From: Jan S. <jst...@re...> - 2012-11-30 15:32:44
|
Hi, I'm occasionally getting core files from clone03/clone06 testcases. The testcase itself gives PASS, it is the child which is randomly crashing. It seems to occur more on single cpu systems. For example: Core was generated by `clone03'. Program terminated with signal 11, Segmentation fault. #0 0x0000000000402bfd in tst_print (tcid=0x403d0e "clone03", tnum=1, ttype=2, tmesg=0x14c6070 "unexpected signal 15 received (pid = 17427).") at tst_res.c:412 412 { (gdb) bt #0 0x0000000000402bfd in tst_print (tcid=0x403d0e "clone03", tnum=1, ttype=2, tmesg=0x14c6070 "unexpected signal 15 received (pid = 17427).") at tst_res.c:412 #1 0x00000000004031be in tst_res (ttype=2, fname=<value optimized out>, arg_fmt=<value optimized out>) at tst_res.c:316 #2 0x0000000000403761 in tst_brk (ttype=2, fname=0x0, func=0x4013d0 <cleanup>, arg_fmt=<value optimized out>) at tst_res.c:640 #3 0x0000000000403960 in tst_brkm (ttype=2, func=0x4013d0 <cleanup>, arg_fmt=<value optimized out>) at tst_res.c:698 #4 0x0000000000403b45 in def_handler (sig=15) at tst_sig.c:248 #5 <signal handler called> #6 0x00000037940db650 in __write_nocancel () at ../sysdeps/unix/syscall-template.S:82 #7 0x000000000040169e in child_fn () at clone03.c:208 #8 0x00000037940e890d in clone () at ../sysdeps/unix/sysv/linux/x86_64/clone.S:115 Dump of assembler code for function tst_print: 0x0000000000402bd0 <+0>: mov %rbx,-0x30(%rsp) 0x0000000000402bd5 <+5>: mov %rbp,-0x28(%rsp) 0x0000000000402bda <+10>: mov %edx,%ebx 0x0000000000402bdc <+12>: mov %r12,-0x20(%rsp) 0x0000000000402be1 <+17>: mov %r13,-0x18(%rsp) 0x0000000000402be6 <+22>: mov %rdi,%r12 0x0000000000402be9 <+25>: mov %r14,-0x10(%rsp) 0x0000000000402bee <+30>: mov %r15,-0x8(%rsp) 0x0000000000402bf3 <+35>: sub $0x2858,%rsp 0x0000000000402bfa <+42>: mov %esi,%r14d => 0x0000000000402bfd <+45>: mov %rcx,0x18(%rsp) (gdb) p $rsp $1 = (void *) 0x14c3800 (gdb) x/1x $rsp 0x14c3800: Cannot access memory at address 0x14c3800 It looks like it receives SIGTERM and while handling SIGTERM it hits SIGSEGV. I don't know what is source of that SIGTERM. I was looking into the second part and looks like the stack for child is not large enough. I modified clone03.c (see attached clone03_poison.patch) to get some extra empty buffer before the child's stack, which was set to pattern 0xDE. Before: |-------------------------------| child_stack child_stack+CHILD_STACK_SIZE After: |---------------------|-------------------------------| poision_start child_stack child_stack+CHILD_STACK_SIZE Now if I start clone03 and kill it I can randomly reproduce the SIGSEGV (attached clone03_kill.sh). The backtrace usually looks like: ... (random place) #5 0x000000000040324e in tst_res (ttype=2, fname=<value optimized out>, arg_fmt=<value optimized out>) at tst_res.c:316 #6 0x00000000004037f1 in tst_brk (ttype=2, fname=0x0, func=0x401420 <cleanup>, arg_fmt=<value optimized out>) at tst_res.c:640 #7 0x00000000004039f0 in tst_brkm (ttype=2, func=0x401420 <cleanup>, arg_fmt=<value optimized out>) at tst_res.c:698 #8 0x0000000000403bd5 in def_handler (sig=13) at tst_sig.c:248 #9 <signal handler called> #10 0x0000003327cdb650 in __write_nocancel () at ../sysdeps/unix/syscall-template.S:82 #11 0x000000000040172e in child_fn () at clone03.c:212 #12 0x0000003327ce890d in clone () at ../sysdeps/unix/sysv/linux/x86_64/clone.S:115 (gdb) p poison_start $1 = (void *) 0xa02010 (gdb) p child_stack $2 = (void *) 0xa03010 (gdb) x/16x poison_start 0xa02010: 0xdededede 0xdededede 0xdededede 0xdededede 0xa02020: 0xdededede 0xdededede 0xdededede 0xdededede 0xa02030: 0xdededede 0xdededede 0xdededede 0xdededede 0xa02040: 0xdededede 0xdededede 0xdededede 0xdededede ... (gdb) 0xa02490: 0xdededede 0xdededede 0xdededede 0xdededede 0xa024a0: 0x00000018 0x00000030 0x00a02800 0x00000000 0xa024b0: 0x00a02740 0x00000000 0xdededede 0xdededede 0xa024c0: 0xdededede 0xdededede 0x27409296 0x00000033 The above shows that 0xDE pattern has been overwritten. Extending child stack helps with the second part: SIGSEGV #define CHILD_STACK_SIZE 16384*4 but I have no idea, where is that first SIGTERM coming from. Any ideas? Regards, Jan |