1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright (C) 1992 Krishna Balasubramanian and Linus Torvalds
4 * Copyright (C) 1999 Ingo Molnar <mingo@redhat.com>
5 * Copyright (C) 2002 Andi Kleen
7 * This handles calls from both 32bit and 64bit mode.
15 #include <linux/errno.h>
16 #include <linux/gfp.h>
17 #include <linux/sched.h>
18 #include <linux/string.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>
29 #include <asm/mmu_context.h>
30 #include <asm/pgtable_areas.h>
32 /* This is a multiple of PAGE_SIZE. */
33 #define LDT_SLOT_STRIDE (LDT_ENTRIES * LDT_ENTRY_SIZE)
35 static inline void *ldt_slot_va(int slot)
37 return (void *)(LDT_BASE_ADDR + LDT_SLOT_STRIDE * slot);
40 void load_mm_ldt(struct mm_struct *mm)
42 struct ldt_struct *ldt;
44 /* READ_ONCE synchronizes with smp_store_release */
45 ldt = READ_ONCE(mm->context.ldt);
48 * Any change to mm->context.ldt is followed by an IPI to all
49 * CPUs with the mm active. The LDT will not be freed until
50 * after the IPI is handled by all such CPUs. This means that,
51 * if the ldt_struct changes before we return, the values we see
52 * will be safe, and the new values will be loaded before we run
55 * NB: don't try to convert this to use RCU without extreme care.
56 * We would still need IRQs off, because we don't want to change
57 * the local LDT after an IPI loaded a newer value than the one
62 if (static_cpu_has(X86_FEATURE_PTI)) {
63 if (WARN_ON_ONCE((unsigned long)ldt->slot > 1)) {
65 * Whoops -- either the new LDT isn't mapped
66 * (if slot == -1) or is mapped into a bogus
74 * If page table isolation is enabled, ldt->entries
75 * will not be mapped in the userspace pagetables.
76 * Tell the CPU to access the LDT through the alias
77 * at ldt_slot_va(ldt->slot).
79 set_ldt(ldt_slot_va(ldt->slot), ldt->nr_entries);
81 set_ldt(ldt->entries, ldt->nr_entries);
88 void switch_ldt(struct mm_struct *prev, struct mm_struct *next)
91 * Load the LDT if either the old or new mm had an LDT.
93 * An mm will never go from having an LDT to not having an LDT. Two
94 * mms never share an LDT, so we don't gain anything by checking to
95 * see whether the LDT changed. There's also no guarantee that
96 * prev->context.ldt actually matches LDTR, but, if LDTR is non-NULL,
97 * then prev->context.ldt will also be non-NULL.
99 * If we really cared, we could optimize the case where prev == next
100 * and we're exiting lazy mode. Most of the time, if this happens,
101 * we don't actually need to reload LDTR, but modify_ldt() is mostly
102 * used by legacy code and emulators where we don't need this level of
105 * This uses | instead of || because it generates better code.
107 if (unlikely((unsigned long)prev->context.ldt |
108 (unsigned long)next->context.ldt))
111 DEBUG_LOCKS_WARN_ON(preemptible());
114 static void refresh_ldt_segments(void)
120 * Make sure that the cached DS and ES descriptors match the updated
123 savesegment(ds, sel);
124 if ((sel & SEGMENT_TI_MASK) == SEGMENT_LDT)
125 loadsegment(ds, sel);
127 savesegment(es, sel);
128 if ((sel & SEGMENT_TI_MASK) == SEGMENT_LDT)
129 loadsegment(es, sel);
133 /* context.lock is held by the task which issued the smp function call */
134 static void flush_ldt(void *__mm)
136 struct mm_struct *mm = __mm;
138 if (this_cpu_read(cpu_tlbstate.loaded_mm) != mm)
143 refresh_ldt_segments();
146 /* The caller must call finalize_ldt_struct on the result. LDT starts zeroed. */
147 static struct ldt_struct *alloc_ldt_struct(unsigned int num_entries)
149 struct ldt_struct *new_ldt;
150 unsigned int alloc_size;
152 if (num_entries > LDT_ENTRIES)
155 new_ldt = kmalloc(sizeof(struct ldt_struct), GFP_KERNEL);
159 BUILD_BUG_ON(LDT_ENTRY_SIZE != sizeof(struct desc_struct));
160 alloc_size = num_entries * LDT_ENTRY_SIZE;
163 * Xen is very picky: it requires a page-aligned LDT that has no
164 * trailing nonzero bytes in any page that contains LDT descriptors.
165 * Keep it simple: zero the whole allocation and never allocate less
168 if (alloc_size > PAGE_SIZE)
169 new_ldt->entries = vzalloc(alloc_size);
171 new_ldt->entries = (void *)get_zeroed_page(GFP_KERNEL);
173 if (!new_ldt->entries) {
178 /* The new LDT isn't aliased for PTI yet. */
181 new_ldt->nr_entries = num_entries;
185 #ifdef CONFIG_PAGE_TABLE_ISOLATION
187 static void do_sanity_check(struct mm_struct *mm,
188 bool had_kernel_mapping,
189 bool had_user_mapping)
191 if (mm->context.ldt) {
193 * We already had an LDT. The top-level entry should already
194 * have been allocated and synchronized with the usermode
197 WARN_ON(!had_kernel_mapping);
198 if (boot_cpu_has(X86_FEATURE_PTI))
199 WARN_ON(!had_user_mapping);
202 * This is the first time we're mapping an LDT for this process.
203 * Sync the pgd to the usermode tables.
205 WARN_ON(had_kernel_mapping);
206 if (boot_cpu_has(X86_FEATURE_PTI))
207 WARN_ON(had_user_mapping);
211 #ifdef CONFIG_X86_PAE
213 static pmd_t *pgd_to_pmd_walk(pgd_t *pgd, unsigned long va)
221 p4d = p4d_offset(pgd, va);
225 pud = pud_offset(p4d, va);
229 return pmd_offset(pud, va);
232 static void map_ldt_struct_to_user(struct mm_struct *mm)
234 pgd_t *k_pgd = pgd_offset(mm, LDT_BASE_ADDR);
235 pgd_t *u_pgd = kernel_to_user_pgdp(k_pgd);
236 pmd_t *k_pmd, *u_pmd;
238 k_pmd = pgd_to_pmd_walk(k_pgd, LDT_BASE_ADDR);
239 u_pmd = pgd_to_pmd_walk(u_pgd, LDT_BASE_ADDR);
241 if (boot_cpu_has(X86_FEATURE_PTI) && !mm->context.ldt)
242 set_pmd(u_pmd, *k_pmd);
245 static void sanity_check_ldt_mapping(struct mm_struct *mm)
247 pgd_t *k_pgd = pgd_offset(mm, LDT_BASE_ADDR);
248 pgd_t *u_pgd = kernel_to_user_pgdp(k_pgd);
249 bool had_kernel, had_user;
250 pmd_t *k_pmd, *u_pmd;
252 k_pmd = pgd_to_pmd_walk(k_pgd, LDT_BASE_ADDR);
253 u_pmd = pgd_to_pmd_walk(u_pgd, LDT_BASE_ADDR);
254 had_kernel = (k_pmd->pmd != 0);
255 had_user = (u_pmd->pmd != 0);
257 do_sanity_check(mm, had_kernel, had_user);
260 #else /* !CONFIG_X86_PAE */
262 static void map_ldt_struct_to_user(struct mm_struct *mm)
264 pgd_t *pgd = pgd_offset(mm, LDT_BASE_ADDR);
266 if (boot_cpu_has(X86_FEATURE_PTI) && !mm->context.ldt)
267 set_pgd(kernel_to_user_pgdp(pgd), *pgd);
270 static void sanity_check_ldt_mapping(struct mm_struct *mm)
272 pgd_t *pgd = pgd_offset(mm, LDT_BASE_ADDR);
273 bool had_kernel = (pgd->pgd != 0);
274 bool had_user = (kernel_to_user_pgdp(pgd)->pgd != 0);
276 do_sanity_check(mm, had_kernel, had_user);
279 #endif /* CONFIG_X86_PAE */
282 * If PTI is enabled, this maps the LDT into the kernelmode and
283 * usermode tables for the given mm.
286 map_ldt_struct(struct mm_struct *mm, struct ldt_struct *ldt, int slot)
293 if (!boot_cpu_has(X86_FEATURE_PTI))
297 * Any given ldt_struct should have map_ldt_struct() called at most
300 WARN_ON(ldt->slot != -1);
302 /* Check if the current mappings are sane */
303 sanity_check_ldt_mapping(mm);
305 is_vmalloc = is_vmalloc_addr(ldt->entries);
307 nr_pages = DIV_ROUND_UP(ldt->nr_entries * LDT_ENTRY_SIZE, PAGE_SIZE);
309 for (i = 0; i < nr_pages; i++) {
310 unsigned long offset = i << PAGE_SHIFT;
311 const void *src = (char *)ldt->entries + offset;
316 va = (unsigned long)ldt_slot_va(slot) + offset;
317 pfn = is_vmalloc ? vmalloc_to_pfn(src) :
318 page_to_pfn(virt_to_page(src));
320 * Treat the PTI LDT range as a *userspace* range.
321 * get_locked_pte() will allocate all needed pagetables
322 * and account for them in this mm.
324 ptep = get_locked_pte(mm, va, &ptl);
328 * Map it RO so the easy to find address is not a primary
329 * target via some kernel interface which misses a
332 pte_prot = __pgprot(__PAGE_KERNEL_RO & ~_PAGE_GLOBAL);
333 /* Filter out unsuppored __PAGE_KERNEL* bits: */
334 pgprot_val(pte_prot) &= __supported_pte_mask;
335 pte = pfn_pte(pfn, pte_prot);
336 set_pte_at(mm, va, ptep, pte);
337 pte_unmap_unlock(ptep, ptl);
340 /* Propagate LDT mapping to the user page-table */
341 map_ldt_struct_to_user(mm);
347 static void unmap_ldt_struct(struct mm_struct *mm, struct ldt_struct *ldt)
355 /* LDT map/unmap is only required for PTI */
356 if (!boot_cpu_has(X86_FEATURE_PTI))
359 nr_pages = DIV_ROUND_UP(ldt->nr_entries * LDT_ENTRY_SIZE, PAGE_SIZE);
361 for (i = 0; i < nr_pages; i++) {
362 unsigned long offset = i << PAGE_SHIFT;
366 va = (unsigned long)ldt_slot_va(ldt->slot) + offset;
367 ptep = get_locked_pte(mm, va, &ptl);
368 pte_clear(mm, va, ptep);
369 pte_unmap_unlock(ptep, ptl);
372 va = (unsigned long)ldt_slot_va(ldt->slot);
373 flush_tlb_mm_range(mm, va, va + nr_pages * PAGE_SIZE, PAGE_SHIFT, false);
376 #else /* !CONFIG_PAGE_TABLE_ISOLATION */
379 map_ldt_struct(struct mm_struct *mm, struct ldt_struct *ldt, int slot)
384 static void unmap_ldt_struct(struct mm_struct *mm, struct ldt_struct *ldt)
387 #endif /* CONFIG_PAGE_TABLE_ISOLATION */
389 static void free_ldt_pgtables(struct mm_struct *mm)
391 #ifdef CONFIG_PAGE_TABLE_ISOLATION
392 struct mmu_gather tlb;
393 unsigned long start = LDT_BASE_ADDR;
394 unsigned long end = LDT_END_ADDR;
396 if (!boot_cpu_has(X86_FEATURE_PTI))
399 tlb_gather_mmu(&tlb, mm, start, end);
400 free_pgd_range(&tlb, start, end, start, end);
401 tlb_finish_mmu(&tlb, start, end);
405 /* After calling this, the LDT is immutable. */
406 static void finalize_ldt_struct(struct ldt_struct *ldt)
408 paravirt_alloc_ldt(ldt->entries, ldt->nr_entries);
411 static void install_ldt(struct mm_struct *mm, struct ldt_struct *ldt)
413 mutex_lock(&mm->context.lock);
415 /* Synchronizes with READ_ONCE in load_mm_ldt. */
416 smp_store_release(&mm->context.ldt, ldt);
418 /* Activate the LDT for all CPUs using currents mm. */
419 on_each_cpu_mask(mm_cpumask(mm), flush_ldt, mm, true);
421 mutex_unlock(&mm->context.lock);
424 static void free_ldt_struct(struct ldt_struct *ldt)
429 paravirt_free_ldt(ldt->entries, ldt->nr_entries);
430 if (ldt->nr_entries * LDT_ENTRY_SIZE > PAGE_SIZE)
431 vfree_atomic(ldt->entries);
433 free_page((unsigned long)ldt->entries);
438 * Called on fork from arch_dup_mmap(). Just copy the current LDT state,
439 * the new task is not running, so nothing can be installed.
441 int ldt_dup_context(struct mm_struct *old_mm, struct mm_struct *mm)
443 struct ldt_struct *new_ldt;
449 mutex_lock(&old_mm->context.lock);
450 if (!old_mm->context.ldt)
453 new_ldt = alloc_ldt_struct(old_mm->context.ldt->nr_entries);
459 memcpy(new_ldt->entries, old_mm->context.ldt->entries,
460 new_ldt->nr_entries * LDT_ENTRY_SIZE);
461 finalize_ldt_struct(new_ldt);
463 retval = map_ldt_struct(mm, new_ldt, 0);
465 free_ldt_pgtables(mm);
466 free_ldt_struct(new_ldt);
469 mm->context.ldt = new_ldt;
472 mutex_unlock(&old_mm->context.lock);
477 * No need to lock the MM as we are the last user
479 * 64bit: Don't touch the LDT register - we're already in the next thread.
481 void destroy_context_ldt(struct mm_struct *mm)
483 free_ldt_struct(mm->context.ldt);
484 mm->context.ldt = NULL;
487 void ldt_arch_exit_mmap(struct mm_struct *mm)
489 free_ldt_pgtables(mm);
492 static int read_ldt(void __user *ptr, unsigned long bytecount)
494 struct mm_struct *mm = current->mm;
495 unsigned long entries_size;
498 down_read(&mm->context.ldt_usr_sem);
500 if (!mm->context.ldt) {
505 if (bytecount > LDT_ENTRY_SIZE * LDT_ENTRIES)
506 bytecount = LDT_ENTRY_SIZE * LDT_ENTRIES;
508 entries_size = mm->context.ldt->nr_entries * LDT_ENTRY_SIZE;
509 if (entries_size > bytecount)
510 entries_size = bytecount;
512 if (copy_to_user(ptr, mm->context.ldt->entries, entries_size)) {
517 if (entries_size != bytecount) {
518 /* Zero-fill the rest and pretend we read bytecount bytes. */
519 if (clear_user(ptr + entries_size, bytecount - entries_size)) {
527 up_read(&mm->context.ldt_usr_sem);
531 static int read_default_ldt(void __user *ptr, unsigned long bytecount)
533 /* CHECKME: Can we use _one_ random number ? */
535 unsigned long size = 5 * sizeof(struct desc_struct);
537 unsigned long size = 128;
539 if (bytecount > size)
541 if (clear_user(ptr, bytecount))
546 static int write_ldt(void __user *ptr, unsigned long bytecount, int oldmode)
548 struct mm_struct *mm = current->mm;
549 struct ldt_struct *new_ldt, *old_ldt;
550 unsigned int old_nr_entries, new_nr_entries;
551 struct user_desc ldt_info;
552 struct desc_struct ldt;
556 if (bytecount != sizeof(ldt_info))
559 if (copy_from_user(&ldt_info, ptr, sizeof(ldt_info)))
563 if (ldt_info.entry_number >= LDT_ENTRIES)
565 if (ldt_info.contents == 3) {
568 if (ldt_info.seg_not_present == 0)
572 if ((oldmode && !ldt_info.base_addr && !ldt_info.limit) ||
573 LDT_empty(&ldt_info)) {
574 /* The user wants to clear the entry. */
575 memset(&ldt, 0, sizeof(ldt));
577 if (!IS_ENABLED(CONFIG_X86_16BIT) && !ldt_info.seg_32bit) {
582 fill_ldt(&ldt, &ldt_info);
587 if (down_write_killable(&mm->context.ldt_usr_sem))
590 old_ldt = mm->context.ldt;
591 old_nr_entries = old_ldt ? old_ldt->nr_entries : 0;
592 new_nr_entries = max(ldt_info.entry_number + 1, old_nr_entries);
595 new_ldt = alloc_ldt_struct(new_nr_entries);
600 memcpy(new_ldt->entries, old_ldt->entries, old_nr_entries * LDT_ENTRY_SIZE);
602 new_ldt->entries[ldt_info.entry_number] = ldt;
603 finalize_ldt_struct(new_ldt);
606 * If we are using PTI, map the new LDT into the userspace pagetables.
607 * If there is already an LDT, use the other slot so that other CPUs
608 * will continue to use the old LDT until install_ldt() switches
609 * them over to the new LDT.
611 error = map_ldt_struct(mm, new_ldt, old_ldt ? !old_ldt->slot : 0);
614 * This only can fail for the first LDT setup. If an LDT is
615 * already installed then the PTE page is already
616 * populated. Mop up a half populated page table.
618 if (!WARN_ON_ONCE(old_ldt))
619 free_ldt_pgtables(mm);
620 free_ldt_struct(new_ldt);
624 install_ldt(mm, new_ldt);
625 unmap_ldt_struct(mm, old_ldt);
626 free_ldt_struct(old_ldt);
630 up_write(&mm->context.ldt_usr_sem);
635 SYSCALL_DEFINE3(modify_ldt, int , func , void __user * , ptr ,
636 unsigned long , bytecount)
642 ret = read_ldt(ptr, bytecount);
645 ret = write_ldt(ptr, bytecount, 1);
648 ret = read_default_ldt(ptr, bytecount);
651 ret = write_ldt(ptr, bytecount, 0);
655 * The SYSCALL_DEFINE() macros give us an 'unsigned long'
656 * return type, but tht ABI for sys_modify_ldt() expects
657 * 'int'. This cast gives us an int-sized value in %rax
658 * for the return code. The 'unsigned' is necessary so
659 * the compiler does not try to sign-extend the negative
660 * return codes into the high half of the register when
661 * taking the value from int->long.
663 return (unsigned int)ret;