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