[PATCH 0/2] Fix mas_alloc_cyclic retry

[PATCH 0/2] Fix mas_alloc_cyclic retry

[Liam R. Howlett]

This is a follow-up patch and test for Yang Erkun.  lore seems to have
eaten the original email [1], so I'll add akpm's link as well [2].

Patch 1 applies on top of Yang Erkun's fix, so feel free to squash it
in.

Patch 2 adds a test to the maple tree infrastructure that will catch the
issue if it comes back.

[1] https://lore.kernel.org/all/20241214093005.72284-1-yangerkun@huaweicloud.com/T/#u
[2] https://lore.kernel.org/all/20241216060600.287B4C4CED0@smtp.kernel.org/

Liam R. Howlett (2):
  maple_tree: Fix mas_alloc_cyclic() second search
  test_maple_tree: Test exhausted upper limit of mtree_alloc_cyclic()

 lib/maple_tree.c      |  3 +--
 lib/test_maple_tree.c | 28 ++++++++++++++++++++++++++++
 2 files changed, 29 insertions(+), 2 deletions(-)

-- 
2.43.0

[PATCH 1/2] maple_tree: Fix mas_alloc_cyclic() second search

[Liam R. Howlett]

From: "Liam R. Howlett" <Liam.Howlett@Oracle.com>

The first search may leave the maple state in an error state.  Reset the
maple state before the second search so that the search has a chance of
executing correctly after an exhausted first search.

Link: https://lore.kernel.org/all/20241216060600.287B4C4CED0@smtp.kernel.org/
Fixes: 9b6713cc7522 ("maple_tree: Add mtree_alloc_cyclic()")
Cc: Yang Erkun <yangerkun@huaweicloud.com>
Cc: chuck.lever@oracle.com
Cc: brauner@kernel.org
Signed-off-by: Liam R. Howlett <Liam.Howlett@Oracle.com>
---
 lib/maple_tree.c | 3 +--
 1 file changed, 1 insertion(+), 2 deletions(-)

diff --git a/lib/maple_tree.c b/lib/maple_tree.c
index 940d30653407b..f7153ade1be5f 100644
--- a/lib/maple_tree.c
+++ b/lib/maple_tree.c
@@ -4335,7 +4335,6 @@ int mas_alloc_cyclic(struct ma_state *mas, unsigned long *startp,
 {
 	unsigned long min = range_lo;
 	int ret = 0;
-	struct ma_state m = *mas;
 
 	range_lo = max(min, *next);
 	ret = mas_empty_area(mas, range_lo, range_hi, 1);
@@ -4344,7 +4343,7 @@ int mas_alloc_cyclic(struct ma_state *mas, unsigned long *startp,
 		ret = 1;
 	}
 	if (ret < 0 && range_lo > min) {
-		*mas = m;
+		mas_reset(mas);
 		ret = mas_empty_area(mas, min, range_hi, 1);
 		if (ret == 0)
 			ret = 1;
-- 
2.43.0

Re: [PATCH 1/2] maple_tree: Fix mas_alloc_cyclic() second search

[yangerkun]

在 2024/12/17 3:01, Liam R. Howlett 写道:
> From: "Liam R. Howlett" <Liam.Howlett@Oracle.com>
> 
> The first search may leave the maple state in an error state.  Reset the
> maple state before the second search so that the search has a chance of
> executing correctly after an exhausted first search.
> 
> Link: https://lore.kernel.org/all/20241216060600.287B4C4CED0@smtp.kernel.org/
> Fixes: 9b6713cc7522 ("maple_tree: Add mtree_alloc_cyclic()")
> Cc: Yang Erkun <yangerkun@huaweicloud.com>
> Cc: chuck.lever@oracle.com
> Cc: brauner@kernel.org
> Signed-off-by: Liam R. Howlett <Liam.Howlett@Oracle.com>
> ---
>   lib/maple_tree.c | 3 +--
>   1 file changed, 1 insertion(+), 2 deletions(-)

This looks more clear!

Reviewed-by: Yang Erkun <yangerkun@huawei.com>

> 
> diff --git a/lib/maple_tree.c b/lib/maple_tree.c
> index 940d30653407b..f7153ade1be5f 100644
> --- a/lib/maple_tree.c
> +++ b/lib/maple_tree.c
> @@ -4335,7 +4335,6 @@ int mas_alloc_cyclic(struct ma_state *mas, unsigned long *startp,
>   {
>   	unsigned long min = range_lo;
>   	int ret = 0;
> -	struct ma_state m = *mas;
>   
>   	range_lo = max(min, *next);
>   	ret = mas_empty_area(mas, range_lo, range_hi, 1);
> @@ -4344,7 +4343,7 @@ int mas_alloc_cyclic(struct ma_state *mas, unsigned long *startp,
>   		ret = 1;
>   	}
>   	if (ret < 0 && range_lo > min) {
> -		*mas = m;
> +		mas_reset(mas);
>   		ret = mas_empty_area(mas, min, range_hi, 1);
>   		if (ret == 0)
>   			ret = 1;

[PATCH 2/2] test_maple_tree: Test exhausted upper limit of mtree_alloc_cyclic()

[Liam R. Howlett]

From: "Liam R. Howlett" <Liam.Howlett@Oracle.com>

When the upper bound of the search is exhausted, the maple state may be
returned in an error state of -EBUSY.  This means maple state needs to
be reset before the second search in mas_alloc_cylic() to ensure the
search happens.  This test ensures the issue is not recreated.

Cc: Yang Erkun <yangerkun@huaweicloud.com>
Cc: chuck.lever@oracle.com
Cc: brauner@kernel.org
Signed-off-by: Liam R. Howlett <Liam.Howlett@Oracle.com>
---
 lib/test_maple_tree.c | 28 ++++++++++++++++++++++++++++
 1 file changed, 28 insertions(+)

diff --git a/lib/test_maple_tree.c b/lib/test_maple_tree.c
index 72bda304b5952..13e2a10d7554d 100644
--- a/lib/test_maple_tree.c
+++ b/lib/test_maple_tree.c
@@ -3738,6 +3738,34 @@ static noinline void __init alloc_cyclic_testing(struct maple_tree *mt)
 	}
 
 	mtree_destroy(mt);
+
+	/*
+	 * Issue with reverse search was discovered
+	 * https://lore.kernel.org/all/20241216060600.287B4C4CED0@smtp.kernel.org/
+	 * Exhausting the allocation area and forcing the search to wrap needs a
+	 * mas_reset() in mas_alloc_cyclic().
+	 */
+	next = 0;
+	mt_init_flags(mt, MT_FLAGS_ALLOC_RANGE);
+	for (int i = 0; i < 1023; i++) {
+		mtree_alloc_cyclic(mt, &location, mt, 2, 1024, &next, GFP_KERNEL);
+		MT_BUG_ON(mt, i != location - 2);
+		MT_BUG_ON(mt, i != next - 3);
+		MT_BUG_ON(mt, mtree_load(mt, location) != mt);
+	}
+	mtree_erase(mt, 123);
+	MT_BUG_ON(mt, mtree_load(mt, 123) != NULL);
+	mtree_alloc_cyclic(mt, &location, mt, 2, 1024, &next, GFP_KERNEL);
+	MT_BUG_ON(mt, 123 != location);
+	MT_BUG_ON(mt, 124 != next);
+	MT_BUG_ON(mt, mtree_load(mt, location) != mt);
+	mtree_erase(mt, 100);
+	mtree_alloc_cyclic(mt, &location, mt, 2, 1024, &next, GFP_KERNEL);
+	MT_BUG_ON(mt, 100 != location);
+	MT_BUG_ON(mt, 101 != next);
+	MT_BUG_ON(mt, mtree_load(mt, location) != mt);
+	mtree_destroy(mt);
+
 	/* Overflow test */
 	next = ULONG_MAX - 1;
 	ret = mtree_alloc_cyclic(mt, &location, mt, 2, ULONG_MAX, &next, GFP_KERNEL);
-- 
2.43.0

Re: [PATCH 2/2] test_maple_tree: Test exhausted upper limit of mtree_alloc_cyclic()

[yangerkun]

在 2024/12/17 3:01, Liam R. Howlett 写道:
> From: "Liam R. Howlett" <Liam.Howlett@Oracle.com>
> 
> When the upper bound of the search is exhausted, the maple state may be
> returned in an error state of -EBUSY.  This means maple state needs to
> be reset before the second search in mas_alloc_cylic() to ensure the
> search happens.  This test ensures the issue is not recreated.
> 
> Cc: Yang Erkun <yangerkun@huaweicloud.com>
> Cc: chuck.lever@oracle.com
> Cc: brauner@kernel.org
> Signed-off-by: Liam R. Howlett <Liam.Howlett@Oracle.com>
> ---
>   lib/test_maple_tree.c | 28 ++++++++++++++++++++++++++++
>   1 file changed, 28 insertions(+)
> 

Thanks for the testcase!

Reviewed-by: Yang Erkun <yangerkun@huawei.com>

> diff --git a/lib/test_maple_tree.c b/lib/test_maple_tree.c
> index 72bda304b5952..13e2a10d7554d 100644
> --- a/lib/test_maple_tree.c
> +++ b/lib/test_maple_tree.c
> @@ -3738,6 +3738,34 @@ static noinline void __init alloc_cyclic_testing(struct maple_tree *mt)
>   	}
>   
>   	mtree_destroy(mt);
> +
> +	/*
> +	 * Issue with reverse search was discovered
> +	 * https://lore.kernel.org/all/20241216060600.287B4C4CED0@smtp.kernel.org/
> +	 * Exhausting the allocation area and forcing the search to wrap needs a
> +	 * mas_reset() in mas_alloc_cyclic().
> +	 */
> +	next = 0;
> +	mt_init_flags(mt, MT_FLAGS_ALLOC_RANGE);
> +	for (int i = 0; i < 1023; i++) {
> +		mtree_alloc_cyclic(mt, &location, mt, 2, 1024, &next, GFP_KERNEL);
> +		MT_BUG_ON(mt, i != location - 2);
> +		MT_BUG_ON(mt, i != next - 3);
> +		MT_BUG_ON(mt, mtree_load(mt, location) != mt);
> +	}
> +	mtree_erase(mt, 123);
> +	MT_BUG_ON(mt, mtree_load(mt, 123) != NULL);
> +	mtree_alloc_cyclic(mt, &location, mt, 2, 1024, &next, GFP_KERNEL);
> +	MT_BUG_ON(mt, 123 != location);
> +	MT_BUG_ON(mt, 124 != next);
> +	MT_BUG_ON(mt, mtree_load(mt, location) != mt);
> +	mtree_erase(mt, 100);
> +	mtree_alloc_cyclic(mt, &location, mt, 2, 1024, &next, GFP_KERNEL);
> +	MT_BUG_ON(mt, 100 != location);
> +	MT_BUG_ON(mt, 101 != next);
> +	MT_BUG_ON(mt, mtree_load(mt, location) != mt);
> +	mtree_destroy(mt);
> +
>   	/* Overflow test */
>   	next = ULONG_MAX - 1;
>   	ret = mtree_alloc_cyclic(mt, &location, mt, 2, ULONG_MAX, &next, GFP_KERNEL);