Commit a25a5a45 (Fix "breakpoint always-inserted off"; remove
"breakpoint always-inserted auto") regressed non-stop remote
debugging.
This was exposed by mi-nsintrall.exp intermittently failing with a
spurious SIGTRAP.
The problem is that when debugging with "target remote", new threads
the target has spawned but have never reported a stop aren't visible
to GDB until it explicitly resyncs its thread list with the target's.
For example, in a program like this:
int
main (void)
{
pthread_t child_thread;
pthread_create (&child_thread, NULL, child_function, NULL);
return 0; <<<< set breakpoint here
}
If the user sets a breakpoint at the "return" statement, and runs the
program, when that breakpoint hit is reported, GDB is only aware of
the main thread. So if we base the decision to remove or insert
breakpoints from the target based on whether all the threads we know
about are stopped, we'll miss that child_thread is running, and thus
we'll remove breakpoints from the target, even through they should
still remain inserted, otherwise child_thread will miss them.
The break-while-running.exp test actually should also be exposing this
thread-list-out-of-synch problem. That test sets a breakpoint while
the main thread is stopped, but other threads are running. Because
other threads are running, the breakpoint is supposed to be inserted
immediately. But, unless something forces a refetch of the thread
list, like, e.g., "info threads", GDB won't be aware of the other
threads that had been spawned by the main thread, and so won't insert
new or old breakpoints in the target. And it turns out that the test
is exactly doing an explicit "info threads", masking out the
problem... This commit adjust the test to exercise the case of not
issuing "info threads". The test then fails without the GDB fix.
In the ni-nsintrall.exp case, what happens is that several threads hit
the same breakpoint, and when the first thread reports the stop,
because GDB wasn't aware other threads exist, all threads known to GDB
are found stopped, so GDB removes the breakpoints from the target.
The other threads follow up with SIGTRAPs too for that same
breakpoint, which has already been removed. For the first few
threads, the moribund breakpoints machinery suppresses the SIGTRAPs,
but after a few events (precisely '3 * thread_count () + 1' at the
time the breakpoint was removed, see update_global_location_list), the
moribund breakpoint machinery is no longer aware of the removed
breakpoint, and the SIGTRAP is reported as a spurious stop.
The fix is naturally then to stop assuming that if no thread in the
list is executing, then the target is fully stopped. We can't know
that until we fully sync the thread list. Because updating the thread
list on every stop would be too much RSP traffic, I chose instead to
update it whenever we're about to present a stop to the user.
Actually updating the thread list at that point happens to be an item
I had added to the local/remote parity wiki page a while ago:
Native GNU/Linux debugging adds new threads to the thread list as
the program creates them "The [New Thread foo] messages". Remote
debugging can't do that, and it's arguable whether we shouldn't even
stop native debugging from doing that, as it hinders inferior
performance. However, a related issue is that with remote targets
(and gdbserver), even after the program stops, the user still needs
to do "info threads" to pull an updated thread list. This, should
most likely be addressed, so that GDB pulls the list itself, perhaps
just before presenting a stop to the user.
With that in place, the need to delay "Program received signal FOO"
was actually caught by the manythreads.exp test. Without that bit, I
was getting:
[Thread 0x7ffff7f13700 (LWP 4499) exited]
[New Thread 0x7ffff7f0b700 (LWP 4500)]
^C
Program received signal SIGINT, Interrupt.
[New Thread 0x7ffff7f03700 (LWP 4501)] <<< new output
[Switching to Thread 0x7ffff7f0b700 (LWP 4500)]
__GI___nptl_death_event () at events.c:31
31 {
(gdb) FAIL: gdb.threads/manythreads.exp: stop threads 1
That is, I was now getting "New Thread" lines after the "Program
received signal" line, and the test doesn't expect them. As the
number of new threads discovered before and after the "Program
received signal" output is unbounded, it's much nicer to defer
"Program received signal" until after synching the thread list, thus
close to the "switching to thread" output and "current frame/source"
info:
[Thread 0x7ffff7863700 (LWP 7647) exited]
^C[New Thread 0x7ffff786b700 (LWP 7648)]
Program received signal SIGINT, Interrupt.
[Switching to Thread 0x7ffff7fc4740 (LWP 6243)]
__GI___nptl_create_event () at events.c:25
25 {
(gdb) PASS: gdb.threads/manythreads.exp: stop threads 1
Tested on x86_64 Fedora 20, native and gdbserver.
gdb/
2014-10-02 Pedro Alves <palves@redhat.com>
* breakpoint.c (breakpoints_should_be_inserted_now): Use
threads_are_executing.
* breakpoint.h (breakpoints_should_be_inserted_now): Add
describing comment.
* gdbthread.h (threads_are_executing): Declare.
(handle_signal_stop) <random signals>: Don't print about the
signal here if stopping.
(end_stepping_range): Don't notify observers here.
(normal_stop): Update the thread list. If stopped by a random
signal or a stepping range ended, notify observers.
* thread.c (threads_executing): New global.
(init_thread_list): Clear 'threads_executing'.
(set_executing): Set or clear 'threads_executing'.
(threads_are_executing): New function.
(update_threads_executing): New function.
(update_thread_list): Use it.
gdb/testsuite/
2014-10-02 Pedro Alves <palves@redhat.com>
* gdb.threads/break-while-running.exp (test): Add new
'update_thread_list' argument. Skip "info threads" if false.
(top level): Add new 'update_thread_list' axis.
By default, GDB removes all breakpoints from the target when the
target stops and the prompt is given back to the user. This is useful
in case GDB crashes while the user is interacting, as otherwise,
there's a higher chance breakpoints would be left planted on the
target.
But, as long as any thread is running free, we need to make sure to
keep breakpoints inserted, lest a thread misses a breakpoint. With
that in mind, in preparation for non-stop mode, we added a "breakpoint
always-inserted on" mode. This traded off the extra crash protection
for never having threads miss breakpoints, and in addition is more
efficient if there's a ton of breakpoints to remove/insert at each
user command (e.g., at each "step").
When we added non-stop mode, and for a period, we required users to
manually set "always-inserted on" when they enabled non-stop mode, as
otherwise GDB removes all breakpoints from the target as soon as any
thread stops, which means the other threads still running will miss
breakpoints. The test added by this patch exercises this.
That soon revealed a nuisance, and so later we added an extra
"breakpoint always-inserted auto" mode, that made GDB behave like
"always-inserted on" when non-stop was enabled, and "always-inserted
off" when non-stop was disabled. "auto" was made the default at the
same time.
In hindsight, this "auto" setting was unnecessary, and not the ideal
solution. Non-stop mode does depends on breakpoints always-inserted
mode, but only as long as any thread is running. If no thread is
running, no breakpoint can be missed. The same is true for all-stop
too. E.g., if, in all-stop mode, and the user does:
(gdb) c&
(gdb) b foo
That breakpoint at "foo" should be inserted immediately, but it
currently isn't -- currently it'll end up inserted only if the target
happens to trip on some event, and is re-resumed, e.g., an internal
breakpoint triggers that doesn't cause a user-visible stop, and so we
end up in keep_going calling insert_breakpoints. The test added by
this patch also covers this.
IOW, no matter whether in non-stop or all-stop, if the target fully
stops, we can remove breakpoints. And no matter whether in all-stop
or non-stop, if any thread is running in the target, then we need
breakpoints to be immediately inserted. And then, if the target has
global breakpoints, we need to keep breakpoints even when the target
is stopped.
So with that in mind, and aiming at reducing all-stop vs non-stop
differences for all-stop-on-stop-of-non-stop, this patch fixes
"breakpoint always-inserted off" to not remove breakpoints from the
target until it fully stops, and then removes the "auto" setting as
unnecessary. I propose removing it straight away rather than keeping
it as an alias, unless someone complains they have scripts that need
it and that can't adjust.
Tested on x86_64 Fedora 20.
gdb/
2014-09-22 Pedro Alves <palves@redhat.com>
* NEWS: Mention merge of "breakpoint always-inserted" modes "off"
and "auto" merged.
* breakpoint.c (enum ugll_insert_mode): New enum.
(always_inserted_mode): Now a plain boolean.
(show_always_inserted_mode): No longer handle AUTO_BOOLEAN_AUTO.
(breakpoints_always_inserted_mode): Delete.
(breakpoints_should_be_inserted_now): New function.
(insert_breakpoints): Pass UGLL_INSERT to
update_global_location_list instead of calling
insert_breakpoint_locations manually.
(create_solib_event_breakpoint_1): New, factored out from ...
(create_solib_event_breakpoint): ... this.
(create_and_insert_solib_event_breakpoint): Use
create_solib_event_breakpoint_1 instead of calling
insert_breakpoint_locations manually.
(update_global_location_list): Change parameter type from boolean
to enum ugll_insert_mode. All callers adjusted. Adjust to use
breakpoints_should_be_inserted_now and handle UGLL_INSERT.
(update_global_location_list_nothrow): Change parameter type from
boolean to enum ugll_insert_mode.
(_initialize_breakpoint): "breakpoint always-inserted" option is
now a boolean command. Update help text.
* breakpoint.h (breakpoints_always_inserted_mode): Delete declaration.
(breakpoints_should_be_inserted_now): New declaration.
* infrun.c (handle_inferior_event) <TARGET_WAITKIND_LOADED>:
Remove breakpoints_always_inserted_mode check.
(normal_stop): Adjust to use breakpoints_should_be_inserted_now.
* remote.c (remote_start_remote): Likewise.
gdb/doc/
2014-09-22 Pedro Alves <palves@redhat.com>
* gdb.texinfo (Set Breaks): Document that "set breakpoint
always-inserted off" is the default mode now. Delete
documentation of "set breakpoint always-inserted auto".
gdb/testsuite/
2014-09-22 Pedro Alves <palves@redhat.com>
* gdb.threads/break-while-running.exp: New file.
* gdb.threads/break-while-running.c: New file.
Pedro Alves