Re: adjust prefetch in free_one_pgd()

[Linus Torvalds]

On Sat, 3 Aug 2002, David Woodhouse wrote:
>
> If you prefetch off the end of your object, and the thing you end up
> prefetching is a DMA buffer which we'd just nuked from our cache to allow
> the device to DMA into it, something's going to be unhappy regardless of
> whether your prefetch faults or not.

Yeah, there are broken prefetches around. So what else is new?

My personal opinion is that if a prefetch has semantic meanings outside
the "speed up subsequent accesses", it should not be exposed to the rest
of the kernel (it might still be useful inside architecture-specific
routines like optimized memcpy etc).

If your caches aren't speculation-safe, then prefetch isn't safe to use in
a generic manner. Our current use is probably ok right now, but if we get
future code where doing a prefetch on a pointer that we do not trust is a
good idea, I would personally suggest just disabling prefetch on machines
where that is broken.

Examples of this might be using prefetch on user-supplied addresses (which
might in turn have IO mappings or other issues). Clearly it's not worth
doing a VMA lookup to see if the address is prefetch-safe, so either we
speed up non-broken machines, or we leave everybody slow.

I'd rather speed up non-broken machines and let the broken hardware
hopefully slowly die away.

		Linus