Fix PR18360 - internal error when using "interrupt -a"

If you do "interrupt -a" just while some thread is stepping over a
breakpoint, gdb trips on an internal error.

The test added by this patch manages to trigger this consistently by
spawning a few threads that are constantly tripping on a conditional
breakpoint whose condition always evaluates to false.  With current
gdb, you get:

~~~
 interrupt -a
 .../src/gdb/inline-frame.c:343: internal-error: void skip_inline_frames(ptid_t): Assertion `find_inline_frame_state (ptid) == NULL' failed.
 A problem internal to GDB has been detected,
 further debugging may prove unreliable.
 Quit this debugging session? (y or n) FAIL: gdb.threads/interrupt-while-step-over.exp: displaced-stepping=on: iter=0: interrupt -a (GDB internal error)
[...]
 .../src/gdb/inline-frame.c:343: internal-error: void skip_inline_frames(ptid_t): Assertion `find_inline_frame_state (ptid) == NULL' failed.
 A problem internal to GDB has been detected,
 further debugging may prove unreliable.
 Quit this debugging session? (y or n) FAIL: gdb.threads/interrupt-while-step-over.exp: displaced-stepping=off: iter=0: wait for stops (GDB internal error)
~~~

The assertion triggers because we're processing a stop for a thread
that had already stopped before and thus had already its inline-frame
state filled in.

Calling handle_inferior_event_1 directly within a
"thread_stop_requested" observer is something that I've wanted to get
rid of before, for being fragile.  Nowadays, infrun is aware of
threads with pending events, so we can use that instead, and let the
normal fetch_inferior_event -> handle_inferior_event code path handle
the forced stop.

The change to finish_step_over is necessary because sometimes a thread
that was told to PTRACE_SINGLESTEP reports back a SIGSTOP instead of a
SIGTRAP (i.e., we tell it to single-step, and then interrupt it quick
enough that on the kernel side the thread dequeues the SIGTOP before
ever having had a chance of executing the instruction to be stepped).
SIGSTOP gets translated to a GDB_SIGNAL_0.  And then finish_step_over
would miss calling clear_step_over_info, and thus miss restarting the
other threads (which in this case of threads with pending events,
means setting their "resumed" flag, so their pending events can be
consumed).

And now that we always restart threads in finish_step_over, we no
longer need to do that in handle_signal_stop.

Tested on x86_64 Fedora 23, native and gdbserver.

gdb/ChangeLog:
2017-03-08  Pedro Alves  <palves@redhat.com>

	PR gdb/18360
	* infrun.c (start_step_over, do_target_resume, resume)
	(restart_threads): Assert we're not resuming a thread that is
	meant to be stopped.
	(infrun_thread_stop_requested_callback): Delete.
	(infrun_thread_stop_requested): If the thread is internally
	stopped, queue a pending stop event and clear the thread's
	inline-frame state.
	(handle_stop_requested): New function.
	(handle_syscall_event, handle_inferior_event_1): Use
	handle_stop_requested.
	(handle_stop_requested): New function.
	(handle_signal_stop): Set the thread's stop_signal here instead of
	at caller.
	(finish_step_over): Clear step over info unconditionally.
	(handle_signal_stop): If the user had interrupted the event
	thread, consider the stop a random signal.
	(handle_signal_stop) <signal arrived while stepping over
	breakpoint>: Don't restart threads here.
	(stop_waiting): Don't clear step-over info here.

gdb/testsuite/ChangeLog:
2017-03-08  Pedro Alves  <palves@redhat.com>

	PR gdb/18360
	* gdb.threads/interrupt-while-step-over.c: New file.
	* gdb.threads/interrupt-while-step-over.exp: New file.

0 files changed, 0 insertions, 0 deletions