[CIFS] Reduce chance of list corruption in find_writable_file
authorSteve French <sfrench@us.ibm.com>
Tue, 2 Oct 2007 01:11:08 +0000 (01:11 +0000)
committerSteve French <sfrench@us.ibm.com>
Tue, 2 Oct 2007 01:11:08 +0000 (01:11 +0000)
When find_writable_file is racing with close and the session
to the server goes down, Shaggy noticed that there was a
chance that an open file in the list of files off the inode
could have been freed by close since cifs_reconnect can
block (the spinlock thus not held). This means that
we have to start over at the beginning of the list in some
cases.

There is a 2nd change that needs to be made later
(pointed out by Jeremy Allison and Shaggy) in order to
prevent cifs_close ever freeing the cifs per file info
when a write is pending.  Although we delay close from
freeing this memory for sufficiently long for all known
cases, ultimately on a very, very slow write
overlapping a close pending we need to allow close to return
(without freeing the cifs file info) and defer freeing the
memory to be the responsibility of the (sloooow) write
thread (presumably have to look at every place wrtPending
is decremented - and add a flag for deferred free for
after wrtPending goes to zero).

Acked-by: Shaggy <shaggy@us.ibm.com>
Acked-by: Shirish Pargaonkar <shirishp@us.ibm.com>
Signed-off-by: Steve French <sfrench@us.ibm.com>
fs/cifs/file.c

index 79254919386539d5a4883f44b96a1581dcebbec0..780c0e3e4703c2c1ce6de2e73ba3076b4d82a595 100644 (file)
@@ -1042,6 +1042,7 @@ struct cifsFileInfo *find_writable_file(struct cifsInodeInfo *cifs_inode)
        }
 
        read_lock(&GlobalSMBSeslock);
+refind_writable:
        list_for_each_entry(open_file, &cifs_inode->openFileList, flist) {
                if (open_file->closePend)
                        continue;
@@ -1049,26 +1050,49 @@ struct cifsFileInfo *find_writable_file(struct cifsInodeInfo *cifs_inode)
                    ((open_file->pfile->f_flags & O_RDWR) ||
                     (open_file->pfile->f_flags & O_WRONLY))) {
                        atomic_inc(&open_file->wrtPending);
+
+                       if (!open_file->invalidHandle) {
+                               /* found a good writable file */
+                               read_unlock(&GlobalSMBSeslock);
+                               return open_file;
+                       }
+       
                        read_unlock(&GlobalSMBSeslock);
-                       if (open_file->invalidHandle) {
-                               rc = cifs_reopen_file(open_file->pfile, FALSE);
-                               /* if it fails, try another handle - might be */
-                               /* dangerous to hold up writepages with retry */
-                               if (rc) {
-                                       cFYI(1, ("wp failed on reopen file"));
+                       /* Had to unlock since following call can block */
+                       rc = cifs_reopen_file(open_file->pfile, FALSE);
+                       if (!rc) { 
+                               if (!open_file->closePend)
+                                       return open_file;
+                               else { /* start over in case this was deleted */
+                                      /* since the list could be modified */
                                        read_lock(&GlobalSMBSeslock);
-                                       /* can not use this handle, no write
-                                       pending on this one after all */
                                        atomic_dec(&open_file->wrtPending);
-                                       continue;
+                                       goto refind_writable;
                                }
                        }
-                       if (open_file->closePend) {
-                               read_lock(&GlobalSMBSeslock);
-                               atomic_dec(&open_file->wrtPending);
-                               continue;
-                       }
-                       return open_file;
+
+                       /* if it fails, try another handle if possible -
+                       (we can not do this if closePending since
+                       loop could be modified - in which case we
+                       have to start at the beginning of the list
+                       again. Note that it would be bad
+                       to hold up writepages here (rather than
+                       in caller) with continuous retries */
+                       cFYI(1, ("wp failed on reopen file"));
+                       read_lock(&GlobalSMBSeslock);
+                       /* can not use this handle, no write
+                          pending on this one after all */
+                       atomic_dec(&open_file->wrtPending);
+                       
+                       if (open_file->closePend) /* list could have changed */
+                               goto refind_writable;
+                       /* else we simply continue to the next entry. Thus
+                          we do not loop on reopen errors.  If we
+                          can not reopen the file, for example if we
+                          reconnected to a server with another client
+                          racing to delete or lock the file we would not
+                          make progress if we restarted before the beginning
+                          of the loop here. */
                }
        }
        read_unlock(&GlobalSMBSeslock);