arch/ppc64/kernel/pSeries_smp.c

   1 /*
   2  * SMP support for pSeries and BPA machines.
   3  *
   4  * Dave Engebretsen, Peter Bergner, and
   5  * Mike Corrigan {engebret|bergner|mikec}@us.ibm.com
   6  *
   7  * Plus various changes from other IBM teams...
   8  *
   9  *      This program is free software; you can redistribute it and/or
  10  *      modify it under the terms of the GNU General Public License
  11  *      as published by the Free Software Foundation; either version
  12  *      2 of the License, or (at your option) any later version.
  13  */
  14
  15 #undef DEBUG
  16
  17 #include <linux/config.h>
  18 #include <linux/kernel.h>
  19 #include <linux/module.h>
  20 #include <linux/sched.h>
  21 #include <linux/smp.h>
  22 #include <linux/interrupt.h>
  23 #include <linux/delay.h>
  24 #include <linux/init.h>
  25 #include <linux/spinlock.h>
  26 #include <linux/cache.h>
  27 #include <linux/err.h>
  28 #include <linux/sysdev.h>
  29 #include <linux/cpu.h>
  30
  31 #include <asm/ptrace.h>
  32 #include <asm/atomic.h>
  33 #include <asm/irq.h>
  34 #include <asm/page.h>
  35 #include <asm/pgtable.h>
  36 #include <asm/io.h>
  37 #include <asm/prom.h>
  38 #include <asm/smp.h>
  39 #include <asm/paca.h>
  40 #include <asm/time.h>
  41 #include <asm/machdep.h>
  42 #include <asm/xics.h>
  43 #include <asm/cputable.h>
  44 #include <asm/system.h>
  45 #include <asm/rtas.h>
  46 #include <asm/plpar_wrappers.h>
  47 #include <asm/pSeries_reconfig.h>
  48
  49 #include "mpic.h"
  50 #include "bpa_iic.h"
  51
  52 #ifdef DEBUG
  53 #define DBG(fmt...) udbg_printf(fmt)
  54 #else
  55 #define DBG(fmt...)
  56 #endif
  57
  58 /*
  59  * The primary thread of each non-boot processor is recorded here before
  60  * smp init.
  61  */
  62 static cpumask_t of_spin_map;
  63
  64 extern void pSeries_secondary_smp_init(unsigned long);
  65
  66 #ifdef CONFIG_HOTPLUG_CPU
  67
  68 /* Get state of physical CPU.
  69  * Return codes:
  70  *      0       - The processor is in the RTAS stopped state
  71  *      1       - stop-self is in progress
  72  *      2       - The processor is not in the RTAS stopped state
  73  *      -1      - Hardware Error
  74  *      -2      - Hardware Busy, Try again later.
  75  */
  76 static int query_cpu_stopped(unsigned int pcpu)
  77 {
  78         int cpu_status;
  79         int status, qcss_tok;
  80
  81         qcss_tok = rtas_token("query-cpu-stopped-state");
  82         if (qcss_tok == RTAS_UNKNOWN_SERVICE)
  83                 return -1;
  84         status = rtas_call(qcss_tok, 1, 2, &cpu_status, pcpu);
  85         if (status != 0) {
  86                 printk(KERN_ERR
  87                        "RTAS query-cpu-stopped-state failed: %i\n", status);
  88                 return status;
  89         }
  90
  91         return cpu_status;
  92 }
  93
  94 int pSeries_cpu_disable(void)
  95 {
  96         systemcfg->processorCount--;
  97
  98         /*fix boot_cpuid here*/
  99         if (smp_processor_id() == boot_cpuid)
 100                 boot_cpuid = any_online_cpu(cpu_online_map);
 101
 102         /* FIXME: abstract this to not be platform specific later on */
 103         xics_migrate_irqs_away();
 104         return 0;
 105 }
 106
 107 void pSeries_cpu_die(unsigned int cpu)
 108 {
 109         int tries;
 110         int cpu_status;
 111         unsigned int pcpu = get_hard_smp_processor_id(cpu);
 112
 113         for (tries = 0; tries < 25; tries++) {
 114                 cpu_status = query_cpu_stopped(pcpu);
 115                 if (cpu_status == 0 || cpu_status == -1)
 116                         break;
 117                 msleep(200);
 118         }
 119         if (cpu_status != 0) {
 120                 printk("Querying DEAD? cpu %i (%i) shows %i\n",
 121                        cpu, pcpu, cpu_status);
 122         }
 123
 124         /* Isolation and deallocation are definatly done by
 125          * drslot_chrp_cpu.  If they were not they would be
 126          * done here.  Change isolate state to Isolate and
 127          * change allocation-state to Unusable.
 128          */
 129         paca[cpu].cpu_start = 0;
 130 }
 131
 132 /*
 133  * Update cpu_present_map and paca(s) for a new cpu node.  The wrinkle
 134  * here is that a cpu device node may represent up to two logical cpus
 135  * in the SMT case.  We must honor the assumption in other code that
 136  * the logical ids for sibling SMT threads x and y are adjacent, such
 137  * that x^1 == y and y^1 == x.
 138  */
 139 static int pSeries_add_processor(struct device_node *np)
 140 {
 141         unsigned int cpu;
 142         cpumask_t candidate_map, tmp = CPU_MASK_NONE;
 143         int err = -ENOSPC, len, nthreads, i;
 144         u32 *intserv;
 145
 146         intserv = (u32 *)get_property(np, "ibm,ppc-interrupt-server#s", &len);
 147         if (!intserv)
 148                 return 0;
 149
 150         nthreads = len / sizeof(u32);
 151         for (i = 0; i < nthreads; i++)
 152                 cpu_set(i, tmp);
 153
 154         lock_cpu_hotplug();
 155
 156         BUG_ON(!cpus_subset(cpu_present_map, cpu_possible_map));
 157
 158         /* Get a bitmap of unoccupied slots. */
 159         cpus_xor(candidate_map, cpu_possible_map, cpu_present_map);
 160         if (cpus_empty(candidate_map)) {
 161                 /* If we get here, it most likely means that NR_CPUS is
 162                  * less than the partition's max processors setting.
 163                  */
 164                 printk(KERN_ERR "Cannot add cpu %s; this system configuration"
 165                        " supports %d logical cpus.\n", np->full_name,
 166                        cpus_weight(cpu_possible_map));
 167                 goto out_unlock;
 168         }
 169
 170         while (!cpus_empty(tmp))
 171                 if (cpus_subset(tmp, candidate_map))
 172                         /* Found a range where we can insert the new cpu(s) */
 173                         break;
 174                 else
 175                         cpus_shift_left(tmp, tmp, nthreads);
 176
 177         if (cpus_empty(tmp)) {
 178                 printk(KERN_ERR "Unable to find space in cpu_present_map for"
 179                        " processor %s with %d thread(s)\n", np->name,
 180                        nthreads);
 181                 goto out_unlock;
 182         }
 183
 184         for_each_cpu_mask(cpu, tmp) {
 185                 BUG_ON(cpu_isset(cpu, cpu_present_map));
 186                 cpu_set(cpu, cpu_present_map);
 187                 set_hard_smp_processor_id(cpu, *intserv++);
 188         }
 189         err = 0;
 190 out_unlock:
 191         unlock_cpu_hotplug();
 192         return err;
 193 }
 194
 195 /*
 196  * Update the present map for a cpu node which is going away, and set
 197  * the hard id in the paca(s) to -1 to be consistent with boot time
 198  * convention for non-present cpus.
 199  */
 200 static void pSeries_remove_processor(struct device_node *np)
 201 {
 202         unsigned int cpu;
 203         int len, nthreads, i;
 204         u32 *intserv;
 205
 206         intserv = (u32 *)get_property(np, "ibm,ppc-interrupt-server#s", &len);
 207         if (!intserv)
 208                 return;
 209
 210         nthreads = len / sizeof(u32);
 211
 212         lock_cpu_hotplug();
 213         for (i = 0; i < nthreads; i++) {
 214                 for_each_present_cpu(cpu) {
 215                         if (get_hard_smp_processor_id(cpu) != intserv[i])
 216                                 continue;
 217                         BUG_ON(cpu_online(cpu));
 218                         cpu_clear(cpu, cpu_present_map);
 219                         set_hard_smp_processor_id(cpu, -1);
 220                         break;
 221                 }
 222                 if (cpu == NR_CPUS)
 223                         printk(KERN_WARNING "Could not find cpu to remove "
 224                                "with physical id 0x%x\n", intserv[i]);
 225         }
 226         unlock_cpu_hotplug();
 227 }
 228
 229 static int pSeries_smp_notifier(struct notifier_block *nb, unsigned long action, void *node)
 230 {
 231         int err = NOTIFY_OK;
 232
 233         switch (action) {
 234         case PSERIES_RECONFIG_ADD:
 235                 if (pSeries_add_processor(node))
 236                         err = NOTIFY_BAD;
 237                 break;
 238         case PSERIES_RECONFIG_REMOVE:
 239                 pSeries_remove_processor(node);
 240                 break;
 241         default:
 242                 err = NOTIFY_DONE;
 243                 break;
 244         }
 245         return err;
 246 }
 247
 248 static struct notifier_block pSeries_smp_nb = {
 249         .notifier_call = pSeries_smp_notifier,
 250 };
 251
 252 #endif /* CONFIG_HOTPLUG_CPU */
 253
 254 /**
 255  * smp_startup_cpu() - start the given cpu
 256  *
 257  * At boot time, there is nothing to do for primary threads which were
 258  * started from Open Firmware.  For anything else, call RTAS with the
 259  * appropriate start location.
 260  *
 261  * Returns:
 262  *      0       - failure
 263  *      1       - success
 264  */
 265 static inline int __devinit smp_startup_cpu(unsigned int lcpu)
 266 {
 267         int status;
 268         unsigned long start_here = __pa((u32)*((unsigned long *)
 269                                                pSeries_secondary_smp_init));
 270         unsigned int pcpu;
 271
 272         if (cpu_isset(lcpu, of_spin_map))
 273                 /* Already started by OF and sitting in spin loop */
 274                 return 1;
 275
 276         pcpu = get_hard_smp_processor_id(lcpu);
 277
 278         /* Fixup atomic count: it exited inside IRQ handler. */
 279         paca[lcpu].__current->thread_info->preempt_count        = 0;
 280
 281         status = rtas_call(rtas_token("start-cpu"), 3, 1, NULL,
 282                            pcpu, start_here, lcpu);
 283         if (status != 0) {
 284                 printk(KERN_ERR "start-cpu failed: %i\n", status);
 285                 return 0;
 286         }
 287         return 1;
 288 }
 289
 290 #ifdef CONFIG_XICS
 291 static inline void smp_xics_do_message(int cpu, int msg)
 292 {
 293         set_bit(msg, &xics_ipi_message[cpu].value);
 294         mb();
 295         xics_cause_IPI(cpu);
 296 }
 297
 298 static void smp_xics_message_pass(int target, int msg)
 299 {
 300         unsigned int i;
 301
 302         if (target < NR_CPUS) {
 303                 smp_xics_do_message(target, msg);
 304         } else {
 305                 for_each_online_cpu(i) {
 306                         if (target == MSG_ALL_BUT_SELF
 307                             && i == smp_processor_id())
 308                                 continue;
 309                         smp_xics_do_message(i, msg);
 310                 }
 311         }
 312 }
 313
 314 static int __init smp_xics_probe(void)
 315 {
 316         xics_request_IPIs();
 317
 318         return cpus_weight(cpu_possible_map);
 319 }
 320
 321 static void __devinit smp_xics_setup_cpu(int cpu)
 322 {
 323         if (cpu != boot_cpuid)
 324                 xics_setup_cpu();
 325
 326         if (cur_cpu_spec->firmware_features & FW_FEATURE_SPLPAR)
 327                 vpa_init(cpu);
 328
 329         cpu_clear(cpu, of_spin_map);
 330
 331 }
 332 #endif /* CONFIG_XICS */
 333 #ifdef CONFIG_BPA_IIC
 334 static void smp_iic_message_pass(int target, int msg)
 335 {
 336         unsigned int i;
 337
 338         if (target < NR_CPUS) {
 339                 iic_cause_IPI(target, msg);
 340         } else {
 341                 for_each_online_cpu(i) {
 342                         if (target == MSG_ALL_BUT_SELF
 343                             && i == smp_processor_id())
 344                                 continue;
 345                         iic_cause_IPI(i, msg);
 346                 }
 347         }
 348 }
 349
 350 static int __init smp_iic_probe(void)
 351 {
 352         iic_request_IPIs();
 353
 354         return cpus_weight(cpu_possible_map);
 355 }
 356
 357 static void __devinit smp_iic_setup_cpu(int cpu)
 358 {
 359         if (cpu != boot_cpuid)
 360                 iic_setup_cpu();
 361 }
 362 #endif /* CONFIG_BPA_IIC */
 363
 364 static DEFINE_SPINLOCK(timebase_lock);
 365 static unsigned long timebase = 0;
 366
 367 static void __devinit pSeries_give_timebase(void)
 368 {
 369         spin_lock(&timebase_lock);
 370         rtas_call(rtas_token("freeze-time-base"), 0, 1, NULL);
 371         timebase = get_tb();
 372         spin_unlock(&timebase_lock);
 373
 374         while (timebase)
 375                 barrier();
 376         rtas_call(rtas_token("thaw-time-base"), 0, 1, NULL);
 377 }
 378
 379 static void __devinit pSeries_take_timebase(void)
 380 {
 381         while (!timebase)
 382                 barrier();
 383         spin_lock(&timebase_lock);
 384         set_tb(timebase >> 32, timebase & 0xffffffff);
 385         timebase = 0;
 386         spin_unlock(&timebase_lock);
 387 }
 388
 389 static void __devinit smp_pSeries_kick_cpu(int nr)
 390 {
 391         BUG_ON(nr < 0 || nr >= NR_CPUS);
 392
 393         if (!smp_startup_cpu(nr))
 394                 return;
 395
 396         /*
 397          * The processor is currently spinning, waiting for the
 398          * cpu_start field to become non-zero After we set cpu_start,
 399          * the processor will continue on to secondary_start
 400          */
 401         paca[nr].cpu_start = 1;
 402 }
 403
 404 static int smp_pSeries_cpu_bootable(unsigned int nr)
 405 {
 406         /* Special case - we inhibit secondary thread startup
 407          * during boot if the user requests it.  Odd-numbered
 408          * cpus are assumed to be secondary threads.
 409          */
 410         if (system_state < SYSTEM_RUNNING &&
 411             cpu_has_feature(CPU_FTR_SMT) &&
 412             !smt_enabled_at_boot && nr % 2 != 0)
 413                 return 0;
 414
 415         return 1;
 416 }
 417 #ifdef CONFIG_MPIC
 418 static struct smp_ops_t pSeries_mpic_smp_ops = {
 419         .message_pass   = smp_mpic_message_pass,
 420         .probe          = smp_mpic_probe,
 421         .kick_cpu       = smp_pSeries_kick_cpu,
 422         .setup_cpu      = smp_mpic_setup_cpu,
 423 };
 424 #endif
 425 #ifdef CONFIG_XICS
 426 static struct smp_ops_t pSeries_xics_smp_ops = {
 427         .message_pass   = smp_xics_message_pass,
 428         .probe          = smp_xics_probe,
 429         .kick_cpu       = smp_pSeries_kick_cpu,
 430         .setup_cpu      = smp_xics_setup_cpu,
 431         .cpu_bootable   = smp_pSeries_cpu_bootable,
 432 };
 433 #endif
 434 #ifdef CONFIG_BPA_IIC
 435 static struct smp_ops_t bpa_iic_smp_ops = {
 436         .message_pass   = smp_iic_message_pass,
 437         .probe          = smp_iic_probe,
 438         .kick_cpu       = smp_pSeries_kick_cpu,
 439         .setup_cpu      = smp_iic_setup_cpu,
 440         .cpu_bootable   = smp_pSeries_cpu_bootable,
 441 };
 442 #endif
 443
 444 /* This is called very early */
 445 void __init smp_init_pSeries(void)
 446 {
 447         int i;
 448
 449         DBG(" -> smp_init_pSeries()\n");
 450
 451         switch (ppc64_interrupt_controller) {
 452 #ifdef CONFIG_MPIC
 453         case IC_OPEN_PIC:
 454                 smp_ops = &pSeries_mpic_smp_ops;
 455                 break;
 456 #endif
 457 #ifdef CONFIG_XICS
 458         case IC_PPC_XIC:
 459                 smp_ops = &pSeries_xics_smp_ops;
 460                 break;
 461 #endif
 462 #ifdef CONFIG_BPA_IIC
 463         case IC_BPA_IIC:
 464                 smp_ops = &bpa_iic_smp_ops;
 465                 break;
 466 #endif
 467         default:
 468                 panic("Invalid interrupt controller");
 469         }
 470
 471 #ifdef CONFIG_HOTPLUG_CPU
 472         smp_ops->cpu_disable = pSeries_cpu_disable;
 473         smp_ops->cpu_die = pSeries_cpu_die;
 474
 475         /* Processors can be added/removed only on LPAR */
 476         if (systemcfg->platform == PLATFORM_PSERIES_LPAR)
 477                 pSeries_reconfig_notifier_register(&pSeries_smp_nb);
 478 #endif
 479
 480         /* Mark threads which are still spinning in hold loops. */
 481         if (cpu_has_feature(CPU_FTR_SMT)) {
 482                 for_each_present_cpu(i) {
 483                         if (i % 2 == 0)
 484                                 /*
 485                                  * Even-numbered logical cpus correspond to
 486                                  * primary threads.
 487                                  */
 488                                 cpu_set(i, of_spin_map);
 489                 }
 490         } else {
 491                 of_spin_map = cpu_present_map;
 492         }
 493
 494         cpu_clear(boot_cpuid, of_spin_map);
 495
 496         /* Non-lpar has additional take/give timebase */
 497         if (rtas_token("freeze-time-base") != RTAS_UNKNOWN_SERVICE) {
 498                 smp_ops->give_timebase = pSeries_give_timebase;
 499                 smp_ops->take_timebase = pSeries_take_timebase;
 500         }
 501
 502         DBG(" <- smp_init_pSeries()\n");
 503 }
 504