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