On Wed, Feb 18, 2015 at 10:53 AM, Ingo Molnar <mingo(a)kernel.org> wrote:
* Dan Williams <dan.j.williams(a)intel.com> wrote:
> On Wed, Feb 18, 2015 at 10:30 AM, Ingo Molnar <mingo(a)kernel.org> wrote:
> > * Dan Williams <dan.j.williams(a)intel.com> wrote:
> >> On Tue, Feb 17, 2015 at 12:42 AM, Boaz Harrosh <boaz(a)plexistor.com>
> >> > On 02/17/2015 12:03 AM, Matthew Wilcox wrote:
> >> >> On Mon, Feb 16, 2015 at 01:07:07PM +0200, Boaz Harrosh wrote:
> >> >>> In any way this is a problem for the new type-12 NvDIMM memory
> >> >>> are circulating around. (It is estimated that there are already
> >> >>> thousands NvDIMM chips in active use)
> >> >>
> >> >> Hang on. NV-DIMM chips don't know anyhing about E820
> >> >> tables. They don't have anything in them that says "I
> >> >> am type 12!". How they are reported is up to the
> >> >> BIOS. Just because your BIOS vendor has chosen to
> >> >> report tham as type 12 doesn't mean that any other
> >> >> BIOS vedor is going to have done the same thing.
> >> >>
> >> >> Fortunately, the BIOS people have all got together and
> >> >> decided what they're going to do, and it's not type
> >> >> 12. Unfortunately, I think I'm bound by various
> >> >> agreements to not say what they are going to do until
> >> >> they do. But putting this temporary workaround in the
> >> >> kernel to accomodate one BIOS vendor's unreleased
> >> >> experimental code seems like entirely the wrong idea.
> >> >>
> >> >
> >> > I had a feeling I'm entering an holy war ;-).
> >> >
> >> > I hope you are OK with my first patch. That an unknown
> >> > type need not be reported busy, and behave same as
> >> > "reserved"?
> >> No, it seems the safe thing to do is prevent the
> >> kernel from accessing any memory that it does not know
> >> the side-effects of accessing.
> > Well, except when the kernel does know how to access
> > it: when an nvdimm driver knows about its own memory
> > region and knows how to handle it, right?
> Yes, except that "type-12" is something picked out of the
> air that may be invalidated by a future spec change.
> If firmware wants any driver to handle a memory range it
> can already use E820_RESERVED. The only reason a
> new-type was picked in these early implementations was
> for experiments around reserving nvdimm memory for driver
> use, but also extending it to be covered with struct page
> mappings. Outside of that there is no real driving
> reason for the new type.
But ... if a user is blessed/haunted with such firmware,
why not let new types fall back to 'reserved', which is a
reasonable default that still allows sufficiently aware
Linux drivers to work, right?
> > So is there any practical reason to mark the memory
> > resource as busy in that case, instead of just adding
> > it to the reserved list by default and allowing
> > properly informed drivers to (exclusively) request it?
> I'm not sure we want firmware to repeat this confusion
> going forward. Why support new memory types unless
> defined by ACPI or otherwise sufficiently described by
Because it would make the kernel more functional? We should
always err on the side of allowing more functionality and
not erect roadblocks.
I'm not convinced Linux is better off enabling one-off BIOS
implementations to pick non-standard numbers. Would it not be safer
to at least confirm with the user via a configuration option, "do you
want drivers to access unknown memory types"?