NFSD: Leave open files out of the filecache LRU

There have been reports of problems when running fstests generic/531
against Linux NFS servers with NFSv4. The NFS server that hosts the
test's SCRATCH_DEV suffers from CPU soft lock-ups during the test.
Analysis shows that:

fs/nfsd/filecache.c
 482                 ret = list_lru_walk(&nfsd_file_lru,
 483                                 nfsd_file_lru_cb,
 484                                 &head, LONG_MAX);

causes nfsd_file_gc() to walk the entire length of the filecache LRU
list every time it is called (which is quite frequently). The walk
holds a spinlock the entire time that prevents other nfsd threads
from accessing the filecache.

What's more, for NFSv4 workloads, none of the items that are visited
during this walk may be evicted, since they are all files that are
held OPEN by NFS clients.

Address this by ensuring that open files are not kept on the LRU
list.

Reported-by: Frank van der Linden <fllinden@amazon.com>
Reported-by: Wang Yugui <wangyugui@e16-tech.com>
Link: https://bugzilla.linux-nfs.org/show_bug.cgi?id=386
Suggested-by: Trond Myklebust <trond.myklebust@hammerspace.com>
Reviewed-by: Jeff Layton <jlayton@kernel.org>
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>

2 files changed, 21 insertions, 5 deletions

diff --git a/fs/nfsd/filecache.c b/fs/nfsd/filecache.c
index e831a7b4f916..69d1ab6cfc14 100644
--- a/fs/nfsd/filecache.c
+++ b/fs/nfsd/filecache.c
@@ -263,6 +263,7 @@ nfsd_file_flush(struct nfsd_file *nf)
 
 static void nfsd_file_lru_add(struct nfsd_file *nf)
 {
+	set_bit(NFSD_FILE_REFERENCED, &nf->nf_flags);
 	if (list_lru_add(&nfsd_file_lru, &nf->nf_lru))
 		trace_nfsd_file_lru_add(nf);
 }
@@ -292,7 +293,6 @@ nfsd_file_unhash(struct nfsd_file *nf)
 {
 	if (test_and_clear_bit(NFSD_FILE_HASHED, &nf->nf_flags)) {
 		nfsd_file_do_unhash(nf);
-		nfsd_file_lru_remove(nf);
 		return true;
 	}
 	return false;
@@ -313,6 +313,7 @@ nfsd_file_unhash_and_release_locked(struct nfsd_file *nf, struct list_head *disp
 	if (refcount_dec_not_one(&nf->nf_ref))
 		return true;
 
+	nfsd_file_lru_remove(nf);
 	list_add(&nf->nf_lru, dispose);
 	return true;
 }
@@ -324,6 +325,7 @@ nfsd_file_put_noref(struct nfsd_file *nf)
 
 	if (refcount_dec_and_test(&nf->nf_ref)) {
 		WARN_ON(test_bit(NFSD_FILE_HASHED, &nf->nf_flags));
+		nfsd_file_lru_remove(nf);
 		nfsd_file_free(nf);
 	}
 }
@@ -333,7 +335,7 @@ nfsd_file_put(struct nfsd_file *nf)
 {
 	might_sleep();
 
-	set_bit(NFSD_FILE_REFERENCED, &nf->nf_flags);
+	nfsd_file_lru_add(nf);
 	if (test_bit(NFSD_FILE_HASHED, &nf->nf_flags) == 0) {
 		nfsd_file_flush(nf);
 		nfsd_file_put_noref(nf);
@@ -433,8 +435,18 @@ nfsd_file_dispose_list_delayed(struct list_head *dispose)
 	}
 }
 
-/*
+/**
+ * nfsd_file_lru_cb - Examine an entry on the LRU list
+ * @item: LRU entry to examine
+ * @lru: controlling LRU
+ * @lock: LRU list lock (unused)
+ * @arg: dispose list
+ *
  * Note this can deadlock with nfsd_file_cache_purge.
+ *
+ * Return values:
+ *   %LRU_REMOVED: @item was removed from the LRU
+ *   %LRU_SKIP: @item cannot be evicted
  */
 static enum lru_status
 nfsd_file_lru_cb(struct list_head *item, struct list_lru_one *lru,
@@ -456,8 +468,9 @@ nfsd_file_lru_cb(struct list_head *item, struct list_lru_one *lru,
 	 * That order is deliberate to ensure that we can do this locklessly.
 	 */
 	if (refcount_read(&nf->nf_ref) > 1) {
+		list_lru_isolate(lru, &nf->nf_lru);
 		trace_nfsd_file_gc_in_use(nf);
-		return LRU_SKIP;
+		return LRU_REMOVED;
 	}
 
 	/*
@@ -1014,6 +1027,7 @@ wait_for_construction:
 		goto retry;
 	}
 
+	nfsd_file_lru_remove(nf);
 	this_cpu_inc(nfsd_file_cache_hits);
 
 	status = nfserrno(nfsd_open_break_lease(file_inode(nf->nf_file), may_flags));
@@ -1035,7 +1049,6 @@ open_file:
 	refcount_inc(&nf->nf_ref);
 	__set_bit(NFSD_FILE_HASHED, &nf->nf_flags);
 	__set_bit(NFSD_FILE_PENDING, &nf->nf_flags);
-	nfsd_file_lru_add(nf);
 	hlist_add_head_rcu(&nf->nf_node, &nfsd_file_hashtbl[hashval].nfb_head);
 	++nfsd_file_hashtbl[hashval].nfb_count;
 	nfsd_file_hashtbl[hashval].nfb_maxcount = max(nfsd_file_hashtbl[hashval].nfb_maxcount,
@@ -1060,6 +1073,7 @@ open_file:
 	 */
 	if (status != nfs_ok || inode->i_nlink == 0) {
 		bool do_free;
+		nfsd_file_lru_remove(nf);
 		spin_lock(&nfsd_file_hashtbl[hashval].nfb_lock);
 		do_free = nfsd_file_unhash(nf);
 		spin_unlock(&nfsd_file_hashtbl[hashval].nfb_lock);
diff --git a/fs/nfsd/trace.h b/fs/nfsd/trace.h
index 64a6c577d7b9..9507a6c8eeee 100644
--- a/fs/nfsd/trace.h
+++ b/fs/nfsd/trace.h
@@ -927,7 +927,9 @@ DEFINE_EVENT(nfsd_file_gc_class, name,					\
 	TP_ARGS(nf))
 
 DEFINE_NFSD_FILE_GC_EVENT(nfsd_file_lru_add);
+DEFINE_NFSD_FILE_GC_EVENT(nfsd_file_lru_add_disposed);
 DEFINE_NFSD_FILE_GC_EVENT(nfsd_file_lru_del);
+DEFINE_NFSD_FILE_GC_EVENT(nfsd_file_lru_del_disposed);
 DEFINE_NFSD_FILE_GC_EVENT(nfsd_file_gc_in_use);
 DEFINE_NFSD_FILE_GC_EVENT(nfsd_file_gc_writeback);
 DEFINE_NFSD_FILE_GC_EVENT(nfsd_file_gc_referenced);