[PATCH] ARM: unwind: fix unwind started from IRQ stack in THUMB2 kernel

Haibo Li haibo.li at mediatek.com
Tue Feb 27 23:32:47 PST 2024


When unwinds started from IRQ stack,call_with_stack is the only
place to do stack switch.
It need to identify call_with_stack when unwinding.

In thumb2 kernel,it fails to identify call_with_stack.
In practice,(u32)&call_with_stack has bit 0 set.

	prel31_to_addr(&idx->addr_offset) is 0x806ed518,
	while (u32)&call_with_stack is 0x806ed519.
So it is impossible to do stack switch.

dump_stack from IRQ stack gets below result:
...
 call_timer_fn from __run_timers+0x163/0x25c
 __run_timers from run_timer_softirq+0x15/0x24
 run_timer_softirq from __do_softirq+0xe3/0x232
 __do_softirq from __irq_exit_rcu+0x3f/0xac
 __irq_exit_rcu from irq_exit+0x7/0xe
 irq_exit from call_with_stack+0xd/0x10
...
The stacktrace ends with call_with_stack.

Since bit 0 of pc in thumb2 is always 0,skip bit 0 when do compraing.
Then we get expected stacktrace:
...
 call_timer_fn from __run_timers+0x163/0x25c
 __run_timers from run_timer_softirq+0x15/0x24
 run_timer_softirq from __do_softirq+0xe3/0x232
 __do_softirq from __irq_exit_rcu+0x3f/0xac
 __irq_exit_rcu from irq_exit+0x7/0xe
 irq_exit from call_with_stack+0xd/0x10
 call_with_stack from __irq_svc+0x93/0xb6
Exception stack(0xf0875f60 to 0xf0875fa8)
5f60: ****
5f80: ****
5fa0: ****
 __irq_svc from arch_local_irq_enable+0x2/0x4
 arch_local_irq_enable from do_idle+0xad/0x1f6
...

Signed-off-by: Haibo Li <haibo.li at mediatek.com>
---
 arch/arm/kernel/unwind.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/arch/arm/kernel/unwind.c b/arch/arm/kernel/unwind.c
index 9d2192156087..89e1f440082c 100644
--- a/arch/arm/kernel/unwind.c
+++ b/arch/arm/kernel/unwind.c
@@ -482,7 +482,8 @@ int unwind_frame(struct stackframe *frame)
 
 	ctrl.check_each_pop = 0;
 
-	if (prel31_to_addr(&idx->addr_offset) == (u32)&call_with_stack) {
+	if (prel31_to_addr(&idx->addr_offset) ==
+		((u32)&call_with_stack & (~0x01))) {
 		/*
 		 * call_with_stack() is the only place where we permit SP to
 		 * jump from one stack to another, and since we know it is
-- 
2.18.0




More information about the Linux-mediatek mailing list