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 <linux/thread_info.h>
  23 #include <linux/atomic.h>
  24 #include <linux/jump_label.h>
  25 #include <asm/sections.h>
  26
  27 /*
  28  * Checks if a given pointer and length is contained by the current
  29  * stack frame (if possible).
  30  *
  31  * Returns:
  32  *      NOT_STACK: not at all on the stack
  33  *      GOOD_FRAME: fully within a valid stack frame
  34  *      GOOD_STACK: fully on the stack (when can't do frame-checking)
  35  *      BAD_STACK: error condition (invalid stack position or bad stack frame)
  36  */
  37 static noinline int check_stack_object(const void *obj, unsigned long len)
  38 {
  39         const void * const stack = task_stack_page(current);
  40         const void * const stackend = stack + THREAD_SIZE;
  41         int ret;
  42
  43         /* Object is not on the stack at all. */
  44         if (obj + len <= stack || stackend <= obj)
  45                 return NOT_STACK;
  46
  47         /*
  48          * Reject: object partially overlaps the stack (passing the
  49          * the check above means at least one end is within the stack,
  50          * so if this check fails, the other end is outside the stack).
  51          */
  52         if (obj < stack || stackend < obj + len)
  53                 return BAD_STACK;
  54
  55         /* Check if object is safely within a valid frame. */
  56         ret = arch_within_stack_frames(stack, stackend, obj, len);
  57         if (ret)
  58                 return ret;
  59
  60         return GOOD_STACK;
  61 }
  62
  63 /*
  64  * If these functions are reached, then CONFIG_HARDENED_USERCOPY has found
  65  * an unexpected state during a copy_from_user() or copy_to_user() call.
  66  * There are several checks being performed on the buffer by the
  67  * __check_object_size() function. Normal stack buffer usage should never
  68  * trip the checks, and kernel text addressing will always trip the check.
  69  * For cache objects, it is checking that only the whitelisted range of
  70  * bytes for a given cache is being accessed (via the cache's usersize and
  71  * useroffset fields). To adjust a cache whitelist, use the usercopy-aware
  72  * kmem_cache_create_usercopy() function to create the cache (and
  73  * carefully audit the whitelist range).
  74  */
  75 void usercopy_warn(const char *name, const char *detail, bool to_user,
  76                    unsigned long offset, unsigned long len)
  77 {
  78         WARN_ONCE(1, "Bad or missing usercopy whitelist? Kernel memory %s attempt detected %s %s%s%s%s (offset %lu, size %lu)!\n",
  79                  to_user ? "exposure" : "overwrite",
  80                  to_user ? "from" : "to",
  81                  name ? : "unknown?!",
  82                  detail ? " '" : "", detail ? : "", detail ? "'" : "",
  83                  offset, len);
  84 }
  85
  86 void __noreturn usercopy_abort(const char *name, const char *detail,
  87                                bool to_user, unsigned long offset,
  88                                unsigned long len)
  89 {
  90         pr_emerg("Kernel memory %s attempt detected %s %s%s%s%s (offset %lu, size %lu)!\n",
  91                  to_user ? "exposure" : "overwrite",
  92                  to_user ? "from" : "to",
  93                  name ? : "unknown?!",
  94                  detail ? " '" : "", detail ? : "", detail ? "'" : "",
  95                  offset, len);
  96
  97         /*
  98          * For greater effect, it would be nice to do do_group_exit(),
  99          * but BUG() actually hooks all the lock-breaking and per-arch
 100          * Oops code, so that is used here instead.
 101          */
 102         BUG();
 103 }
 104
 105 /* Returns true if any portion of [ptr,ptr+n) over laps with [low,high). */
 106 static bool overlaps(const unsigned long ptr, unsigned long n,
 107                      unsigned long low, unsigned long high)
 108 {
 109         const unsigned long check_low = ptr;
 110         unsigned long check_high = check_low + n;
 111
 112         /* Does not overlap if entirely above or entirely below. */
 113         if (check_low >= high || check_high <= low)
 114                 return false;
 115
 116         return true;
 117 }
 118
 119 /* Is this address range in the kernel text area? */
 120 static inline void check_kernel_text_object(const unsigned long ptr,
 121                                             unsigned long n, bool to_user)
 122 {
 123         unsigned long textlow = (unsigned long)_stext;
 124         unsigned long texthigh = (unsigned long)_etext;
 125         unsigned long textlow_linear, texthigh_linear;
 126
 127         if (overlaps(ptr, n, textlow, texthigh))
 128                 usercopy_abort("kernel text", NULL, to_user, ptr - textlow, n);
 129
 130         /*
 131          * Some architectures have virtual memory mappings with a secondary
 132          * mapping of the kernel text, i.e. there is more than one virtual
 133          * kernel address that points to the kernel image. It is usually
 134          * when there is a separate linear physical memory mapping, in that
 135          * __pa() is not just the reverse of __va(). This can be detected
 136          * and checked:
 137          */
 138         textlow_linear = (unsigned long)lm_alias(textlow);
 139         /* No different mapping: we're done. */
 140         if (textlow_linear == textlow)
 141                 return;
 142
 143         /* Check the secondary mapping... */
 144         texthigh_linear = (unsigned long)lm_alias(texthigh);
 145         if (overlaps(ptr, n, textlow_linear, texthigh_linear))
 146                 usercopy_abort("linear kernel text", NULL, to_user,
 147                                ptr - textlow_linear, n);
 148 }
 149
 150 static inline void check_bogus_address(const unsigned long ptr, unsigned long n,
 151                                        bool to_user)
 152 {
 153         /* Reject if object wraps past end of memory. */
 154         if (ptr + n < ptr)
 155                 usercopy_abort("wrapped address", NULL, to_user, 0, ptr + n);
 156
 157         /* Reject if NULL or ZERO-allocation. */
 158         if (ZERO_OR_NULL_PTR(ptr))
 159                 usercopy_abort("null address", NULL, to_user, ptr, n);
 160 }
 161
 162 /* Checks for allocs that are marked in some way as spanning multiple pages. */
 163 static inline void check_page_span(const void *ptr, unsigned long n,
 164                                    struct page *page, bool to_user)
 165 {
 166 #ifdef CONFIG_HARDENED_USERCOPY_PAGESPAN
 167         const void *end = ptr + n - 1;
 168         struct page *endpage;
 169         bool is_reserved, is_cma;
 170
 171         /*
 172          * Sometimes the kernel data regions are not marked Reserved (see
 173          * check below). And sometimes [_sdata,_edata) does not cover
 174          * rodata and/or bss, so check each range explicitly.
 175          */
 176
 177         /* Allow reads of kernel rodata region (if not marked as Reserved). */
 178         if (ptr >= (const void *)__start_rodata &&
 179             end <= (const void *)__end_rodata) {
 180                 if (!to_user)
 181                         usercopy_abort("rodata", NULL, to_user, 0, n);
 182                 return;
 183         }
 184
 185         /* Allow kernel data region (if not marked as Reserved). */
 186         if (ptr >= (const void *)_sdata && end <= (const void *)_edata)
 187                 return;
 188
 189         /* Allow kernel bss region (if not marked as Reserved). */
 190         if (ptr >= (const void *)__bss_start &&
 191             end <= (const void *)__bss_stop)
 192                 return;
 193
 194         /* Is the object wholly within one base page? */
 195         if (likely(((unsigned long)ptr & (unsigned long)PAGE_MASK) ==
 196                    ((unsigned long)end & (unsigned long)PAGE_MASK)))
 197                 return;
 198
 199         /* Allow if fully inside the same compound (__GFP_COMP) page. */
 200         endpage = virt_to_head_page(end);
 201         if (likely(endpage == page))
 202                 return;
 203
 204         /*
 205          * Reject if range is entirely either Reserved (i.e. special or
 206          * device memory), or CMA. Otherwise, reject since the object spans
 207          * several independently allocated pages.
 208          */
 209         is_reserved = PageReserved(page);
 210         is_cma = is_migrate_cma_page(page);
 211         if (!is_reserved && !is_cma)
 212                 usercopy_abort("spans multiple pages", NULL, to_user, 0, n);
 213
 214         for (ptr += PAGE_SIZE; ptr <= end; ptr += PAGE_SIZE) {
 215                 page = virt_to_head_page(ptr);
 216                 if (is_reserved && !PageReserved(page))
 217                         usercopy_abort("spans Reserved and non-Reserved pages",
 218                                        NULL, to_user, 0, n);
 219                 if (is_cma && !is_migrate_cma_page(page))
 220                         usercopy_abort("spans CMA and non-CMA pages", NULL,
 221                                        to_user, 0, n);
 222         }
 223 #endif
 224 }
 225
 226 static inline void check_heap_object(const void *ptr, unsigned long n,
 227                                      bool to_user)
 228 {
 229         struct page *page;
 230
 231         if (!virt_addr_valid(ptr))
 232                 return;
 233
 234         page = virt_to_head_page(ptr);
 235
 236         if (PageSlab(page)) {
 237                 /* Check slab allocator for flags and size. */
 238                 __check_heap_object(ptr, n, page, to_user);
 239         } else {
 240                 /* Verify object does not incorrectly span multiple pages. */
 241                 check_page_span(ptr, n, page, to_user);
 242         }
 243 }
 244
 245 static DEFINE_STATIC_KEY_FALSE_RO(bypass_usercopy_checks);
 246
 247 /*
 248  * Validates that the given object is:
 249  * - not bogus address
 250  * - fully contained by stack (or stack frame, when available)
 251  * - fully within SLAB object (or object whitelist area, when available)
 252  * - not in kernel text
 253  */
 254 void __check_object_size(const void *ptr, unsigned long n, bool to_user)
 255 {
 256         if (static_branch_unlikely(&bypass_usercopy_checks))
 257                 return;
 258
 259         /* Skip all tests if size is zero. */
 260         if (!n)
 261                 return;
 262
 263         /* Check for invalid addresses. */
 264         check_bogus_address((const unsigned long)ptr, n, to_user);
 265
 266         /* Check for bad stack object. */
 267         switch (check_stack_object(ptr, n)) {
 268         case NOT_STACK:
 269                 /* Object is not touching the current process stack. */
 270                 break;
 271         case GOOD_FRAME:
 272         case GOOD_STACK:
 273                 /*
 274                  * Object is either in the correct frame (when it
 275                  * is possible to check) or just generally on the
 276                  * process stack (when frame checking not available).
 277                  */
 278                 return;
 279         default:
 280                 usercopy_abort("process stack", NULL, to_user, 0, n);
 281         }
 282
 283         /* Check for bad heap object. */
 284         check_heap_object(ptr, n, to_user);
 285
 286         /* Check for object in kernel to avoid text exposure. */
 287         check_kernel_text_object((const unsigned long)ptr, n, to_user);
 288 }
 289 EXPORT_SYMBOL(__check_object_size);
 290
 291 static bool enable_checks __initdata = true;
 292
 293 static int __init parse_hardened_usercopy(char *str)
 294 {
 295         return strtobool(str, &enable_checks);
 296 }
 297
 298 __setup("hardened_usercopy=", parse_hardened_usercopy);
 299
 300 static int __init set_hardened_usercopy(void)
 301 {
 302         if (enable_checks == false)
 303                 static_branch_enable(&bypass_usercopy_checks);
 304         return 1;
 305 }
 306
 307 late_initcall(set_hardened_usercopy);