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 <asm/pgalloc.h>
14 #include <asm/pgtable.h>
15 #include <asm/setup.h>
16 #include <asm/tlbflush.h>
17 #include <asm/sections.h>
19 static DEFINE_MUTEX(vmem_mutex);
21 struct memory_segment {
22 struct list_head list;
27 static LIST_HEAD(mem_segs);
29 static void __ref *vmem_alloc_pages(unsigned int order)
31 if (slab_is_available())
32 return (void *)__get_free_pages(GFP_KERNEL, order);
33 return alloc_bootmem_pages((1 << order) * PAGE_SIZE);
36 static inline pud_t *vmem_pud_alloc(void)
41 pud = vmem_alloc_pages(2);
44 clear_table((unsigned long *) pud, _REGION3_ENTRY_EMPTY, PAGE_SIZE * 4);
49 static inline pmd_t *vmem_pmd_alloc(void)
54 pmd = vmem_alloc_pages(2);
57 clear_table((unsigned long *) pmd, _SEGMENT_ENTRY_EMPTY, PAGE_SIZE * 4);
62 static pte_t __ref *vmem_pte_alloc(unsigned long address)
66 if (slab_is_available())
67 pte = (pte_t *) page_table_alloc(&init_mm, address);
69 pte = alloc_bootmem(PTRS_PER_PTE * sizeof(pte_t));
72 clear_table((unsigned long *) pte, _PAGE_INVALID,
73 PTRS_PER_PTE * sizeof(pte_t));
78 * Add a physical memory range to the 1:1 mapping.
80 static int vmem_add_mem(unsigned long start, unsigned long size, int ro)
82 unsigned long end = start + size;
83 unsigned long address = start;
90 while (address < end) {
91 pg_dir = pgd_offset_k(address);
92 if (pgd_none(*pg_dir)) {
93 pu_dir = vmem_pud_alloc();
96 pgd_populate(&init_mm, pg_dir, pu_dir);
98 pu_dir = pud_offset(pg_dir, address);
99 #if defined(CONFIG_64BIT) && !defined(CONFIG_DEBUG_PAGEALLOC)
100 if (MACHINE_HAS_EDAT2 && pud_none(*pu_dir) && address &&
101 !(address & ~PUD_MASK) && (address + PUD_SIZE <= end)) {
102 pud_val(*pu_dir) = __pa(address) |
103 _REGION_ENTRY_TYPE_R3 | _REGION3_ENTRY_LARGE |
104 (ro ? _REGION_ENTRY_PROTECT : 0);
109 if (pud_none(*pu_dir)) {
110 pm_dir = vmem_pmd_alloc();
113 pud_populate(&init_mm, pu_dir, pm_dir);
115 pm_dir = pmd_offset(pu_dir, address);
116 #if defined(CONFIG_64BIT) && !defined(CONFIG_DEBUG_PAGEALLOC)
117 if (MACHINE_HAS_EDAT1 && pmd_none(*pm_dir) && address &&
118 !(address & ~PMD_MASK) && (address + PMD_SIZE <= end)) {
119 pmd_val(*pm_dir) = __pa(address) |
120 _SEGMENT_ENTRY | _SEGMENT_ENTRY_LARGE |
121 _SEGMENT_ENTRY_YOUNG |
122 (ro ? _SEGMENT_ENTRY_PROTECT : 0);
127 if (pmd_none(*pm_dir)) {
128 pt_dir = vmem_pte_alloc(address);
131 pmd_populate(&init_mm, pm_dir, pt_dir);
134 pt_dir = pte_offset_kernel(pm_dir, address);
135 pte_val(*pt_dir) = __pa(address) |
136 pgprot_val(ro ? PAGE_KERNEL_RO : PAGE_KERNEL);
137 address += PAGE_SIZE;
141 flush_tlb_kernel_range(start, end);
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 end = start + size;
152 unsigned long address = start;
159 pte_val(pte) = _PAGE_INVALID;
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)) {
176 pm_dir = pmd_offset(pu_dir, address);
177 if (pmd_none(*pm_dir)) {
181 if (pmd_large(*pm_dir)) {
186 pt_dir = pte_offset_kernel(pm_dir, address);
188 address += PAGE_SIZE;
190 flush_tlb_kernel_range(start, end);
194 * Add a backed mem_map array to the virtual mem_map array.
196 int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node)
198 unsigned long address = start;
205 for (address = start; address < end;) {
206 pg_dir = pgd_offset_k(address);
207 if (pgd_none(*pg_dir)) {
208 pu_dir = vmem_pud_alloc();
211 pgd_populate(&init_mm, pg_dir, pu_dir);
214 pu_dir = pud_offset(pg_dir, address);
215 if (pud_none(*pu_dir)) {
216 pm_dir = vmem_pmd_alloc();
219 pud_populate(&init_mm, pu_dir, pm_dir);
222 pm_dir = pmd_offset(pu_dir, address);
223 if (pmd_none(*pm_dir)) {
225 /* Use 1MB frames for vmemmap if available. We always
226 * use large frames even if they are only partially
228 * Otherwise we would have also page tables since
229 * vmemmap_populate gets called for each section
231 if (MACHINE_HAS_EDAT1) {
234 new_page = vmemmap_alloc_block(PMD_SIZE, node);
237 pmd_val(*pm_dir) = __pa(new_page) |
238 _SEGMENT_ENTRY | _SEGMENT_ENTRY_LARGE |
240 address = (address + PMD_SIZE) & PMD_MASK;
244 pt_dir = vmem_pte_alloc(address);
247 pmd_populate(&init_mm, pm_dir, pt_dir);
248 } else if (pmd_large(*pm_dir)) {
249 address = (address + PMD_SIZE) & PMD_MASK;
253 pt_dir = pte_offset_kernel(pm_dir, address);
254 if (pte_none(*pt_dir)) {
255 unsigned long new_page;
257 new_page =__pa(vmem_alloc_pages(0));
261 __pa(new_page) | pgprot_val(PAGE_KERNEL);
263 address += PAGE_SIZE;
265 memset((void *)start, 0, end - start);
268 flush_tlb_kernel_range(start, end);
272 void vmemmap_free(unsigned long start, unsigned long end)
277 * Add memory segment to the segment list if it doesn't overlap with
278 * an already present segment.
280 static int insert_memory_segment(struct memory_segment *seg)
282 struct memory_segment *tmp;
284 if (seg->start + seg->size > VMEM_MAX_PHYS ||
285 seg->start + seg->size < seg->start)
288 list_for_each_entry(tmp, &mem_segs, list) {
289 if (seg->start >= tmp->start + tmp->size)
291 if (seg->start + seg->size <= tmp->start)
295 list_add(&seg->list, &mem_segs);
300 * Remove memory segment from the segment list.
302 static void remove_memory_segment(struct memory_segment *seg)
304 list_del(&seg->list);
307 static void __remove_shared_memory(struct memory_segment *seg)
309 remove_memory_segment(seg);
310 vmem_remove_range(seg->start, seg->size);
313 int vmem_remove_mapping(unsigned long start, unsigned long size)
315 struct memory_segment *seg;
318 mutex_lock(&vmem_mutex);
321 list_for_each_entry(seg, &mem_segs, list) {
322 if (seg->start == start && seg->size == size)
326 if (seg->start != start || seg->size != size)
330 __remove_shared_memory(seg);
333 mutex_unlock(&vmem_mutex);
337 int vmem_add_mapping(unsigned long start, unsigned long size)
339 struct memory_segment *seg;
342 mutex_lock(&vmem_mutex);
344 seg = kzalloc(sizeof(*seg), GFP_KERNEL);
350 ret = insert_memory_segment(seg);
354 ret = vmem_add_mem(start, size, 0);
360 __remove_shared_memory(seg);
364 mutex_unlock(&vmem_mutex);
369 * map whole physical memory to virtual memory (identity mapping)
370 * we reserve enough space in the vmalloc area for vmemmap to hotplug
371 * additional memory segments.
373 void __init vmem_map_init(void)
375 unsigned long ro_start, ro_end;
376 unsigned long start, end;
379 ro_start = PFN_ALIGN((unsigned long)&_stext);
380 ro_end = (unsigned long)&_eshared & PAGE_MASK;
381 for (i = 0; i < MEMORY_CHUNKS; i++) {
382 if (!memory_chunk[i].size)
384 start = memory_chunk[i].addr;
385 end = memory_chunk[i].addr + memory_chunk[i].size;
386 if (start >= ro_end || end <= ro_start)
387 vmem_add_mem(start, end - start, 0);
388 else if (start >= ro_start && end <= ro_end)
389 vmem_add_mem(start, end - start, 1);
390 else if (start >= ro_start) {
391 vmem_add_mem(start, ro_end - start, 1);
392 vmem_add_mem(ro_end, end - ro_end, 0);
393 } else if (end < ro_end) {
394 vmem_add_mem(start, ro_start - start, 0);
395 vmem_add_mem(ro_start, end - ro_start, 1);
397 vmem_add_mem(start, ro_start - start, 0);
398 vmem_add_mem(ro_start, ro_end - ro_start, 1);
399 vmem_add_mem(ro_end, end - ro_end, 0);
405 * Convert memory chunk array to a memory segment list so there is a single
406 * list that contains both r/w memory and shared memory segments.
408 static int __init vmem_convert_memory_chunk(void)
410 struct memory_segment *seg;
413 mutex_lock(&vmem_mutex);
414 for (i = 0; i < MEMORY_CHUNKS; i++) {
415 if (!memory_chunk[i].size)
417 seg = kzalloc(sizeof(*seg), GFP_KERNEL);
419 panic("Out of memory...\n");
420 seg->start = memory_chunk[i].addr;
421 seg->size = memory_chunk[i].size;
422 insert_memory_segment(seg);
424 mutex_unlock(&vmem_mutex);
428 core_initcall(vmem_convert_memory_chunk);