netdev.vger.kernel.org archive mirror
 help / color / mirror / Atom feed
From: "Naveen N. Rao" <naveen.n.rao@linux.vnet.ibm.com>
To: Alexei Starovoitov <ast@fb.com>,
	netdev@vger.kernel.org,
	Sandipan Das <sandipan@linux.vnet.ibm.com>
Cc: Brendan Gregg <brendan.d.gregg@gmail.com>,
	daniel@iogearbox.net, Martin KaFai Lau <kafai@fb.com>,
	Kees Cook <keescook@chromium.org>,
	linux-kernel@vger.kernel.org
Subject: Re: [RFC PATCH] bpf: Add helpers to read useful task_struct members
Date: Mon, 06 Nov 2017 21:25:41 +0530	[thread overview]
Message-ID: <1509982000.092la4257a.naveen@linux.ibm.com> (raw)
In-Reply-To: <4acdc081-341d-ee91-a591-b1d331a8c8d5@fb.com>

Alexei Starovoitov wrote:
> On 11/5/17 2:31 AM, Naveen N. Rao wrote:
>> Hi Alexei,
>>
>> Alexei Starovoitov wrote:
>>> On 11/3/17 3:58 PM, Sandipan Das wrote:
>>>> For added security, the layout of some structures can be
>>>> randomized by enabling CONFIG_GCC_PLUGIN_RANDSTRUCT. One
>>>> such structure is task_struct. To build BPF programs, we
>>>> use Clang which does not support this feature. So, if we
>>>> attempt to read a field of a structure with a randomized
>>>> layout within a BPF program, we do not get the expected
>>>> value because of incorrect offsets. To observe this, it
>>>> is not mandatory to have CONFIG_GCC_PLUGIN_RANDSTRUCT
>>>> enabled because the structure annotations/members added
>>>> for this purpose are enough to cause this. So, all kernel
>>>> builds are affected.
>>>>

...

>>>> diff --git a/include/uapi/linux/bpf.h b/include/uapi/linux/bpf.h
>>>> index f90860d1f897..324508d27bd2 100644
>>>> --- a/include/uapi/linux/bpf.h
>>>> +++ b/include/uapi/linux/bpf.h
>>>> @@ -338,6 +338,16 @@ union bpf_attr {
>>>>   *     @skb: pointer to skb
>>>>   *     Return: classid if != 0
>>>>   *
>>>> + * u64 bpf_get_task_pid_tgid(struct task_struct *task)
>>>> + *     Return: task->tgid << 32 | task->pid
>>>> + *
>>>> + * int bpf_get_task_comm(struct task_struct *task)
>>>> + *     Stores task->comm into buf
>>>> + *     Return: 0 on success or negative error
>>>> + *
>>>> + * u32 bpf_get_task_flags(struct task_struct *task)
>>>> + *     Return: task->flags
>>>> + *
>>>
>>> I don't think it's a solution.
>>> Tracing scripts read other fields too.
>>> Making it work for these 3 fields is a drop in a bucket.
>>
>> Indeed. However...
>>
>>> If randomization is used I think we have to accept
>>> that existing bpf scripts won't be usable.
>>
>> ... the actual issue is that randomization isn't necessary for this to
>> show up. The annotations added to mark off the structure members results
>> in some structure members being moved into an anonymous structure, which
>> would then get padded differently. So, *all* kernels since v4.13 are
>> affected, afaict.
> 
> hmm. why would all 4.13+ be affected?
> It's just an anonymous struct inside task_struct.
> Are you saying that due to clang not adding this 'struct { };' treatment 
> to task_struct?

Yes, that's what it looked like.

> I thought such struct shouldn't change layout.
> If it is we need to fix include/linux/compiler-clang.h to do that
> anon struct as well.

We considered that, but it looked to be very dependent on the version of 
gcc used to build the kernel. But, this may be a simpler approach for 
the shorter term.

> 
>> As such, we wanted to propose this as a short term solution, but I do
>> agree that this doesn't solve the real issue.
>>
>>> Long term solution is to support 'BPF Type Format' or BTF
>>> (which is old C-Type Format) for kernel data structures,
>>> so bcc scripts wouldn't need to use kernel headers and clang.
>>> The proper offsets will be described in BTF.
>>> We were planning to use it initially to describe map key/value,
>>> but it applies for this case as well.
>>> There will be a tool that will take dwarf from vmlinux and
>>> compress it into BTF. Kernel will also be able to verify
>>> that BTF is a valid BTF.
>>
>> This is the first that I've heard about BTF. Can you share more details
>> about it, or point me to some place where it has been discussed?
>>
>> We considered having tools derive the structure offsets from debuginfo,
>> but debuginfo may not always be present on production systems. So, it
>> isn't clear if having that dependency is fine. I'm not sure how BTF will
>> be different.
> 
> It was discussed at this year plumbers:
> https://lwn.net/Articles/734453/
> 
> btw the name BTF is work in progress. We started with CTF, but
> it conflicts with all other meanings of this abbreviation.
> Likely we will call it something different at the end.
> 
> Initial goal was to describe key/map values of bpf maps to
> make debugging easier, but now we want to use this compressed
> type format for tracing as well, since installing kernel headers
> everywhere doesn't scale well while CTF can be embedded in vmlinux

Makes sense, though I'm curious on how you're planning to have this work
without the kernel headers :)

> 
> We were also thinking to improve verifier with CTF knowledge too.
> Like if CTF describes that map value is two u32, but bpf program
> is doing 8-byte access then something is wrong and either warn
> or reject such program.

Sounds good. I look forward to more details/patches on this front once 
you're ready to share more.

Thanks,
- Naveen

  reply	other threads:[~2017-11-06 15:55 UTC|newest]

Thread overview: 18+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2017-11-03  6:58 [RFC PATCH] bpf: Add helpers to read useful task_struct members Sandipan Das
2017-11-04  9:34 ` Alexei Starovoitov
2017-11-04 17:31   ` Naveen N. Rao
2017-11-04 21:10     ` Alexei Starovoitov
2017-11-06 15:55       ` Naveen N. Rao [this message]
2017-11-07  8:08         ` Alexei Starovoitov
2017-11-07  8:37           ` Naveen N. Rao
2017-11-07 21:14             ` Y Song
2017-11-07 21:31               ` Atish Patra
2017-11-07 21:45                 ` Y Song
2017-11-07 21:39               ` Alexei Starovoitov
2017-11-07 21:47                 ` Y Song
2017-11-07 22:04                   ` Alexei Starovoitov
2017-11-07 22:42                     ` Y Song
2017-11-08  0:29                       ` Atish Patra
2017-11-08  1:25                         ` Y Song
2017-11-06  5:16     ` Sandipan Das
2017-11-07  0:16 ` Tushar Dave

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=1509982000.092la4257a.naveen@linux.ibm.com \
    --to=naveen.n.rao@linux.vnet.ibm.com \
    --cc=ast@fb.com \
    --cc=brendan.d.gregg@gmail.com \
    --cc=daniel@iogearbox.net \
    --cc=kafai@fb.com \
    --cc=keescook@chromium.org \
    --cc=linux-kernel@vger.kernel.org \
    --cc=netdev@vger.kernel.org \
    --cc=sandipan@linux.vnet.ibm.com \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).