[PATCH v2] ARM: Fix preemption disable in iwmmxt_task_enable()

[Sebastian Hesselbarth]

commit 431a84b1a4f7d1a0085d5b91330c5053cc8e8b12
 ("ARM: 8034/1: Disable preemption in iwmmxt_task_enable()")
introduced macros {inc,dec}_preempt_count to iwmmxt_task_enable
to make it run with preemption disabled.

Unfortunately, other functions in iwmmxt.S also use concan_{save,dump,load}
sections located in iwmmxt_task_enable() to deal with iWMMXt coprocessor.
This causes an unbalanced preempt_count due to excessive dec_preempt_count
and destroyed return addresses in callers of concan_ labels due to a register
collision:

Linux version 3.16.0-rc3-00062-gd92a333-dirty (jef at armhf) (gcc version 4.8.3 (Debian 4.8.3-4) ) #5 PREEMPT Thu Jul 3 19:46:39 CEST 2014
CPU: ARMv7 Processor [560f5815] revision 5 (ARMv7), cr=10c5387d
CPU: PIPT / VIPT nonaliasing data cache, PIPT instruction cache
Machine model: SolidRun CuBox
...
PJ4 iWMMXt v2 coprocessor enabled.
...
Unable to handle kernel paging request at virtual address fffffffe
pgd = bb25c000
[fffffffe] *pgd=3bfde821, *pte=00000000, *ppte=00000000
Internal error: Oops: 80000007 [#1] PREEMPT ARM
Modules linked in:
CPU: 0 PID: 62 Comm: startpar Not tainted 3.16.0-rc3-00062-gd92a333-dirty #5
task: bb230b80 ti: bb256000 task.ti: bb256000
PC is at 0xfffffffe
LR is at iwmmxt_task_copy+0x44/0x4c
pc : [<fffffffe>]    lr : [<800130ac>]    psr: 40000033
sp : bb257de8  ip : 00000013  fp : bb257ea4
r10: bb256000  r9 : fffffdfe  r8 : 76e898e6
r7 : bb257ec8  r6 : bb256000  r5 : 7ea12760  r4 : 000000a0
r3 : ffffffff  r2 : 00000003  r1 : bb257df8  r0 : 00000000
Flags: nZcv  IRQs on  FIQs on  Mode SVC_32  ISA Thumb  Segment user
Control: 10c5387d  Table: 3b25c019  DAC: 00000015
Process startpar (pid: 62, stack limit = 0xbb256248)

This patch fixes the issue by moving concan_{save,dump,load} into separate
code sections and make iwmmxt_task_enable() call them in the same way the
other functions use concan_ symbols. The test for valid ownership is moved
to concan_save and is safe for the other user of it, iwmmxt_task_disable().
The register collision is also resolved by moving concan_ symbols as
{inc,dec}_preempt_count are now local to iwmmxt_task_enable().

Signed-off-by: Sebastian Hesselbarth <sebastian.hesselbarth at gmail.com>
Reported-by: Jean-Francois Moine <moinejf at free.fr>
Fixes: 431a84b1a4f7 ("ARM: 8034/1: Disable preemption in iwmmxt_task_enable()")
---
The offending commit was intoduced past v3.15-rc1 and the corresponding fix
should also be queued up for stable v3.15+

Changelog
v1->v2:
- return immediately from concan_ instead of branch to 3f, i.e. replace
  'beq 3f' with 'moveq pc, lr' (Suggested by Catalin Marinas)

Cc: Russell King <linux at arm.linux.org.uk>
Cc: Catalin Marinas <catalin.marinas at arm.com>
Cc: Jean-Francois Moine <moinejf at free.fr>
Cc: Jason Cooper <jason at lakedaemon.net>
Cc: linux-arm-kernel at lists.infradead.org
Cc: linux-kernel at vger.kernel.org
---
 arch/arm/kernel/iwmmxt.S | 23 ++++++++++++-----------
 1 file changed, 12 insertions(+), 11 deletions(-)

diff --git a/arch/arm/kernel/iwmmxt.S b/arch/arm/kernel/iwmmxt.S
index a5599cfc43cb..2b32978ae905 100644
--- a/arch/arm/kernel/iwmmxt.S
+++ b/arch/arm/kernel/iwmmxt.S
@@ -94,13 +94,19 @@ ENTRY(iwmmxt_task_enable)
 
 	mrc	p15, 0, r2, c2, c0, 0
 	mov	r2, r2				@ cpwait
+	bl	concan_save
 
-	teq	r1, #0				@ test for last ownership
-	mov	lr, r9				@ normal exit from exception
-	beq	concan_load			@ no owner, skip save
+#ifdef CONFIG_PREEMPT_COUNT
+	get_thread_info r10
+#endif
+4:	dec_preempt_count r10, r3
+	mov	pc, r9				@ normal exit from exception
 
 concan_save:
 
+	teq	r1, #0				@ test for last ownership
+	beq	concan_load			@ no owner, skip save
+
 	tmrc	r2, wCon
 
 	@ CUP? wCx
@@ -138,7 +144,7 @@ concan_dump:
 	wstrd	wR15, [r1, #MMX_WR15]
 
 2:	teq	r0, #0				@ anything to load?
-	beq	3f
+	moveq	pc, lr				@ if not, return
 
 concan_load:
 
@@ -171,14 +177,9 @@ concan_load:
 	@ clear CUP/MUP (only if r1 != 0)
 	teq	r1, #0
 	mov 	r2, #0
-	beq	3f
-	tmcr	wCon, r2
+	moveq	pc, lr
 
-3:
-#ifdef CONFIG_PREEMPT_COUNT
-	get_thread_info r10
-#endif
-4:	dec_preempt_count r10, r3
+	tmcr	wCon, r2
 	mov	pc, lr
 
 /*
-- 
2.0.0