Re: [BUG kernel-5.15] aarch64: __pi_strncmp() out-of-bound error

[Will Deacon]

On Thu, Aug 10, 2023 at 01:23:28PM +0100, Robin Murphy wrote:
> On 2023-08-07 16:32, Mark Rutland wrote:
> > On Mon, Aug 07, 2023 at 12:31:45PM +0000, John Hsu (許永翰) wrote:
> > > [ 7445.269347][  T382] ueventd: Hardware name: MT6886(ENG) (DT)
> > > [ 7445.269354][  T382] ueventd: Call trace:
> > > [ 7445.269359][  T382] ueventd:  dump_backtrace+0x0/0x2a8
> > > [ 7445.269374][  T382] ueventd:  dump_stack_lvl+0x74/0xa4
> > > [ 7445.269384][  T382] ueventd:  dump_stack+0x14/0x1c
> > > [ 7445.269391][  T382] ueventd:  mrdump_common_die+0x32c/0x5ac [mrdump]
> > > [ 7445.269470][  T382] ueventd:  ipanic_die+0x1c/0x28 [mrdump]
> > > [ 7445.269539][  T382] ueventd:  __die+0xbc/0x308
> > > [ 7445.269548][  T382] ueventd:  die+0xd8/0x500
> > > [ 7445.269556][  T382] ueventd:  die_kernel_fault+0x94/0xa8
> > > [ 7445.269565][  T382] ueventd:  __do_kernel_fault+0x1d8/0x214
> > > [ 7445.269571][  T382] ueventd:  do_bad_area+0x40/0x174
> > > [ 7445.269579][  T382] ueventd:  do_translation_fault+0x48/0x54
> > > [ 7445.269585][  T382] ueventd:  do_mem_abort+0x3c/0x100
> > > [ 7445.269592][  T382] ueventd:  el1_abort+0x38/0x54
> > > [ 7445.269602][  T382] ueventd:  el1h_64_sync_handler+0x54/0x88
> > > [ 7445.269610][  T382] ueventd:  el1h_64_sync+0x78/0x7c
> > > [ 7445.269618][  T382] ueventd:  __pi_strncmp+0x1a0/0x1c4
> > > [ 7445.269626][  T382] ueventd:  selinux_genfs_get_sid+0x114/0x220
> > > [ 7445.269636][  T382] ueventd:  inode_doinit_with_dentry+0x3d0/0x598
> > > [ 7445.269644][  T382] ueventd:  selinux_d_instantiate+0x1c/0x24
> > > [ 7445.269652][  T382] ueventd:  d_splice_alias+0x5c/0x280
> > > [ 7445.269662][  T382] ueventd:  kernfs_iop_lookup+0xec/0x21c
> > > [ 7445.269674][  T382] ueventd:  __lookup_slow+0xc4/0x150
> > > [ 7445.269684][  T382] ueventd:  lookup_slow+0x40/0xf0
> > > [ 7445.269690][  T382] ueventd:  walk_component+0x144/0x160
> > > [ 7445.269696][  T382] ueventd:  link_path_walk+0x25c/0x344
> > > [ 7445.269703][  T382] ueventd:  path_lookupat+0x64/0x120
> > > [ 7445.269710][  T382] ueventd:  filename_lookup+0xc4/0x1b0
> > > [ 7445.269718][  T382] ueventd:  user_path_at_empty+0x48/0xb4
> > > [ 7445.269725][  T382] ueventd:  do_faccessat+0xa8/0x1f0
> > > [ 7445.269732][  T382] ueventd:  __arm64_sys_faccessat+0x20/0x28
> > > [ 7445.269738][  T382] ueventd:  invoke_syscall+0x3c/0xf0
> > > [ 7445.269746][  T382] ueventd:  el0_svc_common+0x84/0xe8
> > > [ 7445.269753][  T382] ueventd:  do_el0_svc+0x20/0x84
> > > [ 7445.269759][  T382] ueventd:  el0_svc+0x1c/0x48
> > > [ 7445.269766][  T382] ueventd:  el0t_64_sync_handler+0x7c/0xd8
> > > [ 7445.269773][  T382] ueventd:  el0t_64_sync+0x15c/0x160
> > > We found that we hit this issue when we compare these two strings.
> > > ________________address|_0__1__2__3__4__5__6__7__8__9__A__B__C__D__E__F
> > > __   0123456789ABCDEF
> > >     NSD:FFFFFF80089EDA00|>2F 64 65 76 69 63 65 73 2F 76 69 72 74 75 61
> > > 6C  /devices/virtual
> > >     NSD:FFFFFF80089EDA10| 2F 62 6C 6F 63 6B 2F 00 E0 03 01 AA E1 03 02
> > > AA  /block/.........
> > > ________________address|_0__1__2__3__4__5__6__7__8__9__A__B__C__D__E__F
> > > __    0123456789ABCDEF
> > >     NSD:FFFFFF803FD3EFE0| 00 00 00 00 00 00 00 00 00 00 2F 64 65 76 69
> > > 63    ........../devic
> > >     NSD:FFFFFF803FD3EFF0| 65 73>2F 76 69 72 74 75 61 6C 2F 6D 69 73 63
> > > 00  es/virtual/misc.
> > >     NSD:FFFFFF803FD3F000| ?? ?? ?? ?? ?? ?? ?? ?? ?? ?? ?? ?? ?? ?? ??
> > > ??
> > >     NSD:FFFFFF803FD3F0E0| ?? ?? ?? ?? ?? ?? ?? ?? ?? ?? ?? ?? ?? ?? ??
> > > ??
> > > 
> > > We observe the second string is put at the tail of the first page and
> > > the next page is unreadable.
> > > Thus, we made a simple test as below and it can reproduce this issue.
> 
> I'm not sure there's strictly a bug here. The C standard says:
> 
> "The strncmp function compares not more than n characters (characters that
> follow a null character are not compared) ..."
> 
> so although any characters between the first NULL and n must not be
> considered for the result of the comparison, there doesn't seem to be any
> explicit promise anywhere that they can't be *accessed*. AFAICT what happens
> here is in the request to compare at most 23 characters, it ends up in the
> do_misaligned case, loop_misaligned runs twice and finds no differences or
> NULLs in characters 0-7 and 8-15, so then done_loop loads characters 15-23
> to compare the last 7, and is tripped up by 22-23 not actually existing in
> src2. Possibly the original intent was that this case should have ended up
> in page_end_loop, and the condition for that was slightly off, but I'm not
> sure, and this code is obsolete now anyway.

The long backtrace above worries me, as it suggests that you can trigger
this from userspace. In that case I think it's a bug regardless of what
the C standard says.

Perhaps we should just fallback to the generic strncmp() implementation
in the stable kernel? I couldn't spot any other architectures doing
anything particularly clever, and x86_64 looks like it doesn't select
__HAVE_ARCH_STRNCMP at all.

Will