개념 설명 전체
arch/arm64/mm/mmu.c · v6.18.37
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3 * Based on arch/arm/mm/mmu.c
4 *
5 * Copyright (C) 1995-2005 Russell King
6 * Copyright (C) 2012 ARM Ltd.
7 */
8
9 #include <linux/cache.h>
10 #include <linux/export.h>
11 #include <linux/kernel.h>
12 #include <linux/errno.h>
13 #include <linux/init.h>
14 #include <linux/ioport.h>
15 #include <linux/kexec.h>
16 #include <linux/libfdt.h>
17 #include <linux/mman.h>
18 #include <linux/nodemask.h>
19 #include <linux/memblock.h>
20 #include <linux/memremap.h>
21 #include <linux/memory.h>
22 #include <linux/fs.h>
23 #include <linux/io.h>
24 #include <linux/mm.h>
25 #include <linux/vmalloc.h>
26 #include <linux/set_memory.h>
27 #include <linux/kfence.h>
28 #include <linux/pkeys.h>
29 #include <linux/mm_inline.h>
30 #include <linux/pagewalk.h>
31 #include <linux/stop_machine.h>
32
33 #include <asm/barrier.h>
34 #include <asm/cputype.h>
35 #include <asm/fixmap.h>
36 #include <asm/kasan.h>
37 #include <asm/kernel-pgtable.h>
38 #include <asm/sections.h>
39 #include <asm/setup.h>
40 #include <linux/sizes.h>
41 #include <asm/tlb.h>
42 #include <asm/mmu_context.h>
43 #include <asm/ptdump.h>
44 #include <asm/tlbflush.h>
45 #include <asm/pgalloc.h>
46 #include <asm/kfence.h>
47
48 #define NO_BLOCK_MAPPINGS BIT(0)
49 #define NO_CONT_MAPPINGS BIT(1)
50 #define NO_EXEC_MAPPINGS BIT(2) /* assumes FEAT_HPDS is not used */
51
52 #define INVALID_PHYS_ADDR (-1ULL)
53
54 DEFINE_STATIC_KEY_FALSE(arm64_ptdump_lock_key);
55
56 u64 kimage_voffset __ro_after_init;
57 EXPORT_SYMBOL(kimage_voffset);
58
59 u32 __boot_cpu_mode[] = { BOOT_CPU_MODE_EL2, BOOT_CPU_MODE_EL1 };
60
61 static bool rodata_is_rw __ro_after_init = true;
62
63 /*
64 * The booting CPU updates the failed status @__early_cpu_boot_status,
65 * with MMU turned off.
66 */
67 long __section(".mmuoff.data.write") __early_cpu_boot_status;
68
69 /*
70 * Empty_zero_page is a special page that is used for zero-initialized data
71 * and COW.
72 */
73 unsigned long empty_zero_page[PAGE_SIZE / sizeof(unsigned long)] __page_aligned_bss;
74 EXPORT_SYMBOL(empty_zero_page);
75
76 static DEFINE_SPINLOCK(swapper_pgdir_lock);
77 static DEFINE_MUTEX(fixmap_lock);
78
79 void noinstr set_swapper_pgd(pgd_t *pgdp, pgd_t pgd)
80 {
81 pgd_t *fixmap_pgdp;
82
83 /*
84 * Don't bother with the fixmap if swapper_pg_dir is still mapped
85 * writable in the kernel mapping.
86 */
87 if (rodata_is_rw) {
88 WRITE_ONCE(*pgdp, pgd);
89 dsb(ishst);
90 isb();
91 return;
92 }
93
94 spin_lock(&swapper_pgdir_lock);
95 fixmap_pgdp = pgd_set_fixmap(__pa_symbol(pgdp));
96 WRITE_ONCE(*fixmap_pgdp, pgd);
97 /*
98 * We need dsb(ishst) here to ensure the page-table-walker sees
99 * our new entry before set_p?d() returns. The fixmap's
100 * flush_tlb_kernel_range() via clear_fixmap() does this for us.
101 */
102 pgd_clear_fixmap();
103 spin_unlock(&swapper_pgdir_lock);
104 }
105
106 pgprot_t phys_mem_access_prot(struct file *file, unsigned long pfn,
107 unsigned long size, pgprot_t vma_prot)
108 {
109 if (!pfn_is_map_memory(pfn))
110 return pgprot_noncached(vma_prot);
111 else if (file->f_flags & O_SYNC)
112 return pgprot_writecombine(vma_prot);
113 return vma_prot;
114 }
115 EXPORT_SYMBOL(phys_mem_access_prot);
116
117 static phys_addr_t __init early_pgtable_alloc(enum pgtable_type pgtable_type)
118 {
119 phys_addr_t phys;
120
121 phys = memblock_phys_alloc_range(PAGE_SIZE, PAGE_SIZE, 0,
122 MEMBLOCK_ALLOC_NOLEAKTRACE);
123 if (!phys)
124 panic("Failed to allocate page table page\n");
125
126 return phys;
127 }
128
129 bool pgattr_change_is_safe(pteval_t old, pteval_t new)
130 {
131 /*
132 * The following mapping attributes may be updated in live
133 * kernel mappings without the need for break-before-make.
134 */
135 pteval_t mask = PTE_PXN | PTE_RDONLY | PTE_WRITE | PTE_NG |
136 PTE_SWBITS_MASK;
137
138 /* creating or taking down mappings is always safe */
139 if (!pte_valid(__pte(old)) || !pte_valid(__pte(new)))
140 return true;
141
142 /* A live entry's pfn should not change */
143 if (pte_pfn(__pte(old)) != pte_pfn(__pte(new)))
144 return false;
145
146 /* live contiguous mappings may not be manipulated at all */
147 if ((old | new) & PTE_CONT)
148 return false;
149
150 /* Transitioning from Non-Global to Global is unsafe */
151 if (old & ~new & PTE_NG)
152 return false;
153
154 /*
155 * Changing the memory type between Normal and Normal-Tagged is safe
156 * since Tagged is considered a permission attribute from the
157 * mismatched attribute aliases perspective.
158 */
159 if (((old & PTE_ATTRINDX_MASK) == PTE_ATTRINDX(MT_NORMAL) ||
160 (old & PTE_ATTRINDX_MASK) == PTE_ATTRINDX(MT_NORMAL_TAGGED)) &&
161 ((new & PTE_ATTRINDX_MASK) == PTE_ATTRINDX(MT_NORMAL) ||
162 (new & PTE_ATTRINDX_MASK) == PTE_ATTRINDX(MT_NORMAL_TAGGED)))
163 mask |= PTE_ATTRINDX_MASK;
164
165 return ((old ^ new) & ~mask) == 0;
166 }
167
168 static void init_clear_pgtable(void *table)
169 {
170 clear_page(table);
171
172 /* Ensure the zeroing is observed by page table walks. */
173 dsb(ishst);
174 }
175
176 static void init_pte(pte_t *ptep, unsigned long addr, unsigned long end,
177 phys_addr_t phys, pgprot_t prot)
178 {
179 do {
180 pte_t old_pte = __ptep_get(ptep);
181
182 /*
183 * Required barriers to make this visible to the table walker
184 * are deferred to the end of alloc_init_cont_pte().
185 */
186 __set_pte_nosync(ptep, pfn_pte(__phys_to_pfn(phys), prot));
187
188 /*
189 * After the PTE entry has been populated once, we
190 * only allow updates to the permission attributes.
191 */
192 BUG_ON(!pgattr_change_is_safe(pte_val(old_pte),
193 pte_val(__ptep_get(ptep))));
194
195 phys += PAGE_SIZE;
196 } while (ptep++, addr += PAGE_SIZE, addr != end);
197 }
198
199 static int alloc_init_cont_pte(pmd_t *pmdp, unsigned long addr,
200 unsigned long end, phys_addr_t phys,
201 pgprot_t prot,
202 phys_addr_t (*pgtable_alloc)(enum pgtable_type),
203 int flags)
204 {
205 unsigned long next;
206 pmd_t pmd = READ_ONCE(*pmdp);
207 pte_t *ptep;
208
209 BUG_ON(pmd_sect(pmd));
210 if (pmd_none(pmd)) {
211 pmdval_t pmdval = PMD_TYPE_TABLE | PMD_TABLE_UXN | PMD_TABLE_AF;
212 phys_addr_t pte_phys;
213
214 if (flags & NO_EXEC_MAPPINGS)
215 pmdval |= PMD_TABLE_PXN;
216 BUG_ON(!pgtable_alloc);
217 pte_phys = pgtable_alloc(TABLE_PTE);
218 if (pte_phys == INVALID_PHYS_ADDR)
219 return -ENOMEM;
220 ptep = pte_set_fixmap(pte_phys);
221 init_clear_pgtable(ptep);
222 ptep += pte_index(addr);
223 __pmd_populate(pmdp, pte_phys, pmdval);
224 } else {
225 BUG_ON(pmd_bad(pmd));
226 ptep = pte_set_fixmap_offset(pmdp, addr);
227 }
228
229 do {
230 pgprot_t __prot = prot;
231
232 next = pte_cont_addr_end(addr, end);
233
234 /* use a contiguous mapping if the range is suitably aligned */
235 if ((((addr | next | phys) & ~CONT_PTE_MASK) == 0) &&
236 (flags & NO_CONT_MAPPINGS) == 0)
237 __prot = __pgprot(pgprot_val(prot) | PTE_CONT);
238
239 init_pte(ptep, addr, next, phys, __prot);
240
241 ptep += pte_index(next) - pte_index(addr);
242 phys += next - addr;
243 } while (addr = next, addr != end);
244
245 /*
246 * Note: barriers and maintenance necessary to clear the fixmap slot
247 * ensure that all previous pgtable writes are visible to the table
248 * walker.
249 */
250 pte_clear_fixmap();
251
252 return 0;
253 }
254
255 static int init_pmd(pmd_t *pmdp, unsigned long addr, unsigned long end,
256 phys_addr_t phys, pgprot_t prot,
257 phys_addr_t (*pgtable_alloc)(enum pgtable_type), int flags)
258 {
259 unsigned long next;
260
261 do {
262 pmd_t old_pmd = READ_ONCE(*pmdp);
263
264 next = pmd_addr_end(addr, end);
265
266 /* try section mapping first */
267 if (((addr | next | phys) & ~PMD_MASK) == 0 &&
268 (flags & NO_BLOCK_MAPPINGS) == 0) {
269 pmd_set_huge(pmdp, phys, prot);
270
271 /*
272 * After the PMD entry has been populated once, we
273 * only allow updates to the permission attributes.
274 */
275 BUG_ON(!pgattr_change_is_safe(pmd_val(old_pmd),
276 READ_ONCE(pmd_val(*pmdp))));
277 } else {
278 int ret;
279
280 ret = alloc_init_cont_pte(pmdp, addr, next, phys, prot,
281 pgtable_alloc, flags);
282 if (ret)
283 return ret;
284
285 BUG_ON(pmd_val(old_pmd) != 0 &&
286 pmd_val(old_pmd) != READ_ONCE(pmd_val(*pmdp)));
287 }
288 phys += next - addr;
289 } while (pmdp++, addr = next, addr != end);
290
291 return 0;
292 }
293
294 static int alloc_init_cont_pmd(pud_t *pudp, unsigned long addr,
295 unsigned long end, phys_addr_t phys,
296 pgprot_t prot,
297 phys_addr_t (*pgtable_alloc)(enum pgtable_type),
298 int flags)
299 {
300 int ret;
301 unsigned long next;
302 pud_t pud = READ_ONCE(*pudp);
303 pmd_t *pmdp;
304
305 /*
306 * Check for initial section mappings in the pgd/pud.
307 */
308 BUG_ON(pud_sect(pud));
309 if (pud_none(pud)) {
310 pudval_t pudval = PUD_TYPE_TABLE | PUD_TABLE_UXN | PUD_TABLE_AF;
311 phys_addr_t pmd_phys;
312
313 if (flags & NO_EXEC_MAPPINGS)
314 pudval |= PUD_TABLE_PXN;
315 BUG_ON(!pgtable_alloc);
316 pmd_phys = pgtable_alloc(TABLE_PMD);
317 if (pmd_phys == INVALID_PHYS_ADDR)
318 return -ENOMEM;
319 pmdp = pmd_set_fixmap(pmd_phys);
320 init_clear_pgtable(pmdp);
321 pmdp += pmd_index(addr);
322 __pud_populate(pudp, pmd_phys, pudval);
323 } else {
324 BUG_ON(pud_bad(pud));
325 pmdp = pmd_set_fixmap_offset(pudp, addr);
326 }
327
328 do {
329 pgprot_t __prot = prot;
330
331 next = pmd_cont_addr_end(addr, end);
332
333 /* use a contiguous mapping if the range is suitably aligned */
334 if ((((addr | next | phys) & ~CONT_PMD_MASK) == 0) &&
335 (flags & NO_CONT_MAPPINGS) == 0)
336 __prot = __pgprot(pgprot_val(prot) | PTE_CONT);
337
338 ret = init_pmd(pmdp, addr, next, phys, __prot, pgtable_alloc, flags);
339 if (ret)
340 goto out;
341
342 pmdp += pmd_index(next) - pmd_index(addr);
343 phys += next - addr;
344 } while (addr = next, addr != end);
345
346 out:
347 pmd_clear_fixmap();
348
349 return ret;
350 }
351
352 static int alloc_init_pud(p4d_t *p4dp, unsigned long addr, unsigned long end,
353 phys_addr_t phys, pgprot_t prot,
354 phys_addr_t (*pgtable_alloc)(enum pgtable_type),
355 int flags)
356 {
357 int ret = 0;
358 unsigned long next;
359 p4d_t p4d = READ_ONCE(*p4dp);
360 pud_t *pudp;
361
362 if (p4d_none(p4d)) {
363 p4dval_t p4dval = P4D_TYPE_TABLE | P4D_TABLE_UXN | P4D_TABLE_AF;
364 phys_addr_t pud_phys;
365
366 if (flags & NO_EXEC_MAPPINGS)
367 p4dval |= P4D_TABLE_PXN;
368 BUG_ON(!pgtable_alloc);
369 pud_phys = pgtable_alloc(TABLE_PUD);
370 if (pud_phys == INVALID_PHYS_ADDR)
371 return -ENOMEM;
372 pudp = pud_set_fixmap(pud_phys);
373 init_clear_pgtable(pudp);
374 pudp += pud_index(addr);
375 __p4d_populate(p4dp, pud_phys, p4dval);
376 } else {
377 BUG_ON(p4d_bad(p4d));
378 pudp = pud_set_fixmap_offset(p4dp, addr);
379 }
380
381 do {
382 pud_t old_pud = READ_ONCE(*pudp);
383
384 next = pud_addr_end(addr, end);
385
386 /*
387 * For 4K granule only, attempt to put down a 1GB block
388 */
389 if (pud_sect_supported() &&
390 ((addr | next | phys) & ~PUD_MASK) == 0 &&
391 (flags & NO_BLOCK_MAPPINGS) == 0) {
392 pud_set_huge(pudp, phys, prot);
393
394 /*
395 * After the PUD entry has been populated once, we
396 * only allow updates to the permission attributes.
397 */
398 BUG_ON(!pgattr_change_is_safe(pud_val(old_pud),
399 READ_ONCE(pud_val(*pudp))));
400 } else {
401 ret = alloc_init_cont_pmd(pudp, addr, next, phys, prot,
402 pgtable_alloc, flags);
403 if (ret)
404 goto out;
405
406 BUG_ON(pud_val(old_pud) != 0 &&
407 pud_val(old_pud) != READ_ONCE(pud_val(*pudp)));
408 }
409 phys += next - addr;
410 } while (pudp++, addr = next, addr != end);
411
412 out:
413 pud_clear_fixmap();
414
415 return ret;
416 }
417
418 static int alloc_init_p4d(pgd_t *pgdp, unsigned long addr, unsigned long end,
419 phys_addr_t phys, pgprot_t prot,
420 phys_addr_t (*pgtable_alloc)(enum pgtable_type),
421 int flags)
422 {
423 int ret;
424 unsigned long next;
425 pgd_t pgd = READ_ONCE(*pgdp);
426 p4d_t *p4dp;
427
428 if (pgd_none(pgd)) {
429 pgdval_t pgdval = PGD_TYPE_TABLE | PGD_TABLE_UXN | PGD_TABLE_AF;
430 phys_addr_t p4d_phys;
431
432 if (flags & NO_EXEC_MAPPINGS)
433 pgdval |= PGD_TABLE_PXN;
434 BUG_ON(!pgtable_alloc);
435 p4d_phys = pgtable_alloc(TABLE_P4D);
436 if (p4d_phys == INVALID_PHYS_ADDR)
437 return -ENOMEM;
438 p4dp = p4d_set_fixmap(p4d_phys);
439 init_clear_pgtable(p4dp);
440 p4dp += p4d_index(addr);
441 __pgd_populate(pgdp, p4d_phys, pgdval);
442 } else {
443 BUG_ON(pgd_bad(pgd));
444 p4dp = p4d_set_fixmap_offset(pgdp, addr);
445 }
446
447 do {
448 p4d_t old_p4d = READ_ONCE(*p4dp);
449
450 next = p4d_addr_end(addr, end);
451
452 ret = alloc_init_pud(p4dp, addr, next, phys, prot,
453 pgtable_alloc, flags);
454 if (ret)
455 goto out;
456
457 BUG_ON(p4d_val(old_p4d) != 0 &&
458 p4d_val(old_p4d) != READ_ONCE(p4d_val(*p4dp)));
459
460 phys += next - addr;
461 } while (p4dp++, addr = next, addr != end);
462
463 out:
464 p4d_clear_fixmap();
465
466 return ret;
467 }
468
469 static int __create_pgd_mapping_locked(pgd_t *pgdir, phys_addr_t phys,
470 unsigned long virt, phys_addr_t size,
471 pgprot_t prot,
472 phys_addr_t (*pgtable_alloc)(enum pgtable_type),
473 int flags)
474 {
475 int ret;
476 unsigned long addr, end, next;
477 pgd_t *pgdp = pgd_offset_pgd(pgdir, virt);
478
479 /*
480 * If the virtual and physical address don't have the same offset
481 * within a page, we cannot map the region as the caller expects.
482 */
483 if (WARN_ON((phys ^ virt) & ~PAGE_MASK))
484 return -EINVAL;
485
486 phys &= PAGE_MASK;
487 addr = virt & PAGE_MASK;
488 end = PAGE_ALIGN(virt + size);
489
490 do {
491 next = pgd_addr_end(addr, end);
492 ret = alloc_init_p4d(pgdp, addr, next, phys, prot, pgtable_alloc,
493 flags);
494 if (ret)
495 return ret;
496 phys += next - addr;
497 } while (pgdp++, addr = next, addr != end);
498
499 return 0;
500 }
501
502 static int __create_pgd_mapping(pgd_t *pgdir, phys_addr_t phys,
503 unsigned long virt, phys_addr_t size,
504 pgprot_t prot,
505 phys_addr_t (*pgtable_alloc)(enum pgtable_type),
506 int flags)
507 {
508 int ret;
509
510 mutex_lock(&fixmap_lock);
511 ret = __create_pgd_mapping_locked(pgdir, phys, virt, size, prot,
512 pgtable_alloc, flags);
513 mutex_unlock(&fixmap_lock);
514
515 return ret;
516 }
517
518 static void early_create_pgd_mapping(pgd_t *pgdir, phys_addr_t phys,
519 unsigned long virt, phys_addr_t size,
520 pgprot_t prot,
521 phys_addr_t (*pgtable_alloc)(enum pgtable_type),
522 int flags)
523 {
524 int ret;
525
526 ret = __create_pgd_mapping(pgdir, phys, virt, size, prot, pgtable_alloc,
527 flags);
528 if (ret)
529 panic("Failed to create page tables\n");
530 }
531
532 static phys_addr_t __pgd_pgtable_alloc(struct mm_struct *mm, gfp_t gfp,
533 enum pgtable_type pgtable_type)
534 {
535 /* Page is zeroed by init_clear_pgtable() so don't duplicate effort. */
536 struct ptdesc *ptdesc = pagetable_alloc(gfp & ~__GFP_ZERO, 0);
537 phys_addr_t pa;
538
539 if (!ptdesc)
540 return INVALID_PHYS_ADDR;
541
542 pa = page_to_phys(ptdesc_page(ptdesc));
543
544 switch (pgtable_type) {
545 case TABLE_PTE:
546 BUG_ON(!pagetable_pte_ctor(mm, ptdesc));
547 break;
548 case TABLE_PMD:
549 BUG_ON(!pagetable_pmd_ctor(mm, ptdesc));
550 break;
551 case TABLE_PUD:
552 pagetable_pud_ctor(ptdesc);
553 break;
554 case TABLE_P4D:
555 pagetable_p4d_ctor(ptdesc);
556 break;
557 }
558
559 return pa;
560 }
561
562 static phys_addr_t
563 try_pgd_pgtable_alloc_init_mm(enum pgtable_type pgtable_type, gfp_t gfp)
564 {
565 return __pgd_pgtable_alloc(&init_mm, gfp, pgtable_type);
566 }
567
568 static phys_addr_t __maybe_unused
569 pgd_pgtable_alloc_init_mm(enum pgtable_type pgtable_type)
570 {
571 return __pgd_pgtable_alloc(&init_mm, GFP_PGTABLE_KERNEL, pgtable_type);
572 }
573
574 static phys_addr_t
575 pgd_pgtable_alloc_special_mm(enum pgtable_type pgtable_type)
576 {
577 return __pgd_pgtable_alloc(NULL, GFP_PGTABLE_KERNEL, pgtable_type);
578 }
579
580 static void split_contpte(pte_t *ptep)
581 {
582 int i;
583
584 ptep = PTR_ALIGN_DOWN(ptep, sizeof(*ptep) * CONT_PTES);
585 for (i = 0; i < CONT_PTES; i++, ptep++)
586 __set_pte(ptep, pte_mknoncont(__ptep_get(ptep)));
587 }
588
589 static int split_pmd(pmd_t *pmdp, pmd_t pmd, gfp_t gfp, bool to_cont)
590 {
591 pmdval_t tableprot = PMD_TYPE_TABLE | PMD_TABLE_UXN | PMD_TABLE_AF;
592 unsigned long pfn = pmd_pfn(pmd);
593 pgprot_t prot = pmd_pgprot(pmd);
594 phys_addr_t pte_phys;
595 pte_t *ptep;
596 int i;
597
598 pte_phys = try_pgd_pgtable_alloc_init_mm(TABLE_PTE, gfp);
599 if (pte_phys == INVALID_PHYS_ADDR)
600 return -ENOMEM;
601 ptep = (pte_t *)phys_to_virt(pte_phys);
602
603 if (pgprot_val(prot) & PMD_SECT_PXN)
604 tableprot |= PMD_TABLE_PXN;
605
606 prot = __pgprot((pgprot_val(prot) & ~PTE_TYPE_MASK) | PTE_TYPE_PAGE);
607 if (!pmd_valid(pmd))
608 prot = pte_pgprot(pte_mkinvalid(pfn_pte(0, prot)));
609 prot = __pgprot(pgprot_val(prot) & ~PTE_CONT);
610 if (to_cont)
611 prot = __pgprot(pgprot_val(prot) | PTE_CONT);
612
613 for (i = 0; i < PTRS_PER_PTE; i++, ptep++, pfn++)
614 __set_pte(ptep, pfn_pte(pfn, prot));
615
616 /*
617 * Ensure the pte entries are visible to the table walker by the time
618 * the pmd entry that points to the ptes is visible.
619 */
620 dsb(ishst);
621 __pmd_populate(pmdp, pte_phys, tableprot);
622
623 return 0;
624 }
625
626 static void split_contpmd(pmd_t *pmdp)
627 {
628 int i;
629
630 pmdp = PTR_ALIGN_DOWN(pmdp, sizeof(*pmdp) * CONT_PMDS);
631 for (i = 0; i < CONT_PMDS; i++, pmdp++)
632 set_pmd(pmdp, pmd_mknoncont(pmdp_get(pmdp)));
633 }
634
635 static int split_pud(pud_t *pudp, pud_t pud, gfp_t gfp, bool to_cont)
636 {
637 pudval_t tableprot = PUD_TYPE_TABLE | PUD_TABLE_UXN | PUD_TABLE_AF;
638 unsigned int step = PMD_SIZE >> PAGE_SHIFT;
639 unsigned long pfn = pud_pfn(pud);
640 pgprot_t prot = pud_pgprot(pud);
641 phys_addr_t pmd_phys;
642 pmd_t *pmdp;
643 int i;
644
645 pmd_phys = try_pgd_pgtable_alloc_init_mm(TABLE_PMD, gfp);
646 if (pmd_phys == INVALID_PHYS_ADDR)
647 return -ENOMEM;
648 pmdp = (pmd_t *)phys_to_virt(pmd_phys);
649
650 if (pgprot_val(prot) & PMD_SECT_PXN)
651 tableprot |= PUD_TABLE_PXN;
652
653 prot = __pgprot((pgprot_val(prot) & ~PMD_TYPE_MASK) | PMD_TYPE_SECT);
654 if (!pud_valid(pud))
655 prot = pmd_pgprot(pmd_mkinvalid(pfn_pmd(0, prot)));
656 prot = __pgprot(pgprot_val(prot) & ~PTE_CONT);
657 if (to_cont)
658 prot = __pgprot(pgprot_val(prot) | PTE_CONT);
659
660 for (i = 0; i < PTRS_PER_PMD; i++, pmdp++, pfn += step)
661 set_pmd(pmdp, pfn_pmd(pfn, prot));
662
663 /*
664 * Ensure the pmd entries are visible to the table walker by the time
665 * the pud entry that points to the pmds is visible.
666 */
667 dsb(ishst);
668 __pud_populate(pudp, pmd_phys, tableprot);
669
670 return 0;
671 }
672
673 static int split_kernel_leaf_mapping_locked(unsigned long addr)
674 {
675 pgd_t *pgdp, pgd;
676 p4d_t *p4dp, p4d;
677 pud_t *pudp, pud;
678 pmd_t *pmdp, pmd;
679 pte_t *ptep, pte;
680 int ret = 0;
681
682 /*
683 * PGD: If addr is PGD aligned then addr already describes a leaf
684 * boundary. If not present then there is nothing to split.
685 */
686 if (ALIGN_DOWN(addr, PGDIR_SIZE) == addr)
687 goto out;
688 pgdp = pgd_offset_k(addr);
689 pgd = pgdp_get(pgdp);
690 if (!pgd_present(pgd))
691 goto out;
692
693 /*
694 * P4D: If addr is P4D aligned then addr already describes a leaf
695 * boundary. If not present then there is nothing to split.
696 */
697 if (ALIGN_DOWN(addr, P4D_SIZE) == addr)
698 goto out;
699 p4dp = p4d_offset(pgdp, addr);
700 p4d = p4dp_get(p4dp);
701 if (!p4d_present(p4d))
702 goto out;
703
704 /*
705 * PUD: If addr is PUD aligned then addr already describes a leaf
706 * boundary. If not present then there is nothing to split. Otherwise,
707 * if we have a pud leaf, split to contpmd.
708 */
709 if (ALIGN_DOWN(addr, PUD_SIZE) == addr)
710 goto out;
711 pudp = pud_offset(p4dp, addr);
712 pud = pudp_get(pudp);
713 if (!pud_present(pud))
714 goto out;
715 if (pud_leaf(pud)) {
716 ret = split_pud(pudp, pud, GFP_PGTABLE_KERNEL, true);
717 if (ret)
718 goto out;
719 }
720
721 /*
722 * CONTPMD: If addr is CONTPMD aligned then addr already describes a
723 * leaf boundary. If not present then there is nothing to split.
724 * Otherwise, if we have a contpmd leaf, split to pmd.
725 */
726 if (ALIGN_DOWN(addr, CONT_PMD_SIZE) == addr)
727 goto out;
728 pmdp = pmd_offset(pudp, addr);
729 pmd = pmdp_get(pmdp);
730 if (!pmd_present(pmd))
731 goto out;
732 if (pmd_leaf(pmd)) {
733 if (pmd_cont(pmd))
734 split_contpmd(pmdp);
735 /*
736 * PMD: If addr is PMD aligned then addr already describes a
737 * leaf boundary. Otherwise, split to contpte.
738 */
739 if (ALIGN_DOWN(addr, PMD_SIZE) == addr)
740 goto out;
741 ret = split_pmd(pmdp, pmd, GFP_PGTABLE_KERNEL, true);
742 if (ret)
743 goto out;
744 }
745
746 /*
747 * CONTPTE: If addr is CONTPTE aligned then addr already describes a
748 * leaf boundary. If not present then there is nothing to split.
749 * Otherwise, if we have a contpte leaf, split to pte.
750 */
751 if (ALIGN_DOWN(addr, CONT_PTE_SIZE) == addr)
752 goto out;
753 ptep = pte_offset_kernel(pmdp, addr);
754 pte = __ptep_get(ptep);
755 if (!pte_present(pte))
756 goto out;
757 if (pte_cont(pte))
758 split_contpte(ptep);
759
760 out:
761 return ret;
762 }
763
764 static inline bool force_pte_mapping(void)
765 {
766 const bool bbml2 = system_capabilities_finalized() ?
767 system_supports_bbml2_noabort() : cpu_supports_bbml2_noabort();
768
769 if (debug_pagealloc_enabled())
770 return true;
771 if (bbml2)
772 return false;
773 return rodata_full || arm64_kfence_can_set_direct_map() || is_realm_world();
774 }
775
776 static DEFINE_MUTEX(pgtable_split_lock);
777 static bool linear_map_requires_bbml2;
778
779 int split_kernel_leaf_mapping(unsigned long start, unsigned long end)
780 {
781 int ret;
782
783 /*
784 * If the region is within a pte-mapped area, there is no need to try to
785 * split. Additionally, CONFIG_DEBUG_PAGEALLOC and CONFIG_KFENCE may
786 * change permissions from atomic context so for those cases (which are
787 * always pte-mapped), we must not go any further because taking the
788 * mutex below may sleep. Do not call force_pte_mapping() here because
789 * it could return a confusing result if called from a secondary cpu
790 * prior to finalizing caps. Instead, linear_map_requires_bbml2 gives us
791 * what we need.
792 */
793 if (!linear_map_requires_bbml2 || is_kfence_address((void *)start))
794 return 0;
795
796 if (!system_supports_bbml2_noabort()) {
797 /*
798 * !BBML2_NOABORT systems should not be trying to change
799 * permissions on anything that is not pte-mapped in the first
800 * place. Just return early and let the permission change code
801 * raise a warning if not already pte-mapped.
802 */
803 if (system_capabilities_finalized())
804 return 0;
805
806 /*
807 * Boot-time: split_kernel_leaf_mapping_locked() allocates from
808 * page allocator. Can't split until it's available.
809 */
810 if (WARN_ON(!page_alloc_available))
811 return -EBUSY;
812
813 /*
814 * Boot-time: Started secondary cpus but don't know if they
815 * support BBML2_NOABORT yet. Can't allow splitting in this
816 * window in case they don't.
817 */
818 if (WARN_ON(num_online_cpus() > 1))
819 return -EBUSY;
820 }
821
822 /*
823 * Ensure start and end are at least page-aligned since this is the
824 * finest granularity we can split to.
825 */
826 if (start != PAGE_ALIGN(start) || end != PAGE_ALIGN(end))
827 return -EINVAL;
828
829 mutex_lock(&pgtable_split_lock);
830 arch_enter_lazy_mmu_mode();
831
832 /*
833 * The split_kernel_leaf_mapping_locked() may sleep, it is not a
834 * problem for ARM64 since ARM64's lazy MMU implementation allows
835 * sleeping.
836 *
837 * Optimize for the common case of splitting out a single page from a
838 * larger mapping. Here we can just split on the "least aligned" of
839 * start and end and this will guarantee that there must also be a split
840 * on the more aligned address since the both addresses must be in the
841 * same contpte block and it must have been split to ptes.
842 */
843 if (end - start == PAGE_SIZE) {
844 start = __ffs(start) < __ffs(end) ? start : end;
845 ret = split_kernel_leaf_mapping_locked(start);
846 } else {
847 ret = split_kernel_leaf_mapping_locked(start);
848 if (!ret)
849 ret = split_kernel_leaf_mapping_locked(end);
850 }
851
852 arch_leave_lazy_mmu_mode();
853 mutex_unlock(&pgtable_split_lock);
854 return ret;
855 }
856
857 static int split_to_ptes_pud_entry(pud_t *pudp, unsigned long addr,
858 unsigned long next, struct mm_walk *walk)
859 {
860 gfp_t gfp = *(gfp_t *)walk->private;
861 pud_t pud = pudp_get(pudp);
862 int ret = 0;
863
864 if (pud_leaf(pud))
865 ret = split_pud(pudp, pud, gfp, false);
866
867 return ret;
868 }
869
870 static int split_to_ptes_pmd_entry(pmd_t *pmdp, unsigned long addr,
871 unsigned long next, struct mm_walk *walk)
872 {
873 gfp_t gfp = *(gfp_t *)walk->private;
874 pmd_t pmd = pmdp_get(pmdp);
875 int ret = 0;
876
877 if (pmd_leaf(pmd)) {
878 if (pmd_cont(pmd))
879 split_contpmd(pmdp);
880 ret = split_pmd(pmdp, pmd, gfp, false);
881
882 /*
883 * We have split the pmd directly to ptes so there is no need to
884 * visit each pte to check if they are contpte.
885 */
886 walk->action = ACTION_CONTINUE;
887 }
888
889 return ret;
890 }
891
892 static int split_to_ptes_pte_entry(pte_t *ptep, unsigned long addr,
893 unsigned long next, struct mm_walk *walk)
894 {
895 pte_t pte = __ptep_get(ptep);
896
897 if (pte_cont(pte))
898 split_contpte(ptep);
899
900 return 0;
901 }
902
903 static const struct mm_walk_ops split_to_ptes_ops = {
904 .pud_entry = split_to_ptes_pud_entry,
905 .pmd_entry = split_to_ptes_pmd_entry,
906 .pte_entry = split_to_ptes_pte_entry,
907 };
908
909 static int range_split_to_ptes(unsigned long start, unsigned long end, gfp_t gfp)
910 {
911 int ret;
912
913 arch_enter_lazy_mmu_mode();
914 ret = walk_kernel_page_table_range_lockless(start, end,
915 &split_to_ptes_ops, NULL, &gfp);
916 arch_leave_lazy_mmu_mode();
917
918 return ret;
919 }
920
921 u32 idmap_kpti_bbml2_flag;
922
923 static void __init init_idmap_kpti_bbml2_flag(void)
924 {
925 WRITE_ONCE(idmap_kpti_bbml2_flag, 1);
926 /* Must be visible to other CPUs before stop_machine() is called. */
927 smp_mb();
928 }
929
930 static int __init linear_map_split_to_ptes(void *__unused)
931 {
932 /*
933 * Repainting the linear map must be done by CPU0 (the boot CPU) because
934 * that's the only CPU that we know supports BBML2. The other CPUs will
935 * be held in a waiting area with the idmap active.
936 */
937 if (!smp_processor_id()) {
938 unsigned long lstart = _PAGE_OFFSET(vabits_actual);
939 unsigned long lend = PAGE_END;
940 unsigned long kstart = (unsigned long)lm_alias(_stext);
941 unsigned long kend = (unsigned long)lm_alias(__init_begin);
942 int ret;
943
944 /*
945 * Wait for all secondary CPUs to be put into the waiting area.
946 */
947 smp_cond_load_acquire(&idmap_kpti_bbml2_flag, VAL == num_online_cpus());
948
949 /*
950 * Walk all of the linear map [lstart, lend), except the kernel
951 * linear map alias [kstart, kend), and split all mappings to
952 * PTE. The kernel alias remains static throughout runtime so
953 * can continue to be safely mapped with large mappings.
954 */
955 ret = range_split_to_ptes(lstart, kstart, GFP_ATOMIC);
956 if (!ret)
957 ret = range_split_to_ptes(kend, lend, GFP_ATOMIC);
958 if (ret)
959 panic("Failed to split linear map\n");
960 flush_tlb_kernel_range(lstart, lend);
961
962 /*
963 * Relies on dsb in flush_tlb_kernel_range() to avoid reordering
964 * before any page table split operations.
965 */
966 WRITE_ONCE(idmap_kpti_bbml2_flag, 0);
967 } else {
968 typedef void (wait_split_fn)(void);
969 extern wait_split_fn wait_linear_map_split_to_ptes;
970 wait_split_fn *wait_fn;
971
972 wait_fn = (void *)__pa_symbol(wait_linear_map_split_to_ptes);
973
974 /*
975 * At least one secondary CPU doesn't support BBML2 so cannot
976 * tolerate the size of the live mappings changing. So have the
977 * secondary CPUs wait for the boot CPU to make the changes
978 * with the idmap active and init_mm inactive.
979 */
980 cpu_install_idmap();
981 wait_fn();
982 cpu_uninstall_idmap();
983 }
984
985 return 0;
986 }
987
988 void __init linear_map_maybe_split_to_ptes(void)
989 {
990 if (linear_map_requires_bbml2 && !system_supports_bbml2_noabort()) {
991 init_idmap_kpti_bbml2_flag();
992 stop_machine(linear_map_split_to_ptes, NULL, cpu_online_mask);
993 }
994 }
995
996 /*
997 * This function can only be used to modify existing table entries,
998 * without allocating new levels of table. Note that this permits the
999 * creation of new section or page entries.
1000 */
1001 void __init create_mapping_noalloc(phys_addr_t phys, unsigned long virt,
1002 phys_addr_t size, pgprot_t prot)
1003 {
1004 if (virt < PAGE_OFFSET) {
1005 pr_warn("BUG: not creating mapping for %pa at 0x%016lx - outside kernel range\n",
1006 &phys, virt);
1007 return;
1008 }
1009 early_create_pgd_mapping(init_mm.pgd, phys, virt, size, prot, NULL,
1010 NO_CONT_MAPPINGS);
1011 }
1012
1013 void __init create_pgd_mapping(struct mm_struct *mm, phys_addr_t phys,
1014 unsigned long virt, phys_addr_t size,
1015 pgprot_t prot, bool page_mappings_only)
1016 {
1017 int flags = 0;
1018
1019 BUG_ON(mm == &init_mm);
1020
1021 if (page_mappings_only)
1022 flags = NO_BLOCK_MAPPINGS | NO_CONT_MAPPINGS;
1023
1024 early_create_pgd_mapping(mm->pgd, phys, virt, size, prot,
1025 pgd_pgtable_alloc_special_mm, flags);
1026 }
1027
1028 static void update_mapping_prot(phys_addr_t phys, unsigned long virt,
1029 phys_addr_t size, pgprot_t prot)
1030 {
1031 if (virt < PAGE_OFFSET) {
1032 pr_warn("BUG: not updating mapping for %pa at 0x%016lx - outside kernel range\n",
1033 &phys, virt);
1034 return;
1035 }
1036
1037 early_create_pgd_mapping(init_mm.pgd, phys, virt, size, prot, NULL,
1038 NO_CONT_MAPPINGS);
1039
1040 /* flush the TLBs after updating live kernel mappings */
1041 flush_tlb_kernel_range(virt, virt + size);
1042 }
1043
1044 static void __init __map_memblock(pgd_t *pgdp, phys_addr_t start,
1045 phys_addr_t end, pgprot_t prot, int flags)
1046 {
1047 early_create_pgd_mapping(pgdp, start, __phys_to_virt(start), end - start,
1048 prot, early_pgtable_alloc, flags);
1049 }
1050
1051 void __init mark_linear_text_alias_ro(void)
1052 {
1053 /*
1054 * Remove the write permissions from the linear alias of .text/.rodata
1055 */
1056 update_mapping_prot(__pa_symbol(_text), (unsigned long)lm_alias(_text),
1057 (unsigned long)__init_begin - (unsigned long)_text,
1058 PAGE_KERNEL_RO);
1059 }
1060
1061 #ifdef CONFIG_KFENCE
1062
1063 bool __ro_after_init kfence_early_init = !!CONFIG_KFENCE_SAMPLE_INTERVAL;
1064
1065 /* early_param() will be parsed before map_mem() below. */
1066 static int __init parse_kfence_early_init(char *arg)
1067 {
1068 int val;
1069
1070 if (get_option(&arg, &val))
1071 kfence_early_init = !!val;
1072 return 0;
1073 }
1074 early_param("kfence.sample_interval", parse_kfence_early_init);
1075
1076 static phys_addr_t __init arm64_kfence_alloc_pool(void)
1077 {
1078 phys_addr_t kfence_pool;
1079
1080 if (!kfence_early_init)
1081 return 0;
1082
1083 kfence_pool = memblock_phys_alloc(KFENCE_POOL_SIZE, PAGE_SIZE);
1084 if (!kfence_pool) {
1085 pr_err("failed to allocate kfence pool\n");
1086 kfence_early_init = false;
1087 return 0;
1088 }
1089
1090 /* Temporarily mark as NOMAP. */
1091 memblock_mark_nomap(kfence_pool, KFENCE_POOL_SIZE);
1092
1093 return kfence_pool;
1094 }
1095
1096 static void __init arm64_kfence_map_pool(phys_addr_t kfence_pool, pgd_t *pgdp)
1097 {
1098 if (!kfence_pool)
1099 return;
1100
1101 /* KFENCE pool needs page-level mapping. */
1102 __map_memblock(pgdp, kfence_pool, kfence_pool + KFENCE_POOL_SIZE,
1103 pgprot_tagged(PAGE_KERNEL),
1104 NO_BLOCK_MAPPINGS | NO_CONT_MAPPINGS);
1105 memblock_clear_nomap(kfence_pool, KFENCE_POOL_SIZE);
1106 __kfence_pool = phys_to_virt(kfence_pool);
1107 }
1108
1109 bool arch_kfence_init_pool(void)
1110 {
1111 unsigned long start = (unsigned long)__kfence_pool;
1112 unsigned long end = start + KFENCE_POOL_SIZE;
1113 int ret;
1114
1115 /* Exit early if we know the linear map is already pte-mapped. */
1116 if (force_pte_mapping())
1117 return true;
1118
1119 /* Kfence pool is already pte-mapped for the early init case. */
1120 if (kfence_early_init)
1121 return true;
1122
1123 mutex_lock(&pgtable_split_lock);
1124 ret = range_split_to_ptes(start, end, GFP_PGTABLE_KERNEL);
1125 mutex_unlock(&pgtable_split_lock);
1126
1127 /*
1128 * Since the system supports bbml2_noabort, tlb invalidation is not
1129 * required here; the pgtable mappings have been split to pte but larger
1130 * entries may safely linger in the TLB.
1131 */
1132
1133 return !ret;
1134 }
1135 #else /* CONFIG_KFENCE */
1136
1137 static inline phys_addr_t arm64_kfence_alloc_pool(void) { return 0; }
1138 static inline void arm64_kfence_map_pool(phys_addr_t kfence_pool, pgd_t *pgdp) { }
1139
1140 #endif /* CONFIG_KFENCE */
1141
1142 static void __init map_mem(pgd_t *pgdp)
1143 {
1144 static const u64 direct_map_end = _PAGE_END(VA_BITS_MIN);
1145 phys_addr_t kernel_start = __pa_symbol(_text);
1146 phys_addr_t kernel_end = __pa_symbol(__init_begin);
1147 phys_addr_t start, end;
1148 phys_addr_t early_kfence_pool;
1149 int flags = NO_EXEC_MAPPINGS;
1150 u64 i;
1151
1152 /*
1153 * Setting hierarchical PXNTable attributes on table entries covering
1154 * the linear region is only possible if it is guaranteed that no table
1155 * entries at any level are being shared between the linear region and
1156 * the vmalloc region. Check whether this is true for the PGD level, in
1157 * which case it is guaranteed to be true for all other levels as well.
1158 * (Unless we are running with support for LPA2, in which case the
1159 * entire reduced VA space is covered by a single pgd_t which will have
1160 * been populated without the PXNTable attribute by the time we get here.)
1161 */
1162 BUILD_BUG_ON(pgd_index(direct_map_end - 1) == pgd_index(direct_map_end) &&
1163 pgd_index(_PAGE_OFFSET(VA_BITS_MIN)) != PTRS_PER_PGD - 1);
1164
1165 early_kfence_pool = arm64_kfence_alloc_pool();
1166
1167 linear_map_requires_bbml2 = !force_pte_mapping() && can_set_direct_map();
1168
1169 if (force_pte_mapping())
1170 flags |= NO_BLOCK_MAPPINGS | NO_CONT_MAPPINGS;
1171
1172 /*
1173 * Take care not to create a writable alias for the
1174 * read-only text and rodata sections of the kernel image.
1175 * So temporarily mark them as NOMAP to skip mappings in
1176 * the following for-loop
1177 */
1178 memblock_mark_nomap(kernel_start, kernel_end - kernel_start);
1179
1180 /* map all the memory banks */
1181 for_each_mem_range(i, &start, &end) {
1182 if (start >= end)
1183 break;
1184 /*
1185 * The linear map must allow allocation tags reading/writing
1186 * if MTE is present. Otherwise, it has the same attributes as
1187 * PAGE_KERNEL.
1188 */
1189 __map_memblock(pgdp, start, end, pgprot_tagged(PAGE_KERNEL),
1190 flags);
1191 }
1192
1193 /*
1194 * Map the linear alias of the [_text, __init_begin) interval
1195 * as non-executable now, and remove the write permission in
1196 * mark_linear_text_alias_ro() below (which will be called after
1197 * alternative patching has completed). This makes the contents
1198 * of the region accessible to subsystems such as hibernate,
1199 * but protects it from inadvertent modification or execution.
1200 * Note that contiguous mappings cannot be remapped in this way,
1201 * so we should avoid them here.
1202 */
1203 __map_memblock(pgdp, kernel_start, kernel_end,
1204 PAGE_KERNEL, NO_CONT_MAPPINGS);
1205 memblock_clear_nomap(kernel_start, kernel_end - kernel_start);
1206 arm64_kfence_map_pool(early_kfence_pool, pgdp);
1207 }
1208
1209 void mark_rodata_ro(void)
1210 {
1211 unsigned long section_size;
1212
1213 /*
1214 * mark .rodata as read only. Use __init_begin rather than __end_rodata
1215 * to cover NOTES and EXCEPTION_TABLE.
1216 */
1217 section_size = (unsigned long)__init_begin - (unsigned long)__start_rodata;
1218 WRITE_ONCE(rodata_is_rw, false);
1219 update_mapping_prot(__pa_symbol(__start_rodata), (unsigned long)__start_rodata,
1220 section_size, PAGE_KERNEL_RO);
1221 /* mark the range between _text and _stext as read only. */
1222 update_mapping_prot(__pa_symbol(_text), (unsigned long)_text,
1223 (unsigned long)_stext - (unsigned long)_text,
1224 PAGE_KERNEL_RO);
1225 }
1226
1227 static void __init declare_vma(struct vm_struct *vma,
1228 void *va_start, void *va_end,
1229 unsigned long vm_flags)
1230 {
1231 phys_addr_t pa_start = __pa_symbol(va_start);
1232 unsigned long size = va_end - va_start;
1233
1234 BUG_ON(!PAGE_ALIGNED(pa_start));
1235 BUG_ON(!PAGE_ALIGNED(size));
1236
1237 if (!(vm_flags & VM_NO_GUARD))
1238 size += PAGE_SIZE;
1239
1240 vma->addr = va_start;
1241 vma->phys_addr = pa_start;
1242 vma->size = size;
1243 vma->flags = VM_MAP | vm_flags;
1244 vma->caller = __builtin_return_address(0);
1245
1246 vm_area_add_early(vma);
1247 }
1248
1249 #ifdef CONFIG_UNMAP_KERNEL_AT_EL0
1250 #define KPTI_NG_TEMP_VA (-(1UL << PMD_SHIFT))
1251
1252 static phys_addr_t kpti_ng_temp_alloc __initdata;
1253
1254 static phys_addr_t __init kpti_ng_pgd_alloc(enum pgtable_type type)
1255 {
1256 kpti_ng_temp_alloc -= PAGE_SIZE;
1257 return kpti_ng_temp_alloc;
1258 }
1259
1260 static int __init __kpti_install_ng_mappings(void *__unused)
1261 {
1262 typedef void (kpti_remap_fn)(int, int, phys_addr_t, unsigned long);
1263 extern kpti_remap_fn idmap_kpti_install_ng_mappings;
1264 kpti_remap_fn *remap_fn;
1265
1266 int cpu = smp_processor_id();
1267 int levels = CONFIG_PGTABLE_LEVELS;
1268 int order = order_base_2(levels);
1269 u64 kpti_ng_temp_pgd_pa = 0;
1270 pgd_t *kpti_ng_temp_pgd;
1271 u64 alloc = 0;
1272
1273 if (levels == 5 && !pgtable_l5_enabled())
1274 levels = 4;
1275 else if (levels == 4 && !pgtable_l4_enabled())
1276 levels = 3;
1277
1278 remap_fn = (void *)__pa_symbol(idmap_kpti_install_ng_mappings);
1279
1280 if (!cpu) {
1281 int ret;
1282
1283 alloc = __get_free_pages(GFP_ATOMIC | __GFP_ZERO, order);
1284 kpti_ng_temp_pgd = (pgd_t *)(alloc + (levels - 1) * PAGE_SIZE);
1285 kpti_ng_temp_alloc = kpti_ng_temp_pgd_pa = __pa(kpti_ng_temp_pgd);
1286
1287 //
1288 // Create a minimal page table hierarchy that permits us to map
1289 // the swapper page tables temporarily as we traverse them.
1290 //
1291 // The physical pages are laid out as follows:
1292 //
1293 // +--------+-/-------+-/------ +-/------ +-\\\--------+
1294 // : PTE[] : | PMD[] : | PUD[] : | P4D[] : ||| PGD[] :
1295 // +--------+-\-------+-\------ +-\------ +-///--------+
1296 // ^
1297 // The first page is mapped into this hierarchy at a PMD_SHIFT
1298 // aligned virtual address, so that we can manipulate the PTE
1299 // level entries while the mapping is active. The first entry
1300 // covers the PTE[] page itself, the remaining entries are free
1301 // to be used as a ad-hoc fixmap.
1302 //
1303 ret = __create_pgd_mapping_locked(kpti_ng_temp_pgd, __pa(alloc),
1304 KPTI_NG_TEMP_VA, PAGE_SIZE, PAGE_KERNEL,
1305 kpti_ng_pgd_alloc, 0);
1306 if (ret)
1307 panic("Failed to create page tables\n");
1308 }
1309
1310 cpu_install_idmap();
1311 remap_fn(cpu, num_online_cpus(), kpti_ng_temp_pgd_pa, KPTI_NG_TEMP_VA);
1312 cpu_uninstall_idmap();
1313
1314 if (!cpu) {
1315 free_pages(alloc, order);
1316 arm64_use_ng_mappings = true;
1317 }
1318
1319 return 0;
1320 }
1321
1322 void __init kpti_install_ng_mappings(void)
1323 {
1324 /* Check whether KPTI is going to be used */
1325 if (!arm64_kernel_unmapped_at_el0())
1326 return;
1327
1328 /*
1329 * We don't need to rewrite the page-tables if either we've done
1330 * it already or we have KASLR enabled and therefore have not
1331 * created any global mappings at all.
1332 */
1333 if (arm64_use_ng_mappings)
1334 return;
1335
1336 init_idmap_kpti_bbml2_flag();
1337 stop_machine(__kpti_install_ng_mappings, NULL, cpu_online_mask);
1338 }
1339
1340 static pgprot_t __init kernel_exec_prot(void)
1341 {
1342 return rodata_enabled ? PAGE_KERNEL_ROX : PAGE_KERNEL_EXEC;
1343 }
1344
1345 static int __init map_entry_trampoline(void)
1346 {
1347 int i;
1348
1349 if (!arm64_kernel_unmapped_at_el0())
1350 return 0;
1351
1352 pgprot_t prot = kernel_exec_prot();
1353 phys_addr_t pa_start = __pa_symbol(__entry_tramp_text_start);
1354
1355 /* The trampoline is always mapped and can therefore be global */
1356 pgprot_val(prot) &= ~PTE_NG;
1357
1358 /* Map only the text into the trampoline page table */
1359 memset(tramp_pg_dir, 0, PGD_SIZE);
1360 early_create_pgd_mapping(tramp_pg_dir, pa_start, TRAMP_VALIAS,
1361 entry_tramp_text_size(), prot,
1362 pgd_pgtable_alloc_init_mm, NO_BLOCK_MAPPINGS);
1363
1364 /* Map both the text and data into the kernel page table */
1365 for (i = 0; i < DIV_ROUND_UP(entry_tramp_text_size(), PAGE_SIZE); i++)
1366 __set_fixmap(FIX_ENTRY_TRAMP_TEXT1 - i,
1367 pa_start + i * PAGE_SIZE, prot);
1368
1369 if (IS_ENABLED(CONFIG_RELOCATABLE))
1370 __set_fixmap(FIX_ENTRY_TRAMP_TEXT1 - i,
1371 pa_start + i * PAGE_SIZE, PAGE_KERNEL_RO);
1372
1373 return 0;
1374 }
1375 core_initcall(map_entry_trampoline);
1376 #endif
1377
1378 /*
1379 * Declare the VMA areas for the kernel
1380 */
1381 static void __init declare_kernel_vmas(void)
1382 {
1383 static struct vm_struct vmlinux_seg[KERNEL_SEGMENT_COUNT];
1384
1385 declare_vma(&vmlinux_seg[0], _text, _etext, VM_NO_GUARD);
1386 declare_vma(&vmlinux_seg[1], __start_rodata, __inittext_begin, VM_NO_GUARD);
1387 declare_vma(&vmlinux_seg[2], __inittext_begin, __inittext_end, VM_NO_GUARD);
1388 declare_vma(&vmlinux_seg[3], __initdata_begin, __initdata_end, VM_NO_GUARD);
1389 declare_vma(&vmlinux_seg[4], _data, _end, 0);
1390 }
1391
1392 void __pi_map_range(phys_addr_t *pte, u64 start, u64 end, phys_addr_t pa,
1393 pgprot_t prot, int level, pte_t *tbl, bool may_use_cont,
1394 u64 va_offset);
1395
1396 static u8 idmap_ptes[IDMAP_LEVELS - 1][PAGE_SIZE] __aligned(PAGE_SIZE) __ro_after_init,
1397 kpti_bbml2_ptes[IDMAP_LEVELS - 1][PAGE_SIZE] __aligned(PAGE_SIZE) __ro_after_init;
1398
1399 static void __init create_idmap(void)
1400 {
1401 phys_addr_t start = __pa_symbol(__idmap_text_start);
1402 phys_addr_t end = __pa_symbol(__idmap_text_end);
1403 phys_addr_t ptep = __pa_symbol(idmap_ptes);
1404
1405 __pi_map_range(&ptep, start, end, start, PAGE_KERNEL_ROX,
1406 IDMAP_ROOT_LEVEL, (pte_t *)idmap_pg_dir, false,
1407 __phys_to_virt(ptep) - ptep);
1408
1409 if (linear_map_requires_bbml2 ||
1410 (IS_ENABLED(CONFIG_UNMAP_KERNEL_AT_EL0) && !arm64_use_ng_mappings)) {
1411 phys_addr_t pa = __pa_symbol(&idmap_kpti_bbml2_flag);
1412
1413 /*
1414 * The KPTI G-to-nG conversion code needs a read-write mapping
1415 * of its synchronization flag in the ID map. This is also used
1416 * when splitting the linear map to ptes if a secondary CPU
1417 * doesn't support bbml2.
1418 */
1419 ptep = __pa_symbol(kpti_bbml2_ptes);
1420 __pi_map_range(&ptep, pa, pa + sizeof(u32), pa, PAGE_KERNEL,
1421 IDMAP_ROOT_LEVEL, (pte_t *)idmap_pg_dir, false,
1422 __phys_to_virt(ptep) - ptep);
1423 }
1424 }
1425
1426 void __init paging_init(void)
1427 {
1428 map_mem(swapper_pg_dir);
1429
1430 memblock_allow_resize();
1431
1432 create_idmap();
1433 declare_kernel_vmas();
1434 }
1435
1436 #ifdef CONFIG_MEMORY_HOTPLUG
1437 static void free_hotplug_page_range(struct page *page, size_t size,
1438 struct vmem_altmap *altmap)
1439 {
1440 if (altmap) {
1441 vmem_altmap_free(altmap, size >> PAGE_SHIFT);
1442 } else {
1443 WARN_ON(PageReserved(page));
1444 __free_pages(page, get_order(size));
1445 }
1446 }
1447
1448 static void free_hotplug_pgtable_page(struct page *page)
1449 {
1450 pagetable_dtor(page_ptdesc(page));
1451 free_hotplug_page_range(page, PAGE_SIZE, NULL);
1452 }
1453
1454 static bool pgtable_range_aligned(unsigned long start, unsigned long end,
1455 unsigned long floor, unsigned long ceiling,
1456 unsigned long mask)
1457 {
1458 start &= mask;
1459 if (start < floor)
1460 return false;
1461
1462 if (ceiling) {
1463 ceiling &= mask;
1464 if (!ceiling)
1465 return false;
1466 }
1467
1468 if (end - 1 > ceiling - 1)
1469 return false;
1470 return true;
1471 }
1472
1473 static void unmap_hotplug_pte_range(pmd_t *pmdp, unsigned long addr,
1474 unsigned long end, bool free_mapped,
1475 struct vmem_altmap *altmap)
1476 {
1477 pte_t *ptep, pte;
1478
1479 do {
1480 ptep = pte_offset_kernel(pmdp, addr);
1481 pte = __ptep_get(ptep);
1482 if (pte_none(pte))
1483 continue;
1484
1485 WARN_ON(!pte_present(pte));
1486 __pte_clear(&init_mm, addr, ptep);
1487 if (free_mapped) {
1488 /* CONT blocks are not supported in the vmemmap */
1489 WARN_ON(pte_cont(pte));
1490 flush_tlb_kernel_range(addr, addr + PAGE_SIZE);
1491 free_hotplug_page_range(pte_page(pte),
1492 PAGE_SIZE, altmap);
1493 }
1494 /* unmap_hotplug_range() flushes TLB for !free_mapped */
1495 } while (addr += PAGE_SIZE, addr < end);
1496 }
1497
1498 static void unmap_hotplug_pmd_range(pud_t *pudp, unsigned long addr,
1499 unsigned long end, bool free_mapped,
1500 struct vmem_altmap *altmap)
1501 {
1502 unsigned long next;
1503 pmd_t *pmdp, pmd;
1504
1505 do {
1506 next = pmd_addr_end(addr, end);
1507 pmdp = pmd_offset(pudp, addr);
1508 pmd = READ_ONCE(*pmdp);
1509 if (pmd_none(pmd))
1510 continue;
1511
1512 WARN_ON(!pmd_present(pmd));
1513 if (pmd_sect(pmd)) {
1514 pmd_clear(pmdp);
1515 if (free_mapped) {
1516 /* CONT blocks are not supported in the vmemmap */
1517 WARN_ON(pmd_cont(pmd));
1518 flush_tlb_kernel_range(addr, addr + PMD_SIZE);
1519 free_hotplug_page_range(pmd_page(pmd),
1520 PMD_SIZE, altmap);
1521 }
1522 /* unmap_hotplug_range() flushes TLB for !free_mapped */
1523 continue;
1524 }
1525 WARN_ON(!pmd_table(pmd));
1526 unmap_hotplug_pte_range(pmdp, addr, next, free_mapped, altmap);
1527 } while (addr = next, addr < end);
1528 }
1529
1530 static void unmap_hotplug_pud_range(p4d_t *p4dp, unsigned long addr,
1531 unsigned long end, bool free_mapped,
1532 struct vmem_altmap *altmap)
1533 {
1534 unsigned long next;
1535 pud_t *pudp, pud;
1536
1537 do {
1538 next = pud_addr_end(addr, end);
1539 pudp = pud_offset(p4dp, addr);
1540 pud = READ_ONCE(*pudp);
1541 if (pud_none(pud))
1542 continue;
1543
1544 WARN_ON(!pud_present(pud));
1545 if (pud_sect(pud)) {
1546 pud_clear(pudp);
1547 if (free_mapped) {
1548 flush_tlb_kernel_range(addr, addr + PUD_SIZE);
1549 free_hotplug_page_range(pud_page(pud),
1550 PUD_SIZE, altmap);
1551 }
1552 /* unmap_hotplug_range() flushes TLB for !free_mapped */
1553 continue;
1554 }
1555 WARN_ON(!pud_table(pud));
1556 unmap_hotplug_pmd_range(pudp, addr, next, free_mapped, altmap);
1557 } while (addr = next, addr < end);
1558 }
1559
1560 static void unmap_hotplug_p4d_range(pgd_t *pgdp, unsigned long addr,
1561 unsigned long end, bool free_mapped,
1562 struct vmem_altmap *altmap)
1563 {
1564 unsigned long next;
1565 p4d_t *p4dp, p4d;
1566
1567 do {
1568 next = p4d_addr_end(addr, end);
1569 p4dp = p4d_offset(pgdp, addr);
1570 p4d = READ_ONCE(*p4dp);
1571 if (p4d_none(p4d))
1572 continue;
1573
1574 WARN_ON(!p4d_present(p4d));
1575 unmap_hotplug_pud_range(p4dp, addr, next, free_mapped, altmap);
1576 } while (addr = next, addr < end);
1577 }
1578
1579 static void unmap_hotplug_range(unsigned long addr, unsigned long end,
1580 bool free_mapped, struct vmem_altmap *altmap)
1581 {
1582 unsigned long start = addr;
1583 unsigned long next;
1584 pgd_t *pgdp, pgd;
1585
1586 /*
1587 * altmap can only be used as vmemmap mapping backing memory.
1588 * In case the backing memory itself is not being freed, then
1589 * altmap is irrelevant. Warn about this inconsistency when
1590 * encountered.
1591 */
1592 WARN_ON(!free_mapped && altmap);
1593
1594 do {
1595 next = pgd_addr_end(addr, end);
1596 pgdp = pgd_offset_k(addr);
1597 pgd = READ_ONCE(*pgdp);
1598 if (pgd_none(pgd))
1599 continue;
1600
1601 WARN_ON(!pgd_present(pgd));
1602 unmap_hotplug_p4d_range(pgdp, addr, next, free_mapped, altmap);
1603 } while (addr = next, addr < end);
1604
1605 if (!free_mapped)
1606 flush_tlb_kernel_range(start, end);
1607 }
1608
1609 static void free_empty_pte_table(pmd_t *pmdp, unsigned long addr,
1610 unsigned long end, unsigned long floor,
1611 unsigned long ceiling)
1612 {
1613 pte_t *ptep, pte;
1614 unsigned long i, start = addr;
1615
1616 do {
1617 ptep = pte_offset_kernel(pmdp, addr);
1618 pte = __ptep_get(ptep);
1619
1620 /*
1621 * This is just a sanity check here which verifies that
1622 * pte clearing has been done by earlier unmap loops.
1623 */
1624 WARN_ON(!pte_none(pte));
1625 } while (addr += PAGE_SIZE, addr < end);
1626
1627 if (!pgtable_range_aligned(start, end, floor, ceiling, PMD_MASK))
1628 return;
1629
1630 /*
1631 * Check whether we can free the pte page if the rest of the
1632 * entries are empty. Overlap with other regions have been
1633 * handled by the floor/ceiling check.
1634 */
1635 ptep = pte_offset_kernel(pmdp, 0UL);
1636 for (i = 0; i < PTRS_PER_PTE; i++) {
1637 if (!pte_none(__ptep_get(&ptep[i])))
1638 return;
1639 }
1640
1641 pmd_clear(pmdp);
1642 __flush_tlb_kernel_pgtable(start);
1643 free_hotplug_pgtable_page(virt_to_page(ptep));
1644 }
1645
1646 static void free_empty_pmd_table(pud_t *pudp, unsigned long addr,
1647 unsigned long end, unsigned long floor,
1648 unsigned long ceiling)
1649 {
1650 pmd_t *pmdp, pmd;
1651 unsigned long i, next, start = addr;
1652
1653 do {
1654 next = pmd_addr_end(addr, end);
1655 pmdp = pmd_offset(pudp, addr);
1656 pmd = READ_ONCE(*pmdp);
1657 if (pmd_none(pmd))
1658 continue;
1659
1660 WARN_ON(!pmd_present(pmd) || !pmd_table(pmd) || pmd_sect(pmd));
1661 free_empty_pte_table(pmdp, addr, next, floor, ceiling);
1662 } while (addr = next, addr < end);
1663
1664 if (CONFIG_PGTABLE_LEVELS <= 2)
1665 return;
1666
1667 if (!pgtable_range_aligned(start, end, floor, ceiling, PUD_MASK))
1668 return;
1669
1670 /*
1671 * Check whether we can free the pmd page if the rest of the
1672 * entries are empty. Overlap with other regions have been
1673 * handled by the floor/ceiling check.
1674 */
1675 pmdp = pmd_offset(pudp, 0UL);
1676 for (i = 0; i < PTRS_PER_PMD; i++) {
1677 if (!pmd_none(READ_ONCE(pmdp[i])))
1678 return;
1679 }
1680
1681 pud_clear(pudp);
1682 __flush_tlb_kernel_pgtable(start);
1683 free_hotplug_pgtable_page(virt_to_page(pmdp));
1684 }
1685
1686 static void free_empty_pud_table(p4d_t *p4dp, unsigned long addr,
1687 unsigned long end, unsigned long floor,
1688 unsigned long ceiling)
1689 {
1690 pud_t *pudp, pud;
1691 unsigned long i, next, start = addr;
1692
1693 do {
1694 next = pud_addr_end(addr, end);
1695 pudp = pud_offset(p4dp, addr);
1696 pud = READ_ONCE(*pudp);
1697 if (pud_none(pud))
1698 continue;
1699
1700 WARN_ON(!pud_present(pud) || !pud_table(pud) || pud_sect(pud));
1701 free_empty_pmd_table(pudp, addr, next, floor, ceiling);
1702 } while (addr = next, addr < end);
1703
1704 if (!pgtable_l4_enabled())
1705 return;
1706
1707 if (!pgtable_range_aligned(start, end, floor, ceiling, P4D_MASK))
1708 return;
1709
1710 /*
1711 * Check whether we can free the pud page if the rest of the
1712 * entries are empty. Overlap with other regions have been
1713 * handled by the floor/ceiling check.
1714 */
1715 pudp = pud_offset(p4dp, 0UL);
1716 for (i = 0; i < PTRS_PER_PUD; i++) {
1717 if (!pud_none(READ_ONCE(pudp[i])))
1718 return;
1719 }
1720
1721 p4d_clear(p4dp);
1722 __flush_tlb_kernel_pgtable(start);
1723 free_hotplug_pgtable_page(virt_to_page(pudp));
1724 }
1725
1726 static void free_empty_p4d_table(pgd_t *pgdp, unsigned long addr,
1727 unsigned long end, unsigned long floor,
1728 unsigned long ceiling)
1729 {
1730 p4d_t *p4dp, p4d;
1731 unsigned long i, next, start = addr;
1732
1733 do {
1734 next = p4d_addr_end(addr, end);
1735 p4dp = p4d_offset(pgdp, addr);
1736 p4d = READ_ONCE(*p4dp);
1737 if (p4d_none(p4d))
1738 continue;
1739
1740 WARN_ON(!p4d_present(p4d));
1741 free_empty_pud_table(p4dp, addr, next, floor, ceiling);
1742 } while (addr = next, addr < end);
1743
1744 if (!pgtable_l5_enabled())
1745 return;
1746
1747 if (!pgtable_range_aligned(start, end, floor, ceiling, PGDIR_MASK))
1748 return;
1749
1750 /*
1751 * Check whether we can free the p4d page if the rest of the
1752 * entries are empty. Overlap with other regions have been
1753 * handled by the floor/ceiling check.
1754 */
1755 p4dp = p4d_offset(pgdp, 0UL);
1756 for (i = 0; i < PTRS_PER_P4D; i++) {
1757 if (!p4d_none(READ_ONCE(p4dp[i])))
1758 return;
1759 }
1760
1761 pgd_clear(pgdp);
1762 __flush_tlb_kernel_pgtable(start);
1763 free_hotplug_pgtable_page(virt_to_page(p4dp));
1764 }
1765
1766 static void free_empty_tables(unsigned long addr, unsigned long end,
1767 unsigned long floor, unsigned long ceiling)
1768 {
1769 unsigned long next;
1770 pgd_t *pgdp, pgd;
1771
1772 do {
1773 next = pgd_addr_end(addr, end);
1774 pgdp = pgd_offset_k(addr);
1775 pgd = READ_ONCE(*pgdp);
1776 if (pgd_none(pgd))
1777 continue;
1778
1779 WARN_ON(!pgd_present(pgd));
1780 free_empty_p4d_table(pgdp, addr, next, floor, ceiling);
1781 } while (addr = next, addr < end);
1782 }
1783 #endif
1784
1785 void __meminit vmemmap_set_pmd(pmd_t *pmdp, void *p, int node,
1786 unsigned long addr, unsigned long next)
1787 {
1788 pmd_set_huge(pmdp, __pa(p), __pgprot(PROT_SECT_NORMAL));
1789 }
1790
1791 int __meminit vmemmap_check_pmd(pmd_t *pmdp, int node,
1792 unsigned long addr, unsigned long next)
1793 {
1794 vmemmap_verify((pte_t *)pmdp, node, addr, next);
1795
1796 return pmd_sect(READ_ONCE(*pmdp));
1797 }
1798
1799 int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node,
1800 struct vmem_altmap *altmap)
1801 {
1802 WARN_ON((start < VMEMMAP_START) || (end > VMEMMAP_END));
1803 /* [start, end] should be within one section */
1804 WARN_ON_ONCE(end - start > PAGES_PER_SECTION * sizeof(struct page));
1805
1806 if (!IS_ENABLED(CONFIG_ARM64_4K_PAGES) ||
1807 (end - start < PAGES_PER_SECTION * sizeof(struct page)))
1808 return vmemmap_populate_basepages(start, end, node, altmap);
1809 else
1810 return vmemmap_populate_hugepages(start, end, node, altmap);
1811 }
1812
1813 #ifdef CONFIG_MEMORY_HOTPLUG
1814 void vmemmap_free(unsigned long start, unsigned long end,
1815 struct vmem_altmap *altmap)
1816 {
1817 WARN_ON((start < VMEMMAP_START) || (end > VMEMMAP_END));
1818
1819 unmap_hotplug_range(start, end, true, altmap);
1820 free_empty_tables(start, end, VMEMMAP_START, VMEMMAP_END);
1821 }
1822 #endif /* CONFIG_MEMORY_HOTPLUG */
1823
1824 int pud_set_huge(pud_t *pudp, phys_addr_t phys, pgprot_t prot)
1825 {
1826 pud_t new_pud = pfn_pud(__phys_to_pfn(phys), mk_pud_sect_prot(prot));
1827
1828 /* Only allow permission changes for now */
1829 if (!pgattr_change_is_safe(READ_ONCE(pud_val(*pudp)),
1830 pud_val(new_pud)))
1831 return 0;
1832
1833 VM_BUG_ON(phys & ~PUD_MASK);
1834 set_pud(pudp, new_pud);
1835 return 1;
1836 }
1837
1838 int pmd_set_huge(pmd_t *pmdp, phys_addr_t phys, pgprot_t prot)
1839 {
1840 pmd_t new_pmd = pfn_pmd(__phys_to_pfn(phys), mk_pmd_sect_prot(prot));
1841
1842 /* Only allow permission changes for now */
1843 if (!pgattr_change_is_safe(READ_ONCE(pmd_val(*pmdp)),
1844 pmd_val(new_pmd)))
1845 return 0;
1846
1847 VM_BUG_ON(phys & ~PMD_MASK);
1848 set_pmd(pmdp, new_pmd);
1849 return 1;
1850 }
1851
1852 #ifndef __PAGETABLE_P4D_FOLDED
1853 void p4d_clear_huge(p4d_t *p4dp)
1854 {
1855 }
1856 #endif
1857
1858 int pud_clear_huge(pud_t *pudp)
1859 {
1860 if (!pud_sect(READ_ONCE(*pudp)))
1861 return 0;
1862 pud_clear(pudp);
1863 return 1;
1864 }
1865
1866 int pmd_clear_huge(pmd_t *pmdp)
1867 {
1868 if (!pmd_sect(READ_ONCE(*pmdp)))
1869 return 0;
1870 pmd_clear(pmdp);
1871 return 1;
1872 }
1873
1874 static int __pmd_free_pte_page(pmd_t *pmdp, unsigned long addr,
1875 bool acquire_mmap_lock)
1876 {
1877 pte_t *table;
1878 pmd_t pmd;
1879
1880 pmd = READ_ONCE(*pmdp);
1881
1882 if (!pmd_table(pmd)) {
1883 VM_WARN_ON(1);
1884 return 1;
1885 }
1886
1887 /* See comment in pud_free_pmd_page for static key logic */
1888 table = pte_offset_kernel(pmdp, addr);
1889 pmd_clear(pmdp);
1890 __flush_tlb_kernel_pgtable(addr);
1891 if (static_branch_unlikely(&arm64_ptdump_lock_key) && acquire_mmap_lock) {
1892 mmap_read_lock(&init_mm);
1893 mmap_read_unlock(&init_mm);
1894 }
1895
1896 pte_free_kernel(NULL, table);
1897 return 1;
1898 }
1899
1900 int pmd_free_pte_page(pmd_t *pmdp, unsigned long addr)
1901 {
1902 /* If ptdump is walking the pagetables, acquire init_mm.mmap_lock */
1903 return __pmd_free_pte_page(pmdp, addr, /* acquire_mmap_lock = */ true);
1904 }
1905
1906 int pud_free_pmd_page(pud_t *pudp, unsigned long addr)
1907 {
1908 pmd_t *table;
1909 pmd_t *pmdp;
1910 pud_t pud;
1911 unsigned long next, end;
1912
1913 pud = READ_ONCE(*pudp);
1914
1915 if (!pud_table(pud)) {
1916 VM_WARN_ON(1);
1917 return 1;
1918 }
1919
1920 table = pmd_offset(pudp, addr);
1921
1922 /*
1923 * Our objective is to prevent ptdump from reading a PMD table which has
1924 * been freed. In this race, if pud_free_pmd_page observes the key on
1925 * (which got flipped by ptdump) then the mmap lock sequence here will,
1926 * as a result of the mmap write lock/unlock sequence in ptdump, give
1927 * us the correct synchronization. If not, this means that ptdump has
1928 * yet not started walking the pagetables - the sequence of barriers
1929 * issued by __flush_tlb_kernel_pgtable() guarantees that ptdump will
1930 * observe an empty PUD.
1931 */
1932 pud_clear(pudp);
1933 __flush_tlb_kernel_pgtable(addr);
1934 if (static_branch_unlikely(&arm64_ptdump_lock_key)) {
1935 mmap_read_lock(&init_mm);
1936 mmap_read_unlock(&init_mm);
1937 }
1938
1939 pmdp = table;
1940 next = addr;
1941 end = addr + PUD_SIZE;
1942 do {
1943 if (pmd_present(pmdp_get(pmdp)))
1944 /*
1945 * PMD has been isolated, so ptdump won't see it. No
1946 * need to acquire init_mm.mmap_lock.
1947 */
1948 __pmd_free_pte_page(pmdp, next, /* acquire_mmap_lock = */ false);
1949 } while (pmdp++, next += PMD_SIZE, next != end);
1950
1951 pmd_free(NULL, table);
1952 return 1;
1953 }
1954
1955 #ifdef CONFIG_MEMORY_HOTPLUG
1956 static void __remove_pgd_mapping(pgd_t *pgdir, unsigned long start, u64 size)
1957 {
1958 unsigned long end = start + size;
1959
1960 WARN_ON(pgdir != init_mm.pgd);
1961 WARN_ON((start < PAGE_OFFSET) || (end > PAGE_END));
1962
1963 unmap_hotplug_range(start, end, false, NULL);
1964 free_empty_tables(start, end, PAGE_OFFSET, PAGE_END);
1965 }
1966
1967 struct range arch_get_mappable_range(void)
1968 {
1969 struct range mhp_range;
1970 phys_addr_t start_linear_pa = __pa(_PAGE_OFFSET(vabits_actual));
1971 phys_addr_t end_linear_pa = __pa(PAGE_END - 1);
1972
1973 if (IS_ENABLED(CONFIG_RANDOMIZE_BASE)) {
1974 /*
1975 * Check for a wrap, it is possible because of randomized linear
1976 * mapping the start physical address is actually bigger than
1977 * the end physical address. In this case set start to zero
1978 * because [0, end_linear_pa] range must still be able to cover
1979 * all addressable physical addresses.
1980 */
1981 if (start_linear_pa > end_linear_pa)
1982 start_linear_pa = 0;
1983 }
1984
1985 WARN_ON(start_linear_pa > end_linear_pa);
1986
1987 /*
1988 * Linear mapping region is the range [PAGE_OFFSET..(PAGE_END - 1)]
1989 * accommodating both its ends but excluding PAGE_END. Max physical
1990 * range which can be mapped inside this linear mapping range, must
1991 * also be derived from its end points.
1992 */
1993 mhp_range.start = start_linear_pa;
1994 mhp_range.end = end_linear_pa;
1995
1996 return mhp_range;
1997 }
1998
1999 int arch_add_memory(int nid, u64 start, u64 size,
2000 struct mhp_params *params)
2001 {
2002 int ret, flags = NO_EXEC_MAPPINGS;
2003
2004 VM_BUG_ON(!mhp_range_allowed(start, size, true));
2005
2006 if (force_pte_mapping())
2007 flags |= NO_BLOCK_MAPPINGS | NO_CONT_MAPPINGS;
2008
2009 ret = __create_pgd_mapping(swapper_pg_dir, start, __phys_to_virt(start),
2010 size, params->pgprot, pgd_pgtable_alloc_init_mm,
2011 flags);
2012 if (ret)
2013 goto err;
2014
2015 memblock_clear_nomap(start, size);
2016
2017 ret = __add_pages(nid, start >> PAGE_SHIFT, size >> PAGE_SHIFT,
2018 params);
2019 if (ret)
2020 goto err;
2021
2022 /* Address of hotplugged memory can be smaller */
2023 max_pfn = max(max_pfn, PFN_UP(start + size));
2024 max_low_pfn = max_pfn;
2025
2026 return 0;
2027
2028 err:
2029 __remove_pgd_mapping(swapper_pg_dir,
2030 __phys_to_virt(start), size);
2031 return ret;
2032 }
2033
2034 void arch_remove_memory(u64 start, u64 size, struct vmem_altmap *altmap)
2035 {
2036 unsigned long start_pfn = start >> PAGE_SHIFT;
2037 unsigned long nr_pages = size >> PAGE_SHIFT;
2038
2039 __remove_pages(start_pfn, nr_pages, altmap);
2040 __remove_pgd_mapping(swapper_pg_dir, __phys_to_virt(start), size);
2041 }
2042
2043 /*
2044 * This memory hotplug notifier helps prevent boot memory from being
2045 * inadvertently removed as it blocks pfn range offlining process in
2046 * __offline_pages(). Hence this prevents both offlining as well as
2047 * removal process for boot memory which is initially always online.
2048 * In future if and when boot memory could be removed, this notifier
2049 * should be dropped and free_hotplug_page_range() should handle any
2050 * reserved pages allocated during boot.
2051 */
2052 static int prevent_bootmem_remove_notifier(struct notifier_block *nb,
2053 unsigned long action, void *data)
2054 {
2055 struct mem_section *ms;
2056 struct memory_notify *arg = data;
2057 unsigned long end_pfn = arg->start_pfn + arg->nr_pages;
2058 unsigned long pfn = arg->start_pfn;
2059
2060 if ((action != MEM_GOING_OFFLINE) && (action != MEM_OFFLINE))
2061 return NOTIFY_OK;
2062
2063 for (; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
2064 unsigned long start = PFN_PHYS(pfn);
2065 unsigned long end = start + (1UL << PA_SECTION_SHIFT);
2066
2067 ms = __pfn_to_section(pfn);
2068 if (!early_section(ms))
2069 continue;
2070
2071 if (action == MEM_GOING_OFFLINE) {
2072 /*
2073 * Boot memory removal is not supported. Prevent
2074 * it via blocking any attempted offline request
2075 * for the boot memory and just report it.
2076 */
2077 pr_warn("Boot memory [%lx %lx] offlining attempted\n", start, end);
2078 return NOTIFY_BAD;
2079 } else if (action == MEM_OFFLINE) {
2080 /*
2081 * This should have never happened. Boot memory
2082 * offlining should have been prevented by this
2083 * very notifier. Probably some memory removal
2084 * procedure might have changed which would then
2085 * require further debug.
2086 */
2087 pr_err("Boot memory [%lx %lx] offlined\n", start, end);
2088
2089 /*
2090 * Core memory hotplug does not process a return
2091 * code from the notifier for MEM_OFFLINE events.
2092 * The error condition has been reported. Return
2093 * from here as if ignored.
2094 */
2095 return NOTIFY_DONE;
2096 }
2097 }
2098 return NOTIFY_OK;
2099 }
2100
2101 static struct notifier_block prevent_bootmem_remove_nb = {
2102 .notifier_call = prevent_bootmem_remove_notifier,
2103 };
2104
2105 /*
2106 * This ensures that boot memory sections on the platform are online
2107 * from early boot. Memory sections could not be prevented from being
2108 * offlined, unless for some reason they are not online to begin with.
2109 * This helps validate the basic assumption on which the above memory
2110 * event notifier works to prevent boot memory section offlining and
2111 * its possible removal.
2112 */
2113 static void validate_bootmem_online(void)
2114 {
2115 phys_addr_t start, end, addr;
2116 struct mem_section *ms;
2117 u64 i;
2118
2119 /*
2120 * Scanning across all memblock might be expensive
2121 * on some big memory systems. Hence enable this
2122 * validation only with DEBUG_VM.
2123 */
2124 if (!IS_ENABLED(CONFIG_DEBUG_VM))
2125 return;
2126
2127 for_each_mem_range(i, &start, &end) {
2128 for (addr = start; addr < end; addr += (1UL << PA_SECTION_SHIFT)) {
2129 ms = __pfn_to_section(PHYS_PFN(addr));
2130
2131 /*
2132 * All memory ranges in the system at this point
2133 * should have been marked as early sections.
2134 */
2135 WARN_ON(!early_section(ms));
2136
2137 /*
2138 * Memory notifier mechanism here to prevent boot
2139 * memory offlining depends on the fact that each
2140 * early section memory on the system is initially
2141 * online. Otherwise a given memory section which
2142 * is already offline will be overlooked and can
2143 * be removed completely. Call out such sections.
2144 */
2145 if (!online_section(ms))
2146 pr_err("Boot memory [%llx %llx] is offline, can be removed\n",
2147 addr, addr + (1UL << PA_SECTION_SHIFT));
2148 }
2149 }
2150 }
2151
2152 static int __init prevent_bootmem_remove_init(void)
2153 {
2154 int ret = 0;
2155
2156 if (!IS_ENABLED(CONFIG_MEMORY_HOTREMOVE))
2157 return ret;
2158
2159 validate_bootmem_online();
2160 ret = register_memory_notifier(&prevent_bootmem_remove_nb);
2161 if (ret)
2162 pr_err("%s: Notifier registration failed %d\n", __func__, ret);
2163
2164 return ret;
2165 }
2166 early_initcall(prevent_bootmem_remove_init);
2167 #endif
2168
2169 pte_t modify_prot_start_ptes(struct vm_area_struct *vma, unsigned long addr,
2170 pte_t *ptep, unsigned int nr)
2171 {
2172 pte_t pte = get_and_clear_ptes(vma->vm_mm, addr, ptep, nr);
2173
2174 if (alternative_has_cap_unlikely(ARM64_WORKAROUND_2645198)) {
2175 /*
2176 * Break-before-make (BBM) is required for all user space mappings
2177 * when the permission changes from executable to non-executable
2178 * in cases where cpu is affected with errata #2645198.
2179 */
2180 if (pte_accessible(vma->vm_mm, pte) && pte_user_exec(pte))
2181 __flush_tlb_range(vma, addr, nr * PAGE_SIZE,
2182 PAGE_SIZE, true, 3);
2183 }
2184
2185 return pte;
2186 }
2187
2188 pte_t ptep_modify_prot_start(struct vm_area_struct *vma, unsigned long addr, pte_t *ptep)
2189 {
2190 return modify_prot_start_ptes(vma, addr, ptep, 1);
2191 }
2192
2193 void modify_prot_commit_ptes(struct vm_area_struct *vma, unsigned long addr,
2194 pte_t *ptep, pte_t old_pte, pte_t pte,
2195 unsigned int nr)
2196 {
2197 set_ptes(vma->vm_mm, addr, ptep, pte, nr);
2198 }
2199
2200 void ptep_modify_prot_commit(struct vm_area_struct *vma, unsigned long addr, pte_t *ptep,
2201 pte_t old_pte, pte_t pte)
2202 {
2203 modify_prot_commit_ptes(vma, addr, ptep, old_pte, pte, 1);
2204 }
2205
2206 /*
2207 * Atomically replaces the active TTBR1_EL1 PGD with a new VA-compatible PGD,
2208 * avoiding the possibility of conflicting TLB entries being allocated.
2209 */
2210 void __cpu_replace_ttbr1(pgd_t *pgdp, bool cnp)
2211 {
2212 typedef void (ttbr_replace_func)(phys_addr_t);
2213 extern ttbr_replace_func idmap_cpu_replace_ttbr1;
2214 ttbr_replace_func *replace_phys;
2215 unsigned long daif;
2216
2217 /* phys_to_ttbr() zeros lower 2 bits of ttbr with 52-bit PA */
2218 phys_addr_t ttbr1 = phys_to_ttbr(virt_to_phys(pgdp));
2219
2220 if (cnp)
2221 ttbr1 |= TTBR_CNP_BIT;
2222
2223 replace_phys = (void *)__pa_symbol(idmap_cpu_replace_ttbr1);
2224
2225 cpu_install_idmap();
2226
2227 /*
2228 * We really don't want to take *any* exceptions while TTBR1 is
2229 * in the process of being replaced so mask everything.
2230 */
2231 daif = local_daif_save();
2232 replace_phys(ttbr1);
2233 local_daif_restore(daif);
2234
2235 cpu_uninstall_idmap();
2236 }
2237
2238 #ifdef CONFIG_ARCH_HAS_PKEYS
2239 int arch_set_user_pkey_access(struct task_struct *tsk, int pkey, unsigned long init_val)
2240 {
2241 u64 new_por;
2242 u64 old_por;
2243
2244 if (!system_supports_poe())
2245 return -ENOSPC;
2246
2247 /*
2248 * This code should only be called with valid 'pkey'
2249 * values originating from in-kernel users. Complain
2250 * if a bad value is observed.
2251 */
2252 if (WARN_ON_ONCE(pkey >= arch_max_pkey()))
2253 return -EINVAL;
2254
2255 /* Set the bits we need in POR: */
2256 new_por = POE_RWX;
2257 if (init_val & PKEY_DISABLE_WRITE)
2258 new_por &= ~POE_W;
2259 if (init_val & PKEY_DISABLE_ACCESS)
2260 new_por &= ~POE_RW;
2261 if (init_val & PKEY_DISABLE_READ)
2262 new_por &= ~POE_R;
2263 if (init_val & PKEY_DISABLE_EXECUTE)
2264 new_por &= ~POE_X;
2265
2266 /* Shift the bits in to the correct place in POR for pkey: */
2267 new_por = POR_ELx_PERM_PREP(pkey, new_por);
2268
2269 /* Get old POR and mask off any old bits in place: */
2270 old_por = read_sysreg_s(SYS_POR_EL0);
2271 old_por &= ~(POE_MASK << POR_ELx_PERM_SHIFT(pkey));
2272
2273 /* Write old part along with new part: */
2274 write_sysreg_s(old_por | new_por, SYS_POR_EL0);
2275
2276 return 0;
2277 }
2278 #endif