Re: Kunit test failures with cpumask tests on parisc

From: Helge Deller
Date: Mon Feb 05 2024 - 04:29:02 EST


On 2/5/24 07:53, Guenter Roeck wrote:
when running cpumask Kunit tests on parisc/parisc64 in qemu,
I get the following errors.

KTAP version 1
# Subtest: cpumask
# module: cpumask_kunit
1..6
# test_cpumask_weight: EXPECTATION FAILED at lib/cpumask_kunit.c:68
Expected nr_cpu_ids == cpumask_weight(((const struct cpumask *)&__cpu_possible_mask)), but
nr_cpu_ids == 16 (0x10)
cpumask_weight(((const struct cpumask *)&__cpu_possible_mask)) == 1 (0x1)
cpu_possible_mask contains CPUs 0
not ok 1 test_cpumask_weight
# test_cpumask_first: EXPECTATION FAILED at lib/cpumask_kunit.c:79
Expected nr_cpu_ids <= cpumask_first_zero(((const struct cpumask *)&__cpu_possible_mask)), but
nr_cpu_ids == 16 (0x10)
cpumask_first_zero(((const struct cpumask *)&__cpu_possible_mask)) == 1 (0x1)
cpu_possible_mask contains CPUs 0
not ok 2 test_cpumask_first
# test_cpumask_last: EXPECTATION FAILED at lib/cpumask_kunit.c:87
Expected nr_cpu_ids - 1 == cpumask_last(((const struct cpumask *)&__cpu_possible_mask)), but
nr_cpu_ids - 1 == 15 (0xf)
cpumask_last(((const struct cpumask *)&__cpu_possible_mask)) == 0 (0x0)
cpu_possible_mask contains CPUs 0
not ok 3 test_cpumask_last
# test_cpumask_next: EXPECTATION FAILED at lib/cpumask_kunit.c:94
Expected nr_cpu_ids <= cpumask_next_zero(-1, ((const struct cpumask *)&__cpu_possible_mask)), but
nr_cpu_ids == 16 (0x10)
cpumask_next_zero(-1, ((const struct cpumask *)&__cpu_possible_mask)) == 1 (0x1)
cpu_possible_mask contains CPUs 0
not ok 4 test_cpumask_next
ok 5 test_cpumask_iterators
ok 6 test_cpumask_iterators_builtin
# cpumask: pass:2 fail:4 skip:0 total:6
# Totals: pass:2 fail:4 skip:0 total:6
not ok 5 cpumask

It appears that parisc sets __cpu_possible_mask to the number of online CPUs,
which is limited in qemu and doesn't match CONFIG_NR_CPUS. Is this a problem
with the unit test or with the parisc architecture, or does the unit test
simply not apply for parisc ?

Thank you for finding and reporting this!
It's a bug (or a misunderstanding) in the parisc kernel.
Reverting commit 0921244f6f4f ("parisc: Only list existing CPUs in cpu_possible_mask")
fixes the KUnit test.

Furthermore the revert fixes the issue that CPU hot-unplugging doesn't
work and which I just was starting to debug:
https://lore.kernel.org/lkml/Zb0mbHlIud_bqftx@xxxxxxxxxxxxxxx/t/

Helge