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("osc", this_char)))||
  68                      (!*mds && (*mds = ll_read_opt("mdc", 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 *osc = NULL;
  85         char *mdc = 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, &osc, &mdc);
 106
 107         if (!osc) {
 108                 CERROR("no osc\n");
 109                 GOTO(out_free, sb = NULL);
 110         }
 111
 112         if (!mdc) {
 113                 CERROR("no mdc\n");
 114                 GOTO(out_free, sb = NULL);
 115         }
 116
 117         obd = class_uuid2obd(mdc);
 118         if (!obd) {
 119                 CERROR("MDC %s: not setup or attached\n", mdc);
 120                 GOTO(out_free, sb = NULL);
 121         }
 122
 123 #if 0
 124         err = connmgr_connect(ptlrpc_connmgr, sbi->ll_mdc_conn);
 125         if (err) {
 126                 CERROR("cannot connect to MDC: rc = %d\n", err);
 127                 GOTO(out_rpc, sb = NULL);
 128         }
 129 #endif
 130
 131 #warning shaver: might need a cluuid here
 132         err = obd_connect(&sbi->ll_mdc_conn, obd, NULL);
 133         if (err) {
 134                 CERROR("cannot connect to %s: rc = %d\n", mdc, err);
 135                 GOTO(out_free, sb = NULL);
 136         }
 137         sbi2mdc(sbi)->cl_conn->c_level = LUSTRE_CONN_FULL;
 138
 139         obd = class_uuid2obd(osc);
 140         if (!obd) {
 141                 CERROR("OSC %s: not setup or attached\n", osc);
 142                 GOTO(out_mdc, sb = NULL);
 143         }
 144 #warning shaver: might need a cluuid here
 145         err = obd_connect(&sbi->ll_osc_conn, obd, NULL);
 146         if (err) {
 147                 CERROR("cannot connect to %s: rc = %d\n", osc, err);
 148                 GOTO(out_mdc, sb = NULL);
 149         }
 150
 151         /* XXX: need to store the last_* values somewhere */
 152         err = mdc_getstatus(&sbi->ll_mdc_conn, &rootfid, &last_committed,
 153                             &last_rcvd, &last_xid, &request);
 154         ptlrpc_req_finished(request);
 155         if (err) {
 156                 CERROR("cannot mds_connect: rc = %d\n", err);
 157                 GOTO(out_mdc, sb = NULL);
 158         }
 159         CDEBUG(D_SUPER, "rootfid %Ld\n", (unsigned long long)rootfid.id);
 160         sbi->ll_rootino = rootfid.id;
 161
 162         memset(&sfs, 0, sizeof(sfs));
 163         request = NULL;
 164         err = mdc_statfs(&sbi->ll_mdc_conn, &sfs, &request);
 165         ptlrpc_req_finished(request);
 166         sb->s_blocksize = sfs.f_bsize;
 167         sb->s_blocksize_bits = log2(sfs.f_bsize);
 168         sb->s_magic = LL_SUPER_MAGIC;
 169         sb->s_maxbytes = (1ULL << (32 + 9)) - sfs.f_bsize;
 170
 171         sb->s_op = &ll_super_operations;
 172
 173         /* make root inode */
 174         request = NULL;
 175         err = mdc_getattr(&sbi->ll_mdc_conn, sbi->ll_rootino, S_IFDIR,
 176                           OBD_MD_FLNOTOBD|OBD_MD_FLBLOCKS, 0, &request);
 177         if (err) {
 178                 CERROR("mdc_getattr failed for root: rc = %d\n", err);
 179                 GOTO(out_request, sb = NULL);
 180         }
 181
 182         /* initialize committed transaction callback daemon */
 183         spin_lock_init(&sbi->ll_commitcbd_lock);
 184         init_waitqueue_head(&sbi->ll_commitcbd_waitq);
 185         init_waitqueue_head(&sbi->ll_commitcbd_ctl_waitq);
 186         sbi->ll_commitcbd_flags = 0;
 187         err = ll_commitcbd_setup(sbi);
 188         if (err) {
 189                 CERROR("failed to start commit callback daemon: rc = %d\n",err);
 190                 GOTO(out_request, sb = NULL);
 191         }
 192
 193         md.body = lustre_msg_buf(request->rq_repmsg, 0);
 194         md.md = NULL;
 195         root = iget4(sb, sbi->ll_rootino, NULL, &md);
 196
 197         if (root) {
 198                 sb->s_root = d_alloc_root(root);
 199         } else {
 200                 CERROR("lustre_lite: bad iget4 for root\n");
 201                 GOTO(out_cdb, sb = NULL);
 202         }
 203
 204         ptlrpc_req_finished(request);
 205
 206 out_dev:
 207         if (mdc)
 208                 OBD_FREE(mdc, strlen(mdc) + 1);
 209         if (osc)
 210                 OBD_FREE(osc, strlen(osc) + 1);
 211
 212         RETURN(sb);
 213
 214 out_cdb:
 215         ll_commitcbd_cleanup(sbi);
 216 out_request:
 217         ptlrpc_req_finished(request);
 218         obd_disconnect(&sbi->ll_osc_conn);
 219 out_mdc:
 220         obd_disconnect(&sbi->ll_mdc_conn);
 221 out_free:
 222         OBD_FREE(sbi, sizeof(*sbi));
 223
 224         MOD_DEC_USE_COUNT;
 225         goto out_dev;
 226 } /* ll_read_super */
 227
 228 static void ll_put_super(struct super_block *sb)
 229 {
 230         struct ll_sb_info *sbi = ll_s2sbi(sb);
 231         ENTRY;
 232         ll_commitcbd_cleanup(sbi);
 233         obd_disconnect(&sbi->ll_osc_conn);
 234         obd_disconnect(&sbi->ll_mdc_conn);
 235         OBD_FREE(sbi, sizeof(*sbi));
 236
 237         MOD_DEC_USE_COUNT;
 238         EXIT;
 239 } /* ll_put_super */
 240
 241 static void ll_clear_inode(struct inode *inode)
 242 {
 243         if (atomic_read(&inode->i_count) == 0) {
 244                 struct ll_inode_info *lli = ll_i2info(inode);
 245                 struct lov_stripe_md *md = lli->lli_smd;
 246
 247                 if (md) {
 248                         OBD_FREE(md, md->lmd_easize);
 249                         lli->lli_smd = NULL;
 250                 }
 251                 if (lli->lli_symlink_name) {
 252                         OBD_FREE(lli->lli_symlink_name,
 253                                  strlen(lli->lli_symlink_name) + 1);
 254                         lli->lli_symlink_name = NULL;
 255                 }
 256         }
 257 }
 258
 259 static void ll_delete_inode(struct inode *inode)
 260 {
 261         if (S_ISREG(inode->i_mode)) {
 262                 int err;
 263                 struct obdo *oa;
 264                 struct lov_stripe_md *md = ll_i2info(inode)->lli_smd;
 265
 266                 if (!md)
 267                         GOTO(out, -EINVAL);
 268
 269                 if (md->lmd_object_id == 0) {
 270                         CERROR("This really happens\n");
 271                         /* No obdo was ever created */
 272                         GOTO(out, 0);
 273                 }
 274
 275                 oa = obdo_alloc();
 276                 if (oa == NULL)
 277                         GOTO(out, -ENOMEM);
 278
 279                 oa->o_id = md->lmd_object_id;
 280                 oa->o_easize = md->lmd_easize;
 281                 oa->o_mode = inode->i_mode;
 282                 oa->o_valid = OBD_MD_FLID | OBD_MD_FLEASIZE | OBD_MD_FLMODE;
 283
 284                 err = obd_destroy(ll_i2obdconn(inode), oa, md);
 285                 obdo_free(oa);
 286                 CDEBUG(D_SUPER, "obd destroy of %Ld error %d\n",
 287                        md->lmd_object_id, err);
 288         }
 289 out:
 290         clear_inode(inode);
 291 }
 292
 293 /* like inode_setattr, but doesn't mark the inode dirty */
 294 static int ll_attr2inode(struct inode * inode, struct iattr * attr, int trunc)
 295 {
 296         unsigned int ia_valid = attr->ia_valid;
 297         int error = 0;
 298
 299         if ((ia_valid & ATTR_SIZE) && trunc) {
 300                 error = vmtruncate(inode, attr->ia_size);
 301                 if (error)
 302                         goto out;
 303         } else if (ia_valid & ATTR_SIZE)
 304                 inode->i_size = attr->ia_size;
 305
 306         if (ia_valid & ATTR_UID)
 307                 inode->i_uid = attr->ia_uid;
 308         if (ia_valid & ATTR_GID)
 309                 inode->i_gid = attr->ia_gid;
 310         if (ia_valid & ATTR_ATIME)
 311                 inode->i_atime = attr->ia_atime;
 312         if (ia_valid & ATTR_MTIME)
 313                 inode->i_mtime = attr->ia_mtime;
 314         if (ia_valid & ATTR_CTIME)
 315                 inode->i_ctime = attr->ia_ctime;
 316         if (ia_valid & ATTR_MODE) {
 317                 inode->i_mode = attr->ia_mode;
 318                 if (!in_group_p(inode->i_gid) && !capable(CAP_FSETID))
 319                         inode->i_mode &= ~S_ISGID;
 320         }
 321 out:
 322         return error;
 323 }
 324
 325 int ll_inode_setattr(struct inode *inode, struct iattr *attr, int do_trunc)
 326 {
 327         struct ptlrpc_request *request = NULL;
 328         struct ll_sb_info *sbi = ll_i2sbi(inode);
 329         int err;
 330
 331         ENTRY;
 332
 333         /* change incore inode */
 334         ll_attr2inode(inode, attr, do_trunc);
 335
 336         err = mdc_setattr(&sbi->ll_mdc_conn, inode, attr, &request);
 337         if (err)
 338                 CERROR("mdc_setattr fails (%d)\n", err);
 339
 340         ptlrpc_req_finished(request);
 341
 342         RETURN(err);
 343 }
 344
 345 int ll_setattr(struct dentry *de, struct iattr *attr)
 346 {
 347         int rc = inode_change_ok(de->d_inode, attr);
 348
 349         if (rc)
 350                 return rc;
 351
 352         return ll_inode_setattr(de->d_inode, attr, 1);
 353 }
 354
 355 static int ll_statfs(struct super_block *sb, struct statfs *sfs)
 356 {
 357         struct ptlrpc_request *request = NULL;
 358         struct ll_sb_info *sbi = ll_s2sbi(sb);
 359         int rc;
 360         ENTRY;
 361
 362         memset(sfs, 0, sizeof(*sfs));
 363         rc = mdc_statfs(&sbi->ll_mdc_conn, sfs, &request);
 364         ptlrpc_req_finished(request);
 365         if (rc)
 366                 CERROR("obd_statfs fails: rc = %d\n", rc);
 367         else
 368                 CDEBUG(D_SUPER, "statfs shows blocks %ld/%ld objects %ld/%ld\n",
 369                        sfs->f_bavail, sfs->f_blocks, sfs->f_files,sfs->f_ffree);
 370
 371         /* temporary until mds_statfs returns statfs info for all OSTs */
 372         if (!rc) {
 373                 struct statfs obd_sfs;
 374
 375                 rc = obd_statfs(&sbi->ll_osc_conn, &obd_sfs);
 376                 if (rc) {
 377                         CERROR("obd_statfs fails: rc = %d\n", rc);
 378                         GOTO(out, rc);
 379                 }
 380                 CDEBUG(D_SUPER, "obd_statfs returns blocks %ld/%ld, "
 381                        "objects %ld/%ld\n",
 382                        obd_sfs.f_bavail, obd_sfs.f_blocks,
 383                        obd_sfs.f_ffree, obd_sfs.f_files);
 384
 385                 sfs->f_bfree = obd_sfs.f_bfree;
 386                 sfs->f_bavail = obd_sfs.f_bavail;
 387                 sfs->f_blocks = obd_sfs.f_blocks;
 388                 if (obd_sfs.f_ffree < sfs->f_ffree)
 389                         sfs->f_ffree = obd_sfs.f_ffree;
 390         }
 391
 392 out:
 393         RETURN(rc);
 394 }
 395
 396 inline int ll_stripe_md_size(struct super_block *sb)
 397 {
 398         struct client_obd *mdc = sbi2mdc(ll_s2sbi(sb));
 399         return mdc->cl_max_mdsize;
 400 }
 401
 402 static void ll_read_inode2(struct inode *inode, void *opaque)
 403 {
 404         struct ll_inode_md *md = opaque;
 405         struct mds_body *body = md->body;
 406         struct ll_inode_info *ii = ll_i2info(inode);
 407         ENTRY;
 408
 409         sema_init(&ii->lli_open_sem, 1);
 410
 411         /* core attributes first */
 412         if (body->valid & OBD_MD_FLID)
 413                 inode->i_ino = body->ino;
 414         if (body->valid & OBD_MD_FLATIME)
 415                 inode->i_atime = body->atime;
 416         if (body->valid & OBD_MD_FLMTIME)
 417                 inode->i_mtime = body->mtime;
 418         if (body->valid & OBD_MD_FLCTIME)
 419                 inode->i_ctime = body->ctime;
 420         if (body->valid & OBD_MD_FLSIZE)
 421                 inode->i_size = body->size;
 422         if (body->valid & OBD_MD_FLMODE)
 423                 inode->i_mode = body->mode;
 424         if (body->valid & OBD_MD_FLUID)
 425                 inode->i_uid = body->uid;
 426         if (body->valid & OBD_MD_FLGID)
 427                 inode->i_gid = body->gid;
 428         if (body->valid & OBD_MD_FLFLAGS)
 429                 inode->i_flags = body->flags;
 430         if (body->valid & OBD_MD_FLNLINK)
 431                 inode->i_nlink = body->nlink;
 432         if (body->valid & OBD_MD_FLGENER)
 433                 inode->i_generation = body->generation;
 434         if (body->valid & OBD_MD_FLRDEV)
 435                 inode->i_rdev = body->extra;
 436         //if (body->valid & OBD_MD_FLEASIZE)
 437         if (md && md->md && md->md->lmd_stripe_count) {
 438                 struct lov_stripe_md *smd = md->md;
 439                 int size = ll_stripe_md_size(inode->i_sb);
 440                 if (md->md->lmd_easize != size) {
 441                         CERROR("Striping metadata size error %ld\n",
 442                                inode->i_ino);
 443                         LBUG();
 444                 }
 445                 OBD_ALLOC(ii->lli_smd, size);
 446                 if (!ii->lli_smd){
 447                         CERROR("No memory for %d\n", size);
 448                         LBUG();
 449                 }
 450                 memcpy(ii->lli_smd, smd, size);
 451         }
 452
 453         /* OIDEBUG(inode); */
 454
 455         if (S_ISREG(inode->i_mode)) {
 456                 inode->i_op = &ll_file_inode_operations;
 457                 inode->i_fop = &ll_file_operations;
 458                 inode->i_mapping->a_ops = &ll_aops;
 459                 EXIT;
 460         } else if (S_ISDIR(inode->i_mode)) {
 461                 inode->i_op = &ll_dir_inode_operations;
 462                 inode->i_fop = &ll_dir_operations;
 463                 inode->i_mapping->a_ops = &ll_dir_aops;
 464                 EXIT;
 465         } else if (S_ISLNK(inode->i_mode)) {
 466                 inode->i_op = &ll_fast_symlink_inode_operations;
 467                 EXIT;
 468         } else {
 469                 init_special_inode(inode, inode->i_mode, inode->i_rdev);
 470                 EXIT;
 471         }
 472
 473         return;
 474 }
 475
 476 /* exported operations */
 477 struct super_operations ll_super_operations =
 478 {
 479         read_inode2: ll_read_inode2,
 480         clear_inode: ll_clear_inode,
 481         delete_inode: ll_delete_inode,
 482         put_super: ll_put_super,
 483         statfs: ll_statfs
 484 };
 485
 486 struct file_system_type lustre_lite_fs_type = {
 487         "lustre_lite", 0, ll_read_super, NULL
 488 };
 489
 490 static int __init init_lustre_lite(void)
 491 {
 492         printk(KERN_INFO "Lustre Lite 0.0.1, info@clusterfs.com\n");
 493         ll_file_data_slab = kmem_cache_create("ll_file_data",
 494                                               sizeof(struct ll_file_data), 0,
 495                                                SLAB_HWCACHE_ALIGN, NULL, NULL);
 496         if (ll_file_data_slab == NULL)
 497                 return -ENOMEM;
 498         return register_filesystem(&lustre_lite_fs_type);
 499 }
 500
 501 static void __exit exit_lustre_lite(void)
 502 {
 503         unregister_filesystem(&lustre_lite_fs_type);
 504         kmem_cache_destroy(ll_file_data_slab);
 505 }
 506
 507 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
 508 MODULE_DESCRIPTION("Lustre Lite Client File System v1.0");
 509 MODULE_LICENSE("GPL");
 510
 511 module_init(init_lustre_lite);
 512 module_exit(exit_lustre_lite);