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/slab.h>
36 #include <linux/vmalloc.h>
37
38 #include "ipath_verbs.h"
39
40 /**
41 * ipath_post_srq_receive - post a receive on a shared receive queue
42 * @ibsrq: the SRQ to post the receive on
43 * @wr: the list of work requests to post
44 * @bad_wr: the first WR to cause a problem is put here
45 *
46 * This may be called from interrupt context.
47 */
ipath_post_srq_receive(struct ib_srq * ibsrq,struct ib_recv_wr * wr,struct ib_recv_wr ** bad_wr)48 int ipath_post_srq_receive(struct ib_srq *ibsrq, struct ib_recv_wr *wr,
49 struct ib_recv_wr **bad_wr)
50 {
51 struct ipath_srq *srq = to_isrq(ibsrq);
52 struct ipath_rwq *wq;
53 unsigned long flags;
54 int ret;
55
56 for (; wr; wr = wr->next) {
57 struct ipath_rwqe *wqe;
58 u32 next;
59 int i;
60
61 if ((unsigned) wr->num_sge > srq->rq.max_sge) {
62 *bad_wr = wr;
63 ret = -EINVAL;
64 goto bail;
65 }
66
67 spin_lock_irqsave(&srq->rq.lock, flags);
68 wq = srq->rq.wq;
69 next = wq->head + 1;
70 if (next >= srq->rq.size)
71 next = 0;
72 if (next == wq->tail) {
73 spin_unlock_irqrestore(&srq->rq.lock, flags);
74 *bad_wr = wr;
75 ret = -ENOMEM;
76 goto bail;
77 }
78
79 wqe = get_rwqe_ptr(&srq->rq, wq->head);
80 wqe->wr_id = wr->wr_id;
81 wqe->num_sge = wr->num_sge;
82 for (i = 0; i < wr->num_sge; i++)
83 wqe->sg_list[i] = wr->sg_list[i];
84 /* Make sure queue entry is written before the head index. */
85 smp_wmb();
86 wq->head = next;
87 spin_unlock_irqrestore(&srq->rq.lock, flags);
88 }
89 ret = 0;
90
91 bail:
92 return ret;
93 }
94
95 /**
96 * ipath_create_srq - create a shared receive queue
97 * @ibpd: the protection domain of the SRQ to create
98 * @srq_init_attr: the attributes of the SRQ
99 * @udata: data from libipathverbs when creating a user SRQ
100 */
ipath_create_srq(struct ib_pd * ibpd,struct ib_srq_init_attr * srq_init_attr,struct ib_udata * udata)101 struct ib_srq *ipath_create_srq(struct ib_pd *ibpd,
102 struct ib_srq_init_attr *srq_init_attr,
103 struct ib_udata *udata)
104 {
105 struct ipath_ibdev *dev = to_idev(ibpd->device);
106 struct ipath_srq *srq;
107 u32 sz;
108 struct ib_srq *ret;
109
110 if (srq_init_attr->srq_type != IB_SRQT_BASIC) {
111 ret = ERR_PTR(-ENOSYS);
112 goto done;
113 }
114
115 if (srq_init_attr->attr.max_wr == 0) {
116 ret = ERR_PTR(-EINVAL);
117 goto done;
118 }
119
120 if ((srq_init_attr->attr.max_sge > ib_ipath_max_srq_sges) ||
121 (srq_init_attr->attr.max_wr > ib_ipath_max_srq_wrs)) {
122 ret = ERR_PTR(-EINVAL);
123 goto done;
124 }
125
126 srq = kmalloc(sizeof(*srq), GFP_KERNEL);
127 if (!srq) {
128 ret = ERR_PTR(-ENOMEM);
129 goto done;
130 }
131
132 /*
133 * Need to use vmalloc() if we want to support large #s of entries.
134 */
135 srq->rq.size = srq_init_attr->attr.max_wr + 1;
136 srq->rq.max_sge = srq_init_attr->attr.max_sge;
137 sz = sizeof(struct ib_sge) * srq->rq.max_sge +
138 sizeof(struct ipath_rwqe);
139 srq->rq.wq = vmalloc_user(sizeof(struct ipath_rwq) + srq->rq.size * sz);
140 if (!srq->rq.wq) {
141 ret = ERR_PTR(-ENOMEM);
142 goto bail_srq;
143 }
144
145 /*
146 * Return the address of the RWQ as the offset to mmap.
147 * See ipath_mmap() for details.
148 */
149 if (udata && udata->outlen >= sizeof(__u64)) {
150 int err;
151 u32 s = sizeof(struct ipath_rwq) + srq->rq.size * sz;
152
153 srq->ip =
154 ipath_create_mmap_info(dev, s,
155 ibpd->uobject->context,
156 srq->rq.wq);
157 if (!srq->ip) {
158 ret = ERR_PTR(-ENOMEM);
159 goto bail_wq;
160 }
161
162 err = ib_copy_to_udata(udata, &srq->ip->offset,
163 sizeof(srq->ip->offset));
164 if (err) {
165 ret = ERR_PTR(err);
166 goto bail_ip;
167 }
168 } else
169 srq->ip = NULL;
170
171 /*
172 * ib_create_srq() will initialize srq->ibsrq.
173 */
174 spin_lock_init(&srq->rq.lock);
175 srq->rq.wq->head = 0;
176 srq->rq.wq->tail = 0;
177 srq->limit = srq_init_attr->attr.srq_limit;
178
179 spin_lock(&dev->n_srqs_lock);
180 if (dev->n_srqs_allocated == ib_ipath_max_srqs) {
181 spin_unlock(&dev->n_srqs_lock);
182 ret = ERR_PTR(-ENOMEM);
183 goto bail_ip;
184 }
185
186 dev->n_srqs_allocated++;
187 spin_unlock(&dev->n_srqs_lock);
188
189 if (srq->ip) {
190 spin_lock_irq(&dev->pending_lock);
191 list_add(&srq->ip->pending_mmaps, &dev->pending_mmaps);
192 spin_unlock_irq(&dev->pending_lock);
193 }
194
195 ret = &srq->ibsrq;
196 goto done;
197
198 bail_ip:
199 kfree(srq->ip);
200 bail_wq:
201 vfree(srq->rq.wq);
202 bail_srq:
203 kfree(srq);
204 done:
205 return ret;
206 }
207
208 /**
209 * ipath_modify_srq - modify a shared receive queue
210 * @ibsrq: the SRQ to modify
211 * @attr: the new attributes of the SRQ
212 * @attr_mask: indicates which attributes to modify
213 * @udata: user data for ipathverbs.so
214 */
ipath_modify_srq(struct ib_srq * ibsrq,struct ib_srq_attr * attr,enum ib_srq_attr_mask attr_mask,struct ib_udata * udata)215 int ipath_modify_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr,
216 enum ib_srq_attr_mask attr_mask,
217 struct ib_udata *udata)
218 {
219 struct ipath_srq *srq = to_isrq(ibsrq);
220 struct ipath_rwq *wq;
221 int ret = 0;
222
223 if (attr_mask & IB_SRQ_MAX_WR) {
224 struct ipath_rwq *owq;
225 struct ipath_rwqe *p;
226 u32 sz, size, n, head, tail;
227
228 /* Check that the requested sizes are below the limits. */
229 if ((attr->max_wr > ib_ipath_max_srq_wrs) ||
230 ((attr_mask & IB_SRQ_LIMIT) ?
231 attr->srq_limit : srq->limit) > attr->max_wr) {
232 ret = -EINVAL;
233 goto bail;
234 }
235
236 sz = sizeof(struct ipath_rwqe) +
237 srq->rq.max_sge * sizeof(struct ib_sge);
238 size = attr->max_wr + 1;
239 wq = vmalloc_user(sizeof(struct ipath_rwq) + size * sz);
240 if (!wq) {
241 ret = -ENOMEM;
242 goto bail;
243 }
244
245 /* Check that we can write the offset to mmap. */
246 if (udata && udata->inlen >= sizeof(__u64)) {
247 __u64 offset_addr;
248 __u64 offset = 0;
249
250 ret = ib_copy_from_udata(&offset_addr, udata,
251 sizeof(offset_addr));
252 if (ret)
253 goto bail_free;
254 udata->outbuf =
255 (void __user *) (unsigned long) offset_addr;
256 ret = ib_copy_to_udata(udata, &offset,
257 sizeof(offset));
258 if (ret)
259 goto bail_free;
260 }
261
262 spin_lock_irq(&srq->rq.lock);
263 /*
264 * validate head pointer value and compute
265 * the number of remaining WQEs.
266 */
267 owq = srq->rq.wq;
268 head = owq->head;
269 if (head >= srq->rq.size)
270 head = 0;
271 tail = owq->tail;
272 if (tail >= srq->rq.size)
273 tail = 0;
274 n = head;
275 if (n < tail)
276 n += srq->rq.size - tail;
277 else
278 n -= tail;
279 if (size <= n) {
280 ret = -EINVAL;
281 goto bail_unlock;
282 }
283 n = 0;
284 p = wq->wq;
285 while (tail != head) {
286 struct ipath_rwqe *wqe;
287 int i;
288
289 wqe = get_rwqe_ptr(&srq->rq, tail);
290 p->wr_id = wqe->wr_id;
291 p->num_sge = wqe->num_sge;
292 for (i = 0; i < wqe->num_sge; i++)
293 p->sg_list[i] = wqe->sg_list[i];
294 n++;
295 p = (struct ipath_rwqe *)((char *) p + sz);
296 if (++tail >= srq->rq.size)
297 tail = 0;
298 }
299 srq->rq.wq = wq;
300 srq->rq.size = size;
301 wq->head = n;
302 wq->tail = 0;
303 if (attr_mask & IB_SRQ_LIMIT)
304 srq->limit = attr->srq_limit;
305 spin_unlock_irq(&srq->rq.lock);
306
307 vfree(owq);
308
309 if (srq->ip) {
310 struct ipath_mmap_info *ip = srq->ip;
311 struct ipath_ibdev *dev = to_idev(srq->ibsrq.device);
312 u32 s = sizeof(struct ipath_rwq) + size * sz;
313
314 ipath_update_mmap_info(dev, ip, s, wq);
315
316 /*
317 * Return the offset to mmap.
318 * See ipath_mmap() for details.
319 */
320 if (udata && udata->inlen >= sizeof(__u64)) {
321 ret = ib_copy_to_udata(udata, &ip->offset,
322 sizeof(ip->offset));
323 if (ret)
324 goto bail;
325 }
326
327 spin_lock_irq(&dev->pending_lock);
328 if (list_empty(&ip->pending_mmaps))
329 list_add(&ip->pending_mmaps,
330 &dev->pending_mmaps);
331 spin_unlock_irq(&dev->pending_lock);
332 }
333 } else if (attr_mask & IB_SRQ_LIMIT) {
334 spin_lock_irq(&srq->rq.lock);
335 if (attr->srq_limit >= srq->rq.size)
336 ret = -EINVAL;
337 else
338 srq->limit = attr->srq_limit;
339 spin_unlock_irq(&srq->rq.lock);
340 }
341 goto bail;
342
343 bail_unlock:
344 spin_unlock_irq(&srq->rq.lock);
345 bail_free:
346 vfree(wq);
347 bail:
348 return ret;
349 }
350
ipath_query_srq(struct ib_srq * ibsrq,struct ib_srq_attr * attr)351 int ipath_query_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr)
352 {
353 struct ipath_srq *srq = to_isrq(ibsrq);
354
355 attr->max_wr = srq->rq.size - 1;
356 attr->max_sge = srq->rq.max_sge;
357 attr->srq_limit = srq->limit;
358 return 0;
359 }
360
361 /**
362 * ipath_destroy_srq - destroy a shared receive queue
363 * @ibsrq: the SRQ to destroy
364 */
ipath_destroy_srq(struct ib_srq * ibsrq)365 int ipath_destroy_srq(struct ib_srq *ibsrq)
366 {
367 struct ipath_srq *srq = to_isrq(ibsrq);
368 struct ipath_ibdev *dev = to_idev(ibsrq->device);
369
370 spin_lock(&dev->n_srqs_lock);
371 dev->n_srqs_allocated--;
372 spin_unlock(&dev->n_srqs_lock);
373 if (srq->ip)
374 kref_put(&srq->ip->ref, ipath_release_mmap_info);
375 else
376 vfree(srq->rq.wq);
377 kfree(srq);
378
379 return 0;
380 }
381