--- /dev/null
+// SPDX-License-Identifier: GPL-2.0
+/* Copyright (c) 2021 Facebook */
+#include <test_progs.h>
+#include <network_helpers.h>
+#include "for_each_hash_map_elem.skel.h"
+
+static unsigned int duration;
+
+static void test_hash_map(void)
+{
+       int i, err, hashmap_fd, max_entries, percpu_map_fd;
+       struct for_each_hash_map_elem *skel;
+       __u64 *percpu_valbuf = NULL;
+       __u32 key, num_cpus, retval;
+       __u64 val;
+
+       skel = for_each_hash_map_elem__open_and_load();
+       if (!ASSERT_OK_PTR(skel, "for_each_hash_map_elem__open_and_load"))
+               return;
+
+       hashmap_fd = bpf_map__fd(skel->maps.hashmap);
+       max_entries = bpf_map__max_entries(skel->maps.hashmap);
+       for (i = 0; i < max_entries; i++) {
+               key = i;
+               val = i + 1;
+               err = bpf_map_update_elem(hashmap_fd, &key, &val, BPF_ANY);
+               if (!ASSERT_OK(err, "map_update"))
+                       goto out;
+       }
+
+       num_cpus = bpf_num_possible_cpus();
+       percpu_map_fd = bpf_map__fd(skel->maps.percpu_map);
+       percpu_valbuf = malloc(sizeof(__u64) * num_cpus);
+       if (!ASSERT_OK_PTR(percpu_valbuf, "percpu_valbuf"))
+               goto out;
+
+       key = 1;
+       for (i = 0; i < num_cpus; i++)
+               percpu_valbuf[i] = i + 1;
+       err = bpf_map_update_elem(percpu_map_fd, &key, percpu_valbuf, BPF_ANY);
+       if (!ASSERT_OK(err, "percpu_map_update"))
+               goto out;
+
+       err = bpf_prog_test_run(bpf_program__fd(skel->progs.test_pkt_access),
+                               1, &pkt_v4, sizeof(pkt_v4), NULL, NULL,
+                               &retval, &duration);
+       if (CHECK(err || retval, "ipv4", "err %d errno %d retval %d\n",
+                 err, errno, retval))
+               goto out;
+
+       ASSERT_EQ(skel->bss->hashmap_output, 4, "hashmap_output");
+       ASSERT_EQ(skel->bss->hashmap_elems, max_entries, "hashmap_elems");
+
+       key = 1;
+       err = bpf_map_lookup_elem(hashmap_fd, &key, &val);
+       ASSERT_ERR(err, "hashmap_lookup");
+
+       ASSERT_EQ(skel->bss->percpu_called, 1, "percpu_called");
+       ASSERT_LT(skel->bss->cpu, num_cpus, "num_cpus");
+       ASSERT_EQ(skel->bss->percpu_map_elems, 1, "percpu_map_elems");
+       ASSERT_EQ(skel->bss->percpu_key, 1, "percpu_key");
+       ASSERT_EQ(skel->bss->percpu_val, skel->bss->cpu + 1, "percpu_val");
+       ASSERT_EQ(skel->bss->percpu_output, 100, "percpu_output");
+out:
+       free(percpu_valbuf);
+       for_each_hash_map_elem__destroy(skel);
+}
+
+void test_for_each(void)
+{
+       if (test__start_subtest("hash_map"))
+               test_hash_map();
+}
 
--- /dev/null
+// SPDX-License-Identifier: GPL-2.0
+/* Copyright (c) 2021 Facebook */
+#include "vmlinux.h"
+#include <bpf/bpf_helpers.h>
+
+char _license[] SEC("license") = "GPL";
+
+struct {
+       __uint(type, BPF_MAP_TYPE_HASH);
+       __uint(max_entries, 3);
+       __type(key, __u32);
+       __type(value, __u64);
+} hashmap SEC(".maps");
+
+struct {
+       __uint(type, BPF_MAP_TYPE_PERCPU_HASH);
+       __uint(max_entries, 1);
+       __type(key, __u32);
+       __type(value, __u64);
+} percpu_map SEC(".maps");
+
+struct callback_ctx {
+       struct __sk_buff *ctx;
+       int input;
+       int output;
+};
+
+static __u64
+check_hash_elem(struct bpf_map *map, __u32 *key, __u64 *val,
+               struct callback_ctx *data)
+{
+       struct __sk_buff *skb = data->ctx;
+       __u32 k;
+       __u64 v;
+
+       if (skb) {
+               k = *key;
+               v = *val;
+               if (skb->len == 10000 && k == 10 && v == 10)
+                       data->output = 3; /* impossible path */
+               else
+                       data->output = 4;
+       } else {
+               data->output = data->input;
+               bpf_map_delete_elem(map, key);
+       }
+
+       return 0;
+}
+
+__u32 cpu = 0;
+__u32 percpu_called = 0;
+__u32 percpu_key = 0;
+__u64 percpu_val = 0;
+int percpu_output = 0;
+
+static __u64
+check_percpu_elem(struct bpf_map *map, __u32 *key, __u64 *val,
+                 struct callback_ctx *unused)
+{
+       struct callback_ctx data;
+
+       percpu_called++;
+       cpu = bpf_get_smp_processor_id();
+       percpu_key = *key;
+       percpu_val = *val;
+
+       data.ctx = 0;
+       data.input = 100;
+       data.output = 0;
+       bpf_for_each_map_elem(&hashmap, check_hash_elem, &data, 0);
+       percpu_output = data.output;
+
+       return 0;
+}
+
+int hashmap_output = 0;
+int hashmap_elems = 0;
+int percpu_map_elems = 0;
+
+SEC("classifier")
+int test_pkt_access(struct __sk_buff *skb)
+{
+       struct callback_ctx data;
+
+       data.ctx = skb;
+       data.input = 10;
+       data.output = 0;
+       hashmap_elems = bpf_for_each_map_elem(&hashmap, check_hash_elem, &data, 0);
+       hashmap_output = data.output;
+
+       percpu_map_elems = bpf_for_each_map_elem(&percpu_map, check_percpu_elem,
+                                                (void *)0, 0);
+       return 0;
+}
 
        ___ok;                                                          \
 })
 
+#define ASSERT_LT(actual, expected, name) ({                           \
+       static int duration = 0;                                        \
+       typeof(actual) ___act = (actual);                               \
+       typeof(expected) ___exp = (expected);                           \
+       bool ___ok = ___act < ___exp;                                   \
+       CHECK(!___ok, (name),                                           \
+             "unexpected %s: actual %lld >= expected %lld\n",          \
+             (name), (long long)(___act), (long long)(___exp));        \
+       ___ok;                                                          \
+})
+
 #define ASSERT_STREQ(actual, expected, name) ({                                \
        static int duration = 0;                                        \
        const char *___act = actual;                                    \