Fix read/truncate race
NeilBrown [Tue, 17 Jul 2007 11:03:04 +0000 (04:03 -0700)]
do_generic_mapping_read currently samples the i_size at the start and doesn't
do so again unless it needs to call ->readpage to load a page.  After
->readpage it has to re-sample i_size as a truncate may have caused that page
to be filled with zeros, and the read() call should not see these.

However there are other activities that might cause ->readpage to be called on
a page between the time that do_generic_mapping_read samples i_size and when
it finds that it has an uptodate page.  These include at least read-ahead and
possibly another thread performing a read.

So do_generic_mapping_read must sample i_size *after* it has an uptodate page.
 Thus the current sampling at the start and after a read can be replaced with
a sampling before the copy-out.

The same change applied to __generic_file_splice_read.

Note that this fixes any race with truncate_complete_page, but does not fix a
possible race with truncate_partial_page.  If a partial truncate happens after
do_generic_mapping_read samples i_size and before the copy_out, the nuls that
truncate_partial_page place in the page could be copied out incorrectly.

I think the best fix for that is to *not* zero out parts of the page in
truncate_partial_page, but rather to zero out the tail of a page when
increasing i_size.

Signed-off-by: Neil Brown <neilb@suse.de>
Cc: Jens Axboe <jens.axboe@oracle.com>
Acked-by: Nick Piggin <npiggin@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm/filemap.c

index 100b99c..5d5449f 100644 (file)
@@ -867,13 +867,11 @@ void do_generic_mapping_read(struct address_space *mapping,
 {
        struct inode *inode = mapping->host;
        unsigned long index;
-       unsigned long end_index;
        unsigned long offset;
        unsigned long last_index;
        unsigned long next_index;
        unsigned long prev_index;
        unsigned int prev_offset;
-       loff_t isize;
        struct page *cached_page;
        int error;
        struct file_ra_state ra = *_ra;
@@ -886,27 +884,12 @@ void do_generic_mapping_read(struct address_space *mapping,
        last_index = (*ppos + desc->count + PAGE_CACHE_SIZE-1) >> PAGE_CACHE_SHIFT;
        offset = *ppos & ~PAGE_CACHE_MASK;
 
-       isize = i_size_read(inode);
-       if (!isize)
-               goto out;
-
-       end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
        for (;;) {
                struct page *page;
+               unsigned long end_index;
+               loff_t isize;
                unsigned long nr, ret;
 
-               /* nr is the maximum number of bytes to copy from this page */
-               nr = PAGE_CACHE_SIZE;
-               if (index >= end_index) {
-                       if (index > end_index)
-                               goto out;
-                       nr = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
-                       if (nr <= offset) {
-                               goto out;
-                       }
-               }
-               nr = nr - offset;
-
                cond_resched();
                if (index == next_index)
                        next_index = page_cache_readahead(mapping, &ra, filp,
@@ -921,6 +904,32 @@ find_page:
                if (!PageUptodate(page))
                        goto page_not_up_to_date;
 page_ok:
+               /*
+                * i_size must be checked after we know the page is Uptodate.
+                *
+                * Checking i_size after the check allows us to calculate
+                * the correct value for "nr", which means the zero-filled
+                * part of the page is not copied back to userspace (unless
+                * another truncate extends the file - this is desired though).
+                */
+
+               isize = i_size_read(inode);
+               end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
+               if (unlikely(!isize || index > end_index)) {
+                       page_cache_release(page);
+                       goto out;
+               }
+
+               /* nr is the maximum number of bytes to copy from this page */
+               nr = PAGE_CACHE_SIZE;
+               if (index == end_index) {
+                       nr = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
+                       if (nr <= offset) {
+                               page_cache_release(page);
+                               goto out;
+                       }
+               }
+               nr = nr - offset;
 
                /* If users can be writing to this page using arbitrary
                 * virtual addresses, take care about potential aliasing
@@ -1007,31 +1016,6 @@ readpage:
                        unlock_page(page);
                }
 
-               /*
-                * i_size must be checked after we have done ->readpage.
-                *
-                * Checking i_size after the readpage allows us to calculate
-                * the correct value for "nr", which means the zero-filled
-                * part of the page is not copied back to userspace (unless
-                * another truncate extends the file - this is desired though).
-                */
-               isize = i_size_read(inode);
-               end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
-               if (unlikely(!isize || index > end_index)) {
-                       page_cache_release(page);
-                       goto out;
-               }
-
-               /* nr is the maximum number of bytes to copy from this page */
-               nr = PAGE_CACHE_SIZE;
-               if (index == end_index) {
-                       nr = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
-                       if (nr <= offset) {
-                               page_cache_release(page);
-                               goto out;
-                       }
-               }
-               nr = nr - offset;
                goto page_ok;
 
 readpage_error: