Re: [PATCH 01/26] mm: asi: Make some utility functions noinstr compatible

[Junaid Shahid <junaids@google.com>]

On 11/1/24 1:27 PM, Thomas Gleixner wrote:
> On Thu, Oct 31 2024 at 18:44, Junaid Shahid wrote:
>> On 10/29/24 12:12 PM, Thomas Gleixner wrote:
>>>
>>> I doubt that it works as you want it to work.
>>>
>>> +	inline notrace __attribute((__section__(".noinstr.text")))	\
>>>
>>> So this explicitely puts the inline into the .noinstr.text section,
>>> which means when it is used in .text the compiler will generate an out-of
>>> line function in the .noinstr.text section and insert a call into the
>>> usage site. That's independent of the size of the inline.
>>>
>>
>> Oh, that's interesting. IIRC I had seen regular (.text) inline functions get
>> inlined into .noinstr.text callers. I assume the difference is that here the
>> section is marked explicitly rather than being implicit?
> 
> Correct. Inlines without any section attribute are free to be inlined in
> any section, but if the compiler decides to uninline them, then it
> sticks the uninlined version into the default section ".text".
> 
> The other problem there is that an out of line version can be
> instrumented if not explicitely forbidden.
> 
> That's why we mark them __always_inline, which forces the compiler to
> inline it into the usage site unconditionally.
> 
>> In any case, I guess we could just mark these functions as plain
>> noinstr.
> 
> No. Some of them are used in hotpath '.text'. 'noinstr' prevents them to
> be actually inlined then as I explained to you before.
> 
>> (Unless there happens to be some other way to indicate to the compiler to place
>> any non-inlined copy of the function in .noinstr.text but still allow inlining
>> into .text if it makes sense optimization-wise.)
> 
> Ideally the compilers would provide
> 
>          __attribute__(force_caller_section)
> 
> which makes them place an out of line inline into the section of the
> function from which it is called. But we can't have useful things or
> they are so badly documented that I can't find them ...
> 
> What actually works by some definition of "works" is:
> 
>         static __always_inline void __foo(void) { }
> 
>         static inline void foo(void)
>         {
>                  __(foo);
>         }
> 
>         static inline noinstr void foo_noinstr(void)
>         {
>                  __(foo);
>         }
> 
> The problem is that both GCC and clang optimize foo[_noinstr]() away and
> then follow the __always_inline directive of __foo() even if I make
> __foo() insanely large and have a gazillion of different functions
> marked noinline invoking foo() or foo_noinstr(), unless I add -fno-inline
> to the command line.
> 
> Which means it's not much different from just having '__always_inline
> foo()' without the wrappers....
> 
> Compilers clearly lack a --do-what-I-mean command line option.
> 
> Now if I'm truly nasty then both compilers do what I mean even without a
> magic command line option:
> 
>         static __always_inline void __foo(void) { }
> 
>         static __maybe_unused void foo(void)
>         {
>                  __(foo);
>         }
> 
>         static __maybe_unused noinstr void foo_noinstr(void)
>         {
>                  __(foo);
>         }
> 
> If there is a single invocation of either foo() or foo_noinstr() and
> they are small enough then the compiler inlines them, unless -fno-inline
> is on the command line. If there are multiple invocations and/or foo
> gets big enough then both compilers out of line them. The out of line
> wrappers with __foo() inlined in them end always up in the correct
> section.
> 
> I actually really like the programming model as it is very clear about
> the intention of usage and it allows static checkers to validate.
> 
> Thoughts?
> 

Thank you for the details. Yes, I think the last scheme that you described with 
separate wrappers for regular and noinst usage makes sense. IIRC the existing 
static validation wouldn't catch it if someone mistakenly called the .text 
version of the function from noinstr code and it just happened to get inlined. 
Perhaps we should add the -fno-inline compiler option with 
CONFIG_NOINSTR_VALIDATION?

Thanks,
Junaid