Re: [PATCH v2 09/13] powerpc/ftrace: Implement CONFIG_DYNAMIC_FTRACE_WITH_ARGS

["Naveen N. Rao" <naveen.n.rao@linux.vnet.ibm.com>]

Michael Ellerman wrote:
> Christophe Leroy <christophe.leroy@csgroup.eu> writes:
>> Le 14/02/2022 à 16:25, Naveen N. Rao a écrit :
>>> Christophe Leroy wrote:
>>>> Implement CONFIG_DYNAMIC_FTRACE_WITH_ARGS. It accelerates the call
>>>> of livepatching.
>>>>
>>>> Also note that powerpc being the last one to convert to
>>>> CONFIG_DYNAMIC_FTRACE_WITH_ARGS, it will now be possible to remove
>>>> klp_arch_set_pc() on all architectures.
>>>>
>>>> Signed-off-by: Christophe Leroy <christophe.leroy@csgroup.eu>
>>>> ---
>>>>  arch/powerpc/Kconfig                 |  1 +
>>>>  arch/powerpc/include/asm/ftrace.h    | 17 +++++++++++++++++
>>>>  arch/powerpc/include/asm/livepatch.h |  4 +---
>>>>  3 files changed, 19 insertions(+), 3 deletions(-)
>>>>
>>>> diff --git a/arch/powerpc/Kconfig b/arch/powerpc/Kconfig
>>>> index cdac2115eb00..e2b1792b2aae 100644
>>>> --- a/arch/powerpc/Kconfig
>>>> +++ b/arch/powerpc/Kconfig
>>>> @@ -210,6 +210,7 @@ config PPC
>>>>      select HAVE_DEBUG_KMEMLEAK
>>>>      select HAVE_DEBUG_STACKOVERFLOW
>>>>      select HAVE_DYNAMIC_FTRACE
>>>> +    select HAVE_DYNAMIC_FTRACE_WITH_ARGS    if MPROFILE_KERNEL || PPC32
>>>>      select HAVE_DYNAMIC_FTRACE_WITH_REGS    if MPROFILE_KERNEL || PPC32
>>>>      select HAVE_EBPF_JIT
>>>>      select HAVE_EFFICIENT_UNALIGNED_ACCESS    if !(CPU_LITTLE_ENDIAN 
>>>> && POWER7_CPU)
>>>> diff --git a/arch/powerpc/include/asm/ftrace.h 
>>>> b/arch/powerpc/include/asm/ftrace.h
>>>> index b3f6184f77ea..45c3d6f11daa 100644
>>>> --- a/arch/powerpc/include/asm/ftrace.h
>>>> +++ b/arch/powerpc/include/asm/ftrace.h
>>>> @@ -22,6 +22,23 @@ static inline unsigned long 
>>>> ftrace_call_adjust(unsigned long addr)
>>>>  struct dyn_arch_ftrace {
>>>>      struct module *mod;
>>>>  };
>>>> +
>>>> +#ifdef CONFIG_DYNAMIC_FTRACE_WITH_ARGS
>>>> +struct ftrace_regs {
>>>> +    struct pt_regs regs;
>>>> +};
>>>> +
>>>> +static __always_inline struct pt_regs *arch_ftrace_get_regs(struct 
>>>> ftrace_regs *fregs)
>>>> +{
>>>> +    return &fregs->regs;
>>>> +}
>>> 
>>> I think this is wrong. We need to differentiate between ftrace_caller() 
>>> and ftrace_regs_caller() here, and only return pt_regs if coming in 
>>> through ftrace_regs_caller() (i.e., FL_SAVE_REGS is set).
>>
>> Not sure I follow you.
>>
>> This is based on 5740a7c71ab6 ("s390/ftrace: add 
>> HAVE_DYNAMIC_FTRACE_WITH_ARGS support")
>>
>> It's all the point of HAVE_DYNAMIC_FTRACE_WITH_ARGS, have the regs also 
>> with ftrace_caller().
>>
>> Sure you only have the params, but that's the same on s390, so what did 
>> I miss ?

It looks like s390 is special since it apparently saves all registers 
even for ftrace_caller: 
https://lore.kernel.org/all/YbipdU5X4HNDWIni@osiris/

As I understand it, the reason ftrace_get_regs() was introduced was to 
be able to only return the pt_regs, if _all_ registers were saved into 
it, which we don't do when coming in through ftrace_caller(). See the 
x86 implementation (commit 02a474ca266a47 ("ftrace/x86: Allow for 
arguments to be passed in to ftrace_regs by default"), which returns 
pt_regs conditionally.

> 
> I already have this series in next, I can pull it out, but I'd rather
> not.

Yeah, I'm sorry about the late review on this one.

> 
> I'll leave it in for now, hopefully you two can agree overnight my time
> whether this is a big problem or something we can fix with a fixup
> patch.

I think changes to this particular patch can be added as an incremental 
patch. If anything, pt_regs won't have all valid registers, but no one 
should depend on it without also setting FL_SAVE_REGS anyway.

I was concerned about patch 8 though, where we are missing saving r1 
into pt_regs. That gets used in patch 11, and will be used during 
unwinding when the function_graph tracer is active. But, this should 
still just result in us being unable to unwind the stack, so I think 
that can also be an incremental patch.


Thanks,
Naveen