lustre/mds/handler.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  *  lustre/mds/handler.c
   5  *  Lustre Metadata Server (mds) request handler
   6  *
   7  *  Copyright (c) 2001-2003 Cluster File Systems, Inc.
   8  *   Author: Peter Braam <braam@clusterfs.com>
   9  *   Author: Andreas Dilger <adilger@clusterfs.com>
  10  *   Author: Phil Schwan <phil@clusterfs.com>
  11  *   Author: Mike Shaver <shaver@clusterfs.com>
  12  *
  13  *   This file is part of Lustre, http://www.lustre.org.
  14  *
  15  *   Lustre is free software; you can redistribute it and/or
  16  *   modify it under the terms of version 2 of the GNU General Public
  17  *   License as published by the Free Software Foundation.
  18  *
  19  *   Lustre is distributed in the hope that it will be useful,
  20  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  21  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  22  *   GNU General Public License for more details.
  23  *
  24  *   You should have received a copy of the GNU General Public License
  25  *   along with Lustre; if not, write to the Free Software
  26  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  27  */
  28
  29 #ifndef EXPORT_SYMTAB
  30 # define EXPORT_SYMTAB
  31 #endif
  32 #define DEBUG_SUBSYSTEM S_MDS
  33
  34 #include <linux/module.h>
  35 #include <linux/lustre_mds.h>
  36 #include <linux/lustre_dlm.h>
  37 #include <linux/init.h>
  38 #include <linux/obd_class.h>
  39 #include <linux/random.h>
  40 #include <linux/fs.h>
  41 #include <linux/jbd.h>
  42 #include <linux/ext3_fs.h>
  43 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0))
  44 # include <linux/smp_lock.h>
  45 # include <linux/buffer_head.h>
  46 # include <linux/workqueue.h>
  47 # include <linux/mount.h>
  48 #else
  49 # include <linux/locks.h>
  50 #endif
  51 #include <linux/obd_lov.h>
  52 #include <linux/lustre_mds.h>
  53 #include <linux/lustre_fsfilt.h>
  54 #include <linux/lprocfs_status.h>
  55 #include <linux/lustre_commit_confd.h>
  56
  57 #include "mds_internal.h"
  58
  59 static int mds_intent_policy(struct ldlm_namespace *ns,
  60                              struct ldlm_lock **lockp, void *req_cookie,
  61                              ldlm_mode_t mode, int flags, void *data);
  62 static int mds_postsetup(struct obd_device *obd);
  63 static int mds_cleanup(struct obd_device *obd, int flags);
  64
  65 /* Assumes caller has already pushed into the kernel filesystem context */
  66 static int mds_sendpage(struct ptlrpc_request *req, struct file *file,
  67                         loff_t offset, int count)
  68 {
  69         struct ptlrpc_bulk_desc *desc;
  70         struct l_wait_info lwi;
  71         struct page **pages;
  72         int rc = 0, npages, i, tmpcount, tmpsize = 0;
  73         ENTRY;
  74
  75         LASSERT((offset & (PAGE_SIZE - 1)) == 0); /* I'm dubious about this */
  76
  77         npages = (count + PAGE_SIZE - 1) >> PAGE_SHIFT;
  78         OBD_ALLOC(pages, sizeof(*pages) * npages);
  79         if (!pages)
  80                 GOTO(out, rc = -ENOMEM);
  81
  82         desc = ptlrpc_prep_bulk_exp(req, npages, BULK_PUT_SOURCE,
  83                                     MDS_BULK_PORTAL);
  84         if (desc == NULL)
  85                 GOTO(out_free, rc = -ENOMEM);
  86
  87         for (i = 0, tmpcount = count; i < npages; i++, tmpcount -= tmpsize) {
  88                 tmpsize = tmpcount > PAGE_SIZE ? PAGE_SIZE : tmpcount;
  89
  90                 pages[i] = alloc_pages(GFP_KERNEL, 0);
  91                 if (pages[i] == NULL)
  92                         GOTO(cleanup_buf, rc = -ENOMEM);
  93
  94                 ptlrpc_prep_bulk_page(desc, pages[i], 0, tmpsize);
  95         }
  96
  97         for (i = 0, tmpcount = count; i < npages; i++, tmpcount -= tmpsize) {
  98                 tmpsize = tmpcount > PAGE_SIZE ? PAGE_SIZE : tmpcount;
  99                 CDEBUG(D_EXT2, "reading %u@%llu from dir %lu (size %llu)\n",
 100                        tmpsize, offset, file->f_dentry->d_inode->i_ino,
 101                        file->f_dentry->d_inode->i_size);
 102
 103                 rc = fsfilt_readpage(req->rq_export->exp_obd, file,
 104                                      kmap(pages[i]), tmpsize, &offset);
 105                 kunmap(pages[i]);
 106
 107                 if (rc != tmpsize)
 108                         GOTO(cleanup_buf, rc = -EIO);
 109         }
 110
 111         LASSERT(desc->bd_nob == count);
 112
 113         rc = ptlrpc_start_bulk_transfer(desc);
 114         if (rc)
 115                 GOTO(cleanup_buf, rc);
 116
 117         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
 118                 CERROR("obd_fail_loc=%x, fail operation rc=%d\n",
 119                        OBD_FAIL_MDS_SENDPAGE, rc);
 120                 GOTO(abort_bulk, rc);
 121         }
 122
 123         lwi = LWI_TIMEOUT(obd_timeout * HZ / 4, NULL, NULL);
 124         rc = l_wait_event(desc->bd_waitq, !ptlrpc_bulk_active(desc), &lwi);
 125         LASSERT (rc == 0 || rc == -ETIMEDOUT);
 126
 127         if (rc == 0) {
 128                 if (desc->bd_success &&
 129                     desc->bd_nob_transferred == count)
 130                         GOTO(cleanup_buf, rc);
 131
 132                 rc = -ETIMEDOUT; /* XXX should this be a different errno? */
 133         }
 134
 135         DEBUG_REQ(D_ERROR, req, "bulk failed: %s %d(%d), evicting %s@%s\n",
 136                   (rc == -ETIMEDOUT) ? "timeout" : "network error",
 137                   desc->bd_nob_transferred, count,
 138                   req->rq_export->exp_client_uuid.uuid,
 139                   req->rq_export->exp_connection->c_remote_uuid.uuid);
 140
 141         ptlrpc_fail_export(req->rq_export);
 142
 143         EXIT;
 144  abort_bulk:
 145         ptlrpc_abort_bulk (desc);
 146  cleanup_buf:
 147         for (i = 0; i < npages; i++)
 148                 if (pages[i])
 149                         __free_pages(pages[i], 0);
 150
 151         ptlrpc_free_bulk(desc);
 152  out_free:
 153         OBD_FREE(pages, sizeof(*pages) * npages);
 154  out:
 155         return rc;
 156 }
 157
 158 int mds_lock_mode_for_dir(struct obd_device *obd,
 159                               struct dentry *dentry, int mode)
 160 {
 161         int ret_mode;
 162
 163         /* any dir access needs couple locks:
 164          * 1) on part of dir we gonna lookup/modify in
 165          * 2) on a whole dir to protect it from concurrent splitting
 166          *    and to flush client's cache for readdir()
 167          * so, for a given mode and dentry this routine decides what
 168          * lock mode to use for lock #2:
 169          * 1) if caller's gonna lookup in dir then we need to protect
 170          *    dir from being splitted only - LCK_CR
 171          * 2) if caller's gonna modify dir then we need to protect
 172          *    dir from being splitted and to flush cache - LCK_CW
 173          * 3) if caller's gonna modify dir and that dir seems ready
 174          *    for splitting then we need to protect it from any
 175          *    type of access (lookup/modify/split) - LCK_EX -bzzz */
 176
 177         if (mode == LCK_PR) {
 178                 ret_mode = LCK_CR;
 179         } else if (mode == LCK_PW) {
 180                 /* caller gonna modify directory.we use concurrent
 181                    write lock here to retract client's cache for readdir */
 182                 ret_mode = LCK_CW;
 183                 if (mds_splitting_expected(obd, dentry)) {
 184                         /* splitting possible. serialize any access */
 185                         CERROR("%s: gonna split %lu/%lu\n",
 186                                obd->obd_name,
 187                                (unsigned long) dentry->d_inode->i_ino,
 188                                (unsigned long) dentry->d_inode->i_generation);
 189                         ret_mode = LCK_EX;
 190                 }
 191         } else {
 192                 CWARN("unexpected lock mode %d\n", mode);
 193                 ret_mode = LCK_EX;
 194         }
 195         return ret_mode;
 196 }
 197
 198 /* only valid locked dentries or errors should be returned */
 199 struct dentry *mds_fid2locked_dentry(struct obd_device *obd, struct ll_fid *fid,
 200                                      struct vfsmount **mnt, int lock_mode,
 201                                      struct lustre_handle *lockh, int *mode,
 202                                      char *name, int namelen, __u64 lockpart)
 203 {
 204         struct mds_obd *mds = &obd->u.mds;
 205         struct dentry *de = mds_fid2dentry(mds, fid, mnt), *retval = de;
 206         struct ldlm_res_id res_id = { .name = {0} };
 207         int flags = 0, rc;
 208         ldlm_policy_data_t policy = { .l_inodebits = { lockpart } };
 209
 210         ENTRY;
 211
 212         if (IS_ERR(de))
 213                 RETURN(de);
 214
 215         res_id.name[0] = de->d_inode->i_ino;
 216         res_id.name[1] = de->d_inode->i_generation;
 217         lockh[1].cookie = 0;
 218 #ifdef S_PDIROPS
 219         if (name && IS_PDIROPS(de->d_inode)) {
 220                 ldlm_policy_data_t cpolicy =
 221                         { .l_inodebits = { MDS_INODELOCK_UPDATE } };
 222                 LASSERT(mode != NULL);
 223                 *mode = mds_lock_mode_for_dir(obd, de, lock_mode);
 224                 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
 225                                 res_id, LDLM_IBITS,
 226                                 &cpolicy, *mode, &flags,
 227                                 mds_blocking_ast,
 228                                 ldlm_completion_ast, NULL, NULL,
 229                                 NULL, 0, NULL, lockh + 1);
 230                 if (rc != ELDLM_OK) {
 231                         l_dput(de);
 232                         RETURN(ERR_PTR(-ENOLCK));
 233                 }
 234                 flags = 0;
 235
 236                 res_id.name[2] = full_name_hash(name, namelen);
 237
 238                 CDEBUG(D_INFO, "take lock on %lu:%u:"LPX64"\n",
 239                        de->d_inode->i_ino, de->d_inode->i_generation,
 240                        res_id.name[2]);
 241         }
 242 #endif
 243         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, res_id,
 244                               LDLM_IBITS, &policy, lock_mode, &flags,
 245                               mds_blocking_ast, ldlm_completion_ast, NULL, NULL,
 246                               NULL, 0, NULL, lockh);
 247         if (rc != ELDLM_OK) {
 248                 l_dput(de);
 249                 retval = ERR_PTR(-EIO); /* XXX translate ldlm code */
 250 #ifdef S_PDIROPS
 251                 if (lockh[1].cookie)
 252                         ldlm_lock_decref(lockh + 1, LCK_CW);
 253 #endif
 254         }
 255
 256         RETURN(retval);
 257 }
 258
 259 #ifndef DCACHE_DISCONNECTED
 260 #define DCACHE_DISCONNECTED DCACHE_NFSD_DISCONNECTED
 261 #endif
 262
 263
 264 /* Look up an entry by inode number. */
 265 /* this function ONLY returns valid dget'd dentries with an initialized inode
 266    or errors */
 267 struct dentry *mds_fid2dentry(struct mds_obd *mds, struct ll_fid *fid,
 268                               struct vfsmount **mnt)
 269 {
 270         char fid_name[32];
 271         unsigned long ino = fid->id;
 272         __u32 generation = fid->generation;
 273         struct inode *inode;
 274         struct dentry *result;
 275
 276         if (ino == 0)
 277                 RETURN(ERR_PTR(-ESTALE));
 278
 279         snprintf(fid_name, sizeof(fid_name), "0x%lx", ino);
 280
 281         CDEBUG(D_DENTRY, "--> mds_fid2dentry: ino/gen %lu/%u, sb %p\n",
 282                ino, generation, mds->mds_sb);
 283
 284         /* under ext3 this is neither supposed to return bad inodes
 285            nor NULL inodes. */
 286         result = ll_lookup_one_len(fid_name, mds->mds_fid_de, strlen(fid_name));
 287         if (IS_ERR(result))
 288                 RETURN(result);
 289
 290         inode = result->d_inode;
 291         if (!inode)
 292                 RETURN(ERR_PTR(-ENOENT));
 293
 294 #if 0
 295         /* here we disabled generation check, as root inode i_generation
 296          * of cache mds and real mds are different. */
 297         if (generation && inode->i_generation != generation) {
 298                 /* we didn't find the right inode.. */
 299                 CERROR("bad inode %lu, link: %lu ct: %d or generation %u/%u\n",
 300                        inode->i_ino, (unsigned long)inode->i_nlink,
 301                        atomic_read(&inode->i_count), inode->i_generation,
 302                        generation);
 303                 dput(result);
 304                 RETURN(ERR_PTR(-ENOENT));
 305         }
 306 #endif
 307
 308         if (mnt) {
 309                 *mnt = mds->mds_vfsmnt;
 310                 mntget(*mnt);
 311         }
 312
 313         RETURN(result);
 314 }
 315
 316
 317 /* Establish a connection to the MDS.
 318  *
 319  * This will set up an export structure for the client to hold state data
 320  * about that client, like open files, the last operation number it did
 321  * on the server, etc.
 322  */
 323 static int mds_connect(struct lustre_handle *conn, struct obd_device *obd,
 324                        struct obd_uuid *cluuid)
 325 {
 326         struct obd_export *exp;
 327         struct mds_export_data *med; /*  */
 328         struct mds_client_data *mcd;
 329         int rc, abort_recovery;
 330         ENTRY;
 331
 332         if (!conn || !obd || !cluuid)
 333                 RETURN(-EINVAL);
 334
 335         /* Check for aborted recovery. */
 336         spin_lock_bh(&obd->obd_processing_task_lock);
 337         abort_recovery = obd->obd_abort_recovery;
 338         spin_unlock_bh(&obd->obd_processing_task_lock);
 339         if (abort_recovery)
 340                 target_abort_recovery(obd);
 341
 342         /* XXX There is a small race between checking the list and adding a
 343          * new connection for the same UUID, but the real threat (list
 344          * corruption when multiple different clients connect) is solved.
 345          *
 346          * There is a second race between adding the export to the list,
 347          * and filling in the client data below.  Hence skipping the case
 348          * of NULL mcd above.  We should already be controlling multiple
 349          * connects at the client, and we can't hold the spinlock over
 350          * memory allocations without risk of deadlocking.
 351          */
 352         rc = class_connect(conn, obd, cluuid);
 353         if (rc)
 354                 RETURN(rc);
 355         exp = class_conn2export(conn);
 356         LASSERT(exp);
 357         med = &exp->exp_mds_data;
 358
 359         OBD_ALLOC(mcd, sizeof(*mcd));
 360         if (!mcd) {
 361                 CERROR("mds: out of memory for client data\n");
 362                 GOTO(out, rc = -ENOMEM);
 363         }
 364
 365         memcpy(mcd->mcd_uuid, cluuid, sizeof(mcd->mcd_uuid));
 366         med->med_mcd = mcd;
 367
 368         rc = mds_client_add(obd, &obd->u.mds, med, -1);
 369         if (rc == 0)
 370                 EXIT;
 371 out:
 372         if (rc) {
 373                 OBD_FREE(mcd, sizeof(*mcd));
 374                 class_disconnect(exp, 0);
 375         }
 376         class_export_put(exp);
 377
 378         return rc;
 379 }
 380
 381 static int mds_init_export(struct obd_export *exp)
 382 {
 383         struct mds_export_data *med = &exp->exp_mds_data;
 384
 385         INIT_LIST_HEAD(&med->med_open_head);
 386         spin_lock_init(&med->med_open_lock);
 387         RETURN(0);
 388 }
 389
 390 static int mds_destroy_export(struct obd_export *export)
 391 {
 392         struct mds_export_data *med;
 393         struct obd_device *obd = export->exp_obd;
 394         struct lvfs_run_ctxt saved;
 395         int rc = 0;
 396         ENTRY;
 397
 398         med = &export->exp_mds_data;
 399         target_destroy_export(export);
 400
 401         if (obd_uuid_equals(&export->exp_client_uuid, &obd->obd_uuid))
 402                 GOTO(out, 0);
 403
 404         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 405
 406         /* Close any open files (which may also cause orphan unlinking). */
 407         spin_lock(&med->med_open_lock);
 408         while (!list_empty(&med->med_open_head)) {
 409                 struct list_head *tmp = med->med_open_head.next;
 410                 struct mds_file_data *mfd =
 411                         list_entry(tmp, struct mds_file_data, mfd_list);
 412                 BDEVNAME_DECLARE_STORAGE(btmp);
 413
 414                 /* bug 1579: fix force-closing for 2.5 */
 415                 struct dentry *dentry = mfd->mfd_dentry;
 416
 417                 list_del(&mfd->mfd_list);
 418                 spin_unlock(&med->med_open_lock);
 419
 420                 /* If you change this message, be sure to update
 421                  * replay_single:test_46 */
 422                 CERROR("force closing client file handle for %*s (%s:%lu)\n",
 423                        dentry->d_name.len, dentry->d_name.name,
 424                        ll_bdevname(dentry->d_inode->i_sb, btmp),
 425                        dentry->d_inode->i_ino);
 426                 rc = mds_mfd_close(NULL, obd, mfd,
 427                                    !(export->exp_flags & OBD_OPT_FAILOVER));
 428
 429                 if (rc)
 430                         CDEBUG(D_INODE, "Error closing file: %d\n", rc);
 431                 spin_lock(&med->med_open_lock);
 432         }
 433         spin_unlock(&med->med_open_lock);
 434         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 435
 436 out:
 437         mds_client_free(export, !(export->exp_flags & OBD_OPT_FAILOVER));
 438
 439         RETURN(rc);
 440 }
 441
 442 static int mds_disconnect(struct obd_export *exp, int flags)
 443 {
 444         struct obd_device *obd;
 445         struct mds_obd *mds;
 446         unsigned long irqflags;
 447         int rc;
 448         ENTRY;
 449
 450         LASSERT(exp);
 451         class_export_get(exp);
 452
 453         obd = class_exp2obd(exp);
 454         if (obd == NULL) {
 455                 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
 456                        exp->exp_handle.h_cookie);
 457                 RETURN(-EINVAL);
 458         }
 459         mds = &obd->u.mds;
 460
 461         if (!(exp->exp_flags & OBD_OPT_REAL_CLIENT)
 462                         && !atomic_read(&mds->mds_real_clients)) {
 463                 /* there was no client at all */
 464                 mds_lmv_disconnect(obd, flags);
 465         }
 466
 467         if ((exp->exp_flags & OBD_OPT_REAL_CLIENT)
 468                         && atomic_dec_and_test(&mds->mds_real_clients)) {
 469                 /* time to drop LMV connections */
 470                 CDEBUG(D_OTHER, "%s: last real client %s disconnected.  "
 471                        "Disconnnect from LMV now\n",
 472                        obd->obd_name, exp->exp_client_uuid.uuid);
 473                 mds_lmv_disconnect(obd, flags);
 474         }
 475
 476         spin_lock_irqsave(&exp->exp_lock, irqflags);
 477         exp->exp_flags = flags;
 478         spin_unlock_irqrestore(&exp->exp_lock, irqflags);
 479
 480         /* Disconnect early so that clients can't keep using export */
 481         rc = class_disconnect(exp, flags);
 482         ldlm_cancel_locks_for_export(exp);
 483
 484         /* complete all outstanding replies */
 485         spin_lock_irqsave(&exp->exp_lock, irqflags);
 486         while (!list_empty(&exp->exp_outstanding_replies)) {
 487                 struct ptlrpc_reply_state *rs =
 488                         list_entry(exp->exp_outstanding_replies.next,
 489                                    struct ptlrpc_reply_state, rs_exp_list);
 490                 struct ptlrpc_service *svc = rs->rs_srv_ni->sni_service;
 491
 492                 spin_lock(&svc->srv_lock);
 493                 list_del_init(&rs->rs_exp_list);
 494                 ptlrpc_schedule_difficult_reply(rs);
 495                 spin_unlock(&svc->srv_lock);
 496         }
 497         spin_unlock_irqrestore(&exp->exp_lock, irqflags);
 498
 499         class_export_put(exp);
 500         RETURN(rc);
 501 }
 502
 503 static int mds_getstatus(struct ptlrpc_request *req)
 504 {
 505         struct mds_obd *mds = mds_req2mds(req);
 506         struct mds_body *body;
 507         int rc, size = sizeof(*body);
 508         ENTRY;
 509
 510         rc = lustre_pack_reply(req, 1, &size, NULL);
 511         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_GETSTATUS_PACK)) {
 512                 CERROR("mds: out of memory for message: size=%d\n", size);
 513                 req->rq_status = -ENOMEM;       /* superfluous? */
 514                 RETURN(-ENOMEM);
 515         }
 516
 517         body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
 518         memcpy(&body->fid1, &mds->mds_rootfid, sizeof(body->fid1));
 519
 520         /* the last_committed and last_xid fields are filled in for all
 521          * replies already - no need to do so here also.
 522          */
 523         RETURN(0);
 524 }
 525
 526 int mds_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
 527                      void *data, int flag)
 528 {
 529         int do_ast;
 530         ENTRY;
 531
 532         if (flag == LDLM_CB_CANCELING) {
 533                 /* Don't need to do anything here. */
 534                 RETURN(0);
 535         }
 536
 537         /* XXX layering violation!  -phil */
 538         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 539         /* Get this: if mds_blocking_ast is racing with mds_intent_policy,
 540          * such that mds_blocking_ast is called just before l_i_p takes the
 541          * ns_lock, then by the time we get the lock, we might not be the
 542          * correct blocking function anymore.  So check, and return early, if
 543          * so. */
 544         if (lock->l_blocking_ast != mds_blocking_ast) {
 545                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 546                 RETURN(0);
 547         }
 548
 549         lock->l_flags |= LDLM_FL_CBPENDING;
 550         do_ast = (!lock->l_readers && !lock->l_writers);
 551         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 552
 553         if (do_ast) {
 554                 struct lustre_handle lockh;
 555                 int rc;
 556
 557                 LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel");
 558                 ldlm_lock2handle(lock, &lockh);
 559                 rc = ldlm_cli_cancel(&lockh);
 560                 if (rc < 0)
 561                         CERROR("ldlm_cli_cancel: %d\n", rc);
 562         } else {
 563                 LDLM_DEBUG(lock, "Lock still has references, will be "
 564                            "cancelled later");
 565         }
 566         RETURN(0);
 567 }
 568
 569 int mds_get_md(struct obd_device *obd, struct inode *inode, void *md,
 570                int *size, int lock)
 571 {
 572         int rc = 0;
 573         int lmm_size;
 574
 575         if (lock)
 576                 down(&inode->i_sem);
 577         rc = fsfilt_get_md(obd, inode, md, *size);
 578         if (lock)
 579                 up(&inode->i_sem);
 580
 581         if (rc < 0) {
 582                 CERROR("Error %d reading eadata for ino %lu\n",
 583                        rc, inode->i_ino);
 584         } else if (rc > 0) {
 585                 lmm_size = rc;
 586
 587                 if (S_ISREG(inode->i_mode))
 588                         rc = mds_convert_lov_ea(obd, inode, md, lmm_size);
 589
 590                 if (rc == 0) {
 591                         *size = lmm_size;
 592                         rc = lmm_size;
 593                 } else if (rc > 0) {
 594                         *size = rc;
 595                 }
 596         }
 597
 598         RETURN (rc);
 599 }
 600
 601
 602 /* Call with lock=1 if you want mds_pack_md to take the i_sem.
 603  * Call with lock=0 if the caller has already taken the i_sem. */
 604 int mds_pack_md(struct obd_device *obd, struct lustre_msg *msg, int offset,
 605                 struct mds_body *body, struct inode *inode, int lock)
 606 {
 607         struct mds_obd *mds = &obd->u.mds;
 608         void *lmm;
 609         int lmm_size;
 610         int rc;
 611         ENTRY;
 612
 613         lmm = lustre_msg_buf(msg, offset, 0);
 614         if (lmm == NULL) {
 615                 /* Some problem with getting eadata when I sized the reply
 616                  * buffer... */
 617                 CDEBUG(D_INFO, "no space reserved for inode %lu MD\n",
 618                        inode->i_ino);
 619                 RETURN(0);
 620         }
 621         lmm_size = msg->buflens[offset];
 622
 623         /* I don't really like this, but it is a sanity check on the client
 624          * MD request.  However, if the client doesn't know how much space
 625          * to reserve for the MD, it shouldn't be bad to have too much space.
 626          */
 627         if (lmm_size > mds->mds_max_mdsize) {
 628                 CWARN("Reading MD for inode %lu of %d bytes > max %d\n",
 629                        inode->i_ino, lmm_size, mds->mds_max_mdsize);
 630                 // RETURN(-EINVAL);
 631         }
 632
 633         rc = mds_get_md(obd, inode, lmm, &lmm_size, lock);
 634         if (rc > 0) {
 635                 if (S_ISDIR(inode->i_mode))
 636                         body->valid |= OBD_MD_FLDIREA;
 637                 else
 638                         body->valid |= OBD_MD_FLEASIZE;
 639                 body->eadatasize = lmm_size;
 640                 rc = 0;
 641         }
 642
 643         RETURN(rc);
 644 }
 645
 646 static int mds_getattr_internal(struct obd_device *obd, struct dentry *dentry,
 647                                 struct ptlrpc_request *req,
 648                                 struct mds_body *reqbody, int reply_off)
 649 {
 650         struct mds_body *body;
 651         struct inode *inode = dentry->d_inode;
 652         int rc = 0;
 653         ENTRY;
 654
 655         if (inode == NULL && !(dentry->d_flags & DCACHE_CROSS_REF))
 656                 RETURN(-ENOENT);
 657
 658         body = lustre_msg_buf(req->rq_repmsg, reply_off, sizeof(*body));
 659         LASSERT(body != NULL);                 /* caller prepped reply */
 660
 661         if (dentry->d_flags & DCACHE_CROSS_REF) {
 662                 CDEBUG(D_OTHER, "cross reference: %lu/%lu/%lu\n",
 663                        (unsigned long) dentry->d_mdsnum,
 664                        (unsigned long) dentry->d_inum,
 665                        (unsigned long) dentry->d_generation);
 666                 body->valid |= OBD_MD_FLID | OBD_MD_MDS;
 667                 body->fid1.id = dentry->d_inum;
 668                 body->fid1.mds = dentry->d_mdsnum;
 669                 body->fid1.generation = dentry->d_generation;
 670                 RETURN(0);
 671         }
 672         mds_pack_inode2fid(obd, &body->fid1, inode);
 673         mds_pack_inode2body(obd, body, inode);
 674
 675         if ((S_ISREG(inode->i_mode) && (reqbody->valid & OBD_MD_FLEASIZE)) ||
 676             (S_ISDIR(inode->i_mode) && (reqbody->valid & OBD_MD_FLDIREA))) {
 677                 rc = mds_pack_md(obd, req->rq_repmsg, reply_off + 1, body,
 678                                  inode, 1);
 679
 680                 /* If we have LOV EA data, the OST holds size, atime, mtime */
 681                 if (!(body->valid & OBD_MD_FLEASIZE) &&
 682                     !(body->valid & OBD_MD_FLDIREA))
 683                         body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
 684                                         OBD_MD_FLATIME | OBD_MD_FLMTIME);
 685         } else if (S_ISLNK(inode->i_mode) &&
 686                    (reqbody->valid & OBD_MD_LINKNAME) != 0) {
 687                 char *symname = lustre_msg_buf(req->rq_repmsg, reply_off + 1,0);
 688                 int len;
 689
 690                 LASSERT (symname != NULL);       /* caller prepped reply */
 691                 len = req->rq_repmsg->buflens[reply_off + 1];
 692
 693                 rc = inode->i_op->readlink(dentry, symname, len);
 694                 if (rc < 0) {
 695                         CERROR("readlink failed: %d\n", rc);
 696                 } else if (rc != len - 1) {
 697                         CERROR ("Unexpected readlink rc %d: expecting %d\n",
 698                                 rc, len - 1);
 699                         rc = -EINVAL;
 700                 } else {
 701                         CDEBUG(D_INODE, "read symlink dest %s\n", symname);
 702                         body->valid |= OBD_MD_LINKNAME;
 703                         body->eadatasize = rc + 1;
 704                         symname[rc] = 0;        /* NULL terminate */
 705                         rc = 0;
 706                 }
 707         }
 708
 709         RETURN(rc);
 710 }
 711
 712 static int mds_getattr_pack_msg_cf(struct ptlrpc_request *req,
 713                                         struct dentry *dentry,
 714                                         int offset)
 715 {
 716         int rc = 0, size[1] = {sizeof(struct mds_body)};
 717         ENTRY;
 718
 719         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK)) {
 720                 CERROR("failed MDS_GETATTR_PACK test\n");
 721                 req->rq_status = -ENOMEM;
 722                 GOTO(out, rc = -ENOMEM);
 723         }
 724
 725         rc = lustre_pack_reply(req, 1, size, NULL);
 726         if (rc) {
 727                 CERROR("out of memory\n");
 728                 GOTO(out, req->rq_status = rc);
 729         }
 730
 731         EXIT;
 732  out:
 733         return(rc);
 734 }
 735
 736 static int mds_getattr_pack_msg(struct ptlrpc_request *req, struct inode *inode,
 737                                 int offset)
 738 {
 739         struct mds_obd *mds = mds_req2mds(req);
 740         struct mds_body *body;
 741         int rc = 0, size[2] = {sizeof(*body)}, bufcount = 1;
 742         ENTRY;
 743
 744         body = lustre_msg_buf(req->rq_reqmsg, offset, sizeof (*body));
 745         LASSERT(body != NULL);                 /* checked by caller */
 746         LASSERT_REQSWABBED(req, offset);       /* swabbed by caller */
 747
 748         if ((S_ISREG(inode->i_mode) && (body->valid & OBD_MD_FLEASIZE)) ||
 749             (S_ISDIR(inode->i_mode) && (body->valid & OBD_MD_FLDIREA))) {
 750                 int rc;
 751                 down(&inode->i_sem);
 752                 rc = fsfilt_get_md(req->rq_export->exp_obd, inode, NULL, 0);
 753                 up(&inode->i_sem);
 754                 CDEBUG(D_INODE, "got %d bytes MD data for inode %lu\n",
 755                        rc, inode->i_ino);
 756                 if (rc < 0) {
 757                         if (rc != -ENODATA)
 758                                 CERROR("error getting inode %lu MD: rc = %d\n",
 759                                        inode->i_ino, rc);
 760                         size[bufcount] = 0;
 761                 } else if (rc > mds->mds_max_mdsize) {
 762                         size[bufcount] = 0;
 763                         CERROR("MD size %d larger than maximum possible %u\n",
 764                                rc, mds->mds_max_mdsize);
 765                 } else {
 766                         size[bufcount] = rc;
 767                 }
 768                 bufcount++;
 769         } else if (S_ISLNK(inode->i_mode) && (body->valid & OBD_MD_LINKNAME)) {
 770                 if (inode->i_size + 1 != body->eadatasize)
 771                         CERROR("symlink size: %Lu, reply space: %d\n",
 772                                inode->i_size + 1, body->eadatasize);
 773                 size[bufcount] = min_t(int, inode->i_size+1, body->eadatasize);
 774                 bufcount++;
 775                 CDEBUG(D_INODE, "symlink size: %Lu, reply space: %d\n",
 776                        inode->i_size + 1, body->eadatasize);
 777         }
 778
 779         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK)) {
 780                 CERROR("failed MDS_GETATTR_PACK test\n");
 781                 req->rq_status = -ENOMEM;
 782                 GOTO(out, rc = -ENOMEM);
 783         }
 784
 785         rc = lustre_pack_reply(req, bufcount, size, NULL);
 786         if (rc) {
 787                 CERROR("out of memory\n");
 788                 GOTO(out, req->rq_status = rc);
 789         }
 790
 791         EXIT;
 792  out:
 793         return(rc);
 794 }
 795
 796 static int mds_getattr_name(int offset, struct ptlrpc_request *req,
 797                             struct lustre_handle *child_lockh, int child_part)
 798 {
 799         struct obd_device *obd = req->rq_export->exp_obd;
 800         struct ldlm_reply *rep = NULL;
 801         struct lvfs_run_ctxt saved;
 802         struct mds_body *body;
 803         struct dentry *dparent = NULL, *dchild = NULL;
 804         struct lvfs_ucred uc;
 805         struct lustre_handle parent_lockh[2];
 806         int namesize, update_mode;
 807         int rc = 0, cleanup_phase = 0, resent_req = 0;
 808         char *name;
 809         ENTRY;
 810
 811         LASSERT(!strcmp(obd->obd_type->typ_name, "mds"));
 812
 813         /* Swab now, before anyone looks inside the request */
 814
 815         body = lustre_swab_reqbuf(req, offset, sizeof(*body),
 816                                   lustre_swab_mds_body);
 817         if (body == NULL) {
 818                 CERROR("Can't swab mds_body\n");
 819                 GOTO(cleanup, rc = -EFAULT);
 820         }
 821
 822         LASSERT_REQSWAB(req, offset + 1);
 823         name = lustre_msg_string(req->rq_reqmsg, offset + 1, 0);
 824         if (name == NULL) {
 825                 CERROR("Can't unpack name\n");
 826                 GOTO(cleanup, rc = -EFAULT);
 827         }
 828         namesize = req->rq_reqmsg->buflens[offset + 1];
 829
 830         LASSERT (offset == 0 || offset == 2);
 831         /* if requests were at offset 2, the getattr reply goes back at 1 */
 832         if (offset) {
 833                 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
 834                 offset = 1;
 835         }
 836
 837         uc.luc_fsuid = body->fsuid;
 838         uc.luc_fsgid = body->fsgid;
 839         uc.luc_cap = body->capability;
 840         uc.luc_suppgid1 = body->suppgid;
 841         uc.luc_suppgid2 = -1;
 842         push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
 843         cleanup_phase = 1; /* kernel context */
 844         intent_set_disposition(rep, DISP_LOOKUP_EXECD);
 845
 846         LASSERT(namesize > 0);
 847         if (namesize == 1) {
 848                 /* we have no dentry here, drop LOOKUP bit */
 849                 child_part &= ~MDS_INODELOCK_LOOKUP;
 850                 CDEBUG(D_OTHER, "%s: request to retrieve attrs for %lu/%lu\n",
 851                        obd->obd_name, (unsigned long) body->fid1.id,
 852                        (unsigned long) body->fid1.generation);
 853                 dchild = mds_fid2locked_dentry(obd, &body->fid1, NULL, LCK_PR,
 854                                                parent_lockh, &update_mode,
 855                                                NULL, 0, child_part);
 856                 if (IS_ERR(dchild)) {
 857                         CERROR("can't find inode: %d\n", (int) PTR_ERR(dchild));
 858                         GOTO(cleanup, rc = PTR_ERR(dchild));
 859                 }
 860                 memcpy(child_lockh, parent_lockh, sizeof(parent_lockh[0]));
 861 #ifdef S_PDIROPS
 862                 if (parent_lockh[1].cookie)
 863                         ldlm_lock_decref(parent_lockh + 1, update_mode);
 864 #endif
 865                 cleanup_phase = 2;
 866                 goto fill_inode;
 867         }
 868
 869         /* FIXME: handle raw lookup */
 870 #if 0
 871         if (body->valid == OBD_MD_FLID) {
 872                 struct mds_body *mds_reply;
 873                 int size = sizeof(*mds_reply);
 874                 ino_t inum;
 875                 // The user requested ONLY the inode number, so do a raw lookup
 876                 rc = lustre_pack_reply(req, 1, &size, NULL);
 877                 if (rc) {
 878                         CERROR("out of memory\n");
 879                         GOTO(cleanup, rc);
 880                 }
 881
 882                 rc = dir->i_op->lookup_raw(dir, name, namesize - 1, &inum);
 883
 884                 mds_reply = lustre_msg_buf(req->rq_repmsg, offset,
 885                                            sizeof(*mds_reply));
 886                 mds_reply->fid1.id = inum;
 887                 mds_reply->valid = OBD_MD_FLID;
 888                 GOTO(cleanup, rc);
 889         }
 890 #endif
 891
 892         if (child_lockh->cookie != 0) {
 893                 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT);
 894                 resent_req = 1;
 895         }
 896
 897         if (resent_req == 0) {
 898                 rc = mds_get_parent_child_locked(obd, &obd->u.mds, &body->fid1,
 899                                                  parent_lockh, &dparent,
 900                                                  LCK_PR, MDS_INODELOCK_LOOKUP,
 901                                                  &update_mode, name, namesize,
 902                                                  child_lockh, &dchild, LCK_PR,
 903                                                  child_part);
 904                 if (rc)
 905                         GOTO(cleanup, rc);
 906         } else {
 907                 struct ldlm_lock *granted_lock;
 908                 struct ll_fid child_fid;
 909                 struct ldlm_resource *res;
 910                 DEBUG_REQ(D_DLMTRACE, req, "resent, not enqueuing new locks");
 911                 granted_lock = ldlm_handle2lock(child_lockh);
 912                 LASSERT(granted_lock);
 913
 914                 res = granted_lock->l_resource;
 915                 child_fid.id = res->lr_name.name[0];
 916                 child_fid.generation = res->lr_name.name[1];
 917                 dchild = mds_fid2dentry(&obd->u.mds, &child_fid, NULL);
 918                 LASSERT(dchild);
 919                 LDLM_LOCK_PUT(granted_lock);
 920         }
 921
 922         cleanup_phase = 2; /* dchild, dparent, locks */
 923
 924 fill_inode:
 925         if (!DENTRY_VALID(dchild)) {
 926                 intent_set_disposition(rep, DISP_LOOKUP_NEG);
 927                 /* in the intent case, the policy clears this error:
 928                    the disposition is enough */
 929                 GOTO(cleanup, rc = -ENOENT);
 930         } else {
 931                 intent_set_disposition(rep, DISP_LOOKUP_POS);
 932         }
 933
 934         if (req->rq_repmsg == NULL) {
 935                 if (dchild->d_flags & DCACHE_CROSS_REF)
 936                         rc = mds_getattr_pack_msg_cf(req, dchild, offset);
 937                 else
 938                         rc = mds_getattr_pack_msg(req, dchild->d_inode, offset);
 939                 if (rc != 0) {
 940                         CERROR ("mds_getattr_pack_msg: %d\n", rc);
 941                         GOTO (cleanup, rc);
 942                 }
 943         }
 944
 945         rc = mds_getattr_internal(obd, dchild, req, body, offset);
 946         GOTO(cleanup, rc); /* returns the lock to the client */
 947
 948  cleanup:
 949         switch (cleanup_phase) {
 950         case 2:
 951                 if (resent_req == 0) {
 952                         if (rc && DENTRY_VALID(dchild))
 953                                 ldlm_lock_decref(child_lockh, LCK_PR);
 954                         if (dparent) {
 955                                 ldlm_lock_decref(parent_lockh, LCK_PR);
 956 #ifdef S_PDIROPS
 957                                 if (parent_lockh[1].cookie != 0)
 958                                         ldlm_lock_decref(parent_lockh + 1,
 959                                                          update_mode);
 960 #endif
 961                         }
 962                         if (dparent)
 963                                 l_dput(dparent);
 964                 }
 965                 l_dput(dchild);
 966         case 1:
 967                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
 968         default: ;
 969         }
 970         return rc;
 971 }
 972
 973 static int mds_getattr(int offset, struct ptlrpc_request *req)
 974 {
 975         struct mds_obd *mds = mds_req2mds(req);
 976         struct obd_device *obd = req->rq_export->exp_obd;
 977         struct lvfs_run_ctxt saved;
 978         struct dentry *de;
 979         struct mds_body *body;
 980         struct lvfs_ucred uc;
 981         int rc = 0;
 982         ENTRY;
 983
 984         body = lustre_swab_reqbuf (req, offset, sizeof (*body),
 985                                    lustre_swab_mds_body);
 986         if (body == NULL) {
 987                 CERROR ("Can't unpack body\n");
 988                 RETURN (-EFAULT);
 989         }
 990
 991         uc.luc_fsuid = body->fsuid;
 992         uc.luc_fsgid = body->fsgid;
 993         uc.luc_cap = body->capability;
 994         push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
 995         de = mds_fid2dentry(mds, &body->fid1, NULL);
 996         if (IS_ERR(de)) {
 997                 rc = req->rq_status = -ENOENT;
 998                 GOTO(out_pop, PTR_ERR(de));
 999         }
1000
1001         rc = mds_getattr_pack_msg(req, de->d_inode, offset);
1002         if (rc != 0) {
1003                 CERROR ("mds_getattr_pack_msg: %d\n", rc);
1004                 GOTO (out_pop, rc);
1005         }
1006
1007         req->rq_status = mds_getattr_internal(obd, de, req, body, 0);
1008
1009         l_dput(de);
1010         GOTO(out_pop, rc);
1011 out_pop:
1012         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1013         return rc;
1014 }
1015
1016
1017 static int mds_obd_statfs(struct obd_device *obd, struct obd_statfs *osfs,
1018                           unsigned long max_age)
1019 {
1020         int rc;
1021
1022         spin_lock(&obd->obd_osfs_lock);
1023         rc = fsfilt_statfs(obd, obd->u.mds.mds_sb, max_age);
1024         if (rc == 0)
1025                 memcpy(osfs, &obd->obd_osfs, sizeof(*osfs));
1026         spin_unlock(&obd->obd_osfs_lock);
1027
1028         return rc;
1029 }
1030
1031 static int mds_statfs(struct ptlrpc_request *req)
1032 {
1033         struct obd_device *obd = req->rq_export->exp_obd;
1034         int rc, size = sizeof(struct obd_statfs);
1035         ENTRY;
1036
1037         rc = lustre_pack_reply(req, 1, &size, NULL);
1038         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_STATFS_PACK)) {
1039                 CERROR("mds: statfs lustre_pack_reply failed: rc = %d\n", rc);
1040                 GOTO(out, rc);
1041         }
1042
1043         /* We call this so that we can cache a bit - 1 jiffie worth */
1044         rc = mds_obd_statfs(obd, lustre_msg_buf(req->rq_repmsg, 0, size),
1045                             jiffies - HZ);
1046         if (rc) {
1047                 CERROR("mds_obd_statfs failed: rc %d\n", rc);
1048                 GOTO(out, rc);
1049         }
1050
1051         EXIT;
1052 out:
1053         req->rq_status = rc;
1054         return 0;
1055 }
1056
1057 static int mds_sync(struct ptlrpc_request *req)
1058 {
1059         struct obd_device *obd = req->rq_export->exp_obd;
1060         struct mds_obd *mds = &obd->u.mds;
1061         struct mds_body *body;
1062         int rc, size = sizeof(*body);
1063         ENTRY;
1064
1065         body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof(*body));
1066         if (body == NULL)
1067                 GOTO(out, rc = -EPROTO);
1068
1069         rc = lustre_pack_reply(req, 1, &size, NULL);
1070         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_SYNC_PACK)) {
1071                 CERROR("fsync lustre_pack_reply failed: rc = %d\n", rc);
1072                 GOTO(out, rc);
1073         }
1074
1075         if (body->fid1.id == 0) {
1076                 /* a fid of zero is taken to mean "sync whole filesystem" */
1077                 rc = fsfilt_sync(obd, mds->mds_sb);
1078                 if (rc)
1079                         GOTO(out, rc);
1080         } else {
1081                 /* just any file to grab fsync method - "file" arg unused */
1082                 struct file *file = mds->mds_rcvd_filp;
1083                 struct dentry *de;
1084
1085                 de = mds_fid2dentry(mds, &body->fid1, NULL);
1086                 if (IS_ERR(de))
1087                         GOTO(out, rc = PTR_ERR(de));
1088
1089                 rc = file->f_op->fsync(NULL, de, 1);
1090                 l_dput(de);
1091                 if (rc)
1092                         GOTO(out, rc);
1093
1094                 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*body));
1095                 mds_pack_inode2fid(obd, &body->fid1, de->d_inode);
1096                 mds_pack_inode2body(obd, body, de->d_inode);
1097         }
1098 out:
1099         req->rq_status = rc;
1100         return 0;
1101 }
1102
1103 /* mds_readpage does not take a DLM lock on the inode, because the client must
1104  * already have a PR lock.
1105  *
1106  * If we were to take another one here, a deadlock will result, if another
1107  * thread is already waiting for a PW lock. */
1108 static int mds_readpage(struct ptlrpc_request *req)
1109 {
1110         struct obd_device *obd = req->rq_export->exp_obd;
1111         struct vfsmount *mnt;
1112         struct dentry *de;
1113         struct file *file;
1114         struct mds_body *body, *repbody;
1115         struct lvfs_run_ctxt saved;
1116         int rc, size = sizeof(*repbody);
1117         struct lvfs_ucred uc;
1118         ENTRY;
1119
1120         rc = lustre_pack_reply(req, 1, &size, NULL);
1121         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_READPAGE_PACK)) {
1122                 CERROR("mds: out of memory\n");
1123                 GOTO(out, rc = -ENOMEM);
1124         }
1125
1126         body = lustre_swab_reqbuf(req, 0, sizeof(*body), lustre_swab_mds_body);
1127         if (body == NULL)
1128                 GOTO (out, rc = -EFAULT);
1129
1130         uc.luc_fsuid = body->fsuid;
1131         uc.luc_fsgid = body->fsgid;
1132         uc.luc_cap = body->capability;
1133         push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1134         de = mds_fid2dentry(&obd->u.mds, &body->fid1, &mnt);
1135         if (IS_ERR(de))
1136                 GOTO(out_pop, rc = PTR_ERR(de));
1137
1138         CDEBUG(D_INODE, "ino %lu\n", de->d_inode->i_ino);
1139
1140         file = dentry_open(de, mnt, O_RDONLY | O_LARGEFILE);
1141         /* note: in case of an error, dentry_open puts dentry */
1142         if (IS_ERR(file))
1143                 GOTO(out_pop, rc = PTR_ERR(file));
1144
1145         /* body->size is actually the offset -eeb */
1146         if ((body->size & (de->d_inode->i_blksize - 1)) != 0) {
1147                 CERROR("offset "LPU64" not on a block boundary of %lu\n",
1148                        body->size, de->d_inode->i_blksize);
1149                 GOTO(out_file, rc = -EFAULT);
1150         }
1151
1152         /* body->nlink is actually the #bytes to read -eeb */
1153         if (body->nlink & (de->d_inode->i_blksize - 1)) {
1154                 CERROR("size %u is not multiple of blocksize %lu\n",
1155                        body->nlink, de->d_inode->i_blksize);
1156                 GOTO(out_file, rc = -EFAULT);
1157         }
1158
1159         repbody = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*repbody));
1160         repbody->size = file->f_dentry->d_inode->i_size;
1161         repbody->valid = OBD_MD_FLSIZE;
1162
1163         /* to make this asynchronous make sure that the handling function
1164            doesn't send a reply when this function completes. Instead a
1165            callback function would send the reply */
1166         /* body->size is actually the offset -eeb */
1167         rc = mds_sendpage(req, file, body->size, body->nlink);
1168
1169 out_file:
1170         filp_close(file, 0);
1171 out_pop:
1172         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1173 out:
1174         req->rq_status = rc;
1175         RETURN(0);
1176 }
1177
1178 int mds_reint(struct ptlrpc_request *req, int offset,
1179               struct lustre_handle *lockh)
1180 {
1181         struct mds_update_record *rec; /* 116 bytes on the stack?  no sir! */
1182         int rc;
1183         ENTRY;
1184
1185         OBD_ALLOC(rec, sizeof(*rec));
1186         if (rec == NULL)
1187                 RETURN(-ENOMEM);
1188
1189         rc = mds_update_unpack(req, offset, rec);
1190         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNPACK)) {
1191                 CERROR("invalid record\n");
1192                 GOTO(out, req->rq_status = -EINVAL);
1193         }
1194         /* rc will be used to interrupt a for loop over multiple records */
1195         rc = mds_reint_rec(rec, offset, req, lockh);
1196  out:
1197         OBD_FREE(rec, sizeof(*rec));
1198         RETURN(rc);
1199 }
1200
1201 static int mds_filter_recovery_request(struct ptlrpc_request *req,
1202                                        struct obd_device *obd, int *process)
1203 {
1204         switch (req->rq_reqmsg->opc) {
1205         case MDS_CONNECT: /* This will never get here, but for completeness. */
1206         case OST_CONNECT: /* This will never get here, but for completeness. */
1207         case MDS_DISCONNECT:
1208         case OST_DISCONNECT:
1209                *process = 1;
1210                RETURN(0);
1211
1212         case MDS_CLOSE:
1213         case MDS_SYNC: /* used in unmounting */
1214         case OBD_PING:
1215         case MDS_REINT:
1216         case LDLM_ENQUEUE:
1217         case OST_CREATE:
1218                 *process = target_queue_recovery_request(req, obd);
1219                 RETURN(0);
1220
1221         default:
1222                 DEBUG_REQ(D_ERROR, req, "not permitted during recovery");
1223                 *process = 0;
1224                 /* XXX what should we set rq_status to here? */
1225                 req->rq_status = -EAGAIN;
1226                 RETURN(ptlrpc_error(req));
1227         }
1228 }
1229
1230 static char *reint_names[] = {
1231         [REINT_SETATTR] "setattr",
1232         [REINT_CREATE]  "create",
1233         [REINT_LINK]    "link",
1234         [REINT_UNLINK]  "unlink",
1235         [REINT_RENAME]  "rename",
1236         [REINT_OPEN]    "open",
1237 };
1238
1239 #define FILTER_VALID_FLAGS (OBD_MD_FLTYPE | OBD_MD_FLMODE | OBD_MD_FLGENER  |\
1240                             OBD_MD_FLSIZE | OBD_MD_FLBLOCKS | OBD_MD_FLBLKSZ|\
1241                             OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME|\
1242                             OBD_MD_FLID)
1243
1244 static void reconstruct_create(struct ptlrpc_request *req)
1245 {
1246         struct mds_export_data *med = &req->rq_export->exp_mds_data;
1247         struct mds_client_data *mcd = med->med_mcd;
1248         struct ost_body *body;
1249
1250         body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*body));
1251
1252         /* copy rc, transno and disp; steal locks */
1253         mds_req_from_mcd(req, mcd);
1254         CERROR("reconstruct reply for x"LPU64"\n", req->rq_xid);
1255 }
1256
1257 static int mdt_obj_create(struct ptlrpc_request *req)
1258 {
1259         struct obd_device *obd = req->rq_export->exp_obd;
1260         struct ldlm_res_id res_id = { .name = {0} };
1261         struct mds_obd *mds = &obd->u.mds;
1262         struct ost_body *body, *repbody;
1263         int rc, size = sizeof(*repbody);
1264         char fidname[LL_FID_NAMELEN];
1265         struct inode *parent_inode;
1266         struct lustre_handle lockh;
1267         struct lvfs_run_ctxt saved;
1268         ldlm_policy_data_t policy;
1269         struct dentry *new = NULL;
1270         struct dentry_params dp;
1271         int mealen, flags = 0;
1272         unsigned int tmpname;
1273         struct lvfs_ucred uc;
1274         struct mea *mea;
1275         void *handle;
1276         ENTRY;
1277
1278         DEBUG_REQ(D_HA, req, "create remote object");
1279
1280         parent_inode = mds->mds_objects_dir->d_inode;
1281
1282         body = lustre_swab_reqbuf(req, 0, sizeof(*body),
1283                                   lustre_swab_ost_body);
1284         if (body == NULL)
1285                 RETURN(-EFAULT);
1286
1287         MDS_CHECK_RESENT(req, reconstruct_create(req));
1288
1289         uc.luc_fsuid = body->oa.o_uid;
1290         uc.luc_fsgid = body->oa.o_gid;
1291
1292         push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1293
1294         rc = lustre_pack_reply(req, 1, &size, NULL);
1295         if (rc)
1296                 RETURN(rc);
1297
1298         repbody = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*repbody));
1299
1300         if (body->oa.o_flags & OBD_FL_RECREATE_OBJS) {
1301                 /* this is re-create request from MDS holding directory name.
1302                  * we have to lookup given ino/generation first. if it exists
1303                  * (good case) then there is nothing to do. if it does not
1304                  * then we have to recreate it */
1305                 struct ll_fid fid;
1306                 fid.id = body->oa.o_id;
1307                 fid.generation = body->oa.o_generation;
1308                 new = mds_fid2dentry(mds, &fid, NULL);
1309                 if (!IS_ERR(new) && new->d_inode) {
1310                         CWARN("mkdir() repairing is on its way: %lu/%lu\n",
1311                               (unsigned long) fid.id,
1312                               (unsigned long) fid.generation);
1313                         obdo_from_inode(&repbody->oa, new->d_inode,
1314                                         FILTER_VALID_FLAGS);
1315                         repbody->oa.o_id = new->d_inode->i_ino;
1316                         repbody->oa.o_generation = new->d_inode->i_generation;
1317                         repbody->oa.o_valid |= OBD_MD_FLID | OBD_MD_FLGENER;
1318                         GOTO(cleanup2, rc = 0);
1319                 }
1320                 CWARN("hmm. for some reason dir %lu/%lu (or reply) got lost\n",
1321                       (unsigned long) fid.id, (unsigned long) fid.generation);
1322                 LASSERT(new->d_inode == NULL ||
1323                         new->d_inode->i_generation != fid.generation);
1324                 l_dput(new);
1325         }
1326
1327         down(&parent_inode->i_sem);
1328         handle = fsfilt_start(obd, parent_inode, FSFILT_OP_MKDIR, NULL);
1329         LASSERT(!IS_ERR(handle));
1330
1331 repeat:
1332         tmpname = ll_insecure_random_int();
1333         rc = sprintf(fidname, "%u", tmpname);
1334         new = lookup_one_len(fidname, mds->mds_objects_dir, rc);
1335         if (IS_ERR(new)) {
1336                 CERROR("%s: can't lookup new inode (%s) for mkdir: %d\n",
1337                        obd->obd_name, fidname, (int) PTR_ERR(new));
1338                 fsfilt_commit(obd, new->d_inode, handle, 0);
1339                 up(&parent_inode->i_sem);
1340                 RETURN(PTR_ERR(new));
1341         } else if (new->d_inode) {
1342                 CERROR("%s: name exists. repeat\n", obd->obd_name);
1343                 goto repeat;
1344         }
1345
1346         new->d_fsdata = (void *) &dp;
1347         dp.p_inum = 0;
1348         dp.p_ptr = req;
1349
1350         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1351                 DEBUG_REQ(D_HA, req, "replay create obj %lu/%lu",
1352                           (unsigned long) body->oa.o_id,
1353                           (unsigned long) body->oa.o_generation);
1354                 dp.p_inum = body->oa.o_id;
1355                 dp.p_generation = body->oa.o_generation;
1356         }
1357         rc = vfs_mkdir(parent_inode, new, body->oa.o_mode);
1358         if (rc == 0) {
1359                 obdo_from_inode(&repbody->oa, new->d_inode, FILTER_VALID_FLAGS);
1360                 repbody->oa.o_id = new->d_inode->i_ino;
1361                 repbody->oa.o_generation = new->d_inode->i_generation;
1362                 repbody->oa.o_valid |= OBD_MD_FLID | OBD_MD_FLGENER;
1363
1364                 rc = fsfilt_del_dir_entry(obd, new);
1365                 up(&parent_inode->i_sem);
1366
1367                 if (rc) {
1368                         CERROR("can't remove name for object: %d\n", rc);
1369                         GOTO(cleanup, rc);
1370                 }
1371
1372                 /* this lock should be taken to serialize MDS modifications
1373                  * in failure case */
1374                 res_id.name[0] = new->d_inode->i_ino;
1375                 res_id.name[1] = new->d_inode->i_generation;
1376                 policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
1377                 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1378                                 res_id, LDLM_IBITS, &policy,
1379                                 LCK_EX, &flags, mds_blocking_ast,
1380                                 ldlm_completion_ast, NULL, NULL,
1381                                 NULL, 0, NULL, &lockh);
1382                 if (rc != ELDLM_OK)
1383                         GOTO(cleanup, rc);
1384
1385                 CDEBUG(D_OTHER, "created dirobj: %lu/%lu mode %o\n",
1386                                 (unsigned long) new->d_inode->i_ino,
1387                                 (unsigned long) new->d_inode->i_generation,
1388                                 (unsigned) new->d_inode->i_mode);
1389         } else {
1390                 up(&parent_inode->i_sem);
1391                 CERROR("%s: can't create dirobj: %d\n", obd->obd_name, rc);
1392         }
1393
1394         if (rc == 0 && body->oa.o_valid & OBD_MD_FLID) {
1395                 /* this is new object for splitted dir. we have to
1396                  * prevent recursive splitting on it -bzzz */
1397                 mealen = obd_size_diskmd(mds->mds_lmv_exp, NULL);
1398                 OBD_ALLOC(mea, mealen);
1399                 if (mea == NULL)
1400                         GOTO(cleanup, rc = -ENOMEM);
1401                 mea->mea_count = 0;
1402                 down(&new->d_inode->i_sem);
1403                 rc = fsfilt_set_md(obd, new->d_inode, handle, mea, mealen);
1404                 up(&new->d_inode->i_sem);
1405                 OBD_FREE(mea, mealen);
1406         }
1407
1408 cleanup:
1409         rc = mds_finish_transno(mds, parent_inode, handle, req, rc, 0);
1410         if (rc == 0)
1411                 ptlrpc_save_lock(req, &lockh, LCK_EX);
1412         else
1413                 ldlm_lock_decref(&lockh, LCK_EX);
1414 cleanup2:
1415         l_dput(new);
1416         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1417         RETURN(rc);
1418 }
1419
1420 static int mdt_get_info(struct ptlrpc_request *req)
1421 {
1422         char *key;
1423         struct obd_export *exp = req->rq_export;
1424         int keylen, rc = 0, size = sizeof(obd_id);
1425         obd_id *reply;
1426         ENTRY;
1427
1428         key = lustre_msg_buf(req->rq_reqmsg, 0, 1);
1429         if (key == NULL) {
1430                 DEBUG_REQ(D_HA, req, "no get_info key");
1431                 RETURN(-EFAULT);
1432         }
1433         keylen = req->rq_reqmsg->buflens[0];
1434
1435         if (keylen < strlen("mdsize") || memcmp(key, "mdsize", 6) != 0)
1436                 RETURN(-EPROTO);
1437
1438         rc = lustre_pack_reply(req, 1, &size, NULL);
1439         if (rc)
1440                 RETURN(rc);
1441
1442         reply = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*reply));
1443         rc = obd_get_info(exp, keylen, key, &size, reply);
1444         req->rq_repmsg->status = 0;
1445         RETURN(rc);
1446 }
1447
1448 static int mds_set_info(struct obd_export *exp, __u32 keylen,
1449                         void *key, __u32 vallen, void *val)
1450 {
1451         struct obd_device *obd;
1452         struct mds_obd *mds;
1453         int    rc = 0;
1454         ENTRY;
1455
1456         obd = class_exp2obd(exp);
1457         if (obd == NULL) {
1458                 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
1459                        exp->exp_handle.h_cookie);
1460                 RETURN(-EINVAL);
1461         }
1462
1463         mds = &obd->u.mds;
1464         if (keylen == strlen("mds_num") &&
1465             memcmp(key, "mds_num", keylen) == 0) {
1466                 int valsize;
1467                 __u32 group;
1468                 CDEBUG(D_IOCTL, "set mds num %d\n", *(int*)val);
1469                 mds->mds_num = *(int*)val;
1470                 group = FILTER_GROUP_FIRST_MDS + mds->mds_num;
1471                 valsize = sizeof(group);
1472                 /*mds number has been changed, so the corresponding obdfilter exp
1473                  *need to be changed too*/
1474                 rc = obd_set_info(mds->mds_osc_exp, strlen("mds_conn"), "mds_conn",
1475                           valsize, &group);
1476                 RETURN(rc);
1477         } else if (keylen == strlen("client") &&
1478                    memcmp(key, "client", keylen) == 0) {
1479                 if (!(exp->exp_flags & OBD_OPT_REAL_CLIENT)) {
1480                         atomic_inc(&mds->mds_real_clients);
1481                         CDEBUG(D_OTHER,"%s: peer from %s is real client (%d)\n",
1482                                obd->obd_name, exp->exp_client_uuid.uuid,
1483                                atomic_read(&mds->mds_real_clients));
1484                         exp->exp_flags |= OBD_OPT_REAL_CLIENT;
1485                 }
1486                 if (mds->mds_lmv_name) {
1487                         rc = mds_lmv_connect(obd, mds->mds_lmv_name);
1488                         LASSERT(rc == 0);
1489                 }
1490                 RETURN(0);
1491         }
1492         CDEBUG(D_IOCTL, "invalid key\n");
1493         RETURN(-EINVAL);
1494 }
1495
1496 static int mdt_set_info(struct ptlrpc_request *req)
1497 {
1498         char *key, *val;
1499         struct obd_export *exp = req->rq_export;
1500         int keylen, rc = 0, vallen;
1501         ENTRY;
1502
1503         key = lustre_msg_buf(req->rq_reqmsg, 0, 1);
1504         if (key == NULL) {
1505                 DEBUG_REQ(D_HA, req, "no set_info key");
1506                 RETURN(-EFAULT);
1507         }
1508         keylen = req->rq_reqmsg->buflens[0];
1509
1510         if (keylen == strlen("mds_num") &&
1511             memcmp(key, "mds_num", keylen) == 0) {
1512                 rc = lustre_pack_reply(req, 0, NULL, NULL);
1513                 if (rc)
1514                         RETURN(rc);
1515                 val = lustre_msg_buf(req->rq_reqmsg, 1, 0);
1516
1517                 vallen = req->rq_reqmsg->buflens[1];
1518
1519                 rc = obd_set_info(exp, keylen, key, vallen, val);
1520                 req->rq_repmsg->status = 0;
1521                 RETURN(rc);
1522         } else if (keylen == strlen("client") &&
1523                    memcmp(key, "client", keylen) == 0) {
1524                 rc = lustre_pack_reply(req, 0, NULL, NULL);
1525                 if (rc)
1526                         RETURN(rc);
1527                 rc = obd_set_info(exp, keylen, key, sizeof(obd_id), NULL);
1528                 req->rq_repmsg->status = 0;
1529                 RETURN(rc);
1530         }
1531         CDEBUG(D_IOCTL, "invalid key\n");
1532         RETURN(-EINVAL);
1533 }
1534
1535 extern int ost_brw_write(struct ptlrpc_request *, struct obd_trans_info *);
1536 int mds_handle(struct ptlrpc_request *req)
1537 {
1538         int should_process, fail = OBD_FAIL_MDS_ALL_REPLY_NET;
1539         int rc = 0;
1540         struct mds_obd *mds = NULL; /* quell gcc overwarning */
1541         struct obd_device *obd = NULL;
1542         ENTRY;
1543
1544         OBD_FAIL_RETURN(OBD_FAIL_MDS_ALL_REQUEST_NET | OBD_FAIL_ONCE, 0);
1545
1546         LASSERT(current->journal_info == NULL);
1547         /* XXX identical to OST */
1548         if (req->rq_reqmsg->opc != MDS_CONNECT) {
1549                 struct mds_export_data *med;
1550                 int recovering, abort_recovery;
1551
1552                 if (req->rq_export == NULL) {
1553                         CERROR("lustre_mds: operation %d on unconnected MDS\n",
1554                                req->rq_reqmsg->opc);
1555                         req->rq_status = -ENOTCONN;
1556                         GOTO(out, rc = -ENOTCONN);
1557                 }
1558
1559                 med = &req->rq_export->exp_mds_data;
1560                 obd = req->rq_export->exp_obd;
1561                 mds = &obd->u.mds;
1562
1563                 /* sanity check: if the xid matches, the request must
1564                  * be marked as a resent or replayed */
1565                 if (req->rq_xid == med->med_mcd->mcd_last_xid)
1566                         LASSERTF(lustre_msg_get_flags(req->rq_reqmsg) &
1567                                  (MSG_RESENT | MSG_REPLAY),
1568                                  "rq_xid "LPU64" matches last_xid, "
1569                                  "expected RESENT flag\n",
1570                                  req->rq_xid);
1571                 /* else: note the opposite is not always true; a
1572                  * RESENT req after a failover will usually not match
1573                  * the last_xid, since it was likely never
1574                  * committed. A REPLAYed request will almost never
1575                  * match the last xid, however it could for a
1576                  * committed, but still retained, open. */
1577
1578                 /* Check for aborted recovery. */
1579                 spin_lock_bh(&obd->obd_processing_task_lock);
1580                 abort_recovery = obd->obd_abort_recovery;
1581                 recovering = obd->obd_recovering;
1582                 spin_unlock_bh(&obd->obd_processing_task_lock);
1583                 if (abort_recovery) {
1584                         target_abort_recovery(obd);
1585                 } else if (recovering) {
1586                         rc = mds_filter_recovery_request(req, obd,
1587                                                          &should_process);
1588                         if (rc || !should_process)
1589                                 RETURN(rc);
1590                 }
1591         }
1592
1593         switch (req->rq_reqmsg->opc) {
1594         case MDS_CONNECT:
1595                 DEBUG_REQ(D_INODE, req, "connect");
1596                 OBD_FAIL_RETURN(OBD_FAIL_MDS_CONNECT_NET, 0);
1597                 rc = target_handle_connect(req, mds_handle);
1598                 if (!rc)
1599                         /* Now that we have an export, set mds. */
1600                         mds = mds_req2mds(req);
1601                 break;
1602
1603         case MDS_DISCONNECT:
1604                 DEBUG_REQ(D_INODE, req, "disconnect");
1605                 OBD_FAIL_RETURN(OBD_FAIL_MDS_DISCONNECT_NET, 0);
1606                 rc = target_handle_disconnect(req);
1607                 req->rq_status = rc;            /* superfluous? */
1608                 break;
1609
1610         case MDS_GETSTATUS:
1611                 DEBUG_REQ(D_INODE, req, "getstatus");
1612                 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETSTATUS_NET, 0);
1613                 rc = mds_getstatus(req);
1614                 break;
1615
1616         case MDS_GETATTR:
1617                 DEBUG_REQ(D_INODE, req, "getattr");
1618                 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_NET, 0);
1619                 rc = mds_getattr(0, req);
1620                 break;
1621
1622         case MDS_GETATTR_NAME: {
1623                 struct lustre_handle lockh;
1624                 DEBUG_REQ(D_INODE, req, "getattr_name");
1625                 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_NAME_NET, 0);
1626
1627                 /* If this request gets a reconstructed reply, we won't be
1628                  * acquiring any new locks in mds_getattr_name, so we don't
1629                  * want to cancel.
1630                  */
1631                 lockh.cookie = 0;
1632                 rc = mds_getattr_name(0, req, &lockh, MDS_INODELOCK_UPDATE);
1633                 /* this non-intent call (from an ioctl) is special */
1634                 req->rq_status = rc;
1635                 if (rc == 0 && lockh.cookie)
1636                         ldlm_lock_decref(&lockh, LCK_PR);
1637                 break;
1638         }
1639         case MDS_STATFS:
1640                 DEBUG_REQ(D_INODE, req, "statfs");
1641                 OBD_FAIL_RETURN(OBD_FAIL_MDS_STATFS_NET, 0);
1642                 rc = mds_statfs(req);
1643                 break;
1644
1645         case MDS_READPAGE:
1646                 DEBUG_REQ(D_INODE, req, "readpage");
1647                 OBD_FAIL_RETURN(OBD_FAIL_MDS_READPAGE_NET, 0);
1648                 rc = mds_readpage(req);
1649
1650                 if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_MDS_SENDPAGE)) {
1651                         if (req->rq_reply_state) {
1652                                 lustre_free_reply_state (req->rq_reply_state);
1653                                 req->rq_reply_state = NULL;
1654                         }
1655                         RETURN(0);
1656                 }
1657
1658                 break;
1659
1660         case MDS_REINT: {
1661                 __u32 *opcp = lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*opcp));
1662                 __u32  opc;
1663                 int size[3] = {sizeof(struct mds_body), mds->mds_max_mdsize,
1664                                mds->mds_max_cookiesize};
1665                 int bufcount;
1666
1667                 /* NB only peek inside req now; mds_reint() will swab it */
1668                 if (opcp == NULL) {
1669                         CERROR ("Can't inspect opcode\n");
1670                         rc = -EINVAL;
1671                         break;
1672                 }
1673                 opc = *opcp;
1674                 if (lustre_msg_swabbed (req->rq_reqmsg))
1675                         __swab32s(&opc);
1676
1677                 DEBUG_REQ(D_INODE, req, "reint %d (%s)", opc,
1678                           (opc < sizeof(reint_names) / sizeof(reint_names[0]) ||
1679                            reint_names[opc] == NULL) ? reint_names[opc] :
1680                                                        "unknown opcode");
1681
1682                 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET, 0);
1683
1684                 if (opc == REINT_UNLINK)
1685                         bufcount = 3;
1686                 else if (opc == REINT_OPEN || opc == REINT_RENAME)
1687                         bufcount = 2;
1688                 else
1689                         bufcount = 1;
1690
1691                 rc = lustre_pack_reply(req, bufcount, size, NULL);
1692                 if (rc)
1693                         break;
1694
1695                 rc = mds_reint(req, 0, NULL);
1696                 fail = OBD_FAIL_MDS_REINT_NET_REP;
1697                 break;
1698         }
1699
1700         case MDS_CLOSE:
1701                 DEBUG_REQ(D_INODE, req, "close");
1702                 OBD_FAIL_RETURN(OBD_FAIL_MDS_CLOSE_NET, 0);
1703                 rc = mds_close(req);
1704                 break;
1705
1706         case MDS_DONE_WRITING:
1707                 DEBUG_REQ(D_INODE, req, "done_writing");
1708                 OBD_FAIL_RETURN(OBD_FAIL_MDS_DONE_WRITING_NET, 0);
1709                 rc = mds_done_writing(req);
1710                 break;
1711
1712         case MDS_PIN:
1713                 DEBUG_REQ(D_INODE, req, "pin");
1714                 OBD_FAIL_RETURN(OBD_FAIL_MDS_PIN_NET, 0);
1715                 rc = mds_pin(req);
1716                 break;
1717
1718         case MDS_SYNC:
1719                 DEBUG_REQ(D_INODE, req, "sync");
1720                 OBD_FAIL_RETURN(OBD_FAIL_MDS_SYNC_NET, 0);
1721                 rc = mds_sync(req);
1722                 break;
1723
1724         case OBD_PING:
1725                 DEBUG_REQ(D_INODE, req, "ping");
1726                 rc = target_handle_ping(req);
1727                 break;
1728
1729         case OBD_LOG_CANCEL:
1730                 CDEBUG(D_INODE, "log cancel\n");
1731                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOG_CANCEL_NET, 0);
1732                 rc = -ENOTSUPP; /* la la la */
1733                 break;
1734
1735         case LDLM_ENQUEUE:
1736                 DEBUG_REQ(D_INODE, req, "enqueue");
1737                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_ENQUEUE, 0);
1738                 rc = ldlm_handle_enqueue(req, ldlm_server_completion_ast,
1739                                          ldlm_server_blocking_ast, NULL);
1740                 break;
1741         case LDLM_CONVERT:
1742                 DEBUG_REQ(D_INODE, req, "convert");
1743                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CONVERT, 0);
1744                 rc = ldlm_handle_convert(req);
1745                 break;
1746         case LDLM_BL_CALLBACK:
1747         case LDLM_CP_CALLBACK:
1748                 DEBUG_REQ(D_INODE, req, "callback");
1749                 CERROR("callbacks should not happen on MDS\n");
1750                 LBUG();
1751                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_BL_CALLBACK, 0);
1752                 break;
1753         case LLOG_ORIGIN_HANDLE_CREATE:
1754                 DEBUG_REQ(D_INODE, req, "llog_init");
1755                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1756                 rc = llog_origin_handle_create(req);
1757                 break;
1758         case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
1759                 DEBUG_REQ(D_INODE, req, "llog next block");
1760                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1761                 rc = llog_origin_handle_next_block(req);
1762                 break;
1763         case LLOG_ORIGIN_HANDLE_PREV_BLOCK:
1764                 DEBUG_REQ(D_INODE, req, "llog prev block");
1765                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1766                 rc = llog_origin_handle_prev_block(req);
1767                 break;
1768         case LLOG_ORIGIN_HANDLE_READ_HEADER:
1769                 DEBUG_REQ(D_INODE, req, "llog read header");
1770                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1771                 rc = llog_origin_handle_read_header(req);
1772                 break;
1773         case LLOG_ORIGIN_HANDLE_CLOSE:
1774                 DEBUG_REQ(D_INODE, req, "llog close");
1775                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1776                 rc = llog_origin_handle_close(req);
1777                 break;
1778         case OST_CREATE:
1779                 DEBUG_REQ(D_INODE, req, "ost_create");
1780                 rc = mdt_obj_create(req);
1781                 break;
1782         case OST_GET_INFO:
1783                 DEBUG_REQ(D_INODE, req, "get_info");
1784                 rc = mdt_get_info(req);
1785                 break;
1786         case OST_SET_INFO:
1787                 DEBUG_REQ(D_INODE, req, "set_info");
1788                 rc = mdt_set_info(req);
1789                 break;
1790         case OST_WRITE:
1791                 CDEBUG(D_INODE, "write\n");
1792                 OBD_FAIL_RETURN(OBD_FAIL_OST_BRW_NET, 0);
1793                 rc = ost_brw_write(req, NULL);
1794                 LASSERT(current->journal_info == NULL);
1795                 /* mdt_brw sends its own replies */
1796                 RETURN(rc);
1797                 break;
1798         case LLOG_CATINFO:
1799                 DEBUG_REQ(D_INODE, req, "llog catinfo");
1800                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1801                 rc = llog_catinfo(req);
1802                 break;
1803         default:
1804                 req->rq_status = -ENOTSUPP;
1805                 rc = ptlrpc_error(req);
1806                 RETURN(rc);
1807         }
1808
1809         LASSERT(current->journal_info == NULL);
1810
1811         EXIT;
1812
1813         /* If we're DISCONNECTing, the mds_export_data is already freed */
1814         if (!rc && req->rq_reqmsg->opc != MDS_DISCONNECT) {
1815                 struct mds_export_data *med = &req->rq_export->exp_mds_data;
1816                 struct obd_device *obd = list_entry(mds, struct obd_device,
1817                                                     u.mds);
1818                 req->rq_repmsg->last_xid =
1819                         le64_to_cpu(med->med_mcd->mcd_last_xid);
1820
1821                 if (!obd->obd_no_transno) {
1822                         req->rq_repmsg->last_committed =
1823                                 obd->obd_last_committed;
1824                 } else {
1825                         DEBUG_REQ(D_IOCTL, req,
1826                                   "not sending last_committed update");
1827                 }
1828                 CDEBUG(D_INFO, "last_transno "LPU64", last_committed "LPU64
1829                        ", xid "LPU64"\n",
1830                        mds->mds_last_transno, obd->obd_last_committed,
1831                        req->rq_xid);
1832         }
1833  out:
1834
1835         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LAST_REPLAY) {
1836                 if (obd && obd->obd_recovering) {
1837                         DEBUG_REQ(D_HA, req, "LAST_REPLAY, queuing reply");
1838                         return target_queue_final_reply(req, rc);
1839                 }
1840                 /* Lost a race with recovery; let the error path DTRT. */
1841                 rc = req->rq_status = -ENOTCONN;
1842         }
1843
1844         target_send_reply(req, rc, fail);
1845         return 0;
1846 }
1847
1848 /* Update the server data on disk.  This stores the new mount_count and
1849  * also the last_rcvd value to disk.  If we don't have a clean shutdown,
1850  * then the server last_rcvd value may be less than that of the clients.
1851  * This will alert us that we may need to do client recovery.
1852  *
1853  * Also assumes for mds_last_transno that we are not modifying it (no locking).
1854  */
1855 int mds_update_server_data(struct obd_device *obd, int force_sync)
1856 {
1857         struct mds_obd *mds = &obd->u.mds;
1858         struct mds_server_data *msd = mds->mds_server_data;
1859         struct file *filp = mds->mds_rcvd_filp;
1860         struct lvfs_run_ctxt saved;
1861         loff_t off = 0;
1862         int rc;
1863         ENTRY;
1864
1865         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1866         msd->msd_last_transno = cpu_to_le64(mds->mds_last_transno);
1867
1868         CDEBUG(D_SUPER, "MDS mount_count is "LPU64", last_transno is "LPU64"\n",
1869                mds->mds_mount_count, mds->mds_last_transno);
1870         rc = fsfilt_write_record(obd, filp, msd, sizeof(*msd), &off,force_sync);
1871         if (rc)
1872                 CERROR("error writing MDS server data: rc = %d\n", rc);
1873         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1874
1875         RETURN(rc);
1876 }
1877
1878 /* mount the file system (secretly) */
1879 static int mds_setup(struct obd_device *obd, obd_count len, void *buf)
1880 {
1881         struct lustre_cfg* lcfg = buf;
1882         struct mds_obd *mds = &obd->u.mds;
1883         char *options = NULL;
1884         struct vfsmount *mnt;
1885         unsigned long page;
1886         int rc = 0;
1887         ENTRY;
1888
1889         dev_clear_rdonly(2);
1890
1891         if (!lcfg->lcfg_inlbuf1 || !lcfg->lcfg_inlbuf2)
1892                 RETURN(rc = -EINVAL);
1893
1894         obd->obd_fsops = fsfilt_get_ops(lcfg->lcfg_inlbuf2);
1895         if (IS_ERR(obd->obd_fsops))
1896                 RETURN(rc = PTR_ERR(obd->obd_fsops));
1897
1898         mds->mds_max_mdsize = sizeof(struct lov_mds_md);
1899
1900         page = __get_free_page(GFP_KERNEL);
1901         if (!page)
1902                 RETURN(-ENOMEM);
1903
1904         options = (char *)page;
1905         memset(options, 0, PAGE_SIZE);
1906
1907         /* here we use "iopen_nopriv" hardcoded, because it affects MDS utility
1908          * and the rest of options are passed by mount options. Probably this
1909          * should be moved to somewhere else like startup scripts or lconf. */
1910         sprintf(options, "iopen_nopriv");
1911
1912         if (lcfg->lcfg_inllen4 > 0 && lcfg->lcfg_inlbuf4)
1913                 sprintf(options + strlen(options), ",%s",
1914                         lcfg->lcfg_inlbuf4);
1915
1916         /* we have to know mdsnum before touching underlying fs -bzzz */
1917         if (lcfg->lcfg_inllen5 > 0 && lcfg->lcfg_inlbuf5 &&
1918             strcmp(lcfg->lcfg_inlbuf5, "dumb")) {
1919                 class_uuid_t uuid;
1920
1921                 CDEBUG(D_OTHER, "MDS: %s is master for %s\n",
1922                        obd->obd_name, lcfg->lcfg_inlbuf5);
1923
1924                 generate_random_uuid(uuid);
1925                 class_uuid_unparse(uuid, &mds->mds_lmv_uuid);
1926
1927                 OBD_ALLOC(mds->mds_lmv_name, lcfg->lcfg_inllen5);
1928                 if (mds->mds_lmv_name == NULL)
1929                         RETURN(rc = -ENOMEM);
1930
1931                 memcpy(mds->mds_lmv_name, lcfg->lcfg_inlbuf5,
1932                        lcfg->lcfg_inllen5);
1933
1934                 rc = mds_lmv_connect(obd, mds->mds_lmv_name);
1935                 if (rc) {
1936                         OBD_FREE(mds->mds_lmv_name, lcfg->lcfg_inllen5);
1937                         GOTO(err_ops, rc);
1938                 }
1939         }
1940
1941         /* FIXME-WANGDI: this should be reworked when we will use lmv along
1942          * with cobd, because correct mdsnum is set in mds_lmv_connect(). */
1943         if (lcfg->lcfg_inllen6 > 0 && lcfg->lcfg_inlbuf6 && !mds->mds_lmv_obd &&
1944             strcmp(lcfg->lcfg_inlbuf6, "dumb")) {
1945                 if (!memcmp(lcfg->lcfg_inlbuf6, "master", strlen("master")) &&
1946                     mds->mds_num == 0) {
1947                         mds->mds_num = REAL_MDS_NUMBER;
1948                 } else if (!memcmp(lcfg->lcfg_inlbuf6, "cache", strlen("cache")) &&
1949                            mds->mds_num == 0) {
1950                         mds->mds_num = CACHE_MDS_NUMBER;
1951                 }
1952         }
1953
1954         mnt = do_kern_mount(lcfg->lcfg_inlbuf2, 0,
1955                             lcfg->lcfg_inlbuf1, options);
1956
1957         free_page(page);
1958
1959         if (IS_ERR(mnt)) {
1960                 rc = PTR_ERR(mnt);
1961                 CERROR("do_kern_mount failed: rc = %d\n", rc);
1962                 GOTO(err_ops, rc);
1963         }
1964
1965         CDEBUG(D_SUPER, "%s: mnt = %p\n", lcfg->lcfg_inlbuf1, mnt);
1966
1967         sema_init(&mds->mds_orphan_recovery_sem, 1);
1968         sema_init(&mds->mds_epoch_sem, 1);
1969         spin_lock_init(&mds->mds_transno_lock);
1970         mds->mds_max_cookiesize = sizeof(struct llog_cookie);
1971         atomic_set(&mds->mds_open_count, 0);
1972         atomic_set(&mds->mds_real_clients, 0);
1973
1974         obd->obd_namespace = ldlm_namespace_new(obd->obd_name,
1975                                                 LDLM_NAMESPACE_SERVER);
1976         if (obd->obd_namespace == NULL) {
1977                 mds_cleanup(obd, 0);
1978                 GOTO(err_put, rc = -ENOMEM);
1979         }
1980         ldlm_register_intent(obd->obd_namespace, mds_intent_policy);
1981
1982         rc = mds_fs_setup(obd, mnt);
1983         if (rc) {
1984                 CERROR("MDS filesystem method init failed: rc = %d\n", rc);
1985                 GOTO(err_ns, rc);
1986         }
1987
1988         rc = llog_start_commit_thread();
1989         if (rc < 0)
1990                 GOTO(err_fs, rc);
1991
1992         if (lcfg->lcfg_inllen3 > 0 && lcfg->lcfg_inlbuf3) {
1993                 class_uuid_t uuid;
1994
1995                 generate_random_uuid(uuid);
1996                 class_uuid_unparse(uuid, &mds->mds_lov_uuid);
1997
1998                 OBD_ALLOC(mds->mds_profile, lcfg->lcfg_inllen3);
1999                 if (mds->mds_profile == NULL)
2000                         GOTO(err_fs, rc = -ENOMEM);
2001
2002                 memcpy(mds->mds_profile, lcfg->lcfg_inlbuf3,
2003                        lcfg->lcfg_inllen3);
2004         }
2005
2006         ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
2007                            "mds_ldlm_client", &obd->obd_ldlm_client);
2008         obd->obd_replayable = 1;
2009
2010         rc = mds_postsetup(obd);
2011         if (rc)
2012                 GOTO(err_fs, rc);
2013
2014         RETURN(0);
2015
2016 err_fs:
2017         /* No extra cleanup needed for llog_init_commit_thread() */
2018         mds_fs_cleanup(obd, 0);
2019 err_ns:
2020         ldlm_namespace_free(obd->obd_namespace, 0);
2021         obd->obd_namespace = NULL;
2022 err_put:
2023         unlock_kernel();
2024         mntput(mds->mds_vfsmnt);
2025         mds->mds_sb = 0;
2026         lock_kernel();
2027 err_ops:
2028         fsfilt_put_ops(obd->obd_fsops);
2029         return rc;
2030 }
2031
2032 static int mds_postsetup(struct obd_device *obd)
2033 {
2034         struct mds_obd *mds = &obd->u.mds;
2035         int rc = 0;
2036         ENTRY;
2037
2038         rc = obd_llog_setup(obd, &obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT,
2039                             obd, 0, NULL, &llog_lvfs_ops);
2040         if (rc)
2041                 RETURN(rc);
2042
2043         /* This check for @dumb string is needed to handle mounting MDS
2044            with smfs. Read lconf:MDSDEV.write_conf() for more detail
2045            explanation. */
2046         if (mds->mds_profile && strcmp(mds->mds_profile, "dumb")) {
2047                 struct lvfs_run_ctxt saved;
2048                 struct lustre_profile *lprof;
2049                 struct config_llog_instance cfg;
2050
2051                 cfg.cfg_instance = NULL;
2052                 cfg.cfg_uuid = mds->mds_lov_uuid;
2053                 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2054                 rc = class_config_parse_llog(llog_get_context(&obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT),
2055                                              mds->mds_profile, &cfg);
2056                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2057                 if (rc)
2058                         GOTO(err_llog, rc);
2059
2060                 lprof = class_get_profile(mds->mds_profile);
2061                 if (lprof == NULL) {
2062                         CERROR("No profile found: %s\n", mds->mds_profile);
2063                         GOTO(err_cleanup, rc = -ENOENT);
2064                 }
2065                 rc = mds_lov_connect(obd, lprof->lp_osc);
2066                 if (rc)
2067                         GOTO(err_cleanup, rc);
2068
2069                 rc = mds_lmv_postsetup(obd);
2070                 if (rc)
2071                         GOTO(err_cleanup, rc);
2072         }
2073
2074         RETURN(rc);
2075
2076 err_cleanup:
2077         mds_lov_clean(obd);
2078 err_llog:
2079         obd_llog_cleanup(llog_get_context(&obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT));
2080         RETURN(rc);
2081 }
2082
2083 static int mds_postrecov(struct obd_device *obd)
2084
2085 {
2086         struct llog_ctxt *ctxt;
2087         int rc, rc2;
2088         ENTRY;
2089
2090         LASSERT(!obd->obd_recovering);
2091         ctxt = llog_get_context(&obd->obd_llogs, LLOG_UNLINK_ORIG_CTXT);
2092         LASSERT(ctxt != NULL);
2093
2094         rc = llog_connect(ctxt, obd->u.mds.mds_lov_desc.ld_tgt_count,
2095                           NULL, NULL, NULL);
2096         if (rc != 0) {
2097                 CERROR("faild at llog_origin_connect: %d\n", rc);
2098         }
2099
2100         rc = mds_cleanup_orphans(obd);
2101
2102         rc2 = mds_lov_set_nextid(obd);
2103         if (rc2 == 0)
2104                 rc2 = rc;
2105         RETURN(rc2);
2106 }
2107
2108 int mds_lov_clean(struct obd_device *obd)
2109 {
2110         struct mds_obd *mds = &obd->u.mds;
2111
2112         if (mds->mds_profile) {
2113                 char * cln_prof;
2114                 struct config_llog_instance cfg;
2115                 struct lvfs_run_ctxt saved;
2116                 int len = strlen(mds->mds_profile) + sizeof("-clean") + 1;
2117
2118                 OBD_ALLOC(cln_prof, len);
2119                 sprintf(cln_prof, "%s-clean", mds->mds_profile);
2120
2121                 cfg.cfg_instance = NULL;
2122                 cfg.cfg_uuid = mds->mds_lov_uuid;
2123
2124                 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2125                 class_config_parse_llog(llog_get_context(&obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT),
2126                                         cln_prof, &cfg);
2127                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2128
2129                 OBD_FREE(cln_prof, len);
2130                 OBD_FREE(mds->mds_profile, strlen(mds->mds_profile) + 1);
2131                 mds->mds_profile = NULL;
2132         }
2133         RETURN(0);
2134 }
2135
2136 int mds_lmv_clean(struct obd_device *obd)
2137 {
2138         struct mds_obd *mds = &obd->u.mds;
2139
2140         if (mds->mds_lmv_name) {
2141                 OBD_FREE(mds->mds_lmv_name, strlen(mds->mds_lmv_name) + 1);
2142                 mds->mds_lmv_name = NULL;
2143         }
2144         RETURN(0);
2145 }
2146
2147 static int mds_precleanup(struct obd_device *obd, int flags)
2148 {
2149         int rc = 0;
2150         ENTRY;
2151
2152         mds_lmv_clean(obd);
2153         mds_lov_disconnect(obd, flags);
2154         mds_lov_clean(obd);
2155         obd_llog_cleanup(llog_get_context(&obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT));
2156         RETURN(rc);
2157 }
2158
2159 static int mds_cleanup(struct obd_device *obd, int flags)
2160 {
2161         struct mds_obd *mds = &obd->u.mds;
2162         ENTRY;
2163
2164         if (mds->mds_sb == NULL)
2165                 RETURN(0);
2166
2167         mds_update_server_data(obd, 1);
2168         if (mds->mds_lov_objids != NULL) {
2169                 OBD_FREE(mds->mds_lov_objids,
2170                          mds->mds_lov_desc.ld_tgt_count * sizeof(obd_id));
2171         }
2172         mds_fs_cleanup(obd, flags);
2173
2174         unlock_kernel();
2175
2176         /* 2 seems normal on mds, (may_umount() also expects 2
2177           fwiw), but we only see 1 at this point in obdfilter. */
2178         if (atomic_read(&obd->u.mds.mds_vfsmnt->mnt_count) > 2)
2179                 CERROR("%s: mount busy, mnt_count %d != 2\n", obd->obd_name,
2180                        atomic_read(&obd->u.mds.mds_vfsmnt->mnt_count));
2181
2182         mntput(mds->mds_vfsmnt);
2183
2184         mds->mds_sb = 0;
2185
2186         ldlm_namespace_free(obd->obd_namespace, flags & OBD_OPT_FORCE);
2187
2188         spin_lock_bh(&obd->obd_processing_task_lock);
2189         if (obd->obd_recovering) {
2190                 target_cancel_recovery_timer(obd);
2191                 obd->obd_recovering = 0;
2192         }
2193         spin_unlock_bh(&obd->obd_processing_task_lock);
2194
2195         lock_kernel();
2196         dev_clear_rdonly(2);
2197         fsfilt_put_ops(obd->obd_fsops);
2198
2199         RETURN(0);
2200 }
2201
2202 static void fixup_handle_for_resent_req(struct ptlrpc_request *req,
2203                                         struct ldlm_lock *new_lock,
2204                                         struct lustre_handle *lockh)
2205 {
2206         struct obd_export *exp = req->rq_export;
2207         struct obd_device *obd = exp->exp_obd;
2208         struct ldlm_request *dlmreq =
2209                 lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*dlmreq));
2210         struct lustre_handle remote_hdl = dlmreq->lock_handle1;
2211         struct list_head *iter;
2212
2213         if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))
2214                 return;
2215
2216         l_lock(&obd->obd_namespace->ns_lock);
2217         list_for_each(iter, &exp->exp_ldlm_data.led_held_locks) {
2218                 struct ldlm_lock *lock;
2219                 lock = list_entry(iter, struct ldlm_lock, l_export_chain);
2220                 if (lock == new_lock)
2221                         continue;
2222                 if (lock->l_remote_handle.cookie == remote_hdl.cookie) {
2223                         lockh->cookie = lock->l_handle.h_cookie;
2224                         DEBUG_REQ(D_HA, req, "restoring lock cookie "LPX64,
2225                                   lockh->cookie);
2226                         l_unlock(&obd->obd_namespace->ns_lock);
2227                         return;
2228                 }
2229         }
2230         l_unlock(&obd->obd_namespace->ns_lock);
2231
2232         /* If the xid matches, then we know this is a resent request,
2233          * and allow it. (It's probably an OPEN, for which we don't
2234          * send a lock */
2235         if (req->rq_xid == exp->exp_mds_data.med_mcd->mcd_last_xid)
2236                 return;
2237
2238         /* This remote handle isn't enqueued, so we never received or
2239          * processed this request.  Clear MSG_RESENT, because it can
2240          * be handled like any normal request now. */
2241
2242         lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
2243
2244         DEBUG_REQ(D_HA, req, "no existing lock with rhandle "LPX64,
2245                   remote_hdl.cookie);
2246 }
2247
2248 int intent_disposition(struct ldlm_reply *rep, int flag)
2249 {
2250         if (!rep)
2251                 return 0;
2252         return (rep->lock_policy_res1 & flag);
2253 }
2254
2255 void intent_set_disposition(struct ldlm_reply *rep, int flag)
2256 {
2257         if (!rep)
2258                 return;
2259         rep->lock_policy_res1 |= flag;
2260 }
2261
2262 static int mds_intent_policy(struct ldlm_namespace *ns,
2263                              struct ldlm_lock **lockp, void *req_cookie,
2264                              ldlm_mode_t mode, int flags, void *data)
2265 {
2266         struct ptlrpc_request *req = req_cookie;
2267         struct ldlm_lock *lock = *lockp;
2268         struct ldlm_intent *it;
2269         struct mds_obd *mds = &req->rq_export->exp_obd->u.mds;
2270         struct ldlm_reply *rep;
2271         struct lustre_handle lockh = { 0 };
2272         struct ldlm_lock *new_lock;
2273         int getattr_part = MDS_INODELOCK_UPDATE;
2274         int rc, offset = 2, repsize[4] = {sizeof(struct ldlm_reply),
2275                                           sizeof(struct mds_body),
2276                                           mds->mds_max_mdsize,
2277                                           mds->mds_max_cookiesize};
2278         ENTRY;
2279
2280         LASSERT(req != NULL);
2281
2282         if (req->rq_reqmsg->bufcount <= 1) {
2283                 /* No intent was provided */
2284                 int size = sizeof(struct ldlm_reply);
2285                 rc = lustre_pack_reply(req, 1, &size, NULL);
2286                 LASSERT(rc == 0);
2287                 RETURN(0);
2288         }
2289
2290         it = lustre_swab_reqbuf(req, 1, sizeof(*it), lustre_swab_ldlm_intent);
2291         if (it == NULL) {
2292                 CERROR("Intent missing\n");
2293                 RETURN(req->rq_status = -EFAULT);
2294         }
2295
2296         LDLM_DEBUG(lock, "intent policy, opc: %s", ldlm_it2str(it->opc));
2297
2298         rc = lustre_pack_reply(req, it->opc == IT_UNLINK ? 4 : 3, repsize,
2299                                NULL);
2300         if (rc)
2301                 RETURN(req->rq_status = rc);
2302
2303         rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
2304         intent_set_disposition(rep, DISP_IT_EXECD);
2305
2306         fixup_handle_for_resent_req(req, lock, &lockh);
2307
2308         /* execute policy */
2309         switch ((long)it->opc) {
2310         case IT_OPEN:
2311         case IT_CREAT|IT_OPEN:
2312                 /* XXX swab here to assert that an mds_open reint
2313                  * packet is following */
2314                 rep->lock_policy_res2 = mds_reint(req, offset, &lockh);
2315 #if 0
2316                 /* We abort the lock if the lookup was negative and
2317                  * we did not make it to the OPEN portion */
2318                 if (!intent_disposition(rep, DISP_LOOKUP_EXECD))
2319                         RETURN(ELDLM_LOCK_ABORTED);
2320                 if (intent_disposition(rep, DISP_LOOKUP_NEG) &&
2321                     !intent_disposition(rep, DISP_OPEN_OPEN))
2322 #endif
2323                         if (rep->lock_policy_res2) {
2324                                 /* mds_open returns ENOLCK where it
2325                                  * should return zero, but it has no
2326                                  * lock to return */
2327                                 if (rep->lock_policy_res2 == ENOLCK)
2328                                         rep->lock_policy_res2 = 0;
2329                                 RETURN(ELDLM_LOCK_ABORTED);
2330                         }
2331                 break;
2332         case IT_LOOKUP:
2333                 getattr_part = MDS_INODELOCK_LOOKUP;
2334         case IT_GETATTR:
2335                 getattr_part |= MDS_INODELOCK_LOOKUP;
2336         case IT_READDIR:
2337                 rep->lock_policy_res2 = mds_getattr_name(offset, req, &lockh,
2338                                                          getattr_part);
2339                 /* FIXME: LDLM can set req->rq_status. MDS sets
2340                    policy_res{1,2} with disposition and status.
2341                    - replay: returns 0 & req->status is old status
2342                    - otherwise: returns req->status */
2343                 if (intent_disposition(rep, DISP_LOOKUP_NEG))
2344                         rep->lock_policy_res2 = 0;
2345                 if (!intent_disposition(rep, DISP_LOOKUP_POS) ||
2346                     rep->lock_policy_res2)
2347                         RETURN(ELDLM_LOCK_ABORTED);
2348                 if (req->rq_status != 0) {
2349                         LBUG();
2350                         rep->lock_policy_res2 = req->rq_status;
2351                         RETURN(ELDLM_LOCK_ABORTED);
2352                 }
2353                 break;
2354         default:
2355                 CERROR("Unhandled intent "LPD64"\n", it->opc);
2356                 LBUG();
2357         }
2358
2359         /* By this point, whatever function we called above must have either
2360          * filled in 'lockh', been an intent replay, or returned an error.  We
2361          * want to allow replayed RPCs to not get a lock, since we would just
2362          * drop it below anyways because lock replay is done separately by the
2363          * client afterwards.  For regular RPCs we want to give the new lock to
2364          * the client instead of whatever lock it was about to get. */
2365         new_lock = ldlm_handle2lock(&lockh);
2366         if (new_lock == NULL && (flags & LDLM_FL_INTENT_ONLY))
2367                 RETURN(0);
2368
2369         LASSERT(new_lock != NULL);
2370
2371         /* If we've already given this lock to a client once, then we should
2372          * have no readers or writers.  Otherwise, we should have one reader
2373          * _or_ writer ref (which will be zeroed below) before returning the
2374          * lock to a client. */
2375         if (new_lock->l_export == req->rq_export) {
2376                 LASSERT(new_lock->l_readers + new_lock->l_writers == 0);
2377         } else {
2378                 LASSERT(new_lock->l_export == NULL);
2379                 LASSERT(new_lock->l_readers + new_lock->l_writers == 1);
2380         }
2381
2382         *lockp = new_lock;
2383
2384         if (new_lock->l_export == req->rq_export) {
2385                 /* Already gave this to the client, which means that we
2386                  * reconstructed a reply. */
2387                 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) &
2388                         MSG_RESENT);
2389                 RETURN(ELDLM_LOCK_REPLACED);
2390         }
2391
2392         /* Fixup the lock to be given to the client */
2393         l_lock(&new_lock->l_resource->lr_namespace->ns_lock);
2394         new_lock->l_readers = 0;
2395         new_lock->l_writers = 0;
2396
2397         new_lock->l_export = class_export_get(req->rq_export);
2398         list_add(&new_lock->l_export_chain,
2399                  &new_lock->l_export->exp_ldlm_data.led_held_locks);
2400
2401         new_lock->l_blocking_ast = lock->l_blocking_ast;
2402         new_lock->l_completion_ast = lock->l_completion_ast;
2403
2404         memcpy(&new_lock->l_remote_handle, &lock->l_remote_handle,
2405                sizeof(lock->l_remote_handle));
2406
2407         new_lock->l_flags &= ~LDLM_FL_LOCAL;
2408
2409         LDLM_LOCK_PUT(new_lock);
2410         l_unlock(&new_lock->l_resource->lr_namespace->ns_lock);
2411
2412         RETURN(ELDLM_LOCK_REPLACED);
2413 }
2414
2415 int mds_attach(struct obd_device *dev, obd_count len, void *data)
2416 {
2417         struct lprocfs_static_vars lvars;
2418
2419         lprocfs_init_multi_vars(0, &lvars);
2420         return lprocfs_obd_attach(dev, lvars.obd_vars);
2421 }
2422
2423 int mds_detach(struct obd_device *dev)
2424 {
2425         return lprocfs_obd_detach(dev);
2426 }
2427
2428 int mdt_attach(struct obd_device *dev, obd_count len, void *data)
2429 {
2430         struct lprocfs_static_vars lvars;
2431
2432         lprocfs_init_multi_vars(1, &lvars);
2433         return lprocfs_obd_attach(dev, lvars.obd_vars);
2434 }
2435
2436 int mdt_detach(struct obd_device *dev)
2437 {
2438         return lprocfs_obd_detach(dev);
2439 }
2440
2441 static int mdt_setup(struct obd_device *obd, obd_count len, void *buf)
2442 {
2443         struct mds_obd *mds = &obd->u.mds;
2444         int rc = 0;
2445         ENTRY;
2446
2447         mds->mds_service =
2448                 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
2449                                 MDS_REQUEST_PORTAL, MDC_REPLY_PORTAL,
2450                                 mds_handle, "mds", obd->obd_proc_entry);
2451
2452         if (!mds->mds_service) {
2453                 CERROR("failed to start service\n");
2454                 RETURN(-ENOMEM);
2455         }
2456
2457         rc = ptlrpc_start_n_threads(obd, mds->mds_service, MDT_NUM_THREADS,
2458                                     "ll_mdt");
2459         if (rc)
2460                 GOTO(err_thread, rc);
2461
2462         mds->mds_setattr_service =
2463                 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
2464                                 MDS_SETATTR_PORTAL, MDC_REPLY_PORTAL,
2465                                 mds_handle, "mds_setattr",
2466                                 obd->obd_proc_entry);
2467         if (!mds->mds_setattr_service) {
2468                 CERROR("failed to start getattr service\n");
2469                 GOTO(err_thread, rc = -ENOMEM);
2470         }
2471
2472         rc = ptlrpc_start_n_threads(obd, mds->mds_setattr_service,
2473                                     MDT_NUM_THREADS, "ll_mdt_attr");
2474         if (rc)
2475                 GOTO(err_thread2, rc);
2476
2477         mds->mds_readpage_service =
2478                 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
2479                                 MDS_READPAGE_PORTAL, MDC_REPLY_PORTAL,
2480                                 mds_handle, "mds_readpage",
2481                                 obd->obd_proc_entry);
2482         if (!mds->mds_readpage_service) {
2483                 CERROR("failed to start readpage service\n");
2484                 GOTO(err_thread2, rc = -ENOMEM);
2485         }
2486
2487         rc = ptlrpc_start_n_threads(obd, mds->mds_readpage_service,
2488                                     MDT_NUM_THREADS, "ll_mdt_rdpg");
2489
2490         if (rc)
2491                 GOTO(err_thread3, rc);
2492
2493         RETURN(0);
2494
2495 err_thread3:
2496         ptlrpc_unregister_service(mds->mds_readpage_service);
2497 err_thread2:
2498         ptlrpc_unregister_service(mds->mds_setattr_service);
2499 err_thread:
2500         ptlrpc_unregister_service(mds->mds_service);
2501         return rc;
2502 }
2503
2504 static int mdt_cleanup(struct obd_device *obd, int flags)
2505 {
2506         struct mds_obd *mds = &obd->u.mds;
2507         ENTRY;
2508
2509         ptlrpc_stop_all_threads(mds->mds_readpage_service);
2510         ptlrpc_unregister_service(mds->mds_readpage_service);
2511
2512         ptlrpc_stop_all_threads(mds->mds_setattr_service);
2513         ptlrpc_unregister_service(mds->mds_setattr_service);
2514
2515         ptlrpc_stop_all_threads(mds->mds_service);
2516         ptlrpc_unregister_service(mds->mds_service);
2517
2518         RETURN(0);
2519 }
2520
2521 static struct dentry *mds_lvfs_fid2dentry(__u64 id, __u32 gen, __u64 gr,
2522                                           void *data)
2523 {
2524         struct obd_device *obd = data;
2525         struct ll_fid fid;
2526         fid.id = id;
2527         fid.generation = gen;
2528         return mds_fid2dentry(&obd->u.mds, &fid, NULL);
2529 }
2530
2531 static int mds_get_info(struct obd_export *exp, __u32 keylen,
2532                         void *key, __u32 *vallen, void *val)
2533 {
2534         struct obd_device *obd;
2535         struct mds_obd *mds;
2536         ENTRY;
2537
2538         obd = class_exp2obd(exp);
2539         if (obd == NULL) {
2540                 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
2541                        exp->exp_handle.h_cookie);
2542                 RETURN(-EINVAL);
2543         }
2544
2545         if (keylen >= strlen("reint_log") && memcmp(key, "reint_log", 9) == 0) {
2546                 /*Get log_context handle*/
2547                 unsigned long *llh_handle = val;
2548                 *vallen = sizeof(unsigned long);
2549                 *llh_handle = (unsigned long)obd->obd_llog_ctxt[LLOG_REINT_ORIG_CTXT];
2550                 RETURN(0);
2551         }
2552         if (keylen >= strlen("cache_sb") && memcmp(key, "cache_sb", 8) == 0) {
2553                 /*Get log_context handle*/
2554                 unsigned long *sb = val;
2555                 *vallen = sizeof(unsigned long);
2556                 *sb = (unsigned long)obd->u.mds.mds_sb;
2557                 RETURN(0);
2558         }
2559
2560         mds = &obd->u.mds;
2561         keylen == strlen("mdsize");
2562         if (keylen && memcmp(key, "mdsize", keylen) == 0) {
2563                 __u32 *mdsize = val;
2564                 *vallen = sizeof(*mdsize);
2565                 *mdsize = mds->mds_max_mdsize;
2566                 RETURN(0);
2567         }
2568
2569         CDEBUG(D_IOCTL, "invalid key\n");
2570         RETURN(-EINVAL);
2571
2572 }
2573 struct lvfs_callback_ops mds_lvfs_ops = {
2574         l_fid2dentry:     mds_lvfs_fid2dentry,
2575 };
2576
2577 int mds_preprw(int cmd, struct obd_export *exp, struct obdo *oa,
2578                 int objcount, struct obd_ioobj *obj,
2579                 int niocount, struct niobuf_remote *nb,
2580                 struct niobuf_local *res,
2581                 struct obd_trans_info *oti);
2582 int mds_commitrw(int cmd, struct obd_export *exp, struct obdo *oa,
2583                  int objcount, struct obd_ioobj *obj, int niocount,
2584                  struct niobuf_local *res, struct obd_trans_info *oti,
2585                  int rc);
2586
2587 /* use obd ops to offer management infrastructure */
2588 static struct obd_ops mds_obd_ops = {
2589         .o_owner           = THIS_MODULE,
2590         .o_attach          = mds_attach,
2591         .o_detach          = mds_detach,
2592         .o_connect         = mds_connect,
2593         .o_init_export     = mds_init_export,
2594         .o_destroy_export  = mds_destroy_export,
2595         .o_disconnect      = mds_disconnect,
2596         .o_setup           = mds_setup,
2597         .o_precleanup      = mds_precleanup,
2598         .o_cleanup         = mds_cleanup,
2599         .o_postrecov       = mds_postrecov,
2600         .o_statfs          = mds_obd_statfs,
2601         .o_iocontrol       = mds_iocontrol,
2602         .o_create          = mds_obd_create,
2603         .o_destroy         = mds_obd_destroy,
2604         .o_llog_init       = mds_llog_init,
2605         .o_llog_finish     = mds_llog_finish,
2606         .o_notify          = mds_notify,
2607         .o_get_info        = mds_get_info,
2608         .o_set_info        = mds_set_info,
2609         .o_preprw          = mds_preprw,
2610         .o_commitrw        = mds_commitrw,
2611 };
2612
2613 static struct obd_ops mdt_obd_ops = {
2614         .o_owner           = THIS_MODULE,
2615         .o_attach          = mdt_attach,
2616         .o_detach          = mdt_detach,
2617         .o_setup           = mdt_setup,
2618         .o_cleanup         = mdt_cleanup,
2619         .o_attach          = mdt_attach,
2620         .o_detach          = mdt_detach,
2621 };
2622
2623 static int __init mds_init(void)
2624 {
2625         struct lprocfs_static_vars lvars;
2626
2627         lprocfs_init_multi_vars(0, &lvars);
2628         class_register_type(&mds_obd_ops, NULL, lvars.module_vars,
2629                             LUSTRE_MDS_NAME);
2630         lprocfs_init_multi_vars(1, &lvars);
2631         class_register_type(&mdt_obd_ops, NULL, lvars.module_vars,
2632                             LUSTRE_MDT_NAME);
2633
2634         return 0;
2635 }
2636
2637 static void /*__exit*/ mds_exit(void)
2638 {
2639         class_unregister_type(LUSTRE_MDS_NAME);
2640         class_unregister_type(LUSTRE_MDT_NAME);
2641 }
2642
2643 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
2644 MODULE_DESCRIPTION("Lustre Metadata Server (MDS)");
2645 MODULE_LICENSE("GPL");
2646
2647 module_init(mds_init);
2648 module_exit(mds_exit);