/* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
* vim:expandtab:shiftwidth=8:tabstop=8:
*
- * Copyright (C) 2002, 2003 Cluster File Systems, Inc.
+ * GPL HEADER START
*
- * This file is part of the Lustre file system, http://www.lustre.org
- * Lustre is a trademark of Cluster File Systems, Inc.
+ * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
*
- * You may have signed or agreed to another license before downloading
- * this software. If so, you are bound by the terms and conditions
- * of that agreement, and the following does not apply to you. See the
- * LICENSE file included with this distribution for more information.
+ * 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.
*
- * If you did not agree to a different license, then this copy of Lustre
- * is open source software; you can redistribute it and/or modify it
- * under the terms of version 2 of the GNU General Public License 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).
*
- * In either case, Lustre 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
- * license text for more details.
+ * You should have received a copy of the GNU General Public License
+ * version 2 along with this program; If not, see
+ * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
+ *
+ * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
+ * CA 95054 USA or visit www.sun.com if you need additional information or
+ * have any questions.
+ *
+ * GPL HEADER END
+ */
+/*
+ * Copyright 2008 Sun Microsystems, Inc. All rights reserved
+ * Use is subject to license terms.
+ */
+/*
+ * This file is part of Lustre, http://www.lustre.org/
+ * Lustre is a trademark of Sun Microsystems, Inc.
*/
#ifndef EXPORT_SYMTAB
/* #define QOS_DEBUG 1 */
#define D_QOS D_OTHER
-#define TGT_BAVAIL(i) (lov->lov_tgts[i]->ltd_exp->exp_obd->obd_osfs.os_bavail*\
- lov->lov_tgts[i]->ltd_exp->exp_obd->obd_osfs.os_bsize)
-#define TGT_FFREE(i) (lov->lov_tgts[i]->ltd_exp->exp_obd->obd_osfs.os_ffree)
+#define TGT_BAVAIL(i) (lov->lov_tgts[i]->ltd_exp->exp_obd->obd_osfs.os_bavail *\
+ lov->lov_tgts[i]->ltd_exp->exp_obd->obd_osfs.os_bsize)
+#define TGT_FFREE(i) (lov->lov_tgts[i]->ltd_exp->exp_obd->obd_osfs.os_ffree)
int qos_add_tgt(struct obd_device *obd, __u32 index)
list_add_tail(&oss->lqo_oss_list, &temposs->lqo_oss_list);
lov->lov_qos.lq_dirty = 1;
- lov->lov_qos.lq_dirty_rr = 1;
+ lov->lov_qos.lq_rr.lqr_dirty = 1;
CDEBUG(D_QOS, "add tgt %s to OSS %s (%d OSTs)\n",
obd_uuid2str(&lov->lov_tgts[index]->ltd_uuid),
int rc = 0;
ENTRY;
+ if (!lov->lov_tgts[index])
+ RETURN(0);
+
down_write(&lov->lov_qos.lq_rw_sem);
oss = lov->lov_tgts[index]->ltd_qos.ltq_oss;
}
lov->lov_qos.lq_dirty = 1;
- lov->lov_qos.lq_dirty_rr = 1;
+ lov->lov_qos.lq_rr.lqr_dirty = 1;
out:
up_write(&lov->lov_qos.lq_rw_sem);
RETURN(rc);
__u64 ba_max, ba_min, temp;
__u32 num_active;
int rc, i, prio_wide;
+ time_t now, age;
ENTRY;
if (!lov->lov_qos.lq_dirty)
ba_min = (__u64)(-1);
ba_max = 0;
+ now = cfs_time_current_sec();
/* Calculate OST penalty per object */
/* (lov ref taken in alloc_qos) */
for (i = 0; i < lov->desc.ld_tgt_count; i++) {
lov->lov_tgts[i]->ltd_qos.ltq_penalty_per_obj =
(temp * prio_wide) >> 8;
- if (lov->lov_qos.lq_reset == 0)
+ age = (now - lov->lov_tgts[i]->ltd_qos.ltq_used) >> 3;
+ if (lov->lov_qos.lq_reset || age > 32 * lov->desc.ld_qos_maxage)
lov->lov_tgts[i]->ltd_qos.ltq_penalty = 0;
+ else if (age > lov->desc.ld_qos_maxage)
+ /* Decay the penalty by half for every 8x the update
+ * interval that the device has been idle. That gives
+ * lots of time for the statfs information to be
+ * updated (which the penalty is only a proxy for),
+ * and avoids penalizing OSS/OSTs under light load. */
+ lov->lov_tgts[i]->ltd_qos.ltq_penalty >>=
+ (age / lov->desc.ld_qos_maxage);
}
num_active = lov->lov_qos.lq_active_oss_count - 1;
/* If there's only 1 OSS, we can't penalize it, so instead
we have to double the OST penalty */
num_active = 1;
- for (i = 0; i < lov->desc.ld_tgt_count; i++)
+ for (i = 0; i < lov->desc.ld_tgt_count; i++) {
+ if (lov->lov_tgts[i] == NULL)
+ continue;
lov->lov_tgts[i]->ltd_qos.ltq_penalty_per_obj <<= 1;
+ }
}
/* Per-OSS penalty is prio * oss_avail / oss_osts / (num_oss - 1) / 2 */
temp = oss->lqo_bavail >> 1;
do_div(temp, oss->lqo_ost_count * num_active);
oss->lqo_penalty_per_obj = (temp * prio_wide) >> 8;
- if (lov->lov_qos.lq_reset == 0)
+
+ age = (now - oss->lqo_used) >> 3;
+ if (lov->lov_qos.lq_reset || age > 32 * lov->desc.ld_qos_maxage)
oss->lqo_penalty = 0;
+ else if (age > lov->desc.ld_qos_maxage)
+ /* Decay the penalty by half for every 8x the update
+ * interval that the device has been idle. That gives
+ * lots of time for the statfs information to be
+ * updated (which the penalty is only a proxy for),
+ * and avoids penalizing OSS/OSTs under light load. */
+ oss->lqo_penalty >>= (age / lov->desc.ld_qos_maxage);
}
lov->lov_qos.lq_dirty = 0;
/* Difference is less than 20% */
lov->lov_qos.lq_same_space = 1;
/* Reset weights for the next time we enter qos mode */
- lov->lov_qos.lq_reset = 0;
+ lov->lov_qos.lq_reset = 1;
}
rc = 0;
}
/* We just used this index for a stripe; adjust everyone's weights */
-static int qos_used(struct lov_obd *lov, __u32 index, __u64 *total_wt)
+static int qos_used(struct lov_obd *lov, struct ost_pool *osts,
+ __u32 index, __u64 *total_wt)
{
struct lov_qos_oss *oss;
- int i;
+ int j;
ENTRY;
/* Don't allocate from this stripe anymore, until the next alloc_qos */
lov->lov_tgts[index]->ltd_qos.ltq_penalty >>= 1;
oss->lqo_penalty >>= 1;
+ /* mark the OSS and OST as recently used */
+ lov->lov_tgts[index]->ltd_qos.ltq_used =
+ oss->lqo_used = cfs_time_current_sec();
+
/* Set max penalties for this OST and OSS */
lov->lov_tgts[index]->ltd_qos.ltq_penalty +=
lov->lov_tgts[index]->ltd_qos.ltq_penalty_per_obj *
*total_wt = 0;
/* Decrease all OST penalties */
- for (i = 0; i < lov->desc.ld_tgt_count; i++) {
+ for (j = 0; j < osts->op_count; j++) {
+ int i;
+
+ i = osts->op_array[j];
if (!lov->lov_tgts[i] || !lov->lov_tgts[i]->ltd_active)
continue;
if (lov->lov_tgts[i]->ltd_qos.ltq_penalty <
*total_wt += lov->lov_tgts[i]->ltd_qos.ltq_weight;
#ifdef QOS_DEBUG
- CDEBUG(D_QOS, "recalc tgt %d avail="LPU64
+ CDEBUG(D_QOS, "recalc tgt %d usable=%d avail="LPU64
" ostppo="LPU64" ostp="LPU64" ossppo="LPU64
" ossp="LPU64" wt="LPU64"\n",
- i, TGT_BAVAIL(i) >> 10,
+ i, lov->lov_tgts[i]->ltd_qos.ltq_usable,
+ TGT_BAVAIL(i) >> 10,
lov->lov_tgts[i]->ltd_qos.ltq_penalty_per_obj >> 10,
lov->lov_tgts[i]->ltd_qos.ltq_penalty >> 10,
lov->lov_tgts[i]->ltd_qos.ltq_oss->lqo_penalty_per_obj>>10,
- lov->lov_tgts[i]->ltd_qos.ltq_oss->lqo_penalty>>10,
- lov->lov_tgts[i]->ltd_qos.ltq_weight>>10);
+ lov->lov_tgts[i]->ltd_qos.ltq_oss->lqo_penalty >> 10,
+ lov->lov_tgts[i]->ltd_qos.ltq_weight >> 10);
#endif
}
#define LOV_QOS_EMPTY ((__u32)-1)
/* compute optimal round-robin order, based on OSTs per OSS */
-static int qos_calc_rr(struct lov_obd *lov)
+static int qos_calc_rr(struct lov_obd *lov, struct ost_pool *src_pool,
+ struct lov_qos_rr *lqr)
{
struct lov_qos_oss *oss;
- unsigned ost_count, placed;
- int i;
+ unsigned placed, real_count;
+ int i, rc;
ENTRY;
- if (!lov->lov_qos.lq_dirty_rr) {
- LASSERT(lov->lov_qos.lq_rr_size);
+ if (!lqr->lqr_dirty) {
+ LASSERT(lqr->lqr_pool.op_size);
RETURN(0);
}
+ /* Do actual allocation. */
down_write(&lov->lov_qos.lq_rw_sem);
- ost_count = lov->desc.ld_tgt_count;
-
- if (lov->lov_qos.lq_rr_size)
- OBD_FREE(lov->lov_qos.lq_rr_array, lov->lov_qos.lq_rr_size);
- lov->lov_qos.lq_rr_size = ost_count *
- sizeof(lov->lov_qos.lq_rr_array[0]);
- OBD_ALLOC(lov->lov_qos.lq_rr_array, lov->lov_qos.lq_rr_size);
- if (!lov->lov_qos.lq_rr_array) {
- lov->lov_qos.lq_rr_size = 0;
+
+ /*
+ * Check again. While we were sleeping on @lq_rw_sem something could
+ * change.
+ */
+ if (!lqr->lqr_dirty) {
+ LASSERT(lqr->lqr_pool.op_size);
up_write(&lov->lov_qos.lq_rw_sem);
- RETURN(-ENOMEM);
+ RETURN(0);
}
- for (i = 0; i < ost_count; i++)
- lov->lov_qos.lq_rr_array[i] = LOV_QOS_EMPTY;
+ real_count = src_pool->op_count;
+
+ /* Zero the pool array */
+ /* alloc_rr is holding a read lock on the pool, so nobody is adding/
+ deleting from the pool. The lq_rw_sem insures that nobody else
+ is reading. */
+ lqr->lqr_pool.op_count = real_count;
+ rc = lov_ost_pool_extend(&lqr->lqr_pool, real_count);
+ if (rc) {
+ up_write(&lov->lov_qos.lq_rw_sem);
+ RETURN(rc);
+ }
+ for (i = 0; i < lqr->lqr_pool.op_count; i++)
+ lqr->lqr_pool.op_array[i] = LOV_QOS_EMPTY;
/* Place all the OSTs from 1 OSS at the same time. */
placed = 0;
list_for_each_entry(oss, &lov->lov_qos.lq_oss_list, lqo_oss_list) {
int j = 0;
- for (i = 0; i < ost_count; i++) {
- if (lov->lov_tgts[i]->ltd_qos.ltq_oss == oss) {
+ for (i = 0; i < lqr->lqr_pool.op_count; i++) {
+ if (lov->lov_tgts[src_pool->op_array[i]] &&
+ (lov->lov_tgts[src_pool->op_array[i]]->ltd_qos.ltq_oss == oss)) {
/* Evenly space these OSTs across arrayspace */
- int next = j * ost_count / oss->lqo_ost_count;
- while (lov->lov_qos.lq_rr_array[next] !=
+ int next = j * lqr->lqr_pool.op_count / oss->lqo_ost_count;
+ while (lqr->lqr_pool.op_array[next] !=
LOV_QOS_EMPTY)
- next = (next + 1) % ost_count;
- lov->lov_qos.lq_rr_array[next] = i;
+ next = (next + 1) % lqr->lqr_pool.op_count;
+ lqr->lqr_pool.op_array[next] = src_pool->op_array[i];
j++;
placed++;
- }
+ }
}
- LASSERT(j == oss->lqo_ost_count);
}
- lov->lov_qos.lq_dirty_rr = 0;
+ lqr->lqr_dirty = 0;
up_write(&lov->lov_qos.lq_rw_sem);
- if (placed != ost_count) {
+ if (placed != real_count) {
/* This should never happen */
- LCONSOLE_ERROR("Failed to place all OSTs in the round-robin "
- "list (%d of %d).\n", placed, ost_count);
- for (i = 0; i < ost_count; i++) {
+ LCONSOLE_ERROR_MSG(0x14e, "Failed to place all OSTs in the "
+ "round-robin list (%d of %d).\n",
+ placed, real_count);
+ for (i = 0; i < lqr->lqr_pool.op_count; i++) {
LCONSOLE(D_WARNING, "rr #%d ost idx=%d\n", i,
- lov->lov_qos.lq_rr_array[i]);
+ lqr->lqr_pool.op_array[i]);
}
- lov->lov_qos.lq_dirty_rr = 1;
+ lqr->lqr_dirty = 1;
RETURN(-EAGAIN);
}
-
+
#ifdef QOS_DEBUG
- for (i = 0; i < ost_count; i++) {
- LCONSOLE(D_QOS, "rr #%d ost idx=%d\n", i,
- lov->lov_qos.lq_rr_array[i]);
+ for (i = 0; i < lqr->lqr_pool.op_count; i++) {
+ LCONSOLE(D_QOS, "rr #%d ost idx=%d\n", i,
+ lqr->lqr_pool.op_array[i]);
}
#endif
RETURN(rc);
}
+static int min_stripe_count(int stripe_cnt, int flags)
+{
+ return (flags & LOV_USES_DEFAULT_STRIPE ?
+ stripe_cnt - (stripe_cnt / 4) : stripe_cnt);
+}
+
#define LOV_CREATE_RESEED_MULT 4
#define LOV_CREATE_RESEED_MIN 1000
/* Allocate objects on osts with round-robin algorithm */
-static int alloc_rr(struct lov_obd *lov, int *idx_arr, int *stripe_cnt)
+static int alloc_rr(struct lov_obd *lov, int *idx_arr, int *stripe_cnt,
+ char *poolname, int flags)
{
- unsigned array_idx, ost_count = lov->desc.ld_tgt_count;
- unsigned ost_active_count = lov->desc.ld_active_tgt_count;
- int i, *idx_pos = idx_arr;
+ unsigned array_idx;
+ int i, rc, *idx_pos;
__u32 ost_idx;
+ int ost_start_idx_temp;
+ int speed = 0;
+ int stripe_cnt_min = min_stripe_count(*stripe_cnt, flags);
+ struct pool_desc *pool;
+ struct ost_pool *osts;
+ struct lov_qos_rr *lqr;
ENTRY;
- i = qos_calc_rr(lov);
- if (i)
- RETURN(i);
-
- if (--lov->lov_start_count <= 0) {
- lov->lov_start_idx = ll_rand() % ost_count;
- lov->lov_start_count =
- (LOV_CREATE_RESEED_MIN / max(ost_active_count, 1U) +
- LOV_CREATE_RESEED_MULT) * max(ost_active_count, 1U);
- } else if (*stripe_cnt >= ost_active_count ||
- lov->lov_start_idx > ost_count) {
+ pool = lov_find_pool(lov, poolname);
+ if (pool == NULL) {
+ osts = &(lov->lov_packed);
+ lqr = &(lov->lov_qos.lq_rr);
+ } else {
+ read_lock(&pool_tgt_rwlock(pool));
+ osts = &(pool->pool_obds);
+ lqr = &(pool->pool_rr);
+ }
+
+ rc = qos_calc_rr(lov, osts, lqr);
+ if (rc)
+ GOTO(out, rc);
+
+ if (--lqr->lqr_start_count <= 0) {
+ lqr->lqr_start_idx = ll_rand() % osts->op_count;
+ lqr->lqr_start_count =
+ (LOV_CREATE_RESEED_MIN / max(osts->op_count, 1U) +
+ LOV_CREATE_RESEED_MULT) * max(osts->op_count, 1U);
+ } else if (stripe_cnt_min >= osts->op_count ||
+ lqr->lqr_start_idx > osts->op_count) {
/* If we have allocated from all of the OSTs, slowly
- precess the next start */
- lov->lov_start_idx %= ost_count;
- ++lov->lov_offset_idx;
+ * precess the next start if the OST/stripe count isn't
+ * already doing this for us. */
+ lqr->lqr_start_idx %= osts->op_count;
+ if (*stripe_cnt > 1 && (osts->op_count % (*stripe_cnt)) != 1)
+ ++lqr->lqr_offset_idx;
}
- array_idx = (lov->lov_start_idx + lov->lov_offset_idx) % ost_count;
+ down_read(&lov->lov_qos.lq_rw_sem);
+ ost_start_idx_temp = lqr->lqr_start_idx;
+
+repeat_find:
+ array_idx = (lqr->lqr_start_idx + lqr->lqr_offset_idx) % osts->op_count;
+ idx_pos = idx_arr;
#ifdef QOS_DEBUG
- CDEBUG(D_QOS, "want %d startidx %d startcnt %d offset %d arrayidx %d\n",
- *stripe_cnt, lov->lov_start_idx, lov->lov_start_count,
- lov->lov_offset_idx, array_idx);
+ CDEBUG(D_QOS, "pool '%s' want %d startidx %d startcnt %d offset %d "
+ "active %d count %d arrayidx %d\n", poolname,
+ *stripe_cnt, lqr->lqr_start_idx, lqr->lqr_start_count,
+ lqr->lqr_offset_idx, osts->op_count, osts->op_count, array_idx);
#endif
- down_read(&lov->lov_qos.lq_rw_sem);
- for (i = 0; i < ost_count; i++, array_idx=(array_idx + 1) % ost_count) {
- ++lov->lov_start_idx;
- ost_idx = lov->lov_qos.lq_rr_array[array_idx];
+ for (i = 0; i < osts->op_count;
+ i++, array_idx=(array_idx + 1) % osts->op_count) {
+ ++lqr->lqr_start_idx;
+ ost_idx = lqr->lqr_pool.op_array[array_idx];
#ifdef QOS_DEBUG
CDEBUG(D_QOS, "#%d strt %d act %d strp %d ary %d idx %d\n",
- i, lov->lov_start_idx,
- lov->lov_tgts[ost_idx] ?
+ i, lqr->lqr_start_idx,
+ ((ost_idx != LOV_QOS_EMPTY) && lov->lov_tgts[ost_idx]) ?
lov->lov_tgts[ost_idx]->ltd_active : 0,
idx_pos - idx_arr, array_idx, ost_idx);
#endif
- if (!lov->lov_tgts[ost_idx] ||
+ if ((ost_idx == LOV_QOS_EMPTY) || !lov->lov_tgts[ost_idx] ||
!lov->lov_tgts[ost_idx]->ltd_active)
continue;
+
+ /* Fail Check before osc_precreate() is called
+ so we can only 'fail' single OSC. */
+ if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_PRECREATE) && ost_idx == 0)
+ continue;
+
+ /* Drop slow OSCs if we can */
+ if (obd_precreate(lov->lov_tgts[ost_idx]->ltd_exp) > speed)
+ continue;
+
*idx_pos = ost_idx;
idx_pos++;
/* We have enough stripes */
if (idx_pos - idx_arr == *stripe_cnt)
break;
}
+ if ((speed < 2) && (idx_pos - idx_arr < stripe_cnt_min)) {
+ /* Try again, allowing slower OSCs */
+ speed++;
+ lqr->lqr_start_idx = ost_start_idx_temp;
+ goto repeat_find;
+ }
+
up_read(&lov->lov_qos.lq_rw_sem);
*stripe_cnt = idx_pos - idx_arr;
- RETURN(0);
+out:
+ if (pool != NULL)
+ read_unlock(&pool_tgt_rwlock(pool));
+ RETURN(rc);
}
/* alloc objects on osts with specific stripe offset */
static int alloc_specific(struct lov_obd *lov, struct lov_stripe_md *lsm,
int *idx_arr)
{
- unsigned ost_idx, ost_count = lov->desc.ld_tgt_count;
- int i, *idx_pos = idx_arr;
+ unsigned ost_idx, array_idx, ost_count;
+ int i, rc, *idx_pos;
+ int speed = 0;
+ struct pool_desc *pool;
+ struct ost_pool *osts;
ENTRY;
- ost_idx = lsm->lsm_oinfo[0]->loi_ost_idx;
- for (i = 0; i < ost_count; i++, ost_idx = (ost_idx + 1) % ost_count) {
+ pool = lov_find_pool(lov, lsm->lsm_pool_name);
+ if (pool == NULL) {
+ osts = &(lov->lov_packed);
+ } else {
+ read_lock(&pool_tgt_rwlock(pool));
+ osts = &(pool->pool_obds);
+ }
+
+ ost_count = osts->op_count;
+
+repeat_find:
+ /* search loi_ost_idx in ost array */
+ array_idx = 0;
+ for (i = 0; i < ost_count; i++) {
+ if (osts->op_array[i] == lsm->lsm_oinfo[0]->loi_ost_idx) {
+ array_idx = i;
+ break;
+ }
+ }
+ if (i == ost_count) {
+ CERROR("Start index %d not found in pool '%s'\n",
+ lsm->lsm_oinfo[0]->loi_ost_idx, lsm->lsm_pool_name);
+ GOTO(out, rc = -EINVAL);
+ }
+
+ idx_pos = idx_arr;
+ for (i = 0; i < ost_count;
+ i++, array_idx = (array_idx + 1) % ost_count) {
+ ost_idx = osts->op_array[array_idx];
+
if (!lov->lov_tgts[ost_idx] ||
!lov->lov_tgts[ost_idx]->ltd_active) {
continue;
}
+
+ /* Fail Check before osc_precreate() is called
+ so we can only 'fail' single OSC. */
+ if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_PRECREATE) && ost_idx == 0)
+ continue;
+
+ /* Drop slow OSCs if we can, but not for requested start idx.
+ *
+ * This means "if OSC is slow and it is not the requested
+ * start OST, then it can be skipped, otherwise skip it only
+ * if it is inactive/recovering/out-of-space." */
+ if ((obd_precreate(lov->lov_tgts[ost_idx]->ltd_exp) > speed) &&
+ (i != 0 || speed >= 2))
+ continue;
+
*idx_pos = ost_idx;
idx_pos++;
- /* got enough ost */
+ /* We have enough stripes */
if (idx_pos - idx_arr == lsm->lsm_stripe_count)
- RETURN(0);
+ GOTO(out, rc = 0);
}
+ if (speed < 2) {
+ /* Try again, allowing slower OSCs */
+ speed++;
+ goto repeat_find;
+ }
+
/* If we were passed specific striping params, then a failure to
* meet those requirements is an error, since we can't reallocate
* that memory (it might be part of a larger array or something).
*
* We can only get here if lsm_stripe_count was originally > 1.
*/
- CERROR("can't lstripe objid "LPX64": have "LPSZ" want %u\n",
- lsm->lsm_object_id, idx_pos - idx_arr, lsm->lsm_stripe_count);
- RETURN(-EFBIG);
+ CERROR("can't lstripe objid "LPX64": have %d want %u\n",
+ lsm->lsm_object_id, (int)(idx_pos - idx_arr),
+ lsm->lsm_stripe_count);
+ rc = -EFBIG;
+out:
+ if (pool != NULL)
+ read_unlock(&pool_tgt_rwlock(pool));
+ RETURN(rc);
}
/* Alloc objects on osts with optimization based on:
- free space
- network resources (shared OSS's)
*/
-static int alloc_qos(struct obd_export *exp, int *idx_arr, int *stripe_cnt)
+static int alloc_qos(struct obd_export *exp, int *idx_arr, int *stripe_cnt,
+ char *poolname, int flags)
{
struct lov_obd *lov = &exp->exp_obd->u.lov;
static time_t last_warn = 0;
time_t now = cfs_time_current_sec();
__u64 total_bavail, total_weight = 0;
- __u32 ost_count;
int nfound, good_osts, i, warn = 0, rc = 0;
+ int stripe_cnt_min = min_stripe_count(*stripe_cnt, flags);
+ struct pool_desc *pool;
+ struct ost_pool *osts;
+ struct lov_qos_rr *lqr;
ENTRY;
+ if (stripe_cnt_min < 1)
+ RETURN(-EINVAL);
+
+ pool = lov_find_pool(lov, poolname);
+ if (pool == NULL) {
+ osts = &(lov->lov_packed);
+ lqr = &(lov->lov_qos.lq_rr);
+ } else {
+ read_lock(&pool_tgt_rwlock(pool));
+ osts = &(pool->pool_obds);
+ lqr = &(pool->pool_rr);
+ }
+
lov_getref(exp->exp_obd);
+
+ /* Detect -EAGAIN early, before expensive lock is taken. */
+ if (!lov->lov_qos.lq_dirty && lov->lov_qos.lq_same_space)
+ GOTO(out_nolock, rc = -EAGAIN);
+
+ /* Do actual allocation, use write lock here. */
down_write(&lov->lov_qos.lq_rw_sem);
- ost_count = lov->desc.ld_tgt_count;
+ /*
+ * Check again, while we were sleeping on @lq_rw_sem things could
+ * change.
+ */
+ if (!lov->lov_qos.lq_dirty && lov->lov_qos.lq_same_space)
+ GOTO(out, rc = -EAGAIN);
if (lov->desc.ld_active_tgt_count < 2)
GOTO(out, rc = -EAGAIN);
if (cfs_time_sub(now, last_warn) > 60 * 30)
warn = 1;
/* Find all the OSTs that are valid stripe candidates */
- for (i = 0; i < ost_count; i++) {
+ for (i = 0; i < osts->op_count; i++) {
__u64 bavail;
- if (!lov->lov_tgts[i] || !lov->lov_tgts[i]->ltd_active)
+ if (!lov->lov_tgts[osts->op_array[i]] ||
+ !lov->lov_tgts[osts->op_array[i]]->ltd_active)
continue;
- bavail = TGT_BAVAIL(i);
+ bavail = TGT_BAVAIL(osts->op_array[i]);
if (!bavail) {
if (warn) {
CDEBUG(D_QOS, "no free space on %s\n",
- obd_uuid2str(&lov->lov_tgts[i]->ltd_uuid));
+ obd_uuid2str(&lov->lov_tgts[osts->op_array[i]]->ltd_uuid));
last_warn = now;
}
continue;
}
- if (!TGT_FFREE(i)) {
+ if (!TGT_FFREE(osts->op_array[i])) {
if (warn) {
CDEBUG(D_QOS, "no free inodes on %s\n",
- obd_uuid2str(&lov->lov_tgts[i]->ltd_uuid));
+ obd_uuid2str(&lov->lov_tgts[osts->op_array[i]]->ltd_uuid));
last_warn = now;
}
continue;
}
- lov->lov_tgts[i]->ltd_qos.ltq_usable = 1;
- qos_calc_weight(lov, i);
+ /* Fail Check before osc_precreate() is called
+ so we can only 'fail' single OSC. */
+ if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_PRECREATE) && osts->op_array[i] == 0)
+ continue;
+
+ if (obd_precreate(lov->lov_tgts[osts->op_array[i]]->ltd_exp) > 2)
+ continue;
+
+ lov->lov_tgts[osts->op_array[i]]->ltd_qos.ltq_usable = 1;
+ qos_calc_weight(lov, osts->op_array[i]);
total_bavail += bavail;
- total_weight += lov->lov_tgts[i]->ltd_qos.ltq_weight;
+ total_weight += lov->lov_tgts[osts->op_array[i]]->ltd_qos.ltq_weight;
good_osts++;
}
+#ifdef QOS_DEBUG
+ CDEBUG(D_QOS, "found %d good osts\n", good_osts);
+#endif
+
+ if (good_osts < stripe_cnt_min)
+ GOTO(out, rc = -EAGAIN);
+
if (!total_bavail)
GOTO(out, rc = -ENOSPC);
- /* if we don't have enough good OSTs, we reduce the stripe count. */
+ /* We have enough osts */
if (good_osts < *stripe_cnt)
*stripe_cnt = good_osts;
/* On average, this will hit larger-weighted osts more often.
0-weight osts will always get used last (only when rand=0).*/
- for (i = 0; i < ost_count; i++) {
- if (!lov->lov_tgts[i]->ltd_qos.ltq_usable)
+ for (i = 0; i < osts->op_count; i++) {
+ if (!lov->lov_tgts[osts->op_array[i]] ||
+ !lov->lov_tgts[osts->op_array[i]]->ltd_qos.ltq_usable)
continue;
- cur_weight += lov->lov_tgts[i]->ltd_qos.ltq_weight;
+
+ cur_weight += lov->lov_tgts[osts->op_array[i]]->ltd_qos.ltq_weight;
+#ifdef QOS_DEBUG
+ CDEBUG(D_QOS, "stripe_cnt=%d nfound=%d cur_weight="LPU64
+ " rand="LPU64" total_weight="LPU64"\n",
+ *stripe_cnt, nfound, cur_weight, rand, total_weight);
+#endif
if (cur_weight >= rand) {
#ifdef QOS_DEBUG
CDEBUG(D_QOS, "assigned stripe=%d to idx=%d\n",
- nfound, i);
+ nfound, osts->op_array[i]);
#endif
- idx_arr[nfound++] = i;
- qos_used(lov, i, &total_weight);
+ idx_arr[nfound++] = osts->op_array[i];
+ qos_used(lov, osts, osts->op_array[i], &total_weight);
rc = 0;
break;
}
out:
up_write(&lov->lov_qos.lq_rw_sem);
+out_nolock:
+ if (pool != NULL)
+ read_unlock(&pool_tgt_rwlock(pool));
+
if (rc == -EAGAIN)
- rc = alloc_rr(lov, idx_arr, stripe_cnt);
+ rc = alloc_rr(lov, idx_arr, stripe_cnt, poolname, flags);
lov_putref(exp->exp_obd);
RETURN(rc);
/* return new alloced stripe count on success */
static int alloc_idx_array(struct obd_export *exp, struct lov_stripe_md *lsm,
- int newea, int **idx_arr, int *arr_cnt)
+ int newea, int **idx_arr, int *arr_cnt, int flags)
{
struct lov_obd *lov = &exp->exp_obd->u.lov;
int stripe_cnt = lsm->lsm_stripe_count;
tmp_arr[i] = -1;
if (newea ||
- lsm->lsm_oinfo[0]->loi_ost_idx >= lov->desc.ld_tgt_count)
- rc = alloc_qos(exp, tmp_arr, &stripe_cnt);
+ lsm->lsm_oinfo[0]->loi_ost_idx >= lov->desc.ld_tgt_count)
+ rc = alloc_qos(exp, tmp_arr, &stripe_cnt,
+ lsm->lsm_pool_name, flags);
else
rc = alloc_specific(lov, lsm, tmp_arr);
struct obdo *src_oa = set->set_oi->oi_oa;
struct obd_trans_info *oti = set->set_oti;
int i, stripes, rc = 0, newea = 0;
- int *idx_arr, idx_cnt = 0;
+ int flag = LOV_USES_ASSIGNED_STRIPE;
+ int *idx_arr = NULL, idx_cnt = 0;
ENTRY;
LASSERT(src_oa->o_valid & OBD_MD_FLID);
+ LASSERT(src_oa->o_valid & OBD_MD_FLGROUP);
if (set->set_oi->oi_md == NULL) {
- int stripe_cnt = lov_get_stripecnt(lov, 0);
+ int stripes_def = lov_get_stripecnt(lov, 0);
/* If the MDS file was truncated up to some size, stripe over
* enough OSTs to allow the file to be created at that size.
/* Find a small number of stripes we can use
(up to # of active osts). */
stripes = 1;
- lov_getref(exp->exp_obd);
for (i = 0; i < lov->desc.ld_tgt_count; i++) {
if (!lov->lov_tgts[i] ||
!lov->lov_tgts[i]->ltd_active)
break;
stripes++;
}
- lov_putref(exp->exp_obd);
- if (stripes < stripe_cnt)
- stripes = stripe_cnt;
+ if (stripes < stripes_def)
+ stripes = stripes_def;
} else {
- stripes = stripe_cnt;
+ flag = LOV_USES_DEFAULT_STRIPE;
+ stripes = stripes_def;
}
rc = lov_alloc_memmd(&set->set_oi->oi_md, stripes,
LOV_MAGIC);
if (rc < 0)
GOTO(out_err, rc);
- rc = 0;
newea = 1;
+ rc = 0;
}
lsm = set->set_oi->oi_md;
lsm->lsm_object_id = src_oa->o_id;
+ lsm->lsm_object_gr = src_oa->o_gr;
+
if (!lsm->lsm_stripe_size)
lsm->lsm_stripe_size = lov->desc.ld_default_stripe_size;
if (!lsm->lsm_pattern) {
lsm->lsm_pattern = lov->desc.ld_pattern;
}
- stripes = alloc_idx_array(exp, lsm, newea, &idx_arr, &idx_cnt);
- LASSERT(stripes <= lsm->lsm_stripe_count);
+ stripes = alloc_idx_array(exp, lsm, newea, &idx_arr, &idx_cnt, flag);
if (stripes <= 0)
GOTO(out_err, rc = stripes ? stripes : -EIO);
+ LASSERTF(stripes <= lsm->lsm_stripe_count,"requested %d allocated %d\n",
+ lsm->lsm_stripe_count, stripes);
for (i = 0; i < stripes; i++) {
struct lov_request *req;
if (req->rq_oi.oi_md == NULL)
GOTO(out_err, rc = -ENOMEM);
- req->rq_oi.oi_oa = obdo_alloc();
+ OBDO_ALLOC(req->rq_oi.oi_oa);
if (req->rq_oi.oi_oa == NULL)
GOTO(out_err, rc = -ENOMEM);
CDEBUG(D_INODE, "stripe %d has size "LPU64"/"LPU64"\n",
i, req->rq_oi.oi_oa->o_size, src_oa->o_size);
}
-
}
LASSERT(set->set_count == stripes);
out_err:
if (newea && rc)
obd_free_memmd(exp, &set->set_oi->oi_md);
- free_idx_array(idx_arr, idx_cnt);
+ if (idx_arr)
+ free_idx_array(idx_arr, idx_cnt);
EXIT;
return rc;
}
ENTRY;
lov->lov_qos.lq_dirty = 1;
}
-
git://git.whamcloud.com / fs / lustre-release.git / blobdiff