xref: /linux/drivers/infiniband/hw/hns/hns_roce_hem.c (revision bdd1a21b52557ea8f61d0a5dc2f77151b576eb70)
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 <linux/platform_device.h>
35 #include "hns_roce_device.h"
36 #include "hns_roce_hem.h"
37 #include "hns_roce_common.h"
38 
39 #define HEM_INDEX_BUF			BIT(0)
40 #define HEM_INDEX_L0			BIT(1)
41 #define HEM_INDEX_L1			BIT(2)
42 struct hns_roce_hem_index {
43 	u64 buf;
44 	u64 l0;
45 	u64 l1;
46 	u32 inited; /* indicate which index is available */
47 };
48 
49 bool hns_roce_check_whether_mhop(struct hns_roce_dev *hr_dev, u32 type)
50 {
51 	int hop_num = 0;
52 
53 	switch (type) {
54 	case HEM_TYPE_QPC:
55 		hop_num = hr_dev->caps.qpc_hop_num;
56 		break;
57 	case HEM_TYPE_MTPT:
58 		hop_num = hr_dev->caps.mpt_hop_num;
59 		break;
60 	case HEM_TYPE_CQC:
61 		hop_num = hr_dev->caps.cqc_hop_num;
62 		break;
63 	case HEM_TYPE_SRQC:
64 		hop_num = hr_dev->caps.srqc_hop_num;
65 		break;
66 	case HEM_TYPE_SCCC:
67 		hop_num = hr_dev->caps.sccc_hop_num;
68 		break;
69 	case HEM_TYPE_QPC_TIMER:
70 		hop_num = hr_dev->caps.qpc_timer_hop_num;
71 		break;
72 	case HEM_TYPE_CQC_TIMER:
73 		hop_num = hr_dev->caps.cqc_timer_hop_num;
74 		break;
75 	case HEM_TYPE_GMV:
76 		hop_num = hr_dev->caps.gmv_hop_num;
77 		break;
78 	default:
79 		return false;
80 	}
81 
82 	return hop_num ? true : false;
83 }
84 
85 static bool hns_roce_check_hem_null(struct hns_roce_hem **hem, u64 hem_idx,
86 				    u32 bt_chunk_num, u64 hem_max_num)
87 {
88 	u64 start_idx = round_down(hem_idx, bt_chunk_num);
89 	u64 check_max_num = start_idx + bt_chunk_num;
90 	u64 i;
91 
92 	for (i = start_idx; (i < check_max_num) && (i < hem_max_num); i++)
93 		if (i != hem_idx && hem[i])
94 			return false;
95 
96 	return true;
97 }
98 
99 static bool hns_roce_check_bt_null(u64 **bt, u64 ba_idx, u32 bt_chunk_num)
100 {
101 	u64 start_idx = round_down(ba_idx, bt_chunk_num);
102 	int i;
103 
104 	for (i = 0; i < bt_chunk_num; i++)
105 		if (i != ba_idx && bt[start_idx + i])
106 			return false;
107 
108 	return true;
109 }
110 
111 static int hns_roce_get_bt_num(u32 table_type, u32 hop_num)
112 {
113 	if (check_whether_bt_num_3(table_type, hop_num))
114 		return 3;
115 	else if (check_whether_bt_num_2(table_type, hop_num))
116 		return 2;
117 	else if (check_whether_bt_num_1(table_type, hop_num))
118 		return 1;
119 	else
120 		return 0;
121 }
122 
123 static int get_hem_table_config(struct hns_roce_dev *hr_dev,
124 				struct hns_roce_hem_mhop *mhop,
125 				u32 type)
126 {
127 	struct device *dev = hr_dev->dev;
128 
129 	switch (type) {
130 	case HEM_TYPE_QPC:
131 		mhop->buf_chunk_size = 1 << (hr_dev->caps.qpc_buf_pg_sz
132 					     + PAGE_SHIFT);
133 		mhop->bt_chunk_size = 1 << (hr_dev->caps.qpc_ba_pg_sz
134 					     + PAGE_SHIFT);
135 		mhop->ba_l0_num = hr_dev->caps.qpc_bt_num;
136 		mhop->hop_num = hr_dev->caps.qpc_hop_num;
137 		break;
138 	case HEM_TYPE_MTPT:
139 		mhop->buf_chunk_size = 1 << (hr_dev->caps.mpt_buf_pg_sz
140 					     + PAGE_SHIFT);
141 		mhop->bt_chunk_size = 1 << (hr_dev->caps.mpt_ba_pg_sz
142 					     + PAGE_SHIFT);
143 		mhop->ba_l0_num = hr_dev->caps.mpt_bt_num;
144 		mhop->hop_num = hr_dev->caps.mpt_hop_num;
145 		break;
146 	case HEM_TYPE_CQC:
147 		mhop->buf_chunk_size = 1 << (hr_dev->caps.cqc_buf_pg_sz
148 					     + PAGE_SHIFT);
149 		mhop->bt_chunk_size = 1 << (hr_dev->caps.cqc_ba_pg_sz
150 					    + PAGE_SHIFT);
151 		mhop->ba_l0_num = hr_dev->caps.cqc_bt_num;
152 		mhop->hop_num = hr_dev->caps.cqc_hop_num;
153 		break;
154 	case HEM_TYPE_SCCC:
155 		mhop->buf_chunk_size = 1 << (hr_dev->caps.sccc_buf_pg_sz
156 					     + PAGE_SHIFT);
157 		mhop->bt_chunk_size = 1 << (hr_dev->caps.sccc_ba_pg_sz
158 					    + PAGE_SHIFT);
159 		mhop->ba_l0_num = hr_dev->caps.sccc_bt_num;
160 		mhop->hop_num = hr_dev->caps.sccc_hop_num;
161 		break;
162 	case HEM_TYPE_QPC_TIMER:
163 		mhop->buf_chunk_size = 1 << (hr_dev->caps.qpc_timer_buf_pg_sz
164 					     + PAGE_SHIFT);
165 		mhop->bt_chunk_size = 1 << (hr_dev->caps.qpc_timer_ba_pg_sz
166 					    + PAGE_SHIFT);
167 		mhop->ba_l0_num = hr_dev->caps.qpc_timer_bt_num;
168 		mhop->hop_num = hr_dev->caps.qpc_timer_hop_num;
169 		break;
170 	case HEM_TYPE_CQC_TIMER:
171 		mhop->buf_chunk_size = 1 << (hr_dev->caps.cqc_timer_buf_pg_sz
172 					     + PAGE_SHIFT);
173 		mhop->bt_chunk_size = 1 << (hr_dev->caps.cqc_timer_ba_pg_sz
174 					    + PAGE_SHIFT);
175 		mhop->ba_l0_num = hr_dev->caps.cqc_timer_bt_num;
176 		mhop->hop_num = hr_dev->caps.cqc_timer_hop_num;
177 		break;
178 	case HEM_TYPE_SRQC:
179 		mhop->buf_chunk_size = 1 << (hr_dev->caps.srqc_buf_pg_sz
180 					     + PAGE_SHIFT);
181 		mhop->bt_chunk_size = 1 << (hr_dev->caps.srqc_ba_pg_sz
182 					     + PAGE_SHIFT);
183 		mhop->ba_l0_num = hr_dev->caps.srqc_bt_num;
184 		mhop->hop_num = hr_dev->caps.srqc_hop_num;
185 		break;
186 	case HEM_TYPE_GMV:
187 		mhop->buf_chunk_size = 1 << (hr_dev->caps.gmv_buf_pg_sz +
188 					     PAGE_SHIFT);
189 		mhop->bt_chunk_size = 1 << (hr_dev->caps.gmv_ba_pg_sz +
190 					    PAGE_SHIFT);
191 		mhop->ba_l0_num = hr_dev->caps.gmv_bt_num;
192 		mhop->hop_num = hr_dev->caps.gmv_hop_num;
193 		break;
194 	default:
195 		dev_err(dev, "table %u not support multi-hop addressing!\n",
196 			type);
197 		return -EINVAL;
198 	}
199 
200 	return 0;
201 }
202 
203 int hns_roce_calc_hem_mhop(struct hns_roce_dev *hr_dev,
204 			   struct hns_roce_hem_table *table, unsigned long *obj,
205 			   struct hns_roce_hem_mhop *mhop)
206 {
207 	struct device *dev = hr_dev->dev;
208 	u32 chunk_ba_num;
209 	u32 chunk_size;
210 	u32 table_idx;
211 	u32 bt_num;
212 
213 	if (get_hem_table_config(hr_dev, mhop, table->type))
214 		return -EINVAL;
215 
216 	if (!obj)
217 		return 0;
218 
219 	/*
220 	 * QPC/MTPT/CQC/SRQC/SCCC alloc hem for buffer pages.
221 	 * MTT/CQE alloc hem for bt pages.
222 	 */
223 	bt_num = hns_roce_get_bt_num(table->type, mhop->hop_num);
224 	chunk_ba_num = mhop->bt_chunk_size / BA_BYTE_LEN;
225 	chunk_size = table->type < HEM_TYPE_MTT ? mhop->buf_chunk_size :
226 			      mhop->bt_chunk_size;
227 	table_idx = *obj / (chunk_size / table->obj_size);
228 	switch (bt_num) {
229 	case 3:
230 		mhop->l2_idx = table_idx & (chunk_ba_num - 1);
231 		mhop->l1_idx = table_idx / chunk_ba_num & (chunk_ba_num - 1);
232 		mhop->l0_idx = (table_idx / chunk_ba_num) / chunk_ba_num;
233 		break;
234 	case 2:
235 		mhop->l1_idx = table_idx & (chunk_ba_num - 1);
236 		mhop->l0_idx = table_idx / chunk_ba_num;
237 		break;
238 	case 1:
239 		mhop->l0_idx = table_idx;
240 		break;
241 	default:
242 		dev_err(dev, "table %u not support hop_num = %u!\n",
243 			table->type, mhop->hop_num);
244 		return -EINVAL;
245 	}
246 	if (mhop->l0_idx >= mhop->ba_l0_num)
247 		mhop->l0_idx %= mhop->ba_l0_num;
248 
249 	return 0;
250 }
251 
252 static struct hns_roce_hem *hns_roce_alloc_hem(struct hns_roce_dev *hr_dev,
253 					       int npages,
254 					       unsigned long hem_alloc_size,
255 					       gfp_t gfp_mask)
256 {
257 	struct hns_roce_hem_chunk *chunk = NULL;
258 	struct hns_roce_hem *hem;
259 	struct scatterlist *mem;
260 	int order;
261 	void *buf;
262 
263 	WARN_ON(gfp_mask & __GFP_HIGHMEM);
264 
265 	hem = kmalloc(sizeof(*hem),
266 		      gfp_mask & ~(__GFP_HIGHMEM | __GFP_NOWARN));
267 	if (!hem)
268 		return NULL;
269 
270 	INIT_LIST_HEAD(&hem->chunk_list);
271 
272 	order = get_order(hem_alloc_size);
273 
274 	while (npages > 0) {
275 		if (!chunk) {
276 			chunk = kmalloc(sizeof(*chunk),
277 				gfp_mask & ~(__GFP_HIGHMEM | __GFP_NOWARN));
278 			if (!chunk)
279 				goto fail;
280 
281 			sg_init_table(chunk->mem, HNS_ROCE_HEM_CHUNK_LEN);
282 			chunk->npages = 0;
283 			chunk->nsg = 0;
284 			memset(chunk->buf, 0, sizeof(chunk->buf));
285 			list_add_tail(&chunk->list, &hem->chunk_list);
286 		}
287 
288 		while (1 << order > npages)
289 			--order;
290 
291 		/*
292 		 * Alloc memory one time. If failed, don't alloc small block
293 		 * memory, directly return fail.
294 		 */
295 		mem = &chunk->mem[chunk->npages];
296 		buf = dma_alloc_coherent(hr_dev->dev, PAGE_SIZE << order,
297 				&sg_dma_address(mem), gfp_mask);
298 		if (!buf)
299 			goto fail;
300 
301 		chunk->buf[chunk->npages] = buf;
302 		sg_dma_len(mem) = PAGE_SIZE << order;
303 
304 		++chunk->npages;
305 		++chunk->nsg;
306 		npages -= 1 << order;
307 	}
308 
309 	return hem;
310 
311 fail:
312 	hns_roce_free_hem(hr_dev, hem);
313 	return NULL;
314 }
315 
316 void hns_roce_free_hem(struct hns_roce_dev *hr_dev, struct hns_roce_hem *hem)
317 {
318 	struct hns_roce_hem_chunk *chunk, *tmp;
319 	int i;
320 
321 	if (!hem)
322 		return;
323 
324 	list_for_each_entry_safe(chunk, tmp, &hem->chunk_list, list) {
325 		for (i = 0; i < chunk->npages; ++i)
326 			dma_free_coherent(hr_dev->dev,
327 				   sg_dma_len(&chunk->mem[i]),
328 				   chunk->buf[i],
329 				   sg_dma_address(&chunk->mem[i]));
330 		kfree(chunk);
331 	}
332 
333 	kfree(hem);
334 }
335 
336 static int calc_hem_config(struct hns_roce_dev *hr_dev,
337 			   struct hns_roce_hem_table *table, unsigned long obj,
338 			   struct hns_roce_hem_mhop *mhop,
339 			   struct hns_roce_hem_index *index)
340 {
341 	struct ib_device *ibdev = &hr_dev->ib_dev;
342 	unsigned long mhop_obj = obj;
343 	u32 l0_idx, l1_idx, l2_idx;
344 	u32 chunk_ba_num;
345 	u32 bt_num;
346 	int ret;
347 
348 	ret = hns_roce_calc_hem_mhop(hr_dev, table, &mhop_obj, mhop);
349 	if (ret)
350 		return ret;
351 
352 	l0_idx = mhop->l0_idx;
353 	l1_idx = mhop->l1_idx;
354 	l2_idx = mhop->l2_idx;
355 	chunk_ba_num = mhop->bt_chunk_size / BA_BYTE_LEN;
356 	bt_num = hns_roce_get_bt_num(table->type, mhop->hop_num);
357 	switch (bt_num) {
358 	case 3:
359 		index->l1 = l0_idx * chunk_ba_num + l1_idx;
360 		index->l0 = l0_idx;
361 		index->buf = l0_idx * chunk_ba_num * chunk_ba_num +
362 			     l1_idx * chunk_ba_num + l2_idx;
363 		break;
364 	case 2:
365 		index->l0 = l0_idx;
366 		index->buf = l0_idx * chunk_ba_num + l1_idx;
367 		break;
368 	case 1:
369 		index->buf = l0_idx;
370 		break;
371 	default:
372 		ibdev_err(ibdev, "table %u not support mhop.hop_num = %u!\n",
373 			  table->type, mhop->hop_num);
374 		return -EINVAL;
375 	}
376 
377 	if (unlikely(index->buf >= table->num_hem)) {
378 		ibdev_err(ibdev, "table %u exceed hem limt idx %llu, max %lu!\n",
379 			  table->type, index->buf, table->num_hem);
380 		return -EINVAL;
381 	}
382 
383 	return 0;
384 }
385 
386 static void free_mhop_hem(struct hns_roce_dev *hr_dev,
387 			  struct hns_roce_hem_table *table,
388 			  struct hns_roce_hem_mhop *mhop,
389 			  struct hns_roce_hem_index *index)
390 {
391 	u32 bt_size = mhop->bt_chunk_size;
392 	struct device *dev = hr_dev->dev;
393 
394 	if (index->inited & HEM_INDEX_BUF) {
395 		hns_roce_free_hem(hr_dev, table->hem[index->buf]);
396 		table->hem[index->buf] = NULL;
397 	}
398 
399 	if (index->inited & HEM_INDEX_L1) {
400 		dma_free_coherent(dev, bt_size, table->bt_l1[index->l1],
401 				  table->bt_l1_dma_addr[index->l1]);
402 		table->bt_l1[index->l1] = NULL;
403 	}
404 
405 	if (index->inited & HEM_INDEX_L0) {
406 		dma_free_coherent(dev, bt_size, table->bt_l0[index->l0],
407 				  table->bt_l0_dma_addr[index->l0]);
408 		table->bt_l0[index->l0] = NULL;
409 	}
410 }
411 
412 static int alloc_mhop_hem(struct hns_roce_dev *hr_dev,
413 			  struct hns_roce_hem_table *table,
414 			  struct hns_roce_hem_mhop *mhop,
415 			  struct hns_roce_hem_index *index)
416 {
417 	u32 bt_size = mhop->bt_chunk_size;
418 	struct device *dev = hr_dev->dev;
419 	struct hns_roce_hem_iter iter;
420 	gfp_t flag;
421 	u64 bt_ba;
422 	u32 size;
423 	int ret;
424 
425 	/* alloc L1 BA's chunk */
426 	if ((check_whether_bt_num_3(table->type, mhop->hop_num) ||
427 	     check_whether_bt_num_2(table->type, mhop->hop_num)) &&
428 	     !table->bt_l0[index->l0]) {
429 		table->bt_l0[index->l0] = dma_alloc_coherent(dev, bt_size,
430 					    &table->bt_l0_dma_addr[index->l0],
431 					    GFP_KERNEL);
432 		if (!table->bt_l0[index->l0]) {
433 			ret = -ENOMEM;
434 			goto out;
435 		}
436 		index->inited |= HEM_INDEX_L0;
437 	}
438 
439 	/* alloc L2 BA's chunk */
440 	if (check_whether_bt_num_3(table->type, mhop->hop_num) &&
441 	    !table->bt_l1[index->l1])  {
442 		table->bt_l1[index->l1] = dma_alloc_coherent(dev, bt_size,
443 					    &table->bt_l1_dma_addr[index->l1],
444 					    GFP_KERNEL);
445 		if (!table->bt_l1[index->l1]) {
446 			ret = -ENOMEM;
447 			goto err_alloc_hem;
448 		}
449 		index->inited |= HEM_INDEX_L1;
450 		*(table->bt_l0[index->l0] + mhop->l1_idx) =
451 					       table->bt_l1_dma_addr[index->l1];
452 	}
453 
454 	/*
455 	 * alloc buffer space chunk for QPC/MTPT/CQC/SRQC/SCCC.
456 	 * alloc bt space chunk for MTT/CQE.
457 	 */
458 	size = table->type < HEM_TYPE_MTT ? mhop->buf_chunk_size : bt_size;
459 	flag = (table->lowmem ? GFP_KERNEL : GFP_HIGHUSER) | __GFP_NOWARN;
460 	table->hem[index->buf] = hns_roce_alloc_hem(hr_dev, size >> PAGE_SHIFT,
461 						    size, flag);
462 	if (!table->hem[index->buf]) {
463 		ret = -ENOMEM;
464 		goto err_alloc_hem;
465 	}
466 
467 	index->inited |= HEM_INDEX_BUF;
468 	hns_roce_hem_first(table->hem[index->buf], &iter);
469 	bt_ba = hns_roce_hem_addr(&iter);
470 	if (table->type < HEM_TYPE_MTT) {
471 		if (mhop->hop_num == 2)
472 			*(table->bt_l1[index->l1] + mhop->l2_idx) = bt_ba;
473 		else if (mhop->hop_num == 1)
474 			*(table->bt_l0[index->l0] + mhop->l1_idx) = bt_ba;
475 	} else if (mhop->hop_num == 2) {
476 		*(table->bt_l0[index->l0] + mhop->l1_idx) = bt_ba;
477 	}
478 
479 	return 0;
480 err_alloc_hem:
481 	free_mhop_hem(hr_dev, table, mhop, index);
482 out:
483 	return ret;
484 }
485 
486 static int set_mhop_hem(struct hns_roce_dev *hr_dev,
487 			struct hns_roce_hem_table *table, unsigned long obj,
488 			struct hns_roce_hem_mhop *mhop,
489 			struct hns_roce_hem_index *index)
490 {
491 	struct ib_device *ibdev = &hr_dev->ib_dev;
492 	int step_idx;
493 	int ret = 0;
494 
495 	if (index->inited & HEM_INDEX_L0) {
496 		ret = hr_dev->hw->set_hem(hr_dev, table, obj, 0);
497 		if (ret) {
498 			ibdev_err(ibdev, "set HEM step 0 failed!\n");
499 			goto out;
500 		}
501 	}
502 
503 	if (index->inited & HEM_INDEX_L1) {
504 		ret = hr_dev->hw->set_hem(hr_dev, table, obj, 1);
505 		if (ret) {
506 			ibdev_err(ibdev, "set HEM step 1 failed!\n");
507 			goto out;
508 		}
509 	}
510 
511 	if (index->inited & HEM_INDEX_BUF) {
512 		if (mhop->hop_num == HNS_ROCE_HOP_NUM_0)
513 			step_idx = 0;
514 		else
515 			step_idx = mhop->hop_num;
516 		ret = hr_dev->hw->set_hem(hr_dev, table, obj, step_idx);
517 		if (ret)
518 			ibdev_err(ibdev, "set HEM step last failed!\n");
519 	}
520 out:
521 	return ret;
522 }
523 
524 static int hns_roce_table_mhop_get(struct hns_roce_dev *hr_dev,
525 				   struct hns_roce_hem_table *table,
526 				   unsigned long obj)
527 {
528 	struct ib_device *ibdev = &hr_dev->ib_dev;
529 	struct hns_roce_hem_index index = {};
530 	struct hns_roce_hem_mhop mhop = {};
531 	int ret;
532 
533 	ret = calc_hem_config(hr_dev, table, obj, &mhop, &index);
534 	if (ret) {
535 		ibdev_err(ibdev, "calc hem config failed!\n");
536 		return ret;
537 	}
538 
539 	mutex_lock(&table->mutex);
540 	if (table->hem[index.buf]) {
541 		refcount_inc(&table->hem[index.buf]->refcount);
542 		goto out;
543 	}
544 
545 	ret = alloc_mhop_hem(hr_dev, table, &mhop, &index);
546 	if (ret) {
547 		ibdev_err(ibdev, "alloc mhop hem failed!\n");
548 		goto out;
549 	}
550 
551 	/* set HEM base address to hardware */
552 	if (table->type < HEM_TYPE_MTT) {
553 		ret = set_mhop_hem(hr_dev, table, obj, &mhop, &index);
554 		if (ret) {
555 			ibdev_err(ibdev, "set HEM address to HW failed!\n");
556 			goto err_alloc;
557 		}
558 	}
559 
560 	refcount_set(&table->hem[index.buf]->refcount, 1);
561 	goto out;
562 
563 err_alloc:
564 	free_mhop_hem(hr_dev, table, &mhop, &index);
565 out:
566 	mutex_unlock(&table->mutex);
567 	return ret;
568 }
569 
570 int hns_roce_table_get(struct hns_roce_dev *hr_dev,
571 		       struct hns_roce_hem_table *table, unsigned long obj)
572 {
573 	struct device *dev = hr_dev->dev;
574 	unsigned long i;
575 	int ret = 0;
576 
577 	if (hns_roce_check_whether_mhop(hr_dev, table->type))
578 		return hns_roce_table_mhop_get(hr_dev, table, obj);
579 
580 	i = obj / (table->table_chunk_size / table->obj_size);
581 
582 	mutex_lock(&table->mutex);
583 
584 	if (table->hem[i]) {
585 		refcount_inc(&table->hem[i]->refcount);
586 		goto out;
587 	}
588 
589 	table->hem[i] = hns_roce_alloc_hem(hr_dev,
590 				       table->table_chunk_size >> PAGE_SHIFT,
591 				       table->table_chunk_size,
592 				       (table->lowmem ? GFP_KERNEL :
593 					GFP_HIGHUSER) | __GFP_NOWARN);
594 	if (!table->hem[i]) {
595 		ret = -ENOMEM;
596 		goto out;
597 	}
598 
599 	/* Set HEM base address(128K/page, pa) to Hardware */
600 	if (hr_dev->hw->set_hem(hr_dev, table, obj, HEM_HOP_STEP_DIRECT)) {
601 		hns_roce_free_hem(hr_dev, table->hem[i]);
602 		table->hem[i] = NULL;
603 		ret = -ENODEV;
604 		dev_err(dev, "set HEM base address to HW failed.\n");
605 		goto out;
606 	}
607 
608 	refcount_set(&table->hem[i]->refcount, 1);
609 out:
610 	mutex_unlock(&table->mutex);
611 	return ret;
612 }
613 
614 static void clear_mhop_hem(struct hns_roce_dev *hr_dev,
615 			   struct hns_roce_hem_table *table, unsigned long obj,
616 			   struct hns_roce_hem_mhop *mhop,
617 			   struct hns_roce_hem_index *index)
618 {
619 	struct ib_device *ibdev = &hr_dev->ib_dev;
620 	u32 hop_num = mhop->hop_num;
621 	u32 chunk_ba_num;
622 	int step_idx;
623 
624 	index->inited = HEM_INDEX_BUF;
625 	chunk_ba_num = mhop->bt_chunk_size / BA_BYTE_LEN;
626 	if (check_whether_bt_num_2(table->type, hop_num)) {
627 		if (hns_roce_check_hem_null(table->hem, index->buf,
628 					    chunk_ba_num, table->num_hem))
629 			index->inited |= HEM_INDEX_L0;
630 	} else if (check_whether_bt_num_3(table->type, hop_num)) {
631 		if (hns_roce_check_hem_null(table->hem, index->buf,
632 					    chunk_ba_num, table->num_hem)) {
633 			index->inited |= HEM_INDEX_L1;
634 			if (hns_roce_check_bt_null(table->bt_l1, index->l1,
635 						   chunk_ba_num))
636 				index->inited |= HEM_INDEX_L0;
637 		}
638 	}
639 
640 	if (table->type < HEM_TYPE_MTT) {
641 		if (hop_num == HNS_ROCE_HOP_NUM_0)
642 			step_idx = 0;
643 		else
644 			step_idx = hop_num;
645 
646 		if (hr_dev->hw->clear_hem(hr_dev, table, obj, step_idx))
647 			ibdev_warn(ibdev, "failed to clear hop%u HEM.\n", hop_num);
648 
649 		if (index->inited & HEM_INDEX_L1)
650 			if (hr_dev->hw->clear_hem(hr_dev, table, obj, 1))
651 				ibdev_warn(ibdev, "failed to clear HEM step 1.\n");
652 
653 		if (index->inited & HEM_INDEX_L0)
654 			if (hr_dev->hw->clear_hem(hr_dev, table, obj, 0))
655 				ibdev_warn(ibdev, "failed to clear HEM step 0.\n");
656 	}
657 }
658 
659 static void hns_roce_table_mhop_put(struct hns_roce_dev *hr_dev,
660 				    struct hns_roce_hem_table *table,
661 				    unsigned long obj,
662 				    int check_refcount)
663 {
664 	struct ib_device *ibdev = &hr_dev->ib_dev;
665 	struct hns_roce_hem_index index = {};
666 	struct hns_roce_hem_mhop mhop = {};
667 	int ret;
668 
669 	ret = calc_hem_config(hr_dev, table, obj, &mhop, &index);
670 	if (ret) {
671 		ibdev_err(ibdev, "calc hem config failed!\n");
672 		return;
673 	}
674 
675 	if (!check_refcount)
676 		mutex_lock(&table->mutex);
677 	else if (!refcount_dec_and_mutex_lock(&table->hem[index.buf]->refcount,
678 					      &table->mutex))
679 		return;
680 
681 	clear_mhop_hem(hr_dev, table, obj, &mhop, &index);
682 	free_mhop_hem(hr_dev, table, &mhop, &index);
683 
684 	mutex_unlock(&table->mutex);
685 }
686 
687 void hns_roce_table_put(struct hns_roce_dev *hr_dev,
688 			struct hns_roce_hem_table *table, unsigned long obj)
689 {
690 	struct device *dev = hr_dev->dev;
691 	unsigned long i;
692 
693 	if (hns_roce_check_whether_mhop(hr_dev, table->type)) {
694 		hns_roce_table_mhop_put(hr_dev, table, obj, 1);
695 		return;
696 	}
697 
698 	i = obj / (table->table_chunk_size / table->obj_size);
699 
700 	if (!refcount_dec_and_mutex_lock(&table->hem[i]->refcount,
701 					 &table->mutex))
702 		return;
703 
704 	if (hr_dev->hw->clear_hem(hr_dev, table, obj, HEM_HOP_STEP_DIRECT))
705 		dev_warn(dev, "failed to clear HEM base address.\n");
706 
707 	hns_roce_free_hem(hr_dev, table->hem[i]);
708 	table->hem[i] = NULL;
709 
710 	mutex_unlock(&table->mutex);
711 }
712 
713 void *hns_roce_table_find(struct hns_roce_dev *hr_dev,
714 			  struct hns_roce_hem_table *table,
715 			  unsigned long obj, dma_addr_t *dma_handle)
716 {
717 	struct hns_roce_hem_chunk *chunk;
718 	struct hns_roce_hem_mhop mhop;
719 	struct hns_roce_hem *hem;
720 	unsigned long mhop_obj = obj;
721 	unsigned long obj_per_chunk;
722 	unsigned long idx_offset;
723 	int offset, dma_offset;
724 	void *addr = NULL;
725 	u32 hem_idx = 0;
726 	int length;
727 	int i, j;
728 
729 	if (!table->lowmem)
730 		return NULL;
731 
732 	mutex_lock(&table->mutex);
733 
734 	if (!hns_roce_check_whether_mhop(hr_dev, table->type)) {
735 		obj_per_chunk = table->table_chunk_size / table->obj_size;
736 		hem = table->hem[obj / obj_per_chunk];
737 		idx_offset = obj % obj_per_chunk;
738 		dma_offset = offset = idx_offset * table->obj_size;
739 	} else {
740 		u32 seg_size = 64; /* 8 bytes per BA and 8 BA per segment */
741 
742 		if (hns_roce_calc_hem_mhop(hr_dev, table, &mhop_obj, &mhop))
743 			goto out;
744 		/* mtt mhop */
745 		i = mhop.l0_idx;
746 		j = mhop.l1_idx;
747 		if (mhop.hop_num == 2)
748 			hem_idx = i * (mhop.bt_chunk_size / BA_BYTE_LEN) + j;
749 		else if (mhop.hop_num == 1 ||
750 			 mhop.hop_num == HNS_ROCE_HOP_NUM_0)
751 			hem_idx = i;
752 
753 		hem = table->hem[hem_idx];
754 		dma_offset = offset = obj * seg_size % mhop.bt_chunk_size;
755 		if (mhop.hop_num == 2)
756 			dma_offset = offset = 0;
757 	}
758 
759 	if (!hem)
760 		goto out;
761 
762 	list_for_each_entry(chunk, &hem->chunk_list, list) {
763 		for (i = 0; i < chunk->npages; ++i) {
764 			length = sg_dma_len(&chunk->mem[i]);
765 			if (dma_handle && dma_offset >= 0) {
766 				if (length > (u32)dma_offset)
767 					*dma_handle = sg_dma_address(
768 						&chunk->mem[i]) + dma_offset;
769 				dma_offset -= length;
770 			}
771 
772 			if (length > (u32)offset) {
773 				addr = chunk->buf[i] + offset;
774 				goto out;
775 			}
776 			offset -= length;
777 		}
778 	}
779 
780 out:
781 	mutex_unlock(&table->mutex);
782 	return addr;
783 }
784 
785 int hns_roce_init_hem_table(struct hns_roce_dev *hr_dev,
786 			    struct hns_roce_hem_table *table, u32 type,
787 			    unsigned long obj_size, unsigned long nobj,
788 			    int use_lowmem)
789 {
790 	unsigned long obj_per_chunk;
791 	unsigned long num_hem;
792 
793 	if (!hns_roce_check_whether_mhop(hr_dev, type)) {
794 		table->table_chunk_size = hr_dev->caps.chunk_sz;
795 		obj_per_chunk = table->table_chunk_size / obj_size;
796 		num_hem = DIV_ROUND_UP(nobj, obj_per_chunk);
797 
798 		table->hem = kcalloc(num_hem, sizeof(*table->hem), GFP_KERNEL);
799 		if (!table->hem)
800 			return -ENOMEM;
801 	} else {
802 		struct hns_roce_hem_mhop mhop = {};
803 		unsigned long buf_chunk_size;
804 		unsigned long bt_chunk_size;
805 		unsigned long bt_chunk_num;
806 		unsigned long num_bt_l0;
807 		u32 hop_num;
808 
809 		if (get_hem_table_config(hr_dev, &mhop, type))
810 			return -EINVAL;
811 
812 		buf_chunk_size = mhop.buf_chunk_size;
813 		bt_chunk_size = mhop.bt_chunk_size;
814 		num_bt_l0 = mhop.ba_l0_num;
815 		hop_num = mhop.hop_num;
816 
817 		obj_per_chunk = buf_chunk_size / obj_size;
818 		num_hem = DIV_ROUND_UP(nobj, obj_per_chunk);
819 		bt_chunk_num = bt_chunk_size / BA_BYTE_LEN;
820 
821 		if (type >= HEM_TYPE_MTT)
822 			num_bt_l0 = bt_chunk_num;
823 
824 		table->hem = kcalloc(num_hem, sizeof(*table->hem),
825 					 GFP_KERNEL);
826 		if (!table->hem)
827 			goto err_kcalloc_hem_buf;
828 
829 		if (check_whether_bt_num_3(type, hop_num)) {
830 			unsigned long num_bt_l1;
831 
832 			num_bt_l1 = DIV_ROUND_UP(num_hem, bt_chunk_num);
833 			table->bt_l1 = kcalloc(num_bt_l1,
834 					       sizeof(*table->bt_l1),
835 					       GFP_KERNEL);
836 			if (!table->bt_l1)
837 				goto err_kcalloc_bt_l1;
838 
839 			table->bt_l1_dma_addr = kcalloc(num_bt_l1,
840 						 sizeof(*table->bt_l1_dma_addr),
841 						 GFP_KERNEL);
842 
843 			if (!table->bt_l1_dma_addr)
844 				goto err_kcalloc_l1_dma;
845 		}
846 
847 		if (check_whether_bt_num_2(type, hop_num) ||
848 			check_whether_bt_num_3(type, hop_num)) {
849 			table->bt_l0 = kcalloc(num_bt_l0, sizeof(*table->bt_l0),
850 					       GFP_KERNEL);
851 			if (!table->bt_l0)
852 				goto err_kcalloc_bt_l0;
853 
854 			table->bt_l0_dma_addr = kcalloc(num_bt_l0,
855 						 sizeof(*table->bt_l0_dma_addr),
856 						 GFP_KERNEL);
857 			if (!table->bt_l0_dma_addr)
858 				goto err_kcalloc_l0_dma;
859 		}
860 	}
861 
862 	table->type = type;
863 	table->num_hem = num_hem;
864 	table->obj_size = obj_size;
865 	table->lowmem = use_lowmem;
866 	mutex_init(&table->mutex);
867 
868 	return 0;
869 
870 err_kcalloc_l0_dma:
871 	kfree(table->bt_l0);
872 	table->bt_l0 = NULL;
873 
874 err_kcalloc_bt_l0:
875 	kfree(table->bt_l1_dma_addr);
876 	table->bt_l1_dma_addr = NULL;
877 
878 err_kcalloc_l1_dma:
879 	kfree(table->bt_l1);
880 	table->bt_l1 = NULL;
881 
882 err_kcalloc_bt_l1:
883 	kfree(table->hem);
884 	table->hem = NULL;
885 
886 err_kcalloc_hem_buf:
887 	return -ENOMEM;
888 }
889 
890 static void hns_roce_cleanup_mhop_hem_table(struct hns_roce_dev *hr_dev,
891 					    struct hns_roce_hem_table *table)
892 {
893 	struct hns_roce_hem_mhop mhop;
894 	u32 buf_chunk_size;
895 	u64 obj;
896 	int i;
897 
898 	if (hns_roce_calc_hem_mhop(hr_dev, table, NULL, &mhop))
899 		return;
900 	buf_chunk_size = table->type < HEM_TYPE_MTT ? mhop.buf_chunk_size :
901 					mhop.bt_chunk_size;
902 
903 	for (i = 0; i < table->num_hem; ++i) {
904 		obj = i * buf_chunk_size / table->obj_size;
905 		if (table->hem[i])
906 			hns_roce_table_mhop_put(hr_dev, table, obj, 0);
907 	}
908 
909 	kfree(table->hem);
910 	table->hem = NULL;
911 	kfree(table->bt_l1);
912 	table->bt_l1 = NULL;
913 	kfree(table->bt_l1_dma_addr);
914 	table->bt_l1_dma_addr = NULL;
915 	kfree(table->bt_l0);
916 	table->bt_l0 = NULL;
917 	kfree(table->bt_l0_dma_addr);
918 	table->bt_l0_dma_addr = NULL;
919 }
920 
921 void hns_roce_cleanup_hem_table(struct hns_roce_dev *hr_dev,
922 				struct hns_roce_hem_table *table)
923 {
924 	struct device *dev = hr_dev->dev;
925 	unsigned long i;
926 
927 	if (hns_roce_check_whether_mhop(hr_dev, table->type)) {
928 		hns_roce_cleanup_mhop_hem_table(hr_dev, table);
929 		return;
930 	}
931 
932 	for (i = 0; i < table->num_hem; ++i)
933 		if (table->hem[i]) {
934 			if (hr_dev->hw->clear_hem(hr_dev, table,
935 			    i * table->table_chunk_size / table->obj_size, 0))
936 				dev_err(dev, "Clear HEM base address failed.\n");
937 
938 			hns_roce_free_hem(hr_dev, table->hem[i]);
939 		}
940 
941 	kfree(table->hem);
942 }
943 
944 void hns_roce_cleanup_hem(struct hns_roce_dev *hr_dev)
945 {
946 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_SRQ)
947 		hns_roce_cleanup_hem_table(hr_dev,
948 					   &hr_dev->srq_table.table);
949 	hns_roce_cleanup_hem_table(hr_dev, &hr_dev->cq_table.table);
950 	if (hr_dev->caps.qpc_timer_entry_sz)
951 		hns_roce_cleanup_hem_table(hr_dev,
952 					   &hr_dev->qpc_timer_table);
953 	if (hr_dev->caps.cqc_timer_entry_sz)
954 		hns_roce_cleanup_hem_table(hr_dev,
955 					   &hr_dev->cqc_timer_table);
956 	if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_QP_FLOW_CTRL)
957 		hns_roce_cleanup_hem_table(hr_dev,
958 					   &hr_dev->qp_table.sccc_table);
959 	if (hr_dev->caps.trrl_entry_sz)
960 		hns_roce_cleanup_hem_table(hr_dev,
961 					   &hr_dev->qp_table.trrl_table);
962 
963 	if (hr_dev->caps.gmv_entry_sz)
964 		hns_roce_cleanup_hem_table(hr_dev, &hr_dev->gmv_table);
965 
966 	hns_roce_cleanup_hem_table(hr_dev, &hr_dev->qp_table.irrl_table);
967 	hns_roce_cleanup_hem_table(hr_dev, &hr_dev->qp_table.qp_table);
968 	hns_roce_cleanup_hem_table(hr_dev, &hr_dev->mr_table.mtpt_table);
969 }
970 
971 struct hns_roce_hem_item {
972 	struct list_head list; /* link all hems in the same bt level */
973 	struct list_head sibling; /* link all hems in last hop for mtt */
974 	void *addr;
975 	dma_addr_t dma_addr;
976 	size_t count; /* max ba numbers */
977 	int start; /* start buf offset in this hem */
978 	int end; /* end buf offset in this hem */
979 };
980 
981 /* All HEM items are linked in a tree structure */
982 struct hns_roce_hem_head {
983 	struct list_head branch[HNS_ROCE_MAX_BT_REGION];
984 	struct list_head root;
985 	struct list_head leaf;
986 };
987 
988 static struct hns_roce_hem_item *
989 hem_list_alloc_item(struct hns_roce_dev *hr_dev, int start, int end, int count,
990 		    bool exist_bt, int bt_level)
991 {
992 	struct hns_roce_hem_item *hem;
993 
994 	hem = kzalloc(sizeof(*hem), GFP_KERNEL);
995 	if (!hem)
996 		return NULL;
997 
998 	if (exist_bt) {
999 		hem->addr = dma_alloc_coherent(hr_dev->dev, count * BA_BYTE_LEN,
1000 					       &hem->dma_addr, GFP_KERNEL);
1001 		if (!hem->addr) {
1002 			kfree(hem);
1003 			return NULL;
1004 		}
1005 	}
1006 
1007 	hem->count = count;
1008 	hem->start = start;
1009 	hem->end = end;
1010 	INIT_LIST_HEAD(&hem->list);
1011 	INIT_LIST_HEAD(&hem->sibling);
1012 
1013 	return hem;
1014 }
1015 
1016 static void hem_list_free_item(struct hns_roce_dev *hr_dev,
1017 			       struct hns_roce_hem_item *hem, bool exist_bt)
1018 {
1019 	if (exist_bt)
1020 		dma_free_coherent(hr_dev->dev, hem->count * BA_BYTE_LEN,
1021 				  hem->addr, hem->dma_addr);
1022 	kfree(hem);
1023 }
1024 
1025 static void hem_list_free_all(struct hns_roce_dev *hr_dev,
1026 			      struct list_head *head, bool exist_bt)
1027 {
1028 	struct hns_roce_hem_item *hem, *temp_hem;
1029 
1030 	list_for_each_entry_safe(hem, temp_hem, head, list) {
1031 		list_del(&hem->list);
1032 		hem_list_free_item(hr_dev, hem, exist_bt);
1033 	}
1034 }
1035 
1036 static void hem_list_link_bt(struct hns_roce_dev *hr_dev, void *base_addr,
1037 			     u64 table_addr)
1038 {
1039 	*(u64 *)(base_addr) = table_addr;
1040 }
1041 
1042 /* assign L0 table address to hem from root bt */
1043 static void hem_list_assign_bt(struct hns_roce_dev *hr_dev,
1044 			       struct hns_roce_hem_item *hem, void *cpu_addr,
1045 			       u64 phy_addr)
1046 {
1047 	hem->addr = cpu_addr;
1048 	hem->dma_addr = (dma_addr_t)phy_addr;
1049 }
1050 
1051 static inline bool hem_list_page_is_in_range(struct hns_roce_hem_item *hem,
1052 					     int offset)
1053 {
1054 	return (hem->start <= offset && offset <= hem->end);
1055 }
1056 
1057 static struct hns_roce_hem_item *hem_list_search_item(struct list_head *ba_list,
1058 						      int page_offset)
1059 {
1060 	struct hns_roce_hem_item *hem, *temp_hem;
1061 	struct hns_roce_hem_item *found = NULL;
1062 
1063 	list_for_each_entry_safe(hem, temp_hem, ba_list, list) {
1064 		if (hem_list_page_is_in_range(hem, page_offset)) {
1065 			found = hem;
1066 			break;
1067 		}
1068 	}
1069 
1070 	return found;
1071 }
1072 
1073 static bool hem_list_is_bottom_bt(int hopnum, int bt_level)
1074 {
1075 	/*
1076 	 * hopnum    base address table levels
1077 	 * 0		L0(buf)
1078 	 * 1		L0 -> buf
1079 	 * 2		L0 -> L1 -> buf
1080 	 * 3		L0 -> L1 -> L2 -> buf
1081 	 */
1082 	return bt_level >= (hopnum ? hopnum - 1 : hopnum);
1083 }
1084 
1085 /*
1086  * calc base address entries num
1087  * @hopnum: num of mutihop addressing
1088  * @bt_level: base address table level
1089  * @unit: ba entries per bt page
1090  */
1091 static u32 hem_list_calc_ba_range(int hopnum, int bt_level, int unit)
1092 {
1093 	u32 step;
1094 	int max;
1095 	int i;
1096 
1097 	if (hopnum <= bt_level)
1098 		return 0;
1099 	/*
1100 	 * hopnum  bt_level   range
1101 	 * 1	      0       unit
1102 	 * ------------
1103 	 * 2	      0       unit * unit
1104 	 * 2	      1       unit
1105 	 * ------------
1106 	 * 3	      0       unit * unit * unit
1107 	 * 3	      1       unit * unit
1108 	 * 3	      2       unit
1109 	 */
1110 	step = 1;
1111 	max = hopnum - bt_level;
1112 	for (i = 0; i < max; i++)
1113 		step = step * unit;
1114 
1115 	return step;
1116 }
1117 
1118 /*
1119  * calc the root ba entries which could cover all regions
1120  * @regions: buf region array
1121  * @region_cnt: array size of @regions
1122  * @unit: ba entries per bt page
1123  */
1124 int hns_roce_hem_list_calc_root_ba(const struct hns_roce_buf_region *regions,
1125 				   int region_cnt, int unit)
1126 {
1127 	struct hns_roce_buf_region *r;
1128 	int total = 0;
1129 	int step;
1130 	int i;
1131 
1132 	for (i = 0; i < region_cnt; i++) {
1133 		r = (struct hns_roce_buf_region *)&regions[i];
1134 		if (r->hopnum > 1) {
1135 			step = hem_list_calc_ba_range(r->hopnum, 1, unit);
1136 			if (step > 0)
1137 				total += (r->count + step - 1) / step;
1138 		} else {
1139 			total += r->count;
1140 		}
1141 	}
1142 
1143 	return total;
1144 }
1145 
1146 static int hem_list_alloc_mid_bt(struct hns_roce_dev *hr_dev,
1147 				 const struct hns_roce_buf_region *r, int unit,
1148 				 int offset, struct list_head *mid_bt,
1149 				 struct list_head *btm_bt)
1150 {
1151 	struct hns_roce_hem_item *hem_ptrs[HNS_ROCE_MAX_BT_LEVEL] = { NULL };
1152 	struct list_head temp_list[HNS_ROCE_MAX_BT_LEVEL];
1153 	struct hns_roce_hem_item *cur, *pre;
1154 	const int hopnum = r->hopnum;
1155 	int start_aligned;
1156 	int distance;
1157 	int ret = 0;
1158 	int max_ofs;
1159 	int level;
1160 	u32 step;
1161 	int end;
1162 
1163 	if (hopnum <= 1)
1164 		return 0;
1165 
1166 	if (hopnum > HNS_ROCE_MAX_BT_LEVEL) {
1167 		dev_err(hr_dev->dev, "invalid hopnum %d!\n", hopnum);
1168 		return -EINVAL;
1169 	}
1170 
1171 	if (offset < r->offset) {
1172 		dev_err(hr_dev->dev, "invalid offset %d, min %u!\n",
1173 			offset, r->offset);
1174 		return -EINVAL;
1175 	}
1176 
1177 	distance = offset - r->offset;
1178 	max_ofs = r->offset + r->count - 1;
1179 	for (level = 0; level < hopnum; level++)
1180 		INIT_LIST_HEAD(&temp_list[level]);
1181 
1182 	/* config L1 bt to last bt and link them to corresponding parent */
1183 	for (level = 1; level < hopnum; level++) {
1184 		cur = hem_list_search_item(&mid_bt[level], offset);
1185 		if (cur) {
1186 			hem_ptrs[level] = cur;
1187 			continue;
1188 		}
1189 
1190 		step = hem_list_calc_ba_range(hopnum, level, unit);
1191 		if (step < 1) {
1192 			ret = -EINVAL;
1193 			goto err_exit;
1194 		}
1195 
1196 		start_aligned = (distance / step) * step + r->offset;
1197 		end = min_t(int, start_aligned + step - 1, max_ofs);
1198 		cur = hem_list_alloc_item(hr_dev, start_aligned, end, unit,
1199 					  true, level);
1200 		if (!cur) {
1201 			ret = -ENOMEM;
1202 			goto err_exit;
1203 		}
1204 		hem_ptrs[level] = cur;
1205 		list_add(&cur->list, &temp_list[level]);
1206 		if (hem_list_is_bottom_bt(hopnum, level))
1207 			list_add(&cur->sibling, &temp_list[0]);
1208 
1209 		/* link bt to parent bt */
1210 		if (level > 1) {
1211 			pre = hem_ptrs[level - 1];
1212 			step = (cur->start - pre->start) / step * BA_BYTE_LEN;
1213 			hem_list_link_bt(hr_dev, pre->addr + step,
1214 					 cur->dma_addr);
1215 		}
1216 	}
1217 
1218 	list_splice(&temp_list[0], btm_bt);
1219 	for (level = 1; level < hopnum; level++)
1220 		list_splice(&temp_list[level], &mid_bt[level]);
1221 
1222 	return 0;
1223 
1224 err_exit:
1225 	for (level = 1; level < hopnum; level++)
1226 		hem_list_free_all(hr_dev, &temp_list[level], true);
1227 
1228 	return ret;
1229 }
1230 
1231 static struct hns_roce_hem_item *
1232 alloc_root_hem(struct hns_roce_dev *hr_dev, int unit, int *max_ba_num,
1233 	       const struct hns_roce_buf_region *regions, int region_cnt)
1234 {
1235 	const struct hns_roce_buf_region *r;
1236 	struct hns_roce_hem_item *hem;
1237 	int ba_num;
1238 	int offset;
1239 
1240 	ba_num = hns_roce_hem_list_calc_root_ba(regions, region_cnt, unit);
1241 	if (ba_num < 1)
1242 		return ERR_PTR(-ENOMEM);
1243 
1244 	if (ba_num > unit)
1245 		return ERR_PTR(-ENOBUFS);
1246 
1247 	offset = regions[0].offset;
1248 	/* indicate to last region */
1249 	r = &regions[region_cnt - 1];
1250 	hem = hem_list_alloc_item(hr_dev, offset, r->offset + r->count - 1,
1251 				  ba_num, true, 0);
1252 	if (!hem)
1253 		return ERR_PTR(-ENOMEM);
1254 
1255 	*max_ba_num = ba_num;
1256 
1257 	return hem;
1258 }
1259 
1260 static int alloc_fake_root_bt(struct hns_roce_dev *hr_dev, void *cpu_base,
1261 			      u64 phy_base, const struct hns_roce_buf_region *r,
1262 			      struct list_head *branch_head,
1263 			      struct list_head *leaf_head)
1264 {
1265 	struct hns_roce_hem_item *hem;
1266 
1267 	hem = hem_list_alloc_item(hr_dev, r->offset, r->offset + r->count - 1,
1268 				  r->count, false, 0);
1269 	if (!hem)
1270 		return -ENOMEM;
1271 
1272 	hem_list_assign_bt(hr_dev, hem, cpu_base, phy_base);
1273 	list_add(&hem->list, branch_head);
1274 	list_add(&hem->sibling, leaf_head);
1275 
1276 	return r->count;
1277 }
1278 
1279 static int setup_middle_bt(struct hns_roce_dev *hr_dev, void *cpu_base,
1280 			   int unit, const struct hns_roce_buf_region *r,
1281 			   const struct list_head *branch_head)
1282 {
1283 	struct hns_roce_hem_item *hem, *temp_hem;
1284 	int total = 0;
1285 	int offset;
1286 	int step;
1287 
1288 	step = hem_list_calc_ba_range(r->hopnum, 1, unit);
1289 	if (step < 1)
1290 		return -EINVAL;
1291 
1292 	/* if exist mid bt, link L1 to L0 */
1293 	list_for_each_entry_safe(hem, temp_hem, branch_head, list) {
1294 		offset = (hem->start - r->offset) / step * BA_BYTE_LEN;
1295 		hem_list_link_bt(hr_dev, cpu_base + offset, hem->dma_addr);
1296 		total++;
1297 	}
1298 
1299 	return total;
1300 }
1301 
1302 static int
1303 setup_root_hem(struct hns_roce_dev *hr_dev, struct hns_roce_hem_list *hem_list,
1304 	       int unit, int max_ba_num, struct hns_roce_hem_head *head,
1305 	       const struct hns_roce_buf_region *regions, int region_cnt)
1306 {
1307 	const struct hns_roce_buf_region *r;
1308 	struct hns_roce_hem_item *root_hem;
1309 	void *cpu_base;
1310 	u64 phy_base;
1311 	int i, total;
1312 	int ret;
1313 
1314 	root_hem = list_first_entry(&head->root,
1315 				    struct hns_roce_hem_item, list);
1316 	if (!root_hem)
1317 		return -ENOMEM;
1318 
1319 	total = 0;
1320 	for (i = 0; i < region_cnt && total < max_ba_num; i++) {
1321 		r = &regions[i];
1322 		if (!r->count)
1323 			continue;
1324 
1325 		/* all regions's mid[x][0] shared the root_bt's trunk */
1326 		cpu_base = root_hem->addr + total * BA_BYTE_LEN;
1327 		phy_base = root_hem->dma_addr + total * BA_BYTE_LEN;
1328 
1329 		/* if hopnum is 0 or 1, cut a new fake hem from the root bt
1330 		 * which's address share to all regions.
1331 		 */
1332 		if (hem_list_is_bottom_bt(r->hopnum, 0))
1333 			ret = alloc_fake_root_bt(hr_dev, cpu_base, phy_base, r,
1334 						 &head->branch[i], &head->leaf);
1335 		else
1336 			ret = setup_middle_bt(hr_dev, cpu_base, unit, r,
1337 					      &hem_list->mid_bt[i][1]);
1338 
1339 		if (ret < 0)
1340 			return ret;
1341 
1342 		total += ret;
1343 	}
1344 
1345 	list_splice(&head->leaf, &hem_list->btm_bt);
1346 	list_splice(&head->root, &hem_list->root_bt);
1347 	for (i = 0; i < region_cnt; i++)
1348 		list_splice(&head->branch[i], &hem_list->mid_bt[i][0]);
1349 
1350 	return 0;
1351 }
1352 
1353 static int hem_list_alloc_root_bt(struct hns_roce_dev *hr_dev,
1354 				  struct hns_roce_hem_list *hem_list, int unit,
1355 				  const struct hns_roce_buf_region *regions,
1356 				  int region_cnt)
1357 {
1358 	struct hns_roce_hem_item *root_hem;
1359 	struct hns_roce_hem_head head;
1360 	int max_ba_num;
1361 	int ret;
1362 	int i;
1363 
1364 	root_hem = hem_list_search_item(&hem_list->root_bt, regions[0].offset);
1365 	if (root_hem)
1366 		return 0;
1367 
1368 	max_ba_num = 0;
1369 	root_hem = alloc_root_hem(hr_dev, unit, &max_ba_num, regions,
1370 				  region_cnt);
1371 	if (IS_ERR(root_hem))
1372 		return PTR_ERR(root_hem);
1373 
1374 	/* List head for storing all allocated HEM items */
1375 	INIT_LIST_HEAD(&head.root);
1376 	INIT_LIST_HEAD(&head.leaf);
1377 	for (i = 0; i < region_cnt; i++)
1378 		INIT_LIST_HEAD(&head.branch[i]);
1379 
1380 	hem_list->root_ba = root_hem->dma_addr;
1381 	list_add(&root_hem->list, &head.root);
1382 	ret = setup_root_hem(hr_dev, hem_list, unit, max_ba_num, &head, regions,
1383 			     region_cnt);
1384 	if (ret) {
1385 		for (i = 0; i < region_cnt; i++)
1386 			hem_list_free_all(hr_dev, &head.branch[i], false);
1387 
1388 		hem_list_free_all(hr_dev, &head.root, true);
1389 	}
1390 
1391 	return ret;
1392 }
1393 
1394 /* construct the base address table and link them by address hop config */
1395 int hns_roce_hem_list_request(struct hns_roce_dev *hr_dev,
1396 			      struct hns_roce_hem_list *hem_list,
1397 			      const struct hns_roce_buf_region *regions,
1398 			      int region_cnt, unsigned int bt_pg_shift)
1399 {
1400 	const struct hns_roce_buf_region *r;
1401 	int ofs, end;
1402 	int unit;
1403 	int ret;
1404 	int i;
1405 
1406 	if (region_cnt > HNS_ROCE_MAX_BT_REGION) {
1407 		dev_err(hr_dev->dev, "invalid region region_cnt %d!\n",
1408 			region_cnt);
1409 		return -EINVAL;
1410 	}
1411 
1412 	unit = (1 << bt_pg_shift) / BA_BYTE_LEN;
1413 	for (i = 0; i < region_cnt; i++) {
1414 		r = &regions[i];
1415 		if (!r->count)
1416 			continue;
1417 
1418 		end = r->offset + r->count;
1419 		for (ofs = r->offset; ofs < end; ofs += unit) {
1420 			ret = hem_list_alloc_mid_bt(hr_dev, r, unit, ofs,
1421 						    hem_list->mid_bt[i],
1422 						    &hem_list->btm_bt);
1423 			if (ret) {
1424 				dev_err(hr_dev->dev,
1425 					"alloc hem trunk fail ret=%d!\n", ret);
1426 				goto err_alloc;
1427 			}
1428 		}
1429 	}
1430 
1431 	ret = hem_list_alloc_root_bt(hr_dev, hem_list, unit, regions,
1432 				     region_cnt);
1433 	if (ret)
1434 		dev_err(hr_dev->dev, "alloc hem root fail ret=%d!\n", ret);
1435 	else
1436 		return 0;
1437 
1438 err_alloc:
1439 	hns_roce_hem_list_release(hr_dev, hem_list);
1440 
1441 	return ret;
1442 }
1443 
1444 void hns_roce_hem_list_release(struct hns_roce_dev *hr_dev,
1445 			       struct hns_roce_hem_list *hem_list)
1446 {
1447 	int i, j;
1448 
1449 	for (i = 0; i < HNS_ROCE_MAX_BT_REGION; i++)
1450 		for (j = 0; j < HNS_ROCE_MAX_BT_LEVEL; j++)
1451 			hem_list_free_all(hr_dev, &hem_list->mid_bt[i][j],
1452 					  j != 0);
1453 
1454 	hem_list_free_all(hr_dev, &hem_list->root_bt, true);
1455 	INIT_LIST_HEAD(&hem_list->btm_bt);
1456 	hem_list->root_ba = 0;
1457 }
1458 
1459 void hns_roce_hem_list_init(struct hns_roce_hem_list *hem_list)
1460 {
1461 	int i, j;
1462 
1463 	INIT_LIST_HEAD(&hem_list->root_bt);
1464 	INIT_LIST_HEAD(&hem_list->btm_bt);
1465 	for (i = 0; i < HNS_ROCE_MAX_BT_REGION; i++)
1466 		for (j = 0; j < HNS_ROCE_MAX_BT_LEVEL; j++)
1467 			INIT_LIST_HEAD(&hem_list->mid_bt[i][j]);
1468 }
1469 
1470 void *hns_roce_hem_list_find_mtt(struct hns_roce_dev *hr_dev,
1471 				 struct hns_roce_hem_list *hem_list,
1472 				 int offset, int *mtt_cnt, u64 *phy_addr)
1473 {
1474 	struct list_head *head = &hem_list->btm_bt;
1475 	struct hns_roce_hem_item *hem, *temp_hem;
1476 	void *cpu_base = NULL;
1477 	u64 phy_base = 0;
1478 	int nr = 0;
1479 
1480 	list_for_each_entry_safe(hem, temp_hem, head, sibling) {
1481 		if (hem_list_page_is_in_range(hem, offset)) {
1482 			nr = offset - hem->start;
1483 			cpu_base = hem->addr + nr * BA_BYTE_LEN;
1484 			phy_base = hem->dma_addr + nr * BA_BYTE_LEN;
1485 			nr = hem->end + 1 - offset;
1486 			break;
1487 		}
1488 	}
1489 
1490 	if (mtt_cnt)
1491 		*mtt_cnt = nr;
1492 
1493 	if (phy_addr)
1494 		*phy_addr = phy_base;
1495 
1496 	return cpu_base;
1497 }
1498