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 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 transferred
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
497 * \retval 0 for success
498 * \retval negative error number on failure
500 static int osp_declare_attr_get(const struct lu_env *env, struct dt_object *dt)
502 struct osp_object *obj = dt2osp_obj(dt);
503 struct osp_device *osp = lu2osp_dev(dt->do_lu.lo_dev);
506 if (obj->opo_ooa == NULL) {
507 rc = osp_oac_init(obj);
512 mutex_lock(&osp->opd_async_requests_mutex);
513 rc = osp_insert_async_request(env, OUT_ATTR_GET, obj, 0, NULL, NULL,
514 &obj->opo_ooa->ooa_attr,
515 osp_attr_get_interpterer);
516 mutex_unlock(&osp->opd_async_requests_mutex);
522 * Implement OSP layer dt_object_operations::do_attr_get() interface.
524 * Get attribute from the specified MDT/OST object.
526 * If the attribute is in the OSP object attributes cache, then return
527 * the cached attribute directly. Otherwise it will trigger an OUT RPC
528 * to the peer to get the attribute synchronously, if successful, add it
529 * to the OSP attributes cache. (\see lustre/osp/osp_trans.c for OUT RPC.)
531 * \param[in] env pointer to the thread context
532 * \param[in] dt pointer to the OSP layer dt_object
533 * \param[out] attr pointer to the buffer to hold the output attribute
535 * \retval 0 for success
536 * \retval negative error number on failure
538 int osp_attr_get(const struct lu_env *env, struct dt_object *dt,
539 struct lu_attr *attr)
541 struct osp_device *osp = lu2osp_dev(dt->do_lu.lo_dev);
542 struct osp_object *obj = dt2osp_obj(dt);
543 struct dt_device *dev = &osp->opd_dt_dev;
544 struct dt_update_request *update;
545 struct object_update_reply *reply;
546 struct ptlrpc_request *req = NULL;
550 if (is_ost_obj(&dt->do_lu) && obj->opo_non_exist)
553 if (obj->opo_ooa != NULL) {
554 spin_lock(&obj->opo_lock);
555 if (obj->opo_ooa->ooa_attr.la_valid != 0) {
556 *attr = obj->opo_ooa->ooa_attr;
557 spin_unlock(&obj->opo_lock);
561 spin_unlock(&obj->opo_lock);
564 update = dt_update_request_create(dev);
566 RETURN(PTR_ERR(update));
568 rc = osp_update_rpc_pack(env, attr_get, update, OUT_ATTR_GET,
569 lu_object_fid(&dt->do_lu));
571 CERROR("%s: Insert update error "DFID": rc = %d\n",
572 dev->dd_lu_dev.ld_obd->obd_name,
573 PFID(lu_object_fid(&dt->do_lu)), rc);
578 rc = osp_remote_sync(env, osp, update, &req);
581 osp2lu_obj(obj)->lo_header->loh_attr &= ~LOHA_EXISTS;
582 obj->opo_non_exist = 1;
584 CERROR("%s:osp_attr_get update error "DFID": rc = %d\n",
585 dev->dd_lu_dev.ld_obd->obd_name,
586 PFID(lu_object_fid(&dt->do_lu)), rc);
592 osp2lu_obj(obj)->lo_header->loh_attr |= LOHA_EXISTS;
593 obj->opo_non_exist = 0;
594 reply = req_capsule_server_sized_get(&req->rq_pill,
595 &RMF_OUT_UPDATE_REPLY,
596 OUT_UPDATE_REPLY_SIZE);
597 if (reply == NULL || reply->ourp_magic != UPDATE_REPLY_MAGIC)
598 GOTO(out, rc = -EPROTO);
600 rc = osp_get_attr_from_reply(env, reply, req, attr, obj, 0);
608 ptlrpc_req_finished(req);
610 dt_update_request_destroy(update);
616 * Implement OSP layer dt_object_operations::do_declare_attr_set() interface.
618 * If the transaction is not remote one, then declare the credits that will
619 * be used for the subsequent llog record for the object's attributes.
621 * \param[in] env pointer to the thread context
622 * \param[in] dt pointer to the OSP layer dt_object
623 * \param[in] attr pointer to the attribute to be set
624 * \param[in] th pointer to the transaction handler
626 * \retval 0 for success
627 * \retval negative error number on failure
629 static int osp_declare_attr_set(const struct lu_env *env, struct dt_object *dt,
630 const struct lu_attr *attr, struct thandle *th)
632 struct osp_device *d = lu2osp_dev(dt->do_lu.lo_dev);
633 struct osp_object *o = dt2osp_obj(dt);
636 if (is_only_remote_trans(th))
637 return osp_md_declare_attr_set(env, dt, attr, th);
639 * Usually we don't allow server stack to manipulate size
640 * but there is a special case when striping is created
641 * late, after stripeless file got truncated to non-zero.
643 * In this case we do the following:
645 * 1) grab id in declare - this can lead to leaked OST objects
646 * but we don't currently have proper mechanism and the only
647 * options we have are to do truncate RPC holding transaction
648 * open (very bad) or to grab id in declare at cost of leaked
649 * OST object in same very rare unfortunate case (just bad)
650 * notice 1.6-2.0 do assignment outside of running transaction
651 * all the time, meaning many more chances for leaked objects.
653 * 2) send synchronous truncate RPC with just assigned id
656 /* there are few places in MDD code still passing NULL
657 * XXX: to be fixed soon */
661 if (attr->la_valid & LA_SIZE && attr->la_size > 0 &&
662 fid_is_zero(lu_object_fid(&o->opo_obj.do_lu))) {
663 LASSERT(!dt_object_exists(dt));
664 osp_object_assign_fid(env, d, o);
665 rc = osp_object_truncate(env, dt, attr->la_size);
670 if (!(attr->la_valid & (LA_UID | LA_GID)))
673 /* track all UID/GID changes via llog */
674 rc = osp_sync_declare_add(env, o, MDS_SETATTR64_REC, th);
680 * Implement OSP layer dt_object_operations::do_attr_set() interface.
682 * Set attribute to the specified OST object.
684 * If the transaction is a remote one, then add OUT_ATTR_SET sub-request
685 * in the OUT RPC that will be flushed when the remote transaction stop.
686 * Otherwise, it will generate a MDS_SETATTR64_REC record in the llog that
687 * will be handled by a dedicated thread asynchronously.
689 * If the attribute entry exists in the OSP object attributes cache,
690 * then update the cached attribute according to given attribute.
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_attr_set(const struct lu_env *env, struct dt_object *dt,
701 const struct lu_attr *attr, struct thandle *th)
703 struct osp_object *o = dt2osp_obj(dt);
707 /* we're interested in uid/gid changes only */
708 if (!(attr->la_valid & (LA_UID | LA_GID)))
711 if (!is_only_remote_trans(th)) {
712 rc = osp_sync_add(env, o, MDS_SETATTR64_REC, th, attr);
713 /* XXX: send new uid/gid to OST ASAP? */
717 /* It is for OST-object attr_set directly without updating
718 * local MDT-object attribute. It is usually used by LFSCK. */
719 rc = osp_md_attr_set(env, dt, attr, th);
720 CDEBUG(D_INFO, "(1) set attr "DFID": rc = %d\n",
721 PFID(&dt->do_lu.lo_header->loh_fid), rc);
723 if (rc != 0 || o->opo_ooa == NULL)
726 /* Update the OSP object attributes cache. */
727 la = &o->opo_ooa->ooa_attr;
728 spin_lock(&o->opo_lock);
729 if (attr->la_valid & LA_UID) {
730 la->la_uid = attr->la_uid;
731 la->la_valid |= LA_UID;
734 if (attr->la_valid & LA_GID) {
735 la->la_gid = attr->la_gid;
736 la->la_valid |= LA_GID;
738 spin_unlock(&o->opo_lock);
745 * Interpreter function for getting OSP object extended attribute asynchronously
747 * Called to interpret the result of an async mode RPC for getting the
748 * OSP object extended attribute.
750 * \param[in] env pointer to the thread context
751 * \param[in] reply pointer to the RPC reply
752 * \param[in] req pointer to the RPC request
753 * \param[in] obj pointer to the OSP object
754 * \param[out] data pointer to OSP object attributes cache
755 * \param[in] index the index of the attribute buffer in the reply
756 * \param[in] rc the result for handling the RPC
758 * \retval 0 for success
759 * \retval negative error number on failure
761 static int osp_xattr_get_interpterer(const struct lu_env *env,
762 struct object_update_reply *reply,
763 struct ptlrpc_request *req,
764 struct osp_object *obj,
765 void *data, int index, int rc)
767 struct osp_object_attr *ooa = obj->opo_ooa;
768 struct osp_xattr_entry *oxe = data;
769 struct lu_buf *rbuf = &osp_env_info(env)->osi_lb2;
771 LASSERT(ooa != NULL);
774 size_t len = sizeof(*oxe) + oxe->oxe_namelen + 1;
776 rc = object_update_result_data_get(reply, rbuf, index);
777 if (rc < 0 || rbuf->lb_len > (oxe->oxe_buflen - len)) {
778 spin_lock(&obj->opo_lock);
780 spin_unlock(&obj->opo_lock);
781 osp_oac_xattr_put(oxe);
783 return rc < 0 ? rc : -ERANGE;
786 spin_lock(&obj->opo_lock);
787 oxe->oxe_vallen = rbuf->lb_len;
788 memcpy(oxe->oxe_value, rbuf->lb_buf, rbuf->lb_len);
791 spin_unlock(&obj->opo_lock);
792 } else if (rc == -ENOENT || rc == -ENODATA) {
793 spin_lock(&obj->opo_lock);
796 spin_unlock(&obj->opo_lock);
798 spin_lock(&obj->opo_lock);
800 spin_unlock(&obj->opo_lock);
803 osp_oac_xattr_put(oxe);
809 * Implement OSP dt_object_operations::do_declare_xattr_get() interface.
811 * Declare that the caller will get extended attribute from the specified
814 * This function will add an OUT_XATTR_GET sub-request to the per OSP
815 * based shared asynchronous request queue with the interpreter function:
816 * osp_xattr_get_interpterer().
818 * \param[in] env pointer to the thread context
819 * \param[in] dt pointer to the OSP layer dt_object
820 * \param[out] buf pointer to the lu_buf to hold the extended attribute
821 * \param[in] name the name for the expected extended attribute
823 * \retval 0 for success
824 * \retval negative error number on failure
826 static int osp_declare_xattr_get(const struct lu_env *env, struct dt_object *dt,
827 struct lu_buf *buf, const char *name)
829 struct osp_object *obj = dt2osp_obj(dt);
830 struct osp_device *osp = lu2osp_dev(dt->do_lu.lo_dev);
831 struct osp_xattr_entry *oxe;
832 __u16 namelen = strlen(name);
835 LASSERT(buf != NULL);
836 LASSERT(name != NULL);
838 /* If only for xattr size, return directly. */
839 if (unlikely(buf->lb_len == 0))
842 if (obj->opo_ooa == NULL) {
843 rc = osp_oac_init(obj);
848 oxe = osp_oac_xattr_find_or_add(obj, name, buf->lb_len);
852 mutex_lock(&osp->opd_async_requests_mutex);
853 rc = osp_insert_async_request(env, OUT_XATTR_GET, obj, 1,
854 &namelen, (const void **)&name, oxe,
855 osp_xattr_get_interpterer);
857 mutex_unlock(&osp->opd_async_requests_mutex);
858 osp_oac_xattr_put(oxe);
860 struct dt_update_request *update;
862 /* XXX: Currently, we trigger the batched async OUT
863 * RPC via dt_declare_xattr_get(). It is not
864 * perfect solution, but works well now.
866 * We will improve it in the future. */
867 update = osp->opd_async_requests;
868 if (update != NULL && update->dur_buf.ub_req != NULL &&
869 update->dur_buf.ub_req->ourq_count > 0) {
870 osp->opd_async_requests = NULL;
871 mutex_unlock(&osp->opd_async_requests_mutex);
872 rc = osp_unplug_async_request(env, osp, update);
874 mutex_unlock(&osp->opd_async_requests_mutex);
882 * Implement OSP layer dt_object_operations::do_xattr_get() interface.
884 * Get extended attribute from the specified MDT/OST object.
886 * If the extended attribute is in the OSP object attributes cache, then
887 * return the cached extended attribute directly. Otherwise it will get
888 * the extended attribute synchronously, if successful, add it to the OSP
889 * attributes cache. (\see lustre/osp/osp_trans.c for OUT RPC.)
891 * There is a race condition: some other thread has added the named extended
892 * attributed entry to the OSP object attributes cache during the current
893 * OUT_XATTR_GET handling. If such case happens, the OSP will replace the
894 * (just) existing extended attribute entry with the new replied one.
896 * \param[in] env pointer to the thread context
897 * \param[in] dt pointer to the OSP layer dt_object
898 * \param[out] buf pointer to the lu_buf to hold the extended attribute
899 * \param[in] name the name for the expected extended attribute
901 * \retval 0 for success
902 * \retval negative error number on failure
904 int osp_xattr_get(const struct lu_env *env, struct dt_object *dt,
905 struct lu_buf *buf, const char *name)
907 struct osp_device *osp = lu2osp_dev(dt->do_lu.lo_dev);
908 struct osp_object *obj = dt2osp_obj(dt);
909 struct dt_device *dev = &osp->opd_dt_dev;
910 struct lu_buf *rbuf = &osp_env_info(env)->osi_lb2;
911 struct dt_update_request *update = NULL;
912 struct ptlrpc_request *req = NULL;
913 struct object_update_reply *reply;
914 struct osp_xattr_entry *oxe = NULL;
915 const char *dname = dt->do_lu.lo_dev->ld_obd->obd_name;
919 LASSERT(buf != NULL);
920 LASSERT(name != NULL);
922 if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_NETWORK) &&
923 osp->opd_index == cfs_fail_val) {
924 if (is_ost_obj(&dt->do_lu)) {
925 if (osp_dev2node(osp) == cfs_fail_val)
928 if (strcmp(name, XATTR_NAME_LINK) == 0)
933 if (unlikely(obj->opo_non_exist))
936 oxe = osp_oac_xattr_find(obj, name, false);
938 spin_lock(&obj->opo_lock);
939 if (oxe->oxe_ready) {
941 GOTO(unlock, rc = -ENODATA);
943 if (buf->lb_buf == NULL)
944 GOTO(unlock, rc = oxe->oxe_vallen);
946 if (buf->lb_len < oxe->oxe_vallen)
947 GOTO(unlock, rc = -ERANGE);
949 memcpy(buf->lb_buf, oxe->oxe_value, oxe->oxe_vallen);
951 GOTO(unlock, rc = oxe->oxe_vallen);
954 spin_unlock(&obj->opo_lock);
955 osp_oac_xattr_put(oxe);
959 spin_unlock(&obj->opo_lock);
962 update = dt_update_request_create(dev);
964 GOTO(out, rc = PTR_ERR(update));
966 rc = osp_update_rpc_pack(env, xattr_get, update, OUT_XATTR_GET,
967 lu_object_fid(&dt->do_lu), name);
969 CERROR("%s: Insert update error "DFID": rc = %d\n",
970 dname, PFID(lu_object_fid(&dt->do_lu)), rc);
974 rc = osp_remote_sync(env, osp, update, &req);
977 dt->do_lu.lo_header->loh_attr &= ~LOHA_EXISTS;
978 obj->opo_non_exist = 1;
981 if (obj->opo_ooa == NULL)
985 oxe = osp_oac_xattr_find_or_add(obj, name, buf->lb_len);
988 CWARN("%s: Fail to add xattr (%s) to cache for "
989 DFID" (1): rc = %d\n", dname, name,
990 PFID(lu_object_fid(&dt->do_lu)), rc);
995 spin_lock(&obj->opo_lock);
996 if (rc == -ENOENT || rc == -ENODATA) {
1002 spin_unlock(&obj->opo_lock);
1007 reply = req_capsule_server_sized_get(&req->rq_pill,
1008 &RMF_OUT_UPDATE_REPLY,
1009 OUT_UPDATE_REPLY_SIZE);
1010 if (reply->ourp_magic != UPDATE_REPLY_MAGIC) {
1011 CERROR("%s: Wrong version %x expected %x "DFID": rc = %d\n",
1012 dname, reply->ourp_magic, UPDATE_REPLY_MAGIC,
1013 PFID(lu_object_fid(&dt->do_lu)), -EPROTO);
1015 GOTO(out, rc = -EPROTO);
1018 rc = object_update_result_data_get(reply, rbuf, 0);
1022 if (buf->lb_buf == NULL)
1023 GOTO(out, rc = rbuf->lb_len);
1025 if (unlikely(buf->lb_len < rbuf->lb_len))
1026 GOTO(out, rc = -ERANGE);
1028 memcpy(buf->lb_buf, rbuf->lb_buf, rbuf->lb_len);
1030 if (obj->opo_ooa == NULL)
1034 oxe = osp_oac_xattr_find_or_add(obj, name, rbuf->lb_len);
1036 CWARN("%s: Fail to add xattr (%s) to "
1037 "cache for "DFID" (2): rc = %d\n",
1038 dname, name, PFID(lu_object_fid(&dt->do_lu)), rc);
1044 if (oxe->oxe_buflen - oxe->oxe_namelen - 1 < rbuf->lb_len) {
1045 struct osp_xattr_entry *old = oxe;
1046 struct osp_xattr_entry *tmp;
1048 tmp = osp_oac_xattr_replace(obj, &old, rbuf->lb_len);
1049 osp_oac_xattr_put(oxe);
1052 CWARN("%s: Fail to update xattr (%s) to "
1053 "cache for "DFID": rc = %d\n",
1054 dname, name, PFID(lu_object_fid(&dt->do_lu)), rc);
1055 spin_lock(&obj->opo_lock);
1057 spin_unlock(&obj->opo_lock);
1062 /* Drop the ref for entry on list. */
1063 osp_oac_xattr_put(old);
1066 spin_lock(&obj->opo_lock);
1067 oxe->oxe_vallen = rbuf->lb_len;
1068 memcpy(oxe->oxe_value, rbuf->lb_buf, rbuf->lb_len);
1071 spin_unlock(&obj->opo_lock);
1077 ptlrpc_req_finished(req);
1079 if (update != NULL && !IS_ERR(update))
1080 dt_update_request_destroy(update);
1083 osp_oac_xattr_put(oxe);
1089 * Implement OSP layer dt_object_operations::do_declare_xattr_set() interface.
1091 * Declare that the caller will set extended attribute to the specified
1094 * If it is non-remote transaction, it will add an OUT_XATTR_SET sub-request
1095 * to the OUT RPC that will be flushed when the transaction start. And if the
1096 * OSP attributes cache is initialized, then check whether the name extended
1097 * attribute entry exists in the cache or not. If yes, replace it; otherwise,
1098 * add the extended attribute to the cache.
1100 * \param[in] env pointer to the thread context
1101 * \param[in] dt pointer to the OSP layer dt_object
1102 * \param[in] buf pointer to the lu_buf to hold the extended attribute
1103 * \param[in] name the name of the extended attribute to be set
1104 * \param[in] flag to indicate the detailed set operation: LU_XATTR_CREATE
1105 * or LU_XATTR_REPLACE or others
1106 * \param[in] th pointer to the transaction handler
1108 * \retval 0 for success
1109 * \retval negative error number on failure
1111 int osp_declare_xattr_set(const struct lu_env *env, struct dt_object *dt,
1112 const struct lu_buf *buf, const char *name,
1113 int flag, struct thandle *th)
1115 return osp_trans_update_request_create(th);
1119 * Implement OSP layer dt_object_operations::do_xattr_set() interface.
1121 * Set extended attribute to the specified MDT/OST object.
1123 * Add an OUT_XATTR_SET sub-request into the OUT RPC that will be flushed in
1124 * the transaction stop. And if the OSP attributes cache is initialized, then
1125 * check whether the name extended attribute entry exists in the cache or not.
1126 * If yes, replace it; otherwise, add the extended attribute to the cache.
1128 * \param[in] env pointer to the thread context
1129 * \param[in] dt pointer to the OSP layer dt_object
1130 * \param[in] buf pointer to the lu_buf to hold the extended attribute
1131 * \param[in] name the name of the extended attribute to be set
1132 * \param[in] fl to indicate the detailed set operation: LU_XATTR_CREATE
1133 * or LU_XATTR_REPLACE or others
1134 * \param[in] th pointer to the transaction handler
1136 * \retval 0 for success
1137 * \retval negative error number on failure
1139 int osp_xattr_set(const struct lu_env *env, struct dt_object *dt,
1140 const struct lu_buf *buf, const char *name, int fl,
1143 struct osp_object *o = dt2osp_obj(dt);
1144 struct dt_update_request *update;
1145 struct osp_xattr_entry *oxe;
1149 update = thandle_to_dt_update_request(th);
1150 LASSERT(update != NULL);
1152 CDEBUG(D_INODE, DFID" set xattr '%s' with size %zd\n",
1153 PFID(lu_object_fid(&dt->do_lu)), name, buf->lb_len);
1155 rc = osp_update_rpc_pack(env, xattr_set, update, OUT_XATTR_SET,
1156 lu_object_fid(&dt->do_lu), buf, name, fl);
1157 if (rc != 0 || o->opo_ooa == NULL)
1160 oxe = osp_oac_xattr_find_or_add(o, name, buf->lb_len);
1162 CWARN("%s: cannot cache xattr '%s' of "DFID"\n",
1163 dt->do_lu.lo_dev->ld_obd->obd_name,
1164 name, PFID(lu_object_fid(&dt->do_lu)));
1169 if (oxe->oxe_buflen - oxe->oxe_namelen - 1 < buf->lb_len) {
1170 struct osp_xattr_entry *old = oxe;
1171 struct osp_xattr_entry *tmp;
1173 tmp = osp_oac_xattr_replace(o, &old, buf->lb_len);
1174 osp_oac_xattr_put(oxe);
1177 CWARN("%s: cannot update cached xattr '%s' of "DFID"\n",
1178 dt->do_lu.lo_dev->ld_obd->obd_name,
1179 name, PFID(lu_object_fid(&dt->do_lu)));
1180 spin_lock(&o->opo_lock);
1182 spin_unlock(&o->opo_lock);
1187 /* Drop the ref for entry on list. */
1188 osp_oac_xattr_put(old);
1191 spin_lock(&o->opo_lock);
1192 oxe->oxe_vallen = buf->lb_len;
1193 memcpy(oxe->oxe_value, buf->lb_buf, buf->lb_len);
1196 spin_unlock(&o->opo_lock);
1197 osp_oac_xattr_put(oxe);
1203 * Implement OSP layer dt_object_operations::do_declare_xattr_del() interface.
1205 * Declare that the caller will delete extended attribute on the specified
1208 * If it is non-remote transaction, it will add an OUT_XATTR_DEL sub-request
1209 * to the OUT RPC that will be flushed when the transaction start. And if the
1210 * name extended attribute entry exists in the OSP attributes cache, then remove
1211 * it from the cache.
1213 * \param[in] env pointer to the thread context
1214 * \param[in] dt pointer to the OSP layer dt_object
1215 * \param[in] name the name of the extended attribute to be set
1216 * \param[in] th pointer to the transaction handler
1218 * \retval 0 for success
1219 * \retval negative error number on failure
1221 int osp_declare_xattr_del(const struct lu_env *env, struct dt_object *dt,
1222 const char *name, struct thandle *th)
1224 return osp_trans_update_request_create(th);
1228 * Implement OSP layer dt_object_operations::do_xattr_del() interface.
1230 * Delete extended attribute on the specified MDT/OST object.
1232 * If it is remote transaction, it will add an OUT_XATTR_DEL sub-request into
1233 * the OUT RPC that will be flushed when the transaction stop. And if the name
1234 * extended attribute entry exists in the OSP attributes cache, then remove it
1237 * \param[in] env pointer to the thread context
1238 * \param[in] dt pointer to the OSP layer dt_object
1239 * \param[in] name the name of the extended attribute to be set
1240 * \param[in] th pointer to the transaction handler
1242 * \retval 0 for success
1243 * \retval negative error number on failure
1245 int osp_xattr_del(const struct lu_env *env, struct dt_object *dt,
1246 const char *name, struct thandle *th)
1248 struct dt_update_request *update;
1249 const struct lu_fid *fid = lu_object_fid(&dt->do_lu);
1250 struct osp_object *o = dt2osp_obj(dt);
1251 struct osp_xattr_entry *oxe;
1254 update = thandle_to_dt_update_request(th);
1255 LASSERT(update != NULL);
1257 rc = osp_update_rpc_pack(env, xattr_del, update, OUT_XATTR_DEL,
1259 if (rc != 0 || o->opo_ooa == NULL)
1262 oxe = osp_oac_xattr_find(o, name, true);
1264 /* Drop the ref for entry on list. */
1265 osp_oac_xattr_put(oxe);
1271 * Implement OSP layer dt_object_operations::do_declare_create() interface.
1273 * Declare that the caller will create the OST object.
1275 * If the transaction is a remote transaction and the FID for the OST-object
1276 * has been assigned already, then handle it as creating (remote) MDT object
1277 * via osp_md_declare_object_create(). This function is usually used for LFSCK
1278 * to re-create the lost OST object. Otherwise, if it is not replay case, the
1279 * OSP will reserve pre-created object for the subsequent create operation;
1280 * if the MDT side cached pre-created objects are less than some threshold,
1281 * then it will wakeup the pre-create thread.
1283 * \param[in] env pointer to the thread context
1284 * \param[in] dt pointer to the OSP layer dt_object
1285 * \param[in] attr the attribute for the object to be created
1286 * \param[in] hint pointer to the hint for creating the object, such as
1288 * \param[in] dof pointer to the dt_object_format for help the creation
1289 * \param[in] th pointer to the transaction handler
1291 * \retval 0 for success
1292 * \retval negative error number on failure
1294 static int osp_declare_object_create(const struct lu_env *env,
1295 struct dt_object *dt,
1296 struct lu_attr *attr,
1297 struct dt_allocation_hint *hint,
1298 struct dt_object_format *dof,
1301 struct osp_thread_info *osi = osp_env_info(env);
1302 struct osp_device *d = lu2osp_dev(dt->do_lu.lo_dev);
1303 struct osp_object *o = dt2osp_obj(dt);
1304 const struct lu_fid *fid = lu_object_fid(&dt->do_lu);
1305 struct thandle *local_th;
1310 if (is_only_remote_trans(th) && !fid_is_zero(fid)) {
1311 LASSERT(fid_is_sane(fid));
1313 rc = osp_md_declare_object_create(env, dt, attr, hint, dof, th);
1318 /* should happen to non-0 OSP only so that at least one object
1319 * has been already declared in the scenario and LOD should
1321 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_CREATE_FAIL) && d->opd_index == 1)
1324 LASSERT(d->opd_last_used_oid_file);
1327 * There can be gaps in precreated ids and record to unlink llog
1328 * XXX: we do not handle gaps yet, implemented before solution
1329 * was found to be racy, so we disabled that. there is no
1330 * point in making useless but expensive llog declaration.
1332 /* rc = osp_sync_declare_add(env, o, MDS_UNLINK64_REC, th); */
1334 local_th = osp_get_storage_thandle(env, th, d);
1335 if (IS_ERR(local_th))
1336 RETURN(PTR_ERR(local_th));
1338 if (unlikely(!fid_is_zero(fid))) {
1339 /* replay case: caller knows fid */
1340 osi->osi_off = sizeof(osi->osi_id) * d->opd_index;
1341 osi->osi_lb.lb_len = sizeof(osi->osi_id);
1342 osi->osi_lb.lb_buf = NULL;
1344 rc = dt_declare_record_write(env, d->opd_last_used_oid_file,
1345 &osi->osi_lb, osi->osi_off,
1351 * in declaration we need to reserve object so that we don't block
1352 * awaiting precreation RPC to complete
1354 rc = osp_precreate_reserve(env, d);
1356 * we also need to declare update to local "last used id" file for
1357 * recovery if object isn't used for a reason, we need to release
1358 * reservation, this can be made in osd_object_release()
1361 /* mark id is reserved: in create we don't want to talk
1363 LASSERT(o->opo_reserved == 0);
1364 o->opo_reserved = 1;
1366 /* common for all OSPs file hystorically */
1367 osi->osi_off = sizeof(osi->osi_id) * d->opd_index;
1368 osi->osi_lb.lb_len = sizeof(osi->osi_id);
1369 osi->osi_lb.lb_buf = NULL;
1370 rc = dt_declare_record_write(env, d->opd_last_used_oid_file,
1371 &osi->osi_lb, osi->osi_off,
1374 /* not needed in the cache anymore */
1375 set_bit(LU_OBJECT_HEARD_BANSHEE,
1376 &dt->do_lu.lo_header->loh_flags);
1382 * Implement OSP layer dt_object_operations::do_create() interface.
1384 * Create the OST object.
1386 * If the transaction is a remote transaction and the FID for the OST-object
1387 * has been assigned already, then handle it as handling MDT object via the
1388 * osp_md_object_create(). For other cases, the OSP will assign FID to the
1389 * object to be created, and update last_used Object ID (OID) file.
1391 * \param[in] env pointer to the thread context
1392 * \param[in] dt pointer to the OSP layer dt_object
1393 * \param[in] attr the attribute for the object to be created
1394 * \param[in] hint pointer to the hint for creating the object, such as
1396 * \param[in] dof pointer to the dt_object_format for help the creation
1397 * \param[in] th pointer to the transaction handler
1399 * \retval 0 for success
1400 * \retval negative error number on failure
1402 static int osp_object_create(const struct lu_env *env, struct dt_object *dt,
1403 struct lu_attr *attr,
1404 struct dt_allocation_hint *hint,
1405 struct dt_object_format *dof, struct thandle *th)
1407 struct osp_thread_info *osi = osp_env_info(env);
1408 struct osp_device *d = lu2osp_dev(dt->do_lu.lo_dev);
1409 struct osp_object *o = dt2osp_obj(dt);
1411 struct lu_fid *fid = &osi->osi_fid;
1412 struct thandle *local_th;
1415 if (is_only_remote_trans(th) &&
1416 !fid_is_zero(lu_object_fid(&dt->do_lu))) {
1417 LASSERT(fid_is_sane(lu_object_fid(&dt->do_lu)));
1419 rc = osp_md_object_create(env, dt, attr, hint, dof, th);
1421 o->opo_non_exist = 0;
1426 o->opo_non_exist = 0;
1427 if (o->opo_reserved) {
1428 /* regular case, fid is assigned holding transaction open */
1429 osp_object_assign_fid(env, d, o);
1432 memcpy(fid, lu_object_fid(&dt->do_lu), sizeof(*fid));
1434 LASSERTF(fid_is_sane(fid), "fid for osp_object %p is insane"DFID"!\n",
1437 if (!o->opo_reserved) {
1438 /* special case, id was assigned outside of transaction
1439 * see comments in osp_declare_attr_set */
1440 LASSERT(d->opd_pre != NULL);
1441 spin_lock(&d->opd_pre_lock);
1442 osp_update_last_fid(d, fid);
1443 spin_unlock(&d->opd_pre_lock);
1446 CDEBUG(D_INODE, "fid for osp_object %p is "DFID"\n", o, PFID(fid));
1448 /* If the precreate ends, it means it will be ready to rollover to
1449 * the new sequence soon, all the creation should be synchronized,
1450 * otherwise during replay, the replay fid will be inconsistent with
1451 * last_used/create fid */
1452 if (osp_precreate_end_seq(env, d) && osp_is_fid_client(d))
1455 local_th = osp_get_storage_thandle(env, th, d);
1456 if (IS_ERR(local_th))
1457 RETURN(PTR_ERR(local_th));
1459 * it's OK if the import is inactive by this moment - id was created
1460 * by OST earlier, we just need to maintain it consistently on the disk
1461 * once import is reconnected, OSP will claim this and other objects
1462 * used and OST either keep them, if they exist or recreate
1465 /* we might have lost precreated objects */
1466 if (unlikely(d->opd_gap_count) > 0) {
1467 LASSERT(d->opd_pre != NULL);
1468 spin_lock(&d->opd_pre_lock);
1469 if (d->opd_gap_count > 0) {
1470 int count = d->opd_gap_count;
1472 ostid_set_id(&osi->osi_oi,
1473 fid_oid(&d->opd_gap_start_fid));
1474 d->opd_gap_count = 0;
1475 spin_unlock(&d->opd_pre_lock);
1477 CDEBUG(D_HA, "Writing gap "DFID"+%d in llog\n",
1478 PFID(&d->opd_gap_start_fid), count);
1479 /* real gap handling is disabled intil ORI-692 will be
1480 * fixed, now we only report gaps */
1482 spin_unlock(&d->opd_pre_lock);
1486 /* Only need update last_used oid file, seq file will only be update
1487 * during seq rollover */
1488 osp_objid_buf_prep(&osi->osi_lb, &osi->osi_off,
1489 &d->opd_last_used_fid.f_oid, d->opd_index);
1491 rc = dt_record_write(env, d->opd_last_used_oid_file, &osi->osi_lb,
1492 &osi->osi_off, local_th);
1494 CDEBUG(D_HA, "%s: Wrote last used FID: "DFID", index %d: %d\n",
1495 d->opd_obd->obd_name, PFID(fid), d->opd_index, rc);
1501 * Implement OSP layer dt_object_operations::do_declare_destroy() interface.
1503 * Declare that the caller will destroy the specified OST object.
1505 * The OST object destroy will be handled via llog asynchronously. This
1506 * function will declare the credits for generating MDS_UNLINK64_REC llog.
1508 * \param[in] env pointer to the thread context
1509 * \param[in] dt pointer to the OSP layer dt_object to be destroyed
1510 * \param[in] th pointer to the transaction handler
1512 * \retval 0 for success
1513 * \retval negative error number on failure
1515 int osp_declare_object_destroy(const struct lu_env *env,
1516 struct dt_object *dt, struct thandle *th)
1518 struct osp_object *o = dt2osp_obj(dt);
1519 struct osp_device *osp = lu2osp_dev(dt->do_lu.lo_dev);
1524 LASSERT(!osp->opd_connect_mdt);
1525 rc = osp_sync_declare_add(env, o, MDS_UNLINK64_REC, th);
1531 * Implement OSP layer dt_object_operations::do_destroy() interface.
1533 * Destroy the specified OST object.
1535 * The OSP generates a MDS_UNLINK64_REC record in the llog. There
1536 * will be some dedicated thread to handle the llog asynchronously.
1538 * It also marks the object as non-cached.
1540 * \param[in] env pointer to the thread context
1541 * \param[in] dt pointer to the OSP layer dt_object to be destroyed
1542 * \param[in] th pointer to the transaction handler
1544 * \retval 0 for success
1545 * \retval negative error number on failure
1547 static int osp_object_destroy(const struct lu_env *env, struct dt_object *dt,
1550 struct osp_object *o = dt2osp_obj(dt);
1551 struct osp_device *osp = lu2osp_dev(dt->do_lu.lo_dev);
1556 o->opo_non_exist = 1;
1558 LASSERT(!osp->opd_connect_mdt);
1559 /* once transaction is committed put proper command on
1560 * the queue going to our OST. */
1561 rc = osp_sync_add(env, o, MDS_UNLINK64_REC, th, NULL);
1565 /* not needed in cache any more */
1566 set_bit(LU_OBJECT_HEARD_BANSHEE, &dt->do_lu.lo_header->loh_flags);
1571 static int osp_orphan_index_lookup(const struct lu_env *env,
1572 struct dt_object *dt,
1574 const struct dt_key *key)
1579 static int osp_orphan_index_declare_insert(const struct lu_env *env,
1580 struct dt_object *dt,
1581 const struct dt_rec *rec,
1582 const struct dt_key *key,
1583 struct thandle *handle)
1588 static int osp_orphan_index_insert(const struct lu_env *env,
1589 struct dt_object *dt,
1590 const struct dt_rec *rec,
1591 const struct dt_key *key,
1592 struct thandle *handle,
1598 static int osp_orphan_index_declare_delete(const struct lu_env *env,
1599 struct dt_object *dt,
1600 const struct dt_key *key,
1601 struct thandle *handle)
1606 static int osp_orphan_index_delete(const struct lu_env *env,
1607 struct dt_object *dt,
1608 const struct dt_key *key,
1609 struct thandle *handle)
1615 * Initialize the OSP layer index iteration.
1617 * \param[in] env pointer to the thread context
1618 * \param[in] dt pointer to the index object to be iterated
1619 * \param[in] attr unused
1621 * \retval pointer to the iteration structure
1622 * \retval negative error number on failure
1624 struct dt_it *osp_it_init(const struct lu_env *env, struct dt_object *dt,
1631 return ERR_PTR(-ENOMEM);
1633 it->ooi_pos_ent = -1;
1635 it->ooi_attr = attr;
1637 return (struct dt_it *)it;
1641 * Finalize the OSP layer index iteration.
1643 * \param[in] env pointer to the thread context
1644 * \param[in] di pointer to the iteration structure
1646 void osp_it_fini(const struct lu_env *env, struct dt_it *di)
1648 struct osp_it *it = (struct osp_it *)di;
1649 struct page **pages = it->ooi_pages;
1650 int npages = it->ooi_total_npages;
1653 if (pages != NULL) {
1654 for (i = 0; i < npages; i++) {
1655 if (pages[i] != NULL) {
1656 if (pages[i] == it->ooi_cur_page) {
1658 it->ooi_cur_page = NULL;
1660 __free_page(pages[i]);
1663 OBD_FREE(pages, npages * sizeof(*pages));
1669 * Get more records for the iteration from peer.
1671 * The new records will be filled in an array of pages. The OSP side
1672 * allows 1MB bulk data to be transferred.
1674 * \param[in] env pointer to the thread context
1675 * \param[in] it pointer to the iteration structure
1677 * \retval 0 for success
1678 * \retval negative error number on failure
1680 static int osp_it_fetch(const struct lu_env *env, struct osp_it *it)
1682 struct lu_device *dev = it->ooi_obj->do_lu.lo_dev;
1683 struct osp_device *osp = lu2osp_dev(dev);
1684 struct page **pages;
1685 struct ptlrpc_request *req = NULL;
1686 struct ptlrpc_bulk_desc *desc;
1687 struct idx_info *ii;
1694 npages = min_t(unsigned int, OFD_MAX_BRW_SIZE, 1 << 20);
1695 npages /= PAGE_CACHE_SIZE;
1697 OBD_ALLOC(pages, npages * sizeof(*pages));
1701 it->ooi_pages = pages;
1702 it->ooi_total_npages = npages;
1703 for (i = 0; i < npages; i++) {
1704 pages[i] = alloc_page(GFP_IOFS);
1705 if (pages[i] == NULL)
1709 req = ptlrpc_request_alloc(osp->opd_obd->u.cli.cl_import,
1714 rc = ptlrpc_request_pack(req, LUSTRE_OBD_VERSION, OBD_IDX_READ);
1716 ptlrpc_request_free(req);
1720 req->rq_request_portal = OUT_PORTAL;
1721 ii = req_capsule_client_get(&req->rq_pill, &RMF_IDX_INFO);
1722 memset(ii, 0, sizeof(*ii));
1723 if (fid_is_last_id(lu_object_fid(&it->ooi_obj->do_lu))) {
1724 /* LFSCK will iterate orphan object[FID_SEQ_LAYOUT_BTREE,
1725 * ost_index, 0] with LAST_ID FID, so it needs to replace
1726 * the FID with orphan FID here */
1727 ii->ii_fid.f_seq = FID_SEQ_LAYOUT_RBTREE;
1728 ii->ii_fid.f_oid = osp->opd_index;
1729 ii->ii_fid.f_ver = 0;
1730 ii->ii_flags = II_FL_NOHASH;
1731 ii->ii_attrs = osp_dev2node(osp);
1733 ii->ii_fid = *lu_object_fid(&it->ooi_obj->do_lu);
1734 ii->ii_flags = II_FL_NOHASH | II_FL_NOKEY | II_FL_VARKEY |
1736 ii->ii_attrs = it->ooi_attr;
1738 ii->ii_magic = IDX_INFO_MAGIC;
1739 ii->ii_count = npages * LU_PAGE_COUNT;
1740 ii->ii_hash_start = it->ooi_next;
1742 ptlrpc_at_set_req_timeout(req);
1744 desc = ptlrpc_prep_bulk_imp(req, npages, 1, BULK_PUT_SINK,
1747 ptlrpc_request_free(req);
1751 for (i = 0; i < npages; i++)
1752 ptlrpc_prep_bulk_page_pin(desc, pages[i], 0, PAGE_CACHE_SIZE);
1754 ptlrpc_request_set_replen(req);
1755 rc = ptlrpc_queue_wait(req);
1759 rc = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk,
1760 req->rq_bulk->bd_nob_transferred);
1765 ii = req_capsule_server_get(&req->rq_pill, &RMF_IDX_INFO);
1766 if (ii->ii_magic != IDX_INFO_MAGIC)
1767 GOTO(out, rc = -EPROTO);
1769 npages = (ii->ii_count + LU_PAGE_COUNT - 1) >>
1770 (PAGE_CACHE_SHIFT - LU_PAGE_SHIFT);
1771 if (npages > it->ooi_total_npages) {
1772 CERROR("%s: returned more pages than expected, %u > %u\n",
1773 osp->opd_obd->obd_name, npages, it->ooi_total_npages);
1774 GOTO(out, rc = -EINVAL);
1777 it->ooi_valid_npages = npages;
1778 if (ptlrpc_rep_need_swab(req))
1781 it->ooi_next = ii->ii_hash_end;
1784 ptlrpc_req_finished(req);
1790 * Move the iteration cursor to the next lu_page.
1792 * One system page (PAGE_SIZE) may contain multiple lu_page (4KB),
1793 * that depends on the LU_PAGE_COUNT. If it is not the last lu_page
1794 * in current system page, then move the iteration cursor to the next
1795 * lu_page in current system page. Otherwise, if there are more system
1796 * pages in the cache, then move the iteration cursor to the next system
1797 * page. If all the cached records (pages) have been iterated, then fetch
1798 * more records via osp_it_fetch().
1800 * \param[in] env pointer to the thread context
1801 * \param[in] di pointer to the iteration structure
1803 * \retval positive for end of the directory
1804 * \retval 0 for success
1805 * \retval negative error number on failure
1807 int osp_it_next_page(const struct lu_env *env, struct dt_it *di)
1809 struct osp_it *it = (struct osp_it *)di;
1810 struct lu_idxpage *idxpage;
1811 struct page **pages;
1817 idxpage = it->ooi_cur_idxpage;
1818 if (idxpage != NULL) {
1819 if (idxpage->lip_nr == 0)
1822 if (it->ooi_pos_ent < idxpage->lip_nr) {
1823 CDEBUG(D_INFO, "ooi_pos %d nr %d\n",
1824 (int)it->ooi_pos_ent, (int)idxpage->lip_nr);
1827 it->ooi_cur_idxpage = NULL;
1828 it->ooi_pos_lu_page++;
1831 if (it->ooi_pos_lu_page < LU_PAGE_COUNT) {
1832 it->ooi_cur_idxpage = (void *)it->ooi_cur_page +
1833 LU_PAGE_SIZE * it->ooi_pos_lu_page;
1835 lustre_swab_lip_header(it->ooi_cur_idxpage);
1836 if (it->ooi_cur_idxpage->lip_magic != LIP_MAGIC) {
1837 struct osp_device *osp =
1838 lu2osp_dev(it->ooi_obj->do_lu.lo_dev);
1840 CERROR("%s: invalid magic (%x != %x) for page "
1841 "%d/%d while read layout orphan index\n",
1842 osp->opd_obd->obd_name,
1843 it->ooi_cur_idxpage->lip_magic,
1844 LIP_MAGIC, it->ooi_pos_page,
1845 it->ooi_pos_lu_page);
1846 /* Skip this lu_page next time. */
1847 it->ooi_pos_ent = idxpage->lip_nr - 1;
1850 it->ooi_pos_ent = -1;
1854 kunmap(it->ooi_cur_page);
1855 it->ooi_cur_page = NULL;
1859 pages = it->ooi_pages;
1860 if (it->ooi_pos_page < it->ooi_valid_npages) {
1861 it->ooi_cur_page = kmap(pages[it->ooi_pos_page]);
1862 it->ooi_pos_lu_page = 0;
1866 for (i = 0; i < it->ooi_total_npages; i++) {
1867 if (pages[i] != NULL)
1868 __free_page(pages[i]);
1870 OBD_FREE(pages, it->ooi_total_npages * sizeof(*pages));
1872 it->ooi_pos_page = 0;
1873 it->ooi_total_npages = 0;
1874 it->ooi_valid_npages = 0;
1877 it->ooi_cur_page = NULL;
1878 it->ooi_cur_idxpage = NULL;
1879 it->ooi_pages = NULL;
1882 if (it->ooi_next == II_END_OFF)
1885 rc = osp_it_fetch(env, it);
1893 * Move the iteration cursor to the next record.
1895 * If there are more records in the lu_page, then move the iteration
1896 * cursor to the next record directly. Otherwise, move the iteration
1897 * cursor to the record in the next lu_page via osp_it_next_page()
1899 * \param[in] env pointer to the thread context
1900 * \param[in] di pointer to the iteration structure
1902 * \retval positive for end of the directory
1903 * \retval 0 for success
1904 * \retval negative error number on failure
1906 static int osp_orphan_it_next(const struct lu_env *env, struct dt_it *di)
1908 struct osp_it *it = (struct osp_it *)di;
1909 struct lu_idxpage *idxpage;
1914 idxpage = it->ooi_cur_idxpage;
1915 if (idxpage != NULL) {
1916 if (idxpage->lip_nr == 0)
1920 if (it->ooi_pos_ent < idxpage->lip_nr) {
1922 (struct lu_orphan_ent *)idxpage->lip_entries +
1925 lustre_swab_orphan_ent(it->ooi_ent);
1930 rc = osp_it_next_page(env, di);
1937 int osp_it_get(const struct lu_env *env, struct dt_it *di,
1938 const struct dt_key *key)
1943 void osp_it_put(const struct lu_env *env, struct dt_it *di)
1947 static struct dt_key *osp_orphan_it_key(const struct lu_env *env,
1948 const struct dt_it *di)
1950 struct osp_it *it = (struct osp_it *)di;
1951 struct lu_orphan_ent *ent = (struct lu_orphan_ent *)it->ooi_ent;
1953 if (likely(ent != NULL))
1954 return (struct dt_key *)(&ent->loe_key);
1959 static int osp_orphan_it_key_size(const struct lu_env *env,
1960 const struct dt_it *di)
1962 return sizeof(struct lu_fid);
1965 static int osp_orphan_it_rec(const struct lu_env *env, const struct dt_it *di,
1966 struct dt_rec *rec, __u32 attr)
1968 struct osp_it *it = (struct osp_it *)di;
1969 struct lu_orphan_ent *ent = (struct lu_orphan_ent *)it->ooi_ent;
1971 if (likely(ent != NULL)) {
1972 *(struct lu_orphan_rec *)rec = ent->loe_rec;
1979 __u64 osp_it_store(const struct lu_env *env, const struct dt_it *di)
1981 struct osp_it *it = (struct osp_it *)di;
1983 return it->ooi_next;
1987 * Locate the iteration cursor to the specified position (cookie).
1989 * \param[in] env pointer to the thread context
1990 * \param[in] di pointer to the iteration structure
1991 * \param[in] hash the specified position
1993 * \retval positive number for locating to the exactly position
1995 * \retval 0 for arriving at the end of the iteration
1996 * \retval negative error number on failure
1998 int osp_orphan_it_load(const struct lu_env *env, const struct dt_it *di,
2001 struct osp_it *it = (struct osp_it *)di;
2004 it->ooi_next = hash;
2005 rc = osp_orphan_it_next(env, (struct dt_it *)di);
2015 int osp_it_key_rec(const struct lu_env *env, const struct dt_it *di,
2021 static const struct dt_index_operations osp_orphan_index_ops = {
2022 .dio_lookup = osp_orphan_index_lookup,
2023 .dio_declare_insert = osp_orphan_index_declare_insert,
2024 .dio_insert = osp_orphan_index_insert,
2025 .dio_declare_delete = osp_orphan_index_declare_delete,
2026 .dio_delete = osp_orphan_index_delete,
2028 .init = osp_it_init,
2029 .fini = osp_it_fini,
2030 .next = osp_orphan_it_next,
2033 .key = osp_orphan_it_key,
2034 .key_size = osp_orphan_it_key_size,
2035 .rec = osp_orphan_it_rec,
2036 .store = osp_it_store,
2037 .load = osp_orphan_it_load,
2038 .key_rec = osp_it_key_rec,
2043 * Implement OSP layer dt_object_operations::do_index_try() interface.
2045 * Negotiate the index type.
2047 * If the target index is an IDIF object, then use osp_orphan_index_ops.
2048 * Otherwise, assign osp_md_index_ops to the dt_object::do_index_ops.
2049 * (\see lustre/include/lustre_fid.h for IDIF.)
2051 * \param[in] env pointer to the thread context
2052 * \param[in] dt pointer to the OSP layer dt_object
2053 * \param[in] feat unused
2055 * \retval 0 for success
2057 static int osp_index_try(const struct lu_env *env,
2058 struct dt_object *dt,
2059 const struct dt_index_features *feat)
2061 const struct lu_fid *fid = lu_object_fid(&dt->do_lu);
2063 if (fid_is_last_id(fid) && fid_is_idif(fid))
2064 dt->do_index_ops = &osp_orphan_index_ops;
2066 dt->do_index_ops = &osp_md_index_ops;
2070 static struct dt_object_operations osp_obj_ops = {
2071 .do_declare_attr_get = osp_declare_attr_get,
2072 .do_attr_get = osp_attr_get,
2073 .do_declare_attr_set = osp_declare_attr_set,
2074 .do_attr_set = osp_attr_set,
2075 .do_declare_xattr_get = osp_declare_xattr_get,
2076 .do_xattr_get = osp_xattr_get,
2077 .do_declare_xattr_set = osp_declare_xattr_set,
2078 .do_xattr_set = osp_xattr_set,
2079 .do_declare_create = osp_declare_object_create,
2080 .do_create = osp_object_create,
2081 .do_declare_destroy = osp_declare_object_destroy,
2082 .do_destroy = osp_object_destroy,
2083 .do_index_try = osp_index_try,
2087 * Implement OSP layer lu_object_operations::loo_object_init() interface.
2089 * Initialize the object.
2091 * If it is a remote MDT object, then call do_attr_get() to fetch
2092 * the attribute from the peer.
2094 * \param[in] env pointer to the thread context
2095 * \param[in] o pointer to the OSP layer lu_object
2096 * \param[in] conf unused
2098 * \retval 0 for success
2099 * \retval negative error number on failure
2101 static int osp_object_init(const struct lu_env *env, struct lu_object *o,
2102 const struct lu_object_conf *conf)
2104 struct osp_object *po = lu2osp_obj(o);
2108 spin_lock_init(&po->opo_lock);
2109 o->lo_header->loh_attr |= LOHA_REMOTE;
2111 if (is_ost_obj(o)) {
2112 po->opo_obj.do_ops = &osp_obj_ops;
2114 struct lu_attr *la = &osp_env_info(env)->osi_attr;
2116 po->opo_obj.do_ops = &osp_md_obj_ops;
2117 po->opo_obj.do_body_ops = &osp_md_body_ops;
2118 rc = po->opo_obj.do_ops->do_attr_get(env, lu2dt_obj(o), la);
2120 o->lo_header->loh_attr |=
2121 LOHA_EXISTS | (la->la_mode & S_IFMT);
2122 if (rc == -ENOENT) {
2123 po->opo_non_exist = 1;
2126 init_rwsem(&po->opo_sem);
2132 * Implement OSP layer lu_object_operations::loo_object_free() interface.
2134 * Finalize the object.
2136 * If the OSP object has attributes cache, then destroy the cache.
2137 * Free the object finally.
2139 * \param[in] env pointer to the thread context
2140 * \param[in] o pointer to the OSP layer lu_object
2142 static void osp_object_free(const struct lu_env *env, struct lu_object *o)
2144 struct osp_object *obj = lu2osp_obj(o);
2145 struct lu_object_header *h = o->lo_header;
2147 dt_object_fini(&obj->opo_obj);
2148 lu_object_header_fini(h);
2149 if (obj->opo_ooa != NULL) {
2150 struct osp_xattr_entry *oxe;
2151 struct osp_xattr_entry *tmp;
2154 list_for_each_entry_safe(oxe, tmp,
2155 &obj->opo_ooa->ooa_xattr_list,
2157 list_del(&oxe->oxe_list);
2158 count = atomic_read(&oxe->oxe_ref);
2159 LASSERTF(count == 1,
2160 "Still has %d users on the xattr entry %.*s\n",
2161 count-1, (int)oxe->oxe_namelen, oxe->oxe_buf);
2163 OBD_FREE(oxe, oxe->oxe_buflen);
2165 OBD_FREE_PTR(obj->opo_ooa);
2167 OBD_SLAB_FREE_PTR(obj, osp_object_kmem);
2171 * Implement OSP layer lu_object_operations::loo_object_release() interface.
2173 * Cleanup (not free) the object.
2175 * If it is a reserved object but failed to be created, or it is an OST
2176 * object, then mark the object as non-cached.
2178 * \param[in] env pointer to the thread context
2179 * \param[in] o pointer to the OSP layer lu_object
2181 static void osp_object_release(const struct lu_env *env, struct lu_object *o)
2183 struct osp_object *po = lu2osp_obj(o);
2184 struct osp_device *d = lu2osp_dev(o->lo_dev);
2189 * release reservation if object was declared but not created
2190 * this may require lu_object_put() in LOD
2192 if (unlikely(po->opo_reserved)) {
2193 LASSERT(d->opd_pre != NULL);
2194 LASSERT(d->opd_pre_reserved > 0);
2195 spin_lock(&d->opd_pre_lock);
2196 d->opd_pre_reserved--;
2197 spin_unlock(&d->opd_pre_lock);
2199 /* not needed in cache any more */
2200 set_bit(LU_OBJECT_HEARD_BANSHEE, &o->lo_header->loh_flags);
2204 /* XXX: Currently, NOT cache OST-object on MDT because:
2205 * 1. it is not often accessed on MDT.
2206 * 2. avoid up layer (such as LFSCK) to load too many
2207 * once-used OST-objects. */
2208 set_bit(LU_OBJECT_HEARD_BANSHEE, &o->lo_header->loh_flags);
2213 static int osp_object_print(const struct lu_env *env, void *cookie,
2214 lu_printer_t p, const struct lu_object *l)
2216 const struct osp_object *o = lu2osp_obj((struct lu_object *)l);
2218 return (*p)(env, cookie, LUSTRE_OSP_NAME"-object@%p", o);
2221 static int osp_object_invariant(const struct lu_object *o)
2226 struct lu_object_operations osp_lu_obj_ops = {
2227 .loo_object_init = osp_object_init,
2228 .loo_object_free = osp_object_free,
2229 .loo_object_release = osp_object_release,
2230 .loo_object_print = osp_object_print,
2231 .loo_object_invariant = osp_object_invariant