[Devel] [PATCH hci-8.1 0/6] cgroup/dmem: backport fixes and apply a lower dmem.max at once

[Pavel Tikhomirov]

Reviewed-by: Pavel Tikhomirov <ptikhomirov at virtuozzo.com>

First 5 are clean cherry-pick, 6 has reasonable rework as reclaim
logic is excess in our kernel. All fixes are from stable and look
needed for proper dmem.max functioning.

On 11/2/26 00:51, Mirian Shilakadze wrote:
> Per-container GPU memory limits (VSTOR-135023) enable the dmem cgroup
> controller for Containers. Our dmem.c is upstream as of 8821f36333e2
> (6.14) and lacks four upstream fixes tagged for stable. The first one
> matters most: with dmem enabled for Containers, root in a Container can
> write a region name without a value to dmem.max of its own child
> cgroup, which dereferences NULL and, with CONFIG_PANIC_ON_OOPS=y,
> panics the node.
> 
> Patch 6 is an adapted 747c4bb450ad: a dmem.max below the current usage
> takes effect at once, as upstream, instead of being dropped while the
> write reports success. Its reclaim callback is left out, see the rebase
> note. For the NVIDIA driver this is the same as upstream. For amdgpu
> and xe a lower limit evicts at their next allocation instead of at the
> write. Patch 5 is the stable TTM fix that lets that eviction go on past
> the first buffer, upstream has it before 747c4bb450ad.
> 
> Tested on bg-sdc2-e18-04 (H200, NVIDIA 615.71.10 open modules) with the
> koji 211.52.1.22.14 SRPM plus this series, normal and debug kernels:
> - a region without a value in dmem.max returns -EINVAL, from the host
>   and from Container root
> - a lower dmem.max takes effect at once: while usage is above it a new
>   CUDA context fails, after release allocations stop at the new limit.
>   The stock kernel keeps the old limit and lets both through.
> - NVIDIA driver unload with lockdep active no longer warns about RCU in
>   dmem_cgroup_unregister_region(), as the stock debug kernel did, and
>   reload cycles give no KASAN reports
> - a failed pool preallocation (failslab, debug kernel) makes the write
>   return -ENOMEM
> Patch 5 and the amdgpu and xe paths are build-tested only, the host has
> no such GPU.
> 
> https://virtuozzo.atlassian.net/browse/VSTOR-152813
> 
> Chen Ridong (3):
>   ms/cgroup/dmem: fix NULL pointer dereference when setting max
>   ms/cgroup/dmem: avoid rcu warning when unregister region
>   ms/cgroup/dmem: avoid pool UAF
> 
> Guopeng Zhang (1):
>   ms/cgroup/dmem: Return -ENOMEM on failed pool preallocation
> 
> Thomas Hellström (2):
>   ms/drm/ttm: Convert -EAGAIN from dmem_cgroup_try_charge to -ENOSPC
>   ms/cgroup/dmem: Add reclaim callback for lowering max below current
>     usage
> 
>  drivers/gpu/drm/ttm/ttm_bo.c       |  2 +-
>  drivers/gpu/drm/ttm/ttm_resource.c |  5 +-
>  kernel/cgroup/dmem.c               | 76 ++++++++++++++++++++++++++----
>  3 files changed, 73 insertions(+), 10 deletions(-)
> 

-- 
Best regards, Pavel Tikhomirov
Senior Software Developer, Virtuozzo.