Re: [PATCH] clone.2: add CLONE_PIDFD entry

["Michael Kerrisk (man-pages)" <mtk.manpages@gmail.com>]

Hello Christian,

On 9/16/19 9:40 AM, Christian Brauner wrote:
> On Wed, Sep 11, 2019 at 10:58:57AM +0200, Michael Kerrisk (man-pages) wrote:
>> Hello Christian,
>>
>> On 5/11/19 8:49 AM, Christian Brauner wrote:
>>> From: Christian Brauner <christian@brauner.io>
>>>
>>> Add an entry for CLONE_PIDFD. This flag is available starting with
>>> kernel 5.2. If specified, a process file descriptor ("pidfd") referring
>>> to the child process will be returned in the ptid argument.
>>
>> I've applied this patch in a local branch, and made some minor edits
> 
> Thank you! :)
> 
>> and added a piece. And I have some questions. See below.
>>
>>> Signed-off-by: Christian Brauner <christian@brauner.io>
>>> ---

[...]

>>> Note, that the kernel verifies that the value for
>>> +.I ptid
>>> +is zero. If it is not an error will be returned. This ensures that
>>> +.I ptid
>>> +can potentially be used to specify additional options for
>>> +.B CLONE_PIDFD
>>> +in the future.
>>
>> This piece is no longer true, right? At least I can't see such 
> 
> Correct.

Thanks. Page amended.

>> a check in the kernel code, and my testing doesn't yield an error
>> when ptid != 0 before the call.(No need to send me a patch; if I'm
>> correct just let me know and I'll edit out this piece.)
>>
>>> +.IP
>>> +Since the
>>> +.I ptid
>>> +argument is used to return the pidfd,
>>> +.B CLONE_PIDFD
>>> +cannot be used with
>>> +.B CLONE_PARENT_SETTID.
>>> +.IP
>>> +It is currently not possible to use this flag together with
>>> +.B CLONE_THREAD.
>>> +This means that the process identified by the pidfd will always be a
>>> +thread-group leader.
>>> +.IP
>>> +For a while there was a
>>> +.B CLONE_DETACHED
>>> +flag. This flag is usually ignored when passed along with other flags.
>>> +However, when passed alongside
>>> +.B CLONE_PIDFD
>>> +an error will be returned. This ensures that this flag can be reused
>>> +for further pidfd features in the future.
>>> +.TP
>>>  .BR CLONE_PTRACE " (since Linux 2.2)"
>>>  If
>>>  .B CLONE_PTRACE
>>> @@ -1122,6 +1158,21 @@ For example, on aarch64,
>>>  .I child_stack
>>>  must be a multiple of 16.
>>>  .TP
>>> +.B EINVAL
>>> +.B CLONE_PIDFD
>>> +was specified together with
>>> +.B CLONE_DETACHED.
>>> +.TP
>>> +.B EINVAL
>>> +.B CLONE_PIDFD
>>> +was specified together with
>>> +.B CLONE_PARENT_SETTID.
>>> +.TP
>>> +.B EINVAL
>>> +.B CLONE_PIDFD
>>> +was specified together with
>>> +.B CLONE_THREAD.
>>> +.TP
>>>  .B ENOMEM
>>>  Cannot allocate sufficient memory to allocate a task structure for the
>>>  child, or to copy those parts of the caller's context that need to be
>>
>> One other piece seems to be missing: the returned file descriptor can
>> be fed to poll()/select()/epoll and the FD will test as readable when
>> the child terminates. Right? Did that functionality also land in
>> kernel 5.2? And did it get implemented as a separate commit, or did
>> the behavior just fall naturally out of the implementation of pidfd's?
>> Let me know the details, and I will craft a patch.
> 
> It landed in 5.3. The relevant commit is:
> https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=b53b0b9d9a613c418057f6cb921c2f40a6f78c24
> and belongs to the following merge:
> https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=5450e8a316a64cddcbc15f90733ebc78aa736545

Thanks for that info. One other questions springs to mind.
I haven't looked at the source or tried testing this,
but can anything actually be read() from a PIDFD? Presumably,
it might be useful to have data generated on the FD, since
different values could (ultimately) be used to distinguish
between terminate/stopp/continue transitions.

>> Also, as far as I can see (from testing) the FD only gives pollable
>> events on process termination, not on other process transitions such
>> as stop and continue. Right? (Are there any plans to implement such
> 
> Correct.
> 
>> functionality for stop/contine transitions?
> 
> Yes, at some point we will likely want this.
> 
>>
>> By the way, when do you expect the pidfd-wait functionality to land 
>> in the kernel?
> 
> I've sent a PR for 5.4:
> https://lkml.org/lkml/2019/9/10/682
> which contains the P_PIDFD extension to waitid().

Thanks for that pointer. I see that the code is
now merged.

> Thanks for the work, Michael!

You're welcome!

Cheers,

Michael



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