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/obd_ost.h>
  53 #include <linux/lustre_mds.h>
  54 #include <linux/lustre_fsfilt.h>
  55 #include <linux/lustre_snap.h>
  56 #include <linux/lprocfs_status.h>
  57 #include <linux/lustre_commit_confd.h>
  58
  59 #include "mds_internal.h"
  60
  61 static int mds_intent_policy(struct ldlm_namespace *ns,
  62                              struct ldlm_lock **lockp, void *req_cookie,
  63                              ldlm_mode_t mode, int flags, void *data);
  64 static int mds_postsetup(struct obd_device *obd);
  65 static int mds_cleanup(struct obd_device *obd, int flags);
  66
  67 /* Assumes caller has already pushed into the kernel filesystem context */
  68 static int mds_sendpage(struct ptlrpc_request *req, struct file *file,
  69                         loff_t offset, int count)
  70 {
  71         struct ptlrpc_bulk_desc *desc;
  72         struct l_wait_info lwi;
  73         struct page **pages;
  74         int rc = 0, npages, i, tmpcount, tmpsize = 0;
  75         ENTRY;
  76
  77         LASSERT((offset & (PAGE_SIZE - 1)) == 0); /* I'm dubious about this */
  78
  79         npages = (count + PAGE_SIZE - 1) >> PAGE_SHIFT;
  80         OBD_ALLOC(pages, sizeof(*pages) * npages);
  81         if (!pages)
  82                 GOTO(out, rc = -ENOMEM);
  83
  84         desc = ptlrpc_prep_bulk_exp(req, npages, BULK_PUT_SOURCE,
  85                                     MDS_BULK_PORTAL);
  86         if (desc == NULL)
  87                 GOTO(out_free, rc = -ENOMEM);
  88
  89         for (i = 0, tmpcount = count; i < npages; i++, tmpcount -= tmpsize) {
  90                 tmpsize = tmpcount > PAGE_SIZE ? PAGE_SIZE : tmpcount;
  91
  92                 pages[i] = alloc_pages(GFP_KERNEL, 0);
  93                 if (pages[i] == NULL)
  94                         GOTO(cleanup_buf, rc = -ENOMEM);
  95
  96                 ptlrpc_prep_bulk_page(desc, pages[i], 0, tmpsize);
  97         }
  98
  99         for (i = 0, tmpcount = count; i < npages; i++, tmpcount -= tmpsize) {
 100                 tmpsize = tmpcount > PAGE_SIZE ? PAGE_SIZE : tmpcount;
 101                 CDEBUG(D_EXT2, "reading %u@%llu from dir %lu (size %llu)\n",
 102                        tmpsize, offset, file->f_dentry->d_inode->i_ino,
 103                        file->f_dentry->d_inode->i_size);
 104
 105                 rc = fsfilt_readpage(req->rq_export->exp_obd, file,
 106                                      kmap(pages[i]), tmpsize, &offset);
 107                 kunmap(pages[i]);
 108
 109                 if (rc != tmpsize)
 110                         GOTO(cleanup_buf, rc = -EIO);
 111         }
 112
 113         LASSERT(desc->bd_nob == count);
 114
 115         rc = ptlrpc_start_bulk_transfer(desc);
 116         if (rc)
 117                 GOTO(cleanup_buf, rc);
 118
 119         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
 120                 CERROR("obd_fail_loc=%x, fail operation rc=%d\n",
 121                        OBD_FAIL_MDS_SENDPAGE, rc);
 122                 GOTO(abort_bulk, rc);
 123         }
 124
 125         lwi = LWI_TIMEOUT(obd_timeout * HZ / 4, NULL, NULL);
 126         rc = l_wait_event(desc->bd_waitq, !ptlrpc_bulk_active(desc), &lwi);
 127         LASSERT (rc == 0 || rc == -ETIMEDOUT);
 128
 129         if (rc == 0) {
 130                 if (desc->bd_success &&
 131                     desc->bd_nob_transferred == count)
 132                         GOTO(cleanup_buf, rc);
 133
 134                 rc = -ETIMEDOUT; /* XXX should this be a different errno? */
 135         }
 136
 137         DEBUG_REQ(D_ERROR, req, "bulk failed: %s %d(%d), evicting %s@%s\n",
 138                   (rc == -ETIMEDOUT) ? "timeout" : "network error",
 139                   desc->bd_nob_transferred, count,
 140                   req->rq_export->exp_client_uuid.uuid,
 141                   req->rq_export->exp_connection->c_remote_uuid.uuid);
 142
 143         ptlrpc_fail_export(req->rq_export);
 144
 145         EXIT;
 146  abort_bulk:
 147         ptlrpc_abort_bulk (desc);
 148  cleanup_buf:
 149         for (i = 0; i < npages; i++)
 150                 if (pages[i])
 151                         __free_pages(pages[i], 0);
 152
 153         ptlrpc_free_bulk(desc);
 154  out_free:
 155         OBD_FREE(pages, sizeof(*pages) * npages);
 156  out:
 157         return rc;
 158 }
 159
 160 int mds_lock_mode_for_dir(struct obd_device *obd,
 161                               struct dentry *dentry, int mode)
 162 {
 163         int ret_mode, split;
 164
 165         /* any dir access needs couple locks:
 166          * 1) on part of dir we gonna lookup/modify in
 167          * 2) on a whole dir to protect it from concurrent splitting
 168          *    and to flush client's cache for readdir()
 169          * so, for a given mode and dentry this routine decides what
 170          * lock mode to use for lock #2:
 171          * 1) if caller's gonna lookup in dir then we need to protect
 172          *    dir from being splitted only - LCK_CR
 173          * 2) if caller's gonna modify dir then we need to protect
 174          *    dir from being splitted and to flush cache - LCK_CW
 175          * 3) if caller's gonna modify dir and that dir seems ready
 176          *    for splitting then we need to protect it from any
 177          *    type of access (lookup/modify/split) - LCK_EX -bzzz */
 178
 179         split = mds_splitting_expected(obd, dentry);
 180         if (split == MDS_NO_SPLITTABLE) {
 181                 /* this inode won't be splitted. so we need not to protect from
 182                  * just flush client's cache on modification */
 183                 ret_mode = 0;
 184                 if (mode == LCK_PW)
 185                         ret_mode = LCK_CW;
 186         } else {
 187                 if (mode == LCK_PR) {
 188                         ret_mode = LCK_CR;
 189                 } else if (mode == LCK_PW) {
 190                         /* caller gonna modify directory.we use concurrent
 191                            write lock here to retract client's cache for readdir */
 192                         ret_mode = LCK_CW;
 193                         if (split == MDS_EXPECT_SPLIT) {
 194                                 /* splitting possible. serialize any access
 195                                  * the idea is that first one seen dir is
 196                                  * splittable is given exclusive lock and
 197                                  * split directory. caller passes lock mode
 198                                  * to mds_try_to_split_dir() and splitting
 199                                  * would be done with exclusive lock only -bzzz */
 200                                 CDEBUG(D_OTHER, "%s: gonna split %u/%u\n",
 201                                        obd->obd_name,
 202                                        (unsigned) dentry->d_inode->i_ino,
 203                                        (unsigned) dentry->d_inode->i_generation);
 204                                 ret_mode = LCK_EX;
 205                         }
 206                 } else {
 207                         CWARN("unexpected lock mode %d\n", mode);
 208                         ret_mode = LCK_EX;
 209                 }
 210         }
 211         return ret_mode;
 212 }
 213
 214 /* only valid locked dentries or errors should be returned */
 215 struct dentry *mds_fid2locked_dentry(struct obd_device *obd, struct ll_fid *fid,
 216                                      struct vfsmount **mnt, int lock_mode,
 217                                      struct lustre_handle *lockh, int *mode,
 218                                      char *name, int namelen, __u64 lockpart)
 219 {
 220         struct mds_obd *mds = &obd->u.mds;
 221         struct dentry *de = mds_fid2dentry(mds, fid, mnt), *retval = de;
 222         struct ldlm_res_id res_id = { .name = {0} };
 223         int flags = 0, rc;
 224         ldlm_policy_data_t policy = { .l_inodebits = { lockpart } };
 225
 226         ENTRY;
 227
 228         if (IS_ERR(de))
 229                 RETURN(de);
 230
 231         res_id.name[0] = de->d_inode->i_ino;
 232         res_id.name[1] = de->d_inode->i_generation;
 233         lockh[1].cookie = 0;
 234 #ifdef S_PDIROPS
 235         if (name && IS_PDIROPS(de->d_inode)) {
 236                 ldlm_policy_data_t cpolicy =
 237                         { .l_inodebits = { MDS_INODELOCK_UPDATE } };
 238                 LASSERT(mode != NULL);
 239                 *mode = mds_lock_mode_for_dir(obd, de, lock_mode);
 240                 if (*mode) {
 241                         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
 242                                               res_id, LDLM_IBITS,
 243                                               &cpolicy, *mode, &flags,
 244                                               mds_blocking_ast,
 245                                               ldlm_completion_ast, NULL, NULL,
 246                                               NULL, 0, NULL, lockh + 1);
 247                         if (rc != ELDLM_OK) {
 248                                 l_dput(de);
 249                                 RETURN(ERR_PTR(-ENOLCK));
 250                         }
 251                 }
 252                 flags = 0;
 253
 254                 res_id.name[2] = full_name_hash(name, namelen);
 255
 256                 CDEBUG(D_INFO, "take lock on %lu:%u:"LPX64"\n",
 257                        de->d_inode->i_ino, de->d_inode->i_generation,
 258                        res_id.name[2]);
 259         }
 260 #else
 261 #warning "No PDIROPS support in the kernel"
 262 #endif
 263         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, res_id,
 264                               LDLM_IBITS, &policy, lock_mode, &flags,
 265                               mds_blocking_ast, ldlm_completion_ast, NULL, NULL,
 266                               NULL, 0, NULL, lockh);
 267         if (rc != ELDLM_OK) {
 268                 l_dput(de);
 269                 retval = ERR_PTR(-EIO); /* XXX translate ldlm code */
 270 #ifdef S_PDIROPS
 271                 if (lockh[1].cookie)
 272                         ldlm_lock_decref(lockh + 1, LCK_CW);
 273 #endif
 274         }
 275
 276         RETURN(retval);
 277 }
 278
 279 #ifndef DCACHE_DISCONNECTED
 280 #define DCACHE_DISCONNECTED DCACHE_NFSD_DISCONNECTED
 281 #endif
 282
 283
 284 /* Look up an entry by inode number. */
 285 /* this function ONLY returns valid dget'd dentries with an initialized inode
 286    or errors */
 287 struct dentry *mds_fid2dentry(struct mds_obd *mds, struct ll_fid *fid,
 288                               struct vfsmount **mnt)
 289 {
 290         char fid_name[32];
 291         unsigned long ino = fid->id;
 292         __u32 generation = fid->generation;
 293         struct inode *inode;
 294         struct dentry *result;
 295
 296         if (ino == 0)
 297                 RETURN(ERR_PTR(-ESTALE));
 298
 299         snprintf(fid_name, sizeof(fid_name), "0x%lx", ino);
 300
 301         CDEBUG(D_DENTRY, "--> mds_fid2dentry: ino/gen %lu/%u, sb %p\n",
 302                ino, generation, mds->mds_sb);
 303
 304         /* under ext3 this is neither supposed to return bad inodes
 305            nor NULL inodes. */
 306         result = ll_lookup_one_len(fid_name, mds->mds_fid_de, strlen(fid_name));
 307         if (IS_ERR(result))
 308                 RETURN(result);
 309
 310         inode = result->d_inode;
 311         if (!inode)
 312                 RETURN(ERR_PTR(-ENOENT));
 313
 314         /* here we disabled generation check, as root inode i_generation
 315          * of cache mds and real mds are different. */
 316         if (inode->i_ino != mds->mds_rootfid.id && generation &&
 317                         inode->i_generation != generation) {
 318                 /* we didn't find the right inode.. */
 319                 CERROR("bad inode %lu, link: %lu ct: %d or generation %u/%u\n",
 320                        inode->i_ino, (unsigned long)inode->i_nlink,
 321                        atomic_read(&inode->i_count), inode->i_generation,
 322                        generation);
 323                 dput(result);
 324                 RETURN(ERR_PTR(-ENOENT));
 325         }
 326
 327         if (mnt) {
 328                 *mnt = mds->mds_vfsmnt;
 329                 mntget(*mnt);
 330         }
 331
 332         RETURN(result);
 333 }
 334
 335
 336 /* Establish a connection to the MDS.
 337  *
 338  * This will set up an export structure for the client to hold state data
 339  * about that client, like open files, the last operation number it did
 340  * on the server, etc.
 341  */
 342 static int mds_connect(struct lustre_handle *conn, struct obd_device *obd,
 343                        struct obd_uuid *cluuid, unsigned long connect_flags)
 344 {
 345         struct obd_export *exp;
 346         struct mds_export_data *med; /*  */
 347         struct mds_client_data *mcd;
 348         int rc;
 349         ENTRY;
 350
 351         if (!conn || !obd || !cluuid)
 352                 RETURN(-EINVAL);
 353
 354         /* XXX There is a small race between checking the list and adding a
 355          * new connection for the same UUID, but the real threat (list
 356          * corruption when multiple different clients connect) is solved.
 357          *
 358          * There is a second race between adding the export to the list,
 359          * and filling in the client data below.  Hence skipping the case
 360          * of NULL mcd above.  We should already be controlling multiple
 361          * connects at the client, and we can't hold the spinlock over
 362          * memory allocations without risk of deadlocking.
 363          */
 364         rc = class_connect(conn, obd, cluuid);
 365         if (rc)
 366                 RETURN(rc);
 367         exp = class_conn2export(conn);
 368         LASSERT(exp);
 369         med = &exp->exp_mds_data;
 370
 371         OBD_ALLOC(mcd, sizeof(*mcd));
 372         if (!mcd) {
 373                 CERROR("mds: out of memory for client data\n");
 374                 GOTO(out, rc = -ENOMEM);
 375         }
 376
 377         memcpy(mcd->mcd_uuid, cluuid, sizeof(mcd->mcd_uuid));
 378         med->med_mcd = mcd;
 379
 380         rc = mds_client_add(obd, &obd->u.mds, med, -1);
 381         if (rc)
 382                 GOTO(out, rc);
 383
 384         if (!(connect_flags & OBD_OPT_MDS_CONNECTION)) {
 385                 struct mds_obd *mds = &obd->u.mds;
 386                 if (!(exp->exp_flags & OBD_OPT_REAL_CLIENT)) {
 387                         atomic_inc(&mds->mds_real_clients);
 388                         CDEBUG(D_OTHER,"%s: peer from %s is real client (%d)\n",
 389                                obd->obd_name, exp->exp_client_uuid.uuid,
 390                                atomic_read(&mds->mds_real_clients));
 391                         exp->exp_flags |= OBD_OPT_REAL_CLIENT;
 392                 }
 393                 if (mds->mds_lmv_name)
 394                         rc = mds_lmv_connect(obd, mds->mds_lmv_name);
 395         }
 396         EXIT;
 397 out:
 398         if (rc) {
 399                 OBD_FREE(mcd, sizeof(*mcd));
 400                 class_disconnect(exp, 0);
 401         }
 402         class_export_put(exp);
 403
 404         return rc;
 405 }
 406
 407 static int mds_init_export(struct obd_export *exp)
 408 {
 409         struct mds_export_data *med = &exp->exp_mds_data;
 410
 411         INIT_LIST_HEAD(&med->med_open_head);
 412         spin_lock_init(&med->med_open_lock);
 413         RETURN(0);
 414 }
 415
 416 static int mds_destroy_export(struct obd_export *export)
 417 {
 418         struct mds_export_data *med;
 419         struct obd_device *obd = export->exp_obd;
 420         struct lvfs_run_ctxt saved;
 421         int rc = 0;
 422         ENTRY;
 423
 424         med = &export->exp_mds_data;
 425         target_destroy_export(export);
 426
 427         if (obd_uuid_equals(&export->exp_client_uuid, &obd->obd_uuid))
 428                 GOTO(out, 0);
 429
 430         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 431
 432         /* Close any open files (which may also cause orphan unlinking). */
 433         spin_lock(&med->med_open_lock);
 434         while (!list_empty(&med->med_open_head)) {
 435                 struct list_head *tmp = med->med_open_head.next;
 436                 struct mds_file_data *mfd =
 437                         list_entry(tmp, struct mds_file_data, mfd_list);
 438                 BDEVNAME_DECLARE_STORAGE(btmp);
 439
 440                 /* bug 1579: fix force-closing for 2.5 */
 441                 struct dentry *dentry = mfd->mfd_dentry;
 442
 443                 list_del(&mfd->mfd_list);
 444                 spin_unlock(&med->med_open_lock);
 445
 446                 /* If you change this message, be sure to update
 447                  * replay_single:test_46 */
 448                 CERROR("force closing client file handle for %*s (%s:%lu)\n",
 449                        dentry->d_name.len, dentry->d_name.name,
 450                        ll_bdevname(dentry->d_inode->i_sb, btmp),
 451                        dentry->d_inode->i_ino);
 452                 rc = mds_mfd_close(NULL, obd, mfd,
 453                                    !(export->exp_flags & OBD_OPT_FAILOVER));
 454
 455                 if (rc)
 456                         CDEBUG(D_INODE, "Error closing file: %d\n", rc);
 457                 spin_lock(&med->med_open_lock);
 458         }
 459         spin_unlock(&med->med_open_lock);
 460         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 461
 462 out:
 463         mds_client_free(export, !(export->exp_flags & OBD_OPT_FAILOVER));
 464
 465         RETURN(rc);
 466 }
 467
 468 static int mds_disconnect(struct obd_export *exp, int flags)
 469 {
 470         struct obd_device *obd;
 471         struct mds_obd *mds;
 472         unsigned long irqflags;
 473         int rc;
 474         ENTRY;
 475
 476         LASSERT(exp);
 477         class_export_get(exp);
 478
 479         obd = class_exp2obd(exp);
 480         if (obd == NULL) {
 481                 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
 482                        exp->exp_handle.h_cookie);
 483                 RETURN(-EINVAL);
 484         }
 485         mds = &obd->u.mds;
 486
 487         if (!(exp->exp_flags & OBD_OPT_REAL_CLIENT)
 488                         && !atomic_read(&mds->mds_real_clients)) {
 489                 /* there was no client at all */
 490                 mds_lmv_disconnect(obd, flags);
 491         }
 492
 493         if ((exp->exp_flags & OBD_OPT_REAL_CLIENT)
 494                         && atomic_dec_and_test(&mds->mds_real_clients)) {
 495                 /* time to drop LMV connections */
 496                 CDEBUG(D_OTHER, "%s: last real client %s disconnected.  "
 497                        "Disconnnect from LMV now\n",
 498                        obd->obd_name, exp->exp_client_uuid.uuid);
 499                 mds_lmv_disconnect(obd, flags);
 500         }
 501
 502         spin_lock_irqsave(&exp->exp_lock, irqflags);
 503         exp->exp_flags = flags;
 504         spin_unlock_irqrestore(&exp->exp_lock, irqflags);
 505
 506         /* Disconnect early so that clients can't keep using export */
 507         rc = class_disconnect(exp, flags);
 508         ldlm_cancel_locks_for_export(exp);
 509
 510         /* complete all outstanding replies */
 511         spin_lock_irqsave(&exp->exp_lock, irqflags);
 512         while (!list_empty(&exp->exp_outstanding_replies)) {
 513                 struct ptlrpc_reply_state *rs =
 514                         list_entry(exp->exp_outstanding_replies.next,
 515                                    struct ptlrpc_reply_state, rs_exp_list);
 516                 struct ptlrpc_service *svc = rs->rs_srv_ni->sni_service;
 517
 518                 spin_lock(&svc->srv_lock);
 519                 list_del_init(&rs->rs_exp_list);
 520                 ptlrpc_schedule_difficult_reply(rs);
 521                 spin_unlock(&svc->srv_lock);
 522         }
 523         spin_unlock_irqrestore(&exp->exp_lock, irqflags);
 524
 525         class_export_put(exp);
 526         RETURN(rc);
 527 }
 528
 529 static int mds_getstatus(struct ptlrpc_request *req)
 530 {
 531         struct mds_obd *mds = mds_req2mds(req);
 532         struct mds_body *body;
 533         int rc, size = sizeof(*body);
 534         ENTRY;
 535
 536         rc = lustre_pack_reply(req, 1, &size, NULL);
 537         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_GETSTATUS_PACK)) {
 538                 CERROR("mds: out of memory for message: size=%d\n", size);
 539                 req->rq_status = -ENOMEM;       /* superfluous? */
 540                 RETURN(-ENOMEM);
 541         }
 542
 543         body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
 544         memcpy(&body->fid1, &mds->mds_rootfid, sizeof(body->fid1));
 545
 546         /* the last_committed and last_xid fields are filled in for all
 547          * replies already - no need to do so here also.
 548          */
 549         RETURN(0);
 550 }
 551
 552 int mds_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
 553                      void *data, int flag)
 554 {
 555         int do_ast;
 556         ENTRY;
 557
 558         if (flag == LDLM_CB_CANCELING) {
 559                 /* Don't need to do anything here. */
 560                 RETURN(0);
 561         }
 562
 563         /* XXX layering violation!  -phil */
 564         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 565         /* Get this: if mds_blocking_ast is racing with mds_intent_policy,
 566          * such that mds_blocking_ast is called just before l_i_p takes the
 567          * ns_lock, then by the time we get the lock, we might not be the
 568          * correct blocking function anymore.  So check, and return early, if
 569          * so. */
 570         if (lock->l_blocking_ast != mds_blocking_ast) {
 571                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 572                 RETURN(0);
 573         }
 574
 575         lock->l_flags |= LDLM_FL_CBPENDING;
 576         do_ast = (!lock->l_readers && !lock->l_writers);
 577         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 578
 579         if (do_ast) {
 580                 struct lustre_handle lockh;
 581                 int rc;
 582
 583                 LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel");
 584                 ldlm_lock2handle(lock, &lockh);
 585                 rc = ldlm_cli_cancel(&lockh);
 586                 if (rc < 0)
 587                         CERROR("ldlm_cli_cancel: %d\n", rc);
 588         } else {
 589                 LDLM_DEBUG(lock, "Lock still has references, will be "
 590                            "cancelled later");
 591         }
 592         RETURN(0);
 593 }
 594
 595 int mds_get_md(struct obd_device *obd, struct inode *inode, void *md,
 596                int *size, int lock)
 597 {
 598         int rc = 0;
 599         int lmm_size;
 600
 601         if (lock)
 602                 down(&inode->i_sem);
 603         rc = fsfilt_get_md(obd, inode, md, *size);
 604         if (lock)
 605                 up(&inode->i_sem);
 606
 607         if (rc < 0) {
 608                 CERROR("Error %d reading eadata for ino %lu\n",
 609                        rc, inode->i_ino);
 610         } else if (rc > 0) {
 611                 lmm_size = rc;
 612
 613                 if (S_ISREG(inode->i_mode))
 614                         rc = mds_convert_lov_ea(obd, inode, md, lmm_size);
 615                 if (S_ISDIR(inode->i_mode))
 616                         rc = mds_convert_mea_ea(obd, inode, md, lmm_size);
 617
 618                 if (rc == 0) {
 619                         *size = lmm_size;
 620                         rc = lmm_size;
 621                 } else if (rc > 0) {
 622                         *size = rc;
 623                 }
 624         }
 625
 626         RETURN (rc);
 627 }
 628
 629
 630 /* Call with lock=1 if you want mds_pack_md to take the i_sem.
 631  * Call with lock=0 if the caller has already taken the i_sem. */
 632 int mds_pack_md(struct obd_device *obd, struct lustre_msg *msg, int offset,
 633                 struct mds_body *body, struct inode *inode, int lock)
 634 {
 635         struct mds_obd *mds = &obd->u.mds;
 636         void *lmm;
 637         int lmm_size;
 638         int rc;
 639         ENTRY;
 640
 641         lmm = lustre_msg_buf(msg, offset, 0);
 642         if (lmm == NULL) {
 643                 /* Some problem with getting eadata when I sized the reply
 644                  * buffer... */
 645                 CDEBUG(D_INFO, "no space reserved for inode %lu MD\n",
 646                        inode->i_ino);
 647                 RETURN(0);
 648         }
 649         lmm_size = msg->buflens[offset];
 650
 651         /* I don't really like this, but it is a sanity check on the client
 652          * MD request.  However, if the client doesn't know how much space
 653          * to reserve for the MD, it shouldn't be bad to have too much space.
 654          */
 655         if (lmm_size > mds->mds_max_mdsize) {
 656                 CWARN("Reading MD for inode %lu of %d bytes > max %d\n",
 657                        inode->i_ino, lmm_size, mds->mds_max_mdsize);
 658                 // RETURN(-EINVAL);
 659         }
 660
 661         rc = mds_get_md(obd, inode, lmm, &lmm_size, lock);
 662         if (rc > 0) {
 663                 if (S_ISDIR(inode->i_mode))
 664                         body->valid |= OBD_MD_FLDIREA;
 665                 else
 666                         body->valid |= OBD_MD_FLEASIZE;
 667                 body->eadatasize = lmm_size;
 668                 rc = 0;
 669         }
 670
 671         RETURN(rc);
 672 }
 673
 674 static int mds_getattr_internal(struct obd_device *obd, struct dentry *dentry,
 675                                 struct ptlrpc_request *req,
 676                                 struct mds_body *reqbody, int reply_off)
 677 {
 678         struct mds_body *body;
 679         struct inode *inode = dentry->d_inode;
 680         int rc = 0;
 681         ENTRY;
 682
 683         if (inode == NULL && !(dentry->d_flags & DCACHE_CROSS_REF))
 684                 RETURN(-ENOENT);
 685
 686         body = lustre_msg_buf(req->rq_repmsg, reply_off, sizeof(*body));
 687         LASSERT(body != NULL);                 /* caller prepped reply */
 688
 689         if (dentry->d_flags & DCACHE_CROSS_REF) {
 690                 CDEBUG(D_OTHER, "cross reference: %lu/%lu/%lu\n",
 691                        (unsigned long) dentry->d_mdsnum,
 692                        (unsigned long) dentry->d_inum,
 693                        (unsigned long) dentry->d_generation);
 694                 body->valid |= OBD_MD_FLID | OBD_MD_MDS;
 695                 body->fid1.id = dentry->d_inum;
 696                 body->fid1.mds = dentry->d_mdsnum;
 697                 body->fid1.generation = dentry->d_generation;
 698                 RETURN(0);
 699         }
 700         mds_pack_inode2fid(obd, &body->fid1, inode);
 701         mds_pack_inode2body(obd, body, inode);
 702
 703         if ((S_ISREG(inode->i_mode) && (reqbody->valid & OBD_MD_FLEASIZE)) ||
 704             (S_ISDIR(inode->i_mode) && (reqbody->valid & OBD_MD_FLDIREA))) {
 705                 rc = mds_pack_md(obd, req->rq_repmsg, reply_off + 1, body,
 706                                  inode, 1);
 707
 708                 /* If we have LOV EA data, the OST holds size, atime, mtime */
 709                 if (!(body->valid & OBD_MD_FLEASIZE) &&
 710                     !(body->valid & OBD_MD_FLDIREA))
 711                         body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
 712                                         OBD_MD_FLATIME | OBD_MD_FLMTIME);
 713         } else if (S_ISLNK(inode->i_mode) &&
 714                    (reqbody->valid & OBD_MD_LINKNAME) != 0) {
 715                 char *symname = lustre_msg_buf(req->rq_repmsg, reply_off + 1,0);
 716                 int len;
 717
 718                 LASSERT (symname != NULL);       /* caller prepped reply */
 719                 len = req->rq_repmsg->buflens[reply_off + 1];
 720
 721                 rc = inode->i_op->readlink(dentry, symname, len);
 722                 if (rc < 0) {
 723                         CERROR("readlink failed: %d\n", rc);
 724                 } else if (rc != len - 1) {
 725                         CERROR ("Unexpected readlink rc %d: expecting %d\n",
 726                                 rc, len - 1);
 727                         rc = -EINVAL;
 728                 } else {
 729                         CDEBUG(D_INODE, "read symlink dest %s\n", symname);
 730                         body->valid |= OBD_MD_LINKNAME;
 731                         body->eadatasize = rc + 1;
 732                         symname[rc] = 0;        /* NULL terminate */
 733                         rc = 0;
 734                 }
 735         }
 736
 737         RETURN(rc);
 738 }
 739
 740 static int mds_getattr_pack_msg_cf(struct ptlrpc_request *req,
 741                                         struct dentry *dentry,
 742                                         int offset)
 743 {
 744         int rc = 0, size[1] = {sizeof(struct mds_body)};
 745         ENTRY;
 746
 747         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK)) {
 748                 CERROR("failed MDS_GETATTR_PACK test\n");
 749                 req->rq_status = -ENOMEM;
 750                 GOTO(out, rc = -ENOMEM);
 751         }
 752
 753         rc = lustre_pack_reply(req, 1, size, NULL);
 754         if (rc) {
 755                 CERROR("out of memory\n");
 756                 GOTO(out, req->rq_status = rc);
 757         }
 758
 759         EXIT;
 760  out:
 761         return(rc);
 762 }
 763
 764 static int mds_getattr_pack_msg(struct ptlrpc_request *req, struct inode *inode,
 765                                 int offset)
 766 {
 767         struct mds_obd *mds = mds_req2mds(req);
 768         struct mds_body *body;
 769         int rc = 0, size[2] = {sizeof(*body)}, bufcount = 1;
 770         ENTRY;
 771
 772         body = lustre_msg_buf(req->rq_reqmsg, offset, sizeof (*body));
 773         LASSERT(body != NULL);                 /* checked by caller */
 774         LASSERT_REQSWABBED(req, offset);       /* swabbed by caller */
 775
 776         if ((S_ISREG(inode->i_mode) && (body->valid & OBD_MD_FLEASIZE)) ||
 777             (S_ISDIR(inode->i_mode) && (body->valid & OBD_MD_FLDIREA))) {
 778                 int rc;
 779                 down(&inode->i_sem);
 780                 rc = fsfilt_get_md(req->rq_export->exp_obd, inode, NULL, 0);
 781                 up(&inode->i_sem);
 782                 CDEBUG(D_INODE, "got %d bytes MD data for inode %lu\n",
 783                        rc, inode->i_ino);
 784                 if (rc < 0) {
 785                         if (rc != -ENODATA)
 786                                 CERROR("error getting inode %lu MD: rc = %d\n",
 787                                        inode->i_ino, rc);
 788                         size[bufcount] = 0;
 789                 } else if (rc > mds->mds_max_mdsize) {
 790                         size[bufcount] = 0;
 791                         CERROR("MD size %d larger than maximum possible %u\n",
 792                                rc, mds->mds_max_mdsize);
 793                 } else {
 794                         size[bufcount] = rc;
 795                 }
 796                 bufcount++;
 797         } else if (S_ISLNK(inode->i_mode) && (body->valid & OBD_MD_LINKNAME)) {
 798                 if (inode->i_size + 1 != body->eadatasize)
 799                         CERROR("symlink size: %Lu, reply space: %d\n",
 800                                inode->i_size + 1, body->eadatasize);
 801                 size[bufcount] = min_t(int, inode->i_size+1, body->eadatasize);
 802                 bufcount++;
 803                 CDEBUG(D_INODE, "symlink size: %Lu, reply space: %d\n",
 804                        inode->i_size + 1, body->eadatasize);
 805         }
 806
 807         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK)) {
 808                 CERROR("failed MDS_GETATTR_PACK test\n");
 809                 req->rq_status = -ENOMEM;
 810                 GOTO(out, rc = -ENOMEM);
 811         }
 812
 813         rc = lustre_pack_reply(req, bufcount, size, NULL);
 814         if (rc) {
 815                 CERROR("out of memory\n");
 816                 GOTO(out, req->rq_status = rc);
 817         }
 818
 819         EXIT;
 820  out:
 821         return(rc);
 822 }
 823
 824 int mds_check_mds_num(struct obd_device *obd, struct inode* inode,
 825                       char *name, int namelen)
 826 {
 827         struct mea *mea = NULL;
 828         int mea_size, rc = 0;
 829         ENTRY;
 830
 831         rc = mds_get_lmv_attr(obd, inode, &mea, &mea_size);
 832         if (rc)
 833                 RETURN(rc);
 834         if (mea != NULL) {
 835                 /* dir is already splitted, check if requested filename
 836                  * should live at this MDS or at another one */
 837                 int i;
 838                 i = mea_name2idx(mea, name, namelen - 1);
 839                 if (mea->mea_master != mea->mea_fids[i].mds) {
 840                         CDEBUG(D_OTHER,
 841                                "inapropriate MDS(%d) for %s. should be %d(%d)\n",
 842                                mea->mea_master, name, mea->mea_fids[i].mds, i);
 843                         rc = -ERESTART;
 844                 }
 845         }
 846
 847         if (mea)
 848                 OBD_FREE(mea, mea_size);
 849         RETURN(rc);
 850 }
 851
 852 static int mds_getattr_name(int offset, struct ptlrpc_request *req,
 853                             struct lustre_handle *child_lockh, int child_part)
 854 {
 855         struct obd_device *obd = req->rq_export->exp_obd;
 856         struct ldlm_reply *rep = NULL;
 857         struct lvfs_run_ctxt saved;
 858         struct mds_body *body;
 859         struct dentry *dparent = NULL, *dchild = NULL;
 860         struct lvfs_ucred uc;
 861         struct lustre_handle parent_lockh[2];
 862         int namesize, update_mode;
 863         int rc = 0, cleanup_phase = 0, resent_req = 0;
 864         struct clonefs_info *clone_info = NULL;
 865         char *name;
 866         ENTRY;
 867
 868         LASSERT(!strcmp(obd->obd_type->typ_name, LUSTRE_MDS_NAME));
 869
 870         MDS_UPDATE_COUNTER((&obd->u.mds), MDS_GETATTR_NAME_COUNT);
 871
 872         /* Swab now, before anyone looks inside the request */
 873         body = lustre_swab_reqbuf(req, offset, sizeof(*body),
 874                                   lustre_swab_mds_body);
 875         if (body == NULL) {
 876                 CERROR("Can't swab mds_body\n");
 877                 GOTO(cleanup, rc = -EFAULT);
 878         }
 879
 880         LASSERT_REQSWAB(req, offset + 1);
 881         name = lustre_msg_string(req->rq_reqmsg, offset + 1, 0);
 882         if (name == NULL) {
 883                 CERROR("Can't unpack name\n");
 884                 GOTO(cleanup, rc = -EFAULT);
 885         }
 886 #if CONFIG_SNAPFS
 887         clone_info = lustre_swab_reqbuf(req, offset + 2, sizeof(*clone_info),
 888                                         lustre_swab_clonefs_info);
 889         if (clone_info) {
 890                 CDEBUG(D_INFO,"getattr name %s clone_info index %d \n", name,
 891                        clone_info->clone_index);
 892         }
 893 #endif
 894         namesize = req->rq_reqmsg->buflens[offset + 1];
 895
 896         LASSERT (offset == 0 || offset == 2);
 897         /* if requests were at offset 2, the getattr reply goes back at 1 */
 898         if (offset) {
 899                 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
 900                 offset = 1;
 901         }
 902
 903         uc.luc_fsuid = body->fsuid;
 904         uc.luc_fsgid = body->fsgid;
 905         uc.luc_cap = body->capability;
 906         uc.luc_suppgid1 = body->suppgid;
 907         uc.luc_suppgid2 = -1;
 908         push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
 909         cleanup_phase = 1; /* kernel context */
 910         intent_set_disposition(rep, DISP_LOOKUP_EXECD);
 911
 912         LASSERT(namesize > 0);
 913         if (namesize == 1) {
 914                 /* we have no dentry here, drop LOOKUP bit */
 915                 child_part &= ~MDS_INODELOCK_LOOKUP;
 916                 CDEBUG(D_OTHER, "%s: request to retrieve attrs for %lu/%lu\n",
 917                        obd->obd_name, (unsigned long) body->fid1.id,
 918                        (unsigned long) body->fid1.generation);
 919                 dchild = mds_fid2locked_dentry(obd, &body->fid1, NULL, LCK_PR,
 920                                                parent_lockh, &update_mode,
 921                                                NULL, 0, child_part);
 922                 if (IS_ERR(dchild)) {
 923                         CERROR("can't find inode: %d\n", (int) PTR_ERR(dchild));
 924                         GOTO(cleanup, rc = PTR_ERR(dchild));
 925                 }
 926                 memcpy(child_lockh, parent_lockh, sizeof(parent_lockh[0]));
 927 #ifdef S_PDIROPS
 928                 if (parent_lockh[1].cookie)
 929                         ldlm_lock_decref(parent_lockh + 1, update_mode);
 930 #endif
 931                 cleanup_phase = 2;
 932                 goto fill_inode;
 933         }
 934
 935 #if HAVE_LOOKUP_RAW
 936         /* FIXME: handle raw lookup */
 937         if (body->valid == OBD_MD_FLID) {
 938                 struct mds_obd *mds = &obd->u.mds;
 939                 struct mds_body *mds_reply;
 940                 int size = sizeof(*mds_reply);
 941                 struct inode *dir;
 942                 ino_t inum;
 943                 dparent = mds_fid2dentry(mds, &body->fid1, NULL);
 944                 if (IS_ERR(dparent)) {
 945                         rc = PTR_ERR(dparent);
 946                         GOTO(cleanup, rc);
 947                 }
 948                 // The user requested ONLY the inode number, so do a raw lookup
 949                 rc = lustre_pack_reply(req, 1, &size, NULL);
 950                 if (rc) {
 951                         CERROR("out of memory\n");
 952                         l_dput(dparent);
 953                         GOTO(cleanup, rc);
 954                 }
 955                 dir  = dparent->d_inode;
 956                 LASSERT(dir->i_op->lookup_raw != NULL);
 957                 rc = dir->i_op->lookup_raw(dir, name, namesize - 1, &inum);
 958                 l_dput(dparent);
 959                 mds_reply = lustre_msg_buf(req->rq_repmsg, offset,
 960                                            sizeof(*mds_reply));
 961                 mds_reply->fid1.id = inum;
 962                 mds_reply->valid = OBD_MD_FLID;
 963                 GOTO(cleanup, rc);
 964         }
 965 #endif
 966
 967         if (child_lockh->cookie != 0) {
 968                 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT);
 969                 resent_req = 1;
 970         }
 971
 972         if (resent_req == 0) {
 973                 rc = mds_get_parent_child_locked(obd, &obd->u.mds, &body->fid1,
 974                                                  parent_lockh, &dparent,
 975                                                  LCK_PR, MDS_INODELOCK_LOOKUP,
 976                                                  &update_mode, name, namesize,
 977                                                  child_lockh, &dchild, LCK_PR,
 978                                                  child_part, clone_info);
 979                 if (rc)
 980                         GOTO(cleanup, rc);
 981
 982                 /* let's make sure this name should leave on this mds node */
 983                 rc = mds_check_mds_num(obd, dparent->d_inode, name, namesize);
 984                 if (rc)
 985                         GOTO(cleanup, rc);
 986         } else {
 987                 struct ldlm_lock *granted_lock;
 988                 struct ll_fid child_fid;
 989                 struct ldlm_resource *res;
 990                 DEBUG_REQ(D_DLMTRACE, req, "resent, not enqueuing new locks");
 991                 granted_lock = ldlm_handle2lock(child_lockh);
 992                 LASSERT(granted_lock);
 993
 994                 res = granted_lock->l_resource;
 995                 child_fid.id = res->lr_name.name[0];
 996                 child_fid.generation = res->lr_name.name[1];
 997                 dchild = mds_fid2dentry(&obd->u.mds, &child_fid, NULL);
 998                 LASSERT(dchild);
 999                 LDLM_LOCK_PUT(granted_lock);
1000         }
1001
1002         cleanup_phase = 2; /* dchild, dparent, locks */
1003
1004 fill_inode:
1005
1006         if (!DENTRY_VALID(dchild)) {
1007                 intent_set_disposition(rep, DISP_LOOKUP_NEG);
1008                 /* in the intent case, the policy clears this error:
1009                    the disposition is enough */
1010                 rc = -ENOENT;
1011                 GOTO(cleanup, rc);
1012         } else {
1013                 intent_set_disposition(rep, DISP_LOOKUP_POS);
1014         }
1015
1016         if (req->rq_repmsg == NULL) {
1017                 if (dchild->d_flags & DCACHE_CROSS_REF)
1018                         rc = mds_getattr_pack_msg_cf(req, dchild, offset);
1019                 else
1020                         rc = mds_getattr_pack_msg(req, dchild->d_inode, offset);
1021                 if (rc != 0) {
1022                         CERROR ("mds_getattr_pack_msg: %d\n", rc);
1023                         GOTO (cleanup, rc);
1024                 }
1025         }
1026
1027         rc = mds_getattr_internal(obd, dchild, req, body, offset);
1028         GOTO(cleanup, rc); /* returns the lock to the client */
1029
1030  cleanup:
1031         switch (cleanup_phase) {
1032         case 2:
1033                 if (resent_req == 0) {
1034                         if (rc && DENTRY_VALID(dchild))
1035                                 ldlm_lock_decref(child_lockh, LCK_PR);
1036                         if (dparent) {
1037                                 ldlm_lock_decref(parent_lockh, LCK_PR);
1038 #ifdef S_PDIROPS
1039                                 if (parent_lockh[1].cookie != 0)
1040                                         ldlm_lock_decref(parent_lockh + 1,
1041                                                          update_mode);
1042 #endif
1043                         }
1044                         if (dparent)
1045                                 l_dput(dparent);
1046                 }
1047                 l_dput(dchild);
1048         case 1:
1049                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1050         default: ;
1051         }
1052         return rc;
1053 }
1054
1055 static int mds_getattr(int offset, struct ptlrpc_request *req)
1056 {
1057         struct mds_obd *mds = mds_req2mds(req);
1058         struct obd_device *obd = req->rq_export->exp_obd;
1059         struct lvfs_run_ctxt saved;
1060         struct dentry *de;
1061         struct mds_body *body;
1062         struct lvfs_ucred uc;
1063         int rc = 0;
1064         ENTRY;
1065
1066         body = lustre_swab_reqbuf (req, offset, sizeof (*body),
1067                                    lustre_swab_mds_body);
1068         if (body == NULL) {
1069                 CERROR ("Can't unpack body\n");
1070                 RETURN (-EFAULT);
1071         }
1072
1073         MDS_UPDATE_COUNTER(mds, MDS_GETATTR_COUNT);
1074
1075         uc.luc_fsuid = body->fsuid;
1076         uc.luc_fsgid = body->fsgid;
1077         uc.luc_cap = body->capability;
1078         push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1079         de = mds_fid2dentry(mds, &body->fid1, NULL);
1080         if (IS_ERR(de)) {
1081                 rc = req->rq_status = -ENOENT;
1082                 GOTO(out_pop, PTR_ERR(de));
1083         }
1084
1085         rc = mds_getattr_pack_msg(req, de->d_inode, offset);
1086         if (rc != 0) {
1087                 CERROR ("mds_getattr_pack_msg: %d\n", rc);
1088                 GOTO (out_pop, rc);
1089         }
1090
1091         req->rq_status = mds_getattr_internal(obd, de, req, body, 0);
1092
1093         l_dput(de);
1094         GOTO(out_pop, rc);
1095 out_pop:
1096         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1097         return rc;
1098 }
1099
1100
1101 static int mds_obd_statfs(struct obd_device *obd, struct obd_statfs *osfs,
1102                           unsigned long max_age)
1103 {
1104         int rc;
1105
1106         spin_lock(&obd->obd_osfs_lock);
1107         rc = fsfilt_statfs(obd, obd->u.mds.mds_sb, max_age);
1108         if (rc == 0)
1109                 memcpy(osfs, &obd->obd_osfs, sizeof(*osfs));
1110         spin_unlock(&obd->obd_osfs_lock);
1111
1112         return rc;
1113 }
1114
1115 static int mds_statfs(struct ptlrpc_request *req)
1116 {
1117         struct obd_device *obd = req->rq_export->exp_obd;
1118         int rc, size = sizeof(struct obd_statfs);
1119         ENTRY;
1120
1121         rc = lustre_pack_reply(req, 1, &size, NULL);
1122         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_STATFS_PACK)) {
1123                 CERROR("mds: statfs lustre_pack_reply failed: rc = %d\n", rc);
1124                 GOTO(out, rc);
1125         }
1126
1127         MDS_UPDATE_COUNTER((&obd->u.mds), MDS_STATFS_COUNT);
1128
1129         /* We call this so that we can cache a bit - 1 jiffie worth */
1130         rc = mds_obd_statfs(obd, lustre_msg_buf(req->rq_repmsg, 0, size),
1131                             jiffies - HZ);
1132         if (rc) {
1133                 CERROR("mds_obd_statfs failed: rc %d\n", rc);
1134                 GOTO(out, rc);
1135         }
1136
1137         EXIT;
1138 out:
1139         req->rq_status = rc;
1140         return 0;
1141 }
1142
1143 static int mds_sync(struct ptlrpc_request *req)
1144 {
1145         struct obd_device *obd = req->rq_export->exp_obd;
1146         struct mds_obd *mds = &obd->u.mds;
1147         struct mds_body *body;
1148         int rc, size = sizeof(*body);
1149         ENTRY;
1150
1151         body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof(*body));
1152         if (body == NULL)
1153                 GOTO(out, rc = -EPROTO);
1154
1155         rc = lustre_pack_reply(req, 1, &size, NULL);
1156         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_SYNC_PACK)) {
1157                 CERROR("fsync lustre_pack_reply failed: rc = %d\n", rc);
1158                 GOTO(out, rc);
1159         }
1160
1161         if (body->fid1.id == 0) {
1162                 /* a fid of zero is taken to mean "sync whole filesystem" */
1163                 rc = fsfilt_sync(obd, mds->mds_sb);
1164                 if (rc)
1165                         GOTO(out, rc);
1166         } else {
1167                 /* just any file to grab fsync method - "file" arg unused */
1168                 struct file *file = mds->mds_rcvd_filp;
1169                 struct dentry *de;
1170
1171                 de = mds_fid2dentry(mds, &body->fid1, NULL);
1172                 if (IS_ERR(de))
1173                         GOTO(out, rc = PTR_ERR(de));
1174
1175                 rc = file->f_op->fsync(NULL, de, 1);
1176                 l_dput(de);
1177                 if (rc)
1178                         GOTO(out, rc);
1179
1180                 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*body));
1181                 mds_pack_inode2fid(obd, &body->fid1, de->d_inode);
1182                 mds_pack_inode2body(obd, body, de->d_inode);
1183         }
1184 out:
1185         req->rq_status = rc;
1186         return 0;
1187 }
1188
1189 /* mds_readpage does not take a DLM lock on the inode, because the client must
1190  * already have a PR lock.
1191  *
1192  * If we were to take another one here, a deadlock will result, if another
1193  * thread is already waiting for a PW lock. */
1194 static int mds_readpage(struct ptlrpc_request *req)
1195 {
1196         struct obd_device *obd = req->rq_export->exp_obd;
1197         struct vfsmount *mnt;
1198         struct dentry *de;
1199         struct file *file;
1200         struct mds_body *body, *repbody;
1201         struct lvfs_run_ctxt saved;
1202         int rc, size = sizeof(*repbody);
1203         struct lvfs_ucred uc;
1204         ENTRY;
1205
1206         rc = lustre_pack_reply(req, 1, &size, NULL);
1207         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_READPAGE_PACK)) {
1208                 CERROR("mds: out of memory\n");
1209                 GOTO(out, rc = -ENOMEM);
1210         }
1211
1212         body = lustre_swab_reqbuf(req, 0, sizeof(*body), lustre_swab_mds_body);
1213         if (body == NULL)
1214                 GOTO (out, rc = -EFAULT);
1215
1216         uc.luc_fsuid = body->fsuid;
1217         uc.luc_fsgid = body->fsgid;
1218         uc.luc_cap = body->capability;
1219         push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1220         de = mds_fid2dentry(&obd->u.mds, &body->fid1, &mnt);
1221         if (IS_ERR(de))
1222                 GOTO(out_pop, rc = PTR_ERR(de));
1223
1224         CDEBUG(D_INODE, "ino %lu\n", de->d_inode->i_ino);
1225
1226         file = dentry_open(de, mnt, O_RDONLY | O_LARGEFILE);
1227         /* note: in case of an error, dentry_open puts dentry */
1228         if (IS_ERR(file))
1229                 GOTO(out_pop, rc = PTR_ERR(file));
1230
1231         /* body->size is actually the offset -eeb */
1232         if ((body->size & (de->d_inode->i_blksize - 1)) != 0) {
1233                 CERROR("offset "LPU64" not on a block boundary of %lu\n",
1234                        body->size, de->d_inode->i_blksize);
1235                 GOTO(out_file, rc = -EFAULT);
1236         }
1237
1238         /* body->nlink is actually the #bytes to read -eeb */
1239         if (body->nlink & (de->d_inode->i_blksize - 1)) {
1240                 CERROR("size %u is not multiple of blocksize %lu\n",
1241                        body->nlink, de->d_inode->i_blksize);
1242                 GOTO(out_file, rc = -EFAULT);
1243         }
1244
1245         repbody = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*repbody));
1246         repbody->size = file->f_dentry->d_inode->i_size;
1247         repbody->valid = OBD_MD_FLSIZE;
1248
1249         /* to make this asynchronous make sure that the handling function
1250            doesn't send a reply when this function completes. Instead a
1251            callback function would send the reply */
1252         /* body->size is actually the offset -eeb */
1253         rc = mds_sendpage(req, file, body->size, body->nlink);
1254
1255 out_file:
1256         filp_close(file, 0);
1257 out_pop:
1258         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1259 out:
1260         req->rq_status = rc;
1261         RETURN(0);
1262 }
1263
1264 int mds_reint(struct ptlrpc_request *req, int offset,
1265               struct lustre_handle *lockh)
1266 {
1267         struct mds_update_record *rec; /* 116 bytes on the stack?  no sir! */
1268         int rc;
1269         ENTRY;
1270
1271         OBD_ALLOC(rec, sizeof(*rec));
1272         if (rec == NULL)
1273                 RETURN(-ENOMEM);
1274
1275         rc = mds_update_unpack(req, offset, rec);
1276         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNPACK)) {
1277                 CERROR("invalid record\n");
1278                 GOTO(out, req->rq_status = -EINVAL);
1279         }
1280         /* rc will be used to interrupt a for loop over multiple records */
1281         rc = mds_reint_rec(rec, offset, req, lockh);
1282  out:
1283         OBD_FREE(rec, sizeof(*rec));
1284         RETURN(rc);
1285 }
1286
1287 static int mds_filter_recovery_request(struct ptlrpc_request *req,
1288                                        struct obd_device *obd, int *process)
1289 {
1290         switch (req->rq_reqmsg->opc) {
1291         case MDS_CONNECT: /* This will never get here, but for completeness. */
1292         case OST_CONNECT: /* This will never get here, but for completeness. */
1293         case MDS_DISCONNECT:
1294         case OST_DISCONNECT:
1295                *process = 1;
1296                RETURN(0);
1297
1298         case MDS_CLOSE:
1299         case MDS_SYNC: /* used in unmounting */
1300         case OBD_PING:
1301         case MDS_REINT:
1302         case LDLM_ENQUEUE:
1303         case OST_CREATE:
1304                 *process = target_queue_recovery_request(req, obd);
1305                 RETURN(0);
1306
1307         default:
1308                 DEBUG_REQ(D_ERROR, req, "not permitted during recovery");
1309                 *process = 0;
1310                 /* XXX what should we set rq_status to here? */
1311                 req->rq_status = -EAGAIN;
1312                 RETURN(ptlrpc_error(req));
1313         }
1314 }
1315
1316 static char *reint_names[] = {
1317         [REINT_SETATTR] "setattr",
1318         [REINT_CREATE]  "create",
1319         [REINT_LINK]    "link",
1320         [REINT_UNLINK]  "unlink",
1321         [REINT_RENAME]  "rename",
1322         [REINT_OPEN]    "open",
1323 };
1324
1325 #define FILTER_VALID_FLAGS (OBD_MD_FLTYPE | OBD_MD_FLMODE | OBD_MD_FLGENER  |\
1326                             OBD_MD_FLSIZE | OBD_MD_FLBLOCKS | OBD_MD_FLBLKSZ|\
1327                             OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME|\
1328                             OBD_MD_FLID)
1329
1330 static void reconstruct_create(struct ptlrpc_request *req)
1331 {
1332         struct mds_export_data *med = &req->rq_export->exp_mds_data;
1333         struct mds_obd *mds = &req->rq_export->exp_obd->u.mds;
1334         struct mds_client_data *mcd = med->med_mcd;
1335         struct dentry *dentry;
1336         struct ost_body *body;
1337         struct ll_fid fid;
1338         ENTRY;
1339
1340         /* copy rc, transno and disp; steal locks */
1341         mds_req_from_mcd(req, mcd);
1342         if (req->rq_status) {
1343                 EXIT;
1344                 return;
1345         }
1346
1347         fid.mds = 0;
1348         fid.id = mcd->mcd_last_data;
1349         fid.generation = 0;
1350
1351         LASSERT(fid.id != 0);
1352
1353         dentry = mds_fid2dentry(mds, &fid, NULL);
1354         if (IS_ERR(dentry)) {
1355                 CERROR("can't find inode "LPU64"\n", fid.id);
1356                 req->rq_status = PTR_ERR(dentry);
1357                 EXIT;
1358                 return;
1359         }
1360
1361         CWARN("reconstruct reply for x"LPU64" (remote ino) "LPU64" -> %lu/%u\n",
1362               req->rq_xid, fid.id, dentry->d_inode->i_ino,
1363               dentry->d_inode->i_generation);
1364
1365         body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
1366         obdo_from_inode(&body->oa, dentry->d_inode, FILTER_VALID_FLAGS);
1367         body->oa.o_id = dentry->d_inode->i_ino;
1368         body->oa.o_generation = dentry->d_inode->i_generation;
1369         body->oa.o_valid |= OBD_MD_FLID | OBD_MD_FLGENER;
1370
1371         l_dput(dentry);
1372         EXIT;
1373         return;
1374 }
1375
1376 static int mdt_obj_create(struct ptlrpc_request *req)
1377 {
1378         struct obd_device *obd = req->rq_export->exp_obd;
1379         struct mds_obd *mds = &obd->u.mds;
1380         struct ost_body *body, *repbody;
1381         char fidname[LL_FID_NAMELEN];
1382         struct inode *parent_inode;
1383         struct lvfs_run_ctxt saved;
1384         struct dentry *new = NULL;
1385         struct dentry_params dp;
1386         int mealen, flags = 0, rc, size = sizeof(*repbody), cleanup_phase = 0;
1387         struct lvfs_ucred uc;
1388         struct mea *mea;
1389         void *handle = NULL;
1390         unsigned long cr_inum = 0;
1391         ENTRY;
1392
1393         DEBUG_REQ(D_HA, req, "create remote object");
1394
1395         parent_inode = mds->mds_unnamed_dir->d_inode;
1396
1397         body = lustre_swab_reqbuf(req, 0, sizeof(*body), lustre_swab_ost_body);
1398         if (body == NULL)
1399                 RETURN(-EFAULT);
1400
1401         rc = lustre_pack_reply(req, 1, &size, NULL);
1402         if (rc)
1403                 RETURN(rc);
1404
1405         MDS_CHECK_RESENT(req, reconstruct_create(req));
1406
1407         uc.luc_fsuid = body->oa.o_uid;
1408         uc.luc_fsgid = body->oa.o_gid;
1409
1410         push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1411
1412         repbody = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*repbody));
1413
1414         if (body->oa.o_flags & OBD_FL_RECREATE_OBJS) {
1415                 /* this is re-create request from MDS holding directory name.
1416                  * we have to lookup given ino/generation first. if it exists
1417                  * (good case) then there is nothing to do. if it does not
1418                  * then we have to recreate it */
1419                 struct ll_fid fid;
1420                 fid.id = body->oa.o_id;
1421                 fid.generation = body->oa.o_generation;
1422                 new = mds_fid2dentry(mds, &fid, NULL);
1423                 if (!IS_ERR(new) && new->d_inode) {
1424                         CDEBUG(D_HA, "mkdir() repairing is on its way: %lu/%lu\n",
1425                                (unsigned long) fid.id,
1426                                (unsigned long) fid.generation);
1427                         obdo_from_inode(&repbody->oa, new->d_inode,
1428                                         FILTER_VALID_FLAGS);
1429                         repbody->oa.o_id = new->d_inode->i_ino;
1430                         repbody->oa.o_generation = new->d_inode->i_generation;
1431                         repbody->oa.o_valid |= OBD_MD_FLID | OBD_MD_FLGENER;
1432                         cleanup_phase = 1;
1433                         cr_inum = new->d_inode->i_ino;
1434                         GOTO(cleanup, rc = 0);
1435                 }
1436                 CWARN("hmm. for some reason dir %lu/%lu (or reply) got lost\n",
1437                       (unsigned long) fid.id, (unsigned long) fid.generation);
1438                 LASSERT(new->d_inode == NULL ||
1439                         new->d_inode->i_generation != fid.generation);
1440                 l_dput(new);
1441         }
1442
1443         down(&parent_inode->i_sem);
1444         handle = fsfilt_start(obd, parent_inode, FSFILT_OP_MKDIR, NULL);
1445         if (IS_ERR(handle)) {
1446                 up(&parent_inode->i_sem);
1447                 GOTO(cleanup, rc = PTR_ERR(handle));
1448         }
1449         cleanup_phase = 1; /* transaction */
1450
1451 repeat:
1452         rc = sprintf(fidname, "%u.%u", ll_insecure_random_int(), current->pid);
1453         new = lookup_one_len(fidname, mds->mds_unnamed_dir, rc);
1454         if (IS_ERR(new)) {
1455                 CERROR("%s: can't lookup new inode (%s) for mkdir: %d\n",
1456                        obd->obd_name, fidname, (int) PTR_ERR(new));
1457                 fsfilt_commit(obd, mds->mds_sb, new->d_inode, handle, 0);
1458                 up(&parent_inode->i_sem);
1459                 RETURN(PTR_ERR(new));
1460         } else if (new->d_inode) {
1461                 CERROR("%s: name exists. repeat\n", obd->obd_name);
1462                 goto repeat;
1463         }
1464
1465         new->d_fsdata = (void *) &dp;
1466         dp.p_inum = 0;
1467         dp.p_ptr = req;
1468
1469         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1470                 DEBUG_REQ(D_HA, req, "replay create obj %lu/%lu",
1471                           (unsigned long) body->oa.o_id,
1472                           (unsigned long) body->oa.o_generation);
1473                 dp.p_inum = body->oa.o_id;
1474                 dp.p_generation = body->oa.o_generation;
1475         }
1476         rc = vfs_mkdir(parent_inode, new, body->oa.o_mode);
1477         if (rc == 0) {
1478                 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1479                         LASSERTF(body->oa.o_id == new->d_inode->i_ino,
1480                                  "BUG 3550: failed to recreate obj "
1481                                  LPU64" -> %lu\n",
1482                                  body->oa.o_id, new->d_inode->i_ino);
1483                         LASSERTF(body->oa.o_generation ==
1484                                  new->d_inode->i_generation,
1485                                  "BUG 3550: failed to recreate obj/gen "
1486                                  LPU64"/%u -> %lu/%u\n",
1487                                  body->oa.o_id, body->oa.o_generation,
1488                                  new->d_inode->i_ino,
1489                                  new->d_inode->i_generation);
1490                 }
1491
1492                 obdo_from_inode(&repbody->oa, new->d_inode, FILTER_VALID_FLAGS);
1493                 repbody->oa.o_id = new->d_inode->i_ino;
1494                 repbody->oa.o_generation = new->d_inode->i_generation;
1495                 repbody->oa.o_valid |= OBD_MD_FLID | OBD_MD_FLGENER;
1496
1497                 rc = fsfilt_del_dir_entry(obd, new);
1498                 up(&parent_inode->i_sem);
1499
1500                 if (rc) {
1501                         CERROR("can't remove name for object: %d\n", rc);
1502                         GOTO(cleanup, rc);
1503                 }
1504
1505                 cleanup_phase = 2; /* created directory object */
1506
1507                 cr_inum = new->d_inode->i_ino;
1508                 CDEBUG(D_OTHER, "created dirobj: %lu/%u mode %o\n",
1509                        new->d_inode->i_ino, new->d_inode->i_generation,
1510                        new->d_inode->i_mode);
1511         } else {
1512                 up(&parent_inode->i_sem);
1513                 CERROR("%s: can't create dirobj: %d\n", obd->obd_name, rc);
1514                 GOTO(cleanup, rc);
1515         }
1516
1517         if (body->oa.o_valid & OBD_MD_FLID) {
1518                 /* this is new object for splitted dir. we have to
1519                  * prevent recursive splitting on it -bzzz */
1520                 mealen = obd_size_diskmd(mds->mds_lmv_exp, NULL);
1521                 OBD_ALLOC(mea, mealen);
1522                 if (mea == NULL)
1523                         GOTO(cleanup, rc = -ENOMEM);
1524                 mea->mea_magic = MEA_MAGIC_ALL_CHARS;
1525                 mea->mea_master = 0;
1526                 mea->mea_count = 0;
1527                 down(&new->d_inode->i_sem);
1528                 rc = fsfilt_set_md(obd, new->d_inode, handle, mea, mealen);
1529                 up(&new->d_inode->i_sem);
1530                 OBD_FREE(mea, mealen);
1531                 CDEBUG(D_OTHER, "%s: mark non-splittable %lu/%u - %d\n",
1532                        obd->obd_name, new->d_inode->i_ino,
1533                        new->d_inode->i_generation, flags);
1534         } else if (body->oa.o_easize) {
1535                 /* we pass LCK_EX to split routine to signal that we have
1536                  * exclusive access to the directory. simple because nobody
1537                  * knows it already exists -bzzz */
1538                 mds_try_to_split_dir(obd, new, NULL, body->oa.o_easize, LCK_EX);
1539         }
1540
1541 cleanup:
1542         switch (cleanup_phase) {
1543         case 2: /* object has been created, but we'll may want to replay it later */
1544                 if (rc == 0)
1545                         ptlrpc_require_repack(req);
1546         case 1: /* transaction */
1547                 rc = mds_finish_transno(mds, parent_inode, handle, req, rc, cr_inum);
1548         }
1549
1550         l_dput(new);
1551         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1552         RETURN(rc);
1553 }
1554
1555 static int mdt_get_info(struct ptlrpc_request *req)
1556 {
1557         char *key;
1558         struct obd_export *exp = req->rq_export;
1559         int keylen, rc = 0, size = sizeof(obd_id);
1560         obd_id *reply;
1561         ENTRY;
1562
1563         key = lustre_msg_buf(req->rq_reqmsg, 0, 1);
1564         if (key == NULL) {
1565                 DEBUG_REQ(D_HA, req, "no get_info key");
1566                 RETURN(-EFAULT);
1567         }
1568         keylen = req->rq_reqmsg->buflens[0];
1569
1570         if (keylen < strlen("mdsize") || memcmp(key, "mdsize", 6) != 0)
1571                 RETURN(-EPROTO);
1572
1573         rc = lustre_pack_reply(req, 1, &size, NULL);
1574         if (rc)
1575                 RETURN(rc);
1576
1577         reply = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*reply));
1578         rc = obd_get_info(exp, keylen, key, &size, reply);
1579         req->rq_repmsg->status = 0;
1580         RETURN(rc);
1581 }
1582
1583 static int mds_set_info(struct obd_export *exp, __u32 keylen,
1584                         void *key, __u32 vallen, void *val)
1585 {
1586         struct obd_device *obd;
1587         struct mds_obd *mds;
1588         int    rc = 0;
1589         ENTRY;
1590
1591         obd = class_exp2obd(exp);
1592         if (obd == NULL) {
1593                 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
1594                        exp->exp_handle.h_cookie);
1595                 RETURN(-EINVAL);
1596         }
1597
1598 #define KEY_IS(str) \
1599         (keylen == strlen(str) && memcmp(key, str, keylen) == 0)
1600
1601         mds = &obd->u.mds;
1602         if (KEY_IS("mds_num")) {
1603                 int valsize;
1604                 __u32 group;
1605                 CDEBUG(D_IOCTL, "set mds num %d\n", *(int*)val);
1606                 mds->mds_num = *(int*)val;
1607                 group = FILTER_GROUP_FIRST_MDS + mds->mds_num;
1608                 valsize = sizeof(group);
1609                 /*mds number has been changed, so the corresponding obdfilter exp
1610                  *need to be changed too*/
1611                 rc = obd_set_info(mds->mds_osc_exp, strlen("mds_conn"), "mds_conn",
1612                           valsize, &group);
1613                 RETURN(rc);
1614         }
1615         CDEBUG(D_IOCTL, "invalid key\n");
1616         RETURN(-EINVAL);
1617 }
1618
1619 static int mdt_set_info(struct ptlrpc_request *req)
1620 {
1621         char *key, *val;
1622         struct obd_export *exp = req->rq_export;
1623         int keylen, rc = 0, vallen;
1624         ENTRY;
1625
1626         key = lustre_msg_buf(req->rq_reqmsg, 0, 1);
1627         if (key == NULL) {
1628                 DEBUG_REQ(D_HA, req, "no set_info key");
1629                 RETURN(-EFAULT);
1630         }
1631         keylen = req->rq_reqmsg->buflens[0];
1632
1633         if (keylen == strlen("mds_num") &&
1634             memcmp(key, "mds_num", keylen) == 0) {
1635                 rc = lustre_pack_reply(req, 0, NULL, NULL);
1636                 if (rc)
1637                         RETURN(rc);
1638                 val = lustre_msg_buf(req->rq_reqmsg, 1, 0);
1639
1640                 vallen = req->rq_reqmsg->buflens[1];
1641
1642                 rc = obd_set_info(exp, keylen, key, vallen, val);
1643                 req->rq_repmsg->status = 0;
1644                 RETURN(rc);
1645         } else if (keylen == strlen("client") &&
1646                    memcmp(key, "client", keylen) == 0) {
1647                 rc = lustre_pack_reply(req, 0, NULL, NULL);
1648                 if (rc)
1649                         RETURN(rc);
1650                 rc = obd_set_info(exp, keylen, key, sizeof(obd_id), NULL);
1651                 req->rq_repmsg->status = 0;
1652                 RETURN(rc);
1653         }
1654         CDEBUG(D_IOCTL, "invalid key\n");
1655         RETURN(-EINVAL);
1656 }
1657
1658 int mds_handle(struct ptlrpc_request *req)
1659 {
1660         int should_process, fail = OBD_FAIL_MDS_ALL_REPLY_NET;
1661         int rc = 0;
1662         struct mds_obd *mds = NULL; /* quell gcc overwarning */
1663         struct obd_device *obd = NULL;
1664         ENTRY;
1665
1666         OBD_FAIL_RETURN(OBD_FAIL_MDS_ALL_REQUEST_NET | OBD_FAIL_ONCE, 0);
1667
1668         LASSERT(current->journal_info == NULL);
1669         /* XXX identical to OST */
1670         if (req->rq_reqmsg->opc != MDS_CONNECT) {
1671                 struct mds_export_data *med;
1672                 int recovering;
1673
1674                 if (req->rq_export == NULL) {
1675                         CERROR("lustre_mds: operation %d on unconnected MDS\n",
1676                                req->rq_reqmsg->opc);
1677                         req->rq_status = -ENOTCONN;
1678                         GOTO(out, rc = -ENOTCONN);
1679                 }
1680
1681                 med = &req->rq_export->exp_mds_data;
1682                 obd = req->rq_export->exp_obd;
1683                 mds = &obd->u.mds;
1684
1685                 /* sanity check: if the xid matches, the request must
1686                  * be marked as a resent or replayed */
1687                 if (req->rq_xid == med->med_mcd->mcd_last_xid)
1688                         LASSERTF(lustre_msg_get_flags(req->rq_reqmsg) &
1689                                  (MSG_RESENT | MSG_REPLAY),
1690                                  "rq_xid "LPU64" matches last_xid, "
1691                                  "expected RESENT flag\n",
1692                                  req->rq_xid);
1693                 /* else: note the opposite is not always true; a
1694                  * RESENT req after a failover will usually not match
1695                  * the last_xid, since it was likely never
1696                  * committed. A REPLAYed request will almost never
1697                  * match the last xid, however it could for a
1698                  * committed, but still retained, open. */
1699
1700                 spin_lock_bh(&obd->obd_processing_task_lock);
1701                 recovering = obd->obd_recovering;
1702                 spin_unlock_bh(&obd->obd_processing_task_lock);
1703                 if (recovering) {
1704                         rc = mds_filter_recovery_request(req, obd,
1705                                                          &should_process);
1706                         if (rc || should_process == 0) {
1707                                 RETURN(rc);
1708                         } else if (should_process < 0) {
1709                                 req->rq_status = should_process;
1710                                 rc = ptlrpc_error(req);
1711                                 RETURN(rc);
1712                         }
1713                 }
1714         }
1715
1716         switch (req->rq_reqmsg->opc) {
1717         case MDS_CONNECT:
1718                 DEBUG_REQ(D_INODE, req, "connect");
1719                 OBD_FAIL_RETURN(OBD_FAIL_MDS_CONNECT_NET, 0);
1720                 rc = target_handle_connect(req);
1721                 if (!rc)
1722                         /* Now that we have an export, set mds. */
1723                         mds = mds_req2mds(req);
1724                 break;
1725
1726         case MDS_DISCONNECT:
1727                 DEBUG_REQ(D_INODE, req, "disconnect");
1728                 OBD_FAIL_RETURN(OBD_FAIL_MDS_DISCONNECT_NET, 0);
1729                 rc = target_handle_disconnect(req);
1730                 req->rq_status = rc;            /* superfluous? */
1731                 break;
1732
1733         case MDS_GETSTATUS:
1734                 DEBUG_REQ(D_INODE, req, "getstatus");
1735                 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETSTATUS_NET, 0);
1736                 rc = mds_getstatus(req);
1737                 break;
1738
1739         case MDS_GETATTR:
1740                 DEBUG_REQ(D_INODE, req, "getattr");
1741                 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_NET, 0);
1742                 rc = mds_getattr(0, req);
1743                 break;
1744
1745         case MDS_GETATTR_NAME: {
1746                 struct lustre_handle lockh;
1747                 DEBUG_REQ(D_INODE, req, "getattr_name");
1748                 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_NAME_NET, 0);
1749
1750                 /* If this request gets a reconstructed reply, we won't be
1751                  * acquiring any new locks in mds_getattr_name, so we don't
1752                  * want to cancel.
1753                  */
1754                 lockh.cookie = 0;
1755                 rc = mds_getattr_name(0, req, &lockh, MDS_INODELOCK_UPDATE);
1756                 /* this non-intent call (from an ioctl) is special */
1757                 req->rq_status = rc;
1758                 if (rc == 0 && lockh.cookie)
1759                         ldlm_lock_decref(&lockh, LCK_PR);
1760                 break;
1761         }
1762         case MDS_STATFS:
1763                 DEBUG_REQ(D_INODE, req, "statfs");
1764                 OBD_FAIL_RETURN(OBD_FAIL_MDS_STATFS_NET, 0);
1765                 rc = mds_statfs(req);
1766                 break;
1767
1768         case MDS_READPAGE:
1769                 DEBUG_REQ(D_INODE, req, "readpage");
1770                 OBD_FAIL_RETURN(OBD_FAIL_MDS_READPAGE_NET, 0);
1771                 rc = mds_readpage(req);
1772
1773                 if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_MDS_SENDPAGE)) {
1774                         if (req->rq_reply_state) {
1775                                 lustre_free_reply_state (req->rq_reply_state);
1776                                 req->rq_reply_state = NULL;
1777                         }
1778                         RETURN(0);
1779                 }
1780
1781                 break;
1782
1783         case MDS_REINT: {
1784                 __u32 *opcp = lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*opcp));
1785                 __u32  opc;
1786                 int size[3] = {sizeof(struct mds_body), mds->mds_max_mdsize,
1787                                mds->mds_max_cookiesize};
1788                 int bufcount;
1789
1790                 /* NB only peek inside req now; mds_reint() will swab it */
1791                 if (opcp == NULL) {
1792                         CERROR ("Can't inspect opcode\n");
1793                         rc = -EINVAL;
1794                         break;
1795                 }
1796                 opc = *opcp;
1797                 if (lustre_msg_swabbed (req->rq_reqmsg))
1798                         __swab32s(&opc);
1799
1800                 DEBUG_REQ(D_INODE, req, "reint %d (%s)", opc,
1801                           (opc < sizeof(reint_names) / sizeof(reint_names[0]) ||
1802                            reint_names[opc] == NULL) ? reint_names[opc] :
1803                                                        "unknown opcode");
1804
1805                 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET, 0);
1806
1807                 if (opc == REINT_UNLINK || opc == REINT_RENAME)
1808                         bufcount = 3;
1809                 else if (opc == REINT_OPEN)
1810                         bufcount = 2;
1811                 else
1812                         bufcount = 1;
1813
1814                 rc = lustre_pack_reply(req, bufcount, size, NULL);
1815                 if (rc)
1816                         break;
1817
1818                 rc = mds_reint(req, 0, NULL);
1819                 fail = OBD_FAIL_MDS_REINT_NET_REP;
1820                 break;
1821         }
1822
1823         case MDS_CLOSE:
1824                 DEBUG_REQ(D_INODE, req, "close");
1825                 OBD_FAIL_RETURN(OBD_FAIL_MDS_CLOSE_NET, 0);
1826                 rc = mds_close(req);
1827                 break;
1828
1829         case MDS_DONE_WRITING:
1830                 DEBUG_REQ(D_INODE, req, "done_writing");
1831                 OBD_FAIL_RETURN(OBD_FAIL_MDS_DONE_WRITING_NET, 0);
1832                 rc = mds_done_writing(req);
1833                 break;
1834
1835         case MDS_PIN:
1836                 DEBUG_REQ(D_INODE, req, "pin");
1837                 OBD_FAIL_RETURN(OBD_FAIL_MDS_PIN_NET, 0);
1838                 rc = mds_pin(req);
1839                 break;
1840
1841         case MDS_SYNC:
1842                 DEBUG_REQ(D_INODE, req, "sync");
1843                 OBD_FAIL_RETURN(OBD_FAIL_MDS_SYNC_NET, 0);
1844                 rc = mds_sync(req);
1845                 break;
1846
1847         case OBD_PING:
1848                 DEBUG_REQ(D_INODE, req, "ping");
1849                 rc = target_handle_ping(req);
1850                 break;
1851
1852         case OBD_LOG_CANCEL:
1853                 CDEBUG(D_INODE, "log cancel\n");
1854                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOG_CANCEL_NET, 0);
1855                 rc = -ENOTSUPP; /* la la la */
1856                 break;
1857
1858         case LDLM_ENQUEUE:
1859                 DEBUG_REQ(D_INODE, req, "enqueue");
1860                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_ENQUEUE, 0);
1861                 rc = ldlm_handle_enqueue(req, ldlm_server_completion_ast,
1862                                          ldlm_server_blocking_ast, NULL);
1863                 break;
1864         case LDLM_CONVERT:
1865                 DEBUG_REQ(D_INODE, req, "convert");
1866                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CONVERT, 0);
1867                 rc = ldlm_handle_convert(req);
1868                 break;
1869         case LDLM_BL_CALLBACK:
1870         case LDLM_CP_CALLBACK:
1871                 DEBUG_REQ(D_INODE, req, "callback");
1872                 CERROR("callbacks should not happen on MDS\n");
1873                 LBUG();
1874                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_BL_CALLBACK, 0);
1875                 break;
1876         case LLOG_ORIGIN_HANDLE_OPEN:
1877                 DEBUG_REQ(D_INODE, req, "llog_init");
1878                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1879                 rc = llog_origin_handle_open(req);
1880                 break;
1881         case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
1882                 DEBUG_REQ(D_INODE, req, "llog next block");
1883                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1884                 rc = llog_origin_handle_next_block(req);
1885                 break;
1886         case LLOG_ORIGIN_HANDLE_PREV_BLOCK:
1887                 DEBUG_REQ(D_INODE, req, "llog prev block");
1888                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1889                 rc = llog_origin_handle_prev_block(req);
1890                 break;
1891         case LLOG_ORIGIN_HANDLE_READ_HEADER:
1892                 DEBUG_REQ(D_INODE, req, "llog read header");
1893                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1894                 rc = llog_origin_handle_read_header(req);
1895                 break;
1896         case LLOG_ORIGIN_HANDLE_CLOSE:
1897                 DEBUG_REQ(D_INODE, req, "llog close");
1898                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1899                 rc = llog_origin_handle_close(req);
1900                 break;
1901         case OST_CREATE:
1902                 DEBUG_REQ(D_INODE, req, "ost_create");
1903                 rc = mdt_obj_create(req);
1904                 break;
1905         case OST_GET_INFO:
1906                 DEBUG_REQ(D_INODE, req, "get_info");
1907                 rc = mdt_get_info(req);
1908                 break;
1909         case OST_SET_INFO:
1910                 DEBUG_REQ(D_INODE, req, "set_info");
1911                 rc = mdt_set_info(req);
1912                 break;
1913         case OST_WRITE:
1914                 CDEBUG(D_INODE, "write\n");
1915                 OBD_FAIL_RETURN(OBD_FAIL_OST_BRW_NET, 0);
1916                 rc = ost_brw_write(req, NULL);
1917                 LASSERT(current->journal_info == NULL);
1918                 /* mdt_brw sends its own replies */
1919                 RETURN(rc);
1920                 break;
1921         case LLOG_CATINFO:
1922                 DEBUG_REQ(D_INODE, req, "llog catinfo");
1923                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1924                 rc = llog_catinfo(req);
1925                 break;
1926         default:
1927                 req->rq_status = -ENOTSUPP;
1928                 rc = ptlrpc_error(req);
1929                 RETURN(rc);
1930         }
1931
1932         LASSERT(current->journal_info == NULL);
1933
1934         EXIT;
1935
1936         /* If we're DISCONNECTing, the mds_export_data is already freed */
1937         if (!rc && req->rq_reqmsg->opc != MDS_DISCONNECT) {
1938                 struct mds_export_data *med = &req->rq_export->exp_mds_data;
1939                 struct obd_device *obd = list_entry(mds, struct obd_device,
1940                                                     u.mds);
1941                 req->rq_repmsg->last_xid =
1942                         le64_to_cpu(med->med_mcd->mcd_last_xid);
1943
1944                 if (!obd->obd_no_transno) {
1945                         req->rq_repmsg->last_committed =
1946                                 obd->obd_last_committed;
1947                 } else {
1948                         DEBUG_REQ(D_IOCTL, req,
1949                                   "not sending last_committed update");
1950                 }
1951                 CDEBUG(D_INFO, "last_transno "LPU64", last_committed "LPU64
1952                        ", xid "LPU64"\n",
1953                        mds->mds_last_transno, obd->obd_last_committed,
1954                        req->rq_xid);
1955         }
1956  out:
1957
1958         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LAST_REPLAY) {
1959                 if (obd && obd->obd_recovering) {
1960                         DEBUG_REQ(D_HA, req, "LAST_REPLAY, queuing reply");
1961                         return target_queue_final_reply(req, rc);
1962                 }
1963                 /* Lost a race with recovery; let the error path DTRT. */
1964                 rc = req->rq_status = -ENOTCONN;
1965         }
1966
1967         target_send_reply(req, rc, fail);
1968         return 0;
1969 }
1970
1971 /* Update the server data on disk.  This stores the new mount_count and
1972  * also the last_rcvd value to disk.  If we don't have a clean shutdown,
1973  * then the server last_rcvd value may be less than that of the clients.
1974  * This will alert us that we may need to do client recovery.
1975  *
1976  * Also assumes for mds_last_transno that we are not modifying it (no locking).
1977  */
1978 int mds_update_server_data(struct obd_device *obd, int force_sync)
1979 {
1980         struct mds_obd *mds = &obd->u.mds;
1981         struct mds_server_data *msd = mds->mds_server_data;
1982         struct file *filp = mds->mds_rcvd_filp;
1983         struct lvfs_run_ctxt saved;
1984         loff_t off = 0;
1985         int rc;
1986         ENTRY;
1987
1988         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1989         msd->msd_last_transno = cpu_to_le64(mds->mds_last_transno);
1990
1991         CDEBUG(D_SUPER, "MDS mount_count is "LPU64", last_transno is "LPU64"\n",
1992                mds->mds_mount_count, mds->mds_last_transno);
1993         rc = fsfilt_write_record(obd, filp, msd, sizeof(*msd), &off,force_sync);
1994         if (rc)
1995                 CERROR("error writing MDS server data: rc = %d\n", rc);
1996         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1997
1998         RETURN(rc);
1999 }
2000
2001 /* mount the file system (secretly) */
2002 static int mds_setup(struct obd_device *obd, obd_count len, void *buf)
2003 {
2004         struct lustre_cfg* lcfg = buf;
2005         struct mds_obd *mds = &obd->u.mds;
2006         char *options = NULL;
2007         struct vfsmount *mnt;
2008         unsigned long page;
2009         int rc = 0;
2010         ENTRY;
2011
2012         dev_clear_rdonly(2);
2013
2014         if (!lcfg->lcfg_inlbuf1 || !lcfg->lcfg_inlbuf2)
2015                 RETURN(rc = -EINVAL);
2016
2017         obd->obd_fsops = fsfilt_get_ops(lcfg->lcfg_inlbuf2);
2018         if (IS_ERR(obd->obd_fsops))
2019                 RETURN(rc = PTR_ERR(obd->obd_fsops));
2020
2021         mds->mds_max_mdsize = sizeof(struct lov_mds_md);
2022
2023         page = __get_free_page(GFP_KERNEL);
2024         if (!page)
2025                 RETURN(-ENOMEM);
2026
2027         options = (char *)page;
2028         memset(options, 0, PAGE_SIZE);
2029
2030         /* here we use "iopen_nopriv" hardcoded, because it affects MDS utility
2031          * and the rest of options are passed by mount options. Probably this
2032          * should be moved to somewhere else like startup scripts or lconf. */
2033         sprintf(options, "iopen_nopriv");
2034
2035         if (lcfg->lcfg_inllen4 > 0 && lcfg->lcfg_inlbuf4)
2036                 sprintf(options + strlen(options), ",%s",
2037                         lcfg->lcfg_inlbuf4);
2038
2039         /* we have to know mdsnum before touching underlying fs -bzzz */
2040         sema_init(&mds->mds_lmv_sem, 1);
2041         mds->mds_lmv_connected = 0;
2042         if (lcfg->lcfg_inllen5 > 0 && lcfg->lcfg_inlbuf5 &&
2043             strcmp(lcfg->lcfg_inlbuf5, "dumb")) {
2044                 class_uuid_t uuid;
2045
2046                 CDEBUG(D_OTHER, "MDS: %s is master for %s\n",
2047                        obd->obd_name, lcfg->lcfg_inlbuf5);
2048
2049                 generate_random_uuid(uuid);
2050                 class_uuid_unparse(uuid, &mds->mds_lmv_uuid);
2051
2052                 OBD_ALLOC(mds->mds_lmv_name, lcfg->lcfg_inllen5);
2053                 if (mds->mds_lmv_name == NULL)
2054                         RETURN(rc = -ENOMEM);
2055
2056                 memcpy(mds->mds_lmv_name, lcfg->lcfg_inlbuf5,
2057                        lcfg->lcfg_inllen5);
2058
2059                 rc = mds_lmv_connect(obd, mds->mds_lmv_name);
2060                 if (rc) {
2061                         OBD_FREE(mds->mds_lmv_name, lcfg->lcfg_inllen5);
2062                         GOTO(err_ops, rc);
2063                 }
2064         }
2065
2066         /* FIXME-WANGDI: this should be reworked when we will use lmv along
2067          * with cobd, because correct mdsnum is set in mds_lmv_connect(). */
2068         if (lcfg->lcfg_inllen6 > 0 && lcfg->lcfg_inlbuf6 && !mds->mds_lmv_obd &&
2069             strcmp(lcfg->lcfg_inlbuf6, "dumb")) {
2070                 if (!memcmp(lcfg->lcfg_inlbuf6, "master", strlen("master")) &&
2071                     mds->mds_num == 0) {
2072                         mds->mds_num = REAL_MDS_NUMBER;
2073                 } else if (!memcmp(lcfg->lcfg_inlbuf6, "cache", strlen("cache")) &&
2074                            mds->mds_num == 0) {
2075                         mds->mds_num = CACHE_MDS_NUMBER;
2076                 }
2077         }
2078
2079         mnt = do_kern_mount(lcfg->lcfg_inlbuf2, 0,
2080                             lcfg->lcfg_inlbuf1, options);
2081
2082         free_page(page);
2083
2084         if (IS_ERR(mnt)) {
2085                 rc = PTR_ERR(mnt);
2086                 CERROR("do_kern_mount failed: rc = %d\n", rc);
2087                 GOTO(err_ops, rc);
2088         }
2089
2090         CDEBUG(D_SUPER, "%s: mnt = %p\n", lcfg->lcfg_inlbuf1, mnt);
2091
2092         sema_init(&mds->mds_orphan_recovery_sem, 1);
2093         sema_init(&mds->mds_epoch_sem, 1);
2094         spin_lock_init(&mds->mds_transno_lock);
2095         mds->mds_max_cookiesize = sizeof(struct llog_cookie);
2096         atomic_set(&mds->mds_real_clients, 0);
2097
2098         obd->obd_namespace = ldlm_namespace_new(obd->obd_name,
2099                                                 LDLM_NAMESPACE_SERVER);
2100         if (obd->obd_namespace == NULL) {
2101                 mds_cleanup(obd, 0);
2102                 GOTO(err_put, rc = -ENOMEM);
2103         }
2104         ldlm_register_intent(obd->obd_namespace, mds_intent_policy);
2105
2106         rc = mds_fs_setup(obd, mnt);
2107         if (rc) {
2108                 CERROR("MDS filesystem method init failed: rc = %d\n", rc);
2109                 GOTO(err_ns, rc);
2110         }
2111
2112         rc = llog_start_commit_thread();
2113         if (rc < 0)
2114                 GOTO(err_fs, rc);
2115
2116         /* this check for @dumb string is needed to handle mounting MDS with
2117          * smfs. Read lconf:MDSDEV.write_conf() for more details. */
2118         if (lcfg->lcfg_inllen3 > 0 && lcfg->lcfg_inlbuf3 &&
2119             strcmp(lcfg->lcfg_inlbuf3, "dumb")) {
2120                 class_uuid_t uuid;
2121
2122                 generate_random_uuid(uuid);
2123                 class_uuid_unparse(uuid, &mds->mds_lov_uuid);
2124
2125                 OBD_ALLOC(mds->mds_profile, lcfg->lcfg_inllen3);
2126                 if (mds->mds_profile == NULL)
2127                         GOTO(err_fs, rc = -ENOMEM);
2128
2129                 memcpy(mds->mds_profile, lcfg->lcfg_inlbuf3,
2130                        lcfg->lcfg_inllen3);
2131         }
2132
2133         ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
2134                            "mds_ldlm_client", &obd->obd_ldlm_client);
2135         obd->obd_replayable = 1;
2136
2137         mds->mds_counters = lprocfs_alloc_mds_counters();
2138
2139         rc = mds_postsetup(obd);
2140         if (rc)
2141                 GOTO(err_fs, rc);
2142
2143         RETURN(0);
2144
2145 err_fs:
2146         /* No extra cleanup needed for llog_init_commit_thread() */
2147         mds_fs_cleanup(obd, 0);
2148 err_ns:
2149         ldlm_namespace_free(obd->obd_namespace, 0);
2150         obd->obd_namespace = NULL;
2151 err_put:
2152         unlock_kernel();
2153         mntput(mds->mds_vfsmnt);
2154         mds->mds_sb = 0;
2155         lock_kernel();
2156 err_ops:
2157         fsfilt_put_ops(obd->obd_fsops);
2158         return rc;
2159 }
2160
2161 static int mds_postsetup(struct obd_device *obd)
2162 {
2163         struct mds_obd *mds = &obd->u.mds;
2164         int rc = 0;
2165         ENTRY;
2166
2167         rc = obd_llog_setup(obd, &obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT,
2168                             obd, 0, NULL, &llog_lvfs_ops);
2169         if (rc)
2170                 RETURN(rc);
2171
2172         if (mds->mds_profile) {
2173                 struct llog_ctxt *lgctxt;
2174                 struct lvfs_run_ctxt saved;
2175                 struct lustre_profile *lprof;
2176                 struct config_llog_instance cfg;
2177
2178                 cfg.cfg_instance = NULL;
2179                 cfg.cfg_uuid = mds->mds_lov_uuid;
2180                 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2181
2182                 lgctxt = llog_get_context(&obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT);
2183                 if (!lgctxt) {
2184                         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2185                         GOTO(err_llog, rc = -EINVAL);
2186                 }
2187
2188                 rc = class_config_process_llog(lgctxt, mds->mds_profile, &cfg);
2189                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2190
2191                 if (rc)
2192                         GOTO(err_llog, rc);
2193
2194                 lprof = class_get_profile(mds->mds_profile);
2195                 if (lprof == NULL) {
2196                         CERROR("No profile found: %s\n", mds->mds_profile);
2197                         GOTO(err_cleanup, rc = -ENOENT);
2198                 }
2199                 rc = mds_lov_connect(obd, lprof->lp_osc);
2200                 if (rc)
2201                         GOTO(err_cleanup, rc);
2202
2203                 rc = mds_lmv_postsetup(obd);
2204                 if (rc)
2205                         GOTO(err_cleanup, rc);
2206         }
2207
2208         RETURN(rc);
2209
2210 err_cleanup:
2211         mds_lov_clean(obd);
2212 err_llog:
2213         obd_llog_cleanup(llog_get_context(&obd->obd_llogs,
2214                                           LLOG_CONFIG_ORIG_CTXT));
2215         RETURN(rc);
2216 }
2217
2218 int mds_postrecov(struct obd_device *obd)
2219 {
2220         struct mds_obd *mds = &obd->u.mds;
2221         struct llog_ctxt *ctxt;
2222         int rc, item = 0;
2223         ENTRY;
2224
2225         LASSERT(!obd->obd_recovering);
2226         ctxt = llog_get_context(&obd->obd_llogs, LLOG_UNLINK_ORIG_CTXT);
2227         LASSERT(ctxt != NULL);
2228
2229         /* set nextid first, so we are sure it happens */
2230         rc = mds_lov_set_nextid(obd);
2231         if (rc) {
2232                 CERROR("%s: mds_lov_set_nextid failed\n", obd->obd_name);
2233                 GOTO(out, rc);
2234         }
2235
2236         /* clean PENDING dir */
2237         rc = mds_cleanup_orphans(obd);
2238         if (rc < 0)
2239                 GOTO(out, rc);
2240         item = rc;
2241
2242         rc = llog_connect(ctxt, obd->u.mds.mds_lov_desc.ld_tgt_count,
2243                           NULL, NULL, NULL);
2244         if (rc) {
2245                 CERROR("%s: failed at llog_origin_connect: %d\n",
2246                        obd->obd_name, rc);
2247                 GOTO(out, rc);
2248         }
2249
2250         /* remove the orphaned precreated objects */
2251         rc = mds_lov_clearorphans(mds, NULL /* all OSTs */);
2252         if (rc)
2253                 GOTO(err_llog, rc);
2254
2255 out:
2256         RETURN(rc < 0 ? rc : item);
2257
2258 err_llog:
2259         /* cleanup all llogging subsystems */
2260         rc = obd_llog_finish(obd, &obd->obd_llogs,
2261                              mds->mds_lov_desc.ld_tgt_count);
2262         if (rc)
2263                 CERROR("%s: failed to cleanup llogging subsystems\n",
2264                         obd->obd_name);
2265         goto out;
2266 }
2267
2268 int mds_lov_clean(struct obd_device *obd)
2269 {
2270         struct mds_obd *mds = &obd->u.mds;
2271
2272         if (mds->mds_profile) {
2273                 char * cln_prof;
2274                 struct config_llog_instance cfg;
2275                 struct lvfs_run_ctxt saved;
2276                 int len = strlen(mds->mds_profile) + sizeof("-clean") + 1;
2277
2278                 OBD_ALLOC(cln_prof, len);
2279                 sprintf(cln_prof, "%s-clean", mds->mds_profile);
2280
2281                 cfg.cfg_instance = NULL;
2282                 cfg.cfg_uuid = mds->mds_lov_uuid;
2283
2284                 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2285                 class_config_process_llog(llog_get_context(&obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT),
2286                                           cln_prof, &cfg);
2287                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2288
2289                 OBD_FREE(cln_prof, len);
2290                 OBD_FREE(mds->mds_profile, strlen(mds->mds_profile) + 1);
2291                 mds->mds_profile = NULL;
2292         }
2293         RETURN(0);
2294 }
2295
2296 int mds_lmv_clean(struct obd_device *obd)
2297 {
2298         struct mds_obd *mds = &obd->u.mds;
2299
2300         if (mds->mds_lmv_name) {
2301                 OBD_FREE(mds->mds_lmv_name, strlen(mds->mds_lmv_name) + 1);
2302                 mds->mds_lmv_name = NULL;
2303         }
2304         RETURN(0);
2305 }
2306
2307 static int mds_precleanup(struct obd_device *obd, int flags)
2308 {
2309         int rc = 0;
2310         ENTRY;
2311
2312         mds_lmv_clean(obd);
2313         mds_lov_disconnect(obd, flags);
2314         mds_lov_clean(obd);
2315         obd_llog_cleanup(llog_get_context(&obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT));
2316         RETURN(rc);
2317 }
2318
2319 static int mds_cleanup(struct obd_device *obd, int flags)
2320 {
2321         struct mds_obd *mds = &obd->u.mds;
2322         ENTRY;
2323
2324         if (mds->mds_sb == NULL)
2325                 RETURN(0);
2326
2327         mds_update_server_data(obd, 1);
2328         if (mds->mds_lov_objids != NULL) {
2329                 OBD_FREE(mds->mds_lov_objids,
2330                          mds->mds_lov_desc.ld_tgt_count * sizeof(obd_id));
2331         }
2332         mds_fs_cleanup(obd, flags);
2333
2334         unlock_kernel();
2335
2336         /* 2 seems normal on mds, (may_umount() also expects 2
2337           fwiw), but we only see 1 at this point in obdfilter. */
2338         if (atomic_read(&obd->u.mds.mds_vfsmnt->mnt_count) > 2)
2339                 CERROR("%s: mount busy, mnt_count %d != 2\n", obd->obd_name,
2340                        atomic_read(&obd->u.mds.mds_vfsmnt->mnt_count));
2341
2342         mntput(mds->mds_vfsmnt);
2343
2344         mds->mds_sb = 0;
2345
2346         ldlm_namespace_free(obd->obd_namespace, flags & OBD_OPT_FORCE);
2347
2348         if (mds->mds_counters) {
2349                 lprocfs_free_mds_counters(mds->mds_counters);
2350         }
2351
2352         spin_lock_bh(&obd->obd_processing_task_lock);
2353         if (obd->obd_recovering) {
2354                 target_cancel_recovery_timer(obd);
2355                 obd->obd_recovering = 0;
2356         }
2357         spin_unlock_bh(&obd->obd_processing_task_lock);
2358
2359         lock_kernel();
2360         dev_clear_rdonly(2);
2361         fsfilt_put_ops(obd->obd_fsops);
2362
2363         RETURN(0);
2364 }
2365
2366 static void fixup_handle_for_resent_req(struct ptlrpc_request *req,
2367                                         struct ldlm_lock *new_lock,
2368                                         struct lustre_handle *lockh)
2369 {
2370         struct obd_export *exp = req->rq_export;
2371         struct obd_device *obd = exp->exp_obd;
2372         struct ldlm_request *dlmreq =
2373                 lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*dlmreq));
2374         struct lustre_handle remote_hdl = dlmreq->lock_handle1;
2375         struct list_head *iter;
2376
2377         if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))
2378                 return;
2379
2380         l_lock(&obd->obd_namespace->ns_lock);
2381         list_for_each(iter, &exp->exp_ldlm_data.led_held_locks) {
2382                 struct ldlm_lock *lock;
2383                 lock = list_entry(iter, struct ldlm_lock, l_export_chain);
2384                 if (lock == new_lock)
2385                         continue;
2386                 if (lock->l_remote_handle.cookie == remote_hdl.cookie) {
2387                         lockh->cookie = lock->l_handle.h_cookie;
2388                         DEBUG_REQ(D_HA, req, "restoring lock cookie "LPX64,
2389                                   lockh->cookie);
2390                         l_unlock(&obd->obd_namespace->ns_lock);
2391                         return;
2392                 }
2393         }
2394         l_unlock(&obd->obd_namespace->ns_lock);
2395
2396         /* If the xid matches, then we know this is a resent request,
2397          * and allow it. (It's probably an OPEN, for which we don't
2398          * send a lock */
2399         if (req->rq_xid == exp->exp_mds_data.med_mcd->mcd_last_xid)
2400                 return;
2401
2402         /* This remote handle isn't enqueued, so we never received or
2403          * processed this request.  Clear MSG_RESENT, because it can
2404          * be handled like any normal request now. */
2405
2406         lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
2407
2408         DEBUG_REQ(D_HA, req, "no existing lock with rhandle "LPX64,
2409                   remote_hdl.cookie);
2410 }
2411
2412 int intent_disposition(struct ldlm_reply *rep, int flag)
2413 {
2414         if (!rep)
2415                 return 0;
2416         return (rep->lock_policy_res1 & flag);
2417 }
2418
2419 void intent_set_disposition(struct ldlm_reply *rep, int flag)
2420 {
2421         if (!rep)
2422                 return;
2423         rep->lock_policy_res1 |= flag;
2424 }
2425
2426 static int mds_intent_policy(struct ldlm_namespace *ns,
2427                              struct ldlm_lock **lockp, void *req_cookie,
2428                              ldlm_mode_t mode, int flags, void *data)
2429 {
2430         struct ptlrpc_request *req = req_cookie;
2431         struct ldlm_lock *lock = *lockp;
2432         struct ldlm_intent *it;
2433         struct mds_obd *mds = &req->rq_export->exp_obd->u.mds;
2434         struct ldlm_reply *rep;
2435         struct lustre_handle lockh = { 0 };
2436         struct ldlm_lock *new_lock;
2437         int getattr_part = MDS_INODELOCK_UPDATE;
2438         int rc, offset = 2, repsize[4] = {sizeof(struct ldlm_reply),
2439                                           sizeof(struct mds_body),
2440                                           mds->mds_max_mdsize,
2441                                           mds->mds_max_cookiesize};
2442         ENTRY;
2443
2444         LASSERT(req != NULL);
2445
2446         if (req->rq_reqmsg->bufcount <= 1) {
2447                 /* No intent was provided */
2448                 int size = sizeof(struct ldlm_reply);
2449                 rc = lustre_pack_reply(req, 1, &size, NULL);
2450                 LASSERT(rc == 0);
2451                 RETURN(0);
2452         }
2453
2454         it = lustre_swab_reqbuf(req, 1, sizeof(*it), lustre_swab_ldlm_intent);
2455         if (it == NULL) {
2456                 CERROR("Intent missing\n");
2457                 RETURN(req->rq_status = -EFAULT);
2458         }
2459
2460         LDLM_DEBUG(lock, "intent policy, opc: %s", ldlm_it2str(it->opc));
2461
2462         rc = lustre_pack_reply(req, 3, repsize, NULL);
2463         if (rc)
2464                 RETURN(req->rq_status = rc);
2465
2466         rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
2467         intent_set_disposition(rep, DISP_IT_EXECD);
2468
2469         fixup_handle_for_resent_req(req, lock, &lockh);
2470
2471         /* execute policy */
2472         switch ((long)it->opc) {
2473         case IT_OPEN:
2474         case IT_CREAT|IT_OPEN:
2475                 /* XXX swab here to assert that an mds_open reint
2476                  * packet is following */
2477                 rep->lock_policy_res2 = mds_reint(req, offset, &lockh);
2478 #if 0
2479                 /* We abort the lock if the lookup was negative and
2480                  * we did not make it to the OPEN portion */
2481                 if (!intent_disposition(rep, DISP_LOOKUP_EXECD))
2482                         RETURN(ELDLM_LOCK_ABORTED);
2483                 if (intent_disposition(rep, DISP_LOOKUP_NEG) &&
2484                     !intent_disposition(rep, DISP_OPEN_OPEN))
2485 #endif
2486                 /* IT_OPEN may return lock on cross-node dentry
2487                  * that we want to hold during attr retrival -bzzz */
2488                 if (rc != 0 || lockh.cookie == 0)
2489                         RETURN(ELDLM_LOCK_ABORTED);
2490                 break;
2491         case IT_LOOKUP:
2492                 getattr_part = MDS_INODELOCK_LOOKUP;
2493         case IT_CHDIR:
2494         case IT_GETATTR:
2495                 getattr_part |= MDS_INODELOCK_LOOKUP;
2496         case IT_READDIR:
2497                 rep->lock_policy_res2 = mds_getattr_name(offset, req, &lockh,
2498                                                          getattr_part);
2499                 /* FIXME: LDLM can set req->rq_status. MDS sets
2500                    policy_res{1,2} with disposition and status.
2501                    - replay: returns 0 & req->status is old status
2502                    - otherwise: returns req->status */
2503                 if (intent_disposition(rep, DISP_LOOKUP_NEG))
2504                         rep->lock_policy_res2 = 0;
2505                 if (!intent_disposition(rep, DISP_LOOKUP_POS) ||
2506                     rep->lock_policy_res2)
2507                         RETURN(ELDLM_LOCK_ABORTED);
2508                 if (req->rq_status != 0) {
2509                         LBUG();
2510                         rep->lock_policy_res2 = req->rq_status;
2511                         RETURN(ELDLM_LOCK_ABORTED);
2512                 }
2513                 break;
2514         case IT_UNLINK:
2515                 rc = mds_lock_and_check_slave(offset, req, &lockh);
2516                 if ((rep->lock_policy_res2 = rc)) {
2517                         if (rc == ENOLCK)
2518                                 rep->lock_policy_res2 = 0;
2519                         RETURN(ELDLM_LOCK_ABORTED);
2520                 }
2521                 break;
2522         default:
2523                 CERROR("Unhandled intent "LPD64"\n", it->opc);
2524                 LBUG();
2525         }
2526
2527         /* By this point, whatever function we called above must have either
2528          * filled in 'lockh', been an intent replay, or returned an error.  We
2529          * want to allow replayed RPCs to not get a lock, since we would just
2530          * drop it below anyways because lock replay is done separately by the
2531          * client afterwards.  For regular RPCs we want to give the new lock to
2532          * the client instead of whatever lock it was about to get. */
2533         new_lock = ldlm_handle2lock(&lockh);
2534         if (new_lock == NULL && (flags & LDLM_FL_INTENT_ONLY))
2535                 RETURN(0);
2536
2537         LASSERT(new_lock != NULL);
2538
2539         /* If we've already given this lock to a client once, then we should
2540          * have no readers or writers.  Otherwise, we should have one reader
2541          * _or_ writer ref (which will be zeroed below) before returning the
2542          * lock to a client. */
2543         if (new_lock->l_export == req->rq_export) {
2544                 LASSERT(new_lock->l_readers + new_lock->l_writers == 0);
2545         } else {
2546                 LASSERT(new_lock->l_export == NULL);
2547                 LASSERT(new_lock->l_readers + new_lock->l_writers == 1);
2548         }
2549
2550         *lockp = new_lock;
2551
2552         if (new_lock->l_export == req->rq_export) {
2553                 /* Already gave this to the client, which means that we
2554                  * reconstructed a reply. */
2555                 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) &
2556                         MSG_RESENT);
2557                 RETURN(ELDLM_LOCK_REPLACED);
2558         }
2559
2560         /* Fixup the lock to be given to the client */
2561         l_lock(&new_lock->l_resource->lr_namespace->ns_lock);
2562         new_lock->l_readers = 0;
2563         new_lock->l_writers = 0;
2564
2565         new_lock->l_export = class_export_get(req->rq_export);
2566         list_add(&new_lock->l_export_chain,
2567                  &new_lock->l_export->exp_ldlm_data.led_held_locks);
2568
2569         new_lock->l_blocking_ast = lock->l_blocking_ast;
2570         new_lock->l_completion_ast = lock->l_completion_ast;
2571
2572         memcpy(&new_lock->l_remote_handle, &lock->l_remote_handle,
2573                sizeof(lock->l_remote_handle));
2574
2575         new_lock->l_flags &= ~LDLM_FL_LOCAL;
2576
2577         LDLM_LOCK_PUT(new_lock);
2578         l_unlock(&new_lock->l_resource->lr_namespace->ns_lock);
2579
2580         RETURN(ELDLM_LOCK_REPLACED);
2581 }
2582
2583 int mds_attach(struct obd_device *dev, obd_count len, void *data)
2584 {
2585         struct lprocfs_static_vars lvars;
2586
2587         lprocfs_init_multi_vars(0, &lvars);
2588         return lprocfs_obd_attach(dev, lvars.obd_vars);
2589 }
2590
2591 int mds_detach(struct obd_device *dev)
2592 {
2593         return lprocfs_obd_detach(dev);
2594 }
2595
2596 int mdt_attach(struct obd_device *dev, obd_count len, void *data)
2597 {
2598         struct lprocfs_static_vars lvars;
2599
2600         lprocfs_init_multi_vars(1, &lvars);
2601         return lprocfs_obd_attach(dev, lvars.obd_vars);
2602 }
2603
2604 int mdt_detach(struct obd_device *dev)
2605 {
2606         return lprocfs_obd_detach(dev);
2607 }
2608
2609 static int mdt_setup(struct obd_device *obd, obd_count len, void *buf)
2610 {
2611         struct mds_obd *mds = &obd->u.mds;
2612         int rc = 0;
2613         ENTRY;
2614
2615         mds->mds_service =
2616                 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
2617                                 MDS_REQUEST_PORTAL, MDC_REPLY_PORTAL,
2618                                 mds_handle, "mds", obd->obd_proc_entry);
2619
2620         if (!mds->mds_service) {
2621                 CERROR("failed to start service\n");
2622                 RETURN(-ENOMEM);
2623         }
2624
2625         rc = ptlrpc_start_n_threads(obd, mds->mds_service, MDT_NUM_THREADS,
2626                                     "ll_mdt");
2627         if (rc)
2628                 GOTO(err_thread, rc);
2629
2630         mds->mds_setattr_service =
2631                 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
2632                                 MDS_SETATTR_PORTAL, MDC_REPLY_PORTAL,
2633                                 mds_handle, "mds_setattr",
2634                                 obd->obd_proc_entry);
2635         if (!mds->mds_setattr_service) {
2636                 CERROR("failed to start getattr service\n");
2637                 GOTO(err_thread, rc = -ENOMEM);
2638         }
2639
2640         rc = ptlrpc_start_n_threads(obd, mds->mds_setattr_service,
2641                                     MDT_NUM_THREADS, "ll_mdt_attr");
2642         if (rc)
2643                 GOTO(err_thread2, rc);
2644
2645         mds->mds_readpage_service =
2646                 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
2647                                 MDS_READPAGE_PORTAL, MDC_REPLY_PORTAL,
2648                                 mds_handle, "mds_readpage",
2649                                 obd->obd_proc_entry);
2650         if (!mds->mds_readpage_service) {
2651                 CERROR("failed to start readpage service\n");
2652                 GOTO(err_thread2, rc = -ENOMEM);
2653         }
2654
2655         rc = ptlrpc_start_n_threads(obd, mds->mds_readpage_service,
2656                                     MDT_NUM_THREADS, "ll_mdt_rdpg");
2657
2658         if (rc)
2659                 GOTO(err_thread3, rc);
2660
2661         RETURN(0);
2662
2663 err_thread3:
2664         ptlrpc_unregister_service(mds->mds_readpage_service);
2665 err_thread2:
2666         ptlrpc_unregister_service(mds->mds_setattr_service);
2667 err_thread:
2668         ptlrpc_unregister_service(mds->mds_service);
2669         return rc;
2670 }
2671
2672 static int mdt_cleanup(struct obd_device *obd, int flags)
2673 {
2674         struct mds_obd *mds = &obd->u.mds;
2675         ENTRY;
2676
2677         ptlrpc_stop_all_threads(mds->mds_readpage_service);
2678         ptlrpc_unregister_service(mds->mds_readpage_service);
2679
2680         ptlrpc_stop_all_threads(mds->mds_setattr_service);
2681         ptlrpc_unregister_service(mds->mds_setattr_service);
2682
2683         ptlrpc_stop_all_threads(mds->mds_service);
2684         ptlrpc_unregister_service(mds->mds_service);
2685
2686         RETURN(0);
2687 }
2688
2689 static struct dentry *mds_lvfs_fid2dentry(__u64 id, __u32 gen, __u64 gr,
2690                                           void *data)
2691 {
2692         struct obd_device *obd = data;
2693         struct ll_fid fid;
2694         fid.id = id;
2695         fid.generation = gen;
2696         return mds_fid2dentry(&obd->u.mds, &fid, NULL);
2697 }
2698
2699 static int mds_get_info(struct obd_export *exp, __u32 keylen,
2700                         void *key, __u32 *vallen, void *val)
2701 {
2702         struct obd_device *obd;
2703         struct mds_obd *mds;
2704         ENTRY;
2705
2706         obd = class_exp2obd(exp);
2707         if (obd == NULL) {
2708                 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
2709                        exp->exp_handle.h_cookie);
2710                 RETURN(-EINVAL);
2711         }
2712
2713         if (keylen >= strlen("reint_log") && memcmp(key, "reint_log", 9) == 0) {
2714                 /*Get log_context handle*/
2715                 unsigned long *llh_handle = val;
2716                 *vallen = sizeof(unsigned long);
2717                 *llh_handle = (unsigned long)obd->obd_llog_ctxt[LLOG_REINT_ORIG_CTXT];
2718                 RETURN(0);
2719         }
2720         if (keylen >= strlen("cache_sb") && memcmp(key, "cache_sb", 8) == 0) {
2721                 /*Get log_context handle*/
2722                 unsigned long *sb = val;
2723                 *vallen = sizeof(unsigned long);
2724                 *sb = (unsigned long)obd->u.mds.mds_sb;
2725                 RETURN(0);
2726         }
2727
2728         mds = &obd->u.mds;
2729         if (keylen >= strlen("mdsize") && memcmp(key, "mdsize", keylen) == 0) {
2730                 __u32 *mdsize = val;
2731                 *vallen = sizeof(*mdsize);
2732                 *mdsize = mds->mds_max_mdsize;
2733                 RETURN(0);
2734         }
2735
2736         CDEBUG(D_IOCTL, "invalid key\n");
2737         RETURN(-EINVAL);
2738
2739 }
2740 struct lvfs_callback_ops mds_lvfs_ops = {
2741         l_fid2dentry:     mds_lvfs_fid2dentry,
2742 };
2743
2744 int mds_preprw(int cmd, struct obd_export *exp, struct obdo *oa,
2745                 int objcount, struct obd_ioobj *obj,
2746                 int niocount, struct niobuf_remote *nb,
2747                 struct niobuf_local *res,
2748                 struct obd_trans_info *oti);
2749 int mds_commitrw(int cmd, struct obd_export *exp, struct obdo *oa,
2750                  int objcount, struct obd_ioobj *obj, int niocount,
2751                  struct niobuf_local *res, struct obd_trans_info *oti,
2752                  int rc);
2753
2754 /* use obd ops to offer management infrastructure */
2755 static struct obd_ops mds_obd_ops = {
2756         .o_owner           = THIS_MODULE,
2757         .o_attach          = mds_attach,
2758         .o_detach          = mds_detach,
2759         .o_connect         = mds_connect,
2760         .o_init_export     = mds_init_export,
2761         .o_destroy_export  = mds_destroy_export,
2762         .o_disconnect      = mds_disconnect,
2763         .o_setup           = mds_setup,
2764         .o_precleanup      = mds_precleanup,
2765         .o_cleanup         = mds_cleanup,
2766         .o_postrecov       = mds_postrecov,
2767         .o_statfs          = mds_obd_statfs,
2768         .o_iocontrol       = mds_iocontrol,
2769         .o_create          = mds_obd_create,
2770         .o_destroy         = mds_obd_destroy,
2771         .o_llog_init       = mds_llog_init,
2772         .o_llog_finish     = mds_llog_finish,
2773         .o_notify          = mds_notify,
2774         .o_get_info        = mds_get_info,
2775         .o_set_info        = mds_set_info,
2776         .o_preprw          = mds_preprw,
2777         .o_commitrw        = mds_commitrw,
2778 };
2779
2780 static struct obd_ops mdt_obd_ops = {
2781         .o_owner           = THIS_MODULE,
2782         .o_attach          = mdt_attach,
2783         .o_detach          = mdt_detach,
2784         .o_setup           = mdt_setup,
2785         .o_cleanup         = mdt_cleanup,
2786         .o_attach          = mdt_attach,
2787         .o_detach          = mdt_detach,
2788 };
2789
2790 static int __init mds_init(void)
2791 {
2792         struct lprocfs_static_vars lvars;
2793
2794         lprocfs_init_multi_vars(0, &lvars);
2795         class_register_type(&mds_obd_ops, NULL, lvars.module_vars,
2796                             LUSTRE_MDS_NAME);
2797         lprocfs_init_multi_vars(1, &lvars);
2798         class_register_type(&mdt_obd_ops, NULL, lvars.module_vars,
2799                             LUSTRE_MDT_NAME);
2800
2801         return 0;
2802 }
2803
2804 static void /*__exit*/ mds_exit(void)
2805 {
2806         class_unregister_type(LUSTRE_MDS_NAME);
2807         class_unregister_type(LUSTRE_MDT_NAME);
2808 }
2809
2810 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
2811 MODULE_DESCRIPTION("Lustre Metadata Server (MDS)");
2812 MODULE_LICENSE("GPL");
2813
2814 module_init(mds_init);
2815 module_exit(mds_exit);