Re: [RFC] bpf: offload: report device information for offloaded programs

[Jakub Kicinski]

Hi Kirill,

On Thu, 30 Nov 2017 16:19:13 +0300, Kirill Tkhai wrote:
> > @@ -164,6 +166,38 @@ int bpf_prog_offload_compile(struct bpf_prog *prog)
> > return bpf_prog_offload_translate(prog);
> >  }
> >  
> > +int bpf_prog_offload_info_fill(struct bpf_prog_info *info,
> > +  struct bpf_prog *prog)
> > +{
> > +   struct bpf_dev_offload *offload = prog->aux->offload;
> > +   struct inode *ns_inode;
> > +   struct path ns_path;
> > +   struct net *net;
> > +   int ret = 0;
> > +   void *ptr;
> > +
> > +   info->dev_bound = 1;
> > +
> > +   rtnl_lock();  
> 
> rtnl_lock() is too big lock and it is already overused in kernel.
> Can't we use smaller lock in this driver to protect bpf_prog_offload_devs?
> I suppose rwlock would be appropriate for that.
> 
> (Then, we may completely remove rtnl_lock() from bpf_prog_offload_init()
> and use readlocked dev_base_lock for __dev_get_by_index() instead and
> the new small_rwlock to link in the list.
> 
> Not sure about bpf_prog_offload_verifier_prep() and 
> bpf_prog_offload_translate()
> and which context expect net_device_ops->ndo_bpf users. Either they need rtnl
> or not).

Thanks for the comments, removing the use of rtnl_lock is definitely on
my todo list!

Re: [RFC] bpf: offload: report device information for offloaded programs

[Kirill Tkhai]

Hi, Jakub,

please, read comments below.

On 30.11.2017 03:22, Jakub Kicinski wrote:
> Report to the user ifindex and namespace information of offloaded
> programs.  Always set dev_bound to true if program was loaded for
> a device which has been since removed.  Specify the namespace
> using dev/inode combination.
> 
> Signed-off-by: Jakub Kicinski 
> Reviewed-by: Simon Horman 
> Reviewed-by: Quentin Monnet 
> ---
>  fs/nsfs.c  |  2 +-
>  include/linux/bpf.h|  2 ++
>  include/linux/proc_ns.h|  1 +
>  include/uapi/linux/bpf.h   |  5 +
>  kernel/bpf/offload.c   | 34 ++
>  kernel/bpf/syscall.c   |  6 ++
>  tools/include/uapi/linux/bpf.h |  5 +
>  7 files changed, 54 insertions(+), 1 deletion(-)

[snip]

> diff --git a/kernel/bpf/offload.c b/kernel/bpf/offload.c
> index 8455b89d1bbf..da98349c647d 100644
> --- a/kernel/bpf/offload.c
> +++ b/kernel/bpf/offload.c
> @@ -16,9 +16,11 @@
>  #include 
>  #include 
>  #include 
> +#include 
>  #include 
>  #include 
>  #include 
> +#include 
>  #include 
>  
>  /* protected by RTNL */
> @@ -164,6 +166,38 @@ int bpf_prog_offload_compile(struct bpf_prog *prog)
>   return bpf_prog_offload_translate(prog);
>  }
>  
> +int bpf_prog_offload_info_fill(struct bpf_prog_info *info,
> +struct bpf_prog *prog)
> +{
> + struct bpf_dev_offload *offload = prog->aux->offload;
> + struct inode *ns_inode;
> + struct path ns_path;
> + struct net *net;
> + int ret = 0;
> + void *ptr;
> +
> + info->dev_bound = 1;
> +
> + rtnl_lock();

rtnl_lock() is too big lock and it is already overused in kernel.
Can't we use smaller lock in this driver to protect bpf_prog_offload_devs?
I suppose rwlock would be appropriate for that.

(Then, we may completely remove rtnl_lock() from bpf_prog_offload_init()
and use readlocked dev_base_lock for __dev_get_by_index() instead and
the new small_rwlock to link in the list.

Not sure about bpf_prog_offload_verifier_prep() and bpf_prog_offload_translate()
and which context expect net_device_ops->ndo_bpf users. Either they need rtnl
or not).

Then the below hunk:

> + if (!offload->netdev)
> + goto out;
> +
> + net = dev_net(offload->netdev);
> + get_net(net); /* __ns_get_path() drops the reference */

will be:

read_lock(_rwlock);
if (!offload->netdev)
goto out;

net = dev_net(offload->netdev);
get_net(net); /* __ns_get_path() drops the reference */
read_unlock(_rwlock);

and rtnl_lock() won't be touched.

> + ptr = __ns_get_path(_path, >ns);
> + ret = PTR_ERR_OR_ZERO(ptr);
> + if (ret)
> + goto out;
> + ns_inode = ns_path.dentry->d_inode;
> +
> + info->ns_dev = new_encode_dev(ns_inode->i_sb->s_dev);
> + info->ns_inode = ns_inode->i_ino;
> + info->ifindex = offload->netdev->ifindex;
> +out:
> + rtnl_unlock();
> + return ret;
> +}
> +

[snip]

Kirill