Windows: WindowsConsoleIO produces mojibake for strings longer than 32 KiB

[andy-ms]

BPO	37871
Nosy	@gpshead, @pfmoore, @tjguk, @zware, @eryksun, @zooba, @Andy-MS

^{Note: these values reflect the state of the issue at the time it was migrated and might not reflect the current state.}

Show more details

GitHub fields:

assignee = None
closed_at = None
created_at = <Date 2019-08-16.00:36:04.457>
labels = ['type-bug', '3.8', '3.9', 'expert-IO', '3.10', 'OS-windows']
title = 'Windows: WindowsConsoleIO produces mojibake for strings longer than 32 KiB'
updated_at = <Date 2021-03-20.23:35:16.331>
user = 'https://github.com/andy-ms'

bugs.python.org fields:

activity = <Date 2021-03-20.23:35:16.331>
actor = 'eryksun'
assignee = 'none'
closed = False
closed_date = None
closer = None
components = ['Windows', 'IO']
creation = <Date 2019-08-16.00:36:04.457>
creator = 'anhans'
dependencies = []
files = []
hgrepos = []
issue_num = 37871
keywords = []
message_count = 5.0
messages = ['349837', '349844', '349872', '389191', '389199']
nosy_count = 7.0
nosy_names = ['gregory.p.smith', 'paul.moore', 'tim.golden', 'zach.ware', 'eryksun', 'steve.dower', 'anhans']
pr_nums = []
priority = 'normal'
resolution = None
stage = 'needs patch'
status = 'open'
superseder = None
type = 'behavior'
url = 'https://bugs.python.org/issue37871'
versions = ['Python 3.8', 'Python 3.9', 'Python 3.10']

Linked PRs

• gh-82052: Don't send partial UTF-8 sequences to the Windows API #101103
• [3.11] gh-82052: Don't send partial UTF-8 sequences to the Windows API (GH-101103) #101107
• [3.10] gh-82052: Don't send partial UTF-8 sequences to the Windows API (GH-101103) #101108

[andy-ms]

# To reproduce:
# Put this text in a file a.py and run py a.py.
# Or just run: py -c "print(('é' * 40 + '\n') * 473)"
# Scroll up for a while. One of the lines will be:
# éééééééééééééééééééé��ééééééééééééééééééé
# (You can spot this because it's slightly longer than the other lines.)
# The error is consistently on line 237, column 21 (1-indexed).

# The error reproduces on Windows but not Linux. Tested in both powershell and CMD.
# (Failed to reproduce on either a real Linux machine or on Ubuntu with WSL.)
# On Windows, the error reproduces every time consistently.

# There is no error if N = 472 or 474.
N = 473
# There is no error if W = 39 or 41.
# (I tested with console windows of varying sizes, all well over 40 characters.)
W = 40
# There is no error if ch = "e" with no accent.
# There is still an error for other unicode characters like "Ö" or "ü".
ch = "é"
# There is no error without newlines.
s = (ch * W + "\n") * N
# Assert the string itself is correct.
assert all(c in (ch, "\n") for c in s)
print(s)

# There is no error if we use N separate print statements
# instead of printing a single string with N newlines.

# Similar scripts written in Groovy, JS and Ruby have no error.
# Groovy: System.out.println(("é" * 40 + "\n") * 473)
# JS: console.log(("é".repeat(40) + "\n").repeat(473))
# Ruby: puts(("é" * 40 + "\n") * 473)

[eryksun]

To be compatible with Windows 7, _io__WindowsConsoleIO_write_impl in Modules/_io/winconsoleio.c is forced to write to the console in chunks that do not exceed 32 KiB. It does so by repeatedly dividing the length to decode by 2 until the decoded buffer size is small enough.

    wlen = MultiByteToWideChar(CP_UTF8, 0, b->buf, len, NULL, 0);
    while (wlen > 32766 / sizeof(wchar_t)) {
        len /= 2;
        wlen = MultiByteToWideChar(CP_UTF8, 0, b->buf, len, NULL, 0);
    }

With ('é' * 40 + '\n') * 473, encoded as UTF-8, we have 473 82-byte lines (note that "\n" has been translated to "\r\n"). This is 38,786 bytes, which is too much for a single write, so it splits it in two.

    >>> 38786 // 2
    19393
    >>> 19393 // 82
    236
    >>> 19393 % 82
    41

This means line 237 ends up with 20 'é' characters (UTF-8 b'\xc3\xa9') and one partial character sequjence, b'\xc3'. When this buffer is passed to MultiByteToWideChar to decode from UTF-8 to UTF-16, the partial sequence gets decoded as the replacement character U+FFFD. For the next write, the remaining b'\xa9' byte also gets decoded as U+FFFD.

To avoid this, _io__WindowsConsoleIO_write_impl could decode the whole buffer in one pass, and slice that up into writes that are less than 32 KiB. Or it could ensure that its UTF-8 slices are always at character boundaries.

[zooba]

I'd rather keep encoding incrementally, and reduce the length of each attempt until the last UTF-8 character does not have its top bit set (i.e. is the final character in a multi-byte sequence).

Otherwise the people who like to print >2GB worth of data to the console will complain about the memory error :)

[gpshead]

Steve's approach makes sense and should be robust.

side note: do we need to care about Windows 7 anymore in 3.10 given that microsoft no longer supports it?

[eryksun]

side note: do we need to care about Windows 7 anymore in
3.10 given that microsoft no longer supports it?

If the fix comes in time for Python 3.8, then it needs to support Windows 7. For Python 3.9+, the 32 KiB limit can be removed.

The console documentation still includes the misleading disclaimer about "available heap". This refers to a relatively small block of shared memory (64 KiB IIRC) that's overlayed by a heap, not the default process heap. Shared memory is used by system LPC ports to efficiently pass large messages between a system server (e.g. csrss.exe, conhost.exe) and a client process. The console API used to use an LPC port, but in Windows 8.1+ it uses a driver instead, so none of the "available heap" warnings apply anymore. Microsoft should clarify the docs to stress that the warning is for Windows 7 and earlier.