This is the mail archive of the
rda@sourceware.org
mailing list for the rda project.
[RFC] Improve performance of multi-threaded debugging
- From: Kevin Buettner <kevinb at redhat dot com>
- To: rda at sources dot redhat dot com
- Date: Wed, 14 Sep 2005 15:04:39 -0700
- Subject: [RFC] Improve performance of multi-threaded debugging
- Organization: Red Hat
As things stand now, the thread list is fetched each time rda checks
the status of the program. This doesn't sound like such a burden until
you realize that some decent sized data structures are read via the
ptrace() interface. On slower machines, it can take a significant amount
of time to single step or continue primarily due to this overhead.
The patch below fetches the thread list only when it knows for certain
that something has changed.
The only fly in the ointment is that the signal based event model only
knows about thread creation, but not about thread death. So it won't
catch thread death until some new thread is created. I'm not sure
what the implications of this are in practice.
Comments?
* thread-db.c (ALWAYS_UPDATE_THREAD_LIST): Define to be 0.
(handle_thread_db_event): Update thread list upon receipt of
TD_CREATE or TD_DEATH events.
(thread_db_check_child_state): Potentially disable, depending upon
value of ALWAYS_UPDATE_THREAD_LIST, the thread list update.
(thread_db_check_child_state): Update thread list for signal based
event model too.
Index: thread-db.c
===================================================================
RCS file: /cvs/cvsfiles/devo/rda/unix/thread-db.c,v
retrieving revision 1.25.2.1
diff -u -p -r1.25.2.1 thread-db.c
--- thread-db.c 3 Aug 2005 05:20:58 -0000 1.25.2.1
+++ thread-db.c 14 Sep 2005 21:31:48 -0000
@@ -48,6 +48,8 @@
int thread_db_noisy = 0;
int proc_service_noisy = 0;
+#define ALWAYS_UPDATE_THREAD_LIST 0
+
/*
* A tiny local symbol table.
*
@@ -1779,6 +1781,7 @@ handle_thread_db_event (struct child_pro
struct gdbserv_thread *thread = process->event_thread;
lwpid_t lwp;
union wait w;
+ int do_update = 0;
/* We need to be actually using the event interface. */
if (! using_thread_db_events)
@@ -1812,13 +1815,16 @@ handle_thread_db_event (struct child_pro
break;
}
- /* The only messages we're concerned with are TD_CREATE and
- TD_DEATH.
+ if (msg.event == TD_CREATE || msg.event == TD_DEATH)
+ do_update = 1;
+ }
- Every time thread_db_check_child_state gets a wait status
- from waitpid, we call update_thread_list, so our list is
- always up to date; we don't actually need to do anything with
- these messages for our own sake. */
+ if (do_update)
+ {
+#if !ALWAYS_UPDATE_THREAD_LIST
+ /* Update the thread list. */
+ update_thread_list (process);
+#endif
}
/* Disable the event breakpoints while we step the thread across them. */
@@ -2122,9 +2128,11 @@ thread_db_check_child_state (struct chil
process->stop_signal,
(unsigned long) debug_get_pc (process->serv, process->pid));
+#if ALWAYS_UPDATE_THREAD_LIST
/* Update the thread list, and attach to (and thereby stop)
any new threads we find. */
update_thread_list (process);
+#endif
process->event_thread = thread_list_lookup_by_lid (process->pid);
@@ -2170,6 +2178,12 @@ thread_db_check_child_state (struct chil
process->stop_signal = restart_signal;
else /* not main thread */
process->stop_signal = 0;
+
+#if !ALWAYS_UPDATE_THREAD_LIST
+ /* Update the thread list. */
+ update_thread_list (process);
+#endif
+
}
process->signal_to_send = process->stop_signal;
currentvec->continue_program (serv);