Re: [PATCH bpf-next 1/4] bpf: unprivileged BPF access via /dev/bpf

From: Song Liu <songliubraving@fb.com>
To: Greg KH <gregkh@linuxfoundation.org>
Cc: Daniel Borkmann <daniel@iogearbox.net>,
	Networking <netdev@vger.kernel.org>, bpf <bpf@vger.kernel.org>,
	Alexei Starovoitov <ast@kernel.org>,
	Kernel Team <Kernel-team@fb.com>,
	"jannh@google.com" <jannh@google.com>
Subject: Re: [PATCH bpf-next 1/4] bpf: unprivileged BPF access via /dev/bpf
Date: Thu, 27 Jun 2019 16:51:20 +0000	[thread overview]
Message-ID: <48E35F58-0DAD-40BA-993F-8AB76587A93B@fb.com> (raw)
In-Reply-To: <20190627163723.GA9643@kroah.com>

> On Jun 27, 2019, at 9:37 AM, Greg KH <gregkh@linuxfoundation.org> wrote:
> 
> On Thu, Jun 27, 2019 at 01:00:03AM +0000, Song Liu wrote:
>> 
>> 
>>> On Jun 26, 2019, at 5:08 PM, Greg KH <gregkh@linuxfoundation.org> wrote:
>>> 
>>> On Wed, Jun 26, 2019 at 03:17:47PM +0000, Song Liu wrote:
>>>>>> +static struct miscdevice bpf_dev = {
>>>>>> +	.minor		= MISC_DYNAMIC_MINOR,
>>>>>> +	.name		= "bpf",
>>>>>> +	.fops		= &bpf_chardev_ops,
>>>>>> +	.mode		= 0440,
>>>>>> +	.nodename	= "bpf",
>>>>> 
>>>>> Here's what kvm does:
>>>>> 
>>>>> static struct miscdevice kvm_dev = {
>>>>>      KVM_MINOR,
>>>>>      "kvm",
>>>>>      &kvm_chardev_ops,
>>>>> };
>>> 
>>> Ick, I thought we converted all of these to named initializers a long
>>> time ago :)
>>> 
>>>>> Is there an actual reason that mode is not 0 by default in bpf case? Why
>>>>> we need to define nodename?
>>>> 
>>>> Based on my understanding, mode of 0440 is what we want. If we leave it 
>>>> as 0, it will use default value of 0600. I guess we can just set it to 
>>>> 0440, as user space can change it later anyway. 
>>> 
>>> Don't rely on userspace changing it, set it to what you want the
>>> permissions to be in the kernel here, otherwise you have to create a new
>>> udev rule and get it merged into all of the distros.  Just do it right
>>> the first time and there is no need for it.
>>> 
>>> What is wrong with 0600 for this?  Why 0440?
>> 
>> We would like root to own the device, and let users in a certain group 
>> to be able to open it. So 0440 is what we need. 
> 
> But you are doing a "write" ioctl here, right?  So don't you really need

By "write", you meant that we are modifying a bit in task_struct, right?
In that sense, we probably need 0220?

> 0660 at the least?  And if you "know" the group id, I think you can
> specify it too so udev doesn't have to do a ton of work, but that only
> works for groups that all distros number the same.

I don't think we know the group id yet. 

> 
> And why again is this an ioctl instead of a syscall?  What is so magic
> about the file descriptor here?

We want to control the permission of this operation via this device. 
Users that can open the device would be able to run the ioctl. I think 
syscall cannot achieve control like this, unless we introduce something 
like CAP_BPF_ADMIN?

Thanks,
Song