[openib-general] Re: [PATCH][RFC/v1][9/12] Add InfiniBand userspace MAD support

[Roland Dreier]

    >> Add a driver that provides a character special device for each
    >> InfiniBand port.  This device allows userspace to send and
    >> receive MADs via write() and read() (with some control
    >> operations implemented as ioctls).

    Greg> Do you really need these ioctls?

    Greg> This could be in a sysfs file, right?

The API version definitely can be, good point.

    Greg> You are letting any user, with any privilege register or
    Greg> unregister an "agent"?

They have to be able to open the device node.  We could add a check
that they have it open for writing but there's not really much point
in opening this device read-only.

    Greg> And shouldn't you lock your list of agent ids when adding or
    Greg> removing one, or are you relying on the BKL of the ioctl
    Greg> call?  If so, please document this.

Each file has an "agent_mutex" rwsem that protects this... the global
list of agents handled by the lower level API is protected by its own locking.

    Greg> Also, these "agents" seem to be a type of filter, right?  Is
    Greg> there no other way to implement this than an ioctl?

ioctl seems to be the least bad way to me.  This really feels like a
legitimate use of ioctl to me -- we use read/write to handle passing
data through our file descriptor, and ioctl for control of the
properties of the descriptor.

What would you suggest as an ioctl replacement?

Thanks,
 Roland