To be able to run selftests without any hardware required we
need a software model.  The model can also serve as an example
implementation for those implementing actual HW offloads.
The dummy driver have previously been extended to test SR-IOV,
but the general consensus seems to be against adding further
features to it.

Add a new driver for purposes of software modelling only.
eBPF and SR-IOV will be added here shortly, others are invited
to further extend the driver with their offload models.
Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: Simon Horman <simon.horman@netronome.com>
Reviewed-by: Quentin Monnet <quentin.monnet@netronome.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>

Since day one of XDP drivers had to remember to free the program
on the remove path.  This leads to code duplication and is error
prone.  Make the stack query the installed programs on unregister
and if something is installed, remove the program.  Freeing of
program attached to XDP generic is moved from free_netdev() as well.

Because the remove will now be called before notifiers are
invoked, BPF offload state of the program will not get destroyed
before uninstall.
Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: Simon Horman <simon.horman@netronome.com>
Reviewed-by: Quentin Monnet <quentin.monnet@netronome.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>

Some drivers enforce that flags on program replacement and
removal must match the flags passed on install.  This leaves
the possibility open to enable simultaneous loading
of XDP programs both to HW and DRV.

Allow such drivers to report the flags back to the stack.
Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: Simon Horman <simon.horman@netronome.com>
Reviewed-by: Quentin Monnet <quentin.monnet@netronome.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>

The output parameters will get unwieldy if we want to add more
information about the program.  Simply pass the entire
struct netdev_bpf in.
Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: Simon Horman <simon.horman@netronome.com>
Reviewed-by: Quentin Monnet <quentin.monnet@netronome.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>

Jiong Wang says:

====================
Currently, compiler will lower memcpy function call in XDP/eBPF C program
into a sequence of eBPF load/store pairs for some scenarios.

Compiler is thinking this "inline" optimiation is beneficial as it could
avoid function call and also increase code locality.

However, Netronome NPU is not an tranditional load/store architecture that
doing a sequence of individual load/store actions are not efficient.

This patch set tries to identify the load/store sequences composed of
load/store pairs that comes from memcpy lowering, then accelerates them
through NPU's Command Push Pull (CPP) instruction.

This patch set registered an new optimization pass before doing the actual
JIT work, it traverse through eBPF IR, once found candidate sequence then
record the memory copy source, destination and length information in the
first load instruction starting the sequence and marks all remaining
instructions in the sequence into skipable status. Later, when JITing the
first load instructoin, optimal instructions will be generated using those
record information.

For this safety of this transformation:

  - jump into the middle of the sequence will cancel the optimization.

  - overlapped memory access will cancel the optimization.

  - the load destination register still contains the same value as before
    the transformation.
====================
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>

This patch add the optimization frontend, but adding a new eBPF IR scan
pass "nfp_bpf_opt_ldst_gather".

The pass will traverse the IR to recognize the load/store pairs sequences
that come from lowering of memory copy builtins.

The gathered memory copy information will be kept in the meta info
structure of the first load instruction in the sequence and will be
consumed by the optimization backend added in the previous patches.

NOTE: a sequence with cross memory access doesn't qualify this
optimization, i.e. if one load in the sequence will load from place that
has been written by previous store. This is because when we turn the
sequence into single CPP operation, we are reading all contents at once
into NFP transfer registers, then write them out as a whole. This is not
identical with what the original load/store sequence is doing.

Detecting cross memory access for two random pointers will be difficult,
fortunately under XDP/eBPF's restrictied runtime environment, the copy
normally happen among map, packet data and stack, they do not overlap with
each other.

And for cases supported by NFP, cross memory access will only happen on
PTR_TO_PACKET. Fortunately for this, there is ID information that we could
do accurate memory alias check.
Signed-off-by: Jiong Wang <jiong.wang@netronome.com>
Reviewed-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>

When the gathered copy length is bigger than 32-bytes and within 128-bytes
(the maximum length a single CPP Pull/Push request can finish), the
strategy of read/write are changeed into:

  * Read.
      - use direct reference mode when length is within 32-bytes.
      - use indirect mode when length is bigger than 32-bytes.

  * Write.
      - length <= 8-bytes
        use write8 (direct_ref).
      - length <= 32-byte and 4-bytes aligned
        use write32 (direct_ref).
      - length <= 32-bytes but not 4-bytes aligned
        use write8 (indirect_ref).
      - length > 32-bytes and 4-bytes aligned
        use write32 (indirect_ref).
      - length > 32-bytes and not 4-bytes aligned and <= 40-bytes
        use write32 (direct_ref) to finish the first 32-bytes.
        use write8 (direct_ref) to finish all remaining hanging part.
      - length > 32-bytes and not 4-bytes aligned
        use write32 (indirect_ref) to finish those 4-byte aligned parts.
        use write8 (direct_ref) to finish all remaining hanging part.
Signed-off-by: Jiong Wang <jiong.wang@netronome.com>
Reviewed-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>

For NFP, we want to re-group a sequence of load/store pairs lowered from
memcpy/memmove into single memory bulk operation which then could be
accelerated using NFP CPP bus.

This patch extends the existing load/store auxiliary information by adding
two new fields:

	struct bpf_insn *paired_st;
	s16 ldst_gather_len;

Both fields are supposed to be carried by the the load instruction at the
head of the sequence. "paired_st" is the corresponding store instruction at
the head and "ldst_gather_len" is the gathered length.

If "ldst_gather_len" is negative, then the sequence is doing memory
load/store in descending order, otherwise it is in ascending order. We need
this information to detect overlapped memory access.

This patch then optimize memory bulk copy when the copy length is within
32-bytes.

The strategy of read/write used is:

  * Read.
    Use read32 (direct_ref), always.

  * Write.
    - length <= 8-bytes
      write8 (direct_ref).
    - length <= 32-bytes and is 4-byte aligned
      write32 (direct_ref).
    - length <= 32-bytes but is not 4-byte aligned
      write8 (indirect_ref).

NOTE: the optimization should not change program semantics. The destination
register of the last load instruction should contain the same value before
and after this optimization.
Signed-off-by: Jiong Wang <jiong.wang@netronome.com>
Reviewed-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>

It is usual that we need to check if one BPF insn is for loading/storeing
data from/to memory.

Therefore, it makes sense to factor out related code to become common
helper functions.
Signed-off-by: Jiong Wang <jiong.wang@netronome.com>
Reviewed-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>

Add support for emitting commands with field overwrites.
Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: Jiong Wang <jiong.wang@netronome.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>

When immed is used with No-Dest, the emitter should use reg.dst instead of
reg.areg for the destination, using the latter will actually encode
register zero.
Signed-off-by: Jiong Wang <jiong.wang@netronome.com>
Reviewed-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>

The NFP normally requires the source operands to be difference addressing
modes, but we should rule out the very special NN_REG_NONE type.

There are instruction that ignores both A/B operands, for example:

  local_csr_rd

For these instructions, we might pass the same operand type, NN_REG_NONE,
for both A/B operands.

NOTE: in current NFP ISA, it is only possible for instructions with
unrestricted operands to take none operands, but in case there is new and
similar instructoin in restricted form, they would follow similar rules,
so swreg_to_restricted is updated as well.
Signed-off-by: Jiong Wang <jiong.wang@netronome.com>
Reviewed-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>

If any of the shift insns in the ld/shift sequence is jump destination,
don't do combination.
Signed-off-by: Jiong Wang <jiong.wang@netronome.com>
Reviewed-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>

If the mask insn in the ld/mask pair is jump destination, then don't do
combination.
Signed-off-by: Jiong Wang <jiong.wang@netronome.com>
Reviewed-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>

NFP eBPF offload JIT engine is doing some instruction combine based
optimizations which however must not be safe if the combined sequences
are across basic block boarders.

Currently, there are post checks during fixing jump destinations. If the
jump destination is found to be eBPF insn that has been combined into
another one, then JIT engine will raise error and abort.

This is not optimal. The JIT engine ought to disable the optimization on
such cross-bb-border sequences instead of abort.

As there is no control flow information in eBPF infrastructure that we
can't do basic block based optimizations, this patch extends the existing
jump destination record pass to also flag the jump destination, then in
instruction combine passes we could skip the optimizations if insns in the
sequence are jump targets.
Suggested-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: Jiong Wang <jiong.wang@netronome.com>
Reviewed-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>

eBPF insns are internally organized as dual-list inside NFP offload JIT.
Random access to an insn needs to be done by either forward or backward
traversal along the list.

One place we need to do such traversal is at nfp_fixup_branches where one
traversal is needed for each jump insn to find the destination. Such
traversals could be avoided if jump destinations are collected through a
single travesal in a pre-scan pass, and such information could also be
useful in other places where jump destination info are needed.

This patch adds such jump destination collection in nfp_prog_prepare.
Suggested-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: Jiong Wang <jiong.wang@netronome.com>
Reviewed-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>

This patch adds support for backward jump on NFP.

  - restrictions on backward jump in various functions have been removed.
  - nfp_fixup_branches now supports backward jump.

There is one thing to note, currently an input eBPF JMP insn may generate
several NFP insns, for example,

  NFP imm move insn A \
  NFP compare insn  B  --> 3 NFP insn jited from eBPF JMP insn M
  NFP branch insn   C /
  ---
  NFP insn X           --> 1 NFP insn jited from eBPF insn N
  ---
  ...

therefore, we are doing sanity check to make sure the last jited insn from
an eBPF JMP is a NFP branch instruction.

Once backward jump is allowed, it is possible an eBPF JMP insn is at the
end of the program. This is however causing trouble for the sanity check.
Because the sanity check requires the end index of the NFP insns jited from
one eBPF insn while only the start index is recorded before this patch that
we can only get the end index by:

  start_index_of_the_next_eBPF_insn - 1

or for the above example:

  start_index_of_eBPF_insn_N (which is the index of NFP insn X) - 1

nfp_fixup_branches was using nfp_for_each_insn_walk2 to expose *next* insn
to each iteration during the traversal so the last index could be
calculated from which. Now, it needs some extra code to handle the last
insn. Meanwhile, the use of walk2 is actually unnecessary, we could simply
use generic single instruction walk to do this, the next insn could be
easily calculated using list_next_entry.

So, this patch migrates the jump fixup traversal method to
*list_for_each_entry*, this simplifies the code logic a little bit.

The other thing to note is a new state variable "last_bpf_off" is
introduced to track the index of the last jited NFP insn. This is necessary
because NFP is generating special purposes epilogue sequences, so the index
of the last jited NFP insn is *not* always nfp_prog->prog_len - 1.
Suggested-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: Jiong Wang <jiong.wang@netronome.com>
Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>

Since commit 3a025e1d ("Add optional check for bad kernel-doc
comments") when built with W=1 build will complain about kdoc errors.
Fix the kdoc issues we have.  kdoc is still confused by defines in
nfp_net_ctrl.h but those are not really errors.
Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>

Alexei Starovoitov says:

====================
Small set of verifier improvements and cleanups which is
necessary for bigger patch set of bpf-to-bpf calls coming later.
See individual patches for details.
Tested on x86 and arm64 hw.
====================
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>

since verifier started to print liveness state of the registers
adjust expected output of test_align.
Now this test checks for both proper alignment handling by verifier
and correctness of liveness marks.
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Acked-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>

don't pass large struct bpf_reg_state by value.
Instead pass it by pointer.
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Acked-by: John Fastabend <john.fastabend@gmail.com>
Acked-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>

verifier knows how to trim paths that are known not to be
taken at run-time when register containing run-time constant
is compared with another constant.
It was done only for JEQ comparison.
Extend it to include JNE as well.
More cases can be added in the future.

                     before  after
bpf_lb-DLB_L3.o       2270    2051
bpf_lb-DLB_L4.o       3682    3287
bpf_lb-DUNKNOWN.o     1110    1080
bpf_lxc-DDROP_ALL.o   27876   24980
bpf_lxc-DUNKNOWN.o    38780   34308
bpf_netdev.o          16937   15404
bpf_overlay.o         7929    7191
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Acked-by: John Fastabend <john.fastabend@gmail.com>
Acked-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>

registers with pointers filled from stack were missing live_written marks
which caused liveness propagation to unnecessary mark more registers as
live_read and miss state pruning opportunities later on.

                     before  after
bpf_lb-DLB_L3.o       2285   2270
bpf_lb-DLB_L4.o       3723   3682
bpf_lb-DUNKNOWN.o     1110   1110
bpf_lxc-DDROP_ALL.o   27954  27876
bpf_lxc-DUNKNOWN.o    38954  38780
bpf_netdev.o          16943  16937
bpf_overlay.o         7929   7929
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Acked-by: John Fastabend <john.fastabend@gmail.com>
Acked-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>

when verifier hits an internal bug don't mark register R10==FP as uninit,
since it's read only register and it's not technically correct to let
verifier run further, since it may assume that R10 has valid auxiliary state.

While developing subsequent patches this issue was discovered,
though the code eventually changed that aux reg state doesn't have
pointers any more it is still safer to avoid clearing readonly register.
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Acked-by: John Fastabend <john.fastabend@gmail.com>
Acked-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>

let verifier print register and stack liveness information
into verifier log
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Acked-by: John Fastabend <john.fastabend@gmail.com>
Acked-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>

fix incorrect stack state prints in print_verifier_state()

Fixes: 638f5b90 ("bpf: reduce verifier memory consumption")
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Acked-by: John Fastabend <john.fastabend@gmail.com>
Acked-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>

Attach flag 1 == BPF_F_ALLOW_OVERRIDE; attach flag 2 == BPF_F_ALLOW_MULTI.
Update the calls to bpf_prog_attach() in test_cgrp2_attach2.c to use the
names over the magic numbers.

Fixes: 39323e78 ("samples/bpf: add multi-prog cgroup test case")
Signed-off-by: David Ahern <dsahern@gmail.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>

Utilize the much more capable b53_get_tag_protocol() which takes care of
all Broadcom switches specifics to resolve which port can have Broadcom
tags enabled or not.
Signed-off-by: Florian Fainelli <f.fainelli@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

Since commit e32ea7e7 ("soreuseport: fast reuseport UDP socket
selection") and commit c125e80b ("soreuseport: fast reuseport
TCP socket selection") the relevant reuseport socket matching the current
packet is selected by the reuseport_select_sock() call. The only
exceptions are invalid BPF filters/filters returning out-of-range
indices.
In the latter case the code implicitly falls back to using the hash
demultiplexing, but instead of selecting the socket inside the
reuseport_select_sock() function, it relies on the hash selection
logic introduced with the early soreuseport implementation.

With this patch, in case of a BPF filter returning a bad socket
index value, we fall back to hash-based selection inside the
reuseport_select_sock() body, so that we can drop some duplicate
code in the ipv4 and ipv6 stack.

This also allows faster lookup in the above scenario and will allow
us to avoid computing the hash value for successful, BPF based
demultiplexing - in a later patch.
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Acked-by: Craig Gallek <kraig@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

This is not supported anymore, devices needing a MAC address
just assign one at random, it's just a driver pecularity.
Signed-off-by: Linus Walleij <linus.walleij@linaro.org>
Reviewed-by: Andrew Lunn <andrew@lunn.ch>
Signed-off-by: David S. Miller <davem@davemloft.net>

David Miller says:

====================
net: Significantly shrink the size of routes.

Through a combination of several things, our route structures are
larger than they need to be.

Mostly this stems from having members in dst_entry which are only used
by one class of routes.  So the majority of the work in this series is
about "un-commoning" these members and pushing them into the type
specific structures.

Unfortunately, IPSEC needed the most surgery.  The majority of the
changes here had to do with bundle creation and management.

The other issue is the refcount alignment in dst_entry.  Once we get
rid of the not-so-common members, it really opens the door to removing
that alignment entirely.

I think the new layout looks really nice, so I'll reproduce it here:

	struct net_device       *dev;
	struct  dst_ops	        *ops;
	unsigned long		_metrics;
	unsigned long           expires;
	struct xfrm_state	*xfrm;
	int			(*input)(struct sk_buff *);
	int			(*output)(struct net *net, struct sock *sk, struct sk_buff *skb);
	unsigned short		flags;
	short			obsolete;
	unsigned short		header_len;
	unsigned short		trailer_len;
	atomic_t		__refcnt;
	int			__use;
	unsigned long		lastuse;
	struct lwtunnel_state   *lwtstate;
	struct rcu_head		rcu_head;
	short			error;
	short			__pad;
	__u32			tclassid;

(This is for 64-bit, on 32-bit the __refcnt comes at the very end)

So, the good news:

1) struct dst_entry shrinks from 160 to 112 bytes.

2) struct rtable shrinks from 216 to 168 bytes.

3) struct rt6_info shrinks from 384 to 320 bytes.

Enjoy.

v2:
	Collapse some patches logically based upon feedback.
	Fix the strange patch #7.

v3:	xfrm_dst_path() needs inline keyword
	Properly align __refcnt on 32-bit.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>

There are no more users.
Signed-off-by: David S. Miller <davem@davemloft.net>
Reviewed-by: Eric Dumazet <edumazet@google.com>

While building ipsec bundles, blocks of xfrm dsts are linked together
using dst->next from bottom to the top.

The only thing this is used for is initializing the pmtu values of the
xfrm stack, and for updating the mtu values at xfrm_bundle_ok() time.

The bundle pmtu entries must be processed in this order so that pmtu
values lower in the stack of routes can propagate up to the higher
ones.

Avoid using dst->next by simply maintaining an array of dst pointers
as we already do for the xfrm_state objects when building the bundle.
Signed-off-by: David S. Miller <davem@davemloft.net>
Reviewed-by: Eric Dumazet <edumazet@google.com>

We have padding to try and align the refcount on a separate cache
line.  But after several simplifications the padding has increased
substantially.

So now it's easy to change the layout to get rid of the padding
entirely.

We group the write-heavy __refcnt and __use with less often used
items such as the rcu_head and the error code.
Signed-off-by: David S. Miller <davem@davemloft.net>
Reviewed-by: Eric Dumazet <edumazet@google.com>

The first member of an IPSEC route bundle chain sets it's dst->path to
the underlying ipv4/ipv6 route that carries the bundle.

Stated another way, if one were to follow the xfrm_dst->child chain of
the bundle, the final non-NULL pointer would be the path and point to
either an ipv4 or an ipv6 route.

This is largely used to make sure that PMTU events propagate down to
the correct ipv4 or ipv6 route.

When we don't have the top of an IPSEC bundle 'dst->path == dst'.

Move it down into xfrm_dst and key off of dst->xfrm.
Signed-off-by: David S. Miller <davem@davemloft.net>
Reviewed-by: Eric Dumazet <edumazet@google.com>

The dst->from value is only used by ipv6 routes to track where
a route "came from".

Any time we clone or copy a core ipv6 route in the ipv6 routing
tables, we have the copy/clone's ->from point to the base route.

This is used to handle route expiration properly.

Only ipv6 uses this mechanism, and only ipv6 code references
it.  So it is safe to move it into rt6_info.
Signed-off-by: David S. Miller <davem@davemloft.net>
Reviewed-by: Eric Dumazet <edumazet@google.com>

XFRM bundle child chains look like this:

	xdst1 --> xdst2 --> xdst3 --> path_dst

All of xdstN are xfrm_dst objects and xdst->u.dst.xfrm is non-NULL.
The final child pointer in the chain, here called 'path_dst', is some
other kind of route such as an ipv4 or ipv6 one.

The xfrm output path pops routes, one at a time, via the child
pointer, until we hit one which has a dst->xfrm pointer which
is NULL.

We can easily preserve the above mechanisms with child sitting
only in the xfrm_dst structure.  All children in the chain
before we break out of the xfrm_output() loop have dst->xfrm
non-NULL and are therefore xfrm_dst objects.

Since we break out of the loop when we find dst->xfrm NULL, we
will not try to dereference 'dst' as if it were an xfrm_dst.
Signed-off-by: David S. Miller <davem@davemloft.net>

This will make a future change moving the dst->child pointer less
invasive.
Signed-off-by: David S. Miller <davem@davemloft.net>
Reviewed-by: Eric Dumazet <edumazet@google.com>

Only IPSEC routes have a non-NULL dst->child pointer.  And IPSEC
routes are identified by a non-NULL dst->xfrm pointer.
Signed-off-by: David S. Miller <davem@davemloft.net>

Signed-off-by: David S. Miller <davem@davemloft.net>
Reviewed-by: Eric Dumazet <edumazet@google.com>