gdbserver: handle running threads in qXfer:threads:read

On some systems, the gdb.multi/multi-target.exp testcase occasionally
fails like so:

 Running src/gdb/testsuite/gdb.multi/multi-target.exp ...
 FAIL: gdb.multi/multi-target.exp: info-inferiors: multi_process=on: inferior 1: info connections
 FAIL: gdb.multi/multi-target.exp: info-inferiors: multi_process=on: inferior 1: info inferiors
 FAIL: gdb.multi/multi-target.exp: info-inferiors: multi_process=on: inferior 2: info connections
 FAIL: gdb.multi/multi-target.exp: info-inferiors: multi_process=on: inferior 2: info inferiors
 FAIL: gdb.multi/multi-target.exp: info-inferiors: multi_process=on: inferior 3: inferior 3
 ... many more cascading fails.

The problem starts when the testcase runs an inferior against GDBserver:

 (gdb) run
 Starting program: build/gdb/testsuite/outputs/gdb.multi/multi-target/multi-target
 Reading /lib64/ld-linux-x86-64.so.2 from remote target...
 warning: File transfers from remote targets can be slow. Use "set sysroot" to access files locally instead.
 Reading /lib64/ld-linux-x86-64.so.2 from remote target...
 Reading /lib64/ld-2.31.so from remote target...
 Reading /lib64/.debug/ld-2.31.so from remote target...
 Reading /usr/lib/debug//lib64/ld-2.31.so from remote target...
 Reading /usr/lib/debug/lib64//ld-2.31.so from remote target...
 Reading target:/usr/lib/debug/lib64//ld-2.31.so from remote target...
 Reading /lib/x86_64-linux-gnu/libpthread.so.0 from remote target...
 Reading /lib/x86_64-linux-gnu/libc.so.6 from remote target...
 Reading /lib/x86_64-linux-gnu/libc-2.31.so from remote target...
 Reading /lib/x86_64-linux-gnu/.debug/libc-2.31.so from remote target...
 Reading /usr/lib/debug//lib/x86_64-linux-gnu/libc-2.31.so from remote target...
 Reading /usr/lib/debug//lib/x86_64-linux-gnu/libc-2.31.so from remote target...
 Remote connection closed
 ...

Note the "Remote connection closed" message.  That means GDBserver
exited abruptly.

I traced it down to the fact that GDB fetches the thread list from
GDBserver while the main thread of the process is still running.  On
my main system where I wrote the testcase, I have not observed the
failure because it is slow enough that the thread stops before
GDBserver fetches the thread list in the problem scenario which I'll
describe below.

With some --remote-debug logging from GDBserver side, we see the last
packets before the connection closes:

 ...
 getpkt ("vCont;c");  [no ack sent]
 putpkt ("$OK#9a"); [noack mode]
 getpkt ("Tp10f9a.10f9a");  [no ack sent]
 putpkt ("$OK#9a"); [noack mode]
 getpkt ("Hgp0.0");  [no ack sent]
 putpkt ("$OK#9a"); [noack mode]
 getpkt ("qXfer:threads:read::0,1000");  [no ack sent]

Note the vCont;c , which sets the program running, and then a
qXfer:threads:read packet at the end.

The problem happens when the thread list refresh (qXfer:threads:read)
is sent just while the main thread is running and it still hasn't
initialized its libpthread id internally.  In that state, the main
thread's lwp will remain with the thread_known flag clear.  See in
find_one_thread:

  /* If the new thread ID is zero, a final thread ID will be available
     later.  Do not enable thread debugging yet.  */
  if (ti.ti_tid == 0)
    return 0;

Now, back in server.cc, to handle the qXfer:threads:read, we reach
handle_qxfer_threads -> handle_qxfer_threads_proper, and the latter
then calls handle_qxfer_threads_worker for each known thread.  In
handle_qxfer_threads_worker, we call target_thread_handle.  This ends
up in thread_db_thread_handle, here:

  if (!lwp->thread_known && !find_one_thread (thread->id))
    return false;

Since the thread ID isn't known yet, we call find_one_thread.  This
calls into libthread_db.so, which accesses memory.  Because the
current thread is running, that fails and we throw an error, here:

  /* Get information about this thread.  */
  err = thread_db->td_ta_map_lwp2thr_p (thread_db->thread_agent, lwpid, &th);
  if (err != TD_OK)
    error ("Cannot get thread handle for LWP %d: %s",
	   lwpid, thread_db_err_str (err));

The current design is that whenever GDB-facing packets/requests need
to accesses memory, server.cc is supposed to prepare the target for
the access.  See gdb_read_memory / gdb_write_memory.  This preparation
means pausing threads if in non-stop mode (someday we could lift this
requirement, but we will still need to pause to access registers or do
other related ptrace accesses like PTRACE_GET_THREAD_AREA).  Note that
the multi-target.exp testcase forces "maint set target-non-stop on".

So the fix here is to prepare the target to access memory when
handling qXfer:threads:read too.

gdbserver/ChangeLog:

	* inferiors.cc (switch_to_process): New, moved here from
	thread-db.cc, and made extern.
	* inferiors.h (switch_to_process): Declare.
	* server.cc: Include "gdbsupport/scoped_restore.h".
	(handle_qxfer_threads_proper): Now returns bool.  Prepare to
	access memory around target calls.
	(handle_qxfer_threads): Handle errors.
	* thread-db.cc (switch_to_process): Moved to inferiors.cc.

1 files changed, 44 insertions, 5 deletions

diff --git a/gdbserver/server.cc b/gdbserver/server.cc
index ab5363e..aadcb9b 100644
--- a/gdbserver/server.cc
+++ b/gdbserver/server.cc
@@ -48,6 +48,7 @@
 #include "gdbsupport/selftest.h"
 #include "gdbsupport/scope-exit.h"
 #include "gdbsupport/gdb_select.h"
+#include "gdbsupport/scoped_restore.h"
 
 #define require_running_or_return(BUF)		\
   if (!target_running ())			\
@@ -1678,19 +1679,54 @@ handle_qxfer_threads_worker (thread_info *thread, struct buffer *buffer)
   buffer_xml_printf (buffer, "/>\n");
 }
 
-/* Helper for handle_qxfer_threads.  */
+/* Helper for handle_qxfer_threads.  Return true on success, false
+   otherwise.  */
 
-static void
+static bool
 handle_qxfer_threads_proper (struct buffer *buffer)
 {
+  client_state &cs = get_client_state ();
+
+  scoped_restore save_current_thread
+    = make_scoped_restore (&current_thread);
+  scoped_restore save_current_general_thread
+    = make_scoped_restore (&cs.general_thread);
+
   buffer_grow_str (buffer, "<threads>\n");
 
-  for_each_thread ([&] (thread_info *thread)
+  process_info *error_proc = find_process ([&] (process_info *process)
     {
-      handle_qxfer_threads_worker (thread, buffer);
+      /* The target may need to access memory and registers (e.g. via
+	 libthread_db) to fetch thread properties.  Prepare for memory
+	 access here, so that we potentially pause threads just once
+	 for all accesses.  Note that even if someday we stop needing
+	 to pause threads to access memory, we will need to be able to
+	 access registers, or other ptrace accesses like
+	 PTRACE_GET_THREAD_AREA.  */
+
+      /* Need to switch to each process in turn, because
+	 prepare_to_access_memory prepares for an access in the
+	 current process pointed to by general_thread.  */
+      switch_to_process (process);
+      cs.general_thread = current_thread->id;
+
+      int res = prepare_to_access_memory ();
+      if (res == 0)
+	{
+	  for_each_thread (process->pid, [&] (thread_info *thread)
+	    {
+	      handle_qxfer_threads_worker (thread, buffer);
+	    });
+
+	  done_accessing_memory ();
+	  return false;
+	}
+      else
+	return true;
     });
 
   buffer_grow_str0 (buffer, "</threads>\n");
+  return error_proc == nullptr;
 }
 
 /* Handle qXfer:threads:read.  */
@@ -1719,11 +1755,14 @@ handle_qxfer_threads (const char *annex,
 
       buffer_init (&buffer);
 
-      handle_qxfer_threads_proper (&buffer);
+      bool res = handle_qxfer_threads_proper (&buffer);
 
       result = buffer_finish (&buffer);
       result_length = strlen (result);
       buffer_free (&buffer);
+
+      if (!res)
+	return -1;
     }
 
   if (offset >= result_length)