Merge branch 'master' of master.kernel.org:/pub/scm/linux/kernel/git/davem/net-2.6
[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         paravirt_alloc_pt(page_to_pfn(base));
64         for (i = 0; i < PTRS_PER_PTE; i++, addr += PAGE_SIZE) {
65                set_pte(&pbase[i], pfn_pte(addr >> PAGE_SHIFT,
66                                           addr == address ? prot : ref_prot));
67         }
68         return base;
69
70
71 static void cache_flush_page(struct page *p)
72
73         unsigned long adr = (unsigned long)page_address(p);
74         int i;
75         for (i = 0; i < PAGE_SIZE; i += boot_cpu_data.x86_clflush_size)
76                 asm volatile("clflush (%0)" :: "r" (adr + i));
77 }
78
79 static void flush_kernel_map(void *arg)
80 {
81         struct list_head *lh = (struct list_head *)arg;
82         struct page *p;
83
84         /* High level code is not ready for clflush yet */
85         if (0 && cpu_has_clflush) {
86                 list_for_each_entry (p, lh, lru)
87                         cache_flush_page(p);
88         } else if (boot_cpu_data.x86_model >= 4)
89                 wbinvd();
90
91         /* Flush all to work around Errata in early athlons regarding 
92          * large page flushing. 
93          */
94         __flush_tlb_all();      
95 }
96
97 static void set_pmd_pte(pte_t *kpte, unsigned long address, pte_t pte) 
98
99         struct page *page;
100         unsigned long flags;
101
102         set_pte_atomic(kpte, pte);      /* change init_mm */
103         if (SHARED_KERNEL_PMD)
104                 return;
105
106         spin_lock_irqsave(&pgd_lock, flags);
107         for (page = pgd_list; page; page = (struct page *)page->index) {
108                 pgd_t *pgd;
109                 pud_t *pud;
110                 pmd_t *pmd;
111                 pgd = (pgd_t *)page_address(page) + pgd_index(address);
112                 pud = pud_offset(pgd, address);
113                 pmd = pmd_offset(pud, address);
114                 set_pte_atomic((pte_t *)pmd, pte);
115         }
116         spin_unlock_irqrestore(&pgd_lock, flags);
117 }
118
119 /* 
120  * No more special protections in this 2/4MB area - revert to a
121  * large page again. 
122  */
123 static inline void revert_page(struct page *kpte_page, unsigned long address)
124 {
125         pgprot_t ref_prot;
126         pte_t *linear;
127
128         ref_prot =
129         ((address & LARGE_PAGE_MASK) < (unsigned long)&_etext)
130                 ? PAGE_KERNEL_LARGE_EXEC : PAGE_KERNEL_LARGE;
131
132         linear = (pte_t *)
133                 pmd_offset(pud_offset(pgd_offset_k(address), address), address);
134         set_pmd_pte(linear,  address,
135                     pfn_pte((__pa(address) & LARGE_PAGE_MASK) >> PAGE_SHIFT,
136                             ref_prot));
137 }
138
139 static int
140 __change_page_attr(struct page *page, pgprot_t prot)
141
142         pte_t *kpte; 
143         unsigned long address;
144         struct page *kpte_page;
145
146         BUG_ON(PageHighMem(page));
147         address = (unsigned long)page_address(page);
148
149         kpte = lookup_address(address);
150         if (!kpte)
151                 return -EINVAL;
152         kpte_page = virt_to_page(kpte);
153         if (pgprot_val(prot) != pgprot_val(PAGE_KERNEL)) { 
154                 if (!pte_huge(*kpte)) {
155                         set_pte_atomic(kpte, mk_pte(page, prot)); 
156                 } else {
157                         pgprot_t ref_prot;
158                         struct page *split;
159
160                         ref_prot =
161                         ((address & LARGE_PAGE_MASK) < (unsigned long)&_etext)
162                                 ? PAGE_KERNEL_EXEC : PAGE_KERNEL;
163                         split = split_large_page(address, prot, ref_prot);
164                         if (!split)
165                                 return -ENOMEM;
166                         set_pmd_pte(kpte,address,mk_pte(split, ref_prot));
167                         kpte_page = split;
168                 }
169                 page_private(kpte_page)++;
170         } else if (!pte_huge(*kpte)) {
171                 set_pte_atomic(kpte, mk_pte(page, PAGE_KERNEL));
172                 BUG_ON(page_private(kpte_page) == 0);
173                 page_private(kpte_page)--;
174         } else
175                 BUG();
176
177         /*
178          * If the pte was reserved, it means it was created at boot
179          * time (not via split_large_page) and in turn we must not
180          * replace it with a largepage.
181          */
182         if (!PageReserved(kpte_page)) {
183                 if (cpu_has_pse && (page_private(kpte_page) == 0)) {
184                         ClearPagePrivate(kpte_page);
185                         paravirt_release_pt(page_to_pfn(kpte_page));
186                         list_add(&kpte_page->lru, &df_list);
187                         revert_page(kpte_page, address);
188                 }
189         }
190         return 0;
191
192
193 static inline void flush_map(struct list_head *l)
194 {
195         on_each_cpu(flush_kernel_map, l, 1, 1);
196 }
197
198 /*
199  * Change the page attributes of an page in the linear mapping.
200  *
201  * This should be used when a page is mapped with a different caching policy
202  * than write-back somewhere - some CPUs do not like it when mappings with
203  * different caching policies exist. This changes the page attributes of the
204  * in kernel linear mapping too.
205  * 
206  * The caller needs to ensure that there are no conflicting mappings elsewhere.
207  * This function only deals with the kernel linear map.
208  * 
209  * Caller must call global_flush_tlb() after this.
210  */
211 int change_page_attr(struct page *page, int numpages, pgprot_t prot)
212 {
213         int err = 0; 
214         int i; 
215         unsigned long flags;
216
217         spin_lock_irqsave(&cpa_lock, flags);
218         for (i = 0; i < numpages; i++, page++) { 
219                 err = __change_page_attr(page, prot);
220                 if (err) 
221                         break; 
222         }       
223         spin_unlock_irqrestore(&cpa_lock, flags);
224         return err;
225 }
226
227 void global_flush_tlb(void)
228 {
229         struct list_head l;
230         struct page *pg, *next;
231
232         BUG_ON(irqs_disabled());
233
234         spin_lock_irq(&cpa_lock);
235         list_replace_init(&df_list, &l);
236         spin_unlock_irq(&cpa_lock);
237         flush_map(&l);
238         list_for_each_entry_safe(pg, next, &l, lru) {
239                 __free_page(pg);
240         }
241 }
242
243 #ifdef CONFIG_DEBUG_PAGEALLOC
244 void kernel_map_pages(struct page *page, int numpages, int enable)
245 {
246         if (PageHighMem(page))
247                 return;
248         if (!enable)
249                 debug_check_no_locks_freed(page_address(page),
250                                            numpages * PAGE_SIZE);
251
252         /* the return value is ignored - the calls cannot fail,
253          * large pages are disabled at boot time.
254          */
255         change_page_attr(page, numpages, enable ? PAGE_KERNEL : __pgprot(0));
256         /* we should perform an IPI and flush all tlbs,
257          * but that can deadlock->flush only current cpu.
258          */
259         __flush_tlb_all();
260 }
261 #endif
262
263 EXPORT_SYMBOL(change_page_attr);
264 EXPORT_SYMBOL(global_flush_tlb);