Whamcloud - gitweb
LU-7638 recovery: do not abort update recovery.
[fs/lustre-release.git] / lustre / mdt / mdt_handler.c
1 /*
2  * GPL HEADER START
3  *
4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
5  *
6  * This program is free software; you can redistribute it and/or modify
7  * it under the terms of the GNU General Public License version 2 only,
8  * as published by the Free Software Foundation.
9  *
10  * This program is distributed in the hope that it will be useful, but
11  * WITHOUT ANY WARRANTY; without even the implied warranty of
12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
13  * General Public License version 2 for more details (a copy is included
14  * in the LICENSE file that accompanied this code).
15  *
16  * You should have received a copy of the GNU General Public License
17  * version 2 along with this program; If not, see
18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
19  *
20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21  * CA 95054 USA or visit www.sun.com if you need additional information or
22  * have any questions.
23  *
24  * GPL HEADER END
25  */
26 /*
27  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
28  * Use is subject to license terms.
29  *
30  * Copyright (c) 2010, 2015, Intel Corporation.
31  */
32 /*
33  * This file is part of Lustre, http://www.lustre.org/
34  * Lustre is a trademark of Sun Microsystems, Inc.
35  *
36  * lustre/mdt/mdt_handler.c
37  *
38  * Lustre Metadata Target (mdt) request handler
39  *
40  * Author: Peter Braam <braam@clusterfs.com>
41  * Author: Andreas Dilger <adilger@clusterfs.com>
42  * Author: Phil Schwan <phil@clusterfs.com>
43  * Author: Mike Shaver <shaver@clusterfs.com>
44  * Author: Nikita Danilov <nikita@clusterfs.com>
45  * Author: Huang Hua <huanghua@clusterfs.com>
46  * Author: Yury Umanets <umka@clusterfs.com>
47  */
48
49 #define DEBUG_SUBSYSTEM S_MDS
50
51 #include <linux/module.h>
52
53 #include <dt_object.h>
54 #include <lustre_acl.h>
55 #include <lustre_export.h>
56 #include <lustre_ioctl.h>
57 #include <lustre_lfsck.h>
58 #include <lustre_log.h>
59 #include <lustre_net.h>
60 #include <lustre_nodemap.h>
61 #include <lustre_mds.h>
62 #include <lustre_param.h>
63 #include <lustre_quota.h>
64 #include <lustre_swab.h>
65 #include <obd.h>
66 #include <obd_support.h>
67
68 #include "mdt_internal.h"
69
70
71 static unsigned int max_mod_rpcs_per_client = 8;
72 CFS_MODULE_PARM(max_mod_rpcs_per_client, "i", uint, 0644,
73                 "maximum number of modify RPCs in flight allowed per client");
74
75 mdl_mode_t mdt_mdl_lock_modes[] = {
76         [LCK_MINMODE] = MDL_MINMODE,
77         [LCK_EX]      = MDL_EX,
78         [LCK_PW]      = MDL_PW,
79         [LCK_PR]      = MDL_PR,
80         [LCK_CW]      = MDL_CW,
81         [LCK_CR]      = MDL_CR,
82         [LCK_NL]      = MDL_NL,
83         [LCK_GROUP]   = MDL_GROUP
84 };
85
86 enum ldlm_mode mdt_dlm_lock_modes[] = {
87         [MDL_MINMODE]   = LCK_MINMODE,
88         [MDL_EX]        = LCK_EX,
89         [MDL_PW]        = LCK_PW,
90         [MDL_PR]        = LCK_PR,
91         [MDL_CW]        = LCK_CW,
92         [MDL_CR]        = LCK_CR,
93         [MDL_NL]        = LCK_NL,
94         [MDL_GROUP]     = LCK_GROUP
95 };
96
97 static struct mdt_device *mdt_dev(struct lu_device *d);
98 static int mdt_unpack_req_pack_rep(struct mdt_thread_info *info, __u32 flags);
99
100 static const struct lu_object_operations mdt_obj_ops;
101
102 /* Slab for MDT object allocation */
103 static struct kmem_cache *mdt_object_kmem;
104
105 /* For HSM restore handles */
106 struct kmem_cache *mdt_hsm_cdt_kmem;
107
108 /* For HSM request handles */
109 struct kmem_cache *mdt_hsm_car_kmem;
110
111 static struct lu_kmem_descr mdt_caches[] = {
112         {
113                 .ckd_cache = &mdt_object_kmem,
114                 .ckd_name  = "mdt_obj",
115                 .ckd_size  = sizeof(struct mdt_object)
116         },
117         {
118                 .ckd_cache      = &mdt_hsm_cdt_kmem,
119                 .ckd_name       = "mdt_cdt_restore_handle",
120                 .ckd_size       = sizeof(struct cdt_restore_handle)
121         },
122         {
123                 .ckd_cache      = &mdt_hsm_car_kmem,
124                 .ckd_name       = "mdt_cdt_agent_req",
125                 .ckd_size       = sizeof(struct cdt_agent_req)
126         },
127         {
128                 .ckd_cache = NULL
129         }
130 };
131
132 __u64 mdt_get_disposition(struct ldlm_reply *rep, __u64 op_flag)
133 {
134         if (!rep)
135                 return 0;
136         return rep->lock_policy_res1 & op_flag;
137 }
138
139 void mdt_clear_disposition(struct mdt_thread_info *info,
140                            struct ldlm_reply *rep, __u64 op_flag)
141 {
142         if (info) {
143                 info->mti_opdata &= ~op_flag;
144                 tgt_opdata_clear(info->mti_env, op_flag);
145         }
146         if (rep)
147                 rep->lock_policy_res1 &= ~op_flag;
148 }
149
150 void mdt_set_disposition(struct mdt_thread_info *info,
151                          struct ldlm_reply *rep, __u64 op_flag)
152 {
153         if (info) {
154                 info->mti_opdata |= op_flag;
155                 tgt_opdata_set(info->mti_env, op_flag);
156         }
157         if (rep)
158                 rep->lock_policy_res1 |= op_flag;
159 }
160
161 void mdt_lock_reg_init(struct mdt_lock_handle *lh, enum ldlm_mode lm)
162 {
163         lh->mlh_pdo_hash = 0;
164         lh->mlh_reg_mode = lm;
165         lh->mlh_rreg_mode = lm;
166         lh->mlh_type = MDT_REG_LOCK;
167 }
168
169 void mdt_lock_pdo_init(struct mdt_lock_handle *lh, enum ldlm_mode lock_mode,
170                        const struct lu_name *lname)
171 {
172         lh->mlh_reg_mode = lock_mode;
173         lh->mlh_rreg_mode = lock_mode;
174         lh->mlh_type = MDT_PDO_LOCK;
175
176         if (lu_name_is_valid(lname)) {
177                 lh->mlh_pdo_hash = full_name_hash(lname->ln_name,
178                                                   lname->ln_namelen);
179                 /* XXX Workaround for LU-2856
180                  *
181                  * Zero is a valid return value of full_name_hash, but
182                  * several users of mlh_pdo_hash assume a non-zero
183                  * hash value. We therefore map zero onto an
184                  * arbitrary, but consistent value (1) to avoid
185                  * problems further down the road. */
186                 if (unlikely(lh->mlh_pdo_hash == 0))
187                         lh->mlh_pdo_hash = 1;
188         } else {
189                 lh->mlh_pdo_hash = 0;
190         }
191 }
192
193 static void mdt_lock_pdo_mode(struct mdt_thread_info *info, struct mdt_object *o,
194                               struct mdt_lock_handle *lh)
195 {
196         mdl_mode_t mode;
197         ENTRY;
198
199         /*
200          * Any dir access needs couple of locks:
201          *
202          * 1) on part of dir we gonna take lookup/modify;
203          *
204          * 2) on whole dir to protect it from concurrent splitting and/or to
205          * flush client's cache for readdir().
206          *
207          * so, for a given mode and object this routine decides what lock mode
208          * to use for lock #2:
209          *
210          * 1) if caller's gonna lookup in dir then we need to protect dir from
211          * being splitted only - LCK_CR
212          *
213          * 2) if caller's gonna modify dir then we need to protect dir from
214          * being splitted and to flush cache - LCK_CW
215          *
216          * 3) if caller's gonna modify dir and that dir seems ready for
217          * splitting then we need to protect it from any type of access
218          * (lookup/modify/split) - LCK_EX --bzzz
219          */
220
221         LASSERT(lh->mlh_reg_mode != LCK_MINMODE);
222         LASSERT(lh->mlh_pdo_mode == LCK_MINMODE);
223
224         /*
225          * Ask underlaying level its opinion about preferable PDO lock mode
226          * having access type passed as regular lock mode:
227          *
228          * - MDL_MINMODE means that lower layer does not want to specify lock
229          * mode;
230          *
231          * - MDL_NL means that no PDO lock should be taken. This is used in some
232          * cases. Say, for non-splittable directories no need to use PDO locks
233          * at all.
234          */
235         mode = mdo_lock_mode(info->mti_env, mdt_object_child(o),
236                              mdt_dlm_mode2mdl_mode(lh->mlh_reg_mode));
237
238         if (mode != MDL_MINMODE) {
239                 lh->mlh_pdo_mode = mdt_mdl_mode2dlm_mode(mode);
240         } else {
241                 /*
242                  * Lower layer does not want to specify locking mode. We do it
243                  * our selves. No special protection is needed, just flush
244                  * client's cache on modification and allow concurrent
245                  * mondification.
246                  */
247                 switch (lh->mlh_reg_mode) {
248                 case LCK_EX:
249                         lh->mlh_pdo_mode = LCK_EX;
250                         break;
251                 case LCK_PR:
252                         lh->mlh_pdo_mode = LCK_CR;
253                         break;
254                 case LCK_PW:
255                         lh->mlh_pdo_mode = LCK_CW;
256                         break;
257                 default:
258                         CERROR("Not expected lock type (0x%x)\n",
259                                (int)lh->mlh_reg_mode);
260                         LBUG();
261                 }
262         }
263
264         LASSERT(lh->mlh_pdo_mode != LCK_MINMODE);
265         EXIT;
266 }
267
268 static int mdt_getstatus(struct tgt_session_info *tsi)
269 {
270         struct mdt_thread_info  *info = tsi2mdt_info(tsi);
271         struct mdt_device       *mdt = info->mti_mdt;
272         struct mdt_body         *repbody;
273         int                      rc;
274
275         ENTRY;
276
277         rc = mdt_check_ucred(info);
278         if (rc)
279                 GOTO(out, rc = err_serious(rc));
280
281         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETSTATUS_PACK))
282                 GOTO(out, rc = err_serious(-ENOMEM));
283
284         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
285         repbody->mbo_fid1 = mdt->mdt_md_root_fid;
286         repbody->mbo_valid |= OBD_MD_FLID;
287
288         EXIT;
289 out:
290         mdt_thread_info_fini(info);
291         return rc;
292 }
293
294 static int mdt_statfs(struct tgt_session_info *tsi)
295 {
296         struct ptlrpc_request           *req = tgt_ses_req(tsi);
297         struct mdt_thread_info          *info = tsi2mdt_info(tsi);
298         struct md_device                *next = info->mti_mdt->mdt_child;
299         struct ptlrpc_service_part      *svcpt;
300         struct obd_statfs               *osfs;
301         int                             rc;
302
303         ENTRY;
304
305         svcpt = req->rq_rqbd->rqbd_svcpt;
306
307         /* This will trigger a watchdog timeout */
308         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_STATFS_LCW_SLEEP,
309                          (MDT_SERVICE_WATCHDOG_FACTOR *
310                           at_get(&svcpt->scp_at_estimate)) + 1);
311
312         rc = mdt_check_ucred(info);
313         if (rc)
314                 GOTO(out, rc = err_serious(rc));
315
316         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_STATFS_PACK))
317                 GOTO(out, rc = err_serious(-ENOMEM));
318
319         osfs = req_capsule_server_get(info->mti_pill, &RMF_OBD_STATFS);
320         if (!osfs)
321                 GOTO(out, rc = -EPROTO);
322
323         /** statfs information are cached in the mdt_device */
324         if (cfs_time_before_64(info->mti_mdt->mdt_osfs_age,
325                                cfs_time_shift_64(-OBD_STATFS_CACHE_SECONDS))) {
326                 /** statfs data is too old, get up-to-date one */
327                 rc = next->md_ops->mdo_statfs(info->mti_env, next, osfs);
328                 if (rc)
329                         GOTO(out, rc);
330                 spin_lock(&info->mti_mdt->mdt_osfs_lock);
331                 info->mti_mdt->mdt_osfs = *osfs;
332                 info->mti_mdt->mdt_osfs_age = cfs_time_current_64();
333                 spin_unlock(&info->mti_mdt->mdt_osfs_lock);
334         } else {
335                 /** use cached statfs data */
336                 spin_lock(&info->mti_mdt->mdt_osfs_lock);
337                 *osfs = info->mti_mdt->mdt_osfs;
338                 spin_unlock(&info->mti_mdt->mdt_osfs_lock);
339         }
340
341         if (rc == 0)
342                 mdt_counter_incr(req, LPROC_MDT_STATFS);
343 out:
344         mdt_thread_info_fini(info);
345         RETURN(rc);
346 }
347
348 #ifdef CONFIG_FS_POSIX_ACL
349 /*
350  * Pack ACL data into the reply. UIDs/GIDs are mapped and filtered by nodemap.
351  *
352  * \param       info    thread info object
353  * \param       repbody reply to pack ACLs into
354  * \param       o       mdt object of file to examine
355  * \param       nodemap nodemap of client to reply to
356  * \retval      0       success
357  * \retval      -errno  error getting or parsing ACL from disk
358  */
359 int mdt_pack_acl2body(struct mdt_thread_info *info, struct mdt_body *repbody,
360                       struct mdt_object *o, struct lu_nodemap *nodemap)
361 {
362         const struct lu_env     *env = info->mti_env;
363         struct md_object        *next = mdt_object_child(o);
364         struct lu_buf           *buf = &info->mti_buf;
365         int rc;
366
367         buf->lb_buf = req_capsule_server_get(info->mti_pill, &RMF_ACL);
368         buf->lb_len = req_capsule_get_size(info->mti_pill, &RMF_ACL,
369                                            RCL_SERVER);
370         if (buf->lb_len == 0)
371                 return 0;
372
373         rc = mo_xattr_get(env, next, buf, XATTR_NAME_ACL_ACCESS);
374         if (rc < 0) {
375                 if (rc == -ENODATA) {
376                         repbody->mbo_aclsize = 0;
377                         repbody->mbo_valid |= OBD_MD_FLACL;
378                         rc = 0;
379                 } else if (rc == -EOPNOTSUPP) {
380                         rc = 0;
381                 } else {
382                         CERROR("%s: unable to read "DFID" ACL: rc = %d\n",
383                                mdt_obd_name(info->mti_mdt),
384                                PFID(mdt_object_fid(o)), rc);
385                 }
386         } else {
387                 rc = nodemap_map_acl(nodemap, buf->lb_buf,
388                                      rc, NODEMAP_FS_TO_CLIENT);
389                 /* if all ACLs mapped out, rc is still >= 0 */
390                 if (rc < 0) {
391                         CERROR("%s: nodemap_map_acl unable to parse "DFID
392                                " ACL: rc = %d\n", mdt_obd_name(info->mti_mdt),
393                                PFID(mdt_object_fid(o)), rc);
394                 } else {
395                         repbody->mbo_aclsize = rc;
396                         repbody->mbo_valid |= OBD_MD_FLACL;
397                         rc = 0;
398                 }
399         }
400         return rc;
401 }
402 #endif
403
404 void mdt_pack_attr2body(struct mdt_thread_info *info, struct mdt_body *b,
405                         const struct lu_attr *attr, const struct lu_fid *fid)
406 {
407         struct md_attr          *ma = &info->mti_attr;
408         struct obd_export       *exp = info->mti_exp;
409         struct lu_nodemap       *nodemap = exp->exp_target_data.ted_nodemap;
410
411         LASSERT(ma->ma_valid & MA_INODE);
412
413         if (attr->la_valid & LA_ATIME) {
414                 b->mbo_atime = attr->la_atime;
415                 b->mbo_valid |= OBD_MD_FLATIME;
416         }
417         if (attr->la_valid & LA_MTIME) {
418                 b->mbo_mtime = attr->la_mtime;
419                 b->mbo_valid |= OBD_MD_FLMTIME;
420         }
421         if (attr->la_valid & LA_CTIME) {
422                 b->mbo_ctime = attr->la_ctime;
423                 b->mbo_valid |= OBD_MD_FLCTIME;
424         }
425         if (attr->la_valid & LA_FLAGS) {
426                 b->mbo_flags = attr->la_flags;
427                 b->mbo_valid |= OBD_MD_FLFLAGS;
428         }
429         if (attr->la_valid & LA_NLINK) {
430                 b->mbo_nlink = attr->la_nlink;
431                 b->mbo_valid |= OBD_MD_FLNLINK;
432         }
433         if (attr->la_valid & LA_UID) {
434                 b->mbo_uid = nodemap_map_id(nodemap, NODEMAP_UID,
435                                             NODEMAP_FS_TO_CLIENT,
436                                             attr->la_uid);
437                 b->mbo_valid |= OBD_MD_FLUID;
438         }
439         if (attr->la_valid & LA_GID) {
440                 b->mbo_gid = nodemap_map_id(nodemap, NODEMAP_GID,
441                                             NODEMAP_FS_TO_CLIENT,
442                                             attr->la_gid);
443                 b->mbo_valid |= OBD_MD_FLGID;
444         }
445         b->mbo_mode = attr->la_mode;
446         if (attr->la_valid & LA_MODE)
447                 b->mbo_valid |= OBD_MD_FLMODE;
448         if (attr->la_valid & LA_TYPE)
449                 b->mbo_valid |= OBD_MD_FLTYPE;
450
451         if (fid != NULL) {
452                 b->mbo_fid1 = *fid;
453                 b->mbo_valid |= OBD_MD_FLID;
454                 CDEBUG(D_INODE, DFID": nlink=%d, mode=%o, valid="LPX64"\n",
455                        PFID(fid), b->mbo_nlink, b->mbo_mode, b->mbo_valid);
456         }
457
458         if (info != NULL)
459                 mdt_body_reverse_idmap(info, b);
460
461         if (!(attr->la_valid & LA_TYPE))
462                 return;
463
464         b->mbo_rdev   = attr->la_rdev;
465         b->mbo_size   = attr->la_size;
466         b->mbo_blocks = attr->la_blocks;
467
468         if (!S_ISREG(attr->la_mode)) {
469                 b->mbo_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS | OBD_MD_FLRDEV;
470         } else if (ma->ma_need & MA_LOV && !(ma->ma_valid & MA_LOV)) {
471                 /* means no objects are allocated on osts. */
472                 LASSERT(!(ma->ma_valid & MA_LOV));
473                 /* just ignore blocks occupied by extend attributes on MDS */
474                 b->mbo_blocks = 0;
475                 /* if no object is allocated on osts, the size on mds is valid.
476                  * b=22272 */
477                 b->mbo_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
478         } else if ((ma->ma_valid & MA_LOV) && ma->ma_lmm != NULL &&
479                    ma->ma_lmm->lmm_pattern & LOV_PATTERN_F_RELEASED) {
480                 /* A released file stores its size on MDS. */
481                 /* But return 1 block for released file, unless tools like tar
482                  * will consider it fully sparse. (LU-3864)
483                  */
484                 if (unlikely(b->mbo_size == 0))
485                         b->mbo_blocks = 0;
486                 else
487                         b->mbo_blocks = 1;
488                 b->mbo_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
489         }
490
491         if (fid != NULL && (b->mbo_valid & OBD_MD_FLSIZE))
492                 CDEBUG(D_VFSTRACE, DFID": returning size %llu\n",
493                        PFID(fid), (unsigned long long)b->mbo_size);
494 }
495
496 static inline int mdt_body_has_lov(const struct lu_attr *la,
497                                    const struct mdt_body *body)
498 {
499         return (S_ISREG(la->la_mode) && (body->mbo_valid & OBD_MD_FLEASIZE)) ||
500                (S_ISDIR(la->la_mode) && (body->mbo_valid & OBD_MD_FLDIREA));
501 }
502
503 void mdt_client_compatibility(struct mdt_thread_info *info)
504 {
505         struct mdt_body       *body;
506         struct ptlrpc_request *req = mdt_info_req(info);
507         struct obd_export     *exp = req->rq_export;
508         struct md_attr        *ma = &info->mti_attr;
509         struct lu_attr        *la = &ma->ma_attr;
510         ENTRY;
511
512         if (exp_connect_layout(exp))
513                 /* the client can deal with 16-bit lmm_stripe_count */
514                 RETURN_EXIT;
515
516         body = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
517
518         if (!mdt_body_has_lov(la, body))
519                 RETURN_EXIT;
520
521         /* now we have a reply with a lov for a client not compatible with the
522          * layout lock so we have to clean the layout generation number */
523         if (S_ISREG(la->la_mode))
524                 ma->ma_lmm->lmm_layout_gen = 0;
525         EXIT;
526 }
527
528 static int mdt_attr_get_eabuf_size(struct mdt_thread_info *info,
529                                    struct mdt_object *o)
530 {
531         const struct lu_env *env = info->mti_env;
532         int rc, rc2;
533
534         rc = mo_xattr_get(env, mdt_object_child(o), &LU_BUF_NULL,
535                           XATTR_NAME_LOV);
536
537         if (rc == -ENODATA)
538                 rc = 0;
539
540         if (rc < 0)
541                 goto out;
542
543         /* Is it a directory? Let's check for the LMV as well */
544         if (S_ISDIR(lu_object_attr(&mdt_object_child(o)->mo_lu))) {
545                 rc2 = mo_xattr_get(env, mdt_object_child(o), &LU_BUF_NULL,
546                                    XATTR_NAME_LMV);
547
548                 if (rc2 == -ENODATA)
549                         rc2 = mo_xattr_get(env, mdt_object_child(o),
550                                            &LU_BUF_NULL,
551                                            XATTR_NAME_DEFAULT_LMV);
552
553                 if ((rc2 < 0 && rc2 != -ENODATA) || (rc2 > rc))
554                         rc = rc2;
555         }
556
557 out:
558         return rc;
559 }
560
561 int mdt_big_xattr_get(struct mdt_thread_info *info, struct mdt_object *o,
562                       const char *name)
563 {
564         const struct lu_env *env = info->mti_env;
565         int rc;
566         ENTRY;
567
568         LASSERT(info->mti_big_lmm_used == 0);
569         rc = mo_xattr_get(env, mdt_object_child(o), &LU_BUF_NULL, name);
570         if (rc < 0)
571                 RETURN(rc);
572
573         /* big_lmm may need to be grown */
574         if (info->mti_big_lmmsize < rc) {
575                 int size = size_roundup_power2(rc);
576
577                 if (info->mti_big_lmmsize > 0) {
578                         /* free old buffer */
579                         LASSERT(info->mti_big_lmm);
580                         OBD_FREE_LARGE(info->mti_big_lmm,
581                                        info->mti_big_lmmsize);
582                         info->mti_big_lmm = NULL;
583                         info->mti_big_lmmsize = 0;
584                 }
585
586                 OBD_ALLOC_LARGE(info->mti_big_lmm, size);
587                 if (info->mti_big_lmm == NULL)
588                         RETURN(-ENOMEM);
589                 info->mti_big_lmmsize = size;
590         }
591         LASSERT(info->mti_big_lmmsize >= rc);
592
593         info->mti_buf.lb_buf = info->mti_big_lmm;
594         info->mti_buf.lb_len = info->mti_big_lmmsize;
595         rc = mo_xattr_get(env, mdt_object_child(o), &info->mti_buf, name);
596
597         RETURN(rc);
598 }
599
600 int mdt_stripe_get(struct mdt_thread_info *info, struct mdt_object *o,
601                    struct md_attr *ma, const char *name)
602 {
603         struct md_object *next = mdt_object_child(o);
604         struct lu_buf    *buf = &info->mti_buf;
605         int rc;
606
607         if (strcmp(name, XATTR_NAME_LOV) == 0) {
608                 buf->lb_buf = ma->ma_lmm;
609                 buf->lb_len = ma->ma_lmm_size;
610                 LASSERT(!(ma->ma_valid & MA_LOV));
611         } else if (strcmp(name, XATTR_NAME_LMV) == 0) {
612                 buf->lb_buf = ma->ma_lmv;
613                 buf->lb_len = ma->ma_lmv_size;
614                 LASSERT(!(ma->ma_valid & MA_LMV));
615         } else if (strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0) {
616                 buf->lb_buf = ma->ma_lmv;
617                 buf->lb_len = ma->ma_lmv_size;
618                 LASSERT(!(ma->ma_valid & MA_LMV_DEF));
619         } else {
620                 return -EINVAL;
621         }
622
623         rc = mo_xattr_get(info->mti_env, next, buf, name);
624         if (rc > 0) {
625
626 got:
627                 if (strcmp(name, XATTR_NAME_LOV) == 0) {
628                         if (info->mti_big_lmm_used)
629                                 ma->ma_lmm = info->mti_big_lmm;
630
631                         /* NOT return LOV EA with hole to old client. */
632                         if (unlikely(le32_to_cpu(ma->ma_lmm->lmm_pattern) &
633                                      LOV_PATTERN_F_HOLE) &&
634                             !(exp_connect_flags(info->mti_exp) &
635                               OBD_CONNECT_LFSCK)) {
636                                 return -EIO;
637                         } else {
638                                 ma->ma_lmm_size = rc;
639                                 ma->ma_valid |= MA_LOV;
640                         }
641                 } else if (strcmp(name, XATTR_NAME_LMV) == 0) {
642                         if (info->mti_big_lmm_used)
643                                 ma->ma_lmv = info->mti_big_lmm;
644
645                         ma->ma_lmv_size = rc;
646                         ma->ma_valid |= MA_LMV;
647                 } else if (strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0) {
648                         ma->ma_lmv_size = rc;
649                         ma->ma_valid |= MA_LMV_DEF;
650                 }
651
652                 /* Update mdt_max_mdsize so all clients will be aware that */
653                 if (info->mti_mdt->mdt_max_mdsize < rc)
654                         info->mti_mdt->mdt_max_mdsize = rc;
655
656                 rc = 0;
657         } else if (rc == -ENODATA) {
658                 /* no LOV EA */
659                 rc = 0;
660         } else if (rc == -ERANGE) {
661                 /* Default LMV has fixed size, so it must be able to fit
662                  * in the original buffer */
663                 if (strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0)
664                         return rc;
665                 rc = mdt_big_xattr_get(info, o, name);
666                 if (rc > 0) {
667                         info->mti_big_lmm_used = 1;
668                         goto got;
669                 }
670         }
671
672         return rc;
673 }
674
675 static int mdt_attr_get_pfid(struct mdt_thread_info *info,
676                              struct mdt_object *o, struct lu_fid *pfid)
677 {
678         struct lu_buf           *buf = &info->mti_buf;
679         struct link_ea_header   *leh;
680         struct link_ea_entry    *lee;
681         int                      rc;
682         ENTRY;
683
684         buf->lb_buf = info->mti_big_lmm;
685         buf->lb_len = info->mti_big_lmmsize;
686         rc = mo_xattr_get(info->mti_env, mdt_object_child(o),
687                           buf, XATTR_NAME_LINK);
688         /* ignore errors, MA_PFID won't be set and it is
689          * up to the caller to treat this as an error */
690         if (rc == -ERANGE || buf->lb_len == 0) {
691                 rc = mdt_big_xattr_get(info, o, XATTR_NAME_LINK);
692                 buf->lb_buf = info->mti_big_lmm;
693                 buf->lb_len = info->mti_big_lmmsize;
694         }
695
696         if (rc < 0)
697                 RETURN(rc);
698         if (rc < sizeof(*leh)) {
699                 CERROR("short LinkEA on "DFID": rc = %d\n",
700                        PFID(mdt_object_fid(o)), rc);
701                 RETURN(-ENODATA);
702         }
703
704         leh = (struct link_ea_header *) buf->lb_buf;
705         lee = (struct link_ea_entry *)(leh + 1);
706         if (leh->leh_magic == __swab32(LINK_EA_MAGIC)) {
707                 leh->leh_magic = LINK_EA_MAGIC;
708                 leh->leh_reccount = __swab32(leh->leh_reccount);
709                 leh->leh_len = __swab64(leh->leh_len);
710         }
711         if (leh->leh_magic != LINK_EA_MAGIC)
712                 RETURN(-EINVAL);
713         if (leh->leh_reccount == 0)
714                 RETURN(-ENODATA);
715
716         memcpy(pfid, &lee->lee_parent_fid, sizeof(*pfid));
717         fid_be_to_cpu(pfid, pfid);
718
719         RETURN(0);
720 }
721
722 int mdt_attr_get_complex(struct mdt_thread_info *info,
723                          struct mdt_object *o, struct md_attr *ma)
724 {
725         const struct lu_env *env = info->mti_env;
726         struct md_object    *next = mdt_object_child(o);
727         struct lu_buf       *buf = &info->mti_buf;
728         int                  need = ma->ma_need;
729         int                  rc = 0, rc2;
730         u32                  mode;
731         ENTRY;
732
733         ma->ma_valid = 0;
734
735         if (mdt_object_exists(o) == 0)
736                 GOTO(out, rc = -ENOENT);
737         mode = lu_object_attr(&next->mo_lu);
738
739         if (need & MA_INODE) {
740                 ma->ma_need = MA_INODE;
741                 rc = mo_attr_get(env, next, ma);
742                 if (rc)
743                         GOTO(out, rc);
744                 ma->ma_valid |= MA_INODE;
745         }
746
747         if (need & MA_PFID) {
748                 rc = mdt_attr_get_pfid(info, o, &ma->ma_pfid);
749                 if (rc == 0)
750                         ma->ma_valid |= MA_PFID;
751                 /* ignore this error, parent fid is not mandatory */
752                 rc = 0;
753         }
754
755         if (need & MA_LOV && (S_ISREG(mode) || S_ISDIR(mode))) {
756                 rc = mdt_stripe_get(info, o, ma, XATTR_NAME_LOV);
757                 if (rc)
758                         GOTO(out, rc);
759         }
760
761         if (need & MA_LMV && S_ISDIR(mode)) {
762                 rc = mdt_stripe_get(info, o, ma, XATTR_NAME_LMV);
763                 if (rc != 0)
764                         GOTO(out, rc);
765         }
766
767         if (need & MA_LMV_DEF && S_ISDIR(mode)) {
768                 rc = mdt_stripe_get(info, o, ma, XATTR_NAME_DEFAULT_LMV);
769                 if (rc != 0)
770                         GOTO(out, rc);
771         }
772
773         if (need & MA_HSM && S_ISREG(mode)) {
774                 buf->lb_buf = info->mti_xattr_buf;
775                 buf->lb_len = sizeof(info->mti_xattr_buf);
776                 CLASSERT(sizeof(struct hsm_attrs) <=
777                          sizeof(info->mti_xattr_buf));
778                 rc2 = mo_xattr_get(info->mti_env, next, buf, XATTR_NAME_HSM);
779                 rc2 = lustre_buf2hsm(info->mti_xattr_buf, rc2, &ma->ma_hsm);
780                 if (rc2 == 0)
781                         ma->ma_valid |= MA_HSM;
782                 else if (rc2 < 0 && rc2 != -ENODATA)
783                         GOTO(out, rc = rc2);
784         }
785
786 #ifdef CONFIG_FS_POSIX_ACL
787         if (need & MA_ACL_DEF && S_ISDIR(mode)) {
788                 buf->lb_buf = ma->ma_acl;
789                 buf->lb_len = ma->ma_acl_size;
790                 rc2 = mo_xattr_get(env, next, buf, XATTR_NAME_ACL_DEFAULT);
791                 if (rc2 > 0) {
792                         ma->ma_acl_size = rc2;
793                         ma->ma_valid |= MA_ACL_DEF;
794                 } else if (rc2 == -ENODATA) {
795                         /* no ACLs */
796                         ma->ma_acl_size = 0;
797                 } else
798                         GOTO(out, rc = rc2);
799         }
800 #endif
801 out:
802         ma->ma_need = need;
803         CDEBUG(D_INODE, "after getattr rc = %d, ma_valid = "LPX64" ma_lmm=%p\n",
804                rc, ma->ma_valid, ma->ma_lmm);
805         RETURN(rc);
806 }
807
808 static int mdt_getattr_internal(struct mdt_thread_info *info,
809                                 struct mdt_object *o, int ma_need)
810 {
811         struct md_object        *next = mdt_object_child(o);
812         const struct mdt_body   *reqbody = info->mti_body;
813         struct ptlrpc_request   *req = mdt_info_req(info);
814         struct md_attr          *ma = &info->mti_attr;
815         struct lu_attr          *la = &ma->ma_attr;
816         struct req_capsule      *pill = info->mti_pill;
817         const struct lu_env     *env = info->mti_env;
818         struct mdt_body         *repbody;
819         struct lu_buf           *buffer = &info->mti_buf;
820         struct obd_export       *exp = info->mti_exp;
821         struct lu_nodemap       *nodemap = exp->exp_target_data.ted_nodemap;
822         int                      rc;
823         int                      is_root;
824         ENTRY;
825
826         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK))
827                 RETURN(err_serious(-ENOMEM));
828
829         repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
830
831         ma->ma_valid = 0;
832
833         if (mdt_object_remote(o)) {
834                 /* This object is located on remote node.*/
835                 /* Return -ENOTSUPP for old client */
836                 if (!mdt_is_dne_client(req->rq_export))
837                         GOTO(out, rc = -ENOTSUPP);
838
839                 repbody->mbo_fid1 = *mdt_object_fid(o);
840                 repbody->mbo_valid = OBD_MD_FLID | OBD_MD_MDS;
841                 GOTO(out, rc = 0);
842         }
843
844         if (reqbody->mbo_eadatasize > 0) {
845                 buffer->lb_buf = req_capsule_server_get(pill, &RMF_MDT_MD);
846                 if (buffer->lb_buf == NULL)
847                         GOTO(out, rc = -EPROTO);
848                 buffer->lb_len = req_capsule_get_size(pill, &RMF_MDT_MD,
849                                                       RCL_SERVER);
850         } else {
851                 buffer->lb_buf = NULL;
852                 buffer->lb_len = 0;
853                 ma_need &= ~(MA_LOV | MA_LMV);
854                 CDEBUG(D_INFO, "%s: RPC from %s: does not need LOVEA.\n",
855                        mdt_obd_name(info->mti_mdt),
856                        req->rq_export->exp_client_uuid.uuid);
857         }
858
859         /* If it is dir object and client require MEA, then we got MEA */
860         if (S_ISDIR(lu_object_attr(&next->mo_lu)) &&
861             (reqbody->mbo_valid & (OBD_MD_MEA | OBD_MD_DEFAULT_MEA))) {
862                 /* Assumption: MDT_MD size is enough for lmv size. */
863                 ma->ma_lmv = buffer->lb_buf;
864                 ma->ma_lmv_size = buffer->lb_len;
865                 ma->ma_need = MA_INODE;
866                 if (ma->ma_lmv_size > 0) {
867                         if (reqbody->mbo_valid & OBD_MD_MEA)
868                                 ma->ma_need |= MA_LMV;
869                         else if (reqbody->mbo_valid & OBD_MD_DEFAULT_MEA)
870                                 ma->ma_need |= MA_LMV_DEF;
871                 }
872         } else {
873                 ma->ma_lmm = buffer->lb_buf;
874                 ma->ma_lmm_size = buffer->lb_len;
875                 ma->ma_need = MA_INODE | MA_HSM;
876                 if (ma->ma_lmm_size > 0)
877                         ma->ma_need |= MA_LOV;
878         }
879
880         if (S_ISDIR(lu_object_attr(&next->mo_lu)) &&
881             reqbody->mbo_valid & OBD_MD_FLDIREA  &&
882             lustre_msg_get_opc(req->rq_reqmsg) == MDS_GETATTR) {
883                 /* get default stripe info for this dir. */
884                 ma->ma_need |= MA_LOV_DEF;
885         }
886         ma->ma_need |= ma_need;
887
888         rc = mdt_attr_get_complex(info, o, ma);
889         if (unlikely(rc)) {
890                 CERROR("%s: getattr error for "DFID": rc = %d\n",
891                        mdt_obd_name(info->mti_mdt),
892                        PFID(mdt_object_fid(o)), rc);
893                 RETURN(rc);
894         }
895
896         /* if file is released, check if a restore is running */
897         if (ma->ma_valid & MA_HSM) {
898                 repbody->mbo_valid |= OBD_MD_TSTATE;
899                 if ((ma->ma_hsm.mh_flags & HS_RELEASED) &&
900                     mdt_hsm_restore_is_running(info, mdt_object_fid(o)))
901                         repbody->mbo_t_state = MS_RESTORE;
902         }
903
904         is_root = lu_fid_eq(mdt_object_fid(o), &info->mti_mdt->mdt_md_root_fid);
905
906         /* the Lustre protocol supposes to return default striping
907          * on the user-visible root if explicitly requested */
908         if ((ma->ma_valid & MA_LOV) == 0 && S_ISDIR(la->la_mode) &&
909             (ma->ma_need & MA_LOV_DEF && is_root) && ma->ma_need & MA_LOV) {
910                 struct lu_fid      rootfid;
911                 struct mdt_object *root;
912                 struct mdt_device *mdt = info->mti_mdt;
913
914                 rc = dt_root_get(env, mdt->mdt_bottom, &rootfid);
915                 if (rc)
916                         RETURN(rc);
917                 root = mdt_object_find(env, mdt, &rootfid);
918                 if (IS_ERR(root))
919                         RETURN(PTR_ERR(root));
920                 rc = mdt_stripe_get(info, root, ma, XATTR_NAME_LOV);
921                 mdt_object_put(info->mti_env, root);
922                 if (unlikely(rc)) {
923                         CERROR("%s: getattr error for "DFID": rc = %d\n",
924                                mdt_obd_name(info->mti_mdt),
925                                PFID(mdt_object_fid(o)), rc);
926                         RETURN(rc);
927                 }
928         }
929
930         if (likely(ma->ma_valid & MA_INODE))
931                 mdt_pack_attr2body(info, repbody, la, mdt_object_fid(o));
932         else
933                 RETURN(-EFAULT);
934
935         if (mdt_body_has_lov(la, reqbody)) {
936                 if (ma->ma_valid & MA_LOV) {
937                         LASSERT(ma->ma_lmm_size);
938                         repbody->mbo_eadatasize = ma->ma_lmm_size;
939                         if (S_ISDIR(la->la_mode))
940                                 repbody->mbo_valid |= OBD_MD_FLDIREA;
941                         else
942                                 repbody->mbo_valid |= OBD_MD_FLEASIZE;
943                         mdt_dump_lmm(D_INFO, ma->ma_lmm, repbody->mbo_valid);
944                 }
945                 if (ma->ma_valid & MA_LMV) {
946                         /* Return -ENOTSUPP for old client */
947                         if (!mdt_is_striped_client(req->rq_export))
948                                 RETURN(-ENOTSUPP);
949
950                         LASSERT(S_ISDIR(la->la_mode));
951                         mdt_dump_lmv(D_INFO, ma->ma_lmv);
952                         repbody->mbo_eadatasize = ma->ma_lmv_size;
953                         repbody->mbo_valid |= (OBD_MD_FLDIREA|OBD_MD_MEA);
954                 }
955                 if (ma->ma_valid & MA_LMV_DEF) {
956                         /* Return -ENOTSUPP for old client */
957                         if (!mdt_is_striped_client(req->rq_export))
958                                 RETURN(-ENOTSUPP);
959                         LASSERT(S_ISDIR(la->la_mode));
960                         repbody->mbo_eadatasize = ma->ma_lmv_size;
961                         repbody->mbo_valid |= (OBD_MD_FLDIREA |
962                                                OBD_MD_DEFAULT_MEA);
963                 }
964         } else if (S_ISLNK(la->la_mode) &&
965                    reqbody->mbo_valid & OBD_MD_LINKNAME) {
966                 buffer->lb_buf = ma->ma_lmm;
967                 /* eadatasize from client includes NULL-terminator, so
968                  * there is no need to read it */
969                 buffer->lb_len = reqbody->mbo_eadatasize - 1;
970                 rc = mo_readlink(env, next, buffer);
971                 if (unlikely(rc <= 0)) {
972                         CERROR("%s: readlink failed for "DFID": rc = %d\n",
973                                mdt_obd_name(info->mti_mdt),
974                                PFID(mdt_object_fid(o)), rc);
975                         rc = -EFAULT;
976                 } else {
977                         int print_limit = min_t(int, PAGE_CACHE_SIZE - 128, rc);
978
979                         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_READLINK_EPROTO))
980                                 rc -= 2;
981                         repbody->mbo_valid |= OBD_MD_LINKNAME;
982                         /* we need to report back size with NULL-terminator
983                          * because client expects that */
984                         repbody->mbo_eadatasize = rc + 1;
985                         if (repbody->mbo_eadatasize != reqbody->mbo_eadatasize)
986                                 CDEBUG(D_INODE, "%s: Read shorter symlink %d "
987                                        "on "DFID ", expected %d\n",
988                                        mdt_obd_name(info->mti_mdt),
989                                        rc, PFID(mdt_object_fid(o)),
990                                        reqbody->mbo_eadatasize - 1);
991                         /* NULL terminate */
992                         ((char *)ma->ma_lmm)[rc] = 0;
993
994                         /* If the total CDEBUG() size is larger than a page, it
995                          * will print a warning to the console, avoid this by
996                          * printing just the last part of the symlink. */
997                         CDEBUG(D_INODE, "symlink dest %s%.*s, len = %d\n",
998                                print_limit < rc ? "..." : "", print_limit,
999                                (char *)ma->ma_lmm + rc - print_limit, rc);
1000                         rc = 0;
1001                 }
1002         }
1003
1004         if (reqbody->mbo_valid & OBD_MD_FLMODEASIZE) {
1005                 repbody->mbo_max_mdsize = info->mti_mdt->mdt_max_mdsize;
1006                 repbody->mbo_valid |= OBD_MD_FLMODEASIZE;
1007                 CDEBUG(D_INODE, "changing the max MD size to %u\n",
1008                        repbody->mbo_max_mdsize);
1009         }
1010
1011         if (exp_connect_rmtclient(info->mti_exp) &&
1012             reqbody->mbo_valid & OBD_MD_FLRMTPERM) {
1013                 void *buf = req_capsule_server_get(pill, &RMF_ACL);
1014
1015                 /* mdt_getattr_lock only */
1016                 rc = mdt_pack_remote_perm(info, o, buf);
1017                 if (rc) {
1018                         repbody->mbo_valid &= ~OBD_MD_FLRMTPERM;
1019                         repbody->mbo_aclsize = 0;
1020                         RETURN(rc);
1021                 } else {
1022                         repbody->mbo_valid |= OBD_MD_FLRMTPERM;
1023                         repbody->mbo_aclsize = sizeof(struct mdt_remote_perm);
1024                 }
1025         }
1026 #ifdef CONFIG_FS_POSIX_ACL
1027         else if ((exp_connect_flags(req->rq_export) & OBD_CONNECT_ACL) &&
1028                  (reqbody->mbo_valid & OBD_MD_FLACL))
1029                 rc = mdt_pack_acl2body(info, repbody, o, nodemap);
1030 #endif
1031
1032 out:
1033         if (rc == 0)
1034                 mdt_counter_incr(req, LPROC_MDT_GETATTR);
1035
1036         RETURN(rc);
1037 }
1038
1039 static int mdt_getattr(struct tgt_session_info *tsi)
1040 {
1041         struct mdt_thread_info  *info = tsi2mdt_info(tsi);
1042         struct mdt_object       *obj = info->mti_object;
1043         struct req_capsule      *pill = info->mti_pill;
1044         struct mdt_body         *reqbody;
1045         struct mdt_body         *repbody;
1046         int rc, rc2;
1047         ENTRY;
1048
1049         reqbody = req_capsule_client_get(pill, &RMF_MDT_BODY);
1050         LASSERT(reqbody);
1051         LASSERT(obj != NULL);
1052         LASSERT(lu_object_assert_exists(&obj->mot_obj));
1053
1054         /* Unlike intent case where we need to pre-fill out buffers early on
1055          * in intent policy for ldlm reasons, here we can have a much better
1056          * guess at EA size by just reading it from disk.
1057          * Exceptions are readdir and (missing) directory striping */
1058         /* Readlink */
1059         if (reqbody->mbo_valid & OBD_MD_LINKNAME) {
1060                 /* No easy way to know how long is the symlink, but it cannot
1061                  * be more than PATH_MAX, so we allocate +1 */
1062                 rc = PATH_MAX + 1;
1063
1064         /* A special case for fs ROOT: getattr there might fetch
1065          * default EA for entire fs, not just for this dir!
1066          */
1067         } else if (lu_fid_eq(mdt_object_fid(obj),
1068                              &info->mti_mdt->mdt_md_root_fid) &&
1069                    (reqbody->mbo_valid & OBD_MD_FLDIREA) &&
1070                    (lustre_msg_get_opc(mdt_info_req(info)->rq_reqmsg) ==
1071                                                                  MDS_GETATTR)) {
1072                 /* Should the default strping be bigger, mdt_fix_reply
1073                  * will reallocate */
1074                 rc = DEF_REP_MD_SIZE;
1075         } else {
1076                 /* Read the actual EA size from disk */
1077                 rc = mdt_attr_get_eabuf_size(info, obj);
1078         }
1079
1080         if (rc < 0)
1081                 GOTO(out, rc = err_serious(rc));
1082
1083         req_capsule_set_size(pill, &RMF_MDT_MD, RCL_SERVER, rc);
1084
1085         rc = req_capsule_server_pack(pill);
1086         if (unlikely(rc != 0))
1087                 GOTO(out, rc = err_serious(rc));
1088
1089         repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
1090         LASSERT(repbody != NULL);
1091         repbody->mbo_eadatasize = 0;
1092         repbody->mbo_aclsize = 0;
1093
1094         if (reqbody->mbo_valid & OBD_MD_FLRMTPERM)
1095                 rc = mdt_init_ucred(info, reqbody);
1096         else
1097                 rc = mdt_check_ucred(info);
1098         if (unlikely(rc))
1099                 GOTO(out_shrink, rc);
1100
1101         info->mti_cross_ref = !!(reqbody->mbo_valid & OBD_MD_FLCROSSREF);
1102
1103         rc = mdt_getattr_internal(info, obj, 0);
1104         if (reqbody->mbo_valid & OBD_MD_FLRMTPERM)
1105                 mdt_exit_ucred(info);
1106         EXIT;
1107 out_shrink:
1108         mdt_client_compatibility(info);
1109         rc2 = mdt_fix_reply(info);
1110         if (rc == 0)
1111                 rc = rc2;
1112 out:
1113         mdt_thread_info_fini(info);
1114         return rc;
1115 }
1116
1117 /**
1118  * Exchange MOF_LOV_CREATED flags between two objects after a
1119  * layout swap. No assumption is made on whether o1 or o2 have
1120  * created objects or not.
1121  *
1122  * \param[in,out] o1    First swap layout object
1123  * \param[in,out] o2    Second swap layout object
1124  */
1125 static void mdt_swap_lov_flag(struct mdt_object *o1, struct mdt_object *o2)
1126 {
1127         __u64   o1_flags;
1128
1129         mutex_lock(&o1->mot_lov_mutex);
1130         mutex_lock(&o2->mot_lov_mutex);
1131
1132         o1_flags = o1->mot_flags;
1133         o1->mot_flags = (o1->mot_flags & ~MOF_LOV_CREATED) |
1134                         (o2->mot_flags & MOF_LOV_CREATED);
1135
1136         o2->mot_flags = (o2->mot_flags & ~MOF_LOV_CREATED) |
1137                         (o1_flags & MOF_LOV_CREATED);
1138
1139         mutex_unlock(&o2->mot_lov_mutex);
1140         mutex_unlock(&o1->mot_lov_mutex);
1141 }
1142
1143 static int mdt_swap_layouts(struct tgt_session_info *tsi)
1144 {
1145         struct mdt_thread_info  *info;
1146         struct ptlrpc_request   *req = tgt_ses_req(tsi);
1147         struct obd_export       *exp = req->rq_export;
1148         struct mdt_object       *o1, *o2, *o;
1149         struct mdt_lock_handle  *lh1, *lh2;
1150         struct mdc_swap_layouts *msl;
1151         int                      rc;
1152         ENTRY;
1153
1154         /* client does not support layout lock, so layout swaping
1155          * is disabled.
1156          * FIXME: there is a problem for old clients which don't support
1157          * layout lock yet. If those clients have already opened the file
1158          * they won't be notified at all so that old layout may still be
1159          * used to do IO. This can be fixed after file release is landed by
1160          * doing exclusive open and taking full EX ibits lock. - Jinshan */
1161         if (!exp_connect_layout(exp))
1162                 RETURN(-EOPNOTSUPP);
1163
1164         info = tsi2mdt_info(tsi);
1165
1166         if (info->mti_dlm_req != NULL)
1167                 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
1168
1169         o1 = info->mti_object;
1170         o = o2 = mdt_object_find(info->mti_env, info->mti_mdt,
1171                                 &info->mti_body->mbo_fid2);
1172         if (IS_ERR(o))
1173                 GOTO(out, rc = PTR_ERR(o));
1174
1175         if (mdt_object_remote(o) || !mdt_object_exists(o)) /* remote object */
1176                 GOTO(put, rc = -ENOENT);
1177
1178         rc = lu_fid_cmp(&info->mti_body->mbo_fid1, &info->mti_body->mbo_fid2);
1179         if (unlikely(rc == 0)) /* same file, you kidding me? no-op. */
1180                 GOTO(put, rc);
1181
1182         if (rc < 0)
1183                 swap(o1, o2);
1184
1185         /* permission check. Make sure the calling process having permission
1186          * to write both files. */
1187         rc = mo_permission(info->mti_env, NULL, mdt_object_child(o1), NULL,
1188                                 MAY_WRITE);
1189         if (rc < 0)
1190                 GOTO(put, rc);
1191
1192         rc = mo_permission(info->mti_env, NULL, mdt_object_child(o2), NULL,
1193                                 MAY_WRITE);
1194         if (rc < 0)
1195                 GOTO(put, rc);
1196
1197         msl = req_capsule_client_get(info->mti_pill, &RMF_SWAP_LAYOUTS);
1198         if (msl == NULL)
1199                 GOTO(put, rc = -EPROTO);
1200
1201         lh1 = &info->mti_lh[MDT_LH_NEW];
1202         mdt_lock_reg_init(lh1, LCK_EX);
1203         lh2 = &info->mti_lh[MDT_LH_OLD];
1204         mdt_lock_reg_init(lh2, LCK_EX);
1205
1206         rc = mdt_object_lock(info, o1, lh1, MDS_INODELOCK_LAYOUT |
1207                              MDS_INODELOCK_XATTR);
1208         if (rc < 0)
1209                 GOTO(put, rc);
1210
1211         rc = mdt_object_lock(info, o2, lh2, MDS_INODELOCK_LAYOUT |
1212                              MDS_INODELOCK_XATTR);
1213         if (rc < 0)
1214                 GOTO(unlock1, rc);
1215
1216         rc = mo_swap_layouts(info->mti_env, mdt_object_child(o1),
1217                              mdt_object_child(o2), msl->msl_flags);
1218         if (rc < 0)
1219                 GOTO(unlock2, rc);
1220
1221         mdt_swap_lov_flag(o1, o2);
1222
1223 unlock2:
1224         mdt_object_unlock(info, o2, lh2, rc);
1225 unlock1:
1226         mdt_object_unlock(info, o1, lh1, rc);
1227 put:
1228         mdt_object_put(info->mti_env, o);
1229 out:
1230         mdt_thread_info_fini(info);
1231         RETURN(rc);
1232 }
1233
1234 static int mdt_raw_lookup(struct mdt_thread_info *info,
1235                           struct mdt_object *parent,
1236                           const struct lu_name *lname,
1237                           struct ldlm_reply *ldlm_rep)
1238 {
1239         struct lu_fid   *child_fid = &info->mti_tmp_fid1;
1240         int              rc;
1241         ENTRY;
1242
1243         LASSERT(!info->mti_cross_ref);
1244
1245         /* Only got the fid of this obj by name */
1246         fid_zero(child_fid);
1247         rc = mdo_lookup(info->mti_env, mdt_object_child(info->mti_object),
1248                         lname, child_fid, &info->mti_spec);
1249         if (rc == 0) {
1250                 struct mdt_body *repbody;
1251
1252                 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1253                 repbody->mbo_fid1 = *child_fid;
1254                 repbody->mbo_valid = OBD_MD_FLID;
1255                 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_POS);
1256         } else if (rc == -ENOENT) {
1257                 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_NEG);
1258         }
1259
1260         RETURN(rc);
1261 }
1262
1263 /*
1264  * UPDATE lock should be taken against parent, and be released before exit;
1265  * child_bits lock should be taken against child, and be returned back:
1266  *            (1)normal request should release the child lock;
1267  *            (2)intent request will grant the lock to client.
1268  */
1269 static int mdt_getattr_name_lock(struct mdt_thread_info *info,
1270                                  struct mdt_lock_handle *lhc,
1271                                  __u64 child_bits,
1272                                  struct ldlm_reply *ldlm_rep)
1273 {
1274         struct ptlrpc_request  *req       = mdt_info_req(info);
1275         struct mdt_body        *reqbody   = NULL;
1276         struct mdt_object      *parent    = info->mti_object;
1277         struct mdt_object      *child;
1278         struct lu_fid          *child_fid = &info->mti_tmp_fid1;
1279         struct lu_name         *lname     = NULL;
1280         struct mdt_lock_handle *lhp       = NULL;
1281         struct ldlm_lock       *lock;
1282         bool                    is_resent;
1283         bool                    try_layout;
1284         int                     ma_need = 0;
1285         int                     rc;
1286         ENTRY;
1287
1288         is_resent = lustre_handle_is_used(&lhc->mlh_reg_lh);
1289         LASSERT(ergo(is_resent,
1290                      lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT));
1291
1292         if (parent == NULL)
1293                 RETURN(-ENOENT);
1294
1295         if (info->mti_cross_ref) {
1296                 /* Only getattr on the child. Parent is on another node. */
1297                 mdt_set_disposition(info, ldlm_rep,
1298                                     DISP_LOOKUP_EXECD | DISP_LOOKUP_POS);
1299                 child = parent;
1300                 CDEBUG(D_INODE, "partial getattr_name child_fid = "DFID", "
1301                        "ldlm_rep = %p\n",
1302                        PFID(mdt_object_fid(child)), ldlm_rep);
1303
1304                 rc = mdt_check_resent_lock(info, child, lhc);
1305                 if (rc < 0) {
1306                         RETURN(rc);
1307                 } else if (rc > 0) {
1308                         mdt_lock_handle_init(lhc);
1309                         mdt_lock_reg_init(lhc, LCK_PR);
1310
1311                         /*
1312                          * Object's name is on another MDS, no lookup or layout
1313                          * lock is needed here but update lock is.
1314                          */
1315                         child_bits &= ~(MDS_INODELOCK_LOOKUP |
1316                                         MDS_INODELOCK_LAYOUT);
1317                         child_bits |= MDS_INODELOCK_PERM | MDS_INODELOCK_UPDATE;
1318
1319                         rc = mdt_object_lock(info, child, lhc, child_bits);
1320                         if (rc < 0)
1321                                 RETURN(rc);
1322                 }
1323
1324                 /* Finally, we can get attr for child. */
1325                 if (!mdt_object_exists(child)) {
1326                         LU_OBJECT_DEBUG(D_INFO, info->mti_env,
1327                                         &child->mot_obj,
1328                                         "remote object doesn't exist.\n");
1329                         mdt_object_unlock(info, child, lhc, 1);
1330                         RETURN(-ENOENT);
1331                 }
1332
1333                 rc = mdt_getattr_internal(info, child, 0);
1334                 if (unlikely(rc != 0))
1335                         mdt_object_unlock(info, child, lhc, 1);
1336
1337                 RETURN(rc);
1338         }
1339
1340         lname = &info->mti_name;
1341         mdt_name_unpack(info->mti_pill, &RMF_NAME, lname, MNF_FIX_ANON);
1342
1343         if (lu_name_is_valid(lname)) {
1344                 CDEBUG(D_INODE, "getattr with lock for "DFID"/"DNAME", "
1345                        "ldlm_rep = %p\n", PFID(mdt_object_fid(parent)),
1346                        PNAME(lname), ldlm_rep);
1347         } else {
1348                 reqbody = req_capsule_client_get(info->mti_pill, &RMF_MDT_BODY);
1349                 if (unlikely(reqbody == NULL))
1350                         RETURN(err_serious(-EPROTO));
1351
1352                 *child_fid = reqbody->mbo_fid2;
1353
1354                 if (unlikely(!fid_is_sane(child_fid)))
1355                         RETURN(err_serious(-EINVAL));
1356
1357                 CDEBUG(D_INODE, "getattr with lock for "DFID"/"DFID", "
1358                        "ldlm_rep = %p\n",
1359                        PFID(mdt_object_fid(parent)),
1360                        PFID(&reqbody->mbo_fid2), ldlm_rep);
1361         }
1362
1363         mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_EXECD);
1364
1365         if (unlikely(!mdt_object_exists(parent)) && lu_name_is_valid(lname)) {
1366                 LU_OBJECT_DEBUG(D_INODE, info->mti_env,
1367                                 &parent->mot_obj,
1368                                 "Parent doesn't exist!\n");
1369                 RETURN(-ESTALE);
1370         }
1371
1372         if (mdt_object_remote(parent)) {
1373                 CERROR("%s: parent "DFID" is on remote target\n",
1374                        mdt_obd_name(info->mti_mdt),
1375                        PFID(mdt_object_fid(parent)));
1376                 RETURN(-EIO);
1377         }
1378
1379         if (lu_name_is_valid(lname)) {
1380                 /* Always allow to lookup ".." */
1381                 if (unlikely(lname->ln_namelen == 2 &&
1382                              lname->ln_name[0] == '.' &&
1383                              lname->ln_name[1] == '.'))
1384                         info->mti_spec.sp_permitted = 1;
1385
1386                 if (info->mti_body->mbo_valid == OBD_MD_FLID) {
1387                         rc = mdt_raw_lookup(info, parent, lname, ldlm_rep);
1388
1389                         RETURN(rc);
1390                 }
1391
1392                 /* step 1: lock parent only if parent is a directory */
1393                 if (S_ISDIR(lu_object_attr(&parent->mot_obj))) {
1394                         lhp = &info->mti_lh[MDT_LH_PARENT];
1395                         mdt_lock_pdo_init(lhp, LCK_PR, lname);
1396                         rc = mdt_object_lock(info, parent, lhp,
1397                                              MDS_INODELOCK_UPDATE);
1398                         if (unlikely(rc != 0))
1399                                 RETURN(rc);
1400                 }
1401
1402                 /* step 2: lookup child's fid by name */
1403                 fid_zero(child_fid);
1404                 rc = mdo_lookup(info->mti_env, mdt_object_child(parent), lname,
1405                                 child_fid, &info->mti_spec);
1406                 if (rc == -ENOENT)
1407                         mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_NEG);
1408
1409                 if (rc != 0)
1410                         GOTO(out_parent, rc);
1411         }
1412
1413         mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_POS);
1414
1415         /*
1416          *step 3: find the child object by fid & lock it.
1417          *        regardless if it is local or remote.
1418          *
1419          *Note: LU-3240 (commit 762f2114d282a98ebfa4dbbeea9298a8088ad24e)
1420          *      set parent dir fid the same as child fid in getattr by fid case
1421          *      we should not lu_object_find() the object again, could lead
1422          *      to hung if there is a concurrent unlink destroyed the object.
1423          */
1424         if (lu_fid_eq(mdt_object_fid(parent), child_fid)) {
1425                 mdt_object_get(info->mti_env, parent);
1426                 child = parent;
1427         } else {
1428                 child = mdt_object_find(info->mti_env, info->mti_mdt,
1429                                         child_fid);
1430         }
1431
1432         if (unlikely(IS_ERR(child)))
1433                 GOTO(out_parent, rc = PTR_ERR(child));
1434
1435         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RESEND, obd_timeout * 2);
1436         if (!mdt_object_exists(child)) {
1437                 LU_OBJECT_DEBUG(D_INODE, info->mti_env,
1438                                 &child->mot_obj,
1439                                 "Object doesn't exist!\n");
1440                 GOTO(out_child, rc = -ENOENT);
1441         }
1442
1443         rc = mdt_check_resent_lock(info, child, lhc);
1444         if (rc < 0) {
1445                 GOTO(out_child, rc);
1446         } else if (rc > 0) {
1447                 mdt_lock_handle_init(lhc);
1448                 mdt_lock_reg_init(lhc, LCK_PR);
1449                 try_layout = false;
1450
1451                 if (!(child_bits & MDS_INODELOCK_UPDATE) &&
1452                       mdt_object_exists(child) && !mdt_object_remote(child)) {
1453                         struct md_attr *ma = &info->mti_attr;
1454
1455                         ma->ma_valid = 0;
1456                         ma->ma_need = MA_INODE;
1457                         rc = mdt_attr_get_complex(info, child, ma);
1458                         if (unlikely(rc != 0))
1459                                 GOTO(out_child, rc);
1460
1461                         /* If the file has not been changed for some time, we
1462                          * return not only a LOOKUP lock, but also an UPDATE
1463                          * lock and this might save us RPC on later STAT. For
1464                          * directories, it also let negative dentry cache start
1465                          * working for this dir. */
1466                         if (ma->ma_valid & MA_INODE &&
1467                             ma->ma_attr.la_valid & LA_CTIME &&
1468                             info->mti_mdt->mdt_namespace->ns_ctime_age_limit +
1469                                 ma->ma_attr.la_ctime < cfs_time_current_sec())
1470                                 child_bits |= MDS_INODELOCK_UPDATE;
1471                 }
1472
1473                 /* layout lock must be granted in a best-effort way
1474                  * for IT operations */
1475                 LASSERT(!(child_bits & MDS_INODELOCK_LAYOUT));
1476                 if (!OBD_FAIL_CHECK(OBD_FAIL_MDS_NO_LL_GETATTR) &&
1477                     exp_connect_layout(info->mti_exp) &&
1478                     S_ISREG(lu_object_attr(&child->mot_obj)) &&
1479                     !mdt_object_remote(child) && ldlm_rep != NULL) {
1480                         /* try to grant layout lock for regular file. */
1481                         try_layout = true;
1482                 }
1483
1484                 rc = 0;
1485                 if (try_layout) {
1486                         child_bits |= MDS_INODELOCK_LAYOUT;
1487                         /* try layout lock, it may fail to be granted due to
1488                          * contention at LOOKUP or UPDATE */
1489                         if (!mdt_object_lock_try(info, child, lhc,
1490                                                  child_bits)) {
1491                                 child_bits &= ~MDS_INODELOCK_LAYOUT;
1492                                 LASSERT(child_bits != 0);
1493                                 rc = mdt_object_lock(info, child, lhc,
1494                                                      child_bits);
1495                         } else {
1496                                 ma_need |= MA_LOV;
1497                         }
1498                 } else {
1499                         /* Do not enqueue the UPDATE lock from MDT(cross-MDT),
1500                          * client will enqueue the lock to the remote MDT */
1501                         if (mdt_object_remote(child))
1502                                 child_bits &= ~MDS_INODELOCK_UPDATE;
1503                         rc = mdt_object_lock(info, child, lhc, child_bits);
1504                 }
1505                 if (unlikely(rc != 0))
1506                         GOTO(out_child, rc);
1507         }
1508
1509         lock = ldlm_handle2lock(&lhc->mlh_reg_lh);
1510
1511         /* finally, we can get attr for child. */
1512         rc = mdt_getattr_internal(info, child, ma_need);
1513         if (unlikely(rc != 0)) {
1514                 mdt_object_unlock(info, child, lhc, 1);
1515         } else if (lock) {
1516                 /* Debugging code. */
1517                 LDLM_DEBUG(lock, "Returning lock to client");
1518                 LASSERTF(fid_res_name_eq(mdt_object_fid(child),
1519                                          &lock->l_resource->lr_name),
1520                          "Lock res_id: "DLDLMRES", fid: "DFID"\n",
1521                          PLDLMRES(lock->l_resource),
1522                          PFID(mdt_object_fid(child)));
1523         }
1524         if (lock)
1525                 LDLM_LOCK_PUT(lock);
1526
1527         EXIT;
1528 out_child:
1529         mdt_object_put(info->mti_env, child);
1530 out_parent:
1531         if (lhp)
1532                 mdt_object_unlock(info, parent, lhp, 1);
1533         return rc;
1534 }
1535
1536 /* normal handler: should release the child lock */
1537 static int mdt_getattr_name(struct tgt_session_info *tsi)
1538 {
1539         struct mdt_thread_info  *info = tsi2mdt_info(tsi);
1540         struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_CHILD];
1541         struct mdt_body        *reqbody;
1542         struct mdt_body        *repbody;
1543         int rc, rc2;
1544         ENTRY;
1545
1546         reqbody = req_capsule_client_get(info->mti_pill, &RMF_MDT_BODY);
1547         LASSERT(reqbody != NULL);
1548         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1549         LASSERT(repbody != NULL);
1550
1551         info->mti_cross_ref = !!(reqbody->mbo_valid & OBD_MD_FLCROSSREF);
1552         repbody->mbo_eadatasize = 0;
1553         repbody->mbo_aclsize = 0;
1554
1555         rc = mdt_init_ucred_intent_getattr(info, reqbody);
1556         if (unlikely(rc))
1557                 GOTO(out_shrink, rc);
1558
1559         rc = mdt_getattr_name_lock(info, lhc, MDS_INODELOCK_UPDATE, NULL);
1560         if (lustre_handle_is_used(&lhc->mlh_reg_lh)) {
1561                 ldlm_lock_decref(&lhc->mlh_reg_lh, lhc->mlh_reg_mode);
1562                 lhc->mlh_reg_lh.cookie = 0;
1563         }
1564         mdt_exit_ucred(info);
1565         EXIT;
1566 out_shrink:
1567         mdt_client_compatibility(info);
1568         rc2 = mdt_fix_reply(info);
1569         if (rc == 0)
1570                 rc = rc2;
1571         mdt_thread_info_fini(info);
1572         return rc;
1573 }
1574
1575 static int mdt_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
1576                          void *karg, void *uarg);
1577
1578 static int mdt_set_info(struct tgt_session_info *tsi)
1579 {
1580         struct ptlrpc_request   *req = tgt_ses_req(tsi);
1581         char                    *key;
1582         void                    *val;
1583         int                      keylen, vallen, rc = 0;
1584
1585         ENTRY;
1586
1587         key = req_capsule_client_get(tsi->tsi_pill, &RMF_SETINFO_KEY);
1588         if (key == NULL) {
1589                 DEBUG_REQ(D_HA, req, "no set_info key");
1590                 RETURN(err_serious(-EFAULT));
1591         }
1592
1593         keylen = req_capsule_get_size(tsi->tsi_pill, &RMF_SETINFO_KEY,
1594                                       RCL_CLIENT);
1595
1596         val = req_capsule_client_get(tsi->tsi_pill, &RMF_SETINFO_VAL);
1597         if (val == NULL) {
1598                 DEBUG_REQ(D_HA, req, "no set_info val");
1599                 RETURN(err_serious(-EFAULT));
1600         }
1601
1602         vallen = req_capsule_get_size(tsi->tsi_pill, &RMF_SETINFO_VAL,
1603                                       RCL_CLIENT);
1604
1605         /* Swab any part of val you need to here */
1606         if (KEY_IS(KEY_READ_ONLY)) {
1607                 spin_lock(&req->rq_export->exp_lock);
1608                 if (*(__u32 *)val)
1609                         *exp_connect_flags_ptr(req->rq_export) |=
1610                                 OBD_CONNECT_RDONLY;
1611                 else
1612                         *exp_connect_flags_ptr(req->rq_export) &=
1613                                 ~OBD_CONNECT_RDONLY;
1614                 spin_unlock(&req->rq_export->exp_lock);
1615         } else if (KEY_IS(KEY_CHANGELOG_CLEAR)) {
1616                 struct changelog_setinfo *cs = val;
1617
1618                 if (vallen != sizeof(*cs)) {
1619                         CERROR("%s: bad changelog_clear setinfo size %d\n",
1620                                tgt_name(tsi->tsi_tgt), vallen);
1621                         RETURN(-EINVAL);
1622                 }
1623                 if (ptlrpc_req_need_swab(req)) {
1624                         __swab64s(&cs->cs_recno);
1625                         __swab32s(&cs->cs_id);
1626                 }
1627
1628                 rc = mdt_iocontrol(OBD_IOC_CHANGELOG_CLEAR, req->rq_export,
1629                                    vallen, val, NULL);
1630         } else if (KEY_IS(KEY_EVICT_BY_NID)) {
1631                 if (vallen > 0)
1632                         obd_export_evict_by_nid(req->rq_export->exp_obd, val);
1633         } else {
1634                 RETURN(-EINVAL);
1635         }
1636         RETURN(rc);
1637 }
1638
1639 static int mdt_readpage(struct tgt_session_info *tsi)
1640 {
1641         struct mdt_thread_info  *info = mdt_th_info(tsi->tsi_env);
1642         struct mdt_object       *object = mdt_obj(tsi->tsi_corpus);
1643         struct lu_rdpg          *rdpg = &info->mti_u.rdpg.mti_rdpg;
1644         const struct mdt_body   *reqbody = tsi->tsi_mdt_body;
1645         struct mdt_body         *repbody;
1646         int                      rc;
1647         int                      i;
1648
1649         ENTRY;
1650
1651         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_READPAGE_PACK))
1652                 RETURN(err_serious(-ENOMEM));
1653
1654         repbody = req_capsule_server_get(tsi->tsi_pill, &RMF_MDT_BODY);
1655         if (repbody == NULL || reqbody == NULL)
1656                 RETURN(err_serious(-EFAULT));
1657
1658         /*
1659          * prepare @rdpg before calling lower layers and transfer itself. Here
1660          * reqbody->size contains offset of where to start to read and
1661          * reqbody->nlink contains number bytes to read.
1662          */
1663         rdpg->rp_hash = reqbody->mbo_size;
1664         if (rdpg->rp_hash != reqbody->mbo_size) {
1665                 CERROR("Invalid hash: "LPX64" != "LPX64"\n",
1666                        rdpg->rp_hash, reqbody->mbo_size);
1667                 RETURN(-EFAULT);
1668         }
1669
1670         rdpg->rp_attrs = reqbody->mbo_mode;
1671         if (exp_connect_flags(tsi->tsi_exp) & OBD_CONNECT_64BITHASH)
1672                 rdpg->rp_attrs |= LUDA_64BITHASH;
1673         rdpg->rp_count  = min_t(unsigned int, reqbody->mbo_nlink,
1674                                 exp_max_brw_size(tsi->tsi_exp));
1675         rdpg->rp_npages = (rdpg->rp_count + PAGE_CACHE_SIZE - 1) >>
1676                           PAGE_CACHE_SHIFT;
1677         OBD_ALLOC(rdpg->rp_pages, rdpg->rp_npages * sizeof rdpg->rp_pages[0]);
1678         if (rdpg->rp_pages == NULL)
1679                 RETURN(-ENOMEM);
1680
1681         for (i = 0; i < rdpg->rp_npages; ++i) {
1682                 rdpg->rp_pages[i] = alloc_page(GFP_IOFS);
1683                 if (rdpg->rp_pages[i] == NULL)
1684                         GOTO(free_rdpg, rc = -ENOMEM);
1685         }
1686
1687         /* call lower layers to fill allocated pages with directory data */
1688         rc = mo_readpage(tsi->tsi_env, mdt_object_child(object), rdpg);
1689         if (rc < 0)
1690                 GOTO(free_rdpg, rc);
1691
1692         /* send pages to client */
1693         rc = tgt_sendpage(tsi, rdpg, rc);
1694
1695         EXIT;
1696 free_rdpg:
1697
1698         for (i = 0; i < rdpg->rp_npages; i++)
1699                 if (rdpg->rp_pages[i] != NULL)
1700                         __free_page(rdpg->rp_pages[i]);
1701         OBD_FREE(rdpg->rp_pages, rdpg->rp_npages * sizeof rdpg->rp_pages[0]);
1702
1703         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE))
1704                 RETURN(0);
1705
1706         return rc;
1707 }
1708
1709 static int mdt_reint_internal(struct mdt_thread_info *info,
1710                               struct mdt_lock_handle *lhc,
1711                               __u32 op)
1712 {
1713         struct req_capsule      *pill = info->mti_pill;
1714         struct mdt_body         *repbody;
1715         int                      rc = 0, rc2;
1716
1717         ENTRY;
1718
1719         rc = mdt_reint_unpack(info, op);
1720         if (rc != 0) {
1721                 CERROR("Can't unpack reint, rc %d\n", rc);
1722                 RETURN(err_serious(rc));
1723         }
1724
1725         /* for replay (no_create) lmm is not needed, client has it already */
1726         if (req_capsule_has_field(pill, &RMF_MDT_MD, RCL_SERVER))
1727                 req_capsule_set_size(pill, &RMF_MDT_MD, RCL_SERVER,
1728                                      DEF_REP_MD_SIZE);
1729
1730         /* llog cookies are always 0, the field is kept for compatibility */
1731         if (req_capsule_has_field(pill, &RMF_LOGCOOKIES, RCL_SERVER))
1732                 req_capsule_set_size(pill, &RMF_LOGCOOKIES, RCL_SERVER, 0);
1733
1734         rc = req_capsule_server_pack(pill);
1735         if (rc != 0) {
1736                 CERROR("Can't pack response, rc %d\n", rc);
1737                 RETURN(err_serious(rc));
1738         }
1739
1740         if (req_capsule_has_field(pill, &RMF_MDT_BODY, RCL_SERVER)) {
1741                 repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
1742                 LASSERT(repbody);
1743                 repbody->mbo_eadatasize = 0;
1744                 repbody->mbo_aclsize = 0;
1745         }
1746
1747         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_REINT_DELAY, 10);
1748
1749         /* for replay no cookkie / lmm need, because client have this already */
1750         if (info->mti_spec.no_create)
1751                 if (req_capsule_has_field(pill, &RMF_MDT_MD, RCL_SERVER))
1752                         req_capsule_set_size(pill, &RMF_MDT_MD, RCL_SERVER, 0);
1753
1754         rc = mdt_init_ucred_reint(info);
1755         if (rc)
1756                 GOTO(out_shrink, rc);
1757
1758         rc = mdt_fix_attr_ucred(info, op);
1759         if (rc != 0)
1760                 GOTO(out_ucred, rc = err_serious(rc));
1761
1762         rc = mdt_check_resent(info, mdt_reconstruct, lhc);
1763         if (rc < 0) {
1764                 GOTO(out_ucred, rc);
1765         } else if (rc == 1) {
1766                 DEBUG_REQ(D_INODE, mdt_info_req(info), "resent opt.");
1767                 rc = lustre_msg_get_status(mdt_info_req(info)->rq_repmsg);
1768                 GOTO(out_ucred, rc);
1769         }
1770         rc = mdt_reint_rec(info, lhc);
1771         EXIT;
1772 out_ucred:
1773         mdt_exit_ucred(info);
1774 out_shrink:
1775         mdt_client_compatibility(info);
1776         rc2 = mdt_fix_reply(info);
1777         if (rc == 0)
1778                 rc = rc2;
1779         return rc;
1780 }
1781
1782 static long mdt_reint_opcode(struct ptlrpc_request *req,
1783                              const struct req_format **fmt)
1784 {
1785         struct mdt_device       *mdt;
1786         struct mdt_rec_reint    *rec;
1787         long                     opc;
1788
1789         rec = req_capsule_client_get(&req->rq_pill, &RMF_REC_REINT);
1790         if (rec != NULL) {
1791                 opc = rec->rr_opcode;
1792                 DEBUG_REQ(D_INODE, req, "reint opt = %ld", opc);
1793                 if (opc < REINT_MAX && fmt[opc] != NULL)
1794                         req_capsule_extend(&req->rq_pill, fmt[opc]);
1795                 else {
1796                         mdt = mdt_exp2dev(req->rq_export);
1797                         CERROR("%s: Unsupported opcode '%ld' from client '%s':"
1798                                " rc = %d\n", req->rq_export->exp_obd->obd_name,
1799                                opc, mdt->mdt_ldlm_client->cli_name, -EFAULT);
1800                         opc = err_serious(-EFAULT);
1801                 }
1802         } else {
1803                 opc = err_serious(-EFAULT);
1804         }
1805         return opc;
1806 }
1807
1808 static int mdt_reint(struct tgt_session_info *tsi)
1809 {
1810         long opc;
1811         int  rc;
1812         static const struct req_format *reint_fmts[REINT_MAX] = {
1813                 [REINT_SETATTR]  = &RQF_MDS_REINT_SETATTR,
1814                 [REINT_CREATE]   = &RQF_MDS_REINT_CREATE,
1815                 [REINT_LINK]     = &RQF_MDS_REINT_LINK,
1816                 [REINT_UNLINK]   = &RQF_MDS_REINT_UNLINK,
1817                 [REINT_RENAME]   = &RQF_MDS_REINT_RENAME,
1818                 [REINT_OPEN]     = &RQF_MDS_REINT_OPEN,
1819                 [REINT_SETXATTR] = &RQF_MDS_REINT_SETXATTR,
1820                 [REINT_RMENTRY]  = &RQF_MDS_REINT_UNLINK,
1821                 [REINT_MIGRATE]  = &RQF_MDS_REINT_RENAME
1822         };
1823
1824         ENTRY;
1825
1826         opc = mdt_reint_opcode(tgt_ses_req(tsi), reint_fmts);
1827         if (opc >= 0) {
1828                 struct mdt_thread_info *info = tsi2mdt_info(tsi);
1829                 /*
1830                  * No lock possible here from client to pass it to reint code
1831                  * path.
1832                  */
1833                 rc = mdt_reint_internal(info, NULL, opc);
1834                 mdt_thread_info_fini(info);
1835         } else {
1836                 rc = opc;
1837         }
1838
1839         tsi->tsi_reply_fail_id = OBD_FAIL_MDS_REINT_NET_REP;
1840         RETURN(rc);
1841 }
1842
1843 /* this should sync the whole device */
1844 static int mdt_device_sync(const struct lu_env *env, struct mdt_device *mdt)
1845 {
1846         struct dt_device *dt = mdt->mdt_bottom;
1847         int rc;
1848         ENTRY;
1849
1850         rc = dt->dd_ops->dt_sync(env, dt);
1851         RETURN(rc);
1852 }
1853
1854 /* this should sync this object */
1855 static int mdt_object_sync(struct mdt_thread_info *info)
1856 {
1857         struct md_object *next;
1858         int rc;
1859         ENTRY;
1860
1861         if (!mdt_object_exists(info->mti_object)) {
1862                 CWARN("%s: non existing object "DFID": rc = %d\n",
1863                       mdt_obd_name(info->mti_mdt),
1864                       PFID(mdt_object_fid(info->mti_object)), -ESTALE);
1865                 RETURN(-ESTALE);
1866         }
1867         next = mdt_object_child(info->mti_object);
1868         rc = mo_object_sync(info->mti_env, next);
1869
1870         RETURN(rc);
1871 }
1872
1873 static int mdt_sync(struct tgt_session_info *tsi)
1874 {
1875         struct ptlrpc_request   *req = tgt_ses_req(tsi);
1876         struct req_capsule      *pill = tsi->tsi_pill;
1877         struct mdt_body         *body;
1878         int                      rc;
1879
1880         ENTRY;
1881
1882         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SYNC_PACK))
1883                 RETURN(err_serious(-ENOMEM));
1884
1885         if (fid_seq(&tsi->tsi_mdt_body->mbo_fid1) == 0) {
1886                 rc = mdt_device_sync(tsi->tsi_env, mdt_exp2dev(tsi->tsi_exp));
1887         } else {
1888                 struct mdt_thread_info *info = tsi2mdt_info(tsi);
1889
1890                 /* sync an object */
1891                 rc = mdt_object_sync(info);
1892                 if (rc == 0) {
1893                         const struct lu_fid *fid;
1894                         struct lu_attr *la = &info->mti_attr.ma_attr;
1895
1896                         info->mti_attr.ma_need = MA_INODE;
1897                         info->mti_attr.ma_valid = 0;
1898                         rc = mdt_attr_get_complex(info, info->mti_object,
1899                                                   &info->mti_attr);
1900                         if (rc == 0) {
1901                                 body = req_capsule_server_get(pill,
1902                                                               &RMF_MDT_BODY);
1903                                 fid = mdt_object_fid(info->mti_object);
1904                                 mdt_pack_attr2body(info, body, la, fid);
1905                         }
1906                 }
1907                 mdt_thread_info_fini(info);
1908         }
1909         if (rc == 0)
1910                 mdt_counter_incr(req, LPROC_MDT_SYNC);
1911
1912         RETURN(rc);
1913 }
1914
1915 /*
1916  * Handle quota control requests to consult current usage/limit, but also
1917  * to configure quota enforcement
1918  */
1919 static int mdt_quotactl(struct tgt_session_info *tsi)
1920 {
1921         struct obd_export       *exp  = tsi->tsi_exp;
1922         struct req_capsule      *pill = tsi->tsi_pill;
1923         struct obd_quotactl     *oqctl, *repoqc;
1924         int                      id, rc;
1925         struct mdt_device       *mdt = mdt_exp2dev(exp);
1926         struct lu_device        *qmt = mdt->mdt_qmt_dev;
1927         struct lu_nodemap       *nodemap = exp->exp_target_data.ted_nodemap;
1928         ENTRY;
1929
1930         oqctl = req_capsule_client_get(pill, &RMF_OBD_QUOTACTL);
1931         if (oqctl == NULL)
1932                 RETURN(err_serious(-EPROTO));
1933
1934         rc = req_capsule_server_pack(pill);
1935         if (rc)
1936                 RETURN(err_serious(rc));
1937
1938         switch (oqctl->qc_cmd) {
1939                 /* master quotactl */
1940         case Q_SETINFO:
1941         case Q_SETQUOTA:
1942                 if (!nodemap_can_setquota(nodemap))
1943                         RETURN(-EPERM);
1944         case Q_GETINFO:
1945         case Q_GETQUOTA:
1946                 if (qmt == NULL)
1947                         RETURN(-EOPNOTSUPP);
1948                 /* slave quotactl */
1949         case Q_GETOINFO:
1950         case Q_GETOQUOTA:
1951                 break;
1952         default:
1953                 CERROR("Unsupported quotactl command: %d\n", oqctl->qc_cmd);
1954                 RETURN(-EFAULT);
1955         }
1956
1957         /* map uid/gid for remote client */
1958         id = oqctl->qc_id;
1959         if (exp_connect_rmtclient(exp)) {
1960                 struct lustre_idmap_table *idmap;
1961
1962                 idmap = exp->exp_mdt_data.med_idmap;
1963
1964                 if (unlikely(oqctl->qc_cmd != Q_GETQUOTA &&
1965                              oqctl->qc_cmd != Q_GETINFO))
1966                         RETURN(-EPERM);
1967
1968                 if (oqctl->qc_type == USRQUOTA)
1969                         id = lustre_idmap_lookup_uid(NULL, idmap, 0,
1970                                                      oqctl->qc_id);
1971                 else if (oqctl->qc_type == GRPQUOTA)
1972                         id = lustre_idmap_lookup_gid(NULL, idmap, 0,
1973                                                      oqctl->qc_id);
1974                 else
1975                         RETURN(-EINVAL);
1976
1977                 if (id == CFS_IDMAP_NOTFOUND) {
1978                         CDEBUG(D_QUOTA, "no mapping for id %u\n", oqctl->qc_id);
1979                         RETURN(-EACCES);
1980                 }
1981         }
1982
1983         if (oqctl->qc_type == USRQUOTA)
1984                 id = nodemap_map_id(nodemap, NODEMAP_UID,
1985                                     NODEMAP_CLIENT_TO_FS, id);
1986         else if (oqctl->qc_type == GRPQUOTA)
1987                 id = nodemap_map_id(nodemap, NODEMAP_UID,
1988                                     NODEMAP_CLIENT_TO_FS, id);
1989
1990         repoqc = req_capsule_server_get(pill, &RMF_OBD_QUOTACTL);
1991         if (repoqc == NULL)
1992                 RETURN(err_serious(-EFAULT));
1993
1994         if (oqctl->qc_id != id)
1995                 swap(oqctl->qc_id, id);
1996
1997         switch (oqctl->qc_cmd) {
1998
1999         case Q_GETINFO:
2000         case Q_SETINFO:
2001         case Q_SETQUOTA:
2002         case Q_GETQUOTA:
2003                 /* forward quotactl request to QMT */
2004                 rc = qmt_hdls.qmth_quotactl(tsi->tsi_env, qmt, oqctl);
2005                 break;
2006
2007         case Q_GETOINFO:
2008         case Q_GETOQUOTA:
2009                 /* slave quotactl */
2010                 rc = lquotactl_slv(tsi->tsi_env, tsi->tsi_tgt->lut_bottom,
2011                                    oqctl);
2012                 break;
2013
2014         default:
2015                 CERROR("Unsupported quotactl command: %d\n", oqctl->qc_cmd);
2016                 RETURN(-EFAULT);
2017         }
2018
2019         if (oqctl->qc_id != id)
2020                 swap(oqctl->qc_id, id);
2021
2022         *repoqc = *oqctl;
2023         RETURN(rc);
2024 }
2025
2026 /** clone llog ctxt from child (mdd)
2027  * This allows remote llog (replicator) access.
2028  * We can either pass all llog RPCs (eg mdt_llog_create) on to child where the
2029  * context was originally set up, or we can handle them directly.
2030  * I choose the latter, but that means I need any llog
2031  * contexts set up by child to be accessable by the mdt.  So we clone the
2032  * context into our context list here.
2033  */
2034 static int mdt_llog_ctxt_clone(const struct lu_env *env, struct mdt_device *mdt,
2035                                int idx)
2036 {
2037         struct md_device  *next = mdt->mdt_child;
2038         struct llog_ctxt *ctxt;
2039         int rc;
2040
2041         if (!llog_ctxt_null(mdt2obd_dev(mdt), idx))
2042                 return 0;
2043
2044         rc = next->md_ops->mdo_llog_ctxt_get(env, next, idx, (void **)&ctxt);
2045         if (rc || ctxt == NULL) {
2046                 return 0;
2047         }
2048
2049         rc = llog_group_set_ctxt(&mdt2obd_dev(mdt)->obd_olg, ctxt, idx);
2050         if (rc)
2051                 CERROR("Can't set mdt ctxt %d\n", rc);
2052
2053         return rc;
2054 }
2055
2056 static int mdt_llog_ctxt_unclone(const struct lu_env *env,
2057                                  struct mdt_device *mdt, int idx)
2058 {
2059         struct llog_ctxt *ctxt;
2060
2061         ctxt = llog_get_context(mdt2obd_dev(mdt), idx);
2062         if (ctxt == NULL)
2063                 return 0;
2064         /* Put once for the get we just did, and once for the clone */
2065         llog_ctxt_put(ctxt);
2066         llog_ctxt_put(ctxt);
2067         return 0;
2068 }
2069
2070 /*
2071  * sec context handlers
2072  */
2073 static int mdt_sec_ctx_handle(struct tgt_session_info *tsi)
2074 {
2075         int rc;
2076
2077         rc = mdt_handle_idmap(tsi);
2078         if (unlikely(rc)) {
2079                 struct ptlrpc_request   *req = tgt_ses_req(tsi);
2080                 __u32                    opc;
2081
2082                 opc = lustre_msg_get_opc(req->rq_reqmsg);
2083                 if (opc == SEC_CTX_INIT || opc == SEC_CTX_INIT_CONT)
2084                         sptlrpc_svc_ctx_invalidate(req);
2085         }
2086
2087         CFS_FAIL_TIMEOUT(OBD_FAIL_SEC_CTX_HDL_PAUSE, cfs_fail_val);
2088
2089         return rc;
2090 }
2091
2092 /*
2093  * quota request handlers
2094  */
2095 static int mdt_quota_dqacq(struct tgt_session_info *tsi)
2096 {
2097         struct mdt_device       *mdt = mdt_exp2dev(tsi->tsi_exp);
2098         struct lu_device        *qmt = mdt->mdt_qmt_dev;
2099         int                      rc;
2100         ENTRY;
2101
2102         if (qmt == NULL)
2103                 RETURN(err_serious(-EOPNOTSUPP));
2104
2105         rc = qmt_hdls.qmth_dqacq(tsi->tsi_env, qmt, tgt_ses_req(tsi));
2106         RETURN(rc);
2107 }
2108
2109 struct mdt_object *mdt_object_new(const struct lu_env *env,
2110                                   struct mdt_device *d,
2111                                   const struct lu_fid *f)
2112 {
2113         struct lu_object_conf conf = { .loc_flags = LOC_F_NEW };
2114         struct lu_object *o;
2115         struct mdt_object *m;
2116         ENTRY;
2117
2118         CDEBUG(D_INFO, "Allocate object for "DFID"\n", PFID(f));
2119         o = lu_object_find(env, &d->mdt_lu_dev, f, &conf);
2120         if (unlikely(IS_ERR(o)))
2121                 m = (struct mdt_object *)o;
2122         else
2123                 m = mdt_obj(o);
2124         RETURN(m);
2125 }
2126
2127 struct mdt_object *mdt_object_find(const struct lu_env *env,
2128                                    struct mdt_device *d,
2129                                    const struct lu_fid *f)
2130 {
2131         struct lu_object *o;
2132         struct mdt_object *m;
2133         ENTRY;
2134
2135         CDEBUG(D_INFO, "Find object for "DFID"\n", PFID(f));
2136         o = lu_object_find(env, &d->mdt_lu_dev, f, NULL);
2137         if (unlikely(IS_ERR(o)))
2138                 m = (struct mdt_object *)o;
2139         else
2140                 m = mdt_obj(o);
2141
2142         RETURN(m);
2143 }
2144
2145 /**
2146  * Asyncronous commit for mdt device.
2147  *
2148  * Pass asynchonous commit call down the MDS stack.
2149  *
2150  * \param env environment
2151  * \param mdt the mdt device
2152  */
2153 static void mdt_device_commit_async(const struct lu_env *env,
2154                                     struct mdt_device *mdt)
2155 {
2156         struct dt_device *dt = mdt->mdt_bottom;
2157         int rc;
2158
2159         rc = dt->dd_ops->dt_commit_async(env, dt);
2160         if (unlikely(rc != 0))
2161                 CWARN("%s: async commit start failed: rc = %d\n",
2162                       mdt_obd_name(mdt), rc);
2163 }
2164
2165 /**
2166  * Mark the lock as "synchonous".
2167  *
2168  * Mark the lock to deffer transaction commit to the unlock time.
2169  *
2170  * \param lock the lock to mark as "synchonous"
2171  *
2172  * \see mdt_is_lock_sync
2173  * \see mdt_save_lock
2174  */
2175 static inline void mdt_set_lock_sync(struct ldlm_lock *lock)
2176 {
2177         lock->l_ast_data = (void*)1;
2178 }
2179
2180 /**
2181  * Check whehter the lock "synchonous" or not.
2182  *
2183  * \param lock the lock to check
2184  * \retval 1 the lock is "synchonous"
2185  * \retval 0 the lock isn't "synchronous"
2186  *
2187  * \see mdt_set_lock_sync
2188  * \see mdt_save_lock
2189  */
2190 static inline int mdt_is_lock_sync(struct ldlm_lock *lock)
2191 {
2192         return lock->l_ast_data != NULL;
2193 }
2194
2195 /**
2196  * Blocking AST for mdt locks.
2197  *
2198  * Starts transaction commit if in case of COS lock conflict or
2199  * deffers such a commit to the mdt_save_lock.
2200  *
2201  * \param lock the lock which blocks a request or cancelling lock
2202  * \param desc unused
2203  * \param data unused
2204  * \param flag indicates whether this cancelling or blocking callback
2205  * \retval 0
2206  * \see ldlm_blocking_ast_nocheck
2207  */
2208 int mdt_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
2209                      void *data, int flag)
2210 {
2211         struct obd_device *obd = ldlm_lock_to_ns(lock)->ns_obd;
2212         struct mdt_device *mdt = mdt_dev(obd->obd_lu_dev);
2213         int rc;
2214         ENTRY;
2215
2216         if (flag == LDLM_CB_CANCELING)
2217                 RETURN(0);
2218         lock_res_and_lock(lock);
2219         if (lock->l_blocking_ast != mdt_blocking_ast) {
2220                 unlock_res_and_lock(lock);
2221                 RETURN(0);
2222         }
2223         if (mdt_cos_is_enabled(mdt) &&
2224             lock->l_req_mode & (LCK_PW | LCK_EX) &&
2225             lock->l_blocking_lock != NULL &&
2226             lock->l_client_cookie != lock->l_blocking_lock->l_client_cookie) {
2227                 mdt_set_lock_sync(lock);
2228         }
2229         rc = ldlm_blocking_ast_nocheck(lock);
2230
2231         /* There is no lock conflict if l_blocking_lock == NULL,
2232          * it indicates a blocking ast sent from ldlm_lock_decref_internal
2233          * when the last reference to a local lock was released */
2234         if (lock->l_req_mode == LCK_COS && lock->l_blocking_lock != NULL) {
2235                 struct lu_env env;
2236
2237                 rc = lu_env_init(&env, LCT_LOCAL);
2238                 if (unlikely(rc != 0))
2239                         CWARN("%s: lu_env initialization failed, cannot "
2240                               "start asynchronous commit: rc = %d\n",
2241                               obd->obd_name, rc);
2242                 else
2243                         mdt_device_commit_async(&env, mdt);
2244                 lu_env_fini(&env);
2245         }
2246         RETURN(rc);
2247 }
2248
2249 /* Used for cross-MDT lock */
2250 int mdt_remote_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
2251                             void *data, int flag)
2252 {
2253         struct lustre_handle lockh;
2254         int               rc;
2255
2256         switch (flag) {
2257         case LDLM_CB_BLOCKING:
2258                 ldlm_lock2handle(lock, &lockh);
2259                 rc = ldlm_cli_cancel(&lockh,
2260                         ldlm_is_atomic_cb(lock) ? 0 : LCF_ASYNC);
2261                 if (rc < 0) {
2262                         CDEBUG(D_INODE, "ldlm_cli_cancel: %d\n", rc);
2263                         RETURN(rc);
2264                 }
2265                 break;
2266         case LDLM_CB_CANCELING:
2267                 LDLM_DEBUG(lock, "Revoke remote lock\n");
2268                 break;
2269         default:
2270                 LBUG();
2271         }
2272         RETURN(0);
2273 }
2274
2275 int mdt_check_resent_lock(struct mdt_thread_info *info,
2276                           struct mdt_object *mo,
2277                           struct mdt_lock_handle *lhc)
2278 {
2279         /* the lock might already be gotten in ldlm_handle_enqueue() */
2280         if (unlikely(lustre_handle_is_used(&lhc->mlh_reg_lh))) {
2281                 struct ptlrpc_request *req = mdt_info_req(info);
2282                 struct ldlm_lock      *lock;
2283
2284                 lock = ldlm_handle2lock(&lhc->mlh_reg_lh);
2285                 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT);
2286                 if (lock == NULL) {
2287                         /* Lock is pinned by ldlm_handle_enqueue0() as it is
2288                          * a resend case, however, it could be already destroyed
2289                          * due to client eviction or a raced cancel RPC. */
2290                         LDLM_DEBUG_NOLOCK("Invalid lock handle "LPX64"\n",
2291                                           lhc->mlh_reg_lh.cookie);
2292                         RETURN(-ESTALE);
2293                 }
2294
2295                 if (!fid_res_name_eq(mdt_object_fid(mo),
2296                                      &lock->l_resource->lr_name)) {
2297                         CWARN("%s: Although resent, but still not "
2298                               "get child lock:"DFID"\n",
2299                               info->mti_exp->exp_obd->obd_name,
2300                               PFID(mdt_object_fid(mo)));
2301                         LDLM_LOCK_PUT(lock);
2302                         RETURN(-EPROTO);
2303                 }
2304                 LDLM_LOCK_PUT(lock);
2305                 return 0;
2306         }
2307         return 1;
2308 }
2309
2310 int mdt_remote_object_lock(struct mdt_thread_info *mti, struct mdt_object *o,
2311                            const struct lu_fid *fid, struct lustre_handle *lh,
2312                            enum ldlm_mode mode, __u64 ibits, bool nonblock)
2313 {
2314         struct ldlm_enqueue_info *einfo = &mti->mti_einfo;
2315         union ldlm_policy_data *policy = &mti->mti_policy;
2316         struct ldlm_res_id *res_id = &mti->mti_res_id;
2317         int rc = 0;
2318         ENTRY;
2319
2320         LASSERT(mdt_object_remote(o));
2321
2322         fid_build_reg_res_name(fid, res_id);
2323
2324         memset(einfo, 0, sizeof(*einfo));
2325         einfo->ei_type = LDLM_IBITS;
2326         einfo->ei_mode = mode;
2327         einfo->ei_cb_bl = mdt_remote_blocking_ast;
2328         einfo->ei_cb_cp = ldlm_completion_ast;
2329         einfo->ei_enq_slave = 0;
2330         einfo->ei_res_id = res_id;
2331         if (nonblock)
2332                 einfo->ei_nonblock = 1;
2333
2334         memset(policy, 0, sizeof(*policy));
2335         policy->l_inodebits.bits = ibits;
2336
2337         rc = mo_object_lock(mti->mti_env, mdt_object_child(o), lh, einfo,
2338                             policy);
2339         RETURN(rc);
2340 }
2341
2342 static int mdt_object_local_lock(struct mdt_thread_info *info,
2343                                  struct mdt_object *o,
2344                                  struct mdt_lock_handle *lh, __u64 ibits,
2345                                  bool nonblock)
2346 {
2347         struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
2348         union ldlm_policy_data *policy = &info->mti_policy;
2349         struct ldlm_res_id *res_id = &info->mti_res_id;
2350         __u64 dlmflags;
2351         int rc;
2352         ENTRY;
2353
2354         LASSERT(!lustre_handle_is_used(&lh->mlh_reg_lh));
2355         LASSERT(!lustre_handle_is_used(&lh->mlh_pdo_lh));
2356         LASSERT(lh->mlh_reg_mode != LCK_MINMODE);
2357         LASSERT(lh->mlh_type != MDT_NUL_LOCK);
2358
2359         /* Only enqueue LOOKUP lock for remote object */
2360         if (mdt_object_remote(o))
2361                 LASSERT(ibits == MDS_INODELOCK_LOOKUP);
2362
2363         if (lh->mlh_type == MDT_PDO_LOCK) {
2364                 /* check for exists after object is locked */
2365                 if (mdt_object_exists(o) == 0) {
2366                         /* Non-existent object shouldn't have PDO lock */
2367                         RETURN(-ESTALE);
2368                 } else {
2369                         /* Non-dir object shouldn't have PDO lock */
2370                         if (!S_ISDIR(lu_object_attr(&o->mot_obj)))
2371                                 RETURN(-ENOTDIR);
2372                 }
2373         }
2374
2375         memset(policy, 0, sizeof(*policy));
2376         fid_build_reg_res_name(mdt_object_fid(o), res_id);
2377
2378         dlmflags = LDLM_FL_ATOMIC_CB;
2379         if (nonblock)
2380                 dlmflags |= LDLM_FL_BLOCK_NOWAIT;
2381
2382         /*
2383          * Take PDO lock on whole directory and build correct @res_id for lock
2384          * on part of directory.
2385          */
2386         if (lh->mlh_pdo_hash != 0) {
2387                 LASSERT(lh->mlh_type == MDT_PDO_LOCK);
2388                 mdt_lock_pdo_mode(info, o, lh);
2389                 if (lh->mlh_pdo_mode != LCK_NL) {
2390                         /*
2391                          * Do not use LDLM_FL_LOCAL_ONLY for parallel lock, it
2392                          * is never going to be sent to client and we do not
2393                          * want it slowed down due to possible cancels.
2394                          */
2395                         policy->l_inodebits.bits = MDS_INODELOCK_UPDATE;
2396                         rc = mdt_fid_lock(ns, &lh->mlh_pdo_lh, lh->mlh_pdo_mode,
2397                                           policy, res_id, dlmflags,
2398                                           info->mti_exp == NULL ? NULL :
2399                                           &info->mti_exp->exp_handle.h_cookie);
2400                         if (unlikely(rc != 0))
2401                                 GOTO(out_unlock, rc);
2402                 }
2403
2404                 /*
2405                  * Finish res_id initializing by name hash marking part of
2406                  * directory which is taking modification.
2407                  */
2408                 res_id->name[LUSTRE_RES_ID_HSH_OFF] = lh->mlh_pdo_hash;
2409         }
2410
2411         policy->l_inodebits.bits = ibits;
2412
2413         /*
2414          * Use LDLM_FL_LOCAL_ONLY for this lock. We do not know yet if it is
2415          * going to be sent to client. If it is - mdt_intent_policy() path will
2416          * fix it up and turn FL_LOCAL flag off.
2417          */
2418         rc = mdt_fid_lock(ns, &lh->mlh_reg_lh, lh->mlh_reg_mode, policy,
2419                           res_id, LDLM_FL_LOCAL_ONLY | dlmflags,
2420                           info->mti_exp == NULL ? NULL :
2421                           &info->mti_exp->exp_handle.h_cookie);
2422 out_unlock:
2423         if (rc != 0)
2424                 mdt_object_unlock(info, o, lh, 1);
2425         else if (unlikely(OBD_FAIL_PRECHECK(OBD_FAIL_MDS_PDO_LOCK)) &&
2426                    lh->mlh_pdo_hash != 0 &&
2427                    (lh->mlh_reg_mode == LCK_PW || lh->mlh_reg_mode == LCK_EX))
2428                 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PDO_LOCK, 15);
2429
2430         RETURN(rc);
2431 }
2432
2433 static int
2434 mdt_object_lock_internal(struct mdt_thread_info *info, struct mdt_object *o,
2435                          struct mdt_lock_handle *lh, __u64 ibits,
2436                          bool nonblock)
2437 {
2438         struct mdt_lock_handle *local_lh = NULL;
2439         int rc;
2440         ENTRY;
2441
2442         if (!mdt_object_remote(o))
2443                 return mdt_object_local_lock(info, o, lh, ibits, nonblock);
2444
2445         /* XXX do not support PERM/LAYOUT/XATTR lock for remote object yet */
2446         ibits &= ~(MDS_INODELOCK_PERM | MDS_INODELOCK_LAYOUT |
2447                    MDS_INODELOCK_XATTR);
2448
2449         /* Only enqueue LOOKUP lock for remote object */
2450         if (ibits & MDS_INODELOCK_LOOKUP) {
2451                 rc = mdt_object_local_lock(info, o, lh,
2452                                            MDS_INODELOCK_LOOKUP,
2453                                            nonblock);
2454                 if (rc != ELDLM_OK)
2455                         RETURN(rc);
2456
2457                 local_lh = lh;
2458         }
2459
2460         if (ibits & MDS_INODELOCK_UPDATE) {
2461                 /* Sigh, PDO needs to enqueue 2 locks right now, but
2462                  * enqueue RPC can only request 1 lock, to avoid extra
2463                  * RPC, so it will instead enqueue EX lock for remote
2464                  * object anyway XXX*/
2465                 if (lh->mlh_type == MDT_PDO_LOCK &&
2466                     lh->mlh_pdo_hash != 0) {
2467                         CDEBUG(D_INFO, "%s: "DFID" convert PDO lock to"
2468                                "EX lock.\n", mdt_obd_name(info->mti_mdt),
2469                                PFID(mdt_object_fid(o)));
2470                         lh->mlh_pdo_hash = 0;
2471                         lh->mlh_rreg_mode = LCK_EX;
2472                         lh->mlh_type = MDT_REG_LOCK;
2473                 }
2474                 rc = mdt_remote_object_lock(info, o, mdt_object_fid(o),
2475                                             &lh->mlh_rreg_lh,
2476                                             lh->mlh_rreg_mode,
2477                                             MDS_INODELOCK_UPDATE, nonblock);
2478                 if (rc != ELDLM_OK) {
2479                         if (local_lh != NULL)
2480                                 mdt_object_unlock(info, o, local_lh, rc);
2481                         RETURN(rc);
2482                 }
2483         }
2484
2485         RETURN(0);
2486 }
2487
2488 int mdt_object_lock(struct mdt_thread_info *info, struct mdt_object *o,
2489                     struct mdt_lock_handle *lh, __u64 ibits)
2490 {
2491         return mdt_object_lock_internal(info, o, lh, ibits, false);
2492 }
2493
2494 int mdt_object_lock_try(struct mdt_thread_info *info, struct mdt_object *o,
2495                         struct mdt_lock_handle *lh, __u64 ibits)
2496 {
2497         struct mdt_lock_handle tmp = *lh;
2498         int rc;
2499
2500         rc = mdt_object_lock_internal(info, o, &tmp, ibits, true);
2501         if (rc == 0)
2502                 *lh = tmp;
2503
2504         return rc == 0;
2505 }
2506
2507 /**
2508  * Save a lock within request object.
2509  *
2510  * Keep the lock referenced until whether client ACK or transaction
2511  * commit happens or release the lock immediately depending on input
2512  * parameters. If COS is ON, a write lock is converted to COS lock
2513  * before saving.
2514  *
2515  * \param info thead info object
2516  * \param h lock handle
2517  * \param mode lock mode
2518  * \param decref force immediate lock releasing
2519  */
2520 static void mdt_save_lock(struct mdt_thread_info *info, struct lustre_handle *h,
2521                           enum ldlm_mode mode, int decref)
2522 {
2523         ENTRY;
2524
2525         if (lustre_handle_is_used(h)) {
2526                 if (decref || !info->mti_has_trans ||
2527                     !(mode & (LCK_PW | LCK_EX))) {
2528                         mdt_fid_unlock(h, mode);
2529                 } else {
2530                         struct mdt_device *mdt = info->mti_mdt;
2531                         struct ldlm_lock *lock = ldlm_handle2lock(h);
2532                         struct ptlrpc_request *req = mdt_info_req(info);
2533                         int no_ack = 0;
2534
2535                         LASSERTF(lock != NULL, "no lock for cookie "LPX64"\n",
2536                                  h->cookie);
2537                         /* there is no request if mdt_object_unlock() is called
2538                          * from mdt_export_cleanup()->mdt_add_dirty_flag() */
2539                         if (likely(req != NULL)) {
2540                                 CDEBUG(D_HA, "request = %p reply state = %p"
2541                                        " transno = "LPD64"\n", req,
2542                                        req->rq_reply_state, req->rq_transno);
2543                                 if (mdt_cos_is_enabled(mdt)) {
2544                                         no_ack = 1;
2545                                         ldlm_lock_downgrade(lock, LCK_COS);
2546                                         mode = LCK_COS;
2547                                 }
2548                                 ptlrpc_save_lock(req, h, mode, no_ack);
2549                         } else {
2550                                 ldlm_lock_decref(h, mode);
2551                         }
2552                         if (mdt_is_lock_sync(lock)) {
2553                                 CDEBUG(D_HA, "found sync-lock,"
2554                                        " async commit started\n");
2555                                 mdt_device_commit_async(info->mti_env,
2556                                                         mdt);
2557                         }
2558                         LDLM_LOCK_PUT(lock);
2559                 }
2560                 h->cookie = 0ull;
2561         }
2562
2563         EXIT;
2564 }
2565
2566 /**
2567  * Unlock mdt object.
2568  *
2569  * Immeditely release the regular lock and the PDO lock or save the
2570  * lock in request and keep them referenced until client ACK or
2571  * transaction commit.
2572  *
2573  * \param info thread info object
2574  * \param o mdt object
2575  * \param lh mdt lock handle referencing regular and PDO locks
2576  * \param decref force immediate lock releasing
2577  */
2578 void mdt_object_unlock(struct mdt_thread_info *info, struct mdt_object *o,
2579                        struct mdt_lock_handle *lh, int decref)
2580 {
2581         ENTRY;
2582
2583         mdt_save_lock(info, &lh->mlh_pdo_lh, lh->mlh_pdo_mode, decref);
2584         mdt_save_lock(info, &lh->mlh_reg_lh, lh->mlh_reg_mode, decref);
2585
2586         if (lustre_handle_is_used(&lh->mlh_rreg_lh))
2587                 ldlm_lock_decref(&lh->mlh_rreg_lh, lh->mlh_rreg_mode);
2588
2589         EXIT;
2590 }
2591
2592 struct mdt_object *mdt_object_find_lock(struct mdt_thread_info *info,
2593                                         const struct lu_fid *f,
2594                                         struct mdt_lock_handle *lh,
2595                                         __u64 ibits)
2596 {
2597         struct mdt_object *o;
2598
2599         o = mdt_object_find(info->mti_env, info->mti_mdt, f);
2600         if (!IS_ERR(o)) {
2601                 int rc;
2602
2603                 rc = mdt_object_lock(info, o, lh, ibits);
2604                 if (rc != 0) {
2605                         mdt_object_put(info->mti_env, o);
2606                         o = ERR_PTR(rc);
2607                 }
2608         }
2609         return o;
2610 }
2611
2612 void mdt_object_unlock_put(struct mdt_thread_info * info,
2613                            struct mdt_object * o,
2614                            struct mdt_lock_handle *lh,
2615                            int decref)
2616 {
2617         mdt_object_unlock(info, o, lh, decref);
2618         mdt_object_put(info->mti_env, o);
2619 }
2620
2621 /*
2622  * Generic code handling requests that have struct mdt_body passed in:
2623  *
2624  *  - extract mdt_body from request and save it in @info, if present;
2625  *
2626  *  - create lu_object, corresponding to the fid in mdt_body, and save it in
2627  *  @info;
2628  *
2629  *  - if HABEO_CORPUS flag is set for this request type check whether object
2630  *  actually exists on storage (lu_object_exists()).
2631  *
2632  */
2633 static int mdt_body_unpack(struct mdt_thread_info *info, __u32 flags)
2634 {
2635         const struct mdt_body    *body;
2636         struct mdt_object        *obj;
2637         const struct lu_env      *env;
2638         struct req_capsule       *pill;
2639         int                       rc;
2640         ENTRY;
2641
2642         env = info->mti_env;
2643         pill = info->mti_pill;
2644
2645         body = info->mti_body = req_capsule_client_get(pill, &RMF_MDT_BODY);
2646         if (body == NULL)
2647                 RETURN(-EFAULT);
2648
2649         if (!(body->mbo_valid & OBD_MD_FLID))
2650                 RETURN(0);
2651
2652         if (!fid_is_sane(&body->mbo_fid1)) {
2653                 CERROR("Invalid fid: "DFID"\n", PFID(&body->mbo_fid1));
2654                 RETURN(-EINVAL);
2655         }
2656
2657         obj = mdt_object_find(env, info->mti_mdt, &body->mbo_fid1);
2658         if (!IS_ERR(obj)) {
2659                 if ((flags & HABEO_CORPUS) && !mdt_object_exists(obj)) {
2660                         mdt_object_put(env, obj);
2661                         rc = -ENOENT;
2662                 } else {
2663                         info->mti_object = obj;
2664                         rc = 0;
2665                 }
2666         } else
2667                 rc = PTR_ERR(obj);
2668
2669         RETURN(rc);
2670 }
2671
2672 static int mdt_unpack_req_pack_rep(struct mdt_thread_info *info, __u32 flags)
2673 {
2674         struct req_capsule *pill = info->mti_pill;
2675         int rc;
2676         ENTRY;
2677
2678         if (req_capsule_has_field(pill, &RMF_MDT_BODY, RCL_CLIENT))
2679                 rc = mdt_body_unpack(info, flags);
2680         else
2681                 rc = 0;
2682
2683         if (rc == 0 && (flags & HABEO_REFERO)) {
2684                 /* Pack reply. */
2685                 if (req_capsule_has_field(pill, &RMF_MDT_MD, RCL_SERVER))
2686                         req_capsule_set_size(pill, &RMF_MDT_MD, RCL_SERVER,
2687                                              DEF_REP_MD_SIZE);
2688                 if (req_capsule_has_field(pill, &RMF_LOGCOOKIES, RCL_SERVER))
2689                         req_capsule_set_size(pill, &RMF_LOGCOOKIES,
2690                                              RCL_SERVER, 0);
2691
2692                 rc = req_capsule_server_pack(pill);
2693         }
2694         RETURN(rc);
2695 }
2696
2697 void mdt_lock_handle_init(struct mdt_lock_handle *lh)
2698 {
2699         lh->mlh_type = MDT_NUL_LOCK;
2700         lh->mlh_reg_lh.cookie = 0ull;
2701         lh->mlh_reg_mode = LCK_MINMODE;
2702         lh->mlh_pdo_lh.cookie = 0ull;
2703         lh->mlh_pdo_mode = LCK_MINMODE;
2704         lh->mlh_rreg_lh.cookie = 0ull;
2705         lh->mlh_rreg_mode = LCK_MINMODE;
2706 }
2707
2708 void mdt_lock_handle_fini(struct mdt_lock_handle *lh)
2709 {
2710         LASSERT(!lustre_handle_is_used(&lh->mlh_reg_lh));
2711         LASSERT(!lustre_handle_is_used(&lh->mlh_pdo_lh));
2712 }
2713
2714 /*
2715  * Initialize fields of struct mdt_thread_info. Other fields are left in
2716  * uninitialized state, because it's too expensive to zero out whole
2717  * mdt_thread_info (> 1K) on each request arrival.
2718  */
2719 void mdt_thread_info_init(struct ptlrpc_request *req,
2720                           struct mdt_thread_info *info)
2721 {
2722         int i;
2723
2724         info->mti_pill = &req->rq_pill;
2725
2726         /* lock handle */
2727         for (i = 0; i < ARRAY_SIZE(info->mti_lh); i++)
2728                 mdt_lock_handle_init(&info->mti_lh[i]);
2729
2730         /* mdt device: it can be NULL while CONNECT */
2731         if (req->rq_export) {
2732                 info->mti_mdt = mdt_dev(req->rq_export->exp_obd->obd_lu_dev);
2733                 info->mti_exp = req->rq_export;
2734         } else
2735                 info->mti_mdt = NULL;
2736         info->mti_env = req->rq_svc_thread->t_env;
2737         info->mti_transno = lustre_msg_get_transno(req->rq_reqmsg);
2738
2739         memset(&info->mti_attr, 0, sizeof(info->mti_attr));
2740         info->mti_big_buf = LU_BUF_NULL;
2741         info->mti_body = NULL;
2742         info->mti_object = NULL;
2743         info->mti_dlm_req = NULL;
2744         info->mti_has_trans = 0;
2745         info->mti_cross_ref = 0;
2746         info->mti_opdata = 0;
2747         info->mti_big_lmm_used = 0;
2748
2749         info->mti_spec.no_create = 0;
2750         info->mti_spec.sp_rm_entry = 0;
2751         info->mti_spec.sp_permitted = 0;
2752         info->mti_spec.sp_migrate_close = 0;
2753
2754         info->mti_spec.u.sp_ea.eadata = NULL;
2755         info->mti_spec.u.sp_ea.eadatalen = 0;
2756 }
2757
2758 void mdt_thread_info_fini(struct mdt_thread_info *info)
2759 {
2760         int i;
2761
2762         if (info->mti_object != NULL) {
2763                 mdt_object_put(info->mti_env, info->mti_object);
2764                 info->mti_object = NULL;
2765         }
2766
2767         for (i = 0; i < ARRAY_SIZE(info->mti_lh); i++)
2768                 mdt_lock_handle_fini(&info->mti_lh[i]);
2769         info->mti_env = NULL;
2770         info->mti_pill = NULL;
2771         info->mti_exp = NULL;
2772
2773         if (unlikely(info->mti_big_buf.lb_buf != NULL))
2774                 lu_buf_free(&info->mti_big_buf);
2775 }
2776
2777 struct mdt_thread_info *tsi2mdt_info(struct tgt_session_info *tsi)
2778 {
2779         struct mdt_thread_info  *mti;
2780
2781         mti = mdt_th_info(tsi->tsi_env);
2782         LASSERT(mti != NULL);
2783
2784         mdt_thread_info_init(tgt_ses_req(tsi), mti);
2785         if (tsi->tsi_corpus != NULL) {
2786                 mti->mti_object = mdt_obj(tsi->tsi_corpus);
2787                 lu_object_get(tsi->tsi_corpus);
2788         }
2789         mti->mti_body = tsi->tsi_mdt_body;
2790         mti->mti_dlm_req = tsi->tsi_dlm_req;
2791
2792         return mti;
2793 }
2794
2795 static int mdt_tgt_connect(struct tgt_session_info *tsi)
2796 {
2797         struct ptlrpc_request   *req = tgt_ses_req(tsi);
2798         int                      rc;
2799
2800         ENTRY;
2801
2802         if (OBD_FAIL_CHECK(OBD_FAIL_TGT_DELAY_CONDITIONAL) &&
2803             cfs_fail_val ==
2804             tsi2mdt_info(tsi)->mti_mdt->mdt_seq_site.ss_node_id) {
2805                 set_current_state(TASK_UNINTERRUPTIBLE);
2806                 schedule_timeout(msecs_to_jiffies(3 * MSEC_PER_SEC));
2807         }
2808
2809         rc = tgt_connect(tsi);
2810         if (rc != 0)
2811                 RETURN(rc);
2812
2813         rc = mdt_init_idmap(tsi);
2814         if (rc != 0)
2815                 GOTO(err, rc);
2816         RETURN(0);
2817 err:
2818         obd_disconnect(class_export_get(req->rq_export));
2819         return rc;
2820 }
2821
2822 enum mdt_it_code {
2823         MDT_IT_OPEN,
2824         MDT_IT_OCREAT,
2825         MDT_IT_CREATE,
2826         MDT_IT_GETATTR,
2827         MDT_IT_READDIR,
2828         MDT_IT_LOOKUP,
2829         MDT_IT_UNLINK,
2830         MDT_IT_TRUNC,
2831         MDT_IT_GETXATTR,
2832         MDT_IT_LAYOUT,
2833         MDT_IT_QUOTA,
2834         MDT_IT_NR
2835 };
2836
2837 static int mdt_intent_getattr(enum mdt_it_code opcode,
2838                               struct mdt_thread_info *info,
2839                               struct ldlm_lock **,
2840                               __u64);
2841
2842 static int mdt_intent_getxattr(enum mdt_it_code opcode,
2843                                 struct mdt_thread_info *info,
2844                                 struct ldlm_lock **lockp,
2845                                 __u64 flags);
2846
2847 static int mdt_intent_layout(enum mdt_it_code opcode,
2848                              struct mdt_thread_info *info,
2849                              struct ldlm_lock **,
2850                              __u64);
2851 static int mdt_intent_reint(enum mdt_it_code opcode,
2852                             struct mdt_thread_info *info,
2853                             struct ldlm_lock **,
2854                             __u64);
2855
2856 static struct mdt_it_flavor {
2857         const struct req_format *it_fmt;
2858         __u32                    it_flags;
2859         int                    (*it_act)(enum mdt_it_code ,
2860                                          struct mdt_thread_info *,
2861                                          struct ldlm_lock **,
2862                                          __u64);
2863         long                     it_reint;
2864 } mdt_it_flavor[] = {
2865         [MDT_IT_OPEN]     = {
2866                 .it_fmt   = &RQF_LDLM_INTENT,
2867                 /*.it_flags = HABEO_REFERO,*/
2868                 .it_flags = 0,
2869                 .it_act   = mdt_intent_reint,
2870                 .it_reint = REINT_OPEN
2871         },
2872         [MDT_IT_OCREAT]   = {
2873                 .it_fmt   = &RQF_LDLM_INTENT,
2874                 /*
2875                  * OCREAT is not a MUTABOR request as if the file
2876                  * already exists.
2877                  * We do the extra check of OBD_CONNECT_RDONLY in
2878                  * mdt_reint_open() when we really need to create
2879                  * the object.
2880                  */
2881                 .it_flags = 0,
2882                 .it_act   = mdt_intent_reint,
2883                 .it_reint = REINT_OPEN
2884         },
2885         [MDT_IT_CREATE]   = {
2886                 .it_fmt   = &RQF_LDLM_INTENT,
2887                 .it_flags = MUTABOR,
2888                 .it_act   = mdt_intent_reint,
2889                 .it_reint = REINT_CREATE
2890         },
2891         [MDT_IT_GETATTR]  = {
2892                 .it_fmt   = &RQF_LDLM_INTENT_GETATTR,
2893                 .it_flags = HABEO_REFERO,
2894                 .it_act   = mdt_intent_getattr
2895         },
2896         [MDT_IT_READDIR]  = {
2897                 .it_fmt   = NULL,
2898                 .it_flags = 0,
2899                 .it_act   = NULL
2900         },
2901         [MDT_IT_LOOKUP]   = {
2902                 .it_fmt   = &RQF_LDLM_INTENT_GETATTR,
2903                 .it_flags = HABEO_REFERO,
2904                 .it_act   = mdt_intent_getattr
2905         },
2906         [MDT_IT_UNLINK]   = {
2907                 .it_fmt   = &RQF_LDLM_INTENT_UNLINK,
2908                 .it_flags = MUTABOR,
2909                 .it_act   = NULL,
2910                 .it_reint = REINT_UNLINK
2911         },
2912         [MDT_IT_TRUNC]    = {
2913                 .it_fmt   = NULL,
2914                 .it_flags = MUTABOR,
2915                 .it_act   = NULL
2916         },
2917         [MDT_IT_GETXATTR] = {
2918                 .it_fmt   = &RQF_LDLM_INTENT_GETXATTR,
2919                 .it_flags = HABEO_CORPUS,
2920                 .it_act   = mdt_intent_getxattr
2921         },
2922         [MDT_IT_LAYOUT] = {
2923                 .it_fmt   = &RQF_LDLM_INTENT_LAYOUT,
2924                 .it_flags = 0,
2925                 .it_act   = mdt_intent_layout
2926         }
2927 };
2928
2929 static int
2930 mdt_intent_lock_replace(struct mdt_thread_info *info,
2931                         struct ldlm_lock **lockp,
2932                         struct mdt_lock_handle *lh,
2933                         __u64 flags)
2934 {
2935         struct ptlrpc_request  *req = mdt_info_req(info);
2936         struct ldlm_lock       *lock = *lockp;
2937         struct ldlm_lock       *new_lock;
2938
2939         /* If possible resent found a lock, @lh is set to its handle */
2940         new_lock = ldlm_handle2lock_long(&lh->mlh_reg_lh, 0);
2941
2942         if (new_lock == NULL && (flags & LDLM_FL_INTENT_ONLY)) {
2943                 lh->mlh_reg_lh.cookie = 0;
2944                 RETURN(0);
2945         }
2946
2947         LASSERTF(new_lock != NULL,
2948                  "lockh "LPX64"\n", lh->mlh_reg_lh.cookie);
2949
2950         /*
2951          * If we've already given this lock to a client once, then we should
2952          * have no readers or writers.  Otherwise, we should have one reader
2953          * _or_ writer ref (which will be zeroed below) before returning the
2954          * lock to a client.
2955          */
2956         if (new_lock->l_export == req->rq_export) {
2957                 LASSERT(new_lock->l_readers + new_lock->l_writers == 0);
2958         } else {
2959                 LASSERT(new_lock->l_export == NULL);
2960                 LASSERT(new_lock->l_readers + new_lock->l_writers == 1);
2961         }
2962
2963         *lockp = new_lock;
2964
2965         if (new_lock->l_export == req->rq_export) {
2966                 /*
2967                  * Already gave this to the client, which means that we
2968                  * reconstructed a reply.
2969                  */
2970                 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) &
2971                         MSG_RESENT);
2972
2973                 LDLM_LOCK_RELEASE(new_lock);
2974                 lh->mlh_reg_lh.cookie = 0;
2975                 RETURN(ELDLM_LOCK_REPLACED);
2976         }
2977
2978         /*
2979          * Fixup the lock to be given to the client.
2980          */
2981         lock_res_and_lock(new_lock);
2982         /* Zero new_lock->l_readers and new_lock->l_writers without triggering
2983          * possible blocking AST. */
2984         while (new_lock->l_readers > 0) {
2985                 lu_ref_del(&new_lock->l_reference, "reader", new_lock);
2986                 lu_ref_del(&new_lock->l_reference, "user", new_lock);
2987                 new_lock->l_readers--;
2988         }
2989         while (new_lock->l_writers > 0) {
2990                 lu_ref_del(&new_lock->l_reference, "writer", new_lock);
2991                 lu_ref_del(&new_lock->l_reference, "user", new_lock);
2992                 new_lock->l_writers--;
2993         }
2994
2995         new_lock->l_export = class_export_lock_get(req->rq_export, new_lock);
2996         new_lock->l_blocking_ast = lock->l_blocking_ast;
2997         new_lock->l_completion_ast = lock->l_completion_ast;
2998         new_lock->l_remote_handle = lock->l_remote_handle;
2999         new_lock->l_flags &= ~LDLM_FL_LOCAL;
3000
3001         unlock_res_and_lock(new_lock);
3002
3003         cfs_hash_add(new_lock->l_export->exp_lock_hash,
3004                      &new_lock->l_remote_handle,
3005                      &new_lock->l_exp_hash);
3006
3007         LDLM_LOCK_RELEASE(new_lock);
3008         lh->mlh_reg_lh.cookie = 0;
3009
3010         RETURN(ELDLM_LOCK_REPLACED);
3011 }
3012
3013 static void mdt_intent_fixup_resent(struct mdt_thread_info *info,
3014                                     struct ldlm_lock *new_lock,
3015                                     struct mdt_lock_handle *lh,
3016                                     __u64 flags)
3017 {
3018         struct ptlrpc_request  *req = mdt_info_req(info);
3019         struct ldlm_request    *dlmreq;
3020
3021         if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))
3022                 return;
3023
3024         dlmreq = req_capsule_client_get(info->mti_pill, &RMF_DLM_REQ);
3025
3026         /* Check if this is a resend case (MSG_RESENT is set on RPC) and a
3027          * lock was found by ldlm_handle_enqueue(); if so @lh must be
3028          * initialized. */
3029         if (flags & LDLM_FL_RESENT) {
3030                 lh->mlh_reg_lh.cookie = new_lock->l_handle.h_cookie;
3031                 lh->mlh_reg_mode = new_lock->l_granted_mode;
3032
3033                 LDLM_DEBUG(new_lock, "Restoring lock cookie");
3034                 DEBUG_REQ(D_DLMTRACE, req, "restoring lock cookie "LPX64,
3035                           lh->mlh_reg_lh.cookie);
3036                 return;
3037         }
3038
3039         /*
3040          * If the xid matches, then we know this is a resent request, and allow
3041          * it. (It's probably an OPEN, for which we don't send a lock.
3042          */
3043         if (req_can_reconstruct(req, NULL))
3044                 return;
3045
3046         /*
3047          * This remote handle isn't enqueued, so we never received or processed
3048          * this request.  Clear MSG_RESENT, because it can be handled like any
3049          * normal request now.
3050          */
3051         lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
3052
3053         DEBUG_REQ(D_DLMTRACE, req, "no existing lock with rhandle "LPX64,
3054                   dlmreq->lock_handle[0].cookie);
3055 }
3056
3057 static int mdt_intent_getxattr(enum mdt_it_code opcode,
3058                                 struct mdt_thread_info *info,
3059                                 struct ldlm_lock **lockp,
3060                                 __u64 flags)
3061 {
3062         struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_RMT];
3063         struct ldlm_reply      *ldlm_rep = NULL;
3064         int rc, grc;
3065
3066         /*
3067          * Initialize lhc->mlh_reg_lh either from a previously granted lock
3068          * (for the resend case) or a new lock. Below we will use it to
3069          * replace the original lock.
3070          */
3071         mdt_intent_fixup_resent(info, *lockp, lhc, flags);
3072         if (!lustre_handle_is_used(&lhc->mlh_reg_lh)) {
3073                 mdt_lock_reg_init(lhc, (*lockp)->l_req_mode);
3074                 rc = mdt_object_lock(info, info->mti_object, lhc,
3075                                      MDS_INODELOCK_XATTR);
3076                 if (rc)
3077                         return rc;
3078         }
3079
3080         grc = mdt_getxattr(info);
3081
3082         rc = mdt_intent_lock_replace(info, lockp, lhc, flags);
3083
3084         if (mdt_info_req(info)->rq_repmsg != NULL)
3085                 ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
3086         if (ldlm_rep == NULL)
3087                 RETURN(err_serious(-EFAULT));
3088
3089         ldlm_rep->lock_policy_res2 = grc;
3090
3091         return rc;
3092 }
3093
3094 static int mdt_intent_getattr(enum mdt_it_code opcode,
3095                               struct mdt_thread_info *info,
3096                               struct ldlm_lock **lockp,
3097                               __u64 flags)
3098 {
3099         struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_RMT];
3100         __u64                   child_bits;
3101         struct ldlm_reply      *ldlm_rep;
3102         struct mdt_body        *reqbody;
3103         struct mdt_body        *repbody;
3104         int                     rc, rc2;
3105         ENTRY;
3106
3107         reqbody = req_capsule_client_get(info->mti_pill, &RMF_MDT_BODY);
3108         LASSERT(reqbody);
3109
3110         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
3111         LASSERT(repbody);
3112
3113         info->mti_cross_ref = !!(reqbody->mbo_valid & OBD_MD_FLCROSSREF);
3114         repbody->mbo_eadatasize = 0;
3115         repbody->mbo_aclsize = 0;
3116
3117         switch (opcode) {
3118         case MDT_IT_LOOKUP:
3119                 child_bits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_PERM;
3120                 break;
3121         case MDT_IT_GETATTR:
3122                 child_bits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE |
3123                              MDS_INODELOCK_PERM;
3124                 break;
3125         default:
3126                 CERROR("Unsupported intent (%d)\n", opcode);
3127                 GOTO(out_shrink, rc = -EINVAL);
3128         }
3129
3130         rc = mdt_init_ucred_intent_getattr(info, reqbody);
3131         if (rc)
3132                 GOTO(out_shrink, rc);
3133
3134         ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
3135         mdt_set_disposition(info, ldlm_rep, DISP_IT_EXECD);
3136
3137         /* Get lock from request for possible resent case. */
3138         mdt_intent_fixup_resent(info, *lockp, lhc, flags);
3139
3140         rc = mdt_getattr_name_lock(info, lhc, child_bits, ldlm_rep);
3141         ldlm_rep->lock_policy_res2 = clear_serious(rc);
3142
3143         if (mdt_get_disposition(ldlm_rep, DISP_LOOKUP_NEG))
3144                 ldlm_rep->lock_policy_res2 = 0;
3145         if (!mdt_get_disposition(ldlm_rep, DISP_LOOKUP_POS) ||
3146             ldlm_rep->lock_policy_res2) {
3147                 lhc->mlh_reg_lh.cookie = 0ull;
3148                 GOTO(out_ucred, rc = ELDLM_LOCK_ABORTED);
3149         }
3150
3151         rc = mdt_intent_lock_replace(info, lockp, lhc, flags);
3152         EXIT;
3153 out_ucred:
3154         mdt_exit_ucred(info);
3155 out_shrink:
3156         mdt_client_compatibility(info);
3157         rc2 = mdt_fix_reply(info);
3158         if (rc == 0)
3159                 rc = rc2;
3160         return rc;
3161 }
3162
3163 static int mdt_intent_layout(enum mdt_it_code opcode,
3164                              struct mdt_thread_info *info,
3165                              struct ldlm_lock **lockp,
3166                              __u64 flags)
3167 {
3168         struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_LAYOUT];
3169         struct layout_intent *layout;
3170         struct lu_fid *fid;
3171         struct mdt_object *obj = NULL;
3172         int layout_size = 0;
3173         int rc = 0;
3174         ENTRY;
3175
3176         if (opcode != MDT_IT_LAYOUT) {
3177                 CERROR("%s: Unknown intent (%d)\n", mdt_obd_name(info->mti_mdt),
3178                         opcode);
3179                 RETURN(-EINVAL);
3180         }
3181
3182         layout = req_capsule_client_get(info->mti_pill, &RMF_LAYOUT_INTENT);
3183         if (layout == NULL)
3184                 RETURN(-EPROTO);
3185
3186         if (layout->li_opc != LAYOUT_INTENT_ACCESS) {
3187                 CERROR("%s: Unsupported layout intent opc %d\n",
3188                        mdt_obd_name(info->mti_mdt), layout->li_opc);
3189                 RETURN(-EINVAL);
3190         }
3191
3192         fid = &info->mti_tmp_fid2;
3193         fid_extract_from_res_name(fid, &(*lockp)->l_resource->lr_name);
3194
3195         /* Get lock from request for possible resent case. */
3196         mdt_intent_fixup_resent(info, *lockp, lhc, flags);
3197
3198         obj = mdt_object_find(info->mti_env, info->mti_mdt, fid);
3199         if (IS_ERR(obj))