2 * Copyright IBM Corp. 2006
3 * Author(s): Heiko Carstens <heiko.carstens@de.ibm.com>
6 #include <linux/bootmem.h>
9 #include <linux/module.h>
10 #include <linux/list.h>
11 #include <linux/hugetlb.h>
12 #include <linux/slab.h>
13 #include <linux/memblock.h>
14 #include <asm/cacheflush.h>
15 #include <asm/pgalloc.h>
16 #include <asm/pgtable.h>
17 #include <asm/setup.h>
18 #include <asm/tlbflush.h>
19 #include <asm/sections.h>
21 static DEFINE_MUTEX(vmem_mutex);
23 struct memory_segment {
24 struct list_head list;
29 static LIST_HEAD(mem_segs);
31 static void __ref *vmem_alloc_pages(unsigned int order)
33 unsigned long size = PAGE_SIZE << order;
35 if (slab_is_available())
36 return (void *)__get_free_pages(GFP_KERNEL, order);
37 return alloc_bootmem_align(size, size);
40 static inline pud_t *vmem_pud_alloc(void)
44 pud = vmem_alloc_pages(2);
47 clear_table((unsigned long *) pud, _REGION3_ENTRY_EMPTY, PAGE_SIZE * 4);
51 pmd_t *vmem_pmd_alloc(void)
55 pmd = vmem_alloc_pages(2);
58 clear_table((unsigned long *) pmd, _SEGMENT_ENTRY_EMPTY, PAGE_SIZE * 4);
62 pte_t __ref *vmem_pte_alloc(void)
66 if (slab_is_available())
67 pte = (pte_t *) page_table_alloc(&init_mm);
69 pte = alloc_bootmem_align(PTRS_PER_PTE * sizeof(pte_t),
70 PTRS_PER_PTE * sizeof(pte_t));
73 clear_table((unsigned long *) pte, _PAGE_INVALID,
74 PTRS_PER_PTE * sizeof(pte_t));
79 * Add a physical memory range to the 1:1 mapping.
81 static int vmem_add_mem(unsigned long start, unsigned long size)
83 unsigned long pages4k, pages1m, pages2g;
84 unsigned long end = start + size;
85 unsigned long address = start;
92 pages4k = pages1m = pages2g = 0;
93 while (address < end) {
94 pg_dir = pgd_offset_k(address);
95 if (pgd_none(*pg_dir)) {
96 pu_dir = vmem_pud_alloc();
99 pgd_populate(&init_mm, pg_dir, pu_dir);
101 pu_dir = pud_offset(pg_dir, address);
102 if (MACHINE_HAS_EDAT2 && pud_none(*pu_dir) && address &&
103 !(address & ~PUD_MASK) && (address + PUD_SIZE <= end) &&
104 !debug_pagealloc_enabled()) {
105 pud_val(*pu_dir) = address | pgprot_val(REGION3_KERNEL);
110 if (pud_none(*pu_dir)) {
111 pm_dir = vmem_pmd_alloc();
114 pud_populate(&init_mm, pu_dir, pm_dir);
116 pm_dir = pmd_offset(pu_dir, address);
117 if (MACHINE_HAS_EDAT1 && pmd_none(*pm_dir) && address &&
118 !(address & ~PMD_MASK) && (address + PMD_SIZE <= end) &&
119 !debug_pagealloc_enabled()) {
120 pmd_val(*pm_dir) = address | pgprot_val(SEGMENT_KERNEL);
125 if (pmd_none(*pm_dir)) {
126 pt_dir = vmem_pte_alloc();
129 pmd_populate(&init_mm, pm_dir, pt_dir);
132 pt_dir = pte_offset_kernel(pm_dir, address);
133 pte_val(*pt_dir) = address | pgprot_val(PAGE_KERNEL);
134 address += PAGE_SIZE;
139 update_page_count(PG_DIRECT_MAP_4K, pages4k);
140 update_page_count(PG_DIRECT_MAP_1M, pages1m);
141 update_page_count(PG_DIRECT_MAP_2G, pages2g);
146 * Remove a physical memory range from the 1:1 mapping.
147 * Currently only invalidates page table entries.
149 static void vmem_remove_range(unsigned long start, unsigned long size)
151 unsigned long pages4k, pages1m, pages2g;
152 unsigned long end = start + size;
153 unsigned long address = start;
159 pages4k = pages1m = pages2g = 0;
160 while (address < end) {
161 pg_dir = pgd_offset_k(address);
162 if (pgd_none(*pg_dir)) {
163 address += PGDIR_SIZE;
166 pu_dir = pud_offset(pg_dir, address);
167 if (pud_none(*pu_dir)) {
171 if (pud_large(*pu_dir)) {
177 pm_dir = pmd_offset(pu_dir, address);
178 if (pmd_none(*pm_dir)) {
182 if (pmd_large(*pm_dir)) {
188 pt_dir = pte_offset_kernel(pm_dir, address);
189 pte_clear(&init_mm, address, pt_dir);
190 address += PAGE_SIZE;
193 flush_tlb_kernel_range(start, end);
194 update_page_count(PG_DIRECT_MAP_4K, -pages4k);
195 update_page_count(PG_DIRECT_MAP_1M, -pages1m);
196 update_page_count(PG_DIRECT_MAP_2G, -pages2g);
200 * Add a backed mem_map array to the virtual mem_map array.
202 int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node)
204 unsigned long address = start;
211 for (address = start; address < end;) {
212 pg_dir = pgd_offset_k(address);
213 if (pgd_none(*pg_dir)) {
214 pu_dir = vmem_pud_alloc();
217 pgd_populate(&init_mm, pg_dir, pu_dir);
220 pu_dir = pud_offset(pg_dir, address);
221 if (pud_none(*pu_dir)) {
222 pm_dir = vmem_pmd_alloc();
225 pud_populate(&init_mm, pu_dir, pm_dir);
228 pm_dir = pmd_offset(pu_dir, address);
229 if (pmd_none(*pm_dir)) {
230 /* Use 1MB frames for vmemmap if available. We always
231 * use large frames even if they are only partially
233 * Otherwise we would have also page tables since
234 * vmemmap_populate gets called for each section
236 if (MACHINE_HAS_EDAT1) {
239 new_page = vmemmap_alloc_block(PMD_SIZE, node);
242 pmd_val(*pm_dir) = __pa(new_page) |
243 _SEGMENT_ENTRY | _SEGMENT_ENTRY_LARGE;
244 address = (address + PMD_SIZE) & PMD_MASK;
247 pt_dir = vmem_pte_alloc();
250 pmd_populate(&init_mm, pm_dir, pt_dir);
251 } else if (pmd_large(*pm_dir)) {
252 address = (address + PMD_SIZE) & PMD_MASK;
256 pt_dir = pte_offset_kernel(pm_dir, address);
257 if (pte_none(*pt_dir)) {
260 new_page = vmemmap_alloc_block(PAGE_SIZE, node);
264 __pa(new_page) | pgprot_val(PAGE_KERNEL);
266 address += PAGE_SIZE;
273 void vmemmap_free(unsigned long start, unsigned long end)
278 * Add memory segment to the segment list if it doesn't overlap with
279 * an already present segment.
281 static int insert_memory_segment(struct memory_segment *seg)
283 struct memory_segment *tmp;
285 if (seg->start + seg->size > VMEM_MAX_PHYS ||
286 seg->start + seg->size < seg->start)
289 list_for_each_entry(tmp, &mem_segs, list) {
290 if (seg->start >= tmp->start + tmp->size)
292 if (seg->start + seg->size <= tmp->start)
296 list_add(&seg->list, &mem_segs);
301 * Remove memory segment from the segment list.
303 static void remove_memory_segment(struct memory_segment *seg)
305 list_del(&seg->list);
308 static void __remove_shared_memory(struct memory_segment *seg)
310 remove_memory_segment(seg);
311 vmem_remove_range(seg->start, seg->size);
314 int vmem_remove_mapping(unsigned long start, unsigned long size)
316 struct memory_segment *seg;
319 mutex_lock(&vmem_mutex);
322 list_for_each_entry(seg, &mem_segs, list) {
323 if (seg->start == start && seg->size == size)
327 if (seg->start != start || seg->size != size)
331 __remove_shared_memory(seg);
334 mutex_unlock(&vmem_mutex);
338 int vmem_add_mapping(unsigned long start, unsigned long size)
340 struct memory_segment *seg;
343 mutex_lock(&vmem_mutex);
345 seg = kzalloc(sizeof(*seg), GFP_KERNEL);
351 ret = insert_memory_segment(seg);
355 ret = vmem_add_mem(start, size);
361 __remove_shared_memory(seg);
365 mutex_unlock(&vmem_mutex);
370 * map whole physical memory to virtual memory (identity mapping)
371 * we reserve enough space in the vmalloc area for vmemmap to hotplug
372 * additional memory segments.
374 void __init vmem_map_init(void)
376 unsigned long size = _eshared - _stext;
377 struct memblock_region *reg;
379 for_each_memblock(memory, reg)
380 vmem_add_mem(reg->base, reg->size);
381 set_memory_ro((unsigned long)_stext, size >> PAGE_SHIFT);
382 pr_info("Write protected kernel read-only data: %luk\n", size >> 10);
386 * Convert memblock.memory to a memory segment list so there is a single
387 * list that contains all memory segments.
389 static int __init vmem_convert_memory_chunk(void)
391 struct memblock_region *reg;
392 struct memory_segment *seg;
394 mutex_lock(&vmem_mutex);
395 for_each_memblock(memory, reg) {
396 seg = kzalloc(sizeof(*seg), GFP_KERNEL);
398 panic("Out of memory...\n");
399 seg->start = reg->base;
400 seg->size = reg->size;
401 insert_memory_segment(seg);
403 mutex_unlock(&vmem_mutex);
407 core_initcall(vmem_convert_memory_chunk);