Merge branch 'release' of git://lm-sensors.org/kernel/mhoffman/hwmon-2.6
[linux-2.6] / drivers / infiniband / hw / mlx4 / srq.c
1 /*
2  * Copyright (c) 2007 Cisco Systems, Inc. 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
33 #include <linux/mlx4/qp.h>
34 #include <linux/mlx4/srq.h>
35
36 #include "mlx4_ib.h"
37 #include "user.h"
38
39 static void *get_wqe(struct mlx4_ib_srq *srq, int n)
40 {
41         int offset = n << srq->msrq.wqe_shift;
42
43         if (srq->buf.nbufs == 1)
44                 return srq->buf.u.direct.buf + offset;
45         else
46                 return srq->buf.u.page_list[offset >> PAGE_SHIFT].buf +
47                         (offset & (PAGE_SIZE - 1));
48 }
49
50 static void mlx4_ib_srq_event(struct mlx4_srq *srq, enum mlx4_event type)
51 {
52         struct ib_event event;
53         struct ib_srq *ibsrq = &to_mibsrq(srq)->ibsrq;
54
55         if (ibsrq->event_handler) {
56                 event.device      = ibsrq->device;
57                 event.element.srq = ibsrq;
58                 switch (type) {
59                 case MLX4_EVENT_TYPE_SRQ_LIMIT:
60                         event.event = IB_EVENT_SRQ_LIMIT_REACHED;
61                         break;
62                 case MLX4_EVENT_TYPE_SRQ_CATAS_ERROR:
63                         event.event = IB_EVENT_SRQ_ERR;
64                         break;
65                 default:
66                         printk(KERN_WARNING "mlx4_ib: Unexpected event type %d "
67                                "on SRQ %06x\n", type, srq->srqn);
68                         return;
69                 }
70
71                 ibsrq->event_handler(&event, ibsrq->srq_context);
72         }
73 }
74
75 struct ib_srq *mlx4_ib_create_srq(struct ib_pd *pd,
76                                   struct ib_srq_init_attr *init_attr,
77                                   struct ib_udata *udata)
78 {
79         struct mlx4_ib_dev *dev = to_mdev(pd->device);
80         struct mlx4_ib_srq *srq;
81         struct mlx4_wqe_srq_next_seg *next;
82         int desc_size;
83         int buf_size;
84         int err;
85         int i;
86
87         /* Sanity check SRQ size before proceeding */
88         if (init_attr->attr.max_wr  >= dev->dev->caps.max_srq_wqes ||
89             init_attr->attr.max_sge >  dev->dev->caps.max_srq_sge)
90                 return ERR_PTR(-EINVAL);
91
92         srq = kmalloc(sizeof *srq, GFP_KERNEL);
93         if (!srq)
94                 return ERR_PTR(-ENOMEM);
95
96         mutex_init(&srq->mutex);
97         spin_lock_init(&srq->lock);
98         srq->msrq.max    = roundup_pow_of_two(init_attr->attr.max_wr + 1);
99         srq->msrq.max_gs = init_attr->attr.max_sge;
100
101         desc_size = max(32UL,
102                         roundup_pow_of_two(sizeof (struct mlx4_wqe_srq_next_seg) +
103                                            srq->msrq.max_gs *
104                                            sizeof (struct mlx4_wqe_data_seg)));
105         srq->msrq.wqe_shift = ilog2(desc_size);
106
107         buf_size = srq->msrq.max * desc_size;
108
109         if (pd->uobject) {
110                 struct mlx4_ib_create_srq ucmd;
111
112                 if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd)) {
113                         err = -EFAULT;
114                         goto err_srq;
115                 }
116
117                 srq->umem = ib_umem_get(pd->uobject->context, ucmd.buf_addr,
118                                         buf_size, 0);
119                 if (IS_ERR(srq->umem)) {
120                         err = PTR_ERR(srq->umem);
121                         goto err_srq;
122                 }
123
124                 err = mlx4_mtt_init(dev->dev, ib_umem_page_count(srq->umem),
125                                     ilog2(srq->umem->page_size), &srq->mtt);
126                 if (err)
127                         goto err_buf;
128
129                 err = mlx4_ib_umem_write_mtt(dev, &srq->mtt, srq->umem);
130                 if (err)
131                         goto err_mtt;
132
133                 err = mlx4_ib_db_map_user(to_mucontext(pd->uobject->context),
134                                           ucmd.db_addr, &srq->db);
135                 if (err)
136                         goto err_mtt;
137         } else {
138                 err = mlx4_ib_db_alloc(dev, &srq->db, 0);
139                 if (err)
140                         goto err_srq;
141
142                 *srq->db.db = 0;
143
144                 if (mlx4_buf_alloc(dev->dev, buf_size, PAGE_SIZE * 2, &srq->buf)) {
145                         err = -ENOMEM;
146                         goto err_db;
147                 }
148
149                 srq->head    = 0;
150                 srq->tail    = srq->msrq.max - 1;
151                 srq->wqe_ctr = 0;
152
153                 for (i = 0; i < srq->msrq.max; ++i) {
154                         next = get_wqe(srq, i);
155                         next->next_wqe_index =
156                                 cpu_to_be16((i + 1) & (srq->msrq.max - 1));
157                 }
158
159                 err = mlx4_mtt_init(dev->dev, srq->buf.npages, srq->buf.page_shift,
160                                     &srq->mtt);
161                 if (err)
162                         goto err_buf;
163
164                 err = mlx4_buf_write_mtt(dev->dev, &srq->mtt, &srq->buf);
165                 if (err)
166                         goto err_mtt;
167
168                 srq->wrid = kmalloc(srq->msrq.max * sizeof (u64), GFP_KERNEL);
169                 if (!srq->wrid) {
170                         err = -ENOMEM;
171                         goto err_mtt;
172                 }
173         }
174
175         err = mlx4_srq_alloc(dev->dev, to_mpd(pd)->pdn, &srq->mtt,
176                              srq->db.dma, &srq->msrq);
177         if (err)
178                 goto err_wrid;
179
180         srq->msrq.event = mlx4_ib_srq_event;
181
182         if (pd->uobject)
183                 if (ib_copy_to_udata(udata, &srq->msrq.srqn, sizeof (__u32))) {
184                         err = -EFAULT;
185                         goto err_wrid;
186                 }
187
188         init_attr->attr.max_wr = srq->msrq.max - 1;
189
190         return &srq->ibsrq;
191
192 err_wrid:
193         if (pd->uobject)
194                 mlx4_ib_db_unmap_user(to_mucontext(pd->uobject->context), &srq->db);
195         else
196                 kfree(srq->wrid);
197
198 err_mtt:
199         mlx4_mtt_cleanup(dev->dev, &srq->mtt);
200
201 err_buf:
202         if (pd->uobject)
203                 ib_umem_release(srq->umem);
204         else
205                 mlx4_buf_free(dev->dev, buf_size, &srq->buf);
206
207 err_db:
208         if (!pd->uobject)
209                 mlx4_ib_db_free(dev, &srq->db);
210
211 err_srq:
212         kfree(srq);
213
214         return ERR_PTR(err);
215 }
216
217 int mlx4_ib_modify_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr,
218                        enum ib_srq_attr_mask attr_mask, struct ib_udata *udata)
219 {
220         struct mlx4_ib_dev *dev = to_mdev(ibsrq->device);
221         struct mlx4_ib_srq *srq = to_msrq(ibsrq);
222         int ret;
223
224         /* We don't support resizing SRQs (yet?) */
225         if (attr_mask & IB_SRQ_MAX_WR)
226                 return -EINVAL;
227
228         if (attr_mask & IB_SRQ_LIMIT) {
229                 if (attr->srq_limit >= srq->msrq.max)
230                         return -EINVAL;
231
232                 mutex_lock(&srq->mutex);
233                 ret = mlx4_srq_arm(dev->dev, &srq->msrq, attr->srq_limit);
234                 mutex_unlock(&srq->mutex);
235
236                 if (ret)
237                         return ret;
238         }
239
240         return 0;
241 }
242
243 int mlx4_ib_query_srq(struct ib_srq *ibsrq, struct ib_srq_attr *srq_attr)
244 {
245         struct mlx4_ib_dev *dev = to_mdev(ibsrq->device);
246         struct mlx4_ib_srq *srq = to_msrq(ibsrq);
247         int ret;
248         int limit_watermark;
249
250         ret = mlx4_srq_query(dev->dev, &srq->msrq, &limit_watermark);
251         if (ret)
252                 return ret;
253
254         srq_attr->srq_limit = be16_to_cpu(limit_watermark);
255         srq_attr->max_wr    = srq->msrq.max - 1;
256         srq_attr->max_sge   = srq->msrq.max_gs;
257
258         return 0;
259 }
260
261 int mlx4_ib_destroy_srq(struct ib_srq *srq)
262 {
263         struct mlx4_ib_dev *dev = to_mdev(srq->device);
264         struct mlx4_ib_srq *msrq = to_msrq(srq);
265
266         mlx4_srq_free(dev->dev, &msrq->msrq);
267         mlx4_mtt_cleanup(dev->dev, &msrq->mtt);
268
269         if (srq->uobject) {
270                 mlx4_ib_db_unmap_user(to_mucontext(srq->uobject->context), &msrq->db);
271                 ib_umem_release(msrq->umem);
272         } else {
273                 kfree(msrq->wrid);
274                 mlx4_buf_free(dev->dev, msrq->msrq.max << msrq->msrq.wqe_shift,
275                               &msrq->buf);
276                 mlx4_ib_db_free(dev, &msrq->db);
277         }
278
279         kfree(msrq);
280
281         return 0;
282 }
283
284 void mlx4_ib_free_srq_wqe(struct mlx4_ib_srq *srq, int wqe_index)
285 {
286         struct mlx4_wqe_srq_next_seg *next;
287
288         /* always called with interrupts disabled. */
289         spin_lock(&srq->lock);
290
291         next = get_wqe(srq, srq->tail);
292         next->next_wqe_index = cpu_to_be16(wqe_index);
293         srq->tail = wqe_index;
294
295         spin_unlock(&srq->lock);
296 }
297
298 int mlx4_ib_post_srq_recv(struct ib_srq *ibsrq, struct ib_recv_wr *wr,
299                           struct ib_recv_wr **bad_wr)
300 {
301         struct mlx4_ib_srq *srq = to_msrq(ibsrq);
302         struct mlx4_wqe_srq_next_seg *next;
303         struct mlx4_wqe_data_seg *scat;
304         unsigned long flags;
305         int err = 0;
306         int nreq;
307         int i;
308
309         spin_lock_irqsave(&srq->lock, flags);
310
311         for (nreq = 0; wr; ++nreq, wr = wr->next) {
312                 if (unlikely(wr->num_sge > srq->msrq.max_gs)) {
313                         err = -EINVAL;
314                         *bad_wr = wr;
315                         break;
316                 }
317
318                 if (unlikely(srq->head == srq->tail)) {
319                         err = -ENOMEM;
320                         *bad_wr = wr;
321                         break;
322                 }
323
324                 srq->wrid[srq->head] = wr->wr_id;
325
326                 next      = get_wqe(srq, srq->head);
327                 srq->head = be16_to_cpu(next->next_wqe_index);
328                 scat      = (struct mlx4_wqe_data_seg *) (next + 1);
329
330                 for (i = 0; i < wr->num_sge; ++i) {
331                         scat[i].byte_count = cpu_to_be32(wr->sg_list[i].length);
332                         scat[i].lkey       = cpu_to_be32(wr->sg_list[i].lkey);
333                         scat[i].addr       = cpu_to_be64(wr->sg_list[i].addr);
334                 }
335
336                 if (i < srq->msrq.max_gs) {
337                         scat[i].byte_count = 0;
338                         scat[i].lkey       = cpu_to_be32(MLX4_INVALID_LKEY);
339                         scat[i].addr       = 0;
340                 }
341         }
342
343         if (likely(nreq)) {
344                 srq->wqe_ctr += nreq;
345
346                 /*
347                  * Make sure that descriptors are written before
348                  * doorbell record.
349                  */
350                 wmb();
351
352                 *srq->db.db = cpu_to_be32(srq->wqe_ctr);
353         }
354
355         spin_unlock_irqrestore(&srq->lock, flags);
356
357         return err;
358 }