[PATCH] freepgt: hugetlb_free_pgd_range
[linux-2.6.git] / arch / ppc64 / mm / hugetlbpage.c
1 /*
2  * PPC64 (POWER4) Huge TLB Page Support for Kernel.
3  *
4  * Copyright (C) 2003 David Gibson, IBM Corporation.
5  *
6  * Based on the IA-32 version:
7  * Copyright (C) 2002, Rohit Seth <rohit.seth@intel.com>
8  */
9
10 #include <linux/init.h>
11 #include <linux/fs.h>
12 #include <linux/mm.h>
13 #include <linux/hugetlb.h>
14 #include <linux/pagemap.h>
15 #include <linux/smp_lock.h>
16 #include <linux/slab.h>
17 #include <linux/err.h>
18 #include <linux/sysctl.h>
19 #include <asm/mman.h>
20 #include <asm/pgalloc.h>
21 #include <asm/tlb.h>
22 #include <asm/tlbflush.h>
23 #include <asm/mmu_context.h>
24 #include <asm/machdep.h>
25 #include <asm/cputable.h>
26 #include <asm/tlb.h>
27
28 #include <linux/sysctl.h>
29
30 #define HUGEPGDIR_SHIFT         (HPAGE_SHIFT + PAGE_SHIFT - 3)
31 #define HUGEPGDIR_SIZE          (1UL << HUGEPGDIR_SHIFT)
32 #define HUGEPGDIR_MASK          (~(HUGEPGDIR_SIZE-1))
33
34 #define HUGEPTE_INDEX_SIZE      9
35 #define HUGEPGD_INDEX_SIZE      10
36
37 #define PTRS_PER_HUGEPTE        (1 << HUGEPTE_INDEX_SIZE)
38 #define PTRS_PER_HUGEPGD        (1 << HUGEPGD_INDEX_SIZE)
39
40 static inline int hugepgd_index(unsigned long addr)
41 {
42         return (addr & ~REGION_MASK) >> HUGEPGDIR_SHIFT;
43 }
44
45 static pgd_t *hugepgd_offset(struct mm_struct *mm, unsigned long addr)
46 {
47         int index;
48
49         if (! mm->context.huge_pgdir)
50                 return NULL;
51
52
53         index = hugepgd_index(addr);
54         BUG_ON(index >= PTRS_PER_HUGEPGD);
55         return mm->context.huge_pgdir + index;
56 }
57
58 static inline pte_t *hugepte_offset(pgd_t *dir, unsigned long addr)
59 {
60         int index;
61
62         if (pgd_none(*dir))
63                 return NULL;
64
65         index = (addr >> HPAGE_SHIFT) % PTRS_PER_HUGEPTE;
66         return (pte_t *)pgd_page(*dir) + index;
67 }
68
69 static pgd_t *hugepgd_alloc(struct mm_struct *mm, unsigned long addr)
70 {
71         BUG_ON(! in_hugepage_area(mm->context, addr));
72
73         if (! mm->context.huge_pgdir) {
74                 pgd_t *new;
75                 spin_unlock(&mm->page_table_lock);
76                 /* Don't use pgd_alloc(), because we want __GFP_REPEAT */
77                 new = kmem_cache_alloc(zero_cache, GFP_KERNEL | __GFP_REPEAT);
78                 BUG_ON(memcmp(new, empty_zero_page, PAGE_SIZE));
79                 spin_lock(&mm->page_table_lock);
80
81                 /*
82                  * Because we dropped the lock, we should re-check the
83                  * entry, as somebody else could have populated it..
84                  */
85                 if (mm->context.huge_pgdir)
86                         pgd_free(new);
87                 else
88                         mm->context.huge_pgdir = new;
89         }
90         return hugepgd_offset(mm, addr);
91 }
92
93 static pte_t *hugepte_alloc(struct mm_struct *mm, pgd_t *dir,
94                             unsigned long addr)
95 {
96         if (! pgd_present(*dir)) {
97                 pte_t *new;
98
99                 spin_unlock(&mm->page_table_lock);
100                 new = kmem_cache_alloc(zero_cache, GFP_KERNEL | __GFP_REPEAT);
101                 BUG_ON(memcmp(new, empty_zero_page, PAGE_SIZE));
102                 spin_lock(&mm->page_table_lock);
103                 /*
104                  * Because we dropped the lock, we should re-check the
105                  * entry, as somebody else could have populated it..
106                  */
107                 if (pgd_present(*dir)) {
108                         if (new)
109                                 kmem_cache_free(zero_cache, new);
110                 } else {
111                         struct page *ptepage;
112
113                         if (! new)
114                                 return NULL;
115                         ptepage = virt_to_page(new);
116                         ptepage->mapping = (void *) mm;
117                         ptepage->index = addr & HUGEPGDIR_MASK;
118                         pgd_populate(mm, dir, new);
119                 }
120         }
121
122         return hugepte_offset(dir, addr);
123 }
124
125 static pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
126 {
127         pgd_t *pgd;
128
129         BUG_ON(! in_hugepage_area(mm->context, addr));
130
131         pgd = hugepgd_offset(mm, addr);
132         if (! pgd)
133                 return NULL;
134
135         return hugepte_offset(pgd, addr);
136 }
137
138 static pte_t *huge_pte_alloc(struct mm_struct *mm, unsigned long addr)
139 {
140         pgd_t *pgd;
141
142         BUG_ON(! in_hugepage_area(mm->context, addr));
143
144         pgd = hugepgd_alloc(mm, addr);
145         if (! pgd)
146                 return NULL;
147
148         return hugepte_alloc(mm, pgd, addr);
149 }
150
151 static void set_huge_pte(struct mm_struct *mm, struct vm_area_struct *vma,
152                          unsigned long addr, struct page *page,
153                          pte_t *ptep, int write_access)
154 {
155         pte_t entry;
156
157         add_mm_counter(mm, rss, HPAGE_SIZE / PAGE_SIZE);
158         if (write_access) {
159                 entry =
160                     pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
161         } else {
162                 entry = pte_wrprotect(mk_pte(page, vma->vm_page_prot));
163         }
164         entry = pte_mkyoung(entry);
165         entry = pte_mkhuge(entry);
166
167         set_pte_at(mm, addr, ptep, entry);
168 }
169
170 /*
171  * This function checks for proper alignment of input addr and len parameters.
172  */
173 int is_aligned_hugepage_range(unsigned long addr, unsigned long len)
174 {
175         if (len & ~HPAGE_MASK)
176                 return -EINVAL;
177         if (addr & ~HPAGE_MASK)
178                 return -EINVAL;
179         if (! (within_hugepage_low_range(addr, len)
180                || within_hugepage_high_range(addr, len)) )
181                 return -EINVAL;
182         return 0;
183 }
184
185 static void flush_segments(void *parm)
186 {
187         u16 segs = (unsigned long) parm;
188         unsigned long i;
189
190         asm volatile("isync" : : : "memory");
191
192         for (i = 0; i < 16; i++) {
193                 if (! (segs & (1U << i)))
194                         continue;
195                 asm volatile("slbie %0" : : "r" (i << SID_SHIFT));
196         }
197
198         asm volatile("isync" : : : "memory");
199 }
200
201 static int prepare_low_seg_for_htlb(struct mm_struct *mm, unsigned long seg)
202 {
203         unsigned long start = seg << SID_SHIFT;
204         unsigned long end = (seg+1) << SID_SHIFT;
205         struct vm_area_struct *vma;
206         unsigned long addr;
207         struct mmu_gather *tlb;
208
209         BUG_ON(seg >= 16);
210
211         /* Check no VMAs are in the region */
212         vma = find_vma(mm, start);
213         if (vma && (vma->vm_start < end))
214                 return -EBUSY;
215
216         /* Clean up any leftover PTE pages in the region */
217         spin_lock(&mm->page_table_lock);
218         tlb = tlb_gather_mmu(mm, 0);
219         for (addr = start; addr < end; addr += PMD_SIZE) {
220                 pgd_t *pgd = pgd_offset(mm, addr);
221                 pmd_t *pmd;
222                 struct page *page;
223                 pte_t *pte;
224                 int i;
225
226                 if (pgd_none(*pgd))
227                         continue;
228                 pmd = pmd_offset(pgd, addr);
229                 if (!pmd || pmd_none(*pmd))
230                         continue;
231                 if (pmd_bad(*pmd)) {
232                         pmd_ERROR(*pmd);
233                         pmd_clear(pmd);
234                         continue;
235                 }
236                 pte = (pte_t *)pmd_page_kernel(*pmd);
237                 /* No VMAs, so there should be no PTEs, check just in case. */
238                 for (i = 0; i < PTRS_PER_PTE; i++) {
239                         BUG_ON(!pte_none(*pte));
240                         pte++;
241                 }
242                 page = pmd_page(*pmd);
243                 pmd_clear(pmd);
244                 mm->nr_ptes--;
245                 dec_page_state(nr_page_table_pages);
246                 pte_free_tlb(tlb, page);
247         }
248         tlb_finish_mmu(tlb, start, end);
249         spin_unlock(&mm->page_table_lock);
250
251         return 0;
252 }
253
254 static int open_low_hpage_segs(struct mm_struct *mm, u16 newsegs)
255 {
256         unsigned long i;
257
258         newsegs &= ~(mm->context.htlb_segs);
259         if (! newsegs)
260                 return 0; /* The segments we want are already open */
261
262         for (i = 0; i < 16; i++)
263                 if ((1 << i) & newsegs)
264                         if (prepare_low_seg_for_htlb(mm, i) != 0)
265                                 return -EBUSY;
266
267         mm->context.htlb_segs |= newsegs;
268
269         /* update the paca copy of the context struct */
270         get_paca()->context = mm->context;
271
272         /* the context change must make it to memory before the flush,
273          * so that further SLB misses do the right thing. */
274         mb();
275         on_each_cpu(flush_segments, (void *)(unsigned long)newsegs, 0, 1);
276
277         return 0;
278 }
279
280 int prepare_hugepage_range(unsigned long addr, unsigned long len)
281 {
282         if (within_hugepage_high_range(addr, len))
283                 return 0;
284         else if ((addr < 0x100000000UL) && ((addr+len) < 0x100000000UL)) {
285                 int err;
286                 /* Yes, we need both tests, in case addr+len overflows
287                  * 64-bit arithmetic */
288                 err = open_low_hpage_segs(current->mm,
289                                           LOW_ESID_MASK(addr, len));
290                 if (err)
291                         printk(KERN_DEBUG "prepare_hugepage_range(%lx, %lx)"
292                                " failed (segs: 0x%04hx)\n", addr, len,
293                                LOW_ESID_MASK(addr, len));
294                 return err;
295         }
296
297         return -EINVAL;
298 }
299
300 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
301                         struct vm_area_struct *vma)
302 {
303         pte_t *src_pte, *dst_pte, entry;
304         struct page *ptepage;
305         unsigned long addr = vma->vm_start;
306         unsigned long end = vma->vm_end;
307         int err = -ENOMEM;
308
309         while (addr < end) {
310                 dst_pte = huge_pte_alloc(dst, addr);
311                 if (!dst_pte)
312                         goto out;
313
314                 src_pte = huge_pte_offset(src, addr);
315                 entry = *src_pte;
316                 
317                 ptepage = pte_page(entry);
318                 get_page(ptepage);
319                 add_mm_counter(dst, rss, HPAGE_SIZE / PAGE_SIZE);
320                 set_pte_at(dst, addr, dst_pte, entry);
321
322                 addr += HPAGE_SIZE;
323         }
324
325         err = 0;
326  out:
327         return err;
328 }
329
330 int
331 follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
332                     struct page **pages, struct vm_area_struct **vmas,
333                     unsigned long *position, int *length, int i)
334 {
335         unsigned long vpfn, vaddr = *position;
336         int remainder = *length;
337
338         WARN_ON(!is_vm_hugetlb_page(vma));
339
340         vpfn = vaddr/PAGE_SIZE;
341         while (vaddr < vma->vm_end && remainder) {
342                 if (pages) {
343                         pte_t *pte;
344                         struct page *page;
345
346                         pte = huge_pte_offset(mm, vaddr);
347
348                         /* hugetlb should be locked, and hence, prefaulted */
349                         WARN_ON(!pte || pte_none(*pte));
350
351                         page = &pte_page(*pte)[vpfn % (HPAGE_SIZE/PAGE_SIZE)];
352
353                         WARN_ON(!PageCompound(page));
354
355                         get_page(page);
356                         pages[i] = page;
357                 }
358
359                 if (vmas)
360                         vmas[i] = vma;
361
362                 vaddr += PAGE_SIZE;
363                 ++vpfn;
364                 --remainder;
365                 ++i;
366         }
367
368         *length = remainder;
369         *position = vaddr;
370
371         return i;
372 }
373
374 struct page *
375 follow_huge_addr(struct mm_struct *mm, unsigned long address, int write)
376 {
377         pte_t *ptep;
378         struct page *page;
379
380         if (! in_hugepage_area(mm->context, address))
381                 return ERR_PTR(-EINVAL);
382
383         ptep = huge_pte_offset(mm, address);
384         page = pte_page(*ptep);
385         if (page)
386                 page += (address % HPAGE_SIZE) / PAGE_SIZE;
387
388         return page;
389 }
390
391 int pmd_huge(pmd_t pmd)
392 {
393         return 0;
394 }
395
396 struct page *
397 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
398                 pmd_t *pmd, int write)
399 {
400         BUG();
401         return NULL;
402 }
403
404 void unmap_hugepage_range(struct vm_area_struct *vma,
405                           unsigned long start, unsigned long end)
406 {
407         struct mm_struct *mm = vma->vm_mm;
408         unsigned long addr;
409         pte_t *ptep;
410         struct page *page;
411
412         WARN_ON(!is_vm_hugetlb_page(vma));
413         BUG_ON((start % HPAGE_SIZE) != 0);
414         BUG_ON((end % HPAGE_SIZE) != 0);
415
416         for (addr = start; addr < end; addr += HPAGE_SIZE) {
417                 pte_t pte;
418
419                 ptep = huge_pte_offset(mm, addr);
420                 if (!ptep || pte_none(*ptep))
421                         continue;
422
423                 pte = *ptep;
424                 page = pte_page(pte);
425                 pte_clear(mm, addr, ptep);
426
427                 put_page(page);
428         }
429         add_mm_counter(mm, rss, -((end - start) >> PAGE_SHIFT));
430         flush_tlb_pending();
431 }
432
433 int hugetlb_prefault(struct address_space *mapping, struct vm_area_struct *vma)
434 {
435         struct mm_struct *mm = current->mm;
436         unsigned long addr;
437         int ret = 0;
438
439         WARN_ON(!is_vm_hugetlb_page(vma));
440         BUG_ON((vma->vm_start % HPAGE_SIZE) != 0);
441         BUG_ON((vma->vm_end % HPAGE_SIZE) != 0);
442
443         spin_lock(&mm->page_table_lock);
444         for (addr = vma->vm_start; addr < vma->vm_end; addr += HPAGE_SIZE) {
445                 unsigned long idx;
446                 pte_t *pte = huge_pte_alloc(mm, addr);
447                 struct page *page;
448
449                 if (!pte) {
450                         ret = -ENOMEM;
451                         goto out;
452                 }
453                 if (! pte_none(*pte))
454                         continue;
455
456                 idx = ((addr - vma->vm_start) >> HPAGE_SHIFT)
457                         + (vma->vm_pgoff >> (HPAGE_SHIFT - PAGE_SHIFT));
458                 page = find_get_page(mapping, idx);
459                 if (!page) {
460                         /* charge the fs quota first */
461                         if (hugetlb_get_quota(mapping)) {
462                                 ret = -ENOMEM;
463                                 goto out;
464                         }
465                         page = alloc_huge_page();
466                         if (!page) {
467                                 hugetlb_put_quota(mapping);
468                                 ret = -ENOMEM;
469                                 goto out;
470                         }
471                         ret = add_to_page_cache(page, mapping, idx, GFP_ATOMIC);
472                         if (! ret) {
473                                 unlock_page(page);
474                         } else {
475                                 hugetlb_put_quota(mapping);
476                                 free_huge_page(page);
477                                 goto out;
478                         }
479                 }
480                 set_huge_pte(mm, vma, addr, page, pte, vma->vm_flags & VM_WRITE);
481         }
482 out:
483         spin_unlock(&mm->page_table_lock);
484         return ret;
485 }
486
487 /* Because we have an exclusive hugepage region which lies within the
488  * normal user address space, we have to take special measures to make
489  * non-huge mmap()s evade the hugepage reserved regions. */
490 unsigned long arch_get_unmapped_area(struct file *filp, unsigned long addr,
491                                      unsigned long len, unsigned long pgoff,
492                                      unsigned long flags)
493 {
494         struct mm_struct *mm = current->mm;
495         struct vm_area_struct *vma;
496         unsigned long start_addr;
497
498         if (len > TASK_SIZE)
499                 return -ENOMEM;
500
501         if (addr) {
502                 addr = PAGE_ALIGN(addr);
503                 vma = find_vma(mm, addr);
504                 if (((TASK_SIZE - len) >= addr)
505                     && (!vma || (addr+len) <= vma->vm_start)
506                     && !is_hugepage_only_range(mm, addr,len))
507                         return addr;
508         }
509         start_addr = addr = mm->free_area_cache;
510
511 full_search:
512         vma = find_vma(mm, addr);
513         while (TASK_SIZE - len >= addr) {
514                 BUG_ON(vma && (addr >= vma->vm_end));
515
516                 if (touches_hugepage_low_range(mm, addr, len)) {
517                         addr = ALIGN(addr+1, 1<<SID_SHIFT);
518                         vma = find_vma(mm, addr);
519                         continue;
520                 }
521                 if (touches_hugepage_high_range(addr, len)) {
522                         addr = TASK_HPAGE_END;
523                         vma = find_vma(mm, addr);
524                         continue;
525                 }
526                 if (!vma || addr + len <= vma->vm_start) {
527                         /*
528                          * Remember the place where we stopped the search:
529                          */
530                         mm->free_area_cache = addr + len;
531                         return addr;
532                 }
533                 addr = vma->vm_end;
534                 vma = vma->vm_next;
535         }
536
537         /* Make sure we didn't miss any holes */
538         if (start_addr != TASK_UNMAPPED_BASE) {
539                 start_addr = addr = TASK_UNMAPPED_BASE;
540                 goto full_search;
541         }
542         return -ENOMEM;
543 }
544
545 /*
546  * This mmap-allocator allocates new areas top-down from below the
547  * stack's low limit (the base):
548  *
549  * Because we have an exclusive hugepage region which lies within the
550  * normal user address space, we have to take special measures to make
551  * non-huge mmap()s evade the hugepage reserved regions.
552  */
553 unsigned long
554 arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
555                           const unsigned long len, const unsigned long pgoff,
556                           const unsigned long flags)
557 {
558         struct vm_area_struct *vma, *prev_vma;
559         struct mm_struct *mm = current->mm;
560         unsigned long base = mm->mmap_base, addr = addr0;
561         int first_time = 1;
562
563         /* requested length too big for entire address space */
564         if (len > TASK_SIZE)
565                 return -ENOMEM;
566
567         /* dont allow allocations above current base */
568         if (mm->free_area_cache > base)
569                 mm->free_area_cache = base;
570
571         /* requesting a specific address */
572         if (addr) {
573                 addr = PAGE_ALIGN(addr);
574                 vma = find_vma(mm, addr);
575                 if (TASK_SIZE - len >= addr &&
576                                 (!vma || addr + len <= vma->vm_start)
577                                 && !is_hugepage_only_range(mm, addr,len))
578                         return addr;
579         }
580
581 try_again:
582         /* make sure it can fit in the remaining address space */
583         if (mm->free_area_cache < len)
584                 goto fail;
585
586         /* either no address requested or cant fit in requested address hole */
587         addr = (mm->free_area_cache - len) & PAGE_MASK;
588         do {
589 hugepage_recheck:
590                 if (touches_hugepage_low_range(mm, addr, len)) {
591                         addr = (addr & ((~0) << SID_SHIFT)) - len;
592                         goto hugepage_recheck;
593                 } else if (touches_hugepage_high_range(addr, len)) {
594                         addr = TASK_HPAGE_BASE - len;
595                 }
596
597                 /*
598                  * Lookup failure means no vma is above this address,
599                  * i.e. return with success:
600                  */
601                 if (!(vma = find_vma_prev(mm, addr, &prev_vma)))
602                         return addr;
603
604                 /*
605                  * new region fits between prev_vma->vm_end and
606                  * vma->vm_start, use it:
607                  */
608                 if (addr+len <= vma->vm_start &&
609                                 (!prev_vma || (addr >= prev_vma->vm_end)))
610                         /* remember the address as a hint for next time */
611                         return (mm->free_area_cache = addr);
612                 else
613                         /* pull free_area_cache down to the first hole */
614                         if (mm->free_area_cache == vma->vm_end)
615                                 mm->free_area_cache = vma->vm_start;
616
617                 /* try just below the current vma->vm_start */
618                 addr = vma->vm_start-len;
619         } while (len <= vma->vm_start);
620
621 fail:
622         /*
623          * if hint left us with no space for the requested
624          * mapping then try again:
625          */
626         if (first_time) {
627                 mm->free_area_cache = base;
628                 first_time = 0;
629                 goto try_again;
630         }
631         /*
632          * A failed mmap() very likely causes application failure,
633          * so fall back to the bottom-up function here. This scenario
634          * can happen with large stack limits and large mmap()
635          * allocations.
636          */
637         mm->free_area_cache = TASK_UNMAPPED_BASE;
638         addr = arch_get_unmapped_area(filp, addr0, len, pgoff, flags);
639         /*
640          * Restore the topdown base:
641          */
642         mm->free_area_cache = base;
643
644         return addr;
645 }
646
647 static unsigned long htlb_get_low_area(unsigned long len, u16 segmask)
648 {
649         unsigned long addr = 0;
650         struct vm_area_struct *vma;
651
652         vma = find_vma(current->mm, addr);
653         while (addr + len <= 0x100000000UL) {
654                 BUG_ON(vma && (addr >= vma->vm_end)); /* invariant */
655
656                 if (! __within_hugepage_low_range(addr, len, segmask)) {
657                         addr = ALIGN(addr+1, 1<<SID_SHIFT);
658                         vma = find_vma(current->mm, addr);
659                         continue;
660                 }
661
662                 if (!vma || (addr + len) <= vma->vm_start)
663                         return addr;
664                 addr = ALIGN(vma->vm_end, HPAGE_SIZE);
665                 /* Depending on segmask this might not be a confirmed
666                  * hugepage region, so the ALIGN could have skipped
667                  * some VMAs */
668                 vma = find_vma(current->mm, addr);
669         }
670
671         return -ENOMEM;
672 }
673
674 static unsigned long htlb_get_high_area(unsigned long len)
675 {
676         unsigned long addr = TASK_HPAGE_BASE;
677         struct vm_area_struct *vma;
678
679         vma = find_vma(current->mm, addr);
680         for (vma = find_vma(current->mm, addr);
681              addr + len <= TASK_HPAGE_END;
682              vma = vma->vm_next) {
683                 BUG_ON(vma && (addr >= vma->vm_end)); /* invariant */
684                 BUG_ON(! within_hugepage_high_range(addr, len));
685
686                 if (!vma || (addr + len) <= vma->vm_start)
687                         return addr;
688                 addr = ALIGN(vma->vm_end, HPAGE_SIZE);
689                 /* Because we're in a hugepage region, this alignment
690                  * should not skip us over any VMAs */
691         }
692
693         return -ENOMEM;
694 }
695
696 unsigned long hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
697                                         unsigned long len, unsigned long pgoff,
698                                         unsigned long flags)
699 {
700         if (len & ~HPAGE_MASK)
701                 return -EINVAL;
702
703         if (!cpu_has_feature(CPU_FTR_16M_PAGE))
704                 return -EINVAL;
705
706         if (test_thread_flag(TIF_32BIT)) {
707                 int lastshift = 0;
708                 u16 segmask, cursegs = current->mm->context.htlb_segs;
709
710                 /* First see if we can do the mapping in the existing
711                  * low hpage segments */
712                 addr = htlb_get_low_area(len, cursegs);
713                 if (addr != -ENOMEM)
714                         return addr;
715
716                 for (segmask = LOW_ESID_MASK(0x100000000UL-len, len);
717                      ! lastshift; segmask >>=1) {
718                         if (segmask & 1)
719                                 lastshift = 1;
720
721                         addr = htlb_get_low_area(len, cursegs | segmask);
722                         if ((addr != -ENOMEM)
723                             && open_low_hpage_segs(current->mm, segmask) == 0)
724                                 return addr;
725                 }
726                 printk(KERN_DEBUG "hugetlb_get_unmapped_area() unable to open"
727                        " enough segments\n");
728                 return -ENOMEM;
729         } else {
730                 return htlb_get_high_area(len);
731         }
732 }
733
734 void hugetlb_mm_free_pgd(struct mm_struct *mm)
735 {
736         int i;
737         pgd_t *pgdir;
738
739         spin_lock(&mm->page_table_lock);
740
741         pgdir = mm->context.huge_pgdir;
742         if (! pgdir)
743                 goto out;
744
745         mm->context.huge_pgdir = NULL;
746
747         /* cleanup any hugepte pages leftover */
748         for (i = 0; i < PTRS_PER_HUGEPGD; i++) {
749                 pgd_t *pgd = pgdir + i;
750
751                 if (! pgd_none(*pgd)) {
752                         pte_t *pte = (pte_t *)pgd_page(*pgd);
753                         struct page *ptepage = virt_to_page(pte);
754
755                         ptepage->mapping = NULL;
756
757                         BUG_ON(memcmp(pte, empty_zero_page, PAGE_SIZE));
758                         kmem_cache_free(zero_cache, pte);
759                 }
760                 pgd_clear(pgd);
761         }
762
763         BUG_ON(memcmp(pgdir, empty_zero_page, PAGE_SIZE));
764         kmem_cache_free(zero_cache, pgdir);
765
766  out:
767         spin_unlock(&mm->page_table_lock);
768 }
769
770 int hash_huge_page(struct mm_struct *mm, unsigned long access,
771                    unsigned long ea, unsigned long vsid, int local)
772 {
773         pte_t *ptep;
774         unsigned long va, vpn;
775         pte_t old_pte, new_pte;
776         unsigned long hpteflags, prpn;
777         long slot;
778         int err = 1;
779
780         spin_lock(&mm->page_table_lock);
781
782         ptep = huge_pte_offset(mm, ea);
783
784         /* Search the Linux page table for a match with va */
785         va = (vsid << 28) | (ea & 0x0fffffff);
786         vpn = va >> HPAGE_SHIFT;
787
788         /*
789          * If no pte found or not present, send the problem up to
790          * do_page_fault
791          */
792         if (unlikely(!ptep || pte_none(*ptep)))
793                 goto out;
794
795 /*      BUG_ON(pte_bad(*ptep)); */
796
797         /* 
798          * Check the user's access rights to the page.  If access should be
799          * prevented then send the problem up to do_page_fault.
800          */
801         if (unlikely(access & ~pte_val(*ptep)))
802                 goto out;
803         /*
804          * At this point, we have a pte (old_pte) which can be used to build
805          * or update an HPTE. There are 2 cases:
806          *
807          * 1. There is a valid (present) pte with no associated HPTE (this is 
808          *      the most common case)
809          * 2. There is a valid (present) pte with an associated HPTE. The
810          *      current values of the pp bits in the HPTE prevent access
811          *      because we are doing software DIRTY bit management and the
812          *      page is currently not DIRTY. 
813          */
814
815
816         old_pte = *ptep;
817         new_pte = old_pte;
818
819         hpteflags = 0x2 | (! (pte_val(new_pte) & _PAGE_RW));
820         /* _PAGE_EXEC -> HW_NO_EXEC since it's inverted */
821         hpteflags |= ((pte_val(new_pte) & _PAGE_EXEC) ? 0 : HW_NO_EXEC);
822
823         /* Check if pte already has an hpte (case 2) */
824         if (unlikely(pte_val(old_pte) & _PAGE_HASHPTE)) {
825                 /* There MIGHT be an HPTE for this pte */
826                 unsigned long hash, slot;
827
828                 hash = hpt_hash(vpn, 1);
829                 if (pte_val(old_pte) & _PAGE_SECONDARY)
830                         hash = ~hash;
831                 slot = (hash & htab_hash_mask) * HPTES_PER_GROUP;
832                 slot += (pte_val(old_pte) & _PAGE_GROUP_IX) >> 12;
833
834                 if (ppc_md.hpte_updatepp(slot, hpteflags, va, 1, local) == -1)
835                         pte_val(old_pte) &= ~_PAGE_HPTEFLAGS;
836         }
837
838         if (likely(!(pte_val(old_pte) & _PAGE_HASHPTE))) {
839                 unsigned long hash = hpt_hash(vpn, 1);
840                 unsigned long hpte_group;
841
842                 prpn = pte_pfn(old_pte);
843
844 repeat:
845                 hpte_group = ((hash & htab_hash_mask) *
846                               HPTES_PER_GROUP) & ~0x7UL;
847
848                 /* Update the linux pte with the HPTE slot */
849                 pte_val(new_pte) &= ~_PAGE_HPTEFLAGS;
850                 pte_val(new_pte) |= _PAGE_HASHPTE;
851
852                 /* Add in WIMG bits */
853                 /* XXX We should store these in the pte */
854                 hpteflags |= _PAGE_COHERENT;
855
856                 slot = ppc_md.hpte_insert(hpte_group, va, prpn, 0,
857                                           hpteflags, 0, 1);
858
859                 /* Primary is full, try the secondary */
860                 if (unlikely(slot == -1)) {
861                         pte_val(new_pte) |= _PAGE_SECONDARY;
862                         hpte_group = ((~hash & htab_hash_mask) *
863                                       HPTES_PER_GROUP) & ~0x7UL; 
864                         slot = ppc_md.hpte_insert(hpte_group, va, prpn,
865                                                   1, hpteflags, 0, 1);
866                         if (slot == -1) {
867                                 if (mftb() & 0x1)
868                                         hpte_group = ((hash & htab_hash_mask) * HPTES_PER_GROUP) & ~0x7UL;
869
870                                 ppc_md.hpte_remove(hpte_group);
871                                 goto repeat;
872                         }
873                 }
874
875                 if (unlikely(slot == -2))
876                         panic("hash_huge_page: pte_insert failed\n");
877
878                 pte_val(new_pte) |= (slot<<12) & _PAGE_GROUP_IX;
879
880                 /* 
881                  * No need to use ldarx/stdcx here because all who
882                  * might be updating the pte will hold the
883                  * page_table_lock
884                  */
885                 *ptep = new_pte;
886         }
887
888         err = 0;
889
890  out:
891         spin_unlock(&mm->page_table_lock);
892
893         return err;
894 }