Re: [PATCH v3 2/3] nfs4: handle async processing of F_SETLK with FL_SLEEP

[Vasily Averin]

On 16.01.2022 15:44, Vasily Averin wrote:
> On 03.01.2022 22:53, J. Bruce Fields wrote:
>> On Wed, Dec 29, 2021 at 11:24:43AM +0300, Vasily Averin wrote:
>>> nfsd and lockd use F_SETLK cmd with the FL_SLEEP flag set to request
>>> asynchronous processing of blocking locks.
>>>
>>> Currently nfs4 use locks_lock_inode_wait() function which is blocked
>>> for such requests. To handle them correctly FL_SLEEP flag should be
>>> temporarily reset before executing the locks_lock_inode_wait() function.
>>>
>>> Additionally block flag is forced to set, to translate blocking lock to
>>> remote nfs server, expecting it supports async processing of the blocking
>>> locks too.
>>
>> But this on its own isn't enough for the client to support asynchronous
>> blocking locks, right?  Don't we also need the logic that calls knfsd's
>> lm_notify when it gets a CB_NOTIFY_LOCK from the server?
> 
> No, I think this should be enough.
> We are here a nfs client,
> we can get F_SETLK with FL_SLEEP from nfsd only (i.e. in re-export case)
> we need to avoid blocking if lock is already taken, 
> so we need to call locks_lock_inode_wait without FL_SLEEP,
> then we submit _sleeping_ request to NFS server (i.e. set )data->arg.block = 1)
> and waiting for reply from server.
> 
> Here we rely that server will NOT block on such request too, so our reply wel not be blocked too.

Now I think this assumption is wrong.
We cannot guarantee that NFS server will process our sleeping request asynchronously.
yes, new version of knfsd will do it.
however there are a lot of other NFS servers, that can process this request synchronously and wait till locked fail will be unlocked.

All we can do here is just drop FL_SLEEP and handle incoming async request (F_SETLK with FL_SLEEP) like a regular non-blocking F_SETLK.
Thank you,
	Vasily Averin

> Under "block" I mean that handler can sleep or process request for a very long time 
> but it will NOT BE BLOCKED if lock is taken already, it WILL NOT WAIT when lock will be released,
> it just return some error in this case.
> 
> I think it is correct.
> Do you think I am wrong or maybe I missed something? 
> 
> Thank you,
> 	Vasily Averin
> 
> However I noticed now that past is incorrect, 
> temporally dropped FL_SLEEP should be restored back in _nfs4_proc_setlk before _nfs4_do_setlk() call.
> I'll fix it in next version of this patch-set.
> 
>>> https://bugzilla.kernel.org/show_bug.cgi?id=215383
>>> Signed-off-by: Vasily Averin <vvs@xxxxxxxxxxxxx>
>>> ---
>>>  fs/nfs/nfs4proc.c | 5 ++++-
>>>  1 file changed, 4 insertions(+), 1 deletion(-)
>>>
>>> diff --git a/fs/nfs/nfs4proc.c b/fs/nfs/nfs4proc.c
>>> index ee3bc79f6ca3..9b1380c4223c 100644
>>> --- a/fs/nfs/nfs4proc.c
>>> +++ b/fs/nfs/nfs4proc.c
>>> @@ -7094,7 +7094,7 @@ static int _nfs4_do_setlk(struct nfs4_state *state, int cmd, struct file_lock *f
>>>  			recovery_type == NFS_LOCK_NEW ? GFP_KERNEL : GFP_NOFS);
>>>  	if (data == NULL)
>>>  		return -ENOMEM;
>>> -	if (IS_SETLKW(cmd))
>>> +	if (IS_SETLKW(cmd) || (fl->fl_flags & FL_SLEEP))
>>>  		data->arg.block = 1;
>>>  	nfs4_init_sequence(&data->arg.seq_args, &data->res.seq_res, 1,
>>>  				recovery_type > NFS_LOCK_NEW);
>>> @@ -7200,6 +7200,9 @@ static int _nfs4_proc_setlk(struct nfs4_state *state, int cmd, struct file_lock
>>>  	int status;
>>>  
>>>  	request->fl_flags |= FL_ACCESS;
>>> +	if (((fl_flags & FL_SLEEP_POSIX) == FL_SLEEP_POSIX) && IS_SETLK(cmd))
>>> +		request->fl_flags &= ~FL_SLEEP;
>>> +
>>>  	status = locks_lock_inode_wait(state->inode, request);
>>>  	if (status < 0)
>>>  		goto out;
>>> -- 
>>> 2.25.1
>