Whamcloud - gitweb
LU-4277 lod: handle os_state as a flag, check READONLY
[fs/lustre-release.git] / lustre / lod / lod_qos.c
index 2a70c09..51d6603 100644 (file)
@@ -23,7 +23,7 @@
  * Copyright  2009 Sun Microsystems, Inc. All rights reserved
  * Use is subject to license terms.
  *
- * Copyright (c) 2011, 2012, Whamcloud, Inc.
+ * Copyright (c) 2012, 2013, Intel Corporation.
  */
 /*
  * This file is part of Lustre, http://www.lustre.org/
@@ -59,7 +59,7 @@
 #define TGT_BAVAIL(i) (OST_TGT(lod,i)->ltd_statfs.os_bavail * \
                       OST_TGT(lod,i)->ltd_statfs.os_bsize)
 
-int qos_add_tgt(struct lod_device *lod, struct lod_ost_desc *ost_desc)
+int qos_add_tgt(struct lod_device *lod, struct lod_tgt_desc *ost_desc)
 {
        struct lov_qos_oss *oss = NULL, *temposs;
        struct obd_export  *exp = ost_desc->ltd_exp;
@@ -67,7 +67,7 @@ int qos_add_tgt(struct lod_device *lod, struct lod_ost_desc *ost_desc)
        cfs_list_t         *list;
        ENTRY;
 
-       cfs_down_write(&lod->lod_qos.lq_rw_sem);
+       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
@@ -114,17 +114,17 @@ int qos_add_tgt(struct lod_device *lod, struct lod_ost_desc *ost_desc)
        lod->lod_qos.lq_rr.lqr_dirty = 1;
 
 out:
-       cfs_up_write(&lod->lod_qos.lq_rw_sem);
+       up_write(&lod->lod_qos.lq_rw_sem);
        RETURN(rc);
 }
 
-int qos_del_tgt(struct lod_device *lod, struct lod_ost_desc *ost_desc)
+int qos_del_tgt(struct lod_device *lod, struct lod_tgt_desc *ost_desc)
 {
        struct lov_qos_oss *oss;
        int                 rc = 0;
        ENTRY;
 
-       cfs_down_write(&lod->lod_qos.lq_rw_sem);
+       down_write(&lod->lod_qos.lq_rw_sem);
        oss = ost_desc->ltd_qos.ltq_oss;
        if (!oss)
                GOTO(out, rc = -ENOENT);
@@ -141,14 +141,14 @@ int qos_del_tgt(struct lod_device *lod, struct lod_ost_desc *ost_desc)
        lod->lod_qos.lq_dirty = 1;
        lod->lod_qos.lq_rr.lqr_dirty = 1;
 out:
-       cfs_up_write(&lod->lod_qos.lq_rw_sem);
+       up_write(&lod->lod_qos.lq_rw_sem);
        RETURN(rc);
 }
 
 static int lod_statfs_and_check(const struct lu_env *env, struct lod_device *d,
                                int index, struct obd_statfs *sfs)
 {
-       struct lod_ost_desc *ost;
+       struct lod_tgt_desc *ost;
        int                  rc;
 
        LASSERT(d);
@@ -156,14 +156,18 @@ static int lod_statfs_and_check(const struct lu_env *env, struct lod_device *d,
        LASSERT(ost);
 
        rc = dt_statfs(env, ost->ltd_ost, sfs);
-       if (rc)
-               return rc;
+       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 (unlikely(sfs->os_blocks == 0 && ost->ltd_active)) {
+       if (rc != 0 && ost->ltd_active) {
                /* turned inactive? */
-               cfs_spin_lock(&d->lod_desc_lock);
-               if (sfs->os_blocks == 0 && ost->ltd_active) {
+               spin_lock(&d->lod_desc_lock);
+               if (ost->ltd_active) {
                        ost->ltd_active = 0;
                        LASSERT(d->lod_desc.ld_active_tgt_count > 0);
                        d->lod_desc.ld_active_tgt_count--;
@@ -172,12 +176,14 @@ static int lod_statfs_and_check(const struct lu_env *env, struct lod_device *d,
                        CDEBUG(D_CONFIG, "%s: turns inactive\n",
                               ost->ltd_exp->exp_obd->obd_name);
                }
-               cfs_spin_unlock(&d->lod_desc_lock);
-       } else if (unlikely(sfs->os_blocks && ost->ltd_active == 0)) {
+               spin_unlock(&d->lod_desc_lock);
+       } else if (rc == 0 && ost->ltd_active == 0) {
                /* turned active? */
-               LASSERT(d->lod_desc.ld_active_tgt_count < d->lod_ostnr);
-               cfs_spin_lock(&d->lod_desc_lock);
-               if (sfs->os_blocks && ost->ltd_active == 0) {
+               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;
                        d->lod_desc.ld_active_tgt_count++;
                        d->lod_qos.lq_dirty = 1;
@@ -185,19 +191,12 @@ static int lod_statfs_and_check(const struct lu_env *env, struct lod_device *d,
                        CDEBUG(D_CONFIG, "%s: turns active\n",
                               ost->ltd_exp->exp_obd->obd_name);
                }
-               cfs_spin_unlock(&d->lod_desc_lock);
+               spin_unlock(&d->lod_desc_lock);
        }
 
-       return rc;
+       RETURN(rc);
 }
 
-/*
- * Update statfs data if the current osfs age is older than max_age.
- * If wait is not set, it means that we are called from lov_create()
- * and we should just issue the rpcs without waiting for them to complete.
- * If wait is set, we are called from alloc_qos() and we just have
- * to wait for the request set to complete.
- */
 static void lod_qos_statfs_update(const struct lu_env *env,
                                  struct lod_device *lod)
 {
@@ -207,13 +206,13 @@ static void lod_qos_statfs_update(const struct lu_env *env,
        __u64              max_age, avail;
        ENTRY;
 
-       max_age = cfs_time_shift_64(-2*lod->lod_desc.ld_qos_maxage);
+       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;
 
-       cfs_down_write(&lod->lod_qos.lq_rw_sem);
+       down_write(&lod->lod_qos.lq_rw_sem);
        if (cfs_time_beforeq_64(max_age, obd->obd_osfs_age))
                GOTO(out, rc = 0);
 
@@ -222,11 +221,8 @@ static void lod_qos_statfs_update(const struct lu_env *env,
                avail = OST_TGT(lod,idx)->ltd_statfs.os_bavail;
                rc = lod_statfs_and_check(env, lod, idx,
                                          &OST_TGT(lod,idx)->ltd_statfs);
-               if (rc) {
-                       /* XXX: disable this OST till next refresh? */
-                       CERROR("can't refresh statfs: %d\n", rc);
+               if (rc)
                        break;
-               }
                if (OST_TGT(lod,idx)->ltd_statfs.os_bavail != avail)
                        /* recalculate weigths */
                        lod->lod_qos.lq_dirty = 1;
@@ -234,7 +230,8 @@ static void lod_qos_statfs_update(const struct lu_env *env,
        obd->obd_osfs_age = cfs_time_current_64();
 
 out:
-       cfs_up_write(&lod->lod_qos.lq_rw_sem);
+       up_write(&lod->lod_qos.lq_rw_sem);
+       EXIT;
 }
 
 /* Recalculate per-object penalties for OSSs and OSTs,
@@ -260,9 +257,11 @@ static int lod_qos_calc_ppo(struct lod_device *lod)
                                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. */
+       /*
+        * 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);
@@ -291,7 +290,7 @@ static int lod_qos_calc_ppo(struct lod_device *lod)
 
                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)
+                   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 the penalty by half for every 8x the update
@@ -371,7 +370,7 @@ static int lod_qos_calc_weight(struct lod_device *lod, int i)
 static int lod_qos_used(struct lod_device *lod, struct ost_pool *osts,
                        __u32 index, __u64 *total_wt)
 {
-       struct lod_ost_desc *ost;
+       struct lod_tgt_desc *ost;
        struct lov_qos_oss  *oss;
        int j;
        ENTRY;
@@ -451,7 +450,7 @@ static int lod_qos_calc_rr(struct lod_device *lod, struct ost_pool *src_pool,
                           struct lov_qos_rr *lqr)
 {
        struct lov_qos_oss  *oss;
-       struct lod_ost_desc *ost;
+       struct lod_tgt_desc *ost;
        unsigned placed, real_count;
        int i, rc;
        ENTRY;
@@ -462,7 +461,7 @@ static int lod_qos_calc_rr(struct lod_device *lod, struct ost_pool *src_pool,
        }
 
        /* Do actual allocation. */
-       cfs_down_write(&lod->lod_qos.lq_rw_sem);
+       down_write(&lod->lod_qos.lq_rw_sem);
 
        /*
         * Check again. While we were sleeping on @lq_rw_sem something could
@@ -470,7 +469,7 @@ static int lod_qos_calc_rr(struct lod_device *lod, struct ost_pool *src_pool,
         */
        if (!lqr->lqr_dirty) {
                LASSERT(lqr->lqr_pool.op_size);
-               cfs_up_write(&lod->lod_qos.lq_rw_sem);
+               up_write(&lod->lod_qos.lq_rw_sem);
                RETURN(0);
        }
 
@@ -483,7 +482,7 @@ static int lod_qos_calc_rr(struct lod_device *lod, struct ost_pool *src_pool,
        lqr->lqr_pool.op_count = real_count;
        rc = lod_ost_pool_extend(&lqr->lqr_pool, real_count);
        if (rc) {
-               cfs_up_write(&lod->lod_qos.lq_rw_sem);
+               up_write(&lod->lod_qos.lq_rw_sem);
                RETURN(rc);
        }
        for (i = 0; i < lqr->lqr_pool.op_count; i++)
@@ -518,7 +517,7 @@ static int lod_qos_calc_rr(struct lod_device *lod, struct ost_pool *src_pool,
        }
 
        lqr->lqr_dirty = 0;
-       cfs_up_write(&lod->lod_qos.lq_rw_sem);
+       up_write(&lod->lod_qos.lq_rw_sem);
 
        if (placed != real_count) {
                /* This should never happen */
@@ -557,7 +556,7 @@ static struct dt_object *lod_qos_declare_object_on(const struct lu_env *env,
                                                   int ost_idx,
                                                   struct thandle *th)
 {
-       struct lod_ost_desc *ost;
+       struct lod_tgt_desc *ost;
        struct lu_object *o, *n;
        struct lu_device *nd;
        struct dt_object *dt;
@@ -621,7 +620,7 @@ static int inline lod_qos_dev_is_full(struct obd_statfs *msfs)
 
        /* the minimum of 0.1% used blocks and 1GB bytes. */
        used = min_t(__u64, (msfs->os_blocks - msfs->os_bfree) >> 10,
-                       1 << (31 - cfs_ffs(bs)));
+                       1 << (31 - ffs(bs)));
        return (msfs->os_bavail < used);
 }
 
@@ -637,7 +636,7 @@ static inline int lod_qos_ost_in_use_clear(const struct lu_env *env, int stripes
                CERROR("can't allocate memory for ost-in-use array\n");
                return -ENOMEM;
        }
-       memset(info->lti_ea_store, 0, sizeof(int) * stripes);
+       memset(info->lti_ea_store, -1, sizeof(int) * stripes);
        return 0;
 }
 
@@ -663,9 +662,10 @@ static int lod_qos_is_ost_used(const struct lu_env *env, int ost, int stripes)
        return 0;
 }
 
-/* Allocate objects on osts with round-robin algorithm */
+/* Allocate objects on OSTs with round-robin algorithm */
 static int lod_alloc_rr(const struct lu_env *env, struct lod_object *lo,
-                       int flags, struct thandle *th)
+                       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;
@@ -687,7 +687,7 @@ static int lod_alloc_rr(const struct lu_env *env, struct lod_object *lo,
                pool = lod_find_pool(m, lo->ldo_pool);
 
        if (pool != NULL) {
-               cfs_down_read(&pool_tgt_rw_sem(pool));
+               down_read(&pool_tgt_rw_sem(pool));
                osts = &(pool->pool_obds);
                lqr = &(pool->pool_rr);
        } else {
@@ -717,7 +717,7 @@ static int lod_alloc_rr(const struct lu_env *env, struct lod_object *lo,
                if (stripe_cnt > 1 && (osts->op_count % stripe_cnt) != 1)
                        ++lqr->lqr_offset_idx;
        }
-       cfs_down_read(&m->lod_qos.lq_rw_sem);
+       down_read(&m->lod_qos.lq_rw_sem);
        ost_start_idx_temp = lqr->lqr_start_idx;
 
 repeat_find:
@@ -731,8 +731,8 @@ repeat_find:
                  lqr->lqr_offset_idx, osts->op_count, osts->op_count,
                  array_idx);
 
-       for (i = 0; i < osts->op_count;
-                       i++, array_idx = (array_idx + 1) % osts->op_count) {
+       for (i = 0; i < osts->op_count && stripe_idx < lo->ldo_stripenr;
+            i++, array_idx = (array_idx + 1) % osts->op_count) {
                ++lqr->lqr_start_idx;
                ost_idx = lqr->lqr_pool.op_array[array_idx];
 
@@ -741,7 +741,7 @@ repeat_find:
                          stripe_idx, array_idx, ost_idx);
 
                if ((ost_idx == LOV_QOS_EMPTY) ||
-                               !cfs_bitmap_check(m->lod_ost_bitmap, ost_idx))
+                   !cfs_bitmap_check(m->lod_ost_bitmap, ost_idx))
                        continue;
 
                /* Fail Check before osc_precreate() is called
@@ -752,15 +752,6 @@ repeat_find:
                rc = lod_statfs_and_check(env, m, ost_idx, sfs);
                if (rc) {
                        /* this OSP doesn't feel well */
-                       CERROR("can't statfs #%u: %d\n", ost_idx, rc);
-                       continue;
-               }
-
-               /*
-                * skip empty devices - usually it means inactive device
-                */
-               if (sfs->os_blocks == 0) {
-                       QOS_DEBUG("#%d: inactive\n", ost_idx);
                        continue;
                }
 
@@ -776,7 +767,7 @@ repeat_find:
                 * We expect number of precreated objects in f_ffree at
                 * the first iteration, skip OSPs with no objects ready
                 */
-               if (sfs->os_ffree == 0 && speed == 0) {
+               if (sfs->os_fprecreated == 0 && speed == 0) {
                        QOS_DEBUG("#%d: precreation is empty\n", ost_idx);
                        continue;
                }
@@ -784,7 +775,7 @@ repeat_find:
                /*
                 * try to use another OSP if this one is degraded
                 */
-               if (sfs->os_state == OS_STATE_DEGRADED && speed == 0) {
+               if (sfs->os_state & OS_STATE_DEGRADED && speed < 2) {
                        QOS_DEBUG("#%d: degraded\n", ost_idx);
                        continue;
                }
@@ -807,12 +798,9 @@ repeat_find:
                 * We've successfuly declared (reserved) an object
                 */
                lod_qos_ost_in_use(env, stripe_idx, ost_idx);
-               lo->ldo_stripe[stripe_idx] = o;
+               stripe[stripe_idx] = o;
                stripe_idx++;
 
-               /* We have enough stripes */
-               if (stripe_idx == lo->ldo_stripenr)
-                       break;
        }
        if ((speed < 2) && (stripe_idx < stripe_cnt_min)) {
                /* Try again, allowing slower OSCs */
@@ -821,7 +809,7 @@ repeat_find:
                goto repeat_find;
        }
 
-       cfs_up_read(&m->lod_qos.lq_rw_sem);
+       up_read(&m->lod_qos.lq_rw_sem);
 
        if (stripe_idx) {
                lo->ldo_stripenr = stripe_idx;
@@ -834,7 +822,7 @@ repeat_find:
 
 out:
        if (pool != NULL) {
-               cfs_up_read(&pool_tgt_rw_sem(pool));
+               up_read(&pool_tgt_rw_sem(pool));
                /* put back ref got by lod_find_pool() */
                lod_pool_putref(pool);
        }
@@ -844,7 +832,8 @@ out:
 
 /* alloc objects on osts with specific stripe offset */
 static int lod_alloc_specific(const struct lu_env *env, struct lod_object *lo,
-                             int flags, struct thandle *th)
+                             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;
@@ -856,11 +845,15 @@ static int lod_alloc_specific(const struct lu_env *env, struct lod_object *lo,
        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) {
-               cfs_down_read(&pool_tgt_rw_sem(pool));
+               down_read(&pool_tgt_rw_sem(pool));
                osts = &(pool->pool_obds);
        } else {
                osts = &(m->lod_pool_info);
@@ -896,6 +889,12 @@ repeat_find:
                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
@@ -905,22 +904,15 @@ repeat_find:
                rc = lod_statfs_and_check(env, m, ost_idx, sfs);
                if (rc) {
                        /* this OSP doesn't feel well */
-                       CERROR("can't statfs #%u: %d\n", ost_idx, rc);
                        continue;
                }
 
                /*
-                * skip empty devices - usually it means inactive device
-                */
-               if (sfs->os_blocks == 0)
-                       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_ffree == 0 && speed == 0)
+               if (i != 0 && sfs->os_fprecreated == 0 && speed == 0)
                        continue;
 
                o = lod_qos_declare_object_on(env, m, ost_idx, th);
@@ -933,7 +925,8 @@ repeat_find:
                /*
                 * We've successfuly declared (reserved) an object
                 */
-               lo->ldo_stripe[stripe_num] = o;
+               lod_qos_ost_in_use(env, stripe_num, ost_idx);
+               stripe[stripe_num] = o;
                stripe_num++;
 
                /* We have enough stripes */
@@ -958,7 +951,7 @@ repeat_find:
        rc = -EFBIG;
 out:
        if (pool != NULL) {
-               cfs_up_read(&pool_tgt_rw_sem(pool));
+               up_read(&pool_tgt_rw_sem(pool));
                /* put back ref got by lod_find_pool() */
                lod_pool_putref(pool);
        }
@@ -983,16 +976,17 @@ static inline int lod_qos_is_usable(struct lod_device *lod)
        return 1;
 }
 
-/* Alloc objects on osts with optimization based on:
+/* Alloc objects on OSTs with optimization based on:
    - free space
    - network resources (shared OSS's)
  */
 static int lod_alloc_qos(const struct lu_env *env, struct lod_object *lo,
-                        int flags, struct thandle *th)
+                        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_ost_desc *ost;
+       struct lod_tgt_desc *ost;
        struct dt_object    *o;
        __u64                total_weight = 0;
        int                  nfound, good_osts, i, rc = 0;
@@ -1010,7 +1004,7 @@ static int lod_alloc_qos(const struct lu_env *env, struct lod_object *lo,
                pool = lod_find_pool(m, lo->ldo_pool);
 
        if (pool != NULL) {
-               cfs_down_read(&pool_tgt_rw_sem(pool));
+               down_read(&pool_tgt_rw_sem(pool));
                osts = &(pool->pool_obds);
        } else {
                osts = &(m->lod_pool_info);
@@ -1021,7 +1015,7 @@ static int lod_alloc_qos(const struct lu_env *env, struct lod_object *lo,
                GOTO(out_nolock, rc = -EAGAIN);
 
        /* Do actual allocation, use write lock here. */
-       cfs_down_write(&m->lod_qos.lq_rw_sem);
+       down_write(&m->lod_qos.lq_rw_sem);
 
        /*
         * Check again, while we were sleeping on @lq_rw_sem things could
@@ -1047,17 +1041,10 @@ static int lod_alloc_qos(const struct lu_env *env, struct lod_object *lo,
                rc = lod_statfs_and_check(env, m, osts->op_array[i], sfs);
                if (rc) {
                        /* this OSP doesn't feel well */
-                       CERROR("can't statfs #%u: %d\n", i, rc);
                        continue;
                }
 
                /*
-                * skip empty devices - usually it means inactive device
-                */
-               if (sfs->os_blocks == 0)
-                       continue;
-
-               /*
                 * skip full devices
                 */
                if (lod_qos_dev_is_full(sfs))
@@ -1156,11 +1143,16 @@ static int lod_alloc_qos(const struct lu_env *env, struct lod_object *lo,
                                          idx, (int) PTR_ERR(o));
                                continue;
                        }
-                       lo->ldo_stripe[nfound++] = o;
+                       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)) {
@@ -1173,9 +1165,9 @@ static int lod_alloc_qos(const struct lu_env *env, struct lod_object *lo,
                 */
                LCONSOLE_INFO("wanted %d, found %d\n", stripe_cnt, nfound);
                for (i = 0; i < nfound; i++) {
-                       LASSERT(lo->ldo_stripe[i]);
-                       lu_object_put(env, &lo->ldo_stripe[i]->do_lu);
-                       lo->ldo_stripe[i] = NULL;
+                       LASSERT(stripe[i] != NULL);
+                       lu_object_put(env, &stripe[i]->do_lu);
+                       stripe[i] = NULL;
                }
 
                /* makes sense to rebalance next time */
@@ -1186,11 +1178,11 @@ static int lod_alloc_qos(const struct lu_env *env, struct lod_object *lo,
        }
 
 out:
-       cfs_up_write(&m->lod_qos.lq_rw_sem);
+       up_write(&m->lod_qos.lq_rw_sem);
 
 out_nolock:
        if (pool != NULL) {
-               cfs_up_read(&pool_tgt_rw_sem(pool));
+               up_read(&pool_tgt_rw_sem(pool));
                /* put back ref got by lod_find_pool() */
                lod_pool_putref(pool);
        }
@@ -1223,41 +1215,40 @@ static int lod_use_defined_striping(const struct lu_env *env,
                                    struct lod_object *mo,
                                    const struct lu_buf *buf)
 {
-       struct lod_device      *d = lu2lod_dev(lod2lu_obj(mo)->lo_dev);
        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;
+       int                     rc = 0;
        ENTRY;
 
-       rc = lod_verify_striping(d, buf, 1);
-       if (rc)
-               RETURN(rc);
-
        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);
        }
 
-       /*
-        * LOD shouldn't be aware of recovery at all,
-        * but this track recovery status (to some extent)
-        * to be do additional checks like this one
-        */
-       LASSERT(d->lod_recovery_completed == 0);
-
+       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));
 
-       rc = lod_initialize_objects(env, mo, objs);
+       if (mo->ldo_stripenr > 0)
+               rc = lod_initialize_objects(env, mo, objs);
 
 out:
        RETURN(rc);
@@ -1281,10 +1272,14 @@ static int lod_qos_parse_config(const struct lu_env *env,
        v1 = buf->lb_buf;
        magic = v1->lmm_magic;
 
-       if (magic == __swab32(LOV_USER_MAGIC_V1))
+       if (magic == __swab32(LOV_USER_MAGIC_V1)) {
                lustre_swab_lov_user_md_v1(v1);
-       else if (magic == __swab32(LOV_USER_MAGIC_V3))
+               magic = v1->lmm_magic;
+       } else if (magic == __swab32(LOV_USER_MAGIC_V3)) {
+               v3 = buf->lb_buf;
                lustre_swab_lov_user_md_v3(v3);
+               magic = v3->lmm_magic;
+       }
 
        if (unlikely(magic != LOV_MAGIC_V1 && magic != LOV_MAGIC_V3)) {
                /* try to use as fully defined striping */
@@ -1297,10 +1292,14 @@ static int lod_qos_parse_config(const struct lu_env *env,
                RETURN(-EINVAL);
        }
 
-       if (v1->lmm_pattern != 0 && v1->lmm_pattern != LOV_PATTERN_RAID0) {
+       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("invalid pattern: %x\n", v1->lmm_pattern);
                RETURN(-EINVAL);
        }
+       lo->ldo_pattern = v1->lmm_pattern;
 
        if (v1->lmm_stripe_size)
                lo->ldo_stripe_size = v1->lmm_stripe_size;
@@ -1330,6 +1329,9 @@ static int lod_qos_parse_config(const struct lu_env *env,
                v3 = buf->lb_buf;
                lod_object_set_pool(lo, v3->lmm_pool_name);
 
+               /* In the function below, .hs_keycmp resolves to
+                * pool_hashkey_keycmp() */
+               /* coverity[overrun-buffer-val] */
                pool = lod_find_pool(d, v3->lmm_pool_name);
                if (pool != NULL) {
                        if (lo->ldo_def_stripe_offset !=
@@ -1351,6 +1353,12 @@ static int lod_qos_parse_config(const struct lu_env *env,
        } else
                lod_object_set_pool(lo, NULL);
 
+       /* 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);
 }
 
@@ -1361,9 +1369,11 @@ 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);
-       int flag = LOV_USES_ASSIGNED_STRIPE;
-       int i, rc = 0;
+       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);
@@ -1386,29 +1396,51 @@ int lod_qos_prep_create(const struct lu_env *env, struct lod_object *lo,
        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)) {
                /*
                 * 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);
-               i = sizeof(struct dt_object *) * lo->ldo_stripenr;
-               OBD_ALLOC(lo->ldo_stripe, i);
-               if (lo->ldo_stripe == NULL)
+
+               stripe_len = lo->ldo_stripenr;
+               OBD_ALLOC(stripe, sizeof(stripe[0]) * stripe_len);
+               if (stripe == NULL)
                        GOTO(out, rc = -ENOMEM);
-               lo->ldo_stripes_allocated = lo->ldo_stripenr;
 
-               lod_getref(d);
+               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 (lo->ldo_def_stripe_offset >= d->lod_desc.ld_tgt_count) {
-                       lod_qos_statfs_update(env, d);
-                       rc = lod_alloc_qos(env, lo, flag, th);
+                       rc = lod_alloc_qos(env, lo, stripe, flag, th);
                        if (rc == -EAGAIN)
-                               rc = lod_alloc_rr(env, lo, flag, th);
-               } else
-                       rc = lod_alloc_specific(env, lo, flag, th);
-               lod_putref(d);
+                               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);
+               } else {
+                       lo->ldo_stripe = stripe;
+                       lo->ldo_stripes_allocated = stripe_len;
+               }
        } else {
                /*
                 * lod_qos_parse_config() found supplied buf as a predefined