On 04/18/2013 07:51 PM, David Miller wrote:
> From: David Miller <davem@xxxxxxxxxxxxx>
> Date: Wed, 17 Apr 2013 18:44:45 -0400 (EDT)
>
>> It's not that difficult, I implemented a patch and will post it
>> after I'm done debugging it.
>
> Here's what I've been stress testing.
>
> This deals with a whole slew of problems all in one go.
>
> 1) Use generic smp_call_function_many() so that we don't have to write
> our own synchronization mechanism. This function does not return
> until all cross call siblings finish their work.
>
> 2) Use the paravirtualization hooks arch_{enter,leave}_lazy_mmu_mode,
> to guard when we actually do batching. This also makes sure we do
> batching inside of the page table locks which will be beneficial in
> the future.
>
> This idea is taken from powerpc.
>
> 3) When the batcher is not active, we flush a page at a time,
> synchronously.
>
> 4) Another side effect is that the batch cross call now runs with a
> full environment rather than within an interrupt vector trap
> handler. There has been talk of adding support for batched TLB
> flushes in the sun4v hypervisor, so that we don't have to perform a
> full hypervisor trap for every page, and having more registers to
> work with will facilitate being able to take advantage of that.
>
> 5) We no longer flush the batch from switch_to() which means we don't
> do it with interrupts disabled and the runqueue locks held.
>
> This passed a loop of make -j128 kernel builds on a SPARC-T4 as well
> as a gcc bootstrap and testsuite run. I plan to do some quick testing
> on a cheetah based system tomorrow in order to make sure the non-sun4v
> code paths work fine.
>
> Dave, can you see if this patch fixes your test case?
It does fix it. I do have one concern though...
> diff --git a/arch/sparc/kernel/smp_64.c b/arch/sparc/kernel/smp_64.c
> index 537eb66..33bd996 100644
> --- a/arch/sparc/kernel/smp_64.c
> +++ b/arch/sparc/kernel/smp_64.c
> @@ -849,7 +849,7 @@ void smp_tsb_sync(struct mm_struct *mm)
> }
>
> extern unsigned long xcall_flush_tlb_mm;
> -extern unsigned long xcall_flush_tlb_pending;
> +extern unsigned long xcall_flush_tlb_page;
> extern unsigned long xcall_flush_tlb_kernel_range;
> extern unsigned long xcall_fetch_glob_regs;
> extern unsigned long xcall_fetch_glob_pmu;
> @@ -1074,23 +1074,47 @@ local_flush_and_out:
> put_cpu();
> }
>
> +struct tlb_pending_info {
> + unsigned long ctx;
> + unsigned long nr;
> + unsigned long *vaddrs;
> +};
> +
> +static void tlb_pending_func(void *info)
> +{
> + struct tlb_pending_info *t = info;
> +
> + __flush_tlb_pending(t->ctx, t->nr, t->vaddrs);
> +}
> +
> void smp_flush_tlb_pending(struct mm_struct *mm, unsigned long nr, unsigned long *vaddrs)
> {
> u32 ctx = CTX_HWBITS(mm->context);
> + struct tlb_pending_info info;
> int cpu = get_cpu();
>
> + info.ctx = ctx;
> + info.nr = nr;
> + info.vaddrs = vaddrs;
> +
> if (mm == current->mm && atomic_read(&mm->mm_users) == 1)
> cpumask_copy(mm_cpumask(mm), cpumask_of(cpu));
> else
> - smp_cross_call_masked(&xcall_flush_tlb_pending,
> - ctx, nr, (unsigned long) vaddrs,
> - mm_cpumask(mm));
> + smp_call_function_many(mm_cpumask(mm), tlb_pending_func,
> + &info, 1);
>
> __flush_tlb_pending(ctx, nr, vaddrs);
>
> put_cpu();
> }
>
> +void smp_flush_tlb_page(unsigned long context, unsigned long vaddr)
> +{
> + smp_cross_call(&xcall_flush_tlb_page,
> + context, vaddr, 0);
> + __flush_tlb_page(context, vaddr);
> +}
> +
Why not pass mm into smp_flush_tlb_page() and use mm_cpumask(mm) as is
done in smp_flush_tlb_pending()? I don't see why we wouldn't want to
duplicate that logic to avoid cross calls to every cpu for every
non-batched flush.
> void smp_flush_tlb_kernel_range(unsigned long start, unsigned long end)
> {
> start &= PAGE_MASK;
--
To unsubscribe from this list: send the line "unsubscribe sparclinux" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
