selftests/bpf: Add write to hashmap to array_map iter test
authorDave Marchevsky <davemarchevsky@fb.com>
Thu, 20 Oct 2022 16:07:21 +0000 (09:07 -0700)
committerAlexei Starovoitov <ast@kernel.org>
Sat, 22 Oct 2022 02:23:34 +0000 (19:23 -0700)
Modify iter prog in existing bpf_iter_bpf_array_map.c, which currently
dumps arraymap key/val, to also do a write of (val, key) into a
newly-added hashmap. Confirm that the write succeeds as expected by
modifying the userspace runner program.

Before a change added in an earlier commit - considering PTR_TO_BUF reg
a valid input to helpers which expect MAP_{KEY,VAL} - the verifier
would've rejected this prog change due to type mismatch. Since using
current iter's key/val to access a separate map is a reasonable usecase,
let's add support for it.

Note that the test prog cannot directly write (val, key) into hashmap
via bpf_map_update_elem when both come from iter context because key is
marked MEM_RDONLY. This is due to bpf_map_update_elem - and other basic
map helpers - taking ARG_PTR_TO_MAP_{KEY,VALUE} w/o MEM_RDONLY type
flag. bpf_map_{lookup,update,delete}_elem don't modify their
input key/val so it should be possible to tag their args READONLY, but
due to the ubiquitous use of these helpers and verifier checks for
type == MAP_VALUE, such a change is nontrivial and seems better to
address in a followup series.

Also fixup some 'goto's in test runner's map checking loop.

Signed-off-by: Dave Marchevsky <davemarchevsky@fb.com>
Acked-by: Andrii Nakryiko <andrii@kernel.org>
Link: https://lore.kernel.org/r/20221020160721.4030492-4-davemarchevsky@fb.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
tools/testing/selftests/bpf/prog_tests/bpf_iter.c
tools/testing/selftests/bpf/progs/bpf_iter_bpf_array_map.c

index c39d40f..6f8ed61 100644 (file)
@@ -941,10 +941,10 @@ static void test_bpf_array_map(void)
 {
        __u64 val, expected_val = 0, res_first_val, first_val = 0;
        DECLARE_LIBBPF_OPTS(bpf_iter_attach_opts, opts);
-       __u32 expected_key = 0, res_first_key;
+       __u32 key, expected_key = 0, res_first_key;
+       int err, i, map_fd, hash_fd, iter_fd;
        struct bpf_iter_bpf_array_map *skel;
        union bpf_iter_link_info linfo;
-       int err, i, map_fd, iter_fd;
        struct bpf_link *link;
        char buf[64] = {};
        int len, start;
@@ -1001,12 +1001,20 @@ static void test_bpf_array_map(void)
        if (!ASSERT_EQ(skel->bss->val_sum, expected_val, "val_sum"))
                goto close_iter;
 
+       hash_fd = bpf_map__fd(skel->maps.hashmap1);
        for (i = 0; i < bpf_map__max_entries(skel->maps.arraymap1); i++) {
                err = bpf_map_lookup_elem(map_fd, &i, &val);
-               if (!ASSERT_OK(err, "map_lookup"))
-                       goto out;
-               if (!ASSERT_EQ(i, val, "invalid_val"))
-                       goto out;
+               if (!ASSERT_OK(err, "map_lookup arraymap1"))
+                       goto close_iter;
+               if (!ASSERT_EQ(i, val, "invalid_val arraymap1"))
+                       goto close_iter;
+
+               val = i + 4;
+               err = bpf_map_lookup_elem(hash_fd, &val, &key);
+               if (!ASSERT_OK(err, "map_lookup hashmap1"))
+                       goto close_iter;
+               if (!ASSERT_EQ(key, val - 4, "invalid_val hashmap1"))
+                       goto close_iter;
        }
 
 close_iter:
index 6286023..c5969ca 100644 (file)
@@ -19,13 +19,20 @@ struct {
        __type(value, __u64);
 } arraymap1 SEC(".maps");
 
+struct {
+       __uint(type, BPF_MAP_TYPE_HASH);
+       __uint(max_entries, 10);
+       __type(key, __u64);
+       __type(value, __u32);
+} hashmap1 SEC(".maps");
+
 __u32 key_sum = 0;
 __u64 val_sum = 0;
 
 SEC("iter/bpf_map_elem")
 int dump_bpf_array_map(struct bpf_iter__bpf_map_elem *ctx)
 {
-       __u32 *key = ctx->key;
+       __u32 *hmap_val, *key = ctx->key;
        __u64 *val = ctx->value;
 
        if (key == (void *)0 || val == (void *)0)
@@ -35,6 +42,18 @@ int dump_bpf_array_map(struct bpf_iter__bpf_map_elem *ctx)
        bpf_seq_write(ctx->meta->seq, val, sizeof(__u64));
        key_sum += *key;
        val_sum += *val;
+
+       /* workaround - It's necessary to do this convoluted (val, key)
+        * write into hashmap1, instead of simply doing
+        *   bpf_map_update_elem(&hashmap1, val, key, BPF_ANY);
+        * because key has MEM_RDONLY flag and bpf_map_update elem expects
+        * types without this flag
+        */
+       bpf_map_update_elem(&hashmap1, val, val, BPF_ANY);
+       hmap_val = bpf_map_lookup_elem(&hashmap1, val);
+       if (hmap_val)
+               *hmap_val = *key;
+
        *val = *key;
        return 0;
 }