[Intel-gfx] [PATCH 5/8] drm: Implement method to free unused pages

[Arunpravin]

On 04/11/21 12:46 am, Matthew Auld wrote:
> On 25/10/2021 14:00, Arunpravin wrote:
>> On contiguous allocation, we round up the size
>> to the *next* power of 2, implement a function
>> to free the unused pages after the newly allocate block.
>>
>> Signed-off-by: Arunpravin <Arunpravin.PaneerSelvam at amd.com>
> 
> Ideally this gets added with some user, so we can see it in action? 
> Maybe squash the next patch here?
[Arun] ok
> 
>> ---
>>   drivers/gpu/drm/drm_buddy.c | 103 ++++++++++++++++++++++++++++++++++++
>>   include/drm/drm_buddy.h     |   4 ++
>>   2 files changed, 107 insertions(+)
>>
>> diff --git a/drivers/gpu/drm/drm_buddy.c b/drivers/gpu/drm/drm_buddy.c
>> index 9d3547bcc5da..0da8510736eb 100644
>> --- a/drivers/gpu/drm/drm_buddy.c
>> +++ b/drivers/gpu/drm/drm_buddy.c
>> @@ -284,6 +284,109 @@ static inline bool contains(u64 s1, u64 e1, u64 s2, u64 e2)
>>   	return s1 <= s2 && e1 >= e2;
>>   }
>>   
>> +/**
>> + * drm_buddy_free_unused_pages - free unused pages
>> + *
>> + * @mm: DRM buddy manager
>> + * @actual_size: original size requested
>> + * @blocks: output list head to add allocated blocks
>> + *
>> + * For contiguous allocation, we round up the size to the nearest
>> + * power of two value, drivers consume *actual* size, so remaining
>> + * portions are unused and it can be freed.
>> + *
>> + * Returns:
>> + * 0 on success, error code on failure.
>> + */
>> +int drm_buddy_free_unused_pages(struct drm_buddy_mm *mm,
> 
> drm_buddy_block_trim?
[Arun] ok
> 
>> +				u64 actual_size,
> 
> new_size?
[Arun] ok
> 
>> +				struct list_head *blocks)
>> +{
>> +	struct drm_buddy_block *block;
>> +	struct drm_buddy_block *buddy;
>> +	u64 actual_start;
>> +	u64 actual_end;
>> +	LIST_HEAD(dfs);
>> +	u64 count = 0;
>> +	int err;
>> +
>> +	if (!list_is_singular(blocks))
>> +		return -EINVAL;
>> +
>> +	block = list_first_entry_or_null(blocks,
>> +					 struct drm_buddy_block,
>> +					 link);
>> +
>> +	if (!block)
>> +		return -EINVAL;
> 
> list_is_singular() already ensures that I guess?
[Arun] yes it checks the list empty status, I will remove 'if (!block)'
check
> 
> 
>> +
>> +	if (actual_size > drm_buddy_block_size(mm, block))
>> +		return -EINVAL;
>> +
>> +	if (actual_size == drm_buddy_block_size(mm, block))
>> +		return 0;
> 
> Probably need to check the alignment of the actual_size, and also check 
> that it is non-zero?
[Arun] ok
> 
>> +
>> +	list_del(&block->link);
>> +
>> +	actual_start = drm_buddy_block_offset(block);
>> +	actual_end = actual_start + actual_size - 1;
>> +
>> +	if (drm_buddy_block_is_allocated(block))
> 
> That should rather be a programmer error.
[Arun] ok, I will check for the allocation status and return -EINVAL if
the block is not allocated.
> 
>> +		mark_free(mm, block);
>> +
>> +	list_add(&block->tmp_link, &dfs);
>> +
>> +	while (1) {
>> +		block = list_first_entry_or_null(&dfs,
>> +						 struct drm_buddy_block,
>> +						 tmp_link);
>> +
>> +		if (!block)
>> +			break;
>> +
>> +		list_del(&block->tmp_link);
>> +
>> +		if (count == actual_size)
>> +			return 0;
> 
> 
> Check for overlaps somewhere here to avoid needless searching and splitting?
[Arun] ok
> 
>> +
>> +		if (contains(actual_start, actual_end, drm_buddy_block_offset(block),
>> +			(drm_buddy_block_offset(block) + drm_buddy_block_size(mm, block) - 1))) {
> 
> Could maybe record the start/end for better readability?
[Arun] ok
> 
>> +			BUG_ON(!drm_buddy_block_is_free(block));
>> +
>> +			/* Allocate only required blocks */
>> +			mark_allocated(block);
>> +			mm->avail -= drm_buddy_block_size(mm, block);
>> +			list_add_tail(&block->link, blocks);
>> +			count += drm_buddy_block_size(mm, block);
>> +			continue;
>> +		}
>> +
>> +		if (drm_buddy_block_order(block) == 0)
>> +			continue;
> 
> Should be impossible with overlaps check added.
[Arun] yes, I will remove
> 
>> +
>> +		if (!drm_buddy_block_is_split(block)) {
> 
> That should always be true.
[Arun] ok
> 
>> +			err = split_block(mm, block);
>> +
>> +			if (unlikely(err))
>> +				goto err_undo;
>> +		}
>> +
>> +		list_add(&block->right->tmp_link, &dfs);
>> +		list_add(&block->left->tmp_link, &dfs);
>> +	}
>> +
>> +	return -ENOSPC;
> 
> 
> Would it make sense to factor out part of the alloc_range for this? It 
> looks roughly the same.
[Arun] This function gets called for non-range allocations (0..max_size)
as well on contiguous allocation. alloc_range() is called only for range
allocations.
> 
>> +
>> +err_undo:
>> +	buddy = get_buddy(block);
>> +	if (buddy &&
>> +	    (drm_buddy_block_is_free(block) &&
>> +	     drm_buddy_block_is_free(buddy)))
>> +		__drm_buddy_free(mm, block);
>> +	return err;
> 
> 
> Where do we add the block back to the original list? Did we not just 
> leak it?
[Arun] we are adding back to the original list if contains() check
becomes true. we are adding all the blocks within the actual_start and
actual_end, and remaining blocks are freed (added to free list).
> 
> 
>> +}
>> +EXPORT_SYMBOL(drm_buddy_free_unused_pages);
>> +
>>   static struct drm_buddy_block *
>>   alloc_range(struct drm_buddy_mm *mm,
>>   	    u64 start, u64 end,
>> diff --git a/include/drm/drm_buddy.h b/include/drm/drm_buddy.h
>> index cd8021d2d6e7..1dfc80c88e1f 100644
>> --- a/include/drm/drm_buddy.h
>> +++ b/include/drm/drm_buddy.h
>> @@ -145,6 +145,10 @@ int drm_buddy_alloc(struct drm_buddy_mm *mm,
>>   		    struct list_head *blocks,
>>   		    unsigned long flags);
>>   
>> +int drm_buddy_free_unused_pages(struct drm_buddy_mm *mm,
>> +				u64 actual_size,
>> +				struct list_head *blocks);
>> +
>>   void drm_buddy_free(struct drm_buddy_mm *mm, struct drm_buddy_block *block);
>>   
>>   void drm_buddy_free_list(struct drm_buddy_mm *mm, struct list_head *objects);
>>