X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Flod%2Flod_qos.c;h=09a063004b1609431ebb4a32be76c8afaf3d1ded;hp=bf09e1f78de4335fa0d7adbcd4441be00ab943c7;hb=083d62ee6de5ac6cee95c1d2f86b62b75034093b;hpb=aeaa120d0d460ecce81a6d5d37ae0dc13936f590

diff --git a/lustre/lod/lod_qos.c b/lustre/lod/lod_qos.c
index bf09e1f..09a0630 100644
--- a/lustre/lod/lod_qos.c
+++ b/lustre/lod/lod_qos.c
@@ -23,7 +23,7 @@
  * Copyright  2009 Sun Microsystems, Inc. All rights reserved
  * Use is subject to license terms.
  *
- * Copyright (c) 2012, 2016, Intel Corporation.
+ * Copyright (c) 2012, 2017, Intel Corporation.
  */
 /*
  * This file is part of Lustre, http://www.lustre.org/
@@ -77,6 +77,7 @@ int qos_add_tgt(struct lod_device *lod, struct lod_tgt_desc *ost_desc)
 	struct obd_export  *exp = ost_desc->ltd_exp;
 	int		    rc = 0, found = 0;
 	struct list_head   *list;
+	__u32 id = 0;
 	ENTRY;
 
 	down_write(&lod->lod_qos.lq_rw_sem);
@@ -91,6 +92,8 @@ int qos_add_tgt(struct lod_device *lod, struct lod_tgt_desc *ost_desc)
 			found++;
 			break;
 		}
+		if (oss->lqo_id > id)
+			id = oss->lqo_id;
 	}
 
 	if (!found) {
@@ -99,6 +102,8 @@ int qos_add_tgt(struct lod_device *lod, struct lod_tgt_desc *ost_desc)
 			GOTO(out, rc = -ENOMEM);
 		memcpy(&oss->lqo_uuid, &exp->exp_connection->c_remote_uuid,
 		       sizeof(oss->lqo_uuid));
+		++id;
+		oss->lqo_id = id;
 	} else {
 		/* Assume we have to move this one */
 		list_del(&oss->lqo_oss_list);
@@ -211,6 +216,10 @@ static int lod_statfs_and_check(const struct lu_env *env, struct lod_device *d,
 	if (sfs->os_state & OS_STATE_READONLY)
 		rc = -EROFS;
 
+	/* object precreation is skipped on the OST with max_create_count=0 */
+	if (sfs->os_state & OS_STATE_NOPRECREATE)
+		rc = -ENOBUFS;
+
 	/* check whether device has changed state (active, inactive) */
 	if (rc != 0 && ost->ltd_active) {
 		/* turned inactive? */
@@ -259,24 +268,25 @@ static int lod_statfs_and_check(const struct lu_env *env, struct lod_device *d,
  * \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)
+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;
+	struct ost_pool *osts = &(lod->lod_pool_info);
+	time64_t max_age;
+	unsigned int i;
+	u64 avail;
+	int idx;
 	ENTRY;
 
-	max_age = cfs_time_shift_64(-2 * lod->lod_desc.ld_qos_maxage);
+	max_age = ktime_get_seconds() - 2 * lod->lod_desc.ld_qos_maxage;
 
-	if (cfs_time_beforeq_64(max_age, obd->obd_osfs_age))
+	if (obd->obd_osfs_age > max_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))
+
+	if (obd->obd_osfs_age > max_age)
 		goto out;
 
 	for (i = 0; i < osts->op_count; i++) {
@@ -289,7 +299,7 @@ static void lod_qos_statfs_update(const struct lu_env *env,
 			/* recalculate weigths */
 			lod->lod_qos.lq_dirty = 1;
 	}
-	obd->obd_osfs_age = cfs_time_current_64();
+	obd->obd_osfs_age = ktime_get_seconds();
 
 out:
 	up_write(&lod->lod_qos.lq_rw_sem);
@@ -319,7 +329,7 @@ static int lod_qos_calc_ppo(struct lod_device *lod)
 	__u32		    num_active;
 	unsigned int	    i;
 	int		    rc, prio_wide;
-	time_t		    now, age;
+	time64_t	    now, age;
 	ENTRY;
 
 	if (!lod->lod_qos.lq_dirty)
@@ -343,9 +353,10 @@ static int lod_qos_calc_ppo(struct lod_device *lod)
 
 	ba_min = (__u64)(-1);
 	ba_max = 0;
-	now = cfs_time_current_sec();
+	now = ktime_get_real_seconds();
 	/* Calculate OST penalty per object
-	 * (lod ref taken in lod_qos_prep_create()) */
+	 * (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);
@@ -481,7 +492,7 @@ static int lod_qos_used(struct lod_device *lod, struct ost_pool *osts,
 	oss->lqo_penalty >>= 1;
 
 	/* mark the OSS and OST as recently used */
-	ost->ltd_qos.ltq_used = oss->lqo_used = cfs_time_current_sec();
+	ost->ltd_qos.ltq_used = oss->lqo_used = ktime_get_real_seconds();
 
 	/* Set max penalties for this OST and OSS */
 	ost->ltd_qos.ltq_penalty +=
@@ -819,61 +830,122 @@ static int lod_qos_is_ost_used(const struct lu_env *env, int ost, __u32 stripes)
 	return 0;
 }
 
+static inline bool
+lod_obj_is_ost_use_skip_cb(const struct lu_env *env, struct lod_object *lo,
+			   int comp_idx, struct lod_obj_stripe_cb_data *data)
+{
+	struct lod_layout_component *comp = &lo->ldo_comp_entries[comp_idx];
+
+	return comp->llc_ost_indices == NULL;
+}
+
+static inline int
+lod_obj_is_ost_use_cb(const struct lu_env *env, struct lod_object *lo,
+		      int comp_idx, struct lod_obj_stripe_cb_data *data)
+{
+	struct lod_layout_component *comp = &lo->ldo_comp_entries[comp_idx];
+	int i;
+
+	for (i = 0; i < comp->llc_stripe_count; i++) {
+		if (comp->llc_ost_indices[i] == data->locd_ost_index) {
+			data->locd_ost_index = -1;
+			return -EEXIST;
+		}
+	}
+
+	return 0;
+}
+
 /**
  * Check is OST used in a composite layout
  *
- * \param[in] inuse	all inuse ost indexs
+ * \param[in] lo	lod object
  * \param[in] ost	OST target index to check
  *
- * \retval 0		not used
- * \retval 1		used
+ * \retval false	not used
+ * \retval true		used
  */
-static inline int lod_comp_is_ost_used(struct ost_pool *inuse, int ost)
+static inline bool lod_comp_is_ost_used(const struct lu_env *env,
+				       struct lod_object *lo, int ost)
 {
-	__u32 j;
-	LASSERT(inuse != NULL);
+	struct lod_obj_stripe_cb_data data = { { 0 } };
 
-	if (inuse->op_size == 0)
-		return 0;
+	data.locd_ost_index = ost;
+	data.locd_comp_skip_cb = lod_obj_is_ost_use_skip_cb;
+	data.locd_comp_cb = lod_obj_is_ost_use_cb;
 
-	LASSERT(inuse->op_count * sizeof(inuse->op_array[0]) <= inuse->op_size);
-	for (j = 0; j < inuse->op_count; j++) {
-		if (inuse->op_array[j] == ost)
-			return 1;
-	}
-	return 0;
+	(void)lod_obj_for_each_stripe(env, lo, NULL, &data);
+
+	return data.locd_ost_index == -1;
 }
 
-/**
- * Mark the given target as used for a composite layout
- *
- * \param[in] inuse	inuse ost index array
- * \param[in] idx	index in the array
- */
-static inline void lod_comp_ost_in_use(struct ost_pool *inuse, int ost)
+static inline void lod_avoid_update(struct lod_object *lo,
+				    struct lod_avoid_guide *lag)
+{
+	if (!lod_is_flr(lo))
+		return;
+
+	lag->lag_ost_avail--;
+}
+
+static inline bool lod_should_avoid_ost(struct lod_object *lo,
+					struct lod_avoid_guide *lag,
+					__u32 index)
 {
-	LASSERT(inuse != NULL);
-	if (inuse->op_size && !lod_comp_is_ost_used(inuse, ost)) {
-		LASSERT(inuse->op_count * sizeof(inuse->op_array[0]) <
-			inuse->op_size);
-		inuse->op_array[inuse->op_count] = ost;
-		inuse->op_count++;
+	struct lod_device *lod = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
+	struct lod_tgt_desc *ost = OST_TGT(lod, index);
+	struct lod_qos_oss *lqo = ost->ltd_qos.ltq_oss;
+	bool used = false;
+	int i;
+
+	if (!cfs_bitmap_check(lod->lod_ost_bitmap, index))
+		return true;
+
+	/**
+	 * we've tried our best, all available OSTs have been used in
+	 * overlapped components in the other mirror
+	 */
+	if (lag->lag_ost_avail == 0)
+		return false;
+
+	/* check OSS use */
+	for (i = 0; i < lag->lag_oaa_count; i++) {
+		if (lag->lag_oss_avoid_array[i] == lqo->lqo_id) {
+			used = true;
+			break;
+		}
 	}
+	/**
+	 * if the OSS which OST[index] resides has not been used, we'd like to
+	 * use it
+	 */
+	if (!used)
+		return false;
+
+	/* if the OSS has been used, check whether the OST has been used */
+	if (!cfs_bitmap_check(lag->lag_ost_avoid_bitmap, index))
+		used = false;
+	else
+		QOS_DEBUG("OST%d: has been used in overlapped component "
+			  "in other mirror\n", index);
+	return used;
 }
 
 static int lod_check_and_reserve_ost(const struct lu_env *env,
-				     struct lod_device *m,
+				     struct lod_object *lo,
 				     struct obd_statfs *sfs, __u32 ost_idx,
 				     __u32 speed, __u32 *s_idx,
 				     struct dt_object **stripe,
-				     struct thandle *th,
-				     struct ost_pool *inuse)
+				     __u32 *ost_indices,
+				     struct thandle *th)
 {
+	struct lod_device *lod = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
+	struct lod_avoid_guide *lag = &lod_env_info(env)->lti_avoid;
 	struct dt_object   *o;
 	__u32 stripe_idx = *s_idx;
 	int rc;
 
-	rc = lod_statfs_and_check(env, m, ost_idx, sfs);
+	rc = lod_statfs_and_check(env, lod, ost_idx, sfs);
 	if (rc) {
 		/* this OSP doesn't feel well */
 		goto out_return;
@@ -900,18 +972,27 @@ static int lod_check_and_reserve_ost(const struct lu_env *env,
 	 * try not allocate on OST which has been used by other
 	 * component
 	 */
-	if (speed == 0 && lod_comp_is_ost_used(inuse, ost_idx)) {
+	if (speed == 0 && lod_comp_is_ost_used(env, lo, ost_idx)) {
 		QOS_DEBUG("#%d: used by other component\n", ost_idx);
 		goto out_return;
 	}
 
+	/**
+	 * try not allocate OSTs used by conflicting component of other mirrors
+	 * for the first and second time.
+	 */
+	if (speed < 2 && lod_should_avoid_ost(lo, lag, ost_idx)) {
+		QOS_DEBUG("#%d: used by overlapped component of other mirror\n",
+			  ost_idx);
+		goto out_return;
+	}
 	/*
 	 * do not put >1 objects on a single OST
 	 */
 	if (lod_qos_is_ost_used(env, ost_idx, stripe_idx))
 		goto out_return;
 
-	o = lod_qos_declare_object_on(env, m, ost_idx, th);
+	o = lod_qos_declare_object_on(env, lod, 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));
@@ -922,9 +1003,10 @@ static int lod_check_and_reserve_ost(const struct lu_env *env,
 	/*
 	 * We've successfully declared (reserved) an object
 	 */
+	lod_avoid_update(lo, lag);
 	lod_qos_ost_in_use(env, stripe_idx, ost_idx);
-	lod_comp_ost_in_use(inuse, ost_idx);
 	stripe[stripe_idx] = o;
+	ost_indices[stripe_idx] = ost_idx;
 	OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_LOV_CREATE_RACE, 2);
 	stripe_idx++;
 	*s_idx = stripe_idx;
@@ -947,22 +1029,21 @@ out_return:
  * 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
- * \param[in] comp_idx	index of ldo_comp_entries
- * \param[in|out] inuse	array of inuse ost index
+ * \param[in] env		execution environment for this thread
+ * \param[in] lo		LOD object
+ * \param[out] stripe		striping created
+ * \param[out] ost_indices	ost indices of striping created
+ * \param[in] flags		allocation flags (0 or LOV_USES_DEFAULT_STRIPE)
+ * \param[in] th		transaction handle
+ * \param[in] comp_idx		index of ldo_comp_entries
  *
  * \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, int comp_idx,
-			struct ost_pool *inuse)
+			struct dt_object **stripe, __u32 *ost_indices,
+			int flags, struct thandle *th, int comp_idx)
 {
 	struct lod_layout_component *lod_comp;
 	struct lod_device *m = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
@@ -1048,8 +1129,9 @@ repeat_find:
 			continue;
 
 		spin_unlock(&lqr->lqr_alloc);
-		rc = lod_check_and_reserve_ost(env, m, sfs, ost_idx, speed,
-					       &stripe_idx, stripe, th, inuse);
+		rc = lod_check_and_reserve_ost(env, lo, sfs, ost_idx, speed,
+					       &stripe_idx, stripe, ost_indices,
+					       th);
 		spin_lock(&lqr->lqr_alloc);
 
 		if (rc != 0 && OST_TGT(m, ost_idx)->ltd_connecting)
@@ -1102,12 +1184,12 @@ out:
  * 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] th	transaction handle
- * \param[in] comp_idx	index of ldo_comp_entries
- * \param[in|out] inuse	array of inuse ost index
+ * \param[in] env		execution environment for this thread
+ * \param[in] lo		LOD object
+ * \param[out] stripe		striping created
+ * \param[out] ost_indices	ost indices of striping created
+ * \param[in] th		transaction handle
+ * \param[in] comp_idx		index of ldo_comp_entries
  *
  * \retval 0		on success
  * \retval -ENODEV	OST index does not exist on file system
@@ -1115,8 +1197,8 @@ out:
  * \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 thandle *th,
-			      int comp_idx, struct ost_pool *inuse)
+			      struct dt_object **stripe, __u32 *ost_indices,
+			      struct thandle *th, int comp_idx)
 {
 	struct lod_layout_component *lod_comp;
 	struct lod_device	*m = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
@@ -1137,6 +1219,10 @@ static int lod_alloc_ost_list(const struct lu_env *env, struct lod_object *lo,
 	if (rc < 0)
 		RETURN(rc);
 
+	if (lod_comp->llc_stripe_offset == LOV_OFFSET_DEFAULT)
+		lod_comp->llc_stripe_offset =
+				lod_comp->llc_ostlist.op_array[0];
+
 	for (i = 0; i < lod_comp->llc_stripe_count; i++) {
 		if (lod_comp->llc_ostlist.op_array[i] ==
 		    lod_comp->llc_stripe_offset) {
@@ -1185,8 +1271,8 @@ static int lod_alloc_ost_list(const struct lu_env *env, struct lod_object *lo,
 		 * We've successfully declared (reserved) an object
 		 */
 		lod_qos_ost_in_use(env, stripe_count, ost_idx);
-		lod_comp_ost_in_use(inuse, ost_idx);
 		stripe[stripe_count] = o;
+		ost_indices[stripe_count] = ost_idx;
 		stripe_count++;
 	}
 
@@ -1205,13 +1291,13 @@ static int lod_alloc_ost_list(const struct lu_env *env, struct lod_object *lo,
  * 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
- * \param[in] comp_idx	index of ldo_comp_entries
- * \param[in|out]inuse	array of inuse ost index
+ * \param[in] env		execution environment for this thread
+ * \param[in] lo		LOD object
+ * \param[out] stripe		striping created
+ * \param[out] ost_indices	ost indices of striping created
+ * \param[in] flags		not used
+ * \param[in] th		transaction handle
+ * \param[in] comp_idx		index of ldo_comp_entries
  *
  * \retval 0		on success
  * \retval -ENOSPC	if no OST objects are available at all
@@ -1220,9 +1306,8 @@ static int lod_alloc_ost_list(const struct lu_env *env, struct lod_object *lo,
  * \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, int comp_idx,
-			      struct ost_pool *inuse)
+			      struct dt_object **stripe, __u32 *ost_indices,
+			      int flags, struct thandle *th, int comp_idx)
 {
 	struct lod_layout_component *lod_comp;
 	struct lod_device *m = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
@@ -1293,7 +1378,7 @@ repeat_find:
 		 * try not allocate on the OST used by other component
 		 */
 		if (speed == 0 && i != 0 &&
-		    lod_comp_is_ost_used(inuse, ost_idx))
+		    lod_comp_is_ost_used(env, lo, ost_idx))
 			continue;
 
 		/* Drop slow OSCs if we can, but not for requested start idx.
@@ -1309,9 +1394,9 @@ repeat_find:
 		}
 
 		/*
-		 * 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
+		 * We expect number of precreated objects 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;
@@ -1327,8 +1412,8 @@ repeat_find:
 		 * We've successfully declared (reserved) an object
 		 */
 		lod_qos_ost_in_use(env, stripe_num, ost_idx);
-		lod_comp_ost_in_use(inuse, ost_idx);
 		stripe[stripe_num] = o;
+		ost_indices[stripe_num] = ost_idx;
 		stripe_num++;
 
 		/* We have enough stripes */
@@ -1411,13 +1496,13 @@ static inline int lod_qos_is_usable(struct lod_device *lod)
  * 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
- * \param[in] comp_idx	index of ldo_comp_entries
- * \param[in|out]inuse	array of inuse ost index
+ * \param[in] env		execution environment for this thread
+ * \param[in] lo		LOD object
+ * \param[out] stripe		striping created
+ * \param[out] ost_indices	ost indices of striping created
+ * \param[in] flags		0 or LOV_USES_DEFAULT_STRIPE
+ * \param[in] th		transaction handle
+ * \param[in] comp_idx		index of ldo_comp_entries
  *
  * \retval 0		on success
  * \retval -EAGAIN	not enough OSTs are found for specified stripe count
@@ -1425,20 +1510,20 @@ static inline int lod_qos_is_usable(struct lod_device *lod)
  * \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, int comp_idx,
-			 struct ost_pool *inuse)
+			 struct dt_object **stripe, __u32 *ost_indices,
+			 int flags, struct thandle *th, int comp_idx)
 {
 	struct lod_layout_component *lod_comp;
 	struct lod_device *lod = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
 	struct obd_statfs *sfs = &lod_env_info(env)->lti_osfs;
+	struct lod_avoid_guide *lag = &lod_env_info(env)->lti_avoid;
 	struct lod_tgt_desc *ost;
 	struct dt_object *o;
 	__u64 total_weight = 0;
 	struct pool_desc *pool = NULL;
 	struct ost_pool *osts;
 	unsigned int i;
-	__u32	nfound, good_osts, stripe_count, stripe_count_min;
+	__u32 nfound, good_osts, stripe_count, stripe_count_min;
 	int rc = 0;
 	ENTRY;
 
@@ -1559,7 +1644,7 @@ static int lod_alloc_qos(const struct lu_env *env, struct lod_object *lo,
 		for (i = 0; i < osts->op_count; i++) {
 			__u32 idx = osts->op_array[i];
 
-			if (!cfs_bitmap_check(lod->lod_ost_bitmap, idx))
+			if (lod_should_avoid_ost(lo, lag, idx))
 				continue;
 
 			ost = OST_TGT(lod, idx);
@@ -1581,7 +1666,7 @@ static int lod_alloc_qos(const struct lu_env *env, struct lod_object *lo,
 			 * do not put >1 objects on a single OST
 			 */
 			if (lod_qos_is_ost_used(env, idx, nfound) ||
-			    lod_comp_is_ost_used(inuse, idx))
+			    lod_comp_is_ost_used(env, lo, idx))
 				continue;
 
 			o = lod_qos_declare_object_on(env, lod, idx, th);
@@ -1591,10 +1676,12 @@ static int lod_alloc_qos(const struct lu_env *env, struct lod_object *lo,
 				continue;
 			}
 
+			lod_avoid_update(lo, lag);
 			lod_qos_ost_in_use(env, nfound, idx);
-			lod_comp_ost_in_use(inuse, idx);
-			stripe[nfound++] = o;
+			stripe[nfound] = o;
+			ost_indices[nfound] = idx;
 			lod_qos_used(lod, osts, idx, &total_weight);
+			nfound++;
 			rc = 0;
 			break;
 		}
@@ -1620,9 +1707,6 @@ static int lod_alloc_qos(const struct lu_env *env, struct lod_object *lo,
 			dt_object_put(env, stripe[i]);
 			stripe[i] = NULL;
 		}
-		LASSERTF(nfound <= inuse->op_count,
-			 "nfound:%d, op_count:%u\n", nfound, inuse->op_count);
-		inuse->op_count -= nfound;
 
 		/* makes sense to rebalance next time */
 		lod->lod_qos.lq_dirty = 1;
@@ -1726,29 +1810,38 @@ int lod_use_defined_striping(const struct lu_env *env,
 	struct lov_ost_data_v1 *objs;
 	__u32	magic;
 	__u16	comp_cnt;
+	__u16	mirror_cnt;
 	int	rc = 0, i;
 	ENTRY;
 
-	magic = le32_to_cpu(v1->lmm_magic) & ~LOV_MAGIC_DEF;
+	mutex_lock(&mo->ldo_layout_mutex);
+	lod_striping_free_nolock(env, mo);
+
+	magic = le32_to_cpu(v1->lmm_magic) & ~LOV_MAGIC_DEFINED;
 
 	if (magic != LOV_MAGIC_V1 && magic != LOV_MAGIC_V3 &&
 	    magic != LOV_MAGIC_COMP_V1)
-		RETURN(-EINVAL);
+		GOTO(unlock, rc = -EINVAL);
 
 	if (magic == LOV_MAGIC_COMP_V1) {
 		comp_v1 = buf->lb_buf;
 		comp_cnt = le16_to_cpu(comp_v1->lcm_entry_count);
 		if (comp_cnt == 0)
-			RETURN(-EINVAL);
+			GOTO(unlock, rc = -EINVAL);
+		mirror_cnt = le16_to_cpu(comp_v1->lcm_mirror_count) + 1;
+		mo->ldo_flr_state = le16_to_cpu(comp_v1->lcm_flags) &
+					LCM_FL_FLR_MASK;
 		mo->ldo_is_composite = 1;
 	} else {
 		mo->ldo_is_composite = 0;
 		comp_cnt = 1;
+		mirror_cnt = 0;
 	}
+	mo->ldo_layout_gen = le16_to_cpu(v1->lmm_layout_gen);
 
-	rc = lod_alloc_comp_entries(mo, comp_cnt);
+	rc = lod_alloc_comp_entries(mo, mirror_cnt, comp_cnt);
 	if (rc)
-		RETURN(rc);
+		GOTO(unlock, rc);
 
 	for (i = 0; i < comp_cnt; i++) {
 		struct lu_extent *ext;
@@ -1806,9 +1899,14 @@ int lod_use_defined_striping(const struct lu_env *env,
 				GOTO(out, rc);
 		}
 	}
+
+	rc = lod_fill_mirrors(mo);
+	GOTO(out, rc);
 out:
 	if (rc)
-		lod_object_free_striping(env, mo);
+		lod_striping_free_nolock(env, mo);
+unlock:
+	mutex_unlock(&mo->ldo_layout_mutex);
 
 	RETURN(rc);
 }
@@ -1839,26 +1937,34 @@ int lod_qos_parse_config(const struct lu_env *env, struct lod_object *lo,
 	struct lov_user_md_v1	*v1 = NULL;
 	struct lov_user_md_v3	*v3 = NULL;
 	struct lov_comp_md_v1	*comp_v1 = NULL;
+	char	def_pool[LOV_MAXPOOLNAME + 1];
 	__u32	magic;
 	__u16	comp_cnt;
+	__u16	mirror_cnt;
 	int	i, rc;
 	ENTRY;
 
 	if (buf == NULL || buf->lb_buf == NULL || buf->lb_len == 0)
 		RETURN(0);
 
-	rc = lod_verify_striping(d, buf, false, 0);
-	if (rc)
-		RETURN(-EINVAL);
+	memset(def_pool, 0, sizeof(def_pool));
+	if (lo->ldo_comp_entries != NULL)
+		lod_layout_get_pool(lo->ldo_comp_entries, lo->ldo_comp_cnt,
+				    def_pool, sizeof(def_pool));
 
+	/* free default striping info */
 	lod_free_comp_entries(lo);
 
+	rc = lod_verify_striping(d, lo, buf, false);
+	if (rc)
+		RETURN(-EINVAL);
+
 	v3 = buf->lb_buf;
 	v1 = buf->lb_buf;
 	comp_v1 = buf->lb_buf;
 	magic = v1->lmm_magic;
 
-	if (unlikely(le32_to_cpu(magic) & LOV_MAGIC_DEF)) {
+	if (unlikely(le32_to_cpu(magic) & LOV_MAGIC_DEFINED)) {
 		/* try to use as fully defined striping */
 		rc = lod_use_defined_striping(env, lo, buf);
 		RETURN(rc);
@@ -1903,16 +2009,22 @@ int lod_qos_parse_config(const struct lu_env *env, struct lod_object *lo,
 		comp_cnt = comp_v1->lcm_entry_count;
 		if (comp_cnt == 0)
 			RETURN(-EINVAL);
+		mirror_cnt =  comp_v1->lcm_mirror_count + 1;
+		if (mirror_cnt > 1)
+			lo->ldo_flr_state = LCM_FL_RDONLY;
 		lo->ldo_is_composite = 1;
 	} else {
 		comp_cnt = 1;
+		mirror_cnt = 0;
 		lo->ldo_is_composite = 0;
 	}
 
-	rc = lod_alloc_comp_entries(lo, comp_cnt);
+	rc = lod_alloc_comp_entries(lo, mirror_cnt, comp_cnt);
 	if (rc)
 		RETURN(rc);
 
+	LASSERT(lo->ldo_comp_entries);
+
 	for (i = 0; i < comp_cnt; i++) {
 		struct pool_desc	*pool;
 		struct lu_extent	*ext;
@@ -1925,38 +2037,28 @@ int lod_qos_parse_config(const struct lu_env *env, struct lod_object *lo,
 					comp_v1->lcm_entries[i].lcme_offset);
 			ext = &comp_v1->lcm_entries[i].lcme_extent;
 			lod_comp->llc_extent = *ext;
+			lod_comp->llc_flags =
+				comp_v1->lcm_entries[i].lcme_flags &
+					LCME_USER_FLAGS;
 		}
 
 		pool_name = NULL;
 		if (v1->lmm_magic == LOV_USER_MAGIC_V3 ||
 		    v1->lmm_magic == LOV_USER_MAGIC_SPECIFIC) {
-			int j;
-
 			v3 = (struct lov_user_md_v3 *)v1;
 			if (v3->lmm_pool_name[0] != '\0')
 				pool_name = v3->lmm_pool_name;
 
 			if (v3->lmm_magic == LOV_USER_MAGIC_SPECIFIC) {
-				if (v3->lmm_stripe_offset == LOV_OFFSET_DEFAULT)
-					v3->lmm_stripe_offset =
-						v3->lmm_objects[0].l_ost_idx;
-
-				/* copy ost list from lmm */
-				lod_comp->llc_ostlist.op_count =
-					v3->lmm_stripe_count;
-				lod_comp->llc_ostlist.op_size =
-					v3->lmm_stripe_count * sizeof(__u32);
-				OBD_ALLOC(lod_comp->llc_ostlist.op_array,
-					  lod_comp->llc_ostlist.op_size);
-				if (!lod_comp->llc_ostlist.op_array)
-					GOTO(free_comp, rc = -ENOMEM);
-
-				for (j = 0; j < v3->lmm_stripe_count; j++)
-					lod_comp->llc_ostlist.op_array[j] =
-						v3->lmm_objects[j].l_ost_idx;
+				rc = lod_comp_copy_ost_lists(lod_comp, v3);
+				if (rc)
+					GOTO(free_comp, rc);
 			}
 		}
 
+		if (pool_name == NULL && def_pool[0] != '\0')
+			pool_name = def_pool;
+
 		if (v1->lmm_pattern == 0)
 			v1->lmm_pattern = LOV_PATTERN_RAID0;
 		if (lov_pattern(v1->lmm_pattern) != LOV_PATTERN_RAID0 &&
@@ -2018,6 +2120,138 @@ free_comp:
 }
 
 /**
+ * prepare enough OST avoidance bitmap space
+ */
+int lod_prepare_avoidance(const struct lu_env *env, struct lod_object *lo)
+{
+	struct lod_device *lod = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
+	struct lod_tgt_descs *ltds = &lod->lod_ost_descs;
+	struct lod_avoid_guide *lag = &lod_env_info(env)->lti_avoid;
+	struct cfs_bitmap *bitmap = NULL;
+	__u32 *new_oss = NULL;
+
+	lag->lag_ost_avail = ltds->ltd_tgtnr;
+
+	/* reset OSS avoid guide array */
+	lag->lag_oaa_count = 0;
+	if (lag->lag_oss_avoid_array && lag->lag_oaa_size < ltds->ltd_tgtnr) {
+		OBD_FREE(lag->lag_oss_avoid_array,
+			 sizeof(__u32) * lag->lag_oaa_size);
+		lag->lag_oss_avoid_array = NULL;
+		lag->lag_oaa_size = 0;
+	}
+
+	/* init OST avoid guide bitmap */
+	if (lag->lag_ost_avoid_bitmap) {
+		if (ltds->ltd_tgtnr <= lag->lag_ost_avoid_bitmap->size) {
+			CFS_RESET_BITMAP(lag->lag_ost_avoid_bitmap);
+		} else {
+			CFS_FREE_BITMAP(lag->lag_ost_avoid_bitmap);
+			lag->lag_ost_avoid_bitmap = NULL;
+		}
+	}
+
+	if (!lag->lag_ost_avoid_bitmap) {
+		bitmap = CFS_ALLOCATE_BITMAP(ltds->ltd_tgtnr);
+		if (!bitmap)
+			return -ENOMEM;
+	}
+
+	if (!lag->lag_oss_avoid_array) {
+		/**
+		 * usually there are multiple OSTs in one OSS, but we don't
+		 * know the exact OSS number, so we choose a safe option,
+		 * using OST count to allocate the array to store the OSS
+		 * id.
+		 */
+		OBD_ALLOC(new_oss, sizeof(*new_oss) * ltds->ltd_tgtnr);
+		if (!new_oss) {
+			CFS_FREE_BITMAP(bitmap);
+			return -ENOMEM;
+		}
+	}
+
+	if (new_oss) {
+		lag->lag_oss_avoid_array = new_oss;
+		lag->lag_oaa_size = ltds->ltd_tgtnr;
+	}
+	if (bitmap)
+		lag->lag_ost_avoid_bitmap = bitmap;
+
+	return 0;
+}
+
+/**
+ * Collect information of used OSTs and OSSs in the overlapped components
+ * of other mirrors
+ */
+void lod_collect_avoidance(struct lod_object *lo, struct lod_avoid_guide *lag,
+			   int comp_idx)
+{
+	struct lod_device *lod = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
+	struct lod_layout_component *lod_comp = &lo->ldo_comp_entries[comp_idx];
+	struct cfs_bitmap *bitmap = lag->lag_ost_avoid_bitmap;
+	int i, j;
+
+	/* iterate mirrors */
+	for (i = 0; i < lo->ldo_mirror_count; i++) {
+		struct lod_layout_component *comp;
+
+		/**
+		 * skip mirror containing component[comp_idx], we only
+		 * collect OSTs info of conflicting component in other mirrors,
+		 * so that during read, if OSTs of a mirror's component are
+		 * not available, we still have other mirror with different
+		 * OSTs to read the data.
+		 */
+		comp = &lo->ldo_comp_entries[lo->ldo_mirrors[i].lme_start];
+		if (comp->llc_id == LCME_ID_INVAL ||
+		    mirror_id_of(comp->llc_id) ==
+						mirror_id_of(lod_comp->llc_id))
+			continue;
+
+		/* iterate components of a mirror */
+		lod_foreach_mirror_comp(comp, lo, i) {
+			/* skip non-overlapped or un-instantiated components */
+			if (!lu_extent_is_overlapped(&comp->llc_extent,
+						     &lod_comp->llc_extent) ||
+			    !lod_comp_inited(comp) || !comp->llc_stripe)
+				continue;
+
+			/**
+			 * collect used OSTs index and OSS info from a
+			 * component
+			 */
+			for (j = 0; j < comp->llc_stripe_count; j++) {
+				struct lod_tgt_desc *ost;
+				struct lod_qos_oss *lqo;
+				int k;
+
+				ost = OST_TGT(lod, comp->llc_ost_indices[j]);
+				lqo = ost->ltd_qos.ltq_oss;
+
+				if (cfs_bitmap_check(bitmap, ost->ltd_index))
+					continue;
+
+				cfs_bitmap_set(bitmap, ost->ltd_index);
+				lag->lag_ost_avail--;
+
+				for (k = 0; k < lag->lag_oaa_count; k++) {
+					if (lag->lag_oss_avoid_array[k] ==
+					    lqo->lqo_id)
+						break;
+				}
+				if (k == lag->lag_oaa_count) {
+					lag->lag_oss_avoid_array[k] =
+								lqo->lqo_id;
+					lag->lag_oaa_count++;
+				}
+			}
+		}
+	}
+}
+
+/**
  * Create a striping for an obejct.
  *
  * The function creates a new striping for the object. The function tries QoS
@@ -2031,21 +2265,22 @@ free_comp:
  * \param[in] attr	attributes OST objects will be declared with
  * \param[in] th	transaction handle
  * \param[in] comp_idx	index of ldo_comp_entries
- * \param[in|out] inuse	array of inuse ost index
  *
  * \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, struct thandle *th,
-			int comp_idx, struct ost_pool *inuse)
+			int comp_idx)
 {
 	struct lod_layout_component *lod_comp;
 	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 = 0;
+	struct lod_avoid_guide *lag = &lod_env_info(env)->lti_avoid;
+	struct dt_object **stripe = NULL;
+	__u32 *ost_indices = NULL;
 	ENTRY;
 
 	LASSERT(lo);
@@ -2078,6 +2313,9 @@ int lod_qos_prep_create(const struct lu_env *env, struct lod_object *lo,
 		OBD_ALLOC(stripe, sizeof(stripe[0]) * stripe_len);
 		if (stripe == NULL)
 			GOTO(out, rc = -ENOMEM);
+		OBD_ALLOC(ost_indices, sizeof(*ost_indices) * stripe_len);
+		if (!ost_indices)
+			GOTO(out, rc = -ENOMEM);
 
 		lod_getref(&d->lod_ost_descs);
 		/* XXX: support for non-0 files w/o objects */
@@ -2085,29 +2323,38 @@ int lod_qos_prep_create(const struct lu_env *env, struct lod_object *lo,
 				d->lod_desc.ld_tgt_count, stripe_len);
 
 		if (lod_comp->llc_ostlist.op_array) {
-			rc = lod_alloc_ost_list(env, lo, stripe, th, comp_idx,
-						inuse);
+			rc = lod_alloc_ost_list(env, lo, stripe, ost_indices,
+						th, comp_idx);
 		} else if (lod_comp->llc_stripe_offset == LOV_OFFSET_DEFAULT) {
-			rc = lod_alloc_qos(env, lo, stripe, flag, th,
-					   comp_idx, inuse);
+			/**
+			 * collect OSTs and OSSs used in other mirrors whose
+			 * components cross the ldo_comp_entries[comp_idx]
+			 */
+			rc = lod_prepare_avoidance(env, lo);
+			if (rc)
+				GOTO(put_ldts, rc);
+
+			lod_collect_avoidance(lo, lag, comp_idx);
+
+			rc = lod_alloc_qos(env, lo, stripe, ost_indices, flag,
+					   th, comp_idx);
 			if (rc == -EAGAIN)
-				rc = lod_alloc_rr(env, lo, stripe, flag, th,
-						  comp_idx, inuse);
+				rc = lod_alloc_rr(env, lo, stripe, ost_indices,
+						  flag, th, comp_idx);
 		} else {
-			rc = lod_alloc_specific(env, lo, stripe, flag, th,
-						comp_idx, inuse);
+			rc = lod_alloc_specific(env, lo, stripe, ost_indices,
+						flag, th, comp_idx);
 		}
+put_ldts:
 		lod_putref(d, &d->lod_ost_descs);
-
 		if (rc < 0) {
 			for (i = 0; i < stripe_len; i++)
 				if (stripe[i] != NULL)
 					dt_object_put(env, stripe[i]);
-
-			OBD_FREE(stripe, sizeof(stripe[0]) * stripe_len);
 			lod_comp->llc_stripe_count = 0;
 		} else {
 			lod_comp->llc_stripe = stripe;
+			lod_comp->llc_ost_indices = ost_indices;
 			lod_comp->llc_stripes_allocated = stripe_len;
 		}
 	} else {
@@ -2138,88 +2385,14 @@ int lod_qos_prep_create(const struct lu_env *env, struct lod_object *lo,
 	}
 
 out:
-	RETURN(rc);
-}
-
-int lod_obj_stripe_set_inuse_cb(const struct lu_env *env,
-				struct lod_object *lo,
-				struct dt_object *dt, struct thandle *th,
-				int stripe_idx,
-				struct lod_obj_stripe_cb_data *data)
-{
-	struct lod_thread_info	*info = lod_env_info(env);
-	struct lod_device	*d = lu2lod_dev(lod2lu_obj(lo)->lo_dev);
-	struct lu_fid	*fid = &info->lti_fid;
-	__u32	index;
-	int	rc, type = LU_SEQ_RANGE_OST;
-
-	*fid = *lu_object_fid(&dt->do_lu);
-	rc = lod_fld_lookup(env, d, fid, &index, &type);
 	if (rc < 0) {
-		CERROR("%s: fail to locate "DFID": rc = %d\n",
-		       lod2obd(d)->obd_name, PFID(fid), rc);
-		return rc;
+		if (stripe)
+			OBD_FREE(stripe, sizeof(stripe[0]) * stripe_len);
+		if (ost_indices)
+			OBD_FREE(ost_indices,
+				 sizeof(*ost_indices) * stripe_len);
 	}
-	lod_comp_ost_in_use(data->locd_inuse, index);
-	return 0;
-}
-
-/**
- * Resize per-thread ost list to hold OST target index list already used.
- *
- * \param[in,out] inuse		structure contains ost list array
- * \param[in] cnt		total stripe count of all components
- * \param[in] max		array's max size if @max > 0
- *
- * \retval 0		on success
- * \retval -ENOMEM	reallocation failed
- */
-static int lod_inuse_resize(struct ost_pool *inuse, __u16 cnt, __u16 max)
-{
-	__u32 *array;
-	__u32 new = cnt * sizeof(inuse->op_array[0]);
-
-	inuse->op_count = 0;
-
-	if (new <= inuse->op_size)
-		return 0;
-
-	if (max)
-		new = min_t(__u32, new, max);
-
-	OBD_ALLOC(array, new);
-	if (!array)
-		return -ENOMEM;
-
-	if (inuse->op_array)
-		OBD_FREE(inuse->op_array, inuse->op_size);
-
-	inuse->op_array = array;
-	inuse->op_size = new;
-
-	return 0;
-}
-
-int lod_prepare_inuse(const struct lu_env *env, struct lod_object *lo)
-{
-	struct lod_thread_info *info = lod_env_info(env);
-	struct lod_device *d = lu2lod_dev(lod2lu_obj(lo)->lo_dev);
-	struct ost_pool *inuse = &info->lti_inuse_osts;
-	struct lod_obj_stripe_cb_data data;
-	__u32 stripe_count = 0;
-	int i;
-	int rc;
-
-	for (i = 0; i < lo->ldo_comp_cnt; i++)
-		stripe_count += lod_comp_entry_stripe_count(lo,
-					&lo->ldo_comp_entries[i], false);
-	rc = lod_inuse_resize(inuse, stripe_count, d->lod_osd_max_easize);
-	if (rc)
-		return rc;
-
-	data.locd_inuse = inuse;
-	return lod_obj_for_each_stripe(env, lo, NULL,
-				       lod_obj_stripe_set_inuse_cb, &data);
+	RETURN(rc);
 }
 
 int lod_prepare_create(const struct lu_env *env, struct lod_object *lo,
@@ -2227,10 +2400,7 @@ int lod_prepare_create(const struct lu_env *env, struct lod_object *lo,
 		       struct thandle *th)
 
 {
-	struct lod_thread_info *info = lod_env_info(env);
 	struct lod_device *d = lu2lod_dev(lod2lu_obj(lo)->lo_dev);
-	struct ost_pool inuse_osts = { 0 };
-	struct ost_pool *inuse = &inuse_osts;
 	uint64_t size = 0;
 	int i;
 	int rc;
@@ -2259,14 +2429,6 @@ int lod_prepare_create(const struct lu_env *env, struct lod_object *lo,
 	if (attr->la_valid & LA_SIZE)
 		size = attr->la_size;
 
-	/* only prepare inuse if multiple components to be created */
-	if (size && lo->ldo_is_composite) {
-		rc = lod_prepare_inuse(env, lo);
-		if (rc)
-			RETURN(rc);
-		inuse = &info->lti_inuse_osts;
-	}
-
 	/**
 	 * prepare OST object creation for the component covering file's
 	 * size, the 1st component (including plain layout file) is always
@@ -2281,7 +2443,7 @@ int lod_prepare_create(const struct lu_env *env, struct lod_object *lo,
 		CDEBUG(D_QOS, "%lld [%lld, %lld)\n",
 		       size, extent->e_start, extent->e_end);
 		if (!lo->ldo_is_composite || size >= extent->e_start) {
-			rc = lod_qos_prep_create(env, lo, attr, th, i, inuse);
+			rc = lod_qos_prep_create(env, lo, attr, th, i);
 			if (rc)
 				break;
 		}