Merge branch 'mainline' into upstream-linus
[linux-2.6] / drivers / infiniband / hw / ipath / ipath_srq.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
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 = -ENOMEM;
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                 wq->head = next;
84                 spin_unlock_irqrestore(&srq->rq.lock, flags);
85         }
86         ret = 0;
87
88 bail:
89         return ret;
90 }
91
92 /**
93  * ipath_create_srq - create a shared receive queue
94  * @ibpd: the protection domain of the SRQ to create
95  * @attr: the attributes of the SRQ
96  * @udata: not used by the InfiniPath verbs driver
97  */
98 struct ib_srq *ipath_create_srq(struct ib_pd *ibpd,
99                                 struct ib_srq_init_attr *srq_init_attr,
100                                 struct ib_udata *udata)
101 {
102         struct ipath_ibdev *dev = to_idev(ibpd->device);
103         struct ipath_srq *srq;
104         u32 sz;
105         struct ib_srq *ret;
106
107         if (dev->n_srqs_allocated == ib_ipath_max_srqs) {
108                 ret = ERR_PTR(-ENOMEM);
109                 goto done;
110         }
111
112         if (srq_init_attr->attr.max_wr == 0) {
113                 ret = ERR_PTR(-EINVAL);
114                 goto done;
115         }
116
117         if ((srq_init_attr->attr.max_sge > ib_ipath_max_srq_sges) ||
118             (srq_init_attr->attr.max_wr > ib_ipath_max_srq_wrs)) {
119                 ret = ERR_PTR(-EINVAL);
120                 goto done;
121         }
122
123         srq = kmalloc(sizeof(*srq), GFP_KERNEL);
124         if (!srq) {
125                 ret = ERR_PTR(-ENOMEM);
126                 goto done;
127         }
128
129         /*
130          * Need to use vmalloc() if we want to support large #s of entries.
131          */
132         srq->rq.size = srq_init_attr->attr.max_wr + 1;
133         srq->rq.max_sge = srq_init_attr->attr.max_sge;
134         sz = sizeof(struct ib_sge) * srq->rq.max_sge +
135                 sizeof(struct ipath_rwqe);
136         srq->rq.wq = vmalloc_user(sizeof(struct ipath_rwq) + srq->rq.size * sz);
137         if (!srq->rq.wq) {
138                 ret = ERR_PTR(-ENOMEM);
139                 goto bail_srq;
140         }
141
142         /*
143          * Return the address of the RWQ as the offset to mmap.
144          * See ipath_mmap() for details.
145          */
146         if (udata && udata->outlen >= sizeof(__u64)) {
147                 struct ipath_mmap_info *ip;
148                 __u64 offset = (__u64) srq->rq.wq;
149                 int err;
150
151                 err = ib_copy_to_udata(udata, &offset, sizeof(offset));
152                 if (err) {
153                         ret = ERR_PTR(err);
154                         goto bail_wq;
155                 }
156
157                 /* Allocate info for ipath_mmap(). */
158                 ip = kmalloc(sizeof(*ip), GFP_KERNEL);
159                 if (!ip) {
160                         ret = ERR_PTR(-ENOMEM);
161                         goto bail_wq;
162                 }
163                 srq->ip = ip;
164                 ip->context = ibpd->uobject->context;
165                 ip->obj = srq->rq.wq;
166                 kref_init(&ip->ref);
167                 ip->mmap_cnt = 0;
168                 ip->size = PAGE_ALIGN(sizeof(struct ipath_rwq) +
169                                       srq->rq.size * sz);
170                 spin_lock_irq(&dev->pending_lock);
171                 ip->next = dev->pending_mmaps;
172                 dev->pending_mmaps = ip;
173                 spin_unlock_irq(&dev->pending_lock);
174         } else
175                 srq->ip = NULL;
176
177         /*
178          * ib_create_srq() will initialize srq->ibsrq.
179          */
180         spin_lock_init(&srq->rq.lock);
181         srq->rq.wq->head = 0;
182         srq->rq.wq->tail = 0;
183         srq->rq.max_sge = srq_init_attr->attr.max_sge;
184         srq->limit = srq_init_attr->attr.srq_limit;
185
186         dev->n_srqs_allocated++;
187
188         ret = &srq->ibsrq;
189         goto done;
190
191 bail_wq:
192         vfree(srq->rq.wq);
193
194 bail_srq:
195         kfree(srq);
196
197 done:
198         return ret;
199 }
200
201 /**
202  * ipath_modify_srq - modify a shared receive queue
203  * @ibsrq: the SRQ to modify
204  * @attr: the new attributes of the SRQ
205  * @attr_mask: indicates which attributes to modify
206  * @udata: user data for ipathverbs.so
207  */
208 int ipath_modify_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr,
209                      enum ib_srq_attr_mask attr_mask,
210                      struct ib_udata *udata)
211 {
212         struct ipath_srq *srq = to_isrq(ibsrq);
213         int ret = 0;
214
215         if (attr_mask & IB_SRQ_MAX_WR) {
216                 struct ipath_rwq *owq;
217                 struct ipath_rwq *wq;
218                 struct ipath_rwqe *p;
219                 u32 sz, size, n, head, tail;
220
221                 /* Check that the requested sizes are below the limits. */
222                 if ((attr->max_wr > ib_ipath_max_srq_wrs) ||
223                     ((attr_mask & IB_SRQ_LIMIT) ?
224                      attr->srq_limit : srq->limit) > attr->max_wr) {
225                         ret = -EINVAL;
226                         goto bail;
227                 }
228
229                 sz = sizeof(struct ipath_rwqe) +
230                         srq->rq.max_sge * sizeof(struct ib_sge);
231                 size = attr->max_wr + 1;
232                 wq = vmalloc_user(sizeof(struct ipath_rwq) + size * sz);
233                 if (!wq) {
234                         ret = -ENOMEM;
235                         goto bail;
236                 }
237
238                 /*
239                  * Return the address of the RWQ as the offset to mmap.
240                  * See ipath_mmap() for details.
241                  */
242                 if (udata && udata->inlen >= sizeof(__u64)) {
243                         __u64 offset_addr;
244                         __u64 offset = (__u64) wq;
245
246                         ret = ib_copy_from_udata(&offset_addr, udata,
247                                                  sizeof(offset_addr));
248                         if (ret) {
249                                 vfree(wq);
250                                 goto bail;
251                         }
252                         udata->outbuf = (void __user *) offset_addr;
253                         ret = ib_copy_to_udata(udata, &offset,
254                                                sizeof(offset));
255                         if (ret) {
256                                 vfree(wq);
257                                 goto bail;
258                         }
259                 }
260
261                 spin_lock_irq(&srq->rq.lock);
262                 /*
263                  * validate head pointer value and compute
264                  * the number of remaining WQEs.
265                  */
266                 owq = srq->rq.wq;
267                 head = owq->head;
268                 if (head >= srq->rq.size)
269                         head = 0;
270                 tail = owq->tail;
271                 if (tail >= srq->rq.size)
272                         tail = 0;
273                 n = head;
274                 if (n < tail)
275                         n += srq->rq.size - tail;
276                 else
277                         n -= tail;
278                 if (size <= n) {
279                         spin_unlock_irq(&srq->rq.lock);
280                         vfree(wq);
281                         ret = -EINVAL;
282                         goto bail;
283                 }
284                 n = 0;
285                 p = wq->wq;
286                 while (tail != head) {
287                         struct ipath_rwqe *wqe;
288                         int i;
289
290                         wqe = get_rwqe_ptr(&srq->rq, tail);
291                         p->wr_id = wqe->wr_id;
292                         p->num_sge = wqe->num_sge;
293                         for (i = 0; i < wqe->num_sge; i++)
294                                 p->sg_list[i] = wqe->sg_list[i];
295                         n++;
296                         p = (struct ipath_rwqe *)((char *) p + sz);
297                         if (++tail >= srq->rq.size)
298                                 tail = 0;
299                 }
300                 srq->rq.wq = wq;
301                 srq->rq.size = size;
302                 wq->head = n;
303                 wq->tail = 0;
304                 if (attr_mask & IB_SRQ_LIMIT)
305                         srq->limit = attr->srq_limit;
306                 spin_unlock_irq(&srq->rq.lock);
307
308                 vfree(owq);
309
310                 if (srq->ip) {
311                         struct ipath_mmap_info *ip = srq->ip;
312                         struct ipath_ibdev *dev = to_idev(srq->ibsrq.device);
313
314                         ip->obj = wq;
315                         ip->size = PAGE_ALIGN(sizeof(struct ipath_rwq) +
316                                               size * sz);
317                         spin_lock_irq(&dev->pending_lock);
318                         ip->next = dev->pending_mmaps;
319                         dev->pending_mmaps = ip;
320                         spin_unlock_irq(&dev->pending_lock);
321                 }
322         } else if (attr_mask & IB_SRQ_LIMIT) {
323                 spin_lock_irq(&srq->rq.lock);
324                 if (attr->srq_limit >= srq->rq.size)
325                         ret = -EINVAL;
326                 else
327                         srq->limit = attr->srq_limit;
328                 spin_unlock_irq(&srq->rq.lock);
329         }
330
331 bail:
332         return ret;
333 }
334
335 int ipath_query_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr)
336 {
337         struct ipath_srq *srq = to_isrq(ibsrq);
338
339         attr->max_wr = srq->rq.size - 1;
340         attr->max_sge = srq->rq.max_sge;
341         attr->srq_limit = srq->limit;
342         return 0;
343 }
344
345 /**
346  * ipath_destroy_srq - destroy a shared receive queue
347  * @ibsrq: the SRQ to destroy
348  */
349 int ipath_destroy_srq(struct ib_srq *ibsrq)
350 {
351         struct ipath_srq *srq = to_isrq(ibsrq);
352         struct ipath_ibdev *dev = to_idev(ibsrq->device);
353
354         dev->n_srqs_allocated--;
355         vfree(srq->rq.wq);
356         kfree(srq);
357
358         return 0;
359 }