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 <linux/sched/task.h>
  21 #include <linux/sched/task_stack.h>
  22 #include <asm/sections.h>
  23
  24 enum {
  25         BAD_STACK = -1,
  26         NOT_STACK = 0,
  27         GOOD_FRAME,
  28         GOOD_STACK,
  29 };
  30
  31 /*
  32  * Checks if a given pointer and length is contained by the current
  33  * stack frame (if possible).
  34  *
  35  * Returns:
  36  *      NOT_STACK: not at all on the stack
  37  *      GOOD_FRAME: fully within a valid stack frame
  38  *      GOOD_STACK: fully on the stack (when can't do frame-checking)
  39  *      BAD_STACK: error condition (invalid stack position or bad stack frame)
  40  */
  41 static noinline int check_stack_object(const void *obj, unsigned long len)
  42 {
  43         const void * const stack = task_stack_page(current);
  44         const void * const stackend = stack + THREAD_SIZE;
  45         int ret;
  46
  47         /* Object is not on the stack at all. */
  48         if (obj + len <= stack || stackend <= obj)
  49                 return NOT_STACK;
  50
  51         /*
  52          * Reject: object partially overlaps the stack (passing the
  53          * the check above means at least one end is within the stack,
  54          * so if this check fails, the other end is outside the stack).
  55          */
  56         if (obj < stack || stackend < obj + len)
  57                 return BAD_STACK;
  58
  59         /* Check if object is safely within a valid frame. */
  60         ret = arch_within_stack_frames(stack, stackend, obj, len);
  61         if (ret)
  62                 return ret;
  63
  64         return GOOD_STACK;
  65 }
  66
  67 static void report_usercopy(const void *ptr, unsigned long len,
  68                             bool to_user, const char *type)
  69 {
  70         pr_emerg("kernel memory %s attempt detected %s %p (%s) (%lu bytes)\n",
  71                 to_user ? "exposure" : "overwrite",
  72                 to_user ? "from" : "to", ptr, type ? : "unknown", len);
  73         /*
  74          * For greater effect, it would be nice to do do_group_exit(),
  75          * but BUG() actually hooks all the lock-breaking and per-arch
  76          * Oops code, so that is used here instead.
  77          */
  78         BUG();
  79 }
  80
  81 /* Returns true if any portion of [ptr,ptr+n) over laps with [low,high). */
  82 static bool overlaps(const void *ptr, unsigned long n, unsigned long low,
  83                      unsigned long high)
  84 {
  85         unsigned long check_low = (uintptr_t)ptr;
  86         unsigned long check_high = check_low + n;
  87
  88         /* Does not overlap if entirely above or entirely below. */
  89         if (check_low >= high || check_high <= low)
  90                 return false;
  91
  92         return true;
  93 }
  94
  95 /* Is this address range in the kernel text area? */
  96 static inline const char *check_kernel_text_object(const void *ptr,
  97                                                    unsigned long n)
  98 {
  99         unsigned long textlow = (unsigned long)_stext;
 100         unsigned long texthigh = (unsigned long)_etext;
 101         unsigned long textlow_linear, texthigh_linear;
 102
 103         if (overlaps(ptr, n, textlow, texthigh))
 104                 return "<kernel text>";
 105
 106         /*
 107          * Some architectures have virtual memory mappings with a secondary
 108          * mapping of the kernel text, i.e. there is more than one virtual
 109          * kernel address that points to the kernel image. It is usually
 110          * when there is a separate linear physical memory mapping, in that
 111          * __pa() is not just the reverse of __va(). This can be detected
 112          * and checked:
 113          */
 114         textlow_linear = (unsigned long)lm_alias(textlow);
 115         /* No different mapping: we're done. */
 116         if (textlow_linear == textlow)
 117                 return NULL;
 118
 119         /* Check the secondary mapping... */
 120         texthigh_linear = (unsigned long)lm_alias(texthigh);
 121         if (overlaps(ptr, n, textlow_linear, texthigh_linear))
 122                 return "<linear kernel text>";
 123
 124         return NULL;
 125 }
 126
 127 static inline const char *check_bogus_address(const void *ptr, unsigned long n)
 128 {
 129         /* Reject if object wraps past end of memory. */
 130         if ((unsigned long)ptr + n < (unsigned long)ptr)
 131                 return "<wrapped address>";
 132
 133         /* Reject if NULL or ZERO-allocation. */
 134         if (ZERO_OR_NULL_PTR(ptr))
 135                 return "<null>";
 136
 137         return NULL;
 138 }
 139
 140 /* Checks for allocs that are marked in some way as spanning multiple pages. */
 141 static inline const char *check_page_span(const void *ptr, unsigned long n,
 142                                           struct page *page, bool to_user)
 143 {
 144 #ifdef CONFIG_HARDENED_USERCOPY_PAGESPAN
 145         const void *end = ptr + n - 1;
 146         struct page *endpage;
 147         bool is_reserved, is_cma;
 148
 149         /*
 150          * Sometimes the kernel data regions are not marked Reserved (see
 151          * check below). And sometimes [_sdata,_edata) does not cover
 152          * rodata and/or bss, so check each range explicitly.
 153          */
 154
 155         /* Allow reads of kernel rodata region (if not marked as Reserved). */
 156         if (ptr >= (const void *)__start_rodata &&
 157             end <= (const void *)__end_rodata) {
 158                 if (!to_user)
 159                         return "<rodata>";
 160                 return NULL;
 161         }
 162
 163         /* Allow kernel data region (if not marked as Reserved). */
 164         if (ptr >= (const void *)_sdata && end <= (const void *)_edata)
 165                 return NULL;
 166
 167         /* Allow kernel bss region (if not marked as Reserved). */
 168         if (ptr >= (const void *)__bss_start &&
 169             end <= (const void *)__bss_stop)
 170                 return NULL;
 171
 172         /* Is the object wholly within one base page? */
 173         if (likely(((unsigned long)ptr & (unsigned long)PAGE_MASK) ==
 174                    ((unsigned long)end & (unsigned long)PAGE_MASK)))
 175                 return NULL;
 176
 177         /* Allow if fully inside the same compound (__GFP_COMP) page. */
 178         endpage = virt_to_head_page(end);
 179         if (likely(endpage == page))
 180                 return NULL;
 181
 182         /*
 183          * Reject if range is entirely either Reserved (i.e. special or
 184          * device memory), or CMA. Otherwise, reject since the object spans
 185          * several independently allocated pages.
 186          */
 187         is_reserved = PageReserved(page);
 188         is_cma = is_migrate_cma_page(page);
 189         if (!is_reserved && !is_cma)
 190                 return "<spans multiple pages>";
 191
 192         for (ptr += PAGE_SIZE; ptr <= end; ptr += PAGE_SIZE) {
 193                 page = virt_to_head_page(ptr);
 194                 if (is_reserved && !PageReserved(page))
 195                         return "<spans Reserved and non-Reserved pages>";
 196                 if (is_cma && !is_migrate_cma_page(page))
 197                         return "<spans CMA and non-CMA pages>";
 198         }
 199 #endif
 200
 201         return NULL;
 202 }
 203
 204 static inline const char *check_heap_object(const void *ptr, unsigned long n,
 205                                             bool to_user)
 206 {
 207         struct page *page;
 208
 209         /*
 210          * Some architectures (arm64) return true for virt_addr_valid() on
 211          * vmalloced addresses. Work around this by checking for vmalloc
 212          * first.
 213          *
 214          * We also need to check for module addresses explicitly since we
 215          * may copy static data from modules to userspace
 216          */
 217         if (is_vmalloc_or_module_addr(ptr))
 218                 return NULL;
 219
 220         if (!virt_addr_valid(ptr))
 221                 return NULL;
 222
 223         page = virt_to_head_page(ptr);
 224
 225         /* Check slab allocator for flags and size. */
 226         if (PageSlab(page))
 227                 return __check_heap_object(ptr, n, page);
 228
 229         /* Verify object does not incorrectly span multiple pages. */
 230         return check_page_span(ptr, n, page, to_user);
 231 }
 232
 233 /*
 234  * Validates that the given object is:
 235  * - not bogus address
 236  * - known-safe heap or stack object
 237  * - not in kernel text
 238  */
 239 void __check_object_size(const void *ptr, unsigned long n, bool to_user)
 240 {
 241         const char *err;
 242
 243         /* Skip all tests if size is zero. */
 244         if (!n)
 245                 return;
 246
 247         /* Check for invalid addresses. */
 248         err = check_bogus_address(ptr, n);
 249         if (err)
 250                 goto report;
 251
 252         /* Check for bad heap object. */
 253         err = check_heap_object(ptr, n, to_user);
 254         if (err)
 255                 goto report;
 256
 257         /* Check for bad stack object. */
 258         switch (check_stack_object(ptr, n)) {
 259         case NOT_STACK:
 260                 /* Object is not touching the current process stack. */
 261                 break;
 262         case GOOD_FRAME:
 263         case GOOD_STACK:
 264                 /*
 265                  * Object is either in the correct frame (when it
 266                  * is possible to check) or just generally on the
 267                  * process stack (when frame checking not available).
 268                  */
 269                 return;
 270         default:
 271                 err = "<process stack>";
 272                 goto report;
 273         }
 274
 275         /* Check for object in kernel to avoid text exposure. */
 276         err = check_kernel_text_object(ptr, n);
 277         if (!err)
 278                 return;
 279
 280 report:
 281         report_usercopy(ptr, n, to_user, err);
 282 }
 283 EXPORT_SYMBOL(__check_object_size);