lustre/llite/super.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * Lustre Light Super operations
   5  *
   6  * This code is issued under the GNU General Public License.
   7  * See the file COPYING in this distribution
   8  *
   9  * Copryright (C) 2002 Cluster File Systems, Inc.
  10  */
  11
  12 #define DEBUG_SUBSYSTEM S_LLITE
  13
  14 #include <linux/module.h>
  15 #include <linux/lustre_lite.h>
  16 #include <linux/lustre_ha.h>
  17 #include <linux/lustre_dlm.h>
  18 #include <linux/init.h>
  19
  20 kmem_cache_t *ll_file_data_slab;
  21 extern struct address_space_operations ll_aops;
  22 extern struct address_space_operations ll_dir_aops;
  23 struct super_operations ll_super_operations;
  24
  25 extern int ll_recover(struct ptlrpc_client *);
  26 extern int ll_commitcbd_setup(struct ll_sb_info *);
  27 extern int ll_commitcbd_cleanup(struct ll_sb_info *);
  28
  29 static char *ll_read_opt(const char *opt, char *data)
  30 {
  31         char *value;
  32         char *retval;
  33         ENTRY;
  34
  35         CDEBUG(D_SUPER, "option: %s, data %s\n", opt, data);
  36         if ( strncmp(opt, data, strlen(opt)) )
  37                 RETURN(NULL);
  38         if ( (value = strchr(data, '=')) == NULL )
  39                 RETURN(NULL);
  40
  41         value++;
  42         OBD_ALLOC(retval, strlen(value) + 1);
  43         if ( !retval ) {
  44                 CERROR("out of memory!\n");
  45                 RETURN(NULL);
  46         }
  47
  48         memcpy(retval, value, strlen(value)+1);
  49         CDEBUG(D_SUPER, "Assigned option: %s, value %s\n", opt, retval);
  50         RETURN(retval);
  51 }
  52
  53 static void ll_options(char *options, char **ost, char **mds)
  54 {
  55         char *this_char;
  56         ENTRY;
  57
  58         if (!options) {
  59                 EXIT;
  60                 return;
  61         }
  62
  63         for (this_char = strtok (options, ",");
  64              this_char != NULL;
  65              this_char = strtok (NULL, ",")) {
  66                 CDEBUG(D_SUPER, "this_char %s\n", this_char);
  67                 if ( (!*ost && (*ost = ll_read_opt("ost", this_char)))||
  68                      (!*mds && (*mds = ll_read_opt("mds", this_char))) )
  69                         continue;
  70         }
  71         EXIT;
  72 }
  73
  74 #ifndef log2
  75 #define log2(n) ffz(~(n))
  76 #endif
  77
  78 static struct super_block * ll_read_super(struct super_block *sb,
  79                                           void *data, int silent)
  80 {
  81         struct inode *root = 0;
  82         struct obd_device *obd;
  83         struct ll_sb_info *sbi;
  84         char *ost = NULL;
  85         char *mds = NULL;
  86         int err;
  87         struct ll_fid rootfid;
  88         struct statfs sfs;
  89         __u64 last_committed, last_rcvd;
  90         __u32 last_xid;
  91         struct ptlrpc_request *request = NULL;
  92         struct ll_inode_md md;
  93
  94         ENTRY;
  95         MOD_INC_USE_COUNT;
  96
  97         OBD_ALLOC(sbi, sizeof(*sbi));
  98         if (!sbi) {
  99                 MOD_DEC_USE_COUNT;
 100                 RETURN(NULL);
 101         }
 102
 103         sb->u.generic_sbp = sbi;
 104
 105         ll_options(data, &ost, &mds);
 106
 107         if (!ost) {
 108                 CERROR("no ost\n");
 109                 GOTO(out_free, sb = NULL);
 110         }
 111
 112         if (!mds) {
 113                 CERROR("no mds\n");
 114                 GOTO(out_free, sb = NULL);
 115         }
 116
 117         obd = class_uuid2obd(mds);
 118         if (!obd) {
 119                 CERROR("MDS %s: not setup or attached\n", mds);
 120                 GOTO(out_free, sb = NULL);
 121         }
 122
 123 #if 0
 124         err = connmgr_connect(ptlrpc_connmgr, sbi->ll_mds_conn);
 125         if (err) {
 126                 CERROR("cannot connect to MDS: rc = %d\n", err);
 127                 GOTO(out_rpc, sb = NULL);
 128         }
 129 #endif
 130
 131         err = obd_connect(&sbi->ll_mdc_conn, obd);
 132         if (err) {
 133                 CERROR("cannot connect to %s: rc = %d\n", mds, err);
 134                 GOTO(out_free, sb = NULL);
 135         }
 136         sbi2mdc(sbi)->mdc_conn->c_level = LUSTRE_CONN_FULL;
 137
 138         obd = class_uuid2obd(ost);
 139         if (!obd) {
 140                 CERROR("OST %s: not setup or attached\n", ost);
 141                 GOTO(out_mdc, sb = NULL);
 142         }
 143         err = obd_connect(&sbi->ll_osc_conn, obd);
 144         if (err) {
 145                 CERROR("cannot connect to %s: rc = %d\n", ost, err);
 146                 GOTO(out_mdc, sb = NULL);
 147         }
 148
 149         /* XXX: need to store the last_* values somewhere */
 150         err = mdc_getstatus(&sbi->ll_mdc_conn, &rootfid, &last_committed,
 151                             &last_rcvd, &last_xid, &request);
 152         if (err) {
 153                 CERROR("cannot mds_connect: rc = %d\n", err);
 154                 GOTO(out_request, sb = NULL);
 155         }
 156         CDEBUG(D_SUPER, "rootfid %Ld\n", (unsigned long long)rootfid.id);
 157         sbi->ll_rootino = rootfid.id;
 158
 159         memset(&sfs, 0, sizeof(sfs));
 160         err = mdc_statfs(&sbi->ll_mdc_conn, &sfs, &request);
 161         sb->s_blocksize = sfs.f_bsize;
 162         sb->s_blocksize_bits = log2(sfs.f_bsize);
 163         sb->s_magic = LL_SUPER_MAGIC;
 164         sb->s_maxbytes = (1ULL << (32 + 9)) - sfs.f_bsize;
 165         ptlrpc_req_finished(request);
 166
 167         sb->s_op = &ll_super_operations;
 168
 169         /* make root inode */
 170         err = mdc_getattr(&sbi->ll_mdc_conn, sbi->ll_rootino, S_IFDIR,
 171                           OBD_MD_FLNOTOBD|OBD_MD_FLBLOCKS, 0, &request);
 172         if (err) {
 173                 CERROR("mdc_getattr failed for root: rc = %d\n", err);
 174                 GOTO(out_request, sb = NULL);
 175         }
 176
 177         /* initialize committed transaction callback daemon */
 178         spin_lock_init(&sbi->ll_commitcbd_lock);
 179         init_waitqueue_head(&sbi->ll_commitcbd_waitq);
 180         init_waitqueue_head(&sbi->ll_commitcbd_ctl_waitq);
 181         sbi->ll_commitcbd_flags = 0;
 182         err = ll_commitcbd_setup(sbi);
 183         if (err) {
 184                 CERROR("failed to start commit callback daemon: rc = %d\n",err);
 185                 GOTO(out_request, sb = NULL);
 186         }
 187
 188         md.body = lustre_msg_buf(request->rq_repmsg, 0);
 189         md.md = NULL;
 190         root = iget4(sb, sbi->ll_rootino, NULL, &md);
 191
 192         if (root) {
 193                 sb->s_root = d_alloc_root(root);
 194         } else {
 195                 CERROR("lustre_lite: bad iget4 for root\n");
 196                 GOTO(out_cdb, sb = NULL);
 197         }
 198
 199         ptlrpc_free_req(request);
 200
 201 out_dev:
 202         if (mds)
 203                 OBD_FREE(mds, strlen(mds) + 1);
 204         if (ost)
 205                 OBD_FREE(ost, strlen(ost) + 1);
 206
 207         RETURN(sb);
 208
 209 out_cdb:
 210         ll_commitcbd_cleanup(sbi);
 211 out_request:
 212         ptlrpc_free_req(request);
 213         obd_disconnect(&sbi->ll_osc_conn);
 214 out_mdc:
 215         obd_disconnect(&sbi->ll_mdc_conn);
 216 out_free:
 217         OBD_FREE(sbi, sizeof(*sbi));
 218
 219         MOD_DEC_USE_COUNT;
 220         goto out_dev;
 221 } /* ll_read_super */
 222
 223 static void ll_put_super(struct super_block *sb)
 224 {
 225         struct ll_sb_info *sbi = ll_s2sbi(sb);
 226         ENTRY;
 227         ll_commitcbd_cleanup(sbi);
 228         obd_disconnect(&sbi->ll_osc_conn);
 229         obd_disconnect(&sbi->ll_mdc_conn);
 230         OBD_FREE(sbi, sizeof(*sbi));
 231
 232         MOD_DEC_USE_COUNT;
 233         EXIT;
 234 } /* ll_put_super */
 235
 236 static void ll_clear_inode(struct inode *inode)
 237 {
 238         if (atomic_read(&inode->i_count) == 0) {
 239                 struct ll_inode_info *lli = ll_i2info(inode);
 240                 struct lov_stripe_md *md = lli->lli_smd;
 241
 242                 if (md) {
 243                         OBD_FREE(md, md->lmd_size);
 244                         lli->lli_smd = NULL;
 245                 }
 246                 if (lli->lli_symlink_name) {
 247                         OBD_FREE(lli->lli_symlink_name,
 248                                  strlen(lli->lli_symlink_name) + 1);
 249                         lli->lli_symlink_name = NULL;
 250                 }
 251         }
 252 }
 253
 254 static void ll_delete_inode(struct inode *inode)
 255 {
 256         if (S_ISREG(inode->i_mode)) {
 257                 int err;
 258                 struct obdo oa;
 259                 struct lov_stripe_md *md = ll_i2info(inode)->lli_smd;
 260
 261                if (!md)
 262                         GOTO(out, -EINVAL);
 263
 264                 oa.o_id = md->lmd_object_id;
 265                 oa.o_easize = md->lmd_size;
 266                 if (oa.o_id == 0) {
 267                         CERROR("This really happens\n");
 268                         /* No obdo was ever created */
 269                         GOTO(out, 0);
 270                 }
 271
 272                 err = obd_destroy(ll_i2obdconn(inode), &oa, md);
 273                 CDEBUG(D_SUPER, "obd destroy of %Ld error %d\n",
 274                        md->lmd_object_id, err);
 275         }
 276 out:
 277         clear_inode(inode);
 278 }
 279
 280 /* like inode_setattr, but doesn't mark the inode dirty */
 281 static int ll_attr2inode(struct inode * inode, struct iattr * attr, int trunc)
 282 {
 283         unsigned int ia_valid = attr->ia_valid;
 284         int error = 0;
 285
 286         if ((ia_valid & ATTR_SIZE) && trunc) {
 287                 error = vmtruncate(inode, attr->ia_size);
 288                 if (error)
 289                         goto out;
 290         } else if (ia_valid & ATTR_SIZE)
 291                 inode->i_size = attr->ia_size;
 292
 293         if (ia_valid & ATTR_UID)
 294                 inode->i_uid = attr->ia_uid;
 295         if (ia_valid & ATTR_GID)
 296                 inode->i_gid = attr->ia_gid;
 297         if (ia_valid & ATTR_ATIME)
 298                 inode->i_atime = attr->ia_atime;
 299         if (ia_valid & ATTR_MTIME)
 300                 inode->i_mtime = attr->ia_mtime;
 301         if (ia_valid & ATTR_CTIME)
 302                 inode->i_ctime = attr->ia_ctime;
 303         if (ia_valid & ATTR_MODE) {
 304                 inode->i_mode = attr->ia_mode;
 305                 if (!in_group_p(inode->i_gid) && !capable(CAP_FSETID))
 306                         inode->i_mode &= ~S_ISGID;
 307         }
 308 out:
 309         return error;
 310 }
 311
 312 int ll_inode_setattr(struct inode *inode, struct iattr *attr, int do_trunc)
 313 {
 314         struct ptlrpc_request *request = NULL;
 315         struct ll_sb_info *sbi = ll_i2sbi(inode);
 316         int err;
 317
 318         ENTRY;
 319
 320         /* change incore inode */
 321         ll_attr2inode(inode, attr, do_trunc);
 322
 323         err = mdc_setattr(&sbi->ll_mdc_conn, inode, attr, &request);
 324         if (err)
 325                 CERROR("mdc_setattr fails (%d)\n", err);
 326
 327         ptlrpc_req_finished(request);
 328
 329         RETURN(err);
 330 }
 331
 332 int ll_setattr(struct dentry *de, struct iattr *attr)
 333 {
 334         int rc = inode_change_ok(de->d_inode, attr);
 335
 336         if (rc)
 337                 return rc;
 338
 339         return ll_inode_setattr(de->d_inode, attr, 1);
 340 }
 341
 342 static int ll_statfs(struct super_block *sb, struct statfs *sfs)
 343 {
 344         struct ptlrpc_request *request = NULL;
 345         struct ll_sb_info *sbi = ll_s2sbi(sb);
 346         int rc;
 347         ENTRY;
 348
 349         memset(sfs, 0, sizeof(*sfs));
 350         rc = mdc_statfs(&sbi->ll_mdc_conn, sfs, &request);
 351         ptlrpc_req_finished(request);
 352         if (rc)
 353                 CERROR("obd_statfs fails: rc = %d\n", rc);
 354         else
 355                 CDEBUG(D_SUPER, "statfs shows blocks %ld/%ld objects %ld/%ld\n",
 356                        sfs->f_bavail, sfs->f_blocks, sfs->f_files,sfs->f_ffree);
 357
 358         /* temporary until mds_statfs returns statfs info for all OSTs */
 359         if (!rc) {
 360                 struct statfs obd_sfs;
 361
 362                 rc = obd_statfs(&sbi->ll_osc_conn, &obd_sfs);
 363                 if (rc) {
 364                         CERROR("obd_statfs fails: rc = %d\n", rc);
 365                         GOTO(out, rc);
 366                 }
 367                 CDEBUG(D_SUPER, "obd_statfs returns blocks %ld/%ld, "
 368                        "objects %ld/%ld\n",
 369                        obd_sfs.f_bavail, obd_sfs.f_blocks,
 370                        obd_sfs.f_ffree, obd_sfs.f_files);
 371
 372                 sfs->f_bfree = obd_sfs.f_bfree;
 373                 sfs->f_bavail = obd_sfs.f_bavail;
 374                 sfs->f_blocks = obd_sfs.f_blocks;
 375                 if (obd_sfs.f_ffree < sfs->f_ffree)
 376                         sfs->f_ffree = obd_sfs.f_ffree;
 377         }
 378
 379 out:
 380         RETURN(rc);
 381 }
 382
 383 inline int ll_stripe_md_size(struct super_block *sb)
 384 {
 385         struct mdc_obd *mdc = sbi2mdc(ll_s2sbi(sb));
 386         return mdc->mdc_max_mdsize;
 387 }
 388
 389 static void ll_to_inode(struct inode *dst, struct ll_inode_md *md)
 390 {
 391         struct mds_body *body = md->body;
 392         struct ll_inode_info *ii = ll_i2info(dst);
 393
 394         /* core attributes first */
 395         if (body->valid & OBD_MD_FLID)
 396                 dst->i_ino = body->ino;
 397         if (body->valid & OBD_MD_FLATIME)
 398                 dst->i_atime = body->atime;
 399         if (body->valid & OBD_MD_FLMTIME)
 400                 dst->i_mtime = body->mtime;
 401         if (body->valid & OBD_MD_FLCTIME)
 402                 dst->i_ctime = body->ctime;
 403         if (body->valid & OBD_MD_FLSIZE)
 404                 dst->i_size = body->size;
 405         if (body->valid & OBD_MD_FLMODE)
 406                 dst->i_mode = body->mode;
 407         if (body->valid & OBD_MD_FLUID)
 408                 dst->i_uid = body->uid;
 409         if (body->valid & OBD_MD_FLGID)
 410                 dst->i_gid = body->gid;
 411         if (body->valid & OBD_MD_FLFLAGS)
 412                 dst->i_flags = body->flags;
 413         if (body->valid & OBD_MD_FLNLINK)
 414                 dst->i_nlink = body->nlink;
 415         if (body->valid & OBD_MD_FLGENER)
 416                 dst->i_generation = body->generation;
 417         if (body->valid & OBD_MD_FLRDEV)
 418                 dst->i_rdev = body->extra;
 419         if (md && md->md && md->md->lmd_stripe_count) {
 420                 struct lov_stripe_md *smd = md->md;
 421                 int size = ll_stripe_md_size(dst->i_sb);
 422                 if (md->md->lmd_size != size) {
 423                         CERROR("Striping metadata size error %ld\n",
 424                                dst->i_ino);
 425                         LBUG();
 426                 }
 427                 OBD_ALLOC(ii->lli_smd, size);
 428                 if (!ii->lli_smd){
 429                         CERROR("No memory for %d\n", size);
 430                         LBUG();
 431                 }
 432                 memcpy(ii->lli_smd, smd, size);
 433         }
 434 } /* ll_to_inode */
 435
 436 static void ll_read_inode2(struct inode *inode, void *opaque)
 437 {
 438         struct ll_inode_md *md = opaque;
 439
 440         ENTRY;
 441         ll_to_inode(inode, md);
 442
 443         /* OIDEBUG(inode); */
 444
 445         if (S_ISREG(inode->i_mode)) {
 446                 inode->i_op = &ll_file_inode_operations;
 447                 inode->i_fop = &ll_file_operations;
 448                 inode->i_mapping->a_ops = &ll_aops;
 449                 EXIT;
 450         } else if (S_ISDIR(inode->i_mode)) {
 451                 inode->i_op = &ll_dir_inode_operations;
 452                 inode->i_fop = &ll_dir_operations;
 453                 inode->i_mapping->a_ops = &ll_dir_aops;
 454                 EXIT;
 455         } else if (S_ISLNK(inode->i_mode)) {
 456                 inode->i_op = &ll_fast_symlink_inode_operations;
 457                 EXIT;
 458         } else {
 459                 init_special_inode(inode, inode->i_mode,
 460                                    ((int *)ll_i2info(inode)->lli_inline)[0]);
 461                 EXIT;
 462         }
 463
 464         return;
 465 }
 466
 467 /* exported operations */
 468 struct super_operations ll_super_operations =
 469 {
 470         read_inode2: ll_read_inode2,
 471         clear_inode: ll_clear_inode,
 472         delete_inode: ll_delete_inode,
 473         put_super: ll_put_super,
 474         statfs: ll_statfs
 475 };
 476
 477 struct file_system_type lustre_lite_fs_type = {
 478         "lustre_lite", 0, ll_read_super, NULL
 479 };
 480
 481 static int __init init_lustre_lite(void)
 482 {
 483         printk(KERN_INFO "Lustre Lite 0.0.1, info@clusterfs.com\n");
 484         ll_file_data_slab = kmem_cache_create("ll_file_data",
 485                                               sizeof(struct ll_file_data), 0,
 486                                                SLAB_HWCACHE_ALIGN, NULL, NULL);
 487         if (ll_file_data_slab == NULL)
 488                 return -ENOMEM;
 489         return register_filesystem(&lustre_lite_fs_type);
 490 }
 491
 492 static void __exit exit_lustre_lite(void)
 493 {
 494         unregister_filesystem(&lustre_lite_fs_type);
 495         kmem_cache_destroy(ll_file_data_slab);
 496 }
 497
 498 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
 499 MODULE_DESCRIPTION("Lustre Lite Client File System v1.0");
 500 MODULE_LICENSE("GPL");
 501
 502 module_init(init_lustre_lite);
 503 module_exit(exit_lustre_lite);