4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
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.
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).
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
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
27 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2014, Intel Corporation.
33 * lustre/osp/osp_object.c
35 * Lustre OST Proxy Device (OSP) is the agent on the local MDT for the OST
38 * OSP object attributes cache
39 * ---------------------------
40 * OSP object is the stub of the remote OST-object or MDT-object. Both the
41 * attribute and the extended attributes are stored on the peer side remotely.
42 * It is inefficient to send RPC to peer to fetch those attributes when every
43 * get_attr()/get_xattr() called. For a large system, the LFSCK synchronous
44 * mode scanning is prohibitively inefficient.
46 * So the OSP maintains the OSP object attributes cache to cache some
47 * attributes on the local MDT. The cache is organized against the OSP
50 * struct osp_xattr_entry {
51 * struct list_head oxe_list;
57 * unsigned int oxe_exist:1,
62 * struct osp_object_attr {
63 * struct lu_attr ooa_attr;
64 * struct list_head ooa_xattr_list;
69 * struct osp_object_attr *opo_ooa;
70 * spinlock_t opo_lock;
74 * The basic attributes, such as owner/mode/flags, are stored in the
75 * osp_object_attr::ooa_attr. The extended attributes will be stored
76 * as osp_xattr_entry. Every extended attribute has an independent
77 * osp_xattr_entry, and all the osp_xattr_entry are linked into the
78 * osp_object_attr::ooa_xattr_list. The OSP object attributes cache
79 * is protected by the osp_object::opo_lock.
81 * Not all OSP objects have an attributes cache because maintaining
82 * the cache requires some resources. Currently, the OSP object
83 * attributes cache will be initialized when the attributes or the
84 * extended attributes are pre-fetched via osp_declare_attr_get()
85 * or osp_declare_xattr_get(). That is usually for LFSCK purpose,
86 * but it also can be shared by others.
88 * Author: Alex Zhuravlev <alexey.zhuravlev@intel.com>
89 * Author: Mikhail Pershin <mike.tappro@intel.com>
92 #define DEBUG_SUBSYSTEM S_MDS
94 #include "osp_internal.h"
96 static inline __u32 osp_dev2node(struct osp_device *osp)
98 return osp->opd_storage->dd_lu_dev.ld_site->ld_seq_site->ss_node_id;
101 static inline bool is_ost_obj(struct lu_object *lo)
103 return !lu2osp_dev(lo->lo_dev)->opd_connect_mdt;
107 * Assign FID to the OST object.
109 * This function will assign the FID to the OST object of a striped file.
111 * \param[in] env pointer to the thread context
112 * \param[in] d pointer to the OSP device
113 * \param[in] o pointer to the OSP object that the FID will be
116 static void osp_object_assign_fid(const struct lu_env *env,
117 struct osp_device *d, struct osp_object *o)
119 struct osp_thread_info *osi = osp_env_info(env);
121 LASSERT(fid_is_zero(lu_object_fid(&o->opo_obj.do_lu)));
122 LASSERT(o->opo_reserved);
125 osp_precreate_get_fid(env, d, &osi->osi_fid);
127 lu_object_assign_fid(env, &o->opo_obj.do_lu, &osi->osi_fid);
131 * Initialize the OSP object attributes cache.
133 * \param[in] obj pointer to the OSP object
135 * \retval 0 for success
136 * \retval negative error number on failure
138 static int osp_oac_init(struct osp_object *obj)
140 struct osp_object_attr *ooa;
146 INIT_LIST_HEAD(&ooa->ooa_xattr_list);
147 spin_lock(&obj->opo_lock);
148 if (likely(obj->opo_ooa == NULL)) {
150 spin_unlock(&obj->opo_lock);
152 spin_unlock(&obj->opo_lock);
160 * Find the named extended attribute in the OSP object attributes cache.
162 * The caller should take the osp_object::opo_lock before calling
165 * \param[in] ooa pointer to the OSP object attributes cache
166 * \param[in] name the name of the extended attribute
167 * \param[in] namelen the name length of the extended attribute
169 * \retval pointer to the found extended attribute entry
170 * \retval NULL if the specified extended attribute is not
173 static struct osp_xattr_entry *
174 osp_oac_xattr_find_locked(struct osp_object_attr *ooa,
175 const char *name, size_t namelen)
177 struct osp_xattr_entry *oxe;
179 list_for_each_entry(oxe, &ooa->ooa_xattr_list, oxe_list) {
180 if (namelen == oxe->oxe_namelen &&
181 strncmp(name, oxe->oxe_buf, namelen) == 0)
189 * Find the named extended attribute in the OSP object attributes cache.
191 * Call osp_oac_xattr_find_locked() with the osp_object::opo_lock held.
193 * \param[in] obj pointer to the OSP object
194 * \param[in] name the name of the extended attribute
195 * \param[in] unlink true if the extended attribute entry is to be removed
198 * \retval pointer to the found extended attribute entry
199 * \retval NULL if the specified extended attribute is not
202 static struct osp_xattr_entry *osp_oac_xattr_find(struct osp_object *obj,
203 const char *name, bool unlink)
205 struct osp_xattr_entry *oxe = NULL;
207 spin_lock(&obj->opo_lock);
208 if (obj->opo_ooa != NULL) {
209 oxe = osp_oac_xattr_find_locked(obj->opo_ooa, name,
213 list_del_init(&oxe->oxe_list);
215 atomic_inc(&oxe->oxe_ref);
218 spin_unlock(&obj->opo_lock);
224 * Find the named extended attribute in the OSP object attributes cache.
226 * If it is not in the cache, then add an empty entry (that will be
227 * filled later) to cache with the given name.
229 * \param[in] obj pointer to the OSP object
230 * \param[in] name the name of the extended attribute
231 * \param[in] len the length of the extended attribute value
233 * \retval pointer to the found or new-created extended
235 * \retval NULL if the specified extended attribute is not in the
236 * cache or fail to add new empty entry to the cache.
238 static struct osp_xattr_entry *
239 osp_oac_xattr_find_or_add(struct osp_object *obj, const char *name, size_t len)
241 struct osp_object_attr *ooa = obj->opo_ooa;
242 struct osp_xattr_entry *oxe;
243 struct osp_xattr_entry *tmp = NULL;
244 size_t namelen = strlen(name);
245 size_t size = sizeof(*oxe) + namelen + 1 + len;
247 LASSERT(ooa != NULL);
249 oxe = osp_oac_xattr_find(obj, name, false);
253 OBD_ALLOC(oxe, size);
254 if (unlikely(oxe == NULL))
257 INIT_LIST_HEAD(&oxe->oxe_list);
258 oxe->oxe_buflen = size;
259 oxe->oxe_namelen = namelen;
260 memcpy(oxe->oxe_buf, name, namelen);
261 oxe->oxe_value = oxe->oxe_buf + namelen + 1;
262 /* One ref is for the caller, the other is for the entry on the list. */
263 atomic_set(&oxe->oxe_ref, 2);
265 spin_lock(&obj->opo_lock);
266 tmp = osp_oac_xattr_find_locked(ooa, name, namelen);
268 list_add_tail(&oxe->oxe_list, &ooa->ooa_xattr_list);
270 atomic_inc(&tmp->oxe_ref);
271 spin_unlock(&obj->opo_lock);
282 * Add the given extended attribute to the OSP object attributes cache.
284 * If there is an old extended attributed entry with the same name,
285 * remove it from the cache and return it via the parameter \a poxe.
287 * \param[in] obj pointer to the OSP object
288 * \param[in,out] poxe double pointer to the OSP object extended attribute
289 * entry: the new extended attribute entry is transfered
290 * via such pointer target, and if old the extended
291 * attribute entry exists, then it will be returned back
292 * via such pointer target.
293 * \param[in] len the length of the (new) extended attribute value
295 * \retval pointer to the new extended attribute entry
296 * \retval NULL for failure cases.
298 static struct osp_xattr_entry *
299 osp_oac_xattr_replace(struct osp_object *obj,
300 struct osp_xattr_entry **poxe, size_t len)
302 struct osp_object_attr *ooa = obj->opo_ooa;
303 struct osp_xattr_entry *oxe;
304 size_t namelen = (*poxe)->oxe_namelen;
305 size_t size = sizeof(*oxe) + namelen + 1 + len;
307 LASSERT(ooa != NULL);
309 OBD_ALLOC(oxe, size);
310 if (unlikely(oxe == NULL))
313 INIT_LIST_HEAD(&oxe->oxe_list);
314 oxe->oxe_buflen = size;
315 oxe->oxe_namelen = namelen;
316 memcpy(oxe->oxe_buf, (*poxe)->oxe_buf, namelen);
317 oxe->oxe_value = oxe->oxe_buf + namelen + 1;
318 /* One ref is for the caller, the other is for the entry on the list. */
319 atomic_set(&oxe->oxe_ref, 2);
321 spin_lock(&obj->opo_lock);
322 *poxe = osp_oac_xattr_find_locked(ooa, oxe->oxe_buf, namelen);
323 LASSERT(*poxe != NULL);
325 list_del_init(&(*poxe)->oxe_list);
326 list_add_tail(&oxe->oxe_list, &ooa->ooa_xattr_list);
327 spin_unlock(&obj->opo_lock);
333 * Release reference from the OSP object extended attribute entry.
335 * If it is the last reference, then free the entry.
337 * \param[in] oxe pointer to the OSP object extended attribute entry.
339 static inline void osp_oac_xattr_put(struct osp_xattr_entry *oxe)
341 if (atomic_dec_and_test(&oxe->oxe_ref)) {
342 LASSERT(list_empty(&oxe->oxe_list));
344 OBD_FREE(oxe, oxe->oxe_buflen);
349 * Parse the OSP object attribute from the RPC reply.
351 * If the attribute is valid, then it will be added to the OSP object
354 * \param[in] env pointer to the thread context
355 * \param[in] reply pointer to the RPC reply
356 * \param[in] req pointer to the RPC request
357 * \param[out] attr pointer to buffer to hold the output attribute
358 * \param[in] obj pointer to the OSP object
359 * \param[in] index the index of the attribute buffer in the reply
361 * \retval 0 for success
362 * \retval negative error number on failure
364 static int osp_get_attr_from_reply(const struct lu_env *env,
365 struct object_update_reply *reply,
366 struct ptlrpc_request *req,
367 struct lu_attr *attr,
368 struct osp_object *obj, int index)
370 struct osp_thread_info *osi = osp_env_info(env);
371 struct lu_buf *rbuf = &osi->osi_lb2;
372 struct obdo *lobdo = &osi->osi_obdo;
376 rc = object_update_result_data_get(reply, rbuf, index);
380 wobdo = rbuf->lb_buf;
381 if (rbuf->lb_len != sizeof(*wobdo))
384 LASSERT(req != NULL);
385 if (ptlrpc_req_need_swab(req))
386 lustre_swab_obdo(wobdo);
388 lustre_get_wire_obdo(NULL, lobdo, wobdo);
389 spin_lock(&obj->opo_lock);
390 if (obj->opo_ooa != NULL) {
391 la_from_obdo(&obj->opo_ooa->ooa_attr, lobdo, lobdo->o_valid);
393 *attr = obj->opo_ooa->ooa_attr;
395 LASSERT(attr != NULL);
397 la_from_obdo(attr, lobdo, lobdo->o_valid);
399 spin_unlock(&obj->opo_lock);
405 * Interpreter function for getting OSP object attribute asynchronously.
407 * Called to interpret the result of an async mode RPC for getting the
408 * OSP object attribute.
410 * \param[in] env pointer to the thread context
411 * \param[in] reply pointer to the RPC reply
412 * \param[in] req pointer to the RPC request
413 * \param[in] obj pointer to the OSP object
414 * \param[out] data pointer to buffer to hold the output attribute
415 * \param[in] index the index of the attribute buffer in the reply
416 * \param[in] rc the result for handling the RPC
418 * \retval 0 for success
419 * \retval negative error number on failure
421 static int osp_attr_get_interpterer(const struct lu_env *env,
422 struct object_update_reply *reply,
423 struct ptlrpc_request *req,
424 struct osp_object *obj,
425 void *data, int index, int rc)
427 struct lu_attr *attr = data;
429 LASSERT(obj->opo_ooa != NULL);
432 osp2lu_obj(obj)->lo_header->loh_attr |= LOHA_EXISTS;
433 obj->opo_non_exist = 0;
435 return osp_get_attr_from_reply(env, reply, req, NULL, obj,
439 osp2lu_obj(obj)->lo_header->loh_attr &= ~LOHA_EXISTS;
440 obj->opo_non_exist = 1;
443 spin_lock(&obj->opo_lock);
445 spin_unlock(&obj->opo_lock);
452 * Implement OSP layer dt_object_operations::do_declare_attr_get() interface.
454 * Declare that the caller will get attribute from the specified OST object.
456 * This function adds an Object Unified Target (OUT) sub-request to the per-OSP
457 * based shared asynchronous request queue. The osp_attr_get_interpterer()
458 * is registered as the interpreter function to handle the result of this
461 * \param[in] env pointer to the thread context
462 * \param[in] dt pointer to the OSP layer dt_object
463 * \param[in] capa the capability for this operation
465 * \retval 0 for success
466 * \retval negative error number on failure
468 static int osp_declare_attr_get(const struct lu_env *env, struct dt_object *dt,
469 struct lustre_capa *capa)
471 struct osp_object *obj = dt2osp_obj(dt);
472 struct osp_device *osp = lu2osp_dev(dt->do_lu.lo_dev);
475 if (obj->opo_ooa == NULL) {
476 rc = osp_oac_init(obj);
481 mutex_lock(&osp->opd_async_requests_mutex);
482 rc = osp_insert_async_request(env, OUT_ATTR_GET, obj, 0, NULL, NULL,
483 &obj->opo_ooa->ooa_attr,
484 osp_attr_get_interpterer);
485 mutex_unlock(&osp->opd_async_requests_mutex);
491 * Implement OSP layer dt_object_operations::do_attr_get() interface.
493 * Get attribute from the specified MDT/OST object.
495 * If the attribute is in the OSP object attributes cache, then return
496 * the cached attribute directly. Otherwise it will trigger an OUT RPC
497 * to the peer to get the attribute synchronously, if successful, add it
498 * to the OSP attributes cache. (\see lustre/osp/osp_trans.c for OUT RPC.)
500 * \param[in] env pointer to the thread context
501 * \param[in] dt pointer to the OSP layer dt_object
502 * \param[out] attr pointer to the buffer to hold the output attribute
503 * \param[in] capa the capability for this operation
505 * \retval 0 for success
506 * \retval negative error number on failure
508 int osp_attr_get(const struct lu_env *env, struct dt_object *dt,
509 struct lu_attr *attr, struct lustre_capa *capa)
511 struct osp_device *osp = lu2osp_dev(dt->do_lu.lo_dev);
512 struct osp_object *obj = dt2osp_obj(dt);
513 struct dt_device *dev = &osp->opd_dt_dev;
514 struct dt_update_request *update;
515 struct object_update_reply *reply;
516 struct ptlrpc_request *req = NULL;
520 if (is_ost_obj(&dt->do_lu) && obj->opo_non_exist)
523 if (obj->opo_ooa != NULL) {
524 spin_lock(&obj->opo_lock);
525 if (obj->opo_ooa->ooa_attr.la_valid != 0) {
526 *attr = obj->opo_ooa->ooa_attr;
527 spin_unlock(&obj->opo_lock);
531 spin_unlock(&obj->opo_lock);
534 update = dt_update_request_create(dev);
536 RETURN(PTR_ERR(update));
538 rc = out_attr_get_pack(env, &update->dur_buf,
539 lu_object_fid(&dt->do_lu));
541 CERROR("%s: Insert update error "DFID": rc = %d\n",
542 dev->dd_lu_dev.ld_obd->obd_name,
543 PFID(lu_object_fid(&dt->do_lu)), rc);
548 rc = out_remote_sync(env, osp->opd_obd->u.cli.cl_import, update, &req);
551 osp2lu_obj(obj)->lo_header->loh_attr &= ~LOHA_EXISTS;
552 obj->opo_non_exist = 1;
554 CERROR("%s:osp_attr_get update error "DFID": rc = %d\n",
555 dev->dd_lu_dev.ld_obd->obd_name,
556 PFID(lu_object_fid(&dt->do_lu)), rc);
562 osp2lu_obj(obj)->lo_header->loh_attr |= LOHA_EXISTS;
563 obj->opo_non_exist = 0;
564 reply = req_capsule_server_sized_get(&req->rq_pill,
565 &RMF_OUT_UPDATE_REPLY,
566 OUT_UPDATE_REPLY_SIZE);
567 if (reply == NULL || reply->ourp_magic != UPDATE_REPLY_MAGIC)
568 GOTO(out, rc = -EPROTO);
570 rc = osp_get_attr_from_reply(env, reply, req, attr, obj, 0);
578 ptlrpc_req_finished(req);
580 dt_update_request_destroy(update);
585 static int __osp_attr_set(const struct lu_env *env, struct dt_object *dt,
586 const struct lu_attr *attr, struct thandle *th)
588 struct osp_device *d = lu2osp_dev(dt->do_lu.lo_dev);
589 struct osp_object *o = dt2osp_obj(dt);
595 * Usually we don't allow server stack to manipulate size
596 * but there is a special case when striping is created
597 * late, after stripeless file got truncated to non-zero.
599 * In this case we do the following:
601 * 1) grab id in declare - this can lead to leaked OST objects
602 * but we don't currently have proper mechanism and the only
603 * options we have are to do truncate RPC holding transaction
604 * open (very bad) or to grab id in declare at cost of leaked
605 * OST object in same very rare unfortunate case (just bad)
606 * notice 1.6-2.0 do assignment outside of running transaction
607 * all the time, meaning many more chances for leaked objects.
609 * 2) send synchronous truncate RPC with just assigned id
612 /* there are few places in MDD code still passing NULL
613 * XXX: to be fixed soon */
617 if (attr->la_valid & LA_SIZE && attr->la_size > 0 &&
618 fid_is_zero(lu_object_fid(&o->opo_obj.do_lu))) {
619 LASSERT(!dt_object_exists(dt));
620 osp_object_assign_fid(env, d, o);
621 rc = osp_object_truncate(env, dt, attr->la_size);
627 /* no need in logging for new objects being created */
630 if (!(attr->la_valid & (LA_UID | LA_GID)))
633 if (!is_only_remote_trans(th))
635 * track all UID/GID changes via llog
637 rc = osp_sync_declare_add(env, o, MDS_SETATTR64_REC, th);
639 /* It is for OST-object attr_set directly without updating
640 * local MDT-object attribute. It is usually used by LFSCK. */
641 rc = osp_md_declare_attr_set(env, dt, attr, th);
643 if (rc != 0 || o->opo_ooa == NULL)
646 /* Update the OSP object attributes cache. */
647 la = &o->opo_ooa->ooa_attr;
648 spin_lock(&o->opo_lock);
649 if (attr->la_valid & LA_UID) {
650 la->la_uid = attr->la_uid;
651 la->la_valid |= LA_UID;
654 if (attr->la_valid & LA_GID) {
655 la->la_gid = attr->la_gid;
656 la->la_valid |= LA_GID;
658 spin_unlock(&o->opo_lock);
664 * Implement OSP layer dt_object_operations::do_declare_attr_set() interface.
665 * XXX: NOT prepare set_{attr,xattr} RPC for remote transaction.
667 * According to our current transaction/dt_object_lock framework (to make
668 * the cross-MDTs modification for DNE1 to be workable), the transaction
669 * sponsor will start the transaction firstly, then try to acquire related
670 * dt_object_lock if needed. Under such rules, if we want to prepare the
671 * set_{attr,xattr} RPC in the RPC declare phase, then related attr/xattr
672 * should be known without dt_object_lock. But such condition maybe not
673 * true for some remote transaction case. For example:
675 * For linkEA repairing (by LFSCK) case, before the LFSCK thread obtained
676 * the dt_object_lock on the target MDT-object, it cannot know whether
677 * the MDT-object has linkEA or not, neither invalid or not.
679 * Since the LFSCK thread cannot hold dt_object_lock before the (remote)
680 * transaction start (otherwise there will be some potential deadlock),
681 * it cannot prepare related RPC for repairing during the declare phase
682 * as other normal transactions do.
684 * To resolve the trouble, we will make OSP to prepare related RPC
685 * (set_attr/set_xattr/del_xattr) after remote transaction started,
686 * and trigger the remote updating (RPC sending) when trans_stop.
687 * Then the up layer users, such as LFSCK, can follow the general
688 * rule to handle trans_start/dt_object_lock for repairing linkEA
689 * inconsistency without distinguishing remote MDT-object.
691 * In fact, above solution for remote transaction should be the normal
692 * model without considering DNE1. The trouble brought by DNE1 will be
693 * resolved in DNE2. At that time, this patch can be removed.
695 * \param[in] env pointer to the thread context
696 * \param[in] dt pointer to the OSP layer dt_object
697 * \param[in] attr pointer to the attribute to be set
698 * \param[in] th pointer to the transaction handler
700 * \retval 0 for success
701 * \retval negative error number on failure
703 static int osp_declare_attr_set(const struct lu_env *env, struct dt_object *dt,
704 const struct lu_attr *attr, struct thandle *th)
708 if (!is_only_remote_trans(th))
709 rc = __osp_attr_set(env, dt, attr, th);
715 * Implement OSP layer dt_object_operations::do_attr_set() interface.
717 * Set attribute to the specified OST object.
719 * If the transaction is a remote transaction, then related modification
720 * sub-request has been added in the declare phase and related OUT RPC
721 * has been triggered at transaction start. Otherwise it will generate
722 * a MDS_SETATTR64_REC record in the llog. There is a dedicated thread
723 * to handle the llog asynchronously.
725 * If the attribute entry exists in the OSP object attributes cache,
726 * then update the cached attribute according to given attribute.
728 * \param[in] env pointer to the thread context
729 * \param[in] dt pointer to the OSP layer dt_object
730 * \param[in] attr pointer to the attribute to be set
731 * \param[in] th pointer to the transaction handler
732 * \param[in] capa the capability for this operation
734 * \retval 0 for success
735 * \retval negative error number on failure
737 static int osp_attr_set(const struct lu_env *env, struct dt_object *dt,
738 const struct lu_attr *attr, struct thandle *th,
739 struct lustre_capa *capa)
741 struct osp_object *o = dt2osp_obj(dt);
745 if (is_only_remote_trans(th)) {
746 rc = __osp_attr_set(env, dt, attr, th);
751 /* we're interested in uid/gid changes only */
752 if (!(attr->la_valid & (LA_UID | LA_GID)))
755 /* new object, the very first ->attr_set()
756 * initializing attributes needs no logging
757 * all subsequent one are subject to the
758 * logging and synchronization with OST */
764 if (!is_only_remote_trans(th))
766 * once transaction is committed put proper command on
767 * the queue going to our OST
769 rc = osp_sync_add(env, o, MDS_SETATTR64_REC, th, attr);
770 /* XXX: send new uid/gid to OST ASAP? */
772 /* It is for OST-object attr_set directly without updating
773 * local MDT-object attribute. It is usually used by LFSCK. */
774 rc = osp_md_attr_set(env, dt, attr, th, capa);
780 * Interpreter function for getting OSP object extended attribute asynchronously
782 * Called to interpret the result of an async mode RPC for getting the
783 * OSP object extended attribute.
785 * \param[in] env pointer to the thread context
786 * \param[in] reply pointer to the RPC reply
787 * \param[in] req pointer to the RPC request
788 * \param[in] obj pointer to the OSP object
789 * \param[out] data pointer to OSP object attributes cache
790 * \param[in] index the index of the attribute buffer in the reply
791 * \param[in] rc the result for handling the RPC
793 * \retval 0 for success
794 * \retval negative error number on failure
796 static int osp_xattr_get_interpterer(const struct lu_env *env,
797 struct object_update_reply *reply,
798 struct ptlrpc_request *req,
799 struct osp_object *obj,
800 void *data, int index, int rc)
802 struct osp_object_attr *ooa = obj->opo_ooa;
803 struct osp_xattr_entry *oxe = data;
804 struct lu_buf *rbuf = &osp_env_info(env)->osi_lb2;
806 LASSERT(ooa != NULL);
809 size_t len = sizeof(*oxe) + oxe->oxe_namelen + 1;
811 rc = object_update_result_data_get(reply, rbuf, index);
812 if (rc < 0 || rbuf->lb_len > (oxe->oxe_buflen - len)) {
813 spin_lock(&obj->opo_lock);
815 spin_unlock(&obj->opo_lock);
816 osp_oac_xattr_put(oxe);
818 return rc < 0 ? rc : -ERANGE;
821 spin_lock(&obj->opo_lock);
822 oxe->oxe_vallen = rbuf->lb_len;
823 memcpy(oxe->oxe_value, rbuf->lb_buf, rbuf->lb_len);
826 spin_unlock(&obj->opo_lock);
827 } else if (rc == -ENOENT || rc == -ENODATA) {
828 spin_lock(&obj->opo_lock);
831 spin_unlock(&obj->opo_lock);
833 spin_lock(&obj->opo_lock);
835 spin_unlock(&obj->opo_lock);
838 osp_oac_xattr_put(oxe);
844 * Implement OSP dt_object_operations::do_declare_xattr_get() interface.
846 * Declare that the caller will get extended attribute from the specified
849 * This function will add an OUT_XATTR_GET sub-request to the per OSP
850 * based shared asynchronous request queue with the interpreter function:
851 * osp_xattr_get_interpterer().
853 * \param[in] env pointer to the thread context
854 * \param[in] dt pointer to the OSP layer dt_object
855 * \param[out] buf pointer to the lu_buf to hold the extended attribute
856 * \param[in] name the name for the expected extended attribute
857 * \param[in] capa the capability for this operation
859 * \retval 0 for success
860 * \retval negative error number on failure
862 static int osp_declare_xattr_get(const struct lu_env *env, struct dt_object *dt,
863 struct lu_buf *buf, const char *name,
864 struct lustre_capa *capa)
866 struct osp_object *obj = dt2osp_obj(dt);
867 struct osp_device *osp = lu2osp_dev(dt->do_lu.lo_dev);
868 struct osp_xattr_entry *oxe;
869 __u16 namelen = strlen(name);
872 LASSERT(buf != NULL);
873 LASSERT(name != NULL);
875 /* If only for xattr size, return directly. */
876 if (unlikely(buf->lb_len == 0))
879 if (obj->opo_ooa == NULL) {
880 rc = osp_oac_init(obj);
885 oxe = osp_oac_xattr_find_or_add(obj, name, buf->lb_len);
889 mutex_lock(&osp->opd_async_requests_mutex);
890 rc = osp_insert_async_request(env, OUT_XATTR_GET, obj, 1,
891 &namelen, (const void **)&name, oxe,
892 osp_xattr_get_interpterer);
894 mutex_unlock(&osp->opd_async_requests_mutex);
895 osp_oac_xattr_put(oxe);
897 struct dt_update_request *update;
899 /* XXX: Currently, we trigger the batched async OUT
900 * RPC via dt_declare_xattr_get(). It is not
901 * perfect solution, but works well now.
903 * We will improve it in the future. */
904 update = osp->opd_async_requests;
905 if (update != NULL && update->dur_buf.ub_req != NULL &&
906 update->dur_buf.ub_req->ourq_count > 0) {
907 osp->opd_async_requests = NULL;
908 mutex_unlock(&osp->opd_async_requests_mutex);
909 rc = osp_unplug_async_request(env, osp, update);
911 mutex_unlock(&osp->opd_async_requests_mutex);
919 * Implement OSP layer dt_object_operations::do_xattr_get() interface.
921 * Get extended attribute from the specified MDT/OST object.
923 * If the extended attribute is in the OSP object attributes cache, then
924 * return the cached extended attribute directly. Otherwise it will get
925 * the extended attribute synchronously, if successful, add it to the OSP
926 * attributes cache. (\see lustre/osp/osp_trans.c for OUT RPC.)
928 * There is a race condition: some other thread has added the named extended
929 * attributed entry to the OSP object attributes cache during the current
930 * OUT_XATTR_GET handling. If such case happens, the OSP will replace the
931 * (just) existing extended attribute entry with the new replied one.
933 * \param[in] env pointer to the thread context
934 * \param[in] dt pointer to the OSP layer dt_object
935 * \param[out] buf pointer to the lu_buf to hold the extended attribute
936 * \param[in] name the name for the expected extended attribute
937 * \param[in] capa the capability for this operation
939 * \retval 0 for success
940 * \retval negative error number on failure
942 int osp_xattr_get(const struct lu_env *env, struct dt_object *dt,
943 struct lu_buf *buf, const char *name,
944 struct lustre_capa *capa)
946 struct osp_device *osp = lu2osp_dev(dt->do_lu.lo_dev);
947 struct osp_object *obj = dt2osp_obj(dt);
948 struct dt_device *dev = &osp->opd_dt_dev;
949 struct lu_buf *rbuf = &osp_env_info(env)->osi_lb2;
950 struct dt_update_request *update = NULL;
951 struct ptlrpc_request *req = NULL;
952 struct object_update_reply *reply;
953 struct osp_xattr_entry *oxe = NULL;
954 const char *dname = dt->do_lu.lo_dev->ld_obd->obd_name;
958 LASSERT(buf != NULL);
959 LASSERT(name != NULL);
961 if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_NETWORK) &&
962 osp->opd_index == cfs_fail_val) {
963 if (is_ost_obj(&dt->do_lu)) {
964 if (osp_dev2node(osp) == cfs_fail_val)
967 if (strcmp(name, XATTR_NAME_LINK) == 0)
972 if (unlikely(obj->opo_non_exist))
975 oxe = osp_oac_xattr_find(obj, name, false);
977 spin_lock(&obj->opo_lock);
978 if (oxe->oxe_ready) {
980 GOTO(unlock, rc = -ENODATA);
982 if (buf->lb_buf == NULL)
983 GOTO(unlock, rc = oxe->oxe_vallen);
985 if (buf->lb_len < oxe->oxe_vallen)
986 GOTO(unlock, rc = -ERANGE);
988 memcpy(buf->lb_buf, oxe->oxe_value, oxe->oxe_vallen);
990 GOTO(unlock, rc = oxe->oxe_vallen);
993 spin_unlock(&obj->opo_lock);
994 osp_oac_xattr_put(oxe);
998 spin_unlock(&obj->opo_lock);
1001 update = dt_update_request_create(dev);
1003 GOTO(out, rc = PTR_ERR(update));
1005 rc = out_xattr_get_pack(env, &update->dur_buf,
1006 lu_object_fid(&dt->do_lu), name);
1008 CERROR("%s: Insert update error "DFID": rc = %d\n",
1009 dname, PFID(lu_object_fid(&dt->do_lu)), rc);
1013 rc = out_remote_sync(env, osp->opd_obd->u.cli.cl_import, update, &req);
1015 if (rc == -ENOENT) {
1016 dt->do_lu.lo_header->loh_attr &= ~LOHA_EXISTS;
1017 obj->opo_non_exist = 1;
1020 if (obj->opo_ooa == NULL)
1024 oxe = osp_oac_xattr_find_or_add(obj, name, buf->lb_len);
1027 CWARN("%s: Fail to add xattr (%s) to cache for "
1028 DFID" (1): rc = %d\n", dname, name,
1029 PFID(lu_object_fid(&dt->do_lu)), rc);
1034 spin_lock(&obj->opo_lock);
1035 if (rc == -ENOENT || rc == -ENODATA) {
1041 spin_unlock(&obj->opo_lock);
1046 reply = req_capsule_server_sized_get(&req->rq_pill,
1047 &RMF_OUT_UPDATE_REPLY,
1048 OUT_UPDATE_REPLY_SIZE);
1049 if (reply->ourp_magic != UPDATE_REPLY_MAGIC) {
1050 CERROR("%s: Wrong version %x expected %x "DFID": rc = %d\n",
1051 dname, reply->ourp_magic, UPDATE_REPLY_MAGIC,
1052 PFID(lu_object_fid(&dt->do_lu)), -EPROTO);
1054 GOTO(out, rc = -EPROTO);
1057 rc = object_update_result_data_get(reply, rbuf, 0);
1061 if (buf->lb_buf == NULL)
1062 GOTO(out, rc = rbuf->lb_len);
1064 if (unlikely(buf->lb_len < rbuf->lb_len))
1065 GOTO(out, rc = -ERANGE);
1067 memcpy(buf->lb_buf, rbuf->lb_buf, rbuf->lb_len);
1069 if (obj->opo_ooa == NULL)
1073 oxe = osp_oac_xattr_find_or_add(obj, name, rbuf->lb_len);
1075 CWARN("%s: Fail to add xattr (%s) to "
1076 "cache for "DFID" (2): rc = %d\n",
1077 dname, name, PFID(lu_object_fid(&dt->do_lu)), rc);
1083 if (oxe->oxe_buflen - oxe->oxe_namelen - 1 < rbuf->lb_len) {
1084 struct osp_xattr_entry *old = oxe;
1085 struct osp_xattr_entry *tmp;
1087 tmp = osp_oac_xattr_replace(obj, &old, rbuf->lb_len);
1088 osp_oac_xattr_put(oxe);
1091 CWARN("%s: Fail to update xattr (%s) to "
1092 "cache for "DFID": rc = %d\n",
1093 dname, name, PFID(lu_object_fid(&dt->do_lu)), rc);
1094 spin_lock(&obj->opo_lock);
1096 spin_unlock(&obj->opo_lock);
1101 /* Drop the ref for entry on list. */
1102 osp_oac_xattr_put(old);
1105 spin_lock(&obj->opo_lock);
1106 oxe->oxe_vallen = rbuf->lb_len;
1107 memcpy(oxe->oxe_value, rbuf->lb_buf, rbuf->lb_len);
1110 spin_unlock(&obj->opo_lock);
1116 ptlrpc_req_finished(req);
1118 if (update != NULL && !IS_ERR(update))
1119 dt_update_request_destroy(update);
1122 osp_oac_xattr_put(oxe);
1127 static int __osp_xattr_set(const struct lu_env *env, struct dt_object *dt,
1128 const struct lu_buf *buf, const char *name,
1129 int flag, struct thandle *th)
1131 struct osp_object *o = dt2osp_obj(dt);
1132 struct dt_update_request *update;
1133 struct osp_xattr_entry *oxe;
1137 LASSERT(buf->lb_len > 0 && buf->lb_buf != NULL);
1139 update = dt_update_request_find_or_create(th, dt);
1140 if (IS_ERR(update)) {
1141 CERROR("%s: Get OSP update buf failed "DFID": rc = %d\n",
1142 dt->do_lu.lo_dev->ld_obd->obd_name,
1143 PFID(lu_object_fid(&dt->do_lu)),
1144 (int)PTR_ERR(update));
1146 RETURN(PTR_ERR(update));
1149 rc = out_xattr_set_pack(env, &update->dur_buf,
1150 lu_object_fid(&dt->do_lu),
1151 buf, name, flag, update->dur_batchid);
1152 if (rc != 0 || o->opo_ooa == NULL)
1155 oxe = osp_oac_xattr_find_or_add(o, name, buf->lb_len);
1157 CWARN("%s: Fail to add xattr (%s) to cache for "DFID,
1158 dt->do_lu.lo_dev->ld_obd->obd_name,
1159 name, PFID(lu_object_fid(&dt->do_lu)));
1164 if (oxe->oxe_buflen - oxe->oxe_namelen - 1 < buf->lb_len) {
1165 struct osp_xattr_entry *old = oxe;
1166 struct osp_xattr_entry *tmp;
1168 tmp = osp_oac_xattr_replace(o, &old, buf->lb_len);
1169 osp_oac_xattr_put(oxe);
1172 CWARN("%s: Fail to update xattr (%s) to cache for "DFID,
1173 dt->do_lu.lo_dev->ld_obd->obd_name,
1174 name, PFID(lu_object_fid(&dt->do_lu)));
1175 spin_lock(&o->opo_lock);
1177 spin_unlock(&o->opo_lock);
1182 /* Drop the ref for entry on list. */
1183 osp_oac_xattr_put(old);
1186 spin_lock(&o->opo_lock);
1187 oxe->oxe_vallen = buf->lb_len;
1188 memcpy(oxe->oxe_value, buf->lb_buf, buf->lb_len);
1191 spin_unlock(&o->opo_lock);
1192 osp_oac_xattr_put(oxe);
1198 * Implement OSP layer dt_object_operations::do_declare_xattr_set() interface.
1200 * Declare that the caller will set extended attribute to the specified
1203 * This function will add an OUT_XATTR_SET sub-request to the per
1204 * OSP-transaction based request queue which will be flushed when
1205 * the transaction starts.
1207 * \param[in] env pointer to the thread context
1208 * \param[in] dt pointer to the OSP layer dt_object
1209 * \param[in] buf pointer to the lu_buf to hold the extended attribute
1210 * \param[in] name the name of the extended attribute to be set
1211 * \param[in] flag to indicate the detailed set operation: LU_XATTR_CREATE
1212 * or LU_XATTR_REPLACE or others
1213 * \param[in] th pointer to the transaction handler
1215 * \retval 0 for success
1216 * \retval negative error number on failure
1218 int osp_declare_xattr_set(const struct lu_env *env, struct dt_object *dt,
1219 const struct lu_buf *buf, const char *name,
1220 int flag, struct thandle *th)
1224 /* Please check the comment in osp_attr_set() for handling
1225 * remote transaction. */
1226 if (!is_only_remote_trans(th))
1227 rc = __osp_xattr_set(env, dt, buf, name, flag, th);
1233 * Implement OSP layer dt_object_operations::do_xattr_set() interface.
1235 * Set extended attribute to the specified MDT/OST object.
1237 * The real modification sub-request has been added in the declare phase
1238 * and related (OUT) RPC has been triggered when transaction start.
1240 * If the OSP attributes cache is initialized, then check whether the name
1241 * extended attribute entry exists in the cache or not. If yes, replace it;
1242 * otherwise, add the extended attribute to the cache.
1244 * \param[in] env pointer to the thread context
1245 * \param[in] dt pointer to the OSP layer dt_object
1246 * \param[in] buf pointer to the lu_buf to hold the extended attribute
1247 * \param[in] name the name of the extended attribute to be set
1248 * \param[in] fl to indicate the detailed set operation: LU_XATTR_CREATE
1249 * or LU_XATTR_REPLACE or others
1250 * \param[in] th pointer to the transaction handler
1251 * \param[in] capa the capability for this operation
1253 * \retval 0 for success
1254 * \retval negative error number on failure
1256 int osp_xattr_set(const struct lu_env *env, struct dt_object *dt,
1257 const struct lu_buf *buf, const char *name, int fl,
1258 struct thandle *th, struct lustre_capa *capa)
1262 CDEBUG(D_INFO, "xattr %s set object "DFID"\n", name,
1263 PFID(&dt->do_lu.lo_header->loh_fid));
1265 /* Please check the comment in osp_attr_set() for handling
1266 * remote transaction. */
1267 if (is_only_remote_trans(th))
1268 rc = __osp_xattr_set(env, dt, buf, name, fl, th);
1273 static int __osp_xattr_del(const struct lu_env *env, struct dt_object *dt,
1274 const char *name, struct thandle *th)
1276 struct dt_update_request *update;
1277 const struct lu_fid *fid;
1278 struct osp_object *o = dt2osp_obj(dt);
1279 struct osp_xattr_entry *oxe;
1282 update = dt_update_request_find_or_create(th, dt);
1284 return PTR_ERR(update);
1286 fid = lu_object_fid(&dt->do_lu);
1288 rc = out_xattr_del_pack(env, &update->dur_buf, fid, name,
1289 update->dur_batchid);
1291 if (rc != 0 || o->opo_ooa == NULL)
1294 oxe = osp_oac_xattr_find(o, name, true);
1296 /* Drop the ref for entry on list. */
1297 osp_oac_xattr_put(oxe);
1303 * Implement OSP layer dt_object_operations::do_declare_xattr_del() interface.
1305 * Declare that the caller will delete extended attribute on the specified
1308 * This function will add an OUT_XATTR_DEL sub-request to the per
1309 * OSP-transaction based request queue which will be flushed when
1310 * transaction start.
1312 * \param[in] env pointer to the thread context
1313 * \param[in] dt pointer to the OSP layer dt_object
1314 * \param[in] name the name of the extended attribute to be set
1315 * \param[in] th pointer to the transaction handler
1317 * \retval 0 for success
1318 * \retval negative error number on failure
1320 int osp_declare_xattr_del(const struct lu_env *env, struct dt_object *dt,
1321 const char *name, struct thandle *th)
1325 /* Please check the comment in osp_attr_set() for handling
1326 * remote transaction. */
1327 if (!is_only_remote_trans(th))
1328 rc = __osp_xattr_del(env, dt, name, th);
1334 * Implement OSP layer dt_object_operations::do_xattr_del() interface.
1336 * Delete extended attribute on the specified MDT/OST object.
1338 * The real modification sub-request has been added in the declare phase
1339 * and related (OUT) RPC has been triggered when transaction start.
1341 * If the name extended attribute entry exists in the OSP attributes
1342 * cache, then remove it from the cache.
1344 * \param[in] env pointer to the thread context
1345 * \param[in] dt pointer to the OSP layer dt_object
1346 * \param[in] name the name of the extended attribute to be set
1347 * \param[in] th pointer to the transaction handler
1348 * \param[in] capa the capability for this operation
1350 * \retval 0 for success
1351 * \retval negative error number on failure
1353 int osp_xattr_del(const struct lu_env *env, struct dt_object *dt,
1354 const char *name, struct thandle *th,
1355 struct lustre_capa *capa)
1359 CDEBUG(D_INFO, "xattr %s del object "DFID"\n", name,
1360 PFID(&dt->do_lu.lo_header->loh_fid));
1362 /* Please check the comment in osp_attr_set() for handling
1363 * remote transaction. */
1364 if (is_only_remote_trans(th))
1365 rc = __osp_xattr_del(env, dt, name, th);
1371 * Implement OSP layer dt_object_operations::do_declare_create() interface.
1373 * Declare that the caller will create the OST object.
1375 * If the transaction is a remote transaction (please refer to the
1376 * comment of osp_trans_create() for remote transaction), then the FID
1377 * for the OST object has been assigned already, and will be handled
1378 * as create (remote) MDT object via osp_md_declare_object_create().
1379 * This function is usually used for LFSCK to re-create the lost OST
1380 * object. Otherwise, if it is not replay case, the OSP will reserve
1381 * pre-created object for the subsequent create operation; if the MDT
1382 * side cached pre-created objects are less than some threshold, then
1383 * it will wakeup the pre-create thread.
1385 * \param[in] env pointer to the thread context
1386 * \param[in] dt pointer to the OSP layer dt_object
1387 * \param[in] attr the attribute for the object to be created
1388 * \param[in] hint pointer to the hint for creating the object, such as
1390 * \param[in] dof pointer to the dt_object_format for help the creation
1391 * \param[in] th pointer to the transaction handler
1393 * \retval 0 for success
1394 * \retval negative error number on failure
1396 static int osp_declare_object_create(const struct lu_env *env,
1397 struct dt_object *dt,
1398 struct lu_attr *attr,
1399 struct dt_allocation_hint *hint,
1400 struct dt_object_format *dof,
1403 struct osp_thread_info *osi = osp_env_info(env);
1404 struct osp_device *d = lu2osp_dev(dt->do_lu.lo_dev);
1405 struct osp_object *o = dt2osp_obj(dt);
1406 const struct lu_fid *fid = lu_object_fid(&dt->do_lu);
1411 if (is_only_remote_trans(th)) {
1412 LASSERT(fid_is_sane(fid));
1414 rc = osp_md_declare_object_create(env, dt, attr, hint, dof, th);
1419 /* should happen to non-0 OSP only so that at least one object
1420 * has been already declared in the scenario and LOD should
1422 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_CREATE_FAIL) && d->opd_index == 1)
1425 LASSERT(d->opd_last_used_oid_file);
1428 * There can be gaps in precreated ids and record to unlink llog
1429 * XXX: we do not handle gaps yet, implemented before solution
1430 * was found to be racy, so we disabled that. there is no
1431 * point in making useless but expensive llog declaration.
1433 /* rc = osp_sync_declare_add(env, o, MDS_UNLINK64_REC, th); */
1435 if (unlikely(!fid_is_zero(fid))) {
1436 /* replay case: caller knows fid */
1437 osi->osi_off = sizeof(osi->osi_id) * d->opd_index;
1438 osi->osi_lb.lb_len = sizeof(osi->osi_id);
1439 osi->osi_lb.lb_buf = NULL;
1440 rc = dt_declare_record_write(env, d->opd_last_used_oid_file,
1441 &osi->osi_lb, osi->osi_off, th);
1446 * in declaration we need to reserve object so that we don't block
1447 * awaiting precreation RPC to complete
1449 rc = osp_precreate_reserve(env, d);
1451 * we also need to declare update to local "last used id" file for
1452 * recovery if object isn't used for a reason, we need to release
1453 * reservation, this can be made in osd_object_release()
1456 /* mark id is reserved: in create we don't want to talk
1458 LASSERT(o->opo_reserved == 0);
1459 o->opo_reserved = 1;
1461 /* common for all OSPs file hystorically */
1462 osi->osi_off = sizeof(osi->osi_id) * d->opd_index;
1463 osi->osi_lb.lb_len = sizeof(osi->osi_id);
1464 osi->osi_lb.lb_buf = NULL;
1465 rc = dt_declare_record_write(env, d->opd_last_used_oid_file,
1466 &osi->osi_lb, osi->osi_off, th);
1468 /* not needed in the cache anymore */
1469 set_bit(LU_OBJECT_HEARD_BANSHEE,
1470 &dt->do_lu.lo_header->loh_flags);
1476 * Implement OSP layer dt_object_operations::do_create() interface.
1478 * Create the OST object.
1480 * For remote transaction case, the real create sub-request has been
1481 * added in the declare phase and related (OUT) RPC has been triggered
1482 * at transaction start. Here, like creating (remote) MDT object, the
1483 * OSP will mark the object existence via osp_md_object_create().
1485 * For non-remote transaction case, the OSP will assign FID to the
1486 * object to be created, and update last_used Object ID (OID) file.
1488 * \param[in] env pointer to the thread context
1489 * \param[in] dt pointer to the OSP layer dt_object
1490 * \param[in] attr the attribute for the object to be created
1491 * \param[in] hint pointer to the hint for creating the object, such as
1493 * \param[in] dof pointer to the dt_object_format for help the creation
1494 * \param[in] th pointer to the transaction handler
1496 * \retval 0 for success
1497 * \retval negative error number on failure
1499 static int osp_object_create(const struct lu_env *env, struct dt_object *dt,
1500 struct lu_attr *attr,
1501 struct dt_allocation_hint *hint,
1502 struct dt_object_format *dof, struct thandle *th)
1504 struct osp_thread_info *osi = osp_env_info(env);
1505 struct osp_device *d = lu2osp_dev(dt->do_lu.lo_dev);
1506 struct osp_object *o = dt2osp_obj(dt);
1508 struct lu_fid *fid = &osi->osi_fid;
1511 if (is_only_remote_trans(th)) {
1512 LASSERT(fid_is_sane(lu_object_fid(&dt->do_lu)));
1514 rc = osp_md_object_create(env, dt, attr, hint, dof, th);
1516 o->opo_non_exist = 0;
1521 o->opo_non_exist = 0;
1522 if (o->opo_reserved) {
1523 /* regular case, fid is assigned holding transaction open */
1524 osp_object_assign_fid(env, d, o);
1527 memcpy(fid, lu_object_fid(&dt->do_lu), sizeof(*fid));
1529 LASSERTF(fid_is_sane(fid), "fid for osp_object %p is insane"DFID"!\n",
1532 if (!o->opo_reserved) {
1533 /* special case, id was assigned outside of transaction
1534 * see comments in osp_declare_attr_set */
1535 LASSERT(d->opd_pre != NULL);
1536 spin_lock(&d->opd_pre_lock);
1537 osp_update_last_fid(d, fid);
1538 spin_unlock(&d->opd_pre_lock);
1541 CDEBUG(D_INODE, "fid for osp_object %p is "DFID"\n", o, PFID(fid));
1543 /* If the precreate ends, it means it will be ready to rollover to
1544 * the new sequence soon, all the creation should be synchronized,
1545 * otherwise during replay, the replay fid will be inconsistent with
1546 * last_used/create fid */
1547 if (osp_precreate_end_seq(env, d) && osp_is_fid_client(d))
1551 * it's OK if the import is inactive by this moment - id was created
1552 * by OST earlier, we just need to maintain it consistently on the disk
1553 * once import is reconnected, OSP will claim this and other objects
1554 * used and OST either keep them, if they exist or recreate
1557 /* we might have lost precreated objects */
1558 if (unlikely(d->opd_gap_count) > 0) {
1559 LASSERT(d->opd_pre != NULL);
1560 spin_lock(&d->opd_pre_lock);
1561 if (d->opd_gap_count > 0) {
1562 int count = d->opd_gap_count;
1564 ostid_set_id(&osi->osi_oi,
1565 fid_oid(&d->opd_gap_start_fid));
1566 d->opd_gap_count = 0;
1567 spin_unlock(&d->opd_pre_lock);
1569 CDEBUG(D_HA, "Writing gap "DFID"+%d in llog\n",
1570 PFID(&d->opd_gap_start_fid), count);
1571 /* real gap handling is disabled intil ORI-692 will be
1572 * fixed, now we only report gaps */
1574 spin_unlock(&d->opd_pre_lock);
1578 /* new object, the very first ->attr_set()
1579 * initializing attributes needs no logging */
1582 /* Only need update last_used oid file, seq file will only be update
1583 * during seq rollover */
1584 osp_objid_buf_prep(&osi->osi_lb, &osi->osi_off,
1585 &d->opd_last_used_fid.f_oid, d->opd_index);
1587 rc = dt_record_write(env, d->opd_last_used_oid_file, &osi->osi_lb,
1590 CDEBUG(D_HA, "%s: Wrote last used FID: "DFID", index %d: %d\n",
1591 d->opd_obd->obd_name, PFID(fid), d->opd_index, rc);
1597 * Implement OSP layer dt_object_operations::do_declare_destroy() interface.
1599 * Declare that the caller will destroy the specified OST object.
1601 * The OST object destroy will be handled via llog asynchronously. This
1602 * function will declare the credits for generating MDS_UNLINK64_REC llog.
1604 * \param[in] env pointer to the thread context
1605 * \param[in] dt pointer to the OSP layer dt_object to be destroyed
1606 * \param[in] th pointer to the transaction handler
1608 * \retval 0 for success
1609 * \retval negative error number on failure
1611 int osp_declare_object_destroy(const struct lu_env *env,
1612 struct dt_object *dt, struct thandle *th)
1614 struct osp_object *o = dt2osp_obj(dt);
1620 * track objects to be destroyed via llog
1622 rc = osp_sync_declare_add(env, o, MDS_UNLINK64_REC, th);
1628 * Implement OSP layer dt_object_operations::do_destroy() interface.
1630 * Destroy the specified OST object.
1632 * The OSP generates a MDS_UNLINK64_REC record in the llog. There
1633 * will be some dedicated thread to handle the llog asynchronously.
1635 * It also marks the object as non-cached.
1637 * \param[in] env pointer to the thread context
1638 * \param[in] dt pointer to the OSP layer dt_object to be destroyed
1639 * \param[in] th pointer to the transaction handler
1641 * \retval 0 for success
1642 * \retval negative error number on failure
1644 int osp_object_destroy(const struct lu_env *env, struct dt_object *dt,
1647 struct osp_object *o = dt2osp_obj(dt);
1652 o->opo_non_exist = 1;
1654 * once transaction is committed put proper command on
1655 * the queue going to our OST
1657 rc = osp_sync_add(env, o, MDS_UNLINK64_REC, th, NULL);
1659 /* not needed in cache any more */
1660 set_bit(LU_OBJECT_HEARD_BANSHEE, &dt->do_lu.lo_header->loh_flags);
1665 static int osp_orphan_index_lookup(const struct lu_env *env,
1666 struct dt_object *dt,
1668 const struct dt_key *key,
1669 struct lustre_capa *capa)
1674 static int osp_orphan_index_declare_insert(const struct lu_env *env,
1675 struct dt_object *dt,
1676 const struct dt_rec *rec,
1677 const struct dt_key *key,
1678 struct thandle *handle)
1683 static int osp_orphan_index_insert(const struct lu_env *env,
1684 struct dt_object *dt,
1685 const struct dt_rec *rec,
1686 const struct dt_key *key,
1687 struct thandle *handle,
1688 struct lustre_capa *capa,
1694 static int osp_orphan_index_declare_delete(const struct lu_env *env,
1695 struct dt_object *dt,
1696 const struct dt_key *key,
1697 struct thandle *handle)
1702 static int osp_orphan_index_delete(const struct lu_env *env,
1703 struct dt_object *dt,
1704 const struct dt_key *key,
1705 struct thandle *handle,
1706 struct lustre_capa *capa)
1712 * Initialize the OSP layer index iteration.
1714 * \param[in] env pointer to the thread context
1715 * \param[in] dt pointer to the index object to be iterated
1716 * \param[in] attr unused
1717 * \param[in] capa the capability for this operation
1719 * \retval pointer to the iteration structure
1720 * \retval negative error number on failure
1722 struct dt_it *osp_it_init(const struct lu_env *env, struct dt_object *dt,
1723 __u32 attr, struct lustre_capa *capa)
1729 return ERR_PTR(-ENOMEM);
1731 it->ooi_pos_ent = -1;
1734 return (struct dt_it *)it;
1738 * Finalize the OSP layer index iteration.
1740 * \param[in] env pointer to the thread context
1741 * \param[in] di pointer to the iteration structure
1743 void osp_it_fini(const struct lu_env *env, struct dt_it *di)
1745 struct osp_it *it = (struct osp_it *)di;
1746 struct page **pages = it->ooi_pages;
1747 int npages = it->ooi_total_npages;
1750 if (pages != NULL) {
1751 for (i = 0; i < npages; i++) {
1752 if (pages[i] != NULL) {
1753 if (pages[i] == it->ooi_cur_page) {
1755 it->ooi_cur_page = NULL;
1757 __free_page(pages[i]);
1760 OBD_FREE(pages, npages * sizeof(*pages));
1766 * Get more records for the iteration from peer.
1768 * The new records will be filled in an array of pages. The OSP side
1769 * allows 1MB bulk data to be transfered.
1771 * \param[in] env pointer to the thread context
1772 * \param[in] it pointer to the iteration structure
1774 * \retval 0 for success
1775 * \retval negative error number on failure
1777 static int osp_it_fetch(const struct lu_env *env, struct osp_it *it)
1779 struct lu_device *dev = it->ooi_obj->do_lu.lo_dev;
1780 struct osp_device *osp = lu2osp_dev(dev);
1781 struct page **pages;
1782 struct ptlrpc_request *req = NULL;
1783 struct ptlrpc_bulk_desc *desc;
1784 struct idx_info *ii;
1791 npages = min_t(unsigned int, OFD_MAX_BRW_SIZE, 1 << 20);
1792 npages /= PAGE_CACHE_SIZE;
1794 OBD_ALLOC(pages, npages * sizeof(*pages));
1798 it->ooi_pages = pages;
1799 it->ooi_total_npages = npages;
1800 for (i = 0; i < npages; i++) {
1801 pages[i] = alloc_page(GFP_IOFS);
1802 if (pages[i] == NULL)
1806 req = ptlrpc_request_alloc(osp->opd_obd->u.cli.cl_import,
1811 rc = ptlrpc_request_pack(req, LUSTRE_OBD_VERSION, OBD_IDX_READ);
1813 ptlrpc_request_free(req);
1817 req->rq_request_portal = OUT_PORTAL;
1818 ii = req_capsule_client_get(&req->rq_pill, &RMF_IDX_INFO);
1819 memset(ii, 0, sizeof(*ii));
1820 if (fid_is_last_id(lu_object_fid(&it->ooi_obj->do_lu))) {
1821 /* LFSCK will iterate orphan object[FID_SEQ_LAYOUT_BTREE,
1822 * ost_index, 0] with LAST_ID FID, so it needs to replace
1823 * the FID with orphan FID here */
1824 ii->ii_fid.f_seq = FID_SEQ_LAYOUT_RBTREE;
1825 ii->ii_fid.f_oid = osp->opd_index;
1826 ii->ii_fid.f_ver = 0;
1827 ii->ii_flags = II_FL_NOHASH;
1829 ii->ii_fid = *lu_object_fid(&it->ooi_obj->do_lu);
1830 ii->ii_flags = II_FL_NOHASH | II_FL_NOKEY | II_FL_VARKEY |
1833 ii->ii_magic = IDX_INFO_MAGIC;
1834 ii->ii_count = npages * LU_PAGE_COUNT;
1835 ii->ii_hash_start = it->ooi_next;
1836 ii->ii_attrs = osp_dev2node(osp);
1838 ptlrpc_at_set_req_timeout(req);
1840 desc = ptlrpc_prep_bulk_imp(req, npages, 1, BULK_PUT_SINK,
1843 ptlrpc_request_free(req);
1847 for (i = 0; i < npages; i++)
1848 ptlrpc_prep_bulk_page_pin(desc, pages[i], 0, PAGE_CACHE_SIZE);
1850 ptlrpc_request_set_replen(req);
1851 rc = ptlrpc_queue_wait(req);
1855 rc = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk,
1856 req->rq_bulk->bd_nob_transferred);
1861 ii = req_capsule_server_get(&req->rq_pill, &RMF_IDX_INFO);
1862 if (ii->ii_magic != IDX_INFO_MAGIC)
1863 GOTO(out, rc = -EPROTO);
1865 npages = (ii->ii_count + LU_PAGE_COUNT - 1) >>
1866 (PAGE_CACHE_SHIFT - LU_PAGE_SHIFT);
1867 if (npages > it->ooi_total_npages) {
1868 CERROR("%s: returned more pages than expected, %u > %u\n",
1869 osp->opd_obd->obd_name, npages, it->ooi_total_npages);
1870 GOTO(out, rc = -EINVAL);
1873 it->ooi_valid_npages = npages;
1874 if (ptlrpc_rep_need_swab(req))
1877 it->ooi_next = ii->ii_hash_end;
1880 ptlrpc_req_finished(req);
1886 * Move the iteration cursor to the next lu_page.
1888 * One system page (PAGE_SIZE) may contain multiple lu_page (4KB),
1889 * that depends on the LU_PAGE_COUNT. If it is not the last lu_page
1890 * in current system page, then move the iteration cursor to the next
1891 * lu_page in current system page. Otherwise, if there are more system
1892 * pages in the cache, then move the iteration cursor to the next system
1893 * page. If all the cached records (pages) have been iterated, then fetch
1894 * more records via osp_it_fetch().
1896 * \param[in] env pointer to the thread context
1897 * \param[in] di pointer to the iteration structure
1899 * \retval 0 for success
1900 * \retval negative error number on failure
1902 int osp_it_next_page(const struct lu_env *env, struct dt_it *di)
1904 struct osp_it *it = (struct osp_it *)di;
1905 struct lu_idxpage *idxpage;
1906 struct page **pages;
1912 idxpage = it->ooi_cur_idxpage;
1913 if (idxpage != NULL) {
1914 if (idxpage->lip_nr == 0)
1917 if (it->ooi_pos_ent < idxpage->lip_nr) {
1918 CDEBUG(D_INFO, "ooi_pos %d nr %d\n",
1919 (int)it->ooi_pos_ent, (int)idxpage->lip_nr);
1922 it->ooi_cur_idxpage = NULL;
1923 it->ooi_pos_lu_page++;
1925 if (it->ooi_pos_lu_page < LU_PAGE_COUNT) {
1926 it->ooi_cur_idxpage = (void *)it->ooi_cur_page +
1927 LU_PAGE_SIZE * it->ooi_pos_lu_page;
1929 lustre_swab_lip_header(it->ooi_cur_idxpage);
1930 if (it->ooi_cur_idxpage->lip_magic != LIP_MAGIC) {
1931 struct osp_device *osp =
1932 lu2osp_dev(it->ooi_obj->do_lu.lo_dev);
1934 CERROR("%s: invalid magic (%x != %x) for page "
1935 "%d/%d while read layout orphan index\n",
1936 osp->opd_obd->obd_name,
1937 it->ooi_cur_idxpage->lip_magic,
1938 LIP_MAGIC, it->ooi_pos_page,
1939 it->ooi_pos_lu_page);
1940 /* Skip this lu_page next time. */
1941 it->ooi_pos_ent = idxpage->lip_nr - 1;
1944 it->ooi_pos_ent = -1;
1948 kunmap(it->ooi_cur_page);
1949 it->ooi_cur_page = NULL;
1953 pages = it->ooi_pages;
1954 if (it->ooi_pos_page < it->ooi_valid_npages) {
1955 it->ooi_cur_page = kmap(pages[it->ooi_pos_page]);
1956 it->ooi_pos_lu_page = 0;
1960 for (i = 0; i < it->ooi_total_npages; i++) {
1961 if (pages[i] != NULL)
1962 __free_page(pages[i]);
1964 OBD_FREE(pages, it->ooi_total_npages * sizeof(*pages));
1966 it->ooi_pos_page = 0;
1967 it->ooi_total_npages = 0;
1968 it->ooi_valid_npages = 0;
1971 it->ooi_cur_page = NULL;
1972 it->ooi_cur_idxpage = NULL;
1973 it->ooi_pages = NULL;
1976 if (it->ooi_next == II_END_OFF)
1979 rc = osp_it_fetch(env, it);
1987 * Move the iteration cursor to the next record.
1989 * If there are more records in the lu_page, then move the iteration
1990 * cursor to the next record directly. Otherwise, move the iteration
1991 * cursor to the record in the next lu_page via osp_it_next_page()
1993 * \param[in] env pointer to the thread context
1994 * \param[in] di pointer to the iteration structure
1996 * \retval 0 for success
1997 * \retval negative error number on failure
1999 int osp_orphan_it_next(const struct lu_env *env, struct dt_it *di)
2001 struct osp_it *it = (struct osp_it *)di;
2002 struct lu_idxpage *idxpage;
2007 idxpage = it->ooi_cur_idxpage;
2008 if (idxpage != NULL) {
2009 if (idxpage->lip_nr == 0)
2013 if (it->ooi_pos_ent < idxpage->lip_nr) {
2015 (struct lu_orphan_ent *)idxpage->lip_entries +
2018 lustre_swab_orphan_ent(it->ooi_ent);
2023 rc = osp_it_next_page(env, di);
2030 int osp_it_get(const struct lu_env *env, struct dt_it *di,
2031 const struct dt_key *key)
2036 void osp_it_put(const struct lu_env *env, struct dt_it *di)
2040 struct dt_key *osp_orphan_it_key(const struct lu_env *env,
2041 const struct dt_it *di)
2043 struct osp_it *it = (struct osp_it *)di;
2044 struct lu_orphan_ent *ent = (struct lu_orphan_ent *)it->ooi_ent;
2046 if (likely(ent != NULL))
2047 return (struct dt_key *)(&ent->loe_key);
2052 int osp_orphan_it_key_size(const struct lu_env *env, const struct dt_it *di)
2054 return sizeof(struct lu_fid);
2057 int osp_orphan_it_rec(const struct lu_env *env, const struct dt_it *di,
2058 struct dt_rec *rec, __u32 attr)
2060 struct osp_it *it = (struct osp_it *)di;
2061 struct lu_orphan_ent *ent = (struct lu_orphan_ent *)it->ooi_ent;
2063 if (likely(ent != NULL)) {
2064 *(struct lu_orphan_rec *)rec = ent->loe_rec;
2071 __u64 osp_it_store(const struct lu_env *env, const struct dt_it *di)
2073 struct osp_it *it = (struct osp_it *)di;
2075 return it->ooi_next;
2079 * Locate the iteration cursor to the specified position (cookie).
2081 * \param[in] env pointer to the thread context
2082 * \param[in] di pointer to the iteration structure
2083 * \param[in] hash the specified position
2085 * \retval positive number for locating to the exactly position
2087 * \retval 0 for arriving at the end of the iteration
2088 * \retval negative error number on failure
2090 int osp_orphan_it_load(const struct lu_env *env, const struct dt_it *di,
2093 struct osp_it *it = (struct osp_it *)di;
2096 it->ooi_next = hash;
2097 rc = osp_orphan_it_next(env, (struct dt_it *)di);
2107 int osp_it_key_rec(const struct lu_env *env, const struct dt_it *di,
2113 static const struct dt_index_operations osp_orphan_index_ops = {
2114 .dio_lookup = osp_orphan_index_lookup,
2115 .dio_declare_insert = osp_orphan_index_declare_insert,
2116 .dio_insert = osp_orphan_index_insert,
2117 .dio_declare_delete = osp_orphan_index_declare_delete,
2118 .dio_delete = osp_orphan_index_delete,
2120 .init = osp_it_init,
2121 .fini = osp_it_fini,
2122 .next = osp_orphan_it_next,
2125 .key = osp_orphan_it_key,
2126 .key_size = osp_orphan_it_key_size,
2127 .rec = osp_orphan_it_rec,
2128 .store = osp_it_store,
2129 .load = osp_orphan_it_load,
2130 .key_rec = osp_it_key_rec,
2135 * Implement OSP layer dt_object_operations::do_index_try() interface.
2137 * Negotiate the index type.
2139 * If the target index is an IDIF object, then use osp_orphan_index_ops.
2140 * Otherwise, assign osp_md_index_ops to the dt_object::do_index_ops.
2141 * (\see lustre/include/lustre_fid.h for IDIF.)
2143 * \param[in] env pointer to the thread context
2144 * \param[in] dt pointer to the OSP layer dt_object
2145 * \param[in] feat unused
2147 * \retval 0 for success
2149 static int osp_index_try(const struct lu_env *env,
2150 struct dt_object *dt,
2151 const struct dt_index_features *feat)
2153 const struct lu_fid *fid = lu_object_fid(&dt->do_lu);
2155 if (fid_is_last_id(fid) && fid_is_idif(fid))
2156 dt->do_index_ops = &osp_orphan_index_ops;
2158 dt->do_index_ops = &osp_md_index_ops;
2162 struct dt_object_operations osp_obj_ops = {
2163 .do_declare_attr_get = osp_declare_attr_get,
2164 .do_attr_get = osp_attr_get,
2165 .do_declare_attr_set = osp_declare_attr_set,
2166 .do_attr_set = osp_attr_set,
2167 .do_declare_xattr_get = osp_declare_xattr_get,
2168 .do_xattr_get = osp_xattr_get,
2169 .do_declare_xattr_set = osp_declare_xattr_set,
2170 .do_xattr_set = osp_xattr_set,
2171 .do_declare_create = osp_declare_object_create,
2172 .do_create = osp_object_create,
2173 .do_declare_destroy = osp_declare_object_destroy,
2174 .do_destroy = osp_object_destroy,
2175 .do_index_try = osp_index_try,
2179 * Implement OSP layer lu_object_operations::loo_object_init() interface.
2181 * Initialize the object.
2183 * If it is a remote MDT object, then call do_attr_get() to fetch
2184 * the attribute from the peer.
2186 * \param[in] env pointer to the thread context
2187 * \param[in] o pointer to the OSP layer lu_object
2188 * \param[in] conf unused
2190 * \retval 0 for success
2191 * \retval negative error number on failure
2193 static int osp_object_init(const struct lu_env *env, struct lu_object *o,
2194 const struct lu_object_conf *conf)
2196 struct osp_object *po = lu2osp_obj(o);
2200 spin_lock_init(&po->opo_lock);
2201 o->lo_header->loh_attr |= LOHA_REMOTE;
2203 if (is_ost_obj(o)) {
2204 po->opo_obj.do_ops = &osp_obj_ops;
2206 struct lu_attr *la = &osp_env_info(env)->osi_attr;
2208 po->opo_obj.do_ops = &osp_md_obj_ops;
2209 po->opo_obj.do_body_ops = &osp_md_body_ops;
2210 rc = po->opo_obj.do_ops->do_attr_get(env, lu2dt_obj(o),
2213 o->lo_header->loh_attr |=
2214 LOHA_EXISTS | (la->la_mode & S_IFMT);
2215 if (rc == -ENOENT) {
2216 po->opo_non_exist = 1;
2219 init_rwsem(&po->opo_sem);
2225 * Implement OSP layer lu_object_operations::loo_object_free() interface.
2227 * Finalize the object.
2229 * If the OSP object has attributes cache, then destroy the cache.
2230 * Free the object finally.
2232 * \param[in] env pointer to the thread context
2233 * \param[in] o pointer to the OSP layer lu_object
2235 static void osp_object_free(const struct lu_env *env, struct lu_object *o)
2237 struct osp_object *obj = lu2osp_obj(o);
2238 struct lu_object_header *h = o->lo_header;
2240 dt_object_fini(&obj->opo_obj);
2241 lu_object_header_fini(h);
2242 if (obj->opo_ooa != NULL) {
2243 struct osp_xattr_entry *oxe;
2244 struct osp_xattr_entry *tmp;
2247 list_for_each_entry_safe(oxe, tmp,
2248 &obj->opo_ooa->ooa_xattr_list,
2250 list_del(&oxe->oxe_list);
2251 count = atomic_read(&oxe->oxe_ref);
2252 LASSERTF(count == 1,
2253 "Still has %d users on the xattr entry %.*s\n",
2254 count-1, (int)oxe->oxe_namelen, oxe->oxe_buf);
2256 OBD_FREE(oxe, oxe->oxe_buflen);
2258 OBD_FREE_PTR(obj->opo_ooa);
2260 OBD_SLAB_FREE_PTR(obj, osp_object_kmem);
2264 * Implement OSP layer lu_object_operations::loo_object_release() interface.
2266 * Cleanup (not free) the object.
2268 * If it is a reserved object but failed to be created, or it is an OST
2269 * object, then mark the object as non-cached.
2271 * \param[in] env pointer to the thread context
2272 * \param[in] o pointer to the OSP layer lu_object
2274 static void osp_object_release(const struct lu_env *env, struct lu_object *o)
2276 struct osp_object *po = lu2osp_obj(o);
2277 struct osp_device *d = lu2osp_dev(o->lo_dev);
2282 * release reservation if object was declared but not created
2283 * this may require lu_object_put() in LOD
2285 if (unlikely(po->opo_reserved)) {
2286 LASSERT(d->opd_pre != NULL);
2287 LASSERT(d->opd_pre_reserved > 0);
2288 spin_lock(&d->opd_pre_lock);
2289 d->opd_pre_reserved--;
2290 spin_unlock(&d->opd_pre_lock);
2292 /* not needed in cache any more */
2293 set_bit(LU_OBJECT_HEARD_BANSHEE, &o->lo_header->loh_flags);
2297 /* XXX: Currently, NOT cache OST-object on MDT because:
2298 * 1. it is not often accessed on MDT.
2299 * 2. avoid up layer (such as LFSCK) to load too many
2300 * once-used OST-objects. */
2301 set_bit(LU_OBJECT_HEARD_BANSHEE, &o->lo_header->loh_flags);
2306 static int osp_object_print(const struct lu_env *env, void *cookie,
2307 lu_printer_t p, const struct lu_object *l)
2309 const struct osp_object *o = lu2osp_obj((struct lu_object *)l);
2311 return (*p)(env, cookie, LUSTRE_OSP_NAME"-object@%p", o);
2314 static int osp_object_invariant(const struct lu_object *o)
2319 struct lu_object_operations osp_lu_obj_ops = {
2320 .loo_object_init = osp_object_init,
2321 .loo_object_free = osp_object_free,
2322 .loo_object_release = osp_object_release,
2323 .loo_object_print = osp_object_print,
2324 .loo_object_invariant = osp_object_invariant