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