X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fosp%2Fosp_object.c;h=bd47f5b6775b192078d0c0585d28c2e04b686f0d;hp=966d75190bba13582b9f73179db8e6fefec7b614;hb=82c6e42d6137f39a1f2394b7bc6e8d600eb36181;hpb=cd3aa902bfcf529a6774df3a207a2c3287c994fb diff --git a/lustre/osp/osp_object.c b/lustre/osp/osp_object.c index 966d751..bd47f5b 100644 --- a/lustre/osp/osp_object.c +++ b/lustre/osp/osp_object.c @@ -15,11 +15,7 @@ * * You should have received a copy of the GNU General Public License * version 2 along with this program; If not, see - * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf - * - * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, - * CA 95054 USA or visit www.sun.com if you need additional information or - * have any questions. + * http://www.gnu.org/licenses/gpl-2.0.html * * GPL HEADER END */ @@ -27,15 +23,92 @@ * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved. * Use is subject to license terms. * - * Copyright (c) 2012, 2013, Intel Corporation. + * Copyright (c) 2012, 2017, Intel Corporation. */ /* - * This file is part of Lustre, http://www.lustre.org/ - * Lustre is a trademark of Sun Microsystems, Inc. - * * lustre/osp/osp_object.c * - * Lustre OST Proxy Device + * Lustre OST Proxy Device (OSP) is the agent on the local MDT for the OST + * or remote MDT. + * + * OSP object attributes cache + * --------------------------- + * OSP object is the stub of the remote OST-object or MDT-object. Both the + * attribute and the extended attributes are stored on the peer side remotely. + * It is inefficient to send RPC to peer to fetch those attributes when every + * get_attr()/get_xattr() called. For a large system, the LFSCK synchronous + * mode scanning is prohibitively inefficient. + * + * So the OSP maintains the OSP object attributes cache to cache some + * attributes on the local MDT. The cache is organized against the OSP + * object as follows: + * + * struct osp_xattr_entry { + * struct list_head oxe_list; + * atomic_t oxe_ref; + * void *oxe_value; + * int oxe_buflen; + * int oxe_namelen; + * int oxe_vallen; + * unsigned int oxe_exist:1, + * oxe_ready:1; + * char oxe_buf[0]; + * }; + * + * struct osp_object { + * ... + * struct lu_attr opo_attr; + * struct list_head opo_xattr_list; + * spinlock_t opo_lock; + * ... + * }; + * + * The basic attributes, such as owner/mode/flags, are stored in the + * osp_object::opo_attr. The extended attributes will be stored + * as osp_xattr_entry. Every extended attribute has an independent + * osp_xattr_entry, and all the osp_xattr_entry are linked into the + * osp_object::opo_xattr_list. The OSP object attributes cache + * is protected by the osp_object::opo_lock. + * + * Not all OSP objects have an attributes cache because maintaining + * the cache requires some resources. Currently, the OSP object + * attributes cache will be initialized when the attributes or the + * extended attributes are pre-fetched via osp_declare_attr_get() + * or osp_declare_xattr_get(). That is usually for LFSCK purpose, + * but it also can be shared by others. + * + * + * XXX: NOT prepare out RPC for remote transaction. ((please refer to the + * comment of osp_trans_create() for remote transaction) + * + * According to our current transaction/dt_object_lock framework (to make + * the cross-MDTs modification for DNE1 to be workable), the transaction + * sponsor will start the transaction firstly, then try to acquire related + * dt_object_lock if needed. Under such rules, if we want to prepare the + * OUT RPC in the transaction declare phase, then related attr/xattr + * should be known without dt_object_lock. But such condition maybe not + * true for some remote transaction case. For example: + * + * For linkEA repairing (by LFSCK) case, before the LFSCK thread obtained + * the dt_object_lock on the target MDT-object, it cannot know whether + * the MDT-object has linkEA or not, neither invalid or not. + * + * Since the LFSCK thread cannot hold dt_object_lock before the remote + * transaction start (otherwise there will be some potential deadlock), + * it cannot prepare related OUT RPC for repairing during the declare + * phase as other normal transactions do. + * + * To resolve the trouble, we will make OSP to prepare related OUT RPC + * after remote transaction started, and trigger the remote updating + * (send RPC) when trans_stop. Then the up layer users, such as LFSCK, + * can follow the general rule to handle trans_start/dt_object_lock + * for repairing linkEA inconsistency without distinguishing remote + * MDT-object. + * + * In fact, above solution for remote transaction should be the normal + * model without considering DNE1. The trouble brought by DNE1 will be + * resolved in DNE2. At that time, this patch can be removed. + * * * Author: Alex Zhuravlev * Author: Mikhail Pershin @@ -43,15 +116,50 @@ #define DEBUG_SUBSYSTEM S_MDS +#include +#include + #include "osp_internal.h" +static inline __u32 osp_dev2node(struct osp_device *osp) +{ + return osp->opd_storage->dd_lu_dev.ld_site->ld_seq_site->ss_node_id; +} + +static inline const char *osp_dto2name(struct osp_object *obj) +{ + return obj->opo_obj.do_lu.lo_dev->ld_obd->obd_name; +} + static inline bool is_ost_obj(struct lu_object *lo) { return !lu2osp_dev(lo->lo_dev)->opd_connect_mdt; } +static inline void __osp_oac_xattr_assignment(struct osp_object *obj, + struct osp_xattr_entry *oxe, + const struct lu_buf *buf) +{ + if (buf->lb_len > 0) + memcpy(oxe->oxe_value, buf->lb_buf, buf->lb_len); + + oxe->oxe_vallen = buf->lb_len; + oxe->oxe_exist = 1; + oxe->oxe_ready = 1; +} + +/** + * Assign FID to the OST object. + * + * This function will assign the FID to the OST object of a striped file. + * + * \param[in] env pointer to the thread context + * \param[in] d pointer to the OSP device + * \param[in] o pointer to the OSP object that the FID will be + * assigned to + */ static void osp_object_assign_fid(const struct lu_env *env, - struct osp_device *d, struct osp_object *o) + struct osp_device *d, struct osp_object *o) { struct osp_thread_info *osi = osp_env_info(env); @@ -64,79 +172,115 @@ static void osp_object_assign_fid(const struct lu_env *env, lu_object_assign_fid(env, &o->opo_obj.do_lu, &osi->osi_fid); } -static int osp_oac_init(struct osp_object *obj) -{ - struct osp_object_attr *ooa; +#define OXE_DEFAULT_LEN 16 - OBD_ALLOC_PTR(ooa); - if (ooa == NULL) - return -ENOMEM; +/** + * Release reference from the OSP object extended attribute entry. + * + * If it is the last reference, then free the entry. + * + * \param[in] oxe pointer to the OSP object extended attribute entry. + */ +static inline void osp_oac_xattr_put(struct osp_xattr_entry *oxe) +{ + if (atomic_dec_and_test(&oxe->oxe_ref)) { + LASSERT(list_empty(&oxe->oxe_list)); - INIT_LIST_HEAD(&ooa->ooa_xattr_list); - spin_lock(&obj->opo_lock); - if (likely(obj->opo_ooa == NULL)) { - obj->opo_ooa = ooa; - spin_unlock(&obj->opo_lock); - } else { - spin_unlock(&obj->opo_lock); - OBD_FREE_PTR(ooa); + OBD_FREE_LARGE(oxe, oxe->oxe_buflen); } - - return 0; } +/** + * Find the named extended attribute in the OSP object attributes cache. + * + * The caller should take the osp_object::opo_lock before calling + * this function. + * + * \param[in] obj pointer to the OSP object + * \param[in] name the name of the extended attribute + * \param[in] namelen the name length of the extended attribute + * + * \retval pointer to the found extended attribute entry + * \retval NULL if the specified extended attribute is not + * in the cache + */ static struct osp_xattr_entry * -osp_oac_xattr_find_locked(struct osp_object_attr *ooa, - const char *name, size_t namelen, bool unlink) +osp_oac_xattr_find_locked(struct osp_object *obj, const char *name, + size_t namelen) { struct osp_xattr_entry *oxe; - list_for_each_entry(oxe, &ooa->ooa_xattr_list, oxe_list) { + list_for_each_entry(oxe, &obj->opo_xattr_list, oxe_list) { if (namelen == oxe->oxe_namelen && - strncmp(name, oxe->oxe_buf, namelen) == 0) { - if (unlink) - list_del_init(&oxe->oxe_list); - else - atomic_inc(&oxe->oxe_ref); - + strncmp(name, oxe->oxe_buf, namelen) == 0) return oxe; - } } return NULL; } +/** + * Find the named extended attribute in the OSP object attributes cache. + * + * Call osp_oac_xattr_find_locked() with the osp_object::opo_lock held. + * + * \param[in] obj pointer to the OSP object + * \param[in] name the name of the extended attribute + * \param[in] unlink true if the extended attribute entry is to be removed + * from the cache + * + * \retval pointer to the found extended attribute entry + * \retval NULL if the specified extended attribute is not + * in the cache + */ static struct osp_xattr_entry *osp_oac_xattr_find(struct osp_object *obj, - const char *name) + const char *name, bool unlink) { struct osp_xattr_entry *oxe = NULL; spin_lock(&obj->opo_lock); - if (obj->opo_ooa != NULL) - oxe = osp_oac_xattr_find_locked(obj->opo_ooa, name, - strlen(name), false); + oxe = osp_oac_xattr_find_locked(obj, name, strlen(name)); + if (oxe) { + if (unlink) + list_del_init(&oxe->oxe_list); + else + atomic_inc(&oxe->oxe_ref); + } spin_unlock(&obj->opo_lock); return oxe; } +/** + * Find the named extended attribute in the OSP object attributes cache. + * + * If it is not in the cache, then add an empty entry (that will be + * filled later) to cache with the given name. + * + * \param[in] obj pointer to the OSP object + * \param[in] name the name of the extended attribute + * \param[in] len the length of the extended attribute value + * + * \retval pointer to the found or new-created extended + * attribute entry + * \retval NULL if the specified extended attribute is not in the + * cache or fail to add new empty entry to the cache. + */ static struct osp_xattr_entry * osp_oac_xattr_find_or_add(struct osp_object *obj, const char *name, size_t len) { - struct osp_object_attr *ooa = obj->opo_ooa; struct osp_xattr_entry *oxe; - struct osp_xattr_entry *tmp = NULL; - size_t namelen = strlen(name); - size_t size = sizeof(*oxe) + namelen + 1 + len; + struct osp_xattr_entry *tmp = NULL; + size_t namelen = strlen(name); + size_t size = sizeof(*oxe) + namelen + 1 + + (len ? len : OXE_DEFAULT_LEN); - LASSERT(ooa != NULL); - - oxe = osp_oac_xattr_find(obj, name); - if (oxe != NULL) + oxe = osp_oac_xattr_find(obj, name, false); + if (oxe) return oxe; - OBD_ALLOC(oxe, size); - if (unlikely(oxe == NULL)) + OBD_ALLOC_LARGE(oxe, size); + if (unlikely(!oxe)) return NULL; INIT_LIST_HEAD(&oxe->oxe_list); @@ -148,64 +292,120 @@ osp_oac_xattr_find_or_add(struct osp_object *obj, const char *name, size_t len) atomic_set(&oxe->oxe_ref, 2); spin_lock(&obj->opo_lock); - tmp = osp_oac_xattr_find_locked(ooa, name, namelen, false); - if (tmp == NULL) - list_add_tail(&oxe->oxe_list, &ooa->ooa_xattr_list); + tmp = osp_oac_xattr_find_locked(obj, name, namelen); + if (!tmp) + list_add_tail(&oxe->oxe_list, &obj->opo_xattr_list); + else + atomic_inc(&tmp->oxe_ref); spin_unlock(&obj->opo_lock); - if (tmp != NULL) { - OBD_FREE(oxe, size); + if (tmp) { + OBD_FREE_LARGE(oxe, size); oxe = tmp; } return oxe; } +/** + * Assign the cached OST-object's EA with the given value. + * + * If the current EA entry in cache has not enough space to hold the new + * value, remove it, create a new one, then assign with the given value. + * + * \param[in] obj pointer to the OSP object + * \param[in] oxe pointer to the cached EA entry to be assigned + * \param[in] buf pointer to the buffer with new EA value + * + * \retval pointer to the new created EA entry in cache if + * current entry is not big enough; otherwise, the + * input 'oxe' will be returned. + */ static struct osp_xattr_entry * -osp_oac_xattr_replace(struct osp_object *obj, - struct osp_xattr_entry **poxe, size_t len) +osp_oac_xattr_assignment(struct osp_object *obj, struct osp_xattr_entry *oxe, + const struct lu_buf *buf) { - struct osp_object_attr *ooa = obj->opo_ooa; - struct osp_xattr_entry *old = *poxe; - struct osp_xattr_entry *oxe; - struct osp_xattr_entry *tmp = NULL; - size_t namelen = old->oxe_namelen; - size_t size = sizeof(*oxe) + namelen + 1 + len; - - LASSERT(ooa != NULL); - - OBD_ALLOC(oxe, size); - if (unlikely(oxe == NULL)) - return NULL; - - INIT_LIST_HEAD(&oxe->oxe_list); - oxe->oxe_buflen = size; - oxe->oxe_namelen = namelen; - memcpy(oxe->oxe_buf, old->oxe_buf, namelen); - oxe->oxe_value = oxe->oxe_buf + namelen + 1; - /* One ref is for the caller, the other is for the entry on the list. */ - atomic_set(&oxe->oxe_ref, 2); + struct osp_xattr_entry *new = NULL; + struct osp_xattr_entry *old = NULL; + int namelen = oxe->oxe_namelen; + size_t size = sizeof(*oxe) + namelen + 1 + buf->lb_len; + bool unlink_only = false; + + if (oxe->oxe_buflen < size) { + OBD_ALLOC_LARGE(new, size); + if (likely(new)) { + INIT_LIST_HEAD(&new->oxe_list); + new->oxe_buflen = size; + new->oxe_namelen = namelen; + memcpy(new->oxe_buf, oxe->oxe_buf, namelen); + new->oxe_value = new->oxe_buf + namelen + 1; + /* One ref is for the caller, + * the other is for the entry on the list. */ + atomic_set(&new->oxe_ref, 2); + __osp_oac_xattr_assignment(obj, new, buf); + } else { + unlink_only = true; + CWARN("%s: cannot update cached xattr %.*s of "DFID"\n", + osp_dto2name(obj), namelen, oxe->oxe_buf, + PFID(lu_object_fid(&obj->opo_obj.do_lu))); + } + } spin_lock(&obj->opo_lock); - tmp = osp_oac_xattr_find_locked(ooa, oxe->oxe_buf, namelen, true); - list_add_tail(&oxe->oxe_list, &ooa->ooa_xattr_list); - spin_unlock(&obj->opo_lock); + old = osp_oac_xattr_find_locked(obj, oxe->oxe_buf, namelen); + if (likely(old)) { + if (new) { + /* Unlink the 'old'. */ + list_del_init(&old->oxe_list); - *poxe = tmp; - LASSERT(tmp != NULL); + /* Drop the ref for 'old' on list. */ + osp_oac_xattr_put(old); - return oxe; -} + /* Drop the ref for current using. */ + osp_oac_xattr_put(oxe); + oxe = new; -static inline void osp_oac_xattr_put(struct osp_xattr_entry *oxe) -{ - if (atomic_dec_and_test(&oxe->oxe_ref)) { - LASSERT(list_empty(&oxe->oxe_list)); + /* Insert 'new' into list. */ + list_add_tail(&new->oxe_list, &obj->opo_xattr_list); + } else if (unlink_only) { + /* Unlink the 'old'. */ + list_del_init(&old->oxe_list); + + /* Drop the ref for 'old' on list. */ + osp_oac_xattr_put(old); + } else { + __osp_oac_xattr_assignment(obj, oxe, buf); + } + } else if (new) { + /* Drop the ref for current using. */ + osp_oac_xattr_put(oxe); + oxe = new; - OBD_FREE(oxe, oxe->oxe_buflen); + /* Someone unlinked the 'old' by race, + * insert the 'new' one into list. */ + list_add_tail(&new->oxe_list, &obj->opo_xattr_list); } + spin_unlock(&obj->opo_lock); + + return oxe; } +/** + * Parse the OSP object attribute from the RPC reply. + * + * If the attribute is valid, then it will be added to the OSP object + * attributes cache. + * + * \param[in] env pointer to the thread context + * \param[in] reply pointer to the RPC reply + * \param[in] req pointer to the RPC request + * \param[out] attr pointer to buffer to hold the output attribute + * \param[in] obj pointer to the OSP object + * \param[in] index the index of the attribute buffer in the reply + * + * \retval 0 for success + * \retval negative error number on failure + */ static int osp_get_attr_from_reply(const struct lu_env *env, struct object_update_reply *reply, struct ptlrpc_request *req, @@ -231,21 +431,34 @@ static int osp_get_attr_from_reply(const struct lu_env *env, lustre_swab_obdo(wobdo); lustre_get_wire_obdo(NULL, lobdo, wobdo); - spin_lock(&obj->opo_lock); - if (obj->opo_ooa != NULL) { - la_from_obdo(&obj->opo_ooa->ooa_attr, lobdo, lobdo->o_valid); - if (attr != NULL) - *attr = obj->opo_ooa->ooa_attr; - } else { - LASSERT(attr != NULL); - - la_from_obdo(attr, lobdo, lobdo->o_valid); + if (obj) { + spin_lock(&obj->opo_lock); + la_from_obdo(&obj->opo_attr, lobdo, lobdo->o_valid); + spin_unlock(&obj->opo_lock); } - spin_unlock(&obj->opo_lock); + if (attr) + la_from_obdo(attr, lobdo, lobdo->o_valid); return 0; } +/** + * Interpreter function for getting OSP object attribute asynchronously. + * + * Called to interpret the result of an async mode RPC for getting the + * OSP object attribute. + * + * \param[in] env pointer to the thread context + * \param[in] reply pointer to the RPC reply + * \param[in] req pointer to the RPC request + * \param[in] obj pointer to the OSP object + * \param[out] data pointer to buffer to hold the output attribute + * \param[in] index the index of the attribute buffer in the reply + * \param[in] rc the result for handling the RPC + * + * \retval 0 for success + * \retval negative error number on failure + */ static int osp_attr_get_interpterer(const struct lu_env *env, struct object_update_reply *reply, struct ptlrpc_request *req, @@ -254,8 +467,6 @@ static int osp_attr_get_interpterer(const struct lu_env *env, { struct lu_attr *attr = data; - LASSERT(obj->opo_ooa != NULL); - if (rc == 0) { osp2lu_obj(obj)->lo_header->loh_attr |= LOHA_EXISTS; obj->opo_non_exist = 0; @@ -276,73 +487,107 @@ static int osp_attr_get_interpterer(const struct lu_env *env, return 0; } -static int osp_declare_attr_get(const struct lu_env *env, struct dt_object *dt, - struct lustre_capa *capa) +/** + * Implement OSP layer dt_object_operations::do_declare_attr_get() interface. + * + * Declare that the caller will get attribute from the specified OST object. + * + * This function adds an Object Unified Target (OUT) sub-request to the per-OSP + * based shared asynchronous request queue. The osp_attr_get_interpterer() + * is registered as the interpreter function to handle the result of this + * sub-request. + * + * \param[in] env pointer to the thread context + * \param[in] dt pointer to the OSP layer dt_object + * + * \retval 0 for success + * \retval negative error number on failure + */ +static int osp_declare_attr_get(const struct lu_env *env, struct dt_object *dt) { struct osp_object *obj = dt2osp_obj(dt); struct osp_device *osp = lu2osp_dev(dt->do_lu.lo_dev); int rc = 0; - if (obj->opo_ooa == NULL) { - rc = osp_oac_init(obj); - if (rc != 0) - return rc; - } - mutex_lock(&osp->opd_async_requests_mutex); rc = osp_insert_async_request(env, OUT_ATTR_GET, obj, 0, NULL, NULL, - &obj->opo_ooa->ooa_attr, + &obj->opo_attr, sizeof(struct obdo), osp_attr_get_interpterer); mutex_unlock(&osp->opd_async_requests_mutex); return rc; } +/** + * Implement OSP layer dt_object_operations::do_attr_get() interface. + * + * Get attribute from the specified MDT/OST object. + * + * If the attribute is in the OSP object attributes cache, then return + * the cached attribute directly. Otherwise it will trigger an OUT RPC + * to the peer to get the attribute synchronously, if successful, add it + * to the OSP attributes cache. (\see lustre/osp/osp_trans.c for OUT RPC.) + * + * \param[in] env pointer to the thread context + * \param[in] dt pointer to the OSP layer dt_object + * \param[out] attr pointer to the buffer to hold the output attribute + * + * \retval 0 for success + * \retval negative error number on failure + */ int osp_attr_get(const struct lu_env *env, struct dt_object *dt, - struct lu_attr *attr, struct lustre_capa *capa) + struct lu_attr *attr) { struct osp_device *osp = lu2osp_dev(dt->do_lu.lo_dev); struct osp_object *obj = dt2osp_obj(dt); struct dt_device *dev = &osp->opd_dt_dev; - struct dt_update_request *update; + struct osp_update_request *update; struct object_update_reply *reply; struct ptlrpc_request *req = NULL; - int rc = 0; + int invalidated, cache = 0, rc = 0; ENTRY; if (is_ost_obj(&dt->do_lu) && obj->opo_non_exist) RETURN(-ENOENT); - if (obj->opo_ooa != NULL) { - spin_lock(&obj->opo_lock); - if (obj->opo_ooa->ooa_attr.la_valid != 0) { - *attr = obj->opo_ooa->ooa_attr; - spin_unlock(&obj->opo_lock); - - RETURN(0); - } + spin_lock(&obj->opo_lock); + if (obj->opo_attr.la_valid != 0 && !obj->opo_stale) { + *attr = obj->opo_attr; spin_unlock(&obj->opo_lock); + + RETURN(0); } + spin_unlock(&obj->opo_lock); - update = out_create_update_req(dev); + update = osp_update_request_create(dev); if (IS_ERR(update)) RETURN(PTR_ERR(update)); - rc = out_insert_update(env, update, OUT_ATTR_GET, - lu_object_fid(&dt->do_lu), 0, NULL, NULL); + rc = OSP_UPDATE_RPC_PACK(env, out_attr_get_pack, update, + lu_object_fid(&dt->do_lu)); if (rc != 0) { CERROR("%s: Insert update error "DFID": rc = %d\n", dev->dd_lu_dev.ld_obd->obd_name, PFID(lu_object_fid(&dt->do_lu)), rc); - GOTO(out, rc); + GOTO(out_req, rc); + } + + invalidated = atomic_read(&obj->opo_invalidate_seq); + + rc = osp_remote_sync(env, osp, update, &req); + + down_read(&obj->opo_invalidate_sem); + if (invalidated == atomic_read(&obj->opo_invalidate_seq)) { + /* no invalited has came so far, we can cache the attrs */ + cache = 1; } - rc = out_remote_sync(env, osp->opd_obd->u.cli.cl_import, update, &req); if (rc != 0) { if (rc == -ENOENT) { osp2lu_obj(obj)->lo_header->loh_attr &= ~LOHA_EXISTS; - obj->opo_non_exist = 1; + if (cache) + obj->opo_non_exist = 1; } else { CERROR("%s:osp_attr_get update error "DFID": rc = %d\n", dev->dd_lu_dev.ld_obd->obd_name, @@ -360,35 +605,57 @@ int osp_attr_get(const struct lu_env *env, struct dt_object *dt, if (reply == NULL || reply->ourp_magic != UPDATE_REPLY_MAGIC) GOTO(out, rc = -EPROTO); - rc = osp_get_attr_from_reply(env, reply, req, attr, obj, 0); + rc = osp_get_attr_from_reply(env, reply, req, attr, + cache ? obj : NULL, 0); if (rc != 0) GOTO(out, rc); - GOTO(out, rc = 0); + spin_lock(&obj->opo_lock); + if (cache) + obj->opo_stale = 0; + spin_unlock(&obj->opo_lock); + + GOTO(out, rc); out: + up_read(&obj->opo_invalidate_sem); + +out_req: if (req != NULL) ptlrpc_req_finished(req); - out_destroy_update_req(update); + osp_update_request_destroy(env, update); return rc; } +/** + * Implement OSP layer dt_object_operations::do_declare_attr_set() interface. + * + * If the transaction is not remote one, then declare the credits that will + * be used for the subsequent llog record for the object's attributes. + * + * \param[in] env pointer to the thread context + * \param[in] dt pointer to the OSP layer dt_object + * \param[in] attr pointer to the attribute to be set + * \param[in] th pointer to the transaction handler + * + * \retval 0 for success + * \retval negative error number on failure + */ static int osp_declare_attr_set(const struct lu_env *env, struct dt_object *dt, const struct lu_attr *attr, struct thandle *th) { struct osp_device *d = lu2osp_dev(dt->do_lu.lo_dev); struct osp_object *o = dt2osp_obj(dt); - struct lu_attr *la; - int rc = 0; - - ENTRY; + int rc; + if (is_only_remote_trans(th)) + return osp_md_declare_attr_set(env, dt, attr, th); /* * Usually we don't allow server stack to manipulate size * but there is a special case when striping is created - * late, after stripless file got truncated to non-zero. + * late, after stripeless file got truncated to non-zero. * * In this case we do the following: * @@ -413,179 +680,238 @@ static int osp_declare_attr_set(const struct lu_env *env, struct dt_object *dt, LASSERT(!dt_object_exists(dt)); osp_object_assign_fid(env, d, o); rc = osp_object_truncate(env, dt, attr->la_size); - if (rc) + if (rc != 0) RETURN(rc); } - if (o->opo_new) { - /* no need in logging for new objects being created */ + if (!(attr->la_valid & LA_REMOTE_ATTR_SET)) RETURN(0); - } - if (!(attr->la_valid & (LA_UID | LA_GID))) - RETURN(0); - - if (!is_only_remote_trans(th)) - /* - * track all UID/GID changes via llog - */ - rc = osp_sync_declare_add(env, o, MDS_SETATTR64_REC, th); - else - /* It is for OST-object attr_set directly without updating - * local MDT-object attribute. It is usually used by LFSCK. */ - rc = osp_md_declare_attr_set(env, dt, attr, th); + /* track all UID/GID, projid, and layout version changes via llog */ + rc = osp_sync_declare_add(env, o, MDS_SETATTR64_REC, th); - if (rc != 0 || o->opo_ooa == NULL) - RETURN(rc); - - la = &o->opo_ooa->ooa_attr; - spin_lock(&o->opo_lock); - if (attr->la_valid & LA_UID) { - la->la_uid = attr->la_uid; - la->la_valid |= LA_UID; - } - - if (attr->la_valid & LA_GID) { - la->la_gid = attr->la_gid; - la->la_valid |= LA_GID; - } - spin_unlock(&o->opo_lock); - - RETURN(0); + return 0; } +/** + * Implement OSP layer dt_object_operations::do_attr_set() interface. + * + * Set attribute to the specified OST object. + * + * If the transaction is a remote one, then add OUT_ATTR_SET sub-request + * in the OUT RPC that will be flushed when the remote transaction stop. + * Otherwise, it will generate a MDS_SETATTR64_REC record in the llog that + * will be handled by a dedicated thread asynchronously. + * + * If the attribute entry exists in the OSP object attributes cache, + * then update the cached attribute according to given attribute. + * + * \param[in] env pointer to the thread context + * \param[in] dt pointer to the OSP layer dt_object + * \param[in] attr pointer to the attribute to be set + * \param[in] th pointer to the transaction handler + * + * \retval 0 for success + * \retval negative error number on failure + */ static int osp_attr_set(const struct lu_env *env, struct dt_object *dt, - const struct lu_attr *attr, struct thandle *th, - struct lustre_capa *capa) + const struct lu_attr *attr, struct thandle *th) { struct osp_object *o = dt2osp_obj(dt); int rc = 0; - ENTRY; - /* we're interested in uid/gid changes only */ - if (!(attr->la_valid & (LA_UID | LA_GID))) + /* we're interested in uid/gid/projid/layout version changes only */ + if (!(attr->la_valid & LA_REMOTE_ATTR_SET)) RETURN(0); - /* new object, the very first ->attr_set() - * initializing attributes needs no logging - * all subsequent one are subject to the - * logging and synchronization with OST */ - if (o->opo_new) { - o->opo_new = 0; - RETURN(0); - } + if (!is_only_remote_trans(th)) { + if (attr->la_flags & LUSTRE_SET_SYNC_FL) { + struct ptlrpc_request *req = NULL; + struct osp_update_request *update = NULL; + struct osp_device *osp = lu2osp_dev(dt->do_lu.lo_dev); + + update = osp_update_request_create(&osp->opd_dt_dev); + if (IS_ERR(update)) + RETURN(PTR_ERR(update)); + + rc = OSP_UPDATE_RPC_PACK(env, out_attr_set_pack, update, + lu_object_fid(&dt->do_lu), + attr); + if (rc != 0) { + CERROR("%s: update error "DFID": rc = %d\n", + osp->opd_obd->obd_name, + PFID(lu_object_fid(&dt->do_lu)), rc); + + osp_update_request_destroy(env, update); + RETURN(rc); + } + + rc = osp_remote_sync(env, osp, update, &req); + if (req != NULL) + ptlrpc_req_finished(req); + + osp_update_request_destroy(env, update); + } else { + rc = osp_sync_add(env, o, MDS_SETATTR64_REC, th, attr); + /* XXX: send new uid/gid to OST ASAP? */ + } + } else { + struct lu_attr *la; - if (!is_only_remote_trans(th)) - /* - * once transaction is committed put proper command on - * the queue going to our OST - */ - rc = osp_sync_add(env, o, MDS_SETATTR64_REC, th, attr); - /* XXX: send new uid/gid to OST ASAP? */ - else /* It is for OST-object attr_set directly without updating * local MDT-object attribute. It is usually used by LFSCK. */ - rc = osp_md_attr_set(env, dt, attr, th, capa); + rc = osp_md_attr_set(env, dt, attr, th); + CDEBUG(D_INFO, "(1) set attr "DFID": rc = %d\n", + PFID(&dt->do_lu.lo_header->loh_fid), rc); + + if (rc != 0) + RETURN(rc); + + /* Update the OSP object attributes cache. */ + la = &o->opo_attr; + spin_lock(&o->opo_lock); + if (attr->la_valid & LA_UID) { + la->la_uid = attr->la_uid; + la->la_valid |= LA_UID; + } + + if (attr->la_valid & LA_GID) { + la->la_gid = attr->la_gid; + la->la_valid |= LA_GID; + } + if (attr->la_valid & LA_PROJID) { + la->la_projid = attr->la_projid; + la->la_valid |= LA_PROJID; + } + spin_unlock(&o->opo_lock); + } RETURN(rc); } +/** + * Interpreter function for getting OSP object extended attribute asynchronously + * + * Called to interpret the result of an async mode RPC for getting the + * OSP object extended attribute. + * + * \param[in] env pointer to the thread context + * \param[in] reply pointer to the RPC reply + * \param[in] req pointer to the RPC request + * \param[in] obj pointer to the OSP object + * \param[out] data pointer to OSP object attributes cache + * \param[in] index the index of the attribute buffer in the reply + * \param[in] rc the result for handling the RPC + * + * \retval 0 for success + * \retval negative error number on failure + */ static int osp_xattr_get_interpterer(const struct lu_env *env, struct object_update_reply *reply, struct ptlrpc_request *req, struct osp_object *obj, void *data, int index, int rc) { - struct osp_object_attr *ooa = obj->opo_ooa; - struct osp_xattr_entry *oxe = data; - struct lu_buf *rbuf = &osp_env_info(env)->osi_lb2; - - LASSERT(ooa != NULL); + struct osp_xattr_entry *oxe = data; - if (rc == 0) { + spin_lock(&obj->opo_lock); + if (rc >= 0) { + struct lu_buf *rbuf = &osp_env_info(env)->osi_lb2; size_t len = sizeof(*oxe) + oxe->oxe_namelen + 1; rc = object_update_result_data_get(reply, rbuf, index); - if (rc < 0 || rbuf->lb_len > (oxe->oxe_buflen - len)) { - spin_lock(&obj->opo_lock); - oxe->oxe_ready = 0; - spin_unlock(&obj->opo_lock); - osp_oac_xattr_put(oxe); + if (rc == -ENOENT || rc == -ENODATA || rc == 0) { + oxe->oxe_exist = 0; + oxe->oxe_ready = 1; + goto unlock; + } - return rc < 0 ? rc : -ERANGE; + if (unlikely(rc < 0) || + rbuf->lb_len > (oxe->oxe_buflen - len)) { + oxe->oxe_ready = 0; + goto unlock; } - spin_lock(&obj->opo_lock); - oxe->oxe_vallen = rbuf->lb_len; - memcpy(oxe->oxe_value, rbuf->lb_buf, rbuf->lb_len); - oxe->oxe_exist = 1; - oxe->oxe_ready = 1; - spin_unlock(&obj->opo_lock); + __osp_oac_xattr_assignment(obj, oxe, rbuf); } else if (rc == -ENOENT || rc == -ENODATA) { - spin_lock(&obj->opo_lock); oxe->oxe_exist = 0; oxe->oxe_ready = 1; - spin_unlock(&obj->opo_lock); } else { - spin_lock(&obj->opo_lock); oxe->oxe_ready = 0; - spin_unlock(&obj->opo_lock); } +unlock: + spin_unlock(&obj->opo_lock); + + /* Put the reference obtained in the osp_declare_xattr_get(). */ osp_oac_xattr_put(oxe); return 0; } +/** + * Implement OSP dt_object_operations::do_declare_xattr_get() interface. + * + * Declare that the caller will get extended attribute from the specified + * OST object. + * + * This function will add an OUT_XATTR_GET sub-request to the per OSP + * based shared asynchronous request queue with the interpreter function: + * osp_xattr_get_interpterer(). + * + * \param[in] env pointer to the thread context + * \param[in] dt pointer to the OSP layer dt_object + * \param[out] buf pointer to the lu_buf to hold the extended attribute + * \param[in] name the name for the expected extended attribute + * + * \retval 0 for success + * \retval negative error number on failure + */ static int osp_declare_xattr_get(const struct lu_env *env, struct dt_object *dt, - struct lu_buf *buf, const char *name, - struct lustre_capa *capa) + struct lu_buf *buf, const char *name) { struct osp_object *obj = dt2osp_obj(dt); struct osp_device *osp = lu2osp_dev(dt->do_lu.lo_dev); struct osp_xattr_entry *oxe; - int namelen = strlen(name); int rc = 0; + __u16 len; LASSERT(buf != NULL); LASSERT(name != NULL); - /* If only for xattr size, return directly. */ if (unlikely(buf->lb_len == 0)) - return 0; - - if (obj->opo_ooa == NULL) { - rc = osp_oac_init(obj); - if (rc != 0) - return rc; - } + return -EINVAL; oxe = osp_oac_xattr_find_or_add(obj, name, buf->lb_len); if (oxe == NULL) return -ENOMEM; + len = strlen(name) + 1; mutex_lock(&osp->opd_async_requests_mutex); rc = osp_insert_async_request(env, OUT_XATTR_GET, obj, 1, - &namelen, &name, oxe, + &len, (const void **)&name, + oxe, buf->lb_len, osp_xattr_get_interpterer); if (rc != 0) { mutex_unlock(&osp->opd_async_requests_mutex); osp_oac_xattr_put(oxe); } else { - struct dt_update_request *update; + struct osp_update_request *our; + struct osp_update_request_sub *ours; /* XXX: Currently, we trigger the batched async OUT * RPC via dt_declare_xattr_get(). It is not * perfect solution, but works well now. * * We will improve it in the future. */ - update = osp->opd_async_requests; - if (update != NULL && update->dur_req != NULL && - update->dur_req->ourq_count > 0) { + our = osp->opd_async_requests; + ours = osp_current_object_update_request(our); + if (ours != NULL && ours->ours_req != NULL && + ours->ours_req->ourq_count > 0) { osp->opd_async_requests = NULL; mutex_unlock(&osp->opd_async_requests_mutex); - rc = osp_unplug_async_request(env, osp, update); + rc = osp_unplug_async_request(env, osp, our); } else { mutex_unlock(&osp->opd_async_requests_mutex); } @@ -594,30 +920,64 @@ static int osp_declare_xattr_get(const struct lu_env *env, struct dt_object *dt, return rc; } +/** + * Implement OSP layer dt_object_operations::do_xattr_get() interface. + * + * Get extended attribute from the specified MDT/OST object. + * + * If the extended attribute is in the OSP object attributes cache, then + * return the cached extended attribute directly. Otherwise it will get + * the extended attribute synchronously, if successful, add it to the OSP + * attributes cache. (\see lustre/osp/osp_trans.c for OUT RPC.) + * + * There is a race condition: some other thread has added the named extended + * attributed entry to the OSP object attributes cache during the current + * OUT_XATTR_GET handling. If such case happens, the OSP will replace the + * (just) existing extended attribute entry with the new replied one. + * + * \param[in] env pointer to the thread context + * \param[in] dt pointer to the OSP layer dt_object + * \param[out] buf pointer to the lu_buf to hold the extended attribute + * \param[in] name the name for the expected extended attribute + * + * \retval 0 for success + * \retval negative error number on failure + */ int osp_xattr_get(const struct lu_env *env, struct dt_object *dt, - struct lu_buf *buf, const char *name, - struct lustre_capa *capa) + struct lu_buf *buf, const char *name) { struct osp_device *osp = lu2osp_dev(dt->do_lu.lo_dev); struct osp_object *obj = dt2osp_obj(dt); struct dt_device *dev = &osp->opd_dt_dev; struct lu_buf *rbuf = &osp_env_info(env)->osi_lb2; - struct dt_update_request *update = NULL; + struct osp_update_request *update = NULL; struct ptlrpc_request *req = NULL; struct object_update_reply *reply; struct osp_xattr_entry *oxe = NULL; - const char *dname = dt->do_lu.lo_dev->ld_obd->obd_name; - int namelen; - int rc = 0; + const char *dname = osp_dto2name(obj); + int invalidated, rc = 0; ENTRY; LASSERT(buf != NULL); LASSERT(name != NULL); + if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_NETWORK) && + osp->opd_index == cfs_fail_val) { + if (is_ost_obj(&dt->do_lu)) { + if (osp_dev2node(osp) == cfs_fail_val) + RETURN(-ENOTCONN); + } else { + if (strcmp(name, XATTR_NAME_LINK) == 0) + RETURN(-ENOTCONN); + } + } + if (unlikely(obj->opo_non_exist)) RETURN(-ENOENT); - oxe = osp_oac_xattr_find(obj, name); + invalidated = atomic_read(&obj->opo_invalidate_seq); + + oxe = osp_oac_xattr_find(obj, name, false); if (oxe != NULL) { spin_lock(&obj->opo_lock); if (oxe->oxe_ready) { @@ -630,7 +990,8 @@ int osp_xattr_get(const struct lu_env *env, struct dt_object *dt, if (buf->lb_len < oxe->oxe_vallen) GOTO(unlock, rc = -ERANGE); - memcpy(buf->lb_buf, oxe->oxe_value, oxe->oxe_vallen); + memcpy(buf->lb_buf, oxe->oxe_value, + oxe->oxe_vallen); GOTO(unlock, rc = oxe->oxe_vallen); @@ -642,25 +1003,47 @@ unlock: } spin_unlock(&obj->opo_lock); } - - update = out_create_update_req(dev); + update = osp_update_request_create(dev); if (IS_ERR(update)) - GOTO(out, rc = PTR_ERR(update)); + GOTO(out_req, rc = PTR_ERR(update)); - namelen = strlen(name) + 1; - rc = out_insert_update(env, update, OUT_XATTR_GET, - lu_object_fid(&dt->do_lu), 1, &namelen, &name); + rc = OSP_UPDATE_RPC_PACK(env, out_xattr_get_pack, update, + lu_object_fid(&dt->do_lu), name, buf->lb_len); if (rc != 0) { CERROR("%s: Insert update error "DFID": rc = %d\n", dname, PFID(lu_object_fid(&dt->do_lu)), rc); + GOTO(out_req, rc); + } + + rc = osp_remote_sync(env, osp, update, &req); + down_read(&obj->opo_invalidate_sem); + if (invalidated != atomic_read(&obj->opo_invalidate_seq)) { + /* invalidated has been requested, we can't cache the result */ + if (rc < 0) { + if (rc == -ENOENT) + dt->do_lu.lo_header->loh_attr &= ~LOHA_EXISTS; + GOTO(out, rc); + } + reply = req_capsule_server_sized_get(&req->rq_pill, + &RMF_OUT_UPDATE_REPLY, + OUT_UPDATE_REPLY_SIZE); + if (reply->ourp_magic != UPDATE_REPLY_MAGIC) { + CERROR("%s: Wrong version %x expected %x "DFID + ": rc = %d\n", dname, reply->ourp_magic, + UPDATE_REPLY_MAGIC, + PFID(lu_object_fid(&dt->do_lu)), -EPROTO); + GOTO(out, rc = -EPROTO); + } + rc = object_update_result_data_get(reply, rbuf, 0); GOTO(out, rc); } - rc = out_remote_sync(env, osp->opd_obd->u.cli.cl_import, update, &req); - if (rc != 0) { - if (obj->opo_ooa == NULL) - GOTO(out, rc); + if (rc < 0) { + if (rc == -ENOENT) { + dt->do_lu.lo_header->loh_attr &= ~LOHA_EXISTS; + obj->opo_non_exist = 1; + } if (oxe == NULL) oxe = osp_oac_xattr_find_or_add(obj, name, buf->lb_len); @@ -697,23 +1080,39 @@ unlock: } rc = object_update_result_data_get(reply, rbuf, 0); - if (rc < 0) - GOTO(out, rc); + if (rc < 0 || rbuf->lb_len == 0) { + if (oxe == NULL && rc == -ENODATA) { + oxe = osp_oac_xattr_find_or_add(obj, name, buf->lb_len); + if (oxe == NULL) { + rc = -ENOMEM; + CWARN("%s: Fail to add xattr (%s) to cache for " + DFID" (1): rc = %d\n", dname, name, + PFID(lu_object_fid(&dt->do_lu)), rc); + GOTO(out, rc); + } + } - if (buf->lb_buf == NULL) - GOTO(out, rc = rbuf->lb_len); + if (oxe) { + spin_lock(&obj->opo_lock); + if (unlikely(rc == -ENODATA)) { + oxe->oxe_exist = 0; + oxe->oxe_ready = 1; + } else { + oxe->oxe_ready = 0; + } + spin_unlock(&obj->opo_lock); + } - if (unlikely(buf->lb_len < rbuf->lb_len)) - GOTO(out, rc = -ERANGE); + GOTO(out, rc); + } - memcpy(buf->lb_buf, rbuf->lb_buf, rbuf->lb_len); - rc = rbuf->lb_len; - if (obj->opo_ooa == NULL) + /* For detecting EA size. */ + if (!buf->lb_buf) GOTO(out, rc); - if (oxe == NULL) { + if (!oxe) { oxe = osp_oac_xattr_find_or_add(obj, name, rbuf->lb_len); - if (oxe == NULL) { + if (!oxe) { CWARN("%s: Fail to add xattr (%s) to " "cache for "DFID" (2): rc = %d\n", dname, name, PFID(lu_object_fid(&dt->do_lu)), rc); @@ -722,185 +1121,304 @@ unlock: } } - if (oxe->oxe_buflen - oxe->oxe_namelen - 1 < rbuf->lb_len) { - struct osp_xattr_entry *old = oxe; - struct osp_xattr_entry *tmp; - - tmp = osp_oac_xattr_replace(obj, &old, rbuf->lb_len); - osp_oac_xattr_put(oxe); - oxe = tmp; - if (tmp == NULL) { - CWARN("%s: Fail to update xattr (%s) to " - "cache for "DFID": rc = %d\n", - dname, name, PFID(lu_object_fid(&dt->do_lu)), rc); - spin_lock(&obj->opo_lock); - old->oxe_ready = 0; - spin_unlock(&obj->opo_lock); - - GOTO(out, rc); - } - - /* Drop the ref for entry on list. */ - osp_oac_xattr_put(old); - } - - spin_lock(&obj->opo_lock); - oxe->oxe_vallen = rbuf->lb_len; - memcpy(oxe->oxe_value, rbuf->lb_buf, rbuf->lb_len); - oxe->oxe_exist = 1; - oxe->oxe_ready = 1; - spin_unlock(&obj->opo_lock); + oxe = osp_oac_xattr_assignment(obj, oxe, rbuf); GOTO(out, rc); out: - if (req != NULL) + up_read(&obj->opo_invalidate_sem); + +out_req: + if (rc > 0 && buf->lb_buf) { + if (unlikely(buf->lb_len < rbuf->lb_len)) + rc = -ERANGE; + else + memcpy(buf->lb_buf, rbuf->lb_buf, rbuf->lb_len); + } + + if (req) ptlrpc_req_finished(req); - if (update != NULL && !IS_ERR(update)) - out_destroy_update_req(update); + if (update && !IS_ERR(update)) + osp_update_request_destroy(env, update); - if (oxe != NULL) + if (oxe) osp_oac_xattr_put(oxe); return rc; } +/** + * Implement OSP layer dt_object_operations::do_declare_xattr_set() interface. + * + * Declare that the caller will set extended attribute to the specified + * MDT/OST object. + * + * If it is non-remote transaction, it will add an OUT_XATTR_SET sub-request + * to the OUT RPC that will be flushed when the transaction start. And if the + * OSP attributes cache is initialized, then check whether the name extended + * attribute entry exists in the cache or not. If yes, replace it; otherwise, + * add the extended attribute to the cache. + * + * \param[in] env pointer to the thread context + * \param[in] dt pointer to the OSP layer dt_object + * \param[in] buf pointer to the lu_buf to hold the extended attribute + * \param[in] name the name of the extended attribute to be set + * \param[in] flag to indicate the detailed set operation: LU_XATTR_CREATE + * or LU_XATTR_REPLACE or others + * \param[in] th pointer to the transaction handler + * + * \retval 0 for success + * \retval negative error number on failure + */ int osp_declare_xattr_set(const struct lu_env *env, struct dt_object *dt, const struct lu_buf *buf, const char *name, int flag, struct thandle *th) { - struct osp_object *o = dt2osp_obj(dt); - struct dt_update_request *update; - struct lu_fid *fid; - struct osp_xattr_entry *oxe; - int sizes[3] = {strlen(name), buf->lb_len, - sizeof(int)}; - char *bufs[3] = {(char *)name, (char *)buf->lb_buf }; - int rc; + return osp_trans_update_request_create(th); +} + +/** + * Implement OSP layer dt_object_operations::do_xattr_set() interface. + * + * Set extended attribute to the specified MDT/OST object. + * + * Add an OUT_XATTR_SET sub-request into the OUT RPC that will be flushed in + * the transaction stop. And if the OSP attributes cache is initialized, then + * check whether the name extended attribute entry exists in the cache or not. + * If yes, replace it; otherwise, add the extended attribute to the cache. + * + * \param[in] env pointer to the thread context + * \param[in] dt pointer to the OSP layer dt_object + * \param[in] buf pointer to the lu_buf to hold the extended attribute + * \param[in] name the name of the extended attribute to be set + * \param[in] fl to indicate the detailed set operation: LU_XATTR_CREATE + * or LU_XATTR_REPLACE or others + * \param[in] th pointer to the transaction handler + * + * \retval 0 for success + * \retval negative error number on failure + */ +int osp_xattr_set(const struct lu_env *env, struct dt_object *dt, + const struct lu_buf *buf, const char *name, int fl, + struct thandle *th) +{ + struct osp_object *o = dt2osp_obj(dt); + struct osp_update_request *update; + struct osp_xattr_entry *oxe; + int rc; + ENTRY; - LASSERT(buf->lb_len > 0 && buf->lb_buf != NULL); + update = thandle_to_osp_update_request(th); + LASSERT(update != NULL); - update = out_find_create_update_loc(th, dt); - if (IS_ERR(update)) { - CERROR("%s: Get OSP update buf failed "DFID": rc = %d\n", - dt->do_lu.lo_dev->ld_obd->obd_name, - PFID(lu_object_fid(&dt->do_lu)), - (int)PTR_ERR(update)); + CDEBUG(D_INODE, DFID" set xattr '%s' with size %zd\n", + PFID(lu_object_fid(&dt->do_lu)), name, buf->lb_len); - return PTR_ERR(update); - } + rc = OSP_UPDATE_RPC_PACK(env, out_xattr_set_pack, update, + lu_object_fid(&dt->do_lu), buf, name, fl); + if (rc != 0) + RETURN(rc); - flag = cpu_to_le32(flag); - bufs[2] = (char *)&flag; + /* Do not cache linkEA that may be self-adjusted by peers + * under EA overflow case. */ + if (strcmp(name, XATTR_NAME_LINK) == 0) { + oxe = osp_oac_xattr_find(o, name, true); + if (oxe != NULL) + osp_oac_xattr_put(oxe); - fid = (struct lu_fid *)lu_object_fid(&dt->do_lu); - rc = out_insert_update(env, update, OUT_XATTR_SET, fid, - ARRAY_SIZE(sizes), sizes, (const char **)bufs); - if (rc != 0 || o->opo_ooa == NULL) - return rc; + RETURN(0); + } oxe = osp_oac_xattr_find_or_add(o, name, buf->lb_len); if (oxe == NULL) { - CWARN("%s: Fail to add xattr (%s) to cache for "DFID - ": rc = %d\n", dt->do_lu.lo_dev->ld_obd->obd_name, - name, PFID(lu_object_fid(&dt->do_lu)), rc); + CWARN("%s: cannot cache xattr '%s' of "DFID"\n", + osp_dto2name(o), name, PFID(lu_object_fid(&dt->do_lu))); - return 0; + RETURN(0); } - if (oxe->oxe_buflen - oxe->oxe_namelen - 1 < buf->lb_len) { - struct osp_xattr_entry *old = oxe; - struct osp_xattr_entry *tmp; - - tmp = osp_oac_xattr_replace(o, &old, buf->lb_len); + oxe = osp_oac_xattr_assignment(o, oxe, buf); + if (oxe) osp_oac_xattr_put(oxe); - oxe = tmp; - if (tmp == NULL) { - CWARN("%s: Fail to update xattr (%s) to cache for "DFID - ": rc = %d\n", dt->do_lu.lo_dev->ld_obd->obd_name, - name, PFID(lu_object_fid(&dt->do_lu)), rc); - spin_lock(&o->opo_lock); - old->oxe_ready = 0; - spin_unlock(&o->opo_lock); - - return 0; - } - /* Drop the ref for entry on list. */ - osp_oac_xattr_put(old); - } + RETURN(0); +} - spin_lock(&o->opo_lock); - oxe->oxe_vallen = buf->lb_len; - memcpy(oxe->oxe_value, buf->lb_buf, buf->lb_len); - oxe->oxe_exist = 1; - oxe->oxe_ready = 1; - spin_unlock(&o->opo_lock); - osp_oac_xattr_put(oxe); +/** + * Implement OSP layer dt_object_operations::do_declare_xattr_del() interface. + * + * Declare that the caller will delete extended attribute on the specified + * MDT/OST object. + * + * If it is non-remote transaction, it will add an OUT_XATTR_DEL sub-request + * to the OUT RPC that will be flushed when the transaction start. And if the + * name extended attribute entry exists in the OSP attributes cache, then remove + * it from the cache. + * + * \param[in] env pointer to the thread context + * \param[in] dt pointer to the OSP layer dt_object + * \param[in] name the name of the extended attribute to be set + * \param[in] th pointer to the transaction handler + * + * \retval 0 for success + * \retval negative error number on failure + */ +int osp_declare_xattr_del(const struct lu_env *env, struct dt_object *dt, + const char *name, struct thandle *th) +{ + return osp_trans_update_request_create(th); +} + +/** + * Implement OSP layer dt_object_operations::do_xattr_del() interface. + * + * Delete extended attribute on the specified MDT/OST object. + * + * If it is remote transaction, it will add an OUT_XATTR_DEL sub-request into + * the OUT RPC that will be flushed when the transaction stop. And if the name + * extended attribute entry exists in the OSP attributes cache, then remove it + * from the cache. + * + * \param[in] env pointer to the thread context + * \param[in] dt pointer to the OSP layer dt_object + * \param[in] name the name of the extended attribute to be set + * \param[in] th pointer to the transaction handler + * + * \retval 0 for success + * \retval negative error number on failure + */ +int osp_xattr_del(const struct lu_env *env, struct dt_object *dt, + const char *name, struct thandle *th) +{ + struct osp_update_request *update; + const struct lu_fid *fid = lu_object_fid(&dt->do_lu); + struct osp_object *o = dt2osp_obj(dt); + struct osp_xattr_entry *oxe; + int rc; + + update = thandle_to_osp_update_request(th); + LASSERT(update != NULL); + + rc = OSP_UPDATE_RPC_PACK(env, out_xattr_del_pack, update, fid, name); + if (rc != 0) + return rc; + + oxe = osp_oac_xattr_find(o, name, true); + if (oxe != NULL) + /* Drop the ref for entry on list. */ + osp_oac_xattr_put(oxe); return 0; } -int osp_xattr_set(const struct lu_env *env, struct dt_object *dt, - const struct lu_buf *buf, const char *name, int fl, - struct thandle *th, struct lustre_capa *capa) +void osp_obj_invalidate_cache(struct osp_object *obj) { - CDEBUG(D_INFO, "xattr %s set object "DFID"\n", name, - PFID(&dt->do_lu.lo_header->loh_fid)); + struct osp_xattr_entry *oxe; + struct osp_xattr_entry *tmp; - return 0; + spin_lock(&obj->opo_lock); + list_for_each_entry_safe(oxe, tmp, &obj->opo_xattr_list, oxe_list) { + oxe->oxe_ready = 0; + list_del_init(&oxe->oxe_list); + osp_oac_xattr_put(oxe); + } + obj->opo_attr.la_valid = 0; + spin_unlock(&obj->opo_lock); } -int osp_declare_xattr_del(const struct lu_env *env, struct dt_object *dt, - const char *name, struct thandle *th) +/** + * Implement OSP layer dt_object_operations::do_invalidate() interface. + * + * Invalidate attributes cached on the specified MDT/OST object. + * + * \param[in] env pointer to the thread context + * \param[in] dt pointer to the OSP layer dt_object + * + * \retval 0 for success + * \retval negative error number on failure + */ +int osp_invalidate(const struct lu_env *env, struct dt_object *dt) { - struct dt_update_request *update; - const struct lu_fid *fid; - int size = strlen(name); - int rc; + struct osp_object *obj = dt2osp_obj(dt); + ENTRY; - update = out_find_create_update_loc(th, dt); - if (IS_ERR(update)) - return PTR_ERR(update); + CDEBUG(D_HA, "Invalidate osp_object "DFID"\n", + PFID(lu_object_fid(&dt->do_lu))); - fid = lu_object_fid(&dt->do_lu); + /* serialize attr/EA set vs. invalidation */ + down_write(&obj->opo_invalidate_sem); - rc = out_insert_update(env, update, OUT_XATTR_DEL, fid, 1, &size, - (const char **)&name); + /* this should invalidate all in-flights */ + atomic_inc(&obj->opo_invalidate_seq); - return rc; + spin_lock(&obj->opo_lock); + /* do not mark new objects stale */ + if (obj->opo_attr.la_valid) + obj->opo_stale = 1; + obj->opo_non_exist = 0; + spin_unlock(&obj->opo_lock); + + osp_obj_invalidate_cache(obj); + + up_write(&obj->opo_invalidate_sem); + + RETURN(0); } -int osp_xattr_del(const struct lu_env *env, struct dt_object *dt, - const char *name, struct thandle *th, - struct lustre_capa *capa) +bool osp_check_stale(struct dt_object *dt) { - CDEBUG(D_INFO, "xattr %s del object "DFID"\n", name, - PFID(&dt->do_lu.lo_header->loh_fid)); + struct osp_object *obj = dt2osp_obj(dt); - return 0; + if (is_ost_obj(&dt->do_lu) && obj->opo_non_exist) + return true; + + return obj->opo_stale; } -static int osp_declare_object_create(const struct lu_env *env, - struct dt_object *dt, - struct lu_attr *attr, - struct dt_allocation_hint *hint, - struct dt_object_format *dof, - struct thandle *th) + +/** + * Implement OSP layer dt_object_operations::do_declare_create() interface. + * + * Declare that the caller will create the OST object. + * + * If the transaction is a remote transaction and the FID for the OST-object + * has been assigned already, then handle it as creating (remote) MDT object + * via osp_md_declare_create(). This function is usually used for LFSCK + * to re-create the lost OST object. Otherwise, if it is not replay case, the + * OSP will reserve pre-created object for the subsequent create operation; + * if the MDT side cached pre-created objects are less than some threshold, + * then it will wakeup the pre-create thread. + * + * \param[in] env pointer to the thread context + * \param[in] dt pointer to the OSP layer dt_object + * \param[in] attr the attribute for the object to be created + * \param[in] hint pointer to the hint for creating the object, such as + * the parent object + * \param[in] dof pointer to the dt_object_format for help the creation + * \param[in] th pointer to the transaction handler + * + * \retval 0 for success + * \retval negative error number on failure + */ +static int osp_declare_create(const struct lu_env *env, struct dt_object *dt, + struct lu_attr *attr, + struct dt_allocation_hint *hint, + struct dt_object_format *dof, struct thandle *th) { struct osp_thread_info *osi = osp_env_info(env); struct osp_device *d = lu2osp_dev(dt->do_lu.lo_dev); struct osp_object *o = dt2osp_obj(dt); const struct lu_fid *fid = lu_object_fid(&dt->do_lu); + struct thandle *local_th; int rc = 0; ENTRY; - if (is_only_remote_trans(th)) { + if (is_only_remote_trans(th) && !fid_is_zero(fid)) { LASSERT(fid_is_sane(fid)); - rc = osp_md_declare_object_create(env, dt, attr, hint, dof, th); + rc = osp_md_declare_create(env, dt, attr, hint, dof, th); RETURN(rc); } @@ -921,13 +1439,17 @@ static int osp_declare_object_create(const struct lu_env *env, */ /* rc = osp_sync_declare_add(env, o, MDS_UNLINK64_REC, th); */ + local_th = osp_get_storage_thandle(env, th, d); + if (IS_ERR(local_th)) + RETURN(PTR_ERR(local_th)); + if (unlikely(!fid_is_zero(fid))) { /* replay case: caller knows fid */ - osi->osi_off = sizeof(osi->osi_id) * d->opd_index; - osi->osi_lb.lb_len = sizeof(osi->osi_id); - osi->osi_lb.lb_buf = NULL; + osp_objid_buf_prep(&osi->osi_lb, &osi->osi_off, NULL, + d->opd_index); rc = dt_declare_record_write(env, d->opd_last_used_oid_file, - &osi->osi_lb, osi->osi_off, th); + &osi->osi_lb, osi->osi_off, + local_th); RETURN(rc); } @@ -948,11 +1470,11 @@ static int osp_declare_object_create(const struct lu_env *env, o->opo_reserved = 1; /* common for all OSPs file hystorically */ - osi->osi_off = sizeof(osi->osi_id) * d->opd_index; - osi->osi_lb.lb_len = sizeof(osi->osi_id); - osi->osi_lb.lb_buf = NULL; + osp_objid_buf_prep(&osi->osi_lb, &osi->osi_off, NULL, + d->opd_index); rc = dt_declare_record_write(env, d->opd_last_used_oid_file, - &osi->osi_lb, osi->osi_off, th); + &osi->osi_lb, osi->osi_off, + local_th); } else { /* not needed in the cache anymore */ set_bit(LU_OBJECT_HEARD_BANSHEE, @@ -961,22 +1483,44 @@ static int osp_declare_object_create(const struct lu_env *env, RETURN(rc); } -static int osp_object_create(const struct lu_env *env, struct dt_object *dt, - struct lu_attr *attr, - struct dt_allocation_hint *hint, - struct dt_object_format *dof, struct thandle *th) +/** + * Implement OSP layer dt_object_operations::do_create() interface. + * + * Create the OST object. + * + * If the transaction is a remote transaction and the FID for the OST-object + * has been assigned already, then handle it as handling MDT object via the + * osp_md_create(). For other cases, the OSP will assign FID to the + * object to be created, and update last_used Object ID (OID) file. + * + * \param[in] env pointer to the thread context + * \param[in] dt pointer to the OSP layer dt_object + * \param[in] attr the attribute for the object to be created + * \param[in] hint pointer to the hint for creating the object, such as + * the parent object + * \param[in] dof pointer to the dt_object_format for help the creation + * \param[in] th pointer to the transaction handler + * + * \retval 0 for success + * \retval negative error number on failure + */ +static int osp_create(const struct lu_env *env, struct dt_object *dt, + struct lu_attr *attr, struct dt_allocation_hint *hint, + struct dt_object_format *dof, struct thandle *th) { struct osp_thread_info *osi = osp_env_info(env); struct osp_device *d = lu2osp_dev(dt->do_lu.lo_dev); struct osp_object *o = dt2osp_obj(dt); int rc = 0; struct lu_fid *fid = &osi->osi_fid; + struct thandle *local_th; ENTRY; - if (is_only_remote_trans(th)) { + if (is_only_remote_trans(th) && + !fid_is_zero(lu_object_fid(&dt->do_lu))) { LASSERT(fid_is_sane(lu_object_fid(&dt->do_lu))); - rc = osp_md_object_create(env, dt, attr, hint, dof, th); + rc = osp_md_create(env, dt, attr, hint, dof, th); if (rc == 0) o->opo_non_exist = 0; @@ -985,7 +1529,7 @@ static int osp_object_create(const struct lu_env *env, struct dt_object *dt, o->opo_non_exist = 0; if (o->opo_reserved) { - /* regular case, fid is assigned holding trunsaction open */ + /* regular case, fid is assigned holding transaction open */ osp_object_assign_fid(env, d, o); } @@ -1012,6 +1556,9 @@ static int osp_object_create(const struct lu_env *env, struct dt_object *dt, if (osp_precreate_end_seq(env, d) && osp_is_fid_client(d)) th->th_sync = 1; + local_th = osp_get_storage_thandle(env, th, d); + if (IS_ERR(local_th)) + RETURN(PTR_ERR(local_th)); /* * it's OK if the import is inactive by this moment - id was created * by OST earlier, we just need to maintain it consistently on the disk @@ -1026,12 +1573,16 @@ static int osp_object_create(const struct lu_env *env, struct dt_object *dt, if (d->opd_gap_count > 0) { int count = d->opd_gap_count; - ostid_set_id(&osi->osi_oi, - fid_oid(&d->opd_gap_start_fid)); + rc = ostid_set_id(&osi->osi_oi, + fid_oid(&d->opd_gap_start_fid)); + if (rc) { + spin_unlock(&d->opd_pre_lock); + RETURN(rc); + } d->opd_gap_count = 0; spin_unlock(&d->opd_pre_lock); - CDEBUG(D_HA, "Writting gap "DFID"+%d in llog\n", + CDEBUG(D_HA, "Writing gap "DFID"+%d in llog\n", PFID(&d->opd_gap_start_fid), count); /* real gap handling is disabled intil ORI-692 will be * fixed, now we only report gaps */ @@ -1040,17 +1591,13 @@ static int osp_object_create(const struct lu_env *env, struct dt_object *dt, } } - /* new object, the very first ->attr_set() - * initializing attributes needs no logging */ - o->opo_new = 1; - /* Only need update last_used oid file, seq file will only be update * during seq rollover */ osp_objid_buf_prep(&osi->osi_lb, &osi->osi_off, - &d->opd_last_used_fid.f_oid, d->opd_index); + &d->opd_last_id, d->opd_index); rc = dt_record_write(env, d->opd_last_used_oid_file, &osi->osi_lb, - &osi->osi_off, th); + &osi->osi_off, local_th); CDEBUG(D_HA, "%s: Wrote last used FID: "DFID", index %d: %d\n", d->opd_obd->obd_name, PFID(fid), d->opd_index, rc); @@ -1058,36 +1605,75 @@ static int osp_object_create(const struct lu_env *env, struct dt_object *dt, RETURN(rc); } -int osp_declare_object_destroy(const struct lu_env *env, - struct dt_object *dt, struct thandle *th) +/** + * Implement OSP layer dt_object_operations::do_declare_destroy() interface. + * + * Declare that the caller will destroy the specified OST object. + * + * The OST object destroy will be handled via llog asynchronously. This + * function will declare the credits for generating MDS_UNLINK64_REC llog. + * + * \param[in] env pointer to the thread context + * \param[in] dt pointer to the OSP layer dt_object to be destroyed + * \param[in] th pointer to the transaction handler + * + * \retval 0 for success + * \retval negative error number on failure + */ +int osp_declare_destroy(const struct lu_env *env, struct dt_object *dt, + struct thandle *th) { struct osp_object *o = dt2osp_obj(dt); + struct osp_device *osp = lu2osp_dev(dt->do_lu.lo_dev); int rc = 0; ENTRY; - /* - * track objects to be destroyed via llog - */ - rc = osp_sync_declare_add(env, o, MDS_UNLINK64_REC, th); + LASSERT(!osp->opd_connect_mdt); + + if (!OBD_FAIL_CHECK(OBD_FAIL_LFSCK_LOST_MDTOBJ)) + rc = osp_sync_declare_add(env, o, MDS_UNLINK64_REC, th); RETURN(rc); } -int osp_object_destroy(const struct lu_env *env, struct dt_object *dt, +/** + * Implement OSP layer dt_object_operations::do_destroy() interface. + * + * Destroy the specified OST object. + * + * The OSP generates a MDS_UNLINK64_REC record in the llog. There + * will be some dedicated thread to handle the llog asynchronously. + * + * It also marks the object as non-cached. + * + * \param[in] env pointer to the thread context + * \param[in] dt pointer to the OSP layer dt_object to be destroyed + * \param[in] th pointer to the transaction handler + * + * \retval 0 for success + * \retval negative error number on failure + */ +static int osp_destroy(const struct lu_env *env, struct dt_object *dt, struct thandle *th) { struct osp_object *o = dt2osp_obj(dt); + struct osp_device *osp = lu2osp_dev(dt->do_lu.lo_dev); int rc = 0; ENTRY; o->opo_non_exist = 1; - /* - * once transaction is committed put proper command on - * the queue going to our OST - */ - rc = osp_sync_add(env, o, MDS_UNLINK64_REC, th, NULL); + + LASSERT(!osp->opd_connect_mdt); + + if (!OBD_FAIL_CHECK(OBD_FAIL_LFSCK_LOST_MDTOBJ)) { + /* once transaction is committed put proper command on + * the queue going to our OST. */ + rc = osp_sync_add(env, o, MDS_UNLINK64_REC, th, NULL); + if (rc < 0) + RETURN(rc); + } /* not needed in cache any more */ set_bit(LU_OBJECT_HEARD_BANSHEE, &dt->do_lu.lo_header->loh_flags); @@ -1098,8 +1684,7 @@ int osp_object_destroy(const struct lu_env *env, struct dt_object *dt, static int osp_orphan_index_lookup(const struct lu_env *env, struct dt_object *dt, struct dt_rec *rec, - const struct dt_key *key, - struct lustre_capa *capa) + const struct dt_key *key) { return -EOPNOTSUPP; } @@ -1117,9 +1702,7 @@ static int osp_orphan_index_insert(const struct lu_env *env, struct dt_object *dt, const struct dt_rec *rec, const struct dt_key *key, - struct thandle *handle, - struct lustre_capa *capa, - int ignore_quota) + struct thandle *handle) { return -EOPNOTSUPP; } @@ -1135,14 +1718,23 @@ static int osp_orphan_index_declare_delete(const struct lu_env *env, static int osp_orphan_index_delete(const struct lu_env *env, struct dt_object *dt, const struct dt_key *key, - struct thandle *handle, - struct lustre_capa *capa) + struct thandle *handle) { return -EOPNOTSUPP; } +/** + * Initialize the OSP layer index iteration. + * + * \param[in] env pointer to the thread context + * \param[in] dt pointer to the index object to be iterated + * \param[in] attr unused + * + * \retval pointer to the iteration structure + * \retval negative error number on failure + */ struct dt_it *osp_it_init(const struct lu_env *env, struct dt_object *dt, - __u32 attr, struct lustre_capa *capa) + __u32 attr) { struct osp_it *it; @@ -1152,10 +1744,17 @@ struct dt_it *osp_it_init(const struct lu_env *env, struct dt_object *dt, it->ooi_pos_ent = -1; it->ooi_obj = dt; + it->ooi_attr = attr; return (struct dt_it *)it; } +/** + * Finalize the OSP layer index iteration. + * + * \param[in] env pointer to the thread context + * \param[in] di pointer to the iteration structure + */ void osp_it_fini(const struct lu_env *env, struct dt_it *di) { struct osp_it *it = (struct osp_it *)di; @@ -1173,11 +1772,23 @@ void osp_it_fini(const struct lu_env *env, struct dt_it *di) __free_page(pages[i]); } } - OBD_FREE(pages, npages * sizeof(*pages)); + OBD_FREE_PTR_ARRAY(pages, npages); } OBD_FREE_PTR(it); } +/** + * Get more records for the iteration from peer. + * + * The new records will be filled in an array of pages. The OSP side + * allows 1MB bulk data to be transferred. + * + * \param[in] env pointer to the thread context + * \param[in] it pointer to the iteration structure + * + * \retval 0 for success + * \retval negative error number on failure + */ static int osp_it_fetch(const struct lu_env *env, struct osp_it *it) { struct lu_device *dev = it->ooi_obj->do_lu.lo_dev; @@ -1193,16 +1804,16 @@ static int osp_it_fetch(const struct lu_env *env, struct osp_it *it) /* 1MB bulk */ npages = min_t(unsigned int, OFD_MAX_BRW_SIZE, 1 << 20); - npages /= PAGE_CACHE_SIZE; + npages /= PAGE_SIZE; - OBD_ALLOC(pages, npages * sizeof(*pages)); + OBD_ALLOC_PTR_ARRAY(pages, npages); if (pages == NULL) RETURN(-ENOMEM); it->ooi_pages = pages; it->ooi_total_npages = npages; for (i = 0; i < npages; i++) { - pages[i] = alloc_page(GFP_IOFS); + pages[i] = alloc_page(GFP_NOFS); if (pages[i] == NULL) RETURN(-ENOMEM); } @@ -1218,6 +1829,7 @@ static int osp_it_fetch(const struct lu_env *env, struct osp_it *it) RETURN(rc); } + osp_set_req_replay(osp, req); req->rq_request_portal = OUT_PORTAL; ii = req_capsule_client_get(&req->rq_pill, &RMF_IDX_INFO); memset(ii, 0, sizeof(*ii)); @@ -1229,28 +1841,29 @@ static int osp_it_fetch(const struct lu_env *env, struct osp_it *it) ii->ii_fid.f_oid = osp->opd_index; ii->ii_fid.f_ver = 0; ii->ii_flags = II_FL_NOHASH; + ii->ii_attrs = osp_dev2node(osp); } else { ii->ii_fid = *lu_object_fid(&it->ooi_obj->do_lu); ii->ii_flags = II_FL_NOHASH | II_FL_NOKEY | II_FL_VARKEY | II_FL_VARREC; + ii->ii_attrs = it->ooi_attr; } ii->ii_magic = IDX_INFO_MAGIC; ii->ii_count = npages * LU_PAGE_COUNT; ii->ii_hash_start = it->ooi_next; - ii->ii_attrs = - osp->opd_storage->dd_lu_dev.ld_site->ld_seq_site->ss_node_id; ptlrpc_at_set_req_timeout(req); - desc = ptlrpc_prep_bulk_imp(req, npages, 1, BULK_PUT_SINK, - MDS_BULK_PORTAL); - if (desc == NULL) { - ptlrpc_request_free(req); - RETURN(-ENOMEM); - } + desc = ptlrpc_prep_bulk_imp(req, npages, 1, + PTLRPC_BULK_PUT_SINK, + MDS_BULK_PORTAL, + &ptlrpc_bulk_kiov_pin_ops); + if (desc == NULL) + GOTO(out, rc = -ENOMEM); for (i = 0; i < npages; i++) - ptlrpc_prep_bulk_page_pin(desc, pages[i], 0, PAGE_CACHE_SIZE); + desc->bd_frag_ops->add_kiov_frag(desc, pages[i], 0, + PAGE_SIZE); ptlrpc_request_set_replen(req); rc = ptlrpc_queue_wait(req); @@ -1268,13 +1881,14 @@ static int osp_it_fetch(const struct lu_env *env, struct osp_it *it) GOTO(out, rc = -EPROTO); npages = (ii->ii_count + LU_PAGE_COUNT - 1) >> - (PAGE_CACHE_SHIFT - LU_PAGE_SHIFT); + (PAGE_SHIFT - LU_PAGE_SHIFT); if (npages > it->ooi_total_npages) { CERROR("%s: returned more pages than expected, %u > %u\n", osp->opd_obd->obd_name, npages, it->ooi_total_npages); GOTO(out, rc = -EINVAL); } + it->ooi_rec_size = ii->ii_recsize; it->ooi_valid_npages = npages; if (ptlrpc_rep_need_swab(req)) it->ooi_swab = 1; @@ -1287,6 +1901,24 @@ out: return rc; } +/** + * Move the iteration cursor to the next lu_page. + * + * One system page (PAGE_SIZE) may contain multiple lu_page (4KB), + * that depends on the LU_PAGE_COUNT. If it is not the last lu_page + * in current system page, then move the iteration cursor to the next + * lu_page in current system page. Otherwise, if there are more system + * pages in the cache, then move the iteration cursor to the next system + * page. If all the cached records (pages) have been iterated, then fetch + * more records via osp_it_fetch(). + * + * \param[in] env pointer to the thread context + * \param[in] di pointer to the iteration structure + * + * \retval positive for end of the directory + * \retval 0 for success + * \retval negative error number on failure + */ int osp_it_next_page(const struct lu_env *env, struct dt_it *di) { struct osp_it *it = (struct osp_it *)di; @@ -1309,6 +1941,7 @@ again2: } it->ooi_cur_idxpage = NULL; it->ooi_pos_lu_page++; + again1: if (it->ooi_pos_lu_page < LU_PAGE_COUNT) { it->ooi_cur_idxpage = (void *)it->ooi_cur_page + @@ -1349,7 +1982,7 @@ again0: if (pages[i] != NULL) __free_page(pages[i]); } - OBD_FREE(pages, it->ooi_total_npages * sizeof(*pages)); + OBD_FREE_PTR_ARRAY(pages, it->ooi_total_npages); it->ooi_pos_page = 0; it->ooi_total_npages = 0; @@ -1371,7 +2004,21 @@ again0: RETURN(rc); } -int osp_orphan_it_next(const struct lu_env *env, struct dt_it *di) +/** + * Move the iteration cursor to the next record. + * + * If there are more records in the lu_page, then move the iteration + * cursor to the next record directly. Otherwise, move the iteration + * cursor to the record in the next lu_page via osp_it_next_page() + * + * \param[in] env pointer to the thread context + * \param[in] di pointer to the iteration structure + * + * \retval positive for end of the directory + * \retval 0 for success + * \retval negative error number on failure + */ +static int osp_orphan_it_next(const struct lu_env *env, struct dt_it *di) { struct osp_it *it = (struct osp_it *)di; struct lu_idxpage *idxpage; @@ -1386,11 +2033,27 @@ again: it->ooi_pos_ent++; if (it->ooi_pos_ent < idxpage->lip_nr) { - it->ooi_ent = + if (it->ooi_rec_size == + sizeof(struct lu_orphan_rec_v3)) { + it->ooi_ent = + (struct lu_orphan_ent_v3 *)idxpage->lip_entries+ + it->ooi_pos_ent; + if (it->ooi_swab) + lustre_swab_orphan_ent_v3(it->ooi_ent); + } else if (it->ooi_rec_size == + sizeof(struct lu_orphan_rec_v2)) { + it->ooi_ent = + (struct lu_orphan_ent_v2 *)idxpage->lip_entries+ + it->ooi_pos_ent; + if (it->ooi_swab) + lustre_swab_orphan_ent_v2(it->ooi_ent); + } else { + it->ooi_ent = (struct lu_orphan_ent *)idxpage->lip_entries + it->ooi_pos_ent; - if (it->ooi_swab) - lustre_swab_orphan_ent(it->ooi_ent); + if (it->ooi_swab) + lustre_swab_orphan_ent(it->ooi_ent); + } RETURN(0); } } @@ -1412,8 +2075,8 @@ void osp_it_put(const struct lu_env *env, struct dt_it *di) { } -struct dt_key *osp_orphan_it_key(const struct lu_env *env, - const struct dt_it *di) +static struct dt_key *osp_orphan_it_key(const struct lu_env *env, + const struct dt_it *di) { struct osp_it *it = (struct osp_it *)di; struct lu_orphan_ent *ent = (struct lu_orphan_ent *)it->ooi_ent; @@ -1424,19 +2087,35 @@ struct dt_key *osp_orphan_it_key(const struct lu_env *env, return NULL; } -int osp_orphan_it_key_size(const struct lu_env *env, const struct dt_it *di) +static int osp_orphan_it_key_size(const struct lu_env *env, + const struct dt_it *di) { return sizeof(struct lu_fid); } -int osp_orphan_it_rec(const struct lu_env *env, const struct dt_it *di, - struct dt_rec *rec, __u32 attr) +static int osp_orphan_it_rec(const struct lu_env *env, const struct dt_it *di, + struct dt_rec *rec, __u32 attr) { - struct osp_it *it = (struct osp_it *)di; - struct lu_orphan_ent *ent = (struct lu_orphan_ent *)it->ooi_ent; + struct osp_it *it = (struct osp_it *)di; + + if (likely(it->ooi_ent)) { + if (it->ooi_rec_size == sizeof(struct lu_orphan_rec_v3)) { + struct lu_orphan_ent_v3 *ent = + (struct lu_orphan_ent_v3 *)it->ooi_ent; + + *(struct lu_orphan_rec_v3 *)rec = ent->loe_rec; + } else if (it->ooi_rec_size == + sizeof(struct lu_orphan_rec_v2)) { + struct lu_orphan_ent_v2 *ent = + (struct lu_orphan_ent_v2 *)it->ooi_ent; + + *(struct lu_orphan_rec_v2 *)rec = ent->loe_rec; + } else { + struct lu_orphan_ent *ent = + (struct lu_orphan_ent *)it->ooi_ent; - if (likely(ent != NULL)) { - *(struct lu_orphan_rec *)rec = ent->loe_rec; + *(struct lu_orphan_rec *)rec = ent->loe_rec; + } return 0; } @@ -1451,16 +2130,22 @@ __u64 osp_it_store(const struct lu_env *env, const struct dt_it *di) } /** - * \retval +1: locate to the exactly position - * \retval 0: cannot locate to the exactly position, - * call next() to move to a valid position. - * \retval -ve: on error + * Locate the iteration cursor to the specified position (cookie). + * + * \param[in] env pointer to the thread context + * \param[in] di pointer to the iteration structure + * \param[in] hash the specified position + * + * \retval positive number for locating to the exactly position + * or the next + * \retval 0 for arriving at the end of the iteration + * \retval negative error number on failure */ int osp_orphan_it_load(const struct lu_env *env, const struct dt_it *di, - __u64 hash) + __u64 hash) { struct osp_it *it = (struct osp_it *)di; - int rc; + int rc; it->ooi_next = hash; rc = osp_orphan_it_next(env, (struct dt_it *)di); @@ -1500,6 +2185,21 @@ static const struct dt_index_operations osp_orphan_index_ops = { } }; +/** + * Implement OSP layer dt_object_operations::do_index_try() interface. + * + * Negotiate the index type. + * + * If the target index is an IDIF object, then use osp_orphan_index_ops. + * Otherwise, assign osp_md_index_ops to the dt_object::do_index_ops. + * (\see lustre/include/lustre_fid.h for IDIF.) + * + * \param[in] env pointer to the thread context + * \param[in] dt pointer to the OSP layer dt_object + * \param[in] feat unused + * + * \retval 0 for success + */ static int osp_index_try(const struct lu_env *env, struct dt_object *dt, const struct dt_index_features *feat) @@ -1513,7 +2213,7 @@ static int osp_index_try(const struct lu_env *env, return 0; } -struct dt_object_operations osp_obj_ops = { +static struct dt_object_operations osp_obj_ops = { .do_declare_attr_get = osp_declare_attr_get, .do_attr_get = osp_attr_get, .do_declare_attr_set = osp_declare_attr_set, @@ -1522,13 +2222,28 @@ struct dt_object_operations osp_obj_ops = { .do_xattr_get = osp_xattr_get, .do_declare_xattr_set = osp_declare_xattr_set, .do_xattr_set = osp_xattr_set, - .do_declare_create = osp_declare_object_create, - .do_create = osp_object_create, - .do_declare_destroy = osp_declare_object_destroy, - .do_destroy = osp_object_destroy, + .do_declare_create = osp_declare_create, + .do_create = osp_create, + .do_declare_destroy = osp_declare_destroy, + .do_destroy = osp_destroy, .do_index_try = osp_index_try, }; +/** + * Implement OSP layer lu_object_operations::loo_object_init() interface. + * + * Initialize the object. + * + * If it is a remote MDT object, then call do_attr_get() to fetch + * the attribute from the peer. + * + * \param[in] env pointer to the thread context + * \param[in] o pointer to the OSP layer lu_object + * \param[in] conf unused + * + * \retval 0 for success + * \retval negative error number on failure + */ static int osp_object_init(const struct lu_env *env, struct lu_object *o, const struct lu_object_conf *conf) { @@ -1538,6 +2253,9 @@ static int osp_object_init(const struct lu_env *env, struct lu_object *o, spin_lock_init(&po->opo_lock); o->lo_header->loh_attr |= LOHA_REMOTE; + INIT_LIST_HEAD(&po->opo_xattr_list); + INIT_LIST_HEAD(&po->opo_invalidate_cb_list); + init_rwsem(&po->opo_invalidate_sem); if (is_ost_obj(o)) { po->opo_obj.do_ops = &osp_obj_ops; @@ -1546,48 +2264,78 @@ static int osp_object_init(const struct lu_env *env, struct lu_object *o, po->opo_obj.do_ops = &osp_md_obj_ops; po->opo_obj.do_body_ops = &osp_md_body_ops; - rc = po->opo_obj.do_ops->do_attr_get(env, lu2dt_obj(o), - la, NULL); - if (rc == 0) - o->lo_header->loh_attr |= - LOHA_EXISTS | (la->la_mode & S_IFMT); - if (rc == -ENOENT) { + + if (conf != NULL && conf->loc_flags & LOC_F_NEW) { po->opo_non_exist = 1; - rc = 0; + } else { + rc = po->opo_obj.do_ops->do_attr_get(env, lu2dt_obj(o), + la); + if (rc == 0) + o->lo_header->loh_attr |= + LOHA_EXISTS | (la->la_mode & S_IFMT); + if (rc == -ENOENT) { + po->opo_non_exist = 1; + rc = 0; + } } init_rwsem(&po->opo_sem); } RETURN(rc); } +static void osp_object_free_rcu(struct rcu_head *head) +{ + struct osp_object *obj = container_of(head, struct osp_object, + opo_header.loh_rcu); + + kmem_cache_free(osp_object_kmem, obj); +} + +/** + * Implement OSP layer lu_object_operations::loo_object_free() interface. + * + * Finalize the object. + * + * If the OSP object has attributes cache, then destroy the cache. + * Free the object finally. + * + * \param[in] env pointer to the thread context + * \param[in] o pointer to the OSP layer lu_object + */ static void osp_object_free(const struct lu_env *env, struct lu_object *o) { struct osp_object *obj = lu2osp_obj(o); struct lu_object_header *h = o->lo_header; + struct osp_xattr_entry *oxe; + struct osp_xattr_entry *tmp; + int count; dt_object_fini(&obj->opo_obj); lu_object_header_fini(h); - if (obj->opo_ooa != NULL) { - struct osp_xattr_entry *oxe; - struct osp_xattr_entry *tmp; - int count; - - list_for_each_entry_safe(oxe, tmp, - &obj->opo_ooa->ooa_xattr_list, - oxe_list) { - list_del(&oxe->oxe_list); - count = atomic_read(&oxe->oxe_ref); - LASSERTF(count == 1, - "Still has %d users on the xattr entry %.*s\n", - count-1, (int)oxe->oxe_namelen, oxe->oxe_buf); - - OBD_FREE(oxe, oxe->oxe_buflen); - } - OBD_FREE_PTR(obj->opo_ooa); + list_for_each_entry_safe(oxe, tmp, &obj->opo_xattr_list, oxe_list) { + list_del(&oxe->oxe_list); + count = atomic_read(&oxe->oxe_ref); + LASSERTF(count == 1, + "Still has %d users on the xattr entry %.*s\n", + count-1, (int)oxe->oxe_namelen, oxe->oxe_buf); + + OBD_FREE_LARGE(oxe, oxe->oxe_buflen); } - OBD_SLAB_FREE_PTR(obj, osp_object_kmem); + OBD_FREE_PRE(obj, sizeof(*obj), "slab-freed"); + call_rcu(&obj->opo_header.loh_rcu, osp_object_free_rcu); } +/** + * Implement OSP layer lu_object_operations::loo_object_release() interface. + * + * Cleanup (not free) the object. + * + * If it is a reserved object but failed to be created, or it is an OST + * object, then mark the object as non-cached. + * + * \param[in] env pointer to the thread context + * \param[in] o pointer to the OSP layer lu_object + */ static void osp_object_release(const struct lu_env *env, struct lu_object *o) { struct osp_object *po = lu2osp_obj(o); @@ -1606,6 +2354,14 @@ static void osp_object_release(const struct lu_env *env, struct lu_object *o) d->opd_pre_reserved--; spin_unlock(&d->opd_pre_lock); + /* + * Check that osp_precreate_cleanup_orphans is not blocked + * due to opd_pre_reserved > 0. + */ + if (unlikely(d->opd_pre_reserved == 0 && + (d->opd_pre_recovering || d->opd_pre_status))) + wake_up(&d->opd_pre_waitq); + /* not needed in cache any more */ set_bit(LU_OBJECT_HEARD_BANSHEE, &o->lo_header->loh_flags); }