11:25 p.m.
On Fri, May 5, 2017 at 1:39 PM, Kani, Toshimitsu <toshi.kani(a)hpe.com> wrote:
On Fri, 2017-04-28 at 12:39 -0700, Dan Williams wrote:
> The pmem driver has a need to transfer data with a persistent memory
> destination and be able to rely on the fact that the destination
> writes are not cached. It is sufficient for the writes to be flushed
> to a cpu-store-buffer (non-temporal / "movnt" in x86 terms), as we
> expect userspace to call fsync() to ensure data-writes have reached a
> power-fail-safe zone in the platform. The fsync() triggers a REQ_FUA
> or REQ_FLUSH to the pmem driver which will turn around and fence
> previous writes with an "sfence".
>
> Implement a __copy_from_user_inatomic_wt, memcpy_page_wt, and
> memcpy_wt, that guarantee that the destination buffer is not dirty in
> the cpu cache on completion. The new copy_from_iter_wt and sub-
> routines will be used to replace the "pmem api" (include/linux/pmem.h
> + arch/x86/include/asm/pmem.h). The availability of
> copy_from_iter_wt() and memcpy_wt() are gated by the
> CONFIG_ARCH_HAS_UACCESS_WT config symbol, and fallback to
> copy_from_iter_nocache() and plain memcpy() otherwise.
>
> This is meant to satisfy the concern from Linus that if a driver
> wants to do something beyond the normal nocache semantics it should
> be something private to that driver [1], and Al's concern that
> anything uaccess related belongs with the rest of the uaccess code
> [2].
>
> [1]: https://lists.01.org/pipermail/linux-nvdimm/2017-January/008364.
> html
> [2]: https://lists.01.org/pipermail/linux-nvdimm/2017-April/009942.ht
> ml
>
> Cc: <x86(a)kernel.org>
> Cc: Jan Kara <jack(a)suse.cz>
> Cc: Jeff Moyer <jmoyer(a)redhat.com>
> Cc: Ingo Molnar <mingo(a)redhat.com>
> Cc: Christoph Hellwig <hch(a)lst.de>
> Cc: "H. Peter Anvin" <hpa(a)zytor.com>
> Cc: Al Viro <viro(a)zeniv.linux.org.uk>
> Cc: Thomas Gleixner <tglx(a)linutronix.de>
> Cc: Matthew Wilcox <mawilcox(a)microsoft.com>
> Cc: Ross Zwisler <ross.zwisler(a)linux.intel.com>
> Signed-off-by: Dan Williams <dan.j.williams(a)intel.com>
> ---
> Changes since the initial RFC:
> * s/writethru/wt/ since we already have ioremap_wt(),
> set_memory_wt(), etc. (Ingo)
Sorry I should have said earlier, but I think the term "wt" is
misleading. Non-temporal stores used in memcpy_wt() provide WC
semantics, not WT semantics.
The non-temporal stores do, but memcpy_wt() is using a combination of
non-temporal stores and explicit cache flushing.
How about using "nocache" as it's been
used in __copy_user_nocache()?
The difference in my mind is that the "_nocache" suffix indicates
opportunistic / optional cache pollution avoidance whereas "_wt"
strictly arranges for caches not to contain dirty data upon completion
of the routine. For example, non-temporal stores on older x86 cpus
could potentially leave dirty data in the cache, so memcpy_wt on those
cpus would need to use explicit cache flushing.