arch/x86/kernel/setup_percpu.c

   1 #include <linux/kernel.h>
   2 #include <linux/module.h>
   3 #include <linux/init.h>
   4 #include <linux/bootmem.h>
   5 #include <linux/percpu.h>
   6 #include <linux/kexec.h>
   7 #include <linux/crash_dump.h>
   8 #include <linux/smp.h>
   9 #include <linux/topology.h>
  10 #include <asm/sections.h>
  11 #include <asm/processor.h>
  12 #include <asm/setup.h>
  13 #include <asm/mpspec.h>
  14 #include <asm/apicdef.h>
  15 #include <asm/highmem.h>
  16 #include <asm/proto.h>
  17 #include <asm/cpumask.h>
  18
  19 #ifdef CONFIG_DEBUG_PER_CPU_MAPS
  20 # define DBG(x...) printk(KERN_DEBUG x)
  21 #else
  22 # define DBG(x...)
  23 #endif
  24
  25 /*
  26  * Could be inside CONFIG_HAVE_SETUP_PER_CPU_AREA with other stuff but
  27  * voyager wants cpu_number too.
  28  */
  29 #ifdef CONFIG_SMP
  30 DEFINE_PER_CPU(int, cpu_number);
  31 EXPORT_PER_CPU_SYMBOL(cpu_number);
  32 #endif
  33
  34 #ifdef CONFIG_X86_LOCAL_APIC
  35 unsigned int num_processors;
  36 unsigned disabled_cpus __cpuinitdata;
  37 /* Processor that is doing the boot up */
  38 unsigned int boot_cpu_physical_apicid = -1U;
  39 EXPORT_SYMBOL(boot_cpu_physical_apicid);
  40 unsigned int max_physical_apicid;
  41
  42 /* Bitmask of physically existing CPUs */
  43 physid_mask_t phys_cpu_present_map;
  44 #endif
  45
  46 /*
  47  * Map cpu index to physical APIC ID
  48  */
  49 DEFINE_EARLY_PER_CPU(u16, x86_cpu_to_apicid, BAD_APICID);
  50 DEFINE_EARLY_PER_CPU(u16, x86_bios_cpu_apicid, BAD_APICID);
  51 EXPORT_EARLY_PER_CPU_SYMBOL(x86_cpu_to_apicid);
  52 EXPORT_EARLY_PER_CPU_SYMBOL(x86_bios_cpu_apicid);
  53
  54 #if defined(CONFIG_NUMA) && defined(CONFIG_X86_64)
  55 #define X86_64_NUMA     1       /* (used later) */
  56 DEFINE_PER_CPU(int, node_number) = 0;
  57 EXPORT_PER_CPU_SYMBOL(node_number);
  58
  59 /*
  60  * Map cpu index to node index
  61  */
  62 DEFINE_EARLY_PER_CPU(int, x86_cpu_to_node_map, NUMA_NO_NODE);
  63 EXPORT_EARLY_PER_CPU_SYMBOL(x86_cpu_to_node_map);
  64
  65 /*
  66  * Which logical CPUs are on which nodes
  67  */
  68 cpumask_t *node_to_cpumask_map;
  69 EXPORT_SYMBOL(node_to_cpumask_map);
  70
  71 /*
  72  * Setup node_to_cpumask_map
  73  */
  74 static void __init setup_node_to_cpumask_map(void);
  75
  76 #else
  77 static inline void setup_node_to_cpumask_map(void) { }
  78 #endif
  79
  80 /*
  81  * Define load_pda_offset() and per-cpu __pda for x86_64.
  82  * load_pda_offset() is responsible for loading the offset of pda into
  83  * %gs.
  84  *
  85  * On SMP, pda offset also duals as percpu base address and thus it
  86  * should be at the start of per-cpu area.  To achieve this, it's
  87  * preallocated in vmlinux_64.lds.S directly instead of using
  88  * DEFINE_PER_CPU().
  89  */
  90 #ifdef CONFIG_X86_64
  91 void __cpuinit load_pda_offset(int cpu)
  92 {
  93         /* Memory clobbers used to order pda/percpu accesses */
  94         mb();
  95         wrmsrl(MSR_GS_BASE, cpu_pda(cpu));
  96         mb();
  97 }
  98 #ifndef CONFIG_SMP
  99 DEFINE_PER_CPU(struct x8664_pda, __pda);
 100 #endif
 101 EXPORT_PER_CPU_SYMBOL(__pda);
 102 #endif /* CONFIG_SMP && CONFIG_X86_64 */
 103
 104 #ifdef CONFIG_X86_64
 105
 106 /* correctly size the local cpu masks */
 107 static void setup_cpu_local_masks(void)
 108 {
 109         alloc_bootmem_cpumask_var(&cpu_initialized_mask);
 110         alloc_bootmem_cpumask_var(&cpu_callin_mask);
 111         alloc_bootmem_cpumask_var(&cpu_callout_mask);
 112         alloc_bootmem_cpumask_var(&cpu_sibling_setup_mask);
 113 }
 114
 115 #else /* CONFIG_X86_32 */
 116
 117 static inline void setup_cpu_local_masks(void)
 118 {
 119 }
 120
 121 #endif /* CONFIG_X86_32 */
 122
 123 #ifdef CONFIG_HAVE_SETUP_PER_CPU_AREA
 124 /*
 125  * Copy data used in early init routines from the initial arrays to the
 126  * per cpu data areas.  These arrays then become expendable and the
 127  * *_early_ptr's are zeroed indicating that the static arrays are gone.
 128  */
 129 static void __init setup_per_cpu_maps(void)
 130 {
 131         int cpu;
 132
 133         for_each_possible_cpu(cpu) {
 134                 per_cpu(x86_cpu_to_apicid, cpu) =
 135                                 early_per_cpu_map(x86_cpu_to_apicid, cpu);
 136                 per_cpu(x86_bios_cpu_apicid, cpu) =
 137                                 early_per_cpu_map(x86_bios_cpu_apicid, cpu);
 138 #ifdef X86_64_NUMA
 139                 per_cpu(x86_cpu_to_node_map, cpu) =
 140                                 early_per_cpu_map(x86_cpu_to_node_map, cpu);
 141 #endif
 142         }
 143
 144         /* indicate the early static arrays will soon be gone */
 145         early_per_cpu_ptr(x86_cpu_to_apicid) = NULL;
 146         early_per_cpu_ptr(x86_bios_cpu_apicid) = NULL;
 147 #ifdef X86_64_NUMA
 148         early_per_cpu_ptr(x86_cpu_to_node_map) = NULL;
 149 #endif
 150 }
 151
 152 #ifdef CONFIG_X86_64
 153 unsigned long __per_cpu_offset[NR_CPUS] __read_mostly = {
 154         [0] = (unsigned long)__per_cpu_load,
 155 };
 156 #else
 157 unsigned long __per_cpu_offset[NR_CPUS] __read_mostly;
 158 #endif
 159 EXPORT_SYMBOL(__per_cpu_offset);
 160
 161 /*
 162  * Great future plan:
 163  * Declare PDA itself and support (irqstack,tss,pgd) as per cpu data.
 164  * Always point %gs to its beginning
 165  */
 166 void __init setup_per_cpu_areas(void)
 167 {
 168         ssize_t size, old_size;
 169         char *ptr;
 170         int cpu;
 171         unsigned long align = 1;
 172
 173         /* Copy section for each CPU (we discard the original) */
 174         old_size = PERCPU_ENOUGH_ROOM;
 175         align = max_t(unsigned long, PAGE_SIZE, align);
 176         size = roundup(old_size, align);
 177
 178         pr_info("NR_CPUS:%d nr_cpumask_bits:%d nr_cpu_ids:%d nr_node_ids:%d\n",
 179                 NR_CPUS, nr_cpumask_bits, nr_cpu_ids, nr_node_ids);
 180
 181         pr_info("PERCPU: Allocating %zd bytes of per cpu data\n", size);
 182
 183         for_each_possible_cpu(cpu) {
 184 #ifndef CONFIG_NEED_MULTIPLE_NODES
 185                 ptr = __alloc_bootmem(size, align,
 186                                  __pa(MAX_DMA_ADDRESS));
 187 #else
 188                 int node = early_cpu_to_node(cpu);
 189                 if (!node_online(node) || !NODE_DATA(node)) {
 190                         ptr = __alloc_bootmem(size, align,
 191                                          __pa(MAX_DMA_ADDRESS));
 192                         pr_info("cpu %d has no node %d or node-local memory\n",
 193                                 cpu, node);
 194                         pr_debug("per cpu data for cpu%d at %016lx\n",
 195                                  cpu, __pa(ptr));
 196                 } else {
 197                         ptr = __alloc_bootmem_node(NODE_DATA(node), size, align,
 198                                                         __pa(MAX_DMA_ADDRESS));
 199                         pr_debug("per cpu data for cpu%d on node%d at %016lx\n",
 200                                 cpu, node, __pa(ptr));
 201                 }
 202 #endif
 203
 204                 memcpy(ptr, __per_cpu_load, __per_cpu_end - __per_cpu_start);
 205                 per_cpu_offset(cpu) = ptr - __per_cpu_start;
 206                 per_cpu(this_cpu_off, cpu) = per_cpu_offset(cpu);
 207                 per_cpu(cpu_number, cpu) = cpu;
 208 #ifdef CONFIG_X86_64
 209                 per_cpu(irq_stack_ptr, cpu) =
 210                         (char *)per_cpu(irq_stack, cpu) + IRQ_STACK_SIZE - 64;
 211                 /*
 212                  * CPU0 modified pda in the init data area, reload pda
 213                  * offset for CPU0 and clear the area for others.
 214                  */
 215                 if (cpu == 0)
 216                         load_pda_offset(0);
 217                 else
 218                         memset(cpu_pda(cpu), 0, sizeof(*cpu_pda(cpu)));
 219 #endif
 220
 221                 DBG("PERCPU: cpu %4d %p\n", cpu, ptr);
 222         }
 223
 224         /* Setup percpu data maps */
 225         setup_per_cpu_maps();
 226
 227         /* Setup node to cpumask map */
 228         setup_node_to_cpumask_map();
 229
 230         /* Setup cpu initialized, callin, callout masks */
 231         setup_cpu_local_masks();
 232 }
 233
 234 #endif
 235
 236 #ifdef X86_64_NUMA
 237
 238 /*
 239  * Allocate node_to_cpumask_map based on number of available nodes
 240  * Requires node_possible_map to be valid.
 241  *
 242  * Note: node_to_cpumask() is not valid until after this is done.
 243  * (Use CONFIG_DEBUG_PER_CPU_MAPS to check this.)
 244  */
 245 static void __init setup_node_to_cpumask_map(void)
 246 {
 247         unsigned int node, num = 0;
 248         cpumask_t *map;
 249
 250         /* setup nr_node_ids if not done yet */
 251         if (nr_node_ids == MAX_NUMNODES) {
 252                 for_each_node_mask(node, node_possible_map)
 253                         num = node;
 254                 nr_node_ids = num + 1;
 255         }
 256
 257         /* allocate the map */
 258         map = alloc_bootmem_low(nr_node_ids * sizeof(cpumask_t));
 259         DBG("node_to_cpumask_map at %p for %d nodes\n", map, nr_node_ids);
 260
 261         pr_debug("Node to cpumask map at %p for %d nodes\n",
 262                  map, nr_node_ids);
 263
 264         /* node_to_cpumask() will now work */
 265         node_to_cpumask_map = map;
 266 }
 267
 268 void __cpuinit numa_set_node(int cpu, int node)
 269 {
 270         int *cpu_to_node_map = early_per_cpu_ptr(x86_cpu_to_node_map);
 271
 272         /* early setting, no percpu area yet */
 273         if (cpu_to_node_map) {
 274                 cpu_to_node_map[cpu] = node;
 275                 return;
 276         }
 277
 278 #ifdef CONFIG_DEBUG_PER_CPU_MAPS
 279         if (cpu >= nr_cpu_ids || !per_cpu_offset(cpu)) {
 280                 printk(KERN_ERR "numa_set_node: invalid cpu# (%d)\n", cpu);
 281                 dump_stack();
 282                 return;
 283         }
 284 #endif
 285         per_cpu(x86_cpu_to_node_map, cpu) = node;
 286
 287         if (node != NUMA_NO_NODE)
 288                 per_cpu(node_number, cpu) = node;
 289 }
 290
 291 void __cpuinit numa_clear_node(int cpu)
 292 {
 293         numa_set_node(cpu, NUMA_NO_NODE);
 294 }
 295
 296 #ifndef CONFIG_DEBUG_PER_CPU_MAPS
 297
 298 void __cpuinit numa_add_cpu(int cpu)
 299 {
 300         cpu_set(cpu, node_to_cpumask_map[early_cpu_to_node(cpu)]);
 301 }
 302
 303 void __cpuinit numa_remove_cpu(int cpu)
 304 {
 305         cpu_clear(cpu, node_to_cpumask_map[early_cpu_to_node(cpu)]);
 306 }
 307
 308 #else /* CONFIG_DEBUG_PER_CPU_MAPS */
 309
 310 /*
 311  * --------- debug versions of the numa functions ---------
 312  */
 313 static void __cpuinit numa_set_cpumask(int cpu, int enable)
 314 {
 315         int node = early_cpu_to_node(cpu);
 316         cpumask_t *mask;
 317         char buf[64];
 318
 319         if (node_to_cpumask_map == NULL) {
 320                 printk(KERN_ERR "node_to_cpumask_map NULL\n");
 321                 dump_stack();
 322                 return;
 323         }
 324
 325         mask = &node_to_cpumask_map[node];
 326         if (enable)
 327                 cpu_set(cpu, *mask);
 328         else
 329                 cpu_clear(cpu, *mask);
 330
 331         cpulist_scnprintf(buf, sizeof(buf), mask);
 332         printk(KERN_DEBUG "%s cpu %d node %d: mask now %s\n",
 333                 enable ? "numa_add_cpu" : "numa_remove_cpu", cpu, node, buf);
 334 }
 335
 336 void __cpuinit numa_add_cpu(int cpu)
 337 {
 338         numa_set_cpumask(cpu, 1);
 339 }
 340
 341 void __cpuinit numa_remove_cpu(int cpu)
 342 {
 343         numa_set_cpumask(cpu, 0);
 344 }
 345
 346 int cpu_to_node(int cpu)
 347 {
 348         if (early_per_cpu_ptr(x86_cpu_to_node_map)) {
 349                 printk(KERN_WARNING
 350                         "cpu_to_node(%d): usage too early!\n", cpu);
 351                 dump_stack();
 352                 return early_per_cpu_ptr(x86_cpu_to_node_map)[cpu];
 353         }
 354         return per_cpu(x86_cpu_to_node_map, cpu);
 355 }
 356 EXPORT_SYMBOL(cpu_to_node);
 357
 358 /*
 359  * Same function as cpu_to_node() but used if called before the
 360  * per_cpu areas are setup.
 361  */
 362 int early_cpu_to_node(int cpu)
 363 {
 364         if (early_per_cpu_ptr(x86_cpu_to_node_map))
 365                 return early_per_cpu_ptr(x86_cpu_to_node_map)[cpu];
 366
 367         if (!per_cpu_offset(cpu)) {
 368                 printk(KERN_WARNING
 369                         "early_cpu_to_node(%d): no per_cpu area!\n", cpu);
 370                 dump_stack();
 371                 return NUMA_NO_NODE;
 372         }
 373         return per_cpu(x86_cpu_to_node_map, cpu);
 374 }
 375
 376
 377 /* empty cpumask */
 378 static const cpumask_t cpu_mask_none;
 379
 380 /*
 381  * Returns a pointer to the bitmask of CPUs on Node 'node'.
 382  */
 383 const cpumask_t *cpumask_of_node(int node)
 384 {
 385         if (node_to_cpumask_map == NULL) {
 386                 printk(KERN_WARNING
 387                         "cpumask_of_node(%d): no node_to_cpumask_map!\n",
 388                         node);
 389                 dump_stack();
 390                 return (const cpumask_t *)&cpu_online_map;
 391         }
 392         if (node >= nr_node_ids) {
 393                 printk(KERN_WARNING
 394                         "cpumask_of_node(%d): node > nr_node_ids(%d)\n",
 395                         node, nr_node_ids);
 396                 dump_stack();
 397                 return &cpu_mask_none;
 398         }
 399         return &node_to_cpumask_map[node];
 400 }
 401 EXPORT_SYMBOL(cpumask_of_node);
 402
 403 /*
 404  * Returns a bitmask of CPUs on Node 'node'.
 405  *
 406  * Side note: this function creates the returned cpumask on the stack
 407  * so with a high NR_CPUS count, excessive stack space is used.  The
 408  * node_to_cpumask_ptr function should be used whenever possible.
 409  */
 410 cpumask_t node_to_cpumask(int node)
 411 {
 412         if (node_to_cpumask_map == NULL) {
 413                 printk(KERN_WARNING
 414                         "node_to_cpumask(%d): no node_to_cpumask_map!\n", node);
 415                 dump_stack();
 416                 return cpu_online_map;
 417         }
 418         if (node >= nr_node_ids) {
 419                 printk(KERN_WARNING
 420                         "node_to_cpumask(%d): node > nr_node_ids(%d)\n",
 421                         node, nr_node_ids);
 422                 dump_stack();
 423                 return cpu_mask_none;
 424         }
 425         return node_to_cpumask_map[node];
 426 }
 427 EXPORT_SYMBOL(node_to_cpumask);
 428
 429 /*
 430  * --------- end of debug versions of the numa functions ---------
 431  */
 432
 433 #endif /* CONFIG_DEBUG_PER_CPU_MAPS */
 434
 435 #endif /* X86_64_NUMA */
 436