Re: [PATCH] Fix read and access watchpoint can't stop

[Hev <r@hev.cc>]

Hello,

On Thu, Jun 18, 2020 at 12:38 AM Eli Zaretskii <eliz@gnu.org> wrote:
>
> > From: Heiher <r@hev.cc>
> > Date: Thu, 18 Jun 2020 00:17:08 +0800
> > Cc: Heiher <r@hev.cc>
> >
> > --- a/gdb/breakpoint.c
> > +++ b/gdb/breakpoint.c
> > @@ -4952,7 +4952,7 @@ bpstat_check_watchpoint (bpstat bs)
> >          this watchpoint.  */
> >       must_check_value = 1;
> >        else if (b->watchpoint_triggered == watch_triggered_unknown
> > -            && b->type == bp_hardware_watchpoint)
> > +            && is_hardware_watchpoint (bs->breakpoint_at))
> >       /* We were stopped by a hardware watchpoint, but the target could
> >          not report the data address.  We must check the watchpoint's
> >          value.  Access and read watchpoints are out of luck; without
>
> The comment below the line you suggest to modify explains why we don't
> handle access and read watchpoints here.  So I don't understand the
> rationale for this change (it might be a good idea to explain it in
> more detail in the log message).  What am I missing?
>
> Thanks.

This problem only occurs on the target that target_stopped_data_address()
returns false, and b->watchpoint_triggered is set to
watch_triggered_unknown when watchpoint is triggered. so, In
bpstat_check_watchpoint(), the last branch needs to be hit to enable check
value:

 4961       int must_check_value = 0;
 4962
 4963       if (b->type == bp_watchpoint)
 4964     /* For a software watchpoint, we must always check the
 4965        watched value.  */
 4966     must_check_value = 1;
 4967       else if (b->watchpoint_triggered == watch_triggered_yes)
 4968     /* We have a hardware watchpoint (read, write, or access)
 4969        and the target earlier reported an address watched by
 4970        this watchpoint.  */
 4971     must_check_value = 1;
 4972       else if (b->watchpoint_triggered == watch_triggered_unknown
 4973            && b->type == bp_hardware_watchpoint)
 4974     /* We were stopped by a hardware watchpoint, but the target could
 4975        not report the data address.  We must check the watchpoint's
 4976        value.  Access and read watchpoints are out of luck; without
 4977        a data address, we can't figure it out.  */
 4978     must_check_value = 1;

However, the b->type is the actual watchpoint type, bp_hardware_watchpoint
for write-only, bp_read_watchpoint for read-only, and bp_access_watchpoint
for read-write, so it will not hit when read-only and read-write watching.

On mips we don't know the low order 3 bits of the data address, so we must
return false in stopped_data_address(). This is why we have not found the
problem on other platforms (x86/Aarch64/...).

-- 
Best regards!
Hev
https://hev.cc