Re: [PATCH v2 01/10] maple_tree: Rework mtree_alloc_{range,rrange}()

["Liam R. Howlett" <Liam.Howlett@Oracle.com>]

* Peng Zhang <zhangpeng.00@bytedance.com> [230518 02:10]:
> 
> 
> 在 2023/5/18 02:17, Liam R. Howlett 写道:
> > * Peng Zhang <zhangpeng.00@bytedance.com> [230517 04:58]:
> > > Use mas_empty_area{_rev}() to refactor mtree_alloc_{range,rrange}()
> > > 
> > > Signed-off-by: Peng Zhang <zhangpeng.00@bytedance.com>
> > > ---
> > >   lib/maple_tree.c | 63 ++++++++++++++++++++++++------------------------
> > >   1 file changed, 32 insertions(+), 31 deletions(-)
> > > 
> > > diff --git a/lib/maple_tree.c b/lib/maple_tree.c
> > > index 4eb220008f72..e1820e90f167 100644
> > > --- a/lib/maple_tree.c
> > > +++ b/lib/maple_tree.c
> > > @@ -6493,32 +6493,31 @@ int mtree_alloc_range(struct maple_tree *mt, unsigned long *startp,
> > >   {
> > >   	int ret = 0;
> > > -	MA_STATE(mas, mt, min, min);
> > > +	MA_STATE(mas, mt, 0, 0);
> > >   	if (!mt_is_alloc(mt))
> > >   		return -EINVAL;
> > >   	if (WARN_ON_ONCE(mt_is_reserved(entry)))
> > >   		return -EINVAL;
> > > -	if (min > max)
> > > -		return -EINVAL;
> > > -
> > > -	if (max < size)
> > > -		return -EINVAL;
> > > -
> > > -	if (!size)
> > > -		return -EINVAL;
> > > -
> > >   	mtree_lock(mt);
> > >   retry:
> > > -	mas.offset = 0;
> > > -	mas.index = min;
> > > -	mas.last = max - size + 1;
> > > -	ret = mas_alloc(&mas, entry, size, startp);
> > > -	if (mas_nomem(&mas, gfp))
> > > -		goto retry;
> > > -
> > > +	ret = mas_empty_area(&mas, min, max, size);
> > > +	if (!ret) {
> > > +		mas_insert(&mas, entry);
> > > +		/*
> > > +		 * mas_nomem() may release the lock, causing the allocated area
> > > +		 * to be unavailable, so try to allocate a free area again.
> > > +		 */
> > > +		if (mas_nomem(&mas, gfp))
> > > +			goto retry;
> > > +	}
> > >   	mtree_unlock(mt);
> > > +	if (!ret) {
> > 
> > Checking for a mas_is_err() should probably be outside if (!ret)
> > statement.  If mas_insert() returns something besides ENOMEM, we will
> > not detect the error.  I'm not sure if this is possible today since this
> > should never return an -EEXISTS, but having it this way doesn't add much
> > to the overhead.
> I don't think there will be error that can't be detected here.
> In fact, there are two sources of errors:
> 
> 1. mas_empty_area(), the error number is in the variable ret,
>    and may also be in mas->node, but ret must contain all errors.
> 
> 2. mas_insert(), the error number is in mas->node
> 
> When we check errors, we should first check errors from
> mas_empty_area(). If there is no error in mas_empty_area(), we
> will check errors from mas_insert().
> So, mas_is_err() is inside the if (!ret) statement, no problem here.
> 
> Of course, even if mas_insert() returns -EEXISTS, it can be detected
> under the current encoding, because "if (!ret)" is true in this case.
> But I don't think this can happen, if it happens, it's a bug of maple
> tree.

Right, thanks.

> 
> I don't think it's good to put mas_is_err() outside, because the error
> number stored in mas->node may come from mas_empty_area(). We should use
> the ret variable to detect the error from mas_empty_area() first.

Yeah, I would structure it as a 'goto' to undo the locking to avoid the
if (!ret) nesting, and move the unlock to just before the return:

if (ret)
	goto unlock;

...
	*startp = mas.index;

unlock:
	mtree_unlock(mt);
	return ret;

mas_empty_area() is an external interface and so already decodes the
error in the node using mas_is_err(), so this can be dropped completely.

I don't think holding the lock for the one extra assignment will make a
difference.

> > 
> > > +		if (mas_is_err(&mas))
> > > +			return xa_err(mas.node);
> > > +		*startp = mas.index;
> > > +	}
> > >   	return ret;
> > >   }
> > >   EXPORT_SYMBOL(mtree_alloc_range);
> > > @@ -6529,29 +6528,31 @@ int mtree_alloc_rrange(struct maple_tree *mt, unsigned long *startp,
> > >   {
> > >   	int ret = 0;
> > > -	MA_STATE(mas, mt, min, max - size + 1);
> > > +	MA_STATE(mas, mt, 0, 0);
> > >   	if (!mt_is_alloc(mt))
> > >   		return -EINVAL;
> > >   	if (WARN_ON_ONCE(mt_is_reserved(entry)))
> > >   		return -EINVAL;
> > > -	if (min > max)
> > > -		return -EINVAL;
> > > -
> > > -	if (max < size - 1)
> > > -		return -EINVAL;
> > > -
> > > -	if (!size)
> > > -		return -EINVAL;
> > > -
> > >   	mtree_lock(mt);
> > >   retry:
> > > -	ret = mas_rev_alloc(&mas, min, max, entry, size, startp);
> > > -	if (mas_nomem(&mas, gfp))
> > > -		goto retry;
> > > -
> > > +	ret = mas_empty_area_rev(&mas, min, max, size);
> > > +	if (!ret) {
> > > +		mas_insert(&mas, entry);
> > > +		/*
> > > +		 * mas_nomem() may release the lock, causing the allocated area
> > > +		 * to be unavailable, so try to allocate a free area again.
> > > +		 */
> > > +		if (mas_nomem(&mas, gfp))
> > > +			goto retry;
> > > +	}
> > >   	mtree_unlock(mt);
> > > +	if (!ret) {
> > 
> > Same here.
> > 
> > > +		if (mas_is_err(&mas))
> > > +			return xa_err(mas.node);
> > > +		*startp = mas.index;
> > > +	}
> > >   	return ret;
> > >   }
> > >   EXPORT_SYMBOL(mtree_alloc_rrange);
> > > -- 
> > > 2.20.1
> > >