3 * Copyright (C) 1995-1996 Gary Thomas (gdt@linuxppc.org)
5 * Modifications by Paul Mackerras (PowerMac) (paulus@cs.anu.edu.au)
6 * and Cort Dougan (PReP) (cort@cs.nmt.edu)
7 * Copyright (C) 1996 Paul Mackerras
9 * Derived from "arch/i386/mm/init.c"
10 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
12 * Dave Engebretsen <engebret@us.ibm.com>
13 * Rework for PPC64 port.
15 * This program is free software; you can redistribute it and/or
16 * modify it under the terms of the GNU General Public License
17 * as published by the Free Software Foundation; either version
18 * 2 of the License, or (at your option) any later version.
24 #include <linux/signal.h>
25 #include <linux/sched.h>
26 #include <linux/kernel.h>
27 #include <linux/errno.h>
28 #include <linux/string.h>
29 #include <linux/types.h>
30 #include <linux/mman.h>
32 #include <linux/swap.h>
33 #include <linux/stddef.h>
34 #include <linux/vmalloc.h>
35 #include <linux/init.h>
36 #include <linux/delay.h>
37 #include <linux/highmem.h>
38 #include <linux/idr.h>
39 #include <linux/nodemask.h>
40 #include <linux/module.h>
41 #include <linux/poison.h>
42 #include <linux/memblock.h>
43 #include <linux/hugetlb.h>
44 #include <linux/slab.h>
46 #include <asm/pgalloc.h>
51 #include <asm/mmu_context.h>
52 #include <asm/pgtable.h>
54 #include <asm/uaccess.h>
56 #include <asm/machdep.h>
59 #include <asm/processor.h>
60 #include <asm/mmzone.h>
61 #include <asm/cputable.h>
62 #include <asm/sections.h>
63 #include <asm/iommu.h>
68 #ifdef CONFIG_PPC_STD_MMU_64
69 #if PGTABLE_RANGE > USER_VSID_RANGE
70 #warning Limited user VSID range means pagetable space is wasted
73 #if (TASK_SIZE_USER64 < PGTABLE_RANGE) && (TASK_SIZE_USER64 < USER_VSID_RANGE)
74 #warning TASK_SIZE is smaller than it needs to be.
76 #endif /* CONFIG_PPC_STD_MMU_64 */
78 phys_addr_t memstart_addr = ~0;
79 EXPORT_SYMBOL_GPL(memstart_addr);
80 phys_addr_t kernstart_addr;
81 EXPORT_SYMBOL_GPL(kernstart_addr);
83 static void pgd_ctor(void *addr)
85 memset(addr, 0, PGD_TABLE_SIZE);
88 static void pmd_ctor(void *addr)
90 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
91 memset(addr, 0, PMD_TABLE_SIZE * 2);
93 memset(addr, 0, PMD_TABLE_SIZE);
97 struct kmem_cache *pgtable_cache[MAX_PGTABLE_INDEX_SIZE];
100 * Create a kmem_cache() for pagetables. This is not used for PTE
101 * pages - they're linked to struct page, come from the normal free
102 * pages pool and have a different entry size (see real_pte_t) to
103 * everything else. Caches created by this function are used for all
104 * the higher level pagetables, and for hugepage pagetables.
106 void pgtable_cache_add(unsigned shift, void (*ctor)(void *))
109 unsigned long table_size = sizeof(void *) << shift;
110 unsigned long align = table_size;
112 /* When batching pgtable pointers for RCU freeing, we store
113 * the index size in the low bits. Table alignment must be
114 * big enough to fit it.
116 * Likewise, hugeapge pagetable pointers contain a (different)
117 * shift value in the low bits. All tables must be aligned so
118 * as to leave enough 0 bits in the address to contain it. */
119 unsigned long minalign = max(MAX_PGTABLE_INDEX_SIZE + 1,
120 HUGEPD_SHIFT_MASK + 1);
121 struct kmem_cache *new;
123 /* It would be nice if this was a BUILD_BUG_ON(), but at the
124 * moment, gcc doesn't seem to recognize is_power_of_2 as a
125 * constant expression, so so much for that. */
126 BUG_ON(!is_power_of_2(minalign));
127 BUG_ON((shift < 1) || (shift > MAX_PGTABLE_INDEX_SIZE));
129 if (PGT_CACHE(shift))
130 return; /* Already have a cache of this size */
132 align = max_t(unsigned long, align, minalign);
133 name = kasprintf(GFP_KERNEL, "pgtable-2^%d", shift);
134 new = kmem_cache_create(name, table_size, align, 0, ctor);
136 pgtable_cache[shift - 1] = new;
137 pr_debug("Allocated pgtable cache for order %d\n", shift);
141 void pgtable_cache_init(void)
143 pgtable_cache_add(PGD_INDEX_SIZE, pgd_ctor);
144 pgtable_cache_add(PMD_CACHE_INDEX, pmd_ctor);
145 if (!PGT_CACHE(PGD_INDEX_SIZE) || !PGT_CACHE(PMD_CACHE_INDEX))
146 panic("Couldn't allocate pgtable caches");
147 /* In all current configs, when the PUD index exists it's the
148 * same size as either the pgd or pmd index. Verify that the
149 * initialization above has also created a PUD cache. This
150 * will need re-examiniation if we add new possibilities for
151 * the pagetable layout. */
152 BUG_ON(PUD_INDEX_SIZE && !PGT_CACHE(PUD_INDEX_SIZE));
155 #ifdef CONFIG_SPARSEMEM_VMEMMAP
157 * Given an address within the vmemmap, determine the pfn of the page that
158 * represents the start of the section it is within. Note that we have to
159 * do this by hand as the proffered address may not be correctly aligned.
160 * Subtraction of non-aligned pointers produces undefined results.
162 static unsigned long __meminit vmemmap_section_start(unsigned long page)
164 unsigned long offset = page - ((unsigned long)(vmemmap));
166 /* Return the pfn of the start of the section. */
167 return (offset / sizeof(struct page)) & PAGE_SECTION_MASK;
171 * Check if this vmemmap page is already initialised. If any section
172 * which overlaps this vmemmap page is initialised then this page is
173 * initialised already.
175 static int __meminit vmemmap_populated(unsigned long start, int page_size)
177 unsigned long end = start + page_size;
178 start = (unsigned long)(pfn_to_page(vmemmap_section_start(start)));
180 for (; start < end; start += (PAGES_PER_SECTION * sizeof(struct page)))
181 if (pfn_valid(page_to_pfn((struct page *)start)))
187 /* On hash-based CPUs, the vmemmap is bolted in the hash table.
189 * On Book3E CPUs, the vmemmap is currently mapped in the top half of
190 * the vmalloc space using normal page tables, though the size of
191 * pages encoded in the PTEs can be different
194 #ifdef CONFIG_PPC_BOOK3E
195 static void __meminit vmemmap_create_mapping(unsigned long start,
196 unsigned long page_size,
199 /* Create a PTE encoding without page size */
200 unsigned long i, flags = _PAGE_PRESENT | _PAGE_ACCESSED |
203 /* PTEs only contain page size encodings up to 32M */
204 BUG_ON(mmu_psize_defs[mmu_vmemmap_psize].enc > 0xf);
206 /* Encode the size in the PTE */
207 flags |= mmu_psize_defs[mmu_vmemmap_psize].enc << 8;
209 /* For each PTE for that area, map things. Note that we don't
210 * increment phys because all PTEs are of the large size and
211 * thus must have the low bits clear
213 for (i = 0; i < page_size; i += PAGE_SIZE)
214 BUG_ON(map_kernel_page(start + i, phys, flags));
217 #ifdef CONFIG_MEMORY_HOTPLUG
218 static void vmemmap_remove_mapping(unsigned long start,
219 unsigned long page_size)
223 #else /* CONFIG_PPC_BOOK3E */
224 static void __meminit vmemmap_create_mapping(unsigned long start,
225 unsigned long page_size,
228 int mapped = htab_bolt_mapping(start, start + page_size, phys,
229 pgprot_val(PAGE_KERNEL),
235 #ifdef CONFIG_MEMORY_HOTPLUG
236 static void vmemmap_remove_mapping(unsigned long start,
237 unsigned long page_size)
239 int mapped = htab_remove_mapping(start, start + page_size,
246 #endif /* CONFIG_PPC_BOOK3E */
248 struct vmemmap_backing *vmemmap_list;
249 static struct vmemmap_backing *next;
251 static int num_freed;
253 static __meminit struct vmemmap_backing * vmemmap_list_alloc(int node)
255 struct vmemmap_backing *vmem_back;
256 /* get from freed entries first */
265 /* allocate a page when required and hand out chunks */
267 next = vmemmap_alloc_block(PAGE_SIZE, node);
268 if (unlikely(!next)) {
272 num_left = PAGE_SIZE / sizeof(struct vmemmap_backing);
280 static __meminit void vmemmap_list_populate(unsigned long phys,
284 struct vmemmap_backing *vmem_back;
286 vmem_back = vmemmap_list_alloc(node);
287 if (unlikely(!vmem_back)) {
292 vmem_back->phys = phys;
293 vmem_back->virt_addr = start;
294 vmem_back->list = vmemmap_list;
296 vmemmap_list = vmem_back;
299 int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node)
301 unsigned long page_size = 1 << mmu_psize_defs[mmu_vmemmap_psize].shift;
303 /* Align to the page size of the linear mapping. */
304 start = _ALIGN_DOWN(start, page_size);
306 pr_debug("vmemmap_populate %lx..%lx, node %d\n", start, end, node);
308 for (; start < end; start += page_size) {
311 if (vmemmap_populated(start, page_size))
314 p = vmemmap_alloc_block(page_size, node);
318 vmemmap_list_populate(__pa(p), start, node);
320 pr_debug(" * %016lx..%016lx allocated at %p\n",
321 start, start + page_size, p);
323 vmemmap_create_mapping(start, page_size, __pa(p));
329 #ifdef CONFIG_MEMORY_HOTPLUG
330 static unsigned long vmemmap_list_free(unsigned long start)
332 struct vmemmap_backing *vmem_back, *vmem_back_prev;
334 vmem_back_prev = vmem_back = vmemmap_list;
336 /* look for it with prev pointer recorded */
337 for (; vmem_back; vmem_back = vmem_back->list) {
338 if (vmem_back->virt_addr == start)
340 vmem_back_prev = vmem_back;
343 if (unlikely(!vmem_back)) {
348 /* remove it from vmemmap_list */
349 if (vmem_back == vmemmap_list) /* remove head */
350 vmemmap_list = vmem_back->list;
352 vmem_back_prev->list = vmem_back->list;
354 /* next point to this freed entry */
355 vmem_back->list = next;
359 return vmem_back->phys;
362 void __ref vmemmap_free(unsigned long start, unsigned long end)
364 unsigned long page_size = 1 << mmu_psize_defs[mmu_vmemmap_psize].shift;
366 start = _ALIGN_DOWN(start, page_size);
368 pr_debug("vmemmap_free %lx...%lx\n", start, end);
370 for (; start < end; start += page_size) {
374 * the section has already be marked as invalid, so
375 * vmemmap_populated() true means some other sections still
376 * in this page, so skip it.
378 if (vmemmap_populated(start, page_size))
381 addr = vmemmap_list_free(start);
383 struct page *page = pfn_to_page(addr >> PAGE_SHIFT);
385 if (PageReserved(page)) {
386 /* allocated from bootmem */
387 if (page_size < PAGE_SIZE) {
389 * this shouldn't happen, but if it is
390 * the case, leave the memory there
394 unsigned int nr_pages =
395 1 << get_order(page_size);
397 free_reserved_page(page++);
400 free_pages((unsigned long)(__va(addr)),
401 get_order(page_size));
403 vmemmap_remove_mapping(start, page_size);
408 void register_page_bootmem_memmap(unsigned long section_nr,
409 struct page *start_page, unsigned long size)
414 * We do not have access to the sparsemem vmemmap, so we fallback to
415 * walking the list of sparsemem blocks which we already maintain for
416 * the sake of crashdump. In the long run, we might want to maintain
417 * a tree if performance of that linear walk becomes a problem.
419 * realmode_pfn_to_page functions can fail due to:
420 * 1) As real sparsemem blocks do not lay in RAM continously (they
421 * are in virtual address space which is not available in the real mode),
422 * the requested page struct can be split between blocks so get_page/put_page
424 * 2) When huge pages are used, the get_page/put_page API will fail
425 * in real mode as the linked addresses in the page struct are virtual
428 struct page *realmode_pfn_to_page(unsigned long pfn)
430 struct vmemmap_backing *vmem_back;
432 unsigned long page_size = 1 << mmu_psize_defs[mmu_vmemmap_psize].shift;
433 unsigned long pg_va = (unsigned long) pfn_to_page(pfn);
435 for (vmem_back = vmemmap_list; vmem_back; vmem_back = vmem_back->list) {
436 if (pg_va < vmem_back->virt_addr)
439 /* After vmemmap_list entry free is possible, need check all */
440 if ((pg_va + sizeof(struct page)) <=
441 (vmem_back->virt_addr + page_size)) {
442 page = (struct page *) (vmem_back->phys + pg_va -
443 vmem_back->virt_addr);
448 /* Probably that page struct is split between real pages */
451 EXPORT_SYMBOL_GPL(realmode_pfn_to_page);
453 #elif defined(CONFIG_FLATMEM)
455 struct page *realmode_pfn_to_page(unsigned long pfn)
457 struct page *page = pfn_to_page(pfn);
460 EXPORT_SYMBOL_GPL(realmode_pfn_to_page);
462 #endif /* CONFIG_SPARSEMEM_VMEMMAP/CONFIG_FLATMEM */