]> git.baikalelectronics.ru Git - kernel.git/commitdiff
nfsd: Containerise filecache laundrette
authorTrond Myklebust <trondmy@gmail.com>
Mon, 6 Jan 2020 18:18:05 +0000 (13:18 -0500)
committerJ. Bruce Fields <bfields@redhat.com>
Wed, 22 Jan 2020 21:25:40 +0000 (16:25 -0500)
Ensure that if the filecache laundrette gets stuck, it only affects
the knfsd instances of one container.

The notifier callbacks can be called from various contexts so avoid
using synchonous filesystem operations that might deadlock.

Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>
fs/nfsd/filecache.c
fs/nfsd/filecache.h
fs/nfsd/nfssvc.c

index c048e3071db76f6a77093e480ce4590c3e928851..e71af553c2ed0e9294a2a8279f5448ea4da768a6 100644 (file)
@@ -44,6 +44,17 @@ struct nfsd_fcache_bucket {
 
 static DEFINE_PER_CPU(unsigned long, nfsd_file_cache_hits);
 
+struct nfsd_fcache_disposal {
+       struct list_head list;
+       struct work_struct work;
+       struct net *net;
+       spinlock_t lock;
+       struct list_head freeme;
+       struct rcu_head rcu;
+};
+
+struct workqueue_struct *nfsd_filecache_wq __read_mostly;
+
 static struct kmem_cache               *nfsd_file_slab;
 static struct kmem_cache               *nfsd_file_mark_slab;
 static struct nfsd_fcache_bucket       *nfsd_file_hashtbl;
@@ -52,32 +63,21 @@ static long                         nfsd_file_lru_flags;
 static struct fsnotify_group           *nfsd_file_fsnotify_group;
 static atomic_long_t                   nfsd_filecache_count;
 static struct delayed_work             nfsd_filecache_laundrette;
+static DEFINE_SPINLOCK(laundrette_lock);
+static LIST_HEAD(laundrettes);
 
-enum nfsd_file_laundrette_ctl {
-       NFSD_FILE_LAUNDRETTE_NOFLUSH = 0,
-       NFSD_FILE_LAUNDRETTE_MAY_FLUSH
-};
+static void nfsd_file_gc(void);
 
 static void
-nfsd_file_schedule_laundrette(enum nfsd_file_laundrette_ctl ctl)
+nfsd_file_schedule_laundrette(void)
 {
        long count = atomic_long_read(&nfsd_filecache_count);
 
        if (count == 0 || test_bit(NFSD_FILE_SHUTDOWN, &nfsd_file_lru_flags))
                return;
 
-       /* Be more aggressive about scanning if over the threshold */
-       if (count > NFSD_FILE_LRU_THRESHOLD)
-               mod_delayed_work(system_wq, &nfsd_filecache_laundrette, 0);
-       else
-               schedule_delayed_work(&nfsd_filecache_laundrette, NFSD_LAUNDRETTE_DELAY);
-
-       if (ctl == NFSD_FILE_LAUNDRETTE_NOFLUSH)
-               return;
-
-       /* ...and don't delay flushing if we're out of control */
-       if (count >= NFSD_FILE_LRU_LIMIT)
-               flush_delayed_work(&nfsd_filecache_laundrette);
+       queue_delayed_work(system_wq, &nfsd_filecache_laundrette,
+                       NFSD_LAUNDRETTE_DELAY);
 }
 
 static void
@@ -312,7 +312,9 @@ nfsd_file_put(struct nfsd_file *nf)
 
        set_bit(NFSD_FILE_REFERENCED, &nf->nf_flags);
        if (nfsd_file_put_noref(nf) == 1 && is_hashed && unused)
-               nfsd_file_schedule_laundrette(NFSD_FILE_LAUNDRETTE_MAY_FLUSH);
+               nfsd_file_schedule_laundrette();
+       if (atomic_long_read(&nfsd_filecache_count) >= NFSD_FILE_LRU_LIMIT)
+               nfsd_file_gc();
 }
 
 struct nfsd_file *
@@ -353,6 +355,58 @@ nfsd_file_dispose_list_sync(struct list_head *dispose)
                flush_delayed_fput();
 }
 
+static void
+nfsd_file_list_remove_disposal(struct list_head *dst,
+               struct nfsd_fcache_disposal *l)
+{
+       spin_lock(&l->lock);
+       list_splice_init(&l->freeme, dst);
+       spin_unlock(&l->lock);
+}
+
+static void
+nfsd_file_list_add_disposal(struct list_head *files, struct net *net)
+{
+       struct nfsd_fcache_disposal *l;
+
+       rcu_read_lock();
+       list_for_each_entry_rcu(l, &laundrettes, list) {
+               if (l->net == net) {
+                       spin_lock(&l->lock);
+                       list_splice_tail_init(files, &l->freeme);
+                       spin_unlock(&l->lock);
+                       queue_work(nfsd_filecache_wq, &l->work);
+                       break;
+               }
+       }
+       rcu_read_unlock();
+}
+
+static void
+nfsd_file_list_add_pernet(struct list_head *dst, struct list_head *src,
+               struct net *net)
+{
+       struct nfsd_file *nf, *tmp;
+
+       list_for_each_entry_safe(nf, tmp, src, nf_lru) {
+               if (nf->nf_net == net)
+                       list_move_tail(&nf->nf_lru, dst);
+       }
+}
+
+static void
+nfsd_file_dispose_list_delayed(struct list_head *dispose)
+{
+       LIST_HEAD(list);
+       struct nfsd_file *nf;
+
+       while(!list_empty(dispose)) {
+               nf = list_first_entry(dispose, struct nfsd_file, nf_lru);
+               nfsd_file_list_add_pernet(&list, dispose, nf->nf_net);
+               nfsd_file_list_add_disposal(&list, nf->nf_net);
+       }
+}
+
 /*
  * Note this can deadlock with nfsd_file_cache_purge.
  */
@@ -399,17 +453,40 @@ out_skip:
        return LRU_SKIP;
 }
 
-static void
-nfsd_file_lru_dispose(struct list_head *head)
+static unsigned long
+nfsd_file_lru_walk_list(struct shrink_control *sc)
 {
+       LIST_HEAD(head);
        struct nfsd_file *nf;
+       unsigned long ret;
 
-       list_for_each_entry(nf, head, nf_lru) {
+       if (sc)
+               ret = list_lru_shrink_walk(&nfsd_file_lru, sc,
+                               nfsd_file_lru_cb, &head);
+       else
+               ret = list_lru_walk(&nfsd_file_lru,
+                               nfsd_file_lru_cb,
+                               &head, LONG_MAX);
+       list_for_each_entry(nf, &head, nf_lru) {
                spin_lock(&nfsd_file_hashtbl[nf->nf_hashval].nfb_lock);
                nfsd_file_do_unhash(nf);
                spin_unlock(&nfsd_file_hashtbl[nf->nf_hashval].nfb_lock);
        }
-       nfsd_file_dispose_list(head);
+       nfsd_file_dispose_list_delayed(&head);
+       return ret;
+}
+
+static void
+nfsd_file_gc(void)
+{
+       nfsd_file_lru_walk_list(NULL);
+}
+
+static void
+nfsd_file_gc_worker(struct work_struct *work)
+{
+       nfsd_file_gc();
+       nfsd_file_schedule_laundrette();
 }
 
 static unsigned long
@@ -421,12 +498,7 @@ nfsd_file_lru_count(struct shrinker *s, struct shrink_control *sc)
 static unsigned long
 nfsd_file_lru_scan(struct shrinker *s, struct shrink_control *sc)
 {
-       LIST_HEAD(head);
-       unsigned long ret;
-
-       ret = list_lru_shrink_walk(&nfsd_file_lru, sc, nfsd_file_lru_cb, &head);
-       nfsd_file_lru_dispose(&head);
-       return ret;
+       return nfsd_file_lru_walk_list(sc);
 }
 
 static struct shrinker nfsd_file_shrinker = {
@@ -488,7 +560,7 @@ nfsd_file_close_inode(struct inode *inode)
 
        __nfsd_file_close_inode(inode, hashval, &dispose);
        trace_nfsd_file_close_inode(inode, hashval, !list_empty(&dispose));
-       nfsd_file_dispose_list(&dispose);
+       nfsd_file_dispose_list_delayed(&dispose);
 }
 
 /**
@@ -504,16 +576,11 @@ static void
 nfsd_file_delayed_close(struct work_struct *work)
 {
        LIST_HEAD(head);
+       struct nfsd_fcache_disposal *l = container_of(work,
+                       struct nfsd_fcache_disposal, work);
 
-       list_lru_walk(&nfsd_file_lru, nfsd_file_lru_cb, &head, LONG_MAX);
-
-       if (test_and_clear_bit(NFSD_FILE_LRU_RESCAN, &nfsd_file_lru_flags))
-               nfsd_file_schedule_laundrette(NFSD_FILE_LAUNDRETTE_NOFLUSH);
-
-       if (!list_empty(&head)) {
-               nfsd_file_lru_dispose(&head);
-               flush_delayed_fput();
-       }
+       nfsd_file_list_remove_disposal(&head, l);
+       nfsd_file_dispose_list(&head);
 }
 
 static int
@@ -574,6 +641,10 @@ nfsd_file_cache_init(void)
        if (nfsd_file_hashtbl)
                return 0;
 
+       nfsd_filecache_wq = alloc_workqueue("nfsd_filecache", 0, 0);
+       if (!nfsd_filecache_wq)
+               goto out;
+
        nfsd_file_hashtbl = kcalloc(NFSD_FILE_HASH_SIZE,
                                sizeof(*nfsd_file_hashtbl), GFP_KERNEL);
        if (!nfsd_file_hashtbl) {
@@ -627,7 +698,7 @@ nfsd_file_cache_init(void)
                spin_lock_init(&nfsd_file_hashtbl[i].nfb_lock);
        }
 
-       INIT_DELAYED_WORK(&nfsd_filecache_laundrette, nfsd_file_delayed_close);
+       INIT_DELAYED_WORK(&nfsd_filecache_laundrette, nfsd_file_gc_worker);
 out:
        return ret;
 out_notifier:
@@ -643,6 +714,8 @@ out_err:
        nfsd_file_mark_slab = NULL;
        kfree(nfsd_file_hashtbl);
        nfsd_file_hashtbl = NULL;
+       destroy_workqueue(nfsd_filecache_wq);
+       nfsd_filecache_wq = NULL;
        goto out;
 }
 
@@ -681,6 +754,88 @@ nfsd_file_cache_purge(struct net *net)
        }
 }
 
+static struct nfsd_fcache_disposal *
+nfsd_alloc_fcache_disposal(struct net *net)
+{
+       struct nfsd_fcache_disposal *l;
+
+       l = kmalloc(sizeof(*l), GFP_KERNEL);
+       if (!l)
+               return NULL;
+       INIT_WORK(&l->work, nfsd_file_delayed_close);
+       l->net = net;
+       spin_lock_init(&l->lock);
+       INIT_LIST_HEAD(&l->freeme);
+       return l;
+}
+
+static void
+nfsd_free_fcache_disposal(struct nfsd_fcache_disposal *l)
+{
+       rcu_assign_pointer(l->net, NULL);
+       cancel_work_sync(&l->work);
+       nfsd_file_dispose_list(&l->freeme);
+       kfree_rcu(l, rcu);
+}
+
+static void
+nfsd_add_fcache_disposal(struct nfsd_fcache_disposal *l)
+{
+       spin_lock(&laundrette_lock);
+       list_add_tail_rcu(&l->list, &laundrettes);
+       spin_unlock(&laundrette_lock);
+}
+
+static void
+nfsd_del_fcache_disposal(struct nfsd_fcache_disposal *l)
+{
+       spin_lock(&laundrette_lock);
+       list_del_rcu(&l->list);
+       spin_unlock(&laundrette_lock);
+}
+
+static int
+nfsd_alloc_fcache_disposal_net(struct net *net)
+{
+       struct nfsd_fcache_disposal *l;
+
+       l = nfsd_alloc_fcache_disposal(net);
+       if (!l)
+               return -ENOMEM;
+       nfsd_add_fcache_disposal(l);
+       return 0;
+}
+
+static void
+nfsd_free_fcache_disposal_net(struct net *net)
+{
+       struct nfsd_fcache_disposal *l;
+
+       rcu_read_lock();
+       list_for_each_entry_rcu(l, &laundrettes, list) {
+               if (l->net != net)
+                       continue;
+               nfsd_del_fcache_disposal(l);
+               rcu_read_unlock();
+               nfsd_free_fcache_disposal(l);
+               return;
+       }
+       rcu_read_unlock();
+}
+
+int
+nfsd_file_cache_start_net(struct net *net)
+{
+       return nfsd_alloc_fcache_disposal_net(net);
+}
+
+void
+nfsd_file_cache_shutdown_net(struct net *net)
+{
+       nfsd_file_cache_purge(net);
+       nfsd_free_fcache_disposal_net(net);
+}
+
 void
 nfsd_file_cache_shutdown(void)
 {
@@ -705,6 +860,8 @@ nfsd_file_cache_shutdown(void)
        nfsd_file_mark_slab = NULL;
        kfree(nfsd_file_hashtbl);
        nfsd_file_hashtbl = NULL;
+       destroy_workqueue(nfsd_filecache_wq);
+       nfsd_filecache_wq = NULL;
 }
 
 static bool
@@ -872,7 +1029,8 @@ open_file:
        nfsd_file_hashtbl[hashval].nfb_maxcount = max(nfsd_file_hashtbl[hashval].nfb_maxcount,
                        nfsd_file_hashtbl[hashval].nfb_count);
        spin_unlock(&nfsd_file_hashtbl[hashval].nfb_lock);
-       atomic_long_inc(&nfsd_filecache_count);
+       if (atomic_long_inc_return(&nfsd_filecache_count) >= NFSD_FILE_LRU_THRESHOLD)
+               nfsd_file_gc();
 
        nf->nf_mark = nfsd_file_mark_find_or_create(nf);
        if (nf->nf_mark)
index 851d9abf54c25a35638a7e63ba0350de9ea3664e..79a7d6808d979970d33a41509d3587b8496ccb22 100644 (file)
@@ -51,6 +51,8 @@ struct nfsd_file {
 int nfsd_file_cache_init(void);
 void nfsd_file_cache_purge(struct net *);
 void nfsd_file_cache_shutdown(void);
+int nfsd_file_cache_start_net(struct net *net);
+void nfsd_file_cache_shutdown_net(struct net *net);
 void nfsd_file_put(struct nfsd_file *nf);
 struct nfsd_file *nfsd_file_get(struct nfsd_file *nf);
 void nfsd_file_close_inode_sync(struct inode *inode);
index 7900f3494ecb72cb180a04c2915328e906ca4af3..3b77b904212d8973fde5ab9baf411ea638ffba53 100644 (file)
@@ -400,13 +400,18 @@ static int nfsd_startup_net(int nrservs, struct net *net, const struct cred *cre
                nn->lockd_up = true;
        }
 
-       ret = nfs4_state_start_net(net);
+       ret = nfsd_file_cache_start_net(net);
        if (ret)
                goto out_lockd;
+       ret = nfs4_state_start_net(net);
+       if (ret)
+               goto out_filecache;
 
        nn->nfsd_net_up = true;
        return 0;
 
+out_filecache:
+       nfsd_file_cache_shutdown_net(net);
 out_lockd:
        if (nn->lockd_up) {
                lockd_down(net);
@@ -421,7 +426,7 @@ static void nfsd_shutdown_net(struct net *net)
 {
        struct nfsd_net *nn = net_generic(net, nfsd_net_id);
 
-       nfsd_file_cache_purge(net);
+       nfsd_file_cache_shutdown_net(net);
        nfs4_state_shutdown_net(net);
        if (nn->lockd_up) {
                lockd_down(net);