[SRU][J][PATCH 0/2] rcu stalls with many storage key guests (LP: 1975582)

frank.heimes at canonical.com frank.heimes at canonical.com
Fri Jun 10 12:45:28 UTC 2022


BugLink: https://bugs.launchpad.net/bugs/1975582

SRU Justification:

[Impact] 

 * Ubuntu on s390x KVM environments with lots of large guests with storage
   keys can be affected by rcu stalls.

 * These rcu stalls can cause the system to crash/dump.

[Fix]

 * 3ae11dbcfac9 3ae11dbcfac906a8c3a480e98660a823130dc16a "s390/mm: use non-quiescing sske for KVM switch to keyed guest"

 * 6d5946274df1 6d5946274df1fff539a7eece458a43be733d1db8 "s390/gmap: voluntarily schedule during key setting"

[Test Plan]

 * There is no trigger or direct test or re-creation of the 
   problem situation possible, but...

 * and IBM z13 or LinuxONE (or never) LPAR is needed that
   runs Ubuntu Server 20.04 LTS or 18.04 LTS with HWE kernel
   and acts as KVM host with again several large guests running
   on top with storage groups.

 * Let such a system running for days under significant load
   and watch the logs for rcu issues.

 * Prior to the submission of this SRU patched test kernels
   for focal 5.4 and bionic hwe-5.4 were created and tested.
   They ran for days at a staging environemnt at IBM
   without further issues.

 * The modifications are all limited to s390x.

 * A test kernel was build (see below) that ran in a test environment
   at IBM under appropriate load for several days.

[Where problems could occur]

 * Due to the change for the KVM switch to keyed guest
   from classic sske to non-quiescing sske
   the KVM behaviour might have changed and the storage keys harmed.

 * The now more generous scheduling while setting keys
   has an impact on the guest memory management and mapping
   which will lead to a different performance.

 * This, with the introduction of __s390_enable_skey_pmd and
   cond_resched, might increase the overhead in certain situations,
   but eventually improves the responsiveness over time,
   hence avoid rcu stalls.

[Other Info]
 
 * Since the patches are upstream in 5.19-rc1,
   they will be included in the kernel that is planned for kinetic (5.19).

 * Hence this is an SRU to jammy, impish and focal.

Christian Borntraeger (2):
  s390/gmap: voluntarily schedule during key setting
  s390/mm: use non-quiescing sske for KVM switch to keyed guest

 arch/s390/mm/gmap.c    | 14 ++++++++++++++
 arch/s390/mm/pgtable.c |  2 +-
 2 files changed, 15 insertions(+), 1 deletion(-)

-- 
2.25.1



More information about the kernel-team mailing list