1 // SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB
2 /*
3  * Copyright (c) 2018 Hisilicon Limited.
4  */
5 
6 #include <linux/pci.h>
7 #include <rdma/ib_umem.h>
8 #include "hns_roce_device.h"
9 #include "hns_roce_cmd.h"
10 #include "hns_roce_hem.h"
11 
12 void hns_roce_srq_event(struct hns_roce_dev *hr_dev, u32 srqn, int event_type)
13 {
14 	struct hns_roce_srq_table *srq_table = &hr_dev->srq_table;
15 	struct hns_roce_srq *srq;
16 
17 	xa_lock(&srq_table->xa);
18 	srq = xa_load(&srq_table->xa, srqn & (hr_dev->caps.num_srqs - 1));
19 	if (srq)
20 		refcount_inc(&srq->refcount);
21 	xa_unlock(&srq_table->xa);
22 
23 	if (!srq) {
24 		dev_warn(hr_dev->dev, "Async event for bogus SRQ %08x\n", srqn);
25 		return;
26 	}
27 
28 	srq->event(srq, event_type);
29 
30 	if (refcount_dec_and_test(&srq->refcount))
31 		complete(&srq->free);
32 }
33 
34 static void hns_roce_ib_srq_event(struct hns_roce_srq *srq,
35 				  enum hns_roce_event event_type)
36 {
37 	struct hns_roce_dev *hr_dev = to_hr_dev(srq->ibsrq.device);
38 	struct ib_srq *ibsrq = &srq->ibsrq;
39 	struct ib_event event;
40 
41 	if (ibsrq->event_handler) {
42 		event.device      = ibsrq->device;
43 		event.element.srq = ibsrq;
44 		switch (event_type) {
45 		case HNS_ROCE_EVENT_TYPE_SRQ_LIMIT_REACH:
46 			event.event = IB_EVENT_SRQ_LIMIT_REACHED;
47 			break;
48 		case HNS_ROCE_EVENT_TYPE_SRQ_CATAS_ERROR:
49 			event.event = IB_EVENT_SRQ_ERR;
50 			break;
51 		default:
52 			dev_err(hr_dev->dev,
53 			   "hns_roce:Unexpected event type 0x%x on SRQ %06lx\n",
54 			   event_type, srq->srqn);
55 			return;
56 		}
57 
58 		ibsrq->event_handler(&event, ibsrq->srq_context);
59 	}
60 }
61 
62 static int hns_roce_hw_create_srq(struct hns_roce_dev *dev,
63 				  struct hns_roce_cmd_mailbox *mailbox,
64 				  unsigned long srq_num)
65 {
66 	return hns_roce_cmd_mbox(dev, mailbox->dma, 0, srq_num, 0,
67 				 HNS_ROCE_CMD_CREATE_SRQ,
68 				 HNS_ROCE_CMD_TIMEOUT_MSECS);
69 }
70 
71 static int hns_roce_hw_destroy_srq(struct hns_roce_dev *dev,
72 				   struct hns_roce_cmd_mailbox *mailbox,
73 				   unsigned long srq_num)
74 {
75 	return hns_roce_cmd_mbox(dev, 0, mailbox ? mailbox->dma : 0, srq_num,
76 				 mailbox ? 0 : 1, HNS_ROCE_CMD_DESTROY_SRQ,
77 				 HNS_ROCE_CMD_TIMEOUT_MSECS);
78 }
79 
80 static int alloc_srqc(struct hns_roce_dev *hr_dev, struct hns_roce_srq *srq)
81 {
82 	struct hns_roce_srq_table *srq_table = &hr_dev->srq_table;
83 	struct hns_roce_ida *srq_ida = &hr_dev->srq_table.srq_ida;
84 	struct ib_device *ibdev = &hr_dev->ib_dev;
85 	struct hns_roce_cmd_mailbox *mailbox;
86 	int ret;
87 	int id;
88 
89 	id = ida_alloc_range(&srq_ida->ida, srq_ida->min, srq_ida->max,
90 			     GFP_KERNEL);
91 	if (id < 0) {
92 		ibdev_err(ibdev, "failed to alloc srq(%d).\n", id);
93 		return -ENOMEM;
94 	}
95 	srq->srqn = (unsigned long)id;
96 
97 	ret = hns_roce_table_get(hr_dev, &srq_table->table, srq->srqn);
98 	if (ret) {
99 		ibdev_err(ibdev, "failed to get SRQC table, ret = %d.\n", ret);
100 		goto err_out;
101 	}
102 
103 	ret = xa_err(xa_store(&srq_table->xa, srq->srqn, srq, GFP_KERNEL));
104 	if (ret) {
105 		ibdev_err(ibdev, "failed to store SRQC, ret = %d.\n", ret);
106 		goto err_put;
107 	}
108 
109 	mailbox = hns_roce_alloc_cmd_mailbox(hr_dev);
110 	if (IS_ERR_OR_NULL(mailbox)) {
111 		ibdev_err(ibdev, "failed to alloc mailbox for SRQC.\n");
112 		ret = -ENOMEM;
113 		goto err_xa;
114 	}
115 
116 	ret = hr_dev->hw->write_srqc(srq, mailbox->buf);
117 	if (ret) {
118 		ibdev_err(ibdev, "failed to write SRQC.\n");
119 		goto err_mbox;
120 	}
121 
122 	ret = hns_roce_hw_create_srq(hr_dev, mailbox, srq->srqn);
123 	if (ret) {
124 		ibdev_err(ibdev, "failed to config SRQC, ret = %d.\n", ret);
125 		goto err_mbox;
126 	}
127 
128 	hns_roce_free_cmd_mailbox(hr_dev, mailbox);
129 
130 	return 0;
131 
132 err_mbox:
133 	hns_roce_free_cmd_mailbox(hr_dev, mailbox);
134 err_xa:
135 	xa_erase(&srq_table->xa, srq->srqn);
136 err_put:
137 	hns_roce_table_put(hr_dev, &srq_table->table, srq->srqn);
138 err_out:
139 	ida_free(&srq_ida->ida, id);
140 
141 	return ret;
142 }
143 
144 static void free_srqc(struct hns_roce_dev *hr_dev, struct hns_roce_srq *srq)
145 {
146 	struct hns_roce_srq_table *srq_table = &hr_dev->srq_table;
147 	int ret;
148 
149 	ret = hns_roce_hw_destroy_srq(hr_dev, NULL, srq->srqn);
150 	if (ret)
151 		dev_err(hr_dev->dev, "DESTROY_SRQ failed (%d) for SRQN %06lx\n",
152 			ret, srq->srqn);
153 
154 	xa_erase(&srq_table->xa, srq->srqn);
155 
156 	if (refcount_dec_and_test(&srq->refcount))
157 		complete(&srq->free);
158 	wait_for_completion(&srq->free);
159 
160 	hns_roce_table_put(hr_dev, &srq_table->table, srq->srqn);
161 	ida_free(&srq_table->srq_ida.ida, (int)srq->srqn);
162 }
163 
164 static int alloc_srq_idx(struct hns_roce_dev *hr_dev, struct hns_roce_srq *srq,
165 			 struct ib_udata *udata, unsigned long addr)
166 {
167 	struct hns_roce_idx_que *idx_que = &srq->idx_que;
168 	struct ib_device *ibdev = &hr_dev->ib_dev;
169 	struct hns_roce_buf_attr buf_attr = {};
170 	int ret;
171 
172 	srq->idx_que.entry_shift = ilog2(HNS_ROCE_IDX_QUE_ENTRY_SZ);
173 
174 	buf_attr.page_shift = hr_dev->caps.idx_buf_pg_sz + PAGE_SHIFT;
175 	buf_attr.region[0].size = to_hr_hem_entries_size(srq->wqe_cnt,
176 					srq->idx_que.entry_shift);
177 	buf_attr.region[0].hopnum = hr_dev->caps.idx_hop_num;
178 	buf_attr.region_count = 1;
179 
180 	ret = hns_roce_mtr_create(hr_dev, &idx_que->mtr, &buf_attr,
181 				  hr_dev->caps.idx_ba_pg_sz + PAGE_SHIFT,
182 				  udata, addr);
183 	if (ret) {
184 		ibdev_err(ibdev,
185 			  "failed to alloc SRQ idx mtr, ret = %d.\n", ret);
186 		return ret;
187 	}
188 
189 	if (!udata) {
190 		idx_que->bitmap = bitmap_zalloc(srq->wqe_cnt, GFP_KERNEL);
191 		if (!idx_que->bitmap) {
192 			ibdev_err(ibdev, "failed to alloc SRQ idx bitmap.\n");
193 			ret = -ENOMEM;
194 			goto err_idx_mtr;
195 		}
196 	}
197 
198 	idx_que->head = 0;
199 	idx_que->tail = 0;
200 
201 	return 0;
202 err_idx_mtr:
203 	hns_roce_mtr_destroy(hr_dev, &idx_que->mtr);
204 
205 	return ret;
206 }
207 
208 static void free_srq_idx(struct hns_roce_dev *hr_dev, struct hns_roce_srq *srq)
209 {
210 	struct hns_roce_idx_que *idx_que = &srq->idx_que;
211 
212 	bitmap_free(idx_que->bitmap);
213 	idx_que->bitmap = NULL;
214 	hns_roce_mtr_destroy(hr_dev, &idx_que->mtr);
215 }
216 
217 static int alloc_srq_wqe_buf(struct hns_roce_dev *hr_dev,
218 			     struct hns_roce_srq *srq,
219 			     struct ib_udata *udata, unsigned long addr)
220 {
221 	struct ib_device *ibdev = &hr_dev->ib_dev;
222 	struct hns_roce_buf_attr buf_attr = {};
223 	int ret;
224 
225 	srq->wqe_shift = ilog2(roundup_pow_of_two(max(HNS_ROCE_SGE_SIZE,
226 						      HNS_ROCE_SGE_SIZE *
227 						      srq->max_gs)));
228 
229 	buf_attr.page_shift = hr_dev->caps.srqwqe_buf_pg_sz + PAGE_SHIFT;
230 	buf_attr.region[0].size = to_hr_hem_entries_size(srq->wqe_cnt,
231 							 srq->wqe_shift);
232 	buf_attr.region[0].hopnum = hr_dev->caps.srqwqe_hop_num;
233 	buf_attr.region_count = 1;
234 
235 	ret = hns_roce_mtr_create(hr_dev, &srq->buf_mtr, &buf_attr,
236 				  hr_dev->caps.srqwqe_ba_pg_sz + PAGE_SHIFT,
237 				  udata, addr);
238 	if (ret)
239 		ibdev_err(ibdev,
240 			  "failed to alloc SRQ buf mtr, ret = %d.\n", ret);
241 
242 	return ret;
243 }
244 
245 static void free_srq_wqe_buf(struct hns_roce_dev *hr_dev,
246 			     struct hns_roce_srq *srq)
247 {
248 	hns_roce_mtr_destroy(hr_dev, &srq->buf_mtr);
249 }
250 
251 static int alloc_srq_wrid(struct hns_roce_dev *hr_dev, struct hns_roce_srq *srq)
252 {
253 	srq->wrid = kvmalloc_array(srq->wqe_cnt, sizeof(u64), GFP_KERNEL);
254 	if (!srq->wrid)
255 		return -ENOMEM;
256 
257 	return 0;
258 }
259 
260 static void free_srq_wrid(struct hns_roce_srq *srq)
261 {
262 	kfree(srq->wrid);
263 	srq->wrid = NULL;
264 }
265 
266 static u32 proc_srq_sge(struct hns_roce_dev *dev, struct hns_roce_srq *hr_srq,
267 			bool user)
268 {
269 	u32 max_sge = dev->caps.max_srq_sges;
270 
271 	if (dev->pci_dev->revision >= PCI_REVISION_ID_HIP09)
272 		return max_sge;
273 
274 	/* Reserve SGEs only for HIP08 in kernel; The userspace driver will
275 	 * calculate number of max_sge with reserved SGEs when allocating wqe
276 	 * buf, so there is no need to do this again in kernel. But the number
277 	 * may exceed the capacity of SGEs recorded in the firmware, so the
278 	 * kernel driver should just adapt the value accordingly.
279 	 */
280 	if (user)
281 		max_sge = roundup_pow_of_two(max_sge + 1);
282 	else
283 		hr_srq->rsv_sge = 1;
284 
285 	return max_sge;
286 }
287 
288 static int set_srq_basic_param(struct hns_roce_srq *srq,
289 			       struct ib_srq_init_attr *init_attr,
290 			       struct ib_udata *udata)
291 {
292 	struct hns_roce_dev *hr_dev = to_hr_dev(srq->ibsrq.device);
293 	struct ib_srq_attr *attr = &init_attr->attr;
294 	u32 max_sge;
295 
296 	max_sge = proc_srq_sge(hr_dev, srq, !!udata);
297 	if (attr->max_wr > hr_dev->caps.max_srq_wrs ||
298 	    attr->max_sge > max_sge) {
299 		ibdev_err(&hr_dev->ib_dev,
300 			  "invalid SRQ attr, depth = %u, sge = %u.\n",
301 			  attr->max_wr, attr->max_sge);
302 		return -EINVAL;
303 	}
304 
305 	attr->max_wr = max_t(u32, attr->max_wr, HNS_ROCE_MIN_SRQ_WQE_NUM);
306 	srq->wqe_cnt = roundup_pow_of_two(attr->max_wr);
307 	srq->max_gs = roundup_pow_of_two(attr->max_sge + srq->rsv_sge);
308 
309 	attr->max_wr = srq->wqe_cnt;
310 	attr->max_sge = srq->max_gs - srq->rsv_sge;
311 	attr->srq_limit = 0;
312 
313 	return 0;
314 }
315 
316 static void set_srq_ext_param(struct hns_roce_srq *srq,
317 			      struct ib_srq_init_attr *init_attr)
318 {
319 	srq->cqn = ib_srq_has_cq(init_attr->srq_type) ?
320 		   to_hr_cq(init_attr->ext.cq)->cqn : 0;
321 
322 	srq->xrcdn = (init_attr->srq_type == IB_SRQT_XRC) ?
323 		     to_hr_xrcd(init_attr->ext.xrc.xrcd)->xrcdn : 0;
324 }
325 
326 static int set_srq_param(struct hns_roce_srq *srq,
327 			 struct ib_srq_init_attr *init_attr,
328 			 struct ib_udata *udata)
329 {
330 	int ret;
331 
332 	ret = set_srq_basic_param(srq, init_attr, udata);
333 	if (ret)
334 		return ret;
335 
336 	set_srq_ext_param(srq, init_attr);
337 
338 	return 0;
339 }
340 
341 static int alloc_srq_buf(struct hns_roce_dev *hr_dev, struct hns_roce_srq *srq,
342 			 struct ib_udata *udata)
343 {
344 	struct hns_roce_ib_create_srq ucmd = {};
345 	int ret;
346 
347 	if (udata) {
348 		ret = ib_copy_from_udata(&ucmd, udata,
349 					 min(udata->inlen, sizeof(ucmd)));
350 		if (ret) {
351 			ibdev_err(&hr_dev->ib_dev,
352 				  "failed to copy SRQ udata, ret = %d.\n",
353 				  ret);
354 			return ret;
355 		}
356 	}
357 
358 	ret = alloc_srq_idx(hr_dev, srq, udata, ucmd.que_addr);
359 	if (ret)
360 		return ret;
361 
362 	ret = alloc_srq_wqe_buf(hr_dev, srq, udata, ucmd.buf_addr);
363 	if (ret)
364 		goto err_idx;
365 
366 	if (!udata) {
367 		ret = alloc_srq_wrid(hr_dev, srq);
368 		if (ret)
369 			goto err_wqe_buf;
370 	}
371 
372 	return 0;
373 
374 err_wqe_buf:
375 	free_srq_wqe_buf(hr_dev, srq);
376 err_idx:
377 	free_srq_idx(hr_dev, srq);
378 
379 	return ret;
380 }
381 
382 static void free_srq_buf(struct hns_roce_dev *hr_dev, struct hns_roce_srq *srq)
383 {
384 	free_srq_wrid(srq);
385 	free_srq_wqe_buf(hr_dev, srq);
386 	free_srq_idx(hr_dev, srq);
387 }
388 
389 int hns_roce_create_srq(struct ib_srq *ib_srq,
390 			struct ib_srq_init_attr *init_attr,
391 			struct ib_udata *udata)
392 {
393 	struct hns_roce_dev *hr_dev = to_hr_dev(ib_srq->device);
394 	struct hns_roce_ib_create_srq_resp resp = {};
395 	struct hns_roce_srq *srq = to_hr_srq(ib_srq);
396 	int ret;
397 
398 	mutex_init(&srq->mutex);
399 	spin_lock_init(&srq->lock);
400 
401 	ret = set_srq_param(srq, init_attr, udata);
402 	if (ret)
403 		return ret;
404 
405 	ret = alloc_srq_buf(hr_dev, srq, udata);
406 	if (ret)
407 		return ret;
408 
409 	ret = alloc_srqc(hr_dev, srq);
410 	if (ret)
411 		goto err_srq_buf;
412 
413 	if (udata) {
414 		resp.srqn = srq->srqn;
415 		if (ib_copy_to_udata(udata, &resp,
416 				     min(udata->outlen, sizeof(resp)))) {
417 			ret = -EFAULT;
418 			goto err_srqc;
419 		}
420 	}
421 
422 	srq->db_reg = hr_dev->reg_base + SRQ_DB_REG;
423 	srq->event = hns_roce_ib_srq_event;
424 	refcount_set(&srq->refcount, 1);
425 	init_completion(&srq->free);
426 
427 	return 0;
428 
429 err_srqc:
430 	free_srqc(hr_dev, srq);
431 err_srq_buf:
432 	free_srq_buf(hr_dev, srq);
433 
434 	return ret;
435 }
436 
437 int hns_roce_destroy_srq(struct ib_srq *ibsrq, struct ib_udata *udata)
438 {
439 	struct hns_roce_dev *hr_dev = to_hr_dev(ibsrq->device);
440 	struct hns_roce_srq *srq = to_hr_srq(ibsrq);
441 
442 	free_srqc(hr_dev, srq);
443 	free_srq_buf(hr_dev, srq);
444 	return 0;
445 }
446 
447 void hns_roce_init_srq_table(struct hns_roce_dev *hr_dev)
448 {
449 	struct hns_roce_srq_table *srq_table = &hr_dev->srq_table;
450 	struct hns_roce_ida *srq_ida = &srq_table->srq_ida;
451 
452 	xa_init(&srq_table->xa);
453 
454 	ida_init(&srq_ida->ida);
455 	srq_ida->max = hr_dev->caps.num_srqs - 1;
456 	srq_ida->min = hr_dev->caps.reserved_srqs;
457 }
458