Bug 1550471

We don't do anything special at all in sos to interrogate particular HCAs or ports. We just run the vanilla infiniband commands:

            "ibv_devices",
            "ibv_devinfo",
            "ibstat",
            "ibstatus",
            "ibhosts",
            "iblinkinfo",
            "sminfo",
            "perfquery"

It's not clear from the information you have provided which command you are referring to.

If additional options or arguments are required to fetch information for all HCAs and ports then we can add those.

On the other hand, if the above commands are expected to get information for all devices (which seems reasonable since no device was specified) then this is looking more like a bug in the infiniband tools.

(In reply to Bryn M. Reeves from comment #3)

>             "ibv_devices",
>             "ibv_devinfo",
>             "ibstat",
>             "ibstatus",

Those four commands report information about *ALL* HCA ports for all HCA. But I could suggest run "ibv_devices -v" replace "ibv_devices".

>             "ibhosts",
>             "iblinkinfo",
>             "sminfo",
>             "perfquery"


Those four commands only collect information from the first port.

> It's not clear from the information you have provided which command you are
> referring to.
> 
> If additional options or arguments are required to fetch information for all
> HCAs and ports then we can add those.
> 
> On the other hand, if the above commands are expected to get information for
> all devices (which seems reasonable since no device was specified) then this
> is looking more like a bug in the infiniband tools.

Yes, I agree that. We need add some switches to such infiniband tools to collect info for all ports.

BTW, I can't find any infiniband-diags commands in the sosrport file. Those information is _VERY_ useful to get topology information of the fabric. Is it possible to add such tools?

> run "ibv_devices -v" replace "ibv_devices".
> infiniband-diags commands

No problem: these are both trivial changes to the plugin that we can include in 3.6 (due in rhel-7.6).

We'll hold off on other changes for now until the question over whether to do this in the IB tools or sos is decided.

(In reply to Bryn M. Reeves from comment #5)
> > run "ibv_devices -v" replace "ibv_devices".
> > infiniband-diags commands
> 
> No problem: these are both trivial changes to the plugin that we can include
> in 3.6 (due in rhel-7.6).
> 
> We'll hold off on other changes for now until the question over whether to
> do this in the IB tools or sos is decided.

Raising relevant needinfo:

Waiting to be specified what changes are required from sosreport side, i.e. what particular commands sosreport shall newly collect (or what cmds shall it update, like the "ibv_devices -v").

bouncing the needinfo - much welcomed if the BZ shall go to 7.6.

Well, bit by this issue again in yesterday when NetApp file other mutlipath related bug.

>             "ibhosts",
>             "iblinkinfo",
>             "sminfo",
>             "perfquery"

We have two options to scan all active ports.

Option A: Modify all of those four C programs. Use "umad_get_cas_names"/"umad_get_ca_portguids"/"umad_get_port" to get all active ports. And then iterate over all active ports.

Option B: create a python wrap to get all active ports. And then run those commands iterate over all active ports.

iblinkinfo -C <CA> -P <Port NUM>
ibhosts    -C <CA> -P <Port NUM>
sminfo     -C <CA> -P <Port NUM>
perfquery  -C <CA> -P <Port NUM>

It is easy to get the CA list, just read the dir of "/sys/class/infiniband".

[root@rdma-master ~]$ ls /sys/class/infiniband
hfi1_0  hfi1_1  i40iw0  i40iw1  mlx4_0  qib0

It is also easy to get the port list, just read the dir of "/sys/class/infiniband/<CA>/ports/ .

[root@rdma-master ~]$ ls /sys/class/infiniband/*/ports/
/sys/class/infiniband/hfi1_0/ports/:
1

/sys/class/infiniband/hfi1_1/ports/:
1

/sys/class/infiniband/i40iw0/ports/:
1

/sys/class/infiniband/i40iw1/ports/:
1

/sys/class/infiniband/mlx4_0/ports/:
1  2

/sys/class/infiniband/qib0/ports/:
1


Now, read the file /sys/class/infiniband/<CA>/ports/<NUM>/state to get the state of port.

[root@rdma-master ~]$ tail  /sys/class/infiniband/*/ports/*/state 
==> /sys/class/infiniband/hfi1_0/ports/1/state <==
4: ACTIVE

==> /sys/class/infiniband/hfi1_1/ports/1/state <==
1: DOWN

==> /sys/class/infiniband/i40iw0/ports/1/state <==
1: DOWN

==> /sys/class/infiniband/i40iw1/ports/1/state <==
4: ACTIVE

==> /sys/class/infiniband/mlx4_0/ports/1/state <==
4: ACTIVE

==> /sys/class/infiniband/mlx4_0/ports/2/state <==
4: ACTIVE

==> /sys/class/infiniband/qib0/ports/1/state <==
4: ACTIVE


Is it doable for Option B?

B) shouldnt be a problem, but that info can be collected more easily - we dont need to identify and cat the files, we can easily grab whole /sys/class/infiniband directory ;-) or even /sys/class/infiniband/*/ports directories.

Is collecting any file with /sys/class/infiniband/*/ports mask sufficient? You can try it by updating /usr/lib/python2.7/site-packages/sos/plugins/infiniband.py to:


        self.add_copy_spec([
            "/etc/ofed/openib.conf",
            "/etc/ofed/opensm.conf",
            "/etc/rdma",                       # dont forget adding comma here
            "/sys/class/infiniband/*/ports"    # add this line
        ])

Is this change the only required, please?

No, it will not work. Option B is not to collect /sys/class/infiniband directory. After we get the active port list from those directory, we will run those four commands over the active ports.

For example:

iblinkinfo -C mlx4_0 -P 1
iblinkinfo -C mlx4_0 -P 2

I will create a python patch for you to review. But I'm not good and python programming. so the patch will be very ugly.

Created attachment 1416687 [details]
Loop all active IB ports

Thanks, with some minor improvements, I created PR https://github.com/sosreport/sos/pull/1262 for the same.

devel_ack+ for 7.6

Hello,
same question here - could you please do OtherQE here as well? :)

(In reply to Pavel Moravec from comment #14)
> Hello,
> same question here - could you please do OtherQE here as well? :)

infiniband-qe@ will take it

[root@rdma-virt-01 infiniband]$ ls
ibhosts_-C_mlx4_0_-P_1  iblinkinfo_-C_mlx4_0_-P_1  ibstat    ibv_devices     perfquery_-C_mlx4_0_-P_1  sminfo_-C_mlx4_0_-P_1
ibhosts_-C_mlx4_0_-P_2  iblinkinfo_-C_mlx4_0_-P_2  ibstatus  ibv_devinfo_-v  perfquery_-C_mlx4_0_-P_2  sminfo_-C_mlx4_0_-P_2
[root@rdma-virt-01 infiniband]$ cat sminfo_-C_mlx4_0_-P_1 
sminfo: sm lid 2 sm guid 0xf4521403007be131, activity count 12743948 priority 15 state 3 SMINFO_MASTER
[root@rdma-virt-01 infiniband]$ cat sminfo_-C_mlx4_0_-P_2
sminfo: sm lid 34 sm guid 0x1175000078b68e, activity count 1484999 priority 15 state 3 SMINFO_MASTER
[root@rdma-virt-01 infiniband]$ cat ibstat
CA 'mlx4_0'
	CA type: MT4103
	Number of ports: 2
	Firmware version: 2.40.7000
	Hardware version: 0
	Node GUID: 0xe41d2d03001d6790
	System image GUID: 0xe41d2d03001d6793
	Port 1:
		State: Active
		Physical state: LinkUp
		Rate: 56
		Base lid: 102
		LMC: 1
		SM lid: 2
		Capability mask: 0x02514868
		Port GUID: 0xe41d2d03001d6791
		Link layer: InfiniBand
	Port 2:
		State: Active
		Physical state: LinkUp
		Rate: 40
		Base lid: 74
		LMC: 1
		SM lid: 34
		Capability mask: 0x02514868
		Port GUID: 0xe41d2d03001d6792
		Link layer: InfiniBand
CA 'mlx4_1'
	CA type: MT4103
	Number of ports: 1
	Firmware version: 2.40.7000
	Hardware version: 0
	Node GUID: 0x0002c90300ee5860
	System image GUID: 0x0002c90300ee5863
	Port 1:
		State: Active
		Physical state: LinkUp
		Rate: 40
		Base lid: 0
		LMC: 0
		SM lid: 0
		Capability mask: 0x04010000
		Port GUID: 0x0202c9fffeee5860
		Link layer: Ethernet
[root@rdma-virt-01 infiniband]$ rpm -q sos
sos-3.6-3.el7.noarch

Since the problem described in this bug report should be
resolved in a recent advisory, it has been closed with a
resolution of ERRATA.

For information on the advisory, and where to find the updated
files, follow the link below.

If the solution does not work for you, open a new bug report.

https://access.redhat.com/errata/RHEA-2018:3144