This is the mail archive of the gdb-patches@sourceware.org mailing list for the GDB project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: [PATCH 1/2] Fix error when GDB connects to GDBserver with qC disabled


On 01/23/2013 08:35 AM, Yao Qi wrote:
> GDB gets an internal error when it connects to GDBserver started with
> '--disable-packet=qC'.
> 
> Sending packet: $QNonStop:0#8c...Packet received: OK
> Sending packet: $?#3f...Packet received: T0505:00000000;04:00f0ffbf;08:b0c2e44c;thread:p4255.4255;core:1;
> Sending packet: $Hc-1#09...Packet received: E01
> Sending packet: $qC#b4...Packet received:
> Sending packet: $qAttached:a410#bf...Packet received: E01
> Packet qAttached (query-attached) is supported
> warning: Remote failure reply: E01
> Sending packet: $qOffsets#4b...Packet received:
> ../../../git/gdb/target.c:3248: internal-error: Can't determine the current address space of thread Thread 16981
> 
> When start remote, the call chain is as follows,
> 
> remote_start_remote
>   add_current_inferior_and_thread <--[1]
>   ...
>   start_remote
>     wait_for_inferior
>       remote_wait_as
>         process_stop_reply
>           get_thread_arch_regcache   <--[2]
>           remote_notice_new_inferior <--[3]
> 
> GDB sends packet "qC" in [1] and adds the thread/inferior if the remote
> stubs understands "qC".  In [2], GDB looks for the inferior to build a
> regcache, and notices a new inferior in [3].  As we can see, GDB assumes
> that the inferior can be found in [2].  Once the remote stub doesn't
> support "qC", GDB can't look for the inferior in [2], and emits an
> internal error.
> 
> This patch fix this internal error by exchanging the order of [2] and
> [3].
> 

Thanks.

> gdb:
> 
> 2013-01-22  Yao Qi  <yao@codesourcery.com>
> 
> 	* remote.c (process_stop_reply): Call remote_notice_new_inferior
> 	earlier.
> ---
>  gdb/remote.c |   12 +++++++++---
>  1 files changed, 9 insertions(+), 3 deletions(-)
> 
> diff --git a/gdb/remote.c b/gdb/remote.c
> index 7ea9597..3271ca0 100644
> --- a/gdb/remote.c
> +++ b/gdb/remote.c
> @@ -5690,11 +5690,18 @@ process_stop_reply (struct stop_reply *stop_reply,
>        /* Expedited registers.  */
>        if (stop_reply->regcache)
>  	{
> -	  struct regcache *regcache
> -	    = get_thread_arch_regcache (ptid, target_gdbarch ());
> +	  struct regcache *regcache;
>  	  cached_reg_t *reg;
>  	  int ix;
>  
> +	  /* Add the inferior earlier, because the following
> +	     'get_thread_arch_regcache' calls
> +	     'target_thread_address_space', which requires the
> +	     inferior should be found in the inferior list.  */
> +	  remote_notice_new_inferior (ptid, 0);
> +
> +	  regcache = get_thread_arch_regcache (ptid,
> +					       target_gdbarch ());
>  	  for (ix = 0;
>  	       VEC_iterate(cached_reg_t, stop_reply->regcache, ix, reg);
>  	       ix++)
> @@ -5705,7 +5712,6 @@ process_stop_reply (struct stop_reply *stop_reply,
>        remote_stopped_by_watchpoint_p = stop_reply->stopped_by_watchpoint_p;
>        remote_watch_data_address = stop_reply->watch_data_address;
>  
> -      remote_notice_new_inferior (ptid, 0);
>        demand_private_info (ptid)->core = stop_reply->core;
>      }

Two things don't look right here.

First, this removes the remote_notice_new_inferior call in
the case the target doesn't send any expedited registers.
That is, the call is conditional on "if (stop_reply->regcache)".

Second, the point of the get_thread_arch_regcache call is to
pre-fill the regcache with the expedite register values, before
any other code needs one of the likely registers in the expedited
set (usually PC, SP, FP), thus saving a roundtrip.
I haven't checked if notice_new_inferior (the core function) ends
needing to fetch up registers; it's possible it ends up fetching
the whole g set anyway, but still, it doesn't feel right.

I had another idea.  Right after the initial connection, we
query the target for its state, with the ? packet.  We
store the resulting wait status / stop reply aside, and
query the target for the current thread, using qC, which
fails, so we fake a ptid for the target's thread.
We then later, after the initial setup, end up consuming that
set-aside wait status, parsing the T stop reply, which contains
a "thread" "register" (which was the thread the target would
have replied to qC).  We get into trouble because the ptid
in that stop reply doesn't match our faked up ptid in the initial
setup, although the target threads are the same...

The idea is that, we had the T stop reply handy all along.
We might as well extract the thread's ptid from it, and
avoid all the resulting issues.

qC is also used after vRun, in order to discover the
new process'es main thread.  But, vRun's reply is
also a wait status, just like '?''s, which is quite
convenient.

This means that if we have a "Txx thread: ptid" reply, then we
don't really need qC...  That's what the patch below does.
The patch still tries qC first, but I'm thinking we can flip that
around, and only try qC if the stop reply didn't include a thread.
I looked back in history trying to find the rationale for qC, but
I couldn't find it.  The packet handling seems to have been added
in gdb-4.18 (1999), and I see that in that same release,
"Txx thread: ptid" didn't exist yet, which explains why nobody
though of doing this then.

I ran the testsuite with it against a gdbserver with qC disabled,
and saw no regressions.

2013-01-24  Pedro Alves  <palves@redhat.com>

	* remote.c (stop_reply_extract_thread): New.
	(add_current_inferior_and_thread): New parameter 'wait_status'.
	Handle it.
	(remote_start_remote): Pass wait status to
	add_current_inferior_and_thread.
	(extended_remote_run): Update comment.
	(extended_remote_create_inferior_1): Pass wait status to
	add_current_inferior_and_thread.
---

diff --git a/gdb/remote.c b/gdb/remote.c
index 7ea9597..56f1a49 100644
--- a/gdb/remote.c
+++ b/gdb/remote.c
@@ -3224,13 +3224,57 @@ send_interrupt_sequence (void)
 		    interrupt_sequence_mode);
 }
 
+
+/* If STOP_REPLY is a T stop reply, look for the "thread" register,
+   and extract the PTID.  Returns NULL_PTID if not found.  */
+
+static ptid_t
+stop_reply_extract_thread (char *stop_reply)
+{
+  if (stop_reply[0] == 'T' && strlen (stop_reply) > 3)
+    {
+      char *p;
+
+      /* Txx r:val ; r:val (...)  */
+      p = &stop_reply[3];
+
+      /* Look for "register" named "thread".  */
+      while (*p != '\0')
+	{
+	  char *p1;
+
+	  p1 = strchr (p, ':');
+	  if (p1 == NULL)
+	    return null_ptid;
+
+	  if (strncmp (p, "thread", p1 - p) == 0)
+	    return read_ptid (++p1, &p);
+
+	  p1 = strchr (p, ';');
+	  if (p1 == NULL)
+	    return null_ptid;
+	  p1++;
+
+	  p = p1;
+	}
+    }
+
+  return null_ptid;
+}
+
 /* Query the remote target for which is the current thread/process,
    add it to our tables, and update INFERIOR_PTID.  The caller is
    responsible for setting the state such that the remote end is ready
-   to return the current thread.  */
+   to return the current thread.
+
+   This function is called after handling the '?' or 'vRun' packets,
+   whose response is a stop reply from which we can also try
+   extracting the thread.  If the target doesn't support the explicit
+   qC query, we infer the current thread from that stop reply, passed
+   in in WAIT_STATUS, which may be NULL.  */
 
 static void
-add_current_inferior_and_thread (void)
+add_current_inferior_and_thread (char *wait_status)
 {
   struct remote_state *rs = get_remote_state ();
   int fake_pid_p = 0;
@@ -3238,8 +3282,14 @@ add_current_inferior_and_thread (void)
 
   inferior_ptid = null_ptid;
 
-  /* Now, if we have thread information, update inferior_ptid.  */
+  /* Now, if we have thread information, update inferior_ptid.  First
+     ask the remote which is the current thread, with qC.  If that's
+     not supported, and we have a stop reply handy, maybe it's a T
+     stop reply with a "thread" register.  */
   ptid = remote_current_thread (inferior_ptid);
+  if (ptid_equal (ptid, null_ptid) && wait_status != NULL)
+    ptid = stop_reply_extract_thread (wait_status);
+
   if (!ptid_equal (ptid, null_ptid))
     {
       if (!remote_multi_process_p (rs))
@@ -3400,7 +3450,7 @@ remote_start_remote (int from_tty, struct target_ops *target, int extended_p)
       /* Let the stub know that we want it to return the thread.  */
       set_continue_thread (minus_one_ptid);
 
-      add_current_inferior_and_thread ();
+      add_current_inferior_and_thread (wait_status);
 
       /* init_wait_for_inferior should be called before get_offsets in order
 	 to manage `inserted' flag in bp loc in a correct state.
@@ -7836,7 +7886,7 @@ extended_remote_run (char *args)
 
   if (packet_ok (rs->buf, &remote_protocol_packets[PACKET_vRun]) == PACKET_OK)
     {
-      /* We have a wait response; we don't need it, though.  All is well.  */
+      /* We have a wait response.  All is well.  */
       return 0;
     }
   else if (remote_protocol_packets[PACKET_vRun].support == PACKET_DISABLE)
@@ -7863,6 +7913,10 @@ static void
 extended_remote_create_inferior_1 (char *exec_file, char *args,
 				   char **env, int from_tty)
 {
+  int run_worked;
+  char *stop_reply;
+  struct remote_state *rs = get_remote_state ();
+
   /* If running asynchronously, register the target file descriptor
      with the event loop.  */
   if (target_can_async_p ())
@@ -7873,7 +7927,8 @@ extended_remote_create_inferior_1 (char *exec_file, char *args,
     extended_remote_disable_randomization (disable_randomization);
 
   /* Now restart the remote server.  */
-  if (extended_remote_run (args) == -1)
+  run_worked = extended_remote_run (args) != -1;
+  if (!run_worked)
     {
       /* vRun was not supported.  Fail if we need it to do what the
 	 user requested.  */
@@ -7895,7 +7950,9 @@ extended_remote_create_inferior_1 (char *exec_file, char *args,
       init_wait_for_inferior ();
     }
 
-  add_current_inferior_and_thread ();
+  /* vRun's success return is a stop reply.  */
+  stop_reply = run_worked ? rs->buf : NULL;
+  add_current_inferior_and_thread (stop_reply);
 
   /* Get updated offsets, if the stub uses qOffsets.  */
   get_offsets ();


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]