Merge ../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 #include <asm/pgalloc.h>
19
20 static unsigned long page_table_shareable(struct vm_area_struct *svma,
21                                 struct vm_area_struct *vma,
22                                 unsigned long addr, pgoff_t idx)
23 {
24         unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
25                                 svma->vm_start;
26         unsigned long sbase = saddr & PUD_MASK;
27         unsigned long s_end = sbase + PUD_SIZE;
28
29         /*
30          * match the virtual addresses, permission and the alignment of the
31          * page table page.
32          */
33         if (pmd_index(addr) != pmd_index(saddr) ||
34             vma->vm_flags != svma->vm_flags ||
35             sbase < svma->vm_start || svma->vm_end < s_end)
36                 return 0;
37
38         return saddr;
39 }
40
41 static int vma_shareable(struct vm_area_struct *vma, unsigned long addr)
42 {
43         unsigned long base = addr & PUD_MASK;
44         unsigned long end = base + PUD_SIZE;
45
46         /*
47          * check on proper vm_flags and page table alignment
48          */
49         if (vma->vm_flags & VM_MAYSHARE &&
50             vma->vm_start <= base && end <= vma->vm_end)
51                 return 1;
52         return 0;
53 }
54
55 /*
56  * search for a shareable pmd page for hugetlb.
57  */
58 static void huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
59 {
60         struct vm_area_struct *vma = find_vma(mm, addr);
61         struct address_space *mapping = vma->vm_file->f_mapping;
62         pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
63                         vma->vm_pgoff;
64         struct prio_tree_iter iter;
65         struct vm_area_struct *svma;
66         unsigned long saddr;
67         pte_t *spte = NULL;
68
69         if (!vma_shareable(vma, addr))
70                 return;
71
72         spin_lock(&mapping->i_mmap_lock);
73         vma_prio_tree_foreach(svma, &iter, &mapping->i_mmap, idx, idx) {
74                 if (svma == vma)
75                         continue;
76
77                 saddr = page_table_shareable(svma, vma, addr, idx);
78                 if (saddr) {
79                         spte = huge_pte_offset(svma->vm_mm, saddr);
80                         if (spte) {
81                                 get_page(virt_to_page(spte));
82                                 break;
83                         }
84                 }
85         }
86
87         if (!spte)
88                 goto out;
89
90         spin_lock(&mm->page_table_lock);
91         if (pud_none(*pud))
92                 pud_populate(mm, pud, (pmd_t *)((unsigned long)spte & PAGE_MASK));
93         else
94                 put_page(virt_to_page(spte));
95         spin_unlock(&mm->page_table_lock);
96 out:
97         spin_unlock(&mapping->i_mmap_lock);
98 }
99
100 /*
101  * unmap huge page backed by shared pte.
102  *
103  * Hugetlb pte page is ref counted at the time of mapping.  If pte is shared
104  * indicated by page_count > 1, unmap is achieved by clearing pud and
105  * decrementing the ref count. If count == 1, the pte page is not shared.
106  *
107  * called with vma->vm_mm->page_table_lock held.
108  *
109  * returns: 1 successfully unmapped a shared pte page
110  *          0 the underlying pte page is not shared, or it is the last user
111  */
112 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
113 {
114         pgd_t *pgd = pgd_offset(mm, *addr);
115         pud_t *pud = pud_offset(pgd, *addr);
116
117         BUG_ON(page_count(virt_to_page(ptep)) == 0);
118         if (page_count(virt_to_page(ptep)) == 1)
119                 return 0;
120
121         pud_clear(pud);
122         put_page(virt_to_page(ptep));
123         *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
124         return 1;
125 }
126
127 pte_t *huge_pte_alloc(struct mm_struct *mm,
128                         unsigned long addr, unsigned long sz)
129 {
130         pgd_t *pgd;
131         pud_t *pud;
132         pte_t *pte = NULL;
133
134         pgd = pgd_offset(mm, addr);
135         pud = pud_alloc(mm, pgd, addr);
136         if (pud) {
137                 if (sz == PUD_SIZE) {
138                         pte = (pte_t *)pud;
139                 } else {
140                         BUG_ON(sz != PMD_SIZE);
141                         if (pud_none(*pud))
142                                 huge_pmd_share(mm, addr, pud);
143                         pte = (pte_t *) pmd_alloc(mm, pud, addr);
144                 }
145         }
146         BUG_ON(pte && !pte_none(*pte) && !pte_huge(*pte));
147
148         return pte;
149 }
150
151 pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
152 {
153         pgd_t *pgd;
154         pud_t *pud;
155         pmd_t *pmd = NULL;
156
157         pgd = pgd_offset(mm, addr);
158         if (pgd_present(*pgd)) {
159                 pud = pud_offset(pgd, addr);
160                 if (pud_present(*pud)) {
161                         if (pud_large(*pud))
162                                 return (pte_t *)pud;
163                         pmd = pmd_offset(pud, addr);
164                 }
165         }
166         return (pte_t *) pmd;
167 }
168
169 #if 0   /* This is just for testing */
170 struct page *
171 follow_huge_addr(struct mm_struct *mm, unsigned long address, int write)
172 {
173         unsigned long start = address;
174         int length = 1;
175         int nr;
176         struct page *page;
177         struct vm_area_struct *vma;
178
179         vma = find_vma(mm, addr);
180         if (!vma || !is_vm_hugetlb_page(vma))
181                 return ERR_PTR(-EINVAL);
182
183         pte = huge_pte_offset(mm, address);
184
185         /* hugetlb should be locked, and hence, prefaulted */
186         WARN_ON(!pte || pte_none(*pte));
187
188         page = &pte_page(*pte)[vpfn % (HPAGE_SIZE/PAGE_SIZE)];
189
190         WARN_ON(!PageHead(page));
191
192         return page;
193 }
194
195 int pmd_huge(pmd_t pmd)
196 {
197         return 0;
198 }
199
200 int pud_huge(pud_t pud)
201 {
202         return 0;
203 }
204
205 struct page *
206 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
207                 pmd_t *pmd, int write)
208 {
209         return NULL;
210 }
211
212 #else
213
214 struct page *
215 follow_huge_addr(struct mm_struct *mm, unsigned long address, int write)
216 {
217         return ERR_PTR(-EINVAL);
218 }
219
220 int pmd_huge(pmd_t pmd)
221 {
222         return !!(pmd_val(pmd) & _PAGE_PSE);
223 }
224
225 int pud_huge(pud_t pud)
226 {
227         return !!(pud_val(pud) & _PAGE_PSE);
228 }
229
230 struct page *
231 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
232                 pmd_t *pmd, int write)
233 {
234         struct page *page;
235
236         page = pte_page(*(pte_t *)pmd);
237         if (page)
238                 page += ((address & ~PMD_MASK) >> PAGE_SHIFT);
239         return page;
240 }
241
242 struct page *
243 follow_huge_pud(struct mm_struct *mm, unsigned long address,
244                 pud_t *pud, int write)
245 {
246         struct page *page;
247
248         page = pte_page(*(pte_t *)pud);
249         if (page)
250                 page += ((address & ~PUD_MASK) >> PAGE_SHIFT);
251         return page;
252 }
253
254 #endif
255
256 /* x86_64 also uses this file */
257
258 #ifdef HAVE_ARCH_HUGETLB_UNMAPPED_AREA
259 static unsigned long hugetlb_get_unmapped_area_bottomup(struct file *file,
260                 unsigned long addr, unsigned long len,
261                 unsigned long pgoff, unsigned long flags)
262 {
263         struct hstate *h = hstate_file(file);
264         struct mm_struct *mm = current->mm;
265         struct vm_area_struct *vma;
266         unsigned long start_addr;
267
268         if (len > mm->cached_hole_size) {
269                 start_addr = mm->free_area_cache;
270         } else {
271                 start_addr = TASK_UNMAPPED_BASE;
272                 mm->cached_hole_size = 0;
273         }
274
275 full_search:
276         addr = ALIGN(start_addr, huge_page_size(h));
277
278         for (vma = find_vma(mm, addr); ; vma = vma->vm_next) {
279                 /* At this point:  (!vma || addr < vma->vm_end). */
280                 if (TASK_SIZE - len < addr) {
281                         /*
282                          * Start a new search - just in case we missed
283                          * some holes.
284                          */
285                         if (start_addr != TASK_UNMAPPED_BASE) {
286                                 start_addr = TASK_UNMAPPED_BASE;
287                                 mm->cached_hole_size = 0;
288                                 goto full_search;
289                         }
290                         return -ENOMEM;
291                 }
292                 if (!vma || addr + len <= vma->vm_start) {
293                         mm->free_area_cache = addr + len;
294                         return addr;
295                 }
296                 if (addr + mm->cached_hole_size < vma->vm_start)
297                         mm->cached_hole_size = vma->vm_start - addr;
298                 addr = ALIGN(vma->vm_end, huge_page_size(h));
299         }
300 }
301
302 static unsigned long hugetlb_get_unmapped_area_topdown(struct file *file,
303                 unsigned long addr0, unsigned long len,
304                 unsigned long pgoff, unsigned long flags)
305 {
306         struct hstate *h = hstate_file(file);
307         struct mm_struct *mm = current->mm;
308         struct vm_area_struct *vma, *prev_vma;
309         unsigned long base = mm->mmap_base, addr = addr0;
310         unsigned long largest_hole = mm->cached_hole_size;
311         int first_time = 1;
312
313         /* don't allow allocations above current base */
314         if (mm->free_area_cache > base)
315                 mm->free_area_cache = base;
316
317         if (len <= largest_hole) {
318                 largest_hole = 0;
319                 mm->free_area_cache  = base;
320         }
321 try_again:
322         /* make sure it can fit in the remaining address space */
323         if (mm->free_area_cache < len)
324                 goto fail;
325
326         /* either no address requested or cant fit in requested address hole */
327         addr = (mm->free_area_cache - len) & huge_page_mask(h);
328         do {
329                 /*
330                  * Lookup failure means no vma is above this address,
331                  * i.e. return with success:
332                  */
333                 if (!(vma = find_vma_prev(mm, addr, &prev_vma)))
334                         return addr;
335
336                 /*
337                  * new region fits between prev_vma->vm_end and
338                  * vma->vm_start, use it:
339                  */
340                 if (addr + len <= vma->vm_start &&
341                             (!prev_vma || (addr >= prev_vma->vm_end))) {
342                         /* remember the address as a hint for next time */
343                         mm->cached_hole_size = largest_hole;
344                         return (mm->free_area_cache = addr);
345                 } else {
346                         /* pull free_area_cache down to the first hole */
347                         if (mm->free_area_cache == vma->vm_end) {
348                                 mm->free_area_cache = vma->vm_start;
349                                 mm->cached_hole_size = largest_hole;
350                         }
351                 }
352
353                 /* remember the largest hole we saw so far */
354                 if (addr + largest_hole < vma->vm_start)
355                         largest_hole = vma->vm_start - addr;
356
357                 /* try just below the current vma->vm_start */
358                 addr = (vma->vm_start - len) & huge_page_mask(h);
359         } while (len <= vma->vm_start);
360
361 fail:
362         /*
363          * if hint left us with no space for the requested
364          * mapping then try again:
365          */
366         if (first_time) {
367                 mm->free_area_cache = base;
368                 largest_hole = 0;
369                 first_time = 0;
370                 goto try_again;
371         }
372         /*
373          * A failed mmap() very likely causes application failure,
374          * so fall back to the bottom-up function here. This scenario
375          * can happen with large stack limits and large mmap()
376          * allocations.
377          */
378         mm->free_area_cache = TASK_UNMAPPED_BASE;
379         mm->cached_hole_size = ~0UL;
380         addr = hugetlb_get_unmapped_area_bottomup(file, addr0,
381                         len, pgoff, flags);
382
383         /*
384          * Restore the topdown base:
385          */
386         mm->free_area_cache = base;
387         mm->cached_hole_size = ~0UL;
388
389         return addr;
390 }
391
392 unsigned long
393 hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
394                 unsigned long len, unsigned long pgoff, unsigned long flags)
395 {
396         struct hstate *h = hstate_file(file);
397         struct mm_struct *mm = current->mm;
398         struct vm_area_struct *vma;
399
400         if (len & ~huge_page_mask(h))
401                 return -EINVAL;
402         if (len > TASK_SIZE)
403                 return -ENOMEM;
404
405         if (flags & MAP_FIXED) {
406                 if (prepare_hugepage_range(file, addr, len))
407                         return -EINVAL;
408                 return addr;
409         }
410
411         if (addr) {
412                 addr = ALIGN(addr, huge_page_size(h));
413                 vma = find_vma(mm, addr);
414                 if (TASK_SIZE - len >= addr &&
415                     (!vma || addr + len <= vma->vm_start))
416                         return addr;
417         }
418         if (mm->get_unmapped_area == arch_get_unmapped_area)
419                 return hugetlb_get_unmapped_area_bottomup(file, addr, len,
420                                 pgoff, flags);
421         else
422                 return hugetlb_get_unmapped_area_topdown(file, addr, len,
423                                 pgoff, flags);
424 }
425
426 #endif /*HAVE_ARCH_HUGETLB_UNMAPPED_AREA*/
427
428 #ifdef CONFIG_X86_64
429 static __init int setup_hugepagesz(char *opt)
430 {
431         unsigned long ps = memparse(opt, &opt);
432         if (ps == PMD_SIZE) {
433                 hugetlb_add_hstate(PMD_SHIFT - PAGE_SHIFT);
434         } else if (ps == PUD_SIZE && cpu_has_gbpages) {
435                 hugetlb_add_hstate(PUD_SHIFT - PAGE_SHIFT);
436         } else {
437                 printk(KERN_ERR "hugepagesz: Unsupported page size %lu M\n",
438                         ps >> 20);
439                 return 0;
440         }
441         return 1;
442 }
443 __setup("hugepagesz=", setup_hugepagesz);
444 #endif