RE: [RFC PATCH v12 00/17] dlb: introduce DLB device driver

-----Original Message-----
From: Andrew Lunn <andrew@lunn.ch>
Sent: Tuesday, December 21, 2021 4:41 AM
To: Chen, Mike Ximing <redacted>
Cc: linux-kernel@vger.kernel.org; arnd@arndb.de; gregkh@linuxfoundation.org; Williams, Dan J
[off-list ref]; pierre-louis.bossart@linux.intel.com; netdev@vger.kernel.org;
davem@davemloft.net; kuba@kernel.org
Subject: Re: [RFC PATCH v12 00/17] dlb: introduce DLB device driver

quoted

1. Before a scheduling domain is created/enabled, a set of parameters
are passed to the kernel driver via configfs attribute files in an
configfs domain directory (say $domain) created by user. Each
attribute file corresponds to a configuration parameter of the domain.
After writing to all the attribute files, user writes 1 to "create"
attribute, which triggers an action (i.e., domain creation) in the
kernel driver. Since multiple processes/users can access the $domain
directory, multiple users can write to the attribute files at the same
time.  How do we guarantee an atomic update/configuration of a domain?
In other words, if user A wants to set attributes 1 and 2, how can we
prevent user B from changing attribute 1 and 2 before user A writes 1
to "create"? A configfs directory with individual attribute files
seems to not be able to provide atomic configuration in this case. One
option to solve this issue could be write a structured data (with a
set of parameters) to a single attribute file. This would guarantee the atomic configuration, but may not

be a conventional configfs operation.

How about throw away configfs and use netlink? Messages are atomic, and you can add an arbitrary
number of attributes to a single netlink message. It will also make your code more network like, since
nothing else in the network stack uses configfs, as far as i know.

Hi Andrew,
As I explained in my other response, DLB is not a network accelerator and DLB
driver is not a part of network stack. We would obviously prefer to resolve the
atomic update and resource reset at tear-down Issues within the configfs
framework if possible. But I will take a look at the netlink implementations.

Thanks for the suggestion
Mike