crash in rb_insert_color()

From: Zoltan Menyhart
Date: Fri Feb 01 2008 - 12:27:46 EST


I have got a crash on an 8 processor ia64 machine with the kernel 2.6.18.8:

[<a00000010003cb30>] ia64_fault+0x14b0/0x1600
sp=e00000019ea5f930 bsp=e00000019ea51408
r32 : 000000000000001a r33 : 0000000400000020 r34 : 0000000000000010
r35 : 000000000003800f r36 : e00000019ea5fbb8 r37 : e00000019ea5fbb0
r38 : a00000010027f2e0 r39 : e00000019ea5f9c0 r40 : 000000000000000b
r41 : a00000010000c4e0 r42 : 0000000000000005 r43 : a000000100a90050
[<a00000010000c4e0>] ia64_leave_kernel+0x0/0x280
sp=e00000019ea5fb60 bsp=e00000019ea51408
[<a00000010027f2e0>] rb_insert_color+0x60/0x380
sp=e00000019ea5fd30 bsp=e00000019ea513b8
r32 : e0000004acb19448 r33 : e00000052a7ba100 r34 : e0000002799a5908
r35 : 0000000000000000 r36 : a0000002046f7810 r37 : 0000000000000690
r38 : 6f732e7365725f73 r39 : e0000004acb1c708 r40 : e00000052a7ba100
[<a0000002046f7810>] ext3_htree_store_dirent+0x270/0x2e0 [ext3]
sp=e00000019ea5fd30 bsp=e00000019ea51350
r32 : e00000052a7ba100 r33 : 0000000059bf4f42 r34 : 0000000026f5c228
r35 : e00000082bb9063c r36 : e0000004acb19440 r37 : e0000002799a5918
r38 : 000000000000003d r39 : e00000082bb90642 r40 : e0000004acb19444
r41 : e0000002799a5908 r42 : a000000204708a70 r43 : 0000000000000611
r44 : a000000204716448
[<a000000204708a70>] htree_dirblock_to_tree+0x1b0/0x280 [ext3]
sp=e00000019ea5fd30 bsp=e00000019ea512f0
r32 : e000000126191780 r33 : e00000082bb907f8 r34 : e00000082bb9063c
r35 : e00000019ea5fd50 r36 : 0000000000000000 r37 : 0000000000000000
r38 : e00000014ff15b68 r39 : 0000000000000018 r40 : e00000019ea5fd54
r41 : a000000204708c40 r42 : 0000000000000795 r43 : a000000204716448
[<a000000204708c40>] ext3_htree_fill_tree+0x100/0x460 [ext3]
sp=e00000019ea5fd40 bsp=e00000019ea51278
r32 : e000000126191780 r33 : 0000000000000000 r34 : 0000000000000000
r35 : e00000052a7ba128 r36 : a0000002046f82b0 r37 : e00000014ff060b8
r38 : 0000000000000000 r39 : 00000000000000d0 r40 : 60000000000101cc
r41 : 600ffffffa5ff5b0 r42 : 600ffffffa5ff618 r43 : 000000000000f000
r44 : a0000002046f82f0 r45 : 0000000000000fa6 r46 : a000000204716448
[<a0000002046f82f0>] ext3_readdir+0x890/0xcc0 [ext3]
sp=e00000019ea5fda0 bsp=e00000019ea51178
r32 : e000000126191780 r33 : e00000019ea5fe20 r34 : a000000100656570
r35 : e00000052a7ba108 r36 : e00000052a7ba100 r37 : e00000052a7ba124
r38 : e00000052a7ba120 r39 : e00000014ff06160 r40 : e000000126191830
r41 : e0000001261917b8 r42 : e00000052a7ba128 r43 : e000000126191790
r44 : e00000052a7ba118 r45 : e00000014ff060b8 r46 : e00000052a7ba124
r47 : e00000087a0b5980 r48 : e00000087a0b59a0 r49 : e00000052a7ba108
r50 : 0000000000000000 r51 : 0000000000000000 r52 : 0000000000000000
r53 : 0000000000000793 r54 : e00000019ea5fe00 r55 : 0000000000000001
r56 : 0000000000000793 r57 : 0000000000000000 r58 : 200000000003cc7c
r59 : a00000010017bd60 r60 : 000000000000058e r61 : a000000204716448
r62 : 000000001555a559
[<a00000010017bd60>] vfs_readdir+0x1a0/0x200
sp=e00000019ea5fe10 bsp=e00000019ea51120
r32 : e000000126191780 r33 : a000000100656570 r34 : e00000019ea5fe20
r35 : fffffffffffffffe r36 : e0000001261917a0 r37 : e00000014ff060b8
r38 : e000000126191790 r39 : e00000014ff06178 r40 : a00000010017c540
r41 : 0000000000000915 r42 : a000000100a90050

Here is the beginning of rb_insert_color():

void rb_insert_color(struct rb_node *node, struct rb_root *root)
{
struct rb_node *parent, *gparent;

while ((parent = rb_parent(node)) && rb_is_red(parent)) {
gparent = rb_parent(parent);
if (parent == gparent->rb_left) {

When it crashes, I've got the following nodes:

p/x *((struct rb_node *) 0xe0000004acb19448) // node
{
rb_parent_color = 0xe0000002799a5908,
rb_right = 0x0,
rb_left = 0x0
}
p/x *((struct rb_node *) 0xe0000002799a5908) // parent = rb_parent(node)
{
rb_parent_color = 0x0,
rb_right = 0x0,
rb_left = 0xa000000100650000 // seems to be valid
}

Note that "node" is not necessarily the same as the argument of
rb_insert_color() was.

gparent = rb_parent(parent) is NULL, and the kernel obviously traps
at if (parent == gparent->rb_left).

Does the algorithm guarantee that if a node is red, it must have a parent?
I guess yes.
(I have not seen any modification in rb_insert_color() since my 2.6.18.8.)

Is mutex_lock(&inode->i_mutex) in vfs_readdir() enough to protect the tree?
I guess yes.

I have got no idea how it can happen.
Has something been corrected since?

Regards,

Zoltan Menyhart


--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/