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>
Change TCP_DEFER_ACCEPT implementation so that it transitions a
connection to ESTABLISHED after handshake is complete instead of
leaving it in SYN-RECV until some data arrvies. Place connection in
accept queue when first data packet arrives from slow path.
Benefits:
- established connection is now reset if it never makes it
to the accept queue
- diagnostic state of established matches with the packet traces
showing completed handshake
- TCP_DEFER_ACCEPT timeouts are expressed in seconds and can now be
enforced with reasonable accuracy instead of rounding up to next
exponential back-off of syn-ack retry.
Signed-off-by: Patrick McManus <mcmanus@ducksong.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This way we can remove TCP and DCCP specific versions of
sk->sk_prot->get_port: both v4 and v6 use inet_csk_get_port
sk->sk_prot->hash: inet_hash is directly used, only v6 need
a specific version to deal with mapped sockets
sk->sk_prot->unhash: both v4 and v6 use inet_hash directly
struct inet_connection_sock_af_ops also gets a new member, bind_conflict, so
that inet_csk_get_port can find the per family routine.
Now only the lookup routines receive as a parameter a struct inet_hashtable.
With this we further reuse code, reducing the difference among INET transport
protocols.
Eventually work has to be done on UDP and SCTP to make them share this
infrastructure and get as a bonus inet_diag interfaces so that iproute can be
used with these protocols.
net-2.6/net/ipv4/inet_hashtables.c:
struct proto | +8
struct inet_connection_sock_af_ops | +8
2 structs changed
__inet_hash_nolisten | +18
__inet_hash | -210
inet_put_port | +8
inet_bind_bucket_create | +1
__inet_hash_connect | -8
5 functions changed, 27 bytes added, 218 bytes removed, diff: -191
net-2.6/net/core/sock.c:
proto_seq_show | +3
1 function changed, 3 bytes added, diff: +3
net-2.6/net/ipv4/inet_connection_sock.c:
inet_csk_get_port | +15
1 function changed, 15 bytes added, diff: +15
net-2.6/net/ipv4/tcp.c:
tcp_set_state | -7
1 function changed, 7 bytes removed, diff: -7
net-2.6/net/ipv4/tcp_ipv4.c:
tcp_v4_get_port | -31
tcp_v4_hash | -48
tcp_v4_destroy_sock | -7
tcp_v4_syn_recv_sock | -2
tcp_unhash | -179
5 functions changed, 267 bytes removed, diff: -267
net-2.6/net/ipv6/inet6_hashtables.c:
__inet6_hash | +8
1 function changed, 8 bytes added, diff: +8
net-2.6/net/ipv4/inet_hashtables.c:
inet_unhash | +190
inet_hash | +242
2 functions changed, 432 bytes added, diff: +432
vmlinux:
16 functions changed, 485 bytes added, 492 bytes removed, diff: -7
/home/acme/git/net-2.6/net/ipv6/tcp_ipv6.c:
tcp_v6_get_port | -31
tcp_v6_hash | -7
tcp_v6_syn_recv_sock | -9
3 functions changed, 47 bytes removed, diff: -47
/home/acme/git/net-2.6/net/dccp/proto.c:
dccp_destroy_sock | -7
dccp_unhash | -179
dccp_hash | -49
dccp_set_state | -7
dccp_done | +1
5 functions changed, 1 bytes added, 242 bytes removed, diff: -241
/home/acme/git/net-2.6/net/dccp/ipv4.c:
dccp_v4_get_port | -31
dccp_v4_request_recv_sock | -2
2 functions changed, 33 bytes removed, diff: -33
/home/acme/git/net-2.6/net/dccp/ipv6.c:
dccp_v6_get_port | -31
dccp_v6_hash | -7
dccp_v6_request_recv_sock | +5
3 functions changed, 5 bytes added, 38 bytes removed, diff: -33
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The snd_up check should be enough. I suspect this has been
there to provide a minor optimization in clean_rtx_queue which
used to have a small if (!->sacked) block which could skip
snd_up check among the other work.
Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch introduces new memory accounting functions for each network
protocol. Most of them are renamed from memory accounting functions
for stream protocols. At the same time, some stream memory accounting
functions are removed since other functions do same thing.
Renaming:
sk_stream_free_skb() -> sk_wmem_free_skb()
__sk_stream_mem_reclaim() -> __sk_mem_reclaim()
sk_stream_mem_reclaim() -> sk_mem_reclaim()
sk_stream_mem_schedule -> __sk_mem_schedule()
sk_stream_pages() -> sk_mem_pages()
sk_stream_rmem_schedule() -> sk_rmem_schedule()
sk_stream_wmem_schedule() -> sk_wmem_schedule()
sk_charge_skb() -> sk_mem_charge()
Removeing
sk_stream_rfree(): consolidates into sock_rfree()
sk_stream_set_owner_r(): consolidates into skb_set_owner_r()
sk_stream_mem_schedule()
The following functions are added.
sk_has_account(): check if the protocol supports accounting
sk_mem_uncharge(): do the opposite of sk_mem_charge()
In addition, to achieve consolidation, updating sk_wmem_queued is
removed from sk_mem_charge().
Next, to consolidate memory accounting functions, this patch adds
memory accounting calls to network core functions. Moreover, present
memory accounting call is renamed to new accounting call.
Finally we replace present memory accounting calls with new interface
in TCP and SCTP.
Signed-off-by: Takahiro Yasui <tyasui@redhat.com>
Signed-off-by: Hideo Aoki <haoki@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The sizeof(struct tcp_skb_cb) should not be less than the
sizeof(skb->cb). This is checked in net/ipv4/tcp.c, but
this check can be made more gracefully.
Signed-off-by: Pavel Emelyanov <xemul@openvz.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
The 3rd argument is always zero (according to grep :) Eliminate
it and merge the function with sk_stream_alloc_skb.
This saves 44 more bytes, and together with the previous patch
we have:
add/remove: 1/0 grow/shrink: 0/8 up/down: 183/-751 (-568)
function old new delta
sk_stream_alloc_skb - 183 +183
ip_rt_init 529 525 -4
arp_ignore 112 107 -5
__inet_lookup_listener 284 274 -10
tcp_sendmsg 2583 2481 -102
tcp_sendpage 1449 1300 -149
tso_fragment 417 258 -159
tcp_fragment 1149 988 -161
__tcp_push_pending_frames 1998 1837 -161
Signed-off-by: Pavel Emelyanov <xemul@openvz.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
This function seems too big for inlining. Indeed, it saves
half-a-kilo when uninlined:
add/remove: 1/0 grow/shrink: 0/7 up/down: 195/-719 (-524)
function old new delta
sk_stream_alloc_pskb - 195 +195
ip_rt_init 529 525 -4
__inet_lookup_listener 284 274 -10
tcp_sendmsg 2583 2486 -97
tcp_sendpage 1449 1305 -144
tso_fragment 417 267 -150
tcp_fragment 1149 992 -157
__tcp_push_pending_frames 1998 1841 -157
Signed-off-by: Pavel Emelyanov <xemul@openvz.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
As done two years ago on IP route cache table (commit
22c047ccbc) , we can avoid using one
lock per hash bucket for the huge TCP/DCCP hash tables.
On a typical x86_64 platform, this saves about 2MB or 4MB of ram, for
litle performance differences. (we hit a different cache line for the
rwlock, but then the bucket cache line have a better sharing factor
among cpus, since we dirty it less often). For netstat or ss commands
that want a full scan of hash table, we perform fewer memory accesses.
Using a 'small' table of hashed rwlocks should be more than enough to
provide correct SMP concurrency between different buckets, without
using too much memory. Sizing of this table depends on
num_possible_cpus() and various CONFIG settings.
This patch provides some locking abstraction that may ease a future
work using a different model for TCP/DCCP table.
Signed-off-by: Eric Dumazet <dada1@cosmosbay.com>
Acked-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
On systems with a very large amount of memory, the heuristics in
alloc_large_system_hash() result in a very large TCP established hash
table: 16 millions of entries for a 128 GB ia64 system. This makes
reading from /proc/net/tcp pretty slow (well over a second) and as a
result netstat is slow on these machines. I know that /proc/net/tcp is
deprecated in favor of tcp_diag, however at the moment netstat only
knows of the former.
I am skeptical that such a large TCP established hash is often needed.
Just because a system has a lot of memory doesn't imply that it will
have several millions of concurrent TCP connections. Thus I believe
that we should put an arbitrary high limit to the size of the TCP
established hash by default. Users who really need a bigger hash can
always use the thash_entries boot parameter to get more.
I propose 2 millions of entries as the arbitrary high limit. This
makes /proc/net/tcp reasonably fast on the system in question (0.2 s)
while being still large enough for me to be confident that network
performance won't suffer.
This is just one way to limit the hash size, there are others; I am not
familiar enough with the TCP code to decide which is best. Thus, I
would welcome the proposals of alternatives.
[ 2 million is still too large, thus I've modified the limit in the
change to be '512 * 1024'. -DaveM ]
Signed-off-by: Jean Delvare <jdelvare@suse.de>
Signed-off-by: David S. Miller <davem@davemloft.net>
The task_struct->pid member is going to be deprecated, so start
using the helpers (task_pid_nr/task_pid_vnr/task_pid_nr_ns) in
the kernel.
The first thing to start with is the pid, printed to dmesg - in
this case we may safely use task_pid_nr(). Besides, printks produce
more (much more) than a half of all the explicit pid usage.
[akpm@linux-foundation.org: git-drm went and changed lots of stuff]
Signed-off-by: Pavel Emelyanov <xemul@openvz.org>
Cc: Dave Airlie <airlied@linux.ie>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Return some useful information such as the maximum listen backlog and
the current listen backlog in the tcp_info structure and
INET_DIAG_INFO.
Signed-off-by: Rick Jones <rick.jones2@hp.com>
Signed-off-by: Eric Dumazet <dada1@cosmosbay.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Hopefully captured all single statement cases under net/. I'm
not too sure if there is some policy about #includes that are
"guaranteed" (ie., in the current tree) to be available through
some other #included header, so I just added linux/kernel.h to
each changed file that didn't #include it previously.
Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>
Previously code had IsReno/IsFack defined as macros that were
local to tcp_input.c though sack_ok field has user elsewhere too
for the same purpose. This changes them to static inlines as
preferred according the current coding style and unifies the
access to sack_ok across multiple files. Magic bitops of sack_ok
for FACK and DSACK are also abstracted to functions with
appropriate names.
Note:
- One sack_ok = 1 remains but that's self explanary, i.e., it
enables sack
- Couple of !IsReno cases are changed to tcp_is_sack
- There were no users for IsDSack => I dropped it
Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>
As discovered by Evegniy Polyakov, if we try to sendmsg after
a connection reset, we can do incredibly stupid things.
The core issue is that inet_sendmsg() tries to autobind the
socket, but we should never do that for TCP. Instead we should
just go straight into TCP's sendmsg() code which will do all
of the necessary state and pending socket error checks.
TCP's sendpage already directly vectors to tcp_sendpage(), so this
merely brings sendmsg() in line with that.
Signed-off-by: David S. Miller <davem@davemloft.net>
Slab destructors were no longer supported after Christoph's
c59def9f22 change. They've been
BUGs for both slab and slub, and slob never supported them
either.
This rips out support for the dtor pointer from kmem_cache_create()
completely and fixes up every single callsite in the kernel (there were
about 224, not including the slab allocator definitions themselves,
or the documentation references).
Signed-off-by: Paul Mundt <lethal@linux-sh.org>
net/ipv4/tcp.c: In function 'tcp_recvmsg':
net/ipv4/tcp.c:1111: warning: unused variable 'available'
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Chris Leech <christopher.leech@intel.com>
The performance wins come with having the DMA copy engine doing the copies
in parallel with the context switch. If there is enough data ready on the
socket at recv time just use a regular copy.
Signed-off-by: Chris Leech <christopher.leech@intel.com>
tcp_read_sock() currently assumes that the recv_actor() only returns
number of bytes copied. For network splice receive, we may have to
return an error in some cases. So allow the actor to return a negative
error value.
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This diff changes the default port range used for outgoing connections,
from "use 32768-61000 in most cases, but use N-4999 on small boxes
(where N is a multiple of 1024, depending on just *how* small the box
is)" to just "use 32768-61000 in all cases".
I don't believe there are any drawbacks to this change, and it keeps
outgoing connection ports farther away from the mess of
IANA-registered ports.
Signed-off-by: Mark Glines <mark@glines.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
tcp_out_of_resources() and tcp_close() perform the
same checking of number of orphan sockets. Move this
code into common place.
Signed-off-by: Pavel Emelianov <xemul@openvz.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
Remove includes of <linux/smp_lock.h> where it is not used/needed.
Suggested by Al Viro.
Builds cleanly on x86_64, i386, alpha, ia64, powerpc, sparc,
sparc64, and arm (all 59 defconfigs).
Signed-off-by: Randy Dunlap <randy.dunlap@oracle.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
When the server drops its connection, NFS client reconnects using the
same socket after disconnecting. If the new connection's SYN,ACK
doesn't contain the TCP timestamp option and the old connection's did,
tp->tcp_header_len is recomputed assuming no timestamp header but
tp->rx_opt.tstamp_ok remains set. Then tcp_build_and_update_options()
adds in a timestamp option past the end of the allocated TCP header,
overwriting TCP data, or when the data is in skb_shinfo(skb)->frags[],
overwriting skb_shinfo(skb) causing a crash soon after. (The issue was
debugged from such a crash.)
Similarly, wscale_ok and sack_ok also get set based on the SYN,ACK
packet but not reset on disconnect, since they are zeroed out at
initialization. The patch zeroes out the entire tp->rx_opt struct in
tcp_disconnect() to avoid this sort of problem.
Signed-off-by: Srinivas Aji <Aji_Srinivas@emc.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This updates references to drafts in comments which must be about 10
years old. Internet draft draft-ietf-tcpimpl-prob-03.txt expired in 1998
and was replaced by RFC 2525 in March 1999.
Section 3.10 of the draft maps almost identically into section 2.17 of RFC
2525: both are entitled "Failure to RST on close with data pending", the
differences in text body amount to a typo and minor sentence change.
Signed-off-by: Gerrit Renker <gerrit@erg.abdn.ac.uk>
Signed-off-by: David S. Miller <davem@davemloft.net>
The function is quite big and has several call sites and nothing
to collapse by compiler optimization on inlining.
Besides it's nicer to read in a in .c file.
Signed-off-by: Andi Kleen <ak@suse.de>
Signed-off-by: David S. Miller <davem@davemloft.net>
Spring cleaning time...
There seems to be a lot of places in the network code that have
extra bogus semicolons after conditionals. Most commonly is a
bogus semicolon after: switch() { }
Signed-off-by: Stephen Hemminger <shemminger@linux-foundation.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
tcp_memory_pressure and tcp_socket currently share a cache line with tcp_memory_allocated, tcp_sockets_allocated.
(Very hot cache line)
It makes sense to declare these variables as __read_mostly, to avoid false sharing on SMP.
ffffffff8081d9c0 B tcp_orphan_count
ffffffff8081d9c4 B tcp_memory_allocated
ffffffff8081d9c8 B tcp_sockets_allocated
ffffffff8081d9cc B tcp_memory_pressure
ffffffff8081d9d0 b tcp_md5sig_users
ffffffff8081d9d8 b tcp_md5sig_pool
ffffffff8081d9e0 b warntime.31570
ffffffff8081d9e8 b tcp_socket
Signed-off-by: Eric Dumazet <dada1@cosmosbay.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
So that it is also an offset from skb->head, reduces its size from 8 to 4 bytes
on 64bit architectures, allowing us to combine the 4 bytes hole left by the
layer headers conversion, reducing struct sk_buff size to 256 bytes, i.e. 4
64byte cachelines, and since the sk_buff slab cache is SLAB_HWCACHE_ALIGN...
:-)
Many calculations that previously required that skb->{transport,network,
mac}_header be first converted to a pointer now can be done directly, being
meaningful as offsets or pointers.
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
For the places where we need a pointer to the transport header, it is
still legal to touch skb->h.raw directly if just adding to,
subtracting from or setting it to another layer header.
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This allows the write queue implementation to be changed,
for example, to one which allows fast interval searching.
Signed-off-by: David S. Miller <davem@davemloft.net>
Change tcp_mem initialization function. The fraction of total memory
is now a continuous function of memory size, and independent of page
size.
Signed-off-by: John Heffner <jheffner@psc.edu>
Signed-off-by: David S. Miller <davem@davemloft.net>
The locking calls assumed that these code paths were only
invoked in software interrupt context, but that isn't true.
Therefore we need to use spin_{lock,unlock}_bh() throughout.
Signed-off-by: David S. Miller <davem@davemloft.net>
ehash table layout is currently this one :
First half of this table is used by sockets not in TIME_WAIT state
Second half of it is used by sockets in TIME_WAIT state.
This is non optimal because of for a given hash or socket, the two chain heads
are located in separate cache lines.
Moreover the locks of the second half are never used.
If instead of this halving, we use two list heads in inet_ehash_bucket instead
of only one, we probably can avoid one cache miss, and reduce ram usage,
particularly if sizeof(rwlock_t) is big (various CONFIG_DEBUG_SPINLOCK,
CONFIG_DEBUG_LOCK_ALLOC settings). So we still halves the table but we keep
together related chains to speedup lookups and socket state change.
In this patch I did not try to align struct inet_ehash_bucket, but a future
patch could try to make this structure have a convenient size (a power of two
or a multiple of L1_CACHE_SIZE).
I guess rwlock will just vanish as soon as RCU is plugged into ehash :) , so
maybe we dont need to scratch our heads to align the bucket...
Note : In case struct inet_ehash_bucket is not a power of two, we could
probably change alloc_large_system_hash() (in case it use __get_free_pages())
to free the unused space. It currently allocates a big zone, but the last
quarter of it could be freed. Again, this should be a temporary 'problem'.
Patch tested on ipv4 tcp only, but should be OK for IPV6 and DCCP.
Signed-off-by: Eric Dumazet <dada1@cosmosbay.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
It should call tcp_free_md5sig_pool() not __tcp_free_md5sig_pool()
so that it does proper refcounting.
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch contains the following possible cleanups:
- make the following needlessly global functions statis:
- ipv4/tcp.c: __tcp_alloc_md5sig_pool()
- ipv4/tcp_ipv4.c: tcp_v4_reqsk_md5_lookup()
- ipv4/udplite.c: udplite_rcv()
- ipv4/udplite.c: udplite_err()
- make the following needlessly global structs static:
- ipv4/tcp_ipv4.c: tcp_request_sock_ipv4_ops
- ipv4/tcp_ipv4.c: tcp_sock_ipv4_specific
- ipv6/tcp_ipv6.c: tcp_request_sock_ipv6_ops
- net/ipv{4,6}/udplite.c: remove inline's from static functions
(gcc should know best when to inline them)
Signed-off-by: Adrian Bunk <bunk@stusta.de>
Signed-off-by: David S. Miller <davem@davemloft.net>
Fix up tcp_mem initial settings to take into account the size of the
hash entries (different on SMP and non-SMP systems).
Signed-off-by: John Heffner <jheffner@psc.edu>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch removes consideration of high memory when determining TCP
hash table sizes. Taking into account high memory results in tcp_mem
values that are too large.
Signed-off-by: John Heffner <jheffner@psc.edu>
Signed-off-by: David S. Miller <davem@davemloft.net>
It does not affect either mss-sized connections (obviously) or
connections controlled by Nagle (because there is only one small
segment in flight).
The idea is to record the fact that a small segment arrives on a
connection, where one small segment has already been received and
still not-ACKed. In this case ACK is forced after tcp_recvmsg() drains
receive buffer.
In other words, it is a "soft" each-2nd-segment ACK, which is enough
to preserve ACK clock even when ABC is enabled.
Signed-off-by: Alexey Kuznetsov <kuznet@ms2.inr.ac.ru>
Signed-off-by: David S. Miller <davem@davemloft.net>
Change net/core, ipv4 and ipv6 sysctl variables to __read_mostly.
Couldn't actually measure any performance increase while testing (.3%
I consider noise), but seems like the right thing to do.
Signed-off-by: Brian Haley <brian.haley@hp.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
