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