#include <obd_class.h>
#include <obd_lov.h>
+#include <lustre/lustre_idl.h>
#include "lov_internal.h"
/* #define QOS_DEBUG 1 */
#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)
/* 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 */
- if (temp < ba_min) {
- /* Difference is less than 20% */
+ if ((ba_max * (256 - lov->lov_qos.lq_threshold_rr)) >> 8 < ba_min) {
lov->lov_qos.lq_same_space = 1;
/* Reset weights for the next time we enter qos mode */
lov->lov_qos.lq_reset = 1;
continue;
/* check if objects has been created on this ost */
for (stripe = 0; stripe < lsm->lsm_stripe_count; stripe++) {
+ /* we try send create to this ost but he is failed */
if (stripe == req->rq_stripe)
continue;
+ /* already have object at this stripe */
if (ost_idx == lsm->lsm_oinfo[stripe]->loi_ost_idx)
break;
}
-
if (stripe >= lsm->lsm_stripe_count) {
req->rq_idx = ost_idx;
rc = obd_create(lov->lov_tgts[ost_idx]->ltd_exp,
if (pool != NULL) {
up_read(&pool_tgt_rw_sem(pool));
/* put back ref got by lov_find_pool() */
- lh_put(lov->lov_pools_hash_body, &pool->pool_hash);
+ lov_pool_putref(pool);
}
RETURN(rc);
if (pool != NULL) {
up_read(&pool_tgt_rw_sem(pool));
/* put back ref got by lov_find_pool() */
- lh_put(lov->lov_pools_hash_body, &pool->pool_hash);
+ lov_pool_putref(pool);
}
RETURN(rc);
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;
- int nfound, good_osts, i, warn = 0, rc = 0;
+ __u64 total_weight = 0;
+ int nfound, good_osts, i, rc = 0;
int stripe_cnt_min = min_stripe_count(*stripe_cnt, flags);
struct pool_desc *pool;
struct ost_pool *osts;
lqr = &(pool->pool_rr);
}
- lov_getref(exp->exp_obd);
+ obd_getref(exp->exp_obd);
+
+ /* wait for fresh statfs info if needed, the rpcs are sent in
+ * lov_create() */
+ qos_statfs_update(exp->exp_obd,
+ cfs_time_shift_64(-2 * lov->desc.ld_qos_maxage), 1);
/* Detect -EAGAIN early, before expensive lock is taken. */
if (!lov->lov_qos.lq_dirty && lov->lov_qos.lq_same_space)
if (rc)
GOTO(out, rc);
- total_bavail = 0;
good_osts = 0;
- /* Warn users about zero available space/inode every 30 min */
- if (cfs_time_sub(now, last_warn) > 60 * 30)
- warn = 1;
/* Find all the OSTs that are valid stripe candidates */
for (i = 0; i < osts->op_count; i++) {
- __u64 bavail;
-
if (!lov->lov_tgts[osts->op_array[i]] ||
!lov->lov_tgts[osts->op_array[i]]->ltd_active)
continue;
- 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[osts->op_array[i]]->ltd_uuid));
- last_warn = now;
- }
- continue;
- }
- if (!TGT_FFREE(osts->op_array[i])) {
- if (warn) {
- CDEBUG(D_QOS, "no free inodes on %s\n",
- obd_uuid2str(&lov->lov_tgts[osts->op_array[i]]->ltd_uuid));
- last_warn = now;
- }
- continue;
- }
/* Fail Check before osc_precreate() is called
so we can only 'fail' single OSC. */
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[osts->op_array[i]]->ltd_qos.ltq_weight;
good_osts++;
if (good_osts < stripe_cnt_min)
GOTO(out, rc = -EAGAIN);
- if (!total_bavail)
- GOTO(out, rc = -ENOSPC);
-
/* We have enough osts */
if (good_osts < *stripe_cnt)
*stripe_cnt = good_osts;
if (pool != NULL) {
up_read(&pool_tgt_rw_sem(pool));
/* put back ref got by lov_find_pool() */
- lh_put(lov->lov_pools_hash_body, &pool->pool_hash);
+ lov_pool_putref(pool);
}
if (rc == -EAGAIN)
rc = alloc_rr(lov, idx_arr, stripe_cnt, poolname, flags);
- lov_putref(exp->exp_obd);
+ obd_putref(exp->exp_obd);
RETURN(rc);
}
req->rq_stripe = i;
/* create data objects with "parent" OA */
memcpy(req->rq_oi.oi_oa, src_oa, sizeof(*req->rq_oi.oi_oa));
+ req->rq_oi.oi_cb_up = cb_create_update;
/* XXX When we start creating objects on demand, we need to
* make sure that we always create the object on the
if (stripes < lsm->lsm_stripe_count)
qos_shrink_lsm(set);
+ if (OBD_FAIL_CHECK(OBD_FAIL_MDS_LOV_PREP_CREATE)) {
+ qos_shrink_lsm(set);
+ rc = -EIO;
+ }
if (oti && (src_oa->o_valid & OBD_MD_FLCOOKIE)) {
oti_alloc_cookies(oti, set->set_count);
ENTRY;
lov->lov_qos.lq_dirty = 1;
}
+
+void qos_statfs_done(struct lov_obd *lov)
+{
+ LASSERT(lov->lov_qos.lq_statfs_in_progress);
+ down_write(&lov->lov_qos.lq_rw_sem);
+ lov->lov_qos.lq_statfs_in_progress = 0;
+ /* wake up any threads waiting for the statfs rpcs to complete */
+ cfs_waitq_signal(&lov->lov_qos.lq_statfs_waitq);
+ up_write(&lov->lov_qos.lq_rw_sem);
+}
+
+static int qos_statfs_ready(struct obd_device *obd, __u64 max_age)
+{
+ struct lov_obd *lov = &obd->u.lov;
+ int rc;
+ ENTRY;
+ down_read(&lov->lov_qos.lq_rw_sem);
+ rc = lov->lov_qos.lq_statfs_in_progress == 0 ||
+ cfs_time_beforeq_64(max_age, obd->obd_osfs_age);
+ up_read(&lov->lov_qos.lq_rw_sem);
+ RETURN(rc);
+}
+
+/*
+ * Update statfs data if the current osfs age is older than max_age.
+ * If wait is not set, it means that we are called from lov_create()
+ * and we should just issue the rpcs without waiting for them to complete.
+ * If wait is set, we are called from alloc_qos() and we just have
+ * to wait for the request set to complete.
+ */
+void qos_statfs_update(struct obd_device *obd, __u64 max_age, int wait)
+{
+ struct lov_obd *lov = &obd->u.lov;
+ struct obd_info *oinfo;
+ int rc = 0;
+ struct ptlrpc_request_set *set = NULL;
+ ENTRY;
+
+ if (cfs_time_beforeq_64(max_age, obd->obd_osfs_age))
+ /* statfs data are quite recent, don't need to refresh it */
+ RETURN_EXIT;
+
+ if (!wait && lov->lov_qos.lq_statfs_in_progress)
+ /* statfs already in progress */
+ RETURN_EXIT;
+
+ down_write(&lov->lov_qos.lq_rw_sem);
+ if (lov->lov_qos.lq_statfs_in_progress) {
+ up_write(&lov->lov_qos.lq_rw_sem);
+ GOTO(out, rc = 0);
+ }
+ /* no statfs in flight, send rpcs */
+ lov->lov_qos.lq_statfs_in_progress = 1;
+ up_write(&lov->lov_qos.lq_rw_sem);
+
+ if (wait)
+ CDEBUG(D_QOS, "%s: did not manage to get fresh statfs data "
+ "in a timely manner (osfs age "LPU64", max age "LPU64")"
+ ", sending new statfs rpcs\n",
+ obd_uuid2str(&lov->desc.ld_uuid), obd->obd_osfs_age,
+ max_age);
+
+ /* need to send statfs rpcs */
+ CDEBUG(D_QOS, "sending new statfs requests\n");
+ memset(lov->lov_qos.lq_statfs_data, 0,
+ sizeof(*lov->lov_qos.lq_statfs_data));
+ oinfo = &lov->lov_qos.lq_statfs_data->lsd_oi;
+ oinfo->oi_osfs = &lov->lov_qos.lq_statfs_data->lsd_statfs;
+ oinfo->oi_flags = OBD_STATFS_NODELAY;
+ set = ptlrpc_prep_set();
+ if (!set)
+ GOTO(out_failed, rc = -ENOMEM);
+
+ rc = obd_statfs_async(obd, oinfo, max_age, set);
+ if (rc || list_empty(&set->set_requests)) {
+ if (rc)
+ CWARN("statfs failed with %d\n", rc);
+ GOTO(out_failed, rc);
+ }
+ /* send requests via ptlrpcd */
+ oinfo->oi_flags |= OBD_STATFS_PTLRPCD;
+ ptlrpcd_add_rqset(set);
+ GOTO(out, rc);
+
+out_failed:
+ down_write(&lov->lov_qos.lq_rw_sem);
+ lov->lov_qos.lq_statfs_in_progress = 0;
+ /* wake up any threads waiting for the statfs rpcs to complete */
+ cfs_waitq_signal(&lov->lov_qos.lq_statfs_waitq);
+ up_write(&lov->lov_qos.lq_rw_sem);
+ wait = 0;
+out:
+ if (set)
+ ptlrpc_set_destroy(set);
+ if (wait) {
+ struct l_wait_info lwi = { 0 };
+ CDEBUG(D_QOS, "waiting for statfs requests to complete\n");
+ l_wait_event(lov->lov_qos.lq_statfs_waitq,
+ qos_statfs_ready(obd, max_age), &lwi);
+ if (cfs_time_before_64(obd->obd_osfs_age, max_age))
+ CDEBUG(D_QOS, "%s: still no fresh statfs data after "
+ "waiting (osfs age "LPU64", max age "
+ LPU64")\n",
+ obd_uuid2str(&lov->desc.ld_uuid),
+ obd->obd_osfs_age, max_age);
+ }
+}