#include <obd_lov.h>
#include "lov_internal.h"
-/* #define QOS_DEBUG 1 */
+/* #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_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)
ENTRY;
/* We only need this QOS struct on MDT, not clients - but we may not
- have registered the LOV's observer yet, so there's no way to know */
+ * have registered the LOV's observer yet, so there's no way to know */
if (!exp || !exp->exp_connection) {
CERROR("Missing connection\n");
RETURN(-ENOTCONN);
down_write(&lov->lov_qos.lq_rw_sem);
mutex_down(&lov->lov_lock);
list_for_each_entry(oss, &lov->lov_qos.lq_oss_list, lqo_oss_list) {
- if (obd_uuid_equals(&oss->lqo_uuid,
+ if (obd_uuid_equals(&oss->lqo_uuid,
&exp->exp_connection->c_remote_uuid)) {
found++;
break;
if (!found) {
OBD_ALLOC_PTR(oss);
- if (!oss)
+ if (!oss)
GOTO(out, rc = -ENOMEM);
memcpy(&oss->lqo_uuid,
&exp->exp_connection->c_remote_uuid,
/* Assume we have to move this one */
list_del(&oss->lqo_oss_list);
}
-
+
oss->lqo_ost_count++;
lov->lov_tgts[index]->ltd_qos.ltq_oss = oss;
/* Add sorted by # of OSTs. Find the first entry that we're
bigger than... */
list_for_each_entry(temposs, &lov->lov_qos.lq_oss_list, lqo_oss_list) {
- if (oss->lqo_ost_count > temposs->lqo_ost_count)
+ if (oss->lqo_ost_count > temposs->lqo_ost_count)
break;
}
/* ...and add before it. If we're the first or smallest, temposs
lov->lov_qos.lq_dirty = 1;
lov->lov_qos.lq_dirty_rr = 1;
- CDEBUG(D_QOS, "add tgt %s to OSS %s (%d OSTs)\n",
+ CDEBUG(D_QOS, "add tgt %s to OSS %s (%d OSTs)\n",
obd_uuid2str(&lov->lov_tgts[index]->ltd_uuid),
obd_uuid2str(&oss->lqo_uuid),
oss->lqo_ost_count);
oss->lqo_ost_count--;
if (oss->lqo_ost_count == 0) {
- CDEBUG(D_QOS, "removing OSS %s\n",
+ CDEBUG(D_QOS, "removing OSS %s\n",
obd_uuid2str(&oss->lqo_uuid));
list_del(&oss->lqo_oss_list);
OBD_FREE_PTR(oss);
}
-
+
lov->lov_qos.lq_dirty = 1;
lov->lov_qos.lq_dirty_rr = 1;
out:
RETURN(rc);
}
-/* Recalculate per-object penalties for OSSs and OSTs,
- depends on size of each ost in an oss */
+/* Recalculate per-object penalties for OSSs and OSTs,
+ depends on size of each ost in an oss */
static int qos_calc_ppo(struct obd_device *obd)
{
struct lov_obd *lov = &obd->u.lov;
int rc, i, prio_wide;
ENTRY;
- if (!lov->lov_qos.lq_dirty)
+ if (!lov->lov_qos.lq_dirty)
GOTO(out, rc = 0);
- num_active = lov->desc.ld_active_tgt_count - 1;
+ num_active = lov->desc.ld_active_tgt_count - 1;
if (num_active < 1)
GOTO(out, rc = -EAGAIN);
continue;
ba_min = min(temp, ba_min);
ba_max = max(temp, ba_max);
-
+
/* Count the number of usable OSS's */
if (lov->lov_tgts[i]->ltd_qos.ltq_oss->lqo_bavail == 0)
lov->lov_qos.lq_active_oss_count++;
/* per-OST penalty is prio * TGT_bavail / (num_ost - 1) / 2 */
temp >>= 1;
do_div(temp, num_active);
- lov->lov_tgts[i]->ltd_qos.ltq_penalty_per_obj =
+ lov->lov_tgts[i]->ltd_qos.ltq_penalty_per_obj =
(temp * prio_wide) >> 8;
- if (lov->lov_qos.lq_reset == 0)
+ if (lov->lov_qos.lq_reset == 0)
lov->lov_tgts[i]->ltd_qos.ltq_penalty = 0;
}
/* 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 */
list_for_each_entry(oss, &lov->lov_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;
- if (lov->lov_qos.lq_reset == 0)
+ if (lov->lov_qos.lq_reset == 0)
oss->lqo_penalty = 0;
}
lov->lov_qos.lq_dirty = 0;
lov->lov_qos.lq_reset = 0;
- /* If each ost has almost same free space,
+ /* If each ost has almost same free space,
* do rr allocation for better creation performance */
lov->lov_qos.lq_same_space = 0;
temp = ba_max - ba_min;
- ba_min = (ba_min * 51) >> 8; /* 51/256 = .20 */
+ ba_min = (ba_min * 51) >> 8; /* 51/256 = .20 */
if (temp < ba_min) {
- /* Difference is less than 20% */
+ /* 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;
/* Final ost weight = TGT_BAVAIL - ost_penalty - oss_penalty */
temp = TGT_BAVAIL(i);
- temp2 = lov->lov_tgts[i]->ltd_qos.ltq_penalty +
+ temp2 = lov->lov_tgts[i]->ltd_qos.ltq_penalty +
lov->lov_tgts[i]->ltd_qos.ltq_oss->lqo_penalty;
- if (temp < temp2)
+ if (temp < temp2)
lov->lov_tgts[i]->ltd_qos.ltq_weight = 0;
else
lov->lov_tgts[i]->ltd_qos.ltq_weight = temp - temp2;
lov->lov_tgts[index]->ltd_qos.ltq_penalty +=
lov->lov_tgts[index]->ltd_qos.ltq_penalty_per_obj *
lov->desc.ld_active_tgt_count;
- oss->lqo_penalty += oss->lqo_penalty_per_obj *
+ oss->lqo_penalty += oss->lqo_penalty_per_obj *
lov->lov_qos.lq_active_oss_count;
/* Decrease all OSS penalties */
list_for_each_entry(oss, &lov->lov_qos.lq_oss_list, lqo_oss_list) {
- if (oss->lqo_penalty < oss->lqo_penalty_per_obj)
+ 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 (i = 0; i < lov->desc.ld_tgt_count; i++) {
- if (!lov->lov_tgts[i] || !lov->lov_tgts[i]->ltd_active)
+ if (!lov->lov_tgts[i] || !lov->lov_tgts[i]->ltd_active)
continue;
if (lov->lov_tgts[i]->ltd_qos.ltq_penalty <
lov->lov_tgts[i]->ltd_qos.ltq_penalty_per_obj)
else
lov->lov_tgts[i]->ltd_qos.ltq_penalty -=
lov->lov_tgts[i]->ltd_qos.ltq_penalty_per_obj;
-
+
qos_calc_weight(lov, i);
/* Recalc the total weight of usable osts */
CDEBUG(D_QOS, "recalc tgt %d avail="LPU64
" ostppo="LPU64" ostp="LPU64" ossppo="LPU64
" ossp="LPU64" wt="LPU64"\n",
- i, TGT_BAVAIL(i) >> 10,
+ i, 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_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);
#endif
RETURN(0);
}
- /* Do actuall allocation. */
+ /* 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)
+ 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 *
+ 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) {
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] &&
- lov->lov_tgts[i]->ltd_qos.ltq_oss == oss) {
+ if (lov->lov_tgts[i] &&
+ lov->lov_tgts[i]->ltd_qos.ltq_oss == oss) {
/* Evenly space these OSTs across arrayspace */
int next = j * ost_count / oss->lqo_ost_count;
LASSERT(next < ost_count);
lov->lov_qos.lq_rr_array[i]);
}
#endif
-
+
RETURN(0);
}
}
}
- CWARN("using fewer stripes for object "LPX64": old %u new %u\n",
+ CWARN("using fewer stripes for object "LPU64": old %u new %u\n",
lsm->lsm_object_id, lsm->lsm_stripe_count, set->set_count);
LASSERT(lsm->lsm_stripe_count >= set->set_count);
ost_idx = (req->rq_idx + lsm->lsm_stripe_count) % ost_count;
for (i = 0; i < ost_count; i++, ost_idx = (ost_idx + 1) % ost_count) {
- if (!lov->lov_tgts[ost_idx] ||
- !lov->lov_tgts[ost_idx]->ltd_active)
+ if (!lov->lov_tgts[ost_idx] ||
+ !lov->lov_tgts[ost_idx]->ltd_active)
continue;
/* check if objects has been created on this ost */
for (stripe = 0; stripe < lsm->lsm_stripe_count; stripe++) {
} else if (stripe_cnt_min >= ost_active_count ||
lov->lov_start_idx > ost_count) {
/* If we have allocated from all of the OSTs, slowly
- precess the next start */
+ * precess the next start if the OST/stripe count isn't
+ * already doing this for us. */
lov->lov_start_idx %= ost_count;
if (*stripe_cnt > 1 && (ost_active_count % (*stripe_cnt)) != 1)
++lov->lov_offset_idx;
array_idx = (lov->lov_start_idx + lov->lov_offset_idx) % ost_count;
idx_pos = idx_arr;
#ifdef QOS_DEBUG
- CDEBUG(D_QOS, "want %d startidx %d startcnt %d offset %d arrayidx %d\n",
- stripe_cnt_min, lov->lov_start_idx, lov->lov_start_count,
- lov->lov_offset_idx, array_idx);
+ CDEBUG(D_QOS, "want %d startidx %d startcnt %d offset %d active %d "
+ "count %d arrayidx %d\n",
+ stripe_cnt, lov->lov_start_idx, lov->lov_start_count,
+ lov->lov_offset_idx, ost_active_count, ost_count, array_idx);
#endif
for (i = 0; i < ost_count; i++, array_idx=(array_idx + 1) % ost_count) {
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)
ENTRY;
repeat_find:
- idx_pos = idx_arr;
ost_idx = lsm->lsm_oinfo[0]->loi_ost_idx;
+ idx_pos = idx_arr;
for (i = 0; i < ost_count; i++, ost_idx = (ost_idx + 1) % ost_count) {
- if (!lov->lov_tgts[ost_idx] ||
+ if (!lov->lov_tgts[ost_idx] ||
!lov->lov_tgts[ost_idx]->ltd_active) {
continue;
}
if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_PRECREATE) && ost_idx == 0)
continue;
- /* Drop slow OSCs if we can */
+ /* Drop slow OSCs if we can, but not for requested start idx */
if ((obd_precreate(lov->lov_tgts[ost_idx]->ltd_exp) > speed) &&
(i != 0 || speed < 2))
continue;
ENTRY;
if (stripe_cnt_min < 1)
- GOTO(out, rc = -EINVAL);
+ GOTO(out_nolock, rc = -EINVAL);
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, rc = -EAGAIN);
-
- /* Do actuall allocation, use write lock here. */
+ GOTO(out_nolock, rc = -EAGAIN);
+
+ /* Do actual allocation, use write lock here. */
down_write(&lov->lov_qos.lq_rw_sem);
- /*
+ /*
* 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) {
- up_write(&lov->lov_qos.lq_rw_sem);
+ if (!lov->lov_qos.lq_dirty && lov->lov_qos.lq_same_space)
GOTO(out, rc = -EAGAIN);
- }
+
ost_count = lov->desc.ld_tgt_count;
- if (lov->desc.ld_active_tgt_count < 2)
- GOTO(out_up_write, rc = -EAGAIN);
+ if (lov->desc.ld_active_tgt_count < 2)
+ GOTO(out, rc = -EAGAIN);
rc = qos_calc_ppo(exp->exp_obd);
if (rc)
- GOTO(out_up_write, rc);
+ GOTO(out, rc);
total_bavail = 0;
good_osts = 0;
bavail = TGT_BAVAIL(i);
if (!bavail) {
if (warn) {
- CDEBUG(D_QOS, "no free space on %s\n",
- obd_uuid2str(&lov->lov_tgts[i]->ltd_uuid));
+ CDEBUG(D_QOS, "no free space on %s\n",
+ obd_uuid2str(&lov->lov_tgts[i]->ltd_uuid));
last_warn = now;
}
continue;
}
if (!TGT_FFREE(i)) {
if (warn) {
- CDEBUG(D_QOS, "no free inodes on %s\n",
- obd_uuid2str(&lov->lov_tgts[i]->ltd_uuid));
+ CDEBUG(D_QOS, "no free inodes on %s\n",
+ obd_uuid2str(&lov->lov_tgts[i]->ltd_uuid));
last_warn = now;
}
continue;
}
if (good_osts < stripe_cnt_min)
- GOTO(out_up_write, rc = -EAGAIN);
+ GOTO(out, rc = -EAGAIN);
if (!total_bavail)
- GOTO(out_up_write, rc = -ENOSPC);
+ GOTO(out, rc = -ENOSPC);
/* We have enough osts */
if (good_osts < *stripe_cnt)
*stripe_cnt = good_osts;
if (!*stripe_cnt)
- GOTO(out_up_write, rc = -EAGAIN);
+ GOTO(out, rc = -EAGAIN);
/* Find enough OSTs with weighted random allocation. */
nfound = 0;
while (nfound < *stripe_cnt) {
- __u64 rand, cur_weight = 0;
+ __u64 rand, cur_weight;
+ cur_weight = 0;
rc = -ENODEV;
if (total_weight) {
rand |= ll_rand() % (unsigned)total_weight;
else
rand |= ll_rand();
+
#else
rand = ((__u64)ll_rand() << 32 | ll_rand()) %
total_weight;
}
}
LASSERT(nfound == *stripe_cnt);
-
-out_up_write:
- up_write(&lov->lov_qos.lq_rw_sem);
-
+
out:
+ up_write(&lov->lov_qos.lq_rw_sem);
+
+out_nolock:
if (rc == -EAGAIN)
rc = alloc_rr(lov, idx_arr, stripe_cnt, 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,
+static int alloc_idx_array(struct obd_export *exp, struct lov_stripe_md *lsm,
int newea, int **idx_arr, int *arr_cnt, int flags)
{
struct lov_obd *lov = &exp->exp_obd->u.lov;
for (i = 0; i < *arr_cnt; i++)
tmp_arr[i] = -1;
- if (newea ||
- lsm->lsm_oinfo[0]->loi_ost_idx >= lov->desc.ld_tgt_count)
+ if (newea ||
+ lsm->lsm_oinfo[0]->loi_ost_idx >= lov->desc.ld_tgt_count)
rc = alloc_qos(exp, tmp_arr, &stripe_cnt, 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 = NULL, 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);
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.
+ * enough OSTs to allow the file to be created at that size.
* This may mean we use more than the default # of stripes. */
if (src_oa->o_valid & OBD_MD_FLSIZE) {
obd_size min_bavail = LUSTRE_STRIPE_MAXBYTES;
-
- /* Find a small number of stripes we can use
+
+ /* 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] ||
+ if (!lov->lov_tgts[i] ||
!lov->lov_tgts[i]->ltd_active)
continue;
min_bavail = min(min_bavail, TGT_BAVAIL(i));
stripes = stripes_def;
}
- rc = lov_alloc_memmd(&set->set_oi->oi_md, stripes,
+ rc = lov_alloc_memmd(&set->set_oi->oi_md, stripes,
lov->desc.ld_pattern ?
lov->desc.ld_pattern : LOV_PATTERN_RAID0,
LOV_MAGIC);
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;
int ost_idx = idx_arr[i];
* stripe which holds the existing file size.
*/
if (src_oa->o_valid & OBD_MD_FLSIZE) {
- req->rq_oi.oi_oa->o_size =
+ req->rq_oi.oi_oa->o_size =
lov_size_to_stripe(lsm, src_oa->o_size, i);
CDEBUG(D_INODE, "stripe %d has size "LPU64"/"LPU64"\n",
git://git.whamcloud.com / fs / lustre-release.git / commitdiff