Eric Dumazet says:
====================
tcp/dccp: lockless listener
TCP listener refactoring : this is becoming interesting !
This patch series takes the steps to use normal TCP/DCCP ehash
table to store SYN_RECV requests, instead of the private per-listener
hash table we had until now.
SYNACK skb are now attached to their syn_recv request socket,
so that we no longer heavily modify listener sk_wmem_alloc.
listener lock is no longer held in fast path, including
SYNCOOKIE mode.
During my tests, my server was able to process 3,500,000
SYN packets per second on one listener and still had available
cpu cycles.
That is about 2 to 3 order of magnitude what we had with older kernels.
This effort started two years ago and I am pleased to reach expectations.
We'll probably extend SO_REUSEPORT to add proper cpu/numa affinities,
so that heavy duty TCP servers can get proper siloing thanks to multi-queues
NIC.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>
Everything should now be ready to finally allow SYN
packets processing without holding listener lock.
Tested:
3.5 Mpps SYNFLOOD. Plenty of cpu cycles available.
Next bottleneck is the refcount taken on listener,
that could be avoided if we remove SLAB_DESTROY_BY_RCU
strict semantic for listeners, and use regular RCU.
13.18% [kernel] [k] __inet_lookup_listener
9.61% [kernel] [k] tcp_conn_request
8.16% [kernel] [k] sha_transform
5.30% [kernel] [k] inet_reqsk_alloc
4.22% [kernel] [k] sock_put
3.74% [kernel] [k] tcp_make_synack
2.88% [kernel] [k] ipt_do_table
2.56% [kernel] [k] memcpy_erms
2.53% [kernel] [k] sock_wfree
2.40% [kernel] [k] tcp_v4_rcv
2.08% [kernel] [k] fib_table_lookup
1.84% [kernel] [k] tcp_openreq_init_rwin
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
If a listener with thousands of children in accept queue
is dismantled, it can take a while to close all of them.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This control variable was set at first listen(fd, backlog)
call, but not updated if application tried to increase or decrease
backlog. It made sense at the time listener had a non resizeable
hash table.
Also rounding to powers of two was not very friendly.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
It is enough to check listener sk_state, no need for an extra
condition.
max_qlen_log can be moved into struct request_sock_queue
We can remove syn_wait_lock and the alignment it enforced.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
If a listen backlog is very big (to avoid syncookies), then
the listener sk->sk_wmem_alloc is the main source of false
sharing, as we need to touch it twice per SYNACK re-transmit
and TX completion.
(One SYN packet takes listener lock once, but up to 6 SYNACK
are generated)
By attaching the skb to the request socket, we remove this
source of contention.
Tested:
listen(fd, 10485760); // single listener (no SO_REUSEPORT)
16 RX/TX queue NIC
Sustain a SYNFLOOD attack of ~320,000 SYN per second,
Sending ~1,400,000 SYNACK per second.
Perf profiles now show listener spinlock being next bottleneck.
20.29% [kernel] [k] queued_spin_lock_slowpath
10.06% [kernel] [k] __inet_lookup_established
5.12% [kernel] [k] reqsk_timer_handler
3.22% [kernel] [k] get_next_timer_interrupt
3.00% [kernel] [k] tcp_make_synack
2.77% [kernel] [k] ipt_do_table
2.70% [kernel] [k] run_timer_softirq
2.50% [kernel] [k] ip_finish_output
2.04% [kernel] [k] cascade
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
inet6_csk_search_req() and inet6_csk_reqsk_queue_hash_add()
no longer exist.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
We no longer use hash_rnd, nr_table_entries and syn_table[]
For a listener with a backlog of 10 millions sockets, this
saves 80 MBytes of vmalloced memory.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
In this patch, we insert request sockets into TCP/DCCP
regular ehash table (where ESTABLISHED and TIMEWAIT sockets
are) instead of using the per listener hash table.
ACK packets find SYN_RECV pseudo sockets without having
to find and lock the listener.
In nominal conditions, this halves pressure on listener lock.
Note that this will allow for SO_REUSEPORT refinements,
so that we can select a listener using cpu/numa affinities instead
of the prior 'consistent hash', since only SYN packets will
apply this selection logic.
We will shrink listen_sock in the following patch to ease
code review.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Ying Cai <ycai@google.com>
Cc: Willem de Bruijn <willemb@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
When request sockets are no longer in a per listener hash table
but on regular TCP ehash, we need to access listener uid
through req->rsk_listener
get_openreq6() also gets a const for its request socket argument.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Once listener is lockless, its sk_state can change anytime.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
We'll soon have to call tcp_v[46]_inbound_md5_hash() twice.
Also add const attribute to the socket, as it might be the
unlocked listener for SYN packets.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
We plan to use generic functions to insert request sockets
into ehash table.
sk_prot needs to be set (to retrieve sk_prot->h.hashinfo)
sk_node needs to be cleared.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This fixes a typo : We want to store the NAPI id on child socket.
Presumably nobody really uses busy polling, on short lived flows.
Fixes: 3d97379a67 ("tcp: move sk_mark_napi_id() at the right place")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
long term plan is to remove struct listen_sock when its hash
table is no longer there.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
qlen_inc & young_inc were protected by listener lock,
while qlen_dec & young_dec were atomic fields.
Everything needs to be atomic for upcoming lockless listener.
Also move qlen/young in request_sock_queue as we'll get rid
of struct listen_sock eventually.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
struct request_sock_queue fields are currently protected
by the listener 'lock' (not a real spinlock)
We need to add a private spinlock instead, so that softirq handlers
creating children do not have to worry with backlog notion
that the listener 'lock' carries.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Pull MMC fixes from Ulf Hansson:
"Here are some mmc fixes intended for v4.3 rc4:
MMC core:
- Allow users of mmc_of_parse() to succeed when CONFIG_GPIOLIB is
unset
- Prevent infinite loop of re-tuning for CRC-errors for CMD19 and
CMD21
MMC host:
- pxamci: Fix issues with card detect
- sunxi: Fix clk-delay settings"
* tag 'mmc-v4.3-rc3' of git://git.linaro.org/people/ulf.hansson/mmc:
mmc: core: fix dead loop of mmc_retune
mmc: pxamci: fix card detect with slot-gpio API
mmc: sunxi: Fix clk-delay settings
mmc: core: Don't return an error for CD/WP GPIOs when GPIOLIB is unset
Pull IOVA fixes from David Woodhouse:
"The main fix here is the first one, fixing the over-allocation of
size-aligned requests. The other patches simply make the existing
IOVA code available to users other than the Intel VT-d driver, with no
functional change.
I concede the latter really *should* have been submitted during the
merge window, but since it's basically risk-free and people are
waiting to build on top of it and it's my fault I didn't get it in, I
(and they) would be grateful if you'd take it"
* git://git.infradead.org/intel-iommu:
iommu: Make the iova library a module
iommu: iova: Export symbols
iommu: iova: Move iova cache management to the iova library
iommu/iova: Avoid over-allocating when size-aligned
Merge misc fixes from Andrew Morton:
"12 fixes"
* emailed patches from Andrew Morton <akpm@linux-foundation.org>:
dmapool: fix overflow condition in pool_find_page()
thermal: avoid division by zero in power allocator
memcg: remove pcp_counter_lock
kprobes: use _do_fork() in samples to make them work again
drivers/input/joystick/Kconfig: zhenhua.c needs BITREVERSE
memcg: make mem_cgroup_read_stat() unsigned
memcg: fix dirty page migration
dax: fix NULL pointer in __dax_pmd_fault()
mm: hugetlbfs: skip shared VMAs when unmapping private pages to satisfy a fault
mm/slab: fix unexpected index mapping result of kmalloc_size(INDEX_NODE+1)
userfaultfd: remove kernel header include from uapi header
arch/x86/include/asm/efi.h: fix build failure
Pull power management and ACPI fixes from Rafael Wysocki:
"These are fixes mostly, for a few changes made in this cycle (the
intel_idle driver, the OPP library, the ACPI EC driver, turbostat) and
for some issues that have just been discovered (ACPI PCI IRQ
management, PCI power management documentation, turbostat), with a
couple of cleanups on top of them.
Specifics:
- intel_idle driver fixup for the recently added Skylake chips
support (Len Brown).
- Operating Performance Points (OPP) library fix related to the
recently added support for new DT bindings and a fix for a typo in
a comment (Viresh Kumar, Stephen Boyd).
- ACPI EC driver fix for a recently introduced memory leak in an
error code path (Lv Zheng).
- ACPI PCI IRQ management fix for the issue where an ISA IRQ is
shared with a PCI device which requires it to be configured in a
different way and may cause an interrupt storm to happen as a
result with an extra ACPI SCI IRQ handling simplification on top of
it (Jiang Liu).
- Update of the PCI power management documentation that became
outdated and started to actively confuse the readers to make it
actually reflect the code (Rafael J Wysocki).
- turbostat fixes including an IVB Xeon regression fix (related to
the --debug command line option), Skylake adjustment for the TSC
running at a frequency that doesn't match the base one exactly, and
a Knights Landing quirk to account for the fact that it only
updates APERF and MPERF every 1024 clock cycles plus bumping up the
turbostat version number (Len Brown, Hubert Chrzaniuk)"
* tag 'pm+acpi-4.3-rc4' of git://git.kernel.org/pub/scm/linux/kernel/git/rafael/linux-pm:
tools/power turbosat: update version number
tools/power turbostat: SKL: Adjust for TSC difference from base frequency
tools/power turbostat: KNL workaround for %Busy and Avg_MHz
tools/power turbostat: IVB Xeon: fix --debug regression
ACPI / PCI: Remove duplicated penalty on SCI IRQ
ACPI, PCI, irq: Do not share PCI IRQ with ISA IRQ
ACPI / EC: Fix a memory leak issue in acpi_ec_query()
PM / OPP: Fix typo modifcation -> modification
PCI / PM: Update runtime PM documentation for PCI devices
PM / OPP: of_property_count_u32_elems() can return errors
intel_idle: Skylake Client Support - updated
Pull networking fixes from David Miller:
1) Fix regression in SKB partial checksum handling, from Pravin B
Shalar.
2) Fix VLAN inside of VXLAN handling in i40e driver, from Jesse
Brandeburg.
3) Cure softlockups during accept() in SCTP, from Karl Heiss.
4) MSG_PEEK should return multiple SKBs worth of data in AF_UNIX, from
Aaron Conole.
5) IPV6 erroneously ignores output interface specifier in lookup key for
route lookups, fix from David Ahern.
6) In Marvell DSA driver, forward unknown frames to CPU port, from
Andrew Lunn.
7) Mission flow flag initializations in some code paths, from David
Ahern.
* git://git.kernel.org/pub/scm/linux/kernel/git/davem/net:
net: Initialize flow flags in input path
net: dsa: fix preparation of a port STP update
testptp: Silence compiler warnings on ppc64
net/mlx4: Handle return codes in mlx4_qp_attach_common
dsa: mv88e6xxx: Enable forwarding for unknown to the CPU port
skbuff: Fix skb checksum partial check.
net: ipv6: Add RT6_LOOKUP_F_IFACE flag if oif is set
net sysfs: Print link speed as signed integer
bna: fix error handling
af_unix: return data from multiple SKBs on recv() with MSG_PEEK flag
af_unix: Convert the unix_sk macro to an inline function for type safety
net: sctp: Don't use 64 kilobyte lookup table for four elements
l2tp: protect tunnel->del_work by ref_count
net/ibm/emac: bump version numbers for correct work with ethtool
sctp: Prevent soft lockup when sctp_accept() is called during a timeout event
sctp: Whitespace fix
i40e/i40evf: check for stopped admin queue
i40e: fix VLAN inside VXLAN
r8169: fix handling rtl_readphy result
net: hisilicon: fix handling platform_get_irq result
Jiri Pirko
David S. Miller
Eric Dumazet
Linus Torvalds