Messages in this thread | | | From | Andy Lutomirski <> | Date | Wed, 25 Mar 2015 08:17:58 -0700 | Subject | Re: [PATCH] x86: vdso32/syscall.S: do not load __USER32_DS to %ss |
| |
On Wed, Mar 25, 2015 at 8:03 AM, Denys Vlasenko <dvlasenk@redhat.com> wrote: > On 03/25/2015 10:28 AM, Ingo Molnar wrote: >> >> * Andy Lutomirski <luto@amacapital.net> wrote: >> >>> Now we can do a fun hack on top. On Intel, we have >>> sysenter/sysexitl and, on AMD, we have syscall/sysretl. But, if I >>> read the docs right, Intel has sysretl, too. So we can ditch >>> sysexit entirely, since this mechanism no longer has any need to >>> keep the entry and exit conventions matching. >> >> So this only affects 32-bit vdsos, because on 64-bit both Intel and >> AMD have and use SYSCALL/SYSRET. >> >> So my question would be: what's the performance difference between >> INT80 and sysenter entries on 32-bit, on modern CPUs? >> >> If it's not too horrible (say below 100 cycles) then we could say that >> we start out the simplification and robustification by switching Intel >> over to INT80 + SYSRET on 32-bit, and once we know the 32-bit SYSRET >> and all the other simplifications work fine we implement the >> SYSENTER-hack on top of that? > > int 0x80 is about 250 cycles slower than syscall/sysenter. > (I mean, the instruction per se, not the full round-trip). > This looks too horrible to ignore :(
Agreed.
> > >> Is there any user-space code that relies on being able to execute an >> open coded SYSENTER, or are we shielded via the vDSO? > > Userspace can't use open-coded sysenter. It will return to a different > address. > > Userspace _can_ do this: > > my_sysenter: > push %ecx > push %edx > push %ebp > movl %esp,%ebp > sysenter > /* end of my_sysenter() */ > > ... > ... > ... > > call my_sysenter > > but this depends on matching stack layout with one used by vDSO. > >
I'd be rather surprised if anyone does that. It'll die with SIGILL on AMD systems. Similarly, open-coded syscall instructions in 32-bit code will die with SIGILL on Intel systems.
Gee thanks, anyone.
<with time machine>The correct way to do this ought to have been straightforward. Syscall should have stashed eip/rip in r8, r9, or r10, and sysenter shouldn't exist in long mode. All of this mess would just disappear completely.</with time machine>
--Andy
| |