[uml-devel] [PATCH] fix for reboot panic

[Steve Bennett <steveb@snapgear.com>]

[-- Attachment #1: Type: text/plain, Size: 2684 bytes --]

With a certain configuration, when I run con=xterm
rebooting a UML instance causes a panic.
This is against 2.4.22-6

I traced this to the following problem.

During reboot, close_chan() is called to close all the vty channels.
When a channel is connected to an xterm, this involves killing the
xterm process, which generates a SIGINT on the channel.
This causes chan_interrupt() to close the channel itself, calling close_chan()
and then free_chan().
When the original close_chan() is resumed, the channel memory is now invalid.

The attached patch ensures that no access is required to the channel after
calling the channel close handler.

(gdb) where
#0  close_chan (chans=0xa0230088) at chan_kern.c:178
#1  0xa00bda49 in chan_interrupt (chans=0xa0230088, task=0xa02300bc,
     tty=0xa09b9000, irq=2, dev=0xa0230080) at chan_kern.c:495
#2  0xa00be298 in line_interrupt (irq=2, data=0xa0230080, unused=0xa3a98274)
     at line.c:28
#3  0xa007aa0e in handle_IRQ_event (irq=2, regs=0xa3a98274, action=0xa033efe0)
     at irq.c:160
#4  0xa007ac1f in do_IRQ (irq=2, regs=0xa3a98274) at irq.c:323
#5  0xa007b47c in sigio_handler (sig=29, regs=0xa3a98274) at irq_user.c:73
#6  0xa0086b4a in sig_handler_common_tt (sig=29, sc_ptr=0xa3a9b748)
     at trap_user.c:37
#7  0xa0083023 in sig_handler (sig=29, sc=
       {gs = 0, __gsh = 0, fs = 0, __fsh = 0, es = 43, __esh = 0, ds = 43, __dsh 
= 0, edi = 2745809964, esi = 33, ebp = 2745809468, esp = 2745809440,
#8  <signal handler called>
#9  0xa018de51 in kill ()
#10 0xa008ac72 in os_kill_process (pid=16800, reap_child=1) at process.c:87
#11 0xa00bba8c in xterm_close (fd=33, d=0xa037ce40) at xterm.c:159
#12 0xa00bccb6 in close_chan (chans=0xa0230088) at chan_kern.c:182
#13 0xa00be9bc in close_lines (lines=0xa0230080, nlines=8) at line.c:311
#14 0xa00bc74d in console_exit () at stdio_console.c:244
#15 0xa007d882 in do_uml_exitcalls () at process_kern.c:327
#16 0xa007e7ca in machine_restart (__unused=0x0) at reboot.c:47
#17 0xa00211df in sys_reboot (magic1=-18751827, magic2=672274793,
     cmd=19088743, arg=0xf) at sys.c:308
#18 0xa00857be in execute_syscall_tt (r=0xa3a98274) at syscall_kern.c:120
#19 0xa00857fe in syscall_handler_tt (sig=12, regs=0xa3a98274)
     at syscall_user.c:37
#20 0xa0086b4a in sig_handler_common_tt (sig=12, sc_ptr=0xa3a9bd28)
     at trap_user.c:37
#21 0xa0083023 in sig_handler (sig=12, sc=
       {gs = 0, __gsh = 0, fs = 0, __fsh = 0, es = 43, __esh = 0, ds = 43, __dsh 
= 0, edi = 4276215469, esi = 15, ebp = 2684354108, esp = 2684354028,
     at trap_user.c:101
(gdb)

-- 
Steve Bennett                 http://www.SnapGear.com/
steveb@snapgear.com           Custom Embedded Solutions + Security

[-- Attachment #2: uml-reboot-fix.patch --]
[-- Type: text/plain, Size: 1036 bytes --]

diff  -u linux-2.4.x/arch/um/drivers/chan_kern.c
--- linux-2.4.x/arch/um/drivers/chan_kern.c.orig
+++ linux-2.4.x/arch/um/drivers/chan_kern.c	2003-12-03 10:49:11.000000000 +1000
@@ -175,13 +175,21 @@
 	 * state.  Then, the first one opened will have the original state,
 	 * so it must be the last closed.
 	 */
-        for(ele = chans->prev; ele != chans; ele = ele->prev){
-                chan = list_entry(ele, struct chan, list);
-		if(!chan->opened) continue;
-		if(chan->ops->close != NULL)
-			(*chan->ops->close)(chan->fd, chan->data);
+	for(ele = chans->prev; ele != chans; ele = prev){
+		/* Note that the close function can trigger a signal
+		 * which will close this channel from under us, so
+		 * we need to do everything for the channel
+		 * before we call close */
+		int fd;
+		prev = ele->prev;
+
+		chan = list_entry(ele, struct chan, list);
+		if (!chan->opened) continue;
+		fd = chan->fd;
 		chan->opened = 0;
 		chan->fd = -1;
+		if(chan->ops->close != NULL)
+			(*chan->ops->close)(fd, chan->data);
 	}
 }