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