Re: [PATCH v3] dma-contiguous: setup default numa cma area if not configured explicitly
From: Robin Murphy
Date: Fri May 08 2026 - 08:58:56 EST
On 2026-05-08 12:46 pm, David Hildenbrand (Arm) wrote:
On 5/6/26 17:46, Feng Tang wrote:
On Fri, May 01, 2026 at 08:51:39PM +0200, David Hildenbrand (Arm) wrote:
On 4/28/26 10:37, Feng Tang wrote:
Hi David,
Hi!
[...]
Right for default kernel configs.
In kernel/dma/Kconfig:
config CMA_SIZE_MBYTES
int "Size in Mega Bytes"
depends on !CMA_SIZE_SEL_PERCENTAGE
default 0 if X86
default 16
config CMA_SIZE_PERCENTAGE
int "Percentage of total memory"
depends on !CMA_SIZE_SEL_MBYTES
default 0 if X86
default 10
Yes. I thought about that, and didn't have good solution, and used this
given it's on a multi-numa-node machine, which may not be too bad
regarding memory usage.
It sounds wrong given the existing config options.
Yes, it is confusing.
Robin did concern about the memory usage for embedded/small devices in
v2 review, and we change to v3 to not affect them.
I agree :)
This sounds good to me, if no objection from others. Maybe default 64MB
or more. One good part is, all these setup is under protection of
CONFIG_DMA_NUMA_CMA.
I cannot do the heavy thinking here because -EBUSY, but maybe you want a config
option similar to CMA_SIZE_MBYTES/CMA_SIZE_PERCENTAGE that either controls that
these sizes will be split over NUMA nodes, or another one, that sets the default
for pernuma.
Maybe a CMA_NUMA_SIZE_MBYTES?
Maybe, I'm hoping some CMA DMA people have the capacity to provide input.
But really that _is_ pretty much the idea here - we're effecting a kernel-level default "pernuma" value, which just happens to also be CMA_SIZE_*. But in the process we also need to tweak the "pernuma" behaviour itself to work as a default, since quietly forcing the current opt-in behaviour on single node systems could only hurt them - multiple default CMA areas on the same node offers no performance benefit, while reducing non-movable allocation capacity which could well be detrimental.
Indeed I am rather assuming that actual NUMA systems should have enough memory that this isn't a big deal, but I don't believe that's particularly unreasonable. End users should still be able to override with "numa_cma=0:0" if they don't want it, the only potential gap is if distros want to ship kernels with DMA_NUMA_CMA enabled for command-line opt-in but _without_ this new default behaviour. For that we could perhaps add something like:
config CMA_SIZE_PERNUMA
bool "Default CMA area per NUMA node"
depends on DMA_NUMA_CMA
default y
help
On systems with more than one NUMA node, the selected CMA
area size will be also allocated on each additional node,
so that most devices may have benefit from better DMA
locality without an explicit command-line opt-in.
Thanks,
Robin.
[...]
My thought was 'struct cma' already have 'nid' member, and when CONFIG_NUMA=y,
it may be useful to save the 'nid' info instead of NUMA_NO_NODE for the default
cma area (cmdline like cma=XXG@YYG could make it on different node)
Ah, yeah. It's a bit nasty that we have to handle the default area like that.
Another sign that we probably shouldn't deal with the default area :)
Yep, in v2 I didn't touch the default area, while Robin had a concern
that the v2 approach will bindly add an extra per-numa cma area for
the node which already has the default area, which will hurt those
small/embedded devices which has limited number of memory. Adding
the nid check is trying to keep the behavior of one node device
unchanged.
Per my understanding, it won't. There is a cma_validate_zones() to prevent it
from crossing zones.
It's a bit confusing, because it ignores other nids.
I might have missed your point. Do you mean one cma are could have
multiple ranges?
I don't know, it's confusing :)