mm/usercopy.c

   1 /*
   2  * This implements the various checks for CONFIG_HARDENED_USERCOPY*,
   3  * which are designed to protect kernel memory from needless exposure
   4  * and overwrite under many unintended conditions. This code is based
   5  * on PAX_USERCOPY, which is:
   6  *
   7  * Copyright (C) 2001-2016 PaX Team, Bradley Spengler, Open Source
   8  * Security Inc.
   9  *
  10  * This program is free software; you can redistribute it and/or modify
  11  * it under the terms of the GNU General Public License version 2 as
  12  * published by the Free Software Foundation.
  13  *
  14  */
  15 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  16
  17 #include <linux/mm.h>
  18 #include <linux/slab.h>
  19 #include <asm/sections.h>
  20
  21 enum {
  22         BAD_STACK = -1,
  23         NOT_STACK = 0,
  24         GOOD_FRAME,
  25         GOOD_STACK,
  26 };
  27
  28 /*
  29  * Checks if a given pointer and length is contained by the current
  30  * stack frame (if possible).
  31  *
  32  * Returns:
  33  *      NOT_STACK: not at all on the stack
  34  *      GOOD_FRAME: fully within a valid stack frame
  35  *      GOOD_STACK: fully on the stack (when can't do frame-checking)
  36  *      BAD_STACK: error condition (invalid stack position or bad stack frame)
  37  */
  38 static noinline int check_stack_object(const void *obj, unsigned long len)
  39 {
  40         const void * const stack = task_stack_page(current);
  41         const void * const stackend = stack + THREAD_SIZE;
  42         int ret;
  43
  44         /* Object is not on the stack at all. */
  45         if (obj + len <= stack || stackend <= obj)
  46                 return NOT_STACK;
  47
  48         /*
  49          * Reject: object partially overlaps the stack (passing the
  50          * the check above means at least one end is within the stack,
  51          * so if this check fails, the other end is outside the stack).
  52          */
  53         if (obj < stack || stackend < obj + len)
  54                 return BAD_STACK;
  55
  56         /* Check if object is safely within a valid frame. */
  57         ret = arch_within_stack_frames(stack, stackend, obj, len);
  58         if (ret)
  59                 return ret;
  60
  61         return GOOD_STACK;
  62 }
  63
  64 static void report_usercopy(const void *ptr, unsigned long len,
  65                             bool to_user, const char *type)
  66 {
  67         pr_emerg("kernel memory %s attempt detected %s %p (%s) (%lu bytes)\n",
  68                 to_user ? "exposure" : "overwrite",
  69                 to_user ? "from" : "to", ptr, type ? : "unknown", len);
  70         /*
  71          * For greater effect, it would be nice to do do_group_exit(),
  72          * but BUG() actually hooks all the lock-breaking and per-arch
  73          * Oops code, so that is used here instead.
  74          */
  75         BUG();
  76 }
  77
  78 /* Returns true if any portion of [ptr,ptr+n) over laps with [low,high). */
  79 static bool overlaps(const void *ptr, unsigned long n, unsigned long low,
  80                      unsigned long high)
  81 {
  82         unsigned long check_low = (uintptr_t)ptr;
  83         unsigned long check_high = check_low + n;
  84
  85         /* Does not overlap if entirely above or entirely below. */
  86         if (check_low >= high || check_high <= low)
  87                 return false;
  88
  89         return true;
  90 }
  91
  92 /* Is this address range in the kernel text area? */
  93 static inline const char *check_kernel_text_object(const void *ptr,
  94                                                    unsigned long n)
  95 {
  96         unsigned long textlow = (unsigned long)_stext;
  97         unsigned long texthigh = (unsigned long)_etext;
  98         unsigned long textlow_linear, texthigh_linear;
  99
 100         if (overlaps(ptr, n, textlow, texthigh))
 101                 return "<kernel text>";
 102
 103         /*
 104          * Some architectures have virtual memory mappings with a secondary
 105          * mapping of the kernel text, i.e. there is more than one virtual
 106          * kernel address that points to the kernel image. It is usually
 107          * when there is a separate linear physical memory mapping, in that
 108          * __pa() is not just the reverse of __va(). This can be detected
 109          * and checked:
 110          */
 111         textlow_linear = (unsigned long)__va(__pa(textlow));
 112         /* No different mapping: we're done. */
 113         if (textlow_linear == textlow)
 114                 return NULL;
 115
 116         /* Check the secondary mapping... */
 117         texthigh_linear = (unsigned long)__va(__pa(texthigh));
 118         if (overlaps(ptr, n, textlow_linear, texthigh_linear))
 119                 return "<linear kernel text>";
 120
 121         return NULL;
 122 }
 123
 124 static inline const char *check_bogus_address(const void *ptr, unsigned long n)
 125 {
 126         /* Reject if object wraps past end of memory. */
 127         if ((unsigned long)ptr + (n - 1) < (unsigned long)ptr)
 128                 return "<wrapped address>";
 129
 130         /* Reject if NULL or ZERO-allocation. */
 131         if (ZERO_OR_NULL_PTR(ptr))
 132                 return "<null>";
 133
 134         return NULL;
 135 }
 136
 137 /* Checks for allocs that are marked in some way as spanning multiple pages. */
 138 static inline const char *check_page_span(const void *ptr, unsigned long n,
 139                                           struct page *page, bool to_user)
 140 {
 141 #ifdef CONFIG_HARDENED_USERCOPY_PAGESPAN
 142         const void *end = ptr + n - 1;
 143         struct page *endpage;
 144         bool is_reserved, is_cma;
 145
 146         /*
 147          * Sometimes the kernel data regions are not marked Reserved (see
 148          * check below). And sometimes [_sdata,_edata) does not cover
 149          * rodata and/or bss, so check each range explicitly.
 150          */
 151
 152         /* Allow reads of kernel rodata region (if not marked as Reserved). */
 153         if (ptr >= (const void *)__start_rodata &&
 154             end <= (const void *)__end_rodata) {
 155                 if (!to_user)
 156                         return "<rodata>";
 157                 return NULL;
 158         }
 159
 160         /* Allow kernel data region (if not marked as Reserved). */
 161         if (ptr >= (const void *)_sdata && end <= (const void *)_edata)
 162                 return NULL;
 163
 164         /* Allow kernel bss region (if not marked as Reserved). */
 165         if (ptr >= (const void *)__bss_start &&
 166             end <= (const void *)__bss_stop)
 167                 return NULL;
 168
 169         /* Is the object wholly within one base page? */
 170         if (likely(((unsigned long)ptr & (unsigned long)PAGE_MASK) ==
 171                    ((unsigned long)end & (unsigned long)PAGE_MASK)))
 172                 return NULL;
 173
 174         /* Allow if fully inside the same compound (__GFP_COMP) page. */
 175         endpage = virt_to_head_page(end);
 176         if (likely(endpage == page))
 177                 return NULL;
 178
 179         /*
 180          * Reject if range is entirely either Reserved (i.e. special or
 181          * device memory), or CMA. Otherwise, reject since the object spans
 182          * several independently allocated pages.
 183          */
 184         is_reserved = PageReserved(page);
 185         is_cma = is_migrate_cma_page(page);
 186         if (!is_reserved && !is_cma)
 187                 return "<spans multiple pages>";
 188
 189         for (ptr += PAGE_SIZE; ptr <= end; ptr += PAGE_SIZE) {
 190                 page = virt_to_head_page(ptr);
 191                 if (is_reserved && !PageReserved(page))
 192                         return "<spans Reserved and non-Reserved pages>";
 193                 if (is_cma && !is_migrate_cma_page(page))
 194                         return "<spans CMA and non-CMA pages>";
 195         }
 196 #endif
 197
 198         return NULL;
 199 }
 200
 201 static inline const char *check_heap_object(const void *ptr, unsigned long n,
 202                                             bool to_user)
 203 {
 204         struct page *page;
 205
 206         /*
 207          * Some architectures (arm64) return true for virt_addr_valid() on
 208          * vmalloced addresses. Work around this by checking for vmalloc
 209          * first.
 210          *
 211          * We also need to check for module addresses explicitly since we
 212          * may copy static data from modules to userspace
 213          */
 214         if (is_vmalloc_or_module_addr(ptr))
 215                 return NULL;
 216
 217         if (!virt_addr_valid(ptr))
 218                 return NULL;
 219
 220         page = virt_to_head_page(ptr);
 221
 222         /* Check slab allocator for flags and size. */
 223         if (PageSlab(page))
 224                 return __check_heap_object(ptr, n, page);
 225
 226         /* Verify object does not incorrectly span multiple pages. */
 227         return check_page_span(ptr, n, page, to_user);
 228 }
 229
 230 /*
 231  * Validates that the given object is:
 232  * - not bogus address
 233  * - known-safe heap or stack object
 234  * - not in kernel text
 235  */
 236 void __check_object_size(const void *ptr, unsigned long n, bool to_user)
 237 {
 238         const char *err;
 239
 240         /* Skip all tests if size is zero. */
 241         if (!n)
 242                 return;
 243
 244         /* Check for invalid addresses. */
 245         err = check_bogus_address(ptr, n);
 246         if (err)
 247                 goto report;
 248
 249         /* Check for bad heap object. */
 250         err = check_heap_object(ptr, n, to_user);
 251         if (err)
 252                 goto report;
 253
 254         /* Check for bad stack object. */
 255         switch (check_stack_object(ptr, n)) {
 256         case NOT_STACK:
 257                 /* Object is not touching the current process stack. */
 258                 break;
 259         case GOOD_FRAME:
 260         case GOOD_STACK:
 261                 /*
 262                  * Object is either in the correct frame (when it
 263                  * is possible to check) or just generally on the
 264                  * process stack (when frame checking not available).
 265                  */
 266                 return;
 267         default:
 268                 err = "<process stack>";
 269                 goto report;
 270         }
 271
 272         /* Check for object in kernel to avoid text exposure. */
 273         err = check_kernel_text_object(ptr, n);
 274         if (!err)
 275                 return;
 276
 277 report:
 278         report_usercopy(ptr, n, to_user, err);
 279 }
 280 EXPORT_SYMBOL(__check_object_size);