| Age | Commit message (Collapse) | Author | Files | Lines |
|---|
|
One down side of using TCP rx zerocopy is one extra TLB miss
per page after the mapping operation.
While if the application is using hugepages, the non zerocopy
recvmsg() will not have to pay these TLB costs.
This patch allows server side to use huge pages for
the non zero copy case, to allow fair comparisons when
both solutions use optimal conditions.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Arjun Roy <arjunroy@google.com>
Cc: Soheil Hassas Yeganeh <soheil@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
There are significant gains using huge pages when
available, as shown in [1].
This patch adds mmap_large_buffer() and uses it
in client side (tx path of this reference tool)
Following patch will use the feature for server side.
[1] https://patchwork.ozlabs.org/project/netdev/patch/20200820154359.1806305-1-edumazet@google.com/
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Arjun Roy <arjunroy@google.com>
Cc: Soheil Hassas Yeganeh <soheil@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
When TCP_ZEROCOPY_RECEIVE operation has been added,
I made the mistake of automatically un-mapping prior
content before mapping new pages.
This has the unfortunate effect of adding potentially long
MMU operations (like TLB flushes) while socket lock is held.
Using madvise(MADV_DONTNEED) right after pages has been used
has two benefits :
1) This releases pages sooner, allowing pages to be recycled
if they were part of a page pool in a NIC driver.
2) No more long unmap operations while preventing immediate
processing of incoming packets.
The cost of the added system call is small enough.
Arjun will submit a kernel patch allowing to opt out from
the unmap attempt in tcp_zerocopy_receive()
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Arjun Roy <arjunroy@google.com>
Cc: Soheil Hassas Yeganeh <soheil@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
When size_t maps to unsigned int (e.g. on 32-bit powerpc), then the
comparison with 1<<35 is always true. Clang 9 threw:
warning: result of comparison of constant 34359738368 with \
expression of type 'size_t' (aka 'unsigned int') is always true \
[-Wtautological-constant-out-of-range-compare]
while (total < FILE_SZ) {
Tested: make -C tools/testing/selftests TARGETS="net" run_tests
Fixes: 192dc405f308 ("selftests: net: add tcp_mmap program")
Signed-off-by: Tanner Love <tannerlove@google.com>
Acked-by: Willem de Bruijn <willemb@google.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
Since chunk_size is no longer an integer, we can not
use it directly as an argument of setsockopt().
This patch should fix tcp_mmap for Big Endian kernels.
Fixes: 597b01edafac ("selftests: net: avoid ptl lock contention in tcp_mmap")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Soheil Hassas Yeganeh <soheil@google.com>
Cc: Arjun Roy <arjunroy@google.com>
Acked-by: Soheil Hassas Yeganeh <soheil@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
We added fields in tcp_zerocopy_receive structure,
so make sure to clear all fields to not pass garbage to the kernel.
We were lucky because recent additions added 'out' parameters,
still we need to clean our reference implementation, before folks
copy/paste it.
Fixes: c8856c051454 ("tcp-zerocopy: Return inq along with tcp receive zerocopy.")
Fixes: 33946518d493 ("tcp-zerocopy: Return sk_err (if set) along with tcp receive zerocopy.")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Arjun Roy <arjunroy@google.com>
Cc: Soheil Hassas Yeganeh <soheil@google.com>
Acked-by: Soheil Hassas Yeganeh <soheil@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
Pull networking updates from David Miller:
"Another merge window, another pull full of stuff:
1) Support alternative names for network devices, from Jiri Pirko.
2) Introduce per-netns netdev notifiers, also from Jiri Pirko.
3) Support MSG_PEEK in vsock/virtio, from Matias Ezequiel Vara
Larsen.
4) Allow compiling out the TLS TOE code, from Jakub Kicinski.
5) Add several new tracepoints to the kTLS code, also from Jakub.
6) Support set channels ethtool callback in ena driver, from Sameeh
Jubran.
7) New SCTP events SCTP_ADDR_ADDED, SCTP_ADDR_REMOVED,
SCTP_ADDR_MADE_PRIM, and SCTP_SEND_FAILED_EVENT. From Xin Long.
8) Add XDP support to mvneta driver, from Lorenzo Bianconi.
9) Lots of netfilter hw offload fixes, cleanups and enhancements,
from Pablo Neira Ayuso.
10) PTP support for aquantia chips, from Egor Pomozov.
11) Add UDP segmentation offload support to igb, ixgbe, and i40e. From
Josh Hunt.
12) Add smart nagle to tipc, from Jon Maloy.
13) Support L2 field rewrite by TC offloads in bnxt_en, from Venkat
Duvvuru.
14) Add a flow mask cache to OVS, from Tonghao Zhang.
15) Add XDP support to ice driver, from Maciej Fijalkowski.
16) Add AF_XDP support to ice driver, from Krzysztof Kazimierczak.
17) Support UDP GSO offload in atlantic driver, from Igor Russkikh.
18) Support it in stmmac driver too, from Jose Abreu.
19) Support TIPC encryption and auth, from Tuong Lien.
20) Introduce BPF trampolines, from Alexei Starovoitov.
21) Make page_pool API more numa friendly, from Saeed Mahameed.
22) Introduce route hints to ipv4 and ipv6, from Paolo Abeni.
23) Add UDP segmentation offload to cxgb4, Rahul Lakkireddy"
* git://git.kernel.org/pub/scm/linux/kernel/git/netdev/net-next: (1857 commits)
libbpf: Fix usage of u32 in userspace code
mm: Implement no-MMU variant of vmalloc_user_node_flags
slip: Fix use-after-free Read in slip_open
net: dsa: sja1105: fix sja1105_parse_rgmii_delays()
macvlan: schedule bc_work even if error
enetc: add support Credit Based Shaper(CBS) for hardware offload
net: phy: add helpers phy_(un)lock_mdio_bus
mdio_bus: don't use managed reset-controller
ax88179_178a: add ethtool_op_get_ts_info()
mlxsw: spectrum_router: Fix use of uninitialized adjacency index
mlxsw: spectrum_router: After underlay moves, demote conflicting tunnels
bpf: Simplify __bpf_arch_text_poke poke type handling
bpf: Introduce BPF_TRACE_x helper for the tracing tests
bpf: Add bpf_jit_blinding_enabled for !CONFIG_BPF_JIT
bpf, testing: Add various tail call test cases
bpf, x86: Emit patchable direct jump as tail call
bpf: Constant map key tracking for prog array pokes
bpf: Add poke dependency tracking for prog array maps
bpf: Add initial poke descriptor table for jit images
bpf: Move owner type, jited info into array auxiliary data
...
|
|
tcp_mmap is used as a reference program for TCP rx zerocopy,
so it is important to point out some potential issues.
If multiple threads are concurrently using getsockopt(...
TCP_ZEROCOPY_RECEIVE), there is a chance the low-level mm
functions compete on shared ptl lock, if vma are arbitrary placed.
Instead of letting the mm layer place the chunks back to back,
this patch enforces an alignment so that each thread uses
a different ptl lock.
Performance measured on a 100 Gbit NIC, with 8 tcp_mmap clients
launched at the same time :
$ for f in {1..8}; do ./tcp_mmap -H 2002:a05:6608:290:: & done
In the following run, we reproduce the old behavior by requesting no alignment :
$ tcp_mmap -sz -C $((128*1024)) -a 4096
received 32768 MB (100 % mmap'ed) in 9.69532 s, 28.3516 Gbit
cpu usage user:0.08634 sys:3.86258, 120.511 usec per MB, 171839 c-switches
received 32768 MB (100 % mmap'ed) in 25.4719 s, 10.7914 Gbit
cpu usage user:0.055268 sys:21.5633, 659.745 usec per MB, 9065 c-switches
received 32768 MB (100 % mmap'ed) in 28.5419 s, 9.63069 Gbit
cpu usage user:0.057401 sys:23.8761, 730.392 usec per MB, 14987 c-switches
received 32768 MB (100 % mmap'ed) in 28.655 s, 9.59268 Gbit
cpu usage user:0.059689 sys:23.8087, 728.406 usec per MB, 18509 c-switches
received 32768 MB (100 % mmap'ed) in 28.7808 s, 9.55074 Gbit
cpu usage user:0.066042 sys:23.4632, 718.056 usec per MB, 24702 c-switches
received 32768 MB (100 % mmap'ed) in 28.8259 s, 9.5358 Gbit
cpu usage user:0.056547 sys:23.6628, 723.858 usec per MB, 23518 c-switches
received 32768 MB (100 % mmap'ed) in 28.8808 s, 9.51767 Gbit
cpu usage user:0.059357 sys:23.8515, 729.703 usec per MB, 14691 c-switches
received 32768 MB (100 % mmap'ed) in 28.8879 s, 9.51534 Gbit
cpu usage user:0.047115 sys:23.7349, 725.769 usec per MB, 21773 c-switches
New behavior (automatic alignment based on Hugepagesize),
we can see the system overhead being dramatically reduced.
$ tcp_mmap -sz -C $((128*1024))
received 32768 MB (100 % mmap'ed) in 13.5339 s, 20.3103 Gbit
cpu usage user:0.122644 sys:3.4125, 107.884 usec per MB, 168567 c-switches
received 32768 MB (100 % mmap'ed) in 16.0335 s, 17.1439 Gbit
cpu usage user:0.132428 sys:3.55752, 112.608 usec per MB, 188557 c-switches
received 32768 MB (100 % mmap'ed) in 17.5506 s, 15.6621 Gbit
cpu usage user:0.155405 sys:3.24889, 103.891 usec per MB, 226652 c-switches
received 32768 MB (100 % mmap'ed) in 19.1924 s, 14.3222 Gbit
cpu usage user:0.135352 sys:3.35583, 106.542 usec per MB, 207404 c-switches
received 32768 MB (100 % mmap'ed) in 22.3649 s, 12.2906 Gbit
cpu usage user:0.142429 sys:3.53187, 112.131 usec per MB, 250225 c-switches
received 32768 MB (100 % mmap'ed) in 22.5336 s, 12.1986 Gbit
cpu usage user:0.140654 sys:3.61971, 114.757 usec per MB, 253754 c-switches
received 32768 MB (100 % mmap'ed) in 22.5483 s, 12.1906 Gbit
cpu usage user:0.134035 sys:3.55952, 112.718 usec per MB, 252997 c-switches
received 32768 MB (100 % mmap'ed) in 22.6442 s, 12.139 Gbit
cpu usage user:0.126173 sys:3.71251, 117.147 usec per MB, 253728 c-switches
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Soheil Hassas Yeganeh <soheil@google.com>
Cc: Arjun Roy <arjunroy@google.com>
Acked-by: Soheil Hassas Yeganeh <soheil@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
Since we do not plan using pthread_join() in the server do_accept()
loop, we better create detached threads, or risk increasing memory
footprint over time.
Fixes: 192dc405f308 ("selftests: net: add tcp_mmap program")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
Use size_t and ssize_t correctly for counting send file size
instead of unsigned long and long, because long is 32bit on
32bit arch, which is not enough for counting long file size (>4GB).
Signed-off-by: Masami Hiramatsu <mhiramat@kernel.org>
Cc: Eric Dumazet <edumazet@google.com>
Cc: David S. Miller <davem@davemloft.net>
Signed-off-by: Shuah Khan <skhan@linuxfoundation.org>
|
|
Based on 1 normalized pattern(s):
license gplv2 this program is free software you can redistribute it
and or modify it under the terms and conditions of the gnu general
public license version 2 as published by the free software
foundation this program is distributed in the hope it will be useful
but without any warranty without even the implied warranty of
merchantability or fitness for a particular purpose see the gnu
general public license for more details you should have received a
copy of the gnu general public license along with this program if
not write to the free software foundation inc 51 franklin st fifth
floor boston ma 02110 1301 usa
extracted by the scancode license scanner the SPDX license identifier
GPL-2.0-only
has been chosen to replace the boilerplate/reference in 4 file(s).
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Reviewed-by: Alexios Zavras <alexios.zavras@intel.com>
Reviewed-by: Steve Winslow <swinslow@gmail.com>
Reviewed-by: Allison Randal <allison@lohutok.net>
Cc: linux-spdx@vger.kernel.org
Link: https://lkml.kernel.org/r/20190528171440.038486796@linutronix.de
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
|
|
use actual protocol family passed by user rather than hardcoded
AF_INTE6 to cerate sockets.
current code is not working for IPv4.
Signed-off-by: Maninder Singh <maninder1.s@samsung.com>
Signed-off-by: Vaneet Narang <v.narang@samsung.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
After prior kernel change, mmap() on TCP socket only reserves VMA.
We have to use getsockopt(fd, IPPROTO_TCP, TCP_ZEROCOPY_RECEIVE, ...)
to perform the transfert of pages from skbs in TCP receive queue into such VMA.
struct tcp_zerocopy_receive {
__u64 address; /* in: address of mapping */
__u32 length; /* in/out: number of bytes to map/mapped */
__u32 recv_skip_hint; /* out: amount of bytes to skip */
};
After a successful getsockopt(...TCP_ZEROCOPY_RECEIVE...), @length contains
number of bytes that were mapped, and @recv_skip_hint contains number of bytes
that should be read using conventional read()/recv()/recvmsg() system calls,
to skip a sequence of bytes that can not be mapped, because not properly page
aligned.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Andy Lutomirski <luto@kernel.org>
Acked-by: Soheil Hassas Yeganeh <soheil@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
This is a reference program showing how mmap() can be used
on TCP flows to implement receive zero copy.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
Eric Dumazet
Tanner Love
Linus Torvalds
Masami Hiramatsu
Thomas Gleixner
Maninder Singh