Re: [External] Re: [PATCH] RISC-V: add uniprocessor flush_tlb_range() support

[Jisheng Zhang]

On Mon, Jan 29, 2024 at 07:02:10PM +0800, yunhui cui wrote:
> Hi Jisheng,
> 
> On Mon, Jan 29, 2024 at 5:51 PM Jisheng Zhang <jszhang@xxxxxxxxxx> wrote:
> >
> > On Mon, Jan 29, 2024 at 04:26:57PM +0800, yunhui cui wrote:
> > > Hi Jisheng,
> > >
> > > On Mon, Jan 29, 2024 at 4:02 PM Jisheng Zhang <jszhang@xxxxxxxxxx> wrote:
> > > >
> > > > On Thu, Jan 25, 2024 at 02:20:44PM +0800, Yunhui Cui wrote:
> > > > > Add support for flush_tlb_range() to improve TLB performance for
> > > > > UP systems. In order to avoid the mutual inclusion of tlbflush.h
> > > > > and hugetlb.h, the UP part is also implemented in tlbflush.c.
> > > >
> > > > Hi Yunhui,
> > > >
> > > > IIRC, Samuel sent similar patch series a few weeks ago.
> > > >
> > > > https://lore.kernel.org/linux-riscv/20240102220134.3229156-1-samuel.holland@xxxxxxxxxx/
> > > >
> > > > After that series, do you still need this patch?
> > >
> > > Thank you for your reminder. I didn't find it before I mailed my
> > > patch. I just looked at the content of this patch. I understand that
> > > my patch is needed. For a single core, a more concise TLB flush logic
> > > is needed, and it is helpful to improve performance.
> >
> > Currently, riscv UP flush_tlb_range still use flush all TLB entries,
> > obviously it's is a big hammer, this is what your patch is trying to
> > optimize. I'm not sure whether I understand your code correctly or not.
> > Let me know if I misunderstand your code.
> >
> > After patch5 of the Samuel's series, __flush_tlb_range is unified for
> > SMP and UP, so that UP can also benefit from recent improvements, such
> > as range flush rather than all.
> 
> In my opinion, UP does not need to combine some SMP if... else,
> on_each_cpu(...) logic, which is also a manifestation of performance

Hi Yunhui,

IIRC, the compiler will optimise out the unnecessary logic under UP, I
may misread the code. But if no, indeed, there's improvement room.
However, even in this case, IMHO, it's better if you can base on
Samuel's series.
Anyway, the optimization(range tlb entries rather than *all* entries under
UP case) you want to do has been implemented. While I'm not sure whether
we can rely on the compiler to optimize out all unnecessary logics.

Thanks
> improvement. what do you think?
> Thanks,
> Yunhui