blob: e41056174bf81ad96397500ad89b19754a9bc0ff [file] [log] [blame]
Yehuda Sadeh3d14c5d2010-04-06 15:14:15 -07001#include <linux/ceph/ceph_debug.h>
Sage Weila8e63b72009-10-06 11:31:13 -07002
3#include <linux/exportfs.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +09004#include <linux/slab.h>
Sage Weila8e63b72009-10-06 11:31:13 -07005#include <asm/unaligned.h>
6
7#include "super.h"
Yehuda Sadeh3d14c5d2010-04-06 15:14:15 -07008#include "mds_client.h"
Sage Weila8e63b72009-10-06 11:31:13 -07009
10/*
11 * NFS export support
12 *
13 * NFS re-export of a ceph mount is, at present, only semireliable.
14 * The basic issue is that the Ceph architectures doesn't lend itself
15 * well to generating filehandles that will remain valid forever.
16 *
17 * So, we do our best. If you're lucky, your inode will be in the
18 * client's cache. If it's not, and you have a connectable fh, then
19 * the MDS server may be able to find it for you. Otherwise, you get
20 * ESTALE.
21 *
22 * There are ways to this more reliable, but in the non-connectable fh
23 * case, we won't every work perfectly, and in the connectable case,
24 * some changes are needed on the MDS side to work better.
25 */
26
27/*
28 * Basic fh
29 */
30struct ceph_nfs_fh {
31 u64 ino;
32} __attribute__ ((packed));
33
34/*
35 * Larger 'connectable' fh that includes parent ino and name hash.
36 * Use this whenever possible, as it works more reliably.
37 */
38struct ceph_nfs_confh {
39 u64 ino, parent_ino;
40 u32 parent_name_hash;
41} __attribute__ ((packed));
42
43static int ceph_encode_fh(struct dentry *dentry, u32 *rawfh, int *max_len,
44 int connectable)
45{
Aneesh Kumar K.V92923dc2010-10-05 16:03:41 +053046 int type;
Sage Weila8e63b72009-10-06 11:31:13 -070047 struct ceph_nfs_fh *fh = (void *)rawfh;
48 struct ceph_nfs_confh *cfh = (void *)rawfh;
49 struct dentry *parent = dentry->d_parent;
50 struct inode *inode = dentry->d_inode;
Aneesh Kumar K.V92923dc2010-10-05 16:03:41 +053051 int connected_handle_length = sizeof(*cfh)/4;
52 int handle_length = sizeof(*fh)/4;
Sage Weila8e63b72009-10-06 11:31:13 -070053
54 /* don't re-export snaps */
55 if (ceph_snap(inode) != CEPH_NOSNAP)
56 return -EINVAL;
57
Aneesh Kumar K.V92923dc2010-10-05 16:03:41 +053058 if (*max_len >= connected_handle_length) {
Sage Weila8e63b72009-10-06 11:31:13 -070059 dout("encode_fh %p connectable\n", dentry);
60 cfh->ino = ceph_ino(dentry->d_inode);
61 cfh->parent_ino = ceph_ino(parent->d_inode);
Sage Weil6c0f3af2010-11-16 11:14:34 -080062 cfh->parent_name_hash = ceph_dentry_hash(parent);
Aneesh Kumar K.V92923dc2010-10-05 16:03:41 +053063 *max_len = connected_handle_length;
Sage Weila8e63b72009-10-06 11:31:13 -070064 type = 2;
Aneesh Kumar K.V92923dc2010-10-05 16:03:41 +053065 } else if (*max_len >= handle_length) {
Aneesh Kumar K.Vbba0cd02010-10-05 16:03:42 +053066 if (connectable) {
67 *max_len = connected_handle_length;
Aneesh Kumar K.V92923dc2010-10-05 16:03:41 +053068 return 255;
Aneesh Kumar K.Vbba0cd02010-10-05 16:03:42 +053069 }
Sage Weila8e63b72009-10-06 11:31:13 -070070 dout("encode_fh %p\n", dentry);
71 fh->ino = ceph_ino(dentry->d_inode);
Aneesh Kumar K.V92923dc2010-10-05 16:03:41 +053072 *max_len = handle_length;
Sage Weila8e63b72009-10-06 11:31:13 -070073 type = 1;
74 } else {
Aneesh Kumar K.Vbba0cd02010-10-05 16:03:42 +053075 *max_len = handle_length;
Aneesh Kumar K.V92923dc2010-10-05 16:03:41 +053076 return 255;
Sage Weila8e63b72009-10-06 11:31:13 -070077 }
78 return type;
79}
80
81/*
82 * convert regular fh to dentry
83 *
84 * FIXME: we should try harder by querying the mds for the ino.
85 */
86static struct dentry *__fh_to_dentry(struct super_block *sb,
87 struct ceph_nfs_fh *fh)
88{
89 struct inode *inode;
90 struct dentry *dentry;
91 struct ceph_vino vino;
92 int err;
93
94 dout("__fh_to_dentry %llx\n", fh->ino);
95 vino.ino = fh->ino;
96 vino.snap = CEPH_NOSNAP;
97 inode = ceph_find_inode(sb, vino);
98 if (!inode)
99 return ERR_PTR(-ESTALE);
100
101 dentry = d_obtain_alias(inode);
Dan Carpenter0d509c92010-04-21 12:31:13 +0200102 if (IS_ERR(dentry)) {
Sage Weila8e63b72009-10-06 11:31:13 -0700103 pr_err("fh_to_dentry %llx -- inode %p but ENOMEM\n",
104 fh->ino, inode);
105 iput(inode);
Dan Carpenter0d509c92010-04-21 12:31:13 +0200106 return dentry;
Sage Weila8e63b72009-10-06 11:31:13 -0700107 }
108 err = ceph_init_dentry(dentry);
109
110 if (err < 0) {
111 iput(inode);
112 return ERR_PTR(err);
113 }
114 dout("__fh_to_dentry %llx %p dentry %p\n", fh->ino, inode, dentry);
115 return dentry;
116}
117
118/*
119 * convert connectable fh to dentry
120 */
121static struct dentry *__cfh_to_dentry(struct super_block *sb,
122 struct ceph_nfs_confh *cfh)
123{
Yehuda Sadeh3d14c5d2010-04-06 15:14:15 -0700124 struct ceph_mds_client *mdsc = ceph_sb_to_client(sb)->mdsc;
Sage Weila8e63b72009-10-06 11:31:13 -0700125 struct inode *inode;
126 struct dentry *dentry;
127 struct ceph_vino vino;
128 int err;
129
130 dout("__cfh_to_dentry %llx (%llx/%x)\n",
131 cfh->ino, cfh->parent_ino, cfh->parent_name_hash);
132
133 vino.ino = cfh->ino;
134 vino.snap = CEPH_NOSNAP;
135 inode = ceph_find_inode(sb, vino);
136 if (!inode) {
137 struct ceph_mds_request *req;
138
139 req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_LOOKUPHASH,
140 USE_ANY_MDS);
141 if (IS_ERR(req))
Julia Lawall7e34bc52010-05-22 12:01:14 +0200142 return ERR_CAST(req);
Sage Weila8e63b72009-10-06 11:31:13 -0700143
144 req->r_ino1 = vino;
145 req->r_ino2.ino = cfh->parent_ino;
146 req->r_ino2.snap = CEPH_NOSNAP;
147 req->r_path2 = kmalloc(16, GFP_NOFS);
148 snprintf(req->r_path2, 16, "%d", cfh->parent_name_hash);
149 req->r_num_caps = 1;
150 err = ceph_mdsc_do_request(mdsc, NULL, req);
151 ceph_mdsc_put_request(req);
152 inode = ceph_find_inode(sb, vino);
153 if (!inode)
154 return ERR_PTR(err ? err : -ESTALE);
155 }
156
157 dentry = d_obtain_alias(inode);
Dan Carpenter0d509c92010-04-21 12:31:13 +0200158 if (IS_ERR(dentry)) {
Sage Weila8e63b72009-10-06 11:31:13 -0700159 pr_err("cfh_to_dentry %llx -- inode %p but ENOMEM\n",
160 cfh->ino, inode);
161 iput(inode);
Dan Carpenter0d509c92010-04-21 12:31:13 +0200162 return dentry;
Sage Weila8e63b72009-10-06 11:31:13 -0700163 }
164 err = ceph_init_dentry(dentry);
165 if (err < 0) {
166 iput(inode);
167 return ERR_PTR(err);
168 }
169 dout("__cfh_to_dentry %llx %p dentry %p\n", cfh->ino, inode, dentry);
170 return dentry;
171}
172
173static struct dentry *ceph_fh_to_dentry(struct super_block *sb, struct fid *fid,
174 int fh_len, int fh_type)
175{
176 if (fh_type == 1)
177 return __fh_to_dentry(sb, (struct ceph_nfs_fh *)fid->raw);
178 else
179 return __cfh_to_dentry(sb, (struct ceph_nfs_confh *)fid->raw);
180}
181
182/*
183 * get parent, if possible.
184 *
185 * FIXME: we could do better by querying the mds to discover the
186 * parent.
187 */
188static struct dentry *ceph_fh_to_parent(struct super_block *sb,
189 struct fid *fid,
190 int fh_len, int fh_type)
191{
192 struct ceph_nfs_confh *cfh = (void *)fid->raw;
193 struct ceph_vino vino;
194 struct inode *inode;
195 struct dentry *dentry;
196 int err;
197
198 if (fh_type == 1)
199 return ERR_PTR(-ESTALE);
200
201 pr_debug("fh_to_parent %llx/%d\n", cfh->parent_ino,
202 cfh->parent_name_hash);
203
204 vino.ino = cfh->ino;
205 vino.snap = CEPH_NOSNAP;
206 inode = ceph_find_inode(sb, vino);
207 if (!inode)
208 return ERR_PTR(-ESTALE);
209
210 dentry = d_obtain_alias(inode);
Dan Carpenter0d509c92010-04-21 12:31:13 +0200211 if (IS_ERR(dentry)) {
Sage Weila8e63b72009-10-06 11:31:13 -0700212 pr_err("fh_to_parent %llx -- inode %p but ENOMEM\n",
213 cfh->ino, inode);
214 iput(inode);
Dan Carpenter0d509c92010-04-21 12:31:13 +0200215 return dentry;
Sage Weila8e63b72009-10-06 11:31:13 -0700216 }
217 err = ceph_init_dentry(dentry);
218 if (err < 0) {
219 iput(inode);
220 return ERR_PTR(err);
221 }
222 dout("fh_to_parent %llx %p dentry %p\n", cfh->ino, inode, dentry);
223 return dentry;
224}
225
226const struct export_operations ceph_export_ops = {
227 .encode_fh = ceph_encode_fh,
228 .fh_to_dentry = ceph_fh_to_dentry,
229 .fh_to_parent = ceph_fh_to_parent,
230};