linux.git/Documentation/sysctl/net.txt, branch v4.19

bpf, x86_32: add eBPF JIT compiler for ia32

2018-05-03T16:15:25+00:00

The JIT compiler emits ia32 bit instructions. Currently, It supports eBPF
only. Classic BPF is supported because of the conversion by BPF core.

Almost all instructions from eBPF ISA supported except the following:
BPF_ALU64 | BPF_DIV | BPF_K
BPF_ALU64 | BPF_DIV | BPF_X
BPF_ALU64 | BPF_MOD | BPF_K
BPF_ALU64 | BPF_MOD | BPF_X
BPF_STX | BPF_XADD | BPF_W
BPF_STX | BPF_XADD | BPF_DW

It doesn't support BPF_JMP|BPF_CALL with BPF_PSEUDO_CALL at the moment.

IA32 has few general purpose registers, EAX|EDX|ECX|EBX|ESI|EDI. I use
EAX|EDX|ECX|EBX as temporary registers to simulate instructions in eBPF
ISA, and allocate ESI|EDI to BPF_REG_AX for constant blinding, all others
eBPF registers, R0-R10, are simulated through scratch space on stack.

The reasons behind the hardware registers allocation policy are:
1:MUL need EAX:EDX, shift operation need ECX, so they aren't fit
  for general eBPF 64bit register simulation.
2:We need at least 4 registers to simulate most eBPF ISA operations
  on registers operands instead of on register&memory operands.
3:We need to put BPF_REG_AX on hardware registers, or constant blinding
  will degrade jit performance heavily.

Tested on PC (Intel(R) Core(TM) i5-5200U CPU).
Testing results on i5-5200U:
1) test_bpf: Summary: 349 PASSED, 0 FAILED, [319/341 JIT'ed]
2) test_progs: Summary: 83 PASSED, 0 FAILED.
3) test_lpm: OK
4) test_lru_map: OK
5) test_verifier: Summary: 828 PASSED, 0 FAILED.

Above tests are all done in following two conditions separately:
1:bpf_jit_enable=1 and bpf_jit_harden=0
2:bpf_jit_enable=1 and bpf_jit_harden=2

Below are some numbers for this jit implementation:
Note:
  I run test_progs in kselftest 100 times continuously for every condition,
  the numbers are in format: total/times=avg.
  The numbers that test_bpf reports show almost the same relation.

a:jit_enable=0 and jit_harden=0            b:jit_enable=1 and jit_harden=0
  test_pkt_access:PASS:ipv4:15622/100=156    test_pkt_access:PASS:ipv4:10674/100=106
  test_pkt_access:PASS:ipv6:9130/100=91      test_pkt_access:PASS:ipv6:4855/100=48
  test_xdp:PASS:ipv4:240198/100=2401         test_xdp:PASS:ipv4:138912/100=1389
  test_xdp:PASS:ipv6:137326/100=1373         test_xdp:PASS:ipv6:68542/100=685
  test_l4lb:PASS:ipv4:61100/100=611          test_l4lb:PASS:ipv4:37302/100=373
  test_l4lb:PASS:ipv6:101000/100=1010        test_l4lb:PASS:ipv6:55030/100=550

c:jit_enable=1 and jit_harden=2
  test_pkt_access:PASS:ipv4:10558/100=105
  test_pkt_access:PASS:ipv6:5092/100=50
  test_xdp:PASS:ipv4:131902/100=1319
  test_xdp:PASS:ipv6:77932/100=779
  test_l4lb:PASS:ipv4:38924/100=389
  test_l4lb:PASS:ipv6:57520/100=575

The numbers show we get 30%~50% improvement.

See Documentation/networking/filter.txt for more information.

Changelog:

 Changes v5-v6:
 1:Add do {} while (0) to RETPOLINE_RAX_BPF_JIT for
   consistence reason.
 2:Clean up non-standard comments, reported by Daniel Borkmann.
 3:Fix a memory leak issue, repoted by Daniel Borkmann.

 Changes v4-v5:
 1:Delete is_on_stack, BPF_REG_AX is the only one
   on real hardware registers, so just check with
   it.
 2:Apply commit 1612a981b766 ("bpf, x64: fix JIT emission
   for dead code"), suggested by Daniel Borkmann.

 Changes v3-v4:
 1:Fix changelog in commit.
   I install llvm-6.0, then test_progs willn't report errors.
   I submit another patch:
   "bpf: fix misaligned access for BPF_PROG_TYPE_PERF_EVENT program type on x86_32 platform"
   to fix another problem, after that patch, test_verifier willn't report errors too.
 2:Fix clear r0[1] twice unnecessarily in *BPF_IND|BPF_ABS* simulation.

 Changes v2-v3:
 1:Move BPF_REG_AX to real hardware registers for performance reason.
 3:Using bpf_load_pointer instead of bpf_jit32.S, suggested by Daniel Borkmann.
 4:Delete partial codes in 1c2a088a6626, suggested by Daniel Borkmann.
 5:Some bug fixes and comments improvement.

 Changes v1-v2:
 1:Fix bug in emit_ia32_neg64.
 2:Fix bug in emit_ia32_arsh_r64.
 3:Delete filename in top level comment, suggested by Thomas Gleixner.
 4:Delete unnecessary boiler plate text, suggested by Thomas Gleixner.
 5:Rewrite some words in changelog.
 6:CodingSytle improvement and a little more comments.

Signed-off-by: Wang YanQing 
Signed-off-by: Daniel Borkmann

net: do not create fallback tunnels for non-default namespaces

2018-03-09T16:23:11+00:00

fallback tunnels (like tunl0, gre0, gretap0, erspan0, sit0,
ip6tnl0, ip6gre0) are automatically created when the corresponding
module is loaded.

These tunnels are also automatically created when a new network
namespace is created, at a great cost.

In many cases, netns are used for isolation purposes, and these
extra network devices are a waste of resources. We are using
thousands of netns per host, and hit the netns creation/delete
bottleneck a lot. (Many thanks to Kirill for recent work on this)

Add a new sysctl so that we can opt-out from this automatic creation.

Note that these tunnels are still created for the initial namespace,
to be the least intrusive for typical setups.

Tested:
lpk43:~# cat add_del_unshare.sh
for i in `seq 1 40`
do
 (for j in `seq 1 100` ; do  unshare -n /bin/true >/dev/null ; done) &
done
wait

lpk43:~# echo 0 >/proc/sys/net/core/fb_tunnels_only_for_init_net
lpk43:~# time ./add_del_unshare.sh

real	0m37.521s
user	0m0.886s
sys	7m7.084s
lpk43:~# echo 1 >/proc/sys/net/core/fb_tunnels_only_for_init_net
lpk43:~# time ./add_del_unshare.sh

real	0m4.761s
user	0m0.851s
sys	1m8.343s
lpk43:~#

Signed-off-by: Eric Dumazet 
Signed-off-by: David S. Miller

net: Clarify dev_weight documentation for LRO and GRO_HW.

2017-12-20T18:04:01+00:00

Reported-by: Marcelo Ricardo Leitner 
Signed-off-by: Michael Chan 
Signed-off-by: David S. Miller

bpf, doc: Add arm32 as arch supporting eBPF JIT

2017-08-24T05:40:12+00:00

As eBPF JIT support for arm32 was added recently with
commit 39c13c204bb1150d401e27d41a9d8b332be47c49, it seems appropriate to
add arm32 as arch with support for eBPF JIT in bpf and sysctl docs as well.

Signed-off-by: Shubham Bansal 
Acked-by: Alexei Starovoitov 
Acked-by: Daniel Borkmann 
Signed-off-by: David S. Miller

bpf, doc: also add s390x as arch to sysctl description

2017-08-21T02:45:09+00:00

Looks like this was accidentally missed, so still add s390x
as supported eBPF JIT arch to bpf_jit_enable.

Fixes: 014cd0a368dc ("bpf: Update sysctl documentation to list all supported architectures")
Signed-off-by: Daniel Borkmann 
Acked-by: Alexei Starovoitov 
Signed-off-by: David S. Miller

bpf, doc: improve sysctl knob description

2017-08-18T18:00:41+00:00

Current context speaking of tcpdump filters is out of date these
days, so lets improve the sysctl description for the BPF knobs
a bit.

Signed-off-by: Daniel Borkmann 
Signed-off-by: David S. Miller

bpf: Update sysctl documentation to list all supported architectures

2017-08-17T17:09:28+00:00

The sysctl documentation states that the JIT is only available on
x86_64, which is no longer correct.

Update the list, and break it out to indicate which architectures
support the cBPF JIT (via HAVE_CBPF_JIT) or the eBPF JIT
(HAVE_EBPF_JIT).

Signed-off-by: Michael Ellerman 
Acked-by: Daniel Borkmann 
Signed-off-by: David S. Miller

Replace 2 jiffies with sysctl netdev_budget_usecs to enable softirq tuning

2017-04-21T17:22:34+00:00

Constants used for tuning are generally a bad idea, especially as hardware
changes over time. Replace the constant 2 jiffies with sysctl variable
netdev_budget_usecs to enable sysadmins to tune the softirq processing.
Also document the variable.

For example, a very fast machine might tune this to 1000 microseconds,
while my regression testing 486DX-25 needs it to be 4000 microseconds on
a nearly idle network to prevent time_squeeze from being incremented.

Version 2: changed jiffies to microseconds for predictable units.

Signed-off-by: Matthew Whitehead 
Signed-off-by: David S. Miller

bpf: make jited programs visible in traces

2017-02-17T18:40:05+00:00

Long standing issue with JITed programs is that stack traces from
function tracing check whether a given address is kernel code
through {__,}kernel_text_address(), which checks for code in core
kernel, modules and dynamically allocated ftrace trampolines. But
what is still missing is BPF JITed programs (interpreted programs
are not an issue as __bpf_prog_run() will be attributed to them),
thus when a stack trace is triggered, the code walking the stack
won't see any of the JITed ones. The same for address correlation
done from user space via reading /proc/kallsyms. This is read by
tools like perf, but the latter is also useful for permanent live
tracing with eBPF itself in combination with stack maps when other
eBPF types are part of the callchain. See offwaketime example on
dumping stack from a map.

This work tries to tackle that issue by making the addresses and
symbols known to the kernel. The lookup from *kernel_text_address()
is implemented through a latched RB tree that can be read under
RCU in fast-path that is also shared for symbol/size/offset lookup
for a specific given address in kallsyms. The slow-path iteration
through all symbols in the seq file done via RCU list, which holds
a tiny fraction of all exported ksyms, usually below 0.1 percent.
Function symbols are exported as bpf_prog_, in order to aide
debugging and attribution. This facility is currently enabled for
root-only when bpf_jit_kallsyms is set to 1, and disabled if hardening
is active in any mode. The rationale behind this is that still a lot
of systems ship with world read permissions on kallsyms thus addresses
should not get suddenly exposed for them. If that situation gets
much better in future, we always have the option to change the
default on this. Likewise, unprivileged programs are not allowed
to add entries there either, but that is less of a concern as most
such programs types relevant in this context are for root-only anyway.
If enabled, call graphs and stack traces will then show a correct
attribution; one example is illustrated below, where the trace is
now visible in tooling such as perf script --kallsyms=/proc/kallsyms
and friends.

Before:

  7fff8166889d bpf_clone_redirect+0x80007f0020ed (/lib/modules/4.9.0-rc8+/build/vmlinux)
         f5d80 __sendmsg_nocancel+0xffff006451f1a007 (/usr/lib64/libc-2.18.so)

After:

  7fff816688b7 bpf_clone_redirect+0x80007f002107 (/lib/modules/4.9.0-rc8+/build/vmlinux)
  7fffa0575728 bpf_prog_33c45a467c9e061a+0x8000600020fb (/lib/modules/4.9.0-rc8+/build/vmlinux)
  7fffa07ef1fc cls_bpf_classify+0x8000600020dc (/lib/modules/4.9.0-rc8+/build/vmlinux)
  7fff81678b68 tc_classify+0x80007f002078 (/lib/modules/4.9.0-rc8+/build/vmlinux)
  7fff8164d40b __netif_receive_skb_core+0x80007f0025fb (/lib/modules/4.9.0-rc8+/build/vmlinux)
  7fff8164d718 __netif_receive_skb+0x80007f002018 (/lib/modules/4.9.0-rc8+/build/vmlinux)
  7fff8164e565 process_backlog+0x80007f002095 (/lib/modules/4.9.0-rc8+/build/vmlinux)
  7fff8164dc71 net_rx_action+0x80007f002231 (/lib/modules/4.9.0-rc8+/build/vmlinux)
  7fff81767461 __softirqentry_text_start+0x80007f0020d1 (/lib/modules/4.9.0-rc8+/build/vmlinux)
  7fff817658ac do_softirq_own_stack+0x80007f00201c (/lib/modules/4.9.0-rc8+/build/vmlinux)
  7fff810a2c20 do_softirq+0x80007f002050 (/lib/modules/4.9.0-rc8+/build/vmlinux)
  7fff810a2cb5 __local_bh_enable_ip+0x80007f002085 (/lib/modules/4.9.0-rc8+/build/vmlinux)
  7fff8168d452 ip_finish_output2+0x80007f002152 (/lib/modules/4.9.0-rc8+/build/vmlinux)
  7fff8168ea3d ip_finish_output+0x80007f00217d (/lib/modules/4.9.0-rc8+/build/vmlinux)
  7fff8168f2af ip_output+0x80007f00203f (/lib/modules/4.9.0-rc8+/build/vmlinux)
  [...]
  7fff81005854 do_syscall_64+0x80007f002054 (/lib/modules/4.9.0-rc8+/build/vmlinux)
  7fff817649eb return_from_SYSCALL_64+0x80007f002000 (/lib/modules/4.9.0-rc8+/build/vmlinux)
         f5d80 __sendmsg_nocancel+0xffff01c484812007 (/usr/lib64/libc-2.18.so)

Signed-off-by: Daniel Borkmann 
Acked-by: Alexei Starovoitov 
Cc: linux-kernel@vger.kernel.org
Signed-off-by: David S. Miller

net: dev_weight: TX/RX orthogonality

2016-12-29T20:38:35+00:00

Oftenly, introducing side effects on packet processing on the other half
of the stack by adjusting one of TX/RX via sysctl is not desirable.
There are cases of demand for asymmetric, orthogonal configurability.

This holds true especially for nodes where RPS for RFS usage on top is
configured and therefore use the 'old dev_weight'. This is quite a
common base configuration setup nowadays, even with NICs of superior processing
support (e.g. aRFS).

A good example use case are nodes acting as noSQL data bases with a
large number of tiny requests and rather fewer but large packets as responses.
It's affordable to have large budget and rx dev_weights for the
requests. But as a side effect having this large a number on TX
processed in one run can overwhelm drivers.

This patch therefore introduces an independent configurability via sysctl to
userland.

Signed-off-by: Matthias Tafelmeier 
Signed-off-by: David S. Miller