powerpc: Make sure IPI handlers see data written by IPI senders

We have been observing hangs, both of KVM guest vcpu tasks and more
generally, where a process that is woken doesn't properly wake up and
continue to run, but instead sticks in TASK_WAKING state. This
happens because the update of rq->wake_list in ttwu_queue_remote()
is not ordered with the update of ipi_message in
smp_muxed_ipi_message_pass(), and the reading of rq->wake_list in
scheduler_ipi() is not ordered with the reading of ipi_message in
smp_ipi_demux(). Thus it is possible for the IPI receiver not to see
the updated rq->wake_list and therefore conclude that there is nothing
for it to do.

In order to make sure that anything done before smp_send_reschedule()
is ordered before anything done in the resulting call to scheduler_ipi(),
this adds barriers in smp_muxed_message_pass() and smp_ipi_demux().
The barrier in smp_muxed_message_pass() is a full barrier to ensure that
there is a full ordering between the smp_send_reschedule() caller and
scheduler_ipi(). In smp_ipi_demux(), we use xchg() rather than
xchg_local() because xchg() includes release and acquire barriers.
Using xchg() rather than xchg_local() makes sense given that
ipi_message is not just accessed locally.

This moves the barrier between setting the message and calling the
cause_ipi() function into the individual cause_ipi implementations.
Most of them -- those that used outb, out_8 or similar -- already had
a full barrier because out_8 etc. include a sync before the MMIO
store. This adds an explicit barrier in the two remaining cases.

These changes made no measurable difference to the speed of IPIs as
measured using a simple ping-pong latency test across two CPUs on
different cores of a POWER7 machine.

The analysis of the reason why processes were not waking up properly
is due to Milton Miller.

Cc: [e-mail] # v3.0+
Reported-by: Milton Miller <[e-mail]>
Signed-off-by: Paul Mackerras <[e-mail]>
Signed-off-by: Benjamin Herrenschmidt <[e-mail]>

arch/powerpc/kernel/dbell.c

···
 
 void doorbell_cause_ipi(int cpu, unsigned long data)
 {
+	/* Order previous accesses vs. msgsnd, which is treated as a store */
+	mb();
 	ppc_msgsnd(PPC_DBELL, 0, data);
 }
 

arch/powerpc/kernel/smp.c

···
 	struct cpu_messages *info = &per_cpu(ipi_message, cpu);
 	char *message = (char *)&info->messages;
 
+	/*
+	 * Order previous accesses before accesses in the IPI handler.
+	 */
+	smp_mb();
 	message[msg] = 1;
-	mb();
+	/*
+	 * cause_ipi functions are required to include a full barrier
+	 * before doing whatever causes the IPI.
+	 */
 	smp_ops->cause_ipi(cpu, info->data);
 }
 
···
 	mb();	/* order any irq clear */
 
 	do {
-		all = xchg_local(&info->messages, 0);
+		all = xchg(&info->messages, 0);
 
 #ifdef __BIG_ENDIAN
 		if (all & (1 << (24 - 8 * PPC_MSG_CALL_FUNCTION)))

arch/powerpc/sysdev/xics/icp-hv.c

···
 static inline void icp_hv_set_qirr(int n_cpu , u8 value)
 {
 	int hw_cpu = get_hard_smp_processor_id(n_cpu);
-	long rc = plpar_hcall_norets(H_IPI, hw_cpu, value);
+	long rc;
+
+	/* Make sure all previous accesses are ordered before IPI sending */
+	mb();
+	rc = plpar_hcall_norets(H_IPI, hw_cpu, value);
 	if (rc != H_SUCCESS) {
 		pr_err("%s: bad return code qirr cpu=%d hw_cpu=%d mfrr=0x%x "
 			"returned %ld\n", __func__, n_cpu, hw_cpu, value, rc);