* GPL HEADER END
*/
/*
- * Copyright (c) 2012 Intel, Inc.
+ * Copyright (c) 2012, 2013, Intel Corporation.
* Use is subject to license terms.
*
* Author: Johann Lombardi <johann.lombardi@intel.com>
* Author: Niu Yawei <yawei.niu@intel.com>
*/
-#ifndef EXPORT_SYMTAB
-# define EXPORT_SYMTAB
-#endif
-
#define DEBUG_SUBSYSTEM S_LQUOTA
#include "qmt_internal.h"
LASSERT(lqe_is_master(lqe));
lqe->lqe_revoke_time = 0;
- cfs_init_rwsem(&lqe->lqe_sem);
+ init_rwsem(&lqe->lqe_sem);
}
/*
struct qmt_pool_info *pool = (struct qmt_pool_info *)arg;
libcfs_debug_vmsg2(msgdata, fmt, args,
- "qmt:%s pool:%d-%s id:"LPU64" enforced:%d hard:"LPU64
- " soft:"LPU64" granted:"LPU64" time:"LPU64" qunit:"
- LPU64" edquot:%d revoke:"LPU64"\n",
+ "qmt:%s pool:%d-%s id:%llu enforced:%d hard:%llu"
+ " soft:%llu granted:%llu time:%llu qunit:"
+ "%llu edquot:%d may_rel:%llu revoke:%llu\n",
pool->qpi_qmt->qmt_svname,
pool->qpi_key & 0x0000ffff,
RES_NAME(pool->qpi_key >> 16),
lqe->lqe_id.qid_uid, lqe->lqe_enforced,
lqe->lqe_hardlimit, lqe->lqe_softlimit,
lqe->lqe_granted, lqe->lqe_gracetime,
- lqe->lqe_qunit, lqe->lqe_edquot,
+ lqe->lqe_qunit, lqe->lqe_edquot, lqe->lqe_may_rel,
lqe->lqe_revoke_time);
}
LQUOTA_DEBUG(lqe, "write glb");
- if (!lqe->lqe_enforced && lqe->lqe_granted == 0 &&
- lqe->lqe_id.qid_uid != 0) {
- /* quota isn't enforced any more for this entry and there is no
- * more space granted to slaves, let's just remove the entry
- * from the index */
- rec = NULL;
- } else {
- rec = &qti->qti_glb_rec;
+ /* never delete the entry even when the id isn't enforced and
+ * no any guota granted, otherwise, this entry will not be
+ * synced to slave during the reintegration. */
+ rec = &qti->qti_glb_rec;
- /* fill global index with updated quota settings */
- rec->qbr_granted = lqe->lqe_granted;
- rec->qbr_hardlimit = lqe->lqe_hardlimit;
- rec->qbr_softlimit = lqe->lqe_softlimit;
- rec->qbr_time = lqe->lqe_gracetime;
- }
+ /* fill global index with updated quota settings */
+ rec->qbr_granted = lqe->lqe_granted;
+ rec->qbr_hardlimit = lqe->lqe_hardlimit;
+ rec->qbr_softlimit = lqe->lqe_softlimit;
+ rec->qbr_time = lqe->lqe_gracetime;
/* write new quota settings */
rc = lquota_disk_write(env, th, LQE_GLB_OBJ(lqe), &lqe->lqe_id,
RETURN(rc);
}
- LQUOTA_DEBUG(lqe, "successful slv read "LPU64, *granted);
+ LQUOTA_DEBUG(lqe, "successful slv read %llu", *granted);
RETURN(0);
}
LASSERT(lqe_is_master(lqe));
LASSERT(lqe_is_locked(lqe));
- LQUOTA_DEBUG(lqe, "write slv "DFID" granted:"LPU64,
+ LQUOTA_DEBUG(lqe, "write slv "DFID" granted:%llu",
PFID(lu_object_fid(&slv_obj->do_lu)), granted);
- if (granted == 0) {
- /* this slave does not own any quota space for this ID any more,
- * so let's just remove the entry from the index */
- rec = NULL;
- } else {
- rec = &qti->qti_slv_rec;
+ /* never delete the entry, otherwise, it'll not be transferred
+ * to slave during reintegration. */
+ rec = &qti->qti_slv_rec;
- /* updated space granted to this slave */
- rec->qsr_granted = granted;
- }
+ /* updated space granted to this slave */
+ rec->qsr_granted = granted;
/* write new granted space */
rc = lquota_disk_write(env, th, slv_obj, &lqe->lqe_id,
(struct dt_rec *)rec, flags, ver);
if (rc) {
LQUOTA_ERROR(lqe, "failed to update slave index "DFID" granted:"
- LPU64, PFID(lu_object_fid(&slv_obj->do_lu)),
+ "%llu", PFID(lu_object_fid(&slv_obj->do_lu)),
granted);
RETURN(rc);
}
RETURN(-EINVAL);
RETURN(0);
}
+
+/*
+ * Set/clear edquot flag after quota space allocation/release or settings
+ * change. Slaves will be notified of changes via glimpse on per-ID lock
+ *
+ * \param lqe - is the quota entry to check
+ * \param now - is the current time in second used for grace time managment
+ */
+void qmt_adjust_edquot(struct lquota_entry *lqe, __u64 now)
+{
+ struct qmt_pool_info *pool = lqe2qpi(lqe);
+ ENTRY;
+
+ if (!lqe->lqe_enforced || lqe->lqe_id.qid_uid == 0)
+ RETURN_EXIT;
+
+ if (!lqe->lqe_edquot) {
+ /* space exhausted flag not set, let's check whether it is time
+ * to set the flag */
+
+ if (!qmt_space_exhausted(lqe, now))
+ /* the qmt still has available space */
+ RETURN_EXIT;
+
+ /* See comment in qmt_adjust_qunit(). LU-4139 */
+ if (qmt_hard_exhausted(lqe) ||
+ pool->qpi_key >> 16 != LQUOTA_RES_DT) {
+ /* we haven't reached the minimal qunit yet so there is
+ * still hope that the rebalancing process might free
+ * up some quota space */
+ if (lqe->lqe_qunit != pool->qpi_least_qunit)
+ RETURN_EXIT;
+
+ /* least qunit value not sent to all slaves yet */
+ if (lqe->lqe_revoke_time == 0)
+ RETURN_EXIT;
+
+ /* Let's give more time to slave to release space */
+ if (lqe->lqe_may_rel != 0 &&
+ cfs_time_before_64(cfs_time_shift_64(
+ -QMT_REBA_TIMEOUT),
+ lqe->lqe_revoke_time))
+ RETURN_EXIT;
+ } else {
+ if (lqe->lqe_qunit > pool->qpi_soft_least_qunit)
+ RETURN_EXIT;
+ }
+
+ /* set edquot flag */
+ lqe->lqe_edquot = true;
+ } else {
+ /* space exhausted flag set, let's check whether it is time to
+ * clear it */
+
+ if (qmt_space_exhausted(lqe, now))
+ /* the qmt still has not space */
+ RETURN_EXIT;
+
+ if (lqe->lqe_hardlimit != 0 &&
+ lqe->lqe_granted + pool->qpi_least_qunit >
+ lqe->lqe_hardlimit)
+ /* we clear the flag only once at least one least qunit
+ * is available */
+ RETURN_EXIT;
+
+ /* clear edquot flag */
+ lqe->lqe_edquot = false;
+ }
+
+ LQUOTA_DEBUG(lqe, "changing edquot flag");
+
+ /* let's notify slave by issuing glimpse on per-ID lock.
+ * the rebalance thread will take care of this */
+ qmt_id_lock_notify(pool->qpi_qmt, lqe);
+ EXIT;
+}
+
+/* Using least_qunit when over block softlimit will seriously impact the
+ * write performance, we need to do some special tweaking on that. */
+static __u64 qmt_calc_softlimit(struct lquota_entry *lqe, bool *oversoft)
+{
+ struct qmt_pool_info *pool = lqe2qpi(lqe);
+
+ LASSERT(lqe->lqe_softlimit != 0);
+ *oversoft = false;
+ /* No need to do special tweaking for inode limit */
+ if (pool->qpi_key >> 16 != LQUOTA_RES_DT)
+ return lqe->lqe_softlimit;
+
+ if (lqe->lqe_granted <= lqe->lqe_softlimit +
+ pool->qpi_soft_least_qunit) {
+ return lqe->lqe_softlimit;
+ } else if (lqe->lqe_hardlimit != 0) {
+ *oversoft = true;
+ return lqe->lqe_hardlimit;
+ } else {
+ *oversoft = true;
+ return 0;
+ }
+}
+
+/*
+ * Try to grant more quota space back to slave.
+ *
+ * \param lqe - is the quota entry for which we would like to allocate more
+ * space
+ * \param granted - is how much was already granted as part of the request
+ * processing
+ * \param spare - is how much unused quota space the slave already owns
+ *
+ * \retval return how additional space can be granted to the slave
+ */
+__u64 qmt_alloc_expand(struct lquota_entry *lqe, __u64 granted, __u64 spare)
+{
+ struct qmt_pool_info *pool = lqe2qpi(lqe);
+ __u64 remaining, qunit;
+ int slv_cnt;
+
+ LASSERT(lqe->lqe_enforced && lqe->lqe_qunit != 0);
+
+ slv_cnt = lqe2qpi(lqe)->qpi_slv_nr[lqe->lqe_site->lqs_qtype];
+ qunit = lqe->lqe_qunit;
+
+ /* See comment in qmt_adjust_qunit(). LU-4139. */
+ if (lqe->lqe_softlimit != 0) {
+ bool oversoft;
+ remaining = qmt_calc_softlimit(lqe, &oversoft);
+ if (remaining == 0)
+ remaining = lqe->lqe_granted +
+ pool->qpi_soft_least_qunit;
+ } else {
+ remaining = lqe->lqe_hardlimit;
+ }
+
+ if (lqe->lqe_granted >= remaining)
+ RETURN(0);
+
+ remaining -= lqe->lqe_granted;
+
+ do {
+ if (spare >= qunit)
+ break;
+
+ granted &= (qunit - 1);
+
+ if (remaining > (slv_cnt * qunit) >> 1) {
+ /* enough room to grant more space w/o additional
+ * shrinking ... at least for now */
+ remaining -= (slv_cnt * qunit) >> 1;
+ } else if (qunit != pool->qpi_least_qunit) {
+ qunit >>= 2;
+ continue;
+ }
+
+ granted &= (qunit - 1);
+ if (spare > 0)
+ RETURN(min_t(__u64, qunit - spare, remaining));
+ else
+ RETURN(min_t(__u64, qunit - granted, remaining));
+ } while (qunit >= pool->qpi_least_qunit);
+
+ RETURN(0);
+}
+
+/*
+ * Adjust qunit size according to quota limits and total granted count.
+ * The caller must have locked the lqe.
+ *
+ * \param env - the environment passed by the caller
+ * \param lqe - is the qid entry to be adjusted
+ */
+void qmt_adjust_qunit(const struct lu_env *env, struct lquota_entry *lqe)
+{
+ struct qmt_pool_info *pool = lqe2qpi(lqe);
+ int slv_cnt;
+ __u64 qunit, limit, qunit2 = 0;
+ ENTRY;
+
+ LASSERT(lqe_is_locked(lqe));
+
+ if (!lqe->lqe_enforced || lqe->lqe_id.qid_uid == 0)
+ /* no quota limits */
+ RETURN_EXIT;
+
+ /* record how many slaves have already registered */
+ slv_cnt = pool->qpi_slv_nr[lqe->lqe_site->lqs_qtype];
+ if (slv_cnt == 0)
+ /* wait for at least one slave to join */
+ RETURN_EXIT;
+
+ /* Qunit calculation is based on soft limit, if any, hard limit
+ * otherwise. This means that qunit is shrunk to the minimum when
+ * beyond the soft limit. This will impact performance, but that's the
+ * price of an accurate grace time management. */
+ if (lqe->lqe_softlimit != 0) {
+ bool oversoft;
+ /* As a compromise of write performance and the grace time
+ * accuracy, the block qunit size will be shrunk to
+ * qpi_soft_least_qunit when over softlimit. LU-4139. */
+ limit = qmt_calc_softlimit(lqe, &oversoft);
+ if (oversoft)
+ qunit2 = pool->qpi_soft_least_qunit;
+ if (limit == 0)
+ GOTO(done, qunit = qunit2);
+ } else if (lqe->lqe_hardlimit != 0) {
+ limit = lqe->lqe_hardlimit;
+ } else {
+ LQUOTA_ERROR(lqe, "enforced bit set, but neither hard nor soft "
+ "limit are set");
+ RETURN_EXIT;
+ }
+
+ qunit = lqe->lqe_qunit == 0 ? pool->qpi_least_qunit : lqe->lqe_qunit;
+
+ /* The qunit value is computed as follows: limit / (2 * slv_cnt).
+ * Then 75% of the quota space can be granted with current qunit value.
+ * The remaining 25% are then used with reduced qunit size (by a factor
+ * of 4) which is then divided in a similar manner.
+ *
+ * |---------------------limit---------------------|
+ * |-------limit / 2-------|-limit / 4-|-limit / 4-|
+ * |qunit|qunit|qunit|qunit| | |
+ * |----slv_cnt * qunit----| | |
+ * |-grow limit-| | | |
+ * |--------------shrink limit---------| |
+ * |---space granted in qunit chunks---|-remaining-|
+ * / \
+ * / \
+ * / \
+ * / \
+ * / \
+ * qunit >>= 2; |qunit*slv_cnt|qunit*slv_cnt|
+ * |---space in qunit---|remain|
+ * ... */
+ if (qunit == pool->qpi_least_qunit ||
+ limit >= lqe->lqe_granted + ((slv_cnt * qunit) >> 1)) {
+ /* current qunit value still fits, let's see if we can afford to
+ * increase qunit now ...
+ * To increase qunit again, we have to be under 25% */
+ while (qunit && limit >= lqe->lqe_granted + 6 * qunit * slv_cnt)
+ qunit <<= 2;
+
+ if (!qunit) {
+ qunit = limit;
+ do_div(qunit, 2 * slv_cnt);
+ }
+
+ } else {
+ /* shrink qunit until we find a suitable value */
+ while (qunit > pool->qpi_least_qunit &&
+ limit < lqe->lqe_granted + ((slv_cnt * qunit) >> 1))
+ qunit >>= 2;
+ }
+
+ if (qunit2 && qunit > qunit2)
+ qunit = qunit2;
+done:
+ if (lqe->lqe_qunit == qunit)
+ /* keep current qunit */
+ RETURN_EXIT;
+
+ LQUOTA_DEBUG(lqe, "%s qunit to %llu",
+ lqe->lqe_qunit < qunit ? "increasing" : "decreasing",
+ qunit);
+
+ /* store new qunit value */
+ swap(lqe->lqe_qunit, qunit);
+
+ /* reset revoke time */
+ lqe->lqe_revoke_time = 0;
+
+ if (lqe->lqe_qunit < qunit)
+ /* let's notify slave of qunit shrinking */
+ qmt_id_lock_notify(pool->qpi_qmt, lqe);
+ else if (lqe->lqe_qunit == pool->qpi_least_qunit)
+ /* initial qunit value is the smallest one */
+ lqe->lqe_revoke_time = cfs_time_current_64();
+ EXIT;
+}
+
+/*
+ * Adjust qunit & edquot flag in case it wasn't initialized already (e.g.
+ * limit set while no slaves were connected yet)
+ */
+void qmt_revalidate(const struct lu_env *env, struct lquota_entry *lqe)
+{
+ if (lqe->lqe_qunit == 0) {
+ /* lqe was read from disk, but neither qunit, nor edquot flag
+ * were initialized */
+ qmt_adjust_qunit(env, lqe);
+ if (lqe->lqe_qunit != 0)
+ qmt_adjust_edquot(lqe, cfs_time_current_sec());
+ }
+}
git://git.whamcloud.com / fs / lustre-release.git / blobdiff