1# SPDX-License-Identifier: GPL-2.0-only 2menu "Kernel hardening options" 3 4config GCC_PLUGIN_STRUCTLEAK 5 bool 6 help 7 While the kernel is built with warnings enabled for any missed 8 stack variable initializations, this warning is silenced for 9 anything passed by reference to another function, under the 10 occasionally misguided assumption that the function will do 11 the initialization. As this regularly leads to exploitable 12 flaws, this plugin is available to identify and zero-initialize 13 such variables, depending on the chosen level of coverage. 14 15 This plugin was originally ported from grsecurity/PaX. More 16 information at: 17 * https://grsecurity.net/ 18 * https://pax.grsecurity.net/ 19 20menu "Memory initialization" 21 22config CC_HAS_AUTO_VAR_INIT_PATTERN 23 def_bool $(cc-option,-ftrivial-auto-var-init=pattern) 24 25config CC_HAS_AUTO_VAR_INIT_ZERO_BARE 26 def_bool $(cc-option,-ftrivial-auto-var-init=zero) 27 28config CC_HAS_AUTO_VAR_INIT_ZERO_ENABLER 29 # Clang 16 and later warn about using the -enable flag, but it 30 # is required before then. 31 def_bool $(cc-option,-ftrivial-auto-var-init=zero -enable-trivial-auto-var-init-zero-knowing-it-will-be-removed-from-clang) 32 depends on !CC_HAS_AUTO_VAR_INIT_ZERO_BARE 33 34config CC_HAS_AUTO_VAR_INIT_ZERO 35 def_bool CC_HAS_AUTO_VAR_INIT_ZERO_BARE || CC_HAS_AUTO_VAR_INIT_ZERO_ENABLER 36 37choice 38 prompt "Initialize kernel stack variables at function entry" 39 default GCC_PLUGIN_STRUCTLEAK_BYREF_ALL if COMPILE_TEST && GCC_PLUGINS 40 default INIT_STACK_ALL_PATTERN if COMPILE_TEST && CC_HAS_AUTO_VAR_INIT_PATTERN 41 default INIT_STACK_ALL_ZERO if CC_HAS_AUTO_VAR_INIT_ZERO 42 default INIT_STACK_NONE 43 help 44 This option enables initialization of stack variables at 45 function entry time. This has the possibility to have the 46 greatest coverage (since all functions can have their 47 variables initialized), but the performance impact depends 48 on the function calling complexity of a given workload's 49 syscalls. 50 51 This chooses the level of coverage over classes of potentially 52 uninitialized variables. The selected class of variable will be 53 initialized before use in a function. 54 55 config INIT_STACK_NONE 56 bool "no automatic stack variable initialization (weakest)" 57 help 58 Disable automatic stack variable initialization. 59 This leaves the kernel vulnerable to the standard 60 classes of uninitialized stack variable exploits 61 and information exposures. 62 63 config GCC_PLUGIN_STRUCTLEAK_USER 64 bool "zero-init structs marked for userspace (weak)" 65 # Plugin can be removed once the kernel only supports GCC 12+ 66 depends on GCC_PLUGINS && !CC_HAS_AUTO_VAR_INIT_ZERO 67 select GCC_PLUGIN_STRUCTLEAK 68 help 69 Zero-initialize any structures on the stack containing 70 a __user attribute. This can prevent some classes of 71 uninitialized stack variable exploits and information 72 exposures, like CVE-2013-2141: 73 https://git.kernel.org/linus/b9e146d8eb3b9eca 74 75 config GCC_PLUGIN_STRUCTLEAK_BYREF 76 bool "zero-init structs passed by reference (strong)" 77 # Plugin can be removed once the kernel only supports GCC 12+ 78 depends on GCC_PLUGINS && !CC_HAS_AUTO_VAR_INIT_ZERO 79 depends on !(KASAN && KASAN_STACK) 80 select GCC_PLUGIN_STRUCTLEAK 81 help 82 Zero-initialize any structures on the stack that may 83 be passed by reference and had not already been 84 explicitly initialized. This can prevent most classes 85 of uninitialized stack variable exploits and information 86 exposures, like CVE-2017-1000410: 87 https://git.kernel.org/linus/06e7e776ca4d3654 88 89 As a side-effect, this keeps a lot of variables on the 90 stack that can otherwise be optimized out, so combining 91 this with CONFIG_KASAN_STACK can lead to a stack overflow 92 and is disallowed. 93 94 config GCC_PLUGIN_STRUCTLEAK_BYREF_ALL 95 bool "zero-init everything passed by reference (very strong)" 96 # Plugin can be removed once the kernel only supports GCC 12+ 97 depends on GCC_PLUGINS && !CC_HAS_AUTO_VAR_INIT_ZERO 98 depends on !(KASAN && KASAN_STACK) 99 select GCC_PLUGIN_STRUCTLEAK 100 help 101 Zero-initialize any stack variables that may be passed 102 by reference and had not already been explicitly 103 initialized. This is intended to eliminate all classes 104 of uninitialized stack variable exploits and information 105 exposures. 106 107 As a side-effect, this keeps a lot of variables on the 108 stack that can otherwise be optimized out, so combining 109 this with CONFIG_KASAN_STACK can lead to a stack overflow 110 and is disallowed. 111 112 config INIT_STACK_ALL_PATTERN 113 bool "pattern-init everything (strongest)" 114 depends on CC_HAS_AUTO_VAR_INIT_PATTERN 115 depends on !KMSAN 116 help 117 Initializes everything on the stack (including padding) 118 with a specific debug value. This is intended to eliminate 119 all classes of uninitialized stack variable exploits and 120 information exposures, even variables that were warned about 121 having been left uninitialized. 122 123 Pattern initialization is known to provoke many existing bugs 124 related to uninitialized locals, e.g. pointers receive 125 non-NULL values, buffer sizes and indices are very big. The 126 pattern is situation-specific; Clang on 64-bit uses 0xAA 127 repeating for all types and padding except float and double 128 which use 0xFF repeating (-NaN). Clang on 32-bit uses 0xFF 129 repeating for all types and padding. 130 131 config INIT_STACK_ALL_ZERO 132 bool "zero-init everything (strongest and safest)" 133 depends on CC_HAS_AUTO_VAR_INIT_ZERO 134 depends on !KMSAN 135 help 136 Initializes everything on the stack (including padding) 137 with a zero value. This is intended to eliminate all 138 classes of uninitialized stack variable exploits and 139 information exposures, even variables that were warned 140 about having been left uninitialized. 141 142 Zero initialization provides safe defaults for strings 143 (immediately NUL-terminated), pointers (NULL), indices 144 (index 0), and sizes (0 length), so it is therefore more 145 suitable as a production security mitigation than pattern 146 initialization. 147 148endchoice 149 150config GCC_PLUGIN_STRUCTLEAK_VERBOSE 151 bool "Report forcefully initialized variables" 152 depends on GCC_PLUGIN_STRUCTLEAK 153 depends on !COMPILE_TEST # too noisy 154 help 155 This option will cause a warning to be printed each time the 156 structleak plugin finds a variable it thinks needs to be 157 initialized. Since not all existing initializers are detected 158 by the plugin, this can produce false positive warnings. 159 160config GCC_PLUGIN_STACKLEAK 161 bool "Poison kernel stack before returning from syscalls" 162 depends on GCC_PLUGINS 163 depends on HAVE_ARCH_STACKLEAK 164 help 165 This option makes the kernel erase the kernel stack before 166 returning from system calls. This has the effect of leaving 167 the stack initialized to the poison value, which both reduces 168 the lifetime of any sensitive stack contents and reduces 169 potential for uninitialized stack variable exploits or information 170 exposures (it does not cover functions reaching the same stack 171 depth as prior functions during the same syscall). This blocks 172 most uninitialized stack variable attacks, with the performance 173 impact being driven by the depth of the stack usage, rather than 174 the function calling complexity. 175 176 The performance impact on a single CPU system kernel compilation 177 sees a 1% slowdown, other systems and workloads may vary and you 178 are advised to test this feature on your expected workload before 179 deploying it. 180 181 This plugin was ported from grsecurity/PaX. More information at: 182 * https://grsecurity.net/ 183 * https://pax.grsecurity.net/ 184 185config GCC_PLUGIN_STACKLEAK_VERBOSE 186 bool "Report stack depth analysis instrumentation" if EXPERT 187 depends on GCC_PLUGIN_STACKLEAK 188 depends on !COMPILE_TEST # too noisy 189 help 190 This option will cause a warning to be printed each time the 191 stackleak plugin finds a function it thinks needs to be 192 instrumented. This is useful for comparing coverage between 193 builds. 194 195config STACKLEAK_TRACK_MIN_SIZE 196 int "Minimum stack frame size of functions tracked by STACKLEAK" 197 default 100 198 range 0 4096 199 depends on GCC_PLUGIN_STACKLEAK 200 help 201 The STACKLEAK gcc plugin instruments the kernel code for tracking 202 the lowest border of the kernel stack (and for some other purposes). 203 It inserts the stackleak_track_stack() call for the functions with 204 a stack frame size greater than or equal to this parameter. 205 If unsure, leave the default value 100. 206 207config STACKLEAK_METRICS 208 bool "Show STACKLEAK metrics in the /proc file system" 209 depends on GCC_PLUGIN_STACKLEAK 210 depends on PROC_FS 211 help 212 If this is set, STACKLEAK metrics for every task are available in 213 the /proc file system. In particular, /proc/<pid>/stack_depth 214 shows the maximum kernel stack consumption for the current and 215 previous syscalls. Although this information is not precise, it 216 can be useful for estimating the STACKLEAK performance impact for 217 your workloads. 218 219config STACKLEAK_RUNTIME_DISABLE 220 bool "Allow runtime disabling of kernel stack erasing" 221 depends on GCC_PLUGIN_STACKLEAK 222 help 223 This option provides 'stack_erasing' sysctl, which can be used in 224 runtime to control kernel stack erasing for kernels built with 225 CONFIG_GCC_PLUGIN_STACKLEAK. 226 227config INIT_ON_ALLOC_DEFAULT_ON 228 bool "Enable heap memory zeroing on allocation by default" 229 depends on !KMSAN 230 help 231 This has the effect of setting "init_on_alloc=1" on the kernel 232 command line. This can be disabled with "init_on_alloc=0". 233 When "init_on_alloc" is enabled, all page allocator and slab 234 allocator memory will be zeroed when allocated, eliminating 235 many kinds of "uninitialized heap memory" flaws, especially 236 heap content exposures. The performance impact varies by 237 workload, but most cases see <1% impact. Some synthetic 238 workloads have measured as high as 7%. 239 240config INIT_ON_FREE_DEFAULT_ON 241 bool "Enable heap memory zeroing on free by default" 242 depends on !KMSAN 243 help 244 This has the effect of setting "init_on_free=1" on the kernel 245 command line. This can be disabled with "init_on_free=0". 246 Similar to "init_on_alloc", when "init_on_free" is enabled, 247 all page allocator and slab allocator memory will be zeroed 248 when freed, eliminating many kinds of "uninitialized heap memory" 249 flaws, especially heap content exposures. The primary difference 250 with "init_on_free" is that data lifetime in memory is reduced, 251 as anything freed is wiped immediately, making live forensics or 252 cold boot memory attacks unable to recover freed memory contents. 253 The performance impact varies by workload, but is more expensive 254 than "init_on_alloc" due to the negative cache effects of 255 touching "cold" memory areas. Most cases see 3-5% impact. Some 256 synthetic workloads have measured as high as 8%. 257 258config CC_HAS_ZERO_CALL_USED_REGS 259 def_bool $(cc-option,-fzero-call-used-regs=used-gpr) 260 # https://github.com/ClangBuiltLinux/linux/issues/1766 261 # https://github.com/llvm/llvm-project/issues/59242 262 depends on !CC_IS_CLANG || CLANG_VERSION > 150006 263 264config ZERO_CALL_USED_REGS 265 bool "Enable register zeroing on function exit" 266 depends on CC_HAS_ZERO_CALL_USED_REGS 267 help 268 At the end of functions, always zero any caller-used register 269 contents. This helps ensure that temporary values are not 270 leaked beyond the function boundary. This means that register 271 contents are less likely to be available for side channels 272 and information exposures. Additionally, this helps reduce the 273 number of useful ROP gadgets by about 20% (and removes compiler 274 generated "write-what-where" gadgets) in the resulting kernel 275 image. This has a less than 1% performance impact on most 276 workloads. Image size growth depends on architecture, and should 277 be evaluated for suitability. For example, x86_64 grows by less 278 than 1%, and arm64 grows by about 5%. 279 280endmenu 281 282config CC_HAS_RANDSTRUCT 283 def_bool $(cc-option,-frandomize-layout-seed-file=/dev/null) 284 # Randstruct was first added in Clang 15, but it isn't safe to use until 285 # Clang 16 due to https://github.com/llvm/llvm-project/issues/60349 286 depends on !CC_IS_CLANG || CLANG_VERSION >= 160000 287 288choice 289 prompt "Randomize layout of sensitive kernel structures" 290 default RANDSTRUCT_FULL if COMPILE_TEST && (GCC_PLUGINS || CC_HAS_RANDSTRUCT) 291 default RANDSTRUCT_NONE 292 help 293 If you enable this, the layouts of structures that are entirely 294 function pointers (and have not been manually annotated with 295 __no_randomize_layout), or structures that have been explicitly 296 marked with __randomize_layout, will be randomized at compile-time. 297 This can introduce the requirement of an additional information 298 exposure vulnerability for exploits targeting these structure 299 types. 300 301 Enabling this feature will introduce some performance impact, 302 slightly increase memory usage, and prevent the use of forensic 303 tools like Volatility against the system (unless the kernel 304 source tree isn't cleaned after kernel installation). 305 306 The seed used for compilation is in scripts/basic/randomize.seed. 307 It remains after a "make clean" to allow for external modules to 308 be compiled with the existing seed and will be removed by a 309 "make mrproper" or "make distclean". This file should not be made 310 public, or the structure layout can be determined. 311 312 config RANDSTRUCT_NONE 313 bool "Disable structure layout randomization" 314 help 315 Build normally: no structure layout randomization. 316 317 config RANDSTRUCT_FULL 318 bool "Fully randomize structure layout" 319 depends on CC_HAS_RANDSTRUCT || GCC_PLUGINS 320 select MODVERSIONS if MODULES 321 help 322 Fully randomize the member layout of sensitive 323 structures as much as possible, which may have both a 324 memory size and performance impact. 325 326 One difference between the Clang and GCC plugin 327 implementations is the handling of bitfields. The GCC 328 plugin treats them as fully separate variables, 329 introducing sometimes significant padding. Clang tries 330 to keep adjacent bitfields together, but with their bit 331 ordering randomized. 332 333 config RANDSTRUCT_PERFORMANCE 334 bool "Limit randomization of structure layout to cache-lines" 335 depends on GCC_PLUGINS 336 select MODVERSIONS if MODULES 337 help 338 Randomization of sensitive kernel structures will make a 339 best effort at restricting randomization to cacheline-sized 340 groups of members. It will further not randomize bitfields 341 in structures. This reduces the performance hit of RANDSTRUCT 342 at the cost of weakened randomization. 343endchoice 344 345config RANDSTRUCT 346 def_bool !RANDSTRUCT_NONE 347 348config GCC_PLUGIN_RANDSTRUCT 349 def_bool GCC_PLUGINS && RANDSTRUCT 350 help 351 Use GCC plugin to randomize structure layout. 352 353 This plugin was ported from grsecurity/PaX. More 354 information at: 355 * https://grsecurity.net/ 356 * https://pax.grsecurity.net/ 357 358endmenu 359