[LU-10937] Use sysfs to fix up sptlrpc handling.

• Type: Bug
• Resolution: Unresolved
• Priority: Major
• Affects Version/s: Lustre 2.12.0
• Environment: Lustre with sptlrpc/gss enabled.
• Severity: 3
• Assignee: James A Simmons
• Reporter: James A Simmons

Description

Several problems exist for the GSS / sptlrpc code. The problems are:

1) lctl set_param  P does not work with sptlrpc  LU-7183

2) Specific mgs bingings for sptlrpc is broken. LU-9034 / LU-9086 / LU-9567

    With the move to kobjects with sysfs we could the kobject instead.

3) After the move to sysfs we can use udev events instead of polling the proc files like what is now done in for example svgssd.

 

Issue Links

• is related to: LU-9567 sptlrpc rules are not being updated
• is related to: LU-9034 Separate the config logs between different MGCs
• is related to: LU-9086 obd_config.c:1258:class_process_config()) no device for:
• is related to: LU-7183 lctl set_param -P does not work for sptlrpc flavor
• is related to: LU-7004 fix "lctl set_param -P" to allow deprecation of "lctl conf_param"

Comments

[Gerrit Updater]

Oleg Drokin ([e-mail]) merged in patch https://review.whamcloud.com/33760/
Subject: LU-10937 sptlrpc: split sptlrpc_process_config()
Project: fs/lustre-release
Branch: master
Current Patch Set:
Commit: 0ff7d548eb7bf0c04836c7d6809cac163e0ffc2c

[Gerrit Updater]

James Simmons ([e-mail]) uploaded a new patch: https://review.whamcloud.com/33760
Subject: LU-10937 sptlrpc: split sptlrpc_process_config()
Project: fs/lustre-release
Branch: master
Current Patch Set: 1
Commit: 62c61bb1c4692f1c608047210befd9cb9dedc449

[James A Simmons]

I started to do a code review and I have noticed a bunch of potential bugs in general. This is just for the lctl conf_param case. The first question would be are the targets, the first *. field, just the following types:

lctl conf_param lustre.srpc.****

lctl conf_param lustre-OST0000.srpc.***

lctl conf_param lustre-MDT0000.srpc.***

for example.

If that is the case we can simply replace obdname2fsname() with server_name2fsname(). If server_name2fsname() returns an error then we know the target is a file system.

Next bug I noticed is that if we supply an obd device as a target plus a direction we don't validate the direction. The following should fail but doesn't

lctl conf_param lustre-MDT0000.srpc.flavor.default.cli2ost=skpi.

Lastly I noticed the network type supplied can easily break. Consider the case of a file system named test that are Cray clients and you have routers in between that convert to o2ib1 with infiniband storage backend. So if you do

lctl conf_param test.srpc.flavor.o2ib1=skpi

Does this filter so only the server back end is updated to skpi? I noticed we don't really test what LNet network interface is in use when setting the rule. Is it valid do a partial setup in this case?

 

[Gerrit Updater]

Oleg Drokin ([e-mail]) merged in patch https://review.whamcloud.com/33311/
Subject: LU-10937 mgc: restore mgc binding for sptlrpc
Project: fs/lustre-release
Branch: master
Current Patch Set:
Commit: ca9300e53dc2b7bcaaa5482bb4234cce7d9a344e

[James A Simmons]

Developers from Cray reported to me that the band aid fix landed for LU-9567 actually can cause an MSG server failover to crash the node. So I looked this after noon and figured out how to final resolve LU-9034. This will be needed for the LTS release.

[Gerrit Updater]

James Simmons ([e-mail]) uploaded a new patch: https://review.whamcloud.com/33311
Subject: LU-10937 mgc: restore mgc binding for sptlrpc
Project: fs/lustre-release
Branch: master
Current Patch Set: 1
Commit: 2347c82412b29c0777f5907551164d5cc27b80d4