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
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
60 static 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
66 static 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. */
73 static enum pti_mode {
74 PTI_AUTO = 0,
75 PTI_FORCE_OFF,
76 PTI_FORCE_ON
77 } pti_mode;
78
79 void __init pti_check_boottime_disable(void)
80 {
81 if (hypervisor_is_type(X86_HYPER_XEN_PV)) {
82 pti_mode = PTI_FORCE_OFF;
83 pti_print_if_insecure("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("disabled on command line.");
91 return;
92 }
93
94 if (pti_mode == PTI_FORCE_ON)
95 pti_print_if_secure("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
103 static 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 }
115 early_param("pti", pti_parse_cmdline);
116
117 static int __init pti_parse_cmdline_nopti(char *arg)
118 {
119 pti_mode = PTI_FORCE_OFF;
120 return 0;
121 }
122 early_param("nopti", pti_parse_cmdline_nopti);
123
124 pgd_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(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 */
171 static 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)) {
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_leaf(*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 */
199 static 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_leaf(*p4d) != 0);
210 if (p4d_none(*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(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_leaf(*pud)) {
221 WARN_ON(1);
222 return NULL;
223 }
224 if (pud_none(*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(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 */
244 static pte_t *pti_user_pagetable_walk_pte(unsigned long address, bool late_text)
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 /* Large PMD mapping found */
255 if (pmd_leaf(*pmd)) {
256 /* Clear the PMD if we hit a large mapping from the first round */
257 if (late_text) {
258 set_pmd(pmd, __pmd(0));
259 } else {
260 WARN_ON_ONCE(1);
261 return NULL;
262 }
263 }
264
265 if (pmd_none(*pmd)) {
266 unsigned long new_pte_page = __get_free_page(gfp);
267 if (!new_pte_page)
268 return NULL;
269
270 set_pmd(pmd, __pmd(_KERNPG_TABLE | __pa(new_pte_page)));
271 }
272
273 pte = pte_offset_kernel(pmd, address);
274 if (pte_flags(*pte) & _PAGE_USER) {
275 WARN_ONCE(1, "attempt to walk to user pte\n");
276 return NULL;
277 }
278 return pte;
279 }
280
281 #ifdef CONFIG_X86_VSYSCALL_EMULATION
282 static void __init pti_setup_vsyscall(void)
283 {
284 pte_t *pte, *target_pte;
285 unsigned int level;
286
287 pte = lookup_address(VSYSCALL_ADDR, &level);
288 if (!pte || WARN_ON(level != PG_LEVEL_4K) || pte_none(*pte))
289 return;
290
291 target_pte = pti_user_pagetable_walk_pte(VSYSCALL_ADDR, false);
292 if (WARN_ON(!target_pte))
293 return;
294
295 *target_pte = *pte;
296 set_vsyscall_pgtable_user_bits(kernel_to_user_pgdp(swapper_pg_dir));
297 }
298 #else
299 static void __init pti_setup_vsyscall(void) { }
300 #endif
301
302 enum pti_clone_level {
303 PTI_CLONE_PMD,
304 PTI_CLONE_PTE,
305 };
306
307 static void
308 pti_clone_pgtable(unsigned long start, unsigned long end,
309 enum pti_clone_level level, bool late_text)
310 {
311 unsigned long addr;
312
313 /*
314 * Clone the populated PMDs which cover start to end. These PMD areas
315 * can have holes.
316 */
317 for (addr = start; addr < end;) {
318 pte_t *pte, *target_pte;
319 pmd_t *pmd, *target_pmd;
320 pgd_t *pgd;
321 p4d_t *p4d;
322 pud_t *pud;
323
324 /* Overflow check */
325 if (addr < start)
326 break;
327
328 pgd = pgd_offset_k(addr);
329 if (WARN_ON(pgd_none(*pgd)))
330 return;
331 p4d = p4d_offset(pgd, addr);
332 if (WARN_ON(p4d_none(*p4d)))
333 return;
334
335 pud = pud_offset(p4d, addr);
336 if (pud_none(*pud)) {
337 WARN_ON_ONCE(addr & ~PUD_MASK);
338 addr = round_up(addr + 1, PUD_SIZE);
339 continue;
340 }
341
342 pmd = pmd_offset(pud, addr);
343 if (pmd_none(*pmd)) {
344 WARN_ON_ONCE(addr & ~PMD_MASK);
345 addr = round_up(addr + 1, PMD_SIZE);
346 continue;
347 }
348
349 if (pmd_leaf(*pmd) || level == PTI_CLONE_PMD) {
350 target_pmd = pti_user_pagetable_walk_pmd(addr);
351 if (WARN_ON(!target_pmd))
352 return;
353
354 /*
355 * Only clone present PMDs. This ensures only setting
356 * _PAGE_GLOBAL on present PMDs. This should only be
357 * called on well-known addresses anyway, so a non-
358 * present PMD would be a surprise.
359 */
360 if (WARN_ON(!(pmd_flags(*pmd) & _PAGE_PRESENT)))
361 return;
362
363 /*
364 * Setting 'target_pmd' below creates a mapping in both
365 * the user and kernel page tables. It is effectively
366 * global, so set it as global in both copies. Note:
367 * the X86_FEATURE_PGE check is not _required_ because
368 * the CPU ignores _PAGE_GLOBAL when PGE is not
369 * supported. The check keeps consistency with
370 * code that only set this bit when supported.
371 */
372 if (boot_cpu_has(X86_FEATURE_PGE))
373 *pmd = pmd_set_flags(*pmd, _PAGE_GLOBAL);
374
375 /*
376 * Copy the PMD. That is, the kernelmode and usermode
377 * tables will share the last-level page tables of this
378 * address range
379 */
380 *target_pmd = *pmd;
381
382 addr = round_up(addr + 1, PMD_SIZE);
383
384 } else if (level == PTI_CLONE_PTE) {
385
386 /* Walk the page-table down to the pte level */
387 pte = pte_offset_kernel(pmd, addr);
388 if (pte_none(*pte)) {
389 addr = round_up(addr + 1, PAGE_SIZE);
390 continue;
391 }
392
393 /* Only clone present PTEs */
394 if (WARN_ON(!(pte_flags(*pte) & _PAGE_PRESENT)))
395 return;
396
397 /* Allocate PTE in the user page-table */
398 target_pte = pti_user_pagetable_walk_pte(addr, late_text);
399 if (WARN_ON(!target_pte))
400 return;
401
402 /* Set GLOBAL bit in both PTEs */
403 if (boot_cpu_has(X86_FEATURE_PGE))
404 *pte = pte_set_flags(*pte, _PAGE_GLOBAL);
405
406 /* Clone the PTE */
407 *target_pte = *pte;
408
409 addr = round_up(addr + 1, PAGE_SIZE);
410
411 } else {
412 BUG();
413 }
414 }
415 }
416
417 #ifdef CONFIG_X86_64
418 /*
419 * Clone a single p4d (i.e. a top-level entry on 4-level systems and a
420 * next-level entry on 5-level systems.
421 */
422 static void __init pti_clone_p4d(unsigned long addr)
423 {
424 p4d_t *kernel_p4d, *user_p4d;
425 pgd_t *kernel_pgd;
426
427 user_p4d = pti_user_pagetable_walk_p4d(addr);
428 if (!user_p4d)
429 return;
430
431 kernel_pgd = pgd_offset_k(addr);
432 kernel_p4d = p4d_offset(kernel_pgd, addr);
433 *user_p4d = *kernel_p4d;
434 }
435
436 /*
437 * Clone the CPU_ENTRY_AREA and associated data into the user space visible
438 * page table.
439 */
440 static void __init pti_clone_user_shared(void)
441 {
442 unsigned int cpu;
443
444 pti_clone_p4d(CPU_ENTRY_AREA_BASE);
445
446 for_each_possible_cpu(cpu) {
447 /*
448 * The SYSCALL64 entry code needs one word of scratch space
449 * in which to spill a register. It lives in the sp2 slot
450 * of the CPU's TSS.
451 *
452 * This is done for all possible CPUs during boot to ensure
453 * that it's propagated to all mms.
454 */
455
456 unsigned long va = (unsigned long)&per_cpu(cpu_tss_rw, cpu);
457 phys_addr_t pa = per_cpu_ptr_to_phys((void *)va);
458 pte_t *target_pte;
459
460 target_pte = pti_user_pagetable_walk_pte(va, false);
461 if (WARN_ON(!target_pte))
462 return;
463
464 *target_pte = pfn_pte(pa >> PAGE_SHIFT, PAGE_KERNEL);
465 }
466 }
467
468 #else /* CONFIG_X86_64 */
469
470 /*
471 * On 32 bit PAE systems with 1GB of Kernel address space there is only
472 * one pgd/p4d for the whole kernel. Cloning that would map the whole
473 * address space into the user page-tables, making PTI useless. So clone
474 * the page-table on the PMD level to prevent that.
475 */
476 static void __init pti_clone_user_shared(void)
477 {
478 unsigned long start, end;
479
480 start = CPU_ENTRY_AREA_BASE;
481 end = start + (PAGE_SIZE * CPU_ENTRY_AREA_PAGES);
482
483 pti_clone_pgtable(start, end, PTI_CLONE_PMD, false);
484 }
485 #endif /* CONFIG_X86_64 */
486
487 /*
488 * Clone the ESPFIX P4D into the user space visible page table
489 */
490 static void __init pti_setup_espfix64(void)
491 {
492 #ifdef CONFIG_X86_ESPFIX64
493 pti_clone_p4d(ESPFIX_BASE_ADDR);
494 #endif
495 }
496
497 /*
498 * Clone the populated PMDs of the entry text and force it RO.
499 */
500 static void pti_clone_entry_text(bool late)
501 {
502 pti_clone_pgtable((unsigned long) __entry_text_start,
503 (unsigned long) __entry_text_end,
504 PTI_LEVEL_KERNEL_IMAGE, late);
505 }
506
507 /*
508 * Global pages and PCIDs are both ways to make kernel TLB entries
509 * live longer, reduce TLB misses and improve kernel performance.
510 * But, leaving all kernel text Global makes it potentially accessible
511 * to Meltdown-style attacks which make it trivial to find gadgets or
512 * defeat KASLR.
513 *
514 * Only use global pages when it is really worth it.
515 */
516 static inline bool pti_kernel_image_global_ok(void)
517 {
518 /*
519 * Systems with PCIDs get little benefit from global
520 * kernel text and are not worth the downsides.
521 */
522 if (cpu_feature_enabled(X86_FEATURE_PCID))
523 return false;
524
525 /*
526 * Only do global kernel image for pti=auto. Do the most
527 * secure thing (not global) if pti=on specified.
528 */
529 if (pti_mode != PTI_AUTO)
530 return false;
531
532 /*
533 * K8 may not tolerate the cleared _PAGE_RW on the userspace
534 * global kernel image pages. Do the safe thing (disable
535 * global kernel image). This is unlikely to ever be
536 * noticed because PTI is disabled by default on AMD CPUs.
537 */
538 if (boot_cpu_has(X86_FEATURE_K8))
539 return false;
540
541 /*
542 * RANDSTRUCT derives its hardening benefits from the
543 * attacker's lack of knowledge about the layout of kernel
544 * data structures. Keep the kernel image non-global in
545 * cases where RANDSTRUCT is in use to help keep the layout a
546 * secret.
547 */
548 if (IS_ENABLED(CONFIG_RANDSTRUCT))
549 return false;
550
551 return true;
552 }
553
554 /*
555 * For some configurations, map all of kernel text into the user page
556 * tables. This reduces TLB misses, especially on non-PCID systems.
557 */
558 static void pti_clone_kernel_text(void)
559 {
560 /*
561 * rodata is part of the kernel image and is normally
562 * readable on the filesystem or on the web. But, do not
563 * clone the areas past rodata, they might contain secrets.
564 */
565 unsigned long start = PFN_ALIGN(_text);
566 unsigned long end_clone = (unsigned long)__end_rodata_aligned;
567 unsigned long end_global = PFN_ALIGN((unsigned long)_etext);
568
569 if (!pti_kernel_image_global_ok())
570 return;
571
572 pr_debug("mapping partial kernel image into user address space\n");
573
574 /*
575 * Note that this will undo _some_ of the work that
576 * pti_set_kernel_image_nonglobal() did to clear the
577 * global bit.
578 */
579 pti_clone_pgtable(start, end_clone, PTI_LEVEL_KERNEL_IMAGE, false);
580
581 /*
582 * pti_clone_pgtable() will set the global bit in any PMDs
583 * that it clones, but we also need to get any PTEs in
584 * the last level for areas that are not huge-page-aligned.
585 */
586
587 /* Set the global bit for normal non-__init kernel text: */
588 set_memory_global(start, (end_global - start) >> PAGE_SHIFT);
589 }
590
591 static void pti_set_kernel_image_nonglobal(void)
592 {
593 /*
594 * The identity map is created with PMDs, regardless of the
595 * actual length of the kernel. We need to clear
596 * _PAGE_GLOBAL up to a PMD boundary, not just to the end
597 * of the image.
598 */
599 unsigned long start = PFN_ALIGN(_text);
600 unsigned long end = ALIGN((unsigned long)_end, PMD_SIZE);
601
602 /*
603 * This clears _PAGE_GLOBAL from the entire kernel image.
604 * pti_clone_kernel_text() map put _PAGE_GLOBAL back for
605 * areas that are mapped to userspace.
606 */
607 set_memory_nonglobal(start, (end - start) >> PAGE_SHIFT);
608 }
609
610 /*
611 * Initialize kernel page table isolation
612 */
613 void __init pti_init(void)
614 {
615 if (!boot_cpu_has(X86_FEATURE_PTI))
616 return;
617
618 pr_info("enabled\n");
619
620 #ifdef CONFIG_X86_32
621 /*
622 * We check for X86_FEATURE_PCID here. But the init-code will
623 * clear the feature flag on 32 bit because the feature is not
624 * supported on 32 bit anyway. To print the warning we need to
625 * check with cpuid directly again.
626 */
627 if (cpuid_ecx(0x1) & BIT(17)) {
628 /* Use printk to work around pr_fmt() */
629 printk(KERN_WARNING "\n");
630 printk(KERN_WARNING "************************************************************\n");
631 printk(KERN_WARNING "** WARNING! WARNING! WARNING! WARNING! WARNING! WARNING! **\n");
632 printk(KERN_WARNING "** **\n");
633 printk(KERN_WARNING "** You are using 32-bit PTI on a 64-bit PCID-capable CPU. **\n");
634 printk(KERN_WARNING "** Your performance will increase dramatically if you **\n");
635 printk(KERN_WARNING "** switch to a 64-bit kernel! **\n");
636 printk(KERN_WARNING "** **\n");
637 printk(KERN_WARNING "** WARNING! WARNING! WARNING! WARNING! WARNING! WARNING! **\n");
638 printk(KERN_WARNING "************************************************************\n");
639 }
640 #endif
641
642 pti_clone_user_shared();
643
644 /* Undo all global bits from the init pagetables in head_64.S: */
645 pti_set_kernel_image_nonglobal();
646
647 /* Replace some of the global bits just for shared entry text: */
648 /*
649 * This is very early in boot. Device and Late initcalls can do
650 * modprobe before free_initmem() and mark_readonly(). This
651 * pti_clone_entry_text() allows those user-mode-helpers to function,
652 * but notably the text is still RW.
653 */
654 pti_clone_entry_text(false);
655 pti_setup_espfix64();
656 pti_setup_vsyscall();
657 }
658
659 /*
660 * Finalize the kernel mappings in the userspace page-table. Some of the
661 * mappings for the kernel image might have changed since pti_init()
662 * cloned them. This is because parts of the kernel image have been
663 * mapped RO and/or NX. These changes need to be cloned again to the
664 * userspace page-table.
665 */
666 void pti_finalize(void)
667 {
668 if (!boot_cpu_has(X86_FEATURE_PTI))
669 return;
670 /*
671 * This is after free_initmem() (all initcalls are done) and we've done
672 * mark_readonly(). Text is now NX which might've split some PMDs
673 * relative to the early clone.
674 */
675 pti_clone_entry_text(true);
676 pti_clone_kernel_text();
677
678 debug_checkwx_user();
679 }
680
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.