Re: [PATCH v5 1/3] drm/buddy: Optimize free block management with RB tree

[Arunpravin Paneer Selvam <arunpravin.paneerselvam@amd.com>]

On 9/9/2025 7:35 PM, Peter Zijlstra wrote:
> On Tue, Sep 09, 2025 at 02:04:30PM +0200, Christian König wrote:
>> Hi Arun,
>>
>> On 09.09.25 11:56, Arunpravin Paneer Selvam wrote:
>> [SNIP]
>>
>>> +/**
>>> + * rbtree_for_each_entry_safe - iterate in-order over rb_root safe against removal
>>> + *
>>> + * @pos:	the 'type *' to use as a loop cursor
>>> + * @n:		another 'type *' to use as temporary storage
>>> + * @root:	'rb_root *' of the rbtree
>>> + * @member:	the name of the rb_node field within 'type'
>>> + */
>>> +#define rbtree_for_each_entry_safe(pos, n, root, member) \
>>> +	for ((pos) = rb_entry_safe(rb_first(root), typeof(*(pos)), member), \
>>> +	     (n) = (pos) ? rb_entry_safe(rb_next(&(pos)->member), typeof(*(pos)), member) : NULL; \
>>> +	     (pos); \
>>> +	     (pos) = (n), \
>>> +	     (n) = (pos) ? rb_entry_safe(rb_next(&(pos)->member), typeof(*(pos)), member) : NULL)
>> As far as I know exactly that operation does not work on an R/B tree.
>>
>> See the _safe() variants of the for_each_ macros are usually used to iterate over a container while being able to remove entries.
>>
>> But because of the potential re-balance storing just the next entry is not sufficient for an R/B tree to do that as far as I know.
>>
>> Please explain how exactly you want to use this macro.
> So I don't much like these iterators; I've said so before. Either we
> should introduce a properly threaded rb-tree (where the NULL child
> pointers encode a linked list), or simply keep a list_head next to the
> rb_node and use that.
>
> The rb_{next,prev}() things are O(ln n), in the worst case they do a
> full traversal up the tree and a full traversal down the other branch.
>
> That said; given 'next' will remain an existing node, only the 'pos'
> node gets removed, rb_next() will still work correctly, even in the face
> of rebalance.
Sorry for the delay. I have been discussing with Christian and testing a 
few code
changes. Maintaining a sorted list_head alongside each rb_node is expensive,
which is the main reason we are moving from a list to an rbtree. In the 
force_merge()
function, we only call this during normal allocation to iterate once or 
twice and merge
the required blocks, not the entire tree. Therefore, rb_prev is 
sufficient, and want to
avoid adding unnecessary complexity for this simple operation. 
Therefore, I have removed
all the newly added macros in v7.

A full traversal of force_merge() is only needed during the buddy 
allocator's fini() operation, and in that
case, any slowness or timing overhead is not critical.

Thanks,
Arun.
>
>