139218ff4SKees Cook /* SPDX-License-Identifier: GPL-2.0-only */ 239218ff4SKees Cook #ifndef _LINUX_RANDOMIZE_KSTACK_H 339218ff4SKees Cook #define _LINUX_RANDOMIZE_KSTACK_H 439218ff4SKees Cook 58cb37a59SMarco Elver #ifdef CONFIG_RANDOMIZE_KSTACK_OFFSET 639218ff4SKees Cook #include <linux/kernel.h> 739218ff4SKees Cook #include <linux/jump_label.h> 839218ff4SKees Cook #include <linux/percpu-defs.h> 939218ff4SKees Cook 1039218ff4SKees Cook DECLARE_STATIC_KEY_MAYBE(CONFIG_RANDOMIZE_KSTACK_OFFSET_DEFAULT, 1139218ff4SKees Cook randomize_kstack_offset); 1239218ff4SKees Cook DECLARE_PER_CPU(u32, kstack_offset); 1339218ff4SKees Cook 1439218ff4SKees Cook /* 1539218ff4SKees Cook * Do not use this anywhere else in the kernel. This is used here because 1639218ff4SKees Cook * it provides an arch-agnostic way to grow the stack with correct 1739218ff4SKees Cook * alignment. Also, since this use is being explicitly masked to a max of 1839218ff4SKees Cook * 10 bits, stack-clash style attacks are unlikely. For more details see 1939218ff4SKees Cook * "VLAs" in Documentation/process/deprecated.rst 20efa90c11SMarco Elver * 21efa90c11SMarco Elver * The normal __builtin_alloca() is initialized with INIT_STACK_ALL (currently 22efa90c11SMarco Elver * only with Clang and not GCC). Initializing the unused area on each syscall 23efa90c11SMarco Elver * entry is expensive, and generating an implicit call to memset() may also be 24efa90c11SMarco Elver * problematic (such as in noinstr functions). Therefore, if the compiler 25efa90c11SMarco Elver * supports it (which it should if it initializes allocas), always use the 26efa90c11SMarco Elver * "uninitialized" variant of the builtin. 2739218ff4SKees Cook */ 28efa90c11SMarco Elver #if __has_builtin(__builtin_alloca_uninitialized) 29efa90c11SMarco Elver #define __kstack_alloca __builtin_alloca_uninitialized 30efa90c11SMarco Elver #else 31efa90c11SMarco Elver #define __kstack_alloca __builtin_alloca 32efa90c11SMarco Elver #endif 33efa90c11SMarco Elver 3439218ff4SKees Cook /* 35*872bb37fSKees Cook * Use, at most, 6 bits of entropy (on 64-bit; 8 on 32-bit). This cap is 36*872bb37fSKees Cook * to keep the "VLA" from being unbounded (see above). Additionally clear 37*872bb37fSKees Cook * the bottom 4 bits (on 64-bit systems, 2 for 32-bit), since stack 38*872bb37fSKees Cook * alignment will always be at least word size. This makes the compiler 39*872bb37fSKees Cook * code gen better when it is applying the actual per-arch alignment to 40*872bb37fSKees Cook * the final offset. The resulting randomness is reasonable without overly 41*872bb37fSKees Cook * constraining usable stack space. 4239218ff4SKees Cook */ 43*872bb37fSKees Cook #ifdef CONFIG_64BIT 44*872bb37fSKees Cook #define KSTACK_OFFSET_MAX(x) ((x) & 0b1111110000) 45*872bb37fSKees Cook #else 46*872bb37fSKees Cook #define KSTACK_OFFSET_MAX(x) ((x) & 0b1111111100) 47*872bb37fSKees Cook #endif 4839218ff4SKees Cook 491ff29758SKees Cook /** 501ff29758SKees Cook * add_random_kstack_offset - Increase stack utilization by previously 511ff29758SKees Cook * chosen random offset 521ff29758SKees Cook * 531ff29758SKees Cook * This should be used in the syscall entry path when interrupts and 5439218ff4SKees Cook * preempt are disabled, and after user registers have been stored to 551ff29758SKees Cook * the stack. For testing the resulting entropy, please see: 561ff29758SKees Cook * tools/testing/selftests/lkdtm/stack-entropy.sh 5739218ff4SKees Cook */ 5839218ff4SKees Cook #define add_random_kstack_offset() do { \ 5939218ff4SKees Cook if (static_branch_maybe(CONFIG_RANDOMIZE_KSTACK_OFFSET_DEFAULT, \ 6039218ff4SKees Cook &randomize_kstack_offset)) { \ 6139218ff4SKees Cook u32 offset = raw_cpu_read(kstack_offset); \ 62efa90c11SMarco Elver u8 *ptr = __kstack_alloca(KSTACK_OFFSET_MAX(offset)); \ 6339218ff4SKees Cook /* Keep allocation even after "ptr" loses scope. */ \ 642515dd6cSNick Desaulniers asm volatile("" :: "r"(ptr) : "memory"); \ 6539218ff4SKees Cook } \ 6639218ff4SKees Cook } while (0) 6739218ff4SKees Cook 681ff29758SKees Cook /** 691ff29758SKees Cook * choose_random_kstack_offset - Choose the random offset for the next 701ff29758SKees Cook * add_random_kstack_offset() 711ff29758SKees Cook * 721ff29758SKees Cook * This should only be used during syscall exit when interrupts and 731ff29758SKees Cook * preempt are disabled. This position in the syscall flow is done to 741ff29758SKees Cook * frustrate attacks from userspace attempting to learn the next offset: 751ff29758SKees Cook * - Maximize the timing uncertainty visible from userspace: if the 761ff29758SKees Cook * offset is chosen at syscall entry, userspace has much more control 771ff29758SKees Cook * over the timing between choosing offsets. "How long will we be in 781ff29758SKees Cook * kernel mode?" tends to be more difficult to predict than "how long 791ff29758SKees Cook * will we be in user mode?" 801ff29758SKees Cook * - Reduce the lifetime of the new offset sitting in memory during 811ff29758SKees Cook * kernel mode execution. Exposure of "thread-local" memory content 821ff29758SKees Cook * (e.g. current, percpu, etc) tends to be easier than arbitrary 831ff29758SKees Cook * location memory exposure. 841ff29758SKees Cook */ 8539218ff4SKees Cook #define choose_random_kstack_offset(rand) do { \ 8639218ff4SKees Cook if (static_branch_maybe(CONFIG_RANDOMIZE_KSTACK_OFFSET_DEFAULT, \ 8739218ff4SKees Cook &randomize_kstack_offset)) { \ 8839218ff4SKees Cook u32 offset = raw_cpu_read(kstack_offset); \ 899c573cd3SKees Cook offset = ror32(offset, 5) ^ (rand); \ 9039218ff4SKees Cook raw_cpu_write(kstack_offset, offset); \ 9139218ff4SKees Cook } \ 9239218ff4SKees Cook } while (0) 938cb37a59SMarco Elver #else /* CONFIG_RANDOMIZE_KSTACK_OFFSET */ 948cb37a59SMarco Elver #define add_random_kstack_offset() do { } while (0) 958cb37a59SMarco Elver #define choose_random_kstack_offset(rand) do { } while (0) 968cb37a59SMarco Elver #endif /* CONFIG_RANDOMIZE_KSTACK_OFFSET */ 9739218ff4SKees Cook 9839218ff4SKees Cook #endif 99