lguest: optimize by coding restore_flags and irq_enable in assembler.
The downside of the last patch which made restore_flags and irq_enable
check interrupts is that they are now too big to be patched directly
into the callsites, so the C versions are always used.
But the C versions go via PV_CALLEE_SAVE_REGS_THUNK which saves all
the registers. In fact, we don't need any registers in the fast path,
so we can do better than this if we actually code them in assembler.
The results are in the noise, but since it's about the same amount of
code, it's worth applying.
1GB Guest->Host: input(suppressed),output(suppressed)
Before:
Seconds: 0:16.53
Packets: 377268,753673
Interrupts: 22461,24297
Notifications: 1(5245),21303(732370)
Net IRQs triggered: 377023(245),42578(711095)
After:
Seconds: 0:16.48
Packets: 377289,753673
Interrupts: 22281,24465
Notifications: 1(5245),21296(732377)
Net IRQs triggered: 377060(229),42564(711109)
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
diff --git a/arch/x86/lguest/boot.c b/arch/x86/lguest/boot.c
index 37b8c1d..514f4d0 100644
--- a/arch/x86/lguest/boot.c
+++ b/arch/x86/lguest/boot.c
@@ -179,7 +179,7 @@
paravirt_end_context_switch(next);
}
-/*G:033
+/*G:032
* After that diversion we return to our first native-instruction
* replacements: four functions for interrupt control.
*
@@ -199,41 +199,28 @@
{
return lguest_data.irq_enabled;
}
-PV_CALLEE_SAVE_REGS_THUNK(save_fl);
-
-/* restore_flags() just sets the flags back to the value given. */
-static void restore_fl(unsigned long flags)
-{
- lguest_data.irq_enabled = flags;
- mb();
- /* Null hcall forces interrupt delivery now, if irq_pending is
- * set to X86_EFLAGS_IF (ie. an interrupt is pending, and flags
- * enables interrupts. */
- if (flags & lguest_data.irq_pending)
- kvm_hypercall0(LHCALL_SEND_INTERRUPTS);
-}
-PV_CALLEE_SAVE_REGS_THUNK(restore_fl);
/* Interrupts go off... */
static void irq_disable(void)
{
lguest_data.irq_enabled = 0;
}
+
+/* Let's pause a moment. Remember how I said these are called so often?
+ * Jeremy Fitzhardinge optimized them so hard early in 2009 that he had to
+ * break some rules. In particular, these functions are assumed to save their
+ * own registers if they need to: normal C functions assume they can trash the
+ * eax register. To use normal C functions, we use
+ * PV_CALLEE_SAVE_REGS_THUNK(), which pushes %eax onto the stack, calls the
+ * C function, then restores it. */
+PV_CALLEE_SAVE_REGS_THUNK(save_fl);
PV_CALLEE_SAVE_REGS_THUNK(irq_disable);
-
-/* Interrupts go on... */
-static void irq_enable(void)
-{
- lguest_data.irq_enabled = X86_EFLAGS_IF;
- mb();
- /* Null hcall forces interrupt delivery now. */
- if (lguest_data.irq_pending)
- kvm_hypercall0(LHCALL_SEND_INTERRUPTS);
-
-}
-PV_CALLEE_SAVE_REGS_THUNK(irq_enable);
-
/*:*/
+
+/* These are in i386_head.S */
+extern void lg_irq_enable(void);
+extern void lg_restore_fl(unsigned long flags);
+
/*M:003 Note that we don't check for outstanding interrupts when we re-enable
* them (or when we unmask an interrupt). This seems to work for the moment,
* since interrupts are rare and we'll just get the interrupt on the next timer
@@ -1041,9 +1028,9 @@
/* interrupt-related operations */
pv_irq_ops.init_IRQ = lguest_init_IRQ;
pv_irq_ops.save_fl = PV_CALLEE_SAVE(save_fl);
- pv_irq_ops.restore_fl = PV_CALLEE_SAVE(restore_fl);
+ pv_irq_ops.restore_fl = __PV_IS_CALLEE_SAVE(lg_restore_fl);
pv_irq_ops.irq_disable = PV_CALLEE_SAVE(irq_disable);
- pv_irq_ops.irq_enable = PV_CALLEE_SAVE(irq_enable);
+ pv_irq_ops.irq_enable = __PV_IS_CALLEE_SAVE(lg_irq_enable);
pv_irq_ops.safe_halt = lguest_safe_halt;
/* init-time operations */