1. Connection Abort before
accept Returns
server receiving an RST for an ESTABLISHED connection before accept is called.
The three-way handshake completes, the connection is established, and then the client TCP sends an RST (reset). On the server side, the connection is queued by its TCP, waiting for the server process to call accept when the RST arrives. Sometime later, the server process calls accept.
Unfortunately, what happens to the aborted connection is implementation-dependent. Berkeley-derived implementations handle the aborted connection completely within the kernel, and the server process never sees it.
POSIX specifies that the return must be ECONNABORTED ("software caused connection abort") instead.
（server端的accept函数返回ECONNABORTED）
2. Termination of Server Process
the receipt of the FIN by the client TCP only indicates that the server process has closed its end of the connection and will not be sending any more data. The receipt of the FIN does not tell the client TCP that the server process has terminated.
When the server TCP receives the data from the client, it responds with an RST since the process that had that socket open has terminated. We can verify that the RST was sent by watching the packets with tcpdump.
But if the RST arrives first, the result is an ECONNRESET ("Connection reset by peer") error return from readline.
3. SIGPIPE Signal
The rule that applies is: When a process writes to a socket that has received an RST, the SIGPIPE signal is sent to the process. The default action of this signal is to terminate the process, so the process must catch the signal to avoid being involuntarily terminated.
If the process either catches the signal and returns from the signal handler, or ignores the signal, the write operation returns EPIPE.
It is okay to write to a socket that has received a FIN, but it is an error to write to a socket that has received an RST.
Be aware, however, that if multiple sockets are in use, the delivery of the signal will not tell us which socket encountered the error. If we need to know which write caused the error, then we must either ignore the signal or return from the signal handler and handle EPIPE from the write.
4. Crashing of Server Host (crash 是指机器还在运行，但是没有任何反应的情况)
Assuming the server host crashed and there were no responses at all to the client's data segments, the error is ETIMEDOUT. But if some intermediate router determined that the server host was unreachable and responded with an ICMP "destination unreachable' message, the error is either EHOSTUNREACH or ENETUNREACH.
Although our client discovers (eventually) that the peer is down or unreachable, there are times when we want to detect this quicker than having to wait nine minutes. The solution is to place a timeout on the call to readline.
The scenario that we just discussed detects that the server host has crashed only when we send data to that host. If we want to detect the crashing of the server host even if we are not actively sending it data, another technique is required. We will discuss the SO_KEEPALIVE socket option in Section 7.5.
5. Crashing and Rebooting of Server Host
Our client is blocked in the call to readline when the RST is received, causing readline to return the error ECONNRESET.
6. Shutdown of Server Host
When a Unix system is shut down, the init process normally sends the SIGTERM signal to all processes (we can catch this signal), waits some fixed amount of time (often between 5 and 20 seconds), and then sends the SIGKILL signal (which we cannot catch) to any processes still running. This gives all running processes a short amount of time to clean up and terminate. If we do not catch SIGTERM and terminate, our server will be terminated by the SIGKILL signal.
When the process terminates, all open descriptors are closed.
7. Data Format
The problem is that the two binary integers are sent across the socket in little-endian format by the client, but interpreted as big-endian integers by the server. There are really three potential problems with this example:
1. Different implementations store binary numbers in different formats. The most common formats are big-endian and little-endian, as we described in Section 3.4.
2. Different implementations can store the same C datatype differently. For example, most 32-bit Unix systems use 32 bits for a long but 64-bit systems typically use 64 bits for the same datatype (Figure 1.17). There is no guarantee that a short, int, or long is of any certain size.
3. Different implementations pack structures differently, depending on the number of bits used for the various datatypes and the alignment restrictions of the machine. Therefore, it is never wise to send binary structures across a socket.
There are two common solutions to this data format problem:
1. Pass all numeric data as text strings. This assumes that both hosts have the same character set.
2. Explicitly define the binary formats of the supported datatypes (number of bits, big- or little-endian) and pass all data between the client and server in this format. RPC packages normally use this technique.
阅读(1293) | 评论(0) | 转发(0) |
