Re: 2.4.7pre8aa1

[Andrea Arcangeli <andrea@suse.de>]

On Fri, Jul 20, 2001 at 03:37:49AM +0200, Andrea Arcangeli wrote:
> On Thu, Jul 19, 2001 at 06:45:38PM -0400, Jeff Dike wrote:
> > > Only in 2.4.7pre6aa1: 51_uml-ac-to-aa-2.bz2
> > > Only in 2.4.7pre8aa1/: 51_uml-ac-to-aa-3.bz2
> > >         Moved part of it in the tux directory so it can compile
> > >         without tux (in reality I got errno compilation error
> > >         but it's low prio and I'll sort it out later, Jeff Dike any
> > >         hint is welcome ;).
> > 
> > This is the patch I sent to Alan a while back which works around the problem.
> > 
> > rmk suggested a better way which I'll add at some point.
> > 
> > 				Jeff
> > 
> > 
> > diff -Naur -X exclude-files ac_cur/arch/um/Makefile ac/arch/um/Makefile
> > --- ac_cur/arch/um/Makefile	Mon Jul  9 13:05:03 2001
> > +++ ac/arch/um/Makefile	Mon Jul  9 13:26:21 2001
> > @@ -20,6 +20,8 @@
> >  LINK_PROFILE = $(PROFILE) -Wl,--wrap,__monstartup
> >  endif
> >  
> > +CFLAGS := $(subst -fno-common,,$(CFLAGS))
> > +
> >  SUBDIRS += $(ARCH_DIR)/fs $(ARCH_DIR)/drivers $(ARCH_DIR)/kernel \
> >  	$(ARCH_DIR)/sys-$(SUBARCH)
> 
> works fine thanks! Of course I agree with rmk it would be better not to

I should have said "compiles" fine (not "works" fine :).

__initdata is broken in uml and the kernel deadlocks because the wait
list is empty in complete() despite wait_for_completion actually
registered correctly into it. This because wait_for_completion runs in
a different address space than complete() and the virtual memory is not
shared across the two address spaces (it is not rempped in a MAP_SHARED
so it generates a cow). The registration is basically only private to
the entity that is registrating and it will never get visible to the
waker task that will do nothing. This is a severe bug not just for the
completion code in 2.4.7. here the fix (plus the ptrace bit):

--- ./arch/um/kernel/ptrace.c.~1~	Fri Jul 20 17:23:16 2001
+++ ./arch/um/kernel/ptrace.c	Fri Jul 20 17:43:56 2001
@@ -43,7 +43,7 @@
 	if (request == PTRACE_ATTACH) {
 		if (child == current)
 			goto out_tsk;
-		if ((!child->dumpable ||
+		if ((!ptrace_dumpable(child) ||
 		     (current->uid != child->euid) ||
 		     (current->uid != child->suid) ||
 		     (current->uid != child->uid) ||
--- ./arch/um/link.ld.in.~1~	Fri Jul 20 17:23:16 2001
+++ ./arch/um/link.ld.in	Mon Jul 23 03:06:40 2001
@@ -45,7 +45,6 @@
   . = ALIGN(4096);		/* Init code and data */
   __init_begin = .;
   .text.init : { *(.text.init) }
-  .data.init : { *(.data.init) }
   . = ALIGN(16);
   __setup_start = .;
   .setup.init : { *(.setup.init) }
@@ -92,6 +91,7 @@
     . = ALIGN(16384);		/* init_task */
     *(.data.init_task)
     *(.data)
+    *(.data.init)
     *(.gnu.linkonce.d*)
     CONSTRUCTORS
   }


(I didn't update the init_begin/end logic since there's no logic at all
in uml because we don't free that memory anyways so there's no side
effect because of that change)

> disable -fno-common but this is ok for now ;) (after all we would catch
> any potential important name collision during the compiles of the other
> targets)
> 
> Andrea


Andrea