On Tue, 25 May 2021 19:32:42 +0100,
Mark Rutland [off-list ref] wrote:

This series (based on v5.13-rc2) migrates most of the remaining
exception triage assembly to C. All the exception vectors are given C
handlers, so that we can defer all decision making to C code, and the
assembly code can be made simpler and more uniform. At the same time,
I've tried to consolidate all the entry sequencing (e.g. reading
exception registers and calling accounting code) in entry-common.c so
that this is easier to maintain.

I was recently informed that `noinstr` wasn't protecting entry sequences
from KCOV instrumentation, so I've refactored things so that we can
avoid this by preventing KCOV instrumentation for the entirety of
entry-common.c. I've done likewise for the low-level idle sequences
which have the same problems with instrumentation when RCU isn't
watching, etc.

I've stopped short of converting the ret_to_user / work_pending loop.
Converting this cleanly will probably need something like the wrappers
generated by SYSCALL_DEFINE() to handle the common entry/exit logic, and
this is easier to build once all the handlers have been converted to C.
Similar is true for portions of kernel_entry and kernel_exit that could
be converted to C.

It should also be possible to generate the vectors and their associated
assembly handlers in one go by placing these in separate sections and
using .pushsection and .popsection. I've held off doing this for now as
this probably requires some changes to the linker script, and regardless
it should be easier to make that change atop this series.

So far this has seen some light boot testing, and a day's worth of fuzzing
under Syzkaller, which I intend to leave to soak for a while.

I've pushed the series to my arm64/entry/rework branch on kernel.org:

  https://git.kernel.org/pub/scm/linux/kernel/git/mark/linux.git/log/?h=arm64/entry/rework
  git://git.kernel.org/pub/scm/linux/kernel/git/mark/linux.git arm64/entry/rework

I've tagged this version as arm64-entry-rework-20210525.

Since v1 [1]
* Rebase to v5.13-rc2
* Fold NMI entry/exit sequencing into entry-common.c
* Make NMI entry/exit helpers private to entry-comomn.c
* Prevent KCOV instrumentation of entry-common.c
* Prevent KCOV instrumentation of idle code

Since v2 [2]:
* Correct commit message description in patch 2
* Use `el0t` prefix for EL0 handlers for consistency with `el1h` and `el1t`
* Remove `user_exit_irqoff` macro
* Remove leftover bad_mode() prototype
* Add patch to replace el1_inv() with a direct call to __panic_unhandled()
* Accumulate Reviewed-by tags
* Typo fixes

[1] https://lore.kernel.org/r/20210510155621.52811-1-mark.rutland@arm.com (local)
[2] https://lore.kernel.org/r/20210519123902.2452-1-mark.rutland@arm.com (local)

I've been through the patches and couldn't spot anything wrong,
although I had to apply the patches and look at the final result.
I also tested it lightly on a few odd machines, and nothing caught
fire. So FWIW:

Acked-by: Marc Zyngier <maz@kernel.org>

	M.

-- 
Without deviation from the norm, progress is not possible.

_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel