Re: [PATCH v10 12/12] drivers/block/pmem: Map NVDIMM with ioremap_wt()

[Toshi Kani]

On Fri, 2015-05-29 at 11:19 -0700, Dan Williams wrote:
> On Fri, May 29, 2015 at 8:03 AM, Toshi Kani <toshi.kani@xxxxxx> wrote:
> > On Fri, 2015-05-29 at 07:43 -0700, Dan Williams wrote:
> >> On Fri, May 29, 2015 at 2:11 AM, Borislav Petkov <bp@xxxxxxxxx> wrote:
> >> > On Wed, May 27, 2015 at 09:19:04AM -0600, Toshi Kani wrote:
> >> >> The pmem driver maps NVDIMM with ioremap_nocache() as we cannot
 :
> >> >> -     pmem->virt_addr = ioremap_nocache(pmem->phys_addr, pmem->size);
> >> >> +     pmem->virt_addr = ioremap_wt(pmem->phys_addr, pmem->size);
> >> >>       if (!pmem->virt_addr)
> >> >>               goto out_release_region;
> >> >
> >> > Dan, Ross, what about this one?
> >> >
> >> > ACK to pick it up as a temporary solution?
> >>
> >> I see that is_new_memtype_allowed() is updated to disallow some
> >> combinations, but the manual seems to imply any mixing of memory types
> >> is unsupported.  Which worries me even in the current code where we
> >> have uncached mappings in the driver, and potentially cached DAX
> >> mappings handed out to userspace.
> >
> > is_new_memtype_allowed() is not to allow some combinations of mixing of
> > memory types.  When it is allowed, the requested type of ioremap_xxx()
> > is changed to match with the existing map type, so that mixing of memory
> > types does not happen.
> 
> Yes, but now if the caller was expecting one memory type and gets
> another one that is something I think the driver would want to know.
> At a minimum I don't think we want to get emails about pmem driver
> performance problems when someone's platform is silently degrading WB
> to UC for example.

The pmem driver creates an ioremap map to an NVDIMM range first.  So,
there will be no conflict at this point, unless there is a conflicting
driver claiming the same NVDIMM range.

DAX then uses the pmem driver (or other byte-addressable driver) to
mount a file system and creates a separate user-space mapping for
mmap().  So, a (silent) map-type conflict will happen at this point,
which may not be protected by the ioremap itself.

> > DAX uses vm_insert_mixed(), which does not even check the existing map
> > type to the physical address.
> 
> Right, I think that's a problem...
> 
> >> A general quibble separate from this patch is that we don't have a way
> >> of knowing if ioremap() will reject or change our requested memory
> >> type.  Shouldn't the driver be explicitly requesting a known valid
> >> type in advance?
> >
> > I agree we need a solution here.
> >
> >> Lastly we now have the PMEM API patches from Ross out for review where
> >> he is assuming cached mappings with non-temporal writes:
> >> https://lists.01.org/pipermail/linux-nvdimm/2015-May/000929.html.
> >> This gives us WC semantics on writes which I believe has the nice
> >> property of reducing the number of write transactions to memory.
> >> Also, the numbers in the paper seem to be assuming DAX operation, but
> >> this ioremap_wt() is in the driver and typically behind a file system.
> >> Are the numbers relevant to that usage mode?
> >
> > I have not looked into the Ross's changes yet, but they do not seem to
> > replace the use of ioremap_nocache().  If his changes can use WB type
> > reliably, yes, we do not need a temporary solution of using ioremap_wt()
> > in this driver.
> 
> Hmm, yes you're right, it seems those patches did not change the
> implementation to use ioremap_cache()... which happens to not be
> implemented on all architectures.  I'll take a look.

Thanks,
-Toshi

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/