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 <linux/atomic.h>
  25 #include <linux/jump_label.h>
  26 #include <asm/sections.h>
  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 /*
  65  * If these functions are reached, then CONFIG_HARDENED_USERCOPY has found
  66  * an unexpected state during a copy_from_user() or copy_to_user() call.
  67  * There are several checks being performed on the buffer by the
  68  * __check_object_size() function. Normal stack buffer usage should never
  69  * trip the checks, and kernel text addressing will always trip the check.
  70  * For cache objects, it is checking that only the whitelisted range of
  71  * bytes for a given cache is being accessed (via the cache's usersize and
  72  * useroffset fields). To adjust a cache whitelist, use the usercopy-aware
  73  * kmem_cache_create_usercopy() function to create the cache (and
  74  * carefully audit the whitelist range).
  75  */
  76 void usercopy_warn(const char *name, const char *detail, bool to_user,
  77                    unsigned long offset, unsigned long len)
  78 {
  79         WARN_ONCE(1, "Bad or missing usercopy whitelist? Kernel memory %s attempt detected %s %s%s%s%s (offset %lu, size %lu)!\n",
  80                  to_user ? "exposure" : "overwrite",
  81                  to_user ? "from" : "to",
  82                  name ? : "unknown?!",
  83                  detail ? " '" : "", detail ? : "", detail ? "'" : "",
  84                  offset, len);
  85 }
  86
  87 void __noreturn usercopy_abort(const char *name, const char *detail,
  88                                bool to_user, unsigned long offset,
  89                                unsigned long len)
  90 {
  91         pr_emerg("Kernel memory %s attempt detected %s %s%s%s%s (offset %lu, size %lu)!\n",
  92                  to_user ? "exposure" : "overwrite",
  93                  to_user ? "from" : "to",
  94                  name ? : "unknown?!",
  95                  detail ? " '" : "", detail ? : "", detail ? "'" : "",
  96                  offset, len);
  97
  98         /*
  99          * For greater effect, it would be nice to do do_group_exit(),
 100          * but BUG() actually hooks all the lock-breaking and per-arch
 101          * Oops code, so that is used here instead.
 102          */
 103         BUG();
 104 }
 105
 106 /* Returns true if any portion of [ptr,ptr+n) over laps with [low,high). */
 107 static bool overlaps(const unsigned long ptr, unsigned long n,
 108                      unsigned long low, unsigned long high)
 109 {
 110         const unsigned long check_low = ptr;
 111         unsigned long check_high = check_low + n;
 112
 113         /* Does not overlap if entirely above or entirely below. */
 114         if (check_low >= high || check_high <= low)
 115                 return false;
 116
 117         return true;
 118 }
 119
 120 /* Is this address range in the kernel text area? */
 121 static inline void check_kernel_text_object(const unsigned long ptr,
 122                                             unsigned long n, bool to_user)
 123 {
 124         unsigned long textlow = (unsigned long)_stext;
 125         unsigned long texthigh = (unsigned long)_etext;
 126         unsigned long textlow_linear, texthigh_linear;
 127
 128         if (overlaps(ptr, n, textlow, texthigh))
 129                 usercopy_abort("kernel text", NULL, to_user, ptr - textlow, n);
 130
 131         /*
 132          * Some architectures have virtual memory mappings with a secondary
 133          * mapping of the kernel text, i.e. there is more than one virtual
 134          * kernel address that points to the kernel image. It is usually
 135          * when there is a separate linear physical memory mapping, in that
 136          * __pa() is not just the reverse of __va(). This can be detected
 137          * and checked:
 138          */
 139         textlow_linear = (unsigned long)lm_alias(textlow);
 140         /* No different mapping: we're done. */
 141         if (textlow_linear == textlow)
 142                 return;
 143
 144         /* Check the secondary mapping... */
 145         texthigh_linear = (unsigned long)lm_alias(texthigh);
 146         if (overlaps(ptr, n, textlow_linear, texthigh_linear))
 147                 usercopy_abort("linear kernel text", NULL, to_user,
 148                                ptr - textlow_linear, n);
 149 }
 150
 151 static inline void check_bogus_address(const unsigned long ptr, unsigned long n,
 152                                        bool to_user)
 153 {
 154         /* Reject if object wraps past end of memory. */
 155         if (ptr + (n - 1) < ptr)
 156                 usercopy_abort("wrapped address", NULL, to_user, 0, ptr + n);
 157
 158         /* Reject if NULL or ZERO-allocation. */
 159         if (ZERO_OR_NULL_PTR(ptr))
 160                 usercopy_abort("null address", NULL, to_user, ptr, n);
 161 }
 162
 163 /* Checks for allocs that are marked in some way as spanning multiple pages. */
 164 static inline void check_page_span(const void *ptr, unsigned long n,
 165                                    struct page *page, bool to_user)
 166 {
 167 #ifdef CONFIG_HARDENED_USERCOPY_PAGESPAN
 168         const void *end = ptr + n - 1;
 169         struct page *endpage;
 170         bool is_reserved, is_cma;
 171
 172         /*
 173          * Sometimes the kernel data regions are not marked Reserved (see
 174          * check below). And sometimes [_sdata,_edata) does not cover
 175          * rodata and/or bss, so check each range explicitly.
 176          */
 177
 178         /* Allow reads of kernel rodata region (if not marked as Reserved). */
 179         if (ptr >= (const void *)__start_rodata &&
 180             end <= (const void *)__end_rodata) {
 181                 if (!to_user)
 182                         usercopy_abort("rodata", NULL, to_user, 0, n);
 183                 return;
 184         }
 185
 186         /* Allow kernel data region (if not marked as Reserved). */
 187         if (ptr >= (const void *)_sdata && end <= (const void *)_edata)
 188                 return;
 189
 190         /* Allow kernel bss region (if not marked as Reserved). */
 191         if (ptr >= (const void *)__bss_start &&
 192             end <= (const void *)__bss_stop)
 193                 return;
 194
 195         /* Is the object wholly within one base page? */
 196         if (likely(((unsigned long)ptr & (unsigned long)PAGE_MASK) ==
 197                    ((unsigned long)end & (unsigned long)PAGE_MASK)))
 198                 return;
 199
 200         /* Allow if fully inside the same compound (__GFP_COMP) page. */
 201         endpage = virt_to_head_page(end);
 202         if (likely(endpage == page))
 203                 return;
 204
 205         /*
 206          * Reject if range is entirely either Reserved (i.e. special or
 207          * device memory), or CMA. Otherwise, reject since the object spans
 208          * several independently allocated pages.
 209          */
 210         is_reserved = PageReserved(page);
 211         is_cma = is_migrate_cma_page(page);
 212         if (!is_reserved && !is_cma)
 213                 usercopy_abort("spans multiple pages", NULL, to_user, 0, n);
 214
 215         for (ptr += PAGE_SIZE; ptr <= end; ptr += PAGE_SIZE) {
 216                 page = virt_to_head_page(ptr);
 217                 if (is_reserved && !PageReserved(page))
 218                         usercopy_abort("spans Reserved and non-Reserved pages",
 219                                        NULL, to_user, 0, n);
 220                 if (is_cma && !is_migrate_cma_page(page))
 221                         usercopy_abort("spans CMA and non-CMA pages", NULL,
 222                                        to_user, 0, n);
 223         }
 224 #endif
 225 }
 226
 227 static inline void check_heap_object(const void *ptr, unsigned long n,
 228                                      bool to_user)
 229 {
 230         struct page *page;
 231
 232         if (!virt_addr_valid(ptr))
 233                 return;
 234
 235         /*
 236          * When CONFIG_HIGHMEM=y, kmap_to_page() will give either the
 237          * highmem page or fallback to virt_to_page(). The following
 238          * is effectively a highmem-aware virt_to_head_page().
 239          */
 240         page = compound_head(kmap_to_page((void *)ptr));
 241
 242         if (PageSlab(page)) {
 243                 /* Check slab allocator for flags and size. */
 244                 __check_heap_object(ptr, n, page, to_user);
 245         } else {
 246                 /* Verify object does not incorrectly span multiple pages. */
 247                 check_page_span(ptr, n, page, to_user);
 248         }
 249 }
 250
 251 static DEFINE_STATIC_KEY_FALSE_RO(bypass_usercopy_checks);
 252
 253 /*
 254  * Validates that the given object is:
 255  * - not bogus address
 256  * - fully contained by stack (or stack frame, when available)
 257  * - fully within SLAB object (or object whitelist area, when available)
 258  * - not in kernel text
 259  */
 260 void __check_object_size(const void *ptr, unsigned long n, bool to_user)
 261 {
 262         if (static_branch_unlikely(&bypass_usercopy_checks))
 263                 return;
 264
 265         /* Skip all tests if size is zero. */
 266         if (!n)
 267                 return;
 268
 269         /* Check for invalid addresses. */
 270         check_bogus_address((const unsigned long)ptr, n, to_user);
 271
 272         /* Check for bad stack object. */
 273         switch (check_stack_object(ptr, n)) {
 274         case NOT_STACK:
 275                 /* Object is not touching the current process stack. */
 276                 break;
 277         case GOOD_FRAME:
 278         case GOOD_STACK:
 279                 /*
 280                  * Object is either in the correct frame (when it
 281                  * is possible to check) or just generally on the
 282                  * process stack (when frame checking not available).
 283                  */
 284                 return;
 285         default:
 286                 usercopy_abort("process stack", NULL, to_user, 0, n);
 287         }
 288
 289         /* Check for bad heap object. */
 290         check_heap_object(ptr, n, to_user);
 291
 292         /* Check for object in kernel to avoid text exposure. */
 293         check_kernel_text_object((const unsigned long)ptr, n, to_user);
 294 }
 295 EXPORT_SYMBOL(__check_object_size);
 296
 297 static bool enable_checks __initdata = true;
 298
 299 static int __init parse_hardened_usercopy(char *str)
 300 {
 301         return strtobool(str, &enable_checks);
 302 }
 303
 304 __setup("hardened_usercopy=", parse_hardened_usercopy);
 305
 306 static int __init set_hardened_usercopy(void)
 307 {
 308         if (enable_checks == false)
 309                 static_branch_enable(&bypass_usercopy_checks);
 310         return 1;
 311 }
 312
 313 late_initcall(set_hardened_usercopy);