1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
4 * Lustre Light Super operations
6 * This code is issued under the GNU General Public License.
7 * See the file COPYING in this distribution
9 * Copryright (C) 2002 Cluster File Systems, Inc.
12 #define DEBUG_SUBSYSTEM S_LLITE
14 #include <linux/module.h>
15 #include <linux/random.h>
16 #include <linux/lustre_lite.h>
17 #include <linux/lustre_ha.h>
18 #include <linux/obd_lov.h>
19 #include <linux/lustre_dlm.h>
20 #include <linux/init.h>
23 kmem_cache_t *ll_file_data_slab;
24 extern struct address_space_operations ll_aops;
25 extern struct address_space_operations ll_dir_aops;
26 struct super_operations ll_super_operations;
28 extern int ll_recover(struct recovd_data *, int);
29 extern int ll_commitcbd_setup(struct ll_sb_info *);
30 extern int ll_commitcbd_cleanup(struct ll_sb_info *);
32 static char *ll_read_opt(const char *opt, char *data)
38 CDEBUG(D_SUPER, "option: %s, data %s\n", opt, data);
39 if ( strncmp(opt, data, strlen(opt)) )
41 if ( (value = strchr(data, '=')) == NULL )
45 OBD_ALLOC(retval, strlen(value) + 1);
47 CERROR("out of memory!\n");
51 memcpy(retval, value, strlen(value)+1);
52 CDEBUG(D_SUPER, "Assigned option: %s, value %s\n", opt, retval);
56 static int ll_set_opt(const char *opt, char *data, int fl)
60 CDEBUG(D_SUPER, "option: %s, data %s\n", opt, data);
61 if ( strncmp(opt, data, strlen(opt)) )
67 static void ll_options(char *options, char **ost, char **mds, int *flags)
77 for (this_char = strtok (options, ",");
79 this_char = strtok (NULL, ",")) {
80 CDEBUG(D_SUPER, "this_char %s\n", this_char);
81 if ( (!*ost && (*ost = ll_read_opt("osc", this_char)))||
82 (!*mds && (*mds = ll_read_opt("mdc", this_char)))||
83 (!(*flags & LL_SBI_NOLCK) && ((*flags) = (*flags) |
84 ll_set_opt("nolock", this_char, LL_SBI_NOLCK))) )
91 #define log2(n) ffz(~(n))
94 static struct super_block * ll_read_super(struct super_block *sb,
95 void *data, int silent)
97 struct inode *root = 0;
98 struct obd_device *obd;
99 struct ll_sb_info *sbi;
103 struct ll_fid rootfid;
104 struct obd_statfs osfs;
105 __u64 last_committed;
107 struct ptlrpc_request *request = NULL;
108 struct ptlrpc_connection *mdc_conn;
109 struct ll_read_inode2_cookie lic;
115 OBD_ALLOC(sbi, sizeof(*sbi));
121 INIT_LIST_HEAD(&sbi->ll_conn_chain);
122 sbi->ll_mount_epoch = 0;
123 generate_random_uuid(uuid);
124 class_uuid_unparse(uuid, sbi->ll_sb_uuid);
126 sb->u.generic_sbp = sbi;
128 ll_options(data, &osc, &mdc, &sbi->ll_flags);
132 GOTO(out_free, sb = NULL);
137 GOTO(out_free, sb = NULL);
140 obd = class_uuid2obd(mdc);
142 CERROR("MDC %s: not setup or attached\n", mdc);
143 GOTO(out_free, sb = NULL);
146 err = obd_connect(&sbi->ll_mdc_conn, obd, sbi->ll_sb_uuid);
148 CERROR("cannot connect to %s: rc = %d\n", mdc, err);
149 GOTO(out_free, sb = NULL);
152 #warning Peter: is this the right place to raise the connection level?
153 mdc_conn = sbi2mdc(sbi)->cl_import.imp_connection;
154 mdc_conn->c_level = LUSTRE_CONN_FULL;
155 list_add(&mdc_conn->c_sb_chain, &sbi->ll_conn_chain);
157 obd = class_uuid2obd(osc);
159 CERROR("OSC %s: not setup or attached\n", osc);
160 GOTO(out_mdc, sb = NULL);
163 err = obd_connect(&sbi->ll_osc_conn, obd, sbi->ll_sb_uuid);
165 CERROR("cannot connect to %s: rc = %d\n", osc, err);
166 GOTO(out_mdc, sb = NULL);
169 /* XXX: need to store the last_* values somewhere */
170 err = mdc_getstatus(&sbi->ll_mdc_conn, &rootfid, &last_committed,
171 &last_xid, &request);
172 ptlrpc_req_finished(request);
174 CERROR("cannot mds_connect: rc = %d\n", err);
175 GOTO(out_mdc, sb = NULL);
177 CDEBUG(D_SUPER, "rootfid %Ld\n", (unsigned long long)rootfid.id);
178 sbi->ll_rootino = rootfid.id;
180 memset(&osfs, 0, sizeof(osfs));
182 err = mdc_statfs(&sbi->ll_mdc_conn, &osfs, &request);
183 ptlrpc_req_finished(request);
184 sb->s_blocksize = osfs.os_bsize;
185 sb->s_blocksize_bits = log2(osfs.os_bsize);
186 sb->s_magic = LL_SUPER_MAGIC;
187 sb->s_maxbytes = (1ULL << (32 + 9)) - osfs.os_bsize;
189 sb->s_op = &ll_super_operations;
191 /* make root inode */
193 err = mdc_getattr(&sbi->ll_mdc_conn, sbi->ll_rootino, S_IFDIR,
194 OBD_MD_FLNOTOBD|OBD_MD_FLBLOCKS, 0, &request);
196 CERROR("mdc_getattr failed for root: rc = %d\n", err);
197 GOTO(out_request, sb = NULL);
200 /* initialize committed transaction callback daemon */
201 spin_lock_init(&sbi->ll_commitcbd_lock);
202 init_waitqueue_head(&sbi->ll_commitcbd_waitq);
203 init_waitqueue_head(&sbi->ll_commitcbd_ctl_waitq);
204 sbi->ll_commitcbd_flags = 0;
205 err = ll_commitcbd_setup(sbi);
207 CERROR("failed to start commit callback daemon: rc = %d\n",err);
208 GOTO(out_request, sb = NULL);
211 lic.lic_body = lustre_msg_buf(request->rq_repmsg, 0);
213 root = iget4(sb, sbi->ll_rootino, NULL, &lic);
216 sb->s_root = d_alloc_root(root);
218 CERROR("lustre_lite: bad iget4 for root\n");
219 GOTO(out_cdb, sb = NULL);
222 ptlrpc_req_finished(request);
226 OBD_FREE(mdc, strlen(mdc) + 1);
228 OBD_FREE(osc, strlen(osc) + 1);
233 ll_commitcbd_cleanup(sbi);
235 ptlrpc_req_finished(request);
236 obd_disconnect(&sbi->ll_osc_conn);
238 obd_disconnect(&sbi->ll_mdc_conn);
240 OBD_FREE(sbi, sizeof(*sbi));
244 } /* ll_read_super */
246 static void ll_put_super(struct super_block *sb)
248 struct ll_sb_info *sbi = ll_s2sbi(sb);
251 list_del(&sbi->ll_conn_chain);
252 ll_commitcbd_cleanup(sbi);
253 obd_disconnect(&sbi->ll_osc_conn);
254 obd_disconnect(&sbi->ll_mdc_conn);
255 OBD_FREE(sbi, sizeof(*sbi));
261 static void ll_clear_inode(struct inode *inode)
265 if (atomic_read(&inode->i_count) == 0) {
266 struct ll_inode_info *lli = ll_i2info(inode);
267 struct lov_stripe_md *lsm = lli->lli_smd;
268 char *symlink_name = lli->lli_symlink_name;
271 OBD_FREE(lsm, ll_ost_easize(inode->i_sb));
275 OBD_FREE(symlink_name, strlen(symlink_name) + 1);
276 lli->lli_symlink_name = NULL;
281 static void ll_delete_inode(struct inode *inode)
283 if (S_ISREG(inode->i_mode)) {
286 struct lov_stripe_md *lsm = ll_i2info(inode)->lli_smd;
291 if (lsm->lsm_object_id == 0) {
292 CERROR("This really happens\n");
293 /* No obdo was ever created */
301 oa->o_id = lsm->lsm_object_id;
302 oa->o_easize = ll_mds_easize(inode->i_sb);
303 oa->o_mode = inode->i_mode;
304 oa->o_valid = OBD_MD_FLID | OBD_MD_FLEASIZE | OBD_MD_FLTYPE;
306 err = obd_destroy(ll_i2obdconn(inode), oa, lsm);
308 CDEBUG(D_SUPER, "obd destroy of objid "LPX64" error %d\n",
309 lsm->lsm_object_id, err);
315 /* like inode_setattr, but doesn't mark the inode dirty */
316 static int ll_attr2inode(struct inode * inode, struct iattr * attr, int trunc)
318 unsigned int ia_valid = attr->ia_valid;
321 if ((ia_valid & ATTR_SIZE) && trunc) {
322 error = vmtruncate(inode, attr->ia_size);
325 } else if (ia_valid & ATTR_SIZE)
326 inode->i_size = attr->ia_size;
328 if (ia_valid & ATTR_UID)
329 inode->i_uid = attr->ia_uid;
330 if (ia_valid & ATTR_GID)
331 inode->i_gid = attr->ia_gid;
332 if (ia_valid & ATTR_ATIME)
333 inode->i_atime = attr->ia_atime;
334 if (ia_valid & ATTR_MTIME)
335 inode->i_mtime = attr->ia_mtime;
336 if (ia_valid & ATTR_CTIME)
337 inode->i_ctime = attr->ia_ctime;
338 if (ia_valid & ATTR_MODE) {
339 inode->i_mode = attr->ia_mode;
340 if (!in_group_p(inode->i_gid) && !capable(CAP_FSETID))
341 inode->i_mode &= ~S_ISGID;
347 int ll_inode_setattr(struct inode *inode, struct iattr *attr, int do_trunc)
349 struct ptlrpc_request *request = NULL;
350 struct ll_sb_info *sbi = ll_i2sbi(inode);
355 /* change incore inode */
356 ll_attr2inode(inode, attr, do_trunc);
358 err = mdc_setattr(&sbi->ll_mdc_conn, inode, attr, &request);
360 CERROR("mdc_setattr fails (%d)\n", err);
362 ptlrpc_req_finished(request);
367 int ll_setattr(struct dentry *de, struct iattr *attr)
369 int rc = inode_change_ok(de->d_inode, attr);
374 return ll_inode_setattr(de->d_inode, attr, 1);
377 static int ll_statfs(struct super_block *sb, struct statfs *sfs)
379 struct ptlrpc_request *request = NULL;
380 struct ll_sb_info *sbi = ll_s2sbi(sb);
381 struct obd_statfs osfs;
385 memset(sfs, 0, sizeof(*sfs));
386 rc = mdc_statfs(&sbi->ll_mdc_conn, &osfs, &request);
387 statfs_unpack(sfs, &osfs);
388 ptlrpc_req_finished(request);
390 CERROR("mdc_statfs fails: rc = %d\n", rc);
392 CDEBUG(D_SUPER, "mdc_statfs shows blocks "LPU64"/"LPU64
393 " objects "LPU64"/"LPU64"\n",
394 osfs.os_bavail, osfs.os_blocks,
395 osfs.os_ffree, osfs.os_files);
397 /* temporary until mds_statfs returns statfs info for all OSTs */
399 rc = obd_statfs(&sbi->ll_osc_conn, &osfs);
401 CERROR("obd_statfs fails: rc = %d\n", rc);
404 CDEBUG(D_SUPER, "obd_statfs shows blocks "LPU64"/"LPU64
405 " objects "LPU64"/"LPU64"\n",
406 osfs.os_bavail, osfs.os_blocks,
407 osfs.os_ffree, osfs.os_files);
409 while (osfs.os_blocks > ~0UL) {
412 osfs.os_blocks >>= 1;
414 osfs.os_bavail >>= 1;
416 sfs->f_blocks = osfs.os_blocks;
417 sfs->f_bfree = osfs.os_bfree;
418 sfs->f_bavail = osfs.os_bavail;
419 if (osfs.os_ffree < (__u64)sfs->f_ffree)
420 sfs->f_ffree = osfs.os_ffree;
427 void ll_update_inode(struct inode *inode, struct mds_body *body)
429 if (body->valid & OBD_MD_FLID)
430 inode->i_ino = body->ino;
431 if (body->valid & OBD_MD_FLATIME)
432 inode->i_atime = body->atime;
433 if (body->valid & OBD_MD_FLMTIME)
434 inode->i_mtime = body->mtime;
435 if (body->valid & OBD_MD_FLCTIME)
436 inode->i_ctime = body->ctime;
437 if (body->valid & OBD_MD_FLMODE)
438 inode->i_mode = (inode->i_mode & S_IFMT)|(body->mode & ~S_IFMT);
439 if (body->valid & OBD_MD_FLTYPE)
440 inode->i_mode = (inode->i_mode & ~S_IFMT)|(body->mode & S_IFMT);
441 if (body->valid & OBD_MD_FLUID)
442 inode->i_uid = body->uid;
443 if (body->valid & OBD_MD_FLGID)
444 inode->i_gid = body->gid;
445 if (body->valid & OBD_MD_FLFLAGS)
446 inode->i_flags = body->flags;
447 if (body->valid & OBD_MD_FLNLINK)
448 inode->i_nlink = body->nlink;
449 if (body->valid & OBD_MD_FLGENER)
450 inode->i_generation = body->generation;
451 if (body->valid & OBD_MD_FLRDEV)
452 inode->i_rdev = body->extra;
453 if (body->valid & OBD_MD_FLSIZE)
454 inode->i_size = body->size;
457 static void ll_read_inode2(struct inode *inode, void *opaque)
459 struct ll_read_inode2_cookie *lic = opaque;
460 struct mds_body *body = lic->lic_body;
461 struct ll_inode_info *lli = ll_i2info(inode);
464 sema_init(&lli->lli_open_sem, 1);
465 lli->lli_mount_epoch = ll_i2sbi(inode)->ll_mount_epoch;
467 /* core attributes first */
468 ll_update_inode(inode, body);
470 //if (body->valid & OBD_MD_FLEASIZE)
471 if (lic && lic->lic_lmm) {
472 struct lov_mds_md *lmm = lic->lic_lmm;
475 if (lmm->lmm_easize != ll_mds_easize(inode->i_sb)) {
476 CERROR("Striping metadata size error %ld\n",
480 size = ll_ost_easize(inode->i_sb);
481 OBD_ALLOC(lli->lli_smd, size);
483 CERROR("No memory for %d\n", size);
486 lov_unpackmd(lli->lli_smd, lmm);
491 /* Get the authoritative file size */
492 if (lli->lli_smd && (inode->i_mode & S_IFREG)) {
495 rc = ll_file_size(inode, lli->lli_smd);
497 CERROR("ll_file_size: %d\n", rc);
498 /* FIXME: need to somehow prevent inode creation */
503 /* OIDEBUG(inode); */
505 if (S_ISREG(inode->i_mode)) {
506 inode->i_op = &ll_file_inode_operations;
507 inode->i_fop = &ll_file_operations;
508 inode->i_mapping->a_ops = &ll_aops;
510 } else if (S_ISDIR(inode->i_mode)) {
511 inode->i_op = &ll_dir_inode_operations;
512 inode->i_fop = &ll_dir_operations;
513 inode->i_mapping->a_ops = &ll_dir_aops;
515 } else if (S_ISLNK(inode->i_mode)) {
516 inode->i_op = &ll_fast_symlink_inode_operations;
519 init_special_inode(inode, inode->i_mode, inode->i_rdev);
524 /* exported operations */
525 struct super_operations ll_super_operations =
527 read_inode2: ll_read_inode2,
528 clear_inode: ll_clear_inode,
529 delete_inode: ll_delete_inode,
530 put_super: ll_put_super,
534 struct file_system_type lustre_lite_fs_type = {
535 "lustre_lite", 0, ll_read_super, NULL
538 static int __init init_lustre_lite(void)
540 printk(KERN_INFO "Lustre Lite 0.5.14, info@clusterfs.com\n");
541 ll_file_data_slab = kmem_cache_create("ll_file_data",
542 sizeof(struct ll_file_data), 0,
543 SLAB_HWCACHE_ALIGN, NULL, NULL);
544 if (ll_file_data_slab == NULL)
546 return register_filesystem(&lustre_lite_fs_type);
549 static void __exit exit_lustre_lite(void)
551 unregister_filesystem(&lustre_lite_fs_type);
552 kmem_cache_destroy(ll_file_data_slab);
555 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
556 MODULE_DESCRIPTION("Lustre Lite Client File System v1.0");
557 MODULE_LICENSE("GPL");
559 module_init(init_lustre_lite);
560 module_exit(exit_lustre_lite);