xref: /linux/drivers/infiniband/hw/hns/hns_roce_hem.c (revision 55d0969c451159cff86949b38c39171cab962069)
1 /*
2  * Copyright (c) 2016 Hisilicon Limited.
3  * Copyright (c) 2007, 2008 Mellanox Technologies. All rights reserved.
4  *
5  * This software is available to you under a choice of one of two
6  * licenses.  You may choose to be licensed under the terms of the GNU
7  * General Public License (GPL) Version 2, available from the file
8  * COPYING in the main directory of this source tree, or the
9  * OpenIB.org BSD license below:
10  *
11  *     Redistribution and use in source and binary forms, with or
12  *     without modification, are permitted provided that the following
13  *     conditions are met:
14  *
15  *      - Redistributions of source code must retain the above
16  *        copyright notice, this list of conditions and the following
17  *        disclaimer.
18  *
19  *      - Redistributions in binary form must reproduce the above
20  *        copyright notice, this list of conditions and the following
21  *        disclaimer in the documentation and/or other materials
22  *        provided with the distribution.
23  *
24  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
25  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
26  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
27  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
28  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
29  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
30  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
31  * SOFTWARE.
32  */
33 
34 #include "hns_roce_device.h"
35 #include "hns_roce_hem.h"
36 #include "hns_roce_common.h"
37 
38 #define HEM_INDEX_BUF			BIT(0)
39 #define HEM_INDEX_L0			BIT(1)
40 #define HEM_INDEX_L1			BIT(2)
41 struct hns_roce_hem_index {
42 	u64 buf;
43 	u64 l0;
44 	u64 l1;
45 	u32 inited; /* indicate which index is available */
46 };
47 
48 bool hns_roce_check_whether_mhop(struct hns_roce_dev *hr_dev, u32 type)
49 {
50 	int hop_num = 0;
51 
52 	switch (type) {
53 	case HEM_TYPE_QPC:
54 		hop_num = hr_dev->caps.qpc_hop_num;
55 		break;
56 	case HEM_TYPE_MTPT:
57 		hop_num = hr_dev->caps.mpt_hop_num;
58 		break;
59 	case HEM_TYPE_CQC:
60 		hop_num = hr_dev->caps.cqc_hop_num;
61 		break;
62 	case HEM_TYPE_SRQC:
63 		hop_num = hr_dev->caps.srqc_hop_num;
64 		break;
65 	case HEM_TYPE_SCCC:
66 		hop_num = hr_dev->caps.sccc_hop_num;
67 		break;
68 	case HEM_TYPE_QPC_TIMER:
69 		hop_num = hr_dev->caps.qpc_timer_hop_num;
70 		break;
71 	case HEM_TYPE_CQC_TIMER:
72 		hop_num = hr_dev->caps.cqc_timer_hop_num;
73 		break;
74 	case HEM_TYPE_GMV:
75 		hop_num = hr_dev->caps.gmv_hop_num;
76 		break;
77 	default:
78 		return false;
79 	}
80 
81 	return hop_num;
82 }
83 
84 static bool hns_roce_check_hem_null(struct hns_roce_hem **hem, u64 hem_idx,
85 				    u32 bt_chunk_num, u64 hem_max_num)
86 {
87 	u64 start_idx = round_down(hem_idx, bt_chunk_num);
88 	u64 check_max_num = start_idx + bt_chunk_num;
89 	u64 i;
90 
91 	for (i = start_idx; (i < check_max_num) && (i < hem_max_num); i++)
92 		if (i != hem_idx && hem[i])
93 			return false;
94 
95 	return true;
96 }
97 
98 static bool hns_roce_check_bt_null(u64 **bt, u64 ba_idx, u32 bt_chunk_num)
99 {
100 	u64 start_idx = round_down(ba_idx, bt_chunk_num);
101 	int i;
102 
103 	for (i = 0; i < bt_chunk_num; i++)
104 		if (i != ba_idx && bt[start_idx + i])
105 			return false;
106 
107 	return true;
108 }
109 
110 static int hns_roce_get_bt_num(u32 table_type, u32 hop_num)
111 {
112 	if (check_whether_bt_num_3(table_type, hop_num))
113 		return 3;
114 	else if (check_whether_bt_num_2(table_type, hop_num))
115 		return 2;
116 	else if (check_whether_bt_num_1(table_type, hop_num))
117 		return 1;
118 	else
119 		return 0;
120 }
121 
122 static int get_hem_table_config(struct hns_roce_dev *hr_dev,
123 				struct hns_roce_hem_mhop *mhop,
124 				u32 type)
125 {
126 	struct device *dev = hr_dev->dev;
127 
128 	switch (type) {
129 	case HEM_TYPE_QPC:
130 		mhop->buf_chunk_size = 1 << (hr_dev->caps.qpc_buf_pg_sz
131 					     + PAGE_SHIFT);
132 		mhop->bt_chunk_size = 1 << (hr_dev->caps.qpc_ba_pg_sz
133 					     + PAGE_SHIFT);
134 		mhop->ba_l0_num = hr_dev->caps.qpc_bt_num;
135 		mhop->hop_num = hr_dev->caps.qpc_hop_num;
136 		break;
137 	case HEM_TYPE_MTPT:
138 		mhop->buf_chunk_size = 1 << (hr_dev->caps.mpt_buf_pg_sz
139 					     + PAGE_SHIFT);
140 		mhop->bt_chunk_size = 1 << (hr_dev->caps.mpt_ba_pg_sz
141 					     + PAGE_SHIFT);
142 		mhop->ba_l0_num = hr_dev->caps.mpt_bt_num;
143 		mhop->hop_num = hr_dev->caps.mpt_hop_num;
144 		break;
145 	case HEM_TYPE_CQC:
146 		mhop->buf_chunk_size = 1 << (hr_dev->caps.cqc_buf_pg_sz
147 					     + PAGE_SHIFT);
148 		mhop->bt_chunk_size = 1 << (hr_dev->caps.cqc_ba_pg_sz
149 					    + PAGE_SHIFT);
150 		mhop->ba_l0_num = hr_dev->caps.cqc_bt_num;
151 		mhop->hop_num = hr_dev->caps.cqc_hop_num;
152 		break;
153 	case HEM_TYPE_SCCC:
154 		mhop->buf_chunk_size = 1 << (hr_dev->caps.sccc_buf_pg_sz
155 					     + PAGE_SHIFT);
156 		mhop->bt_chunk_size = 1 << (hr_dev->caps.sccc_ba_pg_sz
157 					    + PAGE_SHIFT);
158 		mhop->ba_l0_num = hr_dev->caps.sccc_bt_num;
159 		mhop->hop_num = hr_dev->caps.sccc_hop_num;
160 		break;
161 	case HEM_TYPE_QPC_TIMER:
162 		mhop->buf_chunk_size = 1 << (hr_dev->caps.qpc_timer_buf_pg_sz
163 					     + PAGE_SHIFT);
164 		mhop->bt_chunk_size = 1 << (hr_dev->caps.qpc_timer_ba_pg_sz
165 					    + PAGE_SHIFT);
166 		mhop->ba_l0_num = hr_dev->caps.qpc_timer_bt_num;
167 		mhop->hop_num = hr_dev->caps.qpc_timer_hop_num;
168 		break;
169 	case HEM_TYPE_CQC_TIMER:
170 		mhop->buf_chunk_size = 1 << (hr_dev->caps.cqc_timer_buf_pg_sz
171 					     + PAGE_SHIFT);
172 		mhop->bt_chunk_size = 1 << (hr_dev->caps.cqc_timer_ba_pg_sz
173 					    + PAGE_SHIFT);
174 		mhop->ba_l0_num = hr_dev->caps.cqc_timer_bt_num;
175 		mhop->hop_num = hr_dev->caps.cqc_timer_hop_num;
176 		break;
177 	case HEM_TYPE_SRQC:
178 		mhop->buf_chunk_size = 1 << (hr_dev->caps.srqc_buf_pg_sz
179 					     + PAGE_SHIFT);
180 		mhop->bt_chunk_size = 1 << (hr_dev->caps.srqc_ba_pg_sz
181 					     + PAGE_SHIFT);
182 		mhop->ba_l0_num = hr_dev->caps.srqc_bt_num;
183 		mhop->hop_num = hr_dev->caps.srqc_hop_num;
184 		break;
185 	case HEM_TYPE_GMV:
186 		mhop->buf_chunk_size = 1 << (hr_dev->caps.gmv_buf_pg_sz +
187 					     PAGE_SHIFT);
188 		mhop->bt_chunk_size = 1 << (hr_dev->caps.gmv_ba_pg_sz +
189 					    PAGE_SHIFT);
190 		mhop->ba_l0_num = hr_dev->caps.gmv_bt_num;
191 		mhop->hop_num = hr_dev->caps.gmv_hop_num;
192 		break;
193 	default:
194 		dev_err(dev, "table %u not support multi-hop addressing!\n",
195 			type);
196 		return -EINVAL;
197 	}
198 
199 	return 0;
200 }
201 
202 int hns_roce_calc_hem_mhop(struct hns_roce_dev *hr_dev,
203 			   struct hns_roce_hem_table *table, unsigned long *obj,
204 			   struct hns_roce_hem_mhop *mhop)
205 {
206 	struct device *dev = hr_dev->dev;
207 	u32 chunk_ba_num;
208 	u32 chunk_size;
209 	u32 table_idx;
210 	u32 bt_num;
211 
212 	if (get_hem_table_config(hr_dev, mhop, table->type))
213 		return -EINVAL;
214 
215 	if (!obj)
216 		return 0;
217 
218 	/*
219 	 * QPC/MTPT/CQC/SRQC/SCCC alloc hem for buffer pages.
220 	 * MTT/CQE alloc hem for bt pages.
221 	 */
222 	bt_num = hns_roce_get_bt_num(table->type, mhop->hop_num);
223 	chunk_ba_num = mhop->bt_chunk_size / BA_BYTE_LEN;
224 	chunk_size = table->type < HEM_TYPE_MTT ? mhop->buf_chunk_size :
225 			      mhop->bt_chunk_size;
226 	table_idx = *obj / (chunk_size / table->obj_size);
227 	switch (bt_num) {
228 	case 3:
229 		mhop->l2_idx = table_idx & (chunk_ba_num - 1);
230 		mhop->l1_idx = table_idx / chunk_ba_num & (chunk_ba_num - 1);
231 		mhop->l0_idx = (table_idx / chunk_ba_num) / chunk_ba_num;
232 		break;
233 	case 2:
234 		mhop->l1_idx = table_idx & (chunk_ba_num - 1);
235 		mhop->l0_idx = table_idx / chunk_ba_num;
236 		break;
237 	case 1:
238 		mhop->l0_idx = table_idx;
239 		break;
240 	default:
241 		dev_err(dev, "table %u not support hop_num = %u!\n",
242 			table->type, mhop->hop_num);
243 		return -EINVAL;
244 	}
245 	if (mhop->l0_idx >= mhop->ba_l0_num)
246 		mhop->l0_idx %= mhop->ba_l0_num;
247 
248 	return 0;
249 }
250 
251 static struct hns_roce_hem *hns_roce_alloc_hem(struct hns_roce_dev *hr_dev,
252 					       unsigned long hem_alloc_size,
253 					       gfp_t gfp_mask)
254 {
255 	struct hns_roce_hem *hem;
256 	int order;
257 	void *buf;
258 
259 	WARN_ON(gfp_mask & __GFP_HIGHMEM);
260 
261 	order = get_order(hem_alloc_size);
262 	if (PAGE_SIZE << order != hem_alloc_size) {
263 		dev_err(hr_dev->dev, "invalid hem_alloc_size: %lu!\n",
264 			hem_alloc_size);
265 		return NULL;
266 	}
267 
268 	hem = kmalloc(sizeof(*hem),
269 		      gfp_mask & ~(__GFP_HIGHMEM | __GFP_NOWARN));
270 	if (!hem)
271 		return NULL;
272 
273 	buf = dma_alloc_coherent(hr_dev->dev, hem_alloc_size,
274 				 &hem->dma, gfp_mask);
275 	if (!buf)
276 		goto fail;
277 
278 	hem->buf = buf;
279 	hem->size = hem_alloc_size;
280 
281 	return hem;
282 
283 fail:
284 	kfree(hem);
285 	return NULL;
286 }
287 
288 void hns_roce_free_hem(struct hns_roce_dev *hr_dev, struct hns_roce_hem *hem)
289 {
290 	if (!hem)
291 		return;
292 
293 	dma_free_coherent(hr_dev->dev, hem->size, hem->buf, hem->dma);
294 
295 	kfree(hem);
296 }
297 
298 static int calc_hem_config(struct hns_roce_dev *hr_dev,
299 			   struct hns_roce_hem_table *table, unsigned long obj,
300 			   struct hns_roce_hem_mhop *mhop,
301 			   struct hns_roce_hem_index *index)
302 {
303 	struct ib_device *ibdev = &hr_dev->ib_dev;
304 	unsigned long mhop_obj = obj;
305 	u32 l0_idx, l1_idx, l2_idx;
306 	u32 chunk_ba_num;
307 	u32 bt_num;
308 	int ret;
309 
310 	ret = hns_roce_calc_hem_mhop(hr_dev, table, &mhop_obj, mhop);
311 	if (ret)
312 		return ret;
313 
314 	l0_idx = mhop->l0_idx;
315 	l1_idx = mhop->l1_idx;
316 	l2_idx = mhop->l2_idx;
317 	chunk_ba_num = mhop->bt_chunk_size / BA_BYTE_LEN;
318 	bt_num = hns_roce_get_bt_num(table->type, mhop->hop_num);
319 	switch (bt_num) {
320 	case 3:
321 		index->l1 = l0_idx * chunk_ba_num + l1_idx;
322 		index->l0 = l0_idx;
323 		index->buf = l0_idx * chunk_ba_num * chunk_ba_num +
324 			     l1_idx * chunk_ba_num + l2_idx;
325 		break;
326 	case 2:
327 		index->l0 = l0_idx;
328 		index->buf = l0_idx * chunk_ba_num + l1_idx;
329 		break;
330 	case 1:
331 		index->buf = l0_idx;
332 		break;
333 	default:
334 		ibdev_err(ibdev, "table %u not support mhop.hop_num = %u!\n",
335 			  table->type, mhop->hop_num);
336 		return -EINVAL;
337 	}
338 
339 	if (unlikely(index->buf >= table->num_hem)) {
340 		ibdev_err(ibdev, "table %u exceed hem limt idx %llu, max %lu!\n",
341 			  table->type, index->buf, table->num_hem);
342 		return -EINVAL;
343 	}
344 
345 	return 0;
346 }
347 
348 static void free_mhop_hem(struct hns_roce_dev *hr_dev,
349 			  struct hns_roce_hem_table *table,
350 			  struct hns_roce_hem_mhop *mhop,
351 			  struct hns_roce_hem_index *index)
352 {
353 	u32 bt_size = mhop->bt_chunk_size;
354 	struct device *dev = hr_dev->dev;
355 
356 	if (index->inited & HEM_INDEX_BUF) {
357 		hns_roce_free_hem(hr_dev, table->hem[index->buf]);
358 		table->hem[index->buf] = NULL;
359 	}
360 
361 	if (index->inited & HEM_INDEX_L1) {
362 		dma_free_coherent(dev, bt_size, table->bt_l1[index->l1],
363 				  table->bt_l1_dma_addr[index->l1]);
364 		table->bt_l1[index->l1] = NULL;
365 	}
366 
367 	if (index->inited & HEM_INDEX_L0) {
368 		dma_free_coherent(dev, bt_size, table->bt_l0[index->l0],
369 				  table->bt_l0_dma_addr[index->l0]);
370 		table->bt_l0[index->l0] = NULL;
371 	}
372 }
373 
374 static int alloc_mhop_hem(struct hns_roce_dev *hr_dev,
375 			  struct hns_roce_hem_table *table,
376 			  struct hns_roce_hem_mhop *mhop,
377 			  struct hns_roce_hem_index *index)
378 {
379 	u32 bt_size = mhop->bt_chunk_size;
380 	struct device *dev = hr_dev->dev;
381 	gfp_t flag;
382 	u64 bt_ba;
383 	u32 size;
384 	int ret;
385 
386 	/* alloc L1 BA's chunk */
387 	if ((check_whether_bt_num_3(table->type, mhop->hop_num) ||
388 	     check_whether_bt_num_2(table->type, mhop->hop_num)) &&
389 	     !table->bt_l0[index->l0]) {
390 		table->bt_l0[index->l0] = dma_alloc_coherent(dev, bt_size,
391 					    &table->bt_l0_dma_addr[index->l0],
392 					    GFP_KERNEL);
393 		if (!table->bt_l0[index->l0]) {
394 			ret = -ENOMEM;
395 			goto out;
396 		}
397 		index->inited |= HEM_INDEX_L0;
398 	}
399 
400 	/* alloc L2 BA's chunk */
401 	if (check_whether_bt_num_3(table->type, mhop->hop_num) &&
402 	    !table->bt_l1[index->l1])  {
403 		table->bt_l1[index->l1] = dma_alloc_coherent(dev, bt_size,
404 					    &table->bt_l1_dma_addr[index->l1],
405 					    GFP_KERNEL);
406 		if (!table->bt_l1[index->l1]) {
407 			ret = -ENOMEM;
408 			goto err_alloc_hem;
409 		}
410 		index->inited |= HEM_INDEX_L1;
411 		*(table->bt_l0[index->l0] + mhop->l1_idx) =
412 					       table->bt_l1_dma_addr[index->l1];
413 	}
414 
415 	/*
416 	 * alloc buffer space chunk for QPC/MTPT/CQC/SRQC/SCCC.
417 	 * alloc bt space chunk for MTT/CQE.
418 	 */
419 	size = table->type < HEM_TYPE_MTT ? mhop->buf_chunk_size : bt_size;
420 	flag = GFP_KERNEL | __GFP_NOWARN;
421 	table->hem[index->buf] = hns_roce_alloc_hem(hr_dev, size, flag);
422 	if (!table->hem[index->buf]) {
423 		ret = -ENOMEM;
424 		goto err_alloc_hem;
425 	}
426 
427 	index->inited |= HEM_INDEX_BUF;
428 	bt_ba = table->hem[index->buf]->dma;
429 
430 	if (table->type < HEM_TYPE_MTT) {
431 		if (mhop->hop_num == 2)
432 			*(table->bt_l1[index->l1] + mhop->l2_idx) = bt_ba;
433 		else if (mhop->hop_num == 1)
434 			*(table->bt_l0[index->l0] + mhop->l1_idx) = bt_ba;
435 	} else if (mhop->hop_num == 2) {
436 		*(table->bt_l0[index->l0] + mhop->l1_idx) = bt_ba;
437 	}
438 
439 	return 0;
440 err_alloc_hem:
441 	free_mhop_hem(hr_dev, table, mhop, index);
442 out:
443 	return ret;
444 }
445 
446 static int set_mhop_hem(struct hns_roce_dev *hr_dev,
447 			struct hns_roce_hem_table *table, unsigned long obj,
448 			struct hns_roce_hem_mhop *mhop,
449 			struct hns_roce_hem_index *index)
450 {
451 	struct ib_device *ibdev = &hr_dev->ib_dev;
452 	u32 step_idx;
453 	int ret = 0;
454 
455 	if (index->inited & HEM_INDEX_L0) {
456 		ret = hr_dev->hw->set_hem(hr_dev, table, obj, 0);
457 		if (ret) {
458 			ibdev_err(ibdev, "set HEM step 0 failed!\n");
459 			goto out;
460 		}
461 	}
462 
463 	if (index->inited & HEM_INDEX_L1) {
464 		ret = hr_dev->hw->set_hem(hr_dev, table, obj, 1);
465 		if (ret) {
466 			ibdev_err(ibdev, "set HEM step 1 failed!\n");
467 			goto out;
468 		}
469 	}
470 
471 	if (index->inited & HEM_INDEX_BUF) {
472 		if (mhop->hop_num == HNS_ROCE_HOP_NUM_0)
473 			step_idx = 0;
474 		else
475 			step_idx = mhop->hop_num;
476 		ret = hr_dev->hw->set_hem(hr_dev, table, obj, step_idx);
477 		if (ret)
478 			ibdev_err(ibdev, "set HEM step last failed!\n");
479 	}
480 out:
481 	return ret;
482 }
483 
484 static int hns_roce_table_mhop_get(struct hns_roce_dev *hr_dev,
485 				   struct hns_roce_hem_table *table,
486 				   unsigned long obj)
487 {
488 	struct ib_device *ibdev = &hr_dev->ib_dev;
489 	struct hns_roce_hem_index index = {};
490 	struct hns_roce_hem_mhop mhop = {};
491 	int ret;
492 
493 	ret = calc_hem_config(hr_dev, table, obj, &mhop, &index);
494 	if (ret) {
495 		ibdev_err(ibdev, "calc hem config failed!\n");
496 		return ret;
497 	}
498 
499 	mutex_lock(&table->mutex);
500 	if (table->hem[index.buf]) {
501 		refcount_inc(&table->hem[index.buf]->refcount);
502 		goto out;
503 	}
504 
505 	ret = alloc_mhop_hem(hr_dev, table, &mhop, &index);
506 	if (ret) {
507 		ibdev_err(ibdev, "alloc mhop hem failed!\n");
508 		goto out;
509 	}
510 
511 	/* set HEM base address to hardware */
512 	if (table->type < HEM_TYPE_MTT) {
513 		ret = set_mhop_hem(hr_dev, table, obj, &mhop, &index);
514 		if (ret) {
515 			ibdev_err(ibdev, "set HEM address to HW failed!\n");
516 			goto err_alloc;
517 		}
518 	}
519 
520 	refcount_set(&table->hem[index.buf]->refcount, 1);
521 	goto out;
522 
523 err_alloc:
524 	free_mhop_hem(hr_dev, table, &mhop, &index);
525 out:
526 	mutex_unlock(&table->mutex);
527 	return ret;
528 }
529 
530 int hns_roce_table_get(struct hns_roce_dev *hr_dev,
531 		       struct hns_roce_hem_table *table, unsigned long obj)
532 {
533 	struct device *dev = hr_dev->dev;
534 	unsigned long i;
535 	int ret = 0;
536 
537 	if (hns_roce_check_whether_mhop(hr_dev, table->type))
538 		return hns_roce_table_mhop_get(hr_dev, table, obj);
539 
540 	i = obj / (table->table_chunk_size / table->obj_size);
541 
542 	mutex_lock(&table->mutex);
543 
544 	if (table->hem[i]) {
545 		refcount_inc(&table->hem[i]->refcount);
546 		goto out;
547 	}
548 
549 	table->hem[i] = hns_roce_alloc_hem(hr_dev,
550 				       table->table_chunk_size,
551 				       GFP_KERNEL | __GFP_NOWARN);
552 	if (!table->hem[i]) {
553 		ret = -ENOMEM;
554 		goto out;
555 	}
556 
557 	/* Set HEM base address(128K/page, pa) to Hardware */
558 	ret = hr_dev->hw->set_hem(hr_dev, table, obj, HEM_HOP_STEP_DIRECT);
559 	if (ret) {
560 		hns_roce_free_hem(hr_dev, table->hem[i]);
561 		table->hem[i] = NULL;
562 		dev_err(dev, "set HEM base address to HW failed, ret = %d.\n",
563 			ret);
564 		goto out;
565 	}
566 
567 	refcount_set(&table->hem[i]->refcount, 1);
568 out:
569 	mutex_unlock(&table->mutex);
570 	return ret;
571 }
572 
573 static void clear_mhop_hem(struct hns_roce_dev *hr_dev,
574 			   struct hns_roce_hem_table *table, unsigned long obj,
575 			   struct hns_roce_hem_mhop *mhop,
576 			   struct hns_roce_hem_index *index)
577 {
578 	struct ib_device *ibdev = &hr_dev->ib_dev;
579 	u32 hop_num = mhop->hop_num;
580 	u32 chunk_ba_num;
581 	u32 step_idx;
582 	int ret;
583 
584 	index->inited = HEM_INDEX_BUF;
585 	chunk_ba_num = mhop->bt_chunk_size / BA_BYTE_LEN;
586 	if (check_whether_bt_num_2(table->type, hop_num)) {
587 		if (hns_roce_check_hem_null(table->hem, index->buf,
588 					    chunk_ba_num, table->num_hem))
589 			index->inited |= HEM_INDEX_L0;
590 	} else if (check_whether_bt_num_3(table->type, hop_num)) {
591 		if (hns_roce_check_hem_null(table->hem, index->buf,
592 					    chunk_ba_num, table->num_hem)) {
593 			index->inited |= HEM_INDEX_L1;
594 			if (hns_roce_check_bt_null(table->bt_l1, index->l1,
595 						   chunk_ba_num))
596 				index->inited |= HEM_INDEX_L0;
597 		}
598 	}
599 
600 	if (table->type < HEM_TYPE_MTT) {
601 		if (hop_num == HNS_ROCE_HOP_NUM_0)
602 			step_idx = 0;
603 		else
604 			step_idx = hop_num;
605 
606 		ret = hr_dev->hw->clear_hem(hr_dev, table, obj, step_idx);
607 		if (ret)
608 			ibdev_warn(ibdev, "failed to clear hop%u HEM, ret = %d.\n",
609 				   hop_num, ret);
610 
611 		if (index->inited & HEM_INDEX_L1) {
612 			ret = hr_dev->hw->clear_hem(hr_dev, table, obj, 1);
613 			if (ret)
614 				ibdev_warn(ibdev, "failed to clear HEM step 1, ret = %d.\n",
615 					   ret);
616 		}
617 
618 		if (index->inited & HEM_INDEX_L0) {
619 			ret = hr_dev->hw->clear_hem(hr_dev, table, obj, 0);
620 			if (ret)
621 				ibdev_warn(ibdev, "failed to clear HEM step 0, ret = %d.\n",
622 					   ret);
623 		}
624 	}
625 }
626 
627 static void hns_roce_table_mhop_put(struct hns_roce_dev *hr_dev,
628 				    struct hns_roce_hem_table *table,
629 				    unsigned long obj,
630 				    int check_refcount)
631 {
632 	struct ib_device *ibdev = &hr_dev->ib_dev;
633 	struct hns_roce_hem_index index = {};
634 	struct hns_roce_hem_mhop mhop = {};
635 	int ret;
636 
637 	ret = calc_hem_config(hr_dev, table, obj, &mhop, &index);
638 	if (ret) {
639 		ibdev_err(ibdev, "calc hem config failed!\n");
640 		return;
641 	}
642 
643 	if (!check_refcount)
644 		mutex_lock(&table->mutex);
645 	else if (!refcount_dec_and_mutex_lock(&table->hem[index.buf]->refcount,
646 					      &table->mutex))
647 		return;
648 
649 	clear_mhop_hem(hr_dev, table, obj, &mhop, &index);
650 	free_mhop_hem(hr_dev, table, &mhop, &index);
651 
652 	mutex_unlock(&table->mutex);
653 }
654 
655 void hns_roce_table_put(struct hns_roce_dev *hr_dev,
656 			struct hns_roce_hem_table *table, unsigned long obj)
657 {
658 	struct device *dev = hr_dev->dev;
659 	unsigned long i;
660 	int ret;
661 
662 	if (hns_roce_check_whether_mhop(hr_dev, table->type)) {
663 		hns_roce_table_mhop_put(hr_dev, table, obj, 1);
664 		return;
665 	}
666 
667 	i = obj / (table->table_chunk_size / table->obj_size);
668 
669 	if (!refcount_dec_and_mutex_lock(&table->hem[i]->refcount,
670 					 &table->mutex))
671 		return;
672 
673 	ret = hr_dev->hw->clear_hem(hr_dev, table, obj, HEM_HOP_STEP_DIRECT);
674 	if (ret)
675 		dev_warn(dev, "failed to clear HEM base address, ret = %d.\n",
676 			 ret);
677 
678 	hns_roce_free_hem(hr_dev, table->hem[i]);
679 	table->hem[i] = NULL;
680 
681 	mutex_unlock(&table->mutex);
682 }
683 
684 void *hns_roce_table_find(struct hns_roce_dev *hr_dev,
685 			  struct hns_roce_hem_table *table,
686 			  unsigned long obj, dma_addr_t *dma_handle)
687 {
688 	struct hns_roce_hem_mhop mhop;
689 	struct hns_roce_hem *hem;
690 	unsigned long mhop_obj = obj;
691 	unsigned long obj_per_chunk;
692 	unsigned long idx_offset;
693 	int offset, dma_offset;
694 	void *addr = NULL;
695 	u32 hem_idx = 0;
696 	int i, j;
697 
698 	mutex_lock(&table->mutex);
699 
700 	if (!hns_roce_check_whether_mhop(hr_dev, table->type)) {
701 		obj_per_chunk = table->table_chunk_size / table->obj_size;
702 		hem = table->hem[obj / obj_per_chunk];
703 		idx_offset = obj % obj_per_chunk;
704 		dma_offset = offset = idx_offset * table->obj_size;
705 	} else {
706 		u32 seg_size = 64; /* 8 bytes per BA and 8 BA per segment */
707 
708 		if (hns_roce_calc_hem_mhop(hr_dev, table, &mhop_obj, &mhop))
709 			goto out;
710 		/* mtt mhop */
711 		i = mhop.l0_idx;
712 		j = mhop.l1_idx;
713 		if (mhop.hop_num == 2)
714 			hem_idx = i * (mhop.bt_chunk_size / BA_BYTE_LEN) + j;
715 		else if (mhop.hop_num == 1 ||
716 			 mhop.hop_num == HNS_ROCE_HOP_NUM_0)
717 			hem_idx = i;
718 
719 		hem = table->hem[hem_idx];
720 		dma_offset = offset = obj * seg_size % mhop.bt_chunk_size;
721 		if (mhop.hop_num == 2)
722 			dma_offset = offset = 0;
723 	}
724 
725 	if (!hem)
726 		goto out;
727 
728 	*dma_handle = hem->dma + dma_offset;
729 	addr = hem->buf + offset;
730 
731 out:
732 	mutex_unlock(&table->mutex);
733 	return addr;
734 }
735 
736 int hns_roce_init_hem_table(struct hns_roce_dev *hr_dev,
737 			    struct hns_roce_hem_table *table, u32 type,
738 			    unsigned long obj_size, unsigned long nobj)
739 {
740 	unsigned long obj_per_chunk;
741 	unsigned long num_hem;
742 
743 	if (!hns_roce_check_whether_mhop(hr_dev, type)) {
744 		table->table_chunk_size = hr_dev->caps.chunk_sz;
745 		obj_per_chunk = table->table_chunk_size / obj_size;
746 		num_hem = DIV_ROUND_UP(nobj, obj_per_chunk);
747 
748 		table->hem = kcalloc(num_hem, sizeof(*table->hem), GFP_KERNEL);
749 		if (!table->hem)
750 			return -ENOMEM;
751 	} else {
752 		struct hns_roce_hem_mhop mhop = {};
753 		unsigned long buf_chunk_size;
754 		unsigned long bt_chunk_size;
755 		unsigned long bt_chunk_num;
756 		unsigned long num_bt_l0;
757 		u32 hop_num;
758 
759 		if (get_hem_table_config(hr_dev, &mhop, type))
760 			return -EINVAL;
761 
762 		buf_chunk_size = mhop.buf_chunk_size;
763 		bt_chunk_size = mhop.bt_chunk_size;
764 		num_bt_l0 = mhop.ba_l0_num;
765 		hop_num = mhop.hop_num;
766 
767 		obj_per_chunk = buf_chunk_size / obj_size;
768 		num_hem = DIV_ROUND_UP(nobj, obj_per_chunk);
769 		bt_chunk_num = bt_chunk_size / BA_BYTE_LEN;
770 
771 		if (type >= HEM_TYPE_MTT)
772 			num_bt_l0 = bt_chunk_num;
773 
774 		table->hem = kcalloc(num_hem, sizeof(*table->hem),
775 					 GFP_KERNEL);
776 		if (!table->hem)
777 			goto err_kcalloc_hem_buf;
778 
779 		if (check_whether_bt_num_3(type, hop_num)) {
780 			unsigned long num_bt_l1;
781 
782 			num_bt_l1 = DIV_ROUND_UP(num_hem, bt_chunk_num);
783 			table->bt_l1 = kcalloc(num_bt_l1,
784 					       sizeof(*table->bt_l1),
785 					       GFP_KERNEL);
786 			if (!table->bt_l1)
787 				goto err_kcalloc_bt_l1;
788 
789 			table->bt_l1_dma_addr = kcalloc(num_bt_l1,
790 						 sizeof(*table->bt_l1_dma_addr),
791 						 GFP_KERNEL);
792 
793 			if (!table->bt_l1_dma_addr)
794 				goto err_kcalloc_l1_dma;
795 		}
796 
797 		if (check_whether_bt_num_2(type, hop_num) ||
798 			check_whether_bt_num_3(type, hop_num)) {
799 			table->bt_l0 = kcalloc(num_bt_l0, sizeof(*table->bt_l0),
800 					       GFP_KERNEL);
801 			if (!table->bt_l0)
802 				goto err_kcalloc_bt_l0;
803 
804 			table->bt_l0_dma_addr = kcalloc(num_bt_l0,
805 						 sizeof(*table->bt_l0_dma_addr),
806 						 GFP_KERNEL);
807 			if (!table->bt_l0_dma_addr)
808 				goto err_kcalloc_l0_dma;
809 		}
810 	}
811 
812 	table->type = type;
813 	table->num_hem = num_hem;
814 	table->obj_size = obj_size;
815 	mutex_init(&table->mutex);
816 
817 	return 0;
818 
819 err_kcalloc_l0_dma:
820 	kfree(table->bt_l0);
821 	table->bt_l0 = NULL;
822 
823 err_kcalloc_bt_l0:
824 	kfree(table->bt_l1_dma_addr);
825 	table->bt_l1_dma_addr = NULL;
826 
827 err_kcalloc_l1_dma:
828 	kfree(table->bt_l1);
829 	table->bt_l1 = NULL;
830 
831 err_kcalloc_bt_l1:
832 	kfree(table->hem);
833 	table->hem = NULL;
834 
835 err_kcalloc_hem_buf:
836 	return -ENOMEM;
837 }
838 
839 static void hns_roce_cleanup_mhop_hem_table(struct hns_roce_dev *hr_dev,
840 					    struct hns_roce_hem_table *table)
841 {
842 	struct hns_roce_hem_mhop mhop;
843 	u32 buf_chunk_size;
844 	u64 obj;
845 	int i;
846 
847 	if (hns_roce_calc_hem_mhop(hr_dev, table, NULL, &mhop))
848 		return;
849 	buf_chunk_size = table->type < HEM_TYPE_MTT ? mhop.buf_chunk_size :
850 					mhop.bt_chunk_size;
851 
852 	for (i = 0; i < table->num_hem; ++i) {
853 		obj = i * buf_chunk_size / table->obj_size;
854 		if (table->hem[i])
855 			hns_roce_table_mhop_put(hr_dev, table, obj, 0);
856 	}
857 
858 	kfree(table->hem);
859 	table->hem = NULL;
860 	kfree(table->bt_l1);
861 	table->bt_l1 = NULL;
862 	kfree(table->bt_l1_dma_addr);
863 	table->bt_l1_dma_addr = NULL;
864 	kfree(table->bt_l0);
865 	table->bt_l0 = NULL;
866 	kfree(table->bt_l0_dma_addr);
867 	table->bt_l0_dma_addr = NULL;
868 }
869 
870 void hns_roce_cleanup_hem_table(struct hns_roce_dev *hr_dev,
871 				struct hns_roce_hem_table *table)
872 {
873 	struct device *dev = hr_dev->dev;
874 	unsigned long i;
875 	int obj;
876 	int ret;
877 
878 	if (hns_roce_check_whether_mhop(hr_dev, table->type)) {
879 		hns_roce_cleanup_mhop_hem_table(hr_dev, table);
880 		mutex_destroy(&table->mutex);
881 		return;
882 	}
883 
884 	for (i = 0; i < table->num_hem; ++i)
885 		if (table->hem[i]) {
886 			obj = i * table->table_chunk_size / table->obj_size;
887 			ret = hr_dev->hw->clear_hem(hr_dev, table, obj, 0);
888 			if (ret)
889 				dev_err(dev, "clear HEM base address failed, ret = %d.\n",
890 					ret);
891 
892 			hns_roce_free_hem(hr_dev, table->hem[i]);
893 		}
894 
895 	mutex_destroy(&table->mutex);
896 	kfree(table->hem);
897 }
898 
899 void hns_roce_cleanup_hem(struct hns_roce_dev *hr_dev)
900 {
901 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_SRQ)
902 		hns_roce_cleanup_hem_table(hr_dev,
903 					   &hr_dev->srq_table.table);
904 	hns_roce_cleanup_hem_table(hr_dev, &hr_dev->cq_table.table);
905 	if (hr_dev->caps.qpc_timer_entry_sz)
906 		hns_roce_cleanup_hem_table(hr_dev,
907 					   &hr_dev->qpc_timer_table);
908 	if (hr_dev->caps.cqc_timer_entry_sz)
909 		hns_roce_cleanup_hem_table(hr_dev,
910 					   &hr_dev->cqc_timer_table);
911 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_QP_FLOW_CTRL)
912 		hns_roce_cleanup_hem_table(hr_dev,
913 					   &hr_dev->qp_table.sccc_table);
914 	if (hr_dev->caps.trrl_entry_sz)
915 		hns_roce_cleanup_hem_table(hr_dev,
916 					   &hr_dev->qp_table.trrl_table);
917 
918 	if (hr_dev->caps.gmv_entry_sz)
919 		hns_roce_cleanup_hem_table(hr_dev, &hr_dev->gmv_table);
920 
921 	hns_roce_cleanup_hem_table(hr_dev, &hr_dev->qp_table.irrl_table);
922 	hns_roce_cleanup_hem_table(hr_dev, &hr_dev->qp_table.qp_table);
923 	hns_roce_cleanup_hem_table(hr_dev, &hr_dev->mr_table.mtpt_table);
924 }
925 
926 struct hns_roce_hem_item {
927 	struct list_head list; /* link all hems in the same bt level */
928 	struct list_head sibling; /* link all hems in last hop for mtt */
929 	void *addr;
930 	dma_addr_t dma_addr;
931 	size_t count; /* max ba numbers */
932 	int start; /* start buf offset in this hem */
933 	int end; /* end buf offset in this hem */
934 };
935 
936 /* All HEM items are linked in a tree structure */
937 struct hns_roce_hem_head {
938 	struct list_head branch[HNS_ROCE_MAX_BT_REGION];
939 	struct list_head root;
940 	struct list_head leaf;
941 };
942 
943 static struct hns_roce_hem_item *
944 hem_list_alloc_item(struct hns_roce_dev *hr_dev, int start, int end, int count,
945 		    bool exist_bt)
946 {
947 	struct hns_roce_hem_item *hem;
948 
949 	hem = kzalloc(sizeof(*hem), GFP_KERNEL);
950 	if (!hem)
951 		return NULL;
952 
953 	if (exist_bt) {
954 		hem->addr = dma_alloc_coherent(hr_dev->dev, count * BA_BYTE_LEN,
955 					       &hem->dma_addr, GFP_KERNEL);
956 		if (!hem->addr) {
957 			kfree(hem);
958 			return NULL;
959 		}
960 	}
961 
962 	hem->count = count;
963 	hem->start = start;
964 	hem->end = end;
965 	INIT_LIST_HEAD(&hem->list);
966 	INIT_LIST_HEAD(&hem->sibling);
967 
968 	return hem;
969 }
970 
971 static void hem_list_free_item(struct hns_roce_dev *hr_dev,
972 			       struct hns_roce_hem_item *hem, bool exist_bt)
973 {
974 	if (exist_bt)
975 		dma_free_coherent(hr_dev->dev, hem->count * BA_BYTE_LEN,
976 				  hem->addr, hem->dma_addr);
977 	kfree(hem);
978 }
979 
980 static void hem_list_free_all(struct hns_roce_dev *hr_dev,
981 			      struct list_head *head, bool exist_bt)
982 {
983 	struct hns_roce_hem_item *hem, *temp_hem;
984 
985 	list_for_each_entry_safe(hem, temp_hem, head, list) {
986 		list_del(&hem->list);
987 		hem_list_free_item(hr_dev, hem, exist_bt);
988 	}
989 }
990 
991 static void hem_list_link_bt(void *base_addr, u64 table_addr)
992 {
993 	*(u64 *)(base_addr) = table_addr;
994 }
995 
996 /* assign L0 table address to hem from root bt */
997 static void hem_list_assign_bt(struct hns_roce_hem_item *hem, void *cpu_addr,
998 			       u64 phy_addr)
999 {
1000 	hem->addr = cpu_addr;
1001 	hem->dma_addr = (dma_addr_t)phy_addr;
1002 }
1003 
1004 static inline bool hem_list_page_is_in_range(struct hns_roce_hem_item *hem,
1005 					     int offset)
1006 {
1007 	return (hem->start <= offset && offset <= hem->end);
1008 }
1009 
1010 static struct hns_roce_hem_item *hem_list_search_item(struct list_head *ba_list,
1011 						      int page_offset)
1012 {
1013 	struct hns_roce_hem_item *hem, *temp_hem;
1014 	struct hns_roce_hem_item *found = NULL;
1015 
1016 	list_for_each_entry_safe(hem, temp_hem, ba_list, list) {
1017 		if (hem_list_page_is_in_range(hem, page_offset)) {
1018 			found = hem;
1019 			break;
1020 		}
1021 	}
1022 
1023 	return found;
1024 }
1025 
1026 static bool hem_list_is_bottom_bt(int hopnum, int bt_level)
1027 {
1028 	/*
1029 	 * hopnum    base address table levels
1030 	 * 0		L0(buf)
1031 	 * 1		L0 -> buf
1032 	 * 2		L0 -> L1 -> buf
1033 	 * 3		L0 -> L1 -> L2 -> buf
1034 	 */
1035 	return bt_level >= (hopnum ? hopnum - 1 : hopnum);
1036 }
1037 
1038 /*
1039  * calc base address entries num
1040  * @hopnum: num of mutihop addressing
1041  * @bt_level: base address table level
1042  * @unit: ba entries per bt page
1043  */
1044 static u64 hem_list_calc_ba_range(int hopnum, int bt_level, int unit)
1045 {
1046 	u64 step;
1047 	int max;
1048 	int i;
1049 
1050 	if (hopnum <= bt_level)
1051 		return 0;
1052 	/*
1053 	 * hopnum  bt_level   range
1054 	 * 1	      0       unit
1055 	 * ------------
1056 	 * 2	      0       unit * unit
1057 	 * 2	      1       unit
1058 	 * ------------
1059 	 * 3	      0       unit * unit * unit
1060 	 * 3	      1       unit * unit
1061 	 * 3	      2       unit
1062 	 */
1063 	step = 1;
1064 	max = hopnum - bt_level;
1065 	for (i = 0; i < max; i++)
1066 		step = step * unit;
1067 
1068 	return step;
1069 }
1070 
1071 /*
1072  * calc the root ba entries which could cover all regions
1073  * @regions: buf region array
1074  * @region_cnt: array size of @regions
1075  * @unit: ba entries per bt page
1076  */
1077 int hns_roce_hem_list_calc_root_ba(const struct hns_roce_buf_region *regions,
1078 				   int region_cnt, int unit)
1079 {
1080 	struct hns_roce_buf_region *r;
1081 	int total = 0;
1082 	u64 step;
1083 	int i;
1084 
1085 	for (i = 0; i < region_cnt; i++) {
1086 		r = (struct hns_roce_buf_region *)&regions[i];
1087 		if (r->hopnum > 1) {
1088 			step = hem_list_calc_ba_range(r->hopnum, 1, unit);
1089 			if (step > 0)
1090 				total += (r->count + step - 1) / step;
1091 		} else {
1092 			total += r->count;
1093 		}
1094 	}
1095 
1096 	return total;
1097 }
1098 
1099 static int hem_list_alloc_mid_bt(struct hns_roce_dev *hr_dev,
1100 				 const struct hns_roce_buf_region *r, int unit,
1101 				 int offset, struct list_head *mid_bt,
1102 				 struct list_head *btm_bt)
1103 {
1104 	struct hns_roce_hem_item *hem_ptrs[HNS_ROCE_MAX_BT_LEVEL] = { NULL };
1105 	struct list_head temp_list[HNS_ROCE_MAX_BT_LEVEL];
1106 	struct hns_roce_hem_item *cur, *pre;
1107 	const int hopnum = r->hopnum;
1108 	int start_aligned;
1109 	int distance;
1110 	int ret = 0;
1111 	int max_ofs;
1112 	int level;
1113 	u64 step;
1114 	int end;
1115 
1116 	if (hopnum <= 1)
1117 		return 0;
1118 
1119 	if (hopnum > HNS_ROCE_MAX_BT_LEVEL) {
1120 		dev_err(hr_dev->dev, "invalid hopnum %d!\n", hopnum);
1121 		return -EINVAL;
1122 	}
1123 
1124 	if (offset < r->offset) {
1125 		dev_err(hr_dev->dev, "invalid offset %d, min %u!\n",
1126 			offset, r->offset);
1127 		return -EINVAL;
1128 	}
1129 
1130 	distance = offset - r->offset;
1131 	max_ofs = r->offset + r->count - 1;
1132 	for (level = 0; level < hopnum; level++)
1133 		INIT_LIST_HEAD(&temp_list[level]);
1134 
1135 	/* config L1 bt to last bt and link them to corresponding parent */
1136 	for (level = 1; level < hopnum; level++) {
1137 		if (!hem_list_is_bottom_bt(hopnum, level)) {
1138 			cur = hem_list_search_item(&mid_bt[level], offset);
1139 			if (cur) {
1140 				hem_ptrs[level] = cur;
1141 				continue;
1142 			}
1143 		}
1144 
1145 		step = hem_list_calc_ba_range(hopnum, level, unit);
1146 		if (step < 1) {
1147 			ret = -EINVAL;
1148 			goto err_exit;
1149 		}
1150 
1151 		start_aligned = (distance / step) * step + r->offset;
1152 		end = min_t(u64, start_aligned + step - 1, max_ofs);
1153 		cur = hem_list_alloc_item(hr_dev, start_aligned, end, unit,
1154 					  true);
1155 		if (!cur) {
1156 			ret = -ENOMEM;
1157 			goto err_exit;
1158 		}
1159 		hem_ptrs[level] = cur;
1160 		list_add(&cur->list, &temp_list[level]);
1161 		if (hem_list_is_bottom_bt(hopnum, level))
1162 			list_add(&cur->sibling, &temp_list[0]);
1163 
1164 		/* link bt to parent bt */
1165 		if (level > 1) {
1166 			pre = hem_ptrs[level - 1];
1167 			step = (cur->start - pre->start) / step * BA_BYTE_LEN;
1168 			hem_list_link_bt(pre->addr + step, cur->dma_addr);
1169 		}
1170 	}
1171 
1172 	list_splice(&temp_list[0], btm_bt);
1173 	for (level = 1; level < hopnum; level++)
1174 		list_splice(&temp_list[level], &mid_bt[level]);
1175 
1176 	return 0;
1177 
1178 err_exit:
1179 	for (level = 1; level < hopnum; level++)
1180 		hem_list_free_all(hr_dev, &temp_list[level], true);
1181 
1182 	return ret;
1183 }
1184 
1185 static struct hns_roce_hem_item *
1186 alloc_root_hem(struct hns_roce_dev *hr_dev, int unit, int *max_ba_num,
1187 	       const struct hns_roce_buf_region *regions, int region_cnt)
1188 {
1189 	const struct hns_roce_buf_region *r;
1190 	struct hns_roce_hem_item *hem;
1191 	int ba_num;
1192 	int offset;
1193 
1194 	ba_num = hns_roce_hem_list_calc_root_ba(regions, region_cnt, unit);
1195 	if (ba_num < 1)
1196 		return ERR_PTR(-ENOMEM);
1197 
1198 	if (ba_num > unit)
1199 		return ERR_PTR(-ENOBUFS);
1200 
1201 	offset = regions[0].offset;
1202 	/* indicate to last region */
1203 	r = &regions[region_cnt - 1];
1204 	hem = hem_list_alloc_item(hr_dev, offset, r->offset + r->count - 1,
1205 				  ba_num, true);
1206 	if (!hem)
1207 		return ERR_PTR(-ENOMEM);
1208 
1209 	*max_ba_num = ba_num;
1210 
1211 	return hem;
1212 }
1213 
1214 static int alloc_fake_root_bt(struct hns_roce_dev *hr_dev, void *cpu_base,
1215 			      u64 phy_base, const struct hns_roce_buf_region *r,
1216 			      struct list_head *branch_head,
1217 			      struct list_head *leaf_head)
1218 {
1219 	struct hns_roce_hem_item *hem;
1220 
1221 	hem = hem_list_alloc_item(hr_dev, r->offset, r->offset + r->count - 1,
1222 				  r->count, false);
1223 	if (!hem)
1224 		return -ENOMEM;
1225 
1226 	hem_list_assign_bt(hem, cpu_base, phy_base);
1227 	list_add(&hem->list, branch_head);
1228 	list_add(&hem->sibling, leaf_head);
1229 
1230 	return r->count;
1231 }
1232 
1233 static int setup_middle_bt(struct hns_roce_dev *hr_dev, void *cpu_base,
1234 			   int unit, const struct hns_roce_buf_region *r,
1235 			   const struct list_head *branch_head)
1236 {
1237 	struct hns_roce_hem_item *hem, *temp_hem;
1238 	int total = 0;
1239 	int offset;
1240 	u64 step;
1241 
1242 	step = hem_list_calc_ba_range(r->hopnum, 1, unit);
1243 	if (step < 1)
1244 		return -EINVAL;
1245 
1246 	/* if exist mid bt, link L1 to L0 */
1247 	list_for_each_entry_safe(hem, temp_hem, branch_head, list) {
1248 		offset = (hem->start - r->offset) / step * BA_BYTE_LEN;
1249 		hem_list_link_bt(cpu_base + offset, hem->dma_addr);
1250 		total++;
1251 	}
1252 
1253 	return total;
1254 }
1255 
1256 static int
1257 setup_root_hem(struct hns_roce_dev *hr_dev, struct hns_roce_hem_list *hem_list,
1258 	       int unit, int max_ba_num, struct hns_roce_hem_head *head,
1259 	       const struct hns_roce_buf_region *regions, int region_cnt)
1260 {
1261 	const struct hns_roce_buf_region *r;
1262 	struct hns_roce_hem_item *root_hem;
1263 	void *cpu_base;
1264 	u64 phy_base;
1265 	int i, total;
1266 	int ret;
1267 
1268 	root_hem = list_first_entry(&head->root,
1269 				    struct hns_roce_hem_item, list);
1270 	if (!root_hem)
1271 		return -ENOMEM;
1272 
1273 	total = 0;
1274 	for (i = 0; i < region_cnt && total < max_ba_num; i++) {
1275 		r = &regions[i];
1276 		if (!r->count)
1277 			continue;
1278 
1279 		/* all regions's mid[x][0] shared the root_bt's trunk */
1280 		cpu_base = root_hem->addr + total * BA_BYTE_LEN;
1281 		phy_base = root_hem->dma_addr + total * BA_BYTE_LEN;
1282 
1283 		/* if hopnum is 0 or 1, cut a new fake hem from the root bt
1284 		 * which's address share to all regions.
1285 		 */
1286 		if (hem_list_is_bottom_bt(r->hopnum, 0))
1287 			ret = alloc_fake_root_bt(hr_dev, cpu_base, phy_base, r,
1288 						 &head->branch[i], &head->leaf);
1289 		else
1290 			ret = setup_middle_bt(hr_dev, cpu_base, unit, r,
1291 					      &hem_list->mid_bt[i][1]);
1292 
1293 		if (ret < 0)
1294 			return ret;
1295 
1296 		total += ret;
1297 	}
1298 
1299 	list_splice(&head->leaf, &hem_list->btm_bt);
1300 	list_splice(&head->root, &hem_list->root_bt);
1301 	for (i = 0; i < region_cnt; i++)
1302 		list_splice(&head->branch[i], &hem_list->mid_bt[i][0]);
1303 
1304 	return 0;
1305 }
1306 
1307 static int hem_list_alloc_root_bt(struct hns_roce_dev *hr_dev,
1308 				  struct hns_roce_hem_list *hem_list, int unit,
1309 				  const struct hns_roce_buf_region *regions,
1310 				  int region_cnt)
1311 {
1312 	struct hns_roce_hem_item *root_hem;
1313 	struct hns_roce_hem_head head;
1314 	int max_ba_num;
1315 	int ret;
1316 	int i;
1317 
1318 	root_hem = hem_list_search_item(&hem_list->root_bt, regions[0].offset);
1319 	if (root_hem)
1320 		return 0;
1321 
1322 	max_ba_num = 0;
1323 	root_hem = alloc_root_hem(hr_dev, unit, &max_ba_num, regions,
1324 				  region_cnt);
1325 	if (IS_ERR(root_hem))
1326 		return PTR_ERR(root_hem);
1327 
1328 	/* List head for storing all allocated HEM items */
1329 	INIT_LIST_HEAD(&head.root);
1330 	INIT_LIST_HEAD(&head.leaf);
1331 	for (i = 0; i < region_cnt; i++)
1332 		INIT_LIST_HEAD(&head.branch[i]);
1333 
1334 	hem_list->root_ba = root_hem->dma_addr;
1335 	list_add(&root_hem->list, &head.root);
1336 	ret = setup_root_hem(hr_dev, hem_list, unit, max_ba_num, &head, regions,
1337 			     region_cnt);
1338 	if (ret) {
1339 		for (i = 0; i < region_cnt; i++)
1340 			hem_list_free_all(hr_dev, &head.branch[i], false);
1341 
1342 		hem_list_free_all(hr_dev, &head.root, true);
1343 	}
1344 
1345 	return ret;
1346 }
1347 
1348 /* construct the base address table and link them by address hop config */
1349 int hns_roce_hem_list_request(struct hns_roce_dev *hr_dev,
1350 			      struct hns_roce_hem_list *hem_list,
1351 			      const struct hns_roce_buf_region *regions,
1352 			      int region_cnt, unsigned int bt_pg_shift)
1353 {
1354 	const struct hns_roce_buf_region *r;
1355 	int ofs, end;
1356 	int unit;
1357 	int ret;
1358 	int i;
1359 
1360 	if (region_cnt > HNS_ROCE_MAX_BT_REGION) {
1361 		dev_err(hr_dev->dev, "invalid region region_cnt %d!\n",
1362 			region_cnt);
1363 		return -EINVAL;
1364 	}
1365 
1366 	unit = (1 << bt_pg_shift) / BA_BYTE_LEN;
1367 	for (i = 0; i < region_cnt; i++) {
1368 		r = &regions[i];
1369 		if (!r->count)
1370 			continue;
1371 
1372 		end = r->offset + r->count;
1373 		for (ofs = r->offset; ofs < end; ofs += unit) {
1374 			ret = hem_list_alloc_mid_bt(hr_dev, r, unit, ofs,
1375 						    hem_list->mid_bt[i],
1376 						    &hem_list->btm_bt);
1377 			if (ret) {
1378 				dev_err(hr_dev->dev,
1379 					"alloc hem trunk fail ret = %d!\n", ret);
1380 				goto err_alloc;
1381 			}
1382 		}
1383 	}
1384 
1385 	ret = hem_list_alloc_root_bt(hr_dev, hem_list, unit, regions,
1386 				     region_cnt);
1387 	if (ret)
1388 		dev_err(hr_dev->dev, "alloc hem root fail ret = %d!\n", ret);
1389 	else
1390 		return 0;
1391 
1392 err_alloc:
1393 	hns_roce_hem_list_release(hr_dev, hem_list);
1394 
1395 	return ret;
1396 }
1397 
1398 void hns_roce_hem_list_release(struct hns_roce_dev *hr_dev,
1399 			       struct hns_roce_hem_list *hem_list)
1400 {
1401 	int i, j;
1402 
1403 	for (i = 0; i < HNS_ROCE_MAX_BT_REGION; i++)
1404 		for (j = 0; j < HNS_ROCE_MAX_BT_LEVEL; j++)
1405 			hem_list_free_all(hr_dev, &hem_list->mid_bt[i][j],
1406 					  j != 0);
1407 
1408 	hem_list_free_all(hr_dev, &hem_list->root_bt, true);
1409 	INIT_LIST_HEAD(&hem_list->btm_bt);
1410 	hem_list->root_ba = 0;
1411 }
1412 
1413 void hns_roce_hem_list_init(struct hns_roce_hem_list *hem_list)
1414 {
1415 	int i, j;
1416 
1417 	INIT_LIST_HEAD(&hem_list->root_bt);
1418 	INIT_LIST_HEAD(&hem_list->btm_bt);
1419 	for (i = 0; i < HNS_ROCE_MAX_BT_REGION; i++)
1420 		for (j = 0; j < HNS_ROCE_MAX_BT_LEVEL; j++)
1421 			INIT_LIST_HEAD(&hem_list->mid_bt[i][j]);
1422 }
1423 
1424 void *hns_roce_hem_list_find_mtt(struct hns_roce_dev *hr_dev,
1425 				 struct hns_roce_hem_list *hem_list,
1426 				 int offset, int *mtt_cnt)
1427 {
1428 	struct list_head *head = &hem_list->btm_bt;
1429 	struct hns_roce_hem_item *hem, *temp_hem;
1430 	void *cpu_base = NULL;
1431 	int nr = 0;
1432 
1433 	list_for_each_entry_safe(hem, temp_hem, head, sibling) {
1434 		if (hem_list_page_is_in_range(hem, offset)) {
1435 			nr = offset - hem->start;
1436 			cpu_base = hem->addr + nr * BA_BYTE_LEN;
1437 			nr = hem->end + 1 - offset;
1438 			break;
1439 		}
1440 	}
1441 
1442 	if (mtt_cnt)
1443 		*mtt_cnt = nr;
1444 
1445 	return cpu_base;
1446 }
1447