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
 135 #endif /* CONFIG_SMP && CONFIG_X86_64 */
 136
 137 #ifdef CONFIG_X86_64
 138
 139 /* correctly size the local cpu masks */
 140 static void setup_cpu_local_masks(void)
 141 {
 142         alloc_bootmem_cpumask_var(&cpu_initialized_mask);
 143         alloc_bootmem_cpumask_var(&cpu_callin_mask);
 144         alloc_bootmem_cpumask_var(&cpu_callout_mask);
 145         alloc_bootmem_cpumask_var(&cpu_sibling_setup_mask);
 146 }
 147
 148 #else /* CONFIG_X86_32 */
 149
 150 static inline void setup_cpu_local_masks(void)
 151 {
 152 }
 153
 154 #endif /* CONFIG_X86_32 */
 155
 156 /*
 157  * Great future plan:
 158  * Declare PDA itself and support (irqstack,tss,pgd) as per cpu data.
 159  * Always point %gs to its beginning
 160  */
 161 void __init setup_per_cpu_areas(void)
 162 {
 163         ssize_t size, old_size;
 164         char *ptr;
 165         int cpu;
 166         unsigned long align = 1;
 167
 168         /* Setup cpu_pda map */
 169         setup_cpu_pda_map();
 170
 171         /* Copy section for each CPU (we discard the original) */
 172         old_size = PERCPU_ENOUGH_ROOM;
 173         align = max_t(unsigned long, PAGE_SIZE, align);
 174         size = roundup(old_size, align);
 175
 176         pr_info("NR_CPUS:%d nr_cpumask_bits:%d nr_cpu_ids:%d nr_node_ids:%d\n",
 177                 NR_CPUS, nr_cpumask_bits, nr_cpu_ids, nr_node_ids);
 178
 179         pr_info("PERCPU: Allocating %zd bytes of per cpu data\n", size);
 180
 181         for_each_possible_cpu(cpu) {
 182 #ifndef CONFIG_NEED_MULTIPLE_NODES
 183                 ptr = __alloc_bootmem(size, align,
 184                                  __pa(MAX_DMA_ADDRESS));
 185 #else
 186                 int node = early_cpu_to_node(cpu);
 187                 if (!node_online(node) || !NODE_DATA(node)) {
 188                         ptr = __alloc_bootmem(size, align,
 189                                          __pa(MAX_DMA_ADDRESS));
 190                         pr_info("cpu %d has no node %d or node-local memory\n",
 191                                 cpu, node);
 192                         pr_debug("per cpu data for cpu%d at %016lx\n",
 193                                  cpu, __pa(ptr));
 194                 } else {
 195                         ptr = __alloc_bootmem_node(NODE_DATA(node), size, align,
 196                                                         __pa(MAX_DMA_ADDRESS));
 197                         pr_debug("per cpu data for cpu%d on node%d at %016lx\n",
 198                                 cpu, node, __pa(ptr));
 199                 }
 200 #endif
 201                 per_cpu_offset(cpu) = ptr - __per_cpu_start;
 202                 memcpy(ptr, __per_cpu_start, __per_cpu_end - __per_cpu_start);
 203         }
 204
 205         /* Setup percpu data maps */
 206         setup_per_cpu_maps();
 207
 208         /* Setup node to cpumask map */
 209         setup_node_to_cpumask_map();
 210
 211         /* Setup cpu initialized, callin, callout masks */
 212         setup_cpu_local_masks();
 213 }
 214
 215 #endif
 216
 217 #ifdef X86_64_NUMA
 218
 219 /*
 220  * Allocate node_to_cpumask_map based on number of available nodes
 221  * Requires node_possible_map to be valid.
 222  *
 223  * Note: node_to_cpumask() is not valid until after this is done.
 224  */
 225 static void __init setup_node_to_cpumask_map(void)
 226 {
 227         unsigned int node, num = 0;
 228         cpumask_t *map;
 229
 230         /* setup nr_node_ids if not done yet */
 231         if (nr_node_ids == MAX_NUMNODES) {
 232                 for_each_node_mask(node, node_possible_map)
 233                         num = node;
 234                 nr_node_ids = num + 1;
 235         }
 236
 237         /* allocate the map */
 238         map = alloc_bootmem_low(nr_node_ids * sizeof(cpumask_t));
 239
 240         pr_debug("Node to cpumask map at %p for %d nodes\n",
 241                  map, nr_node_ids);
 242
 243         /* node_to_cpumask() will now work */
 244         node_to_cpumask_map = map;
 245 }
 246
 247 void __cpuinit numa_set_node(int cpu, int node)
 248 {
 249         int *cpu_to_node_map = early_per_cpu_ptr(x86_cpu_to_node_map);
 250
 251         if (cpu_pda(cpu) && node != NUMA_NO_NODE)
 252                 cpu_pda(cpu)->nodenumber = node;
 253
 254         if (cpu_to_node_map)
 255                 cpu_to_node_map[cpu] = node;
 256
 257         else if (per_cpu_offset(cpu))
 258                 per_cpu(x86_cpu_to_node_map, cpu) = node;
 259
 260         else
 261                 pr_debug("Setting node for non-present cpu %d\n", cpu);
 262 }
 263
 264 void __cpuinit numa_clear_node(int cpu)
 265 {
 266         numa_set_node(cpu, NUMA_NO_NODE);
 267 }
 268
 269 #ifndef CONFIG_DEBUG_PER_CPU_MAPS
 270
 271 void __cpuinit numa_add_cpu(int cpu)
 272 {
 273         cpu_set(cpu, node_to_cpumask_map[early_cpu_to_node(cpu)]);
 274 }
 275
 276 void __cpuinit numa_remove_cpu(int cpu)
 277 {
 278         cpu_clear(cpu, node_to_cpumask_map[cpu_to_node(cpu)]);
 279 }
 280
 281 #else /* CONFIG_DEBUG_PER_CPU_MAPS */
 282
 283 /*
 284  * --------- debug versions of the numa functions ---------
 285  */
 286 static void __cpuinit numa_set_cpumask(int cpu, int enable)
 287 {
 288         int node = cpu_to_node(cpu);
 289         cpumask_t *mask;
 290         char buf[64];
 291
 292         if (node_to_cpumask_map == NULL) {
 293                 printk(KERN_ERR "node_to_cpumask_map NULL\n");
 294                 dump_stack();
 295                 return;
 296         }
 297
 298         mask = &node_to_cpumask_map[node];
 299         if (enable)
 300                 cpu_set(cpu, *mask);
 301         else
 302                 cpu_clear(cpu, *mask);
 303
 304         cpulist_scnprintf(buf, sizeof(buf), mask);
 305         printk(KERN_DEBUG "%s cpu %d node %d: mask now %s\n",
 306                 enable? "numa_add_cpu":"numa_remove_cpu", cpu, node, buf);
 307  }
 308
 309 void __cpuinit numa_add_cpu(int cpu)
 310 {
 311         numa_set_cpumask(cpu, 1);
 312 }
 313
 314 void __cpuinit numa_remove_cpu(int cpu)
 315 {
 316         numa_set_cpumask(cpu, 0);
 317 }
 318
 319 int cpu_to_node(int cpu)
 320 {
 321         if (early_per_cpu_ptr(x86_cpu_to_node_map)) {
 322                 printk(KERN_WARNING
 323                         "cpu_to_node(%d): usage too early!\n", cpu);
 324                 dump_stack();
 325                 return early_per_cpu_ptr(x86_cpu_to_node_map)[cpu];
 326         }
 327         return per_cpu(x86_cpu_to_node_map, cpu);
 328 }
 329 EXPORT_SYMBOL(cpu_to_node);
 330
 331 /*
 332  * Same function as cpu_to_node() but used if called before the
 333  * per_cpu areas are setup.
 334  */
 335 int early_cpu_to_node(int cpu)
 336 {
 337         if (early_per_cpu_ptr(x86_cpu_to_node_map))
 338                 return early_per_cpu_ptr(x86_cpu_to_node_map)[cpu];
 339
 340         if (!per_cpu_offset(cpu)) {
 341                 printk(KERN_WARNING
 342                         "early_cpu_to_node(%d): no per_cpu area!\n", cpu);
 343                 dump_stack();
 344                 return NUMA_NO_NODE;
 345         }
 346         return per_cpu(x86_cpu_to_node_map, cpu);
 347 }
 348
 349
 350 /* empty cpumask */
 351 static const cpumask_t cpu_mask_none;
 352
 353 /*
 354  * Returns a pointer to the bitmask of CPUs on Node 'node'.
 355  */
 356 const cpumask_t *cpumask_of_node(int node)
 357 {
 358         if (node_to_cpumask_map == NULL) {
 359                 printk(KERN_WARNING
 360                         "cpumask_of_node(%d): no node_to_cpumask_map!\n",
 361                         node);
 362                 dump_stack();
 363                 return (const cpumask_t *)&cpu_online_map;
 364         }
 365         if (node >= nr_node_ids) {
 366                 printk(KERN_WARNING
 367                         "cpumask_of_node(%d): node > nr_node_ids(%d)\n",
 368                         node, nr_node_ids);
 369                 dump_stack();
 370                 return &cpu_mask_none;
 371         }
 372         return &node_to_cpumask_map[node];
 373 }
 374 EXPORT_SYMBOL(cpumask_of_node);
 375
 376 /*
 377  * Returns a bitmask of CPUs on Node 'node'.
 378  *
 379  * Side note: this function creates the returned cpumask on the stack
 380  * so with a high NR_CPUS count, excessive stack space is used.  The
 381  * node_to_cpumask_ptr function should be used whenever possible.
 382  */
 383 cpumask_t node_to_cpumask(int node)
 384 {
 385         if (node_to_cpumask_map == NULL) {
 386                 printk(KERN_WARNING
 387                         "node_to_cpumask(%d): no node_to_cpumask_map!\n", node);
 388                 dump_stack();
 389                 return cpu_online_map;
 390         }
 391         if (node >= nr_node_ids) {
 392                 printk(KERN_WARNING
 393                         "node_to_cpumask(%d): node > nr_node_ids(%d)\n",
 394                         node, nr_node_ids);
 395                 dump_stack();
 396                 return cpu_mask_none;
 397         }
 398         return node_to_cpumask_map[node];
 399 }
 400 EXPORT_SYMBOL(node_to_cpumask);
 401
 402 /*
 403  * --------- end of debug versions of the numa functions ---------
 404  */
 405
 406 #endif /* CONFIG_DEBUG_PER_CPU_MAPS */
 407
 408 #endif /* X86_64_NUMA */
 409