xref: /linux/drivers/accel/habanalabs/common/mmu/mmu_v1.c (revision d26a3a6ce7e02f9c056ad992bcd9624735022337)
1 // SPDX-License-Identifier: GPL-2.0
2 
3 /*
4  * Copyright 2016-2019 HabanaLabs, Ltd.
5  * All Rights Reserved.
6  */
7 
8 #include "../habanalabs.h"
9 #include "../../include/hw_ip/mmu/mmu_general.h"
10 
11 #include <linux/slab.h>
12 
13 #define MMU_V1_MAX_HOPS	(MMU_HOP4 + 1)
14 
15 static inline u64 get_phys_addr(struct hl_ctx *ctx, u64 shadow_addr);
16 
17 static struct pgt_info *get_pgt_info(struct hl_ctx *ctx, u64 hop_addr)
18 {
19 	struct pgt_info *pgt_info = NULL;
20 
21 	hash_for_each_possible(ctx->mmu_shadow_hash, pgt_info, node,
22 				(unsigned long) hop_addr)
23 		if (hop_addr == pgt_info->shadow_addr)
24 			break;
25 
26 	return pgt_info;
27 }
28 
29 static void _free_hop(struct hl_ctx *ctx, struct pgt_info *pgt_info)
30 {
31 	struct hl_device *hdev = ctx->hdev;
32 
33 	gen_pool_free(hdev->mmu_priv.dr.mmu_pgt_pool, pgt_info->phys_addr,
34 			hdev->asic_prop.mmu_hop_table_size);
35 	hash_del(&pgt_info->node);
36 	kfree((u64 *) (uintptr_t) pgt_info->shadow_addr);
37 	kfree(pgt_info);
38 }
39 
40 static void free_hop(struct hl_ctx *ctx, u64 hop_addr)
41 {
42 	struct pgt_info *pgt_info = get_pgt_info(ctx, hop_addr);
43 
44 	_free_hop(ctx, pgt_info);
45 }
46 
47 static u64 alloc_hop(struct hl_ctx *ctx)
48 {
49 	struct hl_device *hdev = ctx->hdev;
50 	struct asic_fixed_properties *prop = &hdev->asic_prop;
51 	struct pgt_info *pgt_info;
52 	u64 phys_addr, shadow_addr;
53 
54 	pgt_info = kmalloc(sizeof(*pgt_info), GFP_KERNEL);
55 	if (!pgt_info)
56 		return ULLONG_MAX;
57 
58 	phys_addr = (u64) gen_pool_alloc(hdev->mmu_priv.dr.mmu_pgt_pool,
59 					prop->mmu_hop_table_size);
60 	if (!phys_addr) {
61 		dev_err(hdev->dev, "failed to allocate page\n");
62 		goto pool_add_err;
63 	}
64 
65 	shadow_addr = (u64) (uintptr_t) kzalloc(prop->mmu_hop_table_size,
66 						GFP_KERNEL);
67 	if (!shadow_addr)
68 		goto shadow_err;
69 
70 	pgt_info->phys_addr = phys_addr;
71 	pgt_info->shadow_addr = shadow_addr;
72 	pgt_info->ctx = ctx;
73 	pgt_info->num_of_ptes = 0;
74 	hash_add(ctx->mmu_shadow_hash, &pgt_info->node, shadow_addr);
75 
76 	return shadow_addr;
77 
78 shadow_err:
79 	gen_pool_free(hdev->mmu_priv.dr.mmu_pgt_pool, phys_addr,
80 			prop->mmu_hop_table_size);
81 pool_add_err:
82 	kfree(pgt_info);
83 
84 	return ULLONG_MAX;
85 }
86 
87 static inline u64 get_phys_hop0_addr(struct hl_ctx *ctx)
88 {
89 	return ctx->hdev->asic_prop.mmu_pgt_addr +
90 			(ctx->asid * ctx->hdev->asic_prop.mmu_hop_table_size);
91 }
92 
93 static inline u64 get_hop0_addr(struct hl_ctx *ctx)
94 {
95 	return (u64) (uintptr_t) ctx->hdev->mmu_priv.dr.mmu_shadow_hop0 +
96 			(ctx->asid * ctx->hdev->asic_prop.mmu_hop_table_size);
97 }
98 
99 static void flush(struct hl_ctx *ctx)
100 {
101 	/* flush all writes from all cores to reach PCI */
102 	mb();
103 	ctx->hdev->asic_funcs->read_pte(ctx->hdev, get_phys_hop0_addr(ctx));
104 }
105 
106 /* transform the value to physical address when writing to H/W */
107 static inline void write_pte(struct hl_ctx *ctx, u64 shadow_pte_addr, u64 val)
108 {
109 	/*
110 	 * The value to write is actually the address of the next shadow hop +
111 	 * flags at the 12 LSBs.
112 	 * Hence in order to get the value to write to the physical PTE, we
113 	 * clear the 12 LSBs and translate the shadow hop to its associated
114 	 * physical hop, and add back the original 12 LSBs.
115 	 */
116 	u64 phys_val = get_phys_addr(ctx, val & HOP_PHYS_ADDR_MASK) |
117 				(val & FLAGS_MASK);
118 
119 	ctx->hdev->asic_funcs->write_pte(ctx->hdev,
120 					get_phys_addr(ctx, shadow_pte_addr),
121 					phys_val);
122 
123 	*(u64 *) (uintptr_t) shadow_pte_addr = val;
124 }
125 
126 /* do not transform the value to physical address when writing to H/W */
127 static inline void write_final_pte(struct hl_ctx *ctx, u64 shadow_pte_addr,
128 					u64 val)
129 {
130 	ctx->hdev->asic_funcs->write_pte(ctx->hdev,
131 					get_phys_addr(ctx, shadow_pte_addr),
132 					val);
133 	*(u64 *) (uintptr_t) shadow_pte_addr = val;
134 }
135 
136 /* clear the last and present bits */
137 static inline void clear_pte(struct hl_ctx *ctx, u64 pte_addr)
138 {
139 	/* no need to transform the value to physical address */
140 	write_final_pte(ctx, pte_addr, 0);
141 }
142 
143 static inline void get_pte(struct hl_ctx *ctx, u64 hop_addr)
144 {
145 	get_pgt_info(ctx, hop_addr)->num_of_ptes++;
146 }
147 
148 /*
149  * put_pte - decrement the num of ptes and free the hop if possible
150  *
151  * @ctx: pointer to the context structure
152  * @hop_addr: addr of the hop
153  *
154  * This function returns the number of ptes left on this hop. If the number is
155  * 0, it means the pte was freed.
156  */
157 static inline int put_pte(struct hl_ctx *ctx, u64 hop_addr)
158 {
159 	struct pgt_info *pgt_info = get_pgt_info(ctx, hop_addr);
160 	int num_of_ptes_left;
161 
162 	pgt_info->num_of_ptes--;
163 
164 	/*
165 	 * Need to save the number of ptes left because free_hop might free
166 	 * the pgt_info
167 	 */
168 	num_of_ptes_left = pgt_info->num_of_ptes;
169 	if (!num_of_ptes_left)
170 		_free_hop(ctx, pgt_info);
171 
172 	return num_of_ptes_left;
173 }
174 
175 static inline u64 get_hop_pte_addr(struct hl_ctx *ctx, struct hl_mmu_properties *mmu_prop,
176 					u64 *hop_addr_arr, u64 virt_addr, enum mmu_hop_num hop_idx)
177 {
178 	u64 mask, shift;
179 
180 	mask = mmu_prop->hop_masks[hop_idx];
181 	shift = mmu_prop->hop_shifts[hop_idx];
182 	return hop_addr_arr[hop_idx] +
183 			ctx->hdev->asic_prop.mmu_pte_size * ((virt_addr & mask) >> shift);
184 }
185 
186 static inline u64 get_alloc_next_hop_addr(struct hl_ctx *ctx, u64 curr_pte,
187 						bool *is_new_hop)
188 {
189 	u64 hop_addr = hl_mmu_get_next_hop_addr(ctx, curr_pte);
190 
191 	if (hop_addr == ULLONG_MAX) {
192 		hop_addr = alloc_hop(ctx);
193 		*is_new_hop = (hop_addr != ULLONG_MAX);
194 	}
195 
196 	return hop_addr;
197 }
198 
199 /* translates shadow address inside hop to a physical address */
200 static inline u64 get_phys_addr(struct hl_ctx *ctx, u64 shadow_addr)
201 {
202 	u64 page_mask = (ctx->hdev->asic_prop.mmu_hop_table_size - 1);
203 	u64 shadow_hop_addr = shadow_addr & ~page_mask;
204 	u64 pte_offset = shadow_addr & page_mask;
205 	u64 phys_hop_addr;
206 
207 	if (shadow_hop_addr != get_hop0_addr(ctx))
208 		phys_hop_addr = get_pgt_info(ctx, shadow_hop_addr)->phys_addr;
209 	else
210 		phys_hop_addr = get_phys_hop0_addr(ctx);
211 
212 	return phys_hop_addr + pte_offset;
213 }
214 
215 static int dram_default_mapping_init(struct hl_ctx *ctx)
216 {
217 	struct hl_device *hdev = ctx->hdev;
218 	struct asic_fixed_properties *prop = &hdev->asic_prop;
219 	u64 num_of_hop3, total_hops, hop0_addr, hop1_addr, hop2_addr,
220 		hop2_pte_addr, hop3_pte_addr, pte_val;
221 	int rc, i, j, hop3_allocated = 0;
222 
223 	if ((!prop->dram_supports_virtual_memory) ||
224 			(!hdev->dram_default_page_mapping) ||
225 			(ctx->asid == HL_KERNEL_ASID_ID))
226 		return 0;
227 
228 	num_of_hop3 = prop->dram_size_for_default_page_mapping;
229 	do_div(num_of_hop3, prop->dram_page_size);
230 	do_div(num_of_hop3, HOP_PTE_ENTRIES_512);
231 
232 	/* add hop1 and hop2 */
233 	total_hops = num_of_hop3 + 2;
234 
235 	ctx->dram_default_hops = kzalloc(HL_PTE_SIZE * total_hops,  GFP_KERNEL);
236 	if (!ctx->dram_default_hops)
237 		return -ENOMEM;
238 
239 	hop0_addr = get_hop0_addr(ctx);
240 
241 	hop1_addr = alloc_hop(ctx);
242 	if (hop1_addr == ULLONG_MAX) {
243 		dev_err(hdev->dev, "failed to alloc hop 1\n");
244 		rc = -ENOMEM;
245 		goto hop1_err;
246 	}
247 
248 	ctx->dram_default_hops[total_hops - 1] = hop1_addr;
249 
250 	hop2_addr = alloc_hop(ctx);
251 	if (hop2_addr == ULLONG_MAX) {
252 		dev_err(hdev->dev, "failed to alloc hop 2\n");
253 		rc = -ENOMEM;
254 		goto hop2_err;
255 	}
256 
257 	ctx->dram_default_hops[total_hops - 2] = hop2_addr;
258 
259 	for (i = 0 ; i < num_of_hop3 ; i++) {
260 		ctx->dram_default_hops[i] = alloc_hop(ctx);
261 		if (ctx->dram_default_hops[i] == ULLONG_MAX) {
262 			dev_err(hdev->dev, "failed to alloc hop 3, i: %d\n", i);
263 			rc = -ENOMEM;
264 			goto hop3_err;
265 		}
266 		hop3_allocated++;
267 	}
268 
269 	/* need only pte 0 in hops 0 and 1 */
270 	pte_val = (hop1_addr & HOP_PHYS_ADDR_MASK) | PAGE_PRESENT_MASK;
271 	write_pte(ctx, hop0_addr, pte_val);
272 
273 	pte_val = (hop2_addr & HOP_PHYS_ADDR_MASK) | PAGE_PRESENT_MASK;
274 	write_pte(ctx, hop1_addr, pte_val);
275 	get_pte(ctx, hop1_addr);
276 
277 	hop2_pte_addr = hop2_addr;
278 	for (i = 0 ; i < num_of_hop3 ; i++) {
279 		pte_val = (ctx->dram_default_hops[i] & HOP_PHYS_ADDR_MASK) |
280 				PAGE_PRESENT_MASK;
281 		write_pte(ctx, hop2_pte_addr, pte_val);
282 		get_pte(ctx, hop2_addr);
283 		hop2_pte_addr += HL_PTE_SIZE;
284 	}
285 
286 	pte_val = (prop->mmu_dram_default_page_addr & HOP_PHYS_ADDR_MASK) |
287 			LAST_MASK | PAGE_PRESENT_MASK;
288 
289 	for (i = 0 ; i < num_of_hop3 ; i++) {
290 		hop3_pte_addr = ctx->dram_default_hops[i];
291 		for (j = 0 ; j < HOP_PTE_ENTRIES_512 ; j++) {
292 			write_final_pte(ctx, hop3_pte_addr, pte_val);
293 			get_pte(ctx, ctx->dram_default_hops[i]);
294 			hop3_pte_addr += HL_PTE_SIZE;
295 		}
296 	}
297 
298 	flush(ctx);
299 
300 	return 0;
301 
302 hop3_err:
303 	for (i = 0 ; i < hop3_allocated ; i++)
304 		free_hop(ctx, ctx->dram_default_hops[i]);
305 
306 	free_hop(ctx, hop2_addr);
307 hop2_err:
308 	free_hop(ctx, hop1_addr);
309 hop1_err:
310 	kfree(ctx->dram_default_hops);
311 
312 	return rc;
313 }
314 
315 static void dram_default_mapping_fini(struct hl_ctx *ctx)
316 {
317 	struct hl_device *hdev = ctx->hdev;
318 	struct asic_fixed_properties *prop = &hdev->asic_prop;
319 	u64 num_of_hop3, total_hops, hop0_addr, hop1_addr, hop2_addr,
320 		hop2_pte_addr, hop3_pte_addr;
321 	int i, j;
322 
323 	if ((!prop->dram_supports_virtual_memory) ||
324 			(!hdev->dram_default_page_mapping) ||
325 			(ctx->asid == HL_KERNEL_ASID_ID))
326 		return;
327 
328 	num_of_hop3 = prop->dram_size_for_default_page_mapping;
329 	do_div(num_of_hop3, prop->dram_page_size);
330 	do_div(num_of_hop3, HOP_PTE_ENTRIES_512);
331 
332 	hop0_addr = get_hop0_addr(ctx);
333 	/* add hop1 and hop2 */
334 	total_hops = num_of_hop3 + 2;
335 	hop1_addr = ctx->dram_default_hops[total_hops - 1];
336 	hop2_addr = ctx->dram_default_hops[total_hops - 2];
337 
338 	for (i = 0 ; i < num_of_hop3 ; i++) {
339 		hop3_pte_addr = ctx->dram_default_hops[i];
340 		for (j = 0 ; j < HOP_PTE_ENTRIES_512 ; j++) {
341 			clear_pte(ctx, hop3_pte_addr);
342 			put_pte(ctx, ctx->dram_default_hops[i]);
343 			hop3_pte_addr += HL_PTE_SIZE;
344 		}
345 	}
346 
347 	hop2_pte_addr = hop2_addr;
348 	for (i = 0 ; i < num_of_hop3 ; i++) {
349 		clear_pte(ctx, hop2_pte_addr);
350 		put_pte(ctx, hop2_addr);
351 		hop2_pte_addr += HL_PTE_SIZE;
352 	}
353 
354 	clear_pte(ctx, hop1_addr);
355 	put_pte(ctx, hop1_addr);
356 	clear_pte(ctx, hop0_addr);
357 
358 	kfree(ctx->dram_default_hops);
359 
360 	flush(ctx);
361 }
362 
363 /**
364  * hl_mmu_v1_init() - initialize the MMU module.
365  * @hdev: habanalabs device structure.
366  *
367  * This function does the following:
368  * - Create a pool of pages for pgt_infos.
369  * - Create a shadow table for pgt
370  *
371  * Return: 0 for success, non-zero for failure.
372  */
373 static int hl_mmu_v1_init(struct hl_device *hdev)
374 {
375 	struct asic_fixed_properties *prop = &hdev->asic_prop;
376 	int rc;
377 
378 	hdev->mmu_priv.dr.mmu_pgt_pool =
379 			gen_pool_create(__ffs(prop->mmu_hop_table_size), -1);
380 
381 	if (!hdev->mmu_priv.dr.mmu_pgt_pool) {
382 		dev_err(hdev->dev, "Failed to create page gen pool\n");
383 		return -ENOMEM;
384 	}
385 
386 	rc = gen_pool_add(hdev->mmu_priv.dr.mmu_pgt_pool, prop->mmu_pgt_addr +
387 			prop->mmu_hop0_tables_total_size,
388 			prop->mmu_pgt_size - prop->mmu_hop0_tables_total_size,
389 			-1);
390 	if (rc) {
391 		dev_err(hdev->dev, "Failed to add memory to page gen pool\n");
392 		goto err_pool_add;
393 	}
394 
395 	hdev->mmu_priv.dr.mmu_shadow_hop0 = kvcalloc(prop->max_asid, prop->mmu_hop_table_size,
396 										GFP_KERNEL);
397 	if (ZERO_OR_NULL_PTR(hdev->mmu_priv.dr.mmu_shadow_hop0)) {
398 		rc = -ENOMEM;
399 		goto err_pool_add;
400 	}
401 
402 	/* MMU H/W init will be done in device hw_init() */
403 
404 	return 0;
405 
406 err_pool_add:
407 	gen_pool_destroy(hdev->mmu_priv.dr.mmu_pgt_pool);
408 
409 	return rc;
410 }
411 
412 /**
413  * hl_mmu_v1_fini() - release the MMU module.
414  * @hdev: habanalabs device structure.
415  *
416  * This function does the following:
417  * - Disable MMU in H/W.
418  * - Free the pgt_infos pool.
419  *
420  * All contexts should be freed before calling this function.
421  */
422 static void hl_mmu_v1_fini(struct hl_device *hdev)
423 {
424 	/* MMU H/W fini was already done in device hw_fini() */
425 
426 	if (!ZERO_OR_NULL_PTR(hdev->mmu_priv.dr.mmu_shadow_hop0)) {
427 		kvfree(hdev->mmu_priv.dr.mmu_shadow_hop0);
428 		gen_pool_destroy(hdev->mmu_priv.dr.mmu_pgt_pool);
429 
430 		/* Make sure that if we arrive here again without init was
431 		 * called we won't cause kernel panic. This can happen for
432 		 * example if we fail during hard reset code at certain points
433 		 */
434 		hdev->mmu_priv.dr.mmu_shadow_hop0 = NULL;
435 	}
436 }
437 
438 /**
439  * hl_mmu_v1_ctx_init() - initialize a context for using the MMU module.
440  * @ctx: pointer to the context structure to initialize.
441  *
442  * Initialize a mutex to protect the concurrent mapping flow, a hash to hold all
443  * page tables hops related to this context.
444  * Return: 0 on success, non-zero otherwise.
445  */
446 static int hl_mmu_v1_ctx_init(struct hl_ctx *ctx)
447 {
448 	hash_init(ctx->mmu_shadow_hash);
449 	return dram_default_mapping_init(ctx);
450 }
451 
452 /*
453  * hl_mmu_ctx_fini - disable a ctx from using the mmu module
454  *
455  * @ctx: pointer to the context structure
456  *
457  * This function does the following:
458  * - Free any pgts which were not freed yet
459  * - Free the mutex
460  * - Free DRAM default page mapping hops
461  */
462 static void hl_mmu_v1_ctx_fini(struct hl_ctx *ctx)
463 {
464 	struct hl_device *hdev = ctx->hdev;
465 	struct pgt_info *pgt_info;
466 	struct hlist_node *tmp;
467 	int i;
468 
469 	dram_default_mapping_fini(ctx);
470 
471 	if (!hash_empty(ctx->mmu_shadow_hash))
472 		dev_err(hdev->dev, "ctx %d is freed while it has pgts in use\n",
473 			ctx->asid);
474 
475 	hash_for_each_safe(ctx->mmu_shadow_hash, i, tmp, pgt_info, node) {
476 		dev_err_ratelimited(hdev->dev,
477 			"pgt_info of addr 0x%llx of asid %d was not destroyed, num_ptes: %d\n",
478 			pgt_info->phys_addr, ctx->asid, pgt_info->num_of_ptes);
479 		_free_hop(ctx, pgt_info);
480 	}
481 }
482 
483 static int hl_mmu_v1_unmap(struct hl_ctx *ctx,
484 				u64 virt_addr, bool is_dram_addr)
485 {
486 	u64 hop_addr[MMU_V1_MAX_HOPS] = {0}, hop_pte_addr[MMU_V1_MAX_HOPS] = {0}, curr_pte = 0;
487 	struct hl_device *hdev = ctx->hdev;
488 	struct asic_fixed_properties *prop = &hdev->asic_prop;
489 	struct hl_mmu_properties *mmu_prop;
490 	bool is_huge, clear_hop3 = true;
491 	int hop_idx;
492 
493 	/* shifts and masks are the same in PMMU and HPMMU, use one of them */
494 	mmu_prop = is_dram_addr ? &prop->dmmu : &prop->pmmu;
495 
496 	for (hop_idx = MMU_HOP0; hop_idx < MMU_HOP4; hop_idx++) {
497 		if (hop_idx == MMU_HOP0) {
498 			hop_addr[hop_idx] = get_hop0_addr(ctx);
499 		} else {
500 			hop_addr[hop_idx] = hl_mmu_get_next_hop_addr(ctx, curr_pte);
501 			if (hop_addr[hop_idx] == ULLONG_MAX)
502 				goto not_mapped;
503 		}
504 
505 		hop_pte_addr[hop_idx] =
506 				get_hop_pte_addr(ctx, mmu_prop, hop_addr, virt_addr, hop_idx);
507 
508 		curr_pte = *(u64 *) (uintptr_t) hop_pte_addr[hop_idx];
509 	}
510 
511 	is_huge = curr_pte & mmu_prop->last_mask;
512 
513 	if (is_dram_addr && !is_huge) {
514 		dev_err(hdev->dev, "DRAM unmapping should use huge pages only\n");
515 		return -EFAULT;
516 	}
517 
518 	if (!is_huge) {
519 		hop_idx = MMU_HOP4;
520 		hop_addr[hop_idx] = hl_mmu_get_next_hop_addr(ctx, curr_pte);
521 		if (hop_addr[hop_idx] == ULLONG_MAX)
522 			goto not_mapped;
523 
524 		hop_pte_addr[hop_idx] =
525 				get_hop_pte_addr(ctx, mmu_prop, hop_addr, virt_addr, hop_idx);
526 		curr_pte = *(u64 *) (uintptr_t) hop_pte_addr[hop_idx];
527 		clear_hop3 = false;
528 	}
529 
530 	if (hdev->dram_default_page_mapping && is_dram_addr) {
531 		u64 default_pte = (prop->mmu_dram_default_page_addr &
532 				HOP_PHYS_ADDR_MASK) | mmu_prop->last_mask |
533 					PAGE_PRESENT_MASK;
534 		if (curr_pte == default_pte) {
535 			dev_err(hdev->dev,
536 				"DRAM: hop3 PTE points to zero page, can't unmap, va: 0x%llx\n",
537 					virt_addr);
538 			goto not_mapped;
539 		}
540 
541 		if (!(curr_pte & PAGE_PRESENT_MASK)) {
542 			dev_err(hdev->dev,
543 				"DRAM: hop3 PTE is cleared! can't unmap, va: 0x%llx\n",
544 					virt_addr);
545 			goto not_mapped;
546 		}
547 
548 		hop_idx = MMU_HOP3;
549 		write_final_pte(ctx, hop_pte_addr[hop_idx], default_pte);
550 		put_pte(ctx, hop_addr[hop_idx]);
551 	} else {
552 		if (!(curr_pte & PAGE_PRESENT_MASK))
553 			goto not_mapped;
554 
555 		if (hop_addr[MMU_HOP4])
556 			clear_pte(ctx, hop_pte_addr[MMU_HOP4]);
557 		else
558 			clear_pte(ctx, hop_pte_addr[MMU_HOP3]);
559 
560 		if (hop_addr[MMU_HOP4] && !put_pte(ctx, hop_addr[MMU_HOP4]))
561 			clear_hop3 = true;
562 
563 		if (!clear_hop3)
564 			goto mapped;
565 
566 		for (hop_idx = MMU_HOP3; hop_idx >= 0; hop_idx--) {
567 			clear_pte(ctx, hop_pte_addr[hop_idx]);
568 
569 			if (hop_idx == MMU_HOP0)
570 				break;
571 
572 			if (put_pte(ctx, hop_addr[hop_idx]))
573 				goto mapped;
574 		}
575 	}
576 
577 mapped:
578 	return 0;
579 
580 not_mapped:
581 	dev_err(hdev->dev, "virt addr 0x%llx is not mapped to phys addr\n",
582 		virt_addr);
583 
584 	return -EINVAL;
585 }
586 
587 static int hl_mmu_v1_map(struct hl_ctx *ctx, u64 virt_addr, u64 phys_addr,
588 			u32 page_size, bool is_dram_addr)
589 {
590 	u64 hop_addr[MMU_V1_MAX_HOPS] = {0}, hop_pte_addr[MMU_V1_MAX_HOPS] = {0}, curr_pte = 0;
591 	struct hl_device *hdev = ctx->hdev;
592 	struct asic_fixed_properties *prop = &hdev->asic_prop;
593 	struct hl_mmu_properties *mmu_prop;
594 	bool is_huge, hop_new[MMU_V1_MAX_HOPS] = {false};
595 	int num_hops, hop_idx, prev_hop, rc = -ENOMEM;
596 
597 	/*
598 	 * This mapping function can map a page or a huge page. For huge page
599 	 * there are only 3 hops rather than 4. Currently the DRAM allocation
600 	 * uses huge pages only but user memory could have been allocated with
601 	 * one of the two page sizes. Since this is a common code for all the
602 	 * three cases, we need this hugs page check.
603 	 */
604 	if (is_dram_addr) {
605 		mmu_prop = &prop->dmmu;
606 		is_huge = true;
607 	} else if (page_size == prop->pmmu_huge.page_size) {
608 		mmu_prop = &prop->pmmu_huge;
609 		is_huge = true;
610 	} else {
611 		mmu_prop = &prop->pmmu;
612 		is_huge = false;
613 	}
614 
615 	num_hops = is_huge ? (MMU_V1_MAX_HOPS - 1) : MMU_V1_MAX_HOPS;
616 
617 	for (hop_idx = MMU_HOP0; hop_idx < num_hops; hop_idx++) {
618 		if (hop_idx == MMU_HOP0) {
619 			hop_addr[hop_idx] = get_hop0_addr(ctx);
620 		} else {
621 			hop_addr[hop_idx] =
622 					get_alloc_next_hop_addr(ctx, curr_pte, &hop_new[hop_idx]);
623 			if (hop_addr[hop_idx] == ULLONG_MAX)
624 				goto err;
625 		}
626 
627 		hop_pte_addr[hop_idx] =
628 				get_hop_pte_addr(ctx, mmu_prop, hop_addr, virt_addr, hop_idx);
629 		curr_pte = *(u64 *) (uintptr_t) hop_pte_addr[hop_idx];
630 	}
631 
632 	if (hdev->dram_default_page_mapping && is_dram_addr) {
633 		u64 default_pte = (prop->mmu_dram_default_page_addr &
634 					HOP_PHYS_ADDR_MASK) | mmu_prop->last_mask |
635 						PAGE_PRESENT_MASK;
636 
637 		if (curr_pte != default_pte) {
638 			dev_err(hdev->dev,
639 				"DRAM: mapping already exists for virt_addr 0x%llx\n",
640 					virt_addr);
641 			rc = -EINVAL;
642 			goto err;
643 		}
644 
645 		for (hop_idx = MMU_HOP1; hop_idx < num_hops; hop_idx++) {
646 			if (hop_new[hop_idx]) {
647 				dev_err(hdev->dev, "DRAM mapping should not allocate more hops\n");
648 				rc = -EFAULT;
649 				goto err;
650 			}
651 		}
652 	} else if (curr_pte & PAGE_PRESENT_MASK) {
653 		dev_err(hdev->dev,
654 			"mapping already exists for virt_addr 0x%llx\n",
655 				virt_addr);
656 
657 		for (hop_idx = MMU_HOP0; hop_idx < num_hops; hop_idx++)
658 			dev_dbg(hdev->dev, "hop%d pte: 0x%llx (0x%llx)\n", hop_idx,
659 					*(u64 *) (uintptr_t) hop_pte_addr[hop_idx],
660 					hop_pte_addr[hop_idx]);
661 
662 		rc = -EINVAL;
663 		goto err;
664 	}
665 
666 	curr_pte = (phys_addr & HOP_PHYS_ADDR_MASK) | mmu_prop->last_mask
667 			| PAGE_PRESENT_MASK;
668 
669 	write_final_pte(ctx, hop_pte_addr[num_hops - 1], curr_pte);
670 
671 	for (hop_idx = MMU_HOP1; hop_idx < num_hops; hop_idx++) {
672 		prev_hop = hop_idx - 1;
673 
674 		if (hop_new[hop_idx]) {
675 			curr_pte = (hop_addr[hop_idx] & HOP_PHYS_ADDR_MASK) | PAGE_PRESENT_MASK;
676 			write_pte(ctx, hop_pte_addr[prev_hop], curr_pte);
677 			if (hop_idx != MMU_HOP1)
678 				get_pte(ctx, hop_addr[prev_hop]);
679 		}
680 	}
681 
682 	get_pte(ctx, hop_addr[num_hops - 1]);
683 
684 	return 0;
685 
686 err:
687 	for (hop_idx = num_hops; hop_idx > MMU_HOP0; hop_idx--) {
688 		if (hop_new[hop_idx])
689 			free_hop(ctx, hop_addr[hop_idx]);
690 	}
691 
692 	return rc;
693 }
694 
695 /*
696  * hl_mmu_v1_swap_out - marks all mapping of the given ctx as swapped out
697  *
698  * @ctx: pointer to the context structure
699  *
700  */
701 static void hl_mmu_v1_swap_out(struct hl_ctx *ctx)
702 {
703 
704 }
705 
706 /*
707  * hl_mmu_v1_swap_in - marks all mapping of the given ctx as swapped in
708  *
709  * @ctx: pointer to the context structure
710  *
711  */
712 static void hl_mmu_v1_swap_in(struct hl_ctx *ctx)
713 {
714 
715 }
716 
717 static int hl_mmu_v1_get_tlb_info(struct hl_ctx *ctx, u64 virt_addr,
718 				struct hl_mmu_hop_info *hops)
719 {
720 	struct hl_device *hdev = ctx->hdev;
721 	struct asic_fixed_properties *prop = &hdev->asic_prop;
722 	struct hl_mmu_properties *mmu_prop;
723 	bool is_dram_addr, is_pmmu_addr, is_pmmu_h_addr, is_huge;
724 	int i, used_hops;
725 
726 	is_dram_addr = hl_mem_area_inside_range(virt_addr, prop->dmmu.page_size,
727 						prop->dmmu.start_addr,
728 						prop->dmmu.end_addr);
729 	is_pmmu_addr = hl_mem_area_inside_range(virt_addr, prop->pmmu.page_size,
730 						prop->pmmu.start_addr,
731 						prop->pmmu.end_addr);
732 	is_pmmu_h_addr = hl_mem_area_inside_range(virt_addr,
733 						prop->pmmu_huge.page_size,
734 						prop->pmmu_huge.start_addr,
735 						prop->pmmu_huge.end_addr);
736 	if (is_dram_addr) {
737 		mmu_prop = &prop->dmmu;
738 		is_huge = true;
739 	} else if (is_pmmu_addr) {
740 		mmu_prop = &prop->pmmu;
741 		is_huge = false;
742 	} else if (is_pmmu_h_addr) {
743 		mmu_prop = &prop->pmmu_huge;
744 		is_huge = true;
745 	} else {
746 		return -EINVAL;
747 	}
748 
749 	used_hops = mmu_prop->num_hops;
750 
751 	/* huge pages use lesser hops */
752 	if (is_huge)
753 		used_hops--;
754 
755 	hops->hop_info[0].hop_addr = get_phys_hop0_addr(ctx);
756 	hops->hop_info[0].hop_pte_addr =
757 			hl_mmu_get_hop_pte_phys_addr(ctx, mmu_prop, 0,
758 					hops->hop_info[0].hop_addr, virt_addr);
759 	hops->hop_info[0].hop_pte_val =
760 			hdev->asic_funcs->read_pte(hdev,
761 						hops->hop_info[0].hop_pte_addr);
762 
763 	for (i = 1 ; i < used_hops ; i++) {
764 		hops->hop_info[i].hop_addr =
765 			hl_mmu_get_next_hop_addr(ctx,
766 					hops->hop_info[i - 1].hop_pte_val);
767 		if (hops->hop_info[i].hop_addr == ULLONG_MAX)
768 			return -EFAULT;
769 
770 		hops->hop_info[i].hop_pte_addr =
771 				hl_mmu_get_hop_pte_phys_addr(ctx, mmu_prop, i,
772 						hops->hop_info[i].hop_addr,
773 						virt_addr);
774 		hops->hop_info[i].hop_pte_val =
775 				hdev->asic_funcs->read_pte(hdev,
776 						hops->hop_info[i].hop_pte_addr);
777 
778 		if (!(hops->hop_info[i].hop_pte_val & PAGE_PRESENT_MASK))
779 			return -EFAULT;
780 
781 		if (hops->hop_info[i].hop_pte_val & mmu_prop->last_mask)
782 			break;
783 	}
784 
785 	/* if passed over all hops then no last hop was found */
786 	if (i == mmu_prop->num_hops)
787 		return -EFAULT;
788 
789 	if (!(hops->hop_info[i].hop_pte_val & PAGE_PRESENT_MASK))
790 		return -EFAULT;
791 
792 	hops->used_hops = i + 1;
793 
794 	return 0;
795 }
796 
797 /*
798  * hl_mmu_v1_prepare - prepare mmu  for working with mmu v1
799  *
800  * @hdev: pointer to the device structure
801  */
802 void hl_mmu_v1_set_funcs(struct hl_device *hdev, struct hl_mmu_funcs *mmu)
803 {
804 	mmu->init = hl_mmu_v1_init;
805 	mmu->fini = hl_mmu_v1_fini;
806 	mmu->ctx_init = hl_mmu_v1_ctx_init;
807 	mmu->ctx_fini = hl_mmu_v1_ctx_fini;
808 	mmu->map = hl_mmu_v1_map;
809 	mmu->unmap = hl_mmu_v1_unmap;
810 	mmu->flush = flush;
811 	mmu->swap_out = hl_mmu_v1_swap_out;
812 	mmu->swap_in = hl_mmu_v1_swap_in;
813 	mmu->get_tlb_info = hl_mmu_v1_get_tlb_info;
814 }
815