Bug#1042980: gnome-shell: FTBFS on mips64el, mipsel: perf-* tests fail

[YunQiang Su]

YunQiang Su <wzssyqa at gmail.com> 于2023年8月4日周五 19:40写道：
>
> Simon McVittie <smcv at debian.org> 于2023年8月4日周五 19:26写道：
> >
> > Control: forwarded -1 https://gitlab.gnome.org/GNOME/gnome-shell/-/issues/6877
> >
> > On Fri, 04 Aug 2023 at 15:50:32 +0800, YunQiang Su wrote:
> > > 156   if (rpath)
> > > 157     paths = g_strsplit (strtab + rpath->d_un.d_val, ":", -1);
> > > 158   else
> > > 159     paths = g_strsplit (strtab + runpath->d_un.d_val, ":", -1);
> > >      // <- segfault here due to
> >
> > Thanks, that's very useful information.
> >
> > After some printf debugging on eller, I think what is happening here is
> > that GNOME Shell is assuming that the d_un.d_val of the DT_STRTAB is
> > an absolute pointer, but on mips64el for whatever reason it's only an
> > offset, and it needs to be added to the base address of the executable
> > to get a real pointer.
> >
> > If I understand correctly, the value in the header in the binary on-disk
> > is just an offset, and on most architectures the dynamic linker overwrites
> > the offset with (base_address + offset) during loading/relocation - but
> > for whatever reason, mips64el isn't doing that.
> >
>
> Yes. It is this case. I guess it should be a glibc's bug, while during years,
> it is a feature now (;
> We cannot change this behaviour now.
>
> I am working on a patch for it, the real base address can be get by:
>     ((getauxval(AT_PHDR) >> 8) <<8)
>
> > Instead of doing this low-level ELF manipulation, I'm testing a patch which
> > uses an environment variable to propagate the search path into the
> > executable - that seems less likely to go wrong on unusual architectures.
> >
> >     smcv
>

FYI: this patch can fix the segfault problem in
maybe_add_rpath_introspection_paths

--- gnome-shell-44.3.orig/src/main.c
+++ gnome-shell-44.3/src/main.c
@@ -7,6 +7,7 @@
 #endif
 #include <stdlib.h>
 #include <string.h>
+#include <sys/auxv.h>

 #include <cogl-pango/cogl-pango.h>
 #include <clutter/clutter.h>
@@ -129,6 +130,7 @@ shell_dbus_init (gboolean replace)
 }

 #ifdef HAVE_EXE_INTROSPECTION
+__attribute__((optimize("O0")))
 static void
 maybe_add_rpath_introspection_paths (void)
 {
@@ -150,6 +152,15 @@ maybe_add_rpath_introspection_paths (voi
         strtab = (const char *) dyn->d_un.d_val;
     }

+#if defined(__mips)
+  /* The d_val of DT_STRTAB, contains the offset of .dynstr
+   * and the start of elf is loaded, instead of absolute address.
+   * getauxval(AT_PHDR) is near enough with the start, and we are sure
+   * that the elf will be loaded page-aligned.
+   */
+  strtab = strtab + ((getauxval(AT_PHDR) >> 8) << 8);
+#endif
+
   if ((!rpath && !runpath) || !strtab)
     return;


I am continue working on (EE) failed to write to Xwayland fd: Broken
pipe problem.

>
>
> --
> YunQiang Su



-- 
YunQiang Su