After initial handshake a NEO node checks the identification of its peer
by sending the 'RequestIdentification' packet. In case the peer is a
secondary master it responds with 'NotPrimaryMaster'. Before this patch
'Dial' ignored the 'NotPrimaryMaster' packet and simply returned a
general error. Now - after this patch - 'Dial' returns an instance of
'proto.NotPrimaryMaster' (which implements 'Error').
This helps a caller to correctly handle the secondary-master-case, which
otherwise is impossible to differentiate from any other error
possibility.

When a client receives 'NotPrimaryMaster' from a secondary master, the
situation is similar to the situation when we receive an error: the
other node tells us, don't connect with me, connect with someone else.
Finally the peer even closes the connection.

Due to this similarity in structure (& because it helps us later to
teach NEO/go to correctly handle 'NotPrimaryMaster' with minimal
changes), we implement 'Error' for 'proto.NotPrimaryMaster'. Now
'NotPrimaryMaster' can be treated like an error.

In a NEO URI more than one master node can be specified, because a NEO
cluster may have more masters than one. But before this patch
'openClientByURL' always assumed that the given URL only specifies one
master. Now the host is split into potentially > 1 master nodes. It
therefore works now in the same way as the Python implementation [1].

[1] https://lab.nexedi.com/nexedi/neoppod/blob/342168cd/neo/client/zodburi.py#L64

Before this patch Client.URL didn't contain more than one master node.
This can be problematic in case we have a cluster with > 1 master nodes
and the printed master is a secondary master (which may be down). In
this case the user who trusts the "URL" attribute to connect to the
cluster won't succeed, because the primary master can't be reached.

Some NEO clusters have more than one master to gain a higher availability.
Before this patch NEO/go Node type only handled one master address. This
commit adjusts the node type and related bits so that it can support more than
one master node.

/reviewed-by @kirr
/reviewed-on !5

* t-fix-flaky-testload:
  fixup! neonet/newlink: Fix lost conn in encoding detector
  go/neo/neonet: Fix client handshake not to accept server encoding if it is different from what client indicated
  go/neo/neonet: Demonstrate problem in handshake with NEO/py
  go/neo/neonet: Dedicate an error type to indicate "protocol version mismatch" as handshake failure cause
  fixup! client_test: Keep NEO srv logs if test fails
  fixup! client_test/NEOSrv += LogContent for better debug
  neonet/newlink: Fix lost conn in encoding detector
  client_test: Keep NEO srv logs if test fails
  client_test += print NEO server log if >=1 test(s) failed
  client_test/NEOSrv += LogContent for better debug

* master:
  go/neo/neonet: Fix client handshake not to accept server encoding if it is different from what client indicated
  go/neo/neonet: Demonstrate problem in handshake with NEO/py
  go/neo/neonet: Dedicate an error type to indicate "protocol version mismatch" as handshake failure cause

go/neo/neonet: Fix client handshake not to accept server encoding if it is different from what client indicated

If the peers encoding is different than our encoding two different
scenarios can happen, because the handshake order is undefined (e.g.
we don't know if our handshake is received before the peer sends its
handshake):

1. Our handshake is received before peer sends its handshake, NEO/py
closes connection if it sees unexpected magic, version, etc.

2. The client already sends a handshake before it proceeds our handshake.
In this case it initally sends us it version, we can extract its encoding,
and only later, once it proceeded our handshake with the bad encoding,
it closes the connection.

Before this patch case (2) wasn't handled correctly by the automatic
encoding detection of 'DialLink'. 'DialLink' simply accepted the
different-than-expected encoding, but once the peer proceeded the nodes
handshake the peer closed the connection and the initially established
and returned link was immediately closed again. Due to this it was good
luck whether connecting with a peer different with an encoding different
from the expected one worked or didn't work (it depended on which handshake
was faster). Now 'DialLink' should reliably find the correct encoding
and return a stable link.

--------

kirr: this is based on the following original patch by Levin: f6b59772

I updated documentation throughout correspondingly and also added
corresponding handshake-specific test in the previous patch.

See kirr/neo!5 and b2da69e2
(go/neo/neonet: Demonstrate problem in handshake with NEO/py) for more
in-depth description of the problem.

Levin Zimmerman discovered that sometimes NEO/py accepts our handshake
hello with encoding 'M', then replies its owh handshake ehlo with
encoding 'N' and then further terminates the connection. In other words
it looks like that the handshake went successful, but it actually did
not and NEO/py terminates the link after some time.

This manifests itself e.g. as infrequent TestLoad failures on t branch
with the following output:

=== RUN   TestLoad/py/!ssl
I: runneo.py: /tmp/neo776618506/1 !ssl: started master(s): 127.0.0.1:21151
=== RUN   TestLoad/py/!ssl/enc=N(dialTryOrder=N,M)
    client_test.go:598: skip: does not excercise client redial
=== RUN   TestLoad/py/!ssl/enc=N(dialTryOrder=M,N)
    xtesting.go:330: load 0285cbac258bf266:0000000000000000: returned err unexpected:
        have: neo://127.0.0.1:21151/1: load 0285cbac258bf266:0000000000000000: dial S1: dial 127.0.0.1:40345 (STORAGE): 127.0.0.1:56678 - 127.0.0.1:40345: request identification: 127.0.0.1:56678 - 127.0.0.1:40345 .1: recv: EOF
        want: nil
    xtesting.go:330: load 0285cbac258bf266:0000000000000000: returned tid unexpected: 0000000000000000  ; want: 0285cbac258bf266
    xtesting.go:330: load 0285cbac258bf266:0000000000000000: returned buf = nil
    xtesting.go:339: load 0285cbac258bf265:0000000000000000: returned err unexpected:
        have: neo://127.0.0.1:21151/1: load 0285cbac258bf265:0000000000000000: dial S1: dial 127.0.0.1:40345 (STORAGE): 127.0.0.1:56688 - 127.0.0.1:40345: request identification: 127.0.0.1:56688 - 127.0.0.1:40345 .1: recv: EOF
        want: neo://127.0.0.1:21151/1: load 0285cbac258bf265:0000000000000000: 0000000000000000: object was not yet created
    ...
    client_test.go:588: NEO log tail:

        log file 'storage_0.log' tail:

        2023-07-17 17:21:57.1519 DEBUG     S1         connection completed for <ServerConnection(nid=None, address=127.0.0.1:51230, handler=IdentificationHandler, fd=20, server) at 7f3583fd4f50> (from 127.0.0.1:40345)
        2023-07-17 17:21:57.1537 WARNING   S1         Protocol version mismatch with <ServerConnection(nid=None, address=127.0.0.1:51230, handler=IdentificationHandler, fd=20, server) at 7f3583fd4f50>
        2023-07-17 17:21:57.1548 DEBUG     S1         connection closed for <ServerConnection(nid=None, address=127.0.0.1:51230, handler=IdentificationHandler, closed, server) at 7f3583fd4f50>
        2023-07-17 17:21:57.1551 WARNING   S1         A connection was lost during identification
        2023-07-17 17:22:00.1582 DEBUG     S1         accepted a connection from 127.0.0.1:51236
        2023-07-17 17:22:00.1585 DEBUG     S1         connection completed for <ServerConnection(nid=None, address=127.0.0.1:51236, handler=IdentificationHandler, fd=20, server) at 7f3583fd4e90> (from 127.0.0.1:40345)
        2023-07-17 17:22:00.1604 WARNING   S1         Protocol version mismatch with <ServerConnection(nid=None, address=127.0.0.1:51236, handler=IdentificationHandler, fd=20, server) at 7f3583fd4e90>
        2023-07-17 17:22:00.1622 DEBUG     S1         connection closed for <ServerConnection(nid=None, address=127.0.0.1:51236, handler=IdentificationHandler, closed, server) at 7f3583fd4e90>
        2023-07-17 17:22:00.1625 WARNING   S1         A connection was lost during identification
        2023-07-17 17:22:03.1663 DEBUG     S1         accepted a connection from 127.0.0.1:51238
        2023-07-17 17:22:03.1666 DEBUG     S1         connection completed for <ServerConnection(nid=None, address=127.0.0.1:51238, handler=IdentificationHandler, fd=20, server) at 7f3583fd4d10> (from 127.0.0.1:40345)
        2023-07-17 17:22:03.1674 WARNING   S1         Protocol version mismatch with <ServerConnection(nid=None, address=127.0.0.1:51238, handler=IdentificationHandler, fd=20, server) at 7f3583fd4d10>
        2023-07-17 17:22:03.1688 DEBUG     S1         connection closed for <ServerConnection(nid=None, address=127.0.0.1:51238, handler=IdentificationHandler, closed, server) at 7f3583fd4d10>
        2023-07-17 17:22:03.1691 WARNING   S1         A connection was lost during identification
        2023-07-17 17:22:06.1714 DEBUG     S1         accepted a connection from 127.0.0.1:57072
        2023-07-17 17:22:06.1719 DEBUG     S1         connection completed for <ServerConnection(nid=None, address=127.0.0.1:57072, handler=IdentificationHandler, fd=20, server) at 7f3583fd4b50> (from 127.0.0.1:40345)
        2023-07-17 17:22:06.1727 WARNING   S1         Protocol version mismatch with <ServerConnection(nid=None, address=127.0.0.1:57072, handler=IdentificationHandler, fd=20, server) at 7f3583fd4b50>
        2023-07-17 17:22:06.1738 DEBUG     S1         connection closed for <ServerConnection(nid=None, address=127.0.0.1:57072, handler=IdentificationHandler, closed, server) at 7f3583fd4b50>
        2023-07-17 17:22:06.1738 WARNING   S1         A connection was lost during identification

        log file 'master_0.log' tail:

        2023-07-17 17:21:21.0799 PACKET    M1         #0x0012 NotifyNodeInformation          > A1 (127.0.0.1:37906)
        2023-07-17 17:21:21.0799 PACKET    M1          ! C0 | CLIENT |  | RUNNING | 2023-07-17 14:21:21.079838
        2023-07-17 17:21:21.0800 PACKET    M1         #0x0102 NotifyNodeInformation          > S1 (127.0.0.1:37918)
        2023-07-17 17:21:21.0800 PACKET    M1          ! C0 | CLIENT |  | RUNNING | 2023-07-17 14:21:21.079838
        2023-07-17 17:21:21.0801 DEBUG     M1         Handler changed on <ServerConnection(nid=None, address=127.0.0.1:37966, handler=ClientServiceHandler, fd=18, server) at 7f3584245910>
        2023-07-17 17:21:21.0802 PACKET    M1         #0x0001 AnswerRequestIdentification    > C0 (127.0.0.1:37966)
        2023-07-17 17:21:21.0804 PACKET    M1         #0x0000 NotifyNodeInformation          > C0 (127.0.0.1:37966)
        2023-07-17 17:21:21.0804 PACKET    M1          ! C0 | CLIENT  |                 | RUNNING | 2023-07-17 14:21:21.079838
        2023-07-17 17:21:21.0804 PACKET    M1          ! M1 | MASTER  | 127.0.0.1:21151 | RUNNING | None
        2023-07-17 17:21:21.0804 PACKET    M1          ! S1 | STORAGE | 127.0.0.1:40345 | RUNNING | 2023-07-17 14:21:18.737469
        2023-07-17 17:21:21.0805 PACKET    M1         #0x0002 NotifyPartitionTable           > C0 (127.0.0.1:37966)
        2023-07-17 17:21:21.0810 PACKET    M1         #0x0003 LastTransaction                < C0 (127.0.0.1:37966)
        2023-07-17 17:21:21.0811 PACKET    M1         #0x0003 AnswerLastTransaction          > C0 (127.0.0.1:37966)
        2023-07-17 17:22:06.2053 DEBUG     M1         <SocketConnectorIPv4 at 0x7f3584252d10 fileno 18 ('127.0.0.1', 21151), opened from ('127.0.0.1', 37966)> closed in recv
        2023-07-17 17:22:06.2056 DEBUG     M1         connection closed for <ServerConnection(nid=C0, address=127.0.0.1:37966, handler=ClientServiceHandler, closed, server) at 7f3584245910>
        2023-07-17 17:22:06.2058 PACKET    M1         #0x0014 NotifyNodeInformation          > A1 (127.0.0.1:37906)
        2023-07-17 17:22:06.2058 PACKET    M1          ! C0 | CLIENT |  | UNKNOWN | 2023-07-17 14:21:21.079838
        2023-07-17 17:22:06.2059 PACKET    M1         #0x0104 NotifyNodeInformation          > S1 (127.0.0.1:37918)
        2023-07-17 17:22:06.2059 PACKET    M1          ! C0 | CLIENT |  | UNKNOWN | 2023-07-17 14:21:21.079838

The problem is due to that my analysis from e407f725 (go/neo/neonet:
Rework handshake to differentiate client and server parts) turned out to
be incorrect. Quoting that patch:

    -> Rework handshake so that client always sends its hello first, and
    only then the server side replies. This matches actual NEO/py behaviour:

    https://lab.nexedi.com/nexedi/neoppod/blob/v1.12-67-g261dd4b4/neo/lib/connector.py#L293-294

    even though the "NEO protocol" states that

    	Handshake transmissions are not ordered with respect to each other and can go in parallel.

    	( https://neo.nexedi.com/P-NEO-Protocol.Specification.2019?portal_skin=CI_slideshow#/9/2 )

    If I recall correctly that sentence was authored by me in 2018 based on
    previous understanding of should-be full symmetry in-between client and
    server.

so here "This matches actual NEO/py behaviour" was wrong: even though
https://lab.nexedi.com/nexedi/neoppod/blob/v1.12-67-g261dd4b4/neo/lib/connector.py#L293-294
indeed says that

    #       The NEO protocol is such that a client connection is always the
    #       first to send a packet, as soon as the connection is established,

in reality it is not the case as SocketConnector always queues handshake
hello upon its creation before receiving anything from remote side:
https://lab.nexedi.com/nexedi/neoppod/blob/v1.12-93-gfd87e153/neo/lib/connector.py#L77-78 .
In practice this leads to that in non-SSL case NEO/py server might be
fast enough to send its prepared hello before receiving hello from us.
Levin also explains at kirr/neo!5 (comment 187429):

    I think what happens is this:

    the NEO protocol doesn't specify in which order handshakes happen after
    initial dial. If the peer sends a handshake before receiving our
    handshake and if this peers handshake is received by us, 'DialLink'
    assumes everything is fine (no err is returned), it breaks the loop and
    returns the link. But then, very little time later, when the peer
    finally receives our handshake, this looks strange for the peer and it
    closes the connection.

    So in my understanding this should be fixed by explicitly comparing the
    encodings between our expected one and what the peer provided us. If
    encodings don't match we should retry with a new encoding in order to
    prevent the peer from closing the connection.

    For me this also explains why sometimes the tests passed and sometimes
    didn't: it depended on which node was faster ('race condition').

-> In this patch we add correspondig handshake test that demonstrates
this problem. It currently fails as

--- FAIL: TestHandshake (0.01s)
    --- FAIL: TestHandshake/enc=N (0.00s)
        newlink_test.go:154: handshake encoding mismatch: client: unexpected error:
            have: <nil>
                 "<nil>"
            want: &neonet._HandshakeError{LocalRole:1, LocalAddr:net.pipeAddr{}, RemoteAddr:net.pipeAddr{}, Err:(*neonet._EncodingMismatchError)(0xc0000a4190)}
                 "pipe - pipe: handshake (client): protocol encoding mismatch: peer = 'M'  ; our side = 'N'"
    --- FAIL: TestHandshake/enc=M (0.00s)
        newlink_test.go:154: handshake encoding mismatch: client: unexpected error:
            have: <nil>
                 "<nil>"
            want: &neonet._HandshakeError{LocalRole:1, LocalAddr:net.pipeAddr{}, RemoteAddr:net.pipeAddr{}, Err:(*neonet._EncodingMismatchError)(0xc0001a22cc)}
                 "pipe - pipe: handshake (client): protocol encoding mismatch: peer = 'N'  ; our side = 'M'"

We will fix it in the next patch.

/reported-by @levin.zimmermann
/reported-on kirr/neo!5

go/neo/neonet: Dedicate an error type to indicate "protocol version mismatch" as handshake failure cause

We will soon need to detect if a handshake failure was due to mismatch
of protocol encodings and that would require introduction of dedicated
error type for that cause. As a preparatory step first refactor "version
mismatch cause" to follow the same style for symmetry.

LogContent -> LogTail
py/neolog  -> `python -m neo.scripts.neolog`
FIXME for glog
FIXME for admin_0.log and other log files

If the peers encoding is different than our encoding two different
scenarios can happen, because the handshake order is undefined (e.g.
we don't know if our handshake is received before the peer sends its
handshake):

1. Our handshake is received before peer sends its handshake, NEO/py
closes connection if it sees unexpected magic, version, etc.

2. The client already sends a handshake before it proceeds our handshake.
In this case it initally sends us it version, we can extract its encoding,
and only later, once it proceeded our handshake with the bad encoding,
it closes the connection.

Before this patch case (2) wasn't handled correctly by the automatic
encoding detection of 'DialLink'. 'DialLink' simply accepted the
different-than-expected encoding, but once the peer proceeded the nodes
handshake the peer closed the connection and the initially established
and returned link was immediately closed again. Due to this it was good
luck whether connecting with a peer different with an encoding different
from the expected one worked or didn't work (it depended on which handshake
was faster). Now 'DialLink' should reliably find the correct encoding
and return a stable link.

If a test fails we may want to invest further why this test fails and
reading the NEO server log files can be an efficient way for these
investigations. Therefore we should keep log files if a test fails.

If tests fail we may want to check NEO logs. The new function LogContent
of the NEOSrv interface provides an API to fetch the relevant parts of
the log file of a NEO server. This API can be used by test code in order
to automatically print relevant server log information if a test fails.

For example option `compress` controls kind of compression that _client_
performs when saving data to server. Similarly cache-size, logfile and
read-only adjust on-client behaviour, not server.

From nexedi/neoppod!18 (comment 124725) :

    In general the most correct thing to do is:

    - use host part for where to connect (host:port, list of host ports, UNIX socket, etc)
    - use path part to identify a database or other on-server resource
    - use query part for parameters that are passed to remote server (e.g. `storage` in case of ZEO)
    - use fragment part for local parameters that are not passed to remote server (e.g. local `logfile`)
    - use credentials part for things required to authenticate/encrypt.

    To normalize an URL wcfs client would drop credentials and fragment, but keep host, path and query.

    Fragments are documented not to be sent to remote side and to be evaluated by local side only.

-> Move options that control client behaviour to fragment.

Because list of masters and cluster name must be already present in
netloc and path. Previously e.g.

	neo://α,β,γ/db?master_nodes=a,b,c"

would mean to use master nodes {a,b,c} not {α,β,γ}. Now it is treated as
invalid URI to remove ambiguity. Same for cluster name.

* master:
  go/neo: Expand user prefix in TLS key/cert paths
  go/neo/interal += xpath, xfilepath

This patch fixes a discrepancy between NEO/py and NEO/go: NEO/py expands
the '~' and the '~username' prefix in the file path of the TLS
certificate/key files [1]. This syntax is used in NEO/py SlapOS SR [2].
We need to fix this discrepancy in NEO/go in order to use TLS encryption
with NEO + WCFS.

[1] https://lab.nexedi.com/nexedi/neoppod/blob/7c539f0f/neo/lib/config.py#L149
and https://lab.nexedi.com/nexedi/neoppod/blob/fa63d856/neo/lib/app.py#L25-31

[2] https://lab.nexedi.com/nexedi/slapos/blob/397726e1/stack/erp5/instance-zodb-base.cfg.in#L18-20
and https://lab.nexedi.com/nexedi/slapos/blob/a8150a1a/software/neoppod/instance-neo-input-schema.json#L62

/reviewed-by @kirr
/reviewed-on !1

The xfilepath package supports resolving filepaths with a user prefix
to absolute paths: it converts '~' and '~username' to $HOME of user
(as it's done by for  instance bash). No builtin golang module supports this
functionality [1]. We need this functionality in order to imitate the
behaviour of NEO/py in NEO/go [2].

---

[1] https://stackoverflow.com/questions/47261719/how-can-i-resolve-a-relative-path-to-absolute-path-in-golang)
[2] nexedi/slapos!1307 (comment 17574)

/reviewed-by @kirr
/reviewed-on kirr/neo!1

* master:
  go/neo/t/neotest: Use python -c 'print ...' in a way that works on both py2 and py3
  go/neo/t/tzodb.py: Fix zhash for Python3

Without parenthesis it was failing on py3:

	(neo) (py3.venv) (g.env) kirr@deca:~/src/neo/src/lab.nexedi.com/kirr/neo/go/neo/t$ ./neotest info-local
	date:   Wed, 18 May 2022 11:05:50 +0300
	xnode:  kirr@deca.navytux.spb.ru (2401:5180:0:af::1 192.168.0.3 (+ 1·ipv4))
	uname:  Linux deca 5.10.0-13-amd64 #1 SMP Debian 5.10.106-1 (2022-03-17) x86_64 GNU/Linux
	cpu:    Intel(R) Core(TM) i7-7600U CPU @ 2.80GHz
	  File "<string>", line 1
	    print '%.2fGHz' % (400000 / 1E6)
	          ^
	SyntaxError: invalid syntax

On py3 dict.keys() returns iterator instead of list:

    $ ./tzodb.py zhash
    Traceback (most recent call last):
      File "/home/kirr/src/neo/src/lab.nexedi.com/kirr/neo/go/neo/t/./tzodb.py", line 141, in <module>
        main()
      File "/home/kirr/src/neo/src/lab.nexedi.com/kirr/neo/go/neo/t/./tzodb.py", line 138, in main
        zhash()
      File "/home/kirr/src/neo/src/lab.nexedi.com/kirr/neo/go/neo/t/./tzodb.py", line 59, in zhash
        optv, argv = getopt(sys.argv[2:], "h", ["help", "check=", "bench="] + hashRegistry.keys())
    TypeError: can only concatenate list (not "dict_keys") to list

* master:
  go/internal/xurl: New package (draft)

With xurl.ParseQuery being simpler analog of url.ParseQuery.
Simpler: It returns regular map instead of url.Values by not allowing duplicates.

Instead of requiring users to explicitly mark their tests as finished in
particular subject-to-deadlock places, tracetest was fixed to
unconditionally detect "extra sends" deadlocks automatically:

go123@3b19f68c

go get: upgraded github.com/fsnotify/fsnotify v1.4.10-0.20200417215612-7f4cf4dd2b52 => v1.5.1
go get: upgraded github.com/golang/glog v0.0.0-20160126235308-23def4e6c14b => v1.0.0
go get: upgraded github.com/gwenn/gosqlite v0.0.0-20201008200117-82e079acf5b6 => v0.0.0-20211101095637-b18efb2e44c8
go get: upgraded github.com/gwenn/yacr v0.0.0-20200112083327-bbe82c1f4d60 => v0.0.0-20211101095056-492fb0c571bc
go get: upgraded github.com/kisielk/og-rek v1.1.0 => v1.2.0
go get: upgraded github.com/soheilhy/cmux v0.1.5-0.20210205191134-5ec6847320e5 => v0.1.5
go get: upgraded github.com/tinylib/msgp v1.1.5 => v1.1.6
go get: upgraded golang.org/x/net v0.0.0-20210226172049-e18ecbb05110 => v0.0.0-20211111160137-58aab5ef257a
go get: upgraded golang.org/x/sys v0.0.0-20210301091718-77cc2087c03b => v0.0.0-20211111213525-f221eed1c01e
go get: upgraded golang.org/x/text v0.3.5 => v0.3.7
go get: upgraded lab.nexedi.com/kirr/go123 v0.0.0-20210128150852-c20e95f0f789 => v0.0.0-20210906140734-c9eb28d9e408

Highlights:

- og-rek brings support for pickle protocol 5
- go123  brings in support for Go1.17

* master:
  go/zodb/btree: Change V<op> family to also provide visited node key coverage on visit callback
  go/zodb/btree: Add KeyRange type
  go/zodb/btree: Introduce constants for min/max key value
  go/zodb/btree: tests: Don't forget to close storage
  go/zodb/btree: Cosmetics

WCFS needs to know key coverage for every visited node.

WARNING: this is API change.

KeyRange represents [lo,hi) key range. It simplifies working with ranges
of keys. We will use it in the next commit.

KeyRange originated in WCFS and was copied from there:

https://lab.nexedi.com/kirr/wendelin.core/blob/57be0126/wcfs/internal/xbtree/blib/keyrange.go

We were already using math.Min<Key> in one place, but the number of such
places is going to increase. -> Keep min/max definition in only one place.

NOTE: db was already being closed in the test's code.

* t+btree:
  .
  .
  .
  .
  .
  .
  .
  X Start reworking BTree to provide keycov on visit callback