We currently use a per socket order-0 page cache for tcp_sendmsg()
operations.
This page is used to build fragments for skbs.
Its done to increase probability of coalescing small write() into
single segments in skbs still in write queue (not yet sent)
But it wastes a lot of memory for applications handling many mostly
idle sockets, since each socket holds one page in sk->sk_sndmsg_page
Its also quite inefficient to build TSO 64KB packets, because we need
about 16 pages per skb on arches where PAGE_SIZE = 4096, so we hit
page allocator more than wanted.
This patch adds a per task frag allocator and uses bigger pages,
if available. An automatic fallback is done in case of memory pressure.
(up to 32768 bytes per frag, thats order-3 pages on x86)
This increases TCP stream performance by 20% on loopback device,
but also benefits on other network devices, since 8x less frags are
mapped on transmit and unmapped on tx completion. Alexander Duyck
mentioned a probable performance win on systems with IOMMU enabled.
Its possible some SG enabled hardware cant cope with bigger fragments,
but their ndo_start_xmit() should already handle this, splitting a
fragment in sub fragments, since some arches have PAGE_SIZE=65536
Successfully tested on various ethernet devices.
(ixgbe, igb, bnx2x, tg3, mellanox mlx4)
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Ben Hutchings <bhutchings@solarflare.com>
Cc: Vijay Subramanian <subramanian.vijay@gmail.com>
Cc: Alexander Duyck <alexander.h.duyck@intel.com>
Tested-by: Vijay Subramanian <subramanian.vijay@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
IPv4 conntrack defragments incoming packet at the PRE_ROUTING hook and
(in case of forwarded packets) refragments them at POST_ROUTING
independent of the IP_DF flag. Refragmentation uses the dst_mtu() of
the local route without caring about the original fragment sizes,
thereby breaking PMTUD.
This patch fixes this by keeping track of the largest received fragment
with IP_DF set and generates an ICMP fragmentation required error during
refragmentation if that size exceeds the MTU.
Signed-off-by: Patrick McHardy <kaber@trash.net>
Acked-by: Eric Dumazet <edumazet@google.com>
Acked-by: David S. Miller <davem@davemloft.net>
Christian Casteyde reported a kmemcheck 32-bit read from uninitialized
memory in __ip_select_ident().
It turns out that __ip_make_skb() called ip_select_ident() before
properly initializing iph->daddr.
This is a bug uncovered by commit 1d861aa4b3 (inet: Minimize use of
cached route inetpeer.)
Addresses https://bugzilla.kernel.org/show_bug.cgi?id=46131
Reported-by: Christian Casteyde <casteyde.christian@free.fr>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Stephen Hemminger <shemminger@vyatta.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
ip_send_skb() can send orphaned skb, so we must pass the net pointer to
avoid possible NULL dereference in error path.
Bug added by commit 3a7c384ffd (ipv4: tcp: unicast_sock should not
land outside of TCP stack)
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
commit be9f4a44e7 (ipv4: tcp: remove per net tcp_sock) added a
selinux regression, reported and bisected by John Stultz
selinux_ip_postroute_compat() expect to find a valid sk->sk_security
pointer, but this field is NULL for unicast_sock
It turns out that unicast_sock are really temporary stuff to be able
to reuse part of IP stack (ip_append_data()/ip_push_pending_frames())
Fact is that frames sent by ip_send_unicast_reply() should be orphaned
to not fool LSM.
Note IPv6 never had this problem, as tcp_v6_send_response() doesnt use a
fake socket at all. I'll probably implement tcp_v4_send_response() to
remove these unicast_sock in linux-3.7
Reported-by: John Stultz <johnstul@us.ibm.com>
Bisected-by: John Stultz <johnstul@us.ibm.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Paul Moore <paul@paul-moore.com>
Cc: Eric Paris <eparis@parisplace.org>
Cc: "Serge E. Hallyn" <serge@hallyn.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
__neigh_create() returns either a pointer to struct neighbour or PTR_ERR().
But the caller expects it to return either a pointer or NULL. Replace
the NULL check with IS_ERR() check.
The bug was introduced in a263b30936
("ipv4: Make neigh lookups directly in output packet path.").
Signed-off-by: Vasily Kulikov <segoon@openwall.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The ipv4 routing cache is non-deterministic, performance wise, and is
subject to reasonably easy to launch denial of service attacks.
The routing cache works great for well behaved traffic, and the world
was a much friendlier place when the tradeoffs that led to the routing
cache's design were considered.
What it boils down to is that the performance of the routing cache is
a product of the traffic patterns seen by a system rather than being a
product of the contents of the routing tables. The former of which is
controllable by external entitites.
Even for "well behaved" legitimate traffic, high volume sites can see
hit rates in the routing cache of only ~%10.
The general flow of this patch series is that first the routing cache
is removed. We build a completely new rtable entry every lookup
request.
Next we make some simplifications due to the fact that removing the
routing cache causes several members of struct rtable to become no
longer necessary.
Then we need to make some amends such that we can legally cache
pre-constructed routes in the FIB nexthops. Firstly, we need to
invalidate routes which are hit with nexthop exceptions. Secondly we
have to change the semantics of rt->rt_gateway such that zero means
that the destination is on-link and non-zero otherwise.
Now that the preparations are ready, we start caching precomputed
routes in the FIB nexthops. Output and input routes need different
kinds of care when determining if we can legally do such caching or
not. The details are in the commit log messages for those changes.
The patch series then winds down with some more struct rtable
simplifications and other tidy ups that remove unnecessary overhead.
On a SPARC-T3 output route lookups are ~876 cycles. Input route
lookups are ~1169 cycles with rpfilter disabled, and about ~1468
cycles with rpfilter enabled.
These measurements were taken with the kbench_mod test module in the
net_test_tools GIT tree:
git://git.kernel.org/pub/scm/linux/kernel/git/davem/net_test_tools.git
That GIT tree also includes a udpflood tester tool and stresses
route lookups on packet output.
For example, on the same SPARC-T3 system we can run:
time ./udpflood -l 10000000 10.2.2.11
with routing cache:
real 1m21.955s user 0m6.530s sys 1m15.390s
without routing cache:
real 1m31.678s user 0m6.520s sys 1m25.140s
Performance undoubtedly can easily be improved further.
For example fib_table_lookup() performs a lot of excessive
computations with all the masking and shifting, some of it
conditionalized to deal with edge cases.
Also, Eric's no-ref optimization for input route lookups can be
re-instated for the FIB nexthop caching code path. I would be really
pleased if someone would work on that.
In fact anyone suitable motivated can just fire up perf on the loading
of the test net_test_tools benchmark kernel module. I spend much of
my time going:
bash# perf record insmod ./kbench_mod.ko dst=172.30.42.22 src=74.128.0.1 iif=2
bash# perf report
Thanks to helpful feedback from Joe Perches, Eric Dumazet, Ben
Hutchings, and others.
Signed-off-by: David S. Miller <davem@davemloft.net>
Set unicast_sock uc_ttl to -1 so that we select the right ttl,
instead of sending packets with a 0 ttl.
Bug added in commit be9f4a44e7 (ipv4: tcp: remove per net tcp_sock)
Signed-off-by: Hiroaki SHIMODA <shimoda.hiroaki@gmail.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
In order to allow prefixed routes, we have to adjust how rt_gateway
is set and interpreted.
The new interpretation is:
1) rt_gateway == 0, destination is on-link, nexthop is iph->daddr
2) rt_gateway != 0, destination requires a nexthop gateway
Abstract the fetching of the proper nexthop value using a new
inline helper, rt_nexthop(), as suggested by Joe Perches.
Signed-off-by: David S. Miller <davem@davemloft.net>
Tested-by: Vijay Subramanian <subramanian.vijay@gmail.com>
tcp_v4_send_reset() and tcp_v4_send_ack() use a single socket
per network namespace.
This leads to bad behavior on multiqueue NICS, because many cpus
contend for the socket lock and once socket lock is acquired, extra
false sharing on various socket fields slow down the operations.
To better resist to attacks, we use a percpu socket. Each cpu can
run without contention, using appropriate memory (local node)
Additional features :
1) We also mirror the queue_mapping of the incoming skb, so that
answers use the same queue if possible.
2) Setting SOCK_USE_WRITE_QUEUE socket flag speedup sock_wfree()
3) We now limit the number of in-flight RST/ACK [1] packets
per cpu, instead of per namespace, and we honor the sysctl_wmem_default
limit dynamically. (Prior to this patch, sysctl_wmem_default value was
copied at boot time, so any further change would not affect tcp_sock
limit)
[1] These packets are only generated when no socket was matched for
the incoming packet.
Reported-by: Bill Sommerfeld <wsommerfeld@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Tom Herbert <therbert@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
When a dst_confirm() happens, mark the confirmation as pending in the
dst. Then on the next packet out, when we have the neigh in-hand, do
the update.
This removes the dependency in dst_confirm() of dst's having an
attached neigh.
While we're here, remove the explicit 'dst' NULL check, all except 2
or 3 call sites ensure it's not NULL. So just fix those cases up.
Signed-off-by: David S. Miller <davem@davemloft.net>
Rename it to ip_send_unicast_reply() and add explicit 'saddr'
argument.
This removed one of the few users of rt->rt_spec_dst.
Signed-off-by: David S. Miller <davem@davemloft.net>
Add dev_loopback_xmit() in order to deduplicate functions
ip_dev_loopback_xmit() (in net/ipv4/ip_output.c) and
ip6_dev_loopback_xmit() (in net/ipv6/ip6_output.c).
I was about to reinvent the wheel when I noticed that
ip_dev_loopback_xmit() and ip6_dev_loopback_xmit() do exactly what I
need and are not IP-only functions, but they were not available to reuse
elsewhere.
ip6_dev_loopback_xmit() does not have line "skb_dst_force(skb);", but I
understand that this is harmless, and should be in dev_loopback_xmit().
Signed-off-by: Michel Machado <michel@digirati.com.br>
CC: "David S. Miller" <davem@davemloft.net>
CC: Alexey Kuznetsov <kuznet@ms2.inr.ac.ru>
CC: James Morris <jmorris@namei.org>
CC: Hideaki YOSHIFUJI <yoshfuji@linux-ipv6.org>
CC: Patrick McHardy <kaber@trash.net>
CC: Eric Dumazet <edumazet@google.com>
CC: Jiri Pirko <jpirko@redhat.com>
CC: "Michał Mirosław" <mirq-linux@rere.qmqm.pl>
CC: Ben Hutchings <bhutchings@solarflare.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Remove some dropwatch/drop_monitor false positives.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Standardize the net core ratelimited logging functions.
Coalesce formats, align arguments.
Change a printk then vprintk sequence to use printf extension %pV.
Signed-off-by: Joe Perches <joe@perches.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Remove all #inclusions of asm/system.h preparatory to splitting and killing
it. Performed with the following command:
perl -p -i -e 's!^#\s*include\s*<asm/system[.]h>.*\n!!' `grep -Irl '^#\s*include\s*<asm/system[.]h>' *`
Signed-off-by: David Howells <dhowells@redhat.com>
To reflect the fact that a refrence is not obtained to the
resulting neighbour entry.
Signed-off-by: David S. Miller <davem@davemloft.net>
Acked-by: Roland Dreier <roland@purestorage.com>
gcc compiler is smart enough to use a single load/store if we
memcpy(dptr, sptr, 8) on x86_64, regardless of
CONFIG_CC_OPTIMIZE_FOR_SIZE
In IP header, daddr immediately follows saddr, this wont change in the
future. We only need to make sure our flowi4 (saddr,daddr) fields wont
break the rule.
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
There is a long standing bug in linux tcp stack, about ACK messages sent
on behalf of TIME_WAIT sockets.
In the IP header of the ACK message, we choose to reflect TOS field of
incoming message, and this might break some setups.
Example of things that were broken :
- Routing using TOS as a selector
- Firewalls
- Trafic classification / shaping
We now remember in timewait structure the inet tos field and use it in
ACK generation, and route lookup.
Notes :
- We still reflect incoming TOS in RST messages.
- We could extend MuraliRaja Muniraju patch to report TOS value in
netlink messages for TIME_WAIT sockets.
- A patch is needed for IPv6
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
To ease skb->truesize sanitization, its better to be able to localize
all references to skb frags size.
Define accessors : skb_frag_size() to fetch frag size, and
skb_frag_size_{set|add|sub}() to manipulate it.
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Make sure skb dst has reference when moving to
another context. Currently, I don't see protocols that can
hit it when sending broadcasts/multicasts to loopback using
noref dsts, so it is just a precaution.
Signed-off-by: Julian Anastasov <ja@ssi.bg>
Signed-off-by: David S. Miller <davem@davemloft.net>
Gergely Kalman reported crashes in check_peer_redir().
It appears commit f39925dbde (ipv4: Cache learned redirect
information in inetpeer.) added a race, leading to possible NULL ptr
dereference.
Since we can now change dst neighbour, we should make sure a reader can
safely use a neighbour.
Add RCU protection to dst neighbour, and make sure check_peer_redir()
can be called safely by different cpus in parallel.
As neighbours are already freed after one RCU grace period, this patch
should not add typical RCU penalty (cache cold effects)
Many thanks to Gergely for providing a pretty report pointing to the
bug.
Reported-by: Gergely Kalman <synapse@hippy.csoma.elte.hu>
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Because the ip fragment offset field counts 8-byte chunks, ip
fragments other than the last must contain a multiple of 8 bytes of
payload. ip_ufo_append_data wasn't respecting this constraint and,
depending on the MTU and ip option sizes, could create malformed
non-final fragments.
Google-Bug-Id: 5009328
Signed-off-by: Bill Sommerfeld <wsommerfeld@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Now that there is a one-to-one correspondance between neighbour
and hh_cache entries, we no longer need:
1) dynamic allocation
2) attachment to dst->hh
3) refcounting
Initialization of the hh_cache entry is indicated by hh_len
being non-zero, and such initialization is always done with
the neighbour's lock held as a writer.
Signed-off-by: David S. Miller <davem@davemloft.net>
We might call ip_ufo_append_data() for packets that will be IPsec
transformed later. This function should be used just for real
udp packets. So we check for rt->dst.header_len which is only
nonzero on IPsec handling and call ip_ufo_append_data() just
if rt->dst.header_len is zero.
Signed-off-by: Steffen Klassert <steffen.klassert@secunet.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
ip_append_data() builds packets based on the mtu from dst_mtu(rt->dst.path).
On IPsec the effective mtu is lower because we need to add the protocol
headers and trailers later when we do the IPsec transformations. So after
the IPsec transformations the packet might be too big, which leads to a
slowpath fragmentation then. This patch fixes this by building the packets
based on the lower IPsec mtu from dst_mtu(&rt->dst) and adapts the exthdr
handling to this.
Signed-off-by: Steffen Klassert <steffen.klassert@secunet.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Git commit 59104f06 (ip: take care of last fragment in ip_append_data)
added a check to see if we exceed the mtu when we add trailer_len.
However, the mtu is already subtracted by the trailer length when the
xfrm transfomation bundles are set up. So IPsec packets with mtu
size get fragmented, or if the DF bit is set the packets will not
be send even though they match the mtu perfectly fine. This patch
actually reverts commit 59104f06.
Signed-off-by: Steffen Klassert <steffen.klassert@secunet.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
There are enough instances of this:
iph->frag_off & htons(IP_MF | IP_OFFSET)
that a helper function is probably warranted.
Signed-off-by: Paul Gortmaker <paul.gortmaker@windriver.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
We assume that transhdrlen is positive on the first fragment
which is wrong for raw packets. So we don't add exthdrlen to the
packet size for raw packets. This leads to a reallocation on IPsec
because we have not enough headroom on the skb to place the IPsec
headers. This patch fixes this by adding exthdrlen to the packet
size whenever the send queue of the socket is empty. This issue was
introduced with git commit 1470ddf7 (inet: Remove explicit write
references to sk/inet in ip_append_data)
Signed-off-by: Steffen Klassert <steffen.klassert@secunet.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This way ip_output.c no longer needs rt->rt_{src,dst}.
We already have these keys sitting, ready and waiting, on the stack or
in a socket structure.
Signed-off-by: David S. Miller <davem@davemloft.net>
This allows us to acquire the exact route keying information from the
protocol, however that might be managed.
It handles all of the possibilities, from the simplest case of storing
the key in inet->cork.fl to the more complex setup SCTP has where
individual transports determine the flow.
Signed-off-by: David S. Miller <davem@davemloft.net>
All invokers of ip_queue_xmit() must make certain that the
socket is locked. All of SCTP, TCP, DCCP, and L2TP now make
sure this is the case.
Therefore we can use the cork flow during output route lookup in
ip_queue_xmit() when the socket route check fails.
Signed-off-by: David S. Miller <davem@davemloft.net>
ip_setup_cork() explicitly initializes every member of
inet_cork except flags, addr, and opt. So we can simply
set those three members to zero instead of using a
memset() via an empty struct assignment.
Signed-off-by: David S. Miller <davem@davemloft.net>
Acked-by: Eric Dumazet <eric.dumazet@gmail.com>
When we fast path datagram sends to avoid locking by putting
the inet_cork on the stack we use up lots of space that isn't
necessary.
This is because inet_cork contains a "struct flowi" which isn't
used in these code paths.
Split inet_cork to two parts, "inet_cork" and "inet_cork_full".
Only the latter of which has the "struct flowi" and is what is
stored in inet_sock.
Signed-off-by: David S. Miller <davem@davemloft.net>
Acked-by: Eric Dumazet <eric.dumazet@gmail.com>
Instead of rt->rt_{dst,src}
The only tricky part is source route option handling.
If the source route option is enabled we can't just use plain 'daddr',
we have to use opt->opt.faddr.
Signed-off-by: David S. Miller <davem@davemloft.net>
We lack proper synchronization to manipulate inet->opt ip_options
Problem is ip_make_skb() calls ip_setup_cork() and
ip_setup_cork() possibly makes a copy of ipc->opt (struct ip_options),
without any protection against another thread manipulating inet->opt.
Another thread can change inet->opt pointer and free old one under us.
Use RCU to protect inet->opt (changed to inet->inet_opt).
Instead of handling atomic refcounts, just copy ip_options when
necessary, to avoid cache line dirtying.
We cant insert an rcu_head in struct ip_options since its included in
skb->cb[], so this patch is large because I had to introduce a new
ip_options_rcu structure.
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Cc: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
