ceph: nfs re-export support
authorSage Weil <sage@newdream.net>
Tue, 6 Oct 2009 18:31:13 +0000 (11:31 -0700)
committerSage Weil <sage@newdream.net>
Tue, 6 Oct 2009 18:31:13 +0000 (11:31 -0700)
Basic NFS re-export support is included.  This mostly works.  However,
Ceph's MDS design precludes the ability to generate a (small)
filehandle that will be valid forever, so this is of limited utility.

Signed-off-by: Sage Weil <sage@newdream.net>
fs/ceph/export.c [new file with mode: 0644]

diff --git a/fs/ceph/export.c b/fs/ceph/export.c
new file mode 100644 (file)
index 0000000..fc68e39
--- /dev/null
@@ -0,0 +1,223 @@
+#include "ceph_debug.h"
+
+#include <linux/exportfs.h>
+#include <asm/unaligned.h>
+
+#include "super.h"
+
+/*
+ * NFS export support
+ *
+ * NFS re-export of a ceph mount is, at present, only semireliable.
+ * The basic issue is that the Ceph architectures doesn't lend itself
+ * well to generating filehandles that will remain valid forever.
+ *
+ * So, we do our best.  If you're lucky, your inode will be in the
+ * client's cache.  If it's not, and you have a connectable fh, then
+ * the MDS server may be able to find it for you.  Otherwise, you get
+ * ESTALE.
+ *
+ * There are ways to this more reliable, but in the non-connectable fh
+ * case, we won't every work perfectly, and in the connectable case,
+ * some changes are needed on the MDS side to work better.
+ */
+
+/*
+ * Basic fh
+ */
+struct ceph_nfs_fh {
+       u64 ino;
+} __attribute__ ((packed));
+
+/*
+ * Larger 'connectable' fh that includes parent ino and name hash.
+ * Use this whenever possible, as it works more reliably.
+ */
+struct ceph_nfs_confh {
+       u64 ino, parent_ino;
+       u32 parent_name_hash;
+} __attribute__ ((packed));
+
+static int ceph_encode_fh(struct dentry *dentry, u32 *rawfh, int *max_len,
+                         int connectable)
+{
+       struct ceph_nfs_fh *fh = (void *)rawfh;
+       struct ceph_nfs_confh *cfh = (void *)rawfh;
+       struct dentry *parent = dentry->d_parent;
+       struct inode *inode = dentry->d_inode;
+       int type;
+
+       /* don't re-export snaps */
+       if (ceph_snap(inode) != CEPH_NOSNAP)
+               return -EINVAL;
+
+       if (*max_len >= sizeof(*cfh)) {
+               dout("encode_fh %p connectable\n", dentry);
+               cfh->ino = ceph_ino(dentry->d_inode);
+               cfh->parent_ino = ceph_ino(parent->d_inode);
+               cfh->parent_name_hash = parent->d_name.hash;
+               *max_len = sizeof(*cfh);
+               type = 2;
+       } else if (*max_len > sizeof(*fh)) {
+               if (connectable)
+                       return -ENOSPC;
+               dout("encode_fh %p\n", dentry);
+               fh->ino = ceph_ino(dentry->d_inode);
+               *max_len = sizeof(*fh);
+               type = 1;
+       } else {
+               return -ENOSPC;
+       }
+       return type;
+}
+
+/*
+ * convert regular fh to dentry
+ *
+ * FIXME: we should try harder by querying the mds for the ino.
+ */
+static struct dentry *__fh_to_dentry(struct super_block *sb,
+                                    struct ceph_nfs_fh *fh)
+{
+       struct inode *inode;
+       struct dentry *dentry;
+       struct ceph_vino vino;
+       int err;
+
+       dout("__fh_to_dentry %llx\n", fh->ino);
+       vino.ino = fh->ino;
+       vino.snap = CEPH_NOSNAP;
+       inode = ceph_find_inode(sb, vino);
+       if (!inode)
+               return ERR_PTR(-ESTALE);
+
+       dentry = d_obtain_alias(inode);
+       if (!dentry) {
+               pr_err("fh_to_dentry %llx -- inode %p but ENOMEM\n",
+                      fh->ino, inode);
+               iput(inode);
+               return ERR_PTR(-ENOMEM);
+       }
+       err = ceph_init_dentry(dentry);
+
+       if (err < 0) {
+               iput(inode);
+               return ERR_PTR(err);
+       }
+       dout("__fh_to_dentry %llx %p dentry %p\n", fh->ino, inode, dentry);
+       return dentry;
+}
+
+/*
+ * convert connectable fh to dentry
+ */
+static struct dentry *__cfh_to_dentry(struct super_block *sb,
+                                     struct ceph_nfs_confh *cfh)
+{
+       struct ceph_mds_client *mdsc = &ceph_client(sb)->mdsc;
+       struct inode *inode;
+       struct dentry *dentry;
+       struct ceph_vino vino;
+       int err;
+
+       dout("__cfh_to_dentry %llx (%llx/%x)\n",
+            cfh->ino, cfh->parent_ino, cfh->parent_name_hash);
+
+       vino.ino = cfh->ino;
+       vino.snap = CEPH_NOSNAP;
+       inode = ceph_find_inode(sb, vino);
+       if (!inode) {
+               struct ceph_mds_request *req;
+
+               req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_LOOKUPHASH,
+                                              USE_ANY_MDS);
+               if (IS_ERR(req))
+                       return ERR_PTR(PTR_ERR(req));
+
+               req->r_ino1 = vino;
+               req->r_ino2.ino = cfh->parent_ino;
+               req->r_ino2.snap = CEPH_NOSNAP;
+               req->r_path2 = kmalloc(16, GFP_NOFS);
+               snprintf(req->r_path2, 16, "%d", cfh->parent_name_hash);
+               req->r_num_caps = 1;
+               err = ceph_mdsc_do_request(mdsc, NULL, req);
+               ceph_mdsc_put_request(req);
+               inode = ceph_find_inode(sb, vino);
+               if (!inode)
+                       return ERR_PTR(err ? err : -ESTALE);
+       }
+
+       dentry = d_obtain_alias(inode);
+       if (!dentry) {
+               pr_err("cfh_to_dentry %llx -- inode %p but ENOMEM\n",
+                      cfh->ino, inode);
+               iput(inode);
+               return ERR_PTR(-ENOMEM);
+       }
+       err = ceph_init_dentry(dentry);
+       if (err < 0) {
+               iput(inode);
+               return ERR_PTR(err);
+       }
+       dout("__cfh_to_dentry %llx %p dentry %p\n", cfh->ino, inode, dentry);
+       return dentry;
+}
+
+static struct dentry *ceph_fh_to_dentry(struct super_block *sb, struct fid *fid,
+                                       int fh_len, int fh_type)
+{
+       if (fh_type == 1)
+               return __fh_to_dentry(sb, (struct ceph_nfs_fh *)fid->raw);
+       else
+               return __cfh_to_dentry(sb, (struct ceph_nfs_confh *)fid->raw);
+}
+
+/*
+ * get parent, if possible.
+ *
+ * FIXME: we could do better by querying the mds to discover the
+ * parent.
+ */
+static struct dentry *ceph_fh_to_parent(struct super_block *sb,
+                                        struct fid *fid,
+                                       int fh_len, int fh_type)
+{
+       struct ceph_nfs_confh *cfh = (void *)fid->raw;
+       struct ceph_vino vino;
+       struct inode *inode;
+       struct dentry *dentry;
+       int err;
+
+       if (fh_type == 1)
+               return ERR_PTR(-ESTALE);
+
+       pr_debug("fh_to_parent %llx/%d\n", cfh->parent_ino,
+                cfh->parent_name_hash);
+
+       vino.ino = cfh->ino;
+       vino.snap = CEPH_NOSNAP;
+       inode = ceph_find_inode(sb, vino);
+       if (!inode)
+               return ERR_PTR(-ESTALE);
+
+       dentry = d_obtain_alias(inode);
+       if (!dentry) {
+               pr_err("fh_to_parent %llx -- inode %p but ENOMEM\n",
+                      cfh->ino, inode);
+               iput(inode);
+               return ERR_PTR(-ENOMEM);
+       }
+       err = ceph_init_dentry(dentry);
+       if (err < 0) {
+               iput(inode);
+               return ERR_PTR(err);
+       }
+       dout("fh_to_parent %llx %p dentry %p\n", cfh->ino, inode, dentry);
+       return dentry;
+}
+
+const struct export_operations ceph_export_ops = {
+       .encode_fh = ceph_encode_fh,
+       .fh_to_dentry = ceph_fh_to_dentry,
+       .fh_to_parent = ceph_fh_to_parent,
+};