This commit eliminates the use of the link 'stale_limit' & 'prev_from'
(besides the already removed - 'stale_cnt') variables in the detection
of repeated retransmit failures as there is no proper way to initialize
them to avoid a false detection, i.e. it is not really a retransmission
failure but due to a garbage values in the variables.
Instead, a jiffies variable will be added to individual skbs (like the
way we restrict the skb retransmissions) in order to mark the first skb
retransmit time. Later on, at the next retransmissions, the timestamp
will be checked to see if the skb in the link transmq is "too stale",
that is, the link tolerance time has passed, so that a link reset will
be ordered. Note, just checking on the first skb in the queue is fine
enough since it must be the oldest one.
A counter is also added to keep track the actual skb retransmissions'
number for later checking when the failure happens.
The downside of this approach is that the skb->cb[] buffer is about to
be exhausted, however it is always able to allocate another memory area
and keep a reference to it when needed.
Fixes: 77cf8edbc0 ("tipc: simplify stale link failure criteria")
Reported-by: Hoang Le <hoang.h.le@dektech.com.au>
Acked-by: Ying Xue <ying.xue@windriver.com>
Acked-by: Jon Maloy <jon.maloy@ericsson.com>
Signed-off-by: Tuong Lien <tuong.t.lien@dektech.com.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
In conjunction with changing the interfaces' MTU (e.g. especially in
the case of a bonding) where the TIPC links are brought up and down
in a short time, a couple of issues were detected with the current link
changeover mechanism:
1) When one link is up but immediately forced down again, the failover
procedure will be carried out in order to failover all the messages in
the link's transmq queue onto the other working link. The link and node
state is also set to FAILINGOVER as part of the process. The message
will be transmited in form of a FAILOVER_MSG, so its size is plus of 40
bytes (= the message header size). There is no problem if the original
message size is not larger than the link's MTU - 40, and indeed this is
the max size of a normal payload messages. However, in the situation
above, because the link has just been up, the messages in the link's
transmq are almost SYNCH_MSGs which had been generated by the link
synching procedure, then their size might reach the max value already!
When the FAILOVER_MSG is built on the top of such a SYNCH_MSG, its size
will exceed the link's MTU. As a result, the messages are dropped
silently and the failover procedure will never end up, the link will
not be able to exit the FAILINGOVER state, so cannot be re-established.
2) The same scenario above can happen more easily in case the MTU of
the links is set differently or when changing. In that case, as long as
a large message in the failure link's transmq queue was built and
fragmented with its link's MTU > the other link's one, the issue will
happen (there is no need of a link synching in advance).
3) The link synching procedure also faces with the same issue but since
the link synching is only started upon receipt of a SYNCH_MSG, dropping
the message will not result in a state deadlock, but it is not expected
as design.
The 1) & 3) issues are resolved by the last commit that only a dummy
SYNCH_MSG (i.e. without data) is generated at the link synching, so the
size of a FAILOVER_MSG if any then will never exceed the link's MTU.
For the 2) issue, the only solution is trying to fragment the messages
in the failure link's transmq queue according to the working link's MTU
so they can be failovered then. A new function is made to accomplish
this, it will still be a TUNNEL PROTOCOL/FAILOVER MSG but if the
original message size is too large, it will be fragmented & reassembled
at the receiving side.
Acked-by: Ying Xue <ying.xue@windriver.com>
Acked-by: Jon Maloy <jon.maloy@ericsson.com>
Signed-off-by: Tuong Lien <tuong.t.lien@dektech.com.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
This commit along with the next one are to resolve the issues with the
link changeover mechanism. See that commit for details.
Basically, for the link synching, from now on, we will send only one
single ("dummy") SYNCH message to peer. The SYNCH message does not
contain any data, just a header conveying the synch point to the peer.
A new node capability flag ("TIPC_TUNNEL_ENHANCED") is introduced for
backward compatible!
Acked-by: Ying Xue <ying.xue@windriver.com>
Acked-by: Jon Maloy <jon.maloy@ericsson.com>
Suggested-by: Jon Maloy <jon.maloy@ericsson.com>
Signed-off-by: Tuong Lien <tuong.t.lien@dektech.com.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
We rename the inline function msg_get_wrapped() to the more
comprehensible msg_inner_hdr().
Signed-off-by: Jon Maloy <jon.maloy@ericsson.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
For unicast transmission, the current NACK sending althorithm is over-
active that forces the sending side to retransmit a packet that is not
really lost but just arrived at the receiving side with some delay, or
even retransmit same packets that have already been retransmitted
before. As a result, many duplicates are observed also under normal
condition, ie. without packet loss.
One example case is: node1 transmits 1 2 3 4 10 5 6 7 8 9, when node2
receives packet #10, it puts into the deferdq. When the packet #5 comes
it sends NACK with gap [6 - 9]. However, shortly after that, when
packet #6 arrives, it pulls out packet #10 from the deferfq, but it is
still out of order, so it makes another NACK with gap [7 - 9] and so on
... Finally, node1 has to retransmit the packets 5 6 7 8 9 a number of
times, but in fact all the packets are not lost at all, so duplicates!
This commit reduces duplicates by changing the condition to send NACK,
also restricting the retransmissions on individual packets via a timer
of about 1ms. However, it also needs to say that too tricky condition
for NACKs or too long timeout value for retransmissions will result in
performance reducing! The criterias in this commit are found to be
effective for both the requirements to reduce duplicates but not affect
performance.
The tipc_link_rcv() is also improved to only dequeue skb from the link
deferdq if it is expected (ie. its seqno <= rcv_nxt).
Acked-by: Ying Xue <ying.xue@windriver.com>
Acked-by: Jon Maloy <jon.maloy@ericsson.com>
Signed-off-by: Tuong Lien <tuong.t.lien@dektech.com.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
During unicast link transmission, it's observed very often that because
of one or a few lost/dis-ordered packets, the sending side will fastly
reach the send window limit and must wait for the packets to be arrived
at the receiving side or in the worst case, a retransmission must be
done first. The sending side cannot release a lot of subsequent packets
in its transmq even though all of them might have already been received
by the receiving side.
That is, one or two packets dis-ordered/lost and dozens of packets have
to wait, this obviously reduces the overall throughput!
This commit introduces an algorithm to overcome this by using "Gap ACK
blocks". Basically, a Gap ACK block will consist of <ack, gap> numbers
that describes the link deferdq where packets have been got by the
receiving side but with gaps, for example:
link deferdq: [1 2 3 4 10 11 13 14 15 20]
--> Gap ACK blocks: <4, 5>, <11, 1>, <15, 4>, <20, 0>
The Gap ACK blocks will be sent to the sending side along with the
traditional ACK or NACK message. Immediately when receiving the message
the sending side will now not only release from its transmq the packets
ack-ed by the ACK but also by the Gap ACK blocks! So, more packets can
be enqueued and transmitted.
In addition, the sending side can now do "multi-retransmissions"
according to the Gaps reported in the Gap ACK blocks.
The new algorithm as verified helps greatly improve the TIPC throughput
especially under packet loss condition.
So far, a maximum of 32 blocks is quite enough without any "Too few Gap
ACK blocks" reports with a 5.0% packet loss rate, however this number
can be increased in the furture if needed.
Also, the patch is backward compatible.
Acked-by: Ying Xue <ying.xue@windriver.com>
Acked-by: Jon Maloy <jon.maloy@ericsson.com>
Signed-off-by: Tuong Lien <tuong.t.lien@dektech.com.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
Currently, a multicast stream may start out using replicast, because
there are few destinations, and then it should ideally switch to
L2/broadcast IGMP/multicast when the number of destinations grows beyond
a certain limit. The opposite should happen when the number decreases
below the limit.
To eliminate the risk of message reordering caused by method change,
a sending socket must stick to a previously selected method until it
enters an idle period of 5 seconds. Means there is a 5 seconds pause
in the traffic from the sender socket.
If the sender never makes such a pause, the method will never change,
and transmission may become very inefficient as the cluster grows.
With this commit, we allow such a switch between replicast and
broadcast without any need for a traffic pause.
Solution is to send a dummy message with only the header, also with
the SYN bit set, via broadcast or replicast. For the data message,
the SYN bit is set and sending via replicast or broadcast (inverse
method with dummy).
Then, at receiving side any messages follow first SYN bit message
(data or dummy message), they will be held in deferred queue until
another pair (dummy or data message) arrived in other link.
v2: reverse christmas tree declaration
Acked-by: Jon Maloy <jon.maloy@ericsson.com>
Signed-off-by: Hoang Le <hoang.h.le@dektech.com.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
When a link endpoint is re-created (e.g. after a node reboot or
interface reset), the link session number is varied by random, the peer
endpoint will be synced with this new session number before the link is
re-established.
However, there is a shortcoming in this mechanism that can lead to the
link never re-established or faced with a failure then. It happens when
the peer endpoint is ready in ESTABLISHING state, the 'peer_session' as
well as the 'in_session' flag have been set, but suddenly this link
endpoint leaves. When it comes back with a random session number, there
are two situations possible:
1/ If the random session number is larger than (or equal to) the
previous one, the peer endpoint will be updated with this new session
upon receipt of a RESET_MSG from this endpoint, and the link can be re-
established as normal. Otherwise, all the RESET_MSGs from this endpoint
will be rejected by the peer. In turn, when this link endpoint receives
one ACTIVATE_MSG from the peer, it will move to ESTABLISHED and start
to send STATE_MSGs, but again these messages will be dropped by the
peer due to wrong session.
The peer link endpoint can still become ESTABLISHED after receiving a
traffic message from this endpoint (e.g. a BCAST_PROTOCOL or
NAME_DISTRIBUTOR), but since all the STATE_MSGs are invalid, the link
will be forced down sooner or later!
Even in case the random session number is larger than the previous one,
it can be that the ACTIVATE_MSG from the peer arrives first, and this
link endpoint moves quickly to ESTABLISHED without sending out any
RESET_MSG yet. Consequently, the peer link will not be updated with the
new session number, and the same link failure scenario as above will
happen.
2/ Another situation can be that, the peer link endpoint was reset due
to any reasons in the meantime, its link state was set to RESET from
ESTABLISHING but still in session, i.e. the 'in_session' flag is not
reset...
Now, if the random session number from this endpoint is less than the
previous one, all the RESET_MSGs from this endpoint will be rejected by
the peer. In the other direction, when this link endpoint receives a
RESET_MSG from the peer, it moves to ESTABLISHING and starts to send
ACTIVATE_MSGs, but all these messages will be rejected by the peer too.
As a result, the link cannot be re-established but gets stuck with this
link endpoint in state ESTABLISHING and the peer in RESET!
Solution:
===========
This link endpoint should not go directly to ESTABLISHED when getting
ACTIVATE_MSG from the peer which may belong to the old session if the
link was re-created. To ensure the session to be correct before the
link is re-established, the peer endpoint in ESTABLISHING state will
send back the last session number in ACTIVATE_MSG for a verification at
this endpoint. Then, if needed, a new and more appropriate session
number will be regenerated to force a re-synch first.
In addition, when a link in ESTABLISHING state is reset, its state will
move to RESET according to the link FSM, along with resetting the
'in_session' flag (and the other data) as a normal link reset, it will
also be deleted if requested.
The solution is backward compatible.
Acked-by: Jon Maloy <jon.maloy@ericsson.com>
Acked-by: Ying Xue <ying.xue@windriver.com>
Signed-off-by: Tuong Lien <tuong.t.lien@dektech.com.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
Currently, the broadcast retransmission algorithm is using the
'prev_retr' field in struct tipc_link to time stamp the latest broadcast
retransmission occasion. This helps to restrict retransmission of
individual broadcast packets to max once per 10 milliseconds, even
though all other criteria for retransmission are met.
We now move this time stamp to the control block of each individual
packet, and remove other limiting criteria. This simplifies the
retransmission algorithm, and eliminates any risk of logical errors
in selecting which packets can be retransmitted.
Acked-by: Ying Xue <ying.xue@windriver.com>
Signed-off-by: LUU Duc Canh <canh.d.luu@dektech.com.au>
Signed-off-by: Jon Maloy <jon.maloy@ericsson.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Default socket receive buffer size for a listener socket is 2Mb. For
each arriving empty SYN, the linux kernel allocates a 768 bytes buffer.
This means that a listener socket can serve maximum 2700 simultaneous
empty connection setup requests before it hits a receive buffer
overflow, and much fewer if the SYN is carrying any significant
amount of data.
When this happens the setup request is rejected, and the client
receives an ECONNREFUSED error.
This commit mitigates this problem by letting the client socket try to
retransmit the SYN message multiple times when it sees it rejected with
the code TIPC_ERR_OVERLOAD. Retransmission is done at random intervals
in the range of [100 ms, setup_timeout / 4], as many times as there is
room for within the setup timeout limit.
Signed-off-by: Tung Nguyen <tung.q.nguyen@dektech.com.au>
Acked-by: Ying Xue <ying.xue@windriver.com>
Signed-off-by: Jon Maloy <jon.maloy@ericsson.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Messages intended for intitating a connection are currently
indistinguishable from regular datagram messages. The TIPC
protocol specification defines bit 17 in word 0 as a SYN bit
to allow sanity check of such messages in the listening socket,
but this has so far never been implemented.
We do that in this commit.
Acked-by: Ying Xue <ying.xue@windriver.com>
Signed-off-by: Jon Maloy <jon.maloy@ericsson.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
When a 32-bit node address is generated from a 128-bit identifier,
there is a risk of collisions which must be discovered and handled.
We do this as follows:
- We don't apply the generated address immediately to the node, but do
instead initiate a 1 sec trial period to allow other cluster members
to discover and handle such collisions.
- During the trial period the node periodically sends out a new type
of message, DSC_TRIAL_MSG, using broadcast or emulated broadcast,
to all the other nodes in the cluster.
- When a node is receiving such a message, it must check that the
presented 32-bit identifier either is unused, or was used by the very
same peer in a previous session. In both cases it accepts the request
by not responding to it.
- If it finds that the same node has been up before using a different
address, it responds with a DSC_TRIAL_FAIL_MSG containing that
address.
- If it finds that the address has already been taken by some other
node, it generates a new, unused address and returns it to the
requester.
- During the trial period the requesting node must always be prepared
to accept a failure message, i.e., a message where a peer suggests a
different (or equal) address to the one tried. In those cases it
must apply the suggested value as trial address and restart the trial
period.
This algorithm ensures that in the vast majority of cases a node will
have the same address before and after a reboot. If a legacy user
configures the address explicitly, there will be no trial period and
messages, so this protocol addition is completely backwards compatible.
Acked-by: Ying Xue <ying.xue@windriver.com>
Signed-off-by: Jon Maloy <jon.maloy@ericsson.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
When sending node local messages the code is using an 'mtu' of 66060
bytes to avoid unnecessary fragmentation. During situations of low
memory tipc_msg_build() may sometimes fail to allocate such large
buffers, resulting in unnecessary send failures. This can easily be
remedied by falling back to a smaller MTU, and then reassemble the
buffer chain as if the message were arriving from a remote node.
At the same time, we change the initial MTU setting of the broadcast
link to a lower value, so that large messages always are fragmented
into smaller buffers even when we run in single node mode. Apart from
obtaining the same advantage as for the 'fallback' solution above, this
turns out to give a significant performance improvement. This can
probably be explained with the __pskb_copy() operation performed on the
buffer for each recipient during reception. We found the optimal value
for this, considering the most relevant skb pool, to be 3744 bytes.
Acked-by: Ying Xue <ying.xue@ericsson.com>
Signed-off-by: Jon Maloy <jon.maloy@ericsson.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The socket level flow control is based on the assumption that incoming
buffers meet the condition (skb->truesize / roundup(skb->len) <= 4),
where the latter value is rounded off upwards to the nearest 1k number.
This does empirically hold true for the device drivers we know, but we
cannot trust that it will always be so, e.g., in a system with jumbo
frames and very small packets.
We now introduce a check for this condition at packet arrival, and if
we find it to be false, we copy the packet to a new, smaller buffer,
where the condition will be true. We expect this to affect only a small
fraction of all incoming packets, if at all.
Acked-by: Ying Xue <ying.xue@windriver.com>
Signed-off-by: Jon Maloy <jon.maloy@ericsson.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Currently, the TIPC RPS dissector is based only on the incoming packets'
source node address, hence steering all traffic from a node to the same
core. We have seen that this makes the links vulnerable to starvation
and unnecessary resets when we turn down the link tolerance to very low
values.
To reduce the risk of this happening, we exempt probe and probe replies
packets from the convergence to one core per source node. Instead, we do
the opposite, - we try to diverge those packets across as many cores as
possible, by randomizing the flow selector key.
To make such packets identifiable to the dissector, we add a new
'is_keepalive' bit to word 0 of the LINK_PROTOCOL header. This bit is
set both for PROBE and PROBE_REPLY messages, and only for those.
It should be noted that these packets are not part of any flow anyway,
and only constitute a minuscule fraction of all packets sent across a
link. Hence, there is no risk that this will affect overall performance.
Acked-by: Ying Xue <ying.xue@windriver.com>
Signed-off-by: Jon Maloy <jon.maloy@ericsson.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
We already have point-to-multipoint flow control within a group. But
we even need the opposite; -a scheme which can handle that potentially
hundreds of sources may try to send messages to the same destination
simultaneously without causing buffer overflow at the recipient. This
commit adds such a mechanism.
The algorithm works as follows:
- When a member detects a new, joining member, it initially set its
state to JOINED and advertises a minimum window to the new member.
This window is chosen so that the new member can send exactly one
maximum sized message, or several smaller ones, to the recipient
before it must stop and wait for an additional advertisement. This
minimum window ADV_IDLE is set to 65 1kB blocks.
- When a member receives the first data message from a JOINED member,
it changes the state of the latter to ACTIVE, and advertises a larger
window ADV_ACTIVE = 12 x ADV_IDLE blocks to the sender, so it can
continue sending with minimal disturbances to the data flow.
- The active members are kept in a dedicated linked list. Each time a
message is received from an active member, it will be moved to the
tail of that list. This way, we keep a record of which members have
been most (tail) and least (head) recently active.
- There is a maximum number (16) of permitted simultaneous active
senders per receiver. When this limit is reached, the receiver will
not advertise anything immediately to a new sender, but instead put
it in a PENDING state, and add it to a corresponding queue. At the
same time, it will pick the least recently active member, send it an
advertisement RECLAIM message, and set this member to state
RECLAIMING.
- The reclaimee member has to respond with a REMIT message, meaning that
it goes back to a send window of ADV_IDLE, and returns its unused
advertised blocks beyond that value to the reclaiming member.
- When the reclaiming member receives the REMIT message, it unlinks
the reclaimee from its active list, resets its state to JOINED, and
notes that it is now back at ADV_IDLE advertised blocks to that
member. If there are still unread data messages sent out by
reclaimee before the REMIT, the member goes into an intermediate
state REMITTED, where it stays until the said messages have been
consumed.
- The returned advertised blocks can now be re-advertised to the
pending member, which is now set to state ACTIVE and added to
the active member list.
- To be proactive, i.e., to minimize the risk that any member will
end up in the pending queue, we start reclaiming resources already
when the number of active members exceeds 3/4 of the permitted
maximum.
Signed-off-by: Jon Maloy <jon.maloy@ericsson.com>
Acked-by: Ying Xue <ying.xue@windriver.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
We need a mechanism guaranteeing that group unicasts sent out from a
socket are not bypassed by later sent broadcasts from the same socket.
We do this as follows:
- Each time a unicast is sent, we set a the broadcast method for the
socket to "replicast" and "mandatory". This forces the first
subsequent broadcast message to follow the same network and data path
as the preceding unicast to a destination, hence preventing it from
overtaking the latter.
- In order to make the 'same data path' statement above true, we let
group unicasts pass through the multicast link input queue, instead
of as previously through the unicast link input queue.
- In the first broadcast following a unicast, we set a new header flag,
requiring all recipients to immediately acknowledge its reception.
- During the period before all the expected acknowledges are received,
the socket refuses to accept any more broadcast attempts, i.e., by
blocking or returning EAGAIN. This period should typically not be
longer than a few microseconds.
- When all acknowledges have been received, the sending socket will
open up for subsequent broadcasts, this time giving the link layer
freedom to itself select the best transmission method.
- The forced and/or abrupt transmission method changes described above
may lead to broadcasts arriving out of order to the recipients. We
remedy this by introducing code that checks and if necessary
re-orders such messages at the receiving end.
Signed-off-by: Jon Maloy <jon.maloy@ericsson.com>
Acked-by: Ying Xue <ying.xue@windriver.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The previously introduced message transport to all group members is
based on the tipc multicast service, but is logically a broadcast
service within the group, and that is what we call it.
We now add functionality for sending messages to all group members
having a certain identity. Correspondingly, we call this feature 'group
multicast'. The service is using unicast when only one destination is
found, otherwise it will use the bearer broadcast service to transfer
the messages. In the latter case, the receiving members filter arriving
messages by looking at the intended destination instance. If there is
no match, the message will be dropped, while still being considered
received and read as seen by the flow control mechanism.
Signed-off-by: Jon Maloy <jon.maloy@ericsson.com>
Acked-by: Ying Xue <ying.xue@windriver.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
We now make it possible to send connectionless unicast messages
within a communication group. To send a message, the sender can use
either a direct port address, aka port identity, or an indirect port
name to be looked up.
This type of messages are subject to the same start synchronization
and flow control mechanism as group broadcast messages.
Signed-off-by: Jon Maloy <jon.maloy@ericsson.com>
Acked-by: Ying Xue <ying.xue@windriver.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
We introduce an end-to-end flow control mechanism for group broadcast
messages. This ensures that no messages are ever lost because of
destination receive buffer overflow, with minimal impact on performance.
For now, the algorithm is based on the assumption that there is only one
active transmitter at any moment in time.
Signed-off-by: Jon Maloy <jon.maloy@ericsson.com>
Acked-by: Ying Xue <ying.xue@windriver.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Like with any other service, group members' availability can be
subscribed for by connecting to be topology server. However, because
the events arrive via a different socket than the member socket, there
is a real risk that membership events my arrive out of synch with the
actual JOIN/LEAVE action. I.e., it is possible to receive the first
messages from a new member before the corresponding JOIN event arrives,
just as it is possible to receive the last messages from a leaving
member after the LEAVE event has already been received.
Since each member socket is internally also subscribing for membership
events, we now fix this problem by passing those events on to the user
via the member socket. We leverage the already present member synch-
ronization protocol to guarantee correct message/event order. An event
is delivered to the user as an empty message where the two source
addresses identify the new/lost member. Furthermore, we set the MSG_OOB
bit in the message flags to mark it as an event. If the event is an
indication about a member loss we also set the MSG_EOR bit, so it can
be distinguished from a member addition event.
Signed-off-by: Jon Maloy <jon.maloy@ericsson.com>
Acked-by: Ying Xue <ying.xue@windriver.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
With group communication, it becomes important for a message receiver to
identify not only from which socket (identfied by a node:port tuple) the
message was sent, but also the logical identity (type:instance) of the
sending member.
We fix this by adding a second instance of struct sockaddr_tipc to the
source address area when a message is read. The extra address struct
is filled in with data found in the received message header (type,) and
in the local member representation struct (instance.)
Signed-off-by: Jon Maloy <jon.maloy@ericsson.com>
Acked-by: Ying Xue <ying.xue@windriver.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
As a preparation for introducing flow control for multicast and datagram
messaging we need a more strictly defined framework than we have now. A
socket must be able keep track of exactly how many and which other
sockets it is allowed to communicate with at any moment, and keep the
necessary state for those.
We therefore introduce a new concept we have named Communication Group.
Sockets can join a group via a new setsockopt() call TIPC_GROUP_JOIN.
The call takes four parameters: 'type' serves as group identifier,
'instance' serves as an logical member identifier, and 'scope' indicates
the visibility of the group (node/cluster/zone). Finally, 'flags' makes
it possible to set certain properties for the member. For now, there is
only one flag, indicating if the creator of the socket wants to receive
a copy of broadcast or multicast messages it is sending via the socket,
and if wants to be eligible as destination for its own anycasts.
A group is closed, i.e., sockets which have not joined a group will
not be able to send messages to or receive messages from members of
the group, and vice versa.
Any member of a group can send multicast ('group broadcast') messages
to all group members, optionally including itself, using the primitive
send(). The messages are received via the recvmsg() primitive. A socket
can only be member of one group at a time.
Signed-off-by: Jon Maloy <jon.maloy@ericsson.com>
Acked-by: Ying Xue <ying.xue@windriver.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
In the following commits we will need to handle multiple incoming and
rejected/returned buffers in the function socket.c::filter_rcv().
As a preparation for this, we generalize the function by handling
buffer queues instead of individual buffers. We also introduce a
help function tipc_skb_reject(), and rename filter_rcv() to
tipc_sk_filter_rcv() in line with other functions in socket.c.
Signed-off-by: Jon Maloy <jon.maloy@ericsson.com>
Acked-by: Ying Xue <ying.xue@windriver.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
As preparation for introducing communication groups, we add the ability
to issue topology subscriptions and receive topology events from kernel
space. This will make it possible for group member sockets to keep track
of other group members.
Signed-off-by: Jon Maloy <jon.maloy@ericsson.com>
Acked-by: Ying Xue <ying.xue@windriver.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
TIPC multicast messages are currently carried over a reliable
'broadcast link', making use of the underlying media's ability to
transport packets as L2 broadcast or IP multicast to all nodes in
the cluster.
When the used bearer is lacking that ability, we can instead emulate
the broadcast service by replicating and sending the packets over as
many unicast links as needed to reach all identified destinations.
We now introduce a new TIPC link-level 'replicast' service that does
this.
Reviewed-by: Parthasarathy Bhuvaragan <parthasarathy.bhuvaragan@ericsson.com>
Acked-by: Ying Xue <ying.xue@windriver.com>
Signed-off-by: Jon Maloy <jon.maloy@ericsson.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Until now, we allocate memory always with GFP_ATOMIC flag.
When the system is under memory pressure and a user tries to send,
the send fails due to low memory. However, the user application
can wait for free memory if we allocate it using GFP_KERNEL flag.
In this commit, we use allocate memory with GFP_KERNEL for all user
allocation.
Reported-by: Rune Torgersen <runet@innovsys.com>
Acked-by: Jon Maloy <jon.maloy@ericsson.com>
Signed-off-by: Parthasarathy Bhuvaragan <parthasarathy.bhuvaragan@ericsson.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The socket code currently handles link congestion by either blocking
and trying to send again when the congestion has abated, or just
returning to the user with -EAGAIN and let him re-try later.
This mechanism is prone to starvation, because the wakeup algorithm is
non-atomic. During the time the link issues a wakeup signal, until the
socket wakes up and re-attempts sending, other senders may have come
in between and occupied the free buffer space in the link. This in turn
may lead to a socket having to make many send attempts before it is
successful. In extremely loaded systems we have observed latency times
of several seconds before a low-priority socket is able to send out a
message.
In this commit, we simplify this mechanism and reduce the risk of the
described scenario happening. When a message is attempted sent via a
congested link, we now let it be added to the link's backlog queue
anyway, thus permitting an oversubscription of one message per source
socket. We still create a wakeup item and return an error code, hence
instructing the sender to block or stop sending. Only when enough space
has been freed up in the link's backlog queue do we issue a wakeup event
that allows the sender to continue with the next message, if any.
The fact that a socket now can consider a message sent even when the
link returns a congestion code means that the sending socket code can
be simplified. Also, since this is a good opportunity to get rid of the
obsolete 'mtu change' condition in the three socket send functions, we
now choose to refactor those functions completely.
Signed-off-by: Parthasarathy Bhuvaragan <parthasarathy.bhuvaragan@ericsson.com>
Acked-by: Ying Xue <ying.xue@windriver.com>
Signed-off-by: Jon Maloy <jon.maloy@ericsson.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
In this commit, we rename handle to bytes_read indicating the
purpose of the member.
Signed-off-by: Parthasarathy Bhuvaragan <parthasarathy.bhuvaragan@ericsson.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
In commit 2d18ac4ba7 ("tipc: extend broadcast link initialization
criteria") we tried to fix a problem with the initial synchronization
of broadcast link acknowledge values. Unfortunately that solution is
not sufficient to solve the issue.
We have seen it happen that LINK_PROTOCOL/STATE packets with a valid
non-zero unicast acknowledge number may bypass BCAST_PROTOCOL
initialization, NAME_DISTRIBUTOR and other STATE packets with invalid
broadcast acknowledge numbers, leading to premature opening of the
broadcast link. When the bypassed packets finally arrive, they are
inadvertently accepted, and the already correctly initialized
acknowledge number in the broadcast receive link is overwritten by
the invalid (zero) value of the said packets. After this the broadcast
link goes stale.
We now fix this by marking the packets where we know the acknowledge
value is or may be invalid, and then ignoring the acks from those.
To this purpose, we claim an unused bit in the header to indicate that
the value is invalid. We set the bit to 1 in the initial BCAST_PROTOCOL
synchronization packet and all initial ("bulk") NAME_DISTRIBUTOR
packets, plus those LINK_PROTOCOL packets sent out before the broadcast
links are fully synchronized.
This minor protocol update is fully backwards compatible.
Reported-by: John Thompson <thompa.atl@gmail.com>
Tested-by: John Thompson <thompa.atl@gmail.com>
Signed-off-by: Jon Maloy <jon.maloy@ericsson.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
When we send broadcasts in clusters of more 70-80 nodes, we sometimes
see the broadcast link resetting because of an excessive number of
retransmissions. This is caused by a combination of two factors:
1) A 'NACK crunch", where loss of broadcast packets is discovered
and NACK'ed by several nodes simultaneously, leading to multiple
redundant broadcast retransmissions.
2) The fact that the NACKS as such also are sent as broadcast, leading
to excessive load and packet loss on the transmitting switch/bridge.
This commit deals with the latter problem, by moving sending of
broadcast nacks from the dedicated BCAST_PROTOCOL/NACK message type
to regular unicast LINK_PROTOCOL/STATE messages. We allocate 10 unused
bits in word 8 of the said message for this purpose, and introduce a
new capability bit, TIPC_BCAST_STATE_NACK in order to keep the change
backwards compatible.
Reviewed-by: Ying Xue <ying.xue@windriver.com>
Signed-off-by: Jon Maloy <jon.maloy@ericsson.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
When extracting an individual message from a received "bundle" buffer,
we just create a clone of the base buffer, and adjust it to point into
the right position of the linearized data area of the latter. This works
well for regular message reception, but during periods of extremely high
load it may happen that an extracted buffer, e.g, a connection probe, is
reversed and forwarded through an external interface while the preceding
extracted message is still unhandled. When this happens, the header or
data area of the preceding message will be partially overwritten by a
MAC header, leading to unpredicatable consequences, such as a link
reset.
We now fix this by ensuring that the msg_reverse() function never
returns a cloned buffer, and that the returned buffer always contains
sufficient valid head and tail room to be forwarded.
Reported-by: Erik Hugne <erik.hugne@gmail.com>
Acked-by: Ying Xue <ying.xue@windriver.com>
Signed-off-by: Jon Maloy <jon.maloy@ericsson.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
There are two flow control mechanisms in TIPC; one at link level that
handles network congestion, burst control, and retransmission, and one
at connection level which' only remaining task is to prevent overflow
in the receiving socket buffer. In TIPC, the latter task has to be
solved end-to-end because messages can not be thrown away once they
have been accepted and delivered upwards from the link layer, i.e, we
can never permit the receive buffer to overflow.
Currently, this algorithm is message based. A counter in the receiving
socket keeps track of number of consumed messages, and sends a dedicated
acknowledge message back to the sender for each 256 consumed message.
A counter at the sending end keeps track of the sent, not yet
acknowledged messages, and blocks the sender if this number ever reaches
512 unacknowledged messages. When the missing acknowledge arrives, the
socket is then woken up for renewed transmission. This works well for
keeping the message flow running, as it almost never happens that a
sender socket is blocked this way.
A problem with the current mechanism is that it potentially is very
memory consuming. Since we don't distinguish between small and large
messages, we have to dimension the socket receive buffer according
to a worst-case of both. I.e., the window size must be chosen large
enough to sustain a reasonable throughput even for the smallest
messages, while we must still consider a scenario where all messages
are of maximum size. Hence, the current fix window size of 512 messages
and a maximum message size of 66k results in a receive buffer of 66 MB
when truesize(66k) = 131k is taken into account. It is possible to do
much better.
This commit introduces an algorithm where we instead use 1024-byte
blocks as base unit. This unit, always rounded upwards from the
actual message size, is used when we advertise windows as well as when
we count and acknowledge transmitted data. The advertised window is
based on the configured receive buffer size in such a way that even
the worst-case truesize/msgsize ratio always is covered. Since the
smallest possible message size (from a flow control viewpoint) now is
1024 bytes, we can safely assume this ratio to be less than four, which
is the value we are now using.
This way, we have been able to reduce the default receive buffer size
from 66 MB to 2 MB with maintained performance.
In order to keep this solution backwards compatible, we introduce a
new capability bit in the discovery protocol, and use this throughout
the message sending/reception path to always select the right unit.
Acked-by: Ying Xue <ying.xue@windriver.com>
Signed-off-by: Jon Maloy <jon.maloy@ericsson.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
When a link endpoint is going down locally, e.g., because its interface
is being stopped, it will spontaneously send out a RESET message to
its peer, informing it about this fact. This saves the peer from
detecting the failure via probing, and hence gives both speedier and
less resource consuming failure detection on the peer side.
According to the link FSM, a receiver of a RESET message, ignoring the
reason for it, must now consider the sender ready to come back up, and
starts periodically sending out ACTIVATE messages to the peer in order
to re-establish the link. Also, according to the FSM, the receiver of
an ACTIVATE message can now go directly to state ESTABLISHED and start
sending regular traffic packets. This is a well-proven and robust FSM.
However, in the case of a reboot, there is a small possibilty that link
endpoint on the rebooted node may have been re-created with a new bearer
identity between the moment it sent its (pre-boot) RESET and the moment
it receives the ACTIVATE from the peer. The new bearer identity cannot
be known by the peer according to this scenario, since traffic headers
don't convey such information. This is a problem, because both endpoints
need to know the correct value of the peer's bearer id at any moment in
time in order to be able to produce correct link events for their users.
The only way to guarantee this is to enforce a full setup message
exchange (RESET + ACTIVATE) even after the reboot, since those messages
carry the bearer idientity in their header.
In this commit we do this by introducing and setting a "stopping" bit in
the header of the spontaneously generated RESET messages, informing the
peer that the sender will not be immediately ready to re-establish the
link. A receiver seeing this bit must act as if this were a locally
detected connectivity failure, and hence has to go through a full two-
way setup message exchange before any link can be re-established.
Although never reported, this problem seems to have always been around.
This protocol addition is fully backwards compatible.
Acked-by: Ying Xue <ying.xue@windriver.com>
Signed-off-by: Jon Maloy <jon.maloy@ericsson.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Resetting a bearer/interface, with the consequence of resetting all its
pertaining links, is not an atomic action. This becomes particularly
evident in very large clusters, where a lot of traffic may happen on the
remaining links while we are busy shutting them down. In extreme cases,
we may even see links being re-created and re-established before we are
finished with the job.
To solve this, we now introduce a solution where we temporarily detach
the bearer from the interface when the bearer is reset. This inhibits
all packet reception, while sending still is possible. For the latter,
we use the fact that the device's user pointer now is zero to filter out
which packets can be sent during this situation; i.e., outgoing RESET
messages only. This filtering serves to speed up the neighbors'
detection of the loss event, and saves us from unnecessary probing.
Acked-by: Ying Xue <ying.xue@windriver.com>
Signed-off-by: Jon Maloy <jon.maloy@ericsson.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The code path for receiving broadcast packets is currently distinct
from the unicast path. This leads to unnecessary code and data
duplication, something that can be avoided with some effort.
We now introduce separate per-peer tipc_link instances for handling
broadcast packet reception. Each receive link keeps a pointer to the
common, single, broadcast link instance, and can hence handle release
and retransmission of send buffers as if they belonged to the own
instance.
Furthermore, we let each unicast link instance keep a reference to both
the pertaining broadcast receive link, and to the common send link.
This makes it possible for the unicast links to easily access data for
broadcast link synchronization, as well as for carrying acknowledges for
received broadcast packets.
Signed-off-by: Jon Maloy <jon.maloy@ericsson.com>
Reviewed-by: Ying Xue <ying.xue@windriver.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This commit simplifies the broadcast link transmission function, by
leveraging previous changes to the link transmission function and the
broadcast transmission link life cycle.
Signed-off-by: Jon Maloy <jon.maloy@ericsson.com>
Reviewed-by: Ying Xue <ying.xue@windriver.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Realizing that unicast is just a special case of broadcast, we also see
that we can go in the other direction, i.e., that modest changes to the
current unicast link can make it generic enough to support broadcast.
The following changes are introduced here:
- A new counter ("ackers") in struct tipc_link, to indicate how many
peers need to ack a packet before it can be released.
- A corresponding counter in the skb user area, to keep track of how
many peers a are left to ack before a buffer can be released.
- A new counter ("acked"), to keep persistent track of how far a peer
has acked at the moment, i.e., where in the transmission queue to
start updating buffers when the next ack arrives. This is to avoid
double acknowledgements from a peer, with inadvertent relase of
packets as a result.
- A more generic tipc_link_retrans() function, where retransmit starts
from a given sequence number, instead of the first packet in the
transmision queue. This is to minimize the number of retransmitted
packets on the broadcast media.
When the new functionality is taken into use in the next commits,
we expect it to have minimal effect on unicast mode performance.
Signed-off-by: Jon Maloy <jon.maloy@ericsson.com>
Reviewed-by: Ying Xue <ying.xue@windriver.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Conflicts:
drivers/net/usb/asix_common.c
net/ipv4/inet_connection_sock.c
net/switchdev/switchdev.c
In the inet_connection_sock.c case the request socket hashing scheme
is completely different in net-next.
The other two conflicts were overlapping changes.
Signed-off-by: David S. Miller <davem@davemloft.net>
After the previous commits, we are guaranteed that no packets
of type LINK_PROTOCOL or with illegal sequence numbers will be
attempted added to the link deferred queue. This makes it possible to
make some simplifications to the sorting algorithm in the function
tipc_skb_queue_sorted().
We also alter the function so that it will drop packets if one with
the same seqeunce number is already present in the queue. This is
necessary because we have identified weird packet sequences, involving
duplicate packets, where a legitimate in-sequence packet may advance to
the head of the queue without being detected and de-queued.
Finally, we make this function outline, since it will now be called only
in exceptional cases.
Signed-off-by: Jon Maloy <jon.maloy@ericsson.com>
Acked-by: Ying Xue <ying.xue@windriver.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
In commit e3eea1eb47 ("tipc: clean up handling of message priorities")
we introduced a field in the packet header for keeping track of the
priority of fragments, since this value is not present in the specified
protocol header. Since the value so far only is used at the transmitting
end of the link, we have not yet officially defined it as part of the
protocol.
Unfortunately, the field we use for keeping this value, bits 13-15 in
in word 5, has turned out to be a poor choice; it is already used by the
broadcast protocol for carrying the 'network id' field of the sending
node. Since packet fragments also need to be transported across the
broadcast protocol, the risk of conflict is obvious, and we see this
happen when we use network identities larger than 2^13-1. This has
escaped our testing because we have so far only been using small network
id values.
We now move this field to bits 0-2 in word 9, a field that is guaranteed
to be unused by all involved protocols.
Fixes: e3eea1eb47 ("tipc: clean up handling of message priorities")
Signed-off-by: Jon Maloy <jon.maloy@ericsson.com>
Acked-by: Ying Xue <ying.xue@windriver.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
After the most recent changes, all access calls to a link which
may entail addition of messages to the link's input queue are
postpended by an explicit call to tipc_sk_rcv(), using a reference
to the correct queue.
This means that the potentially hazardous implicit delivery, using
tipc_node_unlock() in combination with a binary flag and a cached
queue pointer, now has become redundant.
This commit removes this implicit delivery mechanism both for regular
data messages and for binding table update messages.
Tested-by: Ying Xue <ying.xue@windriver.com>
Signed-off-by: Jon Maloy <jon.maloy@ericsson.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
In order to facilitate future improvements to the locking structure, we
want to make resetting and establishing of links non-atomic. I.e., the
functions tipc_node_link_up() and tipc_node_link_down() should be called
from outside the node lock context, and grab/release the node lock
themselves. This requires that we can freeze the link state from the
moment it is set to RESETTING or PEER_RESET in one lock context until
it is set to RESET or ESTABLISHING in a later context. The recently
introduced link FSM makes this possible, so we are now ready to introduce
the above change.
This commit implements this.
Tested-by: Ying Xue <ying.xue@windriver.com>
Signed-off-by: Jon Maloy <jon.maloy@ericsson.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Link failover and synchronization have until now been handled by the
links themselves, forcing them to have knowledge about and to access
parallel links in order to make the two algorithms work correctly.
In this commit, we move the control part of this functionality to the
link aggregation level in node.c, which is the right location for this.
As a result, the two algorithms become easier to follow, and the link
implementation becomes simpler.
Tested-by: Ying Xue <ying.xue@windriver.com>
Signed-off-by: Jon Maloy <jon.maloy@ericsson.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
When a message is received in a socket, one of the call chains
tipc_sk_rcv()->tipc_sk_enqueue()->filter_rcv()(->tipc_sk_proto_rcv())
or
tipc_sk_backlog_rcv()->filter_rcv()(->tipc_sk_proto_rcv())
are followed. At each of these levels we may encounter situations
where the message may need to be rejected, or a new message
produced for transfer back to the sender. Despite recent
improvements, the current code for doing this is perceived
as awkward and hard to follow.
Leveraging the two previous commits in this series, we now
introduce a more uniform handling of such situations. We
let each of the functions in the chain itself produce/reverse
the message to be returned to the sender, but also perform the
actual forwarding. This simplifies the necessary logics within
each function.
Reviewed-by: Ying Xue <ying.xue@windriver.com>
Signed-off-by: Jon Maloy <jon.maloy@ericsson.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Currently, we use the code sequence
if (msg_reverse())
tipc_link_xmit_skb()
at numerous locations in socket.c. The preparation of arguments
for these calls, as well as the sequence itself, makes the code
unecessarily complex.
In this commit, we introduce a new function, tipc_sk_respond(),
that performs this call combination. We also replace some, but not
yet all, of these explicit call sequences with calls to the new
function. Notably, we let the function tipc_sk_proto_rcv() use
the new function to directly send out PROBE_REPLY messages,
instead of deferring this to the calling tipc_sk_rcv() function,
as we do now.
Reviewed-by: Ying Xue <ying.xue@windriver.com>
Signed-off-by: Jon Maloy <jon.maloy@ericsson.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The shortest TIPC message header, for cluster local CONNECTED messages,
is 24 bytes long. With this format, the fields "dest_node" and
"orig_node" are optimized away, since they in reality are redundant
in this particular case.
However, the absence of these fields leads to code inconsistencies
that are difficult to handle in some cases, especially when we need
to reverse or reject messages at the socket layer.
In this commit, we concentrate the handling of the absent fields
to one place, by letting the function tipc_msg_reverse() reallocate
the buffer and expand the header to 32 bytes when necessary. This
means that the socket code now can assume that the two previously
absent fields are present in the header when a message needs to be
rejected. This opens up for some further simplifications of the
socket code.
Reviewed-by: Ying Xue <ying.xue@windriver.com>
Signed-off-by: Jon Maloy <jon.maloy@ericsson.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
We convert packet/message reception according to the same principle
we have been using for message sending and timeout handling:
We move the function tipc_rcv() to node.c, hence handling the initial
packet reception at the link aggregation level. The function grabs
the node lock, selects the receiving link, and accesses it via a new
call tipc_link_rcv(). This function appends buffers to the input
queue for delivery upwards, but it may also append outgoing packets
to the xmit queue, just as we do during regular message sending. The
latter will happen when buffers are forwarded from the link backlog,
or when retransmission is requested.
Upon return of this function, and after having released the node lock,
tipc_rcv() delivers/tranmsits the contents of those queues, but it may
also perform actions such as link activation or reset, as indicated by
the return flags from the link.
This reduces the number of cpu cycles spent inside the node spinlock,
and reduces contention on that lock.
Reviewed-by: Ying Xue <ying.xue@windriver.com>
Signed-off-by: Jon Maloy <jon.maloy@ericsson.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
David S. Miller