1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright IBM Corp. 2006
6 #include <linux/memory_hotplug.h>
7 #include <linux/memblock.h>
10 #include <linux/init.h>
11 #include <linux/list.h>
12 #include <linux/hugetlb.h>
13 #include <linux/slab.h>
14 #include <linux/sort.h>
15 #include <asm/page-states.h>
16 #include <asm/cacheflush.h>
17 #include <asm/nospec-branch.h>
18 #include <asm/ctlreg.h>
19 #include <asm/pgalloc.h>
20 #include <asm/setup.h>
21 #include <asm/tlbflush.h>
22 #include <asm/sections.h>
23 #include <asm/set_memory.h>
25 static DEFINE_MUTEX(vmem_mutex);
27 static void __ref *vmem_alloc_pages(unsigned int order)
29 unsigned long size = PAGE_SIZE << order;
31 if (slab_is_available())
32 return (void *)__get_free_pages(GFP_KERNEL, order);
33 return memblock_alloc(size, size);
36 static void vmem_free_pages(unsigned long addr, int order, struct vmem_altmap *altmap)
39 vmem_altmap_free(altmap, 1 << order);
42 /* We don't expect boot memory to be removed ever. */
43 if (!slab_is_available() ||
44 WARN_ON_ONCE(PageReserved(virt_to_page((void *)addr))))
46 free_pages(addr, order);
49 void *vmem_crst_alloc(unsigned long val)
53 table = vmem_alloc_pages(CRST_ALLOC_ORDER);
56 crst_table_init(table, val);
57 __arch_set_page_dat(table, 1UL << CRST_ALLOC_ORDER);
61 pte_t __ref *vmem_pte_alloc(void)
63 unsigned long size = PTRS_PER_PTE * sizeof(pte_t);
66 if (slab_is_available())
67 pte = (pte_t *) page_table_alloc(&init_mm);
69 pte = (pte_t *) memblock_alloc(size, size);
72 memset64((u64 *)pte, _PAGE_INVALID, PTRS_PER_PTE);
73 __arch_set_page_dat(pte, 1);
77 static void vmem_pte_free(unsigned long *table)
79 /* We don't expect boot memory to be removed ever. */
80 if (!slab_is_available() ||
81 WARN_ON_ONCE(PageReserved(virt_to_page(table))))
83 page_table_free(&init_mm, table);
86 #define PAGE_UNUSED 0xFD
89 * The unused vmemmap range, which was not yet memset(PAGE_UNUSED) ranges
90 * from unused_sub_pmd_start to next PMD_SIZE boundary.
92 static unsigned long unused_sub_pmd_start;
94 static void vmemmap_flush_unused_sub_pmd(void)
96 if (!unused_sub_pmd_start)
98 memset((void *)unused_sub_pmd_start, PAGE_UNUSED,
99 ALIGN(unused_sub_pmd_start, PMD_SIZE) - unused_sub_pmd_start);
100 unused_sub_pmd_start = 0;
103 static void vmemmap_mark_sub_pmd_used(unsigned long start, unsigned long end)
106 * As we expect to add in the same granularity as we remove, it's
107 * sufficient to mark only some piece used to block the memmap page from
108 * getting removed (just in case the memmap never gets initialized,
109 * e.g., because the memory block never gets onlined).
111 memset((void *)start, 0, sizeof(struct page));
114 static void vmemmap_use_sub_pmd(unsigned long start, unsigned long end)
117 * We only optimize if the new used range directly follows the
118 * previously unused range (esp., when populating consecutive sections).
120 if (unused_sub_pmd_start == start) {
121 unused_sub_pmd_start = end;
122 if (likely(IS_ALIGNED(unused_sub_pmd_start, PMD_SIZE)))
123 unused_sub_pmd_start = 0;
126 vmemmap_flush_unused_sub_pmd();
127 vmemmap_mark_sub_pmd_used(start, end);
130 static void vmemmap_use_new_sub_pmd(unsigned long start, unsigned long end)
132 unsigned long page = ALIGN_DOWN(start, PMD_SIZE);
134 vmemmap_flush_unused_sub_pmd();
136 /* Could be our memmap page is filled with PAGE_UNUSED already ... */
137 vmemmap_mark_sub_pmd_used(start, end);
139 /* Mark the unused parts of the new memmap page PAGE_UNUSED. */
140 if (!IS_ALIGNED(start, PMD_SIZE))
141 memset((void *)page, PAGE_UNUSED, start - page);
143 * We want to avoid memset(PAGE_UNUSED) when populating the vmemmap of
144 * consecutive sections. Remember for the last added PMD the last
145 * unused range in the populated PMD.
147 if (!IS_ALIGNED(end, PMD_SIZE))
148 unused_sub_pmd_start = end;
151 /* Returns true if the PMD is completely unused and can be freed. */
152 static bool vmemmap_unuse_sub_pmd(unsigned long start, unsigned long end)
154 unsigned long page = ALIGN_DOWN(start, PMD_SIZE);
156 vmemmap_flush_unused_sub_pmd();
157 memset((void *)start, PAGE_UNUSED, end - start);
158 return !memchr_inv((void *)page, PAGE_UNUSED, PMD_SIZE);
161 /* __ref: we'll only call vmemmap_alloc_block() via vmemmap_populate() */
162 static int __ref modify_pte_table(pmd_t *pmd, unsigned long addr,
163 unsigned long end, bool add, bool direct,
164 struct vmem_altmap *altmap)
166 unsigned long prot, pages = 0;
170 prot = pgprot_val(PAGE_KERNEL);
172 prot &= ~_PAGE_NOEXEC;
174 pte = pte_offset_kernel(pmd, addr);
175 for (; addr < end; addr += PAGE_SIZE, pte++) {
180 vmem_free_pages((unsigned long)pfn_to_virt(pte_pfn(*pte)), get_order(PAGE_SIZE), altmap);
181 pte_clear(&init_mm, addr, pte);
182 } else if (pte_none(*pte)) {
184 void *new_page = vmemmap_alloc_block_buf(PAGE_SIZE, NUMA_NO_NODE, altmap);
188 set_pte(pte, __pte(__pa(new_page) | prot));
190 set_pte(pte, __pte(__pa(addr) | prot));
200 update_page_count(PG_DIRECT_MAP_4K, add ? pages : -pages);
204 static void try_free_pte_table(pmd_t *pmd, unsigned long start)
209 /* We can safely assume this is fully in 1:1 mapping & vmemmap area */
210 pte = pte_offset_kernel(pmd, start);
211 for (i = 0; i < PTRS_PER_PTE; i++, pte++) {
215 vmem_pte_free((unsigned long *) pmd_deref(*pmd));
219 /* __ref: we'll only call vmemmap_alloc_block() via vmemmap_populate() */
220 static int __ref modify_pmd_table(pud_t *pud, unsigned long addr,
221 unsigned long end, bool add, bool direct,
222 struct vmem_altmap *altmap)
224 unsigned long next, prot, pages = 0;
229 prot = pgprot_val(SEGMENT_KERNEL);
231 prot &= ~_SEGMENT_ENTRY_NOEXEC;
233 pmd = pmd_offset(pud, addr);
234 for (; addr < end; addr = next, pmd++) {
235 next = pmd_addr_end(addr, end);
239 if (pmd_leaf(*pmd)) {
240 if (IS_ALIGNED(addr, PMD_SIZE) &&
241 IS_ALIGNED(next, PMD_SIZE)) {
243 vmem_free_pages(pmd_deref(*pmd), get_order(PMD_SIZE), altmap);
246 } else if (!direct && vmemmap_unuse_sub_pmd(addr, next)) {
247 vmem_free_pages(pmd_deref(*pmd), get_order(PMD_SIZE), altmap);
252 } else if (pmd_none(*pmd)) {
253 if (IS_ALIGNED(addr, PMD_SIZE) &&
254 IS_ALIGNED(next, PMD_SIZE) &&
255 MACHINE_HAS_EDAT1 && direct &&
256 !debug_pagealloc_enabled()) {
257 set_pmd(pmd, __pmd(__pa(addr) | prot));
260 } else if (!direct && MACHINE_HAS_EDAT1) {
264 * Use 1MB frames for vmemmap if available. We
265 * always use large frames even if they are only
266 * partially used. Otherwise we would have also
267 * page tables since vmemmap_populate gets
268 * called for each section separately.
270 new_page = vmemmap_alloc_block_buf(PMD_SIZE, NUMA_NO_NODE, altmap);
272 set_pmd(pmd, __pmd(__pa(new_page) | prot));
273 if (!IS_ALIGNED(addr, PMD_SIZE) ||
274 !IS_ALIGNED(next, PMD_SIZE)) {
275 vmemmap_use_new_sub_pmd(addr, next);
280 pte = vmem_pte_alloc();
283 pmd_populate(&init_mm, pmd, pte);
284 } else if (pmd_leaf(*pmd)) {
286 vmemmap_use_sub_pmd(addr, next);
289 ret = modify_pte_table(pmd, addr, next, add, direct, altmap);
293 try_free_pte_table(pmd, addr & PMD_MASK);
298 update_page_count(PG_DIRECT_MAP_1M, add ? pages : -pages);
302 static void try_free_pmd_table(pud_t *pud, unsigned long start)
307 pmd = pmd_offset(pud, start);
308 for (i = 0; i < PTRS_PER_PMD; i++, pmd++)
311 vmem_free_pages(pud_deref(*pud), CRST_ALLOC_ORDER, NULL);
315 static int modify_pud_table(p4d_t *p4d, unsigned long addr, unsigned long end,
316 bool add, bool direct, struct vmem_altmap *altmap)
318 unsigned long next, prot, pages = 0;
323 prot = pgprot_val(REGION3_KERNEL);
325 prot &= ~_REGION_ENTRY_NOEXEC;
326 pud = pud_offset(p4d, addr);
327 for (; addr < end; addr = next, pud++) {
328 next = pud_addr_end(addr, end);
332 if (pud_leaf(*pud)) {
333 if (IS_ALIGNED(addr, PUD_SIZE) &&
334 IS_ALIGNED(next, PUD_SIZE)) {
340 } else if (pud_none(*pud)) {
341 if (IS_ALIGNED(addr, PUD_SIZE) &&
342 IS_ALIGNED(next, PUD_SIZE) &&
343 MACHINE_HAS_EDAT2 && direct &&
344 !debug_pagealloc_enabled()) {
345 set_pud(pud, __pud(__pa(addr) | prot));
349 pmd = vmem_crst_alloc(_SEGMENT_ENTRY_EMPTY);
352 pud_populate(&init_mm, pud, pmd);
353 } else if (pud_leaf(*pud)) {
356 ret = modify_pmd_table(pud, addr, next, add, direct, altmap);
360 try_free_pmd_table(pud, addr & PUD_MASK);
365 update_page_count(PG_DIRECT_MAP_2G, add ? pages : -pages);
369 static void try_free_pud_table(p4d_t *p4d, unsigned long start)
374 pud = pud_offset(p4d, start);
375 for (i = 0; i < PTRS_PER_PUD; i++, pud++) {
379 vmem_free_pages(p4d_deref(*p4d), CRST_ALLOC_ORDER, NULL);
383 static int modify_p4d_table(pgd_t *pgd, unsigned long addr, unsigned long end,
384 bool add, bool direct, struct vmem_altmap *altmap)
391 p4d = p4d_offset(pgd, addr);
392 for (; addr < end; addr = next, p4d++) {
393 next = p4d_addr_end(addr, end);
397 } else if (p4d_none(*p4d)) {
398 pud = vmem_crst_alloc(_REGION3_ENTRY_EMPTY);
401 p4d_populate(&init_mm, p4d, pud);
403 ret = modify_pud_table(p4d, addr, next, add, direct, altmap);
407 try_free_pud_table(p4d, addr & P4D_MASK);
414 static void try_free_p4d_table(pgd_t *pgd, unsigned long start)
419 p4d = p4d_offset(pgd, start);
420 for (i = 0; i < PTRS_PER_P4D; i++, p4d++) {
424 vmem_free_pages(pgd_deref(*pgd), CRST_ALLOC_ORDER, NULL);
428 static int modify_pagetable(unsigned long start, unsigned long end, bool add,
429 bool direct, struct vmem_altmap *altmap)
431 unsigned long addr, next;
436 if (WARN_ON_ONCE(!PAGE_ALIGNED(start | end)))
438 /* Don't mess with any tables not fully in 1:1 mapping & vmemmap area */
439 if (WARN_ON_ONCE(end > VMALLOC_START))
441 for (addr = start; addr < end; addr = next) {
442 next = pgd_addr_end(addr, end);
443 pgd = pgd_offset_k(addr);
448 } else if (pgd_none(*pgd)) {
449 p4d = vmem_crst_alloc(_REGION2_ENTRY_EMPTY);
452 pgd_populate(&init_mm, pgd, p4d);
454 ret = modify_p4d_table(pgd, addr, next, add, direct, altmap);
458 try_free_p4d_table(pgd, addr & PGDIR_MASK);
463 flush_tlb_kernel_range(start, end);
467 static int add_pagetable(unsigned long start, unsigned long end, bool direct,
468 struct vmem_altmap *altmap)
470 return modify_pagetable(start, end, true, direct, altmap);
473 static int remove_pagetable(unsigned long start, unsigned long end, bool direct,
474 struct vmem_altmap *altmap)
476 return modify_pagetable(start, end, false, direct, altmap);
480 * Add a physical memory range to the 1:1 mapping.
482 static int vmem_add_range(unsigned long start, unsigned long size)
484 start = (unsigned long)__va(start);
485 return add_pagetable(start, start + size, true, NULL);
489 * Remove a physical memory range from the 1:1 mapping.
491 static void vmem_remove_range(unsigned long start, unsigned long size)
493 start = (unsigned long)__va(start);
494 remove_pagetable(start, start + size, true, NULL);
498 * Add a backed mem_map array to the virtual mem_map array.
500 int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node,
501 struct vmem_altmap *altmap)
505 mutex_lock(&vmem_mutex);
506 /* We don't care about the node, just use NUMA_NO_NODE on allocations */
507 ret = add_pagetable(start, end, false, altmap);
509 remove_pagetable(start, end, false, altmap);
510 mutex_unlock(&vmem_mutex);
514 #ifdef CONFIG_MEMORY_HOTPLUG
516 void vmemmap_free(unsigned long start, unsigned long end,
517 struct vmem_altmap *altmap)
519 mutex_lock(&vmem_mutex);
520 remove_pagetable(start, end, false, altmap);
521 mutex_unlock(&vmem_mutex);
526 void vmem_remove_mapping(unsigned long start, unsigned long size)
528 mutex_lock(&vmem_mutex);
529 vmem_remove_range(start, size);
530 mutex_unlock(&vmem_mutex);
533 struct range arch_get_mappable_range(void)
535 struct range mhp_range;
538 mhp_range.end = max_mappable - 1;
542 int vmem_add_mapping(unsigned long start, unsigned long size)
544 struct range range = arch_get_mappable_range();
547 if (start < range.start ||
548 start + size > range.end + 1 ||
549 start + size < start)
552 mutex_lock(&vmem_mutex);
553 ret = vmem_add_range(start, size);
555 vmem_remove_range(start, size);
556 mutex_unlock(&vmem_mutex);
561 * Allocate new or return existing page-table entry, but do not map it
562 * to any physical address. If missing, allocate segment- and region-
563 * table entries along. Meeting a large segment- or region-table entry
564 * while traversing is an error, since the function is expected to be
565 * called against virtual regions reserved for 4KB mappings only.
567 pte_t *vmem_get_alloc_pte(unsigned long addr, bool alloc)
576 pgd = pgd_offset_k(addr);
577 if (pgd_none(*pgd)) {
580 p4d = vmem_crst_alloc(_REGION2_ENTRY_EMPTY);
583 pgd_populate(&init_mm, pgd, p4d);
585 p4d = p4d_offset(pgd, addr);
586 if (p4d_none(*p4d)) {
589 pud = vmem_crst_alloc(_REGION3_ENTRY_EMPTY);
592 p4d_populate(&init_mm, p4d, pud);
594 pud = pud_offset(p4d, addr);
595 if (pud_none(*pud)) {
598 pmd = vmem_crst_alloc(_SEGMENT_ENTRY_EMPTY);
601 pud_populate(&init_mm, pud, pmd);
602 } else if (WARN_ON_ONCE(pud_leaf(*pud))) {
605 pmd = pmd_offset(pud, addr);
606 if (pmd_none(*pmd)) {
609 pte = vmem_pte_alloc();
612 pmd_populate(&init_mm, pmd, pte);
613 } else if (WARN_ON_ONCE(pmd_leaf(*pmd))) {
616 ptep = pte_offset_kernel(pmd, addr);
621 int __vmem_map_4k_page(unsigned long addr, unsigned long phys, pgprot_t prot, bool alloc)
625 if (!IS_ALIGNED(addr, PAGE_SIZE))
627 ptep = vmem_get_alloc_pte(addr, alloc);
630 __ptep_ipte(addr, ptep, 0, 0, IPTE_GLOBAL);
631 pte = mk_pte_phys(phys, prot);
636 int vmem_map_4k_page(unsigned long addr, unsigned long phys, pgprot_t prot)
640 mutex_lock(&vmem_mutex);
641 rc = __vmem_map_4k_page(addr, phys, prot, true);
642 mutex_unlock(&vmem_mutex);
646 void vmem_unmap_4k_page(unsigned long addr)
650 mutex_lock(&vmem_mutex);
651 ptep = virt_to_kpte(addr);
652 __ptep_ipte(addr, ptep, 0, 0, IPTE_GLOBAL);
653 pte_clear(&init_mm, addr, ptep);
654 mutex_unlock(&vmem_mutex);
657 void __init vmem_map_init(void)
659 __set_memory_rox(_stext, _etext);
660 __set_memory_ro(_etext, __end_rodata);
661 __set_memory_rox(_sinittext, _einittext);
662 __set_memory_rox(__stext_amode31, __etext_amode31);
664 * If the BEAR-enhancement facility is not installed the first
665 * prefix page is used to return to the previous context with
666 * an LPSWE instruction and therefore must be executable.
668 if (!static_key_enabled(&cpu_has_bear))
670 if (debug_pagealloc_enabled()) {
672 * Use RELOC_HIDE() as long as __va(0) translates to NULL,
673 * since performing pointer arithmetic on a NULL pointer
674 * has undefined behavior and generates compiler warnings.
676 __set_memory_4k(__va(0), RELOC_HIDE(__va(0), ident_map_size));
679 system_ctl_set_bit(0, CR0_INSTRUCTION_EXEC_PROTECTION_BIT);
680 pr_info("Write protected kernel read-only data: %luk\n",
681 (unsigned long)(__end_rodata - _stext) >> 10);