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 "offline_states.h"
37
/* This version can't take the spinlock, because it never returns */
38
static struct rtas_args rtas_stop_self_args = {
39
.token = RTAS_UNKNOWN_SERVICE,
42
.rets = &rtas_stop_self_args.args[0],
45
static DEFINE_PER_CPU(enum cpu_state_vals, preferred_offline_state) =
47
static DEFINE_PER_CPU(enum cpu_state_vals, current_state) = CPU_STATE_OFFLINE;
49
static enum cpu_state_vals default_offline_state = CPU_STATE_OFFLINE;
51
static int cede_offline_enabled __read_mostly = 1;
54
* Enable/disable cede_offline when available.
56
static int __init setup_cede_offline(char *str)
58
if (!strcmp(str, "off"))
59
cede_offline_enabled = 0;
60
else if (!strcmp(str, "on"))
61
cede_offline_enabled = 1;
67
__setup("cede_offline=", setup_cede_offline);
69
enum cpu_state_vals get_cpu_current_state(int cpu)
71
return per_cpu(current_state, cpu);
74
void set_cpu_current_state(int cpu, enum cpu_state_vals state)
76
per_cpu(current_state, cpu) = state;
79
enum cpu_state_vals get_preferred_offline_state(int cpu)
81
return per_cpu(preferred_offline_state, cpu);
84
void set_preferred_offline_state(int cpu, enum cpu_state_vals state)
86
per_cpu(preferred_offline_state, cpu) = state;
89
void set_default_offline_state(int cpu)
91
per_cpu(preferred_offline_state, cpu) = default_offline_state;
94
static void rtas_stop_self(void)
96
struct rtas_args *args = &rtas_stop_self_args;
100
BUG_ON(args->token == RTAS_UNKNOWN_SERVICE);
102
printk("cpu %u (hwid %u) Ready to die...\n",
103
smp_processor_id(), hard_smp_processor_id());
104
enter_rtas(__pa(args));
106
panic("Alas, I survived.\n");
109
static void pseries_mach_cpu_die(void)
111
unsigned int cpu = smp_processor_id();
112
unsigned int hwcpu = hard_smp_processor_id();
113
u8 cede_latency_hint = 0;
119
if (get_preferred_offline_state(cpu) == CPU_STATE_INACTIVE) {
120
set_cpu_current_state(cpu, CPU_STATE_INACTIVE);
121
if (ppc_md.suspend_disable_cpu)
122
ppc_md.suspend_disable_cpu();
124
cede_latency_hint = 2;
126
get_lppaca()->idle = 1;
127
if (!get_lppaca()->shared_proc)
128
get_lppaca()->donate_dedicated_cpu = 1;
130
while (get_preferred_offline_state(cpu) == CPU_STATE_INACTIVE) {
131
extended_cede_processor(cede_latency_hint);
134
if (!get_lppaca()->shared_proc)
135
get_lppaca()->donate_dedicated_cpu = 0;
136
get_lppaca()->idle = 0;
138
if (get_preferred_offline_state(cpu) == CPU_STATE_ONLINE) {
139
unregister_slb_shadow(hwcpu);
142
* Call to start_secondary_resume() will not return.
143
* Kernel stack will be reset and start_secondary()
144
* will be called to continue the online operation.
146
start_secondary_resume();
150
/* Requested state is CPU_STATE_OFFLINE at this point */
151
WARN_ON(get_preferred_offline_state(cpu) != CPU_STATE_OFFLINE);
153
set_cpu_current_state(cpu, CPU_STATE_OFFLINE);
154
unregister_slb_shadow(hwcpu);
157
/* Should never get here... */
162
static int pseries_cpu_disable(void)
164
int cpu = smp_processor_id();
166
set_cpu_online(cpu, false);
167
vdso_data->processorCount--;
169
/*fix boot_cpuid here*/
170
if (cpu == boot_cpuid)
171
boot_cpuid = cpumask_any(cpu_online_mask);
173
/* FIXME: abstract this to not be platform specific later on */
174
xics_migrate_irqs_away();
179
* pseries_cpu_die: Wait for the cpu to die.
180
* @cpu: logical processor id of the CPU whose death we're awaiting.
182
* This function is called from the context of the thread which is performing
183
* the cpu-offline. Here we wait for long enough to allow the cpu in question
184
* to self-destroy so that the cpu-offline thread can send the CPU_DEAD
187
* OTOH, pseries_mach_cpu_die() is called by the @cpu when it wants to
190
static void pseries_cpu_die(unsigned int cpu)
194
unsigned int pcpu = get_hard_smp_processor_id(cpu);
196
if (get_preferred_offline_state(cpu) == CPU_STATE_INACTIVE) {
198
for (tries = 0; tries < 5000; tries++) {
199
if (get_cpu_current_state(cpu) == CPU_STATE_INACTIVE) {
205
} else if (get_preferred_offline_state(cpu) == CPU_STATE_OFFLINE) {
207
for (tries = 0; tries < 25; tries++) {
208
cpu_status = smp_query_cpu_stopped(pcpu);
209
if (cpu_status == QCSS_STOPPED ||
210
cpu_status == QCSS_HARDWARE_ERROR)
216
if (cpu_status != 0) {
217
printk("Querying DEAD? cpu %i (%i) shows %i\n",
218
cpu, pcpu, cpu_status);
221
/* Isolation and deallocation are definitely done by
222
* drslot_chrp_cpu. If they were not they would be
223
* done here. Change isolate state to Isolate and
224
* change allocation-state to Unusable.
226
paca[cpu].cpu_start = 0;
230
* Update cpu_present_mask and paca(s) for a new cpu node. The wrinkle
231
* here is that a cpu device node may represent up to two logical cpus
232
* in the SMT case. We must honor the assumption in other code that
233
* the logical ids for sibling SMT threads x and y are adjacent, such
234
* that x^1 == y and y^1 == x.
236
static int pseries_add_processor(struct device_node *np)
239
cpumask_var_t candidate_mask, tmp;
240
int err = -ENOSPC, len, nthreads, i;
243
intserv = of_get_property(np, "ibm,ppc-interrupt-server#s", &len);
247
zalloc_cpumask_var(&candidate_mask, GFP_KERNEL);
248
zalloc_cpumask_var(&tmp, GFP_KERNEL);
250
nthreads = len / sizeof(u32);
251
for (i = 0; i < nthreads; i++)
252
cpumask_set_cpu(i, tmp);
254
cpu_maps_update_begin();
256
BUG_ON(!cpumask_subset(cpu_present_mask, cpu_possible_mask));
258
/* Get a bitmap of unoccupied slots. */
259
cpumask_xor(candidate_mask, cpu_possible_mask, cpu_present_mask);
260
if (cpumask_empty(candidate_mask)) {
261
/* If we get here, it most likely means that NR_CPUS is
262
* less than the partition's max processors setting.
264
printk(KERN_ERR "Cannot add cpu %s; this system configuration"
265
" supports %d logical cpus.\n", np->full_name,
266
cpumask_weight(cpu_possible_mask));
270
while (!cpumask_empty(tmp))
271
if (cpumask_subset(tmp, candidate_mask))
272
/* Found a range where we can insert the new cpu(s) */
275
cpumask_shift_left(tmp, tmp, nthreads);
277
if (cpumask_empty(tmp)) {
278
printk(KERN_ERR "Unable to find space in cpu_present_mask for"
279
" processor %s with %d thread(s)\n", np->name,
284
for_each_cpu(cpu, tmp) {
285
BUG_ON(cpu_present(cpu));
286
set_cpu_present(cpu, true);
287
set_hard_smp_processor_id(cpu, *intserv++);
291
cpu_maps_update_done();
292
free_cpumask_var(candidate_mask);
293
free_cpumask_var(tmp);
298
* Update the present map for a cpu node which is going away, and set
299
* the hard id in the paca(s) to -1 to be consistent with boot time
300
* convention for non-present cpus.
302
static void pseries_remove_processor(struct device_node *np)
305
int len, nthreads, i;
308
intserv = of_get_property(np, "ibm,ppc-interrupt-server#s", &len);
312
nthreads = len / sizeof(u32);
314
cpu_maps_update_begin();
315
for (i = 0; i < nthreads; i++) {
316
for_each_present_cpu(cpu) {
317
if (get_hard_smp_processor_id(cpu) != intserv[i])
319
BUG_ON(cpu_online(cpu));
320
set_cpu_present(cpu, false);
321
set_hard_smp_processor_id(cpu, -1);
324
if (cpu >= nr_cpu_ids)
325
printk(KERN_WARNING "Could not find cpu to remove "
326
"with physical id 0x%x\n", intserv[i]);
328
cpu_maps_update_done();
331
static int pseries_smp_notifier(struct notifier_block *nb,
332
unsigned long action, void *node)
337
case PSERIES_RECONFIG_ADD:
338
err = pseries_add_processor(node);
340
case PSERIES_RECONFIG_REMOVE:
341
pseries_remove_processor(node);
344
return notifier_from_errno(err);
347
static struct notifier_block pseries_smp_nb = {
348
.notifier_call = pseries_smp_notifier,
351
#define MAX_CEDE_LATENCY_LEVELS 4
352
#define CEDE_LATENCY_PARAM_LENGTH 10
353
#define CEDE_LATENCY_PARAM_MAX_LENGTH \
354
(MAX_CEDE_LATENCY_LEVELS * CEDE_LATENCY_PARAM_LENGTH * sizeof(char))
355
#define CEDE_LATENCY_TOKEN 45
357
static char cede_parameters[CEDE_LATENCY_PARAM_MAX_LENGTH];
359
static int parse_cede_parameters(void)
361
memset(cede_parameters, 0, CEDE_LATENCY_PARAM_MAX_LENGTH);
362
return rtas_call(rtas_token("ibm,get-system-parameter"), 3, 1,
365
__pa(cede_parameters),
366
CEDE_LATENCY_PARAM_MAX_LENGTH);
369
static int __init pseries_cpu_hotplug_init(void)
371
struct device_node *np;
376
for_each_node_by_name(np, "interrupt-controller") {
377
typep = of_get_property(np, "compatible", NULL);
378
if (strstr(typep, "open-pic")) {
381
printk(KERN_INFO "CPU Hotplug not supported on "
382
"systems using MPIC\n");
387
rtas_stop_self_args.token = rtas_token("stop-self");
388
qcss_tok = rtas_token("query-cpu-stopped-state");
390
if (rtas_stop_self_args.token == RTAS_UNKNOWN_SERVICE ||
391
qcss_tok == RTAS_UNKNOWN_SERVICE) {
392
printk(KERN_INFO "CPU Hotplug not supported by firmware "
397
ppc_md.cpu_die = pseries_mach_cpu_die;
398
smp_ops->cpu_disable = pseries_cpu_disable;
399
smp_ops->cpu_die = pseries_cpu_die;
401
/* Processors can be added/removed only on LPAR */
402
if (firmware_has_feature(FW_FEATURE_LPAR)) {
403
pSeries_reconfig_notifier_register(&pseries_smp_nb);
404
cpu_maps_update_begin();
405
if (cede_offline_enabled && parse_cede_parameters() == 0) {
406
default_offline_state = CPU_STATE_INACTIVE;
407
for_each_online_cpu(cpu)
408
set_default_offline_state(cpu);
410
cpu_maps_update_done();
415
arch_initcall(pseries_cpu_hotplug_init);