4:23 p.m.
On Thu, Apr 30, 2015 at 6:21 PM, Rafael J. Wysocki <rjw(a)rjwysocki.net> wrote:
On Thursday, April 30, 2015 05:39:06 PM Dan Williams wrote:
> On Thu, Apr 30, 2015 at 4:23 PM, Rafael J. Wysocki <rjw(a)rjwysocki.net> wrote:
[..]
> >> +if ND_DEVICES
> >> +
> >> +config LIBND
> >> + tristate "LIBND: libnd device driver support"
> >> + help
> >> + Platform agnostic device model for a libnd bus. Publishes
> >> + resources for a PMEM (persistent-memory) driver and/or BLK
> >> + (sliding mmio window(s)) driver to attach. Exposes a device
> >> + topology under a "ndX" bus device, a
"/dev/ndctlX" bus-ioctl
> >> + message passing interface, and a "/dev/nmemX" dimm-ioctl
> >> + message interface for each memory device registered on the
> >> + bus. instance. A userspace library "ndctl" provides an
API
> >> + to enumerate/manage this subsystem.
> >> +
> >> +config ND_ACPI
> >> + tristate "ACPI: NFIT to libnd bus support"
> >> + select LIBND
> >> + depends on ACPI
> >> + help
> >> + Infrastructure to probe ACPI 6 compliant platforms for
> >> + NVDIMMs (NFIT) and register a libnd device tree. In
> >> + addition to storage devices this also enables libnd craft
> >> + ACPI._DSM messages for platform/dimm configuration.
> >
> > I'm wondering if the two CONFIG options above really need to be
user-selectable?
> >
> > For example, what reason people (who've already selected ND_DEVICES) may
have
> > for not selecting ND_ACPI if ACPI is set?
>
>
> Later on in the series we introduce ND_E820 which supports creating a
> libnd-bus from e820-type-12 memory ranges on pre-NFIT systems. I'm
> also considering a configfs defined libnd-bus because e820 types are
> not nearly enough information to safely define nvdimm resources
> outside of NFIT.
I hope these are not mutually exclusive with ND_ACPI? Otherwise distros
will have problems with supporting them in one kernel.
You can have ND_E820 support and ND_ACPI support in the same system.
Likely an NFIT enabled system will never have e820-type-12 ranges, but
if a user messes up and uses the new memmap=ss!nn command line to
overlap NFIT-defined memory then the request_mem_region() calls in the
driver will collide. First to load wins in that scenario.
If ND_E820 and ND_ACPI aren't mutually exclusive, I still
don't see a good
enough reason for asking users about ND_ACPI. Why would I ever say "No"
here if I said "Yes" or "Module" to ND_DEVICES?
I agree that if the user selects ND_DEVICES then ND_ACPI should
probably default on, but otherwise turning it off is a useful option.
If you know your system is pre-ACPI-6 then why bother including
support?
> >> +
> >> +endif
> >> diff --git a/drivers/block/nd/Makefile b/drivers/block/nd/Makefile
> >> new file mode 100644
> >> index 000000000000..944b5947c0cb
> >> --- /dev/null
> >> +++ b/drivers/block/nd/Makefile
> >> @@ -0,0 +1,6 @@
> >> +obj-$(CONFIG_LIBND) += libnd.o
> >> +obj-$(CONFIG_ND_ACPI) += nd_acpi.o
> >> +
> >> +nd_acpi-y := acpi.o
> >> +
> >> +libnd-y := core.o
> >
> > OK, so it looks like no modules, just built-in code, right?
> >
>
> Um, no, both CONFIG_ND_ACPI and CONFIG_LIBND can be =m.
OK
[cut]
> >> +static int nd_acpi_remove(struct acpi_device *adev)
> >> +{
> >> + struct acpi_nfit_desc *acpi_desc =
dev_get_drvdata(&adev->dev);
> >> +
> >> + nd_bus_unregister(acpi_desc->nd_bus);
> >> + return 0;
> >> +}
> >> +
> >> +static void nd_acpi_notify(struct acpi_device *adev, u32 event)
> >> +{
> >> + /* TODO: handle ACPI_NOTIFY_BUS_CHECK notification */
> >> + dev_dbg(&adev->dev, "%s: event: %d\n", __func__,
event);
> >> +}
> >> +
> >> +static const struct acpi_device_id nd_acpi_ids[] = {
> >> + { "ACPI0012", 0 },
> >> + { "", 0 },
> >> +};
> >> +MODULE_DEVICE_TABLE(acpi, nd_acpi_ids);
> >> +
> >> +static struct acpi_driver nd_acpi_driver = {
> >> + .name = KBUILD_MODNAME,
> >> + .ids = nd_acpi_ids,
> >> + .flags = ACPI_DRIVER_ALL_NOTIFY_EVENTS,
> >> + .ops = {
> >> + .add = nd_acpi_add,
> >> + .remove = nd_acpi_remove,
> >> + .notify = nd_acpi_notify
> >> + },
> >> +};
> >
> > Since this is going to be non-modular built-in code, please use an ACPI
> > scan handler instead of using a driver here. acpi_memhotplug.c does that,
> > you can use it as an example, but I guess you don't need to enable hotplug
> > for it to start with.
>
>
> No, you misunderstood, this will certainly be modular and loaded on-demand.
OK
So please drop the .notify thing at least for now. It most likely doesn't do
what you need anyway.
The .notify handler will eventually be filled in to handle hot-add of
NFIT structures, but yes I'll drop it for now.