1 // SPDX-License-Identifier: GPL-2.0-only
3 * This implements the various checks for CONFIG_HARDENED_USERCOPY*,
4 * which are designed to protect kernel memory from needless exposure
5 * and overwrite under many unintended conditions. This code is based
6 * on PAX_USERCOPY, which is:
8 * Copyright (C) 2001-2016 PaX Team, Bradley Spengler, Open Source
11 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
14 #include <linux/highmem.h>
15 #include <linux/slab.h>
16 #include <linux/sched.h>
17 #include <linux/sched/task.h>
18 #include <linux/sched/task_stack.h>
19 #include <linux/thread_info.h>
20 #include <linux/atomic.h>
21 #include <linux/jump_label.h>
22 #include <asm/sections.h>
26 * Checks if a given pointer and length is contained by the current
27 * stack frame (if possible).
30 * NOT_STACK: not at all on the stack
31 * GOOD_FRAME: fully within a valid stack frame
32 * GOOD_STACK: fully on the stack (when can't do frame-checking)
33 * BAD_STACK: error condition (invalid stack position or bad stack frame)
35 static noinline int check_stack_object(const void *obj, unsigned long len)
37 const void * const stack = task_stack_page(current);
38 const void * const stackend = stack + THREAD_SIZE;
41 /* Object is not on the stack at all. */
42 if (obj + len <= stack || stackend <= obj)
46 * Reject: object partially overlaps the stack (passing the
47 * check above means at least one end is within the stack,
48 * so if this check fails, the other end is outside the stack).
50 if (obj < stack || stackend < obj + len)
53 /* Check if object is safely within a valid frame. */
54 ret = arch_within_stack_frames(stack, stackend, obj, len);
62 * If these functions are reached, then CONFIG_HARDENED_USERCOPY has found
63 * an unexpected state during a copy_from_user() or copy_to_user() call.
64 * There are several checks being performed on the buffer by the
65 * __check_object_size() function. Normal stack buffer usage should never
66 * trip the checks, and kernel text addressing will always trip the check.
67 * For cache objects, it is checking that only the whitelisted range of
68 * bytes for a given cache is being accessed (via the cache's usersize and
69 * useroffset fields). To adjust a cache whitelist, use the usercopy-aware
70 * kmem_cache_create_usercopy() function to create the cache (and
71 * carefully audit the whitelist range).
73 void __noreturn usercopy_abort(const char *name, const char *detail,
74 bool to_user, unsigned long offset,
77 pr_emerg("Kernel memory %s attempt detected %s %s%s%s%s (offset %lu, size %lu)!\n",
78 to_user ? "exposure" : "overwrite",
79 to_user ? "from" : "to",
81 detail ? " '" : "", detail ? : "", detail ? "'" : "",
85 * For greater effect, it would be nice to do do_group_exit(),
86 * but BUG() actually hooks all the lock-breaking and per-arch
87 * Oops code, so that is used here instead.
92 /* Returns true if any portion of [ptr,ptr+n) over laps with [low,high). */
93 static bool overlaps(const unsigned long ptr, unsigned long n,
94 unsigned long low, unsigned long high)
96 const unsigned long check_low = ptr;
97 unsigned long check_high = check_low + n;
99 /* Does not overlap if entirely above or entirely below. */
100 if (check_low >= high || check_high <= low)
106 /* Is this address range in the kernel text area? */
107 static inline void check_kernel_text_object(const unsigned long ptr,
108 unsigned long n, bool to_user)
110 unsigned long textlow = (unsigned long)_stext;
111 unsigned long texthigh = (unsigned long)_etext;
112 unsigned long textlow_linear, texthigh_linear;
114 if (overlaps(ptr, n, textlow, texthigh))
115 usercopy_abort("kernel text", NULL, to_user, ptr - textlow, n);
118 * Some architectures have virtual memory mappings with a secondary
119 * mapping of the kernel text, i.e. there is more than one virtual
120 * kernel address that points to the kernel image. It is usually
121 * when there is a separate linear physical memory mapping, in that
122 * __pa() is not just the reverse of __va(). This can be detected
125 textlow_linear = (unsigned long)lm_alias(textlow);
126 /* No different mapping: we're done. */
127 if (textlow_linear == textlow)
130 /* Check the secondary mapping... */
131 texthigh_linear = (unsigned long)lm_alias(texthigh);
132 if (overlaps(ptr, n, textlow_linear, texthigh_linear))
133 usercopy_abort("linear kernel text", NULL, to_user,
134 ptr - textlow_linear, n);
137 static inline void check_bogus_address(const unsigned long ptr, unsigned long n,
140 /* Reject if object wraps past end of memory. */
141 if (ptr + (n - 1) < ptr)
142 usercopy_abort("wrapped address", NULL, to_user, 0, ptr + n);
144 /* Reject if NULL or ZERO-allocation. */
145 if (ZERO_OR_NULL_PTR(ptr))
146 usercopy_abort("null address", NULL, to_user, ptr, n);
149 /* Checks for allocs that are marked in some way as spanning multiple pages. */
150 static inline void check_page_span(const void *ptr, unsigned long n,
151 struct page *page, bool to_user)
153 #ifdef CONFIG_HARDENED_USERCOPY_PAGESPAN
154 const void *end = ptr + n - 1;
155 struct page *endpage;
156 bool is_reserved, is_cma;
159 * Sometimes the kernel data regions are not marked Reserved (see
160 * check below). And sometimes [_sdata,_edata) does not cover
161 * rodata and/or bss, so check each range explicitly.
164 /* Allow reads of kernel rodata region (if not marked as Reserved). */
165 if (ptr >= (const void *)__start_rodata &&
166 end <= (const void *)__end_rodata) {
168 usercopy_abort("rodata", NULL, to_user, 0, n);
172 /* Allow kernel data region (if not marked as Reserved). */
173 if (ptr >= (const void *)_sdata && end <= (const void *)_edata)
176 /* Allow kernel bss region (if not marked as Reserved). */
177 if (ptr >= (const void *)__bss_start &&
178 end <= (const void *)__bss_stop)
181 /* Is the object wholly within one base page? */
182 if (likely(((unsigned long)ptr & (unsigned long)PAGE_MASK) ==
183 ((unsigned long)end & (unsigned long)PAGE_MASK)))
186 /* Allow if fully inside the same compound (__GFP_COMP) page. */
187 endpage = virt_to_head_page(end);
188 if (likely(endpage == page))
192 * Reject if range is entirely either Reserved (i.e. special or
193 * device memory), or CMA. Otherwise, reject since the object spans
194 * several independently allocated pages.
196 is_reserved = PageReserved(page);
197 is_cma = is_migrate_cma_page(page);
198 if (!is_reserved && !is_cma)
199 usercopy_abort("spans multiple pages", NULL, to_user, 0, n);
201 for (ptr += PAGE_SIZE; ptr <= end; ptr += PAGE_SIZE) {
202 page = virt_to_head_page(ptr);
203 if (is_reserved && !PageReserved(page))
204 usercopy_abort("spans Reserved and non-Reserved pages",
205 NULL, to_user, 0, n);
206 if (is_cma && !is_migrate_cma_page(page))
207 usercopy_abort("spans CMA and non-CMA pages", NULL,
213 static inline void check_heap_object(const void *ptr, unsigned long n,
218 if (!virt_addr_valid(ptr))
222 * When CONFIG_HIGHMEM=y, kmap_to_page() will give either the
223 * highmem page or fallback to virt_to_page(). The following
224 * is effectively a highmem-aware virt_to_slab().
226 folio = page_folio(kmap_to_page((void *)ptr));
228 if (folio_test_slab(folio)) {
229 /* Check slab allocator for flags and size. */
230 __check_heap_object(ptr, n, folio_slab(folio), to_user);
232 /* Verify object does not incorrectly span multiple pages. */
233 check_page_span(ptr, n, folio_page(folio, 0), to_user);
237 static DEFINE_STATIC_KEY_FALSE_RO(bypass_usercopy_checks);
240 * Validates that the given object is:
241 * - not bogus address
242 * - fully contained by stack (or stack frame, when available)
243 * - fully within SLAB object (or object whitelist area, when available)
244 * - not in kernel text
246 void __check_object_size(const void *ptr, unsigned long n, bool to_user)
248 if (static_branch_unlikely(&bypass_usercopy_checks))
251 /* Skip all tests if size is zero. */
255 /* Check for invalid addresses. */
256 check_bogus_address((const unsigned long)ptr, n, to_user);
258 /* Check for bad stack object. */
259 switch (check_stack_object(ptr, n)) {
261 /* Object is not touching the current process stack. */
266 * Object is either in the correct frame (when it
267 * is possible to check) or just generally on the
268 * process stack (when frame checking not available).
272 usercopy_abort("process stack", NULL, to_user, 0, n);
275 /* Check for bad heap object. */
276 check_heap_object(ptr, n, to_user);
278 /* Check for object in kernel to avoid text exposure. */
279 check_kernel_text_object((const unsigned long)ptr, n, to_user);
281 EXPORT_SYMBOL(__check_object_size);
283 static bool enable_checks __initdata = true;
285 static int __init parse_hardened_usercopy(char *str)
287 return strtobool(str, &enable_checks);
290 __setup("hardened_usercopy=", parse_hardened_usercopy);
292 static int __init set_hardened_usercopy(void)
294 if (enable_checks == false)
295 static_branch_enable(&bypass_usercopy_checks);
299 late_initcall(set_hardened_usercopy);