arch/sparc/kernel/leon_smp.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /* leon_smp.c: Sparc-Leon SMP support.
   3  *
   4  * based on sun4m_smp.c
   5  * Copyright (C) 1996 David S. Miller (davem@caip.rutgers.edu)
   6  * Copyright (C) 2009 Daniel Hellstrom (daniel@gaisler.com) Aeroflex Gaisler AB
   7  * Copyright (C) 2009 Konrad Eisele (konrad@gaisler.com) Aeroflex Gaisler AB
   8  */
   9
  10 #include <asm/head.h>
  11
  12 #include <linux/kernel.h>
  13 #include <linux/sched/mm.h>
  14 #include <linux/threads.h>
  15 #include <linux/smp.h>
  16 #include <linux/interrupt.h>
  17 #include <linux/kernel_stat.h>
  18 #include <linux/of.h>
  19 #include <linux/init.h>
  20 #include <linux/spinlock.h>
  21 #include <linux/mm.h>
  22 #include <linux/swap.h>
  23 #include <linux/profile.h>
  24 #include <linux/pm.h>
  25 #include <linux/delay.h>
  26 #include <linux/gfp.h>
  27 #include <linux/cpu.h>
  28 #include <linux/clockchips.h>
  29
  30 #include <asm/cacheflush.h>
  31 #include <asm/tlbflush.h>
  32
  33 #include <asm/ptrace.h>
  34 #include <linux/atomic.h>
  35 #include <asm/irq_regs.h>
  36 #include <asm/traps.h>
  37
  38 #include <asm/delay.h>
  39 #include <asm/irq.h>
  40 #include <asm/page.h>
  41 #include <asm/pgalloc.h>
  42 #include <asm/pgtable.h>
  43 #include <asm/oplib.h>
  44 #include <asm/cpudata.h>
  45 #include <asm/asi.h>
  46 #include <asm/leon.h>
  47 #include <asm/leon_amba.h>
  48 #include <asm/timer.h>
  49
  50 #include "kernel.h"
  51
  52 #include "irq.h"
  53
  54 extern ctxd_t *srmmu_ctx_table_phys;
  55 static int smp_processors_ready;
  56 extern volatile unsigned long cpu_callin_map[NR_CPUS];
  57 extern cpumask_t smp_commenced_mask;
  58 void leon_configure_cache_smp(void);
  59 static void leon_ipi_init(void);
  60
  61 /* IRQ number of LEON IPIs */
  62 int leon_ipi_irq = LEON3_IRQ_IPI_DEFAULT;
  63
  64 static inline unsigned long do_swap(volatile unsigned long *ptr,
  65                                     unsigned long val)
  66 {
  67         __asm__ __volatile__("swapa [%2] %3, %0\n\t" : "=&r"(val)
  68                              : "0"(val), "r"(ptr), "i"(ASI_LEON_DCACHE_MISS)
  69                              : "memory");
  70         return val;
  71 }
  72
  73 void leon_cpu_pre_starting(void *arg)
  74 {
  75         leon_configure_cache_smp();
  76 }
  77
  78 void leon_cpu_pre_online(void *arg)
  79 {
  80         int cpuid = hard_smp_processor_id();
  81
  82         /* Allow master to continue. The master will then give us the
  83          * go-ahead by setting the smp_commenced_mask and will wait without
  84          * timeouts until our setup is completed fully (signified by
  85          * our bit being set in the cpu_online_mask).
  86          */
  87         do_swap(&cpu_callin_map[cpuid], 1);
  88
  89         local_ops->cache_all();
  90         local_ops->tlb_all();
  91
  92         /* Fix idle thread fields. */
  93         __asm__ __volatile__("ld [%0], %%g6\n\t" : : "r"(&current_set[cpuid])
  94                              : "memory" /* paranoid */);
  95
  96         /* Attach to the address space of init_task. */
  97         mmgrab(&init_mm);
  98         current->active_mm = &init_mm;
  99
 100         while (!cpumask_test_cpu(cpuid, &smp_commenced_mask))
 101                 mb();
 102 }
 103
 104 /*
 105  *      Cycle through the processors asking the PROM to start each one.
 106  */
 107
 108 extern struct linux_prom_registers smp_penguin_ctable;
 109
 110 void leon_configure_cache_smp(void)
 111 {
 112         unsigned long cfg = sparc_leon3_get_dcachecfg();
 113         int me = smp_processor_id();
 114
 115         if (ASI_LEON3_SYSCTRL_CFG_SSIZE(cfg) > 4) {
 116                 printk(KERN_INFO "Note: SMP with snooping only works on 4k cache, found %dk(0x%x) on cpu %d, disabling caches\n",
 117                      (unsigned int)ASI_LEON3_SYSCTRL_CFG_SSIZE(cfg),
 118                      (unsigned int)cfg, (unsigned int)me);
 119                 sparc_leon3_disable_cache();
 120         } else {
 121                 if (cfg & ASI_LEON3_SYSCTRL_CFG_SNOOPING) {
 122                         sparc_leon3_enable_snooping();
 123                 } else {
 124                         printk(KERN_INFO "Note: You have to enable snooping in the vhdl model cpu %d, disabling caches\n",
 125                              me);
 126                         sparc_leon3_disable_cache();
 127                 }
 128         }
 129
 130         local_ops->cache_all();
 131         local_ops->tlb_all();
 132 }
 133
 134 static void leon_smp_setbroadcast(unsigned int mask)
 135 {
 136         int broadcast =
 137             ((LEON3_BYPASS_LOAD_PA(&(leon3_irqctrl_regs->mpstatus)) >>
 138               LEON3_IRQMPSTATUS_BROADCAST) & 1);
 139         if (!broadcast) {
 140                 prom_printf("######## !!!! The irqmp-ctrl must have broadcast enabled, smp wont work !!!!! ####### nr cpus: %d\n",
 141                      leon_smp_nrcpus());
 142                 if (leon_smp_nrcpus() > 1) {
 143                         BUG();
 144                 } else {
 145                         prom_printf("continue anyway\n");
 146                         return;
 147                 }
 148         }
 149         LEON_BYPASS_STORE_PA(&(leon3_irqctrl_regs->mpbroadcast), mask);
 150 }
 151
 152 int leon_smp_nrcpus(void)
 153 {
 154         int nrcpu =
 155             ((LEON3_BYPASS_LOAD_PA(&(leon3_irqctrl_regs->mpstatus)) >>
 156               LEON3_IRQMPSTATUS_CPUNR) & 0xf) + 1;
 157         return nrcpu;
 158 }
 159
 160 void __init leon_boot_cpus(void)
 161 {
 162         int nrcpu = leon_smp_nrcpus();
 163         int me = smp_processor_id();
 164
 165         /* Setup IPI */
 166         leon_ipi_init();
 167
 168         printk(KERN_INFO "%d:(%d:%d) cpus mpirq at 0x%x\n", (unsigned int)me,
 169                (unsigned int)nrcpu, (unsigned int)NR_CPUS,
 170                (unsigned int)&(leon3_irqctrl_regs->mpstatus));
 171
 172         leon_enable_irq_cpu(LEON3_IRQ_CROSS_CALL, me);
 173         leon_enable_irq_cpu(LEON3_IRQ_TICKER, me);
 174         leon_enable_irq_cpu(leon_ipi_irq, me);
 175
 176         leon_smp_setbroadcast(1 << LEON3_IRQ_TICKER);
 177
 178         leon_configure_cache_smp();
 179         local_ops->cache_all();
 180
 181 }
 182
 183 int leon_boot_one_cpu(int i, struct task_struct *idle)
 184 {
 185         int timeout;
 186
 187         current_set[i] = task_thread_info(idle);
 188
 189         /* See trampoline.S:leon_smp_cpu_startup for details...
 190          * Initialize the contexts table
 191          * Since the call to prom_startcpu() trashes the structure,
 192          * we need to re-initialize it for each cpu
 193          */
 194         smp_penguin_ctable.which_io = 0;
 195         smp_penguin_ctable.phys_addr = (unsigned int)srmmu_ctx_table_phys;
 196         smp_penguin_ctable.reg_size = 0;
 197
 198         /* whirrr, whirrr, whirrrrrrrrr... */
 199         printk(KERN_INFO "Starting CPU %d : (irqmp: 0x%x)\n", (unsigned int)i,
 200                (unsigned int)&leon3_irqctrl_regs->mpstatus);
 201         local_ops->cache_all();
 202
 203         /* Make sure all IRQs are of from the start for this new CPU */
 204         LEON_BYPASS_STORE_PA(&leon3_irqctrl_regs->mask[i], 0);
 205
 206         /* Wake one CPU */
 207         LEON_BYPASS_STORE_PA(&(leon3_irqctrl_regs->mpstatus), 1 << i);
 208
 209         /* wheee... it's going... */
 210         for (timeout = 0; timeout < 10000; timeout++) {
 211                 if (cpu_callin_map[i])
 212                         break;
 213                 udelay(200);
 214         }
 215         printk(KERN_INFO "Started CPU %d\n", (unsigned int)i);
 216
 217         if (!(cpu_callin_map[i])) {
 218                 printk(KERN_ERR "Processor %d is stuck.\n", i);
 219                 return -ENODEV;
 220         } else {
 221                 leon_enable_irq_cpu(LEON3_IRQ_CROSS_CALL, i);
 222                 leon_enable_irq_cpu(LEON3_IRQ_TICKER, i);
 223                 leon_enable_irq_cpu(leon_ipi_irq, i);
 224         }
 225
 226         local_ops->cache_all();
 227         return 0;
 228 }
 229
 230 void __init leon_smp_done(void)
 231 {
 232
 233         int i, first;
 234         int *prev;
 235
 236         /* setup cpu list for irq rotation */
 237         first = 0;
 238         prev = &first;
 239         for (i = 0; i < NR_CPUS; i++) {
 240                 if (cpu_online(i)) {
 241                         *prev = i;
 242                         prev = &cpu_data(i).next;
 243                 }
 244         }
 245         *prev = first;
 246         local_ops->cache_all();
 247
 248         /* Free unneeded trap tables */
 249         if (!cpu_present(1)) {
 250                 free_reserved_page(virt_to_page(&trapbase_cpu1));
 251         }
 252         if (!cpu_present(2)) {
 253                 free_reserved_page(virt_to_page(&trapbase_cpu2));
 254         }
 255         if (!cpu_present(3)) {
 256                 free_reserved_page(virt_to_page(&trapbase_cpu3));
 257         }
 258         /* Ok, they are spinning and ready to go. */
 259         smp_processors_ready = 1;
 260
 261 }
 262
 263 struct leon_ipi_work {
 264         int single;
 265         int msk;
 266         int resched;
 267 };
 268
 269 static DEFINE_PER_CPU_SHARED_ALIGNED(struct leon_ipi_work, leon_ipi_work);
 270
 271 /* Initialize IPIs on the LEON, in order to save IRQ resources only one IRQ
 272  * is used for all three types of IPIs.
 273  */
 274 static void __init leon_ipi_init(void)
 275 {
 276         int cpu, len;
 277         struct leon_ipi_work *work;
 278         struct property *pp;
 279         struct device_node *rootnp;
 280         struct tt_entry *trap_table;
 281         unsigned long flags;
 282
 283         /* Find IPI IRQ or stick with default value */
 284         rootnp = of_find_node_by_path("/ambapp0");
 285         if (rootnp) {
 286                 pp = of_find_property(rootnp, "ipi_num", &len);
 287                 if (pp && (*(int *)pp->value))
 288                         leon_ipi_irq = *(int *)pp->value;
 289         }
 290         printk(KERN_INFO "leon: SMP IPIs at IRQ %d\n", leon_ipi_irq);
 291
 292         /* Adjust so that we jump directly to smpleon_ipi */
 293         local_irq_save(flags);
 294         trap_table = &sparc_ttable[SP_TRAP_IRQ1 + (leon_ipi_irq - 1)];
 295         trap_table->inst_three += smpleon_ipi - real_irq_entry;
 296         local_ops->cache_all();
 297         local_irq_restore(flags);
 298
 299         for_each_possible_cpu(cpu) {
 300                 work = &per_cpu(leon_ipi_work, cpu);
 301                 work->single = work->msk = work->resched = 0;
 302         }
 303 }
 304
 305 static void leon_send_ipi(int cpu, int level)
 306 {
 307         unsigned long mask;
 308         mask = leon_get_irqmask(level);
 309         LEON3_BYPASS_STORE_PA(&leon3_irqctrl_regs->force[cpu], mask);
 310 }
 311
 312 static void leon_ipi_single(int cpu)
 313 {
 314         struct leon_ipi_work *work = &per_cpu(leon_ipi_work, cpu);
 315
 316         /* Mark work */
 317         work->single = 1;
 318
 319         /* Generate IRQ on the CPU */
 320         leon_send_ipi(cpu, leon_ipi_irq);
 321 }
 322
 323 static void leon_ipi_mask_one(int cpu)
 324 {
 325         struct leon_ipi_work *work = &per_cpu(leon_ipi_work, cpu);
 326
 327         /* Mark work */
 328         work->msk = 1;
 329
 330         /* Generate IRQ on the CPU */
 331         leon_send_ipi(cpu, leon_ipi_irq);
 332 }
 333
 334 static void leon_ipi_resched(int cpu)
 335 {
 336         struct leon_ipi_work *work = &per_cpu(leon_ipi_work, cpu);
 337
 338         /* Mark work */
 339         work->resched = 1;
 340
 341         /* Generate IRQ on the CPU (any IRQ will cause resched) */
 342         leon_send_ipi(cpu, leon_ipi_irq);
 343 }
 344
 345 void leonsmp_ipi_interrupt(void)
 346 {
 347         struct leon_ipi_work *work = this_cpu_ptr(&leon_ipi_work);
 348
 349         if (work->single) {
 350                 work->single = 0;
 351                 smp_call_function_single_interrupt();
 352         }
 353         if (work->msk) {
 354                 work->msk = 0;
 355                 smp_call_function_interrupt();
 356         }
 357         if (work->resched) {
 358                 work->resched = 0;
 359                 smp_resched_interrupt();
 360         }
 361 }
 362
 363 static struct smp_funcall {
 364         smpfunc_t func;
 365         unsigned long arg1;
 366         unsigned long arg2;
 367         unsigned long arg3;
 368         unsigned long arg4;
 369         unsigned long arg5;
 370         unsigned long processors_in[NR_CPUS];   /* Set when ipi entered. */
 371         unsigned long processors_out[NR_CPUS];  /* Set when ipi exited. */
 372 } ccall_info __attribute__((aligned(8)));
 373
 374 static DEFINE_SPINLOCK(cross_call_lock);
 375
 376 /* Cross calls must be serialized, at least currently. */
 377 static void leon_cross_call(smpfunc_t func, cpumask_t mask, unsigned long arg1,
 378                             unsigned long arg2, unsigned long arg3,
 379                             unsigned long arg4)
 380 {
 381         if (smp_processors_ready) {
 382                 register int high = NR_CPUS - 1;
 383                 unsigned long flags;
 384
 385                 spin_lock_irqsave(&cross_call_lock, flags);
 386
 387                 {
 388                         /* If you make changes here, make sure gcc generates proper code... */
 389                         register smpfunc_t f asm("i0") = func;
 390                         register unsigned long a1 asm("i1") = arg1;
 391                         register unsigned long a2 asm("i2") = arg2;
 392                         register unsigned long a3 asm("i3") = arg3;
 393                         register unsigned long a4 asm("i4") = arg4;
 394                         register unsigned long a5 asm("i5") = 0;
 395
 396                         __asm__ __volatile__("std %0, [%6]\n\t"
 397                                              "std %2, [%6 + 8]\n\t"
 398                                              "std %4, [%6 + 16]\n\t" : :
 399                                              "r"(f), "r"(a1), "r"(a2), "r"(a3),
 400                                              "r"(a4), "r"(a5),
 401                                              "r"(&ccall_info.func));
 402                 }
 403
 404                 /* Init receive/complete mapping, plus fire the IPI's off. */
 405                 {
 406                         register int i;
 407
 408                         cpumask_clear_cpu(smp_processor_id(), &mask);
 409                         cpumask_and(&mask, cpu_online_mask, &mask);
 410                         for (i = 0; i <= high; i++) {
 411                                 if (cpumask_test_cpu(i, &mask)) {
 412                                         ccall_info.processors_in[i] = 0;
 413                                         ccall_info.processors_out[i] = 0;
 414                                         leon_send_ipi(i, LEON3_IRQ_CROSS_CALL);
 415
 416                                 }
 417                         }
 418                 }
 419
 420                 {
 421                         register int i;
 422
 423                         i = 0;
 424                         do {
 425                                 if (!cpumask_test_cpu(i, &mask))
 426                                         continue;
 427
 428                                 while (!ccall_info.processors_in[i])
 429                                         barrier();
 430                         } while (++i <= high);
 431
 432                         i = 0;
 433                         do {
 434                                 if (!cpumask_test_cpu(i, &mask))
 435                                         continue;
 436
 437                                 while (!ccall_info.processors_out[i])
 438                                         barrier();
 439                         } while (++i <= high);
 440                 }
 441
 442                 spin_unlock_irqrestore(&cross_call_lock, flags);
 443         }
 444 }
 445
 446 /* Running cross calls. */
 447 void leon_cross_call_irq(void)
 448 {
 449         int i = smp_processor_id();
 450
 451         ccall_info.processors_in[i] = 1;
 452         ccall_info.func(ccall_info.arg1, ccall_info.arg2, ccall_info.arg3,
 453                         ccall_info.arg4, ccall_info.arg5);
 454         ccall_info.processors_out[i] = 1;
 455 }
 456
 457 static const struct sparc32_ipi_ops leon_ipi_ops = {
 458         .cross_call = leon_cross_call,
 459         .resched    = leon_ipi_resched,
 460         .single     = leon_ipi_single,
 461         .mask_one   = leon_ipi_mask_one,
 462 };
 463
 464 void __init leon_init_smp(void)
 465 {
 466         /* Patch ipi15 trap table */
 467         t_nmi[1] = t_nmi[1] + (linux_trap_ipi15_leon - linux_trap_ipi15_sun4m);
 468
 469         sparc32_ipi_ops = &leon_ipi_ops;
 470 }