diff mbox

[patchv2,2/2] Workaround gdbserver<7.7 for setfs

Message ID 20160324223241.GB2548@host1.jankratochvil.net
State New, archived
Headers

Commit Message

Jan Kratochvil March 24, 2016, 10:32 p.m. UTC 
  There was a bug in patchv1.
gdb/ChangeLog
2016-03-24  Jan Kratochvil  <jan.kratochvil@redhat.com>

	* remote.c (packet_ok): Add workaround for PACKET_vFile_setfs.

Comments

Pedro Alves March 30, 2016, 2:17 p.m. UTC | #1 
On 03/24/2016 10:32 PM, Jan Kratochvil wrote:
> There was a bug in patchv1.
> 
> 
> move2.patch
> 

Please include self-contained a commit/rationale along with patch
posts (and reposts).  You had context in your intro to v1 that was
lost in v2.

> 
> gdb/ChangeLog
> 2016-03-24  Jan Kratochvil  <jan.kratochvil@redhat.com>
> 
> 	* remote.c (packet_ok): Add workaround for PACKET_vFile_setfs.
> 
> diff --git a/gdb/remote.c b/gdb/remote.c
> index bb027cf..f80fee8 100644
> --- a/gdb/remote.c
> +++ b/gdb/remote.c
> @@ -1453,7 +1453,15 @@ packet_ok (const char *buf, struct packet_config *config)
>      internal_error (__FILE__, __LINE__,
>  		    _("packet_ok: attempt to use a disabled packet"));
>  
> -  result = packet_check_result (buf);
> +  if (config == &remote_protocol_packets[PACKET_vFile_setfs]
> +      && strcmp (buf, "OK") == 0)
> +    {
> +      /* Workaround gdbserver < 7.7 before its fix from 2013-12-11.  */
> +      result = PACKET_UNKNOWN;
> +    }

This comment could use more detail.

E.g., reading this I'm left wondering, did it always respond OK to
unknown vFile packets, or to all unknown packets?  I think it was
actually the latter.

AFAICS from the commit you pointed at in v1, the "OK" was
gdbserver mistaking any unknown packet for a vStopped packet,
with vStopped being the notification ack for the "%Stop" RSP async
notification.  So it could also happen that gdb sends the setfs
packet while gdbserver had a pending notification, and then
gdbserver replies back a stop reply instead of "OK"...

We may need to guarantee an early enough setfs is attempted.
Is that already the case?

If I'm right and gdbserver mishandled _any_ unknown packet,
then I wonder whether you fix this one, but will trip on another
when you get past initial connection and actually do any serious
debugging?

If not, this may be sufficient.  Otherwise, we may need to come up with
a different workaround, maybe based on sending an early probe packet,
like "MustReplyEmpty", to which well behaved stubs reply empty, just because
that's not a known packet to them.  If a stub replies something other than
empty to that one, then maybe we should disable all other
auto-probed packets...  That may force-disable too much functionality though...

So in sum:

- Expand comment a bit / include commit log with rationale in the patch.

- Give assurance that this is sufficient and that we won't trip on the
  same thing with other packets anyway.  Otherwise we may need to think
  of something else.

Thanks,
Pedro Alves
Jan Kratochvil April 3, 2016, 7:30 p.m. UTC | #2 
On Wed, 30 Mar 2016 16:17:11 +0200, Pedro Alves wrote:
> E.g., reading this I'm left wondering, did it always respond OK to
> unknown vFile packets, or to all unknown packets?  I think it was
> actually the latter.

Yes.


> AFAICS from the commit you pointed at in v1, the "OK" was
> gdbserver mistaking any unknown packet for a vStopped packet,
> with vStopped being the notification ack for the "%Stop" RSP async
> notification.  So it could also happen that gdb sends the setfs
> packet while gdbserver had a pending notification, and then
> gdbserver replies back a stop reply instead of "OK"...

OK, I did not realize this possible regression.


> We may need to guarantee an early enough setfs is attempted.
> Is that already the case?

For linux targets it is because they read:
	/proc/29202/smaps
Although you are right that does not need to be the case for non-linux
targets.  setfs packet seems to be implemented linux-independently.


> If I'm right and gdbserver mishandled _any_ unknown packet,
> then I wonder whether you fix this one, but will trip on another
> when you get past initial connection and actually do any serious
> debugging?

That would mean gdbserver < 7.7 did not work for "any serious debugging".
I have seen the regression only since "setfs" but I admit I did not do "any
serious debugging".


> If not, this may be sufficient.  Otherwise, we may need to come up with
> a different workaround, maybe based on sending an early probe packet,
> like "MustReplyEmpty", to which well behaved stubs reply empty, just because
> that's not a known packet to them.  If a stub replies something other than
> empty to that one, then maybe we should disable all other
> auto-probed packets...  That may force-disable too much functionality though...
> 
> So in sum:

The patch was fixing a common use case with RHEL<=7 targets.  You have
provided out a counterexample that it may hypothetically regress in a racy
case of non-linux FSF gdbserver.  Normally I would find this workaround
applicable only for RHEL GDBs but now that everything needs to be upstream
first the RHEL workaround needs to be implemented in FSF GDB first (where it
does not belong much IMNSHO).  I will therefore try to implement the
"MustReplyEmpty" packet but I have no idea what effect will have your
mentioned "disable all other auto-probed packets".


Thanks,
Jan
diff mbox

Patch

diff --git a/gdb/remote.c b/gdb/remote.c
index bb027cf..f80fee8 100644
--- a/gdb/remote.c
+++ b/gdb/remote.c
@@ -1453,7 +1453,15 @@  packet_ok (const char *buf, struct packet_config *config)
     internal_error (__FILE__, __LINE__,
 		    _("packet_ok: attempt to use a disabled packet"));
 
-  result = packet_check_result (buf);
+  if (config == &remote_protocol_packets[PACKET_vFile_setfs]
+      && strcmp (buf, "OK") == 0)
+    {
+      /* Workaround gdbserver < 7.7 before its fix from 2013-12-11.  */
+      result = PACKET_UNKNOWN;
+    }
+  else
+    result = packet_check_result (buf);
+
   switch (result)
     {
     case PACKET_OK: