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