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                 if (ppc_md.suspend_disable_cpu)
 120                         ppc_md.suspend_disable_cpu();
 121
 122                 cede_latency_hint = 2;
 123
 124                 get_lppaca()->idle = 1;
 125                 if (!get_lppaca()->shared_proc)
 126                         get_lppaca()->donate_dedicated_cpu = 1;
 127
 128                 while (get_preferred_offline_state(cpu) == CPU_STATE_INACTIVE) {
 129                         extended_cede_processor(cede_latency_hint);
 130                 }
 131
 132                 if (!get_lppaca()->shared_proc)
 133                         get_lppaca()->donate_dedicated_cpu = 0;
 134                 get_lppaca()->idle = 0;
 135
 136                 if (get_preferred_offline_state(cpu) == CPU_STATE_ONLINE) {
 137                         unregister_slb_shadow(hwcpu, __pa(get_slb_shadow()));
 138
 139                         /*
 140                          * Call to start_secondary_resume() will not return.
 141                          * Kernel stack will be reset and start_secondary()
 142                          * will be called to continue the online operation.
 143                          */
 144                         start_secondary_resume();
 145                 }
 146         }
 147
 148         /* Requested state is CPU_STATE_OFFLINE at this point */
 149         WARN_ON(get_preferred_offline_state(cpu) != CPU_STATE_OFFLINE);
 150
 151         set_cpu_current_state(cpu, CPU_STATE_OFFLINE);
 152         unregister_slb_shadow(hwcpu, __pa(get_slb_shadow()));
 153         rtas_stop_self();
 154
 155         /* Should never get here... */
 156         BUG();
 157         for(;;);
 158 }
 159
 160 static int pseries_cpu_disable(void)
 161 {
 162         int cpu = smp_processor_id();
 163
 164         set_cpu_online(cpu, false);
 165         vdso_data->processorCount--;
 166
 167         /*fix boot_cpuid here*/
 168         if (cpu == boot_cpuid)
 169                 boot_cpuid = cpumask_any(cpu_online_mask);
 170
 171         /* FIXME: abstract this to not be platform specific later on */
 172         xics_migrate_irqs_away();
 173         return 0;
 174 }
 175
 176 /*
 177  * pseries_cpu_die: Wait for the cpu to die.
 178  * @cpu: logical processor id of the CPU whose death we're awaiting.
 179  *
 180  * This function is called from the context of the thread which is performing
 181  * the cpu-offline. Here we wait for long enough to allow the cpu in question
 182  * to self-destroy so that the cpu-offline thread can send the CPU_DEAD
 183  * notifications.
 184  *
 185  * OTOH, pseries_mach_cpu_die() is called by the @cpu when it wants to
 186  * self-destruct.
 187  */
 188 static void pseries_cpu_die(unsigned int cpu)
 189 {
 190         int tries;
 191         int cpu_status = 1;
 192         unsigned int pcpu = get_hard_smp_processor_id(cpu);
 193
 194         if (get_preferred_offline_state(cpu) == CPU_STATE_INACTIVE) {
 195                 cpu_status = 1;
 196                 for (tries = 0; tries < 5000; tries++) {
 197                         if (get_cpu_current_state(cpu) == CPU_STATE_INACTIVE) {
 198                                 cpu_status = 0;
 199                                 break;
 200                         }
 201                         msleep(1);
 202                 }
 203         } else if (get_preferred_offline_state(cpu) == CPU_STATE_OFFLINE) {
 204
 205                 for (tries = 0; tries < 25; tries++) {
 206                         cpu_status = smp_query_cpu_stopped(pcpu);
 207                         if (cpu_status == QCSS_STOPPED ||
 208                             cpu_status == QCSS_HARDWARE_ERROR)
 209                                 break;
 210                         cpu_relax();
 211                 }
 212         }
 213
 214         if (cpu_status != 0) {
 215                 printk("Querying DEAD? cpu %i (%i) shows %i\n",
 216                        cpu, pcpu, cpu_status);
 217         }
 218
 219         /* Isolation and deallocation are definitely done by
 220          * drslot_chrp_cpu.  If they were not they would be
 221          * done here.  Change isolate state to Isolate and
 222          * change allocation-state to Unusable.
 223          */
 224         paca[cpu].cpu_start = 0;
 225 }
 226
 227 /*
 228  * Update cpu_present_mask and paca(s) for a new cpu node.  The wrinkle
 229  * here is that a cpu device node may represent up to two logical cpus
 230  * in the SMT case.  We must honor the assumption in other code that
 231  * the logical ids for sibling SMT threads x and y are adjacent, such
 232  * that x^1 == y and y^1 == x.
 233  */
 234 static int pseries_add_processor(struct device_node *np)
 235 {
 236         unsigned int cpu;
 237         cpumask_var_t candidate_mask, tmp;
 238         int err = -ENOSPC, len, nthreads, i;
 239         const u32 *intserv;
 240
 241         intserv = of_get_property(np, "ibm,ppc-interrupt-server#s", &len);
 242         if (!intserv)
 243                 return 0;
 244
 245         zalloc_cpumask_var(&candidate_mask, GFP_KERNEL);
 246         zalloc_cpumask_var(&tmp, GFP_KERNEL);
 247
 248         nthreads = len / sizeof(u32);
 249         for (i = 0; i < nthreads; i++)
 250                 cpumask_set_cpu(i, tmp);
 251
 252         cpu_maps_update_begin();
 253
 254         BUG_ON(!cpumask_subset(cpu_present_mask, cpu_possible_mask));
 255
 256         /* Get a bitmap of unoccupied slots. */
 257         cpumask_xor(candidate_mask, cpu_possible_mask, cpu_present_mask);
 258         if (cpumask_empty(candidate_mask)) {
 259                 /* If we get here, it most likely means that NR_CPUS is
 260                  * less than the partition's max processors setting.
 261                  */
 262                 printk(KERN_ERR "Cannot add cpu %s; this system configuration"
 263                        " supports %d logical cpus.\n", np->full_name,
 264                        cpumask_weight(cpu_possible_mask));
 265                 goto out_unlock;
 266         }
 267
 268         while (!cpumask_empty(tmp))
 269                 if (cpumask_subset(tmp, candidate_mask))
 270                         /* Found a range where we can insert the new cpu(s) */
 271                         break;
 272                 else
 273                         cpumask_shift_left(tmp, tmp, nthreads);
 274
 275         if (cpumask_empty(tmp)) {
 276                 printk(KERN_ERR "Unable to find space in cpu_present_mask for"
 277                        " processor %s with %d thread(s)\n", np->name,
 278                        nthreads);
 279                 goto out_unlock;
 280         }
 281
 282         for_each_cpu(cpu, tmp) {
 283                 BUG_ON(cpumask_test_cpu(cpu, cpu_present_mask));
 284                 set_cpu_present(cpu, true);
 285                 set_hard_smp_processor_id(cpu, *intserv++);
 286         }
 287         err = 0;
 288 out_unlock:
 289         cpu_maps_update_done();
 290         free_cpumask_var(candidate_mask);
 291         free_cpumask_var(tmp);
 292         return err;
 293 }
 294
 295 /*
 296  * Update the present map for a cpu node which is going away, and set
 297  * the hard id in the paca(s) to -1 to be consistent with boot time
 298  * convention for non-present cpus.
 299  */
 300 static void pseries_remove_processor(struct device_node *np)
 301 {
 302         unsigned int cpu;
 303         int len, nthreads, i;
 304         const u32 *intserv;
 305
 306         intserv = of_get_property(np, "ibm,ppc-interrupt-server#s", &len);
 307         if (!intserv)
 308                 return;
 309
 310         nthreads = len / sizeof(u32);
 311
 312         cpu_maps_update_begin();
 313         for (i = 0; i < nthreads; i++) {
 314                 for_each_present_cpu(cpu) {
 315                         if (get_hard_smp_processor_id(cpu) != intserv[i])
 316                                 continue;
 317                         BUG_ON(cpu_online(cpu));
 318                         set_cpu_present(cpu, false);
 319                         set_hard_smp_processor_id(cpu, -1);
 320                         break;
 321                 }
 322                 if (cpu >= nr_cpu_ids)
 323                         printk(KERN_WARNING "Could not find cpu to remove "
 324                                "with physical id 0x%x\n", intserv[i]);
 325         }
 326         cpu_maps_update_done();
 327 }
 328
 329 static int pseries_smp_notifier(struct notifier_block *nb,
 330                                 unsigned long action, void *node)
 331 {
 332         int err = NOTIFY_OK;
 333
 334         switch (action) {
 335         case PSERIES_RECONFIG_ADD:
 336                 if (pseries_add_processor(node))
 337                         err = NOTIFY_BAD;
 338                 break;
 339         case PSERIES_RECONFIG_REMOVE:
 340                 pseries_remove_processor(node);
 341                 break;
 342         default:
 343                 err = NOTIFY_DONE;
 344                 break;
 345         }
 346         return err;
 347 }
 348
 349 static struct notifier_block pseries_smp_nb = {
 350         .notifier_call = pseries_smp_notifier,
 351 };
 352
 353 #define MAX_CEDE_LATENCY_LEVELS         4
 354 #define CEDE_LATENCY_PARAM_LENGTH       10
 355 #define CEDE_LATENCY_PARAM_MAX_LENGTH   \
 356         (MAX_CEDE_LATENCY_LEVELS * CEDE_LATENCY_PARAM_LENGTH * sizeof(char))
 357 #define CEDE_LATENCY_TOKEN              45
 358
 359 static char cede_parameters[CEDE_LATENCY_PARAM_MAX_LENGTH];
 360
 361 static int parse_cede_parameters(void)
 362 {
 363         memset(cede_parameters, 0, CEDE_LATENCY_PARAM_MAX_LENGTH);
 364         return rtas_call(rtas_token("ibm,get-system-parameter"), 3, 1,
 365                          NULL,
 366                          CEDE_LATENCY_TOKEN,
 367                          __pa(cede_parameters),
 368                          CEDE_LATENCY_PARAM_MAX_LENGTH);
 369 }
 370
 371 static int __init pseries_cpu_hotplug_init(void)
 372 {
 373         struct device_node *np;
 374         const char *typep;
 375         int cpu;
 376         int qcss_tok;
 377
 378         for_each_node_by_name(np, "interrupt-controller") {
 379                 typep = of_get_property(np, "compatible", NULL);
 380                 if (strstr(typep, "open-pic")) {
 381                         of_node_put(np);
 382
 383                         printk(KERN_INFO "CPU Hotplug not supported on "
 384                                 "systems using MPIC\n");
 385                         return 0;
 386                 }
 387         }
 388
 389         rtas_stop_self_args.token = rtas_token("stop-self");
 390         qcss_tok = rtas_token("query-cpu-stopped-state");
 391
 392         if (rtas_stop_self_args.token == RTAS_UNKNOWN_SERVICE ||
 393                         qcss_tok == RTAS_UNKNOWN_SERVICE) {
 394                 printk(KERN_INFO "CPU Hotplug not supported by firmware "
 395                                 "- disabling.\n");
 396                 return 0;
 397         }
 398
 399         ppc_md.cpu_die = pseries_mach_cpu_die;
 400         smp_ops->cpu_disable = pseries_cpu_disable;
 401         smp_ops->cpu_die = pseries_cpu_die;
 402
 403         /* Processors can be added/removed only on LPAR */
 404         if (firmware_has_feature(FW_FEATURE_LPAR)) {
 405                 pSeries_reconfig_notifier_register(&pseries_smp_nb);
 406                 cpu_maps_update_begin();
 407                 if (cede_offline_enabled && parse_cede_parameters() == 0) {
 408                         default_offline_state = CPU_STATE_INACTIVE;
 409                         for_each_online_cpu(cpu)
 410                                 set_default_offline_state(cpu);
 411                 }
 412                 cpu_maps_update_done();
 413         }
 414
 415         return 0;
 416 }
 417 arch_initcall(pseries_cpu_hotplug_init);