X-Git-Url: https://git.whamcloud.com/?a=blobdiff_plain;f=lustre%2Fosp%2Fosp_precreate.c;h=cac5bf2bf350aba96878874fd401883c1b5857fc;hb=acc918d76856ff14306c543c74e6ceef3865bcbc;hp=9b3febcedfe620645e1738a36fdb67c981200272;hpb=689a3a288ee6b06d13541c4b6cd99ea146c5ad22;p=fs%2Flustre-release.git diff --git a/lustre/osp/osp_precreate.c b/lustre/osp/osp_precreate.c index 9b3febc..cac5bf2 100644 --- a/lustre/osp/osp_precreate.c +++ b/lustre/osp/osp_precreate.c @@ -27,7 +27,7 @@ * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved. * Use is subject to license terms. * - * Copyright (c) 2012, Intel Corporation. + * Copyright (c) 2012, 2014, Intel Corporation. */ /* * This file is part of Lustre, http://www.lustre.org/ @@ -42,11 +42,9 @@ * Author: Di Wang */ -#ifndef EXPORT_SYMTAB -# define EXPORT_SYMTAB -#endif #define DEBUG_SUBSYSTEM S_MDS +#include #include "osp_internal.h" /* @@ -60,7 +58,15 @@ */ /* - * statfs + ** + * Check whether statfs data is expired + * + * OSP device caches statfs data for the target, the function checks + * whether the data is expired or not. + * + * \param[in] d OSP device + * + * \retval 0 - not expired, 1 - expired */ static inline int osp_statfs_need_update(struct osp_device *d) { @@ -68,14 +74,47 @@ static inline int osp_statfs_need_update(struct osp_device *d) d->opd_statfs_fresh_till); } +/* + * OSP tries to maintain pool of available objects so that calls to create + * objects don't block most of time + * + * each time OSP gets connected to OST, we should start from precreation cleanup + */ +static inline bool osp_precreate_running(struct osp_device *d) +{ + return !!(d->opd_pre_thread.t_flags & SVC_RUNNING); +} + +static inline bool osp_precreate_stopped(struct osp_device *d) +{ + return !!(d->opd_pre_thread.t_flags & SVC_STOPPED); +} + static void osp_statfs_timer_cb(unsigned long _d) { struct osp_device *d = (struct osp_device *) _d; LASSERT(d); - cfs_waitq_signal(&d->opd_pre_waitq); + if (d->opd_pre != NULL && osp_precreate_running(d)) + wake_up(&d->opd_pre_waitq); } +/** + * RPC interpret callback for OST_STATFS RPC + * + * An interpretation callback called by ptlrpc for OST_STATFS RPC when it is + * replied by the target. It's used to maintain statfs cache for the target. + * The function fills data from the reply if successful and schedules another + * update. + * + * \param[in] env LU environment provided by the caller + * \param[in] req RPC replied + * \param[in] aa callback data + * \param[in] rc RPC result + * + * \retval 0 on success + * \retval negative negated errno on error + */ static int osp_statfs_interpret(const struct lu_env *env, struct ptlrpc_request *req, union ptlrpc_async_args *aa, int rc) @@ -111,13 +150,25 @@ static int osp_statfs_interpret(const struct lu_env *env, RETURN(0); out: /* couldn't update statfs, try again as soon as possible */ - cfs_waitq_signal(&d->opd_pre_waitq); + if (d->opd_pre != NULL && osp_precreate_running(d)) + wake_up(&d->opd_pre_waitq); + if (req->rq_import_generation == imp->imp_generation) CDEBUG(D_CACHE, "%s: couldn't update statfs: rc = %d\n", d->opd_obd->obd_name, rc); RETURN(rc); } +/** + * Send OST_STATFS RPC + * + * Sends OST_STATFS RPC to refresh cached statfs data for the target. + * Also disables scheduled updates as times OSP may need to refresh + * statfs data before expiration. The function doesn't block, instead + * an interpretation callback osp_statfs_interpret() is used. + * + * \param[in] d OSP device + */ static int osp_statfs_update(struct osp_device *d) { struct ptlrpc_request *req; @@ -156,16 +207,23 @@ static int osp_statfs_update(struct osp_device *d) d->opd_statfs_fresh_till = cfs_time_shift(obd_timeout * 1000); d->opd_statfs_update_in_progress = 1; - ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1); + ptlrpcd_add_req(req); RETURN(0); } -/* - * XXX: there might be a case where removed object(s) do not add free - * space (empty object). if the number of such deletions is high, then - * we can start to update statfs too often - a rpc storm - * TODO: some throttling is needed +/** + * Schedule an immediate update for statfs data + * + * If cached statfs data claim no free space, but OSP has got a request to + * destroy an object (so release some space probably), then we may need to + * refresh cached statfs data sooner than planned. The function checks there + * is no statfs update going and schedules immediate update if so. + * XXX: there might be a case where removed object(s) do not add free space (empty + * object). If the number of such deletions is high, then we can start to update + * statfs too often causing a RPC storm. some throttling is needed... + * + * \param[in] d OSP device where statfs data needs to be refreshed */ void osp_statfs_need_now(struct osp_device *d) { @@ -177,51 +235,41 @@ void osp_statfs_need_now(struct osp_device *d) */ d->opd_statfs_fresh_till = cfs_time_shift(-1); cfs_timer_disarm(&d->opd_statfs_timer); - cfs_waitq_signal(&d->opd_pre_waitq); + wake_up(&d->opd_pre_waitq); } } - -/* - * OSP tries to maintain pool of available objects so that calls to create - * objects don't block most of time +/** + * Return number of precreated objects * - * each time OSP gets connected to OST, we should start from precreation cleanup + * A simple helper to calculate the number of precreated objects on the device. + * + * \param[in] env LU environment provided by the caller + * \param[in] osp OSP device + * + * \retval the number of the precreated objects */ -static inline int osp_precreate_running(struct osp_device *d) -{ - return !!(d->opd_pre_thread.t_flags & SVC_RUNNING); -} - -static inline int osp_precreate_stopped(struct osp_device *d) -{ - return !!(d->opd_pre_thread.t_flags & SVC_STOPPED); -} - static inline int osp_objs_precreated(const struct lu_env *env, struct osp_device *osp) { - struct lu_fid *fid1 = &osp->opd_pre_last_created_fid; - struct lu_fid *fid2 = &osp->opd_pre_used_fid; - - LASSERTF(fid_seq(fid1) == fid_seq(fid2), - "Created fid"DFID" Next fid "DFID"\n", PFID(fid1), PFID(fid2)); - - if (fid_is_idif(fid1)) { - struct ost_id *oi1 = &osp_env_info(env)->osi_oi; - struct ost_id *oi2 = &osp_env_info(env)->osi_oi2; - - LASSERT(fid_is_idif(fid1) && fid_is_idif(fid2)); - ostid_idif_pack(fid1, oi1); - ostid_idif_pack(fid2, oi2); - LASSERT(oi1->oi_id >= oi2->oi_id); - - return oi1->oi_id - oi2->oi_id; - } - - return fid_oid(fid1) - fid_oid(fid2); + return osp_fid_diff(&osp->opd_pre_last_created_fid, + &osp->opd_pre_used_fid); } +/** + * Check pool of precreated objects is nearly empty + * + * We should not wait till the pool of the precreated objects is exhausted, + * because then there will be a long period of OSP being unavailable for the + * new creations due to lenghty precreate RPC. Instead we ask for another + * precreation ahead and hopefully have it ready before the current pool is + * empty. Notice this function relies on an external locking. + * + * \param[in] env LU environment provided by the caller + * \param[in] d OSP device + * + * \retval 0 - current pool is good enough, 1 - time to precreate + */ static inline int osp_precreate_near_empty_nolock(const struct lu_env *env, struct osp_device *d) { @@ -233,6 +281,17 @@ static inline int osp_precreate_near_empty_nolock(const struct lu_env *env, (d->opd_pre_status == 0)); } +/** + * Check pool of precreated objects + * + * This is protected version of osp_precreate_near_empty_nolock(), check that + * for the details. + * + * \param[in] env LU environment provided by the caller + * \param[in] d OSP device + * + * \retval 0 - current pool is good enough, 1 - time to precreate + */ static inline int osp_precreate_near_empty(const struct lu_env *env, struct osp_device *d) { @@ -245,6 +304,18 @@ static inline int osp_precreate_near_empty(const struct lu_env *env, return rc; } +/** + * Check given sequence is empty + * + * Returns a binary result whether the given sequence has some IDs left + * or not. Find the details in osp_fid_end_seq(). This is a lock protected + * version of that function. + * + * \param[in] env LU environment provided by the caller + * \param[in] osp OSP device + * + * \retval 0 - current sequence has no IDs, 1 - otherwise + */ static inline int osp_create_end_seq(const struct lu_env *env, struct osp_device *osp) { @@ -258,7 +329,19 @@ static inline int osp_create_end_seq(const struct lu_env *env, } /** - * Write fid into last_oid/last_seq file. + * Write FID into into last_oid/last_seq file + * + * The function stores the sequence and the in-sequence id into two dedicated + * files. The sync argument can be used to request synchronous commit, so the + * function won't return until the updates are committed. + * + * \param[in] env LU environment provided by the caller + * \param[in] osp OSP device + * \param[in] fid fid where sequence/id is taken + * \param[in] sync update mode: 0 - asynchronously, 1 - synchronously + * + * \retval 0 on success + * \retval negative negated errno on error **/ int osp_write_last_oid_seq_files(struct lu_env *env, struct osp_device *osp, struct lu_fid *fid, int sync) @@ -272,16 +355,15 @@ int osp_write_last_oid_seq_files(struct lu_env *env, struct osp_device *osp, int rc; ENTRY; - /* Note: through f_oid is only 32bits, it will also write - * 64 bits for oid to keep compatiblity with the previous - * version. */ + /* Note: through f_oid is only 32 bits, it will also write 64 bits + * for oid to keep compatibility with the previous version. */ lb_oid->lb_buf = &fid->f_oid; - lb_oid->lb_len = sizeof(obd_id); - oid_off = sizeof(obd_id) * osp->opd_index; + lb_oid->lb_len = sizeof(u64); + oid_off = sizeof(u64) * osp->opd_index; lb_oseq->lb_buf = &fid->f_seq; - lb_oseq->lb_len = sizeof(obd_id); - oseq_off = sizeof(obd_id) * osp->opd_index; + lb_oseq->lb_len = sizeof(u64); + oseq_off = sizeof(u64) * osp->opd_index; th = dt_trans_create(env, osp->opd_storage); if (IS_ERR(th)) @@ -289,12 +371,12 @@ int osp_write_last_oid_seq_files(struct lu_env *env, struct osp_device *osp, th->th_sync |= sync; rc = dt_declare_record_write(env, osp->opd_last_used_oid_file, - lb_oid->lb_len, oid_off, th); + lb_oid, oid_off, th); if (rc != 0) GOTO(out, rc); rc = dt_declare_record_write(env, osp->opd_last_used_seq_file, - lb_oseq->lb_len, oseq_off, th); + lb_oseq, oseq_off, th); if (rc != 0) GOTO(out, rc); @@ -321,7 +403,27 @@ out: RETURN(rc); } -int osp_precreate_rollover_new_seq(struct lu_env *env, struct osp_device *osp) +/** + * Switch to another sequence + * + * When a current sequence has no available IDs left, OSP has to switch to + * another new sequence. OSP requests it using the regular FLDB protocol + * and stores synchronously before that is used in precreated. This is needed + * to basically have the sequences referenced (not orphaned), otherwise it's + * possible that OST has some objects precreated and the clients have data + * written to it, but after MDT failover nobody refers those objects and OSP + * has no idea that the sequence need cleanup to be done. + * While this is very expensive operation, it's supposed to happen very very + * infrequently because sequence has 2^32 or 2^48 objects (depending on type) + * + * \param[in] env LU environment provided by the caller + * \param[in] osp OSP device + * + * \retval 0 on success + * \retval negative negated errno on error + */ +static int osp_precreate_rollover_new_seq(struct lu_env *env, + struct osp_device *osp) { struct lu_fid *fid = &osp_env_info(env)->osi_fid; struct lu_fid *last_fid = &osp->opd_last_used_fid; @@ -363,11 +465,23 @@ int osp_precreate_rollover_new_seq(struct lu_env *env, struct osp_device *osp) } /** - * alloc fids for precreation. - * rc = 0 Success, @grow is the count of real allocation. - * rc = 1 Current seq is used up. - * rc < 0 Other error. - **/ + * Find IDs available in current sequence + * + * The function calculates the highest possible ID and the number of IDs + * available in the current sequence OSP is using. The number is limited + * artifically by the caller (grow param) and the number of IDs available + * in the sequence by nature. The function doesn't require an external + * locking. + * + * \param[in] env LU environment provided by the caller + * \param[in] osp OSP device + * \param[in] fid FID the caller wants to start with + * \param[in] grow how many the caller wants + * \param[out] fid the highest calculated FID + * \param[out] grow the number of available IDs calculated + * + * \retval 0 on success, 1 - the sequence is empty + */ static int osp_precreate_fids(const struct lu_env *env, struct osp_device *osp, struct lu_fid *fid, int *grow) { @@ -381,16 +495,16 @@ static int osp_precreate_fids(const struct lu_env *env, struct osp_device *osp, spin_lock(&osp->opd_pre_lock); last_fid = &osp->opd_pre_last_created_fid; - ostid_idif_pack(last_fid, oi); - end = min(oi->oi_id + *grow, IDIF_MAX_OID); - *grow = end - oi->oi_id; - oi->oi_id += *grow; + fid_to_ostid(last_fid, oi); + end = min(ostid_id(oi) + *grow, IDIF_MAX_OID); + *grow = end - ostid_id(oi); + ostid_set_id(oi, ostid_id(oi) + *grow); spin_unlock(&osp->opd_pre_lock); if (*grow == 0) return 1; - ostid_idif_unpack(oi, fid, osp->opd_index); + ostid_to_fid(fid, oi, osp->opd_index); return 0; } @@ -408,6 +522,21 @@ static int osp_precreate_fids(const struct lu_env *env, struct osp_device *osp, return *grow > 0 ? 0 : 1; } +/** + * Prepare and send precreate RPC + * + * The function finds how many objects should be precreated. Then allocates, + * prepares and schedules precreate RPC synchronously. Upon reply the function + * wake ups the threads waiting for the new objects on this target. If the + * target wasn't able to create all the objects requested, then the next + * precreate will be asking less objects (i.e. slow precreate down). + * + * \param[in] env LU environment provided by the caller + * \param[in] d OSP device + * + * \retval 0 on success + * \retval negative negated errno on error + **/ static int osp_precreate_send(const struct lu_env *env, struct osp_device *d) { struct osp_thread_info *oti = osp_env_info(env); @@ -462,7 +591,7 @@ static int osp_precreate_send(const struct lu_env *env, struct osp_device *d) osp_pre_update_status(d, -ENOSPC); rc = -ENOSPC; } - cfs_waitq_signal(&d->opd_pre_waitq); + wake_up(&d->opd_pre_waitq); GOTO(out_req, rc); } @@ -473,7 +602,7 @@ static int osp_precreate_send(const struct lu_env *env, struct osp_device *d) fid->f_seq = 0; } - fid_ostid_pack(fid, &body->oa.o_oi); + fid_to_ostid(fid, &body->oa.o_oi); body->oa.o_valid = OBD_MD_FLGROUP; ptlrpc_request_set_replen(req); @@ -490,15 +619,15 @@ static int osp_precreate_send(const struct lu_env *env, struct osp_device *d) if (body == NULL) GOTO(out_req, rc = -EPROTO); - fid_ostid_unpack(fid, &body->oa.o_oi, d->opd_index); - LASSERTF(lu_fid_diff(fid, &d->opd_pre_used_fid) > 0, - "reply fid "DFID" pre used fid "DFID"\n", PFID(fid), - PFID(&d->opd_pre_used_fid)); - - CDEBUG(D_HA, "%s: new last_created "DFID"\n", d->opd_obd->obd_name, - PFID(fid)); + ostid_to_fid(fid, &body->oa.o_oi, d->opd_index); + if (osp_fid_diff(fid, &d->opd_pre_used_fid) <= 0) { + CERROR("%s: precreate fid "DFID" < local used fid "DFID + ": rc = %d\n", d->opd_obd->obd_name, + PFID(fid), PFID(&d->opd_pre_used_fid), -ESTALE); + GOTO(out_req, rc = -ESTALE); + } - diff = lu_fid_diff(fid, &d->opd_pre_last_created_fid); + diff = osp_fid_diff(fid, &d->opd_pre_last_created_fid); spin_lock(&d->opd_pre_lock); if (diff < grow) { @@ -513,27 +642,44 @@ static int osp_precreate_send(const struct lu_env *env, struct osp_device *d) d->opd_pre_grow_slow = 0; } + body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY); + fid_to_ostid(fid, &body->oa.o_oi); + d->opd_pre_last_created_fid = *fid; spin_unlock(&d->opd_pre_lock); - CDEBUG(D_OTHER, "current precreated pool: "DFID"-"DFID"\n", - PFID(&d->opd_pre_used_fid), PFID(&d->opd_pre_last_created_fid)); + CDEBUG(D_HA, "%s: current precreated pool: "DFID"-"DFID"\n", + d->opd_obd->obd_name, PFID(&d->opd_pre_used_fid), + PFID(&d->opd_pre_last_created_fid)); out_req: /* now we can wakeup all users awaiting for objects */ osp_pre_update_status(d, rc); - cfs_waitq_signal(&d->opd_pre_user_waitq); + wake_up(&d->opd_pre_user_waitq); ptlrpc_req_finished(req); RETURN(rc); } +/** + * Get last precreated object from target (OST) + * + * Sends synchronous RPC to the target (OST) to learn the last precreated + * object. This later is used to remove all unused objects (cleanup orphan + * procedure). Also, the next object after one we got will be used as a + * starting point for the new precreates. + * + * \param[in] env LU environment provided by the caller + * \param[in] d OSP device + * + * \retval 0 on success + * \retval negative negated errno on error + **/ static int osp_get_lastfid_from_ost(const struct lu_env *env, struct osp_device *d) { - struct ost_id *oi = &osp_env_info(env)->osi_oi; struct ptlrpc_request *req = NULL; struct obd_import *imp; - struct lu_fid *last_fid = &d->opd_last_used_fid; + struct lu_fid *last_fid; char *tmp; int rc; ENTRY; @@ -545,26 +691,22 @@ static int osp_get_lastfid_from_ost(const struct lu_env *env, if (req == NULL) RETURN(-ENOMEM); - req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_KEY, RCL_CLIENT, + req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_KEY, RCL_CLIENT, sizeof(KEY_LAST_FID)); - req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_VAL, RCL_CLIENT, - sizeof(*oi)); - rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GET_INFO); if (rc) { ptlrpc_request_free(req); RETURN(rc); } - tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY); + tmp = req_capsule_client_get(&req->rq_pill, &RMF_GETINFO_KEY); memcpy(tmp, KEY_LAST_FID, sizeof(KEY_LAST_FID)); req->rq_no_delay = req->rq_no_resend = 1; - fid_ostid_pack(last_fid, oi); - ostid_cpu_to_le(oi, oi); - tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_VAL); - memcpy(tmp, oi, sizeof(*oi)); + last_fid = req_capsule_client_get(&req->rq_pill, &RMF_FID); + fid_cpu_to_le(last_fid, &d->opd_last_used_fid); + ptlrpc_request_set_replen(req); rc = ptlrpc_queue_wait(req); @@ -578,14 +720,13 @@ static int osp_get_lastfid_from_ost(const struct lu_env *env, GOTO(out, rc); } - oi = req_capsule_server_get(&req->rq_pill, &RMF_OST_ID); - if (oi == NULL) { + last_fid = req_capsule_server_get(&req->rq_pill, &RMF_FID); + if (last_fid == NULL) { CERROR("%s: Got last_fid failed.\n", d->opd_obd->obd_name); GOTO(out, rc = -EPROTO); } - rc = fid_ostid_unpack(last_fid, oi, d->opd_index); - if (rc != 0 || !fid_is_sane(last_fid)) { + if (!fid_is_sane(last_fid)) { CERROR("%s: Got insane last_fid "DFID"\n", d->opd_obd->obd_name, PFID(last_fid)); GOTO(out, rc = -EPROTO); @@ -605,8 +746,21 @@ out: } /** - * asks OST to clean precreate orphans - * and gets next id for new objects + * Cleanup orphans on OST + * + * This function is called in a contex of a dedicated thread handling + * all the precreation suff. The function waits till local recovery + * is complete, then identify all the unreferenced objects (orphans) + * using the highest ID referenced by a local and the highest object + * precreated by the target. The found range is a subject to removal + * using specially flagged RPC. During this process OSP is marked + * unavailable for new objects. + * + * \param[in] env LU environment provided by the caller + * \param[in] d OSP device + * + * \retval 0 on success + * \retval negative negated errno on error */ static int osp_precreate_cleanup_orphans(struct lu_env *env, struct osp_device *d) @@ -682,7 +836,7 @@ static int osp_precreate_cleanup_orphans(struct lu_env *env, body->oa.o_flags = OBD_FL_DELORPHAN; body->oa.o_valid = OBD_MD_FLFLAGS | OBD_MD_FLGROUP; - fid_ostid_pack(&d->opd_last_used_fid, &body->oa.o_oi); + fid_to_ostid(&d->opd_last_used_fid, &body->oa.o_oi); ptlrpc_request_set_replen(req); @@ -702,13 +856,10 @@ static int osp_precreate_cleanup_orphans(struct lu_env *env, /* * OST provides us with id new pool starts from in body->oa.o_id */ - fid_ostid_unpack(last_fid, &body->oa.o_oi, d->opd_index); - CDEBUG(D_INFO, "%s: last_fid "DFID" server last fid "DFID"\n", - d->opd_obd->obd_name, PFID(&d->opd_last_used_fid), - PFID(last_fid)); + ostid_to_fid(last_fid, &body->oa.o_oi, d->opd_index); spin_lock(&d->opd_pre_lock); - diff = lu_fid_diff(&d->opd_last_used_fid, last_fid); + diff = osp_fid_diff(&d->opd_last_used_fid, last_fid); if (diff > 0) { d->opd_pre_grow_count = OST_MIN_PRECREATE + diff; d->opd_pre_last_created_fid = d->opd_last_used_fid; @@ -733,7 +884,9 @@ out: if (req) ptlrpc_req_finished(req); + spin_lock(&d->opd_pre_lock); d->opd_pre_recovering = 0; + spin_unlock(&d->opd_pre_lock); /* * If rc is zero, the pre-creation window should have been emptied. @@ -751,20 +904,43 @@ out: * this OSP isn't quite functional yet */ osp_pre_update_status(d, rc); } else { - cfs_waitq_signal(&d->opd_pre_user_waitq); + wake_up(&d->opd_pre_user_waitq); } } RETURN(rc); } -/* +/** + * Update precreate status using statfs data + * + * The function decides whether this OSP should be used for new objects. + * IOW, whether this OST is used up or has some free space. Cached statfs + * data is used to make this decision. If the latest result of statfs + * request (rc argument) is not success, then just mark OSP unavailable + * right away. + + * Add a bit of hysteresis so this flag isn't continually flapping, + * and ensure that new files don't get extremely fragmented due to + * only a small amount of available space in the filesystem. + * We want to set the NOSPC flag when there is less than ~0.1% free + * and clear it when there is at least ~0.2% free space, so: + * avail < ~0.1% max max = avail + used + * 1025 * avail < avail + used used = blocks - free + * 1024 * avail < used + * 1024 * avail < blocks - free + * avail < ((blocks - free) >> 10) + * + * On very large disk, say 16TB 0.1% will be 16 GB. We don't want to + * lose that amount of space so in those cases we report no space left + * if their is less than 1 GB left. * the function updates current precreation status used: functional or not * - * rc is a last code from the transport, rc == 0 meaning transport works - * well and users of lod can use objects from this OSP + * \param[in] d OSP device + * \param[in] rc new precreate status for device \a d * - * the status depends on current usage of OST + * \retval 0 on success + * \retval negative negated errno on error */ void osp_pre_update_status(struct osp_device *d, int rc) { @@ -776,20 +952,6 @@ void osp_pre_update_status(struct osp_device *d, int rc) if (rc) goto out; - /* Add a bit of hysteresis so this flag isn't continually flapping, - * and ensure that new files don't get extremely fragmented due to - * only a small amount of available space in the filesystem. - * We want to set the NOSPC flag when there is less than ~0.1% free - * and clear it when there is at least ~0.2% free space, so: - * avail < ~0.1% max max = avail + used - * 1025 * avail < avail + used used = blocks - free - * 1024 * avail < used - * 1024 * avail < blocks - free - * avail < ((blocks - free) >> 10) - * - * On very large disk, say 16TB 0.1% will be 16 GB. We don't want to - * lose that amount of space so in those cases we report no space left - * if their is less than 1 GB left. */ if (likely(msfs->os_type)) { used = min_t(__u64, (msfs->os_blocks - msfs->os_bfree) >> 10, 1 << 30); @@ -807,9 +969,11 @@ void osp_pre_update_status(struct osp_device *d, int rc) d->opd_syn_changes, d->opd_syn_rpc_in_progress); } else if (old == -ENOSPC) { d->opd_pre_status = 0; + spin_lock(&d->opd_pre_lock); d->opd_pre_grow_slow = 0; d->opd_pre_grow_count = OST_MIN_PRECREATE; - cfs_waitq_signal(&d->opd_pre_waitq); + spin_unlock(&d->opd_pre_lock); + wake_up(&d->opd_pre_waitq); CDEBUG(D_INFO, "%s: no space: "LPU64" blocks, "LPU64 " free, "LPU64" used, "LPU64" avail -> %d: " "rc = %d\n", d->opd_obd->obd_name, @@ -819,10 +983,26 @@ void osp_pre_update_status(struct osp_device *d, int rc) } out: - cfs_waitq_signal(&d->opd_pre_user_waitq); + wake_up(&d->opd_pre_user_waitq); } -static int osp_init_pre_fid(struct osp_device *osp) +/** + * Initialize FID for precreation + * + * For a just created new target, a new sequence should be taken. + * The function checks there is no IDIF in use (if the target was + * added with the older version of Lustre), then requests a new + * sequence from FLDB using the regular protocol. Then this new + * sequence is stored on a persisten storage synchronously to prevent + * possible object leakage (for the detail see the description for + * osp_precreate_rollover_new_seq()). + * + * \param[in] osp OSP device + * + * \retval 0 on success + * \retval negative negated errno on error + */ +int osp_init_pre_fid(struct osp_device *osp) { struct lu_env env; struct osp_thread_info *osi; @@ -831,6 +1011,8 @@ static int osp_init_pre_fid(struct osp_device *osp) int rc; ENTRY; + LASSERT(osp->opd_pre != NULL); + /* Return if last_used fid has been initialized */ if (!fid_is_zero(&osp->opd_last_used_fid)) RETURN(0); @@ -876,20 +1058,34 @@ out: RETURN(rc); } +/** + * The core of precreate functionality + * + * The function implements the main precreation loop. Basically it + * involves connecting to the target, precerate FID initialization, + * identifying and removing orphans, then serving precreation. As + * part of the latter, the thread is responsible for statfs data + * updates. The precreation is mostly driven by another threads + * asking for new OST objects - those askers wake the thread when + * the number of precreated objects reach low watermark. + * After a disconnect, the sequence above repeats. This is keep going + * until the thread is requested to stop. + * + * \param[in] _arg private data the thread (OSP device to handle) + * + * \retval 0 on success + * \retval negative negated errno on error + */ static int osp_precreate_thread(void *_arg) { struct osp_device *d = _arg; struct ptlrpc_thread *thread = &d->opd_pre_thread; struct l_wait_info lwi = { 0 }; - char pname[16]; struct lu_env env; int rc; ENTRY; - sprintf(pname, "osp-pre-%u", d->opd_index); - cfs_daemonize(pname); - rc = lu_env_init(&env, d->opd_dt_dev.dd_lu_dev.ld_type->ldt_ctx_tags); if (rc) { CERROR("%s: init env error: rc = %d\n", d->opd_obd->obd_name, @@ -900,7 +1096,7 @@ static int osp_precreate_thread(void *_arg) spin_lock(&d->opd_pre_lock); thread->t_flags = SVC_RUNNING; spin_unlock(&d->opd_pre_lock); - cfs_waitq_signal(&thread->t_ctl_waitq); + wake_up(&thread->t_ctl_waitq); while (osp_precreate_running(d)) { /* @@ -924,19 +1120,18 @@ static int osp_precreate_thread(void *_arg) break; LASSERT(d->opd_obd->u.cli.cl_seq != NULL); - if (d->opd_obd->u.cli.cl_seq->lcs_exp == NULL) { - /* Get new sequence for client first */ - LASSERT(d->opd_exp != NULL); - d->opd_obd->u.cli.cl_seq->lcs_exp = - class_export_get(d->opd_exp); - rc = osp_init_pre_fid(d); - if (rc != 0) { - class_export_put(d->opd_exp); - d->opd_obd->u.cli.cl_seq->lcs_exp = NULL; - CERROR("%s: init pre fid error: rc = %d\n", - d->opd_obd->obd_name, rc); - continue; - } + /* Sigh, fid client is not ready yet */ + if (d->opd_obd->u.cli.cl_seq->lcs_exp == NULL) + continue; + + /* Init fid for osp_precreate if necessary */ + rc = osp_init_pre_fid(d); + if (rc != 0) { + class_export_put(d->opd_exp); + d->opd_obd->u.cli.cl_seq->lcs_exp = NULL; + CERROR("%s: init pre fid error: rc = %d\n", + d->opd_obd->obd_name, rc); + continue; } osp_statfs_update(d); @@ -1002,11 +1197,25 @@ static int osp_precreate_thread(void *_arg) thread->t_flags = SVC_STOPPED; lu_env_fini(&env); - cfs_waitq_signal(&thread->t_ctl_waitq); + wake_up(&thread->t_ctl_waitq); RETURN(0); } +/** + * Check when to stop to wait for precreate objects. + * + * The caller wanting a new OST object can't wait undefinitely. The + * function checks for few conditions including available new OST + * objects, disconnected OST, lack of space with no pending destroys, + * etc. IOW, it checks whether the current OSP state is good to keep + * waiting or it's better to give up. + * + * \param[in] env LU environment provided by the caller + * \param[in] d OSP device + * + * \retval 0 - keep waiting, 1 - no luck + */ static int osp_precreate_ready_condition(const struct lu_env *env, struct osp_device *d) { @@ -1018,11 +1227,23 @@ static int osp_precreate_ready_condition(const struct lu_env *env, if (d->opd_pre_reserved + 1 < osp_objs_precreated(env, d)) return 1; - /* ready if OST reported no space and no destoys in progress */ + /* ready if OST reported no space and no destroys in progress */ if (d->opd_syn_changes + d->opd_syn_rpc_in_progress == 0 && d->opd_pre_status == -ENOSPC) return 1; + /* Bail out I/O fails to OST */ + if (d->opd_pre_status != 0 && + d->opd_pre_status != -EAGAIN && + d->opd_pre_status != -ENODEV && + d->opd_pre_status != -ENOSPC) { + /* DEBUG LU-3230 */ + if (d->opd_pre_status != -EIO) + CERROR("%s: precreate failed opd_pre_status %d\n", + d->opd_obd->obd_name, d->opd_pre_status); + return 1; + } + return 0; } @@ -1030,30 +1251,46 @@ static int osp_precreate_timeout_condition(void *data) { struct osp_device *d = data; - LCONSOLE_WARN("%s: slow creates, last="DFID", next="DFID", " - "reserved="LPU64", syn_changes=%lu, " - "syn_rpc_in_progress=%d, status=%d\n", - d->opd_obd->obd_name, PFID(&d->opd_pre_last_created_fid), - PFID(&d->opd_pre_used_fid), d->opd_pre_reserved, - d->opd_syn_changes, d->opd_syn_rpc_in_progress, - d->opd_pre_status); + CDEBUG(D_HA, "%s: slow creates, last="DFID", next="DFID", " + "reserved="LPU64", syn_changes=%lu, " + "syn_rpc_in_progress=%d, status=%d\n", + d->opd_obd->obd_name, PFID(&d->opd_pre_last_created_fid), + PFID(&d->opd_pre_used_fid), d->opd_pre_reserved, + d->opd_syn_changes, d->opd_syn_rpc_in_progress, + d->opd_pre_status); return 1; } -/* - * called to reserve object in the pool - * return codes: - * ENOSPC - no space on corresponded OST - * EAGAIN - precreation is in progress, try later - * EIO - no access to OST +/** + * Reserve object in precreate pool + * + * When the caller wants to create a new object on this target (target + * represented by the given OSP), it should declare this intention using + * a regular ->dt_declare_create() OSD API method. Then OSP will be trying + * to reserve an object in the existing precreated pool or wait up to + * obd_timeout for the available object to appear in the pool (a dedicated + * thread will be doing real precreation in background). The object can be + * consumed later with osp_precreate_get_fid() or be released with call to + * lu_object_put(). Notice the function doesn't reserve a specific ID, just + * some ID. The actual ID assignment happen in osp_precreate_get_fid(). + * If the space on the target is short and there is a pending object destroy, + * then the function forces local commit to speedup space release (see + * osp_sync.c for the details). + * + * \param[in] env LU environment provided by the caller + * \param[in] d OSP device + * + * \retval 0 on success + * \retval -ENOSPC when no space on OST + * \retval -EAGAIN try later, slow precreation in progress + * \retval -EIO when no access to OST */ int osp_precreate_reserve(const struct lu_env *env, struct osp_device *d) { struct l_wait_info lwi; cfs_time_t expire = cfs_time_shift(obd_timeout); int precreated, rc; - int count = 0; ENTRY; @@ -1069,18 +1306,7 @@ int osp_precreate_reserve(const struct lu_env *env, struct osp_device *d) * - OST can allocate fid sequence. */ while ((rc = d->opd_pre_status) == 0 || rc == -ENOSPC || - rc == -ENODEV || rc == -EAGAIN) { - -#if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 3, 90, 0) - /* - * to address Andreas's concern on possible busy-loop - * between this thread and osp_precreate_send() - */ - if (unlikely(count++ == 1000)) { - osp_precreate_timeout_condition(d); - LBUG(); - } -#endif + rc == -ENODEV || rc == -EAGAIN || rc == -ENOTCONN) { /* * increase number of precreations @@ -1106,7 +1332,7 @@ int osp_precreate_reserve(const struct lu_env *env, struct osp_device *d) /* XXX: don't wake up if precreation is in progress */ if (osp_precreate_near_empty_nolock(env, d) && !osp_precreate_end_seq_nolock(env, d)) - cfs_waitq_signal(&d->opd_pre_waitq); + wake_up(&d->opd_pre_waitq); break; } @@ -1135,7 +1361,7 @@ int osp_precreate_reserve(const struct lu_env *env, struct osp_device *d) } /* XXX: don't wake up if precreation is in progress */ - cfs_waitq_signal(&d->opd_pre_waitq); + wake_up(&d->opd_pre_waitq); lwi = LWI_TIMEOUT(expire - cfs_time_current(), osp_precreate_timeout_condition, d); @@ -1151,8 +1377,27 @@ int osp_precreate_reserve(const struct lu_env *env, struct osp_device *d) RETURN(rc); } -/* - * this function relies on reservation made before +/** + * Get a FID from precreation pool + * + * The function is a companion for osp_precreate_reserve() - it assigns + * a specific FID from the precreate. The function should be called only + * if the call to osp_precreate_reserve() was successful. The function + * updates a local storage to remember the highest object ID referenced + * by the node in the given sequence. + * + * A very importan details: this is supposed to be called once the + * transaction is started, so on-disk update will be atomic with the + * data (like LOVEA) refering this object. Then the object won't be leaked: + * either it's referenced by the committed transaction or it's a subject + * to the orphan cleanup procedure. + * + * \param[in] env LU environment provided by the caller + * \param[in] d OSP device + * \param[out] fid generated FID + * + * \retval 0 on success + * \retval negative negated errno on error */ int osp_precreate_get_fid(const struct lu_env *env, struct osp_device *d, struct lu_fid *fid) @@ -1160,7 +1405,7 @@ int osp_precreate_get_fid(const struct lu_env *env, struct osp_device *d, /* grab next id from the pool */ spin_lock(&d->opd_pre_lock); - LASSERTF(lu_fid_diff(&d->opd_pre_used_fid, + LASSERTF(osp_fid_diff(&d->opd_pre_used_fid, &d->opd_pre_last_created_fid) < 0, "next fid "DFID" last created fid "DFID"\n", PFID(&d->opd_pre_used_fid), @@ -1182,13 +1427,26 @@ int osp_precreate_get_fid(const struct lu_env *env, struct osp_device *d, * osp_precreate_thread() just before orphan cleanup */ if (unlikely(d->opd_pre_reserved == 0 && d->opd_pre_status)) - cfs_waitq_signal(&d->opd_pre_waitq); + wake_up(&d->opd_pre_waitq); return 0; } /* + * Set size regular attribute on an object + * + * When a striping is created late, it's possible that size is already + * initialized on the file. Then the new striping should inherit size + * from the file. The function sets size on the object using the regular + * protocol (OST_PUNCH). + * XXX: should be re-implemented using OUT ? * + * \param[in] env LU environment provided by the caller + * \param[in] dt object + * \param[in] size size to set. + * + * \retval 0 on success + * \retval negative negated errno on error */ int osp_object_truncate(const struct lu_env *env, struct dt_object *dt, __u64 size) @@ -1231,7 +1489,7 @@ int osp_object_truncate(const struct lu_env *env, struct dt_object *dt, if (oa == NULL) GOTO(out, rc = -ENOMEM); - rc = fid_ostid_pack(lu_object_fid(&dt->do_lu), &oa->o_oi); + rc = fid_to_ostid(lu_object_fid(&dt->do_lu), &oa->o_oi); LASSERT(rc == 0); oa->o_size = size; oa->o_blocks = OBD_OBJECT_EOF; @@ -1240,7 +1498,7 @@ int osp_object_truncate(const struct lu_env *env, struct dt_object *dt, body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY); LASSERT(body); - lustre_set_wire_obdo(&body->oa, oa); + lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa); /* XXX: capa support? */ /* osc_pack_capa(req, body, capa); */ @@ -1257,13 +1515,27 @@ out: RETURN(rc); } +/** + * Initialize precreation functionality of OSP + * + * Prepares all the internal structures and starts the precreate thread + * + * \param[in] d OSP device + * + * \retval 0 on success + * \retval negative negated errno on error + */ int osp_init_precreate(struct osp_device *d) { struct l_wait_info lwi = { 0 }; - int rc; + struct task_struct *task; ENTRY; + OBD_ALLOC_PTR(d->opd_pre); + if (d->opd_pre == NULL) + RETURN(-ENOMEM); + /* initially precreation isn't ready */ d->opd_pre_status = -EAGAIN; fid_zero(&d->opd_pre_used_fid); @@ -1278,9 +1550,9 @@ int osp_init_precreate(struct osp_device *d) d->opd_pre_max_grow_count = OST_MAX_PRECREATE; spin_lock_init(&d->opd_pre_lock); - cfs_waitq_init(&d->opd_pre_waitq); - cfs_waitq_init(&d->opd_pre_user_waitq); - cfs_waitq_init(&d->opd_pre_thread.t_ctl_waitq); + init_waitqueue_head(&d->opd_pre_waitq); + init_waitqueue_head(&d->opd_pre_user_waitq); + init_waitqueue_head(&d->opd_pre_thread.t_ctl_waitq); /* * Initialize statfs-related things @@ -1295,10 +1567,11 @@ int osp_init_precreate(struct osp_device *d) /* * start thread handling precreation and statfs updates */ - rc = cfs_create_thread(osp_precreate_thread, d, 0); - if (rc < 0) { - CERROR("can't start precreate thread %d\n", rc); - RETURN(rc); + task = kthread_run(osp_precreate_thread, d, + "osp-pre-%u-%u", d->opd_index, d->opd_group); + if (IS_ERR(task)) { + CERROR("can't start precreate thread %ld\n", PTR_ERR(task)); + RETURN(PTR_ERR(task)); } l_wait_event(d->opd_pre_thread.t_ctl_waitq, @@ -1308,18 +1581,35 @@ int osp_init_precreate(struct osp_device *d) RETURN(0); } +/** + * Finish precreate functionality of OSP + * + * + * Asks all the activity (the thread, update timer) to stop, then + * wait till that is done. + * + * \param[in] d OSP device + */ void osp_precreate_fini(struct osp_device *d) { - struct ptlrpc_thread *thread = &d->opd_pre_thread; + struct ptlrpc_thread *thread; ENTRY; cfs_timer_disarm(&d->opd_statfs_timer); + if (d->opd_pre == NULL) + RETURN_EXIT; + + thread = &d->opd_pre_thread; + thread->t_flags = SVC_STOPPING; - cfs_waitq_signal(&d->opd_pre_waitq); + wake_up(&d->opd_pre_waitq); + + wait_event(thread->t_ctl_waitq, thread->t_flags & SVC_STOPPED); - cfs_wait_event(thread->t_ctl_waitq, thread->t_flags & SVC_STOPPED); + OBD_FREE_PTR(d->opd_pre); + d->opd_pre = NULL; EXIT; }