mm: migration: add migrate_entry_wait_huge()
[linux-3.10.git] / mm / fremap.c
1 /*
2  *   linux/mm/fremap.c
3  * 
4  * Explicit pagetable population and nonlinear (random) mappings support.
5  *
6  * started by Ingo Molnar, Copyright (C) 2002, 2003
7  */
8 #include <linux/export.h>
9 #include <linux/backing-dev.h>
10 #include <linux/mm.h>
11 #include <linux/swap.h>
12 #include <linux/file.h>
13 #include <linux/mman.h>
14 #include <linux/pagemap.h>
15 #include <linux/swapops.h>
16 #include <linux/rmap.h>
17 #include <linux/syscalls.h>
18 #include <linux/mmu_notifier.h>
19
20 #include <asm/mmu_context.h>
21 #include <asm/cacheflush.h>
22 #include <asm/tlbflush.h>
23
24 #include "internal.h"
25
26 static void zap_pte(struct mm_struct *mm, struct vm_area_struct *vma,
27                         unsigned long addr, pte_t *ptep)
28 {
29         pte_t pte = *ptep;
30
31         if (pte_present(pte)) {
32                 struct page *page;
33
34                 flush_cache_page(vma, addr, pte_pfn(pte));
35                 pte = ptep_clear_flush(vma, addr, ptep);
36                 page = vm_normal_page(vma, addr, pte);
37                 if (page) {
38                         if (pte_dirty(pte))
39                                 set_page_dirty(page);
40                         page_remove_rmap(page);
41                         page_cache_release(page);
42                         update_hiwater_rss(mm);
43                         dec_mm_counter(mm, MM_FILEPAGES);
44                 }
45         } else {
46                 if (!pte_file(pte))
47                         free_swap_and_cache(pte_to_swp_entry(pte));
48                 pte_clear_not_present_full(mm, addr, ptep, 0);
49         }
50 }
51
52 /*
53  * Install a file pte to a given virtual memory address, release any
54  * previously existing mapping.
55  */
56 static int install_file_pte(struct mm_struct *mm, struct vm_area_struct *vma,
57                 unsigned long addr, unsigned long pgoff, pgprot_t prot)
58 {
59         int err = -ENOMEM;
60         pte_t *pte;
61         spinlock_t *ptl;
62
63         pte = get_locked_pte(mm, addr, &ptl);
64         if (!pte)
65                 goto out;
66
67         if (!pte_none(*pte))
68                 zap_pte(mm, vma, addr, pte);
69
70         set_pte_at(mm, addr, pte, pgoff_to_pte(pgoff));
71         /*
72          * We don't need to run update_mmu_cache() here because the "file pte"
73          * being installed by install_file_pte() is not a real pte - it's a
74          * non-present entry (like a swap entry), noting what file offset should
75          * be mapped there when there's a fault (in a non-linear vma where
76          * that's not obvious).
77          */
78         pte_unmap_unlock(pte, ptl);
79         err = 0;
80 out:
81         return err;
82 }
83
84 int generic_file_remap_pages(struct vm_area_struct *vma, unsigned long addr,
85                              unsigned long size, pgoff_t pgoff)
86 {
87         struct mm_struct *mm = vma->vm_mm;
88         int err;
89
90         do {
91                 err = install_file_pte(mm, vma, addr, pgoff, vma->vm_page_prot);
92                 if (err)
93                         return err;
94
95                 size -= PAGE_SIZE;
96                 addr += PAGE_SIZE;
97                 pgoff++;
98         } while (size);
99
100         return 0;
101 }
102 EXPORT_SYMBOL(generic_file_remap_pages);
103
104 /**
105  * sys_remap_file_pages - remap arbitrary pages of an existing VM_SHARED vma
106  * @start: start of the remapped virtual memory range
107  * @size: size of the remapped virtual memory range
108  * @prot: new protection bits of the range (see NOTE)
109  * @pgoff: to-be-mapped page of the backing store file
110  * @flags: 0 or MAP_NONBLOCKED - the later will cause no IO.
111  *
112  * sys_remap_file_pages remaps arbitrary pages of an existing VM_SHARED vma
113  * (shared backing store file).
114  *
115  * This syscall works purely via pagetables, so it's the most efficient
116  * way to map the same (large) file into a given virtual window. Unlike
117  * mmap()/mremap() it does not create any new vmas. The new mappings are
118  * also safe across swapout.
119  *
120  * NOTE: the @prot parameter right now is ignored (but must be zero),
121  * and the vma's default protection is used. Arbitrary protections
122  * might be implemented in the future.
123  */
124 SYSCALL_DEFINE5(remap_file_pages, unsigned long, start, unsigned long, size,
125                 unsigned long, prot, unsigned long, pgoff, unsigned long, flags)
126 {
127         struct mm_struct *mm = current->mm;
128         struct address_space *mapping;
129         struct vm_area_struct *vma;
130         int err = -EINVAL;
131         int has_write_lock = 0;
132         vm_flags_t vm_flags = 0;
133
134         if (prot)
135                 return err;
136         /*
137          * Sanitize the syscall parameters:
138          */
139         start = start & PAGE_MASK;
140         size = size & PAGE_MASK;
141
142         /* Does the address range wrap, or is the span zero-sized? */
143         if (start + size <= start)
144                 return err;
145
146         /* Does pgoff wrap? */
147         if (pgoff + (size >> PAGE_SHIFT) < pgoff)
148                 return err;
149
150         /* Can we represent this offset inside this architecture's pte's? */
151 #if PTE_FILE_MAX_BITS < BITS_PER_LONG
152         if (pgoff + (size >> PAGE_SHIFT) >= (1UL << PTE_FILE_MAX_BITS))
153                 return err;
154 #endif
155
156         /* We need down_write() to change vma->vm_flags. */
157         down_read(&mm->mmap_sem);
158  retry:
159         vma = find_vma(mm, start);
160
161         /*
162          * Make sure the vma is shared, that it supports prefaulting,
163          * and that the remapped range is valid and fully within
164          * the single existing vma.
165          */
166         if (!vma || !(vma->vm_flags & VM_SHARED))
167                 goto out;
168
169         if (!vma->vm_ops || !vma->vm_ops->remap_pages)
170                 goto out;
171
172         if (start < vma->vm_start || start + size > vma->vm_end)
173                 goto out;
174
175         /* Must set VM_NONLINEAR before any pages are populated. */
176         if (!(vma->vm_flags & VM_NONLINEAR)) {
177                 /*
178                  * vm_private_data is used as a swapout cursor
179                  * in a VM_NONLINEAR vma.
180                  */
181                 if (vma->vm_private_data)
182                         goto out;
183
184                 /* Don't need a nonlinear mapping, exit success */
185                 if (pgoff == linear_page_index(vma, start)) {
186                         err = 0;
187                         goto out;
188                 }
189
190                 if (!has_write_lock) {
191 get_write_lock:
192                         up_read(&mm->mmap_sem);
193                         down_write(&mm->mmap_sem);
194                         has_write_lock = 1;
195                         goto retry;
196                 }
197                 mapping = vma->vm_file->f_mapping;
198                 /*
199                  * page_mkclean doesn't work on nonlinear vmas, so if
200                  * dirty pages need to be accounted, emulate with linear
201                  * vmas.
202                  */
203                 if (mapping_cap_account_dirty(mapping)) {
204                         unsigned long addr;
205                         struct file *file = get_file(vma->vm_file);
206
207                         addr = mmap_region(file, start, size,
208                                         vma->vm_flags, pgoff);
209                         fput(file);
210                         if (IS_ERR_VALUE(addr)) {
211                                 err = addr;
212                         } else {
213                                 BUG_ON(addr != start);
214                                 err = 0;
215                         }
216                         goto out;
217                 }
218                 mutex_lock(&mapping->i_mmap_mutex);
219                 flush_dcache_mmap_lock(mapping);
220                 vma->vm_flags |= VM_NONLINEAR;
221                 vma_interval_tree_remove(vma, &mapping->i_mmap);
222                 vma_nonlinear_insert(vma, &mapping->i_mmap_nonlinear);
223                 flush_dcache_mmap_unlock(mapping);
224                 mutex_unlock(&mapping->i_mmap_mutex);
225         }
226
227         if (vma->vm_flags & VM_LOCKED) {
228                 /*
229                  * drop PG_Mlocked flag for over-mapped range
230                  */
231                 if (!has_write_lock)
232                         goto get_write_lock;
233                 vm_flags = vma->vm_flags;
234                 munlock_vma_pages_range(vma, start, start + size);
235                 vma->vm_flags = vm_flags;
236         }
237
238         mmu_notifier_invalidate_range_start(mm, start, start + size);
239         err = vma->vm_ops->remap_pages(vma, start, size, pgoff);
240         mmu_notifier_invalidate_range_end(mm, start, start + size);
241
242         /*
243          * We can't clear VM_NONLINEAR because we'd have to do
244          * it after ->populate completes, and that would prevent
245          * downgrading the lock.  (Locks can't be upgraded).
246          */
247
248 out:
249         if (vma)
250                 vm_flags = vma->vm_flags;
251         if (likely(!has_write_lock))
252                 up_read(&mm->mmap_sem);
253         else
254                 up_write(&mm->mmap_sem);
255         if (!err && ((vm_flags & VM_LOCKED) || !(flags & MAP_NONBLOCK)))
256                 mm_populate(start, size);
257
258         return err;
259 }