mirror of https://github.com/python/cpython
148 lines
4.2 KiB
Python
148 lines
4.2 KiB
Python
|
# this module is an OS/2 oriented replacement for the grp standard
|
||
|
# extension module.
|
||
|
|
||
|
# written by Andrew MacIntyre, April 2001.
|
||
|
# released into the public domain "as is", with NO WARRANTY
|
||
|
|
||
|
# note that this implementation checks whether ":" or ";" as used as
|
||
|
# the field separator character.
|
||
|
|
||
|
"""Replacement for grp standard extension module, intended for use on
|
||
|
OS/2 and similar systems which don't normally have an /etc/group file.
|
||
|
|
||
|
The standard Unix group database is an ASCII text file with 4 fields per
|
||
|
record (line), separated by a colon:
|
||
|
- group name (string)
|
||
|
- group password (optional encrypted string)
|
||
|
- group id (integer)
|
||
|
- group members (comma delimited list of userids, with no spaces)
|
||
|
|
||
|
Note that members are only included in the group file for groups that
|
||
|
aren't their primary groups.
|
||
|
(see the section 8.2 of the Python Library Reference)
|
||
|
|
||
|
This implementation differs from the standard Unix implementation by
|
||
|
allowing use of the platform's native path separator character - ';' on OS/2,
|
||
|
DOS and MS-Windows - as the field separator in addition to the Unix
|
||
|
standard ":".
|
||
|
|
||
|
The module looks for the group database at the following locations
|
||
|
(in order first to last):
|
||
|
- ${ETC_GROUP} (or %ETC_GROUP%)
|
||
|
- ${ETC}/group (or %ETC%/group)
|
||
|
- ${PYTHONHOME}/Etc/group (or %PYTHONHOME%/Etc/group)
|
||
|
|
||
|
Classes
|
||
|
-------
|
||
|
|
||
|
None
|
||
|
|
||
|
Functions
|
||
|
---------
|
||
|
|
||
|
getgrgid(gid) - return the record for group-id gid as a 4-tuple
|
||
|
|
||
|
getgrnam(name) - return the record for group 'name' as a 4-tuple
|
||
|
|
||
|
getgrall() - return a list of 4-tuples, each tuple being one record
|
||
|
(NOTE: the order is arbitrary)
|
||
|
|
||
|
Attributes
|
||
|
----------
|
||
|
|
||
|
group_file - the path of the group database file
|
||
|
|
||
|
"""
|
||
|
|
||
|
import os
|
||
|
|
||
|
# try and find the group file
|
||
|
__group_path = []
|
||
|
if os.environ.has_key('ETC_GROUP'):
|
||
|
__group_path.append(os.environ['ETC_GROUP'])
|
||
|
if os.environ.has_key('ETC'):
|
||
|
__group_path.append('%s/group' % os.environ['ETC'])
|
||
|
if os.environ.has_key('PYTHONHOME'):
|
||
|
__group_path.append('%s/Etc/group' % os.environ['PYTHONHOME'])
|
||
|
|
||
|
group_file = None
|
||
|
for __i in __group_path:
|
||
|
try:
|
||
|
__f = open(__i, 'r')
|
||
|
__f.close()
|
||
|
group_file = __i
|
||
|
break
|
||
|
except:
|
||
|
pass
|
||
|
|
||
|
# decide what field separator we can try to use - Unix standard, with
|
||
|
# the platform's path separator as an option. No special field conversion
|
||
|
# handlers are required for the group file.
|
||
|
__field_sep = [':']
|
||
|
if os.pathsep:
|
||
|
if os.pathsep != ':':
|
||
|
__field_sep.append(os.pathsep)
|
||
|
|
||
|
# helper routine to identify which separator character is in use
|
||
|
def __get_field_sep(record):
|
||
|
fs = None
|
||
|
for c in __field_sep:
|
||
|
# there should be 3 delimiter characters (for 4 fields)
|
||
|
if record.count(c) == 3:
|
||
|
fs = c
|
||
|
break
|
||
|
if fs:
|
||
|
return fs
|
||
|
else:
|
||
|
raise KeyError, '>> group database fields not delimited <<'
|
||
|
|
||
|
# read the whole file, parsing each entry into tuple form
|
||
|
# with dictionaries to speed recall by GID or group name
|
||
|
def __read_group_file():
|
||
|
if group_file:
|
||
|
group = open(group_file, 'r')
|
||
|
else:
|
||
|
raise KeyError, '>> no group database <<'
|
||
|
gidx = {}
|
||
|
namx = {}
|
||
|
sep = None
|
||
|
while 1:
|
||
|
entry = group.readline().strip()
|
||
|
if len(entry) > 3:
|
||
|
if sep == None:
|
||
|
sep = __get_field_sep(entry)
|
||
|
fields = entry.split(sep)
|
||
|
fields[2] = int(fields[2])
|
||
|
record = tuple(fields)
|
||
|
if not gidx.has_key(fields[2]):
|
||
|
gidx[fields[2]] = record
|
||
|
if not namx.has_key(fields[0]):
|
||
|
namx[fields[0]] = record
|
||
|
elif len(entry) > 0:
|
||
|
pass # skip empty or malformed records
|
||
|
else:
|
||
|
break
|
||
|
group.close()
|
||
|
if len(gidx) == 0:
|
||
|
raise KeyError
|
||
|
return (gidx, namx)
|
||
|
|
||
|
# return the group database entry by GID
|
||
|
def getgrgid(gid):
|
||
|
g, n = __read_group_file()
|
||
|
return g[gid]
|
||
|
|
||
|
# return the group database entry by group name
|
||
|
def getgrnam(name):
|
||
|
g, n = __read_group_file()
|
||
|
return n[name]
|
||
|
|
||
|
# return all the group database entries
|
||
|
def getgrall():
|
||
|
g, n = __read_group_file()
|
||
|
return g.values()
|
||
|
|
||
|
# test harness
|
||
|
if __name__ == '__main__':
|
||
|
getgrall()
|