On Thu, Feb 25, 2016 at 11:49 AM, Brian Gerst <brge...@gmail.com> wrote: > On Thu, Feb 25, 2016 at 2:39 PM, Andy Lutomirski <l...@amacapital.net> wrote: >> On Thu, Feb 25, 2016 at 11:31 AM, Brian Gerst <brge...@gmail.com> wrote: >>> On Thu, Feb 25, 2016 at 1:30 PM, Linus Torvalds >>> <torva...@linux-foundation.org> wrote: >>>> On Thu, Feb 25, 2016 at 10:20 AM, Andy Lutomirski <l...@amacapital.net> >>>> wrote: >>>>> >>>>> Ideally we'd fix this up and restore flags on sysexit. At least >>>>> failing to restore arithmetic flags isn't an info leak because the >>>>> exit code clobbers them with entirely predictable data. I doubt >>>>> anyone cares all that much if we clobber AC. >>>> >>>> As long as the "clobber AC" is purely about clearing it, it's probably >>>> fine. >>>> >>>> Although there may be programs that set AC in order to actually get >>>> notified about alignment issues (perhaps for portability reasons, >>>> perhaps for small performance reasons). Clearing it will make those >>>> programs still work, but they lose the checking. >>>> >>>>> I wrote a test for NT and the test fails for a different reason: our >>>>> TF handling appears broken as well. (Our sysenter TF handling is >>>>> *crap*, but it seems to work on 64-bit kernels at least.) >>>> >>>> TF should be entirely immaterial for system calls. Why would we care? >>>> We need it for correct handling of real traps, but not for the system >>>> call case afaik. Returning with TF clear is the right thing, since >>>> we're not returning *to* the system call instruction, but the >>>> instruction after. >>>> >>>>> My personal preference would be to add the missing popf. >>>> >>>> I don't mind adding the popf, but it won't help for iopl. Only iret >>>> restores iopl, if I recall correctly (but maybe I don't, and I'm too >>>> lazy to take the 30 seconds to look it up). >>>> >>>> Linus >>> >>> According to the SDM, popf will change IOPL only at CPL0, which is why >>> Xen (which runs at CPL1 on 32-bit) has a paravirt hook for it. >> >> But maybe we can ditch that paravirt hook and just modify regs->flags >> in sys_iopl. Xen never uses sysexit at all: >> >> /* XEN PV guests always use IRET path */ >> ALTERNATIVE "testl %eax, %eax; jz .Lsyscall_32_done", \ >> "jmp .Lsyscall_32_done", X86_FEATURE_XENPV >> >> and if we add the missing popf, we should be good to go. > > IRET won't change IOPL either at CPL != 0, so Xen still needs that hook.
But xen_iret isn't the same thing as IRET at all. We don't use a real IRET instruction to switch between kernel and user mode on Xen PV. --Andy > > -- > Brian Gerst -- Andy Lutomirski AMA Capital Management, LLC
