1// SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB
2/*
3 * Copyright (c) 2013-2018, Mellanox Technologies inc.  All rights reserved.
4 */
5
6#include <linux/mlx5/qp.h>
7#include <linux/slab.h>
8#include <rdma/ib_umem.h>
9#include <rdma/ib_user_verbs.h>
10#include "mlx5_ib.h"
11#include "srq.h"
12
13static void *get_wqe(struct mlx5_ib_srq *srq, int n)
14{
15	return mlx5_frag_buf_get_wqe(&srq->fbc, n);
16}
17
18static void mlx5_ib_srq_event(struct mlx5_core_srq *srq, enum mlx5_event type)
19{
20	struct ib_event event;
21	struct ib_srq *ibsrq = &to_mibsrq(srq)->ibsrq;
22
23	if (ibsrq->event_handler) {
24		event.device      = ibsrq->device;
25		event.element.srq = ibsrq;
26		switch (type) {
27		case MLX5_EVENT_TYPE_SRQ_RQ_LIMIT:
28			event.event = IB_EVENT_SRQ_LIMIT_REACHED;
29			break;
30		case MLX5_EVENT_TYPE_SRQ_CATAS_ERROR:
31			event.event = IB_EVENT_SRQ_ERR;
32			break;
33		default:
34			pr_warn("mlx5_ib: Unexpected event type %d on SRQ %06x\n",
35				type, srq->srqn);
36			return;
37		}
38
39		ibsrq->event_handler(&event, ibsrq->srq_context);
40	}
41}
42
43static int create_srq_user(struct ib_pd *pd, struct mlx5_ib_srq *srq,
44			   struct mlx5_srq_attr *in,
45			   struct ib_udata *udata, int buf_size)
46{
47	struct mlx5_ib_dev *dev = to_mdev(pd->device);
48	struct mlx5_ib_create_srq ucmd = {};
49	struct mlx5_ib_ucontext *ucontext = rdma_udata_to_drv_context(
50		udata, struct mlx5_ib_ucontext, ibucontext);
51	size_t ucmdlen;
52	int err;
53	u32 uidx = MLX5_IB_DEFAULT_UIDX;
54
55	ucmdlen = min(udata->inlen, sizeof(ucmd));
56
57	if (ib_copy_from_udata(&ucmd, udata, ucmdlen)) {
58		mlx5_ib_dbg(dev, "failed copy udata\n");
59		return -EFAULT;
60	}
61
62	if (ucmd.reserved0 || ucmd.reserved1)
63		return -EINVAL;
64
65	if (udata->inlen > sizeof(ucmd) &&
66	    !ib_is_udata_cleared(udata, sizeof(ucmd),
67				 udata->inlen - sizeof(ucmd)))
68		return -EINVAL;
69
70	if (in->type != IB_SRQT_BASIC) {
71		err = get_srq_user_index(ucontext, &ucmd, udata->inlen, &uidx);
72		if (err)
73			return err;
74	}
75
76	srq->wq_sig = !!(ucmd.flags & MLX5_SRQ_FLAG_SIGNATURE);
77
78	srq->umem = ib_umem_get(pd->device, ucmd.buf_addr, buf_size, 0);
79	if (IS_ERR(srq->umem)) {
80		mlx5_ib_dbg(dev, "failed umem get, size %d\n", buf_size);
81		err = PTR_ERR(srq->umem);
82		return err;
83	}
84	in->umem = srq->umem;
85
86	err = mlx5_ib_db_map_user(ucontext, ucmd.db_addr, &srq->db);
87	if (err) {
88		mlx5_ib_dbg(dev, "map doorbell failed\n");
89		goto err_umem;
90	}
91
92	in->uid = (in->type != IB_SRQT_XRC) ?  to_mpd(pd)->uid : 0;
93	if (MLX5_CAP_GEN(dev->mdev, cqe_version) == MLX5_CQE_VERSION_V1 &&
94	    in->type != IB_SRQT_BASIC)
95		in->user_index = uidx;
96
97	return 0;
98
99err_umem:
100	ib_umem_release(srq->umem);
101
102	return err;
103}
104
105static int create_srq_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_srq *srq,
106			     struct mlx5_srq_attr *in, int buf_size)
107{
108	int err;
109	int i;
110	struct mlx5_wqe_srq_next_seg *next;
111
112	err = mlx5_db_alloc(dev->mdev, &srq->db);
113	if (err) {
114		mlx5_ib_warn(dev, "alloc dbell rec failed\n");
115		return err;
116	}
117
118	if (mlx5_frag_buf_alloc_node(dev->mdev, buf_size, &srq->buf,
119				     dev->mdev->priv.numa_node)) {
120		mlx5_ib_dbg(dev, "buf alloc failed\n");
121		err = -ENOMEM;
122		goto err_db;
123	}
124
125	mlx5_init_fbc(srq->buf.frags, srq->msrq.wqe_shift, ilog2(srq->msrq.max),
126		      &srq->fbc);
127
128	srq->head    = 0;
129	srq->tail    = srq->msrq.max - 1;
130	srq->wqe_ctr = 0;
131
132	for (i = 0; i < srq->msrq.max; i++) {
133		next = get_wqe(srq, i);
134		next->next_wqe_index =
135			cpu_to_be16((i + 1) & (srq->msrq.max - 1));
136	}
137
138	mlx5_ib_dbg(dev, "srq->buf.page_shift = %d\n", srq->buf.page_shift);
139	in->pas = kvcalloc(srq->buf.npages, sizeof(*in->pas), GFP_KERNEL);
140	if (!in->pas) {
141		err = -ENOMEM;
142		goto err_buf;
143	}
144	mlx5_fill_page_frag_array(&srq->buf, in->pas);
145
146	srq->wrid = kvmalloc_array(srq->msrq.max, sizeof(u64), GFP_KERNEL);
147	if (!srq->wrid) {
148		err = -ENOMEM;
149		goto err_in;
150	}
151	srq->wq_sig = 0;
152
153	in->log_page_size = srq->buf.page_shift - MLX5_ADAPTER_PAGE_SHIFT;
154	if (MLX5_CAP_GEN(dev->mdev, cqe_version) == MLX5_CQE_VERSION_V1 &&
155	    in->type != IB_SRQT_BASIC)
156		in->user_index = MLX5_IB_DEFAULT_UIDX;
157
158	return 0;
159
160err_in:
161	kvfree(in->pas);
162
163err_buf:
164	mlx5_frag_buf_free(dev->mdev, &srq->buf);
165
166err_db:
167	mlx5_db_free(dev->mdev, &srq->db);
168	return err;
169}
170
171static void destroy_srq_user(struct ib_pd *pd, struct mlx5_ib_srq *srq,
172			     struct ib_udata *udata)
173{
174	mlx5_ib_db_unmap_user(
175		rdma_udata_to_drv_context(
176			udata,
177			struct mlx5_ib_ucontext,
178			ibucontext),
179		&srq->db);
180	ib_umem_release(srq->umem);
181}
182
183
184static void destroy_srq_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_srq *srq)
185{
186	kvfree(srq->wrid);
187	mlx5_frag_buf_free(dev->mdev, &srq->buf);
188	mlx5_db_free(dev->mdev, &srq->db);
189}
190
191int mlx5_ib_create_srq(struct ib_srq *ib_srq,
192		       struct ib_srq_init_attr *init_attr,
193		       struct ib_udata *udata)
194{
195	struct mlx5_ib_dev *dev = to_mdev(ib_srq->device);
196	struct mlx5_ib_srq *srq = to_msrq(ib_srq);
197	size_t desc_size;
198	size_t buf_size;
199	int err;
200	struct mlx5_srq_attr in = {};
201	__u32 max_srq_wqes = 1 << MLX5_CAP_GEN(dev->mdev, log_max_srq_sz);
202
203	if (init_attr->srq_type != IB_SRQT_BASIC &&
204	    init_attr->srq_type != IB_SRQT_XRC &&
205	    init_attr->srq_type != IB_SRQT_TM)
206		return -EOPNOTSUPP;
207
208	/* Sanity check SRQ size before proceeding */
209	if (init_attr->attr.max_wr >= max_srq_wqes) {
210		mlx5_ib_dbg(dev, "max_wr %d, cap %d\n",
211			    init_attr->attr.max_wr,
212			    max_srq_wqes);
213		return -EINVAL;
214	}
215
216	mutex_init(&srq->mutex);
217	spin_lock_init(&srq->lock);
218	srq->msrq.max    = roundup_pow_of_two(init_attr->attr.max_wr + 1);
219	srq->msrq.max_gs = init_attr->attr.max_sge;
220
221	desc_size = sizeof(struct mlx5_wqe_srq_next_seg) +
222		    srq->msrq.max_gs * sizeof(struct mlx5_wqe_data_seg);
223	if (desc_size == 0 || srq->msrq.max_gs > desc_size)
224		return -EINVAL;
225
226	desc_size = roundup_pow_of_two(desc_size);
227	desc_size = max_t(size_t, 32, desc_size);
228	if (desc_size < sizeof(struct mlx5_wqe_srq_next_seg))
229		return -EINVAL;
230
231	srq->msrq.max_avail_gather = (desc_size - sizeof(struct mlx5_wqe_srq_next_seg)) /
232		sizeof(struct mlx5_wqe_data_seg);
233	srq->msrq.wqe_shift = ilog2(desc_size);
234	buf_size = srq->msrq.max * desc_size;
235	if (buf_size < desc_size)
236		return -EINVAL;
237
238	in.type = init_attr->srq_type;
239
240	if (udata)
241		err = create_srq_user(ib_srq->pd, srq, &in, udata, buf_size);
242	else
243		err = create_srq_kernel(dev, srq, &in, buf_size);
244
245	if (err) {
246		mlx5_ib_warn(dev, "create srq %s failed, err %d\n",
247			     udata ? "user" : "kernel", err);
248		return err;
249	}
250
251	in.log_size = ilog2(srq->msrq.max);
252	in.wqe_shift = srq->msrq.wqe_shift - 4;
253	if (srq->wq_sig)
254		in.flags |= MLX5_SRQ_FLAG_WQ_SIG;
255
256	if (init_attr->srq_type == IB_SRQT_XRC && init_attr->ext.xrc.xrcd)
257		in.xrcd = to_mxrcd(init_attr->ext.xrc.xrcd)->xrcdn;
258	else
259		in.xrcd = dev->devr.xrcdn0;
260
261	if (init_attr->srq_type == IB_SRQT_TM) {
262		in.tm_log_list_size =
263			ilog2(init_attr->ext.tag_matching.max_num_tags) + 1;
264		if (in.tm_log_list_size >
265		    MLX5_CAP_GEN(dev->mdev, log_tag_matching_list_sz)) {
266			mlx5_ib_dbg(dev, "TM SRQ max_num_tags exceeding limit\n");
267			err = -EINVAL;
268			goto err_usr_kern_srq;
269		}
270		in.flags |= MLX5_SRQ_FLAG_RNDV;
271	}
272
273	if (ib_srq_has_cq(init_attr->srq_type))
274		in.cqn = to_mcq(init_attr->ext.cq)->mcq.cqn;
275	else
276		in.cqn = to_mcq(dev->devr.c0)->mcq.cqn;
277
278	in.pd = to_mpd(ib_srq->pd)->pdn;
279	in.db_record = srq->db.dma;
280	err = mlx5_cmd_create_srq(dev, &srq->msrq, &in);
281	kvfree(in.pas);
282	if (err) {
283		mlx5_ib_dbg(dev, "create SRQ failed, err %d\n", err);
284		goto err_usr_kern_srq;
285	}
286
287	mlx5_ib_dbg(dev, "create SRQ with srqn 0x%x\n", srq->msrq.srqn);
288
289	srq->msrq.event = mlx5_ib_srq_event;
290	srq->ibsrq.ext.xrc.srq_num = srq->msrq.srqn;
291
292	if (udata) {
293		struct mlx5_ib_create_srq_resp resp = {
294			.srqn = srq->msrq.srqn,
295		};
296
297		if (ib_copy_to_udata(udata, &resp, min(udata->outlen,
298				     sizeof(resp)))) {
299			mlx5_ib_dbg(dev, "copy to user failed\n");
300			err = -EFAULT;
301			goto err_core;
302		}
303	}
304
305	init_attr->attr.max_wr = srq->msrq.max - 1;
306
307	return 0;
308
309err_core:
310	mlx5_cmd_destroy_srq(dev, &srq->msrq);
311
312err_usr_kern_srq:
313	if (udata)
314		destroy_srq_user(ib_srq->pd, srq, udata);
315	else
316		destroy_srq_kernel(dev, srq);
317
318	return err;
319}
320
321int mlx5_ib_modify_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr,
322		       enum ib_srq_attr_mask attr_mask, struct ib_udata *udata)
323{
324	struct mlx5_ib_dev *dev = to_mdev(ibsrq->device);
325	struct mlx5_ib_srq *srq = to_msrq(ibsrq);
326	int ret;
327
328	/* We don't support resizing SRQs yet */
329	if (attr_mask & IB_SRQ_MAX_WR)
330		return -EINVAL;
331
332	if (attr_mask & IB_SRQ_LIMIT) {
333		if (attr->srq_limit >= srq->msrq.max)
334			return -EINVAL;
335
336		mutex_lock(&srq->mutex);
337		ret = mlx5_cmd_arm_srq(dev, &srq->msrq, attr->srq_limit, 1);
338		mutex_unlock(&srq->mutex);
339
340		if (ret)
341			return ret;
342	}
343
344	return 0;
345}
346
347int mlx5_ib_query_srq(struct ib_srq *ibsrq, struct ib_srq_attr *srq_attr)
348{
349	struct mlx5_ib_dev *dev = to_mdev(ibsrq->device);
350	struct mlx5_ib_srq *srq = to_msrq(ibsrq);
351	int ret;
352	struct mlx5_srq_attr *out;
353
354	out = kzalloc(sizeof(*out), GFP_KERNEL);
355	if (!out)
356		return -ENOMEM;
357
358	ret = mlx5_cmd_query_srq(dev, &srq->msrq, out);
359	if (ret)
360		goto out_box;
361
362	srq_attr->srq_limit = out->lwm;
363	srq_attr->max_wr    = srq->msrq.max - 1;
364	srq_attr->max_sge   = srq->msrq.max_gs;
365
366out_box:
367	kfree(out);
368	return ret;
369}
370
371int mlx5_ib_destroy_srq(struct ib_srq *srq, struct ib_udata *udata)
372{
373	struct mlx5_ib_dev *dev = to_mdev(srq->device);
374	struct mlx5_ib_srq *msrq = to_msrq(srq);
375	int ret;
376
377	ret = mlx5_cmd_destroy_srq(dev, &msrq->msrq);
378	if (ret)
379		return ret;
380
381	if (udata)
382		destroy_srq_user(srq->pd, msrq, udata);
383	else
384		destroy_srq_kernel(dev, msrq);
385	return 0;
386}
387
388void mlx5_ib_free_srq_wqe(struct mlx5_ib_srq *srq, int wqe_index)
389{
390	struct mlx5_wqe_srq_next_seg *next;
391
392	/* always called with interrupts disabled. */
393	spin_lock(&srq->lock);
394
395	next = get_wqe(srq, srq->tail);
396	next->next_wqe_index = cpu_to_be16(wqe_index);
397	srq->tail = wqe_index;
398
399	spin_unlock(&srq->lock);
400}
401
402int mlx5_ib_post_srq_recv(struct ib_srq *ibsrq, const struct ib_recv_wr *wr,
403			  const struct ib_recv_wr **bad_wr)
404{
405	struct mlx5_ib_srq *srq = to_msrq(ibsrq);
406	struct mlx5_wqe_srq_next_seg *next;
407	struct mlx5_wqe_data_seg *scat;
408	struct mlx5_ib_dev *dev = to_mdev(ibsrq->device);
409	struct mlx5_core_dev *mdev = dev->mdev;
410	unsigned long flags;
411	int err = 0;
412	int nreq;
413	int i;
414
415	spin_lock_irqsave(&srq->lock, flags);
416
417	if (mdev->state == MLX5_DEVICE_STATE_INTERNAL_ERROR) {
418		err = -EIO;
419		*bad_wr = wr;
420		goto out;
421	}
422
423	for (nreq = 0; wr; nreq++, wr = wr->next) {
424		if (unlikely(wr->num_sge > srq->msrq.max_gs)) {
425			err = -EINVAL;
426			*bad_wr = wr;
427			break;
428		}
429
430		if (unlikely(srq->head == srq->tail)) {
431			err = -ENOMEM;
432			*bad_wr = wr;
433			break;
434		}
435
436		srq->wrid[srq->head] = wr->wr_id;
437
438		next      = get_wqe(srq, srq->head);
439		srq->head = be16_to_cpu(next->next_wqe_index);
440		scat      = (struct mlx5_wqe_data_seg *)(next + 1);
441
442		for (i = 0; i < wr->num_sge; i++) {
443			scat[i].byte_count = cpu_to_be32(wr->sg_list[i].length);
444			scat[i].lkey       = cpu_to_be32(wr->sg_list[i].lkey);
445			scat[i].addr       = cpu_to_be64(wr->sg_list[i].addr);
446		}
447
448		if (i < srq->msrq.max_avail_gather) {
449			scat[i].byte_count = 0;
450			scat[i].lkey = dev->mkeys.terminate_scatter_list_mkey;
451			scat[i].addr       = 0;
452		}
453	}
454
455	if (likely(nreq)) {
456		srq->wqe_ctr += nreq;
457
458		/* Make sure that descriptors are written before
459		 * doorbell record.
460		 */
461		wmb();
462
463		*srq->db.db = cpu_to_be32(srq->wqe_ctr);
464	}
465out:
466	spin_unlock_irqrestore(&srq->lock, flags);
467
468	return err;
469}
470