File Descriptor Management

File descriptors underlie all input/output mechanisms offered by the system. They are used to implementation the FILE *-based functions found in <stdio.h>, and all the file and network communication facilities provided by the Python and Java environments are eventually implemented in them.

File descriptors are small, non-negative integers in userspace, and are backed on the kernel side with complicated data structures which can sometimes grow very large.

Closing Descriptors

If a descriptor is no longer used by a program and is not closed explicitly, its number cannot be reused (which is problematic in itself, see Dealing with the select Limit), and the kernel resources are not freed. Therefore, it is important to close all descriptors at the earliest point in time possible, but not earlier.

Error Handling during Descriptor Close

The close system call is always successful in the sense that the passed file descriptor is never valid after the function has been called. However, close still can return an error, for example if there was a file system failure. But this error is not very useful because the absence of an error does not mean that all caches have been emptied and previous writes have been made durable. Programs which need such guarantees must open files with O_SYNC or use fsync or fdatasync, and may also have to fsync the directory containing the file.

Closing Descriptors and Race Conditions

Unlike process IDs, which are recycle only gradually, the kernel always allocates the lowest unused file descriptor when a new descriptor is created. This means that in a multi-threaded program which constantly opens and closes file descriptors, descriptors are reused very quickly. Unless descriptor closing and other operations on the same file descriptor are synchronized (typically, using a mutex), there will be race conditions and I/O operations will be applied to the wrong file descriptor.

Sometimes, it is necessary to close a file descriptor concurrently, while another thread might be about to use it in a system call. In order to support this, a program needs to create a single special file descriptor, one on which all I/O operations fail. One way to achieve this is to use socketpair, close one of the descriptors, and call shutdown(fd, SHUTRDWR) on the other.

When a descriptor is closed concurrently, the program does not call close on the descriptor. Instead it program uses dup2 to replace the descriptor to be closed with the dummy descriptor created earlier. This way, the kernel will not reuse the descriptor, but it will carry out all other steps associated with calling a descriptor (for instance, if the descriptor refers to a stream socket, the peer will be notified).

This is just a sketch, and many details are missing. Additional data structures are needed to determine when it is safe to really close the descriptor, and proper locking is required for that.

Lingering State after Close

By default, closing a stream socket returns immediately, and the kernel will try to send the data in the background. This means that it is impossible to implement accurate accounting of network-related resource utilization from userspace.

The SO_LINGER socket option alters the behavior of close, so that it will return only after the lingering data has been processed, either by sending it to the peer successfully, or by discarding it after the configured timeout. However, there is no interface which could perform this operation in the background, so a separate userspace thread is needed for each close call, causing scalability issues.

Currently, there is no application-level countermeasure which applies universally. Mitigation is possible with iptables (the connlimit match type in particular) and specialized filtering devices for denial-of-service network traffic.

These problems are not related to the TIME_WAIT state commonly seen in netstat output. The kernel automatically expires such sockets if necessary.

Preventing File Descriptor Leaks to Child Processes

Child processes created with fork share the initial set of file descriptors with their parent process. By default, file descriptors are also preserved if a new process image is created with execve (or any of the other functions such as system or posix_spawn).

Usually, this behavior is not desirable. There are two ways to turn it off, that is, to prevent new process images from inheriting the file descriptors in the parent process:

  • Set the close-on-exec flag on all newly created file descriptors. Traditionally, this flag is controlled by the FD_CLOEXEC flag, using F_GETFD and F_SETFD operations of the fcntl function.

    However, in a multi-threaded process, there is a race condition: a subprocess could have been created between the time the descriptor was created and the FD_CLOEXEC was set. Therefore, many system calls which create descriptors (such as open and openat) now accept the O_CLOEXEC flag (SOCK_CLOEXEC for socket and socketpair), which cause the FD_CLOEXEC flag to be set for the file descriptor in an atomic fashion. In addition, a few new systems calls were introduced, such as pipe2 and dup3.

    The downside of this approach is that every descriptor needs to receive special treatment at the time of creation, otherwise it is not completely effective.

  • After calling fork, but before creating a new process image with execve, all file descriptors which the child process will not need are closed.

    Traditionally, this was implemented as a loop over file descriptors ranging from 3 to 255 and later 1023. But this is only an approximation because it is possible to create file descriptors outside this range easily (see Dealing with the select Limit). Another approach reads /proc/self/fd and closes the unexpected descriptors listed there, but this approach is much slower.

At present, environments which care about file descriptor leakage implement the second approach. OpenJDK 6 and 7 are among them.

Dealing with the select Limit

By default, a user is allowed to open only 1024 files in a single process, but the system administrator can easily change this limit (which is necessary for busy network servers). However, there is another restriction which is more difficult to overcome.

The select function only supports a maximum of FD_SETSIZE file descriptors (that is, the maximum permitted value for a file descriptor is FD_SETSIZE - 1, usually 1023.) If a process opens many files, descriptors may exceed such limits. It is impossible to query such descriptors using select.

If a library which creates many file descriptors is used in the same process as a library which uses select, at least one of them needs to be changed. Calls to select can be replaced with calls to poll or another event handling mechanism. Replacing the select function is the recommended approach.

Alternatively, the library with high descriptor usage can relocate descriptors above the FD_SETSIZE limit using the following procedure.

  • Create the file descriptor fd as usual, preferably with the O_CLOEXEC flag.

  • Before doing anything else with the descriptor fd, invoke:

	  int newfd = fcntl(fd, F_DUPFD_CLOEXEC, (long)FD_SETSIZE);

  • Check that newfd result is non-negative, otherwise close fd and report an error, and return.

  • Close fd and continue to use newfd.

The new descriptor has been allocated above the FD_SETSIZE. Even though this algorithm is racy in the sense that the FD_SETSIZE first descriptors could fill up, a very high degree of physical parallelism is required before this becomes a problem.