2 * Copyright (c) 2013-2015, Mellanox Technologies. All rights reserved.
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:
10 * Redistribution and use in source and binary forms, with or
11 * without modification, are permitted provided that the following
14 * - Redistributions of source code must retain the above
15 * copyright notice, this list of conditions and the following
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.
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
33 #include <linux/module.h>
34 #include <linux/mlx5/qp.h>
35 #include <linux/mlx5/srq.h>
36 #include <linux/slab.h>
37 #include <rdma/ib_umem.h>
38 #include <rdma/ib_user_verbs.h>
43 /* not supported currently */
44 static int srq_signature;
46 static void *get_wqe(struct mlx5_ib_srq *srq, int n)
48 return mlx5_buf_offset(&srq->buf, n << srq->msrq.wqe_shift);
51 static void mlx5_ib_srq_event(struct mlx5_core_srq *srq, enum mlx5_event type)
53 struct ib_event event;
54 struct ib_srq *ibsrq = &to_mibsrq(srq)->ibsrq;
56 if (ibsrq->event_handler) {
57 event.device = ibsrq->device;
58 event.element.srq = ibsrq;
60 case MLX5_EVENT_TYPE_SRQ_RQ_LIMIT:
61 event.event = IB_EVENT_SRQ_LIMIT_REACHED;
63 case MLX5_EVENT_TYPE_SRQ_CATAS_ERROR:
64 event.event = IB_EVENT_SRQ_ERR;
67 pr_warn("mlx5_ib: Unexpected event type %d on SRQ %06x\n",
72 ibsrq->event_handler(&event, ibsrq->srq_context);
76 static int create_srq_user(struct ib_pd *pd, struct mlx5_ib_srq *srq,
77 struct mlx5_create_srq_mbox_in **in,
78 struct ib_udata *udata, int buf_size, int *inlen)
80 struct mlx5_ib_dev *dev = to_mdev(pd->device);
81 struct mlx5_ib_create_srq ucmd;
90 (udata->inlen - sizeof(struct ib_uverbs_cmd_hdr) <
91 sizeof(ucmd)) ? (sizeof(ucmd) -
92 sizeof(ucmd.reserved)) : sizeof(ucmd);
94 if (ib_copy_from_udata(&ucmd, udata, ucmdlen)) {
95 mlx5_ib_dbg(dev, "failed copy udata\n");
99 if (ucmdlen == sizeof(ucmd) &&
103 srq->wq_sig = !!(ucmd.flags & MLX5_SRQ_FLAG_SIGNATURE);
105 srq->umem = ib_umem_get(pd->uobject->context, ucmd.buf_addr, buf_size,
107 if (IS_ERR(srq->umem)) {
108 mlx5_ib_dbg(dev, "failed umem get, size %d\n", buf_size);
109 err = PTR_ERR(srq->umem);
113 mlx5_ib_cont_pages(srq->umem, ucmd.buf_addr, &npages,
114 &page_shift, &ncont, NULL);
115 err = mlx5_ib_get_buf_offset(ucmd.buf_addr, page_shift,
118 mlx5_ib_warn(dev, "bad offset\n");
122 *inlen = sizeof(**in) + sizeof(*(*in)->pas) * ncont;
123 *in = mlx5_vzalloc(*inlen);
129 mlx5_ib_populate_pas(dev, srq->umem, page_shift, (*in)->pas, 0);
131 err = mlx5_ib_db_map_user(to_mucontext(pd->uobject->context),
132 ucmd.db_addr, &srq->db);
134 mlx5_ib_dbg(dev, "map doorbell failed\n");
138 (*in)->ctx.log_pg_sz = page_shift - MLX5_ADAPTER_PAGE_SHIFT;
139 (*in)->ctx.pgoff_cqn = cpu_to_be32(offset << 26);
147 ib_umem_release(srq->umem);
152 static int create_srq_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_srq *srq,
153 struct mlx5_create_srq_mbox_in **in, int buf_size,
158 struct mlx5_wqe_srq_next_seg *next;
162 err = mlx5_db_alloc(dev->mdev, &srq->db);
164 mlx5_ib_warn(dev, "alloc dbell rec failed\n");
168 if (mlx5_buf_alloc(dev->mdev, buf_size, PAGE_SIZE * 2, &srq->buf)) {
169 mlx5_ib_dbg(dev, "buf alloc failed\n");
173 page_shift = srq->buf.page_shift;
176 srq->tail = srq->msrq.max - 1;
179 for (i = 0; i < srq->msrq.max; i++) {
180 next = get_wqe(srq, i);
181 next->next_wqe_index =
182 cpu_to_be16((i + 1) & (srq->msrq.max - 1));
185 npages = DIV_ROUND_UP(srq->buf.npages, 1 << (page_shift - PAGE_SHIFT));
186 mlx5_ib_dbg(dev, "buf_size %d, page_shift %d, npages %d, calc npages %d\n",
187 buf_size, page_shift, srq->buf.npages, npages);
188 *inlen = sizeof(**in) + sizeof(*(*in)->pas) * npages;
189 *in = mlx5_vzalloc(*inlen);
194 mlx5_fill_page_array(&srq->buf, (*in)->pas);
196 srq->wrid = kmalloc(srq->msrq.max * sizeof(u64), GFP_KERNEL);
198 mlx5_ib_dbg(dev, "kmalloc failed %lu\n",
199 (unsigned long)(srq->msrq.max * sizeof(u64)));
203 srq->wq_sig = !!srq_signature;
205 (*in)->ctx.log_pg_sz = page_shift - MLX5_ADAPTER_PAGE_SHIFT;
213 mlx5_buf_free(dev->mdev, &srq->buf);
216 mlx5_db_free(dev->mdev, &srq->db);
220 static void destroy_srq_user(struct ib_pd *pd, struct mlx5_ib_srq *srq)
222 mlx5_ib_db_unmap_user(to_mucontext(pd->uobject->context), &srq->db);
223 ib_umem_release(srq->umem);
227 static void destroy_srq_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_srq *srq)
230 mlx5_buf_free(dev->mdev, &srq->buf);
231 mlx5_db_free(dev->mdev, &srq->db);
234 struct ib_srq *mlx5_ib_create_srq(struct ib_pd *pd,
235 struct ib_srq_init_attr *init_attr,
236 struct ib_udata *udata)
238 struct mlx5_ib_dev *dev = to_mdev(pd->device);
239 struct mlx5_general_caps *gen;
240 struct mlx5_ib_srq *srq;
244 struct mlx5_create_srq_mbox_in *uninitialized_var(in);
245 int uninitialized_var(inlen);
249 gen = &dev->mdev->caps.gen;
250 /* Sanity check SRQ size before proceeding */
251 if (init_attr->attr.max_wr >= gen->max_srq_wqes) {
252 mlx5_ib_dbg(dev, "max_wr %d, cap %d\n",
253 init_attr->attr.max_wr,
255 return ERR_PTR(-EINVAL);
258 srq = kmalloc(sizeof(*srq), GFP_KERNEL);
260 return ERR_PTR(-ENOMEM);
262 mutex_init(&srq->mutex);
263 spin_lock_init(&srq->lock);
264 srq->msrq.max = roundup_pow_of_two(init_attr->attr.max_wr + 1);
265 srq->msrq.max_gs = init_attr->attr.max_sge;
267 desc_size = sizeof(struct mlx5_wqe_srq_next_seg) +
268 srq->msrq.max_gs * sizeof(struct mlx5_wqe_data_seg);
269 desc_size = roundup_pow_of_two(desc_size);
270 desc_size = max_t(int, 32, desc_size);
271 srq->msrq.max_avail_gather = (desc_size - sizeof(struct mlx5_wqe_srq_next_seg)) /
272 sizeof(struct mlx5_wqe_data_seg);
273 srq->msrq.wqe_shift = ilog2(desc_size);
274 buf_size = srq->msrq.max * desc_size;
275 mlx5_ib_dbg(dev, "desc_size 0x%x, req wr 0x%x, srq size 0x%x, max_gs 0x%x, max_avail_gather 0x%x\n",
276 desc_size, init_attr->attr.max_wr, srq->msrq.max, srq->msrq.max_gs,
277 srq->msrq.max_avail_gather);
280 err = create_srq_user(pd, srq, &in, udata, buf_size, &inlen);
282 err = create_srq_kernel(dev, srq, &in, buf_size, &inlen);
285 mlx5_ib_warn(dev, "create srq %s failed, err %d\n",
286 pd->uobject ? "user" : "kernel", err);
290 is_xrc = (init_attr->srq_type == IB_SRQT_XRC);
291 in->ctx.state_log_sz = ilog2(srq->msrq.max);
292 flgs = ((srq->msrq.wqe_shift - 4) | (is_xrc << 5) | (srq->wq_sig << 7)) << 24;
295 xrcdn = to_mxrcd(init_attr->ext.xrc.xrcd)->xrcdn;
296 in->ctx.pgoff_cqn |= cpu_to_be32(to_mcq(init_attr->ext.xrc.cq)->mcq.cqn);
297 } else if (init_attr->srq_type == IB_SRQT_BASIC) {
298 xrcdn = to_mxrcd(dev->devr.x0)->xrcdn;
299 in->ctx.pgoff_cqn |= cpu_to_be32(to_mcq(dev->devr.c0)->mcq.cqn);
302 in->ctx.flags_xrcd = cpu_to_be32((flgs & 0xFF000000) | (xrcdn & 0xFFFFFF));
304 in->ctx.pd = cpu_to_be32(to_mpd(pd)->pdn);
305 in->ctx.db_record = cpu_to_be64(srq->db.dma);
306 err = mlx5_core_create_srq(dev->mdev, &srq->msrq, in, inlen);
309 mlx5_ib_dbg(dev, "create SRQ failed, err %d\n", err);
310 goto err_usr_kern_srq;
313 mlx5_ib_dbg(dev, "create SRQ with srqn 0x%x\n", srq->msrq.srqn);
315 srq->msrq.event = mlx5_ib_srq_event;
316 srq->ibsrq.ext.xrc.srq_num = srq->msrq.srqn;
319 if (ib_copy_to_udata(udata, &srq->msrq.srqn, sizeof(__u32))) {
320 mlx5_ib_dbg(dev, "copy to user failed\n");
325 init_attr->attr.max_wr = srq->msrq.max - 1;
330 mlx5_core_destroy_srq(dev->mdev, &srq->msrq);
334 destroy_srq_user(pd, srq);
336 destroy_srq_kernel(dev, srq);
344 int mlx5_ib_modify_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr,
345 enum ib_srq_attr_mask attr_mask, struct ib_udata *udata)
347 struct mlx5_ib_dev *dev = to_mdev(ibsrq->device);
348 struct mlx5_ib_srq *srq = to_msrq(ibsrq);
351 /* We don't support resizing SRQs yet */
352 if (attr_mask & IB_SRQ_MAX_WR)
355 if (attr_mask & IB_SRQ_LIMIT) {
356 if (attr->srq_limit >= srq->msrq.max)
359 mutex_lock(&srq->mutex);
360 ret = mlx5_core_arm_srq(dev->mdev, &srq->msrq, attr->srq_limit, 1);
361 mutex_unlock(&srq->mutex);
370 int mlx5_ib_query_srq(struct ib_srq *ibsrq, struct ib_srq_attr *srq_attr)
372 struct mlx5_ib_dev *dev = to_mdev(ibsrq->device);
373 struct mlx5_ib_srq *srq = to_msrq(ibsrq);
375 struct mlx5_query_srq_mbox_out *out;
377 out = kzalloc(sizeof(*out), GFP_KERNEL);
381 ret = mlx5_core_query_srq(dev->mdev, &srq->msrq, out);
385 srq_attr->srq_limit = be16_to_cpu(out->ctx.lwm);
386 srq_attr->max_wr = srq->msrq.max - 1;
387 srq_attr->max_sge = srq->msrq.max_gs;
394 int mlx5_ib_destroy_srq(struct ib_srq *srq)
396 struct mlx5_ib_dev *dev = to_mdev(srq->device);
397 struct mlx5_ib_srq *msrq = to_msrq(srq);
399 mlx5_core_destroy_srq(dev->mdev, &msrq->msrq);
402 mlx5_ib_db_unmap_user(to_mucontext(srq->uobject->context), &msrq->db);
403 ib_umem_release(msrq->umem);
405 destroy_srq_kernel(dev, msrq);
412 void mlx5_ib_free_srq_wqe(struct mlx5_ib_srq *srq, int wqe_index)
414 struct mlx5_wqe_srq_next_seg *next;
416 /* always called with interrupts disabled. */
417 spin_lock(&srq->lock);
419 next = get_wqe(srq, srq->tail);
420 next->next_wqe_index = cpu_to_be16(wqe_index);
421 srq->tail = wqe_index;
423 spin_unlock(&srq->lock);
426 int mlx5_ib_post_srq_recv(struct ib_srq *ibsrq, struct ib_recv_wr *wr,
427 struct ib_recv_wr **bad_wr)
429 struct mlx5_ib_srq *srq = to_msrq(ibsrq);
430 struct mlx5_wqe_srq_next_seg *next;
431 struct mlx5_wqe_data_seg *scat;
437 spin_lock_irqsave(&srq->lock, flags);
439 for (nreq = 0; wr; nreq++, wr = wr->next) {
440 if (unlikely(wr->num_sge > srq->msrq.max_gs)) {
446 if (unlikely(srq->head == srq->tail)) {
452 srq->wrid[srq->head] = wr->wr_id;
454 next = get_wqe(srq, srq->head);
455 srq->head = be16_to_cpu(next->next_wqe_index);
456 scat = (struct mlx5_wqe_data_seg *)(next + 1);
458 for (i = 0; i < wr->num_sge; i++) {
459 scat[i].byte_count = cpu_to_be32(wr->sg_list[i].length);
460 scat[i].lkey = cpu_to_be32(wr->sg_list[i].lkey);
461 scat[i].addr = cpu_to_be64(wr->sg_list[i].addr);
464 if (i < srq->msrq.max_avail_gather) {
465 scat[i].byte_count = 0;
466 scat[i].lkey = cpu_to_be32(MLX5_INVALID_LKEY);
472 srq->wqe_ctr += nreq;
474 /* Make sure that descriptors are written before
479 *srq->db.db = cpu_to_be32(srq->wqe_ctr);
482 spin_unlock_irqrestore(&srq->lock, flags);