gro: Optimise TCP packet reception
As this function can be called more than half a million times for
10GbE, it's important to optimise it as much as we can.
This patch uses bit ops to logical ops, as well as open coding
memcmp to exploit alignment properties.
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
Unfortunately simplicity isn't always the best. The fraginfo
interface turned out to be suboptimal. The problem was quite
obvious. For every packet, we have to copy the headers from
the frags structure into skb->head, even though for 99% of the
packets this part is immediately thrown away after the merge.
LRO didn't have this problem because it directly read the headers
from the frags structure.
This patch attempts to address this by creating an interface
that allows GRO to access the headers in the first frag without
having to copy it. Because all drivers that use frags place the
headers in the first frag this optimisation should be enough.
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
tcp_splice_data_recv has two lengths to consider: the len parameter it
gets from tcp_read_sock, which specifies the amount of data in the skb,
and rd_desc->count, which is the amount of data the splice caller still
wants. Currently it passes just the latter to skb_splice_bits, which then
splices min(rd_desc->count, skb->len - offset) bytes.
Most of the time this is fine, except when the skb contains urgent data.
In that case len goes only up to the urgent byte and is less than
skb->len - offset. By ignoring len tcp_splice_data_recv may a) splice
data tcp_read_sock told it not to, b) return to tcp_read_sock a value > len.
Now, tcp_read_sock doesn't handle used > len and leaves the socket in a
bad state (both sk_receive_queue and copied_seq are bad at that point)
resulting in duplicated data and corruption.
Fix by passing min(rd_desc->count, len) to skb_splice_bits.
Signed-off-by: Dimitris Michailidis <dm@chelsio.com>
Acked-by: Eric Dumazet <dada1@cosmosbay.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
When we get a GSO packet from an untrusted source, we need to
ensure that it is sufficiently long so that we don't end up
crashing.
Based on discovery and patch by Ian Campbell.
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Tested-by: Ian Campbell <ian.campbell@citrix.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
As spotted by Willy Tarreau, current splice() from tcp socket to pipe is not
optimal. It processes at most one segment per call.
This results in low performance and very high overhead due to syscall rate
when splicing from interfaces which do not support LRO.
Willy provided a patch inside tcp_splice_read(), but a better fix
is to let tcp_read_sock() process as many segments as possible, so
that tcp_rcv_space_adjust() and tcp_cleanup_rbuf() are called less
often.
With this change, splice() behaves like tcp_recvmsg(), being able
to consume many skbs in one system call. With typical 1460 bytes
of payload per frame, that means splice(SPLICE_F_NONBLOCK) can return
16*1460 = 23360 bytes.
Signed-off-by: Willy Tarreau <w@1wt.eu>
Signed-off-by: Eric Dumazet <dada1@cosmosbay.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
* 'next' of git://git.kernel.org/pub/scm/linux/kernel/git/djbw/async_tx: (22 commits)
ioat: fix self test for multi-channel case
dmaengine: bump initcall level to arch_initcall
dmaengine: advertise all channels on a device to dma_filter_fn
dmaengine: use idr for registering dma device numbers
dmaengine: add a release for dma class devices and dependent infrastructure
ioat: do not perform removal actions at shutdown
iop-adma: enable module removal
iop-adma: kill debug BUG_ON
iop-adma: let devm do its job, don't duplicate free
dmaengine: kill enum dma_state_client
dmaengine: remove 'bigref' infrastructure
dmaengine: kill struct dma_client and supporting infrastructure
dmaengine: replace dma_async_client_register with dmaengine_get
atmel-mci: convert to dma_request_channel and down-level dma_slave
dmatest: convert to dma_request_channel
dmaengine: introduce dma_request_channel and private channels
net_dma: convert to dma_find_channel
dmaengine: provide a common 'issue_pending_all' implementation
dmaengine: centralize channel allocation, introduce dma_find_channel
dmaengine: up-level reference counting to the module level
...
This patch adds GRO support for TCP over IPv6. The code is exactly
the same as the IPv4 version except for the pseudo-header checksum
computation.
Note that I've removed the unused tcphdr argument from tcp_v6_check
rather than invent a bogus value for GRO.
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
Use the general-purpose channel allocation provided by dmaengine.
Reviewed-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
Simply, if a client wants any dmaengine channel then prevent all dmaengine
modules from being removed. Once the clients are done re-enable module
removal.
Why?, beyond reducing complication:
1/ Tracking reference counts per-transaction in an efficient manner, as
is currently done, requires a complicated scheme to avoid cache-line
bouncing effects.
2/ Per-transaction ref-counting gives the false impression that a
dma-driver can be gracefully removed ahead of its user (net, md, or
dma-slave)
3/ None of the in-tree dma-drivers talk to hot pluggable hardware, but
if such an engine were built one day we still would not need to notify
clients of remove events. The driver can simply return NULL to a
->prep() request, something that is much easier for a client to handle.
Reviewed-by: Andrew Morton <akpm@linux-foundation.org>
Acked-by: Maciej Sosnowski <maciej.sosnowski@intel.com>
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
In splice TCP receive, the SPLICE_F_NONBLOCK flag is used
to compute the "timeo" value. So checking it again inside
of the main receive loop to trigger -EAGAIN processing is
entirely unnecessary.
Noticed by Jarek P. and Lennert Buytenhek.
Signed-off-by: David S. Miller <davem@davemloft.net>
Currently, setting SPLICE_F_NONBLOCK on splice from a TCP socket
results in masking of EOF (RDHUP) and error conditions on the socket
by an -EAGAIN return. Move the NONBLOCK check in tcp_splice_read()
to be after the EOF and error checks to fix this.
Signed-off-by: Lennert Buytenhek <buytenh@marvell.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
In order to allow GRO packets without frag_list at all, we need to
store the MSS in the packet itself. The obvious place is gso_size.
The only thing to watch out for is if the packet ends up not being
GRO then we need to clear gso_size before pushing the packet into
the stack.
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
When we converted the protocol atomic counters such as the orphan
count and the total socket count deadlocks were introduced due to
the mismatch in BH status of the spots that used the percpu counter
operations.
Based on the diagnosis and patch by Peter Zijlstra, this patch
fixes these issues by disabling BH where we may be in process
context.
Reported-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
Tested-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch adds the TCP-specific portion of GRO. The criterion for
merging is extremely strict (the TCP header must match exactly apart
from the checksum) so as to allow refragmentation. Otherwise this
is pretty much identical to LRO, except that we support the merging
of ECN packets.
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
Instead of using one atomic_t per protocol, use a percpu_counter
for "orphan_count", to reduce cache line contention on
heavy duty network servers.
Signed-off-by: Eric Dumazet <dada1@cosmosbay.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Instead of using one atomic_t per protocol, use a percpu_counter
for "sockets_allocated", to reduce cache line contention on
heavy duty network servers.
Note : We revert commit (248969ae31
net: af_unix can make unix_nr_socks visbile in /proc),
since it is not anymore used after sock_prot_inuse_add() addition
Signed-off-by: Eric Dumazet <dada1@cosmosbay.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
RCU was added to UDP lookups, using a fast infrastructure :
- sockets kmem_cache use SLAB_DESTROY_BY_RCU and dont pay the
price of call_rcu() at freeing time.
- hlist_nulls permits to use few memory barriers.
This patch uses same infrastructure for TCP/DCCP established
and timewait sockets.
Thanks to SLAB_DESTROY_BY_RCU, no slowdown for applications
using short lived TCP connections. A followup patch, converting
rwlocks to spinlocks will even speedup this case.
__inet_lookup_established() is pretty fast now we dont have to
dirty a contended cache line (read_lock/read_unlock)
Only established and timewait hashtable are converted to RCU
(bind table and listen table are still using traditional locking)
Signed-off-by: Eric Dumazet <dada1@cosmosbay.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Vito Caputo noticed that tcp_recvmsg() returns immediately from
partial reads when MSG_PEEK is used. In particular, this means that
SO_RCVLOWAT is not respected.
Simply remove the test. And this matches the behavior of several
other systems, including BSD.
Signed-off-by: David S. Miller <davem@davemloft.net>
It all started from me noticing that this urgent check in
tcp_clean_rtx_queue is unnecessarily inside the loop. Then
I took a longer look to it and found out that the users of
urg_mode can trivially do without, well almost, there was
one gotcha.
Bonus: those funny people who use urg with >= 2^31 write_seq -
snd_una could now rejoice too (that's the only purpose for the
between being there, otherwise a simple compare would have done
the thing). Not that I assume that the rest of the tcp code
happily lives with such mind-boggling numbers :-). Alas, it
turned out to be impossible to set wmem to such numbers anyway,
yes I really tried a big sendfile after setting some wmem but
nothing happened :-). ...Tcp_wmem is int and so is sk_sndbuf...
So I hacked a bit variable to long and found out that it seems
to work... :-)
Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>
Wrap calling sk->sk_backlog_rcv() in a function. This will allow extending the
generic sk_backlog_rcv behaviour.
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: David S. Miller <davem@davemloft.net>
Removes legacy reinvent-the-wheel type thing. The generic
machinery integrates much better to automated debugging aids
such as kerneloops.org (and others), and is unambiguous due to
better naming. Non-intuively BUG_TRAP() is actually equal to
WARN_ON() rather than BUG_ON() though some might actually be
promoted to BUG_ON() but I left that to future.
I could make at least one BUILD_BUG_ON conversion.
Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>
Currently, the MD5 code assumes that the SKBs are linear and, in the case
that they aren't, happily goes off and hashes off the end of the SKB and
into random memory.
Reported by Stephen Hemminger in [1]. Advice thanks to Stephen and Evgeniy
Polyakov. Also includes a couple of missed route_caps from Stephen's patch
in [2].
[1] http://marc.info/?l=linux-netdev&m=121445989106145&w=2
[2] http://marc.info/?l=linux-netdev&m=121459157816964&w=2
Signed-off-by: Adam Langley <agl@imperialviolet.org>
Acked-by: Stephen Hemminger <shemminger@vyatta.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Proc temporary uses stats from init_net.
BTW, TCP_XXX_STATS are beautiful (w/o do { } while (0) facing) again :)
Signed-off-by: Pavel Emelyanov <xemul@openvz.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
Done with NET_XXX_STATS macros :)
To be continued...
Signed-off-by: Pavel Emelyanov <xemul@openvz.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
The tcp_enter_memory_pressure calls NET_INC_STATS, but doesn't
have where to get the net from.
I decided to add a sk argument, not the net itself, only to factor
all the required sock_net(sk) calls inside the enter_memory_pressure
callback itself.
Signed-off-by: Pavel Emelyanov <xemul@openvz.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
Same as before - the sock is always there to get the net from,
but there are also some places with the net already saved on
the stack.
Signed-off-by: Pavel Emelyanov <xemul@openvz.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
Fortunately (almost) all the TCP code has a sock to get the net from :)
Signed-off-by: Pavel Emelyanov <xemul@openvz.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
Change PULLHUP to POLLHUP in tcp_poll comments and clean up another
comment for grammar and coding style.
Signed-off-by: Will Newton <will.newton@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
<used> should be of type int (not size_t) since recv_actor can return
negative values and it is also used in a < 0 comparison.
Signed-off-by: Octavian Purdila <opurdila@ixiacom.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
alpha:
net/ipv4/tcp.c: In function 'tcp_calc_md5_hash':
net/ipv4/tcp.c:2479: error: implicit declaration of function 'sg_init_table' net/ipv4/tcp.c:2482: error: implicit declaration of function 'sg_set_buf'
net/ipv4/tcp.c:2507: error: implicit declaration of function 'sg_mark_end'
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
The tcp_mem array which contains limits on the total amount of memory
used by TCP sockets is calculated based on nr_all_pages. On a 32 bits
x86 system, we should base this on the number of lowmem pages.
Signed-off-by: Miquel van Smoorenburg <miquels@cistron.nl>
Signed-off-by: David S. Miller <davem@davemloft.net>
This reverts two changesets, ec3c0982a2
("[TCP]: TCP_DEFER_ACCEPT updates - process as established") and
the follow-on bug fix 9ae27e0adb
("tcp: Fix slab corruption with ipv6 and tcp6fuzz").
This change causes several problems, first reported by Ingo Molnar
as a distcc-over-loopback regression where connections were getting
stuck.
Ilpo Järvinen first spotted the locking problems. The new function
added by this code, tcp_defer_accept_check(), only has the
child socket locked, yet it is modifying state of the parent
listening socket.
Fixing that is non-trivial at best, because we can't simply just grab
the parent listening socket lock at this point, because it would
create an ABBA deadlock. The normal ordering is parent listening
socket --> child socket, but this code path would require the
reverse lock ordering.
Next is a problem noticed by Vitaliy Gusev, he noted:
----------------------------------------
>--- a/net/ipv4/tcp_timer.c
>+++ b/net/ipv4/tcp_timer.c
>@@ -481,6 +481,11 @@ static void tcp_keepalive_timer (unsigned long data)
> goto death;
> }
>
>+ if (tp->defer_tcp_accept.request && sk->sk_state == TCP_ESTABLISHED) {
>+ tcp_send_active_reset(sk, GFP_ATOMIC);
>+ goto death;
Here socket sk is not attached to listening socket's request queue. tcp_done()
will not call inet_csk_destroy_sock() (and tcp_v4_destroy_sock() which should
release this sk) as socket is not DEAD. Therefore socket sk will be lost for
freeing.
----------------------------------------
Finally, Alexey Kuznetsov argues that there might not even be any
real value or advantage to these new semantics even if we fix all
of the bugs:
----------------------------------------
Hiding from accept() sockets with only out-of-order data only
is the only thing which is impossible with old approach. Is this really
so valuable? My opinion: no, this is nothing but a new loophole
to consume memory without control.
----------------------------------------
So revert this thing for now.
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch removes CVS keywords that weren't updated for a long time
from comments.
Signed-off-by: Adrian Bunk <bunk@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
We can share most part of the hash calculation code because
the only difference between IPv4 and IPv6 is their pseudo headers.
Signed-off-by: YOSHIFUJI Hideaki <yoshfuji@linux-ipv6.org>
skb_splice_bits temporary drops the socket lock while iterating over
the socket queue in order to break a reverse locking condition which
happens with sendfile. This, however, opens a window of opportunity
for tcp_collapse() to aggregate skbs and thus potentially free the
current skb used in skb_splice_bits and tcp_read_sock.
This patch fixes the problem by (re-)getting the same "logical skb"
after the lock has been temporary dropped.
Based on idea and initial patch from Evgeniy Polyakov.
Signed-off-by: Octavian Purdila <opurdila@ixiacom.com>
Acked-by: Evgeniy Polyakov <johnpol@2ka.mipt.ru>
Signed-off-by: David S. Miller <davem@davemloft.net>
This is a trivial fix to correct function name in a comment in
net/ipv4/tcp.c.
Signed-off-by: Satoru SATOH <satoru.satoh@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
While testing the virtio-net driver on KVM with TSO I noticed
that TSO performance with a 1500 MTU is significantly worse
compared to the performance of non-TSO with a 16436 MTU. The
packet dump shows that most of the packets sent are smaller
than a page.
Looking at the code this actually is quite obvious as it always
stop extending the packet if it's the first packet yet to be
sent and if it's larger than the MSS. Since each extension is
bound by the page size, this means that (given a 1500 MTU) we're
very unlikely to construct packets greater than a page, provided
that the receiver and the path is fast enough so that packets can
always be sent immediately.
The fix is also quite obvious. The push calls inside the loop
is just an optimisation so that we don't end up doing all the
sending at the end of the loop. Therefore there is no specific
reason why it has to do so at MSS boundaries. For TSO, the
most natural extension of this optimisation is to do the pushing
once the skb exceeds the TSO size goal.
This is what the patch does and testing with KVM shows that the
TSO performance with a 1500 MTU easily surpasses that of a 16436
MTU and indeed the packet sizes sent are generally larger than
16436.
I don't see any obvious downsides for slower peers or connections,
but it would be prudent to test this extensively to ensure that
those cases don't regress.
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
