Merge branch 'master' of /pub/scm/linux/kernel/git/torvalds/linux-2.6
[linux-2.6] / arch / x86 / mm / hugetlbpage.c
1 /*
2  * IA-32 Huge TLB Page Support for Kernel.
3  *
4  * Copyright (C) 2002, Rohit Seth <rohit.seth@intel.com>
5  */
6
7 #include <linux/init.h>
8 #include <linux/fs.h>
9 #include <linux/mm.h>
10 #include <linux/hugetlb.h>
11 #include <linux/pagemap.h>
12 #include <linux/slab.h>
13 #include <linux/err.h>
14 #include <linux/sysctl.h>
15 #include <asm/mman.h>
16 #include <asm/tlb.h>
17 #include <asm/tlbflush.h>
18
19 static unsigned long page_table_shareable(struct vm_area_struct *svma,
20                                 struct vm_area_struct *vma,
21                                 unsigned long addr, pgoff_t idx)
22 {
23         unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
24                                 svma->vm_start;
25         unsigned long sbase = saddr & PUD_MASK;
26         unsigned long s_end = sbase + PUD_SIZE;
27
28         /*
29          * match the virtual addresses, permission and the alignment of the
30          * page table page.
31          */
32         if (pmd_index(addr) != pmd_index(saddr) ||
33             vma->vm_flags != svma->vm_flags ||
34             sbase < svma->vm_start || svma->vm_end < s_end)
35                 return 0;
36
37         return saddr;
38 }
39
40 static int vma_shareable(struct vm_area_struct *vma, unsigned long addr)
41 {
42         unsigned long base = addr & PUD_MASK;
43         unsigned long end = base + PUD_SIZE;
44
45         /*
46          * check on proper vm_flags and page table alignment
47          */
48         if (vma->vm_flags & VM_MAYSHARE &&
49             vma->vm_start <= base && end <= vma->vm_end)
50                 return 1;
51         return 0;
52 }
53
54 /*
55  * search for a shareable pmd page for hugetlb.
56  */
57 static void huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
58 {
59         struct vm_area_struct *vma = find_vma(mm, addr);
60         struct address_space *mapping = vma->vm_file->f_mapping;
61         pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
62                         vma->vm_pgoff;
63         struct prio_tree_iter iter;
64         struct vm_area_struct *svma;
65         unsigned long saddr;
66         pte_t *spte = NULL;
67
68         if (!vma_shareable(vma, addr))
69                 return;
70
71         spin_lock(&mapping->i_mmap_lock);
72         vma_prio_tree_foreach(svma, &iter, &mapping->i_mmap, idx, idx) {
73                 if (svma == vma)
74                         continue;
75
76                 saddr = page_table_shareable(svma, vma, addr, idx);
77                 if (saddr) {
78                         spte = huge_pte_offset(svma->vm_mm, saddr);
79                         if (spte) {
80                                 get_page(virt_to_page(spte));
81                                 break;
82                         }
83                 }
84         }
85
86         if (!spte)
87                 goto out;
88
89         spin_lock(&mm->page_table_lock);
90         if (pud_none(*pud))
91                 pud_populate(mm, pud, (unsigned long) spte & PAGE_MASK);
92         else
93                 put_page(virt_to_page(spte));
94         spin_unlock(&mm->page_table_lock);
95 out:
96         spin_unlock(&mapping->i_mmap_lock);
97 }
98
99 /*
100  * unmap huge page backed by shared pte.
101  *
102  * Hugetlb pte page is ref counted at the time of mapping.  If pte is shared
103  * indicated by page_count > 1, unmap is achieved by clearing pud and
104  * decrementing the ref count. If count == 1, the pte page is not shared.
105  *
106  * called with vma->vm_mm->page_table_lock held.
107  *
108  * returns: 1 successfully unmapped a shared pte page
109  *          0 the underlying pte page is not shared, or it is the last user
110  */
111 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
112 {
113         pgd_t *pgd = pgd_offset(mm, *addr);
114         pud_t *pud = pud_offset(pgd, *addr);
115
116         BUG_ON(page_count(virt_to_page(ptep)) == 0);
117         if (page_count(virt_to_page(ptep)) == 1)
118                 return 0;
119
120         pud_clear(pud);
121         put_page(virt_to_page(ptep));
122         *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
123         return 1;
124 }
125
126 pte_t *huge_pte_alloc(struct mm_struct *mm, unsigned long addr)
127 {
128         pgd_t *pgd;
129         pud_t *pud;
130         pte_t *pte = NULL;
131
132         pgd = pgd_offset(mm, addr);
133         pud = pud_alloc(mm, pgd, addr);
134         if (pud) {
135                 if (pud_none(*pud))
136                         huge_pmd_share(mm, addr, pud);
137                 pte = (pte_t *) pmd_alloc(mm, pud, addr);
138         }
139         BUG_ON(pte && !pte_none(*pte) && !pte_huge(*pte));
140
141         return pte;
142 }
143
144 pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
145 {
146         pgd_t *pgd;
147         pud_t *pud;
148         pmd_t *pmd = NULL;
149
150         pgd = pgd_offset(mm, addr);
151         if (pgd_present(*pgd)) {
152                 pud = pud_offset(pgd, addr);
153                 if (pud_present(*pud))
154                         pmd = pmd_offset(pud, addr);
155         }
156         return (pte_t *) pmd;
157 }
158
159 #if 0   /* This is just for testing */
160 struct page *
161 follow_huge_addr(struct mm_struct *mm, unsigned long address, int write)
162 {
163         unsigned long start = address;
164         int length = 1;
165         int nr;
166         struct page *page;
167         struct vm_area_struct *vma;
168
169         vma = find_vma(mm, addr);
170         if (!vma || !is_vm_hugetlb_page(vma))
171                 return ERR_PTR(-EINVAL);
172
173         pte = huge_pte_offset(mm, address);
174
175         /* hugetlb should be locked, and hence, prefaulted */
176         WARN_ON(!pte || pte_none(*pte));
177
178         page = &pte_page(*pte)[vpfn % (HPAGE_SIZE/PAGE_SIZE)];
179
180         WARN_ON(!PageCompound(page));
181
182         return page;
183 }
184
185 int pmd_huge(pmd_t pmd)
186 {
187         return 0;
188 }
189
190 struct page *
191 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
192                 pmd_t *pmd, int write)
193 {
194         return NULL;
195 }
196
197 #else
198
199 struct page *
200 follow_huge_addr(struct mm_struct *mm, unsigned long address, int write)
201 {
202         return ERR_PTR(-EINVAL);
203 }
204
205 int pmd_huge(pmd_t pmd)
206 {
207         return !!(pmd_val(pmd) & _PAGE_PSE);
208 }
209
210 struct page *
211 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
212                 pmd_t *pmd, int write)
213 {
214         struct page *page;
215
216         page = pte_page(*(pte_t *)pmd);
217         if (page)
218                 page += ((address & ~HPAGE_MASK) >> PAGE_SHIFT);
219         return page;
220 }
221 #endif
222
223 /* x86_64 also uses this file */
224
225 #ifdef HAVE_ARCH_HUGETLB_UNMAPPED_AREA
226 static unsigned long hugetlb_get_unmapped_area_bottomup(struct file *file,
227                 unsigned long addr, unsigned long len,
228                 unsigned long pgoff, unsigned long flags)
229 {
230         struct mm_struct *mm = current->mm;
231         struct vm_area_struct *vma;
232         unsigned long start_addr;
233
234         if (len > mm->cached_hole_size) {
235                 start_addr = mm->free_area_cache;
236         } else {
237                 start_addr = TASK_UNMAPPED_BASE;
238                 mm->cached_hole_size = 0;
239         }
240
241 full_search:
242         addr = ALIGN(start_addr, HPAGE_SIZE);
243
244         for (vma = find_vma(mm, addr); ; vma = vma->vm_next) {
245                 /* At this point:  (!vma || addr < vma->vm_end). */
246                 if (TASK_SIZE - len < addr) {
247                         /*
248                          * Start a new search - just in case we missed
249                          * some holes.
250                          */
251                         if (start_addr != TASK_UNMAPPED_BASE) {
252                                 start_addr = TASK_UNMAPPED_BASE;
253                                 mm->cached_hole_size = 0;
254                                 goto full_search;
255                         }
256                         return -ENOMEM;
257                 }
258                 if (!vma || addr + len <= vma->vm_start) {
259                         mm->free_area_cache = addr + len;
260                         return addr;
261                 }
262                 if (addr + mm->cached_hole_size < vma->vm_start)
263                         mm->cached_hole_size = vma->vm_start - addr;
264                 addr = ALIGN(vma->vm_end, HPAGE_SIZE);
265         }
266 }
267
268 static unsigned long hugetlb_get_unmapped_area_topdown(struct file *file,
269                 unsigned long addr0, unsigned long len,
270                 unsigned long pgoff, unsigned long flags)
271 {
272         struct mm_struct *mm = current->mm;
273         struct vm_area_struct *vma, *prev_vma;
274         unsigned long base = mm->mmap_base, addr = addr0;
275         unsigned long largest_hole = mm->cached_hole_size;
276         int first_time = 1;
277
278         /* don't allow allocations above current base */
279         if (mm->free_area_cache > base)
280                 mm->free_area_cache = base;
281
282         if (len <= largest_hole) {
283                 largest_hole = 0;
284                 mm->free_area_cache  = base;
285         }
286 try_again:
287         /* make sure it can fit in the remaining address space */
288         if (mm->free_area_cache < len)
289                 goto fail;
290
291         /* either no address requested or cant fit in requested address hole */
292         addr = (mm->free_area_cache - len) & HPAGE_MASK;
293         do {
294                 /*
295                  * Lookup failure means no vma is above this address,
296                  * i.e. return with success:
297                  */
298                 if (!(vma = find_vma_prev(mm, addr, &prev_vma)))
299                         return addr;
300
301                 /*
302                  * new region fits between prev_vma->vm_end and
303                  * vma->vm_start, use it:
304                  */
305                 if (addr + len <= vma->vm_start &&
306                             (!prev_vma || (addr >= prev_vma->vm_end))) {
307                         /* remember the address as a hint for next time */
308                         mm->cached_hole_size = largest_hole;
309                         return (mm->free_area_cache = addr);
310                 } else {
311                         /* pull free_area_cache down to the first hole */
312                         if (mm->free_area_cache == vma->vm_end) {
313                                 mm->free_area_cache = vma->vm_start;
314                                 mm->cached_hole_size = largest_hole;
315                         }
316                 }
317
318                 /* remember the largest hole we saw so far */
319                 if (addr + largest_hole < vma->vm_start)
320                         largest_hole = vma->vm_start - addr;
321
322                 /* try just below the current vma->vm_start */
323                 addr = (vma->vm_start - len) & HPAGE_MASK;
324         } while (len <= vma->vm_start);
325
326 fail:
327         /*
328          * if hint left us with no space for the requested
329          * mapping then try again:
330          */
331         if (first_time) {
332                 mm->free_area_cache = base;
333                 largest_hole = 0;
334                 first_time = 0;
335                 goto try_again;
336         }
337         /*
338          * A failed mmap() very likely causes application failure,
339          * so fall back to the bottom-up function here. This scenario
340          * can happen with large stack limits and large mmap()
341          * allocations.
342          */
343         mm->free_area_cache = TASK_UNMAPPED_BASE;
344         mm->cached_hole_size = ~0UL;
345         addr = hugetlb_get_unmapped_area_bottomup(file, addr0,
346                         len, pgoff, flags);
347
348         /*
349          * Restore the topdown base:
350          */
351         mm->free_area_cache = base;
352         mm->cached_hole_size = ~0UL;
353
354         return addr;
355 }
356
357 unsigned long
358 hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
359                 unsigned long len, unsigned long pgoff, unsigned long flags)
360 {
361         struct mm_struct *mm = current->mm;
362         struct vm_area_struct *vma;
363
364         if (len & ~HPAGE_MASK)
365                 return -EINVAL;
366         if (len > TASK_SIZE)
367                 return -ENOMEM;
368
369         if (flags & MAP_FIXED) {
370                 if (prepare_hugepage_range(addr, len))
371                         return -EINVAL;
372                 return addr;
373         }
374
375         if (addr) {
376                 addr = ALIGN(addr, HPAGE_SIZE);
377                 vma = find_vma(mm, addr);
378                 if (TASK_SIZE - len >= addr &&
379                     (!vma || addr + len <= vma->vm_start))
380                         return addr;
381         }
382         if (mm->get_unmapped_area == arch_get_unmapped_area)
383                 return hugetlb_get_unmapped_area_bottomup(file, addr, len,
384                                 pgoff, flags);
385         else
386                 return hugetlb_get_unmapped_area_topdown(file, addr, len,
387                                 pgoff, flags);
388 }
389
390 #endif /*HAVE_ARCH_HUGETLB_UNMAPPED_AREA*/
391