Bug 862343
| Summary: | assembled code at __restore_rt wastes space | ||
|---|---|---|---|
| Product: | [Fedora] Fedora | Reporter: | John Reiser <jreiser> |
| Component: | glibc | Assignee: | Jeff Law <law> |
| Status: | CLOSED NOTABUG | QA Contact: | Fedora Extras Quality Assurance <extras-qa> |
| Severity: | unspecified | Docs Contact: | |
| Priority: | unspecified | ||
| Version: | 17 | CC: | fweimer, jakub, law, pfrankli, schwab, spoyarek |
| Target Milestone: | --- | ||
| Target Release: | --- | ||
| Hardware: | x86_64 | ||
| OS: | Linux | ||
| Whiteboard: | |||
| Fixed In Version: | Doc Type: | Bug Fix | |
| Doc Text: | Story Points: | --- | |
| Clone Of: | Environment: | ||
| Last Closed: | 2012-10-02 17:12:04 UTC | Type: | Bug |
| Regression: | --- | Mount Type: | --- |
| Documentation: | --- | CRM: | |
| Verified Versions: | Category: | --- | |
| oVirt Team: | --- | RHEL 7.3 requirements from Atomic Host: | |
| Cloudforms Team: | --- | Target Upstream Version: | |
| Embargoed: | |||
Changing it would break libgcc and unwinding (at least on kernels that don't provide vDSO, but if they provide vDSO, __restore_rt isn't used at all anyway).
gcc/libgcc/config/i386/linux-unwind.h relies on the exact insn sequence:
/* movq $__NR_rt_sigreturn, %rax ; syscall. */
#ifdef __LP64__
#define RT_SIGRETURN_SYSCALL 0x050f0000000fc0c7ULL
#else
#define RT_SIGRETURN_SYSCALL 0x050f40000201c0c7ULL
#endif
if (*(unsigned char *)(pc+0) == 0x48
&& *(unsigned long long *)(pc+1) == RT_SIGRETURN_SYSCALL)
(In reply to comment #1) > if [the kernel] provides vDSO, [then] __restore_rt isn't used at all anyway. That is not true in kernel-3.5.3-1.fc17.x86_64 glibc-2.15-56.fc17.x86_64 __restore_rt is used for many signal returns, even when [vdso] is present and used. Here's a test case: ----- my_test.c #include <signal.h> #include <stdio.h> #include <string.h> void my_handler(int signo, siginfo_t *si, void *v) { struct ucontext *uc = v; printf("signo=%d si=%p uc=%p\n", signo, si, uc); } main() { struct sigaction act; struct sigaction old; act.sa_sigaction = my_handler; memset(&act.sa_mask, 0, sizeof(act.sa_mask)); act.sa_flags = 0; sigaction (SIGRTMIN, &act, &old); kill(getpid(), SIGRTMIN); return 0; } ----- $ gcc -o my_test -g my_test.c $ gdb ./my_test (gdb) b main (gdb) run Breakpoint 1, main () at my_test.c:16 (gdb) x/i __restore_rt 0x3e8f8359a0 <__restore_rt>: mov $0xf,%rax (gdb) b *0x3e8f8359a0 (gdb) c Program received signal SIG34, Real-time event 34. (gdb) c Continuing. signo=34 si=0x7fffffffdb30 uc=0x7fffffffda00 Breakpoint 2, <signal handler called> (gdb) x/i $pc => 0x3e8f8359a0 <__restore_rt>: mov $0xf,%rax <<<<< __restore_rt is used (gdb) info proc process 4312 (gdb) shell cat /proc/4312/maps 7ffff7ffe000-7ffff7fff000 r-xp 00000000 00:00 0 [vdso] 7ffffffde000-7ffffffff000 rw-p 00000000 00:00 0 [stack] ffffffffff600000-ffffffffff601000 r-xp 00000000 00:00 0 [vsyscall] gdb) shell od -Ax -tx4 /proc/4312/auxv 000000 00000021 00000000 f7ffe000 00007fff <<<<< [vdso] is present at 0x7ffff7ffe000 $ grep 'AT_.*33' /usr/include/elf.h #define AT_SYSINFO_EHDR 33 $ |
Description of problem: On x86_64 the assembled instructions at __restore_rt are functionally correct, but waste space. (gdb) x/4i __restore_rt 0x3e8f8359a0 <__restore_rt>: mov $0xf,%rax 0x3e8f8359a7 <__restore_rt+7>: syscall 0x3e8f8359a9 <__restore_rt+9>: nopl 0x0(%rax) 0x3e8f8359b0 <__GI___libc_sigaction>: sub $0xd0,%rsp The first instruction "mov $0xf,%rax" takes seven bytes: (gdb) x/xg __restore_rt 0x3e8f8359a0 <__restore_rt>: 0x0f0000000fc0c748 using a REX prefix byte of 0x48 to force 64-bit width, and the 0xC7 opcode which is "mov 32-bit immediate to any destination". Five bytes would be sufficient: movl $0xf,%eax ### B8 0F 00 00 00 which uses the specific "mov 32-bit immediate to register EAX" and the implicit zero extension from 32 bits into a 64-bit register. Because of the two extra bytes, then the 8-byte alignment for the next routine __GI___libc_sigaction requires an additional 7 bytes for the nopl. Thus __restore_rt occupies 16 bytes when 8 would suffice. [How many other instances are there of such waste?] Version-Release number of selected component (if applicable): glibc-2.15-56.fc17.x86_64 How reproducible: always Steps to Reproduce: 1. $ gdb /lib64/libc.so.6 2. (gdb) x/4i __restore_rt 3. (gdb) x/xg __restore_rt Actual results: The first instruction "mov $0xf,%rax" occupies 7 bytes using REX prefix and 0xC7 opcode. Alignment of the next routine requires a 7-byte NOP. Expected results: The first instruction "movl $0xf,%eax" achieves the same effect in only 5 bytes. Alignment of the next routine requires only a 1-byte NOP. Additional info: The source file is sysdeps/unix/sysv/linux/x86_64/sigaction.c