arch/powerpc/platforms/pseries/hotplug-cpu.c

   1 /*
   2  * pseries CPU Hotplug infrastructure.
   3  *
   4  * Split out from arch/powerpc/platforms/pseries/setup.c
   5  *  arch/powerpc/kernel/rtas.c, and arch/powerpc/platforms/pseries/smp.c
   6  *
   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...
  12  *
  13  * Copyright (C) 2006 Michael Ellerman, IBM Corporation
  14  *
  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.
  19  */
  20
  21 #include <linux/kernel.h>
  22 #include <linux/delay.h>
  23 #include <linux/cpu.h>
  24 #include <asm/system.h>
  25 #include <asm/prom.h>
  26 #include <asm/rtas.h>
  27 #include <asm/firmware.h>
  28 #include <asm/machdep.h>
  29 #include <asm/vdso_datapage.h>
  30 #include <asm/pSeries_reconfig.h>
  31 #include "xics.h"
  32 #include "plpar_wrappers.h"
  33 #include "offline_states.h"
  34
  35 /* This version can't take the spinlock, because it never returns */
  36 static struct rtas_args rtas_stop_self_args = {
  37         .token = RTAS_UNKNOWN_SERVICE,
  38         .nargs = 0,
  39         .nret = 1,
  40         .rets = &rtas_stop_self_args.args[0],
  41 };
  42
  43 static DEFINE_PER_CPU(enum cpu_state_vals, preferred_offline_state) =
  44                                                         CPU_STATE_OFFLINE;
  45 static DEFINE_PER_CPU(enum cpu_state_vals, current_state) = CPU_STATE_OFFLINE;
  46
  47 static enum cpu_state_vals default_offline_state = CPU_STATE_OFFLINE;
  48
  49 static int cede_offline_enabled __read_mostly = 1;
  50
  51 /*
  52  * Enable/disable cede_offline when available.
  53  */
  54 static int __init setup_cede_offline(char *str)
  55 {
  56         if (!strcmp(str, "off"))
  57                 cede_offline_enabled = 0;
  58         else if (!strcmp(str, "on"))
  59                 cede_offline_enabled = 1;
  60         else
  61                 return 0;
  62         return 1;
  63 }
  64
  65 __setup("cede_offline=", setup_cede_offline);
  66
  67 enum cpu_state_vals get_cpu_current_state(int cpu)
  68 {
  69         return per_cpu(current_state, cpu);
  70 }
  71
  72 void set_cpu_current_state(int cpu, enum cpu_state_vals state)
  73 {
  74         per_cpu(current_state, cpu) = state;
  75 }
  76
  77 enum cpu_state_vals get_preferred_offline_state(int cpu)
  78 {
  79         return per_cpu(preferred_offline_state, cpu);
  80 }
  81
  82 void set_preferred_offline_state(int cpu, enum cpu_state_vals state)
  83 {
  84         per_cpu(preferred_offline_state, cpu) = state;
  85 }
  86
  87 void set_default_offline_state(int cpu)
  88 {
  89         per_cpu(preferred_offline_state, cpu) = default_offline_state;
  90 }
  91
  92 static void rtas_stop_self(void)
  93 {
  94         struct rtas_args *args = &rtas_stop_self_args;
  95
  96         local_irq_disable();
  97
  98         BUG_ON(args->token == RTAS_UNKNOWN_SERVICE);
  99
 100         printk("cpu %u (hwid %u) Ready to die...\n",
 101                smp_processor_id(), hard_smp_processor_id());
 102         enter_rtas(__pa(args));
 103
 104         panic("Alas, I survived.\n");
 105 }
 106
 107 static void pseries_mach_cpu_die(void)
 108 {
 109         unsigned int cpu = smp_processor_id();
 110         unsigned int hwcpu = hard_smp_processor_id();
 111         u8 cede_latency_hint = 0;
 112
 113         local_irq_disable();
 114         idle_task_exit();
 115         xics_teardown_cpu();
 116
 117         if (get_preferred_offline_state(cpu) == CPU_STATE_INACTIVE) {
 118                 set_cpu_current_state(cpu, CPU_STATE_INACTIVE);
 119                 cede_latency_hint = 2;
 120
 121                 get_lppaca()->idle = 1;
 122                 if (!get_lppaca()->shared_proc)
 123                         get_lppaca()->donate_dedicated_cpu = 1;
 124
 125                 printk(KERN_INFO
 126                         "cpu %u (hwid %u) ceding for offline with hint %d\n",
 127                         cpu, hwcpu, cede_latency_hint);
 128                 while (get_preferred_offline_state(cpu) == CPU_STATE_INACTIVE) {
 129                         extended_cede_processor(cede_latency_hint);
 130                         printk(KERN_INFO "cpu %u (hwid %u) returned from cede.\n",
 131                                 cpu, hwcpu);
 132                         printk(KERN_INFO
 133                         "Decrementer value = %x Timebase value = %llx\n",
 134                         get_dec(), get_tb());
 135                 }
 136
 137                 printk(KERN_INFO "cpu %u (hwid %u) got prodded to go online\n",
 138                         cpu, hwcpu);
 139
 140                 if (!get_lppaca()->shared_proc)
 141                         get_lppaca()->donate_dedicated_cpu = 0;
 142                 get_lppaca()->idle = 0;
 143         }
 144
 145         if (get_preferred_offline_state(cpu) == CPU_STATE_ONLINE) {
 146                 unregister_slb_shadow(hwcpu, __pa(get_slb_shadow()));
 147
 148                 /*
 149                  * NOTE: Calling start_secondary() here for now to
 150                  * start new context.
 151                  * However, need to do it cleanly by resetting the
 152                  * stack pointer.
 153                  */
 154                 start_secondary();
 155
 156         } else if (get_preferred_offline_state(cpu) == CPU_STATE_OFFLINE) {
 157
 158                 set_cpu_current_state(cpu, CPU_STATE_OFFLINE);
 159                 unregister_slb_shadow(hard_smp_processor_id(),
 160                                         __pa(get_slb_shadow()));
 161                 rtas_stop_self();
 162         }
 163
 164         /* Should never get here... */
 165         BUG();
 166         for(;;);
 167 }
 168
 169 static int qcss_tok;    /* query-cpu-stopped-state token */
 170
 171 /* Get state of physical CPU.
 172  * Return codes:
 173  *      0       - The processor is in the RTAS stopped state
 174  *      1       - stop-self is in progress
 175  *      2       - The processor is not in the RTAS stopped state
 176  *      -1      - Hardware Error
 177  *      -2      - Hardware Busy, Try again later.
 178  */
 179 static int query_cpu_stopped(unsigned int pcpu)
 180 {
 181         int cpu_status, status;
 182
 183         status = rtas_call(qcss_tok, 1, 2, &cpu_status, pcpu);
 184         if (status != 0) {
 185                 printk(KERN_ERR
 186                        "RTAS query-cpu-stopped-state failed: %i\n", status);
 187                 return status;
 188         }
 189
 190         return cpu_status;
 191 }
 192
 193 static int pseries_cpu_disable(void)
 194 {
 195         int cpu = smp_processor_id();
 196
 197         set_cpu_online(cpu, false);
 198         vdso_data->processorCount--;
 199
 200         /*fix boot_cpuid here*/
 201         if (cpu == boot_cpuid)
 202                 boot_cpuid = any_online_cpu(cpu_online_map);
 203
 204         /* FIXME: abstract this to not be platform specific later on */
 205         xics_migrate_irqs_away();
 206         return 0;
 207 }
 208
 209 /*
 210  * pseries_cpu_die: Wait for the cpu to die.
 211  * @cpu: logical processor id of the CPU whose death we're awaiting.
 212  *
 213  * This function is called from the context of the thread which is performing
 214  * the cpu-offline. Here we wait for long enough to allow the cpu in question
 215  * to self-destroy so that the cpu-offline thread can send the CPU_DEAD
 216  * notifications.
 217  *
 218  * OTOH, pseries_mach_cpu_die() is called by the @cpu when it wants to
 219  * self-destruct.
 220  */
 221 static void pseries_cpu_die(unsigned int cpu)
 222 {
 223         int tries;
 224         int cpu_status = 1;
 225         unsigned int pcpu = get_hard_smp_processor_id(cpu);
 226
 227         if (get_preferred_offline_state(cpu) == CPU_STATE_INACTIVE) {
 228                 cpu_status = 1;
 229                 for (tries = 0; tries < 1000; tries++) {
 230                         if (get_cpu_current_state(cpu) == CPU_STATE_INACTIVE) {
 231                                 cpu_status = 0;
 232                                 break;
 233                         }
 234                         cpu_relax();
 235                 }
 236         } else if (get_preferred_offline_state(cpu) == CPU_STATE_OFFLINE) {
 237
 238                 for (tries = 0; tries < 25; tries++) {
 239                         cpu_status = query_cpu_stopped(pcpu);
 240                         if (cpu_status == 0 || cpu_status == -1)
 241                                 break;
 242                         cpu_relax();
 243                 }
 244         }
 245
 246         if (cpu_status != 0) {
 247                 printk("Querying DEAD? cpu %i (%i) shows %i\n",
 248                        cpu, pcpu, cpu_status);
 249         }
 250
 251         /* Isolation and deallocation are definatly done by
 252          * drslot_chrp_cpu.  If they were not they would be
 253          * done here.  Change isolate state to Isolate and
 254          * change allocation-state to Unusable.
 255          */
 256         paca[cpu].cpu_start = 0;
 257 }
 258
 259 /*
 260  * Update cpu_present_map and paca(s) for a new cpu node.  The wrinkle
 261  * here is that a cpu device node may represent up to two logical cpus
 262  * in the SMT case.  We must honor the assumption in other code that
 263  * the logical ids for sibling SMT threads x and y are adjacent, such
 264  * that x^1 == y and y^1 == x.
 265  */
 266 static int pseries_add_processor(struct device_node *np)
 267 {
 268         unsigned int cpu;
 269         cpumask_t candidate_map, tmp = CPU_MASK_NONE;
 270         int err = -ENOSPC, len, nthreads, i;
 271         const u32 *intserv;
 272
 273         intserv = of_get_property(np, "ibm,ppc-interrupt-server#s", &len);
 274         if (!intserv)
 275                 return 0;
 276
 277         nthreads = len / sizeof(u32);
 278         for (i = 0; i < nthreads; i++)
 279                 cpu_set(i, tmp);
 280
 281         cpu_maps_update_begin();
 282
 283         BUG_ON(!cpus_subset(cpu_present_map, cpu_possible_map));
 284
 285         /* Get a bitmap of unoccupied slots. */
 286         cpus_xor(candidate_map, cpu_possible_map, cpu_present_map);
 287         if (cpus_empty(candidate_map)) {
 288                 /* If we get here, it most likely means that NR_CPUS is
 289                  * less than the partition's max processors setting.
 290                  */
 291                 printk(KERN_ERR "Cannot add cpu %s; this system configuration"
 292                        " supports %d logical cpus.\n", np->full_name,
 293                        cpus_weight(cpu_possible_map));
 294                 goto out_unlock;
 295         }
 296
 297         while (!cpus_empty(tmp))
 298                 if (cpus_subset(tmp, candidate_map))
 299                         /* Found a range where we can insert the new cpu(s) */
 300                         break;
 301                 else
 302                         cpus_shift_left(tmp, tmp, nthreads);
 303
 304         if (cpus_empty(tmp)) {
 305                 printk(KERN_ERR "Unable to find space in cpu_present_map for"
 306                        " processor %s with %d thread(s)\n", np->name,
 307                        nthreads);
 308                 goto out_unlock;
 309         }
 310
 311         for_each_cpu_mask(cpu, tmp) {
 312                 BUG_ON(cpu_isset(cpu, cpu_present_map));
 313                 set_cpu_present(cpu, true);
 314                 set_hard_smp_processor_id(cpu, *intserv++);
 315         }
 316         err = 0;
 317 out_unlock:
 318         cpu_maps_update_done();
 319         return err;
 320 }
 321
 322 /*
 323  * Update the present map for a cpu node which is going away, and set
 324  * the hard id in the paca(s) to -1 to be consistent with boot time
 325  * convention for non-present cpus.
 326  */
 327 static void pseries_remove_processor(struct device_node *np)
 328 {
 329         unsigned int cpu;
 330         int len, nthreads, i;
 331         const u32 *intserv;
 332
 333         intserv = of_get_property(np, "ibm,ppc-interrupt-server#s", &len);
 334         if (!intserv)
 335                 return;
 336
 337         nthreads = len / sizeof(u32);
 338
 339         cpu_maps_update_begin();
 340         for (i = 0; i < nthreads; i++) {
 341                 for_each_present_cpu(cpu) {
 342                         if (get_hard_smp_processor_id(cpu) != intserv[i])
 343                                 continue;
 344                         BUG_ON(cpu_online(cpu));
 345                         set_cpu_present(cpu, false);
 346                         set_hard_smp_processor_id(cpu, -1);
 347                         break;
 348                 }
 349                 if (cpu == NR_CPUS)
 350                         printk(KERN_WARNING "Could not find cpu to remove "
 351                                "with physical id 0x%x\n", intserv[i]);
 352         }
 353         cpu_maps_update_done();
 354 }
 355
 356 static int pseries_smp_notifier(struct notifier_block *nb,
 357                                 unsigned long action, void *node)
 358 {
 359         int err = NOTIFY_OK;
 360
 361         switch (action) {
 362         case PSERIES_RECONFIG_ADD:
 363                 if (pseries_add_processor(node))
 364                         err = NOTIFY_BAD;
 365                 break;
 366         case PSERIES_RECONFIG_REMOVE:
 367                 pseries_remove_processor(node);
 368                 break;
 369         default:
 370                 err = NOTIFY_DONE;
 371                 break;
 372         }
 373         return err;
 374 }
 375
 376 static struct notifier_block pseries_smp_nb = {
 377         .notifier_call = pseries_smp_notifier,
 378 };
 379
 380 #define MAX_CEDE_LATENCY_LEVELS         4
 381 #define CEDE_LATENCY_PARAM_LENGTH       10
 382 #define CEDE_LATENCY_PARAM_MAX_LENGTH   \
 383         (MAX_CEDE_LATENCY_LEVELS * CEDE_LATENCY_PARAM_LENGTH * sizeof(char))
 384 #define CEDE_LATENCY_TOKEN              45
 385
 386 static char cede_parameters[CEDE_LATENCY_PARAM_MAX_LENGTH];
 387
 388 static int parse_cede_parameters(void)
 389 {
 390         memset(cede_parameters, 0, CEDE_LATENCY_PARAM_MAX_LENGTH);
 391         return rtas_call(rtas_token("ibm,get-system-parameter"), 3, 1,
 392                          NULL,
 393                          CEDE_LATENCY_TOKEN,
 394                          __pa(cede_parameters),
 395                          CEDE_LATENCY_PARAM_MAX_LENGTH);
 396 }
 397
 398 static int __init pseries_cpu_hotplug_init(void)
 399 {
 400         struct device_node *np;
 401         const char *typep;
 402         int cpu;
 403
 404         for_each_node_by_name(np, "interrupt-controller") {
 405                 typep = of_get_property(np, "compatible", NULL);
 406                 if (strstr(typep, "open-pic")) {
 407                         of_node_put(np);
 408
 409                         printk(KERN_INFO "CPU Hotplug not supported on "
 410                                 "systems using MPIC\n");
 411                         return 0;
 412                 }
 413         }
 414
 415         rtas_stop_self_args.token = rtas_token("stop-self");
 416         qcss_tok = rtas_token("query-cpu-stopped-state");
 417
 418         if (rtas_stop_self_args.token == RTAS_UNKNOWN_SERVICE ||
 419                         qcss_tok == RTAS_UNKNOWN_SERVICE) {
 420                 printk(KERN_INFO "CPU Hotplug not supported by firmware "
 421                                 "- disabling.\n");
 422                 return 0;
 423         }
 424
 425         ppc_md.cpu_die = pseries_mach_cpu_die;
 426         smp_ops->cpu_disable = pseries_cpu_disable;
 427         smp_ops->cpu_die = pseries_cpu_die;
 428
 429         /* Processors can be added/removed only on LPAR */
 430         if (firmware_has_feature(FW_FEATURE_LPAR)) {
 431                 pSeries_reconfig_notifier_register(&pseries_smp_nb);
 432                 cpu_maps_update_begin();
 433                 if (cede_offline_enabled && parse_cede_parameters() == 0) {
 434                         default_offline_state = CPU_STATE_INACTIVE;
 435                         for_each_online_cpu(cpu)
 436                                 set_default_offline_state(cpu);
 437                 }
 438                 cpu_maps_update_done();
 439         }
 440
 441         return 0;
 442 }
 443 arch_initcall(pseries_cpu_hotplug_init);