srq.c revision 331769
1/*
2 * Copyright (c) 2007 Cisco, Inc.  All rights reserved.
3 *
4 * This software is available to you under a choice of one of two
5 * licenses.  You may choose to be licensed under the terms of the GNU
6 * General Public License (GPL) Version 2, available from the file
7 * COPYING in the main directory of this source tree, or the
8 * OpenIB.org BSD license below:
9 *
10 *     Redistribution and use in source and binary forms, with or
11 *     without modification, are permitted provided that the following
12 *     conditions are met:
13 *
14 *      - Redistributions of source code must retain the above
15 *        copyright notice, this list of conditions and the following
16 *        disclaimer.
17 *
18 *      - Redistributions in binary form must reproduce the above
19 *        copyright notice, this list of conditions and the following
20 *        disclaimer in the documentation and/or other materials
21 *        provided with the distribution.
22 *
23 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
30 * SOFTWARE.
31 */
32
33#include <config.h>
34
35#include <stdlib.h>
36#include <pthread.h>
37#include <string.h>
38
39#include "mlx4.h"
40#include "doorbell.h"
41#include "wqe.h"
42#include "mlx4-abi.h"
43
44static void *get_wqe(struct mlx4_srq *srq, int n)
45{
46	return srq->buf.buf + (n << srq->wqe_shift);
47}
48
49void mlx4_free_srq_wqe(struct mlx4_srq *srq, int ind)
50{
51	struct mlx4_wqe_srq_next_seg *next;
52
53	pthread_spin_lock(&srq->lock);
54
55	next = get_wqe(srq, srq->tail);
56	next->next_wqe_index = htobe16(ind);
57	srq->tail = ind;
58
59	pthread_spin_unlock(&srq->lock);
60}
61
62int mlx4_post_srq_recv(struct ibv_srq *ibsrq,
63		       struct ibv_recv_wr *wr,
64		       struct ibv_recv_wr **bad_wr)
65{
66	struct mlx4_srq *srq = to_msrq(ibsrq);
67	struct mlx4_wqe_srq_next_seg *next;
68	struct mlx4_wqe_data_seg *scat;
69	int err = 0;
70	int nreq;
71	int i;
72
73	pthread_spin_lock(&srq->lock);
74
75	for (nreq = 0; wr; ++nreq, wr = wr->next) {
76		if (wr->num_sge > srq->max_gs) {
77			err = -1;
78			*bad_wr = wr;
79			break;
80		}
81
82		if (srq->head == srq->tail) {
83			/* SRQ is full*/
84			err = -1;
85			*bad_wr = wr;
86			break;
87		}
88
89		srq->wrid[srq->head] = wr->wr_id;
90
91		next      = get_wqe(srq, srq->head);
92		srq->head = be16toh(next->next_wqe_index);
93		scat      = (struct mlx4_wqe_data_seg *) (next + 1);
94
95		for (i = 0; i < wr->num_sge; ++i) {
96			scat[i].byte_count = htobe32(wr->sg_list[i].length);
97			scat[i].lkey       = htobe32(wr->sg_list[i].lkey);
98			scat[i].addr       = htobe64(wr->sg_list[i].addr);
99		}
100
101		if (i < srq->max_gs) {
102			scat[i].byte_count = 0;
103			scat[i].lkey       = htobe32(MLX4_INVALID_LKEY);
104			scat[i].addr       = 0;
105		}
106	}
107
108	if (nreq) {
109		srq->counter += nreq;
110
111		/*
112		 * Make sure that descriptors are written before
113		 * we write doorbell record.
114		 */
115		udma_to_device_barrier();
116
117		*srq->db = htobe32(srq->counter);
118	}
119
120	pthread_spin_unlock(&srq->lock);
121
122	return err;
123}
124
125int mlx4_alloc_srq_buf(struct ibv_pd *pd, struct ibv_srq_attr *attr,
126		       struct mlx4_srq *srq)
127{
128	struct mlx4_wqe_srq_next_seg *next;
129	struct mlx4_wqe_data_seg *scatter;
130	int size;
131	int buf_size;
132	int i;
133
134	srq->wrid = malloc(srq->max * sizeof (uint64_t));
135	if (!srq->wrid)
136		return -1;
137
138	size = sizeof (struct mlx4_wqe_srq_next_seg) +
139		srq->max_gs * sizeof (struct mlx4_wqe_data_seg);
140
141	for (srq->wqe_shift = 5; 1 << srq->wqe_shift < size; ++srq->wqe_shift)
142		; /* nothing */
143
144	buf_size = srq->max << srq->wqe_shift;
145
146	if (mlx4_alloc_buf(&srq->buf, buf_size,
147			   to_mdev(pd->context->device)->page_size)) {
148		free(srq->wrid);
149		return -1;
150	}
151
152	memset(srq->buf.buf, 0, buf_size);
153
154	/*
155	 * Now initialize the SRQ buffer so that all of the WQEs are
156	 * linked into the list of free WQEs.
157	 */
158
159	for (i = 0; i < srq->max; ++i) {
160		next = get_wqe(srq, i);
161		next->next_wqe_index = htobe16((i + 1) & (srq->max - 1));
162
163		for (scatter = (void *) (next + 1);
164		     (void *) scatter < (void *) next + (1 << srq->wqe_shift);
165		     ++scatter)
166			scatter->lkey = htobe32(MLX4_INVALID_LKEY);
167	}
168
169	srq->head = 0;
170	srq->tail = srq->max - 1;
171
172	return 0;
173}
174
175void mlx4_init_xsrq_table(struct mlx4_xsrq_table *xsrq_table, int size)
176{
177	memset(xsrq_table, 0, sizeof *xsrq_table);
178	xsrq_table->num_xsrq = size;
179	xsrq_table->shift = ffs(size) - 1 - MLX4_XSRQ_TABLE_BITS;
180	xsrq_table->mask = (1 << xsrq_table->shift) - 1;
181
182	pthread_mutex_init(&xsrq_table->mutex, NULL);
183}
184
185struct mlx4_srq *mlx4_find_xsrq(struct mlx4_xsrq_table *xsrq_table, uint32_t srqn)
186{
187	int index;
188
189	index = (srqn & (xsrq_table->num_xsrq - 1)) >> xsrq_table->shift;
190	if (xsrq_table->xsrq_table[index].refcnt)
191		return xsrq_table->xsrq_table[index].table[srqn & xsrq_table->mask];
192
193	return NULL;
194}
195
196int mlx4_store_xsrq(struct mlx4_xsrq_table *xsrq_table, uint32_t srqn,
197		    struct mlx4_srq *srq)
198{
199	int index, ret = 0;
200
201	index = (srqn & (xsrq_table->num_xsrq - 1)) >> xsrq_table->shift;
202	pthread_mutex_lock(&xsrq_table->mutex);
203	if (!xsrq_table->xsrq_table[index].refcnt) {
204		xsrq_table->xsrq_table[index].table = calloc(xsrq_table->mask + 1,
205							     sizeof(struct mlx4_srq *));
206		if (!xsrq_table->xsrq_table[index].table) {
207			ret = -1;
208			goto out;
209		}
210	}
211
212	xsrq_table->xsrq_table[index].refcnt++;
213	xsrq_table->xsrq_table[index].table[srqn & xsrq_table->mask] = srq;
214
215out:
216	pthread_mutex_unlock(&xsrq_table->mutex);
217	return ret;
218}
219
220void mlx4_clear_xsrq(struct mlx4_xsrq_table *xsrq_table, uint32_t srqn)
221{
222	int index;
223
224	index = (srqn & (xsrq_table->num_xsrq - 1)) >> xsrq_table->shift;
225	pthread_mutex_lock(&xsrq_table->mutex);
226
227	if (--xsrq_table->xsrq_table[index].refcnt)
228		xsrq_table->xsrq_table[index].table[srqn & xsrq_table->mask] = NULL;
229	else
230		free(xsrq_table->xsrq_table[index].table);
231
232	pthread_mutex_unlock(&xsrq_table->mutex);
233}
234
235struct ibv_srq *mlx4_create_xrc_srq(struct ibv_context *context,
236				    struct ibv_srq_init_attr_ex *attr_ex)
237{
238	struct mlx4_create_xsrq cmd;
239	struct mlx4_create_srq_resp resp;
240	struct mlx4_srq *srq;
241	int ret;
242
243	/* Sanity check SRQ size before proceeding */
244	if (attr_ex->attr.max_wr > 1 << 16 || attr_ex->attr.max_sge > 64)
245		return NULL;
246
247	srq = calloc(1, sizeof *srq);
248	if (!srq)
249		return NULL;
250
251	if (pthread_spin_init(&srq->lock, PTHREAD_PROCESS_PRIVATE))
252		goto err;
253
254	srq->max     = align_queue_size(attr_ex->attr.max_wr + 1);
255	srq->max_gs  = attr_ex->attr.max_sge;
256	srq->counter = 0;
257	srq->ext_srq = 1;
258
259	if (mlx4_alloc_srq_buf(attr_ex->pd, &attr_ex->attr, srq))
260		goto err;
261
262	srq->db = mlx4_alloc_db(to_mctx(context), MLX4_DB_TYPE_RQ);
263	if (!srq->db)
264		goto err_free;
265
266	*srq->db = 0;
267
268	cmd.buf_addr = (uintptr_t) srq->buf.buf;
269	cmd.db_addr  = (uintptr_t) srq->db;
270
271	ret = ibv_cmd_create_srq_ex(context, &srq->verbs_srq,
272				    sizeof(srq->verbs_srq),
273				    attr_ex,
274				    &cmd.ibv_cmd, sizeof cmd,
275				    &resp.ibv_resp, sizeof resp);
276	if (ret)
277		goto err_db;
278
279	ret = mlx4_store_xsrq(&to_mctx(context)->xsrq_table,
280			      srq->verbs_srq.srq_num, srq);
281	if (ret)
282		goto err_destroy;
283
284	return &srq->verbs_srq.srq;
285
286err_destroy:
287	ibv_cmd_destroy_srq(&srq->verbs_srq.srq);
288err_db:
289	mlx4_free_db(to_mctx(context), MLX4_DB_TYPE_RQ, srq->db);
290err_free:
291	free(srq->wrid);
292	mlx4_free_buf(&srq->buf);
293err:
294	free(srq);
295	return NULL;
296}
297
298int mlx4_destroy_xrc_srq(struct ibv_srq *srq)
299{
300	struct mlx4_context *mctx = to_mctx(srq->context);
301	struct mlx4_srq *msrq = to_msrq(srq);
302	struct mlx4_cq *mcq;
303	int ret;
304
305	mcq = to_mcq(msrq->verbs_srq.cq);
306	mlx4_cq_clean(mcq, 0, msrq);
307	pthread_spin_lock(&mcq->lock);
308	mlx4_clear_xsrq(&mctx->xsrq_table, msrq->verbs_srq.srq_num);
309	pthread_spin_unlock(&mcq->lock);
310
311	ret = ibv_cmd_destroy_srq(srq);
312	if (ret) {
313		pthread_spin_lock(&mcq->lock);
314		mlx4_store_xsrq(&mctx->xsrq_table, msrq->verbs_srq.srq_num, msrq);
315		pthread_spin_unlock(&mcq->lock);
316		return ret;
317	}
318
319	mlx4_free_db(mctx, MLX4_DB_TYPE_RQ, msrq->db);
320	mlx4_free_buf(&msrq->buf);
321	free(msrq->wrid);
322	free(msrq);
323
324	return 0;
325}
326