On Thu Nov 12, 2020 at 4:27 PM PST, Yonghong Song wrote:
>
>
> On 11/12/20 2:20 PM, Daniel Xu wrote:
> > Hi,
> >
> > I'm looking at the current implementation of
> > bpf_get_ns_current_pid_tgid() and the helper seems to be a bit overly
> > restricting to me. Specifically the following line:
> >
> > if (!ns_match(&pidns->ns, (dev_t)dev, ino))
> > goto clear;
> >
> > Why bail if the inode # does not match? IIUC from the old discussions,
> > it was b/c in the future pidns files might belong to different devices.
> > It's not clear to me (possibly b/c I'm missing something) why the inode
> > has to match as well.
>
> Yes, pidns file might belong to different devices in theory so we need
> to match dev as well.
>
> The inode number needs to match so we can ensure user indeed wants to
> get the *current pidns* tgid/pid.
Right, this double-checking at the API level is what feels strange to
me -- why make the user prove they know what they're doing?
Furthermore, the "proof" restricts flexibility. It's as if
bpf_get_current_task() required a (dev,ino) pair. How would you get the
namespaced pid for a process you don't know about yet? eg when you're
profiling the system.
>
> (dev, ino) input expressed user intention. Without this, in no-process
> context, it will be hard to interpret the results.
But bpf_get_current_pid_tgid() doesn't return errors so this shouldn't
either, right?
>
> >
> > Would it be possible to instead have the helper return the pid/tgid of
> > the current task as viewed _from_ the `dev`/`ino` pidns? If the current
> > task is hidden from the `dev`/`ino` pidns, then return -ENOENT. The use
> > case is for bpftrace symbolize stacks when run inside a container. For
> > example:
> >
> > (in-container)# bpftrace -e 'profile:hz:99 { print(ustack) }'
>
> I think you try to propose something like below:
> - user provides dev/ino
> - the helper will try to go through all pidns'es (not just active
> one), if any match pidns match, returns tgid/pid in that pidns,
> otherwise, returns -ENOENT.
Right, exactly.
>
> The current helper is
> bpf_get_ns_current_pid_tgid
> you want
> bpf_get_ns_pid_tgid
>
> I think it is possible, you need to check
> pid->numbers[pid_level].ns
> for all pid levels. You need to get a reference count for the namespace
> to ensure valid result.
>
> This may work for root inode, but for container inode, it may have
> issues. For example,
> container 1: create, inode 2
> container 1 removed
> container 2: create, inode 2
> If you use inode 2, depending on timing you may accidentally targetting
> wrong container.
Yeah, so maybe an fd to /proc/<pid>/ns/pid or something.
>
> I think you can workaround the issue without this helper. See below.
>
> >
> > This currently does not work b/c bpftrace will generate a prog that gets
> > the root pidns pid, pack it with the stackid, and pass it up to
> > userspace. But b/c bpftrace is running inside the container, the root
> > pidns pid is invalid and symbolization fails.
>
> bpftrace can generate a program takes dev/inode as input parameters in
> map. The bpftrace will supply dev/inode value, by query the current
> system/container, and then run the program.
I don't think it's very feasible to have bpftrace integrate with every
container runtime out there. This also becomes really difficult to
manage if you want to trace N processes. You'd need N maps or N progs.
>
> >
> > What would be nice is if bpftrace could generate a prog that gets the
> > current pid as viewed from bpftrace's pidns. Then symbolization would
> > work.
>
> Despite the above workaround, what you really need is although it is
> running on container, you want to get stack trace interpreted with
> root pid/tgid for symbolization purpose? But you can already achieve
> this with bpf_get_pid_tgid()?
No, this isn't possible when bpftrace runs inside the container. ie
bpftrace is in a pidns along with the tracees. Bpftrace gets the root
pidns pid from the kernel but cannot resolve it to the pidns pid. That
means bpftrace cannot find the executable file to symbolize against.
[...]
Thanks,
Daniel
