Hello world gadget
This is a short getting started guide to write your first gadget. This guide will get you familiar with the key concepts by implementing a simplified version of the "trace open" (opensnoop) tool.
Starting from a template
If you want to create a new repository for your gadget, you can use the gadget-template repository. This is a GitHub tempate repository.
You can also look for examples in gadgets published on Artifact Hub:
Starting from scratch
If you already have a git repository for your project and want to add a gadget to it, you can start from scratch. The rest of this guide assumes you will start from scratch.
The first step is to create an empty folder where the source code of the gadget will be stored:
$ mkdir mygadget
Implementing the eBPF program
The eBPF code contains the source code for the programs that are injected in the kernel to collect
information. Let's create a file called program.bpf.c
and put the following contents in there.
The first thing we need is to include some header files.
// Kernel types definitions
// Check https://blog.aquasec.com/vmlinux.h-ebpf-programs for more details
#include <vmlinux.h>
// eBPF helpers signatures
// Check https://man7.org/linux/man-pages/man7/bpf-helpers.7.html to learn
// more about different available helpers
#include <bpf/bpf_helpers.h>
// Inspektor Gadget buffer
#include <gadget/buffer.h>
// Inspektor Gadget macros
#include <gadget/macros.h>
Then, we have to specify a structure with all the information our gadget will provide. Let's only put the pid for the time being.
struct event {
__u32 pid;
};
Then, create a buffer eBPF map to send events to user space:
// events is the name of the buffer map and 1024 * 256 is its size.
GADGET_TRACER_MAP(events, 1024 * 256);
This macro will automatically create a ring buffer if the kernel supports it. Otherwise, a perf array will be created.
And define a tracer by using the GADGET_TRACER
macro with the following
parameters:
- Tracer's Name:
open
- Buffer Map Name:
events
- Event Structure Name:
event
// Define a tracer
GADGET_TRACER(open, events, event);
After that, we need to define a program that is attached to a hook that provides the information we need, in this case we'll attach to a tracepoint that is called each time the openat() syscall is executed.
This program collects the information to fill the event (only pid for now), and then calls
gadget_submit_buf()
helper to send the event to user space.
SEC("tracepoint/syscalls/sys_enter_openat")
int enter_openat(struct syscall_trace_enter *ctx)
{
struct event *event;
event = gadget_reserve_buf(&events, sizeof(*event));
if (!event)
return 0;
event->pid = bpf_get_current_pid_tgid() >> 32;
gadget_submit_buf(ctx, &events, event, sizeof(*event));
return 0;
}
Finally, it's needed to define the license of the eBPF code.
char LICENSE[] SEC("license") = "GPL";
The full file should look like:
// program.bpf.c
// Kernel types definitions
// Check https://blog.aquasec.com/vmlinux.h-ebpf-programs for more details
#include <vmlinux.h>
// eBPF helpers signatures
// Check https://man7.org/linux/man-pages/man7/bpf-helpers.7.html to learn
// more about different available helpers
#include <bpf/bpf_helpers.h>
// Inspektor Gadget buffer
#include <gadget/buffer.h>
// Inspektor Gadget macros
#include <gadget/macros.h>
struct event {
__u32 pid;
};
// events is the name of the buffer map and 1024 * 256 is its size.
GADGET_TRACER_MAP(events, 1024 * 256);
// Define a tracer
GADGET_TRACER(open, events, event);
SEC("tracepoint/syscalls/sys_enter_openat")
int enter_openat(struct syscall_trace_enter *ctx)
{
struct event *event;
event = gadget_reserve_buf(&events, sizeof(*event));
if (!event)
return 0;
event->pid = bpf_get_current_pid_tgid() >> 32;
gadget_submit_buf(ctx, &events, event, sizeof(*event));
return 0;
}
char LICENSE[] SEC("license") = "GPL";
Check Gadget eBPF API to learn all functions, macros and types that Inspektor Gadget exposes to the eBPF programs.
Building the gadget for the first time
We can now compile our gadget. You don't need to have any build tools installed on the
machine, the image build
by default uses docker to run a
container with all dependencies to compile the code.
$ cd mygadget
$ sudo ig image build -t mygadget:latest .
Successfully built ghcr.io/inspektor-gadget/gadget/mygadget:latest@sha256:dd3f5c357983bb863ef86942e36f4c851933eec4b32ba65ee375acb1c514f628
Take into account that it is possible to customize the build process by defining
a build.yaml
file. Check the Building a Gadget section for
more details.
(Optional) Pushing the gadget image to a container registry
You could push the gadget to a remote container registry. If you're using the same machine for building and running the gadget, this step can be skipped.
$ sudo ig image tag mygadget:latest ghcr.io/my-org/mygadget:latest
Successfully tagged with ghcr.io/my-org/mygadget:latest@sha256:dd3f5c357983bb863ef86942e36f4c851933eec4b32ba65ee375acb1c514f628
$ sudo ig image push ghcr.io/my-org/mygadget:latest
Pushing ghcr.io/my-org/mygadget:latest...
Successfully pushed ghcr.io/my-org/mygadget:latest@sha256:dd3f5c357983bb863ef86942e36f4c851933eec4b32ba65ee375acb1c514f628
For the sake of simplicity this guide doesn't cover signing the gadget image, however, we strongly encourage you to sign them. Please check signing to get more details.
Running the gadget
We're now all set to run our gadget for the first time.
$ sudo ig run mygadget:latest --verify-image=false
PID
1113
1113
1113
1113
1113
1113
1113
1113
1113
1113
1113
1113
1113
1113
1113
1219
220121
Great, our program already shows the PID! Can we improve it further?
Adding more information to the gadget
Let's add some more information to our event, like command and file name.
Add the fields in the event structure.
#define NAME_MAX 255
struct event {
__u32 pid;
char comm[TASK_COMM_LEN];
char filename[NAME_MAX];
};
Now create the logic to fill those fields in the enter_openat
program. Insert them after you have
reserved space for your event structure and before you submit the buffer.
bpf_get_current_comm(event->comm, sizeof(event->comm));
bpf_probe_read_user_str(event->filename, sizeof(event->filename), (const char *)ctx->args[1]);
Build and run the gadget again. Now it provides more information.
$ sudo ig image build -t mygadget:latest .
....
$ sudo ig run mygadget:latest --verify-image=false
PID COMM FILENAME
11305 Chrome_ChildIOT /dev/shm/.org.chromium.…
11305 ThreadPoolForeg /home/mvb/.config/Slack…
11305 Chrome_ChildIOT /dev/shm/.org.chromium.…
11305 ThreadPoolForeg /home/mvb/.config/Slack…
11305 Chrome_ChildIOT /dev/shm/.org.chromium.…
1349 containerd /var/lib/containerd/io.…
1349 containerd /var/lib/containerd/io.…
1349 containerd /var/lib/containerd/io.…
Creating a metadata file
The above formatting is not totally great, the pid column is taking a lot of space while the filename is being trimmed. The metadata file contains extra information about the gadget, among other things, it can be used to specify the format to be used.
An initial version of the metadata file can be created by passing --update-metadata
to the build command:
$ sudo ig image build . -t mygadget --update-metadata
It'll create a gadget.yaml
file:
name: 'TODO: Fill the gadget name'
description: 'TODO: Fill the gadget description'
homepageURL: 'TODO: Fill the gadget homepage URL'
documentationURL: 'TODO: Fill the gadget documentation URL'
sourceURL: 'TODO: Fill the gadget source code URL'
datasources:
open:
fields:
comm:
annotations:
description: 'TODO: Fill field description'
filename:
annotations:
description: 'TODO: Fill field description'
pid:
annotations:
description: 'TODO: Fill field description'
Let's edit the file to customize the output. We define some templates for well-known fields like pid, comm, etc.
name: mygadget
description: Example gadget
homepageURL: http://mygadget.com
documentationURL: https://mygadget.com/docs
sourceURL: https://github.com/my-org/mygadget/
datasources:
open:
fields:
comm:
annotations:
description: Name of the process opening a file
template: comm
filename:
annotations:
description: Path of the file being opened
columns.width: 64
pid:
annotations:
description: PID of the process opening a file
template: pid
Now we can build and run the gadget again
$ sudo ig image build . -t mygadget
...
$ sudo ig run mygadget:latest --verify-image=false
PID COMM FILENAME
1094 systemd-oomd /sys/fs/cgroup/user.slice/user-1001.slice/[email protected]/memor…
1094 systemd-oomd /sys/fs/cgroup/user.slice/user-1001.slice/[email protected]/memor…
1094 systemd-oomd /sys/fs/cgroup/user.slice/user-1001.slice/[email protected]/memor…
1094 systemd-oomd /sys/fs/cgroup/user.slice/user-1001.slice/[email protected]/memor…
1094 systemd-oomd /sys/fs/cgroup/user.slice/user-1001.slice/[email protected]/memor…
1094 systemd-oomd /sys/fs/cgroup/user.slice/user-1001.slice/[email protected]/memor…
1094 systemd-oomd /proc/meminfo
20100 tmux: server /proc/118135/cmdline
1094 systemd-oomd /proc/meminfo
1094 systemd-oomd /proc/meminfo
5803 FSBroker114558 /proc/114558/statm
5803 FSBroker114558 /proc/114558/statm
5803 FSBroker114558 /proc/114558/smaps
20100 tmux: server /proc/118135/cmdline
1094 systemd-oomd /proc/meminfo
118137 ig /proc
Now the output is much better.
Container Enrichment
Our gadget provides information about the process (pid, process name) performing the actions, however there is not any information about the container where that process is running in. In order to provide that information, the gadget needs to provide the mount namespace inode id of the process, and Inspektor Gadget will take care of adding the container information to the event.
Include the following header:
#include <gadget/mntns.h>
Add the following field to the event structure.
struct event {
...
gadget_mntns_id mntns_id;
...
}
and set this field
event->mntns_id = mntns_id;
After adding the gadget_mntns_id
field to the event structure, compiling and running again,
Inspektor Gadget will automatically add the container name column to the output.
$ sudo ig run mygadget:latest --verify-image=false
RUNTIME.CONTAINERNAME MNTNS_ID PID COMM FILENAME
...
mycontainer 4026536168 297810 sh /usr/lib/x86_64-linux-gnu/glibc-hwcaps/x86-…
mycontainer 4026536168 297810 sh /usr/lib/x86_64-linux-gnu/glibc-hwcaps/x86-
...
Event Filtering
The gadget we created provides information about all events happening on the host, however it doesn't allow to filter events by a given container, user, etc. Inspektor Gadget provides a set of helpers files and functions to make this process easy.
Inlcude the filter.h
file:
// Inspektor Gadget filtering helpers
#include <gadget/filter.h>
The gadget_should_discard_data_current
function is provided to determine if a given
event should be traced or not. This function should be called as early as
possible in the program to avoid unnecessary work:
...
if (gadget_should_discard_data_current())
return 0;
event = gadget_reserve_buf(&events, sizeof(*event));
if (!event)
return 0;
...
The gadget now exposes some parameters to filter events:
$ sudo ig run mygadget:latest --verify-image=false -h
Run a gadget
Usage:
ig run [flags]
Flags:
...
-c, --containername string Show only data from containers with that name (default "")
...
--gid uint32 Show only events generated by processes with this gid (default 4294967295)
...
--host Show data from both the host and containers
...
--pid uint32 Show only events generated by processes with this pid (default 0)
...
--tid uint32 Show only events generated by threads with this tid (default 0)
...
--uid uint32 Show only events generated by processes with this uid (default 4294967295)
...
The following command doesn't show any event as there is no container with the specified name:
$ sudo ig run mygadget:latest -c non_existing_container --verify-image=false
RUNTIME.CONTAINERNAME MNTNS_ID PID COMM FILENAME
Updating the gadget
Some times we have to update our gadgets, like adding more fields to the generated event for instance. Let's add the following fields to the event struct:
struct event {
...
__u32 uid;
__u32 gid;
...
}
and add the logic in the eBPF program to fill them:
__u64 uid_gid = bpf_get_current_uid_gid();
event->uid = (__u32)uid_gid;
event->gid = (__u32)(uid_gid >> 32);
Let's build the gadget with the --update-metadata
file, so our new fields are automatically added
to the metadata file. Notice the -v option is used to get debugging messages.
$ sudo ig image build . -t mygadget --update-metadata -v
...
DEBU[0000] Metadata file found, updating it
DEBU[0000] Adding field "mntns_id"
DEBU[0000] Field "pid" already exists, skipping
DEBU[0000] Field "comm" already exists, skipping
DEBU[0000] Field "filename" already exists, skipping
DEBU[0000] Adding field "uid"
DEBU[0000] Adding field "gid"
...
The uid, gid and mntns_id (added in the previous step fields were added to the metadata file:
gid:
annotations:
description: 'TODO: Fill field description'
mntns_id:
annotations:
description: 'TODO: Fill field description'
pid:
annotations:
description: PID of the process opening a file
template: pid
uid:
annotations:
description: 'TODO: Fill field description'
Edit them, build and run the gadget again:
comm:
annotations:
description: Name of the process opening a file
template: comm
filename:
annotations:
description: Path of the file being opened
columns.width: 64
gid:
annotations:
description: Group ID opening the file
template: uid
mntns_id:
annotations:
description: Mount namespace inode id
template: ns
pid:
annotations:
description: PID of the process opening a file
template: pid
uid:
annotations:
description: User ID opening the file
template: uid
$ sudo ig image build . -t mygadget --update-metadata -v
...
$ sudo ig run mygadget:latest --verify-image=false
RUNTIME.CONTAINERN… PID COMM FILENAME UID GID
Now, the UID and GID columns have the expected format. Notice also that the MNTNS_ID column is
not showed because the template ns
hides it by default.
Using helpers to handle common fields
Most of the gadgets have common fields like pid
, comm
, uid
, gid
, etc.
Inspektor Gadget provides some types and common helpers to handle these fields.
Let's refactor the event structure to use them. This is the new full code of the
gadget:
// Kernel types definitions
// Check https://blog.aquasec.com/vmlinux.h-ebpf-programs for more details
#include <vmlinux.h>
// eBPF helpers signatures
// Check https://man7.org/linux/man-pages/man7/bpf-helpers.7.html to learn
// more about different available helpers
#include <bpf/bpf_helpers.h>
// Inspektor Gadget buffer
#include <gadget/buffer.h>
// Helpers to handle common data
#include <gadget/common.h>
// Inspektor Gadget macros
#include <gadget/macros.h>
// Inspektor Gadget filtering
#include <gadget/mntns_filter.h>
// Inspektor Gadget types
#include <gadget/types.h>
#define NAME_MAX 255
struct event {
struct gadget_process proc;
char filename[NAME_MAX];
};
// events is the name of the buffer map and 1024 * 256 is its size.
GADGET_TRACER_MAP(events, 1024 * 256);
// [Optional] Define a tracer
GADGET_TRACER(open, events, event);
SEC("tracepoint/syscalls/sys_enter_openat")
int enter_openat(struct syscall_trace_enter *ctx)
{
struct event *event;
if (gadget_should_discard_data_current())
return 0;
event = gadget_reserve_buf(&events, sizeof(*event));
if (!event)
return 0;
gadget_process_populate(&event->proc);
bpf_probe_read_user_str(event->filename, sizeof(event->filename),
(const char *)ctx->args[1]);
gadget_submit_buf(ctx, &events, event, sizeof(*event));
return 0;
}
char LICENSE[] SEC("license") = "GPL";
And the metadata can be simplified too, as Inspektor Gadget automatically provides field annotations for well-known types:
name: mygadget
description: Example gadget
homepageURL: http://mygadget.com
documentationURL: https://mygadget.com/docs
sourceURL: https://github.com/my-org/mygadget/
datasources:
open:
fields:
filename:
annotations:
description: Path of the file being opened
columns.width: 64
When running the gadget, the comm
, pid
and tid
columns are visible by default:
$ sudo ig run mygadget:latest --verify-image=false
WARN[0001] image signature verification is disabled due to using corresponding option
WARN[0001] image signature verification is disabled due to using corresponding option
RUNTIME.CONTAINERNAME COMM PID TID FILENAME
...
mycontainer cat 51020 51020 /dev/null
And the json output contains all the process information:
{
"filename": "/dev/null",
"k8s": {
"containerName": "",
"hostnetwork": false,
"namespace": "",
"node": "",
"owner": {
"kind": "",
"name": ""
},
"podName": ""
},
"proc": {
"comm": "cat",
"mntns_id": 4026534688,
"parent": {
"comm": "sh",
"pid": 50991
},
"pid": 51092,
"tid": 51092,
"creds": {
"gid": 0,
"group": "root",
"uid": 0,
"user": "root"
}
},
"runtime": {
"containerId": "ff4df9bf30a89c0cc8bd32d1c976875bf5bd8d7c745d09b98358b953ef4d8666",
"containerImageDigest": "sha256:768e5c6f5cb6db0794eec98dc7a967f40631746c32232b78a3105fb946f3ab83",
"containerImageName": "docker.io/library/busybox:latest",
"containerName": "mycontainer",
"runtimeName": "docker"
}
}
Closing
Congratulations! You've implemented your first gadget. Check out our documentation to get more information.