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 free_one_pmd(tlb, pmd+j);
119 pmd_free_tlb(tlb, pmd);
123 * This function clears all user-level page tables of a process - this
124 * is needed by execve(), so that old pages aren't in the way.
126 * Must be called with pagetable lock held.
128 void clear_page_tables(struct mmu_gather *tlb, unsigned long first, int nr)
130 pgd_t * page_dir = tlb->mm->pgd;
134 free_one_pgd(tlb, page_dir);
139 pte_t * pte_alloc_map(struct mm_struct *mm, pmd_t *pmd, unsigned long address)
141 if (!pmd_present(*pmd)) {
144 spin_unlock(&mm->page_table_lock);
145 new = pte_alloc_one(mm, address);
146 spin_lock(&mm->page_table_lock);
151 * Because we dropped the lock, we should re-check the
152 * entry, as somebody else could have populated it..
154 if (pmd_present(*pmd)) {
158 pgtable_add_rmap(new, mm, address);
159 pmd_populate(mm, pmd, new);
162 return pte_offset_map(pmd, address);
165 pte_t * pte_alloc_kernel(struct mm_struct *mm, pmd_t *pmd, unsigned long address)
167 if (!pmd_present(*pmd)) {
170 spin_unlock(&mm->page_table_lock);
171 new = pte_alloc_one_kernel(mm, address);
172 spin_lock(&mm->page_table_lock);
177 * Because we dropped the lock, we should re-check the
178 * entry, as somebody else could have populated it..
180 if (pmd_present(*pmd)) {
181 pte_free_kernel(new);
184 pgtable_add_rmap(virt_to_page(new), mm, address);
185 pmd_populate_kernel(mm, pmd, new);
188 return pte_offset_kernel(pmd, address);
190 #define PTE_TABLE_MASK ((PTRS_PER_PTE-1) * sizeof(pte_t))
191 #define PMD_TABLE_MASK ((PTRS_PER_PMD-1) * sizeof(pmd_t))
194 * copy one vm_area from one task to the other. Assumes the page tables
195 * already present in the new task to be cleared in the whole range
196 * covered by this vma.
198 * 08Jan98 Merged into one routine from several inline routines to reduce
199 * variable count and make things faster. -jj
201 * dst->page_table_lock is held on entry and exit,
202 * but may be dropped within pmd_alloc() and pte_alloc_map().
204 int copy_page_range(struct mm_struct *dst, struct mm_struct *src,
205 struct vm_area_struct *vma)
207 pgd_t * src_pgd, * dst_pgd;
208 unsigned long address = vma->vm_start;
209 unsigned long end = vma->vm_end;
211 struct pte_chain *pte_chain = NULL;
213 if (is_vm_hugetlb_page(vma))
214 return copy_hugetlb_page_range(dst, src, vma);
216 pte_chain = pte_chain_alloc(GFP_ATOMIC);
218 spin_unlock(&dst->page_table_lock);
219 pte_chain = pte_chain_alloc(GFP_KERNEL);
220 spin_lock(&dst->page_table_lock);
225 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
226 src_pgd = pgd_offset(src, address)-1;
227 dst_pgd = pgd_offset(dst, address)-1;
230 pmd_t * src_pmd, * dst_pmd;
232 src_pgd++; dst_pgd++;
236 if (pgd_none(*src_pgd))
237 goto skip_copy_pmd_range;
238 if (pgd_bad(*src_pgd)) {
241 skip_copy_pmd_range: address = (address + PGDIR_SIZE) & PGDIR_MASK;
242 if (!address || (address >= end))
247 src_pmd = pmd_offset(src_pgd, address);
248 dst_pmd = pmd_alloc(dst, dst_pgd, address);
253 pte_t * src_pte, * dst_pte;
257 if (pmd_none(*src_pmd))
258 goto skip_copy_pte_range;
259 if (pmd_bad(*src_pmd)) {
263 address = (address + PMD_SIZE) & PMD_MASK;
266 goto cont_copy_pmd_range;
269 dst_pte = pte_alloc_map(dst, dst_pmd, address);
272 spin_lock(&src->page_table_lock);
273 src_pte = pte_offset_map_nested(src_pmd, address);
275 pte_t pte = *src_pte;
282 goto cont_copy_pte_range_noset;
283 /* pte contains position in swap, so copy. */
284 if (!pte_present(pte)) {
286 swap_duplicate(pte_to_swp_entry(pte));
287 set_pte(dst_pte, pte);
288 goto cont_copy_pte_range_noset;
291 /* the pte points outside of valid memory, the
292 * mapping is assumed to be good, meaningful
293 * and not mapped via rmap - duplicate the
298 page = pfn_to_page(pfn);
300 if (!page || PageReserved(page)) {
301 set_pte(dst_pte, pte);
302 goto cont_copy_pte_range_noset;
306 * If it's a COW mapping, write protect it both
307 * in the parent and the child
310 ptep_set_wrprotect(src_pte);
315 * If it's a shared mapping, mark it clean in
318 if (vma->vm_flags & VM_SHARED)
319 pte = pte_mkclean(pte);
320 pte = pte_mkold(pte);
324 set_pte(dst_pte, pte);
325 pte_chain = page_add_rmap(page, dst_pte,
328 goto cont_copy_pte_range_noset;
329 pte_chain = pte_chain_alloc(GFP_ATOMIC);
331 goto cont_copy_pte_range_noset;
334 * pte_chain allocation failed, and we need to
337 pte_unmap_nested(src_pte);
339 spin_unlock(&src->page_table_lock);
340 spin_unlock(&dst->page_table_lock);
341 pte_chain = pte_chain_alloc(GFP_KERNEL);
342 spin_lock(&dst->page_table_lock);
345 spin_lock(&src->page_table_lock);
346 dst_pte = pte_offset_map(dst_pmd, address);
347 src_pte = pte_offset_map_nested(src_pmd,
349 cont_copy_pte_range_noset:
350 address += PAGE_SIZE;
351 if (address >= end) {
352 pte_unmap_nested(src_pte);
358 } while ((unsigned long)src_pte & PTE_TABLE_MASK);
359 pte_unmap_nested(src_pte-1);
360 pte_unmap(dst_pte-1);
361 spin_unlock(&src->page_table_lock);
366 } while ((unsigned long)src_pmd & PMD_TABLE_MASK);
369 spin_unlock(&src->page_table_lock);
371 pte_chain_free(pte_chain);
374 pte_chain_free(pte_chain);
379 zap_pte_range(struct mmu_gather *tlb, pmd_t * pmd,
380 unsigned long address, unsigned long size)
382 unsigned long offset;
392 ptep = pte_offset_map(pmd, address);
393 offset = address & ~PMD_MASK;
394 if (offset + size > PMD_SIZE)
395 size = PMD_SIZE - offset;
397 for (offset=0; offset < size; ptep++, offset += PAGE_SIZE) {
401 if (pte_present(pte)) {
402 unsigned long pfn = pte_pfn(pte);
404 pte = ptep_get_and_clear(ptep);
405 tlb_remove_tlb_entry(tlb, ptep, address+offset);
406 if (pfn_valid(pfn)) {
407 struct page *page = pfn_to_page(pfn);
408 if (!PageReserved(page)) {
410 set_page_dirty(page);
411 if (page->mapping && pte_young(pte) &&
412 !PageSwapCache(page))
413 mark_page_accessed(page);
415 page_remove_rmap(page, ptep);
416 tlb_remove_page(tlb, page);
421 free_swap_and_cache(pte_to_swp_entry(pte));
429 zap_pmd_range(struct mmu_gather *tlb, pgd_t * dir,
430 unsigned long address, unsigned long size)
442 pmd = pmd_offset(dir, address);
443 end = address + size;
444 if (end > ((address + PGDIR_SIZE) & PGDIR_MASK))
445 end = ((address + PGDIR_SIZE) & PGDIR_MASK);
447 zap_pte_range(tlb, pmd, address, end - address);
448 address = (address + PMD_SIZE) & PMD_MASK;
450 } while (address < end);
453 void unmap_page_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
454 unsigned long address, unsigned long end)
458 if (is_vm_hugetlb_page(vma)) {
459 unmap_hugepage_range(vma, address, end);
463 BUG_ON(address >= end);
465 dir = pgd_offset(vma->vm_mm, address);
466 tlb_start_vma(tlb, vma);
468 zap_pmd_range(tlb, dir, address, end - address);
469 address = (address + PGDIR_SIZE) & PGDIR_MASK;
471 } while (address && (address < end));
472 tlb_end_vma(tlb, vma);
475 /* Dispose of an entire struct mmu_gather per rescheduling point */
476 #if defined(CONFIG_SMP) && defined(CONFIG_PREEMPT)
477 #define ZAP_BLOCK_SIZE (FREE_PTE_NR * PAGE_SIZE)
480 /* For UP, 256 pages at a time gives nice low latency */
481 #if !defined(CONFIG_SMP) && defined(CONFIG_PREEMPT)
482 #define ZAP_BLOCK_SIZE (256 * PAGE_SIZE)
485 /* No preempt: go for the best straight-line efficiency */
486 #if !defined(CONFIG_PREEMPT)
487 #define ZAP_BLOCK_SIZE (~(0UL))
491 * unmap_vmas - unmap a range of memory covered by a list of vma's
492 * @tlbp: address of the caller's struct mmu_gather
493 * @mm: the controlling mm_struct
494 * @vma: the starting vma
495 * @start_addr: virtual address at which to start unmapping
496 * @end_addr: virtual address at which to end unmapping
497 * @nr_accounted: Place number of unmapped pages in vm-accountable vma's here
499 * Returns the number of vma's which were covered by the unmapping.
501 * Unmap all pages in the vma list. Called under page_table_lock.
503 * We aim to not hold page_table_lock for too long (for scheduling latency
504 * reasons). So zap pages in ZAP_BLOCK_SIZE bytecounts. This means we need to
505 * return the ending mmu_gather to the caller.
507 * Only addresses between `start' and `end' will be unmapped.
509 * The VMA list must be sorted in ascending virtual address order.
511 * unmap_vmas() assumes that the caller will flush the whole unmapped address
512 * range after unmap_vmas() returns. So the only responsibility here is to
513 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
514 * drops the lock and schedules.
516 int unmap_vmas(struct mmu_gather **tlbp, struct mm_struct *mm,
517 struct vm_area_struct *vma, unsigned long start_addr,
518 unsigned long end_addr, unsigned long *nr_accounted)
520 unsigned long zap_bytes = ZAP_BLOCK_SIZE;
521 unsigned long tlb_start; /* For tlb_finish_mmu */
522 int tlb_start_valid = 0;
525 if (vma) { /* debug. killme. */
526 if (end_addr <= vma->vm_start)
527 printk("%s: end_addr(0x%08lx) <= vm_start(0x%08lx)\n",
528 __FUNCTION__, end_addr, vma->vm_start);
529 if (start_addr >= vma->vm_end)
530 printk("%s: start_addr(0x%08lx) <= vm_end(0x%08lx)\n",
531 __FUNCTION__, start_addr, vma->vm_end);
534 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next) {
538 start = max(vma->vm_start, start_addr);
539 if (start >= vma->vm_end)
541 end = min(vma->vm_end, end_addr);
542 if (end <= vma->vm_start)
545 if (vma->vm_flags & VM_ACCOUNT)
546 *nr_accounted += (end - start) >> PAGE_SHIFT;
549 while (start != end) {
552 if (is_vm_hugetlb_page(vma))
555 block = min(zap_bytes, end - start);
557 if (!tlb_start_valid) {
562 unmap_page_range(*tlbp, vma, start, start + block);
565 if ((long)zap_bytes > 0)
567 if (need_resched()) {
568 tlb_finish_mmu(*tlbp, tlb_start, start);
569 cond_resched_lock(&mm->page_table_lock);
570 *tlbp = tlb_gather_mmu(mm, 0);
573 zap_bytes = ZAP_BLOCK_SIZE;
575 if (vma->vm_next && vma->vm_next->vm_start < vma->vm_end)
576 printk("%s: VMA list is not sorted correctly!\n",
583 * zap_page_range - remove user pages in a given range
584 * @vma: vm_area_struct holding the applicable pages
585 * @address: starting address of pages to zap
586 * @size: number of bytes to zap
588 void zap_page_range(struct vm_area_struct *vma,
589 unsigned long address, unsigned long size)
591 struct mm_struct *mm = vma->vm_mm;
592 struct mmu_gather *tlb;
593 unsigned long end = address + size;
594 unsigned long nr_accounted = 0;
598 if (is_vm_hugetlb_page(vma)) {
599 zap_hugepage_range(vma, address, size);
604 spin_lock(&mm->page_table_lock);
605 tlb = tlb_gather_mmu(mm, 0);
606 unmap_vmas(&tlb, mm, vma, address, end, &nr_accounted);
607 tlb_finish_mmu(tlb, address, end);
608 spin_unlock(&mm->page_table_lock);
612 * Do a quick page-table lookup for a single page.
613 * mm->page_table_lock must be held.
616 follow_page(struct mm_struct *mm, unsigned long address, int write)
622 struct vm_area_struct *vma;
624 vma = hugepage_vma(mm, address);
626 return follow_huge_addr(mm, vma, address, write);
628 pgd = pgd_offset(mm, address);
629 if (pgd_none(*pgd) || pgd_bad(*pgd))
632 pmd = pmd_offset(pgd, address);
636 return follow_huge_pmd(mm, address, pmd, write);
640 ptep = pte_offset_map(pmd, address);
646 if (pte_present(pte)) {
647 if (!write || (pte_write(pte) && pte_dirty(pte))) {
650 return pfn_to_page(pfn);
659 * Given a physical address, is there a useful struct page pointing to
660 * it? This may become more complex in the future if we start dealing
661 * with IO-aperture pages for direct-IO.
664 static inline struct page *get_page_map(struct page *page)
666 if (!pfn_valid(page_to_pfn(page)))
672 int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
673 unsigned long start, int len, int write, int force,
674 struct page **pages, struct vm_area_struct **vmas)
680 * Require read or write permissions.
681 * If 'force' is set, we only require the "MAY" flags.
683 flags = write ? (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
684 flags &= force ? (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
688 struct vm_area_struct * vma;
690 vma = find_extend_vma(mm, start);
692 #ifdef VSYSCALL_START
693 if (!vma && start >= VSYSCALL_START && start < VSYSCALL_END) {
694 static struct vm_area_struct fixmap_vma = {
695 /* Catch users - if there are any valid
696 ones, we can make this be "&init_mm" or
699 .vm_start = FIXADDR_START,
700 .vm_end = FIXADDR_TOP,
701 .vm_page_prot = PAGE_READONLY,
702 .vm_flags = VM_READ | VM_EXEC,
704 unsigned long pg = start & PAGE_MASK;
708 pgd = pgd_offset_k(pg);
710 return i ? : -EFAULT;
711 pmd = pmd_offset(pgd, pg);
713 return i ? : -EFAULT;
714 pte = pte_offset_kernel(pmd, pg);
715 if (!pte || !pte_present(*pte) || !pte_user(*pte) ||
716 !(write ? pte_write(*pte) : pte_read(*pte)))
717 return i ? : -EFAULT;
719 pages[i] = pte_page(*pte);
723 vmas[i] = &fixmap_vma;
731 if (!vma || (pages && (vma->vm_flags & VM_IO))
732 || !(flags & vma->vm_flags))
733 return i ? : -EFAULT;
735 if (is_vm_hugetlb_page(vma)) {
736 i = follow_hugetlb_page(mm, vma, pages, vmas,
740 spin_lock(&mm->page_table_lock);
743 while (!(map = follow_page(mm, start, write))) {
744 spin_unlock(&mm->page_table_lock);
745 switch (handle_mm_fault(mm,vma,start,write)) {
752 case VM_FAULT_SIGBUS:
753 return i ? i : -EFAULT;
755 return i ? i : -ENOMEM;
759 spin_lock(&mm->page_table_lock);
762 pages[i] = get_page_map(map);
764 spin_unlock(&mm->page_table_lock);
766 page_cache_release(pages[i]);
770 flush_dcache_page(pages[i]);
771 if (!PageReserved(pages[i]))
772 page_cache_get(pages[i]);
779 } while(len && start < vma->vm_end);
780 spin_unlock(&mm->page_table_lock);
786 static void zeromap_pte_range(pte_t * pte, unsigned long address,
787 unsigned long size, pgprot_t prot)
791 address &= ~PMD_MASK;
792 end = address + size;
796 pte_t zero_pte = pte_wrprotect(mk_pte(ZERO_PAGE(address), prot));
797 BUG_ON(!pte_none(*pte));
798 set_pte(pte, zero_pte);
799 address += PAGE_SIZE;
801 } while (address && (address < end));
804 static inline int zeromap_pmd_range(struct mm_struct *mm, pmd_t * pmd, unsigned long address,
805 unsigned long size, pgprot_t prot)
809 address &= ~PGDIR_MASK;
810 end = address + size;
811 if (end > PGDIR_SIZE)
814 pte_t * pte = pte_alloc_map(mm, pmd, address);
817 zeromap_pte_range(pte, address, end - address, prot);
819 address = (address + PMD_SIZE) & PMD_MASK;
821 } while (address && (address < end));
825 int zeromap_page_range(struct vm_area_struct *vma, unsigned long address, unsigned long size, pgprot_t prot)
829 unsigned long beg = address;
830 unsigned long end = address + size;
831 struct mm_struct *mm = vma->vm_mm;
833 dir = pgd_offset(mm, address);
834 flush_cache_range(vma, beg, end);
838 spin_lock(&mm->page_table_lock);
840 pmd_t *pmd = pmd_alloc(mm, dir, address);
844 error = zeromap_pmd_range(mm, pmd, address, end - address, prot);
847 address = (address + PGDIR_SIZE) & PGDIR_MASK;
849 } while (address && (address < end));
850 flush_tlb_range(vma, beg, end);
851 spin_unlock(&mm->page_table_lock);
856 * maps a range of physical memory into the requested pages. the old
857 * mappings are removed. any references to nonexistent pages results
858 * in null mappings (currently treated as "copy-on-access")
860 static inline void remap_pte_range(pte_t * pte, unsigned long address, unsigned long size,
861 unsigned long phys_addr, pgprot_t prot)
866 address &= ~PMD_MASK;
867 end = address + size;
870 pfn = phys_addr >> PAGE_SHIFT;
872 BUG_ON(!pte_none(*pte));
873 if (!pfn_valid(pfn) || PageReserved(pfn_to_page(pfn)))
874 set_pte(pte, pfn_pte(pfn, prot));
875 address += PAGE_SIZE;
878 } while (address && (address < end));
881 static inline int remap_pmd_range(struct mm_struct *mm, pmd_t * pmd, unsigned long address, unsigned long size,
882 unsigned long phys_addr, pgprot_t prot)
884 unsigned long base, end;
886 base = address & PGDIR_MASK;
887 address &= ~PGDIR_MASK;
888 end = address + size;
889 if (end > PGDIR_SIZE)
891 phys_addr -= address;
893 pte_t * pte = pte_alloc_map(mm, pmd, base + address);
896 remap_pte_range(pte, base + address, end - address, address + phys_addr, prot);
898 address = (address + PMD_SIZE) & PMD_MASK;
900 } while (address && (address < end));
904 /* Note: this is only safe if the mm semaphore is held when called. */
905 int remap_page_range(struct vm_area_struct *vma, unsigned long from, unsigned long phys_addr, unsigned long size, pgprot_t prot)
909 unsigned long beg = from;
910 unsigned long end = from + size;
911 struct mm_struct *mm = vma->vm_mm;
914 dir = pgd_offset(mm, from);
915 flush_cache_range(vma, beg, end);
919 spin_lock(&mm->page_table_lock);
921 pmd_t *pmd = pmd_alloc(mm, dir, from);
925 error = remap_pmd_range(mm, pmd, from, end - from, phys_addr + from, prot);
928 from = (from + PGDIR_SIZE) & PGDIR_MASK;
930 } while (from && (from < end));
931 flush_tlb_range(vma, beg, end);
932 spin_unlock(&mm->page_table_lock);
937 * Establish a new mapping:
938 * - flush the old one
939 * - update the page tables
940 * - inform the TLB about the new one
942 * We hold the mm semaphore for reading and vma->vm_mm->page_table_lock
944 static inline void establish_pte(struct vm_area_struct * vma, unsigned long address, pte_t *page_table, pte_t entry)
946 set_pte(page_table, entry);
947 flush_tlb_page(vma, address);
948 update_mmu_cache(vma, address, entry);
952 * We hold the mm semaphore for reading and vma->vm_mm->page_table_lock
954 static inline void break_cow(struct vm_area_struct * vma, struct page * new_page, unsigned long address,
957 invalidate_vcache(address, vma->vm_mm, new_page);
958 flush_cache_page(vma, address);
959 establish_pte(vma, address, page_table, pte_mkwrite(pte_mkdirty(mk_pte(new_page, vma->vm_page_prot))));
963 * This routine handles present pages, when users try to write
964 * to a shared page. It is done by copying the page to a new address
965 * and decrementing the shared-page counter for the old page.
967 * Goto-purists beware: the only reason for goto's here is that it results
968 * in better assembly code.. The "default" path will see no jumps at all.
970 * Note that this routine assumes that the protection checks have been
971 * done by the caller (the low-level page fault routine in most cases).
972 * Thus we can safely just mark it writable once we've done any necessary
975 * We also mark the page dirty at this point even though the page will
976 * change only once the write actually happens. This avoids a few races,
977 * and potentially makes it more efficient.
979 * We hold the mm semaphore and the page_table_lock on entry and exit
980 * with the page_table_lock released.
982 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct * vma,
983 unsigned long address, pte_t *page_table, pmd_t *pmd, pte_t pte)
985 struct page *old_page, *new_page;
986 unsigned long pfn = pte_pfn(pte);
987 struct pte_chain *pte_chain = NULL;
990 if (unlikely(!pfn_valid(pfn))) {
992 * This should really halt the system so it can be debugged or
993 * at least the kernel stops what it's doing before it corrupts
994 * data, but for the moment just pretend this is OOM.
996 pte_unmap(page_table);
997 printk(KERN_ERR "do_wp_page: bogus page at address %08lx\n",
1001 old_page = pfn_to_page(pfn);
1003 if (!TestSetPageLocked(old_page)) {
1004 int reuse = can_share_swap_page(old_page);
1005 unlock_page(old_page);
1007 flush_cache_page(vma, address);
1008 establish_pte(vma, address, page_table,
1009 pte_mkyoung(pte_mkdirty(pte_mkwrite(pte))));
1010 pte_unmap(page_table);
1011 ret = VM_FAULT_MINOR;
1015 pte_unmap(page_table);
1018 * Ok, we need to copy. Oh, well..
1020 page_cache_get(old_page);
1021 spin_unlock(&mm->page_table_lock);
1023 pte_chain = pte_chain_alloc(GFP_KERNEL);
1026 new_page = alloc_page(GFP_HIGHUSER);
1029 copy_cow_page(old_page,new_page,address);
1032 * Re-check the pte - we dropped the lock
1034 spin_lock(&mm->page_table_lock);
1035 page_table = pte_offset_map(pmd, address);
1036 if (pte_same(*page_table, pte)) {
1037 if (PageReserved(old_page))
1039 page_remove_rmap(old_page, page_table);
1040 break_cow(vma, new_page, address, page_table);
1041 pte_chain = page_add_rmap(new_page, page_table, pte_chain);
1042 lru_cache_add_active(new_page);
1044 /* Free the old page.. */
1045 new_page = old_page;
1047 pte_unmap(page_table);
1048 page_cache_release(new_page);
1049 page_cache_release(old_page);
1050 ret = VM_FAULT_MINOR;
1054 page_cache_release(old_page);
1058 spin_unlock(&mm->page_table_lock);
1059 pte_chain_free(pte_chain);
1063 static void vmtruncate_list(struct list_head *head, unsigned long pgoff)
1065 unsigned long start, end, len, diff;
1066 struct vm_area_struct *vma;
1067 struct list_head *curr;
1069 list_for_each(curr, head) {
1070 vma = list_entry(curr, struct vm_area_struct, shared);
1071 start = vma->vm_start;
1075 /* mapping wholly truncated? */
1076 if (vma->vm_pgoff >= pgoff) {
1077 zap_page_range(vma, start, len);
1081 /* mapping wholly unaffected? */
1082 len = len >> PAGE_SHIFT;
1083 diff = pgoff - vma->vm_pgoff;
1087 /* Ok, partially affected.. */
1088 start += diff << PAGE_SHIFT;
1089 len = (len - diff) << PAGE_SHIFT;
1090 zap_page_range(vma, start, len);
1095 * Handle all mappings that got truncated by a "truncate()"
1098 * NOTE! We have to be ready to update the memory sharing
1099 * between the file and the memory map for a potential last
1100 * incomplete page. Ugly, but necessary.
1102 int vmtruncate(struct inode * inode, loff_t offset)
1104 unsigned long pgoff;
1105 struct address_space *mapping = inode->i_mapping;
1106 unsigned long limit;
1108 if (inode->i_size < offset)
1110 inode->i_size = offset;
1111 down(&mapping->i_shared_sem);
1112 if (list_empty(&mapping->i_mmap) && list_empty(&mapping->i_mmap_shared))
1115 pgoff = (offset + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1116 if (!list_empty(&mapping->i_mmap))
1117 vmtruncate_list(&mapping->i_mmap, pgoff);
1118 if (!list_empty(&mapping->i_mmap_shared))
1119 vmtruncate_list(&mapping->i_mmap_shared, pgoff);
1122 up(&mapping->i_shared_sem);
1123 truncate_inode_pages(mapping, offset);
1127 limit = current->rlim[RLIMIT_FSIZE].rlim_cur;
1128 if (limit != RLIM_INFINITY && offset > limit)
1130 if (offset > inode->i_sb->s_maxbytes)
1132 inode->i_size = offset;
1135 if (inode->i_op && inode->i_op->truncate)
1136 inode->i_op->truncate(inode);
1139 send_sig(SIGXFSZ, current, 0);
1145 * Primitive swap readahead code. We simply read an aligned block of
1146 * (1 << page_cluster) entries in the swap area. This method is chosen
1147 * because it doesn't cost us any seek time. We also make sure to queue
1148 * the 'original' request together with the readahead ones...
1150 void swapin_readahead(swp_entry_t entry)
1153 struct page *new_page;
1154 unsigned long offset;
1157 * Get the number of handles we should do readahead io to.
1159 num = valid_swaphandles(entry, &offset);
1160 for (i = 0; i < num; offset++, i++) {
1161 /* Ok, do the async read-ahead now */
1162 new_page = read_swap_cache_async(swp_entry(swp_type(entry),
1166 page_cache_release(new_page);
1168 lru_add_drain(); /* Push any new pages onto the LRU now */
1172 * We hold the mm semaphore and the page_table_lock on entry and
1173 * should release the pagetable lock on exit..
1175 static int do_swap_page(struct mm_struct * mm,
1176 struct vm_area_struct * vma, unsigned long address,
1177 pte_t *page_table, pmd_t *pmd, pte_t orig_pte, int write_access)
1180 swp_entry_t entry = pte_to_swp_entry(orig_pte);
1182 int ret = VM_FAULT_MINOR;
1183 struct pte_chain *pte_chain = NULL;
1185 pte_unmap(page_table);
1186 spin_unlock(&mm->page_table_lock);
1187 page = lookup_swap_cache(entry);
1189 swapin_readahead(entry);
1190 page = read_swap_cache_async(entry);
1193 * Back out if somebody else faulted in this pte while
1194 * we released the page table lock.
1196 spin_lock(&mm->page_table_lock);
1197 page_table = pte_offset_map(pmd, address);
1198 if (pte_same(*page_table, orig_pte))
1201 ret = VM_FAULT_MINOR;
1202 pte_unmap(page_table);
1203 spin_unlock(&mm->page_table_lock);
1207 /* Had to read the page from swap area: Major fault */
1208 ret = VM_FAULT_MAJOR;
1209 inc_page_state(pgmajfault);
1212 mark_page_accessed(page);
1213 pte_chain = pte_chain_alloc(GFP_KERNEL);
1221 * Back out if somebody else faulted in this pte while we
1222 * released the page table lock.
1224 spin_lock(&mm->page_table_lock);
1225 page_table = pte_offset_map(pmd, address);
1226 if (!pte_same(*page_table, orig_pte)) {
1227 pte_unmap(page_table);
1228 spin_unlock(&mm->page_table_lock);
1230 page_cache_release(page);
1231 ret = VM_FAULT_MINOR;
1235 /* The page isn't present yet, go ahead with the fault. */
1239 remove_exclusive_swap_page(page);
1242 pte = mk_pte(page, vma->vm_page_prot);
1243 if (write_access && can_share_swap_page(page))
1244 pte = pte_mkdirty(pte_mkwrite(pte));
1247 flush_icache_page(vma, page);
1248 set_pte(page_table, pte);
1249 pte_chain = page_add_rmap(page, page_table, pte_chain);
1251 /* No need to invalidate - it was non-present before */
1252 update_mmu_cache(vma, address, pte);
1253 pte_unmap(page_table);
1254 spin_unlock(&mm->page_table_lock);
1256 pte_chain_free(pte_chain);
1261 * We are called with the MM semaphore and page_table_lock
1262 * spinlock held to protect against concurrent faults in
1263 * multithreaded programs.
1266 do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
1267 pte_t *page_table, pmd_t *pmd, int write_access,
1271 struct page * page = ZERO_PAGE(addr);
1272 struct pte_chain *pte_chain;
1275 pte_chain = pte_chain_alloc(GFP_ATOMIC);
1277 pte_unmap(page_table);
1278 spin_unlock(&mm->page_table_lock);
1279 pte_chain = pte_chain_alloc(GFP_KERNEL);
1282 spin_lock(&mm->page_table_lock);
1283 page_table = pte_offset_map(pmd, addr);
1286 /* Read-only mapping of ZERO_PAGE. */
1287 entry = pte_wrprotect(mk_pte(ZERO_PAGE(addr), vma->vm_page_prot));
1289 /* ..except if it's a write access */
1291 /* Allocate our own private page. */
1292 pte_unmap(page_table);
1293 spin_unlock(&mm->page_table_lock);
1295 page = alloc_page(GFP_HIGHUSER);
1298 clear_user_highpage(page, addr);
1300 spin_lock(&mm->page_table_lock);
1301 page_table = pte_offset_map(pmd, addr);
1303 if (!pte_none(*page_table)) {
1304 pte_unmap(page_table);
1305 page_cache_release(page);
1306 spin_unlock(&mm->page_table_lock);
1307 ret = VM_FAULT_MINOR;
1311 entry = pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
1312 lru_cache_add_active(page);
1313 mark_page_accessed(page);
1316 set_pte(page_table, entry);
1317 /* ignores ZERO_PAGE */
1318 pte_chain = page_add_rmap(page, page_table, pte_chain);
1319 pte_unmap(page_table);
1321 /* No need to invalidate - it was non-present before */
1322 update_mmu_cache(vma, addr, entry);
1323 spin_unlock(&mm->page_table_lock);
1324 ret = VM_FAULT_MINOR;
1330 pte_chain_free(pte_chain);
1335 * do_no_page() tries to create a new page mapping. It aggressively
1336 * tries to share with existing pages, but makes a separate copy if
1337 * the "write_access" parameter is true in order to avoid the next
1340 * As this is called only for pages that do not currently exist, we
1341 * do not need to flush old virtual caches or the TLB.
1343 * This is called with the MM semaphore held and the page table
1344 * spinlock held. Exit with the spinlock released.
1347 do_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
1348 unsigned long address, int write_access, pte_t *page_table, pmd_t *pmd)
1350 struct page * new_page;
1352 struct pte_chain *pte_chain;
1355 if (!vma->vm_ops || !vma->vm_ops->nopage)
1356 return do_anonymous_page(mm, vma, page_table,
1357 pmd, write_access, address);
1358 pte_unmap(page_table);
1359 spin_unlock(&mm->page_table_lock);
1361 new_page = vma->vm_ops->nopage(vma, address & PAGE_MASK, 0);
1363 /* no page was available -- either SIGBUS or OOM */
1364 if (new_page == NOPAGE_SIGBUS)
1365 return VM_FAULT_SIGBUS;
1366 if (new_page == NOPAGE_OOM)
1367 return VM_FAULT_OOM;
1369 pte_chain = pte_chain_alloc(GFP_KERNEL);
1374 * Should we do an early C-O-W break?
1376 if (write_access && !(vma->vm_flags & VM_SHARED)) {
1377 struct page * page = alloc_page(GFP_HIGHUSER);
1379 page_cache_release(new_page);
1382 copy_user_highpage(page, new_page, address);
1383 page_cache_release(new_page);
1384 lru_cache_add_active(page);
1388 spin_lock(&mm->page_table_lock);
1389 page_table = pte_offset_map(pmd, address);
1392 * This silly early PAGE_DIRTY setting removes a race
1393 * due to the bad i386 page protection. But it's valid
1394 * for other architectures too.
1396 * Note that if write_access is true, we either now have
1397 * an exclusive copy of the page, or this is a shared mapping,
1398 * so we can make it writable and dirty to avoid having to
1399 * handle that later.
1401 /* Only go through if we didn't race with anybody else... */
1402 if (pte_none(*page_table)) {
1404 flush_icache_page(vma, new_page);
1405 entry = mk_pte(new_page, vma->vm_page_prot);
1407 entry = pte_mkwrite(pte_mkdirty(entry));
1408 set_pte(page_table, entry);
1409 pte_chain = page_add_rmap(new_page, page_table, pte_chain);
1410 pte_unmap(page_table);
1412 /* One of our sibling threads was faster, back out. */
1413 pte_unmap(page_table);
1414 page_cache_release(new_page);
1415 spin_unlock(&mm->page_table_lock);
1416 ret = VM_FAULT_MINOR;
1420 /* no need to invalidate: a not-present page shouldn't be cached */
1421 update_mmu_cache(vma, address, entry);
1422 spin_unlock(&mm->page_table_lock);
1423 ret = VM_FAULT_MAJOR;
1428 pte_chain_free(pte_chain);
1433 * Fault of a previously existing named mapping. Repopulate the pte
1434 * from the encoded file_pte if possible. This enables swappable
1437 static int do_file_page(struct mm_struct * mm, struct vm_area_struct * vma,
1438 unsigned long address, int write_access, pte_t *pte, pmd_t *pmd)
1440 unsigned long pgoff;
1443 BUG_ON(!vma->vm_ops || !vma->vm_ops->nopage);
1445 * Fall back to the linear mapping if the fs does not support
1448 if (!vma->vm_ops || !vma->vm_ops->populate ||
1449 (write_access && !(vma->vm_flags & VM_SHARED))) {
1451 return do_no_page(mm, vma, address, write_access, pte, pmd);
1454 pgoff = pte_to_pgoff(*pte);
1457 spin_unlock(&mm->page_table_lock);
1459 err = vma->vm_ops->populate(vma, address & PAGE_MASK, PAGE_SIZE, vma->vm_page_prot, pgoff, 0);
1461 return VM_FAULT_OOM;
1463 return VM_FAULT_SIGBUS;
1464 return VM_FAULT_MAJOR;
1468 * These routines also need to handle stuff like marking pages dirty
1469 * and/or accessed for architectures that don't do it in hardware (most
1470 * RISC architectures). The early dirtying is also good on the i386.
1472 * There is also a hook called "update_mmu_cache()" that architectures
1473 * with external mmu caches can use to update those (ie the Sparc or
1474 * PowerPC hashed page tables that act as extended TLBs).
1476 * Note the "page_table_lock". It is to protect against kswapd removing
1477 * pages from under us. Note that kswapd only ever _removes_ pages, never
1478 * adds them. As such, once we have noticed that the page is not present,
1479 * we can drop the lock early.
1481 * The adding of pages is protected by the MM semaphore (which we hold),
1482 * so we don't need to worry about a page being suddenly been added into
1485 * We enter with the pagetable spinlock held, we are supposed to
1486 * release it when done.
1488 static inline int handle_pte_fault(struct mm_struct *mm,
1489 struct vm_area_struct * vma, unsigned long address,
1490 int write_access, pte_t *pte, pmd_t *pmd)
1495 if (!pte_present(entry)) {
1497 * If it truly wasn't present, we know that kswapd
1498 * and the PTE updates will not touch it later. So
1501 if (pte_none(entry))
1502 return do_no_page(mm, vma, address, write_access, pte, pmd);
1503 if (pte_file(entry))
1504 return do_file_page(mm, vma, address, write_access, pte, pmd);
1505 return do_swap_page(mm, vma, address, pte, pmd, entry, write_access);
1509 if (!pte_write(entry))
1510 return do_wp_page(mm, vma, address, pte, pmd, entry);
1512 entry = pte_mkdirty(entry);
1514 entry = pte_mkyoung(entry);
1515 establish_pte(vma, address, pte, entry);
1517 spin_unlock(&mm->page_table_lock);
1518 return VM_FAULT_MINOR;
1522 * By the time we get here, we already hold the mm semaphore
1524 int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct * vma,
1525 unsigned long address, int write_access)
1530 __set_current_state(TASK_RUNNING);
1531 pgd = pgd_offset(mm, address);
1533 inc_page_state(pgfault);
1535 if (is_vm_hugetlb_page(vma))
1536 return VM_FAULT_SIGBUS; /* mapping truncation does this. */
1539 * We need the page table lock to synchronize with kswapd
1540 * and the SMP-safe atomic PTE updates.
1542 spin_lock(&mm->page_table_lock);
1543 pmd = pmd_alloc(mm, pgd, address);
1546 pte_t * pte = pte_alloc_map(mm, pmd, address);
1548 return handle_pte_fault(mm, vma, address, write_access, pte, pmd);
1550 spin_unlock(&mm->page_table_lock);
1551 return VM_FAULT_OOM;
1555 * Allocate page middle directory.
1557 * We've already handled the fast-path in-line, and we own the
1560 * On a two-level page table, this ends up actually being entirely
1563 pmd_t *__pmd_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
1567 spin_unlock(&mm->page_table_lock);
1568 new = pmd_alloc_one(mm, address);
1569 spin_lock(&mm->page_table_lock);
1574 * Because we dropped the lock, we should re-check the
1575 * entry, as somebody else could have populated it..
1577 if (pgd_present(*pgd)) {
1581 pgd_populate(mm, pgd, new);
1583 return pmd_offset(pgd, address);
1586 int make_pages_present(unsigned long addr, unsigned long end)
1588 int ret, len, write;
1589 struct vm_area_struct * vma;
1591 vma = find_vma(current->mm, addr);
1592 write = (vma->vm_flags & VM_WRITE) != 0;
1595 if (end > vma->vm_end)
1597 len = (end+PAGE_SIZE-1)/PAGE_SIZE-addr/PAGE_SIZE;
1598 ret = get_user_pages(current, current->mm, addr,
1599 len, write, 0, NULL, NULL);
1600 return ret == len ? 0 : -1;
1604 * Map a vmalloc()-space virtual address to the physical page.
1606 struct page * vmalloc_to_page(void * vmalloc_addr)
1608 unsigned long addr = (unsigned long) vmalloc_addr;
1609 struct page *page = NULL;
1610 pgd_t *pgd = pgd_offset_k(addr);
1614 if (!pgd_none(*pgd)) {
1615 pmd = pmd_offset(pgd, addr);
1616 if (!pmd_none(*pmd)) {
1618 ptep = pte_offset_map(pmd, addr);
1620 if (pte_present(pte))
1621 page = pte_page(pte);