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);
626 if (!(attr->la_valid & (LA_UID | LA_GID)))
629 if (!is_only_remote_trans(th)) {
631 * track all UID/GID changes via llog
633 rc = osp_sync_declare_add(env, o, MDS_SETATTR64_REC, th);
635 /* It is for OST-object attr_set directly without updating
636 * local MDT-object attribute. It is usually used by LFSCK. */
637 rc = __osp_md_attr_set(env, dt, attr, th);
640 if (rc != 0 || o->opo_ooa == NULL)
643 /* Update the OSP object attributes cache. */
644 la = &o->opo_ooa->ooa_attr;
645 spin_lock(&o->opo_lock);
646 if (attr->la_valid & LA_UID) {
647 la->la_uid = attr->la_uid;
648 la->la_valid |= LA_UID;
651 if (attr->la_valid & LA_GID) {
652 la->la_gid = attr->la_gid;
653 la->la_valid |= LA_GID;
655 spin_unlock(&o->opo_lock);
661 * Implement OSP layer dt_object_operations::do_declare_attr_set() interface.
662 * XXX: NOT prepare set_{attr,xattr} RPC for remote transaction.
664 * According to our current transaction/dt_object_lock framework (to make
665 * the cross-MDTs modification for DNE1 to be workable), the transaction
666 * sponsor will start the transaction firstly, then try to acquire related
667 * dt_object_lock if needed. Under such rules, if we want to prepare the
668 * set_{attr,xattr} RPC in the RPC declare phase, then related attr/xattr
669 * should be known without dt_object_lock. But such condition maybe not
670 * true for some remote transaction case. For example:
672 * For linkEA repairing (by LFSCK) case, before the LFSCK thread obtained
673 * the dt_object_lock on the target MDT-object, it cannot know whether
674 * the MDT-object has linkEA or not, neither invalid or not.
676 * Since the LFSCK thread cannot hold dt_object_lock before the (remote)
677 * transaction start (otherwise there will be some potential deadlock),
678 * it cannot prepare related RPC for repairing during the declare phase
679 * as other normal transactions do.
681 * To resolve the trouble, we will make OSP to prepare related RPC
682 * (set_attr/set_xattr/del_xattr) after remote transaction started,
683 * and trigger the remote updating (RPC sending) when trans_stop.
684 * Then the up layer users, such as LFSCK, can follow the general
685 * rule to handle trans_start/dt_object_lock for repairing linkEA
686 * inconsistency without distinguishing remote MDT-object.
688 * In fact, above solution for remote transaction should be the normal
689 * model without considering DNE1. The trouble brought by DNE1 will be
690 * resolved in DNE2. At that time, this patch can be removed.
692 * \param[in] env pointer to the thread context
693 * \param[in] dt pointer to the OSP layer dt_object
694 * \param[in] attr pointer to the attribute to be set
695 * \param[in] th pointer to the transaction handler
697 * \retval 0 for success
698 * \retval negative error number on failure
700 static int osp_declare_attr_set(const struct lu_env *env, struct dt_object *dt,
701 const struct lu_attr *attr, struct thandle *th)
705 if (!is_only_remote_trans(th))
706 rc = __osp_attr_set(env, dt, attr, th);
712 * Implement OSP layer dt_object_operations::do_attr_set() interface.
714 * Set attribute to the specified OST object.
716 * If the transaction is a remote transaction, then related modification
717 * sub-request has been added in the declare phase and related OUT RPC
718 * has been triggered at transaction start. Otherwise it will generate
719 * a MDS_SETATTR64_REC record in the llog. There is a dedicated thread
720 * to handle the llog asynchronously.
722 * If the attribute entry exists in the OSP object attributes cache,
723 * then update the cached attribute according to given attribute.
725 * \param[in] env pointer to the thread context
726 * \param[in] dt pointer to the OSP layer dt_object
727 * \param[in] attr pointer to the attribute to be set
728 * \param[in] th pointer to the transaction handler
729 * \param[in] capa the capability for this operation
731 * \retval 0 for success
732 * \retval negative error number on failure
734 static int osp_attr_set(const struct lu_env *env, struct dt_object *dt,
735 const struct lu_attr *attr, struct thandle *th,
736 struct lustre_capa *capa)
738 struct osp_object *o = dt2osp_obj(dt);
742 if (is_only_remote_trans(th)) {
743 rc = __osp_attr_set(env, dt, attr, th);
747 /* we're interested in uid/gid changes only */
748 if (!(attr->la_valid & (LA_UID | LA_GID)))
751 rc = osp_sync_add(env, o, MDS_SETATTR64_REC, th, attr);
752 /* XXX: send new uid/gid to OST ASAP? */
758 * Interpreter function for getting OSP object extended attribute asynchronously
760 * Called to interpret the result of an async mode RPC for getting the
761 * OSP object extended attribute.
763 * \param[in] env pointer to the thread context
764 * \param[in] reply pointer to the RPC reply
765 * \param[in] req pointer to the RPC request
766 * \param[in] obj pointer to the OSP object
767 * \param[out] data pointer to OSP object attributes cache
768 * \param[in] index the index of the attribute buffer in the reply
769 * \param[in] rc the result for handling the RPC
771 * \retval 0 for success
772 * \retval negative error number on failure
774 static int osp_xattr_get_interpterer(const struct lu_env *env,
775 struct object_update_reply *reply,
776 struct ptlrpc_request *req,
777 struct osp_object *obj,
778 void *data, int index, int rc)
780 struct osp_object_attr *ooa = obj->opo_ooa;
781 struct osp_xattr_entry *oxe = data;
782 struct lu_buf *rbuf = &osp_env_info(env)->osi_lb2;
784 LASSERT(ooa != NULL);
787 size_t len = sizeof(*oxe) + oxe->oxe_namelen + 1;
789 rc = object_update_result_data_get(reply, rbuf, index);
790 if (rc < 0 || rbuf->lb_len > (oxe->oxe_buflen - len)) {
791 spin_lock(&obj->opo_lock);
793 spin_unlock(&obj->opo_lock);
794 osp_oac_xattr_put(oxe);
796 return rc < 0 ? rc : -ERANGE;
799 spin_lock(&obj->opo_lock);
800 oxe->oxe_vallen = rbuf->lb_len;
801 memcpy(oxe->oxe_value, rbuf->lb_buf, rbuf->lb_len);
804 spin_unlock(&obj->opo_lock);
805 } else if (rc == -ENOENT || rc == -ENODATA) {
806 spin_lock(&obj->opo_lock);
809 spin_unlock(&obj->opo_lock);
811 spin_lock(&obj->opo_lock);
813 spin_unlock(&obj->opo_lock);
816 osp_oac_xattr_put(oxe);
822 * Implement OSP dt_object_operations::do_declare_xattr_get() interface.
824 * Declare that the caller will get extended attribute from the specified
827 * This function will add an OUT_XATTR_GET sub-request to the per OSP
828 * based shared asynchronous request queue with the interpreter function:
829 * osp_xattr_get_interpterer().
831 * \param[in] env pointer to the thread context
832 * \param[in] dt pointer to the OSP layer dt_object
833 * \param[out] buf pointer to the lu_buf to hold the extended attribute
834 * \param[in] name the name for the expected extended attribute
835 * \param[in] capa the capability for this operation
837 * \retval 0 for success
838 * \retval negative error number on failure
840 static int osp_declare_xattr_get(const struct lu_env *env, struct dt_object *dt,
841 struct lu_buf *buf, const char *name,
842 struct lustre_capa *capa)
844 struct osp_object *obj = dt2osp_obj(dt);
845 struct osp_device *osp = lu2osp_dev(dt->do_lu.lo_dev);
846 struct osp_xattr_entry *oxe;
847 __u16 namelen = strlen(name);
850 LASSERT(buf != NULL);
851 LASSERT(name != NULL);
853 /* If only for xattr size, return directly. */
854 if (unlikely(buf->lb_len == 0))
857 if (obj->opo_ooa == NULL) {
858 rc = osp_oac_init(obj);
863 oxe = osp_oac_xattr_find_or_add(obj, name, buf->lb_len);
867 mutex_lock(&osp->opd_async_requests_mutex);
868 rc = osp_insert_async_request(env, OUT_XATTR_GET, obj, 1,
869 &namelen, (const void **)&name, oxe,
870 osp_xattr_get_interpterer);
872 mutex_unlock(&osp->opd_async_requests_mutex);
873 osp_oac_xattr_put(oxe);
875 struct dt_update_request *update;
877 /* XXX: Currently, we trigger the batched async OUT
878 * RPC via dt_declare_xattr_get(). It is not
879 * perfect solution, but works well now.
881 * We will improve it in the future. */
882 update = osp->opd_async_requests;
883 if (update != NULL && update->dur_buf.ub_req != NULL &&
884 update->dur_buf.ub_req->ourq_count > 0) {
885 osp->opd_async_requests = NULL;
886 mutex_unlock(&osp->opd_async_requests_mutex);
887 rc = osp_unplug_async_request(env, osp, update);
889 mutex_unlock(&osp->opd_async_requests_mutex);
897 * Implement OSP layer dt_object_operations::do_xattr_get() interface.
899 * Get extended attribute from the specified MDT/OST object.
901 * If the extended attribute is in the OSP object attributes cache, then
902 * return the cached extended attribute directly. Otherwise it will get
903 * the extended attribute synchronously, if successful, add it to the OSP
904 * attributes cache. (\see lustre/osp/osp_trans.c for OUT RPC.)
906 * There is a race condition: some other thread has added the named extended
907 * attributed entry to the OSP object attributes cache during the current
908 * OUT_XATTR_GET handling. If such case happens, the OSP will replace the
909 * (just) existing extended attribute entry with the new replied one.
911 * \param[in] env pointer to the thread context
912 * \param[in] dt pointer to the OSP layer dt_object
913 * \param[out] buf pointer to the lu_buf to hold the extended attribute
914 * \param[in] name the name for the expected extended attribute
915 * \param[in] capa the capability for this operation
917 * \retval 0 for success
918 * \retval negative error number on failure
920 int osp_xattr_get(const struct lu_env *env, struct dt_object *dt,
921 struct lu_buf *buf, const char *name,
922 struct lustre_capa *capa)
924 struct osp_device *osp = lu2osp_dev(dt->do_lu.lo_dev);
925 struct osp_object *obj = dt2osp_obj(dt);
926 struct dt_device *dev = &osp->opd_dt_dev;
927 struct lu_buf *rbuf = &osp_env_info(env)->osi_lb2;
928 struct dt_update_request *update = NULL;
929 struct ptlrpc_request *req = NULL;
930 struct object_update_reply *reply;
931 struct osp_xattr_entry *oxe = NULL;
932 const char *dname = dt->do_lu.lo_dev->ld_obd->obd_name;
936 LASSERT(buf != NULL);
937 LASSERT(name != NULL);
939 if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_NETWORK) &&
940 osp->opd_index == cfs_fail_val) {
941 if (is_ost_obj(&dt->do_lu)) {
942 if (osp_dev2node(osp) == cfs_fail_val)
945 if (strcmp(name, XATTR_NAME_LINK) == 0)
950 if (unlikely(obj->opo_non_exist))
953 oxe = osp_oac_xattr_find(obj, name, false);
955 spin_lock(&obj->opo_lock);
956 if (oxe->oxe_ready) {
958 GOTO(unlock, rc = -ENODATA);
960 if (buf->lb_buf == NULL)
961 GOTO(unlock, rc = oxe->oxe_vallen);
963 if (buf->lb_len < oxe->oxe_vallen)
964 GOTO(unlock, rc = -ERANGE);
966 memcpy(buf->lb_buf, oxe->oxe_value, oxe->oxe_vallen);
968 GOTO(unlock, rc = oxe->oxe_vallen);
971 spin_unlock(&obj->opo_lock);
972 osp_oac_xattr_put(oxe);
976 spin_unlock(&obj->opo_lock);
979 update = dt_update_request_create(dev);
981 GOTO(out, rc = PTR_ERR(update));
983 rc = out_xattr_get_pack(env, &update->dur_buf,
984 lu_object_fid(&dt->do_lu), name);
986 CERROR("%s: Insert update error "DFID": rc = %d\n",
987 dname, PFID(lu_object_fid(&dt->do_lu)), rc);
991 rc = out_remote_sync(env, osp->opd_obd->u.cli.cl_import, update, &req);
994 dt->do_lu.lo_header->loh_attr &= ~LOHA_EXISTS;
995 obj->opo_non_exist = 1;
998 if (obj->opo_ooa == NULL)
1002 oxe = osp_oac_xattr_find_or_add(obj, name, buf->lb_len);
1005 CWARN("%s: Fail to add xattr (%s) to cache for "
1006 DFID" (1): rc = %d\n", dname, name,
1007 PFID(lu_object_fid(&dt->do_lu)), rc);
1012 spin_lock(&obj->opo_lock);
1013 if (rc == -ENOENT || rc == -ENODATA) {
1019 spin_unlock(&obj->opo_lock);
1024 reply = req_capsule_server_sized_get(&req->rq_pill,
1025 &RMF_OUT_UPDATE_REPLY,
1026 OUT_UPDATE_REPLY_SIZE);
1027 if (reply->ourp_magic != UPDATE_REPLY_MAGIC) {
1028 CERROR("%s: Wrong version %x expected %x "DFID": rc = %d\n",
1029 dname, reply->ourp_magic, UPDATE_REPLY_MAGIC,
1030 PFID(lu_object_fid(&dt->do_lu)), -EPROTO);
1032 GOTO(out, rc = -EPROTO);
1035 rc = object_update_result_data_get(reply, rbuf, 0);
1039 if (buf->lb_buf == NULL)
1040 GOTO(out, rc = rbuf->lb_len);
1042 if (unlikely(buf->lb_len < rbuf->lb_len))
1043 GOTO(out, rc = -ERANGE);
1045 memcpy(buf->lb_buf, rbuf->lb_buf, rbuf->lb_len);
1047 if (obj->opo_ooa == NULL)
1051 oxe = osp_oac_xattr_find_or_add(obj, name, rbuf->lb_len);
1053 CWARN("%s: Fail to add xattr (%s) to "
1054 "cache for "DFID" (2): rc = %d\n",
1055 dname, name, PFID(lu_object_fid(&dt->do_lu)), rc);
1061 if (oxe->oxe_buflen - oxe->oxe_namelen - 1 < rbuf->lb_len) {
1062 struct osp_xattr_entry *old = oxe;
1063 struct osp_xattr_entry *tmp;
1065 tmp = osp_oac_xattr_replace(obj, &old, rbuf->lb_len);
1066 osp_oac_xattr_put(oxe);
1069 CWARN("%s: Fail to update xattr (%s) to "
1070 "cache for "DFID": rc = %d\n",
1071 dname, name, PFID(lu_object_fid(&dt->do_lu)), rc);
1072 spin_lock(&obj->opo_lock);
1074 spin_unlock(&obj->opo_lock);
1079 /* Drop the ref for entry on list. */
1080 osp_oac_xattr_put(old);
1083 spin_lock(&obj->opo_lock);
1084 oxe->oxe_vallen = rbuf->lb_len;
1085 memcpy(oxe->oxe_value, rbuf->lb_buf, rbuf->lb_len);
1088 spin_unlock(&obj->opo_lock);
1094 ptlrpc_req_finished(req);
1096 if (update != NULL && !IS_ERR(update))
1097 dt_update_request_destroy(update);
1100 osp_oac_xattr_put(oxe);
1105 static int __osp_xattr_set(const struct lu_env *env, struct dt_object *dt,
1106 const struct lu_buf *buf, const char *name,
1107 int flag, struct thandle *th)
1109 struct osp_object *o = dt2osp_obj(dt);
1110 struct dt_update_request *update;
1111 struct osp_xattr_entry *oxe;
1115 LASSERT(buf->lb_len > 0 && buf->lb_buf != NULL);
1117 update = dt_update_request_find_or_create(th, dt);
1118 if (IS_ERR(update)) {
1119 CERROR("%s: Get OSP update buf failed "DFID": rc = %d\n",
1120 dt->do_lu.lo_dev->ld_obd->obd_name,
1121 PFID(lu_object_fid(&dt->do_lu)),
1122 (int)PTR_ERR(update));
1124 RETURN(PTR_ERR(update));
1127 rc = out_xattr_set_pack(env, &update->dur_buf,
1128 lu_object_fid(&dt->do_lu),
1129 buf, name, flag, update->dur_batchid);
1130 if (rc != 0 || o->opo_ooa == NULL)
1133 oxe = osp_oac_xattr_find_or_add(o, name, buf->lb_len);
1135 CWARN("%s: cannot cache xattr '%s' of "DFID"\n",
1136 dt->do_lu.lo_dev->ld_obd->obd_name,
1137 name, PFID(lu_object_fid(&dt->do_lu)));
1142 if (oxe->oxe_buflen - oxe->oxe_namelen - 1 < buf->lb_len) {
1143 struct osp_xattr_entry *old = oxe;
1144 struct osp_xattr_entry *tmp;
1146 tmp = osp_oac_xattr_replace(o, &old, buf->lb_len);
1147 osp_oac_xattr_put(oxe);
1150 CWARN("%s: cannot update cached xattr '%s' of "DFID"\n",
1151 dt->do_lu.lo_dev->ld_obd->obd_name,
1152 name, PFID(lu_object_fid(&dt->do_lu)));
1153 spin_lock(&o->opo_lock);
1155 spin_unlock(&o->opo_lock);
1160 /* Drop the ref for entry on list. */
1161 osp_oac_xattr_put(old);
1164 spin_lock(&o->opo_lock);
1165 oxe->oxe_vallen = buf->lb_len;
1166 memcpy(oxe->oxe_value, buf->lb_buf, buf->lb_len);
1169 spin_unlock(&o->opo_lock);
1170 osp_oac_xattr_put(oxe);
1176 * Implement OSP layer dt_object_operations::do_declare_xattr_set() interface.
1178 * Declare that the caller will set extended attribute to the specified
1181 * This function will add an OUT_XATTR_SET sub-request to the per
1182 * OSP-transaction based request queue which will be flushed when
1183 * the transaction starts.
1185 * \param[in] env pointer to the thread context
1186 * \param[in] dt pointer to the OSP layer dt_object
1187 * \param[in] buf pointer to the lu_buf to hold the extended attribute
1188 * \param[in] name the name of the extended attribute to be set
1189 * \param[in] flag to indicate the detailed set operation: LU_XATTR_CREATE
1190 * or LU_XATTR_REPLACE or others
1191 * \param[in] th pointer to the transaction handler
1193 * \retval 0 for success
1194 * \retval negative error number on failure
1196 int osp_declare_xattr_set(const struct lu_env *env, struct dt_object *dt,
1197 const struct lu_buf *buf, const char *name,
1198 int flag, struct thandle *th)
1202 /* Please check the comment in osp_attr_set() for handling
1203 * remote transaction. */
1204 if (!is_only_remote_trans(th))
1205 rc = __osp_xattr_set(env, dt, buf, name, flag, th);
1211 * Implement OSP layer dt_object_operations::do_xattr_set() interface.
1213 * Set extended attribute to the specified MDT/OST object.
1215 * The real modification sub-request has been added in the declare phase
1216 * and related (OUT) RPC has been triggered when transaction start.
1218 * If the OSP attributes cache is initialized, then check whether the name
1219 * extended attribute entry exists in the cache or not. If yes, replace it;
1220 * otherwise, add the extended attribute to the cache.
1222 * \param[in] env pointer to the thread context
1223 * \param[in] dt pointer to the OSP layer dt_object
1224 * \param[in] buf pointer to the lu_buf to hold the extended attribute
1225 * \param[in] name the name of the extended attribute to be set
1226 * \param[in] fl to indicate the detailed set operation: LU_XATTR_CREATE
1227 * or LU_XATTR_REPLACE or others
1228 * \param[in] th pointer to the transaction handler
1229 * \param[in] capa the capability for this operation
1231 * \retval 0 for success
1232 * \retval negative error number on failure
1234 int osp_xattr_set(const struct lu_env *env, struct dt_object *dt,
1235 const struct lu_buf *buf, const char *name, int fl,
1236 struct thandle *th, struct lustre_capa *capa)
1240 CDEBUG(D_INFO, "xattr %s set object "DFID"\n", name,
1241 PFID(&dt->do_lu.lo_header->loh_fid));
1243 /* Please check the comment in osp_attr_set() for handling
1244 * remote transaction. */
1245 if (is_only_remote_trans(th))
1246 rc = __osp_xattr_set(env, dt, buf, name, fl, th);
1251 static int __osp_xattr_del(const struct lu_env *env, struct dt_object *dt,
1252 const char *name, struct thandle *th)
1254 struct dt_update_request *update;
1255 const struct lu_fid *fid;
1256 struct osp_object *o = dt2osp_obj(dt);
1257 struct osp_xattr_entry *oxe;
1260 update = dt_update_request_find_or_create(th, dt);
1262 return PTR_ERR(update);
1264 fid = lu_object_fid(&dt->do_lu);
1266 rc = out_xattr_del_pack(env, &update->dur_buf, fid, name,
1267 update->dur_batchid);
1269 if (rc != 0 || o->opo_ooa == NULL)
1272 oxe = osp_oac_xattr_find(o, name, true);
1274 /* Drop the ref for entry on list. */
1275 osp_oac_xattr_put(oxe);
1281 * Implement OSP layer dt_object_operations::do_declare_xattr_del() interface.
1283 * Declare that the caller will delete extended attribute on the specified
1286 * This function will add an OUT_XATTR_DEL sub-request to the per
1287 * OSP-transaction based request queue which will be flushed when
1288 * transaction start.
1290 * \param[in] env pointer to the thread context
1291 * \param[in] dt pointer to the OSP layer dt_object
1292 * \param[in] name the name of the extended attribute to be set
1293 * \param[in] th pointer to the transaction handler
1295 * \retval 0 for success
1296 * \retval negative error number on failure
1298 int osp_declare_xattr_del(const struct lu_env *env, struct dt_object *dt,
1299 const char *name, struct thandle *th)
1303 /* Please check the comment in osp_attr_set() for handling
1304 * remote transaction. */
1305 if (!is_only_remote_trans(th))
1306 rc = __osp_xattr_del(env, dt, name, th);
1312 * Implement OSP layer dt_object_operations::do_xattr_del() interface.
1314 * Delete extended attribute on the specified MDT/OST object.
1316 * The real modification sub-request has been added in the declare phase
1317 * and related (OUT) RPC has been triggered when transaction start.
1319 * If the name extended attribute entry exists in the OSP attributes
1320 * cache, then remove it from the cache.
1322 * \param[in] env pointer to the thread context
1323 * \param[in] dt pointer to the OSP layer dt_object
1324 * \param[in] name the name of the extended attribute to be set
1325 * \param[in] th pointer to the transaction handler
1326 * \param[in] capa the capability for this operation
1328 * \retval 0 for success
1329 * \retval negative error number on failure
1331 int osp_xattr_del(const struct lu_env *env, struct dt_object *dt,
1332 const char *name, struct thandle *th,
1333 struct lustre_capa *capa)
1337 CDEBUG(D_INFO, "xattr %s del object "DFID"\n", name,
1338 PFID(&dt->do_lu.lo_header->loh_fid));
1340 /* Please check the comment in osp_attr_set() for handling
1341 * remote transaction. */
1342 if (is_only_remote_trans(th))
1343 rc = __osp_xattr_del(env, dt, name, th);
1349 * Implement OSP layer dt_object_operations::do_declare_create() interface.
1351 * Declare that the caller will create the OST object.
1353 * If the transaction is a remote transaction (please refer to the
1354 * comment of osp_trans_create() for remote transaction), then the FID
1355 * for the OST object has been assigned already, and will be handled
1356 * as create (remote) MDT object via osp_md_declare_object_create().
1357 * This function is usually used for LFSCK to re-create the lost OST
1358 * object. Otherwise, if it is not replay case, the OSP will reserve
1359 * pre-created object for the subsequent create operation; if the MDT
1360 * side cached pre-created objects are less than some threshold, then
1361 * it will wakeup the pre-create thread.
1363 * \param[in] env pointer to the thread context
1364 * \param[in] dt pointer to the OSP layer dt_object
1365 * \param[in] attr the attribute for the object to be created
1366 * \param[in] hint pointer to the hint for creating the object, such as
1368 * \param[in] dof pointer to the dt_object_format for help the creation
1369 * \param[in] th pointer to the transaction handler
1371 * \retval 0 for success
1372 * \retval negative error number on failure
1374 static int osp_declare_object_create(const struct lu_env *env,
1375 struct dt_object *dt,
1376 struct lu_attr *attr,
1377 struct dt_allocation_hint *hint,
1378 struct dt_object_format *dof,
1381 struct osp_thread_info *osi = osp_env_info(env);
1382 struct osp_device *d = lu2osp_dev(dt->do_lu.lo_dev);
1383 struct osp_object *o = dt2osp_obj(dt);
1384 const struct lu_fid *fid = lu_object_fid(&dt->do_lu);
1389 if (is_only_remote_trans(th)) {
1390 LASSERT(fid_is_sane(fid));
1392 rc = osp_md_declare_object_create(env, dt, attr, hint, dof, th);
1397 /* should happen to non-0 OSP only so that at least one object
1398 * has been already declared in the scenario and LOD should
1400 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_CREATE_FAIL) && d->opd_index == 1)
1403 LASSERT(d->opd_last_used_oid_file);
1406 * There can be gaps in precreated ids and record to unlink llog
1407 * XXX: we do not handle gaps yet, implemented before solution
1408 * was found to be racy, so we disabled that. there is no
1409 * point in making useless but expensive llog declaration.
1411 /* rc = osp_sync_declare_add(env, o, MDS_UNLINK64_REC, th); */
1413 if (unlikely(!fid_is_zero(fid))) {
1414 /* replay case: caller knows fid */
1415 osi->osi_off = sizeof(osi->osi_id) * d->opd_index;
1416 osi->osi_lb.lb_len = sizeof(osi->osi_id);
1417 osi->osi_lb.lb_buf = NULL;
1418 rc = dt_declare_record_write(env, d->opd_last_used_oid_file,
1419 &osi->osi_lb, osi->osi_off, th);
1424 * in declaration we need to reserve object so that we don't block
1425 * awaiting precreation RPC to complete
1427 rc = osp_precreate_reserve(env, d);
1429 * we also need to declare update to local "last used id" file for
1430 * recovery if object isn't used for a reason, we need to release
1431 * reservation, this can be made in osd_object_release()
1434 /* mark id is reserved: in create we don't want to talk
1436 LASSERT(o->opo_reserved == 0);
1437 o->opo_reserved = 1;
1439 /* common for all OSPs file hystorically */
1440 osi->osi_off = sizeof(osi->osi_id) * d->opd_index;
1441 osi->osi_lb.lb_len = sizeof(osi->osi_id);
1442 osi->osi_lb.lb_buf = NULL;
1443 rc = dt_declare_record_write(env, d->opd_last_used_oid_file,
1444 &osi->osi_lb, osi->osi_off, th);
1446 /* not needed in the cache anymore */
1447 set_bit(LU_OBJECT_HEARD_BANSHEE,
1448 &dt->do_lu.lo_header->loh_flags);
1454 * Implement OSP layer dt_object_operations::do_create() interface.
1456 * Create the OST object.
1458 * For remote transaction case, the real create sub-request has been
1459 * added in the declare phase and related (OUT) RPC has been triggered
1460 * at transaction start. Here, like creating (remote) MDT object, the
1461 * OSP will mark the object existence via osp_md_object_create().
1463 * For non-remote transaction case, the OSP will assign FID to the
1464 * object to be created, and update last_used Object ID (OID) file.
1466 * \param[in] env pointer to the thread context
1467 * \param[in] dt pointer to the OSP layer dt_object
1468 * \param[in] attr the attribute for the object to be created
1469 * \param[in] hint pointer to the hint for creating the object, such as
1471 * \param[in] dof pointer to the dt_object_format for help the creation
1472 * \param[in] th pointer to the transaction handler
1474 * \retval 0 for success
1475 * \retval negative error number on failure
1477 static int osp_object_create(const struct lu_env *env, struct dt_object *dt,
1478 struct lu_attr *attr,
1479 struct dt_allocation_hint *hint,
1480 struct dt_object_format *dof, struct thandle *th)
1482 struct osp_thread_info *osi = osp_env_info(env);
1483 struct osp_device *d = lu2osp_dev(dt->do_lu.lo_dev);
1484 struct osp_object *o = dt2osp_obj(dt);
1486 struct lu_fid *fid = &osi->osi_fid;
1489 if (is_only_remote_trans(th)) {
1490 LASSERT(fid_is_sane(lu_object_fid(&dt->do_lu)));
1492 rc = osp_md_object_create(env, dt, attr, hint, dof, th);
1494 o->opo_non_exist = 0;
1499 o->opo_non_exist = 0;
1500 if (o->opo_reserved) {
1501 /* regular case, fid is assigned holding transaction open */
1502 osp_object_assign_fid(env, d, o);
1505 memcpy(fid, lu_object_fid(&dt->do_lu), sizeof(*fid));
1507 LASSERTF(fid_is_sane(fid), "fid for osp_object %p is insane"DFID"!\n",
1510 if (!o->opo_reserved) {
1511 /* special case, id was assigned outside of transaction
1512 * see comments in osp_declare_attr_set */
1513 LASSERT(d->opd_pre != NULL);
1514 spin_lock(&d->opd_pre_lock);
1515 osp_update_last_fid(d, fid);
1516 spin_unlock(&d->opd_pre_lock);
1519 CDEBUG(D_INODE, "fid for osp_object %p is "DFID"\n", o, PFID(fid));
1521 /* If the precreate ends, it means it will be ready to rollover to
1522 * the new sequence soon, all the creation should be synchronized,
1523 * otherwise during replay, the replay fid will be inconsistent with
1524 * last_used/create fid */
1525 if (osp_precreate_end_seq(env, d) && osp_is_fid_client(d))
1529 * it's OK if the import is inactive by this moment - id was created
1530 * by OST earlier, we just need to maintain it consistently on the disk
1531 * once import is reconnected, OSP will claim this and other objects
1532 * used and OST either keep them, if they exist or recreate
1535 /* we might have lost precreated objects */
1536 if (unlikely(d->opd_gap_count) > 0) {
1537 LASSERT(d->opd_pre != NULL);
1538 spin_lock(&d->opd_pre_lock);
1539 if (d->opd_gap_count > 0) {
1540 int count = d->opd_gap_count;
1542 ostid_set_id(&osi->osi_oi,
1543 fid_oid(&d->opd_gap_start_fid));
1544 d->opd_gap_count = 0;
1545 spin_unlock(&d->opd_pre_lock);
1547 CDEBUG(D_HA, "Writing gap "DFID"+%d in llog\n",
1548 PFID(&d->opd_gap_start_fid), count);
1549 /* real gap handling is disabled intil ORI-692 will be
1550 * fixed, now we only report gaps */
1552 spin_unlock(&d->opd_pre_lock);
1556 /* Only need update last_used oid file, seq file will only be update
1557 * during seq rollover */
1558 osp_objid_buf_prep(&osi->osi_lb, &osi->osi_off,
1559 &d->opd_last_used_fid.f_oid, d->opd_index);
1561 rc = dt_record_write(env, d->opd_last_used_oid_file, &osi->osi_lb,
1564 CDEBUG(D_HA, "%s: Wrote last used FID: "DFID", index %d: %d\n",
1565 d->opd_obd->obd_name, PFID(fid), d->opd_index, rc);
1571 * Implement OSP layer dt_object_operations::do_declare_destroy() interface.
1573 * Declare that the caller will destroy the specified OST object.
1575 * The OST object destroy will be handled via llog asynchronously. This
1576 * function will declare the credits for generating MDS_UNLINK64_REC llog.
1578 * \param[in] env pointer to the thread context
1579 * \param[in] dt pointer to the OSP layer dt_object to be destroyed
1580 * \param[in] th pointer to the transaction handler
1582 * \retval 0 for success
1583 * \retval negative error number on failure
1585 int osp_declare_object_destroy(const struct lu_env *env,
1586 struct dt_object *dt, struct thandle *th)
1588 struct osp_object *o = dt2osp_obj(dt);
1594 * track objects to be destroyed via llog
1596 rc = osp_sync_declare_add(env, o, MDS_UNLINK64_REC, th);
1602 * Implement OSP layer dt_object_operations::do_destroy() interface.
1604 * Destroy the specified OST object.
1606 * The OSP generates a MDS_UNLINK64_REC record in the llog. There
1607 * will be some dedicated thread to handle the llog asynchronously.
1609 * It also marks the object as non-cached.
1611 * \param[in] env pointer to the thread context
1612 * \param[in] dt pointer to the OSP layer dt_object to be destroyed
1613 * \param[in] th pointer to the transaction handler
1615 * \retval 0 for success
1616 * \retval negative error number on failure
1618 int osp_object_destroy(const struct lu_env *env, struct dt_object *dt,
1621 struct osp_object *o = dt2osp_obj(dt);
1626 o->opo_non_exist = 1;
1628 * once transaction is committed put proper command on
1629 * the queue going to our OST
1631 rc = osp_sync_add(env, o, MDS_UNLINK64_REC, th, NULL);
1633 /* not needed in cache any more */
1634 set_bit(LU_OBJECT_HEARD_BANSHEE, &dt->do_lu.lo_header->loh_flags);
1639 static int osp_orphan_index_lookup(const struct lu_env *env,
1640 struct dt_object *dt,
1642 const struct dt_key *key,
1643 struct lustre_capa *capa)
1648 static int osp_orphan_index_declare_insert(const struct lu_env *env,
1649 struct dt_object *dt,
1650 const struct dt_rec *rec,
1651 const struct dt_key *key,
1652 struct thandle *handle)
1657 static int osp_orphan_index_insert(const struct lu_env *env,
1658 struct dt_object *dt,
1659 const struct dt_rec *rec,
1660 const struct dt_key *key,
1661 struct thandle *handle,
1662 struct lustre_capa *capa,
1668 static int osp_orphan_index_declare_delete(const struct lu_env *env,
1669 struct dt_object *dt,
1670 const struct dt_key *key,
1671 struct thandle *handle)
1676 static int osp_orphan_index_delete(const struct lu_env *env,
1677 struct dt_object *dt,
1678 const struct dt_key *key,
1679 struct thandle *handle,
1680 struct lustre_capa *capa)
1686 * Initialize the OSP layer index iteration.
1688 * \param[in] env pointer to the thread context
1689 * \param[in] dt pointer to the index object to be iterated
1690 * \param[in] attr unused
1691 * \param[in] capa the capability for this operation
1693 * \retval pointer to the iteration structure
1694 * \retval negative error number on failure
1696 struct dt_it *osp_it_init(const struct lu_env *env, struct dt_object *dt,
1697 __u32 attr, struct lustre_capa *capa)
1703 return ERR_PTR(-ENOMEM);
1705 it->ooi_pos_ent = -1;
1707 it->ooi_attr = attr;
1709 return (struct dt_it *)it;
1713 * Finalize the OSP layer index iteration.
1715 * \param[in] env pointer to the thread context
1716 * \param[in] di pointer to the iteration structure
1718 void osp_it_fini(const struct lu_env *env, struct dt_it *di)
1720 struct osp_it *it = (struct osp_it *)di;
1721 struct page **pages = it->ooi_pages;
1722 int npages = it->ooi_total_npages;
1725 if (pages != NULL) {
1726 for (i = 0; i < npages; i++) {
1727 if (pages[i] != NULL) {
1728 if (pages[i] == it->ooi_cur_page) {
1730 it->ooi_cur_page = NULL;
1732 __free_page(pages[i]);
1735 OBD_FREE(pages, npages * sizeof(*pages));
1741 * Get more records for the iteration from peer.
1743 * The new records will be filled in an array of pages. The OSP side
1744 * allows 1MB bulk data to be transfered.
1746 * \param[in] env pointer to the thread context
1747 * \param[in] it pointer to the iteration structure
1749 * \retval 0 for success
1750 * \retval negative error number on failure
1752 static int osp_it_fetch(const struct lu_env *env, struct osp_it *it)
1754 struct lu_device *dev = it->ooi_obj->do_lu.lo_dev;
1755 struct osp_device *osp = lu2osp_dev(dev);
1756 struct page **pages;
1757 struct ptlrpc_request *req = NULL;
1758 struct ptlrpc_bulk_desc *desc;
1759 struct idx_info *ii;
1766 npages = min_t(unsigned int, OFD_MAX_BRW_SIZE, 1 << 20);
1767 npages /= PAGE_CACHE_SIZE;
1769 OBD_ALLOC(pages, npages * sizeof(*pages));
1773 it->ooi_pages = pages;
1774 it->ooi_total_npages = npages;
1775 for (i = 0; i < npages; i++) {
1776 pages[i] = alloc_page(GFP_IOFS);
1777 if (pages[i] == NULL)
1781 req = ptlrpc_request_alloc(osp->opd_obd->u.cli.cl_import,
1786 rc = ptlrpc_request_pack(req, LUSTRE_OBD_VERSION, OBD_IDX_READ);
1788 ptlrpc_request_free(req);
1792 req->rq_request_portal = OUT_PORTAL;
1793 ii = req_capsule_client_get(&req->rq_pill, &RMF_IDX_INFO);
1794 memset(ii, 0, sizeof(*ii));
1795 if (fid_is_last_id(lu_object_fid(&it->ooi_obj->do_lu))) {
1796 /* LFSCK will iterate orphan object[FID_SEQ_LAYOUT_BTREE,
1797 * ost_index, 0] with LAST_ID FID, so it needs to replace
1798 * the FID with orphan FID here */
1799 ii->ii_fid.f_seq = FID_SEQ_LAYOUT_RBTREE;
1800 ii->ii_fid.f_oid = osp->opd_index;
1801 ii->ii_fid.f_ver = 0;
1802 ii->ii_flags = II_FL_NOHASH;
1803 ii->ii_attrs = osp_dev2node(osp);
1805 ii->ii_fid = *lu_object_fid(&it->ooi_obj->do_lu);
1806 ii->ii_flags = II_FL_NOHASH | II_FL_NOKEY | II_FL_VARKEY |
1808 ii->ii_attrs = it->ooi_attr;
1810 ii->ii_magic = IDX_INFO_MAGIC;
1811 ii->ii_count = npages * LU_PAGE_COUNT;
1812 ii->ii_hash_start = it->ooi_next;
1814 ptlrpc_at_set_req_timeout(req);
1816 desc = ptlrpc_prep_bulk_imp(req, npages, 1, BULK_PUT_SINK,
1819 ptlrpc_request_free(req);
1823 for (i = 0; i < npages; i++)
1824 ptlrpc_prep_bulk_page_pin(desc, pages[i], 0, PAGE_CACHE_SIZE);
1826 ptlrpc_request_set_replen(req);
1827 rc = ptlrpc_queue_wait(req);
1831 rc = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk,
1832 req->rq_bulk->bd_nob_transferred);
1837 ii = req_capsule_server_get(&req->rq_pill, &RMF_IDX_INFO);
1838 if (ii->ii_magic != IDX_INFO_MAGIC)
1839 GOTO(out, rc = -EPROTO);
1841 npages = (ii->ii_count + LU_PAGE_COUNT - 1) >>
1842 (PAGE_CACHE_SHIFT - LU_PAGE_SHIFT);
1843 if (npages > it->ooi_total_npages) {
1844 CERROR("%s: returned more pages than expected, %u > %u\n",
1845 osp->opd_obd->obd_name, npages, it->ooi_total_npages);
1846 GOTO(out, rc = -EINVAL);
1849 it->ooi_valid_npages = npages;
1850 if (ptlrpc_rep_need_swab(req))
1853 it->ooi_next = ii->ii_hash_end;
1856 ptlrpc_req_finished(req);
1862 * Move the iteration cursor to the next lu_page.
1864 * One system page (PAGE_SIZE) may contain multiple lu_page (4KB),
1865 * that depends on the LU_PAGE_COUNT. If it is not the last lu_page
1866 * in current system page, then move the iteration cursor to the next
1867 * lu_page in current system page. Otherwise, if there are more system
1868 * pages in the cache, then move the iteration cursor to the next system
1869 * page. If all the cached records (pages) have been iterated, then fetch
1870 * more records via osp_it_fetch().
1872 * \param[in] env pointer to the thread context
1873 * \param[in] di pointer to the iteration structure
1875 * \retval positive for end of the directory
1876 * \retval 0 for success
1877 * \retval negative error number on failure
1879 int osp_it_next_page(const struct lu_env *env, struct dt_it *di)
1881 struct osp_it *it = (struct osp_it *)di;
1882 struct lu_idxpage *idxpage;
1883 struct page **pages;
1889 idxpage = it->ooi_cur_idxpage;
1890 if (idxpage != NULL) {
1891 if (idxpage->lip_nr == 0)
1894 if (it->ooi_pos_ent < idxpage->lip_nr) {
1895 CDEBUG(D_INFO, "ooi_pos %d nr %d\n",
1896 (int)it->ooi_pos_ent, (int)idxpage->lip_nr);
1899 it->ooi_cur_idxpage = NULL;
1900 it->ooi_pos_lu_page++;
1903 if (it->ooi_pos_lu_page < LU_PAGE_COUNT) {
1904 it->ooi_cur_idxpage = (void *)it->ooi_cur_page +
1905 LU_PAGE_SIZE * it->ooi_pos_lu_page;
1907 lustre_swab_lip_header(it->ooi_cur_idxpage);
1908 if (it->ooi_cur_idxpage->lip_magic != LIP_MAGIC) {
1909 struct osp_device *osp =
1910 lu2osp_dev(it->ooi_obj->do_lu.lo_dev);
1912 CERROR("%s: invalid magic (%x != %x) for page "
1913 "%d/%d while read layout orphan index\n",
1914 osp->opd_obd->obd_name,
1915 it->ooi_cur_idxpage->lip_magic,
1916 LIP_MAGIC, it->ooi_pos_page,
1917 it->ooi_pos_lu_page);
1918 /* Skip this lu_page next time. */
1919 it->ooi_pos_ent = idxpage->lip_nr - 1;
1922 it->ooi_pos_ent = -1;
1926 kunmap(it->ooi_cur_page);
1927 it->ooi_cur_page = NULL;
1931 pages = it->ooi_pages;
1932 if (it->ooi_pos_page < it->ooi_valid_npages) {
1933 it->ooi_cur_page = kmap(pages[it->ooi_pos_page]);
1934 it->ooi_pos_lu_page = 0;
1938 for (i = 0; i < it->ooi_total_npages; i++) {
1939 if (pages[i] != NULL)
1940 __free_page(pages[i]);
1942 OBD_FREE(pages, it->ooi_total_npages * sizeof(*pages));
1944 it->ooi_pos_page = 0;
1945 it->ooi_total_npages = 0;
1946 it->ooi_valid_npages = 0;
1949 it->ooi_cur_page = NULL;
1950 it->ooi_cur_idxpage = NULL;
1951 it->ooi_pages = NULL;
1954 if (it->ooi_next == II_END_OFF)
1957 rc = osp_it_fetch(env, it);
1965 * Move the iteration cursor to the next record.
1967 * If there are more records in the lu_page, then move the iteration
1968 * cursor to the next record directly. Otherwise, move the iteration
1969 * cursor to the record in the next lu_page via osp_it_next_page()
1971 * \param[in] env pointer to the thread context
1972 * \param[in] di pointer to the iteration structure
1974 * \retval positive for end of the directory
1975 * \retval 0 for success
1976 * \retval negative error number on failure
1978 int osp_orphan_it_next(const struct lu_env *env, struct dt_it *di)
1980 struct osp_it *it = (struct osp_it *)di;
1981 struct lu_idxpage *idxpage;
1986 idxpage = it->ooi_cur_idxpage;
1987 if (idxpage != NULL) {
1988 if (idxpage->lip_nr == 0)
1992 if (it->ooi_pos_ent < idxpage->lip_nr) {
1994 (struct lu_orphan_ent *)idxpage->lip_entries +
1997 lustre_swab_orphan_ent(it->ooi_ent);
2002 rc = osp_it_next_page(env, di);
2009 int osp_it_get(const struct lu_env *env, struct dt_it *di,
2010 const struct dt_key *key)
2015 void osp_it_put(const struct lu_env *env, struct dt_it *di)
2019 struct dt_key *osp_orphan_it_key(const struct lu_env *env,
2020 const struct dt_it *di)
2022 struct osp_it *it = (struct osp_it *)di;
2023 struct lu_orphan_ent *ent = (struct lu_orphan_ent *)it->ooi_ent;
2025 if (likely(ent != NULL))
2026 return (struct dt_key *)(&ent->loe_key);
2031 int osp_orphan_it_key_size(const struct lu_env *env, const struct dt_it *di)
2033 return sizeof(struct lu_fid);
2036 int osp_orphan_it_rec(const struct lu_env *env, const struct dt_it *di,
2037 struct dt_rec *rec, __u32 attr)
2039 struct osp_it *it = (struct osp_it *)di;
2040 struct lu_orphan_ent *ent = (struct lu_orphan_ent *)it->ooi_ent;
2042 if (likely(ent != NULL)) {
2043 *(struct lu_orphan_rec *)rec = ent->loe_rec;
2050 __u64 osp_it_store(const struct lu_env *env, const struct dt_it *di)
2052 struct osp_it *it = (struct osp_it *)di;
2054 return it->ooi_next;
2058 * Locate the iteration cursor to the specified position (cookie).
2060 * \param[in] env pointer to the thread context
2061 * \param[in] di pointer to the iteration structure
2062 * \param[in] hash the specified position
2064 * \retval positive number for locating to the exactly position
2066 * \retval 0 for arriving at the end of the iteration
2067 * \retval negative error number on failure
2069 int osp_orphan_it_load(const struct lu_env *env, const struct dt_it *di,
2072 struct osp_it *it = (struct osp_it *)di;
2075 it->ooi_next = hash;
2076 rc = osp_orphan_it_next(env, (struct dt_it *)di);
2086 int osp_it_key_rec(const struct lu_env *env, const struct dt_it *di,
2092 static const struct dt_index_operations osp_orphan_index_ops = {
2093 .dio_lookup = osp_orphan_index_lookup,
2094 .dio_declare_insert = osp_orphan_index_declare_insert,
2095 .dio_insert = osp_orphan_index_insert,
2096 .dio_declare_delete = osp_orphan_index_declare_delete,
2097 .dio_delete = osp_orphan_index_delete,
2099 .init = osp_it_init,
2100 .fini = osp_it_fini,
2101 .next = osp_orphan_it_next,
2104 .key = osp_orphan_it_key,
2105 .key_size = osp_orphan_it_key_size,
2106 .rec = osp_orphan_it_rec,
2107 .store = osp_it_store,
2108 .load = osp_orphan_it_load,
2109 .key_rec = osp_it_key_rec,
2114 * Implement OSP layer dt_object_operations::do_index_try() interface.
2116 * Negotiate the index type.
2118 * If the target index is an IDIF object, then use osp_orphan_index_ops.
2119 * Otherwise, assign osp_md_index_ops to the dt_object::do_index_ops.
2120 * (\see lustre/include/lustre_fid.h for IDIF.)
2122 * \param[in] env pointer to the thread context
2123 * \param[in] dt pointer to the OSP layer dt_object
2124 * \param[in] feat unused
2126 * \retval 0 for success
2128 static int osp_index_try(const struct lu_env *env,
2129 struct dt_object *dt,
2130 const struct dt_index_features *feat)
2132 const struct lu_fid *fid = lu_object_fid(&dt->do_lu);
2134 if (fid_is_last_id(fid) && fid_is_idif(fid))
2135 dt->do_index_ops = &osp_orphan_index_ops;
2137 dt->do_index_ops = &osp_md_index_ops;
2141 struct dt_object_operations osp_obj_ops = {
2142 .do_declare_attr_get = osp_declare_attr_get,
2143 .do_attr_get = osp_attr_get,
2144 .do_declare_attr_set = osp_declare_attr_set,
2145 .do_attr_set = osp_attr_set,
2146 .do_declare_xattr_get = osp_declare_xattr_get,
2147 .do_xattr_get = osp_xattr_get,
2148 .do_declare_xattr_set = osp_declare_xattr_set,
2149 .do_xattr_set = osp_xattr_set,
2150 .do_declare_create = osp_declare_object_create,
2151 .do_create = osp_object_create,
2152 .do_declare_destroy = osp_declare_object_destroy,
2153 .do_destroy = osp_object_destroy,
2154 .do_index_try = osp_index_try,
2158 * Implement OSP layer lu_object_operations::loo_object_init() interface.
2160 * Initialize the object.
2162 * If it is a remote MDT object, then call do_attr_get() to fetch
2163 * the attribute from the peer.
2165 * \param[in] env pointer to the thread context
2166 * \param[in] o pointer to the OSP layer lu_object
2167 * \param[in] conf unused
2169 * \retval 0 for success
2170 * \retval negative error number on failure
2172 static int osp_object_init(const struct lu_env *env, struct lu_object *o,
2173 const struct lu_object_conf *conf)
2175 struct osp_object *po = lu2osp_obj(o);
2179 spin_lock_init(&po->opo_lock);
2180 o->lo_header->loh_attr |= LOHA_REMOTE;
2182 if (is_ost_obj(o)) {
2183 po->opo_obj.do_ops = &osp_obj_ops;
2185 struct lu_attr *la = &osp_env_info(env)->osi_attr;
2187 po->opo_obj.do_ops = &osp_md_obj_ops;
2188 po->opo_obj.do_body_ops = &osp_md_body_ops;
2189 rc = po->opo_obj.do_ops->do_attr_get(env, lu2dt_obj(o),
2192 o->lo_header->loh_attr |=
2193 LOHA_EXISTS | (la->la_mode & S_IFMT);
2194 if (rc == -ENOENT) {
2195 po->opo_non_exist = 1;
2198 init_rwsem(&po->opo_sem);
2204 * Implement OSP layer lu_object_operations::loo_object_free() interface.
2206 * Finalize the object.
2208 * If the OSP object has attributes cache, then destroy the cache.
2209 * Free the object finally.
2211 * \param[in] env pointer to the thread context
2212 * \param[in] o pointer to the OSP layer lu_object
2214 static void osp_object_free(const struct lu_env *env, struct lu_object *o)
2216 struct osp_object *obj = lu2osp_obj(o);
2217 struct lu_object_header *h = o->lo_header;
2219 dt_object_fini(&obj->opo_obj);
2220 lu_object_header_fini(h);
2221 if (obj->opo_ooa != NULL) {
2222 struct osp_xattr_entry *oxe;
2223 struct osp_xattr_entry *tmp;
2226 list_for_each_entry_safe(oxe, tmp,
2227 &obj->opo_ooa->ooa_xattr_list,
2229 list_del(&oxe->oxe_list);
2230 count = atomic_read(&oxe->oxe_ref);
2231 LASSERTF(count == 1,
2232 "Still has %d users on the xattr entry %.*s\n",
2233 count-1, (int)oxe->oxe_namelen, oxe->oxe_buf);
2235 OBD_FREE(oxe, oxe->oxe_buflen);
2237 OBD_FREE_PTR(obj->opo_ooa);
2239 OBD_SLAB_FREE_PTR(obj, osp_object_kmem);
2243 * Implement OSP layer lu_object_operations::loo_object_release() interface.
2245 * Cleanup (not free) the object.
2247 * If it is a reserved object but failed to be created, or it is an OST
2248 * object, then mark the object as non-cached.
2250 * \param[in] env pointer to the thread context
2251 * \param[in] o pointer to the OSP layer lu_object
2253 static void osp_object_release(const struct lu_env *env, struct lu_object *o)
2255 struct osp_object *po = lu2osp_obj(o);
2256 struct osp_device *d = lu2osp_dev(o->lo_dev);
2261 * release reservation if object was declared but not created
2262 * this may require lu_object_put() in LOD
2264 if (unlikely(po->opo_reserved)) {
2265 LASSERT(d->opd_pre != NULL);
2266 LASSERT(d->opd_pre_reserved > 0);
2267 spin_lock(&d->opd_pre_lock);
2268 d->opd_pre_reserved--;
2269 spin_unlock(&d->opd_pre_lock);
2271 /* not needed in cache any more */
2272 set_bit(LU_OBJECT_HEARD_BANSHEE, &o->lo_header->loh_flags);
2276 /* XXX: Currently, NOT cache OST-object on MDT because:
2277 * 1. it is not often accessed on MDT.
2278 * 2. avoid up layer (such as LFSCK) to load too many
2279 * once-used OST-objects. */
2280 set_bit(LU_OBJECT_HEARD_BANSHEE, &o->lo_header->loh_flags);
2285 static int osp_object_print(const struct lu_env *env, void *cookie,
2286 lu_printer_t p, const struct lu_object *l)
2288 const struct osp_object *o = lu2osp_obj((struct lu_object *)l);
2290 return (*p)(env, cookie, LUSTRE_OSP_NAME"-object@%p", o);
2293 static int osp_object_invariant(const struct lu_object *o)
2298 struct lu_object_operations osp_lu_obj_ops = {
2299 .loo_object_init = osp_object_init,
2300 .loo_object_free = osp_object_free,
2301 .loo_object_release = osp_object_release,
2302 .loo_object_print = osp_object_print,
2303 .loo_object_invariant = osp_object_invariant