4 files changed, 88 insertions, 3 deletions

diff --git a/fs/ceph/caps.c b/fs/ceph/caps.c
index 789be30d6ee2..2321e5ddb664 100644
--- a/fs/ceph/caps.c
+++ b/fs/ceph/caps.c
@@ -1627,6 +1627,7 @@ void ceph_flush_snaps(struct ceph_inode_info *ci,
 	struct inode *inode = &ci->netfs.inode;
 	struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
 	struct ceph_mds_session *session = NULL;
+	bool need_put = false;
 	int mds;
 
 	dout("ceph_flush_snaps %p\n", inode);
@@ -1671,8 +1672,13 @@ out:
 		ceph_put_mds_session(session);
 	/* we flushed them all; remove this inode from the queue */
 	spin_lock(&mdsc->snap_flush_lock);
+	if (!list_empty(&ci->i_snap_flush_item))
+		need_put = true;
 	list_del_init(&ci->i_snap_flush_item);
 	spin_unlock(&mdsc->snap_flush_lock);
+
+	if (need_put)
+		iput(inode);
 }
 
 /*
diff --git a/fs/ceph/file.c b/fs/ceph/file.c
index f4d8bf7dec88..4285f6cb5d3b 100644
--- a/fs/ceph/file.c
+++ b/fs/ceph/file.c
@@ -1746,6 +1746,69 @@ again:
 }
 
 /*
+ * Wrap filemap_splice_read with checks for cap bits on the inode.
+ * Atomically grab references, so that those bits are not released
+ * back to the MDS mid-read.
+ */
+static ssize_t ceph_splice_read(struct file *in, loff_t *ppos,
+				struct pipe_inode_info *pipe,
+				size_t len, unsigned int flags)
+{
+	struct ceph_file_info *fi = in->private_data;
+	struct inode *inode = file_inode(in);
+	struct ceph_inode_info *ci = ceph_inode(inode);
+	ssize_t ret;
+	int want = 0, got = 0;
+	CEPH_DEFINE_RW_CONTEXT(rw_ctx, 0);
+
+	dout("splice_read %p %llx.%llx %llu~%zu trying to get caps on %p\n",
+	     inode, ceph_vinop(inode), *ppos, len, inode);
+
+	if (ceph_inode_is_shutdown(inode))
+		return -ESTALE;
+
+	if (ceph_has_inline_data(ci) ||
+	    (fi->flags & CEPH_F_SYNC))
+		return copy_splice_read(in, ppos, pipe, len, flags);
+
+	ceph_start_io_read(inode);
+
+	want = CEPH_CAP_FILE_CACHE;
+	if (fi->fmode & CEPH_FILE_MODE_LAZY)
+		want |= CEPH_CAP_FILE_LAZYIO;
+
+	ret = ceph_get_caps(in, CEPH_CAP_FILE_RD, want, -1, &got);
+	if (ret < 0)
+		goto out_end;
+
+	if ((got & (CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO)) == 0) {
+		dout("splice_read/sync %p %llx.%llx %llu~%zu got cap refs on %s\n",
+		     inode, ceph_vinop(inode), *ppos, len,
+		     ceph_cap_string(got));
+
+		ceph_put_cap_refs(ci, got);
+		ceph_end_io_read(inode);
+		return copy_splice_read(in, ppos, pipe, len, flags);
+	}
+
+	dout("splice_read %p %llx.%llx %llu~%zu got cap refs on %s\n",
+	     inode, ceph_vinop(inode), *ppos, len, ceph_cap_string(got));
+
+	rw_ctx.caps = got;
+	ceph_add_rw_context(fi, &rw_ctx);
+	ret = filemap_splice_read(in, ppos, pipe, len, flags);
+	ceph_del_rw_context(fi, &rw_ctx);
+
+	dout("splice_read %p %llx.%llx dropping cap refs on %s = %zd\n",
+	     inode, ceph_vinop(inode), ceph_cap_string(got), ret);
+
+	ceph_put_cap_refs(ci, got);
+out_end:
+	ceph_end_io_read(inode);
+	return ret;
+}
+
+/*
  * Take cap references to avoid releasing caps to MDS mid-write.
  *
  * If we are synchronous, and write with an old snap context, the OSD
@@ -2593,7 +2656,7 @@ const struct file_operations ceph_file_fops = {
 	.lock = ceph_lock,
 	.setlease = simple_nosetlease,
 	.flock = ceph_flock,
-	.splice_read = generic_file_splice_read,
+	.splice_read = ceph_splice_read,
 	.splice_write = iter_file_splice_write,
 	.unlocked_ioctl = ceph_ioctl,
 	.compat_ioctl = compat_ptr_ioctl,
diff --git a/fs/ceph/mds_client.c b/fs/ceph/mds_client.c
index 29cf00220b09..4c0f22acf53d 100644
--- a/fs/ceph/mds_client.c
+++ b/fs/ceph/mds_client.c
@@ -3942,7 +3942,7 @@ static int reconnect_caps_cb(struct inode *inode, int mds, void *arg)
 	struct dentry *dentry;
 	struct ceph_cap *cap;
 	char *path;
-	int pathlen = 0, err = 0;
+	int pathlen = 0, err;
 	u64 pathbase;
 	u64 snap_follows;
 
@@ -3965,6 +3965,7 @@ static int reconnect_caps_cb(struct inode *inode, int mds, void *arg)
 	cap = __get_cap_for_mds(ci, mds);
 	if (!cap) {
 		spin_unlock(&ci->i_ceph_lock);
+		err = 0;
 		goto out_err;
 	}
 	dout(" adding %p ino %llx.%llx cap %p %lld %s\n",
diff --git a/fs/ceph/snap.c b/fs/ceph/snap.c
index 87007203f130..2e73ba62bd7a 100644
--- a/fs/ceph/snap.c
+++ b/fs/ceph/snap.c
@@ -693,8 +693,10 @@ int __ceph_finish_cap_snap(struct ceph_inode_info *ci,
 	     capsnap->size);
 
 	spin_lock(&mdsc->snap_flush_lock);
-	if (list_empty(&ci->i_snap_flush_item))
+	if (list_empty(&ci->i_snap_flush_item)) {
+		ihold(inode);
 		list_add_tail(&ci->i_snap_flush_item, &mdsc->snap_flush_list);
+	}
 	spin_unlock(&mdsc->snap_flush_lock);
 	return 1;  /* caller may want to ceph_flush_snaps */
 }
@@ -1111,6 +1113,19 @@ skip_inode:
 				continue;
 			adjust_snap_realm_parent(mdsc, child, realm->ino);
 		}
+	} else {
+		/*
+		 * In the non-split case both 'num_split_inos' and
+		 * 'num_split_realms' should be 0, making this a no-op.
+		 * However the MDS happens to populate 'split_realms' list
+		 * in one of the UPDATE op cases by mistake.
+		 *
+		 * Skip both lists just in case to ensure that 'p' is
+		 * positioned at the start of realm info, as expected by
+		 * ceph_update_snap_trace().
+		 */
+		p += sizeof(u64) * num_split_inos;
+		p += sizeof(u64) * num_split_realms;
 	}
 
 	/*