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;
we have to double the OST penalty */
num_active = 1;
for (i = 0; i < lov->desc.ld_tgt_count; i++)
- lov->lov_tgts[i]->ltd_qos.ltq_penalty_per_obj <<= 1;
+ if (lov->lov_tgts[i])
+ lov->lov_tgts[i]->ltd_qos.ltq_penalty_per_obj <<= 1;
}
/* Per-OSS penalty is prio * oss_avail / oss_osts / (num_oss - 1) / 2 */
static int qos_calc_rr(struct lov_obd *lov)
{
struct lov_qos_oss *oss;
- unsigned ost_count, placed;
+ unsigned ost_count, placed, real_count;
int i;
ENTRY;
RETURN(-ENOMEM);
}
- for (i = 0; i < ost_count; i++)
+ real_count = 0;
+ for (i = 0; i < ost_count; i++) {
lov->lov_qos.lq_rr_array[i] = LOV_QOS_EMPTY;
+ if (lov->lov_tgts[i])
+ real_count++;
+ }
/* 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) {
+ 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;
while (lov->lov_qos.lq_rr_array[next] !=
lov->lov_qos.lq_rr_array[next] = i;
j++;
placed++;
- }
+ }
}
LASSERT(j == oss->lqo_ost_count);
}
lov->lov_qos.lq_dirty_rr = 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);
+ 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 < ost_count; i++) {
LCONSOLE(D_WARNING, "rr #%d ost idx=%d\n", i,
lov->lov_qos.lq_rr_array[i]);
lov->lov_qos.lq_dirty_rr = 1;
RETURN(-EAGAIN);
}
-
+
#ifdef QOS_DEBUG
for (i = 0; i < ost_count; i++) {
- LCONSOLE(D_QOS, "rr #%d ost idx=%d\n", i,
+ LCONSOLE(D_QOS, "rr #%d ost idx=%d\n", i,
lov->lov_qos.lq_rr_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,
+ 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;
+ int i, *idx_pos;
__u32 ost_idx;
+ int ost_start_idx_temp;
+ int speed = 0;
+ int stripe_cnt_min = min_stripe_count(*stripe_cnt, flags);
ENTRY;
i = qos_calc_rr(lov);
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 ||
+ } 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;
- ++lov->lov_offset_idx;
+ if (*stripe_cnt > 1 && (ost_active_count % (*stripe_cnt)) != 1)
+ ++lov->lov_offset_idx;
}
+ down_read(&lov->lov_qos.lq_rw_sem);
+ ost_start_idx_temp = lov->lov_start_idx;
+
+repeat_find :
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, 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
- 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];
#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] ?
+ ((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++;
+ lov->lov_start_idx = ost_start_idx_temp;
+ goto repeat_find;
+ }
+
up_read(&lov->lov_qos.lq_rw_sem);
*stripe_cnt = idx_pos - idx_arr;
int *idx_arr)
{
unsigned ost_idx, ost_count = lov->desc.ld_tgt_count;
- int i, *idx_pos = idx_arr;
+ int i, *idx_pos;
+ int speed = 0;
ENTRY;
+repeat_find:
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] ||
!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 */
+ 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);
}
+ 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).
- 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,
+ int flags)
{
struct lov_obd *lov = &exp->exp_obd->u.lov;
static time_t last_warn = 0;
__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);
ENTRY;
+ if (stripe_cnt_min < 1)
+ GOTO(out, rc = -EINVAL);
+
lov_getref(exp->exp_obd);
down_write(&lov->lov_qos.lq_rw_sem);
continue;
}
+ /* Fail Check before osc_precreate() is called
+ so we can only 'fail' single OSC. */
+ if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_PRECREATE) && i == 0)
+ continue;
+
+ if (obd_precreate(lov->lov_tgts[i]->ltd_exp) > 2)
+ continue;
+
lov->lov_tgts[i]->ltd_qos.ltq_usable = 1;
qos_calc_weight(lov, i);
total_bavail += bavail;
good_osts++;
}
+ 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;
- if (!*stripe_cnt)
- GOTO(out, rc = -EAGAIN);
-
/* Find enough OSTs with weighted random allocation. */
nfound = 0;
while (nfound < *stripe_cnt) {
/* 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)
+ if (!lov->lov_tgts[i] ||
+ !lov->lov_tgts[i]->ltd_qos.ltq_usable)
continue;
+
cur_weight += lov->lov_tgts[i]->ltd_qos.ltq_weight;
if (cur_weight >= rand) {
#ifdef QOS_DEBUG
up_write(&lov->lov_qos.lq_rw_sem);
if (rc == -EAGAIN)
- rc = alloc_rr(lov, idx_arr, stripe_cnt);
+ 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,
- 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, 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);
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.
}
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_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);