Levin found that when all handshake attempts fail DialLink returns
both link=nil and err=nil which breaks what callers expect and lead to
segmentation fault when accessing that nil link.

-> Add test to demonstrate the problem.

With xfail removed that test currently fails as

    --- FAIL: TestDialLink_AllHandshakeErr (0.00s)
    panic: lab.nexedi.com/kirr/neo/go/neo/neonet.TestDialLink_AllHandshakeErr.gox.func4.1: lab.nexedi.com/kirr/neo/go/neo/neonet.TestDialLink_AllHandshakeErr.func2: DialLink to handshake-rejecting server:
    have: link=<nil> err=<nil>
    want: link=<nil> err=client:1 - server:2: handshake (client): unexpected EOF [recovered]

We will fix the problem in the next patch.

/reported-by @levin.zimmermann
/reported-at kirr/neo!10

I realized a minor mistake I did in writing the tests that needs to
be fixed in order to have reliable test results.

/reviewed-by @kirr
/reviewed-on kirr/neo!9

During our work on 'wendelin.core' URI normalization, Kirill Smelkov
noted a inconsistency between NEO/py and NEO/go URI parser [1]: NEO/py
drops empty query options, while NEO/go preserves them. Let's fix this
inconsistency by adjusting NEO/go to NEO/py behaviour.

[1] nexedi/wendelin.core!28 (comment 212447)

/reviewed-by @kirr
/reviewed-on kirr/neo!9

Hello Kirill,

in nexedi/neoppod!18 and nexedi/neoppod!21 we could find a common solution for a zurl format that previously diverged between NEOgo and NEOpy. The purpose of this MR is to sync again NEOgo and NEOpy zurl format. After merging this, we can continue to sync NEO zurl format in 'wendelin.core' & 'slapos'. Then we finally have unified approach again, which simplifies understanding and reduces unnecessary mental overhead.

As this is strongly related to nexedi/neoppod!21 I thought it'd be a good idea to generally reduce difference and to replace WIP commits with merged NEOpy upstream commits.

Best, Levin

/reviewed-by @kirr
/reviewed-on !7

* lev/sync-zurl:
  client: Don't allow oPtion_nAme in zurl
  app: Remember SSL credentials so that it is possible to retrieve them
  client: Allow to force TLS via neos:// scheme
  client: Don't allow master_nodes and name to be present in options
  Revert "."
  Revert "Y client: Fix URI scheme to move credentials out of query"
  Revert "X Adjust NEO/go to neo:// URL change + py fixups"
  Revert "fixup! Y client: Fix URI scheme to move credentials out of query"
  Revert "Y client: Don't allow master_nodes and name to be present in options"
  go/client/zurl: Sync format to py upstream

Julien notes this is very likely unneeded:
nexedi/neoppod!21 (diffs, comment 195929)

We had it like this since 01a01c8c (client: Add support for zodburi),
but I rechecked zodburi codebase now and it does not do any similar
lowering anywhere.

So drop support for case normalization in zurl options.

/cc @levin.zimmermann
/reviewed-by @jm
/reviewed-on nexedi/neoppod!21

(cherry-picked from commit 798c9f25)

Unfortunately after creating SSL context it is not possible, or at least
I could not find how, to retrieve original credentials with which the
context was created. However wendelin.core needs to be able to take a
client storage, reconstruct zurl to refer to that particular storage,
and pass that zurl to wcfs, so that wcfs, in turn, could access the same
ZODB database.

Given a NEO client instance, it is already possible to retrieve
master_nodes, cluster name, and detect whether SSL is being in use.
However without being able to retrieve original SSL credentials,
reconstructed zurl will not be full and wcfs won't be able to use
exactly the same secrets as python part does.

-> Help wendelin.core by remembering which ca/cert/key were used to
build SSL context.

This information is used by zstor_2zurl in wendelin.core here:

https://lab.nexedi.com/nexedi/wendelin.core/blob/885b3556/lib/zodb.py#L390-418

/cc @levin.zimmermann
/reviewed-by @jm
/reviewed-on nexedi/neoppod!21

Similarly to how it is done with e.g. http:// and https:// - if neos://
is given TLS usage is forced and ca/cert/key must be there either in the
URI itself, or in $NEO_CA, $NEO_CERT and $NEO_KEY environment variables
mimicking the way how e.g. for https:// TLS credentials are taken from
host environment, not from the uri.

The latter might be usability convenience, but is also useful for WCFS
which needs to be able to remove secrets from uri on zurl normalization.

Please see discussion at nexedi/neoppod!18 (comment 184439)
for details.

/cc @levin.zimmermann
/reviewed-by @jm
/reviewed-on nexedi/neoppod!21

(cherry-picked from commit bc3e38ea)

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

	neo://db@α,β,γ?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.

/cc @levin.zimmermann
/reviewed-by @jm
/reviewed-on nexedi/neoppod!21

(cherry-picked from commit 22ccebd6)

This reverts commit a2f192cb. This has
been merged upstream with nexedi/neoppod@17af7f27.
We should rather cherry-pick upstream commit.

This reverts commit b9a42957. In
nexedi/neoppod!18 and
nexedi/neoppod!21 a common
solution for a zurl format was found. This common format keeps
credentials in the query, therefore we should revert patch
b9a42957.

This reverts the py part of kirr/neo@8c974485.
go parts of this patch are handled in 70c0a984.

This reverts commit cf685fb5. This used
to be a divergence between NEO/py and NEO/go, however in nexedi/neoppod!18 and
nexedi/neoppod!21 a common
solution for a zurl format was found. This common format keeps
credentials in the query, therefore we should revert patch
cf685fb5.

This reverts commit 6047f893 in order to
replace it with py upstream commit nexedi/neoppod@22ccebd6.

NEO/go and NEO/py zurl format diverged over time:

- kirr/neo@8c974485

However with nexedi/neoppod!21 a
common solution was found. From there, this patch aims to adjust NEO/go zurl
format to be in sync with NEO/py zurl format again.

* master:
  go/zodb: Handle common options in zurl in generic layer

/reviewed-by @kirr
/reviewed-on kirr/neo!4

* kirr/t+new-uri:
  Revert "Y client: Adjust URI scheme to move client-specific options to fragment"
  fixup! client.go: Fix URI client option parsing for supported + unsupported options
  client.go: Fix URI client option parsing for supported + unsupported options
  fixup! client_test: Add tests for NEO URI parser
  client_test: Add tests for NEO URI parser
  fixup! client: Refactor openClientByURL for easier testing
  client: Refactor openClientByURL for easier testing
  Y go/zodb: Handle common options in zurl in generic layer

Offload drivers from handling options such as ?read-only=1 and force
them to deal with such options only via DriverOptions, never zurl.

See added comment for details.

/reviewed-by @levin.zimmermann
/reviewed-on !4

This reverts commit kirr/neo@4c9414ea.
This patch was added at a time when nexedi/neoppod!18 wasn't
resolved yet, but we already wanted to proceed with WCFS. Now the NEO MR
is resolved and we decided to mostly leave the NEO zurl as it was
originally implemented in nexedi/neoppod!6.
This means we don't need this patch anymore which changed the NEO
zurl format.

readonly is handled by common zodb.OpenDriver.

Before this patch, the parser ignored options which were already supported
by the client (for instance 'read-only') and even raised an error. But the
client can already use this option: as a9246333 describes this should happen in the
local storage URL parser.

Furthermore not-yet-supported client options (for instance compress) broke
the NEO client before this patch. Now these options only raise a warning which
informs the user that they are ignored. Why? We want to use pre-complete NEO
in real-world projects together with NEO/py clusters. Those real-world projects
may already specify options which aren't supported by our NEO/go client yet.
But it doesn't matters so much, because those options are mostly
relevant for other NEO/py cluster clients (e.g. zope nodes). Instead of
filtering those parameters before parsing them to NEO/go in a higher
level (e.g. SlapOS), NEO/go should already support any valid NEO URL
and raise warnings for not yet implemented features.

- use simplified parseURL signature - DriverOptions are not passed nor changed there.
- read-only is handled by generic zodb layer not neo.parseURL .

This test was missing so far. Particularly recent changes of the
NEO URI scheme [1], but also problems with valid old URI [2] stressed
out the necessity for comprehensive NEO URI parser tests.

[1] kirr/neo@4c9414ea
[2] 573514c6 (comment 184417)

- no need to pass DriverOptions into parseURL - it is only zurl that is
  parsed, and also DriverOptions should not be changed by the opener.

- no need to document "If anything fails within this process an error
  and nil are returned." because that is standard omnipresent Go convention.

With all the recent changes of the NEO URI scheme we need to reliably test
the function which parses the URI and convert it into the different
parameter. Testing is much simpler if we can only analyse how the URI
parsing works. Therefore this patch moves NEO URI parsing to an
external function.

Offload drivers from handling options such as ?read-only=1 and force
them to deal with such options only via DriverOptions, never zurl.

See added commend for details.

I missed the following build failure in go/neo/cmd:

    # lab.nexedi.com/kirr/neo/go/neo/cmd/neo
    ./storage.go:128:37: cannot use master (variable of type string) as []string value in argument to neo.NewStorage

See kirr/neo!2 for discussion,
context and details.

/reviewed-by @kirr

* t-with-multiple-master-nodes:
  fixup! client_test: Add nmaster={1,2} to test matrix
  fixup! client_test: Support test cluster /w >1 master
  fixup! TalkMaster: Switch master if dialed M is secondary
  fixup! Node: Add support for NEO cluster with > 1 master
  fixup! Dial: Catch NotPrimaryMaster & return custom error
  fixup! proto: Implement Error for NotPrimaryMaster
  fixup! proto.NotPrimaryMaster: Fix .Primary data type (2)
  fixup! proto.NotPrimaryMaster: Fix .Primary data type (1)
  client_test: Add nmaster={1,2} to test matrix
  client_test: Support test cluster /w >1 master
  proto.NotPrimaryMaster: Fix .Primary data type
  TalkMaster: Switch master if dialed M is secondary
  Dial: Catch NotPrimaryMaster & return custom error
  proto: Implement Error for NotPrimaryMaster
  openClientByURL: Fix for >1 master (split URL host)
  Client.URL: Fix incomplete URL if > 1 master nodes
  Node: Add support for NEO cluster with > 1 master

Actually do test nmaster=2 case.

- show all options in error context and ran test kind

- skip nmaster > 1 for NEO/go as that is currently not implemented on NEO/go server

- use json for interacting with runneo.py, so that we can use
  whatever builtin type for any argument without hardcoding ad-hoc
  handling of specific arguments inside runneo.py.

- adjust comments + cosmetics

- validate received NotPrimaryMaster
- use Address.String() instead of printf with format that works only for ipv6
- add some logging, comments and TODO

- no need to keep Node.MasterAddr anymore - the address of current PM is
  managed by TalkMaster and is provided as part of operational context
  to user functions that TalkMaster runs.

- correct docstrings.

- cosmetics.

Expect NotPrimaryMaster only if we are trying to connect to a master.

Provide details in the error message.

Change .Promary type from int8 back to int32.

5d93e434 says that .Primary type is not NodeID. That is true, but
changing it to int8 was a mistake:

1. PSignedNull is explicitly defined to come with '!l' struct code, which
   according to https://docs.python.org/3/library/struct.html#module-struct
   comes as 4-bytes integer on the wire:

   https://lab.nexedi.com/nexedi/neoppod/blob/v1.12-13-gf2ea4be2/neo/lib/protocol.py#L560-562

2. verifying this via serializing NotPrimaryMaster on NEO/py also
   confirms that .Primary occupies 4 bytes, not one:

   In [1]: from neo.lib.protocol import NotPrimaryMaster

   In [2]: NotPrimaryMaster(0x01020304, [('m111', 111), ('m222', 222)])._body
   Out[2]: '\x01\x02\x03\x04\x00\x00\x00\x02\x00\x00\x00\x04m111\x00o\x00\x00\x00\x04m222\x00\xde'
            ^^^^^^^^^^^^^^^^
            NOTE  NOTE  NOTE

-> change .Primary type back to being 4-bytes integer, but to int32
instead of NodeID because, as 5d93e434 correctly says, .Primary comes as
array index, not a node ID. The following place of NEO/py code
explicitly confirms this:

https://lab.nexedi.com/nexedi/neoppod/blob/v1.12-13-gf2ea4be2/neo/master/handlers/identification.py#L155-159

Add corresponding test.

Rerun `go generate`. As the diff in zproto-marshal.go shows changing
NotPrimaryMaster.Primary type from NodeID to int8 actually does make a
difference. This happens because NodeID type is based on int32 and
changing that to int8 changes how NotPrimaryMaster structure is layed
out in memory and on the wire.

The changes to zproto-marshal.go in 5d93e434 seem to be done by hand
and not matching the change to proto.go even though head of
zproto-marshal.go says

	// Code generated by protogen.go; DO NOT EDIT.

Tests should work with both one master or more than one masters.

Now it's possible to run client tests against a NEO cluster which
has more than one master nodes. We need this adjustement in order
to test NEO/go client modification in order to support more than
one master node.

The '.Primary' attribute of the 'NotPrimaryMaster' packet has been
assigned to 'NodeID' data type. This is incorrect, because the data
doesn't represent the ID of the node, but an index of the
'.KnownMasterList' [1]. In the old protocol NEO/py therefore also
used 'PSignedNull' instead of 'PUUID' [2]. This patches fixes the data
type of '.Primary' and uses 'int8' instead of 'NodeID'. Technically this
doesn't make any difference, but semantically for human beings the code
is easier to understand now.

[1] https://lab.nexedi.com/nexedi/neoppod/blob/c6453626/neo/lib/handler.py#L161
[2] https://lab.nexedi.com/nexedi/neoppod/blob/c6453626/neo/lib/protocol.py#L716

When connecting to a master node, the client needs to try a different
master if the initially tried one is a secondary master node. This
statement wasn't implemented yet before this patch and therefore it was
good luck if the initally tried master was the primary one - and the
connection worked - or if it was a secondary master - and the client
got stuck in re-trying the same node forever. This patch makes NEO/go
usage with clusters of more than one master therefore much more stable.

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.