Merge branch 'master' of git://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux-2.6
[linux-2.6] / drivers / infiniband / hw / mthca / mthca_srq.c
1 /*
2  * Copyright (c) 2005 Cisco Systems. All rights reserved.
3  *
4  * This software is available to you under a choice of one of two
5  * licenses.  You may choose to be licensed under the terms of the GNU
6  * General Public License (GPL) Version 2, available from the file
7  * COPYING in the main directory of this source tree, or the
8  * OpenIB.org BSD license below:
9  *
10  *     Redistribution and use in source and binary forms, with or
11  *     without modification, are permitted provided that the following
12  *     conditions are met:
13  *
14  *      - Redistributions of source code must retain the above
15  *        copyright notice, this list of conditions and the following
16  *        disclaimer.
17  *
18  *      - Redistributions in binary form must reproduce the above
19  *        copyright notice, this list of conditions and the following
20  *        disclaimer in the documentation and/or other materials
21  *        provided with the distribution.
22  *
23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
30  * SOFTWARE.
31  *
32  * $Id: mthca_srq.c 3047 2005-08-10 03:59:35Z roland $
33  */
34
35 #include <linux/slab.h>
36 #include <linux/string.h>
37
38 #include <asm/io.h>
39
40 #include "mthca_dev.h"
41 #include "mthca_cmd.h"
42 #include "mthca_memfree.h"
43 #include "mthca_wqe.h"
44
45 enum {
46         MTHCA_MAX_DIRECT_SRQ_SIZE = 4 * PAGE_SIZE
47 };
48
49 struct mthca_tavor_srq_context {
50         __be64 wqe_base_ds;     /* low 6 bits is descriptor size */
51         __be32 state_pd;
52         __be32 lkey;
53         __be32 uar;
54         __be16 limit_watermark;
55         __be16 wqe_cnt;
56         u32    reserved[2];
57 };
58
59 struct mthca_arbel_srq_context {
60         __be32 state_logsize_srqn;
61         __be32 lkey;
62         __be32 db_index;
63         __be32 logstride_usrpage;
64         __be64 wqe_base;
65         __be32 eq_pd;
66         __be16 limit_watermark;
67         __be16 wqe_cnt;
68         u16    reserved1;
69         __be16 wqe_counter;
70         u32    reserved2[3];
71 };
72
73 static void *get_wqe(struct mthca_srq *srq, int n)
74 {
75         if (srq->is_direct)
76                 return srq->queue.direct.buf + (n << srq->wqe_shift);
77         else
78                 return srq->queue.page_list[(n << srq->wqe_shift) >> PAGE_SHIFT].buf +
79                         ((n << srq->wqe_shift) & (PAGE_SIZE - 1));
80 }
81
82 /*
83  * Return a pointer to the location within a WQE that we're using as a
84  * link when the WQE is in the free list.  We use the imm field
85  * because in the Tavor case, posting a WQE may overwrite the next
86  * segment of the previous WQE, but a receive WQE will never touch the
87  * imm field.  This avoids corrupting our free list if the previous
88  * WQE has already completed and been put on the free list when we
89  * post the next WQE.
90  */
91 static inline int *wqe_to_link(void *wqe)
92 {
93         return (int *) (wqe + offsetof(struct mthca_next_seg, imm));
94 }
95
96 static void mthca_tavor_init_srq_context(struct mthca_dev *dev,
97                                          struct mthca_pd *pd,
98                                          struct mthca_srq *srq,
99                                          struct mthca_tavor_srq_context *context)
100 {
101         memset(context, 0, sizeof *context);
102
103         context->wqe_base_ds = cpu_to_be64(1 << (srq->wqe_shift - 4));
104         context->state_pd    = cpu_to_be32(pd->pd_num);
105         context->lkey        = cpu_to_be32(srq->mr.ibmr.lkey);
106
107         if (pd->ibpd.uobject)
108                 context->uar =
109                         cpu_to_be32(to_mucontext(pd->ibpd.uobject->context)->uar.index);
110         else
111                 context->uar = cpu_to_be32(dev->driver_uar.index);
112 }
113
114 static void mthca_arbel_init_srq_context(struct mthca_dev *dev,
115                                          struct mthca_pd *pd,
116                                          struct mthca_srq *srq,
117                                          struct mthca_arbel_srq_context *context)
118 {
119         int logsize;
120
121         memset(context, 0, sizeof *context);
122
123         logsize = long_log2(srq->max) + srq->wqe_shift;
124         context->state_logsize_srqn = cpu_to_be32(logsize << 24 | srq->srqn);
125         context->lkey = cpu_to_be32(srq->mr.ibmr.lkey);
126         context->db_index = cpu_to_be32(srq->db_index);
127         context->logstride_usrpage = cpu_to_be32((srq->wqe_shift - 4) << 29);
128         if (pd->ibpd.uobject)
129                 context->logstride_usrpage |=
130                         cpu_to_be32(to_mucontext(pd->ibpd.uobject->context)->uar.index);
131         else
132                 context->logstride_usrpage |= cpu_to_be32(dev->driver_uar.index);
133         context->eq_pd = cpu_to_be32(MTHCA_EQ_ASYNC << 24 | pd->pd_num);
134 }
135
136 static void mthca_free_srq_buf(struct mthca_dev *dev, struct mthca_srq *srq)
137 {
138         mthca_buf_free(dev, srq->max << srq->wqe_shift, &srq->queue,
139                        srq->is_direct, &srq->mr);
140         kfree(srq->wrid);
141 }
142
143 static int mthca_alloc_srq_buf(struct mthca_dev *dev, struct mthca_pd *pd,
144                                struct mthca_srq *srq)
145 {
146         struct mthca_data_seg *scatter;
147         void *wqe;
148         int err;
149         int i;
150
151         if (pd->ibpd.uobject)
152                 return 0;
153
154         srq->wrid = kmalloc(srq->max * sizeof (u64), GFP_KERNEL);
155         if (!srq->wrid)
156                 return -ENOMEM;
157
158         err = mthca_buf_alloc(dev, srq->max << srq->wqe_shift,
159                               MTHCA_MAX_DIRECT_SRQ_SIZE,
160                               &srq->queue, &srq->is_direct, pd, 1, &srq->mr);
161         if (err) {
162                 kfree(srq->wrid);
163                 return err;
164         }
165
166         /*
167          * Now initialize the SRQ buffer so that all of the WQEs are
168          * linked into the list of free WQEs.  In addition, set the
169          * scatter list L_Keys to the sentry value of 0x100.
170          */
171         for (i = 0; i < srq->max; ++i) {
172                 wqe = get_wqe(srq, i);
173
174                 *wqe_to_link(wqe) = i < srq->max - 1 ? i + 1 : -1;
175
176                 for (scatter = wqe + sizeof (struct mthca_next_seg);
177                      (void *) scatter < wqe + (1 << srq->wqe_shift);
178                      ++scatter)
179                         scatter->lkey = cpu_to_be32(MTHCA_INVAL_LKEY);
180         }
181
182         srq->last = get_wqe(srq, srq->max - 1);
183
184         return 0;
185 }
186
187 int mthca_alloc_srq(struct mthca_dev *dev, struct mthca_pd *pd,
188                     struct ib_srq_attr *attr, struct mthca_srq *srq)
189 {
190         struct mthca_mailbox *mailbox;
191         u8 status;
192         int ds;
193         int err;
194
195         /* Sanity check SRQ size before proceeding */
196         if (attr->max_wr  > dev->limits.max_srq_wqes ||
197             attr->max_sge > dev->limits.max_srq_sge)
198                 return -EINVAL;
199
200         srq->max      = attr->max_wr;
201         srq->max_gs   = attr->max_sge;
202         srq->counter  = 0;
203
204         if (mthca_is_memfree(dev))
205                 srq->max = roundup_pow_of_two(srq->max + 1);
206         else
207                 srq->max = srq->max + 1;
208
209         ds = max(64UL,
210                  roundup_pow_of_two(sizeof (struct mthca_next_seg) +
211                                     srq->max_gs * sizeof (struct mthca_data_seg)));
212
213         if (!mthca_is_memfree(dev) && (ds > dev->limits.max_desc_sz))
214                 return -EINVAL;
215
216         srq->wqe_shift = long_log2(ds);
217
218         srq->srqn = mthca_alloc(&dev->srq_table.alloc);
219         if (srq->srqn == -1)
220                 return -ENOMEM;
221
222         if (mthca_is_memfree(dev)) {
223                 err = mthca_table_get(dev, dev->srq_table.table, srq->srqn);
224                 if (err)
225                         goto err_out;
226
227                 if (!pd->ibpd.uobject) {
228                         srq->db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_SRQ,
229                                                        srq->srqn, &srq->db);
230                         if (srq->db_index < 0) {
231                                 err = -ENOMEM;
232                                 goto err_out_icm;
233                         }
234                 }
235         }
236
237         mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
238         if (IS_ERR(mailbox)) {
239                 err = PTR_ERR(mailbox);
240                 goto err_out_db;
241         }
242
243         err = mthca_alloc_srq_buf(dev, pd, srq);
244         if (err)
245                 goto err_out_mailbox;
246
247         spin_lock_init(&srq->lock);
248         srq->refcount = 1;
249         init_waitqueue_head(&srq->wait);
250         mutex_init(&srq->mutex);
251
252         if (mthca_is_memfree(dev))
253                 mthca_arbel_init_srq_context(dev, pd, srq, mailbox->buf);
254         else
255                 mthca_tavor_init_srq_context(dev, pd, srq, mailbox->buf);
256
257         err = mthca_SW2HW_SRQ(dev, mailbox, srq->srqn, &status);
258
259         if (err) {
260                 mthca_warn(dev, "SW2HW_SRQ failed (%d)\n", err);
261                 goto err_out_free_buf;
262         }
263         if (status) {
264                 mthca_warn(dev, "SW2HW_SRQ returned status 0x%02x\n",
265                            status);
266                 err = -EINVAL;
267                 goto err_out_free_buf;
268         }
269
270         spin_lock_irq(&dev->srq_table.lock);
271         if (mthca_array_set(&dev->srq_table.srq,
272                             srq->srqn & (dev->limits.num_srqs - 1),
273                             srq)) {
274                 spin_unlock_irq(&dev->srq_table.lock);
275                 goto err_out_free_srq;
276         }
277         spin_unlock_irq(&dev->srq_table.lock);
278
279         mthca_free_mailbox(dev, mailbox);
280
281         srq->first_free = 0;
282         srq->last_free  = srq->max - 1;
283
284         attr->max_wr    = srq->max - 1;
285         attr->max_sge   = srq->max_gs;
286
287         return 0;
288
289 err_out_free_srq:
290         err = mthca_HW2SW_SRQ(dev, mailbox, srq->srqn, &status);
291         if (err)
292                 mthca_warn(dev, "HW2SW_SRQ failed (%d)\n", err);
293         else if (status)
294                 mthca_warn(dev, "HW2SW_SRQ returned status 0x%02x\n", status);
295
296 err_out_free_buf:
297         if (!pd->ibpd.uobject)
298                 mthca_free_srq_buf(dev, srq);
299
300 err_out_mailbox:
301         mthca_free_mailbox(dev, mailbox);
302
303 err_out_db:
304         if (!pd->ibpd.uobject && mthca_is_memfree(dev))
305                 mthca_free_db(dev, MTHCA_DB_TYPE_SRQ, srq->db_index);
306
307 err_out_icm:
308         mthca_table_put(dev, dev->srq_table.table, srq->srqn);
309
310 err_out:
311         mthca_free(&dev->srq_table.alloc, srq->srqn);
312
313         return err;
314 }
315
316 static inline int get_srq_refcount(struct mthca_dev *dev, struct mthca_srq *srq)
317 {
318         int c;
319
320         spin_lock_irq(&dev->srq_table.lock);
321         c = srq->refcount;
322         spin_unlock_irq(&dev->srq_table.lock);
323
324         return c;
325 }
326
327 void mthca_free_srq(struct mthca_dev *dev, struct mthca_srq *srq)
328 {
329         struct mthca_mailbox *mailbox;
330         int err;
331         u8 status;
332
333         mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
334         if (IS_ERR(mailbox)) {
335                 mthca_warn(dev, "No memory for mailbox to free SRQ.\n");
336                 return;
337         }
338
339         err = mthca_HW2SW_SRQ(dev, mailbox, srq->srqn, &status);
340         if (err)
341                 mthca_warn(dev, "HW2SW_SRQ failed (%d)\n", err);
342         else if (status)
343                 mthca_warn(dev, "HW2SW_SRQ returned status 0x%02x\n", status);
344
345         spin_lock_irq(&dev->srq_table.lock);
346         mthca_array_clear(&dev->srq_table.srq,
347                           srq->srqn & (dev->limits.num_srqs - 1));
348         --srq->refcount;
349         spin_unlock_irq(&dev->srq_table.lock);
350
351         wait_event(srq->wait, !get_srq_refcount(dev, srq));
352
353         if (!srq->ibsrq.uobject) {
354                 mthca_free_srq_buf(dev, srq);
355                 if (mthca_is_memfree(dev))
356                         mthca_free_db(dev, MTHCA_DB_TYPE_SRQ, srq->db_index);
357         }
358
359         mthca_table_put(dev, dev->srq_table.table, srq->srqn);
360         mthca_free(&dev->srq_table.alloc, srq->srqn);
361         mthca_free_mailbox(dev, mailbox);
362 }
363
364 int mthca_modify_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr,
365                      enum ib_srq_attr_mask attr_mask, struct ib_udata *udata)
366 {
367         struct mthca_dev *dev = to_mdev(ibsrq->device);
368         struct mthca_srq *srq = to_msrq(ibsrq);
369         int ret;
370         u8 status;
371
372         /* We don't support resizing SRQs (yet?) */
373         if (attr_mask & IB_SRQ_MAX_WR)
374                 return -EINVAL;
375
376         if (attr_mask & IB_SRQ_LIMIT) {
377                 u32 max_wr = mthca_is_memfree(dev) ? srq->max - 1 : srq->max;
378                 if (attr->srq_limit > max_wr)
379                         return -EINVAL;
380
381                 mutex_lock(&srq->mutex);
382                 ret = mthca_ARM_SRQ(dev, srq->srqn, attr->srq_limit, &status);
383                 mutex_unlock(&srq->mutex);
384
385                 if (ret)
386                         return ret;
387                 if (status)
388                         return -EINVAL;
389         }
390
391         return 0;
392 }
393
394 int mthca_query_srq(struct ib_srq *ibsrq, struct ib_srq_attr *srq_attr)
395 {
396         struct mthca_dev *dev = to_mdev(ibsrq->device);
397         struct mthca_srq *srq = to_msrq(ibsrq);
398         struct mthca_mailbox *mailbox;
399         struct mthca_arbel_srq_context *arbel_ctx;
400         struct mthca_tavor_srq_context *tavor_ctx;
401         u8 status;
402         int err;
403
404         mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
405         if (IS_ERR(mailbox))
406                 return PTR_ERR(mailbox);
407
408         err = mthca_QUERY_SRQ(dev, srq->srqn, mailbox, &status);
409         if (err)
410                 goto out;
411
412         if (mthca_is_memfree(dev)) {
413                 arbel_ctx = mailbox->buf;
414                 srq_attr->srq_limit = be16_to_cpu(arbel_ctx->limit_watermark);
415         } else {
416                 tavor_ctx = mailbox->buf;
417                 srq_attr->srq_limit = be16_to_cpu(tavor_ctx->limit_watermark);
418         }
419
420         srq_attr->max_wr  = srq->max - 1;
421         srq_attr->max_sge = srq->max_gs;
422
423 out:
424         mthca_free_mailbox(dev, mailbox);
425
426         return err;
427 }
428
429 void mthca_srq_event(struct mthca_dev *dev, u32 srqn,
430                      enum ib_event_type event_type)
431 {
432         struct mthca_srq *srq;
433         struct ib_event event;
434
435         spin_lock(&dev->srq_table.lock);
436         srq = mthca_array_get(&dev->srq_table.srq, srqn & (dev->limits.num_srqs - 1));
437         if (srq)
438                 ++srq->refcount;
439         spin_unlock(&dev->srq_table.lock);
440
441         if (!srq) {
442                 mthca_warn(dev, "Async event for bogus SRQ %08x\n", srqn);
443                 return;
444         }
445
446         if (!srq->ibsrq.event_handler)
447                 goto out;
448
449         event.device      = &dev->ib_dev;
450         event.event       = event_type;
451         event.element.srq = &srq->ibsrq;
452         srq->ibsrq.event_handler(&event, srq->ibsrq.srq_context);
453
454 out:
455         spin_lock(&dev->srq_table.lock);
456         if (!--srq->refcount)
457                 wake_up(&srq->wait);
458         spin_unlock(&dev->srq_table.lock);
459 }
460
461 /*
462  * This function must be called with IRQs disabled.
463  */
464 void mthca_free_srq_wqe(struct mthca_srq *srq, u32 wqe_addr)
465 {
466         int ind;
467
468         ind = wqe_addr >> srq->wqe_shift;
469
470         spin_lock(&srq->lock);
471
472         if (likely(srq->first_free >= 0))
473                 *wqe_to_link(get_wqe(srq, srq->last_free)) = ind;
474         else
475                 srq->first_free = ind;
476
477         *wqe_to_link(get_wqe(srq, ind)) = -1;
478         srq->last_free = ind;
479
480         spin_unlock(&srq->lock);
481 }
482
483 int mthca_tavor_post_srq_recv(struct ib_srq *ibsrq, struct ib_recv_wr *wr,
484                               struct ib_recv_wr **bad_wr)
485 {
486         struct mthca_dev *dev = to_mdev(ibsrq->device);
487         struct mthca_srq *srq = to_msrq(ibsrq);
488         __be32 doorbell[2];
489         unsigned long flags;
490         int err = 0;
491         int first_ind;
492         int ind;
493         int next_ind;
494         int nreq;
495         int i;
496         void *wqe;
497         void *prev_wqe;
498
499         spin_lock_irqsave(&srq->lock, flags);
500
501         first_ind = srq->first_free;
502
503         for (nreq = 0; wr; wr = wr->next) {
504                 ind = srq->first_free;
505
506                 if (ind < 0) {
507                         mthca_err(dev, "SRQ %06x full\n", srq->srqn);
508                         err = -ENOMEM;
509                         *bad_wr = wr;
510                         break;
511                 }
512
513                 wqe       = get_wqe(srq, ind);
514                 next_ind  = *wqe_to_link(wqe);
515
516                 if (next_ind < 0) {
517                         mthca_err(dev, "SRQ %06x full\n", srq->srqn);
518                         err = -ENOMEM;
519                         *bad_wr = wr;
520                         break;
521                 }
522
523                 prev_wqe  = srq->last;
524                 srq->last = wqe;
525
526                 ((struct mthca_next_seg *) wqe)->nda_op = 0;
527                 ((struct mthca_next_seg *) wqe)->ee_nds = 0;
528                 /* flags field will always remain 0 */
529
530                 wqe += sizeof (struct mthca_next_seg);
531
532                 if (unlikely(wr->num_sge > srq->max_gs)) {
533                         err = -EINVAL;
534                         *bad_wr = wr;
535                         srq->last = prev_wqe;
536                         break;
537                 }
538
539                 for (i = 0; i < wr->num_sge; ++i) {
540                         ((struct mthca_data_seg *) wqe)->byte_count =
541                                 cpu_to_be32(wr->sg_list[i].length);
542                         ((struct mthca_data_seg *) wqe)->lkey =
543                                 cpu_to_be32(wr->sg_list[i].lkey);
544                         ((struct mthca_data_seg *) wqe)->addr =
545                                 cpu_to_be64(wr->sg_list[i].addr);
546                         wqe += sizeof (struct mthca_data_seg);
547                 }
548
549                 if (i < srq->max_gs) {
550                         ((struct mthca_data_seg *) wqe)->byte_count = 0;
551                         ((struct mthca_data_seg *) wqe)->lkey = cpu_to_be32(MTHCA_INVAL_LKEY);
552                         ((struct mthca_data_seg *) wqe)->addr = 0;
553                 }
554
555                 ((struct mthca_next_seg *) prev_wqe)->nda_op =
556                         cpu_to_be32((ind << srq->wqe_shift) | 1);
557                 wmb();
558                 ((struct mthca_next_seg *) prev_wqe)->ee_nds =
559                         cpu_to_be32(MTHCA_NEXT_DBD);
560
561                 srq->wrid[ind]  = wr->wr_id;
562                 srq->first_free = next_ind;
563
564                 ++nreq;
565                 if (unlikely(nreq == MTHCA_TAVOR_MAX_WQES_PER_RECV_DB)) {
566                         nreq = 0;
567
568                         doorbell[0] = cpu_to_be32(first_ind << srq->wqe_shift);
569                         doorbell[1] = cpu_to_be32(srq->srqn << 8);
570
571                         /*
572                          * Make sure that descriptors are written
573                          * before doorbell is rung.
574                          */
575                         wmb();
576
577                         mthca_write64(doorbell,
578                                       dev->kar + MTHCA_RECEIVE_DOORBELL,
579                                       MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
580
581                         first_ind = srq->first_free;
582                 }
583         }
584
585         if (likely(nreq)) {
586                 doorbell[0] = cpu_to_be32(first_ind << srq->wqe_shift);
587                 doorbell[1] = cpu_to_be32((srq->srqn << 8) | nreq);
588
589                 /*
590                  * Make sure that descriptors are written before
591                  * doorbell is rung.
592                  */
593                 wmb();
594
595                 mthca_write64(doorbell,
596                               dev->kar + MTHCA_RECEIVE_DOORBELL,
597                               MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
598         }
599
600         /*
601          * Make sure doorbells don't leak out of SRQ spinlock and
602          * reach the HCA out of order:
603          */
604         mmiowb();
605
606         spin_unlock_irqrestore(&srq->lock, flags);
607         return err;
608 }
609
610 int mthca_arbel_post_srq_recv(struct ib_srq *ibsrq, struct ib_recv_wr *wr,
611                               struct ib_recv_wr **bad_wr)
612 {
613         struct mthca_dev *dev = to_mdev(ibsrq->device);
614         struct mthca_srq *srq = to_msrq(ibsrq);
615         unsigned long flags;
616         int err = 0;
617         int ind;
618         int next_ind;
619         int nreq;
620         int i;
621         void *wqe;
622
623         spin_lock_irqsave(&srq->lock, flags);
624
625         for (nreq = 0; wr; ++nreq, wr = wr->next) {
626                 ind = srq->first_free;
627
628                 if (ind < 0) {
629                         mthca_err(dev, "SRQ %06x full\n", srq->srqn);
630                         err = -ENOMEM;
631                         *bad_wr = wr;
632                         break;
633                 }
634
635                 wqe       = get_wqe(srq, ind);
636                 next_ind  = *wqe_to_link(wqe);
637
638                 if (next_ind < 0) {
639                         mthca_err(dev, "SRQ %06x full\n", srq->srqn);
640                         err = -ENOMEM;
641                         *bad_wr = wr;
642                         break;
643                 }
644
645                 ((struct mthca_next_seg *) wqe)->nda_op =
646                         cpu_to_be32((next_ind << srq->wqe_shift) | 1);
647                 ((struct mthca_next_seg *) wqe)->ee_nds = 0;
648                 /* flags field will always remain 0 */
649
650                 wqe += sizeof (struct mthca_next_seg);
651
652                 if (unlikely(wr->num_sge > srq->max_gs)) {
653                         err = -EINVAL;
654                         *bad_wr = wr;
655                         break;
656                 }
657
658                 for (i = 0; i < wr->num_sge; ++i) {
659                         ((struct mthca_data_seg *) wqe)->byte_count =
660                                 cpu_to_be32(wr->sg_list[i].length);
661                         ((struct mthca_data_seg *) wqe)->lkey =
662                                 cpu_to_be32(wr->sg_list[i].lkey);
663                         ((struct mthca_data_seg *) wqe)->addr =
664                                 cpu_to_be64(wr->sg_list[i].addr);
665                         wqe += sizeof (struct mthca_data_seg);
666                 }
667
668                 if (i < srq->max_gs) {
669                         ((struct mthca_data_seg *) wqe)->byte_count = 0;
670                         ((struct mthca_data_seg *) wqe)->lkey = cpu_to_be32(MTHCA_INVAL_LKEY);
671                         ((struct mthca_data_seg *) wqe)->addr = 0;
672                 }
673
674                 srq->wrid[ind]  = wr->wr_id;
675                 srq->first_free = next_ind;
676         }
677
678         if (likely(nreq)) {
679                 srq->counter += nreq;
680
681                 /*
682                  * Make sure that descriptors are written before
683                  * we write doorbell record.
684                  */
685                 wmb();
686                 *srq->db = cpu_to_be32(srq->counter);
687         }
688
689         spin_unlock_irqrestore(&srq->lock, flags);
690         return err;
691 }
692
693 int mthca_max_srq_sge(struct mthca_dev *dev)
694 {
695         if (mthca_is_memfree(dev))
696                 return dev->limits.max_sg;
697
698         /*
699          * SRQ allocations are based on powers of 2 for Tavor,
700          * (although they only need to be multiples of 16 bytes).
701          *
702          * Therefore, we need to base the max number of sg entries on
703          * the largest power of 2 descriptor size that is <= to the
704          * actual max WQE descriptor size, rather than return the
705          * max_sg value given by the firmware (which is based on WQE
706          * sizes as multiples of 16, not powers of 2).
707          *
708          * If SRQ implementation is changed for Tavor to be based on
709          * multiples of 16, the calculation below can be deleted and
710          * the FW max_sg value returned.
711          */
712         return min_t(int, dev->limits.max_sg,
713                      ((1 << (fls(dev->limits.max_desc_sz) - 1)) -
714                       sizeof (struct mthca_next_seg)) /
715                      sizeof (struct mthca_data_seg));
716 }
717
718 int __devinit mthca_init_srq_table(struct mthca_dev *dev)
719 {
720         int err;
721
722         if (!(dev->mthca_flags & MTHCA_FLAG_SRQ))
723                 return 0;
724
725         spin_lock_init(&dev->srq_table.lock);
726
727         err = mthca_alloc_init(&dev->srq_table.alloc,
728                                dev->limits.num_srqs,
729                                dev->limits.num_srqs - 1,
730                                dev->limits.reserved_srqs);
731         if (err)
732                 return err;
733
734         err = mthca_array_init(&dev->srq_table.srq,
735                                dev->limits.num_srqs);
736         if (err)
737                 mthca_alloc_cleanup(&dev->srq_table.alloc);
738
739         return err;
740 }
741
742 void mthca_cleanup_srq_table(struct mthca_dev *dev)
743 {
744         if (!(dev->mthca_flags & MTHCA_FLAG_SRQ))
745                 return;
746
747         mthca_array_cleanup(&dev->srq_table.srq, dev->limits.num_srqs);
748         mthca_alloc_cleanup(&dev->srq_table.alloc);
749 }