Re: [PATCH 01/13] driver core, libnvdimm: disable manual unbind of dimms while region active

[Linda Knippers <linda.knippers@hpe.com>]

On 6/6/2016 3:46 PM, Dan Williams wrote:
> On Mon, Jun 6, 2016 at 12:36 PM, Linda Knippers <linda.knippers@hpe.com> wrote:
>>
>>
>> On 6/6/2016 3:31 PM, Dan Williams wrote:
>>> On Mon, Jun 6, 2016 at 12:25 PM, Linda Knippers <linda.knippers@hpe.com> wrote:
>>>> On 6/4/2016 4:52 PM, Dan Williams wrote:
>>>>> There are scenarios where we need a middle ground between disabling all
>>>>> manual bind/unbind attempts (via driver->suppress_bind_attrs) and
>>>>> allowing unbind at any userspace-determined time.  Pinning modules takes
>>>>> away one vector for unwanted out-of-sequence device_release_driver()
>>>>> invocations, this new mechanism (via device->suppress_unbind_attr) takes
>>>>> away another.
>>>>>
>>>>> The first user of this mechanism is the libnvdimm sub-system where
>>>>> manual dimm disabling should be prevented while the dimm is active in
>>>>> any region.  Note that there is a 1:N dimm-to-region relationship which
>>>>> is why this is implemented as a disable count rather than a flag.  This
>>>>> forces userspace to disable regions before dimms when manually shutting
>>>>> down a bus topology.
>>>>
>>>> How is this related to deprecating pcommit?
>>>
>>> We need guarantees that the flush hint mappings are valid for the
>>> duration of a pmem namespace being enabled.  I am going to move the
>>> mapping of the flush hint region from per-dimm to per-region.  However
>>> since multiple regions may reference the same dimm the mapping needs
>>> to be reference counted and shared across regions.  This will be
>>> similar to the arrangement we have for BLK-regions that share a
>>> control region mapping.
>>
>> Why are things moving around?  Aren't flush hints defined per NFIT device
>> handle, making them an optional per-dimm thing?
>>
>> I don't understand a lot of this patch series and had the same questions
>> as Jeff.  How does deprecating pcommit, because it's not necessary with ADR,
>> change so much?
> 
> This patch set deprecates pcommit, and introduces the usage of flush
> hints into the pmem path.  The introduction patch used the word
> "usually", I should have fleshed that out to say "usually ADR, or
> explicit use of flush hints".
> 
> A solution to the posted-write-queue flushing needs to be available
> and a platform can choose to use flush hints or ADR.  If the NFIT
> defines an NVDIMM device without hints we assume the platform must
> have ADR.  If the platform NFIT neglects to define an NVDIMM to
> physical address range mapping, we warn about a potentially broken
> BIOS.  Hopefully we can make this clearer in future versions of the
> spec.

You lost me on those last 2 sentences.  An NVDIMM doesn't have to have
an SPA range, but that seems to be unrelated to pcommit or flushes.

-- ljk