There were quite a few overlapping sets of changes here.
Daniel's bug fix for off-by-ones in the new BPF branch instructions,
along with the added allowances for "data_end > ptr + x" forms
collided with the metadata additions.
Along with those three changes came veritifer test cases, which in
their final form I tried to group together properly. If I had just
trimmed GIT's conflict tags as-is, this would have split up the
meta tests unnecessarily.
In the socketmap code, a set of preemption disabling changes
overlapped with the rename of bpf_compute_data_end() to
bpf_compute_data_pointers().
Changes were made to the mv88e6060.c driver set addr method
which got removed in net-next.
The hyperv transport socket layer had a locking change in 'net'
which overlapped with a change of socket state macro usage
in 'net-next'.
Signed-off-by: David S. Miller <davem@davemloft.net>
In preparation to enabling -Wimplicit-fallthrough, mark switch cases
where we are expecting to fall through.
Notice that in some cases I placed the "fall through" comment
on its own line, which is what GCC is expecting to find.
Addresses-Coverity-ID: 115108
Signed-off-by: Gustavo A. R. Silva <garsilva@embeddedor.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Using a linear list to store all skbs in write queue has been okay
for quite a while : O(N) is not too bad when N < 500.
Things get messy when N is the order of 100,000 : Modern TCP stacks
want 10Gbit+ of throughput even with 200 ms RTT flows.
40 ns per cache line miss means a full scan can use 4 ms,
blowing away CPU caches.
SACK processing often can use various hints to avoid parsing
whole retransmit queue. But with high packet losses and/or high
reordering, hints no longer work.
Sender has to process thousands of unfriendly SACK, accumulating
a huge socket backlog, burning a cpu and massively dropping packets.
Using an rb-tree for retransmit queue has been avoided for years
because it added complexity and overhead, but now is the time
to be more resistant and say no to quadratic behavior.
1) RTX queue is no longer part of the write queue : already sent skbs
are stored in one rb-tree.
2) Since reaching the head of write queue no longer needs
sk->sk_send_head, we added an union of sk_send_head and tcp_rtx_queue
Tested:
On receiver :
netem on ingress : delay 150ms 200us loss 1
GRO disabled to force stress and SACK storms.
for f in `seq 1 10`
do
./netperf -H lpaa6 -l30 -- -K bbr -o THROUGHPUT|tail -1
done | awk '{print $0} {sum += $0} END {printf "%7u\n",sum}'
Before patch :
323.87
351.48
339.59
338.62
306.72
204.07
304.93
291.88
202.47
176.88
2840
After patch:
1700.83
2207.98
2070.17
1544.26
2114.76
2124.89
1693.14
1080.91
2216.82
1299.94
18053
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
No need to recompute previous skb, as it will be a bit more
expensive when rtx queue is converted to RB tree.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
With upcoming rb-tree implementation, the checks will trigger
more often, and this is expected.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
It will be a bit more expensive to get the head of rtx queue
once rtx queue is converted to an rb-tree.
We can avoid this extra cost in case tp->lost_skb_hint is set.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Geeralize private netem_rb_to_skb()
TCP rtx queue will soon be converted to rb-tree,
so we will need skb_rbtree_walk() helpers.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch adds a new queue (list) that tracks the sent but not yet
acked or SACKed skbs for a TCP connection. The list is chronologically
ordered by skb->skb_mstamp (the head is the oldest sent skb).
This list will be used to optimize TCP Rack recovery, which checks
an skb's timestamp to judge if it has been lost and needs to be
retransmitted. Since TCP write queue is ordered by sequence instead
of sent time, RACK has to scan over the write queue to catch all
eligible packets to detect lost retransmission, and iterates through
SACKed skbs repeatedly.
Special cares for rare events:
1. TCP repair fakes skb transmission so the send queue needs adjusted
2. SACK reneging would require re-inserting SACKed skbs into the
send queue. For now I believe it's not worth the complexity to
make RACK work perfectly on SACK reneging, so we do nothing here.
3. Fast Open: currently for non-TFO, send-queue correctly queues
the pure SYN packet. For TFO which queues a pure SYN and
then a data packet, send-queue only queues the data packet but
not the pure SYN due to the structure of TFO code. This is okay
because the SYN receiver would never respond with a SACK on a
missing SYN (i.e. SYN is never fast-retransmitted by SACK/RACK).
In order to not grow sk_buff, we use an union for the new list and
_skb_refdst/destructor fields. This is a bit complicated because
we need to make sure _skb_refdst and destructor are properly zeroed
before skb is cloned/copied at transmit, and before being freed.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This commit does a cleanup and moves tcp_rearm_rto() call in the TFO
server case into a previous spot in tcp_rcv_state_process() to make
it more compact.
This is only a cosmetic change.
Suggested-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Wei Wang <weiwan@google.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Acked-by: Yuchung Cheng <ycheng@google.com>
Acked-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Currently in the TCP code, the initialization sequence for cached
metrics, congestion control, BPF, etc, after successful connection
is very inconsistent. This introduces inconsistent bevhavior and is
prone to bugs. The current call sequence is as follows:
(1) for active case (tcp_finish_connect() case):
tcp_mtup_init(sk);
icsk->icsk_af_ops->rebuild_header(sk);
tcp_init_metrics(sk);
tcp_call_bpf(sk, BPF_SOCK_OPS_ACTIVE_ESTABLISHED_CB);
tcp_init_congestion_control(sk);
tcp_init_buffer_space(sk);
(2) for passive case (tcp_rcv_state_process() TCP_SYN_RECV case):
icsk->icsk_af_ops->rebuild_header(sk);
tcp_call_bpf(sk, BPF_SOCK_OPS_PASSIVE_ESTABLISHED_CB);
tcp_init_congestion_control(sk);
tcp_mtup_init(sk);
tcp_init_buffer_space(sk);
tcp_init_metrics(sk);
(3) for TFO passive case (tcp_fastopen_create_child()):
inet_csk(child)->icsk_af_ops->rebuild_header(child);
tcp_init_congestion_control(child);
tcp_mtup_init(child);
tcp_init_metrics(child);
tcp_call_bpf(child, BPF_SOCK_OPS_PASSIVE_ESTABLISHED_CB);
tcp_init_buffer_space(child);
This commit uniforms the above functions to have the following sequence:
tcp_mtup_init(sk);
icsk->icsk_af_ops->rebuild_header(sk);
tcp_init_metrics(sk);
tcp_call_bpf(sk, BPF_SOCK_OPS_ACTIVE/PASSIVE_ESTABLISHED_CB);
tcp_init_congestion_control(sk);
tcp_init_buffer_space(sk);
This sequence is the same as the (1) active case. We pick this sequence
because this order correctly allows BPF to override the settings
including congestion control module and initial cwnd, etc from
the route, and then allows the CC module to see those settings.
Suggested-by: Neal Cardwell <ncardwell@google.com>
Tested-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: Wei Wang <weiwan@google.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Acked-by: Yuchung Cheng <ycheng@google.com>
Acked-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
skb->rbnode shares space with skb->next, skb->prev and skb->tstamp
Current uses (TCP receive ofo queue and netem) need to save/restore
tstamp, while skb->dev is either NULL (TCP) or a constant for a given
queue (netem).
Since we plan using an RB tree for TCP retransmit queue to speedup SACK
processing with large BDP, this patch exchanges skb->dev and
skb->tstamp.
This saves some overhead in both TCP and netem.
v2: removes the swtstamp field from struct tcp_skb_cb
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Soheil Hassas Yeganeh <soheil@google.com>
Cc: Wei Wang <weiwan@google.com>
Cc: Willem de Bruijn <willemb@google.com>
Acked-by: Soheil Hassas Yeganeh <soheil@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Move the #ifdef into the static void function so that the use
of DBGUNDO is validated when FASTRETRANS_DEBUG <= 1.
Remove the now unnecessary #else and #define DBGUNDO.
Signed-off-by: Joe Perches <joe@perches.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This reverts commit 45f119bf93.
Eric Dumazet says:
We found at Google a significant regression caused by
45f119bf93 tcp: remove header prediction
In typical RPC (TCP_RR), when a TCP socket receives data, we now call
tcp_ack() while we used to not call it.
This touches enough cache lines to cause a slowdown.
so problem does not seem to be HP removal itself but the tcp_ack()
call. Therefore, it might be possible to remove HP after all, provided
one finds a way to elide tcp_ack for most cases.
Reported-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Florian Westphal <fw@strlen.de>
Signed-off-by: David S. Miller <davem@davemloft.net>
This change was a followup to the header prediction removal,
so first revert this as a prerequisite to back out hp removal.
Signed-off-by: Florian Westphal <fw@strlen.de>
Signed-off-by: David S. Miller <davem@davemloft.net>
Florian reported UDP xmit drops that could be root caused to the
too small neigh limit.
Current limit is 64 KB, meaning that even a single UDP socket would hit
it, since its default sk_sndbuf comes from net.core.wmem_default
(~212992 bytes on 64bit arches).
Once ARP/ND resolution is in progress, we should allow a little more
packets to be queued, at least for one producer.
Once neigh arp_queue is filled, a rogue socket should hit its sk_sndbuf
limit and either block in sendmsg() or return -EAGAIN.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reported-by: Florian Fainelli <f.fainelli@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
When SOF_TIMESTAMPING_RX_SOFTWARE is enabled for tcp sockets, return the
timestamp corresponding to the highest sequence number data returned.
Previously the skb->tstamp is overwritten when a TCP packet is placed
in the out of order queue. While the packet is in the ooo queue, save the
timestamp in the TCB_SKB_CB. This space is shared with the gso_*
options which are only used on the tx path, and a previously unused 4
byte hole.
When skbs are coalesced either in the sk_receive_queue or the
out_of_order_queue always choose the timestamp of the appended skb to
maintain the invariant of returning the timestamp of the last byte in
the recvmsg buffer.
Signed-off-by: Mike Maloney <maloney@google.com>
Acked-by: Willem de Bruijn <willemb@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Tonghao Zhang <xiangxia.m.yue@gmail.com>
Acked-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
tcp_peer_is_proven needs a proper route to make the
determination, but dst always is NULL. This bug may
be there at the beginning of git tree. This does not
look serious enough to deserve backports to stable
versions.
Signed-off-by: Tonghao Zhang <xiangxia.m.yue@gmail.com>
Acked-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
In some situations tcp_send_loss_probe() can realize that it's unable
to send a loss probe (TLP), and falls back to calling tcp_rearm_rto()
to schedule an RTO timer. In such cases, sometimes tcp_rearm_rto()
realizes that the RTO was eligible to fire immediately or at some
point in the past (delta_us <= 0). Previously in such cases
tcp_rearm_rto() was scheduling such "overdue" RTOs to happen at now +
icsk_rto, which caused needless delays of hundreds of milliseconds
(and non-linear behavior that made reproducible testing
difficult). This commit changes the logic to schedule "overdue" RTOs
ASAP, rather than at now + icsk_rto.
Fixes: 6ba8a3b19e ("tcp: Tail loss probe (TLP)")
Suggested-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The UDP offload conflict is dealt with by simply taking what is
in net-next where we have removed all of the UFO handling code
entirely.
The TCP conflict was a case of local variables in a function
being removed from both net and net-next.
In netvsc we had an assignment right next to where a missing
set of u64 stats sync object inits were added.
Signed-off-by: David S. Miller <davem@davemloft.net>
Using ssthresh to revert cwnd is less reliable when ssthresh is
bounded to 2 packets. This patch uses an existing variable in TCP
"prior_cwnd" that snapshots the cwnd right before entering fast
recovery and RTO recovery in Reno. This fixes the issue discussed
in netdev thread: "A buggy behavior for Linux TCP Reno and HTCP"
https://www.spinics.net/lists/netdev/msg444955.html
Suggested-by: Neal Cardwell <ncardwell@google.com>
Reported-by: Wei Sun <unlcsewsun@gmail.com>
Signed-off-by: Yuchung Cheng <ncardwell@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Fix a TCP loss recovery performance bug raised recently on the netdev
list, in two threads:
(i) July 26, 2017: netdev thread "TCP fast retransmit issues"
(ii) July 26, 2017: netdev thread:
"[PATCH V2 net-next] TLP: Don't reschedule PTO when there's one
outstanding TLP retransmission"
The basic problem is that incoming TCP packets that did not indicate
forward progress could cause the xmit timer (TLP or RTO) to be rearmed
and pushed back in time. In certain corner cases this could result in
the following problems noted in these threads:
- Repeated ACKs coming in with bogus SACKs corrupted by middleboxes
could cause TCP to repeatedly schedule TLPs forever. We kept
sending TLPs after every ~200ms, which elicited bogus SACKs, which
caused more TLPs, ad infinitum; we never fired an RTO to fill in
the holes.
- Incoming data segments could, in some cases, cause us to reschedule
our RTO or TLP timer further out in time, for no good reason. This
could cause repeated inbound data to result in stalls in outbound
data, in the presence of packet loss.
This commit fixes these bugs by changing the TLP and RTO ACK
processing to:
(a) Only reschedule the xmit timer once per ACK.
(b) Only reschedule the xmit timer if tcp_clean_rtx_queue() deems the
ACK indicates sufficient forward progress (a packet was
cumulatively ACKed, or we got a SACK for a packet that was sent
before the most recent retransmit of the write queue head).
This brings us back into closer compliance with the RFCs, since, as
the comment for tcp_rearm_rto() notes, we should only restart the RTO
timer after forward progress on the connection. Previously we were
restarting the xmit timer even in these cases where there was no
forward progress.
As a side benefit, this commit simplifies and speeds up the TCP timer
arming logic. We had been calling inet_csk_reset_xmit_timer() three
times on normal ACKs that cumulatively acknowledged some data:
1) Once near the top of tcp_ack() to switch from TLP timer to RTO:
if (icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)
tcp_rearm_rto(sk);
2) Once in tcp_clean_rtx_queue(), to update the RTO:
if (flag & FLAG_ACKED) {
tcp_rearm_rto(sk);
3) Once in tcp_ack() after tcp_fastretrans_alert() to switch from RTO
to TLP:
if (icsk->icsk_pending == ICSK_TIME_RETRANS)
tcp_schedule_loss_probe(sk);
This commit, by only rescheduling the xmit timer once per ACK,
simplifies the code and reduces CPU overhead.
This commit was tested in an A/B test with Google web server
traffic. SNMP stats and request latency metrics were within noise
levels, substantiating that for normal web traffic patterns this is a
rare issue. This commit was also tested with packetdrill tests to
verify that it fixes the timer behavior in the corner cases discussed
in the netdev threads mentioned above.
This patch is a bug fix patch intended to be queued for -stable
relases.
Fixes: 6ba8a3b19e ("tcp: Tail loss probe (TLP)")
Reported-by: Klavs Klavsen <kl@vsen.dk>
Reported-by: Mao Wenan <maowenan@huawei.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Nandita Dukkipati <nanditad@google.com>
Acked-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Pure refactor. This helper will be required in the xmit timer fix
later in the patch series. (Because the TLP logic will want to make
this calculation.)
Fixes: 6ba8a3b19e ("tcp: Tail loss probe (TLP)")
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Nandita Dukkipati <nanditad@google.com>
Acked-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Commit 45f119bf93 ("tcp: remove header prediction") introduced a
minor bug: the sk_state_change() and sk_wake_async() notifications for
a completed active connection happen twice: once in this new spot
inside tcp_finish_connect() and once in the existing code in
tcp_rcv_synsent_state_process() immediately after it calls
tcp_finish_connect(). This commit remoes the duplicate POLL_OUT
notifications.
Fixes: 45f119bf93 ("tcp: remove header prediction")
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Cc: Florian Westphal <fw@strlen.de>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Yuchung Cheng <ycheng@google.com>
Acked-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
If the sender switches the congestion control during ECN-triggered
cwnd-reduction state (CA_CWR), upon exiting recovery cwnd is set to
the ssthresh value calculated by the previous congestion control. If
the previous congestion control is BBR that always keep ssthresh
to TCP_INIFINITE_SSTHRESH, cwnd ends up being infinite. The safe
step is to avoid assigning invalid ssthresh value when recovery ends.
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Acked-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Commit c13ee2a4f0 ("tcp: reindent two spots after prequeue removal")
removed code in tcp_data_queue().
We can go a little farther, removing an always true test,
and removing initializers for fragstolen and eaten variables.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Florian Westphal <fw@strlen.de>
Signed-off-by: David S. Miller <davem@davemloft.net>
re-indent tcp_ack, and remove CA_ACK_SLOWPATH; it is always set now.
Signed-off-by: Florian Westphal <fw@strlen.de>
Signed-off-by: David S. Miller <davem@davemloft.net>
Like prequeue, I am not sure this is overly useful nowadays.
If we receive a train of packets, GRO will aggregate them if the
headers are the same (HP predates GRO by several years) so we don't
get a per-packet benefit, only a per-aggregated-packet one.
Signed-off-by: Florian Westphal <fw@strlen.de>
Signed-off-by: David S. Miller <davem@davemloft.net>
These two branches are now always true, remove the conditional.
objdiff shows no changes.
Signed-off-by: Florian Westphal <fw@strlen.de>
Signed-off-by: David S. Miller <davem@davemloft.net>
prequeue is a tcp receive optimization that moves part of rx processing
from bh to process context.
This only works if the socket being processed belongs to a process that
is blocked in recv on that socket.
In practice, this doesn't happen anymore that often because nowadays
servers tend to use an event driven (epoll) model.
Even normal client applications (web browsers) commonly use many tcp
connections in parallel.
This has measureable impact only in netperf (which uses plain recv and
thus allows prequeue use) from host to locally running vm (~4%), however,
there were no changes when using netperf between two physical hosts with
ixgbe interfaces.
Signed-off-by: Florian Westphal <fw@strlen.de>
Signed-off-by: David S. Miller <davem@davemloft.net>
The last (4th) argument of tcp_rcv_established() is redundant as it
always equals to skb->len and the skb itself is always passed as 2th
agrument. There is no reason to have it.
Signed-off-by: Ilya V. Matveychikov <matvejchikov@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Added support for changing congestion control for SOCK_OPS bpf
programs through the setsockopt bpf helper function. It also adds
a new SOCK_OPS op, BPF_SOCK_OPS_NEEDS_ECN, that is needed for
congestion controls, like dctcp, that need to enable ECN in the
SYN packets.
Signed-off-by: Lawrence Brakmo <brakmo@fb.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Added callbacks to BPF SOCK_OPS type program before an active
connection is intialized and after a passive or active connection is
established.
The following patch demostrates how they can be used to set send and
receive buffer sizes.
Signed-off-by: Lawrence Brakmo <brakmo@fb.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch adds support for setting a per connection SYN and
SYN_ACK RTOs from within a BPF_SOCK_OPS program. For example,
to set small RTOs when it is known both hosts are within a
datacenter.
Signed-off-by: Lawrence Brakmo <brakmo@fb.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
We want to move some TCP sysctls to net namespaces in the future.
tcp_window_scaling, tcp_sack and tcp_timestamps being fetched
from tcp_parse_options(), we need to pass an extra parameter.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Currently when a data packet is retransmitted, we do not compute an
RTT sample for congestion control due to Kern's check. Therefore the
congestion control that uses RTT signals may not receive any update
during loss recovery which could last many round trips. For example,
BBR and Vegas may not be able to update its min RTT estimation if the
network path has shortened until it recovers from losses. This patch
mitigates that by using TCP timestamp options for RTT measurement
for congestion control. Note that we already use timestamps for
RTT estimation.
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Soheil Hassas Yeganeh <soheil@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Paul Fiterau Brostean reported :
<quote>
Linux TCP stack we analyze exhibits behavior that seems odd to me.
The scenario is as follows (all packets have empty payloads, no window
scaling, rcv/snd window size should not be a factor):
TEST HARNESS (CLIENT) LINUX SERVER
1. - LISTEN (server listen,
then accepts)
2. - --> <SEQ=100><CTL=SYN> --> SYN-RECEIVED
3. - <-- <SEQ=300><ACK=101><CTL=SYN,ACK> <-- SYN-RECEIVED
4. - --> <SEQ=101><ACK=301><CTL=ACK> --> ESTABLISHED
5. - <-- <SEQ=301><ACK=101><CTL=FIN,ACK> <-- FIN WAIT-1 (server
opts to close the data connection calling "close" on the connection
socket)
6. - --> <SEQ=101><ACK=99999><CTL=FIN,ACK> --> CLOSING (client sends
FIN,ACK with not yet sent acknowledgement number)
7. - <-- <SEQ=302><ACK=102><CTL=ACK> <-- CLOSING (ACK is 102
instead of 101, why?)
... (silence from CLIENT)
8. - <-- <SEQ=301><ACK=102><CTL=FIN,ACK> <-- CLOSING
(retransmission, again ACK is 102)
Now, note that packet 6 while having the expected sequence number,
acknowledges something that wasn't sent by the server. So I would
expect
the packet to maybe prompt an ACK response from the server, and then be
ignored. Yet it is not ignored and actually leads to an increase of the
acknowledgement number in the server's retransmission of the FIN,ACK
packet. The explanation I found is that the FIN in packet 6 was
processed, despite the acknowledgement number being unacceptable.
Further experiments indeed show that the server processes this FIN,
transitioning to CLOSING, then on receiving an ACK for the FIN it had
send in packet 5, the server (or better said connection) transitions
from CLOSING to TIME_WAIT (as signaled by netstat).
</quote>
Indeed, tcp_rcv_state_process() calls tcp_ack() but
does not exploit the @acceptable status but for TCP_SYN_RECV
state.
What we want here is to send a challenge ACK, if not in TCP_SYN_RECV
state. TCP_FIN_WAIT1 state is not the only state we should fix.
Add a FLAG_NO_CHALLENGE_ACK so that tcp_rcv_state_process()
can choose to send a challenge ACK and discard the packet instead
of wrongly change socket state.
With help from Neal Cardwell.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reported-by: Paul Fiterau Brostean <p.fiterau-brostean@science.ru.nl>
Cc: Neal Cardwell <ncardwell@google.com>
Cc: Yuchung Cheng <ycheng@google.com>
Cc: Soheil Hassas Yeganeh <soheil@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Commit bafbb9c732 ("tcp: eliminate negative reordering
in tcp_clean_rtx_queue") fixes an issue for negative
reordering metrics.
To be resilient to such errors, warn and return
when a negative metric is passed to tcp_update_reordering().
Signed-off-by: Soheil Hassas Yeganeh <soheil@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
skbs in (re)transmit queue no longer have a copy of jiffies
at the time of the transmit : skb->skb_mstamp is now in usec unit,
with no correlation to tcp_jiffies32.
We have to convert rto from jiffies to usec, compute a time difference
in usec, then convert the delta to HZ units.
Fixes: 9a568de481 ("tcp: switch TCP TS option (RFC 7323) to 1ms clock")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Acked-by: Soheil Hassas Yeganeh <soheil@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
TCP Timestamps option is defined in RFC 7323
Traditionally on linux, it has been tied to the internal
'jiffies' variable, because it had been a cheap and good enough
generator.
For TCP flows on the Internet, 1 ms resolution would be much better
than 4ms or 10ms (HZ=250 or HZ=100 respectively)
For TCP flows in the DC, Google has used usec resolution for more
than two years with great success [1]
Receive size autotuning (DRS) is indeed more precise and converges
faster to optimal window size.
This patch converts tp->tcp_mstamp to a plain u64 value storing
a 1 usec TCP clock.
This choice will allow us to upstream the 1 usec TS option as
discussed in IETF 97.
[1] https://www.ietf.org/proceedings/97/slides/slides-97-tcpm-tcp-options-for-low-latency-00.pdf
Signed-off-by: Eric Dumazet <edumazet@google.com>
Acked-by: Soheil Hassas Yeganeh <soheil@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
After this patch, all uses of tcp_time_stamp will require
a change when we introduce 1 ms and/or 1 us TCP TS option.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Acked-by: Soheil Hassas Yeganeh <soheil@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This place wants to use tcp_jiffies32, this is good enough.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Acked-by: Soheil Hassas Yeganeh <soheil@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Use tcp_jiffies32 instead of tcp_time_stamp, since
tcp_time_stamp will soon be only used for TCP TS option.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Acked-by: Soheil Hassas Yeganeh <soheil@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Use tcp_jiffies32 instead of tcp_time_stamp to feed
tp->snd_cwnd_stamp.
tcp_time_stamp will soon be a litle bit more expensive
than simply reading 'jiffies'.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Acked-by: Soheil Hassas Yeganeh <soheil@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Use tcp_jiffies32 instead of tcp_time_stamp to feed
tp->lsndtime.
tcp_time_stamp will soon be a litle bit more expensive
than simply reading 'jiffies'.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Acked-by: Soheil Hassas Yeganeh <soheil@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
tcp_ack() can call tcp_fragment() which may dededuct the
value tp->fackets_out when MSS changes. When prior_fackets
is larger than tp->fackets_out, tcp_clean_rtx_queue() can
invoke tcp_update_reordering() with negative values. This
results in absurd tp->reodering values higher than
sysctl_tcp_max_reordering.
Note that tcp_update_reordering indeeds sets tp->reordering
to min(sysctl_tcp_max_reordering, metric), but because
the comparison is signed, a negative metric always wins.
Fixes: c7caf8d3ed ("[TCP]: Fix reord detection due to snd_una covered holes")
Reported-by: Rebecca Isaacs <risaacs@google.com>
Signed-off-by: Soheil Hassas Yeganeh <soheil@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch fixes a bug in splitting an SKB during SACK
processing. Specifically if an skb contains multiple
packets and is only partially sacked in the higher sequences,
tcp_match_sack_to_skb() splits the skb and marks the second fragment
as SACKed.
The current code further attempts rounding up the first fragment
to MSS boundaries. But it misses a boundary condition when the
rounded-up fragment size (pkt_len) is exactly skb size. Spliting
such an skb is pointless and causses a kernel warning and aborts
the SACK processing. This patch universally checks such over-split
before calling tcp_fragment to prevent these unnecessary warnings.
Fixes: adb92db857 ("tcp: Make SACK code to split only at mss boundaries")
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Soheil Hassas Yeganeh <soheil@google.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Whole point of randomization was to hide server uptime, but an attacker
can simply start a syn flood and TCP generates 'old style' timestamps,
directly revealing server jiffies value.
Also, TSval sent by the server to a particular remote address vary
depending on syncookies being sent or not, potentially triggering PAWS
drops for innocent clients.
Lets implement proper randomization, including for SYNcookies.
Also we do not need to export sysctl_tcp_timestamps, since it is not
used from a module.
In v2, I added Florian feedback and contribution, adding tsoff to
tcp_get_cookie_sock().
v3 removed one unused variable in tcp_v4_connect() as Florian spotted.
Fixes: 95a22caee3 ("tcp: randomize tcp timestamp offsets for each connection")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Florian Westphal <fw@strlen.de>
Tested-by: Florian Westphal <fw@strlen.de>
Cc: Yuchung Cheng <ycheng@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Some devices or distributions use HZ=100 or HZ=250
TCP receive buffer autotuning has poor behavior caused by this choice.
Since autotuning happens after 4 ms or 10 ms, short distance flows
get their receive buffer tuned to a very high value, but after an initial
period where it was frozen to (too small) initial value.
With tp->tcp_mstamp introduction, we can switch to high resolution
timestamps almost for free (at the expense of 8 additional bytes per
TCP structure)
Note that some TCP stacks use usec TCP timestamps where this
patch makes even more sense : Many TCP flows have < 500 usec RTT.
Hopefully this finer TS option can be standardized soon.
Tested:
HZ=100 kernel
./netperf -H lpaa24 -t TCP_RR -l 1000 -- -r 10000,10000 &
Peer without patch :
lpaa24:~# ss -tmi dst lpaa23
...
skmem:(r0,rb8388608,...)
rcv_rtt:10 rcv_space:3210000 minrtt:0.017
Peer with the patch :
lpaa23:~# ss -tmi dst lpaa24
...
skmem:(r0,rb428800,...)
rcv_rtt:0.069 rcv_space:30000 minrtt:0.017
We can see saner RCVBUF, and more precise rcv_rtt information.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Acked-by: Soheil Hassas Yeganeh <soheil@google.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
It is no longer needed, everything uses tp->tcp_mstamp instead.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Acked-by: Soheil Hassas Yeganeh <soheil@google.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Following patch will remove ack_time from struct tcp_sacktag_state
Same info is now found in tp->tcp_mstamp
Signed-off-by: Eric Dumazet <edumazet@google.com>
Acked-by: Soheil Hassas Yeganeh <soheil@google.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
No longer needed, since tp->tcp_mstamp holds the information.
This is needed to remove sack_state.ack_time in a following patch.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Acked-by: Soheil Hassas Yeganeh <soheil@google.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
No longer needed, since tp->tcp_mstamp holds the information.
This is needed to remove sack_state.ack_time in a following patch.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Acked-by: Soheil Hassas Yeganeh <soheil@google.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Not used anymore now tp->tcp_mstamp holds the information.
This is needed to remove sack_state.ack_time in a following patch.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Acked-by: Soheil Hassas Yeganeh <soheil@google.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Not used anymore now tp->tcp_mstamp holds the information.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Acked-by: Soheil Hassas Yeganeh <soheil@google.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This is no longer used, since tcp_rack_detect_loss() takes
the timestamp from tp->tcp_mstamp
Signed-off-by: Eric Dumazet <edumazet@google.com>
Acked-by: Soheil Hassas Yeganeh <soheil@google.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
We want to use precise timestamps in TCP stack, but we do not
want to call possibly expensive kernel time services too often.
tp->tcp_mstamp is guaranteed to be updated once per incoming packet.
We will use it in the following patches, removing specific
skb_mstamp_get() calls, and removing ack_time from
struct tcp_sacktag_state.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Acked-by: Soheil Hassas Yeganeh <soheil@google.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This counter records the number of times the firewall blackhole issue is
detected and active TFO is disabled.
Signed-off-by: Wei Wang <weiwan@google.com>
Acked-by: Yuchung Cheng <ycheng@google.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Middlebox firewall issues can potentially cause server's data being
blackholed after a successful 3WHS using TFO. Following are the related
reports from Apple:
https://www.nanog.org/sites/default/files/Paasch_Network_Support.pdf
Slide 31 identifies an issue where the client ACK to the server's data
sent during a TFO'd handshake is dropped.
C ---> syn-data ---> S
C <--- syn/ack ----- S
C (accept & write)
C <---- data ------- S
C ----- ACK -> X S
[retry and timeout]
https://www.ietf.org/proceedings/94/slides/slides-94-tcpm-13.pdf
Slide 5 shows a similar situation that the server's data gets dropped
after 3WHS.
C ---- syn-data ---> S
C <--- syn/ack ----- S
C ---- ack --------> S
S (accept & write)
C? X <- data ------ S
[retry and timeout]
This is the worst failure b/c the client can not detect such behavior to
mitigate the situation (such as disabling TFO). Failing to proceed, the
application (e.g., SSL library) may simply timeout and retry with TFO
again, and the process repeats indefinitely.
The proposed solution is to disable active TFO globally under the
following circumstances:
1. client side TFO socket detects out of order FIN
2. client side TFO socket receives out of order RST
We disable active side TFO globally for 1hr at first. Then if it
happens again, we disable it for 2h, then 4h, 8h, ...
And we reset the timeout to 1hr if a client side TFO sockets not opened
on loopback has successfully received data segs from server.
And we examine this condition during close().
The rational behind it is that when such firewall issue happens,
application running on the client should eventually close the socket as
it is not able to get the data it is expecting. Or application running
on the server should close the socket as it is not able to receive any
response from client.
In both cases, out of order FIN or RST will get received on the client
given that the firewall will not block them as no data are in those
frames.
And we want to disable active TFO globally as it helps if the middle box
is very close to the client and most of the connections are likely to
fail.
Also, add a debug sysctl:
tcp_fastopen_blackhole_detect_timeout_sec:
the initial timeout to use when firewall blackhole issue happens.
This can be set and read.
When setting it to 0, it means to disable the active disable logic.
Signed-off-by: Wei Wang <weiwan@google.com>
Acked-by: Yuchung Cheng <ycheng@google.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
When using TCP FastOpen for an active session, we send one wakeup event
from tcp_finish_connect(), right before the data eventually contained in
the received SYNACK is queued to sk->sk_receive_queue.
This means that depending on machine load or luck, poll() users
might receive POLLOUT events instead of POLLIN|POLLOUT
To fix this, we need to move the call to sk->sk_state_change()
after the (optional) call to tcp_rcv_fastopen_synack()
Signed-off-by: Eric Dumazet <edumazet@google.com>
Acked-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
When a RST packet is processed, we send two wakeup events to interested
polling users.
First one by a sk->sk_error_report(sk) from tcp_reset(),
followed by a sk->sk_state_change(sk) from tcp_done().
Depending on machine load and luck, poll() can either return POLLERR,
or POLLIN|POLLOUT|POLLERR|POLLHUP (this happens on 99 % of the cases)
This is probably fine, but we can avoid the confusion by reordering
things so that we have more TCP fields updated before the first wakeup.
This might even allow us to remove some barriers we added in the past.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Acked-by: Soheil Hassas Yeganeh <soheil@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Conflicts were simply overlapping changes. In the net/ipv4/route.c
case the code had simply moved around a little bit and the same fix
was made in both 'net' and 'net-next'.
In the net/sched/sch_generic.c case a fix in 'net' happened at
the same time that a new argument was added to qdisc_hash_add().
Signed-off-by: David S. Miller <davem@davemloft.net>
The recent extension of F-RTO 89fe18e44 ("tcp: extend F-RTO
to catch more spurious timeouts") interacts badly with certain
broken middle-boxes. These broken boxes modify and falsely raise
the receive window on the ACKs. During a timeout induced recovery,
F-RTO would send new data packets to probe if the timeout is false
or not. Since the receive window is falsely raised, the receiver
would silently drop these F-RTO packets. The recovery would take N
(exponentially backoff) timeouts to repair N packet losses. A TCP
performance killer.
Due to this unfortunate situation, this patch removes this extension
to revert F-RTO back to the RFC specification.
Fixes: 89fe18e44f ("tcp: extend F-RTO to catch more spurious timeouts")
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: Soheil Hassas Yeganeh <soheil@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Mostly simple cases of overlapping changes (adding code nearby,
a function whose name changes, for example).
Signed-off-by: David S. Miller <davem@davemloft.net>
Currently the reordering SNMP counters only increase if a connection
sees a higher degree then it has previously seen. It ignores if the
reordering degree is not greater than the default system threshold.
This significantly under-counts the number of reordering events
and falsely convey that reordering is rare on the network.
This patch properly and faithfully records the number of reordering
events detected by the TCP stack, just like the comment says "this
exciting event is worth to be remembered". Note that even so TCP
still under-estimate the actual reordering events because TCP
requires TS options or certain packet sequences to detect reordering
(i.e. ACKing never-retransmitted sequence in recovery or disordered
state).
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: Soheil Hassas Yeganeh <soheil@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Define one new macro TCP_MAX_WSCALE instead of literal number '14',
and use U16_MAX instead of 65535 as the max value of TCP window.
There is another minor change, use rounddown(space, mss) instead of
(space / mss) * mss;
Signed-off-by: Gao Feng <fgao@ikuai8.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Markus Trippelsdorf reported that after commit dcb17d22e1 ("tcp: warn
on bogus MSS and try to amend it") the kernel started logging the
warning for a NIC driver that doesn't even support GRO.
It was diagnosed that it was possibly caused on connections that were
using TCP Timestamps but some packets lacked the Timestamps option. As
we reduce rcv_mss when timestamps are used, the lack of them would cause
the packets to be bigger than expected, although this is a valid case.
As this warning is more as a hint, getting a clean-cut on the
threshold is probably not worth the execution time spent on it. This
patch thus alleviates the false-positives with 2 quick checks: by
accounting for the entire TCP option space and also checking against the
interface MTU if it's available.
These changes, specially the MTU one, might mask some real positives,
though if they are really happening, it's possible that sooner or later
it will be triggered anyway.
Reported-by: Markus Trippelsdorf <markus@trippelsdorf.de>
Cc: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: Marcelo Ricardo Leitner <marcelo.leitner@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Conflicts:
drivers/net/ethernet/broadcom/genet/bcmmii.c
drivers/net/hyperv/netvsc.c
kernel/bpf/hashtab.c
Almost entirely overlapping changes.
Signed-off-by: David S. Miller <davem@davemloft.net>
icsk_ack.lrcvtime has a 0 value at socket creation time.
tcpi_last_data_recv can have bogus value if no payload is ever received.
This patch initializes icsk_ack.lrcvtime for active sessions
in tcp_finish_connect(), and for passive sessions in
tcp_create_openreq_child()
Signed-off-by: Eric Dumazet <edumazet@google.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The tcp_tw_recycle was already broken for connections
behind NAT, since the per-destination timestamp is not
monotonically increasing for multiple machines behind
a single destination address.
After the randomization of TCP timestamp offsets
in commit 8a5bd45f6616 (tcp: randomize tcp timestamp offsets
for each connection), the tcp_tw_recycle is broken for all
types of connections for the same reason: the timestamps
received from a single machine is not monotonically increasing,
anymore.
Remove tcp_tw_recycle, since it is not functional. Also, remove
the PAWSPassive SNMP counter since it is only used for
tcp_tw_recycle, and simplify tcp_v4_route_req and tcp_v6_route_req
since the strict argument is only set when tcp_tw_recycle is
enabled.
Signed-off-by: Soheil Hassas Yeganeh <soheil@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Cc: Lutz Vieweg <lvml@5t9.de>
Cc: Florian Westphal <fw@strlen.de>
Signed-off-by: David S. Miller <davem@davemloft.net>
Commit 8a5bd45f6616 (tcp: randomize tcp timestamp offsets for each connection)
randomizes TCP timestamps per connection. After this commit,
there is no guarantee that the timestamps received from the
same destination are monotonically increasing. As a result,
the per-destination timestamp cache in TCP metrics (i.e., tcpm_ts
in struct tcp_metrics_block) is broken and cannot be relied upon.
Remove the per-destination timestamp cache and all related code
paths.
Note that this cache was already broken for caching timestamps of
multiple machines behind a NAT sharing the same address.
Signed-off-by: Soheil Hassas Yeganeh <soheil@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Cc: Lutz Vieweg <lvml@5t9.de>
Cc: Florian Westphal <fw@strlen.de>
Signed-off-by: David S. Miller <davem@davemloft.net>
The functions that are returning tcp sequence number also setup
TS offset value, so rename them to better describe their purpose.
No functional changes in this patch.
Suggested-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Alexey Kodanev <alexey.kodanev@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
When same struct dst_entry can be used for many different
neighbours we can not use it for pending confirmations.
Use the new sk_dst_confirm() helper to propagate the
indication from received packets to sock_confirm_neigh().
Reported-by: YueHaibing <yuehaibing@huawei.com>
Fixes: 5110effee8 ("net: Do delayed neigh confirmation.")
Fixes: f2bb4bedf3 ("ipv4: Cache output routes in fib_info nexthops.")
Tested-by: YueHaibing <yuehaibing@huawei.com>
Signed-off-by: Julian Anastasov <ja@ssi.bg>
Acked-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
sock_reset_flag() maps to __clear_bit() not the atomic version clear_bit().
Thus, we need smp_mb(), smp_mb__after_atomic() is not sufficient.
Fixes: 3c7151275c ("tcp: add memory barriers to write space paths")
Cc: Eric Dumazet <eric.dumazet@gmail.com>
Cc: Oleg Nesterov <oleg@redhat.com>
Signed-off-by: Jason Baron <jbaron@akamai.com>
Acked-by: Eric Dumazet <edumazet@google.com>
Reported-by: Oleg Nesterov <oleg@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
tcp_add_backlog() can use skb_condense() helper to get better
gains and less SKB_TRUESIZE() magic. This only happens when socket
backlog has to be used.
Some attacks involve specially crafted out of order tiny TCP packets,
clogging the ofo queue of (many) sockets.
Then later, expensive collapse happens, trying to copy all these skbs
into single ones.
This unfortunately does not work if each skb has no neighbor in TCP
sequence order.
By using skb_condense() if the skb could not be coalesced to a prior
one, we defeat these kind of threats, potentially saving 4K per skb
(or more, since this is one page fragment).
A typical NAPI driver allocates gro packets with GRO_MAX_HEAD bytes
in skb->head, meaning the copy done by skb_condense() is limited to
about 200 bytes.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Using a Mac OSX box as a client connecting to a Linux server, we have found
that when certain applications (such as 'ab'), are abruptly terminated
(via ^C), a FIN is sent followed by a RST packet on tcp connections. The
FIN is accepted by the Linux stack but the RST is sent with the same
sequence number as the FIN, and Linux responds with a challenge ACK per
RFC 5961. The OSX client then sometimes (they are rate-limited) does not
reply with any RST as would be expected on a closed socket.
This results in sockets accumulating on the Linux server left mostly in
the CLOSE_WAIT state, although LAST_ACK and CLOSING are also possible.
This sequence of events can tie up a lot of resources on the Linux server
since there may be a lot of data in write buffers at the time of the RST.
Accepting a RST equal to rcv_nxt - 1, after we have already successfully
processed a FIN, has made a significant difference for us in practice, by
freeing up unneeded resources in a more expedient fashion.
A packetdrill test demonstrating the behavior:
// testing mac osx rst behavior
// Establish a connection
0.000 socket(..., SOCK_STREAM, IPPROTO_TCP) = 3
0.000 setsockopt(3, SOL_SOCKET, SO_REUSEADDR, [1], 4) = 0
0.000 bind(3, ..., ...) = 0
0.000 listen(3, 1) = 0
0.100 < S 0:0(0) win 32768 <mss 1460,nop,wscale 10>
0.100 > S. 0:0(0) ack 1 <mss 1460,nop,wscale 5>
0.200 < . 1:1(0) ack 1 win 32768
0.200 accept(3, ..., ...) = 4
// Client closes the connection
0.300 < F. 1:1(0) ack 1 win 32768
// now send rst with same sequence
0.300 < R. 1:1(0) ack 1 win 32768
// make sure we are in TCP_CLOSE
0.400 %{
assert tcpi_state == 7
}%
Signed-off-by: Jason Baron <jbaron@akamai.com>
Cc: Eric Dumazet <edumazet@google.com>
Acked-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch disables FACK by default as RACK is the successor of FACK
(inspired by the insights behind FACK).
FACK[1] in Linux works as follows: a packet P is deemed lost,
if packet Q of higher sequence is s/acked and P and Q are distant
by at least dupthresh number of packets in sequence space.
FACK is more aggressive than the IETF recommened recovery for SACK
(RFC3517 A Conservative Selective Acknowledgment (SACK)-based Loss
Recovery Algorithm for TCP), because a single SACK may trigger
fast recovery. This obviously won't work well with reordering so
FACK is dynamically disabled upon detecting reordering.
RACK supersedes FACK by using time distance instead of sequence
distance. On reordering, RACK waits for a quarter of RTT receiving
a single SACK before starting recovery. (the timer can be made more
adaptive in the future by measuring reordering distance in time,
but currently RTT/4 seem to work well.) Once the recovery starts,
RACK behaves almost like FACK because it reduces the reodering
window to 1ms, so it fast retransmits quickly. In addition RACK
can detect loss retransmission as it does not care about the packet
sequences (being repeated or not), which is extremely useful when
the connection is going through a traffic policer.
Google server experiments indicate that disabling FACK after enabling
RACK has negligible impact on the overall loss recovery performance
with more reordering events detected. But we still keep the FACK
implementation for backup if RACK has bugs that needs to be disabled.
[1] M. Mathis, J. Mahdavi, "Forward Acknowledgment: Refining
TCP Congestion Control," In Proceedings of SIGCOMM '96, August 1996.
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Acked-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Thin stream DUPACK is to start fast recovery on only one DUPACK
provided the connection is a thin stream (i.e., low inflight). But
this older feature is now subsumed with RACK. If a connection
receives only a single DUPACK, RACK would arm a reordering timer
and soon starts fast recovery instead of timeout if no further
ACKs are received.
The socket option (THIN_DUPACK) is kept as a nop for compatibility.
Note that this patch does not change another thin-stream feature
which enables linear RTO. Although it might be good to generalize
that in the future (i.e., linear RTO for the first say 3 retries).
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Acked-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch removes the (partial) implementation of the aggressive
limited transmit in RFC4653 TCP Non-Congestion Robustness (NCR).
NCR is a mitigation to the problem created by the dynamic
DUPACK threshold. With the current adaptive DUPACK threshold
(tp->reordering) could cause timeouts by preventing fast recovery.
For example, if the last packet of a cwnd burst was reordered, the
threshold will be set to the size of cwnd. But if next application
burst is smaller than threshold and has drops instead of reorderings,
the sender would not trigger fast recovery but instead resorts to a
timeout recovery.
NCR mitigates this issue by checking the number of DUPACKs against
the current flight size additionally. The techniqueue is similar to
the early retransmit RFC.
With RACK loss detection, this mitigation is not needed, because RACK
does not use DUPACK threshold to detect losses. RACK arms a reordering
timer to fire at most a quarter RTT later to start fast recovery.
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Acked-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch removes the support of RFC5827 early retransmit (i.e.,
fast recovery on small inflight with <3 dupacks) because it is
subsumed by the new RACK loss detection. More specifically when
RACK receives DUPACKs, it'll arm a reordering timer to start fast
recovery after a quarter of (min)RTT, hence it covers the early
retransmit except RACK does not limit itself to specific inflight
or dupack numbers.
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Acked-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Forward retransmit is an esoteric feature in RFC3517 (condition(3)
in the NextSeg()). Basically if a packet is not considered lost by
the current criteria (# of dupacks etc), but the congestion window
has room for more packets, then retransmit this packet.
However it actually conflicts with the rest of recovery design. For
example, when reordering is detected we want to be conservative
in retransmitting packets but forward-retransmit feature would
break that to force more retransmission. Also the implementation is
fairly complicated inside the retransmission logic inducing extra
iterations in the write queue. With RACK losses are being detected
timely and this heuristic is no longer necessary. There this patch
removes the feature.
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Acked-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Current F-RTO reverts cwnd reset whenever a never-retransmitted
packet was (s)acked. The timeout can be declared spurious because
the packets acknoledged with this ACK was transmitted before the
timeout, so clearly not all the packets are lost to reset the cwnd.
This nice detection does not really depend F-RTO internals. This
patch applies the detection universally. On Google servers this
change detected 20% more spurious timeouts.
Suggested-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Acked-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch changes two things:
1. Start fast recovery with RACK in addition to other heuristics
(e.g., DUPACK threshold, FACK). Prior to this change RACK
is enabled to detect losses only after the recovery has
started by other algorithms.
2. Disable TCP early retransmit. RACK subsumes the early retransmit
with the new reordering timer feature. A latter patch in this
series removes the early retransmit code.
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Acked-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Currently RACK would mark loss before the undo operations in TCP
loss recovery. This could incorrectly identify real losses as
spurious. For example a sender first experiences a delay spike and
then eventually some packets were lost due to buffer overrun.
In this case, the sender should perform fast recovery b/c not all
the packets were lost.
But the sender may first trigger a (spurious) RTO and reset
cwnd to 1. The following ACKs may used to mark real losses by
tcp_rack_mark_lost. Then in tcp_process_loss this ACK could trigger
F-RTO undo condition and unmark real losses and revert the cwnd
reduction. If there are no more ACKs coming back, eventually the
sender would timeout again instead of performing fast recovery.
The patch fixes this incorrect process by always performing
the undo checks before detecting losses.
Fixes: 4f41b1c58a ("tcp: use RACK to detect losses")
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Acked-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The packets inside a jumbo skb (e.g., TSO) share the same skb
timestamp, even though they are sent sequentially on the wire. Since
RACK is based on time, it can not detect some packets inside the
same skb are lost. However, we can leverage the packet sequence
numbers as extended timestamps to detect losses. Therefore, when
RACK timestamp is identical to skb's timestamp (i.e., one of the
packets of the skb is acked or sacked), we use the sequence numbers
of the acked and unacked packets to break ties.
We can use the same sequence logic to advance RACK xmit time as
well to detect more losses and avoid timeout.
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Acked-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch makes RACK install a reordering timer when it suspects
some packets might be lost, but wants to delay the decision
a little bit to accomodate reordering.
It does not create a new timer but instead repurposes the existing
RTO timer, because both are meant to retransmit packets.
Specifically it arms a timer ICSK_TIME_REO_TIMEOUT when
the RACK timing check fails. The wait time is set to
RACK.RTT + RACK.reo_wnd - (NOW - Packet.xmit_time) + fudge
This translates to expecting a packet (Packet) should take
(RACK.RTT + RACK.reo_wnd + fudge) to deliver after it was sent.
When there are multiple packets that need a timer, we use one timer
with the maximum timeout. Therefore the timer conservatively uses
the maximum window to expire N packets by one timeout, instead of
N timeouts to expire N packets sent at different times.
The fudge factor is 2 jiffies to ensure when the timer fires, all
the suspected packets would exceed the deadline and be marked lost
by tcp_rack_detect_loss(). It has to be at least 1 jiffy because the
clock may tick between calling icsk_reset_xmit_timer(timeout) and
actually hang the timer. The next jiffy is to lower-bound the timeout
to 2 jiffies when reo_wnd is < 1ms.
When the reordering timer fires (tcp_rack_reo_timeout): If we aren't
in Recovery we'll enter fast recovery and force fast retransmit.
This is very similar to the early retransmit (RFC5827) except RACK
is not constrained to only enter recovery for small outstanding
flights.
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Acked-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Record the most recent RTT in RACK. It is often identical to the
"ca_rtt_us" values in tcp_clean_rtx_queue. But when the packet has
been retransmitted, RACK choses to believe the ACK is for the
(latest) retransmitted packet if the RTT is over minimum RTT.
This requires passing the arrival time of the most recent ACK to
RACK routines. The timestamp is now recorded in the "ack_time"
in tcp_sacktag_state during the ACK processing.
This patch does not change the RACK algorithm itself. It only adds
the RTT variable to prepare the next main patch.
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Acked-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Create a new helper tcp_rack_detect_loss to prepare the upcoming
RACK reordering timer patch.
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Acked-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Different namespace application might require different maximal
number of remembered connection requests.
Signed-off-by: Haishuang Yan <yanhaishuang@cmss.chinamobile.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Different namespace application might require fast recycling
TIME-WAIT sockets independently of the host.
Signed-off-by: Haishuang Yan <yanhaishuang@cmss.chinamobile.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
There have been some reports lately about TCP connection stalls caused
by NIC drivers that aren't setting gso_size on aggregated packets on rx
path. This causes TCP to assume that the MSS is actually the size of the
aggregated packet, which is invalid.
Although the proper fix is to be done at each driver, it's often hard
and cumbersome for one to debug, come to such root cause and report/fix
it.
This patch amends this situation in two ways. First, it adds a warning
on when this situation occurs, so it gives a hint to those trying to
debug this. It also limit the maximum probed MSS to the adverised MSS,
as it should never be any higher than that.
The result is that the connection may not have the best performance ever
but it shouldn't stall, and the admin will have a hint on what to look
for.
Tested with virtio by forcing gso_size to 0.
v2: updated msg per David's suggestion
v3: use skb_iif to find the interface and also log its name, per Eric
Dumazet's suggestion. As the skb may be backlogged and the interface
gone by then, we need to check if the number still has a meaning.
v4: use helper tcp_gro_dev_warn() and avoid pr_warn_once inside __once, per
David's suggestion
Cc: Jonathan Maxwell <jmaxwell37@gmail.com>
Signed-off-by: Marcelo Ricardo Leitner <marcelo.leitner@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
