On Wed, Apr 29, 2015 at 11:25 AM, Toshi Kani <toshi.kani(a)hp.com> wrote:
Thanks for the update. This version of the patchset enumerates our NFIT
table properly. :-)
On Tue, 2015-04-28 at 14:25 -0400, Dan Williams wrote:
> In preparation for the pmem driver attaching to pmem-namespaces emitted
> by libnd, convert it to use an ida instead of an always increasing
> atomic index. This provides a bit of stability to pmem device names in
> the presence of driver re-bind events.
> @@ -122,20 +123,26 @@ static struct pmem_device *pmem_alloc(struct device *dev,
struct resource *res)
> struct pmem_device *pmem;
> struct gendisk *disk;
> - int idx, err;
> + int err;
> err = -ENOMEM;
> pmem = kzalloc(sizeof(*pmem), GFP_KERNEL);
> if (!pmem)
> goto out;
> + pmem->id = ida_simple_get(&pmem_ida, 0, 0, GFP_KERNEL);
nd_pmem_probe() is called asynchronously via async_schedule_domain
(). We have seen a case that the region#->pmem# binding becomes
inconsistent across a reboot when there are 8 NVDIMM cards (reported by
Robert Elliott). This leads user to access a wrong device.
I think pmem id needs to be assigned before async_schedule_domain(), and
cascaded to nd_pmem_probe().
I'll take a look at making this better, but it will never be
bulletproof. For the same reason that root=UUID=<uuid> is preferred
over root=/dev/sda userspace should never rely on consistent pmem
device names from boot to boot.