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>
22 kmem_cache_t *ll_file_data_slab;
23 extern struct address_space_operations ll_aops;
24 extern struct address_space_operations ll_dir_aops;
25 struct super_operations ll_super_operations;
27 extern int ll_recover(struct recovd_data *, int);
28 extern int ll_commitcbd_setup(struct ll_sb_info *);
29 extern int ll_commitcbd_cleanup(struct ll_sb_info *);
31 static char *ll_read_opt(const char *opt, char *data)
37 CDEBUG(D_SUPER, "option: %s, data %s\n", opt, data);
38 if ( strncmp(opt, data, strlen(opt)) )
40 if ( (value = strchr(data, '=')) == NULL )
44 OBD_ALLOC(retval, strlen(value) + 1);
46 CERROR("out of memory!\n");
50 memcpy(retval, value, strlen(value)+1);
51 CDEBUG(D_SUPER, "Assigned option: %s, value %s\n", opt, retval);
55 static int ll_set_opt(const char *opt, char *data, int fl)
59 CDEBUG(D_SUPER, "option: %s, data %s\n", opt, data);
60 if ( strncmp(opt, data, strlen(opt)) )
66 static void ll_options(char *options, char **ost, char **mds, int *flags)
76 for (this_char = strtok (options, ",");
78 this_char = strtok (NULL, ",")) {
79 CDEBUG(D_SUPER, "this_char %s\n", this_char);
80 if ( (!*ost && (*ost = ll_read_opt("osc", this_char)))||
81 (!*mds && (*mds = ll_read_opt("mdc", this_char)))||
82 (!(*flags & LL_SBI_NOLCK) && ((*flags) = (*flags) |
83 ll_set_opt("nolock", this_char, LL_SBI_NOLCK))) )
90 #define log2(n) ffz(~(n))
93 static struct super_block * ll_read_super(struct super_block *sb,
94 void *data, int silent)
96 struct inode *root = 0;
97 struct obd_device *obd;
98 struct ll_sb_info *sbi;
102 struct ll_fid rootfid;
103 struct obd_statfs osfs;
104 __u64 last_committed;
106 struct ptlrpc_request *request = NULL;
107 struct ll_read_inode2_cookie lic;
113 OBD_ALLOC(sbi, sizeof(*sbi));
119 generate_random_uuid(uuid);
120 class_uuid_unparse(uuid, sbi->ll_sb_uuid);
122 sb->u.generic_sbp = sbi;
124 ll_options(data, &osc, &mdc, &sbi->ll_flags);
128 GOTO(out_free, sb = NULL);
133 GOTO(out_free, sb = NULL);
136 obd = class_uuid2obd(mdc);
138 CERROR("MDC %s: not setup or attached\n", mdc);
139 GOTO(out_free, sb = NULL);
142 err = obd_connect(&sbi->ll_mdc_conn, obd, sbi->ll_sb_uuid);
144 CERROR("cannot connect to %s: rc = %d\n", mdc, err);
145 GOTO(out_free, sb = NULL);
148 #warning Peter: is this the right place to raise the connection level?
149 sbi2mdc(sbi)->cl_import.imp_connection->c_level = LUSTRE_CONN_FULL;
151 obd = class_uuid2obd(osc);
153 CERROR("OSC %s: not setup or attached\n", osc);
154 GOTO(out_mdc, sb = NULL);
156 err = obd_connect(&sbi->ll_osc_conn, obd, sbi->ll_sb_uuid);
158 CERROR("cannot connect to %s: rc = %d\n", osc, err);
159 GOTO(out_mdc, sb = NULL);
162 /* XXX: need to store the last_* values somewhere */
163 err = mdc_getstatus(&sbi->ll_mdc_conn, &rootfid, &last_committed,
164 &last_xid, &request);
165 ptlrpc_req_finished(request);
167 CERROR("cannot mds_connect: rc = %d\n", err);
168 GOTO(out_mdc, sb = NULL);
170 CDEBUG(D_SUPER, "rootfid %Ld\n", (unsigned long long)rootfid.id);
171 sbi->ll_rootino = rootfid.id;
173 memset(&osfs, 0, sizeof(osfs));
175 err = mdc_statfs(&sbi->ll_mdc_conn, &osfs, &request);
176 ptlrpc_req_finished(request);
177 sb->s_blocksize = osfs.os_bsize;
178 sb->s_blocksize_bits = log2(osfs.os_bsize);
179 sb->s_magic = LL_SUPER_MAGIC;
180 sb->s_maxbytes = (1ULL << (32 + 9)) - osfs.os_bsize;
182 sb->s_op = &ll_super_operations;
184 /* make root inode */
186 err = mdc_getattr(&sbi->ll_mdc_conn, sbi->ll_rootino, S_IFDIR,
187 OBD_MD_FLNOTOBD|OBD_MD_FLBLOCKS, 0, &request);
189 CERROR("mdc_getattr failed for root: rc = %d\n", err);
190 GOTO(out_request, sb = NULL);
193 /* initialize committed transaction callback daemon */
194 spin_lock_init(&sbi->ll_commitcbd_lock);
195 init_waitqueue_head(&sbi->ll_commitcbd_waitq);
196 init_waitqueue_head(&sbi->ll_commitcbd_ctl_waitq);
197 sbi->ll_commitcbd_flags = 0;
198 err = ll_commitcbd_setup(sbi);
200 CERROR("failed to start commit callback daemon: rc = %d\n",err);
201 GOTO(out_request, sb = NULL);
204 lic.lic_body = lustre_msg_buf(request->rq_repmsg, 0);
206 root = iget4(sb, sbi->ll_rootino, NULL, &lic);
209 sb->s_root = d_alloc_root(root);
211 CERROR("lustre_lite: bad iget4 for root\n");
212 GOTO(out_cdb, sb = NULL);
215 ptlrpc_req_finished(request);
219 OBD_FREE(mdc, strlen(mdc) + 1);
221 OBD_FREE(osc, strlen(osc) + 1);
226 ll_commitcbd_cleanup(sbi);
228 ptlrpc_req_finished(request);
229 obd_disconnect(&sbi->ll_osc_conn);
231 obd_disconnect(&sbi->ll_mdc_conn);
233 OBD_FREE(sbi, sizeof(*sbi));
237 } /* ll_read_super */
239 static void ll_put_super(struct super_block *sb)
241 struct ll_sb_info *sbi = ll_s2sbi(sb);
243 ll_commitcbd_cleanup(sbi);
244 obd_disconnect(&sbi->ll_osc_conn);
245 obd_disconnect(&sbi->ll_mdc_conn);
246 OBD_FREE(sbi, sizeof(*sbi));
252 static void ll_clear_inode(struct inode *inode)
254 if (atomic_read(&inode->i_count) == 0) {
255 struct ll_inode_info *lli = ll_i2info(inode);
256 struct lov_stripe_md *lsm = lli->lli_smd;
257 char *symlink_name = lli->lli_symlink_name;
260 OBD_FREE(lsm, ll_ost_easize(inode->i_sb));
264 OBD_FREE(symlink_name, strlen(symlink_name) + 1);
265 lli->lli_symlink_name = NULL;
270 static void ll_delete_inode(struct inode *inode)
272 if (S_ISREG(inode->i_mode)) {
275 struct lov_stripe_md *lsm = ll_i2info(inode)->lli_smd;
280 if (lsm->lsm_object_id == 0) {
281 CERROR("This really happens\n");
282 /* No obdo was ever created */
290 oa->o_id = lsm->lsm_object_id;
291 oa->o_easize = ll_mds_easize(inode->i_sb);
292 oa->o_mode = inode->i_mode;
293 oa->o_valid = OBD_MD_FLID | OBD_MD_FLEASIZE | OBD_MD_FLMODE;
295 err = obd_destroy(ll_i2obdconn(inode), oa, lsm);
297 CDEBUG(D_SUPER, "obd destroy of objid "LPX64" error %d\n",
298 lsm->lsm_object_id, err);
304 /* like inode_setattr, but doesn't mark the inode dirty */
305 static int ll_attr2inode(struct inode * inode, struct iattr * attr, int trunc)
307 unsigned int ia_valid = attr->ia_valid;
310 if ((ia_valid & ATTR_SIZE) && trunc) {
311 error = vmtruncate(inode, attr->ia_size);
314 } else if (ia_valid & ATTR_SIZE)
315 inode->i_size = attr->ia_size;
317 if (ia_valid & ATTR_UID)
318 inode->i_uid = attr->ia_uid;
319 if (ia_valid & ATTR_GID)
320 inode->i_gid = attr->ia_gid;
321 if (ia_valid & ATTR_ATIME)
322 inode->i_atime = attr->ia_atime;
323 if (ia_valid & ATTR_MTIME)
324 inode->i_mtime = attr->ia_mtime;
325 if (ia_valid & ATTR_CTIME)
326 inode->i_ctime = attr->ia_ctime;
327 if (ia_valid & ATTR_MODE) {
328 inode->i_mode = attr->ia_mode;
329 if (!in_group_p(inode->i_gid) && !capable(CAP_FSETID))
330 inode->i_mode &= ~S_ISGID;
336 int ll_inode_setattr(struct inode *inode, struct iattr *attr, int do_trunc)
338 struct ptlrpc_request *request = NULL;
339 struct ll_sb_info *sbi = ll_i2sbi(inode);
344 /* change incore inode */
345 ll_attr2inode(inode, attr, do_trunc);
347 err = mdc_setattr(&sbi->ll_mdc_conn, inode, attr, &request);
349 CERROR("mdc_setattr fails (%d)\n", err);
351 ptlrpc_req_finished(request);
356 int ll_setattr(struct dentry *de, struct iattr *attr)
358 int rc = inode_change_ok(de->d_inode, attr);
363 return ll_inode_setattr(de->d_inode, attr, 1);
366 static int ll_statfs(struct super_block *sb, struct statfs *sfs)
368 struct ptlrpc_request *request = NULL;
369 struct ll_sb_info *sbi = ll_s2sbi(sb);
370 struct obd_statfs osfs;
374 memset(sfs, 0, sizeof(*sfs));
375 rc = mdc_statfs(&sbi->ll_mdc_conn, &osfs, &request);
376 statfs_unpack(sfs, &osfs);
377 ptlrpc_req_finished(request);
379 CERROR("mdc_statfs fails: rc = %d\n", rc);
381 CDEBUG(D_SUPER, "mdc_statfs shows blocks "LPU64"/"LPU64
382 " objects "LPU64"/"LPU64"\n",
383 osfs.os_bavail, osfs.os_blocks,
384 osfs.os_ffree, osfs.os_files);
386 /* temporary until mds_statfs returns statfs info for all OSTs */
388 rc = obd_statfs(&sbi->ll_osc_conn, &osfs);
390 CERROR("obd_statfs fails: rc = %d\n", rc);
393 CDEBUG(D_SUPER, "obd_statfs shows blocks "LPU64"/"LPU64
394 " objects "LPU64"/"LPU64"\n",
395 osfs.os_bavail, osfs.os_blocks,
396 osfs.os_ffree, osfs.os_files);
398 while (osfs.os_blocks > ~0UL) {
401 osfs.os_blocks >>= 1;
403 osfs.os_bavail >>= 1;
405 sfs->f_blocks = osfs.os_blocks;
406 sfs->f_bfree = osfs.os_bfree;
407 sfs->f_bavail = osfs.os_bavail;
408 if (osfs.os_ffree < (__u64)sfs->f_ffree)
409 sfs->f_ffree = osfs.os_ffree;
416 static void ll_read_inode2(struct inode *inode, void *opaque)
418 struct ll_read_inode2_cookie *lic = opaque;
419 struct mds_body *body = lic->lic_body;
420 struct ll_inode_info *lli = ll_i2info(inode);
423 sema_init(&lli->lli_open_sem, 1);
425 /* core attributes first */
426 if (body->valid & OBD_MD_FLID)
427 inode->i_ino = body->ino;
428 if (body->valid & OBD_MD_FLATIME)
429 inode->i_atime = body->atime;
430 if (body->valid & OBD_MD_FLMTIME)
431 inode->i_mtime = body->mtime;
432 if (body->valid & OBD_MD_FLCTIME)
433 inode->i_ctime = body->ctime;
434 if (body->valid & OBD_MD_FLMODE)
435 inode->i_mode = body->mode;
436 if (body->valid & OBD_MD_FLUID)
437 inode->i_uid = body->uid;
438 if (body->valid & OBD_MD_FLGID)
439 inode->i_gid = body->gid;
440 if (body->valid & OBD_MD_FLFLAGS)
441 inode->i_flags = body->flags;
442 if (body->valid & OBD_MD_FLNLINK)
443 inode->i_nlink = body->nlink;
444 if (body->valid & OBD_MD_FLGENER)
445 inode->i_generation = body->generation;
446 if (body->valid & OBD_MD_FLRDEV)
447 inode->i_rdev = body->extra;
448 if (body->valid & OBD_MD_FLSIZE)
449 inode->i_size = body->size;
451 //if (body->valid & OBD_MD_FLEASIZE)
452 if (lic && lic->lic_lmm) {
453 struct lov_mds_md *lmm = lic->lic_lmm;
456 if (lmm->lmm_easize != ll_mds_easize(inode->i_sb)) {
457 CERROR("Striping metadata size error %ld\n",
461 size = ll_ost_easize(inode->i_sb);
462 OBD_ALLOC(lli->lli_smd, size);
464 CERROR("No memory for %d\n", size);
467 lov_unpackmd(lli->lli_smd, lmm);
472 /* Get the authoritative file size */
473 if (lli->lli_smd && (inode->i_mode & S_IFREG)) {
476 rc = ll_file_size(inode, lli->lli_smd);
478 CERROR("ll_file_size: %d\n", rc);
479 /* FIXME: need to somehow prevent inode creation */
484 /* OIDEBUG(inode); */
486 if (S_ISREG(inode->i_mode)) {
487 inode->i_op = &ll_file_inode_operations;
488 inode->i_fop = &ll_file_operations;
489 inode->i_mapping->a_ops = &ll_aops;
491 } else if (S_ISDIR(inode->i_mode)) {
492 inode->i_op = &ll_dir_inode_operations;
493 inode->i_fop = &ll_dir_operations;
494 inode->i_mapping->a_ops = &ll_dir_aops;
496 } else if (S_ISLNK(inode->i_mode)) {
497 inode->i_op = &ll_fast_symlink_inode_operations;
500 init_special_inode(inode, inode->i_mode, inode->i_rdev);
507 /* exported operations */
508 struct super_operations ll_super_operations =
510 read_inode2: ll_read_inode2,
511 clear_inode: ll_clear_inode,
512 delete_inode: ll_delete_inode,
513 put_super: ll_put_super,
517 struct file_system_type lustre_lite_fs_type = {
518 "lustre_lite", 0, ll_read_super, NULL
521 static int __init init_lustre_lite(void)
523 printk(KERN_INFO "Lustre Lite 0.0.1, info@clusterfs.com\n");
524 ll_file_data_slab = kmem_cache_create("ll_file_data",
525 sizeof(struct ll_file_data), 0,
526 SLAB_HWCACHE_ALIGN, NULL, NULL);
527 if (ll_file_data_slab == NULL)
529 return register_filesystem(&lustre_lite_fs_type);
532 static void __exit exit_lustre_lite(void)
534 unregister_filesystem(&lustre_lite_fs_type);
535 kmem_cache_destroy(ll_file_data_slab);
538 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
539 MODULE_DESCRIPTION("Lustre Lite Client File System v1.0");
540 MODULE_LICENSE("GPL");
542 module_init(init_lustre_lite);
543 module_exit(exit_lustre_lite);