[Bug dynamic-link/28937] New DSO dependency sorter does not put new map first if in a cycle

["fweimer at redhat dot com" <sourceware-bugzilla@sourceware.org>]

https://sourceware.org/bugzilla/show_bug.cgi?id=28937

Florian Weimer <fweimer at redhat dot com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
            Summary|New DSO dependency resolver |New DSO dependency sorter
                   |buggyness with dlclose()    |does not put new map first
                   |                            |if in a cycle
           Assignee|unassigned at sourceware dot org   |fweimer at redhat dot com
              Flags|                            |security-
             Status|WAITING                     |ASSIGNED

--- Comment #9 from Florian Weimer <fweimer at redhat dot com> ---
I could get it to boot in UEFI mode, but the disk has to be attached as sata
(not virtio). Thank you for putting together this reproducer, it is really
helpful.

The situation matches the one described in comment 0. There is no dlclose call
for libfreetype.so.6, so libbz2.so.1.0 really can't been unloaded.

A comment in _dl_close_worker alludes to the root cause:

          /* We are always the zeroth entry, and since we don't include
             ourselves in the dependency analysis start at 1.  */
          struct link_map **lp = &l->l_initfini[1];

The assumption is not correct in this case, l_initfini[0] is actually
libbz2.so.1.0 here. The reason is that the new algorithm does not implement
skipping, and due to the cyclic dependency between libfreetype.so.6 and
libharfbuzz.so.0, libfreetype.so.6 is not sorted first naturally.

I think it makes sense to always initialize the newly-loaded object first, but
as the comment in _dl_sort_maps_dfs explains, that algorithm cannot skip the
dependencies directly. But I think we should adjsut _dl_sort_maps_dfs because
the initialization order property seems useful. I'll try to come up with a
patch.

-- 
You are receiving this mail because:
You are on the CC list for the bug.