arch/x86/entry/vdso/vma.c

   1 /*
   2  * Copyright 2007 Andi Kleen, SUSE Labs.
   3  * Subject to the GPL, v.2
   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/slab.h>
  11 #include <linux/init.h>
  12 #include <linux/random.h>
  13 #include <linux/elf.h>
  14 #include <linux/cpu.h>
  15 #include <linux/ptrace.h>
  16 #include <asm/pvclock.h>
  17 #include <asm/vgtod.h>
  18 #include <asm/proto.h>
  19 #include <asm/vdso.h>
  20 #include <asm/vvar.h>
  21 #include <asm/page.h>
  22 #include <asm/desc.h>
  23 #include <asm/cpufeature.h>
  24
  25 #if defined(CONFIG_X86_64)
  26 unsigned int __read_mostly vdso64_enabled = 1;
  27 #endif
  28
  29 void __init init_vdso_image(const struct vdso_image *image)
  30 {
  31         BUG_ON(image->size % PAGE_SIZE != 0);
  32
  33         apply_alternatives((struct alt_instr *)(image->data + image->alt),
  34                            (struct alt_instr *)(image->data + image->alt +
  35                                                 image->alt_len));
  36 }
  37
  38 struct linux_binprm;
  39
  40 static int vdso_fault(const struct vm_special_mapping *sm,
  41                       struct vm_area_struct *vma, struct vm_fault *vmf)
  42 {
  43         const struct vdso_image *image = vma->vm_mm->context.vdso_image;
  44
  45         if (!image || (vmf->pgoff << PAGE_SHIFT) >= image->size)
  46                 return VM_FAULT_SIGBUS;
  47
  48         vmf->page = virt_to_page(image->data + (vmf->pgoff << PAGE_SHIFT));
  49         get_page(vmf->page);
  50         return 0;
  51 }
  52
  53 static void vdso_fix_landing(const struct vdso_image *image,
  54                 struct vm_area_struct *new_vma)
  55 {
  56 #if defined CONFIG_X86_32 || defined CONFIG_IA32_EMULATION
  57         if (in_ia32_syscall() && image == &vdso_image_32) {
  58                 struct pt_regs *regs = current_pt_regs();
  59                 unsigned long vdso_land = image->sym_int80_landing_pad;
  60                 unsigned long old_land_addr = vdso_land +
  61                         (unsigned long)current->mm->context.vdso;
  62
  63                 /* Fixing userspace landing - look at do_fast_syscall_32 */
  64                 if (regs->ip == old_land_addr)
  65                         regs->ip = new_vma->vm_start + vdso_land;
  66         }
  67 #endif
  68 }
  69
  70 static int vdso_mremap(const struct vm_special_mapping *sm,
  71                 struct vm_area_struct *new_vma)
  72 {
  73         unsigned long new_size = new_vma->vm_end - new_vma->vm_start;
  74         const struct vdso_image *image = current->mm->context.vdso_image;
  75
  76         if (image->size != new_size)
  77                 return -EINVAL;
  78
  79         if (WARN_ON_ONCE(current->mm != new_vma->vm_mm))
  80                 return -EFAULT;
  81
  82         vdso_fix_landing(image, new_vma);
  83         current->mm->context.vdso = (void __user *)new_vma->vm_start;
  84
  85         return 0;
  86 }
  87
  88 static int vvar_fault(const struct vm_special_mapping *sm,
  89                       struct vm_area_struct *vma, struct vm_fault *vmf)
  90 {
  91         const struct vdso_image *image = vma->vm_mm->context.vdso_image;
  92         long sym_offset;
  93         int ret = -EFAULT;
  94
  95         if (!image)
  96                 return VM_FAULT_SIGBUS;
  97
  98         sym_offset = (long)(vmf->pgoff << PAGE_SHIFT) +
  99                 image->sym_vvar_start;
 100
 101         /*
 102          * Sanity check: a symbol offset of zero means that the page
 103          * does not exist for this vdso image, not that the page is at
 104          * offset zero relative to the text mapping.  This should be
 105          * impossible here, because sym_offset should only be zero for
 106          * the page past the end of the vvar mapping.
 107          */
 108         if (sym_offset == 0)
 109                 return VM_FAULT_SIGBUS;
 110
 111         if (sym_offset == image->sym_vvar_page) {
 112                 ret = vm_insert_pfn(vma, vmf->address,
 113                                     __pa_symbol(&__vvar_page) >> PAGE_SHIFT);
 114         } else if (sym_offset == image->sym_pvclock_page) {
 115                 struct pvclock_vsyscall_time_info *pvti =
 116                         pvclock_pvti_cpu0_va();
 117                 if (pvti && vclock_was_used(VCLOCK_PVCLOCK)) {
 118                         ret = vm_insert_pfn(
 119                                 vma,
 120                                 vmf->address,
 121                                 __pa(pvti) >> PAGE_SHIFT);
 122                 }
 123         }
 124
 125         if (ret == 0 || ret == -EBUSY)
 126                 return VM_FAULT_NOPAGE;
 127
 128         return VM_FAULT_SIGBUS;
 129 }
 130
 131 static const struct vm_special_mapping vdso_mapping = {
 132         .name = "[vdso]",
 133         .fault = vdso_fault,
 134         .mremap = vdso_mremap,
 135 };
 136 static const struct vm_special_mapping vvar_mapping = {
 137         .name = "[vvar]",
 138         .fault = vvar_fault,
 139 };
 140
 141 /*
 142  * Add vdso and vvar mappings to current process.
 143  * @image          - blob to map
 144  * @addr           - request a specific address (zero to map at free addr)
 145  */
 146 static int map_vdso(const struct vdso_image *image, unsigned long addr)
 147 {
 148         struct mm_struct *mm = current->mm;
 149         struct vm_area_struct *vma;
 150         unsigned long text_start;
 151         int ret = 0;
 152
 153         if (down_write_killable(&mm->mmap_sem))
 154                 return -EINTR;
 155
 156         addr = get_unmapped_area(NULL, addr,
 157                                  image->size - image->sym_vvar_start, 0, 0);
 158         if (IS_ERR_VALUE(addr)) {
 159                 ret = addr;
 160                 goto up_fail;
 161         }
 162
 163         text_start = addr - image->sym_vvar_start;
 164
 165         /*
 166          * MAYWRITE to allow gdb to COW and set breakpoints
 167          */
 168         vma = _install_special_mapping(mm,
 169                                        text_start,
 170                                        image->size,
 171                                        VM_READ|VM_EXEC|
 172                                        VM_MAYREAD|VM_MAYWRITE|VM_MAYEXEC,
 173                                        &vdso_mapping);
 174
 175         if (IS_ERR(vma)) {
 176                 ret = PTR_ERR(vma);
 177                 goto up_fail;
 178         }
 179
 180         vma = _install_special_mapping(mm,
 181                                        addr,
 182                                        -image->sym_vvar_start,
 183                                        VM_READ|VM_MAYREAD|VM_IO|VM_DONTDUMP|
 184                                        VM_PFNMAP,
 185                                        &vvar_mapping);
 186
 187         if (IS_ERR(vma)) {
 188                 ret = PTR_ERR(vma);
 189                 do_munmap(mm, text_start, image->size);
 190         } else {
 191                 current->mm->context.vdso = (void __user *)text_start;
 192                 current->mm->context.vdso_image = image;
 193         }
 194
 195 up_fail:
 196         up_write(&mm->mmap_sem);
 197         return ret;
 198 }
 199
 200 #ifdef CONFIG_X86_64
 201 /*
 202  * Put the vdso above the (randomized) stack with another randomized
 203  * offset.  This way there is no hole in the middle of address space.
 204  * To save memory make sure it is still in the same PTE as the stack
 205  * top.  This doesn't give that many random bits.
 206  *
 207  * Note that this algorithm is imperfect: the distribution of the vdso
 208  * start address within a PMD is biased toward the end.
 209  *
 210  * Only used for the 64-bit and x32 vdsos.
 211  */
 212 static unsigned long vdso_addr(unsigned long start, unsigned len)
 213 {
 214         unsigned long addr, end;
 215         unsigned offset;
 216
 217         /*
 218          * Round up the start address.  It can start out unaligned as a result
 219          * of stack start randomization.
 220          */
 221         start = PAGE_ALIGN(start);
 222
 223         /* Round the lowest possible end address up to a PMD boundary. */
 224         end = (start + len + PMD_SIZE - 1) & PMD_MASK;
 225         if (end >= TASK_SIZE_MAX)
 226                 end = TASK_SIZE_MAX;
 227         end -= len;
 228
 229         if (end > start) {
 230                 offset = get_random_int() % (((end - start) >> PAGE_SHIFT) + 1);
 231                 addr = start + (offset << PAGE_SHIFT);
 232         } else {
 233                 addr = start;
 234         }
 235
 236         /*
 237          * Forcibly align the final address in case we have a hardware
 238          * issue that requires alignment for performance reasons.
 239          */
 240         addr = align_vdso_addr(addr);
 241
 242         return addr;
 243 }
 244
 245 static int map_vdso_randomized(const struct vdso_image *image)
 246 {
 247         unsigned long addr = vdso_addr(current->mm->start_stack, image->size-image->sym_vvar_start);
 248
 249         return map_vdso(image, addr);
 250 }
 251 #endif
 252
 253 int map_vdso_once(const struct vdso_image *image, unsigned long addr)
 254 {
 255         struct mm_struct *mm = current->mm;
 256         struct vm_area_struct *vma;
 257
 258         down_write(&mm->mmap_sem);
 259         /*
 260          * Check if we have already mapped vdso blob - fail to prevent
 261          * abusing from userspace install_speciall_mapping, which may
 262          * not do accounting and rlimit right.
 263          * We could search vma near context.vdso, but it's a slowpath,
 264          * so let's explicitely check all VMAs to be completely sure.
 265          */
 266         for (vma = mm->mmap; vma; vma = vma->vm_next) {
 267                 if (vma_is_special_mapping(vma, &vdso_mapping) ||
 268                                 vma_is_special_mapping(vma, &vvar_mapping)) {
 269                         up_write(&mm->mmap_sem);
 270                         return -EEXIST;
 271                 }
 272         }
 273         up_write(&mm->mmap_sem);
 274
 275         return map_vdso(image, addr);
 276 }
 277
 278 #if defined(CONFIG_X86_32) || defined(CONFIG_IA32_EMULATION)
 279 static int load_vdso32(void)
 280 {
 281         if (vdso32_enabled != 1)  /* Other values all mean "disabled" */
 282                 return 0;
 283
 284         return map_vdso(&vdso_image_32, 0);
 285 }
 286 #endif
 287
 288 #ifdef CONFIG_X86_64
 289 int arch_setup_additional_pages(struct linux_binprm *bprm, int uses_interp)
 290 {
 291         if (!vdso64_enabled)
 292                 return 0;
 293
 294         return map_vdso_randomized(&vdso_image_64);
 295 }
 296
 297 #ifdef CONFIG_COMPAT
 298 int compat_arch_setup_additional_pages(struct linux_binprm *bprm,
 299                                        int uses_interp)
 300 {
 301 #ifdef CONFIG_X86_X32_ABI
 302         if (test_thread_flag(TIF_X32)) {
 303                 if (!vdso64_enabled)
 304                         return 0;
 305                 return map_vdso_randomized(&vdso_image_x32);
 306         }
 307 #endif
 308 #ifdef CONFIG_IA32_EMULATION
 309         return load_vdso32();
 310 #else
 311         return 0;
 312 #endif
 313 }
 314 #endif
 315 #else
 316 int arch_setup_additional_pages(struct linux_binprm *bprm, int uses_interp)
 317 {
 318         return load_vdso32();
 319 }
 320 #endif
 321
 322 #ifdef CONFIG_X86_64
 323 static __init int vdso_setup(char *s)
 324 {
 325         vdso64_enabled = simple_strtoul(s, NULL, 0);
 326         return 0;
 327 }
 328 __setup("vdso=", vdso_setup);
 329 #endif
 330
 331 #ifdef CONFIG_X86_64
 332 static void vgetcpu_cpu_init(void *arg)
 333 {
 334         int cpu = smp_processor_id();
 335         struct desc_struct d = { };
 336         unsigned long node = 0;
 337 #ifdef CONFIG_NUMA
 338         node = cpu_to_node(cpu);
 339 #endif
 340         if (static_cpu_has(X86_FEATURE_RDTSCP))
 341                 write_rdtscp_aux((node << 12) | cpu);
 342
 343         /*
 344          * Store cpu number in limit so that it can be loaded
 345          * quickly in user space in vgetcpu. (12 bits for the CPU
 346          * and 8 bits for the node)
 347          */
 348         d.limit0 = cpu | ((node & 0xf) << 12);
 349         d.limit = node >> 4;
 350         d.type = 5;             /* RO data, expand down, accessed */
 351         d.dpl = 3;              /* Visible to user code */
 352         d.s = 1;                /* Not a system segment */
 353         d.p = 1;                /* Present */
 354         d.d = 1;                /* 32-bit */
 355
 356         write_gdt_entry(get_cpu_gdt_table(cpu), GDT_ENTRY_PER_CPU, &d, DESCTYPE_S);
 357 }
 358
 359 static int vgetcpu_online(unsigned int cpu)
 360 {
 361         return smp_call_function_single(cpu, vgetcpu_cpu_init, NULL, 1);
 362 }
 363
 364 static int __init init_vdso(void)
 365 {
 366         init_vdso_image(&vdso_image_64);
 367
 368 #ifdef CONFIG_X86_X32_ABI
 369         init_vdso_image(&vdso_image_x32);
 370 #endif
 371
 372         /* notifier priority > KVM */
 373         return cpuhp_setup_state(CPUHP_AP_X86_VDSO_VMA_ONLINE,
 374                                  "x86/vdso/vma:online", vgetcpu_online, NULL);
 375 }
 376 subsys_initcall(init_vdso);
 377 #endif /* CONFIG_X86_64 */