Linux-2.6.12-rc2
[linux-2.6.git] / arch / i386 / mm / pageattr.c
1 /* 
2  * Copyright 2002 Andi Kleen, SuSE Labs. 
3  * Thanks to Ben LaHaise for precious feedback.
4  */ 
5
6 #include <linux/config.h>
7 #include <linux/mm.h>
8 #include <linux/sched.h>
9 #include <linux/highmem.h>
10 #include <linux/module.h>
11 #include <linux/slab.h>
12 #include <asm/uaccess.h>
13 #include <asm/processor.h>
14 #include <asm/tlbflush.h>
15
16 static DEFINE_SPINLOCK(cpa_lock);
17 static struct list_head df_list = LIST_HEAD_INIT(df_list);
18
19
20 pte_t *lookup_address(unsigned long address) 
21
22         pgd_t *pgd = pgd_offset_k(address);
23         pud_t *pud;
24         pmd_t *pmd;
25         if (pgd_none(*pgd))
26                 return NULL;
27         pud = pud_offset(pgd, address);
28         if (pud_none(*pud))
29                 return NULL;
30         pmd = pmd_offset(pud, address);
31         if (pmd_none(*pmd))
32                 return NULL;
33         if (pmd_large(*pmd))
34                 return (pte_t *)pmd;
35         return pte_offset_kernel(pmd, address);
36
37
38 static struct page *split_large_page(unsigned long address, pgprot_t prot)
39
40         int i; 
41         unsigned long addr;
42         struct page *base;
43         pte_t *pbase;
44
45         spin_unlock_irq(&cpa_lock);
46         base = alloc_pages(GFP_KERNEL, 0);
47         spin_lock_irq(&cpa_lock);
48         if (!base) 
49                 return NULL;
50
51         address = __pa(address);
52         addr = address & LARGE_PAGE_MASK; 
53         pbase = (pte_t *)page_address(base);
54         for (i = 0; i < PTRS_PER_PTE; i++, addr += PAGE_SIZE) {
55                 pbase[i] = pfn_pte(addr >> PAGE_SHIFT, 
56                                    addr == address ? prot : PAGE_KERNEL);
57         }
58         return base;
59
60
61 static void flush_kernel_map(void *dummy) 
62
63         /* Could use CLFLUSH here if the CPU supports it (Hammer,P4) */
64         if (boot_cpu_data.x86_model >= 4) 
65                 asm volatile("wbinvd":::"memory"); 
66         /* Flush all to work around Errata in early athlons regarding 
67          * large page flushing. 
68          */
69         __flush_tlb_all();      
70 }
71
72 static void set_pmd_pte(pte_t *kpte, unsigned long address, pte_t pte) 
73
74         struct page *page;
75         unsigned long flags;
76
77         set_pte_atomic(kpte, pte);      /* change init_mm */
78         if (PTRS_PER_PMD > 1)
79                 return;
80
81         spin_lock_irqsave(&pgd_lock, flags);
82         for (page = pgd_list; page; page = (struct page *)page->index) {
83                 pgd_t *pgd;
84                 pud_t *pud;
85                 pmd_t *pmd;
86                 pgd = (pgd_t *)page_address(page) + pgd_index(address);
87                 pud = pud_offset(pgd, address);
88                 pmd = pmd_offset(pud, address);
89                 set_pte_atomic((pte_t *)pmd, pte);
90         }
91         spin_unlock_irqrestore(&pgd_lock, flags);
92 }
93
94 /* 
95  * No more special protections in this 2/4MB area - revert to a
96  * large page again. 
97  */
98 static inline void revert_page(struct page *kpte_page, unsigned long address)
99 {
100         pte_t *linear = (pte_t *) 
101                 pmd_offset(pud_offset(pgd_offset_k(address), address), address);
102         set_pmd_pte(linear,  address,
103                     pfn_pte((__pa(address) & LARGE_PAGE_MASK) >> PAGE_SHIFT,
104                             PAGE_KERNEL_LARGE));
105 }
106
107 static int
108 __change_page_attr(struct page *page, pgprot_t prot)
109
110         pte_t *kpte; 
111         unsigned long address;
112         struct page *kpte_page;
113
114         BUG_ON(PageHighMem(page));
115         address = (unsigned long)page_address(page);
116
117         kpte = lookup_address(address);
118         if (!kpte)
119                 return -EINVAL;
120         kpte_page = virt_to_page(kpte);
121         if (pgprot_val(prot) != pgprot_val(PAGE_KERNEL)) { 
122                 if ((pte_val(*kpte) & _PAGE_PSE) == 0) { 
123                         set_pte_atomic(kpte, mk_pte(page, prot)); 
124                 } else {
125                         struct page *split = split_large_page(address, prot); 
126                         if (!split)
127                                 return -ENOMEM;
128                         set_pmd_pte(kpte,address,mk_pte(split, PAGE_KERNEL));
129                         kpte_page = split;
130                 }       
131                 get_page(kpte_page);
132         } else if ((pte_val(*kpte) & _PAGE_PSE) == 0) { 
133                 set_pte_atomic(kpte, mk_pte(page, PAGE_KERNEL));
134                 __put_page(kpte_page);
135         } else
136                 BUG();
137
138         /*
139          * If the pte was reserved, it means it was created at boot
140          * time (not via split_large_page) and in turn we must not
141          * replace it with a largepage.
142          */
143         if (!PageReserved(kpte_page)) {
144                 /* memleak and potential failed 2M page regeneration */
145                 BUG_ON(!page_count(kpte_page));
146
147                 if (cpu_has_pse && (page_count(kpte_page) == 1)) {
148                         list_add(&kpte_page->lru, &df_list);
149                         revert_page(kpte_page, address);
150                 }
151         }
152         return 0;
153
154
155 static inline void flush_map(void)
156 {
157         on_each_cpu(flush_kernel_map, NULL, 1, 1);
158 }
159
160 /*
161  * Change the page attributes of an page in the linear mapping.
162  *
163  * This should be used when a page is mapped with a different caching policy
164  * than write-back somewhere - some CPUs do not like it when mappings with
165  * different caching policies exist. This changes the page attributes of the
166  * in kernel linear mapping too.
167  * 
168  * The caller needs to ensure that there are no conflicting mappings elsewhere.
169  * This function only deals with the kernel linear map.
170  * 
171  * Caller must call global_flush_tlb() after this.
172  */
173 int change_page_attr(struct page *page, int numpages, pgprot_t prot)
174 {
175         int err = 0; 
176         int i; 
177         unsigned long flags;
178
179         spin_lock_irqsave(&cpa_lock, flags);
180         for (i = 0; i < numpages; i++, page++) { 
181                 err = __change_page_attr(page, prot);
182                 if (err) 
183                         break; 
184         }       
185         spin_unlock_irqrestore(&cpa_lock, flags);
186         return err;
187 }
188
189 void global_flush_tlb(void)
190
191         LIST_HEAD(l);
192         struct page *pg, *next;
193
194         BUG_ON(irqs_disabled());
195
196         spin_lock_irq(&cpa_lock);
197         list_splice_init(&df_list, &l);
198         spin_unlock_irq(&cpa_lock);
199         flush_map();
200         list_for_each_entry_safe(pg, next, &l, lru)
201                 __free_page(pg);
202
203
204 #ifdef CONFIG_DEBUG_PAGEALLOC
205 void kernel_map_pages(struct page *page, int numpages, int enable)
206 {
207         if (PageHighMem(page))
208                 return;
209         /* the return value is ignored - the calls cannot fail,
210          * large pages are disabled at boot time.
211          */
212         change_page_attr(page, numpages, enable ? PAGE_KERNEL : __pgprot(0));
213         /* we should perform an IPI and flush all tlbs,
214          * but that can deadlock->flush only current cpu.
215          */
216         __flush_tlb_all();
217 }
218 #endif
219
220 EXPORT_SYMBOL(change_page_attr);
221 EXPORT_SYMBOL(global_flush_tlb);