X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fosp%2Fosp_precreate.c;h=1b77f3006d9d8bf03978668b7fd4ae6cf6525462;hp=db6851e1e2145ad9236cfde4ab837c9a6cd3dea3;hb=63e17799a369e2ff0b140fd41dc5d7d8656d2bf0;hpb=4014ddbb235043e3adc882c6a6df10e7887a7835 diff --git a/lustre/osp/osp_precreate.c b/lustre/osp/osp_precreate.c index db6851e..1b77f30 100644 --- a/lustre/osp/osp_precreate.c +++ b/lustre/osp/osp_precreate.c @@ -27,7 +27,6 @@ */ /* * This file is part of Lustre, http://www.lustre.org/ - * Lustre is a trademark of Sun Microsystems, Inc. * * lustre/osp/osp_precreate.c * @@ -56,8 +55,7 @@ * */ -/* - ** +/** * Check whether statfs data is expired * * OSP device caches statfs data for the target, the function checks @@ -78,25 +76,49 @@ static inline int osp_statfs_need_update(struct osp_device *d) * * each time OSP gets connected to OST, we should start from precreation cleanup */ -static inline bool osp_precreate_running(struct osp_device *d) +static void osp_statfs_timer_cb(cfs_timer_cb_arg_t data) { - return !!(d->opd_pre_thread.t_flags & SVC_RUNNING); + struct osp_device *d = cfs_from_timer(d, data, opd_statfs_timer); + + LASSERT(d); + /* invalidate statfs data so osp_precreate_thread() can refresh */ + d->opd_statfs_fresh_till = ktime_sub_ns(ktime_get(), NSEC_PER_SEC); + if (d->opd_pre_task) + wake_up(&d->opd_pre_waitq); } -static inline bool osp_precreate_stopped(struct osp_device *d) +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) { - return !!(d->opd_pre_thread.t_flags & SVC_STOPPED); + CDEBUG(D_INFO, "%s: Updating status = %d\n", d->opd_obd->obd_name, rc); + if (rc) + d->opd_pre_status = rc; + else + osp_pre_update_msfs(d, msfs); + + wake_up_all(&d->opd_pre_user_waitq); } -static void osp_statfs_timer_cb(cfs_timer_cb_arg_t data) +/* Pass in the old statfs data in case the limits have changed */ +void osp_pre_update_status(struct osp_device *d, int rc) { - struct osp_device *d = cfs_from_timer(d, data, opd_statfs_timer); - - LASSERT(d); - if (osp_precreate_running(d)) - wake_up(&d->opd_pre_waitq); + osp_pre_update_status_msfs(d, &d->opd_statfs, rc); } + /** * RPC interpret callback for OST_STATFS RPC * @@ -119,12 +141,13 @@ static int osp_statfs_interpret(const struct lu_env *env, union ptlrpc_async_args *aa = args; struct obd_import *imp = req->rq_import; struct obd_statfs *msfs; + struct obd_statfs *sfs; 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); @@ -135,10 +158,10 @@ static int osp_statfs_interpret(const struct lu_env *env, if (msfs == NULL) GOTO(out, rc = -EPROTO); - d->opd_statfs = *msfs; - if (d->opd_pre) - osp_pre_update_status(d, rc); + osp_pre_update_status_msfs(d, msfs, 0); + else + d->opd_statfs = *msfs; /* schedule next update */ maxage_ns = d->opd_statfs_maxage * NSEC_PER_SEC; @@ -147,14 +170,20 @@ static int osp_statfs_interpret(const struct lu_env *env, jiffies + cfs_time_seconds(d->opd_statfs_maxage)); d->opd_statfs_update_in_progress = 0; - CDEBUG(D_CACHE, "updated statfs %p\n", d); + sfs = &d->opd_statfs; + CDEBUG(D_CACHE, "%s (%p): %llu blocks, %llu free, %llu avail, " + "%u bsize, %u reserved mb low, %u reserved mb high," + "%llu files, %llu free files\n", d->opd_obd->obd_name, d, + sfs->os_blocks, sfs->os_bfree, sfs->os_bavail, sfs->os_bsize, + d->opd_reserved_mb_low, d->opd_reserved_mb_high, + sfs->os_files, sfs->os_ffree); RETURN(0); out: /* 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)) + if (d->opd_pre && d->opd_pre_task) wake_up(&d->opd_pre_waitq); if (req->rq_import_generation == imp->imp_generation) @@ -206,7 +235,7 @@ static int osp_statfs_update(const struct lu_env *env, struct osp_device *d) ptlrpc_at_set_req_timeout(req); req->rq_interpret_reply = osp_statfs_interpret; - aa = ptlrpc_req_async_args(req); + aa = ptlrpc_req_async_args(aa, req); aa->pointer_arg[0] = d; /* @@ -236,8 +265,8 @@ static int osp_statfs_update(const struct lu_env *env, struct osp_device *d) 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); } + dt_trans_stop(env, d->opd_storage, th); } out: @@ -309,8 +338,8 @@ 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_create_count / 2) && - (d->opd_pre_status == 0)); + return ((window - d->opd_pre_reserved < d->opd_pre_create_count / 2 || + d->opd_force_creation) && (d->opd_pre_status == 0)); } /** @@ -567,9 +596,9 @@ 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 @@ -604,9 +633,11 @@ static int osp_precreate_send(const struct lu_env *env, struct osp_device *d) if (req == NULL) RETURN(-ENOMEM); req->rq_request_portal = OST_CREATE_PORTAL; - /* we should not resend create request - anyway we will have delorphan - * and kill these objects */ - req->rq_no_delay = req->rq_no_resend = 1; + + /* Delorphan happens only with a first MDT-OST connect. resend/replay + * handles objects creation on reconnects, no need to do delorhpan + * in this case. + */ rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE); if (rc) { @@ -615,7 +646,9 @@ static int osp_precreate_send(const struct lu_env *env, struct osp_device *d) } spin_lock(&d->opd_pre_lock); - if (d->opd_pre_create_count > d->opd_pre_max_create_count / 2) + if (d->opd_force_creation) + d->opd_pre_create_count = OST_MIN_PRECREATE; + else 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); @@ -625,15 +658,9 @@ 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 @@ -654,9 +681,11 @@ static int osp_precreate_send(const struct lu_env *env, struct osp_device *d) if (rc) { CERROR("%s: can't precreate: rc = %d\n", d->opd_obd->obd_name, rc); + if (req->rq_net_err) + /* have osp_precreate_reserve() to wait for repeat */ + rc = -ENOTCONN; GOTO(out_req, rc); } - LASSERT(req->rq_transno == 0); body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY); if (body == NULL) @@ -691,6 +720,7 @@ ready: fid_to_ostid(fid, &body->oa.o_oi); d->opd_pre_last_created_fid = *fid; + d->opd_force_creation = false; spin_unlock(&d->opd_pre_lock); CDEBUG(D_HA, "%s: current precreated pool: "DFID"-"DFID"\n", @@ -699,7 +729,9 @@ ready: out_req: /* now we can wakeup all users awaiting for objects */ osp_pre_update_status(d, rc); - wake_up(&d->opd_pre_user_waitq); + + /* pause to let osp_precreate_reserve to go first */ + CFS_FAIL_TIMEOUT(OBD_FAIL_OSP_PRECREATE_PAUSE, 2); ptlrpc_req_finished(req); RETURN(rc); @@ -715,12 +747,13 @@ out_req: * * \param[in] env LU environment provided by the caller * \param[in] d OSP device + * \param[in] update update or not update last used fid * * \retval 0 on success * \retval negative negated errno on error **/ static int osp_get_lastfid_from_ost(const struct lu_env *env, - struct osp_device *d) + struct osp_device *d, bool update) { struct ptlrpc_request *req = NULL; struct obd_import *imp; @@ -756,12 +789,11 @@ static int osp_get_lastfid_from_ost(const struct lu_env *env, rc = ptlrpc_queue_wait(req); if (rc) { - /* bad-bad OST.. let sysadm sort this out */ - if (rc == -ENOTSUPP) { - CERROR("%s: server does not support FID: rc = %d\n", - d->opd_obd->obd_name, -ENOTSUPP); - } - ptlrpc_set_import_active(imp, 0); + /* -EFAULT means reading LAST_FID failed (see ofd_get_info_hld), + * let sysadm sort this * out. + */ + if (rc == -EFAULT) + ptlrpc_set_import_active(imp, 0); GOTO(out, rc); } @@ -779,9 +811,16 @@ static int osp_get_lastfid_from_ost(const struct lu_env *env, /* Only update the last used fid, if the OST has objects for * this sequence, i.e. fid_oid > 0 */ - if (fid_oid(last_fid) > 0) + if (fid_oid(last_fid) > 0 && update) d->opd_last_used_fid = *last_fid; + if (fid_oid(last_fid) == 0 && + fid_seq(last_fid) == fid_seq(&d->opd_last_used_fid)) { + /* reformatted OST, it requires creation request + * to recreate objects + */ + d->opd_force_creation = true; + } CDEBUG(D_HA, "%s: Got last_fid "DFID"\n", d->opd_obd->obd_name, PFID(last_fid)); @@ -815,7 +854,6 @@ static int osp_precreate_cleanup_orphans(struct lu_env *env, struct ptlrpc_request *req = NULL; struct obd_import *imp; struct ost_body *body; - struct l_wait_info lwi = { 0 }; int update_status = 0; int rc; int diff; @@ -823,6 +861,15 @@ static int osp_precreate_cleanup_orphans(struct lu_env *env, ENTRY; /* + * Do cleanup orphans only with a first connection, after that + * all precreate requests uses resend/replay flags to support OST + * failover/reconnect. + */ + if (d->opd_cleanup_orphans_done) { + rc = osp_get_lastfid_from_ost(env, d, false); + RETURN(0); + } + /* * 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 @@ -840,22 +887,24 @@ 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_pre_reserved && d->opd_recovery_completed) || - !osp_precreate_running(d) || d->opd_got_disconnected, - &lwi); - if (!osp_precreate_running(d) || d->opd_got_disconnected) + wait_event_idle(d->opd_pre_waitq, + (!d->opd_pre_reserved && d->opd_recovery_completed) || + !d->opd_pre_task || d->opd_got_disconnected); + if (!d->opd_pre_task || 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)); + OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_DELAY_DELORPHAN, cfs_fail_val); + *last_fid = d->opd_last_used_fid; /* The OSP should already get the valid seq now */ LASSERT(!fid_is_zero(last_fid)); - if (fid_oid(&d->opd_last_used_fid) < 2) { + if (fid_oid(&d->opd_last_used_fid) < 2 || + OBD_FAIL_CHECK(OBD_FAIL_OSP_GET_LAST_FID)) { /* lastfid looks strange... ask OST */ - rc = osp_get_lastfid_from_ost(env, d); + rc = osp_get_lastfid_from_ost(env, d, true); if (rc) GOTO(out, rc); } @@ -929,6 +978,7 @@ out: if (req) ptlrpc_req_finished(req); + /* * If rc is zero, the pre-creation window should have been emptied. * Since waking up the herd would be useless without pre-created @@ -945,12 +995,13 @@ out: * this OSP isn't quite functional yet */ osp_pre_update_status(d, rc); } else { - wake_up(&d->opd_pre_user_waitq); + wake_up_all(&d->opd_pre_user_waitq); } } else { spin_lock(&d->opd_pre_lock); d->opd_pre_recovering = 0; spin_unlock(&d->opd_pre_lock); + d->opd_cleanup_orphans_done = true; } RETURN(rc); @@ -964,99 +1015,106 @@ 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; - - d->opd_pre_status = rc; - if (rc) - goto out; - - 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_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_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); + 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); + } + + available_mb = (msfs->os_bavail * (msfs->os_bsize >> 10)) >> 10; + if (msfs->os_ffree < reserved_ino_low) + msfs->os_state |= OS_STATFS_ENOINO; + else if (msfs->os_ffree <= reserved_ino_high) + msfs->os_state |= old_state & OS_STATFS_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_STATFS_ENOSPC; + else if (available_mb <= d->opd_reserved_mb_high) + msfs->os_state |= old_state & OS_STATFS_ENOSPC; + /* else don't clear flags in new msfs->os_state sent from OST */ + + if (msfs->os_state & (OS_STATFS_ENOINO | OS_STATFS_ENOSPC)) { + d->opd_pre_status = -ENOSPC; + if (!(old_state & (OS_STATFS_ENOINO | OS_STATFS_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_STATFS_ENOINO | OS_STATFS_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); - /* 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; + 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_STATFS_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; } /** @@ -1149,6 +1207,11 @@ out: RETURN(rc); } +struct opt_args { + struct osp_device *opta_dev; + struct lu_env opta_env; + struct completion *opta_started; +}; /** * The core of precreate functionality * @@ -1167,59 +1230,40 @@ out: * \retval 0 on success * \retval negative negated errno on error */ -static int osp_precreate_thread(void *_arg) +static int osp_precreate_thread(void *_args) { - 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; + struct opt_args *args = _args; + struct osp_device *d = args->opta_dev; + struct lu_env *env = &args->opta_env; int rc; ENTRY; - rc = lu_env_init(&env, d->opd_dt_dev.dd_lu_dev.ld_type->ldt_ctx_tags); - if (rc) { - CERROR("%s: init env error: rc = %d\n", d->opd_obd->obd_name, - 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); - } - - spin_lock(&d->opd_pre_lock); - thread->t_flags = SVC_RUNNING; - spin_unlock(&d->opd_pre_lock); - wake_up(&thread->t_ctl_waitq); - - while (osp_precreate_running(d)) { + complete(args->opta_started); + while (!kthread_should_stop()) { /* * need to be connected to OST */ - while (osp_precreate_running(d)) { + while (!kthread_should_stop()) { 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, - &lwi); + wait_event_idle(d->opd_pre_waitq, + kthread_should_stop() || + d->opd_new_connection); if (!d->opd_new_connection) continue; + OBD_FAIL_TIMEOUT(OBD_FAIL_OSP_CON_EVENT_DELAY, + cfs_fail_val); d->opd_new_connection = 0; d->opd_got_disconnected = 0; break; } - if (!osp_precreate_running(d)) + if (kthread_should_stop()) break; if (d->opd_pre) { @@ -1239,9 +1283,13 @@ static int osp_precreate_thread(void *_arg) } } - if (osp_statfs_update(&env, d)) { - l_wait_event(d->opd_pre_waitq, - !osp_precreate_running(d), &lwi2); + if (osp_statfs_update(env, d)) { + if (wait_event_idle_timeout(d->opd_pre_waitq, + kthread_should_stop(), + cfs_time_seconds(5)) == 0) + l_wait_event_abortable( + d->opd_pre_waitq, + kthread_should_stop()); continue; } @@ -1249,10 +1297,9 @@ static int osp_precreate_thread(void *_arg) /* * Clean up orphans or recreate missing objects. */ - rc = osp_precreate_cleanup_orphans(&env, d); + rc = osp_precreate_cleanup_orphans(env, d); if (rc != 0) { - schedule_timeout_interruptible( - msecs_to_jiffies(MSEC_PER_SEC)); + schedule_timeout_interruptible(cfs_time_seconds(1)); continue; } } @@ -1260,14 +1307,14 @@ static int osp_precreate_thread(void *_arg) /* * connected, can handle precreates now */ - while (osp_precreate_running(d)) { - l_wait_event(d->opd_pre_waitq, - !osp_precreate_running(d) || - osp_precreate_near_empty(&env, d) || - osp_statfs_need_update(d) || - d->opd_got_disconnected, &lwi); - - if (!osp_precreate_running(d)) + while (!kthread_should_stop()) { + wait_event_idle(d->opd_pre_waitq, + kthread_should_stop() || + osp_precreate_near_empty(env, d) || + osp_statfs_need_update(d) || + d->opd_got_disconnected); + + if (kthread_should_stop()) break; /* something happened to the connection @@ -1276,32 +1323,37 @@ static int osp_precreate_thread(void *_arg) break; if (osp_statfs_need_update(d)) - if (osp_statfs_update(&env, d)) + if (osp_statfs_update(env, d)) break; if (d->opd_pre == NULL) continue; + if (OBD_FAIL_CHECK(OBD_FAIL_OSP_GET_LAST_FID)) { + d->opd_pre_recovering = 1; + break; + } + /* To avoid handling different seq in precreate/orphan * cleanup, it will hold precreate until current seq is * used up. */ - if (unlikely(osp_precreate_end_seq(&env, d) && - !osp_create_end_seq(&env, d))) + if (unlikely(osp_precreate_end_seq(env, d) && + !osp_create_end_seq(env, d))) continue; - if (unlikely(osp_precreate_end_seq(&env, d) && - osp_create_end_seq(&env, d))) { + if (unlikely(osp_precreate_end_seq(env, d) && + osp_create_end_seq(env, d))) { 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)); - rc = osp_precreate_rollover_new_seq(&env, d); + rc = osp_precreate_rollover_new_seq(env, d); if (rc) continue; } - if (osp_precreate_near_empty(&env, d)) { - rc = osp_precreate_send(&env, d); + if (osp_precreate_near_empty(env, d)) { + rc = osp_precreate_send(env, d); /* osp_precreate_send() sets opd_pre_status * in case of error, that prevent the using of * failed device. */ @@ -1314,9 +1366,8 @@ static int osp_precreate_thread(void *_arg) } } - thread->t_flags = SVC_STOPPED; - lu_env_fini(&env); - wake_up(&thread->t_ctl_waitq); + lu_env_fini(env); + OBD_FREE_PTR(args); RETURN(0); } @@ -1338,6 +1389,19 @@ static int osp_precreate_thread(void *_arg) static int osp_precreate_ready_condition(const struct lu_env *env, struct osp_device *d) { + /* Bail out I/O fails to OST */ + if (d->opd_pre_status != 0 && + d->opd_pre_status != -EAGAIN && + d->opd_pre_status != -ENODEV && + d->opd_pre_status != -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; + } + if (d->opd_pre_recovering) return 0; @@ -1352,38 +1416,9 @@ static int osp_precreate_ready_condition(const struct lu_env *env, d->opd_pre_status == -ENOSPC) return 1; - /* Bail out I/O fails to OST */ - if (d->opd_pre_status != 0 && - d->opd_pre_status != -EAGAIN && - d->opd_pre_status != -ENODEV && - d->opd_pre_status != -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; } -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, 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; -} - /** * Reserve object in precreate pool * @@ -1408,10 +1443,10 @@ static int osp_precreate_timeout_condition(void *data) * \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) +int osp_precreate_reserve(const struct lu_env *env, struct osp_device *d, + bool can_block) { time64_t expire = ktime_get_seconds() + obd_timeout; - struct l_wait_info lwi; int precreated, rc, synced = 0; ENTRY; @@ -1478,8 +1513,9 @@ int osp_precreate_reserve(const struct lu_env *env, struct osp_device *d) synced = 1; } if (atomic_read(&d->opd_sync_rpcs_in_progress)) { - /* just wait till destroys are done */ - /* see l_wait_even() few lines below */ + /* just wait till destroys are done + * see wait_event_idle_timeout() below + */ } if (atomic_read(&d->opd_sync_changes) + atomic_read(&d->opd_sync_rpcs_in_progress) == 0) { @@ -1491,15 +1527,37 @@ 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(cfs_time_seconds(obd_timeout), - osp_precreate_timeout_condition, d); if (ktime_get_seconds() >= expire) { rc = -ETIMEDOUT; break; } - l_wait_event(d->opd_pre_user_waitq, - osp_precreate_ready_condition(env, d), &lwi); + if (!can_block) { + LASSERT(d->opd_pre); + rc = -ENOBUFS; + break; + } + + CDEBUG(D_INFO, "%s: Sleeping on objects\n", + d->opd_obd->obd_name); + if (wait_event_idle_timeout( + d->opd_pre_user_waitq, + osp_precreate_ready_condition(env, d), + cfs_time_seconds(obd_timeout)) == 0) { + 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); + } else { + CDEBUG(D_INFO, "%s: Waked up, status=%d\n", + d->opd_obd->obd_name, d->opd_pre_status); + } } RETURN(rc); @@ -1621,12 +1679,12 @@ 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 !! + * and if we don't resend, then client will also get -EAGAIN !! * (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 + * -EAGAIN/-EAGAIN upon return from ptlrpc_queue_wait() at the * end of this routine */ req->rq_no_resend = req->rq_no_delay = 1; @@ -1656,14 +1714,14 @@ int osp_object_truncate(const struct lu_env *env, struct dt_object *dt, rc = ptlrpc_queue_wait(req); if (rc) { - /* -EWOULDBLOCK/-EAGAIN means OST is unreachable at the moment + /* -EAGAIN/-EWOULDBLOCK 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) { + if (rc == -EAGAIN) { rc = -EINPROGRESS; CDEBUG(D_HA, "returning -EINPROGRESS instead of " "-EWOULDBLOCK/-EAGAIN to allow Client to " @@ -1713,6 +1771,8 @@ int osp_init_precreate(struct osp_device *d) d->opd_pre_max_create_count = OST_MAX_PRECREATE; d->opd_reserved_mb_high = 0; d->opd_reserved_mb_low = 0; + d->opd_cleanup_orphans_done = false; + d->opd_force_creation = false; RETURN(0); } @@ -1741,15 +1801,15 @@ void osp_precreate_fini(struct osp_device *d) int osp_init_statfs(struct osp_device *d) { - struct l_wait_info lwi = { 0 }; - struct task_struct *task; + struct task_struct *task; + struct opt_args *args; + DECLARE_COMPLETION_ONSTACK(started); + int rc; ENTRY; spin_lock_init(&d->opd_pre_lock); init_waitqueue_head(&d->opd_pre_waitq); - thread_set_flags(&d->opd_pre_thread, SVC_INIT); - init_waitqueue_head(&d->opd_pre_thread.t_ctl_waitq); /* * Initialize statfs-related things @@ -1766,35 +1826,48 @@ int osp_init_statfs(struct osp_device *d) if (d->opd_storage->dd_rdonly) RETURN(0); + OBD_ALLOC_PTR(args); + if (!args) + RETURN(0); + args->opta_dev = d; + args->opta_started = &started; + rc = lu_env_init(&args->opta_env, + d->opd_dt_dev.dd_lu_dev.ld_type->ldt_ctx_tags); + if (rc) { + CERROR("%s: init env error: rc = %d\n", d->opd_obd->obd_name, + rc); + OBD_FREE_PTR(args); + RETURN(0); + } + /* * start thread handling precreation and statfs updates */ - task = kthread_run(osp_precreate_thread, d, - "osp-pre-%u-%u", d->opd_index, d->opd_group); + task = kthread_create(osp_precreate_thread, args, + "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)); + lu_env_fini(&args->opta_env); + OBD_FREE_PTR(args); RETURN(PTR_ERR(task)); } - - l_wait_event(d->opd_pre_thread.t_ctl_waitq, - osp_precreate_running(d) || osp_precreate_stopped(d), - &lwi); + d->opd_pre_task = task; + wake_up_process(task); + wait_for_completion(&started); RETURN(0); } void osp_statfs_fini(struct osp_device *d) { - struct ptlrpc_thread *thread = &d->opd_pre_thread; + struct task_struct *task = d->opd_pre_task; ENTRY; del_timer(&d->opd_statfs_timer); - 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)); - } + d->opd_pre_task = NULL; + if (task) + kthread_stop(task); EXIT; }