3 * (C) COPYRIGHT 2010-2016 ARM Limited. All rights reserved.
5 * This program is free software and is provided to you under the terms of the
6 * GNU General Public License version 2 as published by the Free Software
7 * Foundation, and any use by you of this program is subject to the terms
10 * A copy of the licence is included with the program, and can also be obtained
11 * from Free Software Foundation, Inc., 51 Franklin Street, Fifth Floor,
12 * Boston, MA 02110-1301, USA.
21 * @file mali_kbase_mem.c
22 * Base kernel memory APIs
24 #ifdef CONFIG_DMA_SHARED_BUFFER
25 #include <linux/dma-buf.h>
26 #endif /* CONFIG_DMA_SHARED_BUFFER */
28 #include <linux/ump.h>
29 #endif /* CONFIG_UMP */
30 #include <linux/kernel.h>
31 #include <linux/bug.h>
32 #include <linux/compat.h>
33 #include <linux/version.h>
35 #include <mali_kbase_config.h>
36 #include <mali_kbase.h>
37 #include <mali_midg_regmap.h>
38 #include <mali_kbase_cache_policy.h>
39 #include <mali_kbase_hw.h>
40 #include <mali_kbase_hwaccess_time.h>
41 #include <mali_kbase_tlstream.h>
44 * @brief Check the zone compatibility of two regions.
46 static int kbase_region_tracker_match_zone(struct kbase_va_region *reg1,
47 struct kbase_va_region *reg2)
49 return ((reg1->flags & KBASE_REG_ZONE_MASK) ==
50 (reg2->flags & KBASE_REG_ZONE_MASK));
53 KBASE_EXPORT_TEST_API(kbase_region_tracker_match_zone);
55 /* This function inserts a region into the tree. */
56 static void kbase_region_tracker_insert(struct kbase_context *kctx, struct kbase_va_region *new_reg)
58 u64 start_pfn = new_reg->start_pfn;
59 struct rb_node **link = &(kctx->reg_rbtree.rb_node);
60 struct rb_node *parent = NULL;
62 /* Find the right place in the tree using tree search */
64 struct kbase_va_region *old_reg;
67 old_reg = rb_entry(parent, struct kbase_va_region, rblink);
69 /* RBTree requires no duplicate entries. */
70 KBASE_DEBUG_ASSERT(old_reg->start_pfn != start_pfn);
72 if (old_reg->start_pfn > start_pfn)
73 link = &(*link)->rb_left;
75 link = &(*link)->rb_right;
78 /* Put the new node there, and rebalance tree */
79 rb_link_node(&(new_reg->rblink), parent, link);
80 rb_insert_color(&(new_reg->rblink), &(kctx->reg_rbtree));
83 /* Find allocated region enclosing free range. */
84 static struct kbase_va_region *kbase_region_tracker_find_region_enclosing_range_free(
85 struct kbase_context *kctx, u64 start_pfn, size_t nr_pages)
87 struct rb_node *rbnode;
88 struct kbase_va_region *reg;
90 u64 end_pfn = start_pfn + nr_pages;
92 rbnode = kctx->reg_rbtree.rb_node;
94 u64 tmp_start_pfn, tmp_end_pfn;
96 reg = rb_entry(rbnode, struct kbase_va_region, rblink);
97 tmp_start_pfn = reg->start_pfn;
98 tmp_end_pfn = reg->start_pfn + reg->nr_pages;
100 /* If start is lower than this, go left. */
101 if (start_pfn < tmp_start_pfn)
102 rbnode = rbnode->rb_left;
103 /* If end is higher than this, then go right. */
104 else if (end_pfn > tmp_end_pfn)
105 rbnode = rbnode->rb_right;
113 /* Find region enclosing given address. */
114 struct kbase_va_region *kbase_region_tracker_find_region_enclosing_address(struct kbase_context *kctx, u64 gpu_addr)
116 struct rb_node *rbnode;
117 struct kbase_va_region *reg;
118 u64 gpu_pfn = gpu_addr >> PAGE_SHIFT;
120 KBASE_DEBUG_ASSERT(NULL != kctx);
122 lockdep_assert_held(&kctx->reg_lock);
124 rbnode = kctx->reg_rbtree.rb_node;
126 u64 tmp_start_pfn, tmp_end_pfn;
128 reg = rb_entry(rbnode, struct kbase_va_region, rblink);
129 tmp_start_pfn = reg->start_pfn;
130 tmp_end_pfn = reg->start_pfn + reg->nr_pages;
132 /* If start is lower than this, go left. */
133 if (gpu_pfn < tmp_start_pfn)
134 rbnode = rbnode->rb_left;
135 /* If end is higher than this, then go right. */
136 else if (gpu_pfn >= tmp_end_pfn)
137 rbnode = rbnode->rb_right;
145 KBASE_EXPORT_TEST_API(kbase_region_tracker_find_region_enclosing_address);
147 /* Find region with given base address */
148 struct kbase_va_region *kbase_region_tracker_find_region_base_address(struct kbase_context *kctx, u64 gpu_addr)
150 u64 gpu_pfn = gpu_addr >> PAGE_SHIFT;
151 struct rb_node *rbnode;
152 struct kbase_va_region *reg;
154 KBASE_DEBUG_ASSERT(NULL != kctx);
156 lockdep_assert_held(&kctx->reg_lock);
158 rbnode = kctx->reg_rbtree.rb_node;
160 reg = rb_entry(rbnode, struct kbase_va_region, rblink);
161 if (reg->start_pfn > gpu_pfn)
162 rbnode = rbnode->rb_left;
163 else if (reg->start_pfn < gpu_pfn)
164 rbnode = rbnode->rb_right;
173 KBASE_EXPORT_TEST_API(kbase_region_tracker_find_region_base_address);
175 /* Find region meeting given requirements */
176 static struct kbase_va_region *kbase_region_tracker_find_region_meeting_reqs(struct kbase_context *kctx, struct kbase_va_region *reg_reqs, size_t nr_pages, size_t align)
178 struct rb_node *rbnode;
179 struct kbase_va_region *reg;
181 /* Note that this search is a linear search, as we do not have a target
182 address in mind, so does not benefit from the rbtree search */
183 rbnode = rb_first(&(kctx->reg_rbtree));
185 reg = rb_entry(rbnode, struct kbase_va_region, rblink);
186 if ((reg->nr_pages >= nr_pages) &&
187 (reg->flags & KBASE_REG_FREE) &&
188 kbase_region_tracker_match_zone(reg, reg_reqs)) {
189 /* Check alignment */
190 u64 start_pfn = (reg->start_pfn + align - 1) & ~(align - 1);
192 if ((start_pfn >= reg->start_pfn) &&
193 (start_pfn <= (reg->start_pfn + reg->nr_pages - 1)) &&
194 ((start_pfn + nr_pages - 1) <= (reg->start_pfn + reg->nr_pages - 1)))
197 rbnode = rb_next(rbnode);
204 * @brief Remove a region object from the global list.
206 * The region reg is removed, possibly by merging with other free and
207 * compatible adjacent regions. It must be called with the context
208 * region lock held. The associated memory is not released (see
209 * kbase_free_alloced_region). Internal use only.
211 static int kbase_remove_va_region(struct kbase_context *kctx, struct kbase_va_region *reg)
213 struct rb_node *rbprev;
214 struct kbase_va_region *prev = NULL;
215 struct rb_node *rbnext;
216 struct kbase_va_region *next = NULL;
218 int merged_front = 0;
222 /* Try to merge with the previous block first */
223 rbprev = rb_prev(&(reg->rblink));
225 prev = rb_entry(rbprev, struct kbase_va_region, rblink);
226 if ((prev->flags & KBASE_REG_FREE) && kbase_region_tracker_match_zone(prev, reg)) {
227 /* We're compatible with the previous VMA, merge with it */
228 prev->nr_pages += reg->nr_pages;
229 rb_erase(&(reg->rblink), &kctx->reg_rbtree);
235 /* Try to merge with the next block second */
236 /* Note we do the lookup here as the tree may have been rebalanced. */
237 rbnext = rb_next(&(reg->rblink));
239 /* We're compatible with the next VMA, merge with it */
240 next = rb_entry(rbnext, struct kbase_va_region, rblink);
241 if ((next->flags & KBASE_REG_FREE) && kbase_region_tracker_match_zone(next, reg)) {
242 next->start_pfn = reg->start_pfn;
243 next->nr_pages += reg->nr_pages;
244 rb_erase(&(reg->rblink), &kctx->reg_rbtree);
247 /* We already merged with prev, free it */
248 kbase_free_alloced_region(reg);
253 /* If we failed to merge then we need to add a new block */
254 if (!(merged_front || merged_back)) {
256 * We didn't merge anything. Add a new free
257 * placeholder and remove the original one.
259 struct kbase_va_region *free_reg;
261 free_reg = kbase_alloc_free_region(kctx, reg->start_pfn, reg->nr_pages, reg->flags & KBASE_REG_ZONE_MASK);
267 rb_replace_node(&(reg->rblink), &(free_reg->rblink), &(kctx->reg_rbtree));
274 KBASE_EXPORT_TEST_API(kbase_remove_va_region);
277 * @brief Insert a VA region to the list, replacing the current at_reg.
279 static int kbase_insert_va_region_nolock(struct kbase_context *kctx, struct kbase_va_region *new_reg, struct kbase_va_region *at_reg, u64 start_pfn, size_t nr_pages)
283 /* Must be a free region */
284 KBASE_DEBUG_ASSERT((at_reg->flags & KBASE_REG_FREE) != 0);
285 /* start_pfn should be contained within at_reg */
286 KBASE_DEBUG_ASSERT((start_pfn >= at_reg->start_pfn) && (start_pfn < at_reg->start_pfn + at_reg->nr_pages));
287 /* at least nr_pages from start_pfn should be contained within at_reg */
288 KBASE_DEBUG_ASSERT(start_pfn + nr_pages <= at_reg->start_pfn + at_reg->nr_pages);
290 new_reg->start_pfn = start_pfn;
291 new_reg->nr_pages = nr_pages;
293 /* Regions are a whole use, so swap and delete old one. */
294 if (at_reg->start_pfn == start_pfn && at_reg->nr_pages == nr_pages) {
295 rb_replace_node(&(at_reg->rblink), &(new_reg->rblink), &(kctx->reg_rbtree));
296 kbase_free_alloced_region(at_reg);
298 /* New region replaces the start of the old one, so insert before. */
299 else if (at_reg->start_pfn == start_pfn) {
300 at_reg->start_pfn += nr_pages;
301 KBASE_DEBUG_ASSERT(at_reg->nr_pages >= nr_pages);
302 at_reg->nr_pages -= nr_pages;
304 kbase_region_tracker_insert(kctx, new_reg);
306 /* New region replaces the end of the old one, so insert after. */
307 else if ((at_reg->start_pfn + at_reg->nr_pages) == (start_pfn + nr_pages)) {
308 at_reg->nr_pages -= nr_pages;
310 kbase_region_tracker_insert(kctx, new_reg);
312 /* New region splits the old one, so insert and create new */
314 struct kbase_va_region *new_front_reg;
316 new_front_reg = kbase_alloc_free_region(kctx,
318 start_pfn - at_reg->start_pfn,
319 at_reg->flags & KBASE_REG_ZONE_MASK);
322 at_reg->nr_pages -= nr_pages + new_front_reg->nr_pages;
323 at_reg->start_pfn = start_pfn + nr_pages;
325 kbase_region_tracker_insert(kctx, new_front_reg);
326 kbase_region_tracker_insert(kctx, new_reg);
336 * @brief Add a VA region to the list.
338 int kbase_add_va_region(struct kbase_context *kctx,
339 struct kbase_va_region *reg, u64 addr,
340 size_t nr_pages, size_t align)
342 struct kbase_va_region *tmp;
343 u64 gpu_pfn = addr >> PAGE_SHIFT;
346 KBASE_DEBUG_ASSERT(NULL != kctx);
347 KBASE_DEBUG_ASSERT(NULL != reg);
349 lockdep_assert_held(&kctx->reg_lock);
354 /* must be a power of 2 */
355 KBASE_DEBUG_ASSERT((align & (align - 1)) == 0);
356 KBASE_DEBUG_ASSERT(nr_pages > 0);
358 /* Path 1: Map a specific address. Find the enclosing region, which *must* be free. */
360 struct device *dev = kctx->kbdev->dev;
362 KBASE_DEBUG_ASSERT(!(gpu_pfn & (align - 1)));
364 tmp = kbase_region_tracker_find_region_enclosing_range_free(kctx, gpu_pfn, nr_pages);
366 dev_warn(dev, "Enclosing region not found: 0x%08llx gpu_pfn, %zu nr_pages", gpu_pfn, nr_pages);
371 if ((!kbase_region_tracker_match_zone(tmp, reg)) ||
372 (!(tmp->flags & KBASE_REG_FREE))) {
373 dev_warn(dev, "Zone mismatch: %lu != %lu", tmp->flags & KBASE_REG_ZONE_MASK, reg->flags & KBASE_REG_ZONE_MASK);
374 dev_warn(dev, "!(tmp->flags & KBASE_REG_FREE): tmp->start_pfn=0x%llx tmp->flags=0x%lx tmp->nr_pages=0x%zx gpu_pfn=0x%llx nr_pages=0x%zx\n", tmp->start_pfn, tmp->flags, tmp->nr_pages, gpu_pfn, nr_pages);
375 dev_warn(dev, "in function %s (%p, %p, 0x%llx, 0x%zx, 0x%zx)\n", __func__, kctx, reg, addr, nr_pages, align);
380 err = kbase_insert_va_region_nolock(kctx, reg, tmp, gpu_pfn, nr_pages);
382 dev_warn(dev, "Failed to insert va region");
390 /* Path 2: Map any free address which meets the requirements. */
395 * Depending on the zone the allocation request is for
396 * we might need to retry it.
399 tmp = kbase_region_tracker_find_region_meeting_reqs(
400 kctx, reg, nr_pages, align);
402 start_pfn = (tmp->start_pfn + align - 1) &
404 err = kbase_insert_va_region_nolock(kctx, reg,
405 tmp, start_pfn, nr_pages);
410 * If the allocation is not from the same zone as JIT
411 * then don't retry, we're out of VA and there is
412 * nothing which can be done about it.
414 if ((reg->flags & KBASE_REG_ZONE_MASK) !=
415 KBASE_REG_ZONE_CUSTOM_VA)
417 } while (kbase_jit_evict(kctx));
427 KBASE_EXPORT_TEST_API(kbase_add_va_region);
430 * @brief Initialize the internal region tracker data structure.
432 static void kbase_region_tracker_ds_init(struct kbase_context *kctx,
433 struct kbase_va_region *same_va_reg,
434 struct kbase_va_region *exec_reg,
435 struct kbase_va_region *custom_va_reg)
437 kctx->reg_rbtree = RB_ROOT;
438 kbase_region_tracker_insert(kctx, same_va_reg);
440 /* exec and custom_va_reg doesn't always exist */
441 if (exec_reg && custom_va_reg) {
442 kbase_region_tracker_insert(kctx, exec_reg);
443 kbase_region_tracker_insert(kctx, custom_va_reg);
447 void kbase_region_tracker_term(struct kbase_context *kctx)
449 struct rb_node *rbnode;
450 struct kbase_va_region *reg;
453 rbnode = rb_first(&(kctx->reg_rbtree));
455 rb_erase(rbnode, &(kctx->reg_rbtree));
456 reg = rb_entry(rbnode, struct kbase_va_region, rblink);
457 kbase_free_alloced_region(reg);
463 * Initialize the region tracker data structure.
465 int kbase_region_tracker_init(struct kbase_context *kctx)
467 struct kbase_va_region *same_va_reg;
468 struct kbase_va_region *exec_reg = NULL;
469 struct kbase_va_region *custom_va_reg = NULL;
470 size_t same_va_bits = sizeof(void *) * BITS_PER_BYTE;
471 u64 custom_va_size = KBASE_REG_ZONE_CUSTOM_VA_SIZE;
472 u64 gpu_va_limit = (1ULL << kctx->kbdev->gpu_props.mmu.va_bits) >> PAGE_SHIFT;
476 /* Take the lock as kbase_free_alloced_region requires it */
477 kbase_gpu_vm_lock(kctx);
479 #if defined(CONFIG_ARM64)
480 same_va_bits = VA_BITS;
481 #elif defined(CONFIG_X86_64)
483 #elif defined(CONFIG_64BIT)
484 #error Unsupported 64-bit architecture
490 else if (kbase_hw_has_feature(kctx->kbdev, BASE_HW_FEATURE_33BIT_VA))
494 if (kctx->kbdev->gpu_props.mmu.va_bits < same_va_bits) {
499 same_va_pages = (1ULL << (same_va_bits - PAGE_SHIFT)) - 1;
500 /* all have SAME_VA */
501 same_va_reg = kbase_alloc_free_region(kctx, 1,
503 KBASE_REG_ZONE_SAME_VA);
511 /* 32-bit clients have exec and custom VA zones */
512 if (kctx->is_compat) {
514 if (gpu_va_limit <= KBASE_REG_ZONE_CUSTOM_VA_BASE) {
516 goto fail_free_same_va;
518 /* If the current size of TMEM is out of range of the
519 * virtual address space addressable by the MMU then
520 * we should shrink it to fit
522 if ((KBASE_REG_ZONE_CUSTOM_VA_BASE + KBASE_REG_ZONE_CUSTOM_VA_SIZE) >= gpu_va_limit)
523 custom_va_size = gpu_va_limit - KBASE_REG_ZONE_CUSTOM_VA_BASE;
525 exec_reg = kbase_alloc_free_region(kctx,
526 KBASE_REG_ZONE_EXEC_BASE,
527 KBASE_REG_ZONE_EXEC_SIZE,
528 KBASE_REG_ZONE_EXEC);
532 goto fail_free_same_va;
535 custom_va_reg = kbase_alloc_free_region(kctx,
536 KBASE_REG_ZONE_CUSTOM_VA_BASE,
537 custom_va_size, KBASE_REG_ZONE_CUSTOM_VA);
539 if (!custom_va_reg) {
547 kbase_region_tracker_ds_init(kctx, same_va_reg, exec_reg, custom_va_reg);
549 kctx->same_va_end = same_va_pages + 1;
551 kbase_gpu_vm_unlock(kctx);
555 kbase_free_alloced_region(exec_reg);
557 kbase_free_alloced_region(same_va_reg);
559 kbase_gpu_vm_unlock(kctx);
563 int kbase_region_tracker_init_jit(struct kbase_context *kctx, u64 jit_va_pages)
566 struct kbase_va_region *same_va;
567 struct kbase_va_region *custom_va_reg;
573 * Nothing to do for 32-bit clients, JIT uses the existing
579 #if defined(CONFIG_ARM64)
580 same_va_bits = VA_BITS;
581 #elif defined(CONFIG_X86_64)
583 #elif defined(CONFIG_64BIT)
584 #error Unsupported 64-bit architecture
587 if (kbase_hw_has_feature(kctx->kbdev, BASE_HW_FEATURE_33BIT_VA))
590 total_va_size = (1ULL << (same_va_bits - PAGE_SHIFT)) - 1;
592 kbase_gpu_vm_lock(kctx);
595 * Modify the same VA free region after creation. Be careful to ensure
596 * that allocations haven't been made as they could cause an overlap
597 * to happen with existing same VA allocations and the custom VA zone.
599 same_va = kbase_region_tracker_find_region_base_address(kctx,
606 /* The region flag or region size has changed since creation so bail. */
607 if ((!(same_va->flags & KBASE_REG_FREE)) ||
608 (same_va->nr_pages != total_va_size)) {
613 if (same_va->nr_pages < jit_va_pages ||
614 kctx->same_va_end < jit_va_pages) {
619 /* It's safe to adjust the same VA zone now */
620 same_va->nr_pages -= jit_va_pages;
621 kctx->same_va_end -= jit_va_pages;
624 * Create a custom VA zone at the end of the VA for allocations which
625 * JIT can use so it doesn't have to allocate VA from the kernel.
627 custom_va_reg = kbase_alloc_free_region(kctx,
630 KBASE_REG_ZONE_CUSTOM_VA);
631 if (!custom_va_reg) {
633 * The context will be destroyed if we fail here so no point
634 * reverting the change we made to same_va.
640 kbase_region_tracker_insert(kctx, custom_va_reg);
642 kbase_gpu_vm_unlock(kctx);
646 kbase_gpu_vm_unlock(kctx);
653 int kbase_mem_init(struct kbase_device *kbdev)
655 struct kbasep_mem_device *memdev;
657 KBASE_DEBUG_ASSERT(kbdev);
659 memdev = &kbdev->memdev;
660 kbdev->mem_pool_max_size_default = KBASE_MEM_POOL_MAX_SIZE_KCTX;
662 /* Initialize memory usage */
663 atomic_set(&memdev->used_pages, 0);
665 return kbase_mem_pool_init(&kbdev->mem_pool,
666 KBASE_MEM_POOL_MAX_SIZE_KBDEV, kbdev, NULL);
669 void kbase_mem_halt(struct kbase_device *kbdev)
674 void kbase_mem_term(struct kbase_device *kbdev)
676 struct kbasep_mem_device *memdev;
679 KBASE_DEBUG_ASSERT(kbdev);
681 memdev = &kbdev->memdev;
683 pages = atomic_read(&memdev->used_pages);
685 dev_warn(kbdev->dev, "%s: %d pages in use!\n", __func__, pages);
687 kbase_mem_pool_term(&kbdev->mem_pool);
690 KBASE_EXPORT_TEST_API(kbase_mem_term);
696 * @brief Allocate a free region object.
698 * The allocated object is not part of any list yet, and is flagged as
699 * KBASE_REG_FREE. No mapping is allocated yet.
701 * zone is KBASE_REG_ZONE_CUSTOM_VA, KBASE_REG_ZONE_SAME_VA, or KBASE_REG_ZONE_EXEC
704 struct kbase_va_region *kbase_alloc_free_region(struct kbase_context *kctx, u64 start_pfn, size_t nr_pages, int zone)
706 struct kbase_va_region *new_reg;
708 KBASE_DEBUG_ASSERT(kctx != NULL);
710 /* zone argument should only contain zone related region flags */
711 KBASE_DEBUG_ASSERT((zone & ~KBASE_REG_ZONE_MASK) == 0);
712 KBASE_DEBUG_ASSERT(nr_pages > 0);
713 /* 64-bit address range is the max */
714 KBASE_DEBUG_ASSERT(start_pfn + nr_pages <= (U64_MAX / PAGE_SIZE));
716 new_reg = kzalloc(sizeof(*new_reg), GFP_KERNEL);
721 new_reg->cpu_alloc = NULL; /* no alloc bound yet */
722 new_reg->gpu_alloc = NULL; /* no alloc bound yet */
723 new_reg->kctx = kctx;
724 new_reg->flags = zone | KBASE_REG_FREE;
726 new_reg->flags |= KBASE_REG_GROWABLE;
728 new_reg->start_pfn = start_pfn;
729 new_reg->nr_pages = nr_pages;
734 KBASE_EXPORT_TEST_API(kbase_alloc_free_region);
737 * @brief Free a region object.
739 * The described region must be freed of any mapping.
741 * If the region is not flagged as KBASE_REG_FREE, the region's
742 * alloc object will be released.
743 * It is a bug if no alloc object exists for non-free regions.
746 void kbase_free_alloced_region(struct kbase_va_region *reg)
748 if (!(reg->flags & KBASE_REG_FREE)) {
750 * The physical allocation should have been removed from the
751 * eviction list before this function is called. However, in the
752 * case of abnormal process termination or the app leaking the
753 * memory kbase_mem_free_region is not called so it can still be
754 * on the list at termination time of the region tracker.
756 if (!list_empty(®->gpu_alloc->evict_node)) {
758 * Unlink the physical allocation before unmaking it
759 * evictable so that the allocation isn't grown back to
760 * its last backed size as we're going to unmap it
763 reg->cpu_alloc->reg = NULL;
764 if (reg->cpu_alloc != reg->gpu_alloc)
765 reg->gpu_alloc->reg = NULL;
768 * If a region has been made evictable then we must
769 * unmake it before trying to free it.
770 * If the memory hasn't been reclaimed it will be
771 * unmapped and freed below, if it has been reclaimed
772 * then the operations below are no-ops.
774 if (reg->flags & KBASE_REG_DONT_NEED) {
775 KBASE_DEBUG_ASSERT(reg->cpu_alloc->type ==
776 KBASE_MEM_TYPE_NATIVE);
777 kbase_mem_evictable_unmake(reg->gpu_alloc);
782 * Remove the region from the sticky resource metadata
783 * list should it be there.
785 kbase_sticky_resource_release(reg->kctx, NULL,
786 reg->start_pfn << PAGE_SHIFT);
788 kbase_mem_phy_alloc_put(reg->cpu_alloc);
789 kbase_mem_phy_alloc_put(reg->gpu_alloc);
790 /* To detect use-after-free in debug builds */
791 KBASE_DEBUG_CODE(reg->flags |= KBASE_REG_FREE);
796 KBASE_EXPORT_TEST_API(kbase_free_alloced_region);
798 int kbase_gpu_mmap(struct kbase_context *kctx, struct kbase_va_region *reg, u64 addr, size_t nr_pages, size_t align)
803 unsigned long mask = ~KBASE_REG_MEMATTR_MASK;
805 if ((kctx->kbdev->system_coherency == COHERENCY_ACE) &&
806 (reg->flags & KBASE_REG_SHARE_BOTH))
807 attr = KBASE_REG_MEMATTR_INDEX(AS_MEMATTR_INDEX_OUTER_WA);
809 attr = KBASE_REG_MEMATTR_INDEX(AS_MEMATTR_INDEX_WRITE_ALLOC);
811 KBASE_DEBUG_ASSERT(NULL != kctx);
812 KBASE_DEBUG_ASSERT(NULL != reg);
814 err = kbase_add_va_region(kctx, reg, addr, nr_pages, align);
818 if (reg->gpu_alloc->type == KBASE_MEM_TYPE_ALIAS) {
820 struct kbase_mem_phy_alloc *alloc;
822 alloc = reg->gpu_alloc;
823 stride = alloc->imported.alias.stride;
824 KBASE_DEBUG_ASSERT(alloc->imported.alias.aliased);
825 for (i = 0; i < alloc->imported.alias.nents; i++) {
826 if (alloc->imported.alias.aliased[i].alloc) {
827 err = kbase_mmu_insert_pages(kctx,
828 reg->start_pfn + (i * stride),
829 alloc->imported.alias.aliased[i].alloc->pages + alloc->imported.alias.aliased[i].offset,
830 alloc->imported.alias.aliased[i].length,
835 kbase_mem_phy_alloc_gpu_mapped(alloc->imported.alias.aliased[i].alloc);
837 err = kbase_mmu_insert_single_page(kctx,
838 reg->start_pfn + i * stride,
839 page_to_phys(kctx->aliasing_sink_page),
840 alloc->imported.alias.aliased[i].length,
841 (reg->flags & mask) | attr);
848 err = kbase_mmu_insert_pages(kctx, reg->start_pfn,
849 kbase_get_gpu_phy_pages(reg),
850 kbase_reg_current_backed_size(reg),
854 kbase_mem_phy_alloc_gpu_mapped(reg->gpu_alloc);
860 if (reg->gpu_alloc->type == KBASE_MEM_TYPE_ALIAS) {
863 stride = reg->gpu_alloc->imported.alias.stride;
864 KBASE_DEBUG_ASSERT(reg->gpu_alloc->imported.alias.aliased);
866 if (reg->gpu_alloc->imported.alias.aliased[i].alloc) {
867 kbase_mmu_teardown_pages(kctx, reg->start_pfn + (i * stride), reg->gpu_alloc->imported.alias.aliased[i].length);
868 kbase_mem_phy_alloc_gpu_unmapped(reg->gpu_alloc->imported.alias.aliased[i].alloc);
872 kbase_remove_va_region(kctx, reg);
877 KBASE_EXPORT_TEST_API(kbase_gpu_mmap);
879 int kbase_gpu_munmap(struct kbase_context *kctx, struct kbase_va_region *reg)
883 if (reg->start_pfn == 0)
886 if (reg->gpu_alloc && reg->gpu_alloc->type == KBASE_MEM_TYPE_ALIAS) {
889 err = kbase_mmu_teardown_pages(kctx, reg->start_pfn, reg->nr_pages);
890 KBASE_DEBUG_ASSERT(reg->gpu_alloc->imported.alias.aliased);
891 for (i = 0; i < reg->gpu_alloc->imported.alias.nents; i++)
892 if (reg->gpu_alloc->imported.alias.aliased[i].alloc)
893 kbase_mem_phy_alloc_gpu_unmapped(reg->gpu_alloc->imported.alias.aliased[i].alloc);
895 err = kbase_mmu_teardown_pages(kctx, reg->start_pfn, kbase_reg_current_backed_size(reg));
896 kbase_mem_phy_alloc_gpu_unmapped(reg->gpu_alloc);
902 err = kbase_remove_va_region(kctx, reg);
906 static struct kbase_cpu_mapping *kbasep_find_enclosing_cpu_mapping_of_region(const struct kbase_va_region *reg, unsigned long uaddr, size_t size)
908 struct kbase_cpu_mapping *map;
909 struct list_head *pos;
911 KBASE_DEBUG_ASSERT(NULL != reg);
912 KBASE_DEBUG_ASSERT(reg->cpu_alloc);
914 if ((uintptr_t) uaddr + size < (uintptr_t) uaddr) /* overflow check */
917 list_for_each(pos, ®->cpu_alloc->mappings) {
918 map = list_entry(pos, struct kbase_cpu_mapping, mappings_list);
919 if (map->vm_start <= uaddr && map->vm_end >= uaddr + size)
926 KBASE_EXPORT_TEST_API(kbasep_find_enclosing_cpu_mapping_of_region);
928 int kbasep_find_enclosing_cpu_mapping_offset(
929 struct kbase_context *kctx, u64 gpu_addr,
930 unsigned long uaddr, size_t size, u64 * offset)
932 struct kbase_cpu_mapping *map = NULL;
933 const struct kbase_va_region *reg;
936 KBASE_DEBUG_ASSERT(kctx != NULL);
938 kbase_gpu_vm_lock(kctx);
940 reg = kbase_region_tracker_find_region_enclosing_address(kctx, gpu_addr);
941 if (reg && !(reg->flags & KBASE_REG_FREE)) {
942 map = kbasep_find_enclosing_cpu_mapping_of_region(reg, uaddr,
945 *offset = (uaddr - PTR_TO_U64(map->vm_start)) +
946 (map->page_off << PAGE_SHIFT);
951 kbase_gpu_vm_unlock(kctx);
956 KBASE_EXPORT_TEST_API(kbasep_find_enclosing_cpu_mapping_offset);
958 void kbase_sync_single(struct kbase_context *kctx,
959 phys_addr_t cpu_pa, phys_addr_t gpu_pa,
960 off_t offset, size_t size, enum kbase_sync_type sync_fn)
962 struct page *cpu_page;
964 cpu_page = pfn_to_page(PFN_DOWN(cpu_pa));
966 if (likely(cpu_pa == gpu_pa)) {
970 BUG_ON(offset + size > PAGE_SIZE);
972 dma_addr = kbase_dma_addr(cpu_page) + offset;
973 if (sync_fn == KBASE_SYNC_TO_CPU)
974 dma_sync_single_for_cpu(kctx->kbdev->dev, dma_addr,
975 size, DMA_BIDIRECTIONAL);
976 else if (sync_fn == KBASE_SYNC_TO_DEVICE)
977 dma_sync_single_for_device(kctx->kbdev->dev, dma_addr,
978 size, DMA_BIDIRECTIONAL);
982 struct page *gpu_page;
984 if (WARN(!gpu_pa, "No GPU PA found for infinite cache op"))
987 gpu_page = pfn_to_page(PFN_DOWN(gpu_pa));
989 if (sync_fn == KBASE_SYNC_TO_DEVICE) {
990 src = ((unsigned char *)kmap(cpu_page)) + offset;
991 dst = ((unsigned char *)kmap(gpu_page)) + offset;
992 } else if (sync_fn == KBASE_SYNC_TO_CPU) {
993 dma_sync_single_for_cpu(kctx->kbdev->dev,
994 kbase_dma_addr(gpu_page) + offset,
995 size, DMA_BIDIRECTIONAL);
996 src = ((unsigned char *)kmap(gpu_page)) + offset;
997 dst = ((unsigned char *)kmap(cpu_page)) + offset;
999 memcpy(dst, src, size);
1002 if (sync_fn == KBASE_SYNC_TO_DEVICE)
1003 dma_sync_single_for_device(kctx->kbdev->dev,
1004 kbase_dma_addr(gpu_page) + offset,
1005 size, DMA_BIDIRECTIONAL);
1009 static int kbase_do_syncset(struct kbase_context *kctx,
1010 struct base_syncset *set, enum kbase_sync_type sync_fn)
1013 struct basep_syncset *sset = &set->basep_sset;
1014 struct kbase_va_region *reg;
1015 struct kbase_cpu_mapping *map;
1016 unsigned long start;
1018 phys_addr_t *cpu_pa;
1019 phys_addr_t *gpu_pa;
1020 u64 page_off, page_count;
1024 kbase_os_mem_map_lock(kctx);
1025 kbase_gpu_vm_lock(kctx);
1027 /* find the region where the virtual address is contained */
1028 reg = kbase_region_tracker_find_region_enclosing_address(kctx,
1029 sset->mem_handle.basep.handle);
1031 dev_warn(kctx->kbdev->dev, "Can't find region at VA 0x%016llX",
1032 sset->mem_handle.basep.handle);
1037 if (!(reg->flags & KBASE_REG_CPU_CACHED))
1040 start = (uintptr_t)sset->user_addr;
1041 size = (size_t)sset->size;
1043 map = kbasep_find_enclosing_cpu_mapping_of_region(reg, start, size);
1045 dev_warn(kctx->kbdev->dev, "Can't find CPU mapping 0x%016lX for VA 0x%016llX",
1046 start, sset->mem_handle.basep.handle);
1051 offset = start & (PAGE_SIZE - 1);
1052 page_off = map->page_off + ((start - map->vm_start) >> PAGE_SHIFT);
1053 page_count = (size + offset + (PAGE_SIZE - 1)) >> PAGE_SHIFT;
1054 cpu_pa = kbase_get_cpu_phy_pages(reg);
1055 gpu_pa = kbase_get_gpu_phy_pages(reg);
1057 /* Sync first page */
1058 if (cpu_pa[page_off]) {
1059 size_t sz = MIN(((size_t) PAGE_SIZE - offset), size);
1061 kbase_sync_single(kctx, cpu_pa[page_off], gpu_pa[page_off],
1062 offset, sz, sync_fn);
1065 /* Sync middle pages (if any) */
1066 for (i = 1; page_count > 2 && i < page_count - 1; i++) {
1067 /* we grow upwards, so bail on first non-present page */
1068 if (!cpu_pa[page_off + i])
1071 kbase_sync_single(kctx, cpu_pa[page_off + i],
1072 gpu_pa[page_off + i], 0, PAGE_SIZE, sync_fn);
1075 /* Sync last page (if any) */
1076 if (page_count > 1 && cpu_pa[page_off + page_count - 1]) {
1077 size_t sz = ((start + size - 1) & ~PAGE_MASK) + 1;
1079 kbase_sync_single(kctx, cpu_pa[page_off + page_count - 1],
1080 gpu_pa[page_off + page_count - 1], 0, sz,
1085 kbase_gpu_vm_unlock(kctx);
1086 kbase_os_mem_map_unlock(kctx);
1090 int kbase_sync_now(struct kbase_context *kctx, struct base_syncset *syncset)
1093 struct basep_syncset *sset;
1095 KBASE_DEBUG_ASSERT(NULL != kctx);
1096 KBASE_DEBUG_ASSERT(NULL != syncset);
1098 sset = &syncset->basep_sset;
1100 switch (sset->type) {
1101 case BASE_SYNCSET_OP_MSYNC:
1102 err = kbase_do_syncset(kctx, syncset, KBASE_SYNC_TO_DEVICE);
1105 case BASE_SYNCSET_OP_CSYNC:
1106 err = kbase_do_syncset(kctx, syncset, KBASE_SYNC_TO_CPU);
1110 dev_warn(kctx->kbdev->dev, "Unknown msync op %d\n", sset->type);
1117 KBASE_EXPORT_TEST_API(kbase_sync_now);
1119 /* vm lock must be held */
1120 int kbase_mem_free_region(struct kbase_context *kctx, struct kbase_va_region *reg)
1124 KBASE_DEBUG_ASSERT(NULL != kctx);
1125 KBASE_DEBUG_ASSERT(NULL != reg);
1126 lockdep_assert_held(&kctx->reg_lock);
1129 * Unlink the physical allocation before unmaking it evictable so
1130 * that the allocation isn't grown back to its last backed size
1131 * as we're going to unmap it anyway.
1133 reg->cpu_alloc->reg = NULL;
1134 if (reg->cpu_alloc != reg->gpu_alloc)
1135 reg->gpu_alloc->reg = NULL;
1138 * If a region has been made evictable then we must unmake it
1139 * before trying to free it.
1140 * If the memory hasn't been reclaimed it will be unmapped and freed
1141 * below, if it has been reclaimed then the operations below are no-ops.
1143 if (reg->flags & KBASE_REG_DONT_NEED) {
1144 KBASE_DEBUG_ASSERT(reg->cpu_alloc->type ==
1145 KBASE_MEM_TYPE_NATIVE);
1146 kbase_mem_evictable_unmake(reg->gpu_alloc);
1149 err = kbase_gpu_munmap(kctx, reg);
1151 dev_warn(reg->kctx->kbdev->dev, "Could not unmap from the GPU...\n");
1155 /* This will also free the physical pages */
1156 kbase_free_alloced_region(reg);
1162 KBASE_EXPORT_TEST_API(kbase_mem_free_region);
1165 * @brief Free the region from the GPU and unregister it.
1167 * This function implements the free operation on a memory segment.
1168 * It will loudly fail if called with outstanding mappings.
1170 int kbase_mem_free(struct kbase_context *kctx, u64 gpu_addr)
1173 struct kbase_va_region *reg;
1175 KBASE_DEBUG_ASSERT(kctx != NULL);
1177 if (0 == gpu_addr) {
1178 dev_warn(kctx->kbdev->dev, "gpu_addr 0 is reserved for the ringbuffer and it's an error to try to free it using kbase_mem_free\n");
1181 kbase_gpu_vm_lock(kctx);
1183 if (gpu_addr >= BASE_MEM_COOKIE_BASE &&
1184 gpu_addr < BASE_MEM_FIRST_FREE_ADDRESS) {
1185 int cookie = PFN_DOWN(gpu_addr - BASE_MEM_COOKIE_BASE);
1187 reg = kctx->pending_regions[cookie];
1193 /* ask to unlink the cookie as we'll free it */
1195 kctx->pending_regions[cookie] = NULL;
1196 kctx->cookies |= (1UL << cookie);
1198 kbase_free_alloced_region(reg);
1201 /* Validate the region */
1202 reg = kbase_region_tracker_find_region_base_address(kctx, gpu_addr);
1203 if (!reg || (reg->flags & KBASE_REG_FREE)) {
1204 dev_warn(kctx->kbdev->dev, "kbase_mem_free called with nonexistent gpu_addr 0x%llX",
1210 if ((reg->flags & KBASE_REG_ZONE_MASK) == KBASE_REG_ZONE_SAME_VA) {
1211 /* SAME_VA must be freed through munmap */
1212 dev_warn(kctx->kbdev->dev, "%s called on SAME_VA memory 0x%llX", __func__,
1217 err = kbase_mem_free_region(kctx, reg);
1221 kbase_gpu_vm_unlock(kctx);
1225 KBASE_EXPORT_TEST_API(kbase_mem_free);
1227 void kbase_update_region_flags(struct kbase_context *kctx,
1228 struct kbase_va_region *reg, unsigned long flags)
1230 KBASE_DEBUG_ASSERT(NULL != reg);
1231 KBASE_DEBUG_ASSERT((flags & ~((1ul << BASE_MEM_FLAGS_NR_BITS) - 1)) == 0);
1233 reg->flags |= kbase_cache_enabled(flags, reg->nr_pages);
1234 /* all memory is now growable */
1235 reg->flags |= KBASE_REG_GROWABLE;
1237 if (flags & BASE_MEM_GROW_ON_GPF)
1238 reg->flags |= KBASE_REG_PF_GROW;
1240 if (flags & BASE_MEM_PROT_CPU_WR)
1241 reg->flags |= KBASE_REG_CPU_WR;
1243 if (flags & BASE_MEM_PROT_CPU_RD)
1244 reg->flags |= KBASE_REG_CPU_RD;
1246 if (flags & BASE_MEM_PROT_GPU_WR)
1247 reg->flags |= KBASE_REG_GPU_WR;
1249 if (flags & BASE_MEM_PROT_GPU_RD)
1250 reg->flags |= KBASE_REG_GPU_RD;
1252 if (0 == (flags & BASE_MEM_PROT_GPU_EX))
1253 reg->flags |= KBASE_REG_GPU_NX;
1255 if (flags & BASE_MEM_COHERENT_SYSTEM ||
1256 flags & BASE_MEM_COHERENT_SYSTEM_REQUIRED)
1257 reg->flags |= KBASE_REG_SHARE_BOTH;
1258 else if (flags & BASE_MEM_COHERENT_LOCAL)
1259 reg->flags |= KBASE_REG_SHARE_IN;
1261 /* Set up default MEMATTR usage */
1262 if (kctx->kbdev->system_coherency == COHERENCY_ACE &&
1263 (reg->flags & KBASE_REG_SHARE_BOTH)) {
1265 KBASE_REG_MEMATTR_INDEX(AS_MEMATTR_INDEX_DEFAULT_ACE);
1268 KBASE_REG_MEMATTR_INDEX(AS_MEMATTR_INDEX_DEFAULT);
1271 KBASE_EXPORT_TEST_API(kbase_update_region_flags);
1273 int kbase_alloc_phy_pages_helper(
1274 struct kbase_mem_phy_alloc *alloc,
1275 size_t nr_pages_requested)
1277 int new_page_count __maybe_unused;
1278 size_t old_page_count = alloc->nents;
1280 KBASE_DEBUG_ASSERT(alloc->type == KBASE_MEM_TYPE_NATIVE);
1281 KBASE_DEBUG_ASSERT(alloc->imported.kctx);
1283 if (nr_pages_requested == 0)
1284 goto done; /*nothing to do*/
1286 new_page_count = kbase_atomic_add_pages(
1287 nr_pages_requested, &alloc->imported.kctx->used_pages);
1288 kbase_atomic_add_pages(nr_pages_requested, &alloc->imported.kctx->kbdev->memdev.used_pages);
1290 /* Increase mm counters before we allocate pages so that this
1291 * allocation is visible to the OOM killer */
1292 kbase_process_page_usage_inc(alloc->imported.kctx, nr_pages_requested);
1294 if (kbase_mem_pool_alloc_pages(&alloc->imported.kctx->mem_pool,
1295 nr_pages_requested, alloc->pages + old_page_count) != 0)
1299 * Request a zone cache update, this scans only the new pages an
1300 * appends their information to the zone cache. if the update
1301 * fails then clear the cache so we fall-back to doing things
1304 if (kbase_zone_cache_update(alloc, old_page_count) != 0)
1305 kbase_zone_cache_clear(alloc);
1307 kbase_tlstream_aux_pagesalloc(
1308 (u32)alloc->imported.kctx->id,
1309 (u64)new_page_count);
1311 alloc->nents += nr_pages_requested;
1316 kbase_process_page_usage_dec(alloc->imported.kctx, nr_pages_requested);
1317 kbase_atomic_sub_pages(nr_pages_requested, &alloc->imported.kctx->used_pages);
1318 kbase_atomic_sub_pages(nr_pages_requested, &alloc->imported.kctx->kbdev->memdev.used_pages);
1323 int kbase_free_phy_pages_helper(
1324 struct kbase_mem_phy_alloc *alloc,
1325 size_t nr_pages_to_free)
1327 struct kbase_context *kctx = alloc->imported.kctx;
1329 bool reclaimed = (alloc->evicted != 0);
1330 phys_addr_t *start_free;
1331 int new_page_count __maybe_unused;
1333 KBASE_DEBUG_ASSERT(alloc->type == KBASE_MEM_TYPE_NATIVE);
1334 KBASE_DEBUG_ASSERT(alloc->imported.kctx);
1335 KBASE_DEBUG_ASSERT(alloc->nents >= nr_pages_to_free);
1337 /* early out if nothing to do */
1338 if (0 == nr_pages_to_free)
1341 start_free = alloc->pages + alloc->nents - nr_pages_to_free;
1343 syncback = alloc->properties & KBASE_MEM_PHY_ALLOC_ACCESSED_CACHED;
1346 * Clear the zone cache, we don't expect JIT allocations to be
1347 * shrunk in parts so there is no point trying to optimize for that
1348 * by scanning for the changes caused by freeing this memory and
1349 * updating the existing cache entries.
1351 kbase_zone_cache_clear(alloc);
1353 kbase_mem_pool_free_pages(&kctx->mem_pool,
1359 alloc->nents -= nr_pages_to_free;
1362 * If the allocation was not evicted (i.e. evicted == 0) then
1363 * the page accounting needs to be done.
1366 kbase_process_page_usage_dec(kctx, nr_pages_to_free);
1367 new_page_count = kbase_atomic_sub_pages(nr_pages_to_free,
1369 kbase_atomic_sub_pages(nr_pages_to_free,
1370 &kctx->kbdev->memdev.used_pages);
1372 kbase_tlstream_aux_pagesalloc(
1374 (u64)new_page_count);
1380 void kbase_mem_kref_free(struct kref *kref)
1382 struct kbase_mem_phy_alloc *alloc;
1384 alloc = container_of(kref, struct kbase_mem_phy_alloc, kref);
1386 switch (alloc->type) {
1387 case KBASE_MEM_TYPE_NATIVE: {
1388 WARN_ON(!alloc->imported.kctx);
1390 * The physical allocation must have been removed from the
1391 * eviction list before trying to free it.
1393 WARN_ON(!list_empty(&alloc->evict_node));
1394 kbase_free_phy_pages_helper(alloc, alloc->nents);
1397 case KBASE_MEM_TYPE_ALIAS: {
1398 /* just call put on the underlying phy allocs */
1400 struct kbase_aliased *aliased;
1402 aliased = alloc->imported.alias.aliased;
1404 for (i = 0; i < alloc->imported.alias.nents; i++)
1405 if (aliased[i].alloc)
1406 kbase_mem_phy_alloc_put(aliased[i].alloc);
1411 case KBASE_MEM_TYPE_RAW:
1412 /* raw pages, external cleanup */
1415 case KBASE_MEM_TYPE_IMPORTED_UMP:
1416 ump_dd_release(alloc->imported.ump_handle);
1419 #ifdef CONFIG_DMA_SHARED_BUFFER
1420 case KBASE_MEM_TYPE_IMPORTED_UMM:
1421 dma_buf_detach(alloc->imported.umm.dma_buf,
1422 alloc->imported.umm.dma_attachment);
1423 dma_buf_put(alloc->imported.umm.dma_buf);
1426 case KBASE_MEM_TYPE_IMPORTED_USER_BUF:
1427 if (alloc->imported.user_buf.mm)
1428 mmdrop(alloc->imported.user_buf.mm);
1429 kfree(alloc->imported.user_buf.pages);
1431 case KBASE_MEM_TYPE_TB:{
1434 tb = alloc->imported.kctx->jctx.tb;
1435 kbase_device_trace_buffer_uninstall(alloc->imported.kctx);
1440 WARN(1, "Unexecpted free of type %d\n", alloc->type);
1444 /* Free based on allocation type */
1445 if (alloc->properties & KBASE_MEM_PHY_ALLOC_LARGE)
1451 KBASE_EXPORT_TEST_API(kbase_mem_kref_free);
1453 int kbase_alloc_phy_pages(struct kbase_va_region *reg, size_t vsize, size_t size)
1455 KBASE_DEBUG_ASSERT(NULL != reg);
1456 KBASE_DEBUG_ASSERT(vsize > 0);
1458 /* validate user provided arguments */
1459 if (size > vsize || vsize > reg->nr_pages)
1462 /* Prevent vsize*sizeof from wrapping around.
1463 * For instance, if vsize is 2**29+1, we'll allocate 1 byte and the alloc won't fail.
1465 if ((size_t) vsize > ((size_t) -1 / sizeof(*reg->cpu_alloc->pages)))
1468 KBASE_DEBUG_ASSERT(0 != vsize);
1470 if (kbase_alloc_phy_pages_helper(reg->cpu_alloc, size) != 0)
1473 reg->cpu_alloc->reg = reg;
1474 if (reg->cpu_alloc != reg->gpu_alloc) {
1475 if (kbase_alloc_phy_pages_helper(reg->gpu_alloc, size) != 0)
1477 reg->gpu_alloc->reg = reg;
1483 kbase_free_phy_pages_helper(reg->cpu_alloc, size);
1488 KBASE_EXPORT_TEST_API(kbase_alloc_phy_pages);
1490 bool kbase_check_alloc_flags(unsigned long flags)
1492 /* Only known input flags should be set. */
1493 if (flags & ~BASE_MEM_FLAGS_INPUT_MASK)
1496 /* At least one flag should be set */
1500 /* Either the GPU or CPU must be reading from the allocated memory */
1501 if ((flags & (BASE_MEM_PROT_CPU_RD | BASE_MEM_PROT_GPU_RD)) == 0)
1504 /* Either the GPU or CPU must be writing to the allocated memory */
1505 if ((flags & (BASE_MEM_PROT_CPU_WR | BASE_MEM_PROT_GPU_WR)) == 0)
1508 /* GPU cannot be writing to GPU executable memory and cannot grow the memory on page fault. */
1509 if ((flags & BASE_MEM_PROT_GPU_EX) && (flags & (BASE_MEM_PROT_GPU_WR | BASE_MEM_GROW_ON_GPF)))
1512 /* GPU should have at least read or write access otherwise there is no
1513 reason for allocating. */
1514 if ((flags & (BASE_MEM_PROT_GPU_RD | BASE_MEM_PROT_GPU_WR)) == 0)
1517 /* BASE_MEM_IMPORT_SHARED is only valid for imported memory */
1518 if ((flags & BASE_MEM_IMPORT_SHARED) == BASE_MEM_IMPORT_SHARED)
1524 bool kbase_check_import_flags(unsigned long flags)
1526 /* Only known input flags should be set. */
1527 if (flags & ~BASE_MEM_FLAGS_INPUT_MASK)
1530 /* At least one flag should be set */
1534 /* Imported memory cannot be GPU executable */
1535 if (flags & BASE_MEM_PROT_GPU_EX)
1538 /* Imported memory cannot grow on page fault */
1539 if (flags & BASE_MEM_GROW_ON_GPF)
1542 /* GPU should have at least read or write access otherwise there is no
1543 reason for importing. */
1544 if ((flags & (BASE_MEM_PROT_GPU_RD | BASE_MEM_PROT_GPU_WR)) == 0)
1547 /* Secure memory cannot be read by the CPU */
1548 if ((flags & BASE_MEM_SECURE) && (flags & BASE_MEM_PROT_CPU_RD))
1555 * @brief Acquire the per-context region list lock
1557 void kbase_gpu_vm_lock(struct kbase_context *kctx)
1559 KBASE_DEBUG_ASSERT(kctx != NULL);
1560 mutex_lock(&kctx->reg_lock);
1563 KBASE_EXPORT_TEST_API(kbase_gpu_vm_lock);
1566 * @brief Release the per-context region list lock
1568 void kbase_gpu_vm_unlock(struct kbase_context *kctx)
1570 KBASE_DEBUG_ASSERT(kctx != NULL);
1571 mutex_unlock(&kctx->reg_lock);
1574 KBASE_EXPORT_TEST_API(kbase_gpu_vm_unlock);
1576 #ifdef CONFIG_DEBUG_FS
1577 struct kbase_jit_debugfs_data {
1578 int (*func)(struct kbase_jit_debugfs_data *);
1580 struct kbase_context *kctx;
1587 static int kbase_jit_debugfs_common_open(struct inode *inode,
1588 struct file *file, int (*func)(struct kbase_jit_debugfs_data *))
1590 struct kbase_jit_debugfs_data *data;
1592 data = kzalloc(sizeof(*data), GFP_KERNEL);
1597 mutex_init(&data->lock);
1598 data->kctx = (struct kbase_context *) inode->i_private;
1600 file->private_data = data;
1602 return nonseekable_open(inode, file);
1605 static ssize_t kbase_jit_debugfs_common_read(struct file *file,
1606 char __user *buf, size_t len, loff_t *ppos)
1608 struct kbase_jit_debugfs_data *data;
1612 data = (struct kbase_jit_debugfs_data *) file->private_data;
1613 mutex_lock(&data->lock);
1616 size = strnlen(data->buffer, sizeof(data->buffer));
1623 if (data->func(data)) {
1628 size = scnprintf(data->buffer, sizeof(data->buffer),
1629 "%llu,%llu,%llu", data->active_value,
1630 data->pool_value, data->destroy_value);
1633 ret = simple_read_from_buffer(buf, len, ppos, data->buffer, size);
1636 mutex_unlock(&data->lock);
1640 static int kbase_jit_debugfs_common_release(struct inode *inode,
1643 kfree(file->private_data);
1647 #define KBASE_JIT_DEBUGFS_DECLARE(__fops, __func) \
1648 static int __fops ## _open(struct inode *inode, struct file *file) \
1650 return kbase_jit_debugfs_common_open(inode, file, __func); \
1652 static const struct file_operations __fops = { \
1653 .owner = THIS_MODULE, \
1654 .open = __fops ## _open, \
1655 .release = kbase_jit_debugfs_common_release, \
1656 .read = kbase_jit_debugfs_common_read, \
1658 .llseek = generic_file_llseek, \
1661 static int kbase_jit_debugfs_count_get(struct kbase_jit_debugfs_data *data)
1663 struct kbase_context *kctx = data->kctx;
1664 struct list_head *tmp;
1666 mutex_lock(&kctx->jit_lock);
1667 list_for_each(tmp, &kctx->jit_active_head) {
1668 data->active_value++;
1671 list_for_each(tmp, &kctx->jit_pool_head) {
1675 list_for_each(tmp, &kctx->jit_destroy_head) {
1676 data->destroy_value++;
1678 mutex_unlock(&kctx->jit_lock);
1682 KBASE_JIT_DEBUGFS_DECLARE(kbase_jit_debugfs_count_fops,
1683 kbase_jit_debugfs_count_get);
1685 static int kbase_jit_debugfs_vm_get(struct kbase_jit_debugfs_data *data)
1687 struct kbase_context *kctx = data->kctx;
1688 struct kbase_va_region *reg;
1690 mutex_lock(&kctx->jit_lock);
1691 list_for_each_entry(reg, &kctx->jit_active_head, jit_node) {
1692 data->active_value += reg->nr_pages;
1695 list_for_each_entry(reg, &kctx->jit_pool_head, jit_node) {
1696 data->pool_value += reg->nr_pages;
1699 list_for_each_entry(reg, &kctx->jit_destroy_head, jit_node) {
1700 data->destroy_value += reg->nr_pages;
1702 mutex_unlock(&kctx->jit_lock);
1706 KBASE_JIT_DEBUGFS_DECLARE(kbase_jit_debugfs_vm_fops,
1707 kbase_jit_debugfs_vm_get);
1709 static int kbase_jit_debugfs_phys_get(struct kbase_jit_debugfs_data *data)
1711 struct kbase_context *kctx = data->kctx;
1712 struct kbase_va_region *reg;
1714 mutex_lock(&kctx->jit_lock);
1715 list_for_each_entry(reg, &kctx->jit_active_head, jit_node) {
1716 data->active_value += reg->gpu_alloc->nents;
1719 list_for_each_entry(reg, &kctx->jit_pool_head, jit_node) {
1720 data->pool_value += reg->gpu_alloc->nents;
1723 list_for_each_entry(reg, &kctx->jit_destroy_head, jit_node) {
1724 data->destroy_value += reg->gpu_alloc->nents;
1726 mutex_unlock(&kctx->jit_lock);
1730 KBASE_JIT_DEBUGFS_DECLARE(kbase_jit_debugfs_phys_fops,
1731 kbase_jit_debugfs_phys_get);
1733 void kbase_jit_debugfs_add(struct kbase_context *kctx)
1735 /* Debugfs entry for getting the number of JIT allocations. */
1736 debugfs_create_file("mem_jit_count", S_IRUGO, kctx->kctx_dentry,
1737 kctx, &kbase_jit_debugfs_count_fops);
1740 * Debugfs entry for getting the total number of virtual pages
1741 * used by JIT allocations.
1743 debugfs_create_file("mem_jit_vm", S_IRUGO, kctx->kctx_dentry,
1744 kctx, &kbase_jit_debugfs_vm_fops);
1747 * Debugfs entry for getting the number of physical pages used
1748 * by JIT allocations.
1750 debugfs_create_file("mem_jit_phys", S_IRUGO, kctx->kctx_dentry,
1751 kctx, &kbase_jit_debugfs_phys_fops);
1753 #endif /* CONFIG_DEBUG_FS */
1756 * kbase_jit_destroy_worker - Deferred worker which frees JIT allocations
1759 * This function does the work of freeing JIT allocations whose physical
1760 * backing has been released.
1762 static void kbase_jit_destroy_worker(struct work_struct *work)
1764 struct kbase_context *kctx;
1765 struct kbase_va_region *reg;
1767 kctx = container_of(work, struct kbase_context, jit_work);
1769 mutex_lock(&kctx->jit_lock);
1770 if (list_empty(&kctx->jit_destroy_head))
1773 reg = list_first_entry(&kctx->jit_destroy_head,
1774 struct kbase_va_region, jit_node);
1777 list_del(®->jit_node);
1778 mutex_unlock(&kctx->jit_lock);
1780 kbase_gpu_vm_lock(kctx);
1781 kbase_mem_free_region(kctx, reg);
1782 kbase_gpu_vm_unlock(kctx);
1784 mutex_unlock(&kctx->jit_lock);
1788 int kbase_jit_init(struct kbase_context *kctx)
1790 INIT_LIST_HEAD(&kctx->jit_active_head);
1791 INIT_LIST_HEAD(&kctx->jit_pool_head);
1792 INIT_LIST_HEAD(&kctx->jit_destroy_head);
1793 mutex_init(&kctx->jit_lock);
1794 INIT_WORK(&kctx->jit_work, kbase_jit_destroy_worker);
1799 struct kbase_va_region *kbase_jit_allocate(struct kbase_context *kctx,
1800 struct base_jit_alloc_info *info)
1802 struct kbase_va_region *reg = NULL;
1803 struct kbase_va_region *walker;
1804 struct kbase_va_region *temp;
1805 size_t current_diff = SIZE_MAX;
1809 mutex_lock(&kctx->jit_lock);
1811 * Scan the pool for an existing allocation which meets our
1812 * requirements and remove it.
1814 list_for_each_entry_safe(walker, temp, &kctx->jit_pool_head, jit_node) {
1816 if (walker->nr_pages >= info->va_pages) {
1817 size_t min_size, max_size, diff;
1820 * The JIT allocations VA requirements have been
1821 * meet, it's suitable but other allocations
1822 * might be a better fit.
1824 min_size = min_t(size_t, walker->gpu_alloc->nents,
1825 info->commit_pages);
1826 max_size = max_t(size_t, walker->gpu_alloc->nents,
1827 info->commit_pages);
1828 diff = max_size - min_size;
1830 if (current_diff > diff) {
1831 current_diff = diff;
1835 /* The allocation is an exact match, stop looking */
1836 if (current_diff == 0)
1843 * Remove the found region from the pool and add it to the
1846 list_del_init(®->jit_node);
1847 list_add(®->jit_node, &kctx->jit_active_head);
1849 /* Release the jit lock before modifying the allocation */
1850 mutex_unlock(&kctx->jit_lock);
1852 kbase_gpu_vm_lock(kctx);
1854 /* Make the physical backing no longer reclaimable */
1855 if (!kbase_mem_evictable_unmake(reg->gpu_alloc))
1858 /* Grow the backing if required */
1859 if (reg->gpu_alloc->nents < info->commit_pages) {
1861 size_t old_size = reg->gpu_alloc->nents;
1863 /* Allocate some more pages */
1864 delta = info->commit_pages - reg->gpu_alloc->nents;
1865 if (kbase_alloc_phy_pages_helper(reg->gpu_alloc, delta)
1869 if (reg->cpu_alloc != reg->gpu_alloc) {
1870 if (kbase_alloc_phy_pages_helper(
1871 reg->cpu_alloc, delta) != 0) {
1872 kbase_free_phy_pages_helper(
1873 reg->gpu_alloc, delta);
1878 ret = kbase_mem_grow_gpu_mapping(kctx, reg,
1879 info->commit_pages, old_size);
1881 * The grow failed so put the allocation back in the
1882 * pool and return failure.
1887 kbase_gpu_vm_unlock(kctx);
1889 /* No suitable JIT allocation was found so create a new one */
1890 u64 flags = BASE_MEM_PROT_CPU_RD | BASE_MEM_PROT_GPU_RD |
1891 BASE_MEM_PROT_GPU_WR | BASE_MEM_GROW_ON_GPF |
1892 BASE_MEM_COHERENT_LOCAL;
1896 mutex_unlock(&kctx->jit_lock);
1898 reg = kbase_mem_alloc(kctx, info->va_pages, info->commit_pages,
1899 info->extent, &flags, &gpu_addr, &alignment);
1903 mutex_lock(&kctx->jit_lock);
1904 list_add(®->jit_node, &kctx->jit_active_head);
1905 mutex_unlock(&kctx->jit_lock);
1912 * An update to an allocation from the pool failed, chances
1913 * are slim a new allocation would fair any better so return
1914 * the allocation to the pool and return the function with failure.
1916 kbase_gpu_vm_unlock(kctx);
1917 mutex_lock(&kctx->jit_lock);
1918 list_del_init(®->jit_node);
1919 list_add(®->jit_node, &kctx->jit_pool_head);
1920 mutex_unlock(&kctx->jit_lock);
1925 void kbase_jit_free(struct kbase_context *kctx, struct kbase_va_region *reg)
1927 /* The physical backing of memory in the pool is always reclaimable */
1928 down_read(&kctx->process_mm->mmap_sem);
1929 kbase_gpu_vm_lock(kctx);
1930 kbase_mem_evictable_make(reg->gpu_alloc);
1931 kbase_gpu_vm_unlock(kctx);
1932 up_read(&kctx->process_mm->mmap_sem);
1934 mutex_lock(&kctx->jit_lock);
1935 list_del_init(®->jit_node);
1936 list_add(®->jit_node, &kctx->jit_pool_head);
1937 mutex_unlock(&kctx->jit_lock);
1940 void kbase_jit_backing_lost(struct kbase_va_region *reg)
1942 struct kbase_context *kctx = reg->kctx;
1945 * JIT allocations will always be on a list, if the region
1946 * is not on a list then it's not a JIT allocation.
1948 if (list_empty(®->jit_node))
1952 * Freeing the allocation requires locks we might not be able
1953 * to take now, so move the allocation to the free list and kick
1954 * the worker which will do the freeing.
1956 mutex_lock(&kctx->jit_lock);
1957 list_del_init(®->jit_node);
1958 list_add(®->jit_node, &kctx->jit_destroy_head);
1959 mutex_unlock(&kctx->jit_lock);
1961 schedule_work(&kctx->jit_work);
1964 bool kbase_jit_evict(struct kbase_context *kctx)
1966 struct kbase_va_region *reg = NULL;
1968 lockdep_assert_held(&kctx->reg_lock);
1970 /* Free the oldest allocation from the pool */
1971 mutex_lock(&kctx->jit_lock);
1972 if (!list_empty(&kctx->jit_pool_head)) {
1973 reg = list_entry(kctx->jit_pool_head.prev,
1974 struct kbase_va_region, jit_node);
1975 list_del(®->jit_node);
1977 mutex_unlock(&kctx->jit_lock);
1980 kbase_mem_free_region(kctx, reg);
1982 return (reg != NULL);
1985 void kbase_jit_term(struct kbase_context *kctx)
1987 struct kbase_va_region *walker;
1989 /* Free all allocations for this context */
1992 * Flush the freeing of allocations whose backing has been freed
1993 * (i.e. everything in jit_destroy_head).
1995 cancel_work_sync(&kctx->jit_work);
1997 kbase_gpu_vm_lock(kctx);
1998 /* Free all allocations from the pool */
1999 while (!list_empty(&kctx->jit_pool_head)) {
2000 walker = list_first_entry(&kctx->jit_pool_head,
2001 struct kbase_va_region, jit_node);
2002 list_del(&walker->jit_node);
2003 kbase_mem_free_region(kctx, walker);
2006 /* Free all allocations from active list */
2007 while (!list_empty(&kctx->jit_active_head)) {
2008 walker = list_first_entry(&kctx->jit_active_head,
2009 struct kbase_va_region, jit_node);
2010 list_del(&walker->jit_node);
2011 kbase_mem_free_region(kctx, walker);
2013 kbase_gpu_vm_unlock(kctx);
2016 static int kbase_jd_user_buf_map(struct kbase_context *kctx,
2017 struct kbase_va_region *reg)
2020 struct kbase_mem_phy_alloc *alloc;
2021 struct page **pages;
2025 unsigned long address;
2026 struct mm_struct *mm;
2028 unsigned long offset;
2029 unsigned long local_size;
2031 alloc = reg->gpu_alloc;
2032 pa = kbase_get_gpu_phy_pages(reg);
2033 address = alloc->imported.user_buf.address;
2034 mm = alloc->imported.user_buf.mm;
2036 KBASE_DEBUG_ASSERT(alloc->type == KBASE_MEM_TYPE_IMPORTED_USER_BUF);
2038 pages = alloc->imported.user_buf.pages;
2040 #if LINUX_VERSION_CODE < KERNEL_VERSION(4, 6, 0)
2041 pinned_pages = get_user_pages(NULL, mm,
2043 alloc->imported.user_buf.nr_pages,
2044 reg->flags & KBASE_REG_GPU_WR,
2047 pinned_pages = get_user_pages_remote(NULL, mm,
2049 alloc->imported.user_buf.nr_pages,
2050 reg->flags & KBASE_REG_GPU_WR,
2054 if (pinned_pages <= 0)
2055 return pinned_pages;
2057 if (pinned_pages != alloc->imported.user_buf.nr_pages) {
2058 for (i = 0; i < pinned_pages; i++)
2063 dev = kctx->kbdev->dev;
2064 offset = address & ~PAGE_MASK;
2065 local_size = alloc->imported.user_buf.size;
2067 for (i = 0; i < pinned_pages; i++) {
2068 dma_addr_t dma_addr;
2071 min = MIN(PAGE_SIZE - offset, local_size);
2072 dma_addr = dma_map_page(dev, pages[i],
2075 if (dma_mapping_error(dev, dma_addr))
2078 alloc->imported.user_buf.dma_addrs[i] = dma_addr;
2079 pa[i] = page_to_phys(pages[i]);
2085 alloc->nents = pinned_pages;
2087 err = kbase_mmu_insert_pages(kctx, reg->start_pfn, pa,
2088 kbase_reg_current_backed_size(reg),
2097 dma_unmap_page(kctx->kbdev->dev,
2098 alloc->imported.user_buf.dma_addrs[i],
2099 PAGE_SIZE, DMA_BIDIRECTIONAL);
2107 static void kbase_jd_user_buf_unmap(struct kbase_context *kctx,
2108 struct kbase_mem_phy_alloc *alloc, bool writeable)
2111 struct page **pages;
2112 unsigned long size = alloc->imported.user_buf.size;
2114 KBASE_DEBUG_ASSERT(alloc->type == KBASE_MEM_TYPE_IMPORTED_USER_BUF);
2115 pages = alloc->imported.user_buf.pages;
2116 for (i = 0; i < alloc->imported.user_buf.nr_pages; i++) {
2117 unsigned long local_size;
2118 dma_addr_t dma_addr = alloc->imported.user_buf.dma_addrs[i];
2120 local_size = MIN(size, PAGE_SIZE - (dma_addr & ~PAGE_MASK));
2121 dma_unmap_page(kctx->kbdev->dev, dma_addr, local_size,
2124 set_page_dirty_lock(pages[i]);
2134 /* to replace sg_dma_len. */
2135 #define MALI_SG_DMA_LEN(sg) ((sg)->length)
2137 #ifdef CONFIG_DMA_SHARED_BUFFER
2138 static int kbase_jd_umm_map(struct kbase_context *kctx,
2139 struct kbase_va_region *reg)
2141 struct sg_table *sgt;
2142 struct scatterlist *s;
2147 struct kbase_mem_phy_alloc *alloc;
2149 alloc = reg->gpu_alloc;
2151 KBASE_DEBUG_ASSERT(alloc->type == KBASE_MEM_TYPE_IMPORTED_UMM);
2152 KBASE_DEBUG_ASSERT(NULL == alloc->imported.umm.sgt);
2153 sgt = dma_buf_map_attachment(alloc->imported.umm.dma_attachment,
2156 if (IS_ERR_OR_NULL(sgt))
2159 /* save for later */
2160 alloc->imported.umm.sgt = sgt;
2162 pa = kbase_get_gpu_phy_pages(reg);
2163 KBASE_DEBUG_ASSERT(pa);
2165 for_each_sg(sgt->sgl, s, sgt->nents, i) {
2167 size_t pages = PFN_UP(MALI_SG_DMA_LEN(s));
2169 WARN_ONCE(MALI_SG_DMA_LEN(s) & (PAGE_SIZE-1),
2170 "MALI_SG_DMA_LEN(s)=%u is not a multiple of PAGE_SIZE\n",
2171 MALI_SG_DMA_LEN(s));
2173 WARN_ONCE(sg_dma_address(s) & (PAGE_SIZE-1),
2174 "sg_dma_address(s)=%llx is not aligned to PAGE_SIZE\n",
2175 (unsigned long long) sg_dma_address(s));
2177 for (j = 0; (j < pages) && (count < reg->nr_pages); j++,
2179 *pa++ = sg_dma_address(s) + (j << PAGE_SHIFT);
2180 WARN_ONCE(j < pages,
2181 "sg list from dma_buf_map_attachment > dma_buf->size=%zu\n",
2182 alloc->imported.umm.dma_buf->size);
2185 if (WARN_ONCE(count < reg->nr_pages,
2186 "sg list from dma_buf_map_attachment < dma_buf->size=%zu\n",
2187 alloc->imported.umm.dma_buf->size)) {
2192 /* Update nents as we now have pages to map */
2193 alloc->nents = count;
2195 err = kbase_mmu_insert_pages(kctx, reg->start_pfn,
2196 kbase_get_gpu_phy_pages(reg),
2197 kbase_reg_current_backed_size(reg),
2198 reg->flags | KBASE_REG_GPU_WR | KBASE_REG_GPU_RD);
2202 dma_buf_unmap_attachment(alloc->imported.umm.dma_attachment,
2203 alloc->imported.umm.sgt, DMA_BIDIRECTIONAL);
2204 alloc->imported.umm.sgt = NULL;
2210 static void kbase_jd_umm_unmap(struct kbase_context *kctx,
2211 struct kbase_mem_phy_alloc *alloc)
2213 KBASE_DEBUG_ASSERT(kctx);
2214 KBASE_DEBUG_ASSERT(alloc);
2215 KBASE_DEBUG_ASSERT(alloc->imported.umm.dma_attachment);
2216 KBASE_DEBUG_ASSERT(alloc->imported.umm.sgt);
2217 dma_buf_unmap_attachment(alloc->imported.umm.dma_attachment,
2218 alloc->imported.umm.sgt, DMA_BIDIRECTIONAL);
2219 alloc->imported.umm.sgt = NULL;
2222 #endif /* CONFIG_DMA_SHARED_BUFFER */
2224 #if (defined(CONFIG_KDS) && defined(CONFIG_UMP)) \
2225 || defined(CONFIG_DMA_SHARED_BUFFER_USES_KDS)
2226 static void add_kds_resource(struct kds_resource *kds_res,
2227 struct kds_resource **kds_resources, u32 *kds_res_count,
2228 unsigned long *kds_access_bitmap, bool exclusive)
2232 for (i = 0; i < *kds_res_count; i++) {
2233 /* Duplicate resource, ignore */
2234 if (kds_resources[i] == kds_res)
2238 kds_resources[*kds_res_count] = kds_res;
2240 set_bit(*kds_res_count, kds_access_bitmap);
2245 struct kbase_mem_phy_alloc *kbase_map_external_resource(
2246 struct kbase_context *kctx, struct kbase_va_region *reg,
2247 struct mm_struct *locked_mm
2249 , u32 *kds_res_count, struct kds_resource **kds_resources,
2250 unsigned long *kds_access_bitmap, bool exclusive
2256 /* decide what needs to happen for this resource */
2257 switch (reg->gpu_alloc->type) {
2258 case KBASE_MEM_TYPE_IMPORTED_USER_BUF: {
2259 if (reg->gpu_alloc->imported.user_buf.mm != locked_mm)
2262 reg->gpu_alloc->imported.user_buf.current_mapping_usage_count++;
2263 if (1 == reg->gpu_alloc->imported.user_buf.current_mapping_usage_count) {
2264 err = kbase_jd_user_buf_map(kctx, reg);
2266 reg->gpu_alloc->imported.user_buf.current_mapping_usage_count--;
2272 case KBASE_MEM_TYPE_IMPORTED_UMP: {
2273 #if defined(CONFIG_KDS) && defined(CONFIG_UMP)
2274 if (kds_res_count) {
2275 struct kds_resource *kds_res;
2277 kds_res = ump_dd_kds_resource_get(
2278 reg->gpu_alloc->imported.ump_handle);
2280 add_kds_resource(kds_res, kds_resources,
2282 kds_access_bitmap, exclusive);
2284 #endif /*defined(CONFIG_KDS) && defined(CONFIG_UMP) */
2287 #ifdef CONFIG_DMA_SHARED_BUFFER
2288 case KBASE_MEM_TYPE_IMPORTED_UMM: {
2289 #ifdef CONFIG_DMA_SHARED_BUFFER_USES_KDS
2290 if (kds_res_count) {
2291 struct kds_resource *kds_res;
2293 kds_res = get_dma_buf_kds_resource(
2294 reg->gpu_alloc->imported.umm.dma_buf);
2296 add_kds_resource(kds_res, kds_resources,
2298 kds_access_bitmap, exclusive);
2301 reg->gpu_alloc->imported.umm.current_mapping_usage_count++;
2302 if (1 == reg->gpu_alloc->imported.umm.current_mapping_usage_count) {
2303 err = kbase_jd_umm_map(kctx, reg);
2305 reg->gpu_alloc->imported.umm.current_mapping_usage_count--;
2316 return kbase_mem_phy_alloc_get(reg->gpu_alloc);
2321 void kbase_unmap_external_resource(struct kbase_context *kctx,
2322 struct kbase_va_region *reg, struct kbase_mem_phy_alloc *alloc)
2324 switch (alloc->type) {
2325 #ifdef CONFIG_DMA_SHARED_BUFFER
2326 case KBASE_MEM_TYPE_IMPORTED_UMM: {
2327 alloc->imported.umm.current_mapping_usage_count--;
2329 if (0 == alloc->imported.umm.current_mapping_usage_count) {
2330 if (reg && reg->gpu_alloc == alloc)
2331 kbase_mmu_teardown_pages(
2334 kbase_reg_current_backed_size(reg));
2336 kbase_jd_umm_unmap(kctx, alloc);
2340 #endif /* CONFIG_DMA_SHARED_BUFFER */
2341 case KBASE_MEM_TYPE_IMPORTED_USER_BUF: {
2342 alloc->imported.user_buf.current_mapping_usage_count--;
2344 if (0 == alloc->imported.user_buf.current_mapping_usage_count) {
2345 bool writeable = true;
2347 if (reg && reg->gpu_alloc == alloc)
2348 kbase_mmu_teardown_pages(
2351 kbase_reg_current_backed_size(reg));
2353 if (reg && ((reg->flags & KBASE_REG_GPU_WR) == 0))
2356 kbase_jd_user_buf_unmap(kctx, alloc, writeable);
2363 kbase_mem_phy_alloc_put(alloc);
2366 struct kbase_ctx_ext_res_meta *kbase_sticky_resource_acquire(
2367 struct kbase_context *kctx, u64 gpu_addr)
2369 struct kbase_ctx_ext_res_meta *meta = NULL;
2370 struct kbase_ctx_ext_res_meta *walker;
2372 lockdep_assert_held(&kctx->reg_lock);
2375 * Walk the per context external resource metadata list for the
2376 * metadata which matches the region which is being acquired.
2378 list_for_each_entry(walker, &kctx->ext_res_meta_head, ext_res_node) {
2379 if (walker->gpu_addr == gpu_addr) {
2385 /* No metadata exists so create one. */
2387 struct kbase_va_region *reg;
2389 /* Find the region */
2390 reg = kbase_region_tracker_find_region_enclosing_address(
2392 if (NULL == reg || (reg->flags & KBASE_REG_FREE))
2395 /* Allocate the metadata object */
2396 meta = kzalloc(sizeof(*meta), GFP_KERNEL);
2401 * Fill in the metadata object and acquire a reference
2402 * for the physical resource.
2404 meta->alloc = kbase_map_external_resource(kctx, reg, NULL
2414 meta->gpu_addr = reg->start_pfn << PAGE_SHIFT;
2416 list_add(&meta->ext_res_node, &kctx->ext_res_meta_head);
2427 bool kbase_sticky_resource_release(struct kbase_context *kctx,
2428 struct kbase_ctx_ext_res_meta *meta, u64 gpu_addr)
2430 struct kbase_ctx_ext_res_meta *walker;
2431 struct kbase_va_region *reg;
2433 lockdep_assert_held(&kctx->reg_lock);
2435 /* Search of the metadata if one isn't provided. */
2438 * Walk the per context external resource metadata list for the
2439 * metadata which matches the region which is being released.
2441 list_for_each_entry(walker, &kctx->ext_res_meta_head,
2443 if (walker->gpu_addr == gpu_addr) {
2450 /* No metadata so just return. */
2454 /* Drop the physical memory reference and free the metadata. */
2455 reg = kbase_region_tracker_find_region_enclosing_address(
2459 kbase_unmap_external_resource(kctx, reg, meta->alloc);
2460 list_del(&meta->ext_res_node);
2466 int kbase_sticky_resource_init(struct kbase_context *kctx)
2468 INIT_LIST_HEAD(&kctx->ext_res_meta_head);
2473 void kbase_sticky_resource_term(struct kbase_context *kctx)
2475 struct kbase_ctx_ext_res_meta *walker;
2477 lockdep_assert_held(&kctx->reg_lock);
2480 * Free any sticky resources which haven't been unmapped.
2483 * We don't care about refcounts at this point as no future
2484 * references to the meta data will be made.
2485 * Region termination would find these if we didn't free them
2486 * here, but it's more efficient if we do the clean up here.
2488 while (!list_empty(&kctx->ext_res_meta_head)) {
2489 walker = list_first_entry(&kctx->ext_res_meta_head,
2490 struct kbase_ctx_ext_res_meta, ext_res_node);
2492 kbase_sticky_resource_release(kctx, walker, 0);