attr->cap.max_send_wr + 1;
803                 qp->s_max_sge = init_attr->cap.max_send_sge;
804                 if (init_attr->sq_sig_type == IB_SIGNAL_REQ_WR)
805                         qp->s_flags = IPATH_S_SIGNAL_REQ_WR;
806                 else
807                         qp->s_flags = 0;
808                 dev = to_idev(ibpd->device);
809                 err = ipath_alloc_qpn(&dev->qp_table, qp,
810                                       init_attr->qp_type);
811                 if (err) {
812                         ret = ERR_PTR(err);
813                         goto bail_rwq;
814                 }
815                 qp->ip = NULL;
816                 ipath_reset_qp(qp);
817                 break;
818
819         default:
820                 /* Don't support raw QPs */
821                 ret = ERR_PTR(-ENOSYS);
822                 goto bail;
823         }
824
825         init_attr->cap.max_inline_data = 0;
826
827         /*
828          * Return the address of the RWQ as the offset to mmap.
829          * See ipath_mmap() for details.
830          */
831         if (udata && udata->outlen >= sizeof(__u64)) {
832                 struct ipath_mmap_info *ip;
833                 __u64 offset = (__u64) qp->r_rq.wq;
834                 int err;
835
836                 err = ib_copy_to_udata(udata, &offset, sizeof(offset));
837                 if (err) {
838                         ret = ERR_PTR(err);
839                         goto bail_rwq;
840                 }
841
842                 if (qp->r_rq.wq) {
843                         /* Allocate info for ipath_mmap(). */
844                         ip = kmalloc(sizeof(*ip), GFP_KERNEL);
845                         if (!ip) {
846                                 ret = ERR_PTR(-ENOMEM);
847                                 goto bail_rwq;
848                         }
849                         qp->ip = ip;
850                         ip->context = ibpd->uobject->context;
851                         ip->obj = qp->r_rq.wq;
852                         kref_init(&ip->ref);
853                         ip->mmap_cnt = 0;
854                         ip->size = PAGE_ALIGN(sizeof(struct ipath_rwq) +
855                                               qp->r_rq.size * sz);
856                         spin_lock_irq(&dev->pending_lock);
857                         ip->next = dev->pending_mmaps;
858                         dev->pending_mmaps = ip;
859                         spin_unlock_irq(&dev->pending_lock);
860                 }
861         }
862
863         spin_lock(&dev->n_qps_lock);
864         if (dev->n_qps_allocated == ib_ipath_max_qps) {
865                 spin_unlock(&dev->n_qps_lock);
866                 ret = ERR_PTR(-ENOMEM);
867                 goto bail_ip;
868         }
869
870         dev->n_qps_allocated++;
871         spin_unlock(&dev->n_qps_lock);
872
873         ret = &qp->ibqp;
874         goto bail;
875
876 bail_ip:
877         kfree(qp->ip);
878 bail_rwq:
879         vfree(qp->r_rq.wq);
880 bail_qp:
881         kfree(qp);
882 bail_swq:
883         vfree(swq);
884 bail:
885         return ret;
886 }
887
888 /**
889  * ipath_destroy_qp - destroy a queue pair
890  * @ibqp: the queue pair to destroy
891  *
892  * Returns 0 on success.
893  *
894  * Note that this can be called while the QP is actively sending or
895  * receiving!
896  */
897 int ipath_destroy_qp(struct ib_qp *ibqp)
898 {
899         struct ipath_qp *qp = to_iqp(ibqp);
900         struct ipath_ibdev *dev = to_idev(ibqp->device);
901         unsigned long flags;
902
903         spin_lock_irqsave(&qp->s_lock, flags);
904         qp->state = IB_QPS_ERR;
905         spin_unlock_irqrestore(&qp->s_lock, flags);
906         spin_lock(&dev->n_qps_lock);
907         dev->n_qps_allocated--;
908         spin_unlock(&dev->n_qps_lock);
909
910         /* Stop the sending tasklet. */
911         tasklet_kill(&qp->s_task);
912
913         /* Make sure the QP isn't on the timeout list. */
914         spin_lock_irqsave(&dev->pending_lock, flags);
915         if (!list_empty(&qp->timerwait))
916                 list_del_init(&qp->timerwait);
917         if (!list_empty(&qp->piowait))
918                 list_del_init(&qp->piowait);
919         spin_unlock_irqrestore(&dev->pending_lock, flags);
920
921         /*
922          * Make sure that the QP is not in the QPN table so receive
923          * interrupts will discard packets for this QP.  XXX Also remove QP
924          * from multicast table.
925          */
926         if (atomic_read(&qp->refcount) != 0)
927                 ipath_free_qp(&dev->qp_table, qp);
928
929         if (qp->ip)
930                 kref_put(&qp->ip->ref, ipath_release_mmap_info);
931         else
932                 vfree(qp->r_rq.wq);
933         vfree(qp->s_wq);
934         kfree(qp);
935         return 0;
936 }
937
938 /**
939  * ipath_init_qp_table - initialize the QP table for a device
940  * @idev: the device who's QP table we're initializing
941  * @size: the size of the QP table
942  *
943  * Returns 0 on success, otherwise returns an errno.
944  */
945 int ipath_init_qp_table(struct ipath_ibdev *idev, int size)
946 {
947         int i;
948         int ret;
949
950         idev->qp_table.last = 1;        /* QPN 0 and 1 are special. */
951         idev->qp_table.max = size;
952         idev->qp_table.nmaps = 1;
953         idev->qp_table.table = kzalloc(size * sizeof(*idev->qp_table.table),
954                                        GFP_KERNEL);
955         if (idev->qp_table.table == NULL) {
956                 ret = -ENOMEM;
957                 goto bail;
958         }
959
960         for (i = 0; i < ARRAY_SIZE(idev->qp_table.map); i++) {
961                 atomic_set(&idev->qp_table.map[i].n_free, BITS_PER_PAGE);
962                 idev->qp_table.map[i].page = NULL;
963         }
964
965         ret = 0;
966
967 bail:
968         return ret;
969 }
970
971 /**
972  * ipath_sqerror_qp - put a QP's send queue into an error state
973  * @qp: QP who's send queue will be put into an error state
974  * @wc: the WC responsible for putting the QP in this state
975  *
976  * Flushes the send work queue.
977  * The QP s_lock should be held and interrupts disabled.
978  */
979
980 void ipath_sqerror_qp(struct ipath_qp *qp, struct ib_wc *wc)
981 {
982         struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
983         struct ipath_swqe *wqe = get_swqe_ptr(qp, qp->s_last);
984
985         ipath_dbg("Send queue error on QP%d/%d: err: %d\n",
986                   qp->ibqp.qp_num, qp->remote_qpn, wc->status);
987
988         spin_lock(&dev->pending_lock);
989         /* XXX What if its already removed by the timeout code? */
990         if (!list_empty(&qp->timerwait))
991                 list_del_init(&qp->timerwait);
992         if (!list_empty(&qp->piowait))
993                 list_del_init(&qp->piowait);
994         spin_unlock(&dev->pending_lock);
995
996         ipath_cq_enter(to_icq(qp->ibqp.send_cq), wc, 1);
997         if (++qp->s_last >= qp->s_size)
998                 qp->s_last = 0;
999
1000         wc->status = IB_WC_WR_FLUSH_ERR;
1001
1002         while (qp->s_last != qp->s_head) {
1003                 wqe = get_swqe_ptr(qp, qp->s_last);
1004                 wc->wr_id = wqe->wr.wr_id;
1005                 wc->opcode = ib_ipath_wc_opcode[wqe->wr.opcode];
1006                 ipath_cq_enter(to_icq(qp->ibqp.send_cq), wc, 1);
1007                 if (++qp->s_last >= qp->s_size)
1008                         qp->s_last = 0;
1009         }
1010         qp->s_cur = qp->s_tail = qp->s_head;
1011         qp->state = IB_QPS_SQE;
1012 }
1013
1014 /**
1015  * ipath_get_credit - flush the send work queue of a QP
1016  * @qp: the qp who's send work queue to flush
1017  * @aeth: the Acknowledge Extended Transport Header
1018  *
1019  * The QP s_lock should be held.
1020  */
1021 void ipath_get_credit(struct ipath_qp *qp, u32 aeth)
1022 {
1023         u32 credit = (aeth >> IPATH_AETH_CREDIT_SHIFT) & IPATH_AETH_CREDIT_MASK;
1024
1025         /*
1026          * If the credit is invalid, we can send
1027          * as many packets as we like.  Otherwise, we have to
1028          * honor the credit field.
1029          */
1030         if (credit == IPATH_AETH_CREDIT_INVAL)
1031                 qp->s_lsn = (u32) -1;
1032         else if (qp->s_lsn != (u32) -1) {
1033                 /* Compute new LSN (i.e., MSN + credit) */
1034                 credit = (aeth + credit_table[credit]) & IPATH_MSN_MASK;
1035                 if (ipath_cmp24(credit, qp->s_lsn) > 0)
1036                         qp->s_lsn = credit;
1037         }
1038
1039         /* Restart sending if it was blocked due to lack of credits. */
1040         if (qp->s_cur != qp->s_head &&
1041             (qp->s_lsn == (u32) -1 ||
1042              ipath_cmp24(get_swqe_ptr(qp, qp->s_cur)->ssn,
1043                          qp->s_lsn + 1) <= 0))
1044                 tasklet_hi_schedule(&qp->s_task);
1045 }