Re: [PATCH v2] Fix hardware watchpoints in replay mode

[Guinevere Larsen <blarsen@redhat.com>]

On 14/09/2023 18:27, Hannes Domani via Gdb-patches wrote:
> Changes introduced by commit 9e8915c6cee5c37637521b424d723e990e06d597
> caused a regression that meant hardware watchpoint stops would not
> trigger in reverse execution or replay mode.  This was documented in
> PR breakpoints/21969.
> The problem is that record_check_stopped_by_breakpoint always overwrites
> record_full_stop_reason, thus loosing the TARGET_STOPPED_BY_WATCHPOINT
> value which would be checked afterwards.
>
> This commit fixes that by checking if record_full_stop_reason is
> TARGET_STOPPED_BY_BREAKPOINT and, if so, not overriding it.
>
> Bug: https://sourceware.org/bugzilla/show_bug.cgi?id=21969
> ---

Hi! Thanks for the quick turn around for this patch!

>   gdb/record-full.c                           | 4 +++-
>   gdb/testsuite/gdb.reverse/watch-reverse.exp | 4 ++++
>   2 files changed, 7 insertions(+), 1 deletion(-)
>
> diff --git a/gdb/record-full.c b/gdb/record-full.c
> index faf8b595d22..da785ef4b2a 100644
> --- a/gdb/record-full.c
> +++ b/gdb/record-full.c
> @@ -1382,7 +1382,9 @@ record_full_wait_1 (struct target_ops *ops,
>   
>   		      /* check breakpoint */
>   		      tmp_pc = regcache_read_pc (regcache);
> -		      if (record_check_stopped_by_breakpoint
> +		      if (record_full_stop_reason
> +			  != TARGET_STOPPED_BY_WATCHPOINT
> +			  && record_check_stopped_by_breakpoint

So I did some more digging on this, based on your response to the v1 
email, copied here for convenience:

     I just tried this, with this change it suddenly shows SIGTRAP instead of
     hit watchpoints, so I will keep it as it is now.

The reason it shows a SIGTRAP is that record_full_stop_reason ended up 
over-written anyway to stop_no_reason, so you are right that just 
flipping isn't enough. If you were to put this call in an "else if", 
instead of unrelated ifs, the issue stops happening. But I wonder if 
this is the right place to solve it anyway. I think 
record_check_stopped_by_breakpoint should just not overwrite 
record_full_stop_reason at all times.

It seems that since record_check_stopped_by_breakpoint was introduced in 
9e8915c6cee5c37637521b424d723e990e06d597 by Pedro (back in 2015), it was 
already indiscriminately overwriting the reason. I changed that locally 
and see no regressions and it also fixes this issue. Pedro, do you 
happen to remember why you added this back then?

>   			  (aspace, tmp_pc, &record_full_stop_reason))
>   			{
>   			  if (record_debug)
> diff --git a/gdb/testsuite/gdb.reverse/watch-reverse.exp b/gdb/testsuite/gdb.reverse/watch-reverse.exp
> index 6b81a6fdf88..a7126ac9b1d 100644
> --- a/gdb/testsuite/gdb.reverse/watch-reverse.exp
> +++ b/gdb/testsuite/gdb.reverse/watch-reverse.exp
> @@ -118,8 +118,12 @@ gdb_test "continue" \
>       ".*\[Ww\]atchpoint.*ival3.*Old value = 0.*New value = -1.*ival3 = count; ival4 = count;.*" \
>       "watchpoint hit in reverse, fifth time"
>   
> +gdb_test_no_output "disable \$bpnum" "disable non-hw watchpoint"
> +
>   gdb_test_no_output "set can-use-hw-watchpoints 1" "enable hw watchpoints"
>   
> +gdb_test_no_output "enable \$bpnum" "enable hw watchpoint"
> +
>   ###
>   ###
>   ###

This is a better solution! nice catch :)

I did forget to mention in last email (sorry =/ ) but there is a second 
test to do with watchpoints, called watch-precsave.exp, that - from what 
I gather - should always be the same as watch-reverse, but it uses a 
pre-recorded execution instead. I tried the same trick of disabling and 
re-enabling the watchpoints, but it didn't trigger the bug on that one 
(and just blindly clean-restarting and redoing the start also didn't 
work), I think it needs some work too. Or at least some investigation 
about if the bug does trigger on that scenario.

-- 
Cheers,
Guinevere Larsen
She/Her/Hers