selftests/bpf: Cope with 512 bytes limit with bpf_global_percpu_ma
authorYonghong Song <yonghong.song@linux.dev>
Fri, 22 Dec 2023 03:18:07 +0000 (19:18 -0800)
committerAlexei Starovoitov <ast@kernel.org>
Thu, 4 Jan 2024 05:08:26 +0000 (21:08 -0800)
In the previous patch, the maximum data size for bpf_global_percpu_ma
is 512 bytes. This breaks selftest test_bpf_ma. The test is adjusted
in two aspects:
  - Since the maximum allowed data size for bpf_global_percpu_ma is
    512, remove all tests beyond that, names sizes 1024, 2048 and 4096.
  - Previously the percpu data size is bucket_size - 8 in order to
    avoid percpu allocation into the next bucket. This patch removed
    such data size adjustment thanks to Patch 1.

Also, a better way to generate BTF type is used than adding
a member to the value struct.

Acked-by: Hou Tao <houtao1@huawei.com>
Signed-off-by: Yonghong Song <yonghong.song@linux.dev>
Link: https://lore.kernel.org/r/20231222031807.1292853-1-yonghong.song@linux.dev
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
tools/testing/selftests/bpf/prog_tests/test_bpf_ma.c
tools/testing/selftests/bpf/progs/test_bpf_ma.c

index d3491a84b3b98322a9bc0b135b361e923d28eb81..ccae0b31ac6c6d2a77485282bd131be2cb7383a9 100644 (file)
@@ -14,7 +14,8 @@ static void do_bpf_ma_test(const char *name)
        struct test_bpf_ma *skel;
        struct bpf_program *prog;
        struct btf *btf;
-       int i, err;
+       int i, err, id;
+       char tname[32];
 
        skel = test_bpf_ma__open();
        if (!ASSERT_OK_PTR(skel, "open"))
@@ -25,16 +26,21 @@ static void do_bpf_ma_test(const char *name)
                goto out;
 
        for (i = 0; i < ARRAY_SIZE(skel->rodata->data_sizes); i++) {
-               char name[32];
-               int id;
-
-               snprintf(name, sizeof(name), "bin_data_%u", skel->rodata->data_sizes[i]);
-               id = btf__find_by_name_kind(btf, name, BTF_KIND_STRUCT);
-               if (!ASSERT_GT(id, 0, "bin_data"))
+               snprintf(tname, sizeof(tname), "bin_data_%u", skel->rodata->data_sizes[i]);
+               id = btf__find_by_name_kind(btf, tname, BTF_KIND_STRUCT);
+               if (!ASSERT_GT(id, 0, tname))
                        goto out;
                skel->rodata->data_btf_ids[i] = id;
        }
 
+       for (i = 0; i < ARRAY_SIZE(skel->rodata->percpu_data_sizes); i++) {
+               snprintf(tname, sizeof(tname), "percpu_bin_data_%u", skel->rodata->percpu_data_sizes[i]);
+               id = btf__find_by_name_kind(btf, tname, BTF_KIND_STRUCT);
+               if (!ASSERT_GT(id, 0, tname))
+                       goto out;
+               skel->rodata->percpu_data_btf_ids[i] = id;
+       }
+
        prog = bpf_object__find_program_by_name(skel->obj, name);
        if (!ASSERT_OK_PTR(prog, "invalid prog name"))
                goto out;
index b78f4f702ae0f11099a078533240b3459524c704..3494ca30fa7fe7ecd6c0b7c78b66eb72ff0425cd 100644 (file)
@@ -20,6 +20,9 @@ char _license[] SEC("license") = "GPL";
 const unsigned int data_sizes[] = {16, 32, 64, 96, 128, 192, 256, 512, 1024, 2048, 4096};
 const volatile unsigned int data_btf_ids[ARRAY_SIZE(data_sizes)] = {};
 
+const unsigned int percpu_data_sizes[] = {8, 16, 32, 64, 96, 128, 192, 256, 512};
+const volatile unsigned int percpu_data_btf_ids[ARRAY_SIZE(data_sizes)] = {};
+
 int err = 0;
 u32 pid = 0;
 
@@ -27,10 +30,10 @@ u32 pid = 0;
        struct bin_data_##_size { \
                char data[_size - sizeof(void *)]; \
        }; \
+       /* See Commit 5d8d6634ccc, force btf generation for type bin_data_##_size */    \
+       struct bin_data_##_size *__bin_data_##_size; \
        struct map_value_##_size { \
                struct bin_data_##_size __kptr * data; \
-               /* To emit BTF info for bin_data_xx */ \
-               struct bin_data_##_size not_used; \
        }; \
        struct { \
                __uint(type, BPF_MAP_TYPE_ARRAY); \
@@ -40,8 +43,12 @@ u32 pid = 0;
        } array_##_size SEC(".maps")
 
 #define DEFINE_ARRAY_WITH_PERCPU_KPTR(_size) \
+       struct percpu_bin_data_##_size { \
+               char data[_size]; \
+       }; \
+       struct percpu_bin_data_##_size *__percpu_bin_data_##_size; \
        struct map_value_percpu_##_size { \
-               struct bin_data_##_size __percpu_kptr * data; \
+               struct percpu_bin_data_##_size __percpu_kptr * data; \
        }; \
        struct { \
                __uint(type, BPF_MAP_TYPE_ARRAY); \
@@ -114,7 +121,7 @@ static __always_inline void batch_percpu_alloc(struct bpf_map *map, unsigned int
                        return;
                }
                /* per-cpu allocator may not be able to refill in time */
-               new = bpf_percpu_obj_new_impl(data_btf_ids[idx], NULL);
+               new = bpf_percpu_obj_new_impl(percpu_data_btf_ids[idx], NULL);
                if (!new)
                        continue;
 
@@ -179,7 +186,7 @@ DEFINE_ARRAY_WITH_KPTR(1024);
 DEFINE_ARRAY_WITH_KPTR(2048);
 DEFINE_ARRAY_WITH_KPTR(4096);
 
-/* per-cpu kptr doesn't support bin_data_8 which is a zero-sized array */
+DEFINE_ARRAY_WITH_PERCPU_KPTR(8);
 DEFINE_ARRAY_WITH_PERCPU_KPTR(16);
 DEFINE_ARRAY_WITH_PERCPU_KPTR(32);
 DEFINE_ARRAY_WITH_PERCPU_KPTR(64);
@@ -188,9 +195,6 @@ DEFINE_ARRAY_WITH_PERCPU_KPTR(128);
 DEFINE_ARRAY_WITH_PERCPU_KPTR(192);
 DEFINE_ARRAY_WITH_PERCPU_KPTR(256);
 DEFINE_ARRAY_WITH_PERCPU_KPTR(512);
-DEFINE_ARRAY_WITH_PERCPU_KPTR(1024);
-DEFINE_ARRAY_WITH_PERCPU_KPTR(2048);
-DEFINE_ARRAY_WITH_PERCPU_KPTR(4096);
 
 SEC("?fentry/" SYS_PREFIX "sys_nanosleep")
 int test_batch_alloc_free(void *ctx)
@@ -246,20 +250,18 @@ int test_batch_percpu_alloc_free(void *ctx)
        if ((u32)bpf_get_current_pid_tgid() != pid)
                return 0;
 
-       /* Alloc 128 16-bytes per-cpu objects in batch to trigger refilling,
-        * then free 128 16-bytes per-cpu objects in batch to trigger freeing.
+       /* Alloc 128 8-bytes per-cpu objects in batch to trigger refilling,
+        * then free 128 8-bytes per-cpu objects in batch to trigger freeing.
         */
-       CALL_BATCH_PERCPU_ALLOC_FREE(16, 128, 0);
-       CALL_BATCH_PERCPU_ALLOC_FREE(32, 128, 1);
-       CALL_BATCH_PERCPU_ALLOC_FREE(64, 128, 2);
-       CALL_BATCH_PERCPU_ALLOC_FREE(96, 128, 3);
-       CALL_BATCH_PERCPU_ALLOC_FREE(128, 128, 4);
-       CALL_BATCH_PERCPU_ALLOC_FREE(192, 128, 5);
-       CALL_BATCH_PERCPU_ALLOC_FREE(256, 128, 6);
-       CALL_BATCH_PERCPU_ALLOC_FREE(512, 64, 7);
-       CALL_BATCH_PERCPU_ALLOC_FREE(1024, 32, 8);
-       CALL_BATCH_PERCPU_ALLOC_FREE(2048, 16, 9);
-       CALL_BATCH_PERCPU_ALLOC_FREE(4096, 8, 10);
+       CALL_BATCH_PERCPU_ALLOC_FREE(8, 128, 0);
+       CALL_BATCH_PERCPU_ALLOC_FREE(16, 128, 1);
+       CALL_BATCH_PERCPU_ALLOC_FREE(32, 128, 2);
+       CALL_BATCH_PERCPU_ALLOC_FREE(64, 128, 3);
+       CALL_BATCH_PERCPU_ALLOC_FREE(96, 128, 4);
+       CALL_BATCH_PERCPU_ALLOC_FREE(128, 128, 5);
+       CALL_BATCH_PERCPU_ALLOC_FREE(192, 128, 6);
+       CALL_BATCH_PERCPU_ALLOC_FREE(256, 128, 7);
+       CALL_BATCH_PERCPU_ALLOC_FREE(512, 64, 8);
 
        return 0;
 }
@@ -270,20 +272,18 @@ int test_percpu_free_through_map_free(void *ctx)
        if ((u32)bpf_get_current_pid_tgid() != pid)
                return 0;
 
-       /* Alloc 128 16-bytes per-cpu objects in batch to trigger refilling,
+       /* Alloc 128 8-bytes per-cpu objects in batch to trigger refilling,
         * then free these object through map free.
         */
-       CALL_BATCH_PERCPU_ALLOC(16, 128, 0);
-       CALL_BATCH_PERCPU_ALLOC(32, 128, 1);
-       CALL_BATCH_PERCPU_ALLOC(64, 128, 2);
-       CALL_BATCH_PERCPU_ALLOC(96, 128, 3);
-       CALL_BATCH_PERCPU_ALLOC(128, 128, 4);
-       CALL_BATCH_PERCPU_ALLOC(192, 128, 5);
-       CALL_BATCH_PERCPU_ALLOC(256, 128, 6);
-       CALL_BATCH_PERCPU_ALLOC(512, 64, 7);
-       CALL_BATCH_PERCPU_ALLOC(1024, 32, 8);
-       CALL_BATCH_PERCPU_ALLOC(2048, 16, 9);
-       CALL_BATCH_PERCPU_ALLOC(4096, 8, 10);
+       CALL_BATCH_PERCPU_ALLOC(8, 128, 0);
+       CALL_BATCH_PERCPU_ALLOC(16, 128, 1);
+       CALL_BATCH_PERCPU_ALLOC(32, 128, 2);
+       CALL_BATCH_PERCPU_ALLOC(64, 128, 3);
+       CALL_BATCH_PERCPU_ALLOC(96, 128, 4);
+       CALL_BATCH_PERCPU_ALLOC(128, 128, 5);
+       CALL_BATCH_PERCPU_ALLOC(192, 128, 6);
+       CALL_BATCH_PERCPU_ALLOC(256, 128, 7);
+       CALL_BATCH_PERCPU_ALLOC(512, 64, 8);
 
        return 0;
 }