This is the mail archive of the
gdb-patches@sourceware.org
mailing list for the GDB project.
Re: [ppc-linux-nat]: set access flag for h/w watchpoint even if it is only read or write
- From: Eli Zaretskii <eliz at gnu dot org>
- To: Wu Zhou <woodzltc at cn dot ibm dot com>, gdb-patches at sourceware dot org
- Date: Thu, 06 Jul 2006 23:58:58 +0300
- Subject: Re: [ppc-linux-nat]: set access flag for h/w watchpoint even if it is only read or write
- References: <Pine.LNX.4.64.0606092320510.5286@localhost.localdomain> <20060706132020.GB18827@nevyn.them.org>
- Reply-to: Eli Zaretskii <eliz at gnu dot org>
> Date: Thu, 6 Jul 2006 09:20:20 -0400
> From: Daniel Jacobowitz <drow@false.org>
> Cc: gdb-patches@sourceware.org
>
> On Fri, Jun 09, 2006 at 11:40:26PM +0800, Wu Zhou wrote:
> > Hello all,
> >
> > I found a bug in the current ppc-linux h/w watchpoint implementation:
> > when we set read watchpoint to some expression, if there are any write
> > operation to it before a read operation is hit, watchpoint_check will see
> > that its value is changed. So user won't see the watchpoint is hit.
> >
> > I make one change to the SET_DEBUGREG operation: even if it is only
> > read or write watchpoint, we still set access flag. Then, no matter
> > what operation is on the watched address, a SIGTRAP will be triggered.
> > The gdb code itself can determine if it is a write operation or read
> > operation. If it is write, watchpoint_check routine can update the
> > bs->value to the latest.
>
> Eli, you're the most familiar with watchpoint support; do you have any
> comment on this?
I'm sorry, I missed it somehow.
Yes, this problem is known on x86 and elsewhere. The problem is
extremely rare, as reading and writing to the same address in the same
instruction is a hard-to-accomplish treat. Wu, could you show a
real-life example of where this matters?
The solution you suggest, in my experience, is worse than the problem:
it will cause a significant slow-down of the most frequent case, as
Daniel points out.
> Many targets can tell us what sort of watchpoint was hit; we could
> do something like extend target_stopped_data_address to also return
> what sort of watchpoint was triggered, if it knows.
Yes, that would be a welcome addition, I agree.