LU-7309 lod: notify client retry creation

[fs/lustre-release.git] / lustre / lod / lod_qos.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 COPYING file that accompanied this code.
 *
 * You should have received a copy of the GNU General Public License
 * along with this program; if not, write to the Free Software
 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
 *
 * GPL HEADER END
 */
/*
 * Copyright  2009 Sun Microsystems, Inc. All rights reserved
 * Use is subject to license terms.
 *
 * Copyright (c) 2012, 2015, Intel Corporation.
 */
/*
 * This file is part of Lustre, http://www.lustre.org/
 * Lustre is a trademark of Sun Microsystems, Inc.
 *
 * lustre/lod/lod_qos.c
 *
 * Implementation of different allocation algorithm used
 * to distribute objects and data among OSTs.
 */

#define DEBUG_SUBSYSTEM S_LOV

#include <asm/div64.h>
#include <libcfs/libcfs.h>
#include <lustre/lustre_idl.h>
#include <lustre_swab.h>
#include <obd_class.h>

#include "lod_internal.h"

/*
 * force QoS policy (not RR) to be used for testing purposes
 */
#define FORCE_QOS_

#define D_QOS   D_OTHER

#define QOS_DEBUG(fmt, ...)     CDEBUG(D_QOS, fmt, ## __VA_ARGS__)
#define QOS_CONSOLE(fmt, ...)   LCONSOLE(D_QOS, fmt, ## __VA_ARGS__)

#define TGT_BAVAIL(i) (OST_TGT(lod,i)->ltd_statfs.os_bavail * \
                       OST_TGT(lod,i)->ltd_statfs.os_bsize)

/**
 * Add a new target to Quality of Service (QoS) target table.
 *
 * Add a new OST target to the structure representing an OSS. Resort the list
 * of known OSSs by the number of OSTs attached to each OSS. The OSS list is
 * protected internally and no external locking is required.
 *
 * \param[in] lod               LOD device
 * \param[in] ost_desc          OST description
 *
 * \retval 0                    on success
 * \retval -ENOMEM              on error
 */
int qos_add_tgt(struct lod_device *lod, struct lod_tgt_desc *ost_desc)
{
        struct lod_qos_oss *oss = NULL, *temposs;
        struct obd_export  *exp = ost_desc->ltd_exp;
        int                 rc = 0, found = 0;
        struct list_head   *list;
        ENTRY;

        down_write(&lod->lod_qos.lq_rw_sem);
        /*
         * a bit hacky approach to learn NID of corresponding connection
         * but there is no official API to access information like this
         * with OSD API.
         */
        list_for_each_entry(oss, &lod->lod_qos.lq_oss_list, lqo_oss_list) {
                if (obd_uuid_equals(&oss->lqo_uuid,
                                    &exp->exp_connection->c_remote_uuid)) {
                        found++;
                        break;
                }
        }

        if (!found) {
                OBD_ALLOC_PTR(oss);
                if (!oss)
                        GOTO(out, rc = -ENOMEM);
                memcpy(&oss->lqo_uuid, &exp->exp_connection->c_remote_uuid,
                       sizeof(oss->lqo_uuid));
        } else {
                /* Assume we have to move this one */
                list_del(&oss->lqo_oss_list);
        }

        oss->lqo_ost_count++;
        ost_desc->ltd_qos.ltq_oss = oss;

        CDEBUG(D_QOS, "add tgt %s to OSS %s (%d OSTs)\n",
               obd_uuid2str(&ost_desc->ltd_uuid), obd_uuid2str(&oss->lqo_uuid),
               oss->lqo_ost_count);

        /* Add sorted by # of OSTs.  Find the first entry that we're
           bigger than... */
        list = &lod->lod_qos.lq_oss_list;
        list_for_each_entry(temposs, list, lqo_oss_list) {
                if (oss->lqo_ost_count > temposs->lqo_ost_count)
                        break;
        }
        /* ...and add before it.  If we're the first or smallest, temposs
           points to the list head, and we add to the end. */
        list_add_tail(&oss->lqo_oss_list, &temposs->lqo_oss_list);

        lod->lod_qos.lq_dirty = 1;
        lod->lod_qos.lq_rr.lqr_dirty = 1;

out:
        up_write(&lod->lod_qos.lq_rw_sem);
        RETURN(rc);
}

/**
 * Remove OST target from QoS table.
 *
 * Removes given OST target from QoS table and releases related OSS structure
 * if no OSTs remain on the OSS.
 *
 * \param[in] lod               LOD device
 * \param[in] ost_desc          OST description
 *
 * \retval 0                    on success
 * \retval -ENOENT              if no OSS was found
 */
int qos_del_tgt(struct lod_device *lod, struct lod_tgt_desc *ost_desc)
{
        struct lod_qos_oss *oss;
        int                 rc = 0;
        ENTRY;

        down_write(&lod->lod_qos.lq_rw_sem);
        oss = ost_desc->ltd_qos.ltq_oss;
        if (!oss)
                GOTO(out, rc = -ENOENT);

        oss->lqo_ost_count--;
        if (oss->lqo_ost_count == 0) {
                CDEBUG(D_QOS, "removing OSS %s\n",
                       obd_uuid2str(&oss->lqo_uuid));
                list_del(&oss->lqo_oss_list);
                ost_desc->ltd_qos.ltq_oss = NULL;
                OBD_FREE_PTR(oss);
        }

        lod->lod_qos.lq_dirty = 1;
        lod->lod_qos.lq_rr.lqr_dirty = 1;
out:
        up_write(&lod->lod_qos.lq_rw_sem);
        RETURN(rc);
}

/**
 * Check whether the target is available for new OST objects.
 *
 * Request statfs data from the given target and verify it's active and not
 * read-only. If so, then it can be used to place new OST objects. This
 * function also maintains the number of active/inactive targets and sets
 * dirty flags if those numbers change so others can run re-balance procedures.
 * No external locking is required.
 *
 * \param[in] env       execution environment for this thread
 * \param[in] d         LOD device
 * \param[in] index     index of OST target to check
 * \param[out] sfs      buffer for statfs data
 *
 * \retval 0            if the target is good
 * \retval negative     negated errno on error

 */
static int lod_statfs_and_check(const struct lu_env *env, struct lod_device *d,
                                int index, struct obd_statfs *sfs)
{
        struct lod_tgt_desc *ost;
        int                  rc;
        ENTRY;

        LASSERT(d);
        ost = OST_TGT(d,index);
        LASSERT(ost);

        rc = dt_statfs(env, ost->ltd_ost, sfs);
        if (rc && rc != -ENOTCONN)
                CERROR("%s: statfs: rc = %d\n", lod2obd(d)->obd_name, rc);

        /* If the OST is readonly then we can't allocate objects there */
        if (sfs->os_state & OS_STATE_READONLY)
                rc = -EROFS;

        /* check whether device has changed state (active, inactive) */
        if (rc != 0 && ost->ltd_active) {
                /* turned inactive? */
                spin_lock(&d->lod_desc_lock);
                if (ost->ltd_active) {
                        ost->ltd_active = 0;
                        if (rc == -ENOTCONN)
                                ost->ltd_connecting = 1;

                        LASSERT(d->lod_desc.ld_active_tgt_count > 0);
                        d->lod_desc.ld_active_tgt_count--;
                        d->lod_qos.lq_dirty = 1;
                        d->lod_qos.lq_rr.lqr_dirty = 1;
                        CDEBUG(D_CONFIG, "%s: turns inactive\n",
                               ost->ltd_exp->exp_obd->obd_name);
                }
                spin_unlock(&d->lod_desc_lock);
        } else if (rc == 0 && ost->ltd_active == 0) {
                /* turned active? */
                LASSERTF(d->lod_desc.ld_active_tgt_count < d->lod_ostnr,
                         "active tgt count %d, ost nr %d\n",
                         d->lod_desc.ld_active_tgt_count, d->lod_ostnr);
                spin_lock(&d->lod_desc_lock);
                if (ost->ltd_active == 0) {
                        ost->ltd_active = 1;
                        ost->ltd_connecting = 0;
                        d->lod_desc.ld_active_tgt_count++;
                        d->lod_qos.lq_dirty = 1;
                        d->lod_qos.lq_rr.lqr_dirty = 1;
                        CDEBUG(D_CONFIG, "%s: turns active\n",
                               ost->ltd_exp->exp_obd->obd_name);
                }
                spin_unlock(&d->lod_desc_lock);
        }

        RETURN(rc);
}

/**
 * Maintain per-target statfs data.
 *
 * The function refreshes statfs data for all the targets every N seconds.
 * The actual N is controlled via procfs and set to LOV_DESC_QOS_MAXAGE_DEFAULT
 * initially.
 *
 * \param[in] env       execution environment for this thread
 * \param[in] lod       LOD device
 */
static void lod_qos_statfs_update(const struct lu_env *env,
                                  struct lod_device *lod)
{
        struct obd_device *obd = lod2obd(lod);
        struct ost_pool   *osts = &(lod->lod_pool_info);
        unsigned int       i;
        int                idx;
        __u64              max_age, avail;
        ENTRY;

        max_age = cfs_time_shift_64(-2 * lod->lod_desc.ld_qos_maxage);

        if (cfs_time_beforeq_64(max_age, obd->obd_osfs_age))
                /* statfs data are quite recent, don't need to refresh it */
                RETURN_EXIT;

        down_write(&lod->lod_qos.lq_rw_sem);
        if (cfs_time_beforeq_64(max_age, obd->obd_osfs_age))
                goto out;

        for (i = 0; i < osts->op_count; i++) {
                idx = osts->op_array[i];
                avail = OST_TGT(lod,idx)->ltd_statfs.os_bavail;
                if (lod_statfs_and_check(env, lod, idx,
                                         &OST_TGT(lod, idx)->ltd_statfs))
                        continue;
                if (OST_TGT(lod,idx)->ltd_statfs.os_bavail != avail)
                        /* recalculate weigths */
                        lod->lod_qos.lq_dirty = 1;
        }
        obd->obd_osfs_age = cfs_time_current_64();

out:
        up_write(&lod->lod_qos.lq_rw_sem);
        EXIT;
}

/**
 * Calculate per-OST and per-OSS penalties
 *
 * Re-calculate penalties when the configuration changes, active targets
 * change and after statfs refresh (all these are reflected by lq_dirty flag).
 * On every OST and OSS: decay the penalty by half for every 8x the update
 * interval that the device has been idle. That gives lots of time for the
 * statfs information to be updated (which the penalty is only a proxy for),
 * and avoids penalizing OSS/OSTs under light load.
 * See lod_qos_calc_weight() for how penalties are factored into the weight.
 *
 * \param[in] lod       LOD device
 *
 * \retval 0            on success
 * \retval -EAGAIN      the number of OSTs isn't enough
 */
static int lod_qos_calc_ppo(struct lod_device *lod)
{
        struct lod_qos_oss *oss;
        __u64               ba_max, ba_min, temp;
        __u32               num_active;
        unsigned int        i;
        int                 rc, prio_wide;
        time_t              now, age;
        ENTRY;

        if (!lod->lod_qos.lq_dirty)
                GOTO(out, rc = 0);

        num_active = lod->lod_desc.ld_active_tgt_count - 1;
        if (num_active < 1)
                GOTO(out, rc = -EAGAIN);

        /* find bavail on each OSS */
        list_for_each_entry(oss, &lod->lod_qos.lq_oss_list, lqo_oss_list)
                            oss->lqo_bavail = 0;
        lod->lod_qos.lq_active_oss_count = 0;

        /*
         * How badly user wants to select OSTs "widely" (not recently chosen
         * and not on recent OSS's).  As opposed to "freely" (free space
         * avail.) 0-256
         */
        prio_wide = 256 - lod->lod_qos.lq_prio_free;

        ba_min = (__u64)(-1);
        ba_max = 0;
        now = cfs_time_current_sec();
        /* Calculate OST penalty per object
         * (lod ref taken in lod_qos_prep_create()) */
        cfs_foreach_bit(lod->lod_ost_bitmap, i) {
                LASSERT(OST_TGT(lod,i));
                temp = TGT_BAVAIL(i);
                if (!temp)
                        continue;
                ba_min = min(temp, ba_min);
                ba_max = max(temp, ba_max);

                /* Count the number of usable OSS's */
                if (OST_TGT(lod,i)->ltd_qos.ltq_oss->lqo_bavail == 0)
                        lod->lod_qos.lq_active_oss_count++;
                OST_TGT(lod,i)->ltd_qos.ltq_oss->lqo_bavail += temp;

                /* per-OST penalty is prio * TGT_bavail / (num_ost - 1) / 2 */
                temp >>= 1;
                do_div(temp, num_active);
                OST_TGT(lod,i)->ltd_qos.ltq_penalty_per_obj =
                        (temp * prio_wide) >> 8;

                age = (now - OST_TGT(lod,i)->ltd_qos.ltq_used) >> 3;
                if (lod->lod_qos.lq_reset ||
                    age > 32 * lod->lod_desc.ld_qos_maxage)
                        OST_TGT(lod,i)->ltd_qos.ltq_penalty = 0;
                else if (age > lod->lod_desc.ld_qos_maxage)
                        /* Decay OST penalty. */
                        OST_TGT(lod,i)->ltd_qos.ltq_penalty >>=
                                (age / lod->lod_desc.ld_qos_maxage);
        }

        num_active = lod->lod_qos.lq_active_oss_count - 1;
        if (num_active < 1) {
                /* If there's only 1 OSS, we can't penalize it, so instead
                   we have to double the OST penalty */
                num_active = 1;
                cfs_foreach_bit(lod->lod_ost_bitmap, i)
                        OST_TGT(lod,i)->ltd_qos.ltq_penalty_per_obj <<= 1;
        }

        /* Per-OSS penalty is prio * oss_avail / oss_osts / (num_oss - 1) / 2 */
        list_for_each_entry(oss, &lod->lod_qos.lq_oss_list, lqo_oss_list) {
                temp = oss->lqo_bavail >> 1;
                do_div(temp, oss->lqo_ost_count * num_active);
                oss->lqo_penalty_per_obj = (temp * prio_wide) >> 8;

                age = (now - oss->lqo_used) >> 3;
                if (lod->lod_qos.lq_reset ||
                    age > 32 * lod->lod_desc.ld_qos_maxage)
                        oss->lqo_penalty = 0;
                else if (age > lod->lod_desc.ld_qos_maxage)
                        /* Decay OSS penalty. */
                        oss->lqo_penalty >>= age / lod->lod_desc.ld_qos_maxage;
        }

        lod->lod_qos.lq_dirty = 0;
        lod->lod_qos.lq_reset = 0;

        /* If each ost has almost same free space,
         * do rr allocation for better creation performance */
        lod->lod_qos.lq_same_space = 0;
        if ((ba_max * (256 - lod->lod_qos.lq_threshold_rr)) >> 8 < ba_min) {
                lod->lod_qos.lq_same_space = 1;
                /* Reset weights for the next time we enter qos mode */
                lod->lod_qos.lq_reset = 1;
        }
        rc = 0;

out:
#ifndef FORCE_QOS
        if (!rc && lod->lod_qos.lq_same_space)
                RETURN(-EAGAIN);
#endif
        RETURN(rc);
}

/**
 * Calculate weight for a given OST target.
 *
 * The final OST weight is the number of bytes available minus the OST and
 * OSS penalties.  See lod_qos_calc_ppo() for how penalties are calculated.
 *
 * \param[in] lod       LOD device, where OST targets are listed
 * \param[in] i         OST target index
 *
 * \retval              0
 */
static int lod_qos_calc_weight(struct lod_device *lod, int i)
{
        __u64 temp, temp2;

        temp = TGT_BAVAIL(i);
        temp2 = OST_TGT(lod,i)->ltd_qos.ltq_penalty +
                OST_TGT(lod,i)->ltd_qos.ltq_oss->lqo_penalty;
        if (temp < temp2)
                OST_TGT(lod,i)->ltd_qos.ltq_weight = 0;
        else
                OST_TGT(lod,i)->ltd_qos.ltq_weight = temp - temp2;
        return 0;
}

/**
 * Re-calculate weights.
 *
 * The function is called when some OST target was used for a new object. In
 * this case we should re-calculate all the weights to keep new allocations
 * balanced well.
 *
 * \param[in] lod       LOD device
 * \param[in] osts      OST pool where a new object was placed
 * \param[in] index     OST target where a new object was placed
 * \param[out] total_wt new total weight for the pool
 *
 * \retval              0
 */
static int lod_qos_used(struct lod_device *lod, struct ost_pool *osts,
                        __u32 index, __u64 *total_wt)
{
        struct lod_tgt_desc *ost;
        struct lod_qos_oss  *oss;
        unsigned int j;
        ENTRY;

        ost = OST_TGT(lod,index);
        LASSERT(ost);

        /* Don't allocate on this devuce anymore, until the next alloc_qos */
        ost->ltd_qos.ltq_usable = 0;

        oss = ost->ltd_qos.ltq_oss;

        /* Decay old penalty by half (we're adding max penalty, and don't
           want it to run away.) */
        ost->ltd_qos.ltq_penalty >>= 1;
        oss->lqo_penalty >>= 1;

        /* mark the OSS and OST as recently used */
        ost->ltd_qos.ltq_used = oss->lqo_used = cfs_time_current_sec();

        /* Set max penalties for this OST and OSS */
        ost->ltd_qos.ltq_penalty +=
                ost->ltd_qos.ltq_penalty_per_obj * lod->lod_ostnr;
        oss->lqo_penalty += oss->lqo_penalty_per_obj *
                lod->lod_qos.lq_active_oss_count;

        /* Decrease all OSS penalties */
        list_for_each_entry(oss, &lod->lod_qos.lq_oss_list, lqo_oss_list) {
                if (oss->lqo_penalty < oss->lqo_penalty_per_obj)
                        oss->lqo_penalty = 0;
                else
                        oss->lqo_penalty -= oss->lqo_penalty_per_obj;
        }

        *total_wt = 0;
        /* Decrease all OST penalties */
        for (j = 0; j < osts->op_count; j++) {
                int i;

                i = osts->op_array[j];
                if (!cfs_bitmap_check(lod->lod_ost_bitmap, i))
                        continue;

                ost = OST_TGT(lod,i);
                LASSERT(ost);

                if (ost->ltd_qos.ltq_penalty <
                                ost->ltd_qos.ltq_penalty_per_obj)
                        ost->ltd_qos.ltq_penalty = 0;
                else
                        ost->ltd_qos.ltq_penalty -=
                                ost->ltd_qos.ltq_penalty_per_obj;

                lod_qos_calc_weight(lod, i);

                /* Recalc the total weight of usable osts */
                if (ost->ltd_qos.ltq_usable)
                        *total_wt += ost->ltd_qos.ltq_weight;

                QOS_DEBUG("recalc tgt %d usable=%d avail="LPU64
                          " ostppo="LPU64" ostp="LPU64" ossppo="LPU64
                          " ossp="LPU64" wt="LPU64"\n",
                          i, ost->ltd_qos.ltq_usable, TGT_BAVAIL(i) >> 10,
                          ost->ltd_qos.ltq_penalty_per_obj >> 10,
                          ost->ltd_qos.ltq_penalty >> 10,
                          ost->ltd_qos.ltq_oss->lqo_penalty_per_obj >> 10,
                          ost->ltd_qos.ltq_oss->lqo_penalty >> 10,
                          ost->ltd_qos.ltq_weight >> 10);
        }

        RETURN(0);
}

void lod_qos_rr_init(struct lod_qos_rr *lqr)
{
        spin_lock_init(&lqr->lqr_alloc);
        lqr->lqr_dirty = 1;
}


#define LOV_QOS_EMPTY ((__u32)-1)

/**
 * Calculate optimal round-robin order with regard to OSSes.
 *
 * Place all the OSTs from pool \a src_pool in a special array to be used for
 * round-robin (RR) stripe allocation.  The placement algorithm interleaves
 * OSTs from the different OSSs so that RR allocation can balance OSSs evenly.
 * Resorts the targets when the number of active targets changes (because of
 * a new target or activation/deactivation).
 *
 * \param[in] lod       LOD device
 * \param[in] src_pool  OST pool
 * \param[in] lqr       round-robin list
 *
 * \retval 0            on success
 * \retval -ENOMEM      fails to allocate the array
 */
static int lod_qos_calc_rr(struct lod_device *lod, struct ost_pool *src_pool,
                           struct lod_qos_rr *lqr)
{
        struct lod_qos_oss  *oss;
        struct lod_tgt_desc *ost;
        unsigned placed, real_count;
        unsigned int i;
        int rc;
        ENTRY;

        if (!lqr->lqr_dirty) {
                LASSERT(lqr->lqr_pool.op_size);
                RETURN(0);
        }

        /* Do actual allocation. */
        down_write(&lod->lod_qos.lq_rw_sem);

        /*
         * Check again. While we were sleeping on @lq_rw_sem something could
         * change.
         */
        if (!lqr->lqr_dirty) {
                LASSERT(lqr->lqr_pool.op_size);
                up_write(&lod->lod_qos.lq_rw_sem);
                RETURN(0);
        }

        real_count = src_pool->op_count;

        /* Zero the pool array */
        /* alloc_rr is holding a read lock on the pool, so nobody is adding/
           deleting from the pool. The lq_rw_sem insures that nobody else
           is reading. */
        lqr->lqr_pool.op_count = real_count;
        rc = lod_ost_pool_extend(&lqr->lqr_pool, real_count);
        if (rc) {
                up_write(&lod->lod_qos.lq_rw_sem);
                RETURN(rc);
        }
        for (i = 0; i < lqr->lqr_pool.op_count; i++)
                lqr->lqr_pool.op_array[i] = LOV_QOS_EMPTY;

        /* Place all the OSTs from 1 OSS at the same time. */
        placed = 0;
        list_for_each_entry(oss, &lod->lod_qos.lq_oss_list, lqo_oss_list) {
                int j = 0;

                for (i = 0; i < lqr->lqr_pool.op_count; i++) {
                        int next;

                        if (!cfs_bitmap_check(lod->lod_ost_bitmap,
                                                src_pool->op_array[i]))
                                continue;

                        ost = OST_TGT(lod,src_pool->op_array[i]);
                        LASSERT(ost && ost->ltd_ost);
                        if (ost->ltd_qos.ltq_oss != oss)
                                continue;

                        /* Evenly space these OSTs across arrayspace */
                        next = j * lqr->lqr_pool.op_count / oss->lqo_ost_count;
                        while (lqr->lqr_pool.op_array[next] != LOV_QOS_EMPTY)
                                next = (next + 1) % lqr->lqr_pool.op_count;

                        lqr->lqr_pool.op_array[next] = src_pool->op_array[i];
                        j++;
                        placed++;
                }
        }

        lqr->lqr_dirty = 0;
        up_write(&lod->lod_qos.lq_rw_sem);

        if (placed != real_count) {
                /* This should never happen */
                LCONSOLE_ERROR_MSG(0x14e, "Failed to place all OSTs in the "
                                   "round-robin list (%d of %d).\n",
                                   placed, real_count);
                for (i = 0; i < lqr->lqr_pool.op_count; i++) {
                        LCONSOLE(D_WARNING, "rr #%d ost idx=%d\n", i,
                                 lqr->lqr_pool.op_array[i]);
                }
                lqr->lqr_dirty = 1;
                RETURN(-EAGAIN);
        }

#if 0
        for (i = 0; i < lqr->lqr_pool.op_count; i++)
                QOS_CONSOLE("rr #%d ost idx=%d\n", i, lqr->lqr_pool.op_array[i]);
#endif

        RETURN(0);
}

/**
 * Instantiate and declare creation of a new object.
 *
 * The function instantiates LU representation for a new object on the
 * specified device. Also it declares an intention to create that
 * object on the storage target.
 *
 * Note lu_object_anon() is used which is a trick with regard to LU/OSD
 * infrastructure - in the existing precreation framework we can't assign FID
 * at this moment, we do this later once a transaction is started. So the
 * special method instantiates FID-less object in the cache and later it
 * will get a FID and proper placement in LU cache.
 *
 * \param[in] env       execution environment for this thread
 * \param[in] d         LOD device
 * \param[in] ost_idx   OST target index where the object is being created
 * \param[in] th        transaction handle
 *
 * \retval              object ptr on success, ERR_PTR() otherwise
 */
static struct dt_object *lod_qos_declare_object_on(const struct lu_env *env,
                                                   struct lod_device *d,
                                                   __u32 ost_idx,
                                                   struct thandle *th)
{
        struct lod_tgt_desc *ost;
        struct lu_object *o, *n;
        struct lu_device *nd;
        struct dt_object *dt;
        int               rc;
        ENTRY;

        LASSERT(d);
        LASSERT(ost_idx < d->lod_osts_size);
        ost = OST_TGT(d,ost_idx);
        LASSERT(ost);
        LASSERT(ost->ltd_ost);

        nd = &ost->ltd_ost->dd_lu_dev;

        /*
         * allocate anonymous object with zero fid, real fid
         * will be assigned by OSP within transaction
         * XXX: to be fixed with fully-functional OST fids
         */
        o = lu_object_anon(env, nd, NULL);
        if (IS_ERR(o))
                GOTO(out, dt = ERR_PTR(PTR_ERR(o)));

        n = lu_object_locate(o->lo_header, nd->ld_type);
        if (unlikely(n == NULL)) {
                CERROR("can't find slice\n");
                lu_object_put(env, o);
                GOTO(out, dt = ERR_PTR(-EINVAL));
        }

        dt = container_of(n, struct dt_object, do_lu);

        rc = lod_sub_object_declare_create(env, dt, NULL, NULL, NULL, th);
        if (rc < 0) {
                CDEBUG(D_OTHER, "can't declare creation on #%u: %d\n",
                       ost_idx, rc);
                lu_object_put(env, o);
                dt = ERR_PTR(rc);
        }

out:
        RETURN(dt);
}

/**
 * Calculate a minimum acceptable stripe count.
 *
 * Return an acceptable stripe count depending on flag LOV_USES_DEFAULT_STRIPE:
 * all stripes or 3/4 of stripes.
 *
 * \param[in] stripe_cnt        number of stripes requested
 * \param[in] flags             0 or LOV_USES_DEFAULT_STRIPE
 *
 * \retval                      acceptable stripecount
 */
static int min_stripe_count(__u32 stripe_cnt, int flags)
{
        return (flags & LOV_USES_DEFAULT_STRIPE ?
                        stripe_cnt - (stripe_cnt / 4) : stripe_cnt);
}

#define LOV_CREATE_RESEED_MULT 30
#define LOV_CREATE_RESEED_MIN  2000

/**
 * Check if an OST is full.
 *
 * Check whether an OST should be considered full based
 * on the given statfs data.
 *
 * \param[in] msfs      statfs data
 *
 * \retval false        not full
 * \retval true         full
 */
static int inline lod_qos_dev_is_full(struct obd_statfs *msfs)
{
        __u64 used;
        int   bs = msfs->os_bsize;

        LASSERT(((bs - 1) & bs) == 0);

        /* the minimum of 0.1% used blocks and 1GB bytes. */
        used = min_t(__u64, (msfs->os_blocks - msfs->os_bfree) >> 10,
                        1 << (31 - ffs(bs)));
        return (msfs->os_bavail < used);
}

/**
 * Initialize temporary OST-in-use array.
 *
 * Allocate or extend the array used to mark targets already assigned to a new
 * striping so they are not used more than once.
 *
 * \param[in] env       execution environment for this thread
 * \param[in] stripes   number of items needed in the array
 *
 * \retval 0            on success
 * \retval -ENOMEM      on error
 */
static inline int lod_qos_ost_in_use_clear(const struct lu_env *env,
                                           __u32 stripes)
{
        struct lod_thread_info *info = lod_env_info(env);

        if (info->lti_ea_store_size < sizeof(int) * stripes)
                lod_ea_store_resize(info, stripes * sizeof(int));
        if (info->lti_ea_store_size < sizeof(int) * stripes) {
                CERROR("can't allocate memory for ost-in-use array\n");
                return -ENOMEM;
        }
        memset(info->lti_ea_store, -1, sizeof(int) * stripes);
        return 0;
}

/**
 * Remember a target in the array of used targets.
 *
 * Mark the given target as used for a new striping being created. The status
 * of an OST in a striping can be checked with lod_qos_is_ost_used().
 *
 * \param[in] env       execution environment for this thread
 * \param[in] idx       index in the array
 * \param[in] ost       OST target index to mark as used
 */
static inline void lod_qos_ost_in_use(const struct lu_env *env,
                                      int idx, int ost)
{
        struct lod_thread_info *info = lod_env_info(env);
        int *osts = info->lti_ea_store;

        LASSERT(info->lti_ea_store_size >= idx * sizeof(int));
        osts[idx] = ost;
}

/**
 * Check is OST used in a striping.
 *
 * Checks whether OST with the given index is marked as used in the temporary
 * array (see lod_qos_ost_in_use()).
 *
 * \param[in] env       execution environment for this thread
 * \param[in] ost       OST target index to check
 * \param[in] stripes   the number of items used in the array already
 *
 * \retval 0            not used
 * \retval 1            used
 */
static int lod_qos_is_ost_used(const struct lu_env *env, int ost, __u32 stripes)
{
        struct lod_thread_info *info = lod_env_info(env);
        int *osts = info->lti_ea_store;
        __u32 j;

        for (j = 0; j < stripes; j++) {
                if (osts[j] == ost)
                        return 1;
        }
        return 0;
}

static int lod_check_and_reserve_ost(const struct lu_env *env,
                                     struct lod_device *m,
                                     struct obd_statfs *sfs, __u32 ost_idx,
                                     __u32 speed, __u32 *s_idx,
                                     struct dt_object **stripe,
                                     struct thandle *th)
{
        struct dt_object   *o;
        __u32 stripe_idx = *s_idx;
        int rc;

        rc = lod_statfs_and_check(env, m, ost_idx, sfs);
        if (rc) {
                /* this OSP doesn't feel well */
                goto out_return;
        }

        /*
         * skip full devices
         */
        if (lod_qos_dev_is_full(sfs)) {
                QOS_DEBUG("#%d is full\n", ost_idx);
                goto out_return;
        }

        /*
         * We expect number of precreated objects in f_ffree at
         * the first iteration, skip OSPs with no objects ready
         */
        if (sfs->os_fprecreated == 0 && speed == 0) {
                QOS_DEBUG("#%d: precreation is empty\n", ost_idx);
                goto out_return;
        }

        /*
         * try to use another OSP if this one is degraded
         */
        if (sfs->os_state & OS_STATE_DEGRADED && speed < 2) {
                QOS_DEBUG("#%d: degraded\n", ost_idx);
                goto out_return;
        }

        /*
         * do not put >1 objects on a single OST
         */
        if (speed && lod_qos_is_ost_used(env, ost_idx, stripe_idx))
                goto out_return;

        o = lod_qos_declare_object_on(env, m, ost_idx, th);
        if (IS_ERR(o)) {
                CDEBUG(D_OTHER, "can't declare new object on #%u: %d\n",
                       ost_idx, (int) PTR_ERR(o));
                rc = PTR_ERR(o);
                goto out_return;
        }

        /*
         * We've successfully declared (reserved) an object
         */
        lod_qos_ost_in_use(env, stripe_idx, ost_idx);
        stripe[stripe_idx] = o;
        stripe_idx++;
        *s_idx = stripe_idx;

out_return:
        return rc;
}

/**
 * Allocate a striping using round-robin algorithm.
 *
 * Allocates a new striping using round-robin algorithm. The function refreshes
 * all the internal structures (statfs cache, array of available OSTs sorted
 * with regard to OSS, etc). The number of stripes required is taken from the
 * object (must be prepared by the caller), but can change if the flag
 * LOV_USES_DEFAULT_STRIPE is supplied. The caller should ensure nobody else
 * is trying to create a striping on the object in parallel. All the internal
 * structures (like pools, etc) are protected and no additional locking is
 * required. The function succeeds even if a single stripe is allocated. To save
 * time we give priority to targets which already have objects precreated.
 * Full OSTs are skipped (see lod_qos_dev_is_full() for the details).
 *
 * \param[in] env       execution environment for this thread
 * \param[in] lo        LOD object
 * \param[out] stripe   striping created
 * \param[in] flags     allocation flags (0 or LOV_USES_DEFAULT_STRIPE)
 * \param[in] th        transaction handle
 *
 * \retval 0            on success
 * \retval -ENOSPC      if not enough OSTs are found
 * \retval negative     negated errno for other failures
 */
static int lod_alloc_rr(const struct lu_env *env, struct lod_object *lo,
                        struct dt_object **stripe, int flags,
                        struct thandle *th)
{
        struct lod_device *m = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
        struct obd_statfs *sfs = &lod_env_info(env)->lti_osfs;
        struct pool_desc  *pool = NULL;
        struct ost_pool   *osts;
        struct lod_qos_rr *lqr;
        unsigned int       i, array_idx;
        int                rc;
        __u32              ost_start_idx_temp;
        int                speed = 0;
        int                ost_connecting = 0;
        __u32              stripe_idx = 0;
        __u32              stripe_cnt = lo->ldo_stripenr;
        __u32              stripe_cnt_min = min_stripe_count(stripe_cnt, flags);
        __u32              ost_idx;
        ENTRY;

        if (lo->ldo_pool)
                pool = lod_find_pool(m, lo->ldo_pool);

        if (pool != NULL) {
                down_read(&pool_tgt_rw_sem(pool));
                osts = &(pool->pool_obds);
                lqr = &(pool->pool_rr);
        } else {
                osts = &(m->lod_pool_info);
                lqr = &(m->lod_qos.lq_rr);
        }

        rc = lod_qos_calc_rr(m, osts, lqr);
        if (rc)
                GOTO(out, rc);

        rc = lod_qos_ost_in_use_clear(env, lo->ldo_stripenr);
        if (rc)
                GOTO(out, rc);

        down_read(&m->lod_qos.lq_rw_sem);
        spin_lock(&lqr->lqr_alloc);
        if (--lqr->lqr_start_count <= 0) {
                lqr->lqr_start_idx = cfs_rand() % osts->op_count;
                lqr->lqr_start_count =
                        (LOV_CREATE_RESEED_MIN / max(osts->op_count, 1U) +
                         LOV_CREATE_RESEED_MULT) * max(osts->op_count, 1U);
        } else if (stripe_cnt_min >= osts->op_count ||
                        lqr->lqr_start_idx > osts->op_count) {
                /* If we have allocated from all of the OSTs, slowly
                 * precess the next start if the OST/stripe count isn't
                 * already doing this for us. */
                lqr->lqr_start_idx %= osts->op_count;
                if (stripe_cnt > 1 && (osts->op_count % stripe_cnt) != 1)
                        ++lqr->lqr_offset_idx;
        }
        ost_start_idx_temp = lqr->lqr_start_idx;

repeat_find:

        QOS_DEBUG("pool '%s' want %d startidx %d startcnt %d offset %d "
                  "active %d count %d\n",
                  lo->ldo_pool ? lo->ldo_pool : "",
                  stripe_cnt, lqr->lqr_start_idx, lqr->lqr_start_count,
                  lqr->lqr_offset_idx, osts->op_count, osts->op_count);

        for (i = 0; i < osts->op_count && stripe_idx < lo->ldo_stripenr; i++) {
                array_idx = (lqr->lqr_start_idx + lqr->lqr_offset_idx) %
                                osts->op_count;
                ++lqr->lqr_start_idx;
                ost_idx = lqr->lqr_pool.op_array[array_idx];

                QOS_DEBUG("#%d strt %d act %d strp %d ary %d idx %d\n",
                          i, lqr->lqr_start_idx, /* XXX: active*/ 0,
                          stripe_idx, array_idx, ost_idx);

                if ((ost_idx == LOV_QOS_EMPTY) ||
                    !cfs_bitmap_check(m->lod_ost_bitmap, ost_idx))
                        continue;

                /* Fail Check before osc_precreate() is called
                   so we can only 'fail' single OSC. */
                if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_PRECREATE) && ost_idx == 0)
                        continue;

                spin_unlock(&lqr->lqr_alloc);
                rc = lod_check_and_reserve_ost(env, m, sfs, ost_idx, speed,
                                               &stripe_idx, stripe, th);
                spin_lock(&lqr->lqr_alloc);

                if (rc != 0 && OST_TGT(m, ost_idx)->ltd_connecting)
                        ost_connecting = 1;
        }
        if ((speed < 2) && (stripe_idx < stripe_cnt_min)) {
                /* Try again, allowing slower OSCs */
                speed++;
                lqr->lqr_start_idx = ost_start_idx_temp;

                ost_connecting = 0;
                goto repeat_find;
        }

        spin_unlock(&lqr->lqr_alloc);
        up_read(&m->lod_qos.lq_rw_sem);

        if (stripe_idx) {
                lo->ldo_stripenr = stripe_idx;
                /* at least one stripe is allocated */
                rc = 0;
        } else {
                /* nobody provided us with a single object */
                if (ost_connecting)
                        rc = -EINPROGRESS;
                else
                        rc = -ENOSPC;
        }

out:
        if (pool != NULL) {
                up_read(&pool_tgt_rw_sem(pool));
                /* put back ref got by lod_find_pool() */
                lod_pool_putref(pool);
        }

        RETURN(rc);
}

/**
 * Allocate a specific striping layout on a user defined set of OSTs.
 *
 * Allocates new striping using the OST index range provided by the data from
 * the lmm_obejcts contained in the lov_user_md passed to this method. Full
 * OSTs are not considered. The exact order of OSTs requested by the user
 * is respected as much as possible depending on OST status. The number of
 * stripes needed and stripe offset are taken from the object. If that number
 * can not be met, then the function returns a failure and then it's the
 * caller's responsibility to release the stripes allocated. All the internal
 * structures are protected, but no concurrent allocation is allowed on the
 * same objects.
 *
 * \param[in] env       execution environment for this thread
 * \param[in] lo        LOD object
 * \param[out] stripe   striping created
 * \param[in] lum       stripe md to specify list of OSTs
 * \param[in] th        transaction handle
 *
 * \retval 0            on success
 * \retval -ENODEV      OST index does not exist on file system
 * \retval -EINVAL      requested OST index is invalid
 * \retval negative     negated errno on error
 */
static int lod_alloc_ost_list(const struct lu_env *env,
                              struct lod_object *lo, struct dt_object **stripe,
                              struct lov_user_md *lum, struct thandle *th)
{
        struct lod_device       *m = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
        struct obd_statfs       *sfs = &lod_env_info(env)->lti_osfs;
        struct dt_object        *o;
        struct lov_user_md_v3   *v3;
        unsigned int            array_idx = 0;
        int                     stripe_count = 0;
        int                     i;
        int                     rc;
        ENTRY;

        /* for specific OSTs layout */
        LASSERT(lum != NULL && lum->lmm_magic == LOV_USER_MAGIC_SPECIFIC);
        lustre_print_user_md(D_OTHER, lum, __func__);

        rc = lod_qos_ost_in_use_clear(env, lo->ldo_stripenr);
        if (rc < 0)
                RETURN(rc);

        v3 = (struct lov_user_md_v3 *)lum;
        for (i = 0; i < lo->ldo_stripenr; i++) {
                if (v3->lmm_objects[i].l_ost_idx == lo->ldo_def_stripe_offset) {
                        array_idx = i;
                        break;
                }
        }
        if (i == lo->ldo_stripenr) {
                CDEBUG(D_OTHER,
                       "%s: start index %d not in the specified list of OSTs\n",
                       lod2obd(m)->obd_name, lo->ldo_def_stripe_offset);
                RETURN(-EINVAL);
        }

        for (i = 0; i < lo->ldo_stripenr;
             i++, array_idx = (array_idx + 1) % lo->ldo_stripenr) {
                __u32 ost_idx = v3->lmm_objects[array_idx].l_ost_idx;

                if (!cfs_bitmap_check(m->lod_ost_bitmap, ost_idx)) {
                        rc = -ENODEV;
                        break;
                }

                /*
                 * do not put >1 objects on a single OST
                 */
                if (lod_qos_is_ost_used(env, ost_idx, stripe_count)) {
                        rc = -EINVAL;
                        break;
                }

                rc = lod_statfs_and_check(env, m, ost_idx, sfs);
                if (rc < 0) /* this OSP doesn't feel well */
                        break;

                o = lod_qos_declare_object_on(env, m, ost_idx, th);
                if (IS_ERR(o)) {
                        rc = PTR_ERR(o);
                        CDEBUG(D_OTHER,
                               "%s: can't declare new object on #%u: %d\n",
                               lod2obd(m)->obd_name, ost_idx, rc);
                        break;
                }

                /*
                 * We've successfully declared (reserved) an object
                 */
                lod_qos_ost_in_use(env, stripe_count, ost_idx);
                stripe[stripe_count] = o;
                stripe_count++;
        }

        RETURN(rc);
}

/**
 * Allocate a striping on a predefined set of OSTs.
 *
 * Allocates new layout starting from OST index in lo->ldo_def_stripe_offset.
 * Full OSTs are not considered. The exact order of OSTs is not important and
 * varies depending on OST status. The allocation procedure prefers the targets
 * with precreated objects ready. The number of stripes needed and stripe
 * offset are taken from the object. If that number cannot be met, then the
 * function returns an error and then it's the caller's responsibility to
 * release the stripes allocated. All the internal structures are protected,
 * but no concurrent allocation is allowed on the same objects.
 *
 * \param[in] env       execution environment for this thread
 * \param[in] lo        LOD object
 * \param[out] stripe   striping created
 * \param[in] flags     not used
 * \param[in] th        transaction handle
 *
 * \retval 0            on success
 * \retval -ENOSPC      if no OST objects are available at all
 * \retval -EFBIG       if not enough OST objects are found
 * \retval -EINVAL      requested offset is invalid
 * \retval negative     errno on failure
 */
static int lod_alloc_specific(const struct lu_env *env, struct lod_object *lo,
                              struct dt_object **stripe, int flags,
                              struct thandle *th)
{
        struct lod_device *m = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
        struct obd_statfs *sfs = &lod_env_info(env)->lti_osfs;
        struct dt_object  *o;
        __u32              ost_idx;
        unsigned int       i, array_idx, ost_count;
        int                rc, stripe_num = 0;
        int                speed = 0;
        struct pool_desc  *pool = NULL;
        struct ost_pool   *osts;
        ENTRY;

        rc = lod_qos_ost_in_use_clear(env, lo->ldo_stripenr);
        if (rc)
                GOTO(out, rc);

        if (lo->ldo_pool)
                pool = lod_find_pool(m, lo->ldo_pool);

        if (pool != NULL) {
                down_read(&pool_tgt_rw_sem(pool));
                osts = &(pool->pool_obds);
        } else {
                osts = &(m->lod_pool_info);
        }

        ost_count = osts->op_count;

repeat_find:
        /* search loi_ost_idx in ost array */
        array_idx = 0;
        for (i = 0; i < ost_count; i++) {
                if (osts->op_array[i] == lo->ldo_def_stripe_offset) {
                        array_idx = i;
                        break;
                }
        }
        if (i == ost_count) {
                CERROR("Start index %d not found in pool '%s'\n",
                       lo->ldo_def_stripe_offset,
                       lo->ldo_pool ? lo->ldo_pool : "");
                GOTO(out, rc = -EINVAL);
        }

        for (i = 0; i < ost_count;
                        i++, array_idx = (array_idx + 1) % ost_count) {
                ost_idx = osts->op_array[array_idx];

                if (!cfs_bitmap_check(m->lod_ost_bitmap, ost_idx))
                        continue;

                /* Fail Check before osc_precreate() is called
                   so we can only 'fail' single OSC. */
                if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_PRECREATE) && ost_idx == 0)
                        continue;

                /*
                 * do not put >1 objects on a single OST
                 */
                if (lod_qos_is_ost_used(env, ost_idx, stripe_num))
                        continue;

                /* Drop slow OSCs if we can, but not for requested start idx.
                 *
                 * This means "if OSC is slow and it is not the requested
                 * start OST, then it can be skipped, otherwise skip it only
                 * if it is inactive/recovering/out-of-space." */

                rc = lod_statfs_and_check(env, m, ost_idx, sfs);
                if (rc) {
                        /* this OSP doesn't feel well */
                        continue;
                }

                /*
                 * We expect number of precreated objects in f_ffree at
                 * the first iteration, skip OSPs with no objects ready
                 * don't apply this logic to OST specified with stripe_offset
                 */
                if (i != 0 && sfs->os_fprecreated == 0 && speed == 0)
                        continue;

                o = lod_qos_declare_object_on(env, m, ost_idx, th);
                if (IS_ERR(o)) {
                        CDEBUG(D_OTHER, "can't declare new object on #%u: %d\n",
                               ost_idx, (int) PTR_ERR(o));
                        continue;
                }

                /*
                 * We've successfully declared (reserved) an object
                 */
                lod_qos_ost_in_use(env, stripe_num, ost_idx);
                stripe[stripe_num] = o;
                stripe_num++;

                /* We have enough stripes */
                if (stripe_num == lo->ldo_stripenr)
                        GOTO(out, rc = 0);
        }
        if (speed < 2) {
                /* Try again, allowing slower OSCs */
                speed++;
                goto repeat_find;
        }

        /* If we were passed specific striping params, then a failure to
         * meet those requirements is an error, since we can't reallocate
         * that memory (it might be part of a larger array or something).
         */
        CERROR("can't lstripe objid "DFID": have %d want %u\n",
               PFID(lu_object_fid(lod2lu_obj(lo))), stripe_num,
               lo->ldo_stripenr);
        rc = stripe_num == 0 ? -ENOSPC : -EFBIG;
out:
        if (pool != NULL) {
                up_read(&pool_tgt_rw_sem(pool));
                /* put back ref got by lod_find_pool() */
                lod_pool_putref(pool);
        }

        RETURN(rc);
}

/**
 * Check whether QoS allocation should be used.
 *
 * A simple helper to decide when QoS allocation should be used:
 * if it's just a single available target or the used space is
 * evenly distributed among the targets at the moment, then QoS
 * allocation algorithm should not be used.
 *
 * \param[in] lod       LOD device
 *
 * \retval 0            should not be used
 * \retval 1            should be used
 */
static inline int lod_qos_is_usable(struct lod_device *lod)
{
#ifdef FORCE_QOS
        /* to be able to debug QoS code */
        return 1;
#endif

        /* Detect -EAGAIN early, before expensive lock is taken. */
        if (!lod->lod_qos.lq_dirty && lod->lod_qos.lq_same_space)
                return 0;

        if (lod->lod_desc.ld_active_tgt_count < 2)
                return 0;

        return 1;
}

/**
 * Allocate a striping using an algorithm with weights.
 *
 * The function allocates OST objects to create a striping. The algorithm
 * used is based on weights (currently only using the free space), and it's
 * trying to ensure the space is used evenly by OSTs and OSSs. The striping
 * configuration (# of stripes, offset, pool) is taken from the object and
 * is prepared by the caller.
 *
 * If LOV_USES_DEFAULT_STRIPE is not passed and prepared configuration can't
 * be met due to too few OSTs, then allocation fails. If the flag is passed
 * fewer than 3/4 of the requested number of stripes can be allocated, then
 * allocation fails.
 *
 * No concurrent allocation is allowed on the object and this must be ensured
 * by the caller. All the internal structures are protected by the function.
 *
 * The algorithm has two steps: find available OSTs and calculate their
 * weights, then select the OSTs with their weights used as the probability.
 * An OST with a higher weight is proportionately more likely to be selected
 * than one with a lower weight.
 *
 * \param[in] env       execution environment for this thread
 * \param[in] lo        LOD object
 * \param[out] stripe   striping created
 * \param[in] flags     0 or LOV_USES_DEFAULT_STRIPE
 * \param[in] th        transaction handle
 *
 * \retval 0            on success
 * \retval -EAGAIN      not enough OSTs are found for specified stripe count
 * \retval -EINVAL      requested OST index is invalid
 * \retval negative     errno on failure
 */
static int lod_alloc_qos(const struct lu_env *env, struct lod_object *lo,
                         struct dt_object **stripe, int flags,
                         struct thandle *th)
{
        struct lod_device   *m = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
        struct obd_statfs   *sfs = &lod_env_info(env)->lti_osfs;
        struct lod_tgt_desc *ost;
        struct dt_object    *o;
        __u64                total_weight = 0;
        unsigned int         i;
        int                  rc = 0;
        __u32                nfound, good_osts;
        __u32                stripe_cnt = lo->ldo_stripenr;
        __u32                stripe_cnt_min;
        struct pool_desc    *pool = NULL;
        struct ost_pool    *osts;
        ENTRY;

        stripe_cnt_min = min_stripe_count(stripe_cnt, flags);
        if (stripe_cnt_min < 1)
                RETURN(-EINVAL);

        if (lo->ldo_pool)
                pool = lod_find_pool(m, lo->ldo_pool);

        if (pool != NULL) {
                down_read(&pool_tgt_rw_sem(pool));
                osts = &(pool->pool_obds);
        } else {
                osts = &(m->lod_pool_info);
        }

        /* Detect -EAGAIN early, before expensive lock is taken. */
        if (!lod_qos_is_usable(m))
                GOTO(out_nolock, rc = -EAGAIN);

        /* Do actual allocation, use write lock here. */
        down_write(&m->lod_qos.lq_rw_sem);

        /*
         * Check again, while we were sleeping on @lq_rw_sem things could
         * change.
         */
        if (!lod_qos_is_usable(m))
                GOTO(out, rc = -EAGAIN);

        rc = lod_qos_calc_ppo(m);
        if (rc)
                GOTO(out, rc);

        rc = lod_qos_ost_in_use_clear(env, lo->ldo_stripenr);
        if (rc)
                GOTO(out, rc);

        good_osts = 0;
        /* Find all the OSTs that are valid stripe candidates */
        for (i = 0; i < osts->op_count; i++) {
                if (!cfs_bitmap_check(m->lod_ost_bitmap, osts->op_array[i]))
                        continue;

                rc = lod_statfs_and_check(env, m, osts->op_array[i], sfs);
                if (rc) {
                        /* this OSP doesn't feel well */
                        continue;
                }

                /*
                 * skip full devices
                 */
                if (lod_qos_dev_is_full(sfs))
                        continue;

                /* Fail Check before osc_precreate() is called
                   so we can only 'fail' single OSC. */
                if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_PRECREATE) &&
                                   osts->op_array[i] == 0)
                        continue;

                ost = OST_TGT(m,osts->op_array[i]);
                ost->ltd_qos.ltq_usable = 1;
                lod_qos_calc_weight(m, osts->op_array[i]);
                total_weight += ost->ltd_qos.ltq_weight;

                good_osts++;
        }

        QOS_DEBUG("found %d good osts\n", good_osts);

        if (good_osts < stripe_cnt_min)
                GOTO(out, rc = -EAGAIN);

        /* We have enough osts */
        if (good_osts < stripe_cnt)
                stripe_cnt = good_osts;

        /* Find enough OSTs with weighted random allocation. */
        nfound = 0;
        while (nfound < stripe_cnt) {
                __u64 rand, cur_weight;

                cur_weight = 0;
                rc = -ENOSPC;

                if (total_weight) {
#if BITS_PER_LONG == 32
                        rand = cfs_rand() % (unsigned)total_weight;
                        /* If total_weight > 32-bit, first generate the high
                         * 32 bits of the random number, then add in the low
                         * 32 bits (truncated to the upper limit, if needed) */
                        if (total_weight > 0xffffffffULL)
                                rand = (__u64)(cfs_rand() %
                                        (unsigned)(total_weight >> 32)) << 32;
                        else
                                rand = 0;

                        if (rand == (total_weight & 0xffffffff00000000ULL))
                                rand |= cfs_rand() % (unsigned)total_weight;
                        else
                                rand |= cfs_rand();

#else
                        rand = ((__u64)cfs_rand() << 32 | cfs_rand()) %
                                total_weight;
#endif
                } else {
                        rand = 0;
                }

                /* On average, this will hit larger-weighted OSTs more often.
                 * 0-weight OSTs will always get used last (only when rand=0) */
                for (i = 0; i < osts->op_count; i++) {
                        __u32 idx = osts->op_array[i];

                        if (!cfs_bitmap_check(m->lod_ost_bitmap, idx))
                                continue;

                        ost = OST_TGT(m,idx);

                        if (!ost->ltd_qos.ltq_usable)
                                continue;

                        cur_weight += ost->ltd_qos.ltq_weight;
                        QOS_DEBUG("stripe_cnt=%d nfound=%d cur_weight="LPU64
                                  " rand="LPU64" total_weight="LPU64"\n",
                                  stripe_cnt, nfound, cur_weight, rand,
                                  total_weight);

                        if (cur_weight < rand)
                                continue;

                        QOS_DEBUG("stripe=%d to idx=%d\n", nfound, idx);

                        /*
                         * do not put >1 objects on a single OST
                         */
                        if (lod_qos_is_ost_used(env, idx, nfound))
                                continue;
                        lod_qos_ost_in_use(env, nfound, idx);

                        o = lod_qos_declare_object_on(env, m, idx, th);
                        if (IS_ERR(o)) {
                                QOS_DEBUG("can't declare object on #%u: %d\n",
                                          idx, (int) PTR_ERR(o));
                                continue;
                        }
                        stripe[nfound++] = o;
                        lod_qos_used(m, osts, idx, &total_weight);
                        rc = 0;
                        break;
                }

                if (rc) {
                        /* no OST found on this iteration, give up */
                        break;
                }
        }

        if (unlikely(nfound != stripe_cnt)) {
                /*
                 * when the decision to use weighted algorithm was made
                 * we had enough appropriate OSPs, but this state can
                 * change anytime (no space on OST, broken connection, etc)
                 * so it's possible OSP won't be able to provide us with
                 * an object due to just changed state
                 */
                LCONSOLE_INFO("wanted %d, found %d\n", stripe_cnt, nfound);
                for (i = 0; i < nfound; i++) {
                        LASSERT(stripe[i] != NULL);
                        lu_object_put(env, &stripe[i]->do_lu);
                        stripe[i] = NULL;
                }

                /* makes sense to rebalance next time */
                m->lod_qos.lq_dirty = 1;
                m->lod_qos.lq_same_space = 0;

                rc = -EAGAIN;
        }

out:
        up_write(&m->lod_qos.lq_rw_sem);

out_nolock:
        if (pool != NULL) {
                up_read(&pool_tgt_rw_sem(pool));
                /* put back ref got by lod_find_pool() */
                lod_pool_putref(pool);
        }

        RETURN(rc);
}

/**
 * Find largest stripe count the caller can use.
 *
 * Find the maximal possible stripe count not greater than \a stripe_count.
 * Sometimes suggested stripecount can't be reached for a number of reasons:
 * lack of enough active OSTs or the backend does not support EAs that large.
 * If the passed one is 0, then the filesystem's default one is used.
 *
 * \param[in] lod       LOD device
 * \param[in] magic     the format if striping
 * \param[in] stripe_count      count the caller would like to use
 *
 * \retval              the maximum usable stripe count
 */
static __u16 lod_get_stripecnt(struct lod_device *lod, __u32 magic,
                               __u16 stripe_count)
{
        __u32 max_stripes = LOV_MAX_STRIPE_COUNT_OLD;

        if (!stripe_count)
                stripe_count = lod->lod_desc.ld_default_stripe_count;
        if (stripe_count > lod->lod_desc.ld_active_tgt_count)
                stripe_count = lod->lod_desc.ld_active_tgt_count;
        if (!stripe_count)
                stripe_count = 1;

        /* stripe count is based on whether OSD can handle larger EA sizes */
        if (lod->lod_osd_max_easize > 0)
                max_stripes = lov_mds_md_max_stripe_count(
                        lod->lod_osd_max_easize, magic);

        return (stripe_count < max_stripes) ? stripe_count : max_stripes;
}

/**
 * Create in-core respresentation for a fully-defined striping
 *
 * When the caller passes a fully-defined striping (i.e. everything including
 * OST object FIDs are defined), then we still need to instantiate LU-cache
 * with the objects representing the stripes defined. This function completes
 * that task.
 *
 * \param[in] env       execution environment for this thread
 * \param[in] mo        LOD object
 * \param[in] buf       buffer containing the striping
 *
 * \retval 0            on success
 * \retval negative     negated errno on error
 */
static int lod_use_defined_striping(const struct lu_env *env,
                                    struct lod_object *mo,
                                    const struct lu_buf *buf)
{
        struct lov_mds_md_v1   *v1 = buf->lb_buf;
        struct lov_mds_md_v3   *v3 = buf->lb_buf;
        struct lov_ost_data_v1 *objs;
        __u32                   magic;
        int                     rc = 0;
        ENTRY;

        magic = le32_to_cpu(v1->lmm_magic);
        if (magic == LOV_MAGIC_V1_DEF) {
                magic = LOV_MAGIC_V1;
                objs = &v1->lmm_objects[0];
        } else if (magic == LOV_MAGIC_V3_DEF) {
                magic = LOV_MAGIC_V3;
                objs = &v3->lmm_objects[0];
                lod_object_set_pool(mo, v3->lmm_pool_name);
        } else {
                GOTO(out, rc = -EINVAL);
        }

        mo->ldo_pattern = le32_to_cpu(v1->lmm_pattern);
        mo->ldo_stripe_size = le32_to_cpu(v1->lmm_stripe_size);
        mo->ldo_stripenr = le16_to_cpu(v1->lmm_stripe_count);
        mo->ldo_layout_gen = le16_to_cpu(v1->lmm_layout_gen);

        /* fixup for released file before object initialization */
        if (mo->ldo_pattern & LOV_PATTERN_F_RELEASED) {
                mo->ldo_released_stripenr = mo->ldo_stripenr;
                mo->ldo_stripenr = 0;
        }

        LASSERT(buf->lb_len >= lov_mds_md_size(mo->ldo_stripenr, magic));

        if (mo->ldo_stripenr > 0)
                rc = lod_initialize_objects(env, mo, objs);

out:
        RETURN(rc);
}

/**
 * Parse suggested striping configuration.
 *
 * The caller gets a suggested striping configuration from a number of sources
 * including per-directory default and applications. Then it needs to verify
 * the suggested striping is valid, apply missing bits and store the resulting
 * configuration in the object to be used by the allocator later. Must not be
 * called concurrently against the same object. It's OK to provide a
 * fully-defined striping.
 *
 * \param[in] env       execution environment for this thread
 * \param[in] lo        LOD object
 * \param[in] buf       buffer containing the striping
 *
 * \retval 0            on success
 * \retval negative     negated errno on error
 */
static int lod_qos_parse_config(const struct lu_env *env,
                                struct lod_object *lo,
                                const struct lu_buf *buf)
{
        struct lod_device     *d = lu2lod_dev(lod2lu_obj(lo)->lo_dev);
        struct lov_user_md_v1 *v1 = NULL;
        struct lov_user_md_v3 *v3 = NULL;
        char                  *pool_name = NULL;
        __u32                  magic;
        int                    rc;
        unsigned int           size;
        ENTRY;

        if (buf == NULL || buf->lb_buf == NULL || buf->lb_len == 0)
                RETURN(0);

        v3 = buf->lb_buf;
        v1 = buf->lb_buf;
        magic = v1->lmm_magic;

        if (unlikely(magic == LOV_MAGIC_V1_DEF || magic == LOV_MAGIC_V3_DEF)) {
                /* try to use as fully defined striping */
                rc = lod_use_defined_striping(env, lo, buf);
                RETURN(rc);
        }

        switch (magic) {
        case __swab32(LOV_USER_MAGIC_V1):
                lustre_swab_lov_user_md_v1(v1);
                magic = v1->lmm_magic;
                /* fall through */
        case LOV_USER_MAGIC_V1:
                size = sizeof(*v1);
                break;

        case __swab32(LOV_USER_MAGIC_V3):
                lustre_swab_lov_user_md_v3(v3);
                magic = v3->lmm_magic;
                /* fall through */
        case LOV_USER_MAGIC_V3:
                size = sizeof(*v3);
                pool_name = v3->lmm_pool_name;
                break;

        case __swab32(LOV_USER_MAGIC_SPECIFIC):
                lustre_swab_lov_user_md_v3(v3);
                lustre_swab_lov_user_md_objects(v3->lmm_objects,
                                                v3->lmm_stripe_count);
                magic = v3->lmm_magic;
                /* fall through */
        case LOV_USER_MAGIC_SPECIFIC:
                if (v3->lmm_stripe_offset == LOV_OFFSET_DEFAULT)
                        v3->lmm_stripe_offset = v3->lmm_objects[0].l_ost_idx;
                if (v3->lmm_pool_name[0] != '\0')
                        pool_name = v3->lmm_pool_name;
                size = lov_user_md_size(v3->lmm_stripe_count,
                                        LOV_USER_MAGIC_SPECIFIC);
                break;

        default:
                CERROR("%s: unrecognized magic %X\n",
                       lod2obd(d)->obd_name, magic);
                RETURN(-EINVAL);
        }

        if (unlikely(buf->lb_len < size)) {
                CERROR("%s: wrong size: %zd, expect: %u\n",
                       lod2obd(d)->obd_name, buf->lb_len, size);
                RETURN(-EINVAL);
        }

        lustre_print_user_md(D_OTHER, v1, "parse config");

        v1->lmm_magic = magic;
        if (v1->lmm_pattern == 0)
                v1->lmm_pattern = LOV_PATTERN_RAID0;
        if (lov_pattern(v1->lmm_pattern) != LOV_PATTERN_RAID0) {
                CERROR("%s: invalid pattern: %x\n",
                       lod2obd(d)->obd_name, v1->lmm_pattern);
                RETURN(-EINVAL);
        }
        lo->ldo_pattern = v1->lmm_pattern;

        if (v1->lmm_stripe_size > 0)
                lo->ldo_stripe_size = v1->lmm_stripe_size;

        if (lo->ldo_stripe_size & (LOV_MIN_STRIPE_SIZE - 1))
                lo->ldo_stripe_size = LOV_MIN_STRIPE_SIZE;

        if (v1->lmm_stripe_count > 0)
                lo->ldo_stripenr = v1->lmm_stripe_count;

        lo->ldo_def_stripe_offset = v1->lmm_stripe_offset;

        lod_object_set_pool(lo, NULL);
        if (pool_name != NULL) {
                struct pool_desc *pool;

                /* In the function below, .hs_keycmp resolves to
                 * pool_hashkey_keycmp() */
                /* coverity[overrun-buffer-val] */
                pool = lod_find_pool(d, pool_name);
                if (pool != NULL) {
                        if (lo->ldo_def_stripe_offset != LOV_OFFSET_DEFAULT) {
                                rc = lod_check_index_in_pool(
                                               lo->ldo_def_stripe_offset, pool);
                                if (rc < 0) {
                                        lod_pool_putref(pool);
                                        CERROR("%s: invalid offset, %u\n",
                                               lod2obd(d)->obd_name,
                                               lo->ldo_def_stripe_offset);
                                        RETURN(-EINVAL);
                                }
                        }

                        if (lo->ldo_stripenr > pool_tgt_count(pool))
                                lo->ldo_stripenr = pool_tgt_count(pool);

                        lod_pool_putref(pool);
                }

                lod_object_set_pool(lo, pool_name);
        }

        /* fixup for released file */
        if (lo->ldo_pattern & LOV_PATTERN_F_RELEASED) {
                lo->ldo_released_stripenr = lo->ldo_stripenr;
                lo->ldo_stripenr = 0;
        }

        RETURN(0);
}

/**
 * Create a striping for an obejct.
 *
 * The function creates a new striping for the object. A buffer containing
 * configuration hints can be provided optionally. The function tries QoS
 * algorithm first unless free space is distributed evenly among OSTs, but
 * by default RR algorithm is preferred due to internal concurrency (QoS is
 * serialized). The caller must ensure no concurrent calls to the function
 * are made against the same object.
 *
 * \param[in] env       execution environment for this thread
 * \param[in] lo        LOD object
 * \param[in] attr      attributes OST objects will be declared with
 * \param[in] buf       suggested striping configuration or NULL
 * \param[in] th        transaction handle
 *
 * \retval 0            on success
 * \retval negative     negated errno on error
 */
int lod_qos_prep_create(const struct lu_env *env, struct lod_object *lo,
                        struct lu_attr *attr, const struct lu_buf *buf,
                        struct thandle *th)
{
        struct lod_device      *d = lu2lod_dev(lod2lu_obj(lo)->lo_dev);
        struct dt_object      **stripe;
        int                     stripe_len;
        int                     flag = LOV_USES_ASSIGNED_STRIPE;
        int                     i, rc;
        ENTRY;

        LASSERT(lo);

        /* no OST available */
        /* XXX: should we be waiting a bit to prevent failures during
         * cluster initialization? */
        if (d->lod_ostnr == 0)
                GOTO(out, rc = -EIO);

        /*
         * by this time, the object's ldo_stripenr and ldo_stripe_size
         * contain default value for striping: taken from the parent
         * or from filesystem defaults
         *
         * in case the caller is passing lovea with new striping config,
         * we may need to parse lovea and apply new configuration
         */
        rc = lod_qos_parse_config(env, lo, buf);
        if (rc)
                GOTO(out, rc);

        /* A released file is being created */
        if (lo->ldo_stripenr == 0)
                GOTO(out, rc = 0);

        if (likely(lo->ldo_stripe == NULL)) {
                struct lov_user_md *lum = NULL;

                /*
                 * no striping has been created so far
                 */
                LASSERT(lo->ldo_stripenr > 0);
                /*
                 * statfs and check OST targets now, since ld_active_tgt_count
                 * could be changed if some OSTs are [de]activated manually.
                 */
                lod_qos_statfs_update(env, d);
                lo->ldo_stripenr = lod_get_stripecnt(d, LOV_MAGIC,
                                                     lo->ldo_stripenr);

                stripe_len = lo->ldo_stripenr;
                OBD_ALLOC(stripe, sizeof(stripe[0]) * stripe_len);
                if (stripe == NULL)
                        GOTO(out, rc = -ENOMEM);

                lod_getref(&d->lod_ost_descs);
                /* XXX: support for non-0 files w/o objects */
                CDEBUG(D_OTHER, "tgt_count %d stripenr %d\n",
                                d->lod_desc.ld_tgt_count, stripe_len);

                if (buf != NULL && buf->lb_buf != NULL)
                        lum = buf->lb_buf;

                if (lum != NULL && lum->lmm_magic == LOV_USER_MAGIC_SPECIFIC) {
                        rc = lod_alloc_ost_list(env, lo, stripe, lum, th);
                } else if (lo->ldo_def_stripe_offset == LOV_OFFSET_DEFAULT) {
                        rc = lod_alloc_qos(env, lo, stripe, flag, th);
                        if (rc == -EAGAIN)
                                rc = lod_alloc_rr(env, lo, stripe, flag, th);
                } else {
                        rc = lod_alloc_specific(env, lo, stripe, flag, th);
                }
                lod_putref(d, &d->lod_ost_descs);

                if (rc < 0) {
                        for (i = 0; i < stripe_len; i++)
                                if (stripe[i] != NULL)
                                        lu_object_put(env, &stripe[i]->do_lu);

                        OBD_FREE(stripe, sizeof(stripe[0]) * stripe_len);
                        lo->ldo_stripenr = 0;
                } else {
                        lo->ldo_stripe = stripe;
                        lo->ldo_stripes_allocated = stripe_len;
                }
        } else {
                /*
                 * lod_qos_parse_config() found supplied buf as a predefined
                 * striping (not a hint), so it allocated all the object
                 * now we need to create them
                 */
                for (i = 0; i < lo->ldo_stripenr; i++) {
                        struct dt_object  *o;

                        o = lo->ldo_stripe[i];
                        LASSERT(o);

                        rc = lod_sub_object_declare_create(env, o, attr, NULL,
                                                           NULL, th);
                        if (rc < 0) {
                                CERROR("can't declare create: %d\n", rc);
                                break;
                        }
                }
        }

out:
        RETURN(rc);
}