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
57 static void __init pti_print_if_insecure(const char *reason)
59 if (boot_cpu_has_bug(X86_BUG_CPU_MELTDOWN))
60 pr_info("%s\n", reason);
63 static void __init pti_print_if_secure(const char *reason)
65 if (!boot_cpu_has_bug(X86_BUG_CPU_MELTDOWN))
66 pr_info("%s\n", reason);
75 void __init pti_check_boottime_disable(void)
80 /* Assume mode is auto unless overridden. */
83 if (hypervisor_is_type(X86_HYPER_XEN_PV)) {
84 pti_mode = PTI_FORCE_OFF;
85 pti_print_if_insecure("disabled on XEN PV.");
89 ret = cmdline_find_option(boot_command_line, "pti", arg, sizeof(arg));
91 if (ret == 3 && !strncmp(arg, "off", 3)) {
92 pti_mode = PTI_FORCE_OFF;
93 pti_print_if_insecure("disabled on command line.");
96 if (ret == 2 && !strncmp(arg, "on", 2)) {
97 pti_mode = PTI_FORCE_ON;
98 pti_print_if_secure("force enabled on command line.");
101 if (ret == 4 && !strncmp(arg, "auto", 4)) {
107 if (cmdline_find_option_bool(boot_command_line, "nopti")) {
108 pti_mode = PTI_FORCE_OFF;
109 pti_print_if_insecure("disabled on command line.");
114 if (!boot_cpu_has_bug(X86_BUG_CPU_MELTDOWN))
117 setup_force_cpu_cap(X86_FEATURE_PTI);
120 pgd_t __pti_set_user_pgtbl(pgd_t *pgdp, pgd_t pgd)
123 * Changes to the high (kernel) portion of the kernelmode page
124 * tables are not automatically propagated to the usermode tables.
126 * Users should keep in mind that, unlike the kernelmode tables,
127 * there is no vmalloc_fault equivalent for the usermode tables.
128 * Top-level entries added to init_mm's usermode pgd after boot
129 * will not be automatically propagated to other mms.
131 if (!pgdp_maps_userspace(pgdp))
135 * The user page tables get the full PGD, accessible from
138 kernel_to_user_pgdp(pgdp)->pgd = pgd.pgd;
141 * If this is normal user memory, make it NX in the kernel
142 * pagetables so that, if we somehow screw up and return to
143 * usermode with the kernel CR3 loaded, we'll get a page fault
144 * instead of allowing user code to execute with the wrong CR3.
146 * As exceptions, we don't set NX if:
147 * - _PAGE_USER is not set. This could be an executable
148 * EFI runtime mapping or something similar, and the kernel
149 * may execute from it
150 * - we don't have NX support
151 * - we're clearing the PGD (i.e. the new pgd is not present).
153 if ((pgd.pgd & (_PAGE_USER|_PAGE_PRESENT)) == (_PAGE_USER|_PAGE_PRESENT) &&
154 (__supported_pte_mask & _PAGE_NX))
157 /* return the copy of the PGD we want the kernel to use: */
162 * Walk the user copy of the page tables (optionally) trying to allocate
163 * page table pages on the way down.
165 * Returns a pointer to a P4D on success, or NULL on failure.
167 static p4d_t *pti_user_pagetable_walk_p4d(unsigned long address)
169 pgd_t *pgd = kernel_to_user_pgdp(pgd_offset_k(address));
170 gfp_t gfp = (GFP_KERNEL | __GFP_NOTRACK | __GFP_ZERO);
172 if (address < PAGE_OFFSET) {
173 WARN_ONCE(1, "attempt to walk user address\n");
177 if (pgd_none(*pgd)) {
178 unsigned long new_p4d_page = __get_free_page(gfp);
179 if (WARN_ON_ONCE(!new_p4d_page))
182 set_pgd(pgd, __pgd(_KERNPG_TABLE | __pa(new_p4d_page)));
184 BUILD_BUG_ON(pgd_large(*pgd) != 0);
186 return p4d_offset(pgd, address);
190 * Walk the user copy of the page tables (optionally) trying to allocate
191 * page table pages on the way down.
193 * Returns a pointer to a PMD on success, or NULL on failure.
195 static pmd_t *pti_user_pagetable_walk_pmd(unsigned long address)
197 gfp_t gfp = (GFP_KERNEL | __GFP_NOTRACK | __GFP_ZERO);
201 p4d = pti_user_pagetable_walk_p4d(address);
205 BUILD_BUG_ON(p4d_large(*p4d) != 0);
206 if (p4d_none(*p4d)) {
207 unsigned long new_pud_page = __get_free_page(gfp);
208 if (WARN_ON_ONCE(!new_pud_page))
211 set_p4d(p4d, __p4d(_KERNPG_TABLE | __pa(new_pud_page)));
214 pud = pud_offset(p4d, address);
215 /* The user page tables do not use large mappings: */
216 if (pud_large(*pud)) {
220 if (pud_none(*pud)) {
221 unsigned long new_pmd_page = __get_free_page(gfp);
222 if (WARN_ON_ONCE(!new_pmd_page))
225 set_pud(pud, __pud(_KERNPG_TABLE | __pa(new_pmd_page)));
228 return pmd_offset(pud, address);
231 #ifdef CONFIG_X86_VSYSCALL_EMULATION
233 * Walk the shadow copy of the page tables (optionally) trying to allocate
234 * page table pages on the way down. Does not support large pages.
236 * Note: this is only used when mapping *new* kernel data into the
237 * user/shadow page tables. It is never used for userspace data.
239 * Returns a pointer to a PTE on success, or NULL on failure.
241 static __init pte_t *pti_user_pagetable_walk_pte(unsigned long address)
243 gfp_t gfp = (GFP_KERNEL | __GFP_NOTRACK | __GFP_ZERO);
247 pmd = pti_user_pagetable_walk_pmd(address);
251 /* We can't do anything sensible if we hit a large mapping. */
252 if (pmd_large(*pmd)) {
257 if (pmd_none(*pmd)) {
258 unsigned long new_pte_page = __get_free_page(gfp);
262 set_pmd(pmd, __pmd(_KERNPG_TABLE | __pa(new_pte_page)));
265 pte = pte_offset_kernel(pmd, address);
266 if (pte_flags(*pte) & _PAGE_USER) {
267 WARN_ONCE(1, "attempt to walk to user pte\n");
273 static void __init pti_setup_vsyscall(void)
275 pte_t *pte, *target_pte;
278 pte = lookup_address(VSYSCALL_ADDR, &level);
279 if (!pte || WARN_ON(level != PG_LEVEL_4K) || pte_none(*pte))
282 target_pte = pti_user_pagetable_walk_pte(VSYSCALL_ADDR);
283 if (WARN_ON(!target_pte))
287 set_vsyscall_pgtable_user_bits(kernel_to_user_pgdp(swapper_pg_dir));
290 static void __init pti_setup_vsyscall(void) { }
294 pti_clone_pmds(unsigned long start, unsigned long end, pmdval_t clear)
299 * Clone the populated PMDs which cover start to end. These PMD areas
302 for (addr = start; addr < end; addr += PMD_SIZE) {
303 pmd_t *pmd, *target_pmd;
312 pgd = pgd_offset_k(addr);
313 if (WARN_ON(pgd_none(*pgd)))
315 p4d = p4d_offset(pgd, addr);
316 if (WARN_ON(p4d_none(*p4d)))
318 pud = pud_offset(p4d, addr);
321 pmd = pmd_offset(pud, addr);
325 target_pmd = pti_user_pagetable_walk_pmd(addr);
326 if (WARN_ON(!target_pmd))
330 * Only clone present PMDs. This ensures only setting
331 * _PAGE_GLOBAL on present PMDs. This should only be
332 * called on well-known addresses anyway, so a non-
333 * present PMD would be a surprise.
335 if (WARN_ON(!(pmd_flags(*pmd) & _PAGE_PRESENT)))
339 * Setting 'target_pmd' below creates a mapping in both
340 * the user and kernel page tables. It is effectively
341 * global, so set it as global in both copies. Note:
342 * the X86_FEATURE_PGE check is not _required_ because
343 * the CPU ignores _PAGE_GLOBAL when PGE is not
344 * supported. The check keeps consistentency with
345 * code that only set this bit when supported.
347 if (boot_cpu_has(X86_FEATURE_PGE))
348 *pmd = pmd_set_flags(*pmd, _PAGE_GLOBAL);
351 * Copy the PMD. That is, the kernelmode and usermode
352 * tables will share the last-level page tables of this
355 *target_pmd = pmd_clear_flags(*pmd, clear);
361 * Clone a single p4d (i.e. a top-level entry on 4-level systems and a
362 * next-level entry on 5-level systems.
364 static void __init pti_clone_p4d(unsigned long addr)
366 p4d_t *kernel_p4d, *user_p4d;
369 user_p4d = pti_user_pagetable_walk_p4d(addr);
373 kernel_pgd = pgd_offset_k(addr);
374 kernel_p4d = p4d_offset(kernel_pgd, addr);
375 *user_p4d = *kernel_p4d;
379 * Clone the CPU_ENTRY_AREA into the user space visible page table.
381 static void __init pti_clone_user_shared(void)
383 pti_clone_p4d(CPU_ENTRY_AREA_BASE);
386 #else /* CONFIG_X86_64 */
389 * On 32 bit PAE systems with 1GB of Kernel address space there is only
390 * one pgd/p4d for the whole kernel. Cloning that would map the whole
391 * address space into the user page-tables, making PTI useless. So clone
392 * the page-table on the PMD level to prevent that.
394 static void __init pti_clone_user_shared(void)
396 unsigned long start, end;
398 start = CPU_ENTRY_AREA_BASE;
399 end = start + (PAGE_SIZE * CPU_ENTRY_AREA_PAGES);
401 pti_clone_pmds(start, end, 0);
403 #endif /* CONFIG_X86_64 */
406 * Clone the ESPFIX P4D into the user space visible page table
408 static void __init pti_setup_espfix64(void)
410 #ifdef CONFIG_X86_ESPFIX64
411 pti_clone_p4d(ESPFIX_BASE_ADDR);
416 * Clone the populated PMDs of the entry and irqentry text and force it RO.
418 static void pti_clone_entry_text(void)
420 pti_clone_pmds((unsigned long) __entry_text_start,
421 (unsigned long) __irqentry_text_end,
426 * Global pages and PCIDs are both ways to make kernel TLB entries
427 * live longer, reduce TLB misses and improve kernel performance.
428 * But, leaving all kernel text Global makes it potentially accessible
429 * to Meltdown-style attacks which make it trivial to find gadgets or
432 * Only use global pages when it is really worth it.
434 static inline bool pti_kernel_image_global_ok(void)
437 * Systems with PCIDs get litlle benefit from global
438 * kernel text and are not worth the downsides.
440 if (cpu_feature_enabled(X86_FEATURE_PCID))
444 * Only do global kernel image for pti=auto. Do the most
445 * secure thing (not global) if pti=on specified.
447 if (pti_mode != PTI_AUTO)
451 * K8 may not tolerate the cleared _PAGE_RW on the userspace
452 * global kernel image pages. Do the safe thing (disable
453 * global kernel image). This is unlikely to ever be
454 * noticed because PTI is disabled by default on AMD CPUs.
456 if (boot_cpu_has(X86_FEATURE_K8))
460 * RANDSTRUCT derives its hardening benefits from the
461 * attacker's lack of knowledge about the layout of kernel
462 * data structures. Keep the kernel image non-global in
463 * cases where RANDSTRUCT is in use to help keep the layout a
466 if (IS_ENABLED(CONFIG_GCC_PLUGIN_RANDSTRUCT))
473 * For some configurations, map all of kernel text into the user page
474 * tables. This reduces TLB misses, especially on non-PCID systems.
476 static void pti_clone_kernel_text(void)
479 * rodata is part of the kernel image and is normally
480 * readable on the filesystem or on the web. But, do not
481 * clone the areas past rodata, they might contain secrets.
483 unsigned long start = PFN_ALIGN(_text);
484 unsigned long end = (unsigned long)__end_rodata_aligned;
486 if (!pti_kernel_image_global_ok())
489 pr_debug("mapping partial kernel image into user address space\n");
492 * Note that this will undo _some_ of the work that
493 * pti_set_kernel_image_nonglobal() did to clear the
496 pti_clone_pmds(start, end, 0);
500 * This is the only user for it and it is not arch-generic like
501 * the other set_memory.h functions. Just extern it.
503 extern int set_memory_nonglobal(unsigned long addr, int numpages);
504 static void pti_set_kernel_image_nonglobal(void)
507 * The identity map is created with PMDs, regardless of the
508 * actual length of the kernel. We need to clear
509 * _PAGE_GLOBAL up to a PMD boundary, not just to the end
512 unsigned long start = PFN_ALIGN(_text);
513 unsigned long end = ALIGN((unsigned long)_end, PMD_PAGE_SIZE);
515 if (pti_kernel_image_global_ok())
518 set_memory_nonglobal(start, (end - start) >> PAGE_SHIFT);
522 * Initialize kernel page table isolation
524 void __init pti_init(void)
526 if (!static_cpu_has(X86_FEATURE_PTI))
529 pr_info("enabled\n");
533 * We check for X86_FEATURE_PCID here. But the init-code will
534 * clear the feature flag on 32 bit because the feature is not
535 * supported on 32 bit anyway. To print the warning we need to
536 * check with cpuid directly again.
538 if (cpuid_ecx(0x1) && BIT(17)) {
539 /* Use printk to work around pr_fmt() */
540 printk(KERN_WARNING "\n");
541 printk(KERN_WARNING "************************************************************\n");
542 printk(KERN_WARNING "** WARNING! WARNING! WARNING! WARNING! WARNING! WARNING! **\n");
543 printk(KERN_WARNING "** **\n");
544 printk(KERN_WARNING "** You are using 32-bit PTI on a 64-bit PCID-capable CPU. **\n");
545 printk(KERN_WARNING "** Your performance will increase dramatically if you **\n");
546 printk(KERN_WARNING "** switch to a 64-bit kernel! **\n");
547 printk(KERN_WARNING "** **\n");
548 printk(KERN_WARNING "** WARNING! WARNING! WARNING! WARNING! WARNING! WARNING! **\n");
549 printk(KERN_WARNING "************************************************************\n");
553 pti_clone_user_shared();
555 /* Undo all global bits from the init pagetables in head_64.S: */
556 pti_set_kernel_image_nonglobal();
557 /* Replace some of the global bits just for shared entry text: */
558 pti_clone_entry_text();
559 pti_setup_espfix64();
560 pti_setup_vsyscall();
564 * Finalize the kernel mappings in the userspace page-table. Some of the
565 * mappings for the kernel image might have changed since pti_init()
566 * cloned them. This is because parts of the kernel image have been
567 * mapped RO and/or NX. These changes need to be cloned again to the
568 * userspace page-table.
570 void pti_finalize(void)
573 * We need to clone everything (again) that maps parts of the
576 pti_clone_entry_text();
577 pti_clone_kernel_text();