xref: /openbmc/linux/drivers/infiniband/hw/mlx4/srq.c (revision 799a545b)
1 /*
2  * Copyright (c) 2007 Cisco Systems, Inc. All rights reserved.
3  * Copyright (c) 2007, 2008 Mellanox Technologies. 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/mlx4/qp.h>
35 #include <linux/mlx4/srq.h>
36 #include <linux/slab.h>
37 #include <linux/vmalloc.h>
38 
39 #include "mlx4_ib.h"
40 #include "user.h"
41 
42 static void *get_wqe(struct mlx4_ib_srq *srq, int n)
43 {
44 	return mlx4_buf_offset(&srq->buf, n << srq->msrq.wqe_shift);
45 }
46 
47 static void mlx4_ib_srq_event(struct mlx4_srq *srq, enum mlx4_event type)
48 {
49 	struct ib_event event;
50 	struct ib_srq *ibsrq = &to_mibsrq(srq)->ibsrq;
51 
52 	if (ibsrq->event_handler) {
53 		event.device      = ibsrq->device;
54 		event.element.srq = ibsrq;
55 		switch (type) {
56 		case MLX4_EVENT_TYPE_SRQ_LIMIT:
57 			event.event = IB_EVENT_SRQ_LIMIT_REACHED;
58 			break;
59 		case MLX4_EVENT_TYPE_SRQ_CATAS_ERROR:
60 			event.event = IB_EVENT_SRQ_ERR;
61 			break;
62 		default:
63 			pr_warn("Unexpected event type %d "
64 			       "on SRQ %06x\n", type, srq->srqn);
65 			return;
66 		}
67 
68 		ibsrq->event_handler(&event, ibsrq->srq_context);
69 	}
70 }
71 
72 struct ib_srq *mlx4_ib_create_srq(struct ib_pd *pd,
73 				  struct ib_srq_init_attr *init_attr,
74 				  struct ib_udata *udata)
75 {
76 	struct mlx4_ib_dev *dev = to_mdev(pd->device);
77 	struct mlx4_ib_srq *srq;
78 	struct mlx4_wqe_srq_next_seg *next;
79 	struct mlx4_wqe_data_seg *scatter;
80 	u32 cqn;
81 	u16 xrcdn;
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, 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_db_alloc(dev->dev, &srq->db, 0, GFP_KERNEL);
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 				   GFP_KERNEL)) {
146 			err = -ENOMEM;
147 			goto err_db;
148 		}
149 
150 		srq->head    = 0;
151 		srq->tail    = srq->msrq.max - 1;
152 		srq->wqe_ctr = 0;
153 
154 		for (i = 0; i < srq->msrq.max; ++i) {
155 			next = get_wqe(srq, i);
156 			next->next_wqe_index =
157 				cpu_to_be16((i + 1) & (srq->msrq.max - 1));
158 
159 			for (scatter = (void *) (next + 1);
160 			     (void *) scatter < (void *) next + desc_size;
161 			     ++scatter)
162 				scatter->lkey = cpu_to_be32(MLX4_INVALID_LKEY);
163 		}
164 
165 		err = mlx4_mtt_init(dev->dev, srq->buf.npages, srq->buf.page_shift,
166 				    &srq->mtt);
167 		if (err)
168 			goto err_buf;
169 
170 		err = mlx4_buf_write_mtt(dev->dev, &srq->mtt, &srq->buf, GFP_KERNEL);
171 		if (err)
172 			goto err_mtt;
173 
174 		srq->wrid = kmalloc_array(srq->msrq.max, sizeof(u64),
175 					GFP_KERNEL | __GFP_NOWARN);
176 		if (!srq->wrid) {
177 			srq->wrid = __vmalloc(srq->msrq.max * sizeof(u64),
178 					      GFP_KERNEL, PAGE_KERNEL);
179 			if (!srq->wrid) {
180 				err = -ENOMEM;
181 				goto err_mtt;
182 			}
183 		}
184 	}
185 
186 	cqn = (init_attr->srq_type == IB_SRQT_XRC) ?
187 		to_mcq(init_attr->ext.xrc.cq)->mcq.cqn : 0;
188 	xrcdn = (init_attr->srq_type == IB_SRQT_XRC) ?
189 		to_mxrcd(init_attr->ext.xrc.xrcd)->xrcdn :
190 		(u16) dev->dev->caps.reserved_xrcds;
191 	err = mlx4_srq_alloc(dev->dev, to_mpd(pd)->pdn, cqn, xrcdn, &srq->mtt,
192 			     srq->db.dma, &srq->msrq);
193 	if (err)
194 		goto err_wrid;
195 
196 	srq->msrq.event = mlx4_ib_srq_event;
197 	srq->ibsrq.ext.xrc.srq_num = srq->msrq.srqn;
198 
199 	if (pd->uobject)
200 		if (ib_copy_to_udata(udata, &srq->msrq.srqn, sizeof (__u32))) {
201 			err = -EFAULT;
202 			goto err_wrid;
203 		}
204 
205 	init_attr->attr.max_wr = srq->msrq.max - 1;
206 
207 	return &srq->ibsrq;
208 
209 err_wrid:
210 	if (pd->uobject)
211 		mlx4_ib_db_unmap_user(to_mucontext(pd->uobject->context), &srq->db);
212 	else
213 		kvfree(srq->wrid);
214 
215 err_mtt:
216 	mlx4_mtt_cleanup(dev->dev, &srq->mtt);
217 
218 err_buf:
219 	if (pd->uobject)
220 		ib_umem_release(srq->umem);
221 	else
222 		mlx4_buf_free(dev->dev, buf_size, &srq->buf);
223 
224 err_db:
225 	if (!pd->uobject)
226 		mlx4_db_free(dev->dev, &srq->db);
227 
228 err_srq:
229 	kfree(srq);
230 
231 	return ERR_PTR(err);
232 }
233 
234 int mlx4_ib_modify_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr,
235 		       enum ib_srq_attr_mask attr_mask, struct ib_udata *udata)
236 {
237 	struct mlx4_ib_dev *dev = to_mdev(ibsrq->device);
238 	struct mlx4_ib_srq *srq = to_msrq(ibsrq);
239 	int ret;
240 
241 	/* We don't support resizing SRQs (yet?) */
242 	if (attr_mask & IB_SRQ_MAX_WR)
243 		return -EINVAL;
244 
245 	if (attr_mask & IB_SRQ_LIMIT) {
246 		if (attr->srq_limit >= srq->msrq.max)
247 			return -EINVAL;
248 
249 		mutex_lock(&srq->mutex);
250 		ret = mlx4_srq_arm(dev->dev, &srq->msrq, attr->srq_limit);
251 		mutex_unlock(&srq->mutex);
252 
253 		if (ret)
254 			return ret;
255 	}
256 
257 	return 0;
258 }
259 
260 int mlx4_ib_query_srq(struct ib_srq *ibsrq, struct ib_srq_attr *srq_attr)
261 {
262 	struct mlx4_ib_dev *dev = to_mdev(ibsrq->device);
263 	struct mlx4_ib_srq *srq = to_msrq(ibsrq);
264 	int ret;
265 	int limit_watermark;
266 
267 	ret = mlx4_srq_query(dev->dev, &srq->msrq, &limit_watermark);
268 	if (ret)
269 		return ret;
270 
271 	srq_attr->srq_limit = limit_watermark;
272 	srq_attr->max_wr    = srq->msrq.max - 1;
273 	srq_attr->max_sge   = srq->msrq.max_gs;
274 
275 	return 0;
276 }
277 
278 int mlx4_ib_destroy_srq(struct ib_srq *srq)
279 {
280 	struct mlx4_ib_dev *dev = to_mdev(srq->device);
281 	struct mlx4_ib_srq *msrq = to_msrq(srq);
282 
283 	mlx4_srq_free(dev->dev, &msrq->msrq);
284 	mlx4_mtt_cleanup(dev->dev, &msrq->mtt);
285 
286 	if (srq->uobject) {
287 		mlx4_ib_db_unmap_user(to_mucontext(srq->uobject->context), &msrq->db);
288 		ib_umem_release(msrq->umem);
289 	} else {
290 		kvfree(msrq->wrid);
291 		mlx4_buf_free(dev->dev, msrq->msrq.max << msrq->msrq.wqe_shift,
292 			      &msrq->buf);
293 		mlx4_db_free(dev->dev, &msrq->db);
294 	}
295 
296 	kfree(msrq);
297 
298 	return 0;
299 }
300 
301 void mlx4_ib_free_srq_wqe(struct mlx4_ib_srq *srq, int wqe_index)
302 {
303 	struct mlx4_wqe_srq_next_seg *next;
304 
305 	/* always called with interrupts disabled. */
306 	spin_lock(&srq->lock);
307 
308 	next = get_wqe(srq, srq->tail);
309 	next->next_wqe_index = cpu_to_be16(wqe_index);
310 	srq->tail = wqe_index;
311 
312 	spin_unlock(&srq->lock);
313 }
314 
315 int mlx4_ib_post_srq_recv(struct ib_srq *ibsrq, struct ib_recv_wr *wr,
316 			  struct ib_recv_wr **bad_wr)
317 {
318 	struct mlx4_ib_srq *srq = to_msrq(ibsrq);
319 	struct mlx4_wqe_srq_next_seg *next;
320 	struct mlx4_wqe_data_seg *scat;
321 	unsigned long flags;
322 	int err = 0;
323 	int nreq;
324 	int i;
325 	struct mlx4_ib_dev *mdev = to_mdev(ibsrq->device);
326 
327 	spin_lock_irqsave(&srq->lock, flags);
328 	if (mdev->dev->persist->state & MLX4_DEVICE_STATE_INTERNAL_ERROR) {
329 		err = -EIO;
330 		*bad_wr = wr;
331 		nreq = 0;
332 		goto out;
333 	}
334 
335 	for (nreq = 0; wr; ++nreq, wr = wr->next) {
336 		if (unlikely(wr->num_sge > srq->msrq.max_gs)) {
337 			err = -EINVAL;
338 			*bad_wr = wr;
339 			break;
340 		}
341 
342 		if (unlikely(srq->head == srq->tail)) {
343 			err = -ENOMEM;
344 			*bad_wr = wr;
345 			break;
346 		}
347 
348 		srq->wrid[srq->head] = wr->wr_id;
349 
350 		next      = get_wqe(srq, srq->head);
351 		srq->head = be16_to_cpu(next->next_wqe_index);
352 		scat      = (struct mlx4_wqe_data_seg *) (next + 1);
353 
354 		for (i = 0; i < wr->num_sge; ++i) {
355 			scat[i].byte_count = cpu_to_be32(wr->sg_list[i].length);
356 			scat[i].lkey       = cpu_to_be32(wr->sg_list[i].lkey);
357 			scat[i].addr       = cpu_to_be64(wr->sg_list[i].addr);
358 		}
359 
360 		if (i < srq->msrq.max_gs) {
361 			scat[i].byte_count = 0;
362 			scat[i].lkey       = cpu_to_be32(MLX4_INVALID_LKEY);
363 			scat[i].addr       = 0;
364 		}
365 	}
366 
367 	if (likely(nreq)) {
368 		srq->wqe_ctr += nreq;
369 
370 		/*
371 		 * Make sure that descriptors are written before
372 		 * doorbell record.
373 		 */
374 		wmb();
375 
376 		*srq->db.db = cpu_to_be32(srq->wqe_ctr);
377 	}
378 out:
379 
380 	spin_unlock_irqrestore(&srq->lock, flags);
381 
382 	return err;
383 }
384