mlx4_ib_srq.c revision 331769 
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 <dev/mlx4/qp.h>
35#include <dev/mlx4/srq.h>
36#include <linux/slab.h>
37
38#include "mlx4_ib.h"
39#include <rdma/mlx4-abi.h>
40
41static void *get_wqe(struct mlx4_ib_srq *srq, int n)
42{
43	return mlx4_buf_offset(&srq->buf, n << srq->msrq.wqe_shift);
44}
45
46static void mlx4_ib_srq_event(struct mlx4_srq *srq, enum mlx4_event type)
47{
48	struct ib_event event;
49	struct ib_srq *ibsrq = &to_mibsrq(srq)->ibsrq;
50
51	if (ibsrq->event_handler) {
52		event.device      = ibsrq->device;
53		event.element.srq = ibsrq;
54		switch (type) {
55		case MLX4_EVENT_TYPE_SRQ_LIMIT:
56			event.event = IB_EVENT_SRQ_LIMIT_REACHED;
57			break;
58		case MLX4_EVENT_TYPE_SRQ_CATAS_ERROR:
59			event.event = IB_EVENT_SRQ_ERR;
60			break;
61		default:
62			pr_warn("Unexpected event type %d "
63			       "on SRQ %06x\n", type, srq->srqn);
64			return;
65		}
66
67		ibsrq->event_handler(&event, ibsrq->srq_context);
68	}
69}
70
71struct ib_srq *mlx4_ib_create_srq(struct ib_pd *pd,
72				  struct ib_srq_init_attr *init_attr,
73				  struct ib_udata *udata)
74{
75	struct mlx4_ib_dev *dev = to_mdev(pd->device);
76	struct mlx4_ib_srq *srq;
77	struct mlx4_wqe_srq_next_seg *next;
78	struct mlx4_wqe_data_seg *scatter;
79	u32 cqn;
80	u16 xrcdn;
81	int desc_size;
82	int buf_size;
83	int err;
84	int i;
85
86	/* Sanity check SRQ size before proceeding */
87	if (init_attr->attr.max_wr  >= dev->dev->caps.max_srq_wqes ||
88	    init_attr->attr.max_sge >  dev->dev->caps.max_srq_sge)
89		return ERR_PTR(-EINVAL);
90
91	srq = kmalloc(sizeof *srq, GFP_KERNEL);
92	if (!srq)
93		return ERR_PTR(-ENOMEM);
94
95	mutex_init(&srq->mutex);
96	spin_lock_init(&srq->lock);
97	srq->msrq.max    = roundup_pow_of_two(init_attr->attr.max_wr + 1);
98	srq->msrq.max_gs = init_attr->attr.max_sge;
99
100	desc_size = max(32UL,
101			roundup_pow_of_two(sizeof (struct mlx4_wqe_srq_next_seg) +
102					   srq->msrq.max_gs *
103					   sizeof (struct mlx4_wqe_data_seg)));
104	srq->msrq.wqe_shift = ilog2(desc_size);
105
106	buf_size = srq->msrq.max * desc_size;
107
108	if (pd->uobject) {
109		struct mlx4_ib_create_srq ucmd;
110
111		if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd)) {
112			err = -EFAULT;
113			goto err_srq;
114		}
115
116		srq->umem = ib_umem_get(pd->uobject->context, ucmd.buf_addr,
117					buf_size, 0, 0);
118		if (IS_ERR(srq->umem)) {
119			err = PTR_ERR(srq->umem);
120			goto err_srq;
121		}
122
123		err = mlx4_mtt_init(dev->dev, ib_umem_page_count(srq->umem),
124				    ilog2(srq->umem->page_size), &srq->mtt);
125		if (err)
126			goto err_buf;
127
128		err = mlx4_ib_umem_write_mtt(dev, &srq->mtt, srq->umem);
129		if (err)
130			goto err_mtt;
131
132		err = mlx4_ib_db_map_user(to_mucontext(pd->uobject->context),
133					  ucmd.db_addr, &srq->db);
134		if (err)
135			goto err_mtt;
136	} else {
137		err = mlx4_db_alloc(dev->dev, &srq->db, 0, GFP_KERNEL);
138		if (err)
139			goto err_srq;
140
141		*srq->db.db = 0;
142
143		if (mlx4_buf_alloc(dev->dev, buf_size, PAGE_SIZE * 2, &srq->buf,
144				   GFP_KERNEL)) {
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			for (scatter = (void *) (next + 1);
159			     (void *) scatter < (void *) next + desc_size;
160			     ++scatter)
161				scatter->lkey = cpu_to_be32(MLX4_INVALID_LKEY);
162		}
163
164		err = mlx4_mtt_init(dev->dev, srq->buf.npages, srq->buf.page_shift,
165				    &srq->mtt);
166		if (err)
167			goto err_buf;
168
169		err = mlx4_buf_write_mtt(dev->dev, &srq->mtt, &srq->buf, GFP_KERNEL);
170		if (err)
171			goto err_mtt;
172
173		srq->wrid = kmalloc_array(srq->msrq.max, sizeof(u64),
174					GFP_KERNEL | __GFP_NOWARN);
175		if (!srq->wrid) {
176			srq->wrid = __vmalloc(srq->msrq.max * sizeof(u64),
177					      GFP_KERNEL, 0 /*PAGE_KERNEL*/);
178			if (!srq->wrid) {
179				err = -ENOMEM;
180				goto err_mtt;
181			}
182		}
183	}
184
185	cqn = (init_attr->srq_type == IB_SRQT_XRC) ?
186		to_mcq(init_attr->ext.xrc.cq)->mcq.cqn : 0;
187	xrcdn = (init_attr->srq_type == IB_SRQT_XRC) ?
188		to_mxrcd(init_attr->ext.xrc.xrcd)->xrcdn :
189		(u16) dev->dev->caps.reserved_xrcds;
190	err = mlx4_srq_alloc(dev->dev, to_mpd(pd)->pdn, cqn, xrcdn, &srq->mtt,
191			     srq->db.dma, &srq->msrq);
192	if (err)
193		goto err_wrid;
194
195	srq->msrq.event = mlx4_ib_srq_event;
196	srq->ibsrq.ext.xrc.srq_num = srq->msrq.srqn;
197
198	if (pd->uobject)
199		if (ib_copy_to_udata(udata, &srq->msrq.srqn, sizeof (__u32))) {
200			err = -EFAULT;
201			goto err_wrid;
202		}
203
204	init_attr->attr.max_wr = srq->msrq.max - 1;
205
206	return &srq->ibsrq;
207
208err_wrid:
209	if (pd->uobject)
210		mlx4_ib_db_unmap_user(to_mucontext(pd->uobject->context), &srq->db);
211	else
212		kvfree(srq->wrid);
213
214err_mtt:
215	mlx4_mtt_cleanup(dev->dev, &srq->mtt);
216
217err_buf:
218	if (pd->uobject)
219		ib_umem_release(srq->umem);
220	else
221		mlx4_buf_free(dev->dev, buf_size, &srq->buf);
222
223err_db:
224	if (!pd->uobject)
225		mlx4_db_free(dev->dev, &srq->db);
226
227err_srq:
228	kfree(srq);
229
230	return ERR_PTR(err);
231}
232
233int mlx4_ib_modify_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr,
234		       enum ib_srq_attr_mask attr_mask, struct ib_udata *udata)
235{
236	struct mlx4_ib_dev *dev = to_mdev(ibsrq->device);
237	struct mlx4_ib_srq *srq = to_msrq(ibsrq);
238	int ret;
239
240	/* We don't support resizing SRQs (yet?) */
241	if (attr_mask & IB_SRQ_MAX_WR)
242		return -EINVAL;
243
244	if (attr_mask & IB_SRQ_LIMIT) {
245		if (attr->srq_limit >= srq->msrq.max)
246			return -EINVAL;
247
248		mutex_lock(&srq->mutex);
249		ret = mlx4_srq_arm(dev->dev, &srq->msrq, attr->srq_limit);
250		mutex_unlock(&srq->mutex);
251
252		if (ret)
253			return ret;
254	}
255
256	return 0;
257}
258
259int mlx4_ib_query_srq(struct ib_srq *ibsrq, struct ib_srq_attr *srq_attr)
260{
261	struct mlx4_ib_dev *dev = to_mdev(ibsrq->device);
262	struct mlx4_ib_srq *srq = to_msrq(ibsrq);
263	int ret;
264	int limit_watermark;
265
266	ret = mlx4_srq_query(dev->dev, &srq->msrq, &limit_watermark);
267	if (ret)
268		return ret;
269
270	srq_attr->srq_limit = limit_watermark;
271	srq_attr->max_wr    = srq->msrq.max - 1;
272	srq_attr->max_sge   = srq->msrq.max_gs;
273
274	return 0;
275}
276
277int mlx4_ib_destroy_srq(struct ib_srq *srq)
278{
279	struct mlx4_ib_dev *dev = to_mdev(srq->device);
280	struct mlx4_ib_srq *msrq = to_msrq(srq);
281
282	mlx4_srq_free(dev->dev, &msrq->msrq);
283	mlx4_mtt_cleanup(dev->dev, &msrq->mtt);
284
285	if (srq->uobject) {
286		mlx4_ib_db_unmap_user(to_mucontext(srq->uobject->context), &msrq->db);
287		ib_umem_release(msrq->umem);
288	} else {
289		kvfree(msrq->wrid);
290		mlx4_buf_free(dev->dev, msrq->msrq.max << msrq->msrq.wqe_shift,
291			      &msrq->buf);
292		mlx4_db_free(dev->dev, &msrq->db);
293	}
294
295	kfree(msrq);
296
297	return 0;
298}
299
300void mlx4_ib_free_srq_wqe(struct mlx4_ib_srq *srq, int wqe_index)
301{
302	struct mlx4_wqe_srq_next_seg *next;
303
304	/* always called with interrupts disabled. */
305	spin_lock(&srq->lock);
306
307	next = get_wqe(srq, srq->tail);
308	next->next_wqe_index = cpu_to_be16(wqe_index);
309	srq->tail = wqe_index;
310
311	spin_unlock(&srq->lock);
312}
313
314int mlx4_ib_post_srq_recv(struct ib_srq *ibsrq, struct ib_recv_wr *wr,
315			  struct ib_recv_wr **bad_wr)
316{
317	struct mlx4_ib_srq *srq = to_msrq(ibsrq);
318	struct mlx4_wqe_srq_next_seg *next;
319	struct mlx4_wqe_data_seg *scat;
320	unsigned long flags;
321	int err = 0;
322	int nreq;
323	int i;
324	struct mlx4_ib_dev *mdev = to_mdev(ibsrq->device);
325
326	spin_lock_irqsave(&srq->lock, flags);
327	if (mdev->dev->persist->state & MLX4_DEVICE_STATE_INTERNAL_ERROR) {
328		err = -EIO;
329		*bad_wr = wr;
330		nreq = 0;
331		goto out;
332	}
333
334	for (nreq = 0; wr; ++nreq, wr = wr->next) {
335		if (unlikely(wr->num_sge > srq->msrq.max_gs)) {
336			err = -EINVAL;
337			*bad_wr = wr;
338			break;
339		}
340
341		if (unlikely(srq->head == srq->tail)) {
342			err = -ENOMEM;
343			*bad_wr = wr;
344			break;
345		}
346
347		srq->wrid[srq->head] = wr->wr_id;
348
349		next      = get_wqe(srq, srq->head);
350		srq->head = be16_to_cpu(next->next_wqe_index);
351		scat      = (struct mlx4_wqe_data_seg *) (next + 1);
352
353		for (i = 0; i < wr->num_sge; ++i) {
354			scat[i].byte_count = cpu_to_be32(wr->sg_list[i].length);
355			scat[i].lkey       = cpu_to_be32(wr->sg_list[i].lkey);
356			scat[i].addr       = cpu_to_be64(wr->sg_list[i].addr);
357		}
358
359		if (i < srq->msrq.max_gs) {
360			scat[i].byte_count = 0;
361			scat[i].lkey       = cpu_to_be32(MLX4_INVALID_LKEY);
362			scat[i].addr       = 0;
363		}
364	}
365
366	if (likely(nreq)) {
367		srq->wqe_ctr += nreq;
368
369		/*
370		 * Make sure that descriptors are written before
371		 * doorbell record.
372		 */
373		wmb();
374
375		*srq->db.db = cpu_to_be32(srq->wqe_ctr);
376	}
377out:
378
379	spin_unlock_irqrestore(&srq->lock, flags);
380
381	return err;
382}
383