On Wed, 23 Oct 2019, Andy Lutomirski wrote:
> On Wed, Oct 23, 2019 at 5:31 AM Thomas Gleixner <tglx@xxxxxxxxxxxxx> wrote:
> >
> > Interrupt state tracing can be safely done in C code. The few stack
> > operations in assembly do not need to be covered.
> >
> > Remove the now pointless indirection via .Lsyscall_32_done and jump to
> > swapgs_restore_regs_and_return_to_usermode directly.
>
> This doesn't look right.
>
> > #define SYSCALL_EXIT_WORK_FLAGS \
> > @@ -279,6 +282,9 @@ static void syscall_slow_exit_work(struc
> > {
> > struct thread_info *ti;
> >
> > + /* User to kernel transition disabled interrupts. */
> > + trace_hardirqs_off();
> > +
>
> So you just traced IRQs off, but...
>
> > enter_from_user_mode();
> > local_irq_enable();
>
> Now they're on and traced on again?
Yes, because that's what actually happens.
usermode
syscall <- Disables interrupts, but tracing thinks they are on
entry_SYSCALL_64
....
call do_syscall_64
trace_hardirqs_off() <- So before calling anything else, we have to tell
the tracer that interrupts are on, which we did
so far in the ASM code between entry_SYSCALL_64
and 'call do_syscall_64'. I'm merily lifting this
to C-code.
enter_from_user_mode()
local_irq_enable()
> I also don't see how your patch handles the fastpath case.
Hmm?
All syscalls return through:
syscall_return_slowpath()
local_irq_disable()
prepare_exit_to_usermode()
user_enter_irqoff()
mds_user_clear_cpu_buffers()
trace_hardirqs_on()
What am I missing?
> How about the attached patch instead?
^^^^^^ Groan.
>
> user_enter_irqoff();
>
> + /*
> + * The actual return to usermode will almost certainly turn IRQs on.
> + * Trace it here to simplify the asm code.
Why would we return to user from a syscall or interrupt with interrupts
traced as disabled? Also the existing ASM is inconsistent vs. that:
ENTRY(entry_SYSENTER_32) TRACE_IRQS_ON
ENTRY(entry_INT80_32) TRACE_IRQS_IRET
ENTRY(entry_SYSCALL_64) TRACE_IRQS_IRET
ENTRY(ret_from_fork) TRACE_IRQS_ON
GLOBAL(retint_user) TRACE_IRQS_IRETQ
ENTRY(entry_SYSCALL_compat) TRACE_IRQS_ON
ENTRY(entry_INT80_compat) TRACE_IRQS_ON
> + */
> + if (likely(regs->flags & X86_EFLAGS_IF))
> + trace_hardirqs_on();
My variant does this unconditionally and after
mds_user_clear_cpu_buffers().
> mds_user_clear_cpu_buffers();
> }
And your ASM changes keep still all the TRACE_IRQS_OFF invocations in the
various syscall entry pathes, which is what I removed and put as the first
thing into the C functions.
Confused.
Thanks,
tglx
