[RFC PATCH v2 1/8] arm64: Implement stack trace termination record

Madhavan T. Venkataraman madvenka at linux.microsoft.com
Tue Mar 23 12:39:34 GMT 2021



On 3/23/21 5:24 AM, Mark Rutland wrote:
> On Fri, Mar 19, 2021 at 05:03:09PM -0500, Madhavan T. Venkataraman wrote:
>> I solved this by using existing functions logically instead of inventing a
>> dummy function. I initialize pt_regs->stackframe[1] to an existing function
>> so that the stack trace will not show a 0x0 entry as well as the kernel and
>> gdb will show identical stack traces.
>>
>> For all task stack traces including the idle tasks, the stack trace will
>> end at copy_thread() as copy_thread() is the function that initializes the
>> pt_regs and the first stack frame for a task.
> 
> I don't think this is a good idea, as it will mean that copy_thread()
> will appear to be live in every thread, and therefore will not be
> patchable.
> 
> There are other things people need to be aware of when using an external
> debugger (e.g. during EL0<->ELx transitions there are periods when X29
> and X30 contain the EL0 values, and cannot be used to unwind), so I
> don't think there's a strong need to make this look prettier to an
> external debugger.
> 

OK.

>> For EL0 exceptions, the stack trace will end with vectors() as vectors
>> entries call the EL0 handlers.
>>
>> Here are sample stack traces (I only show the ending of each trace):
>>
>> Idle task on primary CPU
>> ========================
>>
>> 		 ...
>> [    0.022557]   start_kernel+0x5b8/0x5f4
>> [    0.022570]   __primary_switched+0xa8/0xb8
>> [    0.022578]   copy_thread+0x0/0x188
>>
>> Idle task on secondary CPU
>> ==========================
>>
>> 		 ...
>> [    0.023397]   secondary_start_kernel+0x188/0x1e0
>> [    0.023406]   __secondary_switched+0x40/0x88
>> [    0.023415]   copy_thread+0x0/0x188
>>
>> All other kernel threads
>> ========================
>>
>> 		 ...
>> [   13.501062]   ret_from_fork+0x10/0x18
>> [   13.507998]   copy_thread+0x0/0x188
>>
>> User threads (EL0 exception)
>> ============
>>
>> write(2) system call example:
>>
>> 		 ...
>> [  521.686148]   vfs_write+0xc8/0x2c0
>> [  521.686156]   ksys_write+0x74/0x108
>> [  521.686161]   __arm64_sys_write+0x24/0x30
>> [  521.686166]   el0_svc_common.constprop.0+0x70/0x1a8
>> [  521.686175]   do_el0_svc+0x2c/0x98
>> [  521.686180]   el0_svc+0x2c/0x70
>> [  521.686188]   el0_sync_handler+0xb0/0xb8
>> [  521.686193]   el0_sync+0x17c/0x180
>> [  521.686198]   vectors+0x0/0x7d8
> 
> [...]
> 
>> If you approve, the above will become RFC Patch v3 1/8 in the next version.
> 
> As above, I don't think we should repurpose an existing function here,
> and my preference is to use 0x0.
> 

OK.

>> Let me know.
>>
>> Also, I could introduce an extra frame in the EL1 exception stack trace that
>> includes vectors so the stack trace would look like this (timer interrupt example):
>>
>> call_timer_fn
>> run_timer_softirq
>> __do_softirq
>> irq_exit
>> __handle_domain_irq
>> gic_handle_irq
>> el1_irq
>> vectors
>>
>> This way, if the unwinder finds vectors, it knows that it is an exception frame.
> 
> I can see this might make it simpler to detect exception boundaries, but
> I suspect that we need other information anyway, so this doesn't become
> all that helpful. For EL0<->EL1 exception boundaries we want to
> successfully terminate a robust stacktrace whereas for EL1<->EL1
> exception boundaries we want to fail a robust stacktrace.
> 
> I reckon we have to figure that out from the el1_* and el0_* entry
> points (which I am working to reduce/simplify as part of the entry
> assembly conversion to C). With that we can terminate unwind at the
> el0_* parts, and reject unwinding across any other bit of .entry.text.
> 

OK. That is fine.

Thanks.

Madhavan



More information about the linux-arm-kernel mailing list