Skip to contentSkip to conversation

Passing references to kobjects between userland and kernel

DISCUSSION Daniele orlandi 2 participants 5 messages Jun 14, 2006 – Jun 20, 2006

lore
Share & tools
lore
#0 lore raw

Hello,

I'm trying to figure out what is the correct way to pass references to kobjects between userland and kernel space.

I have my big object hierarchy of kobjects representing a TDM interconnect graph with channels, crossconnectors, physical ports and so on. The main objects are nodes and archs; archs connect two nodes together.

The hierarchy is exported to sysfs.

Quoted 1 line

From userland I want to tell the kernel "Connect node X to node Y".

The first problem is deciding what interface to use; currently I'm using an ioctl() but I'm open to suggestions.

The other issue is how to identify objects X and Y in the ioctl() request, here are the possible ways I've examined:

1- Path
-------

Every kobject registered in sysfs has its own path. I can use path_lookup() and lookup the object by its kobj.dentry.

Pros: In userland it's easy to identify kobjects by their path in sysfs. The path is unambiguous and already available.

Cons: Passing the pathname to the kernel requires a big ioctl() body (theoretically at least 2*MAX_PATH using a fixed-length structure).

Doing the reverse mapping (from kernel to userspace) is not easy since you need to know from which vfsmount to generate the path with d_path().

2- Relative path
----------------

Pros: No need to consider where sysfs is mounted and would also work if the kobjects are not registeres in sysfs.

Cons: There should be specific functions to parse/generate a pathname that is specific to kobject hierarchy.

3- Unique ID
------------

A unique ID, probably a integer value can be assigned to each object.

Pros: much more compact, do not need parsing

Cons: Needs an 'id' attribute in sysfs for every object and a reverse mapping (done with symlinks?) to allow an userland application to map IDs to paths and vice-versa. Unique ID allocation needs an allocator (list, bitmap, etc) either global or for every namespace.

4- kobject's pointer
------------------
The kobject's pointer could opaquely be seen as a unique identifier.

Pros: Compact, no need for storage.

Cons: Probably insane, exposes kernel internals to the userland, a developer may be tempted to deference the user-provided value directly :)

So, are there any other ways? If not, which one would be advisable among these?

Bye,

Signature · 1 line
  Daniele "Vihai" Orlandi
#1 lore raw
Quoted 5 lines · Daniele orlandi

Hello,

I'm trying to figure out what is the correct way to pass references to kobjects between userland and kernel space.

Use the kobject_uevent() call from kernelspace to let userspace know whatever you want it to. That is what it is there for :)

Quoted 8 lines

I have my big object hierarchy of kobjects representing a TDM interconnect graph with channels, crossconnectors, physical ports and so on. The main objects are nodes and archs; archs connect two nodes together.

The hierarchy is exported to sysfs.

From userland I want to tell the kernel "Connect node X to node Y".

Then use the names of the kobjects within your subdirectory. They have to be unique so you should be safe.

Or use configfs :)

thanks,

greg k-h

#2 lore raw
Quoted 3 lines · Greg KH

Use the kobject_uevent() call from kernelspace to let userspace know whatever you want it to. That is what it is there for :)

kobject_uevent() is fine if I want to asynchronously notify the user space of an event.

What I need is a synchronous bidirectional interface, e.g. I tell the kernel "connect node X with node Y" and I get back the resulting pipeline identifier.

Quoted 1 line

Or use configfs :)

Configfs is very interesting, it's surely well suited for several tasks I'm doing and I will add support for it into my subsystem.

What it's missing is the ability to create non-persistent configurations, bound to a process and disappearing along with the process that created them.

For example, my process creates pipelines between nodes using a ioctl() interface. If the process dies unexpectedly the file descriptor is automatically closed and I can release all the pipelines. I don't see a way I could accomplish the same with configfs or sysfs.

OTOH, for what concerns persistent pipelines, configfs is quite good. It would be even better if I could create links from configfs objects to sysfs objects.

Bye,

Signature · 1 line
  Daniele "Vihai" Orlandi
#3 lore raw
Quoted 11 lines · Daniele Orlandi
Quoted 3 lines · Greg KH

Use the kobject_uevent() call from kernelspace to let userspace know whatever you want it to. That is what it is there for :)

kobject_uevent() is fine if I want to asynchronously notify the user space of an event.

What I need is a synchronous bidirectional interface, e.g. I tell the kernel "connect node X with node Y" and I get back the resulting pipeline identifier.

Why do you feel that this is a requirement? What exactly are you trying to do?

thanks,

greg k-h

#4 lore raw
Quoted 3 lines · Greg KH

Why do you feel that this is a requirement? What exactly are you trying to do?

I have my channels interconnect graph, nicely exported to sysfs, and I just need an interface that an application could use to request connections between two nodes of this graph.

The result of the connection is a pipeline object, also exported to sysfs, that the application will use to do further work on the connection.

Currently I'm using an ioctl() but, given that ioctls seem to be a no-no for future projects, given that there is great effort in having new and better interfaces, I'm trying to figure out what is the best way to do what I need.

thanks,
Bye,

Signature · 1 line
  Daniele Orlandi