* Question about _dl_debug_state and new glibc
@ 2004-10-11 18:54 Randolph Chung
2004-10-11 19:03 ` Kevin Buettner
0 siblings, 1 reply; 6+ messages in thread
From: Randolph Chung @ 2004-10-11 18:54 UTC (permalink / raw)
To: gdb
Hi all,
I've been away from gdb for a bit, so hopefully this is not something
simple I missed while I've been away :)
It appears that newer versions of glibc now marks the _dl_debug_state
function as hidden, making it not visible to gdb. This breaks the
solib tracking code in solib-svr4.c.....
Looking at the docs, it looks like the "proper" way for gdb to do this
is to look up the r_debug symbol and use the r_brk member to locate
_dl_debug_state's address. is there any particular reason why we don't
do this in gdb?
randolph
--
Randolph Chung
Debian GNU/Linux Developer, hppa/ia64 ports
http://www.tausq.org/
^ permalink raw reply [flat|nested] 6+ messages in thread
* Re: Question about _dl_debug_state and new glibc
2004-10-11 18:54 Question about _dl_debug_state and new glibc Randolph Chung
@ 2004-10-11 19:03 ` Kevin Buettner
2004-10-11 19:14 ` Daniel Jacobowitz
0 siblings, 1 reply; 6+ messages in thread
From: Kevin Buettner @ 2004-10-11 19:03 UTC (permalink / raw)
To: Randolph Chung; +Cc: gdb
On Mon, 11 Oct 2004 10:22:07 -0700
Randolph Chung <randolph@tausq.org> wrote:
> I've been away from gdb for a bit, so hopefully this is not something
> simple I missed while I've been away :)
>
> It appears that newer versions of glibc now marks the _dl_debug_state
> function as hidden, making it not visible to gdb. This breaks the
> solib tracking code in solib-svr4.c.....
>
> Looking at the docs, it looks like the "proper" way for gdb to do this
> is to look up the r_debug symbol and use the r_brk member to locate
> _dl_debug_state's address. is there any particular reason why we don't
> do this in gdb?
I can't think of any reason. I'm guessing that it was more expedient
for the original authors of the code to do it the way it was done.
In any case, I know of no reason not to change it so that it's done
"properly".
Patches are welcome...
Kevin
^ permalink raw reply [flat|nested] 6+ messages in thread
* Re: Question about _dl_debug_state and new glibc
2004-10-11 19:03 ` Kevin Buettner
@ 2004-10-11 19:14 ` Daniel Jacobowitz
2004-10-11 20:47 ` Mark Kettenis
2004-10-11 22:00 ` Kevin Buettner
0 siblings, 2 replies; 6+ messages in thread
From: Daniel Jacobowitz @ 2004-10-11 19:14 UTC (permalink / raw)
To: Kevin Buettner; +Cc: Randolph Chung, gdb
On Mon, Oct 11, 2004 at 10:46:47AM -0700, Kevin Buettner wrote:
> On Mon, 11 Oct 2004 10:22:07 -0700
> Randolph Chung <randolph@tausq.org> wrote:
>
> > I've been away from gdb for a bit, so hopefully this is not something
> > simple I missed while I've been away :)
> >
> > It appears that newer versions of glibc now marks the _dl_debug_state
> > function as hidden, making it not visible to gdb. This breaks the
> > solib tracking code in solib-svr4.c.....
> >
> > Looking at the docs, it looks like the "proper" way for gdb to do this
> > is to look up the r_debug symbol and use the r_brk member to locate
> > _dl_debug_state's address. is there any particular reason why we don't
> > do this in gdb?
Only for static executables usually; for dynamic executables it's
supposed to be DT_DEBUG if that's available. Same difference.
> I can't think of any reason. I'm guessing that it was more expedient
> for the original authors of the code to do it the way it was done.
> In any case, I know of no reason not to change it so that it's done
> "properly".
We set the breakpoint on _dl_debug_state before starting the inferior.
_dl_debug_initialize initializes _r_debug somewhat later... I'm not
sure how it is supposed to be used. Do either of you know?
I see it hasn't been exported since March. Blech.
--
Daniel Jacobowitz
^ permalink raw reply [flat|nested] 6+ messages in thread
* Re: Question about _dl_debug_state and new glibc
2004-10-11 19:14 ` Daniel Jacobowitz
@ 2004-10-11 20:47 ` Mark Kettenis
2004-10-12 6:57 ` Daniel Jacobowitz
2004-10-11 22:00 ` Kevin Buettner
1 sibling, 1 reply; 6+ messages in thread
From: Mark Kettenis @ 2004-10-11 20:47 UTC (permalink / raw)
To: drow; +Cc: kevinb, randolph, gdb, libc-alpha
Date: Mon, 11 Oct 2004 13:55:25 -0400
From: Daniel Jacobowitz <drow@false.org>
On Mon, Oct 11, 2004 at 10:46:47AM -0700, Kevin Buettner wrote:
> On Mon, 11 Oct 2004 10:22:07 -0700
> Randolph Chung <randolph@tausq.org> wrote:
>
> > I've been away from gdb for a bit, so hopefully this is not something
> > simple I missed while I've been away :)
> >
> > It appears that newer versions of glibc now marks the _dl_debug_state
> > function as hidden, making it not visible to gdb. This breaks the
> > solib tracking code in solib-svr4.c.....
> >
> > Looking at the docs, it looks like the "proper" way for gdb to do this
> > is to look up the r_debug symbol and use the r_brk member to locate
> > _dl_debug_state's address. is there any particular reason why we don't
> > do this in gdb?
Only for static executables usually; for dynamic executables it's
supposed to be DT_DEBUG if that's available. Same difference.
Hey wake up Daniel. Static executables don't have shared libraries.
> I can't think of any reason. I'm guessing that it was more expedient
> for the original authors of the code to do it the way it was done.
> In any case, I know of no reason not to change it so that it's done
> "properly".
We set the breakpoint on _dl_debug_state before starting the inferior.
_dl_debug_initialize initializes _r_debug somewhat later... I'm not
sure how it is supposed to be used. Do either of you know?
The problem is that even DT_DEBUG isn't initialized before
initialization of the dynamic linker. Theoretically, we could put a
watchpoint on DT_DEBUG, and lookup r_brk if it triggers. However, I
don't think that's very practical on platforms without hardware
watchpoint support.
I suppose this is just something that has never been completely
implemented. Somehow I think the dynamic linker was supposed to trap
just after initializing r_debug and setting DT_DEBUG if it was being
traced, perhaps if it noticed that DT_DEBUG was set to some special
value by the debugger. Perhaps we'll know when "Open Solaris" is
released.
I see it hasn't been exported since March. Blech.
So we should ask the glibc developers to unhide _dl_debug_state.
Mark
^ permalink raw reply [flat|nested] 6+ messages in thread
* Re: Question about _dl_debug_state and new glibc
2004-10-11 19:14 ` Daniel Jacobowitz
2004-10-11 20:47 ` Mark Kettenis
@ 2004-10-11 22:00 ` Kevin Buettner
1 sibling, 0 replies; 6+ messages in thread
From: Kevin Buettner @ 2004-10-11 22:00 UTC (permalink / raw)
To: Daniel Jacobowitz; +Cc: Randolph Chung, gdb
On Mon, 11 Oct 2004 13:55:25 -0400
Daniel Jacobowitz <drow@false.org> wrote:
> On Mon, Oct 11, 2004 at 10:46:47AM -0700, Kevin Buettner wrote:
>
> > I can't think of any reason. I'm guessing that it was more expedient
> > for the original authors of the code to do it the way it was done.
> > In any case, I know of no reason not to change it so that it's done
> > "properly".
>
> We set the breakpoint on _dl_debug_state before starting the inferior.
> _dl_debug_initialize initializes _r_debug somewhat later... I'm not
> sure how it is supposed to be used. Do either of you know?
Ah! Perhaps that's the reason why _dl_debug_state was used directly.
I ran into this problem when I wrote the FR-V's solib code. (Due to
the fact that text and data are independetly relocated, the code in
solib-svr4.c couldn't be used.) The symbol in question was not
exported and could only be accessed from the r_debug struct. But as you
say, it's value wasn't immediately available. So, we set the solib
event breakpoint on the executable's entry point (probably _start).
Once this breakpoint was hit, the data structures in question had been
initialized, and so we reset it to the address determined by examining
the r_debug struct.
I suspect that a similar strategy could be used for solib-svr4.c.
Kevin
^ permalink raw reply [flat|nested] 6+ messages in thread
* Re: Question about _dl_debug_state and new glibc
2004-10-11 20:47 ` Mark Kettenis
@ 2004-10-12 6:57 ` Daniel Jacobowitz
0 siblings, 0 replies; 6+ messages in thread
From: Daniel Jacobowitz @ 2004-10-12 6:57 UTC (permalink / raw)
To: Mark Kettenis; +Cc: kevinb, randolph, gdb, libc-alpha
On Mon, Oct 11, 2004 at 09:13:41PM +0200, Mark Kettenis wrote:
> Date: Mon, 11 Oct 2004 13:55:25 -0400
> From: Daniel Jacobowitz <drow@false.org>
>
> On Mon, Oct 11, 2004 at 10:46:47AM -0700, Kevin Buettner wrote:
> > On Mon, 11 Oct 2004 10:22:07 -0700
> > Randolph Chung <randolph@tausq.org> wrote:
> >
> > > I've been away from gdb for a bit, so hopefully this is not something
> > > simple I missed while I've been away :)
> > >
> > > It appears that newer versions of glibc now marks the _dl_debug_state
> > > function as hidden, making it not visible to gdb. This breaks the
> > > solib tracking code in solib-svr4.c.....
> > >
> > > Looking at the docs, it looks like the "proper" way for gdb to do this
> > > is to look up the r_debug symbol and use the r_brk member to locate
> > > _dl_debug_state's address. is there any particular reason why we don't
> > > do this in gdb?
>
> Only for static executables usually; for dynamic executables it's
> supposed to be DT_DEBUG if that's available. Same difference.
>
> Hey wake up Daniel. Static executables don't have shared libraries.
Oddly enough, in glibc they can - and often do. There's a static
version of the runtime linker included, and things like dlopen work -
and internal libc functionality like NSS (gethostname, etc.) and gconv
use it.
Also, there are architectures where DT_DEBUG is missing. I think MIPS
may be one. Or maybe I'm misremembering this bit.
> I suppose this is just something that has never been completely
> implemented. Somehow I think the dynamic linker was supposed to trap
> just after initializing r_debug and setting DT_DEBUG if it was being
> traced, perhaps if it noticed that DT_DEBUG was set to some special
> value by the debugger. Perhaps we'll know when "Open Solaris" is
> released.
Makes sense to me. Do any of the BSDs support using r_debug this way,
or do they export _dl_debug_state also?
> I see it hasn't been exported since March. Blech.
>
> So we should ask the glibc developers to unhide _dl_debug_state.
I agree.
--
Daniel Jacobowitz
^ permalink raw reply [flat|nested] 6+ messages in thread
end of thread, other threads:[~2004-10-11 19:21 UTC | newest]
Thread overview: 6+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2004-10-11 18:54 Question about _dl_debug_state and new glibc Randolph Chung
2004-10-11 19:03 ` Kevin Buettner
2004-10-11 19:14 ` Daniel Jacobowitz
2004-10-11 20:47 ` Mark Kettenis
2004-10-12 6:57 ` Daniel Jacobowitz
2004-10-11 22:00 ` Kevin Buettner
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).