Public bug reported:
Summary
Set CONFIG_CMA_SIZE_MBYTES=0 for arm64 linux-nvidia kernels.
This keeps CMA support compiled into the kernel, but stops reserving a
default global CMA pool unless one is explicitly requested with the cma=
kernel command-line option.
Rationale
The arm64 linux-nvidia kernels previously reserved CMA by default to reduce
noisy cma_alloc() failures seen during SMMUv3 initialization on large NVIDIA
systems. After further testing, the default CMA pool does not appear to be
required for the affected paths. The DMA/IOMMU allocation paths that try CMA
also have fallback paths through the normal page allocator, and testing
confirms that SMMUv3 queue allocation continues to succeed with cma=0.
Keeping the default CMA pool has become a liability on larger systems
because the pool can still be exhausted, which produces misleading
kernel log noise even though the allocation fallback succeeds. There is
also no existing knob to keep a default CMA pool while suppressing the
failed cma_alloc() messages.
Using CONFIG_CMA_SIZE_MBYTES=0 also aligns the arm64 linux-nvidia
default with amd64 behavior and with other enterprise arm64 distro
kernels, including RHEL and SLES, which use CONFIG_CMA_SIZE_MBYTES=0.
Users that need a default CMA reservation can still opt in with the cma=
kernel command-line parameter.
Compatibility
The main user-visible change is that CMA-backed dma-heap nodes are not created
by default when there is no default CMA area. For example, with CMA reserved
the system exposes:
/dev/dma_heap/default_cma_region
/dev/dma_heap/reserved
/dev/dma_heap/system
With cma=0, only the generic system heap remains:
/dev/dma_heap/system
This is a compatibility consideration for userspace that explicitly opens the
CMA-backed heap names. Generic dma-buf system heap users are unaffected.
Validation
Tested with cma=0 on linux-nvidia arm64 systems and guests, including 4K and
64K kernel combinations.
Observed behavior:
kernel command line contains cma=0
dmesg reports 0K cma-reserved
/proc/meminfo reports CmaTotal: 0 kB and CmaFree: 0 kB
no __cma_alloc failures
no DMA pool exhaustion or DMA out-of-memory messages
SMMUv3 cmdq/evtq/priq allocation succeeds
atomic DMA pool remains separately preallocated
/dev/dma_heap/system remains present
DGX Spark testing also showed a clean boot with cma=0; the only observed
dma-heap difference was the expected disappearance of the CMA-backed heap nodes.
** Affects: linux-nvidia (Ubuntu)
Importance: Undecided
Status: New
** Affects: linux-nvidia-bos (Ubuntu)
Importance: Undecided
Status: New
** Also affects: linux-nvidia (Ubuntu)
Importance: Undecided
Status: New
--
You received this bug notification because you are a member of Ubuntu
Bugs, which is subscribed to Ubuntu.
https://bugs.launchpad.net/bugs/2150898
Title:
UBUNTU: [Config] nvidia: Disable default CMA reservation
To manage notifications about this bug go to:
https://bugs.launchpad.net/ubuntu/+source/linux-nvidia/+bug/2150898/+subscriptions
--
ubuntu-bugs mailing list
[email protected]
https://lists.ubuntu.com/mailman/listinfo/ubuntu-bugs
