# HG changeset patch
# User Anton Shestakov <av6@dwimlabs.net>
# Date 1559902488 -28800
# Node ID 313565dd75e3117d49363e371205f45951523ab1
# Parent  f1466f5ffbf59c8acabf47adf654584bc6f99aff
pick: remove transaction on the whole command (issue6037)

At its core, pick is a pretty straightforward and well-behaving command, it
uses functions already in core hg, it checks that wdir is clean and that
changeset to pick is not public, it checks if there happen to be merge
conflicts and can be --continue'd later, etc.

It is very similar to graft in core (it also uses mergemod.graft function), but
it obsoletes the original changeset. However, graft does not experience this
incorrect behavior from issue 6037.

What happens in the test case for this issue when we pick a revision that
touches both "a" and "b": mergemod.graft() takes the original changeset and
tries to apply it to the wdir, which results in "b" being marked as newly added
and ready to be committed, "a" updated with the new content and being marked as
modified, but "a" also has conflicts. Pick correctly notices this and saves its
state before asking for user intervention. So far so good. However, when the
command raises InterventionRequired to print a user-facing message and exit
while being wrapped in repo.transaction() context manager, the latter partially
undoes what mergemod.graft() did: it unmarks "b" as added. And when user
continues pick, "b" is therefore not tracked and is not included in the
resulting commit.

The transaction is not useful here, because it doesn't touch wdir (it's still
dirty), it doesn't remove pickstate (and other commands will refuse to work
until pick --abort or --continue), it just makes "b" untracked.

The solution is to use repo.transaction() only to wrap code that writes data to
hg store in the final stages of the command after all checks have passed and is
not expected to fail on trivial cases like merge conflicts. For example,
committing the picked changeset. But since pick uses repo.commit() for that,
and because that function already uses a transaction, wrapping it in another
transaction doesn't make sense.

diff -r f1466f5ffbf5 -r 313565dd75e3 CHANGELOG
--- a/CHANGELOG	Sun Dec 23 01:02:36 2018 +0800
+++ b/CHANGELOG	Fri Jun 07 18:14:48 2019 +0800
@@ -1,6 +1,11 @@
 Changelog
 =========
 
+9.0.1 - in progress
+-------------------
+
+  * pick: no longer forget file in case of conflict (issue6037)
+
 9.0.0 -- 2019-06-06
 -------------------
 
diff -r f1466f5ffbf5 -r 313565dd75e3 hgext3rd/evolve/cmdrewrite.py
--- a/hgext3rd/evolve/cmdrewrite.py	Sun Dec 23 01:02:36 2018 +0800
+++ b/hgext3rd/evolve/cmdrewrite.py	Fri Jun 07 18:14:48 2019 +0800
@@ -1431,7 +1431,7 @@
         revs.append(opts['rev'])
 
     overrides = {('ui', 'forcemerge'): opts.get('tool', '')}
-    with repo.wlock(), repo.lock(), repo.transaction('pick'), ui.configoverride(overrides, 'pick'):
+    with repo.wlock(), repo.lock(), ui.configoverride(overrides, 'pick'):
         pickstate = state.cmdstate(repo, path='pickstate')
         pctx = repo['.']
 
diff -r f1466f5ffbf5 -r 313565dd75e3 tests/test-pick.t
--- a/tests/test-pick.t	Sun Dec 23 01:02:36 2018 +0800
+++ b/tests/test-pick.t	Fri Jun 07 18:14:48 2019 +0800
@@ -434,12 +434,12 @@
   continue: hg pick --continue
   $ hg pick --continue
 
-But what's this? b was forgotten and is not in the picked changeset!
+Demonstrate that b was not forgotten and is definitely included in 4
 
-  $ hg status b
-  ? b
+  $ hg status b -A
+  C b
   $ hg glf
-  @  4: apricot and banana (a)
+  @  4: apricot and banana (a b)
   |
   o  3: avocado (a)
   |