On 8/4/23 02:02, Michael Tokarev wrote:
01.08.2023 00:02, Richard Henderson wrote:
Ensure that the stores to rb_left and rb_right are complete before
inserting the new node into the tree. Otherwise a concurrent reader
could see garbage in the new leaf.
Cc: qemu-sta...@nongnu.org
Reviewed-by: Peter Maydell <peter.mayd...@linaro.org>
Signed-off-by: Richard Henderson <richard.hender...@linaro.org>
---
util/interval-tree.c | 6 +++++-
1 file changed, 5 insertions(+), 1 deletion(-)
diff --git a/util/interval-tree.c b/util/interval-tree.c
index 5a0ad21b2d..759562db7d 100644
--- a/util/interval-tree.c
+++ b/util/interval-tree.c
@@ -128,7 +128,11 @@ static inline void rb_link_node(RBNode *node, RBNode *parent,
RBNode **rb_link)
node->rb_parent_color = (uintptr_t)parent;
node->rb_left = node->rb_right = NULL;
- qatomic_set(rb_link, node);
+ /*
+ * Ensure that node is initialized before insertion,
+ * as viewed by a concurrent search.
+ */
+ qatomic_set_mb(rb_link, node);
FWIW, there's no qatomic_set_mb() in 8.0 and before, so this can not be
directly applied to stable-8.0. This commit is missing in 8.0 before
qatomic_set_mb() can be used:
commit 06831001ac8949b0801e0d20c347d97339769a20
Author: Paolo Bonzini <pbonz...@redhat.com>
Date: Fri Mar 3 14:37:51 2023 +0100
atomics: eliminate mb_read/mb_set
I don't think it's a good idea to back-port this commit to stable-8.0.
How do you think we can solve this for 8.0?
The function is called qatomic_mb_set instead of qatomic_set_mb in stable-8.0.
r~
