Inefficient ssl.SSLWantReadError exception slows down very common use-case
Nobody has claimed this yet.
- Dominant language
- Python
- Stars
- 77.2k
- Forks
- 35.9k
- PR merge metrics
- PR metrics pending
Description
Bug report
Bug description:
Event loops like uvloop, asyncio use nonblocking ssl. They typically
- read data from the socket when epoll returns that it is ready
- push data to the incoming MemoryBIO
- read from SSLObject until SSLWantReadError is thrown
- pass read data to the application protocol
when peers are exchanging relatively small messages, SSLObject.read is typically called 2 times . First call returns data, second - throws SSLWantReadError
perf shows that the second call is almost as expensive as the first call because of time spent on constructing new exception object.
Is it possible to optimize exception object creation for the second call?
I tried to avoid the second call by analyzing MemoryBIO.pending and SSLObject.pending values but they can't always reliably tell that we have to wait for more data.
For example, it is possible that incoming MemoryBIO.pending > 0, SSLObject.pending == 0. We call SSLObject.read and it throws because incoming MemoryBIO doesn't have the full ssl frame yet.
Example echo client that replicates internal logic in asyncio/uvloop:
import socket
import ssl
import select
from typing import Optional
ssl_context = ssl.create_default_context()
ssl_context.check_hostname = False
ssl_context.verify_mode = ssl.CERT_NONE
ep = select.epoll(2)
incoming = ssl.MemoryBIO()
outgoing = ssl.MemoryBIO()
sock: Optional[socket.socket] = None
ssl_sock: Optional[ssl.SSLObject] = None
def wait_data():
ep.poll()
try:
while True:
chunk = sock.recv(1024)
incoming.write(chunk)
except BlockingIOError:
pass
def wait_data_until_ssl_read_succeed():
data = bytearray()
try:
wait_data()
# while ssl_sock.pending() > 0 or incoming.pending > 0:
while True:
data += ssl_sock.read()
except ssl.SSLWantReadError:
pass
return data
with socket.socket(socket.AF_INET, socket.SOCK_STREAM) as sock:
ep.register(sock.fileno(), select.EPOLLIN)
ssl_sock = ssl_context.wrap_bio(incoming, outgoing, server_hostname='localhost')
sock.connect(('127.0.0.1', 25000))
sock.setblocking(False)
handshake_complete = False
message_sent = False
msg = b"a" * 256
# do handshake
while True:
try:
ssl_sock.do_handshake()
break
except ssl.SSLWantReadError as ex:
if outgoing.pending > 0:
chunk = outgoing.read(outgoing.pending)
sock.send(chunk)
wait_data()
# send message and wait for reply
while True:
ssl_sock.write(msg)
chunk = outgoing.read(outgoing.pending)
sock.send(chunk)
data = wait_data_until_ssl_read_succeed()
# print(data)
Perf output:
17.41% 0.25% 43 python _ssl.cpython-314-x86_64-linux-gnu.so [.] _ssl__SSLSocket_read
|
--17.16%--_ssl__SSLSocket_read
|
|--8.09%--SSL_read_ex
| |
| --7.74%--0x7b1fbef883f9
| |
| |--4.83%--0x7b1fbefadc22
| | |
| | |--0.93%--0x7b1fbefa6919
| | | |
| | | --0.90%--EVP_DecryptUpdate
| | | |
| | | --0.90%--0x7b1fbec90c8b
| | | |
| | | --0.87%--0x7b1fbec90b45
| | |
| | |--0.79%--0x7b1fbefa64a5
| | | |
| | | --0.61%--EVP_CIPHER_CTX_get_iv_length
| | |
| | |--0.72%--0x7b1fbefa6721
| | | |
| | | --0.60%--EVP_CipherInit_ex
| | |
| | |--0.57%--0x7b1fbefa6750
| | |
| | --0.52%--0x7b1fbefa68f0
| |
| |--1.13%--0x7b1fbefadf94
| | |
| | --0.52%--0x7b1fbefac3c7
| |
| --0.68%--0x7b1fbefad750
|
|--6.21%--PySSL_SetError.constprop.0
| |
| --5.17%--fill_and_set_sslerror
| |
| |--2.83%--PyUnicode_FromFormat
| | |
| | --2.54%--unicode_from_format
| | |
| | --1.23%--__sprintf_chk
| | __vsprintf_internal
| | |
| | --0.99%--__vfprintf_internal
| |
| --1.01%--PyObject_SetAttr
| |
| --0.98%--PyObject_GenericSetAttr
| |
| --0.52%--_PyObjectDict_SetItem
|
--0.60%--SSL_get_error
To reproduce you would need some ssl echo server running on localhost 25000 port. After you have started it, run echo client code under perf.
$ perf record -F 999 -g --call-graph lbr --user-callchains -- python echo_client.py
$ perf report -G -n --stdio
Let it work for 15 seconds and then press Ctrl-C
CPython versions tested on:
CPython main branch
Operating systems tested on:
Linux
Linked PRs
- gh-128391
Contributor guide
First steps
- Read the whole issue, then the project's contributing guide.
- Comment on the issue to say you are picking it up — it saves two people doing the same work.
- Fork the repository and make your change on a branch.
- Open a pull request that references the issue number.
Research direction
Start at SSLObject.read and the _ssl__SSLSocket_read path shown in the perf output, especially PySSL_SetError and fill_and_set_sslerror. Reproduce with the provided echo client and a localhost SSL echo server, then run the supplied perf commands. Done means reducing SSLWantReadError construction overhead without changing nonblocking SSL behavior.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- python
- Domain
- networking, performance
- Issue type
- Bug
- Difficulty
- 4/5
- Estimated time
- 3-5 days
- Activity status
- Stale
- Clarity
- Mostly clear
- Newbie friendliness
- 35/100