ARM: 9251/1: perf: Fix stacktraces for tracepoint events in THUMB2 kernels
Store the frame address where arm_get_current_stackframe() looks for it
(ARM_r7 instead of ARM_fp if CONFIG_THUMB2_KERNEL=y). Otherwise frame->fp
gets set to 0, causing unwind_frame() to fail.
# bpftrace -e 't:sched:sched_switch { @[kstack] = count(); exit(); }'
Attaching 1 probe...
@[
__schedule+1059
]: 1
A typical first unwind instruction is 0x97 (SP = R7), so after executing
it SP ends up being 0 and -URC_FAILURE is returned.
unwind_frame(pc = ac9da7d7 lr = 00000000 sp = c69bdda0 fp = 00000000)
unwind_find_idx(ac9da7d7)
unwind_exec_insn: insn = 00000097
unwind_exec_insn: fp = 00000000 sp = 00000000 lr = 00000000 pc = 00000000
With this patch:
# bpftrace -e 't:sched:sched_switch { @[kstack] = count(); exit(); }'
Attaching 1 probe...
@[
__schedule+1059
__schedule+1059
schedule+79
schedule_hrtimeout_range_clock+163
schedule_hrtimeout_range+17
ep_poll+471
SyS_epoll_wait+111
sys_epoll_pwait+231
__ret_fast_syscall+1
]: 1
Link: https://lore.kernel.org/r/20220920230728.2617421-1-tnovak@fb.com/
Reviewed-by: Linus Walleij <linus.walleij@linaro.org>
Signed-off-by: Tomislav Novak <tnovak@fb.com>
Signed-off-by: Russell King (Oracle) <rmk+kernel@armlinux.org.uk>
diff --git a/arch/arm/include/asm/perf_event.h b/arch/arm/include/asm/perf_event.h
index fe87397..bdbc1e5 100644
--- a/arch/arm/include/asm/perf_event.h
+++ b/arch/arm/include/asm/perf_event.h
@@ -17,7 +17,7 @@
#define perf_arch_fetch_caller_regs(regs, __ip) { \
(regs)->ARM_pc = (__ip); \
- (regs)->ARM_fp = (unsigned long) __builtin_frame_address(0); \
+ frame_pointer((regs)) = (unsigned long) __builtin_frame_address(0); \
(regs)->ARM_sp = current_stack_pointer; \
(regs)->ARM_cpsr = SVC_MODE; \
}
