Re: [kvm-unit-tests PATCH v5 00/29] EFI and ACPI support for arm64

[Andrew Jones <andrew.jones@linux.dev>]

On Fri, Apr 28, 2023 at 01:03:36PM +0100, Nikos Nikoleris wrote:
> Hello,
> 
> This series adds initial support for building arm64 tests as EFI
> apps and running them under QEMU. Much like x86_64, we import external
> dependencies from gnu-efi and adapt them to work with types and other
> assumptions from kvm-unit-tests. In addition, this series adds support
> for enumerating parts of the system using ACPI.
> 
> The first set of patches, moves the existing ACPI code to the common
> lib path. Then, it extends definitions and functions to allow for more
> robust discovery of ACPI tables. We add support for setting up the PSCI
> conduit, discovering the UART, timers, GIC and cpus via ACPI. The code
> retains existing behavior and gives priority to discovery through DT
> when one has been provided.
> 
> In the second set of patches, we add support for getting the command
> line from the EFI shell. This is a requirement for many of the
> existing arm64 tests.
> 
> In the third set of patches, we import code from gnu-efi, make minor
> changes and add an alternative setup sequence from arm64 systems that
> boot through EFI. Finally, we add support in the build system and a
> run script which is used to run an EFI app.
> 
> After this set of patches one can build arm64 EFI tests:
> 
> $> ./configure --enable-efi
> $> make
> 
> And use the run script to run an EFI tests:
> 
> $> ./arm/efi/run ./arm/selftest.efi -smp 2 -m 256 -append "setup smp=2 mem=256"
> 
> Or all tests:
> 
> $> ./run_tests.sh
> 
> There are a few items that this series does not address but they would
> be useful to have:
>  - Support for booting the system from EL2. Currently, we assume that a
>    test starts EL1. This will be required to run EFI tests on sytems
>    that implement EL2.
>  - Support for reading environment variables and populating __envp.
>  - Support for discovering the PCI subsystem using ACPI.
>  - Get rid of other assumptions (e.g., vmalloc area) that don't hold on
>    real HW.
>  - Various fixes related to cache maintaince to better support turn the
>    MMU off.
>  - Switch to a new stack and avoid relying on the one provided by EFI.
> 
> git branch: https://github.com/relokin/kvm-unit-tests/pull/new/target-efi-upstream-v5
> 
> v4: https://lore.kernel.org/kvmarm/20230213101759.2577077-1-nikos.nikoleris@arm.com/
> v3: https://lore.kernel.org/all/20220630100324.3153655-1-nikos.nikoleris@arm.com/
> v2: https://lore.kernel.org/kvm/20220506205605.359830-1-nikos.nikoleris@arm.com/
> 
> Changes in v5:
>  - Minor style changes (thanks Shaoqin).
>  - Avoid including lib/acpi.o to cflatobjs twice (thanks Drew).
>  - Increase NR_INITIAL_MEM_REGIONS to avoid overflows and add check when
>    we run out of space (thanks Shaoqin).
> 
> Changes in v4:
>  - Removed patch that reworks cache maintenance when turning the MMU
>    off. This is not strictly required for EFI tests running with tcg and
>    will be addressed in a separate series by Alex.
>  - Fix compilation for arm (Alex).
>  - Convert ACPI tables to Linux style (Alex).
> 
> Changes in v3:
>  - Addressed feedback from Drew, Alex and Ricardo. Many thanks for the reviews!
>  - Added support for discovering the GIC through ACPI
>  - Added a missing header file (<elf.h>)
>  - Added support for correctly parsing the outcome of tests (./run_tests)
>

Thanks, Nikos!

I'd like to get an ack from either Paolo or Sean on the changes to ACPI,
as they're shared with x86, and there are also some x86 code changes.

Also,

  1) It'd be nice if this worked with DT, too. We can use UEFI with DT
     when adding '-no-acpi' to the QEMU command line. setup_efi() needs
     to learn how to find the dtb and most the '#ifdef CONFIG_EFI's
     would need to change to a new CONFIG_ACPI guard.

  2) The debug bp and ss tests fail with EFI, but not without, for me.

  3) The timer test runs (and succeeds) when run with
     './arm/efi/run ./arm/timer.efi', but not when run with
     './run_tests.sh -g timer'. This is because UEFI takes
     up all the given timeout time (10s), and then the test times out.
     The hackyish fix below resolves it for me. I'll consider posting it
     as a real patch

Thanks,
drew

diff --git a/scripts/arch-run.bash b/scripts/arch-run.bash
index 51e4b97b27d1..72ce718b1170 100644
--- a/scripts/arch-run.bash
+++ b/scripts/arch-run.bash
@@ -94,7 +94,17 @@ run_qemu_status ()
 
 timeout_cmd ()
 {
+	local s
+
 	if [ "$TIMEOUT" ] && [ "$TIMEOUT" != "0" ]; then
+		if [ "$CONFIG_EFI" = 'y' ]; then
+			s=${TIMEOUT: -1}
+			if [ "$s" = 's' ]; then
+				TIMEOUT=${TIMEOUT:0:-1}
+				((TIMEOUT += 10)) # Add 10 seconds for booting UEFI
+				TIMEOUT="${TIMEOUT}s"
+			fi
+		fi
 		echo "timeout -k 1s --foreground $TIMEOUT"
 	fi
 }