Re: RCU bug with v3.17-rc3 ?

[Rabin Vincent <rabin@rab.in>]

On Thu, Oct 09, 2014 at 11:26:56AM -0500, Felipe Balbi wrote:
> alright, it's pretty deterministic however. Always on the same test, no
> matter which USB controller, no matter if backing store is RAM or MMC.
> 
> Those two undefined instructions on the disassembly caught my attention,
> perhaps I'm facing a GCC bug ?

The undefined instructions are just ARM's BUG() implementation.

But did you see the question I asked you yesterday in your other thread?
http://www.spinics.net/lists/arm-kernel/msg368634.html

Here it is again:

  What GCC version are you using?
  
  4.8.1 and 4.8.2 are known to miscompile the ARM kernel and these
  find_get_entry() crashes with 0xffffffff involved smell a lot like the
  earlier reports from kernels build with those compilers:
  
  https://lkml.org/lkml/2014/6/25/456
  https://lkml.org/lkml/2014/6/30/375
  https://lkml.org/lkml/2014/6/30/660
  https://gcc.gnu.org/bugzilla/show_bug.cgi?id=58854
  https://lkml.org/lkml/2014/5/9/330

Also, I didn't see any public email making a definitive link between GCC
PR 58854 that Nathan pointed out in https://lkml.org/lkml/2014/6/30/660
and the earlier find_get_entry() crashes, but I just built GCC 4.8.1 and
an ARM kernel with that, and the GCC bug is clearly seen in
radix_tree_lookup_slot() which returns the pointer which
find_get_entry() is dereferencing:

  <radix_tree_lookup_slot>:
   e1a0c00d  mov     ip, sp
   e92dd800  push    {fp, ip, lr, pc}
   e24cb004  sub     fp, ip, #4
   e24dd008  sub     sp, sp, #8
   e3a02000  mov     r2, #0
   e24b3010  sub     r3, fp, #16
   ebffffc5  bl      c0176ab8 <__radix_tree_lookup>
   e24bd00c  sub     sp, fp, #12		<--- sp moved up
   e3500000  cmp     r0, #0
   151b0010  ldrne   r0, [fp, #-16]		<--- load from under sp 
   e89da800  ldm     sp, {fp, sp, pc}

Please check your compiler to make sure it's not the same problem.