Re: recent SIGBUS/SIGSEGV mips kernel bug

David Daney <ddaney@xxxxxxxxxxxxxxxxxx> · Wed, 20 Oct 2010 09:38:54 -0700

On 10/20/2010 08:22 AM, Camm Maguire wrote:
Greetings!  Recent mips (aka gabrielli sid) appears to have a new
SIGBUS/SIGSEGV bug.  Briefly, attempted writes to pages mprotected
read-only occasionally pass a SIGBUS instead of SIGSEGV to the
sigaction handler.  In such an instance, the code of the siginfo
structure is 128, and the fault address is not filled in.  This
behavior appears to hinge on writing from the floating point registers
into the protected memory.

What processor is it running on?  Specifically does it have hardware 
floating point, or are the floating point instructions being emulated in 
the kernel?

It is conceivable that it is an FPU emulator bug.

David Daney

strace:

30717 mprotect(0x2f59000, 4096, PROT_READ|PROT_EXEC) = 0
30717 mprotect(0x2f5a000, 745472, PROT_READ|PROT_WRITE|PROT_EXEC) = 0
30717 mprotect(0x3010000, 323584, PROT_READ|PROT_EXEC) = 0
30717 mprotect(0x305f000, 548864, PROT_READ|PROT_WRITE|PROT_EXEC) = 0
30717 mprotect(0x30e5000, 1220608, PROT_READ|PROT_EXEC) = 0
30717 mprotect(0x320f000, 53559296, PROT_READ|PROT_WRITE|PROT_EXEC) = 0
30717 mprotect(0x6523000, 53739520, PROT_READ|PROT_EXEC) = 0
30717 mprotect(0x9863000, 60375040, PROT_READ|PROT_WRITE|PROT_EXEC) = 0
30717 mprotect(0x6523000, 4096, PROT_READ|PROT_WRITE|PROT_EXEC) = 0
30717 --- SIGSEGV (Segmentation fault) @ 0 (0) ---  /* good */
30717 mprotect(0x31fb000, 4096, PROT_READ|PROT_WRITE|PROT_EXEC) = 0
30717 rt_sigreturn(0xdc5e80)            = -1 EPERM (Operation not permitted)
30717 --- SIGSEGV (Segmentation fault) @ 0 (0) ---  /* good */
30717 mprotect(0x9862000, 4096, PROT_READ|PROT_WRITE|PROT_EXEC) = 0
30717 rt_sigreturn(0xdc5e80)            = -1 EBADF (Bad file descriptor)
30717 --- SIGBUS (Bus error) @ 0 (0) ---  /* bad handler call */

Here are the gdb details:

(gdb) frame 28
#28 0x004a3e98 in memprotect_handler (sig=10, code=2140765848,
     scp=0x7f997f18, addr=0xfdb6c0 "\b") at sgbc.c:1687
1687	  segmentation_catcher(0);
(gdb) p faddr
$1 = 0x0
(gdb) p/x *((siginfo_t *)code )
$2 = {si_signo = 0xa, si_code = 0x80, si_errno = 0x0, __pad0 = 0x7f997e98,
   _sifields = {_pad = {0x0, 0x0, 0x7f997eb4, 0x7f997fd4, 0x2167b80,
       0x233c5b8, 0xde5e80, 0xbbec80, 0x65f8a80, 0x7f997ed0, 0x4951a8,
       0xbbec80, 0xffffffff, 0x7f997ee0, 0x495170, 0xbbec80, 0x65f8a48,
       0x7f997ef0, 0x4951a8, 0xbbec80, 0xffffffff, 0x7f997f00, 0x495170,
       0xbbec80, 0x31ecc48, 0x7f997f10, 0xfdead4, 0x7f997f18, 0x454ebc},
     _kill = {si_pid = 0x0, si_uid = 0x0}, _timer = {si_tid = 0x0,
       si_overrun = 0x0, si_sigval = {sival_int = 0x7f997eb4,
         sival_ptr = 0x7f997eb4}}, _rt = {si_pid = 0x0, si_uid = 0x0,
       si_sigval = {sival_int = 0x7f997eb4, sival_ptr = 0x7f997eb4}},
     _sigchld = {si_pid = 0x0, si_uid = 0x0, si_status = 0x7f997eb4,
       si_utime = 0x7f997fd4, si_stime = 0x2167b80}, _sigfault = {
       si_addr = 0x0}, _sigpoll = {si_band = 0x0, si_fd = 0x0}}}
(gdb) up
#29<signal handler called>
(gdb) up
#30 0x004484e8 in fSaset1 (x=0x29b5930, i=0, val=0x2d8d870) at array.c:231
231	      (x->lfa.lfa_self[i]) = Mlf(val);
(gdb) p/x x->lfa.lfa_self
$3 = 0x9861ff0
(gdb) disassemble
    0x004484d8<+1352>:	lwc1	$f0,12(v1)
    0x004484dc<+1356>:	move	at,at
    0x004484e0<+1360>:	lwc1	$f1,8(v1)
    0x004484e4<+1364>:	move	at,at
=>  0x004484e8<+1368>:	swc1	$f0,4(v0)
    0x004484ec<+1372>:	swc1	$f1,0(v0)
(gdb) i reg v1
v1: 0x2d8d870
(gdb) i reg f1
f1:  0x3fa2056c flt: 1.26579046
(gdb) i reg f0
f0:  0x8f1afe4f flt: -7.64176427e-30   dbl: 0.035197632283140316
(gdb) i reg v0
v0: 0x9861ff0

How should this be filed?

Take care,