4 * Explicit pagetable population and nonlinear (random) mappings support.
6 * started by Ingo Molnar, Copyright (C) 2002, 2003
10 #include <linux/swap.h>
11 #include <linux/file.h>
12 #include <linux/mman.h>
13 #include <linux/pagemap.h>
14 #include <linux/swapops.h>
15 #include <linux/rmap.h>
16 #include <linux/module.h>
17 #include <linux/vs_memory.h>
19 #include <asm/mmu_context.h>
20 #include <asm/cacheflush.h>
21 #include <asm/tlbflush.h>
23 static inline void zap_pte(struct mm_struct *mm, struct vm_area_struct *vma,
24 unsigned long addr, pte_t *ptep)
30 if (pte_present(pte)) {
31 unsigned long pfn = pte_pfn(pte);
33 flush_cache_page(vma, addr);
34 pte = ptep_clear_flush(vma, addr, ptep);
36 struct page *page = pfn_to_page(pfn);
37 if (!PageReserved(page)) {
40 page_remove_rmap(page);
41 page_cache_release(page);
48 free_swap_and_cache(pte_to_swp_entry(pte));
54 * Install a file page to a given virtual memory address, release any
55 * previously existing mapping.
57 int install_page(struct mm_struct *mm, struct vm_area_struct *vma,
58 unsigned long addr, struct page *page, pgprot_t prot)
68 pgd = pgd_offset(mm, addr);
69 spin_lock(&mm->page_table_lock);
71 if (!vx_rsspages_avail(mm, 1))
74 pmd = pmd_alloc(mm, pgd, addr);
78 pte = pte_alloc_map(mm, pmd, addr);
83 * This page may have been truncated. Tell the
87 inode = vma->vm_file->f_mapping->host;
88 size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
89 if (!page->mapping || page->index >= size)
92 zap_pte(mm, vma, addr, pte);
96 flush_icache_page(vma, page);
97 set_pte(pte, mk_pte(page, prot));
98 page_add_file_rmap(page);
101 update_mmu_cache(vma, addr, pte_val);
105 spin_unlock(&mm->page_table_lock);
108 EXPORT_SYMBOL(install_page);
112 * Install a file pte to a given virtual memory address, release any
113 * previously existing mapping.
115 int install_file_pte(struct mm_struct *mm, struct vm_area_struct *vma,
116 unsigned long addr, unsigned long pgoff, pgprot_t prot)
124 pgd = pgd_offset(mm, addr);
125 spin_lock(&mm->page_table_lock);
127 pmd = pmd_alloc(mm, pgd, addr);
131 pte = pte_alloc_map(mm, pmd, addr);
135 zap_pte(mm, vma, addr, pte);
137 set_pte(pte, pgoff_to_pte(pgoff));
140 update_mmu_cache(vma, addr, pte_val);
141 spin_unlock(&mm->page_table_lock);
145 spin_unlock(&mm->page_table_lock);
151 * sys_remap_file_pages - remap arbitrary pages of a shared backing store
152 * file within an existing vma.
153 * @start: start of the remapped virtual memory range
154 * @size: size of the remapped virtual memory range
155 * @prot: new protection bits of the range
156 * @pgoff: to be mapped page of the backing store file
157 * @flags: 0 or MAP_NONBLOCKED - the later will cause no IO.
159 * this syscall works purely via pagetables, so it's the most efficient
160 * way to map the same (large) file into a given virtual window. Unlike
161 * mmap()/mremap() it does not create any new vmas. The new mappings are
162 * also safe across swapout.
164 * NOTE: the 'prot' parameter right now is ignored, and the vma's default
165 * protection is used. Arbitrary protections might be implemented in the
168 asmlinkage long sys_remap_file_pages(unsigned long start, unsigned long size,
169 unsigned long __prot, unsigned long pgoff, unsigned long flags)
171 struct mm_struct *mm = current->mm;
172 struct address_space *mapping;
173 unsigned long end = start + size;
174 struct vm_area_struct *vma;
176 int has_write_lock = 0;
181 * Sanitize the syscall parameters:
183 start = start & PAGE_MASK;
184 size = size & PAGE_MASK;
186 /* Does the address range wrap, or is the span zero-sized? */
187 if (start + size <= start)
190 /* Can we represent this offset inside this architecture's pte's? */
191 #if PTE_FILE_MAX_BITS < BITS_PER_LONG
192 if (pgoff + (size >> PAGE_SHIFT) >= (1UL << PTE_FILE_MAX_BITS))
196 /* We need down_write() to change vma->vm_flags. */
197 down_read(&mm->mmap_sem);
199 vma = find_vma(mm, start);
202 * Make sure the vma is shared, that it supports prefaulting,
203 * and that the remapped range is valid and fully within
204 * the single existing vma. vm_private_data is used as a
205 * swapout cursor in a VM_NONLINEAR vma (unless VM_RESERVED
206 * or VM_LOCKED, but VM_LOCKED could be revoked later on).
208 if (vma && (vma->vm_flags & VM_SHARED) &&
209 (!vma->vm_private_data ||
210 (vma->vm_flags & (VM_NONLINEAR|VM_RESERVED))) &&
211 vma->vm_ops && vma->vm_ops->populate &&
212 end > start && start >= vma->vm_start &&
213 end <= vma->vm_end) {
215 /* Must set VM_NONLINEAR before any pages are populated. */
216 if (pgoff != linear_page_index(vma, start) &&
217 !(vma->vm_flags & VM_NONLINEAR)) {
218 if (!has_write_lock) {
219 up_read(&mm->mmap_sem);
220 down_write(&mm->mmap_sem);
224 mapping = vma->vm_file->f_mapping;
225 spin_lock(&mapping->i_mmap_lock);
226 flush_dcache_mmap_lock(mapping);
227 vma->vm_flags |= VM_NONLINEAR;
228 vma_prio_tree_remove(vma, &mapping->i_mmap);
229 vma_prio_tree_init(vma);
230 list_add_tail(&vma->shared.vm_set.list,
231 &mapping->i_mmap_nonlinear);
232 flush_dcache_mmap_unlock(mapping);
233 spin_unlock(&mapping->i_mmap_lock);
236 err = vma->vm_ops->populate(vma, start, size,
238 pgoff, flags & MAP_NONBLOCK);
241 * We can't clear VM_NONLINEAR because we'd have to do
242 * it after ->populate completes, and that would prevent
243 * downgrading the lock. (Locks can't be upgraded).
246 if (likely(!has_write_lock))
247 up_read(&mm->mmap_sem);
249 up_write(&mm->mmap_sem);