When ip_queue_xmit calls ip_select_ident_more for IP identity selection
it gives it the wrong packet count for TSO packets. The ip_select_*
functions expect one less than the number of packets, so we need to
subtract one for TSO packets.
This bug was diagnosed and fixed by Tom Young.
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
From: Jesper Juhl <jesper.juhl@gmail.com>
This is the net/ part of the big kfree cleanup patch.
Remove pointless checks for NULL prior to calling kfree() in net/.
Signed-off-by: Jesper Juhl <jesper.juhl@gmail.com>
Cc: "David S. Miller" <davem@davemloft.net>
Cc: Arnaldo Carvalho de Melo <acme@conectiva.com.br>
Acked-by: Marcel Holtmann <marcel@holtmann.org>
Acked-by: YOSHIFUJI Hideaki <yoshfuji@linux-ipv6.org>
Signed-off-by: Andrew Morton <akpm@osdl.org>
There was a fix in 2.6.13 that changed the behaviour of
ip_vs_conn_expire_now function not to put reference to connection,
its callers should hold write lock or connection refcnt. But we
forgot to convert one caller, when the real server for connection
is unavailable caller should put the connection reference. It
happens only when sysctl var expire_nodest_conn is set to 1 and
such connections never expire. Thanks to Roberto Nibali who found
the problem and tested a 2.4.32-rc2 patch, which is equal to this
2.6 version. Patch for 2.4 is already sent to Marcelo.
Signed-off-by: Julian Anastasov <ja@ssi.bg>
Signed-off-by: Roberto Nibali <ratz@drugphish.ch>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch randomizes the port selected on bind() for connections
to help with possible security attacks. It should also be faster
in most cases because there is no need for a global lock.
Signed-off-by: Stephen Hemminger <shemminger@osdl.org>
Signed-off-by: Arnaldo Carvalho de Melo <acme@mandriva.com>
There's a missing dependency from the CONNMARK target to ip_conntrack.
Signed-off-by: Pablo Neira Ayuso <pablo@eurodev.net>
Signed-off-by: Harald Welte <laforge@netfilter.org>
Signed-off-by: Arnaldo Carvalho de Melo <acme@mandriva.com>
It's not necessary to free skb if netlink_unicast() failed.
Signed-off-by: Yasuyuki Kozakai <yasuyuki.kozakai@toshiba.co.jp>
Signed-off-by: Harald Welte <laforge@netfilter.org>
Signed-off-by: Arnaldo Carvalho de Melo <acme@mandriva.com>
The unknown protocol is used as a fallback when a protocol isn't known.
Hence we cannot handle it failing, so don't set ".me". It's OK, since we
only grab a reference from within the same module (iptable_nat.ko), so we
never take the module refcount from 0 to 1.
Also, remove the "protocol is NULL" test: it's never NULL.
Signed-off-by: Rusty Rusty <rusty@rustcorp.com.au>
Signed-off-by: Harald Welte <laforge@netfilter.org>
Signed-off-by: Arnaldo Carvalho de Melo <acme@mandriva.com>
This endianness bug slipped through while changing the 'gre.key' field in the
conntrack tuple from 32bit to 16bit.
None of my tests caught the problem, since the linux pptp client always has
'0' as call id / gre key. Only windows clients actually trigger the bug.
Signed-off-by: Harald Welte <laforge@netfilter.org>
Signed-off-by: Arnaldo Carvalho de Melo <acme@mandriva.com>
This patch fixes compilation of the PPTP conntrack helper when NAT is
configured off.
Signed-off-by: Yasuyuki Kozakai <yasuyuki.kozakai@toshiba.co.jp>
Signed-off-by: Harald Welte <laforge@netfilter.org>
Signed-off-by: Arnaldo Carvalho de Melo <acme@mandriva.com>
The max growth of BIC TCP is too large. Original code was based on
BIC 1.0 and the default there was 32. Later code (2.6.13) included
compensation for delayed acks, and should have reduced the default
value to 16; since normally TCP gets one ack for every two packets sent.
The current value of 32 makes BIC too aggressive and unfair to other
flows.
Submitted-by: Injong Rhee <rhee@eos.ncsu.edu>
Signed-off-by: Stephen Hemminger <shemminger@osdl.org>
Acked-by: Ian McDonald <imcdnzl@gmail.com>
Signed-off-by: Arnaldo Carvalho de Melo <acme@mandriva.com>
And filter mode is exclude.
Further explanation by David Stevens:
Multicast source filters aren't widely used yet, and that's really the only
feature that's affected if an application actually exercises this bug, as far
as I can tell. An ordinary filter-less multicast join should still work, and
only forwarded multicast traffic making use of filters and doing empty-source
filters with the MSFILTER ioctl would be at risk of not getting multicast
traffic forwarded to them because the reports generated would not be based on
the correct counts.
Signed-off-by: Yan Zheng <yanzheng@21cn.com
Acked-by: David L Stevens <dlstevens@us.ibm.com>
Signed-off-by: Arnaldo Carvalho de Melo <acme@mandriva.com>
Like ip_tables already has it for some time, this adds support for
having multiple revisions for each match/target. We steal one byte from
the name in order to accomodate a 8 bit version number.
Signed-off-by: Harald Welte <laforge@netfilter.org>
Signed-off-by: Arnaldo Carvalho de Melo <acme@mandriva.com>
Typo fix: dots appearing after a newline in printk strings.
Signed-off-by: Jean Delvare <khali@linux-fr.org>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
fib_del_ifaddr() dereferences ifa->ifa_dev, so the code already assumes that
ifa->ifa_dev is non-NULL, the check is unnecessary.
Signed-off-by: Jayachandran C. <c.jayachandran at gmail.com>
Signed-off-by: Arnaldo Carvalho de Melo <acme@mandriva.com>
Attached is kernel patch for UDP Fragmentation Offload (UFO) feature.
1. This patch incorporate the review comments by Jeff Garzik.
2. Renamed USO as UFO (UDP Fragmentation Offload)
3. udp sendfile support with UFO
This patches uses scatter-gather feature of skb to generate large UDP
datagram. Below is a "how-to" on changes required in network device
driver to use the UFO interface.
UDP Fragmentation Offload (UFO) Interface:
-------------------------------------------
UFO is a feature wherein the Linux kernel network stack will offload the
IP fragmentation functionality of large UDP datagram to hardware. This
will reduce the overhead of stack in fragmenting the large UDP datagram to
MTU sized packets
1) Drivers indicate their capability of UFO using
dev->features |= NETIF_F_UFO | NETIF_F_HW_CSUM | NETIF_F_SG
NETIF_F_HW_CSUM is required for UFO over ipv6.
2) UFO packet will be submitted for transmission using driver xmit routine.
UFO packet will have a non-zero value for
"skb_shinfo(skb)->ufo_size"
skb_shinfo(skb)->ufo_size will indicate the length of data part in each IP
fragment going out of the adapter after IP fragmentation by hardware.
skb->data will contain MAC/IP/UDP header and skb_shinfo(skb)->frags[]
contains the data payload. The skb->ip_summed will be set to CHECKSUM_HW
indicating that hardware has to do checksum calculation. Hardware should
compute the UDP checksum of complete datagram and also ip header checksum of
each fragmented IP packet.
For IPV6 the UFO provides the fragment identification-id in
skb_shinfo(skb)->ip6_frag_id. The adapter should use this ID for generating
IPv6 fragments.
Signed-off-by: Ananda Raju <ananda.raju@neterion.com>
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au> (forwarded)
Signed-off-by: Arnaldo Carvalho de Melo <acme@mandriva.com>
This bug is responsible for causing the infamous "Treason uncloaked"
messages that's been popping up everywhere since the printk was added.
It has usually been blamed on foreign operating systems. However,
some of those reports implicate Linux as both systems are running
Linux or the TCP connection is going across the loopback interface.
In fact, there really is a bug in the Linux TCP header prediction code
that's been there since at least 2.1.8. This bug was tracked down with
help from Dale Blount.
The effect of this bug ranges from harmless "Treason uncloaked"
messages to hung/aborted TCP connections. The details of the bug
and fix is as follows.
When snd_wnd is updated, we only update pred_flags if
tcp_fast_path_check succeeds. When it fails (for example,
when our rcvbuf is used up), we will leave pred_flags with
an out-of-date snd_wnd value.
When the out-of-date pred_flags happens to match the next incoming
packet we will again hit the fast path and use the current snd_wnd
which will be wrong.
In the case of the treason messages, it just happens that the snd_wnd
cached in pred_flags is zero while tp->snd_wnd is non-zero. Therefore
when a zero-window packet comes in we incorrectly conclude that the
window is non-zero.
In fact if the peer continues to send us zero-window pure ACKs we
will continue making the same mistake. It's only when the peer
transmits a zero-window packet with data attached that we get a
chance to snap out of it. This is what triggers the treason
message at the next retransmit timeout.
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: Arnaldo Carvalho de Melo <acme@mandriva.com>
Fix setting of the broadcast address when the netmask is set via
SIOCSIFNETMASK in Linux 2.6. The code wanted the old value of
ifa->ifa_mask but used it after it had already been overwritten with
the new value.
Signed-off-by: David Engel <gigem@comcast.net>
Signed-off-by: Arnaldo Carvalho de Melo <acme@mandriva.com>
skb_prev is assigned from skb, which cannot be NULL. This patch removes the
unnecessary NULL check.
Signed-off-by: Jayachandran C. <c.jayachandran at gmail.com>
Acked-by: James Morris <jmorris@namei.org>
Signed-off-by: Arnaldo Carvalho de Melo <acme@mandriva.com>
This patch kills a redundant rcu_dereference on fa->fa_info in fib_trie.c.
As this dereference directly follows a list_for_each_entry_rcu line, we
have already taken a read barrier with respect to getting an entry from
the list.
This read barrier guarantees that all values read out of fa are valid.
In particular, the contents of structure pointed to by fa->fa_info is
initialised before fa->fa_info is actually set (see fn_trie_insert);
the setting of fa->fa_info itself is further separated with a write
barrier from the insertion of fa into the list.
Therefore by taking a read barrier after obtaining fa from the list
(which is given by list_for_each_entry_rcu), we can be sure that
fa->fa_info contains a valid pointer, as well as the fact that the
data pointed to by fa->fa_info is itself valid.
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Acked-by: Paul E. McKenney <paulmck@us.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Arnaldo Carvalho de Melo <acme@mandriva.com>
It's fairly simple to resize the hash table, but currently you need to
remove and reinsert the module. That's bad (we lose connection
state). Harald has even offered to write a daemon which sets this
based on load.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Signed-off-by: Harald Welte <laforge@netfilter.org>
Signed-off-by: Arnaldo Carvalho de Melo <acme@mandriva.com>
In 'net' change the explicit use of for-loops and NR_CPUS into the
general for_each_cpu() or for_each_online_cpu() constructs, as
appropriate. This widens the scope of potential future optimizations
of the general constructs, as well as takes advantage of the existing
optimizations of first_cpu() and next_cpu(), which is advantageous
when the true CPU count is much smaller than NR_CPUS.
Signed-off-by: John Hawkes <hawkes@sgi.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Arnaldo Carvalho de Melo <acme@mandriva.com>
IPVS used flag NFC_IPVS_PROPERTY in nfcache but as now nfcache was removed the
new flag 'ipvs_property' still needs to be copied. This patch should be
included in 2.6.14.
Further comments from Harald Welte:
Sorry, seems like the bug was introduced by me.
Signed-off-by: Julian Anastasov <ja@ssi.bg>
Signed-off-by: Harald Welte <laforge@netfilter.org>
Signed-off-by: Arnaldo Carvalho de Melo <acme@mandriva.com>
It is legitimate to call tcp_fragment with len == skb->len since
that is done for FIN packets and the FIN flag counts as one byte.
So we should only check for the len > skb->len case.
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: Arnaldo Carvalho de Melo <acme@mandriva.com>
Original patch by Harald Welte, with feedback from Herbert Xu
and testing by Sbastien Bernard.
EBTABLES, ARP tables, and IP/IP6 tables all assume that cpus
are numbered linearly. That is not necessarily true.
This patch fixes that up by calculating the largest possible
cpu number, and allocating enough per-cpu structure space given
that.
Signed-off-by: David S. Miller <davem@davemloft.net>
This is the second report of this bug. Unfortunately the first
reporter hasn't been able to reproduce it since to provide more
debugging info.
So let's apply this patch for 2.6.14 to
1) Make this non-fatal.
2) Provide the info we need to track it down.
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
This is required to avoid unloading a module that has active timewait
sockets, such as DCCP.
Signed-off-by: Arnaldo Carvalho de Melo <acme@ghostprotocols.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch add support to change the state of the private protocol
information via conntrack_netlink.
Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>
Signed-off-by: Harald Welte <laforge@netfilter.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch adds the ability of changing the state a TCP connection. I know
that this must be used with care but it's required to provide a complete
conntrack creation via conntrack_netlink. So I'll document this aspect on
the upcoming docs.
Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>
Signed-off-by: Harald Welte <laforge@netfilter.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
Initially we used 64bit counters for conntrack-based accounting, since we
had no event mechanism to tell userspace that our counters are about to
overflow. With nfnetlink_conntrack, we now have such a event mechanism and
thus can save 16bytes per connection.
Signed-off-by: Harald Welte <laforge@netfilter.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch fixes the following bugs in ESP:
* Fix transport mode MTU overestimate. This means that the inner MTU
is smaller than it needs be. Worse yet, given an input MTU which
is a multiple of 4 it will always produce an estimate which is not
a multiple of 4.
For example, given a standard ESP/3DES/MD5 transform and an MTU of
1500, the resulting MTU for transport mode is 1462 when it should
be 1464.
The reason for this is because IP header lengths are always a multiple
of 4 for IPv4 and 8 for IPv6.
* Ensure that the block size is at least 4. This is required by RFC2406
and corresponds to what the esp_output function does. At the moment
this only affects crypto_null as its block size is 1.
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch uses the macro ALIGN in all the applicable spots for ESP.
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
To keep consistency, the TCP private protocol information is nested
attributes under CTA_PROTOINFO_TCP. This way the sequence of attributes to
access the TCP state information looks like here below:
CTA_PROTOINFO
CTA_PROTOINFO_TCP
CTA_PROTOINFO_TCP_STATE
instead of:
CTA_PROTOINFO
CTA_PROTOINFO_TCP_STATE
Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>
Signed-off-by: Harald Welte <laforge@netfilter.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
The ID is only required by ICMP type 8 (echo), so it's not
mandatory for all sort of ICMP connections. This patch makes
mandatory only the type and the code for ICMP netlink messages.
Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>
Signed-off-by: Harald Welte <laforge@netfilter.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
When we send "status" from userspace, we forget to convert the endianness.
This patch adds the reqired conversion. Thanks to Pablo Neira for
discovering this.
Signed-off-by: Harald Welte <laforge@netfilter.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
Similar to nfnetlink_queue and ip_queue, we mark ipt_ULOG as obsolete.
This should have been part of the original nfnetlink_log merge, but
I somehow missed it.
Signed-off-by: Harald Welte <laforge@netfilter.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
PPTP should not be selectable without conntrack enabled
Signed-off-by: Harald Welte <laforge@netfilter.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
- added typedef unsigned int __nocast gfp_t;
- replaced __nocast uses for gfp flags with gfp_t - it gives exactly
the same warnings as far as sparse is concerned, doesn't change
generated code (from gcc point of view we replaced unsigned int with
typedef) and documents what's going on far better.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Missing parenthesis in causes BIC to be slow in increasing congestion
window.
Spotted by Injong Rhee.
Signed-off-by: Stephen Hemminger <shemminger@osdl.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
From: Randy Dunlap <rdunlap@xenotime.net>
Fix implicit nocast warnings in ip_vs code:
net/ipv4/ipvs/ip_vs_app.c:631:54: warning: implicit cast to nocast type
Signed-off-by: Randy Dunlap <rdunlap@xenotime.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
The patch below introduces special thresholds to keep root node in the trie
large. This gives a flatter tree at the cost of a modest memory increase.
Overall it seems to be gain and this was also proposed by one the authors
of the paper in recent a seminar.
Main table after loading 123 k routes.
Aver depth: 3.30
Max depth: 9
Root-node size 12 bits
Total size: 4044 kB
With the patch:
Aver depth: 2.78
Max depth: 8
Root-node size 15 bits
Total size: 4150 kB
An increase of 8-10% was seen in forwading performance for an rDoS attack.
Signed-off-by: Robert Olsson <robert.olsson@its.uu.se>
Signed-off-by: David S. Miller <davem@davemloft.net>
The following patch renames __in_dev_get() to __in_dev_get_rtnl() and
introduces __in_dev_get_rcu() to cover the second case.
1) RCU with refcnt should use in_dev_get().
2) RCU without refcnt should use __in_dev_get_rcu().
3) All others must hold RTNL and use __in_dev_get_rtnl().
There is one exception in net/ipv4/route.c which is in fact a pre-existing
race condition. I've marked it as such so that we remember to fix it.
This patch is based on suggestions and prior work by Suzanne Wood and
Paul McKenney.
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
Meelis Roos <mroos@linux.ee> wrote:
> RK> My firewall setup relies on proxyarp working. However, with 2.6.14-rc3,
> RK> it appears to be completely broken. The firewall is 212.18.232.186,
>
> Same here with some kernel between 14-rc2 and 14-rc3 - no reposnse to
> ARP on a proxyarp gateway. Sorry, no exact revison and no more debugging
> yet since it'a a production gateway.
The breakage is caused by the change to use the CB area for flagging
whether a packet has been queued due to proxy_delay. This area gets
cleared every time arp_rcv gets called. Unfortunately packets delayed
due to proxy_delay also go through arp_rcv when they are reprocessed.
In fact, I can't think of a reason why delayed proxy packets should go
through netfilter again at all. So the easiest solution is to bypass
that and go straight to arp_process.
This is essentially what would've happened before netfilter support
was added to ARP.
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
Arnaldo and I agreed it could be applied now, because I have other
pending patches depending on this one (Thank you Arnaldo)
(The other important patch moves skc_refcnt in a separate cache line,
so that the SMP/NUMA performance doesnt suffer from cache line ping pongs)
1) First some performance data :
--------------------------------
tcp_v4_rcv() wastes a *lot* of time in __inet_lookup_established()
The most time critical code is :
sk_for_each(sk, node, &head->chain) {
if (INET_MATCH(sk, acookie, saddr, daddr, ports, dif))
goto hit; /* You sunk my battleship! */
}
The sk_for_each() does use prefetch() hints but only the begining of
"struct sock" is prefetched.
As INET_MATCH first comparison uses inet_sk(__sk)->daddr, wich is far
away from the begining of "struct sock", it has to bring into CPU
cache cold cache line. Each iteration has to use at least 2 cache
lines.
This can be problematic if some chains are very long.
2) The goal
-----------
The idea I had is to change things so that INET_MATCH() may return
FALSE in 99% of cases only using the data already in the CPU cache,
using one cache line per iteration.
3) Description of the patch
---------------------------
Adds a new 'unsigned int skc_hash' field in 'struct sock_common',
filling a 32 bits hole on 64 bits platform.
struct sock_common {
unsigned short skc_family;
volatile unsigned char skc_state;
unsigned char skc_reuse;
int skc_bound_dev_if;
struct hlist_node skc_node;
struct hlist_node skc_bind_node;
atomic_t skc_refcnt;
+ unsigned int skc_hash;
struct proto *skc_prot;
};
Store in this 32 bits field the full hash, not masked by (ehash_size -
1) Using this full hash as the first comparison done in INET_MATCH
permits us immediatly skip the element without touching a second cache
line in case of a miss.
Suppress the sk_hashent/tw_hashent fields since skc_hash (aliased to
sk_hash and tw_hash) already contains the slot number if we mask with
(ehash_size - 1)
File include/net/inet_hashtables.h
64 bits platforms :
#define INET_MATCH(__sk, __hash, __cookie, __saddr, __daddr, __ports, __dif)\
(((__sk)->sk_hash == (__hash))
((*((__u64 *)&(inet_sk(__sk)->daddr)))== (__cookie)) && \
((*((__u32 *)&(inet_sk(__sk)->dport))) == (__ports)) && \
(!((__sk)->sk_bound_dev_if) || ((__sk)->sk_bound_dev_if == (__dif))))
32bits platforms:
#define TCP_IPV4_MATCH(__sk, __hash, __cookie, __saddr, __daddr, __ports, __dif)\
(((__sk)->sk_hash == (__hash)) && \
(inet_sk(__sk)->daddr == (__saddr)) && \
(inet_sk(__sk)->rcv_saddr == (__daddr)) && \
(!((__sk)->sk_bound_dev_if) || ((__sk)->sk_bound_dev_if == (__dif))))
- Adds a prefetch(head->chain.first) in
__inet_lookup_established()/__tcp_v4_check_established() and
__inet6_lookup_established()/__tcp_v6_check_established() and
__dccp_v4_check_established() to bring into cache the first element of the
list, before the {read|write}_lock(&head->lock);
Signed-off-by: Eric Dumazet <dada1@cosmosbay.com>
Acked-by: Arnaldo Carvalho de Melo <acme@ghostprotocols.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
I've found the problem in general. It affects any 64-bit
architecture. The problem occurs when you change the system time.
Suppose that when you boot your system clock is forward by a day.
This gets recorded down in skb_tv_base. You then wind the clock back
by a day. From that point onwards the offset will be negative which
essentially overflows the 32-bit variables they're stored in.
In fact, why don't we just store the real time stamp in those 32-bit
variables? After all, we're not going to overflow for quite a while
yet.
When we do overflow, we'll need a better solution of course.
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
From: Alexey Kuznetsov <kuznet@ms2.inr.ac.ru>
Handle better the case where the sender sends full sized
frames initially, then moves to a mode where it trickles
out small amounts of data at a time.
This known problem is even mentioned in the comments
above tcp_grow_window() in tcp_input.c, specifically:
...
* The scheme does not work when sender sends good segments opening
* window and then starts to feed us spagetti. But it should work
* in common situations. Otherwise, we have to rely on queue collapsing.
...
When the sender gives full sized frames, the "struct sk_buff" overhead
from each packet is small. So we'll advertize a larger window.
If the sender moves to a mode where small segments are sent, this
ratio becomes tilted to the other extreme and we start overrunning
the socket buffer space.
tcp_clamp_window() tries to address this, but it's clamping of
tp->window_clamp is a wee bit too aggressive for this particular case.
Fix confirmed by Ion Badulescu.
Signed-off-by: David S. Miller <davem@davemloft.net>
But retain the comment fix.
Alexey Kuznetsov has explained the situation as follows:
--------------------
I think the fix is incorrect. Look, the RFC function init_cwnd(mss) is
not continuous: f.e. for mss=1095 it needs initial window 1095*4, but
for mss=1096 it is 1096*3. We do not know exactly what mss sender used
for calculations. If we advertised 1096 (and calculate initial window
3*1096), the sender could limit it to some value < 1096 and then it
will need window his_mss*4 > 3*1096 to send initial burst.
See?
So, the honest function for inital rcv_wnd derived from
tcp_init_cwnd() is:
init_rcv_wnd(mss)=
min { init_cwnd(mss1)*mss1 for mss1 <= mss }
It is something sort of:
if (mss < 1096)
return mss*4;
if (mss < 1096*2)
return 1096*4;
return mss*2;
(I just scrablled a graph of piece of paper, it is difficult to see or
to explain without this)
I selected it differently giving more window than it is strictly
required. Initial receive window must be large enough to allow sender
following to the rfc (or just setting initial cwnd to 2) to send
initial burst. But besides that it is arbitrary, so I decided to give
slack space of one segment.
Actually, the logic was:
If mss is low/normal (<=ethernet), set window to receive more than
initial burst allowed by rfc under the worst conditions
i.e. mss*4. This gives slack space of 1 segment for ethernet frames.
For msses slighlty more than ethernet frame, take 3. Try to give slack
space of 1 frame again.
If mss is huge, force 2*mss. No slack space.
Value 1460*3 is really confusing. Minimal one is 1096*2, but besides
that it is an arbitrary value. It was meant to be ~4096. 1460*3 is
just the magic number from RFC, 1460*3 = 1095*4 is the magic :-), so
that I guess hands typed this themselves.
--------------------
Signed-off-by: David S. Miller <davem@davemloft.net>
Herbert Xu