Merge commit 'v2.6.30-rc1' into core/urgent
[linux-2.6] / drivers / infiniband / hw / ipath / ipath_srq.c
1 /*
2  * Copyright (c) 2006, 2007, 2008 QLogic Corporation. All rights reserved.
3  * Copyright (c) 2005, 2006 PathScale, Inc. All rights reserved.
4  *
5  * This software is available to you under a choice of one of two
6  * licenses.  You may choose to be licensed under the terms of the GNU
7  * General Public License (GPL) Version 2, available from the file
8  * COPYING in the main directory of this source tree, or the
9  * OpenIB.org BSD license below:
10  *
11  *     Redistribution and use in source and binary forms, with or
12  *     without modification, are permitted provided that the following
13  *     conditions are met:
14  *
15  *      - Redistributions of source code must retain the above
16  *        copyright notice, this list of conditions and the following
17  *        disclaimer.
18  *
19  *      - Redistributions in binary form must reproduce the above
20  *        copyright notice, this list of conditions and the following
21  *        disclaimer in the documentation and/or other materials
22  *        provided with the distribution.
23  *
24  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
25  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
26  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
27  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
28  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
29  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
30  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
31  * SOFTWARE.
32  */
33
34 #include <linux/err.h>
35 #include <linux/vmalloc.h>
36
37 #include "ipath_verbs.h"
38
39 /**
40  * ipath_post_srq_receive - post a receive on a shared receive queue
41  * @ibsrq: the SRQ to post the receive on
42  * @wr: the list of work requests to post
43  * @bad_wr: the first WR to cause a problem is put here
44  *
45  * This may be called from interrupt context.
46  */
47 int ipath_post_srq_receive(struct ib_srq *ibsrq, struct ib_recv_wr *wr,
48                            struct ib_recv_wr **bad_wr)
49 {
50         struct ipath_srq *srq = to_isrq(ibsrq);
51         struct ipath_rwq *wq;
52         unsigned long flags;
53         int ret;
54
55         for (; wr; wr = wr->next) {
56                 struct ipath_rwqe *wqe;
57                 u32 next;
58                 int i;
59
60                 if ((unsigned) wr->num_sge > srq->rq.max_sge) {
61                         *bad_wr = wr;
62                         ret = -EINVAL;
63                         goto bail;
64                 }
65
66                 spin_lock_irqsave(&srq->rq.lock, flags);
67                 wq = srq->rq.wq;
68                 next = wq->head + 1;
69                 if (next >= srq->rq.size)
70                         next = 0;
71                 if (next == wq->tail) {
72                         spin_unlock_irqrestore(&srq->rq.lock, flags);
73                         *bad_wr = wr;
74                         ret = -ENOMEM;
75                         goto bail;
76                 }
77
78                 wqe = get_rwqe_ptr(&srq->rq, wq->head);
79                 wqe->wr_id = wr->wr_id;
80                 wqe->num_sge = wr->num_sge;
81                 for (i = 0; i < wr->num_sge; i++)
82                         wqe->sg_list[i] = wr->sg_list[i];
83                 /* Make sure queue entry is written before the head index. */
84                 smp_wmb();
85                 wq->head = next;
86                 spin_unlock_irqrestore(&srq->rq.lock, flags);
87         }
88         ret = 0;
89
90 bail:
91         return ret;
92 }
93
94 /**
95  * ipath_create_srq - create a shared receive queue
96  * @ibpd: the protection domain of the SRQ to create
97  * @srq_init_attr: the attributes of the SRQ
98  * @udata: data from libipathverbs when creating a user SRQ
99  */
100 struct ib_srq *ipath_create_srq(struct ib_pd *ibpd,
101                                 struct ib_srq_init_attr *srq_init_attr,
102                                 struct ib_udata *udata)
103 {
104         struct ipath_ibdev *dev = to_idev(ibpd->device);
105         struct ipath_srq *srq;
106         u32 sz;
107         struct ib_srq *ret;
108
109         if (srq_init_attr->attr.max_wr == 0) {
110                 ret = ERR_PTR(-EINVAL);
111                 goto done;
112         }
113
114         if ((srq_init_attr->attr.max_sge > ib_ipath_max_srq_sges) ||
115             (srq_init_attr->attr.max_wr > ib_ipath_max_srq_wrs)) {
116                 ret = ERR_PTR(-EINVAL);
117                 goto done;
118         }
119
120         srq = kmalloc(sizeof(*srq), GFP_KERNEL);
121         if (!srq) {
122                 ret = ERR_PTR(-ENOMEM);
123                 goto done;
124         }
125
126         /*
127          * Need to use vmalloc() if we want to support large #s of entries.
128          */
129         srq->rq.size = srq_init_attr->attr.max_wr + 1;
130         srq->rq.max_sge = srq_init_attr->attr.max_sge;
131         sz = sizeof(struct ib_sge) * srq->rq.max_sge +
132                 sizeof(struct ipath_rwqe);
133         srq->rq.wq = vmalloc_user(sizeof(struct ipath_rwq) + srq->rq.size * sz);
134         if (!srq->rq.wq) {
135                 ret = ERR_PTR(-ENOMEM);
136                 goto bail_srq;
137         }
138
139         /*
140          * Return the address of the RWQ as the offset to mmap.
141          * See ipath_mmap() for details.
142          */
143         if (udata && udata->outlen >= sizeof(__u64)) {
144                 int err;
145                 u32 s = sizeof(struct ipath_rwq) + srq->rq.size * sz;
146
147                 srq->ip =
148                     ipath_create_mmap_info(dev, s,
149                                            ibpd->uobject->context,
150                                            srq->rq.wq);
151                 if (!srq->ip) {
152                         ret = ERR_PTR(-ENOMEM);
153                         goto bail_wq;
154                 }
155
156                 err = ib_copy_to_udata(udata, &srq->ip->offset,
157                                        sizeof(srq->ip->offset));
158                 if (err) {
159                         ret = ERR_PTR(err);
160                         goto bail_ip;
161                 }
162         } else
163                 srq->ip = NULL;
164
165         /*
166          * ib_create_srq() will initialize srq->ibsrq.
167          */
168         spin_lock_init(&srq->rq.lock);
169         srq->rq.wq->head = 0;
170         srq->rq.wq->tail = 0;
171         srq->limit = srq_init_attr->attr.srq_limit;
172
173         spin_lock(&dev->n_srqs_lock);
174         if (dev->n_srqs_allocated == ib_ipath_max_srqs) {
175                 spin_unlock(&dev->n_srqs_lock);
176                 ret = ERR_PTR(-ENOMEM);
177                 goto bail_ip;
178         }
179
180         dev->n_srqs_allocated++;
181         spin_unlock(&dev->n_srqs_lock);
182
183         if (srq->ip) {
184                 spin_lock_irq(&dev->pending_lock);
185                 list_add(&srq->ip->pending_mmaps, &dev->pending_mmaps);
186                 spin_unlock_irq(&dev->pending_lock);
187         }
188
189         ret = &srq->ibsrq;
190         goto done;
191
192 bail_ip:
193         kfree(srq->ip);
194 bail_wq:
195         vfree(srq->rq.wq);
196 bail_srq:
197         kfree(srq);
198 done:
199         return ret;
200 }
201
202 /**
203  * ipath_modify_srq - modify a shared receive queue
204  * @ibsrq: the SRQ to modify
205  * @attr: the new attributes of the SRQ
206  * @attr_mask: indicates which attributes to modify
207  * @udata: user data for ipathverbs.so
208  */
209 int ipath_modify_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr,
210                      enum ib_srq_attr_mask attr_mask,
211                      struct ib_udata *udata)
212 {
213         struct ipath_srq *srq = to_isrq(ibsrq);
214         struct ipath_rwq *wq;
215         int ret = 0;
216
217         if (attr_mask & IB_SRQ_MAX_WR) {
218                 struct ipath_rwq *owq;
219                 struct ipath_rwqe *p;
220                 u32 sz, size, n, head, tail;
221
222                 /* Check that the requested sizes are below the limits. */
223                 if ((attr->max_wr > ib_ipath_max_srq_wrs) ||
224                     ((attr_mask & IB_SRQ_LIMIT) ?
225                      attr->srq_limit : srq->limit) > attr->max_wr) {
226                         ret = -EINVAL;
227                         goto bail;
228                 }
229
230                 sz = sizeof(struct ipath_rwqe) +
231                         srq->rq.max_sge * sizeof(struct ib_sge);
232                 size = attr->max_wr + 1;
233                 wq = vmalloc_user(sizeof(struct ipath_rwq) + size * sz);
234                 if (!wq) {
235                         ret = -ENOMEM;
236                         goto bail;
237                 }
238
239                 /* Check that we can write the offset to mmap. */
240                 if (udata && udata->inlen >= sizeof(__u64)) {
241                         __u64 offset_addr;
242                         __u64 offset = 0;
243
244                         ret = ib_copy_from_udata(&offset_addr, udata,
245                                                  sizeof(offset_addr));
246                         if (ret)
247                                 goto bail_free;
248                         udata->outbuf =
249                                 (void __user *) (unsigned long) offset_addr;
250                         ret = ib_copy_to_udata(udata, &offset,
251                                                sizeof(offset));
252                         if (ret)
253                                 goto bail_free;
254                 }
255
256                 spin_lock_irq(&srq->rq.lock);
257                 /*
258                  * validate head pointer value and compute
259                  * the number of remaining WQEs.
260                  */
261                 owq = srq->rq.wq;
262                 head = owq->head;
263                 if (head >= srq->rq.size)
264                         head = 0;
265                 tail = owq->tail;
266                 if (tail >= srq->rq.size)
267                         tail = 0;
268                 n = head;
269                 if (n < tail)
270                         n += srq->rq.size - tail;
271                 else
272                         n -= tail;
273                 if (size <= n) {
274                         ret = -EINVAL;
275                         goto bail_unlock;
276                 }
277                 n = 0;
278                 p = wq->wq;
279                 while (tail != head) {
280                         struct ipath_rwqe *wqe;
281                         int i;
282
283                         wqe = get_rwqe_ptr(&srq->rq, tail);
284                         p->wr_id = wqe->wr_id;
285                         p->num_sge = wqe->num_sge;
286                         for (i = 0; i < wqe->num_sge; i++)
287                                 p->sg_list[i] = wqe->sg_list[i];
288                         n++;
289                         p = (struct ipath_rwqe *)((char *) p + sz);
290                         if (++tail >= srq->rq.size)
291                                 tail = 0;
292                 }
293                 srq->rq.wq = wq;
294                 srq->rq.size = size;
295                 wq->head = n;
296                 wq->tail = 0;
297                 if (attr_mask & IB_SRQ_LIMIT)
298                         srq->limit = attr->srq_limit;
299                 spin_unlock_irq(&srq->rq.lock);
300
301                 vfree(owq);
302
303                 if (srq->ip) {
304                         struct ipath_mmap_info *ip = srq->ip;
305                         struct ipath_ibdev *dev = to_idev(srq->ibsrq.device);
306                         u32 s = sizeof(struct ipath_rwq) + size * sz;
307
308                         ipath_update_mmap_info(dev, ip, s, wq);
309
310                         /*
311                          * Return the offset to mmap.
312                          * See ipath_mmap() for details.
313                          */
314                         if (udata && udata->inlen >= sizeof(__u64)) {
315                                 ret = ib_copy_to_udata(udata, &ip->offset,
316                                                        sizeof(ip->offset));
317                                 if (ret)
318                                         goto bail;
319                         }
320
321                         spin_lock_irq(&dev->pending_lock);
322                         if (list_empty(&ip->pending_mmaps))
323                                 list_add(&ip->pending_mmaps,
324                                          &dev->pending_mmaps);
325                         spin_unlock_irq(&dev->pending_lock);
326                 }
327         } else if (attr_mask & IB_SRQ_LIMIT) {
328                 spin_lock_irq(&srq->rq.lock);
329                 if (attr->srq_limit >= srq->rq.size)
330                         ret = -EINVAL;
331                 else
332                         srq->limit = attr->srq_limit;
333                 spin_unlock_irq(&srq->rq.lock);
334         }
335         goto bail;
336
337 bail_unlock:
338         spin_unlock_irq(&srq->rq.lock);
339 bail_free:
340         vfree(wq);
341 bail:
342         return ret;
343 }
344
345 int ipath_query_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr)
346 {
347         struct ipath_srq *srq = to_isrq(ibsrq);
348
349         attr->max_wr = srq->rq.size - 1;
350         attr->max_sge = srq->rq.max_sge;
351         attr->srq_limit = srq->limit;
352         return 0;
353 }
354
355 /**
356  * ipath_destroy_srq - destroy a shared receive queue
357  * @ibsrq: the SRQ to destroy
358  */
359 int ipath_destroy_srq(struct ib_srq *ibsrq)
360 {
361         struct ipath_srq *srq = to_isrq(ibsrq);
362         struct ipath_ibdev *dev = to_idev(ibsrq->device);
363
364         spin_lock(&dev->n_srqs_lock);
365         dev->n_srqs_allocated--;
366         spin_unlock(&dev->n_srqs_lock);
367         if (srq->ip)
368                 kref_put(&srq->ip->ref, ipath_release_mmap_info);
369         else
370                 vfree(srq->rq.wq);
371         kfree(srq);
372
373         return 0;
374 }