The rpciod workqueue is on the write-out path for freeing dirty memory, so it is important that it never block waiting for memory to be allocated - this can lead to a deadlock. rpc_execute() - which is often called by an rpciod work item - calls rcp_task_release_client() which can lead to rpc_free_client(). rpc_free_client() makes two calls which could potentially block wating for memory allocation. rpc_clnt_debugfs_unregister() calls into debugfs and will block while any of the debugfs files are being accessed. In particular it can block while any of the 'open' methods are being called and all of these use malloc for one thing or another. So this can deadlock if the memory allocation waits for NFS to complete some writes via rpciod. rpc_clnt_remove_pipedir() can take the inode_lock() and while it isn't obvious that memory allocations can happen while the lock it held, it is safer to assume they might and to not let rpciod call rpc_clnt_remove_pipedir(). So this patch moves these two calls (together with the final kfree() and rpciod_down()) into a work-item to be run from the system work-queue. rpciod can continue its important work, and the final stages of the free can happen whenever they happen. I have seen this deadlock on a 4.12 based kernel where debugfs used synchronize_srcu() when removing objects. synchronize_srcu() requires a workqueue and there were no free workther threads and none could be allocated. While debugsfs no longer uses SRCU, I believe the deadlock is still possible. Signed-off-by: NeilBrown <neilb@xxxxxxx> --- include/linux/sunrpc/clnt.h | 8 +++++++- net/sunrpc/clnt.c | 21 +++++++++++++++++---- 2 files changed, 24 insertions(+), 5 deletions(-) diff --git a/include/linux/sunrpc/clnt.h b/include/linux/sunrpc/clnt.h index ca7e108248e2..7bd124e06b36 100644 --- a/include/linux/sunrpc/clnt.h +++ b/include/linux/sunrpc/clnt.h @@ -71,7 +71,13 @@ struct rpc_clnt { #if IS_ENABLED(CONFIG_SUNRPC_DEBUG) struct dentry *cl_debugfs; /* debugfs directory */ #endif - struct rpc_xprt_iter cl_xpi; + /* cl_work is only needed after cl_xpi is no longer used, + * and that are of similar size + */ + union { + struct rpc_xprt_iter cl_xpi; + struct work_struct cl_work; + }; const struct cred *cl_cred; }; diff --git a/net/sunrpc/clnt.c b/net/sunrpc/clnt.c index 7324b21f923e..a2c215a6980d 100644 --- a/net/sunrpc/clnt.c +++ b/net/sunrpc/clnt.c @@ -880,6 +880,20 @@ EXPORT_SYMBOL_GPL(rpc_shutdown_client); /* * Free an RPC client */ +static void rpc_free_client_work(struct work_struct *work) +{ + struct rpc_clnt *clnt = container_of(work, struct rpc_clnt, cl_work); + + /* These might block on processes that might allocate memory, + * so they cannot be called in rpciod, so they are handled separately + * here. + */ + rpc_clnt_debugfs_unregister(clnt); + rpc_clnt_remove_pipedir(clnt); + + kfree(clnt); + rpciod_down(); +} static struct rpc_clnt * rpc_free_client(struct rpc_clnt *clnt) { @@ -890,17 +904,16 @@ rpc_free_client(struct rpc_clnt *clnt) rcu_dereference(clnt->cl_xprt)->servername); if (clnt->cl_parent != clnt) parent = clnt->cl_parent; - rpc_clnt_debugfs_unregister(clnt); - rpc_clnt_remove_pipedir(clnt); rpc_unregister_client(clnt); rpc_free_iostats(clnt->cl_metrics); clnt->cl_metrics = NULL; xprt_put(rcu_dereference_raw(clnt->cl_xprt)); xprt_iter_destroy(&clnt->cl_xpi); - rpciod_down(); put_cred(clnt->cl_cred); rpc_free_clid(clnt); - kfree(clnt); + + INIT_WORK(&clnt->cl_work, rpc_free_client_work); + schedule_work(&clnt->cl_work); return parent; } -- 2.26.0
Attachment:
signature.asc
Description: PGP signature