[lkml]   [2011]   [Jul]   [8]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
SubjectRe: [RFC] sunrpc: Fix race between work-queue and rpc_killall_tasks.
On 07/08/2011 11:11 AM, Myklebust, Trond wrote:
>> -----Original Message-----
>> From: Ben Greear []
>> Sent: Friday, July 08, 2011 1:19 PM
>> To: Myklebust, Trond
>> Cc:;
>> Subject: Re: [RFC] sunrpc: Fix race between work-queue and
>> rpc_killall_tasks.
>> On 07/06/2011 04:45 PM, Trond Myklebust wrote:
>>> On Wed, 2011-07-06 at 15:49 -0700, wrote:
>>>> From: Ben Greear<>
>>>> The rpc_killall_tasks logic is not locked against
>>>> the work-queue thread, but it still directly modifies
>>>> function pointers and data in the task objects.
>>>> This patch changes the killall-tasks logic to set a flag
>>>> that tells the work-queue thread to terminate the task
>>>> instead of directly calling the terminate logic.
>>>> Signed-off-by: Ben Greear<>
>>>> ---
>>>> NOTE: This needs review, as I am still struggling to understand
>>>> the rpc code, and it's quite possible this patch either doesn't
>>>> fully fix the problem or actually causes other issues. That said,
>>>> my nfs stress test seems to run a bit more stable with this patch
>> applied.
>>> Yes, but I don't see why you are adding a new flag, nor do I see why
>> we
>>> want to keep checking for that flag in the rpc_execute() loop.
>>> rpc_killall_tasks() is not a frequent operation that we want to
>> optimise
>>> for.
>>> How about the following instead?
>> Ok, I looked at your patch closer. I think it can still cause
>> bad race conditions.
>> For instance:
>> Assume that tk_callback is NULL at beginning of while loop in
>> __rpc_execute,
>> and tk_action is rpc_exit_task.
>> While do_action(task) is being called, tk_action is set to NULL in
>> rpc_exit_task.
>> But, right after tk_action is set to NULL in rpc_exit_task, the
>> rpc_killall_tasks
>> method calls rpc_exit, which sets tk_action back to rpc_exit_task.
>> I believe this could cause the xprt_release(task) logic to be called in
>> the
>> work-queue's execution of rpc_exit_task due to tk_action != NULL when
>> it should not be.
> Why would this be a problem? xprt_release() can certainly be called multiple times on an rpc_task. Ditto rpbc_getport_done.
> The only thing that is not re-entrant there is rpcb_map_release, which should only ever be called once whether or not something calls rpc_killall_tasks.

From the trace I posted, this stack trace below is being
called with the void *data object already freed.

One way for this to happen would be to have rpc_exit_task call task->tk_ops->rpc_call_done
more than once (I believe). Two calls to rpc_exit_task could do that, and since the
rpc_exit_task method is assigned to tk_action, I *think* the race I mention above could cause
rpc_exit_task to be called twice.

[<ffffffff81105907>] print_trailer+0x131/0x13a
[<ffffffff81105945>] object_err+0x35/0x3e
[<ffffffff811077b3>] verify_mem_not_deleted+0x7a/0xb7
[<ffffffffa02891e5>] rpcb_getport_done+0x23/0x126 [sunrpc]
[<ffffffffa02810df>] rpc_exit_task+0x3f/0x6d [sunrpc]
[<ffffffffa02814d8>] __rpc_execute+0x80/0x253 [sunrpc]
[<ffffffffa02816ed>] ? rpc_execute+0x42/0x42 [sunrpc]
[<ffffffffa02816fd>] rpc_async_schedule+0x10/0x12 [sunrpc]
[<ffffffff81061343>] process_one_work+0x230/0x41d
[<ffffffff8106128e>] ? process_one_work+0x17b/0x41d
[<ffffffff8106379f>] worker_thread+0x133/0x217
[<ffffffff8106366c>] ? manage_workers+0x191/0x191
[<ffffffff81066f9c>] kthread+0x7d/0x85
[<ffffffff81485ee4>] kernel_thread_helper+0x4/0x10
[<ffffffff8147f0d8>] ? retint_restore_args+0x13/0x13
[<ffffffff81066f1f>] ? __init_kthread_worker+0x56/0x56
[<ffffffff81485ee0>] ? gs_change+0x13/0x13


Ben Greear <>
Candela Technologies Inc

 \ /
  Last update: 2011-07-09 00:05    [W:0.047 / U:0.080 seconds]
©2003-2018 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site