Re: [PATCH] Support for HWbreak/watchpoint accross fork/vfork on arm-native

[Pedro Alves <palves at redhat dot com>]

On 01/30/2014 10:23 PM, Omair Javaid wrote:
> Accidentally missed the patch description and changelog:
> 
> This patch updates arm native support for hardware breakpoints and watchpoints
> to accommodate watchpoint/breakpoints across fork/vfork. Previously for arm
> native a record of breakpoints at thread level was being kept which has been
> changed to a process level record to come in sync with gdb-linux calls to
> threads and process creation/destruction hooks.

Thanks.

This also changes the hwbreak/watchpoint insertion mechanism
to the modern way of marking debug registers as needing update, but
only really updating on resume, which is necessary for supporting
watchpoints in non-stop mode (the current code tries to poke at
running threads with ptrace, which fails).  Please update the
commit log to mention this.

The patch looks overall mostly good, but I have a few comments
below.

> ---
>  gdb/arm-linux-nat.c | 417 ++++++++++++++++++++++++++++++++++------------------
>  1 file changed, 273 insertions(+), 144 deletions(-)
> 
> diff --git a/gdb/arm-linux-nat.c b/gdb/arm-linux-nat.c
> index 6f56634..9a756c1 100644
> --- a/gdb/arm-linux-nat.c
> +++ b/gdb/arm-linux-nat.c
> @@ -787,70 +787,120 @@ struct arm_linux_hw_breakpoint
>    arm_hwbp_control_t control;
>  };
>  
> -/* Structure containing arrays of the break and watch points which are have
> -   active in each thread.
> -
> -   The Linux ptrace interface to hardware break-/watch-points presents the 
> -   values in a vector centred around 0 (which is used fo generic information).
> -   Positive indicies refer to breakpoint addresses/control registers, negative
> -   indices to watchpoint addresses/control registers.
> -
> -   The Linux vector is indexed as follows:
> -      -((i << 1) + 2): Control register for watchpoint i.
> -      -((i << 1) + 1): Address register for watchpoint i.
> -                    0: Information register.
> -       ((i << 1) + 1): Address register for breakpoint i.
> -       ((i << 1) + 2): Control register for breakpoint i.

Why remove this ptrace comment?  It looks quite useful to me.

> -
> -   This structure is used as a per-thread cache of the state stored by the 
> -   kernel, so that we don't need to keep calling into the kernel to find a 
> -   free breakpoint.
> -
> -   We treat break-/watch-points with their enable bit clear as being deleted.
> -   */
> -typedef struct arm_linux_thread_points
> +/* Since we cannot dynamically allocate subfields of arm_linux_process_info,
> +   assume a maximum number of supported break-/watchpoints.  */

> +#define MAX_BPTS 16
> +#define MAX_WPTS 16

I saw no assertion making sure these maximums aren't
overflowed.  see aarch64_linux_get_debug_reg_capacity.

> +  /* Linked list.  */
> +  struct arm_linux_process_info *next;
> +  /* The process identifier.  */
> +  pid_t pid;
> +  /* Hardware breakpoints for this process.  */
> +  struct arm_linux_hw_breakpoint bpts[MAX_BPTS];
> +  /* Hardware watchpoints for this process.  */
> +  struct arm_linux_hw_breakpoint wpts[MAX_WPTS];

It'd be a little cleaner if these two array were moved
to a separate structure, like e.g., in the aarch64 port.

> +/* Get hardware breakpoint state for process PID.  */
>  
> -  return t;
> +static struct arm_linux_hw_breakpoint *
> +arm_linux_get_hwbreak_state (pid_t pid)
> +{
> +  return arm_linux_process_info_get (pid)->bpts;
> +}
> +
> +/* Get watchpoint state for process PID.  */
> +
> +static struct arm_linux_hw_breakpoint *
> +arm_linux_get_watch_state (pid_t pid)
> +{
> +  return arm_linux_process_info_get (pid)->wpts;
>  }

Then these two could be a single:

static struct arm_linux_hw_breakpoint *
arm_linux_get_point_state (pid_t pid)
{
  return &arm_linux_process_info_get (pid)->state;
}

Users could then do:
 arm_linux_get_point_state (pid)->wpts;
 arm_linux_get_point_state (pid)->bpts;


> +static int
> +update_registers_callback (struct lwp_info *lwp, void *arg)
> +{
> +  struct update_registers_data *data = (struct update_registers_data *) arg;
> +
> +  /* Force iterate_over_lwps to return matched lwp_info*.  */
> +  if (arg == NULL)
> +    return 1;

I don't understand this.  It seems nothing passes a NULL arg.

> +
> +  if (lwp->arch_private == NULL)
> +    lwp->arch_private = XCNEW (struct arch_lwp_info);
> +
> +  /* The actual update is done later just before resuming the lwp,
> +     we just mark that the registers need updating.  */
> +  if (data->watch)
> +    lwp->arch_private->wpts_changed[data->index] = 1;
> +  else
> +    lwp->arch_private->bpts_changed[data->index] = 1;
> +
> +  /* If the lwp isn't stopped, force it to momentarily pause, so
> +     we can update its breakpoint registers.  */
> +  if (!lwp->stopped)
> +    linux_stop_lwp (lwp);
> +
> +  return 0;
> +}

>    if (watchpoint)
>      {
>        count = arm_linux_get_hw_watchpoint_count ();
> -      bpts = t->wpts;
> -      dir = -1;
> +      bpts = arm_linux_get_watch_state(pid);

Space before parens.

>      }
>    else
>      {
>        count = arm_linux_get_hw_breakpoint_count ();
> -      bpts = t->bpts;
> -      dir = 1;
> +      bpts = arm_linux_get_hwbreak_state(pid);

Ditto.


>      }

>    if (watchpoint)
>      {
>        count = arm_linux_get_hw_watchpoint_count ();
> -      bpts = t->wpts;
> -      dir = -1;
> +      bpts = arm_linux_get_watch_state(pid);

Ditto.

>      }
>    else
>      {
>        count = arm_linux_get_hw_breakpoint_count ();
> -      bpts = t->bpts;
> -      dir = 1;
> +      bpts = arm_linux_get_hwbreak_state(pid);

Ditto.

>      }
>  


> +arm_linux_prepare_to_resume (struct lwp_info *lwp)
>  {
...
> +  for (i = 0; i < arm_linux_get_hw_breakpoint_count (); i++)
> +    if (arm_lwp_info->bpts_changed[i])
> +      {
> +        errno = 0;
> +        pid = ptid_get_lwp (lwp->ptid);
> +        bpts = arm_linux_get_hwbreak_state (ptid_get_pid (lwp->ptid));

This should probably be moved out of the loop.  As is it's doing
a linear lookup per iteration.

> -	}
> +        if (arm_hwbp_control_is_enabled (bpts[i].control))
> +          if (ptrace (PTRACE_SETHBPREGS, pid,
> +              (PTRACE_TYPE_ARG3) ((i << 1) + 1), &bpts[i].address) < 0)
> +              perror_with_name ("Unexpected error setting breakpoint address");

Missing _() for i18n.  The old code had it.

>  
> -      if (t == NULL)
> -	return;
> +        if (bpts[i].control != 0)
> +          if (ptrace (PTRACE_SETHBPREGS, pid,
> +              (PTRACE_TYPE_ARG3) ((i << 1) + 2), &bpts[i].control) < 0)
> +              perror_with_name ("Unexpected error setting breakpoint");

Ditto.

>  
> -      VEC_unordered_remove (arm_linux_thread_points_p, arm_threads, i);
> +        arm_lwp_info->bpts_changed[i] = 0;
> +      }
>  
> -      xfree (t->bpts);
> -      xfree (t->wpts);
> -      xfree (t);
> -    }
> +  for (i = 0; i < arm_linux_get_hw_watchpoint_count (); i++)
> +    if (arm_lwp_info->wpts_changed[i])
> +      {
> +        errno = 0;
> +        pid = ptid_get_lwp (lwp->ptid);
> +        wpts = arm_linux_get_watch_state (ptid_get_pid (lwp->ptid));

Likewise as comment above.

> +
> +        if (arm_hwbp_control_is_enabled (wpts[i].control))
> +          if (ptrace (PTRACE_SETHBPREGS, pid,
> +              (PTRACE_TYPE_ARG3) -((i << 1) + 1), &wpts[i].address) < 0)
> +              perror_with_name ("Unexpected error setting watchpoint address");

Ditto, etc.

> +static void
> +arm_linux_new_fork (struct lwp_info *parent, pid_t child_pid)
> +{


> +  parent_pid = ptid_get_pid (parent->ptid);
> +  parent_point_state = arm_linux_get_hwbreak_state (parent_pid);
> +  child_point_state = arm_linux_get_hwbreak_state (child_pid);
> +  memcpy (child_point_state, parent_point_state,
> +          sizeof (struct arm_linux_hw_breakpoint) * MAX_BPTS);
> +
> +  parent_point_state = arm_linux_get_watch_state (parent_pid);
> +  child_point_state = arm_linux_get_watch_state (child_pid);
> +  memcpy (child_point_state, parent_point_state,
> +          sizeof (struct arm_linux_hw_breakpoint) * MAX_BPTS);

The latter should be MAX_WPTS.  But with the earlier proposed change,
these two would simply turn into a single:

  parent_pid = ptid_get_pid (parent->ptid);
  parent_point_state = arm_linux_get_point_state (parent_pid);
  child_point_state = arm_linux_get_point_state (child_pid);
  *child_point_state = *parent_point_state;