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/sched/hotplug.h>
28 #include <linux/cpu.h>
30 #include <linux/slab.h>
33 #include <asm/firmware.h>
34 #include <asm/machdep.h>
35 #include <asm/vdso_datapage.h>
38 #include <asm/plpar_wrappers.h>
39 #include <asm/topology.h>
42 #include "offline_states.h"
44 /* This version can't take the spinlock, because it never returns */
45 static int rtas_stop_self_token = RTAS_UNKNOWN_SERVICE;
47 static DEFINE_PER_CPU(enum cpu_state_vals, preferred_offline_state) =
49 static DEFINE_PER_CPU(enum cpu_state_vals, current_state) = CPU_STATE_OFFLINE;
51 static enum cpu_state_vals default_offline_state = CPU_STATE_OFFLINE;
53 static bool cede_offline_enabled __read_mostly = true;
56 * Enable/disable cede_offline when available.
58 static int __init setup_cede_offline(char *str)
60 return (kstrtobool(str, &cede_offline_enabled) == 0);
63 __setup("cede_offline=", setup_cede_offline);
65 enum cpu_state_vals get_cpu_current_state(int cpu)
67 return per_cpu(current_state, cpu);
70 void set_cpu_current_state(int cpu, enum cpu_state_vals state)
72 per_cpu(current_state, cpu) = state;
75 enum cpu_state_vals get_preferred_offline_state(int cpu)
77 return per_cpu(preferred_offline_state, cpu);
80 void set_preferred_offline_state(int cpu, enum cpu_state_vals state)
82 per_cpu(preferred_offline_state, cpu) = state;
85 void set_default_offline_state(int cpu)
87 per_cpu(preferred_offline_state, cpu) = default_offline_state;
90 static void rtas_stop_self(void)
92 static struct rtas_args args;
96 BUG_ON(rtas_stop_self_token == RTAS_UNKNOWN_SERVICE);
98 printk("cpu %u (hwid %u) Ready to die...\n",
99 smp_processor_id(), hard_smp_processor_id());
101 rtas_call_unlocked(&args, rtas_stop_self_token, 0, 1, NULL);
103 panic("Alas, I survived.\n");
106 static void pseries_mach_cpu_die(void)
108 unsigned int cpu = smp_processor_id();
109 unsigned int hwcpu = hard_smp_processor_id();
110 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 (!lppaca_shared_proc(get_lppaca()))
128 get_lppaca()->donate_dedicated_cpu = 1;
130 while (get_preferred_offline_state(cpu) == CPU_STATE_INACTIVE) {
131 while (!prep_irq_for_idle()) {
136 extended_cede_processor(cede_latency_hint);
141 if (!lppaca_shared_proc(get_lppaca()))
142 get_lppaca()->donate_dedicated_cpu = 0;
143 get_lppaca()->idle = 0;
145 if (get_preferred_offline_state(cpu) == CPU_STATE_ONLINE) {
146 unregister_slb_shadow(hwcpu);
150 * Call to start_secondary_resume() will not return.
151 * Kernel stack will be reset and start_secondary()
152 * will be called to continue the online operation.
154 start_secondary_resume();
158 /* Requested state is CPU_STATE_OFFLINE at this point */
159 WARN_ON(get_preferred_offline_state(cpu) != CPU_STATE_OFFLINE);
161 set_cpu_current_state(cpu, CPU_STATE_OFFLINE);
162 unregister_slb_shadow(hwcpu);
165 /* Should never get here... */
170 static int pseries_cpu_disable(void)
172 int cpu = smp_processor_id();
174 set_cpu_online(cpu, false);
175 vdso_data->processorCount--;
177 /*fix boot_cpuid here*/
178 if (cpu == boot_cpuid)
179 boot_cpuid = cpumask_any(cpu_online_mask);
181 /* FIXME: abstract this to not be platform specific later on */
183 xive_smp_disable_cpu();
185 xics_migrate_irqs_away();
190 * pseries_cpu_die: Wait for the cpu to die.
191 * @cpu: logical processor id of the CPU whose death we're awaiting.
193 * This function is called from the context of the thread which is performing
194 * the cpu-offline. Here we wait for long enough to allow the cpu in question
195 * to self-destroy so that the cpu-offline thread can send the CPU_DEAD
198 * OTOH, pseries_mach_cpu_die() is called by the @cpu when it wants to
201 static void pseries_cpu_die(unsigned int cpu)
205 unsigned int pcpu = get_hard_smp_processor_id(cpu);
207 if (get_preferred_offline_state(cpu) == CPU_STATE_INACTIVE) {
209 for (tries = 0; tries < 5000; tries++) {
210 if (get_cpu_current_state(cpu) == CPU_STATE_INACTIVE) {
216 } else if (get_preferred_offline_state(cpu) == CPU_STATE_OFFLINE) {
218 for (tries = 0; tries < 25; tries++) {
219 cpu_status = smp_query_cpu_stopped(pcpu);
220 if (cpu_status == QCSS_STOPPED ||
221 cpu_status == QCSS_HARDWARE_ERROR)
227 if (cpu_status != 0) {
228 printk("Querying DEAD? cpu %i (%i) shows %i\n",
229 cpu, pcpu, cpu_status);
232 /* Isolation and deallocation are definitely done by
233 * drslot_chrp_cpu. If they were not they would be
234 * done here. Change isolate state to Isolate and
235 * change allocation-state to Unusable.
237 paca_ptrs[cpu]->cpu_start = 0;
241 * Update cpu_present_mask and paca(s) for a new cpu node. The wrinkle
242 * here is that a cpu device node may represent up to two logical cpus
243 * in the SMT case. We must honor the assumption in other code that
244 * the logical ids for sibling SMT threads x and y are adjacent, such
245 * that x^1 == y and y^1 == x.
247 static int pseries_add_processor(struct device_node *np)
250 cpumask_var_t candidate_mask, tmp;
251 int err = -ENOSPC, len, nthreads, i;
252 const __be32 *intserv;
254 intserv = of_get_property(np, "ibm,ppc-interrupt-server#s", &len);
258 zalloc_cpumask_var(&candidate_mask, GFP_KERNEL);
259 zalloc_cpumask_var(&tmp, GFP_KERNEL);
261 nthreads = len / sizeof(u32);
262 for (i = 0; i < nthreads; i++)
263 cpumask_set_cpu(i, tmp);
265 cpu_maps_update_begin();
267 BUG_ON(!cpumask_subset(cpu_present_mask, cpu_possible_mask));
269 /* Get a bitmap of unoccupied slots. */
270 cpumask_xor(candidate_mask, cpu_possible_mask, cpu_present_mask);
271 if (cpumask_empty(candidate_mask)) {
272 /* If we get here, it most likely means that NR_CPUS is
273 * less than the partition's max processors setting.
275 printk(KERN_ERR "Cannot add cpu %pOF; this system configuration"
276 " supports %d logical cpus.\n", np,
277 num_possible_cpus());
281 while (!cpumask_empty(tmp))
282 if (cpumask_subset(tmp, candidate_mask))
283 /* Found a range where we can insert the new cpu(s) */
286 cpumask_shift_left(tmp, tmp, nthreads);
288 if (cpumask_empty(tmp)) {
289 printk(KERN_ERR "Unable to find space in cpu_present_mask for"
290 " processor %s with %d thread(s)\n", np->name,
295 for_each_cpu(cpu, tmp) {
296 BUG_ON(cpu_present(cpu));
297 set_cpu_present(cpu, true);
298 set_hard_smp_processor_id(cpu, be32_to_cpu(*intserv++));
302 cpu_maps_update_done();
303 free_cpumask_var(candidate_mask);
304 free_cpumask_var(tmp);
309 * Update the present map for a cpu node which is going away, and set
310 * the hard id in the paca(s) to -1 to be consistent with boot time
311 * convention for non-present cpus.
313 static void pseries_remove_processor(struct device_node *np)
316 int len, nthreads, i;
317 const __be32 *intserv;
320 intserv = of_get_property(np, "ibm,ppc-interrupt-server#s", &len);
324 nthreads = len / sizeof(u32);
326 cpu_maps_update_begin();
327 for (i = 0; i < nthreads; i++) {
328 thread = be32_to_cpu(intserv[i]);
329 for_each_present_cpu(cpu) {
330 if (get_hard_smp_processor_id(cpu) != thread)
332 BUG_ON(cpu_online(cpu));
333 set_cpu_present(cpu, false);
334 set_hard_smp_processor_id(cpu, -1);
335 update_numa_cpu_lookup_table(cpu, -1);
338 if (cpu >= nr_cpu_ids)
339 printk(KERN_WARNING "Could not find cpu to remove "
340 "with physical id 0x%x\n", thread);
342 cpu_maps_update_done();
345 static int dlpar_online_cpu(struct device_node *dn)
349 int len, nthreads, i;
350 const __be32 *intserv;
353 intserv = of_get_property(dn, "ibm,ppc-interrupt-server#s", &len);
357 nthreads = len / sizeof(u32);
359 cpu_maps_update_begin();
360 for (i = 0; i < nthreads; i++) {
361 thread = be32_to_cpu(intserv[i]);
362 for_each_present_cpu(cpu) {
363 if (get_hard_smp_processor_id(cpu) != thread)
365 BUG_ON(get_cpu_current_state(cpu)
366 != CPU_STATE_OFFLINE);
367 cpu_maps_update_done();
368 timed_topology_update(1);
369 find_and_online_cpu_nid(cpu);
370 rc = device_online(get_cpu_device(cpu));
373 cpu_maps_update_begin();
377 if (cpu == num_possible_cpus())
378 printk(KERN_WARNING "Could not find cpu to online "
379 "with physical id 0x%x\n", thread);
381 cpu_maps_update_done();
388 static bool dlpar_cpu_exists(struct device_node *parent, u32 drc_index)
390 struct device_node *child = NULL;
395 /* Assume cpu doesn't exist */
398 for_each_child_of_node(parent, child) {
399 rc = of_property_read_u32(child, "ibm,my-drc-index",
404 if (my_drc_index == drc_index) {
414 static bool valid_cpu_drc_index(struct device_node *parent, u32 drc_index)
423 rc = of_property_read_u32_index(parent, "ibm,drc-indexes",
428 if (drc == drc_index)
435 static ssize_t dlpar_cpu_add(u32 drc_index)
437 struct device_node *dn, *parent;
440 pr_debug("Attempting to add CPU, drc index: %x\n", drc_index);
442 parent = of_find_node_by_path("/cpus");
444 pr_warn("Failed to find CPU root node \"/cpus\"\n");
448 if (dlpar_cpu_exists(parent, drc_index)) {
450 pr_warn("CPU with drc index %x already exists\n", drc_index);
454 if (!valid_cpu_drc_index(parent, drc_index)) {
456 pr_warn("Cannot find CPU (drc index %x) to add.\n", drc_index);
460 rc = dlpar_acquire_drc(drc_index);
462 pr_warn("Failed to acquire DRC, rc: %d, drc index: %x\n",
468 dn = dlpar_configure_connector(cpu_to_be32(drc_index), parent);
470 pr_warn("Failed call to configure-connector, drc index: %x\n",
472 dlpar_release_drc(drc_index);
477 rc = dlpar_attach_node(dn, parent);
479 /* Regardless we are done with parent now */
484 pr_warn("Failed to attach node %s, rc: %d, drc index: %x\n",
485 dn->name, rc, drc_index);
487 rc = dlpar_release_drc(drc_index);
489 dlpar_free_cc_nodes(dn);
494 rc = dlpar_online_cpu(dn);
497 pr_warn("Failed to online cpu %s, rc: %d, drc index: %x\n",
498 dn->name, rc, drc_index);
500 rc = dlpar_detach_node(dn);
502 dlpar_release_drc(drc_index);
507 pr_debug("Successfully added CPU %s, drc index: %x\n", dn->name,
512 static int dlpar_offline_cpu(struct device_node *dn)
516 int len, nthreads, i;
517 const __be32 *intserv;
520 intserv = of_get_property(dn, "ibm,ppc-interrupt-server#s", &len);
524 nthreads = len / sizeof(u32);
526 cpu_maps_update_begin();
527 for (i = 0; i < nthreads; i++) {
528 thread = be32_to_cpu(intserv[i]);
529 for_each_present_cpu(cpu) {
530 if (get_hard_smp_processor_id(cpu) != thread)
533 if (get_cpu_current_state(cpu) == CPU_STATE_OFFLINE)
536 if (get_cpu_current_state(cpu) == CPU_STATE_ONLINE) {
537 set_preferred_offline_state(cpu,
539 cpu_maps_update_done();
540 timed_topology_update(1);
541 rc = device_offline(get_cpu_device(cpu));
544 cpu_maps_update_begin();
550 * The cpu is in CPU_STATE_INACTIVE.
551 * Upgrade it's state to CPU_STATE_OFFLINE.
553 set_preferred_offline_state(cpu, CPU_STATE_OFFLINE);
554 BUG_ON(plpar_hcall_norets(H_PROD, thread)
559 if (cpu == num_possible_cpus())
560 printk(KERN_WARNING "Could not find cpu to offline with physical id 0x%x\n", thread);
562 cpu_maps_update_done();
569 static ssize_t dlpar_cpu_remove(struct device_node *dn, u32 drc_index)
573 pr_debug("Attempting to remove CPU %s, drc index: %x\n",
574 dn->name, drc_index);
576 rc = dlpar_offline_cpu(dn);
578 pr_warn("Failed to offline CPU %s, rc: %d\n", dn->name, rc);
582 rc = dlpar_release_drc(drc_index);
584 pr_warn("Failed to release drc (%x) for CPU %s, rc: %d\n",
585 drc_index, dn->name, rc);
586 dlpar_online_cpu(dn);
590 rc = dlpar_detach_node(dn);
594 pr_warn("Failed to detach CPU %s, rc: %d", dn->name, rc);
596 rc = dlpar_acquire_drc(drc_index);
598 dlpar_online_cpu(dn);
603 pr_debug("Successfully removed CPU, drc index: %x\n", drc_index);
607 static struct device_node *cpu_drc_index_to_dn(u32 drc_index)
609 struct device_node *dn;
613 for_each_node_by_type(dn, "cpu") {
614 rc = of_property_read_u32(dn, "ibm,my-drc-index", &my_index);
618 if (my_index == drc_index)
625 static int dlpar_cpu_remove_by_index(u32 drc_index)
627 struct device_node *dn;
630 dn = cpu_drc_index_to_dn(drc_index);
632 pr_warn("Cannot find CPU (drc index %x) to remove\n",
637 rc = dlpar_cpu_remove(dn, drc_index);
642 static int find_dlpar_cpus_to_remove(u32 *cpu_drcs, int cpus_to_remove)
644 struct device_node *dn;
648 /* We want to find cpus_to_remove + 1 CPUs to ensure we do not
649 * remove the last CPU.
651 for_each_node_by_type(dn, "cpu") {
654 if (cpus_found > cpus_to_remove) {
659 /* Note that cpus_found is always 1 ahead of the index
660 * into the cpu_drcs array, so we use cpus_found - 1
662 rc = of_property_read_u32(dn, "ibm,my-drc-index",
663 &cpu_drcs[cpus_found - 1]);
665 pr_warn("Error occurred getting drc-index for %s\n",
672 if (cpus_found < cpus_to_remove) {
673 pr_warn("Failed to find enough CPUs (%d of %d) to remove\n",
674 cpus_found, cpus_to_remove);
675 } else if (cpus_found == cpus_to_remove) {
676 pr_warn("Cannot remove all CPUs\n");
682 static int dlpar_cpu_remove_by_count(u32 cpus_to_remove)
686 int cpus_removed = 0;
689 pr_debug("Attempting to hot-remove %d CPUs\n", cpus_to_remove);
691 cpu_drcs = kcalloc(cpus_to_remove, sizeof(*cpu_drcs), GFP_KERNEL);
695 cpus_found = find_dlpar_cpus_to_remove(cpu_drcs, cpus_to_remove);
696 if (cpus_found <= cpus_to_remove) {
701 for (i = 0; i < cpus_to_remove; i++) {
702 rc = dlpar_cpu_remove_by_index(cpu_drcs[i]);
709 if (cpus_removed != cpus_to_remove) {
710 pr_warn("CPU hot-remove failed, adding back removed CPUs\n");
712 for (i = 0; i < cpus_removed; i++)
713 dlpar_cpu_add(cpu_drcs[i]);
724 static int find_dlpar_cpus_to_add(u32 *cpu_drcs, u32 cpus_to_add)
726 struct device_node *parent;
730 parent = of_find_node_by_path("/cpus");
732 pr_warn("Could not find CPU root node in device tree\n");
737 /* Search the ibm,drc-indexes array for possible CPU drcs to
738 * add. Note that the format of the ibm,drc-indexes array is
739 * the number of entries in the array followed by the array
740 * of drc values so we start looking at index = 1.
743 while (cpus_found < cpus_to_add) {
746 rc = of_property_read_u32_index(parent, "ibm,drc-indexes",
751 if (dlpar_cpu_exists(parent, drc))
754 cpu_drcs[cpus_found++] = drc;
761 static int dlpar_cpu_add_by_count(u32 cpus_to_add)
768 pr_debug("Attempting to hot-add %d CPUs\n", cpus_to_add);
770 cpu_drcs = kcalloc(cpus_to_add, sizeof(*cpu_drcs), GFP_KERNEL);
774 cpus_found = find_dlpar_cpus_to_add(cpu_drcs, cpus_to_add);
775 if (cpus_found < cpus_to_add) {
776 pr_warn("Failed to find enough CPUs (%d of %d) to add\n",
777 cpus_found, cpus_to_add);
782 for (i = 0; i < cpus_to_add; i++) {
783 rc = dlpar_cpu_add(cpu_drcs[i]);
790 if (cpus_added < cpus_to_add) {
791 pr_warn("CPU hot-add failed, removing any added CPUs\n");
793 for (i = 0; i < cpus_added; i++)
794 dlpar_cpu_remove_by_index(cpu_drcs[i]);
805 int dlpar_cpu(struct pseries_hp_errorlog *hp_elog)
807 u32 count, drc_index;
810 count = hp_elog->_drc_u.drc_count;
811 drc_index = hp_elog->_drc_u.drc_index;
813 lock_device_hotplug();
815 switch (hp_elog->action) {
816 case PSERIES_HP_ELOG_ACTION_REMOVE:
817 if (hp_elog->id_type == PSERIES_HP_ELOG_ID_DRC_COUNT)
818 rc = dlpar_cpu_remove_by_count(count);
819 else if (hp_elog->id_type == PSERIES_HP_ELOG_ID_DRC_INDEX)
820 rc = dlpar_cpu_remove_by_index(drc_index);
824 case PSERIES_HP_ELOG_ACTION_ADD:
825 if (hp_elog->id_type == PSERIES_HP_ELOG_ID_DRC_COUNT)
826 rc = dlpar_cpu_add_by_count(count);
827 else if (hp_elog->id_type == PSERIES_HP_ELOG_ID_DRC_INDEX)
828 rc = dlpar_cpu_add(drc_index);
833 pr_err("Invalid action (%d) specified\n", hp_elog->action);
838 unlock_device_hotplug();
842 #ifdef CONFIG_ARCH_CPU_PROBE_RELEASE
844 static ssize_t dlpar_cpu_probe(const char *buf, size_t count)
849 rc = kstrtou32(buf, 0, &drc_index);
853 rc = dlpar_cpu_add(drc_index);
855 return rc ? rc : count;
858 static ssize_t dlpar_cpu_release(const char *buf, size_t count)
860 struct device_node *dn;
864 dn = of_find_node_by_path(buf);
868 rc = of_property_read_u32(dn, "ibm,my-drc-index", &drc_index);
874 rc = dlpar_cpu_remove(dn, drc_index);
877 return rc ? rc : count;
880 #endif /* CONFIG_ARCH_CPU_PROBE_RELEASE */
882 static int pseries_smp_notifier(struct notifier_block *nb,
883 unsigned long action, void *data)
885 struct of_reconfig_data *rd = data;
889 case OF_RECONFIG_ATTACH_NODE:
890 err = pseries_add_processor(rd->dn);
892 case OF_RECONFIG_DETACH_NODE:
893 pseries_remove_processor(rd->dn);
896 return notifier_from_errno(err);
899 static struct notifier_block pseries_smp_nb = {
900 .notifier_call = pseries_smp_notifier,
903 #define MAX_CEDE_LATENCY_LEVELS 4
904 #define CEDE_LATENCY_PARAM_LENGTH 10
905 #define CEDE_LATENCY_PARAM_MAX_LENGTH \
906 (MAX_CEDE_LATENCY_LEVELS * CEDE_LATENCY_PARAM_LENGTH * sizeof(char))
907 #define CEDE_LATENCY_TOKEN 45
909 static char cede_parameters[CEDE_LATENCY_PARAM_MAX_LENGTH];
911 static int parse_cede_parameters(void)
913 memset(cede_parameters, 0, CEDE_LATENCY_PARAM_MAX_LENGTH);
914 return rtas_call(rtas_token("ibm,get-system-parameter"), 3, 1,
917 __pa(cede_parameters),
918 CEDE_LATENCY_PARAM_MAX_LENGTH);
921 static int __init pseries_cpu_hotplug_init(void)
926 #ifdef CONFIG_ARCH_CPU_PROBE_RELEASE
927 ppc_md.cpu_probe = dlpar_cpu_probe;
928 ppc_md.cpu_release = dlpar_cpu_release;
929 #endif /* CONFIG_ARCH_CPU_PROBE_RELEASE */
931 rtas_stop_self_token = rtas_token("stop-self");
932 qcss_tok = rtas_token("query-cpu-stopped-state");
934 if (rtas_stop_self_token == RTAS_UNKNOWN_SERVICE ||
935 qcss_tok == RTAS_UNKNOWN_SERVICE) {
936 printk(KERN_INFO "CPU Hotplug not supported by firmware "
941 ppc_md.cpu_die = pseries_mach_cpu_die;
942 smp_ops->cpu_disable = pseries_cpu_disable;
943 smp_ops->cpu_die = pseries_cpu_die;
945 /* Processors can be added/removed only on LPAR */
946 if (firmware_has_feature(FW_FEATURE_LPAR)) {
947 of_reconfig_notifier_register(&pseries_smp_nb);
948 cpu_maps_update_begin();
949 if (cede_offline_enabled && parse_cede_parameters() == 0) {
950 default_offline_state = CPU_STATE_INACTIVE;
951 for_each_online_cpu(cpu)
952 set_default_offline_state(cpu);
954 cpu_maps_update_done();
959 machine_arch_initcall(pseries, pseries_cpu_hotplug_init);