arch/x86/kernel/ldt.c

   1 /*
   2  * Copyright (C) 1992 Krishna Balasubramanian and Linus Torvalds
   3  * Copyright (C) 1999 Ingo Molnar <mingo@redhat.com>
   4  * Copyright (C) 2002 Andi Kleen
   5  *
   6  * This handles calls from both 32bit and 64bit mode.
   7  *
   8  * Lock order:
   9  *      contex.ldt_usr_sem
  10  *        mmap_sem
  11  *          context.lock
  12  */
  13
  14 #include <linux/errno.h>
  15 #include <linux/gfp.h>
  16 #include <linux/sched.h>
  17 #include <linux/string.h>
  18 #include <linux/mm.h>
  19 #include <linux/smp.h>
  20 #include <linux/syscalls.h>
  21 #include <linux/slab.h>
  22 #include <linux/vmalloc.h>
  23 #include <linux/uaccess.h>
  24
  25 #include <asm/ldt.h>
  26 #include <asm/tlb.h>
  27 #include <asm/desc.h>
  28 #include <asm/mmu_context.h>
  29 #include <asm/syscalls.h>
  30
  31 static void refresh_ldt_segments(void)
  32 {
  33 #ifdef CONFIG_X86_64
  34         unsigned short sel;
  35
  36         /*
  37          * Make sure that the cached DS and ES descriptors match the updated
  38          * LDT.
  39          */
  40         savesegment(ds, sel);
  41         if ((sel & SEGMENT_TI_MASK) == SEGMENT_LDT)
  42                 loadsegment(ds, sel);
  43
  44         savesegment(es, sel);
  45         if ((sel & SEGMENT_TI_MASK) == SEGMENT_LDT)
  46                 loadsegment(es, sel);
  47 #endif
  48 }
  49
  50 /* context.lock is held by the task which issued the smp function call */
  51 static void flush_ldt(void *__mm)
  52 {
  53         struct mm_struct *mm = __mm;
  54
  55         if (this_cpu_read(cpu_tlbstate.loaded_mm) != mm)
  56                 return;
  57
  58         load_mm_ldt(mm);
  59
  60         refresh_ldt_segments();
  61 }
  62
  63 /* The caller must call finalize_ldt_struct on the result. LDT starts zeroed. */
  64 static struct ldt_struct *alloc_ldt_struct(unsigned int num_entries)
  65 {
  66         struct ldt_struct *new_ldt;
  67         unsigned int alloc_size;
  68
  69         if (num_entries > LDT_ENTRIES)
  70                 return NULL;
  71
  72         new_ldt = kmalloc(sizeof(struct ldt_struct), GFP_KERNEL);
  73         if (!new_ldt)
  74                 return NULL;
  75
  76         BUILD_BUG_ON(LDT_ENTRY_SIZE != sizeof(struct desc_struct));
  77         alloc_size = num_entries * LDT_ENTRY_SIZE;
  78
  79         /*
  80          * Xen is very picky: it requires a page-aligned LDT that has no
  81          * trailing nonzero bytes in any page that contains LDT descriptors.
  82          * Keep it simple: zero the whole allocation and never allocate less
  83          * than PAGE_SIZE.
  84          */
  85         if (alloc_size > PAGE_SIZE)
  86                 new_ldt->entries = vzalloc(alloc_size);
  87         else
  88                 new_ldt->entries = (void *)get_zeroed_page(GFP_KERNEL);
  89
  90         if (!new_ldt->entries) {
  91                 kfree(new_ldt);
  92                 return NULL;
  93         }
  94
  95         /* The new LDT isn't aliased for PTI yet. */
  96         new_ldt->slot = -1;
  97
  98         new_ldt->nr_entries = num_entries;
  99         return new_ldt;
 100 }
 101
 102 /*
 103  * If PTI is enabled, this maps the LDT into the kernelmode and
 104  * usermode tables for the given mm.
 105  *
 106  * There is no corresponding unmap function.  Even if the LDT is freed, we
 107  * leave the PTEs around until the slot is reused or the mm is destroyed.
 108  * This is harmless: the LDT is always in ordinary memory, and no one will
 109  * access the freed slot.
 110  *
 111  * If we wanted to unmap freed LDTs, we'd also need to do a flush to make
 112  * it useful, and the flush would slow down modify_ldt().
 113  */
 114 static int
 115 map_ldt_struct(struct mm_struct *mm, struct ldt_struct *ldt, int slot)
 116 {
 117 #ifdef CONFIG_PAGE_TABLE_ISOLATION
 118         bool is_vmalloc, had_top_level_entry;
 119         unsigned long va;
 120         spinlock_t *ptl;
 121         pgd_t *pgd;
 122         int i;
 123
 124         if (!static_cpu_has(X86_FEATURE_PTI))
 125                 return 0;
 126
 127         /*
 128          * Any given ldt_struct should have map_ldt_struct() called at most
 129          * once.
 130          */
 131         WARN_ON(ldt->slot != -1);
 132
 133         /*
 134          * Did we already have the top level entry allocated?  We can't
 135          * use pgd_none() for this because it doens't do anything on
 136          * 4-level page table kernels.
 137          */
 138         pgd = pgd_offset(mm, LDT_BASE_ADDR);
 139         had_top_level_entry = (pgd->pgd != 0);
 140
 141         is_vmalloc = is_vmalloc_addr(ldt->entries);
 142
 143         for (i = 0; i * PAGE_SIZE < ldt->nr_entries * LDT_ENTRY_SIZE; i++) {
 144                 unsigned long offset = i << PAGE_SHIFT;
 145                 const void *src = (char *)ldt->entries + offset;
 146                 unsigned long pfn;
 147                 pte_t pte, *ptep;
 148
 149                 va = (unsigned long)ldt_slot_va(slot) + offset;
 150                 pfn = is_vmalloc ? vmalloc_to_pfn(src) :
 151                         page_to_pfn(virt_to_page(src));
 152                 /*
 153                  * Treat the PTI LDT range as a *userspace* range.
 154                  * get_locked_pte() will allocate all needed pagetables
 155                  * and account for them in this mm.
 156                  */
 157                 ptep = get_locked_pte(mm, va, &ptl);
 158                 if (!ptep)
 159                         return -ENOMEM;
 160                 /*
 161                  * Map it RO so the easy to find address is not a primary
 162                  * target via some kernel interface which misses a
 163                  * permission check.
 164                  */
 165                 pte = pfn_pte(pfn, __pgprot(__PAGE_KERNEL_RO & ~_PAGE_GLOBAL));
 166                 set_pte_at(mm, va, ptep, pte);
 167                 pte_unmap_unlock(ptep, ptl);
 168         }
 169
 170         if (mm->context.ldt) {
 171                 /*
 172                  * We already had an LDT.  The top-level entry should already
 173                  * have been allocated and synchronized with the usermode
 174                  * tables.
 175                  */
 176                 WARN_ON(!had_top_level_entry);
 177                 if (static_cpu_has(X86_FEATURE_PTI))
 178                         WARN_ON(!kernel_to_user_pgdp(pgd)->pgd);
 179         } else {
 180                 /*
 181                  * This is the first time we're mapping an LDT for this process.
 182                  * Sync the pgd to the usermode tables.
 183                  */
 184                 WARN_ON(had_top_level_entry);
 185                 if (static_cpu_has(X86_FEATURE_PTI)) {
 186                         WARN_ON(kernel_to_user_pgdp(pgd)->pgd);
 187                         set_pgd(kernel_to_user_pgdp(pgd), *pgd);
 188                 }
 189         }
 190
 191         va = (unsigned long)ldt_slot_va(slot);
 192         flush_tlb_mm_range(mm, va, va + LDT_SLOT_STRIDE, 0);
 193
 194         ldt->slot = slot;
 195 #endif
 196         return 0;
 197 }
 198
 199 static void free_ldt_pgtables(struct mm_struct *mm)
 200 {
 201 #ifdef CONFIG_PAGE_TABLE_ISOLATION
 202         struct mmu_gather tlb;
 203         unsigned long start = LDT_BASE_ADDR;
 204         unsigned long end = start + (1UL << PGDIR_SHIFT);
 205
 206         if (!static_cpu_has(X86_FEATURE_PTI))
 207                 return;
 208
 209         tlb_gather_mmu(&tlb, mm, start, end);
 210         free_pgd_range(&tlb, start, end, start, end);
 211         tlb_finish_mmu(&tlb, start, end);
 212 #endif
 213 }
 214
 215 /* After calling this, the LDT is immutable. */
 216 static void finalize_ldt_struct(struct ldt_struct *ldt)
 217 {
 218         paravirt_alloc_ldt(ldt->entries, ldt->nr_entries);
 219 }
 220
 221 static void install_ldt(struct mm_struct *mm, struct ldt_struct *ldt)
 222 {
 223         mutex_lock(&mm->context.lock);
 224
 225         /* Synchronizes with READ_ONCE in load_mm_ldt. */
 226         smp_store_release(&mm->context.ldt, ldt);
 227
 228         /* Activate the LDT for all CPUs using currents mm. */
 229         on_each_cpu_mask(mm_cpumask(mm), flush_ldt, mm, true);
 230
 231         mutex_unlock(&mm->context.lock);
 232 }
 233
 234 static void free_ldt_struct(struct ldt_struct *ldt)
 235 {
 236         if (likely(!ldt))
 237                 return;
 238
 239         paravirt_free_ldt(ldt->entries, ldt->nr_entries);
 240         if (ldt->nr_entries * LDT_ENTRY_SIZE > PAGE_SIZE)
 241                 vfree_atomic(ldt->entries);
 242         else
 243                 free_page((unsigned long)ldt->entries);
 244         kfree(ldt);
 245 }
 246
 247 /*
 248  * Called on fork from arch_dup_mmap(). Just copy the current LDT state,
 249  * the new task is not running, so nothing can be installed.
 250  */
 251 int ldt_dup_context(struct mm_struct *old_mm, struct mm_struct *mm)
 252 {
 253         struct ldt_struct *new_ldt;
 254         int retval = 0;
 255
 256         if (!old_mm)
 257                 return 0;
 258
 259         mutex_lock(&old_mm->context.lock);
 260         if (!old_mm->context.ldt)
 261                 goto out_unlock;
 262
 263         new_ldt = alloc_ldt_struct(old_mm->context.ldt->nr_entries);
 264         if (!new_ldt) {
 265                 retval = -ENOMEM;
 266                 goto out_unlock;
 267         }
 268
 269         memcpy(new_ldt->entries, old_mm->context.ldt->entries,
 270                new_ldt->nr_entries * LDT_ENTRY_SIZE);
 271         finalize_ldt_struct(new_ldt);
 272
 273         retval = map_ldt_struct(mm, new_ldt, 0);
 274         if (retval) {
 275                 free_ldt_pgtables(mm);
 276                 free_ldt_struct(new_ldt);
 277                 goto out_unlock;
 278         }
 279         mm->context.ldt = new_ldt;
 280
 281 out_unlock:
 282         mutex_unlock(&old_mm->context.lock);
 283         return retval;
 284 }
 285
 286 /*
 287  * No need to lock the MM as we are the last user
 288  *
 289  * 64bit: Don't touch the LDT register - we're already in the next thread.
 290  */
 291 void destroy_context_ldt(struct mm_struct *mm)
 292 {
 293         free_ldt_struct(mm->context.ldt);
 294         mm->context.ldt = NULL;
 295 }
 296
 297 void ldt_arch_exit_mmap(struct mm_struct *mm)
 298 {
 299         free_ldt_pgtables(mm);
 300 }
 301
 302 static int read_ldt(void __user *ptr, unsigned long bytecount)
 303 {
 304         struct mm_struct *mm = current->mm;
 305         unsigned long entries_size;
 306         int retval;
 307
 308         down_read(&mm->context.ldt_usr_sem);
 309
 310         if (!mm->context.ldt) {
 311                 retval = 0;
 312                 goto out_unlock;
 313         }
 314
 315         if (bytecount > LDT_ENTRY_SIZE * LDT_ENTRIES)
 316                 bytecount = LDT_ENTRY_SIZE * LDT_ENTRIES;
 317
 318         entries_size = mm->context.ldt->nr_entries * LDT_ENTRY_SIZE;
 319         if (entries_size > bytecount)
 320                 entries_size = bytecount;
 321
 322         if (copy_to_user(ptr, mm->context.ldt->entries, entries_size)) {
 323                 retval = -EFAULT;
 324                 goto out_unlock;
 325         }
 326
 327         if (entries_size != bytecount) {
 328                 /* Zero-fill the rest and pretend we read bytecount bytes. */
 329                 if (clear_user(ptr + entries_size, bytecount - entries_size)) {
 330                         retval = -EFAULT;
 331                         goto out_unlock;
 332                 }
 333         }
 334         retval = bytecount;
 335
 336 out_unlock:
 337         up_read(&mm->context.ldt_usr_sem);
 338         return retval;
 339 }
 340
 341 static int read_default_ldt(void __user *ptr, unsigned long bytecount)
 342 {
 343         /* CHECKME: Can we use _one_ random number ? */
 344 #ifdef CONFIG_X86_32
 345         unsigned long size = 5 * sizeof(struct desc_struct);
 346 #else
 347         unsigned long size = 128;
 348 #endif
 349         if (bytecount > size)
 350                 bytecount = size;
 351         if (clear_user(ptr, bytecount))
 352                 return -EFAULT;
 353         return bytecount;
 354 }
 355
 356 static int write_ldt(void __user *ptr, unsigned long bytecount, int oldmode)
 357 {
 358         struct mm_struct *mm = current->mm;
 359         struct ldt_struct *new_ldt, *old_ldt;
 360         unsigned int old_nr_entries, new_nr_entries;
 361         struct user_desc ldt_info;
 362         struct desc_struct ldt;
 363         int error;
 364
 365         error = -EINVAL;
 366         if (bytecount != sizeof(ldt_info))
 367                 goto out;
 368         error = -EFAULT;
 369         if (copy_from_user(&ldt_info, ptr, sizeof(ldt_info)))
 370                 goto out;
 371
 372         error = -EINVAL;
 373         if (ldt_info.entry_number >= LDT_ENTRIES)
 374                 goto out;
 375         if (ldt_info.contents == 3) {
 376                 if (oldmode)
 377                         goto out;
 378                 if (ldt_info.seg_not_present == 0)
 379                         goto out;
 380         }
 381
 382         if ((oldmode && !ldt_info.base_addr && !ldt_info.limit) ||
 383             LDT_empty(&ldt_info)) {
 384                 /* The user wants to clear the entry. */
 385                 memset(&ldt, 0, sizeof(ldt));
 386         } else {
 387                 if (!IS_ENABLED(CONFIG_X86_16BIT) && !ldt_info.seg_32bit) {
 388                         error = -EINVAL;
 389                         goto out;
 390                 }
 391
 392                 fill_ldt(&ldt, &ldt_info);
 393                 if (oldmode)
 394                         ldt.avl = 0;
 395         }
 396
 397         if (down_write_killable(&mm->context.ldt_usr_sem))
 398                 return -EINTR;
 399
 400         old_ldt       = mm->context.ldt;
 401         old_nr_entries = old_ldt ? old_ldt->nr_entries : 0;
 402         new_nr_entries = max(ldt_info.entry_number + 1, old_nr_entries);
 403
 404         error = -ENOMEM;
 405         new_ldt = alloc_ldt_struct(new_nr_entries);
 406         if (!new_ldt)
 407                 goto out_unlock;
 408
 409         if (old_ldt)
 410                 memcpy(new_ldt->entries, old_ldt->entries, old_nr_entries * LDT_ENTRY_SIZE);
 411
 412         new_ldt->entries[ldt_info.entry_number] = ldt;
 413         finalize_ldt_struct(new_ldt);
 414
 415         /*
 416          * If we are using PTI, map the new LDT into the userspace pagetables.
 417          * If there is already an LDT, use the other slot so that other CPUs
 418          * will continue to use the old LDT until install_ldt() switches
 419          * them over to the new LDT.
 420          */
 421         error = map_ldt_struct(mm, new_ldt, old_ldt ? !old_ldt->slot : 0);
 422         if (error) {
 423                 /*
 424                  * This only can fail for the first LDT setup. If an LDT is
 425                  * already installed then the PTE page is already
 426                  * populated. Mop up a half populated page table.
 427                  */
 428                 if (!WARN_ON_ONCE(old_ldt))
 429                         free_ldt_pgtables(mm);
 430                 free_ldt_struct(new_ldt);
 431                 goto out_unlock;
 432         }
 433
 434         install_ldt(mm, new_ldt);
 435         free_ldt_struct(old_ldt);
 436         error = 0;
 437
 438 out_unlock:
 439         up_write(&mm->context.ldt_usr_sem);
 440 out:
 441         return error;
 442 }
 443
 444 SYSCALL_DEFINE3(modify_ldt, int , func , void __user * , ptr ,
 445                 unsigned long , bytecount)
 446 {
 447         int ret = -ENOSYS;
 448
 449         switch (func) {
 450         case 0:
 451                 ret = read_ldt(ptr, bytecount);
 452                 break;
 453         case 1:
 454                 ret = write_ldt(ptr, bytecount, 1);
 455                 break;
 456         case 2:
 457                 ret = read_default_ldt(ptr, bytecount);
 458                 break;
 459         case 0x11:
 460                 ret = write_ldt(ptr, bytecount, 0);
 461                 break;
 462         }
 463         /*
 464          * The SYSCALL_DEFINE() macros give us an 'unsigned long'
 465          * return type, but tht ABI for sys_modify_ldt() expects
 466          * 'int'.  This cast gives us an int-sized value in %rax
 467          * for the return code.  The 'unsigned' is necessary so
 468          * the compiler does not try to sign-extend the negative
 469          * return codes into the high half of the register when
 470          * taking the value from int->long.
 471          */
 472         return (unsigned int)ret;
 473 }