Em Fri, May 26, 2023 at 02:59:49AM +0300, Eduard Zingerman escreveu:
> When pahole is executed in '-F dwarf --sort' mode there are two places
> where 'struct structure' instance could be added to the rb_tree:
>
> The first is triggered from the following call stack:
>
> print_classes()
> structures__add()
> __structures__add()
> (adds to global pahole.c:structures__tree)
>
> The second is triggered from the following call stack:
>
> print_ordered_classes()
> resort_classes()
> resort_add()
> (adds to local rb_tree instance)
>
> Both places use the same 'struct structure::rb_node' field, so if both
> code pathes are executed the final state of the 'structures__tree'
> might be inconsistent.
>
> For example, this could be observed when DEBUG_CHECK_LEAKS build flag
> is set. Here is the command line snippet that eventually leads to a
> segfault:
>
> $ for i in $(seq 1 100); do \
> echo $i; \
> pahole -F dwarf --flat_arrays --sort --jobs vmlinux > /dev/null \
> || break; \
> done
>
> GDB shows the following stack trace:
>
> Thread 1 "pahole" received signal SIGSEGV, Segmentation fault.
> 0x00007ffff7f819ad in __rb_erase_color (node=0x7fffd4045830, parent=0x0, root=0x5555555672d8 <structures.tree>) at /home/eddy/work/dwarves-fork/rbtree.c:134
> 134 if (parent->rb_left == node)
> (gdb) bt
> #0 0x00007ffff7f819ad in __rb_erase_color (node=0x7fffd4045830, parent=0x0, root=0x5555555672d8 <structures.tree>) at /home/eddy/work/dwarves-fork/rbtree.c:134
> #1 0x00007ffff7f82014 in rb_erase (node=0x7fff21ae5b80, root=0x5555555672d8 <structures.tree>) at /home/eddy/work/dwarves-fork/rbtree.c:275
> #2 0x0000555555559c3d in __structures__delete () at /home/eddy/work/dwarves-fork/pahole.c:440
> #3 0x0000555555559c70 in structures__delete () at /home/eddy/work/dwarves-fork/pahole.c:448
> #4 0x0000555555560bb6 in main (argc=13, argv=0x7fffffffdcd8) at /home/eddy/work/dwarves-fork/pahole.c:3584
>
> This commit modifies resort_classes() to re-use 'structures__tree' and
> to reset 'rb_node' fields before adding structure instances to the
> tree for a second time.
>
> Lock/unlock structures_lock to be consistent with structures_add() and
> structures__delete() code.
>
> Signed-off-by: Eduard Zingerman <eddyz87@xxxxxxxxx>
> ---
> pahole.c | 41 ++++++++++++++++++++++++++++-------------
> 1 file changed, 28 insertions(+), 13 deletions(-)
>
> diff --git a/pahole.c b/pahole.c
> index 6fc4ed6..576733f 100644
> --- a/pahole.c
> +++ b/pahole.c
> @@ -621,9 +621,9 @@ static void print_classes(struct cu *cu)
> }
> }
>
> -static void __print_ordered_classes(struct rb_root *root)
> +static void __print_ordered_classes(void)
> {
> - struct rb_node *next = rb_first(root);
> + struct rb_node *next = rb_first(&structures__tree);
>
> while (next) {
> struct structure *st = rb_entry(next, struct structure, rb_node);
> @@ -660,24 +660,39 @@ static void resort_add(struct rb_root *resorted, struct structure *str)
> rb_insert_color(&str->rb_node, resorted);
> }
>
> -static void resort_classes(struct rb_root *resorted, struct list_head *head)
> +static void resort_classes(void)
> {
> struct structure *str;
>
> - list_for_each_entry(str, head, node)
> - resort_add(resorted, str);
> + pthread_mutex_lock(&structures_lock);
> +
> + /* The need_resort flag is set by type__compare_members()
> + * within the following call stack:
> + *
> + * print_classes()
> + * structures__add()
> + * __structures__add()
> + * type__compare()
> + *
> + * The call to structures__add() registers 'struct structures'
> + * instances in both 'structures__tree' and 'structures__list'.
> + * In order to avoid adding same node to the tree twice reset
> + * both the 'structures__tree' and 'str->rb_node'.
> + */
> + structures__tree = RB_ROOT;
> + list_for_each_entry(str, &structures__list, node) {
> + bzero(&str->rb_node, sizeof(str->rb_node));
Why is this bzero needed?
> + resort_add(&structures__tree, str);
resort_add will call rb_link_node(&str->rb_node, parent, p); and it, in
turn:
static inline void rb_link_node(struct rb_node * node, struct rb_node * parent,
struct rb_node ** rb_link)
{
node->rb_parent_color = (unsigned long )parent;
node->rb_left = node->rb_right = NULL;
*rb_link = node;
}
And:
struct rb_node
{
unsigned long rb_parent_color;
#define RB_RED 0
#define RB_BLACK 1
struct rb_node *rb_right;
struct rb_node *rb_left;
} __attribute__((aligned(sizeof(long))))
So all the fields are being initialized in the operation right after the
bzero(), no?
- Arnaldo
> + }
> +
> + pthread_mutex_unlock(&structures_lock);
> }
>
> static void print_ordered_classes(void)
> {
> - if (!need_resort) {
> - __print_ordered_classes(&structures__tree);
> - } else {
> - struct rb_root resorted = RB_ROOT;
> -
> - resort_classes(&resorted, &structures__list);
> - __print_ordered_classes(&resorted);
> - }
> + if (need_resort)
> + resort_classes();
> + __print_ordered_classes();
> }
>
>
> --
> 2.40.1
>
--
- Arnaldo
