John Fastabend says:

====================

Users of sockmap and skmsg trying to build proxys and other tools
have pointed out to me the error handling can be problematic. If
the proxy is under-provisioned and/or the BPF admin does not have
the ability to update/modify memory provisions on the sockets
its possible data may be dropped. For some things we have retries
so everything works out OK, but for most things this is likely
not great. And things go bad.

The original design dropped memory accounting on the receive
socket as early as possible. We did this early in sk_skb
handling and then charged it to the redirect socket immediately
after running the BPF program.

But, this design caused a fundamental problem. Namely, what should we do
if we redirect to a socket that has already reached its socket memory
limits. For proxy use cases the network admin can tune memory limits.
But, in general we punted on this problem and told folks to simply make
your memory limits high enough to handle your workload. This is not a
really good answer. When deploying into environments where we expect this
to be transparent its no longer the case because we need to tune params.
In fact its really only viable in cases where we have fine grained
control over the application. For example a proxy redirecting from an
ingress socket to an egress socket. The result is I get bug
reports because its surprising for one, but more importantly also breaks
some use cases. So lets fix it.

This series cleans up the different cases so that in many common
modes, such as passing packet up to receive socket, we can simply
use the underlying assumption that the TCP stack already has done
memory accounting.

Next instead of trying to do memory accounting against the socket
we plan to redirect into we keep memory accounting on the receive
socket until the skb can be put on the redirect socket. This means
if we do an egress redirect to a socket and sock_writable() returns
EAGAIN we can requeue the skb on the workqueue and try again. The
same scenario plays out for ingress. If the skb can not be put on
the receive queue of the redirect socket than we simply requeue and
retry. In both cases memory is still accounted for against the
receiving socket.

This also handles head of line blocking. With the above scheme the
skb is on a queue associated with the socket it will be sent/recv'd
on, but the memory accounting is against the received socket. This
means the receive socket can advance to the next skb and avoid head
of line blocking. At least until its receive memory on the socket
runs out. This will put some maximum size on the amount of data any
socket can enqueue giving us bounds on the skb lists so they can't grow
indefinitely.

Overall I think this is a win. Tested with test_sockmap.

These are fixes, but I tagged it for bpf-next considering we are
at -rc8.

v1->v2: Fix uninitialized/unused variables (kernel test robot)
v2->v3: fix typo in patch2 err=0 needs to be <0 so use err=-EIO
---
====================
Signed-off-by: Alexei Starovoitov <ast@kernel.org>

Move skb->sk assignment out of sk_psock_bpf_run() and into individual
callers. Then we can use proper skb_set_owner_r() call to assign a
sk to a skb. This improves things by also charging the truesize against
the sockets sk_rmem_alloc counter. With this done we get some accounting
in place to ensure the memory associated with skbs on the workqueue are
still being accounted for somewhere. Finally, by using skb_set_owner_r
the destructor is setup so we can just let the normal skb_kfree logic
recover the memory. Combined with previous patch dropping skb_orphan()
we now can recover from memory pressure and maintain accounting.

Note, we will charge the skbs against their originating socket even
if being redirected into another socket. Once the skb completes the
redirect op the kfree_skb will give the memory back. This is important
because if we charged the socket we are redirecting to (like it was
done before this series) the sock_writeable() test could fail because
of the skb trying to be sent is already charged against the socket.

Also TLS case is special. Here we wait until we have decided not to
simply PASS the packet up the stack. In the case where we PASS the
packet up the stack we already have an skb which is accounted for on
the TLS socket context.

For the parser case we continue to just set/clear skb->sk this is
because the skb being used here may be combined with other skbs or
turned into multiple skbs depending on the parser logic. For example
the parser could request a payload length greater than skb->len so
that the strparser needs to collect multiple skbs. At any rate
the final result will be handled in the strparser recv callback.

Fixes: 604326b4 ("bpf, sockmap: convert to generic sk_msg interface")
Signed-off-by: John Fastabend <john.fastabend@gmail.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/160226867513.5692.10579573214635925960.stgit@john-Precision-5820-Tower

Calling skb_orphan() is unnecessary in the strp rcv handler because the skb
is from a skb_clone() in __strp_recv. So it never has a destructor or a
sk assigned. Plus its confusing to read because it might hint to the reader
that the skb could have an sk assigned which is not true. Even if we did
have an sk assigned it would be cleaner to simply wait for the upcoming
kfree_skb().

Additionally, move the comment about strparser clone up so its closer to
the logic it is describing and add to it so that it is more complete.

Fixes: 604326b4 ("bpf, sockmap: convert to generic sk_msg interface")
Signed-off-by: John Fastabend <john.fastabend@gmail.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/160226865548.5692.9098315689984599579.stgit@john-Precision-5820-Tower

In the sk_skb redirect case we didn't handle the case where we overrun
the sk_rmem_alloc entry on ingress redirect or sk_wmem_alloc on egress.
Because we didn't have anything implemented we simply dropped the skb.
This meant data could be dropped if socket memory accounting was in
place.

This fixes the above dropped data case by moving the memory checks
later in the code where we actually do the send or recv. This pushes
those checks into the workqueue and allows us to return an EAGAIN error
which in turn allows us to try again later from the workqueue.

Fixes: 51199405 ("bpf: skb_verdict, support SK_PASS on RX BPF path")
Signed-off-by: John Fastabend <john.fastabend@gmail.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/160226863689.5692.13861422742592309285.stgit@john-Precision-5820-Tower

The skb_set_owner_w is unnecessary here. The sendpage call will create a
fresh skb and set the owner correctly from workqueue. Its also not entirely
harmless because it consumes cycles, but also impacts resource accounting
by increasing sk_wmem_alloc. This is charging the socket we are going to
send to for the skb, but we will put it on the workqueue for some time
before this happens so we are artifically inflating sk_wmem_alloc for
this period. Further, we don't know how many skbs will be used to send the
packet or how it will be broken up when sent over the new socket so
charging it with one big sum is also not correct when the workqueue may
break it up if facing memory pressure. Seeing we don't know how/when
this is going to be sent drop the early accounting.

A later patch will do proper accounting charged on receive socket for
the case where skbs get enqueued on the workqueue.

Fixes: 604326b4 ("bpf, sockmap: convert to generic sk_msg interface")
Signed-off-by: John Fastabend <john.fastabend@gmail.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/160226861708.5692.17964237936462425136.stgit@john-Precision-5820-Tower

When we receive an skb and the ingress skb verdict program returns
SK_PASS we currently set the ingress flag and put it on the workqueue
so it can be turned into a sk_msg and put on the sk_msg ingress queue.
Then finally telling userspace with data_ready hook.

Here we observe that if the workqueue is empty then we can try to
convert into a sk_msg type and call data_ready directly without
bouncing through a workqueue. Its a common pattern to have a recv
verdict program for visibility that always returns SK_PASS. In this
case unless there is an ENOMEM error or we overrun the socket we
can avoid the workqueue completely only using it when we fall back
to error cases caused by memory pressure.

By doing this we eliminate another case where data may be dropped
if errors occur on memory limits in workqueue.

Fixes: 51199405 ("bpf: skb_verdict, support SK_PASS on RX BPF path")
Signed-off-by: John Fastabend <john.fastabend@gmail.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/160226859704.5692.12929678876744977669.stgit@john-Precision-5820-Tower

For sk_skb case where skb_verdict program returns SK_PASS to continue to
pass packet up the stack, the memory limits were already checked before
enqueuing in skb_queue_tail from TCP side. So, lets remove the extra checks
here. The theory is if the TCP stack believes we have memory to receive
the packet then lets trust the stack and not double check the limits.

In fact the accounting here can cause a drop if sk_rmem_alloc has increased
after the stack accepted this packet, but before the duplicate check here.
And worse if this happens because TCP stack already believes the data has
been received there is no retransmit.

Fixes: 51199405 ("bpf: skb_verdict, support SK_PASS on RX BPF path")
Signed-off-by: John Fastabend <john.fastabend@gmail.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/160226857664.5692.668205469388498375.stgit@john-Precision-5820-Tower

Move the bpf/bpf-next patch processing queue to patchwork.kernel.org.
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Link: https://lore.kernel.org/bpf/20201011200149.66537-1-alexei.starovoitov@gmail.com

The bpf_fib_lookup() helper performs a neighbour lookup for the destination
IP and returns BPF_FIB_LKUP_NO_NEIGH if this fails, with the expectation
that the BPF program will pass the packet up the stack in this case.
However, with the addition of bpf_redirect_neigh() that can be used instead
to perform the neighbour lookup, at the cost of a bit of duplicated work.

For that we still need the target ifindex, and since bpf_fib_lookup()
already has that at the time it performs the neighbour lookup, there is
really no reason why it can't just return it in any case. So let's just
always return the ifindex if the FIB lookup itself succeeds.
Signed-off-by: Toke Høiland-Jørgensen <toke@redhat.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Cc: David Ahern <dsahern@gmail.com>
Link: https://lore.kernel.org/bpf/20201009184234.134214-1-toke@redhat.com

"Daniel T. Lee" says:

====================
To avoid confusion caused by the increasing fragmentation of the BPF
Loader program, this commit would like to convert the previous bpf_load
loader with the libbpf loader.

Thanks to libbpf's bpf_link interface, managing the tracepoint BPF
program is much easier. bpf_program__attach_tracepoint manages the
enable of tracepoint event and attach of BPF programs to it with a
single interface bpf_link, so there is no need to manage event_fd and
prog_fd separately.

And due to addition of generic bpf_program__attach() to libbpf, it is
now possible to attach BPF programs with __attach() instead of
explicitly calling __attach_<type>().

This patchset refactors xdp_monitor with using this libbpf API, and the
bpf_load is removed and migrated to libbpf. Also, attach_tracepoint()
is replaced with the generic __attach() method in xdp_redirect_cpu.
Moreover, maps in kern program have been converted to BTF-defined map.
---
Changes in v2:
 - added cleanup logic for bpf_link and bpf_object in xdp_monitor
 - program section match with bpf_program__is_<type> instead of strncmp
 - revert BTF key/val type to default of BPF_MAP_TYPE_PERF_EVENT_ARRAY
 - split increment into seperate satement
 - refactor pointer array initialization
 - error code cleanup
====================
Signed-off-by: Alexei Starovoitov <ast@kernel.org>

Most of the samples were converted to use the new BTF-defined MAP as
they moved to libbpf, but some of the samples were missing.

Instead of using the previous BPF MAP definition, this commit refactors
xdp_monitor and xdp_sample_pkts_kern MAP definition with the new
BTF-defined MAP format.

Also, this commit removes the max_entries attribute at PERF_EVENT_ARRAY
map type. The libbpf's bpf_object__create_map() will automatically
set max_entries to the maximum configured number of CPUs on the host.
Signed-off-by: Daniel T. Lee <danieltimlee@gmail.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Acked-by: Andrii Nakryiko <andrii@kernel.org>
Link: https://lore.kernel.org/bpf/20201010181734.1109-4-danieltimlee@gmail.com

>From commit d7a18ea7 ("libbpf: Add generic bpf_program__attach()"),
for some BPF programs, it is now possible to attach BPF programs
with __attach() instead of explicitly calling __attach_<type>().

This commit refactors the __attach_tracepoint() with libbpf's generic
__attach() method. In addition, this refactors the logic of setting
the map FD to simplify the code. Also, the missing removal of
bpf_load.o in Makefile has been fixed.
Signed-off-by: Daniel T. Lee <danieltimlee@gmail.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Acked-by: Andrii Nakryiko <andrii@kernel.org>
Link: https://lore.kernel.org/bpf/20201010181734.1109-3-danieltimlee@gmail.com

To avoid confusion caused by the increasing fragmentation of the BPF
Loader program, this commit would like to change to the libbpf loader
instead of using the bpf_load.

Thanks to libbpf's bpf_link interface, managing the tracepoint BPF
program is much easier. bpf_program__attach_tracepoint manages the
enable of tracepoint event and attach of BPF programs to it with a
single interface bpf_link, so there is no need to manage event_fd and
prog_fd separately.

This commit refactors xdp_monitor with using this libbpf API, and the
bpf_load is removed and migrated to libbpf.
Signed-off-by: Daniel T. Lee <danieltimlee@gmail.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Acked-by: Andrii Nakryiko <andrii@kernel.org>
Link: https://lore.kernel.org/bpf/20201010181734.1109-2-danieltimlee@gmail.com

Daniel Borkmann says:

====================

This series addresses most of the feedback [0] that was to be followed
up from the last series, that is, UAPI helper comment improvements and
getting rid of the ifindex obj file hacks in the selftest by using a
BPF map instead. The __sk_buff data/data_end pointer work, I'm planning
to do in a later round as well as the mem*() BPF improvements we have
in Cilium for libbpf. Next, the series adds two features, i) a helper
called redirect_peer() to improve latency on netns switch, and ii) to
allow map in map with dynamic inner array map sizes. Selftests for each
are added as well. For details, please check individual patches, thanks!

  [0] https://lore.kernel.org/bpf/cover.1601477936.git.daniel@iogearbox.net/

v5 -> v6:
  - Going with Andrii's suggestion to make the misconfigured verifier
    test more robust, and only probe on -EOPNOTSUPP (Andrii)
v4 -> v5:
  - Replace cnt == -EOPNOTSUPP check with cnt < 0; I've used < 0
    here as I think it's useful to keep the existing cnt == 0 ||
    cnt >= ARRAY_SIZE(insn_buf) for error detection (Andrii)
v3 -> v4:
  - Rename new array map flag to BPF_F_INNER_MAP (Alexei)
v2 -> v3:
  - Remove tab that slipped into uapi helper desc (Jakub)
  - Rework map in map for array to error from map_gen_lookup (Andrii)
v1 -> v2:
  - Fixed selftest comment wrt inner1/inner2 value (Yonghong)
====================
Signed-off-by: Alexei Starovoitov <ast@kernel.org>

Extend the test_tc_redirect test and add a small test that exercises the new
redirect_peer() helper for the IPv4 and IPv6 case.
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/20201010234006.7075-7-daniel@iogearbox.net

Rename into test_tc_redirect.sh and move setup and test code into separate
functions so they can be reused for newly added tests in here. Also remove
the crude hack to override ifindex inside the object file via xxd and sed
and just use a simple map instead. Map given iproute2 does not support BTF
fully and therefore neither global data at this point.
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Acked-by: Yonghong Song <yhs@fb.com>
Link: https://lore.kernel.org/bpf/20201010234006.7075-6-daniel@iogearbox.net

Extend the "diff_size" subtest to also include a non-inlined array map variant
where dynamic inner #elems are possible.
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Acked-by: Yonghong Song <yhs@fb.com>
Link: https://lore.kernel.org/bpf/20201010234006.7075-5-daniel@iogearbox.net

Recent work in f4d05259 ("bpf: Add map_meta_equal map ops") and 134fede4
("bpf: Relax max_entries check for most of the inner map types") added support
for dynamic inner max elements for most map-in-map types. Exceptions were maps
like array or prog array where the map_gen_lookup() callback uses the maps'
max_entries field as a constant when emitting instructions.

We recently implemented Maglev consistent hashing into Cilium's load balancer
which uses map-in-map with an outer map being hash and inner being array holding
the Maglev backend table for each service. This has been designed this way in
order to reduce overall memory consumption given the outer hash map allows to
avoid preallocating a large, flat memory area for all services. Also, the
number of service mappings is not always known a-priori.

The use case for dynamic inner array map entries is to further reduce memory
overhead, for example, some services might just have a small number of back
ends while others could have a large number. Right now the Maglev backend table
for small and large number of backends would need to have the same inner array
map entries which adds a lot of unneeded overhead.

Dynamic inner array map entries can be realized by avoiding the inlined code
generation for their lookup. The lookup will still be efficient since it will
be calling into array_map_lookup_elem() directly and thus avoiding retpoline.
The patch adds a BPF_F_INNER_MAP flag to map creation which therefore skips
inline code generation and relaxes array_map_meta_equal() check to ignore both
maps' max_entries. This also still allows to have faster lookups for map-in-map
when BPF_F_INNER_MAP is not specified and hence dynamic max_entries not needed.

Example code generation where inner map is dynamic sized array:

  # bpftool p d x i 125
  int handle__sys_enter(void * ctx):
  ; int handle__sys_enter(void *ctx)
     0: (b4) w1 = 0
  ; int key = 0;
     1: (63) *(u32 *)(r10 -4) = r1
     2: (bf) r2 = r10
  ;
     3: (07) r2 += -4
  ; inner_map = bpf_map_lookup_elem(&outer_arr_dyn, &key);
     4: (18) r1 = map[id:468]
     6: (07) r1 += 272
     7: (61) r0 = *(u32 *)(r2 +0)
     8: (35) if r0 >= 0x3 goto pc+5
     9: (67) r0 <<= 3
    10: (0f) r0 += r1
    11: (79) r0 = *(u64 *)(r0 +0)
    12: (15) if r0 == 0x0 goto pc+1
    13: (05) goto pc+1
    14: (b7) r0 = 0
    15: (b4) w6 = -1
  ; if (!inner_map)
    16: (15) if r0 == 0x0 goto pc+6
    17: (bf) r2 = r10
  ;
    18: (07) r2 += -4
  ; val = bpf_map_lookup_elem(inner_map, &key);
    19: (bf) r1 = r0                               | No inlining but instead
    20: (85) call array_map_lookup_elem#149280     | call to array_map_lookup_elem()
  ; return val ? *val : -1;                        | for inner array lookup.
    21: (15) if r0 == 0x0 goto pc+1
  ; return val ? *val : -1;
    22: (61) r6 = *(u32 *)(r0 +0)
  ; }
    23: (bc) w0 = w6
    24: (95) exit
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Acked-by: Andrii Nakryiko <andrii@kernel.org>
Link: https://lore.kernel.org/bpf/20201010234006.7075-4-daniel@iogearbox.net

Add an efficient ingress to ingress netns switch that can be used out of tc BPF
programs in order to redirect traffic from host ns ingress into a container
veth device ingress without having to go via CPU backlog queue [0]. For local
containers this can also be utilized and path via CPU backlog queue only needs
to be taken once, not twice. On a high level this borrows from ipvlan which does
similar switch in __netif_receive_skb_core() and then iterates via another_round.
This helps to reduce latency for mentioned use cases.

Pod to remote pod with redirect(), TCP_RR [1]:

  # percpu_netperf 10.217.1.33
          RT_LATENCY:         122.450         (per CPU:         122.666         122.401         122.333         122.401 )
        MEAN_LATENCY:         121.210         (per CPU:         121.100         121.260         121.320         121.160 )
      STDDEV_LATENCY:         120.040         (per CPU:         119.420         119.910         125.460         115.370 )
         MIN_LATENCY:          46.500         (per CPU:          47.000          47.000          47.000          45.000 )
         P50_LATENCY:         118.500         (per CPU:         118.000         119.000         118.000         119.000 )
         P90_LATENCY:         127.500         (per CPU:         127.000         128.000         127.000         128.000 )
         P99_LATENCY:         130.750         (per CPU:         131.000         131.000         129.000         132.000 )

    TRANSACTION_RATE:       32666.400         (per CPU:        8152.200        8169.842        8174.439        8169.897 )

Pod to remote pod with redirect_peer(), TCP_RR:

  # percpu_netperf 10.217.1.33
          RT_LATENCY:          44.449         (per CPU:          43.767          43.127          45.279          45.622 )
        MEAN_LATENCY:          45.065         (per CPU:          44.030          45.530          45.190          45.510 )
      STDDEV_LATENCY:          84.823         (per CPU:          66.770          97.290          84.380          90.850 )
         MIN_LATENCY:          33.500         (per CPU:          33.000          33.000          34.000          34.000 )
         P50_LATENCY:          43.250         (per CPU:          43.000          43.000          43.000          44.000 )
         P90_LATENCY:          46.750         (per CPU:          46.000          47.000          47.000          47.000 )
         P99_LATENCY:          52.750         (per CPU:          51.000          54.000          53.000          53.000 )

    TRANSACTION_RATE:       90039.500         (per CPU:       22848.186       23187.089       22085.077       21919.130 )

  [0] https://linuxplumbersconf.org/event/7/contributions/674/attachments/568/1002/plumbers_2020_cilium_load_balancer.pdf
  [1] https://github.com/borkmann/netperf_scripts/blob/master/percpu_netperfSigned-off-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/20201010234006.7075-3-daniel@iogearbox.net

Follow-up to address David's feedback that we should better describe internals
of the bpf_redirect_neigh() helper.
Suggested-by: David Ahern <dsahern@gmail.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Reviewed-by: David Ahern <dsahern@gmail.com>
Link: https://lore.kernel.org/bpf/20201010234006.7075-2-daniel@iogearbox.net

Alexei Starovoitov says:

====================
Make two verifier improvements:

- The llvm register allocator may use two different registers representing
  the same virtual register. Teach the verifier to recognize that.
- Track bounded scalar spill/fill.

The profiler[123] test in patch 3 will fail to load without patches 1 and 2.
The profiler[23] test may fail to load on older llvm due to speculative
code motion nd instruction combining optimizations that are fixed in
https://reviews.llvm.org/D85570

v1 -> v2:
  - fixed 32-bit mov issue spotted by John.
  - allowed r2=r1; r3=r2; sequence as suggested by John.
  - added comments, acks, more tests.
====================
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>

Add asm tests for register allocator tracking logic.
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Andrii Nakryiko <andrii@kernel.org>
Link: https://lore.kernel.org/bpf/20201009011240.48506-5-alexei.starovoitov@gmail.com

The main purpose of the profiler test to check different llvm generation
patterns to make sure the verifier can load these large programs.

Note that profiler.inc.h test doesn't follow strict kernel coding style.
The code was formatted in the kernel style, but variable declarations are
kept as-is to preserve original llvm IR pattern.

profiler1.c should pass with older and newer llvm

profiler[23].c may fail on older llvm that don't have:
https://reviews.llvm.org/D85570
because llvm may do speculative code motion optimization that
will generate code like this:

// r9 is a pointer to map_value
// r7 is a scalar
17:       bf 96 00 00 00 00 00 00 r6 = r9
18:       0f 76 00 00 00 00 00 00 r6 += r7
19:       a5 07 01 00 01 01 00 00 if r7 < 257 goto +1
20:       bf 96 00 00 00 00 00 00 r6 = r9
// r6 is used here

The verifier will reject such code with the error:
"math between map_value pointer and register with unbounded min value is not allowed"
At insn 18 the r7 is indeed unbounded. The later insn 19 checks the bounds and
the insn 20 undoes map_value addition. It is currently impossible for the
verifier to understand such speculative pointer arithmetic. Hence llvm D85570
addresses it on the compiler side.
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Yonghong Song <yhs@fb.com>
Link: https://lore.kernel.org/bpf/20201009011240.48506-4-alexei.starovoitov@gmail.com

Under register pressure the llvm may spill registers with bounds into the stack.
The verifier has to track them through spill/fill otherwise many kinds of bound
errors will be seen. The spill/fill of induction variables was already
happening. This patch extends this logic from tracking spill/fill of a constant
into any bounded register. There is no need to track spill/fill of unbounded,
since no new information will be retrieved from the stack during register fill.

Though extra stack difference could cause state pruning to be less effective, no
adverse affects were seen from this patch on selftests and on cilium programs.
Signed-off-by: Yonghong Song <yhs@fb.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Andrii Nakryiko <andrii@kernel.org>
Acked-by: John Fastabend <john.fastabend@gmail.com>
Link: https://lore.kernel.org/bpf/20201009011240.48506-3-alexei.starovoitov@gmail.com

The llvm register allocator may use two different registers representing the
same virtual register. In such case the following pattern can be observed:
1047: (bf) r9 = r6
1048: (a5) if r6 < 0x1000 goto pc+1
1050: ...
1051: (a5) if r9 < 0x2 goto pc+66
1052: ...
1053: (bf) r2 = r9 /* r2 needs to have upper and lower bounds */

This is normal behavior of greedy register allocator.
The slides 137+ explain why regalloc introduces such register copy:
http://llvm.org/devmtg/2018-04/slides/Yatsina-LLVM%20Greedy%20Register%20Allocator.pdf
There is no way to tell llvm 'not to do this'.
Hence the verifier has to recognize such patterns.

In order to track this information without backtracking allocate ID
for scalars in a similar way as it's done for find_good_pkt_pointers().

When the verifier encounters r9 = r6 assignment it will assign the same ID
to both registers. Later if either register range is narrowed via conditional
jump propagate the register state into the other register.

Clear register ID in adjust_reg_min_max_vals() for any alu instruction. The
register ID is ignored for scalars in regsafe() and doesn't affect state
pruning. mark_reg_unknown() clears the ID. It's used to process call, endian
and other instructions. Hence ID is explicitly cleared only in
adjust_reg_min_max_vals() and in 32-bit mov.
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Andrii Nakryiko <andrii@kernel.org>
Acked-by: John Fastabend <john.fastabend@gmail.com>
Link: https://lore.kernel.org/bpf/20201009011240.48506-2-alexei.starovoitov@gmail.com

Adding support for TCP_NOTSENT_LOWAT sockoption (https://lwn.net/Articles/560082/)
in tcp bpf programs.
Signed-off-by: Nikita V. Shirokov <tehnerd@tehnerd.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Yonghong Song <yhs@fb.com>
Link: https://lore.kernel.org/bpf/20201009070325.226855-1-tehnerd@tehnerd.com

Introduce one cache line worth of padding between the producer and
consumer pointers in all the lockless rings. This so that the HW
adjacency prefetcher will not prefetch the consumer pointer when the
producer pointer is used and vice versa. This improves throughput
performance for the l2fwd sample app with 2% on my machine with HW
prefetching turned on.
Signed-off-by: Magnus Karlsson <magnus.karlsson@intel.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Link: https://lore.kernel.org/bpf/1602166338-21378-1-git-send-email-magnus.karlsson@gmail.com

Andrii Nakryiko says:

====================
Patch set implements logic in libbpf to auto-adjust memory size (1-, 2-, 4-,
8-bytes) of load/store (LD/ST/STX) instructions which have BPF CO-RE field
offset relocation associated with it. In practice this means transparent
handling of 32-bit kernels, both pointer and unsigned integers. Signed
integers are not relocatable with zero-extending loads/stores, so libbpf
poisons them and generates a warning. If/when BPF gets support for
sign-extending loads/stores, it would be possible to automatically relocate
them as well.

All the details are contained in patch #2 comments and commit message.
Patch #3 is a simple change in libbpf to make advanced testing with custom BTF
easier. Patch #4 validates correct uses of auto-resizable loads, as well as
check that libbpf fails invalid uses. Patch #1 skips CO-RE relocation for
programs that had bpf_program__set_autoload(prog, false) set on them, reducing
warnings and noise.

v2->v3:
  - fix copyright (Alexei);
v1->v2:
  - more consistent names for instruction mem size convertion routines (Alexei);
  - extended selftests to use relocatable STX instructions (Alexei);
  - added a fix for skipping CO-RE relocation for non-loadable programs.

Cc: Luka Perkov <luka.perkov@sartura.hr>
Cc: Tony Ambardar <tony.ambardar@gmail.com>
====================
Signed-off-by: Alexei Starovoitov <ast@kernel.org>

Add selftests validating libbpf's auto-resizing of load/store instructions
when used with CO-RE relocations. An explicit and manual approach with using
bpf_core_read() is also demonstrated and tested. Separate BPF program is
supposed to fail due to using signed integers of sizes that differ from
kernel's sizes.

To reliably simulate 32-bit BTF (i.e., the one with sizeof(long) ==
sizeof(void *) == 4), selftest generates its own custom BTF and passes it as
a replacement for real kernel BTF. This allows to test 32/64-bitness mix on
all architectures.
Signed-off-by: Andrii Nakryiko <andrii@kernel.org>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/20201008001025.292064-5-andrii@kernel.org

Use generalized BTF parsing logic, making it possible to parse BTF both from
ELF file, as well as a raw BTF dump. This makes it easier to write custom
tests with manually generated BTFs.
Signed-off-by: Andrii Nakryiko <andrii@kernel.org>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/20201008001025.292064-4-andrii@kernel.org

Add support for patching instructions of the following form:
  - rX = *(T *)(rY + <off>);
  - *(T *)(rX + <off>) = rY;
  - *(T *)(rX + <off>) = <imm>, where T is one of {u8, u16, u32, u64}.

For such instructions, if the actual kernel field recorded in CO-RE relocation
has a different size than the one recorded locally (e.g., from vmlinux.h),
then libbpf will adjust T to an appropriate 1-, 2-, 4-, or 8-byte loads.

In general, such transformation is not always correct and could lead to
invalid final value being loaded or stored. But two classes of cases are
always safe:
  - if both local and target (kernel) types are unsigned integers, but of
  different sizes, then it's OK to adjust load/store instruction according to
  the necessary memory size. Zero-extending nature of such instructions and
  unsignedness make sure that the final value is always correct;
  - pointer size mismatch between BPF target architecture (which is always
  64-bit) and 32-bit host kernel architecture can be similarly resolved
  automatically, because pointer is essentially an unsigned integer. Loading
  32-bit pointer into 64-bit BPF register with zero extension will leave
  correct pointer in the register.

Both cases are necessary to support CO-RE on 32-bit kernels, as `unsigned
long` in vmlinux.h generated from 32-bit kernel is 32-bit, but when compiled
with BPF program for BPF target it will be treated by compiler as 64-bit
integer. Similarly, pointers in vmlinux.h are 32-bit for kernel, but treated
as 64-bit values by compiler for BPF target. Both problems are now resolved by
libbpf for direct memory reads.

But similar transformations are useful in general when kernel fields are
"resized" from, e.g., unsigned int to unsigned long (or vice versa).

Now, similar transformations for signed integers are not safe to perform as
they will result in incorrect sign extension of the value. If such situation
is detected, libbpf will emit helpful message and will poison the instruction.
Not failing immediately means that it's possible to guard the instruction
based on kernel version (or other conditions) and make sure it's not
reachable.

If there is a need to read signed integers that change sizes between different
kernels, it's possible to use BPF_CORE_READ_BITFIELD() macro, which works both
with bitfields and non-bitfield integers of any signedness and handles
sign-extension properly. Also, bpf_core_read() with proper size and/or use of
bpf_core_field_size() relocation could allow to deal with such complicated
situations explicitly, if not so conventiently as direct memory reads.

Selftests added in a separate patch in progs/test_core_autosize.c demonstrate
both direct memory and probed use cases.

BPF_CORE_READ() is not changed and it won't deal with such situations as
automatically as direct memory reads due to the signedness integer
limitations, which are much harder to detect and control with compiler macro
magic. So it's encouraged to utilize direct memory reads as much as possible.
Signed-off-by: Andrii Nakryiko <andrii@kernel.org>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/20201008001025.292064-3-andrii@kernel.org

Bypass CO-RE relocations step for BPF programs that are not going to be
loaded. This allows to have BPF programs compiled in and disabled dynamically
if kernel is not supposed to provide enough relocation information. In such
case, there won't be unnecessary warnings about failed relocations.

Fixes: d9297581 ("libbpf: Support disabling auto-loading BPF programs")
Signed-off-by: Andrii Nakryiko <andrii@kernel.org>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/20201008001025.292064-2-andrii@kernel.org

Fix a compatibility problem when the old XDP_SHARED_UMEM mode is used
together with the xsk_socket__create() call. In the old XDP_SHARED_UMEM
mode, only sharing of the same device and queue id was allowed, and
in this mode, the fill ring and completion ring were shared between
the AF_XDP sockets.

Therefore, it was perfectly fine to call the xsk_socket__create() API
for each socket and not use the new xsk_socket__create_shared() API.
This behavior was ruined by the commit introducing XDP_SHARED_UMEM
support between different devices and/or queue ids. This patch restores
the ability to use xsk_socket__create in these circumstances so that
backward compatibility is not broken.

Fixes: 2f6324a3 ("libbpf: Support shared umems between queues and devices")
Signed-off-by: Magnus Karlsson <magnus.karlsson@intel.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Link: https://lore.kernel.org/bpf/1602070946-11154-1-git-send-email-magnus.karlsson@gmail.com

Reported-by: Samanta Navarro <ferivoz@riseup.net>
Signed-off-by: Jakub Wilk <jwilk@jwilk.net>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Acked-by: Yonghong Song <yhs@fb.com>
Link: https://lore.kernel.org/bpf/20201007055717.7319-1-jwilk@jwilk.net

When CONFIG_NET is not defined, I hit the following build error:
    kernel/trace/bpf_trace.o:(.rodata+0x110): undefined reference to `bpf_prog_test_run_raw_tp'

Commit 1b4d60ec ("bpf: Enable BPF_PROG_TEST_RUN for raw_tracepoint")
added test_run support for raw_tracepoint in /kernel/trace/bpf_trace.c.
But the test_run function bpf_prog_test_run_raw_tp is defined in
net/bpf/test_run.c, only available with CONFIG_NET=y.

Adding a CONFIG_NET guard for
    .test_run = bpf_prog_test_run_raw_tp;
fixed the above build issue.

Fixes: 1b4d60ec ("bpf: Enable BPF_PROG_TEST_RUN for raw_tracepoint")
Signed-off-by: Yonghong Song <yhs@fb.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/20201007062933.3425899-1-yhs@fb.com

Fix build errors in kernel/bpf/verifier.c when CONFIG_NET is
not enabled.

../kernel/bpf/verifier.c:3995:13: error: ‘btf_sock_ids’ undeclared here (not in a function); did you mean ‘bpf_sock_ops’?
  .btf_id = &btf_sock_ids[BTF_SOCK_TYPE_SOCK_COMMON],

../kernel/bpf/verifier.c:3995:26: error: ‘BTF_SOCK_TYPE_SOCK_COMMON’ undeclared here (not in a function); did you mean ‘PTR_TO_SOCK_COMMON’?
  .btf_id = &btf_sock_ids[BTF_SOCK_TYPE_SOCK_COMMON],

Fixes: 1df8f55a ("bpf: Enable bpf_skc_to_* sock casting helper to networking prog type")
Signed-off-by: Randy Dunlap <rdunlap@infradead.org>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Acked-by: Yonghong Song <yhs@fb.com>
Link: https://lore.kernel.org/bpf/20201007021613.13646-1-rdunlap@infradead.org

Commit 4976b718 ("bpf: Introduce pseudo_btf_id") switched
the order of check_subprogs() and resolve_pseudo_ldimm() in
the verifier. Now an empty prog expects to see the error "last
insn is not an the prog of a single invalid ldimm exit or jmp"
instead, because the check for subprogs comes first. It's now
pointless to validate that half of ldimm64 won't be the last
instruction.

Tested:
 # ./test_verifier
 Summary: 1129 PASSED, 537 SKIPPED, 0 FAILED
 and the full set of bpf selftests.

Fixes: 4976b718 ("bpf: Introduce pseudo_btf_id")
Signed-off-by: Hao Luo <haoluo@google.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/20201007022857.2791884-1-haoluo@google.com

Add an option to count the number of interrupts generated per second and
total number of interrupts during the lifetime of the application for a
given interface. This information is extracted from /proc/interrupts. Since
there is no naming convention across drivers, the user must provide the
string which is specific to their interface in the /proc/interrupts file on
the command line.

Usage:

./xdpsock ... -I <irq_str>

eg. for queue 0 of i40e device eth0:

./xdpsock ... -I i40e-eth0-TxRx-0
Signed-off-by: Ciara Loftus <ciara.loftus@intel.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/20201002133612.31536-3-ciara.loftus@intel.com

Categorise and record syscalls issued in the xdpsock sample app. The
categories recorded are:

  rx_empty_polls:    polls when the rx ring is empty
  fill_fail_polls:   polls when failed to get addr from fill ring
  copy_tx_sendtos:   sendtos issued for tx when copy mode enabled
  tx_wakeup_sendtos: sendtos issued when tx ring needs waking up
  opt_polls:         polls issued since the '-p' flag is set

Print the stats using '-a' on the xdpsock command line.
Signed-off-by: Ciara Loftus <ciara.loftus@intel.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Acked-by: Yonghong Song <yhs@fb.com>
Link: https://lore.kernel.org/bpf/20201002133612.31536-2-ciara.loftus@intel.com

New statistics will be added in future commits. In preparation for this,
let's split out the existing statistics into their own struct.
Signed-off-by: Ciara Loftus <ciara.loftus@intel.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Acked-by: Yonghong Song <yhs@fb.com>
Link: https://lore.kernel.org/bpf/20201002133612.31536-1-ciara.loftus@intel.com