| 1 | // SPDX-License-Identifier: GPL-2.0-only |
| 2 | |
| 3 | #include <linux/mm.h> |
| 4 | #include <linux/mman.h> |
| 5 | #include <linux/syscalls.h> |
| 6 | #include <linux/types.h> |
| 7 | |
| 8 | #include <asm/cmpxchg.h> |
| 9 | #include <asm/cpufeature.h> |
| 10 | #include <asm/gcs.h> |
| 11 | #include <asm/page.h> |
| 12 | |
| 13 | static unsigned long alloc_gcs(unsigned long addr, unsigned long size) |
| 14 | { |
| 15 | int flags = MAP_ANONYMOUS | MAP_PRIVATE; |
| 16 | struct mm_struct *mm = current->mm; |
| 17 | unsigned long mapped_addr, unused; |
| 18 | |
| 19 | if (addr) |
| 20 | flags |= MAP_FIXED_NOREPLACE; |
| 21 | |
| 22 | mmap_write_lock(mm); |
| 23 | mapped_addr = do_mmap(NULL, addr, len: size, PROT_READ, flags, |
| 24 | VM_SHADOW_STACK | VM_WRITE, pgoff: 0, populate: &unused, NULL); |
| 25 | mmap_write_unlock(mm); |
| 26 | |
| 27 | return mapped_addr; |
| 28 | } |
| 29 | |
| 30 | static unsigned long gcs_size(unsigned long size) |
| 31 | { |
| 32 | if (size) |
| 33 | return PAGE_ALIGN(size); |
| 34 | |
| 35 | /* Allocate RLIMIT_STACK/2 with limits of PAGE_SIZE..2G */ |
| 36 | size = PAGE_ALIGN(min_t(unsigned long long, |
| 37 | rlimit(RLIMIT_STACK) / 2, SZ_2G)); |
| 38 | return max(PAGE_SIZE, size); |
| 39 | } |
| 40 | |
| 41 | unsigned long gcs_alloc_thread_stack(struct task_struct *tsk, |
| 42 | const struct kernel_clone_args *args) |
| 43 | { |
| 44 | unsigned long addr, size; |
| 45 | |
| 46 | if (!system_supports_gcs()) |
| 47 | return 0; |
| 48 | |
| 49 | if (!task_gcs_el0_enabled(tsk)) |
| 50 | return 0; |
| 51 | |
| 52 | if ((args->flags & (CLONE_VFORK | CLONE_VM)) != CLONE_VM) { |
| 53 | tsk->thread.gcspr_el0 = read_sysreg_s(SYS_GCSPR_EL0); |
| 54 | return 0; |
| 55 | } |
| 56 | |
| 57 | size = args->stack_size / 2; |
| 58 | |
| 59 | size = gcs_size(size); |
| 60 | addr = alloc_gcs(addr: 0, size); |
| 61 | if (IS_ERR_VALUE(addr)) |
| 62 | return addr; |
| 63 | |
| 64 | tsk->thread.gcs_base = addr; |
| 65 | tsk->thread.gcs_size = size; |
| 66 | tsk->thread.gcspr_el0 = addr + size - sizeof(u64); |
| 67 | |
| 68 | return addr; |
| 69 | } |
| 70 | |
| 71 | SYSCALL_DEFINE3(map_shadow_stack, unsigned long, addr, unsigned long, size, unsigned int, flags) |
| 72 | { |
| 73 | unsigned long alloc_size; |
| 74 | unsigned long __user *cap_ptr; |
| 75 | unsigned long cap_val; |
| 76 | int ret = 0; |
| 77 | int cap_offset; |
| 78 | |
| 79 | if (!system_supports_gcs()) |
| 80 | return -EOPNOTSUPP; |
| 81 | |
| 82 | if (flags & ~(SHADOW_STACK_SET_TOKEN | SHADOW_STACK_SET_MARKER)) |
| 83 | return -EINVAL; |
| 84 | |
| 85 | if (!PAGE_ALIGNED(addr)) |
| 86 | return -EINVAL; |
| 87 | |
| 88 | if (size == 8 || !IS_ALIGNED(size, 8)) |
| 89 | return -EINVAL; |
| 90 | |
| 91 | /* |
| 92 | * An overflow would result in attempting to write the restore token |
| 93 | * to the wrong location. Not catastrophic, but just return the right |
| 94 | * error code and block it. |
| 95 | */ |
| 96 | alloc_size = PAGE_ALIGN(size); |
| 97 | if (alloc_size < size) |
| 98 | return -EOVERFLOW; |
| 99 | |
| 100 | addr = alloc_gcs(addr, size: alloc_size); |
| 101 | if (IS_ERR_VALUE(addr)) |
| 102 | return addr; |
| 103 | |
| 104 | /* |
| 105 | * Put a cap token at the end of the allocated region so it |
| 106 | * can be switched to. |
| 107 | */ |
| 108 | if (flags & SHADOW_STACK_SET_TOKEN) { |
| 109 | /* Leave an extra empty frame as a top of stack marker? */ |
| 110 | if (flags & SHADOW_STACK_SET_MARKER) |
| 111 | cap_offset = 2; |
| 112 | else |
| 113 | cap_offset = 1; |
| 114 | |
| 115 | cap_ptr = (unsigned long __user *)(addr + size - |
| 116 | (cap_offset * sizeof(unsigned long))); |
| 117 | cap_val = GCS_CAP(cap_ptr); |
| 118 | |
| 119 | put_user_gcs(cap_val, cap_ptr, &ret); |
| 120 | if (ret != 0) { |
| 121 | vm_munmap(addr, size); |
| 122 | return -EFAULT; |
| 123 | } |
| 124 | |
| 125 | /* |
| 126 | * Ensure the new cap is ordered before standard |
| 127 | * memory accesses to the same location. |
| 128 | */ |
| 129 | gcsb_dsync(); |
| 130 | } |
| 131 | |
| 132 | return addr; |
| 133 | } |
| 134 | |
| 135 | /* |
| 136 | * Apply the GCS mode configured for the specified task to the |
| 137 | * hardware. |
| 138 | */ |
| 139 | void gcs_set_el0_mode(struct task_struct *task) |
| 140 | { |
| 141 | u64 gcscre0_el1 = GCSCRE0_EL1_nTR; |
| 142 | |
| 143 | if (task->thread.gcs_el0_mode & PR_SHADOW_STACK_ENABLE) |
| 144 | gcscre0_el1 |= GCSCRE0_EL1_RVCHKEN | GCSCRE0_EL1_PCRSEL; |
| 145 | |
| 146 | if (task->thread.gcs_el0_mode & PR_SHADOW_STACK_WRITE) |
| 147 | gcscre0_el1 |= GCSCRE0_EL1_STREn; |
| 148 | |
| 149 | if (task->thread.gcs_el0_mode & PR_SHADOW_STACK_PUSH) |
| 150 | gcscre0_el1 |= GCSCRE0_EL1_PUSHMEn; |
| 151 | |
| 152 | write_sysreg_s(gcscre0_el1, SYS_GCSCRE0_EL1); |
| 153 | } |
| 154 | |
| 155 | void gcs_free(struct task_struct *task) |
| 156 | { |
| 157 | if (!system_supports_gcs()) |
| 158 | return; |
| 159 | |
| 160 | if (!task->mm || task->mm != current->mm) |
| 161 | return; |
| 162 | |
| 163 | if (task->thread.gcs_base) |
| 164 | vm_munmap(task->thread.gcs_base, task->thread.gcs_size); |
| 165 | |
| 166 | task->thread.gcspr_el0 = 0; |
| 167 | task->thread.gcs_base = 0; |
| 168 | task->thread.gcs_size = 0; |
| 169 | } |
| 170 | |
| 171 | int arch_set_shadow_stack_status(struct task_struct *task, unsigned long arg) |
| 172 | { |
| 173 | unsigned long gcs, size; |
| 174 | int ret; |
| 175 | |
| 176 | if (!system_supports_gcs()) |
| 177 | return -EINVAL; |
| 178 | |
| 179 | if (is_compat_thread(task_thread_info(task))) |
| 180 | return -EINVAL; |
| 181 | |
| 182 | /* Reject unknown flags */ |
| 183 | if (arg & ~PR_SHADOW_STACK_SUPPORTED_STATUS_MASK) |
| 184 | return -EINVAL; |
| 185 | |
| 186 | ret = gcs_check_locked(task, arg); |
| 187 | if (ret != 0) |
| 188 | return ret; |
| 189 | |
| 190 | /* If we are enabling GCS then make sure we have a stack */ |
| 191 | if (arg & PR_SHADOW_STACK_ENABLE && |
| 192 | !task_gcs_el0_enabled(task)) { |
| 193 | /* Do not allow GCS to be reenabled */ |
| 194 | if (task->thread.gcs_base || task->thread.gcspr_el0) |
| 195 | return -EINVAL; |
| 196 | |
| 197 | if (task != current) |
| 198 | return -EBUSY; |
| 199 | |
| 200 | size = gcs_size(size: 0); |
| 201 | gcs = alloc_gcs(addr: 0, size); |
| 202 | if (!gcs) |
| 203 | return -ENOMEM; |
| 204 | |
| 205 | task->thread.gcspr_el0 = gcs + size - sizeof(u64); |
| 206 | task->thread.gcs_base = gcs; |
| 207 | task->thread.gcs_size = size; |
| 208 | if (task == current) |
| 209 | write_sysreg_s(task->thread.gcspr_el0, |
| 210 | SYS_GCSPR_EL0); |
| 211 | } |
| 212 | |
| 213 | task->thread.gcs_el0_mode = arg; |
| 214 | if (task == current) |
| 215 | gcs_set_el0_mode(task); |
| 216 | |
| 217 | return 0; |
| 218 | } |
| 219 | |
| 220 | int arch_get_shadow_stack_status(struct task_struct *task, |
| 221 | unsigned long __user *arg) |
| 222 | { |
| 223 | if (!system_supports_gcs()) |
| 224 | return -EINVAL; |
| 225 | |
| 226 | if (is_compat_thread(task_thread_info(task))) |
| 227 | return -EINVAL; |
| 228 | |
| 229 | return put_user(task->thread.gcs_el0_mode, arg); |
| 230 | } |
| 231 | |
| 232 | int arch_lock_shadow_stack_status(struct task_struct *task, |
| 233 | unsigned long arg) |
| 234 | { |
| 235 | if (!system_supports_gcs()) |
| 236 | return -EINVAL; |
| 237 | |
| 238 | if (is_compat_thread(task_thread_info(task))) |
| 239 | return -EINVAL; |
| 240 | |
| 241 | /* |
| 242 | * We support locking unknown bits so applications can prevent |
| 243 | * any changes in a future proof manner. |
| 244 | */ |
| 245 | task->thread.gcs_el0_locked |= arg; |
| 246 | |
| 247 | return 0; |
| 248 | } |
| 249 | |