1 // SPDX-License-Identifier: GPL-2.0-or-later
3 * pseries CPU Hotplug infrastructure.
5 * Split out from arch/powerpc/platforms/pseries/setup.c
6 * arch/powerpc/kernel/rtas.c, and arch/powerpc/platforms/pseries/smp.c
8 * Peter Bergner, IBM March 2001.
9 * Copyright (C) 2001 IBM.
10 * Dave Engebretsen, Peter Bergner, and
11 * Mike Corrigan {engebret|bergner|mikec}@us.ibm.com
12 * Plus various changes from other IBM teams...
14 * Copyright (C) 2006 Michael Ellerman, IBM Corporation
17 #define pr_fmt(fmt) "pseries-hotplug-cpu: " fmt
19 #include <linux/kernel.h>
20 #include <linux/interrupt.h>
21 #include <linux/delay.h>
22 #include <linux/sched.h> /* for idle_task_exit */
23 #include <linux/sched/hotplug.h>
24 #include <linux/cpu.h>
26 #include <linux/slab.h>
29 #include <asm/firmware.h>
30 #include <asm/machdep.h>
31 #include <asm/vdso_datapage.h>
34 #include <asm/plpar_wrappers.h>
35 #include <asm/topology.h>
38 #include "offline_states.h"
40 /* This version can't take the spinlock, because it never returns */
41 static int rtas_stop_self_token = RTAS_UNKNOWN_SERVICE;
43 static DEFINE_PER_CPU(enum cpu_state_vals, preferred_offline_state) =
45 static DEFINE_PER_CPU(enum cpu_state_vals, current_state) = CPU_STATE_OFFLINE;
47 static enum cpu_state_vals default_offline_state = CPU_STATE_OFFLINE;
49 static bool cede_offline_enabled __read_mostly = true;
52 * Enable/disable cede_offline when available.
54 static int __init setup_cede_offline(char *str)
56 return (kstrtobool(str, &cede_offline_enabled) == 0);
59 __setup("cede_offline=", setup_cede_offline);
61 enum cpu_state_vals get_cpu_current_state(int cpu)
63 return per_cpu(current_state, cpu);
66 void set_cpu_current_state(int cpu, enum cpu_state_vals state)
68 per_cpu(current_state, cpu) = state;
71 enum cpu_state_vals get_preferred_offline_state(int cpu)
73 return per_cpu(preferred_offline_state, cpu);
76 void set_preferred_offline_state(int cpu, enum cpu_state_vals state)
78 per_cpu(preferred_offline_state, cpu) = state;
81 void set_default_offline_state(int cpu)
83 per_cpu(preferred_offline_state, cpu) = default_offline_state;
86 static void rtas_stop_self(void)
88 static struct rtas_args args;
92 BUG_ON(rtas_stop_self_token == RTAS_UNKNOWN_SERVICE);
94 rtas_call_unlocked(&args, rtas_stop_self_token, 0, 1, NULL);
96 panic("Alas, I survived.\n");
99 static void pseries_mach_cpu_die(void)
101 unsigned int cpu = smp_processor_id();
102 unsigned int hwcpu = hard_smp_processor_id();
103 u8 cede_latency_hint = 0;
112 if (get_preferred_offline_state(cpu) == CPU_STATE_INACTIVE) {
113 set_cpu_current_state(cpu, CPU_STATE_INACTIVE);
114 if (ppc_md.suspend_disable_cpu)
115 ppc_md.suspend_disable_cpu();
117 cede_latency_hint = 2;
119 get_lppaca()->idle = 1;
120 if (!lppaca_shared_proc(get_lppaca()))
121 get_lppaca()->donate_dedicated_cpu = 1;
123 while (get_preferred_offline_state(cpu) == CPU_STATE_INACTIVE) {
124 while (!prep_irq_for_idle()) {
129 extended_cede_processor(cede_latency_hint);
134 if (!lppaca_shared_proc(get_lppaca()))
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);
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 */
176 xive_smp_disable_cpu();
178 xics_migrate_irqs_away();
183 * pseries_cpu_die: Wait for the cpu to die.
184 * @cpu: logical processor id of the CPU whose death we're awaiting.
186 * This function is called from the context of the thread which is performing
187 * the cpu-offline. Here we wait for long enough to allow the cpu in question
188 * to self-destroy so that the cpu-offline thread can send the CPU_DEAD
191 * OTOH, pseries_mach_cpu_die() is called by the @cpu when it wants to
194 static void pseries_cpu_die(unsigned int cpu)
198 unsigned int pcpu = get_hard_smp_processor_id(cpu);
200 if (get_preferred_offline_state(cpu) == CPU_STATE_INACTIVE) {
202 for (tries = 0; tries < 5000; tries++) {
203 if (get_cpu_current_state(cpu) == CPU_STATE_INACTIVE) {
209 } else if (get_preferred_offline_state(cpu) == CPU_STATE_OFFLINE) {
211 for (tries = 0; tries < 25; tries++) {
212 cpu_status = smp_query_cpu_stopped(pcpu);
213 if (cpu_status == QCSS_STOPPED ||
214 cpu_status == QCSS_HARDWARE_ERROR)
220 if (cpu_status != 0) {
221 printk("Querying DEAD? cpu %i (%i) shows %i\n",
222 cpu, pcpu, cpu_status);
225 /* Isolation and deallocation are definitely done by
226 * drslot_chrp_cpu. If they were not they would be
227 * done here. Change isolate state to Isolate and
228 * change allocation-state to Unusable.
230 paca_ptrs[cpu]->cpu_start = 0;
234 * Update cpu_present_mask and paca(s) for a new cpu node. The wrinkle
235 * here is that a cpu device node may represent up to two logical cpus
236 * in the SMT case. We must honor the assumption in other code that
237 * the logical ids for sibling SMT threads x and y are adjacent, such
238 * that x^1 == y and y^1 == x.
240 static int pseries_add_processor(struct device_node *np)
243 cpumask_var_t candidate_mask, tmp;
244 int err = -ENOSPC, len, nthreads, i;
245 const __be32 *intserv;
247 intserv = of_get_property(np, "ibm,ppc-interrupt-server#s", &len);
251 zalloc_cpumask_var(&candidate_mask, GFP_KERNEL);
252 zalloc_cpumask_var(&tmp, GFP_KERNEL);
254 nthreads = len / sizeof(u32);
255 for (i = 0; i < nthreads; i++)
256 cpumask_set_cpu(i, tmp);
258 cpu_maps_update_begin();
260 BUG_ON(!cpumask_subset(cpu_present_mask, cpu_possible_mask));
262 /* Get a bitmap of unoccupied slots. */
263 cpumask_xor(candidate_mask, cpu_possible_mask, cpu_present_mask);
264 if (cpumask_empty(candidate_mask)) {
265 /* If we get here, it most likely means that NR_CPUS is
266 * less than the partition's max processors setting.
268 printk(KERN_ERR "Cannot add cpu %pOF; this system configuration"
269 " supports %d logical cpus.\n", np,
270 num_possible_cpus());
274 while (!cpumask_empty(tmp))
275 if (cpumask_subset(tmp, candidate_mask))
276 /* Found a range where we can insert the new cpu(s) */
279 cpumask_shift_left(tmp, tmp, nthreads);
281 if (cpumask_empty(tmp)) {
282 printk(KERN_ERR "Unable to find space in cpu_present_mask for"
283 " processor %pOFn with %d thread(s)\n", np,
288 for_each_cpu(cpu, tmp) {
289 BUG_ON(cpu_present(cpu));
290 set_cpu_present(cpu, true);
291 set_hard_smp_processor_id(cpu, be32_to_cpu(*intserv++));
295 cpu_maps_update_done();
296 free_cpumask_var(candidate_mask);
297 free_cpumask_var(tmp);
302 * Update the present map for a cpu node which is going away, and set
303 * the hard id in the paca(s) to -1 to be consistent with boot time
304 * convention for non-present cpus.
306 static void pseries_remove_processor(struct device_node *np)
309 int len, nthreads, i;
310 const __be32 *intserv;
313 intserv = of_get_property(np, "ibm,ppc-interrupt-server#s", &len);
317 nthreads = len / sizeof(u32);
319 cpu_maps_update_begin();
320 for (i = 0; i < nthreads; i++) {
321 thread = be32_to_cpu(intserv[i]);
322 for_each_present_cpu(cpu) {
323 if (get_hard_smp_processor_id(cpu) != thread)
325 BUG_ON(cpu_online(cpu));
326 set_cpu_present(cpu, false);
327 set_hard_smp_processor_id(cpu, -1);
328 update_numa_cpu_lookup_table(cpu, -1);
331 if (cpu >= nr_cpu_ids)
332 printk(KERN_WARNING "Could not find cpu to remove "
333 "with physical id 0x%x\n", thread);
335 cpu_maps_update_done();
338 static int dlpar_online_cpu(struct device_node *dn)
342 int len, nthreads, i;
343 const __be32 *intserv;
346 intserv = of_get_property(dn, "ibm,ppc-interrupt-server#s", &len);
350 nthreads = len / sizeof(u32);
352 cpu_maps_update_begin();
353 for (i = 0; i < nthreads; i++) {
354 thread = be32_to_cpu(intserv[i]);
355 for_each_present_cpu(cpu) {
356 if (get_hard_smp_processor_id(cpu) != thread)
358 BUG_ON(get_cpu_current_state(cpu)
359 != CPU_STATE_OFFLINE);
360 cpu_maps_update_done();
361 timed_topology_update(1);
362 find_and_online_cpu_nid(cpu);
363 rc = device_online(get_cpu_device(cpu));
366 cpu_maps_update_begin();
370 if (cpu == num_possible_cpus())
371 printk(KERN_WARNING "Could not find cpu to online "
372 "with physical id 0x%x\n", thread);
374 cpu_maps_update_done();
381 static bool dlpar_cpu_exists(struct device_node *parent, u32 drc_index)
383 struct device_node *child = NULL;
388 /* Assume cpu doesn't exist */
391 for_each_child_of_node(parent, child) {
392 rc = of_property_read_u32(child, "ibm,my-drc-index",
397 if (my_drc_index == drc_index) {
407 static bool valid_cpu_drc_index(struct device_node *parent, u32 drc_index)
416 rc = of_property_read_u32_index(parent, "ibm,drc-indexes",
421 if (drc == drc_index)
428 static ssize_t dlpar_cpu_add(u32 drc_index)
430 struct device_node *dn, *parent;
433 pr_debug("Attempting to add CPU, drc index: %x\n", drc_index);
435 parent = of_find_node_by_path("/cpus");
437 pr_warn("Failed to find CPU root node \"/cpus\"\n");
441 if (dlpar_cpu_exists(parent, drc_index)) {
443 pr_warn("CPU with drc index %x already exists\n", drc_index);
447 if (!valid_cpu_drc_index(parent, drc_index)) {
449 pr_warn("Cannot find CPU (drc index %x) to add.\n", drc_index);
453 rc = dlpar_acquire_drc(drc_index);
455 pr_warn("Failed to acquire DRC, rc: %d, drc index: %x\n",
461 dn = dlpar_configure_connector(cpu_to_be32(drc_index), parent);
463 pr_warn("Failed call to configure-connector, drc index: %x\n",
465 dlpar_release_drc(drc_index);
470 rc = dlpar_attach_node(dn, parent);
472 /* Regardless we are done with parent now */
477 pr_warn("Failed to attach node %pOFn, rc: %d, drc index: %x\n",
480 rc = dlpar_release_drc(drc_index);
482 dlpar_free_cc_nodes(dn);
487 rc = dlpar_online_cpu(dn);
490 pr_warn("Failed to online cpu %pOFn, rc: %d, drc index: %x\n",
493 rc = dlpar_detach_node(dn);
495 dlpar_release_drc(drc_index);
500 pr_debug("Successfully added CPU %pOFn, drc index: %x\n", dn,
505 static int dlpar_offline_cpu(struct device_node *dn)
509 int len, nthreads, i;
510 const __be32 *intserv;
513 intserv = of_get_property(dn, "ibm,ppc-interrupt-server#s", &len);
517 nthreads = len / sizeof(u32);
519 cpu_maps_update_begin();
520 for (i = 0; i < nthreads; i++) {
521 thread = be32_to_cpu(intserv[i]);
522 for_each_present_cpu(cpu) {
523 if (get_hard_smp_processor_id(cpu) != thread)
526 if (get_cpu_current_state(cpu) == CPU_STATE_OFFLINE)
529 if (get_cpu_current_state(cpu) == CPU_STATE_ONLINE) {
530 set_preferred_offline_state(cpu,
532 cpu_maps_update_done();
533 timed_topology_update(1);
534 rc = device_offline(get_cpu_device(cpu));
537 cpu_maps_update_begin();
543 * The cpu is in CPU_STATE_INACTIVE.
544 * Upgrade it's state to CPU_STATE_OFFLINE.
546 set_preferred_offline_state(cpu, CPU_STATE_OFFLINE);
547 BUG_ON(plpar_hcall_norets(H_PROD, thread)
552 if (cpu == num_possible_cpus())
553 printk(KERN_WARNING "Could not find cpu to offline with physical id 0x%x\n", thread);
555 cpu_maps_update_done();
562 static ssize_t dlpar_cpu_remove(struct device_node *dn, u32 drc_index)
566 pr_debug("Attempting to remove CPU %pOFn, drc index: %x\n",
569 rc = dlpar_offline_cpu(dn);
571 pr_warn("Failed to offline CPU %pOFn, rc: %d\n", dn, rc);
575 rc = dlpar_release_drc(drc_index);
577 pr_warn("Failed to release drc (%x) for CPU %pOFn, rc: %d\n",
579 dlpar_online_cpu(dn);
583 rc = dlpar_detach_node(dn);
587 pr_warn("Failed to detach CPU %pOFn, rc: %d", dn, rc);
589 rc = dlpar_acquire_drc(drc_index);
591 dlpar_online_cpu(dn);
596 pr_debug("Successfully removed CPU, drc index: %x\n", drc_index);
600 static struct device_node *cpu_drc_index_to_dn(u32 drc_index)
602 struct device_node *dn;
606 for_each_node_by_type(dn, "cpu") {
607 rc = of_property_read_u32(dn, "ibm,my-drc-index", &my_index);
611 if (my_index == drc_index)
618 static int dlpar_cpu_remove_by_index(u32 drc_index)
620 struct device_node *dn;
623 dn = cpu_drc_index_to_dn(drc_index);
625 pr_warn("Cannot find CPU (drc index %x) to remove\n",
630 rc = dlpar_cpu_remove(dn, drc_index);
635 static int find_dlpar_cpus_to_remove(u32 *cpu_drcs, int cpus_to_remove)
637 struct device_node *dn;
641 /* We want to find cpus_to_remove + 1 CPUs to ensure we do not
642 * remove the last CPU.
644 for_each_node_by_type(dn, "cpu") {
647 if (cpus_found > cpus_to_remove) {
652 /* Note that cpus_found is always 1 ahead of the index
653 * into the cpu_drcs array, so we use cpus_found - 1
655 rc = of_property_read_u32(dn, "ibm,my-drc-index",
656 &cpu_drcs[cpus_found - 1]);
658 pr_warn("Error occurred getting drc-index for %pOFn\n",
665 if (cpus_found < cpus_to_remove) {
666 pr_warn("Failed to find enough CPUs (%d of %d) to remove\n",
667 cpus_found, cpus_to_remove);
668 } else if (cpus_found == cpus_to_remove) {
669 pr_warn("Cannot remove all CPUs\n");
675 static int dlpar_cpu_remove_by_count(u32 cpus_to_remove)
679 int cpus_removed = 0;
682 pr_debug("Attempting to hot-remove %d CPUs\n", cpus_to_remove);
684 cpu_drcs = kcalloc(cpus_to_remove, sizeof(*cpu_drcs), GFP_KERNEL);
688 cpus_found = find_dlpar_cpus_to_remove(cpu_drcs, cpus_to_remove);
689 if (cpus_found <= cpus_to_remove) {
694 for (i = 0; i < cpus_to_remove; i++) {
695 rc = dlpar_cpu_remove_by_index(cpu_drcs[i]);
702 if (cpus_removed != cpus_to_remove) {
703 pr_warn("CPU hot-remove failed, adding back removed CPUs\n");
705 for (i = 0; i < cpus_removed; i++)
706 dlpar_cpu_add(cpu_drcs[i]);
717 static int find_dlpar_cpus_to_add(u32 *cpu_drcs, u32 cpus_to_add)
719 struct device_node *parent;
723 parent = of_find_node_by_path("/cpus");
725 pr_warn("Could not find CPU root node in device tree\n");
730 /* Search the ibm,drc-indexes array for possible CPU drcs to
731 * add. Note that the format of the ibm,drc-indexes array is
732 * the number of entries in the array followed by the array
733 * of drc values so we start looking at index = 1.
736 while (cpus_found < cpus_to_add) {
739 rc = of_property_read_u32_index(parent, "ibm,drc-indexes",
744 if (dlpar_cpu_exists(parent, drc))
747 cpu_drcs[cpus_found++] = drc;
754 static int dlpar_cpu_add_by_count(u32 cpus_to_add)
761 pr_debug("Attempting to hot-add %d CPUs\n", cpus_to_add);
763 cpu_drcs = kcalloc(cpus_to_add, sizeof(*cpu_drcs), GFP_KERNEL);
767 cpus_found = find_dlpar_cpus_to_add(cpu_drcs, cpus_to_add);
768 if (cpus_found < cpus_to_add) {
769 pr_warn("Failed to find enough CPUs (%d of %d) to add\n",
770 cpus_found, cpus_to_add);
775 for (i = 0; i < cpus_to_add; i++) {
776 rc = dlpar_cpu_add(cpu_drcs[i]);
783 if (cpus_added < cpus_to_add) {
784 pr_warn("CPU hot-add failed, removing any added CPUs\n");
786 for (i = 0; i < cpus_added; i++)
787 dlpar_cpu_remove_by_index(cpu_drcs[i]);
798 int dlpar_cpu_readd(int cpu)
800 struct device_node *dn;
805 dev = get_cpu_device(cpu);
808 rc = of_property_read_u32(dn, "ibm,my-drc-index", &drc_index);
810 rc = dlpar_cpu_remove_by_index(drc_index);
812 rc = dlpar_cpu_add(drc_index);
817 int dlpar_cpu(struct pseries_hp_errorlog *hp_elog)
819 u32 count, drc_index;
822 count = hp_elog->_drc_u.drc_count;
823 drc_index = hp_elog->_drc_u.drc_index;
825 lock_device_hotplug();
827 switch (hp_elog->action) {
828 case PSERIES_HP_ELOG_ACTION_REMOVE:
829 if (hp_elog->id_type == PSERIES_HP_ELOG_ID_DRC_COUNT)
830 rc = dlpar_cpu_remove_by_count(count);
831 else if (hp_elog->id_type == PSERIES_HP_ELOG_ID_DRC_INDEX)
832 rc = dlpar_cpu_remove_by_index(drc_index);
836 case PSERIES_HP_ELOG_ACTION_ADD:
837 if (hp_elog->id_type == PSERIES_HP_ELOG_ID_DRC_COUNT)
838 rc = dlpar_cpu_add_by_count(count);
839 else if (hp_elog->id_type == PSERIES_HP_ELOG_ID_DRC_INDEX)
840 rc = dlpar_cpu_add(drc_index);
845 pr_err("Invalid action (%d) specified\n", hp_elog->action);
850 unlock_device_hotplug();
854 #ifdef CONFIG_ARCH_CPU_PROBE_RELEASE
856 static ssize_t dlpar_cpu_probe(const char *buf, size_t count)
861 rc = kstrtou32(buf, 0, &drc_index);
865 rc = dlpar_cpu_add(drc_index);
867 return rc ? rc : count;
870 static ssize_t dlpar_cpu_release(const char *buf, size_t count)
872 struct device_node *dn;
876 dn = of_find_node_by_path(buf);
880 rc = of_property_read_u32(dn, "ibm,my-drc-index", &drc_index);
886 rc = dlpar_cpu_remove(dn, drc_index);
889 return rc ? rc : count;
892 #endif /* CONFIG_ARCH_CPU_PROBE_RELEASE */
894 static int pseries_smp_notifier(struct notifier_block *nb,
895 unsigned long action, void *data)
897 struct of_reconfig_data *rd = data;
901 case OF_RECONFIG_ATTACH_NODE:
902 err = pseries_add_processor(rd->dn);
904 case OF_RECONFIG_DETACH_NODE:
905 pseries_remove_processor(rd->dn);
908 return notifier_from_errno(err);
911 static struct notifier_block pseries_smp_nb = {
912 .notifier_call = pseries_smp_notifier,
915 #define MAX_CEDE_LATENCY_LEVELS 4
916 #define CEDE_LATENCY_PARAM_LENGTH 10
917 #define CEDE_LATENCY_PARAM_MAX_LENGTH \
918 (MAX_CEDE_LATENCY_LEVELS * CEDE_LATENCY_PARAM_LENGTH * sizeof(char))
919 #define CEDE_LATENCY_TOKEN 45
921 static char cede_parameters[CEDE_LATENCY_PARAM_MAX_LENGTH];
923 static int parse_cede_parameters(void)
925 memset(cede_parameters, 0, CEDE_LATENCY_PARAM_MAX_LENGTH);
926 return rtas_call(rtas_token("ibm,get-system-parameter"), 3, 1,
929 __pa(cede_parameters),
930 CEDE_LATENCY_PARAM_MAX_LENGTH);
933 static int __init pseries_cpu_hotplug_init(void)
938 #ifdef CONFIG_ARCH_CPU_PROBE_RELEASE
939 ppc_md.cpu_probe = dlpar_cpu_probe;
940 ppc_md.cpu_release = dlpar_cpu_release;
941 #endif /* CONFIG_ARCH_CPU_PROBE_RELEASE */
943 rtas_stop_self_token = rtas_token("stop-self");
944 qcss_tok = rtas_token("query-cpu-stopped-state");
946 if (rtas_stop_self_token == RTAS_UNKNOWN_SERVICE ||
947 qcss_tok == RTAS_UNKNOWN_SERVICE) {
948 printk(KERN_INFO "CPU Hotplug not supported by firmware "
953 ppc_md.cpu_die = pseries_mach_cpu_die;
954 smp_ops->cpu_disable = pseries_cpu_disable;
955 smp_ops->cpu_die = pseries_cpu_die;
957 /* Processors can be added/removed only on LPAR */
958 if (firmware_has_feature(FW_FEATURE_LPAR)) {
959 of_reconfig_notifier_register(&pseries_smp_nb);
960 cpu_maps_update_begin();
961 if (cede_offline_enabled && parse_cede_parameters() == 0) {
962 default_offline_state = CPU_STATE_INACTIVE;
963 for_each_online_cpu(cpu)
964 set_default_offline_state(cpu);
966 cpu_maps_update_done();
971 machine_arch_initcall(pseries, pseries_cpu_hotplug_init);