APPLIED: [SRU][J][I][F][PATCH 0/1] KVM nesting support leaks too much memory, might result in stalls during cleanup (LP: 1974017)

[Kleber Souza]

On 19.05.22 08:39, frank.heimes at canonical.com wrote:
> BugLink: https://bugs.launchpad.net/bugs/1974017
> 
> SRU Justification:
> 
> [Impact]
> 
>   * If running KVM with nesting support (e.g. 'kvm.nested=1' on the kernel
>     command line), the shadow page table code will produce too many entries
>     in the shadow code.
> 
>   * The below mentioned upstream fix will prevent the entries from being
>     piled up, by checking for existing entries at insert time.
> 
>   * This measurably reduces the list length and is faster than traversing
>     the list at shutdown time only.
> 
> [Fix]
> 
>   * a06afe8383080c630a7a528b8382fc6bb4925b61 a06afe838308
>     "KVM: s390: vsie/gmap: reduce gmap_rmap overhead"
> 
> [Test Plan]
> 
>   * A IBM zSystems or LinuxONE LPAR on a z13 or newer is needed.
> 
>   * Ubuntu focal, impish or jammy needs to be installed
>     and the Ubuntu LPAR setup as (1st level) KVM host,
>     allowing nested virtualization.
> 
>   * Now setup one (or more) KVM virtual machines,
>     with similar Ubuntu releases,
>     and define one or more of them again as (2nd level) KVM host.
> 
>   * Define several KVM virtual machines on this (2nd level) KVM host
>     in a memory constraint fashion,
>     so that a lot of memory mapping is caused.
> 
>   * Let such a system run for a while under load.
> 
>   * Now shutdown one (or more) 2nd level VMs and notice the
>     time it takes.
> 
>   * With the patch in place this time should be considerably
>     quicker than without.
> 
>   * The result is reduced mapping (gmap_rmap) overhead,
>     less danger of leaking memory
>     and a better responding system.
> 
> [Where problems could occur]
> 
>   * In case wrong entries are freed up this will harm the virtual
>     memory management and may even lead to crashes.
> 
>   * In case the pointer handling is not done properly,
>     again crashes may occur.
> 
>   * But with net just five new lines the patch is pretty short, readable
>     and the modifications traceable in arch/s390/mm/gmap.c only.
> 
>   * The changes are limited to s390/mm only,
>     hence don't affect other architectures.
> 
> [Other Info]
>   
>   * The commit was upstream accepted in v5.18-rc6.
> 
>   * Since the planned target kernel for kinetic is 5.19,
>     the kinetic kernel does not need to be patched.
> 
>   * Hence the SRUs are for jammy, impish and focal.
> 
> Christian Borntraeger (1):
>    KVM: s390: vsie/gmap: reduce gmap_rmap overhead
> 
>   arch/s390/mm/gmap.c | 7 +++++++
>   1 file changed, 7 insertions(+)
> 

Applied to focal/impish/jammy:linux.

Thanks,
Kleber