On 15/8/26 02:24, [email protected] wrote: >> diff --git a/tools/testing/selftests/bpf/prog_tests/global_data_init.c >> b/tools/testing/selftests/bpf/prog_tests/global_data_init.c >> index 7d6bda909295..2f0dbeb88934 100644 >> --- a/tools/testing/selftests/bpf/prog_tests/global_data_init.c >> +++ b/tools/testing/selftests/bpf/prog_tests/global_data_init.c > > [ ... ] > >> @@ -126,11 +123,11 @@ static void test_percpu_data_on_cpus(struct bpf_map >> *map, int map_fd, int prog_f >> static void test_global_percpu_data_init(void) >> { >> struct test_global_percpu_data__percpu init_value = {}; >> + const __u32 desired_sz = sysconf(_SC_PAGE_SIZE) * 2; >> struct test_global_percpu_data__percpu *init_data; >> - const __u32 desired_sz = sysconf(_SC_PAGE_SIZE); > > Does multiplying by 2 here break the test on 64K page systems? > > Per-CPU array value sizes are capped by the kernel: > > /* percpu map value size is bound by PCPU_MIN_UNIT_SIZE */ > if (percpu && round_up(attr->value_size, 8) > PCPU_MIN_UNIT_SIZE) > return -E2BIG; > > in kernel/bpf/arraymap.c:array_map_alloc_check(), where PCPU_MIN_UNIT_SIZE > is defined as: > > #define PCPU_MIN_UNIT_SIZE PFN_ALIGN(32 << 10) > > which equals max(32768, PAGE_SIZE). > > The old code was safe by construction: desired_sz == PAGE_SIZE <= > PFN_ALIGN(32768) == PCPU_MIN_UNIT_SIZE. Doubling it breaks that: > > PAGE_SIZE 4096 -> limit 32768, request 8192 OK > PAGE_SIZE 16384 -> limit 32768, request 32768 OK (exactly at limit) > PAGE_SIZE 65536 -> limit 65536, request 131072 -E2BIG > > On kernels built with 64K pages (arm64 CONFIG_ARM64_64K_PAGES, ppc64) > bpf_map__set_value_size() still succeeds (it is purely userspace > bookkeeping), all four assertions that follow it pass, and then map > creation fails inside test_global_percpu_data__load(): > >> err = test_global_percpu_data__load(skel); >> if (!ASSERT_OK(err, "test_global_percpu_data__load")) >> goto out; > > The "global_percpu_data/init" subtest fails, and test_percpu_data_on_cpus() > is never reached. That is a new failure introduced by this patch (with > desired_sz == PAGE_SIZE the same configuration created a 65536-byte per-CPU > value, which is exactly at the limit and accepted). > > Note that the neighbouring resize test in global_map_resize.c deliberately > does not scale a per-CPU map by page size for this reason. At line 46 it > uses: > > new_sz = sizeof(skel->data_percpu_arr->percpu_arr[0]) * > libbpf_num_possible_cpus(); > > A size that is independent of PAGE_SIZE would still force the mmap region > to move, which is what the new skel->percpu_arr reassignment and the > arr[0] == -1 check are there to exercise, without hitting the per-CPU > allocator bound. >
To avoid breaking the test on 64K page systems, run the resize subtest only on 4K page systems. >From maintenance view, I think it's better to add a dedicated subtest to test bpf_map__set_value_size() for percpu data. Then, it will be able to skip the subtest when sysconf(_SC_PAGE_SIZE) != 4K. Thanks, Leon
