Re: [PATCH v9 4/9] clocksource/drivers/arm_arch_timer: use readq to get 64-bit CNTVCT

From: Will Deacon
Date: Thu Jul 28 2016 - 09:53:33 EST


On Tue, Jul 26, 2016 at 09:11:49AM -0500, Timur Tabi wrote:
> Will Deacon wrote:
> >The kernel really needs to support both of those platforms :/
> >
> >For the memory-mapped counter registers, the architecture says:
> >
> > `If the implementation supports 64-bit atomic accesses, then the
> > CNTV_CVAL register must be accessible as an atomic 64-bit value.'
> >
> >which is borderline tautological. If we take the generous reading that
> >this means AArch64 CPUs can use readq (and I'm not completely
> >comfortable with that assertion, particularly as you say that it breaks
> >the model), then you still need to use readq_relaxed here to avoid a
> >DSB. Furthermore, what are you going to do for AArch32? readq doesn't
> >exist over there, and if you use the generic implementation then it's
> >not atomic. In which case, we end up with the current code, as well as a
> >readq_relaxed guarded by a questionable #ifdef that is known to break a
> >supported platform for an unknown performance improvement. Hardly a big
> >win.
>
> I know Fu dropped this patch, and I don't want to kick a dead horse, but I
> was wondering if it would be okay to do this:
>
> static u64 arch_counter_get_cntvct_mem(void)
> {
> #ifdef readq_relaxed
> return readq_relaxed(arch_counter_base + CNTVCT_LO);
> #else
> u32 vct_lo, vct_hi, tmp_hi;
>
> do {
> vct_hi = readl_relaxed(arch_counter_base + CNTVCT_HI);
> vct_lo = readl_relaxed(arch_counter_base + CNTVCT_LO);
> tmp_hi = readl_relaxed(arch_counter_base + CNTVCT_HI);
> } while (vct_hi != tmp_hi);
>
> return ((u64) vct_hi << 32) | vct_lo;
> #endif
> }
>
> readq and readq_relaxed are defined in arch/arm64/include/asm/io.h. Why
> would the function exist if AArch64 CPUs can't use it?
>
> Do we need something like ARCH_HAS_64BIT_ATOMIC_READ in order to decide
> whether readq is safe?

No, I'm still not ok with this. If you want to use readq_relaxed we need
the following guarantees:

1. readq_relaxed is provided by the architecture
2. readq_relaxed is single-copy atomic from the CPU's perspective
3. The memory-mapped timer has been integrated in such a way that it
can be accessed using 64-bit transactions.

(1) is easy, and you have that above. For (2), we just need to avoid
include/linux/io-64-nonatomic-*.h. (3), however, is not something we
can safely probe. If this optimisation really is worthwhile, then we
need to extend the device-tree binding for the counter so that we can
tell the kernel that it's ok to use 64-bit accesses for the counter
without tearing.

I have confirmed this with the architects here at ARM.

Will