On Sat, 10 Aug 2024, Michael Schmitz wrote:
Anyway, if I run dump under strace I see no CLONE_INTO_CGROUP flag:
strace may not be aware of the CLONE_INTO_CGROUP flag yet? How old is
your strace binary?
I don't think strace is the problem. If it was, we should still see all
the flags in the disassembly, in the constant passed to the syscall.
clone(child_stack=NULL, flags=CLONE_IO|SIGCHLD) = -1 EBADF (Bad file
descriptor)
The -EBADF result was introduced into cgroup_css_set_fork() by the
commit above. That should not happen unless CLONE_INTO_CGROUP was set,
but strace says its not. So I don't know what's going on here.
Here's what gdb says, FWIW...
# gdb
GNU gdb (Debian 13.1-3) 13.1
...
(gdb) file /usr/sbin/dump
Reading symbols from /usr/sbin/dump...
Reading symbols from
/usr/lib/debug/.build-id/24/071a827207bee9c025d364137514447279302b.debug...
(gdb) run -0f /dev/null /dev/sda
Starting program: /usr/sbin/dump -0f /dev/null /dev/sda
DUMP: Date of this level 0 dump: Fri Aug 9 23:37:15 2024
DUMP: Dumping /dev/sda (an unlisted file system) to /dev/null
DUMP: Label: none
DUMP: Writing 10 Kilobyte records
DUMP: mapping (Pass I) [regular files]
DUMP: mapping (Pass II) [directories]
DUMP: estimated 3595695 blocks.
DUMP: Context save fork fails in parent 671
[Inferior 1 (process 671) exited with code 03]
(gdb) b fork_clone_io
Breakpoint 1 at 0x80009dbc: file tape.c, line 740.
(gdb) run -0f /dev/null /dev/sda
Starting program: /usr/sbin/dump -0f /dev/null /dev/sda
DUMP: Date of this level 0 dump: Fri Aug 9 23:38:17 2024
DUMP: Dumping /dev/sda (an unlisted file system) to /dev/null
DUMP: Label: none
DUMP: Writing 10 Kilobyte records
DUMP: mapping (Pass I) [regular files]
DUMP: mapping (Pass II) [directories]
DUMP: estimated 3595695 blocks.
Program received signal SIGSEGV, Segmentation fault.
0x00000001 in ?? ()
(gdb) l fork_clone_io
warning: Source file is more recent than executable.
735
736 #ifdef __linux__
737 #if defined(SYS_clone) && defined(CLONE_IO)
738 pid_t
739 fork_clone_io(void)
740 {
741 pid_t res,parent;
742 parent=getppid(); /* az hackety hack... */
743
744 res=syscall(SYS_clone, CLONE_ARGS);
745 getppid();
746 /* as per clone call manpage: caching! */
747 getpid();
748 #ifdef __alpha__
749 syscall(SYS_getxpid);
750 #else
751 syscall(SYS_getpid);
752 #endif
753
754 /* az: clone manpage doesn't say jack about what the
(gdb) disas fork_clone_io
Dump of assembler code for function fork_clone_io:
0x80009dbc <+0>: movel %d3,%sp@-
0x80009dbe <+2>: movel %d2,%sp@-
0x80009dc0 <+4>: bsrl 0x80004200 <getppid@plt>
0x80009dc6 <+10>: movel %d0,%d3
0x80009dc8 <+12>: clrl %sp@-
0x80009dca <+14>: clrl %sp@-
0x80009dcc <+16>: clrl %sp@-
0x80009dce <+18>: movel #-2147483631,%sp@-
0x80009dd4 <+24>: pea 0x78
0x80009dd8 <+28>: bsrl 0x80003fd0 <syscall@plt>
0x80009dde <+34>: movel %d0,%d2
0x80009de0 <+36>: bsrl 0x80004200 <getppid@plt>
0x80009de6 <+42>: bsrl 0x80003c9c <getpid@plt>
0x80009dec <+48>: pea 0x14
0x80009df0 <+52>: bsrl 0x80003fd0 <syscall@plt>
0x80009df6 <+58>: bsrl 0x80004200 <getppid@plt>
0x80009dfc <+64>: lea %sp@(24),%sp
0x80009e00 <+68>: cmpl %d0,%d3
0x80009e02 <+70>: beqs 0x80009e06 <fork_clone_io+74>
0x80009e04 <+72>: clrl %d2
0x80009e06 <+74>: movel %d2,%d0
0x80009e08 <+76>: movel %sp@+,%d2
0x80009e0a <+78>: movel %sp@+,%d3
0x80009e0c <+80>: rts
End of assembler dump.
(gdb)
Is this clone syscall (0x78) really executing sys_clone3()? Also,
Nope, syscall no. 120 calls __sys_clone() which in turn calls
m68k_clone() which emulates sys_clone() (roundabout way due to different
calling conventions on m68k).
clone3 is syscall 435 (calling __sys_clone3() -> m68k_clone3() ->
sys_clone3()).
What confused me was that 'git bisect' fingered what looked like a clone3
patch, but it turns out that this patch affects anything that calls
cgroup_can_fork(), that is, any syscalls that call copy_process().
But as long as syscall() takes care of the calling convention, I see no
reason why that way of calling sys_clone() would fail.
The interesting thing about the calling convention is that the flags make
up a 32-bit quantity when passed to clone as an int, and a 64-bit quantity
when passed to clone3 as struct clone_args.flags.
So I've just added some printk() statements and found that m68k_clone()
messed up the flags in the kernel_clone_args struct: I'm seeing
0xFFFFFFFF80000000 which explains how CLONE_INTO_CGROUP got set.
I'll send a patch.
