Why enforcing sw_breakpoint_from_kind() implementation in GDBserver targets

Simon Marchi simark@simark.ca
Thu Jun 11 03:05:38 GMT 2020


On 2020-06-10 1:47 p.m., Shahab Vahedi via Gdb wrote:
> Now, it has to provide the implementation for "sw_breakpoint_from_kind".
> That is not the whole story.  I also noticed that this  piece  of  newly
> implemented code never gets executed.  This makes sense because we  have
> a setup  that  looks  like  below  (both  entities  are  running  inside
> GNU/Linux):
> 
>  ,------------------------------------------------.
>  | ARC GDB client on x86 machine (cross debugger) |
>  `------------------------------------------------'
>                         /\
>                         ||
>                  remote debugging
>                         ||
>                         \/
>  ,------------------------------------------------.
>  |   ARC GDBserver on ARC board (native server)   |
>  `------------------------------------------------'
> 
> It is the "ARC GDB client" that inserts the breakpoint.  That has always
> been the case.  Else, it would be impossible to  break  while  debugging
> with GDBserver in older implementation (before  rebase).   It  is  worth
> mentioning the "ARC GDB client" does have the  "sw_breakpoint_from_kind"
> implemented [].

I don't understand your "This makes sense because we have a setup that looks
like below", because that looks like a standard GDB/GDBserver setup used for
other architectures.

When a breakpoint is inserted, what's the remote protocol packet used? Is it
Z0, or is it a memory write operation that writes the breakpoint's opcode?  Z0
is the "modern" way that provides more features (like target-side condition
evaluation) and a memory write is the legacy fallback.

sw_breakpoint_from_kind would be used if the Z0 packet was used, to translate
the "kind" into an opcode.  Since you claim that sw_breakpoint_from_kind is
not used, I guess that the breakpoint is inserted with a memory write operation.
I'd look into why that is the case.  GDB tries Z0 first and falls back to the
memory write if Z0 is not supported, so your GDBserver must not support it for
some reason.
> Last  but  not  least,  one  nitpick:  Even  though  I  have  added  the
> implementation of "sw_breakpoint_from_kind", I have never  done  so  for
> "breakpoint_kind_from_pc"    or    "breakpoint_kind_from_current_state".
> These last two are supposed to provide  the  "kind"  that  will  be  the
> input parameter for "sw_breakpoint_from_kind".  Therefore, even  if  the
> new piece of "sw_breakpoint_from_kind" would be executed, that would  be
> problematic.  I'm not sure what can be done about this but I think  _if_
> "sw_breakpoint_from_kind"     should     be    mandatory,     so     are
> "breakpoint_kind_from_pc" and "breakpoint_kind_from_current_state".

As mentioned above, the input for sw_breakpoint_from_kind can come from
the Z0 packet, so it may make sense to have sw_breakpoint_from_kind without
the others.  I am not sure off-hand when the others are used.

Simon


More information about the Gdb mailing list