riscv32 EXT4 splat, 6.8 regression?

[Nam Cao]

On 2024-04-12 Björn Töpel wrote:
> Hi!
> 
> I've been looking at an EXT4 splat on riscv32, that LKFT found [1]:
> 
>   | EXT4-fs (vda): mounted filesystem 13697a42-d10e-4a9e-8e56-cb9083be92f9 ro with ordered data mode. Quota mode: disabled.
>   | VFS: Mounted root (ext4 filesystem) readonly on device 254:0.
>   | Unable to handle kernel NULL pointer dereference at virtual address 00000006
>   | Oops [#1]
>   | Modules linked in:
>   | CPU: 1 PID: 1 Comm: swapper/0 Not tainted 6.8.0 #41
>   | Hardware name: riscv-virtio,qemu (DT)
>   | epc : ext4_search_dir+0x52/0xe4
>   |  ra : __ext4_find_entry+0x1d6/0x578
>   | epc : c035b60e ra : c035b876 sp : c253fc10
>   |  gp : c21a7380 tp : c25c8000 t0 : 44c0657f
>   |  t1 : 0000000c t2 : 1de5b089 s0 : c253fc50
>   |  s1 : 00000000 a0 : fffffffc a1 : fffff000
>   |  a2 : 00000000 a3 : c29c04f8 a4 : c253fd00
>   |  a5 : 00000000 a6 : c253fcfc a7 : fffffff3
>   |  s2 : 00001000 s3 : 00000000 s4 : 00001000
>   |  s5 : c29c04f8 s6 : c292db40 s7 : c253fcfc
>   |  s8 : fffffff7 s9 : c253fd00 s10: fffff000
>   |  s11: c292db40 t3 : 00000007 t4 : 5e8b4525
>   |  t5 : 00000000 t6 : 00000000
>   | status: 00000120 badaddr: 00000006 cause: 0000000d
>   | [<c035b60e>] ext4_search_dir+0x52/0xe4
>   | [<c035b876>] __ext4_find_entry+0x1d6/0x578
>   | [<c035bcaa>] ext4_lookup+0x92/0x200
>   | [<c0295c14>] __lookup_slow+0x8e/0x142
>   | [<c029943a>] walk_component+0x104/0x174
>   | [<c0299f18>] path_lookupat+0x78/0x182
>   | [<c029b24c>] filename_lookup+0x96/0x158
>   | [<c029b346>] kern_path+0x38/0x56
>   | [<c0c1bee4>] init_mount+0x46/0x96
>   | [<c0c2ae1c>] devtmpfs_mount+0x44/0x7a
>   | [<c0c01c26>] prepare_namespace+0x226/0x27c
>   | [<c0c01130>] kernel_init_freeable+0x27e/0x2a0
>   | [<c0b78402>] kernel_init+0x2a/0x158
>   | [<c0b82bf2>] ret_from_fork+0xe/0x20
>   | Code: 84ae a809 d303 0044 949a 0f63 0603 991a fd63 0584 (c603) 0064 
>   | ---[ end trace 0000000000000000 ]---
>   | Kernel panic - not syncing: Attempted to kill init! exitcode=0x0000000b
> 
> This was not present in 6.7. Bisection wasn't really helpful (to me at
> least); I got it down to commit c604110e662a ("Merge tag 'vfs-6.8.misc'
> of git://git.kernel.org/pub/scm/linux/kernel/git/vfs/vfs"), and when I
> revert the commits in the vfs merge the splat went away, but I *really*
> struggle to see how those are related...
> 
> What I see in ext4_search_dir() is that search_buf is 0xfffff000, and at
> some point the address wraps to zero, and boom. I doubt that 0xfffff000
> is a sane address.

I have zero knowledge about file system, but I think it's an integer
overflow problem. The calculation of "dlimit" overflow and dlimit wraps
around, this leads to wrong comparison later on.

I guess that explains why your bisect and Conor's bisect results are
strange: the bug has been here for quite some time, but it only appears
when "dlimit" happens to overflow.

It can be fixed by re-arrange the comparisons a bit. Can you give the
below patch a try?

Best regards,
Nam

diff --git a/fs/ext4/namei.c b/fs/ext4/namei.c
index 05b647e6bc19..71b88b33b676 100644
--- a/fs/ext4/namei.c
+++ b/fs/ext4/namei.c
@@ -1532,15 +1532,13 @@ int ext4_search_dir(struct buffer_head *bh, char *search_buf, int buf_size,
 		    unsigned int offset, struct ext4_dir_entry_2 **res_dir)
 {
 	struct ext4_dir_entry_2 * de;
-	char * dlimit;
 	int de_len;
 
 	de = (struct ext4_dir_entry_2 *)search_buf;
-	dlimit = search_buf + buf_size;
-	while ((char *) de < dlimit - EXT4_BASE_DIR_LEN) {
+	while ((char *) de - search_buf < buf_size - EXT4_BASE_DIR_LEN) {
 		/* this code is executed quadratically often */
 		/* do minimal checking `by hand' */
-		if (de->name + de->name_len <= dlimit &&
+		if (de->name + de->name_len - search_buf <= buf_size &&
 		    ext4_match(dir, fname, de)) {
 			/* found a match - just to be sure, do
 			 * a full check */