blob: 8f0f164bc696d30692990d817cb52af23c457846 [file] [log] [blame]
# persist.py
#
# Implement limited persistence.
#
# Simple interface:
# persist.save() save __main__ module on file (overwrite)
# persist.load() load __main__ module from file (merge)
#
# These use the filename persist.defaultfile, initialized to 'wsrestore.py'.
#
# A raw interface also exists:
# persist.writedict(dict, fp) save dictionary to open file
# persist.readdict(dict, fp) read (merge) dictionary from open file
#
# Internally, the function dump() and a whole bunch of support of functions
# traverse a graph of objects and print them in a restorable form
# (which happens to be a Python module).
#
# XXX Limitations:
# - Volatile objects are dumped as strings:
# - open files, windows etc.
# - Other 'obscure' objects are dumped as strings:
# - classes, instances and methods
# - compiled regular expressions
# - anything else reasonably obscure (e.g., capabilities)
# - type objects for obscure objects
# - It's slow when there are many of lists or dictionaries
# (This could be fixed if there were a quick way to compute a hash
# function of any object, even if recursive)
defaultfile = 'wsrestore.py'
def save():
import __main__
import os
# XXX On SYSV, if len(defaultfile) >= 14, this is wrong!
backup = defaultfile + '~'
try:
os.unlink(backup)
except os.error:
pass
try:
os.rename(defaultfile, backup)
except os.error:
pass
fp = open(defaultfile, 'w')
writedict(__main__.__dict__, fp)
fp.close()
def load():
import __main__
fp = open(defaultfile, 'r')
readdict(__main__.__dict__, fp)
def writedict(dict, fp):
import sys
savestdout = sys.stdout
try:
sys.stdout = fp
dump(dict) # Writes to sys.stdout
finally:
sys.stdout = savestdout
def readdict(dict, fp):
contents = fp.read()
globals = {}
exec(contents, globals)
top = globals['top']
for key in top.keys():
if dict.has_key(key):
print 'warning:', key, 'not overwritten'
else:
dict[key] = top[key]
# Function dump(x) prints (on sys.stdout!) a sequence of Python statements
# that, when executed in an empty environment, will reconstruct the
# contents of an arbitrary dictionary.
import sys
# Name used for objects dict on output.
#
FUNNYNAME = FN = 'A'
# Top-level function. Call with the object you want to dump.
#
def dump(x):
types = {}
stack = [] # Used by test for recursive objects
print FN, '= {}'
topuid = dumpobject(x, types, stack)
print 'top =', FN, '[', `topuid`, ']'
# Generic function to dump any object.
#
dumpswitch = {}
#
def dumpobject(x, types, stack):
typerepr = `type(x)`
if not types.has_key(typerepr):
types[typerepr] = {}
typedict = types[typerepr]
if dumpswitch.has_key(typerepr):
return dumpswitch[typerepr](x, typedict, types, stack)
else:
return dumpbadvalue(x, typedict, types, stack)
# Generic function to dump unknown values.
# This assumes that the Python interpreter prints such values as
# <foo object at xxxxxxxx>.
# The object will be read back as a string: '<foo object at xxxxxxxx>'.
# In some cases it may be possible to fix the dump manually;
# to ease the editing, these cases are labeled with an XXX comment.
#
def dumpbadvalue(x, typedict, types, stack):
xrepr = `x`
if typedict.has_key(xrepr):
return typedict[xrepr]
uid = genuid()
typedict[xrepr] = uid
print FN, '[', `uid`, '] =', `xrepr`, '# XXX'
return uid
# Generic function to dump pure, simple values, except strings
#
def dumpvalue(x, typedict, types, stack):
xrepr = `x`
if typedict.has_key(xrepr):
return typedict[xrepr]
uid = genuid()
typedict[xrepr] = uid
print FN, '[', `uid`, '] =', `x`
return uid
# Functions to dump string objects
#
def dumpstring(x, typedict, types, stack):
# XXX This can break if strings have embedded '\0' bytes
# XXX because of a bug in the dictionary module
if typedict.has_key(x):
return typedict[x]
uid = genuid()
typedict[x] = uid
print FN, '[', `uid`, '] =', `x`
return uid
# Function to dump type objects
#
typeswitch = {}
class some_class:
def method(self): pass
some_instance = some_class()
#
def dumptype(x, typedict, types, stack):
xrepr = `x`
if typedict.has_key(xrepr):
return typedict[xrepr]
uid = genuid()
typedict[xrepr] = uid
if typeswitch.has_key(xrepr):
print FN, '[', `uid`, '] =', typeswitch[xrepr]
elif x == type(sys):
print 'import sys'
print FN, '[', `uid`, '] = type(sys)'
elif x == type(sys.stderr):
print 'import sys'
print FN, '[', `uid`, '] = type(sys.stderr)'
elif x == type(dumptype):
print 'def some_function(): pass'
print FN, '[', `uid`, '] = type(some_function)'
elif x == type(some_class):
print 'class some_class: pass'
print FN, '[', `uid`, '] = type(some_class)'
elif x == type(some_instance):
print 'class another_class: pass'
print 'some_instance = another_class()'
print FN, '[', `uid`, '] = type(some_instance)'
elif x == type(some_instance.method):
print 'class yet_another_class:'
print ' def method(): pass'
print 'another_instance = yet_another_class()'
print FN, '[', `uid`, '] = type(another_instance.method)'
else:
# Unknown type
print FN, '[', `uid`, '] =', `xrepr`, '# XXX'
return uid
# Initialize the typeswitch
#
for x in None, 0, 0.0, '', (), [], {}:
typeswitch[`type(x)`] = 'type(' + `x` + ')'
for s in 'type(0)', 'abs', '[].append':
typeswitch[`type(eval(s))`] = 'type(' + s + ')'
# Dump a tuple object
#
def dumptuple(x, typedict, types, stack):
item_uids = []
xrepr = ''
for item in x:
item_uid = dumpobject(item, types, stack)
item_uids.append(item_uid)
xrepr = xrepr + ' ' + item_uid
del stack[-1:]
if typedict.has_key(xrepr):
return typedict[xrepr]
uid = genuid()
typedict[xrepr] = uid
print FN, '[', `uid`, '] = (',
for item_uid in item_uids:
print FN, '[', `item_uid`, '],',
print ')'
return uid
# Dump a list object
#
def dumplist(x, typedict, types, stack):
# Check for recursion
for x1, uid1 in stack:
if x is x1: return uid1
# Check for occurrence elsewhere in the typedict
for uid1 in typedict.keys():
if x is typedict[uid1]: return uid1
# This uses typedict differently!
uid = genuid()
typedict[uid] = x
print FN, '[', `uid`, '] = []'
stack.append(x, uid)
item_uids = []
for item in x:
item_uid = dumpobject(item, types, stack)
item_uids.append(item_uid)
del stack[-1:]
for item_uid in item_uids:
print FN, '[', `uid`, '].append(', FN, '[', `item_uid`, '])'
return uid
# Dump a dictionary object
#
def dumpdict(x, typedict, types, stack):
# Check for recursion
for x1, uid1 in stack:
if x is x1: return uid1
# Check for occurrence elsewhere in the typedict
for uid1 in typedict.keys():
if x is typedict[uid1]: return uid1
# This uses typedict differently!
uid = genuid()
typedict[uid] = x
print FN, '[', `uid`, '] = {}'
stack.append(x, uid)
item_uids = []
for key in x.keys():
val_uid = dumpobject(x[key], types, stack)
item_uids.append(key, val_uid)
del stack[-1:]
for key, val_uid in item_uids:
print FN, '[', `uid`, '][', `key`, '] =',
print FN, '[', `val_uid`, ']'
return uid
# Dump a module object
#
def dumpmodule(x, typedict, types, stack):
xrepr = `x`
if typedict.has_key(xrepr):
return typedict[xrepr]
from string import split
# `x` has the form <module 'foo'>
name = xrepr[9:-2]
uid = genuid()
typedict[xrepr] = uid
print 'import', name
print FN, '[', `uid`, '] =', name
return uid
# Initialize dumpswitch, a table of functions to dump various objects,
# indexed by `type(x)`.
#
for x in None, 0, 0.0:
dumpswitch[`type(x)`] = dumpvalue
for x, f in ('', dumpstring), (type(0), dumptype), ((), dumptuple), \
([], dumplist), ({}, dumpdict), (sys, dumpmodule):
dumpswitch[`type(x)`] = f
# Generate the next unique id; a string consisting of digits.
# The seed is stored as seed[0].
#
seed = [0]
#
def genuid():
x = seed[0]
seed[0] = seed[0] + 1
return `x`