Re: qemu couldn't connect the unix domain socket

[Li Feng <fengli@smartx.com>]

[-- Attachment #1: Type: text/plain, Size: 4713 bytes --]

On Fri, Oct 29, 2021 at 7:52 PM Stefano Brivio <sbrivio(a)redhat.com> wrote:
>
> On Fri, 29 Oct 2021 19:02:15 +0800
> Li Feng <fengli(a)smartx.com> wrote:
>
> > On Fri, Oct 29, 2021 at 5:34 PM Stefano Brivio <sbrivio(a)redhat.com> wrote:
> > >
> > > On Fri, 29 Oct 2021 16:54:47 +0800
> > > Li Feng <fengli(a)smartx.com> wrote:
> > >
> > > > [...]
> > > >
> > > > Thanks for the detailed explanation.
> > > > I finally found out that the `qrap` was the root cause.
> > > > I patched the qemu, and it works well.
> > >
> > > Have you found out what was the offending syscall? I'll probably hit
> > > this later too, but that would help me double checking what the problem
> > > was.
> >
> > I made a mistake in the previous mail, using `./passt -f` works, but
> > if running in background,
> > it still exits without any output.
> >
> > This is the strace output.
> >
> > ```
> > $ strace -f ./passt
> > ...
> > ...
> > accept(6, NULL, NULL)                   = 7
> > epoll_ctl(5, EPOLL_CTL_ADD, 7, {events=EPOLLIN|EPOLLRDHUP|EPOLLET,
> > data={u32=7, u64=7}}) = 0
> > getrandom("\x40\xfc\xc5\x4a\x29\x3e\xdb\xcd\x25\x92\xc6\xc3\xc7\xcb\x57\x5a",
> > 16, GRND_RANDOM) = 16
> > socket(AF_INET, SOCK_STREAM|SOCK_NONBLOCK, IPPROTO_TCP) = 8
> > socket(AF_INET, SOCK_STREAM|SOCK_NONBLOCK, IPPROTO_TCP) = 9
> > socket(AF_INET, SOCK_STREAM|SOCK_NONBLOCK, IPPROTO_TCP) = 10
> > socket(AF_INET, SOCK_STREAM|SOCK_NONBLOCK, IPPROTO_TCP) = 11
> > socket(AF_INET, SOCK_STREAM|SOCK_NONBLOCK, IPPROTO_TCP) = 12
> > socket(AF_INET, SOCK_STREAM|SOCK_NONBLOCK, IPPROTO_TCP) = 13
> > socket(AF_INET, SOCK_STREAM|SOCK_NONBLOCK, IPPROTO_TCP) = 14
> > socket(AF_INET, SOCK_STREAM|SOCK_NONBLOCK, IPPROTO_TCP) = 15
> > socket(AF_INET, SOCK_STREAM|SOCK_NONBLOCK, IPPROTO_TCP) = 16
> > socket(AF_INET, SOCK_STREAM|SOCK_NONBLOCK, IPPROTO_TCP) = 17
> > socket(AF_INET, SOCK_STREAM|SOCK_NONBLOCK, IPPROTO_TCP) = 18
> > socket(AF_INET, SOCK_STREAM|SOCK_NONBLOCK, IPPROTO_TCP) = 19
> > socket(AF_INET, SOCK_STREAM|SOCK_NONBLOCK, IPPROTO_TCP) = 20
> > socket(AF_INET, SOCK_STREAM|SOCK_NONBLOCK, IPPROTO_TCP) = 21
> > socket(AF_INET, SOCK_STREAM|SOCK_NONBLOCK, IPPROTO_TCP) = 22
> > socket(AF_INET, SOCK_STREAM|SOCK_NONBLOCK, IPPROTO_TCP) = 23
> > socket(AF_INET, SOCK_STREAM|SOCK_NONBLOCK, IPPROTO_TCP) = 24
> > socket(AF_INET, SOCK_STREAM|SOCK_NONBLOCK, IPPROTO_TCP) = 25
> > socket(AF_INET, SOCK_STREAM|SOCK_NONBLOCK, IPPROTO_TCP) = 26
> > socket(AF_INET, SOCK_STREAM|SOCK_NONBLOCK, IPPROTO_TCP) = 27
> > socket(AF_INET, SOCK_STREAM|SOCK_NONBLOCK, IPPROTO_TCP) = 28
> > socket(AF_INET, SOCK_STREAM|SOCK_NONBLOCK, IPPROTO_TCP) = 29
> > socket(AF_INET, SOCK_STREAM|SOCK_NONBLOCK, IPPROTO_TCP) = 30
> > socket(AF_INET, SOCK_STREAM|SOCK_NONBLOCK, IPPROTO_TCP) = 31
> > socket(AF_INET, SOCK_STREAM|SOCK_NONBLOCK, IPPROTO_TCP) = 32
> > socket(AF_INET, SOCK_STREAM|SOCK_NONBLOCK, IPPROTO_TCP) = 33
> > socket(AF_INET, SOCK_STREAM|SOCK_NONBLOCK, IPPROTO_TCP) = 34
> > socket(AF_INET, SOCK_STREAM|SOCK_NONBLOCK, IPPROTO_TCP) = 35
> > socket(AF_INET, SOCK_STREAM|SOCK_NONBLOCK, IPPROTO_TCP) = 36
> > socket(AF_INET, SOCK_STREAM|SOCK_NONBLOCK, IPPROTO_TCP) = 37
> > socket(AF_INET, SOCK_STREAM|SOCK_NONBLOCK, IPPROTO_TCP) = 38
> > socket(AF_INET, SOCK_STREAM|SOCK_NONBLOCK, IPPROTO_TCP) = 39
> > clone(child_stack=NULL,
> > flags=CLONE_CHILD_CLEARTID|CLONE_CHILD_SETTID|SIGCHLDstrace: Process
> > 172939 attached
> > , child_tidptr=0x7f7b89b1da10) = 172939
> > [pid 172921] exit_group(0 <unfinished ...>
> > [pid 172939] set_robust_list(0x7f7b89b1da20, 24 <unfinished ...>
> > [pid 172921] <... exit_group resumed>)  = ?
> > [pid 172921] +++ exited with 0 +++
> > <... set_robust_list resumed>)          = ?
> > +++ killed by SIGSYS (core dumped) +++
> > ```
> > Which is the bad syscall?
>
> Oh, it's set_robust_list(), it's normal that exit_group() doesn't
> return. That new usage probably comes from:
>         https://sourceware.org/git/?p=glibc.git;a=commit;h=9a7565403758f65c07fe3705e966381d9cfd35b6
>
> and that code path is not really needed for passt, so I would have a
> quick try at avoiding it rather than adding a syscall, perhaps with a
> small replacement of daemon() using clone() instead of fork().
>
> Meanwhile, this should work for you:
>
> diff --git a/passt.c b/passt.c
> index 6436a45..2a4ba8b 100644
> --- a/passt.c
> +++ b/passt.c
> @@ -280,3 +280,3 @@ static void pid_file(struct ctx *c) {
>   * #syscalls openat fstat fcntl lseek clone setsid exit_group getpid
> - * #syscalls clock_gettime newfstatat
> + * #syscalls clock_gettime newfstatat set_robust_list
>   * #syscalls:pasta rt_sigreturn
>
> --
> Stefano
>
It works!