* Re: [PATCH 01/26] mm: asi: Make some utility functions noinstr compatible
2024-11-01 20:27 ` [PATCH 01/26] mm: asi: Make some utility functions noinstr compatible Thomas Gleixner
@ 2024-11-05 21:40 ` Junaid Shahid
2024-12-13 14:45 ` Brendan Jackman
1 sibling, 0 replies; 3+ messages in thread
From: Junaid Shahid @ 2024-11-05 21:40 UTC (permalink / raw)
To: Thomas Gleixner, Brendan Jackman, Borislav Petkov
Cc: Ingo Molnar, Dave Hansen, H. Peter Anvin, Andy Lutomirski,
Peter Zijlstra, Sean Christopherson, Paolo Bonzini,
Alexandre Chartre, Jan Setje-Eilers, Catalin Marinas, Will Deacon,
Mark Rutland, Andrew Morton, Mel Gorman, Lorenzo Stoakes,
David Hildenbrand, Vlastimil Babka, Michal Hocko, Juri Lelli,
Vincent Guittot, Dietmar Eggemann, Steven Rostedt,
Valentin Schneider, Paul Turner, Reiji Watanabe, Ofir Weisse,
Yosry Ahmed, Patrick Bellasi, KP Singh, Alexandra Sandulescu,
Matteo Rizzo, Jann Horn, x86, linux-kernel, linux-mm, kvm,
linux-toolchains
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
^ permalink raw reply [flat|nested] 3+ messages in thread* Re: [PATCH 01/26] mm: asi: Make some utility functions noinstr compatible
2024-11-01 20:27 ` [PATCH 01/26] mm: asi: Make some utility functions noinstr compatible Thomas Gleixner
2024-11-05 21:40 ` Junaid Shahid
@ 2024-12-13 14:45 ` Brendan Jackman
1 sibling, 0 replies; 3+ messages in thread
From: Brendan Jackman @ 2024-12-13 14:45 UTC (permalink / raw)
To: Thomas Gleixner
Cc: Junaid Shahid, Borislav Petkov, Ingo Molnar, Dave Hansen,
H. Peter Anvin, Andy Lutomirski, Peter Zijlstra,
Sean Christopherson, Paolo Bonzini, Alexandre Chartre, Liran Alon,
Jan Setje-Eilers, Catalin Marinas, Will Deacon, Mark Rutland,
Andrew Morton, Mel Gorman, Lorenzo Stoakes, David Hildenbrand,
Vlastimil Babka, Michal Hocko, Khalid Aziz, Juri Lelli,
Vincent Guittot, Dietmar Eggemann, Steven Rostedt,
Valentin Schneider, Paul Turner, Reiji Watanabe, Ofir Weisse,
Yosry Ahmed, Patrick Bellasi, KP Singh, Alexandra Sandulescu,
Matteo Rizzo, Jann Horn, x86, linux-kernel, linux-mm, kvm,
linux-toolchains
On Fri, 1 Nov 2024 at 21:27, Thomas Gleixner <tglx@linutronix.de> wrote:
> On Thu, Oct 31 2024 at 18:44, Junaid Shahid wrote:
> 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.
In this experiment did you modify the definition of noinstr to remove
noinline? Otherwise I always get out-of-line calls (as you'd expect
from the noinline).
> 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);
> }
I don't see any difference with __maybe_unused: if the noinline is
there it's never inlined, otherwise with the wrapper technique it's
always inlined regardless of __maybe_unused:
static __always_inline void __big(void)
{
asm volatile(
"nop; nop; nop; nop; nop; nop; nop; nop; nop;"
// and so on
"nop; nop; nop; nop; nop; nop; nop; nop; nop;"
);
}
static inline __section(".noinstr.text") void big_noinstr(void)
{
__big();
}
When I call big_noinstr() from a noinstr function I see:
Dump of assembler code for function asi_exit:
0xffffffff811e0080 <+0>: endbr64
0xffffffff811e0084 <+4>: nop
0xffffffff811e0085 <+5>: nop
...and so on
I'm using GCC 14.2.0.
(I thought maybe this was because I used asm volatile nops to embiggen
the function but I see the same thing with a big stream of volatile C
statements).
I think we might have no choice but to always use
__always_inline/noinline for code that's called from both sections -
seems there's no way to tell the compiler "I don't care if you inline
it, but it we can't cross a section boundary". Am I missing anything?
^ permalink raw reply [flat|nested] 3+ messages in thread