[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Date: Tue, 19 Jan 2021 12:35:26 -0800
From: Rasmus Andersson <rasmus@...ion.se>
To: musl@...ts.openwall.com
Subject: Re: waitpid (wait4) on Linux 5 returns invalid values
Relevant parts from disassembly: (objdump --disassemble --source
--line-numbers a.out)
0000000000401f61 <__syscall_cp_c>:
sccp():
401f61: 48 89 f8 mov %rdi,%rax
401f64: 4d 89 c2 mov %r8,%r10
401f67: 48 89 f7 mov %rsi,%rdi
401f6a: 4d 89 c8 mov %r9,%r8
401f6d: 48 89 d6 mov %rdx,%rsi
401f70: 4c 8b 4c 24 08 mov 0x8(%rsp),%r9
401f75: 48 89 ca mov %rcx,%rdx
401f78: 0f 05 syscall
401f7a: c3 retq
0000000000401f7b <__syscall_cp>:
__syscall_cp():
401f7b: e9 e1 ff ff ff jmpq 401f61 <__syscall_cp_c>
00000000004004fd <waitpid>:
waitpid():
...
40051a: e8 5c 1a 00 00 callq 401f7b <__syscall_cp>
40051f: 48 83 c4 18 add $0x18,%rsp
400523: c3 retq
main():
...
4002a2: e8 56 02 00 00 callq 4004fd <waitpid>
On Tue, Jan 19, 2021 at 12:28 PM Rasmus Andersson <rasmus@...ion.se> wrote:
>
> In the meantime, in case you're curious, here's a program that
> reproduces the issue. I've attached a statically-compiled version too
> that can be disassembled. Running it:
> $ gcc -static a.c && ./a.out
> waitpid returned -10, wstat=0
>
> #include <sys/wait.h>
> #include <stdio.h>
> int main(int argc, char **argv) {
> int wstat = 0;
> pid_t child = waitpid(-1, &wstat, WNOHANG);
> printf("waitpid returned %d, wstat=%d\n", child, wstat);
> return 0;
> }
>
> On Tue, Jan 19, 2021 at 12:17 PM Rasmus Andersson <rasmus@...ion.se> wrote:
> >
> > On Tue, Jan 19, 2021 at 11:33 AM Markus Wichmann <nullplan@....net> wrote:
> > >
> > > On Tue, Jan 19, 2021 at 10:18:04AM -0800, Rasmus Andersson wrote:
> > > > However the wait4 syscall[3] in Linux 5 returns other values,
> > > > specifically it returns errors as negative values. The error that
> > > > trips up programs like runit's runsv is ECHILD (-10) which wait4
> > > > returns when there are no children (i.e. they have exited.)
> > > >
> > >
> > > Hmm... that is very weird. That should not happen. Because the
> > > syscall_cp macro already contains a call to __syscall_ret(), which does
> > > exactly what you propose. So that means, there is something else going
> > > on in your copy of the code. Did you change anything about the source
> > > code? Or can you disassemble the function to see what it does?
> > >
> > Strange! You are right of course; looking at src/internal/syscall.h I
> > indeed see that syscall_cp calls __syscall_ret(__syscall_cp(args))
> > The musl I'm building with comes from https://musl.cc/#binaries which
> > is created from these scripts according to the author:
> > https://git.zv.io/xstatic/mcm It doesn't seem to apply any patches.
> > Anyhow, I'm currently building musl & gcc myself using
> > musl-cross-make. I will see how that pans out and report back.
> > Thank you for your quick response Markus!
Powered by blists - more mailing lists
Confused about mailing lists and their use?
Read about mailing lists on Wikipedia
and check out these
guidelines on proper formatting of your messages.
