In the current TSO implementation, NETIF_F_TSO and ECN cannot be
turned on together in a TCP connection. The problem is that most
hardware that supports TSO does not handle CWR correctly if it is set
in the TSO packet. Correct handling requires CWR to be set in the
first packet only if it is set in the TSO header.
This patch adds the ability to turn on NETIF_F_TSO and ECN using
GSO if necessary to handle TSO packets with CWR set. Hardware
that handles CWR correctly can turn on NETIF_F_TSO_ECN in the dev->
features flag.
All TSO packets with CWR set will have the SKB_GSO_TCPV4_ECN set. If
the output device does not have the NETIF_F_TSO_ECN feature set, GSO
will split the packet up correctly with CWR only set in the first
segment.
With help from Herbert Xu <herbert@gondor.apana.org.au>.
Since ECN can always be enabled with TSO, the SOCK_NO_LARGESEND sock
flag is completely removed.
Signed-off-by: Michael Chan <mchan@broadcom.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Having separate fields in sk_buff for TSO/UFO (tso_size/ufo_size) is not
going to scale if we add any more segmentation methods (e.g., DCCP). So
let's merge them.
They were used to tell the protocol of a packet. This function has been
subsumed by the new gso_type field. This is essentially a set of netdev
feature bits (shifted by 16 bits) that are required to process a specific
skb. As such it's easy to tell whether a given device can process a GSO
skb: you just have to and the gso_type field and the netdev's features
field.
I've made gso_type a conjunction. The idea is that you have a base type
(e.g., SKB_GSO_TCPV4) that can be modified further to support new features.
For example, if we add a hardware TSO type that supports ECN, they would
declare NETIF_F_TSO | NETIF_F_TSO_ECN. All TSO packets with CWR set would
have a gso_type of SKB_GSO_TCPV4 | SKB_GSO_TCPV4_ECN while all other TSO
packets would be SKB_GSO_TCPV4. This means that only the CWR packets need
to be emulated in software.
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
A lot of people have asked for a way to disable tcp_cwnd_restart(),
and it seems reasonable to add a sysctl to do that.
Signed-off-by: David S. Miller <davem@davemloft.net>
Trimming the head of an skb by calling skb_pull can cause the packet
to become unaligned if the length pulled is odd. Since the length is
entirely arbitrary for a FIN packet carrying data, this is actually
quite common.
Unaligned data is not the end of the world, but we should avoid it if
it's easily done. In this case it is trivial. Since we're discarding
all of the head data it doesn't matter whether we move skb->data forward
or back.
However, it is still possible to have unaligned skb->data in general.
So network drivers should be prepared to handle it instead of crashing.
This patch also adds an unlikely marking on len < headlen since partial
ACKs on head data are extremely rare in the wild. As the return value
of __pskb_trim_head is no longer ever NULL that has been removed.
Signed-off-by: Herbert Xu ~{PmV>HI~} <herbert@gondor.apana.org.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
The following unlikely should be replaced by likely because the
condition happens every time unless there is a hard error to transmit
a packet.
Signed-off-by: Hua Zhong <hzhong@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Make sure that we get the full sizeof(struct sk_buff)
plus the data size accounted for in skb->truesize.
This will create invariants that will allow adding
assertion checks on skb->truesize.
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
There is a problem with the TSO packet trimming code. The cause of
this lies in the tcp_fragment() function.
When we allocate a fragment for a completely non-linear packet the
truesize is calculated for a payload length of zero. This means that
truesize could in fact be less than the real payload length.
When that happens the TSO packet trimming can cause truesize to become
negative. This in turn can cause sk_forward_alloc to be -n * PAGE_SIZE
which would trigger the warning.
I've copied the code DaveM used in tso_fragment which should work here.
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch contains the following possible cleanups:
- make the following needlessly global function static:
- arp.c: arp_rcv()
- remove the following unused EXPORT_SYMBOL's:
- devinet.c: devinet_ioctl
- fib_frontend.c: ip_rt_ioctl
- inet_hashtables.c: inet_bind_bucket_create
- inet_hashtables.c: inet_bind_hash
- tcp_input.c: sysctl_tcp_abc
- tcp_ipv4.c: sysctl_tcp_tw_reuse
- tcp_output.c: sysctl_tcp_mtu_probing
- tcp_output.c: sysctl_tcp_base_mss
Signed-off-by: Adrian Bunk <bunk@stusta.de>
Signed-off-by: David S. Miller <davem@davemloft.net>
Back in the dark ages, we had to be conservative and only allow 15-bit
window fields if the window scale option was not negotiated. Some
ancient stacks used a signed 16-bit quantity for the window field of
the TCP header and would get confused.
Those days are long gone, so we can use the full 16-bits by default
now.
There is a sysctl added so that we can still interact with such old
stacks
Signed-off-by: Rick Jones <rick.jones2@hp.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This moves some TCP-specific MTU probing state out of
inet_connection_sock back to tcp_sock.
Signed-off-by: John Heffner <jheffner@psc.edu>
Signed-off-by: David S. Miller <davem@davemloft.net>
Implementation of packetization layer path mtu discovery for TCP, based on
the internet-draft currently found at
<http://www.ietf.org/internet-drafts/draft-ietf-pmtud-method-05.txt>.
Signed-off-by: John Heffner <jheffner@psc.edu>
Signed-off-by: David S. Miller <davem@davemloft.net>
That's >= a full sized TSO frame, so we should always
return 0 in that case.
Based upon a report and initial patch from Lachlan
Andrew, final patch suggested by Herbert Xu.
Signed-off-by: David S. Miller <davem@davemloft.net>
TCP inline usage cleanup:
* get rid of inline in several places
* replace __inline__ with inline where possible
* move functions used in one file out of tcp.h
* let compiler decide on used once cases
On x86_64:
text data bss dec hex filename
3594701 648348 567400 4810449 4966d1 vmlinux.orig
3593133 648580 567400 4809113 496199 vmlinux
On sparc64:
text data bss dec hex filename
2538278 406152 530392 3474822 350586 vmlinux.ORIG
2536382 406384 530392 3473158 34ff06 vmlinux
Signed-off-by: Stephen Hemminger <shemminger@osdl.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
As DCCP needs to be called in the same spots.
Now we have a member in inet_sock (is_icsk), set at sock creation time from
struct inet_protosw->flags (if INET_PROTOSW_ICSK is set, like for TCP and
DCCP) to see if a struct sock instance is a inet_connection_sock for places
like the ones in ip_sockglue.c (v4 and v6) where we previously were looking if
sk_type was SOCK_STREAM, that is insufficient because we now use the same code
for DCCP, that has sk_type SOCK_DCCP.
Signed-off-by: Arnaldo Carvalho de Melo <acme@mandriva.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
And move it to struct inet_connection_sock. DCCP will use it in the
upcoming changesets.
Signed-off-by: Arnaldo Carvalho de Melo <acme@mandriva.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
We have to store the congestion control timestamp on the SKB before we
clone it, not after. Else we get no timestamping information at all.
tcp_transmit_skb() has been reworked so that we can do the timestamp
still in one spot, instead of at all the call sites.
Problem discovered, and initial fix, from Tom Young
<tyo@ee.unimelb.edu.au>.
Signed-off-by: David S. Miller <davem@davemloft.net>
Use "hints" to speed up the SACK processing. Various forms
of this have been used by TCP developers (Web100, STCP, BIC)
to avoid the 2x linear search of outstanding segments.
Signed-off-by: Stephen Hemminger <shemminger@osdl.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
TCP peformance with TSO over networks with delay is awful.
On a 100Mbit link with 150ms delay, we get 4Mbits/sec with TSO and
50Mbits/sec without TSO.
The problem is with TSO, we intentionally do not keep the maximum
number of packets in flight to fill the window, we hold out to until
we can send a MSS chunk. But, we also don't update the congestion window
unless we have filled, as per RFC2861.
This patch replaces the check for the congestion window being full
with something smarter that accounts for TSO.
Signed-off-by: Stephen Hemminger <shemminger@osdl.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
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>
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>
- 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>
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>
Since the introduction of TSO pcount a year ago, it has been possible
for tcp_fragment() to cause packets_out to decrease. Prior to that,
tcp_retrans_try_collapse() was the only way for that to happen on the
retransmission path.
When this happens with Reno, it is possible for sasked_out to become
invalid because it is only an estimate and not tied to any particular
packet on the retransmission queue.
Therefore we need to adjust sacked_out as well as left_out in the Reno
case. The following patch does exactly that.
This bug is pretty difficult to trigger in practice though since you
need a SACKless peer with a retransmission that occurs just as the
cached MTU value expires.
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
The problem is that we're now calling tcp_fragment() in a context
where the packets might be marked as SACKED_ACKED or SACKED_RETRANS.
This was not possible before as you never retransmitted packets that
are so marked.
Because of this, we need to adjust sacked_out and retrans_out in
tcp_fragment(). This is exactly what the following patch does.
We also need to preserve the SACKED_ACKED/SACKED_RETRANS marking
if they exist.
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
The problem is that the SACK fragmenting code may incorrectly call
tcp_fragment() with a length larger than the skb->len. This happens
when the skb on the transmit queue completely falls to the LHS of the
SACK.
And add a BUG() check to tcp_fragment() so we can spot this kind of
error more quickly in the future.
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
There is an extra left_out/lost_out adjustment in tcp_fragment which
means that the lost_out accounting is always wrong. This patch removes
that chunk of code.
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
All we need to do is resegment the queue so that
we record SACK information accurately. The edges
of the SACK blocks guide our resegmenting decisions.
With help from Herbert Xu.
Signed-off-by: David S. Miller <davem@davemloft.net>
Protocols that make extensive use of SKB cloning,
for example TCP, eat at least 2 allocations per
packet sent as a result.
To cut the kmalloc() count in half, we implement
a pre-allocation scheme wherein we allocate
2 sk_buff objects in advance, then use a simple
reference count to free up the memory at the
correct time.
Based upon an initial patch by Thomas Graf and
suggestions from Herbert Xu.
Signed-off-by: David S. Miller <davem@davemloft.net>
This changeset basically moves tcp_sk()->{ca_ops,ca_state,etc} to inet_csk(),
minimal renaming/moving done in this changeset to ease review.
Most of it is just changes of struct tcp_sock * to struct sock * parameters.
With this we move to a state closer to two interesting goals:
1. Generalisation of net/ipv4/tcp_diag.c, becoming inet_diag.c, being used
for any INET transport protocol that has struct inet_hashinfo and are
derived from struct inet_connection_sock. Keeps the userspace API, that will
just not display DCCP sockets, while newer versions of tools can support
DCCP.
2. INET generic transport pluggable Congestion Avoidance infrastructure, using
the current TCP CA infrastructure with DCCP.
Signed-off-by: Arnaldo Carvalho de Melo <acme@mandriva.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Completing the previous changeset, this also generalises tcp_v4_synq_add,
renaming it to inet_csk_reqsk_queue_hash_add, already geing used in the
DCCP tree, which I plan to merge RSN.
Signed-off-by: Arnaldo Carvalho de Melo <acme@ghostprotocols.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
This creates struct inet_connection_sock, moving members out of struct
tcp_sock that are shareable with other INET connection oriented
protocols, such as DCCP, that in my private tree already uses most of
these members.
The functions that operate on these members were renamed, using a
inet_csk_ prefix while not being moved yet to a new file, so as to
ease the review of these changes.
Signed-off-by: Arnaldo Carvalho de Melo <acme@ghostprotocols.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
Remove the "list" member of struct sk_buff, as it is entirely
redundant. All SKB list removal callers know which list the
SKB is on, so storing this in sk_buff does nothing other than
taking up some space.
Two tricky bits were SCTP, which I took care of, and two ATM
drivers which Francois Romieu <romieu@fr.zoreil.com> fixed
up.
Signed-off-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Francois Romieu <romieu@fr.zoreil.com>
If the tail SKB fits into the window, it is still
benefitical to defer until the goal percentage of
the window is available. This give the application
time to feed more data into the send queue and thus
results in larger TSO frames going out.
Patch from Dmitry Yusupov <dima@neterion.com>.
Signed-off-by: David S. Miller <davem@davemloft.net>
1) We send out a normal sized packet with TSO on to start off.
2) ICMP is received indicating a smaller MTU.
3) We send the current sk_send_head which needs to be fragmented
since it was created before the ICMP event. The first fragment
is then sent out.
At this point the remaining fragment is allocated by tcp_fragment.
However, its size is padded to fit the L1 cache-line size therefore
creating tail-room up to 124 bytes long.
This fragment will also be sitting at sk_send_head.
4) tcp_sendmsg is called again and it stores data in the tail-room of
of the fragment.
5) tcp_push_one is called by tcp_sendmsg which then calls tso_fragment
since the packet as a whole exceeds the MTU.
At this point we have a packet that has data in the head area being
fed to tso_fragment which bombs out.
My take on this is that we shouldn't ever call tcp_fragment on a TSO
socket for a packet that is yet to be transmitted since this creates
a packet on sk_send_head that cannot be extended.
So here is a patch to change it so that tso_fragment is always used
in this case.
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
Well I've only found one potential cause for the assertion
failure in tcp_mark_head_lost. First of all, this can only
occur if cnt > 1 since tp->packets_out is never zero here.
If it did hit zero we'd have much bigger problems.
So cnt is equal to fackets_out - reordering. Normally
fackets_out is less than packets_out. The only reason
I've found that might cause fackets_out to exceed packets_out
is if tcp_fragment is called from tcp_retransmit_skb with a
TSO skb and the current MSS is greater than the MSS stored
in the TSO skb. This might occur as the result of an expiring
dst entry.
In that case, packets_out may decrease (line 1380-1381 in
tcp_output.c). However, fackets_out is unchanged which means
that it may in fact exceed packets_out.
Previously tcp_retrans_try_collapse was the only place where
packets_out can go down and it takes care of this by decrementing
fackets_out.
So we should make sure that fackets_out is reduced by an appropriate
amount here as well.
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
tcp_write_xmit caches the cwnd value indirectly in cwnd_quota. When
tcp_transmit_skb reduces the cwnd because of tcp_enter_cwr, the cached
value becomes invalid.
This patch ensures that the cwnd value is always reread after each
tcp_transmit_skb call.
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Cc: "David S. Miller" <davem@davemloft.net>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
MSS changes can be lost since we preemptively initialize the tso_segs count
for an SKB before we %100 commit to sending it out.
So, by the time we send it out, the tso_size information can be stale due
to PMTU events. This mucks up all of the logic in our send engine, and can
even result in the BUG() triggering in tcp_tso_should_defer().
Another problem we have is that we're storing the tp->mss_cache, not the
SACK block normalized MSS, as the tso_size. That's wrong too.
Signed-off-by: David S. Miller <davem@davemloft.net>
Cc: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
From: Victor Fusco <victor@cetuc.puc-rio.br>
Fix the sparse warning "implicit cast to nocast type"
Signed-off-by: Victor Fusco <victor@cetuc.puc-rio.br>
Signed-off-by: Domen Puncer <domen@coderock.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
Congestion window recover after loss depends upon the fact
that if we have a full MSS sized frame at the head of the
send queue, we will send it. TSO deferral can defeat the
ACK clocking necessary to exit cleanly from recovery.
Signed-off-by: David S. Miller <davem@davemloft.net>
Make TSO segment transmit size decisions at send time not earlier.
The basic scheme is that we try to build as large a TSO frame as
possible when pulling in the user data, but the size of the TSO frame
output to the card is determined at transmit time.
This is guided by tp->xmit_size_goal. It is always set to a multiple
of MSS and tells sendmsg/sendpage how large an SKB to try and build.
Later, tcp_write_xmit() and tcp_push_one() chop up the packet if
necessary and conditions warrant. These routines can also decide to
"defer" in order to wait for more ACKs to arrive and thus allow larger
TSO frames to be emitted.
A general observation is that TSO elongates the pipe, thus requiring a
larger congestion window and larger buffering especially at the sender
side. Therefore, it is important that applications 1) get a large
enough socket send buffer (this is accomplished by our dynamic send
buffer expansion code) 2) do large enough writes.
Signed-off-by: David S. Miller <davem@davemloft.net>
tcp_snd_test() is run for every packet output by a single
call to tcp_write_xmit(), but this is not necessary.
For one, the congestion window space needs to only be
calculated one time, then used throughout the duration
of the loop.
This cleanup also makes experimenting with different TSO
packetization schemes much easier.
Signed-off-by: David S. Miller <davem@davemloft.net>
tcp_snd_test() does several different things, use inline
functions to express this more clearly.
1) It initializes the TSO count of SKB, if necessary.
2) It performs the Nagle test.
3) It makes sure the congestion window is adhered to.
4) It makes sure SKB fits into the send window.
This cleanup also sets things up so that things like the
available packets in the congestion window does not need
to be calculated multiple times by packet sending loops
such as tcp_write_xmit().
Signed-off-by: David S. Miller <davem@davemloft.net>
'nonagle' should be passed to the tcp_snd_test() function
as 'TCP_NAGLE_PUSH' if we are checking an SKB not at the
tail of the write_queue. This is because Nagle does not
apply to such frames since we cannot possibly tack more
data onto them.
However, while doing this __tcp_push_pending_frames() makes
all of the packets in the write_queue use this modified
'nonagle' value.
Fix the bug and simplify this function by just calling
tcp_write_xmit() directly if sk_send_head is non-NULL.
As a result, we can now make tcp_data_snd_check() just call
tcp_push_pending_frames() instead of the specialized
__tcp_data_snd_check().
Signed-off-by: David S. Miller <davem@davemloft.net>
tcp_write_xmit() uses tcp_current_mss(), but some of it's callers,
namely __tcp_push_pending_frames(), already has this value available
already.
While we're here, fix the "cur_mss" argument to be "unsigned int"
instead of plain "unsigned".
Signed-off-by: David S. Miller <davem@davemloft.net>
Put the main basic block of work at the top-level of
tabbing, and mark the TCP_CLOSE test with unlikely().
Signed-off-by: David S. Miller <davem@davemloft.net>
