xref: /openbmc/linux/arch/arm64/mm/mmu.c (revision ffd08731)
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/fs.h>
21 #include <linux/io.h>
22 #include <linux/mm.h>
23 #include <linux/vmalloc.h>
24 
25 #include <asm/barrier.h>
26 #include <asm/cputype.h>
27 #include <asm/fixmap.h>
28 #include <asm/kasan.h>
29 #include <asm/kernel-pgtable.h>
30 #include <asm/sections.h>
31 #include <asm/setup.h>
32 #include <linux/sizes.h>
33 #include <asm/tlb.h>
34 #include <asm/mmu_context.h>
35 #include <asm/ptdump.h>
36 #include <asm/tlbflush.h>
37 
38 #define NO_BLOCK_MAPPINGS	BIT(0)
39 #define NO_CONT_MAPPINGS	BIT(1)
40 
41 u64 idmap_t0sz = TCR_T0SZ(VA_BITS);
42 u64 idmap_ptrs_per_pgd = PTRS_PER_PGD;
43 
44 u64 __section(".mmuoff.data.write") vabits_actual;
45 EXPORT_SYMBOL(vabits_actual);
46 
47 u64 kimage_voffset __ro_after_init;
48 EXPORT_SYMBOL(kimage_voffset);
49 
50 /*
51  * Empty_zero_page is a special page that is used for zero-initialized data
52  * and COW.
53  */
54 unsigned long empty_zero_page[PAGE_SIZE / sizeof(unsigned long)] __page_aligned_bss;
55 EXPORT_SYMBOL(empty_zero_page);
56 
57 static pte_t bm_pte[PTRS_PER_PTE] __page_aligned_bss;
58 static pmd_t bm_pmd[PTRS_PER_PMD] __page_aligned_bss __maybe_unused;
59 static pud_t bm_pud[PTRS_PER_PUD] __page_aligned_bss __maybe_unused;
60 
61 static DEFINE_SPINLOCK(swapper_pgdir_lock);
62 
63 void set_swapper_pgd(pgd_t *pgdp, pgd_t pgd)
64 {
65 	pgd_t *fixmap_pgdp;
66 
67 	spin_lock(&swapper_pgdir_lock);
68 	fixmap_pgdp = pgd_set_fixmap(__pa_symbol(pgdp));
69 	WRITE_ONCE(*fixmap_pgdp, pgd);
70 	/*
71 	 * We need dsb(ishst) here to ensure the page-table-walker sees
72 	 * our new entry before set_p?d() returns. The fixmap's
73 	 * flush_tlb_kernel_range() via clear_fixmap() does this for us.
74 	 */
75 	pgd_clear_fixmap();
76 	spin_unlock(&swapper_pgdir_lock);
77 }
78 
79 pgprot_t phys_mem_access_prot(struct file *file, unsigned long pfn,
80 			      unsigned long size, pgprot_t vma_prot)
81 {
82 	if (!pfn_valid(pfn))
83 		return pgprot_noncached(vma_prot);
84 	else if (file->f_flags & O_SYNC)
85 		return pgprot_writecombine(vma_prot);
86 	return vma_prot;
87 }
88 EXPORT_SYMBOL(phys_mem_access_prot);
89 
90 static phys_addr_t __init early_pgtable_alloc(int shift)
91 {
92 	phys_addr_t phys;
93 	void *ptr;
94 
95 	phys = memblock_phys_alloc(PAGE_SIZE, PAGE_SIZE);
96 	if (!phys)
97 		panic("Failed to allocate page table page\n");
98 
99 	/*
100 	 * The FIX_{PGD,PUD,PMD} slots may be in active use, but the FIX_PTE
101 	 * slot will be free, so we can (ab)use the FIX_PTE slot to initialise
102 	 * any level of table.
103 	 */
104 	ptr = pte_set_fixmap(phys);
105 
106 	memset(ptr, 0, PAGE_SIZE);
107 
108 	/*
109 	 * Implicit barriers also ensure the zeroed page is visible to the page
110 	 * table walker
111 	 */
112 	pte_clear_fixmap();
113 
114 	return phys;
115 }
116 
117 static bool pgattr_change_is_safe(u64 old, u64 new)
118 {
119 	/*
120 	 * The following mapping attributes may be updated in live
121 	 * kernel mappings without the need for break-before-make.
122 	 */
123 	static const pteval_t mask = PTE_PXN | PTE_RDONLY | PTE_WRITE | PTE_NG;
124 
125 	/* creating or taking down mappings is always safe */
126 	if (old == 0 || new == 0)
127 		return true;
128 
129 	/* live contiguous mappings may not be manipulated at all */
130 	if ((old | new) & PTE_CONT)
131 		return false;
132 
133 	/* Transitioning from Non-Global to Global is unsafe */
134 	if (old & ~new & PTE_NG)
135 		return false;
136 
137 	return ((old ^ new) & ~mask) == 0;
138 }
139 
140 static void init_pte(pmd_t *pmdp, unsigned long addr, unsigned long end,
141 		     phys_addr_t phys, pgprot_t prot)
142 {
143 	pte_t *ptep;
144 
145 	ptep = pte_set_fixmap_offset(pmdp, addr);
146 	do {
147 		pte_t old_pte = READ_ONCE(*ptep);
148 
149 		set_pte(ptep, pfn_pte(__phys_to_pfn(phys), prot));
150 
151 		/*
152 		 * After the PTE entry has been populated once, we
153 		 * only allow updates to the permission attributes.
154 		 */
155 		BUG_ON(!pgattr_change_is_safe(pte_val(old_pte),
156 					      READ_ONCE(pte_val(*ptep))));
157 
158 		phys += PAGE_SIZE;
159 	} while (ptep++, addr += PAGE_SIZE, addr != end);
160 
161 	pte_clear_fixmap();
162 }
163 
164 static void alloc_init_cont_pte(pmd_t *pmdp, unsigned long addr,
165 				unsigned long end, phys_addr_t phys,
166 				pgprot_t prot,
167 				phys_addr_t (*pgtable_alloc)(int),
168 				int flags)
169 {
170 	unsigned long next;
171 	pmd_t pmd = READ_ONCE(*pmdp);
172 
173 	BUG_ON(pmd_sect(pmd));
174 	if (pmd_none(pmd)) {
175 		phys_addr_t pte_phys;
176 		BUG_ON(!pgtable_alloc);
177 		pte_phys = pgtable_alloc(PAGE_SHIFT);
178 		__pmd_populate(pmdp, pte_phys, PMD_TYPE_TABLE);
179 		pmd = READ_ONCE(*pmdp);
180 	}
181 	BUG_ON(pmd_bad(pmd));
182 
183 	do {
184 		pgprot_t __prot = prot;
185 
186 		next = pte_cont_addr_end(addr, end);
187 
188 		/* use a contiguous mapping if the range is suitably aligned */
189 		if ((((addr | next | phys) & ~CONT_PTE_MASK) == 0) &&
190 		    (flags & NO_CONT_MAPPINGS) == 0)
191 			__prot = __pgprot(pgprot_val(prot) | PTE_CONT);
192 
193 		init_pte(pmdp, addr, next, phys, __prot);
194 
195 		phys += next - addr;
196 	} while (addr = next, addr != end);
197 }
198 
199 static void init_pmd(pud_t *pudp, unsigned long addr, unsigned long end,
200 		     phys_addr_t phys, pgprot_t prot,
201 		     phys_addr_t (*pgtable_alloc)(int), int flags)
202 {
203 	unsigned long next;
204 	pmd_t *pmdp;
205 
206 	pmdp = pmd_set_fixmap_offset(pudp, addr);
207 	do {
208 		pmd_t old_pmd = READ_ONCE(*pmdp);
209 
210 		next = pmd_addr_end(addr, end);
211 
212 		/* try section mapping first */
213 		if (((addr | next | phys) & ~SECTION_MASK) == 0 &&
214 		    (flags & NO_BLOCK_MAPPINGS) == 0) {
215 			pmd_set_huge(pmdp, phys, prot);
216 
217 			/*
218 			 * After the PMD entry has been populated once, we
219 			 * only allow updates to the permission attributes.
220 			 */
221 			BUG_ON(!pgattr_change_is_safe(pmd_val(old_pmd),
222 						      READ_ONCE(pmd_val(*pmdp))));
223 		} else {
224 			alloc_init_cont_pte(pmdp, addr, next, phys, prot,
225 					    pgtable_alloc, flags);
226 
227 			BUG_ON(pmd_val(old_pmd) != 0 &&
228 			       pmd_val(old_pmd) != READ_ONCE(pmd_val(*pmdp)));
229 		}
230 		phys += next - addr;
231 	} while (pmdp++, addr = next, addr != end);
232 
233 	pmd_clear_fixmap();
234 }
235 
236 static void alloc_init_cont_pmd(pud_t *pudp, unsigned long addr,
237 				unsigned long end, phys_addr_t phys,
238 				pgprot_t prot,
239 				phys_addr_t (*pgtable_alloc)(int), int flags)
240 {
241 	unsigned long next;
242 	pud_t pud = READ_ONCE(*pudp);
243 
244 	/*
245 	 * Check for initial section mappings in the pgd/pud.
246 	 */
247 	BUG_ON(pud_sect(pud));
248 	if (pud_none(pud)) {
249 		phys_addr_t pmd_phys;
250 		BUG_ON(!pgtable_alloc);
251 		pmd_phys = pgtable_alloc(PMD_SHIFT);
252 		__pud_populate(pudp, pmd_phys, PUD_TYPE_TABLE);
253 		pud = READ_ONCE(*pudp);
254 	}
255 	BUG_ON(pud_bad(pud));
256 
257 	do {
258 		pgprot_t __prot = prot;
259 
260 		next = pmd_cont_addr_end(addr, end);
261 
262 		/* use a contiguous mapping if the range is suitably aligned */
263 		if ((((addr | next | phys) & ~CONT_PMD_MASK) == 0) &&
264 		    (flags & NO_CONT_MAPPINGS) == 0)
265 			__prot = __pgprot(pgprot_val(prot) | PTE_CONT);
266 
267 		init_pmd(pudp, addr, next, phys, __prot, pgtable_alloc, flags);
268 
269 		phys += next - addr;
270 	} while (addr = next, addr != end);
271 }
272 
273 static inline bool use_1G_block(unsigned long addr, unsigned long next,
274 			unsigned long phys)
275 {
276 	if (PAGE_SHIFT != 12)
277 		return false;
278 
279 	if (((addr | next | phys) & ~PUD_MASK) != 0)
280 		return false;
281 
282 	return true;
283 }
284 
285 static void alloc_init_pud(pgd_t *pgdp, unsigned long addr, unsigned long end,
286 			   phys_addr_t phys, pgprot_t prot,
287 			   phys_addr_t (*pgtable_alloc)(int),
288 			   int flags)
289 {
290 	unsigned long next;
291 	pud_t *pudp;
292 	pgd_t pgd = READ_ONCE(*pgdp);
293 
294 	if (pgd_none(pgd)) {
295 		phys_addr_t pud_phys;
296 		BUG_ON(!pgtable_alloc);
297 		pud_phys = pgtable_alloc(PUD_SHIFT);
298 		__pgd_populate(pgdp, pud_phys, PUD_TYPE_TABLE);
299 		pgd = READ_ONCE(*pgdp);
300 	}
301 	BUG_ON(pgd_bad(pgd));
302 
303 	pudp = pud_set_fixmap_offset(pgdp, addr);
304 	do {
305 		pud_t old_pud = READ_ONCE(*pudp);
306 
307 		next = pud_addr_end(addr, end);
308 
309 		/*
310 		 * For 4K granule only, attempt to put down a 1GB block
311 		 */
312 		if (use_1G_block(addr, next, phys) &&
313 		    (flags & NO_BLOCK_MAPPINGS) == 0) {
314 			pud_set_huge(pudp, phys, prot);
315 
316 			/*
317 			 * After the PUD entry has been populated once, we
318 			 * only allow updates to the permission attributes.
319 			 */
320 			BUG_ON(!pgattr_change_is_safe(pud_val(old_pud),
321 						      READ_ONCE(pud_val(*pudp))));
322 		} else {
323 			alloc_init_cont_pmd(pudp, addr, next, phys, prot,
324 					    pgtable_alloc, flags);
325 
326 			BUG_ON(pud_val(old_pud) != 0 &&
327 			       pud_val(old_pud) != READ_ONCE(pud_val(*pudp)));
328 		}
329 		phys += next - addr;
330 	} while (pudp++, addr = next, addr != end);
331 
332 	pud_clear_fixmap();
333 }
334 
335 static void __create_pgd_mapping(pgd_t *pgdir, phys_addr_t phys,
336 				 unsigned long virt, phys_addr_t size,
337 				 pgprot_t prot,
338 				 phys_addr_t (*pgtable_alloc)(int),
339 				 int flags)
340 {
341 	unsigned long addr, end, next;
342 	pgd_t *pgdp = pgd_offset_raw(pgdir, virt);
343 
344 	/*
345 	 * If the virtual and physical address don't have the same offset
346 	 * within a page, we cannot map the region as the caller expects.
347 	 */
348 	if (WARN_ON((phys ^ virt) & ~PAGE_MASK))
349 		return;
350 
351 	phys &= PAGE_MASK;
352 	addr = virt & PAGE_MASK;
353 	end = PAGE_ALIGN(virt + size);
354 
355 	do {
356 		next = pgd_addr_end(addr, end);
357 		alloc_init_pud(pgdp, addr, next, phys, prot, pgtable_alloc,
358 			       flags);
359 		phys += next - addr;
360 	} while (pgdp++, addr = next, addr != end);
361 }
362 
363 static phys_addr_t __pgd_pgtable_alloc(int shift)
364 {
365 	void *ptr = (void *)__get_free_page(GFP_PGTABLE_KERNEL);
366 	BUG_ON(!ptr);
367 
368 	/* Ensure the zeroed page is visible to the page table walker */
369 	dsb(ishst);
370 	return __pa(ptr);
371 }
372 
373 static phys_addr_t pgd_pgtable_alloc(int shift)
374 {
375 	phys_addr_t pa = __pgd_pgtable_alloc(shift);
376 
377 	/*
378 	 * Call proper page table ctor in case later we need to
379 	 * call core mm functions like apply_to_page_range() on
380 	 * this pre-allocated page table.
381 	 *
382 	 * We don't select ARCH_ENABLE_SPLIT_PMD_PTLOCK if pmd is
383 	 * folded, and if so pgtable_pmd_page_ctor() becomes nop.
384 	 */
385 	if (shift == PAGE_SHIFT)
386 		BUG_ON(!pgtable_pte_page_ctor(phys_to_page(pa)));
387 	else if (shift == PMD_SHIFT)
388 		BUG_ON(!pgtable_pmd_page_ctor(phys_to_page(pa)));
389 
390 	return pa;
391 }
392 
393 /*
394  * This function can only be used to modify existing table entries,
395  * without allocating new levels of table. Note that this permits the
396  * creation of new section or page entries.
397  */
398 static void __init create_mapping_noalloc(phys_addr_t phys, unsigned long virt,
399 				  phys_addr_t size, pgprot_t prot)
400 {
401 	if ((virt >= PAGE_END) && (virt < VMALLOC_START)) {
402 		pr_warn("BUG: not creating mapping for %pa at 0x%016lx - outside kernel range\n",
403 			&phys, virt);
404 		return;
405 	}
406 	__create_pgd_mapping(init_mm.pgd, phys, virt, size, prot, NULL,
407 			     NO_CONT_MAPPINGS);
408 }
409 
410 void __init create_pgd_mapping(struct mm_struct *mm, phys_addr_t phys,
411 			       unsigned long virt, phys_addr_t size,
412 			       pgprot_t prot, bool page_mappings_only)
413 {
414 	int flags = 0;
415 
416 	BUG_ON(mm == &init_mm);
417 
418 	if (page_mappings_only)
419 		flags = NO_BLOCK_MAPPINGS | NO_CONT_MAPPINGS;
420 
421 	__create_pgd_mapping(mm->pgd, phys, virt, size, prot,
422 			     pgd_pgtable_alloc, flags);
423 }
424 
425 static void update_mapping_prot(phys_addr_t phys, unsigned long virt,
426 				phys_addr_t size, pgprot_t prot)
427 {
428 	if ((virt >= PAGE_END) && (virt < VMALLOC_START)) {
429 		pr_warn("BUG: not updating mapping for %pa at 0x%016lx - outside kernel range\n",
430 			&phys, virt);
431 		return;
432 	}
433 
434 	__create_pgd_mapping(init_mm.pgd, phys, virt, size, prot, NULL,
435 			     NO_CONT_MAPPINGS);
436 
437 	/* flush the TLBs after updating live kernel mappings */
438 	flush_tlb_kernel_range(virt, virt + size);
439 }
440 
441 static void __init __map_memblock(pgd_t *pgdp, phys_addr_t start,
442 				  phys_addr_t end, pgprot_t prot, int flags)
443 {
444 	__create_pgd_mapping(pgdp, start, __phys_to_virt(start), end - start,
445 			     prot, early_pgtable_alloc, flags);
446 }
447 
448 void __init mark_linear_text_alias_ro(void)
449 {
450 	/*
451 	 * Remove the write permissions from the linear alias of .text/.rodata
452 	 */
453 	update_mapping_prot(__pa_symbol(_text), (unsigned long)lm_alias(_text),
454 			    (unsigned long)__init_begin - (unsigned long)_text,
455 			    PAGE_KERNEL_RO);
456 }
457 
458 static void __init map_mem(pgd_t *pgdp)
459 {
460 	phys_addr_t kernel_start = __pa_symbol(_text);
461 	phys_addr_t kernel_end = __pa_symbol(__init_begin);
462 	struct memblock_region *reg;
463 	int flags = 0;
464 
465 	if (rodata_full || debug_pagealloc_enabled())
466 		flags = NO_BLOCK_MAPPINGS | NO_CONT_MAPPINGS;
467 
468 	/*
469 	 * Take care not to create a writable alias for the
470 	 * read-only text and rodata sections of the kernel image.
471 	 * So temporarily mark them as NOMAP to skip mappings in
472 	 * the following for-loop
473 	 */
474 	memblock_mark_nomap(kernel_start, kernel_end - kernel_start);
475 #ifdef CONFIG_KEXEC_CORE
476 	if (crashk_res.end)
477 		memblock_mark_nomap(crashk_res.start,
478 				    resource_size(&crashk_res));
479 #endif
480 
481 	/* map all the memory banks */
482 	for_each_memblock(memory, reg) {
483 		phys_addr_t start = reg->base;
484 		phys_addr_t end = start + reg->size;
485 
486 		if (start >= end)
487 			break;
488 		if (memblock_is_nomap(reg))
489 			continue;
490 
491 		__map_memblock(pgdp, start, end, PAGE_KERNEL, flags);
492 	}
493 
494 	/*
495 	 * Map the linear alias of the [_text, __init_begin) interval
496 	 * as non-executable now, and remove the write permission in
497 	 * mark_linear_text_alias_ro() below (which will be called after
498 	 * alternative patching has completed). This makes the contents
499 	 * of the region accessible to subsystems such as hibernate,
500 	 * but protects it from inadvertent modification or execution.
501 	 * Note that contiguous mappings cannot be remapped in this way,
502 	 * so we should avoid them here.
503 	 */
504 	__map_memblock(pgdp, kernel_start, kernel_end,
505 		       PAGE_KERNEL, NO_CONT_MAPPINGS);
506 	memblock_clear_nomap(kernel_start, kernel_end - kernel_start);
507 
508 #ifdef CONFIG_KEXEC_CORE
509 	/*
510 	 * Use page-level mappings here so that we can shrink the region
511 	 * in page granularity and put back unused memory to buddy system
512 	 * through /sys/kernel/kexec_crash_size interface.
513 	 */
514 	if (crashk_res.end) {
515 		__map_memblock(pgdp, crashk_res.start, crashk_res.end + 1,
516 			       PAGE_KERNEL,
517 			       NO_BLOCK_MAPPINGS | NO_CONT_MAPPINGS);
518 		memblock_clear_nomap(crashk_res.start,
519 				     resource_size(&crashk_res));
520 	}
521 #endif
522 }
523 
524 void mark_rodata_ro(void)
525 {
526 	unsigned long section_size;
527 
528 	/*
529 	 * mark .rodata as read only. Use __init_begin rather than __end_rodata
530 	 * to cover NOTES and EXCEPTION_TABLE.
531 	 */
532 	section_size = (unsigned long)__init_begin - (unsigned long)__start_rodata;
533 	update_mapping_prot(__pa_symbol(__start_rodata), (unsigned long)__start_rodata,
534 			    section_size, PAGE_KERNEL_RO);
535 
536 	debug_checkwx();
537 }
538 
539 static void __init map_kernel_segment(pgd_t *pgdp, void *va_start, void *va_end,
540 				      pgprot_t prot, struct vm_struct *vma,
541 				      int flags, unsigned long vm_flags)
542 {
543 	phys_addr_t pa_start = __pa_symbol(va_start);
544 	unsigned long size = va_end - va_start;
545 
546 	BUG_ON(!PAGE_ALIGNED(pa_start));
547 	BUG_ON(!PAGE_ALIGNED(size));
548 
549 	__create_pgd_mapping(pgdp, pa_start, (unsigned long)va_start, size, prot,
550 			     early_pgtable_alloc, flags);
551 
552 	if (!(vm_flags & VM_NO_GUARD))
553 		size += PAGE_SIZE;
554 
555 	vma->addr	= va_start;
556 	vma->phys_addr	= pa_start;
557 	vma->size	= size;
558 	vma->flags	= VM_MAP | vm_flags;
559 	vma->caller	= __builtin_return_address(0);
560 
561 	vm_area_add_early(vma);
562 }
563 
564 static int __init parse_rodata(char *arg)
565 {
566 	int ret = strtobool(arg, &rodata_enabled);
567 	if (!ret) {
568 		rodata_full = false;
569 		return 0;
570 	}
571 
572 	/* permit 'full' in addition to boolean options */
573 	if (strcmp(arg, "full"))
574 		return -EINVAL;
575 
576 	rodata_enabled = true;
577 	rodata_full = true;
578 	return 0;
579 }
580 early_param("rodata", parse_rodata);
581 
582 #ifdef CONFIG_UNMAP_KERNEL_AT_EL0
583 static int __init map_entry_trampoline(void)
584 {
585 	pgprot_t prot = rodata_enabled ? PAGE_KERNEL_ROX : PAGE_KERNEL_EXEC;
586 	phys_addr_t pa_start = __pa_symbol(__entry_tramp_text_start);
587 
588 	/* The trampoline is always mapped and can therefore be global */
589 	pgprot_val(prot) &= ~PTE_NG;
590 
591 	/* Map only the text into the trampoline page table */
592 	memset(tramp_pg_dir, 0, PGD_SIZE);
593 	__create_pgd_mapping(tramp_pg_dir, pa_start, TRAMP_VALIAS, PAGE_SIZE,
594 			     prot, __pgd_pgtable_alloc, 0);
595 
596 	/* Map both the text and data into the kernel page table */
597 	__set_fixmap(FIX_ENTRY_TRAMP_TEXT, pa_start, prot);
598 	if (IS_ENABLED(CONFIG_RANDOMIZE_BASE)) {
599 		extern char __entry_tramp_data_start[];
600 
601 		__set_fixmap(FIX_ENTRY_TRAMP_DATA,
602 			     __pa_symbol(__entry_tramp_data_start),
603 			     PAGE_KERNEL_RO);
604 	}
605 
606 	return 0;
607 }
608 core_initcall(map_entry_trampoline);
609 #endif
610 
611 /*
612  * Create fine-grained mappings for the kernel.
613  */
614 static void __init map_kernel(pgd_t *pgdp)
615 {
616 	static struct vm_struct vmlinux_text, vmlinux_rodata, vmlinux_inittext,
617 				vmlinux_initdata, vmlinux_data;
618 
619 	/*
620 	 * External debuggers may need to write directly to the text
621 	 * mapping to install SW breakpoints. Allow this (only) when
622 	 * explicitly requested with rodata=off.
623 	 */
624 	pgprot_t text_prot = rodata_enabled ? PAGE_KERNEL_ROX : PAGE_KERNEL_EXEC;
625 
626 	/*
627 	 * Only rodata will be remapped with different permissions later on,
628 	 * all other segments are allowed to use contiguous mappings.
629 	 */
630 	map_kernel_segment(pgdp, _text, _etext, text_prot, &vmlinux_text, 0,
631 			   VM_NO_GUARD);
632 	map_kernel_segment(pgdp, __start_rodata, __inittext_begin, PAGE_KERNEL,
633 			   &vmlinux_rodata, NO_CONT_MAPPINGS, VM_NO_GUARD);
634 	map_kernel_segment(pgdp, __inittext_begin, __inittext_end, text_prot,
635 			   &vmlinux_inittext, 0, VM_NO_GUARD);
636 	map_kernel_segment(pgdp, __initdata_begin, __initdata_end, PAGE_KERNEL,
637 			   &vmlinux_initdata, 0, VM_NO_GUARD);
638 	map_kernel_segment(pgdp, _data, _end, PAGE_KERNEL, &vmlinux_data, 0, 0);
639 
640 	if (!READ_ONCE(pgd_val(*pgd_offset_raw(pgdp, FIXADDR_START)))) {
641 		/*
642 		 * The fixmap falls in a separate pgd to the kernel, and doesn't
643 		 * live in the carveout for the swapper_pg_dir. We can simply
644 		 * re-use the existing dir for the fixmap.
645 		 */
646 		set_pgd(pgd_offset_raw(pgdp, FIXADDR_START),
647 			READ_ONCE(*pgd_offset_k(FIXADDR_START)));
648 	} else if (CONFIG_PGTABLE_LEVELS > 3) {
649 		pgd_t *bm_pgdp;
650 		pud_t *bm_pudp;
651 		/*
652 		 * The fixmap shares its top level pgd entry with the kernel
653 		 * mapping. This can really only occur when we are running
654 		 * with 16k/4 levels, so we can simply reuse the pud level
655 		 * entry instead.
656 		 */
657 		BUG_ON(!IS_ENABLED(CONFIG_ARM64_16K_PAGES));
658 		bm_pgdp = pgd_offset_raw(pgdp, FIXADDR_START);
659 		bm_pudp = pud_set_fixmap_offset(bm_pgdp, FIXADDR_START);
660 		pud_populate(&init_mm, bm_pudp, lm_alias(bm_pmd));
661 		pud_clear_fixmap();
662 	} else {
663 		BUG();
664 	}
665 
666 	kasan_copy_shadow(pgdp);
667 }
668 
669 void __init paging_init(void)
670 {
671 	pgd_t *pgdp = pgd_set_fixmap(__pa_symbol(swapper_pg_dir));
672 
673 	map_kernel(pgdp);
674 	map_mem(pgdp);
675 
676 	pgd_clear_fixmap();
677 
678 	cpu_replace_ttbr1(lm_alias(swapper_pg_dir));
679 	init_mm.pgd = swapper_pg_dir;
680 
681 	memblock_free(__pa_symbol(init_pg_dir),
682 		      __pa_symbol(init_pg_end) - __pa_symbol(init_pg_dir));
683 
684 	memblock_allow_resize();
685 }
686 
687 /*
688  * Check whether a kernel address is valid (derived from arch/x86/).
689  */
690 int kern_addr_valid(unsigned long addr)
691 {
692 	pgd_t *pgdp;
693 	pud_t *pudp, pud;
694 	pmd_t *pmdp, pmd;
695 	pte_t *ptep, pte;
696 
697 	if ((((long)addr) >> VA_BITS) != -1UL)
698 		return 0;
699 
700 	pgdp = pgd_offset_k(addr);
701 	if (pgd_none(READ_ONCE(*pgdp)))
702 		return 0;
703 
704 	pudp = pud_offset(pgdp, addr);
705 	pud = READ_ONCE(*pudp);
706 	if (pud_none(pud))
707 		return 0;
708 
709 	if (pud_sect(pud))
710 		return pfn_valid(pud_pfn(pud));
711 
712 	pmdp = pmd_offset(pudp, addr);
713 	pmd = READ_ONCE(*pmdp);
714 	if (pmd_none(pmd))
715 		return 0;
716 
717 	if (pmd_sect(pmd))
718 		return pfn_valid(pmd_pfn(pmd));
719 
720 	ptep = pte_offset_kernel(pmdp, addr);
721 	pte = READ_ONCE(*ptep);
722 	if (pte_none(pte))
723 		return 0;
724 
725 	return pfn_valid(pte_pfn(pte));
726 }
727 #ifdef CONFIG_SPARSEMEM_VMEMMAP
728 #if !ARM64_SWAPPER_USES_SECTION_MAPS
729 int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node,
730 		struct vmem_altmap *altmap)
731 {
732 	return vmemmap_populate_basepages(start, end, node);
733 }
734 #else	/* !ARM64_SWAPPER_USES_SECTION_MAPS */
735 int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node,
736 		struct vmem_altmap *altmap)
737 {
738 	unsigned long addr = start;
739 	unsigned long next;
740 	pgd_t *pgdp;
741 	pud_t *pudp;
742 	pmd_t *pmdp;
743 
744 	do {
745 		next = pmd_addr_end(addr, end);
746 
747 		pgdp = vmemmap_pgd_populate(addr, node);
748 		if (!pgdp)
749 			return -ENOMEM;
750 
751 		pudp = vmemmap_pud_populate(pgdp, addr, node);
752 		if (!pudp)
753 			return -ENOMEM;
754 
755 		pmdp = pmd_offset(pudp, addr);
756 		if (pmd_none(READ_ONCE(*pmdp))) {
757 			void *p = NULL;
758 
759 			p = vmemmap_alloc_block_buf(PMD_SIZE, node);
760 			if (!p)
761 				return -ENOMEM;
762 
763 			pmd_set_huge(pmdp, __pa(p), __pgprot(PROT_SECT_NORMAL));
764 		} else
765 			vmemmap_verify((pte_t *)pmdp, node, addr, next);
766 	} while (addr = next, addr != end);
767 
768 	return 0;
769 }
770 #endif	/* !ARM64_SWAPPER_USES_SECTION_MAPS */
771 void vmemmap_free(unsigned long start, unsigned long end,
772 		struct vmem_altmap *altmap)
773 {
774 }
775 #endif	/* CONFIG_SPARSEMEM_VMEMMAP */
776 
777 static inline pud_t * fixmap_pud(unsigned long addr)
778 {
779 	pgd_t *pgdp = pgd_offset_k(addr);
780 	pgd_t pgd = READ_ONCE(*pgdp);
781 
782 	BUG_ON(pgd_none(pgd) || pgd_bad(pgd));
783 
784 	return pud_offset_kimg(pgdp, addr);
785 }
786 
787 static inline pmd_t * fixmap_pmd(unsigned long addr)
788 {
789 	pud_t *pudp = fixmap_pud(addr);
790 	pud_t pud = READ_ONCE(*pudp);
791 
792 	BUG_ON(pud_none(pud) || pud_bad(pud));
793 
794 	return pmd_offset_kimg(pudp, addr);
795 }
796 
797 static inline pte_t * fixmap_pte(unsigned long addr)
798 {
799 	return &bm_pte[pte_index(addr)];
800 }
801 
802 /*
803  * The p*d_populate functions call virt_to_phys implicitly so they can't be used
804  * directly on kernel symbols (bm_p*d). This function is called too early to use
805  * lm_alias so __p*d_populate functions must be used to populate with the
806  * physical address from __pa_symbol.
807  */
808 void __init early_fixmap_init(void)
809 {
810 	pgd_t *pgdp, pgd;
811 	pud_t *pudp;
812 	pmd_t *pmdp;
813 	unsigned long addr = FIXADDR_START;
814 
815 	pgdp = pgd_offset_k(addr);
816 	pgd = READ_ONCE(*pgdp);
817 	if (CONFIG_PGTABLE_LEVELS > 3 &&
818 	    !(pgd_none(pgd) || pgd_page_paddr(pgd) == __pa_symbol(bm_pud))) {
819 		/*
820 		 * We only end up here if the kernel mapping and the fixmap
821 		 * share the top level pgd entry, which should only happen on
822 		 * 16k/4 levels configurations.
823 		 */
824 		BUG_ON(!IS_ENABLED(CONFIG_ARM64_16K_PAGES));
825 		pudp = pud_offset_kimg(pgdp, addr);
826 	} else {
827 		if (pgd_none(pgd))
828 			__pgd_populate(pgdp, __pa_symbol(bm_pud), PUD_TYPE_TABLE);
829 		pudp = fixmap_pud(addr);
830 	}
831 	if (pud_none(READ_ONCE(*pudp)))
832 		__pud_populate(pudp, __pa_symbol(bm_pmd), PMD_TYPE_TABLE);
833 	pmdp = fixmap_pmd(addr);
834 	__pmd_populate(pmdp, __pa_symbol(bm_pte), PMD_TYPE_TABLE);
835 
836 	/*
837 	 * The boot-ioremap range spans multiple pmds, for which
838 	 * we are not prepared:
839 	 */
840 	BUILD_BUG_ON((__fix_to_virt(FIX_BTMAP_BEGIN) >> PMD_SHIFT)
841 		     != (__fix_to_virt(FIX_BTMAP_END) >> PMD_SHIFT));
842 
843 	if ((pmdp != fixmap_pmd(fix_to_virt(FIX_BTMAP_BEGIN)))
844 	     || pmdp != fixmap_pmd(fix_to_virt(FIX_BTMAP_END))) {
845 		WARN_ON(1);
846 		pr_warn("pmdp %p != %p, %p\n",
847 			pmdp, fixmap_pmd(fix_to_virt(FIX_BTMAP_BEGIN)),
848 			fixmap_pmd(fix_to_virt(FIX_BTMAP_END)));
849 		pr_warn("fix_to_virt(FIX_BTMAP_BEGIN): %08lx\n",
850 			fix_to_virt(FIX_BTMAP_BEGIN));
851 		pr_warn("fix_to_virt(FIX_BTMAP_END):   %08lx\n",
852 			fix_to_virt(FIX_BTMAP_END));
853 
854 		pr_warn("FIX_BTMAP_END:       %d\n", FIX_BTMAP_END);
855 		pr_warn("FIX_BTMAP_BEGIN:     %d\n", FIX_BTMAP_BEGIN);
856 	}
857 }
858 
859 /*
860  * Unusually, this is also called in IRQ context (ghes_iounmap_irq) so if we
861  * ever need to use IPIs for TLB broadcasting, then we're in trouble here.
862  */
863 void __set_fixmap(enum fixed_addresses idx,
864 			       phys_addr_t phys, pgprot_t flags)
865 {
866 	unsigned long addr = __fix_to_virt(idx);
867 	pte_t *ptep;
868 
869 	BUG_ON(idx <= FIX_HOLE || idx >= __end_of_fixed_addresses);
870 
871 	ptep = fixmap_pte(addr);
872 
873 	if (pgprot_val(flags)) {
874 		set_pte(ptep, pfn_pte(phys >> PAGE_SHIFT, flags));
875 	} else {
876 		pte_clear(&init_mm, addr, ptep);
877 		flush_tlb_kernel_range(addr, addr+PAGE_SIZE);
878 	}
879 }
880 
881 void *__init fixmap_remap_fdt(phys_addr_t dt_phys, int *size, pgprot_t prot)
882 {
883 	const u64 dt_virt_base = __fix_to_virt(FIX_FDT);
884 	int offset;
885 	void *dt_virt;
886 
887 	/*
888 	 * Check whether the physical FDT address is set and meets the minimum
889 	 * alignment requirement. Since we are relying on MIN_FDT_ALIGN to be
890 	 * at least 8 bytes so that we can always access the magic and size
891 	 * fields of the FDT header after mapping the first chunk, double check
892 	 * here if that is indeed the case.
893 	 */
894 	BUILD_BUG_ON(MIN_FDT_ALIGN < 8);
895 	if (!dt_phys || dt_phys % MIN_FDT_ALIGN)
896 		return NULL;
897 
898 	/*
899 	 * Make sure that the FDT region can be mapped without the need to
900 	 * allocate additional translation table pages, so that it is safe
901 	 * to call create_mapping_noalloc() this early.
902 	 *
903 	 * On 64k pages, the FDT will be mapped using PTEs, so we need to
904 	 * be in the same PMD as the rest of the fixmap.
905 	 * On 4k pages, we'll use section mappings for the FDT so we only
906 	 * have to be in the same PUD.
907 	 */
908 	BUILD_BUG_ON(dt_virt_base % SZ_2M);
909 
910 	BUILD_BUG_ON(__fix_to_virt(FIX_FDT_END) >> SWAPPER_TABLE_SHIFT !=
911 		     __fix_to_virt(FIX_BTMAP_BEGIN) >> SWAPPER_TABLE_SHIFT);
912 
913 	offset = dt_phys % SWAPPER_BLOCK_SIZE;
914 	dt_virt = (void *)dt_virt_base + offset;
915 
916 	/* map the first chunk so we can read the size from the header */
917 	create_mapping_noalloc(round_down(dt_phys, SWAPPER_BLOCK_SIZE),
918 			dt_virt_base, SWAPPER_BLOCK_SIZE, prot);
919 
920 	if (fdt_magic(dt_virt) != FDT_MAGIC)
921 		return NULL;
922 
923 	*size = fdt_totalsize(dt_virt);
924 	if (*size > MAX_FDT_SIZE)
925 		return NULL;
926 
927 	if (offset + *size > SWAPPER_BLOCK_SIZE)
928 		create_mapping_noalloc(round_down(dt_phys, SWAPPER_BLOCK_SIZE), dt_virt_base,
929 			       round_up(offset + *size, SWAPPER_BLOCK_SIZE), prot);
930 
931 	return dt_virt;
932 }
933 
934 int __init arch_ioremap_p4d_supported(void)
935 {
936 	return 0;
937 }
938 
939 int __init arch_ioremap_pud_supported(void)
940 {
941 	/*
942 	 * Only 4k granule supports level 1 block mappings.
943 	 * SW table walks can't handle removal of intermediate entries.
944 	 */
945 	return IS_ENABLED(CONFIG_ARM64_4K_PAGES) &&
946 	       !IS_ENABLED(CONFIG_ARM64_PTDUMP_DEBUGFS);
947 }
948 
949 int __init arch_ioremap_pmd_supported(void)
950 {
951 	/* See arch_ioremap_pud_supported() */
952 	return !IS_ENABLED(CONFIG_ARM64_PTDUMP_DEBUGFS);
953 }
954 
955 int pud_set_huge(pud_t *pudp, phys_addr_t phys, pgprot_t prot)
956 {
957 	pud_t new_pud = pfn_pud(__phys_to_pfn(phys), mk_pud_sect_prot(prot));
958 
959 	/* Only allow permission changes for now */
960 	if (!pgattr_change_is_safe(READ_ONCE(pud_val(*pudp)),
961 				   pud_val(new_pud)))
962 		return 0;
963 
964 	VM_BUG_ON(phys & ~PUD_MASK);
965 	set_pud(pudp, new_pud);
966 	return 1;
967 }
968 
969 int pmd_set_huge(pmd_t *pmdp, phys_addr_t phys, pgprot_t prot)
970 {
971 	pmd_t new_pmd = pfn_pmd(__phys_to_pfn(phys), mk_pmd_sect_prot(prot));
972 
973 	/* Only allow permission changes for now */
974 	if (!pgattr_change_is_safe(READ_ONCE(pmd_val(*pmdp)),
975 				   pmd_val(new_pmd)))
976 		return 0;
977 
978 	VM_BUG_ON(phys & ~PMD_MASK);
979 	set_pmd(pmdp, new_pmd);
980 	return 1;
981 }
982 
983 int pud_clear_huge(pud_t *pudp)
984 {
985 	if (!pud_sect(READ_ONCE(*pudp)))
986 		return 0;
987 	pud_clear(pudp);
988 	return 1;
989 }
990 
991 int pmd_clear_huge(pmd_t *pmdp)
992 {
993 	if (!pmd_sect(READ_ONCE(*pmdp)))
994 		return 0;
995 	pmd_clear(pmdp);
996 	return 1;
997 }
998 
999 int pmd_free_pte_page(pmd_t *pmdp, unsigned long addr)
1000 {
1001 	pte_t *table;
1002 	pmd_t pmd;
1003 
1004 	pmd = READ_ONCE(*pmdp);
1005 
1006 	if (!pmd_table(pmd)) {
1007 		VM_WARN_ON(1);
1008 		return 1;
1009 	}
1010 
1011 	table = pte_offset_kernel(pmdp, addr);
1012 	pmd_clear(pmdp);
1013 	__flush_tlb_kernel_pgtable(addr);
1014 	pte_free_kernel(NULL, table);
1015 	return 1;
1016 }
1017 
1018 int pud_free_pmd_page(pud_t *pudp, unsigned long addr)
1019 {
1020 	pmd_t *table;
1021 	pmd_t *pmdp;
1022 	pud_t pud;
1023 	unsigned long next, end;
1024 
1025 	pud = READ_ONCE(*pudp);
1026 
1027 	if (!pud_table(pud)) {
1028 		VM_WARN_ON(1);
1029 		return 1;
1030 	}
1031 
1032 	table = pmd_offset(pudp, addr);
1033 	pmdp = table;
1034 	next = addr;
1035 	end = addr + PUD_SIZE;
1036 	do {
1037 		pmd_free_pte_page(pmdp, next);
1038 	} while (pmdp++, next += PMD_SIZE, next != end);
1039 
1040 	pud_clear(pudp);
1041 	__flush_tlb_kernel_pgtable(addr);
1042 	pmd_free(NULL, table);
1043 	return 1;
1044 }
1045 
1046 int p4d_free_pud_page(p4d_t *p4d, unsigned long addr)
1047 {
1048 	return 0;	/* Don't attempt a block mapping */
1049 }
1050 
1051 #ifdef CONFIG_MEMORY_HOTPLUG
1052 int arch_add_memory(int nid, u64 start, u64 size,
1053 			struct mhp_restrictions *restrictions)
1054 {
1055 	int flags = 0;
1056 
1057 	if (rodata_full || debug_pagealloc_enabled())
1058 		flags = NO_BLOCK_MAPPINGS | NO_CONT_MAPPINGS;
1059 
1060 	__create_pgd_mapping(swapper_pg_dir, start, __phys_to_virt(start),
1061 			     size, PAGE_KERNEL, __pgd_pgtable_alloc, flags);
1062 
1063 	memblock_clear_nomap(start, size);
1064 
1065 	return __add_pages(nid, start >> PAGE_SHIFT, size >> PAGE_SHIFT,
1066 			   restrictions);
1067 }
1068 void arch_remove_memory(int nid, u64 start, u64 size,
1069 			struct vmem_altmap *altmap)
1070 {
1071 	unsigned long start_pfn = start >> PAGE_SHIFT;
1072 	unsigned long nr_pages = size >> PAGE_SHIFT;
1073 
1074 	/*
1075 	 * FIXME: Cleanup page tables (also in arch_add_memory() in case
1076 	 * adding fails). Until then, this function should only be used
1077 	 * during memory hotplug (adding memory), not for memory
1078 	 * unplug. ARCH_ENABLE_MEMORY_HOTREMOVE must not be
1079 	 * unlocked yet.
1080 	 */
1081 	__remove_pages(start_pfn, nr_pages, altmap);
1082 }
1083 #endif
1084