lustre/mds/handler.c

   1 /*
   2  *  linux/mds/handler.c
   3  *
   4  *  Lustre Metadata Server (mds) request handler
   5  *
   6  *  Copyright (C) 2001, 2002 Cluster File Systems, Inc.
   7  *
   8  *  This code is issued under the GNU General Public License.
   9  *  See the file COPYING in this distribution
  10  *
  11  *  by Peter Braam <braam@clusterfs.com>
  12  *
  13  *  This server is single threaded at present (but can easily be multi threaded).
  14  *
  15  */
  16
  17
  18 #define EXPORT_SYMTAB
  19
  20 #include <linux/version.h>
  21 #include <linux/module.h>
  22 #include <linux/fs.h>
  23 #include <linux/stat.h>
  24 #include <linux/locks.h>
  25 #include <linux/ext2_fs.h>
  26 #include <linux/quotaops.h>
  27 #include <asm/unistd.h>
  28 #include <asm/uaccess.h>
  29 #include <linux/obd_support.h>
  30 #include <linux/obd.h>
  31 #include <linux/lustre_lib.h>
  32 #include <linux/lustre_idl.h>
  33 #include <linux/lustre_mds.h>
  34 #include <linux/lustre_net.h>
  35 #include <linux/obd_class.h>
  36
  37 // XXX for testing
  38 static struct mds_obd *MDS;
  39
  40 // XXX make this networked!
  41 static int mds_queue_req(struct ptlrpc_request *req)
  42 {
  43         struct ptlrpc_request *srv_req;
  44
  45         if (!MDS) {
  46                 EXIT;
  47                 return -1;
  48         }
  49
  50         srv_req = kmalloc(sizeof(*srv_req), GFP_KERNEL);
  51         if (!srv_req) {
  52                 EXIT;
  53                 return -ENOMEM;
  54         }
  55
  56         printk("---> MDS at %d %p, incoming req %p, srv_req %p\n",
  57                __LINE__, MDS, req, srv_req);
  58
  59         memset(srv_req, 0, sizeof(*req));
  60
  61         /* move the request buffer */
  62         srv_req->rq_reqbuf = req->rq_reqbuf;
  63         srv_req->rq_reqlen    = req->rq_reqlen;
  64         srv_req->rq_obd = MDS;
  65
  66         /* remember where it came from */
  67         srv_req->rq_reply_handle = req;
  68
  69         list_add(&srv_req->rq_list, &MDS->mds_reqs);
  70         wake_up(&MDS->mds_waitq);
  71         return 0;
  72 }
  73
  74 /* XXX do this over the net */
  75 int mds_sendpage(struct ptlrpc_request *req, struct file *file,
  76                     __u64 offset, struct niobuf *dst)
  77 {
  78         int rc;
  79         mm_segment_t oldfs = get_fs();
  80
  81         if (req->rq_peer.peer_nid == 0) {
  82                 /* dst->addr is a user address, but in a different task! */
  83                 set_fs(KERNEL_DS);
  84                 rc = generic_file_read(file, (char *)(long)dst->addr,
  85                                        PAGE_SIZE, &offset);
  86                 set_fs(oldfs);
  87
  88                 if (rc != PAGE_SIZE)
  89                         return -EIO;
  90         } else {
  91                 char *buf;
  92
  93                 buf = kmalloc(PAGE_SIZE, GFP_KERNEL);
  94                 if (!buf) {
  95                         return -ENOMEM;
  96                 }
  97
  98                 set_fs(KERNEL_DS);
  99                 rc = generic_file_read(file, buf, PAGE_SIZE, &offset);
 100                 set_fs(oldfs);
 101
 102                 if (rc != PAGE_SIZE)
 103                         return -EIO;
 104
 105                 req->rq_bulkbuf = buf;
 106                 req->rq_bulklen = PAGE_SIZE;
 107                 rc = ptl_send_buf(req, &req->rq_peer, MDS_BULK_PORTAL, 0);
 108                 init_waitqueue_head(&req->rq_wait_for_bulk);
 109                 sleep_on(&req->rq_wait_for_bulk);
 110                 kfree(buf);
 111                 req->rq_bulklen = 0; /* FIXME: eek. */
 112         }
 113
 114         return 0;
 115 }
 116
 117 /* XXX replace with networking code */
 118 int mds_reply(struct ptlrpc_request *req)
 119 {
 120         struct ptlrpc_request *clnt_req = req->rq_reply_handle;
 121
 122         ENTRY;
 123
 124         if (req->rq_obd->mds_service != NULL) {
 125                 /* This is a request that came from the network via portals. */
 126
 127                 /* FIXME: we need to increment the count of handled events */
 128                 ptl_send_buf(req, &req->rq_peer, MDS_REPLY_PORTAL, 0);
 129         } else {
 130                 /* This is a local request that came from another thread. */
 131
 132                 /* move the reply to the client */
 133                 clnt_req->rq_replen = req->rq_replen;
 134                 clnt_req->rq_repbuf = req->rq_repbuf;
 135                 req->rq_repbuf = NULL;
 136                 req->rq_replen = 0;
 137
 138                 /* free the request buffer */
 139                 kfree(req->rq_reqbuf);
 140                 req->rq_reqbuf = NULL;
 141
 142                 /* wake up the client */
 143                 wake_up_interruptible(&clnt_req->rq_wait_for_rep);
 144         }
 145
 146         EXIT;
 147         return 0;
 148 }
 149
 150 int mds_error(struct ptlrpc_request *req)
 151 {
 152         struct ptlrep_hdr *hdr;
 153
 154         ENTRY;
 155
 156         hdr = kmalloc(sizeof(*hdr), GFP_KERNEL);
 157         if (!hdr) {
 158                 EXIT;
 159                 return -ENOMEM;
 160         }
 161
 162         memset(hdr, 0, sizeof(*hdr));
 163
 164         hdr->seqno = req->rq_reqhdr->seqno;
 165         hdr->status = req->rq_status;
 166         hdr->type = MDS_TYPE_ERR;
 167
 168         req->rq_repbuf = (char *)hdr;
 169         req->rq_replen = sizeof(*hdr);
 170
 171         EXIT;
 172         return mds_reply(req);
 173 }
 174
 175 struct dentry *mds_fid2dentry(struct mds_obd *mds, struct ll_fid *fid, struct vfsmount **mnt)
 176 {
 177         /* stolen from NFS */
 178         struct super_block *sb = mds->mds_sb;
 179         unsigned long ino = fid->id;
 180         //__u32 generation = fid->generation;
 181         __u32 generation = 0;
 182         struct inode *inode;
 183         struct list_head *lp;
 184         struct dentry *result;
 185
 186         if (mnt) {
 187                 *mnt = mntget(mds->mds_vfsmnt);
 188         }
 189
 190         if (ino == 0)
 191                 return ERR_PTR(-ESTALE);
 192
 193         inode = iget(sb, ino);
 194         if (inode == NULL)
 195                 return ERR_PTR(-ENOMEM);
 196
 197         printk("--> mds_fid2dentry: sb %p\n", inode->i_sb);
 198
 199         if (is_bad_inode(inode)
 200             || (generation && inode->i_generation != generation)
 201                 ) {
 202                 /* we didn't find the right inode.. */
 203                 printk(__FUNCTION__
 204                        "bad inode %lu, link: %d ct: %d or version  %u/%u\n",
 205                         inode->i_ino,
 206                         inode->i_nlink, atomic_read(&inode->i_count),
 207                         inode->i_generation,
 208                         generation);
 209                 iput(inode);
 210                 return ERR_PTR(-ESTALE);
 211         }
 212
 213         /* now to find a dentry.
 214          * If possible, get a well-connected one
 215          */
 216         spin_lock(&dcache_lock);
 217         for (lp = inode->i_dentry.next; lp != &inode->i_dentry ; lp=lp->next) {
 218                 result = list_entry(lp,struct dentry, d_alias);
 219                 if (! (result->d_flags & DCACHE_NFSD_DISCONNECTED)) {
 220                         dget_locked(result);
 221                         result->d_vfs_flags |= DCACHE_REFERENCED;
 222                         spin_unlock(&dcache_lock);
 223                         iput(inode);
 224                         return result;
 225                 }
 226         }
 227         spin_unlock(&dcache_lock);
 228         result = d_alloc_root(inode);
 229         if (result == NULL) {
 230                 iput(inode);
 231                 return ERR_PTR(-ENOMEM);
 232         }
 233         result->d_flags |= DCACHE_NFSD_DISCONNECTED;
 234         return result;
 235 }
 236
 237 static inline void mds_get_objid(struct inode *inode, __u64 *id)
 238 {
 239         memcpy(id, &inode->u.ext2_i.i_data, sizeof(*id));
 240 }
 241
 242 int mds_getattr(struct ptlrpc_request *req)
 243 {
 244         struct dentry *de = mds_fid2dentry(req->rq_obd, &req->rq_req.mds->fid1,
 245                                            NULL);
 246         struct inode *inode;
 247         struct mds_rep *rep;
 248         int rc;
 249
 250         rc = mds_pack_rep(NULL, 0, NULL, 0, &req->rq_rephdr, &req->rq_rep.mds,
 251                           &req->rq_replen, &req->rq_repbuf);
 252         if (rc) {
 253                 EXIT;
 254                 printk("mds: out of memory\n");
 255                 req->rq_status = -ENOMEM;
 256                 return 0;
 257         }
 258
 259         req->rq_rephdr->seqno = req->rq_reqhdr->seqno;
 260         rep = req->rq_rep.mds;
 261
 262         if (!de) {
 263                 EXIT;
 264                 req->rq_rephdr->status = -ENOENT;
 265                 return 0;
 266         }
 267
 268         inode = de->d_inode;
 269         rep->ino = inode->i_ino;
 270         rep->atime = inode->i_atime;
 271         rep->ctime = inode->i_ctime;
 272         rep->mtime = inode->i_mtime;
 273         rep->uid = inode->i_uid;
 274         rep->gid = inode->i_gid;
 275         rep->size = inode->i_size;
 276         rep->mode = inode->i_mode;
 277         rep->nlink = inode->i_nlink;
 278         rep->valid = ~0;
 279         mds_get_objid(inode, &rep->objid);
 280         dput(de);
 281         return 0;
 282 }
 283
 284 int mds_readpage(struct ptlrpc_request *req)
 285 {
 286         struct vfsmount *mnt;
 287         struct dentry *de = mds_fid2dentry(req->rq_obd, &req->rq_req.mds->fid1,
 288                                            &mnt);
 289         struct file *file;
 290         struct niobuf *niobuf;
 291         struct mds_rep *rep;
 292         int rc;
 293
 294         printk("mds_readpage: ino %ld\n", de->d_inode->i_ino);
 295         rc = mds_pack_rep(NULL, 0, NULL, 0, &req->rq_rephdr, &req->rq_rep.mds,
 296                           &req->rq_replen, &req->rq_repbuf);
 297         if (rc) {
 298                 EXIT;
 299                 printk("mds: out of memory\n");
 300                 req->rq_status = -ENOMEM;
 301                 return 0;
 302         }
 303
 304         req->rq_rephdr->seqno = req->rq_reqhdr->seqno;
 305         rep = req->rq_rep.mds;
 306
 307         if (IS_ERR(de)) {
 308                 EXIT;
 309                 req->rq_rephdr->status = PTR_ERR(de);
 310                 return 0;
 311         }
 312
 313         file = dentry_open(de, mnt, O_RDONLY | O_LARGEFILE);
 314         /* note: in case of an error, dentry_open puts dentry */
 315         if (IS_ERR(file)) {
 316                 EXIT;
 317                 req->rq_rephdr->status = PTR_ERR(file);
 318                 return 0;
 319         }
 320
 321         niobuf = mds_req_tgt(req->rq_req.mds);
 322
 323         /* to make this asynchronous make sure that the handling function
 324            doesn't send a reply when this function completes. Instead a
 325            callback function would send the reply */
 326         rc = mds_sendpage(req, file, req->rq_req.mds->size, niobuf);
 327
 328         filp_close(file, 0);
 329         req->rq_rephdr->status = rc;
 330         EXIT;
 331         return 0;
 332 }
 333
 334 int mds_reint(struct ptlrpc_request *req)
 335 {
 336         int rc;
 337         char *buf = mds_req_tgt(req->rq_req.mds);
 338         int len = req->rq_req.mds->tgtlen;
 339         struct mds_update_record rec;
 340
 341         rc = mds_update_unpack(buf, len, &rec);
 342         if (rc) {
 343                 printk(__FUNCTION__ ": 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 mds_conn *conn, int len, char *buf)
 353 int mds_handle(struct ptlrpc_request *req)
 354 {
 355         int rc;
 356         struct ptlreq_hdr *hdr;
 357
 358         ENTRY;
 359
 360         hdr = (struct ptlreq_hdr *)req->rq_reqbuf;
 361
 362         if (NTOH__u32(hdr->type) != MDS_TYPE_REQ) {
 363                 printk("lustre_mds: wrong packet type sent %d\n",
 364                        NTOH__u32(hdr->type));
 365                 rc = -EINVAL;
 366                 goto out;
 367         }
 368
 369         rc = mds_unpack_req(req->rq_reqbuf, req->rq_reqlen,
 370                             &req->rq_reqhdr, &req->rq_req.mds);
 371         if (rc) {
 372                 printk("lustre_mds: Invalid request\n");
 373                 EXIT;
 374                 goto out;
 375         }
 376
 377         switch (req->rq_reqhdr->opc) {
 378
 379         case MDS_GETATTR:
 380                 CDEBUG(D_INODE, "getattr\n");
 381                 rc = mds_getattr(req);
 382                 break;
 383
 384         case MDS_READPAGE:
 385                 CDEBUG(D_INODE, "readpage\n");
 386                 rc = mds_readpage(req);
 387                 break;
 388
 389         case MDS_REINT:
 390                 CDEBUG(D_INODE, "reint\n");
 391                 rc = mds_reint(req);
 392                 break;
 393
 394         default:
 395                 return mds_error(req);
 396         }
 397
 398 out:
 399         if (rc) {
 400                 printk(__FUNCTION__ ": no header\n");
 401                 return 0;
 402         }
 403
 404         if( req->rq_status) {
 405                 mds_error(req);
 406         } else {
 407                 CDEBUG(D_INODE, "sending reply\n");
 408                 mds_reply(req);
 409         }
 410
 411         return 0;
 412 }
 413
 414
 415 static void mds_timer_run(unsigned long __data)
 416 {
 417         struct task_struct * p = (struct task_struct *) __data;
 418
 419         wake_up_process(p);
 420 }
 421
 422 int mds_main(void *arg)
 423 {
 424         struct mds_obd *mds = (struct mds_obd *) arg;
 425         struct timer_list timer;
 426
 427         lock_kernel();
 428         daemonize();
 429         spin_lock_irq(&current->sigmask_lock);
 430         sigfillset(&current->blocked);
 431         recalc_sigpending(current);
 432         spin_unlock_irq(&current->sigmask_lock);
 433
 434         sprintf(current->comm, "lustre_mds");
 435
 436         /* Set up an interval timer which can be used to trigger a
 437            wakeup after the interval expires */
 438         init_timer(&timer);
 439         timer.data = (unsigned long) current;
 440         timer.function = mds_timer_run;
 441         mds->mds_timer = &timer;
 442
 443         /* Record that the  thread is running */
 444         mds->mds_thread = current;
 445         wake_up(&mds->mds_done_waitq);
 446
 447         printk(KERN_INFO "lustre_mds starting.  Commit interval %d seconds\n",
 448                         mds->mds_interval / HZ);
 449
 450         /* XXX maintain a list of all managed devices: insert here */
 451
 452         /* And now, wait forever for commit wakeup events. */
 453         while (1) {
 454                 int rc;
 455
 456                 if (mds->mds_flags & MDS_UNMOUNT)
 457                         break;
 458
 459                 wake_up(&mds->mds_done_waitq);
 460                 interruptible_sleep_on(&mds->mds_waitq);
 461
 462                 CDEBUG(D_INODE, "lustre_mds wakes\n");
 463                 CDEBUG(D_INODE, "pick up req here and continue\n");
 464
 465                 if (mds->mds_service != NULL) {
 466                         ptl_event_t ev;
 467
 468                         while (1) {
 469                                 struct ptlrpc_request request;
 470
 471                                 rc = PtlEQGet(mds->mds_service->srv_eq, &ev);
 472                                 if (rc != PTL_OK && rc != PTL_EQ_DROPPED)
 473                                         break;
 474                                 /* FIXME: If we move to an event-driven model,
 475                                  * we should put the request on the stack of
 476                                  * mds_handle instead. */
 477                                 memset(&request, 0, sizeof(request));
 478                                 request.rq_reqbuf = ev.mem_desc.start +
 479                                         ev.offset;
 480                                 request.rq_reqlen = ev.mem_desc.length;
 481                                 request.rq_obd = MDS;
 482                                 request.rq_xid = ev.match_bits;
 483
 484                                 request.rq_peer.peer_nid = ev.initiator.nid;
 485                                 /* FIXME: this NI should be the incoming NI.
 486                                  * We don't know how to find that from here. */
 487                                 request.rq_peer.peer_ni =
 488                                         mds->mds_service->srv_self.peer_ni;
 489                                 rc = mds_handle(&request);
 490                         }
 491                 } else {
 492                         struct ptlrpc_request *request;
 493
 494                         if (list_empty(&mds->mds_reqs)) {
 495                                 CDEBUG(D_INODE, "woke because of timer\n");
 496                         } else {
 497                                 request = list_entry(mds->mds_reqs.next,
 498                                                      struct ptlrpc_request,
 499                                                      rq_list);
 500                                 list_del(&request->rq_list);
 501                                 rc = mds_handle(request);
 502                         }
 503                 }
 504         }
 505
 506         del_timer_sync(mds->mds_timer);
 507
 508         /* XXX maintain a list of all managed devices: cleanup here */
 509
 510         mds->mds_thread = NULL;
 511         wake_up(&mds->mds_done_waitq);
 512         printk("lustre_mds: exiting\n");
 513         return 0;
 514 }
 515
 516 static void mds_stop_srv_thread(struct mds_obd *mds)
 517 {
 518         mds->mds_flags |= MDS_UNMOUNT;
 519
 520         while (mds->mds_thread) {
 521                 wake_up(&mds->mds_waitq);
 522                 sleep_on(&mds->mds_done_waitq);
 523         }
 524 }
 525
 526 static void mds_start_srv_thread(struct mds_obd *mds)
 527 {
 528         init_waitqueue_head(&mds->mds_waitq);
 529         init_waitqueue_head(&mds->mds_done_waitq);
 530         kernel_thread(mds_main, (void *)mds,
 531                       CLONE_VM | CLONE_FS | CLONE_FILES);
 532         while (!mds->mds_thread)
 533                 sleep_on(&mds->mds_done_waitq);
 534 }
 535
 536 /* mount the file system (secretly) */
 537 static int mds_setup(struct obd_device *obddev, obd_count len,
 538                         void *buf)
 539
 540 {
 541         struct obd_ioctl_data* data = buf;
 542         struct mds_obd *mds = &obddev->u.mds;
 543         struct vfsmount *mnt;
 544         struct lustre_peer peer;
 545         int err;
 546         ENTRY;
 547
 548
 549         mnt = do_kern_mount(data->ioc_inlbuf2, 0, data->ioc_inlbuf1, NULL);
 550         err = PTR_ERR(mnt);
 551         if (IS_ERR(mnt)) {
 552                 EXIT;
 553                 return err;
 554         }
 555
 556         mds->mds_sb = mnt->mnt_root->d_inode->i_sb;
 557         if (!obddev->u.mds.mds_sb) {
 558                 EXIT;
 559                 return -ENODEV;
 560         }
 561
 562         mds->mds_vfsmnt = mnt;
 563         obddev->u.mds.mds_fstype = strdup(data->ioc_inlbuf2);
 564
 565         mds->mds_ctxt.pwdmnt = mnt;
 566         mds->mds_ctxt.pwd = mnt->mnt_root;
 567         mds->mds_ctxt.fs = KERNEL_DS;
 568         mds->mds_remote_nid = 0;
 569
 570         INIT_LIST_HEAD(&mds->mds_reqs);
 571         mds->mds_thread = NULL;
 572         mds->mds_flags = 0;
 573         mds->mds_interval = 3 * HZ;
 574         MDS = mds;
 575
 576         spin_lock_init(&obddev->u.mds.mds_lock);
 577
 578         err = kportal_uuid_to_peer("self", &peer);
 579         if (err == 0) {
 580                 mds->mds_service = kmalloc(sizeof(*mds->mds_service),
 581                                                   GFP_KERNEL);
 582                 if (mds->mds_service == NULL)
 583                         return -ENOMEM;
 584                 mds->mds_service->srv_buf_size = 64 * 1024;
 585                 mds->mds_service->srv_portal = MDS_REQUEST_PORTAL;
 586                 memcpy(&mds->mds_service->srv_self, &peer, sizeof(peer));
 587                 mds->mds_service->srv_wait_queue = &mds->mds_waitq;
 588
 589                 rpc_register_service(mds->mds_service, "self");
 590         }
 591
 592         mds_start_srv_thread(mds);
 593
 594         MOD_INC_USE_COUNT;
 595         EXIT;
 596         return 0;
 597 }
 598
 599 static int mds_cleanup(struct obd_device * obddev)
 600 {
 601         struct super_block *sb;
 602         struct mds_obd *mds = &obddev->u.mds;
 603
 604         ENTRY;
 605
 606         if ( !(obddev->obd_flags & OBD_SET_UP) ) {
 607                 EXIT;
 608                 return 0;
 609         }
 610
 611         if ( !list_empty(&obddev->obd_gen_clients) ) {
 612                 printk(KERN_WARNING __FUNCTION__ ": still has clients!\n");
 613                 EXIT;
 614                 return -EBUSY;
 615         }
 616
 617         MDS = NULL;
 618         mds_stop_srv_thread(mds);
 619         sb = mds->mds_sb;
 620         if (!mds->mds_sb){
 621                 EXIT;
 622                 return 0;
 623         }
 624
 625         if (!list_empty(&mds->mds_reqs)) {
 626                 // XXX reply with errors and clean up
 627                 CDEBUG(D_INODE, "Request list not empty!\n");
 628         }
 629
 630         unlock_kernel();
 631         mntput(mds->mds_vfsmnt);
 632         mds->mds_sb = 0;
 633         kfree(mds->mds_fstype);
 634         lock_kernel();
 635
 636
 637         MOD_DEC_USE_COUNT;
 638         EXIT;
 639         return 0;
 640 }
 641
 642 /* use obd ops to offer management infrastructure */
 643 static struct obd_ops mds_obd_ops = {
 644         o_setup:       mds_setup,
 645         o_cleanup:     mds_cleanup,
 646 };
 647
 648 static int __init mds_init(void)
 649 {
 650         obd_register_type(&mds_obd_ops, LUSTRE_MDS_NAME);
 651         return 0;
 652 }
 653
 654 static void __exit mds_exit(void)
 655 {
 656         obd_unregister_type(LUSTRE_MDS_NAME);
 657 }
 658
 659 MODULE_AUTHOR("Peter J. Braam <braam@clusterfs.com>");
 660 MODULE_DESCRIPTION("Lustre Metadata Server (MDS) v0.01");
 661 MODULE_LICENSE("GPL");
 662
 663
 664 // for testing (maybe this stays)
 665 EXPORT_SYMBOL(mds_queue_req);
 666
 667 module_init(mds_init);
 668 module_exit(mds_exit);