Re: [PATCH] KFENCE: Clarify that sample allocations are not following NUMA or memory policies

[Marco Elver <elver@google.com>]

On Thu, 23 Jan 2025 at 23:44, Christoph Lameter via B4 Relay
<devnull+cl.gentwo.org@kernel.org> wrote:
>
> From: Christoph Lameter <cl@linux.com>
>
> KFENCE manages its own pools and redirects regular memory allocations
> to those pools in a sporadic way. The usual memory allocator features
> like NUMA, memory policies and pfmemalloc are not supported.
> This means that one gets surprising object placement with KFENCE that
> may impact performance on some NUMA systems.
>
> Update the description and make KFENCE depend on VM debugging
> having been enabled.

While the documentation updates are fine with me, the Kconfig change
seems overly drastic. What's the motivation? CONFIG_KFENCE is not
enabled by default, and if there's a problem users are free to either
not select it in the first place, or if you cannot unset CONFIG_KFENCE
because you have a prebuilt kernel, set 'kfence.sample_interval=0' in
the kernel cmdline. More commentary below.

> Signed-off-by: Christoph Lameter <cl@linux.com>
> ---
>  Documentation/dev-tools/kfence.rst |  4 +++-
>  lib/Kconfig.kfence                 | 10 ++++++----
>  2 files changed, 9 insertions(+), 5 deletions(-)
>
> diff --git a/Documentation/dev-tools/kfence.rst b/Documentation/dev-tools/kfence.rst
> index 541899353865..27150780d6f5 100644
> --- a/Documentation/dev-tools/kfence.rst
> +++ b/Documentation/dev-tools/kfence.rst
> @@ -8,7 +8,9 @@ Kernel Electric-Fence (KFENCE) is a low-overhead sampling-based memory safety
>  error detector. KFENCE detects heap out-of-bounds access, use-after-free, and
>  invalid-free errors.
>
> -KFENCE is designed to be enabled in production kernels, and has near zero
> +KFENCE is designed to be low overhead but does not implememnt the typical

s/implememnt/implement/

> +memory allocation features for its samples like memory policies, NUMA and
> +management of emergency memory pools. It has near zero
>  performance overhead. Compared to KASAN, KFENCE trades performance for
>  precision. The main motivation behind KFENCE's design, is that with enough
>  total uptime KFENCE will detect bugs in code paths not typically exercised by
> diff --git a/lib/Kconfig.kfence b/lib/Kconfig.kfence
> index 6fbbebec683a..48d2a6a1be08 100644
> --- a/lib/Kconfig.kfence
> +++ b/lib/Kconfig.kfence
> @@ -5,14 +5,14 @@ config HAVE_ARCH_KFENCE
>
>  menuconfig KFENCE
>         bool "KFENCE: low-overhead sampling-based memory safety error detector"
> -       depends on HAVE_ARCH_KFENCE
> +       depends on HAVE_ARCH_KFENCE && DEBUG_VM

This is not going to work. There are plenty deployments of KFENCE in
kernels that do not enable DEBUG_VM, and this will silently disable
KFENCE once those kernels upgrade. And enabling DEBUG_VM is not what
anybody wants, because enabling DEBUG_VM adds features significantly
more expensive than KFENCE, even if disabled they pull in code and
increase .text size.

Nack with the dependency on DEBUG_VM. The documentation change is fine.

>         select STACKTRACE
>         select IRQ_WORK
>         help
>           KFENCE is a low-overhead sampling-based detector of heap out-of-bounds
>           access, use-after-free, and invalid-free errors. KFENCE is designed
> -         to have negligible cost to permit enabling it in production
> -         environments.
> +         to have negligible cost. KFENCE does not support NUMA features
> +         and other memory allocator features for it sample allocations.

s/sample/samples/

>           See <file:Documentation/dev-tools/kfence.rst> for more details.
>
> @@ -21,7 +21,9 @@ menuconfig KFENCE
>           detect, albeit at very different performance profiles. If you can
>           afford to use KASAN, continue using KASAN, for example in test
>           environments. If your kernel targets production use, and cannot
> -         enable KASAN due to its cost, consider using KFENCE.
> +         enable KASAN due to its cost and you are not using NUMA and have
> +         no use of the memory reserve logic of the memory allocators,
> +         consider using KFENCE.

That's just repetition from above, and I think the point here is just
that if you run tests but can't use KASAN, consider KFENCE. In those
cases, users typically would use much higher sampling rates that
_will_ have somewhat noticeable performance impact.

Thanks.