[3/7] Force to insert software single step breakpoint

  GDB doesn't insert software single step breakpoint if the instruction
branches to itself, so that the program can't stop after command "si".

(gdb) b 32
Breakpoint 2 at 0x8680: file git/gdb/testsuite/gdb.base/branch-to-self.c, line 32.
(gdb) c
Continuing.

Breakpoint 2, main () at gdb/git/gdb/testsuite/gdb.base/branch-to-self.c:32
32	  asm (".Lhere: " BRANCH_INSN " .Lhere"); /* loop-line */
(gdb) si
infrun: clear_proceed_status_thread (Thread 3991.3991)
infrun: proceed (addr=0xffffffff, signal=GDB_SIGNAL_DEFAULT)
infrun: step-over queue now empty
infrun: resuming [Thread 3991.3991] for step-over
infrun: skipping breakpoint: stepping past insn at: 0x8680
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
Sending packet: $Z0,8678,4#f3...Packet received: OK
infrun: skipping breakpoint: stepping past insn at: 0x8680
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
Sending packet: $Z0,b6fe86c8,4#82...Packet received: OK
infrun: resume (step=1, signal=GDB_SIGNAL_0), trap_expected=1, current thread [Thread 3991.3991] at 0x868

breakpoint.c:should_be_inserted thinks the breakpoint shouldn't be
inserted, which is wrong.  This patch restrict the condition that only
return false if breakpoint is NOT single step breakpoint.

gdb:

2016-03-23  Yao Qi  <yao.qi@linaro.org>

	* breakpoint.c (should_be_inserted): Don't return 0 if single
	step breakpoint is inserted at the address we're stepping over.
	* gdbarch.sh (software_single_step): Update comments.
	* gdbarch.h: Regenerated.
---
 gdb/breakpoint.c | 9 ++++++++-
 gdb/gdbarch.h    | 5 ++++-
 gdb/gdbarch.sh   | 5 ++++-
 3 files changed, 16 insertions(+), 3 deletions(-)
  

On 03/23/2016 04:09 PM, Yao Qi wrote:
> GDB doesn't insert software single step breakpoint if the instruction
> branches to itself, so that the program can't stop after command "si".
> 
> (gdb) b 32
> Breakpoint 2 at 0x8680: file git/gdb/testsuite/gdb.base/branch-to-self.c, line 32.
> (gdb) c
> Continuing.
> 
> Breakpoint 2, main () at gdb/git/gdb/testsuite/gdb.base/branch-to-self.c:32
> 32	  asm (".Lhere: " BRANCH_INSN " .Lhere"); /* loop-line */
> (gdb) si
> infrun: clear_proceed_status_thread (Thread 3991.3991)
> infrun: proceed (addr=0xffffffff, signal=GDB_SIGNAL_DEFAULT)
> infrun: step-over queue now empty
> infrun: resuming [Thread 3991.3991] for step-over
> infrun: skipping breakpoint: stepping past insn at: 0x8680
> ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
> Sending packet: $Z0,8678,4#f3...Packet received: OK
> infrun: skipping breakpoint: stepping past insn at: 0x8680
> ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
> Sending packet: $Z0,b6fe86c8,4#82...Packet received: OK
> infrun: resume (step=1, signal=GDB_SIGNAL_0), trap_expected=1, current thread [Thread 3991.3991] at 0x868
> 
> breakpoint.c:should_be_inserted thinks the breakpoint shouldn't be
> inserted, which is wrong.  This patch restrict the condition that only
> return false if breakpoint is NOT single step breakpoint.
> 
> gdb:
> 
> 2016-03-23  Yao Qi  <yao.qi@linaro.org>
> 
> 	* breakpoint.c (should_be_inserted): Don't return 0 if single
> 	step breakpoint is inserted at the address we're stepping over.
> 	* gdbarch.sh (software_single_step): Update comments.
> 	* gdbarch.h: Regenerated.
> ---
>  gdb/breakpoint.c | 9 ++++++++-
>  gdb/gdbarch.h    | 5 ++++-
>  gdb/gdbarch.sh   | 5 ++++-
>  3 files changed, 16 insertions(+), 3 deletions(-)
> 
> diff --git a/gdb/breakpoint.c b/gdb/breakpoint.c
> index f99a7ab..9ecfb07 100644
> --- a/gdb/breakpoint.c
> +++ b/gdb/breakpoint.c
> @@ -2219,9 +2219,16 @@ should_be_inserted (struct bp_location *bl)
>      return 0;
>  
>    /* Don't insert a breakpoint if we're trying to step past its
> -     location.  */
> +     location except single step breakpoint, because the single step
> +     breakpoint may be inserted at the location we're trying to step
> +     if the instruction branches to itself.  However, the instruction
> +     won't be executed at all and it may break the semantics of the
> +     instruction, for example, the instruction is a conditional
> +     branch or updates some flags.  We can't fix it unless GDB is able
> +     to emulate the instruction or switch to displaced stepping.  */
>    if ((bl->loc_type == bp_loc_software_breakpoint
>         || bl->loc_type == bp_loc_hardware_breakpoint)
> +      && bl->owner->type != bp_single_step
>        && stepping_past_instruction_at (bl->pspace->aspace,
>  				       bl->address))

Another scenario occurred to me:

 - Thread A is software single-stepping.
 - Thread B hits single-step breakpoint of thread A.
 - We pause all threads and set thread B stepping past the
   single-step breakpoint of thread A.

But if the single-step breakpoint is for another thread, then
we won't actually manage to have thread B step past it, resulting
in spurious re-hits and no-guaranteed forward progress.  See
e.g., non-stop-fair-events.exp:

        # On software single-step targets that don't support displaced
        # stepping, threads keep hitting each others' single-step
        # breakpoints, and then GDB needs to pause all threads to step
        # past those.  The end result is that progress in the main
        # thread will be slower and it may take a bit longer for the
        # signal to be queued; bump the timeout.

Sounds like we may need to look at the single-step breakpoint's thread 
id, and only insert it if it is for the thread that is going to be 
doing the step-over?  We may need to record that in step_over_info and 
pass more info to stepping_past_instruction_at.

> --- a/gdb/gdbarch.sh
> +++ b/gdb/gdbarch.sh
> @@ -609,7 +609,10 @@ m:CORE_ADDR:addr_bits_remove:CORE_ADDR addr:addr::core_addr_identity::0
>  # target can single step.  If not, then implement single step using breakpoints.
>  #
>  # A return value of 1 means that the software_single_step breakpoints
> -# were inserted; 0 means they were not.
> +# were inserted; 0 means they were not.  Multiple breakpoints may be
> +# inserted for some instructions such as conditional branch.  However,
> +# each implementation must always evaluate the condition and only put
> +# the breakpoint at the branch destination if the condition is true.

I'd add:

(...) condition is true, so that we ensure forward progress when 
stepping past a conditional branch to self.

This will help porters evaluate whether that's really necessary
for their ports.

Thanks,
Pedro Alves
  

Pedro Alves <palves@redhat.com> writes:

> Another scenario occurred to me:
>
>  - Thread A is software single-stepping.
>  - Thread B hits single-step breakpoint of thread A.
>  - We pause all threads and set thread B stepping past the
>    single-step breakpoint of thread A.
>
> But if the single-step breakpoint is for another thread, then
> we won't actually manage to have thread B step past it, resulting
> in spurious re-hits and no-guaranteed forward progress.  See
> e.g., non-stop-fair-events.exp:
>
>         # On software single-step targets that don't support displaced
>         # stepping, threads keep hitting each others' single-step
>         # breakpoints, and then GDB needs to pause all threads to step
>         # past those.  The end result is that progress in the main
>         # thread will be slower and it may take a bit longer for the
>         # signal to be queued; bump the timeout.

I finally managed to reproduce that thread id in step_over_info is
different from the thread id of the single-step breakpoint.

GDB now gives the high priority to finishing step over, to avoid
"threads keep hitting each others' single-step breakpoint".  With my
patch applied, single-step breakpoint (of threads other than we are
stepping over) is still inserted even we try to step past the location,
so the step-over can't be finished.

>
> Sounds like we may need to look at the single-step breakpoint's thread 
> id, and only insert it if it is for the thread that is going to be 
> doing the step-over?  We may need to record that in step_over_info and 
> pass more info to stepping_past_instruction_at.

Yes, after I added a new field 'struct thread_info *thread' in
'struct step_over_info', I realize that IWBN to convert 'step_over_info'
to 'the thread we are stepping over', so the fields 'aspace' and 'address'
can be replaced by 'thread', like this,

struct step_over_info
{
  struct thread_info *thread;
  int nonsteppable_watchpoint_p;
};

Is it a good idea?  If there is nothing obviously wrong, I'll post
patches to do this.
  

Pedro Alves <palves@redhat.com> writes:

> Another scenario occurred to me:
>
>  - Thread A is software single-stepping.
>  - Thread B hits single-step breakpoint of thread A.
>  - We pause all threads and set thread B stepping past the
>    single-step breakpoint of thread A.
>
> But if the single-step breakpoint is for another thread, then
> we won't actually manage to have thread B step past it, resulting
> in spurious re-hits and no-guaranteed forward progress.  See
> e.g., non-stop-fair-events.exp:
>
>         # On software single-step targets that don't support displaced
>         # stepping, threads keep hitting each others' single-step
>         # breakpoints, and then GDB needs to pause all threads to step
>         # past those.  The end result is that progress in the main
>         # thread will be slower and it may take a bit longer for the
>         # signal to be queued; bump the timeout.
>
> Sounds like we may need to look at the single-step breakpoint's thread 
> id, and only insert it if it is for the thread that is going to be 
> doing the step-over?  We may need to record that in step_over_info and 
> pass more info to stepping_past_instruction_at.

I think this is about any thread specific breakpoint, instead of
only single-step breakpoint (single-step breakpoint is thread specific
too).  If we are doing step-over for thread A, do we need to insert any
breakpoints specific to other threads?  (my answer is No).
  

On 04/19/2016 03:54 PM, Yao Qi wrote:
> Pedro Alves <palves@redhat.com> writes:
> 
>> Another scenario occurred to me:
>>
>>  - Thread A is software single-stepping.
>>  - Thread B hits single-step breakpoint of thread A.
>>  - We pause all threads and set thread B stepping past the
>>    single-step breakpoint of thread A.
>>
>> But if the single-step breakpoint is for another thread, then
>> we won't actually manage to have thread B step past it, resulting
>> in spurious re-hits and no-guaranteed forward progress.  See
>> e.g., non-stop-fair-events.exp:
>>
>>         # On software single-step targets that don't support displaced
>>         # stepping, threads keep hitting each others' single-step
>>         # breakpoints, and then GDB needs to pause all threads to step
>>         # past those.  The end result is that progress in the main
>>         # thread will be slower and it may take a bit longer for the
>>         # signal to be queued; bump the timeout.
>>
>> Sounds like we may need to look at the single-step breakpoint's thread 
>> id, and only insert it if it is for the thread that is going to be 
>> doing the step-over?  We may need to record that in step_over_info and 
>> pass more info to stepping_past_instruction_at.
> 
> I think this is about any thread specific breakpoint, instead of
> only single-step breakpoint (single-step breakpoint is thread specific
> too).  If we are doing step-over for thread A, do we need to insert any
> breakpoints specific to other threads?  (my answer is No).

Right, we don't need to insert them, because other threads
will remain stopped while thread A is doing the step-over.

However, given that gdb does not remove/re-insert all breakpoints on
internal stops nowadays, removing thread-specific breakpoints of others
threads will be less efficient than leaving them be, I think.  I mean,
you'll get more z0/Z0 traffic than if you leave them inserted.

Thanks,
Pedro Alves