[lkml]   [2019]   [Nov]   [9]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
SubjectRe: For review: documentation of clone3() system call
On Sat, Nov 09, 2019 at 09:09:51AM +0100, Michael Kerrisk (man-pages) wrote:
> [CC += Ingo, in case he has something to add re MAP_STACK; perhaps
> Florian also might have some thoughts]

Good idea.

> >> NAME
> >> clone, __clone2 - create a child process
> >
> > Should this include clone3()?
> >
> >>
> >> /* Prototype for the glibc wrapper function */
> >>
> >> #define _GNU_SOURCE
> >> #include <sched.h>
> >>
> >> int clone(int (*fn)(void *), void *stack, int flags, void *arg, ...
> >> /* pid_t *parent_tid, void *tls, pid_t *child_tid */ );
> >
> > I've always been confused by the "..." for the glibc wrapper. The glibc
> > prototype in bits/sched.h also looks like this:
> I'm not sure that I understand your confusion. The point is, it's a
> variadic function: the extra arguments are only ever touched if the
> relevant flags are specified.

Sure. My confusion is why it is a) variadic in the first place and b)
even if it is can't we just advertise it as non-variadic since these
functions will not grow any additional arguments.

> Or maybe you just meant: "this should not be a variadic function,
> but rather that all 7 arguments should always be specified"? But,
> then I think the point is that clone() (and the underlying syscall)
> did not always have this number of arguments. Before Linux 2.6, the
> wrapper and syscall only had 4 arguments.

Yeah, though I wonder whether this is all so ancient that it doesn't
matter anymore at this point in time.

> >> These system calls create a new process, in a manner similar to
> >> fork(2).
> >>
> >> Unlike fork(2), these system calls allow the child process to
> >> share parts of its execution context with the calling process,
> >
> > Hm, sharing part of the execution context is not the only thing that
> > clone{3}() does.
> True. That text has been in the page for 21 years. It probably needs
> a new coat of paint...
> > Maybe something like:
> >
> > Unlike fork(2), these system calls allow to create a child process with
> > different properties than its parent. For example, these syscalls allow
> > the child to share various parts of the execution context with the
> > calling process such as [...]. They also allow placing the process in a
> > new set of namespaces.
> >
> > Just a thought.
> A good thought...
> I changed the text to read:
> Unlike fork(2), these system calls allow the child to be created
> with various properties that differ from the parent. For example,
> these system calls provide more precise control over what pieces
> of execution context are shared between the calling process and
> the child process. For example, using these system calls, the
> caller can control whether or not the two processes share the vir‐
> tual address space, the table of file descriptors, and the table
> of signal handlers. These system system calls also allow the new
> child process to placed in separate namespaces(7).
> Okay?

Nit: The second and thirs sentence both start with "For example".
Otherwise sounds great.

> >> such as the virtual address space, the table of file descriptors,
> >> and the table of signal handlers. (Note that on this manual page,
> >> "calling process" normally corresponds to "parent process". But
> >> see the description of CLONE_PARENT below.)
> >>
> >> This page describes the following interfaces:
> >>
> >> * The glibc clone() wrapper function and the underlying system
> >> call on which it is based. The main text describes the wrapper
> >> function; the differences for the raw system call are described
> >> toward the end of this page.
> >>
> >> * The newer clone3() system call.
> >>
> >> The clone() wrapper function
> >> When the child process is created with the clone() wrapper func‐
> >> tion, it commences execution by calling the function pointed to by
> >> the argument fn. (This differs from fork(2), where execution con‐
> >> tinues in the child from the point of the fork(2) call.) The arg
> >> argument is passed as the argument of the function fn.
> >>
> >> When the fn(arg) function returns, the child process terminates.
> >> The integer returned by fn is the exit status for the child
> >> process. The child process may also terminate explicitly by call‐
> >> ing exit(2) or after receiving a fatal signal.
> >>
> >> The stack argument specifies the location of the stack used by the
> >> child process. Since the child and calling process may share mem‐
> >> ory, it is not possible for the child process to execute in the
> >> same stack as the calling process. The calling process must
> >> therefore set up memory space for the child stack and pass a
> >> pointer to this space to clone(). Stacks grow downward on all
> >
> > It might be a good idea to advise people to use mmap() to create a
> > stack. The "canonical" way of doing this would usually be something like
> >
> > #define DEFAULT_STACK_SIZE (4 * 1024 * 1024) /* 8 MB usually on Linux */
> >
> > (Yes, the MAP_STACK is usally a noop but people should always include it
> > in case some arch will have weird alignment requirement in which case
> > this flag can be changed to actually do something...)
> So, I'm getting a little bit of an education here, and maybe you are
> going to further educate me. Long ago, I added the documentation of
> MAP_STACK to mmap(2), but I never quite connected the dots.
> However, you say MAP_STACK is *usually* a noop. As far as I can see,
> in current kernels it is *always* a noop. And AFAICS, since it was first
> added in 2.6.27 (2008), it has always been a noop.
> I wonder if it will always be a noop.
> If we go back and look at the commit:
> [[
> commit 2fdc86901d2ab30a12402b46238951d2a7891590
> Author: Ingo Molnar <>
> Date: Wed Aug 13 18:02:18 2008 +0200
> x86: add MAP_STACK mmap flag
> as per this discussion:
> Pardo reported that 64-bit threaded apps, if their stacks exceed the
> combined size of ~4GB, slow down drastically in pthread_create() - because
> glibc uses MAP_32BIT to allocate the stacks. The use of MAP_32BIT is
> a legacy hack - to speed up context switching on certain early model
> 64-bit P4 CPUs.
> So introduce a new flag to be used by glibc instead, to not constrain
> 64-bit apps like this.
> glibc can switch to this new flag straight away - it will be ignored
> by the kernel. If those old CPUs ever matter to anyone, support for
> it can be implemented.
> ]]
> And see also
> So, my understanding from the above is that MAP_STACK was added to
> allow a possible fix on some old architectures, should anyone decide it
> was worth doing the work of implementing it. But so far, after 12 years,
> no one did. It kind of looks like no one ever will (since those old
> architectures become less and less relevant).
> So, AFAICT, while it's not wrong to tell people to use mmap(MAP_STACKED),
> it doesn't provide any benefit (and perhaps never will), and it is a
> more clumsy than plain old malloc().

Apart from MAP_STACK at some point (however unlikely) becoming relevant
I also like that mmap() makes it explicit that the page needs to be
readable (PROT_READ) and writeable (PROT_WRITE) and also every
(relevant) libc I know uses the explicit mmap() in their pthread
implementation. If you prefer the malloc() as an example that's
obviously fine.

> Yeah, but I didn't get mentioned in the commit ;-)

You provided your Ack after I had taken the patch into my tree and I've
already changed it once for Arnd's ack so I didn't want to change it

> >> CLONE_NEWUTS (since Linux 2.6.19)
> >> If CLONE_NEWUTS is set, then create the process in a new
> >> UTS namespace, whose identifiers are initialized by dupli‐
> >> cating the identifiers from the UTS namespace of the call‐
> >> ing process. If this flag is not set, then (as with
> >> fork(2)) the process is created in the same UTS namespace
> >> as the calling process. This flag is intended for the
> >> implementation of containers.
> >>
> >> A UTS namespace is the set of identifiers returned by
> >> uname(2); among these, the domain name and the hostname can
> >> be modified by setdomainname(2) and sethostname(2), respec‐
> >> tively. Changes made to the identifiers in a UTS namespace
> >> are visible to all other processes in the same namespace,
> >> but are not visible to processes in other UTS namespaces.
> >
> > Might again be a little too detailed but that's just my opinion. :)
> I agree. The thing is that the clone(2) text was written long before
> the section 7 namespaces manual pages, and some duplication occurred.
> I've removed this text, and done the same for the corresponding text


> I changed the text here to read:
> CLONE_PID (Linux 2.0 to 2.5.15)
> If CLONE_PID is set, the child process is created with the
> same process ID as the calling process. This is good for
> hacking the system, but otherwise of not much use. From
> Linux 2.3.21 onward, this flag could be specified only by
> the system boot process (PID 0). The flag disappeared com‐
> pletely from the kernel sources in Linux 2.5.16. Subse‐
> quently, the kernel silently ignored this bit if it was
> specified in the flags mask. Much later, the same bit was
> recycled for use as the CLONE_PIDFD flag.

Yip, sounds good.

> CLONE_DETACHED (historical)
> For a while (during the Linux 2.5 development series) there
> was a CLONE_DETACHED flag, which caused the parent not to
> receive a signal when the child terminated. Ultimately,
> the effect of this flag was subsumed under the CLONE_THREAD
> flag and by the time Linux 2.6.0 was released, this flag
> had no effect. Starting in Linux 2.6.2, the need to give
> this flag together with CLONE_THREAD disappeared.
> This flag is still defined, but it is usually ignored when
> calling clone(). However, see the description of
> CLONE_PIDFD for some exceptions.
> And then under CLONE_PIDFD, I have:
> If the obsolete CLONE_DETACHED flag is specified alongside
> CLONE_PIDFD when calling clone(), an error is returned. An
> error also results if CLONE_DETACHED is specified when
> calling clone3(). This error behavior ensures that the bit
> corresponding to CLONE_DETACHED can be reused for further
> PID file descriptor features in the future.
> Okay?


> >> CLONE_SETTLS (since Linux 2.5.32)
> >> The TLS (Thread Local Storage) descriptor is set to tls.
> >>
> >> The interpretation of tls and the resulting effect is
> >> architecture dependent. On x86, tls is interpreted as a
> >> struct user_desc * (see set_thread_area(2)). On x86-64 it
> >> is the new value to be set for the %fs base register (see
> >> the ARCH_SET_FS argument to arch_prctl(2)). On architec‐
> >> tures with a dedicated TLS register, it is the new value of
> >> that register.
> >
> > Probably a gentle warning that this is a very advanced option and
> > usually should not be used by callers other than libraries implementing
> > threading or with specific use cases directly.
> I added:
> Use of this flag requires detailed knowledge and generally it
> should not be used except in libraries implementing threading.

Sounds good.

> >> If any of the threads in a thread group performs an
> >> execve(2), then all threads other than the thread group
> >> leader are terminated, and the new program is executed in
> >
> > s/is executed in/becomes the/?
> Hmmm, a program is not a task, so this doesn't feel quite right.
> Why don't you like the existing text?

Just read odd to me "executed in the thread group leader".

> > I won't do a deep review of the thread section now but you might want to
> > mention that fatal signals always take down the whole thread-group, i.e.
> > SIGKILL, SIGSEGV, etc...
> That point is covered above, in the paragraph that begins: "Signal
> dispositions and actions are process-wide...". Does that not
> suffice?

This is probably too detailed from my end.
Signals can be either thread-group-directed or thread-directed. That's
one aspect. The other aspect is whether they are delivered
thread-specific or thread-group-wide. While signals can be
thread-group-directed kill() they are not necessarily delivered
thread-group-wide. (I can go into details but that's not necessarily too
useful for this manpage.) My point was just that mentioning fatal
signals (e.g. SIGKILL, SIGSTOP) are always delivered thread-group-wide
no matter if they are thread-group-directed or thread-directed. This
might all be covered by your wording already and I just didn't read
close enough.

> >> Note that the glibc clone() wrapper function makes some changes in
> >> the memory pointed to by stack (changes required to set the stack
> >> up correctly for the child) before invoking the clone() system
> >
> > In essence, you can't really use the clone{3}() syscall with a stack
> > argument directly without having to do some assembly.
> (Yes.)
> > User needing to
> > mess with stacks are well-advised to use the glibc wrapper or need to
> > really know what they are doing for _each_ arch they are using the
> > syscall on.
> I understand the issues (I think), but it's not clear to me if
> you mean that some text in the manual page needs changing.

It might be worth mentioning that using it with a stack involves more
than just allocating and passing it because people tend to assume that
using the syscall (not the glibc wrapper) with a stack just works. So
it's similar to the TLS comment we added with the difference that once
we have a glibc wrapper for clone3() the stack argument becomes easier
to use.

> Thanks. I made it:
> In contrast to the glibc wrapper, the raw clone() system call
> accepts NULL as a stack argument (and clone3() likewise allows
> cl_args.stack to be NULL). In this case, the child uses a dupli‐
> cate of the parent's stack. [...]

Sounds good.

> > I wouldn't even mention clone() for ia64 anymore. It will _not_ work
> > correctly at all. ia64 requires stack_size as it expects the stack to be
> > passed pointing to the lowest address but the clone() version for ia64
> > does not have a stack_size argument... So the only way to get clone() to
> > work on ia64 is by using the ia64 specific clone2().
> Fair enough. I removed mention of is-64 here.


> >> long clone(unsigned long flags, void *stack,
> >> int stack_size, /* Size of stack */
> >> int *parent_tid, int *child_tid,
> >> unsigned long tls);
> >
> > The additional argument is stack_size and contrary to what one would
> > expect _ignored_. I.e. on microblaze one still needs to pass stack
> > pointing to the top of the stack.
> I added this sentence:
> Although a stack_size argument is provided, stack must still point
> to the top of the stack.


> I made it:
> EINVAL (clone3() only)
> CLONE_DETACHED was specified in the flags mask.
> EINVAL (clone() only)
> CLONE_PIDFD was specified together with CLONE_DETACHED in
> the flags mask.
> Okay?


> Thanks for the detailed feedback, Christian!

Thanks for the detailed manpage.


 \ /
  Last update: 2019-11-09 17:55    [W:0.105 / U:4.840 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site