X-Git-Url: https://git.whamcloud.com/?a=blobdiff_plain;f=lustre%2Fosp%2Fosp_precreate.c;h=a9ab53c371ae86d73468719ecb76a742141c139a;hb=9b0ebf78f7919a144673edadc4a95bad84fae2d3;hp=7f68596ea21a735638e03403d5611c81b815bae9;hpb=e089a515efae3391709b997be889ebe0f3306e9d;p=fs%2Flustre-release.git diff --git a/lustre/osp/osp_precreate.c b/lustre/osp/osp_precreate.c index 7f68596..a9ab53c 100644 --- a/lustre/osp/osp_precreate.c +++ b/lustre/osp/osp_precreate.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 */ @@ -27,13 +23,13 @@ * 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_sync.c + * lustre/osp/osp_precreate.c * * Lustre OST Proxy Device * @@ -44,6 +40,10 @@ #define DEBUG_SUBSYSTEM S_MDS +#include + +#include + #include "osp_internal.h" /* @@ -52,38 +52,109 @@ * = import is disconnected = * * = import is inactive = - * in this case osp_declare_object_create() returns an error + * in this case osp_declare_create() returns an error * */ -/* - * 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) { - return !cfs_time_before(cfs_time_current(), - d->opd_statfs_fresh_till); + return !ktime_before(ktime_get(), 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) +static void osp_statfs_timer_cb(cfs_timer_cb_arg_t data) { - struct osp_device *d = (struct osp_device *) _d; + struct osp_device *d = cfs_from_timer(d, data, opd_statfs_timer); LASSERT(d); - wake_up(&d->opd_pre_waitq); + if (osp_precreate_running(d)) + wake_up(&d->opd_pre_waitq); +} + +static void osp_pre_update_msfs(struct osp_device *d, struct obd_statfs *msfs); + +/* + * The function updates current precreation status if broken, and + * updates that cached statfs state if functional, then wakes up waiters. + * We don't clear opd_pre_status directly here, but rather leave this + * to osp_pre_update_msfs() to do if everything is OK so that we don't + * have a race to clear opd_pre_status and then set it to -ENOSPC again. + * + * \param[in] d OSP device + * \param[in] msfs statfs data + * \param[in] rc new precreate status for device \a d + */ +static void osp_pre_update_status_msfs(struct osp_device *d, + struct obd_statfs *msfs, int rc) +{ + if (rc) + d->opd_pre_status = rc; + else + osp_pre_update_msfs(d, msfs); + + wake_up(&d->opd_pre_user_waitq); +} + +/* Pass in the old statfs data in case the limits have changed */ +void osp_pre_update_status(struct osp_device *d, int rc) +{ + osp_pre_update_status_msfs(d, &d->opd_statfs, rc); } + +/** + * 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) + struct ptlrpc_request *req, void *args, int rc) { - struct obd_import *imp = req->rq_import; - struct obd_statfs *msfs; - struct osp_device *d; + union ptlrpc_async_args *aa = args; + struct obd_import *imp = req->rq_import; + struct obd_statfs *msfs; + struct osp_device *d; + u64 maxage_ns; ENTRY; - aa = ptlrpc_req_async_args(req); + aa = ptlrpc_req_async_args(aa, req); d = aa->pointer_arg[0]; LASSERT(d); @@ -94,33 +165,51 @@ static int osp_statfs_interpret(const struct lu_env *env, if (msfs == NULL) GOTO(out, rc = -EPROTO); - d->opd_statfs = *msfs; - - osp_pre_update_status(d, rc); + if (d->opd_pre) + osp_pre_update_status_msfs(d, msfs, 0); + else + d->opd_statfs = *msfs; /* schedule next update */ - d->opd_statfs_fresh_till = cfs_time_shift(d->opd_statfs_maxage); - cfs_timer_arm(&d->opd_statfs_timer, d->opd_statfs_fresh_till); + maxage_ns = d->opd_statfs_maxage * NSEC_PER_SEC; + d->opd_statfs_fresh_till = ktime_add_ns(ktime_get(), maxage_ns); + mod_timer(&d->opd_statfs_timer, + jiffies + cfs_time_seconds(d->opd_statfs_maxage)); d->opd_statfs_update_in_progress = 0; CDEBUG(D_CACHE, "updated statfs %p\n", d); RETURN(0); out: - /* couldn't update statfs, try again as soon as possible */ - wake_up(&d->opd_pre_waitq); + /* couldn't update statfs, try again with a small delay */ + d->opd_statfs_fresh_till = ktime_add_ns(ktime_get(), 10 * NSEC_PER_SEC); + d->opd_statfs_update_in_progress = 0; + 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); } -static int osp_statfs_update(struct osp_device *d) +/** + * 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(const struct lu_env *env, struct osp_device *d) { + u64 expire = obd_timeout * 1000 * NSEC_PER_SEC; struct ptlrpc_request *req; struct obd_import *imp; union ptlrpc_async_args *aa; - int rc; + int rc; ENTRY; @@ -129,40 +218,74 @@ static int osp_statfs_update(struct osp_device *d) imp = d->opd_obd->u.cli.cl_import; LASSERT(imp); - req = ptlrpc_request_alloc(imp, &RQF_OST_STATFS); + req = ptlrpc_request_alloc(imp, + d->opd_pre ? &RQF_OST_STATFS : &RQF_MDS_STATFS); if (req == NULL) RETURN(-ENOMEM); - rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS); + rc = ptlrpc_request_pack(req, + d->opd_pre ? LUSTRE_OST_VERSION : LUSTRE_MDS_VERSION, + d->opd_pre ? OST_STATFS : MDS_STATFS); if (rc) { ptlrpc_request_free(req); RETURN(rc); } ptlrpc_request_set_replen(req); - req->rq_request_portal = OST_CREATE_PORTAL; + if (d->opd_pre) + req->rq_request_portal = OST_CREATE_PORTAL; ptlrpc_at_set_req_timeout(req); - req->rq_interpret_reply = (ptlrpc_interpterer_t)osp_statfs_interpret; - aa = ptlrpc_req_async_args(req); + req->rq_interpret_reply = osp_statfs_interpret; + aa = ptlrpc_req_async_args(aa, req); aa->pointer_arg[0] = d; /* * no updates till reply */ - cfs_timer_disarm(&d->opd_statfs_timer); - d->opd_statfs_fresh_till = cfs_time_shift(obd_timeout * 1000); + del_timer(&d->opd_statfs_timer); + d->opd_statfs_fresh_till = ktime_add_ns(ktime_get(), expire); d->opd_statfs_update_in_progress = 1; - ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1); + ptlrpcd_add_req(req); + + /* we still want to sync changes if no new changes are coming */ + if (ktime_before(ktime_get(), d->opd_sync_next_commit_cb)) + GOTO(out, rc); + + if (atomic_read(&d->opd_sync_changes)) { + struct thandle *th; + + th = dt_trans_create(env, d->opd_storage); + if (IS_ERR(th)) { + CERROR("%s: can't sync\n", d->opd_obd->obd_name); + GOTO(out, rc); + } + rc = dt_trans_start_local(env, d->opd_storage, th); + if (rc == 0) { + CDEBUG(D_OTHER, "%s: sync forced, %d changes\n", + d->opd_obd->obd_name, + atomic_read(&d->opd_sync_changes)); + osp_sync_add_commit_cb_1s(env, d, th); + dt_trans_stop(env, d->opd_storage, th); + } + } +out: 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) { @@ -172,53 +295,43 @@ void osp_statfs_need_now(struct osp_device *d) * then we should poll OST immediately once object destroy * is replied */ - d->opd_statfs_fresh_till = cfs_time_shift(-1); - cfs_timer_disarm(&d->opd_statfs_timer); + d->opd_statfs_fresh_till = ktime_sub_ns(ktime_get(), NSEC_PER_SEC); + del_timer(&d->opd_statfs_timer); 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)); - fid_to_ostid(fid1, oi1); - fid_to_ostid(fid2, oi2); - LASSERT(ostid_id(oi1) >= ostid_id(oi2)); - - return ostid_id(oi1) - ostid_id(oi2); - } - - 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) { @@ -226,15 +339,29 @@ static inline int osp_precreate_near_empty_nolock(const struct lu_env *env, /* don't consider new precreation till OST is healty and * has free space */ - return ((window - d->opd_pre_reserved < d->opd_pre_grow_count / 2) && + return ((window - d->opd_pre_reserved < d->opd_pre_create_count / 2) && (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) { int rc; + if (d->opd_pre == NULL) + return 0; + /* XXX: do we really need locking here? */ spin_lock(&d->opd_pre_lock); rc = osp_precreate_near_empty_nolock(env, d); @@ -242,6 +369,18 @@ 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) { @@ -255,7 +394,19 @@ 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) @@ -264,21 +415,23 @@ int osp_write_last_oid_seq_files(struct lu_env *env, struct osp_device *osp, struct lu_buf *lb_oid = &oti->osi_lb; struct lu_buf *lb_oseq = &oti->osi_lb2; loff_t oid_off; + u64 oid; loff_t oseq_off; struct thandle *th; 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. */ - lb_oid->lb_buf = &fid->f_oid; - lb_oid->lb_len = sizeof(obd_id); - oid_off = sizeof(obd_id) * osp->opd_index; + if (osp->opd_storage->dd_rdonly) + RETURN(0); - lb_oseq->lb_buf = &fid->f_seq; - lb_oseq->lb_len = sizeof(obd_id); - oseq_off = sizeof(obd_id) * osp->opd_index; + /* Note: through f_oid is only 32 bits, it will also write 64 bits + * for oid to keep compatibility with the previous version. */ + oid = fid->f_oid; + osp_objid_buf_prep(lb_oid, &oid_off, + &oid, osp->opd_index); + + osp_objseq_buf_prep(lb_oseq, &oseq_off, + &fid->f_seq, osp->opd_index); th = dt_trans_create(env, osp->opd_storage); if (IS_ERR(th)) @@ -286,12 +439,12 @@ 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); @@ -318,7 +471,27 @@ 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; @@ -345,12 +518,13 @@ int osp_precreate_rollover_new_seq(struct lu_env *env, struct osp_device *osp) RETURN(rc); } - LCONSOLE_INFO("%s: update sequence from "LPX64" to "LPX64"\n", + LCONSOLE_INFO("%s: update sequence from %#llx to %#llx\n", osp->opd_obd->obd_name, fid_seq(last_fid), fid_seq(fid)); /* Update last_xxx to the new seq */ spin_lock(&osp->opd_pre_lock); osp->opd_last_used_fid = *fid; + osp_fid_to_obdid(fid, &osp->opd_last_id); osp->opd_gap_start_fid = *fid; osp->opd_pre_used_fid = *fid; osp->opd_pre_last_created_fid = *fid; @@ -360,11 +534,23 @@ 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) { @@ -375,16 +561,17 @@ static int osp_precreate_fids(const struct lu_env *env, struct osp_device *osp, if (fid_is_idif(fid)) { struct lu_fid *last_fid; struct ost_id *oi = &osi->osi_oi; + int rc; spin_lock(&osp->opd_pre_lock); last_fid = &osp->opd_pre_last_created_fid; 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); + rc = ostid_set_id(oi, ostid_id(oi) + *grow); spin_unlock(&osp->opd_pre_lock); - if (*grow == 0) + if (*grow == 0 || rc) return 1; ostid_to_fid(fid, oi, osp->opd_index); @@ -405,6 +592,21 @@ 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 + * wakes up 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 for fewer 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); @@ -443,9 +645,9 @@ static int osp_precreate_send(const struct lu_env *env, struct osp_device *d) } spin_lock(&d->opd_pre_lock); - if (d->opd_pre_grow_count > d->opd_pre_max_grow_count / 2) - d->opd_pre_grow_count = d->opd_pre_max_grow_count / 2; - grow = d->opd_pre_grow_count; + if (d->opd_pre_create_count > d->opd_pre_max_create_count / 2) + d->opd_pre_create_count = d->opd_pre_max_create_count / 2; + grow = d->opd_pre_create_count; spin_unlock(&d->opd_pre_lock); body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY); @@ -453,15 +655,9 @@ static int osp_precreate_send(const struct lu_env *env, struct osp_device *d) *fid = d->opd_pre_last_created_fid; rc = osp_precreate_fids(env, d, fid, &grow); - if (rc == 1) { + if (rc == 1) /* Current seq has been used up*/ - if (!osp_is_fid_client(d)) { - osp_pre_update_status(d, -ENOSPC); - rc = -ENOSPC; - } - wake_up(&d->opd_pre_waitq); - GOTO(out_req, rc); - } + GOTO(out_req, rc = -ENOSPC); if (!osp_is_fid_client(d)) { /* Non-FID client will always send seq 0 because of @@ -475,6 +671,9 @@ static int osp_precreate_send(const struct lu_env *env, struct osp_device *d) ptlrpc_request_set_replen(req); + if (OBD_FAIL_CHECK(OBD_FAIL_OSP_FAKE_PRECREATE)) + GOTO(ready, rc = 0); + rc = ptlrpc_queue_wait(req); if (rc) { CERROR("%s: can't precreate: rc = %d\n", d->opd_obd->obd_name, @@ -488,25 +687,33 @@ static int osp_precreate_send(const struct lu_env *env, struct osp_device *d) GOTO(out_req, rc = -EPROTO); ostid_to_fid(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)); - diff = lu_fid_diff(fid, &d->opd_pre_last_created_fid); +ready: + 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 = osp_fid_diff(fid, &d->opd_pre_last_created_fid); spin_lock(&d->opd_pre_lock); if (diff < grow) { /* the OST has not managed to create all the * objects we asked for */ - d->opd_pre_grow_count = max(diff, OST_MIN_PRECREATE); - d->opd_pre_grow_slow = 1; + d->opd_pre_create_count = max(diff, OST_MIN_PRECREATE); + d->opd_pre_create_slow = 1; } else { /* the OST is able to keep up with the work, - * we could consider increasing grow_count + * we could consider increasing create_count * next time if needed */ - d->opd_pre_grow_slow = 0; + d->opd_pre_create_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); @@ -522,6 +729,20 @@ out_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) { @@ -539,24 +760,22 @@ 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(struct lu_fid)); - 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; - tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_VAL); - fid_cpu_to_le((struct lu_fid *)tmp, &d->opd_last_used_fid); + 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); @@ -596,8 +815,21 @@ 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) @@ -696,12 +928,12 @@ static int osp_precreate_cleanup_orphans(struct lu_env *env, 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_create_count = OST_MIN_PRECREATE + diff; d->opd_pre_last_created_fid = d->opd_last_used_fid; } else { - d->opd_pre_grow_count = OST_MIN_PRECREATE; + d->opd_pre_create_count = OST_MIN_PRECREATE; d->opd_pre_last_created_fid = *last_fid; } /* @@ -711,7 +943,7 @@ static int osp_precreate_cleanup_orphans(struct lu_env *env, LASSERT(fid_oid(&d->opd_pre_last_created_fid) <= LUSTRE_DATA_SEQ_MAX_WIDTH); d->opd_pre_used_fid = d->opd_pre_last_created_fid; - d->opd_pre_grow_slow = 0; + d->opd_pre_create_slow = 0; spin_unlock(&d->opd_pre_lock); CDEBUG(D_HA, "%s: Got last_id "DFID" from OST, last_created "DFID @@ -721,10 +953,6 @@ 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. * Since waking up the herd would be useless without pre-created @@ -743,78 +971,142 @@ out: } else { wake_up(&d->opd_pre_user_waitq); } + } else { + spin_lock(&d->opd_pre_lock); + d->opd_pre_recovering = 0; + spin_unlock(&d->opd_pre_lock); } RETURN(rc); } -/* - * the function updates current precreation status used: functional or not +/** + * 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. * - * rc is a last code from the transport, rc == 0 meaning transport works - * well and users of lod can use objects from this OSP + * The new statfs data is passed in \a msfs and needs to be stored into + * opd_statfs, but only after the various flags in os_state are set, so + * that the new statfs data is not visible without appropriate flags set. + * As such, there is no need to clear the flags here, since this is called + * with new statfs data, and they should not be cleared if sent from OST. * - * the status depends on current usage of OST + * 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 ENOSPC/ENOINO flags unconditionally when there is less than the + * reserved size free, and still copy them from the old state when there + * is less than 2*reserved size free space or inodes. + * + * \param[in] d OSP device + * \param[in] msfs statfs data */ -void osp_pre_update_status(struct osp_device *d, int rc) +static void osp_pre_update_msfs(struct osp_device *d, struct obd_statfs *msfs) { - struct obd_statfs *msfs = &d->opd_statfs; - int old = d->opd_pre_status; - __u64 used; - - d->opd_pre_status = 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); - if ((msfs->os_ffree < 32) || (msfs->os_bavail < used)) { - d->opd_pre_status = -ENOSPC; - if (old != -ENOSPC) - CDEBUG(D_INFO, "%s: status: "LPU64" blocks, " - LPU64" free, "LPU64" used, "LPU64" " - "avail -> %d: rc = %d\n", - d->opd_obd->obd_name, msfs->os_blocks, - msfs->os_bfree, used, msfs->os_bavail, - d->opd_pre_status, rc); - CDEBUG(D_INFO, - "non-commited changes: %lu, in progress: %u\n", - 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; - 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, - msfs->os_blocks, msfs->os_bfree, used, - msfs->os_bavail, d->opd_pre_status, rc); + u32 old_state = d->opd_statfs.os_state; + u32 reserved_ino_low = 32; /* could be tunable in the future */ + u32 reserved_ino_high = reserved_ino_low * 2; + u64 available_mb; + + /* statfs structure not initialized yet */ + if (unlikely(!msfs->os_type)) + return; + + /* if the low and high watermarks have not been initialized yet */ + if (unlikely(d->opd_reserved_mb_high == 0 && + d->opd_reserved_mb_low == 0)) { + /* Use ~0.1% by default to disable object allocation, + * and ~0.2% to enable, size in MB, set both watermark + */ + spin_lock(&d->opd_pre_lock); + if (d->opd_reserved_mb_high == 0 && + d->opd_reserved_mb_low == 0) { + d->opd_reserved_mb_low = ((msfs->os_bsize >> 10) * + msfs->os_blocks) >> 20; + if (d->opd_reserved_mb_low == 0) + d->opd_reserved_mb_low = 1; + d->opd_reserved_mb_high = + (d->opd_reserved_mb_low << 1) + 1; } + spin_unlock(&d->opd_pre_lock); } -out: - wake_up(&d->opd_pre_user_waitq); + available_mb = (msfs->os_bavail * (msfs->os_bsize >> 10)) >> 10; + if (msfs->os_ffree < reserved_ino_low) + msfs->os_state |= OS_STATE_ENOINO; + else if (msfs->os_ffree <= reserved_ino_high) + msfs->os_state |= old_state & OS_STATE_ENOINO; + /* else don't clear flags in new msfs->os_state sent from OST */ + + CDEBUG(D_INFO, + "%s: blocks=%llu free=%llu avail=%llu avail_mb=%llu hwm_mb=%u files=%llu ffree=%llu state=%x: rc = %d\n", + d->opd_obd->obd_name, msfs->os_blocks, msfs->os_bfree, + msfs->os_bavail, available_mb, d->opd_reserved_mb_high, + msfs->os_files, msfs->os_ffree, msfs->os_state, + d->opd_pre_status); + if (available_mb < d->opd_reserved_mb_low) + msfs->os_state |= OS_STATE_ENOSPC; + else if (available_mb <= d->opd_reserved_mb_high) + msfs->os_state |= old_state & OS_STATE_ENOSPC; + /* else don't clear flags in new msfs->os_state sent from OST */ + + if (msfs->os_state & (OS_STATE_ENOINO | OS_STATE_ENOSPC)) { + d->opd_pre_status = -ENOSPC; + if (!(old_state & (OS_STATE_ENOINO | OS_STATE_ENOSPC))) + CDEBUG(D_INFO, "%s: full: state=%x: rc = %x\n", + d->opd_obd->obd_name, msfs->os_state, + d->opd_pre_status); + CDEBUG(D_INFO, "uncommitted changes=%u in_progress=%u\n", + atomic_read(&d->opd_sync_changes), + atomic_read(&d->opd_sync_rpcs_in_progress)); + } else if (old_state & (OS_STATE_ENOINO | OS_STATE_ENOSPC)) { + d->opd_pre_status = 0; + spin_lock(&d->opd_pre_lock); + d->opd_pre_create_slow = 0; + d->opd_pre_create_count = OST_MIN_PRECREATE; + spin_unlock(&d->opd_pre_lock); + wake_up(&d->opd_pre_waitq); + + CDEBUG(D_INFO, + "%s: available: state=%x: rc = %d\n", + d->opd_obd->obd_name, msfs->os_state, + d->opd_pre_status); + } else { + /* we only get here if rc == 0 in the caller */ + d->opd_pre_status = 0; + } + + /* Object precreation skipped on OST if manually disabled */ + if (d->opd_pre_max_create_count == 0) + msfs->os_state |= OS_STATE_NOPRECREATE; + /* else don't clear flags in new msfs->os_state sent from OST */ + + /* copy only new statfs state to make it visible to MDS threads */ + if (&d->opd_statfs != msfs) + d->opd_statfs = *msfs; } -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; @@ -823,9 +1115,29 @@ static int osp_init_pre_fid(struct osp_device *osp) int rc; ENTRY; - /* Return if last_used fid has been initialized */ + LASSERT(osp->opd_pre != NULL); + + /* Let's check if the current last_seq/fid is valid, + * otherwise request new sequence from the controller */ + if (osp_is_fid_client(osp) && osp->opd_group != 0) { + /* Non-MDT0 can only use normal sequence for + * OST objects */ + if (fid_is_norm(&osp->opd_last_used_fid)) + RETURN(0); + } else { + /* Initially MDT0 will start with IDIF, after + * that it will request new sequence from the + * controller */ + if (fid_is_idif(&osp->opd_last_used_fid) || + fid_is_norm(&osp->opd_last_used_fid)) + RETURN(0); + } + if (!fid_is_zero(&osp->opd_last_used_fid)) - RETURN(0); + CWARN("%s: invalid last used fid "DFID + ", try to get new sequence.\n", + osp->opd_obd->obd_name, + PFID(&osp->opd_last_used_fid)); rc = lu_env_init(&env, osp->opd_dt_dev.dd_lu_dev.ld_type->ldt_ctx_tags); if (rc) { @@ -868,11 +1180,31 @@ 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 }; + struct l_wait_info lwi2 = LWI_TIMEOUT(cfs_time_seconds(5), + back_to_sleep, NULL); struct lu_env env; int rc; @@ -882,6 +1214,12 @@ static int osp_precreate_thread(void *_arg) if (rc) { CERROR("%s: init env error: rc = %d\n", d->opd_obd->obd_name, rc); + + spin_lock(&d->opd_pre_lock); + thread->t_flags = SVC_STOPPED; + spin_unlock(&d->opd_pre_lock); + wake_up(&thread->t_ctl_waitq); + RETURN(rc); } @@ -895,6 +1233,10 @@ static int osp_precreate_thread(void *_arg) * need to be connected to OST */ while (osp_precreate_running(d)) { + if ((d->opd_pre == NULL || d->opd_pre_recovering) && + d->opd_imp_connected && + !d->opd_got_disconnected) + break; l_wait_event(d->opd_pre_waitq, !osp_precreate_running(d) || d->opd_new_connection, @@ -911,30 +1253,40 @@ static int osp_precreate_thread(void *_arg) if (!osp_precreate_running(d)) 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); + if (d->opd_pre) { + LASSERT(d->opd_obd->u.cli.cl_seq != NULL); + /* 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); + d->opd_obd->obd_name, rc); continue; } } - osp_statfs_update(d); - - /* - * Clean up orphans or recreate missing objects. - */ - rc = osp_precreate_cleanup_orphans(&env, d); - if (rc != 0) + if (osp_statfs_update(&env, d)) { + l_wait_event(d->opd_pre_waitq, + !osp_precreate_running(d), &lwi2); continue; + } + + if (d->opd_pre) { + /* + * Clean up orphans or recreate missing objects. + */ + rc = osp_precreate_cleanup_orphans(&env, d); + if (rc != 0) { + schedule_timeout_interruptible(cfs_time_seconds(1)); + continue; + } + } + /* * connected, can handle precreates now */ @@ -954,7 +1306,11 @@ static int osp_precreate_thread(void *_arg) break; if (osp_statfs_need_update(d)) - osp_statfs_update(d); + if (osp_statfs_update(&env, d)) + break; + + if (d->opd_pre == NULL) + continue; /* To avoid handling different seq in precreate/orphan * cleanup, it will hold precreate until current seq is @@ -965,7 +1321,7 @@ static int osp_precreate_thread(void *_arg) if (unlikely(osp_precreate_end_seq(&env, d) && osp_create_end_seq(&env, d))) { - LCONSOLE_INFO("%s:"LPX64" is used up." + LCONSOLE_INFO("%s:%#llx is used up." " Update to new seq\n", d->opd_obd->obd_name, fid_seq(&d->opd_pre_last_created_fid)); @@ -995,6 +1351,20 @@ static int osp_precreate_thread(void *_arg) 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) { @@ -1007,13 +1377,23 @@ static int osp_precreate_ready_condition(const struct lu_env *env, return 1; /* ready if OST reported no space and no destroys in progress */ - if (d->opd_syn_changes + d->opd_syn_rpc_in_progress == 0 && + if (atomic_read(&d->opd_sync_changes) + + atomic_read(&d->opd_sync_rpcs_in_progress) == 0 && d->opd_pre_status == -ENOSPC) return 1; /* Bail out I/O fails to OST */ - if (d->opd_pre_status == -EIO) + if (d->opd_pre_status != 0 && + d->opd_pre_status != -EAGAIN && + d->opd_pre_status != -ENODEV && + d->opd_pre_status != -ENOTCONN && + 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; } @@ -1022,29 +1402,47 @@ 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=%llu, sync_changes=%u, " + "sync_rpcs_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, + atomic_read(&d->opd_sync_changes), + atomic_read(&d->opd_sync_rpcs_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; + time64_t expire = ktime_get_seconds() + obd_timeout; + struct l_wait_info lwi; + int precreated, rc, synced = 0; ENTRY; @@ -1052,6 +1450,10 @@ int osp_precreate_reserve(const struct lu_env *env, struct osp_device *d) "Next FID "DFID"\n", PFID(&d->opd_pre_last_created_fid), PFID(&d->opd_pre_used_fid)); + /* opd_pre_max_create_count 0 to not use specified OST. */ + if (d->opd_pre_max_create_count == 0) + RETURN(-ENOBUFS); + /* * wait till: * - preallocation is done @@ -1060,18 +1462,18 @@ 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) { + rc == -ENODEV || rc == -EAGAIN || rc == -ENOTCONN) { /* * increase number of precreations */ precreated = osp_objs_precreated(env, d); - if (d->opd_pre_grow_count < d->opd_pre_max_grow_count && - d->opd_pre_grow_slow == 0 && - precreated <= (d->opd_pre_grow_count / 4 + 1)) { + if (d->opd_pre_create_count < d->opd_pre_max_create_count && + d->opd_pre_create_slow == 0 && + precreated <= (d->opd_pre_create_count / 4 + 1)) { spin_lock(&d->opd_pre_lock); - d->opd_pre_grow_slow = 1; - d->opd_pre_grow_count *= 2; + d->opd_pre_create_slow = 1; + d->opd_pre_create_count *= 2; spin_unlock(&d->opd_pre_lock); } @@ -1099,16 +1501,18 @@ int osp_precreate_reserve(const struct lu_env *env, struct osp_device *d) * wait till that is done - some space might be released */ if (unlikely(rc == -ENOSPC)) { - if (d->opd_syn_changes) { + if (atomic_read(&d->opd_sync_changes) && synced == 0) { /* force local commit to release space */ dt_commit_async(env, d->opd_storage); + osp_sync_check_for_work(d); + synced = 1; } - if (d->opd_syn_rpc_in_progress) { + if (atomic_read(&d->opd_sync_rpcs_in_progress)) { /* just wait till destroys are done */ /* see l_wait_even() few lines below */ } - if (d->opd_syn_changes + - d->opd_syn_rpc_in_progress == 0) { + if (atomic_read(&d->opd_sync_changes) + + atomic_read(&d->opd_sync_rpcs_in_progress) == 0) { /* no hope for free space */ break; } @@ -1117,9 +1521,9 @@ int osp_precreate_reserve(const struct lu_env *env, struct osp_device *d) /* XXX: don't wake up if precreation is in progress */ wake_up(&d->opd_pre_waitq); - lwi = LWI_TIMEOUT(expire - cfs_time_current(), - osp_precreate_timeout_condition, d); - if (cfs_time_aftereq(cfs_time_current(), expire)) { + lwi = LWI_TIMEOUT(cfs_time_seconds(obd_timeout), + osp_precreate_timeout_condition, d); + if (ktime_get_seconds() >= expire) { rc = -ETIMEDOUT; break; } @@ -1131,21 +1535,55 @@ 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) { + struct lu_fid *pre_used_fid = &d->opd_pre_used_fid; /* 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), PFID(&d->opd_pre_last_created_fid)); + /* + * When sequence is used up, new one should be allocated in + * osp_precreate_rollover_new_seq. So ASSERT here to avoid + * objid overflow. + */ + LASSERTF(osp_fid_end_seq(env, pre_used_fid) == 0, + "next fid "DFID" last created fid "DFID"\n", + PFID(&d->opd_pre_used_fid), + PFID(&d->opd_pre_last_created_fid)); + /* Non IDIF fids shoulnd't get here with oid == 0xFFFFFFFF. */ + if (fid_is_idif(pre_used_fid) && + unlikely(fid_oid(pre_used_fid) == LUSTRE_DATA_SEQ_MAX_WIDTH)) + pre_used_fid->f_seq++; + d->opd_pre_used_fid.f_oid++; memcpy(fid, &d->opd_pre_used_fid, sizeof(*fid)); d->opd_pre_reserved--; @@ -1161,14 +1599,28 @@ int osp_precreate_get_fid(const struct lu_env *env, struct osp_device *d, * all reservations are released, see comment in * osp_precreate_thread() just before orphan cleanup */ - if (unlikely(d->opd_pre_reserved == 0 && d->opd_pre_status)) + if (unlikely(d->opd_pre_reserved == 0 && + (d->opd_pre_recovering || d->opd_pre_status))) 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) @@ -1189,8 +1641,6 @@ int osp_object_truncate(const struct lu_env *env, struct dt_object *dt, if (req == NULL) RETURN(-ENOMEM); - /* XXX: capa support? */ - /* osc_set_capa_size(req, &RMF_CAPA1, capa); */ rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_PUNCH); if (rc) { ptlrpc_request_free(req); @@ -1201,6 +1651,13 @@ int osp_object_truncate(const struct lu_env *env, struct dt_object *dt, * XXX: decide how do we do here with resend * if we don't resend, then client may see wrong file size * if we do resend, then MDS thread can get stuck for quite long + * and if we don't resend, then client will also get -EWOULDBLOCK !! + * (see LU-7975 and sanity/test_27F use cases) + * but let's decide not to resend/delay this truncate request to OST + * and allow Client to decide to resend, in a less agressive way from + * after_reply(), by returning -EINPROGRESS instead of + * -EAGAIN/-EWOULDBLOCK upon return from ptlrpc_queue_wait() at the + * end of this routine */ req->rq_no_resend = req->rq_no_delay = 1; @@ -1228,8 +1685,23 @@ int osp_object_truncate(const struct lu_env *env, struct dt_object *dt, ptlrpc_request_set_replen(req); rc = ptlrpc_queue_wait(req); - if (rc) - CERROR("can't punch object: %d\n", rc); + if (rc) { + /* -EWOULDBLOCK/-EAGAIN means OST is unreachable at the moment + * since we have decided not to resend/delay, but this could + * lead to wrong size to be seen at Client side and even process + * trying to open to exit/fail if not itself handling -EAGAIN. + * So it should be better to return -EINPROGRESS instead and + * leave the decision to resend at Client side in after_reply() + */ + if (rc == -EWOULDBLOCK) { + rc = -EINPROGRESS; + CDEBUG(D_HA, "returning -EINPROGRESS instead of " + "-EWOULDBLOCK/-EAGAIN to allow Client to " + "resend\n"); + } else { + CERROR("can't punch object: %d\n", rc); + } + } out: ptlrpc_req_finished(req); if (oa) @@ -1237,46 +1709,98 @@ 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 }; - cfs_task_t *task; - ENTRY; + OBD_ALLOC_PTR(d->opd_pre); + if (d->opd_pre == NULL) + RETURN(-ENOMEM); + /* initially precreation isn't ready */ + init_waitqueue_head(&d->opd_pre_user_waitq); d->opd_pre_status = -EAGAIN; fid_zero(&d->opd_pre_used_fid); d->opd_pre_used_fid.f_oid = 1; fid_zero(&d->opd_pre_last_created_fid); d->opd_pre_last_created_fid.f_oid = 1; + d->opd_last_id = 0; d->opd_pre_reserved = 0; d->opd_got_disconnected = 1; - d->opd_pre_grow_slow = 0; - d->opd_pre_grow_count = OST_MIN_PRECREATE; - d->opd_pre_min_grow_count = OST_MIN_PRECREATE; - d->opd_pre_max_grow_count = OST_MAX_PRECREATE; + d->opd_pre_create_slow = 0; + d->opd_pre_create_count = OST_MIN_PRECREATE; + d->opd_pre_min_create_count = OST_MIN_PRECREATE; + d->opd_pre_max_create_count = OST_MAX_PRECREATE; + d->opd_reserved_mb_high = 0; + d->opd_reserved_mb_low = 0; + + 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) +{ + ENTRY; + + if (d->opd_pre == NULL) + RETURN_EXIT; + + OBD_FREE_PTR(d->opd_pre); + d->opd_pre = NULL; + + EXIT; +} + +int osp_init_statfs(struct osp_device *d) +{ + struct l_wait_info lwi = { 0 }; + struct task_struct *task; + + ENTRY; spin_lock_init(&d->opd_pre_lock); init_waitqueue_head(&d->opd_pre_waitq); - init_waitqueue_head(&d->opd_pre_user_waitq); + thread_set_flags(&d->opd_pre_thread, SVC_INIT); init_waitqueue_head(&d->opd_pre_thread.t_ctl_waitq); /* * Initialize statfs-related things */ - d->opd_statfs_maxage = 5; /* default update interval */ - d->opd_statfs_fresh_till = cfs_time_shift(-1000); - CDEBUG(D_OTHER, "current %llu, fresh till %llu\n", - (unsigned long long)cfs_time_current(), - (unsigned long long)d->opd_statfs_fresh_till); - cfs_timer_init(&d->opd_statfs_timer, osp_statfs_timer_cb, d); + d->opd_statfs_maxage = 5; /* defaultupdate interval */ + d->opd_statfs_fresh_till = ktime_sub_ns(ktime_get(), + 1000 * NSEC_PER_SEC); + CDEBUG(D_OTHER, "current %lldns, fresh till %lldns\n", + ktime_get_ns(), + ktime_to_ns(d->opd_statfs_fresh_till)); + cfs_timer_setup(&d->opd_statfs_timer, osp_statfs_timer_cb, + (unsigned long)d, 0); + + if (d->opd_storage->dd_rdonly) + RETURN(0); /* * start thread handling precreation and statfs updates */ task = kthread_run(osp_precreate_thread, d, - "osp-pre-%u", d->opd_index); + "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)); @@ -1289,19 +1813,18 @@ int osp_init_precreate(struct osp_device *d) RETURN(0); } -void osp_precreate_fini(struct osp_device *d) +void osp_statfs_fini(struct osp_device *d) { struct ptlrpc_thread *thread = &d->opd_pre_thread; - ENTRY; - cfs_timer_disarm(&d->opd_statfs_timer); + del_timer(&d->opd_statfs_timer); - thread->t_flags = SVC_STOPPING; - wake_up(&d->opd_pre_waitq); - - wait_event(thread->t_ctl_waitq, thread->t_flags & SVC_STOPPED); + if (!thread_is_init(thread) && !thread_is_stopped(thread)) { + thread->t_flags = SVC_STOPPING; + wake_up(&d->opd_pre_waitq); + wait_event(thread->t_ctl_waitq, thread_is_stopped(thread)); + } EXIT; } -