xref: /openbmc/linux/drivers/infiniband/hw/mlx5/srq.c (revision 55fd7e02)
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/module.h>
7 #include <linux/mlx5/qp.h>
8 #include <linux/slab.h>
9 #include <rdma/ib_umem.h>
10 #include <rdma/ib_user_verbs.h>
11 #include "mlx5_ib.h"
12 #include "srq.h"
13 
14 static void *get_wqe(struct mlx5_ib_srq *srq, int n)
15 {
16 	return mlx5_frag_buf_get_wqe(&srq->fbc, n);
17 }
18 
19 static void mlx5_ib_srq_event(struct mlx5_core_srq *srq, enum mlx5_event type)
20 {
21 	struct ib_event event;
22 	struct ib_srq *ibsrq = &to_mibsrq(srq)->ibsrq;
23 
24 	if (ibsrq->event_handler) {
25 		event.device      = ibsrq->device;
26 		event.element.srq = ibsrq;
27 		switch (type) {
28 		case MLX5_EVENT_TYPE_SRQ_RQ_LIMIT:
29 			event.event = IB_EVENT_SRQ_LIMIT_REACHED;
30 			break;
31 		case MLX5_EVENT_TYPE_SRQ_CATAS_ERROR:
32 			event.event = IB_EVENT_SRQ_ERR;
33 			break;
34 		default:
35 			pr_warn("mlx5_ib: Unexpected event type %d on SRQ %06x\n",
36 				type, srq->srqn);
37 			return;
38 		}
39 
40 		ibsrq->event_handler(&event, ibsrq->srq_context);
41 	}
42 }
43 
44 static int create_srq_user(struct ib_pd *pd, struct mlx5_ib_srq *srq,
45 			   struct mlx5_srq_attr *in,
46 			   struct ib_udata *udata, int buf_size)
47 {
48 	struct mlx5_ib_dev *dev = to_mdev(pd->device);
49 	struct mlx5_ib_create_srq ucmd = {};
50 	struct mlx5_ib_ucontext *ucontext = rdma_udata_to_drv_context(
51 		udata, struct mlx5_ib_ucontext, ibucontext);
52 	size_t ucmdlen;
53 	int err;
54 	int npages;
55 	int page_shift;
56 	int ncont;
57 	u32 offset;
58 	u32 uidx = MLX5_IB_DEFAULT_UIDX;
59 
60 	ucmdlen = min(udata->inlen, sizeof(ucmd));
61 
62 	if (ib_copy_from_udata(&ucmd, udata, ucmdlen)) {
63 		mlx5_ib_dbg(dev, "failed copy udata\n");
64 		return -EFAULT;
65 	}
66 
67 	if (ucmd.reserved0 || ucmd.reserved1)
68 		return -EINVAL;
69 
70 	if (udata->inlen > sizeof(ucmd) &&
71 	    !ib_is_udata_cleared(udata, sizeof(ucmd),
72 				 udata->inlen - sizeof(ucmd)))
73 		return -EINVAL;
74 
75 	if (in->type != IB_SRQT_BASIC) {
76 		err = get_srq_user_index(ucontext, &ucmd, udata->inlen, &uidx);
77 		if (err)
78 			return err;
79 	}
80 
81 	srq->wq_sig = !!(ucmd.flags & MLX5_SRQ_FLAG_SIGNATURE);
82 
83 	srq->umem = ib_umem_get(pd->device, ucmd.buf_addr, buf_size, 0);
84 	if (IS_ERR(srq->umem)) {
85 		mlx5_ib_dbg(dev, "failed umem get, size %d\n", buf_size);
86 		err = PTR_ERR(srq->umem);
87 		return err;
88 	}
89 
90 	mlx5_ib_cont_pages(srq->umem, ucmd.buf_addr, 0, &npages,
91 			   &page_shift, &ncont, NULL);
92 	err = mlx5_ib_get_buf_offset(ucmd.buf_addr, page_shift,
93 				     &offset);
94 	if (err) {
95 		mlx5_ib_warn(dev, "bad offset\n");
96 		goto err_umem;
97 	}
98 
99 	in->pas = kvcalloc(ncont, sizeof(*in->pas), GFP_KERNEL);
100 	if (!in->pas) {
101 		err = -ENOMEM;
102 		goto err_umem;
103 	}
104 
105 	mlx5_ib_populate_pas(dev, srq->umem, page_shift, in->pas, 0);
106 
107 	err = mlx5_ib_db_map_user(ucontext, udata, ucmd.db_addr, &srq->db);
108 	if (err) {
109 		mlx5_ib_dbg(dev, "map doorbell failed\n");
110 		goto err_in;
111 	}
112 
113 	in->log_page_size = page_shift - MLX5_ADAPTER_PAGE_SHIFT;
114 	in->page_offset = offset;
115 	in->uid = (in->type != IB_SRQT_XRC) ?  to_mpd(pd)->uid : 0;
116 	if (MLX5_CAP_GEN(dev->mdev, cqe_version) == MLX5_CQE_VERSION_V1 &&
117 	    in->type != IB_SRQT_BASIC)
118 		in->user_index = uidx;
119 
120 	return 0;
121 
122 err_in:
123 	kvfree(in->pas);
124 
125 err_umem:
126 	ib_umem_release(srq->umem);
127 
128 	return err;
129 }
130 
131 static int create_srq_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_srq *srq,
132 			     struct mlx5_srq_attr *in, int buf_size)
133 {
134 	int err;
135 	int i;
136 	struct mlx5_wqe_srq_next_seg *next;
137 
138 	err = mlx5_db_alloc(dev->mdev, &srq->db);
139 	if (err) {
140 		mlx5_ib_warn(dev, "alloc dbell rec failed\n");
141 		return err;
142 	}
143 
144 	if (mlx5_frag_buf_alloc_node(dev->mdev, buf_size, &srq->buf,
145 				     dev->mdev->priv.numa_node)) {
146 		mlx5_ib_dbg(dev, "buf alloc failed\n");
147 		err = -ENOMEM;
148 		goto err_db;
149 	}
150 
151 	mlx5_init_fbc(srq->buf.frags, srq->msrq.wqe_shift, ilog2(srq->msrq.max),
152 		      &srq->fbc);
153 
154 	srq->head    = 0;
155 	srq->tail    = srq->msrq.max - 1;
156 	srq->wqe_ctr = 0;
157 
158 	for (i = 0; i < srq->msrq.max; i++) {
159 		next = get_wqe(srq, i);
160 		next->next_wqe_index =
161 			cpu_to_be16((i + 1) & (srq->msrq.max - 1));
162 	}
163 
164 	mlx5_ib_dbg(dev, "srq->buf.page_shift = %d\n", srq->buf.page_shift);
165 	in->pas = kvcalloc(srq->buf.npages, sizeof(*in->pas), GFP_KERNEL);
166 	if (!in->pas) {
167 		err = -ENOMEM;
168 		goto err_buf;
169 	}
170 	mlx5_fill_page_frag_array(&srq->buf, in->pas);
171 
172 	srq->wrid = kvmalloc_array(srq->msrq.max, sizeof(u64), GFP_KERNEL);
173 	if (!srq->wrid) {
174 		err = -ENOMEM;
175 		goto err_in;
176 	}
177 	srq->wq_sig = 0;
178 
179 	in->log_page_size = srq->buf.page_shift - MLX5_ADAPTER_PAGE_SHIFT;
180 	if (MLX5_CAP_GEN(dev->mdev, cqe_version) == MLX5_CQE_VERSION_V1 &&
181 	    in->type != IB_SRQT_BASIC)
182 		in->user_index = MLX5_IB_DEFAULT_UIDX;
183 
184 	return 0;
185 
186 err_in:
187 	kvfree(in->pas);
188 
189 err_buf:
190 	mlx5_frag_buf_free(dev->mdev, &srq->buf);
191 
192 err_db:
193 	mlx5_db_free(dev->mdev, &srq->db);
194 	return err;
195 }
196 
197 static void destroy_srq_user(struct ib_pd *pd, struct mlx5_ib_srq *srq,
198 			     struct ib_udata *udata)
199 {
200 	mlx5_ib_db_unmap_user(
201 		rdma_udata_to_drv_context(
202 			udata,
203 			struct mlx5_ib_ucontext,
204 			ibucontext),
205 		&srq->db);
206 	ib_umem_release(srq->umem);
207 }
208 
209 
210 static void destroy_srq_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_srq *srq)
211 {
212 	kvfree(srq->wrid);
213 	mlx5_frag_buf_free(dev->mdev, &srq->buf);
214 	mlx5_db_free(dev->mdev, &srq->db);
215 }
216 
217 int mlx5_ib_create_srq(struct ib_srq *ib_srq,
218 		       struct ib_srq_init_attr *init_attr,
219 		       struct ib_udata *udata)
220 {
221 	struct mlx5_ib_dev *dev = to_mdev(ib_srq->device);
222 	struct mlx5_ib_srq *srq = to_msrq(ib_srq);
223 	size_t desc_size;
224 	size_t buf_size;
225 	int err;
226 	struct mlx5_srq_attr in = {};
227 	__u32 max_srq_wqes = 1 << MLX5_CAP_GEN(dev->mdev, log_max_srq_sz);
228 
229 	/* Sanity check SRQ size before proceeding */
230 	if (init_attr->attr.max_wr >= max_srq_wqes) {
231 		mlx5_ib_dbg(dev, "max_wr %d, cap %d\n",
232 			    init_attr->attr.max_wr,
233 			    max_srq_wqes);
234 		return -EINVAL;
235 	}
236 
237 	mutex_init(&srq->mutex);
238 	spin_lock_init(&srq->lock);
239 	srq->msrq.max    = roundup_pow_of_two(init_attr->attr.max_wr + 1);
240 	srq->msrq.max_gs = init_attr->attr.max_sge;
241 
242 	desc_size = sizeof(struct mlx5_wqe_srq_next_seg) +
243 		    srq->msrq.max_gs * sizeof(struct mlx5_wqe_data_seg);
244 	if (desc_size == 0 || srq->msrq.max_gs > desc_size)
245 		return -EINVAL;
246 
247 	desc_size = roundup_pow_of_two(desc_size);
248 	desc_size = max_t(size_t, 32, desc_size);
249 	if (desc_size < sizeof(struct mlx5_wqe_srq_next_seg))
250 		return -EINVAL;
251 
252 	srq->msrq.max_avail_gather = (desc_size - sizeof(struct mlx5_wqe_srq_next_seg)) /
253 		sizeof(struct mlx5_wqe_data_seg);
254 	srq->msrq.wqe_shift = ilog2(desc_size);
255 	buf_size = srq->msrq.max * desc_size;
256 	if (buf_size < desc_size)
257 		return -EINVAL;
258 
259 	in.type = init_attr->srq_type;
260 
261 	if (udata)
262 		err = create_srq_user(ib_srq->pd, srq, &in, udata, buf_size);
263 	else
264 		err = create_srq_kernel(dev, srq, &in, buf_size);
265 
266 	if (err) {
267 		mlx5_ib_warn(dev, "create srq %s failed, err %d\n",
268 			     udata ? "user" : "kernel", err);
269 		return err;
270 	}
271 
272 	in.log_size = ilog2(srq->msrq.max);
273 	in.wqe_shift = srq->msrq.wqe_shift - 4;
274 	if (srq->wq_sig)
275 		in.flags |= MLX5_SRQ_FLAG_WQ_SIG;
276 
277 	if (init_attr->srq_type == IB_SRQT_XRC)
278 		in.xrcd = to_mxrcd(init_attr->ext.xrc.xrcd)->xrcdn;
279 	else
280 		in.xrcd = to_mxrcd(dev->devr.x0)->xrcdn;
281 
282 	if (init_attr->srq_type == IB_SRQT_TM) {
283 		in.tm_log_list_size =
284 			ilog2(init_attr->ext.tag_matching.max_num_tags) + 1;
285 		if (in.tm_log_list_size >
286 		    MLX5_CAP_GEN(dev->mdev, log_tag_matching_list_sz)) {
287 			mlx5_ib_dbg(dev, "TM SRQ max_num_tags exceeding limit\n");
288 			err = -EINVAL;
289 			goto err_usr_kern_srq;
290 		}
291 		in.flags |= MLX5_SRQ_FLAG_RNDV;
292 	}
293 
294 	if (ib_srq_has_cq(init_attr->srq_type))
295 		in.cqn = to_mcq(init_attr->ext.cq)->mcq.cqn;
296 	else
297 		in.cqn = to_mcq(dev->devr.c0)->mcq.cqn;
298 
299 	in.pd = to_mpd(ib_srq->pd)->pdn;
300 	in.db_record = srq->db.dma;
301 	err = mlx5_cmd_create_srq(dev, &srq->msrq, &in);
302 	kvfree(in.pas);
303 	if (err) {
304 		mlx5_ib_dbg(dev, "create SRQ failed, err %d\n", err);
305 		goto err_usr_kern_srq;
306 	}
307 
308 	mlx5_ib_dbg(dev, "create SRQ with srqn 0x%x\n", srq->msrq.srqn);
309 
310 	srq->msrq.event = mlx5_ib_srq_event;
311 	srq->ibsrq.ext.xrc.srq_num = srq->msrq.srqn;
312 
313 	if (udata) {
314 		struct mlx5_ib_create_srq_resp resp = {
315 			.srqn = srq->msrq.srqn,
316 		};
317 
318 		if (ib_copy_to_udata(udata, &resp, min(udata->outlen,
319 				     sizeof(resp)))) {
320 			mlx5_ib_dbg(dev, "copy to user failed\n");
321 			err = -EFAULT;
322 			goto err_core;
323 		}
324 	}
325 
326 	init_attr->attr.max_wr = srq->msrq.max - 1;
327 
328 	return 0;
329 
330 err_core:
331 	mlx5_cmd_destroy_srq(dev, &srq->msrq);
332 
333 err_usr_kern_srq:
334 	if (udata)
335 		destroy_srq_user(ib_srq->pd, srq, udata);
336 	else
337 		destroy_srq_kernel(dev, srq);
338 
339 	return err;
340 }
341 
342 int mlx5_ib_modify_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr,
343 		       enum ib_srq_attr_mask attr_mask, struct ib_udata *udata)
344 {
345 	struct mlx5_ib_dev *dev = to_mdev(ibsrq->device);
346 	struct mlx5_ib_srq *srq = to_msrq(ibsrq);
347 	int ret;
348 
349 	/* We don't support resizing SRQs yet */
350 	if (attr_mask & IB_SRQ_MAX_WR)
351 		return -EINVAL;
352 
353 	if (attr_mask & IB_SRQ_LIMIT) {
354 		if (attr->srq_limit >= srq->msrq.max)
355 			return -EINVAL;
356 
357 		mutex_lock(&srq->mutex);
358 		ret = mlx5_cmd_arm_srq(dev, &srq->msrq, attr->srq_limit, 1);
359 		mutex_unlock(&srq->mutex);
360 
361 		if (ret)
362 			return ret;
363 	}
364 
365 	return 0;
366 }
367 
368 int mlx5_ib_query_srq(struct ib_srq *ibsrq, struct ib_srq_attr *srq_attr)
369 {
370 	struct mlx5_ib_dev *dev = to_mdev(ibsrq->device);
371 	struct mlx5_ib_srq *srq = to_msrq(ibsrq);
372 	int ret;
373 	struct mlx5_srq_attr *out;
374 
375 	out = kzalloc(sizeof(*out), GFP_KERNEL);
376 	if (!out)
377 		return -ENOMEM;
378 
379 	ret = mlx5_cmd_query_srq(dev, &srq->msrq, out);
380 	if (ret)
381 		goto out_box;
382 
383 	srq_attr->srq_limit = out->lwm;
384 	srq_attr->max_wr    = srq->msrq.max - 1;
385 	srq_attr->max_sge   = srq->msrq.max_gs;
386 
387 out_box:
388 	kfree(out);
389 	return ret;
390 }
391 
392 void mlx5_ib_destroy_srq(struct ib_srq *srq, struct ib_udata *udata)
393 {
394 	struct mlx5_ib_dev *dev = to_mdev(srq->device);
395 	struct mlx5_ib_srq *msrq = to_msrq(srq);
396 
397 	mlx5_cmd_destroy_srq(dev, &msrq->msrq);
398 
399 	if (srq->uobject) {
400 		mlx5_ib_db_unmap_user(
401 			rdma_udata_to_drv_context(
402 				udata,
403 				struct mlx5_ib_ucontext,
404 				ibucontext),
405 			&msrq->db);
406 		ib_umem_release(msrq->umem);
407 	} else {
408 		destroy_srq_kernel(dev, msrq);
409 	}
410 }
411 
412 void mlx5_ib_free_srq_wqe(struct mlx5_ib_srq *srq, int wqe_index)
413 {
414 	struct mlx5_wqe_srq_next_seg *next;
415 
416 	/* always called with interrupts disabled. */
417 	spin_lock(&srq->lock);
418 
419 	next = get_wqe(srq, srq->tail);
420 	next->next_wqe_index = cpu_to_be16(wqe_index);
421 	srq->tail = wqe_index;
422 
423 	spin_unlock(&srq->lock);
424 }
425 
426 int mlx5_ib_post_srq_recv(struct ib_srq *ibsrq, const struct ib_recv_wr *wr,
427 			  const struct ib_recv_wr **bad_wr)
428 {
429 	struct mlx5_ib_srq *srq = to_msrq(ibsrq);
430 	struct mlx5_wqe_srq_next_seg *next;
431 	struct mlx5_wqe_data_seg *scat;
432 	struct mlx5_ib_dev *dev = to_mdev(ibsrq->device);
433 	struct mlx5_core_dev *mdev = dev->mdev;
434 	unsigned long flags;
435 	int err = 0;
436 	int nreq;
437 	int i;
438 
439 	spin_lock_irqsave(&srq->lock, flags);
440 
441 	if (mdev->state == MLX5_DEVICE_STATE_INTERNAL_ERROR) {
442 		err = -EIO;
443 		*bad_wr = wr;
444 		goto out;
445 	}
446 
447 	for (nreq = 0; wr; nreq++, wr = wr->next) {
448 		if (unlikely(wr->num_sge > srq->msrq.max_gs)) {
449 			err = -EINVAL;
450 			*bad_wr = wr;
451 			break;
452 		}
453 
454 		if (unlikely(srq->head == srq->tail)) {
455 			err = -ENOMEM;
456 			*bad_wr = wr;
457 			break;
458 		}
459 
460 		srq->wrid[srq->head] = wr->wr_id;
461 
462 		next      = get_wqe(srq, srq->head);
463 		srq->head = be16_to_cpu(next->next_wqe_index);
464 		scat      = (struct mlx5_wqe_data_seg *)(next + 1);
465 
466 		for (i = 0; i < wr->num_sge; i++) {
467 			scat[i].byte_count = cpu_to_be32(wr->sg_list[i].length);
468 			scat[i].lkey       = cpu_to_be32(wr->sg_list[i].lkey);
469 			scat[i].addr       = cpu_to_be64(wr->sg_list[i].addr);
470 		}
471 
472 		if (i < srq->msrq.max_avail_gather) {
473 			scat[i].byte_count = 0;
474 			scat[i].lkey       = cpu_to_be32(MLX5_INVALID_LKEY);
475 			scat[i].addr       = 0;
476 		}
477 	}
478 
479 	if (likely(nreq)) {
480 		srq->wqe_ctr += nreq;
481 
482 		/* Make sure that descriptors are written before
483 		 * doorbell record.
484 		 */
485 		wmb();
486 
487 		*srq->db.db = cpu_to_be32(srq->wqe_ctr);
488 	}
489 out:
490 	spin_unlock_irqrestore(&srq->lock, flags);
491 
492 	return err;
493 }
494