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