ARM: fix oops on initial entry to userspace with Thumb2 kernels

Daniel Mack reports an oops at boot with the latest kernels:

  Internal error: Oops - undefined instruction: 0 [#1] SMP THUMB2
  Modules linked in:
  CPU: 0    Not tainted  (3.6.0-11057-g584df1d #145)
  PC is at cpsw_probe+0x45a/0x9ac
  LR is at trace_hardirqs_on_caller+0x8f/0xfc
  pc : [<c03493de>]    lr : [<c005e81f>]    psr: 60000113
  sp : cf055fb0  ip : 00000000  fp : 00000000
  r10: 00000000  r9 : 00000000  r8 : 00000000
  r7 : 00000000  r6 : 00000000  r5 : c0344555  r4 : 00000000
  r3 : cf057a40  r2 : 00000000  r1 : 00000001  r0 : 00000000
  Flags: nZCv  IRQs on  FIQs on  Mode SVC_32  ISA ARM Segment user
  Control: 50c5387d  Table: 8f3f4019  DAC: 00000015
  Process init (pid: 1, stack limit = 0xcf054240)
  Stack: (0xcf055fb0 to 0xcf056000)
  5fa0:                                     00000001 00000000 00000000 00000000
  5fc0: cf055fb0 c000d1a8 00000000 00000000 00000000 00000000 00000000 00000000
  5fe0: 00000000 be9b3f10 00000000 b6f6add0 00000010 00000000 aaaabfaf a8babbaa

The analysis of this is as follows.  In init/main.c, we issue:

	kernel_thread(kernel_init, NULL, CLONE_FS | CLONE_SIGHAND);

This creates a new thread, which falls through to the ret_from_fork
assembly, with r4 set NULL and r5 set to kernel_init.  You can see
this in your oops dump register set - r5 is 0xc0344555, which is the
address of kernel_init plus 1 which marks the function as Thumb code.

Now, let's look at this code a little closer - this is what the
disassembly looks like:

  c000d180 <ret_from_fork>:
  c000d180:       f03a fe08       bl      c0047d94 <schedule_tail>
  c000d184:       2d00            cmp     r5, #0
  c000d186:       bf1e            ittt    ne
  c000d188:       4620            movne   r0, r4
  c000d18a:       46fe            movne   lr, pc <-- XXXXXXX
  c000d18c:       46af            movne   pc, r5
  c000d18e:       46e9            mov     r9, sp
  c000d190:       ea4f 3959       mov.w   r9, r9, lsr #13
  c000d194:       ea4f 3949       mov.w   r9, r9, lsl #13
  c000d198:       e7c8            b.n     c000d12c <ret_to_user>
  c000d19a:       bf00            nop
  c000d19c:       f3af 8000       nop.w

This code was introduced in 9fff2fa0 (arm: switch to saner
kernel_execve() semantics).  I have marked one instruction, and it's
the significant one - I'll come back to that later.

Eventually, having had a successful call to kernel_execve(), kernel_init()
returns zero.

In returning, it uses the value in 'lr' which was set by the instruction
I marked above.  Unfortunately, this causes lr to contain 0xc000d18e -
an even address.  This switches the ISA to ARM on return but with a non
word aligned PC value.

So, what do we end up executing?  Well, not the instructions above - yes
the opcodes, but they don't mean the same thing in ARM mode.  In ARM mode,
it looks like this instead:

  c000d18c:       46e946af        strbtmi r4, [r9], pc, lsr #13
  c000d190:       3959ea4f        ldmdbcc r9, {r0, r1, r2, r3, r6, r9, fp, sp, lr, pc}^
  c000d194:       3949ea4f        stmdbcc r9, {r0, r1, r2, r3, r6, r9, fp, sp, lr, pc}^
  c000d198:       bf00e7c8        svclt   0x0000e7c8
  c000d19c:       8000f3af        andhi   pc, r0, pc, lsr #7
  c000d1a0:       e88db092        stm     sp, {r1, r4, r7, ip, sp, pc}
  c000d1a4:       46e81fff                        ; <UNDEFINED> instruction: 0x46e81fff
  c000d1a8:       8a00f3ef        bhi     0xc004a16c
  c000d1ac:       0a0cf08a        beq     0xc03493dc

I have included more above, because it's relevant.  The PSR flags which
we can see in the oops dump are nZCv, so Z and C are set.

All the above ARM instructions are not executed, except for two.
c000d1a0, which has no writeback, and writes below the current stack
pointer (and that data is lost when we take the next exception.) The
other instruction which is executed is c000d1ac, which takes us to...
0xc03493dc.  However, remember that bit 1 of the PC got set.  So that
makes the PC value 0xc03493de.

And that value is the value we find in the oops dump for PC.  What is
the instruction here when interpreted in ARM mode?

       0:       f71e150c                ; <UNDEFINED> instruction: 0xf71e150c

and there we have our undefined instruction (remember that the 'never'
condition code, 0xf, has been deprecated and is now always executed as
it is now being used for additional instructions.)

This path also nicely explains the state of the stack we see in the oops
dump too.

The above is a consistent and sane story for how we got to the oops
dump, which all stems from the instruction at 0xc000d18a being wrong.
Reported-by: Daniel Mack <zonque@gmail.com>
Tested-by: Daniel Mack <zonque@gmail.com>
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

arch/arm/kernel/entry-common.S

@@ -88,9 +88,9 @@ ENTRY(ret_from_fork)
 	bl	schedule_tail
 	cmp	r5, #0
 	movne	r0, r4
-	movne	lr, pc
+	adrne	lr, BSYM(1f)
 	movne	pc, r5
-	get_thread_info tsk
+1:	get_thread_info tsk
 	b	ret_slow_syscall
 ENDPROC(ret_from_fork)