X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fosp%2Fosp_precreate.c;h=b8061599c1e3c7240b392e7366074a4c1c031382;hp=076cc6a711d592e96d9664f2c60c9433c15804d8;hb=3cce65712d94cffe8f1626545845b95b88aef672;hpb=f625f670afbe954030ff81f0f8522137d6cdd335 diff --git a/lustre/osp/osp_precreate.c b/lustre/osp/osp_precreate.c index 076cc6a..b806159 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, 2014, 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" /* @@ -52,7 +52,7 @@ * = 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 * */ @@ -69,8 +69,7 @@ */ 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); } /* @@ -89,12 +88,12 @@ 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); } @@ -115,12 +114,13 @@ 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; @@ -137,18 +137,23 @@ static int osp_statfs_interpret(const struct lu_env *env, d->opd_statfs = *msfs; - osp_pre_update_status(d, rc); + if (d->opd_pre) + osp_pre_update_status(d, rc); /* 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 */ + /* 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 +173,13 @@ 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 +188,59 @@ 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; + req->rq_interpret_reply = osp_statfs_interpret; aa = ptlrpc_req_async_args(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); } @@ -216,7 +249,7 @@ static int osp_statfs_update(struct osp_device *d) * * 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 then planned. The function checks there + * 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 @@ -232,8 +265,8 @@ 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); } } @@ -276,7 +309,7 @@ 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)); } @@ -296,6 +329,9 @@ 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); @@ -349,20 +385,23 @@ 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; + if (osp->opd_storage->dd_rdonly) + RETURN(0); + /* 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)) @@ -449,12 +488,13 @@ static int osp_precreate_rollover_new_seq(struct lu_env *env, 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; @@ -491,16 +531,17 @@ 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); @@ -574,9 +615,9 @@ 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); @@ -584,15 +625,9 @@ 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 @@ -606,6 +641,9 @@ 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, @@ -619,8 +657,10 @@ 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); + +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); @@ -632,13 +672,13 @@ static int osp_precreate_send(const struct lu_env *env, struct osp_device *d) 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); @@ -860,10 +900,10 @@ static int osp_precreate_cleanup_orphans(struct lu_env *env, spin_lock(&d->opd_pre_lock); 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; } /* @@ -873,7 +913,7 @@ 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 @@ -883,10 +923,6 @@ 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 @@ -905,6 +941,10 @@ 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); @@ -922,17 +962,8 @@ 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. + * 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 @@ -945,42 +976,86 @@ void osp_pre_update_status(struct osp_device *d, int rc) { struct obd_statfs *msfs = &d->opd_statfs; int old = d->opd_pre_status; - __u64 used; + __u64 available; d->opd_pre_status = rc; if (rc) goto out; 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)) { + 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; + + 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); + 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: "LPU64" blocks, " - LPU64" free, "LPU64" used, "LPU64" " - "avail -> %d: rc = %d\n", + 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, used, msfs->os_bavail, + msfs->os_bfree, msfs->os_bavail, + available, d->opd_reserved_mb_high, 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) { + "non-committed changes: %u, in progress: %u\n", + atomic_read(&d->opd_sync_changes), + atomic_read(&d->opd_sync_rpcs_in_progress)); + } else if (unlikely(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; + 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: 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); + + 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); } - } + /* Object precreation is skipped on the OST with + * max_create_count=0. */ + if (d->opd_pre_max_create_count == 0) + msfs->os_state |= OS_STATE_NOPRECREATE; + else + msfs->os_state &= ~OS_STATE_NOPRECREATE; + } out: wake_up(&d->opd_pre_user_waitq); } @@ -1012,9 +1087,27 @@ int osp_init_pre_fid(struct osp_device *osp) LASSERT(osp->opd_pre != NULL); - /* Return if last_used fid has been initialized */ + /* 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) { @@ -1080,6 +1173,8 @@ 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; @@ -1089,6 +1184,12 @@ 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); } @@ -1102,6 +1203,10 @@ 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, @@ -1118,30 +1223,41 @@ 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); + /* 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(&env, d)) { + l_wait_event(d->opd_pre_waitq, + !osp_precreate_running(d), &lwi2); continue; } - osp_statfs_update(d); + 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; + } + } /* - * Clean up orphans or recreate missing objects. - */ - rc = osp_precreate_cleanup_orphans(&env, d); - if (rc != 0) - continue; - /* * connected, can handle precreates now */ while (osp_precreate_running(d)) { @@ -1160,7 +1276,11 @@ 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 @@ -1171,7 +1291,7 @@ 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)); @@ -1227,7 +1347,8 @@ 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 (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; @@ -1235,6 +1356,7 @@ static int osp_precreate_ready_condition(const struct lu_env *env, 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) @@ -1251,11 +1373,12 @@ static int osp_precreate_timeout_condition(void *data) struct osp_device *d = data; CDEBUG(D_HA, "%s: slow creates, last="DFID", next="DFID", " - "reserved="LPU64", syn_changes=%lu, " - "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, - d->opd_syn_changes, 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; @@ -1287,9 +1410,9 @@ 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; @@ -1297,6 +1420,10 @@ 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 @@ -1311,12 +1438,12 @@ int osp_precreate_reserve(const struct lu_env *env, struct osp_device *d) * 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); } @@ -1344,16 +1471,18 @@ 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; } @@ -1362,9 +1491,9 @@ 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; } @@ -1401,6 +1530,7 @@ 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); @@ -1410,6 +1540,20 @@ 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--; @@ -1425,7 +1569,8 @@ 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; @@ -1466,8 +1611,6 @@ 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); @@ -1478,6 +1621,13 @@ 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; @@ -1505,8 +1655,23 @@ 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) @@ -1526,9 +1691,6 @@ 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); @@ -1536,32 +1698,73 @@ 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_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 @@ -1580,36 +1783,18 @@ 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; - + struct ptlrpc_thread *thread = &d->opd_pre_thread; ENTRY; - cfs_timer_disarm(&d->opd_statfs_timer); - - if (d->opd_pre == NULL) - RETURN_EXIT; + del_timer(&d->opd_statfs_timer); - thread = &d->opd_pre_thread; - - thread->t_flags = SVC_STOPPING; - wake_up(&d->opd_pre_waitq); - - wait_event(thread->t_ctl_waitq, thread->t_flags & SVC_STOPPED); - - OBD_FREE_PTR(d->opd_pre); - d->opd_pre = NULL; + 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; } -