gdb 6.3 misses breakpoint on Linux when inferior does clone()

gdb 6.3 misses breakpoint on Linux when inferior does clone()

[Satish Mohan]

Hi
 If the inferior does a clone(...,CLONE_VM,...) on Linux, then gdb deletes breakpoints in the child which indirectly affects the parent because of CLONE_VM. The following sample code illustrates the problem:

int mythread (void * unused) 
{
   printf("my thread running\n");
   return 0;
}

int main (int argc, char **argv) 
{
   int status, pid;
   char * stack = (char *)malloc(4096);
   if ((pid = clone (mythread, (char *)(stack + 4096 - 4), 
                     CLONE_VM | SIGCHLD, 0)) == -1)
   {
       printf("clone failed, errno=%d\n", errno); 
       return -1;
   }

   if (waitpid(pid, &status, 0) == -1) 
   {
       perror("waitpid failed\n");
       return -1;
   }
}

If a breakpoint is set on the waitpid() call above, then gdb misses it. This seems to be because gdb deletes breakpoints in the child after the clone() but doesn't seem to know that CLONE_VM is set, and that the breakpoints will also get deleted from the parent. 

Commenting the detach_breakpoints(child_pid) in child_follow_fork() in linux-nat.c fixes the problem as long as parent and child don't execute the same code after clone. 

Is this a gdb bug ? Could anyone let me know the right fix for this ?

Thanks,
Satish Mohan.

Re: gdb 6.3 misses breakpoint on Linux when inferior does clone()

[Daniel Jacobowitz]

On Thu, May 26, 2005 at 12:40:39PM -0700, Satish Mohan wrote:
> Hi  If the inferior does a clone(...,CLONE_VM,...) on Linux, then gdb
> deletes breakpoints in the child which indirectly affects the parent
> because of CLONE_VM. The following sample code illustrates the
> problem:

The kernel does not provide any interface for GDB to know whether or
not a clone syscall has used CLONE_VM.  Until it does, there's very
little that GDB can do about this.  It assumes that clone using SIGCHLD
is not sharing a VM space, and clone using a different signal is.

-- 
Daniel Jacobowitz
CodeSourcery, LLC

Re: gdb 6.3 misses breakpoint on Linux when inferior does clone()

[Daniel Jacobowitz]

On Sun, Jun 12, 2005 at 01:31:38AM +0200, Mark Kettenis wrote:
>    Date: Wed, 1 Jun 2005 09:23:31 -0400
>    From: Daniel Jacobowitz <drow@false.org>
> 
>    On Tue, May 31, 2005 at 10:05:22AM -0700, tronics93-gdb@yahoo.com wrote:
>    > The following code snippet gets the flags passed in to clone() (on
>    > x86):
>    > 
>    >    unsigned long regs[FRAME_SIZE];
>    > 
>    >    if (ptrace(PTRACE_GETREGS, pid, 0, &regs) < 0)
>    >         perror("ptrace_getregs failed:");
>    > 
>    >    printf("syscall=%d, arg1=%#x, arg2=%#x\n", regs[ORIG_EAX], 
>    >                               regs[EBX], regs[ECX]);
> 
>    Only if you're stopped at the clone.  Hmm.  Good point; perhaps we can
>    recover them from registers when we receive the clone event.
> 
> But having to do that in a machine-dependent manner really sucks...
> 
>    It could probably be implemented if someone spent the time.
> 
> ...so that time is probably better spent on implementing the proper
> machine-independent functionality in the Linux kernel!

Except that what we implement in GDB, we don't need to test for at
runtime and handle dependencies on the kernel version.  For the special
case of CLONE_VM there's a third option, which may be superior: write
to one thread's memory and see where the value shows up.

-- 
Daniel Jacobowitz
CodeSourcery, LLC

Re: gdb 6.3 misses breakpoint on Linux when inferior does clone()

[Mark Kettenis]

   Date: Wed, 1 Jun 2005 09:23:31 -0400
   From: Daniel Jacobowitz <drow@false.org>

   On Tue, May 31, 2005 at 10:05:22AM -0700, tronics93-gdb@yahoo.com wrote:
   > The following code snippet gets the flags passed in to clone() (on
   > x86):
   > 
   >    unsigned long regs[FRAME_SIZE];
   > 
   >    if (ptrace(PTRACE_GETREGS, pid, 0, &regs) < 0)
   >         perror("ptrace_getregs failed:");
   > 
   >    printf("syscall=%d, arg1=%#x, arg2=%#x\n", regs[ORIG_EAX], 
   >                               regs[EBX], regs[ECX]);

   Only if you're stopped at the clone.  Hmm.  Good point; perhaps we can
   recover them from registers when we receive the clone event.

But having to do that in a machine-dependent manner really sucks...

   It could probably be implemented if someone spent the time.

...so that time is probably better spent on implementing the proper
machine-independent functionality in the Linux kernel!

Mark

Re: gdb 6.3 misses breakpoint on Linux when inferior does clone()

[Daniel Jacobowitz]

On Tue, May 31, 2005 at 10:05:22AM -0700, tronics93-gdb@yahoo.com wrote:
> >> The kernel does not provide any interface for GDB to know whether
> or
> >> not a clone syscall has used CLONE_VM.   
> >> 
> >> [Satish] Doesn't ptrace(PTRACE_GETREGS,...) provide that
> information >>? Or is that arch. specific ?
> 
> >No, it does not.
> 
> The following code snippet gets the flags passed in to clone() (on
> x86):
> 
>    unsigned long regs[FRAME_SIZE];
> 
>    if (ptrace(PTRACE_GETREGS, pid, 0, &regs) < 0)
>         perror("ptrace_getregs failed:");
> 
>    printf("syscall=%d, arg1=%#x, arg2=%#x\n", regs[ORIG_EAX], 
>                               regs[EBX], regs[ECX]);

Only if you're stopped at the clone.  Hmm.  Good point; perhaps we can
recover them from registers when we receive the clone event.

It could probably be implemented if someone spent the time.

-- 
Daniel Jacobowitz
CodeSourcery, LLC

Re: gdb 6.3 misses breakpoint on Linux when inferior does clone()

[tronics93-gdb]

>> The kernel does not provide any interface for GDB to know whether
or
>> not a clone syscall has used CLONE_VM.   
>> 
>> [Satish] Doesn't ptrace(PTRACE_GETREGS,...) provide that
information >>? Or is that arch. specific ?

>No, it does not.

The following code snippet gets the flags passed in to clone() (on
x86):

   unsigned long regs[FRAME_SIZE];

   if (ptrace(PTRACE_GETREGS, pid, 0, &regs) < 0)
        perror("ptrace_getregs failed:");

   printf("syscall=%d, arg1=%#x, arg2=%#x\n", regs[ORIG_EAX], 
                              regs[EBX], regs[ECX]);

--tronics93.

Re: gdb 6.3 misses breakpoint on Linux when inferior does clone()

[Daniel Jacobowitz]

On Thu, May 26, 2005 at 01:02:28PM -0700, Satish Mohan wrote:
> 
> 
> -----Original Message-----
> From: Daniel Jacobowitz [mailto:drow@false.org] 
> Sent: Thursday, May 26, 2005 12:44 PM
> To: Satish Mohan
> Cc: gdb@sources.redhat.com
> Subject: Re: gdb 6.3 misses breakpoint on Linux when inferior does clone()
> 
> On Thu, May 26, 2005 at 12:40:39PM -0700, Satish Mohan wrote:
> > Hi  If the inferior does a clone(...,CLONE_VM,...) on Linux, then gdb
> > deletes breakpoints in the child which indirectly affects the parent
> > because of CLONE_VM. The following sample code illustrates the
> > problem:
> 
> The kernel does not provide any interface for GDB to know whether or
> not a clone syscall has used CLONE_VM.   
> 
> [Satish] Doesn't ptrace(PTRACE_GETREGS,...) provide that information ? Or is that arch. specific ?

No, it does not.

-- 
Daniel Jacobowitz
CodeSourcery, LLC

RE: gdb 6.3 misses breakpoint on Linux when inferior does clone()

[Satish Mohan]

-----Original Message-----
From: Daniel Jacobowitz [mailto:drow@false.org] 
Sent: Thursday, May 26, 2005 12:44 PM
To: Satish Mohan
Cc: gdb@sources.redhat.com
Subject: Re: gdb 6.3 misses breakpoint on Linux when inferior does clone()

On Thu, May 26, 2005 at 12:40:39PM -0700, Satish Mohan wrote:
> Hi  If the inferior does a clone(...,CLONE_VM,...) on Linux, then gdb
> deletes breakpoints in the child which indirectly affects the parent
> because of CLONE_VM. The following sample code illustrates the
> problem:

The kernel does not provide any interface for GDB to know whether or
not a clone syscall has used CLONE_VM.   

[Satish] Doesn't ptrace(PTRACE_GETREGS,...) provide that information ? Or is that arch. specific ?

Thx
--satish