[CIFS] fix regression in cifs_write_begin/cifs_write_end

The conversion to write_begin/write_end interfaces had a bug where we
were passing a bad parameter to cifs_readpage_worker. Rather than
passing the page offset of the start of the write, we needed to pass the
offset of the beginning of the page. This was reliably showing up as
data corruption in the fsx-linux test from LTP.

It also became evident that this code was occasionally doing unnecessary
read calls. Optimize those away by using the PG_checked flag to indicate
that the unwritten part of the page has been initialized.

CC: Nick Piggin <npiggin@suse.de>
Acked-by: Dave Kleikamp <shaggy@us.ibm.com>
Signed-off-by: Jeff Layton <jlayton@redhat.com>
Signed-off-by: Steve French <sfrench@us.ibm.com>

diff --git a/fs/cifs/file.c b/fs/cifs/file.c
index b691b893a848a8d4d74c1e0ecce800d561b99b67..f0a81e631ae60bed3ec139833a2a1a63cef45642 100644 (file)
--- a/fs/cifs/file.c
+++ b/fs/cifs/file.c
@@ -1475,7 +1475,11 @@ static int cifs_write_end(struct file *file, struct address_space *mapping,
        cFYI(1, ("write_end for page %p from pos %lld with %d bytes",
                 page, pos, copied));
 
-       if (!PageUptodate(page) && copied == PAGE_CACHE_SIZE)
+       if (PageChecked(page)) {
+               if (copied == len)
+                       SetPageUptodate(page);
+               ClearPageChecked(page);
+       } else if (!PageUptodate(page) && copied == PAGE_CACHE_SIZE)
                SetPageUptodate(page);
 
        if (!PageUptodate(page)) {
@@ -2062,39 +2066,70 @@ static int cifs_write_begin(struct file *file, struct address_space *mapping,
 {
        pgoff_t index = pos >> PAGE_CACHE_SHIFT;
        loff_t offset = pos & (PAGE_CACHE_SIZE - 1);
+       loff_t page_start = pos & PAGE_MASK;
+       loff_t i_size;
+       struct page *page;
+       int rc = 0;
 
        cFYI(1, ("write_begin from %lld len %d", (long long)pos, len));
 
-       *pagep = __grab_cache_page(mapping, index);
-       if (!*pagep)
-               return -ENOMEM;
-
-       if (PageUptodate(*pagep))
-               return 0;
+       page = __grab_cache_page(mapping, index);
+       if (!page) {
+               rc = -ENOMEM;
+               goto out;
+       }
 
-       /* If we are writing a full page it will be up to date,
-          no need to read from the server */
-       if (len == PAGE_CACHE_SIZE && flags & AOP_FLAG_UNINTERRUPTIBLE)
-               return 0;
+       if (PageUptodate(page))
+               goto out;
 
-       if ((file->f_flags & O_ACCMODE) != O_WRONLY) {
-               int rc;
+       /*
+        * If we write a full page it will be up to date, no need to read from
+        * the server. If the write is short, we'll end up doing a sync write
+        * instead.
+        */
+       if (len == PAGE_CACHE_SIZE)
+               goto out;
 
-               /* might as well read a page, it is fast enough */
-               rc = cifs_readpage_worker(file, *pagep, &offset);
+       /*
+        * optimize away the read when we have an oplock, and we're not
+        * expecting to use any of the data we'd be reading in. That
+        * is, when the page lies beyond the EOF, or straddles the EOF
+        * and the write will cover all of the existing data.
+        */
+       if (CIFS_I(mapping->host)->clientCanCacheRead) {
+               i_size = i_size_read(mapping->host);
+               if (page_start >= i_size ||
+                   (offset == 0 && (pos + len) >= i_size)) {
+                       zero_user_segments(page, 0, offset,
+                                          offset + len,
+                                          PAGE_CACHE_SIZE);
+                       /*
+                        * PageChecked means that the parts of the page
+                        * to which we're not writing are considered up
+                        * to date. Once the data is copied to the
+                        * page, it can be set uptodate.
+                        */
+                       SetPageChecked(page);
+                       goto out;
+               }
+       }
 
-               /* we do not need to pass errors back
-                  e.g. if we do not have read access to the file
-                  because cifs_write_end will attempt synchronous writes
-                  -- shaggy */
+       if ((file->f_flags & O_ACCMODE) != O_WRONLY) {
+               /*
+                * might as well read a page, it is fast enough. If we get
+                * an error, we don't need to return it. cifs_write_end will
+                * do a sync write instead since PG_uptodate isn't set.
+                */
+               cifs_readpage_worker(file, page, &page_start);
        } else {
                /* we could try using another file handle if there is one -
                   but how would we lock it to prevent close of that handle
                   racing with this read? In any case
                   this will be written out by write_end so is fine */
        }
-
-       return 0;
+out:
+       *pagep = page;
+       return rc;
 }
 
 const struct address_space_operations cifs_addr_ops = {