2 * pseries CPU Hotplug infrastructure.
4 * Split out from arch/powerpc/platforms/pseries/setup.c
5 * arch/powerpc/kernel/rtas.c, and arch/powerpc/platforms/pseries/smp.c
7 * Peter Bergner, IBM March 2001.
8 * Copyright (C) 2001 IBM.
9 * Dave Engebretsen, Peter Bergner, and
10 * Mike Corrigan {engebret|bergner|mikec}@us.ibm.com
11 * Plus various changes from other IBM teams...
13 * Copyright (C) 2006 Michael Ellerman, IBM Corporation
15 * This program is free software; you can redistribute it and/or
16 * modify it under the terms of the GNU General Public License
17 * as published by the Free Software Foundation; either version
18 * 2 of the License, or (at your option) any later version.
21 #include <linux/kernel.h>
22 #include <linux/interrupt.h>
23 #include <linux/delay.h>
24 #include <linux/sched.h> /* for idle_task_exit */
25 #include <linux/cpu.h>
26 #include <asm/system.h>
29 #include <asm/firmware.h>
30 #include <asm/machdep.h>
31 #include <asm/vdso_datapage.h>
32 #include <asm/pSeries_reconfig.h>
34 #include "plpar_wrappers.h"
35 #include <asm/topology.h>
36 #include "offline_states.h"
38 /* This version can't take the spinlock, because it never returns */
39 static struct rtas_args rtas_stop_self_args = {
40 .token = RTAS_UNKNOWN_SERVICE,
43 .rets = &rtas_stop_self_args.args[0],
46 static DEFINE_PER_CPU(enum cpu_state_vals, preferred_offline_state) =
48 static DEFINE_PER_CPU(enum cpu_state_vals, current_state) = CPU_STATE_OFFLINE;
50 static enum cpu_state_vals default_offline_state = CPU_STATE_OFFLINE;
52 static int cede_offline_enabled __read_mostly = 1;
55 * Enable/disable cede_offline when available.
57 static int __init setup_cede_offline(char *str)
59 if (!strcmp(str, "off"))
60 cede_offline_enabled = 0;
61 else if (!strcmp(str, "on"))
62 cede_offline_enabled = 1;
68 __setup("cede_offline=", setup_cede_offline);
70 enum cpu_state_vals get_cpu_current_state(int cpu)
72 return per_cpu(current_state, cpu);
75 void set_cpu_current_state(int cpu, enum cpu_state_vals state)
77 per_cpu(current_state, cpu) = state;
80 enum cpu_state_vals get_preferred_offline_state(int cpu)
82 return per_cpu(preferred_offline_state, cpu);
85 void set_preferred_offline_state(int cpu, enum cpu_state_vals state)
87 per_cpu(preferred_offline_state, cpu) = state;
90 void set_default_offline_state(int cpu)
92 per_cpu(preferred_offline_state, cpu) = default_offline_state;
95 static void rtas_stop_self(void)
97 struct rtas_args *args = &rtas_stop_self_args;
101 BUG_ON(args->token == RTAS_UNKNOWN_SERVICE);
103 printk("cpu %u (hwid %u) Ready to die...\n",
104 smp_processor_id(), hard_smp_processor_id());
105 enter_rtas(__pa(args));
107 panic("Alas, I survived.\n");
110 static void pseries_mach_cpu_die(void)
112 unsigned int cpu = smp_processor_id();
113 unsigned int hwcpu = hard_smp_processor_id();
114 u8 cede_latency_hint = 0;
120 if (get_preferred_offline_state(cpu) == CPU_STATE_INACTIVE) {
121 set_cpu_current_state(cpu, CPU_STATE_INACTIVE);
122 if (ppc_md.suspend_disable_cpu)
123 ppc_md.suspend_disable_cpu();
125 cede_latency_hint = 2;
127 get_lppaca()->idle = 1;
128 if (!get_lppaca()->shared_proc)
129 get_lppaca()->donate_dedicated_cpu = 1;
131 while (get_preferred_offline_state(cpu) == CPU_STATE_INACTIVE) {
132 extended_cede_processor(cede_latency_hint);
135 if (!get_lppaca()->shared_proc)
136 get_lppaca()->donate_dedicated_cpu = 0;
137 get_lppaca()->idle = 0;
139 if (get_preferred_offline_state(cpu) == CPU_STATE_ONLINE) {
140 unregister_slb_shadow(hwcpu);
143 * Call to start_secondary_resume() will not return.
144 * Kernel stack will be reset and start_secondary()
145 * will be called to continue the online operation.
147 start_secondary_resume();
151 /* Requested state is CPU_STATE_OFFLINE at this point */
152 WARN_ON(get_preferred_offline_state(cpu) != CPU_STATE_OFFLINE);
154 set_cpu_current_state(cpu, CPU_STATE_OFFLINE);
155 unregister_slb_shadow(hwcpu);
158 /* Should never get here... */
163 static int pseries_cpu_disable(void)
165 int cpu = smp_processor_id();
167 set_cpu_online(cpu, false);
168 vdso_data->processorCount--;
170 /*fix boot_cpuid here*/
171 if (cpu == boot_cpuid)
172 boot_cpuid = cpumask_any(cpu_online_mask);
174 /* FIXME: abstract this to not be platform specific later on */
175 xics_migrate_irqs_away();
180 * pseries_cpu_die: Wait for the cpu to die.
181 * @cpu: logical processor id of the CPU whose death we're awaiting.
183 * This function is called from the context of the thread which is performing
184 * the cpu-offline. Here we wait for long enough to allow the cpu in question
185 * to self-destroy so that the cpu-offline thread can send the CPU_DEAD
188 * OTOH, pseries_mach_cpu_die() is called by the @cpu when it wants to
191 static void pseries_cpu_die(unsigned int cpu)
195 unsigned int pcpu = get_hard_smp_processor_id(cpu);
197 if (get_preferred_offline_state(cpu) == CPU_STATE_INACTIVE) {
199 for (tries = 0; tries < 5000; tries++) {
200 if (get_cpu_current_state(cpu) == CPU_STATE_INACTIVE) {
206 } else if (get_preferred_offline_state(cpu) == CPU_STATE_OFFLINE) {
208 for (tries = 0; tries < 25; tries++) {
209 cpu_status = smp_query_cpu_stopped(pcpu);
210 if (cpu_status == QCSS_STOPPED ||
211 cpu_status == QCSS_HARDWARE_ERROR)
217 if (cpu_status != 0) {
218 printk("Querying DEAD? cpu %i (%i) shows %i\n",
219 cpu, pcpu, cpu_status);
222 /* Isolation and deallocation are definitely done by
223 * drslot_chrp_cpu. If they were not they would be
224 * done here. Change isolate state to Isolate and
225 * change allocation-state to Unusable.
227 paca[cpu].cpu_start = 0;
231 * Update cpu_present_mask and paca(s) for a new cpu node. The wrinkle
232 * here is that a cpu device node may represent up to two logical cpus
233 * in the SMT case. We must honor the assumption in other code that
234 * the logical ids for sibling SMT threads x and y are adjacent, such
235 * that x^1 == y and y^1 == x.
237 static int pseries_add_processor(struct device_node *np)
240 cpumask_var_t candidate_mask, tmp;
241 int err = -ENOSPC, len, nthreads, i;
244 intserv = of_get_property(np, "ibm,ppc-interrupt-server#s", &len);
248 zalloc_cpumask_var(&candidate_mask, GFP_KERNEL);
249 zalloc_cpumask_var(&tmp, GFP_KERNEL);
251 nthreads = len / sizeof(u32);
252 for (i = 0; i < nthreads; i++)
253 cpumask_set_cpu(i, tmp);
255 cpu_maps_update_begin();
257 BUG_ON(!cpumask_subset(cpu_present_mask, cpu_possible_mask));
259 /* Get a bitmap of unoccupied slots. */
260 cpumask_xor(candidate_mask, cpu_possible_mask, cpu_present_mask);
261 if (cpumask_empty(candidate_mask)) {
262 /* If we get here, it most likely means that NR_CPUS is
263 * less than the partition's max processors setting.
265 printk(KERN_ERR "Cannot add cpu %s; this system configuration"
266 " supports %d logical cpus.\n", np->full_name,
267 cpumask_weight(cpu_possible_mask));
271 while (!cpumask_empty(tmp))
272 if (cpumask_subset(tmp, candidate_mask))
273 /* Found a range where we can insert the new cpu(s) */
276 cpumask_shift_left(tmp, tmp, nthreads);
278 if (cpumask_empty(tmp)) {
279 printk(KERN_ERR "Unable to find space in cpu_present_mask for"
280 " processor %s with %d thread(s)\n", np->name,
285 for_each_cpu(cpu, tmp) {
286 BUG_ON(cpu_present(cpu));
287 set_cpu_present(cpu, true);
288 set_hard_smp_processor_id(cpu, *intserv++);
292 cpu_maps_update_done();
293 free_cpumask_var(candidate_mask);
294 free_cpumask_var(tmp);
299 * Update the present map for a cpu node which is going away, and set
300 * the hard id in the paca(s) to -1 to be consistent with boot time
301 * convention for non-present cpus.
303 static void pseries_remove_processor(struct device_node *np)
306 int len, nthreads, i;
309 intserv = of_get_property(np, "ibm,ppc-interrupt-server#s", &len);
313 nthreads = len / sizeof(u32);
315 cpu_maps_update_begin();
316 for (i = 0; i < nthreads; i++) {
317 for_each_present_cpu(cpu) {
318 if (get_hard_smp_processor_id(cpu) != intserv[i])
320 BUG_ON(cpu_online(cpu));
321 set_cpu_present(cpu, false);
322 set_hard_smp_processor_id(cpu, -1);
323 update_numa_cpu_lookup_table(cpu, -1);
326 if (cpu >= nr_cpu_ids)
327 printk(KERN_WARNING "Could not find cpu to remove "
328 "with physical id 0x%x\n", intserv[i]);
330 cpu_maps_update_done();
333 static int pseries_smp_notifier(struct notifier_block *nb,
334 unsigned long action, void *node)
339 case PSERIES_RECONFIG_ADD:
340 err = pseries_add_processor(node);
342 case PSERIES_RECONFIG_REMOVE:
343 pseries_remove_processor(node);
346 return notifier_from_errno(err);
349 static struct notifier_block pseries_smp_nb = {
350 .notifier_call = pseries_smp_notifier,
353 #define MAX_CEDE_LATENCY_LEVELS 4
354 #define CEDE_LATENCY_PARAM_LENGTH 10
355 #define CEDE_LATENCY_PARAM_MAX_LENGTH \
356 (MAX_CEDE_LATENCY_LEVELS * CEDE_LATENCY_PARAM_LENGTH * sizeof(char))
357 #define CEDE_LATENCY_TOKEN 45
359 static char cede_parameters[CEDE_LATENCY_PARAM_MAX_LENGTH];
361 static int parse_cede_parameters(void)
363 memset(cede_parameters, 0, CEDE_LATENCY_PARAM_MAX_LENGTH);
364 return rtas_call(rtas_token("ibm,get-system-parameter"), 3, 1,
367 __pa(cede_parameters),
368 CEDE_LATENCY_PARAM_MAX_LENGTH);
371 static int __init pseries_cpu_hotplug_init(void)
373 struct device_node *np;
378 for_each_node_by_name(np, "interrupt-controller") {
379 typep = of_get_property(np, "compatible", NULL);
380 if (strstr(typep, "open-pic")) {
383 printk(KERN_INFO "CPU Hotplug not supported on "
384 "systems using MPIC\n");
389 rtas_stop_self_args.token = rtas_token("stop-self");
390 qcss_tok = rtas_token("query-cpu-stopped-state");
392 if (rtas_stop_self_args.token == RTAS_UNKNOWN_SERVICE ||
393 qcss_tok == RTAS_UNKNOWN_SERVICE) {
394 printk(KERN_INFO "CPU Hotplug not supported by firmware "
399 ppc_md.cpu_die = pseries_mach_cpu_die;
400 smp_ops->cpu_disable = pseries_cpu_disable;
401 smp_ops->cpu_die = pseries_cpu_die;
403 /* Processors can be added/removed only on LPAR */
404 if (firmware_has_feature(FW_FEATURE_LPAR)) {
405 pSeries_reconfig_notifier_register(&pseries_smp_nb);
406 cpu_maps_update_begin();
407 if (cede_offline_enabled && parse_cede_parameters() == 0) {
408 default_offline_state = CPU_STATE_INACTIVE;
409 for_each_online_cpu(cpu)
410 set_default_offline_state(cpu);
412 cpu_maps_update_done();
417 arch_initcall(pseries_cpu_hotplug_init);