]> git.openfabrics.org - ~shefty/rdma-dev.git/blobdiff - fs/xfs/xfs_qm.c
xfs: use a normal shrinker for the dquot freelist
[~shefty/rdma-dev.git] / fs / xfs / xfs_qm.c
index 671f37eae1c7bbe3c9f89fd78b9c51b9195f4e32..c436def733bf892eb324603f688e6d92b92ffc61 100644 (file)
@@ -50,7 +50,6 @@
  */
 struct mutex   xfs_Gqm_lock;
 struct xfs_qm  *xfs_Gqm;
-uint           ndquot;
 
 kmem_zone_t    *qm_dqzone;
 kmem_zone_t    *qm_dqtrxzone;
@@ -93,7 +92,6 @@ xfs_Gqm_init(void)
                goto out_free_udqhash;
 
        hsize /= sizeof(xfs_dqhash_t);
-       ndquot = hsize << 8;
 
        xqm = kmem_zalloc(sizeof(xfs_qm_t), KM_SLEEP);
        xqm->qm_dqhashmask = hsize - 1;
@@ -137,7 +135,6 @@ xfs_Gqm_init(void)
                xqm->qm_dqtrxzone = qm_dqtrxzone;
 
        atomic_set(&xqm->qm_totaldquots, 0);
-       xqm->qm_dqfree_ratio = XFS_QM_DQFREE_RATIO;
        xqm->qm_nrefs = 0;
        return xqm;
 
@@ -1600,216 +1597,150 @@ xfs_qm_init_quotainos(
        return 0;
 }
 
+STATIC void
+xfs_qm_dqfree_one(
+       struct xfs_dquot        *dqp)
+{
+       struct xfs_mount        *mp = dqp->q_mount;
+       struct xfs_quotainfo    *qi = mp->m_quotainfo;
 
+       mutex_lock(&dqp->q_hash->qh_lock);
+       list_del_init(&dqp->q_hashlist);
+       dqp->q_hash->qh_version++;
+       mutex_unlock(&dqp->q_hash->qh_lock);
 
-/*
- * Pop the least recently used dquot off the freelist and recycle it.
- */
-STATIC struct xfs_dquot *
-xfs_qm_dqreclaim_one(void)
+       mutex_lock(&qi->qi_dqlist_lock);
+       list_del_init(&dqp->q_mplist);
+       qi->qi_dquots--;
+       qi->qi_dqreclaims++;
+       mutex_unlock(&qi->qi_dqlist_lock);
+
+       xfs_qm_dqdestroy(dqp);
+}
+
+STATIC void
+xfs_qm_dqreclaim_one(
+       struct xfs_dquot        *dqp,
+       struct list_head        *dispose_list)
 {
-       struct xfs_dquot        *dqp;
-       int                     restarts = 0;
+       struct xfs_mount        *mp = dqp->q_mount;
+       int                     error;
 
-       mutex_lock(&xfs_Gqm->qm_dqfrlist_lock);
-restart:
-       list_for_each_entry(dqp, &xfs_Gqm->qm_dqfrlist, q_freelist) {
-               struct xfs_mount *mp = dqp->q_mount;
+       if (!xfs_dqlock_nowait(dqp))
+               goto out_busy;
 
-               if (!xfs_dqlock_nowait(dqp))
-                       continue;
+       /*
+        * This dquot has acquired a reference in the meantime remove it from
+        * the freelist and try again.
+        */
+       if (dqp->q_nrefs) {
+               xfs_dqunlock(dqp);
 
-               /*
-                * This dquot has already been grabbed by dqlookup.
-                * Remove it from the freelist and try again.
-                */
-               if (dqp->q_nrefs) {
-                       trace_xfs_dqreclaim_want(dqp);
-                       XQM_STATS_INC(xqmstats.xs_qm_dqwants);
-
-                       list_del_init(&dqp->q_freelist);
-                       xfs_Gqm->qm_dqfrlist_cnt--;
-                       restarts++;
-                       goto dqunlock;
-               }
+               trace_xfs_dqreclaim_want(dqp);
+               XQM_STATS_INC(xqmstats.xs_qm_dqwants);
 
-               ASSERT(dqp->q_hash);
-               ASSERT(!list_empty(&dqp->q_mplist));
+               list_del_init(&dqp->q_freelist);
+               xfs_Gqm->qm_dqfrlist_cnt--;
+               return;
+       }
 
-               /*
-                * Try to grab the flush lock. If this dquot is in the process
-                * of getting flushed to disk, we don't want to reclaim it.
-                */
-               if (!xfs_dqflock_nowait(dqp))
-                       goto dqunlock;
+       ASSERT(dqp->q_hash);
+       ASSERT(!list_empty(&dqp->q_mplist));
 
-               /*
-                * We have the flush lock so we know that this is not in the
-                * process of being flushed. So, if this is dirty, flush it
-                * DELWRI so that we don't get a freelist infested with
-                * dirty dquots.
-                */
-               if (XFS_DQ_IS_DIRTY(dqp)) {
-                       int     error;
+       /*
+        * Try to grab the flush lock. If this dquot is in the process of
+        * getting flushed to disk, we don't want to reclaim it.
+        */
+       if (!xfs_dqflock_nowait(dqp))
+               goto out_busy;
 
-                       trace_xfs_dqreclaim_dirty(dqp);
+       /*
+        * We have the flush lock so we know that this is not in the
+        * process of being flushed. So, if this is dirty, flush it
+        * DELWRI so that we don't get a freelist infested with
+        * dirty dquots.
+        */
+       if (XFS_DQ_IS_DIRTY(dqp)) {
+               trace_xfs_dqreclaim_dirty(dqp);
 
-                       /*
-                        * We flush it delayed write, so don't bother
-                        * releasing the freelist lock.
-                        */
-                       error = xfs_qm_dqflush(dqp, SYNC_TRYLOCK);
-                       if (error) {
-                               xfs_warn(mp, "%s: dquot %p flush failed",
-                                       __func__, dqp);
-                       }
-                       goto dqunlock;
+               /*
+                * We flush it delayed write, so don't bother releasing the
+                * freelist lock.
+                */
+               error = xfs_qm_dqflush(dqp, 0);
+               if (error) {
+                       xfs_warn(mp, "%s: dquot %p flush failed",
+                                __func__, dqp);
                }
-               xfs_dqfunlock(dqp);
 
                /*
-                * Prevent lookup now that we are going to reclaim the dquot.
-                * Once XFS_DQ_FREEING is set lookup won't touch the dquot,
-                * thus we can drop the lock now.
+                * Give the dquot another try on the freelist, as the
+                * flushing will take some time.
                 */
-               dqp->dq_flags |= XFS_DQ_FREEING;
-               xfs_dqunlock(dqp);
-
-               mutex_lock(&dqp->q_hash->qh_lock);
-               list_del_init(&dqp->q_hashlist);
-               dqp->q_hash->qh_version++;
-               mutex_unlock(&dqp->q_hash->qh_lock);
-
-               mutex_lock(&mp->m_quotainfo->qi_dqlist_lock);
-               list_del_init(&dqp->q_mplist);
-               mp->m_quotainfo->qi_dquots--;
-               mp->m_quotainfo->qi_dqreclaims++;
-               mutex_unlock(&mp->m_quotainfo->qi_dqlist_lock);
+               goto out_busy;
+       }
+       xfs_dqfunlock(dqp);
 
-               ASSERT(dqp->q_nrefs == 0);
-               list_del_init(&dqp->q_freelist);
-               xfs_Gqm->qm_dqfrlist_cnt--;
+       /*
+        * Prevent lookups now that we are past the point of no return.
+        */
+       dqp->dq_flags |= XFS_DQ_FREEING;
+       xfs_dqunlock(dqp);
 
-               mutex_unlock(&xfs_Gqm->qm_dqfrlist_lock);
-               return dqp;
-dqunlock:
-               xfs_dqunlock(dqp);
-               if (restarts >= XFS_QM_RECLAIM_MAX_RESTARTS)
-                       break;
-               goto restart;
-       }
+       ASSERT(dqp->q_nrefs == 0);
+       list_move_tail(&dqp->q_freelist, dispose_list);
+       xfs_Gqm->qm_dqfrlist_cnt--;
 
-       mutex_unlock(&xfs_Gqm->qm_dqfrlist_lock);
-       return NULL;
-}
+       trace_xfs_dqreclaim_done(dqp);
+       XQM_STATS_INC(xqmstats.xs_qm_dqreclaims);
+       return;
 
-/*
- * Traverse the freelist of dquots and attempt to reclaim a maximum of
- * 'howmany' dquots. This operation races with dqlookup(), and attempts to
- * favor the lookup function ...
- */
-STATIC int
-xfs_qm_shake_freelist(
-       int     howmany)
-{
-       int             nreclaimed = 0;
-       xfs_dquot_t     *dqp;
+out_busy:
+       xfs_dqunlock(dqp);
 
-       if (howmany <= 0)
-               return 0;
+       /*
+        * Move the dquot to the tail of the list so that we don't spin on it.
+        */
+       list_move_tail(&dqp->q_freelist, &xfs_Gqm->qm_dqfrlist);
 
-       while (nreclaimed < howmany) {
-               dqp = xfs_qm_dqreclaim_one();
-               if (!dqp)
-                       return nreclaimed;
-               xfs_qm_dqdestroy(dqp);
-               nreclaimed++;
-       }
-       return nreclaimed;
+       trace_xfs_dqreclaim_busy(dqp);
+       XQM_STATS_INC(xqmstats.xs_qm_dqreclaim_misses);
 }
 
-/*
- * The kmem_shake interface is invoked when memory is running low.
- */
-/* ARGSUSED */
 STATIC int
 xfs_qm_shake(
-       struct shrinker *shrink,
-       struct shrink_control *sc)
+       struct shrinker         *shrink,
+       struct shrink_control   *sc)
 {
-       int     ndqused, nfree, n;
-       gfp_t gfp_mask = sc->gfp_mask;
-
-       if (!kmem_shake_allow(gfp_mask))
-               return 0;
-       if (!xfs_Gqm)
-               return 0;
-
-       nfree = xfs_Gqm->qm_dqfrlist_cnt; /* free dquots */
-       /* incore dquots in all f/s's */
-       ndqused = atomic_read(&xfs_Gqm->qm_totaldquots) - nfree;
-
-       ASSERT(ndqused >= 0);
+       int                     nr_to_scan = sc->nr_to_scan;
+       LIST_HEAD               (dispose_list);
+       struct xfs_dquot        *dqp;
 
-       if (nfree <= ndqused && nfree < ndquot)
+       if ((sc->gfp_mask & (__GFP_FS|__GFP_WAIT)) != (__GFP_FS|__GFP_WAIT))
                return 0;
+       if (!nr_to_scan)
+               goto out;
 
-       ndqused *= xfs_Gqm->qm_dqfree_ratio;    /* target # of free dquots */
-       n = nfree - ndqused - ndquot;           /* # over target */
-
-       return xfs_qm_shake_freelist(MAX(nfree, n));
-}
-
-
-/*------------------------------------------------------------------*/
-
-/*
- * Return a new incore dquot. Depending on the number of
- * dquots in the system, we either allocate a new one on the kernel heap,
- * or reclaim a free one.
- * Return value is B_TRUE if we allocated a new dquot, B_FALSE if we managed
- * to reclaim an existing one from the freelist.
- */
-boolean_t
-xfs_qm_dqalloc_incore(
-       xfs_dquot_t **O_dqpp)
-{
-       xfs_dquot_t     *dqp;
-
-       /*
-        * Check against high water mark to see if we want to pop
-        * a nincompoop dquot off the freelist.
-        */
-       if (atomic_read(&xfs_Gqm->qm_totaldquots) >= ndquot) {
-               /*
-                * Try to recycle a dquot from the freelist.
-                */
-               if ((dqp = xfs_qm_dqreclaim_one())) {
-                       XQM_STATS_INC(xqmstats.xs_qm_dqreclaims);
-                       /*
-                        * Just zero the core here. The rest will get
-                        * reinitialized by caller. XXX we shouldn't even
-                        * do this zero ...
-                        */
-                       memset(&dqp->q_core, 0, sizeof(dqp->q_core));
-                       *O_dqpp = dqp;
-                       return B_FALSE;
-               }
-               XQM_STATS_INC(xqmstats.xs_qm_dqreclaim_misses);
+       mutex_lock(&xfs_Gqm->qm_dqfrlist_lock);
+       while (!list_empty(&xfs_Gqm->qm_dqfrlist)) {
+               if (nr_to_scan-- <= 0)
+                       break;
+               dqp = list_first_entry(&xfs_Gqm->qm_dqfrlist, struct xfs_dquot,
+                                      q_freelist);
+               xfs_qm_dqreclaim_one(dqp, &dispose_list);
        }
+       mutex_unlock(&xfs_Gqm->qm_dqfrlist_lock);
 
-       /*
-        * Allocate a brand new dquot on the kernel heap and return it
-        * to the caller to initialize.
-        */
-       ASSERT(xfs_Gqm->qm_dqzone != NULL);
-       *O_dqpp = kmem_zone_zalloc(xfs_Gqm->qm_dqzone, KM_SLEEP);
-       atomic_inc(&xfs_Gqm->qm_totaldquots);
-
-       return B_TRUE;
+       while (!list_empty(&dispose_list)) {
+               dqp = list_first_entry(&dispose_list, struct xfs_dquot,
+                                      q_freelist);
+               list_del_init(&dqp->q_freelist);
+               xfs_qm_dqfree_one(dqp);
+       }
+out:
+       return (xfs_Gqm->qm_dqfrlist_cnt / 100) * sysctl_vfs_cache_pressure;
 }
 
-
 /*
  * Start a transaction and write the incore superblock changes to
  * disk. flags parameter indicates which fields have changed.