Re: [PATCH] mm/alloc_tag: add the ARCH_NEEDS_WEAK_PER_CPU macro when statically defining the percpu variable alloc_tag_counters.

[Hao Ge]

On 2025/6/11 23:24, Suren Baghdasaryan wrote:
> On Tue, Jun 10, 2025 at 10:27 PM Hao Ge <hao.ge@xxxxxxxxx> wrote:
> > On 2025/6/10 00:39, Suren Baghdasaryan wrote:
> > > On Sun, Jun 8, 2025 at 11:08 PM Hao Ge <hao.ge@xxxxxxxxx> wrote:
> > > > On 2025/5/29 15:35, Hao Ge wrote:
> > > > > From: Hao Ge <gehao@xxxxxxxxxx>
> > > > >
> > > > > Recently discovered this entry while checking kallsyms on ARM64:
> > > > > ffff800083e509c0 D _shared_alloc_tag
> > > > >
> > > > > If ARCH_NEEDS_WEAK_PER_CPU is not defined,there's no need to statically
> > > > > define the percpu variable alloc_tag_counters.
> > > > >
> > > > > Therefore,add therelevant macro guards at the appropriate location.
> > > > >
> > > > > Fixes: 22d407b164ff ("lib: add allocation tagging support for memory allocation profiling")
> > > > > Signed-off-by: Hao Ge <gehao@xxxxxxxxxx>
> > > > > ---
> > > > >     lib/alloc_tag.c | 2 ++
> > > > >     1 file changed, 2 insertions(+)
> > > > >
> > > > > diff --git a/lib/alloc_tag.c b/lib/alloc_tag.c
> > > > > index c7f602fa7b23..d1dab80b70ad 100644
> > > > > --- a/lib/alloc_tag.c
> > > > > +++ b/lib/alloc_tag.c
> > > > > @@ -24,8 +24,10 @@ static bool mem_profiling_support;
> > > > >
> > > > >     static struct codetag_type *alloc_tag_cttype;
> > > > >
> > > > > +#ifdef ARCH_NEEDS_WEAK_PER_CPU
> > > > >     DEFINE_PER_CPU(struct alloc_tag_counters, _shared_alloc_tag);
> > > > >     EXPORT_SYMBOL(_shared_alloc_tag);
> > > > > +#endif /* ARCH_NEEDS_WEAK_PER_CPU */
> > > > >
> > > > >     DEFINE_STATIC_KEY_MAYBE(CONFIG_MEM_ALLOC_PROFILING_ENABLED_BY_DEFAULT,
> > > > >                         mem_alloc_profiling_key);
> > > > Hi Suren
> > > >
> > > >
> > > > I'm sorry to bother you. As mentioned in my commit message,
> > > >
> > > > in fact, on the ARM64 architecture, the _shared_alloc_tag percpu
> > > > variable is not needed.
> > > >
> > > > In my understanding, it will create a copy for each CPU.
> > > >
> > > >     The alloc_tag_counters variable will occupy 16 bytes,
> > > >
> > > > and as the number of CPUs increases, more and more memory will be wasted
> > > > in this segment.
> > > >
> > > > I realized that this modification was a mistake. It resulted in a build
> > > > error, and the link is as follows:
> > > >
> > > > https://lore.kernel.org/all/202506080448.KWN8arrX-lkp@xxxxxxxxx/
> > > >
> > > > After I studied the comments of DECLARE_PER_CPU_SECTION, I roughly
> > > > understood why this is the case.
> > > >
> > > > But so far, I haven't come up with a good way to solve this problem. Do
> > > > you have any suggestions?
> > > Hi Hao,
> > > The problem here is that ARCH_NEEDS_WEAK_PER_CPU is not a Kconfig
> > > option, it gets defined only on 2 architectures and only when building
> > > modules here https://elixir.bootlin.com/linux/v6.15.1/source/arch/alpha/include/asm/percpu.h#L14
> > > and here https://elixir.bootlin.com/linux/v6.15.1/source/arch/s390/include/asm/percpu.h#L21.
> > > A nicer way to deal with that is to make if a Kconfig option which is
> > > enabled only for alpha and s390 and then do something like this:
> > >
> > >    #if defined(MODULE) && defined(ARCH_NEEDS_WEAK_PER_CPU)
> > > #define MODULE_NEEDS_WEAK_PER_CPU
> > > #endif
> > >
> > > and change all the usages of ARCH_NEEDS_WEAK_PER_CPU with
> > > MODULE_NEEDS_WEAK_PER_CPU.
> > > Did I explain the idea clearly?
> > > Thanks,
> > > Suren.
> > Hi Suren
> Hi Hao,
>
> > Thanks for your guidance.
> > I understand this train of thought.
> >
> > I've been thinking about a problem: I only added the
> > ARCH_NEEDS_WEAK_PER_CPU
> >
> > macro to isolate the definition of _shared_alloc_tag.
> >
> > Since s390 defines this macro, why did this build error occur?
Hi Suren
> The problem is that ARCH_NEEDS_WEAK_PER_CPU is not a Kconfig option,
> it's just a definition, for s390 it's here:
> https://elixir.bootlin.com/linux/v6.15.1/source/arch/s390/include/asm/percpu.h#L21
> So, even for s390 if you are building core kernel code (not a module),
> ARCH_NEEDS_WEAK_PER_CPU will be undefined, however if you are building
> a module on s390 then it is defined. So, your change effectively
> results in _shared_alloc_tag being compiled out in the core kernel
> while it's used when you build a module. Therefore during linking
> modules can't link to that symbol in the core kernel. Hope this
> explains the issue.


Thank you so, so, so much! I understand now, and thank you for such a 
detailed explanation.


> The way I would fix this is by making ARCH_NEEDS_WEAK_PER_CPU a
> Kconfig option and enable it for s390 and alpha, would replace old
> definitions from
> https://elixir.bootlin.com/linux/v6.15.1/source/arch/s390/include/asm/percpu.h#L21
> and https://elixir.bootlin.com/linux/v6.15.1/source/arch/alpha/include/asm/percpu.h#L14
> with:
>
> #if defined(MODULE) && defined(ARCH_NEEDS_WEAK_PER_CPU)
> #define MODULE_NEEDS_WEAK_PER_CPU
> #endif
>
> Then use MODULE_NEEDS_WEAK_PER_CPU instead of ARCH_NEEDS_WEAK_PER_CPU
> in all the current places in the kernel code. Lastly, to compile out
> _shared_alloc_tag your current patch should work fine because on s390
> and alpha ARCH_NEEDS_WEAK_PER_CPU will be defined after all these
> changes.
> Does that make sense?


I quite agree with this approach.

Thanks
Best Regards
Hao

> > Could you please help explain it again?
> >
> > Thanks
> > Best Regards
> > Hao
> >
> > > > Thanks
> > > >
> > > > Best Regards
> > > >
> > > > Hao