On 6/1/17 6:55 AM, David Ahern wrote:

On 5/31/17 8:22 PM, Jakub Kicinski wrote:

quoted

On Wed, 31 May 2017 11:58:54 -0700, Martin KaFai Lau wrote:

quoted

This patch series:
1) Introduce ID for both bpf_prog and bpf_map.
2) Add bpf commands to iterate the prog IDs and map
   IDs of the system.
3) Add bpf commands to get a prog/map fd from an ID
4) Add bpf command to get prog/map info from a fd.
   The prog/map info is a jump start in this patchset
   and it is not meant to be a complete list.  They can
   be extended in the future patches.

Hi!  IIUC there were plans at some point to store the BPF byte code in
the kernel, the way it was loaded from the user space and before it
went through the verifier.  Are you still planning on implementing that?
It would be handy for offloads to have the original byte code, things
like call inlining make translations of maps calls a bit challenging,
since we try to run through the verifier again..

Save the instructions:
https://github.com/dsahern/linux/commit/5315578db7b0886a3e8e3bef6a56e828ae2bd2c4

Let userspace read it back:
https://github.com/dsahern/linux/commit/619982064b9b582fc5d4a504feffca1730610c4e

One more in between in the branch:
https://github.com/dsahern/linux/tree/bpf/retrieve-bpf-wip

Now that the id code is in, it needs to be returned on link dumps, route
dumps, etc and accessible via fdinfo. The id can be turned into an fd
and the fd read using the above patches.

what problem are we trying to solve?
If we don't agree on the problem, we won't be able to agree on the solution.

Sounds to me that Jakub wants to translate lookup/update helper calls
for JITing and it's difficult due to inlining optimization.
In such case isn't the easiest solution to disable inlining when
JIT to NFP is on?

Dave's case is to be able to retrieve a cgroup_sock program used
for vrf. Such prog doesn't have any maps and I don't see why
returning real insns is not enough.
As Martin said there is work in progress patches that return IDs
for programs attached to everything: cgroup_sock hook, xdp, etc.
The user can get prog ID, convert to FD and dump all insns.
Why this is not enough for VRF introspection?