arch/x86/kernel/smpboot.c

   1 #include <linux/init.h>
   2 #include <linux/smp.h>
   3 #include <linux/module.h>
   4 #include <linux/sched.h>
   5 #include <linux/percpu.h>
   6 #include <linux/bootmem.h>
   7
   8 #include <asm/nmi.h>
   9 #include <asm/irq.h>
  10 #include <asm/smp.h>
  11 #include <asm/cpu.h>
  12 #include <asm/numa.h>
  13
  14 /* Number of siblings per CPU package */
  15 int smp_num_siblings = 1;
  16 EXPORT_SYMBOL(smp_num_siblings);
  17
  18 /* Last level cache ID of each logical CPU */
  19 DEFINE_PER_CPU(u16, cpu_llc_id) = BAD_APICID;
  20
  21 /* bitmap of online cpus */
  22 cpumask_t cpu_online_map __read_mostly;
  23 EXPORT_SYMBOL(cpu_online_map);
  24
  25 cpumask_t cpu_callin_map;
  26 cpumask_t cpu_callout_map;
  27 cpumask_t cpu_possible_map;
  28 EXPORT_SYMBOL(cpu_possible_map);
  29
  30 /* representing HT siblings of each logical CPU */
  31 DEFINE_PER_CPU(cpumask_t, cpu_sibling_map);
  32 EXPORT_PER_CPU_SYMBOL(cpu_sibling_map);
  33
  34 /* representing HT and core siblings of each logical CPU */
  35 DEFINE_PER_CPU(cpumask_t, cpu_core_map);
  36 EXPORT_PER_CPU_SYMBOL(cpu_core_map);
  37
  38 /* Per CPU bogomips and other parameters */
  39 DEFINE_PER_CPU_SHARED_ALIGNED(struct cpuinfo_x86, cpu_info);
  40 EXPORT_PER_CPU_SYMBOL(cpu_info);
  41
  42 /* ready for x86_64, no harm for x86, since it will overwrite after alloc */
  43 unsigned char *trampoline_base = __va(SMP_TRAMPOLINE_BASE);
  44
  45 /* representing cpus for which sibling maps can be computed */
  46 static cpumask_t cpu_sibling_setup_map;
  47
  48 #ifdef CONFIG_X86_32
  49 /* Set if we find a B stepping CPU */
  50 int __cpuinitdata smp_b_stepping;
  51 #endif
  52
  53 static void __cpuinit smp_apply_quirks(struct cpuinfo_x86 *c)
  54 {
  55 #ifdef CONFIG_X86_32
  56         /*
  57          * Mask B, Pentium, but not Pentium MMX
  58          */
  59         if (c->x86_vendor == X86_VENDOR_INTEL &&
  60             c->x86 == 5 &&
  61             c->x86_mask >= 1 && c->x86_mask <= 4 &&
  62             c->x86_model <= 3)
  63                 /*
  64                  * Remember we have B step Pentia with bugs
  65                  */
  66                 smp_b_stepping = 1;
  67
  68         /*
  69          * Certain Athlons might work (for various values of 'work') in SMP
  70          * but they are not certified as MP capable.
  71          */
  72         if ((c->x86_vendor == X86_VENDOR_AMD) && (c->x86 == 6)) {
  73
  74                 if (num_possible_cpus() == 1)
  75                         goto valid_k7;
  76
  77                 /* Athlon 660/661 is valid. */
  78                 if ((c->x86_model == 6) && ((c->x86_mask == 0) ||
  79                     (c->x86_mask == 1)))
  80                         goto valid_k7;
  81
  82                 /* Duron 670 is valid */
  83                 if ((c->x86_model == 7) && (c->x86_mask == 0))
  84                         goto valid_k7;
  85
  86                 /*
  87                  * Athlon 662, Duron 671, and Athlon >model 7 have capability
  88                  * bit. It's worth noting that the A5 stepping (662) of some
  89                  * Athlon XP's have the MP bit set.
  90                  * See http://www.heise.de/newsticker/data/jow-18.10.01-000 for
  91                  * more.
  92                  */
  93                 if (((c->x86_model == 6) && (c->x86_mask >= 2)) ||
  94                     ((c->x86_model == 7) && (c->x86_mask >= 1)) ||
  95                      (c->x86_model > 7))
  96                         if (cpu_has_mp)
  97                                 goto valid_k7;
  98
  99                 /* If we get here, not a certified SMP capable AMD system. */
 100                 add_taint(TAINT_UNSAFE_SMP);
 101         }
 102
 103 valid_k7:
 104         ;
 105 #endif
 106 }
 107
 108 /*
 109  * The bootstrap kernel entry code has set these up. Save them for
 110  * a given CPU
 111  */
 112
 113 void __cpuinit smp_store_cpu_info(int id)
 114 {
 115         struct cpuinfo_x86 *c = &cpu_data(id);
 116
 117         *c = boot_cpu_data;
 118         c->cpu_index = id;
 119         if (id != 0)
 120                 identify_secondary_cpu(c);
 121         smp_apply_quirks(c);
 122 }
 123
 124
 125 void __cpuinit set_cpu_sibling_map(int cpu)
 126 {
 127         int i;
 128         struct cpuinfo_x86 *c = &cpu_data(cpu);
 129
 130         cpu_set(cpu, cpu_sibling_setup_map);
 131
 132         if (smp_num_siblings > 1) {
 133                 for_each_cpu_mask(i, cpu_sibling_setup_map) {
 134                         if (c->phys_proc_id == cpu_data(i).phys_proc_id &&
 135                             c->cpu_core_id == cpu_data(i).cpu_core_id) {
 136                                 cpu_set(i, per_cpu(cpu_sibling_map, cpu));
 137                                 cpu_set(cpu, per_cpu(cpu_sibling_map, i));
 138                                 cpu_set(i, per_cpu(cpu_core_map, cpu));
 139                                 cpu_set(cpu, per_cpu(cpu_core_map, i));
 140                                 cpu_set(i, c->llc_shared_map);
 141                                 cpu_set(cpu, cpu_data(i).llc_shared_map);
 142                         }
 143                 }
 144         } else {
 145                 cpu_set(cpu, per_cpu(cpu_sibling_map, cpu));
 146         }
 147
 148         cpu_set(cpu, c->llc_shared_map);
 149
 150         if (current_cpu_data.x86_max_cores == 1) {
 151                 per_cpu(cpu_core_map, cpu) = per_cpu(cpu_sibling_map, cpu);
 152                 c->booted_cores = 1;
 153                 return;
 154         }
 155
 156         for_each_cpu_mask(i, cpu_sibling_setup_map) {
 157                 if (per_cpu(cpu_llc_id, cpu) != BAD_APICID &&
 158                     per_cpu(cpu_llc_id, cpu) == per_cpu(cpu_llc_id, i)) {
 159                         cpu_set(i, c->llc_shared_map);
 160                         cpu_set(cpu, cpu_data(i).llc_shared_map);
 161                 }
 162                 if (c->phys_proc_id == cpu_data(i).phys_proc_id) {
 163                         cpu_set(i, per_cpu(cpu_core_map, cpu));
 164                         cpu_set(cpu, per_cpu(cpu_core_map, i));
 165                         /*
 166                          *  Does this new cpu bringup a new core?
 167                          */
 168                         if (cpus_weight(per_cpu(cpu_sibling_map, cpu)) == 1) {
 169                                 /*
 170                                  * for each core in package, increment
 171                                  * the booted_cores for this new cpu
 172                                  */
 173                                 if (first_cpu(per_cpu(cpu_sibling_map, i)) == i)
 174                                         c->booted_cores++;
 175                                 /*
 176                                  * increment the core count for all
 177                                  * the other cpus in this package
 178                                  */
 179                                 if (i != cpu)
 180                                         cpu_data(i).booted_cores++;
 181                         } else if (i != cpu && !c->booted_cores)
 182                                 c->booted_cores = cpu_data(i).booted_cores;
 183                 }
 184         }
 185 }
 186
 187 /* maps the cpu to the sched domain representing multi-core */
 188 cpumask_t cpu_coregroup_map(int cpu)
 189 {
 190         struct cpuinfo_x86 *c = &cpu_data(cpu);
 191         /*
 192          * For perf, we return last level cache shared map.
 193          * And for power savings, we return cpu_core_map
 194          */
 195         if (sched_mc_power_savings || sched_smt_power_savings)
 196                 return per_cpu(cpu_core_map, cpu);
 197         else
 198                 return c->llc_shared_map;
 199 }
 200
 201 /*
 202  * Currently trivial. Write the real->protected mode
 203  * bootstrap into the page concerned. The caller
 204  * has made sure it's suitably aligned.
 205  */
 206
 207 unsigned long __cpuinit setup_trampoline(void)
 208 {
 209         memcpy(trampoline_base, trampoline_data,
 210                trampoline_end - trampoline_data);
 211         return virt_to_phys(trampoline_base);
 212 }
 213
 214 #ifdef CONFIG_X86_32
 215 /*
 216  * We are called very early to get the low memory for the
 217  * SMP bootup trampoline page.
 218  */
 219 void __init smp_alloc_memory(void)
 220 {
 221         trampoline_base = alloc_bootmem_low_pages(PAGE_SIZE);
 222         /*
 223          * Has to be in very low memory so we can execute
 224          * real-mode AP code.
 225          */
 226         if (__pa(trampoline_base) >= 0x9F000)
 227                 BUG();
 228 }
 229 #endif
 230
 231 #ifdef CONFIG_HOTPLUG_CPU
 232 void remove_siblinginfo(int cpu)
 233 {
 234         int sibling;
 235         struct cpuinfo_x86 *c = &cpu_data(cpu);
 236
 237         for_each_cpu_mask(sibling, per_cpu(cpu_core_map, cpu)) {
 238                 cpu_clear(cpu, per_cpu(cpu_core_map, sibling));
 239                 /*/
 240                  * last thread sibling in this cpu core going down
 241                  */
 242                 if (cpus_weight(per_cpu(cpu_sibling_map, cpu)) == 1)
 243                         cpu_data(sibling).booted_cores--;
 244         }
 245
 246         for_each_cpu_mask(sibling, per_cpu(cpu_sibling_map, cpu))
 247                 cpu_clear(cpu, per_cpu(cpu_sibling_map, sibling));
 248         cpus_clear(per_cpu(cpu_sibling_map, cpu));
 249         cpus_clear(per_cpu(cpu_core_map, cpu));
 250         c->phys_proc_id = 0;
 251         c->cpu_core_id = 0;
 252         cpu_clear(cpu, cpu_sibling_setup_map);
 253 }
 254
 255 int additional_cpus __initdata = -1;
 256
 257 static __init int setup_additional_cpus(char *s)
 258 {
 259         return s && get_option(&s, &additional_cpus) ? 0 : -EINVAL;
 260 }
 261 early_param("additional_cpus", setup_additional_cpus);
 262
 263 /*
 264  * cpu_possible_map should be static, it cannot change as cpu's
 265  * are onlined, or offlined. The reason is per-cpu data-structures
 266  * are allocated by some modules at init time, and dont expect to
 267  * do this dynamically on cpu arrival/departure.
 268  * cpu_present_map on the other hand can change dynamically.
 269  * In case when cpu_hotplug is not compiled, then we resort to current
 270  * behaviour, which is cpu_possible == cpu_present.
 271  * - Ashok Raj
 272  *
 273  * Three ways to find out the number of additional hotplug CPUs:
 274  * - If the BIOS specified disabled CPUs in ACPI/mptables use that.
 275  * - The user can overwrite it with additional_cpus=NUM
 276  * - Otherwise don't reserve additional CPUs.
 277  * We do this because additional CPUs waste a lot of memory.
 278  * -AK
 279  */
 280 __init void prefill_possible_map(void)
 281 {
 282         int i;
 283         int possible;
 284
 285         if (additional_cpus == -1) {
 286                 if (disabled_cpus > 0)
 287                         additional_cpus = disabled_cpus;
 288                 else
 289                         additional_cpus = 0;
 290         }
 291         possible = num_processors + additional_cpus;
 292         if (possible > NR_CPUS)
 293                 possible = NR_CPUS;
 294
 295         printk(KERN_INFO "SMP: Allowing %d CPUs, %d hotplug CPUs\n",
 296                 possible, max_t(int, possible - num_processors, 0));
 297
 298         for (i = 0; i < possible; i++)
 299                 cpu_set(i, cpu_possible_map);
 300 }
 301
 302 static void __ref remove_cpu_from_maps(int cpu)
 303 {
 304         cpu_clear(cpu, cpu_online_map);
 305 #ifdef CONFIG_X86_64
 306         cpu_clear(cpu, cpu_callout_map);
 307         cpu_clear(cpu, cpu_callin_map);
 308         /* was set by cpu_init() */
 309         clear_bit(cpu, (unsigned long *)&cpu_initialized);
 310         clear_node_cpumask(cpu);
 311 #endif
 312 }
 313
 314 int __cpu_disable(void)
 315 {
 316         int cpu = smp_processor_id();
 317
 318         /*
 319          * Perhaps use cpufreq to drop frequency, but that could go
 320          * into generic code.
 321          *
 322          * We won't take down the boot processor on i386 due to some
 323          * interrupts only being able to be serviced by the BSP.
 324          * Especially so if we're not using an IOAPIC   -zwane
 325          */
 326         if (cpu == 0)
 327                 return -EBUSY;
 328
 329         if (nmi_watchdog == NMI_LOCAL_APIC)
 330                 stop_apic_nmi_watchdog(NULL);
 331         clear_local_APIC();
 332
 333         /*
 334          * HACK:
 335          * Allow any queued timer interrupts to get serviced
 336          * This is only a temporary solution until we cleanup
 337          * fixup_irqs as we do for IA64.
 338          */
 339         local_irq_enable();
 340         mdelay(1);
 341
 342         local_irq_disable();
 343         remove_siblinginfo(cpu);
 344
 345         /* It's now safe to remove this processor from the online map */
 346         remove_cpu_from_maps(cpu);
 347         fixup_irqs(cpu_online_map);
 348         return 0;
 349 }
 350
 351 void __cpu_die(unsigned int cpu)
 352 {
 353         /* We don't do anything here: idle task is faking death itself. */
 354         unsigned int i;
 355
 356         for (i = 0; i < 10; i++) {
 357                 /* They ack this in play_dead by setting CPU_DEAD */
 358                 if (per_cpu(cpu_state, cpu) == CPU_DEAD) {
 359                         printk(KERN_INFO "CPU %d is now offline\n", cpu);
 360                         if (1 == num_online_cpus())
 361                                 alternatives_smp_switch(0);
 362                         return;
 363                 }
 364                 msleep(100);
 365         }
 366         printk(KERN_ERR "CPU %u didn't die...\n", cpu);
 367 }
 368 #else /* ... !CONFIG_HOTPLUG_CPU */
 369 int __cpu_disable(void)
 370 {
 371         return -ENOSYS;
 372 }
 373
 374 void __cpu_die(unsigned int cpu)
 375 {
 376         /* We said "no" in __cpu_disable */
 377         BUG();
 378 }
 379 #endif
 380
 381 /*
 382  * If the BIOS enumerates physical processors before logical,
 383  * maxcpus=N at enumeration-time can be used to disable HT.
 384  */
 385 static int __init parse_maxcpus(char *arg)
 386 {
 387         extern unsigned int maxcpus;
 388
 389         maxcpus = simple_strtoul(arg, NULL, 0);
 390         return 0;
 391 }
 392 early_param("maxcpus", parse_maxcpus);