* 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/
#define DEBUG_SUBSYSTEM S_LOV
#include <asm/div64.h>
+#include <linux/random.h>
+
#include <libcfs/libcfs.h>
#include <uapi/linux/lustre/lustre_idl.h>
#include <lustre_swab.h>
#define TGT_BAVAIL(i) (OST_TGT(lod,i)->ltd_statfs.os_bavail * \
OST_TGT(lod,i)->ltd_statfs.os_bsize)
-/**
- * Add a new target to Quality of Service (QoS) target table.
- *
- * Add a new OST target to the structure representing an OSS. Resort the list
- * of known OSSs by the number of OSTs attached to each OSS. The OSS list is
- * protected internally and no external locking is required.
- *
- * \param[in] lod LOD device
- * \param[in] ost_desc OST description
- *
- * \retval 0 on success
- * \retval -ENOMEM on error
- */
-int qos_add_tgt(struct lod_device *lod, struct lod_tgt_desc *ost_desc)
+static inline int lod_statfs_check(struct lu_tgt_descs *ltd,
+ struct lu_tgt_desc *tgt)
{
- struct lod_qos_oss *oss = NULL, *temposs;
- struct obd_export *exp = ost_desc->ltd_exp;
- int rc = 0, found = 0;
- struct list_head *list;
- ENTRY;
+ struct obd_statfs *sfs = &tgt->ltd_statfs;
- down_write(&lod->lod_qos.lq_rw_sem);
- /*
- * a bit hacky approach to learn NID of corresponding connection
- * but there is no official API to access information like this
- * with OSD API.
- */
- list_for_each_entry(oss, &lod->lod_qos.lq_oss_list, lqo_oss_list) {
- if (obd_uuid_equals(&oss->lqo_uuid,
- &exp->exp_connection->c_remote_uuid)) {
- found++;
- break;
- }
- }
-
- if (!found) {
- OBD_ALLOC_PTR(oss);
- if (!oss)
- GOTO(out, rc = -ENOMEM);
- memcpy(&oss->lqo_uuid, &exp->exp_connection->c_remote_uuid,
- sizeof(oss->lqo_uuid));
- } else {
- /* Assume we have to move this one */
- list_del(&oss->lqo_oss_list);
- }
-
- oss->lqo_ost_count++;
- ost_desc->ltd_qos.ltq_oss = oss;
-
- CDEBUG(D_QOS, "add tgt %s to OSS %s (%d OSTs)\n",
- obd_uuid2str(&ost_desc->ltd_uuid), obd_uuid2str(&oss->lqo_uuid),
- oss->lqo_ost_count);
-
- /* Add sorted by # of OSTs. Find the first entry that we're
- bigger than... */
- list = &lod->lod_qos.lq_oss_list;
- list_for_each_entry(temposs, list, lqo_oss_list) {
- if (oss->lqo_ost_count > temposs->lqo_ost_count)
- break;
- }
- /* ...and add before it. If we're the first or smallest, temposs
- points to the list head, and we add to the end. */
- list_add_tail(&oss->lqo_oss_list, &temposs->lqo_oss_list);
+ if (((sfs->os_state & OS_STATFS_ENOSPC) ||
+ (!ltd->ltd_is_mdt && sfs->os_state & OS_STATFS_ENOINO &&
+ sfs->os_fprecreated == 0)))
+ return -ENOSPC;
- lod->lod_qos.lq_dirty = 1;
- lod->lod_qos.lq_rr.lqr_dirty = 1;
-
-out:
- up_write(&lod->lod_qos.lq_rw_sem);
- RETURN(rc);
-}
-
-/**
- * Remove OST target from QoS table.
- *
- * Removes given OST target from QoS table and releases related OSS structure
- * if no OSTs remain on the OSS.
- *
- * \param[in] lod LOD device
- * \param[in] ost_desc OST description
- *
- * \retval 0 on success
- * \retval -ENOENT if no OSS was found
- */
-int qos_del_tgt(struct lod_device *lod, struct lod_tgt_desc *ost_desc)
-{
- struct lod_qos_oss *oss;
- int rc = 0;
- ENTRY;
-
- down_write(&lod->lod_qos.lq_rw_sem);
- oss = ost_desc->ltd_qos.ltq_oss;
- if (!oss)
- GOTO(out, rc = -ENOENT);
+ /* If the OST is readonly then we can't allocate objects there */
+ if (sfs->os_state & OS_STATFS_READONLY)
+ return -EROFS;
- oss->lqo_ost_count--;
- if (oss->lqo_ost_count == 0) {
- CDEBUG(D_QOS, "removing OSS %s\n",
- obd_uuid2str(&oss->lqo_uuid));
- list_del(&oss->lqo_oss_list);
- ost_desc->ltd_qos.ltq_oss = NULL;
- OBD_FREE_PTR(oss);
- }
+ /* object precreation is skipped on the OST with max_create_count=0 */
+ if (!ltd->ltd_is_mdt && sfs->os_state & OS_STATFS_NOPRECREATE)
+ return -ENOBUFS;
- lod->lod_qos.lq_dirty = 1;
- lod->lod_qos.lq_rr.lqr_dirty = 1;
-out:
- up_write(&lod->lod_qos.lq_rw_sem);
- RETURN(rc);
+ return 0;
}
/**
- * Check whether the target is available for new OST objects.
+ * Check whether the target is available for new objects.
*
* Request statfs data from the given target and verify it's active and not
- * read-only. If so, then it can be used to place new OST objects. This
+ * read-only. If so, then it can be used to place new objects. This
* function also maintains the number of active/inactive targets and sets
* dirty flags if those numbers change so others can run re-balance procedures.
* No external locking is required.
*
* \param[in] env execution environment for this thread
* \param[in] d LOD device
- * \param[in] index index of OST target to check
- * \param[out] sfs buffer for statfs data
+ * \param[in] ltd target table
+ * \param[in] tgt target
*
* \retval 0 if the target is good
* \retval negative negated errno on error
-
*/
static int lod_statfs_and_check(const struct lu_env *env, struct lod_device *d,
- int index, struct obd_statfs *sfs)
+ struct lu_tgt_descs *ltd,
+ struct lu_tgt_desc *tgt, __u64 reserve)
{
- struct lod_tgt_desc *ost;
- int rc;
+ struct obd_statfs_info info = { 0 };
+ struct lov_desc *desc = <d->ltd_lov_desc;
+ int rc;
ENTRY;
LASSERT(d);
- ost = OST_TGT(d,index);
- LASSERT(ost);
-
- rc = dt_statfs(env, ost->ltd_ost, sfs);
-
- if (rc == 0 && ((sfs->os_state & OS_STATE_ENOSPC) ||
- (sfs->os_state & OS_STATE_ENOINO && sfs->os_fprecreated == 0)))
- RETURN(-ENOSPC);
+ LASSERT(tgt);
+ info.os_enable_pre = 1;
+ rc = dt_statfs_info(env, tgt->ltd_tgt, &tgt->ltd_statfs, &info);
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;
+ if (!rc) {
+ rc = lod_statfs_check(ltd, tgt);
+ if (rc == -ENOSPC)
+ return rc;
+ }
+
+ if (reserve &&
+ (reserve + (info.os_reserved_mb_low << 20) >
+ tgt->ltd_statfs.os_bavail * tgt->ltd_statfs.os_bsize))
+ return -ENOSPC;
/* check whether device has changed state (active, inactive) */
- if (rc != 0 && ost->ltd_active) {
+ if (rc != 0 && tgt->ltd_active) {
/* turned inactive? */
spin_lock(&d->lod_lock);
- if (ost->ltd_active) {
- ost->ltd_active = 0;
+ if (tgt->ltd_active) {
+ tgt->ltd_active = 0;
if (rc == -ENOTCONN)
- ost->ltd_connecting = 1;
+ tgt->ltd_connecting = 1;
- LASSERT(d->lod_desc.ld_active_tgt_count > 0);
- d->lod_desc.ld_active_tgt_count--;
- d->lod_qos.lq_dirty = 1;
- d->lod_qos.lq_rr.lqr_dirty = 1;
+ LASSERT(desc->ld_active_tgt_count > 0);
+ desc->ld_active_tgt_count--;
+ ltd->ltd_qos.lq_dirty = 1;
+ ltd->ltd_qos.lq_rr.lqr_dirty = 1;
CDEBUG(D_CONFIG, "%s: turns inactive\n",
- ost->ltd_exp->exp_obd->obd_name);
+ tgt->ltd_exp->exp_obd->obd_name);
}
spin_unlock(&d->lod_lock);
- } else if (rc == 0 && ost->ltd_active == 0) {
+ } else if (rc == 0 && tgt->ltd_active == 0) {
/* turned active? */
- LASSERTF(d->lod_desc.ld_active_tgt_count < d->lod_ostnr,
- "active tgt count %d, ost nr %d\n",
- d->lod_desc.ld_active_tgt_count, d->lod_ostnr);
+ LASSERTF(desc->ld_active_tgt_count < desc->ld_tgt_count,
+ "active tgt count %d, tgt nr %d\n",
+ desc->ld_active_tgt_count, desc->ld_tgt_count);
spin_lock(&d->lod_lock);
- if (ost->ltd_active == 0) {
- ost->ltd_active = 1;
- ost->ltd_connecting = 0;
- d->lod_desc.ld_active_tgt_count++;
- d->lod_qos.lq_dirty = 1;
- d->lod_qos.lq_rr.lqr_dirty = 1;
+ if (tgt->ltd_active == 0) {
+ tgt->ltd_active = 1;
+ tgt->ltd_connecting = 0;
+ desc->ld_active_tgt_count++;
+ ltd->ltd_qos.lq_dirty = 1;
+ ltd->ltd_qos.lq_rr.lqr_dirty = 1;
CDEBUG(D_CONFIG, "%s: turns active\n",
- ost->ltd_exp->exp_obd->obd_name);
+ tgt->ltd_exp->exp_obd->obd_name);
}
spin_unlock(&d->lod_lock);
}
+ if (rc == -ENOTCONN) {
+ /* In case that the ENOTCONN for inactive OST state is
+ * mistreated as MDT disconnection state by the client,
+ * this error should be changed to someone else.
+ */
+ rc = -EREMOTEIO;
+ }
RETURN(rc);
}
+static int lod_is_tgt_usable(struct lu_tgt_descs *ltd, struct lu_tgt_desc *tgt)
+{
+ int rc;
+
+ rc = lod_statfs_check(ltd, tgt);
+ if (rc)
+ return rc;
+
+ if (!tgt->ltd_active)
+ return -ENOTCONN;
+
+ return 0;
+}
+
/**
* Maintain per-target statfs data.
*
*
* \param[in] env execution environment for this thread
* \param[in] lod LOD device
+ * \param[in] ltd tgt table
*/
-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 lu_tgt_descs *ltd)
{
struct obd_device *obd = lod2obd(lod);
- struct ost_pool *osts = &(lod->lod_pool_info);
- unsigned int i;
- int idx;
- __u64 max_age, avail;
+ struct lu_tgt_desc *tgt;
+ time64_t max_age;
+ u64 avail;
ENTRY;
- max_age = cfs_time_shift_64(-2 * lod->lod_desc.ld_qos_maxage);
+ max_age = ktime_get_seconds() - 2 * ltd->ltd_lov_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))
+ down_write(<d->ltd_qos.lq_rw_sem);
+
+ if (obd->obd_osfs_age > max_age)
goto out;
- for (i = 0; i < osts->op_count; i++) {
- idx = osts->op_array[i];
- avail = OST_TGT(lod,idx)->ltd_statfs.os_bavail;
- if (lod_statfs_and_check(env, lod, idx,
- &OST_TGT(lod, idx)->ltd_statfs))
+ ltd_foreach_tgt(ltd, tgt) {
+ avail = tgt->ltd_statfs.os_bavail;
+ if (lod_statfs_and_check(env, lod, ltd, tgt, 0))
continue;
- if (OST_TGT(lod,idx)->ltd_statfs.os_bavail != avail)
+
+ if (tgt->ltd_statfs.os_bavail != avail)
/* recalculate weigths */
- lod->lod_qos.lq_dirty = 1;
+ ltd->ltd_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);
+ up_write(<d->ltd_qos.lq_rw_sem);
EXIT;
}
-/**
- * Calculate per-OST and per-OSS penalties
- *
- * Re-calculate penalties when the configuration changes, active targets
- * change and after statfs refresh (all these are reflected by lq_dirty flag).
- * On every OST and OSS: decay the penalty by half for every 8x the update
- * interval that the device has been idle. That gives lots of time for the
- * statfs information to be updated (which the penalty is only a proxy for),
- * and avoids penalizing OSS/OSTs under light load.
- * See lod_qos_calc_weight() for how penalties are factored into the weight.
- *
- * \param[in] lod LOD device
- *
- * \retval 0 on success
- * \retval -EAGAIN the number of OSTs isn't enough
- */
-static int lod_qos_calc_ppo(struct lod_device *lod)
-{
- struct lod_qos_oss *oss;
- __u64 ba_max, ba_min, temp;
- __u32 num_active;
- unsigned int i;
- int rc, prio_wide;
- time_t now, age;
- ENTRY;
-
- if (!lod->lod_qos.lq_dirty)
- GOTO(out, rc = 0);
-
- num_active = lod->lod_desc.ld_active_tgt_count - 1;
- if (num_active < 1)
- GOTO(out, rc = -EAGAIN);
-
- /* find bavail on each OSS */
- list_for_each_entry(oss, &lod->lod_qos.lq_oss_list, lqo_oss_list)
- oss->lqo_bavail = 0;
- lod->lod_qos.lq_active_oss_count = 0;
-
- /*
- * How badly user wants to select OSTs "widely" (not recently chosen
- * and not on recent OSS's). As opposed to "freely" (free space
- * avail.) 0-256
- */
- prio_wide = 256 - lod->lod_qos.lq_prio_free;
-
- ba_min = (__u64)(-1);
- ba_max = 0;
- now = cfs_time_current_sec();
- /* Calculate OST penalty per object
- * (lod ref taken in lod_qos_prep_create()) */
- cfs_foreach_bit(lod->lod_ost_bitmap, i) {
- LASSERT(OST_TGT(lod,i));
- temp = TGT_BAVAIL(i);
- if (!temp)
- continue;
- ba_min = min(temp, ba_min);
- ba_max = max(temp, ba_max);
-
- /* Count the number of usable OSS's */
- if (OST_TGT(lod,i)->ltd_qos.ltq_oss->lqo_bavail == 0)
- lod->lod_qos.lq_active_oss_count++;
- OST_TGT(lod,i)->ltd_qos.ltq_oss->lqo_bavail += temp;
-
- /* per-OST penalty is prio * TGT_bavail / (num_ost - 1) / 2 */
- temp >>= 1;
- do_div(temp, num_active);
- OST_TGT(lod,i)->ltd_qos.ltq_penalty_per_obj =
- (temp * prio_wide) >> 8;
-
- age = (now - OST_TGT(lod,i)->ltd_qos.ltq_used) >> 3;
- if (lod->lod_qos.lq_reset ||
- age > 32 * lod->lod_desc.ld_qos_maxage)
- OST_TGT(lod,i)->ltd_qos.ltq_penalty = 0;
- else if (age > lod->lod_desc.ld_qos_maxage)
- /* Decay OST penalty. */
- OST_TGT(lod,i)->ltd_qos.ltq_penalty >>=
- (age / lod->lod_desc.ld_qos_maxage);
- }
-
- num_active = lod->lod_qos.lq_active_oss_count - 1;
- if (num_active < 1) {
- /* If there's only 1 OSS, we can't penalize it, so instead
- we have to double the OST penalty */
- num_active = 1;
- cfs_foreach_bit(lod->lod_ost_bitmap, i)
- OST_TGT(lod,i)->ltd_qos.ltq_penalty_per_obj <<= 1;
- }
-
- /* Per-OSS penalty is prio * oss_avail / oss_osts / (num_oss - 1) / 2 */
- list_for_each_entry(oss, &lod->lod_qos.lq_oss_list, lqo_oss_list) {
- temp = oss->lqo_bavail >> 1;
- do_div(temp, oss->lqo_ost_count * num_active);
- oss->lqo_penalty_per_obj = (temp * prio_wide) >> 8;
-
- age = (now - oss->lqo_used) >> 3;
- if (lod->lod_qos.lq_reset ||
- age > 32 * lod->lod_desc.ld_qos_maxage)
- oss->lqo_penalty = 0;
- else if (age > lod->lod_desc.ld_qos_maxage)
- /* Decay OSS penalty. */
- oss->lqo_penalty >>= age / lod->lod_desc.ld_qos_maxage;
- }
-
- lod->lod_qos.lq_dirty = 0;
- lod->lod_qos.lq_reset = 0;
-
- /* If each ost has almost same free space,
- * do rr allocation for better creation performance */
- lod->lod_qos.lq_same_space = 0;
- if ((ba_max * (256 - lod->lod_qos.lq_threshold_rr)) >> 8 < ba_min) {
- lod->lod_qos.lq_same_space = 1;
- /* Reset weights for the next time we enter qos mode */
- lod->lod_qos.lq_reset = 1;
- }
- rc = 0;
-
-out:
-#ifndef FORCE_QOS
- if (!rc && lod->lod_qos.lq_same_space)
- RETURN(-EAGAIN);
-#endif
- RETURN(rc);
-}
-
-/**
- * Calculate weight for a given OST target.
- *
- * The final OST weight is the number of bytes available minus the OST and
- * OSS penalties. See lod_qos_calc_ppo() for how penalties are calculated.
- *
- * \param[in] lod LOD device, where OST targets are listed
- * \param[in] i OST target index
- *
- * \retval 0
- */
-static int lod_qos_calc_weight(struct lod_device *lod, int i)
-{
- __u64 temp, temp2;
-
- temp = TGT_BAVAIL(i);
- temp2 = OST_TGT(lod,i)->ltd_qos.ltq_penalty +
- OST_TGT(lod,i)->ltd_qos.ltq_oss->lqo_penalty;
- if (temp < temp2)
- OST_TGT(lod,i)->ltd_qos.ltq_weight = 0;
- else
- OST_TGT(lod,i)->ltd_qos.ltq_weight = temp - temp2;
- return 0;
-}
-
-/**
- * Re-calculate weights.
- *
- * The function is called when some OST target was used for a new object. In
- * this case we should re-calculate all the weights to keep new allocations
- * balanced well.
- *
- * \param[in] lod LOD device
- * \param[in] osts OST pool where a new object was placed
- * \param[in] index OST target where a new object was placed
- * \param[out] total_wt new total weight for the pool
- *
- * \retval 0
- */
-static int lod_qos_used(struct lod_device *lod, struct ost_pool *osts,
- __u32 index, __u64 *total_wt)
-{
- struct lod_tgt_desc *ost;
- struct lod_qos_oss *oss;
- unsigned int j;
- ENTRY;
-
- ost = OST_TGT(lod,index);
- LASSERT(ost);
-
- /* Don't allocate on this devuce anymore, until the next alloc_qos */
- ost->ltd_qos.ltq_usable = 0;
-
- oss = ost->ltd_qos.ltq_oss;
-
- /* Decay old penalty by half (we're adding max penalty, and don't
- want it to run away.) */
- ost->ltd_qos.ltq_penalty >>= 1;
- oss->lqo_penalty >>= 1;
-
- /* mark the OSS and OST as recently used */
- ost->ltd_qos.ltq_used = oss->lqo_used = cfs_time_current_sec();
-
- /* Set max penalties for this OST and OSS */
- ost->ltd_qos.ltq_penalty +=
- ost->ltd_qos.ltq_penalty_per_obj * lod->lod_ostnr;
- oss->lqo_penalty += oss->lqo_penalty_per_obj *
- lod->lod_qos.lq_active_oss_count;
-
- /* Decrease all OSS penalties */
- list_for_each_entry(oss, &lod->lod_qos.lq_oss_list, lqo_oss_list) {
- if (oss->lqo_penalty < oss->lqo_penalty_per_obj)
- oss->lqo_penalty = 0;
- else
- oss->lqo_penalty -= oss->lqo_penalty_per_obj;
- }
-
- *total_wt = 0;
- /* Decrease all OST penalties */
- for (j = 0; j < osts->op_count; j++) {
- int i;
-
- i = osts->op_array[j];
- if (!cfs_bitmap_check(lod->lod_ost_bitmap, i))
- continue;
-
- ost = OST_TGT(lod,i);
- LASSERT(ost);
-
- if (ost->ltd_qos.ltq_penalty <
- ost->ltd_qos.ltq_penalty_per_obj)
- ost->ltd_qos.ltq_penalty = 0;
- else
- ost->ltd_qos.ltq_penalty -=
- ost->ltd_qos.ltq_penalty_per_obj;
-
- lod_qos_calc_weight(lod, i);
-
- /* Recalc the total weight of usable osts */
- if (ost->ltd_qos.ltq_usable)
- *total_wt += ost->ltd_qos.ltq_weight;
-
- QOS_DEBUG("recalc tgt %d usable=%d avail=%llu"
- " ostppo=%llu ostp=%llu ossppo=%llu"
- " ossp=%llu wt=%llu\n",
- i, ost->ltd_qos.ltq_usable, TGT_BAVAIL(i) >> 10,
- ost->ltd_qos.ltq_penalty_per_obj >> 10,
- ost->ltd_qos.ltq_penalty >> 10,
- ost->ltd_qos.ltq_oss->lqo_penalty_per_obj >> 10,
- ost->ltd_qos.ltq_oss->lqo_penalty >> 10,
- ost->ltd_qos.ltq_weight >> 10);
- }
-
- RETURN(0);
-}
-
-void lod_qos_rr_init(struct lod_qos_rr *lqr)
-{
- spin_lock_init(&lqr->lqr_alloc);
- lqr->lqr_dirty = 1;
-}
-
-
#define LOV_QOS_EMPTY ((__u32)-1)
/**
* a new target or activation/deactivation).
*
* \param[in] lod LOD device
- * \param[in] src_pool OST pool
+ * \param[in] ltd tgt table
+ * \param[in] src_pool tgt pool
* \param[in] lqr round-robin list
*
* \retval 0 on success
* \retval -ENOMEM fails to allocate the array
*/
-static int lod_qos_calc_rr(struct lod_device *lod, struct ost_pool *src_pool,
- struct lod_qos_rr *lqr)
+static int lod_qos_calc_rr(struct lod_device *lod, struct lu_tgt_descs *ltd,
+ const struct lu_tgt_pool *src_pool,
+ struct lu_qos_rr *lqr)
{
- struct lod_qos_oss *oss;
- struct lod_tgt_desc *ost;
+ struct lu_svr_qos *svr;
+ struct lu_tgt_desc *tgt;
unsigned placed, real_count;
unsigned int i;
int rc;
}
/* Do actual allocation. */
- down_write(&lod->lod_qos.lq_rw_sem);
+ down_write(<d->ltd_qos.lq_rw_sem);
/*
* Check again. While we were sleeping on @lq_rw_sem something could
*/
if (!lqr->lqr_dirty) {
LASSERT(lqr->lqr_pool.op_size);
- up_write(&lod->lod_qos.lq_rw_sem);
+ up_write(<d->ltd_qos.lq_rw_sem);
RETURN(0);
}
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);
+ rc = tgt_pool_extend(&lqr->lqr_pool, real_count);
if (rc) {
- up_write(&lod->lod_qos.lq_rw_sem);
+ up_write(<d->ltd_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. */
+ /* Place all the tgts from 1 svr at the same time. */
placed = 0;
- list_for_each_entry(oss, &lod->lod_qos.lq_oss_list, lqo_oss_list) {
+ list_for_each_entry(svr, <d->ltd_qos.lq_svr_list, lsq_svr_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]))
+ if (!test_bit(src_pool->op_array[i],
+ ltd->ltd_tgt_bitmap))
continue;
- ost = OST_TGT(lod,src_pool->op_array[i]);
- LASSERT(ost && ost->ltd_ost);
- if (ost->ltd_qos.ltq_oss != oss)
+ tgt = LTD_TGT(ltd, src_pool->op_array[i]);
+ LASSERT(tgt && tgt->ltd_tgt);
+ if (tgt->ltd_qos.ltq_svr != svr)
continue;
- /* Evenly space these OSTs across arrayspace */
- next = j * lqr->lqr_pool.op_count / oss->lqo_ost_count;
+ /* Evenly space these tgts across arrayspace */
+ next = j * lqr->lqr_pool.op_count / svr->lsq_tgt_count;
while (lqr->lqr_pool.op_array[next] != LOV_QOS_EMPTY)
next = (next + 1) % lqr->lqr_pool.op_count;
}
lqr->lqr_dirty = 0;
- up_write(&lod->lod_qos.lq_rw_sem);
+ up_write(<d->ltd_qos.lq_rw_sem);
if (placed != real_count) {
/* This should never happen */
- LCONSOLE_ERROR_MSG(0x14e, "Failed to place all OSTs in the "
+ LCONSOLE_ERROR_MSG(0x14e, "Failed to place all tgts 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,
+ LCONSOLE(D_WARNING, "rr #%d tgt idx=%d\n", i,
lqr->lqr_pool.op_array[i]);
}
lqr->lqr_dirty = 1;
ENTRY;
LASSERT(d);
- LASSERT(ost_idx < d->lod_osts_size);
+ LASSERT(ost_idx < d->lod_ost_descs.ltd_tgts_size);
ost = OST_TGT(d,ost_idx);
LASSERT(ost);
- LASSERT(ost->ltd_ost);
+ LASSERT(ost->ltd_tgt);
- nd = &ost->ltd_ost->dd_lu_dev;
+ nd = &ost->ltd_tgt->dd_lu_dev;
/*
* allocate anonymous object with zero fid, real fid
*/
o = lu_object_anon(env, nd, NULL);
if (IS_ERR(o))
- GOTO(out, dt = ERR_PTR(PTR_ERR(o)));
+ GOTO(out, dt = ERR_CAST(o));
n = lu_object_locate(o->lo_header, nd->ld_type);
if (unlikely(n == NULL)) {
#define LOV_CREATE_RESEED_MIN 2000
/**
- * Initialize temporary OST-in-use array.
+ * Initialize temporary tgt-in-use array.
*
* Allocate or extend the array used to mark targets already assigned to a new
* striping so they are not used more than once.
* \retval 0 on success
* \retval -ENOMEM on error
*/
-static inline int lod_qos_ost_in_use_clear(const struct lu_env *env,
+static inline int lod_qos_tgt_in_use_clear(const struct lu_env *env,
__u32 stripes)
{
struct lod_thread_info *info = lod_env_info(env);
if (info->lti_ea_store_size < sizeof(int) * stripes)
lod_ea_store_resize(info, stripes * sizeof(int));
if (info->lti_ea_store_size < sizeof(int) * stripes) {
- CERROR("can't allocate memory for ost-in-use array\n");
+ CERROR("can't allocate memory for tgt-in-use array\n");
return -ENOMEM;
}
memset(info->lti_ea_store, -1, sizeof(int) * stripes);
* Remember a target in the array of used targets.
*
* Mark the given target as used for a new striping being created. The status
- * of an OST in a striping can be checked with lod_qos_is_ost_used().
+ * of an tgt in a striping can be checked with lod_qos_is_tgt_used().
*
* \param[in] env execution environment for this thread
* \param[in] idx index in the array
- * \param[in] ost OST target index to mark as used
+ * \param[in] tgt_idx target index to mark as used
*/
-static inline void lod_qos_ost_in_use(const struct lu_env *env,
- int idx, int ost)
+static inline void lod_qos_tgt_in_use(const struct lu_env *env,
+ int idx, int tgt_idx)
{
struct lod_thread_info *info = lod_env_info(env);
- int *osts = info->lti_ea_store;
+ int *tgts = info->lti_ea_store;
LASSERT(info->lti_ea_store_size >= idx * sizeof(int));
- osts[idx] = ost;
+ tgts[idx] = tgt_idx;
}
/**
- * Check is OST used in a striping.
+ * Check is tgt used in a striping.
*
- * Checks whether OST with the given index is marked as used in the temporary
- * array (see lod_qos_ost_in_use()).
+ * Checks whether tgt with the given index is marked as used in the temporary
+ * array (see lod_qos_tgt_in_use()).
*
* \param[in] env execution environment for this thread
- * \param[in] ost OST target index to check
+ * \param[in] tgt_idx target index to check
* \param[in] stripes the number of items used in the array already
*
* \retval 0 not used
* \retval 1 used
*/
-static int lod_qos_is_ost_used(const struct lu_env *env, int ost, __u32 stripes)
+static int lod_qos_is_tgt_used(const struct lu_env *env, int tgt_idx,
+ __u32 stripes)
{
struct lod_thread_info *info = lod_env_info(env);
- int *osts = info->lti_ea_store;
+ int *tgts = info->lti_ea_store;
__u32 j;
for (j = 0; j < stripes; j++) {
- if (osts[j] == ost)
+ if (tgts[j] == tgt_idx)
return 1;
}
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)
{
- 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++;
+ 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)
+{
+ struct lod_device *lod = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
+ struct lod_tgt_desc *ost = OST_TGT(lod, index);
+ struct lu_svr_qos *lsq = ost->ltd_qos.ltq_svr;
+ bool used = false;
+ int i;
+
+ if (!test_bit(index, lod->lod_ost_bitmap)) {
+ QOS_DEBUG("OST%d: been used in conflicting mirror component\n",
+ 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] == lsq->lsq_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: been used in conflicting mirror component\n",
+ index);
+ return used;
}
static int lod_check_and_reserve_ost(const struct lu_env *env,
- struct lod_device *m,
- struct obd_statfs *sfs, __u32 ost_idx,
- __u32 speed, __u32 *s_idx,
+ struct lod_object *lo,
+ struct lod_layout_component *lod_comp,
+ __u32 ost_idx, __u32 speed, __u32 *s_idx,
struct dt_object **stripe,
+ __u32 *ost_indices,
struct thandle *th,
- struct ost_pool *inuse)
+ bool *overstriped,
+ __u64 reserve)
{
+ 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 lu_tgt_desc *ost = OST_TGT(lod, ost_idx);
struct dt_object *o;
__u32 stripe_idx = *s_idx;
int rc;
- rc = lod_statfs_and_check(env, m, ost_idx, sfs);
- if (rc) {
- /* this OSP doesn't feel well */
- goto out_return;
- }
+ ENTRY;
+
+ rc = lod_statfs_and_check(env, lod, &lod->lod_ost_descs, ost, reserve);
+ if (rc)
+ RETURN(rc);
/*
* We expect number of precreated objects in f_ffree at
* the first iteration, skip OSPs with no objects ready
*/
- if (sfs->os_fprecreated == 0 && speed == 0) {
+ if (ost->ltd_statfs.os_fprecreated == 0 && speed == 0) {
QOS_DEBUG("#%d: precreation is empty\n", ost_idx);
- goto out_return;
+ RETURN(rc);
}
/*
* try to use another OSP if this one is degraded
*/
- if (sfs->os_state & OS_STATE_DEGRADED && speed < 2) {
+ if (ost->ltd_statfs.os_state & OS_STATFS_DEGRADED && speed < 2) {
QOS_DEBUG("#%d: degraded\n", ost_idx);
- goto out_return;
+ RETURN(rc);
}
/*
* try not allocate on OST which has been used by other
* component
*/
- if (speed == 0 && lod_comp_is_ost_used(inuse, ost_idx)) {
- QOS_DEBUG("#%d: used by other component\n", ost_idx);
- goto out_return;
+ if (speed == 0 && lod_comp_is_ost_used(env, lo, ost_idx)) {
+ QOS_DEBUG("iter %d: OST%d used by other component\n",
+ speed, ost_idx);
+ RETURN(rc);
}
- /*
- * do not put >1 objects on a single OST
+ /**
+ * try not allocate OSTs used by conflicting component of other mirrors
+ * for the first and second time.
*/
- if (lod_qos_is_ost_used(env, ost_idx, stripe_idx))
- goto out_return;
+ if (speed < 2 && lod_should_avoid_ost(lo, lag, ost_idx)) {
+ QOS_DEBUG("iter %d: OST%d used by conflicting mirror component\n",
+ speed, ost_idx);
+ RETURN(rc);
+ }
+
+ /* do not put >1 objects on a single OST, except for overstriping */
+ if (lod_qos_is_tgt_used(env, ost_idx, stripe_idx)) {
+ if (lod_comp->llc_pattern & LOV_PATTERN_OVERSTRIPING)
+ *overstriped = true;
+ else
+ RETURN(rc);
+ }
- 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));
rc = PTR_ERR(o);
- goto out_return;
+ RETURN(rc);
}
/*
* We've successfully declared (reserved) an object
*/
- lod_qos_ost_in_use(env, stripe_idx, ost_idx);
- lod_comp_ost_in_use(inuse, ost_idx);
+ lod_avoid_update(lo, lag);
+ lod_qos_tgt_in_use(env, stripe_idx, 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;
-out_return:
- return rc;
+ RETURN(rc);
}
/**
* 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)
+static int lod_ost_alloc_rr(const struct lu_env *env, struct lod_object *lo,
+ struct dt_object **stripe, __u32 *ost_indices,
+ int flags, struct thandle *th, int comp_idx,
+ __u64 reserve)
{
struct lod_layout_component *lod_comp;
struct lod_device *m = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
- struct obd_statfs *sfs = &lod_env_info(env)->lti_osfs;
struct pool_desc *pool = NULL;
- struct ost_pool *osts;
- struct lod_qos_rr *lqr;
- unsigned int i, array_idx;
+ struct lu_tgt_pool *osts;
+ struct lu_qos_rr *lqr;
+ unsigned int i, array_idx;
__u32 ost_start_idx_temp;
__u32 stripe_idx = 0;
__u32 stripe_count, stripe_count_min, ost_idx;
int rc, speed = 0, ost_connecting = 0;
+ int stripes_per_ost = 1;
+ bool overstriped = false;
ENTRY;
LASSERT(lo->ldo_comp_cnt > comp_idx && lo->ldo_comp_entries != NULL);
osts = &(pool->pool_obds);
lqr = &(pool->pool_rr);
} else {
- osts = &(m->lod_pool_info);
- lqr = &(m->lod_qos.lq_rr);
+ osts = &m->lod_ost_descs.ltd_tgt_pool;
+ lqr = &(m->lod_ost_descs.ltd_qos.lq_rr);
}
- rc = lod_qos_calc_rr(m, osts, lqr);
+ rc = lod_qos_calc_rr(m, &m->lod_ost_descs, osts, lqr);
if (rc)
GOTO(out, rc);
- rc = lod_qos_ost_in_use_clear(env, stripe_count);
+ rc = lod_qos_tgt_in_use_clear(env, stripe_count);
if (rc)
GOTO(out, rc);
- down_read(&m->lod_qos.lq_rw_sem);
+ down_read(&m->lod_ost_descs.ltd_qos.lq_rw_sem);
spin_lock(&lqr->lqr_alloc);
if (--lqr->lqr_start_count <= 0) {
- lqr->lqr_start_idx = cfs_rand() % osts->op_count;
+ lqr->lqr_start_idx = prandom_u32_max(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);
stripe_count, lqr->lqr_start_idx, lqr->lqr_start_count,
lqr->lqr_offset_idx, osts->op_count, osts->op_count);
- for (i = 0; i < osts->op_count && stripe_idx < stripe_count; i++) {
+ if (lod_comp->llc_pattern & LOV_PATTERN_OVERSTRIPING)
+ stripes_per_ost =
+ (lod_comp->llc_stripe_count - 1)/osts->op_count + 1;
+
+ for (i = 0; i < osts->op_count * stripes_per_ost
+ && stripe_idx < stripe_count; i++) {
array_idx = (lqr->lqr_start_idx + lqr->lqr_offset_idx) %
osts->op_count;
++lqr->lqr_start_idx;
stripe_idx, array_idx, ost_idx);
if ((ost_idx == LOV_QOS_EMPTY) ||
- !cfs_bitmap_check(m->lod_ost_bitmap, ost_idx))
+ !test_bit(ost_idx, m->lod_ost_bitmap))
continue;
/* Fail Check before osc_precreate() is called
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, lod_comp, ost_idx,
+ speed, &stripe_idx, stripe,
+ ost_indices, th, &overstriped,
+ reserve);
spin_lock(&lqr->lqr_alloc);
if (rc != 0 && OST_TGT(m, ost_idx)->ltd_connecting)
}
spin_unlock(&lqr->lqr_alloc);
- up_read(&m->lod_qos.lq_rw_sem);
+ up_read(&m->lod_ost_descs.ltd_qos.lq_rw_sem);
+
+ /* If there are enough OSTs, a component with overstriping requested
+ * will not actually end up overstriped. The comp should reflect this.
+ */
+ if (!overstriped)
+ lod_comp->llc_pattern &= ~LOV_PATTERN_OVERSTRIPING;
if (stripe_idx) {
lod_comp->llc_stripe_count = stripe_idx;
RETURN(rc);
}
+static int
+lod_qos_mdt_in_use_init(const struct lu_env *env,
+ const struct lu_tgt_descs *ltd,
+ u32 stripe_idx, u32 stripe_count,
+ const struct lu_tgt_pool *pool,
+ struct dt_object **stripes)
+{
+ u32 mdt_idx;
+ struct lu_tgt_desc *mdt;
+ int i, j;
+ int rc;
+
+ rc = lod_qos_tgt_in_use_clear(env, stripe_count);
+ if (rc)
+ return rc;
+
+ /* if stripe_idx > 1, we are splitting directory, mark existing stripes
+ * in_use. Because for either split or creation, stripe 0 is local,
+ * don't mark it in use.
+ */
+ for (i = 1; i < stripe_idx; i++) {
+ LASSERT(stripes[i]);
+ for (j = 0; j < pool->op_count; j++) {
+ mdt_idx = pool->op_array[j];
+
+ if (!test_bit(mdt_idx, ltd->ltd_tgt_bitmap))
+ continue;
+
+ mdt = LTD_TGT(ltd, mdt_idx);
+ if (&mdt->ltd_tgt->dd_lu_dev ==
+ stripes[i]->do_lu.lo_dev)
+ lod_qos_tgt_in_use(env, i, mdt_idx);
+ }
+ }
+
+ return 0;
+}
+
+/**
+ * Allocate a striping using round-robin algorithm.
+ *
+ * Allocates a new striping using round-robin algorithm. The function refreshes
+ * all the internal structures (statfs cache, array of available remote MDTs
+ * sorted with regard to MDS, etc). The number of stripes required is taken from
+ * the object (must be prepared by the caller). The caller should ensure nobody
+ * else is trying to create a striping on the object in parallel. All the
+ * internal structures (like pools, etc) are protected and no additional locking
+ * is required. The function succeeds even if a single stripe is allocated.
+ *
+ * \param[in] env execution environment for this thread
+ * \param[in] lo LOD object
+ * \param[out] stripes striping created
+ *
+ * \retval positive stripe objects allocated, including the first stripe
+ * allocated outside
+ * \retval -ENOSPC if not enough MDTs are found
+ * \retval negative negated errno for other failures
+ */
+int lod_mdt_alloc_rr(const struct lu_env *env, struct lod_object *lo,
+ struct dt_object **stripes, u32 stripe_idx,
+ u32 stripe_count)
+{
+ struct lod_device *lod = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
+ struct lu_tgt_descs *ltd = &lod->lod_mdt_descs;
+ struct lu_tgt_pool *pool;
+ struct lu_qos_rr *lqr;
+ struct lu_tgt_desc *mdt;
+ struct lu_object_conf conf = { .loc_flags = LOC_F_NEW };
+ struct lu_fid fid = { 0 };
+ struct dt_object *dto;
+ unsigned int pool_idx;
+ unsigned int i;
+ u32 saved_idx = stripe_idx;
+ u32 start_mdt;
+ u32 mdt_idx;
+ bool use_degraded = false;
+ int tgt_connecting = 0;
+ int rc;
+
+ ENTRY;
+
+ pool = <d->ltd_tgt_pool;
+ lqr = <d->ltd_qos.lq_rr;
+ rc = lod_qos_calc_rr(lod, ltd, pool, lqr);
+ if (rc)
+ RETURN(rc);
+
+ rc = lod_qos_mdt_in_use_init(env, ltd, stripe_idx, stripe_count, pool,
+ stripes);
+ if (rc)
+ RETURN(rc);
+
+ down_read(<d->ltd_qos.lq_rw_sem);
+ spin_lock(&lqr->lqr_alloc);
+ if (--lqr->lqr_start_count <= 0) {
+ lqr->lqr_start_idx = prandom_u32_max(pool->op_count);
+ lqr->lqr_start_count =
+ (LOV_CREATE_RESEED_MIN / max(pool->op_count, 1U) +
+ LOV_CREATE_RESEED_MULT) * max(pool->op_count, 1U);
+ } else if (stripe_count - 1 >= pool->op_count ||
+ lqr->lqr_start_idx > pool->op_count) {
+ /* If we have allocated from all of the tgts, slowly
+ * precess the next start if the tgt/stripe count isn't
+ * already doing this for us. */
+ lqr->lqr_start_idx %= pool->op_count;
+ if (stripe_count - 1 > 1 &&
+ (pool->op_count % (stripe_count - 1)) != 1)
+ ++lqr->lqr_offset_idx;
+ }
+ start_mdt = lqr->lqr_start_idx;
+
+repeat_find:
+ QOS_DEBUG("want=%d start_idx=%d start_count=%d offset=%d active=%d count=%d\n",
+ stripe_count - 1, lqr->lqr_start_idx, lqr->lqr_start_count,
+ lqr->lqr_offset_idx, pool->op_count, pool->op_count);
+
+ for (i = 0; i < pool->op_count && stripe_idx < stripe_count; i++) {
+ pool_idx = (lqr->lqr_start_idx + lqr->lqr_offset_idx) %
+ pool->op_count;
+ ++lqr->lqr_start_idx;
+ mdt_idx = lqr->lqr_pool.op_array[pool_idx];
+ mdt = LTD_TGT(ltd, mdt_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, pool_idx, mdt_idx);
+
+ if (mdt_idx == LOV_QOS_EMPTY ||
+ !test_bit(mdt_idx, ltd->ltd_tgt_bitmap))
+ continue;
+
+ /* do not put >1 objects on one MDT */
+ if (lod_qos_is_tgt_used(env, mdt_idx, stripe_idx))
+ continue;
+
+ rc = lod_is_tgt_usable(ltd, mdt);
+ if (rc) {
+ if (mdt->ltd_connecting)
+ tgt_connecting = 1;
+ continue;
+ }
+
+ /* try to use another OSP if this one is degraded */
+ if (mdt->ltd_statfs.os_state & OS_STATFS_DEGRADED &&
+ !use_degraded) {
+ QOS_DEBUG("#%d: degraded\n", mdt_idx);
+ continue;
+ }
+ spin_unlock(&lqr->lqr_alloc);
+
+ rc = dt_fid_alloc(env, mdt->ltd_tgt, &fid, NULL, NULL);
+ if (rc < 0) {
+ QOS_DEBUG("#%d: alloc FID failed: %dl\n", mdt_idx, rc);
+ spin_lock(&lqr->lqr_alloc);
+ continue;
+ }
+
+ dto = dt_locate_at(env, mdt->ltd_tgt, &fid,
+ lo->ldo_obj.do_lu.lo_dev->ld_site->ls_top_dev,
+ &conf);
+
+ spin_lock(&lqr->lqr_alloc);
+ if (IS_ERR(dto)) {
+ QOS_DEBUG("can't alloc stripe on #%u: %d\n",
+ mdt->ltd_index, (int) PTR_ERR(dto));
+
+ if (mdt->ltd_connecting)
+ tgt_connecting = 1;
+ continue;
+ }
+
+ lod_qos_tgt_in_use(env, stripe_idx, mdt_idx);
+ stripes[stripe_idx++] = dto;
+ }
+
+ if (!use_degraded && stripe_idx < stripe_count) {
+ /* Try again, allowing slower MDTs */
+ use_degraded = true;
+ lqr->lqr_start_idx = start_mdt;
+
+ tgt_connecting = 0;
+ goto repeat_find;
+ }
+ spin_unlock(&lqr->lqr_alloc);
+ up_read(<d->ltd_qos.lq_rw_sem);
+
+ if (stripe_idx > saved_idx)
+ /* at least one stripe is allocated */
+ RETURN(stripe_idx);
+
+ /* nobody provided us with a single object */
+ if (tgt_connecting)
+ RETURN(-EINPROGRESS);
+
+ RETURN(-ENOSPC);
+}
+
/**
* Allocate a specific striping layout on a user defined set of OSTs.
*
* 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
* \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, __u64 reserve)
{
struct lod_layout_component *lod_comp;
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 int array_idx = 0;
int stripe_count = 0;
LASSERT(lo->ldo_comp_cnt > comp_idx && lo->ldo_comp_entries != NULL);
lod_comp = &lo->ldo_comp_entries[comp_idx];
LASSERT(lod_comp->llc_ostlist.op_array);
+ LASSERT(lod_comp->llc_ostlist.op_count);
- rc = lod_qos_ost_in_use_clear(env, lod_comp->llc_stripe_count);
+ rc = lod_qos_tgt_in_use_clear(env, lod_comp->llc_stripe_count);
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) {
i++, array_idx = (array_idx + 1) % lod_comp->llc_stripe_count) {
__u32 ost_idx = lod_comp->llc_ostlist.op_array[array_idx];
- if (!cfs_bitmap_check(m->lod_ost_bitmap, ost_idx)) {
+ if (!test_bit(ost_idx, m->lod_ost_bitmap)) {
rc = -ENODEV;
break;
}
- /*
- * do not put >1 objects on a single OST
+ /* do not put >1 objects on a single OST, except for
+ * overstriping
*/
- if (lod_qos_is_ost_used(env, ost_idx, stripe_count)) {
+ if (lod_qos_is_tgt_used(env, ost_idx, stripe_count) &&
+ !(lod_comp->llc_pattern & LOV_PATTERN_OVERSTRIPING)) {
rc = -EINVAL;
break;
}
- rc = lod_statfs_and_check(env, m, ost_idx, sfs);
+ rc = lod_statfs_and_check(env, m, &m->lod_ost_descs,
+ LTD_TGT(&m->lod_ost_descs, ost_idx),
+ reserve);
if (rc < 0) /* this OSP doesn't feel well */
break;
/*
* 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);
+ lod_qos_tgt_in_use(env, stripe_count, ost_idx);
stripe[stripe_count] = o;
+ ost_indices[stripe_count] = ost_idx;
stripe_count++;
}
* 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
* \retval -EINVAL requested offset is invalid
* \retval negative errno on failure
*/
-static int lod_alloc_specific(const struct lu_env *env, struct lod_object *lo,
- struct dt_object **stripe, int flags,
- struct thandle *th, int comp_idx,
- struct ost_pool *inuse)
+static int lod_ost_alloc_specific(const struct lu_env *env,
+ struct lod_object *lo,
+ struct dt_object **stripe, __u32 *ost_indices,
+ int flags, struct thandle *th, int comp_idx,
+ __u64 reserve)
{
struct lod_layout_component *lod_comp;
struct lod_device *m = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
- struct obd_statfs *sfs = &lod_env_info(env)->lti_osfs;
- struct dt_object *o;
- __u32 ost_idx;
- unsigned int i, array_idx, ost_count;
- int rc, stripe_num = 0;
- int speed = 0;
- struct pool_desc *pool = NULL;
- struct ost_pool *osts;
+ struct dt_object *o;
+ struct lu_tgt_desc *tgt;
+ __u32 ost_idx;
+ unsigned int i, array_idx, ost_count;
+ int rc, stripe_num = 0;
+ int speed = 0;
+ struct pool_desc *pool = NULL;
+ struct lu_tgt_pool *osts;
+ int stripes_per_ost = 1;
+ bool overstriped = false;
ENTRY;
LASSERT(lo->ldo_comp_cnt > comp_idx && lo->ldo_comp_entries != NULL);
lod_comp = &lo->ldo_comp_entries[comp_idx];
- rc = lod_qos_ost_in_use_clear(env, lod_comp->llc_stripe_count);
+ rc = lod_qos_tgt_in_use_clear(env, lod_comp->llc_stripe_count);
if (rc)
GOTO(out, rc);
down_read(&pool_tgt_rw_sem(pool));
osts = &(pool->pool_obds);
} else {
- osts = &(m->lod_pool_info);
+ osts = &m->lod_ost_descs.ltd_tgt_pool;
}
ost_count = osts->op_count;
GOTO(out, rc = -EINVAL);
}
- for (i = 0; i < ost_count;
+ if (lod_comp->llc_pattern & LOV_PATTERN_OVERSTRIPING)
+ stripes_per_ost =
+ (lod_comp->llc_stripe_count - 1)/ost_count + 1;
+
+ for (i = 0; i < ost_count * stripes_per_ost;
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))
+ if (!test_bit(ost_idx, m->lod_ost_bitmap))
continue;
/* Fail Check before osc_precreate() is called
continue;
/*
- * do not put >1 objects on a single OST
+ * do not put >1 objects on a single OST, except for
+ * overstriping, where it is intended
*/
- if (lod_qos_is_ost_used(env, ost_idx, stripe_num))
- continue;
+ if (lod_qos_is_tgt_used(env, ost_idx, stripe_num)) {
+ if (lod_comp->llc_pattern & LOV_PATTERN_OVERSTRIPING)
+ overstriped = true;
+ else
+ continue;
+ }
/*
* 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;
+ tgt = LTD_TGT(&m->lod_ost_descs, ost_idx);
+
/* 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);
+ rc = lod_statfs_and_check(env, m, &m->lod_ost_descs,
+ tgt, reserve);
if (rc) {
/* this OSP doesn't feel well */
continue;
}
/*
- * We expect number of precreated objects in f_ffree at
- * the first iteration, skip OSPs with no objects ready
- * don't apply this logic to OST specified with stripe_offset
+ * 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)
+ if (i && !tgt->ltd_statfs.os_fprecreated && !speed)
continue;
o = lod_qos_declare_object_on(env, m, ost_idx, th);
/*
* 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);
+ lod_qos_tgt_in_use(env, stripe_num, ost_idx);
stripe[stripe_num] = o;
+ ost_indices[stripe_num] = ost_idx;
stripe_num++;
/* We have enough stripes */
PFID(lu_object_fid(lod2lu_obj(lo))), stripe_num,
lod_comp->llc_stripe_count);
rc = stripe_num == 0 ? -ENOSPC : -EFBIG;
+
+ /* If there are enough OSTs, a component with overstriping requessted
+ * will not actually end up overstriped. The comp should reflect this.
+ */
+ if (rc == 0 && !overstriped)
+ lod_comp->llc_pattern &= ~LOV_PATTERN_OVERSTRIPING;
+
out:
if (pool != NULL) {
up_read(&pool_tgt_rw_sem(pool));
/* put back ref got by lod_find_pool() */
lod_pool_putref(pool);
- }
-
- RETURN(rc);
-}
-
-/**
- * Check whether QoS allocation should be used.
- *
- * A simple helper to decide when QoS allocation should be used:
- * if it's just a single available target or the used space is
- * evenly distributed among the targets at the moment, then QoS
- * allocation algorithm should not be used.
- *
- * \param[in] lod LOD device
- *
- * \retval 0 should not be used
- * \retval 1 should be used
- */
-static inline int lod_qos_is_usable(struct lod_device *lod)
-{
-#ifdef FORCE_QOS
- /* to be able to debug QoS code */
- return 1;
-#endif
-
- /* Detect -EAGAIN early, before expensive lock is taken. */
- if (!lod->lod_qos.lq_dirty && lod->lod_qos.lq_same_space)
- return 0;
-
- if (lod->lod_desc.ld_active_tgt_count < 2)
- return 0;
+ }
- return 1;
+ RETURN(rc);
}
/**
* 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
* \retval -EINVAL requested OST index is invalid
* \retval negative errno on failure
*/
-static int lod_alloc_qos(const struct lu_env *env, struct lod_object *lo,
- struct dt_object **stripe, int flags,
- struct thandle *th, int comp_idx,
- struct ost_pool *inuse)
+static int lod_ost_alloc_qos(const struct lu_env *env, struct lod_object *lo,
+ struct dt_object **stripe, __u32 *ost_indices,
+ int flags, struct thandle *th, int comp_idx,
+ __u64 reserve)
{
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;
+ struct lu_tgt_pool *osts;
unsigned int i;
- __u32 nfound, good_osts, stripe_count, stripe_count_min;
+ __u32 nfound, good_osts, stripe_count, stripe_count_min;
+ bool overstriped = false;
+ int stripes_per_ost = 1;
int rc = 0;
ENTRY;
down_read(&pool_tgt_rw_sem(pool));
osts = &(pool->pool_obds);
} else {
- osts = &(lod->lod_pool_info);
+ osts = &lod->lod_ost_descs.ltd_tgt_pool;
}
/* Detect -EAGAIN early, before expensive lock is taken. */
- if (!lod_qos_is_usable(lod))
+ if (!ltd_qos_is_usable(&lod->lod_ost_descs))
GOTO(out_nolock, rc = -EAGAIN);
+ if (lod_comp->llc_pattern & LOV_PATTERN_OVERSTRIPING)
+ stripes_per_ost =
+ (lod_comp->llc_stripe_count - 1)/osts->op_count + 1;
+
/* Do actual allocation, use write lock here. */
- down_write(&lod->lod_qos.lq_rw_sem);
+ down_write(&lod->lod_ost_descs.ltd_qos.lq_rw_sem);
/*
* Check again, while we were sleeping on @lq_rw_sem things could
* change.
*/
- if (!lod_qos_is_usable(lod))
+ if (!ltd_qos_is_usable(&lod->lod_ost_descs))
GOTO(out, rc = -EAGAIN);
- rc = lod_qos_calc_ppo(lod);
+ rc = ltd_qos_penalties_calc(&lod->lod_ost_descs);
if (rc)
GOTO(out, rc);
- rc = lod_qos_ost_in_use_clear(env, lod_comp->llc_stripe_count);
+ rc = lod_qos_tgt_in_use_clear(env, lod_comp->llc_stripe_count);
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(lod->lod_ost_bitmap, osts->op_array[i]))
+ if (!test_bit(osts->op_array[i], lod->lod_ost_bitmap))
continue;
ost = OST_TGT(lod, osts->op_array[i]);
ost->ltd_qos.ltq_usable = 0;
- rc = lod_statfs_and_check(env, lod, osts->op_array[i], sfs);
+ rc = lod_statfs_and_check(env, lod, &lod->lod_ost_descs,
+ ost, reserve);
if (rc) {
/* this OSP doesn't feel well */
continue;
}
- if (sfs->os_state & OS_STATE_DEGRADED)
+ if (ost->ltd_statfs.os_state & OS_STATFS_DEGRADED)
continue;
/* Fail Check before osc_precreate() is called
- so we can only 'fail' single OSC. */
+ * so we can only 'fail' single OSC.
+ */
if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_PRECREATE) &&
osts->op_array[i] == 0)
continue;
ost->ltd_qos.ltq_usable = 1;
- lod_qos_calc_weight(lod, osts->op_array[i]);
+ lu_tgt_qos_weight_calc(ost);
total_weight += ost->ltd_qos.ltq_weight;
good_osts++;
if (good_osts < stripe_count_min)
GOTO(out, rc = -EAGAIN);
- /* We have enough osts */
- if (good_osts < stripe_count)
- stripe_count = good_osts;
+ /* If we do not have enough OSTs for the requested stripe count, do not
+ * put more stripes per OST than requested.
+ */
+ if (stripe_count / stripes_per_ost > good_osts)
+ stripe_count = good_osts * stripes_per_ost;
/* Find enough OSTs with weighted random allocation. */
nfound = 0;
while (nfound < stripe_count) {
- __u64 rand, cur_weight;
+ 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;
- }
+ rand = lu_prandom_u64_max(total_weight);
/* On average, this will hit larger-weighted OSTs more often.
- * 0-weight OSTs will always get used last (only when rand=0) */
+ * 0-weight OSTs will always get used last (only when rand=0)
+ */
for (i = 0; i < osts->op_count; i++) {
__u32 idx = osts->op_array[i];
- if (!cfs_bitmap_check(lod->lod_ost_bitmap, idx))
+ if (lod_should_avoid_ost(lo, lag, idx))
continue;
ost = OST_TGT(lod, idx);
QOS_DEBUG("stripe=%d to idx=%d\n", nfound, idx);
/*
- * do not put >1 objects on a single OST
+ * do not put >1 objects on a single OST, except for
+ * overstriping
*/
- if (lod_qos_is_ost_used(env, idx, nfound) ||
- lod_comp_is_ost_used(inuse, idx))
+ if ((lod_comp_is_ost_used(env, lo, idx)) &&
+ !(lod_comp->llc_pattern & LOV_PATTERN_OVERSTRIPING))
continue;
+ if (lod_qos_is_tgt_used(env, idx, nfound)) {
+ if (lod_comp->llc_pattern &
+ LOV_PATTERN_OVERSTRIPING)
+ overstriped = true;
+ else
+ continue;
+ }
+
o = lod_qos_declare_object_on(env, lod, idx, th);
if (IS_ERR(o)) {
QOS_DEBUG("can't declare object on #%u: %d\n",
continue;
}
- lod_qos_ost_in_use(env, nfound, idx);
- lod_comp_ost_in_use(inuse, idx);
- stripe[nfound++] = o;
- lod_qos_used(lod, osts, idx, &total_weight);
+ lod_avoid_update(lo, lag);
+ lod_qos_tgt_in_use(env, nfound, idx);
+ stripe[nfound] = o;
+ ost_indices[nfound] = idx;
+ ltd_qos_update(&lod->lod_ost_descs, ost, &total_weight);
+ nfound++;
rc = 0;
break;
}
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;
- lod->lod_qos.lq_same_space = 0;
+ lod->lod_ost_descs.ltd_qos.lq_dirty = 1;
+ lod->lod_ost_descs.ltd_qos.lq_same_space = 0;
rc = -EAGAIN;
}
+ /* If there are enough OSTs, a component with overstriping requessted
+ * will not actually end up overstriped. The comp should reflect this.
+ */
+ if (rc == 0 && !overstriped)
+ lod_comp->llc_pattern &= ~LOV_PATTERN_OVERSTRIPING;
+
out:
- up_write(&lod->lod_qos.lq_rw_sem);
+ up_write(&lod->lod_ost_descs.ltd_qos.lq_rw_sem);
out_nolock:
if (pool != NULL) {
}
/**
- * Find largest stripe count the caller can use.
+ * Allocate a striping using an algorithm with weights.
+ *
+ * The function allocates remote MDT objects to create a striping, the first
+ * object was already allocated on current MDT to ensure master object and
+ * the first object are on the same MDT. The algorithm used is based on weights
+ * (both free space and inodes), and it's trying to ensure the space/inodes are
+ * used evenly by MDTs and MDSs. The striping configuration (# of stripes,
+ * offset, pool) is taken from the object and is prepared by the caller.
+ *
+ * If prepared configuration can't be met due to too few MDTs, then allocation
+ * fails.
+ *
+ * No concurrent allocation is allowed on the object and this must be ensured
+ * by the caller. All the internal structures are protected by the function.
+ *
+ * The algorithm has two steps: find available MDTs and calculate their
+ * weights, then select the MDTs with their weights used as the probability.
+ * An MDT 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[in] stripe_idx starting stripe index to allocate, if it's not
+ * 0, we are restriping directory
+ * \param[in] stripe_count total stripe count
+ * \param[out] stripes striping created
+ *
+ * \retval positive stripes allocated, and it should be equal to
+ * lo->ldo_dir_stripe_count
+ * \retval -EAGAIN not enough tgts are found for specified stripe count
+ * \retval -EINVAL requested MDT index is invalid
+ * \retval negative errno on failure
+ */
+int lod_mdt_alloc_qos(const struct lu_env *env, struct lod_object *lo,
+ struct dt_object **stripes, u32 stripe_idx,
+ u32 stripe_count)
+{
+ struct lod_device *lod = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
+ struct lu_tgt_descs *ltd = &lod->lod_mdt_descs;
+ struct lu_object_conf conf = { .loc_flags = LOC_F_NEW };
+ struct lu_fid fid = { 0 };
+ const struct lu_tgt_pool *pool;
+ struct lu_tgt_desc *mdt;
+ struct dt_object *dto;
+ u64 total_weight = 0;
+ u32 saved_idx = stripe_idx;
+ u32 mdt_idx;
+ unsigned int good_mdts;
+ unsigned int i;
+ int rc = 0;
+
+ ENTRY;
+
+ LASSERT(stripe_idx <= stripe_count);
+ if (stripe_idx == stripe_count)
+ RETURN(stripe_count);
+
+ /* use MDT pool in @ltd, once MDT pool is supported in the future, it
+ * can be passed in as argument like OST object allocation.
+ */
+ pool = <d->ltd_tgt_pool;
+
+ /* Detect -EAGAIN early, before expensive lock is taken. */
+ if (!ltd_qos_is_usable(ltd))
+ RETURN(-EAGAIN);
+
+ rc = lod_qos_mdt_in_use_init(env, ltd, stripe_idx, stripe_count, pool,
+ stripes);
+ if (rc)
+ RETURN(rc);
+
+ /* Do actual allocation, use write lock here. */
+ down_write(<d->ltd_qos.lq_rw_sem);
+
+ /*
+ * Check again, while we were sleeping on @lq_rw_sem things could
+ * change.
+ */
+ if (!ltd_qos_is_usable(ltd))
+ GOTO(unlock, rc = -EAGAIN);
+
+ rc = ltd_qos_penalties_calc(ltd);
+ if (rc)
+ GOTO(unlock, rc);
+
+ good_mdts = 0;
+ /* Find all the MDTs that are valid stripe candidates */
+ for (i = 0; i < pool->op_count; i++) {
+ if (!test_bit(pool->op_array[i], ltd->ltd_tgt_bitmap))
+ continue;
+
+ mdt = LTD_TGT(ltd, pool->op_array[i]);
+ mdt->ltd_qos.ltq_usable = 0;
+
+ rc = lod_is_tgt_usable(ltd, mdt);
+ if (rc)
+ continue;
+
+ if (mdt->ltd_statfs.os_state & OS_STATFS_DEGRADED)
+ continue;
+
+ mdt->ltd_qos.ltq_usable = 1;
+ lu_tgt_qos_weight_calc(mdt);
+ total_weight += mdt->ltd_qos.ltq_weight;
+
+ good_mdts++;
+ }
+
+ QOS_DEBUG("found %d good MDTs\n", good_mdts);
+
+ if (good_mdts < stripe_count - stripe_idx)
+ GOTO(unlock, rc = -EAGAIN);
+
+ /* Find enough MDTs with weighted random allocation. */
+ while (stripe_idx < stripe_count) {
+ u64 rand, cur_weight;
+
+ cur_weight = 0;
+ rc = -ENOSPC;
+
+ rand = lu_prandom_u64_max(total_weight);
+
+ /* On average, this will hit larger-weighted MDTs more often.
+ * 0-weight MDT will always get used last (only when rand=0) */
+ for (i = 0; i < pool->op_count; i++) {
+ int rc2;
+
+ mdt_idx = pool->op_array[i];
+ mdt = LTD_TGT(ltd, mdt_idx);
+
+ if (!mdt->ltd_qos.ltq_usable)
+ continue;
+
+ cur_weight += mdt->ltd_qos.ltq_weight;
+
+ QOS_DEBUG("stripe_count=%d stripe_index=%d cur_weight=%llu rand=%llu total_weight=%llu\n",
+ stripe_count, stripe_idx, cur_weight, rand,
+ total_weight);
+
+ if (cur_weight < rand)
+ continue;
+
+ QOS_DEBUG("stripe=%d to idx=%d\n",
+ stripe_idx, mdt_idx);
+
+ if (lod_qos_is_tgt_used(env, mdt_idx, stripe_idx))
+ continue;
+
+ rc2 = dt_fid_alloc(env, mdt->ltd_tgt, &fid, NULL, NULL);
+ if (rc2 < 0) {
+ QOS_DEBUG("can't alloc FID on #%u: %d\n",
+ mdt_idx, rc2);
+ continue;
+ }
+
+ conf.loc_flags = LOC_F_NEW;
+ dto = dt_locate_at(env, mdt->ltd_tgt, &fid,
+ lo->ldo_obj.do_lu.lo_dev->ld_site->ls_top_dev,
+ &conf);
+ if (IS_ERR(dto)) {
+ QOS_DEBUG("can't alloc stripe on #%u: %d\n",
+ mdt_idx, (int) PTR_ERR(dto));
+ continue;
+ }
+
+ lod_qos_tgt_in_use(env, stripe_idx, mdt_idx);
+ stripes[stripe_idx] = dto;
+ ltd_qos_update(ltd, mdt, &total_weight);
+ stripe_idx++;
+ rc = 0;
+ break;
+ }
+
+ /* no MDT found on this iteration, give up */
+ if (rc)
+ break;
+ }
+
+ if (unlikely(stripe_idx != stripe_count)) {
+ /*
+ * when the decision to use weighted algorithm was made
+ * we had enough appropriate OSPs, but this state can
+ * change anytime (no space on MDT, broken connection, etc)
+ * so it's possible OSP won't be able to provide us with
+ * an object due to just changed state
+ */
+ QOS_DEBUG("%s: wanted %d objects, found only %d\n",
+ lod2obd(lod)->obd_name, stripe_count, stripe_idx);
+ for (i = saved_idx; i < stripe_idx; i++) {
+ LASSERT(stripes[i] != NULL);
+ dt_object_put(env, stripes[i]);
+ stripes[i] = NULL;
+ }
+
+ /* makes sense to rebalance next time */
+ ltd->ltd_qos.lq_dirty = 1;
+ ltd->ltd_qos.lq_same_space = 0;
+
+ rc = -EAGAIN;
+ } else {
+ rc = stripe_idx;
+ }
+
+unlock:
+ up_write(<d->ltd_qos.lq_rw_sem);
+
+ RETURN(rc);
+}
+
+/**
+ * Check stripe count the caller can use.
*
+ * For new layouts (no initialized components), check the total size of the
+ * layout against the maximum EA size from the backing file system. This
+ * stops us from creating a layout which will be too large once initialized.
+ *
+ * For existing layouts (with initialized components):
* Find the maximal possible stripe count not greater than \a stripe_count.
- * Sometimes suggested stripecount can't be reached for a number of reasons:
- * lack of enough active OSTs or the backend does not support EAs that large.
- * If the passed one is 0, then the filesystem's default one is used.
+ * If the provided stripe count is 0, then the filesystem's default is used.
*
* \param[in] lod LOD device
* \param[in] lo The lod_object
* \retval the maximum usable stripe count
*/
__u16 lod_get_stripe_count(struct lod_device *lod, struct lod_object *lo,
- __u16 stripe_count)
+ __u16 stripe_count, bool overstriping)
{
__u32 max_stripes = LOV_MAX_STRIPE_COUNT_OLD;
+ /* max stripe count is based on OSD ea size */
+ unsigned int easize = lod->lod_osd_max_easize;
+ int i;
+
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;
+ stripe_count =
+ lod->lod_ost_descs.ltd_lov_desc.ld_default_stripe_count;
if (!stripe_count)
stripe_count = 1;
+ /* Overstriping allows more stripes than targets */
+ if (stripe_count >
+ lod->lod_ost_descs.ltd_lov_desc.ld_active_tgt_count &&
+ !overstriping)
+ stripe_count =
+ lod->lod_ost_descs.ltd_lov_desc.ld_active_tgt_count;
+
+ if (lo->ldo_is_composite) {
+ struct lod_layout_component *lod_comp;
+ unsigned int header_sz = sizeof(struct lov_comp_md_v1);
+ unsigned int init_comp_sz = 0;
+ unsigned int total_comp_sz = 0;
+ unsigned int comp_sz;
+
+ header_sz += sizeof(struct lov_comp_md_entry_v1) *
+ lo->ldo_comp_cnt;
+
+ for (i = 0; i < lo->ldo_comp_cnt; i++) {
+ lod_comp = &lo->ldo_comp_entries[i];
+ comp_sz = lov_mds_md_size(lod_comp->llc_stripe_count,
+ LOV_MAGIC_V3);
+ total_comp_sz += comp_sz;
+ if (lod_comp->llc_flags & LCME_FL_INIT)
+ init_comp_sz += comp_sz;
+ }
- /* stripe count is based on whether OSD can handle larger EA sizes */
- if (lod->lod_osd_max_easize > 0) {
- unsigned int easize = lod->lod_osd_max_easize;
- int i;
+ if (init_comp_sz > 0)
+ total_comp_sz = init_comp_sz;
- if (lo->ldo_is_composite) {
- struct lod_layout_component *lod_comp;
- unsigned int header_sz = sizeof(struct lov_comp_md_v1);
-
- header_sz += sizeof(struct lov_comp_md_entry_v1) *
- lo->ldo_comp_cnt;
- for (i = 0; i < lo->ldo_comp_cnt; i++) {
- lod_comp = &lo->ldo_comp_entries[i];
- if (lod_comp->llc_flags & LCME_FL_INIT)
- header_sz += lov_mds_md_size(
- lod_comp->llc_stripe_count,
- LOV_MAGIC_V3);
- }
- if (easize > header_sz)
- easize -= header_sz;
- else
- easize = 0;
- }
+ header_sz += total_comp_sz;
- max_stripes = lov_mds_md_max_stripe_count(easize, LOV_MAGIC_V3);
+ if (easize > header_sz)
+ easize -= header_sz;
+ else
+ easize = 0;
}
+ max_stripes = lov_mds_md_max_stripe_count(easize, LOV_MAGIC_V3);
+ max_stripes = (max_stripes == 0) ? 0 : max_stripes - 1;
+
return (stripe_count < max_stripes) ? stripe_count : max_stripes;
}
struct lov_ost_data_v1 *objs;
__u32 magic;
__u16 comp_cnt;
+ __u16 mirror_cnt;
int rc = 0, i;
ENTRY;
+ 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);
+ magic != LOV_MAGIC_COMP_V1 && magic != LOV_MAGIC_FOREIGN)
+ 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 if (magic == LOV_MAGIC_FOREIGN) {
+ struct lov_foreign_md *foreign;
+ size_t length;
+
+ if (buf->lb_len < offsetof(typeof(*foreign), lfm_value)) {
+ CDEBUG(D_LAYOUT,
+ "buf len %zu < min lov_foreign_md size (%zu)\n",
+ buf->lb_len,
+ offsetof(typeof(*foreign), lfm_value));
+ GOTO(out, rc = -EINVAL);
+ }
+ foreign = (struct lov_foreign_md *)buf->lb_buf;
+ length = foreign_size_le(foreign);
+ if (buf->lb_len < length) {
+ CDEBUG(D_LAYOUT,
+ "buf len %zu < this lov_foreign_md size (%zu)\n",
+ buf->lb_len, length);
+ GOTO(out, rc = -EINVAL);
+ }
+
+ /* just cache foreign LOV EA raw */
+ rc = lod_alloc_foreign_lov(mo, length);
+ if (rc)
+ GOTO(out, rc);
+ memcpy(mo->ldo_foreign_lov, buf->lb_buf, length);
+ GOTO(out, rc);
} 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;
if (mo->ldo_is_composite) {
offs = le32_to_cpu(comp_v1->lcm_entries[i].lcme_offset);
v1 = (struct lov_mds_md_v1 *)((char *)comp_v1 + offs);
+ v3 = (struct lov_mds_md_v3 *)v1;
magic = le32_to_cpu(v1->lmm_magic);
ext = &comp_v1->lcm_entries[i].lcme_extent;
lod_comp->llc_extent.e_end = le64_to_cpu(ext->e_end);
lod_comp->llc_flags =
le32_to_cpu(comp_v1->lcm_entries[i].lcme_flags);
+ if (lod_comp->llc_flags & LCME_FL_NOSYNC)
+ lod_comp->llc_timestamp = le64_to_cpu(
+ comp_v1->lcm_entries[i].lcme_timestamp);
lod_comp->llc_id =
le32_to_cpu(comp_v1->lcm_entries[i].lcme_id);
if (lod_comp->llc_id == LCME_ID_INVAL)
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);
}
const struct lu_buf *buf)
{
struct lod_layout_component *lod_comp;
- struct lod_device *d = lu2lod_dev(lod2lu_obj(lo)->lo_dev);
- struct lov_desc *desc = &d->lod_desc;
- struct lov_user_md_v1 *v1 = NULL;
- struct lov_user_md_v3 *v3 = NULL;
- struct lov_comp_md_v1 *comp_v1 = NULL;
- __u32 magic;
- __u16 comp_cnt;
- int i, rc;
+ struct lod_device *d = lu2lod_dev(lod2lu_obj(lo)->lo_dev);
+ struct lov_desc *desc = &d->lod_ost_descs.ltd_lov_desc;
+ struct lov_user_md_v1 *v1 = NULL;
+ struct lov_user_md_v3 *v3 = NULL;
+ struct lov_comp_md_v1 *comp_v1 = NULL;
+ struct lov_foreign_md *lfm = 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, lo, buf, false);
+ 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 */
+ if (lo->ldo_is_foreign)
+ lod_free_foreign_lov(lo);
+ else
+ lod_free_comp_entries(lo);
+
+ rc = lod_verify_striping(env, d, lo, buf, false);
if (rc)
RETURN(-EINVAL);
- lod_free_comp_entries(lo);
-
v3 = buf->lb_buf;
v1 = buf->lb_buf;
comp_v1 = buf->lb_buf;
+ /* {lmm,lfm}_magic position/length work for all LOV formats */
magic = v1->lmm_magic;
if (unlikely(le32_to_cpu(magic) & LOV_MAGIC_DEFINED)) {
/* fall trhough */
case LOV_USER_MAGIC_COMP_V1:
break;
+ case __swab32(LOV_USER_MAGIC_FOREIGN):
+ lfm = buf->lb_buf;
+ __swab32s(&lfm->lfm_magic);
+ __swab32s(&lfm->lfm_length);
+ __swab32s(&lfm->lfm_type);
+ __swab32s(&lfm->lfm_flags);
+ magic = lfm->lfm_magic;
+ /* fall through */
+ case LOV_USER_MAGIC_FOREIGN:
+ if (!lfm)
+ lfm = buf->lb_buf;
+ rc = lod_alloc_foreign_lov(lo, foreign_size(lfm));
+ if (rc)
+ RETURN(rc);
+ memcpy(lo->ldo_foreign_lov, buf->lb_buf, foreign_size(lfm));
+ RETURN(0);
default:
CERROR("%s: unrecognized magic %X\n",
lod2obd(d)->obd_name, magic);
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;
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_CL_COMP_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 &&
- lov_pattern(v1->lmm_pattern) != LOV_PATTERN_MDT) {
+ lov_pattern(v1->lmm_pattern) != LOV_PATTERN_MDT &&
+ lov_pattern(v1->lmm_pattern) !=
+ (LOV_PATTERN_RAID0 | LOV_PATTERN_OVERSTRIPING)) {
CDEBUG(D_LAYOUT, "%s: invalid pattern: %x\n",
lod2obd(d)->obd_name, v1->lmm_pattern);
GOTO(free_comp, rc = -EINVAL);
}
lod_comp->llc_pattern = v1->lmm_pattern;
- lod_comp->llc_stripe_size = desc->ld_default_stripe_size;
- if (v1->lmm_stripe_size)
- lod_comp->llc_stripe_size = v1->lmm_stripe_size;
+ lod_comp->llc_stripe_size = v1->lmm_stripe_size;
+ lod_adjust_stripe_size(lod_comp, desc->ld_default_stripe_size);
lod_comp->llc_stripe_count = desc->ld_default_stripe_count;
if (v1->lmm_stripe_count ||
lov_pattern(v1->lmm_pattern) == LOV_PATTERN_MDT)
lod_comp->llc_stripe_count = v1->lmm_stripe_count;
+ if (lov_pattern(lod_comp->llc_pattern) == LOV_PATTERN_MDT &&
+ lod_comp->llc_stripe_count != 0) {
+ CDEBUG(D_LAYOUT, "%s: invalid stripe count: %u\n",
+ lod2obd(d)->obd_name,
+ lod_comp->llc_stripe_count);
+ GOTO(free_comp, rc = -EINVAL);
+ }
+
lod_comp->llc_stripe_offset = v1->lmm_stripe_offset;
lod_obj_set_pool(lo, i, pool_name);
- LASSERT(ergo(lov_pattern(lod_comp->llc_pattern) ==
- LOV_PATTERN_MDT, lod_comp->llc_stripe_count == 0));
-
if (pool_name == NULL)
continue;
}
}
- if (lod_comp->llc_stripe_count > pool_tgt_count(pool))
+ if (lod_comp->llc_stripe_count > pool_tgt_count(pool) &&
+ !(lod_comp->llc_pattern & LOV_PATTERN_OVERSTRIPING))
lod_comp->llc_stripe_count = pool_tgt_count(pool);
lod_pool_putref(pool);
}
/**
+ * 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_avoid_guide *lag = &lod_env_info(env)->lti_avoid;
+ struct cfs_bitmap *bitmap = NULL;
+ __u32 *new_oss = NULL;
+
+ lag->lag_ost_avail = lod->lod_ost_count;
+
+ /* reset OSS avoid guide array */
+ lag->lag_oaa_count = 0;
+ if (lag->lag_oss_avoid_array &&
+ lag->lag_oaa_size < lod->lod_ost_count) {
+ OBD_FREE_PTR_ARRAY(lag->lag_oss_avoid_array, 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 (lod->lod_ost_count <= 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(lod->lod_ost_count);
+ 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_PTR_ARRAY(new_oss, lod->lod_ost_count);
+ if (!new_oss) {
+ CFS_FREE_BITMAP(bitmap);
+ return -ENOMEM;
+ }
+ }
+
+ if (new_oss) {
+ lag->lag_oss_avoid_array = new_oss;
+ lag->lag_oaa_size = lod->lod_ost_count;
+ }
+ 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,
+ * NOTE: don't use lod_comp_inited(comp) to judge
+ * whether @comp has been inited, since during
+ * declare phase, comp->llc_stripe has been allocated
+ * while it's init flag not been set until the exec
+ * phase.
+ */
+ if (!lu_extent_is_overlapped(&comp->llc_extent,
+ &lod_comp->llc_extent) ||
+ !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 lu_svr_qos *lsq;
+ int k;
+
+ ost = OST_TGT(lod, comp->llc_ost_indices[j]);
+ lsq = ost->ltd_qos.ltq_svr;
+
+ if (cfs_bitmap_check(bitmap, ost->ltd_index))
+ continue;
+
+ QOS_DEBUG("OST%d used in conflicting mirror "
+ "component\n", ost->ltd_index);
+ 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] ==
+ lsq->lsq_id)
+ break;
+ }
+ if (k == lag->lag_oaa_count) {
+ lag->lag_oss_avoid_array[k] =
+ lsq->lsq_id;
+ lag->lag_oaa_count++;
+ }
+ }
+ }
+ }
+}
+
+/**
* Create a striping for an obejct.
*
* The function creates a new striping for the object. The function tries QoS
* \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, __u64 reserve)
{
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);
LASSERT(lo->ldo_comp_cnt > comp_idx && lo->ldo_comp_entries != NULL);
lod_comp = &lo->ldo_comp_entries[comp_idx];
+ LASSERT(!(lod_comp->llc_flags & LCME_FL_EXTENSION));
/* A released component is being created */
if (lod_comp->llc_pattern & LOV_PATTERN_F_RELEASED)
* 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);
+ lod_qos_statfs_update(env, d, &d->lod_ost_descs);
stripe_len = lod_get_stripe_count(d, lo,
- lod_comp->llc_stripe_count);
+ lod_comp->llc_stripe_count,
+ lod_comp->llc_pattern &
+ LOV_PATTERN_OVERSTRIPING);
+
if (stripe_len == 0)
GOTO(out, rc = -ERANGE);
lod_comp->llc_stripe_count = stripe_len;
- OBD_ALLOC(stripe, sizeof(stripe[0]) * stripe_len);
+ OBD_ALLOC_PTR_ARRAY(stripe, stripe_len);
if (stripe == NULL)
GOTO(out, rc = -ENOMEM);
+ OBD_ALLOC_PTR_ARRAY(ost_indices, stripe_len);
+ if (!ost_indices)
+ GOTO(out, rc = -ENOMEM);
+repeat:
lod_getref(&d->lod_ost_descs);
/* XXX: support for non-0 files w/o objects */
CDEBUG(D_OTHER, "tgt_count %d stripe_count %d\n",
- d->lod_desc.ld_tgt_count, stripe_len);
+ d->lod_ost_count, stripe_len);
- if (lod_comp->llc_ostlist.op_array) {
- rc = lod_alloc_ost_list(env, lo, stripe, th, comp_idx,
- inuse);
+ if (lod_comp->llc_ostlist.op_array &&
+ lod_comp->llc_ostlist.op_count) {
+ rc = lod_alloc_ost_list(env, lo, stripe, ost_indices,
+ th, comp_idx, reserve);
} 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);
+
+ QOS_DEBUG("collecting conflict osts for comp[%d]\n",
+ comp_idx);
+ lod_collect_avoidance(lo, lag, comp_idx);
+
+ rc = lod_ost_alloc_qos(env, lo, stripe, ost_indices,
+ flag, th, comp_idx, reserve);
if (rc == -EAGAIN)
- rc = lod_alloc_rr(env, lo, stripe, flag, th,
- comp_idx, inuse);
+ rc = lod_ost_alloc_rr(env, lo, stripe,
+ ost_indices, flag, th,
+ comp_idx, reserve);
} else {
- rc = lod_alloc_specific(env, lo, stripe, flag, th,
- comp_idx, inuse);
+ rc = lod_ost_alloc_specific(env, lo, stripe,
+ ost_indices, flag, th,
+ comp_idx, reserve);
}
+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);
+ /* In case there is no space on any OST, let's ignore
+ * the @reserve space to avoid an error at the init
+ * time, probably the actual IO will be less than the
+ * given @reserve space (aka extension_size). */
+ if (reserve) {
+ reserve = 0;
+ goto repeat;
+ }
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 {
}
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_PTR_ARRAY(stripe, stripe_len);
+ if (ost_indices)
+ OBD_FREE_PTR_ARRAY(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,
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;
/* no OST available */
/* XXX: should we be waiting a bit to prevent failures during
* cluster initialization? */
- if (d->lod_ostnr == 0)
+ if (!d->lod_ost_count)
RETURN(-EIO);
/*
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
lod_comp = &lo->ldo_comp_entries[i];
extent = &lod_comp->llc_extent;
- CDEBUG(D_QOS, "%lld [%lld, %lld)\n",
- size, extent->e_start, extent->e_end);
+ QOS_DEBUG("comp[%d] %lld "DEXT"\n", i, size, PEXT(extent));
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, 0);
if (rc)
break;
}