tests/test-transaction-rollback-on-sigpipe.t

filemerge: add support for partial conflict resolution by external tool

A common class of merge conflicts is in imports/#includes/etc. It's
relatively easy to write a tool that can resolve these conflicts,
perhaps by naively just unioning the statements and leaving any
cleanup to other tools to do later [1]. Such specialized tools cannot
generally resolve all conflicts in a file, of course. Let's therefore
call them "partial merge tools". Note that the internal simplemerge
algorithm is such a partial merge tool - one that only resolves
trivial "conflicts" where one side is unchanged or both sides change
in the same way.

One can also imagine having smarter language-aware partial tools that
merge the AST. It may be useful for such tools to interactively let
the user resolve any conflicts it can't resolve itself. However,
having the option of implementing it as a partial merge tool means
that the developer doesn't *need* to create a UI for it. Instead, the
user can resolve any remaining conflicts with their regular merge tool
(e.g. `:merge3` or `meld).

We don't currently have a way to let the user define such partial
merge tools. That's what this patch addresses. It lets the user
configure partial merge tools to run. Each tool can be configured to
run only on files matching certain patterns (e.g. "*.py"). The tool
takes three inputs (local, base, other) and resolves conflicts by
updating these in place. For example, let's say the inputs are these:


base:
```
import sys

def main():
  print('Hello')
```

local:
```
import os
import sys

def main():
  print('Hi')
```

other:
```
import re
import sys

def main():
  print('Howdy')
```

A partial merge tool could now resolve the conflicting imports by
replacing the import statements in *all* files by the following
snippet, while leaving the remainder of the files unchanged.

```
import os
import re
import sys
```

As a result, simplemerge and any regular merge tool that runs after
the partial merge tool(s) will consider the imports to be
non-conflicting and will only present the conflict in `main()` to the
user.

Differential Revision: https://phab.mercurial-scm.org/D12356

Test that, when an hg push is interrupted and the remote side receives SIGPIPE,
the remote hg is able to successfully roll back the transaction.

  $ hg init -q remote
  $ hg clone -q ssh://user@dummy/`pwd`/remote local
  $ SIGPIPE_REMOTE_DEBUG_FILE="$TESTTMP/DEBUGFILE"
  $ SYNCFILE1="$TESTTMP/SYNCFILE1"
  $ SYNCFILE2="$TESTTMP/SYNCFILE2"
  $ export SIGPIPE_REMOTE_DEBUG_FILE
  $ export SYNCFILE1
  $ export SYNCFILE2
  $ PYTHONUNBUFFERED=1
  $ export PYTHONUNBUFFERED

On the remote end, run hg, piping stdout and stderr through processes that we
know the PIDs of. We will later kill these to simulate an ssh client
disconnecting.

  $ remotecmd="$RUNTESTDIR/testlib/sigpipe-remote.py"

In the pretxnchangegroup hook, kill the PIDs recorded above to simulate ssh
disconnecting. Then exit nonzero, to force a transaction rollback.


  $ cat >remote/.hg/hgrc <<EOF
  > [hooks]
  > pretxnchangegroup.00-break-things=sh "$RUNTESTDIR/testlib/wait-on-file" 10 "$SYNCFILE2" "$SYNCFILE1"
  > pretxnchangegroup.01-output-things=echo "some remote output to be forward to the closed pipe"
  > pretxnchangegroup.02-output-things=echo "some more remote output"
  > EOF

  $ hg --cwd ./remote tip -T '{node|short}\n'
  000000000000
  $ cd local
  $ echo foo > foo ; hg commit -qAm "commit"

(use quiet to avoid flacky output from the server)

  $ hg push --quiet --remotecmd "$remotecmd"
  abort: stream ended unexpectedly (got 0 bytes, expected 4)
  [255]
  $ cat $SIGPIPE_REMOTE_DEBUG_FILE
  SIGPIPE-HELPER: Starting
  SIGPIPE-HELPER: Redirection in place
  SIGPIPE-HELPER: pipes closed in main
  SIGPIPE-HELPER: SYNCFILE1 detected
  SIGPIPE-HELPER: worker killed
  SIGPIPE-HELPER: creating SYNCFILE2
  SIGPIPE-HELPER: Shutting down
  SIGPIPE-HELPER: Server process terminated with status 255 (no-windows !)
  SIGPIPE-HELPER: Server process terminated with status 1 (windows !)
  SIGPIPE-HELPER: Shut down

The remote should be left in a good state
  $ hg --cwd ../remote tip -T '{node|short}\n'
  000000000000

#if windows

XXX-Windows Broken behavior to be fixed

Behavior on Windows is broken and should be fixed. However this is a fairly
corner case situation and no data are being corrupted. This would affect
central repository being hosted on a Windows machine and accessed using ssh.

This was catch as we setup new CI for Windows. Making the test pass on Windows
was enough of a pain that fixing the behavior set aside for now. Dear and
honorable reader, feel free to fix it.

  $ hg --cwd ../remote recover
  rolling back interrupted transaction
  (verify step skipped, run `hg verify` to check your repository content)

#else

  $ hg --cwd ../remote recover
  no interrupted transaction available
  [1]

#endif