On 04/28/16 20:07, Tom Yan wrote:
> Instead of requiring a serial of arbitrary length/format, I think a
> WWN/EUI-64 is more useful/important,

WWN/EUI-64 is not "more important". Section "7.9 Unique Identifier" in
the NVMe spec (Revision 1.2a, October 23, 2015) says that the serial
number is mandatory, while implementing an EUI-64 is optional. Let me
quote it all (emphases mine):

> 7.9 Unique Identifier
>
> Information is returned in the Identify Controller data structure that
> may be used to construct a unique identifier. Specifically, the PCI
> Vendor ID, *Serial Number*, and Model Number fields when combined
> shall form a globally unique value that identifies the NVM subsystem.
> The mechanism used by the vendor to assign Serial Number and Model
> Number values to ensure uniqueness is *outside the scope* of this
> specification.
>
> An NVM subsystem may contain multiple controllers. All of the
> controllers that make up an NVM subsystem share the same NVM subsystem
> identifier (i.e., PCI Vendor ID, Serial Number, and Model Number). The
> Controller ID (CNTLID) value returned in the Identify Controller data
> structure may be used to uniquely identify a controller within an NVM
> subsystem. The Controller ID value when combined with the NVM
> subsystem identifier forms a globally unique value that identifies the
> controller. The mechanism used by the vendor to assign Controller ID
> values is outside the scope of this specification.
>
> The Identify Namespace data structure contains the IEEE Extended
> Unique Identifier (EUI64) and the Namespace Globally Unique Identifier
> (NGUID) fields. EUI64 is an 8-byte EUI-64 identifier and NGUID is a
> 16-byte identifier based on EUI-64. When creating a namespace, the
> controller specifies a globally unique value in the EUI64 or NGUID
> field (the controller may optionally specify a globally unique value
> in both fields). In cases where the 64-bit EUI64 field is unable to
> ensure a globally unique namespace identifier, the EUI64 field shall
> be cleared to 0h. *When not implemented*, these fields contain a value
> of 0h.

The QEMU device model conforms to this:

- The serial number is mandatory, and its generation is unspecified.
(First paragraph quoted.) Accordingly, QEMU forces the user to generate
and provide a serial number.

- The EUI64 is optional (third paragraph); it shall be zero-filled when
not implemented. QEMU conforms.

> not to mention that the WWN/EUI-64
> can pretty much always be used as the serial at the same time.
> 
> Unlike Linux, Windows consider the WWN/EUI-64 as the "serial":

That's Windows's problem. Not the first (and not the last) occasion
where Microsoft interpret a specification creatively.

> "C:\Windows\system32>sg_vpd -i PD1
> Device Identification VPD page:
>   Addressed logical unit:
>     designator type: SCSI name string,  code set: UTF-8
>       SCSI name string:
>       8086QEMU NVMe Ctrl                          00012BDAC262CF831698
> 
> C:\Windows\system32>sg_vpd -p sn PD1
> Unit serial number VPD page:
>   Unit serial number: 0000_0000_0000_0000."
> 
> (https://bugs.launchpad.net/qemu/+bug/1576347/+attachment/4650553/+files/02.PNG)
> 
> UEFI also makes use of the WWN/EUI-64 to generate boot entries for NVMe devices:
> https://bugs.launchpad.net/qemu/+bug/1576347/+attachment/4650554/+files/03.png
> https://bugs.launchpad.net/qemu/+bug/1576347/+attachment/4650555/+files/04.png

The UEFI specification (version 2.6, January 2016) says in "9.3.5.22 NVM
Express namespace messaging device path node":

  Mnemonic:    IEEE Extended Unique Identifier
  Byte Offset: 8
  Byte Length: 8
  Description: This field contains the IEEE Extended Unique
               Identifier (EUI-64). Devices without an EUI-64 value
               must initialize this field with a value of 0.

QEMU conforms.

The device paths visible on your OVMF screenshots are distinguishable
from each other by their Pci() device path nodes. There is no collision.

I recommend reviewing the following commits:

http://git.qemu.org/?p=qemu.git;a=commitdiff;h=a907ec52cc1a
https://github.com/tianocore/edk2/commit/d7c0dfaef26c

The point being: if QEMU grows a capability to store a nonzero EUI64,
and that EUI64 is reflected in the OpenFirmware device path that is
placed into the "bootorder" fw_cfg file, then OVMF will parse it just
fine. However, QEMU is not required to grow such a capability, according
to the NVMe and UEFI specifications. In practice, multiple NVMe devices
can be distinguished from each other by their different PCI B/D/F locations.

Thanks,
Laszlo