mm: avoid livelock on !__GFP_FS allocations
[linux-2.6.git] / mm / fremap.c
index 224cc15..b8e0e2d 100644 (file)
@@ -5,7 +5,7 @@
  *
  * started by Ingo Molnar, Copyright (C) 2002, 2003
  */
-
+#include <linux/backing-dev.h>
 #include <linux/mm.h>
 #include <linux/swap.h>
 #include <linux/file.h>
 #include <linux/rmap.h>
 #include <linux/module.h>
 #include <linux/syscalls.h>
+#include <linux/mmu_notifier.h>
 
 #include <asm/mmu_context.h>
 #include <asm/cacheflush.h>
 #include <asm/tlbflush.h>
 
-static inline void zap_pte(struct mm_struct *mm, struct vm_area_struct *vma,
+#include "internal.h"
+
+static void zap_pte(struct mm_struct *mm, struct vm_area_struct *vma,
                        unsigned long addr, pte_t *ptep)
 {
        pte_t pte = *ptep;
 
-       if (pte_none(pte))
-               return;
        if (pte_present(pte)) {
-               unsigned long pfn = pte_pfn(pte);
                struct page *page;
 
-               flush_cache_page(vma, addr, pfn);
+               flush_cache_page(vma, addr, pte_pfn(pte));
                pte = ptep_clear_flush(vma, addr, ptep);
-               if (unlikely(!pfn_valid(pfn))) {
-                       print_bad_pte(vma, pte, addr);
-                       return;
+               page = vm_normal_page(vma, addr, pte);
+               if (page) {
+                       if (pte_dirty(pte))
+                               set_page_dirty(page);
+                       page_remove_rmap(page);
+                       page_cache_release(page);
+                       update_hiwater_rss(mm);
+                       dec_mm_counter(mm, MM_FILEPAGES);
                }
-               page = pfn_to_page(pfn);
-               if (pte_dirty(pte))
-                       set_page_dirty(page);
-               page_remove_rmap(page);
-               page_cache_release(page);
-               dec_mm_counter(mm, file_rss);
        } else {
                if (!pte_file(pte))
                        free_swap_and_cache(pte_to_swp_entry(pte));
-               pte_clear(mm, addr, ptep);
+               pte_clear_not_present_full(mm, addr, ptep, 0);
        }
 }
 
 /*
- * Install a file page to a given virtual memory address, release any
+ * Install a file pte to a given virtual memory address, release any
  * previously existing mapping.
  */
-int install_page(struct mm_struct *mm, struct vm_area_struct *vma,
-               unsigned long addr, struct page *page, pgprot_t prot)
+static int install_file_pte(struct mm_struct *mm, struct vm_area_struct *vma,
+               unsigned long addr, unsigned long pgoff, pgprot_t prot)
 {
-       struct inode *inode;
-       pgoff_t size;
        int err = -ENOMEM;
        pte_t *pte;
-       pmd_t *pmd;
-       pud_t *pud;
-       pgd_t *pgd;
-       pte_t pte_val;
-
-       BUG_ON(vma->vm_flags & VM_RESERVED);
-
-       pgd = pgd_offset(mm, addr);
-       spin_lock(&mm->page_table_lock);
-       
-       pud = pud_alloc(mm, pgd, addr);
-       if (!pud)
-               goto err_unlock;
+       spinlock_t *ptl;
 
-       pmd = pmd_alloc(mm, pud, addr);
-       if (!pmd)
-               goto err_unlock;
-
-       pte = pte_alloc_map(mm, pmd, addr);
+       pte = get_locked_pte(mm, addr, &ptl);
        if (!pte)
-               goto err_unlock;
+               goto out;
+
+       if (!pte_none(*pte))
+               zap_pte(mm, vma, addr, pte);
 
+       set_pte_at(mm, addr, pte, pgoff_to_pte(pgoff));
        /*
-        * This page may have been truncated. Tell the
-        * caller about it.
+        * We don't need to run update_mmu_cache() here because the "file pte"
+        * being installed by install_file_pte() is not a real pte - it's a
+        * non-present entry (like a swap entry), noting what file offset should
+        * be mapped there when there's a fault (in a non-linear vma where
+        * that's not obvious).
         */
-       err = -EINVAL;
-       inode = vma->vm_file->f_mapping->host;
-       size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
-       if (!page->mapping || page->index >= size)
-               goto err_unlock;
-       err = -ENOMEM;
-       if (page_mapcount(page) > INT_MAX/2)
-               goto err_unlock;
-
-       zap_pte(mm, vma, addr, pte);
-
-       inc_mm_counter(mm, file_rss);
-       flush_icache_page(vma, page);
-       set_pte_at(mm, addr, pte, mk_pte(page, prot));
-       page_add_file_rmap(page);
-       pte_val = *pte;
-       pte_unmap(pte);
-       update_mmu_cache(vma, addr, pte_val);
-
+       pte_unmap_unlock(pte, ptl);
        err = 0;
-err_unlock:
-       spin_unlock(&mm->page_table_lock);
+out:
        return err;
 }
-EXPORT_SYMBOL(install_page);
-
 
-/*
- * Install a file pte to a given virtual memory address, release any
- * previously existing mapping.
- */
-int install_file_pte(struct mm_struct *mm, struct vm_area_struct *vma,
-               unsigned long addr, unsigned long pgoff, pgprot_t prot)
+static int populate_range(struct mm_struct *mm, struct vm_area_struct *vma,
+                       unsigned long addr, unsigned long size, pgoff_t pgoff)
 {
-       int err = -ENOMEM;
-       pte_t *pte;
-       pmd_t *pmd;
-       pud_t *pud;
-       pgd_t *pgd;
-       pte_t pte_val;
+       int err;
 
-       BUG_ON(vma->vm_flags & VM_RESERVED);
+       do {
+               err = install_file_pte(mm, vma, addr, pgoff, vma->vm_page_prot);
+               if (err)
+                       return err;
 
-       pgd = pgd_offset(mm, addr);
-       spin_lock(&mm->page_table_lock);
-       
-       pud = pud_alloc(mm, pgd, addr);
-       if (!pud)
-               goto err_unlock;
+               size -= PAGE_SIZE;
+               addr += PAGE_SIZE;
+               pgoff++;
+       } while (size);
 
-       pmd = pmd_alloc(mm, pud, addr);
-       if (!pmd)
-               goto err_unlock;
+        return 0;
 
-       pte = pte_alloc_map(mm, pmd, addr);
-       if (!pte)
-               goto err_unlock;
-
-       zap_pte(mm, vma, addr, pte);
-
-       set_pte_at(mm, addr, pte, pgoff_to_pte(pgoff));
-       pte_val = *pte;
-       pte_unmap(pte);
-       update_mmu_cache(vma, addr, pte_val);
-       spin_unlock(&mm->page_table_lock);
-       return 0;
-
-err_unlock:
-       spin_unlock(&mm->page_table_lock);
-       return err;
 }
 
-
-/***
- * sys_remap_file_pages - remap arbitrary pages of a shared backing store
- *                        file within an existing vma.
+/**
+ * sys_remap_file_pages - remap arbitrary pages of an existing VM_SHARED vma
  * @start: start of the remapped virtual memory range
  * @size: size of the remapped virtual memory range
- * @prot: new protection bits of the range
- * @pgoff: to be mapped page of the backing store file
+ * @prot: new protection bits of the range (see NOTE)
+ * @pgoff: to-be-mapped page of the backing store file
  * @flags: 0 or MAP_NONBLOCKED - the later will cause no IO.
  *
- * this syscall works purely via pagetables, so it's the most efficient
+ * sys_remap_file_pages remaps arbitrary pages of an existing VM_SHARED vma
+ * (shared backing store file).
+ *
+ * This syscall works purely via pagetables, so it's the most efficient
  * way to map the same (large) file into a given virtual window. Unlike
  * mmap()/mremap() it does not create any new vmas. The new mappings are
  * also safe across swapout.
  *
- * NOTE: the 'prot' parameter right now is ignored, and the vma's default
- * protection is used. Arbitrary protections might be implemented in the
- * future.
+ * NOTE: the @prot parameter right now is ignored (but must be zero),
+ * and the vma's default protection is used. Arbitrary protections
+ * might be implemented in the future.
  */
-asmlinkage long sys_remap_file_pages(unsigned long start, unsigned long size,
-       unsigned long __prot, unsigned long pgoff, unsigned long flags)
+SYSCALL_DEFINE5(remap_file_pages, unsigned long, start, unsigned long, size,
+               unsigned long, prot, unsigned long, pgoff, unsigned long, flags)
 {
        struct mm_struct *mm = current->mm;
        struct address_space *mapping;
-       unsigned long end = start + size;
        struct vm_area_struct *vma;
        int err = -EINVAL;
        int has_write_lock = 0;
 
-       if (__prot)
+       if (prot)
                return err;
        /*
         * Sanitize the syscall parameters:
@@ -200,6 +141,10 @@ asmlinkage long sys_remap_file_pages(unsigned long start, unsigned long size,
        if (start + size <= start)
                return err;
 
+       /* Does pgoff wrap? */
+       if (pgoff + (size >> PAGE_SHIFT) < pgoff)
+               return err;
+
        /* Can we represent this offset inside this architecture's pte's? */
 #if PTE_FILE_MAX_BITS < BITS_PER_LONG
        if (pgoff + (size >> PAGE_SHIFT) >= (1UL << PTE_FILE_MAX_BITS))
@@ -215,45 +160,100 @@ asmlinkage long sys_remap_file_pages(unsigned long start, unsigned long size,
         * Make sure the vma is shared, that it supports prefaulting,
         * and that the remapped range is valid and fully within
         * the single existing vma.  vm_private_data is used as a
-        * swapout cursor in a VM_NONLINEAR vma (unless VM_RESERVED
-        * or VM_LOCKED, but VM_LOCKED could be revoked later on).
+        * swapout cursor in a VM_NONLINEAR vma.
         */
-       if (vma && (vma->vm_flags & VM_SHARED) &&
-               (!vma->vm_private_data ||
-                       (vma->vm_flags & (VM_NONLINEAR|VM_RESERVED))) &&
-               vma->vm_ops && vma->vm_ops->populate &&
-                       end > start && start >= vma->vm_start &&
-                               end <= vma->vm_end) {
-
-               /* Must set VM_NONLINEAR before any pages are populated. */
-               if (pgoff != linear_page_index(vma, start) &&
-                   !(vma->vm_flags & VM_NONLINEAR)) {
-                       if (!has_write_lock) {
-                               up_read(&mm->mmap_sem);
-                               down_write(&mm->mmap_sem);
-                               has_write_lock = 1;
-                               goto retry;
-                       }
-                       mapping = vma->vm_file->f_mapping;
-                       spin_lock(&mapping->i_mmap_lock);
-                       flush_dcache_mmap_lock(mapping);
-                       vma->vm_flags |= VM_NONLINEAR;
-                       vma_prio_tree_remove(vma, &mapping->i_mmap);
-                       vma_nonlinear_insert(vma, &mapping->i_mmap_nonlinear);
-                       flush_dcache_mmap_unlock(mapping);
-                       spin_unlock(&mapping->i_mmap_lock);
+       if (!vma || !(vma->vm_flags & VM_SHARED))
+               goto out;
+
+       if (vma->vm_private_data && !(vma->vm_flags & VM_NONLINEAR))
+               goto out;
+
+       if (!(vma->vm_flags & VM_CAN_NONLINEAR))
+               goto out;
+
+       if (start < vma->vm_start || start + size > vma->vm_end)
+               goto out;
+
+       /* Must set VM_NONLINEAR before any pages are populated. */
+       if (!(vma->vm_flags & VM_NONLINEAR)) {
+               /* Don't need a nonlinear mapping, exit success */
+               if (pgoff == linear_page_index(vma, start)) {
+                       err = 0;
+                       goto out;
                }
 
-               err = vma->vm_ops->populate(vma, start, size,
-                                           vma->vm_page_prot,
-                                           pgoff, flags & MAP_NONBLOCK);
+               if (!has_write_lock) {
+                       up_read(&mm->mmap_sem);
+                       down_write(&mm->mmap_sem);
+                       has_write_lock = 1;
+                       goto retry;
+               }
+               mapping = vma->vm_file->f_mapping;
+               /*
+                * page_mkclean doesn't work on nonlinear vmas, so if
+                * dirty pages need to be accounted, emulate with linear
+                * vmas.
+                */
+               if (mapping_cap_account_dirty(mapping)) {
+                       unsigned long addr;
+                       struct file *file = vma->vm_file;
+
+                       flags &= MAP_NONBLOCK;
+                       get_file(file);
+                       addr = mmap_region(file, start, size,
+                                       flags, vma->vm_flags, pgoff);
+                       fput(file);
+                       if (IS_ERR_VALUE(addr)) {
+                               err = addr;
+                       } else {
+                               BUG_ON(addr != start);
+                               err = 0;
+                       }
+                       goto out;
+               }
+               mutex_lock(&mapping->i_mmap_mutex);
+               flush_dcache_mmap_lock(mapping);
+               vma->vm_flags |= VM_NONLINEAR;
+               vma_prio_tree_remove(vma, &mapping->i_mmap);
+               vma_nonlinear_insert(vma, &mapping->i_mmap_nonlinear);
+               flush_dcache_mmap_unlock(mapping);
+               mutex_unlock(&mapping->i_mmap_mutex);
+       }
 
+       if (vma->vm_flags & VM_LOCKED) {
                /*
-                * We can't clear VM_NONLINEAR because we'd have to do
-                * it after ->populate completes, and that would prevent
-                * downgrading the lock.  (Locks can't be upgraded).
+                * drop PG_Mlocked flag for over-mapped range
                 */
+               vm_flags_t saved_flags = vma->vm_flags;
+               munlock_vma_pages_range(vma, start, start + size);
+               vma->vm_flags = saved_flags;
+       }
+
+       mmu_notifier_invalidate_range_start(mm, start, start + size);
+       err = populate_range(mm, vma, start, size, pgoff);
+       mmu_notifier_invalidate_range_end(mm, start, start + size);
+       if (!err && !(flags & MAP_NONBLOCK)) {
+               if (vma->vm_flags & VM_LOCKED) {
+                       /*
+                        * might be mapping previously unmapped range of file
+                        */
+                       mlock_vma_pages_range(vma, start, start + size);
+               } else {
+                       if (unlikely(has_write_lock)) {
+                               downgrade_write(&mm->mmap_sem);
+                               has_write_lock = 0;
+                       }
+                       make_pages_present(start, start+size);
+               }
        }
+
+       /*
+        * We can't clear VM_NONLINEAR because we'd have to do
+        * it after ->populate completes, and that would prevent
+        * downgrading the lock.  (Locks can't be upgraded).
+        */
+
+out:
        if (likely(!has_write_lock))
                up_read(&mm->mmap_sem);
        else
@@ -261,4 +261,3 @@ asmlinkage long sys_remap_file_pages(unsigned long start, unsigned long size,
 
        return err;
 }
-