Use Map-in-Map in BPF programs via Libbpf
Introduction
Among all BPF map types, two special ones, BPF_MAP_TYPE_ARRAY_OF_MAPS
and BPF_MAP_TYPE_HASH_OF_MAPS
, are more complex than others. As the names imply, they are “map-in-map”, meaning that the value of each entry is also a map.
Due to this indirectness, CRUD (Creat, Read, Update, Delete) operations on a map-in-map are different from those of a regular single-level map. In this post, I will use a simple example to illustrate the basic usage of a map-in-map and some caveats.
Note: Since the only difference between BPF_MAP_TYPE_ARRAY_OF_MAPS
and BPF_MAP_TYPE_HASH_OF_MAPS
is whether the outer map is an array or a hash map, I will only use BPF_MAP_TYPE_ARRAY_OF_MAPS
as a representative. Also, error handling is ignored for simplicity.
Create
During load time, all regular maps defined in a BPF C program are created. However, a map-in-map is different in that you only have to define the outer map in the source code. It makes sense since inner maps are created and inserted into the outer map at runtime by the user. The map meta data is defined in struct bpf_map_def
, same as all other regular map types. For example,
1
2
3
4
5
6
struct bpf_map_def SEC("maps") outer_map = {
.type = BPF_MAP_TYPE_HASH_OF_MAPS,
.key_size = sizeof(__u32),
.value_size = sizeof(__u32), // Must be u32 becuase it is inner map id
.max_entries = 1,
};
Note:
- The value size of an outer map must be
__u32
, which is the size of inner map id (See Lookup section below for more details). struct bpf_map_def
is a helper structure used by BPF C program to describe map attributes to theelf_bpf
loader. You can define it yourself or you can use the one defined inbpf_helpers.h
.
Though you don’t have to define the inner map in BPF C program, the verifier needs to know the inner map definition during load time. Therefore, before calling bpf_object__load
, you must create a dummy inner map and set its file descriptor (fd) to the outer map by calling bpf_map__set_inner_map_fd
. The dummy map fd must be closed after load since its only used to make the verifier happy.
1
2
3
4
5
6
7
8
9
10
11
const char* outer_map_name = "outer_map";
struct bpf_map* outer_map = bpf_object__find_map_by_name(obj, outer_map_name);
int inner_map_fd = bpf_create_map(
BPF_MAP_TYPE_HASH, // type
sizeof(__u32), // key_size
sizeof(__u32), // value_size
8, // max_entries
0); // flag
bpf_map__set_inner_map_fd(outer_map, inner_map_fd);
bpf_object__load(obj);
close(inner_map_fd); // Important!
Insert
Insert into Outer Map
There are 3 steps to insert an entry into the outer map:
- Create a new inner map.
- Insert an entry to the outer map with the the inner map fd as the value.
- Close the inner map fd.
1
2
3
4
5
6
7
8
9
10
int inner_map_fd = bpf_create_map_name(
BPF_MAP_TYPE_HASH, // type
"hechaol_inner_map", // name
sizeof(__u32), // key_size
sizeof(__u32), // value_size
8, // max_entries
0); // flag
__u32 outer_key = 42;
bpf_map_update_elem(outer_map_fd, &outer_key, &inner_map_fd, 0 /* flag */);
close(inner_map_fd); // Important!
Note:
- The value of each entry in an outer map is the id of an inner map (See Lookup section below). However, when you call
bpf_map_update_elem
API, the value you give is the fd of the inner map. - You have to close the inner map fd after insertion to avoid memory leak. Because after map creation, you have a reference (fd) to the map in user space. After insertion, kernel has a reference to the map in kernel space. If you don’t close the map fd now, then after the entry is deleted from the outer map (kernel releases its reference), the inner map resource won’t be cleaned up since you still have the reference even though you may have lost the fd after your insertion function returns.
Inert into Inner Map
As we mentioned above, the value of an entry in the outer map is the id (NOT fd) of the inner map. Though we use inner map fd as the value when calling bpf_map_update_elem
, we get back the inner map id when calling bpf_map_lookup_elem
. To get the map fd, we need to call bpf_map_get_fd_by_id
.
1
2
3
4
5
6
7
8
9
const __u32 outer_key = 42;
__u32 inner_map_id;
bpf_map_lookup_elem(outer_map_fd, &outer_key, &inner_map_id);
int inner_map_fd = bpf_map_get_fd_by_id(inner_map_id);
const __u32 inner_key = 12;
__u32 inner_value;
bpf_map_lookup_elem(inner_map_fd, &inner_key, &inner_value);
// ... Use inner_value;
close(inner_map_fd); // Important!
Note:
- Each call to
bpf_map_get_fd_by_id
returns a new fd. You have to close it after use to avoid memory leak.
Delete
Delete operation is as simple as other regular maps.
1
2
const __u32 outer_key = 42;
bpf_map_delete_elem(outer_map_fd, &outer_key);
All inner map resources should be released after bpf_map_delete_elem
as long as you have avoided all the memory leak risks mentioned above. I would highly recommend you do some stressful tests including bulk insertion/deletion to make sure there is no memory leak, which w had in our project once.
Code
Complete code can be found on github.
Comments powered by Disqus.