NFS: nfs_lock_and_join_requests and nfs_scan_commit_list can deadlock
authorTrond Myklebust <trond.myklebust@primarydata.com>
Sat, 9 Sep 2017 16:28:01 +0000 (12:28 -0400)
committerTrond Myklebust <trond.myklebust@primarydata.com>
Sat, 9 Sep 2017 16:28:01 +0000 (12:28 -0400)
Since the commit list is not ordered, it is possible for nfs_scan_commit_list
to hold a request that nfs_lock_and_join_requests() is waiting for, while
at the same time trying to grab a request that nfs_lock_and_join_requests
already holds.

Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
fs/nfs/pnfs_nfs.c
fs/nfs/write.c

index 303ff17..d03d836 100644 (file)
@@ -91,22 +91,28 @@ static int
 pnfs_generic_transfer_commit_list(struct list_head *src, struct list_head *dst,
                                  struct nfs_commit_info *cinfo, int max)
 {
-       struct nfs_page *req;
+       struct nfs_page *req, *tmp;
        int ret = 0;
 
-       while(!list_empty(src)) {
-               req = list_first_entry(src, struct nfs_page, wb_list);
-
+restart:
+       list_for_each_entry_safe(req, tmp, src, wb_list) {
                kref_get(&req->wb_kref);
                if (!nfs_lock_request(req)) {
                        int status;
+
+                       /* Prevent deadlock with nfs_lock_and_join_requests */
+                       if (!list_empty(dst)) {
+                               nfs_release_request(req);
+                               continue;
+                       }
+                       /* Ensure we make progress to prevent livelock */
                        mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
                        status = nfs_wait_on_request(req);
                        nfs_release_request(req);
                        mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
                        if (status < 0)
                                break;
-                       continue;
+                       goto restart;
                }
                nfs_request_remove_commit_list(req, cinfo);
                clear_bit(PG_COMMIT_TO_DS, &req->wb_flags);
index ae26775..c3f627b 100644 (file)
@@ -1028,21 +1028,28 @@ int
 nfs_scan_commit_list(struct list_head *src, struct list_head *dst,
                     struct nfs_commit_info *cinfo, int max)
 {
-       struct nfs_page *req;
+       struct nfs_page *req, *tmp;
        int ret = 0;
 
-       while(!list_empty(src)) {
-               req = list_first_entry(src, struct nfs_page, wb_list);
+restart:
+       list_for_each_entry_safe(req, tmp, src, wb_list) {
                kref_get(&req->wb_kref);
                if (!nfs_lock_request(req)) {
                        int status;
+
+                       /* Prevent deadlock with nfs_lock_and_join_requests */
+                       if (!list_empty(dst)) {
+                               nfs_release_request(req);
+                               continue;
+                       }
+                       /* Ensure we make progress to prevent livelock */
                        mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
                        status = nfs_wait_on_request(req);
                        nfs_release_request(req);
                        mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
                        if (status < 0)
                                break;
-                       continue;
+                       goto restart;
                }
                nfs_request_remove_commit_list(req, cinfo);
                nfs_list_add_request(req, dst);