Skip to contentSkip to conversation

waitid() return value strangeness when infop is NULL

PATCH In review Michael Kerrisk 7 participants 18 messages Jan 13, 2009 – Jan 14, 2009 1 message not in the archive Man-pages: manual pages for Linux -- Sections 2, 3, 4, 5, and 7

lore
Share & tools
lore
#0 lore raw

Hello Roland,

Vegard sent me a report of a possible bug in the waitid(2) man page, but on closer examination, the issue seems to be in the waitid() system call.

POSIX.1 says of waitid():

RETURN VALUE
If WNOHANG was specified and there are no children to wait for, 0 shall be returned. If waitid() returns due to the change of state of one of its children, 0 shall be returned. Otherwise, -1 shall be returned and errno set to indicate the error.

The Linux waitid() follows this behavior, as long as infop is not NULL, which is what POSIX requires:

The application shall ensure that the infop argument points to a siginfo_t structure.

However, Linux's waitid() does permit infop to be NULL. But in this case, a successful call returns the PID of the waited-for child. Thinking about it, I suppose there is a rationale for doing this, since otherwise the caller of waitid() has no way of obtaining the PID of the waited-for child. Anyway, this behavior should either be documented, or "fixed" (i.e., change waitid() to return 0 on success even if infop is NULL). So, some questions:

a) Is it intentional to support the infop==NULL case for waitid() -- rather than simply giving an error return?

b) If so, what is the rationale for supporting this feature?

c) If the the answer a) is "yes", is it intentional for waitid(idtype,id,NULL,options) to return the PID of the child, rather than returning 0?

I'll post a test program in a follow-up message.

Cheers,

Michael
--
To unsubscribe from this list: send the line "unsubscribe linux-man" in the body of a message to [email protected] More majordomo info at http://vger.kernel.org/majordomo-info.html

#1 lore raw

[Resend with right LKML address.]

Hello Roland,

Vegard sent me a report of a possible bug in the waitid(2) man page, but on closer examination, the issue seems to be in the waitid() system call.

POSIX.1 says of waitid():

RETURN VALUE
If WNOHANG was specified and there are no children to wait for, 0 shall be returned. If waitid() returns due to the change of state of one of its children, 0 shall be returned. Otherwise, -1 shall be returned and errno set to indicate the error.

The Linux waitid() follows this behavior, as long as infop is not NULL, which is what POSIX requires:

The application shall ensure that the infop argument points to a siginfo_t structure.

However, Linux's waitid() does permit infop to be NULL. But in this case, a successful call returns the PID of the waited-for child. Thinking about it, I suppose there is a rationale for doing this, since otherwise the caller of waitid() has no way of obtaining the PID of the waited-for child. Anyway, this behavior should either be documented, or "fixed" (i.e., change waitid() to return 0 on success even if infop is NULL). So, some questions:

a) Is it intentional to support the infop==NULL case for waitid() -- rather than simply giving an error return?

b) If so, what is the rationale for supporting this feature?

c) If the the answer a) is "yes", is it intentional for waitid(idtype,id,NULL,options) to return the PID of the child, rather than returning 0?

I'll post a test program in a follow-up message.

Cheers,

Michael

Signature · 9 lines
Michael Kerrisk
Linux man-pages maintainer; http://www.kernel.org/doc/man-pages/
git://git.kernel.org/pub/scm/docs/man-pages/man-pages.git
man-pages online: http://www.kernel.org/doc/man-pages/online_pages.html
Found a bug? http://www.kernel.org/doc/man-pages/reporting_bugs.html
--
To unsubscribe from this list: send the line "unsubscribe linux-man" in
the body of a message to [email protected]
More majordomo info at  http://vger.kernel.org/majordomo-info.html
#2 lore raw

Test program below. Two test runs, of which the second shows the behavior I'm reporting.

$ ./a.out
Created child with PID 7727
infop = 0xbfe608e0
waitid() returned 0
$ ./a.out n
Created child with PID 7729
infop = (nil)
waitid() returned 7729

===

#define _GNU_SOURCE
#include <unistd.h>
#include <sys/syscall.h>
#include <sys/types.h>
#include <sys/wait.h>
#include <stdlib.h>
#include <string.h>
#include <stdio.h>
#include <signal.h>
#include <unistd.h>

int
main(int argc, char *argv[])

{
    siginfo_t info, *infop;
    int options, idtype;
    long s;
    pid_t child;

    child = fork();
    if (child == 0)
        exit(0);
    printf("Created child with PID %ld\n", (long) child);

    idtype = P_PID;
    options = WEXITED;
    infop = (argc > 1 && strchr(argv[1], 'n') != NULL) ? NULL : &info;

    printf("infop = %p\n", infop);

    s = syscall(SYS_waitid, idtype, child, infop, options, NULL);
    if (s == -1) {
        perror("waitid");
        exit(EXIT_FAILURE);
    }

    printf("waitid() returned %ld\n", s);

    exit(EXIT_SUCCESS);
}

--
To unsubscribe from this list: send the line "unsubscribe linux-man" in the body of a message to [email protected] More majordomo info at http://vger.kernel.org/majordomo-info.html

#4 lore raw

PATCH 1/1 sys_waitid: return -EFAULT for NULL

It's always been invalid to call waitid() with a NULL pointer. It was an oversight that it was allowed (and acts like a wait4() call instead).

---
 kernel/exit.c |    2 ++
 1 files changed, 2 insertions(+), 0 deletions(-)
kernel/exit.c +2 -0
index c7740fa..fa25790 100644
@@ -1760,6 +1760,8 @@ asmlinkage long sys_waitid(int which, pid_t upid,
 	enum pid_type type;
 	long ret;
 
+	if (unlikely(!infop))
+		return -EFAULT;
 	if (options & ~(WNOHANG|WNOWAIT|WEXITED|WSTOPPED|WCONTINUED))
 		return -EINVAL;
 	if (!(options & (WEXITED|WSTOPPED|WCONTINUED)))

--
To unsubscribe from this list: send the line "unsubscribe linux-man" in the body of a message to [email protected] More majordomo info at http://vger.kernel.org/majordomo-info.html

Acked-by on v1 1/1
#5 lore raw

Re: [PATCH] sys_waitid: return -EFAULT for NULL

Quoted 4 lines · Roland McGrath

It's always been invalid to call waitid() with a NULL pointer. It was an oversight that it was allowed (and acts like a wait4() call instead).

Modulo the observation that this change will break any Linux-specific application that violate POSIX.1's requirement that infop not be NULL (*), and rely on the existing Linux behavior for waitd(idtype,id,NULL,options):

(*) It seems unlikely that such applications exist, and we really should make this change for POSIX.1 conformance.

Quoted 18 lines
---
 kernel/exit.c |    2 ++
 1 files changed, 2 insertions(+), 0 deletions(-)
kernel/exit.c +2 -0
index c7740fa..fa25790 100644
@@ -1760,6 +1760,8 @@ asmlinkage long sys_waitid(int which, pid_t upid,
       enum pid_type type;
       long ret;
 
+       if (unlikely(!infop))
+               return -EFAULT;
       if (options & ~(WNOHANG|WNOWAIT|WEXITED|WSTOPPED|WCONTINUED))
               return -EINVAL;
       if (!(options & (WEXITED|WSTOPPED|WCONTINUED)))
Signature · 9 lines
Michael Kerrisk
Linux man-pages maintainer; http://www.kernel.org/doc/man-pages/
git://git.kernel.org/pub/scm/docs/man-pages/man-pages.git
man-pages online: http://www.kernel.org/doc/man-pages/online_pages.html
Found a bug? http://www.kernel.org/doc/man-pages/reporting_bugs.html
--
To unsubscribe from this list: send the line "unsubscribe linux-man" in
the body of a message to [email protected]
More majordomo info at  http://vger.kernel.org/majordomo-info.html
#6 lore raw

Re: [PATCH] sys_waitid: return -EFAULT for NULL

Quoted 11 lines · On Wed, 14 Jan 2009 12:14:20 +1300 "Michael Kerrisk"…
Quoted 4 lines · Roland McGrath

It's always been invalid to call waitid() with a NULL pointer. It was an oversight that it was allowed (and acts like a wait4() call instead).

Modulo the observation that this change will break any Linux-specific application that violate POSIX.1's requirement that infop not be NULL (*), and rely on the existing Linux behavior for waitd(idtype,id,NULL,options):

Well yes. waitid() has been in there since 2.6.9. I assume that it has had this wait4-emulation mode for that amount of time as well?

Quoted 3 lines

(*) It seems unlikely that such applications exist, and we really should make this change for POSIX.1 conformance.

Well, we might get away with it. But formally speaking, we should live with our Linux-specific screwup.

If we _are_ going to make this change then we should merge it as far back in -stable as we can, to reduce the risk that people will develop applications on kernel version A only to have then behave differently on version B.

--
To unsubscribe from this list: send the line "unsubscribe linux-man" in the body of a message to [email protected] More majordomo info at http://vger.kernel.org/majordomo-info.html

#7 lore raw

Re: [PATCH] sys_waitid: return -EFAULT for NULL

Quoted 9 lines · Andrew Morton
Quoted 6 lines · On Wed, 14 Jan 2009 12:14:20 +1300 "Michael Kerrisk"…
Quoted 4 lines · Roland McGrath

It's always been invalid to call waitid() with a NULL pointer. It was an oversight that it was allowed (and acts like a wait4() call instead).

Quoted 8 lines
Quoted 5 lines

Modulo the observation that this change will break any Linux-specific application that violate POSIX.1's requirement that infop not be NULL (*), and rely on the existing Linux behavior for waitd(idtype,id,NULL,options):

Well yes. waitid() has been in there since 2.6.9. I assume that it has had this wait4-emulation mode for that amount of time as well?

AFIACS, yes.

Quoted 5 lines
Quoted 2 lines

(*) It seems unlikely that such applications exist, and we really should make this change for POSIX.1 conformance.

Well, we might get away with it. But formally speaking, we should live with our Linux-specific screwup.

Yes, tricky. On the one hand, we shouldn't break the ABI. On the other hand, POSIX.1 is explicit in disallowing the case that would lead the ABI change made by this patch. (For what it is worth, the man page was released at pretty much the same time as the syscall, and has always documented that the return value on success was 0, and Vegard was the first person to report this case that deviated from the doc.)

Quoted 4 lines

If we _are_ going to make this change then we should merge it as far back in -stable as we can, to reduce the risk that people will develop applications on kernel version A only to have then behave differently on version B.

Ack.

Signature · 9 lines
Michael Kerrisk
Linux man-pages maintainer; http://www.kernel.org/doc/man-pages/
git://git.kernel.org/pub/scm/docs/man-pages/man-pages.git
man-pages online: http://www.kernel.org/doc/man-pages/online_pages.html
Found a bug? http://www.kernel.org/doc/man-pages/reporting_bugs.html
--
To unsubscribe from this list: send the line "unsubscribe linux-man" in
the body of a message to [email protected]
More majordomo info at  http://vger.kernel.org/majordomo-info.html
#8 lore raw

Re: [PATCH] sys_waitid: return -EFAULT for NULL

In reply to Andrew Morton
Quoted 2 lines

Well yes. waitid() has been in there since 2.6.9. I assume that it has had this wait4-emulation mode for that amount of time as well?

Correct.

Quoted 2 lines

Well, we might get away with it. But formally speaking, we should live with our Linux-specific screwup.

I agree with both of you and I don't care which we do.

Thanks,
Roland
--
To unsubscribe from this list: send the line "unsubscribe linux-man" in the body of a message to [email protected] More majordomo info at http://vger.kernel.org/majordomo-info.html

#9 lore raw

Re: [PATCH] sys_waitid: return -EFAULT for NULL

Quoted 9 lines · Roland McGrath
Quoted 2 lines

Well yes. waitid() has been in there since 2.6.9. I assume that it has had this wait4-emulation mode for that amount of time as well?

Correct.

Quoted 2 lines

Well, we might get away with it. But formally speaking, we should live with our Linux-specific screwup.

I agree with both of you and I don't care which we do.

I don't feel strongly about this, but lean towards the idea that we should fix things, as per Roland's patch, and I suspect we can get away with it, for the reasons I mentioned already.

Signature · 9 lines
Michael Kerrisk
Linux man-pages maintainer; http://www.kernel.org/doc/man-pages/
git://git.kernel.org/pub/scm/docs/man-pages/man-pages.git
man-pages online: http://www.kernel.org/doc/man-pages/online_pages.html
Found a bug? http://www.kernel.org/doc/man-pages/reporting_bugs.html
--
To unsubscribe from this list: send the line "unsubscribe linux-man" in
the body of a message to [email protected]
More majordomo info at  http://vger.kernel.org/majordomo-info.html
#10 lore raw

Re: [PATCH] sys_waitid: return -EFAULT for NULL

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

Quoted 3 lines · Michael Kerrisk

I don't feel strongly about this, but lean towards the idea that we should fix things, as per Roland's patch, and I suspect we can get away with it, for the reasons I mentioned already.

Yes, change it, it will only cause harm leaving it. I very much doubt any programmer deliberately uses this case and if it's not deliberate they likely want to know about it.

- --
➧ Ulrich Drepper ➧ Red Hat, Inc. ➧ 444 Castro St ➧ Mountain View, CA ❖ -----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.4.9 (GNU/Linux)

iEYEARECAAYFAkltMMcACgkQ2ijCOnn/RHQ8fQCgnghAaJQ4fkEMZCnQwsaza80X WrIAn2z7tEsiMy8pYJs3gkdzO6CNiKTz
=WaAV
-----END PGP SIGNATURE-----
--
To unsubscribe from this list: send the line "unsubscribe linux-man" in the body of a message to [email protected] More majordomo info at http://vger.kernel.org/majordomo-info.html

#11 lore raw

Re: [PATCH] sys_waitid: return -EFAULT for NULL

In reply to Roland McGrath
Quoted 3 lines · Roland McGrath

It's always been invalid to call waitid() with a NULL pointer. It was an oversight that it was allowed (and acts like a wait4() call instead).

I'm not going to take this.

If it was some new system call, of if there was some downside to out behavior, I might be interested. As it is, our behaviour has zero downside, and changing existing interfaces simply isn't worth it.

The alleged "downsides" are bogus:

- POSIX is not that strict. EFAULT is one of the odd error cases anyway, and even explicit requirements are irrelevant: if somebody wants to get strict conformance paperwork done, you just need to tell where you differ, and you're basically done. But perhaps more important, nobody cares.

- The "portability" argument is totally bogus, since it's not like you compile programs without even testing to another UNIX _anyway_.

So I'm simply not going to potentially break binaries over something that is so _totally_ irrelevant. Document it in the man-page instead.

Linus
--
To unsubscribe from this list: send the line "unsubscribe linux-man" in the body of a message to [email protected] More majordomo info at http://vger.kernel.org/majordomo-info.html

#12 lore raw

Re: [PATCH] sys_waitid: return -EFAULT for NULL

Quoted 4 lines · Linus Torvalds
Quoted 1 line · Roland McGrath
Quoted 8 lines
Quoted 2 lines

It's always been invalid to call waitid() with a NULL pointer. It was an oversight that it was allowed (and acts like a wait4() call instead).

I'm not going to take this.

If it was some new system call, of if there was some downside to out behavior, I might be interested. As it is, our behaviour has zero downside, and changing existing interfaces simply isn't worth it.

It has zero downside for *us*. But it is yet another example of Linux littering the Unix landscape with unnecessary inconsistencies that application writers must deal with. That's a downside for the app writers. (But, given how long the existing behavior has been in the wild, my argument here is somewhat academic...)

Quoted 3 lines

The alleged "downsides" are bogus:

- POSIX is not that strict.

Well, POSIX.1-2001 is fairly clear:

The application shall ensure that the infop argument points to a siginfo_t structure.

(Admittedly, this is a requirement imposed on the application, rather than the implementation, but the standard goes on to say that the implemenation shall fill in the structure pointed to by infop.)

Cheers,

Michael

Quoted 11 lines

EFAULT is one of the odd error cases anyway,
and even explicit requirements are irrelevant: if somebody wants to get strict conformance paperwork done, you just need to tell where you differ, and you're basically done. But perhaps more important, nobody cares.

- The "portability" argument is totally bogus, since it's not like you compile programs without even testing to another UNIX _anyway_.

So I'm simply not going to potentially break binaries over something that is so _totally_ irrelevant. Document it in the man-page instead.

Well, that's doable of course.

Signature · 9 lines
Michael Kerrisk
Linux man-pages maintainer; http://www.kernel.org/doc/man-pages/
git://git.kernel.org/pub/scm/docs/man-pages/man-pages.git
man-pages online: http://www.kernel.org/doc/man-pages/online_pages.html
Found a bug? http://www.kernel.org/doc/man-pages/reporting_bugs.html
--
To unsubscribe from this list: send the line "unsubscribe linux-man" in
the body of a message to [email protected]
More majordomo info at  http://vger.kernel.org/majordomo-info.html
#13 lore raw

Re: [PATCH] sys_waitid: return -EFAULT for NULL

Quoted 4 lines · Michael Kerrisk

It has zero downside for *us*. But it is yet another example of Linux littering the Unix landscape with unnecessary inconsistencies that application writers must deal with.

Bah. Not so. It matters not at all if you try to write portable code.

Linux has extensions. Deal with it. We have literally _thousands_ of things that work on Linux but not on other OS's. The fact is, you can't just recompile and assume something works, and waitid() has absolutely nothing to do with it.

Quoted 4 lines

Well, POSIX.1-2001 is fairly clear:

The application shall ensure that the infop argument points to a siginfo_t structure.

Right. So the application should do that, and Linux does the right thing. Problem solved.

Linus
--
To unsubscribe from this list: send the line "unsubscribe linux-man" in the body of a message to [email protected] More majordomo info at http://vger.kernel.org/majordomo-info.html

#14 lore raw

Re: [PATCH] sys_waitid: return -EFAULT for NULL

Quoted 11 lines · Linus Torvalds
Quoted 4 lines · Michael Kerrisk

It has zero downside for *us*. But it is yet another example of Linux littering the Unix landscape with unnecessary inconsistencies that application writers must deal with.

Bah. Not so. It matters not at all if you try to write portable code.

Linux has extensions.

This a behavior that was unintended by the implementer, wasn't requested by application writers, isn't present on other Unix systems, and isn't specified by the standards.

It isn't an extension. It's an accident.

And the fact that such accidents happen more often than necessary is the real problem, rather than the fact that this API in particular is inconsistent with expectations.

Quoted 5 lines

Deal with it. We have literally _thousands_ of
things that work on Linux but not on other OS's. The fact is, you can't just recompile and assume something works, and waitid() has absolutely nothing to do with it.

Quoted 7 lines
Quoted 4 lines

Well, POSIX.1-2001 is fairly clear:

The application shall ensure that the infop argument points to a siginfo_t structure.

Right. So the application should do that, and Linux does the right thing. Problem solved.

Right. And this fact is why, while I incline to think we should fix the interface, I don't feel strongly about it.

Cheers,

Michael

Signature · 9 lines
Michael Kerrisk
Linux man-pages maintainer; http://www.kernel.org/doc/man-pages/
git://git.kernel.org/pub/scm/docs/man-pages/man-pages.git
man-pages online: http://www.kernel.org/doc/man-pages/online_pages.html
Found a bug? http://www.kernel.org/doc/man-pages/reporting_bugs.html
--
To unsubscribe from this list: send the line "unsubscribe linux-man" in
the body of a message to [email protected]
More majordomo info at  http://vger.kernel.org/majordomo-info.html
#15 lore raw

Re: [PATCH] sys_waitid: return -EFAULT for NULL

Quoted 2 lines · Michael Kerrisk

It isn't an extension. It's an accident.

Bah. It doesn't matter. It's an ABI. Your arguments make no sense - simply because there are other differences that means that if you develop under one OS, you can't expect it to run on another.

Arguing that "accident" is somehow different than "extension" is a totally inane and _idiotic_ argument. Because it's just arguing about the words, not about the end result.

If you're doing a dissertation in English literature, arguing about the words you use is valid. But if you're talking about operating systems, it's just pointless masturbation.

Linus
--
To unsubscribe from this list: send the line "unsubscribe linux-man" in the body of a message to [email protected] More majordomo info at http://vger.kernel.org/majordomo-info.html

#16 lore raw

Re: [PATCH] sys_waitid: return -EFAULT for NULL

Quoted 9 lines · Linus Torvalds
Quoted 2 lines · Michael Kerrisk

It isn't an extension. It's an accident.

Bah. It doesn't matter. It's an ABI. Your arguments make no sense - simply because there are other differences that means that if you develop under one OS, you can't expect it to run on another.

(I don't think implied such an expectation. On the other hand, we can do things to make more or less difficult for writers of portable applications.)

Quoted 7 lines

Arguing that "accident" is somehow different than "extension" is a totally inane and _idiotic_ argument. Because it's just arguing about the words, not about the end result.

If you're doing a dissertation in English literature, arguing about the words you use is valid. But if you're talking about operating systems, it's just pointless masturbation.

I agree that for practical purposes with respect to waitid(), "extension" versus "accident" is only about words: we've had the interface in its current form for so long that changing it shouldn't be done without consideration.

But (and this was why I constrasted "extension" with "accident"), I already made it clear that IMO there was another point that was more important than this specific case:

: And the fact that such accidents happen more often than necessary is : the real problem, rather than the fact that this API in particular is : inconsistent with expectations.

but you responded to the point I found less important.

Cheers,

Michael

Signature · 9 lines
Michael Kerrisk
Linux man-pages maintainer; http://www.kernel.org/doc/man-pages/
git://git.kernel.org/pub/scm/docs/man-pages/man-pages.git
man-pages online: http://www.kernel.org/doc/man-pages/online_pages.html
Found a bug? http://www.kernel.org/doc/man-pages/reporting_bugs.html
--
To unsubscribe from this list: send the line "unsubscribe linux-man" in
the body of a message to [email protected]
More majordomo info at  http://vger.kernel.org/majordomo-info.html

A message that is not in the archive (sent off-list or not yet received): cfd18e0f0901131436ud2cf61bkf0e6b00f4b765bb5-JsoAwUIsXosN+BqQ9rBEUg@public.gmane.org

#18 lore raw
Quoted 2 lines

a) Is it intentional to support the infop==NULL case for waitid() -- rather than simply giving an error return?

No. It's just an artifact of the internals shared with wait4/waitpid. Patch to follow.

Thanks,
Roland
--
To unsubscribe from this list: send the line "unsubscribe linux-man" in the body of a message to [email protected] More majordomo info at http://vger.kernel.org/majordomo-info.html