[PATCH 13/13] nvme: introduce generic per-namespace chardev

[Javier Gonz??lez]

On 12.04.2021 09:44, Christoph Hellwig wrote:
>On Fri, Apr 09, 2021 at 08:24:01PM +0900, Minwoo Im wrote:
>> I think if we don't have a route to the live controller from the
>> multipath node /dev/ng0n1, how does application figure out controller
>> node to request admin commands like Identify Namespace before their
>> own I/O ?
>>
>> We have sysfs, but it does not provide every information about the
>> namespace.  Or is there any charming way to find out the live
>> controller from a head node through sysfs or something that I missed
>> here ? :)
>
>The devie links points to the subsystem, which then points to the
>controllers:
>
>root at testvm:/sys/class/nvme-generic/ng0n1# ls -l device/
>total 0
>-r--r--r-- 1 root root 4096 Apr 12 07:43 firmware_rev
>-rw-r--r-- 1 root root 4096 Apr 12 07:44 iopolicy
>-r--r--r-- 1 root root 4096 Apr 12 07:43 model
>drwxr-xr-x 3 root root    0 Apr 12 07:43 ng0n1
>drwxr-xr-x 3 root root    0 Apr 12 07:43 ng0n2
>lrwxrwxrwx 1 root root    0 Apr 12 07:44 nvme0 -> ../../nvme-fabrics/ctl/nvme0
>drwxr-xr-x 9 root root    0 Apr 12 07:43 nvme0n1
>drwxr-xr-x 9 root root    0 Apr 12 07:43 nvme0n2
>lrwxrwxrwx 1 root root    0 Apr 12 07:44 nvme1 -> ../../nvme-fabrics/ctl/nvme1
>drwxr-xr-x 2 root root    0 Apr 12 07:44 power
>-r--r--r-- 1 root root 4096 Apr 12 07:43 serial
>-r--r--r-- 1 root root 4096 Apr 12 07:44 subsysnqn
>lrwxrwxrwx 1 root root    0 Apr 12 07:43 subsystem -> ../../../../class/nvme-subsystem
>-rw-r--r-- 1 root root 4096 Apr 12 07:43 uevent
>

It seems like a limitation that we have to transverse sysfs to make an
in-application enumeration to submit namespace-specific admin commands.

Ideally, applications that are already using IOCTLs should be able to
directly work with the char device. If this is not the case, I am afraid
that the applicability will be limited.

I understand that identifying the live controller in multipath is
something that we need to improve. The question is that if this is a
reasonable thing for you. If there is no fundamental limitation here, I
think we can work the details of enabling namespace admin commands
through and merge the same support for the block device and the char
device.

What do you think?