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