dev-agent

README

dev-agent

This agent provides access to system's eBPF-programs and maps to perform remote debugging.

Features:

• GraphQL model for maps and programs
• Prometheus metrics exporter (including values of maps entries)

See also CHANGELOG.

Usage

Server

sudo ./phydev server [--help]

GraphQL interface: http://localhost:8080/

Prometheus endpoint: http://localhost:8080/metrics

Schema: pkg/graph/schema.graphqls

Prometheus endpoint

Metrics scrape endpoint for Prometheus: http://localhost:8080/metrics

• program metrics:
  • devagent_ebpf_prog_count - number of eBPF programs by type
  • runtime metrics only available with sysctl -w kernel.bpf_stats_enabled=1:
    • devagent_ebpf_prog_run_count - number of times an eBPF program has been run (by id, name, tag, type)
    • devagent_ebpf_prog_run_time - total time spent running eBPF programs (by id, name, tag, type)
• map metrics:
  • devagent_ebpf_map_count - number of eBPF maps by type
  • if map export is configured (see below):
    • devagent_ebpf_map_entry_count - number of entries in an eBPF map (by id, name, type)
    • devagent_ebpf_map_entry_value - value of an eBPF map entry (by key, cpu, id, name, type)

You can find example of Grafana dashboard in grafana-ebpf-dashboard.json:

Configuring map export

As an example, I'm running this bpftrace program:

sudo bpftrace -e 'tracepoint:raw_syscalls:sys_enter { @SYSCALLNUM[comm] = count(); }'

You could see the name of created map - AT_SYSCALLNUM, and the map content in ebpf-explorer:

By default, dev-agent doesn't export map entries to Prometheus, as it may introduce some performance issues.

Instead, you could set an option --etm -:AT_SYSCALLNUM:string when running server, which will suggest agent which map entries to expose in /metrics.

For this HASH_PER_CPU map, it will export 2 metrics:

# HELP devagent_ebpf_map_entry_count Number of entries in an eBPF map
# TYPE devagent_ebpf_map_entry_count gauge
devagent_ebpf_map_entry_count{id="25",name="AT_SYSCALLNUM",type="PerCPUHash"} 764
# HELP devagent_ebpf_map_entry_value Value of an eBPF map entry
# TYPE devagent_ebpf_map_entry_value gauge
devagent_ebpf_map_entry_value{cpu="0",id="25",key="(anacron)",name="AT_SYSCALLNUM",type="PerCPUHash"} 0
devagent_ebpf_map_entry_value{cpu="0",id="25",key="(fprintd)",name="AT_SYSCALLNUM",type="PerCPUHash"} 0

This is how it may look in Grafana (top 10 processes doing most of syscalls):

Full demo in terminal:

sudo bpftrace -e 'tracepoint:raw_syscalls:sys_enter { @SYSCALLNUM[comm] = count(); }' &
docker run -ti --rm --cap-add CAP_SYS_ADMIN --pid=host -v /sys/fs/bpf:/sys/fs/bpf --pid=host -p 8080:8080 ghcr.io/ebpfdev/dev-agent:v0.0.4 server --etm -:AT_SYSCALLNUM:string
curl http://localhost:8080/metrics | grep devagent_ebpf_map_entry_count

Run ./phydev server --help for more details on this flag.

CLI commands

These are just for debugging purpose, use bpftool instead

List loaded eBPF programs:

sudo ./phydev inspect progs list
> ID      Type    Tag     RunCount        RunTime AvgRunTime
> 3               CGroupDevice    e3dbd137be8d6168        0       0s      0s
> 4               CGroupSKB       6deef7357e7b4530        0       0s      0s
> 125     uprobe__BIO_new Kprobe  0d9ea14e5516f975        0       0s      0s
> 126     socket__http_fi SocketFilter    6b7ab673cb23d3f0        0       0s      0s
> 127     kretprobe__do_s Kprobe  154f35d6575c73f9        0       0s      0s
> 128     uretprobe__SSL_ Kprobe  8737d2e349595de3        0       0s      0s

List loaded eBPF maps:

% sudo ./phydev inspect maps list 
ID      Name    FD      Type    Flags   IsPinned        KeySize ValueSize       MaxEntries
1               3       Hash    0       false   9       1       500
2               4       Hash    0       false   9       1       500
44      do_sendfile_arg 10      Hash    0       false   8       8       1024
61      http_in_flight  27      Hash    0       false   48      112     10000
62      http_notificati 28      PerfEventArray  0       false   4       4       16
63      open_at_args    29      Hash    0       false   8       128     1024

Docker

Instead of ./phydev server, use docker command:

docker run -ti --rm --cap-add CAP_SYS_ADMIN --pid=host -e BPF_DIR=/sys/fs/bpf -v /sys/fs/bpf:/sys/fs/bpf -p 8080:8080 ghcr.io/ebpfdev/dev-agent:v0.0.4 server

Security options breakdown

Required:

• --cap-add CAP_SYS_ADMIN is needed for access BPF maps and programs (CAP_BPF is not yet enough to list existing maps and programs)

Optional:

• --pid=host is needed to determine tracepoint/kprobe attachment, if you skip it, you won't see what tracepoint/kprobe a program is attached to:

• -e BPF_DIR=/sys/fs/bpf -v /sys/fs/bpf:/sys/fs/bpf is needed to determine pinned maps (they will be resolved relative to BPF_DIR tough, so it's better to mount it to the same path) if you skip it, you won't see maps pinned path, and it won't be possible to pin new maps:

Development

Build

go build -o phydev cmd/dev-agent/main.go

Updated generated code

This will update GraphQL code and some other constants:

go install golang.org/x/tools/cmd/stringer@latest
go generate ./...