diff mbox

Fix gdbserver qGetTLSAddr for x86_64 -m32

Message ID 20140410114901.GA16411@host2.jankratochvil.net
State Committed
Headers

Commit Message

Jan Kratochvil April 10, 2014, 11:49 a.m. UTC 
  Hi,

gdbserver makes libthread_db to access uninitialized memory.  Surprisingly it
does not harm normally, even -fsanitize=address works with current gdbserver.
I have found just valgrind detects it as a very first warning for gdbserver:

Syscall param ptrace(addr) contains uninitialised byte(s)
   at 0x3721EECEBE: ptrace (ptrace.c:45)
   by 0x436EE5: ps_get_thread_area (linux-x86-low.c:252)
   by 0x5559D02: __td_ta_lookup_th_unique (td_ta_map_lwp2thr.c:157)
   by 0x5559EC3: td_ta_map_lwp2thr (td_ta_map_lwp2thr.c:207)
   by 0x43F87D: find_one_thread (thread-db.c:281)
   by 0x440038: thread_db_get_tls_address (thread-db.c:505)
   by 0x40F6D0: handle_query (server.c:2004)
   by 0x4124CF: process_serial_event (server.c:3445)
   by 0x4136B6: handle_serial_event (server.c:3889)
   by 0x419571: handle_file_event (event-loop.c:434)
   by 0x418D38: process_event (event-loop.c:189)
   by 0x419AB7: start_event_loop (event-loop.c:552)
   
Reproducible with:
cd gdb/testsuite
g++ -o gdb.threads/tls gdb.threads/tls{,2}.c -m32 -pthread
../gdbserver/gdbserver :1234 gdb.threads/tls
../gdb -batch gdb.threads/tls -ex 'target remote :1234' -ex 'b spin' -ex c -ex 'p a_thread_local'

It is more easily reproducible even without valgrind using s/0x00/0xff/ in the
attached patch.  It will then turn the output of reproducer above:
$1 = 0
->
Cannot find thread-local storage for Thread 29044, executable file .../gdb/testsuite/gdb.threads/tls:
Remote target failed to process qGetTLSAddr request


Thanks,
Jan
gdb/gdbserver/
2014-04-10  Jan Kratochvil  <jan.kratochvil@redhat.com>

	Fix gdbserver qGetTLSAddr for x86_64 -m32.
	* linux-x86-low.c (X86_64_USER_REGS): New.
	(x86_fill_gregset): Call memset for BUF first in x86_64 -m32 case.

Comments

Tom Tromey May 19, 2014, 3:07 p.m. UTC | #1 
>>>>> "Jan" == Jan Kratochvil <jan.kratochvil@redhat.com> writes:

Jan> gdbserver makes libthread_db to access uninitialized memory.
Jan> Surprisingly it does not harm normally, even -fsanitize=address
Jan> works with current gdbserver.

Jan> 2014-04-10  Jan Kratochvil  <jan.kratochvil@redhat.com>

Jan> 	Fix gdbserver qGetTLSAddr for x86_64 -m32.
Jan> 	* linux-x86-low.c (X86_64_USER_REGS): New.
Jan> 	(x86_fill_gregset): Call memset for BUF first in x86_64 -m32 case.

Ok.

Tom
Jan Kratochvil May 19, 2014, 6:21 p.m. UTC | #2 
On Mon, 19 May 2014 17:07:49 +0200, Tom Tromey wrote:
> >>>>> "Jan" == Jan Kratochvil <jan.kratochvil@redhat.com> writes:
> Jan> 2014-04-10  Jan Kratochvil  <jan.kratochvil@redhat.com>
> 
> Jan> 	Fix gdbserver qGetTLSAddr for x86_64 -m32.
> Jan> 	* linux-x86-low.c (X86_64_USER_REGS): New.
> Jan> 	(x86_fill_gregset): Call memset for BUF first in x86_64 -m32 case.
> 
> Ok.

Checked in:
	9e0aa64f5510861b2517c5841b59adde8e423540


Jan
diff mbox

Patch

diff --git a/gdb/gdbserver/linux-x86-low.c b/gdb/gdbserver/linux-x86-low.c
index 33b5f26..1156e58 100644
--- a/gdb/gdbserver/linux-x86-low.c
+++ b/gdb/gdbserver/linux-x86-low.c
@@ -185,6 +185,7 @@  static const int x86_64_regmap[] =
 };
 
 #define X86_64_NUM_REGS (sizeof (x86_64_regmap) / sizeof (x86_64_regmap[0]))
+#define X86_64_USER_REGS (GS + 1)
 
 #else /* ! __x86_64__ */
 
@@ -343,6 +344,10 @@  x86_fill_gregset (struct regcache *regcache, void *buf)
 	  collect_register (regcache, i, ((char *) buf) + x86_64_regmap[i]);
       return;
     }
+
+  /* 32-bit inferior registers need to be zero-extended.
+     Callers would read uninitialized memory otherwise.  */
+  memset (buf, 0x00, X86_64_USER_REGS * 8);
 #endif
 
   for (i = 0; i < I386_NUM_REGS; i++)