X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fosp%2Fosp_precreate.c;h=a9ab53c371ae86d73468719ecb76a742141c139a;hp=a89c9f5ad9e50f8f9475feac53a5529e12a64c79;hb=9b0ebf78f7919a144673edadc4a95bad84fae2d3;hpb=34acfbc2bfe502d18c12ba35771bde7c4a0f7906 diff --git a/lustre/osp/osp_precreate.c b/lustre/osp/osp_precreate.c index a89c9f5..a9ab53c 100644 --- a/lustre/osp/osp_precreate.c +++ b/lustre/osp/osp_precreate.c @@ -23,13 +23,13 @@ * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved. * Use is subject to license terms. * - * Copyright (c) 2012, 2016, 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 * @@ -52,12 +52,11 @@ * = 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 * */ -/* - ** +/** * Check whether statfs data is expired * * OSP device caches statfs data for the target, the function checks @@ -69,8 +68,7 @@ */ 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); } /* @@ -89,15 +87,46 @@ 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); - if (d->opd_pre != NULL && osp_precreate_running(d)) + 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 * @@ -115,16 +144,17 @@ static void osp_statfs_timer_cb(unsigned long _d) * \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); @@ -135,20 +165,25 @@ 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); - mod_timer(&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 */ + /* 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); @@ -168,12 +203,13 @@ out: * * \param[in] d OSP device */ -static int osp_statfs_update(struct osp_device *d) +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; @@ -182,32 +218,59 @@ 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 */ del_timer(&d->opd_statfs_timer); - d->opd_statfs_fresh_till = cfs_time_shift(obd_timeout * 1000); + d->opd_statfs_fresh_till = ktime_add_ns(ktime_get(), expire); d->opd_statfs_update_in_progress = 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); } @@ -232,7 +295,7 @@ 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); + 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); } @@ -296,6 +359,9 @@ static inline int osp_precreate_near_empty(const struct lu_env *env, { 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); @@ -349,6 +415,7 @@ 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; @@ -359,13 +426,12 @@ int osp_write_last_oid_seq_files(struct lu_env *env, struct osp_device *osp, /* 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(u64); - oid_off = sizeof(u64) * osp->opd_index; + oid = fid->f_oid; + osp_objid_buf_prep(lb_oid, &oid_off, + &oid, osp->opd_index); - lb_oseq->lb_buf = &fid->f_seq; - lb_oseq->lb_len = sizeof(u64); - oseq_off = sizeof(u64) * 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)) @@ -458,6 +524,7 @@ static int osp_precreate_rollover_new_seq(struct lu_env *env, /* 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; @@ -530,9 +597,9 @@ static int osp_precreate_fids(const struct lu_env *env, struct osp_device *osp, * * 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 + * 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 less objects (i.e. slow precreate down). + * 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 @@ -577,7 +644,6 @@ static int osp_precreate_send(const struct lu_env *env, struct osp_device *d) RETURN(rc); } - LASSERT(d->opd_pre->osp_pre_delorphan_sent != 0); spin_lock(&d->opd_pre_lock); 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; @@ -589,15 +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 @@ -630,7 +690,7 @@ static int osp_precreate_send(const struct lu_env *env, struct osp_device *d) ready: if (osp_fid_diff(fid, &d->opd_pre_used_fid) <= 0) { - CERROR("%s: precreate fid "DFID" < local used fid "DFID + 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); @@ -783,20 +843,18 @@ static int osp_precreate_cleanup_orphans(struct lu_env *env, int update_status = 0; int rc; int diff; - struct lu_fid fid; ENTRY; /* - * wait for local recovery to finish, so we can cleanup orphans. - * orphans are all objects since "last used" (assigned). - * consider reserved objects as created otherwise we can get into - * a livelock when one blocked thread holding a reservation can - * block recovery. see LU-8367 for the details. in some cases this - * can result in gaps (i.e. leaked objects), but we've got LFSCK... - * - * do not allow new reservations because they may end up getting - * orphans being cleaned up below. so we block new reservations. + * wait for local recovery to finish, so we can cleanup orphans + * orphans are all objects since "last used" (assigned), but + * there might be objects reserved and in some cases they won't + * be used. we can't cleanup them till we're sure they won't be + * used. also can't we allow new reservations because they may + * end up getting orphans being cleaned up below. so we block + * new reservations and wait till all reserved objects either + * user or released. */ spin_lock(&d->opd_pre_lock); d->opd_pre_recovering = 1; @@ -806,12 +864,16 @@ static int osp_precreate_cleanup_orphans(struct lu_env *env, * catch all osp_precreate_reserve() calls who find * "!opd_pre_recovering". */ - l_wait_event(d->opd_pre_waitq, d->opd_recovery_completed || + l_wait_event(d->opd_pre_waitq, + (!d->opd_pre_reserved && d->opd_recovery_completed) || !osp_precreate_running(d) || d->opd_got_disconnected, &lwi); if (!osp_precreate_running(d) || d->opd_got_disconnected) GOTO(out, rc = -EAGAIN); + CDEBUG(D_HA, "%s: going to cleanup orphans since "DFID"\n", + d->opd_obd->obd_name, PFID(&d->opd_last_used_fid)); + *last_fid = d->opd_last_used_fid; /* The OSP should already get the valid seq now */ LASSERT(!fid_is_zero(last_fid)); @@ -840,24 +902,10 @@ static int osp_precreate_cleanup_orphans(struct lu_env *env, if (body == NULL) GOTO(out, rc = -EPROTO); - body->oa.o_flags = 0; + body->oa.o_flags = OBD_FL_DELORPHAN; body->oa.o_valid = OBD_MD_FLFLAGS | OBD_MD_FLGROUP; - /* unless this is the very first DELORPHAN (when we really - * can destroy some orphans), just tell OST to recreate - * missing objects in our precreate pool */ - spin_lock(&d->opd_pre_lock); - if (d->opd_pre->osp_pre_delorphan_sent) { - fid = d->opd_pre_last_created_fid; - } else { - fid = d->opd_last_used_fid; - body->oa.o_flags = OBD_FL_DELORPHAN; - } - spin_unlock(&d->opd_pre_lock); - fid_to_ostid(&fid, &body->oa.o_oi); - - CDEBUG(D_HA, "%s: going to cleanup orphans since "DFID"\n", - d->opd_obd->obd_name, PFID(&fid)); + fid_to_ostid(&d->opd_last_used_fid, &body->oa.o_oi); ptlrpc_request_set_replen(req); @@ -880,10 +928,10 @@ 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 = osp_fid_diff(&fid, last_fid); + diff = osp_fid_diff(&d->opd_last_used_fid, last_fid); if (diff > 0) { d->opd_pre_create_count = OST_MIN_PRECREATE + diff; - d->opd_pre_last_created_fid = *last_fid; + d->opd_pre_last_created_fid = d->opd_last_used_fid; } else { d->opd_pre_create_count = OST_MIN_PRECREATE; d->opd_pre_last_created_fid = *last_fid; @@ -894,11 +942,9 @@ static int osp_precreate_cleanup_orphans(struct lu_env *env, */ LASSERT(fid_oid(&d->opd_pre_last_created_fid) <= LUSTRE_DATA_SEQ_MAX_WIDTH); - if (d->opd_pre->osp_pre_delorphan_sent == 0) - d->opd_pre_used_fid = d->opd_pre_last_created_fid; + d->opd_pre_used_fid = d->opd_pre_last_created_fid; d->opd_pre_create_slow = 0; spin_unlock(&d->opd_pre_lock); - d->opd_pre->osp_pre_delorphan_sent = 1; CDEBUG(D_HA, "%s: Got last_id "DFID" from OST, last_created "DFID "last_used is "DFID"\n", d->opd_obd->obd_name, PFID(last_fid), @@ -942,92 +988,106 @@ out: * 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 ENOSPC when there is less than reserved size - * free and clear it when there is at least 2*reserved size free space. - * the function updates current precreation status used: functional or not * - * \param[in] d OSP device - * \param[in] rc new precreate status for device \a d + * 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. * - * \retval 0 on success - * \retval negative negated errno on error + * 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 available; + 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); + } - d->opd_pre_status = rc; - if (rc) - goto out; + 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); - if (likely(msfs->os_type)) { - 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); - } - /* in MB */ - available = (msfs->os_bavail * (msfs->os_bsize >> 10)) >> 10; - if (msfs->os_ffree < 32) - msfs->os_state |= OS_STATE_ENOINO; - else if (msfs->os_ffree > 64) - msfs->os_state &= ~OS_STATE_ENOINO; - - if (available < d->opd_reserved_mb_low) - msfs->os_state |= OS_STATE_ENOSPC; - else if (available > d->opd_reserved_mb_high) - msfs->os_state &= ~OS_STATE_ENOSPC; - if (msfs->os_state & (OS_STATE_ENOINO | OS_STATE_ENOSPC)) { - d->opd_pre_status = -ENOSPC; - if (old != -ENOSPC) - CDEBUG(D_INFO, "%s: status: %llu blocks, %llu " - "free, %llu avail, %llu MB avail, %u " - "hwm -> %d: rc = %d\n", - d->opd_obd->obd_name, msfs->os_blocks, - msfs->os_bfree, msfs->os_bavail, - available, d->opd_reserved_mb_high, - d->opd_pre_status, rc); - CDEBUG(D_INFO, - "non-committed changes: %u, in progress: %u\n", - atomic_read(&d->opd_syn_changes), - atomic_read(&d->opd_syn_rpc_in_progress)); - } else if (unlikely(old == -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: space available: %llu blocks, %llu" - " free, %llu avail, %lluMB avail, %u lwm" - " -> %d: rc = %d\n", d->opd_obd->obd_name, - msfs->os_blocks, msfs->os_bfree, msfs->os_bavail, - available, d->opd_reserved_mb_low, - d->opd_pre_status, rc); - } + 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; } -out: - wake_up(&d->opd_pre_user_waitq); + + /* 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; } /** @@ -1173,7 +1233,7 @@ static int osp_precreate_thread(void *_arg) * need to be connected to OST */ while (osp_precreate_running(d)) { - if (d->opd_pre_recovering && + if ((d->opd_pre == NULL || d->opd_pre_recovering) && d->opd_imp_connected && !d->opd_got_disconnected) break; @@ -1193,35 +1253,40 @@ static int osp_precreate_thread(void *_arg) if (!osp_precreate_running(d)) break; - 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; + 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); - 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; + } } - if (osp_statfs_update(d)) { + if (osp_statfs_update(&env, d)) { l_wait_event(d->opd_pre_waitq, !osp_precreate_running(d), &lwi2); continue; } - /* - * 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; + 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 */ @@ -1241,9 +1306,12 @@ static int osp_precreate_thread(void *_arg) break; if (osp_statfs_need_update(d)) - if (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 * used up. */ @@ -1309,8 +1377,8 @@ 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 (atomic_read(&d->opd_syn_changes) + - atomic_read(&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; @@ -1335,12 +1403,12 @@ static int osp_precreate_timeout_condition(void *data) struct osp_device *d = data; CDEBUG(D_HA, "%s: slow creates, last="DFID", next="DFID", " - "reserved=%llu, syn_changes=%u, " - "syn_rpc_in_progress=%d, status=%d\n", + "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_syn_changes), - atomic_read(&d->opd_syn_rpc_in_progress), + atomic_read(&d->opd_sync_changes), + atomic_read(&d->opd_sync_rpcs_in_progress), d->opd_pre_status); return 1; @@ -1372,9 +1440,9 @@ static int osp_precreate_timeout_condition(void *data) */ 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; @@ -1386,12 +1454,6 @@ int osp_precreate_reserve(const struct lu_env *env, struct osp_device *d) if (d->opd_pre_max_create_count == 0) RETURN(-ENOBUFS); - if (OBD_FAIL_PRECHECK(OBD_FAIL_MDS_OSP_PRECREATE_WAIT)) { - if (d->opd_index == cfs_fail_val) - OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_OSP_PRECREATE_WAIT, - obd_timeout); - } - /* * wait till: * - preallocation is done @@ -1439,16 +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 (atomic_read(&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 (atomic_read(&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 (atomic_read(&d->opd_syn_changes) + - atomic_read(&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; } @@ -1457,9 +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; } @@ -1496,6 +1560,7 @@ int osp_precreate_reserve(const struct lu_env *env, struct osp_device *d) 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); @@ -1505,6 +1570,20 @@ int osp_precreate_get_fid(const struct lu_env *env, struct osp_device *d, 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--; @@ -1520,7 +1599,8 @@ 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; @@ -1561,8 +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); @@ -1643,9 +1721,6 @@ out: */ int osp_init_precreate(struct osp_device *d) { - struct l_wait_info lwi = { 0 }; - struct task_struct *task; - ENTRY; OBD_ALLOC_PTR(d->opd_pre); @@ -1653,11 +1728,13 @@ int osp_init_precreate(struct osp_device *d) 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_create_slow = 0; @@ -1667,22 +1744,54 @@ int osp_init_precreate(struct osp_device *d) 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); - setup_timer(&d->opd_statfs_timer, osp_statfs_timer_cb, - (unsigned long)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); @@ -1704,34 +1813,18 @@ 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) +void osp_statfs_fini(struct osp_device *d) { struct ptlrpc_thread *thread = &d->opd_pre_thread; ENTRY; del_timer(&d->opd_statfs_timer); - if (d->opd_pre == NULL) - RETURN_EXIT; - 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)); } - OBD_FREE_PTR(d->opd_pre); - d->opd_pre = NULL; - EXIT; } -