1 // SPDX-License-Identifier: GPL-2.0-only 2 /* 3 * Copyright 2007 Andi Kleen, SUSE Labs. 4 * 5 * This contains most of the x86 vDSO kernel-side code. 6 */ 7 #include <linux/mm.h> 8 #include <linux/err.h> 9 #include <linux/sched.h> 10 #include <linux/sched/task_stack.h> 11 #include <linux/slab.h> 12 #include <linux/init.h> 13 #include <linux/random.h> 14 #include <linux/elf.h> 15 #include <linux/cpu.h> 16 #include <linux/ptrace.h> 17 #include <linux/time_namespace.h> 18 19 #include <asm/pvclock.h> 20 #include <asm/vgtod.h> 21 #include <asm/proto.h> 22 #include <asm/vdso.h> 23 #include <asm/vvar.h> 24 #include <asm/tlb.h> 25 #include <asm/page.h> 26 #include <asm/desc.h> 27 #include <asm/cpufeature.h> 28 #include <clocksource/hyperv_timer.h> 29 30 #undef _ASM_X86_VVAR_H 31 #define EMIT_VVAR(name, offset) \ 32 const size_t name ## _offset = offset; 33 #include <asm/vvar.h> 34 35 struct vdso_data *arch_get_vdso_data(void *vvar_page) 36 { 37 return (struct vdso_data *)(vvar_page + _vdso_data_offset); 38 } 39 #undef EMIT_VVAR 40 41 unsigned int vclocks_used __read_mostly; 42 43 #if defined(CONFIG_X86_64) 44 unsigned int __read_mostly vdso64_enabled = 1; 45 #endif 46 47 void __init init_vdso_image(const struct vdso_image *image) 48 { 49 BUG_ON(image->size % PAGE_SIZE != 0); 50 51 apply_alternatives((struct alt_instr *)(image->data + image->alt), 52 (struct alt_instr *)(image->data + image->alt + 53 image->alt_len)); 54 } 55 56 static const struct vm_special_mapping vvar_mapping; 57 struct linux_binprm; 58 59 static vm_fault_t vdso_fault(const struct vm_special_mapping *sm, 60 struct vm_area_struct *vma, struct vm_fault *vmf) 61 { 62 const struct vdso_image *image = vma->vm_mm->context.vdso_image; 63 64 if (!image || (vmf->pgoff << PAGE_SHIFT) >= image->size) 65 return VM_FAULT_SIGBUS; 66 67 vmf->page = virt_to_page(image->data + (vmf->pgoff << PAGE_SHIFT)); 68 get_page(vmf->page); 69 return 0; 70 } 71 72 static void vdso_fix_landing(const struct vdso_image *image, 73 struct vm_area_struct *new_vma) 74 { 75 #if defined CONFIG_X86_32 || defined CONFIG_IA32_EMULATION 76 if (in_ia32_syscall() && image == &vdso_image_32) { 77 struct pt_regs *regs = current_pt_regs(); 78 unsigned long vdso_land = image->sym_int80_landing_pad; 79 unsigned long old_land_addr = vdso_land + 80 (unsigned long)current->mm->context.vdso; 81 82 /* Fixing userspace landing - look at do_fast_syscall_32 */ 83 if (regs->ip == old_land_addr) 84 regs->ip = new_vma->vm_start + vdso_land; 85 } 86 #endif 87 } 88 89 static int vdso_mremap(const struct vm_special_mapping *sm, 90 struct vm_area_struct *new_vma) 91 { 92 const struct vdso_image *image = current->mm->context.vdso_image; 93 94 vdso_fix_landing(image, new_vma); 95 current->mm->context.vdso = (void __user *)new_vma->vm_start; 96 97 return 0; 98 } 99 100 #ifdef CONFIG_TIME_NS 101 /* 102 * The vvar page layout depends on whether a task belongs to the root or 103 * non-root time namespace. Whenever a task changes its namespace, the VVAR 104 * page tables are cleared and then they will re-faulted with a 105 * corresponding layout. 106 * See also the comment near timens_setup_vdso_data() for details. 107 */ 108 int vdso_join_timens(struct task_struct *task, struct time_namespace *ns) 109 { 110 struct mm_struct *mm = task->mm; 111 struct vm_area_struct *vma; 112 VMA_ITERATOR(vmi, mm, 0); 113 114 mmap_read_lock(mm); 115 for_each_vma(vmi, vma) { 116 if (vma_is_special_mapping(vma, &vvar_mapping)) 117 zap_vma_pages(vma); 118 } 119 mmap_read_unlock(mm); 120 121 return 0; 122 } 123 #endif 124 125 static vm_fault_t vvar_fault(const struct vm_special_mapping *sm, 126 struct vm_area_struct *vma, struct vm_fault *vmf) 127 { 128 const struct vdso_image *image = vma->vm_mm->context.vdso_image; 129 unsigned long pfn; 130 long sym_offset; 131 132 if (!image) 133 return VM_FAULT_SIGBUS; 134 135 sym_offset = (long)(vmf->pgoff << PAGE_SHIFT) + 136 image->sym_vvar_start; 137 138 /* 139 * Sanity check: a symbol offset of zero means that the page 140 * does not exist for this vdso image, not that the page is at 141 * offset zero relative to the text mapping. This should be 142 * impossible here, because sym_offset should only be zero for 143 * the page past the end of the vvar mapping. 144 */ 145 if (sym_offset == 0) 146 return VM_FAULT_SIGBUS; 147 148 if (sym_offset == image->sym_vvar_page) { 149 struct page *timens_page = find_timens_vvar_page(vma); 150 151 pfn = __pa_symbol(&__vvar_page) >> PAGE_SHIFT; 152 153 /* 154 * If a task belongs to a time namespace then a namespace 155 * specific VVAR is mapped with the sym_vvar_page offset and 156 * the real VVAR page is mapped with the sym_timens_page 157 * offset. 158 * See also the comment near timens_setup_vdso_data(). 159 */ 160 if (timens_page) { 161 unsigned long addr; 162 vm_fault_t err; 163 164 /* 165 * Optimization: inside time namespace pre-fault 166 * VVAR page too. As on timens page there are only 167 * offsets for clocks on VVAR, it'll be faulted 168 * shortly by VDSO code. 169 */ 170 addr = vmf->address + (image->sym_timens_page - sym_offset); 171 err = vmf_insert_pfn(vma, addr, pfn); 172 if (unlikely(err & VM_FAULT_ERROR)) 173 return err; 174 175 pfn = page_to_pfn(timens_page); 176 } 177 178 return vmf_insert_pfn(vma, vmf->address, pfn); 179 } else if (sym_offset == image->sym_pvclock_page) { 180 struct pvclock_vsyscall_time_info *pvti = 181 pvclock_get_pvti_cpu0_va(); 182 if (pvti && vclock_was_used(VDSO_CLOCKMODE_PVCLOCK)) { 183 return vmf_insert_pfn_prot(vma, vmf->address, 184 __pa(pvti) >> PAGE_SHIFT, 185 pgprot_decrypted(vma->vm_page_prot)); 186 } 187 } else if (sym_offset == image->sym_hvclock_page) { 188 pfn = hv_get_tsc_pfn(); 189 190 if (pfn && vclock_was_used(VDSO_CLOCKMODE_HVCLOCK)) 191 return vmf_insert_pfn(vma, vmf->address, pfn); 192 } else if (sym_offset == image->sym_timens_page) { 193 struct page *timens_page = find_timens_vvar_page(vma); 194 195 if (!timens_page) 196 return VM_FAULT_SIGBUS; 197 198 pfn = __pa_symbol(&__vvar_page) >> PAGE_SHIFT; 199 return vmf_insert_pfn(vma, vmf->address, pfn); 200 } 201 202 return VM_FAULT_SIGBUS; 203 } 204 205 static const struct vm_special_mapping vdso_mapping = { 206 .name = "[vdso]", 207 .fault = vdso_fault, 208 .mremap = vdso_mremap, 209 }; 210 static const struct vm_special_mapping vvar_mapping = { 211 .name = "[vvar]", 212 .fault = vvar_fault, 213 }; 214 215 /* 216 * Add vdso and vvar mappings to current process. 217 * @image - blob to map 218 * @addr - request a specific address (zero to map at free addr) 219 */ 220 static int map_vdso(const struct vdso_image *image, unsigned long addr) 221 { 222 struct mm_struct *mm = current->mm; 223 struct vm_area_struct *vma; 224 unsigned long text_start; 225 int ret = 0; 226 227 if (mmap_write_lock_killable(mm)) 228 return -EINTR; 229 230 addr = get_unmapped_area(NULL, addr, 231 image->size - image->sym_vvar_start, 0, 0); 232 if (IS_ERR_VALUE(addr)) { 233 ret = addr; 234 goto up_fail; 235 } 236 237 text_start = addr - image->sym_vvar_start; 238 239 /* 240 * MAYWRITE to allow gdb to COW and set breakpoints 241 */ 242 vma = _install_special_mapping(mm, 243 text_start, 244 image->size, 245 VM_READ|VM_EXEC| 246 VM_MAYREAD|VM_MAYWRITE|VM_MAYEXEC, 247 &vdso_mapping); 248 249 if (IS_ERR(vma)) { 250 ret = PTR_ERR(vma); 251 goto up_fail; 252 } 253 254 vma = _install_special_mapping(mm, 255 addr, 256 -image->sym_vvar_start, 257 VM_READ|VM_MAYREAD|VM_IO|VM_DONTDUMP| 258 VM_PFNMAP, 259 &vvar_mapping); 260 261 if (IS_ERR(vma)) { 262 ret = PTR_ERR(vma); 263 do_munmap(mm, text_start, image->size, NULL); 264 } else { 265 current->mm->context.vdso = (void __user *)text_start; 266 current->mm->context.vdso_image = image; 267 } 268 269 up_fail: 270 mmap_write_unlock(mm); 271 return ret; 272 } 273 274 #ifdef CONFIG_X86_64 275 /* 276 * Put the vdso above the (randomized) stack with another randomized 277 * offset. This way there is no hole in the middle of address space. 278 * To save memory make sure it is still in the same PTE as the stack 279 * top. This doesn't give that many random bits. 280 * 281 * Note that this algorithm is imperfect: the distribution of the vdso 282 * start address within a PMD is biased toward the end. 283 * 284 * Only used for the 64-bit and x32 vdsos. 285 */ 286 static unsigned long vdso_addr(unsigned long start, unsigned len) 287 { 288 unsigned long addr, end; 289 unsigned offset; 290 291 /* 292 * Round up the start address. It can start out unaligned as a result 293 * of stack start randomization. 294 */ 295 start = PAGE_ALIGN(start); 296 297 /* Round the lowest possible end address up to a PMD boundary. */ 298 end = (start + len + PMD_SIZE - 1) & PMD_MASK; 299 if (end >= TASK_SIZE_MAX) 300 end = TASK_SIZE_MAX; 301 end -= len; 302 303 if (end > start) { 304 offset = get_random_u32_below(((end - start) >> PAGE_SHIFT) + 1); 305 addr = start + (offset << PAGE_SHIFT); 306 } else { 307 addr = start; 308 } 309 310 /* 311 * Forcibly align the final address in case we have a hardware 312 * issue that requires alignment for performance reasons. 313 */ 314 addr = align_vdso_addr(addr); 315 316 return addr; 317 } 318 319 static int map_vdso_randomized(const struct vdso_image *image) 320 { 321 unsigned long addr = vdso_addr(current->mm->start_stack, image->size-image->sym_vvar_start); 322 323 return map_vdso(image, addr); 324 } 325 #endif 326 327 int map_vdso_once(const struct vdso_image *image, unsigned long addr) 328 { 329 struct mm_struct *mm = current->mm; 330 struct vm_area_struct *vma; 331 VMA_ITERATOR(vmi, mm, 0); 332 333 mmap_write_lock(mm); 334 /* 335 * Check if we have already mapped vdso blob - fail to prevent 336 * abusing from userspace install_special_mapping, which may 337 * not do accounting and rlimit right. 338 * We could search vma near context.vdso, but it's a slowpath, 339 * so let's explicitly check all VMAs to be completely sure. 340 */ 341 for_each_vma(vmi, vma) { 342 if (vma_is_special_mapping(vma, &vdso_mapping) || 343 vma_is_special_mapping(vma, &vvar_mapping)) { 344 mmap_write_unlock(mm); 345 return -EEXIST; 346 } 347 } 348 mmap_write_unlock(mm); 349 350 return map_vdso(image, addr); 351 } 352 353 #if defined(CONFIG_X86_32) || defined(CONFIG_IA32_EMULATION) 354 static int load_vdso32(void) 355 { 356 if (vdso32_enabled != 1) /* Other values all mean "disabled" */ 357 return 0; 358 359 return map_vdso(&vdso_image_32, 0); 360 } 361 #endif 362 363 #ifdef CONFIG_X86_64 364 int arch_setup_additional_pages(struct linux_binprm *bprm, int uses_interp) 365 { 366 if (!vdso64_enabled) 367 return 0; 368 369 return map_vdso_randomized(&vdso_image_64); 370 } 371 372 #ifdef CONFIG_COMPAT 373 int compat_arch_setup_additional_pages(struct linux_binprm *bprm, 374 int uses_interp, bool x32) 375 { 376 #ifdef CONFIG_X86_X32_ABI 377 if (x32) { 378 if (!vdso64_enabled) 379 return 0; 380 return map_vdso_randomized(&vdso_image_x32); 381 } 382 #endif 383 #ifdef CONFIG_IA32_EMULATION 384 return load_vdso32(); 385 #else 386 return 0; 387 #endif 388 } 389 #endif 390 #else 391 int arch_setup_additional_pages(struct linux_binprm *bprm, int uses_interp) 392 { 393 return load_vdso32(); 394 } 395 #endif 396 397 bool arch_syscall_is_vdso_sigreturn(struct pt_regs *regs) 398 { 399 #if defined(CONFIG_X86_32) || defined(CONFIG_IA32_EMULATION) 400 const struct vdso_image *image = current->mm->context.vdso_image; 401 unsigned long vdso = (unsigned long) current->mm->context.vdso; 402 403 if (in_ia32_syscall() && image == &vdso_image_32) { 404 if (regs->ip == vdso + image->sym_vdso32_sigreturn_landing_pad || 405 regs->ip == vdso + image->sym_vdso32_rt_sigreturn_landing_pad) 406 return true; 407 } 408 #endif 409 return false; 410 } 411 412 #ifdef CONFIG_X86_64 413 static __init int vdso_setup(char *s) 414 { 415 vdso64_enabled = simple_strtoul(s, NULL, 0); 416 return 1; 417 } 418 __setup("vdso=", vdso_setup); 419 420 static int __init init_vdso(void) 421 { 422 BUILD_BUG_ON(VDSO_CLOCKMODE_MAX >= 32); 423 424 init_vdso_image(&vdso_image_64); 425 426 #ifdef CONFIG_X86_X32_ABI 427 init_vdso_image(&vdso_image_x32); 428 #endif 429 430 return 0; 431 } 432 subsys_initcall(init_vdso); 433 #endif /* CONFIG_X86_64 */ 434