On 9/28/24 12:27, Shuah Khan wrote:
On 9/28/24 11:54, Shuah Khan wrote:
On 9/28/24 11:31, Guenter Roeck wrote:
On 9/27/24 17:08, Guenter Roeck wrote:
On 9/27/24 13:45, Shuah Khan wrote:
On 9/27/24 10:19, Guenter Roeck wrote:
Copying devicetree maintainers.
On Thu, Sep 26, 2024 at 09:39:38PM -0700, Guenter Roeck wrote: > On Thu, Sep 26, 2024 at 09:14:11PM -0700, Guenter Roeck wrote: >> Hi Stephen, >> >> On Thu, Jul 18, 2024 at 02:05:07PM -0700, Stephen Boyd wrote: >>> Test that clks registered with 'struct clk_parent_data' work as >>> intended and can find their parents. >>> >> >> When testing this on arm64, I see the error below. The error is only >> seen if I boot through efi, i.e., with "-bios QEMU_EFI-aarch64.fd" >> qemu parameter. >> >> Any idea what might cause the problem ? >> > I noticed that the new overlay tests fail as well, also with "path '/' not > found". > > [Maybe] answering my own question: I think the problem may be that there > is no devicetree file and thus no devicetree root when booting through > efi (in other words, of_root is NULL). Would it make sense to skip the > tests in that case ? >
The problem is that of_root is not initialized in arm64 boots if ACPI is enabled.
From arch/arm64/kernel/setup.c:setup_arch():
if (acpi_disabled) unflatten_device_tree(); // initializes of_root
ACPI is enabled if the system boots from EFI. This also affects CONFIG_OF_KUNIT_TEST, which explicitly checks if of_root exists and fails the test if it doesn't.
I think those tests need to add a check for this condition, or affected machines won't be able to run those unit tests. The obvious solution would be to check if of_root is set, but then the associated test case in CONFIG_OF_KUNIT_TEST would not make sense.
Any suggestions ?
Would it work if these tests check if acpi_disabled and skip if it isn't disabled? It might be low overhead condition to check from these tests.
acpi_disabled is exported:
arch/arm64/kernel/acpi.c:EXPORT_SYMBOL(acpi_disabled); arch/loongarch/kernel/acpi.c:EXPORT_SYMBOL(acpi_disabled); arch/riscv/kernel/acpi.c:EXPORT_SYMBOL(acpi_disabled); arch/x86/kernel/acpi/boot.c:EXPORT_SYMBOL(acpi_disabled);
I don't think that would work. Looking through the use of acpi_init, I don't think that of_root is always NULL when acpi_init is false; that just happens to be the case on arm64 when booting through efi. However, even arm64 has the following code.
if (acpi_disabled) psci_dt_init(); else psci_acpi_init();
While psci_dt_init() doesn't set of_root, it does try to do a devicetree match. So there must be some other condition where acpi_disabled is set but of_root is set anyway. I just have not found that code path.
I ended up disabling all affected unit tests for arm64. I'll do the same for other architectures if I encounter the problem there as well.
Unfortunately that includes all clock unit tests because the tests requiring devicetree support can not be enabled/disabled separately, but that can't be helped and is still better than "mandatory" failures.
of_root is set in drivers/of/pdt.c when it creates the root node. This could be a definitive test for kunit tests that depend on devicetree support.
That is not always the case, including arm64. It is primarily set in unflatten_devicetree(), which is not called on arm64 unless acpi_is disabled (see above).
It is an exported symbol. drivers/of/base.c exports it.
Yes, checking if of_root is NULL and skipping the test in that case might help, but then there is the of_dtb_root_node_populates_of_root unit test which explicitly fails if of_root is NULL. The comment describing the test is
/* * Test that the 'of_root' global variable is always populated when DT code is * enabled. Remove this test once of_root is removed from global access. */
The devicetree unit test code explicitly assumes that of_root is set if CONFIG_OF_EARLY_FLATTREE is enabled, but that is not always the case (again, of_root is NULL on arm64 unless acpi is disabled).
Guenter