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_HAVE_CPUMASK_OF_CPU_MAP
  84 cpumask_t *cpumask_of_cpu_map __read_mostly;
  85 EXPORT_SYMBOL(cpumask_of_cpu_map);
  86
  87 /* requires nr_cpu_ids to be initialized */
  88 static void __init setup_cpumask_of_cpu(void)
  89 {
  90         int i;
  91
  92         /* alloc_bootmem zeroes memory */
  93         cpumask_of_cpu_map = alloc_bootmem_low(sizeof(cpumask_t) * nr_cpu_ids);
  94         for (i = 0; i < nr_cpu_ids; i++)
  95                 cpu_set(i, cpumask_of_cpu_map[i]);
  96 }
  97 #else
  98 static inline void setup_cpumask_of_cpu(void) { }
  99 #endif
 100
 101 #ifdef CONFIG_X86_32
 102 /*
 103  * Great future not-so-futuristic plan: make i386 and x86_64 do it
 104  * the same way
 105  */
 106 unsigned long __per_cpu_offset[NR_CPUS] __read_mostly;
 107 EXPORT_SYMBOL(__per_cpu_offset);
 108 static inline void setup_cpu_pda_map(void) { }
 109
 110 #elif !defined(CONFIG_SMP)
 111 static inline void setup_cpu_pda_map(void) { }
 112
 113 #else /* CONFIG_SMP && CONFIG_X86_64 */
 114
 115 /*
 116  * Allocate cpu_pda pointer table and array via alloc_bootmem.
 117  */
 118 static void __init setup_cpu_pda_map(void)
 119 {
 120         char *pda;
 121         struct x8664_pda **new_cpu_pda;
 122         unsigned long size;
 123         int cpu;
 124
 125         size = roundup(sizeof(struct x8664_pda), cache_line_size());
 126
 127         /* allocate cpu_pda array and pointer table */
 128         {
 129                 unsigned long tsize = nr_cpu_ids * sizeof(void *);
 130                 unsigned long asize = size * (nr_cpu_ids - 1);
 131
 132                 tsize = roundup(tsize, cache_line_size());
 133                 new_cpu_pda = alloc_bootmem(tsize + asize);
 134                 pda = (char *)new_cpu_pda + tsize;
 135         }
 136
 137         /* initialize pointer table to static pda's */
 138         for_each_possible_cpu(cpu) {
 139                 if (cpu == 0) {
 140                         /* leave boot cpu pda in place */
 141                         new_cpu_pda[0] = cpu_pda(0);
 142                         continue;
 143                 }
 144                 new_cpu_pda[cpu] = (struct x8664_pda *)pda;
 145                 new_cpu_pda[cpu]->in_bootmem = 1;
 146                 pda += size;
 147         }
 148
 149         /* point to new pointer table */
 150         _cpu_pda = new_cpu_pda;
 151 }
 152 #endif
 153
 154 /*
 155  * Great future plan:
 156  * Declare PDA itself and support (irqstack,tss,pgd) as per cpu data.
 157  * Always point %gs to its beginning
 158  */
 159 void __init setup_per_cpu_areas(void)
 160 {
 161         ssize_t size = PERCPU_ENOUGH_ROOM;
 162         char *ptr;
 163         int cpu;
 164
 165         /* Setup cpu_pda map */
 166         setup_cpu_pda_map();
 167
 168         /* Copy section for each CPU (we discard the original) */
 169         size = PERCPU_ENOUGH_ROOM;
 170         printk(KERN_INFO "PERCPU: Allocating %zd bytes of per cpu data\n",
 171                           size);
 172
 173         for_each_possible_cpu(cpu) {
 174 #ifndef CONFIG_NEED_MULTIPLE_NODES
 175                 ptr = alloc_bootmem_pages(size);
 176 #else
 177                 int node = early_cpu_to_node(cpu);
 178                 if (!node_online(node) || !NODE_DATA(node)) {
 179                         ptr = alloc_bootmem_pages(size);
 180                         printk(KERN_INFO
 181                                "cpu %d has no node %d or node-local memory\n",
 182                                 cpu, node);
 183                 }
 184                 else
 185                         ptr = alloc_bootmem_pages_node(NODE_DATA(node), size);
 186 #endif
 187                 per_cpu_offset(cpu) = ptr - __per_cpu_start;
 188                 memcpy(ptr, __per_cpu_start, __per_cpu_end - __per_cpu_start);
 189
 190         }
 191
 192         printk(KERN_DEBUG "NR_CPUS: %d, nr_cpu_ids: %d, nr_node_ids %d\n",
 193                 NR_CPUS, nr_cpu_ids, nr_node_ids);
 194
 195         /* Setup percpu data maps */
 196         setup_per_cpu_maps();
 197
 198         /* Setup node to cpumask map */
 199         setup_node_to_cpumask_map();
 200
 201         /* Setup cpumask_of_cpu map */
 202         setup_cpumask_of_cpu();
 203 }
 204
 205 #endif
 206
 207 #ifdef X86_64_NUMA
 208
 209 /*
 210  * Allocate node_to_cpumask_map based on number of available nodes
 211  * Requires node_possible_map to be valid.
 212  *
 213  * Note: node_to_cpumask() is not valid until after this is done.
 214  */
 215 static void __init setup_node_to_cpumask_map(void)
 216 {
 217         unsigned int node, num = 0;
 218         cpumask_t *map;
 219
 220         /* setup nr_node_ids if not done yet */
 221         if (nr_node_ids == MAX_NUMNODES) {
 222                 for_each_node_mask(node, node_possible_map)
 223                         num = node;
 224                 nr_node_ids = num + 1;
 225         }
 226
 227         /* allocate the map */
 228         map = alloc_bootmem_low(nr_node_ids * sizeof(cpumask_t));
 229
 230         pr_debug(KERN_DEBUG "Node to cpumask map at %p for %d nodes\n",
 231                  map, nr_node_ids);
 232
 233         /* node_to_cpumask() will now work */
 234         node_to_cpumask_map = map;
 235 }
 236
 237 void __cpuinit numa_set_node(int cpu, int node)
 238 {
 239         int *cpu_to_node_map = early_per_cpu_ptr(x86_cpu_to_node_map);
 240
 241         if (cpu_pda(cpu) && node != NUMA_NO_NODE)
 242                 cpu_pda(cpu)->nodenumber = node;
 243
 244         if (cpu_to_node_map)
 245                 cpu_to_node_map[cpu] = node;
 246
 247         else if (per_cpu_offset(cpu))
 248                 per_cpu(x86_cpu_to_node_map, cpu) = node;
 249
 250         else
 251                 pr_debug("Setting node for non-present cpu %d\n", cpu);
 252 }
 253
 254 void __cpuinit numa_clear_node(int cpu)
 255 {
 256         numa_set_node(cpu, NUMA_NO_NODE);
 257 }
 258
 259 #ifndef CONFIG_DEBUG_PER_CPU_MAPS
 260
 261 void __cpuinit numa_add_cpu(int cpu)
 262 {
 263         cpu_set(cpu, node_to_cpumask_map[early_cpu_to_node(cpu)]);
 264 }
 265
 266 void __cpuinit numa_remove_cpu(int cpu)
 267 {
 268         cpu_clear(cpu, node_to_cpumask_map[cpu_to_node(cpu)]);
 269 }
 270
 271 #else /* CONFIG_DEBUG_PER_CPU_MAPS */
 272
 273 /*
 274  * --------- debug versions of the numa functions ---------
 275  */
 276 static void __cpuinit numa_set_cpumask(int cpu, int enable)
 277 {
 278         int node = cpu_to_node(cpu);
 279         cpumask_t *mask;
 280         char buf[64];
 281
 282         if (node_to_cpumask_map == NULL) {
 283                 printk(KERN_ERR "node_to_cpumask_map NULL\n");
 284                 dump_stack();
 285                 return;
 286         }
 287
 288         mask = &node_to_cpumask_map[node];
 289         if (enable)
 290                 cpu_set(cpu, *mask);
 291         else
 292                 cpu_clear(cpu, *mask);
 293
 294         cpulist_scnprintf(buf, sizeof(buf), *mask);
 295         printk(KERN_DEBUG "%s cpu %d node %d: mask now %s\n",
 296                 enable? "numa_add_cpu":"numa_remove_cpu", cpu, node, buf);
 297  }
 298
 299 void __cpuinit numa_add_cpu(int cpu)
 300 {
 301         numa_set_cpumask(cpu, 1);
 302 }
 303
 304 void __cpuinit numa_remove_cpu(int cpu)
 305 {
 306         numa_set_cpumask(cpu, 0);
 307 }
 308
 309 int cpu_to_node(int cpu)
 310 {
 311         if (early_per_cpu_ptr(x86_cpu_to_node_map)) {
 312                 printk(KERN_WARNING
 313                         "cpu_to_node(%d): usage too early!\n", cpu);
 314                 dump_stack();
 315                 return early_per_cpu_ptr(x86_cpu_to_node_map)[cpu];
 316         }
 317         return per_cpu(x86_cpu_to_node_map, cpu);
 318 }
 319 EXPORT_SYMBOL(cpu_to_node);
 320
 321 /*
 322  * Same function as cpu_to_node() but used if called before the
 323  * per_cpu areas are setup.
 324  */
 325 int early_cpu_to_node(int cpu)
 326 {
 327         if (early_per_cpu_ptr(x86_cpu_to_node_map))
 328                 return early_per_cpu_ptr(x86_cpu_to_node_map)[cpu];
 329
 330         if (!per_cpu_offset(cpu)) {
 331                 printk(KERN_WARNING
 332                         "early_cpu_to_node(%d): no per_cpu area!\n", cpu);
 333                 dump_stack();
 334                 return NUMA_NO_NODE;
 335         }
 336         return per_cpu(x86_cpu_to_node_map, cpu);
 337 }
 338
 339
 340 /* empty cpumask */
 341 static const cpumask_t cpu_mask_none;
 342
 343 /*
 344  * Returns a pointer to the bitmask of CPUs on Node 'node'.
 345  */
 346 const cpumask_t *_node_to_cpumask_ptr(int node)
 347 {
 348         if (node_to_cpumask_map == NULL) {
 349                 printk(KERN_WARNING
 350                         "_node_to_cpumask_ptr(%d): no node_to_cpumask_map!\n",
 351                         node);
 352                 dump_stack();
 353                 return (const cpumask_t *)&cpu_online_map;
 354         }
 355         if (node >= nr_node_ids) {
 356                 printk(KERN_WARNING
 357                         "_node_to_cpumask_ptr(%d): node > nr_node_ids(%d)\n",
 358                         node, nr_node_ids);
 359                 dump_stack();
 360                 return &cpu_mask_none;
 361         }
 362         return &node_to_cpumask_map[node];
 363 }
 364 EXPORT_SYMBOL(_node_to_cpumask_ptr);
 365
 366 /*
 367  * Returns a bitmask of CPUs on Node 'node'.
 368  *
 369  * Side note: this function creates the returned cpumask on the stack
 370  * so with a high NR_CPUS count, excessive stack space is used.  The
 371  * node_to_cpumask_ptr function should be used whenever possible.
 372  */
 373 cpumask_t node_to_cpumask(int node)
 374 {
 375         if (node_to_cpumask_map == NULL) {
 376                 printk(KERN_WARNING
 377                         "node_to_cpumask(%d): no node_to_cpumask_map!\n", node);
 378                 dump_stack();
 379                 return cpu_online_map;
 380         }
 381         if (node >= nr_node_ids) {
 382                 printk(KERN_WARNING
 383                         "node_to_cpumask(%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(node_to_cpumask);
 391
 392 /*
 393  * --------- end of debug versions of the numa functions ---------
 394  */
 395
 396 #endif /* CONFIG_DEBUG_PER_CPU_MAPS */
 397
 398 #endif /* X86_64_NUMA */
 399