Pull kmalloc into release branch
[linux-2.6] / 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/mm.h>
7 #include <linux/sched.h>
8 #include <linux/highmem.h>
9 #include <linux/module.h>
10 #include <linux/slab.h>
11 #include <asm/uaccess.h>
12 #include <asm/processor.h>
13 #include <asm/tlbflush.h>
14 #include <asm/pgalloc.h>
15 #include <asm/sections.h>
16
17 static DEFINE_SPINLOCK(cpa_lock);
18 static struct list_head df_list = LIST_HEAD_INIT(df_list);
19
20
21 pte_t *lookup_address(unsigned long address) 
22
23         pgd_t *pgd = pgd_offset_k(address);
24         pud_t *pud;
25         pmd_t *pmd;
26         if (pgd_none(*pgd))
27                 return NULL;
28         pud = pud_offset(pgd, address);
29         if (pud_none(*pud))
30                 return NULL;
31         pmd = pmd_offset(pud, address);
32         if (pmd_none(*pmd))
33                 return NULL;
34         if (pmd_large(*pmd))
35                 return (pte_t *)pmd;
36         return pte_offset_kernel(pmd, address);
37
38
39 static struct page *split_large_page(unsigned long address, pgprot_t prot,
40                                         pgprot_t ref_prot)
41
42         int i; 
43         unsigned long addr;
44         struct page *base;
45         pte_t *pbase;
46
47         spin_unlock_irq(&cpa_lock);
48         base = alloc_pages(GFP_KERNEL, 0);
49         spin_lock_irq(&cpa_lock);
50         if (!base) 
51                 return NULL;
52
53         /*
54          * page_private is used to track the number of entries in
55          * the page table page that have non standard attributes.
56          */
57         SetPagePrivate(base);
58         page_private(base) = 0;
59
60         address = __pa(address);
61         addr = address & LARGE_PAGE_MASK; 
62         pbase = (pte_t *)page_address(base);
63         for (i = 0; i < PTRS_PER_PTE; i++, addr += PAGE_SIZE) {
64                set_pte(&pbase[i], pfn_pte(addr >> PAGE_SHIFT,
65                                           addr == address ? prot : ref_prot));
66         }
67         return base;
68
69
70 static void flush_kernel_map(void *dummy) 
71
72         /* Could use CLFLUSH here if the CPU supports it (Hammer,P4) */
73         if (boot_cpu_data.x86_model >= 4) 
74                 wbinvd();
75         /* Flush all to work around Errata in early athlons regarding 
76          * large page flushing. 
77          */
78         __flush_tlb_all();      
79 }
80
81 static void set_pmd_pte(pte_t *kpte, unsigned long address, pte_t pte) 
82
83         struct page *page;
84         unsigned long flags;
85
86         set_pte_atomic(kpte, pte);      /* change init_mm */
87         if (PTRS_PER_PMD > 1)
88                 return;
89
90         spin_lock_irqsave(&pgd_lock, flags);
91         for (page = pgd_list; page; page = (struct page *)page->index) {
92                 pgd_t *pgd;
93                 pud_t *pud;
94                 pmd_t *pmd;
95                 pgd = (pgd_t *)page_address(page) + pgd_index(address);
96                 pud = pud_offset(pgd, address);
97                 pmd = pmd_offset(pud, address);
98                 set_pte_atomic((pte_t *)pmd, pte);
99         }
100         spin_unlock_irqrestore(&pgd_lock, flags);
101 }
102
103 /* 
104  * No more special protections in this 2/4MB area - revert to a
105  * large page again. 
106  */
107 static inline void revert_page(struct page *kpte_page, unsigned long address)
108 {
109         pgprot_t ref_prot;
110         pte_t *linear;
111
112         ref_prot =
113         ((address & LARGE_PAGE_MASK) < (unsigned long)&_etext)
114                 ? PAGE_KERNEL_LARGE_EXEC : PAGE_KERNEL_LARGE;
115
116         linear = (pte_t *)
117                 pmd_offset(pud_offset(pgd_offset_k(address), address), address);
118         set_pmd_pte(linear,  address,
119                     pfn_pte((__pa(address) & LARGE_PAGE_MASK) >> PAGE_SHIFT,
120                             ref_prot));
121 }
122
123 static int
124 __change_page_attr(struct page *page, pgprot_t prot)
125
126         pte_t *kpte; 
127         unsigned long address;
128         struct page *kpte_page;
129
130         BUG_ON(PageHighMem(page));
131         address = (unsigned long)page_address(page);
132
133         kpte = lookup_address(address);
134         if (!kpte)
135                 return -EINVAL;
136         kpte_page = virt_to_page(kpte);
137         if (pgprot_val(prot) != pgprot_val(PAGE_KERNEL)) { 
138                 if ((pte_val(*kpte) & _PAGE_PSE) == 0) { 
139                         set_pte_atomic(kpte, mk_pte(page, prot)); 
140                 } else {
141                         pgprot_t ref_prot;
142                         struct page *split;
143
144                         ref_prot =
145                         ((address & LARGE_PAGE_MASK) < (unsigned long)&_etext)
146                                 ? PAGE_KERNEL_EXEC : PAGE_KERNEL;
147                         split = split_large_page(address, prot, ref_prot);
148                         if (!split)
149                                 return -ENOMEM;
150                         set_pmd_pte(kpte,address,mk_pte(split, ref_prot));
151                         kpte_page = split;
152                 }
153                 page_private(kpte_page)++;
154         } else if ((pte_val(*kpte) & _PAGE_PSE) == 0) { 
155                 set_pte_atomic(kpte, mk_pte(page, PAGE_KERNEL));
156                 BUG_ON(page_private(kpte_page) == 0);
157                 page_private(kpte_page)--;
158         } else
159                 BUG();
160
161         /*
162          * If the pte was reserved, it means it was created at boot
163          * time (not via split_large_page) and in turn we must not
164          * replace it with a largepage.
165          */
166         if (!PageReserved(kpte_page)) {
167                 if (cpu_has_pse && (page_private(kpte_page) == 0)) {
168                         ClearPagePrivate(kpte_page);
169                         list_add(&kpte_page->lru, &df_list);
170                         revert_page(kpte_page, address);
171                 }
172         }
173         return 0;
174
175
176 static inline void flush_map(void)
177 {
178         on_each_cpu(flush_kernel_map, NULL, 1, 1);
179 }
180
181 /*
182  * Change the page attributes of an page in the linear mapping.
183  *
184  * This should be used when a page is mapped with a different caching policy
185  * than write-back somewhere - some CPUs do not like it when mappings with
186  * different caching policies exist. This changes the page attributes of the
187  * in kernel linear mapping too.
188  * 
189  * The caller needs to ensure that there are no conflicting mappings elsewhere.
190  * This function only deals with the kernel linear map.
191  * 
192  * Caller must call global_flush_tlb() after this.
193  */
194 int change_page_attr(struct page *page, int numpages, pgprot_t prot)
195 {
196         int err = 0; 
197         int i; 
198         unsigned long flags;
199
200         spin_lock_irqsave(&cpa_lock, flags);
201         for (i = 0; i < numpages; i++, page++) { 
202                 err = __change_page_attr(page, prot);
203                 if (err) 
204                         break; 
205         }       
206         spin_unlock_irqrestore(&cpa_lock, flags);
207         return err;
208 }
209
210 void global_flush_tlb(void)
211 {
212         struct list_head l;
213         struct page *pg, *next;
214
215         BUG_ON(irqs_disabled());
216
217         spin_lock_irq(&cpa_lock);
218         list_replace_init(&df_list, &l);
219         spin_unlock_irq(&cpa_lock);
220         flush_map();
221         list_for_each_entry_safe(pg, next, &l, lru)
222                 __free_page(pg);
223 }
224
225 #ifdef CONFIG_DEBUG_PAGEALLOC
226 void kernel_map_pages(struct page *page, int numpages, int enable)
227 {
228         if (PageHighMem(page))
229                 return;
230         if (!enable)
231                 debug_check_no_locks_freed(page_address(page),
232                                            numpages * PAGE_SIZE);
233
234         /* the return value is ignored - the calls cannot fail,
235          * large pages are disabled at boot time.
236          */
237         change_page_attr(page, numpages, enable ? PAGE_KERNEL : __pgprot(0));
238         /* we should perform an IPI and flush all tlbs,
239          * but that can deadlock->flush only current cpu.
240          */
241         __flush_tlb_all();
242 }
243 #endif
244
245 EXPORT_SYMBOL(change_page_attr);
246 EXPORT_SYMBOL(global_flush_tlb);