2 * SMP support for pSeries machines.
4 * Dave Engebretsen, Peter Bergner, and
5 * Mike Corrigan {engebret|bergner|mikec}@us.ibm.com
7 * Plus various changes from other IBM teams...
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.
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>
31 #include <asm/ptrace.h>
32 #include <asm/atomic.h>
35 #include <asm/pgtable.h>
41 #include <asm/machdep.h>
43 #include <asm/cputable.h>
44 #include <asm/firmware.h>
45 #include <asm/system.h>
47 #include <asm/plpar_wrappers.h>
48 #include <asm/pSeries_reconfig.h>
52 #define DBG(fmt...) udbg_printf(fmt)
58 * The primary thread of each non-boot processor is recorded here before
61 static cpumask_t of_spin_map;
63 extern void pSeries_secondary_smp_init(unsigned long);
65 #ifdef CONFIG_HOTPLUG_CPU
67 /* Get state of physical CPU.
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
73 * -2 - Hardware Busy, Try again later.
75 static int query_cpu_stopped(unsigned int pcpu)
80 qcss_tok = rtas_token("query-cpu-stopped-state");
81 if (qcss_tok == RTAS_UNKNOWN_SERVICE)
83 status = rtas_call(qcss_tok, 1, 2, &cpu_status, pcpu);
86 "RTAS query-cpu-stopped-state failed: %i\n", status);
93 int pSeries_cpu_disable(void)
95 int cpu = smp_processor_id();
97 cpu_clear(cpu, cpu_online_map);
98 systemcfg->processorCount--;
100 /*fix boot_cpuid here*/
101 if (cpu == boot_cpuid)
102 boot_cpuid = any_online_cpu(cpu_online_map);
104 /* FIXME: abstract this to not be platform specific later on */
105 xics_migrate_irqs_away();
109 void pSeries_cpu_die(unsigned int cpu)
113 unsigned int pcpu = get_hard_smp_processor_id(cpu);
115 for (tries = 0; tries < 25; tries++) {
116 cpu_status = query_cpu_stopped(pcpu);
117 if (cpu_status == 0 || cpu_status == -1)
121 if (cpu_status != 0) {
122 printk("Querying DEAD? cpu %i (%i) shows %i\n",
123 cpu, pcpu, cpu_status);
126 /* Isolation and deallocation are definatly done by
127 * drslot_chrp_cpu. If they were not they would be
128 * done here. Change isolate state to Isolate and
129 * change allocation-state to Unusable.
131 paca[cpu].cpu_start = 0;
135 * Update cpu_present_map and paca(s) for a new cpu node. The wrinkle
136 * here is that a cpu device node may represent up to two logical cpus
137 * in the SMT case. We must honor the assumption in other code that
138 * the logical ids for sibling SMT threads x and y are adjacent, such
139 * that x^1 == y and y^1 == x.
141 static int pSeries_add_processor(struct device_node *np)
144 cpumask_t candidate_map, tmp = CPU_MASK_NONE;
145 int err = -ENOSPC, len, nthreads, i;
148 intserv = (u32 *)get_property(np, "ibm,ppc-interrupt-server#s", &len);
152 nthreads = len / sizeof(u32);
153 for (i = 0; i < nthreads; i++)
158 BUG_ON(!cpus_subset(cpu_present_map, cpu_possible_map));
160 /* Get a bitmap of unoccupied slots. */
161 cpus_xor(candidate_map, cpu_possible_map, cpu_present_map);
162 if (cpus_empty(candidate_map)) {
163 /* If we get here, it most likely means that NR_CPUS is
164 * less than the partition's max processors setting.
166 printk(KERN_ERR "Cannot add cpu %s; this system configuration"
167 " supports %d logical cpus.\n", np->full_name,
168 cpus_weight(cpu_possible_map));
172 while (!cpus_empty(tmp))
173 if (cpus_subset(tmp, candidate_map))
174 /* Found a range where we can insert the new cpu(s) */
177 cpus_shift_left(tmp, tmp, nthreads);
179 if (cpus_empty(tmp)) {
180 printk(KERN_ERR "Unable to find space in cpu_present_map for"
181 " processor %s with %d thread(s)\n", np->name,
186 for_each_cpu_mask(cpu, tmp) {
187 BUG_ON(cpu_isset(cpu, cpu_present_map));
188 cpu_set(cpu, cpu_present_map);
189 set_hard_smp_processor_id(cpu, *intserv++);
193 unlock_cpu_hotplug();
198 * Update the present map for a cpu node which is going away, and set
199 * the hard id in the paca(s) to -1 to be consistent with boot time
200 * convention for non-present cpus.
202 static void pSeries_remove_processor(struct device_node *np)
205 int len, nthreads, i;
208 intserv = (u32 *)get_property(np, "ibm,ppc-interrupt-server#s", &len);
212 nthreads = len / sizeof(u32);
215 for (i = 0; i < nthreads; i++) {
216 for_each_present_cpu(cpu) {
217 if (get_hard_smp_processor_id(cpu) != intserv[i])
219 BUG_ON(cpu_online(cpu));
220 cpu_clear(cpu, cpu_present_map);
221 set_hard_smp_processor_id(cpu, -1);
225 printk(KERN_WARNING "Could not find cpu to remove "
226 "with physical id 0x%x\n", intserv[i]);
228 unlock_cpu_hotplug();
231 static int pSeries_smp_notifier(struct notifier_block *nb, unsigned long action, void *node)
236 case PSERIES_RECONFIG_ADD:
237 if (pSeries_add_processor(node))
240 case PSERIES_RECONFIG_REMOVE:
241 pSeries_remove_processor(node);
250 static struct notifier_block pSeries_smp_nb = {
251 .notifier_call = pSeries_smp_notifier,
254 #endif /* CONFIG_HOTPLUG_CPU */
257 * smp_startup_cpu() - start the given cpu
259 * At boot time, there is nothing to do for primary threads which were
260 * started from Open Firmware. For anything else, call RTAS with the
261 * appropriate start location.
267 static inline int __devinit smp_startup_cpu(unsigned int lcpu)
270 unsigned long start_here = __pa((u32)*((unsigned long *)
271 pSeries_secondary_smp_init));
275 if (cpu_isset(lcpu, of_spin_map))
276 /* Already started by OF and sitting in spin loop */
279 pcpu = get_hard_smp_processor_id(lcpu);
281 /* Fixup atomic count: it exited inside IRQ handler. */
282 paca[lcpu].__current->thread_info->preempt_count = 0;
285 * If the RTAS start-cpu token does not exist then presume the
286 * cpu is already spinning.
288 start_cpu = rtas_token("start-cpu");
289 if (start_cpu == RTAS_UNKNOWN_SERVICE)
292 status = rtas_call(start_cpu, 3, 1, NULL, pcpu, start_here, lcpu);
294 printk(KERN_ERR "start-cpu failed: %i\n", status);
302 static inline void smp_xics_do_message(int cpu, int msg)
304 set_bit(msg, &xics_ipi_message[cpu].value);
309 static void smp_xics_message_pass(int target, int msg)
313 if (target < NR_CPUS) {
314 smp_xics_do_message(target, msg);
316 for_each_online_cpu(i) {
317 if (target == MSG_ALL_BUT_SELF
318 && i == smp_processor_id())
320 smp_xics_do_message(i, msg);
325 static int __init smp_xics_probe(void)
329 return cpus_weight(cpu_possible_map);
332 static void __devinit smp_xics_setup_cpu(int cpu)
334 if (cpu != boot_cpuid)
337 if (firmware_has_feature(FW_FEATURE_SPLPAR))
340 cpu_clear(cpu, of_spin_map);
343 #endif /* CONFIG_XICS */
345 static DEFINE_SPINLOCK(timebase_lock);
346 static unsigned long timebase = 0;
348 static void __devinit pSeries_give_timebase(void)
350 spin_lock(&timebase_lock);
351 rtas_call(rtas_token("freeze-time-base"), 0, 1, NULL);
353 spin_unlock(&timebase_lock);
357 rtas_call(rtas_token("thaw-time-base"), 0, 1, NULL);
360 static void __devinit pSeries_take_timebase(void)
364 spin_lock(&timebase_lock);
365 set_tb(timebase >> 32, timebase & 0xffffffff);
367 spin_unlock(&timebase_lock);
370 static void __devinit smp_pSeries_kick_cpu(int nr)
372 BUG_ON(nr < 0 || nr >= NR_CPUS);
374 if (!smp_startup_cpu(nr))
378 * The processor is currently spinning, waiting for the
379 * cpu_start field to become non-zero After we set cpu_start,
380 * the processor will continue on to secondary_start
382 paca[nr].cpu_start = 1;
385 static int smp_pSeries_cpu_bootable(unsigned int nr)
387 /* Special case - we inhibit secondary thread startup
388 * during boot if the user requests it. Odd-numbered
389 * cpus are assumed to be secondary threads.
391 if (system_state < SYSTEM_RUNNING &&
392 cpu_has_feature(CPU_FTR_SMT) &&
393 !smt_enabled_at_boot && nr % 2 != 0)
399 static struct smp_ops_t pSeries_mpic_smp_ops = {
400 .message_pass = smp_mpic_message_pass,
401 .probe = smp_mpic_probe,
402 .kick_cpu = smp_pSeries_kick_cpu,
403 .setup_cpu = smp_mpic_setup_cpu,
407 static struct smp_ops_t pSeries_xics_smp_ops = {
408 .message_pass = smp_xics_message_pass,
409 .probe = smp_xics_probe,
410 .kick_cpu = smp_pSeries_kick_cpu,
411 .setup_cpu = smp_xics_setup_cpu,
412 .cpu_bootable = smp_pSeries_cpu_bootable,
416 /* This is called very early */
417 void __init smp_init_pSeries(void)
421 DBG(" -> smp_init_pSeries()\n");
423 switch (ppc64_interrupt_controller) {
426 smp_ops = &pSeries_mpic_smp_ops;
431 smp_ops = &pSeries_xics_smp_ops;
435 panic("Invalid interrupt controller");
438 #ifdef CONFIG_HOTPLUG_CPU
439 smp_ops->cpu_disable = pSeries_cpu_disable;
440 smp_ops->cpu_die = pSeries_cpu_die;
442 /* Processors can be added/removed only on LPAR */
443 if (systemcfg->platform == PLATFORM_PSERIES_LPAR)
444 pSeries_reconfig_notifier_register(&pSeries_smp_nb);
447 /* Mark threads which are still spinning in hold loops. */
448 if (cpu_has_feature(CPU_FTR_SMT)) {
449 for_each_present_cpu(i) {
452 * Even-numbered logical cpus correspond to
455 cpu_set(i, of_spin_map);
458 of_spin_map = cpu_present_map;
461 cpu_clear(boot_cpuid, of_spin_map);
463 /* Non-lpar has additional take/give timebase */
464 if (rtas_token("freeze-time-base") != RTAS_UNKNOWN_SERVICE) {
465 smp_ops->give_timebase = pSeries_give_timebase;
466 smp_ops->take_timebase = pSeries_take_timebase;
469 DBG(" <- smp_init_pSeries()\n");