xfs: add a shrinker to background inode reclaim
authorDave Chinner <dchinner@redhat.com>
Wed, 28 Apr 2010 23:55:50 +0000 (09:55 +1000)
committerAlex Elder <aelder@sgi.com>
Thu, 29 Apr 2010 21:22:13 +0000 (16:22 -0500)
On low memory boxes or those with highmem, kernel can OOM before the
background reclaims inodes via xfssyncd. Add a shrinker to run inode
reclaim so that it inode reclaim is expedited when memory is low.

This is more complex than it needs to be because the VM folk don't
want a context added to the shrinker infrastructure. Hence we need
to add a global list of XFS mount structures so the shrinker can
traverse them.

Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
fs/xfs/linux-2.6/xfs_super.c
fs/xfs/linux-2.6/xfs_sync.c
fs/xfs/linux-2.6/xfs_sync.h
fs/xfs/quota/xfs_qm_syscalls.c
fs/xfs/xfs_ag.h
fs/xfs/xfs_mount.h

index 52e06b487ced9467d83bc28c01376e67c1c49749..29f1edca76de3eca8d6af75ce7a197c6eb82a368 100644 (file)
@@ -1209,6 +1209,7 @@ xfs_fs_put_super(
 
        xfs_unmountfs(mp);
        xfs_freesb(mp);
+       xfs_inode_shrinker_unregister(mp);
        xfs_icsb_destroy_counters(mp);
        xfs_close_devices(mp);
        xfs_dmops_put(mp);
@@ -1622,6 +1623,8 @@ xfs_fs_fill_super(
        if (error)
                goto fail_vnrele;
 
+       xfs_inode_shrinker_register(mp);
+
        kfree(mtpt);
        return 0;
 
@@ -1867,6 +1870,7 @@ init_xfs_fs(void)
                goto out_cleanup_procfs;
 
        vfs_initquota();
+       xfs_inode_shrinker_init();
 
        error = register_filesystem(&xfs_fs_type);
        if (error)
@@ -1894,6 +1898,7 @@ exit_xfs_fs(void)
 {
        vfs_exitquota();
        unregister_filesystem(&xfs_fs_type);
+       xfs_inode_shrinker_destroy();
        xfs_sysctl_unregister();
        xfs_cleanup_procfs();
        xfs_buf_terminate();
index fd96982157593e741ee354c2c228f50c9aa7d1e7..a427c638d9095180e114fe8d64d3f16d66d9201b 100644 (file)
@@ -95,7 +95,8 @@ xfs_inode_ag_walk(
                                           struct xfs_perag *pag, int flags),
        int                     flags,
        int                     tag,
-       int                     exclusive)
+       int                     exclusive,
+       int                     *nr_to_scan)
 {
        uint32_t                first_index;
        int                     last_error = 0;
@@ -134,7 +135,7 @@ restart:
                if (error == EFSCORRUPTED)
                        break;
 
-       } while (1);
+       } while ((*nr_to_scan)--);
 
        if (skipped) {
                delay(1);
@@ -150,12 +151,15 @@ xfs_inode_ag_iterator(
                                           struct xfs_perag *pag, int flags),
        int                     flags,
        int                     tag,
-       int                     exclusive)
+       int                     exclusive,
+       int                     *nr_to_scan)
 {
        int                     error = 0;
        int                     last_error = 0;
        xfs_agnumber_t          ag;
+       int                     nr;
 
+       nr = nr_to_scan ? *nr_to_scan : INT_MAX;
        for (ag = 0; ag < mp->m_sb.sb_agcount; ag++) {
                struct xfs_perag        *pag;
 
@@ -165,14 +169,18 @@ xfs_inode_ag_iterator(
                        continue;
                }
                error = xfs_inode_ag_walk(mp, pag, execute, flags, tag,
-                                               exclusive);
+                                               exclusive, &nr);
                xfs_perag_put(pag);
                if (error) {
                        last_error = error;
                        if (error == EFSCORRUPTED)
                                break;
                }
+               if (nr <= 0)
+                       break;
        }
+       if (nr_to_scan)
+               *nr_to_scan = nr;
        return XFS_ERROR(last_error);
 }
 
@@ -291,7 +299,7 @@ xfs_sync_data(
        ASSERT((flags & ~(SYNC_TRYLOCK|SYNC_WAIT)) == 0);
 
        error = xfs_inode_ag_iterator(mp, xfs_sync_inode_data, flags,
-                                     XFS_ICI_NO_TAG, 0);
+                                     XFS_ICI_NO_TAG, 0, NULL);
        if (error)
                return XFS_ERROR(error);
 
@@ -310,7 +318,7 @@ xfs_sync_attr(
        ASSERT((flags & ~SYNC_WAIT) == 0);
 
        return xfs_inode_ag_iterator(mp, xfs_sync_inode_attr, flags,
-                                    XFS_ICI_NO_TAG, 0);
+                                    XFS_ICI_NO_TAG, 0, NULL);
 }
 
 STATIC int
@@ -673,6 +681,7 @@ __xfs_inode_set_reclaim_tag(
        radix_tree_tag_set(&pag->pag_ici_root,
                           XFS_INO_TO_AGINO(ip->i_mount, ip->i_ino),
                           XFS_ICI_RECLAIM_TAG);
+       pag->pag_ici_reclaimable++;
 }
 
 /*
@@ -705,6 +714,7 @@ __xfs_inode_clear_reclaim_tag(
 {
        radix_tree_tag_clear(&pag->pag_ici_root,
                        XFS_INO_TO_AGINO(mp, ip->i_ino), XFS_ICI_RECLAIM_TAG);
+       pag->pag_ici_reclaimable--;
 }
 
 /*
@@ -854,5 +864,93 @@ xfs_reclaim_inodes(
        int             mode)
 {
        return xfs_inode_ag_iterator(mp, xfs_reclaim_inode, mode,
-                                       XFS_ICI_RECLAIM_TAG, 1);
+                                       XFS_ICI_RECLAIM_TAG, 1, NULL);
+}
+
+/*
+ * Shrinker infrastructure.
+ *
+ * This is all far more complex than it needs to be. It adds a global list of
+ * mounts because the shrinkers can only call a global context. We need to make
+ * the shrinkers pass a context to avoid the need for global state.
+ */
+static LIST_HEAD(xfs_mount_list);
+static struct rw_semaphore xfs_mount_list_lock;
+
+static int
+xfs_reclaim_inode_shrink(
+       int             nr_to_scan,
+       gfp_t           gfp_mask)
+{
+       struct xfs_mount *mp;
+       struct xfs_perag *pag;
+       xfs_agnumber_t  ag;
+       int             reclaimable = 0;
+
+       if (nr_to_scan) {
+               if (!(gfp_mask & __GFP_FS))
+                       return -1;
+
+               down_read(&xfs_mount_list_lock);
+               list_for_each_entry(mp, &xfs_mount_list, m_mplist) {
+                       xfs_inode_ag_iterator(mp, xfs_reclaim_inode, 0,
+                                       XFS_ICI_RECLAIM_TAG, 1, &nr_to_scan);
+                       if (nr_to_scan <= 0)
+                               break;
+               }
+               up_read(&xfs_mount_list_lock);
+       }
+
+       down_read(&xfs_mount_list_lock);
+       list_for_each_entry(mp, &xfs_mount_list, m_mplist) {
+               for (ag = 0; ag < mp->m_sb.sb_agcount; ag++) {
+
+                       pag = xfs_perag_get(mp, ag);
+                       if (!pag->pag_ici_init) {
+                               xfs_perag_put(pag);
+                               continue;
+                       }
+                       reclaimable += pag->pag_ici_reclaimable;
+                       xfs_perag_put(pag);
+               }
+       }
+       up_read(&xfs_mount_list_lock);
+       return reclaimable;
+}
+
+static struct shrinker xfs_inode_shrinker = {
+       .shrink = xfs_reclaim_inode_shrink,
+       .seeks = DEFAULT_SEEKS,
+};
+
+void __init
+xfs_inode_shrinker_init(void)
+{
+       init_rwsem(&xfs_mount_list_lock);
+       register_shrinker(&xfs_inode_shrinker);
+}
+
+void
+xfs_inode_shrinker_destroy(void)
+{
+       ASSERT(list_empty(&xfs_mount_list));
+       unregister_shrinker(&xfs_inode_shrinker);
+}
+
+void
+xfs_inode_shrinker_register(
+       struct xfs_mount        *mp)
+{
+       down_write(&xfs_mount_list_lock);
+       list_add_tail(&mp->m_mplist, &xfs_mount_list);
+       up_write(&xfs_mount_list_lock);
+}
+
+void
+xfs_inode_shrinker_unregister(
+       struct xfs_mount        *mp)
+{
+       down_write(&xfs_mount_list_lock);
+       list_del(&mp->m_mplist);
+       up_write(&xfs_mount_list_lock);
 }
index d480c346cabbb124a0d48ab4e017ef6fefae062e..cdcbaaca9880d04dfa8f2146580b0ef216ed178f 100644 (file)
@@ -53,6 +53,11 @@ void __xfs_inode_clear_reclaim_tag(struct xfs_mount *mp, struct xfs_perag *pag,
 int xfs_sync_inode_valid(struct xfs_inode *ip, struct xfs_perag *pag);
 int xfs_inode_ag_iterator(struct xfs_mount *mp,
        int (*execute)(struct xfs_inode *ip, struct xfs_perag *pag, int flags),
-       int flags, int tag, int write_lock);
+       int flags, int tag, int write_lock, int *nr_to_scan);
+
+void xfs_inode_shrinker_init(void);
+void xfs_inode_shrinker_destroy(void);
+void xfs_inode_shrinker_register(struct xfs_mount *mp);
+void xfs_inode_shrinker_unregister(struct xfs_mount *mp);
 
 #endif
index 5d0ee8d492dbc21f9e04e9dd661bf3c35862e607..50bee07d6b0ed144544861cadc14a14be9b138b6 100644 (file)
@@ -891,7 +891,8 @@ xfs_qm_dqrele_all_inodes(
        uint             flags)
 {
        ASSERT(mp->m_quotainfo);
-       xfs_inode_ag_iterator(mp, xfs_dqrele_inode, flags, XFS_ICI_NO_TAG, 0);
+       xfs_inode_ag_iterator(mp, xfs_dqrele_inode, flags,
+                               XFS_ICI_NO_TAG, 0, NULL);
 }
 
 /*------------------------------------------------------------------------*/
index b1a5a1ff88eac58f3618a977b1c3ce0e23e10b0b..abb8222b88c9a06e90c00784a60564170ba253da 100644 (file)
@@ -223,6 +223,7 @@ typedef struct xfs_perag {
        int             pag_ici_init;   /* incore inode cache initialised */
        rwlock_t        pag_ici_lock;   /* incore inode lock */
        struct radix_tree_root pag_ici_root;    /* incore inode cache root */
+       int             pag_ici_reclaimable;    /* reclaimable inodes */
 #endif
        int             pagb_count;     /* pagb slots in use */
        xfs_perag_busy_t pagb_list[XFS_PAGB_NUM_SLOTS]; /* unstable blocks */
index 4fa0bc7b983ebb0e934b4b877f63c6090fb73afa..9ff48a16a7ee6334a2ab310dee58f5dc5a6ba323 100644 (file)
@@ -259,6 +259,7 @@ typedef struct xfs_mount {
        wait_queue_head_t       m_wait_single_sync_task;
        __int64_t               m_update_flags; /* sb flags we need to update
                                                   on the next remount,rw */
+       struct list_head        m_mplist;       /* inode shrinker mount list */
 } xfs_mount_t;
 
 /*