A deadlock scenario was possible before this patch:

- main thread finished checks, wants to return, goes to deferred wg.Wait();
- that wg.Wait waits for other spawned threads to complete;
- in one of that thread there is extra event being sent to a stream;
- that send is blocked waiting for main thread to receive it, but the
  main thread is waiting in wg.Wait() and only after that it would go to
  call T.closeStreamTab() which marks all the channels as down -> deadlock.

-> Fix it by explicitly detecting deadlocks not only in Recv, but also on Send.

NOTE Contrary to Recv - which is always called from main thread - Send is
always called from non-main threads. T.unsentv and the logic prepared in
the previous patch care to handle that correctly.

If we don't, it won't be automatically the case when Send's will be
adjusted to detect and log deadlocks. And then, if there will be no
ordering, it will be hard to do tests in example_test.go which expects
logging output in particular "main -> everything else" order.

This is possible if the system is incorrectly decomposed into serial
streams. We want to show all pending events in that case, not only the
first random one.

-> Reuse T.mu for serialization.

closeStreamTab changes locking so that acquired t.mu covers whole time
when that function could log anything.

There is no guaranty that the main thread would call Fatal or Fail if a
non-main thread calls fatalfInNonMain. This way it was possible for a
test to succeed if only non-main thread(s) failed somehow.

-> Fix it by marking the test as failed if any non-main thread fails.

Generate g for today state of Go 1.18 (go1.17beta1-2522-g47db3bb443)
Compared to Go1.17 there are small number of changes in _defer struct.

Regenerated files stay without changes for Go1.17 and previous releases.

---- 8< ----
diff --git a/zruntime_g_go1.17.go b/zruntime_g_go1.18.go
index 7a76db6..5e2fafb 100644
--- a/zruntime_g_go1.17.go
+++ b/zruntime_g_go1.18.go
@@ -1,6 +1,6 @@
 // Code generated by g_typedef; DO NOT EDIT.

-// +build go1.17,!go1.18
+// +build go1.18,!go1.19

 package xruntime

@@ -115,7 +115,6 @@ type _panic struct {
 	goexit    bool
 }
 type _defer struct {
-	siz     int32 // includes both arguments and results
 	started bool
 	heap    bool
 	// openDefer indicates that this _defer is for a frame with open-coded
@@ -124,9 +123,9 @@ type _defer struct {
 	openDefer bool
 	sp        uintptr // sp at time of defer
 	pc        uintptr // pc at time of defer
-	fn        *funcval // can be nil for open-coded defers
+	fn        func()  // can be nil for open-coded defers
 	_panic    *_panic // panic that is running defer
-	link      *_defer
+	link      *_defer // next defer on G; can point to either heap or stack!

 	// If openDefer is true, the fields below record values about the stack
 	// frame and associated function that has the open-coded defer(s). sp

Generate g for ~ Go 1.17beta1 (go1.17beta1-2-g770f1de8c5)
Compared to Go1.16 there is one non-cosmetic change to g related to tracking
scheduling latencies:

https://github.com/golang/go/commit/bedfeed54a7a

Regenerated files stay without changes for Go1.16 and previous releases.

---- 8< ----
diff --git a/zruntime_g_go1.16.go b/zruntime_g_go1.17.go
index 9604d6b..7a76db6 100644
--- a/zruntime_g_go1.16.go
+++ b/zruntime_g_go1.17.go
@@ -1,6 +1,6 @@
 // Code generated by g_typedef; DO NOT EDIT.

-// +build go1.16,!go1.17
+// +build go1.17,!go1.18

 package xruntime

@@ -25,7 +25,18 @@ type g struct {
        syscallsp uintptr // if status==Gsyscall, syscallsp = sched.sp to use during gc
        syscallpc uintptr // if status==Gsyscall, syscallpc = sched.pc to use during gc
        stktopsp  uintptr // expected sp at top of stack, to check in traceback
-       param        unsafe.Pointer // passed parameter on wakeup
+       // param is a generic pointer parameter field used to pass
+       // values in particular contexts where other storage for the
+       // parameter would be difficult to find. It is currently used
+       // in three ways:
+       // 1. When a channel operation wakes up a blocked goroutine, it sets param to
+       //    point to the sudog of the completed blocking operation.
+       // 2. By gcAssistAlloc1 to signal back to its caller that the goroutine completed
+       //    the GC cycle. It is unsafe to do so in any other way, because the goroutine's
+       //    stack may have moved in the meantime.
+       // 3. By debugCallWrap to pass parameters to a new goroutine because allocating a
+       //    closure in the runtime is forbidden.
+       param        unsafe.Pointer
        atomicstatus uint32
        stackLock    uint32 // sigprof/scang lock; TODO: fold in to atomicstatus
        goid         int64
@@ -57,6 +68,10 @@ type g struct {

        raceignore     int8     // ignore race detection events
        sysblocktraced bool     // StartTrace has emitted EvGoInSyscall about this goroutine
+       tracking       bool     // whether we're tracking this G for sched latency statistics
+       trackingSeq    uint8    // used to decide whether to track this G
+       runnableStamp  int64    // timestamp of when the G last became runnable, only used when tracking
+       runnableTime   int64    // the amount of time spent runnable, cleared when running, only used when tracking
        sysexitticks   int64    // cputicks when syscall has returned (for tracing)
        traceseq       uint64   // trace event sequencer
        tracelastp     puintptr // last P emitted an event for this goroutine
@@ -142,7 +157,7 @@ type gobuf struct {
        pc   uintptr
        g    guintptr
        ctxt unsafe.Pointer
-       ret  uintreg
+       ret  uintptr
        lr   uintptr
        bp   uintptr // for framepointer-enabled architectures
 }

Tracetest is a package to verify concurrent systems based on synchronous
tracing. It is used in NEO/go tests and originates from what was
initially explained in https://navytux.spb.ru/~kirr/neo.html in

    """On top of that package tracetest provides infrastructure for
    testing concurrent systems..."""

See top-level package overview + example_test.go for details.

Some NEO/go tests need to start a cluster under tracing and then to
continue running test workload with tracing disabled.

NetTrace now returns Tracer instead of general Networker. The reason we
need this is because in the next patch Tracer will implement methods,
additional to just what is required by Networker, to control itself.

Rename the interface that specifies methods to be implemented by trace
receivers. Reason is that in the next patch we'll need to expose
"Tracer" as the tracing networker itself.

Generate g for ~ Go 1.16beta1 (go1.16beta1-34-g4fd9455882)
Compared to Go1.15 there are only cosmetic changes:

    --- a/zruntime_g_go1.15.go
    +++ b/zruntime_g_go1.16.go
    @@ -1,6 +1,6 @@
     // Code generated by g_typedef; DO NOT EDIT.

    -// +build go1.15,!go1.16
    +// +build go1.16,!go1.17

     package xruntime

    @@ -144,7 +144,7 @@ type gobuf struct {
     	ctxt unsafe.Pointer
     	ret  uintreg
     	lr   uintptr
    -	bp   uintptr // for GOEXPERIMENT=framepointer
    +	bp   uintptr // for framepointer-enabled architectures
     }
     type funcval struct {
     	fn uintptr
    @@ -158,6 +158,8 @@ type timer struct {
     	// Timer wakes up at when, and then at when+period, ... (period > 0 only)
     	// each time calling f(arg, now) in the timer goroutine, so f must be
     	// a well-behaved function and not block.
    +	//
    +	// when must be positive on an active timer.
     	when   int64
     	period int64
     	f      func(interface{}, uintptr)

Regenerated files stay without changes for Go1.15 and previous releases.

When a virtnet network - e.g. lonet - is explicitly used in tests, the
open and close sequence is easy to implement:

    network = lonet.Join(...)
    defer network.Close()

    hostα = network.NewHost("α")
    defer hostα.Close()
    hostβ = network.NewHost("β")
    defer hostβ.Close()
    ...

However in other scenarios - for example when the code wants to call
xnet.Networker factory that creates a networker out of several potential
choices according to given parameters, only interface for network
access-point (e.g. Host in lonet/virtnet case) is returned. The network
itself (virtnet.SubNetwork) is not returned to the user. As it is this
makes it impossible to close the network and release its resources.

Instead of adding complexity to such "create-networker" functions to
e.g. change Close of returned networker to also close subnetwork, or
return both networker and a separate close func, let's teach virtnet to
close SubNetwork automatically when last host of this subnetwork is
closed. Make it opt-in since this behaviour is not universally wanted.

This way a factory that is asked to connect to network as e.g. a node on
lonet, could do the following:

    func neonet.Join(...) xnet.Networker {
        ...
        // join lonet "<net>" as host "<node>"
        network = lonet.Join("<net>")
        host = network.NewHost("<node>")
        network.AutoClose() // host.Close will close network
        return host
    }

and network will be closed on host.Close() call.

-> Add SubNetwork.AutoClose to schedule Close to be called after last host on the subnetwork is closed.
-> Mirror this change in virtnet part of lonet.py

Even though there is no need to manually release resources for NetPlain
and NetTLS, for other networkers - for example lonet - releasing
resources is needed. In general programs cannot always use lonet
explicitly, as there might be a factory function that creates a
Networker according to given parameters.

-> Adjust Networker interface to include general Close.
-> Adjust NetPlain to follow added "interrupt-on-close" semantic for uniformity.

Similarly to what we do in package xio (see 7ad867a3 "xio: Add
Reader/Writer/... interfaces - io analogs that add support for
contexts"), add utilities to bind xnet.Listener to context (giving
net.Listener), and, on the other hand, adapt uncontextified net.Listener
to xnet.Listener that handles cancellation.

This utilities are needed because a lot of third-party code wants to
work with net.Listener interface only.

Contrary to xio, we can make it working reliably in both ways (because
Accept already buffers accepting connections starting at OS level).

In the previous patch xnet.Networker was changed: Listen now accepts ctx
and returns xnet.Listener instead of net.Listener.

-> Adapt the code all around to that.

We already handle cancellation in Dial, but Accepting was out of luck
until now. This makes it more difficult for clients to implement and
wrap acceptors where they need to handle cancellations. This also makes
it possible for a test or program to get stuck in Accept loop if it is
not careful enough to manually handle ctx cancel around Accept calls.

-> Fix it in one place - here, in xnet - so that users are offloaded
from all this and can just call Accept(ctx) and rely on underlying
implementation to handle ctx cancel.

This patch:

- introduces xnet.Listener interface, which is like net.Listener, but
  Accept goes with ctx argument.
- changes Networker.Listen signature to return xnet.Listener instead of
  net.Listener. While we are here - changing it - also add ctx argument
  to Listen call itself.
- Adds listenerCtx - which, given net.Listener, provides xnet.Listener
  by wrapping some logic around original.
- Adapts NetPlain, NetTLS and NetTrace to provide updated interface.

We'll fix up everything in other packages to match/use updated interface
in the next patch.

See nexedi/pygolang@21756bd3.

Go1.14.10 and Go1.15.3 changed internal structure of g:

https://github.com/golang/go/commit/878da0bf881b (go1.14)
https://github.com/golang/go/commit/bf79f91d3dc4 (go1.15)

A shorthand for when exception context is just formatted string.
Simplifies

	func doSomething(path string) {
		defer exc.Context(func() interface{} {
			return fmt.Sprintf("doing something %s", path)
		})()
		...

to

	func doSomething(path string) {
		defer exc.Contextf("doing something %s", path)
		...

`go mod init` + build/test that fetched in dependencies.

It was address-of error (the type) instead of &err.

https://github.com/golang/go/issues/36503
https://github.com/golang/go/issues/36448

They are there for 6 months already without being considered for real
and so the progress is very unlikely, but still...

Generate g for ~ Go 1.15beta1 (go1.14beta1-1729-gc0e8e405c0)
Compared to Go1.14 there are no changes.

Regenerated files stay without changes for Go1.14 and previous releases.

Provide race.Enabled for programs to know whether they were built with
race detector or not.

Code originates from https://lab.nexedi.com/kirr/wendelin.core/tree/25c3184d/wcfs/internal/race
See also: https://github.com/golang/go/issues/36477.

So that Go1.13 and golang.org/x/xerrors understand the wrapping:
https://blog.golang.org/go1.13-errors

Preserve Cause for compatibility with github.com/pkg/errors.

Add xsync.WorkGroup that is similar to https://godoc.org/golang.org/x/sync/errgroup,
but amends it with a bit better (imho) design where work context is
explicitly passed into worker (see also [1]). This mirrors sync.WorkGroup that we
already have at Pygolang side [2].

[1] https://github.com/golang/go/issues/34510
[2] https://pypi.org/project/pygolang/#concurrency

I need this to support sysread(/head/watch) cancellation in WCFS
filesystem [1,2,3].

[1] wendelin.core@b17aeb8c
[2] wendelin.core@f05271b1
[3] wendelin.core@5ba816da

Rework Pipe to create (xio.Reader, xio.Writer) instead of (io.Reader,
io.Writer) and teach xio.Pipe{Reader,Writer} to accept ctx argument and
handle cancellation.

I need this to support sysread(/head/watch) cancellation in WCFS
filesystem [1,2,3]. See also [4].

[1] wendelin.core@b17aeb8c
[2] wendelin.core@f05271b1
[3] wendelin.core@5ba816da
[4] https://github.com/golang/go/issues/20280

Just make it compile by trivially adapting the code to refer thing from
io package. Change license reference to the place where Go license lives
in go123 (LICENSE-go, added in ad78da1b "xflag: Add counting flag from
go.git internals").

This is pristine copy of pipe.go + tests from Go go1.14beta1-24-g075c20cea8 .
We'll adapt copied code to xio and add cancellation support in further commits.

Done toprovide uniform API in xio package.
Compensate for API change in places where xio.CountReader was used.

In many cases IO needs to be able to be canceled. For example in WCFS
filesystem I need to cancel handling sysread(/head/watch) when FUSE
INTERRUPT request comes in [1,2,3]. The READ handler for /head/watch
inside WCFS is interally implemented via io.Pipe which does not provide
read/write cancellattion besides "destructive" close.

Standard Go answer for cancellations is via contexts. So as a first step
let's add corresponding interfaces - xio.Reader, xio.Writer etc - that
are io analogs that add support for contexts.

For compatibility with legacy code that work with only io.X (e.g.  only
with io.Reader), in spirit of [4], add BindCtx which binds xio.X
instance with context and converts it into io.X. Add WithCtx -
corresponding inverse operation that converts io.X back into xio.X and
for general io.X adds cancellation handling on a best-effort basis.

[1] wendelin.core@b17aeb8c
[2] wendelin.core@f05271b1
[3] wendelin.core@5ba816da
[4] https://github.com/golang/go/issues/20280
[5] https://github.com/golang/go/issues/16522