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 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#include <asm/bugs.h>
42
43#undef pr_fmt
44#define pr_fmt(fmt) "Kernel/User page tables isolation: " fmt
45
46/* Backporting helper */
47#ifndef __GFP_NOTRACK
48#define __GFP_NOTRACK 0
49#endif
50
51/*
52 * Define the page-table levels we clone for user-space on 32
53 * and 64 bit.
54 */
55#ifdef CONFIG_X86_64
56#define PTI_LEVEL_KERNEL_IMAGE PTI_CLONE_PMD
57#else
58#define PTI_LEVEL_KERNEL_IMAGE PTI_CLONE_PTE
59#endif
60
61static void __init pti_print_if_insecure(const char *reason)
62{
63 if (boot_cpu_has_bug(X86_BUG_CPU_MELTDOWN))
64 pr_info("%s\n", reason);
65}
66
67static void __init pti_print_if_secure(const char *reason)
68{
69 if (!boot_cpu_has_bug(X86_BUG_CPU_MELTDOWN))
70 pr_info("%s\n", reason);
71}
72
73/* Assume mode is auto unless overridden via cmdline below. */
74static enum pti_mode {
75 PTI_AUTO = 0,
76 PTI_FORCE_OFF,
77 PTI_FORCE_ON
78} pti_mode;
79
80void __init pti_check_boottime_disable(void)
81{
82 if (hypervisor_is_type(type: X86_HYPER_XEN_PV)) {
83 pti_mode = PTI_FORCE_OFF;
84 pti_print_if_insecure(reason: "disabled on XEN PV.");
85 return;
86 }
87
88 if (pti_mode == PTI_AUTO &&
89 !cpu_attack_vector_mitigated(v: CPU_MITIGATE_USER_KERNEL))
90 pti_mode = PTI_FORCE_OFF;
91 if (pti_mode == PTI_FORCE_OFF) {
92 pti_print_if_insecure(reason: "disabled on command line.");
93 return;
94 }
95
96 if (pti_mode == PTI_FORCE_ON)
97 pti_print_if_secure(reason: "force enabled on command line.");
98
99 if (pti_mode == PTI_AUTO && !boot_cpu_has_bug(X86_BUG_CPU_MELTDOWN))
100 return;
101
102 setup_force_cpu_cap(X86_FEATURE_PTI);
103
104 if (cpu_feature_enabled(X86_FEATURE_INVLPGB)) {
105 pr_debug("PTI enabled, disabling INVLPGB\n");
106 setup_clear_cpu_cap(X86_FEATURE_INVLPGB);
107 }
108}
109
110static int __init pti_parse_cmdline(char *arg)
111{
112 if (!strcmp(arg, "off"))
113 pti_mode = PTI_FORCE_OFF;
114 else if (!strcmp(arg, "on"))
115 pti_mode = PTI_FORCE_ON;
116 else if (!strcmp(arg, "auto"))
117 pti_mode = PTI_AUTO;
118 else
119 return -EINVAL;
120 return 0;
121}
122early_param("pti", pti_parse_cmdline);
123
124static int __init pti_parse_cmdline_nopti(char *arg)
125{
126 pti_mode = PTI_FORCE_OFF;
127 return 0;
128}
129early_param("nopti", pti_parse_cmdline_nopti);
130
131pgd_t __pti_set_user_pgtbl(pgd_t *pgdp, pgd_t pgd)
132{
133 /*
134 * Changes to the high (kernel) portion of the kernelmode page
135 * tables are not automatically propagated to the usermode tables.
136 *
137 * Users should keep in mind that, unlike the kernelmode tables,
138 * there is no vmalloc_fault equivalent for the usermode tables.
139 * Top-level entries added to init_mm's usermode pgd after boot
140 * will not be automatically propagated to other mms.
141 */
142 if (!pgdp_maps_userspace(ptr: pgdp) || (pgd.pgd & _PAGE_NOPTISHADOW))
143 return pgd;
144
145 /*
146 * The user page tables get the full PGD, accessible from
147 * userspace:
148 */
149 kernel_to_user_pgdp(pgdp)->pgd = pgd.pgd;
150
151 /*
152 * If this is normal user memory, make it NX in the kernel
153 * pagetables so that, if we somehow screw up and return to
154 * usermode with the kernel CR3 loaded, we'll get a page fault
155 * instead of allowing user code to execute with the wrong CR3.
156 *
157 * As exceptions, we don't set NX if:
158 * - _PAGE_USER is not set. This could be an executable
159 * EFI runtime mapping or something similar, and the kernel
160 * may execute from it
161 * - we don't have NX support
162 * - we're clearing the PGD (i.e. the new pgd is not present).
163 */
164 if ((pgd.pgd & (_PAGE_USER|_PAGE_PRESENT)) == (_PAGE_USER|_PAGE_PRESENT) &&
165 (__supported_pte_mask & _PAGE_NX))
166 pgd.pgd |= _PAGE_NX;
167
168 /* return the copy of the PGD we want the kernel to use: */
169 return pgd;
170}
171
172/*
173 * Walk the user copy of the page tables (optionally) trying to allocate
174 * page table pages on the way down.
175 *
176 * Returns a pointer to a P4D on success, or NULL on failure.
177 */
178static p4d_t *pti_user_pagetable_walk_p4d(unsigned long address)
179{
180 pgd_t *pgd = kernel_to_user_pgdp(pgd_offset_k(address));
181 gfp_t gfp = (GFP_KERNEL | __GFP_NOTRACK | __GFP_ZERO);
182
183 if (address < PAGE_OFFSET) {
184 WARN_ONCE(1, "attempt to walk user address\n");
185 return NULL;
186 }
187
188 if (pgd_none(pgd: *pgd)) {
189 unsigned long new_p4d_page = __get_free_page(gfp);
190 if (WARN_ON_ONCE(!new_p4d_page))
191 return NULL;
192
193 set_pgd(pgd, __pgd(_KERNPG_TABLE | __pa(new_p4d_page)));
194 }
195 BUILD_BUG_ON(pgd_leaf(*pgd));
196
197 return p4d_offset(pgd, address);
198}
199
200/*
201 * Walk the user copy of the page tables (optionally) trying to allocate
202 * page table pages on the way down.
203 *
204 * Returns a pointer to a PMD on success, or NULL on failure.
205 */
206static pmd_t *pti_user_pagetable_walk_pmd(unsigned long address)
207{
208 gfp_t gfp = (GFP_KERNEL | __GFP_NOTRACK | __GFP_ZERO);
209 p4d_t *p4d;
210 pud_t *pud;
211
212 p4d = pti_user_pagetable_walk_p4d(address);
213 if (!p4d)
214 return NULL;
215
216 BUILD_BUG_ON(p4d_leaf(*p4d));
217 if (p4d_none(p4d: *p4d)) {
218 unsigned long new_pud_page = __get_free_page(gfp);
219 if (WARN_ON_ONCE(!new_pud_page))
220 return NULL;
221
222 set_p4d(p4d, __p4d(_KERNPG_TABLE | __pa(new_pud_page)));
223 }
224
225 pud = pud_offset(p4d, address);
226 /* The user page tables do not use large mappings: */
227 if (pud_leaf(pud: *pud)) {
228 WARN_ON(1);
229 return NULL;
230 }
231 if (pud_none(pud: *pud)) {
232 unsigned long new_pmd_page = __get_free_page(gfp);
233 if (WARN_ON_ONCE(!new_pmd_page))
234 return NULL;
235
236 set_pud(pud, __pud(_KERNPG_TABLE | __pa(new_pmd_page)));
237 }
238
239 return pmd_offset(pud, address);
240}
241
242/*
243 * Walk the shadow copy of the page tables (optionally) trying to allocate
244 * page table pages on the way down. Does not support large pages.
245 *
246 * Note: this is only used when mapping *new* kernel data into the
247 * user/shadow page tables. It is never used for userspace data.
248 *
249 * Returns a pointer to a PTE on success, or NULL on failure.
250 */
251static pte_t *pti_user_pagetable_walk_pte(unsigned long address, bool late_text)
252{
253 gfp_t gfp = (GFP_KERNEL | __GFP_NOTRACK | __GFP_ZERO);
254 pmd_t *pmd;
255 pte_t *pte;
256
257 pmd = pti_user_pagetable_walk_pmd(address);
258 if (!pmd)
259 return NULL;
260
261 /* Large PMD mapping found */
262 if (pmd_leaf(pte: *pmd)) {
263 /* Clear the PMD if we hit a large mapping from the first round */
264 if (late_text) {
265 set_pmd(pmd, __pmd(0));
266 } else {
267 WARN_ON_ONCE(1);
268 return NULL;
269 }
270 }
271
272 if (pmd_none(pmd: *pmd)) {
273 unsigned long new_pte_page = __get_free_page(gfp);
274 if (!new_pte_page)
275 return NULL;
276
277 set_pmd(pmd, __pmd(_KERNPG_TABLE | __pa(new_pte_page)));
278 }
279
280 pte = pte_offset_kernel(pmd, address);
281 if (pte_flags(pte: *pte) & _PAGE_USER) {
282 WARN_ONCE(1, "attempt to walk to user pte\n");
283 return NULL;
284 }
285 return pte;
286}
287
288#ifdef CONFIG_X86_VSYSCALL_EMULATION
289static void __init pti_setup_vsyscall(void)
290{
291 pte_t *pte, *target_pte;
292 unsigned int level;
293
294 pte = lookup_address(VSYSCALL_ADDR, level: &level);
295 if (!pte || WARN_ON(level != PG_LEVEL_4K) || pte_none(pte: *pte))
296 return;
297
298 target_pte = pti_user_pagetable_walk_pte(VSYSCALL_ADDR, late_text: false);
299 if (WARN_ON(!target_pte))
300 return;
301
302 *target_pte = *pte;
303 set_vsyscall_pgtable_user_bits(kernel_to_user_pgdp(swapper_pg_dir));
304}
305#else
306static void __init pti_setup_vsyscall(void) { }
307#endif
308
309enum pti_clone_level {
310 PTI_CLONE_PMD,
311 PTI_CLONE_PTE,
312};
313
314static void
315pti_clone_pgtable(unsigned long start, unsigned long end,
316 enum pti_clone_level level, bool late_text)
317{
318 unsigned long addr;
319
320 /*
321 * Clone the populated PMDs which cover start to end. These PMD areas
322 * can have holes.
323 */
324 for (addr = start; addr < end;) {
325 pte_t *pte, *target_pte;
326 pmd_t *pmd, *target_pmd;
327 pgd_t *pgd;
328 p4d_t *p4d;
329 pud_t *pud;
330
331 /* Overflow check */
332 if (addr < start)
333 break;
334
335 pgd = pgd_offset_k(addr);
336 if (WARN_ON(pgd_none(*pgd)))
337 return;
338 p4d = p4d_offset(pgd, address: addr);
339 if (WARN_ON(p4d_none(*p4d)))
340 return;
341
342 pud = pud_offset(p4d, address: addr);
343 if (pud_none(pud: *pud)) {
344 WARN_ON_ONCE(addr & ~PUD_MASK);
345 addr = round_up(addr + 1, PUD_SIZE);
346 continue;
347 }
348
349 pmd = pmd_offset(pud, address: addr);
350 if (pmd_none(pmd: *pmd)) {
351 WARN_ON_ONCE(addr & ~PMD_MASK);
352 addr = round_up(addr + 1, PMD_SIZE);
353 continue;
354 }
355
356 if (pmd_leaf(pte: *pmd) || level == PTI_CLONE_PMD) {
357 target_pmd = pti_user_pagetable_walk_pmd(address: addr);
358 if (WARN_ON(!target_pmd))
359 return;
360
361 /*
362 * Only clone present PMDs. This ensures only setting
363 * _PAGE_GLOBAL on present PMDs. This should only be
364 * called on well-known addresses anyway, so a non-
365 * present PMD would be a surprise.
366 */
367 if (WARN_ON(!(pmd_flags(*pmd) & _PAGE_PRESENT)))
368 return;
369
370 /*
371 * Setting 'target_pmd' below creates a mapping in both
372 * the user and kernel page tables. It is effectively
373 * global, so set it as global in both copies. Note:
374 * the X86_FEATURE_PGE check is not _required_ because
375 * the CPU ignores _PAGE_GLOBAL when PGE is not
376 * supported. The check keeps consistency with
377 * code that only set this bit when supported.
378 */
379 if (boot_cpu_has(X86_FEATURE_PGE))
380 *pmd = pmd_set_flags(pmd: *pmd, _PAGE_GLOBAL);
381
382 /*
383 * Copy the PMD. That is, the kernelmode and usermode
384 * tables will share the last-level page tables of this
385 * address range
386 */
387 *target_pmd = *pmd;
388
389 addr = round_up(addr + 1, PMD_SIZE);
390
391 } else if (level == PTI_CLONE_PTE) {
392
393 /* Walk the page-table down to the pte level */
394 pte = pte_offset_kernel(pmd, address: addr);
395 if (pte_none(pte: *pte)) {
396 addr = round_up(addr + 1, PAGE_SIZE);
397 continue;
398 }
399
400 /* Only clone present PTEs */
401 if (WARN_ON(!(pte_flags(*pte) & _PAGE_PRESENT)))
402 return;
403
404 /* Allocate PTE in the user page-table */
405 target_pte = pti_user_pagetable_walk_pte(address: addr, late_text);
406 if (WARN_ON(!target_pte))
407 return;
408
409 /* Set GLOBAL bit in both PTEs */
410 if (boot_cpu_has(X86_FEATURE_PGE))
411 *pte = pte_set_flags(pte: *pte, _PAGE_GLOBAL);
412
413 /* Clone the PTE */
414 *target_pte = *pte;
415
416 addr = round_up(addr + 1, PAGE_SIZE);
417
418 } else {
419 BUG();
420 }
421 }
422}
423
424#ifdef CONFIG_X86_64
425/*
426 * Clone a single p4d (i.e. a top-level entry on 4-level systems and a
427 * next-level entry on 5-level systems.
428 */
429static void __init pti_clone_p4d(unsigned long addr)
430{
431 p4d_t *kernel_p4d, *user_p4d;
432 pgd_t *kernel_pgd;
433
434 user_p4d = pti_user_pagetable_walk_p4d(address: addr);
435 if (!user_p4d)
436 return;
437
438 kernel_pgd = pgd_offset_k(addr);
439 kernel_p4d = p4d_offset(pgd: kernel_pgd, address: addr);
440 *user_p4d = *kernel_p4d;
441}
442
443/*
444 * Clone the CPU_ENTRY_AREA and associated data into the user space visible
445 * page table.
446 */
447static void __init pti_clone_user_shared(void)
448{
449 unsigned int cpu;
450
451 pti_clone_p4d(CPU_ENTRY_AREA_BASE);
452
453 for_each_possible_cpu(cpu) {
454 /*
455 * The SYSCALL64 entry code needs one word of scratch space
456 * in which to spill a register. It lives in the sp2 slot
457 * of the CPU's TSS.
458 *
459 * This is done for all possible CPUs during boot to ensure
460 * that it's propagated to all mms.
461 */
462
463 unsigned long va = (unsigned long)&per_cpu(cpu_tss_rw, cpu);
464 phys_addr_t pa = per_cpu_ptr_to_phys(addr: (void *)va);
465 pte_t *target_pte;
466
467 target_pte = pti_user_pagetable_walk_pte(address: va, late_text: false);
468 if (WARN_ON(!target_pte))
469 return;
470
471 *target_pte = pfn_pte(page_nr: pa >> PAGE_SHIFT, PAGE_KERNEL);
472 }
473}
474
475#else /* CONFIG_X86_64 */
476
477/*
478 * On 32 bit PAE systems with 1GB of Kernel address space there is only
479 * one pgd/p4d for the whole kernel. Cloning that would map the whole
480 * address space into the user page-tables, making PTI useless. So clone
481 * the page-table on the PMD level to prevent that.
482 */
483static void __init pti_clone_user_shared(void)
484{
485 unsigned long start, end;
486
487 start = CPU_ENTRY_AREA_BASE;
488 end = start + (PAGE_SIZE * CPU_ENTRY_AREA_PAGES);
489
490 pti_clone_pgtable(start, end, PTI_CLONE_PMD, false);
491}
492#endif /* CONFIG_X86_64 */
493
494/*
495 * Clone the ESPFIX P4D into the user space visible page table
496 */
497static void __init pti_setup_espfix64(void)
498{
499#ifdef CONFIG_X86_ESPFIX64
500 pti_clone_p4d(ESPFIX_BASE_ADDR);
501#endif
502}
503
504/*
505 * Clone the populated PMDs of the entry text and force it RO.
506 */
507static void pti_clone_entry_text(bool late)
508{
509 pti_clone_pgtable(start: (unsigned long) __entry_text_start,
510 end: (unsigned long) __entry_text_end,
511 PTI_LEVEL_KERNEL_IMAGE, late_text: late);
512}
513
514/*
515 * Global pages and PCIDs are both ways to make kernel TLB entries
516 * live longer, reduce TLB misses and improve kernel performance.
517 * But, leaving all kernel text Global makes it potentially accessible
518 * to Meltdown-style attacks which make it trivial to find gadgets or
519 * defeat KASLR.
520 *
521 * Only use global pages when it is really worth it.
522 */
523static inline bool pti_kernel_image_global_ok(void)
524{
525 /*
526 * Systems with PCIDs get little benefit from global
527 * kernel text and are not worth the downsides.
528 */
529 if (cpu_feature_enabled(X86_FEATURE_PCID))
530 return false;
531
532 /*
533 * Only do global kernel image for pti=auto. Do the most
534 * secure thing (not global) if pti=on specified.
535 */
536 if (pti_mode != PTI_AUTO)
537 return false;
538
539 /*
540 * K8 may not tolerate the cleared _PAGE_RW on the userspace
541 * global kernel image pages. Do the safe thing (disable
542 * global kernel image). This is unlikely to ever be
543 * noticed because PTI is disabled by default on AMD CPUs.
544 */
545 if (boot_cpu_has(X86_FEATURE_K8))
546 return false;
547
548 /*
549 * RANDSTRUCT derives its hardening benefits from the
550 * attacker's lack of knowledge about the layout of kernel
551 * data structures. Keep the kernel image non-global in
552 * cases where RANDSTRUCT is in use to help keep the layout a
553 * secret.
554 */
555 if (IS_ENABLED(CONFIG_RANDSTRUCT))
556 return false;
557
558 return true;
559}
560
561/*
562 * For some configurations, map all of kernel text into the user page
563 * tables. This reduces TLB misses, especially on non-PCID systems.
564 */
565static void pti_clone_kernel_text(void)
566{
567 /*
568 * rodata is part of the kernel image and is normally
569 * readable on the filesystem or on the web. But, do not
570 * clone the areas past rodata, they might contain secrets.
571 */
572 unsigned long start = PFN_ALIGN(_text);
573 unsigned long end_clone = (unsigned long)__end_rodata_aligned;
574 unsigned long end_global = PFN_ALIGN((unsigned long)_etext);
575
576 if (!pti_kernel_image_global_ok())
577 return;
578
579 pr_debug("mapping partial kernel image into user address space\n");
580
581 /*
582 * Note that this will undo _some_ of the work that
583 * pti_set_kernel_image_nonglobal() did to clear the
584 * global bit.
585 */
586 pti_clone_pgtable(start, end: end_clone, PTI_LEVEL_KERNEL_IMAGE, late_text: false);
587
588 /*
589 * pti_clone_pgtable() will set the global bit in any PMDs
590 * that it clones, but we also need to get any PTEs in
591 * the last level for areas that are not huge-page-aligned.
592 */
593
594 /* Set the global bit for normal non-__init kernel text: */
595 set_memory_global(addr: start, numpages: (end_global - start) >> PAGE_SHIFT);
596}
597
598static void pti_set_kernel_image_nonglobal(void)
599{
600 /*
601 * The identity map is created with PMDs, regardless of the
602 * actual length of the kernel. We need to clear
603 * _PAGE_GLOBAL up to a PMD boundary, not just to the end
604 * of the image.
605 */
606 unsigned long start = PFN_ALIGN(_text);
607 unsigned long end = ALIGN((unsigned long)_end, PMD_SIZE);
608
609 /*
610 * This clears _PAGE_GLOBAL from the entire kernel image.
611 * pti_clone_kernel_text() map put _PAGE_GLOBAL back for
612 * areas that are mapped to userspace.
613 */
614 set_memory_nonglobal(addr: start, numpages: (end - start) >> PAGE_SHIFT);
615}
616
617/*
618 * Initialize kernel page table isolation
619 */
620void __init pti_init(void)
621{
622 if (!boot_cpu_has(X86_FEATURE_PTI))
623 return;
624
625 pr_info("enabled\n");
626
627#ifdef CONFIG_X86_32
628 /*
629 * We check for X86_FEATURE_PCID here. But the init-code will
630 * clear the feature flag on 32 bit because the feature is not
631 * supported on 32 bit anyway. To print the warning we need to
632 * check with cpuid directly again.
633 */
634 if (cpuid_ecx(0x1) & BIT(17)) {
635 /* Use printk to work around pr_fmt() */
636 printk(KERN_WARNING "\n");
637 printk(KERN_WARNING "************************************************************\n");
638 printk(KERN_WARNING "** WARNING! WARNING! WARNING! WARNING! WARNING! WARNING! **\n");
639 printk(KERN_WARNING "** **\n");
640 printk(KERN_WARNING "** You are using 32-bit PTI on a 64-bit PCID-capable CPU. **\n");
641 printk(KERN_WARNING "** Your performance will increase dramatically if you **\n");
642 printk(KERN_WARNING "** switch to a 64-bit kernel! **\n");
643 printk(KERN_WARNING "** **\n");
644 printk(KERN_WARNING "** WARNING! WARNING! WARNING! WARNING! WARNING! WARNING! **\n");
645 printk(KERN_WARNING "************************************************************\n");
646 }
647#endif
648
649 pti_clone_user_shared();
650
651 /* Undo all global bits from the init pagetables in head_64.S: */
652 pti_set_kernel_image_nonglobal();
653
654 /* Replace some of the global bits just for shared entry text: */
655 /*
656 * This is very early in boot. Device and Late initcalls can do
657 * modprobe before free_initmem() and mark_readonly(). This
658 * pti_clone_entry_text() allows those user-mode-helpers to function,
659 * but notably the text is still RW.
660 */
661 pti_clone_entry_text(late: false);
662 pti_setup_espfix64();
663 pti_setup_vsyscall();
664}
665
666/*
667 * Finalize the kernel mappings in the userspace page-table. Some of the
668 * mappings for the kernel image might have changed since pti_init()
669 * cloned them. This is because parts of the kernel image have been
670 * mapped RO and/or NX. These changes need to be cloned again to the
671 * userspace page-table.
672 */
673void pti_finalize(void)
674{
675 if (!boot_cpu_has(X86_FEATURE_PTI))
676 return;
677 /*
678 * This is after free_initmem() (all initcalls are done) and we've done
679 * mark_readonly(). Text is now NX which might've split some PMDs
680 * relative to the early clone.
681 */
682 pti_clone_entry_text(late: true);
683 pti_clone_kernel_text();
684
685 debug_checkwx_user();
686}
687