On Sat, 2015-08-15 at 08:44 -0700, Dan Williams wrote:
On Sat, Aug 15, 2015 at 2:19 AM, Christoph Hellwig <hch(a)lst.de>
> On Thu, Aug 13, 2015 at 10:51:11AM -0600, Ross Zwisler wrote:
>> Update the annotation for the kaddr pointer returned by direct_access()
>> so that it is a __pmem pointer. This is consistent with the PMEM driver
>> and with how this direct_access() pointer is used in the DAX code.
> IFF we stick to the __pmem annotations this looks good.
> That beeing said I start to really dislike them. We don't special
> accesors to read/write from pmem, we just need to explicitly commit
> it if we want to make it persistent. So I really don't see the need
> to treat it special and require all the force casts to and from the
I'm not going to put up much of a fight if it's really getting in the way....
That said, while we don't need special accessors we do need guarantees
that anything that has written to a persistent memory address has done
so in a way that wmb_pmem() is able to flush it. It's more of a "I've
audited this code path for wmb_pmem() compatibility so use this api to
write to pmem."
Perhaps a better way to statically check for missed flushes might be
to have acquire_pmem_for_write() + release() annotations and the final
release does a wmb_pmem(), but as far as I can tell the sparse
acquire/release annotations don't stack.
FWIW I've been on the fence about the __pmem annotations, but my current
thought is that we really do need a way of saying that stores to these
pointers need special care for wmb_pmem() to do its thing and that __pmem does
a reasonably good job of that. If we can figure out a cooler way, such as the
write() + release() flow Dan is talking about, great. But I think we need
something to keep us from making errors by storing to PMEM pointers and
leaving data in the processor cache.