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
38#include "mlx4_ib.h"
39#include <rdma/mlx4-abi.h>
40#include <rdma/uverbs_ioctl.h>
41
42static 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
47static 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
72int mlx4_ib_create_srq(struct ib_srq *ib_srq,
73		       struct ib_srq_init_attr *init_attr,
74		       struct ib_udata *udata)
75{
76	struct mlx4_ib_dev *dev = to_mdev(ib_srq->device);
77	struct mlx4_ib_ucontext *ucontext = rdma_udata_to_drv_context(
78		udata, struct mlx4_ib_ucontext, ibucontext);
79	struct mlx4_ib_srq *srq = to_msrq(ib_srq);
80	struct mlx4_wqe_srq_next_seg *next;
81	struct mlx4_wqe_data_seg *scatter;
82	u32 cqn;
83	u16 xrcdn;
84	int desc_size;
85	int buf_size;
86	int err;
87	int i;
88
89	if (init_attr->srq_type != IB_SRQT_BASIC &&
90	    init_attr->srq_type != IB_SRQT_XRC)
91		return -EOPNOTSUPP;
92
93	/* Sanity check SRQ size before proceeding */
94	if (init_attr->attr.max_wr  >= dev->dev->caps.max_srq_wqes ||
95	    init_attr->attr.max_sge >  dev->dev->caps.max_srq_sge)
96		return -EINVAL;
97
98	mutex_init(&srq->mutex);
99	spin_lock_init(&srq->lock);
100	srq->msrq.max    = roundup_pow_of_two(init_attr->attr.max_wr + 1);
101	srq->msrq.max_gs = init_attr->attr.max_sge;
102
103	desc_size = max(32UL,
104			roundup_pow_of_two(sizeof (struct mlx4_wqe_srq_next_seg) +
105					   srq->msrq.max_gs *
106					   sizeof (struct mlx4_wqe_data_seg)));
107	srq->msrq.wqe_shift = ilog2(desc_size);
108
109	buf_size = srq->msrq.max * desc_size;
110
111	if (udata) {
112		struct mlx4_ib_create_srq ucmd;
113
114		if (ib_copy_from_udata(&ucmd, udata, sizeof(ucmd)))
115			return -EFAULT;
116
117		srq->umem =
118			ib_umem_get(ib_srq->device, ucmd.buf_addr, buf_size, 0);
119		if (IS_ERR(srq->umem))
120			return PTR_ERR(srq->umem);
121
122		err = mlx4_mtt_init(
123			dev->dev, ib_umem_num_dma_blocks(srq->umem, PAGE_SIZE),
124			PAGE_SHIFT, &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(udata, ucmd.db_addr, &srq->db);
133		if (err)
134			goto err_mtt;
135	} else {
136		err = mlx4_db_alloc(dev->dev, &srq->db, 0);
137		if (err)
138			return err;
139
140		*srq->db.db = 0;
141
142		if (mlx4_buf_alloc(dev->dev, buf_size, PAGE_SIZE * 2,
143				   &srq->buf)) {
144			err = -ENOMEM;
145			goto err_db;
146		}
147
148		srq->head    = 0;
149		srq->tail    = srq->msrq.max - 1;
150		srq->wqe_ctr = 0;
151
152		for (i = 0; i < srq->msrq.max; ++i) {
153			next = get_wqe(srq, i);
154			next->next_wqe_index =
155				cpu_to_be16((i + 1) & (srq->msrq.max - 1));
156
157			for (scatter = (void *) (next + 1);
158			     (void *) scatter < (void *) next + desc_size;
159			     ++scatter)
160				scatter->lkey = cpu_to_be32(MLX4_INVALID_LKEY);
161		}
162
163		err = mlx4_mtt_init(dev->dev, srq->buf.npages, srq->buf.page_shift,
164				    &srq->mtt);
165		if (err)
166			goto err_buf;
167
168		err = mlx4_buf_write_mtt(dev->dev, &srq->mtt, &srq->buf);
169		if (err)
170			goto err_mtt;
171
172		srq->wrid = kvmalloc_array(srq->msrq.max,
173					   sizeof(u64), GFP_KERNEL);
174		if (!srq->wrid) {
175			err = -ENOMEM;
176			goto err_mtt;
177		}
178	}
179
180	cqn = ib_srq_has_cq(init_attr->srq_type) ?
181		to_mcq(init_attr->ext.cq)->mcq.cqn : 0;
182	xrcdn = (init_attr->srq_type == IB_SRQT_XRC) ?
183		to_mxrcd(init_attr->ext.xrc.xrcd)->xrcdn :
184		(u16) dev->dev->caps.reserved_xrcds;
185	err = mlx4_srq_alloc(dev->dev, to_mpd(ib_srq->pd)->pdn, cqn, xrcdn,
186			     &srq->mtt, srq->db.dma, &srq->msrq);
187	if (err)
188		goto err_wrid;
189
190	srq->msrq.event = mlx4_ib_srq_event;
191	srq->ibsrq.ext.xrc.srq_num = srq->msrq.srqn;
192
193	if (udata)
194		if (ib_copy_to_udata(udata, &srq->msrq.srqn, sizeof (__u32))) {
195			err = -EFAULT;
196			goto err_wrid;
197		}
198
199	init_attr->attr.max_wr = srq->msrq.max - 1;
200
201	return 0;
202
203err_wrid:
204	if (udata)
205		mlx4_ib_db_unmap_user(ucontext, &srq->db);
206	else
207		kvfree(srq->wrid);
208
209err_mtt:
210	mlx4_mtt_cleanup(dev->dev, &srq->mtt);
211
212err_buf:
213	if (!srq->umem)
214		mlx4_buf_free(dev->dev, buf_size, &srq->buf);
215	ib_umem_release(srq->umem);
216
217err_db:
218	if (!udata)
219		mlx4_db_free(dev->dev, &srq->db);
220
221	return err;
222}
223
224int mlx4_ib_modify_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr,
225		       enum ib_srq_attr_mask attr_mask, struct ib_udata *udata)
226{
227	struct mlx4_ib_dev *dev = to_mdev(ibsrq->device);
228	struct mlx4_ib_srq *srq = to_msrq(ibsrq);
229	int ret;
230
231	/* We don't support resizing SRQs (yet?) */
232	if (attr_mask & IB_SRQ_MAX_WR)
233		return -EINVAL;
234
235	if (attr_mask & IB_SRQ_LIMIT) {
236		if (attr->srq_limit >= srq->msrq.max)
237			return -EINVAL;
238
239		mutex_lock(&srq->mutex);
240		ret = mlx4_srq_arm(dev->dev, &srq->msrq, attr->srq_limit);
241		mutex_unlock(&srq->mutex);
242
243		if (ret)
244			return ret;
245	}
246
247	return 0;
248}
249
250int mlx4_ib_query_srq(struct ib_srq *ibsrq, struct ib_srq_attr *srq_attr)
251{
252	struct mlx4_ib_dev *dev = to_mdev(ibsrq->device);
253	struct mlx4_ib_srq *srq = to_msrq(ibsrq);
254	int ret;
255	int limit_watermark;
256
257	ret = mlx4_srq_query(dev->dev, &srq->msrq, &limit_watermark);
258	if (ret)
259		return ret;
260
261	srq_attr->srq_limit = limit_watermark;
262	srq_attr->max_wr    = srq->msrq.max - 1;
263	srq_attr->max_sge   = srq->msrq.max_gs;
264
265	return 0;
266}
267
268int mlx4_ib_destroy_srq(struct ib_srq *srq, struct ib_udata *udata)
269{
270	struct mlx4_ib_dev *dev = to_mdev(srq->device);
271	struct mlx4_ib_srq *msrq = to_msrq(srq);
272
273	mlx4_srq_free(dev->dev, &msrq->msrq);
274	mlx4_mtt_cleanup(dev->dev, &msrq->mtt);
275
276	if (udata) {
277		mlx4_ib_db_unmap_user(
278			rdma_udata_to_drv_context(
279				udata,
280				struct mlx4_ib_ucontext,
281				ibucontext),
282			&msrq->db);
283	} else {
284		kvfree(msrq->wrid);
285		mlx4_buf_free(dev->dev, msrq->msrq.max << msrq->msrq.wqe_shift,
286			      &msrq->buf);
287		mlx4_db_free(dev->dev, &msrq->db);
288	}
289	ib_umem_release(msrq->umem);
290	return 0;
291}
292
293void mlx4_ib_free_srq_wqe(struct mlx4_ib_srq *srq, int wqe_index)
294{
295	struct mlx4_wqe_srq_next_seg *next;
296
297	/* always called with interrupts disabled. */
298	spin_lock(&srq->lock);
299
300	next = get_wqe(srq, srq->tail);
301	next->next_wqe_index = cpu_to_be16(wqe_index);
302	srq->tail = wqe_index;
303
304	spin_unlock(&srq->lock);
305}
306
307int mlx4_ib_post_srq_recv(struct ib_srq *ibsrq, const struct ib_recv_wr *wr,
308			  const struct ib_recv_wr **bad_wr)
309{
310	struct mlx4_ib_srq *srq = to_msrq(ibsrq);
311	struct mlx4_wqe_srq_next_seg *next;
312	struct mlx4_wqe_data_seg *scat;
313	unsigned long flags;
314	int err = 0;
315	int nreq;
316	int i;
317	struct mlx4_ib_dev *mdev = to_mdev(ibsrq->device);
318
319	spin_lock_irqsave(&srq->lock, flags);
320	if (mdev->dev->persist->state & MLX4_DEVICE_STATE_INTERNAL_ERROR) {
321		err = -EIO;
322		*bad_wr = wr;
323		goto out;
324	}
325
326	for (nreq = 0; wr; ++nreq, wr = wr->next) {
327		if (unlikely(wr->num_sge > srq->msrq.max_gs)) {
328			err = -EINVAL;
329			*bad_wr = wr;
330			break;
331		}
332
333		if (unlikely(srq->head == srq->tail)) {
334			err = -ENOMEM;
335			*bad_wr = wr;
336			break;
337		}
338
339		srq->wrid[srq->head] = wr->wr_id;
340
341		next      = get_wqe(srq, srq->head);
342		srq->head = be16_to_cpu(next->next_wqe_index);
343		scat      = (struct mlx4_wqe_data_seg *) (next + 1);
344
345		for (i = 0; i < wr->num_sge; ++i) {
346			scat[i].byte_count = cpu_to_be32(wr->sg_list[i].length);
347			scat[i].lkey       = cpu_to_be32(wr->sg_list[i].lkey);
348			scat[i].addr       = cpu_to_be64(wr->sg_list[i].addr);
349		}
350
351		if (i < srq->msrq.max_gs) {
352			scat[i].byte_count = 0;
353			scat[i].lkey       = cpu_to_be32(MLX4_INVALID_LKEY);
354			scat[i].addr       = 0;
355		}
356	}
357
358	if (likely(nreq)) {
359		srq->wqe_ctr += nreq;
360
361		/*
362		 * Make sure that descriptors are written before
363		 * doorbell record.
364		 */
365		wmb();
366
367		*srq->db.db = cpu_to_be32(srq->wqe_ctr);
368	}
369out:
370
371	spin_unlock_irqrestore(&srq->lock, flags);
372
373	return err;
374}
375