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 FIXADDR_USER_START
694 start >= FIXADDR_USER_START && start < FIXADDR_USER_END) {
695 static struct vm_area_struct fixmap_vma = {
696 /* Catch users - if there are any valid
697 ones, we can make this be "&init_mm" or
700 .vm_start = FIXADDR_USER_START,
701 .vm_end = FIXADDR_USER_END,
702 .vm_page_prot = PAGE_READONLY,
703 .vm_flags = VM_READ | VM_EXEC,
705 unsigned long pg = start & PAGE_MASK;
709 if (write) /* user fixmap pages are read-only */
710 return i ? : -EFAULT;
711 pgd = pgd_offset_k(pg);
713 return i ? : -EFAULT;
714 pmd = pmd_offset(pgd, pg);
716 return i ? : -EFAULT;
717 pte = pte_offset_kernel(pmd, pg);
718 if (!pte || !pte_present(*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 pgoff = (offset + PAGE_SIZE - 1) >> PAGE_SHIFT;
1114 down(&mapping->i_shared_sem);
1115 if (unlikely(!list_empty(&mapping->i_mmap)))
1116 vmtruncate_list(&mapping->i_mmap, pgoff);
1117 if (unlikely(!list_empty(&mapping->i_mmap_shared)))
1118 vmtruncate_list(&mapping->i_mmap_shared, pgoff);
1119 up(&mapping->i_shared_sem);
1120 truncate_inode_pages(mapping, offset);
1124 limit = current->rlim[RLIMIT_FSIZE].rlim_cur;
1125 if (limit != RLIM_INFINITY && offset > limit)
1127 if (offset > inode->i_sb->s_maxbytes)
1129 inode->i_size = offset;
1132 if (inode->i_op && inode->i_op->truncate)
1133 inode->i_op->truncate(inode);
1136 send_sig(SIGXFSZ, current, 0);
1142 * Primitive swap readahead code. We simply read an aligned block of
1143 * (1 << page_cluster) entries in the swap area. This method is chosen
1144 * because it doesn't cost us any seek time. We also make sure to queue
1145 * the 'original' request together with the readahead ones...
1147 void swapin_readahead(swp_entry_t entry)
1150 struct page *new_page;
1151 unsigned long offset;
1154 * Get the number of handles we should do readahead io to.
1156 num = valid_swaphandles(entry, &offset);
1157 for (i = 0; i < num; offset++, i++) {
1158 /* Ok, do the async read-ahead now */
1159 new_page = read_swap_cache_async(swp_entry(swp_type(entry),
1163 page_cache_release(new_page);
1165 lru_add_drain(); /* Push any new pages onto the LRU now */
1169 * We hold the mm semaphore and the page_table_lock on entry and
1170 * should release the pagetable lock on exit..
1172 static int do_swap_page(struct mm_struct * mm,
1173 struct vm_area_struct * vma, unsigned long address,
1174 pte_t *page_table, pmd_t *pmd, pte_t orig_pte, int write_access)
1177 swp_entry_t entry = pte_to_swp_entry(orig_pte);
1179 int ret = VM_FAULT_MINOR;
1180 struct pte_chain *pte_chain = NULL;
1182 pte_unmap(page_table);
1183 spin_unlock(&mm->page_table_lock);
1184 page = lookup_swap_cache(entry);
1186 swapin_readahead(entry);
1187 page = read_swap_cache_async(entry);
1190 * Back out if somebody else faulted in this pte while
1191 * we released the page table lock.
1193 spin_lock(&mm->page_table_lock);
1194 page_table = pte_offset_map(pmd, address);
1195 if (pte_same(*page_table, orig_pte))
1198 ret = VM_FAULT_MINOR;
1199 pte_unmap(page_table);
1200 spin_unlock(&mm->page_table_lock);
1204 /* Had to read the page from swap area: Major fault */
1205 ret = VM_FAULT_MAJOR;
1206 inc_page_state(pgmajfault);
1209 mark_page_accessed(page);
1210 pte_chain = pte_chain_alloc(GFP_KERNEL);
1218 * Back out if somebody else faulted in this pte while we
1219 * released the page table lock.
1221 spin_lock(&mm->page_table_lock);
1222 page_table = pte_offset_map(pmd, address);
1223 if (!pte_same(*page_table, orig_pte)) {
1224 pte_unmap(page_table);
1225 spin_unlock(&mm->page_table_lock);
1227 page_cache_release(page);
1228 ret = VM_FAULT_MINOR;
1232 /* The page isn't present yet, go ahead with the fault. */
1236 remove_exclusive_swap_page(page);
1239 pte = mk_pte(page, vma->vm_page_prot);
1240 if (write_access && can_share_swap_page(page))
1241 pte = pte_mkdirty(pte_mkwrite(pte));
1244 flush_icache_page(vma, page);
1245 set_pte(page_table, pte);
1246 pte_chain = page_add_rmap(page, page_table, pte_chain);
1248 /* No need to invalidate - it was non-present before */
1249 update_mmu_cache(vma, address, pte);
1250 pte_unmap(page_table);
1251 spin_unlock(&mm->page_table_lock);
1253 pte_chain_free(pte_chain);
1258 * We are called with the MM semaphore and page_table_lock
1259 * spinlock held to protect against concurrent faults in
1260 * multithreaded programs.
1263 do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
1264 pte_t *page_table, pmd_t *pmd, int write_access,
1268 struct page * page = ZERO_PAGE(addr);
1269 struct pte_chain *pte_chain;
1272 pte_chain = pte_chain_alloc(GFP_ATOMIC);
1274 pte_unmap(page_table);
1275 spin_unlock(&mm->page_table_lock);
1276 pte_chain = pte_chain_alloc(GFP_KERNEL);
1279 spin_lock(&mm->page_table_lock);
1280 page_table = pte_offset_map(pmd, addr);
1283 /* Read-only mapping of ZERO_PAGE. */
1284 entry = pte_wrprotect(mk_pte(ZERO_PAGE(addr), vma->vm_page_prot));
1286 /* ..except if it's a write access */
1288 /* Allocate our own private page. */
1289 pte_unmap(page_table);
1290 spin_unlock(&mm->page_table_lock);
1292 page = alloc_page(GFP_HIGHUSER);
1295 clear_user_highpage(page, addr);
1297 spin_lock(&mm->page_table_lock);
1298 page_table = pte_offset_map(pmd, addr);
1300 if (!pte_none(*page_table)) {
1301 pte_unmap(page_table);
1302 page_cache_release(page);
1303 spin_unlock(&mm->page_table_lock);
1304 ret = VM_FAULT_MINOR;
1308 entry = pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
1309 lru_cache_add_active(page);
1310 mark_page_accessed(page);
1313 set_pte(page_table, entry);
1314 /* ignores ZERO_PAGE */
1315 pte_chain = page_add_rmap(page, page_table, pte_chain);
1316 pte_unmap(page_table);
1318 /* No need to invalidate - it was non-present before */
1319 update_mmu_cache(vma, addr, entry);
1320 spin_unlock(&mm->page_table_lock);
1321 ret = VM_FAULT_MINOR;
1327 pte_chain_free(pte_chain);
1332 * do_no_page() tries to create a new page mapping. It aggressively
1333 * tries to share with existing pages, but makes a separate copy if
1334 * the "write_access" parameter is true in order to avoid the next
1337 * As this is called only for pages that do not currently exist, we
1338 * do not need to flush old virtual caches or the TLB.
1340 * This is called with the MM semaphore held and the page table
1341 * spinlock held. Exit with the spinlock released.
1344 do_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
1345 unsigned long address, int write_access, pte_t *page_table, pmd_t *pmd)
1347 struct page * new_page;
1349 struct pte_chain *pte_chain;
1352 if (!vma->vm_ops || !vma->vm_ops->nopage)
1353 return do_anonymous_page(mm, vma, page_table,
1354 pmd, write_access, address);
1355 pte_unmap(page_table);
1356 spin_unlock(&mm->page_table_lock);
1358 new_page = vma->vm_ops->nopage(vma, address & PAGE_MASK, 0);
1360 /* no page was available -- either SIGBUS or OOM */
1361 if (new_page == NOPAGE_SIGBUS)
1362 return VM_FAULT_SIGBUS;
1363 if (new_page == NOPAGE_OOM)
1364 return VM_FAULT_OOM;
1366 pte_chain = pte_chain_alloc(GFP_KERNEL);
1371 * Should we do an early C-O-W break?
1373 if (write_access && !(vma->vm_flags & VM_SHARED)) {
1374 struct page * page = alloc_page(GFP_HIGHUSER);
1376 page_cache_release(new_page);
1379 copy_user_highpage(page, new_page, address);
1380 page_cache_release(new_page);
1381 lru_cache_add_active(page);
1385 spin_lock(&mm->page_table_lock);
1386 page_table = pte_offset_map(pmd, address);
1389 * This silly early PAGE_DIRTY setting removes a race
1390 * due to the bad i386 page protection. But it's valid
1391 * for other architectures too.
1393 * Note that if write_access is true, we either now have
1394 * an exclusive copy of the page, or this is a shared mapping,
1395 * so we can make it writable and dirty to avoid having to
1396 * handle that later.
1398 /* Only go through if we didn't race with anybody else... */
1399 if (pte_none(*page_table)) {
1401 flush_icache_page(vma, new_page);
1402 entry = mk_pte(new_page, vma->vm_page_prot);
1404 entry = pte_mkwrite(pte_mkdirty(entry));
1405 set_pte(page_table, entry);
1406 pte_chain = page_add_rmap(new_page, page_table, pte_chain);
1407 pte_unmap(page_table);
1409 /* One of our sibling threads was faster, back out. */
1410 pte_unmap(page_table);
1411 page_cache_release(new_page);
1412 spin_unlock(&mm->page_table_lock);
1413 ret = VM_FAULT_MINOR;
1417 /* no need to invalidate: a not-present page shouldn't be cached */
1418 update_mmu_cache(vma, address, entry);
1419 spin_unlock(&mm->page_table_lock);
1420 ret = VM_FAULT_MAJOR;
1425 pte_chain_free(pte_chain);
1430 * Fault of a previously existing named mapping. Repopulate the pte
1431 * from the encoded file_pte if possible. This enables swappable
1434 static int do_file_page(struct mm_struct * mm, struct vm_area_struct * vma,
1435 unsigned long address, int write_access, pte_t *pte, pmd_t *pmd)
1437 unsigned long pgoff;
1440 BUG_ON(!vma->vm_ops || !vma->vm_ops->nopage);
1442 * Fall back to the linear mapping if the fs does not support
1445 if (!vma->vm_ops || !vma->vm_ops->populate ||
1446 (write_access && !(vma->vm_flags & VM_SHARED))) {
1448 return do_no_page(mm, vma, address, write_access, pte, pmd);
1451 pgoff = pte_to_pgoff(*pte);
1454 spin_unlock(&mm->page_table_lock);
1456 err = vma->vm_ops->populate(vma, address & PAGE_MASK, PAGE_SIZE, vma->vm_page_prot, pgoff, 0);
1458 return VM_FAULT_OOM;
1460 return VM_FAULT_SIGBUS;
1461 return VM_FAULT_MAJOR;
1465 * These routines also need to handle stuff like marking pages dirty
1466 * and/or accessed for architectures that don't do it in hardware (most
1467 * RISC architectures). The early dirtying is also good on the i386.
1469 * There is also a hook called "update_mmu_cache()" that architectures
1470 * with external mmu caches can use to update those (ie the Sparc or
1471 * PowerPC hashed page tables that act as extended TLBs).
1473 * Note the "page_table_lock". It is to protect against kswapd removing
1474 * pages from under us. Note that kswapd only ever _removes_ pages, never
1475 * adds them. As such, once we have noticed that the page is not present,
1476 * we can drop the lock early.
1478 * The adding of pages is protected by the MM semaphore (which we hold),
1479 * so we don't need to worry about a page being suddenly been added into
1482 * We enter with the pagetable spinlock held, we are supposed to
1483 * release it when done.
1485 static inline int handle_pte_fault(struct mm_struct *mm,
1486 struct vm_area_struct * vma, unsigned long address,
1487 int write_access, pte_t *pte, pmd_t *pmd)
1492 if (!pte_present(entry)) {
1494 * If it truly wasn't present, we know that kswapd
1495 * and the PTE updates will not touch it later. So
1498 if (pte_none(entry))
1499 return do_no_page(mm, vma, address, write_access, pte, pmd);
1500 if (pte_file(entry))
1501 return do_file_page(mm, vma, address, write_access, pte, pmd);
1502 return do_swap_page(mm, vma, address, pte, pmd, entry, write_access);
1506 if (!pte_write(entry))
1507 return do_wp_page(mm, vma, address, pte, pmd, entry);
1509 entry = pte_mkdirty(entry);
1511 entry = pte_mkyoung(entry);
1512 establish_pte(vma, address, pte, entry);
1514 spin_unlock(&mm->page_table_lock);
1515 return VM_FAULT_MINOR;
1519 * By the time we get here, we already hold the mm semaphore
1521 int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct * vma,
1522 unsigned long address, int write_access)
1527 __set_current_state(TASK_RUNNING);
1528 pgd = pgd_offset(mm, address);
1530 inc_page_state(pgfault);
1532 if (is_vm_hugetlb_page(vma))
1533 return VM_FAULT_SIGBUS; /* mapping truncation does this. */
1536 * We need the page table lock to synchronize with kswapd
1537 * and the SMP-safe atomic PTE updates.
1539 spin_lock(&mm->page_table_lock);
1540 pmd = pmd_alloc(mm, pgd, address);
1543 pte_t * pte = pte_alloc_map(mm, pmd, address);
1545 return handle_pte_fault(mm, vma, address, write_access, pte, pmd);
1547 spin_unlock(&mm->page_table_lock);
1548 return VM_FAULT_OOM;
1552 * Allocate page middle directory.
1554 * We've already handled the fast-path in-line, and we own the
1557 * On a two-level page table, this ends up actually being entirely
1560 pmd_t *__pmd_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
1564 spin_unlock(&mm->page_table_lock);
1565 new = pmd_alloc_one(mm, address);
1566 spin_lock(&mm->page_table_lock);
1571 * Because we dropped the lock, we should re-check the
1572 * entry, as somebody else could have populated it..
1574 if (pgd_present(*pgd)) {
1578 pgd_populate(mm, pgd, new);
1580 return pmd_offset(pgd, address);
1583 int make_pages_present(unsigned long addr, unsigned long end)
1585 int ret, len, write;
1586 struct vm_area_struct * vma;
1588 vma = find_vma(current->mm, addr);
1589 write = (vma->vm_flags & VM_WRITE) != 0;
1592 if (end > vma->vm_end)
1594 len = (end+PAGE_SIZE-1)/PAGE_SIZE-addr/PAGE_SIZE;
1595 ret = get_user_pages(current, current->mm, addr,
1596 len, write, 0, NULL, NULL);
1597 return ret == len ? 0 : -1;
1601 * Map a vmalloc()-space virtual address to the physical page.
1603 struct page * vmalloc_to_page(void * vmalloc_addr)
1605 unsigned long addr = (unsigned long) vmalloc_addr;
1606 struct page *page = NULL;
1607 pgd_t *pgd = pgd_offset_k(addr);
1611 if (!pgd_none(*pgd)) {
1612 pmd = pmd_offset(pgd, addr);
1613 if (!pmd_none(*pmd)) {
1615 ptep = pte_offset_map(pmd, addr);
1617 if (pte_present(pte))
1618 page = pte_page(pte);