Re: [PATCH, ppc] Fix hw *points for embedded ppc in a threaded environment

[Luis Machado <lgustavo at codesourcery dot com>]

On 04/22/2013 09:02 PM, Pedro Alves wrote:
> I hacked GDB some more, and I have further insight.
> GDB _is_ to blame.  Updated GDB/test hack patch at the bottom.

> So GDB is to blame.  Fixing this, however, would be a different
> story, and doesn't look that simple.
>
> So the test could work around this by making sure that threads
> don't exit until after all watchpoints have been tested.

I did one more fix to the testcase to allow it to run properly when the 
number of threads is greater than the number of available hardware 
watchpoints.

I also cleaned up, adjusted and added some more comments to the expect 
file. NR_BYTES was removed (it was useless), so now we use NR_THREADS to 
allocate the static array.

One more internal variable was added to the testcase to allow GDB to 
share its knowledge on the number of available hardware watchpoints. 
That way the testcase never expects more triggers than what is supposed 
to come.

The question here is how many threads we should spawn. I went with 10 
since that is often greater than the number of hardware watchpoints in a 
target.

How does it look?

Luis

2013-04-23  Luis Machado  <lgustavo@codesourcery.com>

	* ppc-linux-nat.c (ppc_linux_new_thread): Clear the new thread's
	debug state prior to replicating existing hardware watchpoints or
	breakpoints.

	* gdb/testsuite/gdb.threads/wp-replication.c: New file.
	* gdb/testsuite/gdb.threads/wp-replication.exp: New file.

Index: gdb-head/gdb/ppc-linux-nat.c
===================================================================
--- gdb-head.orig/gdb/ppc-linux-nat.c	2013-04-23 14:40:07.418560514 +0200
+++ gdb-head/gdb/ppc-linux-nat.c	2013-04-23 14:40:48.262559788 +0200
@@ -2178,7 +2178,21 @@ ppc_linux_new_thread (struct lwp_info *l
       /* Copy that thread's breakpoints and watchpoints to the new thread.  */
       for (i = 0; i < max_slots_number; i++)
 	if (hw_breaks[i].hw_break)
-	  booke_insert_point (hw_breaks[i].hw_break, tid);
+	  {
+	    /* The ppc Linux kernel causes a thread to inherit its parent
+	       thread's debug state, and that includes any hardware
+	       watchpoints or breakpoints that the parent thread may have set.
+
+	       For this reason, the debug state of the new thread is cleared
+	       before trying to replicate any hardware watchpoints or
+	       breakpoints contained in other threads.  */
+
+	    /* The ppc debug resource accounting is done through "slots".
+	       Ask the kernel the deallocate this specific *point's slot.  */
+	    ptrace (PPC_PTRACE_DELHWDEBUG, tid, 0, hw_breaks[i].slot);
+
+	    booke_insert_point (hw_breaks[i].hw_break, tid);
+	  }
     }
   else
     ptrace (PTRACE_SET_DEBUGREG, tid, 0, saved_dabr_value);
Index: gdb-head/gdb/testsuite/gdb.threads/wp-replication.c
===================================================================
--- /dev/null	1970-01-01 00:00:00.000000000 +0000
+++ gdb-head/gdb/testsuite/gdb.threads/wp-replication.c	2013-04-23 15:29:10.518508221 +0200
@@ -0,0 +1,160 @@
+/* This testcase is part of GDB, the GNU debugger.
+
+   Copyright 2013 Free Software Foundation, Inc.
+
+   This program is free software; you can redistribute it and/or modify
+   it under the terms of the GNU General Public License as published by
+   the Free Software Foundation; either version 3 of the License, or
+   (at your option) any later version.
+
+   This program is distributed in the hope that it will be useful,
+   but WITHOUT ANY WARRANTY; without even the implied warranty of
+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+   GNU General Public License for more details.
+
+   You should have received a copy of the GNU General Public License
+   along with this program.  If not, see <http://www.gnu.org/licenses/>.
+
+   Check that hardware watchpoints get propagated to all existing
+   threads when the hardware watchpoint is created.
+*/
+
+#include <stdio.h>
+#include <unistd.h>
+#include <stdlib.h>
+#include <stdint.h>
+#include <pthread.h>
+
+#ifndef NR_THREADS
+#define NR_THREADS 4 /* Set by the testcase.  */
+#endif
+
+#ifndef X_INCR_COUNT
+#define X_INCR_COUNT 10 /* Set by the testcase.  */
+#endif
+
+void *thread_function (void *arg); /* Function executed by each thread.  */
+
+/* Used to hold threads back until wp-replication.exp is ready.  */
+int test_ready = 0;
+
+/* Used to hold threads back until every thread has had a chance of causing
+   a watchpoint trigger.  This prevents a situation in GDB where it may miss
+   watchpoint triggers when threads exit while other threads are causing
+   watchpoint triggers.  */
+int can_terminate = 0;
+
+/* Used to push the program out of the waiting loop after the
+   testcase is done counting the number of hardware watchpoints
+   available for our target.  */
+int watch_count_done = 0;
+
+/* Number of watchpoints GDB is capable of using (this is provided
+   by GDB during the test run).  */
+int hw_watch_count = 0;
+
+/* Array with elements we can create watchpoints for.  */
+static int watched_data[NR_THREADS];
+pthread_mutex_t data_mutex;
+
+/* Wait function to keep threads busy while the testcase does
+   what it needs to do.  */
+void empty_cycle (void)
+{
+  usleep (1);
+}
+
+int
+main ()
+{
+  int res;
+  pthread_t threads[NR_THREADS];
+  int i;
+
+  while (watch_count_done == 0)
+    {
+      /* GDB will modify the value of "i" at runtime and we will
+	 get past this point.  */
+      empty_cycle ();
+    }
+
+  pthread_mutex_init (&data_mutex, NULL);
+
+  for (i = 0; i < NR_THREADS; i++)
+    {
+      res = pthread_create (&threads[i],
+			    NULL, thread_function,
+			    (void *) (intptr_t) i);
+      if (res != 0)
+	{
+	  fprintf (stderr, "error in thread %d create\n", i);
+	  abort ();
+	}
+    }
+
+  for (i = 0; i < NR_THREADS; ++i)
+    {
+      res = pthread_join (threads[i], NULL);
+      if (res != 0)
+	{
+	  fprintf (stderr, "error in thread %d join\n", i);
+	  abort ();
+	}
+    }
+
+  exit (EXIT_SUCCESS);
+}
+
+/* Easy place for a breakpoint.
+   wp-replication.exp uses this to track when all threads are running
+   instead of, for example, the program keeping track
+   because we don't need the program to know when all threads are running,
+   instead we need gdb to know when all threads are running.
+   There is a delay between when a thread has started and when the thread
+   has been registered with gdb.  */
+
+void
+thread_started ()
+{
+}
+
+void *
+thread_function (void *arg)
+{
+  int i, j;
+  long thread_number = (long) arg;
+
+  thread_started ();
+
+  /* Don't start incrementing X until wp-replication.exp is ready.  */
+  while (!test_ready)
+    usleep (1);
+
+  pthread_mutex_lock (&data_mutex);
+
+  for (i = 0; i < NR_TRIGGERS_PER_THREAD; i++)
+    {
+      for (j = 0; j < hw_watch_count; j++)
+	{
+	  /* For debugging.  */
+	  printf ("Thread %ld changing watch_thread[%d] data"
+	          " from %d -> %d\n", thread_number, j,
+	          watched_data[j], watched_data[j] + 1);
+	  /* The call to usleep is so that when the watchpoint triggers,
+	     the pc is still on the same line.  */
+	  /* Increment the watched data field.  */
+	  watched_data[j]++;
+	  usleep (1);
+	}
+    }
+
+  pthread_mutex_unlock (&data_mutex);
+
+  /* Hold the threads here to work around a problem GDB has evaluating
+     watchpoints right when a DSO event shows up.  Sleep a little longer
+     to avoid consuming lots of cycles.  */
+  while (!can_terminate)
+    usleep (100);
+
+  pthread_exit (NULL);
+}
Index: gdb-head/gdb/testsuite/gdb.threads/wp-replication.exp
===================================================================
--- /dev/null	1970-01-01 00:00:00.000000000 +0000
+++ gdb-head/gdb/testsuite/gdb.threads/wp-replication.exp	2013-04-23 15:13:16.230525175 +0200
@@ -0,0 +1,148 @@
+# This testcase is part of GDB, the GNU debugger.
+
+# Copyright 2013 Free Software Foundation, Inc.
+
+# This program is free software; you can redistribute it and/or modify
+# it under the terms of the GNU General Public License as published by
+# the Free Software Foundation; either version 3 of the License, or
+# (at your option) any later version.
+#
+# This program is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+# GNU General Public License for more details.
+#
+# You should have received a copy of the GNU General Public License
+# along with this program.  If not, see <http://www.gnu.org/licenses/>.
+
+# Check that hardware watchpoints get replicated to all existing
+# threads when the hardware watchpoint is created.
+
+set NR_THREADS 10
+set NR_TRIGGERS_PER_THREAD 2
+
+# This test verifies that a hardware watchpoint gets replicated to
+# every existing thread and is detected properly.  This test is
+# only meaningful on a target with hardware watchpoint support.
+if {[skip_hw_watchpoint_tests]} {
+    return 0
+}
+
+standard_testfile
+if {[gdb_compile_pthreads "${srcdir}/${subdir}/${srcfile}" "${binfile}" executable [list debug "additional_flags=-DNR_THREADS=$NR_THREADS -DNR_TRIGGERS_PER_THREAD=$NR_TRIGGERS_PER_THREAD"]] != "" } {
+    return -1
+}
+
+clean_restart ${binfile}
+
+# Force hardware watchpoints to be used.
+gdb_test_no_output "set can-use-hw-watchpoints 1" ""
+
+# Run to `main' where we begin our tests.
+if ![runto_main] then {
+    gdb_suppress_tests
+}
+
+# First, break at empty_cycle.
+gdb_test "break empty_cycle" \
+  "Breakpoint 2 at .*: file .*${srcfile}, line .*" \
+  "breakpoint on empty_cycle"
+
+# Set some default values.
+set hwatch_count 0
+set done 0
+
+# Count the number of hardware watchpoints available on
+# this target.
+while { $done == 0 } {
+
+  gdb_test "continue" \
+    ".*Breakpoint 2, empty_cycle \\(\\) at .*${srcfile}.*" \
+    "continue to empty_cycle"
+
+  # Some targets do resource counting as we insert watchpoints.
+  # Such targets won't cause a watchpoint insertion failure, but
+  # will switch to software watchpoints silently.  We check for
+  # both cases here.
+  gdb_test_multiple "watch watched_data\[$hwatch_count\]" \
+       "watch watched_data\[$hwatch_count\]" {
+    -re "Hardware watchpoint .*$gdb_prompt $" {
+    }
+    -re "Watchpoint .*$gdb_prompt $" {
+      set done 1
+      break
+    }
+  }
+
+  gdb_test_multiple "continue" "watchpoint created successfully" {
+    -re ".*Breakpoint 2, empty_cycle \\(\\).*$gdb_prompt $" {
+      incr hwatch_count
+    }
+    -re ".*Could not insert hardware watchpoint.*$gdb_prompt $" {
+      set done 1
+      break
+    }
+  }
+}
+
+# Target cannot insert hardware watchpoints.  Maybe it should have reported
+# it does not supported them in the first place.
+if { $hwatch_count == 0} {
+  gdb_exit;
+}
+
+# Set the testcase's internal variable indicating the number of
+# hardware watchpoints the target supports.
+gdb_test_no_output "set var hw_watch_count=${hwatch_count}" \
+		   "set var hw_watch_count=${hwatch_count}"
+
+# At this point, we know how many hardware watchpoints
+# the target supports.  Use that to do further testing.
+delete_breakpoints
+
+# Break out of the empty_cycle loop by changing the
+# controlling variable.
+gdb_test_no_output "set var watch_count_done=1" \
+		   "set var watch_count_done=1"
+
+# Prepare to create all the threads.
+gdb_test "break thread_started" \
+	 "Breakpoint \[0-9\]+ at .*: file .*${srcfile}, line .*" \
+	 "Breakpoint on thread_started"
+
+# Move all threads to where they're supposed to be for testing.
+for { set i 0 } { $i < $NR_THREADS } { incr i } {
+
+    # We want to set the maximum number of hardware watchpoints
+    # and make sure the target can handle that without an error.
+    # That will show us the watchpoints got replicated to all the
+    # threads correctly, and that no new watchpoints got created
+    # in the background for a specific thread.
+    if {$i < $hwatch_count} {
+      gdb_test "watch watched_data\[$i\]" \
+	"Hardware watchpoint .*" \
+	"watch watched_data\[$i\]"
+    } else {
+      verbose -log "Not setting watchpoint for watched_data\[$i\]\n"
+    }
+
+    gdb_test continue "Continuing\\..*Breakpoint \[0-9\]+, thread_started \\(\\) at .*$srcfile.*" \
+    "Break at thread_started"
+}
+
+# Let the threads run and change the watched data, leading
+# to watchpoint triggers.
+gdb_test_no_output "set var test_ready=1" \
+      "set var test_ready=1"
+
+# Set the number of expected watchpoint triggers.
+set TRIGGERS [expr "$NR_THREADS * $hwatch_count * $NR_TRIGGERS_PER_THREAD"]
+
+# Move the threads and hit the watchpoints
+# TRIGGERS times.
+for { set i 0 } { $i < $TRIGGERS } { incr i} {
+    gdb_test continue "Continuing\\..*Hardware watchpoint \[0-9\]+: watched_data\[\[0-9\]+\].*Old value = \[0-9\]+.*New value = \[0-9\]+.*thread_function \\(arg=$hex\\) at .*$srcfile.*" \
+    "Continue to watchpoint trigger ${i} out of ${TRIGGERS} on watched_data"
+}
+
+gdb_exit