Re: [PATCH v6 13/18] arm64: unwind: strip PAC from kernel addresses

[James Morse <james.morse@arm.com>]

Hi Amit,

On 10/03/2020 12:28, Amit Kachhap wrote:
> On 3/10/20 12:33 AM, James Morse wrote:
>> On 06/03/2020 06:35, Amit Daniel Kachhap wrote:
>>> From: Mark Rutland <mark.rutland@arm.com>
>>>
>>> When we enable pointer authentication in the kernel, LR values saved to
>>> the stack will have a PAC which we must strip in order to retrieve the
>>> real return address.
>>>
>>> Strip PACs when unwinding the stack in order to account for this.
>>
>> This patch had me looking at the wider pointer-auth + ftrace interaction...

>>> diff --git a/arch/arm64/kernel/stacktrace.c b/arch/arm64/kernel/stacktrace.c
>>> index a336cb1..b479df7 100644
>>> --- a/arch/arm64/kernel/stacktrace.c
>>> +++ b/arch/arm64/kernel/stacktrace.c
>>> @@ -14,6 +14,7 @@
>>>   #include <linux/stacktrace.h>
>>>     #include <asm/irq.h>
>>> +#include <asm/pointer_auth.h>
>>>   #include <asm/stack_pointer.h>
>>>   #include <asm/stacktrace.h>
>>>   @@ -101,6 +102,8 @@ int notrace unwind_frame(struct task_struct *tsk, struct
>>> stackframe *frame)
>>
>> There is an earlier reader of frame->pc:
>> | #ifdef CONFIG_FUNCTION_GRAPH_TRACER
>> |     if (tsk->ret_stack &&
>> |             (frame->pc == (unsigned long)return_to_handler)) {
>>
>>
>> Which leads down the rat-hole of: does this need ptrauth_strip_insn_pac()?
>>
>> The version of GCC on my desktop supports patchable-function-entry, the function pre-amble
>> has two nops for use by ftrace[0]. This means if prepare_ftrace_return() re-writes the
>> saved LR, it does it before the caller paciasp's it.
>>
>> I think that means if you stack-trace from a function that had been hooked by the
>> function_graph_tracer, you will see the LR with a PAC, meaning the above == won't match.
>>
>>
>> The version of LLVM on my desktop however doesn't support patchable-function-entry, it
>> uses _mcount() to do the ftrace stuff[1]. Here prepare_ftrace_return() overwrites a
>> paciasp'd LR with one that isn't, which will fail.
>>
>>
>> Could the ptrauth_strip_insn_pac() call move above the CONFIG_FUNCTION_GRAPH_TRACER block,

> This may not be required as we never explicitly sign return_to_handler

Doesn't the original caller sign it? (I agree that assembly is tricky to work out)

ftrace_graph_caller passes 'parent' to prepare_ftrace_return() as the LR in regs:
| add	x1, sp, #S_LR

prepare_ftrace_return() may overwrite it with an unsigned value.

ftrace_common_return restores this location to x30:
| ldr	x30, [sp, #S_LR]

Then returns to the first real instruction of the original caller: paciasp.

(when navigating that assembly, there are two stack frames, each with an LR, and one LR in
the regs...)


> and frame->pc may
> store it without any PAC signature for patchable-function-entry.

How does return_to_handler() run? Surely when the original caller pulls the LR off the
stack, it runs:
| autiasp
| ret

Wouldn't autiasp transform an unsigned return_to_handler() to be a bogus address?

I agree the 'unsigned' case does happen if you're using _mcount(), this will be caught by
autiasp, hence we need to depend on HAVE_DYNAMIC_FTRACE_WITH_REGS.


> While testing patchable-function-entry, I had an observation regarding WARN messages,
> 
> [  541.030265] Hardware name: Foundation-v8A (DT)
> [  541.033500] pstate: 60400009 (nZCv daif +PAN -UAO)
> [  541.036880] pc : change_pac_parameters+0x40/0x4c
> [  541.040279] lr : return_to_handler+0x0/0x3c
> [  541.043373] sp : ffff8000126e3d00

(a WARN()ing?, where?! Ah, you mean triggered deliberately to check they look right...)


> Here lr may need some logic to display correct return address although it is unrelated to
> this ptrauth series. (arch/arm64/kernel/process.c +264)

Yes, this happens when a function that has been hooked by ftrace, hits a WARN_ON(),
show_regs() will report the real LR. I don't think that's a problem, its helpful to know
that ftrace has hooked this call.

Presumably return_to_handler() doesn't appear in the call-trace? (that would be a problem)


>> and could we add something like:
>> |    depends on (!FTRACE || HAVE_DYNAMIC_FTRACE_WITH_REGS)
>>
>> to the Kconfig to prevent both FTRACE and PTR_AUTH being enabled unless the compiler has
>> support for patchable-function-entry?
> 
> Yes this is a good condition to have. I feel below condition is more suitable as there is
> issue only with FUNCTION_GRAPH_TRACER,

Er, yes!
Because its callers of prepare_ftrace_return() that have the problem, and that is behind
#ifdef FUNCTION_GRAPH_TRACER.


Thanks,

James

_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel