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/proto.h>
  17 #include <asm/cpumask.h>
  18
  19 #ifdef CONFIG_DEBUG_PER_CPU_MAPS
  20 # define DBG(x...) printk(KERN_DEBUG x)
  21 #else
  22 # define DBG(x...)
  23 #endif
  24
  25 /*
  26  * Could be inside CONFIG_HAVE_SETUP_PER_CPU_AREA with other stuff but
  27  * voyager wants cpu_number too.
  28  */
  29 #ifdef CONFIG_SMP
  30 DEFINE_PER_CPU(int, cpu_number);
  31 EXPORT_PER_CPU_SYMBOL(cpu_number);
  32 #endif
  33
  34 #ifdef CONFIG_X86_LOCAL_APIC
  35 unsigned int num_processors;
  36 unsigned disabled_cpus __cpuinitdata;
  37 /* Processor that is doing the boot up */
  38 unsigned int boot_cpu_physical_apicid = -1U;
  39 EXPORT_SYMBOL(boot_cpu_physical_apicid);
  40 unsigned int max_physical_apicid;
  41
  42 /* Bitmask of physically existing CPUs */
  43 physid_mask_t phys_cpu_present_map;
  44 #endif
  45
  46 /*
  47  * Map cpu index to physical APIC ID
  48  */
  49 DEFINE_EARLY_PER_CPU(u16, x86_cpu_to_apicid, BAD_APICID);
  50 DEFINE_EARLY_PER_CPU(u16, x86_bios_cpu_apicid, BAD_APICID);
  51 EXPORT_EARLY_PER_CPU_SYMBOL(x86_cpu_to_apicid);
  52 EXPORT_EARLY_PER_CPU_SYMBOL(x86_bios_cpu_apicid);
  53
  54 #if defined(CONFIG_NUMA) && defined(CONFIG_X86_64)
  55 #define X86_64_NUMA     1       /* (used later) */
  56
  57 /*
  58  * Map cpu index to node index
  59  */
  60 DEFINE_EARLY_PER_CPU(int, x86_cpu_to_node_map, NUMA_NO_NODE);
  61 EXPORT_EARLY_PER_CPU_SYMBOL(x86_cpu_to_node_map);
  62
  63 /*
  64  * Which logical CPUs are on which nodes
  65  */
  66 cpumask_t *node_to_cpumask_map;
  67 EXPORT_SYMBOL(node_to_cpumask_map);
  68
  69 /*
  70  * Setup node_to_cpumask_map
  71  */
  72 static void __init setup_node_to_cpumask_map(void);
  73
  74 #else
  75 static inline void setup_node_to_cpumask_map(void) { }
  76 #endif
  77
  78 /*
  79  * Define load_pda_offset() and per-cpu __pda for x86_64.
  80  * load_pda_offset() is responsible for loading the offset of pda into
  81  * %gs.
  82  *
  83  * On SMP, pda offset also duals as percpu base address and thus it
  84  * should be at the start of per-cpu area.  To achieve this, it's
  85  * preallocated in vmlinux_64.lds.S directly instead of using
  86  * DEFINE_PER_CPU().
  87  */
  88 #ifdef CONFIG_X86_64
  89 void __cpuinit load_pda_offset(int cpu)
  90 {
  91         /* Memory clobbers used to order pda/percpu accesses */
  92         mb();
  93         wrmsrl(MSR_GS_BASE, cpu_pda(cpu));
  94         mb();
  95 }
  96 #ifndef CONFIG_SMP
  97 DEFINE_PER_CPU(struct x8664_pda, __pda);
  98 #endif
  99 EXPORT_PER_CPU_SYMBOL(__pda);
 100 #endif /* CONFIG_SMP && CONFIG_X86_64 */
 101
 102 #ifdef CONFIG_X86_64
 103
 104 /* correctly size the local cpu masks */
 105 static void setup_cpu_local_masks(void)
 106 {
 107         alloc_bootmem_cpumask_var(&cpu_initialized_mask);
 108         alloc_bootmem_cpumask_var(&cpu_callin_mask);
 109         alloc_bootmem_cpumask_var(&cpu_callout_mask);
 110         alloc_bootmem_cpumask_var(&cpu_sibling_setup_mask);
 111 }
 112
 113 #else /* CONFIG_X86_32 */
 114
 115 static inline void setup_cpu_local_masks(void)
 116 {
 117 }
 118
 119 #endif /* CONFIG_X86_32 */
 120
 121 #ifdef CONFIG_HAVE_SETUP_PER_CPU_AREA
 122 /*
 123  * Copy data used in early init routines from the initial arrays to the
 124  * per cpu data areas.  These arrays then become expendable and the
 125  * *_early_ptr's are zeroed indicating that the static arrays are gone.
 126  */
 127 static void __init setup_per_cpu_maps(void)
 128 {
 129         int cpu;
 130
 131         for_each_possible_cpu(cpu) {
 132                 per_cpu(x86_cpu_to_apicid, cpu) =
 133                                 early_per_cpu_map(x86_cpu_to_apicid, cpu);
 134                 per_cpu(x86_bios_cpu_apicid, cpu) =
 135                                 early_per_cpu_map(x86_bios_cpu_apicid, cpu);
 136 #ifdef X86_64_NUMA
 137                 per_cpu(x86_cpu_to_node_map, cpu) =
 138                                 early_per_cpu_map(x86_cpu_to_node_map, cpu);
 139 #endif
 140         }
 141
 142         /* indicate the early static arrays will soon be gone */
 143         early_per_cpu_ptr(x86_cpu_to_apicid) = NULL;
 144         early_per_cpu_ptr(x86_bios_cpu_apicid) = NULL;
 145 #ifdef X86_64_NUMA
 146         early_per_cpu_ptr(x86_cpu_to_node_map) = NULL;
 147 #endif
 148 }
 149
 150 #ifdef CONFIG_X86_64
 151 unsigned long __per_cpu_offset[NR_CPUS] __read_mostly = {
 152         [0] = (unsigned long)__per_cpu_load,
 153 };
 154 #else
 155 unsigned long __per_cpu_offset[NR_CPUS] __read_mostly;
 156 #endif
 157 EXPORT_SYMBOL(__per_cpu_offset);
 158
 159 /*
 160  * Great future plan:
 161  * Declare PDA itself and support (irqstack,tss,pgd) as per cpu data.
 162  * Always point %gs to its beginning
 163  */
 164 void __init setup_per_cpu_areas(void)
 165 {
 166         ssize_t size, old_size;
 167         char *ptr;
 168         int cpu;
 169         unsigned long align = 1;
 170
 171         /* Copy section for each CPU (we discard the original) */
 172         old_size = PERCPU_ENOUGH_ROOM;
 173         align = max_t(unsigned long, PAGE_SIZE, align);
 174         size = roundup(old_size, align);
 175
 176         pr_info("NR_CPUS:%d nr_cpumask_bits:%d nr_cpu_ids:%d nr_node_ids:%d\n",
 177                 NR_CPUS, nr_cpumask_bits, nr_cpu_ids, nr_node_ids);
 178
 179         pr_info("PERCPU: Allocating %zd bytes of per cpu data\n", size);
 180
 181         for_each_possible_cpu(cpu) {
 182 #ifndef CONFIG_NEED_MULTIPLE_NODES
 183                 ptr = __alloc_bootmem(size, align,
 184                                  __pa(MAX_DMA_ADDRESS));
 185 #else
 186                 int node = early_cpu_to_node(cpu);
 187                 if (!node_online(node) || !NODE_DATA(node)) {
 188                         ptr = __alloc_bootmem(size, align,
 189                                          __pa(MAX_DMA_ADDRESS));
 190                         pr_info("cpu %d has no node %d or node-local memory\n",
 191                                 cpu, node);
 192                         pr_debug("per cpu data for cpu%d at %016lx\n",
 193                                  cpu, __pa(ptr));
 194                 } else {
 195                         ptr = __alloc_bootmem_node(NODE_DATA(node), size, align,
 196                                                         __pa(MAX_DMA_ADDRESS));
 197                         pr_debug("per cpu data for cpu%d on node%d at %016lx\n",
 198                                 cpu, node, __pa(ptr));
 199                 }
 200 #endif
 201
 202                 memcpy(ptr, __per_cpu_load, __per_cpu_end - __per_cpu_start);
 203                 per_cpu_offset(cpu) = ptr - __per_cpu_start;
 204                 per_cpu(this_cpu_off, cpu) = per_cpu_offset(cpu);
 205                 per_cpu(cpu_number, cpu) = cpu;
 206 #ifdef CONFIG_X86_64
 207                 per_cpu(irq_stack_ptr, cpu) =
 208                         (char *)per_cpu(irq_stack, cpu) + IRQ_STACK_SIZE - 64;
 209                 /*
 210                  * CPU0 modified pda in the init data area, reload pda
 211                  * offset for CPU0 and clear the area for others.
 212                  */
 213                 if (cpu == 0)
 214                         load_pda_offset(0);
 215                 else
 216                         memset(cpu_pda(cpu), 0, sizeof(*cpu_pda(cpu)));
 217 #endif
 218
 219                 DBG("PERCPU: cpu %4d %p\n", cpu, ptr);
 220         }
 221
 222         /* Setup percpu data maps */
 223         setup_per_cpu_maps();
 224
 225         /* Setup node to cpumask map */
 226         setup_node_to_cpumask_map();
 227
 228         /* Setup cpu initialized, callin, callout masks */
 229         setup_cpu_local_masks();
 230 }
 231
 232 #endif
 233
 234 #ifdef X86_64_NUMA
 235
 236 /*
 237  * Allocate node_to_cpumask_map based on number of available nodes
 238  * Requires node_possible_map to be valid.
 239  *
 240  * Note: node_to_cpumask() is not valid until after this is done.
 241  * (Use CONFIG_DEBUG_PER_CPU_MAPS to check this.)
 242  */
 243 static void __init setup_node_to_cpumask_map(void)
 244 {
 245         unsigned int node, num = 0;
 246         cpumask_t *map;
 247
 248         /* setup nr_node_ids if not done yet */
 249         if (nr_node_ids == MAX_NUMNODES) {
 250                 for_each_node_mask(node, node_possible_map)
 251                         num = node;
 252                 nr_node_ids = num + 1;
 253         }
 254
 255         /* allocate the map */
 256         map = alloc_bootmem_low(nr_node_ids * sizeof(cpumask_t));
 257         DBG("node_to_cpumask_map at %p for %d nodes\n", map, nr_node_ids);
 258
 259         pr_debug("Node to cpumask map at %p for %d nodes\n",
 260                  map, nr_node_ids);
 261
 262         /* node_to_cpumask() will now work */
 263         node_to_cpumask_map = map;
 264 }
 265
 266 void __cpuinit numa_set_node(int cpu, int node)
 267 {
 268         int *cpu_to_node_map = early_per_cpu_ptr(x86_cpu_to_node_map);
 269
 270         /* early setting, no percpu area yet */
 271         if (cpu_to_node_map) {
 272                 cpu_to_node_map[cpu] = node;
 273                 return;
 274         }
 275
 276 #ifdef CONFIG_DEBUG_PER_CPU_MAPS
 277         if (cpu >= nr_cpu_ids || !per_cpu_offset(cpu)) {
 278                 printk(KERN_ERR "numa_set_node: invalid cpu# (%d)\n", cpu);
 279                 dump_stack();
 280                 return;
 281         }
 282 #endif
 283         per_cpu(x86_cpu_to_node_map, cpu) = node;
 284
 285         if (node != NUMA_NO_NODE)
 286                 cpu_pda(cpu)->nodenumber = node;
 287 }
 288
 289 void __cpuinit numa_clear_node(int cpu)
 290 {
 291         numa_set_node(cpu, NUMA_NO_NODE);
 292 }
 293
 294 #ifndef CONFIG_DEBUG_PER_CPU_MAPS
 295
 296 void __cpuinit numa_add_cpu(int cpu)
 297 {
 298         cpu_set(cpu, node_to_cpumask_map[early_cpu_to_node(cpu)]);
 299 }
 300
 301 void __cpuinit numa_remove_cpu(int cpu)
 302 {
 303         cpu_clear(cpu, node_to_cpumask_map[early_cpu_to_node(cpu)]);
 304 }
 305
 306 #else /* CONFIG_DEBUG_PER_CPU_MAPS */
 307
 308 /*
 309  * --------- debug versions of the numa functions ---------
 310  */
 311 static void __cpuinit numa_set_cpumask(int cpu, int enable)
 312 {
 313         int node = early_cpu_to_node(cpu);
 314         cpumask_t *mask;
 315         char buf[64];
 316
 317         if (node_to_cpumask_map == NULL) {
 318                 printk(KERN_ERR "node_to_cpumask_map NULL\n");
 319                 dump_stack();
 320                 return;
 321         }
 322
 323         mask = &node_to_cpumask_map[node];
 324         if (enable)
 325                 cpu_set(cpu, *mask);
 326         else
 327                 cpu_clear(cpu, *mask);
 328
 329         cpulist_scnprintf(buf, sizeof(buf), mask);
 330         printk(KERN_DEBUG "%s cpu %d node %d: mask now %s\n",
 331                 enable ? "numa_add_cpu" : "numa_remove_cpu", cpu, node, buf);
 332 }
 333
 334 void __cpuinit numa_add_cpu(int cpu)
 335 {
 336         numa_set_cpumask(cpu, 1);
 337 }
 338
 339 void __cpuinit numa_remove_cpu(int cpu)
 340 {
 341         numa_set_cpumask(cpu, 0);
 342 }
 343
 344 int cpu_to_node(int cpu)
 345 {
 346         if (early_per_cpu_ptr(x86_cpu_to_node_map)) {
 347                 printk(KERN_WARNING
 348                         "cpu_to_node(%d): usage too early!\n", cpu);
 349                 dump_stack();
 350                 return early_per_cpu_ptr(x86_cpu_to_node_map)[cpu];
 351         }
 352         return per_cpu(x86_cpu_to_node_map, cpu);
 353 }
 354 EXPORT_SYMBOL(cpu_to_node);
 355
 356 /*
 357  * Same function as cpu_to_node() but used if called before the
 358  * per_cpu areas are setup.
 359  */
 360 int early_cpu_to_node(int cpu)
 361 {
 362         if (early_per_cpu_ptr(x86_cpu_to_node_map))
 363                 return early_per_cpu_ptr(x86_cpu_to_node_map)[cpu];
 364
 365         if (!per_cpu_offset(cpu)) {
 366                 printk(KERN_WARNING
 367                         "early_cpu_to_node(%d): no per_cpu area!\n", cpu);
 368                 dump_stack();
 369                 return NUMA_NO_NODE;
 370         }
 371         return per_cpu(x86_cpu_to_node_map, cpu);
 372 }
 373
 374
 375 /* empty cpumask */
 376 static const cpumask_t cpu_mask_none;
 377
 378 /*
 379  * Returns a pointer to the bitmask of CPUs on Node 'node'.
 380  */
 381 const cpumask_t *cpumask_of_node(int node)
 382 {
 383         if (node_to_cpumask_map == NULL) {
 384                 printk(KERN_WARNING
 385                         "cpumask_of_node(%d): no node_to_cpumask_map!\n",
 386                         node);
 387                 dump_stack();
 388                 return (const cpumask_t *)&cpu_online_map;
 389         }
 390         if (node >= nr_node_ids) {
 391                 printk(KERN_WARNING
 392                         "cpumask_of_node(%d): node > nr_node_ids(%d)\n",
 393                         node, nr_node_ids);
 394                 dump_stack();
 395                 return &cpu_mask_none;
 396         }
 397         return &node_to_cpumask_map[node];
 398 }
 399 EXPORT_SYMBOL(cpumask_of_node);
 400
 401 /*
 402  * Returns a bitmask of CPUs on Node 'node'.
 403  *
 404  * Side note: this function creates the returned cpumask on the stack
 405  * so with a high NR_CPUS count, excessive stack space is used.  The
 406  * node_to_cpumask_ptr function should be used whenever possible.
 407  */
 408 cpumask_t node_to_cpumask(int node)
 409 {
 410         if (node_to_cpumask_map == NULL) {
 411                 printk(KERN_WARNING
 412                         "node_to_cpumask(%d): no node_to_cpumask_map!\n", node);
 413                 dump_stack();
 414                 return cpu_online_map;
 415         }
 416         if (node >= nr_node_ids) {
 417                 printk(KERN_WARNING
 418                         "node_to_cpumask(%d): node > nr_node_ids(%d)\n",
 419                         node, nr_node_ids);
 420                 dump_stack();
 421                 return cpu_mask_none;
 422         }
 423         return node_to_cpumask_map[node];
 424 }
 425 EXPORT_SYMBOL(node_to_cpumask);
 426
 427 /*
 428  * --------- end of debug versions of the numa functions ---------
 429  */
 430
 431 #endif /* CONFIG_DEBUG_PER_CPU_MAPS */
 432
 433 #endif /* X86_64_NUMA */
 434