blob: ca3ab3f9ca70d87c6ab3a421c1e366b5b0d0fdbd [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
Sage Weilc8628682012-04-05 12:07:36 -070043/*
44 * The presence of @parent_inode here tells us whether NFS wants a
45 * connectable file handle. However, we want to make a connectionable
46 * file handle unconditionally so that the MDS gets as much of a hint
47 * as possible. That means we only use @parent_dentry to indicate
48 * whether nfsd wants a connectable fh, and whether we should indicate
49 * failure from a too-small @max_len.
50 */
51static int ceph_encode_fh(struct inode *inode, u32 *rawfh, int *max_len,
52 struct inode *parent_inode)
Sage Weila8e63b72009-10-06 11:31:13 -070053{
Aneesh Kumar K.V92923dc2010-10-05 16:03:41 +053054 int type;
Sage Weila8e63b72009-10-06 11:31:13 -070055 struct ceph_nfs_fh *fh = (void *)rawfh;
56 struct ceph_nfs_confh *cfh = (void *)rawfh;
Aneesh Kumar K.V92923dc2010-10-05 16:03:41 +053057 int connected_handle_length = sizeof(*cfh)/4;
58 int handle_length = sizeof(*fh)/4;
Cyril Roelandtf6af75d2012-12-18 14:21:23 -080059 struct dentry *dentry;
Sage Weilc8628682012-04-05 12:07:36 -070060 struct dentry *parent;
Sage Weila8e63b72009-10-06 11:31:13 -070061
62 /* don't re-export snaps */
63 if (ceph_snap(inode) != CEPH_NOSNAP)
64 return -EINVAL;
65
Cyril Roelandtf6af75d2012-12-18 14:21:23 -080066 dentry = d_find_alias(inode);
67
Sage Weilc8628682012-04-05 12:07:36 -070068 /* if we found an alias, generate a connectable fh */
69 if (*max_len >= connected_handle_length && dentry) {
Sage Weila8e63b72009-10-06 11:31:13 -070070 dout("encode_fh %p connectable\n", dentry);
Sage Weilc8628682012-04-05 12:07:36 -070071 spin_lock(&dentry->d_lock);
72 parent = dentry->d_parent;
73 cfh->ino = ceph_ino(inode);
Sage Weila8e63b72009-10-06 11:31:13 -070074 cfh->parent_ino = ceph_ino(parent->d_inode);
Sage Weile5f86dc2011-07-26 11:30:55 -070075 cfh->parent_name_hash = ceph_dentry_hash(parent->d_inode,
76 dentry);
Aneesh Kumar K.V92923dc2010-10-05 16:03:41 +053077 *max_len = connected_handle_length;
Sage Weila8e63b72009-10-06 11:31:13 -070078 type = 2;
Sage Weilc8628682012-04-05 12:07:36 -070079 spin_unlock(&dentry->d_lock);
Aneesh Kumar K.V92923dc2010-10-05 16:03:41 +053080 } else if (*max_len >= handle_length) {
Sage Weilc8628682012-04-05 12:07:36 -070081 if (parent_inode) {
82 /* nfsd wants connectable */
Aneesh Kumar K.Vbba0cd02010-10-05 16:03:42 +053083 *max_len = connected_handle_length;
Sage Weile5f86dc2011-07-26 11:30:55 -070084 type = 255;
85 } else {
86 dout("encode_fh %p\n", dentry);
Sage Weilc8628682012-04-05 12:07:36 -070087 fh->ino = ceph_ino(inode);
Sage Weile5f86dc2011-07-26 11:30:55 -070088 *max_len = handle_length;
89 type = 1;
Aneesh Kumar K.Vbba0cd02010-10-05 16:03:42 +053090 }
Sage Weila8e63b72009-10-06 11:31:13 -070091 } else {
Aneesh Kumar K.Vbba0cd02010-10-05 16:03:42 +053092 *max_len = handle_length;
Sage Weile5f86dc2011-07-26 11:30:55 -070093 type = 255;
Sage Weila8e63b72009-10-06 11:31:13 -070094 }
David Zafman52eb5a92012-10-18 14:01:43 -070095 if (dentry)
96 dput(dentry);
Sage Weila8e63b72009-10-06 11:31:13 -070097 return type;
98}
99
100/*
101 * convert regular fh to dentry
102 *
103 * FIXME: we should try harder by querying the mds for the ino.
104 */
105static struct dentry *__fh_to_dentry(struct super_block *sb,
Hugh Dickins35c2a7f2012-10-07 20:32:51 -0700106 struct ceph_nfs_fh *fh, int fh_len)
Sage Weila8e63b72009-10-06 11:31:13 -0700107{
Sage Weil3c454cf2011-04-06 09:31:40 -0700108 struct ceph_mds_client *mdsc = ceph_sb_to_client(sb)->mdsc;
Sage Weila8e63b72009-10-06 11:31:13 -0700109 struct inode *inode;
110 struct dentry *dentry;
111 struct ceph_vino vino;
112 int err;
113
Hugh Dickins35c2a7f2012-10-07 20:32:51 -0700114 if (fh_len < sizeof(*fh) / 4)
115 return ERR_PTR(-ESTALE);
116
Sage Weila8e63b72009-10-06 11:31:13 -0700117 dout("__fh_to_dentry %llx\n", fh->ino);
118 vino.ino = fh->ino;
119 vino.snap = CEPH_NOSNAP;
120 inode = ceph_find_inode(sb, vino);
Sage Weil3c454cf2011-04-06 09:31:40 -0700121 if (!inode) {
122 struct ceph_mds_request *req;
123
124 req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_LOOKUPINO,
125 USE_ANY_MDS);
126 if (IS_ERR(req))
127 return ERR_CAST(req);
128
129 req->r_ino1 = vino;
130 req->r_num_caps = 1;
131 err = ceph_mdsc_do_request(mdsc, NULL, req);
Sage Weil45e3d3e2011-04-06 09:35:00 -0700132 inode = req->r_target_inode;
133 if (inode)
Sage Weil70b666c2011-05-27 09:24:26 -0700134 ihold(inode);
Sage Weil3c454cf2011-04-06 09:31:40 -0700135 ceph_mdsc_put_request(req);
Sage Weil3c454cf2011-04-06 09:31:40 -0700136 if (!inode)
137 return ERR_PTR(-ESTALE);
138 }
Sage Weila8e63b72009-10-06 11:31:13 -0700139
140 dentry = d_obtain_alias(inode);
Dan Carpenter0d509c92010-04-21 12:31:13 +0200141 if (IS_ERR(dentry)) {
Sage Weila8e63b72009-10-06 11:31:13 -0700142 pr_err("fh_to_dentry %llx -- inode %p but ENOMEM\n",
143 fh->ino, inode);
144 iput(inode);
Dan Carpenter0d509c92010-04-21 12:31:13 +0200145 return dentry;
Sage Weila8e63b72009-10-06 11:31:13 -0700146 }
147 err = ceph_init_dentry(dentry);
Sage Weila8e63b72009-10-06 11:31:13 -0700148 if (err < 0) {
149 iput(inode);
150 return ERR_PTR(err);
151 }
152 dout("__fh_to_dentry %llx %p dentry %p\n", fh->ino, inode, dentry);
153 return dentry;
154}
155
156/*
157 * convert connectable fh to dentry
158 */
159static struct dentry *__cfh_to_dentry(struct super_block *sb,
Hugh Dickins35c2a7f2012-10-07 20:32:51 -0700160 struct ceph_nfs_confh *cfh, int fh_len)
Sage Weila8e63b72009-10-06 11:31:13 -0700161{
Yehuda Sadeh3d14c5d2010-04-06 15:14:15 -0700162 struct ceph_mds_client *mdsc = ceph_sb_to_client(sb)->mdsc;
Sage Weila8e63b72009-10-06 11:31:13 -0700163 struct inode *inode;
164 struct dentry *dentry;
165 struct ceph_vino vino;
166 int err;
167
Hugh Dickins35c2a7f2012-10-07 20:32:51 -0700168 if (fh_len < sizeof(*cfh) / 4)
169 return ERR_PTR(-ESTALE);
170
Sage Weila8e63b72009-10-06 11:31:13 -0700171 dout("__cfh_to_dentry %llx (%llx/%x)\n",
172 cfh->ino, cfh->parent_ino, cfh->parent_name_hash);
173
174 vino.ino = cfh->ino;
175 vino.snap = CEPH_NOSNAP;
176 inode = ceph_find_inode(sb, vino);
177 if (!inode) {
178 struct ceph_mds_request *req;
179
180 req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_LOOKUPHASH,
181 USE_ANY_MDS);
182 if (IS_ERR(req))
Julia Lawall7e34bc52010-05-22 12:01:14 +0200183 return ERR_CAST(req);
Sage Weila8e63b72009-10-06 11:31:13 -0700184
185 req->r_ino1 = vino;
186 req->r_ino2.ino = cfh->parent_ino;
187 req->r_ino2.snap = CEPH_NOSNAP;
188 req->r_path2 = kmalloc(16, GFP_NOFS);
189 snprintf(req->r_path2, 16, "%d", cfh->parent_name_hash);
190 req->r_num_caps = 1;
191 err = ceph_mdsc_do_request(mdsc, NULL, req);
Sage Weil45e3d3e2011-04-06 09:35:00 -0700192 inode = req->r_target_inode;
193 if (inode)
Sage Weil70b666c2011-05-27 09:24:26 -0700194 ihold(inode);
Sage Weila8e63b72009-10-06 11:31:13 -0700195 ceph_mdsc_put_request(req);
Sage Weila8e63b72009-10-06 11:31:13 -0700196 if (!inode)
197 return ERR_PTR(err ? err : -ESTALE);
198 }
199
200 dentry = d_obtain_alias(inode);
Dan Carpenter0d509c92010-04-21 12:31:13 +0200201 if (IS_ERR(dentry)) {
Sage Weila8e63b72009-10-06 11:31:13 -0700202 pr_err("cfh_to_dentry %llx -- inode %p but ENOMEM\n",
203 cfh->ino, inode);
204 iput(inode);
Dan Carpenter0d509c92010-04-21 12:31:13 +0200205 return dentry;
Sage Weila8e63b72009-10-06 11:31:13 -0700206 }
207 err = ceph_init_dentry(dentry);
208 if (err < 0) {
209 iput(inode);
210 return ERR_PTR(err);
211 }
212 dout("__cfh_to_dentry %llx %p dentry %p\n", cfh->ino, inode, dentry);
213 return dentry;
214}
215
216static struct dentry *ceph_fh_to_dentry(struct super_block *sb, struct fid *fid,
217 int fh_len, int fh_type)
218{
219 if (fh_type == 1)
Hugh Dickins35c2a7f2012-10-07 20:32:51 -0700220 return __fh_to_dentry(sb, (struct ceph_nfs_fh *)fid->raw,
221 fh_len);
Sage Weila8e63b72009-10-06 11:31:13 -0700222 else
Hugh Dickins35c2a7f2012-10-07 20:32:51 -0700223 return __cfh_to_dentry(sb, (struct ceph_nfs_confh *)fid->raw,
224 fh_len);
Sage Weila8e63b72009-10-06 11:31:13 -0700225}
226
227/*
228 * get parent, if possible.
229 *
230 * FIXME: we could do better by querying the mds to discover the
231 * parent.
232 */
233static struct dentry *ceph_fh_to_parent(struct super_block *sb,
234 struct fid *fid,
235 int fh_len, int fh_type)
236{
237 struct ceph_nfs_confh *cfh = (void *)fid->raw;
238 struct ceph_vino vino;
239 struct inode *inode;
240 struct dentry *dentry;
241 int err;
242
243 if (fh_type == 1)
244 return ERR_PTR(-ESTALE);
Hugh Dickins35c2a7f2012-10-07 20:32:51 -0700245 if (fh_len < sizeof(*cfh) / 4)
246 return ERR_PTR(-ESTALE);
Sage Weila8e63b72009-10-06 11:31:13 -0700247
248 pr_debug("fh_to_parent %llx/%d\n", cfh->parent_ino,
249 cfh->parent_name_hash);
250
251 vino.ino = cfh->ino;
252 vino.snap = CEPH_NOSNAP;
253 inode = ceph_find_inode(sb, vino);
254 if (!inode)
255 return ERR_PTR(-ESTALE);
256
257 dentry = d_obtain_alias(inode);
Dan Carpenter0d509c92010-04-21 12:31:13 +0200258 if (IS_ERR(dentry)) {
Sage Weila8e63b72009-10-06 11:31:13 -0700259 pr_err("fh_to_parent %llx -- inode %p but ENOMEM\n",
260 cfh->ino, inode);
261 iput(inode);
Dan Carpenter0d509c92010-04-21 12:31:13 +0200262 return dentry;
Sage Weila8e63b72009-10-06 11:31:13 -0700263 }
264 err = ceph_init_dentry(dentry);
265 if (err < 0) {
266 iput(inode);
267 return ERR_PTR(err);
268 }
269 dout("fh_to_parent %llx %p dentry %p\n", cfh->ino, inode, dentry);
270 return dentry;
271}
272
273const struct export_operations ceph_export_ops = {
274 .encode_fh = ceph_encode_fh,
Sage Weila8e63b72009-10-06 11:31:13 -0700275 .fh_to_dentry = ceph_fh_to_dentry,
276 .fh_to_parent = ceph_fh_to_parent,
277};