[PATCH] fix get_user_pages bug
authorNick Piggin <nickpiggin@yahoo.com.au>
Wed, 3 Aug 2005 10:24:01 +0000 (20:24 +1000)
committerLinus Torvalds <torvalds@g5.osdl.org>
Wed, 3 Aug 2005 16:12:05 +0000 (09:12 -0700)
Checking pte_dirty instead of pte_write in __follow_page is problematic
for s390, and for copy_one_pte which leaves dirty when clearing write.

So revert __follow_page to check pte_write as before, and make
do_wp_page pass back a special extra VM_FAULT_WRITE bit to say it has
done its full job: once get_user_pages receives this value, it no longer
requires pte_write in __follow_page.

But most callers of handle_mm_fault, in the various architectures, have
switch statements which do not expect this new case.  To avoid changing
them all in a hurry, make an inline wrapper function (using the old
name) that masks off the new bit, and use the extended interface with
double underscores.

Yes, we do have a call to do_wp_page from do_swap_page, but no need to
change that: in rare case it's needed, another do_wp_page will follow.

Signed-off-by: Hugh Dickins <hugh@veritas.com>
[ Cleanups by Nick Piggin ]
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
include/linux/mm.h
mm/memory.c

index 6eb7f48..82d7024 100644 (file)
@@ -625,10 +625,16 @@ static inline int page_mapped(struct page *page)
  * Used to decide whether a process gets delivered SIGBUS or
  * just gets major/minor fault counters bumped up.
  */
-#define VM_FAULT_OOM   (-1)
-#define VM_FAULT_SIGBUS        0
-#define VM_FAULT_MINOR 1
-#define VM_FAULT_MAJOR 2
+#define VM_FAULT_OOM   0x00
+#define VM_FAULT_SIGBUS        0x01
+#define VM_FAULT_MINOR 0x02
+#define VM_FAULT_MAJOR 0x03
+
+/* 
+ * Special case for get_user_pages.
+ * Must be in a distinct bit from the above VM_FAULT_ flags.
+ */
+#define VM_FAULT_WRITE 0x10
 
 #define offset_in_page(p)      ((unsigned long)(p) & ~PAGE_MASK)
 
@@ -704,7 +710,13 @@ extern pte_t *FASTCALL(pte_alloc_kernel(struct mm_struct *mm, pmd_t *pmd, unsign
 extern pte_t *FASTCALL(pte_alloc_map(struct mm_struct *mm, pmd_t *pmd, unsigned long address));
 extern int install_page(struct mm_struct *mm, struct vm_area_struct *vma, unsigned long addr, struct page *page, pgprot_t prot);
 extern int install_file_pte(struct mm_struct *mm, struct vm_area_struct *vma, unsigned long addr, unsigned long pgoff, pgprot_t prot);
-extern int handle_mm_fault(struct mm_struct *mm,struct vm_area_struct *vma, unsigned long address, int write_access);
+extern int __handle_mm_fault(struct mm_struct *mm,struct vm_area_struct *vma, unsigned long address, int write_access);
+
+static inline int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma, unsigned long address, int write_access)
+{
+       return __handle_mm_fault(mm, vma, address, write_access) & (~VM_FAULT_WRITE);
+}
+
 extern int make_pages_present(unsigned long addr, unsigned long end);
 extern int access_process_vm(struct task_struct *tsk, unsigned long addr, void *buf, int len, int write);
 void install_arg_page(struct vm_area_struct *, struct page *, unsigned long);
index 2405289..81d7117 100644 (file)
@@ -811,15 +811,18 @@ static struct page *__follow_page(struct mm_struct *mm, unsigned long address,
        pte = *ptep;
        pte_unmap(ptep);
        if (pte_present(pte)) {
-               if (write && !pte_dirty(pte))
+               if (write && !pte_write(pte))
                        goto out;
                if (read && !pte_read(pte))
                        goto out;
                pfn = pte_pfn(pte);
                if (pfn_valid(pfn)) {
                        page = pfn_to_page(pfn);
-                       if (accessed)
+                       if (accessed) {
+                               if (write && !pte_dirty(pte) &&!PageDirty(page))
+                                       set_page_dirty(page);
                                mark_page_accessed(page);
+                       }
                        return page;
                }
        }
@@ -941,10 +944,11 @@ int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
                }
                spin_lock(&mm->page_table_lock);
                do {
+                       int write_access = write;
                        struct page *page;
 
                        cond_resched_lock(&mm->page_table_lock);
-                       while (!(page = follow_page(mm, start, write))) {
+                       while (!(page = follow_page(mm, start, write_access))) {
                                /*
                                 * Shortcut for anonymous pages. We don't want
                                 * to force the creation of pages tables for
@@ -957,7 +961,16 @@ int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
                                        break;
                                }
                                spin_unlock(&mm->page_table_lock);
-                               switch (handle_mm_fault(mm,vma,start,write)) {
+                               switch (__handle_mm_fault(mm, vma, start,
+                                                       write_access)) {
+                               case VM_FAULT_WRITE:
+                                       /*
+                                        * do_wp_page has broken COW when
+                                        * necessary, even if maybe_mkwrite
+                                        * decided not to set pte_write
+                                        */
+                                       write_access = 0;
+                                       /* FALLTHRU */
                                case VM_FAULT_MINOR:
                                        tsk->min_flt++;
                                        break;
@@ -1220,6 +1233,7 @@ static int do_wp_page(struct mm_struct *mm, struct vm_area_struct * vma,
        struct page *old_page, *new_page;
        unsigned long pfn = pte_pfn(pte);
        pte_t entry;
+       int ret;
 
        if (unlikely(!pfn_valid(pfn))) {
                /*
@@ -1247,7 +1261,7 @@ static int do_wp_page(struct mm_struct *mm, struct vm_area_struct * vma,
                        lazy_mmu_prot_update(entry);
                        pte_unmap(page_table);
                        spin_unlock(&mm->page_table_lock);
-                       return VM_FAULT_MINOR;
+                       return VM_FAULT_MINOR|VM_FAULT_WRITE;
                }
        }
        pte_unmap(page_table);
@@ -1274,6 +1288,7 @@ static int do_wp_page(struct mm_struct *mm, struct vm_area_struct * vma,
        /*
         * Re-check the pte - we dropped the lock
         */
+       ret = VM_FAULT_MINOR;
        spin_lock(&mm->page_table_lock);
        page_table = pte_offset_map(pmd, address);
        if (likely(pte_same(*page_table, pte))) {
@@ -1290,12 +1305,13 @@ static int do_wp_page(struct mm_struct *mm, struct vm_area_struct * vma,
 
                /* Free the old page.. */
                new_page = old_page;
+               ret |= VM_FAULT_WRITE;
        }
        pte_unmap(page_table);
        page_cache_release(new_page);
        page_cache_release(old_page);
        spin_unlock(&mm->page_table_lock);
-       return VM_FAULT_MINOR;
+       return ret;
 
 no_new_page:
        page_cache_release(old_page);
@@ -1987,7 +2003,6 @@ static inline int handle_pte_fault(struct mm_struct *mm,
        if (write_access) {
                if (!pte_write(entry))
                        return do_wp_page(mm, vma, address, pte, pmd, entry);
-
                entry = pte_mkdirty(entry);
        }
        entry = pte_mkyoung(entry);
@@ -2002,7 +2017,7 @@ static inline int handle_pte_fault(struct mm_struct *mm,
 /*
  * By the time we get here, we already hold the mm semaphore
  */
-int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct * vma,
+int __handle_mm_fault(struct mm_struct *mm, struct vm_area_struct * vma,
                unsigned long address, int write_access)
 {
        pgd_t *pgd;