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 #define pr_fmt(fmt) "pseries-hotplug-cpu: " fmt
23 #include <linux/kernel.h>
24 #include <linux/interrupt.h>
25 #include <linux/delay.h>
26 #include <linux/sched.h> /* for idle_task_exit */
27 #include <linux/cpu.h>
29 #include <linux/slab.h>
32 #include <asm/firmware.h>
33 #include <asm/machdep.h>
34 #include <asm/vdso_datapage.h>
36 #include <asm/plpar_wrappers.h>
39 #include "offline_states.h"
41 /* This version can't take the spinlock, because it never returns */
42 static int rtas_stop_self_token = RTAS_UNKNOWN_SERVICE;
44 static DEFINE_PER_CPU(enum cpu_state_vals, preferred_offline_state) =
46 static DEFINE_PER_CPU(enum cpu_state_vals, current_state) = CPU_STATE_OFFLINE;
48 static enum cpu_state_vals default_offline_state = CPU_STATE_OFFLINE;
50 static bool cede_offline_enabled __read_mostly = true;
53 * Enable/disable cede_offline when available.
55 static int __init setup_cede_offline(char *str)
57 return (kstrtobool(str, &cede_offline_enabled) == 0);
60 __setup("cede_offline=", setup_cede_offline);
62 enum cpu_state_vals get_cpu_current_state(int cpu)
64 return per_cpu(current_state, cpu);
67 void set_cpu_current_state(int cpu, enum cpu_state_vals state)
69 per_cpu(current_state, cpu) = state;
72 enum cpu_state_vals get_preferred_offline_state(int cpu)
74 return per_cpu(preferred_offline_state, cpu);
77 void set_preferred_offline_state(int cpu, enum cpu_state_vals state)
79 per_cpu(preferred_offline_state, cpu) = state;
82 void set_default_offline_state(int cpu)
84 per_cpu(preferred_offline_state, cpu) = default_offline_state;
87 static void rtas_stop_self(void)
89 static struct rtas_args args;
93 BUG_ON(rtas_stop_self_token == RTAS_UNKNOWN_SERVICE);
95 rtas_call_unlocked(&args, rtas_stop_self_token, 0, 1, NULL);
97 panic("Alas, I survived.\n");
100 static void pseries_mach_cpu_die(void)
102 unsigned int cpu = smp_processor_id();
103 unsigned int hwcpu = hard_smp_processor_id();
104 u8 cede_latency_hint = 0;
110 if (get_preferred_offline_state(cpu) == CPU_STATE_INACTIVE) {
111 set_cpu_current_state(cpu, CPU_STATE_INACTIVE);
112 if (ppc_md.suspend_disable_cpu)
113 ppc_md.suspend_disable_cpu();
115 cede_latency_hint = 2;
117 get_lppaca()->idle = 1;
118 if (!lppaca_shared_proc(get_lppaca()))
119 get_lppaca()->donate_dedicated_cpu = 1;
121 while (get_preferred_offline_state(cpu) == CPU_STATE_INACTIVE) {
122 while (!prep_irq_for_idle()) {
127 extended_cede_processor(cede_latency_hint);
132 if (!lppaca_shared_proc(get_lppaca()))
133 get_lppaca()->donate_dedicated_cpu = 0;
134 get_lppaca()->idle = 0;
136 if (get_preferred_offline_state(cpu) == CPU_STATE_ONLINE) {
137 unregister_slb_shadow(hwcpu);
141 * Call to start_secondary_resume() will not return.
142 * Kernel stack will be reset and start_secondary()
143 * will be called to continue the online operation.
145 start_secondary_resume();
149 /* Requested state is CPU_STATE_OFFLINE at this point */
150 WARN_ON(get_preferred_offline_state(cpu) != CPU_STATE_OFFLINE);
152 set_cpu_current_state(cpu, CPU_STATE_OFFLINE);
153 unregister_slb_shadow(hwcpu);
156 /* Should never get here... */
161 static int pseries_cpu_disable(void)
163 int cpu = smp_processor_id();
165 set_cpu_online(cpu, false);
166 vdso_data->processorCount--;
168 /*fix boot_cpuid here*/
169 if (cpu == boot_cpuid)
170 boot_cpuid = cpumask_any(cpu_online_mask);
172 /* FIXME: abstract this to not be platform specific later on */
173 xics_migrate_irqs_away();
178 * pseries_cpu_die: Wait for the cpu to die.
179 * @cpu: logical processor id of the CPU whose death we're awaiting.
181 * This function is called from the context of the thread which is performing
182 * the cpu-offline. Here we wait for long enough to allow the cpu in question
183 * to self-destroy so that the cpu-offline thread can send the CPU_DEAD
186 * OTOH, pseries_mach_cpu_die() is called by the @cpu when it wants to
189 static void pseries_cpu_die(unsigned int cpu)
193 unsigned int pcpu = get_hard_smp_processor_id(cpu);
195 if (get_preferred_offline_state(cpu) == CPU_STATE_INACTIVE) {
197 for (tries = 0; tries < 5000; tries++) {
198 if (get_cpu_current_state(cpu) == CPU_STATE_INACTIVE) {
204 } else if (get_preferred_offline_state(cpu) == CPU_STATE_OFFLINE) {
206 for (tries = 0; tries < 25; tries++) {
207 cpu_status = smp_query_cpu_stopped(pcpu);
208 if (cpu_status == QCSS_STOPPED ||
209 cpu_status == QCSS_HARDWARE_ERROR)
215 if (cpu_status != 0) {
216 printk("Querying DEAD? cpu %i (%i) shows %i\n",
217 cpu, pcpu, cpu_status);
220 /* Isolation and deallocation are definitely done by
221 * drslot_chrp_cpu. If they were not they would be
222 * done here. Change isolate state to Isolate and
223 * change allocation-state to Unusable.
225 paca[cpu].cpu_start = 0;
229 * Update cpu_present_mask and paca(s) for a new cpu node. The wrinkle
230 * here is that a cpu device node may represent up to two logical cpus
231 * in the SMT case. We must honor the assumption in other code that
232 * the logical ids for sibling SMT threads x and y are adjacent, such
233 * that x^1 == y and y^1 == x.
235 static int pseries_add_processor(struct device_node *np)
238 cpumask_var_t candidate_mask, tmp;
239 int err = -ENOSPC, len, nthreads, i;
240 const __be32 *intserv;
242 intserv = of_get_property(np, "ibm,ppc-interrupt-server#s", &len);
246 zalloc_cpumask_var(&candidate_mask, GFP_KERNEL);
247 zalloc_cpumask_var(&tmp, GFP_KERNEL);
249 nthreads = len / sizeof(u32);
250 for (i = 0; i < nthreads; i++)
251 cpumask_set_cpu(i, tmp);
253 cpu_maps_update_begin();
255 BUG_ON(!cpumask_subset(cpu_present_mask, cpu_possible_mask));
257 /* Get a bitmap of unoccupied slots. */
258 cpumask_xor(candidate_mask, cpu_possible_mask, cpu_present_mask);
259 if (cpumask_empty(candidate_mask)) {
260 /* If we get here, it most likely means that NR_CPUS is
261 * less than the partition's max processors setting.
263 printk(KERN_ERR "Cannot add cpu %s; this system configuration"
264 " supports %d logical cpus.\n", np->full_name,
265 num_possible_cpus());
269 while (!cpumask_empty(tmp))
270 if (cpumask_subset(tmp, candidate_mask))
271 /* Found a range where we can insert the new cpu(s) */
274 cpumask_shift_left(tmp, tmp, nthreads);
276 if (cpumask_empty(tmp)) {
277 printk(KERN_ERR "Unable to find space in cpu_present_mask for"
278 " processor %s with %d thread(s)\n", np->name,
283 for_each_cpu(cpu, tmp) {
284 BUG_ON(cpu_present(cpu));
285 set_cpu_present(cpu, true);
286 set_hard_smp_processor_id(cpu, be32_to_cpu(*intserv++));
290 cpu_maps_update_done();
291 free_cpumask_var(candidate_mask);
292 free_cpumask_var(tmp);
297 * Update the present map for a cpu node which is going away, and set
298 * the hard id in the paca(s) to -1 to be consistent with boot time
299 * convention for non-present cpus.
301 static void pseries_remove_processor(struct device_node *np)
304 int len, nthreads, i;
305 const __be32 *intserv;
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 thread = be32_to_cpu(intserv[i]);
317 for_each_present_cpu(cpu) {
318 if (get_hard_smp_processor_id(cpu) != thread)
320 BUG_ON(cpu_online(cpu));
321 set_cpu_present(cpu, false);
322 set_hard_smp_processor_id(cpu, -1);
325 if (cpu >= nr_cpu_ids)
326 printk(KERN_WARNING "Could not find cpu to remove "
327 "with physical id 0x%x\n", thread);
329 cpu_maps_update_done();
332 static int dlpar_online_cpu(struct device_node *dn)
336 int len, nthreads, i;
337 const __be32 *intserv;
340 intserv = of_get_property(dn, "ibm,ppc-interrupt-server#s", &len);
344 nthreads = len / sizeof(u32);
346 cpu_maps_update_begin();
347 for (i = 0; i < nthreads; i++) {
348 thread = be32_to_cpu(intserv[i]);
349 for_each_present_cpu(cpu) {
350 if (get_hard_smp_processor_id(cpu) != thread)
352 BUG_ON(get_cpu_current_state(cpu)
353 != CPU_STATE_OFFLINE);
354 cpu_maps_update_done();
355 rc = device_online(get_cpu_device(cpu));
358 cpu_maps_update_begin();
362 if (cpu == num_possible_cpus())
363 printk(KERN_WARNING "Could not find cpu to online "
364 "with physical id 0x%x\n", thread);
366 cpu_maps_update_done();
373 static bool dlpar_cpu_exists(struct device_node *parent, u32 drc_index)
375 struct device_node *child = NULL;
380 /* Assume cpu doesn't exist */
383 for_each_child_of_node(parent, child) {
384 rc = of_property_read_u32(child, "ibm,my-drc-index",
389 if (my_drc_index == drc_index) {
399 static bool valid_cpu_drc_index(struct device_node *parent, u32 drc_index)
408 rc = of_property_read_u32_index(parent, "ibm,drc-indexes",
413 if (drc == drc_index)
420 static ssize_t dlpar_cpu_add(u32 drc_index)
422 struct device_node *dn, *parent;
425 pr_debug("Attempting to add CPU, drc index: %x\n", drc_index);
427 parent = of_find_node_by_path("/cpus");
429 pr_warn("Failed to find CPU root node \"/cpus\"\n");
433 if (dlpar_cpu_exists(parent, drc_index)) {
435 pr_warn("CPU with drc index %x already exists\n", drc_index);
439 if (!valid_cpu_drc_index(parent, drc_index)) {
441 pr_warn("Cannot find CPU (drc index %x) to add.\n", drc_index);
445 rc = dlpar_acquire_drc(drc_index);
447 pr_warn("Failed to acquire DRC, rc: %d, drc index: %x\n",
453 dn = dlpar_configure_connector(cpu_to_be32(drc_index), parent);
456 pr_warn("Failed call to configure-connector, drc index: %x\n",
458 dlpar_release_drc(drc_index);
462 rc = dlpar_attach_node(dn);
465 pr_warn("Failed to attach node %s, rc: %d, drc index: %x\n",
466 dn->name, rc, drc_index);
468 rc = dlpar_release_drc(drc_index);
470 dlpar_free_cc_nodes(dn);
475 rc = dlpar_online_cpu(dn);
478 pr_warn("Failed to online cpu %s, rc: %d, drc index: %x\n",
479 dn->name, rc, drc_index);
481 rc = dlpar_detach_node(dn);
483 dlpar_release_drc(drc_index);
488 pr_debug("Successfully added CPU %s, drc index: %x\n", dn->name,
493 static int dlpar_offline_cpu(struct device_node *dn)
497 int len, nthreads, i;
498 const __be32 *intserv;
501 intserv = of_get_property(dn, "ibm,ppc-interrupt-server#s", &len);
505 nthreads = len / sizeof(u32);
507 cpu_maps_update_begin();
508 for (i = 0; i < nthreads; i++) {
509 thread = be32_to_cpu(intserv[i]);
510 for_each_present_cpu(cpu) {
511 if (get_hard_smp_processor_id(cpu) != thread)
514 if (get_cpu_current_state(cpu) == CPU_STATE_OFFLINE)
517 if (get_cpu_current_state(cpu) == CPU_STATE_ONLINE) {
518 set_preferred_offline_state(cpu,
520 cpu_maps_update_done();
521 rc = device_offline(get_cpu_device(cpu));
524 cpu_maps_update_begin();
530 * The cpu is in CPU_STATE_INACTIVE.
531 * Upgrade it's state to CPU_STATE_OFFLINE.
533 set_preferred_offline_state(cpu, CPU_STATE_OFFLINE);
534 BUG_ON(plpar_hcall_norets(H_PROD, thread)
539 if (cpu == num_possible_cpus())
540 printk(KERN_WARNING "Could not find cpu to offline with physical id 0x%x\n", thread);
542 cpu_maps_update_done();
549 static ssize_t dlpar_cpu_remove(struct device_node *dn, u32 drc_index)
553 pr_debug("Attemping to remove CPU %s, drc index: %x\n",
554 dn->name, drc_index);
556 rc = dlpar_offline_cpu(dn);
558 pr_warn("Failed to offline CPU %s, rc: %d\n", dn->name, rc);
562 rc = dlpar_release_drc(drc_index);
564 pr_warn("Failed to release drc (%x) for CPU %s, rc: %d\n",
565 drc_index, dn->name, rc);
566 dlpar_online_cpu(dn);
570 rc = dlpar_detach_node(dn);
574 pr_warn("Failed to detach CPU %s, rc: %d", dn->name, rc);
576 rc = dlpar_acquire_drc(drc_index);
578 dlpar_online_cpu(dn);
583 pr_debug("Successfully removed CPU, drc index: %x\n", drc_index);
587 static struct device_node *cpu_drc_index_to_dn(u32 drc_index)
589 struct device_node *dn;
593 for_each_node_by_type(dn, "cpu") {
594 rc = of_property_read_u32(dn, "ibm,my-drc-index", &my_index);
598 if (my_index == drc_index)
605 static int dlpar_cpu_remove_by_index(u32 drc_index)
607 struct device_node *dn;
610 dn = cpu_drc_index_to_dn(drc_index);
612 pr_warn("Cannot find CPU (drc index %x) to remove\n",
617 rc = dlpar_cpu_remove(dn, drc_index);
622 static int find_dlpar_cpus_to_remove(u32 *cpu_drcs, int cpus_to_remove)
624 struct device_node *dn;
628 /* We want to find cpus_to_remove + 1 CPUs to ensure we do not
629 * remove the last CPU.
631 for_each_node_by_type(dn, "cpu") {
634 if (cpus_found > cpus_to_remove) {
639 /* Note that cpus_found is always 1 ahead of the index
640 * into the cpu_drcs array, so we use cpus_found - 1
642 rc = of_property_read_u32(dn, "ibm,my-drc-index",
643 &cpu_drcs[cpus_found - 1]);
645 pr_warn("Error occurred getting drc-index for %s\n",
652 if (cpus_found < cpus_to_remove) {
653 pr_warn("Failed to find enough CPUs (%d of %d) to remove\n",
654 cpus_found, cpus_to_remove);
655 } else if (cpus_found == cpus_to_remove) {
656 pr_warn("Cannot remove all CPUs\n");
662 static int dlpar_cpu_remove_by_count(u32 cpus_to_remove)
666 int cpus_removed = 0;
669 pr_debug("Attempting to hot-remove %d CPUs\n", cpus_to_remove);
671 cpu_drcs = kcalloc(cpus_to_remove, sizeof(*cpu_drcs), GFP_KERNEL);
675 cpus_found = find_dlpar_cpus_to_remove(cpu_drcs, cpus_to_remove);
676 if (cpus_found <= cpus_to_remove) {
681 for (i = 0; i < cpus_to_remove; i++) {
682 rc = dlpar_cpu_remove_by_index(cpu_drcs[i]);
689 if (cpus_removed != cpus_to_remove) {
690 pr_warn("CPU hot-remove failed, adding back removed CPUs\n");
692 for (i = 0; i < cpus_removed; i++)
693 dlpar_cpu_add(cpu_drcs[i]);
704 static int find_dlpar_cpus_to_add(u32 *cpu_drcs, u32 cpus_to_add)
706 struct device_node *parent;
710 parent = of_find_node_by_path("/cpus");
712 pr_warn("Could not find CPU root node in device tree\n");
717 /* Search the ibm,drc-indexes array for possible CPU drcs to
718 * add. Note that the format of the ibm,drc-indexes array is
719 * the number of entries in the array followed by the array
720 * of drc values so we start looking at index = 1.
723 while (cpus_found < cpus_to_add) {
726 rc = of_property_read_u32_index(parent, "ibm,drc-indexes",
731 if (dlpar_cpu_exists(parent, drc))
734 cpu_drcs[cpus_found++] = drc;
741 static int dlpar_cpu_add_by_count(u32 cpus_to_add)
748 pr_debug("Attempting to hot-add %d CPUs\n", cpus_to_add);
750 cpu_drcs = kcalloc(cpus_to_add, sizeof(*cpu_drcs), GFP_KERNEL);
754 cpus_found = find_dlpar_cpus_to_add(cpu_drcs, cpus_to_add);
755 if (cpus_found < cpus_to_add) {
756 pr_warn("Failed to find enough CPUs (%d of %d) to add\n",
757 cpus_found, cpus_to_add);
762 for (i = 0; i < cpus_to_add; i++) {
763 rc = dlpar_cpu_add(cpu_drcs[i]);
770 if (cpus_added < cpus_to_add) {
771 pr_warn("CPU hot-add failed, removing any added CPUs\n");
773 for (i = 0; i < cpus_added; i++)
774 dlpar_cpu_remove_by_index(cpu_drcs[i]);
785 int dlpar_cpu_readd(int cpu)
787 struct device_node *dn;
792 dev = get_cpu_device(cpu);
795 rc = of_property_read_u32(dn, "ibm,my-drc-index", &drc_index);
797 rc = dlpar_cpu_remove_by_index(drc_index);
799 rc = dlpar_cpu_add(drc_index);
804 int dlpar_cpu(struct pseries_hp_errorlog *hp_elog)
806 u32 count, drc_index;
809 count = hp_elog->_drc_u.drc_count;
810 drc_index = hp_elog->_drc_u.drc_index;
812 lock_device_hotplug();
814 switch (hp_elog->action) {
815 case PSERIES_HP_ELOG_ACTION_REMOVE:
816 if (hp_elog->id_type == PSERIES_HP_ELOG_ID_DRC_COUNT)
817 rc = dlpar_cpu_remove_by_count(count);
818 else if (hp_elog->id_type == PSERIES_HP_ELOG_ID_DRC_INDEX)
819 rc = dlpar_cpu_remove_by_index(drc_index);
823 case PSERIES_HP_ELOG_ACTION_ADD:
824 if (hp_elog->id_type == PSERIES_HP_ELOG_ID_DRC_COUNT)
825 rc = dlpar_cpu_add_by_count(count);
826 else if (hp_elog->id_type == PSERIES_HP_ELOG_ID_DRC_INDEX)
827 rc = dlpar_cpu_add(drc_index);
832 pr_err("Invalid action (%d) specified\n", hp_elog->action);
837 unlock_device_hotplug();
841 #ifdef CONFIG_ARCH_CPU_PROBE_RELEASE
843 static ssize_t dlpar_cpu_probe(const char *buf, size_t count)
848 rc = kstrtou32(buf, 0, &drc_index);
852 rc = dlpar_cpu_add(drc_index);
854 return rc ? rc : count;
857 static ssize_t dlpar_cpu_release(const char *buf, size_t count)
859 struct device_node *dn;
863 dn = of_find_node_by_path(buf);
867 rc = of_property_read_u32(dn, "ibm,my-drc-index", &drc_index);
873 rc = dlpar_cpu_remove(dn, drc_index);
876 return rc ? rc : count;
879 #endif /* CONFIG_ARCH_CPU_PROBE_RELEASE */
881 static int pseries_smp_notifier(struct notifier_block *nb,
882 unsigned long action, void *data)
884 struct of_reconfig_data *rd = data;
888 case OF_RECONFIG_ATTACH_NODE:
889 err = pseries_add_processor(rd->dn);
891 case OF_RECONFIG_DETACH_NODE:
892 pseries_remove_processor(rd->dn);
895 return notifier_from_errno(err);
898 static struct notifier_block pseries_smp_nb = {
899 .notifier_call = pseries_smp_notifier,
902 #define MAX_CEDE_LATENCY_LEVELS 4
903 #define CEDE_LATENCY_PARAM_LENGTH 10
904 #define CEDE_LATENCY_PARAM_MAX_LENGTH \
905 (MAX_CEDE_LATENCY_LEVELS * CEDE_LATENCY_PARAM_LENGTH * sizeof(char))
906 #define CEDE_LATENCY_TOKEN 45
908 static char cede_parameters[CEDE_LATENCY_PARAM_MAX_LENGTH];
910 static int parse_cede_parameters(void)
912 memset(cede_parameters, 0, CEDE_LATENCY_PARAM_MAX_LENGTH);
913 return rtas_call(rtas_token("ibm,get-system-parameter"), 3, 1,
916 __pa(cede_parameters),
917 CEDE_LATENCY_PARAM_MAX_LENGTH);
920 static int __init pseries_cpu_hotplug_init(void)
925 #ifdef CONFIG_ARCH_CPU_PROBE_RELEASE
926 ppc_md.cpu_probe = dlpar_cpu_probe;
927 ppc_md.cpu_release = dlpar_cpu_release;
928 #endif /* CONFIG_ARCH_CPU_PROBE_RELEASE */
930 rtas_stop_self_token = rtas_token("stop-self");
931 qcss_tok = rtas_token("query-cpu-stopped-state");
933 if (rtas_stop_self_token == RTAS_UNKNOWN_SERVICE ||
934 qcss_tok == RTAS_UNKNOWN_SERVICE) {
935 printk(KERN_INFO "CPU Hotplug not supported by firmware "
940 ppc_md.cpu_die = pseries_mach_cpu_die;
941 smp_ops->cpu_disable = pseries_cpu_disable;
942 smp_ops->cpu_die = pseries_cpu_die;
944 /* Processors can be added/removed only on LPAR */
945 if (firmware_has_feature(FW_FEATURE_LPAR)) {
946 of_reconfig_notifier_register(&pseries_smp_nb);
947 cpu_maps_update_begin();
948 if (cede_offline_enabled && parse_cede_parameters() == 0) {
949 default_offline_state = CPU_STATE_INACTIVE;
950 for_each_online_cpu(cpu)
951 set_default_offline_state(cpu);
953 cpu_maps_update_done();
958 machine_arch_initcall(pseries, pseries_cpu_hotplug_init);