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                 while (get_preferred_offline_state(cpu) == CPU_STATE_INACTIVE) {
 126                         extended_cede_processor(cede_latency_hint);
 127                 }
 128
 129                 if (!get_lppaca()->shared_proc)
 130                         get_lppaca()->donate_dedicated_cpu = 0;
 131                 get_lppaca()->idle = 0;
 132
 133                 if (get_preferred_offline_state(cpu) == CPU_STATE_ONLINE) {
 134                         unregister_slb_shadow(hwcpu, __pa(get_slb_shadow()));
 135
 136                         /*
 137                          * Call to start_secondary_resume() will not return.
 138                          * Kernel stack will be reset and start_secondary()
 139                          * will be called to continue the online operation.
 140                          */
 141                         start_secondary_resume();
 142                 }
 143         }
 144
 145         /* Requested state is CPU_STATE_OFFLINE at this point */
 146         WARN_ON(get_preferred_offline_state(cpu) != CPU_STATE_OFFLINE);
 147
 148         set_cpu_current_state(cpu, CPU_STATE_OFFLINE);
 149         unregister_slb_shadow(hwcpu, __pa(get_slb_shadow()));
 150         rtas_stop_self();
 151
 152         /* Should never get here... */
 153         BUG();
 154         for(;;);
 155 }
 156
 157 static int qcss_tok;    /* query-cpu-stopped-state token */
 158
 159 /* Get state of physical CPU.
 160  * Return codes:
 161  *      0       - The processor is in the RTAS stopped state
 162  *      1       - stop-self is in progress
 163  *      2       - The processor is not in the RTAS stopped state
 164  *      -1      - Hardware Error
 165  *      -2      - Hardware Busy, Try again later.
 166  */
 167 static int query_cpu_stopped(unsigned int pcpu)
 168 {
 169         int cpu_status, status;
 170
 171         status = rtas_call(qcss_tok, 1, 2, &cpu_status, pcpu);
 172         if (status != 0) {
 173                 printk(KERN_ERR
 174                        "RTAS query-cpu-stopped-state failed: %i\n", status);
 175                 return status;
 176         }
 177
 178         return cpu_status;
 179 }
 180
 181 static int pseries_cpu_disable(void)
 182 {
 183         int cpu = smp_processor_id();
 184
 185         set_cpu_online(cpu, false);
 186         vdso_data->processorCount--;
 187
 188         /*fix boot_cpuid here*/
 189         if (cpu == boot_cpuid)
 190                 boot_cpuid = any_online_cpu(cpu_online_map);
 191
 192         /* FIXME: abstract this to not be platform specific later on */
 193         xics_migrate_irqs_away();
 194         return 0;
 195 }
 196
 197 /*
 198  * pseries_cpu_die: Wait for the cpu to die.
 199  * @cpu: logical processor id of the CPU whose death we're awaiting.
 200  *
 201  * This function is called from the context of the thread which is performing
 202  * the cpu-offline. Here we wait for long enough to allow the cpu in question
 203  * to self-destroy so that the cpu-offline thread can send the CPU_DEAD
 204  * notifications.
 205  *
 206  * OTOH, pseries_mach_cpu_die() is called by the @cpu when it wants to
 207  * self-destruct.
 208  */
 209 static void pseries_cpu_die(unsigned int cpu)
 210 {
 211         int tries;
 212         int cpu_status = 1;
 213         unsigned int pcpu = get_hard_smp_processor_id(cpu);
 214
 215         if (get_preferred_offline_state(cpu) == CPU_STATE_INACTIVE) {
 216                 cpu_status = 1;
 217                 for (tries = 0; tries < 1000; tries++) {
 218                         if (get_cpu_current_state(cpu) == CPU_STATE_INACTIVE) {
 219                                 cpu_status = 0;
 220                                 break;
 221                         }
 222                         cpu_relax();
 223                 }
 224         } else if (get_preferred_offline_state(cpu) == CPU_STATE_OFFLINE) {
 225
 226                 for (tries = 0; tries < 25; tries++) {
 227                         cpu_status = query_cpu_stopped(pcpu);
 228                         if (cpu_status == 0 || cpu_status == -1)
 229                                 break;
 230                         cpu_relax();
 231                 }
 232         }
 233
 234         if (cpu_status != 0) {
 235                 printk("Querying DEAD? cpu %i (%i) shows %i\n",
 236                        cpu, pcpu, cpu_status);
 237         }
 238
 239         /* Isolation and deallocation are definatly done by
 240          * drslot_chrp_cpu.  If they were not they would be
 241          * done here.  Change isolate state to Isolate and
 242          * change allocation-state to Unusable.
 243          */
 244         paca[cpu].cpu_start = 0;
 245 }
 246
 247 /*
 248  * Update cpu_present_map and paca(s) for a new cpu node.  The wrinkle
 249  * here is that a cpu device node may represent up to two logical cpus
 250  * in the SMT case.  We must honor the assumption in other code that
 251  * the logical ids for sibling SMT threads x and y are adjacent, such
 252  * that x^1 == y and y^1 == x.
 253  */
 254 static int pseries_add_processor(struct device_node *np)
 255 {
 256         unsigned int cpu;
 257         cpumask_t candidate_map, tmp = CPU_MASK_NONE;
 258         int err = -ENOSPC, len, nthreads, i;
 259         const u32 *intserv;
 260
 261         intserv = of_get_property(np, "ibm,ppc-interrupt-server#s", &len);
 262         if (!intserv)
 263                 return 0;
 264
 265         nthreads = len / sizeof(u32);
 266         for (i = 0; i < nthreads; i++)
 267                 cpu_set(i, tmp);
 268
 269         cpu_maps_update_begin();
 270
 271         BUG_ON(!cpus_subset(cpu_present_map, cpu_possible_map));
 272
 273         /* Get a bitmap of unoccupied slots. */
 274         cpus_xor(candidate_map, cpu_possible_map, cpu_present_map);
 275         if (cpus_empty(candidate_map)) {
 276                 /* If we get here, it most likely means that NR_CPUS is
 277                  * less than the partition's max processors setting.
 278                  */
 279                 printk(KERN_ERR "Cannot add cpu %s; this system configuration"
 280                        " supports %d logical cpus.\n", np->full_name,
 281                        cpus_weight(cpu_possible_map));
 282                 goto out_unlock;
 283         }
 284
 285         while (!cpus_empty(tmp))
 286                 if (cpus_subset(tmp, candidate_map))
 287                         /* Found a range where we can insert the new cpu(s) */
 288                         break;
 289                 else
 290                         cpus_shift_left(tmp, tmp, nthreads);
 291
 292         if (cpus_empty(tmp)) {
 293                 printk(KERN_ERR "Unable to find space in cpu_present_map for"
 294                        " processor %s with %d thread(s)\n", np->name,
 295                        nthreads);
 296                 goto out_unlock;
 297         }
 298
 299         for_each_cpu_mask(cpu, tmp) {
 300                 BUG_ON(cpu_isset(cpu, cpu_present_map));
 301                 set_cpu_present(cpu, true);
 302                 set_hard_smp_processor_id(cpu, *intserv++);
 303         }
 304         err = 0;
 305 out_unlock:
 306         cpu_maps_update_done();
 307         return err;
 308 }
 309
 310 /*
 311  * Update the present map for a cpu node which is going away, and set
 312  * the hard id in the paca(s) to -1 to be consistent with boot time
 313  * convention for non-present cpus.
 314  */
 315 static void pseries_remove_processor(struct device_node *np)
 316 {
 317         unsigned int cpu;
 318         int len, nthreads, i;
 319         const u32 *intserv;
 320
 321         intserv = of_get_property(np, "ibm,ppc-interrupt-server#s", &len);
 322         if (!intserv)
 323                 return;
 324
 325         nthreads = len / sizeof(u32);
 326
 327         cpu_maps_update_begin();
 328         for (i = 0; i < nthreads; i++) {
 329                 for_each_present_cpu(cpu) {
 330                         if (get_hard_smp_processor_id(cpu) != intserv[i])
 331                                 continue;
 332                         BUG_ON(cpu_online(cpu));
 333                         set_cpu_present(cpu, false);
 334                         set_hard_smp_processor_id(cpu, -1);
 335                         break;
 336                 }
 337                 if (cpu == NR_CPUS)
 338                         printk(KERN_WARNING "Could not find cpu to remove "
 339                                "with physical id 0x%x\n", intserv[i]);
 340         }
 341         cpu_maps_update_done();
 342 }
 343
 344 static int pseries_smp_notifier(struct notifier_block *nb,
 345                                 unsigned long action, void *node)
 346 {
 347         int err = NOTIFY_OK;
 348
 349         switch (action) {
 350         case PSERIES_RECONFIG_ADD:
 351                 if (pseries_add_processor(node))
 352                         err = NOTIFY_BAD;
 353                 break;
 354         case PSERIES_RECONFIG_REMOVE:
 355                 pseries_remove_processor(node);
 356                 break;
 357         default:
 358                 err = NOTIFY_DONE;
 359                 break;
 360         }
 361         return err;
 362 }
 363
 364 static struct notifier_block pseries_smp_nb = {
 365         .notifier_call = pseries_smp_notifier,
 366 };
 367
 368 #define MAX_CEDE_LATENCY_LEVELS         4
 369 #define CEDE_LATENCY_PARAM_LENGTH       10
 370 #define CEDE_LATENCY_PARAM_MAX_LENGTH   \
 371         (MAX_CEDE_LATENCY_LEVELS * CEDE_LATENCY_PARAM_LENGTH * sizeof(char))
 372 #define CEDE_LATENCY_TOKEN              45
 373
 374 static char cede_parameters[CEDE_LATENCY_PARAM_MAX_LENGTH];
 375
 376 static int parse_cede_parameters(void)
 377 {
 378         memset(cede_parameters, 0, CEDE_LATENCY_PARAM_MAX_LENGTH);
 379         return rtas_call(rtas_token("ibm,get-system-parameter"), 3, 1,
 380                          NULL,
 381                          CEDE_LATENCY_TOKEN,
 382                          __pa(cede_parameters),
 383                          CEDE_LATENCY_PARAM_MAX_LENGTH);
 384 }
 385
 386 static int __init pseries_cpu_hotplug_init(void)
 387 {
 388         struct device_node *np;
 389         const char *typep;
 390         int cpu;
 391
 392         for_each_node_by_name(np, "interrupt-controller") {
 393                 typep = of_get_property(np, "compatible", NULL);
 394                 if (strstr(typep, "open-pic")) {
 395                         of_node_put(np);
 396
 397                         printk(KERN_INFO "CPU Hotplug not supported on "
 398                                 "systems using MPIC\n");
 399                         return 0;
 400                 }
 401         }
 402
 403         rtas_stop_self_args.token = rtas_token("stop-self");
 404         qcss_tok = rtas_token("query-cpu-stopped-state");
 405
 406         if (rtas_stop_self_args.token == RTAS_UNKNOWN_SERVICE ||
 407                         qcss_tok == RTAS_UNKNOWN_SERVICE) {
 408                 printk(KERN_INFO "CPU Hotplug not supported by firmware "
 409                                 "- disabling.\n");
 410                 return 0;
 411         }
 412
 413         ppc_md.cpu_die = pseries_mach_cpu_die;
 414         smp_ops->cpu_disable = pseries_cpu_disable;
 415         smp_ops->cpu_die = pseries_cpu_die;
 416
 417         /* Processors can be added/removed only on LPAR */
 418         if (firmware_has_feature(FW_FEATURE_LPAR)) {
 419                 pSeries_reconfig_notifier_register(&pseries_smp_nb);
 420                 cpu_maps_update_begin();
 421                 if (cede_offline_enabled && parse_cede_parameters() == 0) {
 422                         default_offline_state = CPU_STATE_INACTIVE;
 423                         for_each_online_cpu(cpu)
 424                                 set_default_offline_state(cpu);
 425                 }
 426                 cpu_maps_update_done();
 427         }
 428
 429         return 0;
 430 }
 431 arch_initcall(pseries_cpu_hotplug_init);