Skip to content

Latest commit

 

History

History
320 lines (233 loc) · 12.3 KB

README.md

File metadata and controls

320 lines (233 loc) · 12.3 KB
Raw

🌴 bpftree

🗞️ Why bpftree?

During my working routine, I often use tools like pstree, pidof combined with some /proc queries. Consider a concrete case in which you have just spawned a tail process inside a container and you want to know its full lineage. With tools like pstree you should probably grep the tail process and then try to trace the lineage by hand, nothing impossible but of course quite noisy... Moreover imagine you want to know some details about one of the ancestors, in the best case you have to scan /proc for this info, but in the worst one /proc can't help you either because the information you are searching for is not available... So how bpftree come into play? Using the power of eBPF iterators we can scan all tasks running in the kernel and aggregate them generating graphs like trees and process lineages. Moreover, since we are using eBPF we can scrape almost all data we want from the kernel without strict constraints due to the /proc interface! Let's see bpftree in action.

We said we want to obtain the tail lineage, well, let's ask bpftree to do it for us:

sudo bpftree lineage comm tail
# or short form
sudo bpftree l c tail

Here we are asking bpftree to print the lineage for all processes with comm==tail. No need for pidof tail or some "fancy greps" you can just use the process name.

In this example, we have only one tail process in our system:

📜 Task Lineage for 'comm=tail'
⬇️ [tail] tid: 24996, pid: 24996, rptid: 24776, rppid: 24776
⬇️ [zsh] tid: 24776, pid: 24776, rptid: 24755, rppid: 24755
⬇️ [gnome-terminal-] tid: 24755, pid: 24755, rptid: 1297, rppid: 1297
⬇️ [systemd]💀 tid: 1297, pid: 1297, rptid: 1, rppid: 1
⬇️ [systemd] tid: 1, pid: 1, rptid: 0, rppid: 0

Please note: The symbol 💀 means that the process is a child_subreaper.

Let's consider we want to know some extra info, something like the session id (sid). Let's configure bpftree to print tasks according to a specific format:

sudo bpftree lineage comm tail --format 'tid,pid,sid'
# or short form
sudo bpftree l c tail -f 't,p,s'

Here we are asking the same thing as before but we are telling bpftree to print tasks according to a specific format. Please note that when you specify a custom format all the displayed fields are abbreviated (tid becomes t and so on) this is done to support long lists of fields.

📜 Task Lineage for comm: tail
⬇️ [tail] t: 24996, p: 24996, s: 24776
⬇️ [zsh] t: 24776, p: 24776, s: 24776
⬇️ [gnome-terminal-] t: 24755, p: 24755, s: 24755
⬇️ [systemd]💀 t: 1297, p: 1297, s: 1297
⬇️ [systemd] t: 1, p: 1, s: 1

You can find the full list of fields with all the associated abbreviations here or you can directly use the fields command:

sudo bpftree fields
# or short form
sudo bpftree f

Please note: if you run bpftree inside a container you will see only the processes inside the current pid namespace.

⛓️ Requirements

The "known supported architectures" are:

  • x86_64
  • aarch64
  • s390x

"known supported architectures" means that we tried bpftree on these architectures but others could be supported as well.

Right now the unique BPF requirement is the presence of BPF iterators in the running system. They were introduced in kernel version 5.8 so all newer kernel versions should be supported

✔️ Supported fields

Fields Description
tid,t thread id (tid) of the current task (init namespace)
vtid,vt thread id (tid) of the current task (task namespace)
pid,p process id (pid) of the current task (init namespace)
vpid,vp process id (pid) of the current task (task namespace)
pgid,pg process group id (pgid) of the current task (init namespace)
vpgid,vpg process group id (pgid) of the current task (task namespace)
sid,s session id (sid) of the current task (init namespace)
vsid,vs session id (sid) of the current task (task namespace)
ptid,pt parent thread id (ptid) of the current task (init namespace)
ppid,pp parent process id (ppid) of the current task (init namespace)
rptid,rpt real parent thread id (rptid) of the current task (init namespace)
rppid,rpp real parent process id (rppid) of the current task (init namespace)
comm,c human readeable process name ('task->comm')
reaper,r true if the current process is a child_sub_reaper
ns_level,ns pid namespace level of the actual thread
exepath,e full executable path of the current task
loginuid,lu UID of the user that interacted with a login service
euid,eu Effective UID
cmdline,cmd Command line of the current task

In the Fields column, we have the full field name as a first string and its short name as a second one (e.g. "tid,t", tid is the full name while t is the short one).

To specify the format you can use either the full name or the short one:

## Full
sudo ./bpftree info tid 1 --format "tid,pid" 
## Short
sudo ./bpftree i t 1 -f "t,p"
## They will produce the same result
🗞️ [systemd] t: 1, p: 1

The output always uses the short format to improve readability with long field lists

🎮 Available commands

ℹ️ info

This command shows all tasks that match a certain condition. The condition is expressed through a field, the field name is provided as the first argument while the value is provided as the second one.

sudo bpftree info pid 8056
# or in short form
sudo bpftree i p 8056
ℹ️ Task Info for pid: 8056
🗞️ [plugin_host-3.3] tid: 8056, pid: 8056, rptid: 7972, rppid: 7972
🗞️ {thread_queue} tid: 8061, pid: 8056, rptid: 7972, rppid: 7972
🗞️ {process_status} tid: 8064, pid: 8056, rptid: 7972, rppid: 7972

Here the condition is: "print all tasks with pid==8056". So print all tasks that belong to the thread group with id 8056.

To write a condition you can use all fields shown by the fields command. These same fields can be also used to customize the command output but by default, all tasks are rendered in the default "info format":

[comm] tid: ..., pid: ..., rptid: ..., rppid: ...

Please note: When the rendered task is a leader thread its comm is displayed like this [comm] while for secondary threads we use this notation {comm}.

If you want to obtain also the vtid for all the selected tasks, you have just to customize the format. All fields must be separated by ',' (e.g. 'tid,pid,reaper') and DON'T have white spaces between them.

sudo bpftree info pid 3 --format 'tid,pid,rptid,rppid,vtid'
# or in short form
sudo bpftree i p 3 --format 't,p,rpt,rpp,vt'
ℹ️ Task Info for pid: 8056
🗞️ [plugin_host-3.3] t: 8056, p: 8056, rpt: 7972, rpp: 7972, vt: 8056
🗞️ {thread_queue} t: 8061, p: 8056, rpt: 7972, rpp: 7972, vt: 8061
🗞️ {process_status} t: 8064, p: 8056, rpt: 7972, rpp: 7972, vt: 8064

📜 lineage

This command shows the lineage for all tasks that match a certain condition. The command features are the same as the "info" ones.

sudo bpftree lineage pid 8056
# or in short form
sudo bpftree l p 8056
📜 Task Lineage for pid: 8056
⬇️ [plugin_host-3.3] tid: 8056, pid: 8056, rptid: 7972, rppid: 7972
⬇️ [sublime_text] tid: 7972, pid: 7972, rptid: 1486, rppid: 1486
⬇️ [systemd]💀 tid: 1486, pid: 1486, rptid: 1, rppid: 1
⬇️ [systemd] tid: 1, pid: 1, rptid: 0, rppid: 0

⬇️ {thread_queue} tid: 8061, pid: 8056, rptid: 7972, rppid: 7972
⬇️ [sublime_text] tid: 7972, pid: 7972, rptid: 1486, rppid: 1486
⬇️ [systemd]💀 tid: 1486, pid: 1486, rptid: 1, rppid: 1
⬇️ [systemd] tid: 1, pid: 1, rptid: 0, rppid: 0

⬇️ {process_status} tid: 8064, pid: 8056, rptid: 7972, rppid: 7972
⬇️ [sublime_text] tid: 7972, pid: 7972, rptid: 1486, rppid: 1486
⬇️ [systemd]💀 tid: 1486, pid: 1486, rptid: 1, rppid: 1
⬇️ [systemd] tid: 1, pid: 1, rptid: 0, rppid: 0

bpftree shows the lineage for all the tasks with pid==8056.

Also here the default "lineage format" is:

[comm] tid: ..., pid: ..., rptid: ..., rppid: ...

🌴 tree

This command shows the tree for all tasks that match a certain condition. The command features are the same as the "info" ones.

sudo bpftree tree pid 8056
# or in short form
sudo bpftree t p 8056
🌴 Task Tree for pid: 8056
[plugin_host-3.3] tid: 8056, pid: 8056

{thread_queue} tid: 8061, pid: 8056

{process_status} tid: 8064, pid: 8056

bpftree shows the tree for all the tasks with pid==8056. Here is quite useless since all the involved tasks are tree leaves.

The default "tree format" is:

[comm] tid: ..., pid: ...

but you can always customize it with the --format flag.

🗺️ fields

This command shows all fields that can be used both to write conditions and to customize the output.

sudo bpftree fields
# or short form
sudo bpftree f

📂 files

This command shows all the file descriptors for all tasks that match a certain condition.

# show all fds of the thread with tid=1
sudo ./bpftree files tid 1
# or short form
sudo ./bpftree fds t 1

🔻 dump

This command allows you to save read system tasks into a file. Let's imagine you are using an enterprise tool and you find a bug with a specific system configuration. You could dump a capture file and send it to the support team to analyze it. Thanks to the --capture flag bpftree can replay previously dumped files coming also from systems with different architectures.

# Dump the file
sudo bpftree dump output-file.tree
# Read it with the capture flag
sudo bpftree info tid 1 --capture output-file.tree

On the capture file, you can use all commands listed above like in a live run.

The capture files follow the versioning of the bpftree tool. This means that a particular version of bpftree can read only capture files with the same Major and Minor and with a lower or equal Patch. If needed we could create a wrapper tool that calls the right bpftree version for the provided capture file.

For developers

🏗️ Build from source

As with many go project you have just to type a bunch of commands in the root folder of the project:

go generate ./...
go build .

🏎️ Run

In the root folder of the project:

sudo ./bpftree --help

As an example try to log the full process tree of your system 🕹️

sudo ./bpftree tree tid 1
# or short form
sudo ./bpftree t t 1

🧪 Run tests

In the root folder of the project run:

# Unit tests
sudo go test ./pkg/...

# All tests
sudo -E env "PATH=$PATH" go test ./... -count=1

Run linters

To run golangci-lint

golangci-lint run  --timeout=900s

Note: be sure that the used golangci-lint version is compatible with your go version. Tested with golangci-lint 1.59.1 built with go1.22.3 and go1.22.1

➕ Add a new field

  • BPF side
    1. add the new field to exported_task_info or exported_file_info struct
    2. instrument the code to collect the new field
  • Userspace side for taskInfo
    1. add the new field to taskInfo struct
    2. add a getter method for the new field in task.go
    3. add an enum allowedField for the new field
    4. add a new entry for the field into allowedFieldsSlice
    5. add an entry in the TestGetFieldMatrix for the new getter method
    6. update testdata/script/cmd_fields/cmd_fields.txtar file with the new output of sudo ./bpftree f
    7. update testdata/script/fields/systemd_fields.txtar tests with the new field
  • Userspace side for fileInfo
    1. add the new field to fileInfo struct
    2. add the new field into the output of the String() method