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/interrupt.h>
  23 #include <linux/delay.h>
  24 #include <linux/sched.h>        /* for idle_task_exit */
  25 #include <linux/cpu.h>
  26 #include <asm/system.h>
  27 #include <asm/prom.h>
  28 #include <asm/rtas.h>
  29 #include <asm/firmware.h>
  30 #include <asm/machdep.h>
  31 #include <asm/vdso_datapage.h>
  32 #include <asm/pSeries_reconfig.h>
  33 #include <asm/xics.h>
  34 #include "plpar_wrappers.h"
  35 #include <asm/topology.h>
  36 #include "offline_states.h"
  37
  38 /* This version can't take the spinlock, because it never returns */
  39 static struct rtas_args rtas_stop_self_args = {
  40         .token = RTAS_UNKNOWN_SERVICE,
  41         .nargs = 0,
  42         .nret = 1,
  43         .rets = &rtas_stop_self_args.args[0],
  44 };
  45
  46 static DEFINE_PER_CPU(enum cpu_state_vals, preferred_offline_state) =
  47                                                         CPU_STATE_OFFLINE;
  48 static DEFINE_PER_CPU(enum cpu_state_vals, current_state) = CPU_STATE_OFFLINE;
  49
  50 static enum cpu_state_vals default_offline_state = CPU_STATE_OFFLINE;
  51
  52 static int cede_offline_enabled __read_mostly = 1;
  53
  54 /*
  55  * Enable/disable cede_offline when available.
  56  */
  57 static int __init setup_cede_offline(char *str)
  58 {
  59         if (!strcmp(str, "off"))
  60                 cede_offline_enabled = 0;
  61         else if (!strcmp(str, "on"))
  62                 cede_offline_enabled = 1;
  63         else
  64                 return 0;
  65         return 1;
  66 }
  67
  68 __setup("cede_offline=", setup_cede_offline);
  69
  70 enum cpu_state_vals get_cpu_current_state(int cpu)
  71 {
  72         return per_cpu(current_state, cpu);
  73 }
  74
  75 void set_cpu_current_state(int cpu, enum cpu_state_vals state)
  76 {
  77         per_cpu(current_state, cpu) = state;
  78 }
  79
  80 enum cpu_state_vals get_preferred_offline_state(int cpu)
  81 {
  82         return per_cpu(preferred_offline_state, cpu);
  83 }
  84
  85 void set_preferred_offline_state(int cpu, enum cpu_state_vals state)
  86 {
  87         per_cpu(preferred_offline_state, cpu) = state;
  88 }
  89
  90 void set_default_offline_state(int cpu)
  91 {
  92         per_cpu(preferred_offline_state, cpu) = default_offline_state;
  93 }
  94
  95 static void rtas_stop_self(void)
  96 {
  97         struct rtas_args *args = &rtas_stop_self_args;
  98
  99         local_irq_disable();
 100
 101         BUG_ON(args->token == RTAS_UNKNOWN_SERVICE);
 102
 103         printk("cpu %u (hwid %u) Ready to die...\n",
 104                smp_processor_id(), hard_smp_processor_id());
 105         enter_rtas(__pa(args));
 106
 107         panic("Alas, I survived.\n");
 108 }
 109
 110 static void pseries_mach_cpu_die(void)
 111 {
 112         unsigned int cpu = smp_processor_id();
 113         unsigned int hwcpu = hard_smp_processor_id();
 114         u8 cede_latency_hint = 0;
 115
 116         local_irq_disable();
 117         idle_task_exit();
 118         xics_teardown_cpu();
 119
 120         if (get_preferred_offline_state(cpu) == CPU_STATE_INACTIVE) {
 121                 set_cpu_current_state(cpu, CPU_STATE_INACTIVE);
 122                 if (ppc_md.suspend_disable_cpu)
 123                         ppc_md.suspend_disable_cpu();
 124
 125                 cede_latency_hint = 2;
 126
 127                 get_lppaca()->idle = 1;
 128                 if (!get_lppaca()->shared_proc)
 129                         get_lppaca()->donate_dedicated_cpu = 1;
 130
 131                 while (get_preferred_offline_state(cpu) == CPU_STATE_INACTIVE) {
 132                         extended_cede_processor(cede_latency_hint);
 133                 }
 134
 135                 if (!get_lppaca()->shared_proc)
 136                         get_lppaca()->donate_dedicated_cpu = 0;
 137                 get_lppaca()->idle = 0;
 138
 139                 if (get_preferred_offline_state(cpu) == CPU_STATE_ONLINE) {
 140                         unregister_slb_shadow(hwcpu);
 141
 142                         /*
 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.
 146                          */
 147                         start_secondary_resume();
 148                 }
 149         }
 150
 151         /* Requested state is CPU_STATE_OFFLINE at this point */
 152         WARN_ON(get_preferred_offline_state(cpu) != CPU_STATE_OFFLINE);
 153
 154         set_cpu_current_state(cpu, CPU_STATE_OFFLINE);
 155         unregister_slb_shadow(hwcpu);
 156         rtas_stop_self();
 157
 158         /* Should never get here... */
 159         BUG();
 160         for(;;);
 161 }
 162
 163 static int pseries_cpu_disable(void)
 164 {
 165         int cpu = smp_processor_id();
 166
 167         set_cpu_online(cpu, false);
 168         vdso_data->processorCount--;
 169
 170         /*fix boot_cpuid here*/
 171         if (cpu == boot_cpuid)
 172                 boot_cpuid = cpumask_any(cpu_online_mask);
 173
 174         /* FIXME: abstract this to not be platform specific later on */
 175         xics_migrate_irqs_away();
 176         return 0;
 177 }
 178
 179 /*
 180  * pseries_cpu_die: Wait for the cpu to die.
 181  * @cpu: logical processor id of the CPU whose death we're awaiting.
 182  *
 183  * This function is called from the context of the thread which is performing
 184  * the cpu-offline. Here we wait for long enough to allow the cpu in question
 185  * to self-destroy so that the cpu-offline thread can send the CPU_DEAD
 186  * notifications.
 187  *
 188  * OTOH, pseries_mach_cpu_die() is called by the @cpu when it wants to
 189  * self-destruct.
 190  */
 191 static void pseries_cpu_die(unsigned int cpu)
 192 {
 193         int tries;
 194         int cpu_status = 1;
 195         unsigned int pcpu = get_hard_smp_processor_id(cpu);
 196
 197         if (get_preferred_offline_state(cpu) == CPU_STATE_INACTIVE) {
 198                 cpu_status = 1;
 199                 for (tries = 0; tries < 5000; tries++) {
 200                         if (get_cpu_current_state(cpu) == CPU_STATE_INACTIVE) {
 201                                 cpu_status = 0;
 202                                 break;
 203                         }
 204                         msleep(1);
 205                 }
 206         } else if (get_preferred_offline_state(cpu) == CPU_STATE_OFFLINE) {
 207
 208                 for (tries = 0; tries < 25; tries++) {
 209                         cpu_status = smp_query_cpu_stopped(pcpu);
 210                         if (cpu_status == QCSS_STOPPED ||
 211                             cpu_status == QCSS_HARDWARE_ERROR)
 212                                 break;
 213                         cpu_relax();
 214                 }
 215         }
 216
 217         if (cpu_status != 0) {
 218                 printk("Querying DEAD? cpu %i (%i) shows %i\n",
 219                        cpu, pcpu, cpu_status);
 220         }
 221
 222         /* Isolation and deallocation are definitely done by
 223          * drslot_chrp_cpu.  If they were not they would be
 224          * done here.  Change isolate state to Isolate and
 225          * change allocation-state to Unusable.
 226          */
 227         paca[cpu].cpu_start = 0;
 228 }
 229
 230 /*
 231  * Update cpu_present_mask and paca(s) for a new cpu node.  The wrinkle
 232  * here is that a cpu device node may represent up to two logical cpus
 233  * in the SMT case.  We must honor the assumption in other code that
 234  * the logical ids for sibling SMT threads x and y are adjacent, such
 235  * that x^1 == y and y^1 == x.
 236  */
 237 static int pseries_add_processor(struct device_node *np)
 238 {
 239         unsigned int cpu;
 240         cpumask_var_t candidate_mask, tmp;
 241         int err = -ENOSPC, len, nthreads, i;
 242         const u32 *intserv;
 243
 244         intserv = of_get_property(np, "ibm,ppc-interrupt-server#s", &len);
 245         if (!intserv)
 246                 return 0;
 247
 248         zalloc_cpumask_var(&candidate_mask, GFP_KERNEL);
 249         zalloc_cpumask_var(&tmp, GFP_KERNEL);
 250
 251         nthreads = len / sizeof(u32);
 252         for (i = 0; i < nthreads; i++)
 253                 cpumask_set_cpu(i, tmp);
 254
 255         cpu_maps_update_begin();
 256
 257         BUG_ON(!cpumask_subset(cpu_present_mask, cpu_possible_mask));
 258
 259         /* Get a bitmap of unoccupied slots. */
 260         cpumask_xor(candidate_mask, cpu_possible_mask, cpu_present_mask);
 261         if (cpumask_empty(candidate_mask)) {
 262                 /* If we get here, it most likely means that NR_CPUS is
 263                  * less than the partition's max processors setting.
 264                  */
 265                 printk(KERN_ERR "Cannot add cpu %s; this system configuration"
 266                        " supports %d logical cpus.\n", np->full_name,
 267                        cpumask_weight(cpu_possible_mask));
 268                 goto out_unlock;
 269         }
 270
 271         while (!cpumask_empty(tmp))
 272                 if (cpumask_subset(tmp, candidate_mask))
 273                         /* Found a range where we can insert the new cpu(s) */
 274                         break;
 275                 else
 276                         cpumask_shift_left(tmp, tmp, nthreads);
 277
 278         if (cpumask_empty(tmp)) {
 279                 printk(KERN_ERR "Unable to find space in cpu_present_mask for"
 280                        " processor %s with %d thread(s)\n", np->name,
 281                        nthreads);
 282                 goto out_unlock;
 283         }
 284
 285         for_each_cpu(cpu, tmp) {
 286                 BUG_ON(cpu_present(cpu));
 287                 set_cpu_present(cpu, true);
 288                 set_hard_smp_processor_id(cpu, *intserv++);
 289         }
 290         err = 0;
 291 out_unlock:
 292         cpu_maps_update_done();
 293         free_cpumask_var(candidate_mask);
 294         free_cpumask_var(tmp);
 295         return err;
 296 }
 297
 298 /*
 299  * Update the present map for a cpu node which is going away, and set
 300  * the hard id in the paca(s) to -1 to be consistent with boot time
 301  * convention for non-present cpus.
 302  */
 303 static void pseries_remove_processor(struct device_node *np)
 304 {
 305         unsigned int cpu;
 306         int len, nthreads, i;
 307         const u32 *intserv;
 308
 309         intserv = of_get_property(np, "ibm,ppc-interrupt-server#s", &len);
 310         if (!intserv)
 311                 return;
 312
 313         nthreads = len / sizeof(u32);
 314
 315         cpu_maps_update_begin();
 316         for (i = 0; i < nthreads; i++) {
 317                 for_each_present_cpu(cpu) {
 318                         if (get_hard_smp_processor_id(cpu) != intserv[i])
 319                                 continue;
 320                         BUG_ON(cpu_online(cpu));
 321                         set_cpu_present(cpu, false);
 322                         set_hard_smp_processor_id(cpu, -1);
 323                         update_numa_cpu_lookup_table(cpu, -1);
 324                         break;
 325                 }
 326                 if (cpu >= nr_cpu_ids)
 327                         printk(KERN_WARNING "Could not find cpu to remove "
 328                                "with physical id 0x%x\n", intserv[i]);
 329         }
 330         cpu_maps_update_done();
 331 }
 332
 333 static int pseries_smp_notifier(struct notifier_block *nb,
 334                                 unsigned long action, void *node)
 335 {
 336         int err = 0;
 337
 338         switch (action) {
 339         case PSERIES_RECONFIG_ADD:
 340                 err = pseries_add_processor(node);
 341                 break;
 342         case PSERIES_RECONFIG_REMOVE:
 343                 pseries_remove_processor(node);
 344                 break;
 345         }
 346         return notifier_from_errno(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);