/arch/ppc64/kernel/pSeries_smp.c

  1/*
  2 * SMP support for pSeries machines.
  3 *
  4 * Dave Engebretsen, Peter Bergner, and
  5 * Mike Corrigan {engebret|bergner|mikec}@us.ibm.com
  6 *
  7 * Plus various changes from other IBM teams...
  8 *
  9 *      This program is free software; you can redistribute it and/or
 10 *      modify it under the terms of the GNU General Public License
 11 *      as published by the Free Software Foundation; either version
 12 *      2 of the License, or (at your option) any later version.
 13 */
 14
 15#undef DEBUG
 16
 17#include <linux/config.h>
 18#include <linux/kernel.h>
 19#include <linux/module.h>
 20#include <linux/sched.h>
 21#include <linux/smp.h>
 22#include <linux/interrupt.h>
 23#include <linux/delay.h>
 24#include <linux/init.h>
 25#include <linux/spinlock.h>
 26#include <linux/cache.h>
 27#include <linux/err.h>
 28#include <linux/sysdev.h>
 29#include <linux/cpu.h>
 30
 31#include <asm/ptrace.h>
 32#include <asm/atomic.h>
 33#include <asm/irq.h>
 34#include <asm/page.h>
 35#include <asm/pgtable.h>
 36#include <asm/io.h>
 37#include <asm/prom.h>
 38#include <asm/smp.h>
 39#include <asm/paca.h>
 40#include <asm/time.h>
 41#include <asm/machdep.h>
 42#include <asm/xics.h>
 43#include <asm/cputable.h>
 44#include <asm/system.h>
 45#include <asm/rtas.h>
 46#include <asm/plpar_wrappers.h>
 47#include <asm/pSeries_reconfig.h>
 48
 49#include "mpic.h"
 50
 51#ifdef DEBUG
 52#define DBG(fmt...) udbg_printf(fmt)
 53#else
 54#define DBG(fmt...)
 55#endif
 56
 57/*
 58 * The primary thread of each non-boot processor is recorded here before
 59 * smp init.
 60 */
 61static cpumask_t of_spin_map;
 62
 63extern void pSeries_secondary_smp_init(unsigned long);
 64
 65#ifdef CONFIG_HOTPLUG_CPU
 66
 67/* Get state of physical CPU.
 68 * Return codes:
 69 *	0	- The processor is in the RTAS stopped state
 70 *	1	- stop-self is in progress
 71 *	2	- The processor is not in the RTAS stopped state
 72 *	-1	- Hardware Error
 73 *	-2	- Hardware Busy, Try again later.
 74 */
 75static int query_cpu_stopped(unsigned int pcpu)
 76{
 77	int cpu_status;
 78	int status, qcss_tok;
 79
 80	qcss_tok = rtas_token("query-cpu-stopped-state");
 81	if (qcss_tok == RTAS_UNKNOWN_SERVICE)
 82		return -1;
 83	status = rtas_call(qcss_tok, 1, 2, &cpu_status, pcpu);
 84	if (status != 0) {
 85		printk(KERN_ERR
 86		       "RTAS query-cpu-stopped-state failed: %i\n", status);
 87		return status;
 88	}
 89
 90	return cpu_status;
 91}
 92
 93int pSeries_cpu_disable(void)
 94{
 95	systemcfg->processorCount--;
 96
 97	/*fix boot_cpuid here*/
 98	if (smp_processor_id() == boot_cpuid)
 99		boot_cpuid = any_online_cpu(cpu_online_map);
100
101	/* FIXME: abstract this to not be platform specific later on */
102	xics_migrate_irqs_away();
103	return 0;
104}
105
106void pSeries_cpu_die(unsigned int cpu)
107{
108	int tries;
109	int cpu_status;
110	unsigned int pcpu = get_hard_smp_processor_id(cpu);
111
112	for (tries = 0; tries < 25; tries++) {
113		cpu_status = query_cpu_stopped(pcpu);
114		if (cpu_status == 0 || cpu_status == -1)
115			break;
116		msleep(200);
117	}
118	if (cpu_status != 0) {
119		printk("Querying DEAD? cpu %i (%i) shows %i\n",
120		       cpu, pcpu, cpu_status);
121	}
122
123	/* Isolation and deallocation are definatly done by
124	 * drslot_chrp_cpu.  If they were not they would be
125	 * done here.  Change isolate state to Isolate and
126	 * change allocation-state to Unusable.
127	 */
128	paca[cpu].cpu_start = 0;
129}
130
131/*
132 * Update cpu_present_map and paca(s) for a new cpu node.  The wrinkle
133 * here is that a cpu device node may represent up to two logical cpus
134 * in the SMT case.  We must honor the assumption in other code that
135 * the logical ids for sibling SMT threads x and y are adjacent, such
136 * that x^1 == y and y^1 == x.
137 */
138static int pSeries_add_processor(struct device_node *np)
139{
140	unsigned int cpu;
141	cpumask_t candidate_map, tmp = CPU_MASK_NONE;
142	int err = -ENOSPC, len, nthreads, i;
143	u32 *intserv;
144
145	intserv = (u32 *)get_property(np, "ibm,ppc-interrupt-server#s", &len);
146	if (!intserv)
147		return 0;
148
149	nthreads = len / sizeof(u32);
150	for (i = 0; i < nthreads; i++)
151		cpu_set(i, tmp);
152
153	lock_cpu_hotplug();
154
155	BUG_ON(!cpus_subset(cpu_present_map, cpu_possible_map));
156
157	/* Get a bitmap of unoccupied slots. */
158	cpus_xor(candidate_map, cpu_possible_map, cpu_present_map);
159	if (cpus_empty(candidate_map)) {
160		/* If we get here, it most likely means that NR_CPUS is
161		 * less than the partition's max processors setting.
162		 */
163		printk(KERN_ERR "Cannot add cpu %s; this system configuration"
164		       " supports %d logical cpus.\n", np->full_name,
165		       cpus_weight(cpu_possible_map));
166		goto out_unlock;
167	}
168
169	while (!cpus_empty(tmp))
170		if (cpus_subset(tmp, candidate_map))
171			/* Found a range where we can insert the new cpu(s) */
172			break;
173		else
174			cpus_shift_left(tmp, tmp, nthreads);
175
176	if (cpus_empty(tmp)) {
177		printk(KERN_ERR "Unable to find space in cpu_present_map for"
178		       " processor %s with %d thread(s)\n", np->name,
179		       nthreads);
180		goto out_unlock;
181	}
182
183	for_each_cpu_mask(cpu, tmp) {
184		BUG_ON(cpu_isset(cpu, cpu_present_map));
185		cpu_set(cpu, cpu_present_map);
186		set_hard_smp_processor_id(cpu, *intserv++);
187	}
188	err = 0;
189out_unlock:
190	unlock_cpu_hotplug();
191	return err;
192}
193
194/*
195 * Update the present map for a cpu node which is going away, and set
196 * the hard id in the paca(s) to -1 to be consistent with boot time
197 * convention for non-present cpus.
198 */
199static void pSeries_remove_processor(struct device_node *np)
200{
201	unsigned int cpu;
202	int len, nthreads, i;
203	u32 *intserv;
204
205	intserv = (u32 *)get_property(np, "ibm,ppc-interrupt-server#s", &len);
206	if (!intserv)
207		return;
208
209	nthreads = len / sizeof(u32);
210
211	lock_cpu_hotplug();
212	for (i = 0; i < nthreads; i++) {
213		for_each_present_cpu(cpu) {
214			if (get_hard_smp_processor_id(cpu) != intserv[i])
215				continue;
216			BUG_ON(cpu_online(cpu));
217			cpu_clear(cpu, cpu_present_map);
218			set_hard_smp_processor_id(cpu, -1);
219			break;
220		}
221		if (cpu == NR_CPUS)
222			printk(KERN_WARNING "Could not find cpu to remove "
223			       "with physical id 0x%x\n", intserv[i]);
224	}
225	unlock_cpu_hotplug();
226}
227
228static int pSeries_smp_notifier(struct notifier_block *nb, unsigned long action, void *node)
229{
230	int err = NOTIFY_OK;
231
232	switch (action) {
233	case PSERIES_RECONFIG_ADD:
234		if (pSeries_add_processor(node))
235			err = NOTIFY_BAD;
236		break;
237	case PSERIES_RECONFIG_REMOVE:
238		pSeries_remove_processor(node);
239		break;
240	default:
241		err = NOTIFY_DONE;
242		break;
243	}
244	return err;
245}
246
247static struct notifier_block pSeries_smp_nb = {
248	.notifier_call = pSeries_smp_notifier,
249};
250
251#endif /* CONFIG_HOTPLUG_CPU */
252
253/**
254 * smp_startup_cpu() - start the given cpu
255 *
256 * At boot time, there is nothing to do for primary threads which were
257 * started from Open Firmware.  For anything else, call RTAS with the
258 * appropriate start location.
259 *
260 * Returns:
261 *	0	- failure
262 *	1	- success
263 */
264static inline int __devinit smp_startup_cpu(unsigned int lcpu)
265{
266	int status;
267	unsigned long start_here = __pa((u32)*((unsigned long *)
268					       pSeries_secondary_smp_init));
269	unsigned int pcpu;
270
271	if (cpu_isset(lcpu, of_spin_map))
272		/* Already started by OF and sitting in spin loop */
273		return 1;
274
275	pcpu = get_hard_smp_processor_id(lcpu);
276
277	/* Fixup atomic count: it exited inside IRQ handler. */
278	paca[lcpu].__current->thread_info->preempt_count	= 0;
279
280	status = rtas_call(rtas_token("start-cpu"), 3, 1, NULL,
281			   pcpu, start_here, lcpu);
282	if (status != 0) {
283		printk(KERN_ERR "start-cpu failed: %i\n", status);
284		return 0;
285	}
286	return 1;
287}
288
289static inline void smp_xics_do_message(int cpu, int msg)
290{
291	set_bit(msg, &xics_ipi_message[cpu].value);
292	mb();
293	xics_cause_IPI(cpu);
294}
295
296static void smp_xics_message_pass(int target, int msg)
297{
298	unsigned int i;
299
300	if (target < NR_CPUS) {
301		smp_xics_do_message(target, msg);
302	} else {
303		for_each_online_cpu(i) {
304			if (target == MSG_ALL_BUT_SELF
305			    && i == smp_processor_id())
306				continue;
307			smp_xics_do_message(i, msg);
308		}
309	}
310}
311
312static int __init smp_xics_probe(void)
313{
314	xics_request_IPIs();
315
316	return cpus_weight(cpu_possible_map);
317}
318
319static void __devinit smp_xics_setup_cpu(int cpu)
320{
321	if (cpu != boot_cpuid)
322		xics_setup_cpu();
323
324	if (cur_cpu_spec->firmware_features & FW_FEATURE_SPLPAR)
325		vpa_init(cpu);
326
327	cpu_clear(cpu, of_spin_map);
328
329	/*
330	 * Put the calling processor into the GIQ.  This is really only
331	 * necessary from a secondary thread as the OF start-cpu interface
332	 * performs this function for us on primary threads.
333	 */
334	rtas_set_indicator(GLOBAL_INTERRUPT_QUEUE,
335		(1UL << interrupt_server_size) - 1 - default_distrib_server, 1);
336}
337
338static DEFINE_SPINLOCK(timebase_lock);
339static unsigned long timebase = 0;
340
341static void __devinit pSeries_give_timebase(void)
342{
343	spin_lock(&timebase_lock);
344	rtas_call(rtas_token("freeze-time-base"), 0, 1, NULL);
345	timebase = get_tb();
346	spin_unlock(&timebase_lock);
347
348	while (timebase)
349		barrier();
350	rtas_call(rtas_token("thaw-time-base"), 0, 1, NULL);
351}
352
353static void __devinit pSeries_take_timebase(void)
354{
355	while (!timebase)
356		barrier();
357	spin_lock(&timebase_lock);
358	set_tb(timebase >> 32, timebase & 0xffffffff);
359	timebase = 0;
360	spin_unlock(&timebase_lock);
361}
362
363static void __devinit smp_pSeries_kick_cpu(int nr)
364{
365	BUG_ON(nr < 0 || nr >= NR_CPUS);
366
367	if (!smp_startup_cpu(nr))
368		return;
369
370	/*
371	 * The processor is currently spinning, waiting for the
372	 * cpu_start field to become non-zero After we set cpu_start,
373	 * the processor will continue on to secondary_start
374	 */
375	paca[nr].cpu_start = 1;
376}
377
378static int smp_pSeries_cpu_bootable(unsigned int nr)
379{
380	/* Special case - we inhibit secondary thread startup
381	 * during boot if the user requests it.  Odd-numbered
382	 * cpus are assumed to be secondary threads.
383	 */
384	if (system_state < SYSTEM_RUNNING &&
385	    cur_cpu_spec->cpu_features & CPU_FTR_SMT &&
386	    !smt_enabled_at_boot && nr % 2 != 0)
387		return 0;
388
389	return 1;
390}
391
392static struct smp_ops_t pSeries_mpic_smp_ops = {
393	.message_pass	= smp_mpic_message_pass,
394	.probe		= smp_mpic_probe,
395	.kick_cpu	= smp_pSeries_kick_cpu,
396	.setup_cpu	= smp_mpic_setup_cpu,
397};
398
399static struct smp_ops_t pSeries_xics_smp_ops = {
400	.message_pass	= smp_xics_message_pass,
401	.probe		= smp_xics_probe,
402	.kick_cpu	= smp_pSeries_kick_cpu,
403	.setup_cpu	= smp_xics_setup_cpu,
404	.cpu_bootable	= smp_pSeries_cpu_bootable,
405};
406
407/* This is called very early */
408void __init smp_init_pSeries(void)
409{
410	int i;
411
412	DBG(" -> smp_init_pSeries()\n");
413
414	if (ppc64_interrupt_controller == IC_OPEN_PIC)
415		smp_ops = &pSeries_mpic_smp_ops;
416	else
417		smp_ops = &pSeries_xics_smp_ops;
418
419#ifdef CONFIG_HOTPLUG_CPU
420	smp_ops->cpu_disable = pSeries_cpu_disable;
421	smp_ops->cpu_die = pSeries_cpu_die;
422
423	/* Processors can be added/removed only on LPAR */
424	if (systemcfg->platform == PLATFORM_PSERIES_LPAR)
425		pSeries_reconfig_notifier_register(&pSeries_smp_nb);
426#endif
427
428	/* Mark threads which are still spinning in hold loops. */
429	if (cur_cpu_spec->cpu_features & CPU_FTR_SMT)
430		for_each_present_cpu(i) {
431			if (i % 2 == 0)
432				/*
433				 * Even-numbered logical cpus correspond to
434				 * primary threads.
435				 */
436				cpu_set(i, of_spin_map);
437		}
438	else
439		of_spin_map = cpu_present_map;
440
441	cpu_clear(boot_cpuid, of_spin_map);
442
443	/* Non-lpar has additional take/give timebase */
444	if (rtas_token("freeze-time-base") != RTAS_UNKNOWN_SERVICE) {
445		smp_ops->give_timebase = pSeries_give_timebase;
446		smp_ops->take_timebase = pSeries_take_timebase;
447	}
448
449	DBG(" <- smp_init_pSeries()\n");
450}
451