+/*
+ * ocfs2_file_lock() and ocfs2_file_unlock() map to a single pair of
+ * flock() calls. The locking approach this requires is sufficiently
+ * different from all other cluster lock types that we implement a
+ * seperate path to the "low-level" dlm calls. In particular:
+ *
+ * - No optimization of lock levels is done - we take at exactly
+ * what's been requested.
+ *
+ * - No lock caching is employed. We immediately downconvert to
+ * no-lock at unlock time. This also means flock locks never go on
+ * the blocking list).
+ *
+ * - Since userspace can trivially deadlock itself with flock, we make
+ * sure to allow cancellation of a misbehaving applications flock()
+ * request.
+ *
+ * - Access to any flock lockres doesn't require concurrency, so we
+ * can simplify the code by requiring the caller to guarantee
+ * serialization of dlmglue flock calls.
+ */
+int ocfs2_file_lock(struct file *file, int ex, int trylock)
+{
+ int ret, level = ex ? DLM_LOCK_EX : DLM_LOCK_PR;
+ unsigned int lkm_flags = trylock ? DLM_LKF_NOQUEUE : 0;
+ unsigned long flags;
+ struct ocfs2_file_private *fp = file->private_data;
+ struct ocfs2_lock_res *lockres = &fp->fp_flock;
+ struct ocfs2_super *osb = OCFS2_SB(file->f_mapping->host->i_sb);
+ struct ocfs2_mask_waiter mw;
+
+ ocfs2_init_mask_waiter(&mw);
+
+ if ((lockres->l_flags & OCFS2_LOCK_BUSY) ||
+ (lockres->l_level > DLM_LOCK_NL)) {
+ mlog(ML_ERROR,
+ "File lock \"%s\" has busy or locked state: flags: 0x%lx, "
+ "level: %u\n", lockres->l_name, lockres->l_flags,
+ lockres->l_level);
+ return -EINVAL;
+ }
+
+ spin_lock_irqsave(&lockres->l_lock, flags);
+ if (!(lockres->l_flags & OCFS2_LOCK_ATTACHED)) {
+ lockres_add_mask_waiter(lockres, &mw, OCFS2_LOCK_BUSY, 0);
+ spin_unlock_irqrestore(&lockres->l_lock, flags);
+
+ /*
+ * Get the lock at NLMODE to start - that way we
+ * can cancel the upconvert request if need be.
+ */
+ ret = ocfs2_lock_create(osb, lockres, DLM_LOCK_NL, 0);
+ if (ret < 0) {
+ mlog_errno(ret);
+ goto out;
+ }
+
+ ret = ocfs2_wait_for_mask(&mw);
+ if (ret) {
+ mlog_errno(ret);
+ goto out;
+ }
+ spin_lock_irqsave(&lockres->l_lock, flags);
+ }
+
+ lockres->l_action = OCFS2_AST_CONVERT;
+ lkm_flags |= DLM_LKF_CONVERT;
+ lockres->l_requested = level;
+ lockres_or_flags(lockres, OCFS2_LOCK_BUSY);
+
+ lockres_add_mask_waiter(lockres, &mw, OCFS2_LOCK_BUSY, 0);
+ spin_unlock_irqrestore(&lockres->l_lock, flags);
+
+ ret = ocfs2_dlm_lock(osb->cconn, level, &lockres->l_lksb, lkm_flags,
+ lockres->l_name, OCFS2_LOCK_ID_MAX_LEN - 1,
+ lockres);
+ if (ret) {
+ if (!trylock || (ret != -EAGAIN)) {
+ ocfs2_log_dlm_error("ocfs2_dlm_lock", ret, lockres);
+ ret = -EINVAL;
+ }
+
+ ocfs2_recover_from_dlm_error(lockres, 1);
+ lockres_remove_mask_waiter(lockres, &mw);
+ goto out;
+ }
+
+ ret = ocfs2_wait_for_mask_interruptible(&mw, lockres);
+ if (ret == -ERESTARTSYS) {
+ /*
+ * Userspace can cause deadlock itself with
+ * flock(). Current behavior locally is to allow the
+ * deadlock, but abort the system call if a signal is
+ * received. We follow this example, otherwise a
+ * poorly written program could sit in kernel until
+ * reboot.
+ *
+ * Handling this is a bit more complicated for Ocfs2
+ * though. We can't exit this function with an
+ * outstanding lock request, so a cancel convert is
+ * required. We intentionally overwrite 'ret' - if the
+ * cancel fails and the lock was granted, it's easier
+ * to just bubble sucess back up to the user.
+ */
+ ret = ocfs2_flock_handle_signal(lockres, level);
+ } else if (!ret && (level > lockres->l_level)) {
+ /* Trylock failed asynchronously */
+ BUG_ON(!trylock);
+ ret = -EAGAIN;
+ }
+
+out:
+
+ mlog(0, "Lock: \"%s\" ex: %d, trylock: %d, returns: %d\n",
+ lockres->l_name, ex, trylock, ret);
+ return ret;
+}
+
+void ocfs2_file_unlock(struct file *file)
+{
+ int ret;
+ unsigned int gen;
+ unsigned long flags;
+ struct ocfs2_file_private *fp = file->private_data;
+ struct ocfs2_lock_res *lockres = &fp->fp_flock;
+ struct ocfs2_super *osb = OCFS2_SB(file->f_mapping->host->i_sb);
+ struct ocfs2_mask_waiter mw;
+
+ ocfs2_init_mask_waiter(&mw);
+
+ if (!(lockres->l_flags & OCFS2_LOCK_ATTACHED))
+ return;
+
+ if (lockres->l_level == DLM_LOCK_NL)
+ return;
+
+ mlog(0, "Unlock: \"%s\" flags: 0x%lx, level: %d, act: %d\n",
+ lockres->l_name, lockres->l_flags, lockres->l_level,
+ lockres->l_action);
+
+ spin_lock_irqsave(&lockres->l_lock, flags);
+ /*
+ * Fake a blocking ast for the downconvert code.
+ */
+ lockres_or_flags(lockres, OCFS2_LOCK_BLOCKED);
+ lockres->l_blocking = DLM_LOCK_EX;
+
+ gen = ocfs2_prepare_downconvert(lockres, DLM_LOCK_NL);
+ lockres_add_mask_waiter(lockres, &mw, OCFS2_LOCK_BUSY, 0);
+ spin_unlock_irqrestore(&lockres->l_lock, flags);
+
+ ret = ocfs2_downconvert_lock(osb, lockres, DLM_LOCK_NL, 0, gen);
+ if (ret) {
+ mlog_errno(ret);
+ return;
+ }
+
+ ret = ocfs2_wait_for_mask(&mw);
+ if (ret)
+ mlog_errno(ret);
+}
+
+static void ocfs2_downconvert_on_unlock(struct ocfs2_super *osb,
+ struct ocfs2_lock_res *lockres)