xref: /linux/arch/riscv/mm/init.c (revision edbafe65eef2b58625db1e113fbbfb1fe10c0291)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Copyright (C) 2012 Regents of the University of California
4  * Copyright (C) 2019 Western Digital Corporation or its affiliates.
5  * Copyright (C) 2020 FORTH-ICS/CARV
6  *  Nick Kossifidis <mick@ics.forth.gr>
7  */
8 
9 #include <linux/init.h>
10 #include <linux/mm.h>
11 #include <linux/memblock.h>
12 #include <linux/initrd.h>
13 #include <linux/swap.h>
14 #include <linux/swiotlb.h>
15 #include <linux/sizes.h>
16 #include <linux/of_fdt.h>
17 #include <linux/of_reserved_mem.h>
18 #include <linux/libfdt.h>
19 #include <linux/set_memory.h>
20 #include <linux/dma-map-ops.h>
21 #include <linux/crash_dump.h>
22 #include <linux/hugetlb.h>
23 #include <linux/kfence.h>
24 #include <linux/execmem.h>
25 
26 #include <asm/alternative.h>
27 #include <asm/fixmap.h>
28 #include <asm/io.h>
29 #include <asm/kasan.h>
30 #include <asm/module.h>
31 #include <asm/numa.h>
32 #include <asm/pgtable.h>
33 #include <asm/sections.h>
34 #include <asm/soc.h>
35 #include <asm/sparsemem.h>
36 #include <asm/tlbflush.h>
37 
38 #include "../kernel/head.h"
39 
40 #if defined(CONFIG_64BIT) && defined(CONFIG_MMU)
41 DECLARE_BITMAP(new_valid_map_cpus, NR_CPUS);
42 #endif
43 
44 struct kernel_mapping kernel_map __ro_after_init;
45 EXPORT_SYMBOL(kernel_map);
46 
47 #ifdef CONFIG_64BIT
48 u64 satp_mode __ro_after_init = SATP_MODE_57;
49 #else
50 u64 satp_mode __ro_after_init = SATP_MODE_32;
51 #endif
52 EXPORT_SYMBOL(satp_mode);
53 
54 #ifdef CONFIG_64BIT
55 bool pgtable_l4_enabled __ro_after_init = true;
56 bool pgtable_l5_enabled __ro_after_init = true;
57 EXPORT_SYMBOL(pgtable_l4_enabled);
58 EXPORT_SYMBOL(pgtable_l5_enabled);
59 #endif
60 
61 phys_addr_t phys_ram_base __ro_after_init;
62 EXPORT_SYMBOL(phys_ram_base);
63 
64 #ifdef CONFIG_SPARSEMEM_VMEMMAP
65 #define VMEMMAP_ADDR_ALIGN	max(1ULL << SECTION_SIZE_BITS, \
66 				    PFN_PHYS(MAX_FOLIO_VMEMMAP_ALIGN / \
67 					     sizeof(struct page)))
68 
69 unsigned long vmemmap_start_pfn __ro_after_init;
70 EXPORT_SYMBOL(vmemmap_start_pfn);
71 #endif
72 
73 extern char _start[];
74 void *_dtb_early_va __initdata;
75 uintptr_t _dtb_early_pa __initdata;
76 
77 phys_addr_t dma32_phys_limit __initdata;
78 
79 void __init arch_zone_limits_init(unsigned long *max_zone_pfns)
80 {
81 #ifdef CONFIG_ZONE_DMA32
82 	max_zone_pfns[ZONE_DMA32] = PFN_DOWN(dma32_phys_limit);
83 #endif
84 	max_zone_pfns[ZONE_NORMAL] = max_low_pfn;
85 }
86 
87 #if defined(CONFIG_MMU) && defined(CONFIG_DEBUG_VM)
88 
89 #define LOG2_SZ_1K  ilog2(SZ_1K)
90 #define LOG2_SZ_1M  ilog2(SZ_1M)
91 #define LOG2_SZ_1G  ilog2(SZ_1G)
92 #define LOG2_SZ_1T  ilog2(SZ_1T)
93 
94 static inline void print_mlk(char *name, unsigned long b, unsigned long t)
95 {
96 	pr_notice("%12s : 0x%08lx - 0x%08lx   (%4ld kB)\n", name, b, t,
97 		  (((t) - (b)) >> LOG2_SZ_1K));
98 }
99 
100 static inline void print_mlm(char *name, unsigned long b, unsigned long t)
101 {
102 	pr_notice("%12s : 0x%08lx - 0x%08lx   (%4ld MB)\n", name, b, t,
103 		  (((t) - (b)) >> LOG2_SZ_1M));
104 }
105 
106 static inline void print_mlg(char *name, unsigned long b, unsigned long t)
107 {
108 	pr_notice("%12s : 0x%08lx - 0x%08lx   (%4ld GB)\n", name, b, t,
109 		   (((t) - (b)) >> LOG2_SZ_1G));
110 }
111 
112 #ifdef CONFIG_64BIT
113 static inline void print_mlt(char *name, unsigned long b, unsigned long t)
114 {
115 	pr_notice("%12s : 0x%08lx - 0x%08lx   (%4ld TB)\n", name, b, t,
116 		   (((t) - (b)) >> LOG2_SZ_1T));
117 }
118 #else
119 #define print_mlt(n, b, t) do {} while (0)
120 #endif
121 
122 static inline void print_ml(char *name, unsigned long b, unsigned long t)
123 {
124 	unsigned long diff = t - b;
125 
126 	if (IS_ENABLED(CONFIG_64BIT) && (diff >> LOG2_SZ_1T) >= 10)
127 		print_mlt(name, b, t);
128 	else if ((diff >> LOG2_SZ_1G) >= 10)
129 		print_mlg(name, b, t);
130 	else if ((diff >> LOG2_SZ_1M) >= 10)
131 		print_mlm(name, b, t);
132 	else
133 		print_mlk(name, b, t);
134 }
135 
136 static void __init print_vm_layout(void)
137 {
138 	pr_notice("Virtual kernel memory layout:\n");
139 	print_ml("fixmap", (unsigned long)FIXADDR_START,
140 		(unsigned long)FIXADDR_TOP);
141 	print_ml("pci io", (unsigned long)PCI_IO_START,
142 		(unsigned long)PCI_IO_END);
143 	print_ml("vmemmap", (unsigned long)VMEMMAP_START,
144 		(unsigned long)VMEMMAP_END);
145 	print_ml("vmalloc", (unsigned long)VMALLOC_START,
146 		(unsigned long)VMALLOC_END);
147 #ifdef CONFIG_64BIT
148 	print_ml("modules", (unsigned long)MODULES_VADDR,
149 		(unsigned long)MODULES_END);
150 #endif
151 	print_ml("lowmem", (unsigned long)PAGE_OFFSET,
152 		(unsigned long)high_memory);
153 	if (IS_ENABLED(CONFIG_64BIT)) {
154 #ifdef CONFIG_KASAN
155 		print_ml("kasan", KASAN_SHADOW_START, KASAN_SHADOW_END);
156 #endif
157 
158 		print_ml("kernel", (unsigned long)kernel_map.virt_addr,
159 			 (unsigned long)ADDRESS_SPACE_END);
160 	}
161 }
162 #else
163 static void print_vm_layout(void) { }
164 #endif /* CONFIG_DEBUG_VM */
165 
166 void __init arch_mm_preinit(void)
167 {
168 	bool swiotlb = max_pfn > PFN_DOWN(dma32_phys_limit) &&
169 		       memblock_start_of_DRAM() < dma32_phys_limit;
170 	unsigned int swiotlb_flags = SWIOTLB_VERBOSE;
171 #ifdef CONFIG_FLATMEM
172 	BUG_ON(!mem_map);
173 #endif /* CONFIG_FLATMEM */
174 
175 	if (IS_ENABLED(CONFIG_DMA_BOUNCE_UNALIGNED_KMALLOC) && !swiotlb &&
176 	    dma_cache_alignment != 1) {
177 		/*
178 		 * No 32-bit DMA bouncing needed (either all DRAM is within
179 		 * the 32-bit limit, or it all starts above it), but
180 		 * kmalloc() buffers whose sizes are not cache-line-aligned
181 		 * still require bouncing for non-coherent DMA.  Use
182 		 * SWIOTLB_ANY so that the buffer can be allocated from high
183 		 * memory when DRAM starts above dma32_phys_limit.  Allocate
184 		 * ~1 MB per 1 GB of RAM.
185 		 */
186 		unsigned long size =
187 			DIV_ROUND_UP(memblock_phys_mem_size(), 1024);
188 		swiotlb_adjust_size(min(swiotlb_size_or_default(), size));
189 		swiotlb = true;
190 		swiotlb_flags |= SWIOTLB_ANY;
191 	}
192 
193 	swiotlb_init(swiotlb, swiotlb_flags);
194 
195 	print_vm_layout();
196 }
197 
198 /* Limit the memory size via mem. */
199 static phys_addr_t memory_limit;
200 
201 static int __init early_mem(char *p)
202 {
203 	u64 size;
204 
205 	if (!p)
206 		return 1;
207 
208 	size = memparse(p, &p) & PAGE_MASK;
209 	memory_limit = min_t(u64, size, memory_limit);
210 
211 	pr_notice("Memory limited to %lldMB\n", (u64)memory_limit >> 20);
212 
213 	return 0;
214 }
215 early_param("mem", early_mem);
216 
217 static void __init setup_bootmem(void)
218 {
219 	phys_addr_t vmlinux_end = __pa_symbol(&_end);
220 	phys_addr_t max_mapped_addr;
221 	phys_addr_t phys_ram_end, vmlinux_start;
222 
223 	vmlinux_start = __pa_symbol(&_start);
224 
225 	memblock_enforce_memory_limit(memory_limit);
226 
227 	/*
228 	 * Make sure we align the reservation on PMD_SIZE since we will
229 	 * map the kernel in the linear mapping as read-only: we do not want
230 	 * any allocation to happen between _end and the next pmd aligned page.
231 	 */
232 	if (IS_ENABLED(CONFIG_64BIT) && IS_ENABLED(CONFIG_STRICT_KERNEL_RWX))
233 		vmlinux_end = (vmlinux_end + PMD_SIZE - 1) & PMD_MASK;
234 	/*
235 	 * Reserve from the start of the kernel to the end of the kernel
236 	 */
237 	memblock_reserve(vmlinux_start, vmlinux_end - vmlinux_start);
238 
239 	/*
240 	 * Make sure we align the start of the memory on a PMD boundary so that
241 	 * at worst, we map the linear mapping with PMD mappings.
242 	 */
243 	phys_ram_base = memblock_start_of_DRAM() & PMD_MASK;
244 #ifdef CONFIG_SPARSEMEM_VMEMMAP
245 	vmemmap_start_pfn = round_down(phys_ram_base, VMEMMAP_ADDR_ALIGN) >> PAGE_SHIFT;
246 #endif
247 
248 	/*
249 	 * In 64-bit, any use of __va/__pa before this point is wrong as we
250 	 * did not know the start of DRAM before.
251 	 */
252 	if (IS_ENABLED(CONFIG_64BIT) && IS_ENABLED(CONFIG_MMU))
253 		kernel_map.va_pa_offset = PAGE_OFFSET - phys_ram_base;
254 
255 	/*
256 	 * The size of the linear page mapping may restrict the amount of
257 	 * usable RAM.
258 	 */
259 	if (IS_ENABLED(CONFIG_64BIT) && IS_ENABLED(CONFIG_MMU)) {
260 		max_mapped_addr = __pa(PAGE_OFFSET) + KERN_VIRT_SIZE;
261 		if (memblock_end_of_DRAM() > max_mapped_addr) {
262 			memblock_cap_memory_range(phys_ram_base,
263 						  max_mapped_addr - phys_ram_base);
264 			pr_warn("Physical memory overflows the linear mapping size: region above %pa removed",
265 				&max_mapped_addr);
266 		}
267 	}
268 
269 	/*
270 	 * Reserve physical address space that would be mapped to virtual
271 	 * addresses greater than (void *)(-PAGE_SIZE) because:
272 	 *  - This memory would overlap with ERR_PTR
273 	 *  - This memory belongs to high memory, which is not supported
274 	 *
275 	 * This is not applicable to 64-bit kernel, because virtual addresses
276 	 * after (void *)(-PAGE_SIZE) are not linearly mapped: they are
277 	 * occupied by kernel mapping. Also it is unrealistic for high memory
278 	 * to exist on 64-bit platforms.
279 	 */
280 	if (!IS_ENABLED(CONFIG_64BIT)) {
281 		max_mapped_addr = __va_to_pa_nodebug(-PAGE_SIZE);
282 		memblock_reserve(max_mapped_addr, (phys_addr_t)-max_mapped_addr);
283 	}
284 
285 	phys_ram_end = memblock_end_of_DRAM();
286 	min_low_pfn = PFN_UP(phys_ram_base);
287 	max_low_pfn = max_pfn = PFN_DOWN(phys_ram_end);
288 
289 	dma32_phys_limit = min(4UL * SZ_1G, (unsigned long)PFN_PHYS(max_low_pfn));
290 
291 	reserve_initrd_mem();
292 
293 	/*
294 	 * No allocation should be done before reserving the memory as defined
295 	 * in the device tree, otherwise the allocation could end up in a
296 	 * reserved region.
297 	 */
298 	early_init_fdt_scan_reserved_mem();
299 
300 	/*
301 	 * If DTB is built in, no need to reserve its memblock.
302 	 * Otherwise, do reserve it but avoid using
303 	 * early_init_fdt_reserve_self() since __pa() does
304 	 * not work for DTB pointers that are fixmap addresses
305 	 */
306 	if (!IS_ENABLED(CONFIG_BUILTIN_DTB))
307 		memblock_reserve(dtb_early_pa, fdt_totalsize(dtb_early_va));
308 
309 	dma_contiguous_reserve(dma32_phys_limit);
310 }
311 
312 #ifdef CONFIG_RELOCATABLE
313 extern unsigned long __rela_dyn_start, __rela_dyn_end;
314 
315 static void __init relocate_kernel(void)
316 {
317 	Elf_Rela *rela = (Elf_Rela *)&__rela_dyn_start;
318 	/*
319 	 * This holds the offset between the linked virtual address and the
320 	 * relocated virtual address.
321 	 */
322 	uintptr_t reloc_offset = kernel_map.virt_addr - KERNEL_LINK_ADDR;
323 	/*
324 	 * This holds the offset between kernel linked virtual address and
325 	 * physical address.
326 	 */
327 	uintptr_t va_kernel_link_pa_offset = KERNEL_LINK_ADDR - kernel_map.phys_addr;
328 
329 	for ( ; rela < (Elf_Rela *)&__rela_dyn_end; rela++) {
330 		Elf_Addr addr = (rela->r_offset - va_kernel_link_pa_offset);
331 		Elf_Addr relocated_addr = rela->r_addend;
332 
333 		if (rela->r_info != R_RISCV_RELATIVE)
334 			continue;
335 
336 		/*
337 		 * Make sure to not relocate vdso symbols like rt_sigreturn
338 		 * which are linked from the address 0 in vmlinux since
339 		 * vdso symbol addresses are actually used as an offset from
340 		 * mm->context.vdso in VDSO_OFFSET macro.
341 		 */
342 		if (relocated_addr >= KERNEL_LINK_ADDR)
343 			relocated_addr += reloc_offset;
344 
345 		*(Elf_Addr *)addr = relocated_addr;
346 	}
347 }
348 #endif /* CONFIG_RELOCATABLE */
349 
350 #ifdef CONFIG_MMU
351 struct pt_alloc_ops pt_ops __meminitdata;
352 
353 pgd_t swapper_pg_dir[PTRS_PER_PGD] __page_aligned_bss;
354 pgd_t trampoline_pg_dir[PTRS_PER_PGD] __page_aligned_bss;
355 static pte_t fixmap_pte[PTRS_PER_PTE] __page_aligned_bss;
356 
357 pgd_t early_pg_dir[PTRS_PER_PGD] __initdata __aligned(PAGE_SIZE);
358 
359 static const pgprot_t protection_map[16] = {
360 	[VM_NONE]					= PAGE_NONE,
361 	[VM_READ]					= PAGE_READ,
362 	[VM_WRITE]					= PAGE_SHADOWSTACK,
363 	[VM_WRITE | VM_READ]				= PAGE_COPY,
364 	[VM_EXEC]					= PAGE_EXEC,
365 	[VM_EXEC | VM_READ]				= PAGE_READ_EXEC,
366 	[VM_EXEC | VM_WRITE]				= PAGE_COPY_EXEC,
367 	[VM_EXEC | VM_WRITE | VM_READ]			= PAGE_COPY_EXEC,
368 	[VM_SHARED]					= PAGE_NONE,
369 	[VM_SHARED | VM_READ]				= PAGE_READ,
370 	[VM_SHARED | VM_WRITE]				= PAGE_SHARED,
371 	[VM_SHARED | VM_WRITE | VM_READ]		= PAGE_SHARED,
372 	[VM_SHARED | VM_EXEC]				= PAGE_EXEC,
373 	[VM_SHARED | VM_EXEC | VM_READ]			= PAGE_READ_EXEC,
374 	[VM_SHARED | VM_EXEC | VM_WRITE]		= PAGE_SHARED_EXEC,
375 	[VM_SHARED | VM_EXEC | VM_WRITE | VM_READ]	= PAGE_SHARED_EXEC
376 };
377 DECLARE_VM_GET_PAGE_PROT
378 
379 void __set_fixmap(enum fixed_addresses idx, phys_addr_t phys, pgprot_t prot)
380 {
381 	unsigned long addr = __fix_to_virt(idx);
382 	pte_t *ptep;
383 
384 	BUG_ON(idx <= FIX_HOLE || idx >= __end_of_fixed_addresses);
385 
386 	ptep = &fixmap_pte[pte_index(addr)];
387 
388 	if (pgprot_val(prot))
389 		set_pte(ptep, pfn_pte(phys >> PAGE_SHIFT, prot));
390 	else
391 		pte_clear(&init_mm, addr, ptep);
392 	local_flush_tlb_page(addr);
393 }
394 
395 static inline pte_t *__init get_pte_virt_early(phys_addr_t pa)
396 {
397 	return (pte_t *)((uintptr_t)pa);
398 }
399 
400 static inline pte_t *__init get_pte_virt_fixmap(phys_addr_t pa)
401 {
402 	clear_fixmap(FIX_PTE);
403 	return (pte_t *)set_fixmap_offset(FIX_PTE, pa);
404 }
405 
406 static inline pte_t *__meminit get_pte_virt_late(phys_addr_t pa)
407 {
408 	return (pte_t *) __va(pa);
409 }
410 
411 static inline phys_addr_t __init alloc_pte_early(uintptr_t va)
412 {
413 	/*
414 	 * We only create PMD or PGD early mappings so we
415 	 * should never reach here with MMU disabled.
416 	 */
417 	BUG();
418 }
419 
420 static inline phys_addr_t __init alloc_pte_fixmap(uintptr_t va)
421 {
422 	return memblock_phys_alloc(PAGE_SIZE, PAGE_SIZE);
423 }
424 
425 static phys_addr_t __meminit alloc_pte_late(uintptr_t va)
426 {
427 	struct ptdesc *ptdesc = pagetable_alloc(GFP_KERNEL, 0);
428 
429 	/*
430 	 * We do not know which mm the PTE page is associated to at this point.
431 	 * Passing NULL to the ctor is the safe option, though it may result
432 	 * in unnecessary work (e.g. initialising the ptlock for init_mm).
433 	 */
434 	BUG_ON(!ptdesc || !pagetable_pte_ctor(NULL, ptdesc));
435 	return __pa((pte_t *)ptdesc_address(ptdesc));
436 }
437 
438 static void __meminit create_pte_mapping(pte_t *ptep, uintptr_t va, phys_addr_t pa, phys_addr_t sz,
439 					 pgprot_t prot)
440 {
441 	uintptr_t pte_idx = pte_index(va);
442 
443 	BUG_ON(sz != PAGE_SIZE);
444 
445 	if (pte_none(ptep[pte_idx]))
446 		ptep[pte_idx] = pfn_pte(PFN_DOWN(pa), prot);
447 }
448 
449 #ifndef __PAGETABLE_PMD_FOLDED
450 
451 static pmd_t trampoline_pmd[PTRS_PER_PMD] __page_aligned_bss;
452 static pmd_t fixmap_pmd[PTRS_PER_PMD] __page_aligned_bss;
453 static pmd_t early_pmd[PTRS_PER_PMD] __initdata __aligned(PAGE_SIZE);
454 
455 static p4d_t trampoline_p4d[PTRS_PER_P4D] __page_aligned_bss;
456 static p4d_t fixmap_p4d[PTRS_PER_P4D] __page_aligned_bss;
457 static p4d_t early_p4d[PTRS_PER_P4D] __initdata __aligned(PAGE_SIZE);
458 
459 static pud_t trampoline_pud[PTRS_PER_PUD] __page_aligned_bss;
460 static pud_t fixmap_pud[PTRS_PER_PUD] __page_aligned_bss;
461 static pud_t early_pud[PTRS_PER_PUD] __initdata __aligned(PAGE_SIZE);
462 
463 static pmd_t *__init get_pmd_virt_early(phys_addr_t pa)
464 {
465 	/* Before MMU is enabled */
466 	return (pmd_t *)((uintptr_t)pa);
467 }
468 
469 static pmd_t *__init get_pmd_virt_fixmap(phys_addr_t pa)
470 {
471 	clear_fixmap(FIX_PMD);
472 	return (pmd_t *)set_fixmap_offset(FIX_PMD, pa);
473 }
474 
475 static pmd_t *__meminit get_pmd_virt_late(phys_addr_t pa)
476 {
477 	return (pmd_t *) __va(pa);
478 }
479 
480 static phys_addr_t __init alloc_pmd_early(uintptr_t va)
481 {
482 	BUG_ON((va - kernel_map.virt_addr) >> PUD_SHIFT);
483 
484 	return (uintptr_t)early_pmd;
485 }
486 
487 static phys_addr_t __init alloc_pmd_fixmap(uintptr_t va)
488 {
489 	return memblock_phys_alloc(PAGE_SIZE, PAGE_SIZE);
490 }
491 
492 static phys_addr_t __meminit alloc_pmd_late(uintptr_t va)
493 {
494 	struct ptdesc *ptdesc = pagetable_alloc(GFP_KERNEL, 0);
495 
496 	/* See comment in alloc_pte_late() regarding NULL passed the ctor */
497 	BUG_ON(!ptdesc || !pagetable_pmd_ctor(NULL, ptdesc));
498 	return __pa((pmd_t *)ptdesc_address(ptdesc));
499 }
500 
501 static void __meminit create_pmd_mapping(pmd_t *pmdp,
502 					 uintptr_t va, phys_addr_t pa,
503 					 phys_addr_t sz, pgprot_t prot)
504 {
505 	pte_t *ptep;
506 	phys_addr_t pte_phys;
507 	uintptr_t pmd_idx = pmd_index(va);
508 
509 	if (sz == PMD_SIZE) {
510 		if (pmd_none(pmdp[pmd_idx]))
511 			pmdp[pmd_idx] = pfn_pmd(PFN_DOWN(pa), prot);
512 		return;
513 	}
514 
515 	if (pmd_none(pmdp[pmd_idx])) {
516 		pte_phys = pt_ops.alloc_pte(va);
517 		pmdp[pmd_idx] = pfn_pmd(PFN_DOWN(pte_phys), PAGE_TABLE);
518 		ptep = pt_ops.get_pte_virt(pte_phys);
519 		memset(ptep, 0, PAGE_SIZE);
520 	} else {
521 		pte_phys = PFN_PHYS(_pmd_pfn(pmdp[pmd_idx]));
522 		ptep = pt_ops.get_pte_virt(pte_phys);
523 	}
524 
525 	create_pte_mapping(ptep, va, pa, sz, prot);
526 }
527 
528 static pud_t *__init get_pud_virt_early(phys_addr_t pa)
529 {
530 	return (pud_t *)((uintptr_t)pa);
531 }
532 
533 static pud_t *__init get_pud_virt_fixmap(phys_addr_t pa)
534 {
535 	clear_fixmap(FIX_PUD);
536 	return (pud_t *)set_fixmap_offset(FIX_PUD, pa);
537 }
538 
539 static pud_t *__meminit get_pud_virt_late(phys_addr_t pa)
540 {
541 	return (pud_t *)__va(pa);
542 }
543 
544 static phys_addr_t __init alloc_pud_early(uintptr_t va)
545 {
546 	/* Only one PUD is available for early mapping */
547 	BUG_ON((va - kernel_map.virt_addr) >> PGDIR_SHIFT);
548 
549 	return (uintptr_t)early_pud;
550 }
551 
552 static phys_addr_t __init alloc_pud_fixmap(uintptr_t va)
553 {
554 	return memblock_phys_alloc(PAGE_SIZE, PAGE_SIZE);
555 }
556 
557 static phys_addr_t __meminit alloc_pud_late(uintptr_t va)
558 {
559 	struct ptdesc *ptdesc = pagetable_alloc(GFP_KERNEL, 0);
560 
561 	BUG_ON(!ptdesc);
562 	pagetable_pud_ctor(ptdesc);
563 	return __pa((pud_t *)ptdesc_address(ptdesc));
564 }
565 
566 static p4d_t *__init get_p4d_virt_early(phys_addr_t pa)
567 {
568 	return (p4d_t *)((uintptr_t)pa);
569 }
570 
571 static p4d_t *__init get_p4d_virt_fixmap(phys_addr_t pa)
572 {
573 	clear_fixmap(FIX_P4D);
574 	return (p4d_t *)set_fixmap_offset(FIX_P4D, pa);
575 }
576 
577 static p4d_t *__meminit get_p4d_virt_late(phys_addr_t pa)
578 {
579 	return (p4d_t *)__va(pa);
580 }
581 
582 static phys_addr_t __init alloc_p4d_early(uintptr_t va)
583 {
584 	/* Only one P4D is available for early mapping */
585 	BUG_ON((va - kernel_map.virt_addr) >> PGDIR_SHIFT);
586 
587 	return (uintptr_t)early_p4d;
588 }
589 
590 static phys_addr_t __init alloc_p4d_fixmap(uintptr_t va)
591 {
592 	return memblock_phys_alloc(PAGE_SIZE, PAGE_SIZE);
593 }
594 
595 static phys_addr_t __meminit alloc_p4d_late(uintptr_t va)
596 {
597 	struct ptdesc *ptdesc = pagetable_alloc(GFP_KERNEL, 0);
598 
599 	BUG_ON(!ptdesc);
600 	pagetable_p4d_ctor(ptdesc);
601 	return __pa((p4d_t *)ptdesc_address(ptdesc));
602 }
603 
604 static void __meminit create_pud_mapping(pud_t *pudp, uintptr_t va, phys_addr_t pa, phys_addr_t sz,
605 					 pgprot_t prot)
606 {
607 	pmd_t *nextp;
608 	phys_addr_t next_phys;
609 	uintptr_t pud_index = pud_index(va);
610 
611 	if (sz == PUD_SIZE) {
612 		if (pud_val(pudp[pud_index]) == 0)
613 			pudp[pud_index] = pfn_pud(PFN_DOWN(pa), prot);
614 		return;
615 	}
616 
617 	if (pud_val(pudp[pud_index]) == 0) {
618 		next_phys = pt_ops.alloc_pmd(va);
619 		pudp[pud_index] = pfn_pud(PFN_DOWN(next_phys), PAGE_TABLE);
620 		nextp = pt_ops.get_pmd_virt(next_phys);
621 		memset(nextp, 0, PAGE_SIZE);
622 	} else {
623 		next_phys = PFN_PHYS(_pud_pfn(pudp[pud_index]));
624 		nextp = pt_ops.get_pmd_virt(next_phys);
625 	}
626 
627 	create_pmd_mapping(nextp, va, pa, sz, prot);
628 }
629 
630 static void __meminit create_p4d_mapping(p4d_t *p4dp, uintptr_t va, phys_addr_t pa, phys_addr_t sz,
631 					 pgprot_t prot)
632 {
633 	pud_t *nextp;
634 	phys_addr_t next_phys;
635 	uintptr_t p4d_index = p4d_index(va);
636 
637 	if (sz == P4D_SIZE) {
638 		if (p4d_val(p4dp[p4d_index]) == 0)
639 			p4dp[p4d_index] = pfn_p4d(PFN_DOWN(pa), prot);
640 		return;
641 	}
642 
643 	if (p4d_val(p4dp[p4d_index]) == 0) {
644 		next_phys = pt_ops.alloc_pud(va);
645 		p4dp[p4d_index] = pfn_p4d(PFN_DOWN(next_phys), PAGE_TABLE);
646 		nextp = pt_ops.get_pud_virt(next_phys);
647 		memset(nextp, 0, PAGE_SIZE);
648 	} else {
649 		next_phys = PFN_PHYS(_p4d_pfn(p4dp[p4d_index]));
650 		nextp = pt_ops.get_pud_virt(next_phys);
651 	}
652 
653 	create_pud_mapping(nextp, va, pa, sz, prot);
654 }
655 
656 #define pgd_next_t		p4d_t
657 #define alloc_pgd_next(__va)	(pgtable_l5_enabled ?			\
658 		pt_ops.alloc_p4d(__va) : (pgtable_l4_enabled ?		\
659 		pt_ops.alloc_pud(__va) : pt_ops.alloc_pmd(__va)))
660 #define get_pgd_next_virt(__pa)	(pgtable_l5_enabled ?			\
661 		pt_ops.get_p4d_virt(__pa) : (pgd_next_t *)(pgtable_l4_enabled ?	\
662 		pt_ops.get_pud_virt(__pa) : (pud_t *)pt_ops.get_pmd_virt(__pa)))
663 #define create_pgd_next_mapping(__nextp, __va, __pa, __sz, __prot)	\
664 				(pgtable_l5_enabled ?			\
665 		create_p4d_mapping(__nextp, __va, __pa, __sz, __prot) : \
666 				(pgtable_l4_enabled ?			\
667 		create_pud_mapping((pud_t *)__nextp, __va, __pa, __sz, __prot) :	\
668 		create_pmd_mapping((pmd_t *)__nextp, __va, __pa, __sz, __prot)))
669 #define fixmap_pgd_next		(pgtable_l5_enabled ?			\
670 		(uintptr_t)fixmap_p4d : (pgtable_l4_enabled ?		\
671 		(uintptr_t)fixmap_pud : (uintptr_t)fixmap_pmd))
672 #define trampoline_pgd_next	(pgtable_l5_enabled ?			\
673 		(uintptr_t)trampoline_p4d : (pgtable_l4_enabled ?	\
674 		(uintptr_t)trampoline_pud : (uintptr_t)trampoline_pmd))
675 #else
676 #define pgd_next_t		pte_t
677 #define alloc_pgd_next(__va)	pt_ops.alloc_pte(__va)
678 #define get_pgd_next_virt(__pa)	pt_ops.get_pte_virt(__pa)
679 #define create_pgd_next_mapping(__nextp, __va, __pa, __sz, __prot)	\
680 	create_pte_mapping(__nextp, __va, __pa, __sz, __prot)
681 #define fixmap_pgd_next		((uintptr_t)fixmap_pte)
682 #define create_p4d_mapping(__pmdp, __va, __pa, __sz, __prot) do {} while(0)
683 #define create_pud_mapping(__pmdp, __va, __pa, __sz, __prot) do {} while(0)
684 #define create_pmd_mapping(__pmdp, __va, __pa, __sz, __prot) do {} while(0)
685 #endif /* __PAGETABLE_PMD_FOLDED */
686 
687 void __meminit create_pgd_mapping(pgd_t *pgdp, uintptr_t va, phys_addr_t pa, phys_addr_t sz,
688 				  pgprot_t prot)
689 {
690 	pgd_next_t *nextp;
691 	phys_addr_t next_phys;
692 	uintptr_t pgd_idx = pgd_index(va);
693 
694 	if (sz == PGDIR_SIZE) {
695 		if (pgd_val(pgdp[pgd_idx]) == 0)
696 			pgdp[pgd_idx] = pfn_pgd(PFN_DOWN(pa), prot);
697 		return;
698 	}
699 
700 	if (pgd_val(pgdp[pgd_idx]) == 0) {
701 		next_phys = alloc_pgd_next(va);
702 		pgdp[pgd_idx] = pfn_pgd(PFN_DOWN(next_phys), PAGE_TABLE);
703 		nextp = get_pgd_next_virt(next_phys);
704 		memset(nextp, 0, PAGE_SIZE);
705 	} else {
706 		next_phys = PFN_PHYS(_pgd_pfn(pgdp[pgd_idx]));
707 		nextp = get_pgd_next_virt(next_phys);
708 	}
709 
710 	create_pgd_next_mapping(nextp, va, pa, sz, prot);
711 }
712 
713 static uintptr_t __meminit best_map_size(phys_addr_t pa, uintptr_t va, phys_addr_t size)
714 {
715 	if (debug_pagealloc_enabled())
716 		return PAGE_SIZE;
717 
718 	if (pgtable_l5_enabled &&
719 	    !(pa & (P4D_SIZE - 1)) && !(va & (P4D_SIZE - 1)) && size >= P4D_SIZE)
720 		return P4D_SIZE;
721 
722 	if (pgtable_l4_enabled &&
723 	    !(pa & (PUD_SIZE - 1)) && !(va & (PUD_SIZE - 1)) && size >= PUD_SIZE)
724 		return PUD_SIZE;
725 
726 	if (IS_ENABLED(CONFIG_64BIT) &&
727 	    !(pa & (PMD_SIZE - 1)) && !(va & (PMD_SIZE - 1)) && size >= PMD_SIZE)
728 		return PMD_SIZE;
729 
730 	return PAGE_SIZE;
731 }
732 
733 #ifdef CONFIG_STRICT_KERNEL_RWX
734 static __meminit pgprot_t pgprot_from_va(uintptr_t va)
735 {
736 	if (is_va_kernel_text(va))
737 		return PAGE_KERNEL_READ_EXEC;
738 
739 	/*
740 	 * In 64-bit kernel, the kernel mapping is outside the linear mapping so
741 	 * we must protect its linear mapping alias from being executed and
742 	 * written.
743 	 * And rodata section is marked readonly in mark_rodata_ro.
744 	 */
745 	if (IS_ENABLED(CONFIG_64BIT) && is_va_kernel_lm_alias_text(va))
746 		return PAGE_KERNEL_READ;
747 
748 	return PAGE_KERNEL;
749 }
750 
751 void mark_rodata_ro(void)
752 {
753 	set_kernel_memory(__start_rodata, _data, set_memory_ro);
754 	if (IS_ENABLED(CONFIG_64BIT))
755 		set_kernel_memory(lm_alias(__start_rodata), lm_alias(_data),
756 				  set_memory_ro);
757 }
758 #else
759 static __meminit pgprot_t pgprot_from_va(uintptr_t va)
760 {
761 	if (IS_ENABLED(CONFIG_64BIT) && !is_kernel_mapping(va))
762 		return PAGE_KERNEL;
763 
764 	return PAGE_KERNEL_EXEC;
765 }
766 #endif /* CONFIG_STRICT_KERNEL_RWX */
767 
768 #if defined(CONFIG_64BIT)
769 u64 __pi_set_satp_mode_from_cmdline(uintptr_t dtb_pa);
770 u64 __pi_set_satp_mode_from_fdt(uintptr_t dtb_pa);
771 
772 static void __init disable_pgtable_l5(void)
773 {
774 	pgtable_l5_enabled = false;
775 	kernel_map.page_offset = PAGE_OFFSET_L4;
776 	satp_mode = SATP_MODE_48;
777 }
778 
779 static void __init disable_pgtable_l4(void)
780 {
781 	pgtable_l4_enabled = false;
782 	kernel_map.page_offset = PAGE_OFFSET_L3;
783 	satp_mode = SATP_MODE_39;
784 }
785 
786 static int __init print_no4lvl(char *p)
787 {
788 	pr_info("Disabled 4-level and 5-level paging");
789 	return 0;
790 }
791 early_param("no4lvl", print_no4lvl);
792 
793 static int __init print_no5lvl(char *p)
794 {
795 	pr_info("Disabled 5-level paging");
796 	return 0;
797 }
798 early_param("no5lvl", print_no5lvl);
799 
800 static void __init set_mmap_rnd_bits_max(void)
801 {
802 	mmap_rnd_bits_max = MMAP_VA_BITS - PAGE_SHIFT - 3;
803 }
804 
805 static bool __init is_vaddr_valid(unsigned long va)
806 {
807 	unsigned long up = 0;
808 
809 	switch (satp_mode) {
810 	case SATP_MODE_39:
811 		up = 1UL << 38;
812 		break;
813 	case SATP_MODE_48:
814 		up = 1UL << 47;
815 		break;
816 	case SATP_MODE_57:
817 		up = 1UL << 56;
818 		break;
819 	default:
820 		return false;
821 	}
822 
823 	return (va < up) || (va >= (ULONG_MAX - up + 1));
824 }
825 
826 /*
827  * There is a simple way to determine if 4-level is supported by the
828  * underlying hardware: establish 1:1 mapping in 4-level page table mode
829  * then read SATP to see if the configuration was taken into account
830  * meaning sv48 is supported.
831  * The maximum SATP mode is limited by both the command line and the "mmu-type"
832  * property in the device tree, since some platforms may hang if an unsupported
833  * SATP mode is attempted.
834  */
835 static __init void set_satp_mode(uintptr_t dtb_pa)
836 {
837 	u64 identity_satp, hw_satp;
838 	uintptr_t set_satp_mode_pmd = ((unsigned long)set_satp_mode) & PMD_MASK;
839 	u64 satp_mode_limit = min_not_zero(__pi_set_satp_mode_from_cmdline(dtb_pa),
840 					   __pi_set_satp_mode_from_fdt(dtb_pa));
841 
842 	kernel_map.page_offset = PAGE_OFFSET_L5;
843 
844 	if (satp_mode_limit == SATP_MODE_48) {
845 		disable_pgtable_l5();
846 	} else if (satp_mode_limit == SATP_MODE_39) {
847 		disable_pgtable_l5();
848 		disable_pgtable_l4();
849 		return;
850 	}
851 
852 	create_p4d_mapping(early_p4d,
853 			set_satp_mode_pmd, (uintptr_t)early_pud,
854 			P4D_SIZE, PAGE_TABLE);
855 	create_pud_mapping(early_pud,
856 			   set_satp_mode_pmd, (uintptr_t)early_pmd,
857 			   PUD_SIZE, PAGE_TABLE);
858 	/* Handle the case where set_satp_mode straddles 2 PMDs */
859 	create_pmd_mapping(early_pmd,
860 			   set_satp_mode_pmd, set_satp_mode_pmd,
861 			   PMD_SIZE, PAGE_KERNEL_EXEC);
862 	create_pmd_mapping(early_pmd,
863 			   set_satp_mode_pmd + PMD_SIZE,
864 			   set_satp_mode_pmd + PMD_SIZE,
865 			   PMD_SIZE, PAGE_KERNEL_EXEC);
866 retry:
867 	if (!is_vaddr_valid(set_satp_mode_pmd))
868 		goto out;
869 
870 	create_pgd_mapping(early_pg_dir,
871 			   set_satp_mode_pmd,
872 			   pgtable_l5_enabled ?
873 				(uintptr_t)early_p4d : (uintptr_t)early_pud,
874 			   PGDIR_SIZE, PAGE_TABLE);
875 
876 	identity_satp = PFN_DOWN((uintptr_t)&early_pg_dir) | satp_mode;
877 
878 	local_flush_tlb_all();
879 	csr_write(CSR_SATP, identity_satp);
880 	hw_satp = csr_swap(CSR_SATP, 0ULL);
881 	local_flush_tlb_all();
882 
883 	if (hw_satp != identity_satp) {
884 		if (pgtable_l5_enabled) {
885 			disable_pgtable_l5();
886 			memset(early_pg_dir, 0, PAGE_SIZE);
887 			goto retry;
888 		}
889 		disable_pgtable_l4();
890 	}
891 
892 out:
893 	memset(early_pg_dir, 0, PAGE_SIZE);
894 	memset(early_p4d, 0, PAGE_SIZE);
895 	memset(early_pud, 0, PAGE_SIZE);
896 	memset(early_pmd, 0, PAGE_SIZE);
897 }
898 #endif
899 
900 /*
901  * setup_vm() is called from head.S with MMU-off.
902  *
903  * Following requirements should be honoured for setup_vm() to work
904  * correctly:
905  * 1) It should use PC-relative addressing for accessing kernel symbols.
906  *    To achieve this we always use GCC cmodel=medany.
907  * 2) The compiler instrumentation for FTRACE will not work for setup_vm()
908  *    so disable compiler instrumentation when FTRACE is enabled.
909  *
910  * Currently, the above requirements are honoured by using custom CFLAGS
911  * for init.o in mm/Makefile.
912  */
913 
914 #ifndef __riscv_cmodel_medany
915 #error "setup_vm() is called from head.S before relocate so it should not use absolute addressing."
916 #endif
917 
918 static void __init create_kernel_page_table(pgd_t *pgdir, bool early)
919 {
920 	uintptr_t va, end_va;
921 
922 	end_va = kernel_map.virt_addr + kernel_map.size;
923 	for (va = kernel_map.virt_addr; va < end_va; va += PMD_SIZE)
924 		create_pgd_mapping(pgdir, va,
925 				   kernel_map.phys_addr + (va - kernel_map.virt_addr),
926 				   PMD_SIZE,
927 				   early ?
928 					PAGE_KERNEL_EXEC : pgprot_from_va(va));
929 }
930 
931 /*
932  * Setup a 4MB mapping that encompasses the device tree: for 64-bit kernel,
933  * this means 2 PMD entries whereas for 32-bit kernel, this is only 1 PGDIR
934  * entry.
935  */
936 static void __init create_fdt_early_page_table(uintptr_t fix_fdt_va,
937 					       uintptr_t dtb_pa)
938 {
939 #ifndef CONFIG_BUILTIN_DTB
940 	uintptr_t pa = dtb_pa & ~(PMD_SIZE - 1);
941 
942 	/* Make sure the fdt fixmap address is always aligned on PMD size */
943 	BUILD_BUG_ON(FIX_FDT % (PMD_SIZE / PAGE_SIZE));
944 
945 	/* In 32-bit only, the fdt lies in its own PGD */
946 	if (!IS_ENABLED(CONFIG_64BIT)) {
947 		create_pgd_mapping(early_pg_dir, fix_fdt_va,
948 				   pa, MAX_FDT_SIZE, PAGE_KERNEL);
949 	} else {
950 		create_pmd_mapping(fixmap_pmd, fix_fdt_va,
951 				   pa, PMD_SIZE, PAGE_KERNEL);
952 		create_pmd_mapping(fixmap_pmd, fix_fdt_va + PMD_SIZE,
953 				   pa + PMD_SIZE, PMD_SIZE, PAGE_KERNEL);
954 	}
955 
956 	dtb_early_va = (void *)fix_fdt_va + (dtb_pa & (PMD_SIZE - 1));
957 #else
958 	/*
959 	 * For 64-bit kernel, __va can't be used since it would return a linear
960 	 * mapping address whereas dtb_early_va will be used before
961 	 * setup_vm_final installs the linear mapping. For 32-bit kernel, as the
962 	 * kernel is mapped in the linear mapping, that makes no difference.
963 	 */
964 	dtb_early_va = kernel_mapping_pa_to_va(dtb_pa);
965 #endif
966 
967 	dtb_early_pa = dtb_pa;
968 }
969 
970 /*
971  * MMU is not enabled, the page tables are allocated directly using
972  * early_pmd/pud/p4d and the address returned is the physical one.
973  */
974 static void __init pt_ops_set_early(void)
975 {
976 	pt_ops.alloc_pte = alloc_pte_early;
977 	pt_ops.get_pte_virt = get_pte_virt_early;
978 #ifndef __PAGETABLE_PMD_FOLDED
979 	pt_ops.alloc_pmd = alloc_pmd_early;
980 	pt_ops.get_pmd_virt = get_pmd_virt_early;
981 	pt_ops.alloc_pud = alloc_pud_early;
982 	pt_ops.get_pud_virt = get_pud_virt_early;
983 	pt_ops.alloc_p4d = alloc_p4d_early;
984 	pt_ops.get_p4d_virt = get_p4d_virt_early;
985 #endif
986 }
987 
988 /*
989  * MMU is enabled but page table setup is not complete yet.
990  * fixmap page table alloc functions must be used as a means to temporarily
991  * map the allocated physical pages since the linear mapping does not exist yet.
992  *
993  * Note that this is called with MMU disabled, hence kernel_mapping_pa_to_va,
994  * but it will be used as described above.
995  */
996 static void __init pt_ops_set_fixmap(void)
997 {
998 	pt_ops.alloc_pte = kernel_mapping_pa_to_va(alloc_pte_fixmap);
999 	pt_ops.get_pte_virt = kernel_mapping_pa_to_va(get_pte_virt_fixmap);
1000 #ifndef __PAGETABLE_PMD_FOLDED
1001 	pt_ops.alloc_pmd = kernel_mapping_pa_to_va(alloc_pmd_fixmap);
1002 	pt_ops.get_pmd_virt = kernel_mapping_pa_to_va(get_pmd_virt_fixmap);
1003 	pt_ops.alloc_pud = kernel_mapping_pa_to_va(alloc_pud_fixmap);
1004 	pt_ops.get_pud_virt = kernel_mapping_pa_to_va(get_pud_virt_fixmap);
1005 	pt_ops.alloc_p4d = kernel_mapping_pa_to_va(alloc_p4d_fixmap);
1006 	pt_ops.get_p4d_virt = kernel_mapping_pa_to_va(get_p4d_virt_fixmap);
1007 #endif
1008 }
1009 
1010 /*
1011  * MMU is enabled and page table setup is complete, so from now, we can use
1012  * generic page allocation functions to setup page table.
1013  */
1014 static void __init pt_ops_set_late(void)
1015 {
1016 	pt_ops.alloc_pte = alloc_pte_late;
1017 	pt_ops.get_pte_virt = get_pte_virt_late;
1018 #ifndef __PAGETABLE_PMD_FOLDED
1019 	pt_ops.alloc_pmd = alloc_pmd_late;
1020 	pt_ops.get_pmd_virt = get_pmd_virt_late;
1021 	pt_ops.alloc_pud = alloc_pud_late;
1022 	pt_ops.get_pud_virt = get_pud_virt_late;
1023 	pt_ops.alloc_p4d = alloc_p4d_late;
1024 	pt_ops.get_p4d_virt = get_p4d_virt_late;
1025 #endif
1026 }
1027 
1028 #ifdef CONFIG_RANDOMIZE_BASE
1029 extern bool __init __pi_set_nokaslr_from_cmdline(uintptr_t dtb_pa);
1030 extern u64 __init __pi_get_kaslr_seed(uintptr_t dtb_pa);
1031 extern u64 __init __pi_get_kaslr_seed_zkr(const uintptr_t dtb_pa);
1032 
1033 static int __init print_nokaslr(char *p)
1034 {
1035 	pr_info("Disabled KASLR");
1036 	return 0;
1037 }
1038 early_param("nokaslr", print_nokaslr);
1039 #endif
1040 
1041 asmlinkage void __init setup_vm(uintptr_t dtb_pa)
1042 {
1043 	pmd_t __maybe_unused fix_bmap_spmd, fix_bmap_epmd;
1044 
1045 #ifdef CONFIG_RANDOMIZE_BASE
1046 	if (!__pi_set_nokaslr_from_cmdline(dtb_pa)) {
1047 		u64 kaslr_seed = __pi_get_kaslr_seed_zkr(dtb_pa);
1048 		u32 kernel_size = (uintptr_t)(&_end) - (uintptr_t)(&_start);
1049 		u32 nr_pos;
1050 
1051 		if (kaslr_seed == 0)
1052 			kaslr_seed = __pi_get_kaslr_seed(dtb_pa);
1053 		/*
1054 		 * Compute the number of positions available: we are limited
1055 		 * by the early page table that only has one PUD and we must
1056 		 * be aligned on PMD_SIZE.
1057 		 */
1058 		nr_pos = (PUD_SIZE - kernel_size) / PMD_SIZE;
1059 
1060 		kernel_map.virt_offset = (kaslr_seed % nr_pos) * PMD_SIZE;
1061 	}
1062 #endif
1063 
1064 	kernel_map.virt_addr = KERNEL_LINK_ADDR + kernel_map.virt_offset;
1065 
1066 	kernel_map.phys_addr = (uintptr_t)(&_start);
1067 	kernel_map.size = (uintptr_t)(&_end) - kernel_map.phys_addr;
1068 	kernel_map.va_kernel_pa_offset = kernel_map.virt_addr - kernel_map.phys_addr;
1069 
1070 #if defined(CONFIG_64BIT)
1071 	set_satp_mode(dtb_pa);
1072 	set_mmap_rnd_bits_max();
1073 #endif
1074 
1075 	/*
1076 	 * In 64-bit, we defer the setup of va_pa_offset to setup_bootmem,
1077 	 * where we have the system memory layout: this allows us to align
1078 	 * the physical and virtual mappings and then make use of PUD/P4D/PGD
1079 	 * for the linear mapping. This is only possible because the kernel
1080 	 * mapping lies outside the linear mapping.
1081 	 * In 32-bit however, as the kernel resides in the linear mapping,
1082 	 * setup_vm_final can not change the mapping established here,
1083 	 * otherwise the same kernel addresses would get mapped to different
1084 	 * physical addresses (if the start of dram is different from the
1085 	 * kernel physical address start).
1086 	 */
1087 	kernel_map.va_pa_offset = IS_ENABLED(CONFIG_64BIT) ?
1088 				0UL : PAGE_OFFSET - kernel_map.phys_addr;
1089 
1090 	memory_limit = KERN_VIRT_SIZE;
1091 
1092 	/* Sanity check alignment and size */
1093 	BUG_ON((PAGE_OFFSET % PGDIR_SIZE) != 0);
1094 	BUG_ON((kernel_map.phys_addr % PMD_SIZE) != 0);
1095 
1096 #ifdef CONFIG_64BIT
1097 	/*
1098 	 * The last 4K bytes of the addressable memory can not be mapped because
1099 	 * of IS_ERR_VALUE macro.
1100 	 */
1101 	BUG_ON((kernel_map.virt_addr + kernel_map.size) > ADDRESS_SPACE_END - SZ_4K);
1102 #endif
1103 
1104 #ifdef CONFIG_RELOCATABLE
1105 	/*
1106 	 * Early page table uses only one PUD, which makes it possible
1107 	 * to map PUD_SIZE aligned on PUD_SIZE: if the relocation offset
1108 	 * makes the kernel cross over a PUD_SIZE boundary, raise a bug
1109 	 * since a part of the kernel would not get mapped.
1110 	 */
1111 	if (IS_ENABLED(CONFIG_64BIT))
1112 		BUG_ON(PUD_SIZE - (kernel_map.virt_addr & (PUD_SIZE - 1)) < kernel_map.size);
1113 	relocate_kernel();
1114 #endif
1115 
1116 	apply_early_boot_alternatives();
1117 	pt_ops_set_early();
1118 
1119 	/* Setup early PGD for fixmap */
1120 	create_pgd_mapping(early_pg_dir, FIXADDR_START,
1121 			   fixmap_pgd_next, PGDIR_SIZE, PAGE_TABLE);
1122 
1123 #ifndef __PAGETABLE_PMD_FOLDED
1124 	/* Setup fixmap P4D and PUD */
1125 	if (pgtable_l5_enabled)
1126 		create_p4d_mapping(fixmap_p4d, FIXADDR_START,
1127 				   (uintptr_t)fixmap_pud, P4D_SIZE, PAGE_TABLE);
1128 	/* Setup fixmap PUD and PMD */
1129 	if (pgtable_l4_enabled)
1130 		create_pud_mapping(fixmap_pud, FIXADDR_START,
1131 				   (uintptr_t)fixmap_pmd, PUD_SIZE, PAGE_TABLE);
1132 	create_pmd_mapping(fixmap_pmd, FIXADDR_START,
1133 			   (uintptr_t)fixmap_pte, PMD_SIZE, PAGE_TABLE);
1134 	/* Setup trampoline PGD and PMD */
1135 	create_pgd_mapping(trampoline_pg_dir, kernel_map.virt_addr,
1136 			   trampoline_pgd_next, PGDIR_SIZE, PAGE_TABLE);
1137 	if (pgtable_l5_enabled)
1138 		create_p4d_mapping(trampoline_p4d, kernel_map.virt_addr,
1139 				   (uintptr_t)trampoline_pud, P4D_SIZE, PAGE_TABLE);
1140 	if (pgtable_l4_enabled)
1141 		create_pud_mapping(trampoline_pud, kernel_map.virt_addr,
1142 				   (uintptr_t)trampoline_pmd, PUD_SIZE, PAGE_TABLE);
1143 	create_pmd_mapping(trampoline_pmd, kernel_map.virt_addr,
1144 			   kernel_map.phys_addr, PMD_SIZE, PAGE_KERNEL_EXEC);
1145 #else
1146 	/* Setup trampoline PGD */
1147 	create_pgd_mapping(trampoline_pg_dir, kernel_map.virt_addr,
1148 			   kernel_map.phys_addr, PGDIR_SIZE, PAGE_KERNEL_EXEC);
1149 #endif
1150 
1151 	/*
1152 	 * Setup early PGD covering entire kernel which will allow
1153 	 * us to reach paging_init(). We map all memory banks later
1154 	 * in setup_vm_final() below.
1155 	 */
1156 	create_kernel_page_table(early_pg_dir, true);
1157 
1158 	/* Setup early mapping for FDT early scan */
1159 	create_fdt_early_page_table(__fix_to_virt(FIX_FDT), dtb_pa);
1160 
1161 	/*
1162 	 * Bootime fixmap only can handle PMD_SIZE mapping. Thus, boot-ioremap
1163 	 * range can not span multiple pmds.
1164 	 */
1165 	BUG_ON((__fix_to_virt(FIX_BTMAP_BEGIN) >> PMD_SHIFT)
1166 		     != (__fix_to_virt(FIX_BTMAP_END) >> PMD_SHIFT));
1167 
1168 #ifndef __PAGETABLE_PMD_FOLDED
1169 	/*
1170 	 * Early ioremap fixmap is already created as it lies within first 2MB
1171 	 * of fixmap region. We always map PMD_SIZE. Thus, both FIX_BTMAP_END
1172 	 * FIX_BTMAP_BEGIN should lie in the same pmd. Verify that and warn
1173 	 * the user if not.
1174 	 */
1175 	fix_bmap_spmd = fixmap_pmd[pmd_index(__fix_to_virt(FIX_BTMAP_BEGIN))];
1176 	fix_bmap_epmd = fixmap_pmd[pmd_index(__fix_to_virt(FIX_BTMAP_END))];
1177 	if (pmd_val(fix_bmap_spmd) != pmd_val(fix_bmap_epmd)) {
1178 		WARN_ON(1);
1179 		pr_warn("fixmap btmap start [%08lx] != end [%08lx]\n",
1180 			pmd_val(fix_bmap_spmd), pmd_val(fix_bmap_epmd));
1181 		pr_warn("fix_to_virt(FIX_BTMAP_BEGIN): %08lx\n",
1182 			fix_to_virt(FIX_BTMAP_BEGIN));
1183 		pr_warn("fix_to_virt(FIX_BTMAP_END):   %08lx\n",
1184 			fix_to_virt(FIX_BTMAP_END));
1185 
1186 		pr_warn("FIX_BTMAP_END:       %d\n", FIX_BTMAP_END);
1187 		pr_warn("FIX_BTMAP_BEGIN:     %d\n", FIX_BTMAP_BEGIN);
1188 	}
1189 #endif
1190 
1191 	pt_ops_set_fixmap();
1192 }
1193 
1194 static void __meminit create_linear_mapping_range(phys_addr_t start, phys_addr_t end,
1195 						  uintptr_t fixed_map_size, const pgprot_t *pgprot)
1196 {
1197 	phys_addr_t pa;
1198 	uintptr_t va, map_size;
1199 
1200 	for (pa = start; pa < end; pa += map_size) {
1201 		va = (uintptr_t)__va(pa);
1202 		map_size = fixed_map_size ? fixed_map_size :
1203 					    best_map_size(pa, va, end - pa);
1204 
1205 		create_pgd_mapping(swapper_pg_dir, va, pa, map_size,
1206 				   pgprot ? *pgprot : pgprot_from_va(va));
1207 	}
1208 }
1209 
1210 static void __init create_linear_mapping_page_table(void)
1211 {
1212 	phys_addr_t start, end;
1213 	phys_addr_t kfence_pool __maybe_unused;
1214 	u64 i;
1215 
1216 #ifdef CONFIG_STRICT_KERNEL_RWX
1217 	phys_addr_t ktext_start = __pa_symbol(_start);
1218 	phys_addr_t ktext_size = __init_data_begin - _start;
1219 	phys_addr_t krodata_start = __pa_symbol(__start_rodata);
1220 	phys_addr_t krodata_size = _data - __start_rodata;
1221 
1222 	/* Isolate kernel text and rodata so they don't get mapped with a PUD */
1223 	memblock_mark_nomap(ktext_start,  ktext_size);
1224 	memblock_mark_nomap(krodata_start, krodata_size);
1225 #endif
1226 
1227 #ifdef CONFIG_KFENCE
1228 	/*
1229 	 *  kfence pool must be backed by PAGE_SIZE mappings, so allocate it
1230 	 *  before we setup the linear mapping so that we avoid using hugepages
1231 	 *  for this region.
1232 	 */
1233 	kfence_pool = memblock_phys_alloc(KFENCE_POOL_SIZE, PAGE_SIZE);
1234 	BUG_ON(!kfence_pool);
1235 
1236 	memblock_mark_nomap(kfence_pool, KFENCE_POOL_SIZE);
1237 	__kfence_pool = __va(kfence_pool);
1238 #endif
1239 
1240 	/* Map all memory banks in the linear mapping */
1241 	for_each_mem_range(i, &start, &end) {
1242 		if (start >= end)
1243 			break;
1244 		if (start <= __pa(PAGE_OFFSET) &&
1245 		    __pa(PAGE_OFFSET) < end)
1246 			start = __pa(PAGE_OFFSET);
1247 
1248 		create_linear_mapping_range(start, end, 0, NULL);
1249 	}
1250 
1251 #ifdef CONFIG_STRICT_KERNEL_RWX
1252 	create_linear_mapping_range(ktext_start, ktext_start + ktext_size, 0, NULL);
1253 	create_linear_mapping_range(krodata_start, krodata_start + krodata_size, 0, NULL);
1254 
1255 	memblock_clear_nomap(ktext_start,  ktext_size);
1256 	memblock_clear_nomap(krodata_start, krodata_size);
1257 #endif
1258 
1259 #ifdef CONFIG_KFENCE
1260 	create_linear_mapping_range(kfence_pool, kfence_pool + KFENCE_POOL_SIZE, PAGE_SIZE, NULL);
1261 
1262 	memblock_clear_nomap(kfence_pool, KFENCE_POOL_SIZE);
1263 #endif
1264 }
1265 
1266 static void __init setup_vm_final(void)
1267 {
1268 	/* Setup swapper PGD for fixmap */
1269 #if !defined(CONFIG_64BIT)
1270 	/*
1271 	 * In 32-bit, the device tree lies in a pgd entry, so it must be copied
1272 	 * directly in swapper_pg_dir in addition to the pgd entry that points
1273 	 * to fixmap_pte.
1274 	 */
1275 	unsigned long idx = pgd_index(__fix_to_virt(FIX_FDT));
1276 
1277 	set_pgd(&swapper_pg_dir[idx], early_pg_dir[idx]);
1278 #endif
1279 	create_pgd_mapping(swapper_pg_dir, FIXADDR_START,
1280 			   __pa_symbol(fixmap_pgd_next),
1281 			   PGDIR_SIZE, PAGE_TABLE);
1282 
1283 	/* Map the linear mapping */
1284 	create_linear_mapping_page_table();
1285 
1286 	/* Map the kernel */
1287 	if (IS_ENABLED(CONFIG_64BIT))
1288 		create_kernel_page_table(swapper_pg_dir, false);
1289 
1290 #ifdef CONFIG_KASAN
1291 	kasan_swapper_init();
1292 #endif
1293 
1294 	/* Clear fixmap PTE and PMD mappings */
1295 	clear_fixmap(FIX_PTE);
1296 	clear_fixmap(FIX_PMD);
1297 	clear_fixmap(FIX_PUD);
1298 	clear_fixmap(FIX_P4D);
1299 
1300 	/* Move to swapper page table */
1301 	csr_write(CSR_SATP, PFN_DOWN(__pa_symbol(swapper_pg_dir)) | satp_mode);
1302 	local_flush_tlb_all();
1303 
1304 	pt_ops_set_late();
1305 }
1306 #else
1307 asmlinkage void __init setup_vm(uintptr_t dtb_pa)
1308 {
1309 	dtb_early_va = (void *)dtb_pa;
1310 	dtb_early_pa = dtb_pa;
1311 
1312 #ifdef CONFIG_RELOCATABLE
1313 	kernel_map.virt_addr = (uintptr_t)_start;
1314 	kernel_map.phys_addr = (uintptr_t)_start;
1315 	relocate_kernel();
1316 #endif
1317 }
1318 
1319 static inline void setup_vm_final(void)
1320 {
1321 }
1322 #endif /* CONFIG_MMU */
1323 
1324 /*
1325  * reserve_crashkernel() - reserves memory for crash kernel
1326  *
1327  * This function reserves memory area given in "crashkernel=" kernel command
1328  * line parameter. The memory reserved is used by dump capture kernel when
1329  * primary kernel is crashing.
1330  */
1331 static void __init arch_reserve_crashkernel(void)
1332 {
1333 	unsigned long long low_size = 0;
1334 	unsigned long long crash_base, crash_size;
1335 	bool high = false;
1336 	int ret;
1337 
1338 	if (!IS_ENABLED(CONFIG_CRASH_RESERVE))
1339 		return;
1340 
1341 	ret = parse_crashkernel(boot_command_line, memblock_phys_mem_size(),
1342 				&crash_size, &crash_base,
1343 				&low_size, NULL, &high);
1344 	if (ret)
1345 		return;
1346 
1347 	reserve_crashkernel_generic(crash_size, crash_base, low_size, high);
1348 }
1349 
1350 void __init paging_init(void)
1351 {
1352 	setup_bootmem();
1353 	setup_vm_final();
1354 
1355 	/* Depend on that Linear Mapping is ready */
1356 	memblock_allow_resize();
1357 }
1358 
1359 void __init misc_mem_init(void)
1360 {
1361 	early_memtest(min_low_pfn << PAGE_SHIFT, max_low_pfn << PAGE_SHIFT);
1362 	arch_numa_init();
1363 #ifdef CONFIG_SPARSEMEM_VMEMMAP
1364 	/* The entire VMEMMAP region has been populated. Flush TLB for this region */
1365 	local_flush_tlb_kernel_range(VMEMMAP_START, VMEMMAP_END);
1366 #endif
1367 	arch_reserve_crashkernel();
1368 	memblock_dump_all();
1369 }
1370 
1371 #ifdef CONFIG_SPARSEMEM_VMEMMAP
1372 int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node,
1373 			       struct vmem_altmap *altmap)
1374 {
1375 	WARN_ON((start < VMEMMAP_START) || (end > VMEMMAP_END));
1376 
1377 	/*
1378 	 * Note that SPARSEMEM_VMEMMAP is only selected for rv64 and that we
1379 	 * can't use hugepage mappings for 2-level page table because in case of
1380 	 * memory hotplug, we are not able to update all the page tables with
1381 	 * the new PMDs.
1382 	 */
1383 	return vmemmap_populate_hugepages(start, end, node, altmap);
1384 }
1385 #endif
1386 
1387 #if defined(CONFIG_MMU) && defined(CONFIG_64BIT)
1388 /*
1389  * Pre-allocates page-table pages for a specific area in the kernel
1390  * page-table. Only the level which needs to be synchronized between
1391  * all page-tables is allocated because the synchronization can be
1392  * expensive.
1393  */
1394 static void __init preallocate_pgd_pages_range(unsigned long start, unsigned long end,
1395 					       const char *area)
1396 {
1397 	unsigned long addr;
1398 	const char *lvl;
1399 
1400 	for (addr = start; addr < end && addr >= start; addr = ALIGN(addr + 1, PGDIR_SIZE)) {
1401 		pgd_t *pgd = pgd_offset_k(addr);
1402 		p4d_t *p4d;
1403 		pud_t *pud;
1404 		pmd_t *pmd;
1405 
1406 		lvl = "p4d";
1407 		p4d = p4d_alloc(&init_mm, pgd, addr);
1408 		if (!p4d)
1409 			goto failed;
1410 
1411 		if (pgtable_l5_enabled)
1412 			continue;
1413 
1414 		lvl = "pud";
1415 		pud = pud_alloc(&init_mm, p4d, addr);
1416 		if (!pud)
1417 			goto failed;
1418 
1419 		if (pgtable_l4_enabled)
1420 			continue;
1421 
1422 		lvl = "pmd";
1423 		pmd = pmd_alloc(&init_mm, pud, addr);
1424 		if (!pmd)
1425 			goto failed;
1426 	}
1427 	return;
1428 
1429 failed:
1430 	/*
1431 	 * The pages have to be there now or they will be missing in
1432 	 * process page-tables later.
1433 	 */
1434 	panic("Failed to pre-allocate %s pages for %s area\n", lvl, area);
1435 }
1436 
1437 #define PAGE_END KASAN_SHADOW_START
1438 
1439 void __init pgtable_cache_init(void)
1440 {
1441 	preallocate_pgd_pages_range(VMALLOC_START, VMALLOC_END, "vmalloc");
1442 	if (IS_ENABLED(CONFIG_MODULES))
1443 		preallocate_pgd_pages_range(MODULES_VADDR, MODULES_END, "bpf/modules");
1444 	if (IS_ENABLED(CONFIG_MEMORY_HOTPLUG)) {
1445 		preallocate_pgd_pages_range(VMEMMAP_START, VMEMMAP_END, "vmemmap");
1446 		preallocate_pgd_pages_range(PAGE_OFFSET, PAGE_END, "direct map");
1447 		if (IS_ENABLED(CONFIG_KASAN))
1448 			preallocate_pgd_pages_range(KASAN_SHADOW_START, KASAN_SHADOW_END, "kasan");
1449 	}
1450 }
1451 #endif
1452 
1453 #ifdef CONFIG_EXECMEM
1454 #ifdef CONFIG_MMU
1455 static struct execmem_info execmem_info __ro_after_init;
1456 
1457 struct execmem_info __init *execmem_arch_setup(void)
1458 {
1459 	execmem_info = (struct execmem_info){
1460 		.ranges = {
1461 			[EXECMEM_DEFAULT] = {
1462 				.start	= MODULES_VADDR,
1463 				.end	= MODULES_END,
1464 				.pgprot	= PAGE_KERNEL,
1465 				.alignment = 1,
1466 			},
1467 			[EXECMEM_KPROBES] = {
1468 				.start	= VMALLOC_START,
1469 				.end	= VMALLOC_END,
1470 				.pgprot	= PAGE_KERNEL_READ_EXEC,
1471 				.alignment = 1,
1472 			},
1473 			[EXECMEM_BPF] = {
1474 				.start	= BPF_JIT_REGION_START,
1475 				.end	= BPF_JIT_REGION_END,
1476 				.pgprot	= PAGE_KERNEL,
1477 				.alignment = PAGE_SIZE,
1478 			},
1479 		},
1480 	};
1481 
1482 	return &execmem_info;
1483 }
1484 #endif /* CONFIG_MMU */
1485 #endif /* CONFIG_EXECMEM */
1486 
1487 #ifdef CONFIG_MEMORY_HOTPLUG
1488 static void __meminit free_pte_table(pte_t *pte_start, pmd_t *pmd)
1489 {
1490 	struct page *page = pmd_page(*pmd);
1491 	struct ptdesc *ptdesc = page_ptdesc(page);
1492 	pte_t *pte;
1493 	int i;
1494 
1495 	for (i = 0; i < PTRS_PER_PTE; i++) {
1496 		pte = pte_start + i;
1497 		if (!pte_none(*pte))
1498 			return;
1499 	}
1500 
1501 	pagetable_dtor(ptdesc);
1502 	if (PageReserved(page))
1503 		free_reserved_page(page);
1504 	else
1505 		pagetable_free(ptdesc);
1506 	pmd_clear(pmd);
1507 }
1508 
1509 static void __meminit free_pmd_table(pmd_t *pmd_start, pud_t *pud, bool is_vmemmap)
1510 {
1511 	struct page *page = pud_page(*pud);
1512 	struct ptdesc *ptdesc = page_ptdesc(page);
1513 	pmd_t *pmd;
1514 	int i;
1515 
1516 	for (i = 0; i < PTRS_PER_PMD; i++) {
1517 		pmd = pmd_start + i;
1518 		if (!pmd_none(*pmd))
1519 			return;
1520 	}
1521 
1522 	if (!is_vmemmap)
1523 		pagetable_dtor(ptdesc);
1524 	if (PageReserved(page))
1525 		free_reserved_page(page);
1526 	else
1527 		pagetable_free(ptdesc);
1528 	pud_clear(pud);
1529 }
1530 
1531 static void __meminit free_pud_table(pud_t *pud_start, p4d_t *p4d)
1532 {
1533 	struct page *page = p4d_page(*p4d);
1534 	pud_t *pud;
1535 	int i;
1536 
1537 	for (i = 0; i < PTRS_PER_PUD; i++) {
1538 		pud = pud_start + i;
1539 		if (!pud_none(*pud))
1540 			return;
1541 	}
1542 
1543 	if (PageReserved(page))
1544 		free_reserved_page(page);
1545 	else
1546 		__free_pages(page, 0);
1547 	p4d_clear(p4d);
1548 }
1549 
1550 static void __meminit free_vmemmap_storage(struct page *page, size_t size,
1551 					   struct vmem_altmap *altmap)
1552 {
1553 	int order = get_order(size);
1554 
1555 	if (altmap) {
1556 		vmem_altmap_free(altmap, size >> PAGE_SHIFT);
1557 		return;
1558 	}
1559 
1560 	if (PageReserved(page)) {
1561 		unsigned int nr_pages = 1 << order;
1562 
1563 		while (nr_pages--)
1564 			free_reserved_page(page++);
1565 		return;
1566 	}
1567 
1568 	__free_pages(page, order);
1569 }
1570 
1571 static void __meminit remove_pte_mapping(pte_t *pte_base, unsigned long addr, unsigned long end,
1572 					 bool is_vmemmap, struct vmem_altmap *altmap)
1573 {
1574 	unsigned long next;
1575 	pte_t *ptep, pte;
1576 
1577 	for (; addr < end; addr = next) {
1578 		next = (addr + PAGE_SIZE) & PAGE_MASK;
1579 		if (next > end)
1580 			next = end;
1581 
1582 		ptep = pte_base + pte_index(addr);
1583 		pte = ptep_get(ptep);
1584 		if (!pte_present(*ptep))
1585 			continue;
1586 
1587 		pte_clear(&init_mm, addr, ptep);
1588 		if (is_vmemmap)
1589 			free_vmemmap_storage(pte_page(pte), PAGE_SIZE, altmap);
1590 	}
1591 }
1592 
1593 static void __meminit remove_pmd_mapping(pmd_t *pmd_base, unsigned long addr, unsigned long end,
1594 					 bool is_vmemmap, struct vmem_altmap *altmap)
1595 {
1596 	unsigned long next;
1597 	pte_t *pte_base;
1598 	pmd_t *pmdp, pmd;
1599 
1600 	for (; addr < end; addr = next) {
1601 		next = pmd_addr_end(addr, end);
1602 		pmdp = pmd_base + pmd_index(addr);
1603 		pmd = pmdp_get(pmdp);
1604 		if (!pmd_present(pmd))
1605 			continue;
1606 
1607 		if (pmd_leaf(pmd)) {
1608 			pmd_clear(pmdp);
1609 			if (is_vmemmap)
1610 				free_vmemmap_storage(pmd_page(pmd), PMD_SIZE, altmap);
1611 			continue;
1612 		}
1613 
1614 		pte_base = (pte_t *)pmd_page_vaddr(*pmdp);
1615 		remove_pte_mapping(pte_base, addr, next, is_vmemmap, altmap);
1616 		free_pte_table(pte_base, pmdp);
1617 	}
1618 }
1619 
1620 static void __meminit remove_pud_mapping(pud_t *pud_base, unsigned long addr, unsigned long end,
1621 					 bool is_vmemmap, struct vmem_altmap *altmap)
1622 {
1623 	unsigned long next;
1624 	pud_t *pudp, pud;
1625 	pmd_t *pmd_base;
1626 
1627 	for (; addr < end; addr = next) {
1628 		next = pud_addr_end(addr, end);
1629 		pudp = pgtable_l4_enabled ? pud_base + pud_index(addr) : pud_base;
1630 		pud = pudp_get(pudp);
1631 		if (!pud_present(pud))
1632 			continue;
1633 
1634 		if (pud_leaf(pud)) {
1635 			if (pgtable_l4_enabled) {
1636 				pud_clear(pudp);
1637 				if (is_vmemmap)
1638 					free_vmemmap_storage(pud_page(pud), PUD_SIZE, altmap);
1639 			}
1640 			continue;
1641 		}
1642 
1643 		pmd_base = pmd_offset(pudp, 0);
1644 		remove_pmd_mapping(pmd_base, addr, next, is_vmemmap, altmap);
1645 
1646 		if (pgtable_l4_enabled)
1647 			free_pmd_table(pmd_base, pudp, is_vmemmap);
1648 	}
1649 }
1650 
1651 static void __meminit remove_p4d_mapping(p4d_t *p4d_base, unsigned long addr, unsigned long end,
1652 					 bool is_vmemmap, struct vmem_altmap *altmap)
1653 {
1654 	unsigned long next;
1655 	p4d_t *p4dp, p4d;
1656 	pud_t *pud_base;
1657 
1658 	for (; addr < end; addr = next) {
1659 		next = p4d_addr_end(addr, end);
1660 		p4dp = pgtable_l5_enabled ? p4d_base + p4d_index(addr) : p4d_base;
1661 		p4d = p4dp_get(p4dp);
1662 		if (!p4d_present(p4d))
1663 			continue;
1664 
1665 		if (p4d_leaf(p4d)) {
1666 			if (pgtable_l5_enabled) {
1667 				p4d_clear(p4dp);
1668 				if (is_vmemmap)
1669 					free_vmemmap_storage(p4d_page(p4d), P4D_SIZE, altmap);
1670 			}
1671 			continue;
1672 		}
1673 
1674 		pud_base = pud_offset(p4dp, 0);
1675 		remove_pud_mapping(pud_base, addr, next, is_vmemmap, altmap);
1676 
1677 		if (pgtable_l5_enabled)
1678 			free_pud_table(pud_base, p4dp);
1679 	}
1680 }
1681 
1682 static void __meminit remove_pgd_mapping(unsigned long va, unsigned long end, bool is_vmemmap,
1683 					 struct vmem_altmap *altmap)
1684 {
1685 	unsigned long addr, next;
1686 	p4d_t *p4d_base;
1687 	pgd_t *pgd;
1688 
1689 	for (addr = va; addr < end; addr = next) {
1690 		next = pgd_addr_end(addr, end);
1691 		pgd = pgd_offset_k(addr);
1692 
1693 		if (!pgd_present(*pgd))
1694 			continue;
1695 
1696 		if (pgd_leaf(*pgd))
1697 			continue;
1698 
1699 		p4d_base = p4d_offset(pgd, 0);
1700 		remove_p4d_mapping(p4d_base, addr, next, is_vmemmap, altmap);
1701 	}
1702 
1703 	flush_tlb_all();
1704 }
1705 
1706 static void __meminit remove_linear_mapping(phys_addr_t start, u64 size)
1707 {
1708 	unsigned long va = (unsigned long)__va(start);
1709 	unsigned long end = (unsigned long)__va(start + size);
1710 
1711 	remove_pgd_mapping(va, end, false, NULL);
1712 }
1713 
1714 struct range arch_get_mappable_range(void)
1715 {
1716 	struct range mhp_range;
1717 
1718 	mhp_range.start = __pa(PAGE_OFFSET);
1719 	mhp_range.end = __pa(PAGE_END - 1);
1720 	return mhp_range;
1721 }
1722 
1723 int __ref arch_add_memory(int nid, u64 start, u64 size, struct mhp_params *params)
1724 {
1725 	int ret = 0;
1726 
1727 	create_linear_mapping_range(start, start + size, 0, &params->pgprot);
1728 	ret = __add_pages(nid, start >> PAGE_SHIFT, size >> PAGE_SHIFT, params);
1729 	if (ret) {
1730 		remove_linear_mapping(start, size);
1731 		goto out;
1732 	}
1733 
1734 	max_pfn = PFN_UP(start + size);
1735 	max_low_pfn = max_pfn;
1736 
1737  out:
1738 	flush_tlb_all();
1739 	return ret;
1740 }
1741 
1742 void __ref arch_remove_memory(u64 start, u64 size, struct vmem_altmap *altmap,
1743 			      struct dev_pagemap *pgmap)
1744 {
1745 	__remove_pages(start >> PAGE_SHIFT, size >> PAGE_SHIFT, altmap, pgmap);
1746 	remove_linear_mapping(start, size);
1747 	flush_tlb_all();
1748 }
1749 
1750 void __ref vmemmap_free(unsigned long start, unsigned long end, struct vmem_altmap *altmap)
1751 {
1752 	remove_pgd_mapping(start, end, true, altmap);
1753 }
1754 #endif /* CONFIG_MEMORY_HOTPLUG */
1755