1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * Copyright(c) 2017 Intel Corporation. All rights reserved.
4 *
5 * This code is based in part on work published here:
6 *
7 * https://github.com/IAIK/KAISER
8 *
9 * The original work was written by and and signed off by for the Linux
10 * kernel by:
11 *
12 * Signed-off-by: Richard Fellner <richard.fellner@student.tugraz.at>
13 * Signed-off-by: Moritz Lipp <moritz.lipp@iaik.tugraz.at>
14 * Signed-off-by: Daniel Gruss <daniel.gruss@iaik.tugraz.at>
15 * Signed-off-by: Michael Schwarz <michael.schwarz@iaik.tugraz.at>
16 *
17 * Major changes to the original code by: Dave Hansen <dave.hansen@intel.com>
18 * Mostly rewritten by Thomas Gleixner <tglx@linutronix.de> and
19 * Andy Lutomirsky <luto@amacapital.net>
20 */
21#include <linux/kernel.h>
22#include <linux/errno.h>
23#include <linux/string.h>
24#include <linux/types.h>
25#include <linux/bug.h>
26#include <linux/init.h>
27#include <linux/spinlock.h>
28#include <linux/mm.h>
29#include <linux/uaccess.h>
30#include <linux/cpu.h>
31
32#include <asm/cpufeature.h>
33#include <asm/hypervisor.h>
34#include <asm/vsyscall.h>
35#include <asm/cmdline.h>
36#include <asm/pti.h>
37#include <asm/tlbflush.h>
38#include <asm/desc.h>
39#include <asm/sections.h>
40#include <asm/set_memory.h>
41
42#undef pr_fmt
43#define pr_fmt(fmt) "Kernel/User page tables isolation: " fmt
44
45/* Backporting helper */
46#ifndef __GFP_NOTRACK
47#define __GFP_NOTRACK 0
48#endif
49
50/*
51 * Define the page-table levels we clone for user-space on 32
52 * and 64 bit.
53 */
54#ifdef CONFIG_X86_64
55#define PTI_LEVEL_KERNEL_IMAGE PTI_CLONE_PMD
56#else
57#define PTI_LEVEL_KERNEL_IMAGE PTI_CLONE_PTE
58#endif
59
60static void __init pti_print_if_insecure(const char *reason)
61{
62 if (boot_cpu_has_bug(X86_BUG_CPU_MELTDOWN))
63 pr_info("%s\n", reason);
64}
65
66static void __init pti_print_if_secure(const char *reason)
67{
68 if (!boot_cpu_has_bug(X86_BUG_CPU_MELTDOWN))
69 pr_info("%s\n", reason);
70}
71
72/* Assume mode is auto unless overridden via cmdline below. */
73static enum pti_mode {
74 PTI_AUTO = 0,
75 PTI_FORCE_OFF,
76 PTI_FORCE_ON
77} pti_mode;
78
79void __init pti_check_boottime_disable(void)
80{
81 if (hypervisor_is_type(type: X86_HYPER_XEN_PV)) {
82 pti_mode = PTI_FORCE_OFF;
83 pti_print_if_insecure(reason: "disabled on XEN PV.");
84 return;
85 }
86
87 if (cpu_mitigations_off())
88 pti_mode = PTI_FORCE_OFF;
89 if (pti_mode == PTI_FORCE_OFF) {
90 pti_print_if_insecure(reason: "disabled on command line.");
91 return;
92 }
93
94 if (pti_mode == PTI_FORCE_ON)
95 pti_print_if_secure(reason: "force enabled on command line.");
96
97 if (pti_mode == PTI_AUTO && !boot_cpu_has_bug(X86_BUG_CPU_MELTDOWN))
98 return;
99
100 setup_force_cpu_cap(X86_FEATURE_PTI);
101}
102
103static int __init pti_parse_cmdline(char *arg)
104{
105 if (!strcmp(arg, "off"))
106 pti_mode = PTI_FORCE_OFF;
107 else if (!strcmp(arg, "on"))
108 pti_mode = PTI_FORCE_ON;
109 else if (!strcmp(arg, "auto"))
110 pti_mode = PTI_AUTO;
111 else
112 return -EINVAL;
113 return 0;
114}
115early_param("pti", pti_parse_cmdline);
116
117static int __init pti_parse_cmdline_nopti(char *arg)
118{
119 pti_mode = PTI_FORCE_OFF;
120 return 0;
121}
122early_param("nopti", pti_parse_cmdline_nopti);
123
124pgd_t __pti_set_user_pgtbl(pgd_t *pgdp, pgd_t pgd)
125{
126 /*
127 * Changes to the high (kernel) portion of the kernelmode page
128 * tables are not automatically propagated to the usermode tables.
129 *
130 * Users should keep in mind that, unlike the kernelmode tables,
131 * there is no vmalloc_fault equivalent for the usermode tables.
132 * Top-level entries added to init_mm's usermode pgd after boot
133 * will not be automatically propagated to other mms.
134 */
135 if (!pgdp_maps_userspace(ptr: pgdp))
136 return pgd;
137
138 /*
139 * The user page tables get the full PGD, accessible from
140 * userspace:
141 */
142 kernel_to_user_pgdp(pgdp)->pgd = pgd.pgd;
143
144 /*
145 * If this is normal user memory, make it NX in the kernel
146 * pagetables so that, if we somehow screw up and return to
147 * usermode with the kernel CR3 loaded, we'll get a page fault
148 * instead of allowing user code to execute with the wrong CR3.
149 *
150 * As exceptions, we don't set NX if:
151 * - _PAGE_USER is not set. This could be an executable
152 * EFI runtime mapping or something similar, and the kernel
153 * may execute from it
154 * - we don't have NX support
155 * - we're clearing the PGD (i.e. the new pgd is not present).
156 */
157 if ((pgd.pgd & (_PAGE_USER|_PAGE_PRESENT)) == (_PAGE_USER|_PAGE_PRESENT) &&
158 (__supported_pte_mask & _PAGE_NX))
159 pgd.pgd |= _PAGE_NX;
160
161 /* return the copy of the PGD we want the kernel to use: */
162 return pgd;
163}
164
165/*
166 * Walk the user copy of the page tables (optionally) trying to allocate
167 * page table pages on the way down.
168 *
169 * Returns a pointer to a P4D on success, or NULL on failure.
170 */
171static p4d_t *pti_user_pagetable_walk_p4d(unsigned long address)
172{
173 pgd_t *pgd = kernel_to_user_pgdp(pgd_offset_k(address));
174 gfp_t gfp = (GFP_KERNEL | __GFP_NOTRACK | __GFP_ZERO);
175
176 if (address < PAGE_OFFSET) {
177 WARN_ONCE(1, "attempt to walk user address\n");
178 return NULL;
179 }
180
181 if (pgd_none(pgd: *pgd)) {
182 unsigned long new_p4d_page = __get_free_page(gfp);
183 if (WARN_ON_ONCE(!new_p4d_page))
184 return NULL;
185
186 set_pgd(pgd, __pgd(_KERNPG_TABLE | __pa(new_p4d_page)));
187 }
188 BUILD_BUG_ON(pgd_large(*pgd) != 0);
189
190 return p4d_offset(pgd, address);
191}
192
193/*
194 * Walk the user copy of the page tables (optionally) trying to allocate
195 * page table pages on the way down.
196 *
197 * Returns a pointer to a PMD on success, or NULL on failure.
198 */
199static pmd_t *pti_user_pagetable_walk_pmd(unsigned long address)
200{
201 gfp_t gfp = (GFP_KERNEL | __GFP_NOTRACK | __GFP_ZERO);
202 p4d_t *p4d;
203 pud_t *pud;
204
205 p4d = pti_user_pagetable_walk_p4d(address);
206 if (!p4d)
207 return NULL;
208
209 BUILD_BUG_ON(p4d_large(*p4d) != 0);
210 if (p4d_none(p4d: *p4d)) {
211 unsigned long new_pud_page = __get_free_page(gfp);
212 if (WARN_ON_ONCE(!new_pud_page))
213 return NULL;
214
215 set_p4d(p4dp: p4d, p4d: __p4d(_KERNPG_TABLE | __pa(new_pud_page)));
216 }
217
218 pud = pud_offset(p4d, address);
219 /* The user page tables do not use large mappings: */
220 if (pud_large(pud: *pud)) {
221 WARN_ON(1);
222 return NULL;
223 }
224 if (pud_none(pud: *pud)) {
225 unsigned long new_pmd_page = __get_free_page(gfp);
226 if (WARN_ON_ONCE(!new_pmd_page))
227 return NULL;
228
229 set_pud(pudp: pud, pud: __pud(_KERNPG_TABLE | __pa(new_pmd_page)));
230 }
231
232 return pmd_offset(pud, address);
233}
234
235/*
236 * Walk the shadow copy of the page tables (optionally) trying to allocate
237 * page table pages on the way down. Does not support large pages.
238 *
239 * Note: this is only used when mapping *new* kernel data into the
240 * user/shadow page tables. It is never used for userspace data.
241 *
242 * Returns a pointer to a PTE on success, or NULL on failure.
243 */
244static pte_t *pti_user_pagetable_walk_pte(unsigned long address)
245{
246 gfp_t gfp = (GFP_KERNEL | __GFP_NOTRACK | __GFP_ZERO);
247 pmd_t *pmd;
248 pte_t *pte;
249
250 pmd = pti_user_pagetable_walk_pmd(address);
251 if (!pmd)
252 return NULL;
253
254 /* We can't do anything sensible if we hit a large mapping. */
255 if (pmd_large(pte: *pmd)) {
256 WARN_ON(1);
257 return NULL;
258 }
259
260 if (pmd_none(pmd: *pmd)) {
261 unsigned long new_pte_page = __get_free_page(gfp);
262 if (!new_pte_page)
263 return NULL;
264
265 set_pmd(pmdp: pmd, pmd: __pmd(_KERNPG_TABLE | __pa(new_pte_page)));
266 }
267
268 pte = pte_offset_kernel(pmd, address);
269 if (pte_flags(pte: *pte) & _PAGE_USER) {
270 WARN_ONCE(1, "attempt to walk to user pte\n");
271 return NULL;
272 }
273 return pte;
274}
275
276#ifdef CONFIG_X86_VSYSCALL_EMULATION
277static void __init pti_setup_vsyscall(void)
278{
279 pte_t *pte, *target_pte;
280 unsigned int level;
281
282 pte = lookup_address(VSYSCALL_ADDR, level: &level);
283 if (!pte || WARN_ON(level != PG_LEVEL_4K) || pte_none(pte: *pte))
284 return;
285
286 target_pte = pti_user_pagetable_walk_pte(VSYSCALL_ADDR);
287 if (WARN_ON(!target_pte))
288 return;
289
290 *target_pte = *pte;
291 set_vsyscall_pgtable_user_bits(kernel_to_user_pgdp(swapper_pg_dir));
292}
293#else
294static void __init pti_setup_vsyscall(void) { }
295#endif
296
297enum pti_clone_level {
298 PTI_CLONE_PMD,
299 PTI_CLONE_PTE,
300};
301
302static void
303pti_clone_pgtable(unsigned long start, unsigned long end,
304 enum pti_clone_level level)
305{
306 unsigned long addr;
307
308 /*
309 * Clone the populated PMDs which cover start to end. These PMD areas
310 * can have holes.
311 */
312 for (addr = start; addr < end;) {
313 pte_t *pte, *target_pte;
314 pmd_t *pmd, *target_pmd;
315 pgd_t *pgd;
316 p4d_t *p4d;
317 pud_t *pud;
318
319 /* Overflow check */
320 if (addr < start)
321 break;
322
323 pgd = pgd_offset_k(addr);
324 if (WARN_ON(pgd_none(*pgd)))
325 return;
326 p4d = p4d_offset(pgd, address: addr);
327 if (WARN_ON(p4d_none(*p4d)))
328 return;
329
330 pud = pud_offset(p4d, address: addr);
331 if (pud_none(pud: *pud)) {
332 WARN_ON_ONCE(addr & ~PUD_MASK);
333 addr = round_up(addr + 1, PUD_SIZE);
334 continue;
335 }
336
337 pmd = pmd_offset(pud, address: addr);
338 if (pmd_none(pmd: *pmd)) {
339 WARN_ON_ONCE(addr & ~PMD_MASK);
340 addr = round_up(addr + 1, PMD_SIZE);
341 continue;
342 }
343
344 if (pmd_large(pte: *pmd) || level == PTI_CLONE_PMD) {
345 target_pmd = pti_user_pagetable_walk_pmd(address: addr);
346 if (WARN_ON(!target_pmd))
347 return;
348
349 /*
350 * Only clone present PMDs. This ensures only setting
351 * _PAGE_GLOBAL on present PMDs. This should only be
352 * called on well-known addresses anyway, so a non-
353 * present PMD would be a surprise.
354 */
355 if (WARN_ON(!(pmd_flags(*pmd) & _PAGE_PRESENT)))
356 return;
357
358 /*
359 * Setting 'target_pmd' below creates a mapping in both
360 * the user and kernel page tables. It is effectively
361 * global, so set it as global in both copies. Note:
362 * the X86_FEATURE_PGE check is not _required_ because
363 * the CPU ignores _PAGE_GLOBAL when PGE is not
364 * supported. The check keeps consistency with
365 * code that only set this bit when supported.
366 */
367 if (boot_cpu_has(X86_FEATURE_PGE))
368 *pmd = pmd_set_flags(pmd: *pmd, _PAGE_GLOBAL);
369
370 /*
371 * Copy the PMD. That is, the kernelmode and usermode
372 * tables will share the last-level page tables of this
373 * address range
374 */
375 *target_pmd = *pmd;
376
377 addr += PMD_SIZE;
378
379 } else if (level == PTI_CLONE_PTE) {
380
381 /* Walk the page-table down to the pte level */
382 pte = pte_offset_kernel(pmd, address: addr);
383 if (pte_none(pte: *pte)) {
384 addr += PAGE_SIZE;
385 continue;
386 }
387
388 /* Only clone present PTEs */
389 if (WARN_ON(!(pte_flags(*pte) & _PAGE_PRESENT)))
390 return;
391
392 /* Allocate PTE in the user page-table */
393 target_pte = pti_user_pagetable_walk_pte(address: addr);
394 if (WARN_ON(!target_pte))
395 return;
396
397 /* Set GLOBAL bit in both PTEs */
398 if (boot_cpu_has(X86_FEATURE_PGE))
399 *pte = pte_set_flags(pte: *pte, _PAGE_GLOBAL);
400
401 /* Clone the PTE */
402 *target_pte = *pte;
403
404 addr += PAGE_SIZE;
405
406 } else {
407 BUG();
408 }
409 }
410}
411
412#ifdef CONFIG_X86_64
413/*
414 * Clone a single p4d (i.e. a top-level entry on 4-level systems and a
415 * next-level entry on 5-level systems.
416 */
417static void __init pti_clone_p4d(unsigned long addr)
418{
419 p4d_t *kernel_p4d, *user_p4d;
420 pgd_t *kernel_pgd;
421
422 user_p4d = pti_user_pagetable_walk_p4d(address: addr);
423 if (!user_p4d)
424 return;
425
426 kernel_pgd = pgd_offset_k(addr);
427 kernel_p4d = p4d_offset(pgd: kernel_pgd, address: addr);
428 *user_p4d = *kernel_p4d;
429}
430
431/*
432 * Clone the CPU_ENTRY_AREA and associated data into the user space visible
433 * page table.
434 */
435static void __init pti_clone_user_shared(void)
436{
437 unsigned int cpu;
438
439 pti_clone_p4d(CPU_ENTRY_AREA_BASE);
440
441 for_each_possible_cpu(cpu) {
442 /*
443 * The SYSCALL64 entry code needs one word of scratch space
444 * in which to spill a register. It lives in the sp2 slot
445 * of the CPU's TSS.
446 *
447 * This is done for all possible CPUs during boot to ensure
448 * that it's propagated to all mms.
449 */
450
451 unsigned long va = (unsigned long)&per_cpu(cpu_tss_rw, cpu);
452 phys_addr_t pa = per_cpu_ptr_to_phys(addr: (void *)va);
453 pte_t *target_pte;
454
455 target_pte = pti_user_pagetable_walk_pte(address: va);
456 if (WARN_ON(!target_pte))
457 return;
458
459 *target_pte = pfn_pte(page_nr: pa >> PAGE_SHIFT, PAGE_KERNEL);
460 }
461}
462
463#else /* CONFIG_X86_64 */
464
465/*
466 * On 32 bit PAE systems with 1GB of Kernel address space there is only
467 * one pgd/p4d for the whole kernel. Cloning that would map the whole
468 * address space into the user page-tables, making PTI useless. So clone
469 * the page-table on the PMD level to prevent that.
470 */
471static void __init pti_clone_user_shared(void)
472{
473 unsigned long start, end;
474
475 start = CPU_ENTRY_AREA_BASE;
476 end = start + (PAGE_SIZE * CPU_ENTRY_AREA_PAGES);
477
478 pti_clone_pgtable(start, end, PTI_CLONE_PMD);
479}
480#endif /* CONFIG_X86_64 */
481
482/*
483 * Clone the ESPFIX P4D into the user space visible page table
484 */
485static void __init pti_setup_espfix64(void)
486{
487#ifdef CONFIG_X86_ESPFIX64
488 pti_clone_p4d(ESPFIX_BASE_ADDR);
489#endif
490}
491
492/*
493 * Clone the populated PMDs of the entry text and force it RO.
494 */
495static void pti_clone_entry_text(void)
496{
497 pti_clone_pgtable(start: (unsigned long) __entry_text_start,
498 end: (unsigned long) __entry_text_end,
499 level: PTI_CLONE_PMD);
500}
501
502/*
503 * Global pages and PCIDs are both ways to make kernel TLB entries
504 * live longer, reduce TLB misses and improve kernel performance.
505 * But, leaving all kernel text Global makes it potentially accessible
506 * to Meltdown-style attacks which make it trivial to find gadgets or
507 * defeat KASLR.
508 *
509 * Only use global pages when it is really worth it.
510 */
511static inline bool pti_kernel_image_global_ok(void)
512{
513 /*
514 * Systems with PCIDs get little benefit from global
515 * kernel text and are not worth the downsides.
516 */
517 if (cpu_feature_enabled(X86_FEATURE_PCID))
518 return false;
519
520 /*
521 * Only do global kernel image for pti=auto. Do the most
522 * secure thing (not global) if pti=on specified.
523 */
524 if (pti_mode != PTI_AUTO)
525 return false;
526
527 /*
528 * K8 may not tolerate the cleared _PAGE_RW on the userspace
529 * global kernel image pages. Do the safe thing (disable
530 * global kernel image). This is unlikely to ever be
531 * noticed because PTI is disabled by default on AMD CPUs.
532 */
533 if (boot_cpu_has(X86_FEATURE_K8))
534 return false;
535
536 /*
537 * RANDSTRUCT derives its hardening benefits from the
538 * attacker's lack of knowledge about the layout of kernel
539 * data structures. Keep the kernel image non-global in
540 * cases where RANDSTRUCT is in use to help keep the layout a
541 * secret.
542 */
543 if (IS_ENABLED(CONFIG_RANDSTRUCT))
544 return false;
545
546 return true;
547}
548
549/*
550 * For some configurations, map all of kernel text into the user page
551 * tables. This reduces TLB misses, especially on non-PCID systems.
552 */
553static void pti_clone_kernel_text(void)
554{
555 /*
556 * rodata is part of the kernel image and is normally
557 * readable on the filesystem or on the web. But, do not
558 * clone the areas past rodata, they might contain secrets.
559 */
560 unsigned long start = PFN_ALIGN(_text);
561 unsigned long end_clone = (unsigned long)__end_rodata_aligned;
562 unsigned long end_global = PFN_ALIGN((unsigned long)_etext);
563
564 if (!pti_kernel_image_global_ok())
565 return;
566
567 pr_debug("mapping partial kernel image into user address space\n");
568
569 /*
570 * Note that this will undo _some_ of the work that
571 * pti_set_kernel_image_nonglobal() did to clear the
572 * global bit.
573 */
574 pti_clone_pgtable(start, end: end_clone, PTI_LEVEL_KERNEL_IMAGE);
575
576 /*
577 * pti_clone_pgtable() will set the global bit in any PMDs
578 * that it clones, but we also need to get any PTEs in
579 * the last level for areas that are not huge-page-aligned.
580 */
581
582 /* Set the global bit for normal non-__init kernel text: */
583 set_memory_global(addr: start, numpages: (end_global - start) >> PAGE_SHIFT);
584}
585
586static void pti_set_kernel_image_nonglobal(void)
587{
588 /*
589 * The identity map is created with PMDs, regardless of the
590 * actual length of the kernel. We need to clear
591 * _PAGE_GLOBAL up to a PMD boundary, not just to the end
592 * of the image.
593 */
594 unsigned long start = PFN_ALIGN(_text);
595 unsigned long end = ALIGN((unsigned long)_end, PMD_SIZE);
596
597 /*
598 * This clears _PAGE_GLOBAL from the entire kernel image.
599 * pti_clone_kernel_text() map put _PAGE_GLOBAL back for
600 * areas that are mapped to userspace.
601 */
602 set_memory_nonglobal(addr: start, numpages: (end - start) >> PAGE_SHIFT);
603}
604
605/*
606 * Initialize kernel page table isolation
607 */
608void __init pti_init(void)
609{
610 if (!boot_cpu_has(X86_FEATURE_PTI))
611 return;
612
613 pr_info("enabled\n");
614
615#ifdef CONFIG_X86_32
616 /*
617 * We check for X86_FEATURE_PCID here. But the init-code will
618 * clear the feature flag on 32 bit because the feature is not
619 * supported on 32 bit anyway. To print the warning we need to
620 * check with cpuid directly again.
621 */
622 if (cpuid_ecx(0x1) & BIT(17)) {
623 /* Use printk to work around pr_fmt() */
624 printk(KERN_WARNING "\n");
625 printk(KERN_WARNING "************************************************************\n");
626 printk(KERN_WARNING "** WARNING! WARNING! WARNING! WARNING! WARNING! WARNING! **\n");
627 printk(KERN_WARNING "** **\n");
628 printk(KERN_WARNING "** You are using 32-bit PTI on a 64-bit PCID-capable CPU. **\n");
629 printk(KERN_WARNING "** Your performance will increase dramatically if you **\n");
630 printk(KERN_WARNING "** switch to a 64-bit kernel! **\n");
631 printk(KERN_WARNING "** **\n");
632 printk(KERN_WARNING "** WARNING! WARNING! WARNING! WARNING! WARNING! WARNING! **\n");
633 printk(KERN_WARNING "************************************************************\n");
634 }
635#endif
636
637 pti_clone_user_shared();
638
639 /* Undo all global bits from the init pagetables in head_64.S: */
640 pti_set_kernel_image_nonglobal();
641 /* Replace some of the global bits just for shared entry text: */
642 pti_clone_entry_text();
643 pti_setup_espfix64();
644 pti_setup_vsyscall();
645}
646
647/*
648 * Finalize the kernel mappings in the userspace page-table. Some of the
649 * mappings for the kernel image might have changed since pti_init()
650 * cloned them. This is because parts of the kernel image have been
651 * mapped RO and/or NX. These changes need to be cloned again to the
652 * userspace page-table.
653 */
654void pti_finalize(void)
655{
656 if (!boot_cpu_has(X86_FEATURE_PTI))
657 return;
658 /*
659 * We need to clone everything (again) that maps parts of the
660 * kernel image.
661 */
662 pti_clone_entry_text();
663 pti_clone_kernel_text();
664
665 debug_checkwx_user();
666}
667

source code of linux/arch/x86/mm/pti.c