[PATCH] RISC-V: only flush icache when it has VM_EXEC set

[Alexandre Ghiti]

On Wed, Mar 20, 2024 at 1:48 AM Palmer Dabbelt <palmer at dabbelt.com> wrote:
>
> On Tue, 09 Jan 2024 10:48:59 PST (-0800), cyy at cyyself.name wrote:
> > As I-Cache flush on current RISC-V needs to send IPIs to every CPU cores
> > in the system is very costly, limiting flush_icache_mm to be called only
> > when vma->vm_flags has VM_EXEC can help minimize the frequency of these
> > operations. It improves performance and reduces disturbances when
> > copy_from_user_page is needed such as profiling with perf.
> >
> > For I-D coherence concerns, it will not fail if such a page adds VM_EXEC
> > flags in the future since we have checked it in the __set_pte_at function.
> >
> > Signed-off-by: Yangyu Chen <cyy at cyyself.name>
> > ---
> >  arch/riscv/include/asm/cacheflush.h | 7 +++++--
> >  1 file changed, 5 insertions(+), 2 deletions(-)
> >
> > diff --git a/arch/riscv/include/asm/cacheflush.h b/arch/riscv/include/asm/cacheflush.h
> > index 3cb53c4df27c..915f532dc336 100644
> > --- a/arch/riscv/include/asm/cacheflush.h
> > +++ b/arch/riscv/include/asm/cacheflush.h
> > @@ -33,8 +33,11 @@ static inline void flush_dcache_page(struct page *page)
> >   * so instead we just flush the whole thing.
> >   */
> >  #define flush_icache_range(start, end) flush_icache_all()
> > -#define flush_icache_user_page(vma, pg, addr, len) \
> > -     flush_icache_mm(vma->vm_mm, 0)
> > +#define flush_icache_user_page(vma, pg, addr, len)   \
> > +do {                                                 \
> > +     if (vma->vm_flags & VM_EXEC)                    \
> > +             flush_icache_mm(vma->vm_mm, 0);         \
> > +} while (0)
> >
> >  #ifdef CONFIG_64BIT
> >  #define flush_cache_vmap(start, end) flush_tlb_kernel_range(start, end)
>
> I'm not super worried about the benchmarks, I think we can just
> open-loop assume this is faster by avoiding the flushes.  I do think we
> need a hook into at least tlb_update_vma_flags(), though, to insert the
> fence.i when upgrading a mapping to include VM_EXEC.

I'd say Yangyu is right when he mentions in the commit log: "For I-D
coherence concerns, it will not fail if such a page adds VM_EXEC flags
in the future since we have checked it in the __set_pte_at function.".
If a region indeed becomes executable, the page table will be modified
to reflect that and then will pass in __set_pte_at() which, as Yangyu
mentions, does the right thing. Or am I missing something?

Thanks,

Alex