x86/power/64: Fix kernel text mapping corruption during image restoration
[cascardo/linux.git] / arch / x86 / power / hibernate_64.c
index 009947d..f2b5e6a 100644 (file)
@@ -19,6 +19,7 @@
 #include <asm/mtrr.h>
 #include <asm/sections.h>
 #include <asm/suspend.h>
+#include <asm/tlbflush.h>
 
 /* Defined in hibernate_asm_64.S */
 extern asmlinkage __visible int restore_image(void);
@@ -28,6 +29,7 @@ extern asmlinkage __visible int restore_image(void);
  * kernel's text (this value is passed in the image header).
  */
 unsigned long restore_jump_address __visible;
+unsigned long jump_address_phys;
 
 /*
  * Value of the cr3 register from before the hibernation (this value is passed
@@ -37,7 +39,43 @@ unsigned long restore_cr3 __visible;
 
 pgd_t *temp_level4_pgt __visible;
 
-void *relocated_restore_code __visible;
+unsigned long relocated_restore_code __visible;
+
+static int set_up_temporary_text_mapping(void)
+{
+       pmd_t *pmd;
+       pud_t *pud;
+
+       /*
+        * The new mapping only has to cover the page containing the image
+        * kernel's entry point (jump_address_phys), because the switch over to
+        * it is carried out by relocated code running from a page allocated
+        * specifically for this purpose and covered by the identity mapping, so
+        * the temporary kernel text mapping is only needed for the final jump.
+        * Moreover, in that mapping the virtual address of the image kernel's
+        * entry point must be the same as its virtual address in the image
+        * kernel (restore_jump_address), so the image kernel's
+        * restore_registers() code doesn't find itself in a different area of
+        * the virtual address space after switching over to the original page
+        * tables used by the image kernel.
+        */
+       pud = (pud_t *)get_safe_page(GFP_ATOMIC);
+       if (!pud)
+               return -ENOMEM;
+
+       pmd = (pmd_t *)get_safe_page(GFP_ATOMIC);
+       if (!pmd)
+               return -ENOMEM;
+
+       set_pmd(pmd + pmd_index(restore_jump_address),
+               __pmd((jump_address_phys & PMD_MASK) | __PAGE_KERNEL_LARGE_EXEC));
+       set_pud(pud + pud_index(restore_jump_address),
+               __pud(__pa(pmd) | _KERNPG_TABLE));
+       set_pgd(temp_level4_pgt + pgd_index(restore_jump_address),
+               __pgd(__pa(pud) | _KERNPG_TABLE));
+
+       return 0;
+}
 
 static void *alloc_pgt_page(void *context)
 {
@@ -59,9 +97,10 @@ static int set_up_temporary_mappings(void)
        if (!temp_level4_pgt)
                return -ENOMEM;
 
-       /* It is safe to reuse the original kernel mapping */
-       set_pgd(temp_level4_pgt + pgd_index(__START_KERNEL_map),
-               init_level4_pgt[pgd_index(__START_KERNEL_map)]);
+       /* Prepare a temporary mapping for the kernel text */
+       result = set_up_temporary_text_mapping();
+       if (result)
+               return result;
 
        /* Set up the direct mapping from scratch */
        for (i = 0; i < nr_pfn_mapped; i++) {
@@ -78,19 +117,50 @@ static int set_up_temporary_mappings(void)
        return 0;
 }
 
+static int relocate_restore_code(void)
+{
+       pgd_t *pgd;
+       pud_t *pud;
+
+       relocated_restore_code = get_safe_page(GFP_ATOMIC);
+       if (!relocated_restore_code)
+               return -ENOMEM;
+
+       memcpy((void *)relocated_restore_code, &core_restore_code, PAGE_SIZE);
+
+       /* Make the page containing the relocated code executable */
+       pgd = (pgd_t *)__va(read_cr3()) + pgd_index(relocated_restore_code);
+       pud = pud_offset(pgd, relocated_restore_code);
+       if (pud_large(*pud)) {
+               set_pud(pud, __pud(pud_val(*pud) & ~_PAGE_NX));
+       } else {
+               pmd_t *pmd = pmd_offset(pud, relocated_restore_code);
+
+               if (pmd_large(*pmd)) {
+                       set_pmd(pmd, __pmd(pmd_val(*pmd) & ~_PAGE_NX));
+               } else {
+                       pte_t *pte = pte_offset_kernel(pmd, relocated_restore_code);
+
+                       set_pte(pte, __pte(pte_val(*pte) & ~_PAGE_NX));
+               }
+       }
+       __flush_tlb_all();
+
+       return 0;
+}
+
 int swsusp_arch_resume(void)
 {
        int error;
 
        /* We have got enough memory and from now on we cannot recover */
-       if ((error = set_up_temporary_mappings()))
+       error = set_up_temporary_mappings();
+       if (error)
                return error;
 
-       relocated_restore_code = (void *)get_safe_page(GFP_ATOMIC);
-       if (!relocated_restore_code)
-               return -ENOMEM;
-       memcpy(relocated_restore_code, &core_restore_code,
-              &restore_registers - &core_restore_code);
+       error = relocate_restore_code();
+       if (error)
+               return error;
 
        restore_image();
        return 0;
@@ -109,11 +179,12 @@ int pfn_is_nosave(unsigned long pfn)
 
 struct restore_data_record {
        unsigned long jump_address;
+       unsigned long jump_address_phys;
        unsigned long cr3;
        unsigned long magic;
 };
 
-#define RESTORE_MAGIC  0x0123456789ABCDEFUL
+#define RESTORE_MAGIC  0x123456789ABCDEF0UL
 
 /**
  *     arch_hibernation_header_save - populate the architecture specific part
@@ -126,7 +197,8 @@ int arch_hibernation_header_save(void *addr, unsigned int max_size)
 
        if (max_size < sizeof(struct restore_data_record))
                return -EOVERFLOW;
-       rdr->jump_address = restore_jump_address;
+       rdr->jump_address = (unsigned long)&restore_registers;
+       rdr->jump_address_phys = __pa_symbol(&restore_registers);
        rdr->cr3 = restore_cr3;
        rdr->magic = RESTORE_MAGIC;
        return 0;
@@ -142,6 +214,7 @@ int arch_hibernation_header_restore(void *addr)
        struct restore_data_record *rdr = addr;
 
        restore_jump_address = rdr->jump_address;
+       jump_address_phys = rdr->jump_address_phys;
        restore_cr3 = rdr->cr3;
        return (rdr->magic == RESTORE_MAGIC) ? 0 : -EINVAL;
 }