void lu_qos_rr_init(struct lu_qos_rr *lqr);
int lqos_add_tgt(struct lu_qos *qos, struct lu_tgt_desc *ltd);
int lqos_del_tgt(struct lu_qos *qos, struct lu_tgt_desc *ltd);
+bool lqos_is_usable(struct lu_qos *qos, __u32 active_tgt_nr);
+int lqos_calc_penalties(struct lu_qos *qos, struct lu_tgt_descs *ltd,
+ __u32 active_tgt_nr, __u32 maxage, bool is_mdt);
+void lqos_calc_weight(struct lu_tgt_desc *tgt);
+int lqos_recalc_weight(struct lu_qos *qos, struct lu_tgt_descs *ltd,
+ struct lu_tgt_desc *tgt, __u32 active_tgt_nr,
+ __u64 *total_wt);
u64 lu_prandom_u64_max(u64 ep_ro);
int lu_tgt_descs_init(struct lu_tgt_descs *ltd);
MODULES := lmv
-lmv-objs := lmv_obd.o lmv_intent.o lmv_fld.o lproc_lmv.o lmv_qos.o
+lmv-objs := lmv_obd.o lmv_intent.o lmv_fld.o lproc_lmv.o
@INCLUDE_RULES@
struct lmv_tgt_desc *lmv_locate_tgt(struct lmv_obd *lmv,
struct md_op_data *op_data);
-/* lmv_qos.c */
-struct lu_tgt_desc *lmv_locate_tgt_qos(struct lmv_obd *lmv, __u32 *mdt);
-struct lu_tgt_desc *lmv_locate_tgt_rr(struct lmv_obd *lmv, __u32 *mdt);
-
/* lproc_lmv.c */
int lmv_tunables_init(struct obd_device *obd);
RETURN(rc);
}
+static struct lu_tgt_desc *lmv_locate_tgt_qos(struct lmv_obd *lmv, __u32 *mdt)
+{
+ struct lu_tgt_desc *tgt;
+ __u64 total_weight = 0;
+ __u64 cur_weight = 0;
+ __u64 rand;
+ int rc;
+
+ ENTRY;
+
+ if (!lqos_is_usable(&lmv->lmv_qos, lmv->desc.ld_active_tgt_count))
+ RETURN(ERR_PTR(-EAGAIN));
+
+ down_write(&lmv->lmv_qos.lq_rw_sem);
+
+ if (!lqos_is_usable(&lmv->lmv_qos, lmv->desc.ld_active_tgt_count))
+ GOTO(unlock, tgt = ERR_PTR(-EAGAIN));
+
+ rc = lqos_calc_penalties(&lmv->lmv_qos, &lmv->lmv_mdt_descs,
+ lmv->desc.ld_active_tgt_count,
+ lmv->desc.ld_qos_maxage, true);
+ if (rc)
+ GOTO(unlock, tgt = ERR_PTR(rc));
+
+ lmv_foreach_tgt(lmv, tgt) {
+ tgt->ltd_qos.ltq_usable = 0;
+ if (!tgt->ltd_exp || !tgt->ltd_active)
+ continue;
+
+ tgt->ltd_qos.ltq_usable = 1;
+ lqos_calc_weight(tgt);
+ total_weight += tgt->ltd_qos.ltq_weight;
+ }
+
+ rand = lu_prandom_u64_max(total_weight);
+
+ lmv_foreach_connected_tgt(lmv, tgt) {
+ if (!tgt->ltd_qos.ltq_usable)
+ continue;
+
+ cur_weight += tgt->ltd_qos.ltq_weight;
+ if (cur_weight < rand)
+ continue;
+
+ *mdt = tgt->ltd_index;
+ lqos_recalc_weight(&lmv->lmv_qos, &lmv->lmv_mdt_descs, tgt,
+ lmv->desc.ld_active_tgt_count,
+ &total_weight);
+ GOTO(unlock, rc = 0);
+ }
+
+ /* no proper target found */
+ GOTO(unlock, tgt = ERR_PTR(-EAGAIN));
+unlock:
+ up_write(&lmv->lmv_qos.lq_rw_sem);
+
+ return tgt;
+}
+
+static struct lu_tgt_desc *lmv_locate_tgt_rr(struct lmv_obd *lmv, __u32 *mdt)
+{
+ struct lu_tgt_desc *tgt;
+ int i;
+ int index;
+
+ ENTRY;
+
+ spin_lock(&lmv->lmv_qos.lq_rr.lqr_alloc);
+ for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
+ index = (i + lmv->lmv_qos_rr_index) % lmv->desc.ld_tgt_count;
+ tgt = lmv_tgt(lmv, index);
+ if (!tgt || !tgt->ltd_exp || !tgt->ltd_active)
+ continue;
+
+ *mdt = tgt->ltd_index;
+ lmv->lmv_qos_rr_index = (*mdt + 1) % lmv->desc.ld_tgt_count;
+ spin_unlock(&lmv->lmv_qos.lq_rr.lqr_alloc);
+
+ RETURN(tgt);
+ }
+ spin_unlock(&lmv->lmv_qos.lq_rr.lqr_alloc);
+
+ RETURN(ERR_PTR(-ENODEV));
+}
+
static struct lmv_tgt_desc *
lmv_locate_tgt_by_name(struct lmv_obd *lmv, struct lmv_stripe_md *lsm,
const char *name, int namelen, struct lu_fid *fid,
+++ /dev/null
-/*
- * GPL HEADER START
- *
- * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
- *
- * This program is free software; you can redistribute it and/or modify
- * it under the terms of the GNU General Public License version 2 only,
- * as published by the Free Software Foundation.
- *
- * This program is distributed in the hope that it will be useful, but
- * WITHOUT ANY WARRANTY; without even the implied warranty of
- * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
- * General Public License version 2 for more details (a copy is included
- * in the LICENSE file that accompanied this code).
- *
- * You should have received a copy of the GNU General Public License
- * version 2 along with this program; If not, see
- * http://www.gnu.org/licenses/gpl-2.0.html
- *
- * GPL HEADER END
- */
-/*
- * This file is part of Lustre, http://www.lustre.org/
- *
- * lustre/lmv/lmv_qos.c
- *
- * LMV QoS.
- * These are the only exported functions, they provide some generic
- * infrastructure for object allocation QoS
- *
- */
-
-#define DEBUG_SUBSYSTEM S_LMV
-
-#include <asm/div64.h>
-#include <linux/random.h>
-
-#include <libcfs/libcfs.h>
-#include <uapi/linux/lustre/lustre_idl.h>
-#include <lustre_swab.h>
-#include <obd_class.h>
-
-#include "lmv_internal.h"
-
-static inline __u64 tgt_statfs_bavail(struct lu_tgt_desc *tgt)
-{
- struct obd_statfs *statfs = &tgt->ltd_statfs;
-
- return statfs->os_bavail * statfs->os_bsize;
-}
-
-static inline __u64 tgt_statfs_iavail(struct lu_tgt_desc *tgt)
-{
- return tgt->ltd_statfs.os_ffree;
-}
-
-/**
- * Calculate penalties per-tgt and per-server
- *
- * Re-calculate penalties when the configuration changes, active targets
- * change and after statfs refresh (all these are reflected by lq_dirty flag).
- * On every MDT and MDS: 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 MDS/MDTs under light load.
- * See lmv_qos_calc_weight() for how penalties are factored into the weight.
- *
- * \param[in] lmv LMV device
- *
- * \retval 0 on success
- * \retval -EAGAIN the number of MDTs isn't enough or all MDT spaces are
- * almost the same
- */
-static int lmv_qos_calc_ppts(struct lmv_obd *lmv)
-{
- struct lu_qos *qos = &lmv->lmv_qos;
- struct lu_tgt_desc *tgt;
- struct lu_svr_qos *svr;
- __u64 ba_max, ba_min, ba;
- __u64 ia_max, ia_min, ia;
- __u32 num_active;
- int prio_wide;
- time64_t now, age;
- __u32 maxage = lmv->desc.ld_qos_maxage;
- int rc;
-
- ENTRY;
-
- if (!qos->lq_dirty)
- GOTO(out, rc = 0);
-
- num_active = lmv->desc.ld_active_tgt_count;
- if (num_active < 2)
- GOTO(out, rc = -EAGAIN);
-
- /* find bavail on each server */
- list_for_each_entry(svr, &qos->lq_svr_list, lsq_svr_list) {
- svr->lsq_bavail = 0;
- svr->lsq_iavail = 0;
- }
- qos->lq_active_svr_count = 0;
-
- /*
- * How badly user wants to select targets "widely" (not recently chosen
- * and not on recent MDS's). As opposed to "freely" (free space avail.)
- * 0-256
- */
- prio_wide = 256 - qos->lq_prio_free;
-
- ba_min = (__u64)(-1);
- ba_max = 0;
- ia_min = (__u64)(-1);
- ia_max = 0;
- now = ktime_get_real_seconds();
-
- /* Calculate server penalty per object */
- lmv_foreach_tgt(lmv, tgt) {
- if (!tgt->ltd_exp || !tgt->ltd_active)
- continue;
-
- /* bavail >> 16 to avoid overflow */
- ba = tgt_statfs_bavail(tgt) >> 16;
- if (!ba)
- continue;
-
- ba_min = min(ba, ba_min);
- ba_max = max(ba, ba_max);
-
- /* iavail >> 8 to avoid overflow */
- ia = tgt_statfs_iavail(tgt) >> 8;
- if (!ia)
- continue;
-
- ia_min = min(ia, ia_min);
- ia_max = max(ia, ia_max);
-
- /* Count the number of usable MDS's */
- if (tgt->ltd_qos.ltq_svr->lsq_bavail == 0)
- qos->lq_active_svr_count++;
- tgt->ltd_qos.ltq_svr->lsq_bavail += ba;
- tgt->ltd_qos.ltq_svr->lsq_iavail += ia;
-
- /*
- * per-MDT penalty is
- * prio * bavail * iavail / (num_tgt - 1) / 2
- */
- tgt->ltd_qos.ltq_penalty_per_obj = prio_wide * ba * ia;
- do_div(tgt->ltd_qos.ltq_penalty_per_obj, num_active - 1);
- tgt->ltd_qos.ltq_penalty_per_obj >>= 1;
-
- age = (now - tgt->ltd_qos.ltq_used) >> 3;
- if (qos->lq_reset || age > 32 * maxage)
- tgt->ltd_qos.ltq_penalty = 0;
- else if (age > maxage)
- /* Decay tgt penalty. */
- tgt->ltd_qos.ltq_penalty >>= (age / maxage);
- }
-
- num_active = qos->lq_active_svr_count;
- if (num_active < 2) {
- /*
- * If there's only 1 MDS, we can't penalize it, so instead
- * we have to double the MDT penalty
- */
- num_active = 2;
- lmv_foreach_tgt(lmv, tgt) {
- if (!tgt->ltd_exp || !tgt->ltd_active)
- continue;
-
- tgt->ltd_qos.ltq_penalty_per_obj <<= 1;
- }
- }
-
- /*
- * Per-MDS penalty is
- * prio * bavail * iavail / server_tgts / (num_svr - 1) / 2
- */
- list_for_each_entry(svr, &qos->lq_svr_list, lsq_svr_list) {
- ba = svr->lsq_bavail;
- ia = svr->lsq_iavail;
- svr->lsq_penalty_per_obj = prio_wide * ba * ia;
- do_div(ba, svr->lsq_tgt_count * (num_active - 1));
- svr->lsq_penalty_per_obj >>= 1;
-
- age = (now - svr->lsq_used) >> 3;
- if (qos->lq_reset || age > 32 * maxage)
- svr->lsq_penalty = 0;
- else if (age > maxage)
- /* Decay server penalty. */
- svr->lsq_penalty >>= age / maxage;
- }
-
- qos->lq_dirty = 0;
- qos->lq_reset = 0;
-
- /*
- * If each MDT has almost same free space, do rr allocation for better
- * creation performance
- */
- qos->lq_same_space = 0;
- if ((ba_max * (256 - qos->lq_threshold_rr)) >> 8 < ba_min &&
- (ia_max * (256 - qos->lq_threshold_rr)) >> 8 < ia_min) {
- qos->lq_same_space = 1;
- /* Reset weights for the next time we enter qos mode */
- qos->lq_reset = 1;
- }
- rc = 0;
-
-out:
- if (!rc && qos->lq_same_space)
- RETURN(-EAGAIN);
-
- RETURN(rc);
-}
-
-static inline bool lmv_qos_is_usable(struct lmv_obd *lmv)
-{
- if (!lmv->lmv_qos.lq_dirty && lmv->lmv_qos.lq_same_space)
- return false;
-
- if (lmv->desc.ld_active_tgt_count < 2)
- return false;
-
- return true;
-}
-
-/**
- * Calculate weight for a given MDT.
- *
- * The final MDT weight is bavail >> 16 * iavail >> 8 minus the MDT and MDS
- * penalties. See lmv_qos_calc_ppts() for how penalties are calculated.
- *
- * \param[in] tgt MDT target descriptor
- */
-static void lmv_qos_calc_weight(struct lu_tgt_desc *tgt)
-{
- struct lu_tgt_qos *ltq = &tgt->ltd_qos;
- __u64 temp, temp2;
-
- temp = (tgt_statfs_bavail(tgt) >> 16) * (tgt_statfs_iavail(tgt) >> 8);
- temp2 = ltq->ltq_penalty + ltq->ltq_svr->lsq_penalty;
- if (temp < temp2)
- ltq->ltq_weight = 0;
- else
- ltq->ltq_weight = temp - temp2;
-}
-
-/**
- * Re-calculate weights.
- *
- * The function is called when some 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] lmv LMV device
- * \param[in] tgt target where a new object was placed
- * \param[out] total_wt new total weight for the pool
- *
- * \retval 0
- */
-static int lmv_qos_used(struct lmv_obd *lmv, struct lu_tgt_desc *tgt,
- __u64 *total_wt)
-{
- struct lu_tgt_qos *ltq;
- struct lu_svr_qos *svr;
-
- ENTRY;
-
- ltq = &tgt->ltd_qos;
- LASSERT(ltq);
-
- /* Don't allocate on this device anymore, until the next alloc_qos */
- ltq->ltq_usable = 0;
-
- svr = ltq->ltq_svr;
-
- /*
- * Decay old penalty by half (we're adding max penalty, and don't
- * want it to run away.)
- */
- ltq->ltq_penalty >>= 1;
- svr->lsq_penalty >>= 1;
-
- /* mark the MDS and MDT as recently used */
- ltq->ltq_used = svr->lsq_used = ktime_get_real_seconds();
-
- /* Set max penalties for this MDT and MDS */
- ltq->ltq_penalty += ltq->ltq_penalty_per_obj *
- lmv->desc.ld_active_tgt_count;
- svr->lsq_penalty += svr->lsq_penalty_per_obj *
- lmv->lmv_qos.lq_active_svr_count;
-
- /* Decrease all MDS penalties */
- list_for_each_entry(svr, &lmv->lmv_qos.lq_svr_list, lsq_svr_list) {
- if (svr->lsq_penalty < svr->lsq_penalty_per_obj)
- svr->lsq_penalty = 0;
- else
- svr->lsq_penalty -= svr->lsq_penalty_per_obj;
- }
-
- *total_wt = 0;
- /* Decrease all MDT penalties */
- lmv_foreach_tgt(lmv, tgt) {
- if (!tgt->ltd_exp || !tgt->ltd_active)
- continue;
-
- if (ltq->ltq_penalty < ltq->ltq_penalty_per_obj)
- ltq->ltq_penalty = 0;
- else
- ltq->ltq_penalty -= ltq->ltq_penalty_per_obj;
-
- lmv_qos_calc_weight(tgt);
-
- /* Recalc the total weight of usable osts */
- if (ltq->ltq_usable)
- *total_wt += ltq->ltq_weight;
-
- CDEBUG(D_OTHER, "recalc tgt %d usable=%d avail=%llu"
- " tgtppo=%llu tgtp=%llu svrppo=%llu"
- " svrp=%llu wt=%llu\n",
- tgt->ltd_index, ltq->ltq_usable,
- tgt_statfs_bavail(tgt) >> 10,
- ltq->ltq_penalty_per_obj >> 10,
- ltq->ltq_penalty >> 10,
- ltq->ltq_svr->lsq_penalty_per_obj >> 10,
- ltq->ltq_svr->lsq_penalty >> 10,
- ltq->ltq_weight >> 10);
- }
-
- RETURN(0);
-}
-
-struct lu_tgt_desc *lmv_locate_tgt_qos(struct lmv_obd *lmv, __u32 *mdt)
-{
- struct lu_tgt_desc *tgt;
- __u64 total_weight = 0;
- __u64 cur_weight = 0;
- __u64 rand;
- int rc;
-
- ENTRY;
-
- if (!lmv_qos_is_usable(lmv))
- RETURN(ERR_PTR(-EAGAIN));
-
- down_write(&lmv->lmv_qos.lq_rw_sem);
-
- if (!lmv_qos_is_usable(lmv))
- GOTO(unlock, tgt = ERR_PTR(-EAGAIN));
-
- rc = lmv_qos_calc_ppts(lmv);
- if (rc)
- GOTO(unlock, tgt = ERR_PTR(rc));
-
- lmv_foreach_tgt(lmv, tgt) {
- tgt->ltd_qos.ltq_usable = 0;
- if (!tgt->ltd_exp || !tgt->ltd_active)
- continue;
-
- tgt->ltd_qos.ltq_usable = 1;
- lmv_qos_calc_weight(tgt);
- total_weight += tgt->ltd_qos.ltq_weight;
- }
-
- rand = lu_prandom_u64_max(total_weight);
-
- lmv_foreach_tgt(lmv, tgt) {
- if (!tgt->ltd_qos.ltq_usable)
- continue;
-
- cur_weight += tgt->ltd_qos.ltq_weight;
- if (cur_weight < rand)
- continue;
-
- *mdt = tgt->ltd_index;
- lmv_qos_used(lmv, tgt, &total_weight);
- GOTO(unlock, rc = 0);
- }
-
- /* no proper target found */
- GOTO(unlock, tgt = ERR_PTR(-EAGAIN));
-unlock:
- up_write(&lmv->lmv_qos.lq_rw_sem);
-
- return tgt;
-}
-
-struct lu_tgt_desc *lmv_locate_tgt_rr(struct lmv_obd *lmv, __u32 *mdt)
-{
- struct lu_tgt_desc *tgt;
- int i;
-
- ENTRY;
-
- spin_lock(&lmv->lmv_qos.lq_rr.lqr_alloc);
- for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
- tgt = lmv_tgt(lmv,
- (i + lmv->lmv_qos_rr_index) % lmv->desc.ld_tgt_count);
- if (!tgt || !tgt->ltd_exp || !tgt->ltd_active)
- continue;
-
- *mdt = tgt->ltd_index;
- lmv->lmv_qos_rr_index =
- (i + lmv->lmv_qos_rr_index + 1) %
- lmv->desc.ld_tgt_count;
- spin_unlock(&lmv->lmv_qos.lq_rr.lqr_alloc);
-
- RETURN(tgt);
- }
- spin_unlock(&lmv->lmv_qos.lq_rr.lqr_alloc);
-
- RETURN(ERR_PTR(-ENODEV));
-}
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 lu_svr_qos *oss;
- __u64 ba_max, ba_min, temp;
- __u32 num_active;
- unsigned int i;
- int rc, prio_wide;
- time64_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_svr_list, lsq_svr_list)
- oss->lsq_bavail = 0;
- lod->lod_qos.lq_active_svr_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 = ktime_get_real_seconds();
- /* 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_svr->lsq_bavail == 0)
- lod->lod_qos.lq_active_svr_count++;
- OST_TGT(lod, i)->ltd_qos.ltq_svr->lsq_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_svr_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_svr_list, lsq_svr_list) {
- temp = oss->lsq_bavail >> 1;
- do_div(temp, oss->lsq_tgt_count * num_active);
- oss->lsq_penalty_per_obj = (temp * prio_wide) >> 8;
-
- age = (now - oss->lsq_used) >> 3;
- if (lod->lod_qos.lq_reset ||
- age > 32 * lod->lod_desc.ld_qos_maxage)
- oss->lsq_penalty = 0;
- else if (age > lod->lod_desc.ld_qos_maxage)
- /* Decay OSS penalty. */
- oss->lsq_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_svr->lsq_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 lu_svr_qos *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_svr;
-
- /* 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->lsq_penalty >>= 1;
-
- /* mark the OSS and OST as recently used */
- ost->ltd_qos.ltq_used = oss->lsq_used = ktime_get_real_seconds();
-
- /* Set max penalties for this OST and OSS */
- ost->ltd_qos.ltq_penalty +=
- ost->ltd_qos.ltq_penalty_per_obj * lod->lod_ostnr;
- oss->lsq_penalty += oss->lsq_penalty_per_obj *
- lod->lod_qos.lq_active_svr_count;
-
- /* Decrease all OSS penalties */
- list_for_each_entry(oss, &lod->lod_qos.lq_svr_list, lsq_svr_list) {
- if (oss->lsq_penalty < oss->lsq_penalty_per_obj)
- oss->lsq_penalty = 0;
- else
- oss->lsq_penalty -= oss->lsq_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_svr->lsq_penalty_per_obj >> 10,
- ost->ltd_qos.ltq_svr->lsq_penalty >> 10,
- ost->ltd_qos.ltq_weight >> 10);
- }
-
- RETURN(0);
-}
-
#define LOV_QOS_EMPTY ((__u32)-1)
/**
}
/**
- * Check whether QoS allocation should be used.
- *
- * A simple helper to decide when QoS allocation should be used:
- * if it's just a single available target or the used space is
- * evenly distributed among the targets at the moment, then QoS
- * allocation algorithm should not be used.
- *
- * \param[in] lod LOD device
- *
- * \retval 0 should not be used
- * \retval 1 should be used
- */
-static inline int lod_qos_is_usable(struct lod_device *lod)
-{
-#ifdef FORCE_QOS
- /* to be able to debug QoS code */
- return 1;
-#endif
-
- /* Detect -EAGAIN early, before expensive lock is taken. */
- if (!lod->lod_qos.lq_dirty && lod->lod_qos.lq_same_space)
- return 0;
-
- if (lod->lod_desc.ld_active_tgt_count < 2)
- return 0;
-
- return 1;
-}
-
-/**
* Allocate a striping using an algorithm with weights.
*
* The function allocates OST objects to create a striping. The algorithm
}
/* Detect -EAGAIN early, before expensive lock is taken. */
- if (!lod_qos_is_usable(lod))
+ if (!lqos_is_usable(&lod->lod_qos, lod->lod_desc.ld_active_tgt_count))
GOTO(out_nolock, rc = -EAGAIN);
if (lod_comp->llc_pattern & LOV_PATTERN_OVERSTRIPING)
* Check again, while we were sleeping on @lq_rw_sem things could
* change.
*/
- if (!lod_qos_is_usable(lod))
+ if (!lqos_is_usable(&lod->lod_qos, lod->lod_desc.ld_active_tgt_count))
GOTO(out, rc = -EAGAIN);
- rc = lod_qos_calc_ppo(lod);
+ rc = lqos_calc_penalties(&lod->lod_qos, &lod->lod_ost_descs,
+ lod->lod_desc.ld_active_tgt_count,
+ lod->lod_desc.ld_qos_maxage, false);
if (rc)
GOTO(out, rc);
continue;
ost->ltd_qos.ltq_usable = 1;
- lod_qos_calc_weight(lod, osts->op_array[i]);
+ lqos_calc_weight(ost);
total_weight += ost->ltd_qos.ltq_weight;
good_osts++;
lod_qos_ost_in_use(env, nfound, idx);
stripe[nfound] = o;
ost_indices[nfound] = idx;
- lod_qos_used(lod, osts, idx, &total_weight);
+ lqos_recalc_weight(&lod->lod_qos, &lod->lod_ost_descs,
+ ost,
+ lod->lod_desc.ld_active_tgt_count,
+ &total_weight);
nfound++;
rc = 0;
break;
return rand;
}
EXPORT_SYMBOL(lu_prandom_u64_max);
+
+static inline __u64 tgt_statfs_bavail(struct lu_tgt_desc *tgt)
+{
+ struct obd_statfs *statfs = &tgt->ltd_statfs;
+
+ return statfs->os_bavail * statfs->os_bsize;
+}
+
+static inline __u64 tgt_statfs_iavail(struct lu_tgt_desc *tgt)
+{
+ return tgt->ltd_statfs.os_ffree;
+}
+
+/**
+ * Calculate penalties per-tgt and per-server
+ *
+ * Re-calculate penalties when the configuration changes, active targets
+ * change and after statfs refresh (all these are reflected by lq_dirty flag).
+ * On every tgt and server: 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 server/tgt under light load.
+ * See lqos_calc_weight() for how penalties are factored into the weight.
+ *
+ * \param[in] qos lu_qos
+ * \param[in] ltd lu_tgt_descs
+ * \param[in] active_tgt_nr active tgt number
+ * \param[in] maxage qos max age
+ * \param[in] is_mdt MDT will count inode usage
+ *
+ * \retval 0 on success
+ * \retval -EAGAIN the number of tgt isn't enough or all tgt spaces are
+ * almost the same
+ */
+int lqos_calc_penalties(struct lu_qos *qos, struct lu_tgt_descs *ltd,
+ __u32 active_tgt_nr, __u32 maxage, bool is_mdt)
+{
+ struct lu_tgt_desc *tgt;
+ struct lu_svr_qos *svr;
+ __u64 ba_max, ba_min, ba;
+ __u64 ia_max, ia_min, ia = 1;
+ __u32 num_active;
+ int prio_wide;
+ time64_t now, age;
+ int rc;
+
+ ENTRY;
+
+ if (!qos->lq_dirty)
+ GOTO(out, rc = 0);
+
+ num_active = active_tgt_nr - 1;
+ if (num_active < 1)
+ GOTO(out, rc = -EAGAIN);
+
+ /* find bavail on each server */
+ list_for_each_entry(svr, &qos->lq_svr_list, lsq_svr_list) {
+ svr->lsq_bavail = 0;
+ /* if inode is not counted, set to 1 to ignore */
+ svr->lsq_iavail = is_mdt ? 0 : 1;
+ }
+ qos->lq_active_svr_count = 0;
+
+ /*
+ * How badly user wants to select targets "widely" (not recently chosen
+ * and not on recent MDS's). As opposed to "freely" (free space avail.)
+ * 0-256
+ */
+ prio_wide = 256 - qos->lq_prio_free;
+
+ ba_min = (__u64)(-1);
+ ba_max = 0;
+ ia_min = (__u64)(-1);
+ ia_max = 0;
+ now = ktime_get_real_seconds();
+
+ /* Calculate server penalty per object */
+ ltd_foreach_tgt(ltd, tgt) {
+ if (!tgt->ltd_active)
+ continue;
+
+ /* when inode is counted, bavail >> 16 to avoid overflow */
+ ba = tgt_statfs_bavail(tgt);
+ if (is_mdt)
+ ba >>= 16;
+ else
+ ba >>= 8;
+ if (!ba)
+ continue;
+
+ ba_min = min(ba, ba_min);
+ ba_max = max(ba, ba_max);
+
+ /* Count the number of usable servers */
+ if (tgt->ltd_qos.ltq_svr->lsq_bavail == 0)
+ qos->lq_active_svr_count++;
+ tgt->ltd_qos.ltq_svr->lsq_bavail += ba;
+
+ if (is_mdt) {
+ /* iavail >> 8 to avoid overflow */
+ ia = tgt_statfs_iavail(tgt) >> 8;
+ if (!ia)
+ continue;
+
+ ia_min = min(ia, ia_min);
+ ia_max = max(ia, ia_max);
+
+ tgt->ltd_qos.ltq_svr->lsq_iavail += ia;
+ }
+
+ /*
+ * per-tgt penalty is
+ * prio * bavail * iavail / (num_tgt - 1) / 2
+ */
+ tgt->ltd_qos.ltq_penalty_per_obj = prio_wide * ba * ia;
+ do_div(tgt->ltd_qos.ltq_penalty_per_obj, num_active);
+ tgt->ltd_qos.ltq_penalty_per_obj >>= 1;
+
+ age = (now - tgt->ltd_qos.ltq_used) >> 3;
+ if (qos->lq_reset || age > 32 * maxage)
+ tgt->ltd_qos.ltq_penalty = 0;
+ else if (age > maxage)
+ /* Decay tgt penalty. */
+ tgt->ltd_qos.ltq_penalty >>= (age / maxage);
+ }
+
+ num_active = qos->lq_active_svr_count - 1;
+ if (num_active < 1) {
+ /*
+ * If there's only 1 server, we can't penalize it, so instead
+ * we have to double the tgt penalty
+ */
+ num_active = 1;
+ ltd_foreach_tgt(ltd, tgt) {
+ if (!tgt->ltd_active)
+ continue;
+
+ tgt->ltd_qos.ltq_penalty_per_obj <<= 1;
+ }
+ }
+
+ /*
+ * Per-server penalty is
+ * prio * bavail * iavail / server_tgts / (num_svr - 1) / 2
+ */
+ list_for_each_entry(svr, &qos->lq_svr_list, lsq_svr_list) {
+ ba = svr->lsq_bavail;
+ ia = svr->lsq_iavail;
+ svr->lsq_penalty_per_obj = prio_wide * ba * ia;
+ do_div(ba, svr->lsq_tgt_count * num_active);
+ svr->lsq_penalty_per_obj >>= 1;
+
+ age = (now - svr->lsq_used) >> 3;
+ if (qos->lq_reset || age > 32 * maxage)
+ svr->lsq_penalty = 0;
+ else if (age > maxage)
+ /* Decay server penalty. */
+ svr->lsq_penalty >>= age / maxage;
+ }
+
+ qos->lq_dirty = 0;
+ qos->lq_reset = 0;
+
+ /*
+ * If each tgt has almost same free space, do rr allocation for better
+ * creation performance
+ */
+ qos->lq_same_space = 0;
+ if ((ba_max * (256 - qos->lq_threshold_rr)) >> 8 < ba_min &&
+ (ia_max * (256 - qos->lq_threshold_rr)) >> 8 < ia_min) {
+ qos->lq_same_space = 1;
+ /* Reset weights for the next time we enter qos mode */
+ qos->lq_reset = 1;
+ }
+ rc = 0;
+
+out:
+ if (!rc && qos->lq_same_space)
+ RETURN(-EAGAIN);
+
+ RETURN(rc);
+}
+EXPORT_SYMBOL(lqos_calc_penalties);
+
+bool lqos_is_usable(struct lu_qos *qos, __u32 active_tgt_nr)
+{
+ if (!qos->lq_dirty && qos->lq_same_space)
+ return false;
+
+ if (active_tgt_nr < 2)
+ return false;
+
+ return true;
+}
+EXPORT_SYMBOL(lqos_is_usable);
+
+/**
+ * Calculate weight for a given tgt.
+ *
+ * The final tgt weight is bavail >> 16 * iavail >> 8 minus the tgt and server
+ * penalties. See lqos_calc_ppts() for how penalties are calculated.
+ *
+ * \param[in] tgt target descriptor
+ */
+void lqos_calc_weight(struct lu_tgt_desc *tgt)
+{
+ struct lu_tgt_qos *ltq = &tgt->ltd_qos;
+ __u64 temp, temp2;
+
+ temp = (tgt_statfs_bavail(tgt) >> 16) * (tgt_statfs_iavail(tgt) >> 8);
+ temp2 = ltq->ltq_penalty + ltq->ltq_svr->lsq_penalty;
+ if (temp < temp2)
+ ltq->ltq_weight = 0;
+ else
+ ltq->ltq_weight = temp - temp2;
+}
+EXPORT_SYMBOL(lqos_calc_weight);
+
+/**
+ * Re-calculate weights.
+ *
+ * The function is called when some 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] qos lu_qos
+ * \param[in] ltd lu_tgt_descs
+ * \param[in] tgt target where a new object was placed
+ * \param[in] active_tgt_nr active tgt number
+ * \param[out] total_wt new total weight for the pool
+ *
+ * \retval 0
+ */
+int lqos_recalc_weight(struct lu_qos *qos, struct lu_tgt_descs *ltd,
+ struct lu_tgt_desc *tgt, __u32 active_tgt_nr,
+ __u64 *total_wt)
+{
+ struct lu_tgt_qos *ltq;
+ struct lu_svr_qos *svr;
+
+ ENTRY;
+
+ ltq = &tgt->ltd_qos;
+ LASSERT(ltq);
+
+ /* Don't allocate on this device anymore, until the next alloc_qos */
+ ltq->ltq_usable = 0;
+
+ svr = ltq->ltq_svr;
+
+ /*
+ * Decay old penalty by half (we're adding max penalty, and don't
+ * want it to run away.)
+ */
+ ltq->ltq_penalty >>= 1;
+ svr->lsq_penalty >>= 1;
+
+ /* mark the server and tgt as recently used */
+ ltq->ltq_used = svr->lsq_used = ktime_get_real_seconds();
+
+ /* Set max penalties for this tgt and server */
+ ltq->ltq_penalty += ltq->ltq_penalty_per_obj * active_tgt_nr;
+ svr->lsq_penalty += svr->lsq_penalty_per_obj * active_tgt_nr;
+
+ /* Decrease all MDS penalties */
+ list_for_each_entry(svr, &qos->lq_svr_list, lsq_svr_list) {
+ if (svr->lsq_penalty < svr->lsq_penalty_per_obj)
+ svr->lsq_penalty = 0;
+ else
+ svr->lsq_penalty -= svr->lsq_penalty_per_obj;
+ }
+
+ *total_wt = 0;
+ /* Decrease all tgt penalties */
+ ltd_foreach_tgt(ltd, tgt) {
+ if (!tgt->ltd_active)
+ continue;
+
+ if (ltq->ltq_penalty < ltq->ltq_penalty_per_obj)
+ ltq->ltq_penalty = 0;
+ else
+ ltq->ltq_penalty -= ltq->ltq_penalty_per_obj;
+
+ lqos_calc_weight(tgt);
+
+ /* Recalc the total weight of usable osts */
+ if (ltq->ltq_usable)
+ *total_wt += ltq->ltq_weight;
+
+ CDEBUG(D_OTHER, "recalc tgt %d usable=%d avail=%llu"
+ " tgtppo=%llu tgtp=%llu svrppo=%llu"
+ " svrp=%llu wt=%llu\n",
+ tgt->ltd_index, ltq->ltq_usable,
+ tgt_statfs_bavail(tgt) >> 10,
+ ltq->ltq_penalty_per_obj >> 10,
+ ltq->ltq_penalty >> 10,
+ ltq->ltq_svr->lsq_penalty_per_obj >> 10,
+ ltq->ltq_svr->lsq_penalty >> 10,
+ ltq->ltq_weight >> 10);
+ }
+
+ RETURN(0);
+}
+EXPORT_SYMBOL(lqos_recalc_weight);