Linux-2.6.12-rc2
[linux-2.6.git] / arch / i386 / 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/config.h>
8 #include <linux/init.h>
9 #include <linux/fs.h>
10 #include <linux/mm.h>
11 #include <linux/hugetlb.h>
12 #include <linux/pagemap.h>
13 #include <linux/smp_lock.h>
14 #include <linux/slab.h>
15 #include <linux/err.h>
16 #include <linux/sysctl.h>
17 #include <asm/mman.h>
18 #include <asm/tlb.h>
19 #include <asm/tlbflush.h>
20
21 static pte_t *huge_pte_alloc(struct mm_struct *mm, unsigned long addr)
22 {
23         pgd_t *pgd;
24         pud_t *pud;
25         pmd_t *pmd = NULL;
26
27         pgd = pgd_offset(mm, addr);
28         pud = pud_alloc(mm, pgd, addr);
29         pmd = pmd_alloc(mm, pud, addr);
30         return (pte_t *) pmd;
31 }
32
33 static pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
34 {
35         pgd_t *pgd;
36         pud_t *pud;
37         pmd_t *pmd = NULL;
38
39         pgd = pgd_offset(mm, addr);
40         pud = pud_offset(pgd, addr);
41         pmd = pmd_offset(pud, addr);
42         return (pte_t *) pmd;
43 }
44
45 static void set_huge_pte(struct mm_struct *mm, struct vm_area_struct *vma, struct page *page, pte_t * page_table, int write_access)
46 {
47         pte_t entry;
48
49         add_mm_counter(mm, rss, HPAGE_SIZE / PAGE_SIZE);
50         if (write_access) {
51                 entry =
52                     pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
53         } else
54                 entry = pte_wrprotect(mk_pte(page, vma->vm_page_prot));
55         entry = pte_mkyoung(entry);
56         mk_pte_huge(entry);
57         set_pte(page_table, entry);
58 }
59
60 /*
61  * This function checks for proper alignment of input addr and len parameters.
62  */
63 int is_aligned_hugepage_range(unsigned long addr, unsigned long len)
64 {
65         if (len & ~HPAGE_MASK)
66                 return -EINVAL;
67         if (addr & ~HPAGE_MASK)
68                 return -EINVAL;
69         return 0;
70 }
71
72 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
73                         struct vm_area_struct *vma)
74 {
75         pte_t *src_pte, *dst_pte, entry;
76         struct page *ptepage;
77         unsigned long addr = vma->vm_start;
78         unsigned long end = vma->vm_end;
79
80         while (addr < end) {
81                 dst_pte = huge_pte_alloc(dst, addr);
82                 if (!dst_pte)
83                         goto nomem;
84                 src_pte = huge_pte_offset(src, addr);
85                 entry = *src_pte;
86                 ptepage = pte_page(entry);
87                 get_page(ptepage);
88                 set_pte(dst_pte, entry);
89                 add_mm_counter(dst, rss, HPAGE_SIZE / PAGE_SIZE);
90                 addr += HPAGE_SIZE;
91         }
92         return 0;
93
94 nomem:
95         return -ENOMEM;
96 }
97
98 int
99 follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
100                     struct page **pages, struct vm_area_struct **vmas,
101                     unsigned long *position, int *length, int i)
102 {
103         unsigned long vpfn, vaddr = *position;
104         int remainder = *length;
105
106         WARN_ON(!is_vm_hugetlb_page(vma));
107
108         vpfn = vaddr/PAGE_SIZE;
109         while (vaddr < vma->vm_end && remainder) {
110
111                 if (pages) {
112                         pte_t *pte;
113                         struct page *page;
114
115                         pte = huge_pte_offset(mm, vaddr);
116
117                         /* hugetlb should be locked, and hence, prefaulted */
118                         WARN_ON(!pte || pte_none(*pte));
119
120                         page = &pte_page(*pte)[vpfn % (HPAGE_SIZE/PAGE_SIZE)];
121
122                         WARN_ON(!PageCompound(page));
123
124                         get_page(page);
125                         pages[i] = page;
126                 }
127
128                 if (vmas)
129                         vmas[i] = vma;
130
131                 vaddr += PAGE_SIZE;
132                 ++vpfn;
133                 --remainder;
134                 ++i;
135         }
136
137         *length = remainder;
138         *position = vaddr;
139
140         return i;
141 }
142
143 #if 0   /* This is just for testing */
144 struct page *
145 follow_huge_addr(struct mm_struct *mm, unsigned long address, int write)
146 {
147         unsigned long start = address;
148         int length = 1;
149         int nr;
150         struct page *page;
151         struct vm_area_struct *vma;
152
153         vma = find_vma(mm, addr);
154         if (!vma || !is_vm_hugetlb_page(vma))
155                 return ERR_PTR(-EINVAL);
156
157         pte = huge_pte_offset(mm, address);
158
159         /* hugetlb should be locked, and hence, prefaulted */
160         WARN_ON(!pte || pte_none(*pte));
161
162         page = &pte_page(*pte)[vpfn % (HPAGE_SIZE/PAGE_SIZE)];
163
164         WARN_ON(!PageCompound(page));
165
166         return page;
167 }
168
169 int pmd_huge(pmd_t pmd)
170 {
171         return 0;
172 }
173
174 struct page *
175 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
176                 pmd_t *pmd, int write)
177 {
178         return NULL;
179 }
180
181 #else
182
183 struct page *
184 follow_huge_addr(struct mm_struct *mm, unsigned long address, int write)
185 {
186         return ERR_PTR(-EINVAL);
187 }
188
189 int pmd_huge(pmd_t pmd)
190 {
191         return !!(pmd_val(pmd) & _PAGE_PSE);
192 }
193
194 struct page *
195 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
196                 pmd_t *pmd, int write)
197 {
198         struct page *page;
199
200         page = pte_page(*(pte_t *)pmd);
201         if (page)
202                 page += ((address & ~HPAGE_MASK) >> PAGE_SHIFT);
203         return page;
204 }
205 #endif
206
207 void unmap_hugepage_range(struct vm_area_struct *vma,
208                 unsigned long start, unsigned long end)
209 {
210         struct mm_struct *mm = vma->vm_mm;
211         unsigned long address;
212         pte_t pte, *ptep;
213         struct page *page;
214
215         BUG_ON(start & (HPAGE_SIZE - 1));
216         BUG_ON(end & (HPAGE_SIZE - 1));
217
218         for (address = start; address < end; address += HPAGE_SIZE) {
219                 ptep = huge_pte_offset(mm, address);
220                 if (!ptep)
221                         continue;
222                 pte = ptep_get_and_clear(mm, address, ptep);
223                 if (pte_none(pte))
224                         continue;
225                 page = pte_page(pte);
226                 put_page(page);
227         }
228         add_mm_counter(mm ,rss, -((end - start) >> PAGE_SHIFT));
229         flush_tlb_range(vma, start, end);
230 }
231
232 int hugetlb_prefault(struct address_space *mapping, struct vm_area_struct *vma)
233 {
234         struct mm_struct *mm = current->mm;
235         unsigned long addr;
236         int ret = 0;
237
238         BUG_ON(vma->vm_start & ~HPAGE_MASK);
239         BUG_ON(vma->vm_end & ~HPAGE_MASK);
240
241         spin_lock(&mm->page_table_lock);
242         for (addr = vma->vm_start; addr < vma->vm_end; addr += HPAGE_SIZE) {
243                 unsigned long idx;
244                 pte_t *pte = huge_pte_alloc(mm, addr);
245                 struct page *page;
246
247                 if (!pte) {
248                         ret = -ENOMEM;
249                         goto out;
250                 }
251
252                 if (!pte_none(*pte)) {
253                         pmd_t *pmd = (pmd_t *) pte;
254
255                         page = pmd_page(*pmd);
256                         pmd_clear(pmd);
257                         mm->nr_ptes--;
258                         dec_page_state(nr_page_table_pages);
259                         page_cache_release(page);
260                 }
261
262                 idx = ((addr - vma->vm_start) >> HPAGE_SHIFT)
263                         + (vma->vm_pgoff >> (HPAGE_SHIFT - PAGE_SHIFT));
264                 page = find_get_page(mapping, idx);
265                 if (!page) {
266                         /* charge the fs quota first */
267                         if (hugetlb_get_quota(mapping)) {
268                                 ret = -ENOMEM;
269                                 goto out;
270                         }
271                         page = alloc_huge_page();
272                         if (!page) {
273                                 hugetlb_put_quota(mapping);
274                                 ret = -ENOMEM;
275                                 goto out;
276                         }
277                         ret = add_to_page_cache(page, mapping, idx, GFP_ATOMIC);
278                         if (! ret) {
279                                 unlock_page(page);
280                         } else {
281                                 hugetlb_put_quota(mapping);
282                                 free_huge_page(page);
283                                 goto out;
284                         }
285                 }
286                 set_huge_pte(mm, vma, page, pte, vma->vm_flags & VM_WRITE);
287         }
288 out:
289         spin_unlock(&mm->page_table_lock);
290         return ret;
291 }
292
293 /* x86_64 also uses this file */
294
295 #ifdef HAVE_ARCH_HUGETLB_UNMAPPED_AREA
296 static unsigned long hugetlb_get_unmapped_area_bottomup(struct file *file,
297                 unsigned long addr, unsigned long len,
298                 unsigned long pgoff, unsigned long flags)
299 {
300         struct mm_struct *mm = current->mm;
301         struct vm_area_struct *vma;
302         unsigned long start_addr;
303
304         start_addr = mm->free_area_cache;
305
306 full_search:
307         addr = ALIGN(start_addr, HPAGE_SIZE);
308
309         for (vma = find_vma(mm, addr); ; vma = vma->vm_next) {
310                 /* At this point:  (!vma || addr < vma->vm_end). */
311                 if (TASK_SIZE - len < addr) {
312                         /*
313                          * Start a new search - just in case we missed
314                          * some holes.
315                          */
316                         if (start_addr != TASK_UNMAPPED_BASE) {
317                                 start_addr = TASK_UNMAPPED_BASE;
318                                 goto full_search;
319                         }
320                         return -ENOMEM;
321                 }
322                 if (!vma || addr + len <= vma->vm_start) {
323                         mm->free_area_cache = addr + len;
324                         return addr;
325                 }
326                 addr = ALIGN(vma->vm_end, HPAGE_SIZE);
327         }
328 }
329
330 static unsigned long hugetlb_get_unmapped_area_topdown(struct file *file,
331                 unsigned long addr0, unsigned long len,
332                 unsigned long pgoff, unsigned long flags)
333 {
334         struct mm_struct *mm = current->mm;
335         struct vm_area_struct *vma, *prev_vma;
336         unsigned long base = mm->mmap_base, addr = addr0;
337         int first_time = 1;
338
339         /* don't allow allocations above current base */
340         if (mm->free_area_cache > base)
341                 mm->free_area_cache = base;
342
343 try_again:
344         /* make sure it can fit in the remaining address space */
345         if (mm->free_area_cache < len)
346                 goto fail;
347
348         /* either no address requested or cant fit in requested address hole */
349         addr = (mm->free_area_cache - len) & HPAGE_MASK;
350         do {
351                 /*
352                  * Lookup failure means no vma is above this address,
353                  * i.e. return with success:
354                  */
355                 if (!(vma = find_vma_prev(mm, addr, &prev_vma)))
356                         return addr;
357
358                 /*
359                  * new region fits between prev_vma->vm_end and
360                  * vma->vm_start, use it:
361                  */
362                 if (addr + len <= vma->vm_start &&
363                                 (!prev_vma || (addr >= prev_vma->vm_end)))
364                         /* remember the address as a hint for next time */
365                         return (mm->free_area_cache = addr);
366                 else
367                         /* pull free_area_cache down to the first hole */
368                         if (mm->free_area_cache == vma->vm_end)
369                                 mm->free_area_cache = vma->vm_start;
370
371                 /* try just below the current vma->vm_start */
372                 addr = (vma->vm_start - len) & HPAGE_MASK;
373         } while (len <= vma->vm_start);
374
375 fail:
376         /*
377          * if hint left us with no space for the requested
378          * mapping then try again:
379          */
380         if (first_time) {
381                 mm->free_area_cache = base;
382                 first_time = 0;
383                 goto try_again;
384         }
385         /*
386          * A failed mmap() very likely causes application failure,
387          * so fall back to the bottom-up function here. This scenario
388          * can happen with large stack limits and large mmap()
389          * allocations.
390          */
391         mm->free_area_cache = TASK_UNMAPPED_BASE;
392         addr = hugetlb_get_unmapped_area_bottomup(file, addr0,
393                         len, pgoff, flags);
394
395         /*
396          * Restore the topdown base:
397          */
398         mm->free_area_cache = base;
399
400         return addr;
401 }
402
403 unsigned long
404 hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
405                 unsigned long len, unsigned long pgoff, unsigned long flags)
406 {
407         struct mm_struct *mm = current->mm;
408         struct vm_area_struct *vma;
409
410         if (len & ~HPAGE_MASK)
411                 return -EINVAL;
412         if (len > TASK_SIZE)
413                 return -ENOMEM;
414
415         if (addr) {
416                 addr = ALIGN(addr, HPAGE_SIZE);
417                 vma = find_vma(mm, addr);
418                 if (TASK_SIZE - len >= addr &&
419                     (!vma || addr + len <= vma->vm_start))
420                         return addr;
421         }
422         if (mm->get_unmapped_area == arch_get_unmapped_area)
423                 return hugetlb_get_unmapped_area_bottomup(file, addr, len,
424                                 pgoff, flags);
425         else
426                 return hugetlb_get_unmapped_area_topdown(file, addr, len,
427                                 pgoff, flags);
428 }
429
430 #endif /*HAVE_ARCH_HUGETLB_UNMAPPED_AREA*/
431