Whamcloud - gitweb
009df5584529248dc70e8f92940910d5ba149ae1
[fs/lustre-release.git] / lustre / mdc / mdc_locks.c
1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2  * vim:expandtab:shiftwidth=8:tabstop=8:
3  *
4  * Copyright (C) 2001-2003 Cluster File Systems, Inc.
5  *
6  *   This file is part of Lustre, http://www.sf.net/projects/lustre/
7  *
8  *   Lustre is free software; you can redistribute it and/or
9  *   modify it under the terms of version 2 of the GNU General Public
10  *   License as published by the Free Software Foundation.
11  *
12  *   Lustre is distributed in the hope that it will be useful,
13  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
14  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
15  *   GNU General Public License for more details.
16  *
17  *   You should have received a copy of the GNU General Public License
18  *   along with Lustre; if not, write to the Free Software
19  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
20  */
21
22 #ifndef EXPORT_SYMTAB
23 # define EXPORT_SYMTAB
24 #endif
25 #define DEBUG_SUBSYSTEM S_MDC
26
27 #ifdef __KERNEL__
28 # include <linux/module.h>
29 # include <linux/pagemap.h>
30 # include <linux/miscdevice.h>
31 # include <linux/init.h>
32 #else
33 # include <liblustre.h>
34 #endif
35
36 #include <linux/obd_class.h>
37 #include <linux/lustre_mds.h>
38 #include <linux/lustre_dlm.h>
39 #include <linux/lustre_sec.h>
40 #include <linux/lprocfs_status.h>
41 #include <linux/lustre_acl.h>
42 #include <linux/lustre_lite.h>
43 #include "mdc_internal.h"
44
45 int it_disposition(struct lookup_intent *it, int flag)
46 {
47         return LUSTRE_IT(it)->it_disposition & flag;
48 }
49 EXPORT_SYMBOL(it_disposition);
50
51 void it_set_disposition(struct lookup_intent *it, int flag)
52 {
53         LUSTRE_IT(it)->it_disposition |= flag;
54 }
55 EXPORT_SYMBOL(it_set_disposition);
56
57 static void mdc_id2mdc_data(struct mdc_op_data *data,
58                             struct lustre_id *f1, 
59                             struct lustre_id *f2,
60                             const char *name, 
61                             int namelen, int mode)
62 {
63         LASSERT(data);
64         LASSERT(f1);
65
66         data->id1 = *f1;
67         if (f2)
68                 data->id2 = *f2;
69
70         data->valid = 0;
71         data->name = name;
72         data->namelen = namelen;
73         data->create_mode = mode;
74         data->mod_time = LTIME_S(CURRENT_TIME);
75 }
76
77 static int it_to_lock_mode(struct lookup_intent *it)
78 {
79         /* CREAT needs to be tested before open (both could be set) */
80         if (it->it_op & IT_CREAT)
81                 return LCK_PW;
82         else if (it->it_op & (IT_READDIR | IT_GETATTR | IT_OPEN | IT_LOOKUP |
83                               IT_CHDIR))
84                 return LCK_PR;
85
86         LBUG();
87         RETURN(-EINVAL);
88 }
89
90 int it_open_error(int phase, struct lookup_intent *it)
91 {
92         if (it_disposition(it, DISP_OPEN_OPEN)) {
93                 if (phase == DISP_OPEN_OPEN)
94                         return LUSTRE_IT(it)->it_status;
95                 else
96                         return 0;
97         }
98
99         if (it_disposition(it, DISP_OPEN_CREATE)) {
100                 if (phase == DISP_OPEN_CREATE)
101                         return LUSTRE_IT(it)->it_status;
102                 else
103                         return 0;
104         }
105
106         if (it_disposition(it, DISP_LOOKUP_EXECD)) {
107                 if (phase == DISP_LOOKUP_EXECD)
108                         return LUSTRE_IT(it)->it_status;
109                 else
110                         return 0;
111         }
112
113         if (it_disposition(it, DISP_IT_EXECD)) {
114                 if (phase == DISP_IT_EXECD)
115                         return LUSTRE_IT(it)->it_status;
116                 else
117                         return 0;
118         }
119         CERROR("it disp: %X, status: %d\n", LUSTRE_IT(it)->it_disposition,
120                LUSTRE_IT(it)->it_status);
121         LBUG();
122         return 0;
123 }
124 EXPORT_SYMBOL(it_open_error);
125
126 /* this must be called on a lockh that is known to have a referenced lock */
127 int mdc_set_lock_data(struct obd_export *exp, __u64 *l, void *data)
128 {
129         struct ldlm_lock *lock;
130         struct lustre_handle *lockh = (struct lustre_handle *)l;
131         ENTRY;
132
133         if (!*l) {
134                 EXIT;
135                 return 0;
136         }
137
138         lock = ldlm_handle2lock(lockh);
139
140         LASSERT(lock != NULL);
141         l_lock(&lock->l_resource->lr_namespace->ns_lock);
142 #ifdef __KERNEL__
143         if (lock->l_ast_data && lock->l_ast_data != data) {
144                 struct inode *new_inode = data;
145                 struct inode *old_inode = lock->l_ast_data;
146                 LASSERTF(old_inode->i_state & I_FREEING,
147                          "Found existing inode %p/%lu/%u state %lu in lock: "
148                          "setting data to %p/%lu/%u\n", old_inode,
149                          old_inode->i_ino, old_inode->i_generation,
150                          old_inode->i_state, new_inode, new_inode->i_ino,
151                          new_inode->i_generation);
152         }
153 #endif
154         lock->l_ast_data = data;
155         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
156         LDLM_LOCK_PUT(lock);
157
158         EXIT;
159         return 0;
160 }
161 EXPORT_SYMBOL(mdc_set_lock_data);
162
163 int mdc_change_cbdata(struct obd_export *exp, struct lustre_id *id, 
164                       ldlm_iterator_t it, void *data)
165 {
166         struct ldlm_res_id res_id = { .name = {0} };
167         ENTRY;
168
169         res_id.name[0] = id_fid(id);
170         res_id.name[1] = id_group(id);
171
172         ldlm_change_cbdata(class_exp2obd(exp)->obd_namespace,
173                            &res_id, it, data);
174
175         EXIT;
176         return 0;
177 }
178
179 static inline void
180 mdc_clear_replay_flag(struct ptlrpc_request *req, int rc)
181 {
182         /* Don't hold error requests for replay. */
183         if (req->rq_replay) {
184                 unsigned long irqflags;
185                 spin_lock_irqsave(&req->rq_lock, irqflags);
186                 req->rq_replay = 0;
187                 spin_unlock_irqrestore(&req->rq_lock, irqflags);
188         }
189         if (rc && req->rq_transno != 0) {
190                 DEBUG_REQ(D_ERROR, req, "transno returned on error rc %d", rc);
191                 LBUG();
192         }
193 }
194
195 /* We always reserve enough space in the reply packet for a stripe MD, because
196  * we don't know in advance the file type. */
197 int mdc_enqueue(struct obd_export *exp,
198                 int lock_type,
199                 struct lookup_intent *it,
200                 int lock_mode,
201                 struct mdc_op_data *data,
202                 struct lustre_handle *lockh,
203                 void *lmm,
204                 int lmmsize,
205                 ldlm_completion_callback cb_completion,
206                 ldlm_blocking_callback cb_blocking,
207                 void *cb_data)
208 {
209         struct ptlrpc_request *req;
210         struct ldlm_res_id res_id = {
211                 .name = {id_fid(&data->id1), id_group(&data->id1)}
212         };
213         struct obd_device *obddev = class_exp2obd(exp);
214         ldlm_policy_data_t policy = { .l_inodebits = { MDS_INODELOCK_LOOKUP } };
215         struct ldlm_intent *lit;
216         struct ldlm_request *lockreq;
217         int reqsize[6] = {[MDS_REQ_SECDESC_OFF] = 0,
218                           [MDS_REQ_INTENT_LOCKREQ_OFF] = sizeof(*lockreq),
219                           [MDS_REQ_INTENT_IT_OFF] = sizeof(*lit)};
220         int repsize[5] = {sizeof(struct ldlm_reply),
221                           sizeof(struct mds_body),
222                           obddev->u.cli.cl_max_mds_easize};
223         int req_buffers = 3, reply_buffers = 0;
224         int rc, flags = LDLM_FL_HAS_INTENT;
225         struct ldlm_reply *dlm_rep = NULL;
226         void *eadata;
227         unsigned long irqflags;
228         ENTRY;
229
230 //        LDLM_DEBUG_NOLOCK("mdsintent=%s,name=%s,dir=%lu",
231 //                          ldlm_it2str(it->it_op), it_name, it_inode->i_ino);
232
233         reqsize[0] = lustre_secdesc_size();
234
235         if (it->it_op & IT_OPEN) {
236                 it->it_create_mode |= S_IFREG;
237                 it->it_create_mode &= ~current->fs->umask;
238
239                 reqsize[req_buffers++] = sizeof(struct mds_rec_create);
240                 reqsize[req_buffers++] = data->namelen + 1;
241                 reqsize[req_buffers++] = obddev->u.cli.cl_max_mds_easize;
242
243                 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_DLM_VERSION,
244                                       LDLM_ENQUEUE, req_buffers, reqsize, NULL);
245                 if (!req)
246                         RETURN(-ENOMEM);
247
248                 spin_lock_irqsave (&req->rq_lock, irqflags);
249                 req->rq_replay = 1;
250                 spin_unlock_irqrestore (&req->rq_lock, irqflags);
251
252                 /* pack the intent */
253                 lit = lustre_msg_buf(req->rq_reqmsg, MDS_REQ_INTENT_IT_OFF,
254                                      sizeof (*lit));
255                 lit->opc = (__u64)it->it_op;
256
257                 /* pack the intended request */
258                 mdc_open_pack(req->rq_reqmsg, MDS_REQ_INTENT_REC_OFF, data,
259                               it->it_create_mode, 0, it->it_flags, lmm, lmmsize);
260                 /* get ready for the reply */
261                 repsize[3] = 4;
262                 repsize[4] = xattr_acl_size(LL_ACL_MAX_ENTRIES);
263                 reply_buffers = 5;
264                 req->rq_replen = lustre_msg_size(5, repsize);
265         } else if (it->it_op & (IT_GETATTR | IT_LOOKUP | IT_CHDIR)) {
266                 __u64 valid = data->valid | OBD_MD_FLNOTOBD | OBD_MD_FLEASIZE |
267                             OBD_MD_FLACL_ACCESS;
268
269                 reqsize[req_buffers++] = sizeof(struct mds_body);
270                 reqsize[req_buffers++] = data->namelen + 1;
271
272                 if (it->it_op & IT_GETATTR)
273                         policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
274
275                 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_DLM_VERSION,
276                                       LDLM_ENQUEUE, req_buffers, reqsize, NULL);
277
278                 if (!req)
279                         RETURN(-ENOMEM);
280
281                 /* pack the intent */
282                 lit = lustre_msg_buf(req->rq_reqmsg, MDS_REQ_INTENT_IT_OFF,
283                                      sizeof (*lit));
284                 lit->opc = (__u64)it->it_op;
285
286                 /* pack the intended request */
287                 mdc_getattr_pack(req->rq_reqmsg, MDS_REQ_INTENT_REC_OFF,
288                                  valid, it->it_flags, data);
289                 
290                 /* get ready for the reply */
291                 repsize[3] = 4;
292                 repsize[4] = xattr_acl_size(LL_ACL_MAX_ENTRIES);
293                 reply_buffers = 5;
294                 req->rq_replen = lustre_msg_size(5, repsize);
295         } else if (it->it_op == IT_READDIR) {
296                 policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
297                 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_DLM_VERSION,
298                                       LDLM_ENQUEUE, 2, reqsize, NULL);
299                 
300                 if (!req)
301                         RETURN(-ENOMEM);
302                 /* get ready for the reply */
303                 reply_buffers = 1;
304                 req->rq_replen = lustre_msg_size(1, repsize);
305         } else if (it->it_op == IT_UNLINK) {
306                 reqsize[req_buffers++] = sizeof(struct mds_body);
307                 policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
308                 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_DLM_VERSION,
309                                       LDLM_ENQUEUE, req_buffers, reqsize, NULL);
310                 if (!req)
311                         RETURN(-ENOMEM);
312
313                 /* pack the intended request */
314                 mdc_getattr_pack(req->rq_reqmsg, MDS_REQ_INTENT_REC_OFF,
315                                  0, 0, data);
316
317                 /* pack the intent */
318                 lit = lustre_msg_buf(req->rq_reqmsg, MDS_REQ_INTENT_IT_OFF,
319                                      sizeof (*lit));
320                 lit->opc = (__u64)it->it_op;
321
322                 /* get ready for the reply */
323                 reply_buffers = 3;
324                 req->rq_replen = lustre_msg_size(3, repsize);
325         } else {
326                 LBUG();
327                 RETURN(-EINVAL);
328         }
329
330         lustre_pack_secdesc(req, reqsize[0]);
331
332         mdc_get_rpc_lock(obddev->u.cli.cl_rpc_lock, it);
333         rc = ldlm_cli_enqueue(exp, req, obddev->obd_namespace, res_id,
334                               lock_type, &policy, lock_mode, &flags,cb_blocking,
335                               cb_completion, NULL, cb_data, NULL, 0, NULL,
336                               lockh);
337         mdc_put_rpc_lock(obddev->u.cli.cl_rpc_lock, it);
338
339         /* Similarly, if we're going to replay this request, we don't want to
340          * actually get a lock, just perform the intent. */
341         if (req->rq_transno || req->rq_replay) {
342                 lockreq = lustre_msg_buf(req->rq_reqmsg,
343                                          MDS_REQ_INTENT_LOCKREQ_OFF,
344                                          sizeof (*lockreq));
345                 lockreq->lock_flags |= LDLM_FL_INTENT_ONLY;
346         }
347
348         /* This can go when we're sure that this can never happen */
349         LASSERT(rc != -ENOENT);
350         /* We need dlm_rep to be assigned this early, to check lock mode of
351            returned lock from request to avoid possible race with lock
352            conversion */
353         if (rc == ELDLM_LOCK_ABORTED || !rc) {
354                 dlm_rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*dlm_rep));
355                 LASSERT(dlm_rep != NULL);   /* checked by ldlm_cli_enqueue() */
356         }
357         if (rc == ELDLM_LOCK_ABORTED) {
358                 lock_mode = 0;
359                 memset(lockh, 0, sizeof(*lockh));
360                 rc = 0;
361         } else if (rc != 0) {
362                 CERROR("ldlm_cli_enqueue: %d\n", rc);
363                 LASSERTF(rc < 0, "rc = %d\n", rc);
364                 mdc_clear_replay_flag(req, rc);
365                 ptlrpc_req_finished(req);
366                 RETURN(rc);
367         } else { /* rc = 0 */
368                 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
369                 LASSERT(lock);
370
371                 /* If the server gave us back a different lock mode, we should
372                  * fix up our variables. */
373                 if (dlm_rep->lock_desc.l_req_mode != lock_mode) {
374                         ldlm_lock_addref(lockh, dlm_rep->lock_desc.l_req_mode);
375                         ldlm_lock_decref(lockh, lock_mode);
376                         lock_mode = dlm_rep->lock_desc.l_req_mode;
377                 }
378
379                 ldlm_lock_allow_match(lock);
380                 LDLM_LOCK_PUT(lock);
381         }
382
383         LASSERT_REPSWABBED(req, 0);         /* swabbed by ldlm_cli_enqueue() */
384
385         LUSTRE_IT(it)->it_disposition = (int) dlm_rep->lock_policy_res1;
386         LUSTRE_IT(it)->it_status = (int) dlm_rep->lock_policy_res2;
387         LUSTRE_IT(it)->it_lock_mode = lock_mode;
388         LUSTRE_IT(it)->it_data = req;
389
390         if (LUSTRE_IT(it)->it_status < 0 && req->rq_replay)
391                 mdc_clear_replay_flag(req, LUSTRE_IT(it)->it_status);
392
393         DEBUG_REQ(D_DLMTRACE, req, "disposition: %x, status: %d",
394                   LUSTRE_IT(it)->it_disposition, LUSTRE_IT(it)->it_status);
395
396         /* We know what to expect, so we do any byte flipping required here */
397         LASSERT(reply_buffers == 5 || reply_buffers == 4 || 
398                 reply_buffers == 3 || reply_buffers == 1);
399         if (reply_buffers >= 3) {
400                 struct mds_body *body;
401
402                 body = lustre_swab_repbuf(req, 1, sizeof (*body),
403                                           lustre_swab_mds_body);
404                 if (body == NULL) {
405                         CERROR ("Can't swab mds_body\n");
406                         RETURN (-EPROTO);
407                 }
408
409                 if ((body->valid & OBD_MD_FLEASIZE) != 0) {
410                         /* The eadata is opaque; just check that it is there.
411                          * Eventually, obd_unpackmd() will check the contents */
412                         eadata = lustre_swab_repbuf(req, 2, body->eadatasize,
413                                                     NULL);
414                         if (eadata == NULL) {
415                                 CERROR ("Missing/short eadata\n");
416                                 RETURN (-EPROTO);
417                         }
418                         if (it->it_op & IT_OPEN) {
419                                 void *replayea;
420
421                                 replayea = lustre_msg_buf(req->rq_reqmsg,
422                                                           MDS_REQ_INTENT_REC_OFF + 2,
423                                                           body->eadatasize);
424                                 LASSERT(replayea);
425                                 memcpy(replayea, eadata, body->eadatasize);
426
427                                 LASSERT(req->rq_reqmsg->bufcount == 6);
428                                 req->rq_reqmsg->buflens[5] = body->eadatasize;
429                                 /* If this isn't the last buffer, we might
430                                  * have to shift other data around. */
431                         }
432                 }
433         }
434
435         RETURN(rc);
436 }
437 EXPORT_SYMBOL(mdc_enqueue);
438
439 /* 
440  * This long block is all about fixing up the lock and request state
441  * so that it is correct as of the moment _before_ the operation was
442  * applied; that way, the VFS will think that everything is normal and
443  * call Lustre's regular VFS methods.
444  *
445  * If we're performing a creation, that means that unless the creation
446  * failed with EEXIST, we should fake up a negative dentry.
447  *
448  * For everything else, we want to lookup to succeed.
449  *
450  * One additional note: if CREATE or OPEN succeeded, we add an extra
451  * reference to the request because we need to keep it around until
452  * ll_create/ll_open gets called.
453  *
454  * The server will return to us, in it_disposition, an indication of
455  * exactly what d.lustre->it_status refers to.
456  *
457  * If DISP_OPEN_OPEN is set, then d.lustre->it_status refers to the open() call,
458  * otherwise if DISP_OPEN_CREATE is set, then it status is the
459  * creation failure mode.  In either case, one of DISP_LOOKUP_NEG or
460  * DISP_LOOKUP_POS will be set, indicating whether the child lookup
461  * was successful.
462  *
463  * Else, if DISP_LOOKUP_EXECD then d.lustre->it_status is the rc of the
464  * child lookup.
465  */
466 int mdc_intent_lock(struct obd_export *exp, struct lustre_id *pid, 
467                     const char *name, int len, void *lmm, int lmmsize, 
468                     struct lustre_id *cid, struct lookup_intent *it, 
469                     int lookup_flags, struct ptlrpc_request **reqp,
470                     ldlm_blocking_callback cb_blocking)
471 {
472         struct lustre_handle lockh;
473         struct ptlrpc_request *request;
474         struct mds_body *mds_body;
475         struct lustre_handle old_lock;
476         struct ldlm_lock *lock;
477         int rc = 0;
478         ENTRY;
479         LASSERT(it);
480
481         CDEBUG(D_DLMTRACE, "name: %*s in obj "DLID4", intent: %s flags %#o\n",
482                len, name, OLID4(pid), ldlm_it2str(it->it_op), it->it_flags);
483
484         if (cid && (it->it_op == IT_LOOKUP || it->it_op == IT_GETATTR ||
485                     it->it_op == IT_CHDIR)) {
486                 /* We could just return 1 immediately, but since we should only
487                  * be called in revalidate_it if we already have a lock, let's
488                  * verify that. */
489                 struct ldlm_res_id res_id = {.name = {id_fid(cid),
490                                                       id_group(cid)}};
491                 struct lustre_handle lockh;
492                 ldlm_policy_data_t policy;
493                 int mode;
494
495                 /* For the GETATTR case, ll_revalidate_it issues two separate
496                    queries - for LOOKUP and for UPDATE lock because it cannot
497                    check them together - we might have those two bits to be
498                    present in two separate granted locks */
499                 policy.l_inodebits.bits = (it->it_op == IT_GETATTR) ?
500                         MDS_INODELOCK_UPDATE : MDS_INODELOCK_LOOKUP;
501                 
502                 mode = LCK_PR;
503                 rc = ldlm_lock_match(exp->exp_obd->obd_namespace,
504                                      LDLM_FL_BLOCK_GRANTED, &res_id,
505                                      LDLM_IBITS, &policy, mode,
506                                      &lockh);
507
508                 if (!rc) {
509                         mode = LCK_CR;
510                         rc = ldlm_lock_match(exp->exp_obd->obd_namespace,
511                                              LDLM_FL_BLOCK_GRANTED, &res_id,
512                                              LDLM_IBITS, &policy, mode,
513                                              &lockh);
514                 }
515                 if (!rc) {
516                         mode = LCK_PW;
517                         rc = ldlm_lock_match(exp->exp_obd->obd_namespace,
518                                              LDLM_FL_BLOCK_GRANTED, &res_id,
519                                              LDLM_IBITS, &policy, mode,
520                                              &lockh);
521                 }
522                 if (!rc) {
523                         mode = LCK_CW;
524                         rc = ldlm_lock_match(exp->exp_obd->obd_namespace,
525                                              LDLM_FL_BLOCK_GRANTED, &res_id,
526                                              LDLM_IBITS, &policy, mode,
527                                              &lockh);
528                 }
529                 if (rc) {
530                         if (ptlrpcs_check_cred(exp->exp_obd->u.cli.cl_import)) {
531                                 /* return immediately if no credential held */
532                                 ldlm_lock_decref(&lockh, mode);
533                                 RETURN(-EACCES);
534                         }
535                         memcpy(&LUSTRE_IT(it)->it_lock_handle, &lockh,
536                                sizeof(lockh));
537                         LUSTRE_IT(it)->it_lock_mode = mode;
538                 }
539
540                 /* Only return failure if it was not GETATTR by cid (from
541                    inode_revalidate) */
542                 if (rc || name)
543                         RETURN(rc);
544         }
545
546         /* lookup_it may be called only after revalidate_it has run, because
547          * revalidate_it cannot return errors, only zero.  Returning zero causes
548          * this call to lookup, which *can* return an error.
549          *
550          * We only want to execute the request associated with the intent one
551          * time, however, so don't send the request again.  Instead, skip past
552          * this and use the request from revalidate.  In this case, revalidate
553          * never dropped its reference, so the refcounts are all OK */
554         if (!it_disposition(it, DISP_ENQ_COMPLETE)) {
555                 struct mdc_op_data *op_data;
556
557                 OBD_ALLOC(op_data, sizeof(*op_data));
558                 if (op_data == NULL)
559                         RETURN(-ENOMEM);
560
561                 mdc_id2mdc_data(op_data, pid, cid, name, len, 0);
562
563                 if (name != NULL)
564                         op_data->valid |= OBD_MD_FID;
565
566                 rc = mdc_enqueue(exp, LDLM_IBITS, it, it_to_lock_mode(it),
567                                  op_data, &lockh, lmm, lmmsize,
568                                  ldlm_completion_ast, cb_blocking, NULL);
569                 OBD_FREE(op_data, sizeof(*op_data));
570                 if (rc < 0)
571                         RETURN(rc);
572                 
573                 memcpy(&LUSTRE_IT(it)->it_lock_handle, &lockh, sizeof(lockh));
574         }
575         request = *reqp = LUSTRE_IT(it)->it_data;
576         LASSERT(request != NULL);
577         
578         /* If we're doing an IT_OPEN which did not result in an actual
579          * successful open, then we need to remove the bit which saves this
580          * request for unconditional replay.
581          *
582          * It's important that we do this first!  Otherwise we might exit the
583          * function without doing so, and try to replay a failed create (bug
584          * 3440) */
585         if (it->it_op & IT_OPEN && request->rq_replay &&
586             (!it_disposition(it, DISP_OPEN_OPEN) || LUSTRE_IT(it)->it_status != 0))
587                 mdc_clear_replay_flag(request, LUSTRE_IT(it)->it_status);
588  
589         if (!it_disposition(it, DISP_IT_EXECD)) {
590                 /* The server failed before it even started executing the
591                  * intent, i.e. because it couldn't unpack the request. */
592                 LASSERT(LUSTRE_IT(it)->it_status != 0);
593                 RETURN(LUSTRE_IT(it)->it_status);
594         }
595         rc = it_open_error(DISP_IT_EXECD, it);
596         if (rc)
597                 RETURN(rc);
598
599         mds_body = lustre_msg_buf(request->rq_repmsg, 1, sizeof(*mds_body));
600         LASSERT(mds_body != NULL);      /* mdc_enqueue checked */
601         LASSERT_REPSWABBED(request, 1); /* mdc_enqueue swabbed */
602
603         /* If we were revalidating a fid/name pair, mark the intent in case we
604          * fail and get called again from lookup */
605         if (cid != NULL) {
606                 it_set_disposition(it, DISP_ENQ_COMPLETE);
607                 /* Also: did we find the same inode? */
608                 
609                 /* we have to compare all the fields but type, because MDS can
610                  * return fid/mds/ino/gen if inode lives on another MDS -bzzz */
611                 if (!(lookup_flags & LOOKUP_COBD) && !id_equal(cid, &mds_body->id1))
612                         RETURN(-ESTALE);
613         }
614
615         rc = it_open_error(DISP_LOOKUP_EXECD, it);
616         if (rc)
617                 RETURN(rc);
618
619         /*
620          * keep requests around for the multiple phases of the call this shows
621          * the DISP_XX must guarantee we make it into the call.
622          */
623         if (it_disposition(it, DISP_OPEN_CREATE) &&
624             !it_open_error(DISP_OPEN_CREATE, it))
625                 ptlrpc_request_addref(request); /* balanced in ll_create_node */
626         if (it_disposition(it, DISP_OPEN_OPEN) &&
627             !it_open_error(DISP_OPEN_OPEN, it))
628                 ptlrpc_request_addref(request); /* balanced in ll_file_open */
629
630         if (it->it_op & IT_CREAT) {
631                 /* XXX this belongs in ll_create_it */
632         } else if (it->it_op == IT_OPEN) {
633                 LASSERT(!it_disposition(it, DISP_OPEN_CREATE));
634         } else {
635                 LASSERT(it->it_op & (IT_GETATTR | IT_LOOKUP | IT_CHDIR));
636         }
637
638         /*
639          * if we already have a matching lock, then cancel the new one. We have
640          * to set the data here instead of in mdc_enqueue, because we need to
641          * use the child's inode as the l_ast_data to match, and that's not
642          * available until intent_finish has performed the iget().)
643          */
644         lock = ldlm_handle2lock(&lockh);
645         if (lock) {
646                 ldlm_policy_data_t policy = lock->l_policy_data;
647                 LDLM_DEBUG(lock, "matching against this");
648                 LDLM_LOCK_PUT(lock);
649                 
650                 LASSERTF(id_fid(&mds_body->id1) == lock->l_resource->lr_name.name[0] &&
651                          id_group(&mds_body->id1) == lock->l_resource->lr_name.name[1],
652                          "Invalid lock is returned to client. Lock res_is: %lu/%lu, "
653                          "response res_id: %lu/%lu.\n",
654                          (unsigned long)lock->l_resource->lr_name.name[0],
655                          (unsigned long)lock->l_resource->lr_name.name[1],
656                          (unsigned long)id_fid(&mds_body->id1),
657                          (unsigned long)id_group(&mds_body->id1));
658                 
659                 memcpy(&old_lock, &lockh, sizeof(lockh));
660                 if (ldlm_lock_match(NULL, LDLM_FL_BLOCK_GRANTED, NULL,
661                                     LDLM_IBITS, &policy, LCK_NL, &old_lock)) {
662                         ldlm_lock_decref_and_cancel(&lockh,
663                                                     LUSTRE_IT(it)->it_lock_mode);
664                         memcpy(&lockh, &old_lock, sizeof(old_lock));
665                         memcpy(&LUSTRE_IT(it)->it_lock_handle, &lockh,
666                                sizeof(lockh));
667                 }
668         }
669         CDEBUG(D_DENTRY, "D_IT dentry %*s intent: %s status %d disp %x rc %d\n",
670                len, name, ldlm_it2str(it->it_op), LUSTRE_IT(it)->it_status,
671                LUSTRE_IT(it)->it_disposition, rc);
672
673         RETURN(rc);
674 }
675 EXPORT_SYMBOL(mdc_intent_lock);