arch/powerpc/platforms/85xx/smp.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /*
   3  * Author: Andy Fleming <afleming@freescale.com>
   4  *         Kumar Gala <galak@kernel.crashing.org>
   5  *
   6  * Copyright 2006-2008, 2011-2012, 2015 Freescale Semiconductor Inc.
   7  */
   8
   9 #include <linux/stddef.h>
  10 #include <linux/kernel.h>
  11 #include <linux/sched/hotplug.h>
  12 #include <linux/init.h>
  13 #include <linux/delay.h>
  14 #include <linux/of.h>
  15 #include <linux/kexec.h>
  16 #include <linux/highmem.h>
  17 #include <linux/cpu.h>
  18 #include <linux/fsl/guts.h>
  19 #include <linux/pgtable.h>
  20
  21 #include <asm/machdep.h>
  22 #include <asm/page.h>
  23 #include <asm/mpic.h>
  24 #include <asm/cacheflush.h>
  25 #include <asm/dbell.h>
  26 #include <asm/code-patching.h>
  27 #include <asm/cputhreads.h>
  28 #include <asm/fsl_pm.h>
  29
  30 #include <sysdev/fsl_soc.h>
  31 #include <sysdev/mpic.h>
  32 #include "smp.h"
  33
  34 struct epapr_spin_table {
  35         u32     addr_h;
  36         u32     addr_l;
  37         u32     r3_h;
  38         u32     r3_l;
  39         u32     reserved;
  40         u32     pir;
  41 };
  42
  43 #ifdef CONFIG_HOTPLUG_CPU
  44 static u64 timebase;
  45 static int tb_req;
  46 static int tb_valid;
  47
  48 static void mpc85xx_give_timebase(void)
  49 {
  50         unsigned long flags;
  51
  52         local_irq_save(flags);
  53         hard_irq_disable();
  54
  55         while (!tb_req)
  56                 barrier();
  57         tb_req = 0;
  58
  59         qoriq_pm_ops->freeze_time_base(true);
  60 #ifdef CONFIG_PPC64
  61         /*
  62          * e5500/e6500 have a workaround for erratum A-006958 in place
  63          * that will reread the timebase until TBL is non-zero.
  64          * That would be a bad thing when the timebase is frozen.
  65          *
  66          * Thus, we read it manually, and instead of checking that
  67          * TBL is non-zero, we ensure that TB does not change.  We don't
  68          * do that for the main mftb implementation, because it requires
  69          * a scratch register
  70          */
  71         {
  72                 u64 prev;
  73
  74                 asm volatile("mfspr %0, %1" : "=r" (timebase) :
  75                              "i" (SPRN_TBRL));
  76
  77                 do {
  78                         prev = timebase;
  79                         asm volatile("mfspr %0, %1" : "=r" (timebase) :
  80                                      "i" (SPRN_TBRL));
  81                 } while (prev != timebase);
  82         }
  83 #else
  84         timebase = get_tb();
  85 #endif
  86         mb();
  87         tb_valid = 1;
  88
  89         while (tb_valid)
  90                 barrier();
  91
  92         qoriq_pm_ops->freeze_time_base(false);
  93
  94         local_irq_restore(flags);
  95 }
  96
  97 static void mpc85xx_take_timebase(void)
  98 {
  99         unsigned long flags;
 100
 101         local_irq_save(flags);
 102         hard_irq_disable();
 103
 104         tb_req = 1;
 105         while (!tb_valid)
 106                 barrier();
 107
 108         set_tb(timebase >> 32, timebase & 0xffffffff);
 109         isync();
 110         tb_valid = 0;
 111
 112         local_irq_restore(flags);
 113 }
 114
 115 static void smp_85xx_cpu_offline_self(void)
 116 {
 117         unsigned int cpu = smp_processor_id();
 118
 119         local_irq_disable();
 120         hard_irq_disable();
 121         /* mask all irqs to prevent cpu wakeup */
 122         qoriq_pm_ops->irq_mask(cpu);
 123
 124         idle_task_exit();
 125
 126         mtspr(SPRN_TCR, 0);
 127         mtspr(SPRN_TSR, mfspr(SPRN_TSR));
 128
 129         generic_set_cpu_dead(cpu);
 130
 131         cur_cpu_spec->cpu_down_flush();
 132
 133         qoriq_pm_ops->cpu_die(cpu);
 134
 135         while (1)
 136                 ;
 137 }
 138
 139 static void qoriq_cpu_kill(unsigned int cpu)
 140 {
 141         int i;
 142
 143         for (i = 0; i < 500; i++) {
 144                 if (is_cpu_dead(cpu)) {
 145 #ifdef CONFIG_PPC64
 146                         paca_ptrs[cpu]->cpu_start = 0;
 147 #endif
 148                         return;
 149                 }
 150                 msleep(20);
 151         }
 152         pr_err("CPU%d didn't die...\n", cpu);
 153 }
 154 #endif
 155
 156 /*
 157  * To keep it compatible with old boot program which uses
 158  * cache-inhibit spin table, we need to flush the cache
 159  * before accessing spin table to invalidate any staled data.
 160  * We also need to flush the cache after writing to spin
 161  * table to push data out.
 162  */
 163 static inline void flush_spin_table(void *spin_table)
 164 {
 165         flush_dcache_range((ulong)spin_table,
 166                 (ulong)spin_table + sizeof(struct epapr_spin_table));
 167 }
 168
 169 static inline u32 read_spin_table_addr_l(void *spin_table)
 170 {
 171         flush_dcache_range((ulong)spin_table,
 172                 (ulong)spin_table + sizeof(struct epapr_spin_table));
 173         return in_be32(&((struct epapr_spin_table *)spin_table)->addr_l);
 174 }
 175
 176 #ifdef CONFIG_PPC64
 177 static void wake_hw_thread(void *info)
 178 {
 179         void fsl_secondary_thread_init(void);
 180         unsigned long inia;
 181         int cpu = *(const int *)info;
 182
 183         inia = *(unsigned long *)fsl_secondary_thread_init;
 184         book3e_start_thread(cpu_thread_in_core(cpu), inia);
 185 }
 186 #endif
 187
 188 static int smp_85xx_start_cpu(int cpu)
 189 {
 190         int ret = 0;
 191         struct device_node *np;
 192         const u64 *cpu_rel_addr;
 193         unsigned long flags;
 194         int ioremappable;
 195         int hw_cpu = get_hard_smp_processor_id(cpu);
 196         struct epapr_spin_table __iomem *spin_table;
 197
 198         np = of_get_cpu_node(cpu, NULL);
 199         cpu_rel_addr = of_get_property(np, "cpu-release-addr", NULL);
 200         if (!cpu_rel_addr) {
 201                 pr_err("No cpu-release-addr for cpu %d\n", cpu);
 202                 return -ENOENT;
 203         }
 204
 205         /*
 206          * A secondary core could be in a spinloop in the bootpage
 207          * (0xfffff000), somewhere in highmem, or somewhere in lowmem.
 208          * The bootpage and highmem can be accessed via ioremap(), but
 209          * we need to directly access the spinloop if its in lowmem.
 210          */
 211         ioremappable = *cpu_rel_addr > virt_to_phys(high_memory);
 212
 213         /* Map the spin table */
 214         if (ioremappable)
 215                 spin_table = ioremap_coherent(*cpu_rel_addr,
 216                                               sizeof(struct epapr_spin_table));
 217         else
 218                 spin_table = phys_to_virt(*cpu_rel_addr);
 219
 220         local_irq_save(flags);
 221         hard_irq_disable();
 222
 223         if (qoriq_pm_ops)
 224                 qoriq_pm_ops->cpu_up_prepare(cpu);
 225
 226         /* if cpu is not spinning, reset it */
 227         if (read_spin_table_addr_l(spin_table) != 1) {
 228                 /*
 229                  * We don't set the BPTR register here since it already points
 230                  * to the boot page properly.
 231                  */
 232                 mpic_reset_core(cpu);
 233
 234                 /*
 235                  * wait until core is ready...
 236                  * We need to invalidate the stale data, in case the boot
 237                  * loader uses a cache-inhibited spin table.
 238                  */
 239                 if (!spin_event_timeout(
 240                                 read_spin_table_addr_l(spin_table) == 1,
 241                                 10000, 100)) {
 242                         pr_err("timeout waiting for cpu %d to reset\n",
 243                                 hw_cpu);
 244                         ret = -EAGAIN;
 245                         goto err;
 246                 }
 247         }
 248
 249         flush_spin_table(spin_table);
 250         out_be32(&spin_table->pir, hw_cpu);
 251 #ifdef CONFIG_PPC64
 252         out_be64((u64 *)(&spin_table->addr_h),
 253                 __pa(ppc_function_entry(generic_secondary_smp_init)));
 254 #else
 255 #ifdef CONFIG_PHYS_ADDR_T_64BIT
 256         /*
 257          * We need also to write addr_h to spin table for systems
 258          * in which their physical memory start address was configured
 259          * to above 4G, otherwise the secondary core can not get
 260          * correct entry to start from.
 261          */
 262         out_be32(&spin_table->addr_h, __pa(__early_start) >> 32);
 263 #endif
 264         out_be32(&spin_table->addr_l, __pa(__early_start));
 265 #endif
 266         flush_spin_table(spin_table);
 267 err:
 268         local_irq_restore(flags);
 269
 270         if (ioremappable)
 271                 iounmap(spin_table);
 272
 273         return ret;
 274 }
 275
 276 static int smp_85xx_kick_cpu(int nr)
 277 {
 278         int ret = 0;
 279 #ifdef CONFIG_PPC64
 280         int primary = nr;
 281 #endif
 282
 283         WARN_ON(nr < 0 || nr >= num_possible_cpus());
 284
 285         pr_debug("kick CPU #%d\n", nr);
 286
 287 #ifdef CONFIG_PPC64
 288         if (threads_per_core == 2) {
 289                 if (WARN_ON_ONCE(!cpu_has_feature(CPU_FTR_SMT)))
 290                         return -ENOENT;
 291
 292                 booting_thread_hwid = cpu_thread_in_core(nr);
 293                 primary = cpu_first_thread_sibling(nr);
 294
 295                 if (qoriq_pm_ops)
 296                         qoriq_pm_ops->cpu_up_prepare(nr);
 297
 298                 /*
 299                  * If either thread in the core is online, use it to start
 300                  * the other.
 301                  */
 302                 if (cpu_online(primary)) {
 303                         smp_call_function_single(primary,
 304                                         wake_hw_thread, &nr, 1);
 305                         goto done;
 306                 } else if (cpu_online(primary + 1)) {
 307                         smp_call_function_single(primary + 1,
 308                                         wake_hw_thread, &nr, 1);
 309                         goto done;
 310                 }
 311
 312                 /*
 313                  * If getting here, it means both threads in the core are
 314                  * offline. So start the primary thread, then it will start
 315                  * the thread specified in booting_thread_hwid, the one
 316                  * corresponding to nr.
 317                  */
 318
 319         } else if (threads_per_core == 1) {
 320                 /*
 321                  * If one core has only one thread, set booting_thread_hwid to
 322                  * an invalid value.
 323                  */
 324                 booting_thread_hwid = INVALID_THREAD_HWID;
 325
 326         } else if (threads_per_core > 2) {
 327                 pr_err("Do not support more than 2 threads per CPU.");
 328                 return -EINVAL;
 329         }
 330
 331         ret = smp_85xx_start_cpu(primary);
 332         if (ret)
 333                 return ret;
 334
 335 done:
 336         paca_ptrs[nr]->cpu_start = 1;
 337         generic_set_cpu_up(nr);
 338
 339         return ret;
 340 #else
 341         ret = smp_85xx_start_cpu(nr);
 342         if (ret)
 343                 return ret;
 344
 345         generic_set_cpu_up(nr);
 346
 347         return ret;
 348 #endif
 349 }
 350
 351 struct smp_ops_t smp_85xx_ops = {
 352         .cause_nmi_ipi = NULL,
 353         .kick_cpu = smp_85xx_kick_cpu,
 354         .cpu_bootable = smp_generic_cpu_bootable,
 355 #ifdef CONFIG_HOTPLUG_CPU
 356         .cpu_disable    = generic_cpu_disable,
 357         .cpu_die        = generic_cpu_die,
 358 #endif
 359 #if defined(CONFIG_KEXEC_CORE) && !defined(CONFIG_PPC64)
 360         .give_timebase  = smp_generic_give_timebase,
 361         .take_timebase  = smp_generic_take_timebase,
 362 #endif
 363 };
 364
 365 #ifdef CONFIG_KEXEC_CORE
 366 #ifdef CONFIG_PPC32
 367 atomic_t kexec_down_cpus = ATOMIC_INIT(0);
 368
 369 void mpc85xx_smp_kexec_cpu_down(int crash_shutdown, int secondary)
 370 {
 371         local_irq_disable();
 372
 373         if (secondary) {
 374                 cur_cpu_spec->cpu_down_flush();
 375                 atomic_inc(&kexec_down_cpus);
 376                 /* loop forever */
 377                 while (1);
 378         }
 379 }
 380
 381 static void mpc85xx_smp_kexec_down(void *arg)
 382 {
 383         if (ppc_md.kexec_cpu_down)
 384                 ppc_md.kexec_cpu_down(0,1);
 385 }
 386 #else
 387 void mpc85xx_smp_kexec_cpu_down(int crash_shutdown, int secondary)
 388 {
 389         int cpu = smp_processor_id();
 390         int sibling = cpu_last_thread_sibling(cpu);
 391         bool notified = false;
 392         int disable_cpu;
 393         int disable_threadbit = 0;
 394         long start = mftb();
 395         long now;
 396
 397         local_irq_disable();
 398         hard_irq_disable();
 399         mpic_teardown_this_cpu(secondary);
 400
 401         if (cpu == crashing_cpu && cpu_thread_in_core(cpu) != 0) {
 402                 /*
 403                  * We enter the crash kernel on whatever cpu crashed,
 404                  * even if it's a secondary thread.  If that's the case,
 405                  * disable the corresponding primary thread.
 406                  */
 407                 disable_threadbit = 1;
 408                 disable_cpu = cpu_first_thread_sibling(cpu);
 409         } else if (sibling != crashing_cpu &&
 410                    cpu_thread_in_core(cpu) == 0 &&
 411                    cpu_thread_in_core(sibling) != 0) {
 412                 disable_threadbit = 2;
 413                 disable_cpu = sibling;
 414         }
 415
 416         if (disable_threadbit) {
 417                 while (paca_ptrs[disable_cpu]->kexec_state < KEXEC_STATE_REAL_MODE) {
 418                         barrier();
 419                         now = mftb();
 420                         if (!notified && now - start > 1000000) {
 421                                 pr_info("%s/%d: waiting for cpu %d to enter KEXEC_STATE_REAL_MODE (%d)\n",
 422                                         __func__, smp_processor_id(),
 423                                         disable_cpu,
 424                                         paca_ptrs[disable_cpu]->kexec_state);
 425                                 notified = true;
 426                         }
 427                 }
 428
 429                 if (notified) {
 430                         pr_info("%s: cpu %d done waiting\n",
 431                                 __func__, disable_cpu);
 432                 }
 433
 434                 mtspr(SPRN_TENC, disable_threadbit);
 435                 while (mfspr(SPRN_TENSR) & disable_threadbit)
 436                         cpu_relax();
 437         }
 438 }
 439 #endif
 440
 441 static void mpc85xx_smp_machine_kexec(struct kimage *image)
 442 {
 443 #ifdef CONFIG_PPC32
 444         int timeout = INT_MAX;
 445         int i, num_cpus = num_present_cpus();
 446
 447         if (image->type == KEXEC_TYPE_DEFAULT)
 448                 smp_call_function(mpc85xx_smp_kexec_down, NULL, 0);
 449
 450         while ( (atomic_read(&kexec_down_cpus) != (num_cpus - 1)) &&
 451                 ( timeout > 0 ) )
 452         {
 453                 timeout--;
 454         }
 455
 456         if ( !timeout )
 457                 printk(KERN_ERR "Unable to bring down secondary cpu(s)");
 458
 459         for_each_online_cpu(i)
 460         {
 461                 if ( i == smp_processor_id() ) continue;
 462                 mpic_reset_core(i);
 463         }
 464 #endif
 465
 466         default_machine_kexec(image);
 467 }
 468 #endif /* CONFIG_KEXEC_CORE */
 469
 470 static void smp_85xx_setup_cpu(int cpu_nr)
 471 {
 472         mpic_setup_this_cpu();
 473 }
 474
 475 void __init mpc85xx_smp_init(void)
 476 {
 477         struct device_node *np;
 478
 479
 480         np = of_find_node_by_type(NULL, "open-pic");
 481         if (np) {
 482                 smp_85xx_ops.probe = smp_mpic_probe;
 483                 smp_85xx_ops.setup_cpu = smp_85xx_setup_cpu;
 484                 smp_85xx_ops.message_pass = smp_mpic_message_pass;
 485         } else
 486                 smp_85xx_ops.setup_cpu = NULL;
 487
 488         if (cpu_has_feature(CPU_FTR_DBELL)) {
 489                 /*
 490                  * If left NULL, .message_pass defaults to
 491                  * smp_muxed_ipi_message_pass
 492                  */
 493                 smp_85xx_ops.message_pass = NULL;
 494                 smp_85xx_ops.cause_ipi = doorbell_global_ipi;
 495                 smp_85xx_ops.probe = NULL;
 496         }
 497
 498 #ifdef CONFIG_HOTPLUG_CPU
 499 #ifdef CONFIG_FSL_CORENET_RCPM
 500         fsl_rcpm_init();
 501 #endif
 502
 503 #ifdef CONFIG_FSL_PMC
 504         mpc85xx_setup_pmc();
 505 #endif
 506         if (qoriq_pm_ops) {
 507                 smp_85xx_ops.give_timebase = mpc85xx_give_timebase;
 508                 smp_85xx_ops.take_timebase = mpc85xx_take_timebase;
 509                 smp_85xx_ops.cpu_offline_self = smp_85xx_cpu_offline_self;
 510                 smp_85xx_ops.cpu_die = qoriq_cpu_kill;
 511         }
 512 #endif
 513         smp_ops = &smp_85xx_ops;
 514
 515 #ifdef CONFIG_KEXEC_CORE
 516         ppc_md.kexec_cpu_down = mpc85xx_smp_kexec_cpu_down;
 517         ppc_md.machine_kexec = mpc85xx_smp_machine_kexec;
 518 #endif
 519 }