Jakub Kicinski says:

====================
This series centres on NFP offload of bpf_event_output().  The
first patch allows perf event arrays to be used by offloaded
programs.  Next patch makes the nfp driver keep track of such
arrays to be able to filter FW events referring to maps.
Perf event arrays are not device bound.  Having driver
reimplement and manage the perf array seems brittle and unnecessary.

Patch 4 moves slightly the verifier step which replaces map fds
with map pointers.  This is useful for nfp JIT since we can then
easily replace host pointers with NFP table ids (patch 6).  This
allows us to lift the limitation on map helpers having to be used
with the same map pointer on all paths.  Second use of replacing
fds with real host map pointers is that we can use the host map
pointer as a key for FW events in perf event array offload.

Patch 5 adds perf event output offload support for the NFP.

There are some differences between bpf_event_output() offloaded
and non-offloaded version.  The FW messages which carry events
may get dropped and reordered relatively easily.  The return codes
from the helper are also not guaranteed to match the host.  Users
are warned about some of those discrepancies with a one time
warning message to kernel logs.

bpftool gains an ability to dump perf ring events in a very simple
format.  This was very useful for testing and simple debug, maybe
it will be useful to others?

Last patch is a trivial comment fix.
====================
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>

Comments in the verifier refer to free_bpf_prog_info() which
seems to have never existed in tree.  Replace it with
free_used_maps().
Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: Quentin Monnet <quentin.monnet@netronome.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>

Users of BPF sooner or later discover perf_event_output() helpers
and BPF_MAP_TYPE_PERF_EVENT_ARRAY.  Dumping this array type is
not possible, however, we can add simple reading of perf events.
Create a new event_pipe subcommand for maps, this sub command
will only work with BPF_MAP_TYPE_PERF_EVENT_ARRAY maps.

Parts of the code from samples/bpf/trace_output_user.c.
Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: Quentin Monnet <quentin.monnet@netronome.com>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>

Move the get_possible_cpus() function to shared code.  No functional
changes.
Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: Quentin Monnet <quentin.monnet@netronome.com>
Reviewed-by: Jiong Wang <jiong.wang@netronome.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>

Instead of spelling [hex] BYTES everywhere use DATA as keyword
for generalized value.  This will help us keep the messages
concise when longer command are added in the future.  It will
also be useful once BTF support comes.  We will only have to
change the definition of DATA.
Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: Quentin Monnet <quentin.monnet@netronome.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>

Kernel will now replace map fds with actual pointer before
calling the offload prepare.  We can identify those pointers
and replace them with NFP table IDs instead of loading the
table ID in code generated for CALL instruction.

This allows us to support having the same CALL being used with
different maps.

Since we don't want to change the FW ABI we still need to
move the TID from R1 to portion of R0 before the jump.
Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: Quentin Monnet <quentin.monnet@netronome.com>
Reviewed-by: Jiong Wang <jiong.wang@netronome.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>

Add support for the perf_event_output family of helpers.

The implementation on the NFP will not match the host code exactly.
The state of the host map and rings is unknown to the device, hence
device can't return errors when rings are not installed.  The device
simply packs the data into a firmware notification message and sends
it over to the host, returning success to the program.

There is no notion of a host CPU on the device when packets are being
processed.  Device will only offload programs which set BPF_F_CURRENT_CPU.
Still, if map index doesn't match CPU no error will be returned (see
above).

Dropped/lost firmware notification messages will not cause "lost
events" event on the perf ring, they are only visible via device
error counters.

Firmware notification messages may also get reordered in respect
to the packets which caused their generation.
Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: Quentin Monnet <quentin.monnet@netronome.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>

Offloads may find host map pointers more useful than map fds.
Map pointers can be used to identify the map, while fds are
only valid within the context of loading process.

Jump to skip_full_check on error in case verifier log overflow
has to be handled (replace_map_fd_with_map_ptr() prints to the
log, driver prep may do that too in the future).
Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: Quentin Monnet <quentin.monnet@netronome.com>
Reviewed-by: Jiong Wang <jiong.wang@netronome.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>

bpf_event_output() is useful for offloads to add events to BPF
event rings, export it.  Note that export is placed near the stub
since tracing is optional and kernel/bpf/core.c is always going
to be built.
Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: Quentin Monnet <quentin.monnet@netronome.com>
Reviewed-by: Jiong Wang <jiong.wang@netronome.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>

For asynchronous events originating from the device, like perf event
output, we need to be able to make sure that objects being referred
to by the FW message are valid on the host.  FW events can get queued
and reordered.  Even if we had a FW message "barrier" we should still
protect ourselves from bogus FW output.

Add a reverse-mapping hash table and record in it all raw map pointers
FW may refer to.  Only record neutral maps, i.e. perf event arrays.
These are currently the only objects FW can refer to.  Use RCU protection
on the read side, update side is under RTNL.

Since program vs map destruction order is slightly painful for offload
simply take an extra reference on all the recorded maps to make sure
they don't disappear.
Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: Quentin Monnet <quentin.monnet@netronome.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>

BPF_MAP_TYPE_PERF_EVENT_ARRAY is special as far as offload goes.
The map only holds glue to perf ring, not actual data.  Allow
non-offloaded perf event arrays to be used in offloaded programs.
Offload driver can extract the events from HW and put them in
the map for user space to retrieve.
Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: Quentin Monnet <quentin.monnet@netronome.com>
Reviewed-by: Jiong Wang <jiong.wang@netronome.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>

Jiong Wang says:

====================
This patch set clean up some code logic related with managing subprog
information.

Part of the set are inspried by Edwin's code in his RFC:

  "bpf/verifier: subprog/func_call simplifications"

but with clearer separation so it could be easier to review.

  - Path 1 unifies main prog and subprogs. All of them are registered in
    env->subprog_starts.

  - After patch 1, it is clear that subprog_starts and subprog_stack_depth
    could be merged as both of them now have main and subprog unified.
    Patch 2 therefore does the merge, all subprog information are centred
    at bpf_subprog_info.

  - Patch 3 goes further to introduce a new fake "exit" subprog which
    serves as an ending marker to the subprog list. We could then turn the
    following code snippets across verifier:

       if (env->subprog_cnt == cur_subprog + 1)
               subprog_end = insn_cnt;
       else
               subprog_end = env->subprog_info[cur_subprog + 1].start;

    into:
       subprog_end = env->subprog_info[cur_subprog + 1].start;

There is no functional change by this patch set.
No bpf selftest (both non-jit and jit) regression found after this set.

v2:
  - fixed adjust_subprog_starts to also update fake "exit" subprog start.
  - for John's suggestion on renaming subprog to prog, I could work on
    a follow-up patch if it is recognized as worth the change.
====================
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Alexei Starovoitov <ast@kernel.org>

There are quite a few code snippet like the following in verifier:

       subprog_start = 0;
       if (env->subprog_cnt == cur_subprog + 1)
               subprog_end = insn_cnt;
       else
               subprog_end = env->subprog_info[cur_subprog + 1].start;

The reason is there is no marker in subprog_info array to tell the end of
it.

We could resolve this issue by introducing a faked "ending" subprog.
The special "ending" subprog is with "insn_cnt" as start offset, so it is
serving as the end mark whenever we iterate over all subprogs.
Signed-off-by: Jiong Wang <jiong.wang@netronome.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>

It is better to centre all subprog information fields into one structure.
This structure could later serve as function node in call graph.
Signed-off-by: Jiong Wang <jiong.wang@netronome.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>

Currently, verifier treat main prog and subprog differently. All subprogs
detected are kept in env->subprog_starts while main prog is not kept there.
Instead, main prog is implicitly defined as the prog start at 0.

There is actually no difference between main prog and subprog, it is better
to unify them, and register all progs detected into env->subprog_starts.

This could also help simplifying some code logic.
Signed-off-by: Jiong Wang <jiong.wang@netronome.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>

Daniel Borkmann says:

====================
This set simplifies BPF JITs significantly by moving ld_abs/ld_ind
to native BPF, for details see individual patches. Main rationale
is in patch 'implement ld_abs/ld_ind in native bpf'. Thanks!

v1 -> v2:
  - Added missing seen_lds_abs in LDX_MSH and use X = A
    initially due to being preserved on func call.
  - Added a large batch of cBPF tests into test_bpf.
  - Added x32 removal of LD_ABS/LD_IND, so all JITs are
    covered.
====================
Signed-off-by: Alexei Starovoitov <ast@kernel.org>

Only sync the header from include/uapi/linux/bpf.h.
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>

Since LD_ABS/LD_IND instructions are now removed from the core and
reimplemented through a combination of inlined BPF instructions and
a slow-path helper, we can get rid of the complexity from x32 JIT.
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>

Since LD_ABS/LD_IND instructions are now removed from the core and
reimplemented through a combination of inlined BPF instructions and
a slow-path helper, we can get rid of the complexity from s390x JIT.
Tested on s390x instance on LinuxONE.
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Cc: Michael Holzheu <holzheu@linux.vnet.ibm.com>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>

Since LD_ABS/LD_IND instructions are now removed from the core and
reimplemented through a combination of inlined BPF instructions and
a slow-path helper, we can get rid of the complexity from ppc64 JIT.
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Naveen N. Rao <naveen.n.rao@linux.vnet.ibm.com>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Tested-by: Sandipan Das <sandipan@linux.vnet.ibm.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>

Since LD_ABS/LD_IND instructions are now removed from the core and
reimplemented through a combination of inlined BPF instructions and
a slow-path helper, we can get rid of the complexity from mips64 JIT.
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>

Since LD_ABS/LD_IND instructions are now removed from the core and
reimplemented through a combination of inlined BPF instructions and
a slow-path helper, we can get rid of the complexity from arm32 JIT.
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>

Since LD_ABS/LD_IND instructions are now removed from the core and
reimplemented through a combination of inlined BPF instructions and
a slow-path helper, we can get rid of the complexity from sparc64 JIT.
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Acked-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>

Since LD_ABS/LD_IND instructions are now removed from the core and
reimplemented through a combination of inlined BPF instructions and
a slow-path helper, we can get rid of the complexity from arm64 JIT.
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>

Since LD_ABS/LD_IND instructions are now removed from the core and
reimplemented through a combination of inlined BPF instructions and
a slow-path helper, we can get rid of the complexity from x64 JIT.
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>

This adds a small BPF helper similar to bpf_skb_load_bytes() that
is able to load relative to mac/net header offset from the skb's
linear data. Compared to bpf_skb_load_bytes(), it takes a fifth
argument namely start_header, which is either BPF_HDR_START_MAC
or BPF_HDR_START_NET. This allows for a more flexible alternative
compared to LD_ABS/LD_IND with negative offset. It's enabled for
tc BPF programs as well as sock filter program types where it's
mainly useful in reuseport programs to ease access to lower header
data.

Reference: https://lists.iovisor.org/pipermail/iovisor-dev/2017-March/000698.htmlSigned-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>

The main part of this work is to finally allow removal of LD_ABS
and LD_IND from the BPF core by reimplementing them through native
eBPF instead. Both LD_ABS/LD_IND were carried over from cBPF and
keeping them around in native eBPF caused way more trouble than
actually worth it. To just list some of the security issues in
the past:

  * fdfaf64e ("x86: bpf_jit: support negative offsets")
  * 35607b02 ("sparc: bpf_jit: fix loads from negative offsets")
  * e0ee9c12 ("x86: bpf_jit: fix two bugs in eBPF JIT compiler")
  * 07aee943 ("bpf, sparc: fix usage of wrong reg for load_skb_regs after call")
  * 6d59b7db ("bpf, s390x: do not reload skb pointers in non-skb context")
  * 87338c8e ("bpf, ppc64: do not reload skb pointers in non-skb context")

For programs in native eBPF, LD_ABS/LD_IND are pretty much legacy
these days due to their limitations and more efficient/flexible
alternatives that have been developed over time such as direct
packet access. LD_ABS/LD_IND only cover 1/2/4 byte loads into a
register, the load happens in host endianness and its exception
handling can yield unexpected behavior. The latter is explained
in depth in f6b1b3bf ("bpf: fix subprog verifier bypass by
div/mod by 0 exception") with similar cases of exceptions we had.
In native eBPF more recent program types will disable LD_ABS/LD_IND
altogether through may_access_skb() in verifier, and given the
limitations in terms of exception handling, it's also disabled
in programs that use BPF to BPF calls.

In terms of cBPF, the LD_ABS/LD_IND is used in networking programs
to access packet data. It is not used in seccomp-BPF but programs
that use it for socket filtering or reuseport for demuxing with
cBPF. This is mostly relevant for applications that have not yet
migrated to native eBPF.

The main complexity and source of bugs in LD_ABS/LD_IND is coming
from their implementation in the various JITs. Most of them keep
the model around from cBPF times by implementing a fastpath written
in asm. They use typically two from the BPF program hidden CPU
registers for caching the skb's headlen (skb->len - skb->data_len)
and skb->data. Throughout the JIT phase this requires to keep track
whether LD_ABS/LD_IND are used and if so, the two registers need
to be recached each time a BPF helper would change the underlying
packet data in native eBPF case. At least in eBPF case, available
CPU registers are rare and the additional exit path out of the
asm written JIT helper makes it also inflexible since not all
parts of the JITer are in control from plain C. A LD_ABS/LD_IND
implementation in eBPF therefore allows to significantly reduce
the complexity in JITs with comparable performance results for
them, e.g.:

test_bpf             tcpdump port 22             tcpdump complex
x64      - before    15 21 10                    14 19  18
         - after      7 10 10                     7 10  15
arm64    - before    40 91 92                    40 91 151
         - after     51 64 73                    51 62 113

For cBPF we now track any usage of LD_ABS/LD_IND in bpf_convert_filter()
and cache the skb's headlen and data in the cBPF prologue. The
BPF_REG_TMP gets remapped from R8 to R2 since it's mainly just
used as a local temporary variable. This allows to shrink the
image on x86_64 also for seccomp programs slightly since mapping
to %rsi is not an ereg. In callee-saved R8 and R9 we now track
skb data and headlen, respectively. For normal prologue emission
in the JITs this does not add any extra instructions since R8, R9
are pushed to stack in any case from eBPF side. cBPF uses the
convert_bpf_ld_abs() emitter which probes the fast path inline
already and falls back to bpf_skb_load_helper_{8,16,32}() helper
relying on the cached skb data and headlen as well. R8 and R9
never need to be reloaded due to bpf_helper_changes_pkt_data()
since all skb access in cBPF is read-only. Then, for the case
of native eBPF, we use the bpf_gen_ld_abs() emitter, which calls
the bpf_skb_load_helper_{8,16,32}_no_cache() helper unconditionally,
does neither cache skb data and headlen nor has an inlined fast
path. The reason for the latter is that native eBPF does not have
any extra registers available anyway, but even if there were, it
avoids any reload of skb data and headlen in the first place.
Additionally, for the negative offsets, we provide an alternative
bpf_skb_load_bytes_relative() helper in eBPF which operates
similarly as bpf_skb_load_bytes() and allows for more flexibility.
Tested myself on x64, arm64, s390x, from Sandipan on ppc64.
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>

Remove all eBPF tests involving LD_ABS/LD_IND from test_bpf.ko. Reason
is that the eBPF tests from test_bpf module do not go via BPF verifier
and therefore any instruction rewrites from verifier cannot take place.

Therefore, move them into test_verifier which runs out of user space,
so that verfier can rewrite LD_ABS/LD_IND internally in upcoming patches.
It will have the same effect since runtime tests are also performed from
there. This also allows to finally unexport bpf_skb_vlan_{push,pop}_proto
and keep it internal to core kernel.

Additionally, also add further cBPF LD_ABS/LD_IND test coverage into
test_bpf.ko suite.
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>

No change in functionality, just remove the '__' prefix and replace it
with a 'bpf_' prefix instead. We later on add a couple of more helpers
for cBPF and keeping the scheme with '__' is suboptimal there.
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>

Björn Töpel says:

====================
This patch set introduces a new address family called AF_XDP that is
optimized for high performance packet processing and, in upcoming
patch sets, zero-copy semantics. In this patch set, we have removed
all zero-copy related code in order to make it smaller, simpler and
hopefully more review friendly. This patch set only supports copy-mode
for the generic XDP path (XDP_SKB) for both RX and TX and copy-mode
for RX using the XDP_DRV path. Zero-copy support requires XDP and
driver changes that Jesper Dangaard Brouer is working on. Some of his
work has already been accepted. We will publish our zero-copy support
for RX and TX on top of his patch sets at a later point in time.

An AF_XDP socket (XSK) is created with the normal socket()
syscall. Associated with each XSK are two queues: the RX queue and the
TX queue. A socket can receive packets on the RX queue and it can send
packets on the TX queue. These queues are registered and sized with
the setsockopts XDP_RX_RING and XDP_TX_RING, respectively. It is
mandatory to have at least one of these queues for each socket. In
contrast to AF_PACKET V2/V3 these descriptor queues are separated from
packet buffers. An RX or TX descriptor points to a data buffer in a
memory area called a UMEM. RX and TX can share the same UMEM so that a
packet does not have to be copied between RX and TX. Moreover, if a
packet needs to be kept for a while due to a possible retransmit, the
descriptor that points to that packet can be changed to point to
another and reused right away. This again avoids copying data.

This new dedicated packet buffer area is call a UMEM. It consists of a
number of equally size frames and each frame has a unique frame id. A
descriptor in one of the queues references a frame by referencing its
frame id. The user space allocates memory for this UMEM using whatever
means it feels is most appropriate (malloc, mmap, huge pages,
etc). This memory area is then registered with the kernel using the new
setsockopt XDP_UMEM_REG. The UMEM also has two queues: the FILL queue
and the COMPLETION queue. The fill queue is used by the application to
send down frame ids for the kernel to fill in with RX packet
data. References to these frames will then appear in the RX queue of
the XSK once they have been received. The completion queue, on the
other hand, contains frame ids that the kernel has transmitted
completely and can now be used again by user space, for either TX or
RX. Thus, the frame ids appearing in the completion queue are ids that
were previously transmitted using the TX queue. In summary, the RX and
FILL queues are used for the RX path and the TX and COMPLETION queues
are used for the TX path.

The socket is then finally bound with a bind() call to a device and a
specific queue id on that device, and it is not until bind is
completed that traffic starts to flow. Note that in this patch set,
all packet data is copied out to user-space.

A new feature in this patch set is that the UMEM can be shared between
processes, if desired. If a process wants to do this, it simply skips
the registration of the UMEM and its corresponding two queues, sets a
flag in the bind call and submits the XSK of the process it would like
to share UMEM with as well as its own newly created XSK socket. The
new process will then receive frame id references in its own RX queue
that point to this shared UMEM. Note that since the queue structures
are single-consumer / single-producer (for performance reasons), the
new process has to create its own socket with associated RX and TX
queues, since it cannot share this with the other process. This is
also the reason that there is only one set of FILL and COMPLETION
queues per UMEM. It is the responsibility of a single process to
handle the UMEM. If multiple-producer / multiple-consumer queues are
implemented in the future, this requirement could be relaxed.

How is then packets distributed between these two XSK? We have
introduced a new BPF map called XSKMAP (or BPF_MAP_TYPE_XSKMAP in
full). The user-space application can place an XSK at an arbitrary
place in this map. The XDP program can then redirect a packet to a
specific index in this map and at this point XDP validates that the
XSK in that map was indeed bound to that device and queue number. If
not, the packet is dropped. If the map is empty at that index, the
packet is also dropped. This also means that it is currently mandatory
to have an XDP program loaded (and one XSK in the XSKMAP) to be able
to get any traffic to user space through the XSK.

AF_XDP can operate in two different modes: XDP_SKB and XDP_DRV. If the
driver does not have support for XDP, or XDP_SKB is explicitly chosen
when loading the XDP program, XDP_SKB mode is employed that uses SKBs
together with the generic XDP support and copies out the data to user
space. A fallback mode that works for any network device. On the other
hand, if the driver has support for XDP, it will be used by the AF_XDP
code to provide better performance, but there is still a copy of the
data into user space.

There is a xdpsock benchmarking/test application included that
demonstrates how to use AF_XDP sockets with both private and shared
UMEMs. Say that you would like your UDP traffic from port 4242 to end
up in queue 16, that we will enable AF_XDP on. Here, we use ethtool
for this:

      ethtool -N p3p2 rx-flow-hash udp4 fn
      ethtool -N p3p2 flow-type udp4 src-port 4242 dst-port 4242 \
          action 16

Running the rxdrop benchmark in XDP_DRV mode can then be done
using:

      samples/bpf/xdpsock -i p3p2 -q 16 -r -N

For XDP_SKB mode, use the switch "-S" instead of "-N" and all options
can be displayed with "-h", as usual.

We have run some benchmarks on a dual socket system with two Broadwell
E5 2660 @ 2.0 GHz with hyperthreading turned off. Each socket has 14
cores which gives a total of 28, but only two cores are used in these
experiments. One for TR/RX and one for the user space application. The
memory is DDR4 @ 2133 MT/s (1067 MHz) and the size of each DIMM is
8192MB and with 8 of those DIMMs in the system we have 64 GB of total
memory. The compiler used is gcc (Ubuntu 7.3.0-16ubuntu3) 7.3.0. The
NIC is Intel I40E 40Gbit/s using the i40e driver.

Below are the results in Mpps of the I40E NIC benchmark runs for 64
and 1500 byte packets, generated by a commercial packet generator HW
outputing packets at full 40 Gbit/s line rate. The results are without
retpoline so that we can compare against previous numbers. With
retpoline, the AF_XDP numbers drop with between 10 - 15 percent.

AF_XDP performance 64 byte packets. Results from V2 in parenthesis.
Benchmark   XDP_SKB   XDP_DRV
rxdrop       2.9(3.0)   9.6(9.5)
txpush       2.6(2.5)   NA*
l2fwd        1.9(1.9)   2.5(2.5) (TX using XDP_SKB in both cases)

AF_XDP performance 1500 byte packets:
Benchmark   XDP_SKB   XDP_DRV
rxdrop       2.1(2.2)   3.3(3.3)
l2fwd        1.4(1.4)   1.8(1.8) (TX using XDP_SKB in both cases)

* NA since we have no support for TX using the XDP_DRV infrastructure
  in this patch set. This is for a future patch set since it involves
  changes to the XDP NDOs. Some of this has been upstreamed by Jesper
  Dangaard Brouer.

XDP performance on our system as a base line:

64 byte packets:
XDP stats       CPU     pps         issue-pps
XDP-RX CPU      16      32.3(32.9)M  0

1500 byte packets:
XDP stats       CPU     pps         issue-pps
XDP-RX CPU      16      3.3(3.3)M    0

Changes from V2:

* Fixed a race in XSKMAP map found by Will. The code has been
  completely rearchitected and is now simpler, faster, and hopefully
  also not racy. Please review and check if it holds.

If you would like to diff V2 against V3, you can find them here:
https://github.com/bjoto/linux/tree/af-xdp-v2-on-bpf-next
https://github.com/bjoto/linux/tree/af-xdp-v3-on-bpf-next

The structure of the patch set is as follows:

Patches 1-3: Basic socket and umem plumbing
Patches 4-9: RX support together with the new XSKMAP
Patches 10-13: TX support
Patch 14: Statistics support with getsockopt()
Patch 15: Sample application

We based this patch set on bpf-next commit a3fe1f6f ("tools:
bpftool: change time format for program 'loaded at:' information")

To do for this patch set:

* Syzkaller torture session being worked on

Post-series plan:

* Optimize performance

* Kernel selftest

* Kernel load module support of AF_XDP would be nice. Unclear how to
  achieve this though since our XDP code depends on net/core.

* Support for AF_XDP sockets without an XPD program loaded. In this
  case all the traffic on a queue should go up to the user space socket.

* Daniel Borkmann's suggestion for a "copy to XDP socket, and return
  XDP_PASS" for a tcpdump-like functionality.

* And of course getting to zero-copy support in small increments,
  starting with TX then adding RX.

Thanks: Björn and Magnus
====================
Acked-by: Willem de Bruijn <willemb@google.com>
Acked-by: David S. Miller <davem@davemloft.net>
Acked-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>

This is a sample application for AF_XDP sockets. The application
supports three different modes of operation: rxdrop, txonly and l2fwd.

To show-case a simple round-robin load-balancing between a set of
sockets in an xskmap, set the RR_LB compile time define option to 1 in
"xdpsock.h".

v2: The entries variable was calculated twice in {umem,xq}_nb_avail.
Co-authored-by: Björn Töpel <bjorn.topel@intel.com>
Signed-off-by: Björn Töpel <bjorn.topel@intel.com>
Signed-off-by: Magnus Karlsson <magnus.karlsson@intel.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>

In this commit, a new getsockopt is added: XDP_STATISTICS. This is
used to obtain stats from the sockets.

v2: getsockopt now returns size of stats structure.
Signed-off-by: Magnus Karlsson <magnus.karlsson@intel.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>

Here, Tx support is added. The user fills the Tx queue with frames to
be sent by the kernel, and let's the kernel know using the sendmsg
syscall.
Signed-off-by: Magnus Karlsson <magnus.karlsson@intel.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>

The new dev_direct_xmit will be used by AF_XDP in later commits.
Signed-off-by: Magnus Karlsson <magnus.karlsson@intel.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>

Another setsockopt (XDP_TX_QUEUE) is added to let the process allocate
a queue, where the user process can pass frames to be transmitted by
the kernel.

The mmapping of the queue is done using the XDP_PGOFF_TX_QUEUE offset.
Signed-off-by: Magnus Karlsson <magnus.karlsson@intel.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>

Here, we add another setsockopt for registered user memory (umem)
called XDP_UMEM_COMPLETION_QUEUE. Using this socket option, the
process can ask the kernel to allocate a queue (ring buffer) and also
mmap it (XDP_UMEM_PGOFF_COMPLETION_QUEUE) into the process.

The queue is used to explicitly pass ownership of umem frames from the
kernel to user process. This will be used by the TX path to tell user
space that a certain frame has been transmitted and user space can use
it for something else, if it wishes.
Signed-off-by: Magnus Karlsson <magnus.karlsson@intel.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>

This commit wires up the xskmap to XDP_SKB layer.
Signed-off-by: Björn Töpel <bjorn.topel@intel.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>

This commit wires up the xskmap to XDP_DRV layer.
Signed-off-by: Björn Töpel <bjorn.topel@intel.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>

The xskmap is yet another BPF map, very much inspired by
dev/cpu/sockmap, and is a holder of AF_XDP sockets. A user application
adds AF_XDP sockets into the map, and by using the bpf_redirect_map
helper, an XDP program can redirect XDP frames to an AF_XDP socket.

Note that a socket that is bound to certain ifindex/queue index will
*only* accept XDP frames from that netdev/queue index. If an XDP
program tries to redirect from a netdev/queue index other than what
the socket is bound to, the frame will not be received on the socket.

A socket can reside in multiple maps.

v3: Fixed race and simplified code.
v2: Removed one indirection in map lookup.
Signed-off-by: Björn Töpel <bjorn.topel@intel.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>

Here the actual receive functions of AF_XDP are implemented, that in a
later commit, will be called from the XDP layers.

There's one set of functions for the XDP_DRV side and another for
XDP_SKB (generic).

A new XDP API, xdp_return_buff, is also introduced.

Adding xdp_return_buff, which is analogous to xdp_return_frame, but
acts upon an struct xdp_buff. The API will be used by AF_XDP in future
commits.

Support for the poll syscall is also implemented.

v2: xskq_validate_id did not update cons_tail.
    The entries variable was calculated twice in xskq_nb_avail.
    Squashed xdp_return_buff commit.
Signed-off-by: Björn Töpel <bjorn.topel@intel.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>