I'm trying to use the email.* functions to craft HTTP POST data for file upload. Trying something like

filedata = open("data", "rb").read()
postdata = MIMEMultipart()
fileattachment = MIMEApplication(filedata, _encoder=email.encoders.encode_noop)
postdata.attach(fileattachment)
fp = BytesIO()
g = BytesGenerator(fp)
g.flatten(postdata, unixfrom=False)

fails with

Traceback (most recent call last):
  File "./minetest.py", line 30, in <module>
    g.flatten(postdata, unixfrom=False)
  File "/usr/lib/python3.2/email/generator.py", line 91, in flatten
    self._write(msg)
  File "/usr/lib/python3.2/email/generator.py", line 137, in _write
    self._dispatch(msg)
  File "/usr/lib/python3.2/email/generator.py", line 163, in _dispatch
    meth(msg)
  File "/usr/lib/python3.2/email/generator.py", line 224, in _handle_multipart
    g.flatten(part, unixfrom=False, linesep=self._NL)
  File "/usr/lib/python3.2/email/generator.py", line 91, in flatten
    self._write(msg)
  File "/usr/lib/python3.2/email/generator.py", line 137, in _write
    self._dispatch(msg)
  File "/usr/lib/python3.2/email/generator.py", line 163, in _dispatch
    meth(msg)
  File "/usr/lib/python3.2/email/generator.py", line 192, in _handle_text
    raise TypeError('string payload expected: %s' % type(payload))
TypeError: string payload expected: <class 'bytes'>

This is because BytesGenerator._handle_text() expects str payload in which byte values that are non-printable in ASCII have been replaced by surrogates. The example above creates a bytes payload, however, for which super(BytesGenerator,self)._handle_text(msg) = Generator._handle_text(msg) throws the exception.

Note that using any email.encoders other than encode_noop does not really fit the HTTP POST bill, as those define a Content-Transfer-Encoding which HTTP does not know.

It would seem better to me to let BytesGenerator accept a bytes payload and just copy that to the output, rather than making the application encode the bytes as a string, hopefully in a way that s.encode('ascii', 'surrogateescape') can invert.

E.g., a workaround class I use now does

class FixedBytesGenerator(BytesGenerator):
    def _handle_bytes(self, msg):  
        payload = msg.get_payload()
        if payload is None:
            return
        if isinstance(payload, bytes):
            self._fp.write(payload)   
        elif isinstance(payload, str):
            super(FixedBytesGenerator,self)._handle_text(msg)
        else:
            # Payload is neither bytes not string - this can't be right
            raise TypeError('bytes or str payload expected: %s' % type(payload))
    _writeBody = _handle_bytes

Yes, the way BytesGenerator works is basically a hack to get the email package itself working. Use cases outside the email package were not really considered in the (short) timeframe during which it was implemented.

The longer term plan calls for redoing the way payloads are handled to generalize the whole process. I'd like to see this happen for 3.4, but I'm not sure I'm going to have the time to finish the work (I'm hopeful that I will, though).

In the meantime, while your suggestion is a good one, I'm ambivalent about applying it as a bug fix. It is on the border between a fix and a feature, since the email package in 3.x hasn't ever supported bytes payloads, only encoded payloads.

Hmm. Let me rephrase that. *Internally* it doesn't support bytes payloads, it "encodes" bytes payloads as surrogateescaped ascii, as you have oserved. Which is why this is on the borderline, and could possibly be considered a bug fix, because from an external point of view it does support parsing and generating 8bit payloads.

I need to give it some thought, and perhaps others will weigh in with opinions.

Looking at the documentation, it is clear that (a) what you are trying to do is documented as being correct and (b) it worked in Python2, making this a regression.

I've attached a patch to fix this, which also probably fixes some bugs with BytesGenerator handing of non-text CTE 8bit parts created by BytesParser, but I haven't added tests to confirm that.

Updated patch after review by Ezio and Serhiy.

While related, that is a different bug, so I'd rather open a new issue for it.

New changeset 30f92600df9d by R David Murray in branch '2.7':
bpo-16564: test to confirm behavior that regressed in python3.
http://hg.python.org/cpython/rev/30f92600df9d

New changeset a1a04f76d08c by R David Murray in branch '3.2':
bpo-16564: Fix regression in use of encoders.encode_noop with binary data.
http://hg.python.org/cpython/rev/a1a04f76d08c

New changeset 2b1edefc1e99 by R David Murray in branch '3.3':
Merge: bpo-16564: Fix regression in use of encoders.encode_noop with binary data.
http://hg.python.org/cpython/rev/2b1edefc1e99

New changeset 5a0478bd5f11 by R David Murray in branch 'default':
Merge: bpo-16564: Fix regression in use of encoders.encode_noop with binary data.
http://hg.python.org/cpython/rev/5a0478bd5f11

I've opened bpo-17171 for the similar encode7or8bit problem.

New changeset 64e004737837 by R David Murray in branch '3.3':
bpo-18324: set_payload now correctly handles binary input.
http://hg.python.org/cpython/rev/64e004737837

It seems to me that this issue is not fixed correctly yet. I've tried Python 3.3.2:
~/build/Python-3.3.2$ ./python --version
Python 3.3.2

When modifying the test case in Lib/test/test_email/test_email.py like this:

--- Lib/test/test_email/test_email.py	2013-05-15 18:32:55.000000000 +0200
+++ Lib/test/test_email/test_email_mine.py	2013-09-10 14:22:08.160089440 +0200
@@ -1461,17 +1461,17 @@
         # Issue 16564: This does not produce an RFC valid message, since to be
         # valid it should have a CTE of binary.  But the below works in
         # Python2, and is documented as working this way.
-        bytesdata = b'\xfa\xfb\xfc\xfd\xfe\xff'
+        bytesdata = b'\x0b\xfa\xfb\xfc\xfd\xfe\xff'
         msg = MIMEApplication(bytesdata, _encoder=encoders.encode_noop)
         # Treated as a string, this will be invalid code points.
-        self.assertEqual(msg.get_payload(), '\uFFFD' * len(bytesdata))
+        # self.assertEqual(msg.get_payload(), '\uFFFD' * len(bytesdata))
         self.assertEqual(msg.get_payload(decode=True), bytesdata)
         s = BytesIO()
         g = BytesGenerator(s)
         g.flatten(msg)
         wireform = s.getvalue()
         msg2 = email.message_from_bytes(wireform)
-        self.assertEqual(msg.get_payload(), '\uFFFD' * len(bytesdata))
+        # self.assertEqual(msg.get_payload(), '\uFFFD' * len(bytesdata))
         self.assertEqual(msg2.get_payload(decode=True), bytesdata)

then running:

./python ./Tools/scripts/run_tests.py test_email

results in:

======================================================================
FAIL: test_binary_body_with_encode_noop (test_email_mine.TestMIMEApplication)
----------------------------------------------------------------------

Traceback (most recent call last):
  File "/localdisk/kruppaal/build/Python-3.3.2/Lib/test/test_email/test_email_mine.py", line 1475, in test_binary_body_with_encode_noop
    self.assertEqual(msg2.get_payload(decode=True), bytesdata)
AssertionError: b'\x0b\n\xfa\xfb\xfc\xfd\xfe\xff' != b'\x0b\xfa\xfb\xfc\xfd\xfe\xff'

The '\x0b' byte is incorrectly translated to '\x0b\n', i.e., a New Line character is inserted.

Encoding the bytes array:
bytes(range(256))

results output data (MIME Header stripped):

0000000: 0001 0203 0405 0607 0809 0a0b 0a0c 0a0a ................
0000010: 0e0f 1011 1213 1415 1617 1819 1a1b 1c0a ................
0000020: 1d0a 1e0a 1f20 2122 2324 2526 2728 292a ..... !"#$%&'()*
0000030: 2b2c 2d2e 2f30 3132 3334 3536 3738 393a +,-./0123456789:
0000040: 3b3c 3d3e 3f40 4142 4344 4546 4748 494a ;<=>?@abcdefghij
0000050: 4b4c 4d4e 4f50 5152 5354 5556 5758 595a KLMNOPQRSTUVWXYZ
0000060: 5b5c 5d5e 5f60 6162 6364 6566 6768 696a [\]^_`abcdefghij
0000070: 6b6c 6d6e 6f70 7172 7374 7576 7778 797a klmnopqrstuvwxyz
0000080: 7b7c 7d7e 7f80 8182 8384 8586 8788 898a {|}~............
0000090: 8b8c 8d8e 8f90 9192 9394 9596 9798 999a ................
00000a0: 9b9c 9d9e 9fa0 a1a2 a3a4 a5a6 a7a8 a9aa ................
00000b0: abac adae afb0 b1b2 b3b4 b5b6 b7b8 b9ba ................
00000c0: bbbc bdbe bfc0 c1c2 c3c4 c5c6 c7c8 c9ca ................
00000d0: cbcc cdce cfd0 d1d2 d3d4 d5d6 d7d8 d9da ................
00000e0: dbdc ddde dfe0 e1e2 e3e4 e5e6 e7e8 e9ea ................
00000f0: ebec edee eff0 f1f2 f3f4 f5f6 f7f8 f9fa ................
0000100: fbfc fdfe ff .....

That is, a '\n' is inserted after '\x0b', '\x1c', '\x1d', and '\x1e',
and '\x0d' is replaced by '\n\n'.

That's a different bug, and is probably due to the fact that \x0b is considered a line-ending character by the 'splitlines' method.

Could you please open a new issue for this? It could be that this can't be fixed in Python3 until support for the 'binary' CTE is added.

Opened bpo-19003.