Inquiry on DFS Connection Aggregation and Future Plans

Inquiry on DFS Connection Aggregation and Future Plans

[Alexander Yashkin]

Hello Linux CIFS developers and community,

I am writing to seek your expert advice regarding the configuration of
DFS connections and to understand the project's direction concerning
connection aggregation.

SETUP DESCRIPTION

We have a Windows Server 2022 configured with a DFS Namespace. The
setup is as follows:

DFS Namespace: \\SERVER-FR\Data
    │
    ├── DFS Link: Shared01
    │       │
    │       ├── Target 1: \\SERVER-FR\Shared01
    │       └── Target 2: \\SERVER-UK\Shared01
    │
    └── DFS Link: Shared02
            │
            ├── Target 1: \\SERVER-FR\Shared02
            └── Target 2: \\SERVER-UK\Shared02

Our Linux client, located in the same geographical zone as SERVER-FR,
mounts the DFS root \\SERVER-FR\Data. This server (SERVER-FR) acts as
both the DFS namespace server and hosts the SMB shares for the
targets.

ISSUE: Lack of Connection Aggregation

We are observing a significant proliferation of TCP/SMB connections to
the single server (SERVER-FR), which we believe is suboptimal:
- Mounting the DFS namespace creates 2 connections.
- Accessing Shared01 creates 3 additional connections.
- Accessing Shared02 creates another 3 connections.

This results in a total of 8 connections to a single server for this
minimal setup. In our real-world production environment, with dozens
of DFS targets and thousands of clients, this behavior leads to an
overwhelming number of connections, exhausting system resources on the
SMB server.

QUESTIONS

We would be grateful if you could provide insight on the following:
- Is there a current method or configuration parameter (e.g., in
mount.cifs or via sysfs) to encourage connection aggregation?
Specifically, is it possible to reuse existing SMB connections to the
same server for accessing different DFS targets, rather than
establishing new ones for each?
- Are there any active plans or ongoing development efforts to
implement more efficient connection aggregation for single-server DFS
scenarios, similar to the behavior found in other operating systems?

Thank you for your time and for your continued work on the CIFS module.

Best Regards,
Alexander Yashkin

Re: Inquiry on DFS Connection Aggregation and Future Plans

[Paulo Alcantara]

Alexander Yashkin <alex.aspirine@gmail.com> writes:

> We would be grateful if you could provide insight on the following:
> - Is there a current method or configuration parameter (e.g., in
> mount.cifs or via sysfs) to encourage connection aggregation?
> Specifically, is it possible to reuse existing SMB connections to the
> same server for accessing different DFS targets, rather than
> establishing new ones for each?

Nope.  Reusing existing SMB connections will only work with non-DFS
mounts if 'nosharesock' isn't specified.

The current implementation will always create new DFS connections when
mounting/automounting DFS shares.  Yeah, the lack of shared connections
really sucks but it was easiest path to prevent it from reusing SMB
connections from non-DFS mounts and then potentially performing failover
on such shares.

> - Are there any active plans or ongoing development efforts to
> implement more efficient connection aggregation for single-server DFS
> scenarios, similar to the behavior found in other operating systems?

That's in my TODO list but I haven't had any free time to start working
on it.

For your specific case, those connections could definitely be reused.
More specifically, when automounting the DFS links Shared01 or Shared02,
instead of creating new connections all the way from namespace server to
DFS link target, just grab a reference count of parent mount's
connections and then reuse them to create new connection to target
share.