[native x86 GNU/Linux] Access debug register mirror from the corresponding process.

While reviewing the native AArch64 patch, I noticed a problem:

On 02/06/2013 08:46 PM, Pedro Alves wrote:
>
>> > +static void
>> > +aarch64_linux_prepare_to_resume (struct lwp_info *lwp)
>> > +{
>> > +  struct arch_lwp_info *info = lwp->arch_private;
>> > +
>> > +  /* NULL means this is the main thread still going through the shell,
>> > +     or, no watchpoint has been set yet.  In that case, there's
>> > +     nothing to do.  */
>> > +  if (info == NULL)
>> > +    return;
>> > +
>> > +  if (DR_HAS_CHANGED (info->dr_changed_bp)
>> > +      || DR_HAS_CHANGED (info->dr_changed_wp))
>> > +    {
>> > +      int tid = GET_LWP (lwp->ptid);
>> > +      struct aarch64_debug_reg_state *state = aarch64_get_debug_reg_state ();
> Hmm.  This is always fetching the debug_reg_state of
> the current inferior, but may not be the inferior of lwp.
> I see the same bug on x86.  Sorry about that.  I'll fix it.

A natural fix would be to make xxx_get_debug_reg_state take an
inferior argument, but that doesn't work because of the case where we
detach breakpoints/watchpoints from the child fork, at a time there's
no inferior for the child fork at all.  We do a nasty hack in
i386_inferior_data_get, but that relies on all callers pointing the
current inferior to the correct inferior, which isn't actually being
done by all callers, and I don't think we want to enforce that -- deep
in the bowls of linux-nat.c, there are many cases we resume lwps
behind the scenes, and it's be better to not have that code rely on
global state (as it doesn't today).

The fix is to decouple the watchpoints code from inferiors, making it
track target processes instead.  This way, we can freely keep track of
the watchpoint mirrors for these processes behind the core's back.
Checkpoints also play dirty tricks with swapping the process behind
the inferior, so they get special treatment too in the patch (which
just amounts to calling a new hook).  Instead of the old hack in
i386_inferior_data_get, where we returned a copy of the current
inferior's debug registers mirror, as soon as we detect a fork in the
target, we copy the debug register mirror from the parent to the child
process.

I don't have an old kernel handy to test, but I stepped through gdb doing
the watchpoint removal in the fork child in the watchpoint-fork test
seeing that the debug registers end up cleared in the child.

I didn't find the need for linux_nat_iterate_watchpoint_lwps.  If
we use plain iterate_over_lwps instead, what happens is that
when removing watchpoints, that iterate_over_lwps doesn't actually
iterate over anything, since the fork child is not added to the
lwp list until later, at detach time, in linux_child_follow_fork.
And if we don't iterate over that lwp, we don't mark its debug
registers as needing update.  But linux_child_follow_fork takes
care of doing that explicitly:

	  child_lp = add_lwp (inferior_ptid);
	  child_lp->stopped = 1;
	  child_lp->last_resume_kind = resume_stop;
	  make_cleanup (delete_lwp_cleanup, child_lp);

	  /* CHILD_LP has new PID, therefore linux_nat_new_thread is not called for it.
	     See i386_inferior_data_get for the Linux kernel specifics.
	     Ensure linux_nat_prepare_to_resume will reset the hardware debug
	     registers.  It is done by the linux_nat_new_thread call, which is
	     being skipped in add_lwp above for the first lwp of a pid.  */
	  gdb_assert (num_lwps (GET_PID (child_lp->ptid)) == 1);
	  if (linux_nat_new_thread != NULL)
	    linux_nat_new_thread (child_lp);

	  if (linux_nat_prepare_to_resume != NULL)
	    linux_nat_prepare_to_resume (child_lp);
	  ptrace (PTRACE_DETACH, child_pid, 0, 0);

so unless I'm missing something (quite possible) it ends up all
the same.  But, the !detach-on-fork, and the "follow-fork child" paths
should also call linux_nat_new_thread, and they don't presently.  It
seems to me in those cases we're not clearing debug regs correctly
when that's needed.  Instead of copying that bit that works around
add_lwp bypassing the linux_nat_new_thread call, I thought it'd
be better to add an add_initial_lwp call to be used in the case we
really need to bypass linux_nat_new_thread, and make
add_lwp always call linux_nat_new_thread.

i386_cleanup_dregs is rewritten to forget about the current process
debug mirrors, which takes cares of other i386 ports.  Only a couple
of extra tweaks here and there were needed, as some targets wheren't
actually calling i386_cleanup_dregs.

Tested on Fedora 17 x86_64 -m64/-m32.

GDBserver already fetches the i386_debug_reg_state from the right
process, and, it doesn't handle forks at all, so no fix is needed over
there.

gdb/
2013-02-13  Pedro Alves  <palves@redhat.com>

	* amd64-linux-nat.c (update_debug_registers_callback):
	Update comment.
	(amd64_linux_dr_set_control, amd64_linux_dr_set_addr): Use
	iterate_over_lwps.
	(amd64_linux_prepare_to_resume): Pass the lwp's pid to
	i386_debug_reg_state.
	(amd64_linux_new_fork): New function.
	(_initialize_amd64_linux_nat): Install amd64_linux_new_fork as
	linux_nat_new_fork hook, and i386_forget_process as
	linux_nat_forget_process hook.
	* i386-linux-nat.c (update_debug_registers_callback):
	Update comment.
	(amd64_linux_dr_set_control, amd64_linux_dr_set_addr): Use
	iterate_over_lwps.
	(i386_linux_prepare_to_resume): Pass the lwp's pid to
	i386_debug_reg_state.
	(i386_linux_new_fork): New function.
	(_initialize_i386_linux_nat): Install i386_linux_new_fork as
	linux_nat_new_fork hook, and i386_forget_process as
	linux_nat_forget_process hook.
	* i386-nat.c (i386_init_dregs): Delete.
	(i386_inferior_data, struct i386_inferior_data):
	Delete.
	(struct i386_process_info): New.
	(i386_process_list): New global.
	(i386_find_process_pid, i386_add_process, i386_process_info_get):
	New functions.
	(i386_inferior_data_get): Delete.
	(i386_process_info_get): New function.
	(i386_debug_reg_state): New parameter 'pid'.  Reimplement.
	(i386_forget_process): New function.
	(i386_cleanup_dregs): Rewrite.
	(i386_update_inferior_debug_regs, i386_insert_watchpoint)
	(i386_remove_watchpoint, i386_region_ok_for_watchpoint)
	(i386_stopped_data_address, i386_insert_hw_breakpoint)
	(i386_remove_hw_breakpoint): Adjust to pass the current process id
	to i386_debug_reg_state.
	(i386_use_watchpoints): Don't register inferior data.
	* i386-nat.h (i386_debug_reg_state): Add new 'pid' parameter, and
	adjust comment.
	(i386_forget_process): Declare.
	* linux-fork.c (delete_fork): Call linux_nat_forget_process.
	* linux-nat.c (linux_nat_new_fork, linux_nat_forget_process_hook):
	New static globals.
	(linux_child_follow_fork): Don't call linux_nat_new_thread here.
	(add_initial_lwp): New, factored out from ...
	(add_lwp): ... this.  Don't check the number of lwps before
	calling linux_nat_new_thread.
	(linux_nat_iterate_watchpoint_lwps): Delete.
	(linux_nat_attach): Use add_initial_lwp instead of add_lwp.
	(linux_handle_extended_wait): Call the linux_nat_new_fork hook on
	forks and vforks.
	(linux_nat_wait_1): Use add_initial_lwp instead of add_lwp for the
	initial lwp.
	(linux_nat_kill, linux_nat_mourn_inferior): Call
	linux_nat_forget_process.
	(linux_nat_set_new_fork, linux_nat_set_forget_process)
	(linux_nat_forget_process): New functions.
	* linux-nat.h (linux_nat_iterate_watchpoint_lwps_ftype): Delete
	type.
	(linux_nat_iterate_watchpoint_lwps): Delete declaration.
	(linux_nat_new_fork_ftype, linux_nat_forget_process_ftype): New
	types.
	(linux_nat_set_new_fork, linux_nat_set_forget_process)
	(linux_nat_forget_process): New declarations.

	* amd64fbsd-nat.c (super_mourn_inferior): New global.
	(amd64fbsd_mourn_inferior): New function.
	(_initialize_amd64fbsd_nat): Override to_mourn_inferior.
	* windows-nat.c (windows_detach): Call i386_cleanup_dregs.

1 files changed, 45 insertions, 5 deletions

diff --git a/gdb/i386-linux-nat.c b/gdb/i386-linux-nat.c
index 20cc032780..11b510dc49 100644
--- a/gdb/i386-linux-nat.c
+++ b/gdb/i386-linux-nat.c
@@ -708,8 +708,8 @@ i386_linux_dr_get_status (void)
   return i386_linux_dr_get (inferior_ptid, DR_STATUS);
 }
 
-/* Callback for linux_nat_iterate_watchpoint_lwps.  Update the debug registers
-   of LWP.  */
+/* Callback for iterate_over_lwps.  Update the debug registers of
+   LWP.  */
 
 static int
 update_debug_registers_callback (struct lwp_info *lwp, void *arg)
@@ -735,7 +735,9 @@ update_debug_registers_callback (struct lwp_info *lwp, void *arg)
 static void
 i386_linux_dr_set_control (unsigned long control)
 {
-  linux_nat_iterate_watchpoint_lwps (update_debug_registers_callback, NULL);
+  ptid_t pid_ptid = pid_to_ptid (ptid_get_pid (inferior_ptid));
+
+  iterate_over_lwps (pid_ptid, update_debug_registers_callback, NULL);
 }
 
 /* Set address REGNUM (zero based) to ADDR in all LWPs of the current
@@ -748,7 +750,7 @@ i386_linux_dr_set_addr (int regnum, CORE_ADDR addr)
 
   gdb_assert (regnum >= 0 && regnum <= DR_LASTADDR - DR_FIRSTADDR);
 
-  linux_nat_iterate_watchpoint_lwps (update_debug_registers_callback, NULL);
+  iterate_over_lwps (pid_ptid, update_debug_registers_callback, NULL);
 }
 
 /* Called when resuming a thread.
@@ -767,7 +769,8 @@ i386_linux_prepare_to_resume (struct lwp_info *lwp)
 
   if (lwp->arch_private->debug_registers_changed)
     {
-      struct i386_debug_reg_state *state = i386_debug_reg_state ();
+      struct i386_debug_reg_state *state
+	= i386_debug_reg_state (ptid_get_pid (lwp->ptid));
       int i;
 
       /* See amd64_linux_prepare_to_resume for Linux kernel note on
@@ -803,6 +806,41 @@ i386_linux_new_thread (struct lwp_info *lp)
 
   lp->arch_private = info;
 }
+
+/* linux_nat_new_fork hook.   */
+
+static void
+i386_linux_new_fork (struct lwp_info *parent, pid_t child_pid)
+{
+  pid_t parent_pid;
+  struct i386_debug_reg_state *parent_state;
+  struct i386_debug_reg_state *child_state;
+
+  /* NULL means no watchpoint has ever been set in the parent.  In
+     that case, there's nothing to do.  */
+  if (parent->arch_private == NULL)
+    return;
+
+  /* Linux kernel before 2.6.33 commit
+     72f674d203cd230426437cdcf7dd6f681dad8b0d
+     will inherit hardware debug registers from parent
+     on fork/vfork/clone.  Newer Linux kernels create such tasks with
+     zeroed debug registers.
+
+     GDB core assumes the child inherits the watchpoints/hw
+     breakpoints of the parent, and will remove them all from the
+     forked off process.  Copy the debug registers mirrors into the
+     new process so that all breakpoints and watchpoints can be
+     removed together.  The debug registers mirror will become zeroed
+     in the end before detaching the forked off process, thus making
+     this compatible with older Linux kernels too.  */
+
+  parent_pid = ptid_get_pid (parent->ptid);
+  parent_state = i386_debug_reg_state (parent_pid);
+  child_state = i386_debug_reg_state (child_pid);
+  *child_state = *parent_state;
+}
+
 
 
 /* Called by libthread_db.  Returns a pointer to the thread local
@@ -1044,5 +1082,7 @@ _initialize_i386_linux_nat (void)
   /* Register the target.  */
   linux_nat_add_target (t);
   linux_nat_set_new_thread (t, i386_linux_new_thread);
+  linux_nat_set_new_fork (t, i386_linux_new_fork);
+  linux_nat_set_forget_process (t, i386_forget_process);
   linux_nat_set_prepare_to_resume (t, i386_linux_prepare_to_resume);
 }