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 <asm/smp.h>
   9 #include <asm/percpu.h>
  10 #include <asm/sections.h>
  11 #include <asm/processor.h>
  12 #include <asm/setup.h>
  13 #include <asm/topology.h>
  14 #include <asm/mpspec.h>
  15 #include <asm/apicdef.h>
  16 #include <asm/highmem.h>
  17
  18 #ifdef CONFIG_X86_LOCAL_APIC
  19 unsigned int num_processors;
  20 unsigned disabled_cpus __cpuinitdata;
  21 /* Processor that is doing the boot up */
  22 unsigned int boot_cpu_physical_apicid = -1U;
  23 unsigned int max_physical_apicid;
  24 EXPORT_SYMBOL(boot_cpu_physical_apicid);
  25
  26 /* Bitmask of physically existing CPUs */
  27 physid_mask_t phys_cpu_present_map;
  28 #endif
  29
  30 /* map cpu index to physical APIC ID */
  31 DEFINE_EARLY_PER_CPU(u16, x86_cpu_to_apicid, BAD_APICID);
  32 DEFINE_EARLY_PER_CPU(u16, x86_bios_cpu_apicid, BAD_APICID);
  33 EXPORT_EARLY_PER_CPU_SYMBOL(x86_cpu_to_apicid);
  34 EXPORT_EARLY_PER_CPU_SYMBOL(x86_bios_cpu_apicid);
  35
  36 #if defined(CONFIG_NUMA) && defined(CONFIG_X86_64)
  37 #define X86_64_NUMA     1
  38
  39 /* map cpu index to node index */
  40 DEFINE_EARLY_PER_CPU(int, x86_cpu_to_node_map, NUMA_NO_NODE);
  41 EXPORT_EARLY_PER_CPU_SYMBOL(x86_cpu_to_node_map);
  42
  43 /* which logical CPUs are on which nodes */
  44 cpumask_t *node_to_cpumask_map;
  45 EXPORT_SYMBOL(node_to_cpumask_map);
  46
  47 /* setup node_to_cpumask_map */
  48 static void __init setup_node_to_cpumask_map(void);
  49
  50 #else
  51 static inline void setup_node_to_cpumask_map(void) { }
  52 #endif
  53
  54 #if defined(CONFIG_HAVE_SETUP_PER_CPU_AREA) && defined(CONFIG_X86_SMP)
  55 /*
  56  * Copy data used in early init routines from the initial arrays to the
  57  * per cpu data areas.  These arrays then become expendable and the
  58  * *_early_ptr's are zeroed indicating that the static arrays are gone.
  59  */
  60 static void __init setup_per_cpu_maps(void)
  61 {
  62         int cpu;
  63
  64         for_each_possible_cpu(cpu) {
  65                 per_cpu(x86_cpu_to_apicid, cpu) =
  66                                 early_per_cpu_map(x86_cpu_to_apicid, cpu);
  67                 per_cpu(x86_bios_cpu_apicid, cpu) =
  68                                 early_per_cpu_map(x86_bios_cpu_apicid, cpu);
  69 #ifdef X86_64_NUMA
  70                 per_cpu(x86_cpu_to_node_map, cpu) =
  71                                 early_per_cpu_map(x86_cpu_to_node_map, cpu);
  72 #endif
  73         }
  74
  75         /* indicate the early static arrays will soon be gone */
  76         early_per_cpu_ptr(x86_cpu_to_apicid) = NULL;
  77         early_per_cpu_ptr(x86_bios_cpu_apicid) = NULL;
  78 #ifdef X86_64_NUMA
  79         early_per_cpu_ptr(x86_cpu_to_node_map) = NULL;
  80 #endif
  81 }
  82
  83 #ifdef CONFIG_X86_32
  84 /*
  85  * Great future not-so-futuristic plan: make i386 and x86_64 do it
  86  * the same way
  87  */
  88 unsigned long __per_cpu_offset[NR_CPUS] __read_mostly;
  89 EXPORT_SYMBOL(__per_cpu_offset);
  90 static inline void setup_cpu_pda_map(void) { }
  91
  92 #elif !defined(CONFIG_SMP)
  93 static inline void setup_cpu_pda_map(void) { }
  94
  95 #else /* CONFIG_SMP && CONFIG_X86_64 */
  96
  97 /*
  98  * Allocate cpu_pda pointer table and array via alloc_bootmem.
  99  */
 100 static void __init setup_cpu_pda_map(void)
 101 {
 102         char *pda;
 103         struct x8664_pda **new_cpu_pda;
 104         unsigned long size;
 105         int cpu;
 106
 107         size = roundup(sizeof(struct x8664_pda), cache_line_size());
 108
 109         /* allocate cpu_pda array and pointer table */
 110         {
 111                 unsigned long tsize = nr_cpu_ids * sizeof(void *);
 112                 unsigned long asize = size * (nr_cpu_ids - 1);
 113
 114                 tsize = roundup(tsize, cache_line_size());
 115                 new_cpu_pda = alloc_bootmem(tsize + asize);
 116                 pda = (char *)new_cpu_pda + tsize;
 117         }
 118
 119         /* initialize pointer table to static pda's */
 120         for_each_possible_cpu(cpu) {
 121                 if (cpu == 0) {
 122                         /* leave boot cpu pda in place */
 123                         new_cpu_pda[0] = cpu_pda(0);
 124                         continue;
 125                 }
 126                 new_cpu_pda[cpu] = (struct x8664_pda *)pda;
 127                 new_cpu_pda[cpu]->in_bootmem = 1;
 128                 pda += size;
 129         }
 130
 131         /* point to new pointer table */
 132         _cpu_pda = new_cpu_pda;
 133 }
 134 #endif
 135
 136 /*
 137  * Great future plan:
 138  * Declare PDA itself and support (irqstack,tss,pgd) as per cpu data.
 139  * Always point %gs to its beginning
 140  */
 141 void __init setup_per_cpu_areas(void)
 142 {
 143         ssize_t size = PERCPU_ENOUGH_ROOM;
 144         char *ptr;
 145         int cpu;
 146
 147         /* Setup cpu_pda map */
 148         setup_cpu_pda_map();
 149
 150         /* Copy section for each CPU (we discard the original) */
 151         size = PERCPU_ENOUGH_ROOM;
 152         printk(KERN_INFO "PERCPU: Allocating %zd bytes of per cpu data\n",
 153                           size);
 154
 155         for_each_possible_cpu(cpu) {
 156 #ifndef CONFIG_NEED_MULTIPLE_NODES
 157                 ptr = alloc_bootmem_pages(size);
 158 #else
 159                 int node = early_cpu_to_node(cpu);
 160                 if (!node_online(node) || !NODE_DATA(node)) {
 161                         ptr = alloc_bootmem_pages(size);
 162                         printk(KERN_INFO
 163                                "cpu %d has no node %d or node-local memory\n",
 164                                 cpu, node);
 165                         if (ptr)
 166                                 printk(KERN_DEBUG "per cpu data for cpu%d at %016lx\n",
 167                                          cpu, __pa(ptr));
 168                 }
 169                 else {
 170                         ptr = alloc_bootmem_pages_node(NODE_DATA(node), size);
 171                         if (ptr)
 172                                 printk(KERN_DEBUG "per cpu data for cpu%d on node%d at %016lx\n",
 173                                          cpu, node, __pa(ptr));
 174                 }
 175 #endif
 176                 per_cpu_offset(cpu) = ptr - __per_cpu_start;
 177                 memcpy(ptr, __per_cpu_start, __per_cpu_end - __per_cpu_start);
 178
 179         }
 180
 181         printk(KERN_DEBUG "NR_CPUS: %d, nr_cpu_ids: %d, nr_node_ids %d\n",
 182                 NR_CPUS, nr_cpu_ids, nr_node_ids);
 183
 184         /* Setup percpu data maps */
 185         setup_per_cpu_maps();
 186
 187         /* Setup node to cpumask map */
 188         setup_node_to_cpumask_map();
 189 }
 190
 191 #endif
 192
 193 #ifdef X86_64_NUMA
 194
 195 /*
 196  * Allocate node_to_cpumask_map based on number of available nodes
 197  * Requires node_possible_map to be valid.
 198  *
 199  * Note: node_to_cpumask() is not valid until after this is done.
 200  */
 201 static void __init setup_node_to_cpumask_map(void)
 202 {
 203         unsigned int node, num = 0;
 204         cpumask_t *map;
 205
 206         /* setup nr_node_ids if not done yet */
 207         if (nr_node_ids == MAX_NUMNODES) {
 208                 for_each_node_mask(node, node_possible_map)
 209                         num = node;
 210                 nr_node_ids = num + 1;
 211         }
 212
 213         /* allocate the map */
 214         map = alloc_bootmem_low(nr_node_ids * sizeof(cpumask_t));
 215
 216         pr_debug(KERN_DEBUG "Node to cpumask map at %p for %d nodes\n",
 217                  map, nr_node_ids);
 218
 219         /* node_to_cpumask() will now work */
 220         node_to_cpumask_map = map;
 221 }
 222
 223 void __cpuinit numa_set_node(int cpu, int node)
 224 {
 225         int *cpu_to_node_map = early_per_cpu_ptr(x86_cpu_to_node_map);
 226
 227         if (cpu_pda(cpu) && node != NUMA_NO_NODE)
 228                 cpu_pda(cpu)->nodenumber = node;
 229
 230         if (cpu_to_node_map)
 231                 cpu_to_node_map[cpu] = node;
 232
 233         else if (per_cpu_offset(cpu))
 234                 per_cpu(x86_cpu_to_node_map, cpu) = node;
 235
 236         else
 237                 pr_debug("Setting node for non-present cpu %d\n", cpu);
 238 }
 239
 240 void __cpuinit numa_clear_node(int cpu)
 241 {
 242         numa_set_node(cpu, NUMA_NO_NODE);
 243 }
 244
 245 #ifndef CONFIG_DEBUG_PER_CPU_MAPS
 246
 247 void __cpuinit numa_add_cpu(int cpu)
 248 {
 249         cpu_set(cpu, node_to_cpumask_map[early_cpu_to_node(cpu)]);
 250 }
 251
 252 void __cpuinit numa_remove_cpu(int cpu)
 253 {
 254         cpu_clear(cpu, node_to_cpumask_map[cpu_to_node(cpu)]);
 255 }
 256
 257 #else /* CONFIG_DEBUG_PER_CPU_MAPS */
 258
 259 /*
 260  * --------- debug versions of the numa functions ---------
 261  */
 262 static void __cpuinit numa_set_cpumask(int cpu, int enable)
 263 {
 264         int node = cpu_to_node(cpu);
 265         cpumask_t *mask;
 266         char buf[64];
 267
 268         if (node_to_cpumask_map == NULL) {
 269                 printk(KERN_ERR "node_to_cpumask_map NULL\n");
 270                 dump_stack();
 271                 return;
 272         }
 273
 274         mask = &node_to_cpumask_map[node];
 275         if (enable)
 276                 cpu_set(cpu, *mask);
 277         else
 278                 cpu_clear(cpu, *mask);
 279
 280         cpulist_scnprintf(buf, sizeof(buf), *mask);
 281         printk(KERN_DEBUG "%s cpu %d node %d: mask now %s\n",
 282                 enable? "numa_add_cpu":"numa_remove_cpu", cpu, node, buf);
 283  }
 284
 285 void __cpuinit numa_add_cpu(int cpu)
 286 {
 287         numa_set_cpumask(cpu, 1);
 288 }
 289
 290 void __cpuinit numa_remove_cpu(int cpu)
 291 {
 292         numa_set_cpumask(cpu, 0);
 293 }
 294
 295 int cpu_to_node(int cpu)
 296 {
 297         if (early_per_cpu_ptr(x86_cpu_to_node_map)) {
 298                 printk(KERN_WARNING
 299                         "cpu_to_node(%d): usage too early!\n", cpu);
 300                 dump_stack();
 301                 return early_per_cpu_ptr(x86_cpu_to_node_map)[cpu];
 302         }
 303         return per_cpu(x86_cpu_to_node_map, cpu);
 304 }
 305 EXPORT_SYMBOL(cpu_to_node);
 306
 307 /*
 308  * Same function as cpu_to_node() but used if called before the
 309  * per_cpu areas are setup.
 310  */
 311 int early_cpu_to_node(int cpu)
 312 {
 313         if (early_per_cpu_ptr(x86_cpu_to_node_map))
 314                 return early_per_cpu_ptr(x86_cpu_to_node_map)[cpu];
 315
 316         if (!per_cpu_offset(cpu)) {
 317                 printk(KERN_WARNING
 318                         "early_cpu_to_node(%d): no per_cpu area!\n", cpu);
 319                 dump_stack();
 320                 return NUMA_NO_NODE;
 321         }
 322         return per_cpu(x86_cpu_to_node_map, cpu);
 323 }
 324
 325
 326 /* empty cpumask */
 327 static const cpumask_t cpu_mask_none;
 328
 329 /*
 330  * Returns a pointer to the bitmask of CPUs on Node 'node'.
 331  */
 332 const cpumask_t *_node_to_cpumask_ptr(int node)
 333 {
 334         if (node_to_cpumask_map == NULL) {
 335                 printk(KERN_WARNING
 336                         "_node_to_cpumask_ptr(%d): no node_to_cpumask_map!\n",
 337                         node);
 338                 dump_stack();
 339                 return (const cpumask_t *)&cpu_online_map;
 340         }
 341         if (node >= nr_node_ids) {
 342                 printk(KERN_WARNING
 343                         "_node_to_cpumask_ptr(%d): node > nr_node_ids(%d)\n",
 344                         node, nr_node_ids);
 345                 dump_stack();
 346                 return &cpu_mask_none;
 347         }
 348         return &node_to_cpumask_map[node];
 349 }
 350 EXPORT_SYMBOL(_node_to_cpumask_ptr);
 351
 352 /*
 353  * Returns a bitmask of CPUs on Node 'node'.
 354  *
 355  * Side note: this function creates the returned cpumask on the stack
 356  * so with a high NR_CPUS count, excessive stack space is used.  The
 357  * node_to_cpumask_ptr function should be used whenever possible.
 358  */
 359 cpumask_t node_to_cpumask(int node)
 360 {
 361         if (node_to_cpumask_map == NULL) {
 362                 printk(KERN_WARNING
 363                         "node_to_cpumask(%d): no node_to_cpumask_map!\n", node);
 364                 dump_stack();
 365                 return cpu_online_map;
 366         }
 367         if (node >= nr_node_ids) {
 368                 printk(KERN_WARNING
 369                         "node_to_cpumask(%d): node > nr_node_ids(%d)\n",
 370                         node, nr_node_ids);
 371                 dump_stack();
 372                 return cpu_mask_none;
 373         }
 374         return node_to_cpumask_map[node];
 375 }
 376 EXPORT_SYMBOL(node_to_cpumask);
 377
 378 /*
 379  * --------- end of debug versions of the numa functions ---------
 380  */
 381
 382 #endif /* CONFIG_DEBUG_PER_CPU_MAPS */
 383
 384 #endif /* X86_64_NUMA */
 385