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.
*
{
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;
}
-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);
+/* 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);
+/*:*/
-}
-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
/* 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 */
LGUEST_PATCH(cli, movl $0, lguest_data+LGUEST_DATA_irq_enabled)
LGUEST_PATCH(pushf, movl lguest_data+LGUEST_DATA_irq_enabled, %eax)
-/*:*/
+
+/*G:033 But using those wrappers is inefficient (we'll see why that doesn't
+ * matter for save_fl and irq_disable later). If we write our routines
+ * carefully in assembler, we can avoid clobbering any registers and avoid
+ * jumping through the wrapper functions.
+ *
+ * I skipped over our first piece of assembler, but this one is worth studying
+ * in a bit more detail so I'll describe in easy stages. First, the routine
+ * to enable interrupts: */
+ENTRY(lg_irq_enable)
+ /* The reverse of irq_disable, this sets lguest_data.irq_enabled to
+ * X86_EFLAGS_IF (ie. "Interrupts enabled"). */
+ movl $X86_EFLAGS_IF, lguest_data+LGUEST_DATA_irq_enabled
+ /* But now we need to check if the Host wants to know: there might have
+ * been interrupts waiting to be delivered, in which case it will have
+ * set lguest_data.irq_pending to X86_EFLAGS_IF. If it's not zero, we
+ * jump to send_interrupts, otherwise we're done. */
+ testl $0, lguest_data+LGUEST_DATA_irq_pending
+ jnz send_interrupts
+ /* One cool thing about x86 is that you can do many things without using
+ * a register. In this case, the normal path hasn't needed to save or
+ * restore any registers at all! */
+ ret
+send_interrupts:
+ /* OK, now we need a register: eax is used for the hypercall number,
+ * which is LHCALL_SEND_INTERRUPTS.
+ *
+ * We used not to bother with this pending detection at all, which was
+ * much simpler. Sooner or later the Host would realize it had to
+ * send us an interrupt. But that turns out to make performance 7
+ * times worse on a simple tcp benchmark. So now we do this the hard
+ * way. */
+ pushl %eax
+ movl $LHCALL_SEND_INTERRUPTS, %eax
+ /* This is a vmcall instruction (same thing that KVM uses). Older
+ * assembler versions might not know the "vmcall" instruction, so we
+ * create one manually here. */
+ .byte 0x0f,0x01,0xc1 /* KVM_HYPERCALL */
+ popl %eax
+ ret
+
+/* Finally, the "popf" or "restore flags" routine. The %eax register holds the
+ * flags (in practice, either X86_EFLAGS_IF or 0): if it's X86_EFLAGS_IF we're
+ * enabling interrupts again, if it's 0 we're leaving them off. */
+ENTRY(lg_restore_fl)
+ /* This is just "lguest_data.irq_enabled = flags;" */
+ movl %eax, lguest_data+LGUEST_DATA_irq_enabled
+ /* Now, if the %eax value has enabled interrupts and
+ * lguest_data.irq_pending is set, we want to tell the Host so it can
+ * deliver any outstanding interrupts. Fortunately, both values will
+ * be X86_EFLAGS_IF (ie. 512) in that case, and the "testl"
+ * instruction will AND them together for us. If both are set, we
+ * jump to send_interrupts. */
+ testl lguest_data+LGUEST_DATA_irq_pending, %eax
+ jnz send_interrupts
+ /* Again, the normal path has used no extra registers. Clever, huh? */
+ ret
/* These demark the EIP range where host should never deliver interrupts. */
.global lguest_noirq_start
projects / users / jedix / linux-maple.git / commitdiff