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>
38 #include <linux/cpu.h>
40 #include <asm/cpufeature.h>
41 #include <asm/hypervisor.h>
42 #include <asm/vsyscall.h>
43 #include <asm/cmdline.h>
45 #include <asm/pgtable.h>
46 #include <asm/pgalloc.h>
47 #include <asm/tlbflush.h>
49 #include <asm/sections.h>
52 #define pr_fmt(fmt) "Kernel/User page tables isolation: " fmt
54 /* Backporting helper */
56 #define __GFP_NOTRACK 0
60 * Define the page-table levels we clone for user-space on 32
64 #define PTI_LEVEL_KERNEL_IMAGE PTI_CLONE_PMD
66 #define PTI_LEVEL_KERNEL_IMAGE PTI_CLONE_PTE
69 static void __init pti_print_if_insecure(const char *reason)
71 if (boot_cpu_has_bug(X86_BUG_CPU_MELTDOWN))
72 pr_info("%s\n", reason);
75 static void __init pti_print_if_secure(const char *reason)
77 if (!boot_cpu_has_bug(X86_BUG_CPU_MELTDOWN))
78 pr_info("%s\n", reason);
81 static enum pti_mode {
87 void __init pti_check_boottime_disable(void)
92 /* Assume mode is auto unless overridden. */
95 if (hypervisor_is_type(X86_HYPER_XEN_PV)) {
96 pti_mode = PTI_FORCE_OFF;
97 pti_print_if_insecure("disabled on XEN PV.");
101 ret = cmdline_find_option(boot_command_line, "pti", arg, sizeof(arg));
103 if (ret == 3 && !strncmp(arg, "off", 3)) {
104 pti_mode = PTI_FORCE_OFF;
105 pti_print_if_insecure("disabled on command line.");
108 if (ret == 2 && !strncmp(arg, "on", 2)) {
109 pti_mode = PTI_FORCE_ON;
110 pti_print_if_secure("force enabled on command line.");
113 if (ret == 4 && !strncmp(arg, "auto", 4)) {
119 if (cmdline_find_option_bool(boot_command_line, "nopti") ||
120 cpu_mitigations_off()) {
121 pti_mode = PTI_FORCE_OFF;
122 pti_print_if_insecure("disabled on command line.");
127 if (!boot_cpu_has_bug(X86_BUG_CPU_MELTDOWN))
130 setup_force_cpu_cap(X86_FEATURE_PTI);
133 pgd_t __pti_set_user_pgtbl(pgd_t *pgdp, pgd_t pgd)
136 * Changes to the high (kernel) portion of the kernelmode page
137 * tables are not automatically propagated to the usermode tables.
139 * Users should keep in mind that, unlike the kernelmode tables,
140 * there is no vmalloc_fault equivalent for the usermode tables.
141 * Top-level entries added to init_mm's usermode pgd after boot
142 * will not be automatically propagated to other mms.
144 if (!pgdp_maps_userspace(pgdp))
148 * The user page tables get the full PGD, accessible from
151 kernel_to_user_pgdp(pgdp)->pgd = pgd.pgd;
154 * If this is normal user memory, make it NX in the kernel
155 * pagetables so that, if we somehow screw up and return to
156 * usermode with the kernel CR3 loaded, we'll get a page fault
157 * instead of allowing user code to execute with the wrong CR3.
159 * As exceptions, we don't set NX if:
160 * - _PAGE_USER is not set. This could be an executable
161 * EFI runtime mapping or something similar, and the kernel
162 * may execute from it
163 * - we don't have NX support
164 * - we're clearing the PGD (i.e. the new pgd is not present).
166 if ((pgd.pgd & (_PAGE_USER|_PAGE_PRESENT)) == (_PAGE_USER|_PAGE_PRESENT) &&
167 (__supported_pte_mask & _PAGE_NX))
170 /* return the copy of the PGD we want the kernel to use: */
175 * Walk the user copy of the page tables (optionally) trying to allocate
176 * page table pages on the way down.
178 * Returns a pointer to a P4D on success, or NULL on failure.
180 static p4d_t *pti_user_pagetable_walk_p4d(unsigned long address)
182 pgd_t *pgd = kernel_to_user_pgdp(pgd_offset_k(address));
183 gfp_t gfp = (GFP_KERNEL | __GFP_NOTRACK | __GFP_ZERO);
185 if (address < PAGE_OFFSET) {
186 WARN_ONCE(1, "attempt to walk user address\n");
190 if (pgd_none(*pgd)) {
191 unsigned long new_p4d_page = __get_free_page(gfp);
192 if (WARN_ON_ONCE(!new_p4d_page))
195 set_pgd(pgd, __pgd(_KERNPG_TABLE | __pa(new_p4d_page)));
197 BUILD_BUG_ON(pgd_large(*pgd) != 0);
199 return p4d_offset(pgd, address);
203 * Walk the user copy of the page tables (optionally) trying to allocate
204 * page table pages on the way down.
206 * Returns a pointer to a PMD on success, or NULL on failure.
208 static pmd_t *pti_user_pagetable_walk_pmd(unsigned long address)
210 gfp_t gfp = (GFP_KERNEL | __GFP_NOTRACK | __GFP_ZERO);
214 p4d = pti_user_pagetable_walk_p4d(address);
218 BUILD_BUG_ON(p4d_large(*p4d) != 0);
219 if (p4d_none(*p4d)) {
220 unsigned long new_pud_page = __get_free_page(gfp);
221 if (WARN_ON_ONCE(!new_pud_page))
224 set_p4d(p4d, __p4d(_KERNPG_TABLE | __pa(new_pud_page)));
227 pud = pud_offset(p4d, address);
228 /* The user page tables do not use large mappings: */
229 if (pud_large(*pud)) {
233 if (pud_none(*pud)) {
234 unsigned long new_pmd_page = __get_free_page(gfp);
235 if (WARN_ON_ONCE(!new_pmd_page))
238 set_pud(pud, __pud(_KERNPG_TABLE | __pa(new_pmd_page)));
241 return pmd_offset(pud, address);
245 * Walk the shadow copy of the page tables (optionally) trying to allocate
246 * page table pages on the way down. Does not support large pages.
248 * Note: this is only used when mapping *new* kernel data into the
249 * user/shadow page tables. It is never used for userspace data.
251 * Returns a pointer to a PTE on success, or NULL on failure.
253 static pte_t *pti_user_pagetable_walk_pte(unsigned long address)
255 gfp_t gfp = (GFP_KERNEL | __GFP_NOTRACK | __GFP_ZERO);
259 pmd = pti_user_pagetable_walk_pmd(address);
263 /* We can't do anything sensible if we hit a large mapping. */
264 if (pmd_large(*pmd)) {
269 if (pmd_none(*pmd)) {
270 unsigned long new_pte_page = __get_free_page(gfp);
274 set_pmd(pmd, __pmd(_KERNPG_TABLE | __pa(new_pte_page)));
277 pte = pte_offset_kernel(pmd, address);
278 if (pte_flags(*pte) & _PAGE_USER) {
279 WARN_ONCE(1, "attempt to walk to user pte\n");
285 #ifdef CONFIG_X86_VSYSCALL_EMULATION
286 static void __init pti_setup_vsyscall(void)
288 pte_t *pte, *target_pte;
291 pte = lookup_address(VSYSCALL_ADDR, &level);
292 if (!pte || WARN_ON(level != PG_LEVEL_4K) || pte_none(*pte))
295 target_pte = pti_user_pagetable_walk_pte(VSYSCALL_ADDR);
296 if (WARN_ON(!target_pte))
300 set_vsyscall_pgtable_user_bits(kernel_to_user_pgdp(swapper_pg_dir));
303 static void __init pti_setup_vsyscall(void) { }
306 enum pti_clone_level {
312 pti_clone_pgtable(unsigned long start, unsigned long end,
313 enum pti_clone_level level)
318 * Clone the populated PMDs which cover start to end. These PMD areas
321 for (addr = start; addr < end;) {
322 pte_t *pte, *target_pte;
323 pmd_t *pmd, *target_pmd;
332 pgd = pgd_offset_k(addr);
333 if (WARN_ON(pgd_none(*pgd)))
335 p4d = p4d_offset(pgd, addr);
336 if (WARN_ON(p4d_none(*p4d)))
339 pud = pud_offset(p4d, addr);
340 if (pud_none(*pud)) {
345 pmd = pmd_offset(pud, addr);
346 if (pmd_none(*pmd)) {
351 if (pmd_large(*pmd) || level == PTI_CLONE_PMD) {
352 target_pmd = pti_user_pagetable_walk_pmd(addr);
353 if (WARN_ON(!target_pmd))
357 * Only clone present PMDs. This ensures only setting
358 * _PAGE_GLOBAL on present PMDs. This should only be
359 * called on well-known addresses anyway, so a non-
360 * present PMD would be a surprise.
362 if (WARN_ON(!(pmd_flags(*pmd) & _PAGE_PRESENT)))
366 * Setting 'target_pmd' below creates a mapping in both
367 * the user and kernel page tables. It is effectively
368 * global, so set it as global in both copies. Note:
369 * the X86_FEATURE_PGE check is not _required_ because
370 * the CPU ignores _PAGE_GLOBAL when PGE is not
371 * supported. The check keeps consistentency with
372 * code that only set this bit when supported.
374 if (boot_cpu_has(X86_FEATURE_PGE))
375 *pmd = pmd_set_flags(*pmd, _PAGE_GLOBAL);
378 * Copy the PMD. That is, the kernelmode and usermode
379 * tables will share the last-level page tables of this
386 } else if (level == PTI_CLONE_PTE) {
388 /* Walk the page-table down to the pte level */
389 pte = pte_offset_kernel(pmd, addr);
390 if (pte_none(*pte)) {
395 /* Only clone present PTEs */
396 if (WARN_ON(!(pte_flags(*pte) & _PAGE_PRESENT)))
399 /* Allocate PTE in the user page-table */
400 target_pte = pti_user_pagetable_walk_pte(addr);
401 if (WARN_ON(!target_pte))
404 /* Set GLOBAL bit in both PTEs */
405 if (boot_cpu_has(X86_FEATURE_PGE))
406 *pte = pte_set_flags(*pte, _PAGE_GLOBAL);
421 * Clone a single p4d (i.e. a top-level entry on 4-level systems and a
422 * next-level entry on 5-level systems.
424 static void __init pti_clone_p4d(unsigned long addr)
426 p4d_t *kernel_p4d, *user_p4d;
429 user_p4d = pti_user_pagetable_walk_p4d(addr);
433 kernel_pgd = pgd_offset_k(addr);
434 kernel_p4d = p4d_offset(kernel_pgd, addr);
435 *user_p4d = *kernel_p4d;
439 * Clone the CPU_ENTRY_AREA and associated data into the user space visible
442 static void __init pti_clone_user_shared(void)
446 pti_clone_p4d(CPU_ENTRY_AREA_BASE);
448 for_each_possible_cpu(cpu) {
450 * The SYSCALL64 entry code needs to be able to find the
451 * thread stack and needs one word of scratch space in which
452 * to spill a register. All of this lives in the TSS, in
453 * the sp1 and sp2 slots.
455 * This is done for all possible CPUs during boot to ensure
456 * that it's propagated to all mms. If we were to add one of
457 * these mappings during CPU hotplug, we would need to take
458 * some measure to make sure that every mm that subsequently
459 * ran on that CPU would have the relevant PGD entry in its
460 * pagetables. The usual vmalloc_fault() mechanism would not
461 * work for page faults taken in entry_SYSCALL_64 before RSP
465 unsigned long va = (unsigned long)&per_cpu(cpu_tss_rw, cpu);
466 phys_addr_t pa = per_cpu_ptr_to_phys((void *)va);
469 target_pte = pti_user_pagetable_walk_pte(va);
470 if (WARN_ON(!target_pte))
473 *target_pte = pfn_pte(pa >> PAGE_SHIFT, PAGE_KERNEL);
477 #else /* CONFIG_X86_64 */
480 * On 32 bit PAE systems with 1GB of Kernel address space there is only
481 * one pgd/p4d for the whole kernel. Cloning that would map the whole
482 * address space into the user page-tables, making PTI useless. So clone
483 * the page-table on the PMD level to prevent that.
485 static void __init pti_clone_user_shared(void)
487 unsigned long start, end;
489 start = CPU_ENTRY_AREA_BASE;
490 end = start + (PAGE_SIZE * CPU_ENTRY_AREA_PAGES);
492 pti_clone_pgtable(start, end, PTI_CLONE_PMD);
494 #endif /* CONFIG_X86_64 */
497 * Clone the ESPFIX P4D into the user space visible page table
499 static void __init pti_setup_espfix64(void)
501 #ifdef CONFIG_X86_ESPFIX64
502 pti_clone_p4d(ESPFIX_BASE_ADDR);
507 * Clone the populated PMDs of the entry and irqentry text and force it RO.
509 static void pti_clone_entry_text(void)
511 pti_clone_pgtable((unsigned long) __entry_text_start,
512 (unsigned long) __irqentry_text_end,
517 * Global pages and PCIDs are both ways to make kernel TLB entries
518 * live longer, reduce TLB misses and improve kernel performance.
519 * But, leaving all kernel text Global makes it potentially accessible
520 * to Meltdown-style attacks which make it trivial to find gadgets or
523 * Only use global pages when it is really worth it.
525 static inline bool pti_kernel_image_global_ok(void)
528 * Systems with PCIDs get litlle benefit from global
529 * kernel text and are not worth the downsides.
531 if (cpu_feature_enabled(X86_FEATURE_PCID))
535 * Only do global kernel image for pti=auto. Do the most
536 * secure thing (not global) if pti=on specified.
538 if (pti_mode != PTI_AUTO)
542 * K8 may not tolerate the cleared _PAGE_RW on the userspace
543 * global kernel image pages. Do the safe thing (disable
544 * global kernel image). This is unlikely to ever be
545 * noticed because PTI is disabled by default on AMD CPUs.
547 if (boot_cpu_has(X86_FEATURE_K8))
551 * RANDSTRUCT derives its hardening benefits from the
552 * attacker's lack of knowledge about the layout of kernel
553 * data structures. Keep the kernel image non-global in
554 * cases where RANDSTRUCT is in use to help keep the layout a
557 if (IS_ENABLED(CONFIG_GCC_PLUGIN_RANDSTRUCT))
564 * This is the only user for these and it is not arch-generic
565 * like the other set_memory.h functions. Just extern them.
567 extern int set_memory_nonglobal(unsigned long addr, int numpages);
568 extern int set_memory_global(unsigned long addr, int numpages);
571 * For some configurations, map all of kernel text into the user page
572 * tables. This reduces TLB misses, especially on non-PCID systems.
574 static void pti_clone_kernel_text(void)
577 * rodata is part of the kernel image and is normally
578 * readable on the filesystem or on the web. But, do not
579 * clone the areas past rodata, they might contain secrets.
581 unsigned long start = PFN_ALIGN(_text);
582 unsigned long end_clone = (unsigned long)__end_rodata_aligned;
583 unsigned long end_global = PFN_ALIGN((unsigned long)__stop___ex_table);
585 if (!pti_kernel_image_global_ok())
588 pr_debug("mapping partial kernel image into user address space\n");
591 * Note that this will undo _some_ of the work that
592 * pti_set_kernel_image_nonglobal() did to clear the
595 pti_clone_pgtable(start, end_clone, PTI_LEVEL_KERNEL_IMAGE);
598 * pti_clone_pgtable() will set the global bit in any PMDs
599 * that it clones, but we also need to get any PTEs in
600 * the last level for areas that are not huge-page-aligned.
603 /* Set the global bit for normal non-__init kernel text: */
604 set_memory_global(start, (end_global - start) >> PAGE_SHIFT);
607 static void pti_set_kernel_image_nonglobal(void)
610 * The identity map is created with PMDs, regardless of the
611 * actual length of the kernel. We need to clear
612 * _PAGE_GLOBAL up to a PMD boundary, not just to the end
615 unsigned long start = PFN_ALIGN(_text);
616 unsigned long end = ALIGN((unsigned long)_end, PMD_PAGE_SIZE);
619 * This clears _PAGE_GLOBAL from the entire kernel image.
620 * pti_clone_kernel_text() map put _PAGE_GLOBAL back for
621 * areas that are mapped to userspace.
623 set_memory_nonglobal(start, (end - start) >> PAGE_SHIFT);
627 * Initialize kernel page table isolation
629 void __init pti_init(void)
631 if (!boot_cpu_has(X86_FEATURE_PTI))
634 pr_info("enabled\n");
638 * We check for X86_FEATURE_PCID here. But the init-code will
639 * clear the feature flag on 32 bit because the feature is not
640 * supported on 32 bit anyway. To print the warning we need to
641 * check with cpuid directly again.
643 if (cpuid_ecx(0x1) & BIT(17)) {
644 /* Use printk to work around pr_fmt() */
645 printk(KERN_WARNING "\n");
646 printk(KERN_WARNING "************************************************************\n");
647 printk(KERN_WARNING "** WARNING! WARNING! WARNING! WARNING! WARNING! WARNING! **\n");
648 printk(KERN_WARNING "** **\n");
649 printk(KERN_WARNING "** You are using 32-bit PTI on a 64-bit PCID-capable CPU. **\n");
650 printk(KERN_WARNING "** Your performance will increase dramatically if you **\n");
651 printk(KERN_WARNING "** switch to a 64-bit kernel! **\n");
652 printk(KERN_WARNING "** **\n");
653 printk(KERN_WARNING "** WARNING! WARNING! WARNING! WARNING! WARNING! WARNING! **\n");
654 printk(KERN_WARNING "************************************************************\n");
658 pti_clone_user_shared();
660 /* Undo all global bits from the init pagetables in head_64.S: */
661 pti_set_kernel_image_nonglobal();
662 /* Replace some of the global bits just for shared entry text: */
663 pti_clone_entry_text();
664 pti_setup_espfix64();
665 pti_setup_vsyscall();
669 * Finalize the kernel mappings in the userspace page-table. Some of the
670 * mappings for the kernel image might have changed since pti_init()
671 * cloned them. This is because parts of the kernel image have been
672 * mapped RO and/or NX. These changes need to be cloned again to the
673 * userspace page-table.
675 void pti_finalize(void)
678 * We need to clone everything (again) that maps parts of the
681 pti_clone_entry_text();
682 pti_clone_kernel_text();
684 debug_checkwx_user();