Re: [PATCH] [gdb/build] Fix frame_list position in frame.c

[Simon Marchi <simon.marchi@polymtl.ca>]

On 5/3/23 13:58, Tom de Vries wrote:
> In commit 995a34b1772 ("Guard against frame.c destructors running before
> frame-info.c's") the following problem was addressed.
> 
> The frame_info_ptr destructor:
> ...
>   ~frame_info_ptr ()
>   {
>     frame_list.erase (frame_list.iterator_to (*this));
>   }
> ...
> uses frame_list, which is a static member of class frame_info_ptr,
> instantiated in frame-info.c:
> ...
> intrusive_list<frame_info_ptr> frame_info_ptr::frame_list;
> ...
> 
> Then there's a static frame_info_pointer variable named selected_frame in
> frame.c:
> ...
> static frame_info_ptr selected_frame;
> ...
> 
> Because the destructor of selected_frame uses frame_list, its destructor needs
> to be called before the destructor of frame_list.
> 
> But because they're in different compilation units, the initialization order and
> consequently destruction order is not guarantueed.
> 
> The commit fixed this by handling the case that the destructor of frame_list
> is called first, adding a check on is_linked ():
> ...
>    ~frame_info_ptr ()
>    {
> -    frame_list.erase (frame_list.iterator_to (*this));
> +    /* If this node has static storage, it may be deleted after
> +       frame_list.  Attempting to erase ourselves would then trigger
> +       internal errors, so make sure we are still linked first.  */
> +    if (is_linked ())
> +      frame_list.erase (frame_list.iterator_to (*this));
>    }
> ...
> 
> However, since then frame_list has been moved into frame.c, and
> initialization/destruction order is guarantueed inside a compilation unit.
> 
> Revert aforementioned commit, and fix the destruction order problem by moving
> frame_list before selected_frame.
> 
> Reverting the commit is another way of fixing the already fixed
> Wdangling-pointer warning reported in PR build/30413, in a different way than
> commit 9b0ccb1ebae ("Pass const frame_info_ptr reference for
> skip_[language_]trampoline").
> 
> Tested on x86_64-linux.
> 
> PR build/30413
> Bug: https://sourceware.org/bugzilla/show_bug.cgi?id=30413
> ---
>  gdb/frame.c | 11 +++++++----
>  gdb/frame.h |  9 ++++-----
>  2 files changed, 11 insertions(+), 9 deletions(-)
> 
> diff --git a/gdb/frame.c b/gdb/frame.c
> index 36fb02f3c8e..531eadf3d54 100644
> --- a/gdb/frame.c
> +++ b/gdb/frame.c
> @@ -1733,6 +1733,13 @@ get_current_frame (void)
>  static frame_id selected_frame_id = null_frame_id;
>  static int selected_frame_level = -1;
>  
> +/* See frame.h.  This definition should come before any definition of a static
> +   frame_info_ptr, to ensure that frame_list is destroyed after any static
> +   frame_info_ptr.  This is  necessary because the destructor of frame_info_ptr

Spurious double space.

> +   uses frame_list.  */
> +
> +intrusive_list<frame_info_ptr> frame_info_ptr::frame_list;
> +
>  /* The cached frame_info object pointing to the selected frame.
>     Looked up on demand by get_selected_frame.  */
>  static frame_info_ptr selected_frame;
> @@ -3275,10 +3282,6 @@ maintenance_print_frame_id (const char *args, int from_tty)
>  
>  /* See frame-info-ptr.h.  */
>  
> -intrusive_list<frame_info_ptr> frame_info_ptr::frame_list;
> -
> -/* See frame-info-ptr.h.  */
> -
>  frame_info_ptr::frame_info_ptr (struct frame_info *ptr)
>    : m_ptr (ptr)
>  {
> diff --git a/gdb/frame.h b/gdb/frame.h
> index 6ed8db0af56..ed19dfdc090 100644
> --- a/gdb/frame.h
> +++ b/gdb/frame.h
> @@ -254,11 +254,10 @@ class frame_info_ptr : public intrusive_list_node<frame_info_ptr>
>  
>    ~frame_info_ptr ()
>    {
> -    /* If this node has static storage, it may be deleted after
> -       frame_list.  Attempting to erase ourselves would then trigger
> -       internal errors, so make sure we are still linked first.  */
> -    if (is_linked ())
> -      frame_list.erase (frame_list.iterator_to (*this));
> +    /* If this node has static storage, it should be be deleted before
> +       frame_list.  Verify this by checking that it is still in the list.  */
> +    gdb_assert (is_linked ());
> +    frame_list.erase (frame_list.iterator_to (*this));

The assert is a bit redundant with the assertions in
intrusive_list::erase_element:

    gdb_assert (elem_node->prev != INTRUSIVE_LIST_UNLINKED_VALUE);
    gdb_assert (elem_node->next != INTRUSIVE_LIST_UNLINKED_VALUE);

I would maybe remove the assert, but keep the comment (at least the
first sentence)?

In any case, this LGTM, thanks for doing this.

Approved-By: Simon Marchi <simon.marchi@efficios.com>

Simon