The .eh_frame sections for the nptl single argument syscalls (close and fsync) in glibc 2.3.6 are being created incorrectly. Valgrind compains about this, as discussed here: http://sourceforge.net/mailarchive/forum.php?thread_id=7568951&forum_id=32038 valgrind emits the message: DWARF2 CFI reader: unhandled CFI instruction 0:50 on startup of every program linking to the offending object. This creates a lot of noise. readelf agrees that there is a problem with the sections: $ readelf --debug-dump=frames /lib/tls/libc.so.6 > /dev/null unsupported or unknown DW_CFA_50 unsupported or unknown DW_CFA_50 $ The patch I will attach shortly fixes the problem. At least the error message goes away and to me the CFI instructions now look correct. FWIW, I noticed the problem on Debian stable (sarge) and current unstable (sid) but it doesn't look like a Debian only issue.
Created attachment 903 [details] patch to fix this bug The patch above is far enough removed from the final objects that the correctness of the fix isn't obvious (to me anyway). To verify the fix I took the output object and steped through the FDE instructions to create the CFI table. Then I checked that the CFI table corresponds to the subroutine in question. $ ar -x /usr/lib/nptl/libc.a close.o $ readelf --debug-dump=frames close.o The section .eh_frame contains: 00000000 00000010 00000000 CIE Version: 1 Augmentation: "" Code alignment factor: 1 Data alignment factor: -4 Return address column: 8 DW_CFA_def_cfa: r4 ofs 4 DW_CFA_offset: r8 at cfa-4 DW_CFA_nop DW_CFA_nop 00000014 00000020 00000018 FDE cie=00000000 pc=00000000..00000050 DW_CFA_advance_loc: 12 to 0000000c DW_CFA_register: r3 in r2 DW_CFA_advance_loc: 13 to 00000019 DW_CFA_restore: r3 DW_CFA_advance_loc: 21 to 0000002e DW_CFA_register: r3 in r2 DW_CFA_advance_loc: 13 to 0000003b DW_CFA_restore: r3 DW_CFA_advance_loc: 1 to 0000003c DW_CFA_def_cfa_offset: 8 DW_CFA_advance_loc: 8 to 00000044 DW_CFA_def_cfa_offset: 4 DW_CFA_nop DW_CFA_nop $ In order to decode the above I looked at sections 6.4 and Appendix 5 of http://www.arm.com/pdfs/TIS-DWARF2.pdf. The CFI table described above is show below. I also filled in what I believe to be the CFI to x86 register mapping and the callee saved registers. fp fp ra eax ecx edx ebx esp ebp esi edi eip Loc CFA r0 r1 r2 r3 r4 r5 r6 r7 r8 0 r4+4 u u u s u s s s cfa-4 c r2 19 s 2e r2 3b s 3c r4+8 44 r4+4 key: = (blank) same as last (look up in same column to find value) s = same_value rule u = undefined rule rN = register(N) rule $ objdump -j .text -D close.o close.o: file format elf32-i386 Disassembly of section .text: 00000000 <__libc_close>: 0: 65 83 3d 0c 00 00 00 cmpl $0x0,%gs:0xc 7: 00 8: 75 1b jne 25 <__close_nocancel+0x1b> 0000000a <__close_nocancel>: a: 89 da mov %ebx,%edx c: 8b 5c 24 04 mov 0x4(%esp),%ebx 10: b8 06 00 00 00 mov $0x6,%eax 15: cd 80 int $0x80 17: 89 d3 mov %edx,%ebx 19: 3d 01 f0 ff ff cmp $0xfffff001,%eax 1e: 0f 83 fc ff ff ff jae 20 <__close_nocancel+0x16> 24: c3 ret 25: e8 fc ff ff ff call 26 <__close_nocancel+0x1c> 2a: 89 c1 mov %eax,%ecx 2c: 89 da mov %ebx,%edx 2e: 8b 5c 24 04 mov 0x4(%esp),%ebx 32: b8 06 00 00 00 mov $0x6,%eax 37: cd 80 int $0x80 39: 89 d3 mov %edx,%ebx 3b: 50 push %eax 3c: 89 c8 mov %ecx,%eax 3e: e8 fc ff ff ff call 3f <__close_nocancel+0x35> 43: 58 pop %eax 44: 3d 01 f0 ff ff cmp $0xfffff001,%eax 49: 0f 83 fc ff ff ff jae 4b <__close_nocancel+0x41> 4f: c3 ret
Irrelevant patch. We use a completely different method now. This doesn't mean the patch is correct and would have been applied.