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 <linux/pfn.h>
  11 #include <asm/sections.h>
  12 #include <asm/processor.h>
  13 #include <asm/setup.h>
  14 #include <asm/mpspec.h>
  15 #include <asm/apicdef.h>
  16 #include <asm/highmem.h>
  17 #include <asm/proto.h>
  18 #include <asm/cpumask.h>
  19 #include <asm/cpu.h>
  20 #include <asm/stackprotector.h>
  21
  22 #ifdef CONFIG_DEBUG_PER_CPU_MAPS
  23 # define DBG(x...) printk(KERN_DEBUG x)
  24 #else
  25 # define DBG(x...)
  26 #endif
  27
  28 DEFINE_PER_CPU(int, cpu_number);
  29 EXPORT_PER_CPU_SYMBOL(cpu_number);
  30
  31 #ifdef CONFIG_X86_64
  32 #define BOOT_PERCPU_OFFSET ((unsigned long)__per_cpu_load)
  33 #else
  34 #define BOOT_PERCPU_OFFSET 0
  35 #endif
  36
  37 DEFINE_PER_CPU(unsigned long, this_cpu_off) = BOOT_PERCPU_OFFSET;
  38 EXPORT_PER_CPU_SYMBOL(this_cpu_off);
  39
  40 unsigned long __per_cpu_offset[NR_CPUS] __read_mostly = {
  41         [0 ... NR_CPUS-1] = BOOT_PERCPU_OFFSET,
  42 };
  43 EXPORT_SYMBOL(__per_cpu_offset);
  44
  45 /**
  46  * pcpu_need_numa - determine percpu allocation needs to consider NUMA
  47  *
  48  * If NUMA is not configured or there is only one NUMA node available,
  49  * there is no reason to consider NUMA.  This function determines
  50  * whether percpu allocation should consider NUMA or not.
  51  *
  52  * RETURNS:
  53  * true if NUMA should be considered; otherwise, false.
  54  */
  55 static bool __init pcpu_need_numa(void)
  56 {
  57 #ifdef CONFIG_NEED_MULTIPLE_NODES
  58         pg_data_t *last = NULL;
  59         unsigned int cpu;
  60
  61         for_each_possible_cpu(cpu) {
  62                 int node = early_cpu_to_node(cpu);
  63
  64                 if (node_online(node) && NODE_DATA(node) &&
  65                     last && last != NODE_DATA(node))
  66                         return true;
  67
  68                 last = NODE_DATA(node);
  69         }
  70 #endif
  71         return false;
  72 }
  73
  74 /**
  75  * pcpu_alloc_bootmem - NUMA friendly alloc_bootmem wrapper for percpu
  76  * @cpu: cpu to allocate for
  77  * @size: size allocation in bytes
  78  * @align: alignment
  79  *
  80  * Allocate @size bytes aligned at @align for cpu @cpu.  This wrapper
  81  * does the right thing for NUMA regardless of the current
  82  * configuration.
  83  *
  84  * RETURNS:
  85  * Pointer to the allocated area on success, NULL on failure.
  86  */
  87 static void * __init pcpu_alloc_bootmem(unsigned int cpu, unsigned long size,
  88                                         unsigned long align)
  89 {
  90         const unsigned long goal = __pa(MAX_DMA_ADDRESS);
  91 #ifdef CONFIG_NEED_MULTIPLE_NODES
  92         int node = early_cpu_to_node(cpu);
  93         void *ptr;
  94
  95         if (!node_online(node) || !NODE_DATA(node)) {
  96                 ptr = __alloc_bootmem_nopanic(size, align, goal);
  97                 pr_info("cpu %d has no node %d or node-local memory\n",
  98                         cpu, node);
  99                 pr_debug("per cpu data for cpu%d %lu bytes at %016lx\n",
 100                          cpu, size, __pa(ptr));
 101         } else {
 102                 ptr = __alloc_bootmem_node_nopanic(NODE_DATA(node),
 103                                                    size, align, goal);
 104                 pr_debug("per cpu data for cpu%d %lu bytes on node%d at "
 105                          "%016lx\n", cpu, size, node, __pa(ptr));
 106         }
 107         return ptr;
 108 #else
 109         return __alloc_bootmem_nopanic(size, align, goal);
 110 #endif
 111 }
 112
 113 /*
 114  * Embedding allocator
 115  *
 116  * The first chunk is sized to just contain the static area plus
 117  * PERCPU_DYNAMIC_RESERVE and allocated as a contiguous area using
 118  * bootmem allocator and used as-is without being mapped into vmalloc
 119  * area.  This enables the first chunk to piggy back on the linear
 120  * physical PMD mapping and doesn't add any additional pressure to
 121  * TLB.
 122  */
 123 static void *pcpue_ptr __initdata;
 124 static size_t pcpue_unit_size __initdata;
 125
 126 static struct page * __init pcpue_get_page(unsigned int cpu, int pageno)
 127 {
 128         return virt_to_page(pcpue_ptr + cpu * pcpue_unit_size
 129                             + ((size_t)pageno << PAGE_SHIFT));
 130 }
 131
 132 static ssize_t __init setup_pcpu_embed(size_t static_size)
 133 {
 134         unsigned int cpu;
 135
 136         /*
 137          * If large page isn't supported, there's no benefit in doing
 138          * this.  Also, embedding allocation doesn't play well with
 139          * NUMA.
 140          */
 141         if (!cpu_has_pse || pcpu_need_numa())
 142                 return -EINVAL;
 143
 144         /* allocate and copy */
 145         pcpue_unit_size = PFN_ALIGN(static_size + PERCPU_DYNAMIC_RESERVE);
 146         pcpue_unit_size = max(pcpue_unit_size, PCPU_MIN_UNIT_SIZE);
 147         pcpue_ptr = pcpu_alloc_bootmem(0, num_possible_cpus() * pcpue_unit_size,
 148                                        PAGE_SIZE);
 149         if (!pcpue_ptr)
 150                 return -ENOMEM;
 151
 152         for_each_possible_cpu(cpu)
 153                 memcpy(pcpue_ptr + cpu * pcpue_unit_size, __per_cpu_load,
 154                        static_size);
 155
 156         /* we're ready, commit */
 157         pr_info("PERCPU: Embedded %zu pages at %p, static data %zu bytes\n",
 158                 pcpue_unit_size >> PAGE_SHIFT, pcpue_ptr, static_size);
 159
 160         return pcpu_setup_first_chunk(pcpue_get_page, static_size,
 161                                       pcpue_unit_size,
 162                                       pcpue_unit_size - static_size, pcpue_ptr,
 163                                       NULL);
 164 }
 165
 166 /*
 167  * 4k page allocator
 168  *
 169  * This is the basic allocator.  Static percpu area is allocated
 170  * page-by-page and most of initialization is done by the generic
 171  * setup function.
 172  */
 173 static struct page **pcpu4k_pages __initdata;
 174 static int pcpu4k_nr_static_pages __initdata;
 175
 176 static struct page * __init pcpu4k_get_page(unsigned int cpu, int pageno)
 177 {
 178         if (pageno < pcpu4k_nr_static_pages)
 179                 return pcpu4k_pages[cpu * pcpu4k_nr_static_pages + pageno];
 180         return NULL;
 181 }
 182
 183 static void __init pcpu4k_populate_pte(unsigned long addr)
 184 {
 185         populate_extra_pte(addr);
 186 }
 187
 188 static ssize_t __init setup_pcpu_4k(size_t static_size)
 189 {
 190         size_t pages_size;
 191         unsigned int cpu;
 192         int i, j;
 193         ssize_t ret;
 194
 195         pcpu4k_nr_static_pages = PFN_UP(static_size);
 196
 197         /* unaligned allocations can't be freed, round up to page size */
 198         pages_size = PFN_ALIGN(pcpu4k_nr_static_pages * num_possible_cpus()
 199                                * sizeof(pcpu4k_pages[0]));
 200         pcpu4k_pages = alloc_bootmem(pages_size);
 201
 202         /* allocate and copy */
 203         j = 0;
 204         for_each_possible_cpu(cpu)
 205                 for (i = 0; i < pcpu4k_nr_static_pages; i++) {
 206                         void *ptr;
 207
 208                         ptr = pcpu_alloc_bootmem(cpu, PAGE_SIZE, PAGE_SIZE);
 209                         if (!ptr)
 210                                 goto enomem;
 211
 212                         memcpy(ptr, __per_cpu_load + i * PAGE_SIZE, PAGE_SIZE);
 213                         pcpu4k_pages[j++] = virt_to_page(ptr);
 214                 }
 215
 216         /* we're ready, commit */
 217         pr_info("PERCPU: Allocated %d 4k pages, static data %zu bytes\n",
 218                 pcpu4k_nr_static_pages, static_size);
 219
 220         ret = pcpu_setup_first_chunk(pcpu4k_get_page, static_size, 0, 0, NULL,
 221                                      pcpu4k_populate_pte);
 222         goto out_free_ar;
 223
 224 enomem:
 225         while (--j >= 0)
 226                 free_bootmem(__pa(page_address(pcpu4k_pages[j])), PAGE_SIZE);
 227         ret = -ENOMEM;
 228 out_free_ar:
 229         free_bootmem(__pa(pcpu4k_pages), pages_size);
 230         return ret;
 231 }
 232
 233 static inline void setup_percpu_segment(int cpu)
 234 {
 235 #ifdef CONFIG_X86_32
 236         struct desc_struct gdt;
 237
 238         pack_descriptor(&gdt, per_cpu_offset(cpu), 0xFFFFF,
 239                         0x2 | DESCTYPE_S, 0x8);
 240         gdt.s = 1;
 241         write_gdt_entry(get_cpu_gdt_table(cpu),
 242                         GDT_ENTRY_PERCPU, &gdt, DESCTYPE_S);
 243 #endif
 244 }
 245
 246 /*
 247  * Great future plan:
 248  * Declare PDA itself and support (irqstack,tss,pgd) as per cpu data.
 249  * Always point %gs to its beginning
 250  */
 251 void __init setup_per_cpu_areas(void)
 252 {
 253         size_t static_size = __per_cpu_end - __per_cpu_start;
 254         unsigned int cpu;
 255         unsigned long delta;
 256         size_t pcpu_unit_size;
 257         ssize_t ret;
 258
 259         pr_info("NR_CPUS:%d nr_cpumask_bits:%d nr_cpu_ids:%d nr_node_ids:%d\n",
 260                 NR_CPUS, nr_cpumask_bits, nr_cpu_ids, nr_node_ids);
 261
 262         /* allocate percpu area */
 263         ret = setup_pcpu_embed(static_size);
 264         if (ret < 0)
 265                 ret = setup_pcpu_4k(static_size);
 266         if (ret < 0)
 267                 panic("cannot allocate static percpu area (%zu bytes, err=%zd)",
 268                       static_size, ret);
 269
 270         pcpu_unit_size = ret;
 271
 272         /* alrighty, percpu areas up and running */
 273         delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start;
 274         for_each_possible_cpu(cpu) {
 275                 per_cpu_offset(cpu) = delta + cpu * pcpu_unit_size;
 276                 per_cpu(this_cpu_off, cpu) = per_cpu_offset(cpu);
 277                 per_cpu(cpu_number, cpu) = cpu;
 278                 setup_percpu_segment(cpu);
 279                 setup_stack_canary_segment(cpu);
 280                 /*
 281                  * Copy data used in early init routines from the
 282                  * initial arrays to the per cpu data areas.  These
 283                  * arrays then become expendable and the *_early_ptr's
 284                  * are zeroed indicating that the static arrays are
 285                  * gone.
 286                  */
 287 #ifdef CONFIG_X86_LOCAL_APIC
 288                 per_cpu(x86_cpu_to_apicid, cpu) =
 289                         early_per_cpu_map(x86_cpu_to_apicid, cpu);
 290                 per_cpu(x86_bios_cpu_apicid, cpu) =
 291                         early_per_cpu_map(x86_bios_cpu_apicid, cpu);
 292 #endif
 293 #ifdef CONFIG_X86_64
 294                 per_cpu(irq_stack_ptr, cpu) =
 295                         per_cpu(irq_stack_union.irq_stack, cpu) +
 296                         IRQ_STACK_SIZE - 64;
 297 #ifdef CONFIG_NUMA
 298                 per_cpu(x86_cpu_to_node_map, cpu) =
 299                         early_per_cpu_map(x86_cpu_to_node_map, cpu);
 300 #endif
 301 #endif
 302                 /*
 303                  * Up to this point, the boot CPU has been using .data.init
 304                  * area.  Reload any changed state for the boot CPU.
 305                  */
 306                 if (cpu == boot_cpu_id)
 307                         switch_to_new_gdt(cpu);
 308
 309                 DBG("PERCPU: cpu %4d %p\n", cpu, ptr);
 310         }
 311
 312         /* indicate the early static arrays will soon be gone */
 313 #ifdef CONFIG_X86_LOCAL_APIC
 314         early_per_cpu_ptr(x86_cpu_to_apicid) = NULL;
 315         early_per_cpu_ptr(x86_bios_cpu_apicid) = NULL;
 316 #endif
 317 #if defined(CONFIG_X86_64) && defined(CONFIG_NUMA)
 318         early_per_cpu_ptr(x86_cpu_to_node_map) = NULL;
 319 #endif
 320
 321         /* Setup node to cpumask map */
 322         setup_node_to_cpumask_map();
 323
 324         /* Setup cpu initialized, callin, callout masks */
 325         setup_cpu_local_masks();
 326 }