nvme-cli : attaching a namespace to an undiscovered nvme controller on multi-controller nvme disk

[Nilay Shroff]

Hi,

I have got an NVMe disk with multi-controller support. This disk has two controllers
however kernel could only discover one of the controllers. On this disk if we create
a namespace and attach it to a controller (which is not discovered by the kernel) then
that namespace can't used for IO. In fact, there's no path to reach such a namespace
from kernel.

Please find the details below: 

# lspci 
0018:01:00.0 Non-Volatile memory controller: Samsung Electronics Co Ltd NVMe SSD Controller PM173X

# nvme list -v 
Subsystem        Subsystem-NQN                                                                                    Controllers
---------------- ------------------------------------------------------------------------------------------------ ----------------
nvme-subsys0     nqn.1994-11.com.samsung:nvme:PM1735:2.5-inch:S6EUNA0R500358                                      nvme0

Device   SN                   MN                                       FR       TxPort Asdress        Slot   Subsystem    Namespaces      
-------- -------------------- ---------------------------------------- -------- ------ -------------- ------ ------------ ----------------
nvme0    S6EUNA0R500358       1.6TB NVMe Gen4 U.2 SSD                  REV.SN49 pcie   0018:01:00.0          nvme-subsys0 

Device       Generic      NSID       Usage                      Format           Controllers     
------------ ------------ ---------- -------------------------- ---------------- ----------------

# nvme id-ctrl /dev/nvme0 -H 
NVME Identify Controller:
vid       : 0x144d
ssvid     : 0x1014
sn        : S6EUNA0R500358      
mn        : 1.6TB NVMe Gen4 U.2 SSD                 
fr        : REV.SN49
rab       : 8
ieee      : 002538
cmic      : 0x3
  [3:3] : 0	ANA not supported
  [2:2] : 0	PCI
  [1:1] : 0x1	Multi Controller  <== this is multi-controller disk
  [0:0] : 0x1	Multi Port
<snip>
<snip>

# nvme list-ctrl /dev/nvme0 
num of ctrls present: 2
[   0]:0x41
[   1]:0x42

# cat /sys/class/nvme/nvme0/cntlid 
65

So it's apparent from the above output that kernel has discovered one controller (nvme0)
with cntlid 65(0x41). The other controller with cntlid 0x42 remain undiscovered.
Please note even though kernel couldn't discover both disk controllers, it's possible to 
create and attach namespace to a controller which is not discovered by the kernel.

# nvme create-ns /dev/nvme0 --nsze=0x1749A956 --ncap=0x1749A956 --block-size=4096
create-ns: Success, created nsid:1

Now, attach namespace to a controller which is not discovered/known by the kernel.
# nvme attach-ns /dev/nvme0 -c 0x42 -n 1
attach-ns: Success, nsid:1

# nvme list 
Node                  Generic               SN                   Model                                    Namespace  Usage                      Format           FW Rev  
--------------------- --------------------- -------------------- ---------------------------------------- ---------- -------------------------- ---------------- --------

The nvme list doesn't show any namespace and the reason being the namespace is attached
to a controller which is not discovered by the kernel. The kernel doesn't have any path to 
access this namespace. So in essence now this namespace can't be used for any IO.

Looking at the above output (namespace list is empty) if we try creating a new namespace
then we see the below error.

# nvme create-ns /dev/nvme0 --nsze=0x1749A956 --ncap=0x1749A956 --block-size=4096
NVMe status: Namespace Insufficient Capacity: Creating the namespace requires more free space than is currently available(0x2115)

# nvme id-ctrl /dev/nvme0 -H 
NVME Identify Controller:
vid       : 0x144d
ssvid     : 0x1014
sn        : S6EUNA0R500358      
mn        : 1.6TB NVMe Gen4 U.2 SSD
<snip>
<snip>
tnvmcap   : 1600321314816
[127:0] : 1600321314816
	Total NVM Capacity (TNVMCAP)

unvmcap   : 0
[127:0] : 0
	Unallocated NVM Capacity (UNVMCAP)
<snip>
<snip>

We find from the above output that all available capacity of disk has been used but
nvme list output is empty.

Please note that "kernel is unable to discover both NVMe controllers" is not a kernel bug.
Also I think, this is not a bug with NVMe disk or its firmware. This multi controller NVMe 
disk has a "DualPortEn#" pin and by default, the disk is configured to operate a single x4 port. 
Asserting the "DualPortEn#" pin enables the Dual x2 port mode. When disk operates in Dual x2 mode,
kernel could discover both controllers. However when "DualportEn#" pin is not asserted
disk would only enable one of the controllers and hence kernel could discover one controller.

There could be multiple ways to address this issue. However my proposal would be to address 
it in nvme-cli. As nvme-cli builds the nvme topology it shall be easy for nvme-cli to detect 
this anomaly when a namespace is being attached to undiscovered nvme controller and shows some
WARNING message. The WARNING could be shown for ~10 seconds so that it gives enough time to a 
user to cancel this operation. If the operation is not cancelled then nvme-cli would proceed 
as usual.

Please help review and as usual, comments/feedback/suggestions are most welcome!

Thanks,
--Nilay