Commit | Line | Data |
---|---|---|
d2912cb1 | 1 | // SPDX-License-Identifier: GPL-2.0-only |
f5509cc1 KC |
2 | /* |
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: | |
7 | * | |
8 | * Copyright (C) 2001-2016 PaX Team, Bradley Spengler, Open Source | |
9 | * Security Inc. | |
f5509cc1 KC |
10 | */ |
11 | #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt | |
12 | ||
13 | #include <linux/mm.h> | |
314eed30 | 14 | #include <linux/highmem.h> |
f5509cc1 | 15 | #include <linux/slab.h> |
5b825c3a | 16 | #include <linux/sched.h> |
29930025 IM |
17 | #include <linux/sched/task.h> |
18 | #include <linux/sched/task_stack.h> | |
96dc4f9f | 19 | #include <linux/thread_info.h> |
b5cb15d9 CR |
20 | #include <linux/atomic.h> |
21 | #include <linux/jump_label.h> | |
f5509cc1 KC |
22 | #include <asm/sections.h> |
23 | ||
f5509cc1 KC |
24 | /* |
25 | * Checks if a given pointer and length is contained by the current | |
26 | * stack frame (if possible). | |
27 | * | |
28 | * Returns: | |
29 | * NOT_STACK: not at all on the stack | |
30 | * GOOD_FRAME: fully within a valid stack frame | |
31 | * GOOD_STACK: fully on the stack (when can't do frame-checking) | |
32 | * BAD_STACK: error condition (invalid stack position or bad stack frame) | |
33 | */ | |
34 | static noinline int check_stack_object(const void *obj, unsigned long len) | |
35 | { | |
36 | const void * const stack = task_stack_page(current); | |
37 | const void * const stackend = stack + THREAD_SIZE; | |
38 | int ret; | |
39 | ||
40 | /* Object is not on the stack at all. */ | |
41 | if (obj + len <= stack || stackend <= obj) | |
42 | return NOT_STACK; | |
43 | ||
44 | /* | |
45 | * Reject: object partially overlaps the stack (passing the | |
5ce1be0e | 46 | * check above means at least one end is within the stack, |
f5509cc1 KC |
47 | * so if this check fails, the other end is outside the stack). |
48 | */ | |
49 | if (obj < stack || stackend < obj + len) | |
50 | return BAD_STACK; | |
51 | ||
52 | /* Check if object is safely within a valid frame. */ | |
53 | ret = arch_within_stack_frames(stack, stackend, obj, len); | |
54 | if (ret) | |
55 | return ret; | |
56 | ||
57 | return GOOD_STACK; | |
58 | } | |
59 | ||
b394d468 | 60 | /* |
afcc90f8 KC |
61 | * If these functions are reached, then CONFIG_HARDENED_USERCOPY has found |
62 | * an unexpected state during a copy_from_user() or copy_to_user() call. | |
b394d468 KC |
63 | * There are several checks being performed on the buffer by the |
64 | * __check_object_size() function. Normal stack buffer usage should never | |
65 | * trip the checks, and kernel text addressing will always trip the check. | |
afcc90f8 KC |
66 | * For cache objects, it is checking that only the whitelisted range of |
67 | * bytes for a given cache is being accessed (via the cache's usersize and | |
68 | * useroffset fields). To adjust a cache whitelist, use the usercopy-aware | |
69 | * kmem_cache_create_usercopy() function to create the cache (and | |
70 | * carefully audit the whitelist range). | |
b394d468 | 71 | */ |
afcc90f8 KC |
72 | void usercopy_warn(const char *name, const char *detail, bool to_user, |
73 | unsigned long offset, unsigned long len) | |
74 | { | |
75 | WARN_ONCE(1, "Bad or missing usercopy whitelist? Kernel memory %s attempt detected %s %s%s%s%s (offset %lu, size %lu)!\n", | |
76 | to_user ? "exposure" : "overwrite", | |
77 | to_user ? "from" : "to", | |
78 | name ? : "unknown?!", | |
79 | detail ? " '" : "", detail ? : "", detail ? "'" : "", | |
80 | offset, len); | |
81 | } | |
82 | ||
b394d468 KC |
83 | void __noreturn usercopy_abort(const char *name, const char *detail, |
84 | bool to_user, unsigned long offset, | |
85 | unsigned long len) | |
f5509cc1 | 86 | { |
b394d468 KC |
87 | pr_emerg("Kernel memory %s attempt detected %s %s%s%s%s (offset %lu, size %lu)!\n", |
88 | to_user ? "exposure" : "overwrite", | |
89 | to_user ? "from" : "to", | |
90 | name ? : "unknown?!", | |
91 | detail ? " '" : "", detail ? : "", detail ? "'" : "", | |
92 | offset, len); | |
93 | ||
f5509cc1 KC |
94 | /* |
95 | * For greater effect, it would be nice to do do_group_exit(), | |
96 | * but BUG() actually hooks all the lock-breaking and per-arch | |
97 | * Oops code, so that is used here instead. | |
98 | */ | |
99 | BUG(); | |
100 | } | |
101 | ||
102 | /* Returns true if any portion of [ptr,ptr+n) over laps with [low,high). */ | |
f4e6e289 KC |
103 | static bool overlaps(const unsigned long ptr, unsigned long n, |
104 | unsigned long low, unsigned long high) | |
f5509cc1 | 105 | { |
f4e6e289 | 106 | const unsigned long check_low = ptr; |
f5509cc1 KC |
107 | unsigned long check_high = check_low + n; |
108 | ||
109 | /* Does not overlap if entirely above or entirely below. */ | |
94cd97af | 110 | if (check_low >= high || check_high <= low) |
f5509cc1 KC |
111 | return false; |
112 | ||
113 | return true; | |
114 | } | |
115 | ||
116 | /* Is this address range in the kernel text area? */ | |
f4e6e289 KC |
117 | static inline void check_kernel_text_object(const unsigned long ptr, |
118 | unsigned long n, bool to_user) | |
f5509cc1 KC |
119 | { |
120 | unsigned long textlow = (unsigned long)_stext; | |
121 | unsigned long texthigh = (unsigned long)_etext; | |
122 | unsigned long textlow_linear, texthigh_linear; | |
123 | ||
124 | if (overlaps(ptr, n, textlow, texthigh)) | |
f4e6e289 | 125 | usercopy_abort("kernel text", NULL, to_user, ptr - textlow, n); |
f5509cc1 KC |
126 | |
127 | /* | |
128 | * Some architectures have virtual memory mappings with a secondary | |
129 | * mapping of the kernel text, i.e. there is more than one virtual | |
130 | * kernel address that points to the kernel image. It is usually | |
131 | * when there is a separate linear physical memory mapping, in that | |
132 | * __pa() is not just the reverse of __va(). This can be detected | |
133 | * and checked: | |
134 | */ | |
46f6236a | 135 | textlow_linear = (unsigned long)lm_alias(textlow); |
f5509cc1 KC |
136 | /* No different mapping: we're done. */ |
137 | if (textlow_linear == textlow) | |
f4e6e289 | 138 | return; |
f5509cc1 KC |
139 | |
140 | /* Check the secondary mapping... */ | |
46f6236a | 141 | texthigh_linear = (unsigned long)lm_alias(texthigh); |
f5509cc1 | 142 | if (overlaps(ptr, n, textlow_linear, texthigh_linear)) |
f4e6e289 KC |
143 | usercopy_abort("linear kernel text", NULL, to_user, |
144 | ptr - textlow_linear, n); | |
f5509cc1 KC |
145 | } |
146 | ||
f4e6e289 KC |
147 | static inline void check_bogus_address(const unsigned long ptr, unsigned long n, |
148 | bool to_user) | |
f5509cc1 KC |
149 | { |
150 | /* Reject if object wraps past end of memory. */ | |
95153169 | 151 | if (ptr + (n - 1) < ptr) |
f4e6e289 | 152 | usercopy_abort("wrapped address", NULL, to_user, 0, ptr + n); |
f5509cc1 KC |
153 | |
154 | /* Reject if NULL or ZERO-allocation. */ | |
155 | if (ZERO_OR_NULL_PTR(ptr)) | |
f4e6e289 | 156 | usercopy_abort("null address", NULL, to_user, ptr, n); |
f5509cc1 KC |
157 | } |
158 | ||
8e1f74ea | 159 | /* Checks for allocs that are marked in some way as spanning multiple pages. */ |
f4e6e289 KC |
160 | static inline void check_page_span(const void *ptr, unsigned long n, |
161 | struct page *page, bool to_user) | |
f5509cc1 | 162 | { |
8e1f74ea | 163 | #ifdef CONFIG_HARDENED_USERCOPY_PAGESPAN |
f5509cc1 | 164 | const void *end = ptr + n - 1; |
8e1f74ea | 165 | struct page *endpage; |
f5509cc1 KC |
166 | bool is_reserved, is_cma; |
167 | ||
f5509cc1 KC |
168 | /* |
169 | * Sometimes the kernel data regions are not marked Reserved (see | |
170 | * check below). And sometimes [_sdata,_edata) does not cover | |
171 | * rodata and/or bss, so check each range explicitly. | |
172 | */ | |
173 | ||
174 | /* Allow reads of kernel rodata region (if not marked as Reserved). */ | |
175 | if (ptr >= (const void *)__start_rodata && | |
176 | end <= (const void *)__end_rodata) { | |
177 | if (!to_user) | |
f4e6e289 KC |
178 | usercopy_abort("rodata", NULL, to_user, 0, n); |
179 | return; | |
f5509cc1 KC |
180 | } |
181 | ||
182 | /* Allow kernel data region (if not marked as Reserved). */ | |
183 | if (ptr >= (const void *)_sdata && end <= (const void *)_edata) | |
f4e6e289 | 184 | return; |
f5509cc1 KC |
185 | |
186 | /* Allow kernel bss region (if not marked as Reserved). */ | |
187 | if (ptr >= (const void *)__bss_start && | |
188 | end <= (const void *)__bss_stop) | |
f4e6e289 | 189 | return; |
f5509cc1 KC |
190 | |
191 | /* Is the object wholly within one base page? */ | |
192 | if (likely(((unsigned long)ptr & (unsigned long)PAGE_MASK) == | |
193 | ((unsigned long)end & (unsigned long)PAGE_MASK))) | |
f4e6e289 | 194 | return; |
f5509cc1 | 195 | |
8e1f74ea | 196 | /* Allow if fully inside the same compound (__GFP_COMP) page. */ |
f5509cc1 KC |
197 | endpage = virt_to_head_page(end); |
198 | if (likely(endpage == page)) | |
f4e6e289 | 199 | return; |
f5509cc1 KC |
200 | |
201 | /* | |
202 | * Reject if range is entirely either Reserved (i.e. special or | |
203 | * device memory), or CMA. Otherwise, reject since the object spans | |
204 | * several independently allocated pages. | |
205 | */ | |
206 | is_reserved = PageReserved(page); | |
207 | is_cma = is_migrate_cma_page(page); | |
208 | if (!is_reserved && !is_cma) | |
f4e6e289 | 209 | usercopy_abort("spans multiple pages", NULL, to_user, 0, n); |
f5509cc1 KC |
210 | |
211 | for (ptr += PAGE_SIZE; ptr <= end; ptr += PAGE_SIZE) { | |
212 | page = virt_to_head_page(ptr); | |
213 | if (is_reserved && !PageReserved(page)) | |
f4e6e289 KC |
214 | usercopy_abort("spans Reserved and non-Reserved pages", |
215 | NULL, to_user, 0, n); | |
f5509cc1 | 216 | if (is_cma && !is_migrate_cma_page(page)) |
f4e6e289 KC |
217 | usercopy_abort("spans CMA and non-CMA pages", NULL, |
218 | to_user, 0, n); | |
f5509cc1 | 219 | } |
8e1f74ea | 220 | #endif |
8e1f74ea KC |
221 | } |
222 | ||
f4e6e289 KC |
223 | static inline void check_heap_object(const void *ptr, unsigned long n, |
224 | bool to_user) | |
8e1f74ea KC |
225 | { |
226 | struct page *page; | |
227 | ||
8e1f74ea | 228 | if (!virt_addr_valid(ptr)) |
f4e6e289 | 229 | return; |
8e1f74ea | 230 | |
314eed30 KC |
231 | /* |
232 | * When CONFIG_HIGHMEM=y, kmap_to_page() will give either the | |
233 | * highmem page or fallback to virt_to_page(). The following | |
234 | * is effectively a highmem-aware virt_to_head_page(). | |
235 | */ | |
236 | page = compound_head(kmap_to_page((void *)ptr)); | |
8e1f74ea | 237 | |
f4e6e289 KC |
238 | if (PageSlab(page)) { |
239 | /* Check slab allocator for flags and size. */ | |
240 | __check_heap_object(ptr, n, page, to_user); | |
241 | } else { | |
242 | /* Verify object does not incorrectly span multiple pages. */ | |
243 | check_page_span(ptr, n, page, to_user); | |
244 | } | |
f5509cc1 KC |
245 | } |
246 | ||
b5cb15d9 CR |
247 | static DEFINE_STATIC_KEY_FALSE_RO(bypass_usercopy_checks); |
248 | ||
f5509cc1 KC |
249 | /* |
250 | * Validates that the given object is: | |
251 | * - not bogus address | |
7bff3c06 QC |
252 | * - fully contained by stack (or stack frame, when available) |
253 | * - fully within SLAB object (or object whitelist area, when available) | |
f5509cc1 KC |
254 | * - not in kernel text |
255 | */ | |
256 | void __check_object_size(const void *ptr, unsigned long n, bool to_user) | |
257 | { | |
b5cb15d9 CR |
258 | if (static_branch_unlikely(&bypass_usercopy_checks)) |
259 | return; | |
260 | ||
f5509cc1 KC |
261 | /* Skip all tests if size is zero. */ |
262 | if (!n) | |
263 | return; | |
264 | ||
265 | /* Check for invalid addresses. */ | |
f4e6e289 | 266 | check_bogus_address((const unsigned long)ptr, n, to_user); |
f5509cc1 | 267 | |
f5509cc1 KC |
268 | /* Check for bad stack object. */ |
269 | switch (check_stack_object(ptr, n)) { | |
270 | case NOT_STACK: | |
271 | /* Object is not touching the current process stack. */ | |
272 | break; | |
273 | case GOOD_FRAME: | |
274 | case GOOD_STACK: | |
275 | /* | |
276 | * Object is either in the correct frame (when it | |
277 | * is possible to check) or just generally on the | |
278 | * process stack (when frame checking not available). | |
279 | */ | |
280 | return; | |
281 | default: | |
f4e6e289 | 282 | usercopy_abort("process stack", NULL, to_user, 0, n); |
f5509cc1 KC |
283 | } |
284 | ||
7bff3c06 QC |
285 | /* Check for bad heap object. */ |
286 | check_heap_object(ptr, n, to_user); | |
287 | ||
f5509cc1 | 288 | /* Check for object in kernel to avoid text exposure. */ |
f4e6e289 | 289 | check_kernel_text_object((const unsigned long)ptr, n, to_user); |
f5509cc1 KC |
290 | } |
291 | EXPORT_SYMBOL(__check_object_size); | |
b5cb15d9 CR |
292 | |
293 | static bool enable_checks __initdata = true; | |
294 | ||
295 | static int __init parse_hardened_usercopy(char *str) | |
296 | { | |
297 | return strtobool(str, &enable_checks); | |
298 | } | |
299 | ||
300 | __setup("hardened_usercopy=", parse_hardened_usercopy); | |
301 | ||
302 | static int __init set_hardened_usercopy(void) | |
303 | { | |
304 | if (enable_checks == false) | |
305 | static_branch_enable(&bypass_usercopy_checks); | |
306 | return 1; | |
307 | } | |
308 | ||
309 | late_initcall(set_hardened_usercopy); |