changeset 42963:8502f76dbfd7

transaction: detect an attempt to truncate-to-extend on playback, raise error

On some networked filesystems, writes can have delayed finalization/confirmation
and write races can occur such that a remote modification will "win" and
modifications will be lost. There is no functionality for providing this
feedback to userspace programs (in fact, there's not even functionality for
providing this information to the Linux kernel...), so these programs may see
the files suddenly change.

We've noticed that there have been cases where Mercurial has detected something
has gone wrong and attempts to abort (rolling back the transaction), which is
good. However, when rolling back the transaction, for the append-only files,
we attempt to "truncate" the file back to the size it was in before the hg
transaction started, but end up *extending* it. This may be harmless, but if
this happens to the 00changelog.i file, we get a bunch of nulls on the end of
the file and this causes hg to become *really* confused. :)

If we detect that some modification of the file outside of this Mercurial
process has caused the file to be smaller than the size we are attempting to
truncate to, let's just exit and stop trying to clean up the repository -
continuing will likely just cause more damage.

Differential Revision: https://phab.mercurial-scm.org/D6867

--- a/mercurial/transaction.py	Tue Sep 17 15:09:25 2019 -0700
+++ b/mercurial/transaction.py	Tue Sep 17 14:01:26 2019 -0700
@@ -54,6 +54,10 @@
             checkambig = checkambigfiles and (f, '') in checkambigfiles
             try:
                 fp = opener(f, 'a', checkambig=checkambig)
+                if fp.tell() < o:
+                    raise error.Abort(_(
+                            "attempted to truncate %s to %d bytes, but it was "
+                            "already %d bytes\n") % (f, o, fp.tell()))
                 fp.truncate(o)
                 fp.close()
             except IOError: