4 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
8 * demand-loading started 01.12.91 - seems it is high on the list of
9 * things wanted, and it should be easy to implement. - Linus
13 * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
14 * pages started 02.12.91, seems to work. - Linus.
16 * Tested sharing by executing about 30 /bin/sh: under the old kernel it
17 * would have taken more than the 6M I have free, but it worked well as
20 * Also corrected some "invalidate()"s - I wasn't doing enough of them.
24 * Real VM (paging to/from disk) started 18.12.91. Much more work and
25 * thought has to go into this. Oh, well..
26 * 19.12.91 - works, somewhat. Sometimes I get faults, don't know why.
27 * Found it. Everything seems to work now.
28 * 20.12.91 - Ok, making the swap-device changeable like the root.
32 * 05.04.94 - Multi-page memory management added for v1.1.
33 * Idea by Alex Bligh (alex@cconcepts.co.uk)
35 * 16.07.99 - Support of BIGMEM added by Gerhard Wichert, Siemens AG
36 * (Gerhard.Wichert@pdb.siemens.de)
39 #include <linux/kernel_stat.h>
41 #include <linux/hugetlb.h>
42 #include <linux/mman.h>
43 #include <linux/swap.h>
44 #include <linux/highmem.h>
45 #include <linux/pagemap.h>
46 #include <linux/vcache.h>
47 #include <linux/rmap-locking.h>
49 #include <asm/pgalloc.h>
51 #include <asm/uaccess.h>
53 #include <asm/tlbflush.h>
54 #include <asm/pgtable.h>
56 #include <linux/swapops.h>
58 #ifndef CONFIG_DISCONTIGMEM
59 /* use the per-pgdat data instead for discontigmem - mbligh */
60 unsigned long max_mapnr;
64 unsigned long num_physpages;
66 struct page *highmem_start_page;
69 * We special-case the C-O-W ZERO_PAGE, because it's such
70 * a common occurrence (no need to read the page to know
71 * that it's zero - better for the cache and memory subsystem).
73 static inline void copy_cow_page(struct page * from, struct page * to, unsigned long address)
75 if (from == ZERO_PAGE(address)) {
76 clear_user_highpage(to, address);
79 copy_user_highpage(to, from, address);
83 * Note: this doesn't free the actual pages themselves. That
84 * has been handled earlier when unmapping all the memory regions.
86 static inline void free_one_pmd(struct mmu_gather *tlb, pmd_t * dir)
97 page = pmd_page(*dir);
99 pgtable_remove_rmap(page);
100 pte_free_tlb(tlb, page);
103 static inline void free_one_pgd(struct mmu_gather *tlb, pgd_t * dir)
115 pmd = pmd_offset(dir, 0);
117 for (j = 0; j < PTRS_PER_PMD ; j++) {
118 prefetchw(pmd + j + PREFETCH_STRIDE/sizeof(*pmd));
119 free_one_pmd(tlb, pmd+j);
121 pmd_free_tlb(tlb, pmd);
125 * This function clears all user-level page tables of a process - this
126 * is needed by execve(), so that old pages aren't in the way.
128 * Must be called with pagetable lock held.
130 void clear_page_tables(struct mmu_gather *tlb, unsigned long first, int nr)
132 pgd_t * page_dir = tlb->mm->pgd;
136 free_one_pgd(tlb, page_dir);
141 pte_t * pte_alloc_map(struct mm_struct *mm, pmd_t *pmd, unsigned long address)
143 if (!pmd_present(*pmd)) {
146 spin_unlock(&mm->page_table_lock);
147 new = pte_alloc_one(mm, address);
148 spin_lock(&mm->page_table_lock);
153 * Because we dropped the lock, we should re-check the
154 * entry, as somebody else could have populated it..
156 if (pmd_present(*pmd)) {
160 pgtable_add_rmap(new, mm, address);
161 pmd_populate(mm, pmd, new);
164 return pte_offset_map(pmd, address);
167 pte_t * pte_alloc_kernel(struct mm_struct *mm, pmd_t *pmd, unsigned long address)
169 if (!pmd_present(*pmd)) {
172 spin_unlock(&mm->page_table_lock);
173 new = pte_alloc_one_kernel(mm, address);
174 spin_lock(&mm->page_table_lock);
179 * Because we dropped the lock, we should re-check the
180 * entry, as somebody else could have populated it..
182 if (pmd_present(*pmd)) {
183 pte_free_kernel(new);
186 pgtable_add_rmap(virt_to_page(new), mm, address);
187 pmd_populate_kernel(mm, pmd, new);
190 return pte_offset_kernel(pmd, address);
192 #define PTE_TABLE_MASK ((PTRS_PER_PTE-1) * sizeof(pte_t))
193 #define PMD_TABLE_MASK ((PTRS_PER_PMD-1) * sizeof(pmd_t))
196 * copy one vm_area from one task to the other. Assumes the page tables
197 * already present in the new task to be cleared in the whole range
198 * covered by this vma.
200 * 08Jan98 Merged into one routine from several inline routines to reduce
201 * variable count and make things faster. -jj
203 * dst->page_table_lock is held on entry and exit,
204 * but may be dropped within pmd_alloc() and pte_alloc_map().
206 int copy_page_range(struct mm_struct *dst, struct mm_struct *src,
207 struct vm_area_struct *vma)
209 pgd_t * src_pgd, * dst_pgd;
210 unsigned long address = vma->vm_start;
211 unsigned long end = vma->vm_end;
213 struct pte_chain *pte_chain = NULL;
215 if (is_vm_hugetlb_page(vma))
216 return copy_hugetlb_page_range(dst, src, vma);
218 pte_chain = pte_chain_alloc(GFP_ATOMIC);
220 spin_unlock(&dst->page_table_lock);
221 pte_chain = pte_chain_alloc(GFP_KERNEL);
222 spin_lock(&dst->page_table_lock);
227 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
228 src_pgd = pgd_offset(src, address)-1;
229 dst_pgd = pgd_offset(dst, address)-1;
232 pmd_t * src_pmd, * dst_pmd;
234 src_pgd++; dst_pgd++;
238 if (pgd_none(*src_pgd))
239 goto skip_copy_pmd_range;
240 if (pgd_bad(*src_pgd)) {
243 skip_copy_pmd_range: address = (address + PGDIR_SIZE) & PGDIR_MASK;
244 if (!address || (address >= end))
249 src_pmd = pmd_offset(src_pgd, address);
250 dst_pmd = pmd_alloc(dst, dst_pgd, address);
255 pte_t * src_pte, * dst_pte;
259 if (pmd_none(*src_pmd))
260 goto skip_copy_pte_range;
261 if (pmd_bad(*src_pmd)) {
265 address = (address + PMD_SIZE) & PMD_MASK;
268 goto cont_copy_pmd_range;
271 dst_pte = pte_alloc_map(dst, dst_pmd, address);
274 spin_lock(&src->page_table_lock);
275 src_pte = pte_offset_map_nested(src_pmd, address);
277 pte_t pte = *src_pte;
284 goto cont_copy_pte_range_noset;
285 /* pte contains position in swap, so copy. */
286 if (!pte_present(pte)) {
288 swap_duplicate(pte_to_swp_entry(pte));
289 set_pte(dst_pte, pte);
290 goto cont_copy_pte_range_noset;
293 /* the pte points outside of valid memory, the
294 * mapping is assumed to be good, meaningful
295 * and not mapped via rmap - duplicate the
300 page = pfn_to_page(pfn);
302 if (!page || PageReserved(page)) {
303 set_pte(dst_pte, pte);
304 goto cont_copy_pte_range_noset;
308 * If it's a COW mapping, write protect it both
309 * in the parent and the child
312 ptep_set_wrprotect(src_pte);
317 * If it's a shared mapping, mark it clean in
320 if (vma->vm_flags & VM_SHARED)
321 pte = pte_mkclean(pte);
322 pte = pte_mkold(pte);
326 set_pte(dst_pte, pte);
327 pte_chain = page_add_rmap(page, dst_pte,
330 goto cont_copy_pte_range_noset;
331 pte_chain = pte_chain_alloc(GFP_ATOMIC);
333 goto cont_copy_pte_range_noset;
336 * pte_chain allocation failed, and we need to
339 pte_unmap_nested(src_pte);
341 spin_unlock(&src->page_table_lock);
342 spin_unlock(&dst->page_table_lock);
343 pte_chain = pte_chain_alloc(GFP_KERNEL);
344 spin_lock(&dst->page_table_lock);
347 spin_lock(&src->page_table_lock);
348 dst_pte = pte_offset_map(dst_pmd, address);
349 src_pte = pte_offset_map_nested(src_pmd,
351 cont_copy_pte_range_noset:
352 address += PAGE_SIZE;
353 if (address >= end) {
354 pte_unmap_nested(src_pte);
360 } while ((unsigned long)src_pte & PTE_TABLE_MASK);
361 pte_unmap_nested(src_pte-1);
362 pte_unmap(dst_pte-1);
363 spin_unlock(&src->page_table_lock);
368 } while ((unsigned long)src_pmd & PMD_TABLE_MASK);
371 spin_unlock(&src->page_table_lock);
373 pte_chain_free(pte_chain);
376 pte_chain_free(pte_chain);
381 zap_pte_range(struct mmu_gather *tlb, pmd_t * pmd,
382 unsigned long address, unsigned long size)
384 unsigned long offset;
394 ptep = pte_offset_map(pmd, address);
395 offset = address & ~PMD_MASK;
396 if (offset + size > PMD_SIZE)
397 size = PMD_SIZE - offset;
399 for (offset=0; offset < size; ptep++, offset += PAGE_SIZE) {
403 if (pte_present(pte)) {
404 unsigned long pfn = pte_pfn(pte);
406 pte = ptep_get_and_clear(ptep);
407 tlb_remove_tlb_entry(tlb, ptep, address+offset);
408 if (pfn_valid(pfn)) {
409 struct page *page = pfn_to_page(pfn);
410 if (!PageReserved(page)) {
412 set_page_dirty(page);
413 if (page->mapping && pte_young(pte) &&
414 !PageSwapCache(page))
415 mark_page_accessed(page);
417 page_remove_rmap(page, ptep);
418 tlb_remove_page(tlb, page);
423 free_swap_and_cache(pte_to_swp_entry(pte));
431 zap_pmd_range(struct mmu_gather *tlb, pgd_t * dir,
432 unsigned long address, unsigned long size)
444 pmd = pmd_offset(dir, address);
445 end = address + size;
446 if (end > ((address + PGDIR_SIZE) & PGDIR_MASK))
447 end = ((address + PGDIR_SIZE) & PGDIR_MASK);
449 zap_pte_range(tlb, pmd, address, end - address);
450 address = (address + PMD_SIZE) & PMD_MASK;
452 } while (address < end);
455 void unmap_page_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
456 unsigned long address, unsigned long end)
460 if (is_vm_hugetlb_page(vma)) {
461 unmap_hugepage_range(vma, address, end);
465 BUG_ON(address >= end);
467 dir = pgd_offset(vma->vm_mm, address);
468 tlb_start_vma(tlb, vma);
470 zap_pmd_range(tlb, dir, address, end - address);
471 address = (address + PGDIR_SIZE) & PGDIR_MASK;
473 } while (address && (address < end));
474 tlb_end_vma(tlb, vma);
477 /* Dispose of an entire struct mmu_gather per rescheduling point */
478 #if defined(CONFIG_SMP) && defined(CONFIG_PREEMPT)
479 #define ZAP_BLOCK_SIZE (FREE_PTE_NR * PAGE_SIZE)
482 /* For UP, 256 pages at a time gives nice low latency */
483 #if !defined(CONFIG_SMP) && defined(CONFIG_PREEMPT)
484 #define ZAP_BLOCK_SIZE (256 * PAGE_SIZE)
487 /* No preempt: go for the best straight-line efficiency */
488 #if !defined(CONFIG_PREEMPT)
489 #define ZAP_BLOCK_SIZE (~(0UL))
493 * unmap_vmas - unmap a range of memory covered by a list of vma's
494 * @tlbp: address of the caller's struct mmu_gather
495 * @mm: the controlling mm_struct
496 * @vma: the starting vma
497 * @start_addr: virtual address at which to start unmapping
498 * @end_addr: virtual address at which to end unmapping
499 * @nr_accounted: Place number of unmapped pages in vm-accountable vma's here
501 * Returns the number of vma's which were covered by the unmapping.
503 * Unmap all pages in the vma list. Called under page_table_lock.
505 * We aim to not hold page_table_lock for too long (for scheduling latency
506 * reasons). So zap pages in ZAP_BLOCK_SIZE bytecounts. This means we need to
507 * return the ending mmu_gather to the caller.
509 * Only addresses between `start' and `end' will be unmapped.
511 * The VMA list must be sorted in ascending virtual address order.
513 * unmap_vmas() assumes that the caller will flush the whole unmapped address
514 * range after unmap_vmas() returns. So the only responsibility here is to
515 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
516 * drops the lock and schedules.
518 int unmap_vmas(struct mmu_gather **tlbp, struct mm_struct *mm,
519 struct vm_area_struct *vma, unsigned long start_addr,
520 unsigned long end_addr, unsigned long *nr_accounted)
522 unsigned long zap_bytes = ZAP_BLOCK_SIZE;
523 unsigned long tlb_start; /* For tlb_finish_mmu */
524 int tlb_start_valid = 0;
527 if (vma) { /* debug. killme. */
528 if (end_addr <= vma->vm_start)
529 printk("%s: end_addr(0x%08lx) <= vm_start(0x%08lx)\n",
530 __FUNCTION__, end_addr, vma->vm_start);
531 if (start_addr >= vma->vm_end)
532 printk("%s: start_addr(0x%08lx) <= vm_end(0x%08lx)\n",
533 __FUNCTION__, start_addr, vma->vm_end);
536 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next) {
540 start = max(vma->vm_start, start_addr);
541 if (start >= vma->vm_end)
543 end = min(vma->vm_end, end_addr);
544 if (end <= vma->vm_start)
547 if (vma->vm_flags & VM_ACCOUNT)
548 *nr_accounted += (end - start) >> PAGE_SHIFT;
551 while (start != end) {
554 if (is_vm_hugetlb_page(vma))
557 block = min(zap_bytes, end - start);
559 if (!tlb_start_valid) {
564 unmap_page_range(*tlbp, vma, start, start + block);
567 if ((long)zap_bytes > 0)
569 if (need_resched()) {
570 tlb_finish_mmu(*tlbp, tlb_start, start);
571 cond_resched_lock(&mm->page_table_lock);
572 *tlbp = tlb_gather_mmu(mm, 0);
575 zap_bytes = ZAP_BLOCK_SIZE;
577 if (vma->vm_next && vma->vm_next->vm_start < vma->vm_end)
578 printk("%s: VMA list is not sorted correctly!\n",
585 * zap_page_range - remove user pages in a given range
586 * @vma: vm_area_struct holding the applicable pages
587 * @address: starting address of pages to zap
588 * @size: number of bytes to zap
590 void zap_page_range(struct vm_area_struct *vma,
591 unsigned long address, unsigned long size)
593 struct mm_struct *mm = vma->vm_mm;
594 struct mmu_gather *tlb;
595 unsigned long end = address + size;
596 unsigned long nr_accounted = 0;
600 if (is_vm_hugetlb_page(vma)) {
601 zap_hugepage_range(vma, address, size);
606 spin_lock(&mm->page_table_lock);
607 tlb = tlb_gather_mmu(mm, 0);
608 unmap_vmas(&tlb, mm, vma, address, end, &nr_accounted);
609 tlb_finish_mmu(tlb, address, end);
610 spin_unlock(&mm->page_table_lock);
614 * Do a quick page-table lookup for a single page.
615 * mm->page_table_lock must be held.
618 follow_page(struct mm_struct *mm, unsigned long address, int write)
624 struct vm_area_struct *vma;
626 vma = hugepage_vma(mm, address);
628 return follow_huge_addr(mm, vma, address, write);
630 pgd = pgd_offset(mm, address);
631 if (pgd_none(*pgd) || pgd_bad(*pgd))
634 pmd = pmd_offset(pgd, address);
638 return follow_huge_pmd(mm, address, pmd, write);
642 ptep = pte_offset_map(pmd, address);
648 if (pte_present(pte)) {
649 if (!write || (pte_write(pte) && pte_dirty(pte))) {
652 return pfn_to_page(pfn);
661 * Given a physical address, is there a useful struct page pointing to
662 * it? This may become more complex in the future if we start dealing
663 * with IO-aperture pages for direct-IO.
666 static inline struct page *get_page_map(struct page *page)
668 if (!pfn_valid(page_to_pfn(page)))
674 int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
675 unsigned long start, int len, int write, int force,
676 struct page **pages, struct vm_area_struct **vmas)
682 * Require read or write permissions.
683 * If 'force' is set, we only require the "MAY" flags.
685 flags = write ? (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
686 flags &= force ? (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
690 struct vm_area_struct * vma;
692 vma = find_extend_vma(mm, start);
694 #ifdef VSYSCALL_START
695 if (!vma && start >= VSYSCALL_START && start < VSYSCALL_END) {
696 static struct vm_area_struct fixmap_vma = {
697 /* Catch users - if there are any valid
698 ones, we can make this be "&init_mm" or
701 .vm_start = FIXADDR_START,
702 .vm_end = FIXADDR_TOP,
703 .vm_page_prot = PAGE_READONLY,
704 .vm_flags = VM_READ | VM_EXEC,
706 unsigned long pg = start & PAGE_MASK;
710 pgd = pgd_offset_k(pg);
712 return i ? : -EFAULT;
713 pmd = pmd_offset(pgd, pg);
715 return i ? : -EFAULT;
716 pte = pte_offset_kernel(pmd, pg);
717 if (!pte || !pte_present(*pte) || !pte_user(*pte) ||
718 !(write ? pte_write(*pte) : pte_read(*pte)))
719 return i ? : -EFAULT;
721 pages[i] = pte_page(*pte);
725 vmas[i] = &fixmap_vma;
733 if (!vma || (pages && (vma->vm_flags & VM_IO))
734 || !(flags & vma->vm_flags))
735 return i ? : -EFAULT;
737 if (is_vm_hugetlb_page(vma)) {
738 i = follow_hugetlb_page(mm, vma, pages, vmas,
742 spin_lock(&mm->page_table_lock);
745 while (!(map = follow_page(mm, start, write))) {
746 spin_unlock(&mm->page_table_lock);
747 switch (handle_mm_fault(mm,vma,start,write)) {
754 case VM_FAULT_SIGBUS:
755 return i ? i : -EFAULT;
757 return i ? i : -ENOMEM;
761 spin_lock(&mm->page_table_lock);
764 pages[i] = get_page_map(map);
766 spin_unlock(&mm->page_table_lock);
768 page_cache_release(pages[i]);
772 flush_dcache_page(pages[i]);
773 if (!PageReserved(pages[i]))
774 page_cache_get(pages[i]);
781 } while(len && start < vma->vm_end);
782 spin_unlock(&mm->page_table_lock);
788 static void zeromap_pte_range(pte_t * pte, unsigned long address,
789 unsigned long size, pgprot_t prot)
793 address &= ~PMD_MASK;
794 end = address + size;
798 pte_t zero_pte = pte_wrprotect(mk_pte(ZERO_PAGE(address), prot));
799 BUG_ON(!pte_none(*pte));
800 set_pte(pte, zero_pte);
801 address += PAGE_SIZE;
803 } while (address && (address < end));
806 static inline int zeromap_pmd_range(struct mm_struct *mm, pmd_t * pmd, unsigned long address,
807 unsigned long size, pgprot_t prot)
811 address &= ~PGDIR_MASK;
812 end = address + size;
813 if (end > PGDIR_SIZE)
816 pte_t * pte = pte_alloc_map(mm, pmd, address);
819 zeromap_pte_range(pte, address, end - address, prot);
821 address = (address + PMD_SIZE) & PMD_MASK;
823 } while (address && (address < end));
827 int zeromap_page_range(struct vm_area_struct *vma, unsigned long address, unsigned long size, pgprot_t prot)
831 unsigned long beg = address;
832 unsigned long end = address + size;
833 struct mm_struct *mm = vma->vm_mm;
835 dir = pgd_offset(mm, address);
836 flush_cache_range(vma, beg, end);
840 spin_lock(&mm->page_table_lock);
842 pmd_t *pmd = pmd_alloc(mm, dir, address);
846 error = zeromap_pmd_range(mm, pmd, address, end - address, prot);
849 address = (address + PGDIR_SIZE) & PGDIR_MASK;
851 } while (address && (address < end));
852 flush_tlb_range(vma, beg, end);
853 spin_unlock(&mm->page_table_lock);
858 * maps a range of physical memory into the requested pages. the old
859 * mappings are removed. any references to nonexistent pages results
860 * in null mappings (currently treated as "copy-on-access")
862 static inline void remap_pte_range(pte_t * pte, unsigned long address, unsigned long size,
863 unsigned long phys_addr, pgprot_t prot)
868 address &= ~PMD_MASK;
869 end = address + size;
872 pfn = phys_addr >> PAGE_SHIFT;
874 BUG_ON(!pte_none(*pte));
875 if (!pfn_valid(pfn) || PageReserved(pfn_to_page(pfn)))
876 set_pte(pte, pfn_pte(pfn, prot));
877 address += PAGE_SIZE;
880 } while (address && (address < end));
883 static inline int remap_pmd_range(struct mm_struct *mm, pmd_t * pmd, unsigned long address, unsigned long size,
884 unsigned long phys_addr, pgprot_t prot)
886 unsigned long base, end;
888 base = address & PGDIR_MASK;
889 address &= ~PGDIR_MASK;
890 end = address + size;
891 if (end > PGDIR_SIZE)
893 phys_addr -= address;
895 pte_t * pte = pte_alloc_map(mm, pmd, base + address);
898 remap_pte_range(pte, base + address, end - address, address + phys_addr, prot);
900 address = (address + PMD_SIZE) & PMD_MASK;
902 } while (address && (address < end));
906 /* Note: this is only safe if the mm semaphore is held when called. */
907 int remap_page_range(struct vm_area_struct *vma, unsigned long from, unsigned long phys_addr, unsigned long size, pgprot_t prot)
911 unsigned long beg = from;
912 unsigned long end = from + size;
913 struct mm_struct *mm = vma->vm_mm;
916 dir = pgd_offset(mm, from);
917 flush_cache_range(vma, beg, end);
921 spin_lock(&mm->page_table_lock);
923 pmd_t *pmd = pmd_alloc(mm, dir, from);
927 error = remap_pmd_range(mm, pmd, from, end - from, phys_addr + from, prot);
930 from = (from + PGDIR_SIZE) & PGDIR_MASK;
932 } while (from && (from < end));
933 flush_tlb_range(vma, beg, end);
934 spin_unlock(&mm->page_table_lock);
939 * Establish a new mapping:
940 * - flush the old one
941 * - update the page tables
942 * - inform the TLB about the new one
944 * We hold the mm semaphore for reading and vma->vm_mm->page_table_lock
946 static inline void establish_pte(struct vm_area_struct * vma, unsigned long address, pte_t *page_table, pte_t entry)
948 set_pte(page_table, entry);
949 flush_tlb_page(vma, address);
950 update_mmu_cache(vma, address, entry);
954 * We hold the mm semaphore for reading and vma->vm_mm->page_table_lock
956 static inline void break_cow(struct vm_area_struct * vma, struct page * new_page, unsigned long address,
959 invalidate_vcache(address, vma->vm_mm, new_page);
960 flush_cache_page(vma, address);
961 establish_pte(vma, address, page_table, pte_mkwrite(pte_mkdirty(mk_pte(new_page, vma->vm_page_prot))));
965 * This routine handles present pages, when users try to write
966 * to a shared page. It is done by copying the page to a new address
967 * and decrementing the shared-page counter for the old page.
969 * Goto-purists beware: the only reason for goto's here is that it results
970 * in better assembly code.. The "default" path will see no jumps at all.
972 * Note that this routine assumes that the protection checks have been
973 * done by the caller (the low-level page fault routine in most cases).
974 * Thus we can safely just mark it writable once we've done any necessary
977 * We also mark the page dirty at this point even though the page will
978 * change only once the write actually happens. This avoids a few races,
979 * and potentially makes it more efficient.
981 * We hold the mm semaphore and the page_table_lock on entry and exit
982 * with the page_table_lock released.
984 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct * vma,
985 unsigned long address, pte_t *page_table, pmd_t *pmd, pte_t pte)
987 struct page *old_page, *new_page;
988 unsigned long pfn = pte_pfn(pte);
989 struct pte_chain *pte_chain = NULL;
992 if (unlikely(!pfn_valid(pfn))) {
994 * This should really halt the system so it can be debugged or
995 * at least the kernel stops what it's doing before it corrupts
996 * data, but for the moment just pretend this is OOM.
998 pte_unmap(page_table);
999 printk(KERN_ERR "do_wp_page: bogus page at address %08lx\n",
1003 old_page = pfn_to_page(pfn);
1005 if (!TestSetPageLocked(old_page)) {
1006 int reuse = can_share_swap_page(old_page);
1007 unlock_page(old_page);
1009 flush_cache_page(vma, address);
1010 establish_pte(vma, address, page_table,
1011 pte_mkyoung(pte_mkdirty(pte_mkwrite(pte))));
1012 pte_unmap(page_table);
1013 ret = VM_FAULT_MINOR;
1017 pte_unmap(page_table);
1020 * Ok, we need to copy. Oh, well..
1022 page_cache_get(old_page);
1023 spin_unlock(&mm->page_table_lock);
1025 pte_chain = pte_chain_alloc(GFP_KERNEL);
1028 new_page = alloc_page(GFP_HIGHUSER);
1031 copy_cow_page(old_page,new_page,address);
1034 * Re-check the pte - we dropped the lock
1036 spin_lock(&mm->page_table_lock);
1037 page_table = pte_offset_map(pmd, address);
1038 if (pte_same(*page_table, pte)) {
1039 if (PageReserved(old_page))
1041 page_remove_rmap(old_page, page_table);
1042 break_cow(vma, new_page, address, page_table);
1043 pte_chain = page_add_rmap(new_page, page_table, pte_chain);
1044 lru_cache_add_active(new_page);
1046 /* Free the old page.. */
1047 new_page = old_page;
1049 pte_unmap(page_table);
1050 page_cache_release(new_page);
1051 page_cache_release(old_page);
1052 ret = VM_FAULT_MINOR;
1056 page_cache_release(old_page);
1060 spin_unlock(&mm->page_table_lock);
1061 pte_chain_free(pte_chain);
1065 static void vmtruncate_list(struct list_head *head, unsigned long pgoff)
1067 unsigned long start, end, len, diff;
1068 struct vm_area_struct *vma;
1069 struct list_head *curr;
1071 list_for_each(curr, head) {
1072 vma = list_entry(curr, struct vm_area_struct, shared);
1073 start = vma->vm_start;
1077 /* mapping wholly truncated? */
1078 if (vma->vm_pgoff >= pgoff) {
1079 zap_page_range(vma, start, len);
1083 /* mapping wholly unaffected? */
1084 len = len >> PAGE_SHIFT;
1085 diff = pgoff - vma->vm_pgoff;
1089 /* Ok, partially affected.. */
1090 start += diff << PAGE_SHIFT;
1091 len = (len - diff) << PAGE_SHIFT;
1092 zap_page_range(vma, start, len);
1097 * Handle all mappings that got truncated by a "truncate()"
1100 * NOTE! We have to be ready to update the memory sharing
1101 * between the file and the memory map for a potential last
1102 * incomplete page. Ugly, but necessary.
1104 int vmtruncate(struct inode * inode, loff_t offset)
1106 unsigned long pgoff;
1107 struct address_space *mapping = inode->i_mapping;
1108 unsigned long limit;
1110 if (inode->i_size < offset)
1112 inode->i_size = offset;
1113 down(&mapping->i_shared_sem);
1114 if (list_empty(&mapping->i_mmap) && list_empty(&mapping->i_mmap_shared))
1117 pgoff = (offset + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1118 if (!list_empty(&mapping->i_mmap))
1119 vmtruncate_list(&mapping->i_mmap, pgoff);
1120 if (!list_empty(&mapping->i_mmap_shared))
1121 vmtruncate_list(&mapping->i_mmap_shared, pgoff);
1124 up(&mapping->i_shared_sem);
1125 truncate_inode_pages(mapping, offset);
1129 limit = current->rlim[RLIMIT_FSIZE].rlim_cur;
1130 if (limit != RLIM_INFINITY && offset > limit)
1132 if (offset > inode->i_sb->s_maxbytes)
1134 inode->i_size = offset;
1137 if (inode->i_op && inode->i_op->truncate)
1138 inode->i_op->truncate(inode);
1141 send_sig(SIGXFSZ, current, 0);
1147 * Primitive swap readahead code. We simply read an aligned block of
1148 * (1 << page_cluster) entries in the swap area. This method is chosen
1149 * because it doesn't cost us any seek time. We also make sure to queue
1150 * the 'original' request together with the readahead ones...
1152 void swapin_readahead(swp_entry_t entry)
1155 struct page *new_page;
1156 unsigned long offset;
1159 * Get the number of handles we should do readahead io to.
1161 num = valid_swaphandles(entry, &offset);
1162 for (i = 0; i < num; offset++, i++) {
1163 /* Ok, do the async read-ahead now */
1164 new_page = read_swap_cache_async(swp_entry(swp_type(entry),
1168 page_cache_release(new_page);
1170 lru_add_drain(); /* Push any new pages onto the LRU now */
1174 * We hold the mm semaphore and the page_table_lock on entry and
1175 * should release the pagetable lock on exit..
1177 static int do_swap_page(struct mm_struct * mm,
1178 struct vm_area_struct * vma, unsigned long address,
1179 pte_t *page_table, pmd_t *pmd, pte_t orig_pte, int write_access)
1182 swp_entry_t entry = pte_to_swp_entry(orig_pte);
1184 int ret = VM_FAULT_MINOR;
1185 struct pte_chain *pte_chain = NULL;
1187 pte_unmap(page_table);
1188 spin_unlock(&mm->page_table_lock);
1189 page = lookup_swap_cache(entry);
1191 swapin_readahead(entry);
1192 page = read_swap_cache_async(entry);
1195 * Back out if somebody else faulted in this pte while
1196 * we released the page table lock.
1198 spin_lock(&mm->page_table_lock);
1199 page_table = pte_offset_map(pmd, address);
1200 if (pte_same(*page_table, orig_pte))
1203 ret = VM_FAULT_MINOR;
1204 pte_unmap(page_table);
1205 spin_unlock(&mm->page_table_lock);
1209 /* Had to read the page from swap area: Major fault */
1210 ret = VM_FAULT_MAJOR;
1211 inc_page_state(pgmajfault);
1214 mark_page_accessed(page);
1215 pte_chain = pte_chain_alloc(GFP_KERNEL);
1223 * Back out if somebody else faulted in this pte while we
1224 * released the page table lock.
1226 spin_lock(&mm->page_table_lock);
1227 page_table = pte_offset_map(pmd, address);
1228 if (!pte_same(*page_table, orig_pte)) {
1229 pte_unmap(page_table);
1230 spin_unlock(&mm->page_table_lock);
1232 page_cache_release(page);
1233 ret = VM_FAULT_MINOR;
1237 /* The page isn't present yet, go ahead with the fault. */
1241 remove_exclusive_swap_page(page);
1244 pte = mk_pte(page, vma->vm_page_prot);
1245 if (write_access && can_share_swap_page(page))
1246 pte = pte_mkdirty(pte_mkwrite(pte));
1249 flush_icache_page(vma, page);
1250 set_pte(page_table, pte);
1251 pte_chain = page_add_rmap(page, page_table, pte_chain);
1253 /* No need to invalidate - it was non-present before */
1254 update_mmu_cache(vma, address, pte);
1255 pte_unmap(page_table);
1256 spin_unlock(&mm->page_table_lock);
1258 pte_chain_free(pte_chain);
1263 * We are called with the MM semaphore and page_table_lock
1264 * spinlock held to protect against concurrent faults in
1265 * multithreaded programs.
1268 do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
1269 pte_t *page_table, pmd_t *pmd, int write_access,
1273 struct page * page = ZERO_PAGE(addr);
1274 struct pte_chain *pte_chain;
1277 pte_chain = pte_chain_alloc(GFP_ATOMIC);
1279 pte_unmap(page_table);
1280 spin_unlock(&mm->page_table_lock);
1281 pte_chain = pte_chain_alloc(GFP_KERNEL);
1284 spin_lock(&mm->page_table_lock);
1285 page_table = pte_offset_map(pmd, addr);
1288 /* Read-only mapping of ZERO_PAGE. */
1289 entry = pte_wrprotect(mk_pte(ZERO_PAGE(addr), vma->vm_page_prot));
1291 /* ..except if it's a write access */
1293 /* Allocate our own private page. */
1294 pte_unmap(page_table);
1295 spin_unlock(&mm->page_table_lock);
1297 page = alloc_page(GFP_HIGHUSER);
1300 clear_user_highpage(page, addr);
1302 spin_lock(&mm->page_table_lock);
1303 page_table = pte_offset_map(pmd, addr);
1305 if (!pte_none(*page_table)) {
1306 pte_unmap(page_table);
1307 page_cache_release(page);
1308 spin_unlock(&mm->page_table_lock);
1309 ret = VM_FAULT_MINOR;
1313 entry = pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
1314 lru_cache_add_active(page);
1315 mark_page_accessed(page);
1318 set_pte(page_table, entry);
1319 /* ignores ZERO_PAGE */
1320 pte_chain = page_add_rmap(page, page_table, pte_chain);
1321 pte_unmap(page_table);
1323 /* No need to invalidate - it was non-present before */
1324 update_mmu_cache(vma, addr, entry);
1325 spin_unlock(&mm->page_table_lock);
1326 ret = VM_FAULT_MINOR;
1332 pte_chain_free(pte_chain);
1337 * do_no_page() tries to create a new page mapping. It aggressively
1338 * tries to share with existing pages, but makes a separate copy if
1339 * the "write_access" parameter is true in order to avoid the next
1342 * As this is called only for pages that do not currently exist, we
1343 * do not need to flush old virtual caches or the TLB.
1345 * This is called with the MM semaphore held and the page table
1346 * spinlock held. Exit with the spinlock released.
1349 do_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
1350 unsigned long address, int write_access, pte_t *page_table, pmd_t *pmd)
1352 struct page * new_page;
1354 struct pte_chain *pte_chain;
1357 if (!vma->vm_ops || !vma->vm_ops->nopage)
1358 return do_anonymous_page(mm, vma, page_table,
1359 pmd, write_access, address);
1360 pte_unmap(page_table);
1361 spin_unlock(&mm->page_table_lock);
1363 new_page = vma->vm_ops->nopage(vma, address & PAGE_MASK, 0);
1365 /* no page was available -- either SIGBUS or OOM */
1366 if (new_page == NOPAGE_SIGBUS)
1367 return VM_FAULT_SIGBUS;
1368 if (new_page == NOPAGE_OOM)
1369 return VM_FAULT_OOM;
1371 pte_chain = pte_chain_alloc(GFP_KERNEL);
1376 * Should we do an early C-O-W break?
1378 if (write_access && !(vma->vm_flags & VM_SHARED)) {
1379 struct page * page = alloc_page(GFP_HIGHUSER);
1381 page_cache_release(new_page);
1384 copy_user_highpage(page, new_page, address);
1385 page_cache_release(new_page);
1386 lru_cache_add_active(page);
1390 spin_lock(&mm->page_table_lock);
1391 page_table = pte_offset_map(pmd, address);
1394 * This silly early PAGE_DIRTY setting removes a race
1395 * due to the bad i386 page protection. But it's valid
1396 * for other architectures too.
1398 * Note that if write_access is true, we either now have
1399 * an exclusive copy of the page, or this is a shared mapping,
1400 * so we can make it writable and dirty to avoid having to
1401 * handle that later.
1403 /* Only go through if we didn't race with anybody else... */
1404 if (pte_none(*page_table)) {
1406 flush_icache_page(vma, new_page);
1407 entry = mk_pte(new_page, vma->vm_page_prot);
1409 entry = pte_mkwrite(pte_mkdirty(entry));
1410 set_pte(page_table, entry);
1411 pte_chain = page_add_rmap(new_page, page_table, pte_chain);
1412 pte_unmap(page_table);
1414 /* One of our sibling threads was faster, back out. */
1415 pte_unmap(page_table);
1416 page_cache_release(new_page);
1417 spin_unlock(&mm->page_table_lock);
1418 ret = VM_FAULT_MINOR;
1422 /* no need to invalidate: a not-present page shouldn't be cached */
1423 update_mmu_cache(vma, address, entry);
1424 spin_unlock(&mm->page_table_lock);
1425 ret = VM_FAULT_MAJOR;
1430 pte_chain_free(pte_chain);
1435 * Fault of a previously existing named mapping. Repopulate the pte
1436 * from the encoded file_pte if possible. This enables swappable
1439 static int do_file_page(struct mm_struct * mm, struct vm_area_struct * vma,
1440 unsigned long address, int write_access, pte_t *pte, pmd_t *pmd)
1442 unsigned long pgoff;
1445 BUG_ON(!vma->vm_ops || !vma->vm_ops->nopage);
1447 * Fall back to the linear mapping if the fs does not support
1450 if (!vma->vm_ops || !vma->vm_ops->populate ||
1451 (write_access && !(vma->vm_flags & VM_SHARED))) {
1453 return do_no_page(mm, vma, address, write_access, pte, pmd);
1456 pgoff = pte_to_pgoff(*pte);
1459 spin_unlock(&mm->page_table_lock);
1461 err = vma->vm_ops->populate(vma, address & PAGE_MASK, PAGE_SIZE, vma->vm_page_prot, pgoff, 0);
1463 return VM_FAULT_OOM;
1465 return VM_FAULT_SIGBUS;
1466 return VM_FAULT_MAJOR;
1470 * These routines also need to handle stuff like marking pages dirty
1471 * and/or accessed for architectures that don't do it in hardware (most
1472 * RISC architectures). The early dirtying is also good on the i386.
1474 * There is also a hook called "update_mmu_cache()" that architectures
1475 * with external mmu caches can use to update those (ie the Sparc or
1476 * PowerPC hashed page tables that act as extended TLBs).
1478 * Note the "page_table_lock". It is to protect against kswapd removing
1479 * pages from under us. Note that kswapd only ever _removes_ pages, never
1480 * adds them. As such, once we have noticed that the page is not present,
1481 * we can drop the lock early.
1483 * The adding of pages is protected by the MM semaphore (which we hold),
1484 * so we don't need to worry about a page being suddenly been added into
1487 * We enter with the pagetable spinlock held, we are supposed to
1488 * release it when done.
1490 static inline int handle_pte_fault(struct mm_struct *mm,
1491 struct vm_area_struct * vma, unsigned long address,
1492 int write_access, pte_t *pte, pmd_t *pmd)
1497 if (!pte_present(entry)) {
1499 * If it truly wasn't present, we know that kswapd
1500 * and the PTE updates will not touch it later. So
1503 if (pte_none(entry))
1504 return do_no_page(mm, vma, address, write_access, pte, pmd);
1505 if (pte_file(entry))
1506 return do_file_page(mm, vma, address, write_access, pte, pmd);
1507 return do_swap_page(mm, vma, address, pte, pmd, entry, write_access);
1511 if (!pte_write(entry))
1512 return do_wp_page(mm, vma, address, pte, pmd, entry);
1514 entry = pte_mkdirty(entry);
1516 entry = pte_mkyoung(entry);
1517 establish_pte(vma, address, pte, entry);
1519 spin_unlock(&mm->page_table_lock);
1520 return VM_FAULT_MINOR;
1524 * By the time we get here, we already hold the mm semaphore
1526 int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct * vma,
1527 unsigned long address, int write_access)
1532 __set_current_state(TASK_RUNNING);
1533 pgd = pgd_offset(mm, address);
1535 inc_page_state(pgfault);
1537 if (is_vm_hugetlb_page(vma))
1538 return VM_FAULT_SIGBUS; /* mapping truncation does this. */
1541 * We need the page table lock to synchronize with kswapd
1542 * and the SMP-safe atomic PTE updates.
1544 spin_lock(&mm->page_table_lock);
1545 pmd = pmd_alloc(mm, pgd, address);
1548 pte_t * pte = pte_alloc_map(mm, pmd, address);
1550 return handle_pte_fault(mm, vma, address, write_access, pte, pmd);
1552 spin_unlock(&mm->page_table_lock);
1553 return VM_FAULT_OOM;
1557 * Allocate page middle directory.
1559 * We've already handled the fast-path in-line, and we own the
1562 * On a two-level page table, this ends up actually being entirely
1565 pmd_t *__pmd_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
1569 spin_unlock(&mm->page_table_lock);
1570 new = pmd_alloc_one(mm, address);
1571 spin_lock(&mm->page_table_lock);
1576 * Because we dropped the lock, we should re-check the
1577 * entry, as somebody else could have populated it..
1579 if (pgd_present(*pgd)) {
1583 pgd_populate(mm, pgd, new);
1585 return pmd_offset(pgd, address);
1588 int make_pages_present(unsigned long addr, unsigned long end)
1590 int ret, len, write;
1591 struct vm_area_struct * vma;
1593 vma = find_vma(current->mm, addr);
1594 write = (vma->vm_flags & VM_WRITE) != 0;
1597 if (end > vma->vm_end)
1599 len = (end+PAGE_SIZE-1)/PAGE_SIZE-addr/PAGE_SIZE;
1600 ret = get_user_pages(current, current->mm, addr,
1601 len, write, 0, NULL, NULL);
1602 return ret == len ? 0 : -1;
1606 * Map a vmalloc()-space virtual address to the physical page.
1608 struct page * vmalloc_to_page(void * vmalloc_addr)
1610 unsigned long addr = (unsigned long) vmalloc_addr;
1611 struct page *page = NULL;
1612 pgd_t *pgd = pgd_offset_k(addr);
1616 if (!pgd_none(*pgd)) {
1617 pmd = pmd_offset(pgd, addr);
1618 if (!pmd_none(*pmd)) {
1620 ptep = pte_offset_map(pmd, addr);
1622 if (pte_present(pte))
1623 page = pte_page(pte);