Check for truncated registers in process_g_packet

[Lionel Flandrin]

On Tue, Oct 18, 2016 at 11:49:01AM -0400, Simon Marchi wrote:
> On 16-10-18 07:10 AM, Lionel Flandrin wrote:
> > Hello,
> > 
> > While investigating an unrelated issue in remote.c I noticed that the
> > bound checking for 'g' packets was bogus:
> > 
> > The previous code would only check that the first byte of the register
> > was within bounds before passing the buffer to regcache_raw_supply. If
> > it turned out that the register in the 'g' packet was incomplete then
> > regcache_raw_supply would proceed to memcpy out-of-bounds.
> > 
> > Since the buffer is allocated with alloca it's relatively unlikely to
> > crash (you just end up dumping gdb's stack into the cache) but it's
> > still a bit messy.
> > 
> > I changed this logic to check for truncated registers and raise an
> > error if one is encountered. Hopefully it should make debugging remote
> > stubs a bit easier.
> 
> Hi Lionel,
> 
> This patch looks good to me, a few minor comments below about formatting.
> Someone else with the approval stamp must look at it, but hopefully it will
> save them a bit of work.

Thank you for the feedback, here's the updated patch:

diff --git a/gdb/ChangeLog b/gdb/ChangeLog
index 4b642b8..3ace874 100644
--- a/gdb/ChangeLog
+++ b/gdb/ChangeLog
@@ -1,3 +1,8 @@
+2016-10-18  Lionel Flandrin  <lionel@svkt.org>
+
+	* remote.c (process_g_packet): Detect truncated registers in 'g'
+	packets and raise an error.
+
 2016-10-18  Maciej W. Rozycki  <macro@imgtec.com>
 
 	* i386-tdep.c (i386_mpx_info_bounds): Make sure the architecture
diff --git a/gdb/remote.c b/gdb/remote.c
index af7508a..e1b5ad7 100644
--- a/gdb/remote.c
+++ b/gdb/remote.c
@@ -7163,18 +7163,31 @@ process_g_packet (struct regcache *regcache)
      the 'p' packet must be used.  */
   if (buf_len < 2 * rsa->sizeof_g_packet)
     {
-      rsa->sizeof_g_packet = buf_len / 2;
+      long sizeof_g_packet = buf_len / 2;
 
       for (i = 0; i < gdbarch_num_regs (gdbarch); i++)
 	{
+	  long offset = rsa->regs[i].offset;
+	  long reg_size = register_size (gdbarch, i);
+
 	  if (rsa->regs[i].pnum == -1)
 	    continue;
 
-	  if (rsa->regs[i].offset >= rsa->sizeof_g_packet)
+	  if (offset >= sizeof_g_packet)
 	    rsa->regs[i].in_g_packet = 0;
+	  else if (offset + reg_size > sizeof_g_packet)
+	    error (_("Truncated register %d in remote 'g' packet"), i);
 	  else
 	    rsa->regs[i].in_g_packet = 1;
 	}
+
+      /* Looks valid enough, we can assume this is the correct length
+         for a 'g' packet.  It's important not to adjust
+         rsa->sizeof_g_packet if we have truncated registers otherwise
+         this "if" won't be run the next time the method is called
+         with a packet of the same size and one of the internal errors
+         below will trigger instead.  */
+      rsa->sizeof_g_packet = sizeof_g_packet;
     }
 
   regs = (char *) alloca (rsa->sizeof_g_packet);
@@ -7204,10 +7217,11 @@ process_g_packet (struct regcache *regcache)
   for (i = 0; i < gdbarch_num_regs (gdbarch); i++)
     {
       struct packet_reg *r = &rsa->regs[i];
+      long reg_size = register_size (gdbarch, i);
 
       if (r->in_g_packet)
 	{
-	  if (r->offset * 2 >= strlen (rs->buf))
+	  if ((r->offset + reg_size) * 2 > strlen (rs->buf))
 	    /* This shouldn't happen - we adjusted in_g_packet above.  */
 	    internal_error (__FILE__, __LINE__,
 			    _("unexpected end of 'g' packet reply"));


-- 
Lionel Flandrin