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