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