Whamcloud - gitweb
- extN-wantedi accepts generation as well as ino to create an inode with
[fs/lustre-release.git] / lustre / mds / handler.c
1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2  * vim:expandtab:shiftwidth=8:tabstop=8:
3  *
4  *  lustre/mds/handler.c
5  *  Lustre Metadata Server (mds) request handler
6  *
7  *  Copyright (c) 2001-2003 Cluster File Systems, Inc.
8  *   Author: Peter Braam <braam@clusterfs.com>
9  *   Author: Andreas Dilger <adilger@clusterfs.com>
10  *   Author: Phil Schwan <phil@clusterfs.com>
11  *   Author: Mike Shaver <shaver@clusterfs.com>
12  *
13  *   This file is part of Lustre, http://www.lustre.org.
14  *
15  *   Lustre is free software; you can redistribute it and/or
16  *   modify it under the terms of version 2 of the GNU General Public
17  *   License as published by the Free Software Foundation.
18  *
19  *   Lustre is distributed in the hope that it will be useful,
20  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
21  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
22  *   GNU General Public License for more details.
23  *
24  *   You should have received a copy of the GNU General Public License
25  *   along with Lustre; if not, write to the Free Software
26  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
27  */
28
29 #ifndef EXPORT_SYMTAB
30 # define EXPORT_SYMTAB
31 #endif
32 #define DEBUG_SUBSYSTEM S_MDS
33
34 #include <linux/module.h>
35 #include <linux/lustre_mds.h>
36 #include <linux/lustre_dlm.h>
37 #include <linux/init.h>
38 #include <linux/obd_class.h>
39 #include <linux/random.h>
40 #include <linux/fs.h>
41 #include <linux/jbd.h>
42 #include <linux/ext3_fs.h>
43 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0))
44 # include <linux/smp_lock.h>
45 # include <linux/buffer_head.h>
46 # include <linux/workqueue.h>
47 # include <linux/mount.h>
48 #else
49 # include <linux/locks.h>
50 #endif
51 #include <linux/obd_lov.h>
52 #include <linux/lustre_mds.h>
53 #include <linux/lustre_fsfilt.h>
54 #include <linux/lprocfs_status.h>
55 #include <linux/lustre_commit_confd.h>
56
57 #include "mds_internal.h"
58
59 static int mds_intent_policy(struct ldlm_namespace *ns,
60                              struct ldlm_lock **lockp, void *req_cookie,
61                              ldlm_mode_t mode, int flags, void *data);
62 static int mds_postsetup(struct obd_device *obd);
63 static int mds_cleanup(struct obd_device *obd, int flags);
64
65 /* Assumes caller has already pushed into the kernel filesystem context */
66 static int mds_sendpage(struct ptlrpc_request *req, struct file *file,
67                         loff_t offset, int count)
68 {
69         struct ptlrpc_bulk_desc *desc;
70         struct l_wait_info lwi;
71         struct page **pages;
72         int rc = 0, npages, i, tmpcount, tmpsize = 0;
73         ENTRY;
74
75         LASSERT((offset & (PAGE_SIZE - 1)) == 0); /* I'm dubious about this */
76
77         npages = (count + PAGE_SIZE - 1) >> PAGE_SHIFT;
78         OBD_ALLOC(pages, sizeof(*pages) * npages);
79         if (!pages)
80                 GOTO(out, rc = -ENOMEM);
81
82         desc = ptlrpc_prep_bulk_exp(req, npages, BULK_PUT_SOURCE,
83                                     MDS_BULK_PORTAL);
84         if (desc == NULL)
85                 GOTO(out_free, rc = -ENOMEM);
86
87         for (i = 0, tmpcount = count; i < npages; i++, tmpcount -= tmpsize) {
88                 tmpsize = tmpcount > PAGE_SIZE ? PAGE_SIZE : tmpcount;
89
90                 pages[i] = alloc_pages(GFP_KERNEL, 0);
91                 if (pages[i] == NULL)
92                         GOTO(cleanup_buf, rc = -ENOMEM);
93
94                 ptlrpc_prep_bulk_page(desc, pages[i], 0, tmpsize);
95         }
96
97         for (i = 0, tmpcount = count; i < npages; i++, tmpcount -= tmpsize) {
98                 tmpsize = tmpcount > PAGE_SIZE ? PAGE_SIZE : tmpcount;
99                 CDEBUG(D_EXT2, "reading %u@%llu from dir %lu (size %llu)\n",
100                        tmpsize, offset, file->f_dentry->d_inode->i_ino,
101                        file->f_dentry->d_inode->i_size);
102
103                 rc = fsfilt_readpage(req->rq_export->exp_obd, file,
104                                      page_address(pages[i]), tmpsize, &offset);
105
106                 if (rc != tmpsize)
107                         GOTO(cleanup_buf, rc = -EIO);
108         }
109
110         LASSERT(desc->bd_nob == count);
111
112         rc = ptlrpc_start_bulk_transfer(desc);
113         if (rc)
114                 GOTO(cleanup_buf, rc);
115
116         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
117                 CERROR("obd_fail_loc=%x, fail operation rc=%d\n",
118                        OBD_FAIL_MDS_SENDPAGE, rc);
119                 GOTO(abort_bulk, rc);
120         }
121
122         lwi = LWI_TIMEOUT(obd_timeout * HZ / 4, NULL, NULL);
123         rc = l_wait_event(desc->bd_waitq, !ptlrpc_bulk_active(desc), &lwi);
124         LASSERT (rc == 0 || rc == -ETIMEDOUT);
125
126         if (rc == 0) {
127                 if (desc->bd_success &&
128                     desc->bd_nob_transferred == count)
129                         GOTO(cleanup_buf, rc);
130
131                 rc = -ETIMEDOUT; /* XXX should this be a different errno? */
132         }
133         
134         DEBUG_REQ(D_ERROR, req, "bulk failed: %s %d(%d), evicting %s@%s\n",
135                   (rc == -ETIMEDOUT) ? "timeout" : "network error",
136                   desc->bd_nob_transferred, count,
137                   req->rq_export->exp_client_uuid.uuid,
138                   req->rq_export->exp_connection->c_remote_uuid.uuid);
139
140         ptlrpc_fail_export(req->rq_export);
141
142         EXIT;
143  abort_bulk:
144         ptlrpc_abort_bulk (desc);
145  cleanup_buf:
146         for (i = 0; i < npages; i++)
147                 if (pages[i])
148                         __free_pages(pages[i], 0);
149
150         ptlrpc_free_bulk(desc);
151  out_free:
152         OBD_FREE(pages, sizeof(*pages) * npages);
153  out:
154         return rc;
155 }
156
157 /* only valid locked dentries or errors should be returned */
158 struct dentry *mds_fid2locked_dentry(struct obd_device *obd, struct ll_fid *fid,
159                                      struct vfsmount **mnt, int lock_mode,
160                                      struct lustre_handle *lockh,
161                                      char *name, int namelen, __u64 lockpart)
162 {
163         struct mds_obd *mds = &obd->u.mds;
164         struct dentry *de = mds_fid2dentry(mds, fid, mnt), *retval = de;
165         struct ldlm_res_id res_id = { .name = {0} };
166         int flags = 0, rc;
167         ldlm_policy_data_t policy = { .l_inodebits = { lockpart } };
168
169         ENTRY;
170
171         if (IS_ERR(de))
172                 RETURN(de);
173
174         res_id.name[0] = de->d_inode->i_ino;
175         res_id.name[1] = de->d_inode->i_generation;
176 #ifdef S_PDIROPS
177         lockh[1].cookie = 0;
178         if (name && IS_PDIROPS(de->d_inode)) {
179                 ldlm_policy_data_t cpolicy =
180                         { .l_inodebits = { MDS_INODELOCK_UPDATE } };
181                 /* lock just dir { ino, generation } to flush client cache */
182                 if (lock_mode == LCK_PW) {
183                         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
184                                               res_id, LDLM_IBITS,
185                                               &cpolicy, LCK_CW, &flags,
186                                               mds_blocking_ast,
187                                               ldlm_completion_ast, NULL, NULL,
188                                               NULL, 0, NULL, lockh + 1);
189                         if (rc != ELDLM_OK) {
190                                 l_dput(de);
191                                 RETURN(ERR_PTR(-ENOLCK));
192                         }
193                        flags = 0;
194                 }
195
196                 res_id.name[2] = full_name_hash(name, namelen);
197                 CDEBUG(D_INFO, "take lock on %lu:%u:"LPX64"\n",
198                        de->d_inode->i_ino, de->d_inode->i_generation,
199                        res_id.name[2]);
200         }
201 #endif
202         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, res_id,
203                               LDLM_IBITS, &policy, lock_mode, &flags,
204                               mds_blocking_ast, ldlm_completion_ast, NULL, NULL,
205                               NULL, 0, NULL, lockh);
206         if (rc != ELDLM_OK) {
207                 l_dput(de);
208                 retval = ERR_PTR(-EIO); /* XXX translate ldlm code */
209 #ifdef S_PDIROPS
210                 if (lockh[1].cookie)
211                         ldlm_lock_decref(lockh + 1, LCK_CW);
212 #endif
213         }
214
215         RETURN(retval);
216 }
217
218 #ifndef DCACHE_DISCONNECTED
219 #define DCACHE_DISCONNECTED DCACHE_NFSD_DISCONNECTED
220 #endif
221
222
223 /* Look up an entry by inode number. */
224 /* this function ONLY returns valid dget'd dentries with an initialized inode
225    or errors */
226 struct dentry *mds_fid2dentry(struct mds_obd *mds, struct ll_fid *fid,
227                               struct vfsmount **mnt)
228 {
229         char fid_name[32];
230         unsigned long ino = fid->id;
231         __u32 generation = fid->generation;
232         struct inode *inode;
233         struct dentry *result;
234
235         if (ino == 0)
236                 RETURN(ERR_PTR(-ESTALE));
237
238         snprintf(fid_name, sizeof(fid_name), "0x%lx", ino);
239
240         CDEBUG(D_DENTRY, "--> mds_fid2dentry: ino/gen %lu/%u, sb %p\n",
241                ino, generation, mds->mds_sb);
242
243         /* under ext3 this is neither supposed to return bad inodes
244            nor NULL inodes. */
245         result = ll_lookup_one_len(fid_name, mds->mds_fid_de, strlen(fid_name));
246         if (IS_ERR(result))
247                 RETURN(result);
248
249         inode = result->d_inode;
250         if (!inode)
251                 RETURN(ERR_PTR(-ENOENT));
252
253         if (generation && inode->i_generation != generation) {
254                 /* we didn't find the right inode.. */
255                 CERROR("bad inode %lu, link: %lu ct: %d or generation %u/%u\n",
256                        inode->i_ino, (unsigned long)inode->i_nlink,
257                        atomic_read(&inode->i_count), inode->i_generation,
258                        generation);
259                 dput(result);
260                 RETURN(ERR_PTR(-ENOENT));
261         }
262
263         if (mnt) {
264                 *mnt = mds->mds_vfsmnt;
265                 mntget(*mnt);
266         }
267
268         RETURN(result);
269 }
270
271
272 /* Establish a connection to the MDS.
273  *
274  * This will set up an export structure for the client to hold state data
275  * about that client, like open files, the last operation number it did
276  * on the server, etc.
277  */
278 static int mds_connect(struct lustre_handle *conn, struct obd_device *obd,
279                        struct obd_uuid *cluuid)
280 {
281         struct obd_export *exp;
282         struct mds_export_data *med; /*  */
283         struct mds_client_data *mcd;
284         int rc, abort_recovery;
285         ENTRY;
286
287         if (!conn || !obd || !cluuid)
288                 RETURN(-EINVAL);
289
290         /* Check for aborted recovery. */
291         spin_lock_bh(&obd->obd_processing_task_lock);
292         abort_recovery = obd->obd_abort_recovery;
293         spin_unlock_bh(&obd->obd_processing_task_lock);
294         if (abort_recovery)
295                 target_abort_recovery(obd);
296
297         /* XXX There is a small race between checking the list and adding a
298          * new connection for the same UUID, but the real threat (list
299          * corruption when multiple different clients connect) is solved.
300          *
301          * There is a second race between adding the export to the list,
302          * and filling in the client data below.  Hence skipping the case
303          * of NULL mcd above.  We should already be controlling multiple
304          * connects at the client, and we can't hold the spinlock over
305          * memory allocations without risk of deadlocking.
306          */
307         rc = class_connect(conn, obd, cluuid);
308         if (rc)
309                 RETURN(rc);
310         exp = class_conn2export(conn);
311         LASSERT(exp);
312         med = &exp->exp_mds_data;
313
314         OBD_ALLOC(mcd, sizeof(*mcd));
315         if (!mcd) {
316                 CERROR("mds: out of memory for client data\n");
317                 GOTO(out, rc = -ENOMEM);
318         }
319
320         memcpy(mcd->mcd_uuid, cluuid, sizeof(mcd->mcd_uuid));
321         med->med_mcd = mcd;
322
323         rc = mds_client_add(obd, &obd->u.mds, med, -1);
324         if (rc == 0)
325                 EXIT;
326 out:
327         if (rc) {
328                 OBD_FREE(mcd, sizeof(*mcd));
329                 class_disconnect(exp, 0);
330         }
331         class_export_put(exp);
332
333         return rc;
334 }
335
336 static int mds_init_export(struct obd_export *exp) 
337 {
338         struct mds_export_data *med = &exp->exp_mds_data;
339
340         INIT_LIST_HEAD(&med->med_open_head);
341         spin_lock_init(&med->med_open_lock);
342         RETURN(0);
343 }
344
345 static int mds_destroy_export(struct obd_export *export)
346 {
347         struct mds_export_data *med;
348         struct obd_device *obd = export->exp_obd;
349         struct obd_run_ctxt saved;
350         int rc = 0;
351         ENTRY;
352
353         med = &export->exp_mds_data;
354         target_destroy_export(export);
355
356         if (obd_uuid_equals(&export->exp_client_uuid, &obd->obd_uuid))
357                 GOTO(out, 0);
358
359         push_ctxt(&saved, &obd->obd_ctxt, NULL);
360         /* Close any open files (which may also cause orphan unlinking). */
361         spin_lock(&med->med_open_lock);
362         while (!list_empty(&med->med_open_head)) {
363                 struct list_head *tmp = med->med_open_head.next;
364                 struct mds_file_data *mfd =
365                         list_entry(tmp, struct mds_file_data, mfd_list);
366                 BDEVNAME_DECLARE_STORAGE(btmp);
367
368                 /* bug 1579: fix force-closing for 2.5 */
369                 struct dentry *dentry = mfd->mfd_dentry;
370
371                 list_del(&mfd->mfd_list);
372                 spin_unlock(&med->med_open_lock);
373
374                 CERROR("force closing client file handle for %*s (%s:%lu)\n",
375                        dentry->d_name.len, dentry->d_name.name,
376                        ll_bdevname(dentry->d_inode->i_sb, btmp),
377                        dentry->d_inode->i_ino);
378                 rc = mds_mfd_close(NULL, obd, mfd, 
379                                    !(export->exp_flags & OBD_OPT_FAILOVER));
380
381                 if (rc)
382                         CDEBUG(D_INODE, "Error closing file: %d\n", rc);
383                 spin_lock(&med->med_open_lock);
384         }
385         spin_unlock(&med->med_open_lock);
386         pop_ctxt(&saved, &obd->obd_ctxt, NULL);
387
388 out:
389         mds_client_free(export, !(export->exp_flags & OBD_OPT_FAILOVER));
390
391         RETURN(rc);
392 }
393
394 static int mds_disconnect(struct obd_export *exp, int flags)
395 {
396         struct obd_device *obd;
397         struct mds_obd *mds;
398         unsigned long irqflags;
399         int rc;
400         ENTRY;
401
402         LASSERT(exp);
403         class_export_get(exp);
404
405         obd = class_exp2obd(exp);
406         if (obd == NULL) {
407                 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
408                        exp->exp_handle.h_cookie);
409                 RETURN(-EINVAL);
410         }
411         mds = &obd->u.mds;
412
413         if (!(exp->exp_flags & OBD_OPT_REAL_CLIENT)
414                         && !atomic_read(&mds->mds_real_clients)) {
415                 /* there was no client at all */
416                 mds_lmv_disconnect(obd, flags);
417         }
418
419         if ((exp->exp_flags & OBD_OPT_REAL_CLIENT)
420                         && atomic_dec_and_test(&mds->mds_real_clients)) {
421                 /* time to drop LMV connections */
422                 CDEBUG(D_OTHER, "%s: last real client %s disconnected.  "
423                        "Disconnnect from LMV now\n",
424                        obd->obd_name, exp->exp_client_uuid.uuid);
425                 mds_lmv_disconnect(obd, flags);
426         }
427
428         spin_lock_irqsave(&exp->exp_lock, irqflags);
429         exp->exp_flags = flags;
430         spin_unlock_irqrestore(&exp->exp_lock, irqflags);
431
432         /* Disconnect early so that clients can't keep using export */
433         rc = class_disconnect(exp, flags);
434         ldlm_cancel_locks_for_export(exp);
435
436         /* complete all outstanding replies */
437         spin_lock_irqsave(&exp->exp_lock, irqflags);
438         while (!list_empty(&exp->exp_outstanding_replies)) {
439                 struct ptlrpc_reply_state *rs =
440                         list_entry(exp->exp_outstanding_replies.next,
441                                    struct ptlrpc_reply_state, rs_exp_list);
442                 struct ptlrpc_service *svc = rs->rs_srv_ni->sni_service;
443
444                 spin_lock(&svc->srv_lock);
445                 list_del_init(&rs->rs_exp_list);
446                 ptlrpc_schedule_difficult_reply(rs);
447                 spin_unlock(&svc->srv_lock);
448         }
449         spin_unlock_irqrestore(&exp->exp_lock, irqflags);
450
451         class_export_put(exp);
452         RETURN(rc);
453 }
454
455 static int mds_getstatus(struct ptlrpc_request *req)
456 {
457         struct mds_obd *mds = mds_req2mds(req);
458         struct mds_body *body;
459         int rc, size = sizeof(*body);
460         ENTRY;
461
462         rc = lustre_pack_reply(req, 1, &size, NULL);
463         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_GETSTATUS_PACK)) {
464                 CERROR("mds: out of memory for message: size=%d\n", size);
465                 req->rq_status = -ENOMEM;       /* superfluous? */
466                 RETURN(-ENOMEM);
467         }
468
469         body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
470         memcpy(&body->fid1, &mds->mds_rootfid, sizeof(body->fid1));
471
472         /* the last_committed and last_xid fields are filled in for all
473          * replies already - no need to do so here also.
474          */
475         RETURN(0);
476 }
477
478 int mds_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
479                      void *data, int flag)
480 {
481         int do_ast;
482         ENTRY;
483
484         if (flag == LDLM_CB_CANCELING) {
485                 /* Don't need to do anything here. */
486                 RETURN(0);
487         }
488
489         /* XXX layering violation!  -phil */
490         l_lock(&lock->l_resource->lr_namespace->ns_lock);
491         /* Get this: if mds_blocking_ast is racing with mds_intent_policy,
492          * such that mds_blocking_ast is called just before l_i_p takes the
493          * ns_lock, then by the time we get the lock, we might not be the
494          * correct blocking function anymore.  So check, and return early, if
495          * so. */
496         if (lock->l_blocking_ast != mds_blocking_ast) {
497                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
498                 RETURN(0);
499         }
500
501         lock->l_flags |= LDLM_FL_CBPENDING;
502         do_ast = (!lock->l_readers && !lock->l_writers);
503         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
504
505         if (do_ast) {
506                 struct lustre_handle lockh;
507                 int rc;
508
509                 LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel");
510                 ldlm_lock2handle(lock, &lockh);
511                 rc = ldlm_cli_cancel(&lockh);
512                 if (rc < 0)
513                         CERROR("ldlm_cli_cancel: %d\n", rc);
514         } else {
515                 LDLM_DEBUG(lock, "Lock still has references, will be "
516                            "cancelled later");
517         }
518         RETURN(0);
519 }
520
521 /* Call with lock=1 if you want mds_pack_md to take the i_sem.
522  * Call with lock=0 if the caller has already taken the i_sem. */
523 int mds_pack_md(struct obd_device *obd, struct lustre_msg *msg, int offset,
524                 struct mds_body *body, struct inode *inode, int lock)
525 {
526         struct mds_obd *mds = &obd->u.mds;
527         void *lmm;
528         int lmm_size;
529         int rc;
530         ENTRY;
531
532         lmm = lustre_msg_buf(msg, offset, 0);
533         if (lmm == NULL) {
534                 /* Some problem with getting eadata when I sized the reply
535                  * buffer... */
536                 CDEBUG(D_INFO, "no space reserved for inode %lu MD\n",
537                        inode->i_ino);
538                 RETURN(0);
539         }
540         lmm_size = msg->buflens[offset];
541
542         /* I don't really like this, but it is a sanity check on the client
543          * MD request.  However, if the client doesn't know how much space
544          * to reserve for the MD, it shouldn't be bad to have too much space.
545          */
546         if (lmm_size > mds->mds_max_mdsize) {
547                 CWARN("Reading MD for inode %lu of %d bytes > max %d\n",
548                        inode->i_ino, lmm_size, mds->mds_max_mdsize);
549                 // RETURN(-EINVAL);
550         }
551
552         if (lock)
553                 down(&inode->i_sem);
554         rc = fsfilt_get_md(obd, inode, lmm, lmm_size);
555         if (lock)
556                 up(&inode->i_sem);
557         if (rc < 0) {
558                 CERROR("Error %d reading eadata for ino %lu\n",
559                        rc, inode->i_ino);
560         } else if (rc > 0) {
561                 lmm_size = rc;
562                 if (S_ISREG(inode->i_mode))
563                         rc = mds_convert_lov_ea(obd, inode, lmm, lmm_size);
564                 if (rc > 0)
565                         lmm_size = rc;
566                 body->valid |= OBD_MD_FLEASIZE;
567                 body->eadatasize = lmm_size;
568                 rc = 0;
569         }
570
571         RETURN(rc);
572 }
573
574 static int mds_getattr_internal(struct obd_device *obd, struct dentry *dentry,
575                                 struct ptlrpc_request *req,
576                                 struct mds_body *reqbody, int reply_off)
577 {
578         struct mds_body *body;
579         struct inode *inode = dentry->d_inode;
580         int rc = 0;
581         ENTRY;
582
583         if (inode == NULL && !(dentry->d_flags & DCACHE_CROSS_REF))
584                 RETURN(-ENOENT);
585
586         body = lustre_msg_buf(req->rq_repmsg, reply_off, sizeof(*body));
587         LASSERT(body != NULL);                 /* caller prepped reply */
588
589         if (dentry->d_flags & DCACHE_CROSS_REF) {
590                 CDEBUG(D_OTHER, "cross reference: %lu/%lu/%lu\n",
591                        (unsigned long) dentry->d_mdsnum,
592                        (unsigned long) dentry->d_inum,
593                        (unsigned long) dentry->d_generation);
594                 body->valid |= OBD_MD_FLID | OBD_MD_MDS;
595                 body->fid1.id = dentry->d_inum;
596                 body->fid1.mds = dentry->d_mdsnum;
597                 body->fid1.generation = dentry->d_generation;
598                 RETURN(0);
599         }
600         mds_pack_inode2fid(obd, &body->fid1, inode);
601         mds_pack_inode2body(obd, body, inode);
602
603         if (S_ISREG(inode->i_mode) && (reqbody->valid & OBD_MD_FLEASIZE) != 0) {
604                 rc = mds_pack_md(obd, req->rq_repmsg, reply_off + 1, body,
605                                  inode, 1);
606
607                 /* If we have LOV EA data, the OST holds size, atime, mtime */
608                 if (!(body->valid & OBD_MD_FLEASIZE))
609                         body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
610                                         OBD_MD_FLATIME | OBD_MD_FLMTIME);
611         } else if (S_ISLNK(inode->i_mode) &&
612                    (reqbody->valid & OBD_MD_LINKNAME) != 0) {
613                 char *symname = lustre_msg_buf(req->rq_repmsg, reply_off + 1,0);
614                 int len;
615
616                 LASSERT (symname != NULL);       /* caller prepped reply */
617                 len = req->rq_repmsg->buflens[reply_off + 1];
618
619                 rc = inode->i_op->readlink(dentry, symname, len);
620                 if (rc < 0) {
621                         CERROR("readlink failed: %d\n", rc);
622                 } else if (rc != len - 1) {
623                         CERROR ("Unexpected readlink rc %d: expecting %d\n",
624                                 rc, len - 1);
625                         rc = -EINVAL;
626                 } else {
627                         CDEBUG(D_INODE, "read symlink dest %s\n", symname);
628                         body->valid |= OBD_MD_LINKNAME;
629                         body->eadatasize = rc + 1;
630                         symname[rc] = 0;        /* NULL terminate */
631                         rc = 0;
632                 }
633         } else if (S_ISDIR(inode->i_mode) &&
634                         (reqbody->valid & OBD_MD_FLEASIZE) != 0) {
635                 rc = mds_pack_md(obd, req->rq_repmsg, reply_off + 1, body,
636                                 inode, 1);
637         } 
638
639         RETURN(rc);
640 }
641
642 static int mds_getattr_pack_msg_cf(struct ptlrpc_request *req,
643                                         struct dentry *dentry,
644                                         int offset)
645 {
646         int rc = 0, size[1] = {sizeof(struct mds_body)};
647         ENTRY;
648
649         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK)) {
650                 CERROR("failed MDS_GETATTR_PACK test\n");
651                 req->rq_status = -ENOMEM;
652                 GOTO(out, rc = -ENOMEM);
653         }
654
655         rc = lustre_pack_reply(req, 1, size, NULL);
656         if (rc) {
657                 CERROR("out of memory\n");
658                 GOTO(out, req->rq_status = rc);
659         }
660
661         EXIT;
662  out:
663         return(rc);
664 }
665
666 static int mds_getattr_pack_msg(struct ptlrpc_request *req, struct inode *inode,
667                                 int offset)
668 {
669         struct mds_obd *mds = mds_req2mds(req);
670         struct mds_body *body;
671         int rc = 0, size[2] = {sizeof(*body)}, bufcount = 1;
672         ENTRY;
673
674         body = lustre_msg_buf(req->rq_reqmsg, offset, sizeof (*body));
675         LASSERT(body != NULL);                 /* checked by caller */
676         LASSERT_REQSWABBED(req, offset);       /* swabbed by caller */
677
678         if ((S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode))
679                         && (body->valid & OBD_MD_FLEASIZE)) {
680                 int rc;
681                 down(&inode->i_sem);
682                 rc = fsfilt_get_md(req->rq_export->exp_obd, inode, NULL, 0);
683                 up(&inode->i_sem);
684                 CDEBUG(D_INODE, "got %d bytes MD data for inode %lu\n",
685                        rc, inode->i_ino);
686                 if (rc < 0) {
687                         if (rc != -ENODATA)
688                                 CERROR("error getting inode %lu MD: rc = %d\n",
689                                        inode->i_ino, rc);
690                         size[bufcount] = 0;
691                 } else if (rc > mds->mds_max_mdsize) {
692                         size[bufcount] = 0;
693                         CERROR("MD size %d larger than maximum possible %u\n",
694                                rc, mds->mds_max_mdsize);
695                 } else {
696                         size[bufcount] = rc;
697                 }
698                 bufcount++;
699         } else if (S_ISLNK(inode->i_mode) && (body->valid & OBD_MD_LINKNAME)) {
700                 if (inode->i_size + 1 != body->eadatasize)
701                         CERROR("symlink size: %Lu, reply space: %d\n",
702                                inode->i_size + 1, body->eadatasize);
703                 size[bufcount] = min_t(int, inode->i_size+1, body->eadatasize);
704                 bufcount++;
705                 CDEBUG(D_INODE, "symlink size: %Lu, reply space: %d\n",
706                        inode->i_size + 1, body->eadatasize);
707         }
708
709         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK)) {
710                 CERROR("failed MDS_GETATTR_PACK test\n");
711                 req->rq_status = -ENOMEM;
712                 GOTO(out, rc = -ENOMEM);
713         }
714
715         rc = lustre_pack_reply(req, bufcount, size, NULL);
716         if (rc) {
717                 CERROR("out of memory\n");
718                 GOTO(out, req->rq_status = rc);
719         }
720
721         EXIT;
722  out:
723         return(rc);
724 }
725
726 static int mds_getattr_name(int offset, struct ptlrpc_request *req,
727                             struct lustre_handle *child_lockh, int child_part)
728 {
729         struct obd_device *obd = req->rq_export->exp_obd;
730         struct ldlm_reply *rep = NULL;
731         struct obd_run_ctxt saved;
732         struct mds_body *body;
733         struct dentry *dparent = NULL, *dchild = NULL;
734         struct obd_ucred uc;
735         struct lustre_handle parent_lockh[2];
736         int namesize;
737         int rc = 0, cleanup_phase = 0, resent_req = 0;
738         char *name;
739         ENTRY;
740
741         LASSERT(!strcmp(obd->obd_type->typ_name, "mds"));
742
743         /* Swab now, before anyone looks inside the request */
744
745         body = lustre_swab_reqbuf(req, offset, sizeof(*body),
746                                   lustre_swab_mds_body);
747         if (body == NULL) {
748                 CERROR("Can't swab mds_body\n");
749                 GOTO(cleanup, rc = -EFAULT);
750         }
751
752         LASSERT_REQSWAB(req, offset + 1);
753         name = lustre_msg_string(req->rq_reqmsg, offset + 1, 0);
754         if (name == NULL) {
755                 CERROR("Can't unpack name\n");
756                 GOTO(cleanup, rc = -EFAULT);
757         }
758         namesize = req->rq_reqmsg->buflens[offset + 1];
759
760         LASSERT (offset == 0 || offset == 2);
761         /* if requests were at offset 2, the getattr reply goes back at 1 */
762         if (offset) { 
763                 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
764                 offset = 1;
765         }
766
767         uc.ouc_fsuid = body->fsuid;
768         uc.ouc_fsgid = body->fsgid;
769         uc.ouc_cap = body->capability;
770         uc.ouc_suppgid1 = body->suppgid;
771         uc.ouc_suppgid2 = -1;
772         push_ctxt(&saved, &obd->obd_ctxt, &uc);
773         cleanup_phase = 1; /* kernel context */
774         intent_set_disposition(rep, DISP_LOOKUP_EXECD);
775
776         LASSERT(namesize > 0);
777         if (namesize == 1) {
778                 /* we have no dentry here, drop LOOKUP bit */
779                 child_part &= ~MDS_INODELOCK_LOOKUP;
780                 CDEBUG(D_OTHER, "%s: request to retrieve attrs for %lu/%lu\n",
781                        obd->obd_name, (unsigned long) body->fid1.id,
782                        (unsigned long) body->fid1.generation);
783                 dchild = mds_fid2locked_dentry(obd, &body->fid1, NULL, LCK_PR,
784                                                parent_lockh, NULL, 0, child_part);
785                 if (IS_ERR(dchild)) {
786                         CERROR("can't find inode: %d\n", (int) PTR_ERR(dchild));
787                         GOTO(cleanup, rc = PTR_ERR(dchild));
788                 }
789                 memcpy(child_lockh, parent_lockh, sizeof(parent_lockh[0]));
790 #ifdef S_PDIROPS
791                 if (parent_lockh[1].cookie)
792                         ldlm_lock_decref(parent_lockh + 1, LCK_CW);
793 #endif
794                 cleanup_phase = 2;
795                 goto fill_inode;
796         }
797         
798         /* FIXME: handle raw lookup */
799 #if 0
800         if (body->valid == OBD_MD_FLID) {
801                 struct mds_body *mds_reply;
802                 int size = sizeof(*mds_reply);
803                 ino_t inum;
804                 // The user requested ONLY the inode number, so do a raw lookup
805                 rc = lustre_pack_reply(req, 1, &size, NULL);
806                 if (rc) {
807                         CERROR("out of memory\n");
808                         GOTO(cleanup, rc);
809                 }
810
811                 rc = dir->i_op->lookup_raw(dir, name, namesize - 1, &inum);
812
813                 mds_reply = lustre_msg_buf(req->rq_repmsg, offset,
814                                            sizeof(*mds_reply));
815                 mds_reply->fid1.id = inum;
816                 mds_reply->valid = OBD_MD_FLID;
817                 GOTO(cleanup, rc);
818         }
819 #endif
820
821         if (child_lockh->cookie != 0) {
822                 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT);
823                 resent_req = 1;
824         }
825
826         if (resent_req == 0) {
827                 rc = mds_get_parent_child_locked(obd, &obd->u.mds, &body->fid1,
828                                                  parent_lockh, &dparent,
829                                                  LCK_PR, MDS_INODELOCK_LOOKUP,
830                                                  name, namesize,
831                                                  child_lockh, &dchild, LCK_PR,
832                                                  child_part);
833                 if (rc)
834                         GOTO(cleanup, rc);
835         } else {
836                 struct ldlm_lock *granted_lock;
837                 struct ll_fid child_fid;
838                 struct ldlm_resource *res;
839                 DEBUG_REQ(D_DLMTRACE, req, "resent, not enqueuing new locks");
840                 granted_lock = ldlm_handle2lock(child_lockh);
841                 LASSERT(granted_lock);
842
843                 res = granted_lock->l_resource;
844                 child_fid.id = res->lr_name.name[0];
845                 child_fid.generation = res->lr_name.name[1];
846                 dchild = mds_fid2dentry(&obd->u.mds, &child_fid, NULL);
847                 LASSERT(dchild);
848                 LDLM_LOCK_PUT(granted_lock);
849         }
850
851         cleanup_phase = 2; /* dchild, dparent, locks */
852
853 fill_inode:
854         if (!DENTRY_VALID(dchild)) {
855                 intent_set_disposition(rep, DISP_LOOKUP_NEG);
856                 /* in the intent case, the policy clears this error:
857                    the disposition is enough */
858                 GOTO(cleanup, rc = -ENOENT);
859         } else {
860                 intent_set_disposition(rep, DISP_LOOKUP_POS);
861         }
862
863         if (req->rq_repmsg == NULL) {
864                 if (dchild->d_flags & DCACHE_CROSS_REF)
865                         rc = mds_getattr_pack_msg_cf(req, dchild, offset);
866                 else
867                         rc = mds_getattr_pack_msg(req, dchild->d_inode, offset);
868                 if (rc != 0) {
869                         CERROR ("mds_getattr_pack_msg: %d\n", rc);
870                         GOTO (cleanup, rc);
871                 }
872         }
873
874         rc = mds_getattr_internal(obd, dchild, req, body, offset);
875         GOTO(cleanup, rc); /* returns the lock to the client */
876
877  cleanup:
878         switch (cleanup_phase) {
879         case 2:
880                 if (resent_req == 0) {
881                         if (rc && DENTRY_VALID(dchild))
882                                 ldlm_lock_decref(child_lockh, LCK_PR);
883                         if (dparent) {
884                                 ldlm_lock_decref(parent_lockh, LCK_PR);
885 #ifdef S_PDIROPS
886                                 if (parent_lockh[1].cookie != 0)
887                                         ldlm_lock_decref(parent_lockh + 1,
888                                                         LCK_CW);
889 #endif
890                         }
891                         if (dparent)
892                                 l_dput(dparent);
893                 }
894                 l_dput(dchild);
895         case 1:
896                 pop_ctxt(&saved, &obd->obd_ctxt, &uc);
897         default: ;
898         }
899         return rc;
900 }
901
902 static int mds_getattr(int offset, struct ptlrpc_request *req)
903 {
904         struct mds_obd *mds = mds_req2mds(req);
905         struct obd_device *obd = req->rq_export->exp_obd;
906         struct obd_run_ctxt saved;
907         struct dentry *de;
908         struct mds_body *body;
909         struct obd_ucred uc;
910         int rc = 0;
911         ENTRY;
912
913         body = lustre_swab_reqbuf (req, offset, sizeof (*body),
914                                    lustre_swab_mds_body);
915         if (body == NULL) {
916                 CERROR ("Can't unpack body\n");
917                 RETURN (-EFAULT);
918         }
919
920         uc.ouc_fsuid = body->fsuid;
921         uc.ouc_fsgid = body->fsgid;
922         uc.ouc_cap = body->capability;
923         push_ctxt(&saved, &obd->obd_ctxt, &uc);
924         de = mds_fid2dentry(mds, &body->fid1, NULL);
925         if (IS_ERR(de)) {
926                 rc = req->rq_status = -ENOENT;
927                 GOTO(out_pop, PTR_ERR(de));
928         }
929
930         rc = mds_getattr_pack_msg(req, de->d_inode, offset);
931         if (rc != 0) {
932                 CERROR ("mds_getattr_pack_msg: %d\n", rc);
933                 GOTO (out_pop, rc);
934         }
935
936         req->rq_status = mds_getattr_internal(obd, de, req, body, 0);
937
938         l_dput(de);
939         GOTO(out_pop, rc);
940 out_pop:
941         pop_ctxt(&saved, &obd->obd_ctxt, &uc);
942         return rc;
943 }
944
945
946 static int mds_obd_statfs(struct obd_device *obd, struct obd_statfs *osfs,
947                           unsigned long max_age)
948 {
949         int rc;
950
951         spin_lock(&obd->obd_osfs_lock);
952         rc = fsfilt_statfs(obd, obd->u.mds.mds_sb, max_age);
953         if (rc == 0)
954                 memcpy(osfs, &obd->obd_osfs, sizeof(*osfs));
955         spin_unlock(&obd->obd_osfs_lock);
956
957         return rc;
958 }
959
960 static int mds_statfs(struct ptlrpc_request *req)
961 {
962         struct obd_device *obd = req->rq_export->exp_obd;
963         int rc, size = sizeof(struct obd_statfs);
964         ENTRY;
965
966         rc = lustre_pack_reply(req, 1, &size, NULL);
967         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_STATFS_PACK)) {
968                 CERROR("mds: statfs lustre_pack_reply failed: rc = %d\n", rc);
969                 GOTO(out, rc);
970         }
971
972         /* We call this so that we can cache a bit - 1 jiffie worth */
973         rc = mds_obd_statfs(obd, lustre_msg_buf(req->rq_repmsg, 0, size),
974                             jiffies - HZ);
975         if (rc) {
976                 CERROR("mds_obd_statfs failed: rc %d\n", rc);
977                 GOTO(out, rc);
978         }
979
980         EXIT;
981 out:
982         req->rq_status = rc;
983         return 0;
984 }
985
986 static int mds_sync(struct ptlrpc_request *req)
987 {
988         struct obd_device *obd = req->rq_export->exp_obd;
989         struct mds_obd *mds = &obd->u.mds;
990         struct mds_body *body;
991         int rc, size = sizeof(*body);
992         ENTRY;
993
994         body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof(*body));
995         if (body == NULL)
996                 GOTO(out, rc = -EPROTO);
997
998         rc = lustre_pack_reply(req, 1, &size, NULL);
999         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_SYNC_PACK)) {
1000                 CERROR("fsync lustre_pack_reply failed: rc = %d\n", rc);
1001                 GOTO(out, rc);
1002         }
1003
1004         if (body->fid1.id == 0) {
1005                 /* a fid of zero is taken to mean "sync whole filesystem" */
1006                 rc = fsfilt_sync(obd, mds->mds_sb);
1007                 if (rc)
1008                         GOTO(out, rc);
1009         } else {
1010                 /* just any file to grab fsync method - "file" arg unused */
1011                 struct file *file = mds->mds_rcvd_filp;
1012                 struct dentry *de;
1013
1014                 de = mds_fid2dentry(mds, &body->fid1, NULL);
1015                 if (IS_ERR(de))
1016                         GOTO(out, rc = PTR_ERR(de));
1017
1018                 rc = file->f_op->fsync(NULL, de, 1);
1019                 l_dput(de);
1020                 if (rc)
1021                         GOTO(out, rc);
1022
1023                 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*body));
1024                 mds_pack_inode2fid(obd, &body->fid1, de->d_inode);
1025                 mds_pack_inode2body(obd, body, de->d_inode);
1026         }
1027 out:
1028         req->rq_status = rc;
1029         return 0;
1030 }
1031
1032 /* mds_readpage does not take a DLM lock on the inode, because the client must
1033  * already have a PR lock.
1034  *
1035  * If we were to take another one here, a deadlock will result, if another
1036  * thread is already waiting for a PW lock. */
1037 static int mds_readpage(struct ptlrpc_request *req)
1038 {
1039         struct obd_device *obd = req->rq_export->exp_obd;
1040         struct vfsmount *mnt;
1041         struct dentry *de;
1042         struct file *file;
1043         struct mds_body *body, *repbody;
1044         struct obd_run_ctxt saved;
1045         int rc, size = sizeof(*repbody);
1046         struct obd_ucred uc;
1047         ENTRY;
1048
1049         rc = lustre_pack_reply(req, 1, &size, NULL);
1050         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_READPAGE_PACK)) {
1051                 CERROR("mds: out of memory\n");
1052                 GOTO(out, rc = -ENOMEM);
1053         }
1054
1055         body = lustre_swab_reqbuf(req, 0, sizeof(*body), lustre_swab_mds_body);
1056         if (body == NULL)
1057                 GOTO (out, rc = -EFAULT);
1058
1059         uc.ouc_fsuid = body->fsuid;
1060         uc.ouc_fsgid = body->fsgid;
1061         uc.ouc_cap = body->capability;
1062         push_ctxt(&saved, &obd->obd_ctxt, &uc);
1063         de = mds_fid2dentry(&obd->u.mds, &body->fid1, &mnt);
1064         if (IS_ERR(de))
1065                 GOTO(out_pop, rc = PTR_ERR(de));
1066
1067         CDEBUG(D_INODE, "ino %lu\n", de->d_inode->i_ino);
1068
1069         file = dentry_open(de, mnt, O_RDONLY | O_LARGEFILE);
1070         /* note: in case of an error, dentry_open puts dentry */
1071         if (IS_ERR(file))
1072                 GOTO(out_pop, rc = PTR_ERR(file));
1073
1074         /* body->size is actually the offset -eeb */
1075         if ((body->size & (de->d_inode->i_blksize - 1)) != 0) {
1076                 CERROR("offset "LPU64" not on a block boundary of %lu\n",
1077                        body->size, de->d_inode->i_blksize);
1078                 GOTO(out_file, rc = -EFAULT);
1079         }
1080
1081         /* body->nlink is actually the #bytes to read -eeb */
1082         if (body->nlink & (de->d_inode->i_blksize - 1)) {
1083                 CERROR("size %u is not multiple of blocksize %lu\n",
1084                        body->nlink, de->d_inode->i_blksize);
1085                 GOTO(out_file, rc = -EFAULT);
1086         }
1087
1088         repbody = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*repbody));
1089         repbody->size = file->f_dentry->d_inode->i_size;
1090         repbody->valid = OBD_MD_FLSIZE;
1091
1092         /* to make this asynchronous make sure that the handling function
1093            doesn't send a reply when this function completes. Instead a
1094            callback function would send the reply */
1095         /* body->size is actually the offset -eeb */
1096         rc = mds_sendpage(req, file, body->size, body->nlink);
1097
1098 out_file:
1099         filp_close(file, 0);
1100 out_pop:
1101         pop_ctxt(&saved, &obd->obd_ctxt, &uc);
1102 out:
1103         req->rq_status = rc;
1104         RETURN(0);
1105 }
1106
1107 int mds_reint(struct ptlrpc_request *req, int offset,
1108               struct lustre_handle *lockh)
1109 {
1110         struct mds_update_record *rec; /* 116 bytes on the stack?  no sir! */
1111         int rc;
1112         ENTRY;
1113
1114         OBD_ALLOC(rec, sizeof(*rec));
1115         if (rec == NULL)
1116                 RETURN(-ENOMEM);
1117
1118         rc = mds_update_unpack(req, offset, rec);
1119         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNPACK)) {
1120                 CERROR("invalid record\n");
1121                 GOTO(out, req->rq_status = -EINVAL);
1122         }
1123         /* rc will be used to interrupt a for loop over multiple records */
1124         rc = mds_reint_rec(rec, offset, req, lockh);
1125  out:
1126         OBD_FREE(rec, sizeof(*rec));
1127         RETURN(rc);
1128 }
1129
1130 static int mds_filter_recovery_request(struct ptlrpc_request *req,
1131                                        struct obd_device *obd, int *process)
1132 {
1133         switch (req->rq_reqmsg->opc) {
1134         case MDS_CONNECT: /* This will never get here, but for completeness. */
1135         case OST_CONNECT: /* This will never get here, but for completeness. */
1136         case MDS_DISCONNECT:
1137         case OST_DISCONNECT:
1138                *process = 1;
1139                RETURN(0);
1140
1141         case MDS_CLOSE:
1142         case MDS_SYNC: /* used in unmounting */
1143         case OBD_PING:
1144         case MDS_REINT:
1145         case LDLM_ENQUEUE:
1146         case OST_CREATE:
1147                 *process = target_queue_recovery_request(req, obd);
1148                 RETURN(0);
1149
1150         default:
1151                 DEBUG_REQ(D_ERROR, req, "not permitted during recovery");
1152                 *process = 0;
1153                 /* XXX what should we set rq_status to here? */
1154                 req->rq_status = -EAGAIN;
1155                 RETURN(ptlrpc_error(req));
1156         }
1157 }
1158
1159 static char *reint_names[] = {
1160         [REINT_SETATTR] "setattr",
1161         [REINT_CREATE]  "create",
1162         [REINT_LINK]    "link",
1163         [REINT_UNLINK]  "unlink",
1164         [REINT_RENAME]  "rename",
1165         [REINT_OPEN]    "open",
1166 };
1167
1168 #define FILTER_VALID_FLAGS (OBD_MD_FLTYPE | OBD_MD_FLMODE | OBD_MD_FLGENER  |\
1169                             OBD_MD_FLSIZE | OBD_MD_FLBLOCKS | OBD_MD_FLBLKSZ|\
1170                             OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME|\
1171                             OBD_MD_FLID) 
1172
1173 static void reconstruct_create(struct ptlrpc_request *req)
1174 {
1175         struct mds_export_data *med = &req->rq_export->exp_mds_data;
1176         struct mds_client_data *mcd = med->med_mcd;
1177         struct ost_body *body;
1178
1179         body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*body));
1180
1181         /* copy rc, transno and disp; steal locks */
1182         mds_req_from_mcd(req, mcd);
1183         CERROR("reconstruct reply for x"LPU64"\n", req->rq_xid);
1184 }
1185
1186 static int mdt_obj_create(struct ptlrpc_request *req)
1187 {
1188         struct obd_device *obd = req->rq_export->exp_obd;
1189         struct ldlm_res_id res_id = { .name = {0} };
1190         struct mds_obd *mds = &obd->u.mds;
1191         struct ost_body *body, *repbody;
1192         int rc, size = sizeof(*repbody);
1193         char fidname[LL_FID_NAMELEN];
1194         struct inode *parent_inode;
1195         struct lustre_handle lockh;
1196         struct obd_run_ctxt saved;
1197         ldlm_policy_data_t policy;
1198         struct dentry_params dp;
1199         int mealen, flags = 0;
1200         unsigned int tmpname;
1201         struct obd_ucred uc;
1202         struct dentry *new;
1203         struct mea *mea;
1204         void *handle;
1205         ENTRY;
1206        
1207         DEBUG_REQ(D_HA, req, "create remote object");
1208
1209         parent_inode = mds->mds_objects_dir->d_inode;
1210
1211         body = lustre_swab_reqbuf(req, 0, sizeof(*body),
1212                                   lustre_swab_ost_body);
1213         if (body == NULL)
1214                 RETURN(-EFAULT);
1215
1216         MDS_CHECK_RESENT(req, reconstruct_create(req));
1217
1218         uc.ouc_fsuid = body->oa.o_uid;
1219         uc.ouc_fsgid = body->oa.o_gid;
1220
1221         push_ctxt(&saved, &obd->obd_ctxt, &uc);
1222         
1223         rc = lustre_pack_reply(req, 1, &size, NULL);
1224         if (rc)
1225                 RETURN(rc);
1226
1227         repbody = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*repbody));
1228
1229
1230         down(&parent_inode->i_sem);
1231         handle = fsfilt_start(obd, parent_inode, FSFILT_OP_MKDIR, NULL);
1232         LASSERT(!IS_ERR(handle));
1233
1234 repeat:
1235         tmpname = ll_insecure_random_int();
1236         rc = sprintf(fidname, "%u", tmpname);
1237         new = lookup_one_len(fidname, mds->mds_objects_dir, rc);
1238         if (IS_ERR(new)) {
1239                 CERROR("%s: can't lookup new inode (%s) for mkdir: %d\n",
1240                        obd->obd_name, fidname, (int) PTR_ERR(new));
1241                 fsfilt_commit(obd, new->d_inode, handle, 0);
1242                 up(&parent_inode->i_sem);
1243                 RETURN(PTR_ERR(new));
1244         } else if (new->d_inode) {
1245                 CERROR("%s: name exists. repeat\n", obd->obd_name);
1246                 goto repeat;
1247         }
1248
1249         new->d_fsdata = (void *) &dp;
1250         dp.p_inum = 0;
1251         dp.p_ptr = req;
1252
1253         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1254                 DEBUG_REQ(D_HA, req, "replay create obj %lu/%lu",
1255                           (unsigned long) body->oa.o_id,
1256                           (unsigned long) body->oa.o_generation);
1257                 dp.p_inum = body->oa.o_id;
1258                 dp.p_generation = body->oa.o_generation;
1259         }
1260         rc = vfs_mkdir(parent_inode, new, body->oa.o_mode);
1261         if (rc == 0) {
1262                 obdo_from_inode(&repbody->oa, new->d_inode, FILTER_VALID_FLAGS);
1263                 repbody->oa.o_id = new->d_inode->i_ino;
1264                 repbody->oa.o_generation = new->d_inode->i_generation;
1265                 repbody->oa.o_valid |= OBD_MD_FLID | OBD_MD_FLGENER;
1266
1267                 rc = fsfilt_del_dir_entry(obd, new);
1268                 up(&parent_inode->i_sem);
1269
1270                 if (rc) {
1271                         CERROR("can't remove name for object: %d\n", rc);
1272                         GOTO(cleanup, rc);
1273                 }
1274                         
1275                 /* this lock should be taken to serialize MDS modifications
1276                  * in failure case */
1277                 res_id.name[0] = new->d_inode->i_ino;
1278                 res_id.name[1] = new->d_inode->i_generation;
1279                 policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
1280                 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
1281                                 res_id, LDLM_IBITS, &policy,
1282                                 LCK_EX, &flags, mds_blocking_ast,
1283                                 ldlm_completion_ast, NULL, NULL,
1284                                 NULL, 0, NULL, &lockh);
1285                 if (rc != ELDLM_OK)
1286                         GOTO(cleanup, rc);
1287
1288                 CDEBUG(D_OTHER, "created dirobj: %lu/%lu mode %o\n",
1289                                 (unsigned long) new->d_inode->i_ino,
1290                                 (unsigned long) new->d_inode->i_generation,
1291                                 (unsigned) new->d_inode->i_mode);
1292         } else {
1293                 up(&parent_inode->i_sem);
1294                 CERROR("%s: can't create dirobj: %d\n", obd->obd_name, rc);
1295         }
1296
1297         if (rc == 0 && body->oa.o_valid & OBD_MD_FLID) {
1298                 /* this is new object for splitted dir. we have to
1299                  * prevent recursive splitting on it -bzzz */
1300                 mealen = obd_size_diskmd(mds->mds_lmv_exp, NULL);
1301                 OBD_ALLOC(mea, mealen);
1302                 if (mea == NULL)
1303                         GOTO(cleanup, rc = -ENOMEM);
1304                 mea->mea_count = 0;
1305                 down(&new->d_inode->i_sem);
1306                 rc = fsfilt_set_md(obd, new->d_inode, handle, mea, mealen);
1307                 up(&new->d_inode->i_sem);
1308                 OBD_FREE(mea, mealen);
1309         }
1310
1311 cleanup:
1312         rc = mds_finish_transno(mds, parent_inode, handle, req, rc, 0);
1313         if (rc == 0)
1314                 ptlrpc_save_lock(req, &lockh, LCK_EX);
1315         else
1316                 ldlm_lock_decref(&lockh, LCK_EX);
1317         l_dput(new);
1318         pop_ctxt(&saved, &obd->obd_ctxt, &uc);
1319         RETURN(rc);
1320 }
1321
1322 static int mds_get_info(struct obd_export *exp, __u32 keylen,
1323                            void *key, __u32 *vallen, void *val)
1324 {
1325         struct obd_device *obd;
1326         struct mds_obd *mds;
1327         ENTRY;
1328
1329         obd = class_exp2obd(exp);
1330         if (obd == NULL) {
1331                 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
1332                        exp->exp_handle.h_cookie);
1333                 RETURN(-EINVAL);
1334         }
1335
1336         mds = &obd->u.mds;
1337         keylen == strlen("mdsize");
1338         if (keylen && memcmp(key, "mdsize", keylen) == 0) {
1339                 __u32 *mdsize = val;
1340                 *vallen = sizeof(*mdsize);
1341                 *mdsize = mds->mds_max_mdsize;
1342                 RETURN(0);
1343         }
1344
1345         CDEBUG(D_IOCTL, "invalid key\n");
1346         RETURN(-EINVAL);
1347 }
1348
1349 static int mdt_get_info(struct ptlrpc_request *req)
1350 {
1351         char *key;
1352         struct obd_export *exp = req->rq_export;
1353         int keylen, rc = 0, size = sizeof(obd_id);
1354         obd_id *reply;
1355         ENTRY;
1356
1357         key = lustre_msg_buf(req->rq_reqmsg, 0, 1);
1358         if (key == NULL) {
1359                 DEBUG_REQ(D_HA, req, "no get_info key");
1360                 RETURN(-EFAULT);
1361         }
1362         keylen = req->rq_reqmsg->buflens[0];
1363
1364         if (keylen < strlen("mdsize") || memcmp(key, "mdsize", 6) != 0)
1365                 RETURN(-EPROTO);
1366
1367         rc = lustre_pack_reply(req, 1, &size, NULL);
1368         if (rc)
1369                 RETURN(rc);
1370
1371         reply = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*reply));
1372         rc = obd_get_info(exp, keylen, key, &size, reply);
1373         req->rq_repmsg->status = 0;
1374         RETURN(rc);
1375 }
1376
1377 static int mds_set_info(struct obd_export *exp, __u32 keylen,
1378                            void *key, __u32 vallen, void *val)
1379 {
1380         struct obd_device *obd;
1381         struct mds_obd *mds;
1382         int rc;
1383         ENTRY;
1384
1385         obd = class_exp2obd(exp);
1386         if (obd == NULL) {
1387                 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
1388                        exp->exp_handle.h_cookie);
1389                 RETURN(-EINVAL);
1390         }
1391
1392         mds = &obd->u.mds;
1393         keylen == strlen("client");
1394         if (keylen && memcmp(key, "client", keylen) == 0) {
1395                 if (!(exp->exp_flags & OBD_OPT_REAL_CLIENT)) {
1396                         atomic_inc(&mds->mds_real_clients);
1397                         CDEBUG(D_OTHER, "%s: peer from %s is real client (%d)\n",
1398                                         obd->obd_name,
1399                                         exp->exp_client_uuid.uuid,
1400                                         atomic_read(&mds->mds_real_clients));
1401                         exp->exp_flags |= OBD_OPT_REAL_CLIENT;
1402                 }
1403                 rc = mds_lmv_connect(obd, mds->mds_lmv_name);
1404                 LASSERT(rc == 0);
1405                 RETURN(0);
1406         }
1407
1408         CDEBUG(D_IOCTL, "invalid key\n");
1409         RETURN(-EINVAL);
1410 }
1411
1412 static int mdt_set_info(struct ptlrpc_request *req)
1413 {
1414         char *key;
1415         struct obd_export *exp = req->rq_export;
1416         int keylen, rc = 0, size = sizeof(obd_id);
1417         ENTRY;
1418
1419         key = lustre_msg_buf(req->rq_reqmsg, 0, 1);
1420         if (key == NULL) {
1421                 DEBUG_REQ(D_HA, req, "no get_info key");
1422                 RETURN(-EFAULT);
1423         }
1424         keylen = req->rq_reqmsg->buflens[0];
1425
1426         if ((keylen < strlen("client") || memcmp(key, "client", 6) != 0))
1427                 RETURN(-EPROTO);
1428
1429         rc = lustre_pack_reply(req, 0, NULL, NULL);
1430         if (rc)
1431                 RETURN(rc);
1432         rc = obd_set_info(exp, keylen, key, size, NULL);
1433         req->rq_repmsg->status = 0;
1434         RETURN(rc);
1435 }
1436
1437 extern int ost_brw_write(struct ptlrpc_request *, struct obd_trans_info *);
1438
1439 int mds_handle(struct ptlrpc_request *req)
1440 {
1441         int should_process, fail = OBD_FAIL_MDS_ALL_REPLY_NET;
1442         int rc = 0;
1443         struct mds_obd *mds = NULL; /* quell gcc overwarning */
1444         struct obd_device *obd = NULL;
1445         ENTRY;
1446
1447         OBD_FAIL_RETURN(OBD_FAIL_MDS_ALL_REQUEST_NET | OBD_FAIL_ONCE, 0);
1448
1449         LASSERT(current->journal_info == NULL);
1450         /* XXX identical to OST */
1451         if (req->rq_reqmsg->opc != MDS_CONNECT) {
1452                 struct mds_export_data *med;
1453                 int recovering, abort_recovery;
1454
1455                 if (req->rq_export == NULL) {
1456                         CERROR("lustre_mds: operation %d on unconnected MDS\n",
1457                                req->rq_reqmsg->opc);
1458                         req->rq_status = -ENOTCONN;
1459                         GOTO(out, rc = -ENOTCONN);
1460                 }
1461
1462                 med = &req->rq_export->exp_mds_data;
1463                 obd = req->rq_export->exp_obd;
1464                 mds = &obd->u.mds;
1465
1466                 /* Check for aborted recovery. */
1467                 spin_lock_bh(&obd->obd_processing_task_lock);
1468                 abort_recovery = obd->obd_abort_recovery;
1469                 recovering = obd->obd_recovering;
1470                 spin_unlock_bh(&obd->obd_processing_task_lock);
1471                 if (abort_recovery) {
1472                         target_abort_recovery(obd);
1473                 } else if (recovering) {
1474                         rc = mds_filter_recovery_request(req, obd,
1475                                                          &should_process);
1476                         if (rc || !should_process)
1477                                 RETURN(rc);
1478                 }
1479         }
1480
1481         switch (req->rq_reqmsg->opc) {
1482         case MDS_CONNECT:
1483                 DEBUG_REQ(D_INODE, req, "connect");
1484                 OBD_FAIL_RETURN(OBD_FAIL_MDS_CONNECT_NET, 0);
1485                 rc = target_handle_connect(req, mds_handle);
1486                 if (!rc)
1487                         /* Now that we have an export, set mds. */
1488                         mds = mds_req2mds(req);
1489                 break;
1490
1491         case MDS_DISCONNECT:
1492                 DEBUG_REQ(D_INODE, req, "disconnect");
1493                 OBD_FAIL_RETURN(OBD_FAIL_MDS_DISCONNECT_NET, 0);
1494                 rc = target_handle_disconnect(req);
1495                 req->rq_status = rc;            /* superfluous? */
1496                 break;
1497
1498         case MDS_GETSTATUS:
1499                 DEBUG_REQ(D_INODE, req, "getstatus");
1500                 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETSTATUS_NET, 0);
1501                 rc = mds_getstatus(req);
1502                 break;
1503
1504         case MDS_GETATTR:
1505                 DEBUG_REQ(D_INODE, req, "getattr");
1506                 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_NET, 0);
1507                 rc = mds_getattr(0, req);
1508                 break;
1509
1510         case MDS_GETATTR_NAME: {
1511                 struct lustre_handle lockh;
1512                 DEBUG_REQ(D_INODE, req, "getattr_name");
1513                 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_NAME_NET, 0);
1514
1515                 /* If this request gets a reconstructed reply, we won't be
1516                  * acquiring any new locks in mds_getattr_name, so we don't
1517                  * want to cancel.
1518                  */
1519                 lockh.cookie = 0;
1520                 rc = mds_getattr_name(0, req, &lockh, MDS_INODELOCK_UPDATE);
1521                 /* this non-intent call (from an ioctl) is special */
1522                 req->rq_status = rc;
1523                 if (rc == 0 && lockh.cookie)
1524                         ldlm_lock_decref(&lockh, LCK_PR);
1525                 break;
1526         }
1527         case MDS_STATFS:
1528                 DEBUG_REQ(D_INODE, req, "statfs");
1529                 OBD_FAIL_RETURN(OBD_FAIL_MDS_STATFS_NET, 0);
1530                 rc = mds_statfs(req);
1531                 break;
1532
1533         case MDS_READPAGE:
1534                 DEBUG_REQ(D_INODE, req, "readpage");
1535                 OBD_FAIL_RETURN(OBD_FAIL_MDS_READPAGE_NET, 0);
1536                 rc = mds_readpage(req);
1537
1538                 if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_MDS_SENDPAGE)) {
1539                         if (req->rq_reply_state) {
1540                                 lustre_free_reply_state (req->rq_reply_state);
1541                                 req->rq_reply_state = NULL;
1542                         }
1543                         RETURN(0);
1544                 }
1545
1546                 break;
1547
1548         case MDS_REINT: {
1549                 __u32 *opcp = lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*opcp));
1550                 __u32  opc;
1551                 int size[3] = {sizeof(struct mds_body), mds->mds_max_mdsize,
1552                                mds->mds_max_cookiesize};
1553                 int bufcount;
1554
1555                 /* NB only peek inside req now; mds_reint() will swab it */
1556                 if (opcp == NULL) {
1557                         CERROR ("Can't inspect opcode\n");
1558                         rc = -EINVAL;
1559                         break;
1560                 }
1561                 opc = *opcp;
1562                 if (lustre_msg_swabbed (req->rq_reqmsg))
1563                         __swab32s(&opc);
1564
1565                 DEBUG_REQ(D_INODE, req, "reint %d (%s)", opc,
1566                           (opc < sizeof(reint_names) / sizeof(reint_names[0]) ||
1567                            reint_names[opc] == NULL) ? reint_names[opc] :
1568                                                        "unknown opcode");
1569
1570                 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET, 0);
1571
1572                 if (opc == REINT_UNLINK)
1573                         bufcount = 3;
1574                 else if (opc == REINT_OPEN || opc == REINT_RENAME)
1575                         bufcount = 2;
1576                 else
1577                         bufcount = 1;
1578
1579                 rc = lustre_pack_reply(req, bufcount, size, NULL);
1580                 if (rc)
1581                         break;
1582
1583                 rc = mds_reint(req, 0, NULL);
1584                 fail = OBD_FAIL_MDS_REINT_NET_REP;
1585                 break;
1586         }
1587
1588         case MDS_CLOSE:
1589                 DEBUG_REQ(D_INODE, req, "close");
1590                 OBD_FAIL_RETURN(OBD_FAIL_MDS_CLOSE_NET, 0);
1591                 rc = mds_close(req);
1592                 break;
1593
1594         case MDS_DONE_WRITING:
1595                 DEBUG_REQ(D_INODE, req, "done_writing");
1596                 OBD_FAIL_RETURN(OBD_FAIL_MDS_DONE_WRITING_NET, 0);
1597                 rc = mds_done_writing(req);
1598                 break;
1599
1600         case MDS_PIN:
1601                 DEBUG_REQ(D_INODE, req, "pin");
1602                 OBD_FAIL_RETURN(OBD_FAIL_MDS_PIN_NET, 0);
1603                 rc = mds_pin(req);
1604                 break;
1605
1606         case MDS_SYNC:
1607                 DEBUG_REQ(D_INODE, req, "sync");
1608                 OBD_FAIL_RETURN(OBD_FAIL_MDS_SYNC_NET, 0);
1609                 rc = mds_sync(req);
1610                 break;
1611
1612         case OBD_PING:
1613                 DEBUG_REQ(D_INODE, req, "ping");
1614                 rc = target_handle_ping(req);
1615                 break;
1616
1617         case OBD_LOG_CANCEL:
1618                 CDEBUG(D_INODE, "log cancel\n");
1619                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOG_CANCEL_NET, 0);
1620                 rc = -ENOTSUPP; /* la la la */
1621                 break;
1622
1623         case LDLM_ENQUEUE:
1624                 DEBUG_REQ(D_INODE, req, "enqueue");
1625                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_ENQUEUE, 0);
1626                 rc = ldlm_handle_enqueue(req, ldlm_server_completion_ast,
1627                                          ldlm_server_blocking_ast, NULL);
1628                 break;
1629         case LDLM_CONVERT:
1630                 DEBUG_REQ(D_INODE, req, "convert");
1631                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CONVERT, 0);
1632                 rc = ldlm_handle_convert(req);
1633                 break;
1634         case LDLM_BL_CALLBACK:
1635         case LDLM_CP_CALLBACK:
1636                 DEBUG_REQ(D_INODE, req, "callback");
1637                 CERROR("callbacks should not happen on MDS\n");
1638                 LBUG();
1639                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_BL_CALLBACK, 0);
1640                 break;
1641         case LLOG_ORIGIN_HANDLE_CREATE:
1642                 DEBUG_REQ(D_INODE, req, "llog_init");
1643                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1644                 rc = llog_origin_handle_create(req);
1645                 break;
1646         case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
1647                 DEBUG_REQ(D_INODE, req, "llog next block");
1648                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1649                 rc = llog_origin_handle_next_block(req);
1650                 break;
1651         case LLOG_ORIGIN_HANDLE_READ_HEADER:
1652                 DEBUG_REQ(D_INODE, req, "llog read header");
1653                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1654                 rc = llog_origin_handle_read_header(req);
1655                 break;
1656         case LLOG_ORIGIN_HANDLE_CLOSE:
1657                 DEBUG_REQ(D_INODE, req, "llog close");
1658                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1659                 rc = llog_origin_handle_close(req);
1660                 break;
1661         case OST_CREATE:
1662                 DEBUG_REQ(D_INODE, req, "ost_create");
1663                 rc = mdt_obj_create(req);
1664                 break;
1665         case OST_GET_INFO:
1666                 DEBUG_REQ(D_INODE, req, "get_info");
1667                 rc = mdt_get_info(req);
1668                 break;
1669         case OST_SET_INFO:
1670                 DEBUG_REQ(D_INODE, req, "set_info");
1671                 rc = mdt_set_info(req);
1672                 break;
1673         case OST_WRITE:
1674                 CDEBUG(D_INODE, "write\n");
1675                 OBD_FAIL_RETURN(OBD_FAIL_OST_BRW_NET, 0);
1676                 rc = ost_brw_write(req, NULL);
1677                 LASSERT(current->journal_info == NULL);
1678                 /* mdt_brw sends its own replies */
1679                 RETURN(rc);
1680                 break;
1681         case LLOG_CATINFO:
1682                 DEBUG_REQ(D_INODE, req, "llog catinfo");
1683                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1684                 rc = llog_catinfo(req);
1685                 break;
1686         default:
1687                 req->rq_status = -ENOTSUPP;
1688                 rc = ptlrpc_error(req);
1689                 RETURN(rc);
1690         }
1691
1692         LASSERT(current->journal_info == NULL);
1693
1694         EXIT;
1695
1696         /* If we're DISCONNECTing, the mds_export_data is already freed */
1697         if (!rc && req->rq_reqmsg->opc != MDS_DISCONNECT) {
1698                 struct mds_export_data *med = &req->rq_export->exp_mds_data;
1699                 struct obd_device *obd = list_entry(mds, struct obd_device,
1700                                                     u.mds);
1701                 req->rq_repmsg->last_xid =
1702                         le64_to_cpu(med->med_mcd->mcd_last_xid);
1703
1704                 if (!obd->obd_no_transno) {
1705                         req->rq_repmsg->last_committed =
1706                                 obd->obd_last_committed;
1707                 } else {
1708                         DEBUG_REQ(D_IOCTL, req,
1709                                   "not sending last_committed update");
1710                 }
1711                 CDEBUG(D_INFO, "last_transno "LPU64", last_committed "LPU64
1712                        ", xid "LPU64"\n",
1713                        mds->mds_last_transno, obd->obd_last_committed,
1714                        req->rq_xid);
1715         }
1716  out:
1717
1718         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LAST_REPLAY) {
1719                 if (obd && obd->obd_recovering) {
1720                         DEBUG_REQ(D_HA, req, "LAST_REPLAY, queuing reply");
1721                         return target_queue_final_reply(req, rc);
1722                 }
1723                 /* Lost a race with recovery; let the error path DTRT. */
1724                 rc = req->rq_status = -ENOTCONN;
1725         }
1726
1727         target_send_reply(req, rc, fail);
1728         return 0;
1729 }
1730
1731 /* Update the server data on disk.  This stores the new mount_count and
1732  * also the last_rcvd value to disk.  If we don't have a clean shutdown,
1733  * then the server last_rcvd value may be less than that of the clients.
1734  * This will alert us that we may need to do client recovery.
1735  *
1736  * Also assumes for mds_last_transno that we are not modifying it (no locking).
1737  */
1738 int mds_update_server_data(struct obd_device *obd, int force_sync)
1739 {
1740         struct mds_obd *mds = &obd->u.mds;
1741         struct mds_server_data *msd = mds->mds_server_data;
1742         struct file *filp = mds->mds_rcvd_filp;
1743         struct obd_run_ctxt saved;
1744         loff_t off = 0;
1745         int rc;
1746         ENTRY;
1747
1748         push_ctxt(&saved, &obd->obd_ctxt, NULL);
1749         msd->msd_last_transno = cpu_to_le64(mds->mds_last_transno);
1750
1751         CDEBUG(D_SUPER, "MDS mount_count is "LPU64", last_transno is "LPU64"\n",
1752                mds->mds_mount_count, mds->mds_last_transno);
1753         rc = fsfilt_write_record(obd, filp, msd, sizeof(*msd), &off,force_sync);
1754         if (rc)
1755                 CERROR("error writing MDS server data: rc = %d\n", rc);
1756         pop_ctxt(&saved, &obd->obd_ctxt, NULL);
1757
1758         RETURN(rc);
1759 }
1760
1761
1762 /* mount the file system (secretly) */
1763 static int mds_setup(struct obd_device *obd, obd_count len, void *buf)
1764 {
1765         struct lustre_cfg* lcfg = buf;
1766         struct mds_obd *mds = &obd->u.mds;
1767         struct vfsmount *mnt;
1768         int rc = 0;
1769         unsigned long page;
1770         ENTRY;
1771
1772         dev_clear_rdonly(2);
1773
1774         if (!lcfg->lcfg_inlbuf1 || !lcfg->lcfg_inlbuf2)
1775                 RETURN(rc = -EINVAL);
1776
1777         obd->obd_fsops = fsfilt_get_ops(lcfg->lcfg_inlbuf2);
1778         if (IS_ERR(obd->obd_fsops))
1779                 RETURN(rc = PTR_ERR(obd->obd_fsops));
1780
1781         mds->mds_max_mdsize = sizeof(struct lov_mds_md);
1782
1783         /* we have to know mdsnum before touching underlying fs -bzzz */
1784         if (lcfg->lcfg_inllen4 > 0 && lcfg->lcfg_inlbuf4) {
1785                 class_uuid_t uuid;
1786
1787                 CDEBUG(D_OTHER, "MDS: %s is master for %s\n",
1788                        obd->obd_name, lcfg->lcfg_inlbuf4);
1789
1790                 generate_random_uuid(uuid);
1791                 class_uuid_unparse(uuid, &mds->mds_lmv_uuid);
1792
1793                 OBD_ALLOC(mds->mds_lmv_name, lcfg->lcfg_inllen4);
1794                 if (mds->mds_lmv_name == NULL) 
1795                         RETURN(rc = -ENOMEM);
1796
1797                 memcpy(mds->mds_lmv_name, lcfg->lcfg_inlbuf4,
1798                                 lcfg->lcfg_inllen4);
1799                 rc = mds_lmv_connect(obd, mds->mds_lmv_name);
1800                 if (rc) {
1801                         OBD_FREE(mds->mds_lmv_name, lcfg->lcfg_inllen4);
1802                         GOTO(err_ops, rc);
1803                 }
1804         }
1805
1806         if (!(page = __get_free_page(GFP_KERNEL)))
1807                 RETURN(-ENOMEM);
1808
1809         memset((void *)page, 0, PAGE_SIZE);
1810         sprintf((char *)page, "iopen_nopriv");
1811
1812         mnt = do_kern_mount(lcfg->lcfg_inlbuf2, 0,
1813                             lcfg->lcfg_inlbuf1, (void *)page);
1814         free_page(page);
1815         if (IS_ERR(mnt)) {
1816                 rc = PTR_ERR(mnt);
1817                 CERROR("do_kern_mount failed: rc = %d\n", rc);
1818                 GOTO(err_ops, rc);
1819         }
1820
1821         CDEBUG(D_SUPER, "%s: mnt = %p\n", lcfg->lcfg_inlbuf1, mnt);
1822
1823         sema_init(&mds->mds_orphan_recovery_sem, 1);
1824         sema_init(&mds->mds_epoch_sem, 1);
1825         spin_lock_init(&mds->mds_transno_lock);
1826         mds->mds_max_cookiesize = sizeof(struct llog_cookie);
1827         atomic_set(&mds->mds_open_count, 0);
1828         atomic_set(&mds->mds_real_clients, 0);
1829
1830         obd->obd_namespace = ldlm_namespace_new(obd->obd_name,
1831                                                 LDLM_NAMESPACE_SERVER);
1832         if (obd->obd_namespace == NULL) {
1833                 mds_cleanup(obd, 0);
1834                 GOTO(err_put, rc = -ENOMEM);
1835         }
1836         ldlm_register_intent(obd->obd_namespace, mds_intent_policy);
1837
1838         rc = mds_fs_setup(obd, mnt);
1839         if (rc) {
1840                 CERROR("MDS filesystem method init failed: rc = %d\n", rc);
1841                 GOTO(err_ns, rc);
1842         }
1843
1844         rc = llog_start_commit_thread();
1845         if (rc < 0)
1846                 GOTO(err_fs, rc);
1847         
1848
1849         if (lcfg->lcfg_inllen3 > 0 && lcfg->lcfg_inlbuf3) {
1850                 class_uuid_t uuid;
1851
1852                 generate_random_uuid(uuid);
1853                 class_uuid_unparse(uuid, &mds->mds_lov_uuid);
1854
1855                 OBD_ALLOC(mds->mds_profile, lcfg->lcfg_inllen3);
1856                 if (mds->mds_profile == NULL) 
1857                         GOTO(err_fs, rc = -ENOMEM);
1858
1859                 memcpy(mds->mds_profile, lcfg->lcfg_inlbuf3,
1860                        lcfg->lcfg_inllen3);
1861         }
1862
1863         ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
1864                            "mds_ldlm_client", &obd->obd_ldlm_client);
1865         obd->obd_replayable = 1;
1866
1867         rc = mds_postsetup(obd);
1868         if (rc)
1869                 GOTO(err_fs, rc);
1870         RETURN(0);
1871
1872 err_fs:
1873         /* No extra cleanup needed for llog_init_commit_thread() */
1874         mds_fs_cleanup(obd, 0);
1875 err_ns:
1876         ldlm_namespace_free(obd->obd_namespace, 0);
1877         obd->obd_namespace = NULL;
1878 err_put:
1879         unlock_kernel();
1880         mntput(mds->mds_vfsmnt);
1881         mds->mds_sb = 0;
1882         lock_kernel();
1883 err_ops:
1884         fsfilt_put_ops(obd->obd_fsops);
1885         return rc;
1886 }
1887
1888 static int mds_postsetup(struct obd_device *obd)
1889 {
1890         struct mds_obd *mds = &obd->u.mds;
1891         struct llog_ctxt *ctxt;
1892         int rc = 0;
1893         ENTRY;
1894
1895         rc = llog_setup(obd, &obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT,
1896                         obd, 0, NULL, &llog_lvfs_ops);
1897         if (rc)
1898                 RETURN(rc);
1899
1900         if (mds->mds_profile) {
1901                 struct obd_run_ctxt saved;
1902                 struct lustre_profile *lprof;
1903                 struct config_llog_instance cfg;
1904
1905                 cfg.cfg_instance = NULL;
1906                 cfg.cfg_uuid = mds->mds_lov_uuid;
1907                 push_ctxt(&saved, &obd->obd_ctxt, NULL);
1908                 ctxt = llog_get_context(&obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT); 
1909                 rc = class_config_parse_llog(ctxt, mds->mds_profile, &cfg);
1910                 pop_ctxt(&saved, &obd->obd_ctxt, NULL);
1911                 if (rc)
1912                         GOTO(err_llog, rc);
1913
1914                 lprof = class_get_profile(mds->mds_profile);
1915                 if (lprof == NULL) {
1916                         CERROR("No profile found: %s\n", mds->mds_profile);
1917                         GOTO(err_cleanup, rc = -ENOENT);
1918                 }
1919                 rc = mds_lov_connect(obd, lprof->lp_osc);
1920                 if (rc)
1921                         GOTO(err_cleanup, rc);
1922
1923                 rc = mds_lmv_postsetup(obd);
1924                 if (rc)
1925                         GOTO(err_cleanup, rc);
1926         }
1927
1928         RETURN(rc);
1929
1930 err_cleanup:
1931         mds_lov_clean(obd);
1932 err_llog:
1933         ctxt = llog_get_context(&obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT);
1934         llog_cleanup(ctxt);
1935         RETURN(rc);
1936 }
1937
1938 static int mds_postrecov(struct obd_device *obd) 
1939
1940 {
1941         struct llog_ctxt *ctxt;
1942         int rc, rc2;
1943         ENTRY;
1944
1945         LASSERT(!obd->obd_recovering);
1946         ctxt = llog_get_context(&obd->obd_llogs, LLOG_UNLINK_ORIG_CTXT);
1947         LASSERT(ctxt != NULL);
1948
1949         rc = llog_connect(ctxt, obd->u.mds.mds_lov_desc.ld_tgt_count,
1950                           NULL, NULL, NULL);
1951         if (rc != 0) {
1952                 CERROR("faild at llog_origin_connect: %d\n", rc);
1953         }
1954
1955         rc = mds_cleanup_orphans(obd);
1956
1957         rc2 = mds_lov_set_nextid(obd);
1958         if (rc2 == 0)
1959                 rc2 = rc;
1960         RETURN(rc2);
1961 }
1962
1963 int mds_lov_clean(struct obd_device *obd)
1964 {
1965         struct mds_obd *mds = &obd->u.mds;
1966         struct llog_ctxt *ctxt;
1967
1968         if (mds->mds_profile) {
1969                 char * cln_prof;
1970                 struct config_llog_instance cfg;
1971                 struct obd_run_ctxt saved;
1972                 int len = strlen(mds->mds_profile) + sizeof("-clean") + 1;
1973
1974                 OBD_ALLOC(cln_prof, len);
1975                 sprintf(cln_prof, "%s-clean", mds->mds_profile);
1976
1977                 cfg.cfg_instance = NULL;
1978                 cfg.cfg_uuid = mds->mds_lov_uuid;
1979
1980                 push_ctxt(&saved, &obd->obd_ctxt, NULL);
1981                 ctxt = llog_get_context(&obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT);
1982                 class_config_parse_llog(ctxt, cln_prof, &cfg);
1983                 pop_ctxt(&saved, &obd->obd_ctxt, NULL);
1984
1985                 OBD_FREE(cln_prof, len);
1986                 OBD_FREE(mds->mds_profile, strlen(mds->mds_profile) + 1);
1987                 mds->mds_profile = NULL;
1988         }
1989         RETURN(0);
1990 }
1991
1992 int mds_lmv_clean(struct obd_device *obd)
1993 {
1994         struct mds_obd *mds = &obd->u.mds;
1995
1996         if (mds->mds_lmv_name) {
1997                 OBD_FREE(mds->mds_lmv_name, strlen(mds->mds_lmv_name) + 1);
1998                 mds->mds_lmv_name = NULL;
1999         }
2000         RETURN(0);
2001 }
2002
2003 static int mds_precleanup(struct obd_device *obd, int flags)
2004 {
2005         int rc = 0;
2006         ENTRY;
2007
2008         mds_lmv_clean(obd);
2009         mds_lov_disconnect(obd, flags);
2010         mds_lov_clean(obd);
2011         llog_cleanup(llog_get_context(&obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT));
2012         RETURN(rc);
2013 }
2014
2015 static int mds_cleanup(struct obd_device *obd, int flags)
2016 {
2017         struct mds_obd *mds = &obd->u.mds;
2018         ENTRY;
2019
2020         if (mds->mds_sb == NULL)
2021                 RETURN(0);
2022
2023         mds_update_server_data(obd, 1);
2024         if (mds->mds_lov_objids != NULL) {
2025                 OBD_FREE(mds->mds_lov_objids,
2026                          mds->mds_lov_desc.ld_tgt_count * sizeof(obd_id));
2027         }
2028         mds_fs_cleanup(obd, flags);
2029
2030         unlock_kernel();
2031
2032         /* 2 seems normal on mds, (may_umount() also expects 2
2033           fwiw), but we only see 1 at this point in obdfilter. */
2034         if (atomic_read(&obd->u.mds.mds_vfsmnt->mnt_count) > 2)
2035                 CERROR("%s: mount busy, mnt_count %d != 2\n", obd->obd_name,
2036                        atomic_read(&obd->u.mds.mds_vfsmnt->mnt_count));
2037
2038         mntput(mds->mds_vfsmnt);
2039
2040         mds->mds_sb = 0;
2041
2042         ldlm_namespace_free(obd->obd_namespace, flags & OBD_OPT_FORCE);
2043
2044         spin_lock_bh(&obd->obd_processing_task_lock);
2045         if (obd->obd_recovering) {
2046                 target_cancel_recovery_timer(obd);
2047                 obd->obd_recovering = 0;
2048         }
2049         spin_unlock_bh(&obd->obd_processing_task_lock);
2050
2051         lock_kernel();
2052         dev_clear_rdonly(2);
2053         fsfilt_put_ops(obd->obd_fsops);
2054
2055         RETURN(0);
2056 }
2057
2058 static void fixup_handle_for_resent_req(struct ptlrpc_request *req,
2059                                         struct ldlm_lock *new_lock,
2060                                         struct lustre_handle *lockh)
2061 {
2062         struct obd_export *exp = req->rq_export;
2063         struct obd_device *obd = exp->exp_obd;
2064         struct ldlm_request *dlmreq =
2065                 lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*dlmreq));
2066         struct lustre_handle remote_hdl = dlmreq->lock_handle1;
2067         struct list_head *iter;
2068
2069         if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))
2070                 return;
2071
2072         l_lock(&obd->obd_namespace->ns_lock);
2073         list_for_each(iter, &exp->exp_ldlm_data.led_held_locks) {
2074                 struct ldlm_lock *lock;
2075                 lock = list_entry(iter, struct ldlm_lock, l_export_chain);
2076                 if (lock == new_lock)
2077                         continue;
2078                 if (lock->l_remote_handle.cookie == remote_hdl.cookie) {
2079                         lockh->cookie = lock->l_handle.h_cookie;
2080                         DEBUG_REQ(D_HA, req, "restoring lock cookie "LPX64,
2081                                   lockh->cookie);
2082                         l_unlock(&obd->obd_namespace->ns_lock);
2083                         return;
2084                 }
2085         }
2086         l_unlock(&obd->obd_namespace->ns_lock);
2087
2088         /* This remote handle isn't enqueued, so we never received or
2089          * processed this request.  Clear MSG_RESENT, because it can
2090          * be handled like any normal request now. */
2091
2092         lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
2093         
2094         DEBUG_REQ(D_HA, req, "no existing lock with rhandle "LPX64,
2095                   remote_hdl.cookie);
2096 }
2097
2098 int intent_disposition(struct ldlm_reply *rep, int flag)
2099 {
2100         if (!rep)
2101                 return 0;
2102         return (rep->lock_policy_res1 & flag);
2103 }
2104
2105 void intent_set_disposition(struct ldlm_reply *rep, int flag)
2106 {
2107         if (!rep)
2108                 return;
2109         rep->lock_policy_res1 |= flag;
2110 }
2111
2112 static int mds_intent_policy(struct ldlm_namespace *ns,
2113                              struct ldlm_lock **lockp, void *req_cookie,
2114                              ldlm_mode_t mode, int flags, void *data)
2115 {
2116         struct ptlrpc_request *req = req_cookie;
2117         struct ldlm_lock *lock = *lockp;
2118         struct ldlm_intent *it;
2119         struct mds_obd *mds = &req->rq_export->exp_obd->u.mds;
2120         struct ldlm_reply *rep;
2121         struct lustre_handle lockh = { 0 };
2122         struct ldlm_lock *new_lock;
2123         int getattr_part = MDS_INODELOCK_UPDATE;
2124         int rc, offset = 2, repsize[4] = {sizeof(struct ldlm_reply),
2125                                           sizeof(struct mds_body),
2126                                           mds->mds_max_mdsize,
2127                                           mds->mds_max_cookiesize};
2128         ENTRY;
2129
2130         LASSERT(req != NULL);
2131
2132         if (req->rq_reqmsg->bufcount <= 1) {
2133                 /* No intent was provided */
2134                 int size = sizeof(struct ldlm_reply);
2135                 rc = lustre_pack_reply(req, 1, &size, NULL);
2136                 LASSERT(rc == 0);
2137                 RETURN(0);
2138         }
2139
2140         it = lustre_swab_reqbuf(req, 1, sizeof(*it), lustre_swab_ldlm_intent);
2141         if (it == NULL) {
2142                 CERROR("Intent missing\n");
2143                 RETURN(req->rq_status = -EFAULT);
2144         }
2145
2146         LDLM_DEBUG(lock, "intent policy, opc: %s", ldlm_it2str(it->opc));
2147
2148         rc = lustre_pack_reply(req, it->opc == IT_UNLINK ? 4 : 3, repsize,
2149                                NULL);
2150         if (rc)
2151                 RETURN(req->rq_status = rc);
2152
2153         rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
2154         intent_set_disposition(rep, DISP_IT_EXECD);
2155
2156         fixup_handle_for_resent_req(req, lock, &lockh);
2157
2158         /* execute policy */
2159         switch ((long)it->opc) {
2160         case IT_OPEN:
2161         case IT_CREAT|IT_OPEN:
2162                 /* XXX swab here to assert that an mds_open reint
2163                  * packet is following */
2164                 rep->lock_policy_res2 = mds_reint(req, offset, &lockh);
2165 #if 0
2166                 /* We abort the lock if the lookup was negative and
2167                  * we did not make it to the OPEN portion */
2168                 if (!intent_disposition(rep, DISP_LOOKUP_EXECD))
2169                         RETURN(ELDLM_LOCK_ABORTED);
2170                 if (intent_disposition(rep, DISP_LOOKUP_NEG) &&
2171                     !intent_disposition(rep, DISP_OPEN_OPEN))
2172 #endif 
2173                         if (rep->lock_policy_res2) {
2174                                 /* mds_open returns ENOLCK where it
2175                                  * should return zero, but it has no
2176                                  * lock to return */
2177                                 if (rep->lock_policy_res2 == ENOLCK)
2178                                         rep->lock_policy_res2 = 0;
2179                                 RETURN(ELDLM_LOCK_ABORTED);
2180                         }
2181                 break;
2182         case IT_LOOKUP:
2183                 getattr_part = MDS_INODELOCK_LOOKUP;
2184         case IT_GETATTR:
2185                 getattr_part |= MDS_INODELOCK_LOOKUP;
2186         case IT_READDIR:
2187                 rep->lock_policy_res2 = mds_getattr_name(offset, req, &lockh,
2188                                                          getattr_part);
2189                 /* FIXME: LDLM can set req->rq_status. MDS sets
2190                    policy_res{1,2} with disposition and status.
2191                    - replay: returns 0 & req->status is old status 
2192                    - otherwise: returns req->status */
2193                 if (intent_disposition(rep, DISP_LOOKUP_NEG))
2194                         rep->lock_policy_res2 = 0;
2195                 if (!intent_disposition(rep, DISP_LOOKUP_POS) || 
2196                     rep->lock_policy_res2)
2197                         RETURN(ELDLM_LOCK_ABORTED);
2198                 if (req->rq_status != 0) {
2199                         LBUG();
2200                         rep->lock_policy_res2 = req->rq_status;
2201                         RETURN(ELDLM_LOCK_ABORTED);
2202                 }
2203                 break;
2204         default:
2205                 CERROR("Unhandled intent "LPD64"\n", it->opc);
2206                 LBUG();
2207         }
2208
2209         /* By this point, whatever function we called above must have either
2210          * filled in 'lockh', been an intent replay, or returned an error.  We
2211          * want to allow replayed RPCs to not get a lock, since we would just
2212          * drop it below anyways because lock replay is done separately by the
2213          * client afterwards.  For regular RPCs we want to give the new lock to
2214          * the client instead of whatever lock it was about to get. */
2215         new_lock = ldlm_handle2lock(&lockh);
2216         if (new_lock == NULL && (flags & LDLM_FL_INTENT_ONLY))
2217                 RETURN(0);
2218
2219         LASSERT(new_lock != NULL);
2220
2221         /* If we've already given this lock to a client once, then we should
2222          * have no readers or writers.  Otherwise, we should have one reader
2223          * _or_ writer ref (which will be zeroed below) before returning the
2224          * lock to a client. */
2225         if (new_lock->l_export == req->rq_export) {
2226                 LASSERT(new_lock->l_readers + new_lock->l_writers == 0);
2227         } else {
2228                 LASSERT(new_lock->l_export == NULL);
2229                 LASSERT(new_lock->l_readers + new_lock->l_writers == 1);
2230         }
2231
2232         *lockp = new_lock;
2233
2234         if (new_lock->l_export == req->rq_export) {
2235                 /* Already gave this to the client, which means that we
2236                  * reconstructed a reply. */
2237                 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) &
2238                         MSG_RESENT);
2239                 RETURN(ELDLM_LOCK_REPLACED);
2240         }
2241
2242         /* Fixup the lock to be given to the client */
2243         l_lock(&new_lock->l_resource->lr_namespace->ns_lock);
2244         new_lock->l_readers = 0;
2245         new_lock->l_writers = 0;
2246
2247         new_lock->l_export = class_export_get(req->rq_export);
2248         list_add(&new_lock->l_export_chain,
2249                  &new_lock->l_export->exp_ldlm_data.led_held_locks);
2250
2251         new_lock->l_blocking_ast = lock->l_blocking_ast;
2252         new_lock->l_completion_ast = lock->l_completion_ast;
2253
2254         memcpy(&new_lock->l_remote_handle, &lock->l_remote_handle,
2255                sizeof(lock->l_remote_handle));
2256
2257         new_lock->l_flags &= ~LDLM_FL_LOCAL;
2258
2259         LDLM_LOCK_PUT(new_lock);
2260         l_unlock(&new_lock->l_resource->lr_namespace->ns_lock);
2261
2262         RETURN(ELDLM_LOCK_REPLACED);
2263 }
2264
2265 int mds_attach(struct obd_device *dev, obd_count len, void *data)
2266 {
2267         struct lprocfs_static_vars lvars;
2268
2269         lprocfs_init_multi_vars(0, &lvars);
2270         return lprocfs_obd_attach(dev, lvars.obd_vars);
2271 }
2272
2273 int mds_detach(struct obd_device *dev)
2274 {
2275         return lprocfs_obd_detach(dev);
2276 }
2277
2278 int mdt_attach(struct obd_device *dev, obd_count len, void *data)
2279 {
2280         struct lprocfs_static_vars lvars;
2281
2282         lprocfs_init_multi_vars(1, &lvars);
2283         return lprocfs_obd_attach(dev, lvars.obd_vars);
2284 }
2285
2286 int mdt_detach(struct obd_device *dev)
2287 {
2288         return lprocfs_obd_detach(dev);
2289 }
2290
2291 static int mdt_setup(struct obd_device *obddev, obd_count len, void *buf)
2292 {
2293         struct mds_obd *mds = &obddev->u.mds;
2294         int rc = 0;
2295         ENTRY;
2296
2297         mds->mds_service = 
2298                 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
2299                                 MDS_REQUEST_PORTAL, MDC_REPLY_PORTAL,
2300                                 mds_handle, "mds",
2301                                 obddev->obd_proc_entry);
2302
2303         if (!mds->mds_service) {
2304                 CERROR("failed to start service\n");
2305                 RETURN(rc = -ENOMEM);
2306         }
2307
2308         rc = ptlrpc_start_n_threads(obddev, mds->mds_service, MDT_NUM_THREADS,
2309                                     "ll_mdt");
2310         if (rc)
2311                 GOTO(err_thread, rc);
2312
2313         mds->mds_setattr_service =
2314                 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
2315                                 MDS_SETATTR_PORTAL, MDC_REPLY_PORTAL,
2316                                 mds_handle, "mds_setattr", 
2317                                 obddev->obd_proc_entry);
2318         if (!mds->mds_setattr_service) {
2319                 CERROR("failed to start getattr service\n");
2320                 GOTO(err_thread, rc = -ENOMEM);
2321         }
2322
2323         rc = ptlrpc_start_n_threads(obddev, mds->mds_setattr_service,
2324                                  MDT_NUM_THREADS, "ll_mdt_attr");
2325         if (rc)
2326                 GOTO(err_thread2, rc);
2327                         
2328         mds->mds_readpage_service =
2329                 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
2330                                 MDS_READPAGE_PORTAL, MDC_REPLY_PORTAL,
2331                                 mds_handle, "mds_readpage", 
2332                                 obddev->obd_proc_entry);
2333         if (!mds->mds_readpage_service) {
2334                 CERROR("failed to start readpage service\n");
2335                 GOTO(err_thread2, rc = -ENOMEM);
2336         }
2337
2338         rc = ptlrpc_start_n_threads(obddev, mds->mds_readpage_service,
2339                                     MDT_NUM_THREADS, "ll_mdt_rdpg");
2340
2341         if (rc) 
2342                 GOTO(err_thread3, rc);
2343
2344         RETURN(0);
2345
2346 err_thread3:
2347         ptlrpc_unregister_service(mds->mds_readpage_service);
2348 err_thread2:
2349         ptlrpc_unregister_service(mds->mds_setattr_service);
2350 err_thread:
2351         ptlrpc_unregister_service(mds->mds_service);
2352         return rc;
2353 }
2354
2355
2356 static int mdt_cleanup(struct obd_device *obddev, int flags)
2357 {
2358         struct mds_obd *mds = &obddev->u.mds;
2359         ENTRY;
2360
2361         ptlrpc_stop_all_threads(mds->mds_readpage_service);
2362         ptlrpc_unregister_service(mds->mds_readpage_service);
2363
2364         ptlrpc_stop_all_threads(mds->mds_setattr_service);
2365         ptlrpc_unregister_service(mds->mds_setattr_service);
2366
2367         ptlrpc_stop_all_threads(mds->mds_service);
2368         ptlrpc_unregister_service(mds->mds_service);
2369
2370         RETURN(0);
2371 }
2372
2373 static struct dentry *mds_lvfs_fid2dentry(__u64 id, __u32 gen, __u64 gr, void *data)
2374 {
2375         struct obd_device *obd = data;
2376         struct ll_fid fid;
2377         fid.id = id;
2378         fid.generation = gen;
2379         return mds_fid2dentry(&obd->u.mds, &fid, NULL);
2380 }
2381
2382 struct lvfs_callback_ops mds_lvfs_ops = {
2383         l_fid2dentry:     mds_lvfs_fid2dentry,
2384 };
2385
2386 int mds_preprw(int cmd, struct obd_export *exp, struct obdo *oa,
2387                 int objcount, struct obd_ioobj *obj,
2388                 int niocount, struct niobuf_remote *nb,
2389                 struct niobuf_local *res,
2390                 struct obd_trans_info *oti);
2391 int mds_commitrw(int cmd, struct obd_export *exp, struct obdo *oa,
2392                  int objcount, struct obd_ioobj *obj, int niocount,
2393                  struct niobuf_local *res, struct obd_trans_info *oti,
2394                  int rc);
2395
2396 /* use obd ops to offer management infrastructure */
2397 static struct obd_ops mds_obd_ops = {
2398         o_owner:       THIS_MODULE,
2399         o_attach:      mds_attach,
2400         o_detach:      mds_detach,
2401         o_connect:     mds_connect,
2402         o_init_export:  mds_init_export,
2403         o_destroy_export:  mds_destroy_export,
2404         o_disconnect:  mds_disconnect,
2405         o_setup:       mds_setup,
2406         o_precleanup:  mds_precleanup,
2407         o_cleanup:     mds_cleanup,
2408         o_postrecov:   mds_postrecov,
2409         o_statfs:      mds_obd_statfs,
2410         o_iocontrol:   mds_iocontrol,
2411         o_create:      mds_obd_create,
2412         o_destroy:     mds_obd_destroy,
2413         o_llog_init:   mds_llog_init,
2414         o_llog_finish: mds_llog_finish,
2415         o_notify:      mds_notify,
2416         o_get_info:    mds_get_info,
2417         o_set_info:    mds_set_info,
2418         o_preprw:      mds_preprw, 
2419         o_commitrw:    mds_commitrw,
2420 };
2421
2422 static struct obd_ops mdt_obd_ops = {
2423         o_owner:       THIS_MODULE,
2424         o_attach:      mdt_attach,
2425         o_detach:      mdt_detach,
2426         o_setup:       mdt_setup,
2427         o_cleanup:     mdt_cleanup,
2428 };
2429
2430 static int __init mds_init(void)
2431 {
2432         struct lprocfs_static_vars lvars;
2433
2434         lprocfs_init_multi_vars(0, &lvars);
2435         class_register_type(&mds_obd_ops, NULL, lvars.module_vars, LUSTRE_MDS_NAME);
2436         lprocfs_init_multi_vars(1, &lvars);
2437         class_register_type(&mdt_obd_ops, NULL, lvars.module_vars, LUSTRE_MDT_NAME);
2438
2439         return 0;
2440 }
2441
2442 static void /*__exit*/ mds_exit(void)
2443 {
2444         class_unregister_type(LUSTRE_MDS_NAME);
2445         class_unregister_type(LUSTRE_MDT_NAME);
2446 }
2447
2448 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
2449 MODULE_DESCRIPTION("Lustre Metadata Server (MDS)");
2450 MODULE_LICENSE("GPL");
2451
2452 module_init(mds_init);
2453 module_exit(mds_exit);