How to deal with concurrency

[nsoubelet]

Hello there,
Not sure whether you read this Q&A section, however I'll give it a try.
Basically I implemented a nfs4 to azure files (smb) using this incredible library. However, I am facing a really (basic) issue and I would need to know whether solving it in my implementation FS or there is something in the internals (also what you recommend in this case).

In Azure Files, when opening a file for writing, it is marked as exclusive, meaning that any other operation trying to write into it, will get an error.
However, my problem here is that if I lock the write section using any of the standard concurrency mechanisms, I will be penalizing other threads that want to write into different files, which is a perfectly valid operation.

Therefore my question: is there any functionality to enqueue WriteResult write operations over the same Inode or I should implement that logic in my VirtualFileSystem implementation?
In case there is not a standard way, any suggestions on how to approach the problem?

Thanks! :)

[kofemann]

Hi,

We are always happy to see that our works are also useful in other projects!

Coming to the questions. NFS4.0 protocol has a concept of Share reservations, which defines which
concurrent access is allowed. There is share_access and share_deny. The share_access
specifies the access mode which is wanted, i.e. open for read or open for write, while
deny defines which concurrent access is allowed, i.e. other readers or other writers.
The POSIX standard doesn't provide a way to specify share_deny, thus, typically, the
NFS clients can't specify it. However, nfs4j supports it. The quick-and-dirty workaround
will be to set _args.opopen.share_deny.value to OPEN4_SHARE_DENY_BOTH
if share_access is OPEN4_SHARE_ACCESS_WRITE. The NFS server should be updated to
use the new op:

    private static class AzureOperationExecutor extends MDSOperationExecutor {
    
        @Override
        public AbstractNFSv4Operation getOperation(nfs_argop4 op) {
            if (op.argop == nfs_opnum4.OP_OPEN && 
                  op.opopen.share_access.value == OPEN4_SHARE_ACCESS_WRITE) {
               // tweak share deny
               op.opopen.share_deny.value = OPEN4_SHARE_DENY_BOTH;
            }
            return super.getOperation(op);
        }
    }

     NFSServerV41 nfs4 =
        new NFSServerV41.Builder()
            .withOperationExecutor(new AzureOperationExecutor())
            ...
            .build();

The code above is of course just an idea and probably should make it right :)

A cleaner solution might be exposing some kind of file system capabilities to define the default behaviour.

I hope this helps.

[nsoubelet]

Hello @kofemann, thanks for replying.
That was really helpful. I did not think about adding that logic in the operation executor.
I will investigate that. :)