1/*
2 *  PowerPC version
3 *    Copyright (C) 1995-1996 Gary Thomas (gdt@linuxppc.org)
4 *
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
8 *
9 *  Derived from "arch/i386/mm/init.c"
10 *    Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
11 *
12 *  Dave Engebretsen <engebret@us.ibm.com>
13 *      Rework for PPC64 port.
14 *
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.
19 *
20 */
21
22#undef DEBUG
23
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>
31#include <linux/mm.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>
45
46#include <asm/pgalloc.h>
47#include <asm/page.h>
48#include <asm/prom.h>
49#include <asm/rtas.h>
50#include <asm/io.h>
51#include <asm/mmu_context.h>
52#include <asm/pgtable.h>
53#include <asm/mmu.h>
54#include <asm/uaccess.h>
55#include <asm/smp.h>
56#include <asm/machdep.h>
57#include <asm/tlb.h>
58#include <asm/eeh.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>
64#include <asm/vdso.h>
65
66#include "mmu_decl.h"
67
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
71#endif
72
73#if (TASK_SIZE_USER64 < PGTABLE_RANGE) && (TASK_SIZE_USER64 < USER_VSID_RANGE)
74#warning TASK_SIZE is smaller than it needs to be.
75#endif
76#endif /* CONFIG_PPC_STD_MMU_64 */
77
78phys_addr_t memstart_addr = ~0;
79EXPORT_SYMBOL_GPL(memstart_addr);
80phys_addr_t kernstart_addr;
81EXPORT_SYMBOL_GPL(kernstart_addr);
82
83static void pgd_ctor(void *addr)
84{
85	memset(addr, 0, PGD_TABLE_SIZE);
86}
87
88static void pmd_ctor(void *addr)
89{
90#ifdef CONFIG_TRANSPARENT_HUGEPAGE
91	memset(addr, 0, PMD_TABLE_SIZE * 2);
92#else
93	memset(addr, 0, PMD_TABLE_SIZE);
94#endif
95}
96
97struct kmem_cache *pgtable_cache[MAX_PGTABLE_INDEX_SIZE];
98
99/*
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.
105 */
106void pgtable_cache_add(unsigned shift, void (*ctor)(void *))
107{
108	char *name;
109	unsigned long table_size = sizeof(void *) << shift;
110	unsigned long align = table_size;
111
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.
115	 *
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;
122
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));
128
129	if (PGT_CACHE(shift))
130		return; /* Already have a cache of this size */
131
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);
135	kfree(name);
136	pgtable_cache[shift - 1] = new;
137	pr_debug("Allocated pgtable cache for order %d\n", shift);
138}
139
140
141void pgtable_cache_init(void)
142{
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));
153}
154
155#ifdef CONFIG_SPARSEMEM_VMEMMAP
156/*
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.
161 */
162static unsigned long __meminit vmemmap_section_start(unsigned long page)
163{
164	unsigned long offset = page - ((unsigned long)(vmemmap));
165
166	/* Return the pfn of the start of the section. */
167	return (offset / sizeof(struct page)) & PAGE_SECTION_MASK;
168}
169
170/*
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.
174 */
175static int __meminit vmemmap_populated(unsigned long start, int page_size)
176{
177	unsigned long end = start + page_size;
178	start = (unsigned long)(pfn_to_page(vmemmap_section_start(start)));
179
180	for (; start < end; start += (PAGES_PER_SECTION * sizeof(struct page)))
181		if (pfn_valid(page_to_pfn((struct page *)start)))
182			return 1;
183
184	return 0;
185}
186
187/* On hash-based CPUs, the vmemmap is bolted in the hash table.
188 *
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
192 */
193
194#ifdef CONFIG_PPC_BOOK3E
195static void __meminit vmemmap_create_mapping(unsigned long start,
196					     unsigned long page_size,
197					     unsigned long phys)
198{
199	/* Create a PTE encoding without page size */
200	unsigned long i, flags = _PAGE_PRESENT | _PAGE_ACCESSED |
201		_PAGE_KERNEL_RW;
202
203	/* PTEs only contain page size encodings up to 32M */
204	BUG_ON(mmu_psize_defs[mmu_vmemmap_psize].enc > 0xf);
205
206	/* Encode the size in the PTE */
207	flags |= mmu_psize_defs[mmu_vmemmap_psize].enc << 8;
208
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
212	 */
213	for (i = 0; i < page_size; i += PAGE_SIZE)
214		BUG_ON(map_kernel_page(start + i, phys, flags));
215}
216
217#ifdef CONFIG_MEMORY_HOTPLUG
218static void vmemmap_remove_mapping(unsigned long start,
219				   unsigned long page_size)
220{
221}
222#endif
223#else /* CONFIG_PPC_BOOK3E */
224static void __meminit vmemmap_create_mapping(unsigned long start,
225					     unsigned long page_size,
226					     unsigned long phys)
227{
228	int  mapped = htab_bolt_mapping(start, start + page_size, phys,
229					pgprot_val(PAGE_KERNEL),
230					mmu_vmemmap_psize,
231					mmu_kernel_ssize);
232	BUG_ON(mapped < 0);
233}
234
235#ifdef CONFIG_MEMORY_HOTPLUG
236static void vmemmap_remove_mapping(unsigned long start,
237				   unsigned long page_size)
238{
239	int mapped = htab_remove_mapping(start, start + page_size,
240					 mmu_vmemmap_psize,
241					 mmu_kernel_ssize);
242	BUG_ON(mapped < 0);
243}
244#endif
245
246#endif /* CONFIG_PPC_BOOK3E */
247
248struct vmemmap_backing *vmemmap_list;
249static struct vmemmap_backing *next;
250static int num_left;
251static int num_freed;
252
253static __meminit struct vmemmap_backing * vmemmap_list_alloc(int node)
254{
255	struct vmemmap_backing *vmem_back;
256	/* get from freed entries first */
257	if (num_freed) {
258		num_freed--;
259		vmem_back = next;
260		next = next->list;
261
262		return vmem_back;
263	}
264
265	/* allocate a page when required and hand out chunks */
266	if (!num_left) {
267		next = vmemmap_alloc_block(PAGE_SIZE, node);
268		if (unlikely(!next)) {
269			WARN_ON(1);
270			return NULL;
271		}
272		num_left = PAGE_SIZE / sizeof(struct vmemmap_backing);
273	}
274
275	num_left--;
276
277	return next++;
278}
279
280static __meminit void vmemmap_list_populate(unsigned long phys,
281					    unsigned long start,
282					    int node)
283{
284	struct vmemmap_backing *vmem_back;
285
286	vmem_back = vmemmap_list_alloc(node);
287	if (unlikely(!vmem_back)) {
288		WARN_ON(1);
289		return;
290	}
291
292	vmem_back->phys = phys;
293	vmem_back->virt_addr = start;
294	vmem_back->list = vmemmap_list;
295
296	vmemmap_list = vmem_back;
297}
298
299int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node)
300{
301	unsigned long page_size = 1 << mmu_psize_defs[mmu_vmemmap_psize].shift;
302
303	/* Align to the page size of the linear mapping. */
304	start = _ALIGN_DOWN(start, page_size);
305
306	pr_debug("vmemmap_populate %lx..%lx, node %d\n", start, end, node);
307
308	for (; start < end; start += page_size) {
309		void *p;
310
311		if (vmemmap_populated(start, page_size))
312			continue;
313
314		p = vmemmap_alloc_block(page_size, node);
315		if (!p)
316			return -ENOMEM;
317
318		vmemmap_list_populate(__pa(p), start, node);
319
320		pr_debug("      * %016lx..%016lx allocated at %p\n",
321			 start, start + page_size, p);
322
323		vmemmap_create_mapping(start, page_size, __pa(p));
324	}
325
326	return 0;
327}
328
329#ifdef CONFIG_MEMORY_HOTPLUG
330static unsigned long vmemmap_list_free(unsigned long start)
331{
332	struct vmemmap_backing *vmem_back, *vmem_back_prev;
333
334	vmem_back_prev = vmem_back = vmemmap_list;
335
336	/* look for it with prev pointer recorded */
337	for (; vmem_back; vmem_back = vmem_back->list) {
338		if (vmem_back->virt_addr == start)
339			break;
340		vmem_back_prev = vmem_back;
341	}
342
343	if (unlikely(!vmem_back)) {
344		WARN_ON(1);
345		return 0;
346	}
347
348	/* remove it from vmemmap_list */
349	if (vmem_back == vmemmap_list) /* remove head */
350		vmemmap_list = vmem_back->list;
351	else
352		vmem_back_prev->list = vmem_back->list;
353
354	/* next point to this freed entry */
355	vmem_back->list = next;
356	next = vmem_back;
357	num_freed++;
358
359	return vmem_back->phys;
360}
361
362void __ref vmemmap_free(unsigned long start, unsigned long end)
363{
364	unsigned long page_size = 1 << mmu_psize_defs[mmu_vmemmap_psize].shift;
365
366	start = _ALIGN_DOWN(start, page_size);
367
368	pr_debug("vmemmap_free %lx...%lx\n", start, end);
369
370	for (; start < end; start += page_size) {
371		unsigned long addr;
372
373		/*
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.
377		 */
378		if (vmemmap_populated(start, page_size))
379			continue;
380
381		addr = vmemmap_list_free(start);
382		if (addr) {
383			struct page *page = pfn_to_page(addr >> PAGE_SHIFT);
384
385			if (PageReserved(page)) {
386				/* allocated from bootmem */
387				if (page_size < PAGE_SIZE) {
388					/*
389					 * this shouldn't happen, but if it is
390					 * the case, leave the memory there
391					 */
392					WARN_ON_ONCE(1);
393				} else {
394					unsigned int nr_pages =
395						1 << get_order(page_size);
396					while (nr_pages--)
397						free_reserved_page(page++);
398				}
399			} else
400				free_pages((unsigned long)(__va(addr)),
401							get_order(page_size));
402
403			vmemmap_remove_mapping(start, page_size);
404		}
405	}
406}
407#endif
408void register_page_bootmem_memmap(unsigned long section_nr,
409				  struct page *start_page, unsigned long size)
410{
411}
412
413/*
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.
418 *
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
423 * may fail.
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
426 * too.
427 */
428struct page *realmode_pfn_to_page(unsigned long pfn)
429{
430	struct vmemmap_backing *vmem_back;
431	struct page *page;
432	unsigned long page_size = 1 << mmu_psize_defs[mmu_vmemmap_psize].shift;
433	unsigned long pg_va = (unsigned long) pfn_to_page(pfn);
434
435	for (vmem_back = vmemmap_list; vmem_back; vmem_back = vmem_back->list) {
436		if (pg_va < vmem_back->virt_addr)
437			continue;
438
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);
444			return page;
445		}
446	}
447
448	/* Probably that page struct is split between real pages */
449	return NULL;
450}
451EXPORT_SYMBOL_GPL(realmode_pfn_to_page);
452
453#elif defined(CONFIG_FLATMEM)
454
455struct page *realmode_pfn_to_page(unsigned long pfn)
456{
457	struct page *page = pfn_to_page(pfn);
458	return page;
459}
460EXPORT_SYMBOL_GPL(realmode_pfn_to_page);
461
462#endif /* CONFIG_SPARSEMEM_VMEMMAP/CONFIG_FLATMEM */
463