Re: [PATCH v2 2/4] NFS: release per-net clients lock before calling PipeFS dentries creation

[Stanislav Kinsbursky <skinsbursky@parallels.com>]

27.02.2012 19:59, David Laight пишет:
>
>>   	spin_lock(&nn->nfs_client_lock);
>> -	list_for_each_entry(clp,&nn->nfs_client_list, cl_share_link) {
>> +	list_for_each_entry_safe(clp, tmp,&nn->nfs_client_list,
> cl_share_link) {
>>   		if (clp->rpc_ops !=&nfs_v4_clientops)
>>   			continue;
>> +		atomic_inc(&clp->cl_count);
>> +		spin_unlock(&nn->nfs_client_lock);
>>   		error = __rpc_pipefs_event(clp, event, sb);
>> +		nfs_put_client(clp);
>>   		if (error)
>>   			break;
>> +		spin_lock(&nn->nfs_client_lock);
>>   	}
>>   	spin_unlock(&nn->nfs_client_lock);
>>   	return error;
>
> The locking doesn't look right if the loop breaks on error.
> (Same applied to patch v2 1/4)
>

Thanks for the catch. I'll fix this.

> Although list_fo_each_entry_safe() allows the current entry
> to be freed, I don't believe it allows the 'next' to be freed.
> I doubt there is protection against that happening.
>

We need to use safe macro, because client can be destroyed on nfs_put_client() call.
About "protection against ... the 'next' to be freed" - I dont' think, that we 
need any protection against it. This will be done under nfs_client_lock, and 
current entry list pointers will be updated properly.

> Do you need to use an atomic_inc() for cl_count.
> I'd guess the nfs_client_lock is usually held?
>

Sorry, I don't understand this question.

-- 
Best regards,
Stanislav Kinsbursky