nvme_ctrl_alloc() can't handle duplicate connections

[tbzatek]

Okay, I admit my use case is a little weird and might not reflect real world scenario: loop transport, two (duplicate) connections, resulting in a single subsystem and two controllers in sysfs. The libnvme scanning logic would however only find one controller.

libnvme/src/nvme/tree.c

Lines 1270 to 1271 in ddd725e

	c = nvme_lookup_ctrl(s, transport, traddr,
	host_traddr, host_iface, trsvcid);

Looking at how nvme_lookup_ctrl() is called and how are the keys matched inside I see no way to distinguish seemingly duplicate controller entries, just having different name and sysfs path.

This is a broader issue, posting now in light of ongoing tree scanning logic rewrite in #108 to get some consideration.

---

# old/nvme list -v
NVM Express Subsystems

Subsystem        Subsystem-NQN                                                                                    Controllers
---------------- ------------------------------------------------------------------------------------------------ ----------------
nvme-subsys0     test_subnqn                                                                                      nvme0, nvme1

NVM Express Controllers

Device   SN                   MN                                       FR       TxPort Address        Subsystem    Namespaces      
-------- -------------------- ---------------------------------------- -------- ------ -------------- ------------ ----------------
nvme0    cce1b445357e708d9d62 Linux                                    5.14.17- loop                  nvme-subsys0 nvme0n1, nvme0n2, nvme0n3
nvme1    cce1b445357e708d9d62 Linux                                    5.14.17- loop                  nvme-subsys0 nvme0n1, nvme0n2, nvme0n3

NVM Express Namespaces

Device       NSID     Usage                      Format           Controllers     
------------ -------- -------------------------- ---------------- ----------------
nvme0n1      1          1.07  GB /   1.07  GB    512   B +  0 B   nvme0, nvme1
nvme0n2      2          1.07  GB /   1.07  GB    512   B +  0 B   nvme0, nvme1
nvme0n3      3          1.07  GB /   1.07  GB    512   B +  0 B   nvme0, nvme1

# git-master/nvme list -v
Subsystem        Subsystem-NQN                                                                                    Controllers
---------------- ------------------------------------------------------------------------------------------------ ----------------
nvme-subsys0     test_subnqn                                                                                      nvme1

Device   SN                   MN                                       FR       TxPort Address        Subsystem    Namespaces      
-------- -------------------- ---------------------------------------- -------- ------ -------------- ------------ ----------------
nvme1    cce1b445357e708d9d62 Linux                                    5.14.17- loop   (null)         nvme-subsys0 nvme0n3, nvme0n2, nvme0n1, nvme0n3, nvme0n2, nvme0n1

Device       Generic      NSID     Usage                      Format           Controllers     
------------ ------------ -------- -------------------------- ---------------- ----------------
nvme0n3      ng0n3        3          1.07  GB /   1.07  GB    512   B +  0 B   nvme1
nvme0n2      ng0n2        2          1.07  GB /   1.07  GB    512   B +  0 B   nvme1
nvme0n1      ng0n1        1          1.07  GB /   1.07  GB    512   B +  0 B   nvme1

[keithbusch]

Thanks for the report! I don't think your setup is at all unusual for a fabrics target (even if it is "loop"). I'll take a look.

[hreinecke]

Hmm. That is true, we don't have a way of distinguish between them.
Typically we wouldn't notice it as we don't allow for duplicate connections.

Guess the only way to fix this is to return a list of controllers ...

[igaw]

Ah, in this case we should at least change the current API so it is possible to return several controllers before we do our 1.0 release.

[hreinecke]

I've now prepared PR #219, which should address this issue. Can you test with that?

[hreinecke]

With PR #219 I now get:

/tmp/nvme list -v

Subsystem Subsystem-NQN Controllers

---

nvme-subsys0 nqn.2014-08.org.nvmexpress:NVMf:uuid:95096971-6a11-46a0-8fbc-c301194c5857 nvme2, nvme1, nvme0

Device SN MN FR TxPort Address Subsystem Namespaces

---

nvme2 f51d6e34723a74bea185 Linux 5.16.0-r loop traddr=2 nvme-subsys0
nvme1 f51d6e34723a74bea185 Linux 5.16.0-r loop traddr=2 nvme-subsys0
nvme0 f51d6e34723a74bea185 Linux 5.16.0-r loop traddr=1 nvme-subsys0

[hreinecke]

PR has been merged.

[tbzatek]

Works fine now in my original scenario, thanks.