Whamcloud - gitweb
LU-12616 obclass: fix MDS start/stop race
[fs/lustre-release.git] / lustre / quota / qmt_entry.c
index 6fe3c92..d3ed254 100644 (file)
  * GPL HEADER END
  */
 /*
- * Copyright (c) 2012, Intel Corporation.
+ * Copyright (c) 2012, 2016, 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"
@@ -50,6 +46,55 @@ static void qmt_lqe_init(struct lquota_entry *lqe, void *arg)
        init_rwsem(&lqe->lqe_sem);
 }
 
+/* Apply the default quota setting to the specified quota entry
+ *
+ * \param env          - is the environment passed by the caller
+ * \param pool         - is the quota pool of the quota entry
+ * \param lqe          - is the lquota_entry object to apply default quota on
+ * \param create_record        - if true, an global quota record will be created and
+ *                        write to the disk.
+ *
+ * \retval 0           : success
+ * \retval -ve         : other appropriate errors
+ */
+int qmt_lqe_set_default(const struct lu_env *env, struct qmt_pool_info *pool,
+                       struct lquota_entry *lqe, bool create_record)
+{
+       struct lquota_entry     *lqe_def;
+       int                     rc = 0;
+
+       ENTRY;
+
+       if (lqe->lqe_id.qid_uid == 0)
+               RETURN(0);
+
+       lqe_def = pool->qpi_grace_lqe[lqe->lqe_site->lqs_qtype];
+
+       LQUOTA_DEBUG(lqe, "inherit default quota");
+
+       lqe->lqe_is_default = true;
+       lqe->lqe_hardlimit = lqe_def->lqe_hardlimit;
+       lqe->lqe_softlimit = lqe_def->lqe_softlimit;
+
+       if (create_record) {
+               lqe->lqe_uptodate = true;
+               rc = qmt_set_with_lqe(env, pool->qpi_qmt, lqe, 0, 0,
+                                     LQUOTA_GRACE_FLAG(0, LQUOTA_FLAG_DEFAULT),
+                                     QIF_TIMES, true, false);
+
+               if (rc != 0)
+                       LQUOTA_ERROR(lqe, "failed to create the global quota"
+                                    " record: %d", rc);
+       }
+
+       if (lqe->lqe_hardlimit == 0 && lqe->lqe_softlimit == 0)
+               lqe->lqe_enforced = false;
+       else
+               lqe->lqe_enforced = true;
+
+       RETURN(rc);
+}
+
 /*
  * Update a lquota entry. This is done by reading quota settings from the global
  * index. The lquota entry must be write locked.
@@ -74,22 +119,19 @@ static int qmt_lqe_read(const struct lu_env *env, struct lquota_entry *lqe,
 
        switch (rc) {
        case -ENOENT:
-               /* no such entry, assume quota isn't enforced for this user */
-               lqe->lqe_enforced = false;
+               qmt_lqe_set_default(env, pool, lqe, true);
                break;
        case 0:
                /* copy quota settings from on-disk record */
                lqe->lqe_granted   = qti->qti_glb_rec.qbr_granted;
                lqe->lqe_hardlimit = qti->qti_glb_rec.qbr_hardlimit;
                lqe->lqe_softlimit = qti->qti_glb_rec.qbr_softlimit;
-               lqe->lqe_gracetime = qti->qti_glb_rec.qbr_time;
-
-               if (lqe->lqe_hardlimit == 0 && lqe->lqe_softlimit == 0)
-                       /* {hard,soft}limit=0 means no quota enforced */
-                       lqe->lqe_enforced = false;
-               else
-                       lqe->lqe_enforced  = true;
+               lqe->lqe_gracetime = LQUOTA_GRACE(qti->qti_glb_rec.qbr_time);
 
+               if (lqe->lqe_hardlimit == 0 && lqe->lqe_softlimit == 0 &&
+                   (LQUOTA_FLAG(qti->qti_glb_rec.qbr_time) &
+                    LQUOTA_FLAG_DEFAULT))
+                       qmt_lqe_set_default(env, pool, lqe, false);
                break;
        default:
                LQUOTA_ERROR(lqe, "failed to read quota entry from disk, rc:%d",
@@ -97,6 +139,13 @@ static int qmt_lqe_read(const struct lu_env *env, struct lquota_entry *lqe,
                RETURN(rc);
        }
 
+       if (lqe->lqe_id.qid_uid == 0 ||
+           (lqe->lqe_hardlimit == 0 && lqe->lqe_softlimit == 0))
+               /* {hard,soft}limit=0 means no quota enforced */
+               lqe->lqe_enforced = false;
+       else
+               lqe->lqe_enforced  = true;
+
        LQUOTA_DEBUG(lqe, "read");
        RETURN(0);
 }
@@ -111,22 +160,21 @@ static int qmt_lqe_read(const struct lu_env *env, struct lquota_entry *lqe,
  */
 static void qmt_lqe_debug(struct lquota_entry *lqe, void *arg,
                          struct libcfs_debug_msg_data *msgdata,
-                         const char *fmt, va_list args)
+                         struct va_format *vaf)
 {
        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 may_rel:"LPU64" revoke:"LPU64"\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_may_rel,
-                          lqe->lqe_revoke_time);
+       libcfs_debug_msg(msgdata,
+                        "%pV qmt:%s pool:%s-%s id:%llu enforced:%d hard:%llu soft:%llu granted:%llu time:%llu qunit: %llu edquot:%d may_rel:%llu revoke:%lld default:%s\n",
+                        vaf, pool->qpi_qmt->qmt_svname,
+                        RES_NAME(pool->qpi_rtype),
+                        pool->qpi_name,
+                        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_may_rel,
+                        lqe->lqe_revoke_time,
+                        lqe->lqe_is_default ? "yes" : "no");
 }
 
 /*
@@ -268,9 +316,15 @@ int qmt_glb_write(const struct lu_env *env, struct thandle *th,
 
        /* 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;
+       if (lqe->lqe_is_default) {
+               rec->qbr_hardlimit = 0;
+               rec->qbr_softlimit = 0;
+               rec->qbr_time      = LQUOTA_GRACE_FLAG(0, LQUOTA_FLAG_DEFAULT);
+       } else {
+               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,
@@ -331,7 +385,7 @@ int qmt_slv_read(const struct lu_env *env, struct lquota_entry *lqe,
                RETURN(rc);
        }
 
-       LQUOTA_DEBUG(lqe, "successful slv read "LPU64, *granted);
+       LQUOTA_DEBUG(lqe, "successful slv read %llu", *granted);
 
        RETURN(0);
 }
@@ -364,7 +418,7 @@ int qmt_slv_write(const struct lu_env *env, struct thandle *th,
        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);
 
        /* never delete the entry, otherwise, it'll not be transferred
@@ -379,7 +433,7 @@ int qmt_slv_write(const struct lu_env *env, struct thandle *th,
                               (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);
        }
@@ -427,21 +481,29 @@ void qmt_adjust_edquot(struct lquota_entry *lqe, __u64 now)
                        /* the qmt still has available space */
                        RETURN_EXIT;
 
-               if (lqe->lqe_qunit != pool->qpi_least_qunit)
-                       /* we haven't reached the minimal qunit yet, so there is
-                        * still hope that the rebalancing process might free up
-                        * some quota space */
-                       RETURN_EXIT;
+               /* See comment in qmt_adjust_qunit(). LU-4139 */
+               if (qmt_hard_exhausted(lqe) ||
+                   pool->qpi_rtype != LQUOTA_RES_DT) {
+                       time64_t lapse;
+
+                       /* 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;
 
-               if (lqe->lqe_revoke_time == 0)
                        /* least qunit value not sent to all slaves yet */
-                       RETURN_EXIT;
+                       if (lqe->lqe_revoke_time == 0)
+                               RETURN_EXIT;
 
-               if (lqe->lqe_may_rel != 0 &&
-                   cfs_time_before_64(cfs_time_shift_64(-QMT_REBA_TIMEOUT),
-                                      lqe->lqe_revoke_time))
                        /* Let's give more time to slave to release space */
-                       RETURN_EXIT;
+                       lapse = ktime_get_seconds() - QMT_REBA_TIMEOUT;
+                       if (lqe->lqe_may_rel != 0 && lqe->lqe_revoke_time > lapse)
+                               RETURN_EXIT;
+               } else {
+                       if (lqe->lqe_qunit > pool->qpi_soft_least_qunit)
+                               RETURN_EXIT;
+               }
 
                /* set edquot flag */
                lqe->lqe_edquot = true;
@@ -472,6 +534,30 @@ void qmt_adjust_edquot(struct lquota_entry *lqe, __u64 now)
        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_rtype != 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.
  *
@@ -494,10 +580,16 @@ __u64 qmt_alloc_expand(struct lquota_entry *lqe, __u64 granted, __u64 spare)
        slv_cnt = lqe2qpi(lqe)->qpi_slv_nr[lqe->lqe_site->lqs_qtype];
        qunit   = lqe->lqe_qunit;
 
-       if (lqe->lqe_softlimit != 0)
-               remaining = lqe->lqe_softlimit;
-       else
+       /* 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);
@@ -540,7 +632,7 @@ 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;
+       __u64                    qunit, limit, qunit2 = 0;
        ENTRY;
 
        LASSERT(lqe_is_locked(lqe));
@@ -560,7 +652,15 @@ void qmt_adjust_qunit(const struct lu_env *env, struct lquota_entry *lqe)
         * beyond the soft limit. This will impact performance, but that's the
         * price of an accurate grace time management. */
        if (lqe->lqe_softlimit != 0) {
-               limit = lqe->lqe_softlimit;
+               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 {
@@ -596,8 +696,14 @@ void qmt_adjust_qunit(const struct lu_env *env, struct lquota_entry *lqe)
                /* 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 (limit >= lqe->lqe_granted + 6 * qunit * slv_cnt)
+               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 &&
@@ -605,11 +711,14 @@ void qmt_adjust_qunit(const struct lu_env *env, struct lquota_entry *lqe)
                        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 "LPU64,
+       LQUOTA_DEBUG(lqe, "%s qunit to %llu",
                     lqe->lqe_qunit < qunit ? "increasing" : "decreasing",
                     qunit);
 
@@ -624,7 +733,7 @@ void qmt_adjust_qunit(const struct lu_env *env, struct lquota_entry *lqe)
                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();
+               lqe->lqe_revoke_time = ktime_get_seconds();
        EXIT;
 }
 
@@ -639,6 +748,6 @@ void qmt_revalidate(const struct lu_env *env, struct lquota_entry *lqe)
                 * were initialized */
                qmt_adjust_qunit(env, lqe);
                if (lqe->lqe_qunit != 0)
-                       qmt_adjust_edquot(lqe, cfs_time_current_sec());
+                       qmt_adjust_edquot(lqe, ktime_get_real_seconds());
        }
 }