Whamcloud - gitweb
LU-1303 lod: QoS allocation policy
authorAlex Zhuravlev <alexey.zhuravlev@intel.com>
Thu, 20 Sep 2012 11:56:31 +0000 (15:56 +0400)
committerAndreas Dilger <adilger@whamcloud.com>
Fri, 28 Sep 2012 22:36:48 +0000 (18:36 -0400)
basically a copy of lov_qos.c implemented with OSD API.

the patch introduces a special lu-helpers allowing objects
with zero fid - used by LOD/OSP as fid is being created late.
this is a requirement of current orphan cleanup protocol.
probably will be replace with natural fids generated outside
of OSP/transaction at some point.

Signed-off-by: Alex Zhuravlev <alexey.zhuravlev@intel.com>
Change-Id: I9f41d8a770eb7b8ca04b729bbf0866e5d27fff8a
Reviewed-on: http://review.whamcloud.com/4058
Reviewed-by: wangdi <di.wang@whamcloud.com>
Tested-by: Hudson
Reviewed-by: Mike Pershin <tappro@whamcloud.com>
Reviewed-by: Andreas Dilger <adilger@whamcloud.com>
lustre/include/lu_object.h
lustre/lod/lod_internal.h
lustre/lod/lod_object.c
lustre/lod/lod_qos.c
lustre/obdclass/lu_object.c

index b33b8b0..d47db12 100644 (file)
@@ -1339,6 +1339,9 @@ void lu_kmem_fini(struct lu_kmem_descr *caches);
 
 void lu_object_assign_fid(const struct lu_env *env, struct lu_object *o,
                          const struct lu_fid *fid);
+struct lu_object *lu_object_anon(const struct lu_env *env,
+                                struct lu_device *dev,
+                                const struct lu_object_conf *conf);
 
 /** @} lu */
 #endif /* __LUSTRE_LU_OBJECT_H */
index 03614bf..a271101 100644 (file)
@@ -281,6 +281,9 @@ int lod_pool_add(struct obd_device *obd, char *poolname, char *ostname);
 int lod_pool_remove(struct obd_device *obd, char *poolname, char *ostname);
 
 /* lod_qos.c */
+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);
 int qos_add_tgt(struct lod_device*, struct lod_ost_desc *);
 int qos_del_tgt(struct lod_device *, struct lod_ost_desc *);
 
index 816e6cd..24a07d6 100644 (file)
@@ -654,8 +654,14 @@ int lod_declare_striped_object(const struct lu_env *env, struct dt_object *dt,
                GOTO(out, rc = -ENOMEM);
        }
 
-       /* XXX: there will be a call to QoS here */
-       RETURN(0);
+       /* choose OST and generate appropriate objects */
+       rc = lod_qos_prep_create(env, lo, attr, lovea, th);
+       if (rc) {
+               /* failed to create striping, let's reset
+                * config so that others don't get confused */
+               lod_object_free_striping(env, lo);
+               GOTO(out, rc);
+       }
 
        /*
         * declare storage for striping data
index fbc7299..28e31e4 100644 (file)
@@ -145,3 +145,1272 @@ out:
        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;
+       int                  rc;
+
+       LASSERT(d);
+       ost = OST_TGT(d,index);
+       LASSERT(ost);
+
+       rc = dt_statfs(env, ost->ltd_ost, sfs);
+       if (rc)
+               return rc;
+
+       /* check whether device has changed state (active, inactive) */
+       if (unlikely(sfs->os_blocks == 0 && ost->ltd_active)) {
+               /* turned inactive? */
+               cfs_spin_lock(&d->lod_desc_lock);
+               if (sfs->os_blocks == 0 && ost->ltd_active) {
+                       ost->ltd_active = 0;
+                       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);
+               }
+               cfs_spin_unlock(&d->lod_desc_lock);
+       } else if (unlikely(sfs->os_blocks && 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) {
+                       ost->ltd_active = 1;
+                       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);
+               }
+               cfs_spin_unlock(&d->lod_desc_lock);
+       }
+
+       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)
+{
+       struct obd_device *obd = lod2obd(lod);
+       struct ost_pool   *osts = &(lod->lod_pool_info);
+       int                i, idx, rc = 0;
+       __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;
+
+       cfs_down_write(&lod->lod_qos.lq_rw_sem);
+       if (cfs_time_beforeq_64(max_age, obd->obd_osfs_age))
+               GOTO(out, rc = 0);
+
+       for (i = 0; i < osts->op_count; i++) {
+               idx = osts->op_array[i];
+               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);
+                       break;
+               }
+               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:
+       cfs_up_write(&lod->lod_qos.lq_rw_sem);
+}
+
+/* Recalculate per-object penalties for OSSs and OSTs,
+   depends on size of each ost in an oss */
+static int lod_qos_calc_ppo(struct lod_device *lod)
+{
+       struct lov_qos_oss *oss;
+       __u64               ba_max, ba_min, temp;
+       __u32               num_active;
+       int                 rc, i, 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 */
+       cfs_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;
+               lov_do_div64(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 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. */
+                       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 */
+       cfs_list_for_each_entry(oss, &lod->lod_qos.lq_oss_list, lqo_oss_list) {
+               temp = oss->lqo_bavail >> 1;
+               lov_do_div64(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 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. */
+                       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);
+}
+
+static int lod_qos_calc_weight(struct lod_device *lod, int i)
+{
+       __u64 temp, temp2;
+
+       /* Final ost weight = TGT_BAVAIL - ost_penalty - oss_penalty */
+       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;
+}
+
+/* We just used this index for a stripe; adjust everyone's weights */
+static int lod_qos_used(struct lod_device *lod, struct ost_pool *osts,
+                       __u32 index, __u64 *total_wt)
+{
+       struct lod_ost_desc *ost;
+       struct lov_qos_oss  *oss;
+       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 */
+       cfs_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);
+}
+
+#define LOV_QOS_EMPTY ((__u32)-1)
+/* compute optimal round-robin order, based on OSTs per OSS */
+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;
+       unsigned placed, real_count;
+       int i, rc;
+       ENTRY;
+
+       if (!lqr->lqr_dirty) {
+               LASSERT(lqr->lqr_pool.op_size);
+               RETURN(0);
+       }
+
+       /* Do actual allocation. */
+       cfs_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);
+               cfs_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) {
+               cfs_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;
+       cfs_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;
+       cfs_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);
+}
+
+/**
+ * A helper function to:
+ *   create in-core lu object on the specified OSP
+ *   declare creation of the object
+ * IMPORTANT: at this stage object is anonymouos - it has no fid assigned
+ *            this is a workaround till we have natural FIDs on OST
+ *
+ *            at this point we want to declare (reserve) object for us as
+ *            we can't block at execution (when create method is called).
+ *            otherwise we'd block whole transaction batch
+ */
+static struct dt_object *lod_qos_declare_object_on(const struct lu_env *env,
+                                                  struct lod_device *d,
+                                                  int ost_idx,
+                                                  struct thandle *th)
+{
+       struct lod_ost_desc *ost;
+       struct lu_object *o, *n;
+       struct lu_device *nd;
+       struct dt_object *dt;
+       int               rc;
+       ENTRY;
+
+       LASSERT(d);
+       LASSERT(ost_idx >= 0);
+       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 = dt_declare_create(env, dt, NULL, NULL, NULL, th);
+       if (rc) {
+               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);
+}
+
+static int min_stripe_count(int 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
+
+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 - cfs_ffs(bs)));
+       return (msfs->os_bavail < used);
+}
+
+int lod_ea_store_resize(struct lod_thread_info *info, int size);
+
+static inline int lod_qos_ost_in_use_clear(const struct lu_env *env, int 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, 0, sizeof(int) * stripes);
+       return 0;
+}
+
+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;
+}
+
+static int lod_qos_is_ost_used(const struct lu_env *env, int ost, int stripes)
+{
+       struct lod_thread_info *info = lod_env_info(env);
+       int *osts = info->lti_ea_store;
+       int j;
+
+       for (j = 0; j < stripes; j++) {
+               if (osts[j] == ost)
+                       return 1;
+       }
+       return 0;
+}
+
+/* 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 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 lov_qos_rr *lqr;
+       struct dt_object  *o;
+       unsigned           array_idx;
+       int                i, rc;
+       int                ost_start_idx_temp;
+       int                speed = 0;
+       int                stripe_idx = 0;
+       int                stripe_cnt = lo->ldo_stripenr;
+       int                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) {
+               cfs_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);
+
+       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;
+       }
+       cfs_down_read(&m->lod_qos.lq_rw_sem);
+       ost_start_idx_temp = lqr->lqr_start_idx;
+
+repeat_find:
+       array_idx = (lqr->lqr_start_idx + lqr->lqr_offset_idx) %
+                       osts->op_count;
+
+       QOS_DEBUG("pool '%s' want %d startidx %d startcnt %d offset %d "
+                 "active %d count %d arrayidx %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,
+                 array_idx);
+
+       for (i = 0; i < osts->op_count;
+                       i++, array_idx = (array_idx + 1) % 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;
+
+               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;
+               }
+
+               /*
+                * skip full devices
+                */
+               if (lod_qos_dev_is_full(sfs)) {
+                       QOS_DEBUG("#%d is full\n", ost_idx);
+                       continue;
+               }
+
+               /*
+                * 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) {
+                       QOS_DEBUG("#%d: precreation is empty\n", ost_idx);
+                       continue;
+               }
+
+               /*
+                * try to use another OSP if this one is degraded
+                */
+               if (sfs->os_state == OS_STATE_DEGRADED && speed == 0) {
+                       QOS_DEBUG("#%d: degraded\n", ost_idx);
+                       continue;
+               }
+
+               /*
+                * do not put >1 objects on a single OST
+                */
+               if (speed && lod_qos_is_ost_used(env, ost_idx, stripe_idx))
+                       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));
+                       rc = PTR_ERR(o);
+                       continue;
+               }
+
+               /*
+                * We've successfuly declared (reserved) an object
+                */
+               lod_qos_ost_in_use(env, stripe_idx, ost_idx);
+               lo->ldo_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 */
+               speed++;
+               lqr->lqr_start_idx = ost_start_idx_temp;
+               goto repeat_find;
+       }
+
+       cfs_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 */
+               rc = -ENOSPC;
+       }
+
+out:
+       if (pool != NULL) {
+               cfs_up_read(&pool_tgt_rw_sem(pool));
+               /* put back ref got by lod_find_pool() */
+               lod_pool_putref(pool);
+       }
+
+       RETURN(rc);
+}
+
+/* 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 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;
+       unsigned           ost_idx, array_idx, ost_count;
+       int                i, rc, stripe_num = 0;
+       int                speed = 0;
+       struct pool_desc  *pool = NULL;
+       struct ost_pool   *osts;
+       ENTRY;
+
+       if (lo->ldo_pool)
+               pool = lod_find_pool(m, lo->ldo_pool);
+
+       if (pool != NULL) {
+               cfs_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;
+
+               /* 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 */
+                       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)
+                       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 successfuly declared (reserved) an object
+                */
+               lo->ldo_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).
+        *
+        * We can only get here if lsm_stripe_count was originally > 1.
+        */
+       CERROR("can't lstripe objid "DFID": have %d want %u\n",
+              PFID(lu_object_fid(lod2lu_obj(lo))), stripe_num,
+              lo->ldo_stripenr);
+       rc = -EFBIG;
+out:
+       if (pool != NULL) {
+               cfs_up_read(&pool_tgt_rw_sem(pool));
+               /* put back ref got by lod_find_pool() */
+               lod_pool_putref(pool);
+       }
+
+       RETURN(rc);
+}
+
+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;
+}
+
+/* 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 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 dt_object    *o;
+       __u64                total_weight = 0;
+       int                  nfound, good_osts, i, rc = 0;
+       int                  stripe_cnt = lo->ldo_stripenr;
+       int                  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) {
+               cfs_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. */
+       cfs_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);
+
+       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 */
+                       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))
+                       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++) {
+                       int 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;
+
+                       o = lod_qos_declare_object_on(env, m, idx, th);
+                       if (IS_ERR(o)) {
+                               CERROR("can't declare new object on #%u: %d\n",
+                                      idx, (int) PTR_ERR(o));
+                               continue;
+                       }
+                       lod_qos_ost_in_use(env, nfound, idx);
+                       lo->ldo_stripe[nfound++] = o;
+                       lod_qos_used(m, osts, idx, &total_weight);
+                       rc = 0;
+                       break;
+               }
+
+               /* should never satisfy below condition */
+               if (rc) {
+                       CERROR("Didn't find any OSTs?\n");
+                       break;
+               }
+       }
+       LASSERT(nfound == stripe_cnt);
+
+out:
+       cfs_up_write(&m->lod_qos.lq_rw_sem);
+
+out_nolock:
+       if (pool != NULL) {
+               cfs_up_read(&pool_tgt_rw_sem(pool));
+               /* put back ref got by lod_find_pool() */
+               lod_pool_putref(pool);
+       }
+
+       RETURN(rc);
+}
+
+/* Find the max stripecount we should use */
+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_stripecnt(lod->lod_osd_max_easize,
+                                                  magic);
+
+       return (stripe_count < max_stripes) ? stripe_count : max_stripes;
+}
+
+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;
+       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) {
+               objs = &v1->lmm_objects[0];
+       } else if (magic == LOV_MAGIC_V3_DEF) {
+               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_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);
+       LASSERT(buf->lb_len >= lov_mds_md_size(mo->ldo_stripenr, magic));
+
+       rc = lod_initialize_objects(env, mo, objs);
+
+out:
+       RETURN(rc);
+}
+
+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;
+       struct pool_desc      *pool;
+       __u32                  magic;
+       int                    rc;
+       ENTRY;
+
+       if (buf == NULL || buf->lb_buf == NULL || buf->lb_len == 0)
+               RETURN(0);
+
+       v1 = buf->lb_buf;
+       magic = v1->lmm_magic;
+
+       if (magic == __swab32(LOV_USER_MAGIC_V1))
+               lustre_swab_lov_user_md_v1(v1);
+       else if (magic == __swab32(LOV_USER_MAGIC_V3))
+               lustre_swab_lov_user_md_v3(v3);
+
+       if (unlikely(magic != LOV_MAGIC_V1 && magic != LOV_MAGIC_V3)) {
+               /* try to use as fully defined striping */
+               rc = lod_use_defined_striping(env, lo, buf);
+               RETURN(rc);
+       }
+
+       if (unlikely(buf->lb_len < sizeof(*v1))) {
+               CERROR("wrong size: %u\n", (unsigned) buf->lb_len);
+               RETURN(-EINVAL);
+       }
+
+       if (v1->lmm_pattern != 0 && v1->lmm_pattern != LOV_PATTERN_RAID0) {
+               CERROR("invalid pattern: %x\n", v1->lmm_pattern);
+               RETURN(-EINVAL);
+       }
+
+       if (v1->lmm_stripe_size)
+               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)
+               lo->ldo_stripenr = v1->lmm_stripe_count;
+
+       if ((v1->lmm_stripe_offset >= d->lod_desc.ld_tgt_count) &&
+           (v1->lmm_stripe_offset != (typeof(v1->lmm_stripe_offset))(-1))) {
+               CERROR("invalid offset: %x\n", v1->lmm_stripe_offset);
+               RETURN(-EINVAL);
+       }
+       lo->ldo_def_stripe_offset = v1->lmm_stripe_offset;
+
+       CDEBUG(D_OTHER, "lsm: %u size, %u stripes, %u offset\n",
+              v1->lmm_stripe_size, v1->lmm_stripe_count,
+              v1->lmm_stripe_offset);
+
+       if (v1->lmm_magic == LOV_MAGIC_V3) {
+               if (buf->lb_len < sizeof(*v3)) {
+                       CERROR("wrong size: %u\n", (unsigned) buf->lb_len);
+                       RETURN(-EINVAL);
+               }
+
+               v3 = buf->lb_buf;
+               lod_object_set_pool(lo, v3->lmm_pool_name);
+
+               pool = lod_find_pool(d, v3->lmm_pool_name);
+               if (pool != NULL) {
+                       if (lo->ldo_def_stripe_offset !=
+                           (typeof(v1->lmm_stripe_offset))(-1)) {
+                               rc = lo->ldo_def_stripe_offset;
+                               rc = lod_check_index_in_pool(rc, pool);
+                               if (rc < 0) {
+                                       lod_pool_putref(pool);
+                                       CERROR("invalid offset\n");
+                                       RETURN(-EINVAL);
+                               }
+                       }
+
+                       if (lo->ldo_stripenr > pool_tgt_count(pool))
+                               lo->ldo_stripenr= pool_tgt_count(pool);
+
+                       lod_pool_putref(pool);
+               }
+       } else
+               lod_object_set_pool(lo, NULL);
+
+       RETURN(0);
+}
+
+/*
+ * buf should be NULL or contain striping settings
+ */
+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;
+       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);
+
+       if (likely(lo->ldo_stripe == NULL)) {
+               /*
+                * no striping has been created so far
+                */
+               LASSERT(lo->ldo_stripenr > 0);
+               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)
+                       GOTO(out, rc = -ENOMEM);
+               lo->ldo_stripes_allocated = lo->ldo_stripenr;
+
+               rc = lod_qos_ost_in_use_clear(env, lo->ldo_stripenr);
+               if (rc)
+                       GOTO(out, rc);
+
+               lod_getref(d);
+               /* XXX: support for non-0 files w/o objects */
+               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);
+                       if (rc == -EAGAIN)
+                               rc = lod_alloc_rr(env, lo, flag, th);
+               } else
+                       rc = lod_alloc_specific(env, lo, flag, th);
+               lod_putref(d);
+       } 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 = dt_declare_create(env, o, attr, NULL, NULL, th);
+                       if (rc) {
+                               CERROR("can't declare create: %d\n", rc);
+                               break;
+                       }
+               }
+       }
+
+out:
+       RETURN(rc);
+}
+
index 9eef99e..9eaa1ae 100644 (file)
@@ -75,11 +75,32 @@ void lu_object_put(const struct lu_env *env, struct lu_object *o)
         struct lu_site          *site;
         struct lu_object        *orig;
         cfs_hash_bd_t            bd;
+       const struct lu_fid     *fid;
 
         top  = o->lo_header;
         site = o->lo_dev->ld_site;
         orig = o;
 
+       /*
+        * till we have full fids-on-OST implemented anonymous objects
+        * are possible in OSP. such an object isn't listed in the site
+        * so we should not remove it from the site.
+        */
+       fid = lu_object_fid(o);
+       if (fid_is_zero(fid)) {
+               LASSERT(top->loh_hash.next == NULL
+                       && top->loh_hash.pprev == NULL);
+               LASSERT(cfs_list_empty(&top->loh_lru));
+               if (!cfs_atomic_dec_and_test(&top->loh_ref))
+                       return;
+               cfs_list_for_each_entry_reverse(o, &top->loh_layers, lo_linkage) {
+                       if (o->lo_ops->loo_object_release != NULL)
+                               o->lo_ops->loo_object_release(env, o);
+               }
+               lu_object_free(env, orig);
+               return;
+       }
+
         cfs_hash_bd_get(site->ls_obj_hash, &top->loh_fid, &bd);
         bkt = cfs_hash_bd_extra_get(site->ls_obj_hash, &bd);
 
@@ -2091,3 +2112,22 @@ void lu_object_assign_fid(const struct lu_env *env, struct lu_object *o,
        cfs_hash_bd_unlock(hs, &bd, 1);
 }
 EXPORT_SYMBOL(lu_object_assign_fid);
+
+/**
+ * allocates object with 0 (non-assiged) fid
+ * XXX: temporary solution to be able to assign fid in ->do_create()
+ *      till we have fully-functional OST fids
+ */
+struct lu_object *lu_object_anon(const struct lu_env *env,
+                                struct lu_device *dev,
+                                const struct lu_object_conf *conf)
+{
+       struct lu_fid     fid;
+       struct lu_object *o;
+
+       fid_zero(&fid);
+       o = lu_object_alloc(env, dev, &fid, conf);
+
+       return o;
+}
+EXPORT_SYMBOL(lu_object_anon);