This is the mail archive of the
gdb-patches@sourceware.org
mailing list for the GDB project.
Re: eliminate deprecated_insert_raw_breakpoint. what's left.
- From: Peter Schauer <peterschauer at gmx dot net>
- To: uweigand at de dot ibm dot com (Ulrich Weigand)
- Cc: brobecker at adacore dot com (Joel Brobecker), palves at redhat dot com (Pedro Alves), gdb-patches at sourceware dot org (GDB Patches)
- Date: Tue, 9 Sep 2014 14:37:49 +0200 (CEST)
- Subject: Re: eliminate deprecated_insert_raw_breakpoint. what's left.
- Authentication-results: sourceware.org; auth=none
> Peter Schauer wrote:
>
> > I hope to be able to shed some light on this problem, although it
> > is more than fifteen years ago that I did some work for GDB on AIX.
> >
> > From my notes back then, AIX 3 and AIX 4 had a very peculiar ptrace
> > implementation, where the current ptrace state of the inferior process
> > (including the current process registers) was maintained approximately
> > 512 bytes below the current user stack pointer of the process.
> >
> > This resulted in problems with AIX inferior function calls.
> > If the called function takes one or more large aggregate parameters
> > by value, or if you pass a large amount of parameters, the ptrace
> > area gets corrupted, when the dummy function call parameters are
> > pushed on the user stack, due to this awkward AIX stack layout.
>
> Thanks for providing this background!
>
> > To work around this problem, the execution of a dummy instruction
> > (when altering the stack pointer) caused the kernel to move the ptrace
> > state area further below on the user stack, allowing GDB to write below
> > the current user stack safely.
> > In GDB 6.x, rs6000_push_dummy_call even secured the stack partially during
> > pushing of the arguments, via an additional call of
> > regcache_raw_write_signed to gdbarch_sp_regnum (gdbarch), which is
> > no longer present in current versions of GDB.
>
> Well, I still see this:
> /* Set the stack pointer. According to the ABI, the SP is meant to
> be set _before_ the corresponding stack space is used. On AIX,
> this even applies when the target has been completely stopped!
> Not doing this can lead to conflicts with the kernel which thinks
> that it still has control over this not-yet-allocated stack
> region. */
> regcache_raw_write_signed (regcache, gdbarch_sp_regnum (gdbarch), sp);
>
> and:
> /* This is another instance we need to be concerned about
> securing our stack space. If we write anything underneath %sp
> (r1), we might conflict with the kernel who thinks he is free
> to use this area. So, update %sp first before doing anything
> else. */
>
> regcache_raw_write_signed (regcache,
> gdbarch_sp_regnum (gdbarch), sp);
>
> Are there other instances where this is missing?
Ok, my bad, I was looking at the wrong push_dummy_call implementation
in the current GDB source.
rs6000_push_dummy_call in the new rs6000-aix-tdep.c file in the current GDB
source still contains the code in question from GDB 6.x, there is
nothing missing.
> > Executing the dummy instruction is very fragile, especially if signals
> > get involved during the execution, and it didn't even help, if more
> > than ~100 bytes of parameters were pushed on the user stack on AIX 4.
> > Back then, there was no other choice though.
> >
> > Unfortunately I do not know, if this peculiar AIX stack layout is still
> > used in AIX 5 or later, maybe Ulrich Weigand could tell you more about it.
>
> I don't know off-hand. I'll try to find out.
>
> > I think you could/should zap exec_one_dummy_insn, provided that you test
> > a dummy function call on the oldest AIX version that GDB has to support,
> > with a large aggregate parameter, which is passed by value.
>
> The only version I have ready access to is AIX 7.1, and on this there
> are no testsuite regression (and in fact, quite a number of failures
> seem to go away!) when zapping exec_one_dummy_insn.
+1 for zapping exec_one_dummy_insn.
> I'm not sure which versions we need to / should support in GDB; I guess
> the oldest version where the OS itself is still supported by IBM is 6.1.
Maybe somebody could test if zapping exec_one_dummy_insn on AIX 6.1
has any negative effect, and then be done with it.
But even if that can't be tested, I am all in favour of getting rid
of it, perhaps with a detailed comment in the commit message for the
removal (or adding a link to this thread).
--
Peter Schauer Peter.Schauer@mytum.de