Re: [PATCH RFC v16 03/11] NFSD: Add lm_lock_expired call out

[dai.ngo@oracle.com]

On 3/15/22 10:39 AM, Chuck Lever III wrote:
>
>> On Mar 15, 2022, at 12:26 PM, Dai Ngo <dai.ngo@oracle.com> wrote:
>>
>>
>> On 3/15/22 8:02 AM, J. Bruce Fields wrote:
>>> On Fri, Mar 11, 2022 at 06:13:27PM -0800, Dai Ngo wrote:
>>>> Add callout function nfsd4_lm_lock_expired for lm_lock_expired.
>>>> If lock request has conflict with courtesy client then expire the
>>>> courtesy client and return no conflict to caller.
>>>>
>>>> Signed-off-by: Dai Ngo <dai.ngo@oracle.com>
>>>> ---
>>>>   fs/nfsd/nfs4state.c | 37 +++++++++++++++++++++++++++++++++++++
>>>>   1 file changed, 37 insertions(+)
>>>>
>>>> diff --git a/fs/nfsd/nfs4state.c b/fs/nfsd/nfs4state.c
>>>> index a65d59510681..583ac807e98d 100644
>>>> --- a/fs/nfsd/nfs4state.c
>>>> +++ b/fs/nfsd/nfs4state.c
>>>> @@ -6578,10 +6578,47 @@ nfsd4_lm_notify(struct file_lock *fl)
>>>>   	}
>>>>   }
>>>>   +/**
>>>> + * nfsd4_lm_lock_expired - check if lock conflict can be resolved.
>>>> + *
>>>> + * @fl: pointer to file_lock with a potential conflict
>>>> + * Return values:
>>>> + *   %false: real conflict, lock conflict can not be resolved.
>>>> + *   %true: no conflict, lock conflict was resolved.
>>>> + *
>>>> + * Note that this function is called while the flc_lock is held.
>>>> + */
>>>> +static bool
>>>> +nfsd4_lm_lock_expired(struct file_lock *fl)
>>>> +{
>>>> +	struct nfs4_lockowner *lo;
>>>> +	struct nfs4_client *clp;
>>>> +	bool rc = false;
>>>> +
>>>> +	if (!fl)
>>>> +		return false;
>>>> +	lo = (struct nfs4_lockowner *)fl->fl_owner;
>>>> +	clp = lo->lo_owner.so_client;
>>>> +
>>>> +	/* need to sync with courtesy client trying to reconnect */
>>>> +	spin_lock(&clp->cl_cs_lock);
>>>> +	if (test_bit(NFSD4_CLIENT_EXPIRED, &clp->cl_flags))
>>>> +		rc = true;
>>>> +	else {
>>>> +		if (test_bit(NFSD4_CLIENT_COURTESY, &clp->cl_flags)) {
>>>> +			set_bit(NFSD4_CLIENT_EXPIRED, &clp->cl_flags);
>>>> +			rc =  true;
>>>> +		}
>>>> +	}
>>> I'd prefer:
>>>
>>> 	if (test_bit(NFSD4_CLIENT_COURTESY, &clp->cl_flags))
>>> 		set_bit(NFSD4_CLIENT_EXPIRED, &clp->cl_flags);
>> we also need to set rc to true here.
>>
>>> 	if (test_bit(NFSD4_CLIENT_EXPIRED, &clp->cl_flags))
>>> 		rc = true;
>> With v16 we need to check for NFSD4_CLIENT_EXPIRED first then
>> NFSD4_CLIENT_COURTESY because both flags can be set. In the
>> next patch version, we will clear NFSD4_CLIENT_COURTESY when
>> setting NFSD4_CLIENT_EXPIRED so the order of check does not
>> matter.
>>
>>> Same result, but more compact and straightforward, I think.
>> Chuck wants to replace the bits used for courtesy client in
>> cl_flags with a  separate u8 field so it does not have to use
>> bit operation to set/test.
> Code audit suggested there are really only four unique
> combinations of the bit flags that are used.
>
> Plus, taking a spin_lock and using bitops seems like overkill.
>
> The rules for transitioning between the courtesy states are
> straightforward, but need to be done in a critical section.
> So I suggested storing the courtesy state in a lock-protected
> unsigned int instead of using bit flags.

I will try what Chuck suggested.

>
> If we hate it, we can go back to bit flags.

ok.

-Dai


>>>> +	spin_unlock(&clp->cl_cs_lock);
>>>> +	return rc;
>>>> +}
>>>> +
>>>>   static const struct lock_manager_operations nfsd_posix_mng_ops  = {
>>>>   	.lm_notify = nfsd4_lm_notify,
>>>>   	.lm_get_owner = nfsd4_lm_get_owner,
>>>>   	.lm_put_owner = nfsd4_lm_put_owner,
>>>> +	.lm_lock_expired = nfsd4_lm_lock_expired,
>>>>   };
>>>>     static inline void
>>>> -- 
>>>> 2.9.5
> --
> Chuck Lever
>
>
>