2 * Copyright(c) 2017 Intel Corporation. All rights reserved.
4 * This program is free software; you can redistribute it and/or modify
5 * it under the terms of version 2 of the GNU General Public License as
6 * published by the Free Software Foundation.
8 * This program is distributed in the hope that it will be useful, but
9 * WITHOUT ANY WARRANTY; without even the implied warranty of
10 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
11 * General Public License for more details.
13 * This code is based in part on work published here:
15 * https://github.com/IAIK/KAISER
17 * The original work was written by and and signed off by for the Linux
20 * Signed-off-by: Richard Fellner <richard.fellner@student.tugraz.at>
21 * Signed-off-by: Moritz Lipp <moritz.lipp@iaik.tugraz.at>
22 * Signed-off-by: Daniel Gruss <daniel.gruss@iaik.tugraz.at>
23 * Signed-off-by: Michael Schwarz <michael.schwarz@iaik.tugraz.at>
25 * Major changes to the original code by: Dave Hansen <dave.hansen@intel.com>
26 * Mostly rewritten by Thomas Gleixner <tglx@linutronix.de> and
27 * Andy Lutomirsky <luto@amacapital.net>
29 #include <linux/kernel.h>
30 #include <linux/errno.h>
31 #include <linux/string.h>
32 #include <linux/types.h>
33 #include <linux/bug.h>
34 #include <linux/init.h>
35 #include <linux/spinlock.h>
37 #include <linux/uaccess.h>
39 #include <asm/cpufeature.h>
40 #include <asm/hypervisor.h>
41 #include <asm/vsyscall.h>
42 #include <asm/cmdline.h>
44 #include <asm/pgtable.h>
45 #include <asm/pgalloc.h>
46 #include <asm/tlbflush.h>
50 #define pr_fmt(fmt) "Kernel/User page tables isolation: " fmt
52 /* Backporting helper */
54 #define __GFP_NOTRACK 0
58 * Define the page-table levels we clone for user-space on 32
62 #define PTI_LEVEL_KERNEL_IMAGE PTI_CLONE_PMD
64 #define PTI_LEVEL_KERNEL_IMAGE PTI_CLONE_PTE
67 static void __init pti_print_if_insecure(const char *reason)
69 if (boot_cpu_has_bug(X86_BUG_CPU_MELTDOWN))
70 pr_info("%s\n", reason);
73 static void __init pti_print_if_secure(const char *reason)
75 if (!boot_cpu_has_bug(X86_BUG_CPU_MELTDOWN))
76 pr_info("%s\n", reason);
85 void __init pti_check_boottime_disable(void)
90 /* Assume mode is auto unless overridden. */
93 if (hypervisor_is_type(X86_HYPER_XEN_PV)) {
94 pti_mode = PTI_FORCE_OFF;
95 pti_print_if_insecure("disabled on XEN PV.");
99 ret = cmdline_find_option(boot_command_line, "pti", arg, sizeof(arg));
101 if (ret == 3 && !strncmp(arg, "off", 3)) {
102 pti_mode = PTI_FORCE_OFF;
103 pti_print_if_insecure("disabled on command line.");
106 if (ret == 2 && !strncmp(arg, "on", 2)) {
107 pti_mode = PTI_FORCE_ON;
108 pti_print_if_secure("force enabled on command line.");
111 if (ret == 4 && !strncmp(arg, "auto", 4)) {
117 if (cmdline_find_option_bool(boot_command_line, "nopti")) {
118 pti_mode = PTI_FORCE_OFF;
119 pti_print_if_insecure("disabled on command line.");
124 if (!boot_cpu_has_bug(X86_BUG_CPU_MELTDOWN))
127 setup_force_cpu_cap(X86_FEATURE_PTI);
130 pgd_t __pti_set_user_pgtbl(pgd_t *pgdp, pgd_t pgd)
133 * Changes to the high (kernel) portion of the kernelmode page
134 * tables are not automatically propagated to the usermode tables.
136 * Users should keep in mind that, unlike the kernelmode tables,
137 * there is no vmalloc_fault equivalent for the usermode tables.
138 * Top-level entries added to init_mm's usermode pgd after boot
139 * will not be automatically propagated to other mms.
141 if (!pgdp_maps_userspace(pgdp))
145 * The user page tables get the full PGD, accessible from
148 kernel_to_user_pgdp(pgdp)->pgd = pgd.pgd;
151 * If this is normal user memory, make it NX in the kernel
152 * pagetables so that, if we somehow screw up and return to
153 * usermode with the kernel CR3 loaded, we'll get a page fault
154 * instead of allowing user code to execute with the wrong CR3.
156 * As exceptions, we don't set NX if:
157 * - _PAGE_USER is not set. This could be an executable
158 * EFI runtime mapping or something similar, and the kernel
159 * may execute from it
160 * - we don't have NX support
161 * - we're clearing the PGD (i.e. the new pgd is not present).
163 if ((pgd.pgd & (_PAGE_USER|_PAGE_PRESENT)) == (_PAGE_USER|_PAGE_PRESENT) &&
164 (__supported_pte_mask & _PAGE_NX))
167 /* return the copy of the PGD we want the kernel to use: */
172 * Walk the user copy of the page tables (optionally) trying to allocate
173 * page table pages on the way down.
175 * Returns a pointer to a P4D on success, or NULL on failure.
177 static p4d_t *pti_user_pagetable_walk_p4d(unsigned long address)
179 pgd_t *pgd = kernel_to_user_pgdp(pgd_offset_k(address));
180 gfp_t gfp = (GFP_KERNEL | __GFP_NOTRACK | __GFP_ZERO);
182 if (address < PAGE_OFFSET) {
183 WARN_ONCE(1, "attempt to walk user address\n");
187 if (pgd_none(*pgd)) {
188 unsigned long new_p4d_page = __get_free_page(gfp);
189 if (WARN_ON_ONCE(!new_p4d_page))
192 set_pgd(pgd, __pgd(_KERNPG_TABLE | __pa(new_p4d_page)));
194 BUILD_BUG_ON(pgd_large(*pgd) != 0);
196 return p4d_offset(pgd, address);
200 * Walk the user copy of the page tables (optionally) trying to allocate
201 * page table pages on the way down.
203 * Returns a pointer to a PMD on success, or NULL on failure.
205 static pmd_t *pti_user_pagetable_walk_pmd(unsigned long address)
207 gfp_t gfp = (GFP_KERNEL | __GFP_NOTRACK | __GFP_ZERO);
211 p4d = pti_user_pagetable_walk_p4d(address);
215 BUILD_BUG_ON(p4d_large(*p4d) != 0);
216 if (p4d_none(*p4d)) {
217 unsigned long new_pud_page = __get_free_page(gfp);
218 if (WARN_ON_ONCE(!new_pud_page))
221 set_p4d(p4d, __p4d(_KERNPG_TABLE | __pa(new_pud_page)));
224 pud = pud_offset(p4d, address);
225 /* The user page tables do not use large mappings: */
226 if (pud_large(*pud)) {
230 if (pud_none(*pud)) {
231 unsigned long new_pmd_page = __get_free_page(gfp);
232 if (WARN_ON_ONCE(!new_pmd_page))
235 set_pud(pud, __pud(_KERNPG_TABLE | __pa(new_pmd_page)));
238 return pmd_offset(pud, address);
242 * Walk the shadow copy of the page tables (optionally) trying to allocate
243 * page table pages on the way down. Does not support large pages.
245 * Note: this is only used when mapping *new* kernel data into the
246 * user/shadow page tables. It is never used for userspace data.
248 * Returns a pointer to a PTE on success, or NULL on failure.
250 static __init pte_t *pti_user_pagetable_walk_pte(unsigned long address)
252 gfp_t gfp = (GFP_KERNEL | __GFP_NOTRACK | __GFP_ZERO);
256 pmd = pti_user_pagetable_walk_pmd(address);
260 /* We can't do anything sensible if we hit a large mapping. */
261 if (pmd_large(*pmd)) {
266 if (pmd_none(*pmd)) {
267 unsigned long new_pte_page = __get_free_page(gfp);
271 set_pmd(pmd, __pmd(_KERNPG_TABLE | __pa(new_pte_page)));
274 pte = pte_offset_kernel(pmd, address);
275 if (pte_flags(*pte) & _PAGE_USER) {
276 WARN_ONCE(1, "attempt to walk to user pte\n");
282 #ifdef CONFIG_X86_VSYSCALL_EMULATION
283 static void __init pti_setup_vsyscall(void)
285 pte_t *pte, *target_pte;
288 pte = lookup_address(VSYSCALL_ADDR, &level);
289 if (!pte || WARN_ON(level != PG_LEVEL_4K) || pte_none(*pte))
292 target_pte = pti_user_pagetable_walk_pte(VSYSCALL_ADDR);
293 if (WARN_ON(!target_pte))
297 set_vsyscall_pgtable_user_bits(kernel_to_user_pgdp(swapper_pg_dir));
300 static void __init pti_setup_vsyscall(void) { }
303 enum pti_clone_level {
309 pti_clone_pgtable(unsigned long start, unsigned long end,
310 enum pti_clone_level level)
315 * Clone the populated PMDs which cover start to end. These PMD areas
318 for (addr = start; addr < end;) {
319 pte_t *pte, *target_pte;
320 pmd_t *pmd, *target_pmd;
329 pgd = pgd_offset_k(addr);
330 if (WARN_ON(pgd_none(*pgd)))
332 p4d = p4d_offset(pgd, addr);
333 if (WARN_ON(p4d_none(*p4d)))
336 pud = pud_offset(p4d, addr);
337 if (pud_none(*pud)) {
342 pmd = pmd_offset(pud, addr);
343 if (pmd_none(*pmd)) {
348 if (pmd_large(*pmd) || level == PTI_CLONE_PMD) {
349 target_pmd = pti_user_pagetable_walk_pmd(addr);
350 if (WARN_ON(!target_pmd))
354 * Only clone present PMDs. This ensures only setting
355 * _PAGE_GLOBAL on present PMDs. This should only be
356 * called on well-known addresses anyway, so a non-
357 * present PMD would be a surprise.
359 if (WARN_ON(!(pmd_flags(*pmd) & _PAGE_PRESENT)))
363 * Setting 'target_pmd' below creates a mapping in both
364 * the user and kernel page tables. It is effectively
365 * global, so set it as global in both copies. Note:
366 * the X86_FEATURE_PGE check is not _required_ because
367 * the CPU ignores _PAGE_GLOBAL when PGE is not
368 * supported. The check keeps consistentency with
369 * code that only set this bit when supported.
371 if (boot_cpu_has(X86_FEATURE_PGE))
372 *pmd = pmd_set_flags(*pmd, _PAGE_GLOBAL);
375 * Copy the PMD. That is, the kernelmode and usermode
376 * tables will share the last-level page tables of this
383 } else if (level == PTI_CLONE_PTE) {
385 /* Walk the page-table down to the pte level */
386 pte = pte_offset_kernel(pmd, addr);
387 if (pte_none(*pte)) {
392 /* Only clone present PTEs */
393 if (WARN_ON(!(pte_flags(*pte) & _PAGE_PRESENT)))
396 /* Allocate PTE in the user page-table */
397 target_pte = pti_user_pagetable_walk_pte(addr);
398 if (WARN_ON(!target_pte))
401 /* Set GLOBAL bit in both PTEs */
402 if (boot_cpu_has(X86_FEATURE_PGE))
403 *pte = pte_set_flags(*pte, _PAGE_GLOBAL);
418 * Clone a single p4d (i.e. a top-level entry on 4-level systems and a
419 * next-level entry on 5-level systems.
421 static void __init pti_clone_p4d(unsigned long addr)
423 p4d_t *kernel_p4d, *user_p4d;
426 user_p4d = pti_user_pagetable_walk_p4d(addr);
430 kernel_pgd = pgd_offset_k(addr);
431 kernel_p4d = p4d_offset(kernel_pgd, addr);
432 *user_p4d = *kernel_p4d;
436 * Clone the CPU_ENTRY_AREA into the user space visible page table.
438 static void __init pti_clone_user_shared(void)
440 pti_clone_p4d(CPU_ENTRY_AREA_BASE);
443 #else /* CONFIG_X86_64 */
446 * On 32 bit PAE systems with 1GB of Kernel address space there is only
447 * one pgd/p4d for the whole kernel. Cloning that would map the whole
448 * address space into the user page-tables, making PTI useless. So clone
449 * the page-table on the PMD level to prevent that.
451 static void __init pti_clone_user_shared(void)
453 unsigned long start, end;
455 start = CPU_ENTRY_AREA_BASE;
456 end = start + (PAGE_SIZE * CPU_ENTRY_AREA_PAGES);
458 pti_clone_pgtable(start, end, PTI_CLONE_PMD);
460 #endif /* CONFIG_X86_64 */
463 * Clone the ESPFIX P4D into the user space visible page table
465 static void __init pti_setup_espfix64(void)
467 #ifdef CONFIG_X86_ESPFIX64
468 pti_clone_p4d(ESPFIX_BASE_ADDR);
473 * Clone the populated PMDs of the entry and irqentry text and force it RO.
475 static void pti_clone_entry_text(void)
477 pti_clone_pgtable((unsigned long) __entry_text_start,
478 (unsigned long) __irqentry_text_end,
483 * Global pages and PCIDs are both ways to make kernel TLB entries
484 * live longer, reduce TLB misses and improve kernel performance.
485 * But, leaving all kernel text Global makes it potentially accessible
486 * to Meltdown-style attacks which make it trivial to find gadgets or
489 * Only use global pages when it is really worth it.
491 static inline bool pti_kernel_image_global_ok(void)
494 * Systems with PCIDs get litlle benefit from global
495 * kernel text and are not worth the downsides.
497 if (cpu_feature_enabled(X86_FEATURE_PCID))
501 * Only do global kernel image for pti=auto. Do the most
502 * secure thing (not global) if pti=on specified.
504 if (pti_mode != PTI_AUTO)
508 * K8 may not tolerate the cleared _PAGE_RW on the userspace
509 * global kernel image pages. Do the safe thing (disable
510 * global kernel image). This is unlikely to ever be
511 * noticed because PTI is disabled by default on AMD CPUs.
513 if (boot_cpu_has(X86_FEATURE_K8))
517 * RANDSTRUCT derives its hardening benefits from the
518 * attacker's lack of knowledge about the layout of kernel
519 * data structures. Keep the kernel image non-global in
520 * cases where RANDSTRUCT is in use to help keep the layout a
523 if (IS_ENABLED(CONFIG_GCC_PLUGIN_RANDSTRUCT))
530 * This is the only user for these and it is not arch-generic
531 * like the other set_memory.h functions. Just extern them.
533 extern int set_memory_nonglobal(unsigned long addr, int numpages);
534 extern int set_memory_global(unsigned long addr, int numpages);
537 * For some configurations, map all of kernel text into the user page
538 * tables. This reduces TLB misses, especially on non-PCID systems.
540 static void pti_clone_kernel_text(void)
543 * rodata is part of the kernel image and is normally
544 * readable on the filesystem or on the web. But, do not
545 * clone the areas past rodata, they might contain secrets.
547 unsigned long start = PFN_ALIGN(_text);
548 unsigned long end_clone = (unsigned long)__end_rodata_aligned;
549 unsigned long end_global = PFN_ALIGN((unsigned long)__stop___ex_table);
551 if (!pti_kernel_image_global_ok())
554 pr_debug("mapping partial kernel image into user address space\n");
557 * Note that this will undo _some_ of the work that
558 * pti_set_kernel_image_nonglobal() did to clear the
561 pti_clone_pgtable(start, end_clone, PTI_LEVEL_KERNEL_IMAGE);
564 * pti_clone_pgtable() will set the global bit in any PMDs
565 * that it clones, but we also need to get any PTEs in
566 * the last level for areas that are not huge-page-aligned.
569 /* Set the global bit for normal non-__init kernel text: */
570 set_memory_global(start, (end_global - start) >> PAGE_SHIFT);
573 void pti_set_kernel_image_nonglobal(void)
576 * The identity map is created with PMDs, regardless of the
577 * actual length of the kernel. We need to clear
578 * _PAGE_GLOBAL up to a PMD boundary, not just to the end
581 unsigned long start = PFN_ALIGN(_text);
582 unsigned long end = ALIGN((unsigned long)_end, PMD_PAGE_SIZE);
585 * This clears _PAGE_GLOBAL from the entire kernel image.
586 * pti_clone_kernel_text() map put _PAGE_GLOBAL back for
587 * areas that are mapped to userspace.
589 set_memory_nonglobal(start, (end - start) >> PAGE_SHIFT);
593 * Initialize kernel page table isolation
595 void __init pti_init(void)
597 if (!static_cpu_has(X86_FEATURE_PTI))
600 pr_info("enabled\n");
604 * We check for X86_FEATURE_PCID here. But the init-code will
605 * clear the feature flag on 32 bit because the feature is not
606 * supported on 32 bit anyway. To print the warning we need to
607 * check with cpuid directly again.
609 if (cpuid_ecx(0x1) & BIT(17)) {
610 /* Use printk to work around pr_fmt() */
611 printk(KERN_WARNING "\n");
612 printk(KERN_WARNING "************************************************************\n");
613 printk(KERN_WARNING "** WARNING! WARNING! WARNING! WARNING! WARNING! WARNING! **\n");
614 printk(KERN_WARNING "** **\n");
615 printk(KERN_WARNING "** You are using 32-bit PTI on a 64-bit PCID-capable CPU. **\n");
616 printk(KERN_WARNING "** Your performance will increase dramatically if you **\n");
617 printk(KERN_WARNING "** switch to a 64-bit kernel! **\n");
618 printk(KERN_WARNING "** **\n");
619 printk(KERN_WARNING "** WARNING! WARNING! WARNING! WARNING! WARNING! WARNING! **\n");
620 printk(KERN_WARNING "************************************************************\n");
624 pti_clone_user_shared();
626 /* Undo all global bits from the init pagetables in head_64.S: */
627 pti_set_kernel_image_nonglobal();
628 /* Replace some of the global bits just for shared entry text: */
629 pti_clone_entry_text();
630 pti_setup_espfix64();
631 pti_setup_vsyscall();
635 * Finalize the kernel mappings in the userspace page-table. Some of the
636 * mappings for the kernel image might have changed since pti_init()
637 * cloned them. This is because parts of the kernel image have been
638 * mapped RO and/or NX. These changes need to be cloned again to the
639 * userspace page-table.
641 void pti_finalize(void)
644 * We need to clone everything (again) that maps parts of the
647 pti_clone_entry_text();
648 pti_clone_kernel_text();
650 debug_checkwx_user();