Re: [PATCH RFC 0/8] kasan: hardware tag-based mode for production use on arm64

[Marco Elver <elver@google.com>]

On Wed, 14 Oct 2020 at 22:44, Andrey Konovalov <andreyknvl@google.com> wrote:
> This patchset is not complete (see particular TODOs in the last patch),
> and I haven't performed any benchmarking yet, but I would like to start the
> discussion now and hear people's opinions regarding the questions mentioned
> below.
>
> === Overview
>
> This patchset adopts the existing hardware tag-based KASAN mode [1] for
> use in production as a memory corruption mitigation. Hardware tag-based
> KASAN relies on arm64 Memory Tagging Extension (MTE) [2] to perform memory
> and pointer tagging. Please see [3] and [4] for detailed analysis of how
> MTE helps to fight memory safety problems.
>
> The current plan is reuse CONFIG_KASAN_HW_TAGS for production, but add a
> boot time switch, that allows to choose between a debugging mode, that
> includes all KASAN features as they are, and a production mode, that only
> includes the essentials like tag checking.
>
> It is essential that switching between these modes doesn't require
> rebuilding the kernel with different configs, as this is required by the
> Android GKI initiative [5].
>
> The last patch of this series adds a new boot time parameter called
> kasan_mode, which can have the following values:
>
> - "kasan_mode=on" - only production features
> - "kasan_mode=debug" - all debug features
> - "kasan_mode=off" - no checks at all (not implemented yet)
>
> Currently outlined differences between "on" and "debug":
>
> - "on" doesn't keep track of alloc/free stacks, and therefore doesn't
>   require the additional memory to store those
> - "on" uses asyncronous tag checking (not implemented yet)
>
> === Questions
>
> The intention with this kind of a high level switch is to hide the
> implementation details. Arguably, we could add multiple switches that allow
> to separately control each KASAN or MTE feature, but I'm not sure there's
> much value in that.
>
> Does this make sense? Any preference regarding the name of the parameter
> and its values?

KASAN itself used to be a debugging tool only. So introducing an "on"
mode which no longer follows this convention may be confusing.
Instead, maybe the following might be less confusing:

"full" - current "debug", normal KASAN, all debugging help available.
"opt" - current "on", optimized mode for production.
"on" - automatic selection => chooses "full" if CONFIG_DEBUG_KERNEL,
"opt" otherwise.
"off" - as before.

Also, if there is no other kernel boot parameter named "kasan" yet,
maybe it could just be "kasan=..." ?

> What should be the default when the parameter is not specified? I would
> argue that it should be "debug" (for hardware that supports MTE, otherwise
> "off"), as it's the implied default for all other KASAN modes.

Perhaps we could make this dependent on CONFIG_DEBUG_KERNEL as above.
I do not think that having the full/debug KASAN enabled on production
kernels adds any value because for it to be useful requires somebody
to actually look at the stacktraces; I think that choice should be
made explicitly if it's a production kernel. My guess is that we'll
save explaining performance differences and resulting headaches for
ourselves and others that way.

> Should we somehow control whether to panic the kernel on a tag fault?
> Another boot time parameter perhaps?

It already respects panic_on_warn, correct?

> Any ideas as to how properly estimate the slowdown? As there's no
> MTE-enabled hardware yet, the only way to test these patches is use an
> emulator (like QEMU). The delay that is added by the emulator (for setting
> and checking the tags) is different from the hardware delay, and this skews
> the results.
>
> A question to KASAN maintainers: what would be the best way to support the
> "off" mode? I see two potential approaches: add a check into each kasan
> callback (easier to implement, but we still call kasan callbacks, even
> though they immediately return), or add inline header wrappers that do the
> same.
[...]

Thanks,
-- Marco