@vsbogd,Thank you for your reply. But I find another problem.

I derive a subclass from sockerserver.StreamRequestHandler, and set the rbufsize to 0(As CGIHTTPRequestHandler do), like this demo below:

    testserver.py:
                import socketserver
                class TestRequestHandler(socketserver.StreamRequestHandler):
                    rbufsize = 0  ###simulate CGIHTTPRequestHandler
                    def handle(self):
                        while True:
                            data = self.rfile.read(65536*1024) ###client should send 65536*1024 bytes.
                            print(len(data))
                            if len(data) == 0:
                                print("Connection closed.")
                                break
                s = socketserver.TCPServer(("0.0.0.0", 8001), TestRequestHandler)
                s.serve_forever()

    testclient.py:
                import socket
                data = bytearray(65536*1024)
                for i in range(65536*1024):
                    data[i] = 64        #Whatever you set.
                c = socket.socket(socket.AF_INET, socket.SOCK_STREAM)
                c.connect(("127.0.0.1", 8001))
                c.send(data)

The testserver.py can get the whole 65536*1024 data in every "data = self.rfile.read(65536*1024)" line. The normal output of testserver.py is:

    testserver.py output:
                67108864
                0
                Connection closed.
                67108864
                0
                Connection closed.

In other words, this problem of "rfile.read(nbytes)" cannot be reproduce in this demo.

I dont know why, it seems this is not only the problem of the "rfile.read(nbytes)". I guess the CGIHTTPRequestHandler actually do something that make the "rfile.read(nbytes)" perform weirdly. However, I fail to find such a line in the code.