Bug #9654: Segmentation fault while printing out C level backtrace information

[Rei Odaira]

When SIGSEGV happens, C level backtrace information should be printed out, but the printing-out itself causes another segmentation fault.

$ ./ruby -e 'Process.kill :SEGV, $$'
-e:1: [BUG] Segmentation fault at 0x00584f
ruby 2.2.0dev (2014-03-19) [powerpc64-linux]

-- Control frame information -----------------------------------------------
c:0003 p:---- s:0009 e:000008 CFUNC  :kill
c:0002 p:0015 s:0004 E:00153c EVAL   -e:1 [FINISH]
c:0001 p:0000 s:0002 E:002574 TOP    [FINISH]

-- Ruby level backtrace information ----------------------------------------
-e:1:in `<main>'
-e:1:in `kill'

-- C level backtrace information -------------------------------------------
./ruby(Segmentation fault

This second segmentation fault happens at the following stack context.

(gdb) bt
#0  0x201ba994 in strlen () from /lib/libc.so.6
#1  0x2070cbe0 in kvprintf (fmt=0x207a097d "+0x%lx) [0x%lx] %s:%d\n")
    at addr2line.c:1009
#2  kprintf (fmt=0x207a097d "+0x%lx) [0x%lx] %s:%d\n") at addr2line.c:771
#3  0x2070e4f8 in rb_dump_backtrace_with_lines (num_traces=18, 
    traces=0x2081762c, syms=0x20a7d720) at addr2line.c:677
#4  0x206f3ffc in rb_print_backtrace () at vm_dump.c:690
#5  rb_vm_bugreport () at vm_dump.c:825
#6  0x207621ac in report_bug (file=<value optimized out>, 
    line=<value optimized out>, fmt=0x2079857c "Segmentation fault at %p", 
    args=0x2085f864) at error.c:312
#7  0x207624e4 in rb_bug (fmt=0x2079857c "Segmentation fault at %p")
    at error.c:339
#8  0x206664e0 in sigsegv (sig=<value optimized out>, info=0x2085f8f0, 
    ctx=<value optimized out>) at signal.c:704
#9  <signal handler called>
#10 0x2016674c in kill () from /lib/libc.so.6
#11 0x20702c64 in ruby_kill (pid=<value optimized out>, 
    sig=<value optimized out>) at thread.c:5185
<<<<< snip >>>>>

This error began to occur after this change:
http://www.rubyist.net/~kanemoto/chkbuild/plinux/ruby-trunk/log/20140314T070002Z.diff.html.gz
Due to this error, TestBugReporter#test_bug_reporter_add fails on ppc64 GNU/Linux.
My guess is that the changes in addr2line.c are doing something, but I am not sure.

The second segmentation fault is caused because line->sname points to out-of-range memory.
Tracing rb_dump_backtrace_with_lines() and fill_lines(), I found the sname entry was first set correctly by reading the ./ruby file, but it was later overwritten by some incorrect information while reading the /usr/lib/debug/lib/libc-2.5.so.debug file.
In libc-2.5.so.debug, there seem to be several symbol table entries whose st_size is quite big (~1.5 GB), so those entries happen to cover all the addresses in traces[], which results in overwritting sname at the line 584 of addr2line.c.
I am not familiar with ELF, so I cannot track down further.
Hope this report helps.

[Nobuyoshi Nakada]

• Description updated (diff)
• Category set to core
• Status changed from Open to Assigned
• Assignee set to Yui NARUSE
• Target version set to current: 2.2.0
• Backport changed from 2.0.0: UNKNOWN, 2.1: UNKNOWN to 2.0.0: DONTNEED, 2.1: DONTNEED

[Yui NARUSE]

• Status changed from Assigned to Closed
• % Done changed from 0 to 100

Applied in changeset r45383.

---

• addr2line.c (fill_lines): compare the file names of object in which symbols exist. [Bug #9654]

[Rei Odaira]

Thanks for the quick fix. The second segmentation fault no longer happens.

Unfortunately, I found another problem regarding the C-level backtrace. Here is what I see in my environment (ruby 2.2.0dev (2014-03-23) [powerpc64-linux]):

$ ./ruby -e 'Process.kill :SEGV, $$'
-e:1: [BUG] Segmentation fault at 0x004731
ruby 2.2.0dev (2014-03-23) [powerpc64-linux]

-- Control frame information -----------------------------------------------
c:0003 p:---- s:0009 e:000008 CFUNC  :kill
c:0002 p:0015 s:0004 E:001b24 EVAL   -e:1 [FINISH]
c:0001 p:0000 s:0002 E:002384 TOP    [FINISH]

-- Ruby level backtrace information ----------------------------------------
-e:1:in `<main>'
-e:1:in `kill'

-- C level backtrace information -------------------------------------------
./ruby(rb_vm_bugreport+0x134) [0x20294264] vm_dump.c:58285
./ruby(report_bug+0x1ac) [0x2030250c] error.c:242232
./ruby(rb_bug+0xc4) [0x20302844] error.c:255315
./ruby(sigsegv+0xb0) [0x202064f0] signal.c:15040
(__kernel_sigtramp_rt32+0x0) [0x100360]
./ruby(ruby_kill+0x118) [0x202a2ee8] thread.c:495680
./ruby(rb_f_kill+0x98) [0x202074f8] signal.c:97721
./ruby(call_cfunc_m1+0x2c) [0x2027294c] vm_insnhelper.c:10546
./ruby(vm_call_cfunc_with_frame+0x14c) [0x2027c5ac] vm_insnhelper.c:768981
./ruby(vm_call_method+0x580) [0x20291c60] vm_method.c:2680506
./ruby(vm_exec_core+0x2288) [0x202899a8] vm.inc:1934715
./ruby(vm_exec+0x6f8) [0x2028d8c8] vm.c:2278215
./ruby(rb_iseq_eval_main+0x2f8) [0x2028def8] vm_insnhelper.c:2310457
./ruby(ruby_exec_internal+0xe4) [0x20131884] eval.c:153339
./ruby(ruby_run_node+0xa4) [0x201336c4] eval.c:326460
./ruby(main+0x5c) [0x2012f77c] main.c:1828
<<<<< snip >>>>>

The line numbers are incorrect. In DWARF, line_base is signed 8 bits, but parse_debug_line_cu() reads line_base as char, which is not guaranteed as signed in the C language. The following patch will fix this.

--- addr2line.c (revision 45390)
+++ addr2line.c (working copy)
@@ -294,7 +294,7 @@
     is_stmt = default_is_stmt = *(unsigned char *)p;
     p++;

-    line_base = *(char *)p;
+    line_base = *(signed char *)p;
     p++;

     line_range = *(unsigned char *)p;