arch/x86/kernel/machine_kexec_64.c

   1 /*
   2  * handle transition of Linux booting another kernel
   3  * Copyright (C) 2002-2005 Eric Biederman  <ebiederm@xmission.com>
   4  *
   5  * This source code is licensed under the GNU General Public License,
   6  * Version 2.  See the file COPYING for more details.
   7  */
   8
   9 #include <linux/mm.h>
  10 #include <linux/kexec.h>
  11 #include <linux/string.h>
  12 #include <linux/reboot.h>
  13 #include <linux/numa.h>
  14 #include <linux/ftrace.h>
  15 #include <linux/io.h>
  16 #include <linux/suspend.h>
  17
  18 #include <asm/pgtable.h>
  19 #include <asm/tlbflush.h>
  20 #include <asm/mmu_context.h>
  21
  22 static int init_one_level2_page(struct kimage *image, pgd_t *pgd,
  23                                 unsigned long addr)
  24 {
  25         pud_t *pud;
  26         pmd_t *pmd;
  27         struct page *page;
  28         int result = -ENOMEM;
  29
  30         addr &= PMD_MASK;
  31         pgd += pgd_index(addr);
  32         if (!pgd_present(*pgd)) {
  33                 page = kimage_alloc_control_pages(image, 0);
  34                 if (!page)
  35                         goto out;
  36                 pud = (pud_t *)page_address(page);
  37                 memset(pud, 0, PAGE_SIZE);
  38                 set_pgd(pgd, __pgd(__pa(pud) | _KERNPG_TABLE));
  39         }
  40         pud = pud_offset(pgd, addr);
  41         if (!pud_present(*pud)) {
  42                 page = kimage_alloc_control_pages(image, 0);
  43                 if (!page)
  44                         goto out;
  45                 pmd = (pmd_t *)page_address(page);
  46                 memset(pmd, 0, PAGE_SIZE);
  47                 set_pud(pud, __pud(__pa(pmd) | _KERNPG_TABLE));
  48         }
  49         pmd = pmd_offset(pud, addr);
  50         if (!pmd_present(*pmd))
  51                 set_pmd(pmd, __pmd(addr | __PAGE_KERNEL_LARGE_EXEC));
  52         result = 0;
  53 out:
  54         return result;
  55 }
  56
  57 static void init_level2_page(pmd_t *level2p, unsigned long addr)
  58 {
  59         unsigned long end_addr;
  60
  61         addr &= PAGE_MASK;
  62         end_addr = addr + PUD_SIZE;
  63         while (addr < end_addr) {
  64                 set_pmd(level2p++, __pmd(addr | __PAGE_KERNEL_LARGE_EXEC));
  65                 addr += PMD_SIZE;
  66         }
  67 }
  68
  69 static int init_level3_page(struct kimage *image, pud_t *level3p,
  70                                 unsigned long addr, unsigned long last_addr)
  71 {
  72         unsigned long end_addr;
  73         int result;
  74
  75         result = 0;
  76         addr &= PAGE_MASK;
  77         end_addr = addr + PGDIR_SIZE;
  78         while ((addr < last_addr) && (addr < end_addr)) {
  79                 struct page *page;
  80                 pmd_t *level2p;
  81
  82                 page = kimage_alloc_control_pages(image, 0);
  83                 if (!page) {
  84                         result = -ENOMEM;
  85                         goto out;
  86                 }
  87                 level2p = (pmd_t *)page_address(page);
  88                 init_level2_page(level2p, addr);
  89                 set_pud(level3p++, __pud(__pa(level2p) | _KERNPG_TABLE));
  90                 addr += PUD_SIZE;
  91         }
  92         /* clear the unused entries */
  93         while (addr < end_addr) {
  94                 pud_clear(level3p++);
  95                 addr += PUD_SIZE;
  96         }
  97 out:
  98         return result;
  99 }
 100
 101
 102 static int init_level4_page(struct kimage *image, pgd_t *level4p,
 103                                 unsigned long addr, unsigned long last_addr)
 104 {
 105         unsigned long end_addr;
 106         int result;
 107
 108         result = 0;
 109         addr &= PAGE_MASK;
 110         end_addr = addr + (PTRS_PER_PGD * PGDIR_SIZE);
 111         while ((addr < last_addr) && (addr < end_addr)) {
 112                 struct page *page;
 113                 pud_t *level3p;
 114
 115                 page = kimage_alloc_control_pages(image, 0);
 116                 if (!page) {
 117                         result = -ENOMEM;
 118                         goto out;
 119                 }
 120                 level3p = (pud_t *)page_address(page);
 121                 result = init_level3_page(image, level3p, addr, last_addr);
 122                 if (result)
 123                         goto out;
 124                 set_pgd(level4p++, __pgd(__pa(level3p) | _KERNPG_TABLE));
 125                 addr += PGDIR_SIZE;
 126         }
 127         /* clear the unused entries */
 128         while (addr < end_addr) {
 129                 pgd_clear(level4p++);
 130                 addr += PGDIR_SIZE;
 131         }
 132 out:
 133         return result;
 134 }
 135
 136 static void free_transition_pgtable(struct kimage *image)
 137 {
 138         free_page((unsigned long)image->arch.pud);
 139         free_page((unsigned long)image->arch.pmd);
 140         free_page((unsigned long)image->arch.pte);
 141 }
 142
 143 static int init_transition_pgtable(struct kimage *image, pgd_t *pgd)
 144 {
 145         pud_t *pud;
 146         pmd_t *pmd;
 147         pte_t *pte;
 148         unsigned long vaddr, paddr;
 149         int result = -ENOMEM;
 150
 151         vaddr = (unsigned long)relocate_kernel;
 152         paddr = __pa(page_address(image->control_code_page)+PAGE_SIZE);
 153         pgd += pgd_index(vaddr);
 154         if (!pgd_present(*pgd)) {
 155                 pud = (pud_t *)get_zeroed_page(GFP_KERNEL);
 156                 if (!pud)
 157                         goto err;
 158                 image->arch.pud = pud;
 159                 set_pgd(pgd, __pgd(__pa(pud) | _KERNPG_TABLE));
 160         }
 161         pud = pud_offset(pgd, vaddr);
 162         if (!pud_present(*pud)) {
 163                 pmd = (pmd_t *)get_zeroed_page(GFP_KERNEL);
 164                 if (!pmd)
 165                         goto err;
 166                 image->arch.pmd = pmd;
 167                 set_pud(pud, __pud(__pa(pmd) | _KERNPG_TABLE));
 168         }
 169         pmd = pmd_offset(pud, vaddr);
 170         if (!pmd_present(*pmd)) {
 171                 pte = (pte_t *)get_zeroed_page(GFP_KERNEL);
 172                 if (!pte)
 173                         goto err;
 174                 image->arch.pte = pte;
 175                 set_pmd(pmd, __pmd(__pa(pte) | _KERNPG_TABLE));
 176         }
 177         pte = pte_offset_kernel(pmd, vaddr);
 178         set_pte(pte, pfn_pte(paddr >> PAGE_SHIFT, PAGE_KERNEL_EXEC));
 179         return 0;
 180 err:
 181         free_transition_pgtable(image);
 182         return result;
 183 }
 184
 185
 186 static int init_pgtable(struct kimage *image, unsigned long start_pgtable)
 187 {
 188         pgd_t *level4p;
 189         int result;
 190         level4p = (pgd_t *)__va(start_pgtable);
 191         result = init_level4_page(image, level4p, 0, max_pfn << PAGE_SHIFT);
 192         if (result)
 193                 return result;
 194         /*
 195          * image->start may be outside 0 ~ max_pfn, for example when
 196          * jump back to original kernel from kexeced kernel
 197          */
 198         result = init_one_level2_page(image, level4p, image->start);
 199         if (result)
 200                 return result;
 201         return init_transition_pgtable(image, level4p);
 202 }
 203
 204 static void set_idt(void *newidt, u16 limit)
 205 {
 206         struct desc_ptr curidt;
 207
 208         /* x86-64 supports unaliged loads & stores */
 209         curidt.size    = limit;
 210         curidt.address = (unsigned long)newidt;
 211
 212         __asm__ __volatile__ (
 213                 "lidtq %0\n"
 214                 : : "m" (curidt)
 215                 );
 216 };
 217
 218
 219 static void set_gdt(void *newgdt, u16 limit)
 220 {
 221         struct desc_ptr curgdt;
 222
 223         /* x86-64 supports unaligned loads & stores */
 224         curgdt.size    = limit;
 225         curgdt.address = (unsigned long)newgdt;
 226
 227         __asm__ __volatile__ (
 228                 "lgdtq %0\n"
 229                 : : "m" (curgdt)
 230                 );
 231 };
 232
 233 static void load_segments(void)
 234 {
 235         __asm__ __volatile__ (
 236                 "\tmovl %0,%%ds\n"
 237                 "\tmovl %0,%%es\n"
 238                 "\tmovl %0,%%ss\n"
 239                 "\tmovl %0,%%fs\n"
 240                 "\tmovl %0,%%gs\n"
 241                 : : "a" (__KERNEL_DS) : "memory"
 242                 );
 243 }
 244
 245 int machine_kexec_prepare(struct kimage *image)
 246 {
 247         unsigned long start_pgtable;
 248         int result;
 249
 250         /* Calculate the offsets */
 251         start_pgtable = page_to_pfn(image->control_code_page) << PAGE_SHIFT;
 252
 253         /* Setup the identity mapped 64bit page table */
 254         result = init_pgtable(image, start_pgtable);
 255         if (result)
 256                 return result;
 257
 258         return 0;
 259 }
 260
 261 void machine_kexec_cleanup(struct kimage *image)
 262 {
 263         free_transition_pgtable(image);
 264 }
 265
 266 /*
 267  * Do not allocate memory (or fail in any way) in machine_kexec().
 268  * We are past the point of no return, committed to rebooting now.
 269  */
 270 void machine_kexec(struct kimage *image)
 271 {
 272         unsigned long page_list[PAGES_NR];
 273         void *control_page;
 274         int save_ftrace_enabled;
 275
 276 #ifdef CONFIG_KEXEC_JUMP
 277         if (image->preserve_context)
 278                 save_processor_state();
 279 #endif
 280
 281         save_ftrace_enabled = __ftrace_enabled_save();
 282
 283         /* Interrupts aren't acceptable while we reboot */
 284         local_irq_disable();
 285
 286         if (image->preserve_context) {
 287 #ifdef CONFIG_X86_IO_APIC
 288                 /*
 289                  * We need to put APICs in legacy mode so that we can
 290                  * get timer interrupts in second kernel. kexec/kdump
 291                  * paths already have calls to disable_IO_APIC() in
 292                  * one form or other. kexec jump path also need
 293                  * one.
 294                  */
 295                 disable_IO_APIC();
 296 #endif
 297         }
 298
 299         control_page = page_address(image->control_code_page) + PAGE_SIZE;
 300         memcpy(control_page, relocate_kernel, KEXEC_CONTROL_CODE_MAX_SIZE);
 301
 302         page_list[PA_CONTROL_PAGE] = virt_to_phys(control_page);
 303         page_list[VA_CONTROL_PAGE] = (unsigned long)control_page;
 304         page_list[PA_TABLE_PAGE] =
 305           (unsigned long)__pa(page_address(image->control_code_page));
 306
 307         if (image->type == KEXEC_TYPE_DEFAULT)
 308                 page_list[PA_SWAP_PAGE] = (page_to_pfn(image->swap_page)
 309                                                 << PAGE_SHIFT);
 310
 311         /*
 312          * The segment registers are funny things, they have both a
 313          * visible and an invisible part.  Whenever the visible part is
 314          * set to a specific selector, the invisible part is loaded
 315          * with from a table in memory.  At no other time is the
 316          * descriptor table in memory accessed.
 317          *
 318          * I take advantage of this here by force loading the
 319          * segments, before I zap the gdt with an invalid value.
 320          */
 321         load_segments();
 322         /*
 323          * The gdt & idt are now invalid.
 324          * If you want to load them you must set up your own idt & gdt.
 325          */
 326         set_gdt(phys_to_virt(0), 0);
 327         set_idt(phys_to_virt(0), 0);
 328
 329         /* now call it */
 330         image->start = relocate_kernel((unsigned long)image->head,
 331                                        (unsigned long)page_list,
 332                                        image->start,
 333                                        image->preserve_context);
 334
 335 #ifdef CONFIG_KEXEC_JUMP
 336         if (image->preserve_context)
 337                 restore_processor_state();
 338 #endif
 339
 340         __ftrace_enabled_restore(save_ftrace_enabled);
 341 }
 342
 343 void arch_crash_save_vmcoreinfo(void)
 344 {
 345         VMCOREINFO_SYMBOL(phys_base);
 346         VMCOREINFO_SYMBOL(init_level4_pgt);
 347
 348 #ifdef CONFIG_NUMA
 349         VMCOREINFO_SYMBOL(node_data);
 350         VMCOREINFO_LENGTH(node_data, MAX_NUMNODES);
 351 #endif
 352 }
 353