Re: For review: seccomp_user_notif(2) manual page [v2]

From: Michael Kerrisk (man-pages)
Date: Sat Oct 31 2020 - 12:31:23 EST


Hello Sargun,

Thanks for your reply.

On 10/30/20 9:27 PM, Sargun Dhillon wrote:
> On Thu, Oct 29, 2020 at 09:37:21PM +0100, Michael Kerrisk (man-pages)
> wrote:

[...]

>>> I think I commented in another thread somewhere that the
>>> supervisor is not notified if the syscall is preempted. Therefore
>>> if it is performing a preemptible, long-running syscall, you need
>>> to poll SECCOMP_IOCTL_NOTIF_ID_VALID in the background, otherwise
>>> you can end up in a bad situation -- like leaking resources, or
>>> holding on to file descriptors after the program under
>>> supervision has intended to release them.
>>
>> It's been a long day, and I'm not sure I reallu understand this.
>> Could you outline the scnario in more detail?
>>
> S: Sets up filter + interception for accept T: socket(AF_INET,
> SOCK_STREAM, 0) = 7 T: bind(7, {127.0.0.1, 4444}, ..) T: listen(7,
> 10) T: pidfd_getfd(T, 7) = 7 # For the sake of discussion.

Presumably, the preceding line should have been:

S: pidfd_getfd(T, 7) = 7 # For the sake of discussion.
(s/T:/S:/)

right?

> T: accept(7, ...) S: Intercepts accept S: Does accept in background
> T: Receives signal, and accept(...) responds in EINTR T: close(7) S:
> Still running accept(7, ....), holding port 4444, so if now T
> retries to bind to port 4444, things fail.

Okay -- I understand. Presumably the solution here is not to
block in accept(), but rather to use poll() to monitor both the
notification FD and the listening socket FD?

>>> A very specific example is if you're performing an accept on
>>> behalf of the program generating the notification, and the
>>> program intends to reuse the port. You can get into all sorts of
>>> awkward situations there.
>>
>> [...]
>>
> See above

[...]

>>> In addition, if it is a socket, it inherits the cgroup v1 classid
>>> and netprioidx of the receiving process.
>>>
>>> The argument of this is as follows:
>>>
>>> struct seccomp_notif_addfd { __u64 id; __u32 flags; __u32 srcfd;
>>> __u32 newfd; __u32 newfd_flags; };
>>>
>>> id This is the cookie value that was obtained using
>>> SECCOMP_IOCTL_NOTIF_RECV.
>>>
>>> flags A bitmask that includes zero or more of the
>>> SECCOMP_ADDFD_FLAG_* bits set
>>>
>>> SECCOMP_ADDFD_FLAG_SETFD - Use dup2 (or dup3?) like semantics
>>> when copying the file descriptor.
>>>
>>> srcfd The file descriptor number to copy in the supervisor
>>> process.
>>>
>>> newfd If the SECCOMP_ADDFD_FLAG_SETFD flag is specified this will
>>> be the file descriptor that is used in the dup2 semantics. If
>>> this file descriptor exists in the receiving process, it is
>>> closed and replaced by this file descriptor in an atomic fashion.
>>> If the copy process fails due to a MAC failure, or if srcfd is
>>> invalid, the newfd will not be closed in the receiving process.
>>
>> Great description!
>>
>>> If SECCOMP_ADDFD_FLAG_SETFD it not set, then this value must be
>>> 0.
>>>
>>> newfd_flags The file descriptor flags to set on the file
>>> descriptor after it has been received by the process. The only
>>> flag that can currently be specified is O_CLOEXEC.
>>>
>>> On success, this operation returns the file descriptor number in
>>> the receiving process. On failure, -1 is returned.
>>>
>>> It can fail with the following error codes:
>>>
>>> EINPROGRESS The cookie number specified hasn't been received by
>>> the listener
>>
>> I don't understand this. Can you say more about the scenario?
>>
>
> This should not really happen. But if you do a ADDFD(...), on a
> notification *before* you've received it, you will get this error. So
> for example,
> --> epoll(....) -> returns
> --> RECV(...) cookie id is 777
> --> epoll(...) -> returns
> <-- ioctl(ADDFD, id = 778) # Notice how we haven't done a receive yet
> where we've received a notification for 778.

Got it. Looking also at the source code, I came up with the
following:

EINPROGRESS
The user-space notification specified in the id
field exists but has not yet been fetched (by a
SECCOMP_IOCTL_NOTIF_RECV) or has already been
responded to (by a SECCOMP_IOCTL_NOTIF_SEND).

Does that seem okay?

>>> ENOENT The cookie number is not valid. This can happen if a
>>> response has already been sent, or if the syscall was
>>> interrupted
>>>
>>> EBADF If the file descriptor specified in srcfd is invalid, or if
>>> the fd is out of range of the destination program.
>>
>> The piece "or if the fd is out of range of the destination program"
>> is not clear to me. Can you say some more please.
>>
>
> IIRC the maximum fd range is specific in proc by some sysctl named
> nr_open. It's also evaluated against RLIMITs, and nr_max.
>
> If nr-open (maximum fds open per process, iiirc) is 1000, even if 10
> FDs are open, it wont work if newfd is 1001.

Actually, the relevant limit seems to be just the RLIMIT_NOFILE
resource limit at least in my reading of fs/file.c::replace_fd().
So I made the text

EBADF Allocating the file descriptor in the target would
cause the target's RLIMIT_NOFILE limit to be
exceeded (see getrlimit(2)).


>>> EINVAL If flags or new_flags were unrecognized, or if newfd is
>>> non-zero, and SECCOMP_ADDFD_FLAG_SETFD has not been set.
>>>
>>> EMFILE Too many files are open by the destination process.

I'm not sure that the error can really occur. That's the error
that in most other places occurs when RLIMIT_NOFILE is exceeded.
But I may have missed something. More precisely, when do you think
EMFILE can occur?

[...]

Thanks,

Michael

--
Michael Kerrisk
Linux man-pages maintainer; http://www.kernel.org/doc/man-pages/
Linux/UNIX System Programming Training: http://man7.org/training/