1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright IBM Corp. 2006
4 * Author(s): Heiko Carstens <heiko.carstens@de.ibm.com>
7 #include <linux/bootmem.h>
10 #include <linux/init.h>
11 #include <linux/list.h>
12 #include <linux/hugetlb.h>
13 #include <linux/slab.h>
14 #include <linux/memblock.h>
15 #include <asm/cacheflush.h>
16 #include <asm/pgalloc.h>
17 #include <asm/pgtable.h>
18 #include <asm/setup.h>
19 #include <asm/tlbflush.h>
20 #include <asm/sections.h>
21 #include <asm/set_memory.h>
23 static DEFINE_MUTEX(vmem_mutex);
25 struct memory_segment {
26 struct list_head list;
31 static LIST_HEAD(mem_segs);
33 static void __ref *vmem_alloc_pages(unsigned int order)
35 unsigned long size = PAGE_SIZE << order;
37 if (slab_is_available())
38 return (void *)__get_free_pages(GFP_KERNEL, order);
39 return (void *) memblock_alloc(size, size);
42 void *vmem_crst_alloc(unsigned long val)
46 table = vmem_alloc_pages(CRST_ALLOC_ORDER);
48 crst_table_init(table, val);
52 pte_t __ref *vmem_pte_alloc(void)
54 unsigned long size = PTRS_PER_PTE * sizeof(pte_t);
57 if (slab_is_available())
58 pte = (pte_t *) page_table_alloc(&init_mm);
60 pte = (pte_t *) memblock_alloc(size, size);
63 memset64((u64 *)pte, _PAGE_INVALID, PTRS_PER_PTE);
68 * Add a physical memory range to the 1:1 mapping.
70 static int vmem_add_mem(unsigned long start, unsigned long size)
72 unsigned long pgt_prot, sgt_prot, r3_prot;
73 unsigned long pages4k, pages1m, pages2g;
74 unsigned long end = start + size;
75 unsigned long address = start;
83 pgt_prot = pgprot_val(PAGE_KERNEL);
84 sgt_prot = pgprot_val(SEGMENT_KERNEL);
85 r3_prot = pgprot_val(REGION3_KERNEL);
86 if (!MACHINE_HAS_NX) {
87 pgt_prot &= ~_PAGE_NOEXEC;
88 sgt_prot &= ~_SEGMENT_ENTRY_NOEXEC;
89 r3_prot &= ~_REGION_ENTRY_NOEXEC;
91 pages4k = pages1m = pages2g = 0;
92 while (address < end) {
93 pg_dir = pgd_offset_k(address);
94 if (pgd_none(*pg_dir)) {
95 p4_dir = vmem_crst_alloc(_REGION2_ENTRY_EMPTY);
98 pgd_populate(&init_mm, pg_dir, p4_dir);
100 p4_dir = p4d_offset(pg_dir, address);
101 if (p4d_none(*p4_dir)) {
102 pu_dir = vmem_crst_alloc(_REGION3_ENTRY_EMPTY);
105 p4d_populate(&init_mm, p4_dir, pu_dir);
107 pu_dir = pud_offset(p4_dir, address);
108 if (MACHINE_HAS_EDAT2 && pud_none(*pu_dir) && address &&
109 !(address & ~PUD_MASK) && (address + PUD_SIZE <= end) &&
110 !debug_pagealloc_enabled()) {
111 pud_val(*pu_dir) = address | r3_prot;
116 if (pud_none(*pu_dir)) {
117 pm_dir = vmem_crst_alloc(_SEGMENT_ENTRY_EMPTY);
120 pud_populate(&init_mm, pu_dir, pm_dir);
122 pm_dir = pmd_offset(pu_dir, address);
123 if (MACHINE_HAS_EDAT1 && pmd_none(*pm_dir) && address &&
124 !(address & ~PMD_MASK) && (address + PMD_SIZE <= end) &&
125 !debug_pagealloc_enabled()) {
126 pmd_val(*pm_dir) = address | sgt_prot;
131 if (pmd_none(*pm_dir)) {
132 pt_dir = vmem_pte_alloc();
135 pmd_populate(&init_mm, pm_dir, pt_dir);
138 pt_dir = pte_offset_kernel(pm_dir, address);
139 pte_val(*pt_dir) = address | pgt_prot;
140 address += PAGE_SIZE;
145 update_page_count(PG_DIRECT_MAP_4K, pages4k);
146 update_page_count(PG_DIRECT_MAP_1M, pages1m);
147 update_page_count(PG_DIRECT_MAP_2G, pages2g);
152 * Remove a physical memory range from the 1:1 mapping.
153 * Currently only invalidates page table entries.
155 static void vmem_remove_range(unsigned long start, unsigned long size)
157 unsigned long pages4k, pages1m, pages2g;
158 unsigned long end = start + size;
159 unsigned long address = start;
166 pages4k = pages1m = pages2g = 0;
167 while (address < end) {
168 pg_dir = pgd_offset_k(address);
169 if (pgd_none(*pg_dir)) {
170 address += PGDIR_SIZE;
173 p4_dir = p4d_offset(pg_dir, address);
174 if (p4d_none(*p4_dir)) {
178 pu_dir = pud_offset(p4_dir, address);
179 if (pud_none(*pu_dir)) {
183 if (pud_large(*pu_dir)) {
189 pm_dir = pmd_offset(pu_dir, address);
190 if (pmd_none(*pm_dir)) {
194 if (pmd_large(*pm_dir)) {
200 pt_dir = pte_offset_kernel(pm_dir, address);
201 pte_clear(&init_mm, address, pt_dir);
202 address += PAGE_SIZE;
205 flush_tlb_kernel_range(start, end);
206 update_page_count(PG_DIRECT_MAP_4K, -pages4k);
207 update_page_count(PG_DIRECT_MAP_1M, -pages1m);
208 update_page_count(PG_DIRECT_MAP_2G, -pages2g);
212 * Add a backed mem_map array to the virtual mem_map array.
214 int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node,
215 struct vmem_altmap *altmap)
217 unsigned long pgt_prot, sgt_prot;
218 unsigned long address = start;
226 pgt_prot = pgprot_val(PAGE_KERNEL);
227 sgt_prot = pgprot_val(SEGMENT_KERNEL);
228 if (!MACHINE_HAS_NX) {
229 pgt_prot &= ~_PAGE_NOEXEC;
230 sgt_prot &= ~_SEGMENT_ENTRY_NOEXEC;
232 for (address = start; address < end;) {
233 pg_dir = pgd_offset_k(address);
234 if (pgd_none(*pg_dir)) {
235 p4_dir = vmem_crst_alloc(_REGION2_ENTRY_EMPTY);
238 pgd_populate(&init_mm, pg_dir, p4_dir);
241 p4_dir = p4d_offset(pg_dir, address);
242 if (p4d_none(*p4_dir)) {
243 pu_dir = vmem_crst_alloc(_REGION3_ENTRY_EMPTY);
246 p4d_populate(&init_mm, p4_dir, pu_dir);
249 pu_dir = pud_offset(p4_dir, address);
250 if (pud_none(*pu_dir)) {
251 pm_dir = vmem_crst_alloc(_SEGMENT_ENTRY_EMPTY);
254 pud_populate(&init_mm, pu_dir, pm_dir);
257 pm_dir = pmd_offset(pu_dir, address);
258 if (pmd_none(*pm_dir)) {
259 /* Use 1MB frames for vmemmap if available. We always
260 * use large frames even if they are only partially
262 * Otherwise we would have also page tables since
263 * vmemmap_populate gets called for each section
265 if (MACHINE_HAS_EDAT1) {
268 new_page = vmemmap_alloc_block(PMD_SIZE, node);
271 pmd_val(*pm_dir) = __pa(new_page) | sgt_prot;
272 address = (address + PMD_SIZE) & PMD_MASK;
275 pt_dir = vmem_pte_alloc();
278 pmd_populate(&init_mm, pm_dir, pt_dir);
279 } else if (pmd_large(*pm_dir)) {
280 address = (address + PMD_SIZE) & PMD_MASK;
284 pt_dir = pte_offset_kernel(pm_dir, address);
285 if (pte_none(*pt_dir)) {
288 new_page = vmemmap_alloc_block(PAGE_SIZE, node);
291 pte_val(*pt_dir) = __pa(new_page) | pgt_prot;
293 address += PAGE_SIZE;
300 void vmemmap_free(unsigned long start, unsigned long end,
301 struct vmem_altmap *altmap)
306 * Add memory segment to the segment list if it doesn't overlap with
307 * an already present segment.
309 static int insert_memory_segment(struct memory_segment *seg)
311 struct memory_segment *tmp;
313 if (seg->start + seg->size > VMEM_MAX_PHYS ||
314 seg->start + seg->size < seg->start)
317 list_for_each_entry(tmp, &mem_segs, list) {
318 if (seg->start >= tmp->start + tmp->size)
320 if (seg->start + seg->size <= tmp->start)
324 list_add(&seg->list, &mem_segs);
329 * Remove memory segment from the segment list.
331 static void remove_memory_segment(struct memory_segment *seg)
333 list_del(&seg->list);
336 static void __remove_shared_memory(struct memory_segment *seg)
338 remove_memory_segment(seg);
339 vmem_remove_range(seg->start, seg->size);
342 int vmem_remove_mapping(unsigned long start, unsigned long size)
344 struct memory_segment *seg;
347 mutex_lock(&vmem_mutex);
350 list_for_each_entry(seg, &mem_segs, list) {
351 if (seg->start == start && seg->size == size)
355 if (seg->start != start || seg->size != size)
359 __remove_shared_memory(seg);
362 mutex_unlock(&vmem_mutex);
366 int vmem_add_mapping(unsigned long start, unsigned long size)
368 struct memory_segment *seg;
371 mutex_lock(&vmem_mutex);
373 seg = kzalloc(sizeof(*seg), GFP_KERNEL);
379 ret = insert_memory_segment(seg);
383 ret = vmem_add_mem(start, size);
389 __remove_shared_memory(seg);
393 mutex_unlock(&vmem_mutex);
398 * map whole physical memory to virtual memory (identity mapping)
399 * we reserve enough space in the vmalloc area for vmemmap to hotplug
400 * additional memory segments.
402 void __init vmem_map_init(void)
404 struct memblock_region *reg;
406 for_each_memblock(memory, reg)
407 vmem_add_mem(reg->base, reg->size);
408 __set_memory((unsigned long)_stext,
409 (unsigned long)(_etext - _stext) >> PAGE_SHIFT,
410 SET_MEMORY_RO | SET_MEMORY_X);
411 __set_memory((unsigned long)_etext,
412 (unsigned long)(__end_rodata - _etext) >> PAGE_SHIFT,
414 __set_memory((unsigned long)_sinittext,
415 (unsigned long)(_einittext - _sinittext) >> PAGE_SHIFT,
416 SET_MEMORY_RO | SET_MEMORY_X);
417 pr_info("Write protected kernel read-only data: %luk\n",
418 (unsigned long)(__end_rodata - _stext) >> 10);
422 * Convert memblock.memory to a memory segment list so there is a single
423 * list that contains all memory segments.
425 static int __init vmem_convert_memory_chunk(void)
427 struct memblock_region *reg;
428 struct memory_segment *seg;
430 mutex_lock(&vmem_mutex);
431 for_each_memblock(memory, reg) {
432 seg = kzalloc(sizeof(*seg), GFP_KERNEL);
434 panic("Out of memory...\n");
435 seg->start = reg->base;
436 seg->size = reg->size;
437 insert_memory_segment(seg);
439 mutex_unlock(&vmem_mutex);
443 core_initcall(vmem_convert_memory_chunk);