Re: [RFC PATCH 0/2] sctp: add new getsockopt option SCTP_SOCKOPT_PEELOFF_KERNEL

[Marcelo Ricardo Leitner <marcelo.leitner@gmail.com>]

On 17-06-2015 09:20, Neil Horman wrote:
> On Wed, Jun 17, 2015 at 08:38:10AM -0300, Marcelo Ricardo Leitner wrote:
>> On 17-06-2015 07:21, Neil Horman wrote:
>>> On Tue, Jun 16, 2015 at 07:42:31PM -0300, Marcelo Ricardo Leitner wrote:
>>>> Hi,
>>>>
>>>> I'm trying to remove a direct dependency of dlm module on sctp one.
>>>> Currently dlm code is calling sctp_do_peeloff() directly and only this
>>>> call is causing the load of sctp module together with dlm. For that, we
>>>> have basically 3 options:
>>>> - Doing a module split on dlm
>>>>    - which I'm avoiding because it was already split and was merged (more
>>>>      info on patch2 changelog)
>>>>    - and the sctp code on it is rather small if compared with sctp module
>>>>      itself
>>>> - Using some other infra that gets indirectly activated, like getsockopt()
>>>>    - It was like this before, but the exposed sockopt created a file
>>>>      descriptor for the new socket and that create some serious issues.
>>>>      More info on 2f2d76cc3e93 ("dlm: Do not allocate a fd for peeloff")
>>>> - Doing something like ipv6_stub (which is used by vxlan) or similar
>>>>    - but I don't feel that's a good way out here, it doesn't feel right.
>>>>
>>>> So I'm approaching this by going with 2nd option again but this time
>>>> also creating a new sockopt that is only accessible for kernel users of
>>>> this protocol, so that we are safe to directly return a struct socket *
>>>> via getsockopt() results. This is the tricky part of it of this series.
>>>>
>>>> It smells hacky yes but currently most of sctp calls are wrapped behind
>>>> kernel_*(). Even if we set a flag (like netlink does) saying that this
>>>> is a kernel socket, we still have the issue of getting the function call
>>>> through and returning such non-usual return value.
>>>>
>>>> I kept __user marker on sctp_getsockopt_peeloff_kernel() prototype and
>>>> its helpers just to avoid issues with static checkers.
>>>>
>>>> Kernel path not really tested yet.. mainly willing to know what do you
>>>> think, is this feasible? getsockopt option only reachable by kernel
>>>> itself? Couldn't find any other like this.
>>>>
>>>> Thanks,
>>>> Marcelo
>>>>
>>>> Marcelo Ricardo Leitner (2):
>>>>    sctp: add new getsockopt option SCTP_SOCKOPT_PEELOFF_KERNEL
>>>>    dlm: avoid using sctp_do_peeloff directly
>>>>
>>>>   fs/dlm/lowcomms.c         | 17 ++++++++---------
>>>>   include/uapi/linux/sctp.h | 12 ++++++++++++
>>>>   net/sctp/socket.c         | 39 +++++++++++++++++++++++++++++++++++++++
>>>>   3 files changed, 59 insertions(+), 9 deletions(-)
>>>>
>>>> --
>>>> 2.4.1
>>>>
>>>>
>>>
>>> Why not just use the existing PEELOFF socket option with the kernel_getsockopt
>>> interface, and sockfd_lookup to translate the returned value back to a socket
>>> struct?  That seems less redundant and less hack-ish to me.
>>
>> It was like that before commit 2f2d76cc3e93 ("dlm: Do not allocate a fd for
>> peeloff"), but it caused serious issues due to the fd allocation, so that's
>> what I'm willing to avoid now.
>>
>> References:
>> http://article.gmane.org/gmane.linux.network.drbd/22529
>> https://bugzilla.redhat.com/show_bug.cgi?id=1075629 (this one is closed,
>> sorry)
>>
>>    Marcelo
>>
> Ah, I see.  You're using the new socket option as a differentiator to just skip
> the creation of an FD.

Exactly.

> I get your reasoning, but I'm still not in love with the idea of duplicating
> code paths to avoid that action.  Can we use some data inside the socket
> structure to do this differentiation?  Specifically here I'm thinking of
> sock->file.  IIRC that will be non-null for any sockets created in user space,

I had thought about using some socket flags like netlink does but 
couldn't get around with that. Hadn't thought about sock->file though, 
nice idea.

> but will always be NULL for dlm created sockets (since we use sock_create
> directly to create them.  If that is a sufficient differentiator, then we can
> just optionally allocate the new socket fd for the peeled off socket, iff the
> parent sock->file pointer is non-null.
>
> Thoughts?
> Neil

We can re-use the current code path, by either checking it via 
sock->file or via get_fs(). That will require us to change the option 
arg format so we keep it nice and clean but as it would be kernel-side 
only, it should be ok right? It currently is:

typedef struct {
         sctp_assoc_t associd;
         int sd;
} sctp_peeloff_arg_t;

And we would have to fit a pointer in there, something like:
typedef union {
	struct {
	        sctp_assoc_t associd;
	        int sd;
	};
	void *sock;
} sctp_peeloff_arg_t;

Sounds good?

   Marcelo