arch/x86/kernel/ldt.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (C) 1992 Krishna Balasubramanian and Linus Torvalds
   4  * Copyright (C) 1999 Ingo Molnar <mingo@redhat.com>
   5  * Copyright (C) 2002 Andi Kleen
   6  *
   7  * This handles calls from both 32bit and 64bit mode.
   8  *
   9  * Lock order:
  10  *      contex.ldt_usr_sem
  11  *        mmap_sem
  12  *          context.lock
  13  */
  14
  15 #include <linux/errno.h>
  16 #include <linux/gfp.h>
  17 #include <linux/sched.h>
  18 #include <linux/string.h>
  19 #include <linux/mm.h>
  20 #include <linux/smp.h>
  21 #include <linux/syscalls.h>
  22 #include <linux/slab.h>
  23 #include <linux/vmalloc.h>
  24 #include <linux/uaccess.h>
  25
  26 #include <asm/ldt.h>
  27 #include <asm/tlb.h>
  28 #include <asm/desc.h>
  29 #include <asm/mmu_context.h>
  30 #include <asm/syscalls.h>
  31
  32 static void refresh_ldt_segments(void)
  33 {
  34 #ifdef CONFIG_X86_64
  35         unsigned short sel;
  36
  37         /*
  38          * Make sure that the cached DS and ES descriptors match the updated
  39          * LDT.
  40          */
  41         savesegment(ds, sel);
  42         if ((sel & SEGMENT_TI_MASK) == SEGMENT_LDT)
  43                 loadsegment(ds, sel);
  44
  45         savesegment(es, sel);
  46         if ((sel & SEGMENT_TI_MASK) == SEGMENT_LDT)
  47                 loadsegment(es, sel);
  48 #endif
  49 }
  50
  51 /* context.lock is held by the task which issued the smp function call */
  52 static void flush_ldt(void *__mm)
  53 {
  54         struct mm_struct *mm = __mm;
  55
  56         if (this_cpu_read(cpu_tlbstate.loaded_mm) != mm)
  57                 return;
  58
  59         load_mm_ldt(mm);
  60
  61         refresh_ldt_segments();
  62 }
  63
  64 /* The caller must call finalize_ldt_struct on the result. LDT starts zeroed. */
  65 static struct ldt_struct *alloc_ldt_struct(unsigned int num_entries)
  66 {
  67         struct ldt_struct *new_ldt;
  68         unsigned int alloc_size;
  69
  70         if (num_entries > LDT_ENTRIES)
  71                 return NULL;
  72
  73         new_ldt = kmalloc(sizeof(struct ldt_struct), GFP_KERNEL);
  74         if (!new_ldt)
  75                 return NULL;
  76
  77         BUILD_BUG_ON(LDT_ENTRY_SIZE != sizeof(struct desc_struct));
  78         alloc_size = num_entries * LDT_ENTRY_SIZE;
  79
  80         /*
  81          * Xen is very picky: it requires a page-aligned LDT that has no
  82          * trailing nonzero bytes in any page that contains LDT descriptors.
  83          * Keep it simple: zero the whole allocation and never allocate less
  84          * than PAGE_SIZE.
  85          */
  86         if (alloc_size > PAGE_SIZE)
  87                 new_ldt->entries = vzalloc(alloc_size);
  88         else
  89                 new_ldt->entries = (void *)get_zeroed_page(GFP_KERNEL);
  90
  91         if (!new_ldt->entries) {
  92                 kfree(new_ldt);
  93                 return NULL;
  94         }
  95
  96         /* The new LDT isn't aliased for PTI yet. */
  97         new_ldt->slot = -1;
  98
  99         new_ldt->nr_entries = num_entries;
 100         return new_ldt;
 101 }
 102
 103 /*
 104  * If PTI is enabled, this maps the LDT into the kernelmode and
 105  * usermode tables for the given mm.
 106  *
 107  * There is no corresponding unmap function.  Even if the LDT is freed, we
 108  * leave the PTEs around until the slot is reused or the mm is destroyed.
 109  * This is harmless: the LDT is always in ordinary memory, and no one will
 110  * access the freed slot.
 111  *
 112  * If we wanted to unmap freed LDTs, we'd also need to do a flush to make
 113  * it useful, and the flush would slow down modify_ldt().
 114  */
 115 static int
 116 map_ldt_struct(struct mm_struct *mm, struct ldt_struct *ldt, int slot)
 117 {
 118 #ifdef CONFIG_PAGE_TABLE_ISOLATION
 119         bool is_vmalloc, had_top_level_entry;
 120         unsigned long va;
 121         spinlock_t *ptl;
 122         pgd_t *pgd;
 123         int i;
 124
 125         if (!static_cpu_has(X86_FEATURE_PTI))
 126                 return 0;
 127
 128         /*
 129          * Any given ldt_struct should have map_ldt_struct() called at most
 130          * once.
 131          */
 132         WARN_ON(ldt->slot != -1);
 133
 134         /*
 135          * Did we already have the top level entry allocated?  We can't
 136          * use pgd_none() for this because it doens't do anything on
 137          * 4-level page table kernels.
 138          */
 139         pgd = pgd_offset(mm, LDT_BASE_ADDR);
 140         had_top_level_entry = (pgd->pgd != 0);
 141
 142         is_vmalloc = is_vmalloc_addr(ldt->entries);
 143
 144         for (i = 0; i * PAGE_SIZE < ldt->nr_entries * LDT_ENTRY_SIZE; i++) {
 145                 unsigned long offset = i << PAGE_SHIFT;
 146                 const void *src = (char *)ldt->entries + offset;
 147                 unsigned long pfn;
 148                 pte_t pte, *ptep;
 149
 150                 va = (unsigned long)ldt_slot_va(slot) + offset;
 151                 pfn = is_vmalloc ? vmalloc_to_pfn(src) :
 152                         page_to_pfn(virt_to_page(src));
 153                 /*
 154                  * Treat the PTI LDT range as a *userspace* range.
 155                  * get_locked_pte() will allocate all needed pagetables
 156                  * and account for them in this mm.
 157                  */
 158                 ptep = get_locked_pte(mm, va, &ptl);
 159                 if (!ptep)
 160                         return -ENOMEM;
 161                 /*
 162                  * Map it RO so the easy to find address is not a primary
 163                  * target via some kernel interface which misses a
 164                  * permission check.
 165                  */
 166                 pte = pfn_pte(pfn, __pgprot(__PAGE_KERNEL_RO & ~_PAGE_GLOBAL));
 167                 set_pte_at(mm, va, ptep, pte);
 168                 pte_unmap_unlock(ptep, ptl);
 169         }
 170
 171         if (mm->context.ldt) {
 172                 /*
 173                  * We already had an LDT.  The top-level entry should already
 174                  * have been allocated and synchronized with the usermode
 175                  * tables.
 176                  */
 177                 WARN_ON(!had_top_level_entry);
 178                 if (static_cpu_has(X86_FEATURE_PTI))
 179                         WARN_ON(!kernel_to_user_pgdp(pgd)->pgd);
 180         } else {
 181                 /*
 182                  * This is the first time we're mapping an LDT for this process.
 183                  * Sync the pgd to the usermode tables.
 184                  */
 185                 WARN_ON(had_top_level_entry);
 186                 if (static_cpu_has(X86_FEATURE_PTI)) {
 187                         WARN_ON(kernel_to_user_pgdp(pgd)->pgd);
 188                         set_pgd(kernel_to_user_pgdp(pgd), *pgd);
 189                 }
 190         }
 191
 192         va = (unsigned long)ldt_slot_va(slot);
 193         flush_tlb_mm_range(mm, va, va + LDT_SLOT_STRIDE, 0);
 194
 195         ldt->slot = slot;
 196 #endif
 197         return 0;
 198 }
 199
 200 static void free_ldt_pgtables(struct mm_struct *mm)
 201 {
 202 #ifdef CONFIG_PAGE_TABLE_ISOLATION
 203         struct mmu_gather tlb;
 204         unsigned long start = LDT_BASE_ADDR;
 205         unsigned long end = start + (1UL << PGDIR_SHIFT);
 206
 207         if (!static_cpu_has(X86_FEATURE_PTI))
 208                 return;
 209
 210         tlb_gather_mmu(&tlb, mm, start, end);
 211         free_pgd_range(&tlb, start, end, start, end);
 212         tlb_finish_mmu(&tlb, start, end);
 213 #endif
 214 }
 215
 216 /* After calling this, the LDT is immutable. */
 217 static void finalize_ldt_struct(struct ldt_struct *ldt)
 218 {
 219         paravirt_alloc_ldt(ldt->entries, ldt->nr_entries);
 220 }
 221
 222 static void install_ldt(struct mm_struct *mm, struct ldt_struct *ldt)
 223 {
 224         mutex_lock(&mm->context.lock);
 225
 226         /* Synchronizes with READ_ONCE in load_mm_ldt. */
 227         smp_store_release(&mm->context.ldt, ldt);
 228
 229         /* Activate the LDT for all CPUs using currents mm. */
 230         on_each_cpu_mask(mm_cpumask(mm), flush_ldt, mm, true);
 231
 232         mutex_unlock(&mm->context.lock);
 233 }
 234
 235 static void free_ldt_struct(struct ldt_struct *ldt)
 236 {
 237         if (likely(!ldt))
 238                 return;
 239
 240         paravirt_free_ldt(ldt->entries, ldt->nr_entries);
 241         if (ldt->nr_entries * LDT_ENTRY_SIZE > PAGE_SIZE)
 242                 vfree_atomic(ldt->entries);
 243         else
 244                 free_page((unsigned long)ldt->entries);
 245         kfree(ldt);
 246 }
 247
 248 /*
 249  * Called on fork from arch_dup_mmap(). Just copy the current LDT state,
 250  * the new task is not running, so nothing can be installed.
 251  */
 252 int ldt_dup_context(struct mm_struct *old_mm, struct mm_struct *mm)
 253 {
 254         struct ldt_struct *new_ldt;
 255         int retval = 0;
 256
 257         if (!old_mm)
 258                 return 0;
 259
 260         mutex_lock(&old_mm->context.lock);
 261         if (!old_mm->context.ldt)
 262                 goto out_unlock;
 263
 264         new_ldt = alloc_ldt_struct(old_mm->context.ldt->nr_entries);
 265         if (!new_ldt) {
 266                 retval = -ENOMEM;
 267                 goto out_unlock;
 268         }
 269
 270         memcpy(new_ldt->entries, old_mm->context.ldt->entries,
 271                new_ldt->nr_entries * LDT_ENTRY_SIZE);
 272         finalize_ldt_struct(new_ldt);
 273
 274         retval = map_ldt_struct(mm, new_ldt, 0);
 275         if (retval) {
 276                 free_ldt_pgtables(mm);
 277                 free_ldt_struct(new_ldt);
 278                 goto out_unlock;
 279         }
 280         mm->context.ldt = new_ldt;
 281
 282 out_unlock:
 283         mutex_unlock(&old_mm->context.lock);
 284         return retval;
 285 }
 286
 287 /*
 288  * No need to lock the MM as we are the last user
 289  *
 290  * 64bit: Don't touch the LDT register - we're already in the next thread.
 291  */
 292 void destroy_context_ldt(struct mm_struct *mm)
 293 {
 294         free_ldt_struct(mm->context.ldt);
 295         mm->context.ldt = NULL;
 296 }
 297
 298 void ldt_arch_exit_mmap(struct mm_struct *mm)
 299 {
 300         free_ldt_pgtables(mm);
 301 }
 302
 303 static int read_ldt(void __user *ptr, unsigned long bytecount)
 304 {
 305         struct mm_struct *mm = current->mm;
 306         unsigned long entries_size;
 307         int retval;
 308
 309         down_read(&mm->context.ldt_usr_sem);
 310
 311         if (!mm->context.ldt) {
 312                 retval = 0;
 313                 goto out_unlock;
 314         }
 315
 316         if (bytecount > LDT_ENTRY_SIZE * LDT_ENTRIES)
 317                 bytecount = LDT_ENTRY_SIZE * LDT_ENTRIES;
 318
 319         entries_size = mm->context.ldt->nr_entries * LDT_ENTRY_SIZE;
 320         if (entries_size > bytecount)
 321                 entries_size = bytecount;
 322
 323         if (copy_to_user(ptr, mm->context.ldt->entries, entries_size)) {
 324                 retval = -EFAULT;
 325                 goto out_unlock;
 326         }
 327
 328         if (entries_size != bytecount) {
 329                 /* Zero-fill the rest and pretend we read bytecount bytes. */
 330                 if (clear_user(ptr + entries_size, bytecount - entries_size)) {
 331                         retval = -EFAULT;
 332                         goto out_unlock;
 333                 }
 334         }
 335         retval = bytecount;
 336
 337 out_unlock:
 338         up_read(&mm->context.ldt_usr_sem);
 339         return retval;
 340 }
 341
 342 static int read_default_ldt(void __user *ptr, unsigned long bytecount)
 343 {
 344         /* CHECKME: Can we use _one_ random number ? */
 345 #ifdef CONFIG_X86_32
 346         unsigned long size = 5 * sizeof(struct desc_struct);
 347 #else
 348         unsigned long size = 128;
 349 #endif
 350         if (bytecount > size)
 351                 bytecount = size;
 352         if (clear_user(ptr, bytecount))
 353                 return -EFAULT;
 354         return bytecount;
 355 }
 356
 357 static int write_ldt(void __user *ptr, unsigned long bytecount, int oldmode)
 358 {
 359         struct mm_struct *mm = current->mm;
 360         struct ldt_struct *new_ldt, *old_ldt;
 361         unsigned int old_nr_entries, new_nr_entries;
 362         struct user_desc ldt_info;
 363         struct desc_struct ldt;
 364         int error;
 365
 366         error = -EINVAL;
 367         if (bytecount != sizeof(ldt_info))
 368                 goto out;
 369         error = -EFAULT;
 370         if (copy_from_user(&ldt_info, ptr, sizeof(ldt_info)))
 371                 goto out;
 372
 373         error = -EINVAL;
 374         if (ldt_info.entry_number >= LDT_ENTRIES)
 375                 goto out;
 376         if (ldt_info.contents == 3) {
 377                 if (oldmode)
 378                         goto out;
 379                 if (ldt_info.seg_not_present == 0)
 380                         goto out;
 381         }
 382
 383         if ((oldmode && !ldt_info.base_addr && !ldt_info.limit) ||
 384             LDT_empty(&ldt_info)) {
 385                 /* The user wants to clear the entry. */
 386                 memset(&ldt, 0, sizeof(ldt));
 387         } else {
 388                 if (!IS_ENABLED(CONFIG_X86_16BIT) && !ldt_info.seg_32bit) {
 389                         error = -EINVAL;
 390                         goto out;
 391                 }
 392
 393                 fill_ldt(&ldt, &ldt_info);
 394                 if (oldmode)
 395                         ldt.avl = 0;
 396         }
 397
 398         if (down_write_killable(&mm->context.ldt_usr_sem))
 399                 return -EINTR;
 400
 401         old_ldt       = mm->context.ldt;
 402         old_nr_entries = old_ldt ? old_ldt->nr_entries : 0;
 403         new_nr_entries = max(ldt_info.entry_number + 1, old_nr_entries);
 404
 405         error = -ENOMEM;
 406         new_ldt = alloc_ldt_struct(new_nr_entries);
 407         if (!new_ldt)
 408                 goto out_unlock;
 409
 410         if (old_ldt)
 411                 memcpy(new_ldt->entries, old_ldt->entries, old_nr_entries * LDT_ENTRY_SIZE);
 412
 413         new_ldt->entries[ldt_info.entry_number] = ldt;
 414         finalize_ldt_struct(new_ldt);
 415
 416         /*
 417          * If we are using PTI, map the new LDT into the userspace pagetables.
 418          * If there is already an LDT, use the other slot so that other CPUs
 419          * will continue to use the old LDT until install_ldt() switches
 420          * them over to the new LDT.
 421          */
 422         error = map_ldt_struct(mm, new_ldt, old_ldt ? !old_ldt->slot : 0);
 423         if (error) {
 424                 /*
 425                  * This only can fail for the first LDT setup. If an LDT is
 426                  * already installed then the PTE page is already
 427                  * populated. Mop up a half populated page table.
 428                  */
 429                 if (!WARN_ON_ONCE(old_ldt))
 430                         free_ldt_pgtables(mm);
 431                 free_ldt_struct(new_ldt);
 432                 goto out_unlock;
 433         }
 434
 435         install_ldt(mm, new_ldt);
 436         free_ldt_struct(old_ldt);
 437         error = 0;
 438
 439 out_unlock:
 440         up_write(&mm->context.ldt_usr_sem);
 441 out:
 442         return error;
 443 }
 444
 445 SYSCALL_DEFINE3(modify_ldt, int , func , void __user * , ptr ,
 446                 unsigned long , bytecount)
 447 {
 448         int ret = -ENOSYS;
 449
 450         switch (func) {
 451         case 0:
 452                 ret = read_ldt(ptr, bytecount);
 453                 break;
 454         case 1:
 455                 ret = write_ldt(ptr, bytecount, 1);
 456                 break;
 457         case 2:
 458                 ret = read_default_ldt(ptr, bytecount);
 459                 break;
 460         case 0x11:
 461                 ret = write_ldt(ptr, bytecount, 0);
 462                 break;
 463         }
 464         /*
 465          * The SYSCALL_DEFINE() macros give us an 'unsigned long'
 466          * return type, but tht ABI for sys_modify_ldt() expects
 467          * 'int'.  This cast gives us an int-sized value in %rax
 468          * for the return code.  The 'unsigned' is necessary so
 469          * the compiler does not try to sign-extend the negative
 470          * return codes into the high half of the register when
 471          * taking the value from int->long.
 472          */
 473         return (unsigned int)ret;
 474 }