CIFS: keep FileInfo handle live during oplock break
authorAurelien Aptel <aaptel@suse.com>
Fri, 29 Mar 2019 09:49:12 +0000 (10:49 +0100)
committerSteve French <stfrench@microsoft.com>
Tue, 16 Apr 2019 14:38:38 +0000 (09:38 -0500)
In the oplock break handler, writing pending changes from pages puts
the FileInfo handle. If the refcount reaches zero it closes the handle
and waits for any oplock break handler to return, thus causing a deadlock.

To prevent this situation:

* We add a wait flag to cifsFileInfo_put() to decide whether we should
  wait for running/pending oplock break handlers

* We keep an additionnal reference of the SMB FileInfo handle so that
  for the rest of the handler putting the handle won't close it.
  - The ref is bumped everytime we queue the handler via the
    cifs_queue_oplock_break() helper.
  - The ref is decremented at the end of the handler

This bug was triggered by xfstest 464.

Also important fix to address the various reports of
oops in smb2_push_mandatory_locks

Signed-off-by: Aurelien Aptel <aaptel@suse.com>
Signed-off-by: Steve French <stfrench@microsoft.com>
Reviewed-by: Pavel Shilovsky <pshilov@microsoft.com>
CC: Stable <stable@vger.kernel.org>
fs/cifs/cifsglob.h
fs/cifs/file.c
fs/cifs/misc.c
fs/cifs/smb2misc.c

index 5b18d45857409eb06624a894e2308e6728c39ba3..585ad3207cb120a34c3da24e418dd20a6daf04cf 100644 (file)
@@ -1333,6 +1333,7 @@ cifsFileInfo_get_locked(struct cifsFileInfo *cifs_file)
 }
 
 struct cifsFileInfo *cifsFileInfo_get(struct cifsFileInfo *cifs_file);
+void _cifsFileInfo_put(struct cifsFileInfo *cifs_file, bool wait_oplock_hdlr);
 void cifsFileInfo_put(struct cifsFileInfo *cifs_file);
 
 #define CIFS_CACHE_READ_FLG    1
@@ -1855,6 +1856,7 @@ GLOBAL_EXTERN spinlock_t gidsidlock;
 #endif /* CONFIG_CIFS_ACL */
 
 void cifs_oplock_break(struct work_struct *work);
+void cifs_queue_oplock_break(struct cifsFileInfo *cfile);
 
 extern const struct slow_work_ops cifs_oplock_break_ops;
 extern struct workqueue_struct *cifsiod_wq;
index 89006e044973ec2d97ad784236f66cb447431693..9c0ccc06d172e9371513105a9a19cb9cef68ddc0 100644 (file)
@@ -360,12 +360,30 @@ cifsFileInfo_get(struct cifsFileInfo *cifs_file)
        return cifs_file;
 }
 
-/*
- * Release a reference on the file private data. This may involve closing
- * the filehandle out on the server. Must be called without holding
- * tcon->open_file_lock and cifs_file->file_info_lock.
+/**
+ * cifsFileInfo_put - release a reference of file priv data
+ *
+ * Always potentially wait for oplock handler. See _cifsFileInfo_put().
  */
 void cifsFileInfo_put(struct cifsFileInfo *cifs_file)
+{
+       _cifsFileInfo_put(cifs_file, true);
+}
+
+/**
+ * _cifsFileInfo_put - release a reference of file priv data
+ *
+ * This may involve closing the filehandle @cifs_file out on the
+ * server. Must be called without holding tcon->open_file_lock and
+ * cifs_file->file_info_lock.
+ *
+ * If @wait_for_oplock_handler is true and we are releasing the last
+ * reference, wait for any running oplock break handler of the file
+ * and cancel any pending one. If calling this function from the
+ * oplock break handler, you need to pass false.
+ *
+ */
+void _cifsFileInfo_put(struct cifsFileInfo *cifs_file, bool wait_oplock_handler)
 {
        struct inode *inode = d_inode(cifs_file->dentry);
        struct cifs_tcon *tcon = tlink_tcon(cifs_file->tlink);
@@ -414,7 +432,8 @@ void cifsFileInfo_put(struct cifsFileInfo *cifs_file)
 
        spin_unlock(&tcon->open_file_lock);
 
-       oplock_break_cancelled = cancel_work_sync(&cifs_file->oplock_break);
+       oplock_break_cancelled = wait_oplock_handler ?
+               cancel_work_sync(&cifs_file->oplock_break) : false;
 
        if (!tcon->need_reconnect && !cifs_file->invalidHandle) {
                struct TCP_Server_Info *server = tcon->ses->server;
@@ -4603,6 +4622,7 @@ void cifs_oplock_break(struct work_struct *work)
                                                             cinode);
                cifs_dbg(FYI, "Oplock release rc = %d\n", rc);
        }
+       _cifsFileInfo_put(cfile, false /* do not wait for ourself */);
        cifs_done_oplock_break(cinode);
 }
 
index bee203055b300b1d13a284e9571225f7ba4888c8..1e1626a2cfc393afb5ea6470261e6bd9703a2e40 100644 (file)
@@ -501,8 +501,7 @@ is_valid_oplock_break(char *buffer, struct TCP_Server_Info *srv)
                                           CIFS_INODE_DOWNGRADE_OPLOCK_TO_L2,
                                           &pCifsInode->flags);
 
-                               queue_work(cifsoplockd_wq,
-                                          &netfile->oplock_break);
+                               cifs_queue_oplock_break(netfile);
                                netfile->oplock_break_cancelled = false;
 
                                spin_unlock(&tcon->open_file_lock);
@@ -607,6 +606,28 @@ void cifs_put_writer(struct cifsInodeInfo *cinode)
        spin_unlock(&cinode->writers_lock);
 }
 
+/**
+ * cifs_queue_oplock_break - queue the oplock break handler for cfile
+ *
+ * This function is called from the demultiplex thread when it
+ * receives an oplock break for @cfile.
+ *
+ * Assumes the tcon->open_file_lock is held.
+ * Assumes cfile->file_info_lock is NOT held.
+ */
+void cifs_queue_oplock_break(struct cifsFileInfo *cfile)
+{
+       /*
+        * Bump the handle refcount now while we hold the
+        * open_file_lock to enforce the validity of it for the oplock
+        * break handler. The matching put is done at the end of the
+        * handler.
+        */
+       cifsFileInfo_get(cfile);
+
+       queue_work(cifsoplockd_wq, &cfile->oplock_break);
+}
+
 void cifs_done_oplock_break(struct cifsInodeInfo *cinode)
 {
        clear_bit(CIFS_INODE_PENDING_OPLOCK_BREAK, &cinode->flags);
index 0e3570e40ff8e8d233389063290ccdecdfb44a25..e311f58dc1c82809de0283e434a9aff09356825f 100644 (file)
@@ -555,7 +555,7 @@ smb2_tcon_has_lease(struct cifs_tcon *tcon, struct smb2_lease_break *rsp,
                        clear_bit(CIFS_INODE_DOWNGRADE_OPLOCK_TO_L2,
                                  &cinode->flags);
 
-               queue_work(cifsoplockd_wq, &cfile->oplock_break);
+               cifs_queue_oplock_break(cfile);
                kfree(lw);
                return true;
        }
@@ -712,8 +712,8 @@ smb2_is_valid_oplock_break(char *buffer, struct TCP_Server_Info *server)
                                           CIFS_INODE_DOWNGRADE_OPLOCK_TO_L2,
                                           &cinode->flags);
                                spin_unlock(&cfile->file_info_lock);
-                               queue_work(cifsoplockd_wq,
-                                          &cfile->oplock_break);
+
+                               cifs_queue_oplock_break(cfile);
 
                                spin_unlock(&tcon->open_file_lock);
                                spin_unlock(&cifs_tcp_ses_lock);