On Fri, Mar 27, 2015 at 9:00 PM, Linus Torvalds <torva...@linux-foundation.org> wrote: > On Fri, Mar 27, 2015 at 7:25 AM, Denys Vlasenko <dvlas...@redhat.com> wrote: >> >> Apparently, users *don't* depend on arithmetic flags >> to survive over syscall. They also okay with DF flag >> being cleared. > > Generally, users probably dont' care about many registers at all being > saved, but it's worth noting that the reason system calls save/restore > even caller-saved registers is at least partly in order to avoid any > kernel information leaks. > > I don't believe that user mode will ever reasonably care about the > arithmetic flags being changed, but at the same time I also don't it > is something we should ever consider a "feature" we should try to take > advantage of. Generally we should try to not mess with the flag state, > and I'd *much* rather make the rule be that all the system call return > paths restore flags as much as possible.
"We don't clobber anything" ABI has its appeal. OTOH, fulfilling ABI's promises has cost which hast to be paid on every syscall, regardless whether userspace needed it or not. Example. This is the uclibc implementation of write(): 00000000004acfc4 <__libc_write>: 4acfc4: 53 push %rbx 4acfc5: 48 63 ff movslq %edi,%rdi 4acfc8: b8 01 00 00 00 mov $0x1,%eax 4acfcd: 0f 05 syscall 4acfcf: 48 89 c3 mov %rax,%rbx 4acfd2: 48 81 fb 00 f0 ff ff cmp $0xfffffffffffff000,%rbx 4acfd9: 76 0f jbe 4acfea <__libc_write+0x26> 4acfdb: e8 64 15 00 00 callq 4ae544 <__GI___errno_location> 4acfe0: 89 da mov %ebx,%edx 4acfe2: f7 da neg %edx 4acfe4: 89 10 mov %edx,(%rax) 4acfe6: 48 83 c8 ff or $0xffffffffffffffff,%rax 4acfea: 5b pop %rbx 4acfeb: c3 retq This is a C function. Therefore any its caller assumes that C-clobbered registers can be, indeed, clobbered here, so if that caller uses any of them, it saves/restores them. All efforts by kernel code to save/restore C-clobbered registers, eight of them, are in vain. It's just useless work. Userspace does not benefit from that effort. If our syscall ABI would say that those regs are not preserved, we could have a bit faster syscalls. Any userspace code which really had to have those registers preserved across a particular syscall, could push/pop them itself. -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
