nvme disconnect -d fails to disconnect nvme block device

[pszczerx]

Issue description

The latest version of nvme CLI tool fails to disconnect nvme block device.
Example:
$ nvme disconnect -d /dev/nvme0n1
The command will succeed only if you pass NVMe character device (controller) as an argument.
Example:
$ sudo nvme disconnect -d /dev/nvme0

As far as I can tell, this issue was introduced by commit: 986a935
strcmp() function in line 986a935#diff-43e749b3fb33e89f12c38df0ad81a1c5R203 returns a non-zero return value, because content of char d[64]; array does not contain namespace suffix (i.e. device=nvme0n1 d=nvme0).

Older version of nvme CLI tool used to disconnect the controller when you passed a path to nvme block device.

nvme CLI help snippet

nvme --help                 
nvme-1.9.9.g5221                                                   
usage: nvme <command> [<device>] [<args>]                     
                                                              
The '<device>' may be either an NVMe character device (ex: /dev/nvme0) or an
nvme block device (ex: /dev/nvme0n1). 

Version Information

$ sudo nvme version
nvme version 1.9.9.g5221

Steps to reproduce

$ sudo nvme list | cut -d ' ' -f 1
Node
----------------
/dev/nvme0n1
/dev/nvme0n2
/dev/nvme0n3
/dev/nvme0n4
/dev/nvme0n5
/dev/nvme0n6
/dev/nvme0n7

$ sudo nvme disconnect -d /dev/nvme0n1
Failed to disconnect by device name: /dev/nvme0n1
$ 
$ sudo nvme list | cut -d ' ' -f 1
Node
----------------
/dev/nvme0n1
/dev/nvme0n2
/dev/nvme0n3
/dev/nvme0n4
/dev/nvme0n5
/dev/nvme0n6
/dev/nvme0n7

$ sudo nvme disconnect -d /dev/nvme0q
$ sudo nvme list | cut -d ' ' -f 1
$

Logs from nvme CLI 1.7

$ sudo nvme version
nvme version 1.7
$ sudo nvme list | cut -d ' ' -f 1
Node
----------------
/dev/nvme0n1
/dev/nvme0n2
/dev/nvme0n3
/dev/nvme0n4
/dev/nvme0n5
/dev/nvme0n6
/dev/nvme0n7
$ sudo nvme disconnect -d /dev/nvme0n1
$ sudo nvme list
$ 

Could you clarify whether this change was intentional?

Thanks for help,
Przemek

[keithbusch]

I think that was intentional due to the namespace and character name swizzling. I'll double check with the author.

[keithbusch]

Okay, the removal is done by writing to /sys/class/nvme/nvme/delete_controler, but you can't get to that unique path from a namespace handle with multipathing because the namespace may have more than one controller. We wouldn't be able to figure out the intended one to remove from a namespace handle.

[pszczerx]

Thanks for looking into the issue @keithbusch. If this change was intentional, then I think this issue can be closed.