[bpf-next,v2,00/15] bpf: support creating maps on networking devices

Hi!

This set adds support for creating maps on networking devices.  BPF is
programs+maps, the pure program offload has been around for quite some
time, this patchset adds the map part of the equation.

Maps are allocated on the target device from the start.  There is no
host copy when map is created on the device.  Device maps are represented
by struct bpf_offloaded_map, regardless of type.  Host programs can't
access such maps, access is only possible from a program also loaded
to the same device and/or via the BPF syscall.

Offloaded programs are currently only allowed to perform lookups,
control plane is responsible for populating the maps.

For brevity only infrastructure and basic NFP patches are included.
Target device reporting, netdevsim and tests will follow up as well as
some further optimizations to the NFP code.

v2:
 - leave out the array maps, we will add them trivially later to avoid
   merge conflicts with ongoing spectere&meltdown mitigations.

Jakub Kicinski (15):
  bpf: add map_alloc_check callback
  bpf: hashtab: move attribute validation before allocation
  bpf: hashtab: move checks out of alloc function
  bpf: add helper for copying attrs to struct bpf_map
  bpf: rename bpf_dev_offload -> bpf_prog_offload
  bpf: offload: factor out netdev checking at allocation time
  bpf: offload: add map offload infrastructure
  nfp: bpf: add map data structure
  nfp: bpf: add basic control channel communication
  nfp: bpf: implement helpers for FW map ops
  nfp: bpf: parse function call and map capabilities
  nfp: bpf: add helpers for updating immediate instructions
  nfp: bpf: add verification and codegen for map lookups
  nfp: bpf: add support for reading map memory
  nfp: bpf: implement bpf map offload

 drivers/net/ethernet/netronome/nfp/Makefile        |   1 +
 drivers/net/ethernet/netronome/nfp/bpf/cmsg.c      | 446 +++++++++++++++++++++
 drivers/net/ethernet/netronome/nfp/bpf/fw.h        | 103 +++++
 drivers/net/ethernet/netronome/nfp/bpf/jit.c       | 163 +++++++-
 drivers/net/ethernet/netronome/nfp/bpf/main.c      |  60 ++-
 drivers/net/ethernet/netronome/nfp/bpf/main.h      |  95 ++++-
 drivers/net/ethernet/netronome/nfp/bpf/offload.c   | 106 ++++-
 drivers/net/ethernet/netronome/nfp/bpf/verifier.c  |  47 +++
 drivers/net/ethernet/netronome/nfp/nfp_app.h       |   9 +
 drivers/net/ethernet/netronome/nfp/nfp_asm.c       |  58 +++
 drivers/net/ethernet/netronome/nfp/nfp_asm.h       |   4 +
 drivers/net/ethernet/netronome/nfp/nfp_net.h       |  12 +
 .../net/ethernet/netronome/nfp/nfp_net_common.c    |   7 +
 include/linux/bpf.h                                |  65 ++-
 include/linux/netdevice.h                          |   6 +
 include/uapi/linux/bpf.h                           |   1 +
 kernel/bpf/cpumap.c                                |   8 +-
 kernel/bpf/devmap.c                                |   8 +-
 kernel/bpf/hashtab.c                               | 103 +++--
 kernel/bpf/lpm_trie.c                              |   7 +-
 kernel/bpf/offload.c                               | 224 ++++++++++-
 kernel/bpf/sockmap.c                               |   8 +-
 kernel/bpf/stackmap.c                              |   6 +-
 kernel/bpf/syscall.c                               |  71 +++-
 kernel/bpf/verifier.c                              |   7 +
 tools/include/uapi/linux/bpf.h                     |   1 +
 26 files changed, 1506 insertions(+), 120 deletions(-)
 create mode 100644 drivers/net/ethernet/netronome/nfp/bpf/cmsg.c

On 01/12/2018 05:29 AM, Jakub Kicinski wrote:
> Hi!
> 
> This set adds support for creating maps on networking devices.  BPF is
> programs+maps, the pure program offload has been around for quite some
> time, this patchset adds the map part of the equation.
> 
> Maps are allocated on the target device from the start.  There is no
> host copy when map is created on the device.  Device maps are represented
> by struct bpf_offloaded_map, regardless of type.  Host programs can't
> access such maps, access is only possible from a program also loaded
> to the same device and/or via the BPF syscall.
> 
> Offloaded programs are currently only allowed to perform lookups,
> control plane is responsible for populating the maps.
> 
> For brevity only infrastructure and basic NFP patches are included.
> Target device reporting, netdevsim and tests will follow up as well as
> some further optimizations to the NFP code.
> 
> v2:
>  - leave out the array maps, we will add them trivially later to avoid
>    merge conflicts with ongoing spectere&meltdown mitigations.

Series applied to bpf-next, thanks Jakub!