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/cpumask.h>
  17
  18 #ifdef CONFIG_DEBUG_PER_CPU_MAPS
  19 # define DBG(x...) printk(KERN_DEBUG x)
  20 #else
  21 # define DBG(x...)
  22 #endif
  23
  24 #ifdef CONFIG_X86_LOCAL_APIC
  25 unsigned int num_processors;
  26 unsigned disabled_cpus __cpuinitdata;
  27 /* Processor that is doing the boot up */
  28 unsigned int boot_cpu_physical_apicid = -1U;
  29 EXPORT_SYMBOL(boot_cpu_physical_apicid);
  30 unsigned int max_physical_apicid;
  31
  32 /* Bitmask of physically existing CPUs */
  33 physid_mask_t phys_cpu_present_map;
  34 #endif
  35
  36 /*
  37  * Map cpu index to physical APIC ID
  38  */
  39 DEFINE_EARLY_PER_CPU(u16, x86_cpu_to_apicid, BAD_APICID);
  40 DEFINE_EARLY_PER_CPU(u16, x86_bios_cpu_apicid, BAD_APICID);
  41 EXPORT_EARLY_PER_CPU_SYMBOL(x86_cpu_to_apicid);
  42 EXPORT_EARLY_PER_CPU_SYMBOL(x86_bios_cpu_apicid);
  43
  44 #if defined(CONFIG_NUMA) && defined(CONFIG_X86_64)
  45 #define X86_64_NUMA     1       /* (used later) */
  46
  47 /*
  48  * Map cpu index to node index
  49  */
  50 DEFINE_EARLY_PER_CPU(int, x86_cpu_to_node_map, NUMA_NO_NODE);
  51 EXPORT_EARLY_PER_CPU_SYMBOL(x86_cpu_to_node_map);
  52
  53 /*
  54  * Which logical CPUs are on which nodes
  55  */
  56 cpumask_t *node_to_cpumask_map;
  57 EXPORT_SYMBOL(node_to_cpumask_map);
  58
  59 /*
  60  * Setup node_to_cpumask_map
  61  */
  62 static void __init setup_node_to_cpumask_map(void);
  63
  64 #else
  65 static inline void setup_node_to_cpumask_map(void) { }
  66 #endif
  67
  68 #ifdef CONFIG_HAVE_SETUP_PER_CPU_AREA
  69 /*
  70  * Copy data used in early init routines from the initial arrays to the
  71  * per cpu data areas.  These arrays then become expendable and the
  72  * *_early_ptr's are zeroed indicating that the static arrays are gone.
  73  */
  74 static void __init setup_per_cpu_maps(void)
  75 {
  76         int cpu;
  77
  78         for_each_possible_cpu(cpu) {
  79                 per_cpu(x86_cpu_to_apicid, cpu) =
  80                                 early_per_cpu_map(x86_cpu_to_apicid, cpu);
  81                 per_cpu(x86_bios_cpu_apicid, cpu) =
  82                                 early_per_cpu_map(x86_bios_cpu_apicid, cpu);
  83 #ifdef X86_64_NUMA
  84                 per_cpu(x86_cpu_to_node_map, cpu) =
  85                                 early_per_cpu_map(x86_cpu_to_node_map, cpu);
  86 #endif
  87         }
  88
  89         /* indicate the early static arrays will soon be gone */
  90         early_per_cpu_ptr(x86_cpu_to_apicid) = NULL;
  91         early_per_cpu_ptr(x86_bios_cpu_apicid) = NULL;
  92 #ifdef X86_64_NUMA
  93         early_per_cpu_ptr(x86_cpu_to_node_map) = NULL;
  94 #endif
  95 }
  96
  97 #ifdef CONFIG_X86_32
  98 /*
  99  * Great future not-so-futuristic plan: make i386 and x86_64 do it
 100  * the same way
 101  */
 102 unsigned long __per_cpu_offset[NR_CPUS] __read_mostly;
 103 EXPORT_SYMBOL(__per_cpu_offset);
 104 static inline void setup_cpu_pda_map(void) { }
 105
 106 #elif !defined(CONFIG_SMP)
 107 static inline void setup_cpu_pda_map(void) { }
 108
 109 #else /* CONFIG_SMP && CONFIG_X86_64 */
 110
 111 /*
 112  * Allocate cpu_pda pointer table and array via alloc_bootmem.
 113  */
 114 static void __init setup_cpu_pda_map(void)
 115 {
 116         char *pda;
 117         unsigned long size;
 118         int cpu;
 119
 120         size = roundup(sizeof(struct x8664_pda), cache_line_size());
 121
 122         /* allocate cpu_pda array and pointer table */
 123         {
 124                 unsigned long asize = size * (nr_cpu_ids - 1);
 125
 126                 pda = alloc_bootmem(asize);
 127         }
 128
 129         /* initialize pointer table to static pda's */
 130         for_each_possible_cpu(cpu) {
 131                 if (cpu == 0) {
 132                         /* leave boot cpu pda in place */
 133                         continue;
 134                 }
 135                 cpu_pda(cpu) = (struct x8664_pda *)pda;
 136                 cpu_pda(cpu)->in_bootmem = 1;
 137                 pda += size;
 138         }
 139 }
 140
 141 #endif /* CONFIG_SMP && CONFIG_X86_64 */
 142
 143 #ifdef CONFIG_X86_64
 144
 145 /* correctly size the local cpu masks */
 146 static void setup_cpu_local_masks(void)
 147 {
 148         alloc_bootmem_cpumask_var(&cpu_initialized_mask);
 149         alloc_bootmem_cpumask_var(&cpu_callin_mask);
 150         alloc_bootmem_cpumask_var(&cpu_callout_mask);
 151         alloc_bootmem_cpumask_var(&cpu_sibling_setup_mask);
 152 }
 153
 154 #else /* CONFIG_X86_32 */
 155
 156 static inline void setup_cpu_local_masks(void)
 157 {
 158 }
 159
 160 #endif /* CONFIG_X86_32 */
 161
 162 /*
 163  * Great future plan:
 164  * Declare PDA itself and support (irqstack,tss,pgd) as per cpu data.
 165  * Always point %gs to its beginning
 166  */
 167 void __init setup_per_cpu_areas(void)
 168 {
 169         ssize_t size, old_size;
 170         char *ptr;
 171         int cpu;
 172         unsigned long align = 1;
 173
 174         /* Setup cpu_pda map */
 175         setup_cpu_pda_map();
 176
 177         /* Copy section for each CPU (we discard the original) */
 178         old_size = PERCPU_ENOUGH_ROOM;
 179         align = max_t(unsigned long, PAGE_SIZE, align);
 180         size = roundup(old_size, align);
 181
 182         pr_info("NR_CPUS:%d nr_cpumask_bits:%d nr_cpu_ids:%d nr_node_ids:%d\n",
 183                 NR_CPUS, nr_cpumask_bits, nr_cpu_ids, nr_node_ids);
 184
 185         pr_info("PERCPU: Allocating %zd bytes of per cpu data\n", size);
 186
 187         for_each_possible_cpu(cpu) {
 188 #ifndef CONFIG_NEED_MULTIPLE_NODES
 189                 ptr = __alloc_bootmem(size, align,
 190                                  __pa(MAX_DMA_ADDRESS));
 191 #else
 192                 int node = early_cpu_to_node(cpu);
 193                 if (!node_online(node) || !NODE_DATA(node)) {
 194                         ptr = __alloc_bootmem(size, align,
 195                                          __pa(MAX_DMA_ADDRESS));
 196                         pr_info("cpu %d has no node %d or node-local memory\n",
 197                                 cpu, node);
 198                         pr_debug("per cpu data for cpu%d at %016lx\n",
 199                                  cpu, __pa(ptr));
 200                 } else {
 201                         ptr = __alloc_bootmem_node(NODE_DATA(node), size, align,
 202                                                         __pa(MAX_DMA_ADDRESS));
 203                         pr_debug("per cpu data for cpu%d on node%d at %016lx\n",
 204                                 cpu, node, __pa(ptr));
 205                 }
 206 #endif
 207                 per_cpu_offset(cpu) = ptr - __per_cpu_start;
 208                 memcpy(ptr, __per_cpu_load, __per_cpu_end - __per_cpu_start);
 209
 210                 DBG("PERCPU: cpu %4d %p\n", cpu, ptr);
 211         }
 212
 213         /* Setup percpu data maps */
 214         setup_per_cpu_maps();
 215
 216         /* Setup node to cpumask map */
 217         setup_node_to_cpumask_map();
 218
 219         /* Setup cpu initialized, callin, callout masks */
 220         setup_cpu_local_masks();
 221 }
 222
 223 #endif
 224
 225 #ifdef X86_64_NUMA
 226
 227 /*
 228  * Allocate node_to_cpumask_map based on number of available nodes
 229  * Requires node_possible_map to be valid.
 230  *
 231  * Note: node_to_cpumask() is not valid until after this is done.
 232  * (Use CONFIG_DEBUG_PER_CPU_MAPS to check this.)
 233  */
 234 static void __init setup_node_to_cpumask_map(void)
 235 {
 236         unsigned int node, num = 0;
 237         cpumask_t *map;
 238
 239         /* setup nr_node_ids if not done yet */
 240         if (nr_node_ids == MAX_NUMNODES) {
 241                 for_each_node_mask(node, node_possible_map)
 242                         num = node;
 243                 nr_node_ids = num + 1;
 244         }
 245
 246         /* allocate the map */
 247         map = alloc_bootmem_low(nr_node_ids * sizeof(cpumask_t));
 248         DBG("node_to_cpumask_map at %p for %d nodes\n", map, nr_node_ids);
 249
 250         pr_debug("Node to cpumask map at %p for %d nodes\n",
 251                  map, nr_node_ids);
 252
 253         /* node_to_cpumask() will now work */
 254         node_to_cpumask_map = map;
 255 }
 256
 257 void __cpuinit numa_set_node(int cpu, int node)
 258 {
 259         int *cpu_to_node_map = early_per_cpu_ptr(x86_cpu_to_node_map);
 260
 261         /* early setting, no percpu area yet */
 262         if (cpu_to_node_map) {
 263                 cpu_to_node_map[cpu] = node;
 264                 return;
 265         }
 266
 267 #ifdef CONFIG_DEBUG_PER_CPU_MAPS
 268         if (cpu >= nr_cpu_ids || !per_cpu_offset(cpu)) {
 269                 printk(KERN_ERR "numa_set_node: invalid cpu# (%d)\n", cpu);
 270                 dump_stack();
 271                 return;
 272         }
 273 #endif
 274         per_cpu(x86_cpu_to_node_map, cpu) = node;
 275
 276         if (node != NUMA_NO_NODE)
 277                 cpu_pda(cpu)->nodenumber = node;
 278 }
 279
 280 void __cpuinit numa_clear_node(int cpu)
 281 {
 282         numa_set_node(cpu, NUMA_NO_NODE);
 283 }
 284
 285 #ifndef CONFIG_DEBUG_PER_CPU_MAPS
 286
 287 void __cpuinit numa_add_cpu(int cpu)
 288 {
 289         cpu_set(cpu, node_to_cpumask_map[early_cpu_to_node(cpu)]);
 290 }
 291
 292 void __cpuinit numa_remove_cpu(int cpu)
 293 {
 294         cpu_clear(cpu, node_to_cpumask_map[early_cpu_to_node(cpu)]);
 295 }
 296
 297 #else /* CONFIG_DEBUG_PER_CPU_MAPS */
 298
 299 /*
 300  * --------- debug versions of the numa functions ---------
 301  */
 302 static void __cpuinit numa_set_cpumask(int cpu, int enable)
 303 {
 304         int node = early_cpu_to_node(cpu);
 305         cpumask_t *mask;
 306         char buf[64];
 307
 308         if (node_to_cpumask_map == NULL) {
 309                 printk(KERN_ERR "node_to_cpumask_map NULL\n");
 310                 dump_stack();
 311                 return;
 312         }
 313
 314         mask = &node_to_cpumask_map[node];
 315         if (enable)
 316                 cpu_set(cpu, *mask);
 317         else
 318                 cpu_clear(cpu, *mask);
 319
 320         cpulist_scnprintf(buf, sizeof(buf), mask);
 321         printk(KERN_DEBUG "%s cpu %d node %d: mask now %s\n",
 322                 enable ? "numa_add_cpu" : "numa_remove_cpu", cpu, node, buf);
 323 }
 324
 325 void __cpuinit numa_add_cpu(int cpu)
 326 {
 327         numa_set_cpumask(cpu, 1);
 328 }
 329
 330 void __cpuinit numa_remove_cpu(int cpu)
 331 {
 332         numa_set_cpumask(cpu, 0);
 333 }
 334
 335 int cpu_to_node(int cpu)
 336 {
 337         if (early_per_cpu_ptr(x86_cpu_to_node_map)) {
 338                 printk(KERN_WARNING
 339                         "cpu_to_node(%d): usage too early!\n", cpu);
 340                 dump_stack();
 341                 return early_per_cpu_ptr(x86_cpu_to_node_map)[cpu];
 342         }
 343         return per_cpu(x86_cpu_to_node_map, cpu);
 344 }
 345 EXPORT_SYMBOL(cpu_to_node);
 346
 347 /*
 348  * Same function as cpu_to_node() but used if called before the
 349  * per_cpu areas are setup.
 350  */
 351 int early_cpu_to_node(int cpu)
 352 {
 353         if (early_per_cpu_ptr(x86_cpu_to_node_map))
 354                 return early_per_cpu_ptr(x86_cpu_to_node_map)[cpu];
 355
 356         if (!per_cpu_offset(cpu)) {
 357                 printk(KERN_WARNING
 358                         "early_cpu_to_node(%d): no per_cpu area!\n", cpu);
 359                 dump_stack();
 360                 return NUMA_NO_NODE;
 361         }
 362         return per_cpu(x86_cpu_to_node_map, cpu);
 363 }
 364
 365
 366 /* empty cpumask */
 367 static const cpumask_t cpu_mask_none;
 368
 369 /*
 370  * Returns a pointer to the bitmask of CPUs on Node 'node'.
 371  */
 372 const cpumask_t *cpumask_of_node(int node)
 373 {
 374         if (node_to_cpumask_map == NULL) {
 375                 printk(KERN_WARNING
 376                         "cpumask_of_node(%d): no node_to_cpumask_map!\n",
 377                         node);
 378                 dump_stack();
 379                 return (const cpumask_t *)&cpu_online_map;
 380         }
 381         if (node >= nr_node_ids) {
 382                 printk(KERN_WARNING
 383                         "cpumask_of_node(%d): node > nr_node_ids(%d)\n",
 384                         node, nr_node_ids);
 385                 dump_stack();
 386                 return &cpu_mask_none;
 387         }
 388         return &node_to_cpumask_map[node];
 389 }
 390 EXPORT_SYMBOL(cpumask_of_node);
 391
 392 /*
 393  * Returns a bitmask of CPUs on Node 'node'.
 394  *
 395  * Side note: this function creates the returned cpumask on the stack
 396  * so with a high NR_CPUS count, excessive stack space is used.  The
 397  * node_to_cpumask_ptr function should be used whenever possible.
 398  */
 399 cpumask_t node_to_cpumask(int node)
 400 {
 401         if (node_to_cpumask_map == NULL) {
 402                 printk(KERN_WARNING
 403                         "node_to_cpumask(%d): no node_to_cpumask_map!\n", node);
 404                 dump_stack();
 405                 return cpu_online_map;
 406         }
 407         if (node >= nr_node_ids) {
 408                 printk(KERN_WARNING
 409                         "node_to_cpumask(%d): node > nr_node_ids(%d)\n",
 410                         node, nr_node_ids);
 411                 dump_stack();
 412                 return cpu_mask_none;
 413         }
 414         return node_to_cpumask_map[node];
 415 }
 416 EXPORT_SYMBOL(node_to_cpumask);
 417
 418 /*
 419  * --------- end of debug versions of the numa functions ---------
 420  */
 421
 422 #endif /* CONFIG_DEBUG_PER_CPU_MAPS */
 423
 424 #endif /* X86_64_NUMA */
 425