On Fri, 2023-06-02 at 10:42 -0300, Arnaldo Carvalho de Melo wrote:
> 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?
Right, you are correct.
The 'structures__tree = RB_ROOT' part is still necessary, though.
If you are ok with overall structure of the patch I can resend it w/o bzero().
>
> - 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
> >
>
