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