s390/smp: Switch pcpu_devices to percpu

In preparation of moving the CIF flags from lowcore to pcpu_devices,
convert the pcpu_devices array to use the percpu infrastructure.
This is required because using the pcpu_devices array as it is would
introduce a performance penalty due to the fact that CPU flags for
multiple CPUs would end up in the same cacheline.

Note that a pointer to the pcpu struct of the IPL CPU is still required.
This is because a restart interrupt can be triggered on an offline CPU.
s390 stores the percpu offset in lowcore, but offline CPUs have no
lowcore area allocated. So percpu data cannot be used from an offline
CPU and we need to get the pcpu pointer for the IPL cpu from somewhere
else.
Reviewed-by: Heiko Carstens <hca@linux.ibm.com>
Signed-off-by: Sven Schnelle <svens@linux.ibm.com>
Signed-off-by: Vasily Gorbik <gor@linux.ibm.com>

arch/s390/kernel/smp.c

@@ -83,7 +83,14 @@ struct pcpu {
 };
 
 static u8 boot_core_type;
-static struct pcpu pcpu_devices[NR_CPUS];
+static DEFINE_PER_CPU(struct pcpu, pcpu_devices);
+/*
+ * Pointer to the pcpu area of the boot CPU. This is required when a restart
+ * interrupt is triggered on an offline CPU. For that case accessing percpu
+ * data with the common primitives does not work, since the percpu offset is
+ * stored in a non existent lowcore.
+ */
+static struct pcpu *ipl_pcpu;
 
 unsigned int smp_cpu_mt_shift;
 EXPORT_SYMBOL(smp_cpu_mt_shift);
@@ -174,8 +181,8 @@ static struct pcpu *pcpu_find_address(const struct cpumask *mask, u16 address)
 	int cpu;
 
 	for_each_cpu(cpu, mask)
-		if (pcpu_devices[cpu].address == address)
-			return pcpu_devices + cpu;
+		if (per_cpu(pcpu_devices, cpu).address == address)
+			return &per_cpu(pcpu_devices, cpu);
 	return NULL;
 }
 
@@ -230,13 +237,11 @@ static int pcpu_alloc_lowcore(struct pcpu *pcpu, int cpu)
 	return -ENOMEM;
 }
 
-static void pcpu_free_lowcore(struct pcpu *pcpu)
+static void pcpu_free_lowcore(struct pcpu *pcpu, int cpu)
 {
 	unsigned long async_stack, nodat_stack, mcck_stack;
 	struct lowcore *lc;
-	int cpu;
 
-	cpu = pcpu - pcpu_devices;
 	lc = lowcore_ptr[cpu];
 	nodat_stack = lc->nodat_stack - STACK_INIT_OFFSET;
 	async_stack = lc->async_stack - STACK_INIT_OFFSET;
@@ -277,12 +282,10 @@ static void pcpu_prepare_secondary(struct pcpu *pcpu, int cpu)
 	arch_spin_lock_setup(cpu);
 }
 
-static void pcpu_attach_task(struct pcpu *pcpu, struct task_struct *tsk)
+static void pcpu_attach_task(int cpu, struct task_struct *tsk)
 {
 	struct lowcore *lc;
-	int cpu;
 
-	cpu = pcpu - pcpu_devices;
 	lc = lowcore_ptr[cpu];
 	lc->kernel_stack = (unsigned long)task_stack_page(tsk) + STACK_INIT_OFFSET;
 	lc->current_task = (unsigned long)tsk;
@@ -296,18 +299,16 @@ static void pcpu_attach_task(struct pcpu *pcpu, struct task_struct *tsk)
 	lc->steal_timer = 0;
 }
 
-static void pcpu_start_fn(struct pcpu *pcpu, void (*func)(void *), void *data)
+static void pcpu_start_fn(int cpu, void (*func)(void *), void *data)
 {
 	struct lowcore *lc;
-	int cpu;
 
-	cpu = pcpu - pcpu_devices;
 	lc = lowcore_ptr[cpu];
 	lc->restart_stack = lc->kernel_stack;
 	lc->restart_fn = (unsigned long) func;
 	lc->restart_data = (unsigned long) data;
 	lc->restart_source = -1U;
-	pcpu_sigp_retry(pcpu, SIGP_RESTART, 0);
+	pcpu_sigp_retry(per_cpu_ptr(&pcpu_devices, cpu), SIGP_RESTART, 0);
 }
 
 typedef void (pcpu_delegate_fn)(void *);
@@ -320,14 +321,14 @@ static void __pcpu_delegate(pcpu_delegate_fn *func, void *data)
 	func(data);	/* should not return */
 }
 
-static void pcpu_delegate(struct pcpu *pcpu,
+static void pcpu_delegate(struct pcpu *pcpu, int cpu,
 			  pcpu_delegate_fn *func,
 			  void *data, unsigned long stack)
 {
 	struct lowcore *lc, *abs_lc;
 	unsigned int source_cpu;
 
-	lc = lowcore_ptr[pcpu - pcpu_devices];
+	lc = lowcore_ptr[cpu];
 	source_cpu = stap();
 
 	if (pcpu->address == source_cpu) {
@@ -377,7 +378,7 @@ static int pcpu_set_smt(unsigned int mtid)
 		smp_cpu_mt_shift = 0;
 		while (smp_cpu_mtid >= (1U << smp_cpu_mt_shift))
 			smp_cpu_mt_shift++;
-		pcpu_devices[0].address = stap();
+		per_cpu(pcpu_devices, 0).address = stap();
 	}
 	return cc;
 }
@@ -389,11 +390,10 @@ void smp_call_ipl_cpu(void (*func)(void *), void *data)
 {
 	struct lowcore *lc = lowcore_ptr[0];
 
-	if (pcpu_devices[0].address == stap())
+	if (ipl_pcpu->address == stap())
 		lc = get_lowcore();
 
-	pcpu_delegate(&pcpu_devices[0], func, data,
-		      lc->nodat_stack);
+	pcpu_delegate(ipl_pcpu, 0, func, data, lc->nodat_stack);
 }
 
 int smp_find_processor_id(u16 address)
@@ -401,21 +401,21 @@ int smp_find_processor_id(u16 address)
 	int cpu;
 
 	for_each_present_cpu(cpu)
-		if (pcpu_devices[cpu].address == address)
+		if (per_cpu(pcpu_devices, cpu).address == address)
 			return cpu;
 	return -1;
 }
 
 void schedule_mcck_handler(void)
 {
-	pcpu_ec_call(pcpu_devices + smp_processor_id(), ec_mcck_pending);
+	pcpu_ec_call(this_cpu_ptr(&pcpu_devices), ec_mcck_pending);
 }
 
 bool notrace arch_vcpu_is_preempted(int cpu)
 {
 	if (test_cpu_flag_of(CIF_ENABLED_WAIT, cpu))
 		return false;
-	if (pcpu_running(pcpu_devices + cpu))
+	if (pcpu_running(per_cpu_ptr(&pcpu_devices, cpu)))
 		return false;
 	return true;
 }
@@ -427,7 +427,7 @@ void notrace smp_yield_cpu(int cpu)
 		return;
 	diag_stat_inc_norecursion(DIAG_STAT_X09C);
 	asm volatile("diag %0,0,0x9c"
-		     : : "d" (pcpu_devices[cpu].address));
+		     : : "d" (per_cpu(pcpu_devices, cpu).address));
 }
 EXPORT_SYMBOL_GPL(smp_yield_cpu);
 
@@ -448,7 +448,7 @@ void notrace smp_emergency_stop(void)
 
 	end = get_tod_clock() + (1000000UL << 12);
 	for_each_cpu(cpu, &cpumask) {
-		struct pcpu *pcpu = pcpu_devices + cpu;
+		struct pcpu *pcpu = per_cpu_ptr(&pcpu_devices, cpu);
 		set_bit(ec_stop_cpu, &pcpu->ec_mask);
 		while (__pcpu_sigp(pcpu->address, SIGP_EMERGENCY_SIGNAL,
 				   0, NULL) == SIGP_CC_BUSY &&
@@ -457,7 +457,7 @@ void notrace smp_emergency_stop(void)
 	}
 	while (get_tod_clock() < end) {
 		for_each_cpu(cpu, &cpumask)
-			if (pcpu_stopped(pcpu_devices + cpu))
+			if (pcpu_stopped(per_cpu_ptr(&pcpu_devices, cpu)))
 				cpumask_clear_cpu(cpu, &cpumask);
 		if (cpumask_empty(&cpumask))
 			break;
@@ -472,6 +472,7 @@ NOKPROBE_SYMBOL(smp_emergency_stop);
  */
 void smp_send_stop(void)
 {
+	struct pcpu *pcpu;
 	int cpu;
 
 	/* Disable all interrupts/machine checks */
@@ -487,8 +488,9 @@ void smp_send_stop(void)
 	for_each_online_cpu(cpu) {
 		if (cpu == smp_processor_id())
 			continue;
-		pcpu_sigp_retry(pcpu_devices + cpu, SIGP_STOP, 0);
-		while (!pcpu_stopped(pcpu_devices + cpu))
+		pcpu = per_cpu_ptr(&pcpu_devices, cpu);
+		pcpu_sigp_retry(pcpu, SIGP_STOP, 0);
+		while (!pcpu_stopped(pcpu))
 			cpu_relax();
 	}
 }
@@ -502,7 +504,7 @@ static void smp_handle_ext_call(void)
 	unsigned long bits;
 
 	/* handle bit signal external calls */
-	bits = xchg(&pcpu_devices[smp_processor_id()].ec_mask, 0);
+	bits = this_cpu_xchg(pcpu_devices.ec_mask, 0);
 	if (test_bit(ec_stop_cpu, &bits))
 		smp_stop_cpu();
 	if (test_bit(ec_schedule, &bits))
@@ -527,12 +529,12 @@ void arch_send_call_function_ipi_mask(const struct cpumask *mask)
 	int cpu;
 
 	for_each_cpu(cpu, mask)
-		pcpu_ec_call(pcpu_devices + cpu, ec_call_function_single);
+		pcpu_ec_call(per_cpu_ptr(&pcpu_devices, cpu), ec_call_function_single);
 }
 
 void arch_send_call_function_single_ipi(int cpu)
 {
-	pcpu_ec_call(pcpu_devices + cpu, ec_call_function_single);
+	pcpu_ec_call(per_cpu_ptr(&pcpu_devices, cpu), ec_call_function_single);
 }
 
 /*
@@ -542,13 +544,13 @@ void arch_send_call_function_single_ipi(int cpu)
  */
 void arch_smp_send_reschedule(int cpu)
 {
-	pcpu_ec_call(pcpu_devices + cpu, ec_schedule);
+	pcpu_ec_call(per_cpu_ptr(&pcpu_devices, cpu), ec_schedule);
 }
 
 #ifdef CONFIG_IRQ_WORK
 void arch_irq_work_raise(void)
 {
-	pcpu_ec_call(pcpu_devices + smp_processor_id(), ec_irq_work);
+	pcpu_ec_call(this_cpu_ptr(&pcpu_devices), ec_irq_work);
 }
 #endif
 
@@ -560,7 +562,7 @@ int smp_store_status(int cpu)
 	struct pcpu *pcpu;
 	unsigned long pa;
 
-	pcpu = pcpu_devices + cpu;
+	pcpu = per_cpu_ptr(&pcpu_devices, cpu);
 	lc = lowcore_ptr[cpu];
 	pa = __pa(&lc->floating_pt_save_area);
 	if (__pcpu_sigp_relax(pcpu->address, SIGP_STORE_STATUS_AT_ADDRESS,
@@ -668,17 +670,17 @@ void __init smp_save_dump_secondary_cpus(void)
 
 void smp_cpu_set_polarization(int cpu, int val)
 {
-	pcpu_devices[cpu].polarization = val;
+	per_cpu(pcpu_devices, cpu).polarization = val;
 }
 
 int smp_cpu_get_polarization(int cpu)
 {
-	return pcpu_devices[cpu].polarization;
+	return per_cpu(pcpu_devices, cpu).polarization;
 }
 
 int smp_cpu_get_cpu_address(int cpu)
 {
-	return pcpu_devices[cpu].address;
+	return per_cpu(pcpu_devices, cpu).address;
 }
 
 static void __ref smp_get_core_info(struct sclp_core_info *info, int early)
@@ -717,7 +719,7 @@ static int smp_add_core(struct sclp_core_entry *core, cpumask_t *avail,
 	for (i = 0; (i <= smp_cpu_mtid) && (cpu < nr_cpu_ids); i++) {
 		if (pcpu_find_address(cpu_present_mask, address + i))
 			continue;
-		pcpu = pcpu_devices + cpu;
+		pcpu = per_cpu_ptr(&pcpu_devices, cpu);
 		pcpu->address = address + i;
 		if (configured)
 			pcpu->state = CPU_STATE_CONFIGURED;
@@ -752,7 +754,7 @@ static int __smp_rescan_cpus(struct sclp_core_info *info, bool early)
 	 * that all SMT threads get subsequent logical CPU numbers.
 	 */
 	if (early) {
-		core_id = pcpu_devices[0].address >> smp_cpu_mt_shift;
+		core_id = per_cpu(pcpu_devices, 0).address >> smp_cpu_mt_shift;
 		for (i = 0; i < info->configured; i++) {
 			core = &info->core[i];
 			if (core->core_id == core_id) {
@@ -852,7 +854,7 @@ static void smp_start_secondary(void *cpuvoid)
 /* Upping and downing of CPUs */
 int __cpu_up(unsigned int cpu, struct task_struct *tidle)
 {
-	struct pcpu *pcpu = pcpu_devices + cpu;
+	struct pcpu *pcpu = per_cpu_ptr(&pcpu_devices, cpu);
 	int rc;
 
 	if (pcpu->state != CPU_STATE_CONFIGURED)
@@ -870,8 +872,8 @@ int __cpu_up(unsigned int cpu, struct task_struct *tidle)
 	 */
 	system_ctlreg_lock();
 	pcpu_prepare_secondary(pcpu, cpu);
-	pcpu_attach_task(pcpu, tidle);
-	pcpu_start_fn(pcpu, smp_start_secondary, NULL);
+	pcpu_attach_task(cpu, tidle);
+	pcpu_start_fn(cpu, smp_start_secondary, NULL);
 	/* Wait until cpu puts itself in the online & active maps */
 	while (!cpu_online(cpu))
 		cpu_relax();
@@ -916,10 +918,10 @@ void __cpu_die(unsigned int cpu)
 	struct pcpu *pcpu;
 
 	/* Wait until target cpu is down */
-	pcpu = pcpu_devices + cpu;
+	pcpu = per_cpu_ptr(&pcpu_devices, cpu);
 	while (!pcpu_stopped(pcpu))
 		cpu_relax();
-	pcpu_free_lowcore(pcpu);
+	pcpu_free_lowcore(pcpu, cpu);
 	cpumask_clear_cpu(cpu, mm_cpumask(&init_mm));
 	cpumask_clear_cpu(cpu, &init_mm.context.cpu_attach_mask);
 }
@@ -927,7 +929,7 @@ void __cpu_die(unsigned int cpu)
 void __noreturn cpu_die(void)
 {
 	idle_task_exit();
-	pcpu_sigp_retry(pcpu_devices + smp_processor_id(), SIGP_STOP, 0);
+	pcpu_sigp_retry(this_cpu_ptr(&pcpu_devices), SIGP_STOP, 0);
 	for (;;) ;
 }
 
@@ -957,10 +959,9 @@ void __init smp_prepare_cpus(unsigned int max_cpus)
 
 void __init smp_prepare_boot_cpu(void)
 {
-	struct pcpu *pcpu = pcpu_devices;
-
 	WARN_ON(!cpu_present(0) || !cpu_online(0));
-	pcpu->state = CPU_STATE_CONFIGURED;
+	ipl_pcpu = per_cpu_ptr(&pcpu_devices, 0);
+	ipl_pcpu->state = CPU_STATE_CONFIGURED;
 	get_lowcore()->percpu_offset = __per_cpu_offset[0];
 	smp_cpu_set_polarization(0, POLARIZATION_UNKNOWN);
 }
@@ -969,8 +970,8 @@ void __init smp_setup_processor_id(void)
 {
 	struct lowcore *lc = get_lowcore();
 
-	pcpu_devices[0].address = stap();
 	lc->cpu_nr = 0;
+	per_cpu(pcpu_devices, 0).address = stap();
 	lc->spinlock_lockval = arch_spin_lockval(0);
 	lc->spinlock_index = 0;
 }
@@ -992,7 +993,7 @@ static ssize_t cpu_configure_show(struct device *dev,
 	ssize_t count;
 
 	mutex_lock(&smp_cpu_state_mutex);
-	count = sprintf(buf, "%d\n", pcpu_devices[dev->id].state);
+	count = sprintf(buf, "%d\n", per_cpu(pcpu_devices, dev->id).state);
 	mutex_unlock(&smp_cpu_state_mutex);
 	return count;
 }
@@ -1018,7 +1019,7 @@ static ssize_t cpu_configure_store(struct device *dev,
 	for (i = 0; i <= smp_cpu_mtid; i++)
 		if (cpu_online(cpu + i))
 			goto out;
-	pcpu = pcpu_devices + cpu;
+	pcpu = per_cpu_ptr(&pcpu_devices, cpu);
 	rc = 0;
 	switch (val) {
 	case 0:
@@ -1030,7 +1031,7 @@ static ssize_t cpu_configure_store(struct device *dev,
 		for (i = 0; i <= smp_cpu_mtid; i++) {
 			if (cpu + i >= nr_cpu_ids || !cpu_present(cpu + i))
 				continue;
-			pcpu[i].state = CPU_STATE_STANDBY;
+			per_cpu(pcpu_devices, cpu + i).state = CPU_STATE_STANDBY;
 			smp_cpu_set_polarization(cpu + i,
 						 POLARIZATION_UNKNOWN);
 		}
@@ -1045,7 +1046,7 @@ static ssize_t cpu_configure_store(struct device *dev,
 		for (i = 0; i <= smp_cpu_mtid; i++) {
 			if (cpu + i >= nr_cpu_ids || !cpu_present(cpu + i))
 				continue;
-			pcpu[i].state = CPU_STATE_CONFIGURED;
+			per_cpu(pcpu_devices, cpu + i).state = CPU_STATE_CONFIGURED;
 			smp_cpu_set_polarization(cpu + i,
 						 POLARIZATION_UNKNOWN);
 		}
@@ -1064,7 +1065,7 @@ static DEVICE_ATTR(configure, 0644, cpu_configure_show, cpu_configure_store);
 static ssize_t show_cpu_address(struct device *dev,
 				struct device_attribute *attr, char *buf)
 {
-	return sprintf(buf, "%d\n", pcpu_devices[dev->id].address);
+	return sprintf(buf, "%d\n", per_cpu(pcpu_devices, dev->id).address);
 }
 static DEVICE_ATTR(address, 0444, show_cpu_address, NULL);
 
@@ -1090,14 +1091,14 @@ static struct attribute_group cpu_online_attr_group = {
 
 static int smp_cpu_online(unsigned int cpu)
 {
-	struct cpu *c = &per_cpu(cpu_devices, cpu);
+	struct cpu *c = per_cpu_ptr(&cpu_devices, cpu);
 
 	return sysfs_create_group(&c->dev.kobj, &cpu_online_attr_group);
 }
 
 static int smp_cpu_pre_down(unsigned int cpu)
 {
-	struct cpu *c = &per_cpu(cpu_devices, cpu);
+	struct cpu *c = per_cpu_ptr(&cpu_devices, cpu);
 
 	sysfs_remove_group(&c->dev.kobj, &cpu_online_attr_group);
 	return 0;
@@ -1110,7 +1111,7 @@ bool arch_cpu_is_hotpluggable(int cpu)
 
 int arch_register_cpu(int cpu)
 {
-	struct cpu *c = &per_cpu(cpu_devices, cpu);
+	struct cpu *c = per_cpu_ptr(&cpu_devices, cpu);
 	int rc;
 
 	c->hotpluggable = arch_cpu_is_hotpluggable(cpu);