mm: avoid livelock on !__GFP_FS allocations
[linux-2.6.git] / mm / fremap.c
index 69a37c2..b8e0e2d 100644 (file)
 #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>
 
+#include "internal.h"
+
 static void zap_pte(struct mm_struct *mm, struct vm_area_struct *vma,
                        unsigned long addr, pte_t *ptep)
 {
@@ -34,10 +37,10 @@ static void zap_pte(struct mm_struct *mm, struct vm_area_struct *vma,
                if (page) {
                        if (pte_dirty(pte))
                                set_page_dirty(page);
-                       page_remove_rmap(page, vma);
+                       page_remove_rmap(page);
                        page_cache_release(page);
                        update_hiwater_rss(mm);
-                       dec_mm_counter(mm, file_rss);
+                       dec_mm_counter(mm, MM_FILEPAGES);
                }
        } else {
                if (!pte_file(pte))
@@ -113,16 +116,15 @@ static int populate_range(struct mm_struct *mm, struct vm_area_struct *vma,
  * 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 (but must be zero),
+ * 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;
@@ -139,6 +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))
@@ -165,7 +171,7 @@ asmlinkage long sys_remap_file_pages(unsigned long start, unsigned long size,
        if (!(vma->vm_flags & VM_CAN_NONLINEAR))
                goto out;
 
-       if (end <= start || start < vma->vm_start || end > vma->vm_end)
+       if (start < vma->vm_start || start + size > vma->vm_end)
                goto out;
 
        /* Must set VM_NONLINEAR before any pages are populated. */
@@ -195,7 +201,7 @@ asmlinkage long sys_remap_file_pages(unsigned long start, unsigned long size,
                        flags &= MAP_NONBLOCK;
                        get_file(file);
                        addr = mmap_region(file, start, size,
-                                       flags, vma->vm_flags, pgoff, 1);
+                                       flags, vma->vm_flags, pgoff);
                        fput(file);
                        if (IS_ERR_VALUE(addr)) {
                                err = addr;
@@ -205,22 +211,40 @@ asmlinkage long sys_remap_file_pages(unsigned long start, unsigned long size,
                        }
                        goto out;
                }
-               spin_lock(&mapping->i_mmap_lock);
+               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);
-               spin_unlock(&mapping->i_mmap_lock);
+               mutex_unlock(&mapping->i_mmap_mutex);
        }
 
+       if (vma->vm_flags & VM_LOCKED) {
+               /*
+                * 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 (unlikely(has_write_lock)) {
-                       downgrade_write(&mm->mmap_sem);
-                       has_write_lock = 0;
+               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);
                }
-               make_pages_present(start, start+size);
        }
 
        /*
@@ -237,4 +261,3 @@ out:
 
        return err;
 }
-