1 // SPDX-License-Identifier: GPL-2.0
4 * Copyright 2016-2019 HabanaLabs, Ltd.
8 #include "habanalabs.h"
9 #include "include/hw_ip/mmu/mmu_general.h"
11 #include <linux/genalloc.h>
12 #include <linux/slab.h>
14 static inline u64 get_phys_addr(struct hl_ctx *ctx, u64 shadow_addr);
16 static struct pgt_info *get_pgt_info(struct hl_ctx *ctx, u64 hop_addr)
18 struct pgt_info *pgt_info = NULL;
20 hash_for_each_possible(ctx->mmu_shadow_hash, pgt_info, node,
21 (unsigned long) hop_addr)
22 if (hop_addr == pgt_info->shadow_addr)
28 static void free_hop(struct hl_ctx *ctx, u64 hop_addr)
30 struct hl_device *hdev = ctx->hdev;
31 struct pgt_info *pgt_info = get_pgt_info(ctx, hop_addr);
33 gen_pool_free(hdev->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);
40 static u64 alloc_hop(struct hl_ctx *ctx)
42 struct hl_device *hdev = ctx->hdev;
43 struct asic_fixed_properties *prop = &hdev->asic_prop;
44 struct pgt_info *pgt_info;
45 u64 phys_addr, shadow_addr;
47 pgt_info = kmalloc(sizeof(*pgt_info), GFP_KERNEL);
51 phys_addr = (u64) gen_pool_alloc(hdev->mmu_pgt_pool,
52 prop->mmu_hop_table_size);
54 dev_err(hdev->dev, "failed to allocate page\n");
58 shadow_addr = (u64) (uintptr_t) kzalloc(prop->mmu_hop_table_size,
63 pgt_info->phys_addr = phys_addr;
64 pgt_info->shadow_addr = shadow_addr;
66 pgt_info->num_of_ptes = 0;
67 hash_add(ctx->mmu_shadow_hash, &pgt_info->node, shadow_addr);
72 gen_pool_free(hdev->mmu_pgt_pool, phys_addr, prop->mmu_hop_table_size);
79 static inline u64 get_phys_hop0_addr(struct hl_ctx *ctx)
81 return ctx->hdev->asic_prop.mmu_pgt_addr +
82 (ctx->asid * ctx->hdev->asic_prop.mmu_hop_table_size);
85 static inline u64 get_hop0_addr(struct hl_ctx *ctx)
87 return (u64) (uintptr_t) ctx->hdev->mmu_shadow_hop0 +
88 (ctx->asid * ctx->hdev->asic_prop.mmu_hop_table_size);
91 static inline void flush(struct hl_ctx *ctx)
93 /* flush all writes from all cores to reach PCI */
95 ctx->hdev->asic_funcs->read_pte(ctx->hdev, get_phys_hop0_addr(ctx));
98 /* transform the value to physical address when writing to H/W */
99 static inline void write_pte(struct hl_ctx *ctx, u64 shadow_pte_addr, u64 val)
102 * The value to write is actually the address of the next shadow hop +
103 * flags at the 12 LSBs.
104 * Hence in order to get the value to write to the physical PTE, we
105 * clear the 12 LSBs and translate the shadow hop to its associated
106 * physical hop, and add back the original 12 LSBs.
108 u64 phys_val = get_phys_addr(ctx, val & PTE_PHYS_ADDR_MASK) |
111 ctx->hdev->asic_funcs->write_pte(ctx->hdev,
112 get_phys_addr(ctx, shadow_pte_addr),
115 *(u64 *) (uintptr_t) shadow_pte_addr = val;
118 /* do not transform the value to physical address when writing to H/W */
119 static inline void write_final_pte(struct hl_ctx *ctx, u64 shadow_pte_addr,
122 ctx->hdev->asic_funcs->write_pte(ctx->hdev,
123 get_phys_addr(ctx, shadow_pte_addr),
125 *(u64 *) (uintptr_t) shadow_pte_addr = val;
128 /* clear the last and present bits */
129 static inline void clear_pte(struct hl_ctx *ctx, u64 pte_addr)
131 /* no need to transform the value to physical address */
132 write_final_pte(ctx, pte_addr, 0);
135 static inline void get_pte(struct hl_ctx *ctx, u64 hop_addr)
137 get_pgt_info(ctx, hop_addr)->num_of_ptes++;
141 * put_pte - decrement the num of ptes and free the hop if possible
143 * @ctx: pointer to the context structure
144 * @hop_addr: addr of the hop
146 * This function returns the number of ptes left on this hop. If the number is
147 * 0, it means the pte was freed.
149 static inline int put_pte(struct hl_ctx *ctx, u64 hop_addr)
151 struct pgt_info *pgt_info = get_pgt_info(ctx, hop_addr);
152 int num_of_ptes_left;
154 pgt_info->num_of_ptes--;
157 * Need to save the number of ptes left because free_hop might free
160 num_of_ptes_left = pgt_info->num_of_ptes;
161 if (!num_of_ptes_left)
162 free_hop(ctx, hop_addr);
164 return num_of_ptes_left;
167 static inline u64 get_hopN_pte_addr(struct hl_ctx *ctx, u64 hop_addr,
168 u64 virt_addr, u64 mask, u64 shift)
170 return hop_addr + ctx->hdev->asic_prop.mmu_pte_size *
171 ((virt_addr & mask) >> shift);
174 static inline u64 get_hop0_pte_addr(struct hl_ctx *ctx, u64 hop_addr, u64 vaddr)
176 return get_hopN_pte_addr(ctx, hop_addr, vaddr, HOP0_MASK, HOP0_SHIFT);
179 static inline u64 get_hop1_pte_addr(struct hl_ctx *ctx, u64 hop_addr, u64 vaddr)
181 return get_hopN_pte_addr(ctx, hop_addr, vaddr, HOP1_MASK, HOP1_SHIFT);
184 static inline u64 get_hop2_pte_addr(struct hl_ctx *ctx, u64 hop_addr, u64 vaddr)
186 return get_hopN_pte_addr(ctx, hop_addr, vaddr, HOP2_MASK, HOP2_SHIFT);
189 static inline u64 get_hop3_pte_addr(struct hl_ctx *ctx, u64 hop_addr, u64 vaddr)
191 return get_hopN_pte_addr(ctx, hop_addr, vaddr, HOP3_MASK, HOP3_SHIFT);
194 static inline u64 get_hop4_pte_addr(struct hl_ctx *ctx, u64 hop_addr, u64 vaddr)
196 return get_hopN_pte_addr(ctx, hop_addr, vaddr, HOP4_MASK, HOP4_SHIFT);
199 static inline u64 get_next_hop_addr(struct hl_ctx *ctx, u64 curr_pte)
201 if (curr_pte & PAGE_PRESENT_MASK)
202 return curr_pte & PHYS_ADDR_MASK;
207 static inline u64 get_alloc_next_hop_addr(struct hl_ctx *ctx, u64 curr_pte,
210 u64 hop_addr = get_next_hop_addr(ctx, curr_pte);
212 if (hop_addr == ULLONG_MAX) {
213 hop_addr = alloc_hop(ctx);
214 *is_new_hop = (hop_addr != ULLONG_MAX);
220 /* translates shadow address inside hop to a physical address */
221 static inline u64 get_phys_addr(struct hl_ctx *ctx, u64 shadow_addr)
223 u64 page_mask = (ctx->hdev->asic_prop.mmu_hop_table_size - 1);
224 u64 shadow_hop_addr = shadow_addr & ~page_mask;
225 u64 pte_offset = shadow_addr & page_mask;
228 if (shadow_hop_addr != get_hop0_addr(ctx))
229 phys_hop_addr = get_pgt_info(ctx, shadow_hop_addr)->phys_addr;
231 phys_hop_addr = get_phys_hop0_addr(ctx);
233 return phys_hop_addr + pte_offset;
236 static int dram_default_mapping_init(struct hl_ctx *ctx)
238 struct hl_device *hdev = ctx->hdev;
239 struct asic_fixed_properties *prop = &hdev->asic_prop;
240 u64 num_of_hop3, total_hops, hop0_addr, hop1_addr, hop2_addr,
241 hop2_pte_addr, hop3_pte_addr, pte_val;
242 int rc, i, j, hop3_allocated = 0;
244 if ((!hdev->dram_supports_virtual_memory) ||
245 (!hdev->dram_default_page_mapping) ||
246 (ctx->asid == HL_KERNEL_ASID_ID))
249 num_of_hop3 = prop->dram_size_for_default_page_mapping;
250 do_div(num_of_hop3, prop->dram_page_size);
251 do_div(num_of_hop3, PTE_ENTRIES_IN_HOP);
253 /* add hop1 and hop2 */
254 total_hops = num_of_hop3 + 2;
256 ctx->dram_default_hops = kzalloc(HL_PTE_SIZE * total_hops, GFP_KERNEL);
257 if (!ctx->dram_default_hops)
260 hop0_addr = get_hop0_addr(ctx);
262 hop1_addr = alloc_hop(ctx);
263 if (hop1_addr == ULLONG_MAX) {
264 dev_err(hdev->dev, "failed to alloc hop 1\n");
269 ctx->dram_default_hops[total_hops - 1] = hop1_addr;
271 hop2_addr = alloc_hop(ctx);
272 if (hop2_addr == ULLONG_MAX) {
273 dev_err(hdev->dev, "failed to alloc hop 2\n");
278 ctx->dram_default_hops[total_hops - 2] = hop2_addr;
280 for (i = 0 ; i < num_of_hop3 ; i++) {
281 ctx->dram_default_hops[i] = alloc_hop(ctx);
282 if (ctx->dram_default_hops[i] == ULLONG_MAX) {
283 dev_err(hdev->dev, "failed to alloc hop 3, i: %d\n", i);
290 /* need only pte 0 in hops 0 and 1 */
291 pte_val = (hop1_addr & PTE_PHYS_ADDR_MASK) | PAGE_PRESENT_MASK;
292 write_pte(ctx, hop0_addr, pte_val);
294 pte_val = (hop2_addr & PTE_PHYS_ADDR_MASK) | PAGE_PRESENT_MASK;
295 write_pte(ctx, hop1_addr, pte_val);
296 get_pte(ctx, hop1_addr);
298 hop2_pte_addr = hop2_addr;
299 for (i = 0 ; i < num_of_hop3 ; i++) {
300 pte_val = (ctx->dram_default_hops[i] & PTE_PHYS_ADDR_MASK) |
302 write_pte(ctx, hop2_pte_addr, pte_val);
303 get_pte(ctx, hop2_addr);
304 hop2_pte_addr += HL_PTE_SIZE;
307 pte_val = (prop->mmu_dram_default_page_addr & PTE_PHYS_ADDR_MASK) |
308 LAST_MASK | PAGE_PRESENT_MASK;
310 for (i = 0 ; i < num_of_hop3 ; i++) {
311 hop3_pte_addr = ctx->dram_default_hops[i];
312 for (j = 0 ; j < PTE_ENTRIES_IN_HOP ; j++) {
313 write_final_pte(ctx, hop3_pte_addr, pte_val);
314 get_pte(ctx, ctx->dram_default_hops[i]);
315 hop3_pte_addr += HL_PTE_SIZE;
324 for (i = 0 ; i < hop3_allocated ; i++)
325 free_hop(ctx, ctx->dram_default_hops[i]);
327 free_hop(ctx, hop2_addr);
329 free_hop(ctx, hop1_addr);
331 kfree(ctx->dram_default_hops);
336 static void dram_default_mapping_fini(struct hl_ctx *ctx)
338 struct hl_device *hdev = ctx->hdev;
339 struct asic_fixed_properties *prop = &hdev->asic_prop;
340 u64 num_of_hop3, total_hops, hop0_addr, hop1_addr, hop2_addr,
341 hop2_pte_addr, hop3_pte_addr;
344 if ((!hdev->dram_supports_virtual_memory) ||
345 (!hdev->dram_default_page_mapping) ||
346 (ctx->asid == HL_KERNEL_ASID_ID))
349 num_of_hop3 = prop->dram_size_for_default_page_mapping;
350 do_div(num_of_hop3, prop->dram_page_size);
351 do_div(num_of_hop3, PTE_ENTRIES_IN_HOP);
353 hop0_addr = get_hop0_addr(ctx);
354 /* add hop1 and hop2 */
355 total_hops = num_of_hop3 + 2;
356 hop1_addr = ctx->dram_default_hops[total_hops - 1];
357 hop2_addr = ctx->dram_default_hops[total_hops - 2];
359 for (i = 0 ; i < num_of_hop3 ; i++) {
360 hop3_pte_addr = ctx->dram_default_hops[i];
361 for (j = 0 ; j < PTE_ENTRIES_IN_HOP ; j++) {
362 clear_pte(ctx, hop3_pte_addr);
363 put_pte(ctx, ctx->dram_default_hops[i]);
364 hop3_pte_addr += HL_PTE_SIZE;
368 hop2_pte_addr = hop2_addr;
369 hop2_pte_addr = hop2_addr;
370 for (i = 0 ; i < num_of_hop3 ; i++) {
371 clear_pte(ctx, hop2_pte_addr);
372 put_pte(ctx, hop2_addr);
373 hop2_pte_addr += HL_PTE_SIZE;
376 clear_pte(ctx, hop1_addr);
377 put_pte(ctx, hop1_addr);
378 clear_pte(ctx, hop0_addr);
380 kfree(ctx->dram_default_hops);
386 * hl_mmu_init() - initialize the MMU module.
387 * @hdev: habanalabs device structure.
389 * This function does the following:
390 * - Create a pool of pages for pgt_infos.
391 * - Create a shadow table for pgt
393 * Return: 0 for success, non-zero for failure.
395 int hl_mmu_init(struct hl_device *hdev)
397 struct asic_fixed_properties *prop = &hdev->asic_prop;
400 if (!hdev->mmu_enable)
403 /* MMU H/W init was already done in device hw_init() */
406 gen_pool_create(__ffs(prop->mmu_hop_table_size), -1);
408 if (!hdev->mmu_pgt_pool) {
409 dev_err(hdev->dev, "Failed to create page gen pool\n");
413 rc = gen_pool_add(hdev->mmu_pgt_pool, prop->mmu_pgt_addr +
414 prop->mmu_hop0_tables_total_size,
415 prop->mmu_pgt_size - prop->mmu_hop0_tables_total_size,
418 dev_err(hdev->dev, "Failed to add memory to page gen pool\n");
422 hdev->mmu_shadow_hop0 = kvmalloc_array(prop->max_asid,
423 prop->mmu_hop_table_size,
424 GFP_KERNEL | __GFP_ZERO);
425 if (!hdev->mmu_shadow_hop0) {
433 gen_pool_destroy(hdev->mmu_pgt_pool);
439 * hl_mmu_fini() - release the MMU module.
440 * @hdev: habanalabs device structure.
442 * This function does the following:
443 * - Disable MMU in H/W.
444 * - Free the pgt_infos pool.
446 * All contexts should be freed before calling this function.
448 void hl_mmu_fini(struct hl_device *hdev)
450 if (!hdev->mmu_enable)
453 kvfree(hdev->mmu_shadow_hop0);
454 gen_pool_destroy(hdev->mmu_pgt_pool);
456 /* MMU H/W fini will be done in device hw_fini() */
460 * hl_mmu_ctx_init() - initialize a context for using the MMU module.
461 * @ctx: pointer to the context structure to initialize.
463 * Initialize a mutex to protect the concurrent mapping flow, a hash to hold all
464 * page tables hops related to this context.
465 * Return: 0 on success, non-zero otherwise.
467 int hl_mmu_ctx_init(struct hl_ctx *ctx)
469 struct hl_device *hdev = ctx->hdev;
471 if (!hdev->mmu_enable)
474 mutex_init(&ctx->mmu_lock);
475 hash_init(ctx->mmu_phys_hash);
476 hash_init(ctx->mmu_shadow_hash);
478 return dram_default_mapping_init(ctx);
482 * hl_mmu_ctx_fini - disable a ctx from using the mmu module
484 * @ctx: pointer to the context structure
486 * This function does the following:
487 * - Free any pgts which were not freed yet
489 * - Free DRAM default page mapping hops
491 void hl_mmu_ctx_fini(struct hl_ctx *ctx)
493 struct hl_device *hdev = ctx->hdev;
494 struct pgt_info *pgt_info;
495 struct hlist_node *tmp;
498 if (!hdev->mmu_enable)
501 dram_default_mapping_fini(ctx);
503 if (!hash_empty(ctx->mmu_shadow_hash))
504 dev_err(hdev->dev, "ctx is freed while it has pgts in use\n");
506 hash_for_each_safe(ctx->mmu_shadow_hash, i, tmp, pgt_info, node) {
508 "pgt_info of addr 0x%llx of asid %d was not destroyed, num_ptes: %d\n",
509 pgt_info->phys_addr, ctx->asid, pgt_info->num_of_ptes);
510 free_hop(ctx, pgt_info->shadow_addr);
513 mutex_destroy(&ctx->mmu_lock);
516 static int _hl_mmu_unmap(struct hl_ctx *ctx, u64 virt_addr)
518 struct hl_device *hdev = ctx->hdev;
519 struct asic_fixed_properties *prop = &hdev->asic_prop;
520 u64 hop0_addr = 0, hop0_pte_addr = 0,
521 hop1_addr = 0, hop1_pte_addr = 0,
522 hop2_addr = 0, hop2_pte_addr = 0,
523 hop3_addr = 0, hop3_pte_addr = 0,
524 hop4_addr = 0, hop4_pte_addr = 0,
526 bool is_dram_addr, is_huge, clear_hop3 = true;
528 is_dram_addr = hl_mem_area_inside_range(virt_addr, PAGE_SIZE_2MB,
529 prop->va_space_dram_start_address,
530 prop->va_space_dram_end_address);
532 hop0_addr = get_hop0_addr(ctx);
533 hop0_pte_addr = get_hop0_pte_addr(ctx, hop0_addr, virt_addr);
535 curr_pte = *(u64 *) (uintptr_t) hop0_pte_addr;
537 hop1_addr = get_next_hop_addr(ctx, curr_pte);
539 if (hop1_addr == ULLONG_MAX)
542 hop1_pte_addr = get_hop1_pte_addr(ctx, hop1_addr, virt_addr);
544 curr_pte = *(u64 *) (uintptr_t) hop1_pte_addr;
546 hop2_addr = get_next_hop_addr(ctx, curr_pte);
548 if (hop2_addr == ULLONG_MAX)
551 hop2_pte_addr = get_hop2_pte_addr(ctx, hop2_addr, virt_addr);
553 curr_pte = *(u64 *) (uintptr_t) hop2_pte_addr;
555 hop3_addr = get_next_hop_addr(ctx, curr_pte);
557 if (hop3_addr == ULLONG_MAX)
560 hop3_pte_addr = get_hop3_pte_addr(ctx, hop3_addr, virt_addr);
562 curr_pte = *(u64 *) (uintptr_t) hop3_pte_addr;
564 is_huge = curr_pte & LAST_MASK;
566 if (is_dram_addr && !is_huge) {
568 "DRAM unmapping should use huge pages only\n");
573 hop4_addr = get_next_hop_addr(ctx, curr_pte);
575 if (hop4_addr == ULLONG_MAX)
578 hop4_pte_addr = get_hop4_pte_addr(ctx, hop4_addr, virt_addr);
580 curr_pte = *(u64 *) (uintptr_t) hop4_pte_addr;
585 if (hdev->dram_default_page_mapping && is_dram_addr) {
586 u64 default_pte = (prop->mmu_dram_default_page_addr &
587 PTE_PHYS_ADDR_MASK) | LAST_MASK |
589 if (curr_pte == default_pte) {
591 "DRAM: hop3 PTE points to zero page, can't unmap, va: 0x%llx\n",
596 if (!(curr_pte & PAGE_PRESENT_MASK)) {
598 "DRAM: hop3 PTE is cleared! can't unmap, va: 0x%llx\n",
603 write_final_pte(ctx, hop3_pte_addr, default_pte);
604 put_pte(ctx, hop3_addr);
606 if (!(curr_pte & PAGE_PRESENT_MASK))
610 clear_pte(ctx, hop4_pte_addr);
612 clear_pte(ctx, hop3_pte_addr);
614 if (hop4_addr && !put_pte(ctx, hop4_addr))
620 clear_pte(ctx, hop3_pte_addr);
622 if (put_pte(ctx, hop3_addr))
625 clear_pte(ctx, hop2_pte_addr);
627 if (put_pte(ctx, hop2_addr))
630 clear_pte(ctx, hop1_pte_addr);
632 if (put_pte(ctx, hop1_addr))
635 clear_pte(ctx, hop0_pte_addr);
644 dev_err(hdev->dev, "virt addr 0x%llx is not mapped to phys addr\n",
651 * hl_mmu_unmap - unmaps a virtual addr
653 * @ctx: pointer to the context structure
654 * @virt_addr: virt addr to map from
655 * @page_size: size of the page to unmap
657 * This function does the following:
658 * - Check that the virt addr is mapped
659 * - Unmap the virt addr and frees pgts if possible
660 * - Returns 0 on success, -EINVAL if the given addr is not mapped
662 * Because this function changes the page tables in the device and because it
663 * changes the MMU hash, it must be protected by a lock.
664 * However, because it maps only a single page, the lock should be implemented
665 * in a higher level in order to protect the entire mapping of the memory area
667 int hl_mmu_unmap(struct hl_ctx *ctx, u64 virt_addr, u32 page_size)
669 struct hl_device *hdev = ctx->hdev;
671 u32 real_page_size, npages;
674 if (!hdev->mmu_enable)
678 * The H/W handles mapping of 4KB/2MB page. Hence if the host page size
679 * is bigger, we break it to sub-pages and unmap them separately.
681 if ((page_size % PAGE_SIZE_2MB) == 0) {
682 real_page_size = PAGE_SIZE_2MB;
683 } else if ((page_size % PAGE_SIZE_4KB) == 0) {
684 real_page_size = PAGE_SIZE_4KB;
687 "page size of %u is not 4KB nor 2MB aligned, can't unmap\n",
693 npages = page_size / real_page_size;
694 real_virt_addr = virt_addr;
696 for (i = 0 ; i < npages ; i++) {
697 rc = _hl_mmu_unmap(ctx, real_virt_addr);
701 real_virt_addr += real_page_size;
707 static int _hl_mmu_map(struct hl_ctx *ctx, u64 virt_addr, u64 phys_addr,
710 struct hl_device *hdev = ctx->hdev;
711 struct asic_fixed_properties *prop = &hdev->asic_prop;
712 u64 hop0_addr = 0, hop0_pte_addr = 0,
713 hop1_addr = 0, hop1_pte_addr = 0,
714 hop2_addr = 0, hop2_pte_addr = 0,
715 hop3_addr = 0, hop3_pte_addr = 0,
716 hop4_addr = 0, hop4_pte_addr = 0,
718 bool hop1_new = false, hop2_new = false, hop3_new = false,
719 hop4_new = false, is_huge, is_dram_addr;
723 * This mapping function can map a 4KB/2MB page. For 2MB page there are
724 * only 3 hops rather than 4. Currently the DRAM allocation uses 2MB
725 * pages only but user memory could have been allocated with one of the
726 * two page sizes. Since this is a common code for all the three cases,
727 * we need this hugs page check.
729 is_huge = page_size == PAGE_SIZE_2MB;
731 is_dram_addr = hl_mem_area_inside_range(virt_addr, page_size,
732 prop->va_space_dram_start_address,
733 prop->va_space_dram_end_address);
735 if (is_dram_addr && !is_huge) {
736 dev_err(hdev->dev, "DRAM mapping should use huge pages only\n");
740 hop0_addr = get_hop0_addr(ctx);
741 hop0_pte_addr = get_hop0_pte_addr(ctx, hop0_addr, virt_addr);
742 curr_pte = *(u64 *) (uintptr_t) hop0_pte_addr;
744 hop1_addr = get_alloc_next_hop_addr(ctx, curr_pte, &hop1_new);
745 if (hop1_addr == ULLONG_MAX)
748 hop1_pte_addr = get_hop1_pte_addr(ctx, hop1_addr, virt_addr);
749 curr_pte = *(u64 *) (uintptr_t) hop1_pte_addr;
751 hop2_addr = get_alloc_next_hop_addr(ctx, curr_pte, &hop2_new);
752 if (hop2_addr == ULLONG_MAX)
755 hop2_pte_addr = get_hop2_pte_addr(ctx, hop2_addr, virt_addr);
756 curr_pte = *(u64 *) (uintptr_t) hop2_pte_addr;
758 hop3_addr = get_alloc_next_hop_addr(ctx, curr_pte, &hop3_new);
759 if (hop3_addr == ULLONG_MAX)
762 hop3_pte_addr = get_hop3_pte_addr(ctx, hop3_addr, virt_addr);
763 curr_pte = *(u64 *) (uintptr_t) hop3_pte_addr;
766 hop4_addr = get_alloc_next_hop_addr(ctx, curr_pte, &hop4_new);
767 if (hop4_addr == ULLONG_MAX)
770 hop4_pte_addr = get_hop4_pte_addr(ctx, hop4_addr, virt_addr);
771 curr_pte = *(u64 *) (uintptr_t) hop4_pte_addr;
774 if (hdev->dram_default_page_mapping && is_dram_addr) {
775 u64 default_pte = (prop->mmu_dram_default_page_addr &
776 PTE_PHYS_ADDR_MASK) | LAST_MASK |
779 if (curr_pte != default_pte) {
781 "DRAM: mapping already exists for virt_addr 0x%llx\n",
787 if (hop1_new || hop2_new || hop3_new || hop4_new) {
789 "DRAM mapping should not allocate more hops\n");
793 } else if (curr_pte & PAGE_PRESENT_MASK) {
795 "mapping already exists for virt_addr 0x%llx\n",
798 dev_dbg(hdev->dev, "hop0 pte: 0x%llx (0x%llx)\n",
799 *(u64 *) (uintptr_t) hop0_pte_addr, hop0_pte_addr);
800 dev_dbg(hdev->dev, "hop1 pte: 0x%llx (0x%llx)\n",
801 *(u64 *) (uintptr_t) hop1_pte_addr, hop1_pte_addr);
802 dev_dbg(hdev->dev, "hop2 pte: 0x%llx (0x%llx)\n",
803 *(u64 *) (uintptr_t) hop2_pte_addr, hop2_pte_addr);
804 dev_dbg(hdev->dev, "hop3 pte: 0x%llx (0x%llx)\n",
805 *(u64 *) (uintptr_t) hop3_pte_addr, hop3_pte_addr);
808 dev_dbg(hdev->dev, "hop4 pte: 0x%llx (0x%llx)\n",
809 *(u64 *) (uintptr_t) hop4_pte_addr,
816 curr_pte = (phys_addr & PTE_PHYS_ADDR_MASK) | LAST_MASK
820 write_final_pte(ctx, hop3_pte_addr, curr_pte);
822 write_final_pte(ctx, hop4_pte_addr, curr_pte);
826 (hop1_addr & PTE_PHYS_ADDR_MASK) | PAGE_PRESENT_MASK;
827 write_pte(ctx, hop0_pte_addr, curr_pte);
831 (hop2_addr & PTE_PHYS_ADDR_MASK) | PAGE_PRESENT_MASK;
832 write_pte(ctx, hop1_pte_addr, curr_pte);
833 get_pte(ctx, hop1_addr);
837 (hop3_addr & PTE_PHYS_ADDR_MASK) | PAGE_PRESENT_MASK;
838 write_pte(ctx, hop2_pte_addr, curr_pte);
839 get_pte(ctx, hop2_addr);
844 curr_pte = (hop4_addr & PTE_PHYS_ADDR_MASK) |
846 write_pte(ctx, hop3_pte_addr, curr_pte);
847 get_pte(ctx, hop3_addr);
850 get_pte(ctx, hop4_addr);
852 get_pte(ctx, hop3_addr);
861 free_hop(ctx, hop4_addr);
863 free_hop(ctx, hop3_addr);
865 free_hop(ctx, hop2_addr);
867 free_hop(ctx, hop1_addr);
873 * hl_mmu_map - maps a virtual addr to physical addr
875 * @ctx: pointer to the context structure
876 * @virt_addr: virt addr to map from
877 * @phys_addr: phys addr to map to
878 * @page_size: physical page size
880 * This function does the following:
881 * - Check that the virt addr is not mapped
882 * - Allocate pgts as necessary in order to map the virt addr to the phys
883 * - Returns 0 on success, -EINVAL if addr is already mapped, or -ENOMEM.
885 * Because this function changes the page tables in the device and because it
886 * changes the MMU hash, it must be protected by a lock.
887 * However, because it maps only a single page, the lock should be implemented
888 * in a higher level in order to protect the entire mapping of the memory area
890 int hl_mmu_map(struct hl_ctx *ctx, u64 virt_addr, u64 phys_addr, u32 page_size)
892 struct hl_device *hdev = ctx->hdev;
893 u64 real_virt_addr, real_phys_addr;
894 u32 real_page_size, npages;
895 int i, rc, mapped_cnt = 0;
897 if (!hdev->mmu_enable)
901 * The H/W handles mapping of 4KB/2MB page. Hence if the host page size
902 * is bigger, we break it to sub-pages and map them separately.
904 if ((page_size % PAGE_SIZE_2MB) == 0) {
905 real_page_size = PAGE_SIZE_2MB;
906 } else if ((page_size % PAGE_SIZE_4KB) == 0) {
907 real_page_size = PAGE_SIZE_4KB;
910 "page size of %u is not 4KB nor 2MB aligned, can't map\n",
916 WARN_ONCE((phys_addr & (real_page_size - 1)),
917 "Mapping 0x%llx with page size of 0x%x is erroneous! Address must be divisible by page size",
918 phys_addr, real_page_size);
920 npages = page_size / real_page_size;
921 real_virt_addr = virt_addr;
922 real_phys_addr = phys_addr;
924 for (i = 0 ; i < npages ; i++) {
925 rc = _hl_mmu_map(ctx, real_virt_addr, real_phys_addr,
930 real_virt_addr += real_page_size;
931 real_phys_addr += real_page_size;
938 real_virt_addr = virt_addr;
939 for (i = 0 ; i < mapped_cnt ; i++) {
940 if (_hl_mmu_unmap(ctx, real_virt_addr))
941 dev_warn_ratelimited(hdev->dev,
942 "failed to unmap va: 0x%llx\n", real_virt_addr);
944 real_virt_addr += real_page_size;
951 * hl_mmu_swap_out - marks all mapping of the given ctx as swapped out
953 * @ctx: pointer to the context structure
956 void hl_mmu_swap_out(struct hl_ctx *ctx)
962 * hl_mmu_swap_in - marks all mapping of the given ctx as swapped in
964 * @ctx: pointer to the context structure
967 void hl_mmu_swap_in(struct hl_ctx *ctx)