Re: [PATCH 1/1] x86/oprofile: Fix the calltrace upon profiling some specified events with oprofile

[Robert Richter <robert.richter@amd.com>]

On 27.08.12 09:32:13, Wei.Yang@windriver.com wrote:
> From: Wei Yang <Wei.Yang@windriver.com>
> 
> Upon enabling the call-graph functionality of oprofile, A few minutes
> later the following calltrace will always occur.
> 
> BUG: unable to handle kernel paging request at 656d6153

This is probably the same I found to yesterday. Will test your fix.

-Robert

> IP: [<c10050f5>] print_context_stack+0x55/0x110
> *pde = 00000000
> Oops: 0000 [#1] PREEMPT SMP
> Modules linked in:
> Pid: 0, comm: swapper/0 Not tainted 3.6.0-rc3-WR5.0+snapshot-20120820_standard
> EIP: 0060:[<c10050f5>] EFLAGS: 00010093 CPU: 0
> EIP is at print_context_stack+0x55/0x110
> EAX: 656d7ffc EBX: 656d6153 ECX: c1837ee0 EDX: 656d6153
> ESI: 00000000 EDI: ffffe000 EBP: f600deac ESP: f600de88
> DS: 007b ES: 007b FS: 00d8 GS: 0000 SS: 0068
> CR0: 8005003b CR2: 656d6153 CR3: 01934000 CR4: 000007d0
> DR0: 00000000 DR1: 00000000 DR2: 00000000 DR3: 00000000
> DR6: ffff0ff0 DR7: 00000400
> Process swapper/0 (pid: 0, ti=f600c000 task=c18411a0 task.ti=c1836000)
> Stack:
> 1a7f76ea 656d7ffc 656d6000 c1837ee0 ffffe000 c1837ee0 656d6153 c188e27c
> 656d6000 f600dedc c10040f8 c188e27c f600def0 00000000 f600dec8 c1837ee0
> 00000000 f600ded4 c1837ee0 f600dfc4 0000001f f600df08 c1564d22 00000000
> Call Trace:
> [<c10040f8>] dump_trace+0x68/0xf0
> [<c1564d22>] x86_backtrace+0xb2/0xc0
> [<c1562dd2>] oprofile_add_sample+0xa2/0xc0
> [<c1003fbf>] ? do_softirq+0x6f/0xa0
> [<c1566519>] ppro_check_ctrs+0x79/0x100
> [<c15664a0>] ? ppro_shutdown+0x60/0x60
> [<c156552f>] profile_exceptions_notify+0x8f/0xb0
> [<c1672248>] nmi_handle.isra.0+0x48/0x70
> [<c1672343>] do_nmi+0xd3/0x3c0
> [<c1033d39>] ? __local_bh_enable+0x29/0x70
> [<c1034620>] ? ftrace_define_fields_irq_handler_entry+0x80/0x80
> [<c1671a0d>] nmi_stack_correct+0x28/0x2d
> [<c1034620>] ? ftrace_define_fields_irq_handler_entry+0x80/0x80
> [<c1003fbf>] ? do_softirq+0x6f/0xa0
> <IRQ>
> [<c1034ad5>] irq_exit+0x65/0x70
> [<c16776f9>] smp_apic_timer_interrupt+0x59/0x89
> [<c16717da>] apic_timer_interrupt+0x2a/0x30
> [<c135164d>] ? acpi_idle_enter_bm+0x245/0x273
> [<c14f3a25>] cpuidle_enter+0x15/0x20
> [<c14f4070>] cpuidle_idle_call+0xa0/0x320
> [<c1009705>] cpu_idle+0x55/0xb0
> [<c16519a8>] rest_init+0x6c/0x74
> [<c18a291b>] start_kernel+0x2ec/0x2f3
> [<c18a2467>] ? repair_env_string+0x51/0x51
> [<c18a22a2>] i386_start_kernel+0x78/0x7d
> Code: e0 ff ff 89 7d ec 74 5a 8d b4 26 00 00 00 00 8d bc 27 00 00
> 00 00 39 f3 72 0c 8b 45 f0 8d 64 24 18 5b 5e 5f 5d c3 3b 5d ec 72
> ef <8b> 3b 89 f8 89 7d dc e8 ef 40 04 00 85 c0 74 20 8b 40
> EIP: [<c10050f5>] print_context_stack+0x55/0x110 SS:ESP 0068:f600de88
> CR2: 00000000656d6153
> ---[ end trace 751c9b47c6ff5e29 ]---
> 
> Let's assume a scenario that do_softirq() switches the stack to a soft irq
> stack, and the soft irq stack is totally empty. At this moment, a nmi interrupt
> occurs, subsequently, CPU does not automatically save SS and SP registers
> and switch any stack, but instead only stores EFLAGS, CS and IP to the soft irq
> stack. since the CPU is in kernel mode when the NMI exception occurs.
> the layout of the current soft irq stack is this:
> 
>   +--------------+<-----the top of soft irq
>   |   EFLAGS     |
>   +--------------+
>   |    CS        |
>   +--------------+
>   |    IP        |
>   +--------------+
>   |   SAVE_ALL   |
>   +--------------+
> 
> but the return value of the function kernel_stack_pointer() is'&regs->sp'
> (for x86_32 CPU), which is invoked by the x86_backtrace function. Since
> the type of regs pointer is a pt_regs structure, the return value is not
> in the range of the soft irq stack, as the SP register is not save in the
> soft irq stack. Therefore, we need to check if the return value of the function
> resides in valid range. Additionally, the changes has no impact on the normal
> NMI exception.
> 
> Signed-off-by: Yang Wei <wei.yang@windriver.com>
> ---
>  arch/x86/oprofile/backtrace.c |   10 ++++++++++
>  1 files changed, 10 insertions(+), 0 deletions(-)
> 
> diff --git a/arch/x86/oprofile/backtrace.c b/arch/x86/oprofile/backtrace.c
> index d6aa6e8..a5fca0b 100644
> --- a/arch/x86/oprofile/backtrace.c
> +++ b/arch/x86/oprofile/backtrace.c
> @@ -17,6 +17,11 @@
>  #include <asm/ptrace.h>
>  #include <asm/stacktrace.h>
>  
> +static inline int valid_stack_ptr(struct thread_info *tinfo, void *p)
> +{
> +	void *t = tinfo;
> +	return p > t + sizeof(struct thread_info) && p < t + THREAD_SIZE;
> +}
>  static int backtrace_stack(void *data, char *name)
>  {
>  	/* Yes, we want all stacks */
> @@ -110,9 +115,14 @@ void
>  x86_backtrace(struct pt_regs * const regs, unsigned int depth)
>  {
>  	struct stack_frame *head = (struct stack_frame *)frame_pointer(regs);
> +	struct thread_info *context;
>  
>  	if (!user_mode_vm(regs)) {
>  		unsigned long stack = kernel_stack_pointer(regs);
> +		context = (struct thread_info *)
> +			(stack & (~(THREAD_SIZE - 1)));
> +		if (!valid_stack_ptr(context, (void *)stack))
> +			return;
>  		if (depth)
>  			dump_trace(NULL, regs, (unsigned long *)stack, 0,
>  				   &backtrace_ops, &depth);
> -- 
> 1.7.0.2
> 
> 

-- 
Advanced Micro Devices, Inc.
Operating System Research Center