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.
89 * XXX: NOT prepare out RPC for remote transaction. ((please refer to the
90 * comment of osp_trans_create() for remote transaction)
92 * According to our current transaction/dt_object_lock framework (to make
93 * the cross-MDTs modification for DNE1 to be workable), the transaction
94 * sponsor will start the transaction firstly, then try to acquire related
95 * dt_object_lock if needed. Under such rules, if we want to prepare the
96 * OUT RPC in the transaction declare phase, then related attr/xattr
97 * should be known without dt_object_lock. But such condition maybe not
98 * true for some remote transaction case. For example:
100 * For linkEA repairing (by LFSCK) case, before the LFSCK thread obtained
101 * the dt_object_lock on the target MDT-object, it cannot know whether
102 * the MDT-object has linkEA or not, neither invalid or not.
104 * Since the LFSCK thread cannot hold dt_object_lock before the remote
105 * transaction start (otherwise there will be some potential deadlock),
106 * it cannot prepare related OUT RPC for repairing during the declare
107 * phase as other normal transactions do.
109 * To resolve the trouble, we will make OSP to prepare related OUT RPC
110 * after remote transaction started, and trigger the remote updating
111 * (send RPC) when trans_stop. Then the up layer users, such as LFSCK,
112 * can follow the general rule to handle trans_start/dt_object_lock
113 * for repairing linkEA inconsistency without distinguishing remote
116 * In fact, above solution for remote transaction should be the normal
117 * model without considering DNE1. The trouble brought by DNE1 will be
118 * resolved in DNE2. At that time, this patch can be removed.
121 * Author: Alex Zhuravlev <alexey.zhuravlev@intel.com>
122 * Author: Mikhail Pershin <mike.tappro@intel.com>
125 #define DEBUG_SUBSYSTEM S_MDS
127 #include "osp_internal.h"
129 static inline __u32 osp_dev2node(struct osp_device *osp)
131 return osp->opd_storage->dd_lu_dev.ld_site->ld_seq_site->ss_node_id;
134 static inline bool is_ost_obj(struct lu_object *lo)
136 return !lu2osp_dev(lo->lo_dev)->opd_connect_mdt;
140 * Assign FID to the OST object.
142 * This function will assign the FID to the OST object of a striped file.
144 * \param[in] env pointer to the thread context
145 * \param[in] d pointer to the OSP device
146 * \param[in] o pointer to the OSP object that the FID will be
149 static void osp_object_assign_fid(const struct lu_env *env,
150 struct osp_device *d, struct osp_object *o)
152 struct osp_thread_info *osi = osp_env_info(env);
154 LASSERT(fid_is_zero(lu_object_fid(&o->opo_obj.do_lu)));
155 LASSERT(o->opo_reserved);
158 osp_precreate_get_fid(env, d, &osi->osi_fid);
160 lu_object_assign_fid(env, &o->opo_obj.do_lu, &osi->osi_fid);
164 * Initialize the OSP object attributes cache.
166 * \param[in] obj pointer to the OSP object
168 * \retval 0 for success
169 * \retval negative error number on failure
171 static int osp_oac_init(struct osp_object *obj)
173 struct osp_object_attr *ooa;
179 INIT_LIST_HEAD(&ooa->ooa_xattr_list);
180 spin_lock(&obj->opo_lock);
181 if (likely(obj->opo_ooa == NULL)) {
183 spin_unlock(&obj->opo_lock);
185 spin_unlock(&obj->opo_lock);
193 * Find the named extended attribute in the OSP object attributes cache.
195 * The caller should take the osp_object::opo_lock before calling
198 * \param[in] ooa pointer to the OSP object attributes cache
199 * \param[in] name the name of the extended attribute
200 * \param[in] namelen the name length of the extended attribute
202 * \retval pointer to the found extended attribute entry
203 * \retval NULL if the specified extended attribute is not
206 static struct osp_xattr_entry *
207 osp_oac_xattr_find_locked(struct osp_object_attr *ooa,
208 const char *name, size_t namelen)
210 struct osp_xattr_entry *oxe;
212 list_for_each_entry(oxe, &ooa->ooa_xattr_list, oxe_list) {
213 if (namelen == oxe->oxe_namelen &&
214 strncmp(name, oxe->oxe_buf, namelen) == 0)
222 * Find the named extended attribute in the OSP object attributes cache.
224 * Call osp_oac_xattr_find_locked() with the osp_object::opo_lock held.
226 * \param[in] obj pointer to the OSP object
227 * \param[in] name the name of the extended attribute
228 * \param[in] unlink true if the extended attribute entry is to be removed
231 * \retval pointer to the found extended attribute entry
232 * \retval NULL if the specified extended attribute is not
235 static struct osp_xattr_entry *osp_oac_xattr_find(struct osp_object *obj,
236 const char *name, bool unlink)
238 struct osp_xattr_entry *oxe = NULL;
240 spin_lock(&obj->opo_lock);
241 if (obj->opo_ooa != NULL) {
242 oxe = osp_oac_xattr_find_locked(obj->opo_ooa, name,
246 list_del_init(&oxe->oxe_list);
248 atomic_inc(&oxe->oxe_ref);
251 spin_unlock(&obj->opo_lock);
257 * Find the named extended attribute in the OSP object attributes cache.
259 * If it is not in the cache, then add an empty entry (that will be
260 * filled later) to cache with the given name.
262 * \param[in] obj pointer to the OSP object
263 * \param[in] name the name of the extended attribute
264 * \param[in] len the length of the extended attribute value
266 * \retval pointer to the found or new-created extended
268 * \retval NULL if the specified extended attribute is not in the
269 * cache or fail to add new empty entry to the cache.
271 static struct osp_xattr_entry *
272 osp_oac_xattr_find_or_add(struct osp_object *obj, const char *name, size_t len)
274 struct osp_object_attr *ooa = obj->opo_ooa;
275 struct osp_xattr_entry *oxe;
276 struct osp_xattr_entry *tmp = NULL;
277 size_t namelen = strlen(name);
278 size_t size = sizeof(*oxe) + namelen + 1 + len;
280 LASSERT(ooa != NULL);
282 oxe = osp_oac_xattr_find(obj, name, false);
286 OBD_ALLOC(oxe, size);
287 if (unlikely(oxe == NULL))
290 INIT_LIST_HEAD(&oxe->oxe_list);
291 oxe->oxe_buflen = size;
292 oxe->oxe_namelen = namelen;
293 memcpy(oxe->oxe_buf, name, namelen);
294 oxe->oxe_value = oxe->oxe_buf + namelen + 1;
295 /* One ref is for the caller, the other is for the entry on the list. */
296 atomic_set(&oxe->oxe_ref, 2);
298 spin_lock(&obj->opo_lock);
299 tmp = osp_oac_xattr_find_locked(ooa, name, namelen);
301 list_add_tail(&oxe->oxe_list, &ooa->ooa_xattr_list);
303 atomic_inc(&tmp->oxe_ref);
304 spin_unlock(&obj->opo_lock);
315 * Add the given extended attribute to the OSP object attributes cache.
317 * If there is an old extended attributed entry with the same name,
318 * remove it from the cache and return it via the parameter \a poxe.
320 * \param[in] obj pointer to the OSP object
321 * \param[in,out] poxe double pointer to the OSP object extended attribute
322 * entry: the new extended attribute entry is transfered
323 * via such pointer target, and if old the extended
324 * attribute entry exists, then it will be returned back
325 * via such pointer target.
326 * \param[in] len the length of the (new) extended attribute value
328 * \retval pointer to the new extended attribute entry
329 * \retval NULL for failure cases.
331 static struct osp_xattr_entry *
332 osp_oac_xattr_replace(struct osp_object *obj,
333 struct osp_xattr_entry **poxe, size_t len)
335 struct osp_object_attr *ooa = obj->opo_ooa;
336 struct osp_xattr_entry *oxe;
337 size_t namelen = (*poxe)->oxe_namelen;
338 size_t size = sizeof(*oxe) + namelen + 1 + len;
340 LASSERT(ooa != NULL);
342 OBD_ALLOC(oxe, size);
343 if (unlikely(oxe == NULL))
346 INIT_LIST_HEAD(&oxe->oxe_list);
347 oxe->oxe_buflen = size;
348 oxe->oxe_namelen = namelen;
349 memcpy(oxe->oxe_buf, (*poxe)->oxe_buf, namelen);
350 oxe->oxe_value = oxe->oxe_buf + namelen + 1;
351 /* One ref is for the caller, the other is for the entry on the list. */
352 atomic_set(&oxe->oxe_ref, 2);
354 spin_lock(&obj->opo_lock);
355 *poxe = osp_oac_xattr_find_locked(ooa, oxe->oxe_buf, namelen);
356 LASSERT(*poxe != NULL);
358 list_del_init(&(*poxe)->oxe_list);
359 list_add_tail(&oxe->oxe_list, &ooa->ooa_xattr_list);
360 spin_unlock(&obj->opo_lock);
366 * Release reference from the OSP object extended attribute entry.
368 * If it is the last reference, then free the entry.
370 * \param[in] oxe pointer to the OSP object extended attribute entry.
372 static inline void osp_oac_xattr_put(struct osp_xattr_entry *oxe)
374 if (atomic_dec_and_test(&oxe->oxe_ref)) {
375 LASSERT(list_empty(&oxe->oxe_list));
377 OBD_FREE(oxe, oxe->oxe_buflen);
382 * Parse the OSP object attribute from the RPC reply.
384 * If the attribute is valid, then it will be added to the OSP object
387 * \param[in] env pointer to the thread context
388 * \param[in] reply pointer to the RPC reply
389 * \param[in] req pointer to the RPC request
390 * \param[out] attr pointer to buffer to hold the output attribute
391 * \param[in] obj pointer to the OSP object
392 * \param[in] index the index of the attribute buffer in the reply
394 * \retval 0 for success
395 * \retval negative error number on failure
397 static int osp_get_attr_from_reply(const struct lu_env *env,
398 struct object_update_reply *reply,
399 struct ptlrpc_request *req,
400 struct lu_attr *attr,
401 struct osp_object *obj, int index)
403 struct osp_thread_info *osi = osp_env_info(env);
404 struct lu_buf *rbuf = &osi->osi_lb2;
405 struct obdo *lobdo = &osi->osi_obdo;
409 rc = object_update_result_data_get(reply, rbuf, index);
413 wobdo = rbuf->lb_buf;
414 if (rbuf->lb_len != sizeof(*wobdo))
417 LASSERT(req != NULL);
418 if (ptlrpc_req_need_swab(req))
419 lustre_swab_obdo(wobdo);
421 lustre_get_wire_obdo(NULL, lobdo, wobdo);
422 spin_lock(&obj->opo_lock);
423 if (obj->opo_ooa != NULL) {
424 la_from_obdo(&obj->opo_ooa->ooa_attr, lobdo, lobdo->o_valid);
426 *attr = obj->opo_ooa->ooa_attr;
428 LASSERT(attr != NULL);
430 la_from_obdo(attr, lobdo, lobdo->o_valid);
432 spin_unlock(&obj->opo_lock);
438 * Interpreter function for getting OSP object attribute asynchronously.
440 * Called to interpret the result of an async mode RPC for getting the
441 * OSP object attribute.
443 * \param[in] env pointer to the thread context
444 * \param[in] reply pointer to the RPC reply
445 * \param[in] req pointer to the RPC request
446 * \param[in] obj pointer to the OSP object
447 * \param[out] data pointer to buffer to hold the output attribute
448 * \param[in] index the index of the attribute buffer in the reply
449 * \param[in] rc the result for handling the RPC
451 * \retval 0 for success
452 * \retval negative error number on failure
454 static int osp_attr_get_interpterer(const struct lu_env *env,
455 struct object_update_reply *reply,
456 struct ptlrpc_request *req,
457 struct osp_object *obj,
458 void *data, int index, int rc)
460 struct lu_attr *attr = data;
462 LASSERT(obj->opo_ooa != NULL);
465 osp2lu_obj(obj)->lo_header->loh_attr |= LOHA_EXISTS;
466 obj->opo_non_exist = 0;
468 return osp_get_attr_from_reply(env, reply, req, NULL, obj,
472 osp2lu_obj(obj)->lo_header->loh_attr &= ~LOHA_EXISTS;
473 obj->opo_non_exist = 1;
476 spin_lock(&obj->opo_lock);
478 spin_unlock(&obj->opo_lock);
485 * Implement OSP layer dt_object_operations::do_declare_attr_get() interface.
487 * Declare that the caller will get attribute from the specified OST object.
489 * This function adds an Object Unified Target (OUT) sub-request to the per-OSP
490 * based shared asynchronous request queue. The osp_attr_get_interpterer()
491 * is registered as the interpreter function to handle the result of this
494 * \param[in] env pointer to the thread context
495 * \param[in] dt pointer to the OSP layer dt_object
496 * \param[in] capa the capability for this operation
498 * \retval 0 for success
499 * \retval negative error number on failure
501 static int osp_declare_attr_get(const struct lu_env *env, struct dt_object *dt,
502 struct lustre_capa *capa)
504 struct osp_object *obj = dt2osp_obj(dt);
505 struct osp_device *osp = lu2osp_dev(dt->do_lu.lo_dev);
508 if (obj->opo_ooa == NULL) {
509 rc = osp_oac_init(obj);
514 mutex_lock(&osp->opd_async_requests_mutex);
515 rc = osp_insert_async_request(env, OUT_ATTR_GET, obj, 0, NULL, NULL,
516 &obj->opo_ooa->ooa_attr,
517 osp_attr_get_interpterer);
518 mutex_unlock(&osp->opd_async_requests_mutex);
524 * Implement OSP layer dt_object_operations::do_attr_get() interface.
526 * Get attribute from the specified MDT/OST object.
528 * If the attribute is in the OSP object attributes cache, then return
529 * the cached attribute directly. Otherwise it will trigger an OUT RPC
530 * to the peer to get the attribute synchronously, if successful, add it
531 * to the OSP attributes cache. (\see lustre/osp/osp_trans.c for OUT RPC.)
533 * \param[in] env pointer to the thread context
534 * \param[in] dt pointer to the OSP layer dt_object
535 * \param[out] attr pointer to the buffer to hold the output attribute
536 * \param[in] capa the capability for this operation
538 * \retval 0 for success
539 * \retval negative error number on failure
541 int osp_attr_get(const struct lu_env *env, struct dt_object *dt,
542 struct lu_attr *attr, struct lustre_capa *capa)
544 struct osp_device *osp = lu2osp_dev(dt->do_lu.lo_dev);
545 struct osp_object *obj = dt2osp_obj(dt);
546 struct dt_device *dev = &osp->opd_dt_dev;
547 struct dt_update_request *update;
548 struct object_update_reply *reply;
549 struct ptlrpc_request *req = NULL;
553 if (is_ost_obj(&dt->do_lu) && obj->opo_non_exist)
556 if (obj->opo_ooa != NULL) {
557 spin_lock(&obj->opo_lock);
558 if (obj->opo_ooa->ooa_attr.la_valid != 0) {
559 *attr = obj->opo_ooa->ooa_attr;
560 spin_unlock(&obj->opo_lock);
564 spin_unlock(&obj->opo_lock);
567 update = dt_update_request_create(dev);
569 RETURN(PTR_ERR(update));
571 rc = out_attr_get_pack(env, &update->dur_buf,
572 lu_object_fid(&dt->do_lu));
574 CERROR("%s: Insert update error "DFID": rc = %d\n",
575 dev->dd_lu_dev.ld_obd->obd_name,
576 PFID(lu_object_fid(&dt->do_lu)), rc);
581 rc = osp_remote_sync(env, osp, update, &req, false);
584 osp2lu_obj(obj)->lo_header->loh_attr &= ~LOHA_EXISTS;
585 obj->opo_non_exist = 1;
587 CERROR("%s:osp_attr_get update error "DFID": rc = %d\n",
588 dev->dd_lu_dev.ld_obd->obd_name,
589 PFID(lu_object_fid(&dt->do_lu)), rc);
595 osp2lu_obj(obj)->lo_header->loh_attr |= LOHA_EXISTS;
596 obj->opo_non_exist = 0;
597 reply = req_capsule_server_sized_get(&req->rq_pill,
598 &RMF_OUT_UPDATE_REPLY,
599 OUT_UPDATE_REPLY_SIZE);
600 if (reply == NULL || reply->ourp_magic != UPDATE_REPLY_MAGIC)
601 GOTO(out, rc = -EPROTO);
603 rc = osp_get_attr_from_reply(env, reply, req, attr, obj, 0);
611 ptlrpc_req_finished(req);
613 dt_update_request_destroy(update);
618 static int __osp_attr_set(const struct lu_env *env, struct dt_object *dt,
619 const struct lu_attr *attr, struct thandle *th)
621 struct osp_device *d = lu2osp_dev(dt->do_lu.lo_dev);
622 struct osp_object *o = dt2osp_obj(dt);
628 * Usually we don't allow server stack to manipulate size
629 * but there is a special case when striping is created
630 * late, after stripeless file got truncated to non-zero.
632 * In this case we do the following:
634 * 1) grab id in declare - this can lead to leaked OST objects
635 * but we don't currently have proper mechanism and the only
636 * options we have are to do truncate RPC holding transaction
637 * open (very bad) or to grab id in declare at cost of leaked
638 * OST object in same very rare unfortunate case (just bad)
639 * notice 1.6-2.0 do assignment outside of running transaction
640 * all the time, meaning many more chances for leaked objects.
642 * 2) send synchronous truncate RPC with just assigned id
645 /* there are few places in MDD code still passing NULL
646 * XXX: to be fixed soon */
650 if (attr->la_valid & LA_SIZE && attr->la_size > 0 &&
651 fid_is_zero(lu_object_fid(&o->opo_obj.do_lu))) {
652 LASSERT(!dt_object_exists(dt));
653 osp_object_assign_fid(env, d, o);
654 rc = osp_object_truncate(env, dt, attr->la_size);
659 if (!(attr->la_valid & (LA_UID | LA_GID)))
662 if (!is_only_remote_trans(th)) {
664 * track all UID/GID changes via llog
666 rc = osp_sync_declare_add(env, o, MDS_SETATTR64_REC, th);
668 /* It is for OST-object attr_set directly without updating
669 * local MDT-object attribute. It is usually used by LFSCK. */
670 rc = __osp_md_attr_set(env, dt, attr, th);
673 if (rc != 0 || o->opo_ooa == NULL)
676 /* Update the OSP object attributes cache. */
677 la = &o->opo_ooa->ooa_attr;
678 spin_lock(&o->opo_lock);
679 if (attr->la_valid & LA_UID) {
680 la->la_uid = attr->la_uid;
681 la->la_valid |= LA_UID;
684 if (attr->la_valid & LA_GID) {
685 la->la_gid = attr->la_gid;
686 la->la_valid |= LA_GID;
688 spin_unlock(&o->opo_lock);
694 * Implement OSP layer dt_object_operations::do_declare_attr_set() interface.
696 * If the transaction is not remote one, then declare the credits that will
697 * be used for the subsequent llog record for the object's attributes.
699 * \param[in] env pointer to the thread context
700 * \param[in] dt pointer to the OSP layer dt_object
701 * \param[in] attr pointer to the attribute to be set
702 * \param[in] th pointer to the transaction handler
704 * \retval 0 for success
705 * \retval negative error number on failure
707 static int osp_declare_attr_set(const struct lu_env *env, struct dt_object *dt,
708 const struct lu_attr *attr, struct thandle *th)
712 if (!is_only_remote_trans(th)) {
713 rc = __osp_attr_set(env, dt, attr, th);
715 CDEBUG(D_INFO, "declare set attr "DFID": rc = %d\n",
716 PFID(&dt->do_lu.lo_header->loh_fid), rc);
723 * Implement OSP layer dt_object_operations::do_attr_set() interface.
725 * Set attribute to the specified OST object.
727 * If the transaction is a remote one, then add OUT_ATTR_SET sub-request
728 * in the OUT RPC that will be flushed when the remote transaction stop.
729 * Otherwise, it will generate a MDS_SETATTR64_REC record in the llog that
730 * will be handled by a dedicated thread asynchronously.
732 * If the attribute entry exists in the OSP object attributes cache,
733 * then update the cached attribute according to given attribute.
735 * \param[in] env pointer to the thread context
736 * \param[in] dt pointer to the OSP layer dt_object
737 * \param[in] attr pointer to the attribute to be set
738 * \param[in] th pointer to the transaction handler
739 * \param[in] capa the capability for this operation
741 * \retval 0 for success
742 * \retval negative error number on failure
744 static int osp_attr_set(const struct lu_env *env, struct dt_object *dt,
745 const struct lu_attr *attr, struct thandle *th,
746 struct lustre_capa *capa)
748 struct osp_object *o = dt2osp_obj(dt);
752 if (is_only_remote_trans(th)) {
753 rc = __osp_attr_set(env, dt, attr, th);
755 CDEBUG(D_INFO, "(1) set attr "DFID": rc = %d\n",
756 PFID(&dt->do_lu.lo_header->loh_fid), rc);
761 /* we're interested in uid/gid changes only */
762 if (!(attr->la_valid & (LA_UID | LA_GID)))
765 rc = osp_sync_add(env, o, MDS_SETATTR64_REC, th, attr);
766 /* XXX: send new uid/gid to OST ASAP? */
768 CDEBUG(D_INFO, "(2) set attr "DFID": rc = %d\n",
769 PFID(&dt->do_lu.lo_header->loh_fid), rc);
775 * Interpreter function for getting OSP object extended attribute asynchronously
777 * Called to interpret the result of an async mode RPC for getting the
778 * OSP object extended attribute.
780 * \param[in] env pointer to the thread context
781 * \param[in] reply pointer to the RPC reply
782 * \param[in] req pointer to the RPC request
783 * \param[in] obj pointer to the OSP object
784 * \param[out] data pointer to OSP object attributes cache
785 * \param[in] index the index of the attribute buffer in the reply
786 * \param[in] rc the result for handling the RPC
788 * \retval 0 for success
789 * \retval negative error number on failure
791 static int osp_xattr_get_interpterer(const struct lu_env *env,
792 struct object_update_reply *reply,
793 struct ptlrpc_request *req,
794 struct osp_object *obj,
795 void *data, int index, int rc)
797 struct osp_object_attr *ooa = obj->opo_ooa;
798 struct osp_xattr_entry *oxe = data;
799 struct lu_buf *rbuf = &osp_env_info(env)->osi_lb2;
801 LASSERT(ooa != NULL);
804 size_t len = sizeof(*oxe) + oxe->oxe_namelen + 1;
806 rc = object_update_result_data_get(reply, rbuf, index);
807 if (rc < 0 || rbuf->lb_len > (oxe->oxe_buflen - len)) {
808 spin_lock(&obj->opo_lock);
810 spin_unlock(&obj->opo_lock);
811 osp_oac_xattr_put(oxe);
813 return rc < 0 ? rc : -ERANGE;
816 spin_lock(&obj->opo_lock);
817 oxe->oxe_vallen = rbuf->lb_len;
818 memcpy(oxe->oxe_value, rbuf->lb_buf, rbuf->lb_len);
821 spin_unlock(&obj->opo_lock);
822 } else if (rc == -ENOENT || rc == -ENODATA) {
823 spin_lock(&obj->opo_lock);
826 spin_unlock(&obj->opo_lock);
828 spin_lock(&obj->opo_lock);
830 spin_unlock(&obj->opo_lock);
833 osp_oac_xattr_put(oxe);
839 * Implement OSP dt_object_operations::do_declare_xattr_get() interface.
841 * Declare that the caller will get extended attribute from the specified
844 * This function will add an OUT_XATTR_GET sub-request to the per OSP
845 * based shared asynchronous request queue with the interpreter function:
846 * osp_xattr_get_interpterer().
848 * \param[in] env pointer to the thread context
849 * \param[in] dt pointer to the OSP layer dt_object
850 * \param[out] buf pointer to the lu_buf to hold the extended attribute
851 * \param[in] name the name for the expected extended attribute
852 * \param[in] capa the capability for this operation
854 * \retval 0 for success
855 * \retval negative error number on failure
857 static int osp_declare_xattr_get(const struct lu_env *env, struct dt_object *dt,
858 struct lu_buf *buf, const char *name,
859 struct lustre_capa *capa)
861 struct osp_object *obj = dt2osp_obj(dt);
862 struct osp_device *osp = lu2osp_dev(dt->do_lu.lo_dev);
863 struct osp_xattr_entry *oxe;
864 __u16 namelen = strlen(name);
867 LASSERT(buf != NULL);
868 LASSERT(name != NULL);
870 /* If only for xattr size, return directly. */
871 if (unlikely(buf->lb_len == 0))
874 if (obj->opo_ooa == NULL) {
875 rc = osp_oac_init(obj);
880 oxe = osp_oac_xattr_find_or_add(obj, name, buf->lb_len);
884 mutex_lock(&osp->opd_async_requests_mutex);
885 rc = osp_insert_async_request(env, OUT_XATTR_GET, obj, 1,
886 &namelen, (const void **)&name, oxe,
887 osp_xattr_get_interpterer);
889 mutex_unlock(&osp->opd_async_requests_mutex);
890 osp_oac_xattr_put(oxe);
892 struct dt_update_request *update;
894 /* XXX: Currently, we trigger the batched async OUT
895 * RPC via dt_declare_xattr_get(). It is not
896 * perfect solution, but works well now.
898 * We will improve it in the future. */
899 update = osp->opd_async_requests;
900 if (update != NULL && update->dur_buf.ub_req != NULL &&
901 update->dur_buf.ub_req->ourq_count > 0) {
902 osp->opd_async_requests = NULL;
903 mutex_unlock(&osp->opd_async_requests_mutex);
904 rc = osp_unplug_async_request(env, osp, update);
906 mutex_unlock(&osp->opd_async_requests_mutex);
914 * Implement OSP layer dt_object_operations::do_xattr_get() interface.
916 * Get extended attribute from the specified MDT/OST object.
918 * If the extended attribute is in the OSP object attributes cache, then
919 * return the cached extended attribute directly. Otherwise it will get
920 * the extended attribute synchronously, if successful, add it to the OSP
921 * attributes cache. (\see lustre/osp/osp_trans.c for OUT RPC.)
923 * There is a race condition: some other thread has added the named extended
924 * attributed entry to the OSP object attributes cache during the current
925 * OUT_XATTR_GET handling. If such case happens, the OSP will replace the
926 * (just) existing extended attribute entry with the new replied one.
928 * \param[in] env pointer to the thread context
929 * \param[in] dt pointer to the OSP layer dt_object
930 * \param[out] buf pointer to the lu_buf to hold the extended attribute
931 * \param[in] name the name for the expected extended attribute
932 * \param[in] capa the capability for this operation
934 * \retval 0 for success
935 * \retval negative error number on failure
937 int osp_xattr_get(const struct lu_env *env, struct dt_object *dt,
938 struct lu_buf *buf, const char *name,
939 struct lustre_capa *capa)
941 struct osp_device *osp = lu2osp_dev(dt->do_lu.lo_dev);
942 struct osp_object *obj = dt2osp_obj(dt);
943 struct dt_device *dev = &osp->opd_dt_dev;
944 struct lu_buf *rbuf = &osp_env_info(env)->osi_lb2;
945 struct dt_update_request *update = NULL;
946 struct ptlrpc_request *req = NULL;
947 struct object_update_reply *reply;
948 struct osp_xattr_entry *oxe = NULL;
949 const char *dname = dt->do_lu.lo_dev->ld_obd->obd_name;
953 LASSERT(buf != NULL);
954 LASSERT(name != NULL);
956 if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_NETWORK) &&
957 osp->opd_index == cfs_fail_val) {
958 if (is_ost_obj(&dt->do_lu)) {
959 if (osp_dev2node(osp) == cfs_fail_val)
962 if (strcmp(name, XATTR_NAME_LINK) == 0)
967 if (unlikely(obj->opo_non_exist))
970 oxe = osp_oac_xattr_find(obj, name, false);
972 spin_lock(&obj->opo_lock);
973 if (oxe->oxe_ready) {
975 GOTO(unlock, rc = -ENODATA);
977 if (buf->lb_buf == NULL)
978 GOTO(unlock, rc = oxe->oxe_vallen);
980 if (buf->lb_len < oxe->oxe_vallen)
981 GOTO(unlock, rc = -ERANGE);
983 memcpy(buf->lb_buf, oxe->oxe_value, oxe->oxe_vallen);
985 GOTO(unlock, rc = oxe->oxe_vallen);
988 spin_unlock(&obj->opo_lock);
989 osp_oac_xattr_put(oxe);
993 spin_unlock(&obj->opo_lock);
996 update = dt_update_request_create(dev);
998 GOTO(out, rc = PTR_ERR(update));
1000 rc = out_xattr_get_pack(env, &update->dur_buf,
1001 lu_object_fid(&dt->do_lu), name);
1003 CERROR("%s: Insert update error "DFID": rc = %d\n",
1004 dname, PFID(lu_object_fid(&dt->do_lu)), rc);
1008 rc = osp_remote_sync(env, osp, update, &req, false);
1010 if (rc == -ENOENT) {
1011 dt->do_lu.lo_header->loh_attr &= ~LOHA_EXISTS;
1012 obj->opo_non_exist = 1;
1015 if (obj->opo_ooa == NULL)
1019 oxe = osp_oac_xattr_find_or_add(obj, name, buf->lb_len);
1022 CWARN("%s: Fail to add xattr (%s) to cache for "
1023 DFID" (1): rc = %d\n", dname, name,
1024 PFID(lu_object_fid(&dt->do_lu)), rc);
1029 spin_lock(&obj->opo_lock);
1030 if (rc == -ENOENT || rc == -ENODATA) {
1036 spin_unlock(&obj->opo_lock);
1041 reply = req_capsule_server_sized_get(&req->rq_pill,
1042 &RMF_OUT_UPDATE_REPLY,
1043 OUT_UPDATE_REPLY_SIZE);
1044 if (reply->ourp_magic != UPDATE_REPLY_MAGIC) {
1045 CERROR("%s: Wrong version %x expected %x "DFID": rc = %d\n",
1046 dname, reply->ourp_magic, UPDATE_REPLY_MAGIC,
1047 PFID(lu_object_fid(&dt->do_lu)), -EPROTO);
1049 GOTO(out, rc = -EPROTO);
1052 rc = object_update_result_data_get(reply, rbuf, 0);
1056 if (buf->lb_buf == NULL)
1057 GOTO(out, rc = rbuf->lb_len);
1059 if (unlikely(buf->lb_len < rbuf->lb_len))
1060 GOTO(out, rc = -ERANGE);
1062 memcpy(buf->lb_buf, rbuf->lb_buf, rbuf->lb_len);
1064 if (obj->opo_ooa == NULL)
1068 oxe = osp_oac_xattr_find_or_add(obj, name, rbuf->lb_len);
1070 CWARN("%s: Fail to add xattr (%s) to "
1071 "cache for "DFID" (2): rc = %d\n",
1072 dname, name, PFID(lu_object_fid(&dt->do_lu)), rc);
1078 if (oxe->oxe_buflen - oxe->oxe_namelen - 1 < rbuf->lb_len) {
1079 struct osp_xattr_entry *old = oxe;
1080 struct osp_xattr_entry *tmp;
1082 tmp = osp_oac_xattr_replace(obj, &old, rbuf->lb_len);
1083 osp_oac_xattr_put(oxe);
1086 CWARN("%s: Fail to update xattr (%s) to "
1087 "cache for "DFID": rc = %d\n",
1088 dname, name, PFID(lu_object_fid(&dt->do_lu)), rc);
1089 spin_lock(&obj->opo_lock);
1091 spin_unlock(&obj->opo_lock);
1096 /* Drop the ref for entry on list. */
1097 osp_oac_xattr_put(old);
1100 spin_lock(&obj->opo_lock);
1101 oxe->oxe_vallen = rbuf->lb_len;
1102 memcpy(oxe->oxe_value, rbuf->lb_buf, rbuf->lb_len);
1105 spin_unlock(&obj->opo_lock);
1111 ptlrpc_req_finished(req);
1113 if (update != NULL && !IS_ERR(update))
1114 dt_update_request_destroy(update);
1117 osp_oac_xattr_put(oxe);
1122 static int __osp_xattr_set(const struct lu_env *env, struct dt_object *dt,
1123 const struct lu_buf *buf, const char *name,
1124 int flag, struct thandle *th)
1126 struct osp_object *o = dt2osp_obj(dt);
1127 struct dt_update_request *update;
1128 struct osp_xattr_entry *oxe;
1132 LASSERT(buf->lb_len > 0 && buf->lb_buf != NULL);
1134 update = dt_update_request_find_or_create(th, dt);
1135 if (IS_ERR(update)) {
1136 CERROR("%s: Get OSP update buf failed "DFID": rc = %d\n",
1137 dt->do_lu.lo_dev->ld_obd->obd_name,
1138 PFID(lu_object_fid(&dt->do_lu)),
1139 (int)PTR_ERR(update));
1141 RETURN(PTR_ERR(update));
1144 rc = out_xattr_set_pack(env, &update->dur_buf,
1145 lu_object_fid(&dt->do_lu),
1146 buf, name, flag, update->dur_batchid);
1147 if (rc != 0 || o->opo_ooa == NULL)
1150 oxe = osp_oac_xattr_find_or_add(o, name, buf->lb_len);
1152 CWARN("%s: cannot cache xattr '%s' of "DFID"\n",
1153 dt->do_lu.lo_dev->ld_obd->obd_name,
1154 name, PFID(lu_object_fid(&dt->do_lu)));
1159 if (oxe->oxe_buflen - oxe->oxe_namelen - 1 < buf->lb_len) {
1160 struct osp_xattr_entry *old = oxe;
1161 struct osp_xattr_entry *tmp;
1163 tmp = osp_oac_xattr_replace(o, &old, buf->lb_len);
1164 osp_oac_xattr_put(oxe);
1167 CWARN("%s: cannot update cached xattr '%s' of "DFID"\n",
1168 dt->do_lu.lo_dev->ld_obd->obd_name,
1169 name, PFID(lu_object_fid(&dt->do_lu)));
1170 spin_lock(&o->opo_lock);
1172 spin_unlock(&o->opo_lock);
1177 /* Drop the ref for entry on list. */
1178 osp_oac_xattr_put(old);
1181 spin_lock(&o->opo_lock);
1182 oxe->oxe_vallen = buf->lb_len;
1183 memcpy(oxe->oxe_value, buf->lb_buf, buf->lb_len);
1186 spin_unlock(&o->opo_lock);
1187 osp_oac_xattr_put(oxe);
1193 * Implement OSP layer dt_object_operations::do_declare_xattr_set() interface.
1195 * Declare that the caller will set extended attribute to the specified
1198 * If it is non-remote transaction, it will add an OUT_XATTR_SET sub-request
1199 * to the OUT RPC that will be flushed when the transaction start. And if the
1200 * OSP attributes cache is initialized, then check whether the name extended
1201 * attribute entry exists in the cache or not. If yes, replace it; otherwise,
1202 * add the extended attribute to the cache.
1204 * \param[in] env pointer to the thread context
1205 * \param[in] dt pointer to the OSP layer dt_object
1206 * \param[in] buf pointer to the lu_buf to hold the extended attribute
1207 * \param[in] name the name of the extended attribute to be set
1208 * \param[in] flag to indicate the detailed set operation: LU_XATTR_CREATE
1209 * or LU_XATTR_REPLACE or others
1210 * \param[in] th pointer to the transaction handler
1212 * \retval 0 for success
1213 * \retval negative error number on failure
1215 int osp_declare_xattr_set(const struct lu_env *env, struct dt_object *dt,
1216 const struct lu_buf *buf, const char *name,
1217 int flag, struct thandle *th)
1221 if (!is_only_remote_trans(th)) {
1222 rc = __osp_xattr_set(env, dt, buf, name, flag, th);
1224 CDEBUG(D_INFO, "declare xattr %s set object "DFID": rc = %d\n",
1225 name, PFID(&dt->do_lu.lo_header->loh_fid), rc);
1232 * Implement OSP layer dt_object_operations::do_xattr_set() interface.
1234 * Set extended attribute to the specified MDT/OST object.
1236 * If it is remote transaction, it will add an OUT_XATTR_SET sub-request into
1237 * the OUT RPC that will be flushed when the transaction stop. And if the OSP
1238 * attributes cache is initialized, then check whether the name extended
1239 * attribute entry exists in the cache or not. If yes, replace it; otherwise,
1240 * add the extended attribute to the cache.
1242 * \param[in] env pointer to the thread context
1243 * \param[in] dt pointer to the OSP layer dt_object
1244 * \param[in] buf pointer to the lu_buf to hold the extended attribute
1245 * \param[in] name the name of the extended attribute to be set
1246 * \param[in] fl to indicate the detailed set operation: LU_XATTR_CREATE
1247 * or LU_XATTR_REPLACE or others
1248 * \param[in] th pointer to the transaction handler
1249 * \param[in] capa the capability for this operation
1251 * \retval 0 for success
1252 * \retval negative error number on failure
1254 int osp_xattr_set(const struct lu_env *env, struct dt_object *dt,
1255 const struct lu_buf *buf, const char *name, int fl,
1256 struct thandle *th, struct lustre_capa *capa)
1260 if (is_only_remote_trans(th)) {
1261 rc = __osp_xattr_set(env, dt, buf, name, fl, th);
1263 CDEBUG(D_INFO, "xattr %s set object "DFID": rc = %d\n",
1264 name, PFID(&dt->do_lu.lo_header->loh_fid), rc);
1270 static int __osp_xattr_del(const struct lu_env *env, struct dt_object *dt,
1271 const char *name, struct thandle *th)
1273 struct dt_update_request *update;
1274 const struct lu_fid *fid;
1275 struct osp_object *o = dt2osp_obj(dt);
1276 struct osp_xattr_entry *oxe;
1279 update = dt_update_request_find_or_create(th, dt);
1281 return PTR_ERR(update);
1283 fid = lu_object_fid(&dt->do_lu);
1285 rc = out_xattr_del_pack(env, &update->dur_buf, fid, name,
1286 update->dur_batchid);
1288 if (rc != 0 || o->opo_ooa == NULL)
1291 oxe = osp_oac_xattr_find(o, name, true);
1293 /* Drop the ref for entry on list. */
1294 osp_oac_xattr_put(oxe);
1300 * Implement OSP layer dt_object_operations::do_declare_xattr_del() interface.
1302 * Declare that the caller will delete extended attribute on the specified
1305 * If it is non-remote transaction, it will add an OUT_XATTR_DEL sub-request
1306 * to the OUT RPC that will be flushed when the transaction start. And if the
1307 * name extended attribute entry exists in the OSP attributes cache, then remove
1308 * it from the cache.
1310 * \param[in] env pointer to the thread context
1311 * \param[in] dt pointer to the OSP layer dt_object
1312 * \param[in] name the name of the extended attribute to be set
1313 * \param[in] th pointer to the transaction handler
1315 * \retval 0 for success
1316 * \retval negative error number on failure
1318 int osp_declare_xattr_del(const struct lu_env *env, struct dt_object *dt,
1319 const char *name, struct thandle *th)
1323 if (!is_only_remote_trans(th)) {
1324 rc = __osp_xattr_del(env, dt, name, th);
1326 CDEBUG(D_INFO, "declare xattr %s del object "DFID": rc = %d\n",
1327 name, PFID(&dt->do_lu.lo_header->loh_fid), rc);
1334 * Implement OSP layer dt_object_operations::do_xattr_del() interface.
1336 * Delete extended attribute on the specified MDT/OST object.
1338 * If it is remote transaction, it will add an OUT_XATTR_DEL sub-request into
1339 * the OUT RPC that will be flushed when the transaction stop. And if the name
1340 * extended attribute entry exists in the OSP attributes cache, then remove it
1343 * \param[in] env pointer to the thread context
1344 * \param[in] dt pointer to the OSP layer dt_object
1345 * \param[in] name the name of the extended attribute to be set
1346 * \param[in] th pointer to the transaction handler
1347 * \param[in] capa the capability for this operation
1349 * \retval 0 for success
1350 * \retval negative error number on failure
1352 int osp_xattr_del(const struct lu_env *env, struct dt_object *dt,
1353 const char *name, struct thandle *th,
1354 struct lustre_capa *capa)
1358 if (is_only_remote_trans(th)) {
1359 rc = __osp_xattr_del(env, dt, name, th);
1361 CDEBUG(D_INFO, "xattr %s del object "DFID": rc = %d\n",
1362 name, PFID(&dt->do_lu.lo_header->loh_fid), rc);
1369 * Implement OSP layer dt_object_operations::do_declare_create() interface.
1371 * Declare that the caller will create the OST object.
1373 * If the transaction is a remote transaction and the FID for the OST-object
1374 * has been assigned already, then handle it as creating (remote) MDT object
1375 * via osp_md_declare_object_create(). This function is usually used for LFSCK
1376 * to re-create the lost OST object. Otherwise, if it is not replay case, the
1377 * OSP will reserve pre-created object for the subsequent create operation;
1378 * if the MDT side cached pre-created objects are less than some threshold,
1379 * then it will wakeup the pre-create thread.
1381 * \param[in] env pointer to the thread context
1382 * \param[in] dt pointer to the OSP layer dt_object
1383 * \param[in] attr the attribute for the object to be created
1384 * \param[in] hint pointer to the hint for creating the object, such as
1386 * \param[in] dof pointer to the dt_object_format for help the creation
1387 * \param[in] th pointer to the transaction handler
1389 * \retval 0 for success
1390 * \retval negative error number on failure
1392 static int osp_declare_object_create(const struct lu_env *env,
1393 struct dt_object *dt,
1394 struct lu_attr *attr,
1395 struct dt_allocation_hint *hint,
1396 struct dt_object_format *dof,
1399 struct osp_thread_info *osi = osp_env_info(env);
1400 struct osp_device *d = lu2osp_dev(dt->do_lu.lo_dev);
1401 struct osp_object *o = dt2osp_obj(dt);
1402 const struct lu_fid *fid = lu_object_fid(&dt->do_lu);
1407 if (is_only_remote_trans(th) && !fid_is_zero(fid)) {
1408 LASSERT(fid_is_sane(fid));
1410 rc = osp_md_declare_object_create(env, dt, attr, hint, dof, th);
1415 /* should happen to non-0 OSP only so that at least one object
1416 * has been already declared in the scenario and LOD should
1418 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_CREATE_FAIL) && d->opd_index == 1)
1421 LASSERT(d->opd_last_used_oid_file);
1424 * There can be gaps in precreated ids and record to unlink llog
1425 * XXX: we do not handle gaps yet, implemented before solution
1426 * was found to be racy, so we disabled that. there is no
1427 * point in making useless but expensive llog declaration.
1429 /* rc = osp_sync_declare_add(env, o, MDS_UNLINK64_REC, th); */
1431 if (unlikely(!fid_is_zero(fid))) {
1432 /* replay case: caller knows fid */
1433 osi->osi_off = sizeof(osi->osi_id) * d->opd_index;
1434 osi->osi_lb.lb_len = sizeof(osi->osi_id);
1435 osi->osi_lb.lb_buf = NULL;
1436 rc = dt_declare_record_write(env, d->opd_last_used_oid_file,
1437 &osi->osi_lb, osi->osi_off, th);
1442 * in declaration we need to reserve object so that we don't block
1443 * awaiting precreation RPC to complete
1445 rc = osp_precreate_reserve(env, d);
1447 * we also need to declare update to local "last used id" file for
1448 * recovery if object isn't used for a reason, we need to release
1449 * reservation, this can be made in osd_object_release()
1452 /* mark id is reserved: in create we don't want to talk
1454 LASSERT(o->opo_reserved == 0);
1455 o->opo_reserved = 1;
1457 /* common for all OSPs file hystorically */
1458 osi->osi_off = sizeof(osi->osi_id) * d->opd_index;
1459 osi->osi_lb.lb_len = sizeof(osi->osi_id);
1460 osi->osi_lb.lb_buf = NULL;
1461 rc = dt_declare_record_write(env, d->opd_last_used_oid_file,
1462 &osi->osi_lb, osi->osi_off, th);
1464 /* not needed in the cache anymore */
1465 set_bit(LU_OBJECT_HEARD_BANSHEE,
1466 &dt->do_lu.lo_header->loh_flags);
1472 * Implement OSP layer dt_object_operations::do_create() interface.
1474 * Create the OST object.
1476 * If the transaction is a remote transaction and the FID for the OST-object
1477 * has been assigned already, then handle it as handling MDT object via the
1478 * osp_md_object_create(). For other cases, the OSP will assign FID to the
1479 * object to be created, and update last_used Object ID (OID) file.
1481 * \param[in] env pointer to the thread context
1482 * \param[in] dt pointer to the OSP layer dt_object
1483 * \param[in] attr the attribute for the object to be created
1484 * \param[in] hint pointer to the hint for creating the object, such as
1486 * \param[in] dof pointer to the dt_object_format for help the creation
1487 * \param[in] th pointer to the transaction handler
1489 * \retval 0 for success
1490 * \retval negative error number on failure
1492 static int osp_object_create(const struct lu_env *env, struct dt_object *dt,
1493 struct lu_attr *attr,
1494 struct dt_allocation_hint *hint,
1495 struct dt_object_format *dof, struct thandle *th)
1497 struct osp_thread_info *osi = osp_env_info(env);
1498 struct osp_device *d = lu2osp_dev(dt->do_lu.lo_dev);
1499 struct osp_object *o = dt2osp_obj(dt);
1501 struct lu_fid *fid = &osi->osi_fid;
1504 if (is_only_remote_trans(th) &&
1505 !fid_is_zero(lu_object_fid(&dt->do_lu))) {
1506 LASSERT(fid_is_sane(lu_object_fid(&dt->do_lu)));
1508 rc = osp_md_object_create(env, dt, attr, hint, dof, th);
1510 o->opo_non_exist = 0;
1515 o->opo_non_exist = 0;
1516 if (o->opo_reserved) {
1517 /* regular case, fid is assigned holding transaction open */
1518 osp_object_assign_fid(env, d, o);
1521 memcpy(fid, lu_object_fid(&dt->do_lu), sizeof(*fid));
1523 LASSERTF(fid_is_sane(fid), "fid for osp_object %p is insane"DFID"!\n",
1526 if (!o->opo_reserved) {
1527 /* special case, id was assigned outside of transaction
1528 * see comments in osp_declare_attr_set */
1529 LASSERT(d->opd_pre != NULL);
1530 spin_lock(&d->opd_pre_lock);
1531 osp_update_last_fid(d, fid);
1532 spin_unlock(&d->opd_pre_lock);
1535 CDEBUG(D_INODE, "fid for osp_object %p is "DFID"\n", o, PFID(fid));
1537 /* If the precreate ends, it means it will be ready to rollover to
1538 * the new sequence soon, all the creation should be synchronized,
1539 * otherwise during replay, the replay fid will be inconsistent with
1540 * last_used/create fid */
1541 if (osp_precreate_end_seq(env, d) && osp_is_fid_client(d))
1545 * it's OK if the import is inactive by this moment - id was created
1546 * by OST earlier, we just need to maintain it consistently on the disk
1547 * once import is reconnected, OSP will claim this and other objects
1548 * used and OST either keep them, if they exist or recreate
1551 /* we might have lost precreated objects */
1552 if (unlikely(d->opd_gap_count) > 0) {
1553 LASSERT(d->opd_pre != NULL);
1554 spin_lock(&d->opd_pre_lock);
1555 if (d->opd_gap_count > 0) {
1556 int count = d->opd_gap_count;
1558 ostid_set_id(&osi->osi_oi,
1559 fid_oid(&d->opd_gap_start_fid));
1560 d->opd_gap_count = 0;
1561 spin_unlock(&d->opd_pre_lock);
1563 CDEBUG(D_HA, "Writing gap "DFID"+%d in llog\n",
1564 PFID(&d->opd_gap_start_fid), count);
1565 /* real gap handling is disabled intil ORI-692 will be
1566 * fixed, now we only report gaps */
1568 spin_unlock(&d->opd_pre_lock);
1572 /* Only need update last_used oid file, seq file will only be update
1573 * during seq rollover */
1574 osp_objid_buf_prep(&osi->osi_lb, &osi->osi_off,
1575 &d->opd_last_used_fid.f_oid, d->opd_index);
1577 rc = dt_record_write(env, d->opd_last_used_oid_file, &osi->osi_lb,
1580 CDEBUG(D_HA, "%s: Wrote last used FID: "DFID", index %d: %d\n",
1581 d->opd_obd->obd_name, PFID(fid), d->opd_index, rc);
1587 * Implement OSP layer dt_object_operations::do_declare_destroy() interface.
1589 * Declare that the caller will destroy the specified OST object.
1591 * The OST object destroy will be handled via llog asynchronously. This
1592 * function will declare the credits for generating MDS_UNLINK64_REC llog.
1594 * \param[in] env pointer to the thread context
1595 * \param[in] dt pointer to the OSP layer dt_object to be destroyed
1596 * \param[in] th pointer to the transaction handler
1598 * \retval 0 for success
1599 * \retval negative error number on failure
1601 int osp_declare_object_destroy(const struct lu_env *env,
1602 struct dt_object *dt, struct thandle *th)
1604 struct osp_object *o = dt2osp_obj(dt);
1610 * track objects to be destroyed via llog
1612 rc = osp_sync_declare_add(env, o, MDS_UNLINK64_REC, th);
1618 * Implement OSP layer dt_object_operations::do_destroy() interface.
1620 * Destroy the specified OST object.
1622 * The OSP generates a MDS_UNLINK64_REC record in the llog. There
1623 * will be some dedicated thread to handle the llog asynchronously.
1625 * It also marks the object as non-cached.
1627 * \param[in] env pointer to the thread context
1628 * \param[in] dt pointer to the OSP layer dt_object to be destroyed
1629 * \param[in] th pointer to the transaction handler
1631 * \retval 0 for success
1632 * \retval negative error number on failure
1634 int osp_object_destroy(const struct lu_env *env, struct dt_object *dt,
1637 struct osp_object *o = dt2osp_obj(dt);
1642 o->opo_non_exist = 1;
1644 * once transaction is committed put proper command on
1645 * the queue going to our OST
1647 rc = osp_sync_add(env, o, MDS_UNLINK64_REC, th, NULL);
1649 /* not needed in cache any more */
1650 set_bit(LU_OBJECT_HEARD_BANSHEE, &dt->do_lu.lo_header->loh_flags);
1655 static int osp_orphan_index_lookup(const struct lu_env *env,
1656 struct dt_object *dt,
1658 const struct dt_key *key,
1659 struct lustre_capa *capa)
1664 static int osp_orphan_index_declare_insert(const struct lu_env *env,
1665 struct dt_object *dt,
1666 const struct dt_rec *rec,
1667 const struct dt_key *key,
1668 struct thandle *handle)
1673 static int osp_orphan_index_insert(const struct lu_env *env,
1674 struct dt_object *dt,
1675 const struct dt_rec *rec,
1676 const struct dt_key *key,
1677 struct thandle *handle,
1678 struct lustre_capa *capa,
1684 static int osp_orphan_index_declare_delete(const struct lu_env *env,
1685 struct dt_object *dt,
1686 const struct dt_key *key,
1687 struct thandle *handle)
1692 static int osp_orphan_index_delete(const struct lu_env *env,
1693 struct dt_object *dt,
1694 const struct dt_key *key,
1695 struct thandle *handle,
1696 struct lustre_capa *capa)
1702 * Initialize the OSP layer index iteration.
1704 * \param[in] env pointer to the thread context
1705 * \param[in] dt pointer to the index object to be iterated
1706 * \param[in] attr unused
1707 * \param[in] capa the capability for this operation
1709 * \retval pointer to the iteration structure
1710 * \retval negative error number on failure
1712 struct dt_it *osp_it_init(const struct lu_env *env, struct dt_object *dt,
1713 __u32 attr, struct lustre_capa *capa)
1719 return ERR_PTR(-ENOMEM);
1721 it->ooi_pos_ent = -1;
1723 it->ooi_attr = attr;
1725 return (struct dt_it *)it;
1729 * Finalize the OSP layer index iteration.
1731 * \param[in] env pointer to the thread context
1732 * \param[in] di pointer to the iteration structure
1734 void osp_it_fini(const struct lu_env *env, struct dt_it *di)
1736 struct osp_it *it = (struct osp_it *)di;
1737 struct page **pages = it->ooi_pages;
1738 int npages = it->ooi_total_npages;
1741 if (pages != NULL) {
1742 for (i = 0; i < npages; i++) {
1743 if (pages[i] != NULL) {
1744 if (pages[i] == it->ooi_cur_page) {
1746 it->ooi_cur_page = NULL;
1748 __free_page(pages[i]);
1751 OBD_FREE(pages, npages * sizeof(*pages));
1757 * Get more records for the iteration from peer.
1759 * The new records will be filled in an array of pages. The OSP side
1760 * allows 1MB bulk data to be transfered.
1762 * \param[in] env pointer to the thread context
1763 * \param[in] it pointer to the iteration structure
1765 * \retval 0 for success
1766 * \retval negative error number on failure
1768 static int osp_it_fetch(const struct lu_env *env, struct osp_it *it)
1770 struct lu_device *dev = it->ooi_obj->do_lu.lo_dev;
1771 struct osp_device *osp = lu2osp_dev(dev);
1772 struct page **pages;
1773 struct ptlrpc_request *req = NULL;
1774 struct ptlrpc_bulk_desc *desc;
1775 struct idx_info *ii;
1782 npages = min_t(unsigned int, OFD_MAX_BRW_SIZE, 1 << 20);
1783 npages /= PAGE_CACHE_SIZE;
1785 OBD_ALLOC(pages, npages * sizeof(*pages));
1789 it->ooi_pages = pages;
1790 it->ooi_total_npages = npages;
1791 for (i = 0; i < npages; i++) {
1792 pages[i] = alloc_page(GFP_IOFS);
1793 if (pages[i] == NULL)
1797 req = ptlrpc_request_alloc(osp->opd_obd->u.cli.cl_import,
1802 rc = ptlrpc_request_pack(req, LUSTRE_OBD_VERSION, OBD_IDX_READ);
1804 ptlrpc_request_free(req);
1808 req->rq_request_portal = OUT_PORTAL;
1809 ii = req_capsule_client_get(&req->rq_pill, &RMF_IDX_INFO);
1810 memset(ii, 0, sizeof(*ii));
1811 if (fid_is_last_id(lu_object_fid(&it->ooi_obj->do_lu))) {
1812 /* LFSCK will iterate orphan object[FID_SEQ_LAYOUT_BTREE,
1813 * ost_index, 0] with LAST_ID FID, so it needs to replace
1814 * the FID with orphan FID here */
1815 ii->ii_fid.f_seq = FID_SEQ_LAYOUT_RBTREE;
1816 ii->ii_fid.f_oid = osp->opd_index;
1817 ii->ii_fid.f_ver = 0;
1818 ii->ii_flags = II_FL_NOHASH;
1819 ii->ii_attrs = osp_dev2node(osp);
1821 ii->ii_fid = *lu_object_fid(&it->ooi_obj->do_lu);
1822 ii->ii_flags = II_FL_NOHASH | II_FL_NOKEY | II_FL_VARKEY |
1824 ii->ii_attrs = it->ooi_attr;
1826 ii->ii_magic = IDX_INFO_MAGIC;
1827 ii->ii_count = npages * LU_PAGE_COUNT;
1828 ii->ii_hash_start = it->ooi_next;
1830 ptlrpc_at_set_req_timeout(req);
1832 desc = ptlrpc_prep_bulk_imp(req, npages, 1, BULK_PUT_SINK,
1835 ptlrpc_request_free(req);
1839 for (i = 0; i < npages; i++)
1840 ptlrpc_prep_bulk_page_pin(desc, pages[i], 0, PAGE_CACHE_SIZE);
1842 ptlrpc_request_set_replen(req);
1843 rc = ptlrpc_queue_wait(req);
1847 rc = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk,
1848 req->rq_bulk->bd_nob_transferred);
1853 ii = req_capsule_server_get(&req->rq_pill, &RMF_IDX_INFO);
1854 if (ii->ii_magic != IDX_INFO_MAGIC)
1855 GOTO(out, rc = -EPROTO);
1857 npages = (ii->ii_count + LU_PAGE_COUNT - 1) >>
1858 (PAGE_CACHE_SHIFT - LU_PAGE_SHIFT);
1859 if (npages > it->ooi_total_npages) {
1860 CERROR("%s: returned more pages than expected, %u > %u\n",
1861 osp->opd_obd->obd_name, npages, it->ooi_total_npages);
1862 GOTO(out, rc = -EINVAL);
1865 it->ooi_valid_npages = npages;
1866 if (ptlrpc_rep_need_swab(req))
1869 it->ooi_next = ii->ii_hash_end;
1872 ptlrpc_req_finished(req);
1878 * Move the iteration cursor to the next lu_page.
1880 * One system page (PAGE_SIZE) may contain multiple lu_page (4KB),
1881 * that depends on the LU_PAGE_COUNT. If it is not the last lu_page
1882 * in current system page, then move the iteration cursor to the next
1883 * lu_page in current system page. Otherwise, if there are more system
1884 * pages in the cache, then move the iteration cursor to the next system
1885 * page. If all the cached records (pages) have been iterated, then fetch
1886 * more records via osp_it_fetch().
1888 * \param[in] env pointer to the thread context
1889 * \param[in] di pointer to the iteration structure
1891 * \retval positive for end of the directory
1892 * \retval 0 for success
1893 * \retval negative error number on failure
1895 int osp_it_next_page(const struct lu_env *env, struct dt_it *di)
1897 struct osp_it *it = (struct osp_it *)di;
1898 struct lu_idxpage *idxpage;
1899 struct page **pages;
1905 idxpage = it->ooi_cur_idxpage;
1906 if (idxpage != NULL) {
1907 if (idxpage->lip_nr == 0)
1910 if (it->ooi_pos_ent < idxpage->lip_nr) {
1911 CDEBUG(D_INFO, "ooi_pos %d nr %d\n",
1912 (int)it->ooi_pos_ent, (int)idxpage->lip_nr);
1915 it->ooi_cur_idxpage = NULL;
1916 it->ooi_pos_lu_page++;
1919 if (it->ooi_pos_lu_page < LU_PAGE_COUNT) {
1920 it->ooi_cur_idxpage = (void *)it->ooi_cur_page +
1921 LU_PAGE_SIZE * it->ooi_pos_lu_page;
1923 lustre_swab_lip_header(it->ooi_cur_idxpage);
1924 if (it->ooi_cur_idxpage->lip_magic != LIP_MAGIC) {
1925 struct osp_device *osp =
1926 lu2osp_dev(it->ooi_obj->do_lu.lo_dev);
1928 CERROR("%s: invalid magic (%x != %x) for page "
1929 "%d/%d while read layout orphan index\n",
1930 osp->opd_obd->obd_name,
1931 it->ooi_cur_idxpage->lip_magic,
1932 LIP_MAGIC, it->ooi_pos_page,
1933 it->ooi_pos_lu_page);
1934 /* Skip this lu_page next time. */
1935 it->ooi_pos_ent = idxpage->lip_nr - 1;
1938 it->ooi_pos_ent = -1;
1942 kunmap(it->ooi_cur_page);
1943 it->ooi_cur_page = NULL;
1947 pages = it->ooi_pages;
1948 if (it->ooi_pos_page < it->ooi_valid_npages) {
1949 it->ooi_cur_page = kmap(pages[it->ooi_pos_page]);
1950 it->ooi_pos_lu_page = 0;
1954 for (i = 0; i < it->ooi_total_npages; i++) {
1955 if (pages[i] != NULL)
1956 __free_page(pages[i]);
1958 OBD_FREE(pages, it->ooi_total_npages * sizeof(*pages));
1960 it->ooi_pos_page = 0;
1961 it->ooi_total_npages = 0;
1962 it->ooi_valid_npages = 0;
1965 it->ooi_cur_page = NULL;
1966 it->ooi_cur_idxpage = NULL;
1967 it->ooi_pages = NULL;
1970 if (it->ooi_next == II_END_OFF)
1973 rc = osp_it_fetch(env, it);
1981 * Move the iteration cursor to the next record.
1983 * If there are more records in the lu_page, then move the iteration
1984 * cursor to the next record directly. Otherwise, move the iteration
1985 * cursor to the record in the next lu_page via osp_it_next_page()
1987 * \param[in] env pointer to the thread context
1988 * \param[in] di pointer to the iteration structure
1990 * \retval positive for end of the directory
1991 * \retval 0 for success
1992 * \retval negative error number on failure
1994 static int osp_orphan_it_next(const struct lu_env *env, struct dt_it *di)
1996 struct osp_it *it = (struct osp_it *)di;
1997 struct lu_idxpage *idxpage;
2002 idxpage = it->ooi_cur_idxpage;
2003 if (idxpage != NULL) {
2004 if (idxpage->lip_nr == 0)
2008 if (it->ooi_pos_ent < idxpage->lip_nr) {
2010 (struct lu_orphan_ent *)idxpage->lip_entries +
2013 lustre_swab_orphan_ent(it->ooi_ent);
2018 rc = osp_it_next_page(env, di);
2025 int osp_it_get(const struct lu_env *env, struct dt_it *di,
2026 const struct dt_key *key)
2031 void osp_it_put(const struct lu_env *env, struct dt_it *di)
2035 static struct dt_key *osp_orphan_it_key(const struct lu_env *env,
2036 const struct dt_it *di)
2038 struct osp_it *it = (struct osp_it *)di;
2039 struct lu_orphan_ent *ent = (struct lu_orphan_ent *)it->ooi_ent;
2041 if (likely(ent != NULL))
2042 return (struct dt_key *)(&ent->loe_key);
2047 static int osp_orphan_it_key_size(const struct lu_env *env,
2048 const struct dt_it *di)
2050 return sizeof(struct lu_fid);
2053 static int osp_orphan_it_rec(const struct lu_env *env, const struct dt_it *di,
2054 struct dt_rec *rec, __u32 attr)
2056 struct osp_it *it = (struct osp_it *)di;
2057 struct lu_orphan_ent *ent = (struct lu_orphan_ent *)it->ooi_ent;
2059 if (likely(ent != NULL)) {
2060 *(struct lu_orphan_rec *)rec = ent->loe_rec;
2067 __u64 osp_it_store(const struct lu_env *env, const struct dt_it *di)
2069 struct osp_it *it = (struct osp_it *)di;
2071 return it->ooi_next;
2075 * Locate the iteration cursor to the specified position (cookie).
2077 * \param[in] env pointer to the thread context
2078 * \param[in] di pointer to the iteration structure
2079 * \param[in] hash the specified position
2081 * \retval positive number for locating to the exactly position
2083 * \retval 0 for arriving at the end of the iteration
2084 * \retval negative error number on failure
2086 int osp_orphan_it_load(const struct lu_env *env, const struct dt_it *di,
2089 struct osp_it *it = (struct osp_it *)di;
2092 it->ooi_next = hash;
2093 rc = osp_orphan_it_next(env, (struct dt_it *)di);
2103 int osp_it_key_rec(const struct lu_env *env, const struct dt_it *di,
2109 static const struct dt_index_operations osp_orphan_index_ops = {
2110 .dio_lookup = osp_orphan_index_lookup,
2111 .dio_declare_insert = osp_orphan_index_declare_insert,
2112 .dio_insert = osp_orphan_index_insert,
2113 .dio_declare_delete = osp_orphan_index_declare_delete,
2114 .dio_delete = osp_orphan_index_delete,
2116 .init = osp_it_init,
2117 .fini = osp_it_fini,
2118 .next = osp_orphan_it_next,
2121 .key = osp_orphan_it_key,
2122 .key_size = osp_orphan_it_key_size,
2123 .rec = osp_orphan_it_rec,
2124 .store = osp_it_store,
2125 .load = osp_orphan_it_load,
2126 .key_rec = osp_it_key_rec,
2131 * Implement OSP layer dt_object_operations::do_index_try() interface.
2133 * Negotiate the index type.
2135 * If the target index is an IDIF object, then use osp_orphan_index_ops.
2136 * Otherwise, assign osp_md_index_ops to the dt_object::do_index_ops.
2137 * (\see lustre/include/lustre_fid.h for IDIF.)
2139 * \param[in] env pointer to the thread context
2140 * \param[in] dt pointer to the OSP layer dt_object
2141 * \param[in] feat unused
2143 * \retval 0 for success
2145 static int osp_index_try(const struct lu_env *env,
2146 struct dt_object *dt,
2147 const struct dt_index_features *feat)
2149 const struct lu_fid *fid = lu_object_fid(&dt->do_lu);
2151 if (fid_is_last_id(fid) && fid_is_idif(fid))
2152 dt->do_index_ops = &osp_orphan_index_ops;
2154 dt->do_index_ops = &osp_md_index_ops;
2158 static struct dt_object_operations osp_obj_ops = {
2159 .do_declare_attr_get = osp_declare_attr_get,
2160 .do_attr_get = osp_attr_get,
2161 .do_declare_attr_set = osp_declare_attr_set,
2162 .do_attr_set = osp_attr_set,
2163 .do_declare_xattr_get = osp_declare_xattr_get,
2164 .do_xattr_get = osp_xattr_get,
2165 .do_declare_xattr_set = osp_declare_xattr_set,
2166 .do_xattr_set = osp_xattr_set,
2167 .do_declare_create = osp_declare_object_create,
2168 .do_create = osp_object_create,
2169 .do_declare_destroy = osp_declare_object_destroy,
2170 .do_destroy = osp_object_destroy,
2171 .do_index_try = osp_index_try,
2175 * Implement OSP layer lu_object_operations::loo_object_init() interface.
2177 * Initialize the object.
2179 * If it is a remote MDT object, then call do_attr_get() to fetch
2180 * the attribute from the peer.
2182 * \param[in] env pointer to the thread context
2183 * \param[in] o pointer to the OSP layer lu_object
2184 * \param[in] conf unused
2186 * \retval 0 for success
2187 * \retval negative error number on failure
2189 static int osp_object_init(const struct lu_env *env, struct lu_object *o,
2190 const struct lu_object_conf *conf)
2192 struct osp_object *po = lu2osp_obj(o);
2196 spin_lock_init(&po->opo_lock);
2197 o->lo_header->loh_attr |= LOHA_REMOTE;
2199 if (is_ost_obj(o)) {
2200 po->opo_obj.do_ops = &osp_obj_ops;
2202 struct lu_attr *la = &osp_env_info(env)->osi_attr;
2204 po->opo_obj.do_ops = &osp_md_obj_ops;
2205 po->opo_obj.do_body_ops = &osp_md_body_ops;
2206 rc = po->opo_obj.do_ops->do_attr_get(env, lu2dt_obj(o),
2209 o->lo_header->loh_attr |=
2210 LOHA_EXISTS | (la->la_mode & S_IFMT);
2211 if (rc == -ENOENT) {
2212 po->opo_non_exist = 1;
2215 init_rwsem(&po->opo_sem);
2221 * Implement OSP layer lu_object_operations::loo_object_free() interface.
2223 * Finalize the object.
2225 * If the OSP object has attributes cache, then destroy the cache.
2226 * Free the object finally.
2228 * \param[in] env pointer to the thread context
2229 * \param[in] o pointer to the OSP layer lu_object
2231 static void osp_object_free(const struct lu_env *env, struct lu_object *o)
2233 struct osp_object *obj = lu2osp_obj(o);
2234 struct lu_object_header *h = o->lo_header;
2236 dt_object_fini(&obj->opo_obj);
2237 lu_object_header_fini(h);
2238 if (obj->opo_ooa != NULL) {
2239 struct osp_xattr_entry *oxe;
2240 struct osp_xattr_entry *tmp;
2243 list_for_each_entry_safe(oxe, tmp,
2244 &obj->opo_ooa->ooa_xattr_list,
2246 list_del(&oxe->oxe_list);
2247 count = atomic_read(&oxe->oxe_ref);
2248 LASSERTF(count == 1,
2249 "Still has %d users on the xattr entry %.*s\n",
2250 count-1, (int)oxe->oxe_namelen, oxe->oxe_buf);
2252 OBD_FREE(oxe, oxe->oxe_buflen);
2254 OBD_FREE_PTR(obj->opo_ooa);
2256 OBD_SLAB_FREE_PTR(obj, osp_object_kmem);
2260 * Implement OSP layer lu_object_operations::loo_object_release() interface.
2262 * Cleanup (not free) the object.
2264 * If it is a reserved object but failed to be created, or it is an OST
2265 * object, then mark the object as non-cached.
2267 * \param[in] env pointer to the thread context
2268 * \param[in] o pointer to the OSP layer lu_object
2270 static void osp_object_release(const struct lu_env *env, struct lu_object *o)
2272 struct osp_object *po = lu2osp_obj(o);
2273 struct osp_device *d = lu2osp_dev(o->lo_dev);
2278 * release reservation if object was declared but not created
2279 * this may require lu_object_put() in LOD
2281 if (unlikely(po->opo_reserved)) {
2282 LASSERT(d->opd_pre != NULL);
2283 LASSERT(d->opd_pre_reserved > 0);
2284 spin_lock(&d->opd_pre_lock);
2285 d->opd_pre_reserved--;
2286 spin_unlock(&d->opd_pre_lock);
2288 /* not needed in cache any more */
2289 set_bit(LU_OBJECT_HEARD_BANSHEE, &o->lo_header->loh_flags);
2293 /* XXX: Currently, NOT cache OST-object on MDT because:
2294 * 1. it is not often accessed on MDT.
2295 * 2. avoid up layer (such as LFSCK) to load too many
2296 * once-used OST-objects. */
2297 set_bit(LU_OBJECT_HEARD_BANSHEE, &o->lo_header->loh_flags);
2302 static int osp_object_print(const struct lu_env *env, void *cookie,
2303 lu_printer_t p, const struct lu_object *l)
2305 const struct osp_object *o = lu2osp_obj((struct lu_object *)l);
2307 return (*p)(env, cookie, LUSTRE_OSP_NAME"-object@%p", o);
2310 static int osp_object_invariant(const struct lu_object *o)
2315 struct lu_object_operations osp_lu_obj_ops = {
2316 .loo_object_init = osp_object_init,
2317 .loo_object_free = osp_object_free,
2318 .loo_object_release = osp_object_release,
2319 .loo_object_print = osp_object_print,
2320 .loo_object_invariant = osp_object_invariant