| Age | Commit message (Collapse) | Author |
|---|
|
* anders/diameter/sctp/OTP-10889:
Send unordered on all outbound diameter_sctp streams
Delay rotation of diameter_sctp outbound streams
Exercise unordered delivery in traffic suite
Use unordered delivery on a lone outbound stream in diameter_sctp
|
|
There's no reason for sending ordered since request handling is
concurrent: different processes handling incoming requests can't know in
which order they were received on the transport, and different processes
sending requests can't know the order in which they're sent.
|
|
For the same reason as unordered delivery is delayed in the grandparent
commit. Delivery is ordered only within a stream, so until a second
message is received from the peer, there's no guarantee that a second
outgoing request won't be received before the initial capablities
exchange message. Rotation begins upon reception of a second message
from the peer, messages being sent on stream 0 until then.
|
|
* anders/diameter/config_consistency/OTP-14555:
Let strict_mbit and incoming_maxlen be configured per transport
Let a service configure default transport options
Rename type evaluable -> eval
|
|
* anders/diameter/loopback_any/OTP-14544:
Use loopback/any config in examples suite
Handle loopback/any as local address in diameter_tcp/sctp
|
|
RFC 6733 say the below about head-of-line blocking and SCTP, the
suggestion of unordered sending being new compared to the RFC 3588.
Until now, all delivery in diameter_sctp has been ordered, and
roundrobin over available streams unless the user passes a stream
identifier in an outgoing diameter_packet record. This commit changes
this to use unordered delivery when there's only a single outbound
stream to choose from.
The special case at capabilities exchange is handled by only starting to
send unordered after the second message from the peer has been received.
(First message after capabilities exchange, as the RFC probably means.)
The special case at DPR isn't handled, since there's no knowing the
order in which a peer will answer: a node that sends DPR while it has
other requests outstanding can't expect that the latter will be answered
before DPR, even if delivery is ordered since incoming requests are
handled concurrently. If it wants a guarantee then it simply has to wait
for answers before sending DPR.
A user can force all delivery to be unordered by specifying
{sctp_default_send_params, #sctp_sndrcvinfo{flags = [unordered]}}
as a config option to diameter_sctp, but in this case there's no
handling of a request being sent directly after CEA since there's no
ordered flag to override the default.
RFC 6733:
2.1.1. SCTP Guidelines
Diameter messages SHOULD be mapped into SCTP streams in a way that
avoids head-of-the-line (HOL) blocking. Among different ways of
performing the mapping that fulfill this requirement it is
RECOMMENDED that a Diameter node send every Diameter message (request
or response) over stream zero with the unordered flag set. However,
Diameter nodes MAY select and implement other design alternatives for
avoiding HOL blocking such as using multiple streams with the
unordered flag cleared (as originally instructed in RFC 3588). On
the receiving side, a Diameter entity MUST be ready to receive
Diameter messages over any stream, and it is free to return responses
over a different stream. This way, both sides manage the available
streams in the sending direction, independently of the streams chosen
by the other side to send a particular Diameter message. These
messages can be out-of-order and belong to different Diameter
sessions.
Out-of-order delivery has special concerns during a connection
establishment and termination. When a connection is established, the
responder side sends a CEA message and moves to R-Open state as
specified in Section 5.6. If an application message is sent shortly
after the CEA and delivered out-of-order, the initiator side, still
in Wait-I-CEA state, will discard the application message and close
the connection. In order to avoid this race condition, the receiver
side SHOULD NOT use out-of-order delivery methods until the first
message has been received from the initiator, proving that it has
moved to I-Open state. To trigger such a message, the receiver side
could send a DWR immediately after sending a CEA. Upon reception of
the corresponding DWA, the receiver side should start using out-of-
order delivery methods to counter the HOL blocking.
Another race condition may occur when DPR and DPA messages are used.
Both DPR and DPA are small in size; thus, they may be delivered to
the peer faster than application messages when an out-of-order
delivery mechanism is used. Therefore, it is possible that a DPR/DPA
exchange completes while application messages are still in transit,
resulting in a loss of these messages. An implementation could
mitigate this race condition, for example, using timers, and wait for
a short period of time for pending application level messages to
arrive before proceeding to disconnect the transport connection.
Eventually, lost messages are handled by the retransmission mechanism
described in Section 5.5.4.
|
|
Export the old type as a synonym for backwards compatability. The name
evaluable is a bit too awkward.
|
|
The support is implied by documentation, but wasn't handled in code. Be
consistent in retrieving the address from the sock rather than the
configuration, and in accepting both ip and ifaddr for a local address.
|
|
* anders/diameter/performance/OTP-14521:
Work around unexpected common_test behaviour
Randomly skip groups in traffic suite
Randomly disable traffic counters in traffic suite
Add service_opt() traffic_counters
Fix type spec
Split AVPs at decode
Avoid unnecessary copying of binaries in diameter_tcp
Don't update diameter_tcp state unnecessarily
Don't update diameter_tcp state unnecessarily
Simplify extraction of incoming Diameter messages in diameter_tcp
Restructure/simplify message reception in diameter_peer_fsm
Sleep randomly at the start of (parallel) traffic testcases
Fix ct return value in traffic suite
Fix type spec
Optimize sub-binaries
Optimize sub-binaries
Count AVPs in #diameter_avp.index
Don't extract options unnecessarily at encode
Redo message decode as a single pass
|
|
Since messages are accumulated by appending binaries as of three
commits back, the accumulated binary is prone to being copied, as
discussed in the Efficiency Guide. Matching the Message Length header
as bytes are being accumulated is one cause of this, so work around it
by splitting the binary and extracting the length without a match.
This doesn't feel like something that should be necessary: that matching
a binary would cause an append to copy isn't obvious. The first attempt
at simplifying the accumulation was simply to append an incoming binary
to the current fragment, match against <<_, Len:24, _/binary>> to
extract the length, and then test if there are enough bytes for a
message. This lead to horrible performance (response times for 2 MB
messages approximately 100 times worse than previously), and it wasn't
at all obvious that the reason was the accumulated binary being copied
with each append as a result of the match. Using split_binary avoids the
match context that forces the copying.
|
|
Not with each setopts to re-activate the socket.
|
|
Only reset the fragment after all messages have been extracted.
|
|
Appending to a binary is efficient, so just append message fragments
Only match out the length once per message since doing so for every
packet from TCP causes the binary to be copied.
|
|
* anders/diameter/transport/ERL-332:
Remove irrelevant comment
Add missing setopts after deferred diameter_{tcp,sctp} actions
|
|
|
|
Since the actions can request that a previously passive socket be made
active.
Missed in commits 636a7199 and 373cd07c.
|
|
* anders/diameter/transport/ERL-332: (35 commits)
Capitulate on SCTP vs sparc-sun-solaris2.10
Remove obsolete traffic testcase
Fix dialyzer warnings
Remove client/server string decode from traffic suite
Add diameter_sctp option packet
Add diameter_sctp send/recv callbacks
Let diameter_tcp send/recv callbacks deal in diameter_packet
Randomly select traffic testcases
Exercise diameter_tcp message callbacks in traffic suite
Exercise diameter_{tcp,sctp} sender in traffic suite
Remove upgrade from diameter_traffic
Add diameter_tcp send/recv callbacks
Make diameter_{tcp,sctp} sender configurable
Remove upgrade from diameter_sctp; tweak diameter_tcp to match
Fix incomprehensible dialyzer warning
Simplify acks to transport processes
Strip throttling callbacks from diameter_tcp
Deal with (another) SCTP association id quirk on Solaris
Use binary:copy/2 when generating largish data in test suites
Deal with SCTP association id quirk on Solaris
...
|
|
diameter_sctp.erl:292: Record construction #transport{parent::pid(),mode::{'accept',atom() | pid() | port() | {atom(),atom()}},active::'false',recv::'true',os::0,packet::'true',message_cb::'undefined',send::'false'} violates the declared type of field message_cb::'false' | fun() | maybe_improper_list(fun() | maybe_improper_list(any(),[any()]) | {atom(),atom(),[any()]},[any()]) | {atom(),atom(),[any()]}
diameter_sctp.erl:302: Record construction #transport{mode::{'accept',atom() | pid() | port() | {atom(),atom()}},active::'false',recv::'true',os::0,packet::'true',message_cb::'undefined',send::'false'} violates the declared type of field message_cb::'false' | fun() | maybe_improper_list(fun() | maybe_improper_list(any(),[any()]) | {atom(),atom(),[any()]},[any()]) | {atom(),atom(),[any()]}
|
|
To determine the wrapping of messages passed to recv callbacks and into
diameter. The default passing of the input stream in transport_data is
probably of no practical use, but has been set since time immemorial.
|
|
Corresponding to diameter_tcp callbacks a few commits back. Exercise the
callbacks in the traffic suite.
|
|
To let a recv callback for an incoming request set transport_data and
have it returned in a send callback.
|
|
From the receiver process, that can return binaries to send/receive and
stop the transport process from reading on the socket.
This is still undocumented, and may change.
|
|
With sends still from the receiving process by default, since changing
the default behaviour may well have negative effects. A separate sender
probably implies a greater need for some form of load regulation for
one, since a blocking send would no longer imply that incoming messages
are no longer recevied. Dealing with this could result in the same
deadlock that the sending process intends to avoid, but the user should
be in control over how/when incoming traffic is regulated.
|
|
Added in commit 2afd1fe5. Only rename variables in diameter_tcp, no
functional change.
|
|
This:
diameter_tcp.erl:241: Record construction #transport{parent::'false',ssl::boolean() | maybe_improper_list(),frag::<<>>,tref::'false',flush::'false',pending::0,reset::{1 | 4,0 | 2},throttled::boolean(),q::{0,queue:queue(_)},monitor::'undefined' | pid()} violates the declared type of field parent::pid()
The problem isn't #transport.pid at all, it's #monitor.pid, and the only
relation is that the pid that's assigned to the latter is also (later)
assigned to the former. There is no record construction that assigns
false to #transport.parent.
Introduced in commit 33a535e4.
|
|
Commits starting at 472a080c added a throttle_cb option to diameter_tcp
to let a callback apply backpressure when it decides that additional
requests should not be read. It didn't provide a hook for knowing that
an answer was sent however, which is needed when sends no longer take
place in the receiver process, and is more complicated than it should
be. Strip it all away, in preparation for a simpler incarnation.
|
|
Shutdown events have been seen to get a different association id.
For example, first incoming message with association id = 0:
+ {trace_ts,<6421.268.0>,call,
{diameter_sctp,handle_info,
[{sctp,#Port<6421.2588>,
{10,67,16,179},
44159,
{[{sctp_sndrcvinfo,0,0,[],0,0,0,269950872,269950872,0}],
<<1,0,0,156,128,0,1,1,0,0,0,0,6,193,40,137,6,193,40,137,0,0,
1,8,64,0,0,30,67,45,49,51,52,50,49,55,52,52,49,46,101,114,
108,97,110,103,46,111,114,103,0,0,0,0,1,40,64,0,0,18,101,
114,108,97,110,103,46,111,114,103,0,0,0,0,1,1,64,0,0,14,0,
1,127,0,0,1,0,0,0,0,1,10,64,0,0,12,0,0,48,57,0,0,1,13,0,0,
0,20,79,84,80,47,100,105,97,109,101,116,101,114,0,0,1,22,
64,0,0,12,0,0,0,1,0,0,1,2,64,0,0,12,0,0,0,0,0,0,1,3,64,0,0,
12,0,0,0,3>>}},
{transport,<6421.252.0>,accept,#Port<6421.2588>,true,undefined,
{32,32},
0,undefined}]},
{1493,21505,577938}}
Later, a shutdown event with association id 1536:
+ {trace_ts,<6421.268.0>,call,
{diameter_sctp,handle_info,
[{sctp,#Port<6421.2588>,
{10,67,16,179},
44159,
{[],{sctp_shutdown_event,1536}}},
{transport,<6421.252.0>,accept,#Port<6421.2588>,0,undefined,
{32,32},
2,<6421.304.0>}]},
{1493,21505,746929}}
Both this and the grandparent commit are on this:
$ uname -a
SunOS beren 5.10 Generic_118833-33 sun4v sparc SUNW,Netra-T2000
|
|
In particular, that the association id received in messages on a
one-to-one socket after peeloff may be different from the id received on
the listen socket at comm_up.
This seems odd, since it's then not possible to send until the id is
discovered by reception of an SCTP message containing it, but it's
unclear if this is a bug or a feature, or if it's specific to certain
platforms. Treat it as a feature in this commit, and get the association
id as mentioned, an incoming CER being expected before anything is sent.
Commit da3e5d67 has more history.
|
|
Which dialyzer itself has never complained about.
|
|
Both diameter_tcp and diameter_sctp are susceptible to deadlock since a
peer that blocks send also prevents additional messages from being
received. Send from a process that's paired with the transport process
to avoid this. Use the existing monitor process in the TCP case, add one
in the SCTP case.
This has been the reason for many sporadic testcase failures, mostly in
diameter_traffic_SUITE.
|
|
Should be ssl:close/1.
|
|
Since smooth upgrade won't be supported in this branch.
|
|
|
|
|
|
Commit 5ca5fb71 ensured that they were closed immediately at transport
removal, but in so doing broke their closing at stop service completely,
by removing the timer that caused sockets to be closed even belatedly.
Monitor on the service process to make it happen.
This could still be improved, since stop_service listening ports aren't
closed until after the service process has died. They could be closed
earlier in the case of stop_service.
|
|
The transport interface documented in diameter_transport(3) is used to
start/stop accepting/connecting transport processes: they're started
with a function call, and told to die with their parent process. In the
accepting case, both diameter_tcp and diameter_sctp start a listening
process when the first accepting transport is started. However, there's
no way for a listening process to find out that that it should stop
listening when transport configuration is removed.
Both diameter_tcp and diameter_sctp have used a timer to terminate the
listening process after all existing accepting processes have died as a
consequence of transport removal. The problem with this is that nothing
stops a new client from connecting before this, and also that no new
transport can succeed in opening the same listening port (eg.
reconfiguration) until the old listener dies.
This commit solves the problem by adding diameter_reg:subscribe/2, to
allow callers to subscribe to messages about added/removed associations.
A call to diameter:add_transport/2 results in a new child process that
registers a term that a listening process subscribes to. Transport
removal results in the death of the child, and the resulting
notification to the listener causes the latter to close its socket and
terminate.
This is still an internal interface, but the subscription mechanism
should probably be made external (eg. a diameter:subscribe/1 that can
be used to subscribe to specified messages), so that transport modules
other than diameter's own can make use of it. There is no support for
soft upgrade.
|
|
* anders/diameter/overload/OTP-13330:
Suppress dialyzer warning
Remove dead case clause
Let throttling callback send a throttle message
Acknowledge answers to notification pids when throttling
Throttle properly with TLS
Don't ask throttling callback to receive more unless needed
Let a throttling callback answer a received message
Let a throttling callback discard a received message
Let throttling callback return a notification pid
Make throttling callbacks on message reception
Add diameter_tcp option throttle_cb
|
|
This one:
diameter_tcp.erl:928: (call)
The call diameter_tcp:throttle({'timeout',_},#transport{socket::port() | {'sslsocket',_,_},parent::pid(),module::atom(),frag::binary() | {non_neg_integer(),non_neg_integer(),binary(),[binary()]},ssl::boolean() | [any()],timeout::'infinity' | non_neg_integer(),tref::'false' | reference(),flush::boolean(),throttle_cb::'false' | fun() | maybe_improper_list(fun() | maybe_improper_list(any(),[any()]) | {atom(),atom(),[any()]},[any()]) | {atom(),atom(),[any()]},throttled::'true' | binary()})
will never return since it differs in the 1st argument from the
success typing arguments:
('discard' | 'ok' | binary() | pid() | {'discard' | 'ok' | binary() | pid(),'false' | fun() | [fun() | [any()] | {atom(),atom(),[any()]}] | {atom(),atom(),[any()]}},#transport{socket::port() | {'sslsocket',_,_},parent::pid(),module::atom(),frag::binary() | {non_neg_integer(),non_neg_integer(),binary(),[binary()]},ssl::boolean() | [any()],timeout::'infinity' | non_neg_integer(),tref::'false' | reference(),flush::boolean(),throttle_cb::'false' | fun() | [fun() | [any()] | {atom(),atom(),[any()]}] | {atom(),atom(),[any()]},throttled::binary()})
It's true that the clause doesn't return, because of the throw, and
that's the intention.
|
|
Orphaned in commit 9298872b.
|
|
That is, don't assume that it's only diameter_tcp doing so: allow it to
be received when not throttling. This lets a callback module trigger a
new throttling callback itself, but it's not clear if this will be
useful in practice.
|
|
|
|
In particular, let a callback decide when to receive the initial
message.
|
|
TCP packets can contain more than one message, so only ask to receive
another message if it hasn't already been received.
|
|
As discussed in the parent commit. This is easier said than done in
practice, but there's no harm in allowing it.
|
|
This can be used as a simple form of overload protection, discarding the
message before it's passed into diameter to become one more request
process in a flood. Replying with 3004 would be more appropriate when
the request has been directed at a specific server (the RFC's
requirement) however, and possibly it should be possible for a callback
to do this as well.
|
|
In addition to returning ok or {timeout, Tmo}, let a throttling callback
for message reception return a pid(), which is then notified if the
message in question is either discarded or results in a request process.
Notification is by way of messages of the form
{diameter, discard | {request, pid()}}
where the pid is that of a request process resulting from the received
message. This allows the notification process to keep track of the
maximum number of request processes a peer connection can have given
rise to.
|
|
The callback is now applied to the atom 'false' when asking if another
message should be received on the socket, and to a received binary
message after reception. Throttling on received messages makes it
possible to distinguish between requests and answers.
There is no callback on outgoing messages since these don't have to go
through the transport process, even if they currently do.
|
|
To let a callback module decide whether or to receive another message
from the peer, so that backpressure can be applied when it's
inappropriate. This is to let a callback protect against reading more
than can be processed, which is otherwise possible since diameter_tcp
otherwise always asks for more.
A callback is made after each message, and can answer to continue
reading or to ask again after a timeout. It's each message instead of
each packet partly for simplicity, but also since this should be
sufficiently fine-grained. Per packet would require some interaction
with the fragment timer that flushes partial messages that haven't been
completely received.
|
|
Record field types have been modified due to commit 8ce35b2:
"Take out automatic insertion of 'undefined' from typed record fields".
|
|
The diffs are all about adapting to the OTP 18 time interface. The code
was previously backwards compatible, falling back on the erlang:now/0 if
erlang:monotonic_time/0 is unavailable, but this was seen to be a bad
thing in commit 9c0f2f2c. Use of erlang:now/0 is now removed.
|
