[LU-12824] Unable to add single Infiniband interface to multiple o2ib LNets

• Type: Bug
• Resolution: Fixed
• Priority: Critical
• Fix Version/s: Lustre 2.13.0, Lustre 2.12.4
• Affects Version/s: Lustre 2.13.0, Lustre 2.12.3
• Labels: patch
• Severity: 3
• Assignee: Chris Horn
• Reporter: Chris Horn

Description

Configuring a single IB interface on multiple LNets was broken by

commit 75ab841d92a7109cf9f4da69a58ae4d21d360a4c
Author: James Simmons <[e-mail]>
Date:   Mon Jul 8 10:42:47 2019 -0700

   LU-11893 lnet: consoldate secondary IP address handling

Prior to this commit, when configuring an ib device for multiple LNets, we would only create a single struct ib_dev object. This object was created via a call to kiblnd_create_dev(). That function initializes the ib_dev object with a call to kiblnd_dev_failover(). kiblnd_dev_failover() creates the struct rdma_cm_id object, and calls rdma_bind_addr(). When the ib_dev object is created successfully, it is added to a global list of devices:

        list_add_tail(&dev->ibd_list,
                          &kiblnd_data.kib_devs);

When the interface is added to additional LNets, the kiblnd_startup() routine searches the kiblnd_data.kib_devs list to see if there is an existing ib_dev object for the interface being configured. If it finds one, then that ib_dev object is re-used.
The LU-11893 patch I noted above removed the logic for searching this list for an existing ib_dev object. It always creates a new ib_dev object, which I believe results in the EADDRINUSE.
It should be pretty straight forward to re-introduce the logic for searching the kib_devs list.

Reproducer with kernel module parameter:

[root@snx11922n002 ~]# cat /etc/lustre/ip2nets.dat
o2ib040(ib0) 10.12.0.*;
o2ib041(ib0) 10.12.0.50;
[root@snx11922n002 ~]# modprobe lnet
l[root@snx11922n002 ~]# lctl net up
LNET configure error 100: Network is down
[root@snx11922n002 ~]# dmesg | tail
[604327.506043] alg: No test for adler32 (adler32-zlib)
[604327.512517] alg: No test for crc32 (crc32-table)
[604328.280286] LNet: live_router_check_interval and dead_router_check_interval have been deprecated. Use alive_router_check_interval instead. Ignoring these deprecated parameters.
[604330.561491] LNet: 3809:0:(config.c:1641:lnet_inet_enumerate()) lnet: Ignoring interface eth2: it's down
[604330.591143] LNet: Using FastReg for registration
[604330.614353] LNet: Added LNI 10.12.0.50@o2ib40 [16/2048/0/0]
[604330.621410] LNetError: 3809:0:(o2iblnd.c:2776:kiblnd_dev_failover()) Failed to bind ib0:10.12.0.50 to device(ffff881f96ff8000): -98
[604330.636010] LNetError: 3809:0:(o2iblnd.c:3266:kiblnd_startup()) ko2iblnd: Can't initialize device: rc = -98
[604330.647163] LNetError: 105-4: Error -100 starting up LNI o2ib
[604331.659240] LNet: Removed LNI 10.12.0.50@o2ib40

Reproducer with lnetctl:

[root@snx11922n002 ~]# modprobe lnet
[root@snx11922n002 ~]# lctl mark mark
[root@snx11922n002 ~]# lnetctl lnet configure
[root@snx11922n002 ~]# lnetctl net add --net o2ib040 --if ib0
[root@snx11922n002 ~]# lnetctl net add --net o2ib041 --if ib0
add:
    - net:
          errno: -100
          descr: "cannot add network: Network is down"
[root@snx11922n002 ~]# dmesg | tail
[604760.221364] alg: No test for crc32 (crc32-table)
[604760.983433] LNet: live_router_check_interval and dead_router_check_interval have been deprecated. Use alive_router_check_interval instead. Ignoring these deprecated parameters.
[604763.557036] Lustre: DEBUG MARKER: mark
[604777.372005] LNet: 7487:0:(config.c:1641:lnet_inet_enumerate()) lnet: Ignoring interface eth2: it's down
[604777.382924] LNet: Using FastReg for registration
[604777.402400] LNet: Added LNI 10.12.0.50@o2ib40 [16/2048/0/0]
[604781.025699] LNet: 7528:0:(config.c:1641:lnet_inet_enumerate()) lnet: Ignoring interface eth2: it's down
[604781.036209] LNetError: 7528:0:(o2iblnd.c:2776:kiblnd_dev_failover()) Failed to bind ib0:10.12.0.50 to device(ffff881f96ff8000): -98
[604781.050103] LNetError: 7528:0:(o2iblnd.c:3266:kiblnd_startup()) ko2iblnd: Can't initialize device: rc = -98
[604781.060933] LNetError: 105-4: Error -100 starting up LNI o2ib
[root@snx11922n002 ~]#

Comments

[Gerrit Updater]

Minh Diep ([e-mail]) uploaded a new patch: https://review.whamcloud.com/36547
Subject: LU-12824 o2ib: Record rc in debug log on startup failure
Project: fs/lustre-release
Branch: b2_12
Current Patch Set: 1
Commit: 9d021ae9f819f8a15812c90af33a0604452b4bf9

[Gerrit Updater]

Minh Diep ([e-mail]) uploaded a new patch: https://review.whamcloud.com/36546
Subject: LU-12824 o2ib: Fix whitespace in kiblnd_startup
Project: fs/lustre-release
Branch: b2_12
Current Patch Set: 1
Commit: 1f04b73ce39a9d181d0ba689bbaf993f348ea250

[Gerrit Updater]

Minh Diep ([e-mail]) uploaded a new patch: https://review.whamcloud.com/36545
Subject: LU-12824 o2ib: Reintroduce kiblnd_dev_search
Project: fs/lustre-release
Branch: b2_12
Current Patch Set: 1
Commit: fe6666b21f421d0fd948489ce8d30c007f5d94f1

[Peter Jones]

I have marked it as a candidate for a future 2.12.x release.

[Alex Parga]

Is this fix expected to land for 2.12?

[Peter Jones]

Landed for 2.13

[Gerrit Updater]

Oleg Drokin ([e-mail]) merged in patch https://review.whamcloud.com/36326/
Subject: LU-12824 o2ib: Reintroduce kiblnd_dev_search
Project: fs/lustre-release
Branch: master
Current Patch Set:
Commit: e25e45c612a061031e8b4b5233137fbb57b50cc4

[Gerrit Updater]

Oleg Drokin ([e-mail]) merged in patch https://review.whamcloud.com/36325/
Subject: LU-12824 o2ib: Record rc in debug log on startup failure
Project: fs/lustre-release
Branch: master
Current Patch Set:
Commit: 99f85541a685df82265f18167e91c161c523ce50

[Gerrit Updater]

Oleg Drokin ([e-mail]) merged in patch https://review.whamcloud.com/36324/
Subject: LU-12824 o2ib: Fix whitespace in kiblnd_startup
Project: fs/lustre-release
Branch: master
Current Patch Set:
Commit: 50300e83e4cab3157149107eb735825cc4c3aff1

[Amir Shehata (Inactive)]

This config has historically been supported. LNet is designed to act as a virtual network over the physical network. One use case for this configuration is to segregate LNet traffic going over the same interface.

As to regards interaction with health, since at the LNet level these are two different NIDs, their health values will be managed independently. When there is a failure to send over one of these NIDs, then their health value will be decremented and added on the recovery queue. As far as I can see it should work at the LNet level.

The draw back I see with this type of configuration is performance and security. Performance since you're sharing the same link. Security because traffic is using the same link and you can just sniff traffic on both NIDs.

[James A Simmons]

At ORNL we implemented this differently. It just comes as a surprise that such a setup was possible. Talking to Olaf he had the same reaction. I'm not against supporting such a setup but with LNet health and fail over pairing I wonder what corner cases could exist. We should really exercise this in your LNet test suite  

I have had this happen in the past on other projects. The API is not clearly defined in some area and some company implements something no one expected. Then the change is brought before the standards board to sort it. In this case its Amir.