portage.cache: write md5 instead of mtime (bug 568934)

Change cache modules to write md5 in cache entries, instead of mtime.
Since portage-2.2.27, the relevant cache modules have had the ability
to read cache entries containing either md5 or mtime, therefore this
change is backward-compatible with portage-2.2.27 and later.

Also fix the reconstruct_eclasses function to raise CacheCorruption
when the specified chf_type is md5 and the cache entry contains mtime
data, and optimize __getitem__ to skip reconstruct_eclasses calls when
the entry appears to have a different chf_type.

X-Gentoo-Bug: 568934
X-Gentoo-Bug-url: https://bugs.gentoo.org/show_bug.cgi?id=568934
Acked-by: Alexander Berntsen <bernalex@gentoo.org>

pym/portage/cache/anydbm.py

@@ -36,8 +36,8 @@
 
 class database(fs_template.FsBased):
 
-	validation_chf = 'mtime'
-	chf_types = ('mtime', 'md5')
+	validation_chf = 'md5'
+	chf_types = ('md5', 'mtime')
 
 	autocommits = True
 	cleanse_keys = True

pym/portage/cache/flat_hash.py

@@ -163,5 +163,5 @@ class md5_database(database):
 
 
 class mtime_md5_database(database):
-	validation_chf = 'mtime'
-	chf_types = ('mtime', 'md5')
+	validation_chf = 'md5'
+	chf_types = ('md5', 'mtime')

pym/portage/cache/sqlite.py

@@ -18,8 +18,8 @@
 
 class database(fs_template.FsBased):
 
-	validation_chf = 'mtime'
-	chf_types = ('mtime', 'md5')
+	validation_chf = 'md5'
+	chf_types = ('md5', 'mtime')
 
 	autocommits = False
 	synchronous = False

pym/portage/cache/template.py

@@ -54,6 +54,15 @@ def __getitem__(self, cpv):
 
 		if self.serialize_eclasses and "_eclasses_" in d:
 			for chf_type in chf_types:
+				if '_%s_' % chf_type not in d:
+					# Skip the reconstruct_eclasses call, since it's
+					# a waste of time if it contains a different chf_type
+					# than the current one. In the past, it was possible
+					# for reconstruct_eclasses called with chf_type='md5'
+					# to "successfully" return invalid data here, because
+					# it was unable to distinguish between md5 data and
+					# mtime data.
+					continue
 				try:
 					d["_eclasses_"] = reconstruct_eclasses(cpv, d["_eclasses_"],
 						chf_type, paths=self.store_eclass_paths)
@@ -62,6 +71,9 @@ def __getitem__(self, cpv):
 						raise
 				else:
 					break
+			else:
+				raise cache_errors.CacheCorruption(cpv,
+					'entry does not contain a recognized chf_type')
 
 		elif "_eclasses_" not in d:
 			d["_eclasses_"] = {}
@@ -310,16 +322,31 @@ def serialize_eclasses(eclass_dict, chf_type='mtime', paths=True):
 		for k, v in sorted(eclass_dict.items(), key=_keysorter))
 
 
+def _md5_deserializer(md5):
+	"""
+	Without this validation, it's possible for reconstruct_eclasses to
+	mistakenly interpret mtime data as md5 data, and return an invalid
+	data structure containing strings where ints are expected.
+	"""
+	if len(md5) != 32:
+		raise ValueError('expected 32 hex digits')
+	return md5
+
+
+_chf_deserializers = {
+	'md5': _md5_deserializer,
+	'mtime': long,
+}
+
+
 def reconstruct_eclasses(cpv, eclass_string, chf_type='mtime', paths=True):
 	"""returns a dict when handed a string generated by serialize_eclasses"""
 	eclasses = eclass_string.rstrip().lstrip().split("\t")
 	if eclasses == [""]:
 		# occasionally this occurs in the fs backends.  they suck.
 		return {}
 
-	converter = _unicode
-	if chf_type == 'mtime':
-		converter = long
+	converter = _chf_deserializers.get(chf_type, lambda x: x)
 
 	if paths:
 		if len(eclasses) % 3 != 0:
@@ -340,6 +367,7 @@ def reconstruct_eclasses(cpv, eclass_string, chf_type='mtime', paths=True):
 		raise cache_errors.CacheCorruption(cpv,
 			"_eclasses_ was of invalid len %i" % len(eclasses))
 	except ValueError:
-		raise cache_errors.CacheCorruption(cpv, "_eclasses_ mtime conversion to long failed")
+		raise cache_errors.CacheCorruption(cpv,
+			"_eclasses_ not valid for chf_type {}".format(chf_type))
 	del eclasses
 	return d