Re: [PATCH 1/3] maple_tree: Make maple state reusable after mas_empty_area_rev()

[Peng Zhang <zhangpeng.00@bytedance.com>]

在 2023/4/20 06:54, Liam R. Howlett 写道:
> * Peng Zhang <zhangpeng.00@bytedance.com> [230419 05:02]:
>> 在 2023/4/14 22:57, Liam R. Howlett 写道:
>>> Stop using maple state min/max for the range by passing through pointers
>>> for those values.  This will allow the maple state to be reused without
>>> resetting.
>>>
>>> Also add some logic to fail out early on searching with invalid
>>> arguments.
>>>
>>> Fixes: 54a611b60590 ("Maple Tree: add new data structure")
>>> Signed-off-by: Liam R. Howlett <Liam.Howlett@oracle.com>
>>> ---
>>>    lib/maple_tree.c | 27 +++++++++++++--------------
>>>    1 file changed, 13 insertions(+), 14 deletions(-)
>>>
>>> diff --git a/lib/maple_tree.c b/lib/maple_tree.c
>>> index 4df6a0ce1c1b..ed350aa293b2 100644
>>> --- a/lib/maple_tree.c
>>> +++ b/lib/maple_tree.c
>>> @@ -4938,7 +4938,8 @@ static inline void *mas_prev_entry(struct ma_state *mas, unsigned long min)
>>>     * Return: True if found in a leaf, false otherwise.
>>>     *
>>>     */
>>> -static bool mas_rev_awalk(struct ma_state *mas, unsigned long size)
>>> +static bool mas_rev_awalk(struct ma_state *mas, unsigned long size,
>>> +		unsigned long *gap_min, unsigned long *gap_max)
>>>    {
>>>    	enum maple_type type = mte_node_type(mas->node);
>>>    	struct maple_node *node = mas_mn(mas);
>>> @@ -5003,8 +5004,8 @@ static bool mas_rev_awalk(struct ma_state *mas, unsigned long size)
>>>    	if (unlikely(ma_is_leaf(type))) {
>>>    		mas->offset = offset;
>>> -		mas->min = min;
>>> -		mas->max = min + gap - 1;
>>> +		*gap_min = min;
>>> +		*gap_max = min + gap - 1;
>>>    		return true;
>>>    	}
>>> @@ -5280,6 +5281,9 @@ int mas_empty_area(struct ma_state *mas, unsigned long min,
>>>    	unsigned long *pivots;
>>>    	enum maple_type mt;
>>> +	if (min >= max)
>> This can lead to errors, min == max is valid.
>> I think it's better to change it to this:
>> if (min > max || size == 0 || max - min < size - 1)
> I am not sure what it means to search within a range of one.  I guess
> you would expect it to just return that value if it's empty?
Yes, if min==max, and the value pointed to by this index is empty,
we should return it.
>
> In any case, since we are dealing with pages of data for the VMAs,
> having min == max really makes no sense, even with the subtraction of
> one in the caller to reduce the max, the min and max should be at least
> PAGE_SIZE - 1 apart here.
>
> I think you are right, and I think this needs to be looked at for the
> tree on its own, but I don't think it's a problem for the VMA user. I'll
> write a testcase and ensure a search for a single entry in a single
> entry window works separately.  Thanks for pointing this out.
Yes, it's a problem with the tree itself.
>
>>> +		return -EINVAL;
>>> +
>>>    	if (mas_is_start(mas))
>>>    		mas_start(mas);
>>>    	else if (mas->offset >= 2)
>>> @@ -5334,6 +5338,9 @@ int mas_empty_area_rev(struct ma_state *mas, unsigned long min,
>>>    {
>>>    	struct maple_enode *last = mas->node;
>>> +	if (min >= max)
>> ditto.
> I'll do the search in both directions.
Here is also the same problem with the tree itself as above.
Maybe I didn't understand what you mean. I think the case of
min==max should still be considered in mas_empty_area_rev(),
because mas_empty_area_rev() is a separate interface.
>
>>> +		return -EINVAL;
>>> +
>>>    	if (mas_is_start(mas)) {
>>>    		mas_start(mas);
>>>    		mas->offset = mas_data_end(mas);
>>> @@ -5353,7 +5360,7 @@ int mas_empty_area_rev(struct ma_state *mas, unsigned long min,
>>>    	mas->index = min;
>>>    	mas->last = max;
>>> -	while (!mas_rev_awalk(mas, size)) {
>>> +	while (!mas_rev_awalk(mas, size, &min, &max)) {
>>>    		if (last == mas->node) {
>>>    			if (!mas_rewind_node(mas))
>>>    				return -EBUSY;
>>> @@ -5368,17 +5375,9 @@ int mas_empty_area_rev(struct ma_state *mas, unsigned long min,
>>>    	if (unlikely(mas->offset == MAPLE_NODE_SLOTS))
>>>    		return -EBUSY;
>>> -	/*
>>> -	 * mas_rev_awalk() has set mas->min and mas->max to the gap values.  If
>>> -	 * the maximum is outside the window we are searching, then use the last
>>> -	 * location in the search.
>>> -	 * mas->max and mas->min is the range of the gap.
>>> -	 * mas->index and mas->last are currently set to the search range.
>>> -	 */
>>> -
>>>    	/* Trim the upper limit to the max. */
>>> -	if (mas->max <= mas->last)
>>> -		mas->last = mas->max;
>>> +	if (max <= mas->last)
>>> +		mas->last = max;
>> We can get max as follows, without using pointers to track min, max in
>> mas_rev_awalk().
>>
>> mt = mte_node_type(mas->node); pivots = ma_pivots(mas_mn(mas), mt); max =
>> mas_logical_pivot(mas, pivots, mas->offset, mt);
> Yes, but why would we do this?  We have done all this work already in
> mas_rev_awalk(), and we have to do it there to get the offset in the
> first place.
Yes, both methods will work.
>
>> 	if (max < mas->last) /* The equal sign here can be removed */
> Thanks.  I'll keep this in mind when I revisit the function.  I don't
> want to re-spin the patch for this alone.
>
>> 		mas->last = max;
>>
>>>    	mas->index = mas->last - size + 1;
>>>    	return 0;