xref: /linux/drivers/infiniband/hw/hns/hns_roce_hem.c (revision 17cfcb68af3bc7d5e8ae08779b1853310a2949f3)
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 DMA_ADDR_T_SHIFT		12
40 #define BT_BA_SHIFT			32
41 
42 bool hns_roce_check_whether_mhop(struct hns_roce_dev *hr_dev, u32 type)
43 {
44 	int hop_num = 0;
45 
46 	switch (type) {
47 	case HEM_TYPE_QPC:
48 		hop_num = hr_dev->caps.qpc_hop_num;
49 		break;
50 	case HEM_TYPE_MTPT:
51 		hop_num = hr_dev->caps.mpt_hop_num;
52 		break;
53 	case HEM_TYPE_CQC:
54 		hop_num = hr_dev->caps.cqc_hop_num;
55 		break;
56 	case HEM_TYPE_SRQC:
57 		hop_num = hr_dev->caps.srqc_hop_num;
58 		break;
59 	case HEM_TYPE_SCCC:
60 		hop_num = hr_dev->caps.sccc_hop_num;
61 		break;
62 	case HEM_TYPE_QPC_TIMER:
63 		hop_num = hr_dev->caps.qpc_timer_hop_num;
64 		break;
65 	case HEM_TYPE_CQC_TIMER:
66 		hop_num = hr_dev->caps.cqc_timer_hop_num;
67 		break;
68 	case HEM_TYPE_CQE:
69 		hop_num = hr_dev->caps.cqe_hop_num;
70 		break;
71 	case HEM_TYPE_MTT:
72 		hop_num = hr_dev->caps.mtt_hop_num;
73 		break;
74 	case HEM_TYPE_SRQWQE:
75 		hop_num = hr_dev->caps.srqwqe_hop_num;
76 		break;
77 	case HEM_TYPE_IDX:
78 		hop_num = hr_dev->caps.idx_hop_num;
79 		break;
80 	default:
81 		return false;
82 	}
83 
84 	return hop_num ? true : false;
85 }
86 
87 static bool hns_roce_check_hem_null(struct hns_roce_hem **hem, u64 start_idx,
88 			    u32 bt_chunk_num, u64 hem_max_num)
89 {
90 	u64 check_max_num = start_idx + bt_chunk_num;
91 	u64 i;
92 
93 	for (i = start_idx; (i < check_max_num) && (i < hem_max_num); i++)
94 		if (hem[i])
95 			return false;
96 
97 	return true;
98 }
99 
100 static bool hns_roce_check_bt_null(u64 **bt, u64 start_idx, u32 bt_chunk_num)
101 {
102 	int i;
103 
104 	for (i = 0; i < bt_chunk_num; i++)
105 		if (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_MTT:
187 		mhop->buf_chunk_size = 1 << (hr_dev->caps.mtt_buf_pg_sz
188 					     + PAGE_SHIFT);
189 		mhop->bt_chunk_size = 1 << (hr_dev->caps.mtt_ba_pg_sz
190 					     + PAGE_SHIFT);
191 		mhop->ba_l0_num = mhop->bt_chunk_size / BA_BYTE_LEN;
192 		mhop->hop_num = hr_dev->caps.mtt_hop_num;
193 		break;
194 	case HEM_TYPE_CQE:
195 		mhop->buf_chunk_size = 1 << (hr_dev->caps.cqe_buf_pg_sz
196 					     + PAGE_SHIFT);
197 		mhop->bt_chunk_size = 1 << (hr_dev->caps.cqe_ba_pg_sz
198 					     + PAGE_SHIFT);
199 		mhop->ba_l0_num = mhop->bt_chunk_size / BA_BYTE_LEN;
200 		mhop->hop_num = hr_dev->caps.cqe_hop_num;
201 		break;
202 	case HEM_TYPE_SRQWQE:
203 		mhop->buf_chunk_size = 1 << (hr_dev->caps.srqwqe_buf_pg_sz
204 					    + PAGE_SHIFT);
205 		mhop->bt_chunk_size = 1 << (hr_dev->caps.srqwqe_ba_pg_sz
206 					    + PAGE_SHIFT);
207 		mhop->ba_l0_num = mhop->bt_chunk_size / BA_BYTE_LEN;
208 		mhop->hop_num = hr_dev->caps.srqwqe_hop_num;
209 		break;
210 	case HEM_TYPE_IDX:
211 		mhop->buf_chunk_size = 1 << (hr_dev->caps.idx_buf_pg_sz
212 				       + PAGE_SHIFT);
213 		mhop->bt_chunk_size = 1 << (hr_dev->caps.idx_ba_pg_sz
214 				       + PAGE_SHIFT);
215 		mhop->ba_l0_num = mhop->bt_chunk_size / BA_BYTE_LEN;
216 		mhop->hop_num = hr_dev->caps.idx_hop_num;
217 		break;
218 	default:
219 		dev_err(dev, "Table %d not support multi-hop addressing!\n",
220 			type);
221 		return -EINVAL;
222 	}
223 
224 	return 0;
225 }
226 
227 int hns_roce_calc_hem_mhop(struct hns_roce_dev *hr_dev,
228 			   struct hns_roce_hem_table *table, unsigned long *obj,
229 			   struct hns_roce_hem_mhop *mhop)
230 {
231 	struct device *dev = hr_dev->dev;
232 	u32 chunk_ba_num;
233 	u32 table_idx;
234 	u32 bt_num;
235 	u32 chunk_size;
236 
237 	if (get_hem_table_config(hr_dev, mhop, table->type))
238 		return -EINVAL;
239 
240 	if (!obj)
241 		return 0;
242 
243 	/*
244 	 * QPC/MTPT/CQC/SRQC/SCCC alloc hem for buffer pages.
245 	 * MTT/CQE alloc hem for bt pages.
246 	 */
247 	bt_num = hns_roce_get_bt_num(table->type, mhop->hop_num);
248 	chunk_ba_num = mhop->bt_chunk_size / BA_BYTE_LEN;
249 	chunk_size = table->type < HEM_TYPE_MTT ? mhop->buf_chunk_size :
250 			      mhop->bt_chunk_size;
251 	table_idx = (*obj & (table->num_obj - 1)) /
252 		     (chunk_size / table->obj_size);
253 	switch (bt_num) {
254 	case 3:
255 		mhop->l2_idx = table_idx & (chunk_ba_num - 1);
256 		mhop->l1_idx = table_idx / chunk_ba_num & (chunk_ba_num - 1);
257 		mhop->l0_idx = (table_idx / chunk_ba_num) / chunk_ba_num;
258 		break;
259 	case 2:
260 		mhop->l1_idx = table_idx & (chunk_ba_num - 1);
261 		mhop->l0_idx = table_idx / chunk_ba_num;
262 		break;
263 	case 1:
264 		mhop->l0_idx = table_idx;
265 		break;
266 	default:
267 		dev_err(dev, "Table %d not support hop_num = %d!\n",
268 			     table->type, mhop->hop_num);
269 		return -EINVAL;
270 	}
271 	if (mhop->l0_idx >= mhop->ba_l0_num)
272 		mhop->l0_idx %= mhop->ba_l0_num;
273 
274 	return 0;
275 }
276 
277 static struct hns_roce_hem *hns_roce_alloc_hem(struct hns_roce_dev *hr_dev,
278 					       int npages,
279 					       unsigned long hem_alloc_size,
280 					       gfp_t gfp_mask)
281 {
282 	struct hns_roce_hem_chunk *chunk = NULL;
283 	struct hns_roce_hem *hem;
284 	struct scatterlist *mem;
285 	int order;
286 	void *buf;
287 
288 	WARN_ON(gfp_mask & __GFP_HIGHMEM);
289 
290 	hem = kmalloc(sizeof(*hem),
291 		      gfp_mask & ~(__GFP_HIGHMEM | __GFP_NOWARN));
292 	if (!hem)
293 		return NULL;
294 
295 	hem->refcount = 0;
296 	INIT_LIST_HEAD(&hem->chunk_list);
297 
298 	order = get_order(hem_alloc_size);
299 
300 	while (npages > 0) {
301 		if (!chunk) {
302 			chunk = kmalloc(sizeof(*chunk),
303 				gfp_mask & ~(__GFP_HIGHMEM | __GFP_NOWARN));
304 			if (!chunk)
305 				goto fail;
306 
307 			sg_init_table(chunk->mem, HNS_ROCE_HEM_CHUNK_LEN);
308 			chunk->npages = 0;
309 			chunk->nsg = 0;
310 			memset(chunk->buf, 0, sizeof(chunk->buf));
311 			list_add_tail(&chunk->list, &hem->chunk_list);
312 		}
313 
314 		while (1 << order > npages)
315 			--order;
316 
317 		/*
318 		 * Alloc memory one time. If failed, don't alloc small block
319 		 * memory, directly return fail.
320 		 */
321 		mem = &chunk->mem[chunk->npages];
322 		buf = dma_alloc_coherent(hr_dev->dev, PAGE_SIZE << order,
323 				&sg_dma_address(mem), gfp_mask);
324 		if (!buf)
325 			goto fail;
326 
327 		chunk->buf[chunk->npages] = buf;
328 		sg_dma_len(mem) = PAGE_SIZE << order;
329 
330 		++chunk->npages;
331 		++chunk->nsg;
332 		npages -= 1 << order;
333 	}
334 
335 	return hem;
336 
337 fail:
338 	hns_roce_free_hem(hr_dev, hem);
339 	return NULL;
340 }
341 
342 void hns_roce_free_hem(struct hns_roce_dev *hr_dev, struct hns_roce_hem *hem)
343 {
344 	struct hns_roce_hem_chunk *chunk, *tmp;
345 	int i;
346 
347 	if (!hem)
348 		return;
349 
350 	list_for_each_entry_safe(chunk, tmp, &hem->chunk_list, list) {
351 		for (i = 0; i < chunk->npages; ++i)
352 			dma_free_coherent(hr_dev->dev,
353 				   sg_dma_len(&chunk->mem[i]),
354 				   chunk->buf[i],
355 				   sg_dma_address(&chunk->mem[i]));
356 		kfree(chunk);
357 	}
358 
359 	kfree(hem);
360 }
361 
362 static int hns_roce_set_hem(struct hns_roce_dev *hr_dev,
363 			    struct hns_roce_hem_table *table, unsigned long obj)
364 {
365 	spinlock_t *lock = &hr_dev->bt_cmd_lock;
366 	struct device *dev = hr_dev->dev;
367 	long end;
368 	unsigned long flags;
369 	struct hns_roce_hem_iter iter;
370 	void __iomem *bt_cmd;
371 	__le32 bt_cmd_val[2];
372 	__le32 bt_cmd_h = 0;
373 	__le32 bt_cmd_l = 0;
374 	u64 bt_ba = 0;
375 	int ret = 0;
376 
377 	/* Find the HEM(Hardware Entry Memory) entry */
378 	unsigned long i = (obj & (table->num_obj - 1)) /
379 			  (table->table_chunk_size / table->obj_size);
380 
381 	switch (table->type) {
382 	case HEM_TYPE_QPC:
383 	case HEM_TYPE_MTPT:
384 	case HEM_TYPE_CQC:
385 	case HEM_TYPE_SRQC:
386 		roce_set_field(bt_cmd_h, ROCEE_BT_CMD_H_ROCEE_BT_CMD_MDF_M,
387 			ROCEE_BT_CMD_H_ROCEE_BT_CMD_MDF_S, table->type);
388 		break;
389 	default:
390 		return ret;
391 	}
392 
393 	roce_set_field(bt_cmd_h, ROCEE_BT_CMD_H_ROCEE_BT_CMD_IN_MDF_M,
394 		       ROCEE_BT_CMD_H_ROCEE_BT_CMD_IN_MDF_S, obj);
395 	roce_set_bit(bt_cmd_h, ROCEE_BT_CMD_H_ROCEE_BT_CMD_S, 0);
396 	roce_set_bit(bt_cmd_h, ROCEE_BT_CMD_H_ROCEE_BT_CMD_HW_SYNS_S, 1);
397 
398 	/* Currently iter only a chunk */
399 	for (hns_roce_hem_first(table->hem[i], &iter);
400 	     !hns_roce_hem_last(&iter); hns_roce_hem_next(&iter)) {
401 		bt_ba = hns_roce_hem_addr(&iter) >> DMA_ADDR_T_SHIFT;
402 
403 		spin_lock_irqsave(lock, flags);
404 
405 		bt_cmd = hr_dev->reg_base + ROCEE_BT_CMD_H_REG;
406 
407 		end = HW_SYNC_TIMEOUT_MSECS;
408 		while (end > 0) {
409 			if (!(readl(bt_cmd) >> BT_CMD_SYNC_SHIFT))
410 				break;
411 
412 			mdelay(HW_SYNC_SLEEP_TIME_INTERVAL);
413 			end -= HW_SYNC_SLEEP_TIME_INTERVAL;
414 		}
415 
416 		if (end <= 0) {
417 			dev_err(dev, "Write bt_cmd err,hw_sync is not zero.\n");
418 			spin_unlock_irqrestore(lock, flags);
419 			return -EBUSY;
420 		}
421 
422 		bt_cmd_l = cpu_to_le32(bt_ba);
423 		roce_set_field(bt_cmd_h, ROCEE_BT_CMD_H_ROCEE_BT_CMD_BA_H_M,
424 			       ROCEE_BT_CMD_H_ROCEE_BT_CMD_BA_H_S,
425 			       bt_ba >> BT_BA_SHIFT);
426 
427 		bt_cmd_val[0] = bt_cmd_l;
428 		bt_cmd_val[1] = bt_cmd_h;
429 		hns_roce_write64_k(bt_cmd_val,
430 				   hr_dev->reg_base + ROCEE_BT_CMD_L_REG);
431 		spin_unlock_irqrestore(lock, flags);
432 	}
433 
434 	return ret;
435 }
436 
437 static int hns_roce_table_mhop_get(struct hns_roce_dev *hr_dev,
438 				   struct hns_roce_hem_table *table,
439 				   unsigned long obj)
440 {
441 	struct device *dev = hr_dev->dev;
442 	struct hns_roce_hem_mhop mhop;
443 	struct hns_roce_hem_iter iter;
444 	u32 buf_chunk_size;
445 	u32 bt_chunk_size;
446 	u32 chunk_ba_num;
447 	u32 hop_num;
448 	u32 size;
449 	u32 bt_num;
450 	u64 hem_idx;
451 	u64 bt_l1_idx = 0;
452 	u64 bt_l0_idx = 0;
453 	u64 bt_ba;
454 	unsigned long mhop_obj = obj;
455 	int bt_l1_allocated = 0;
456 	int bt_l0_allocated = 0;
457 	int step_idx;
458 	int ret;
459 
460 	ret = hns_roce_calc_hem_mhop(hr_dev, table, &mhop_obj, &mhop);
461 	if (ret)
462 		return ret;
463 
464 	buf_chunk_size = mhop.buf_chunk_size;
465 	bt_chunk_size = mhop.bt_chunk_size;
466 	hop_num = mhop.hop_num;
467 	chunk_ba_num = bt_chunk_size / BA_BYTE_LEN;
468 
469 	bt_num = hns_roce_get_bt_num(table->type, hop_num);
470 	switch (bt_num) {
471 	case 3:
472 		hem_idx = mhop.l0_idx * chunk_ba_num * chunk_ba_num +
473 			  mhop.l1_idx * chunk_ba_num + mhop.l2_idx;
474 		bt_l1_idx = mhop.l0_idx * chunk_ba_num + mhop.l1_idx;
475 		bt_l0_idx = mhop.l0_idx;
476 		break;
477 	case 2:
478 		hem_idx = mhop.l0_idx * chunk_ba_num + mhop.l1_idx;
479 		bt_l0_idx = mhop.l0_idx;
480 		break;
481 	case 1:
482 		hem_idx = mhop.l0_idx;
483 		break;
484 	default:
485 		dev_err(dev, "Table %d not support hop_num = %d!\n",
486 			     table->type, hop_num);
487 		return -EINVAL;
488 	}
489 
490 	if (unlikely(hem_idx >= table->num_hem)) {
491 		dev_err(dev, "Table %d exceed hem limt idx = %llu,max = %lu!\n",
492 			     table->type, hem_idx, table->num_hem);
493 		return -EINVAL;
494 	}
495 
496 	mutex_lock(&table->mutex);
497 
498 	if (table->hem[hem_idx]) {
499 		++table->hem[hem_idx]->refcount;
500 		goto out;
501 	}
502 
503 	/* alloc L1 BA's chunk */
504 	if ((check_whether_bt_num_3(table->type, hop_num) ||
505 		check_whether_bt_num_2(table->type, hop_num)) &&
506 		!table->bt_l0[bt_l0_idx]) {
507 		table->bt_l0[bt_l0_idx] = dma_alloc_coherent(dev, bt_chunk_size,
508 					    &(table->bt_l0_dma_addr[bt_l0_idx]),
509 					    GFP_KERNEL);
510 		if (!table->bt_l0[bt_l0_idx]) {
511 			ret = -ENOMEM;
512 			goto out;
513 		}
514 		bt_l0_allocated = 1;
515 
516 		/* set base address to hardware */
517 		if (table->type < HEM_TYPE_MTT) {
518 			step_idx = 0;
519 			if (hr_dev->hw->set_hem(hr_dev, table, obj, step_idx)) {
520 				ret = -ENODEV;
521 				dev_err(dev, "set HEM base address to HW failed!\n");
522 				goto err_dma_alloc_l1;
523 			}
524 		}
525 	}
526 
527 	/* alloc L2 BA's chunk */
528 	if (check_whether_bt_num_3(table->type, hop_num) &&
529 	    !table->bt_l1[bt_l1_idx])  {
530 		table->bt_l1[bt_l1_idx] = dma_alloc_coherent(dev, bt_chunk_size,
531 					    &(table->bt_l1_dma_addr[bt_l1_idx]),
532 					    GFP_KERNEL);
533 		if (!table->bt_l1[bt_l1_idx]) {
534 			ret = -ENOMEM;
535 			goto err_dma_alloc_l1;
536 		}
537 		bt_l1_allocated = 1;
538 		*(table->bt_l0[bt_l0_idx] + mhop.l1_idx) =
539 					       table->bt_l1_dma_addr[bt_l1_idx];
540 
541 		/* set base address to hardware */
542 		step_idx = 1;
543 		if (hr_dev->hw->set_hem(hr_dev, table, obj, step_idx)) {
544 			ret = -ENODEV;
545 			dev_err(dev, "set HEM base address to HW failed!\n");
546 			goto err_alloc_hem_buf;
547 		}
548 	}
549 
550 	/*
551 	 * alloc buffer space chunk for QPC/MTPT/CQC/SRQC/SCCC.
552 	 * alloc bt space chunk for MTT/CQE.
553 	 */
554 	size = table->type < HEM_TYPE_MTT ? buf_chunk_size : bt_chunk_size;
555 	table->hem[hem_idx] = hns_roce_alloc_hem(hr_dev,
556 						size >> PAGE_SHIFT,
557 						size,
558 						(table->lowmem ? GFP_KERNEL :
559 						GFP_HIGHUSER) | __GFP_NOWARN);
560 	if (!table->hem[hem_idx]) {
561 		ret = -ENOMEM;
562 		goto err_alloc_hem_buf;
563 	}
564 
565 	hns_roce_hem_first(table->hem[hem_idx], &iter);
566 	bt_ba = hns_roce_hem_addr(&iter);
567 
568 	if (table->type < HEM_TYPE_MTT) {
569 		if (hop_num == 2) {
570 			*(table->bt_l1[bt_l1_idx] + mhop.l2_idx) = bt_ba;
571 			step_idx = 2;
572 		} else if (hop_num == 1) {
573 			*(table->bt_l0[bt_l0_idx] + mhop.l1_idx) = bt_ba;
574 			step_idx = 1;
575 		} else if (hop_num == HNS_ROCE_HOP_NUM_0) {
576 			step_idx = 0;
577 		} else {
578 			ret = -EINVAL;
579 			goto err_dma_alloc_l1;
580 		}
581 
582 		/* set HEM base address to hardware */
583 		if (hr_dev->hw->set_hem(hr_dev, table, obj, step_idx)) {
584 			ret = -ENODEV;
585 			dev_err(dev, "set HEM base address to HW failed!\n");
586 			goto err_alloc_hem_buf;
587 		}
588 	} else if (hop_num == 2) {
589 		*(table->bt_l0[bt_l0_idx] + mhop.l1_idx) = bt_ba;
590 	}
591 
592 	++table->hem[hem_idx]->refcount;
593 	goto out;
594 
595 err_alloc_hem_buf:
596 	if (bt_l1_allocated) {
597 		dma_free_coherent(dev, bt_chunk_size, table->bt_l1[bt_l1_idx],
598 				  table->bt_l1_dma_addr[bt_l1_idx]);
599 		table->bt_l1[bt_l1_idx] = NULL;
600 	}
601 
602 err_dma_alloc_l1:
603 	if (bt_l0_allocated) {
604 		dma_free_coherent(dev, bt_chunk_size, table->bt_l0[bt_l0_idx],
605 				  table->bt_l0_dma_addr[bt_l0_idx]);
606 		table->bt_l0[bt_l0_idx] = NULL;
607 	}
608 
609 out:
610 	mutex_unlock(&table->mutex);
611 	return ret;
612 }
613 
614 int hns_roce_table_get(struct hns_roce_dev *hr_dev,
615 		       struct hns_roce_hem_table *table, unsigned long obj)
616 {
617 	struct device *dev = hr_dev->dev;
618 	int ret = 0;
619 	unsigned long i;
620 
621 	if (hns_roce_check_whether_mhop(hr_dev, table->type))
622 		return hns_roce_table_mhop_get(hr_dev, table, obj);
623 
624 	i = (obj & (table->num_obj - 1)) / (table->table_chunk_size /
625 	     table->obj_size);
626 
627 	mutex_lock(&table->mutex);
628 
629 	if (table->hem[i]) {
630 		++table->hem[i]->refcount;
631 		goto out;
632 	}
633 
634 	table->hem[i] = hns_roce_alloc_hem(hr_dev,
635 				       table->table_chunk_size >> PAGE_SHIFT,
636 				       table->table_chunk_size,
637 				       (table->lowmem ? GFP_KERNEL :
638 					GFP_HIGHUSER) | __GFP_NOWARN);
639 	if (!table->hem[i]) {
640 		ret = -ENOMEM;
641 		goto out;
642 	}
643 
644 	/* Set HEM base address(128K/page, pa) to Hardware */
645 	if (hns_roce_set_hem(hr_dev, table, obj)) {
646 		hns_roce_free_hem(hr_dev, table->hem[i]);
647 		table->hem[i] = NULL;
648 		ret = -ENODEV;
649 		dev_err(dev, "set HEM base address to HW failed.\n");
650 		goto out;
651 	}
652 
653 	++table->hem[i]->refcount;
654 out:
655 	mutex_unlock(&table->mutex);
656 	return ret;
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 device *dev = hr_dev->dev;
665 	struct hns_roce_hem_mhop mhop;
666 	unsigned long mhop_obj = obj;
667 	u32 bt_chunk_size;
668 	u32 chunk_ba_num;
669 	u32 hop_num;
670 	u32 start_idx;
671 	u32 bt_num;
672 	u64 hem_idx;
673 	u64 bt_l1_idx = 0;
674 	int ret;
675 
676 	ret = hns_roce_calc_hem_mhop(hr_dev, table, &mhop_obj, &mhop);
677 	if (ret)
678 		return;
679 
680 	bt_chunk_size = mhop.bt_chunk_size;
681 	hop_num = mhop.hop_num;
682 	chunk_ba_num = bt_chunk_size / BA_BYTE_LEN;
683 
684 	bt_num = hns_roce_get_bt_num(table->type, hop_num);
685 	switch (bt_num) {
686 	case 3:
687 		hem_idx = mhop.l0_idx * chunk_ba_num * chunk_ba_num +
688 			  mhop.l1_idx * chunk_ba_num + mhop.l2_idx;
689 		bt_l1_idx = mhop.l0_idx * chunk_ba_num + mhop.l1_idx;
690 		break;
691 	case 2:
692 		hem_idx = mhop.l0_idx * chunk_ba_num + mhop.l1_idx;
693 		break;
694 	case 1:
695 		hem_idx = mhop.l0_idx;
696 		break;
697 	default:
698 		dev_err(dev, "Table %d not support hop_num = %d!\n",
699 			     table->type, hop_num);
700 		return;
701 	}
702 
703 	mutex_lock(&table->mutex);
704 
705 	if (check_refcount && (--table->hem[hem_idx]->refcount > 0)) {
706 		mutex_unlock(&table->mutex);
707 		return;
708 	}
709 
710 	if (table->type < HEM_TYPE_MTT && hop_num == 1) {
711 		if (hr_dev->hw->clear_hem(hr_dev, table, obj, 1))
712 			dev_warn(dev, "Clear HEM base address failed.\n");
713 	} else if (table->type < HEM_TYPE_MTT && hop_num == 2) {
714 		if (hr_dev->hw->clear_hem(hr_dev, table, obj, 2))
715 			dev_warn(dev, "Clear HEM base address failed.\n");
716 	} else if (table->type < HEM_TYPE_MTT &&
717 		   hop_num == HNS_ROCE_HOP_NUM_0) {
718 		if (hr_dev->hw->clear_hem(hr_dev, table, obj, 0))
719 			dev_warn(dev, "Clear HEM base address failed.\n");
720 	}
721 
722 	/*
723 	 * free buffer space chunk for QPC/MTPT/CQC/SRQC/SCCC.
724 	 * free bt space chunk for MTT/CQE.
725 	 */
726 	hns_roce_free_hem(hr_dev, table->hem[hem_idx]);
727 	table->hem[hem_idx] = NULL;
728 
729 	if (check_whether_bt_num_2(table->type, hop_num)) {
730 		start_idx = mhop.l0_idx * chunk_ba_num;
731 		if (hns_roce_check_hem_null(table->hem, start_idx,
732 					    chunk_ba_num, table->num_hem)) {
733 			if (table->type < HEM_TYPE_MTT &&
734 			    hr_dev->hw->clear_hem(hr_dev, table, obj, 0))
735 				dev_warn(dev, "Clear HEM base address failed.\n");
736 
737 			dma_free_coherent(dev, bt_chunk_size,
738 					  table->bt_l0[mhop.l0_idx],
739 					  table->bt_l0_dma_addr[mhop.l0_idx]);
740 			table->bt_l0[mhop.l0_idx] = NULL;
741 		}
742 	} else if (check_whether_bt_num_3(table->type, hop_num)) {
743 		start_idx = mhop.l0_idx * chunk_ba_num * chunk_ba_num +
744 			    mhop.l1_idx * chunk_ba_num;
745 		if (hns_roce_check_hem_null(table->hem, start_idx,
746 					    chunk_ba_num, table->num_hem)) {
747 			if (hr_dev->hw->clear_hem(hr_dev, table, obj, 1))
748 				dev_warn(dev, "Clear HEM base address failed.\n");
749 
750 			dma_free_coherent(dev, bt_chunk_size,
751 					  table->bt_l1[bt_l1_idx],
752 					  table->bt_l1_dma_addr[bt_l1_idx]);
753 			table->bt_l1[bt_l1_idx] = NULL;
754 
755 			start_idx = mhop.l0_idx * chunk_ba_num;
756 			if (hns_roce_check_bt_null(table->bt_l1, start_idx,
757 						   chunk_ba_num)) {
758 				if (hr_dev->hw->clear_hem(hr_dev, table, obj,
759 							  0))
760 					dev_warn(dev, "Clear HEM base address failed.\n");
761 
762 				dma_free_coherent(dev, bt_chunk_size,
763 					    table->bt_l0[mhop.l0_idx],
764 					    table->bt_l0_dma_addr[mhop.l0_idx]);
765 				table->bt_l0[mhop.l0_idx] = NULL;
766 			}
767 		}
768 	}
769 
770 	mutex_unlock(&table->mutex);
771 }
772 
773 void hns_roce_table_put(struct hns_roce_dev *hr_dev,
774 			struct hns_roce_hem_table *table, unsigned long obj)
775 {
776 	struct device *dev = hr_dev->dev;
777 	unsigned long i;
778 
779 	if (hns_roce_check_whether_mhop(hr_dev, table->type)) {
780 		hns_roce_table_mhop_put(hr_dev, table, obj, 1);
781 		return;
782 	}
783 
784 	i = (obj & (table->num_obj - 1)) /
785 	    (table->table_chunk_size / table->obj_size);
786 
787 	mutex_lock(&table->mutex);
788 
789 	if (--table->hem[i]->refcount == 0) {
790 		/* Clear HEM base address */
791 		if (hr_dev->hw->clear_hem(hr_dev, table, obj, 0))
792 			dev_warn(dev, "Clear HEM base address failed.\n");
793 
794 		hns_roce_free_hem(hr_dev, table->hem[i]);
795 		table->hem[i] = NULL;
796 	}
797 
798 	mutex_unlock(&table->mutex);
799 }
800 
801 void *hns_roce_table_find(struct hns_roce_dev *hr_dev,
802 			  struct hns_roce_hem_table *table,
803 			  unsigned long obj, dma_addr_t *dma_handle)
804 {
805 	struct hns_roce_hem_chunk *chunk;
806 	struct hns_roce_hem_mhop mhop;
807 	struct hns_roce_hem *hem;
808 	void *addr = NULL;
809 	unsigned long mhop_obj = obj;
810 	unsigned long obj_per_chunk;
811 	unsigned long idx_offset;
812 	int offset, dma_offset;
813 	int length;
814 	int i, j;
815 	u32 hem_idx = 0;
816 
817 	if (!table->lowmem)
818 		return NULL;
819 
820 	mutex_lock(&table->mutex);
821 
822 	if (!hns_roce_check_whether_mhop(hr_dev, table->type)) {
823 		obj_per_chunk = table->table_chunk_size / table->obj_size;
824 		hem = table->hem[(obj & (table->num_obj - 1)) / obj_per_chunk];
825 		idx_offset = (obj & (table->num_obj - 1)) % obj_per_chunk;
826 		dma_offset = offset = idx_offset * table->obj_size;
827 	} else {
828 		u32 seg_size = 64; /* 8 bytes per BA and 8 BA per segment */
829 
830 		if (hns_roce_calc_hem_mhop(hr_dev, table, &mhop_obj, &mhop))
831 			goto out;
832 		/* mtt mhop */
833 		i = mhop.l0_idx;
834 		j = mhop.l1_idx;
835 		if (mhop.hop_num == 2)
836 			hem_idx = i * (mhop.bt_chunk_size / BA_BYTE_LEN) + j;
837 		else if (mhop.hop_num == 1 ||
838 			 mhop.hop_num == HNS_ROCE_HOP_NUM_0)
839 			hem_idx = i;
840 
841 		hem = table->hem[hem_idx];
842 		dma_offset = offset = (obj & (table->num_obj - 1)) * seg_size %
843 				       mhop.bt_chunk_size;
844 		if (mhop.hop_num == 2)
845 			dma_offset = offset = 0;
846 	}
847 
848 	if (!hem)
849 		goto out;
850 
851 	list_for_each_entry(chunk, &hem->chunk_list, list) {
852 		for (i = 0; i < chunk->npages; ++i) {
853 			length = sg_dma_len(&chunk->mem[i]);
854 			if (dma_handle && dma_offset >= 0) {
855 				if (length > (u32)dma_offset)
856 					*dma_handle = sg_dma_address(
857 						&chunk->mem[i]) + dma_offset;
858 				dma_offset -= length;
859 			}
860 
861 			if (length > (u32)offset) {
862 				addr = chunk->buf[i] + offset;
863 				goto out;
864 			}
865 			offset -= length;
866 		}
867 	}
868 
869 out:
870 	mutex_unlock(&table->mutex);
871 	return addr;
872 }
873 
874 int hns_roce_table_get_range(struct hns_roce_dev *hr_dev,
875 			     struct hns_roce_hem_table *table,
876 			     unsigned long start, unsigned long end)
877 {
878 	struct hns_roce_hem_mhop mhop;
879 	unsigned long inc = table->table_chunk_size / table->obj_size;
880 	unsigned long i = 0;
881 	int ret;
882 
883 	if (hns_roce_check_whether_mhop(hr_dev, table->type)) {
884 		ret = hns_roce_calc_hem_mhop(hr_dev, table, NULL, &mhop);
885 		if (ret)
886 			goto fail;
887 		inc = mhop.bt_chunk_size / table->obj_size;
888 	}
889 
890 	/* Allocate MTT entry memory according to chunk(128K) */
891 	for (i = start; i <= end; i += inc) {
892 		ret = hns_roce_table_get(hr_dev, table, i);
893 		if (ret)
894 			goto fail;
895 	}
896 
897 	return 0;
898 
899 fail:
900 	while (i > start) {
901 		i -= inc;
902 		hns_roce_table_put(hr_dev, table, i);
903 	}
904 	return ret;
905 }
906 
907 void hns_roce_table_put_range(struct hns_roce_dev *hr_dev,
908 			      struct hns_roce_hem_table *table,
909 			      unsigned long start, unsigned long end)
910 {
911 	struct hns_roce_hem_mhop mhop;
912 	unsigned long inc = table->table_chunk_size / table->obj_size;
913 	unsigned long i;
914 
915 	if (hns_roce_check_whether_mhop(hr_dev, table->type)) {
916 		if (hns_roce_calc_hem_mhop(hr_dev, table, NULL, &mhop))
917 			return;
918 		inc = mhop.bt_chunk_size / table->obj_size;
919 	}
920 
921 	for (i = start; i <= end; i += inc)
922 		hns_roce_table_put(hr_dev, table, i);
923 }
924 
925 int hns_roce_init_hem_table(struct hns_roce_dev *hr_dev,
926 			    struct hns_roce_hem_table *table, u32 type,
927 			    unsigned long obj_size, unsigned long nobj,
928 			    int use_lowmem)
929 {
930 	unsigned long obj_per_chunk;
931 	unsigned long num_hem;
932 
933 	if (!hns_roce_check_whether_mhop(hr_dev, type)) {
934 		table->table_chunk_size = hr_dev->caps.chunk_sz;
935 		obj_per_chunk = table->table_chunk_size / obj_size;
936 		num_hem = (nobj + obj_per_chunk - 1) / obj_per_chunk;
937 
938 		table->hem = kcalloc(num_hem, sizeof(*table->hem), GFP_KERNEL);
939 		if (!table->hem)
940 			return -ENOMEM;
941 	} else {
942 		struct hns_roce_hem_mhop mhop = {};
943 		unsigned long buf_chunk_size;
944 		unsigned long bt_chunk_size;
945 		unsigned long bt_chunk_num;
946 		unsigned long num_bt_l0 = 0;
947 		u32 hop_num;
948 
949 		if (get_hem_table_config(hr_dev, &mhop, type))
950 			return -EINVAL;
951 
952 		buf_chunk_size = mhop.buf_chunk_size;
953 		bt_chunk_size = mhop.bt_chunk_size;
954 		num_bt_l0 = mhop.ba_l0_num;
955 		hop_num = mhop.hop_num;
956 
957 		obj_per_chunk = buf_chunk_size / obj_size;
958 		num_hem = (nobj + obj_per_chunk - 1) / obj_per_chunk;
959 		bt_chunk_num = bt_chunk_size / BA_BYTE_LEN;
960 		if (type >= HEM_TYPE_MTT)
961 			num_bt_l0 = bt_chunk_num;
962 
963 		table->hem = kcalloc(num_hem, sizeof(*table->hem),
964 					 GFP_KERNEL);
965 		if (!table->hem)
966 			goto err_kcalloc_hem_buf;
967 
968 		if (check_whether_bt_num_3(type, hop_num)) {
969 			unsigned long num_bt_l1;
970 
971 			num_bt_l1 = (num_hem + bt_chunk_num - 1) /
972 					     bt_chunk_num;
973 			table->bt_l1 = kcalloc(num_bt_l1,
974 					       sizeof(*table->bt_l1),
975 					       GFP_KERNEL);
976 			if (!table->bt_l1)
977 				goto err_kcalloc_bt_l1;
978 
979 			table->bt_l1_dma_addr = kcalloc(num_bt_l1,
980 						 sizeof(*table->bt_l1_dma_addr),
981 						 GFP_KERNEL);
982 
983 			if (!table->bt_l1_dma_addr)
984 				goto err_kcalloc_l1_dma;
985 		}
986 
987 		if (check_whether_bt_num_2(type, hop_num) ||
988 			check_whether_bt_num_3(type, hop_num)) {
989 			table->bt_l0 = kcalloc(num_bt_l0, sizeof(*table->bt_l0),
990 					       GFP_KERNEL);
991 			if (!table->bt_l0)
992 				goto err_kcalloc_bt_l0;
993 
994 			table->bt_l0_dma_addr = kcalloc(num_bt_l0,
995 						 sizeof(*table->bt_l0_dma_addr),
996 						 GFP_KERNEL);
997 			if (!table->bt_l0_dma_addr)
998 				goto err_kcalloc_l0_dma;
999 		}
1000 	}
1001 
1002 	table->type = type;
1003 	table->num_hem = num_hem;
1004 	table->num_obj = nobj;
1005 	table->obj_size = obj_size;
1006 	table->lowmem = use_lowmem;
1007 	mutex_init(&table->mutex);
1008 
1009 	return 0;
1010 
1011 err_kcalloc_l0_dma:
1012 	kfree(table->bt_l0);
1013 	table->bt_l0 = NULL;
1014 
1015 err_kcalloc_bt_l0:
1016 	kfree(table->bt_l1_dma_addr);
1017 	table->bt_l1_dma_addr = NULL;
1018 
1019 err_kcalloc_l1_dma:
1020 	kfree(table->bt_l1);
1021 	table->bt_l1 = NULL;
1022 
1023 err_kcalloc_bt_l1:
1024 	kfree(table->hem);
1025 	table->hem = NULL;
1026 
1027 err_kcalloc_hem_buf:
1028 	return -ENOMEM;
1029 }
1030 
1031 static void hns_roce_cleanup_mhop_hem_table(struct hns_roce_dev *hr_dev,
1032 					    struct hns_roce_hem_table *table)
1033 {
1034 	struct hns_roce_hem_mhop mhop;
1035 	u32 buf_chunk_size;
1036 	int i;
1037 	u64 obj;
1038 
1039 	if (hns_roce_calc_hem_mhop(hr_dev, table, NULL, &mhop))
1040 		return;
1041 	buf_chunk_size = table->type < HEM_TYPE_MTT ? mhop.buf_chunk_size :
1042 					mhop.bt_chunk_size;
1043 
1044 	for (i = 0; i < table->num_hem; ++i) {
1045 		obj = i * buf_chunk_size / table->obj_size;
1046 		if (table->hem[i])
1047 			hns_roce_table_mhop_put(hr_dev, table, obj, 0);
1048 	}
1049 
1050 	kfree(table->hem);
1051 	table->hem = NULL;
1052 	kfree(table->bt_l1);
1053 	table->bt_l1 = NULL;
1054 	kfree(table->bt_l1_dma_addr);
1055 	table->bt_l1_dma_addr = NULL;
1056 	kfree(table->bt_l0);
1057 	table->bt_l0 = NULL;
1058 	kfree(table->bt_l0_dma_addr);
1059 	table->bt_l0_dma_addr = NULL;
1060 }
1061 
1062 void hns_roce_cleanup_hem_table(struct hns_roce_dev *hr_dev,
1063 				struct hns_roce_hem_table *table)
1064 {
1065 	struct device *dev = hr_dev->dev;
1066 	unsigned long i;
1067 
1068 	if (hns_roce_check_whether_mhop(hr_dev, table->type)) {
1069 		hns_roce_cleanup_mhop_hem_table(hr_dev, table);
1070 		return;
1071 	}
1072 
1073 	for (i = 0; i < table->num_hem; ++i)
1074 		if (table->hem[i]) {
1075 			if (hr_dev->hw->clear_hem(hr_dev, table,
1076 			    i * table->table_chunk_size / table->obj_size, 0))
1077 				dev_err(dev, "Clear HEM base address failed.\n");
1078 
1079 			hns_roce_free_hem(hr_dev, table->hem[i]);
1080 		}
1081 
1082 	kfree(table->hem);
1083 }
1084 
1085 void hns_roce_cleanup_hem(struct hns_roce_dev *hr_dev)
1086 {
1087 	if ((hr_dev->caps.num_idx_segs))
1088 		hns_roce_cleanup_hem_table(hr_dev,
1089 					   &hr_dev->mr_table.mtt_idx_table);
1090 	if (hr_dev->caps.num_srqwqe_segs)
1091 		hns_roce_cleanup_hem_table(hr_dev,
1092 					   &hr_dev->mr_table.mtt_srqwqe_table);
1093 	if (hr_dev->caps.srqc_entry_sz)
1094 		hns_roce_cleanup_hem_table(hr_dev,
1095 					   &hr_dev->srq_table.table);
1096 	hns_roce_cleanup_hem_table(hr_dev, &hr_dev->cq_table.table);
1097 	if (hr_dev->caps.qpc_timer_entry_sz)
1098 		hns_roce_cleanup_hem_table(hr_dev,
1099 					   &hr_dev->qpc_timer_table);
1100 	if (hr_dev->caps.cqc_timer_entry_sz)
1101 		hns_roce_cleanup_hem_table(hr_dev,
1102 					   &hr_dev->cqc_timer_table);
1103 	if (hr_dev->caps.sccc_entry_sz)
1104 		hns_roce_cleanup_hem_table(hr_dev,
1105 					   &hr_dev->qp_table.sccc_table);
1106 	if (hr_dev->caps.trrl_entry_sz)
1107 		hns_roce_cleanup_hem_table(hr_dev,
1108 					   &hr_dev->qp_table.trrl_table);
1109 	hns_roce_cleanup_hem_table(hr_dev, &hr_dev->qp_table.irrl_table);
1110 	hns_roce_cleanup_hem_table(hr_dev, &hr_dev->qp_table.qp_table);
1111 	hns_roce_cleanup_hem_table(hr_dev, &hr_dev->mr_table.mtpt_table);
1112 	if (hns_roce_check_whether_mhop(hr_dev, HEM_TYPE_CQE))
1113 		hns_roce_cleanup_hem_table(hr_dev,
1114 					   &hr_dev->mr_table.mtt_cqe_table);
1115 	hns_roce_cleanup_hem_table(hr_dev, &hr_dev->mr_table.mtt_table);
1116 }
1117 
1118 struct roce_hem_item {
1119 	struct list_head list; /* link all hems in the same bt level */
1120 	struct list_head sibling; /* link all hems in last hop for mtt */
1121 	void *addr;
1122 	dma_addr_t dma_addr;
1123 	size_t count; /* max ba numbers */
1124 	int start; /* start buf offset in this hem */
1125 	int end; /* end buf offset in this hem */
1126 };
1127 
1128 static struct roce_hem_item *hem_list_alloc_item(struct hns_roce_dev *hr_dev,
1129 						   int start, int end,
1130 						   int count, bool exist_bt,
1131 						   int bt_level)
1132 {
1133 	struct roce_hem_item *hem;
1134 
1135 	hem = kzalloc(sizeof(*hem), GFP_KERNEL);
1136 	if (!hem)
1137 		return NULL;
1138 
1139 	if (exist_bt) {
1140 		hem->addr = dma_alloc_coherent(hr_dev->dev,
1141 						   count * BA_BYTE_LEN,
1142 						   &hem->dma_addr, GFP_KERNEL);
1143 		if (!hem->addr) {
1144 			kfree(hem);
1145 			return NULL;
1146 		}
1147 	}
1148 
1149 	hem->count = count;
1150 	hem->start = start;
1151 	hem->end = end;
1152 	INIT_LIST_HEAD(&hem->list);
1153 	INIT_LIST_HEAD(&hem->sibling);
1154 
1155 	return hem;
1156 }
1157 
1158 static void hem_list_free_item(struct hns_roce_dev *hr_dev,
1159 			       struct roce_hem_item *hem, bool exist_bt)
1160 {
1161 	if (exist_bt)
1162 		dma_free_coherent(hr_dev->dev, hem->count * BA_BYTE_LEN,
1163 				  hem->addr, hem->dma_addr);
1164 	kfree(hem);
1165 }
1166 
1167 static void hem_list_free_all(struct hns_roce_dev *hr_dev,
1168 			      struct list_head *head, bool exist_bt)
1169 {
1170 	struct roce_hem_item *hem, *temp_hem;
1171 
1172 	list_for_each_entry_safe(hem, temp_hem, head, list) {
1173 		list_del(&hem->list);
1174 		hem_list_free_item(hr_dev, hem, exist_bt);
1175 	}
1176 }
1177 
1178 static void hem_list_link_bt(struct hns_roce_dev *hr_dev, void *base_addr,
1179 			     u64 table_addr)
1180 {
1181 	*(u64 *)(base_addr) = table_addr;
1182 }
1183 
1184 /* assign L0 table address to hem from root bt */
1185 static void hem_list_assign_bt(struct hns_roce_dev *hr_dev,
1186 			       struct roce_hem_item *hem, void *cpu_addr,
1187 			       u64 phy_addr)
1188 {
1189 	hem->addr = cpu_addr;
1190 	hem->dma_addr = (dma_addr_t)phy_addr;
1191 }
1192 
1193 static inline bool hem_list_page_is_in_range(struct roce_hem_item *hem,
1194 					     int offset)
1195 {
1196 	return (hem->start <= offset && offset <= hem->end);
1197 }
1198 
1199 static struct roce_hem_item *hem_list_search_item(struct list_head *ba_list,
1200 						    int page_offset)
1201 {
1202 	struct roce_hem_item *hem, *temp_hem;
1203 	struct roce_hem_item *found = NULL;
1204 
1205 	list_for_each_entry_safe(hem, temp_hem, ba_list, list) {
1206 		if (hem_list_page_is_in_range(hem, page_offset)) {
1207 			found = hem;
1208 			break;
1209 		}
1210 	}
1211 
1212 	return found;
1213 }
1214 
1215 static bool hem_list_is_bottom_bt(int hopnum, int bt_level)
1216 {
1217 	/*
1218 	 * hopnum    base address table levels
1219 	 * 0		L0(buf)
1220 	 * 1		L0 -> buf
1221 	 * 2		L0 -> L1 -> buf
1222 	 * 3		L0 -> L1 -> L2 -> buf
1223 	 */
1224 	return bt_level >= (hopnum ? hopnum - 1 : hopnum);
1225 }
1226 
1227 /**
1228  * calc base address entries num
1229  * @hopnum: num of mutihop addressing
1230  * @bt_level: base address table level
1231  * @unit: ba entries per bt page
1232  */
1233 static u32 hem_list_calc_ba_range(int hopnum, int bt_level, int unit)
1234 {
1235 	u32 step;
1236 	int max;
1237 	int i;
1238 
1239 	if (hopnum <= bt_level)
1240 		return 0;
1241 	/*
1242 	 * hopnum  bt_level   range
1243 	 * 1	      0       unit
1244 	 * ------------
1245 	 * 2	      0       unit * unit
1246 	 * 2	      1       unit
1247 	 * ------------
1248 	 * 3	      0       unit * unit * unit
1249 	 * 3	      1       unit * unit
1250 	 * 3	      2       unit
1251 	 */
1252 	step = 1;
1253 	max = hopnum - bt_level;
1254 	for (i = 0; i < max; i++)
1255 		step = step * unit;
1256 
1257 	return step;
1258 }
1259 
1260 /**
1261  * calc the root ba entries which could cover all regions
1262  * @regions: buf region array
1263  * @region_cnt: array size of @regions
1264  * @unit: ba entries per bt page
1265  */
1266 int hns_roce_hem_list_calc_root_ba(const struct hns_roce_buf_region *regions,
1267 				   int region_cnt, int unit)
1268 {
1269 	struct hns_roce_buf_region *r;
1270 	int total = 0;
1271 	int step;
1272 	int i;
1273 
1274 	for (i = 0; i < region_cnt; i++) {
1275 		r = (struct hns_roce_buf_region *)&regions[i];
1276 		if (r->hopnum > 1) {
1277 			step = hem_list_calc_ba_range(r->hopnum, 1, unit);
1278 			if (step > 0)
1279 				total += (r->count + step - 1) / step;
1280 		} else {
1281 			total += r->count;
1282 		}
1283 	}
1284 
1285 	return total;
1286 }
1287 
1288 static int hem_list_alloc_mid_bt(struct hns_roce_dev *hr_dev,
1289 				 const struct hns_roce_buf_region *r, int unit,
1290 				 int offset, struct list_head *mid_bt,
1291 				 struct list_head *btm_bt)
1292 {
1293 	struct roce_hem_item *hem_ptrs[HNS_ROCE_MAX_BT_LEVEL] = { NULL };
1294 	struct list_head temp_list[HNS_ROCE_MAX_BT_LEVEL];
1295 	struct roce_hem_item *cur, *pre;
1296 	const int hopnum = r->hopnum;
1297 	int start_aligned;
1298 	int distance;
1299 	int ret = 0;
1300 	int max_ofs;
1301 	int level;
1302 	u32 step;
1303 	int end;
1304 
1305 	if (hopnum <= 1)
1306 		return 0;
1307 
1308 	if (hopnum > HNS_ROCE_MAX_BT_LEVEL) {
1309 		dev_err(hr_dev->dev, "invalid hopnum %d!\n", hopnum);
1310 		return -EINVAL;
1311 	}
1312 
1313 	if (offset < r->offset) {
1314 		dev_err(hr_dev->dev, "invalid offset %d,min %d!\n",
1315 			offset, r->offset);
1316 		return -EINVAL;
1317 	}
1318 
1319 	distance = offset - r->offset;
1320 	max_ofs = r->offset + r->count - 1;
1321 	for (level = 0; level < hopnum; level++)
1322 		INIT_LIST_HEAD(&temp_list[level]);
1323 
1324 	/* config L1 bt to last bt and link them to corresponding parent */
1325 	for (level = 1; level < hopnum; level++) {
1326 		cur = hem_list_search_item(&mid_bt[level], offset);
1327 		if (cur) {
1328 			hem_ptrs[level] = cur;
1329 			continue;
1330 		}
1331 
1332 		step = hem_list_calc_ba_range(hopnum, level, unit);
1333 		if (step < 1) {
1334 			ret = -EINVAL;
1335 			goto err_exit;
1336 		}
1337 
1338 		start_aligned = (distance / step) * step + r->offset;
1339 		end = min_t(int, start_aligned + step - 1, max_ofs);
1340 		cur = hem_list_alloc_item(hr_dev, start_aligned, end, unit,
1341 					  true, level);
1342 		if (!cur) {
1343 			ret = -ENOMEM;
1344 			goto err_exit;
1345 		}
1346 		hem_ptrs[level] = cur;
1347 		list_add(&cur->list, &temp_list[level]);
1348 		if (hem_list_is_bottom_bt(hopnum, level))
1349 			list_add(&cur->sibling, &temp_list[0]);
1350 
1351 		/* link bt to parent bt */
1352 		if (level > 1) {
1353 			pre = hem_ptrs[level - 1];
1354 			step = (cur->start - pre->start) / step * BA_BYTE_LEN;
1355 			hem_list_link_bt(hr_dev, pre->addr + step,
1356 					 cur->dma_addr);
1357 		}
1358 	}
1359 
1360 	list_splice(&temp_list[0], btm_bt);
1361 	for (level = 1; level < hopnum; level++)
1362 		list_splice(&temp_list[level], &mid_bt[level]);
1363 
1364 	return 0;
1365 
1366 err_exit:
1367 	for (level = 1; level < hopnum; level++)
1368 		hem_list_free_all(hr_dev, &temp_list[level], true);
1369 
1370 	return ret;
1371 }
1372 
1373 static int hem_list_alloc_root_bt(struct hns_roce_dev *hr_dev,
1374 				  struct hns_roce_hem_list *hem_list, int unit,
1375 				  const struct hns_roce_buf_region *regions,
1376 				  int region_cnt)
1377 {
1378 	struct roce_hem_item *hem, *temp_hem, *root_hem;
1379 	struct list_head temp_list[HNS_ROCE_MAX_BT_REGION];
1380 	const struct hns_roce_buf_region *r;
1381 	struct list_head temp_root;
1382 	struct list_head temp_btm;
1383 	void *cpu_base;
1384 	u64 phy_base;
1385 	int ret = 0;
1386 	int offset;
1387 	int total;
1388 	int step;
1389 	int i;
1390 
1391 	r = &regions[0];
1392 	root_hem = hem_list_search_item(&hem_list->root_bt, r->offset);
1393 	if (root_hem)
1394 		return 0;
1395 
1396 	INIT_LIST_HEAD(&temp_root);
1397 	total = r->offset;
1398 	/* indicate to last region */
1399 	r = &regions[region_cnt - 1];
1400 	root_hem = hem_list_alloc_item(hr_dev, total, r->offset + r->count - 1,
1401 				       unit, true, 0);
1402 	if (!root_hem)
1403 		return -ENOMEM;
1404 	list_add(&root_hem->list, &temp_root);
1405 
1406 	hem_list->root_ba = root_hem->dma_addr;
1407 
1408 	INIT_LIST_HEAD(&temp_btm);
1409 	for (i = 0; i < region_cnt; i++)
1410 		INIT_LIST_HEAD(&temp_list[i]);
1411 
1412 	total = 0;
1413 	for (i = 0; i < region_cnt && total < unit; i++) {
1414 		r = &regions[i];
1415 		if (!r->count)
1416 			continue;
1417 
1418 		/* all regions's mid[x][0] shared the root_bt's trunk */
1419 		cpu_base = root_hem->addr + total * BA_BYTE_LEN;
1420 		phy_base = root_hem->dma_addr + total * BA_BYTE_LEN;
1421 
1422 		/* if hopnum is 0 or 1, cut a new fake hem from the root bt
1423 		 * which's address share to all regions.
1424 		 */
1425 		if (hem_list_is_bottom_bt(r->hopnum, 0)) {
1426 			hem = hem_list_alloc_item(hr_dev, r->offset,
1427 						  r->offset + r->count - 1,
1428 						  r->count, false, 0);
1429 			if (!hem) {
1430 				ret = -ENOMEM;
1431 				goto err_exit;
1432 			}
1433 			hem_list_assign_bt(hr_dev, hem, cpu_base, phy_base);
1434 			list_add(&hem->list, &temp_list[i]);
1435 			list_add(&hem->sibling, &temp_btm);
1436 			total += r->count;
1437 		} else {
1438 			step = hem_list_calc_ba_range(r->hopnum, 1, unit);
1439 			if (step < 1) {
1440 				ret = -EINVAL;
1441 				goto err_exit;
1442 			}
1443 			/* if exist mid bt, link L1 to L0 */
1444 			list_for_each_entry_safe(hem, temp_hem,
1445 					  &hem_list->mid_bt[i][1], list) {
1446 				offset = hem->start / step * BA_BYTE_LEN;
1447 				hem_list_link_bt(hr_dev, cpu_base + offset,
1448 						 hem->dma_addr);
1449 				total++;
1450 			}
1451 		}
1452 	}
1453 
1454 	list_splice(&temp_btm, &hem_list->btm_bt);
1455 	list_splice(&temp_root, &hem_list->root_bt);
1456 	for (i = 0; i < region_cnt; i++)
1457 		list_splice(&temp_list[i], &hem_list->mid_bt[i][0]);
1458 
1459 	return 0;
1460 
1461 err_exit:
1462 	for (i = 0; i < region_cnt; i++)
1463 		hem_list_free_all(hr_dev, &temp_list[i], false);
1464 
1465 	hem_list_free_all(hr_dev, &temp_root, true);
1466 
1467 	return ret;
1468 }
1469 
1470 /* construct the base address table and link them by address hop config */
1471 int hns_roce_hem_list_request(struct hns_roce_dev *hr_dev,
1472 			      struct hns_roce_hem_list *hem_list,
1473 			      const struct hns_roce_buf_region *regions,
1474 			      int region_cnt)
1475 {
1476 	const struct hns_roce_buf_region *r;
1477 	int ofs, end;
1478 	int ret = 0;
1479 	int unit;
1480 	int i;
1481 
1482 	if (region_cnt > HNS_ROCE_MAX_BT_REGION) {
1483 		dev_err(hr_dev->dev, "invalid region region_cnt %d!\n",
1484 			region_cnt);
1485 		return -EINVAL;
1486 	}
1487 
1488 	unit = (1 << hem_list->bt_pg_shift) / BA_BYTE_LEN;
1489 	for (i = 0; i < region_cnt; i++) {
1490 		r = &regions[i];
1491 		if (!r->count)
1492 			continue;
1493 
1494 		end = r->offset + r->count;
1495 		for (ofs = r->offset; ofs < end; ofs += unit) {
1496 			ret = hem_list_alloc_mid_bt(hr_dev, r, unit, ofs,
1497 						    hem_list->mid_bt[i],
1498 						    &hem_list->btm_bt);
1499 			if (ret) {
1500 				dev_err(hr_dev->dev,
1501 					"alloc hem trunk fail ret=%d!\n", ret);
1502 				goto err_alloc;
1503 			}
1504 		}
1505 	}
1506 
1507 	ret = hem_list_alloc_root_bt(hr_dev, hem_list, unit, regions,
1508 				     region_cnt);
1509 	if (ret)
1510 		dev_err(hr_dev->dev, "alloc hem root fail ret=%d!\n", ret);
1511 	else
1512 		return 0;
1513 
1514 err_alloc:
1515 	hns_roce_hem_list_release(hr_dev, hem_list);
1516 
1517 	return ret;
1518 }
1519 
1520 void hns_roce_hem_list_release(struct hns_roce_dev *hr_dev,
1521 			       struct hns_roce_hem_list *hem_list)
1522 {
1523 	int i, j;
1524 
1525 	for (i = 0; i < HNS_ROCE_MAX_BT_REGION; i++)
1526 		for (j = 0; j < HNS_ROCE_MAX_BT_LEVEL; j++)
1527 			hem_list_free_all(hr_dev, &hem_list->mid_bt[i][j],
1528 					  j != 0);
1529 
1530 	hem_list_free_all(hr_dev, &hem_list->root_bt, true);
1531 	INIT_LIST_HEAD(&hem_list->btm_bt);
1532 	hem_list->root_ba = 0;
1533 }
1534 
1535 void hns_roce_hem_list_init(struct hns_roce_hem_list *hem_list,
1536 			    int bt_page_order)
1537 {
1538 	int i, j;
1539 
1540 	INIT_LIST_HEAD(&hem_list->root_bt);
1541 	INIT_LIST_HEAD(&hem_list->btm_bt);
1542 	for (i = 0; i < HNS_ROCE_MAX_BT_REGION; i++)
1543 		for (j = 0; j < HNS_ROCE_MAX_BT_LEVEL; j++)
1544 			INIT_LIST_HEAD(&hem_list->mid_bt[i][j]);
1545 
1546 	hem_list->bt_pg_shift = bt_page_order;
1547 }
1548 
1549 void *hns_roce_hem_list_find_mtt(struct hns_roce_dev *hr_dev,
1550 				 struct hns_roce_hem_list *hem_list,
1551 				 int offset, int *mtt_cnt, u64 *phy_addr)
1552 {
1553 	struct list_head *head = &hem_list->btm_bt;
1554 	struct roce_hem_item *hem, *temp_hem;
1555 	void *cpu_base = NULL;
1556 	u64 phy_base = 0;
1557 	int nr = 0;
1558 
1559 	list_for_each_entry_safe(hem, temp_hem, head, sibling) {
1560 		if (hem_list_page_is_in_range(hem, offset)) {
1561 			nr = offset - hem->start;
1562 			cpu_base = hem->addr + nr * BA_BYTE_LEN;
1563 			phy_base = hem->dma_addr + nr * BA_BYTE_LEN;
1564 			nr = hem->end + 1 - offset;
1565 			break;
1566 		}
1567 	}
1568 
1569 	if (mtt_cnt)
1570 		*mtt_cnt = nr;
1571 
1572 	if (phy_addr)
1573 		*phy_addr = phy_base;
1574 
1575 	return cpu_base;
1576 }
1577