[fs/lustre-release.git] / lustre / osp / osp_sync.c

/*
 * GPL HEADER START
 *
 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
 *
 * This program is free software; you can redistribute it and/or modify
 * it under the terms of the GNU General Public License version 2 only,
 * as published by the Free Software Foundation.
 *
 * This program is distributed in the hope that it will be useful, but
 * WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
 * General Public License version 2 for more details (a copy is included
 * in the LICENSE file that accompanied this code).
 *
 * You should have received a copy of the GNU General Public License
 * version 2 along with this program; If not, see
 * http://www.gnu.org/licenses/gpl-2.0.html
 *
 * GPL HEADER END
 */
/*
 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
 * Use is subject to license terms.
 *
 * Copyright (c) 2012, 2016, 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 OST Proxy Device
 *
 * Author: Alex Zhuravlev <alexey.zhuravlev@intel.com>
 * Author: Mikhail Pershin <mike.pershin@intel.com>
 */

#define DEBUG_SUBSYSTEM S_MDS

#include <linux/kthread.h>
#include <lustre_log.h>
#include <lustre_update.h>
#include "osp_internal.h"

static int osp_sync_id_traction_init(struct osp_device *d);
static void osp_sync_id_traction_fini(struct osp_device *d);
static __u64 osp_sync_id_get(struct osp_device *d, __u64 id);
static void osp_sync_remove_from_tracker(struct osp_device *d);

/*
 * this is a components of OSP implementing synchronization between MDS and OST
 * it llogs all interesting changes (currently it's uig/gid change and object
 * destroy) atomically, then makes sure changes hit OST storage
 *
 * we have 4 queues of work:
 *
 * the first queue is llog itself, once read a change is stored in 2nd queue
 * in form of RPC (but RPC isn't fired yet).
 *
 * the second queue (opd_sync_waiting_for_commit) holds changes awaiting local
 * commit. once change is committed locally it migrates onto 3rd queue.
 *
 * the third queue (opd_sync_committed_here) holds changes committed locally,
 * but not sent to OST (as the pipe can be full). once pipe becomes non-full
 * we take a change from the queue and fire corresponded RPC.
 *
 * once RPC is reported committed by OST (using regular last_committed mech.)
 * the change jumps into 4th queue (opd_sync_committed_there), now we can
 * cancel corresponded llog record and release RPC
 *
 * opd_sync_changes is a number of unread llog records (to be processed).
 * notice this number doesn't include llog records from previous boots.
 * with OSP_SYNC_THRESHOLD we try to batch processing a bit (TO BE IMPLEMENTED)
 *
 * opd_sync_rpcs_in_progress is total number of requests in above 2-4 queues.
 * we control this with OSP_MAX_RPCS_IN_PROGRESS so that OSP don't consume
 * too much memory -- how to deal with 1000th OSTs ? batching could help?
 *
 * opd_sync_rpcs_in_flight is a number of RPC in flight.
 * we control this with OSP_MAX_RPCS_IN_FLIGHT
 */

/* XXX: do math to learn reasonable threshold
 * should it be ~ number of changes fitting bulk? */

#define OSP_SYNC_THRESHOLD              10
#define OSP_MAX_RPCS_IN_FLIGHT          8
#define OSP_MAX_RPCS_IN_PROGRESS        4096

#define OSP_JOB_MAGIC           0x26112005

struct osp_job_req_args {
        /** bytes reserved for ptlrpc_replay_req() */
        struct ptlrpc_replay_async_args jra_raa;
        struct list_head                jra_committed_link;
        struct list_head                jra_in_flight_link;
        struct llog_cookie              jra_lcookie;
        __u32                           jra_magic;
};

static inline int osp_sync_running(struct osp_device *d)
{
        return !!(d->opd_sync_thread.t_flags & SVC_RUNNING);
}

/**
 * Check status: whether OSP thread has stopped
 *
 * \param[in] d         OSP device
 *
 * \retval 0            still running
 * \retval 1            stopped
 */
static inline int osp_sync_stopped(struct osp_device *d)
{
        return !!(d->opd_sync_thread.t_flags & SVC_STOPPED);
}

/*
 ** Check for new changes to sync
 *
 * \param[in] d         OSP device
 *
 * \retval 1            there are changes
 * \retval 0            there are no changes
 */
static inline int osp_sync_has_new_job(struct osp_device *d)
{
        return ((d->opd_sync_last_processed_id < d->opd_sync_last_used_id) &&
                (d->opd_sync_last_processed_id < d->opd_sync_last_committed_id))
                || (d->opd_sync_prev_done == 0);
}

static inline int osp_sync_in_flight_conflict(struct osp_device *d,
                                             struct llog_rec_hdr *h)
{
        struct osp_job_req_args *jra;
        struct ost_id            ostid;
        int                      conflict = 0;

        if (h == NULL || h->lrh_type == LLOG_GEN_REC ||
            list_empty(&d->opd_sync_in_flight_list))
                return conflict;

        memset(&ostid, 0, sizeof(ostid));
        switch (h->lrh_type) {
        case MDS_UNLINK_REC: {
                struct llog_unlink_rec *unlink = (struct llog_unlink_rec *)h;

                ostid_set_seq(&ostid, unlink->lur_oseq);
                if (ostid_set_id(&ostid, unlink->lur_oid)) {
                        CERROR("Bad %llu to set " DOSTID "\n",
                               (unsigned long long)(unlink->lur_oid),
                               POSTID(&ostid));
                        return 1;
                }
                }
                break;
        case MDS_UNLINK64_REC:
                fid_to_ostid(&((struct llog_unlink64_rec *)h)->lur_fid, &ostid);
                break;
        case MDS_SETATTR64_REC:
                ostid = ((struct llog_setattr64_rec *)h)->lsr_oi;
                break;
        default:
                LBUG();
        }

        spin_lock(&d->opd_sync_lock);
        list_for_each_entry(jra, &d->opd_sync_in_flight_list,
                            jra_in_flight_link) {
                struct ptlrpc_request   *req;
                struct ost_body         *body;

                LASSERT(jra->jra_magic == OSP_JOB_MAGIC);

                req = container_of((void *)jra, struct ptlrpc_request,
                                   rq_async_args);
                body = req_capsule_client_get(&req->rq_pill,
                                              &RMF_OST_BODY);
                LASSERT(body);

                if (memcmp(&ostid, &body->oa.o_oi, sizeof(ostid)) == 0) {
                        conflict = 1;
                        break;
                }
        }
        spin_unlock(&d->opd_sync_lock);

        return conflict;
}

static inline int osp_sync_rpcs_in_progress_low(struct osp_device *d)
{
        return atomic_read(&d->opd_sync_rpcs_in_progress) <
                d->opd_sync_max_rpcs_in_progress;
}

/**
 * Check for room in the network pipe to OST
 *
 * \param[in] d         OSP device
 *
 * \retval 1            there is room
 * \retval 0            no room, the pipe is full
 */
static inline int osp_sync_rpcs_in_flight_low(struct osp_device *d)
{
        return atomic_read(&d->opd_sync_rpcs_in_flight) <
                d->opd_sync_max_rpcs_in_flight;
}

/**
 * Wake up check for the main sync thread
 *
 * \param[in] d         OSP device
 *
 * \retval 1            time to wake up
 * \retval 0            no need to wake up
 */
static inline int osp_sync_has_work(struct osp_device *osp)
{
        /* has new/old changes and low in-progress? */
        if (osp_sync_has_new_job(osp) && osp_sync_rpcs_in_progress_low(osp) &&
            osp_sync_rpcs_in_flight_low(osp) && osp->opd_imp_connected)
                return 1;

        /* has remotely committed? */
        if (!list_empty(&osp->opd_sync_committed_there))
                return 1;

        return 0;
}

void osp_sync_check_for_work(struct osp_device *osp)
{
        if (osp_sync_has_work(osp))
                wake_up(&osp->opd_sync_waitq);
}

static inline __u64 osp_sync_correct_id(struct osp_device *d,
                                        struct llog_rec_hdr *rec)
{
        /*
         * llog use cyclic store with 32 bit lrh_id
         * so overflow lrh_id is possible. Range between
         * last_processed and last_committed is less than
         * 64745 ^ 2 and less than 2^32 - 1
         */
        __u64 correct_id = d->opd_sync_last_committed_id;

        if ((correct_id & 0xffffffffULL) < rec->lrh_id)
                correct_id -= 0x100000000ULL;

        correct_id &= ~0xffffffffULL;
        correct_id |= rec->lrh_id;

        return correct_id;
}
/**
 * Check and return ready-for-new status.
 *
 * The thread processing llog record uses this function to check whether
 * it's time to take another record and process it. The number of conditions
 * must be met: the connection should be ready, RPCs in flight not exceeding
 * the limit, the record is committed locally, etc (see the lines below).
 *
 * \param[in] d         OSP device
 * \param[in] rec       next llog record to process
 *
 * \retval 0            not ready
 * \retval 1            ready
 */
static inline int osp_sync_can_process_new(struct osp_device *d,
                                           struct llog_rec_hdr *rec)
{
        LASSERT(d);

        if (unlikely(atomic_read(&d->opd_sync_barrier) > 0))
                return 0;
        if (unlikely(osp_sync_in_flight_conflict(d, rec)))
                return 0;
        if (!osp_sync_rpcs_in_progress_low(d))
                return 0;
        if (!osp_sync_rpcs_in_flight_low(d))
                return 0;
        if (!d->opd_imp_connected)
                return 0;
        if (d->opd_sync_prev_done == 0)
                return 1;
        if (atomic_read(&d->opd_sync_changes) == 0)
                return 0;
        if (rec == NULL ||
            osp_sync_correct_id(d, rec) <= d->opd_sync_last_committed_id)
                return 1;
        return 0;
}

/**
 * Declare intention to add a new change.
 *
 * With regard to OSD API, we have to declare any changes ahead. In this
 * case we declare an intention to add a llog record representing the
 * change on the local storage.
 *
 * \param[in] env       LU environment provided by the caller
 * \param[in] o         OSP object
 * \param[in] type      type of change: MDS_UNLINK64_REC or MDS_SETATTR64_REC
 * \param[in] th        transaction handle (local)
 *
 * \retval 0            on success
 * \retval negative     negated errno on error
 */
int osp_sync_declare_add(const struct lu_env *env, struct osp_object *o,
                         llog_op_type type, struct thandle *th)
{
        struct osp_thread_info  *osi = osp_env_info(env);
        struct osp_device       *d = lu2osp_dev(o->opo_obj.do_lu.lo_dev);
        struct llog_ctxt        *ctxt;
        struct thandle          *storage_th;
        int                      rc;

        ENTRY;

        /* it's a layering violation, to access internals of th,
         * but we can do this as a sanity check, for a while */
        LASSERT(th->th_top != NULL);
        storage_th = thandle_get_sub_by_dt(env, th->th_top, d->opd_storage);
        if (IS_ERR(storage_th))
                RETURN(PTR_ERR(storage_th));

        switch (type) {
        case MDS_UNLINK64_REC:
                osi->osi_hdr.lrh_len = sizeof(struct llog_unlink64_rec);
                break;
        case MDS_SETATTR64_REC:
                osi->osi_hdr.lrh_len = sizeof(struct llog_setattr64_rec_v2);
                break;
        default:
                LBUG();
        }

        /* we want ->dt_trans_start() to allocate per-thandle structure */
        storage_th->th_tags |= LCT_OSP_THREAD;

        ctxt = llog_get_context(d->opd_obd, LLOG_MDS_OST_ORIG_CTXT);
        LASSERT(ctxt);

        rc = llog_declare_add(env, ctxt->loc_handle, &osi->osi_hdr,
                              storage_th);
        llog_ctxt_put(ctxt);

        RETURN(rc);
}

/**
 * Generate a llog record for a given change.
 *
 * Generates a llog record for the change passed. The change can be of two
 * types: unlink and setattr. The record gets an ID which later will be
 * used to track commit status of the change. For unlink changes, the caller
 * can supply a starting FID and the count of the objects to destroy. For
 * setattr the caller should apply attributes to apply.
 *
 *
 * \param[in] env       LU environment provided by the caller
 * \param[in] d         OSP device
 * \param[in] fid       fid of the object the change should be applied to
 * \param[in] type      type of change: MDS_UNLINK64_REC or MDS_SETATTR64_REC
 * \param[in] count     count of objects to destroy
 * \param[in] th        transaction handle (local)
 * \param[in] attr      attributes for setattr
 *
 * \retval 0            on success
 * \retval negative     negated errno on error
 */
static int osp_sync_add_rec(const struct lu_env *env, struct osp_device *d,
                            const struct lu_fid *fid, llog_op_type type,
                            int count, struct thandle *th,
                            const struct lu_attr *attr)
{
        struct osp_thread_info  *osi = osp_env_info(env);
        struct llog_ctxt        *ctxt;
        struct osp_txn_info     *txn;
        struct thandle          *storage_th;
        int                      rc;

        ENTRY;

        /* it's a layering violation, to access internals of th,
         * but we can do this as a sanity check, for a while */
        LASSERT(th->th_top != NULL);
        storage_th = thandle_get_sub_by_dt(env, th->th_top, d->opd_storage);
        if (IS_ERR(storage_th))
                RETURN(PTR_ERR(storage_th));

        switch (type) {
        case MDS_UNLINK64_REC:
                osi->osi_hdr.lrh_len = sizeof(osi->osi_unlink);
                osi->osi_hdr.lrh_type = MDS_UNLINK64_REC;
                osi->osi_unlink.lur_fid  = *fid;
                osi->osi_unlink.lur_count = count;
                break;
        case MDS_SETATTR64_REC:
                rc = fid_to_ostid(fid, &osi->osi_oi);
                LASSERT(rc == 0);
                osi->osi_hdr.lrh_len = sizeof(osi->osi_setattr);
                osi->osi_hdr.lrh_type = MDS_SETATTR64_REC;
                osi->osi_setattr.lsr_oi  = osi->osi_oi;
                LASSERT(attr);
                osi->osi_setattr.lsr_uid = attr->la_uid;
                osi->osi_setattr.lsr_gid = attr->la_gid;
                osi->osi_setattr.lsr_projid = attr->la_projid;
                osi->osi_setattr.lsr_valid =
                        ((attr->la_valid & LA_UID) ? OBD_MD_FLUID : 0) |
                        ((attr->la_valid & LA_GID) ? OBD_MD_FLGID : 0) |
                        ((attr->la_valid & LA_PROJID) ? OBD_MD_FLPROJID : 0);
                break;
        default:
                LBUG();
        }

        txn = osp_txn_info(&storage_th->th_ctx);
        LASSERT(txn);

        txn->oti_current_id = osp_sync_id_get(d, txn->oti_current_id);
        osi->osi_hdr.lrh_id = (txn->oti_current_id & 0xffffffffULL);
        ctxt = llog_get_context(d->opd_obd, LLOG_MDS_OST_ORIG_CTXT);
        if (ctxt == NULL)
                RETURN(-ENOMEM);

        rc = llog_add(env, ctxt->loc_handle, &osi->osi_hdr, &osi->osi_cookie,
                      storage_th);
        llog_ctxt_put(ctxt);

        if (likely(rc >= 0)) {
                CDEBUG(D_OTHER, "%s: new record "DFID":%x.%u: rc = %d\n",
                       d->opd_obd->obd_name,
                       PFID(&osi->osi_cookie.lgc_lgl.lgl_oi.oi_fid),
                       osi->osi_cookie.lgc_lgl.lgl_ogen,
                       osi->osi_cookie.lgc_index, rc);
                atomic_inc(&d->opd_sync_changes);
        }
        /* return 0 always here, error case just cause no llog record */
        RETURN(0);
}

int osp_sync_add(const struct lu_env *env, struct osp_object *o,
                 llog_op_type type, struct thandle *th,
                 const struct lu_attr *attr)
{
        return osp_sync_add_rec(env, lu2osp_dev(o->opo_obj.do_lu.lo_dev),
                                lu_object_fid(&o->opo_obj.do_lu), type, 1,
                                th, attr);
}

int osp_sync_gap(const struct lu_env *env, struct osp_device *d,
                        struct lu_fid *fid, int lost, struct thandle *th)
{
        return osp_sync_add_rec(env, d, fid, MDS_UNLINK64_REC, lost, th, NULL);
}

/*
 * it's quite obvious we can't maintain all the structures in the memory:
 * while OST is down, MDS can be processing thousands and thousands of unlinks
 * filling persistent llogs and in-core respresentation
 *
 * this doesn't scale at all. so we need basically the following:
 * a) destroy/setattr append llog records
 * b) once llog has grown to X records, we process first Y committed records
 *
 *  once record R is found via llog_process(), it becomes committed after any
 *  subsequent commit callback (at the most)
 */

/**
 * ptlrpc commit callback.
 *
 * The callback is called by PTLRPC when a RPC is reported committed by the
 * target (OST). We register the callback for the every RPC applying a change
 * from the llog. This way we know then the llog records can be cancelled.
 * Notice the callback can be called when OSP is finishing. We can detect this
 * checking that actual transno in the request is less or equal of known
 * committed transno (see osp_sync_process_committed() for the details).
 * XXX: this is pretty expensive and can be improved later using batching.
 *
 * \param[in] req       request
 */
static void osp_sync_request_commit_cb(struct ptlrpc_request *req)
{
        struct osp_device *d = req->rq_cb_data;
        struct osp_job_req_args *jra;

        CDEBUG(D_HA, "commit req %p, transno %llu\n", req, req->rq_transno);

        if (unlikely(req->rq_transno == 0))
                return;

        /* do not do any opd_sync_rpcs_* accounting here
         * it's done in osp_sync_interpret sooner or later */
        LASSERT(d);

        jra = ptlrpc_req_async_args(req);
        LASSERT(jra->jra_magic == OSP_JOB_MAGIC);
        LASSERT(list_empty(&jra->jra_committed_link));

        ptlrpc_request_addref(req);

        spin_lock(&d->opd_sync_lock);
        list_add(&jra->jra_committed_link, &d->opd_sync_committed_there);
        spin_unlock(&d->opd_sync_lock);

        /* XXX: some batching wouldn't hurt */
        wake_up(&d->opd_sync_waitq);
}

/**
 * RPC interpretation callback.
 *
 * The callback is called by ptlrpc when RPC is replied. Now we have to decide
 * whether we should:
 *  - put request on a special list to wait until it's committed by the target,
 *    if the request is successful
 *  - schedule llog record cancel if no target object is found
 *  - try later (essentially after reboot) in case of unexpected error
 *
 * \param[in] env       LU environment provided by the caller
 * \param[in] req       request replied
 * \param[in] aa        callback data
 * \param[in] rc        result of RPC
 *
 * \retval 0            always
 */
static int osp_sync_interpret(const struct lu_env *env,
                              struct ptlrpc_request *req, void *aa, int rc)
{
        struct osp_device *d = req->rq_cb_data;
        struct osp_job_req_args *jra = aa;

        if (jra->jra_magic != OSP_JOB_MAGIC) {
                DEBUG_REQ(D_ERROR, req, "bad magic %u\n", jra->jra_magic);
                LBUG();
        }
        LASSERT(d);

        CDEBUG(D_HA, "reply req %p/%d, rc %d, transno %u\n", req,
               atomic_read(&req->rq_refcount),
               rc, (unsigned) req->rq_transno);
        LASSERT(rc || req->rq_transno);

        if (rc == -ENOENT) {
                /*
                 * we tried to destroy object or update attributes,
                 * but object doesn't exist anymore - cancell llog record
                 */
                LASSERT(req->rq_transno == 0);
                LASSERT(list_empty(&jra->jra_committed_link));

                ptlrpc_request_addref(req);

                spin_lock(&d->opd_sync_lock);
                list_add(&jra->jra_committed_link,
                         &d->opd_sync_committed_there);
                spin_unlock(&d->opd_sync_lock);

                wake_up(&d->opd_sync_waitq);
        } else if (rc) {
                struct obd_import *imp = req->rq_import;
                /*
                 * error happened, we'll try to repeat on next boot ?
                 */
                LASSERTF(req->rq_transno == 0 ||
                         req->rq_import_generation < imp->imp_generation,
                         "transno %llu, rc %d, gen: req %d, imp %d\n",
                         req->rq_transno, rc, req->rq_import_generation,
                         imp->imp_generation);
                if (req->rq_transno == 0) {
                        /* this is the last time we see the request
                         * if transno is not zero, then commit cb
                         * will be called at some point */
                        LASSERT(atomic_read(&d->opd_sync_rpcs_in_progress) > 0);
                        atomic_dec(&d->opd_sync_rpcs_in_progress);
                }

                wake_up(&d->opd_sync_waitq);
        } else if (d->opd_pre != NULL &&
                   unlikely(d->opd_pre_status == -ENOSPC)) {
                /*
                 * if current status is -ENOSPC (lack of free space on OST)
                 * then we should poll OST immediately once object destroy
                 * is replied
                 */
                osp_statfs_need_now(d);
        }

        spin_lock(&d->opd_sync_lock);
        list_del_init(&jra->jra_in_flight_link);
        spin_unlock(&d->opd_sync_lock);
        LASSERT(atomic_read(&d->opd_sync_rpcs_in_flight) > 0);
        atomic_dec(&d->opd_sync_rpcs_in_flight);
        if (unlikely(atomic_read(&d->opd_sync_barrier) > 0))
                wake_up(&d->opd_sync_barrier_waitq);
        CDEBUG(D_OTHER, "%s: %d in flight, %d in progress\n",
               d->opd_obd->obd_name, atomic_read(&d->opd_sync_rpcs_in_flight),
               atomic_read(&d->opd_sync_rpcs_in_progress));

        osp_sync_check_for_work(d);

        return 0;
}

/*
 ** Add request to ptlrpc queue.
 *
 * This is just a tiny helper function to put the request on the sending list
 *
 * \param[in] d         OSP device
 * \param[in] llh       llog handle where the record is stored
 * \param[in] h         llog record
 * \param[in] req       request
 */
static void osp_sync_send_new_rpc(struct osp_device *d,
                                  struct llog_handle *llh,
                                  struct llog_rec_hdr *h,
                                  struct ptlrpc_request *req)
{
        struct osp_job_req_args *jra;

        LASSERT(atomic_read(&d->opd_sync_rpcs_in_flight) <=
                d->opd_sync_max_rpcs_in_flight);

        jra = ptlrpc_req_async_args(req);
        jra->jra_magic = OSP_JOB_MAGIC;
        jra->jra_lcookie.lgc_lgl = llh->lgh_id;
        jra->jra_lcookie.lgc_subsys = LLOG_MDS_OST_ORIG_CTXT;
        jra->jra_lcookie.lgc_index = h->lrh_index;
        INIT_LIST_HEAD(&jra->jra_committed_link);
        spin_lock(&d->opd_sync_lock);
        list_add_tail(&jra->jra_in_flight_link, &d->opd_sync_in_flight_list);
        spin_unlock(&d->opd_sync_lock);

        ptlrpcd_add_req(req);
}


/**
 * Allocate and prepare RPC for a new change.
 *
 * The function allocates and initializes an RPC which will be sent soon to
 * apply the change to the target OST. The request is initialized from the
 * llog record passed. Notice only the fields common to all type of changes
 * are initialized.
 *
 * \param[in] d         OSP device
 * \param[in] op        type of the change
 * \param[in] format    request format to be used
 *
 * \retval pointer              new request on success
 * \retval ERR_PTR(errno)       on error
 */
static struct ptlrpc_request *osp_sync_new_job(struct osp_device *d,
                                               ost_cmd_t op,
                                               const struct req_format *format)
{
        struct ptlrpc_request   *req;
        struct obd_import       *imp;
        int                      rc;

        /* Prepare the request */
        imp = d->opd_obd->u.cli.cl_import;
        LASSERT(imp);

        if (OBD_FAIL_CHECK(OBD_FAIL_OSP_CHECK_ENOMEM))
                RETURN(ERR_PTR(-ENOMEM));

        req = ptlrpc_request_alloc(imp, format);
        if (req == NULL)
                RETURN(ERR_PTR(-ENOMEM));

        rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, op);
        if (rc) {
                ptlrpc_req_finished(req);
                return ERR_PTR(rc);
        }

        req->rq_interpret_reply = osp_sync_interpret;
        req->rq_commit_cb = osp_sync_request_commit_cb;
        req->rq_cb_data = d;

        ptlrpc_request_set_replen(req);

        return req;
}

/**
 * Generate a request for setattr change.
 *
 * The function prepares a new RPC, initializes it with setattr specific
 * bits and send the RPC.
 *
 * \param[in] d         OSP device
 * \param[in] llh       llog handle where the record is stored
 * \param[in] h         llog record
 *
 * \retval 0            on success
 * \retval 1            on invalid record
 * \retval negative     negated errno on error
 */
static int osp_sync_new_setattr_job(struct osp_device *d,
                                    struct llog_handle *llh,
                                    struct llog_rec_hdr *h)
{
        struct llog_setattr64_rec       *rec = (struct llog_setattr64_rec *)h;
        struct ptlrpc_request           *req;
        struct ost_body                 *body;

        ENTRY;
        LASSERT(h->lrh_type == MDS_SETATTR64_REC);

        if (OBD_FAIL_CHECK(OBD_FAIL_OSP_CHECK_INVALID_REC))
                RETURN(1);

        /* lsr_valid can only be 0 or HAVE OBD_MD_{FLUID, FLGID, FLPROJID} set,
         * so no bits other than these should be set. */
        if ((rec->lsr_valid & ~(OBD_MD_FLUID | OBD_MD_FLGID |
            OBD_MD_FLPROJID)) != 0) {
                CERROR("%s: invalid setattr record, lsr_valid:%llu\n",
                        d->opd_obd->obd_name, rec->lsr_valid);
                /* return 1 on invalid record */
                RETURN(1);
        }

        req = osp_sync_new_job(d, OST_SETATTR, &RQF_OST_SETATTR);
        if (IS_ERR(req))
                RETURN(PTR_ERR(req));

        body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
        LASSERT(body);
        body->oa.o_oi = rec->lsr_oi;
        body->oa.o_uid = rec->lsr_uid;
        body->oa.o_gid = rec->lsr_gid;
        body->oa.o_valid = OBD_MD_FLGROUP | OBD_MD_FLID;
        if (h->lrh_len > sizeof(struct llog_setattr64_rec))
                body->oa.o_projid = ((struct llog_setattr64_rec_v2 *)
                                      rec)->lsr_projid;

        /* old setattr record (prior 2.6.0) doesn't have 'valid' stored,
         * we assume that both UID and GID are valid in that case. */
        if (rec->lsr_valid == 0)
                body->oa.o_valid |= (OBD_MD_FLUID | OBD_MD_FLGID);
        else
                body->oa.o_valid |= rec->lsr_valid;

        osp_sync_send_new_rpc(d, llh, h, req);
        RETURN(0);
}

/**
 * Generate a request for unlink change.
 *
 * The function prepares a new RPC, initializes it with unlink(destroy)
 * specific bits and sends the RPC. The function is used to handle
 * llog_unlink_rec which were used in the older versions of Lustre.
 * Current version uses llog_unlink_rec64.
 *
 * \param[in] d         OSP device
 * \param[in] llh       llog handle where the record is stored
 * \param[in] h         llog record
 *
 * \retval 0            on success
 * \retval negative     negated errno on error
 */
static int osp_sync_new_unlink_job(struct osp_device *d,
                                   struct llog_handle *llh,
                                   struct llog_rec_hdr *h)
{
        struct llog_unlink_rec  *rec = (struct llog_unlink_rec *)h;
        struct ptlrpc_request   *req;
        struct ost_body         *body;
        int rc;

        ENTRY;
        LASSERT(h->lrh_type == MDS_UNLINK_REC);

        req = osp_sync_new_job(d, OST_DESTROY, &RQF_OST_DESTROY);
        if (IS_ERR(req))
                RETURN(PTR_ERR(req));

        body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
        LASSERT(body);
        ostid_set_seq(&body->oa.o_oi, rec->lur_oseq);
        rc = ostid_set_id(&body->oa.o_oi, rec->lur_oid);
        if (rc)
                return rc;
        body->oa.o_misc = rec->lur_count;
        body->oa.o_valid = OBD_MD_FLGROUP | OBD_MD_FLID;
        if (rec->lur_count)
                body->oa.o_valid |= OBD_MD_FLOBJCOUNT;

        osp_sync_send_new_rpc(d, llh, h, req);
        RETURN(0);
}

/**
 * Generate a request for unlink change.
 *
 * The function prepares a new RPC, initializes it with unlink(destroy)
 * specific bits and sends the RPC. Depending on the target (MDT or OST)
 * two different protocols are used. For MDT we use OUT (basically OSD API
 * updates transferred via a network). For OST we still use the old
 * protocol (OBD?), originally for compatibility. Later we can start to
 * use OUT for OST as well, this will allow batching and better code
 * unification.
 *
 * \param[in] d         OSP device
 * \param[in] llh       llog handle where the record is stored
 * \param[in] h         llog record
 *
 * \retval 0            on success
 * \retval negative     negated errno on error
 */
static int osp_sync_new_unlink64_job(struct osp_device *d,
                                     struct llog_handle *llh,
                                     struct llog_rec_hdr *h)
{
        struct llog_unlink64_rec        *rec = (struct llog_unlink64_rec *)h;
        struct ptlrpc_request           *req = NULL;
        struct ost_body                 *body;
        int                              rc;

        ENTRY;
        LASSERT(h->lrh_type == MDS_UNLINK64_REC);
        req = osp_sync_new_job(d, OST_DESTROY, &RQF_OST_DESTROY);
        if (IS_ERR(req))
                RETURN(PTR_ERR(req));

        body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
        if (body == NULL)
                RETURN(-EFAULT);
        rc = fid_to_ostid(&rec->lur_fid, &body->oa.o_oi);
        if (rc < 0)
                RETURN(rc);
        body->oa.o_misc = rec->lur_count;
        body->oa.o_valid = OBD_MD_FLGROUP | OBD_MD_FLID |
                           OBD_MD_FLOBJCOUNT;
        osp_sync_send_new_rpc(d, llh, h, req);
        RETURN(0);
}

/**
 * Process llog records.
 *
 * This function is called to process the llog records committed locally.
 * In the recovery model used by OSP we can apply a change to a remote
 * target once corresponding transaction (like posix unlink) is committed
 * locally so can't revert.
 * Depending on the llog record type, a given handler is called that is
 * responsible for preparing and sending the RPC to apply the change.
 * Special record type LLOG_GEN_REC marking a reboot is cancelled right away.
 *
 * \param[in] env       LU environment provided by the caller
 * \param[in] d         OSP device
 * \param[in] llh       llog handle where the record is stored
 * \param[in] rec       llog record
 */
static void osp_sync_process_record(const struct lu_env *env,
                                    struct osp_device *d,
                                    struct llog_handle *llh,
                                    struct llog_rec_hdr *rec)
{
        struct llog_handle      *cathandle = llh->u.phd.phd_cat_handle;
        struct llog_cookie       cookie;
        int                      rc = 0;

        ENTRY;

        cookie.lgc_lgl = llh->lgh_id;
        cookie.lgc_subsys = LLOG_MDS_OST_ORIG_CTXT;
        cookie.lgc_index = rec->lrh_index;

        d->opd_sync_last_catalog_idx = llh->lgh_hdr->llh_cat_idx;

        if (unlikely(rec->lrh_type == LLOG_GEN_REC)) {
                struct llog_gen_rec *gen = (struct llog_gen_rec *)rec;

                /* we're waiting for the record generated by this instance */
                LASSERT(d->opd_sync_prev_done == 0);
                if (!memcmp(&d->opd_sync_generation, &gen->lgr_gen,
                            sizeof(gen->lgr_gen))) {
                        CDEBUG(D_HA, "processed all old entries\n");
                        d->opd_sync_prev_done = 1;
                }

                /* cancel any generation record */
                rc = llog_cat_cancel_records(env, cathandle, 1, &cookie);

                RETURN_EXIT;
        }

        /*
         * now we prepare and fill requests to OST, put them on the queue
         * and fire after next commit callback
         */

        /* notice we increment counters before sending RPC, to be consistent
         * in RPC interpret callback which may happen very quickly */
        atomic_inc(&d->opd_sync_rpcs_in_flight);
        atomic_inc(&d->opd_sync_rpcs_in_progress);

        switch (rec->lrh_type) {
        /* case MDS_UNLINK_REC is kept for compatibility */
        case MDS_UNLINK_REC:
                rc = osp_sync_new_unlink_job(d, llh, rec);
                break;
        case MDS_UNLINK64_REC:
                rc = osp_sync_new_unlink64_job(d, llh, rec);
                break;
        case MDS_SETATTR64_REC:
                rc = osp_sync_new_setattr_job(d, llh, rec);
                break;
        default:
                CERROR("%s: unknown record type: %x\n", d->opd_obd->obd_name,
                       rec->lrh_type);
                /* treat "unknown record type" as "invalid" */
                rc = 1;
                break;
        }

        /* For all kinds of records, not matter successful or not,
         * we should decrease changes and bump last_processed_id.
         */
        if (d->opd_sync_prev_done) {
                __u64 correct_id = osp_sync_correct_id(d, rec);
                LASSERT(atomic_read(&d->opd_sync_changes) > 0);
                LASSERT(correct_id <= d->opd_sync_last_committed_id);
                /* NOTE: it's possible to meet same id if
                 * OST stores few stripes of same file
                 */
                while (1) {
                        /* another thread may be trying to set new value */
                        rmb();
                        if (correct_id > d->opd_sync_last_processed_id) {
                                d->opd_sync_last_processed_id = correct_id;
                                wake_up(&d->opd_sync_barrier_waitq);
                        } else
                                break;
                }
                atomic_dec(&d->opd_sync_changes);
        }
        if (rc != 0) {
                atomic_dec(&d->opd_sync_rpcs_in_flight);
                atomic_dec(&d->opd_sync_rpcs_in_progress);
        }

        CDEBUG(D_OTHER, "%s: %d in flight, %d in progress\n",
               d->opd_obd->obd_name, atomic_read(&d->opd_sync_rpcs_in_flight),
               atomic_read(&d->opd_sync_rpcs_in_progress));

        /* Delete the invalid record */
        if (rc == 1) {
                rc = llog_cat_cancel_records(env, cathandle, 1, &cookie);
                if (rc != 0)
                        CERROR("%s: can't delete invalid record: "
                               "fid = "DFID", rec_id = %u, rc = %d\n",
                               d->opd_obd->obd_name,
                               PFID(lu_object_fid(&cathandle->lgh_obj->do_lu)),
                               rec->lrh_id, rc);
        }

        CDEBUG(D_HA, "found record %x, %d, idx %u, id %u\n",
               rec->lrh_type, rec->lrh_len, rec->lrh_index, rec->lrh_id);

        RETURN_EXIT;
}

/**
 * Cancel llog records for the committed changes.
 *
 * The function walks through the list of the committed RPCs and cancels
 * corresponding llog records. see osp_sync_request_commit_cb() for the
 * details.
 *
 * \param[in] env       LU environment provided by the caller
 * \param[in] d         OSP device
 */
static void osp_sync_process_committed(const struct lu_env *env,
                                       struct osp_device *d)
{
        struct obd_device       *obd = d->opd_obd;
        struct obd_import       *imp = obd->u.cli.cl_import;
        struct ost_body         *body;
        struct ptlrpc_request   *req;
        struct llog_ctxt        *ctxt;
        struct llog_handle      *llh;
        struct list_head         list;
        int                      rc, done = 0;

        ENTRY;

        if (list_empty(&d->opd_sync_committed_there))
                return;

        /*
         * if current status is -ENOSPC (lack of free space on OST)
         * then we should poll OST immediately once object destroy
         * is committed.
         * notice: we do this upon commit as well because some backends
         * (like DMU) do not release space right away.
         */
        if (d->opd_pre != NULL && unlikely(d->opd_pre_status == -ENOSPC))
                osp_statfs_need_now(d);

        /*
         * now cancel them all
         * XXX: can we improve this using some batching?
         *      with batch RPC that'll happen automatically?
         * XXX: can we store ctxt in lod_device and save few cycles ?
         */
        ctxt = llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
        LASSERT(ctxt);

        llh = ctxt->loc_handle;
        LASSERT(llh);

        INIT_LIST_HEAD(&list);
        spin_lock(&d->opd_sync_lock);
        list_splice(&d->opd_sync_committed_there, &list);
        INIT_LIST_HEAD(&d->opd_sync_committed_there);
        spin_unlock(&d->opd_sync_lock);

        while (!list_empty(&list)) {
                struct osp_job_req_args *jra;

                jra = list_entry(list.next, struct osp_job_req_args,
                                 jra_committed_link);
                LASSERT(jra->jra_magic == OSP_JOB_MAGIC);
                list_del_init(&jra->jra_committed_link);

                req = container_of((void *)jra, struct ptlrpc_request,
                                   rq_async_args);
                body = req_capsule_client_get(&req->rq_pill,
                                              &RMF_OST_BODY);
                LASSERT(body);
                /* import can be closing, thus all commit cb's are
                 * called we can check committness directly */
                if (req->rq_import_generation == imp->imp_generation) {
                        rc = llog_cat_cancel_records(env, llh, 1,
                                                     &jra->jra_lcookie);
                        if (rc)
                                CERROR("%s: can't cancel record: %d\n",
                                       obd->obd_name, rc);
                } else {
                        DEBUG_REQ(D_OTHER, req, "imp_committed = %llu",
                                  imp->imp_peer_committed_transno);
                }
                ptlrpc_req_finished(req);
                done++;
        }

        llog_ctxt_put(ctxt);

        LASSERT(atomic_read(&d->opd_sync_rpcs_in_progress) >= done);
        atomic_sub(done, &d->opd_sync_rpcs_in_progress);
        CDEBUG(D_OTHER, "%s: %d in flight, %d in progress\n",
               d->opd_obd->obd_name, atomic_read(&d->opd_sync_rpcs_in_flight),
               atomic_read(&d->opd_sync_rpcs_in_progress));

        osp_sync_check_for_work(d);

        /* wake up the thread if requested to stop:
         * it might be waiting for in-progress to complete */
        if (unlikely(osp_sync_running(d) == 0))
                wake_up(&d->opd_sync_waitq);

        EXIT;
}

/**
 * The core of the syncing mechanism.
 *
 * This is a callback called by the llog processing function. Essentially it
 * suspends llog processing until there is a record to process (it's supposed
 * to be committed locally). The function handles RPCs committed by the target
 * and cancels corresponding llog records.
 *
 * \param[in] env       LU environment provided by the caller
 * \param[in] llh       llog handle we're processing
 * \param[in] rec       current llog record
 * \param[in] data      callback data containing a pointer to the device
 *
 * \retval 0                    to ask the caller (llog_process()) to continue
 * \retval LLOG_PROC_BREAK      to ask the caller to break
 */
static int osp_sync_process_queues(const struct lu_env *env,
                                   struct llog_handle *llh,
                                   struct llog_rec_hdr *rec,
                                   void *data)
{
        struct osp_device       *d = data;

        do {
                struct l_wait_info lwi = { 0 };

                if (!osp_sync_running(d)) {
                        CDEBUG(D_HA, "stop llog processing\n");
                        return LLOG_PROC_BREAK;
                }

                /* process requests committed by OST */
                osp_sync_process_committed(env, d);

                /* if we there are changes to be processed and we have
                 * resources for this ... do now */
                if (osp_sync_can_process_new(d, rec)) {
                        if (llh == NULL) {
                                /* ask llog for another record */
                                CDEBUG(D_HA, "%u changes, %u in progress,"
                                     " %u in flight\n",
                                     atomic_read(&d->opd_sync_changes),
                                     atomic_read(&d->opd_sync_rpcs_in_progress),
                                     atomic_read(&d->opd_sync_rpcs_in_flight));
                                return 0;
                        }
                        osp_sync_process_record(env, d, llh, rec);
                        llh = NULL;
                        rec = NULL;
                }

                if (d->opd_sync_last_processed_id == d->opd_sync_last_used_id)
                        osp_sync_remove_from_tracker(d);

                l_wait_event(d->opd_sync_waitq,
                             !osp_sync_running(d) ||
                             osp_sync_can_process_new(d, rec) ||
                             !list_empty(&d->opd_sync_committed_there),
                             &lwi);
        } while (1);
}

/**
 * OSP sync thread.
 *
 * This thread runs llog_cat_process() scanner calling our callback
 * to process llog records. in the callback we implement tricky
 * state machine as we don't want to start scanning of the llog again
 * and again, also we don't want to process too many records and send
 * too many RPCs a time. so, depending on current load (num of changes
 * being synced to OST) the callback can suspend awaiting for some
 * new conditions, like syncs completed.
 *
 * In order to process llog records left by previous boots and to allow
 * llog_process_thread() to find something (otherwise it'd just exit
 * immediately) we add a special GENERATATION record on each boot.
 *
 * \param[in] _arg      a pointer to thread's arguments
 *
 * \retval 0            on success
 * \retval negative     negated errno on error
 */
static int osp_sync_thread(void *_arg)
{
        struct osp_device       *d = _arg;
        struct ptlrpc_thread    *thread = &d->opd_sync_thread;
        struct l_wait_info       lwi = { 0 };
        struct llog_ctxt        *ctxt;
        struct obd_device       *obd = d->opd_obd;
        struct llog_handle      *llh;
        struct lu_env            env;
        int                      rc, count;
        bool                     wrapped;

        ENTRY;

        rc = lu_env_init(&env, LCT_LOCAL);
        if (rc) {
                CERROR("%s: can't initialize env: rc = %d\n",
                       obd->obd_name, rc);

                spin_lock(&d->opd_sync_lock);
                thread->t_flags = SVC_STOPPED;
                spin_unlock(&d->opd_sync_lock);
                wake_up(&thread->t_ctl_waitq);

                RETURN(rc);
        }

        spin_lock(&d->opd_sync_lock);
        thread->t_flags = SVC_RUNNING;
        spin_unlock(&d->opd_sync_lock);
        wake_up(&thread->t_ctl_waitq);

        ctxt = llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
        if (ctxt == NULL) {
                CERROR("can't get appropriate context\n");
                GOTO(out, rc = -EINVAL);
        }

        llh = ctxt->loc_handle;
        if (llh == NULL) {
                CERROR("can't get llh\n");
                llog_ctxt_put(ctxt);
                GOTO(out, rc = -EINVAL);
        }

        /*
         * Catalog processing stops when it processed last catalog record
         * with index equal to the end of catalog bitmap. Or if it is wrapped,
         * processing stops with index equal to the lgh_last_idx. We need to
         * continue processing.
         */
        d->opd_sync_last_catalog_idx = 0;
        do {
                int     size;

                wrapped = (llh->lgh_hdr->llh_cat_idx >= llh->lgh_last_idx &&
                           llh->lgh_hdr->llh_count > 1);

                rc = llog_cat_process(&env, llh, osp_sync_process_queues, d,
                                      d->opd_sync_last_catalog_idx, 0);

                size = OBD_FAIL_PRECHECK(OBD_FAIL_CAT_RECORDS) ?
                       cfs_fail_val : (LLOG_HDR_BITMAP_SIZE(llh->lgh_hdr) - 1);
                /* processing reaches catalog bottom */
                if (d->opd_sync_last_catalog_idx == size)
                        d->opd_sync_last_catalog_idx = 0;
                else if (wrapped)
                        /* If catalog is wrapped we can`t predict last index of
                         * processing because lgh_last_idx could be changed.
                         * Starting form the next one */
                        d->opd_sync_last_catalog_idx++;

        } while (rc == 0 && (wrapped || d->opd_sync_last_catalog_idx == 0));

        if (rc < 0) {
                CERROR("%s: llog process with osp_sync_process_queues "
                       "failed: %d\n", d->opd_obd->obd_name, rc);
                GOTO(close, rc);
        }
        LASSERTF(rc == 0 || rc == LLOG_PROC_BREAK,
                 "%u changes, %u in progress, %u in flight: %d\n",
                 atomic_read(&d->opd_sync_changes),
                 atomic_read(&d->opd_sync_rpcs_in_progress),
                 atomic_read(&d->opd_sync_rpcs_in_flight), rc);

        /* we don't expect llog_process_thread() to exit till umount */
        LASSERTF(thread->t_flags != SVC_RUNNING,
                 "%u changes, %u in progress, %u in flight\n",
                 atomic_read(&d->opd_sync_changes),
                 atomic_read(&d->opd_sync_rpcs_in_progress),
                 atomic_read(&d->opd_sync_rpcs_in_flight));

        /* wait till all the requests are completed */
        count = 0;
        while (atomic_read(&d->opd_sync_rpcs_in_progress) > 0) {
                osp_sync_process_committed(&env, d);

                lwi = LWI_TIMEOUT(cfs_time_seconds(5), NULL, NULL);
                rc = l_wait_event(d->opd_sync_waitq,
                                atomic_read(&d->opd_sync_rpcs_in_progress) == 0,
                                  &lwi);
                if (rc == -ETIMEDOUT)
                        count++;
                LASSERTF(count < 10, "%s: %d %d %sempty\n",
                         d->opd_obd->obd_name,
                         atomic_read(&d->opd_sync_rpcs_in_progress),
                         atomic_read(&d->opd_sync_rpcs_in_flight),
                         list_empty(&d->opd_sync_committed_there) ? "" : "!");

        }

close:
        llog_cat_close(&env, llh);
        rc = llog_cleanup(&env, ctxt);
        if (rc)
                CERROR("can't cleanup llog: %d\n", rc);
out:
        LASSERTF(atomic_read(&d->opd_sync_rpcs_in_progress) == 0,
                 "%s: %d %d %sempty\n", d->opd_obd->obd_name,
                 atomic_read(&d->opd_sync_rpcs_in_progress),
                 atomic_read(&d->opd_sync_rpcs_in_flight),
                 list_empty(&d->opd_sync_committed_there) ? "" : "!");

        thread->t_flags = SVC_STOPPED;

        wake_up(&thread->t_ctl_waitq);

        lu_env_fini(&env);

        RETURN(0);
}

/**
 * Initialize llog.
 *
 * Initializes the llog. Specific llog to be used depends on the type of the
 * target OSP represents (OST or MDT). The function adds appends a new llog
 * record to mark the place where the records associated with this boot
 * start.
 *
 * \param[in] env       LU environment provided by the caller
 * \param[in] d         OSP device
 *
 * \retval 0            on success
 * \retval negative     negated errno on error
 */
static int osp_sync_llog_init(const struct lu_env *env, struct osp_device *d)
{
        struct osp_thread_info  *osi = osp_env_info(env);
        struct lu_fid           *fid = &osi->osi_fid;
        struct llog_handle      *lgh = NULL;
        struct obd_device       *obd = d->opd_obd;
        struct llog_ctxt        *ctxt;
        int                     rc;

        ENTRY;

        LASSERT(obd);

        /*
         * open llog corresponding to our OST
         */
        OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
        obd->obd_lvfs_ctxt.dt = d->opd_storage;

        lu_local_obj_fid(fid, LLOG_CATALOGS_OID);

        rc = llog_osd_get_cat_list(env, d->opd_storage, d->opd_index, 1,
                                   &osi->osi_cid, fid);
        if (rc < 0) {
                if (rc != -EFAULT) {
                        CERROR("%s: can't get id from catalogs: rc = %d\n",
                               obd->obd_name, rc);
                        RETURN(rc);
                }

                /* After sparse OST indices is supported, the CATALOG file
                 * may become a sparse file that results in failure on
                 * reading. Skip this error as the llog will be created
                 * later */
                memset(&osi->osi_cid, 0, sizeof(osi->osi_cid));
                rc = 0;
        }

        CDEBUG(D_INFO, "%s: Init llog for %d - catid "DFID":%x\n",
               obd->obd_name, d->opd_index,
               PFID(&osi->osi_cid.lci_logid.lgl_oi.oi_fid),
               osi->osi_cid.lci_logid.lgl_ogen);

        rc = llog_setup(env, obd, &obd->obd_olg, LLOG_MDS_OST_ORIG_CTXT,
                        d->opd_storage->dd_lu_dev.ld_obd,
                        &osp_mds_ost_orig_logops);
        if (rc)
                RETURN(rc);

        ctxt = llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
        LASSERT(ctxt);

        if (likely(logid_id(&osi->osi_cid.lci_logid) != 0)) {
                rc = llog_open(env, ctxt, &lgh, &osi->osi_cid.lci_logid, NULL,
                               LLOG_OPEN_EXISTS);
                /* re-create llog if it is missing */
                if (rc == -ENOENT)
                        logid_set_id(&osi->osi_cid.lci_logid, 0);
                else if (rc < 0)
                        GOTO(out_cleanup, rc);
        }

        if (unlikely(logid_id(&osi->osi_cid.lci_logid) == 0)) {
                rc = llog_open_create(env, ctxt, &lgh, NULL, NULL);
                if (rc < 0)
                        GOTO(out_cleanup, rc);
                osi->osi_cid.lci_logid = lgh->lgh_id;
        }

        LASSERT(lgh != NULL);
        ctxt->loc_handle = lgh;

        rc = llog_init_handle(env, lgh, LLOG_F_IS_CAT, NULL);
        if (rc)
                GOTO(out_close, rc);

        rc = llog_osd_put_cat_list(env, d->opd_storage, d->opd_index, 1,
                                   &osi->osi_cid, fid);
        if (rc)
                GOTO(out_close, rc);

        /*
         * put a mark in the llog till which we'll be processing
         * old records restless
         */
        d->opd_sync_generation.mnt_cnt = cfs_time_current();
        d->opd_sync_generation.conn_cnt = cfs_time_current();

        osi->osi_hdr.lrh_type = LLOG_GEN_REC;
        osi->osi_hdr.lrh_len = sizeof(osi->osi_gen);

        memcpy(&osi->osi_gen.lgr_gen, &d->opd_sync_generation,
               sizeof(osi->osi_gen.lgr_gen));

        rc = llog_cat_add(env, lgh, &osi->osi_gen.lgr_hdr, &osi->osi_cookie);
        if (rc < 0)
                GOTO(out_close, rc);
        llog_ctxt_put(ctxt);
        RETURN(0);
out_close:
        llog_cat_close(env, lgh);
out_cleanup:
        llog_cleanup(env, ctxt);
        RETURN(rc);
}

/**
 * Cleanup llog used for syncing.
 *
 * Closes and cleanups the llog. The function is called when the device is
 * shutting down.
 *
 * \param[in] env       LU environment provided by the caller
 * \param[in] d         OSP device
 */
static void osp_sync_llog_fini(const struct lu_env *env, struct osp_device *d)
{
        struct llog_ctxt *ctxt;

        ctxt = llog_get_context(d->opd_obd, LLOG_MDS_OST_ORIG_CTXT);
        if (ctxt) {
                llog_cat_close(env, ctxt->loc_handle);
                llog_cleanup(env, ctxt);
        }
}

/**
 * Initialization of the sync component of OSP.
 *
 * Initializes the llog and starts a new thread to handle the changes to
 * the remote target (OST or MDT).
 *
 * \param[in] env       LU environment provided by the caller
 * \param[in] d         OSP device
 *
 * \retval 0            on success
 * \retval negative     negated errno on error
 */
int osp_sync_init(const struct lu_env *env, struct osp_device *d)
{
        struct l_wait_info       lwi = { 0 };
        struct task_struct      *task;
        int                      rc;

        ENTRY;

        d->opd_sync_max_rpcs_in_flight = OSP_MAX_RPCS_IN_FLIGHT;
        d->opd_sync_max_rpcs_in_progress = OSP_MAX_RPCS_IN_PROGRESS;
        spin_lock_init(&d->opd_sync_lock);
        init_waitqueue_head(&d->opd_sync_waitq);
        init_waitqueue_head(&d->opd_sync_barrier_waitq);
        thread_set_flags(&d->opd_sync_thread, SVC_INIT);
        init_waitqueue_head(&d->opd_sync_thread.t_ctl_waitq);
        INIT_LIST_HEAD(&d->opd_sync_in_flight_list);
        INIT_LIST_HEAD(&d->opd_sync_committed_there);

        if (d->opd_storage->dd_rdonly)
                RETURN(0);

        rc = osp_sync_id_traction_init(d);
        if (rc)
                RETURN(rc);

        /*
         * initialize llog storing changes
         */
        rc = osp_sync_llog_init(env, d);
        if (rc) {
                CERROR("%s: can't initialize llog: rc = %d\n",
                       d->opd_obd->obd_name, rc);
                GOTO(err_id, rc);
        }

        /*
         * Start synchronization thread
         */
        task = kthread_run(osp_sync_thread, d, "osp-syn-%u-%u",
                           d->opd_index, d->opd_group);
        if (IS_ERR(task)) {
                rc = PTR_ERR(task);
                CERROR("%s: cannot start sync thread: rc = %d\n",
                       d->opd_obd->obd_name, rc);
                GOTO(err_llog, rc);
        }

        l_wait_event(d->opd_sync_thread.t_ctl_waitq,
                     osp_sync_running(d) || osp_sync_stopped(d), &lwi);

        RETURN(0);
err_llog:
        osp_sync_llog_fini(env, d);
err_id:
        osp_sync_id_traction_fini(d);
        return rc;
}

/**
 * Stop the syncing thread.
 *
 * Asks the syncing thread to stop and wait until it's stopped.
 *
 * \param[in] d         OSP device
 *
 * \retval              0
 */
int osp_sync_fini(struct osp_device *d)
{
        struct ptlrpc_thread *thread = &d->opd_sync_thread;

        ENTRY;

        if (!thread_is_init(thread) && !thread_is_stopped(thread)) {
                thread->t_flags = SVC_STOPPING;
                wake_up(&d->opd_sync_waitq);
                wait_event(thread->t_ctl_waitq, thread_is_stopped(thread));
        }

        /*
         * unregister transaction callbacks only when sync thread
         * has finished operations with llog
         */
        osp_sync_id_traction_fini(d);

        RETURN(0);
}

static DEFINE_MUTEX(osp_id_tracker_sem);
static struct list_head osp_id_tracker_list =
                LIST_HEAD_INIT(osp_id_tracker_list);

/**
 * OSD commit callback.
 *
 * The function is used as a local OSD commit callback to track the highest
 * committed llog record id. see osp_sync_id_traction_init() for the details.
 *
 * \param[in] th        local transaction handle committed
 * \param[in] cookie    commit callback data (our private structure)
 */
static void osp_sync_tracker_commit_cb(struct thandle *th, void *cookie)
{
        struct osp_id_tracker   *tr = cookie;
        struct osp_device       *d;
        struct osp_txn_info     *txn;

        LASSERT(tr);

        txn = osp_txn_info(&th->th_ctx);
        if (txn == NULL || txn->oti_current_id < tr->otr_committed_id)
                return;

        spin_lock(&tr->otr_lock);
        if (likely(txn->oti_current_id > tr->otr_committed_id)) {
                CDEBUG(D_OTHER, "committed: %llu -> %llu\n",
                       tr->otr_committed_id, txn->oti_current_id);
                tr->otr_committed_id = txn->oti_current_id;

                list_for_each_entry(d, &tr->otr_wakeup_list,
                                    opd_sync_ontrack) {
                        d->opd_sync_last_committed_id = tr->otr_committed_id;
                        wake_up(&d->opd_sync_waitq);
                }
        }
        spin_unlock(&tr->otr_lock);
}

/**
 * Initialize commit tracking mechanism.
 *
 * Some setups may have thousands of OSTs and each will be represented by OSP.
 * Meaning order of magnitute many more changes to apply every second. In order
 * to keep the number of commit callbacks low this mechanism was introduced.
 * The mechanism is very similar to transno used by MDT service: it's an single
 * ID stream which can be assigned by any OSP to its llog records. The tricky
 * part is that ID is stored in per-transaction data and re-used by all the OSPs
 * involved in that transaction. Then all these OSPs are woken up utilizing a single OSD commit callback.
 *
 * The function initializes the data used by the tracker described above.
 * A singler tracker per OSD device is created.
 *
 * \param[in] d         OSP device
 *
 * \retval 0            on success
 * \retval negative     negated errno on error
 */
static int osp_sync_id_traction_init(struct osp_device *d)
{
        struct osp_id_tracker   *tr, *found = NULL;
        int                      rc = 0;

        LASSERT(d);
        LASSERT(d->opd_storage);
        LASSERT(d->opd_sync_tracker == NULL);
        INIT_LIST_HEAD(&d->opd_sync_ontrack);

        mutex_lock(&osp_id_tracker_sem);
        list_for_each_entry(tr, &osp_id_tracker_list, otr_list) {
                if (tr->otr_dev == d->opd_storage) {
                        LASSERT(atomic_read(&tr->otr_refcount));
                        atomic_inc(&tr->otr_refcount);
                        d->opd_sync_tracker = tr;
                        found = tr;
                        break;
                }
        }

        if (found == NULL) {
                rc = -ENOMEM;
                OBD_ALLOC_PTR(tr);
                if (tr) {
                        d->opd_sync_tracker = tr;
                        spin_lock_init(&tr->otr_lock);
                        tr->otr_dev = d->opd_storage;
                        tr->otr_next_id = 1;
                        tr->otr_committed_id = 0;
                        atomic_set(&tr->otr_refcount, 1);
                        INIT_LIST_HEAD(&tr->otr_wakeup_list);
                        list_add(&tr->otr_list, &osp_id_tracker_list);
                        tr->otr_tx_cb.dtc_txn_commit =
                                                osp_sync_tracker_commit_cb;
                        tr->otr_tx_cb.dtc_cookie = tr;
                        tr->otr_tx_cb.dtc_tag = LCT_MD_THREAD;
                        dt_txn_callback_add(d->opd_storage, &tr->otr_tx_cb);
                        rc = 0;
                }
        }
        mutex_unlock(&osp_id_tracker_sem);

        return rc;
}

/**
 * Release commit tracker.
 *
 * Decrease a refcounter on the tracker used by the given OSP device \a d.
 * If no more users left, then the tracker is released.
 *
 * \param[in] d         OSP device
 */
static void osp_sync_id_traction_fini(struct osp_device *d)
{
        struct osp_id_tracker *tr;

        ENTRY;

        LASSERT(d);
        tr = d->opd_sync_tracker;
        if (tr == NULL) {
                EXIT;
                return;
        }

        osp_sync_remove_from_tracker(d);

        mutex_lock(&osp_id_tracker_sem);
        if (atomic_dec_and_test(&tr->otr_refcount)) {
                dt_txn_callback_del(d->opd_storage, &tr->otr_tx_cb);
                LASSERT(list_empty(&tr->otr_wakeup_list));
                list_del(&tr->otr_list);
                OBD_FREE_PTR(tr);
                d->opd_sync_tracker = NULL;
        }
        mutex_unlock(&osp_id_tracker_sem);

        EXIT;
}

/**
 * Generate a new ID on a tracker.
 *
 * Generates a new ID using the tracker associated with the given OSP device
 * \a d, if the given ID \a id is non-zero. Unconditially adds OSP device to
 * the wakeup list, so OSP won't miss when a transaction using the ID is
 * committed.
 *
 * \param[in] d         OSP device
 * \param[in] id        0 or ID generated previously
 *
 * \retval              ID the caller should use
 */
static __u64 osp_sync_id_get(struct osp_device *d, __u64 id)
{
        struct osp_id_tracker *tr;

        tr = d->opd_sync_tracker;
        LASSERT(tr);

        /* XXX: we can improve this introducing per-cpu preallocated ids? */
        spin_lock(&tr->otr_lock);
        if (OBD_FAIL_CHECK(OBD_FAIL_MDS_TRACK_OVERFLOW))
                tr->otr_next_id = 0xfffffff0;

        if (unlikely(tr->otr_next_id <= d->opd_sync_last_used_id)) {
                spin_unlock(&tr->otr_lock);
                CERROR("%s: next %llu, last synced %llu\n",
                       d->opd_obd->obd_name, tr->otr_next_id,
                       d->opd_sync_last_used_id);
                LBUG();
        }

        if (id == 0)
                id = tr->otr_next_id++;
        if (id > d->opd_sync_last_used_id)
                d->opd_sync_last_used_id = id;
        if (list_empty(&d->opd_sync_ontrack))
                list_add(&d->opd_sync_ontrack, &tr->otr_wakeup_list);
        spin_unlock(&tr->otr_lock);
        CDEBUG(D_OTHER, "new id %llu\n", id);

        return id;
}

/**
 * Stop to propagate commit status to OSP.
 *
 * If the OSP does not have any llog records she's waiting to commit, then
 * it is possible to unsubscribe from wakeups from the tracking using this
 * method.
 *
 * \param[in] d         OSP device not willing to wakeup
 */
static void osp_sync_remove_from_tracker(struct osp_device *d)
{
        struct osp_id_tracker *tr;

        tr = d->opd_sync_tracker;
        LASSERT(tr);

        if (list_empty(&d->opd_sync_ontrack))
                return;

        spin_lock(&tr->otr_lock);
        list_del_init(&d->opd_sync_ontrack);
        spin_unlock(&tr->otr_lock);
}