Forced unmount - fail future RPCs too
Subsystems (3)
PATCH 1/1 Forced unmount - fail future RPCs too
Modify NFS forced unmount behavior to not only return -EIO for all currently scheduled RPCs, but to also return -EIO for future RPCs. Only a single forced unmount operation is therefore required to ensure that all application will eventually return to userspace.
---
At EMC I've been looking into handling the case when an NFS server goes down, and that's had me looking at forcibly unmounting NFS filesystems. I've found that it takes two forced unmount operations to fail most NFS filesystem operations (in particularly, read()) back to the application. This is because an NFS forced unmount operation really translates to "fail all RPC operations currently in progress".
Typically, a forced unmount during a read system call will result in cancelling an RPC that was created as part of readahead. Though the readhead RPC is cancelled, another RPC is then started to perform the read explicitly. This requires another forced unmount attempt before the read syscall returns to the userspace application. Additionally, if the read was partially successful, the syscall will return the number of bytes read. If the application attempts a read again to get the rest, two more force unmount operations will be required before the application gets an EIO error. While the application has any of these RPCs scheduled it is unkillable, even with SIGKILL.
My patch changes the behavior from "fail all RPC operations currently in progress" to "fail all current and future RPC operations".
I accomplished this by setting a bit in the RPC client, and checking that bit in the RPC finite-state machine. I put the check in the FSM to avoid needing to check every way that an RPC operation could be scheduled. It also means that I don't need to worry about synchronous versus asynchronous scheduling, etc. This was the best approach for me, given my amount of experience with the NFS code, though perhaps you would have a different approach.
With this patch, a forcible lazy unmount (MNT_FORCE|MNT_DETACH) works particularly well. It ensures that all RPC operations for the filesystem will return -EIO, and that the resources for the filesystem will be cleaned up once all applications release their associated file descriptors (and that the applications will at least occasionally drop back into userspace and handle signals if they fail to release those open files).
Being new to the NFS code, I don't know how many awful coding catastrophes I may have committed, but I hope that this feature can be considered, even if the implementation may need to modifed.
fs/nfs/super.c | 2 +- include/linux/sunrpc/clnt.h | 3 ++- net/sunrpc/sched.c | 10 ++++++++++ 3 files changed, 13 insertions(+), 2 deletions(-)
fs/nfs/super.c +1 -1
@@ -602,7 +602,7 @@ static void nfs_umount_begin(struct vfsmif (!(flags &MNT_FORCE ))return ; -/* -EIO all pending I/O */ +/* -EIO all pending and future I/O */ rpc = server->client_acl;if (!IS_ERR (rpc))rpc_killall_tasks (rpc);
include/linux/sunrpc/clnt.h +2 -1
@@ -43,7 +43,8 @@ struct rpc_clnt { unsigned int cl_softrtry : 1,/* soft timeouts */ cl_intr : 1,/* interruptible */ cl_discrtry : 1,/* disconnect before retry */ - cl_autobind : 1;/* use getport() */ + cl_autobind : 1,/* use getport() */ + cl_shutdown : 1;/* fail with -EIO */struct rpc_rtt * cl_rtt;/* RTO estimator data */
net/sunrpc/clnt.c +0 -0
net/sunrpc/sched.c +10 -0
@@ -663,6 +663,15 @@ static void __rpc_execute(struct rpc_tas }/* +* Fail with -EIO if the client has been shut down. +*/ +if (task->tk_client->cl_shutdown && + !(task->tk_flags &RPC_TASK_KILLED )) { + task->tk_flags |=RPC_TASK_KILLED ; +rpc_exit (task, -EIO ); + } + +/* * Perform the next FSM step. * tk_action may be NULL when the task has been killed * by someone else.
@@ -947,6 +956,7 @@ void rpc_killall_tasks(struct rpc_clnt ** Spin lock all_tasks to prevent changes... */ spin_lock (&clnt->cl_lock); + clnt->cl_shutdown =1 ;list_for_each_entry (rovr, &clnt->cl_tasks, tk_task) {if (!RPC_IS_ACTIVATED (rovr))continue ;
-------------------------------------------------------------------------
This SF.net email is sponsored by: Splunk Inc.
Still grepping through log files to find problems? Stop. Now Search log events and configuration files using AJAX and a browser. Download your FREE copy of Splunk now >> http://get.splunk.com/
_______________________________________________
NFS maillist - [email protected]
https://lists.sourceforge.net/lists/listinfo/nfs