lustre/mds/handler.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  *  linux/mds/handler.c
   5  *
   6  *  Lustre Metadata Server (mds) request handler
   7  *
   8  *  Copyright (C) 2001, 2002 Cluster File Systems, Inc.
   9  *
  10  *  This code is issued under the GNU General Public License.
  11  *  See the file COPYING in this distribution
  12  *
  13  *  by Peter Braam <braam@clusterfs.com>
  14  *
  15  *  This server is single threaded at present (but can easily be multi threaded)
  16  *
  17  */
  18
  19 #define EXPORT_SYMTAB
  20
  21 #include <linux/version.h>
  22 #include <linux/module.h>
  23 #include <linux/fs.h>
  24 #include <linux/stat.h>
  25 #include <linux/locks.h>
  26 #include <linux/quotaops.h>
  27 #include <asm/unistd.h>
  28 #include <asm/uaccess.h>
  29
  30 #define DEBUG_SUBSYSTEM S_MDS
  31
  32 #include <linux/lustre_mds.h>
  33 #include <linux/lustre_lib.h>
  34 #include <linux/lustre_net.h>
  35
  36 int mds_sendpage(struct ptlrpc_request *req, struct file *file,
  37                  __u64 offset, struct niobuf *dst)
  38 {
  39         int rc = 0;
  40         mm_segment_t oldfs = get_fs();
  41
  42         if (req->rq_peer.peer_nid == 0) {
  43                 /* dst->addr is a user address, but in a different task! */
  44                 char *buf = (char *)(long)dst->addr;
  45
  46                 set_fs(KERNEL_DS);
  47                 rc = mds_fs_readpage(&req->rq_obd->u.mds, file, buf, PAGE_SIZE,
  48                                      &offset);
  49                 set_fs(oldfs);
  50
  51                 if (rc != PAGE_SIZE) {
  52                         rc = -EIO;
  53                         GOTO(out, rc);
  54                 }
  55                 EXIT;
  56         } else {
  57                 struct ptlrpc_bulk_desc *bulk;
  58                 char *buf;
  59
  60                 bulk = ptlrpc_prep_bulk(&req->rq_peer);
  61                 if (bulk == NULL) {
  62                         rc = -ENOMEM;
  63                         GOTO(out, rc);
  64                 }
  65
  66                 bulk->b_xid = req->rq_xid;
  67
  68                 OBD_ALLOC(buf, PAGE_SIZE);
  69                 if (!buf) {
  70                         rc = -ENOMEM;
  71                         GOTO(cleanup_bulk, rc);
  72                 }
  73
  74                 set_fs(KERNEL_DS);
  75                 rc = mds_fs_readpage(&req->rq_obd->u.mds, file, buf, PAGE_SIZE,
  76                                      &offset);
  77                 set_fs(oldfs);
  78
  79                 if (rc != PAGE_SIZE) {
  80                         rc = -EIO;
  81                         GOTO(cleanup_buf, rc);
  82                 }
  83
  84                 bulk->b_buf = buf;
  85                 bulk->b_buflen = PAGE_SIZE;
  86
  87                 rc = ptlrpc_send_bulk(bulk, MDS_BULK_PORTAL);
  88                 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
  89                         CERROR("obd_fail_loc=%x, fail operation rc=%d\n",
  90                                OBD_FAIL_MDS_SENDPAGE, rc);
  91                         PtlMDUnlink(bulk->b_md_h);
  92                         GOTO(cleanup_buf, rc);
  93                 }
  94                 wait_event_interruptible(bulk->b_waitq,
  95                                          ptlrpc_check_bulk_sent(bulk));
  96
  97                 if (bulk->b_flags == PTL_RPC_INTR) {
  98                         rc = -EINTR;
  99                         GOTO(cleanup_buf, rc);
 100                 }
 101
 102                 EXIT;
 103         cleanup_buf:
 104                 OBD_FREE(buf, PAGE_SIZE);
 105         cleanup_bulk:
 106                 OBD_FREE(bulk, sizeof(*bulk));
 107         }
 108 out:
 109         return rc;
 110 }
 111
 112 struct dentry *mds_fid2dentry(struct mds_obd *mds, struct ll_fid *fid,
 113                               struct vfsmount **mnt)
 114 {
 115         /* stolen from NFS */
 116         struct super_block *sb = mds->mds_sb;
 117         unsigned long ino = fid->id;
 118         __u32 generation = fid->generation;
 119         struct inode *inode;
 120         struct list_head *lp;
 121         struct dentry *result;
 122
 123         if (ino == 0)
 124                 return ERR_PTR(-ESTALE);
 125
 126         inode = iget(sb, ino);
 127         if (inode == NULL)
 128                 return ERR_PTR(-ENOMEM);
 129
 130         CDEBUG(D_DENTRY, "--> mds_fid2dentry: sb %p\n", inode->i_sb);
 131
 132         if (is_bad_inode(inode) ||
 133             (generation && inode->i_generation != generation)) {
 134                 /* we didn't find the right inode.. */
 135                 CERROR("bad inode %lu, link: %d ct: %d or version  %u/%u\n",
 136                         inode->i_ino,
 137                         inode->i_nlink, atomic_read(&inode->i_count),
 138                         inode->i_generation,
 139                         generation);
 140                 LBUG();
 141                 iput(inode);
 142                 return ERR_PTR(-ESTALE);
 143         }
 144
 145         /* now to find a dentry.
 146          * If possible, get a well-connected one
 147          */
 148         if (mnt)
 149                 *mnt = mds->mds_vfsmnt;
 150         spin_lock(&dcache_lock);
 151         for (lp = inode->i_dentry.next; lp != &inode->i_dentry ; lp=lp->next) {
 152                 result = list_entry(lp,struct dentry, d_alias);
 153                 if (! (result->d_flags & DCACHE_NFSD_DISCONNECTED)) {
 154                         dget_locked(result);
 155                         result->d_vfs_flags |= DCACHE_REFERENCED;
 156                         spin_unlock(&dcache_lock);
 157                         iput(inode);
 158                         if (mnt)
 159                                 mntget(*mnt);
 160                         return result;
 161                 }
 162         }
 163         spin_unlock(&dcache_lock);
 164         result = d_alloc_root(inode);
 165         if (result == NULL) {
 166                 iput(inode);
 167                 return ERR_PTR(-ENOMEM);
 168         }
 169         if (mnt)
 170                 mntget(*mnt);
 171         result->d_flags |= DCACHE_NFSD_DISCONNECTED;
 172         return result;
 173 }
 174
 175 int mds_getattr(struct ptlrpc_request *req)
 176 {
 177         struct dentry *de;
 178         struct inode *inode;
 179         struct mds_body *body;
 180         struct mds_obd *mds = &req->rq_obd->u.mds;
 181         int rc, size = sizeof(*body);
 182         ENTRY;
 183
 184         rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repbuf);
 185         req->rq_repmsg = (struct lustre_msg *)req->rq_repbuf;
 186         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK)) {
 187                 CERROR("mds: out of memory\n");
 188                 req->rq_status = -ENOMEM;
 189                 RETURN(0);
 190         }
 191
 192         body = lustre_msg_buf(req->rq_reqmsg, 0);
 193         de = mds_fid2dentry(mds, &body->fid1, NULL);
 194         if (IS_ERR(de)) {
 195                 req->rq_status = -ENOENT;
 196                 RETURN(0);
 197         }
 198
 199         body = lustre_msg_buf(req->rq_repmsg, 0);
 200         inode = de->d_inode;
 201         body->ino = inode->i_ino;
 202         body->generation = inode->i_generation;
 203         body->atime = inode->i_atime;
 204         body->ctime = inode->i_ctime;
 205         body->mtime = inode->i_mtime;
 206         body->uid = inode->i_uid;
 207         body->gid = inode->i_gid;
 208         body->size = inode->i_size;
 209         body->mode = inode->i_mode;
 210         body->nlink = inode->i_nlink;
 211         body->valid = ~0;
 212         mds_fs_get_objid(mds, inode, &body->objid);
 213         l_dput(de);
 214         RETURN(0);
 215 }
 216
 217 int mds_open(struct ptlrpc_request *req)
 218 {
 219         struct dentry *de;
 220         struct mds_body *body;
 221         struct file *file;
 222         struct vfsmount *mnt;
 223         __u32 flags;
 224         int rc, size = sizeof(*body);
 225         ENTRY;
 226
 227         rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repbuf);
 228         req->rq_repmsg = (struct lustre_msg *)req->rq_repbuf;
 229         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK)) {
 230                 CERROR("mds: out of memory\n");
 231                 req->rq_status = -ENOMEM;
 232                 RETURN(0);
 233         }
 234
 235         body = lustre_msg_buf(req->rq_reqmsg, 0);
 236         de = mds_fid2dentry(&req->rq_obd->u.mds, &body->fid1, &mnt);
 237         if (IS_ERR(de)) {
 238                 req->rq_status = -ENOENT;
 239                 RETURN(0);
 240         }
 241         flags = body->flags;
 242         file = dentry_open(de, mnt, flags);
 243         if (!file || IS_ERR(file)) {
 244                 req->rq_status = -EINVAL;
 245                 RETURN(0);
 246         }
 247
 248         body = lustre_msg_buf(req->rq_repmsg, 0);
 249         body->objid = (__u64) (unsigned long)file;
 250         RETURN(0);
 251 }
 252
 253 int mds_close(struct ptlrpc_request *req)
 254 {
 255         struct dentry *de;
 256         struct mds_body *body;
 257         struct file *file;
 258         struct vfsmount *mnt;
 259         int rc;
 260         ENTRY;
 261
 262         rc = lustre_pack_msg(0, NULL, NULL, &req->rq_replen, &req->rq_repbuf);
 263         req->rq_repmsg = (struct lustre_msg *)req->rq_repbuf;
 264         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK)) {
 265                 CERROR("mds: out of memory\n");
 266                 req->rq_status = -ENOMEM;
 267                 RETURN(0);
 268         }
 269
 270         body = lustre_msg_buf(req->rq_reqmsg, 0);
 271         de = mds_fid2dentry(&req->rq_obd->u.mds, &body->fid1, &mnt);
 272         if (IS_ERR(de)) {
 273                 req->rq_status = -ENOENT;
 274                 RETURN(0);
 275         }
 276
 277         file = (struct file *)(unsigned long)body->objid;
 278         req->rq_status = filp_close(file, 0);
 279         l_dput(de);
 280         mntput(mnt);
 281
 282         RETURN(0);
 283 }
 284
 285 int mds_readpage(struct ptlrpc_request *req)
 286 {
 287         struct vfsmount *mnt;
 288         struct dentry *de;
 289         struct file *file;
 290         struct niobuf *niobuf;
 291         struct mds_body *body;
 292         int rc, size = sizeof(*body);
 293         ENTRY;
 294
 295         rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repbuf);
 296         req->rq_repmsg = (struct lustre_msg *)req->rq_repbuf;
 297         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_READPAGE_PACK)) {
 298                 CERROR("mds: out of memory\n");
 299                 req->rq_status = -ENOMEM;
 300                 RETURN(0);
 301         }
 302
 303         body = lustre_msg_buf(req->rq_reqmsg, 0);
 304         de = mds_fid2dentry(&req->rq_obd->u.mds, &body->fid1, &mnt);
 305         if (IS_ERR(de)) {
 306                 req->rq_status = PTR_ERR(de);
 307                 RETURN(0);
 308         }
 309
 310         CDEBUG(D_INODE, "ino %ld\n", de->d_inode->i_ino);
 311
 312         file = dentry_open(de, mnt, O_RDONLY | O_LARGEFILE);
 313         /* note: in case of an error, dentry_open puts dentry */
 314         if (IS_ERR(file)) {
 315                 req->rq_status = PTR_ERR(file);
 316                 RETURN(0);
 317         }
 318
 319         niobuf = lustre_msg_buf(req->rq_reqmsg, 1);
 320         if (!niobuf) {
 321                 req->rq_status = -EINVAL;
 322                 LBUG();
 323                 RETURN(0);
 324         }
 325
 326         /* to make this asynchronous make sure that the handling function
 327            doesn't send a reply when this function completes. Instead a
 328            callback function would send the reply */
 329         rc = mds_sendpage(req, file, body->size, niobuf);
 330
 331         filp_close(file, 0);
 332         req->rq_status = rc;
 333         RETURN(0);
 334 }
 335
 336 int mds_reint(struct ptlrpc_request *req)
 337 {
 338         int rc;
 339         struct mds_update_record rec;
 340
 341         rc = mds_update_unpack(req, &rec);
 342         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNPACK)) {
 343                 CERROR("invalid record\n");
 344                 req->rq_status = -EINVAL;
 345                 RETURN(0);
 346         }
 347         /* rc will be used to interrupt a for loop over multiple records */
 348         rc = mds_reint_rec(&rec, req);
 349         return 0;
 350 }
 351
 352 int mds_handle(struct obd_device *dev, struct ptlrpc_service *svc,
 353                struct ptlrpc_request *req)
 354 {
 355         int rc;
 356         ENTRY;
 357
 358         rc = lustre_unpack_msg(req->rq_reqbuf, req->rq_reqlen);
 359         req->rq_reqmsg = (struct lustre_msg *)req->rq_reqbuf;
 360         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_HANDLE_UNPACK)) {
 361                 CERROR("lustre_mds: Invalid request\n");
 362                 GOTO(out, rc);
 363         }
 364
 365         if (req->rq_reqmsg->type != PTL_RPC_REQUEST) {
 366                 CERROR("lustre_mds: wrong packet type sent %d\n",
 367                        req->rq_reqmsg->type);
 368                 GOTO(out, rc = -EINVAL);
 369         }
 370
 371         switch (req->rq_reqmsg->opc) {
 372         case MDS_GETATTR:
 373                 CDEBUG(D_INODE, "getattr\n");
 374                 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_NET, 0);
 375                 rc = mds_getattr(req);
 376                 break;
 377
 378         case MDS_READPAGE:
 379                 CDEBUG(D_INODE, "readpage\n");
 380                 OBD_FAIL_RETURN(OBD_FAIL_MDS_READPAGE_NET, 0);
 381                 rc = mds_readpage(req);
 382
 383                 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE))
 384                         return 0;
 385                 break;
 386
 387         case MDS_REINT:
 388                 CDEBUG(D_INODE, "reint\n");
 389                 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET, 0);
 390                 rc = mds_reint(req);
 391                 break;
 392
 393         case MDS_OPEN:
 394                 CDEBUG(D_INODE, "open\n");
 395                 OBD_FAIL_RETURN(OBD_FAIL_MDS_OPEN_NET, 0);
 396                 rc = mds_open(req);
 397                 break;
 398
 399         case MDS_CLOSE:
 400                 CDEBUG(D_INODE, "close\n");
 401                 OBD_FAIL_RETURN(OBD_FAIL_MDS_CLOSE_NET, 0);
 402                 rc = mds_close(req);
 403                 break;
 404
 405         default:
 406                 rc = ptlrpc_error(svc, req);
 407                 RETURN(rc);
 408         }
 409
 410         EXIT;
 411 out:
 412         if (rc) {
 413                 ptlrpc_error(svc, req);
 414         } else {
 415                 CDEBUG(D_NET, "sending reply\n");
 416                 ptlrpc_reply(svc, req);
 417         }
 418
 419         return 0;
 420 }
 421
 422 static int mds_prep(struct obd_device *obddev)
 423 {
 424         struct obd_run_ctxt saved;
 425         struct mds_obd *mds = &obddev->u.mds;
 426         struct super_operations *s_ops;
 427         int err;
 428
 429         mds->mds_service = ptlrpc_init_svc(128 * 1024,
 430                                            MDS_REQUEST_PORTAL, MDC_REPLY_PORTAL,
 431                                            "self", mds_handle);
 432
 433         if (!mds->mds_service) {
 434                 CERROR("failed to start service\n");
 435                 RETURN(-EINVAL);
 436         }
 437
 438         err = ptlrpc_start_thread(obddev, mds->mds_service, "lustre_mds");
 439         if (err) {
 440                 CERROR("cannot start thread\n");
 441                 GOTO(err_svc, err);
 442         }
 443
 444         push_ctxt(&saved, &mds->mds_ctxt);
 445         err = simple_mkdir(current->fs->pwd, "ROOT", 0700);
 446         err = simple_mkdir(current->fs->pwd, "FH", 0700);
 447         pop_ctxt(&saved);
 448
 449         /*
 450          * Replace the client filesystem delete_inode method with our own,
 451          * so that we can clear the object ID before the inode is deleted.
 452          * The fs_delete_inode method will call cl_delete_inode for us.
 453          *
 454          * We need to do this for the MDS superblock only, hence we install
 455          * a modified copy of the original superblock method table.
 456          *
 457          * We still assume that there is only a single MDS client filesystem
 458          * type, as we don't have access to the mds struct in * delete_inode.
 459          */
 460         OBD_ALLOC(s_ops, sizeof(*s_ops));
 461         memcpy(s_ops, mds->mds_sb->s_op, sizeof(*s_ops));
 462         mds->mds_fsops->cl_delete_inode = s_ops->delete_inode;
 463         s_ops->delete_inode = mds->mds_fsops->fs_delete_inode;
 464         mds->mds_sb->s_op = s_ops;
 465
 466         RETURN(0);
 467
 468 err_svc:
 469         rpc_unregister_service(mds->mds_service);
 470         OBD_FREE(mds->mds_service, sizeof(*mds->mds_service));
 471
 472         return(err);
 473 }
 474
 475 /* mount the file system (secretly) */
 476 static int mds_setup(struct obd_device *obddev, obd_count len, void *buf)
 477 {
 478         struct obd_ioctl_data* data = buf;
 479         struct mds_obd *mds = &obddev->u.mds;
 480         struct vfsmount *mnt;
 481         int err = 0;
 482         ENTRY;
 483
 484 #ifdef CONFIG_DEV_RDONLY
 485         dev_clear_rdonly(2);
 486 #endif
 487         mds->mds_fstype = strdup(data->ioc_inlbuf2);
 488
 489         if (!strcmp(mds->mds_fstype, "ext3"))
 490                 mds->mds_fsops = &mds_ext3_fs_ops;
 491         else if (!strcmp(mds->mds_fstype, "ext2"))
 492                 mds->mds_fsops = &mds_ext2_fs_ops;
 493         else {
 494                 CERROR("unsupported MDS filesystem type %s\n", mds->mds_fstype);
 495                 GOTO(err_kfree, (err = -EPERM));
 496         }
 497
 498         MOD_INC_USE_COUNT;
 499         mnt = do_kern_mount(mds->mds_fstype, 0, data->ioc_inlbuf1, NULL);
 500         if (IS_ERR(mnt)) {
 501                 CERROR("do_kern_mount failed: %d\n", err);
 502                 GOTO(err_dec, err = PTR_ERR(mnt));
 503         }
 504
 505         mds->mds_sb = mnt->mnt_root->d_inode->i_sb;
 506         if (!mds->mds_sb)
 507                 GOTO(err_put, (err = -ENODEV));
 508
 509         mds->mds_vfsmnt = mnt;
 510         mds->mds_ctxt.pwdmnt = mnt;
 511         mds->mds_ctxt.pwd = mnt->mnt_root;
 512         mds->mds_ctxt.fs = KERNEL_DS;
 513
 514         err = mds_prep(obddev);
 515         if (err)
 516                 GOTO(err_put, err);
 517
 518         RETURN(0);
 519
 520 err_put:
 521         unlock_kernel();
 522         mntput(mds->mds_vfsmnt);
 523         mds->mds_sb = 0;
 524         lock_kernel();
 525 err_dec:
 526         MOD_DEC_USE_COUNT;
 527 err_kfree:
 528         kfree(mds->mds_fstype);
 529         return err;
 530 }
 531
 532 static int mds_cleanup(struct obd_device * obddev)
 533 {
 534         struct super_operations *s_ops = NULL;
 535         struct super_block *sb;
 536         struct mds_obd *mds = &obddev->u.mds;
 537
 538         ENTRY;
 539
 540         if ( !list_empty(&obddev->obd_gen_clients) ) {
 541                 CERROR("still has clients!\n");
 542                 RETURN(-EBUSY);
 543         }
 544
 545         ptlrpc_stop_thread(mds->mds_service);
 546         rpc_unregister_service(mds->mds_service);
 547         if (!list_empty(&mds->mds_service->srv_reqs)) {
 548                 // XXX reply with errors and clean up
 549                 CERROR("Request list not empty!\n");
 550         }
 551         OBD_FREE(mds->mds_service, sizeof(*mds->mds_service));
 552
 553         sb = mds->mds_sb;
 554         if (!mds->mds_sb)
 555                 RETURN(0);
 556
 557         s_ops = sb->s_op;
 558
 559         unlock_kernel();
 560         mntput(mds->mds_vfsmnt);
 561         mds->mds_sb = 0;
 562         kfree(mds->mds_fstype);
 563         lock_kernel();
 564 #ifdef CONFIG_DEV_RDONLY
 565         dev_clear_rdonly(2);
 566 #endif
 567         OBD_FREE(s_ops, sizeof(*s_ops));
 568
 569         MOD_DEC_USE_COUNT;
 570         RETURN(0);
 571 }
 572
 573 /* use obd ops to offer management infrastructure */
 574 static struct obd_ops mds_obd_ops = {
 575         o_setup:       mds_setup,
 576         o_cleanup:     mds_cleanup,
 577 };
 578
 579 static int __init mds_init(void)
 580 {
 581         obd_register_type(&mds_obd_ops, LUSTRE_MDS_NAME);
 582         return 0;
 583 }
 584
 585 static void __exit mds_exit(void)
 586 {
 587         obd_unregister_type(LUSTRE_MDS_NAME);
 588 }
 589
 590 MODULE_AUTHOR("Peter J. Braam <braam@clusterfs.com>");
 591 MODULE_DESCRIPTION("Lustre Metadata Server (MDS) v0.01");
 592 MODULE_LICENSE("GPL");
 593
 594 module_init(mds_init);
 595 module_exit(mds_exit);