On Mon, 2015-08-17 at 21:10 +0200, Christoph Hellwig wrote:
> #include <linux/uaccess.h>
> +#include <linux/uio.h>
Can we keep this in linux/pmem.h? I'm pretty sure the stubs would need
it as well, and even if they don't it'll keep the includes consistent.
> + size_t len;
> + len = copy_from_iter_nocache((void __force *)addr, bytes, i);
> + /*
> + * copy_from_iter_nocache() on x86 only uses non-temporal stores for
> + * iovec iterators, so for other types (bvec & kvec) we must do a
> + * cache write-back.
Shouldn't we fi that?
I'm not sure - When Al make copy_from_iter_nocache() it was just a copy of
copy_from_iter(), with the iovec case changed to use
__copy_from_user_nocache(). The other cases use memcpy_from_page() and
memcpy(). To have everything do non-temporal stores we'd probably need to
make non-temporal versions of each of those (alluded to by Al's comment in the
copy_from_iter_nocache() commit: "BTW, do we want memcpy_nocache()?").
> + */
> + if (iter_is_iovec(i) == false)
> + __arch_wb_cache_pmem(addr, bytes);
And if not and iter_needs_pmem_wb helper to encode this knowledge would
Maybe this should be the short-term solution, and I'll add a TODO to fix the
copy_from_iter_nocache() implementation as described above so we can always
have non-temporal stores?
> +static inline void arch_clear_pmem(void __pmem *addr, size_t
> + /* TODO: implement the zeroing via non-temporal writes */
> + if (size == PAGE_SIZE && ((unsigned long)addr & ~PAGE_MASK) == 0)
> + clear_page((void __force *)addr);
> + else
> + memset((void __force *)addr, 0, size);
> + __arch_wb_cache_pmem(addr, size);
Please add a local vaiable so that the __force casting is only needed
once. Same for other functions with this pattern.