Linux and glibc API changes

As The Linux Programming Interface went to press in August 2010, it was up to date with the then current versions of the Linux kernel (2.6.35) and glibc (2.12). Because the developers of both the Linux kernel and glibc are committed to maintaining ABI compatibility, virtually all of the details provided in TLPI should remain accurate in the future. However, new features are added to the kernel and glibc with each release. As each new release of the Linux kernel and glibc occurs, this page will attempt to note new interface features that are relevant to the subject area of the book.

• Linux kernel
• 3.10
• 3.9
• 3.8
• 3.7
• 3.6
• 3.5
• 3.4
• 3.3
• 3.2
• 3.1
• 3.0
• 2.6.39
• 2.6.38
• 2.6.37
• 2.6.36
• glibc
  • 2.18
  • 2.17
  • 2.16
  • 2.15
  • 2.14
  • 2.13

Linux kernel API changes

Linux 3.10 (Not yet released)

API changes include the following:

• POSIX clocks and timers now support a new clock, CLOCK_TAI, that measures against International Atomic Time.
• …

See also: LWN articles on the kernel 3.10 merge window (1, 2, 3) and the Kernel Newbies kernel 3.10 summary.

Linux 3.9 (28 Apr 2013)

API changes include the following:

• A new SO_REUSEPORT socket option allows multiple sockets to be bound to a UDP or TCP port. The option improves performance in some network server designs. More information can be found in my LWN.net article The SO_REUSEPORT socket option.
• A new /proc/sys/kernel/sched_rr_timeslice_ms file can be used to view and set the SCHED_RR (realtime scheduling round-robin) quantum as a millisecond value. The default value is 100. Writing 0 to this file resets the quantum to the default value.

See also: LWN articles on the kernel 3.9 merge window (1, 2, 3) and the Kernel Newbies kernel 3.9 summary.

Linux 3.8 (18 Feb 2013)

API changes include the following:

• The ptrace() system call supports a new flag, PTRACE_O_EXITKILL. If a tracing process sets this flag, a SIGKILL signal will be sent to every traced process if the tracing process exits.
• On systems that provide multiple huge page sizes, shmget() and mmap() can now select the desired page size for an allocation. For more information, see my LWN.net article Supporting variable-sized huge pages.
• The user namespaces implementation has been completed, allowing unprivileged process to create and work with user namespaces via clone(CLONE_NEWUSER), unshare(CLONE_NEWUSER), and setns(). See also my series of LWN.net articles on namespaces starting here (you can find a directory of the follow-up articles in the series at the end of that article).
• The setns() and unshare() system calls add support for PID, mount, and user namespaces.
• A new finit_module() system call is added. This system call is like init_module(), but loads the module from an open file descriptor. In addition, the new system call has a flags argument that can be used to modify the behavior of the system call. For more information, see this LWN article and the init_module(2) manual page.
• The ext4 and tmpfs file systems add support for the lseek() SEEK_HOLE and SEEK_DATA operations.

See also: LWN articles on the kernel 3.8 merge window (1, 2) and the Kernel Newbies kernel 3.8 summary.

Linux 3.7 (10 Dec 2012)

API changes include the following:

• The server-side implementation of the TCP Fast Open feature was merged. This complements the implementation of the client-side functionality that was merged in 3.6. To enable server-side (i.e., passive) TCP Fast Open, a TCP server must use setsockopt() to set the TCP_FASTOPEN For more information, see my LWN.net article, TCP Fast Open: expediting web services.
• The epoll_ctl() adds a new flag, EPOLL_CTL_DISABLE, that allows multithreaded applications to safely disable monitoring of a file descriptor.
• The Btrfs file system adds support for hole punching (the fallocate(2) FALLOC_FL_PUNCH_HOLE operation added in Linux 2.6.38).

See also: LWN articles on the kernel 3.7 merge window (1, 2, 3) and the Kernel Newbies kernel 3.7 summary.

Linux 3.6 (30 Sep 2012)

API changes include the following:

• The client-side implementation of the TCP Fast Open feature was merged. This implements a new flag, MSG_FASTOPEN, used with either sendto() or sendmsg() to initiate a TCP fast open. The new /proc/sys/net/ipv4/tcp_fastopen file can be set to enable or disable client (and server) TCP Fast Open functionality. For more information, see my LWN.net article, TCP Fast Open: expediting web services.
• Some restrictions on the creation of hard and soft links were added, in order to improve security. For more information, see this LWN.net article and the documentation of /proc/sys/fs/protected_hardlinks and /proc/sys/fs/protected_symlinks in the proc(5) manual page.
• The fcntl() system call adds support for a new command, F_GETOWNER_UIDS, that can be used to retrieve the real and effective user IDs associated with a previous call to F_SETOWNER. (Those UIDs determine the rules for sending a signal to another process for signal-driven I/O.) The third argument of the call is of type uid_t *, and should point to a two-element array that stores the real user ID and effective user ID. This feature is intended for use by the checkpoint/restore facility and is only provided if the kernel was configured with the CONFIG_CHECKPOINT_RESTORE option.

See also: LWN articles on the kernel 3.6 merge window (1, 2, 3) and the Kernel Newbies kernel 3.6 summary.

Linux 3.5 (21 Jul 2012)

API changes include the following:

• The seccomp filters mechanism was added. This feature is designed to allow security conscious applications to limit the set of system calls that they can make. For further information, see this LWN.net article and the kernel source file Documentation/prctl/seccomp_filter.txt. Among other things, the change adds a new a new PTRACE_O_TRACESECCOMP flag to ptrace(2).
• A new kcmp() system call used to determine whether various kernel objects are shared between tasks. This is useful for the checkpoint-restore facility. Some information can be found in this LWN.net article and in the kcmp(2) manual page.
• Some additional PR_SET_MM_* flags for use with the PR_SET_MM prctl() operation added in Linux 3.3.
• A new epoll EPOLLWAKEUP flag prevents system suspend while epoll events are ready. Use of this flag requires that the caller have the newly added CAP_BLOCK_SUSPEND capability.
• The tmpfs file system adds support for hole punching (the fallocate(2) FALLOC_FL_PUNCH_HOLE operation added in Linux 2.6.38) and the lseek() SEEK_HOLE and SEEK_DATA operations.
• The new prctl() PR_SET_NO_NEW_PRIVS operation prevents execve() from granting privileges. For example, a process will not be able to execute a set-user-ID binary to change its UID or GID if this flag is set. The same is true for file capabilities. A corresponding PR_GET_NO_NEW_PRIVS operation can be used to retrieve the state of this attribute for the caller. Documentation can found in the prctl(2) manual page.
• The new prctl() PR_GET_TID_ADDRESS operation allows the caller to retrieve its clear_child_tid address (see set_tid_address(2)). Documentation can found in the prctl(2) manual page.
• If the kernel was configured with CONFIG_CHECKPOINT_RESTORE, then a new /proc/PID/children file lists the children of a process.

See also: LWN articles on the kernel 3.5 merge window (1, 2, 3) and the Kernel Newbies kernel 3.5 summary.

Linux 3.4 (20 May 2012)

API changes include the following:

• The PR_SET_CHILD_SUBREAPER prctl() operation allows a "service manager" process to mark itself as a sort of 'sub-init', able to stay as the parent for all orphaned processes created by the started services. All SIGCHLD signals will be delivered to the service manager. There is a corresponding PR_GET_CHILD_SUBREAPER prctl() operation. Documentation can be found in the prctl(2) manual page. Planned users of this feature include D-Bus and systemd.
• The madvise() MADV_DONTDUMP operation can be used to specify that an address range should be excluded from core dumps. The MADV_DODUMP operation reverses the effect of MADV_DONTDUMP. Documentation can be found in the madvise(2) manual page.
• The setsockopt() SO_PEEK_OFF allows controlling the offset for peeking at data queued in a socket. (Currently supported for UNIX domain sockets only.) Documentation can found in the socket(7) manual page
• A new prctl() operation, PR_SET_PTRACER, is used with the Yama Linux Security Module to control which processes can ptrace() the calling process. Documentation can found in the prctl(2) manual page and in the kernel source file Documentation/security/Yama.txt.

See also: LWN articles on the kernel 3.4 merge window (1, 2, 3) and the Kernel Newbies kernel 3.4 summary.

Linux 3.3 (18 Mar 2012)

API changes include the following:

• A new prctl() operation, PR_SET_MM, intended for use by the checkpoint/restart facility, allows text, data, and heap sizes to be set to the values in effect at checkpoint time when a process is restored. The caller must have the CAP_SYS_RESOURCE capability. This operation is only supported if the kernel is configured with the CONFIG_CHECKPOINT_RESTORE option. Documentation can found in the prctl(2) manual page.
• Two changes related to the /proc filesystem, as reported on LWN.net:
  • A new /proc/PID/map_files directory contains symbolic links describing the file mappings of the process identified by PID; documentation can be found in the proc(5) manual page.
  • New mount options for the /proc file system can be used to control the visibility of the /proc/PID directories.

See also: LWN articles on the kernel 3.3 merge window (1, 2) and the Kernel Newbies kernel 3.3 summary.

Linux 3.2 (4 Jan 2012)

API changes include the following:

• The process_vm_readv() and process_vm_writev() functions, which provide a technique for fast message passing. Some information can be found on LWN.net here and here (describes an early version of the API), and in the process_vm_readv(2) manual page.
• Files under /proc/sys are now pollable, meaning that applications can use poll(), select(), and epoll to check for changes to sysctl parameters.
• A new /proc/sys/kernel/cap_last_cap file exposes the numerical value of the highest capability supported by the running kernel; this can be used to determine the highest bit that may be set in a capability set.

See also: LWN articles on the kernel 3.2 merge window (1, 2) and the Kernel Newbies kernel 3.2 summary.

Linux 3.1 (24 Oct 2011)

API changes include the following:

• Three new operations are added for the ptrace() system call: PTRACE_SEIZE, PTRACE_INTERRUPT, and PTRACE_LISTEN. Documentation can be found in the ptrace(2) manual page. Some further information can be found here.
• Two new flags for the lseek() system call, SEEK_HOLE and SEEK_DATA, provide the ability to search for holes in sparsely allocated files. Some further information can be found in this LWN.net article and the lseek(2) manual page.
• A new /proc/sys/kernel/shm_rmid_forced file can be used to control the handling of System V shared memory segments that have no attached process. The default value in this file is 0, which provides the traditional behavior: unattached segments remain in existence and can be reattached at a later point in time by another process. If the value in shm_rmid_forced is 1, then the effect is as though an IPC_RMID operation is performed on all shared memory segments that currently exist and that are created in the future. This means that those segments will be destroyed as soon as the last process detaches from them. This can be useful to ensure that shared memory segments are counted against the resource usage and limits of at least one process, but it is nonstandard and has the potential to break applications that depend on the traditional behavior. Further details can be found in the proc(5) manual page.

See also: LWN articles on the kernel 3.1 merge window (1, 2) and the Kernel Newbies kernel 3.1 summary.

Linux 3.0 (?) (21 Jul 2011)

API changes include the following:

• A new setns() system call allows its caller to join the namespace specified by its two arguments—a namespace type (one of a subset of the CLONE_* constants given to clone(2)) and a file descriptor referring to one of the files in a /proc/PID/ns directory. Some further info here, and in the setns(2) manual page contributed by Eric Biederman.
• A new sendmmsg() system call provides multiple message sending facilities (the analog of the recvmmsg(2) system call added in Linux 2.6.33). For more information, see the sendmmsg(2) manual page.
• The timerfd_settime() system call adds a TFD_TIMER_CANCEL_ON_SET flag. If this flag is set for a CLOCK_REALTIME absolute (TFD_TIMER_ABSTIME) timer, then the timer is expired if the clock is reset.
• Two new POSIX clocks: CLOCK_BOOTTIME_ALARM and CLOCK_REALTIME_ALARM. According to the commit message, these clocks behave identically to CLOCK_REALTIME and CLOCK_BOOTTIME, but the _ALARM suffixed clocks will wake the system if it is suspended. Some further details can be found here.
• A new CAP_WAKE_ALARM capability governs the use of the CLOCK_BOOTTIME_ALARM and CLOCK_REALTIME_ALARM clocks.
• The /proc/sys/kernel/core_pattern file adds a new specifier, %E. This specifier is replaced by the pathname of the executable, with slashes replaced by exclamation marks (so that the basename of the resulting core dump filename does not contain slashes). Documentation can be found in the core(5) manual page.

See also: LWN articles on the kernel 3.0 merge window (1, 2) and the Kernel Newbies kernel 3.0 summary.

Linux 2.6.39 (18 May 2011)

API changes include the following:

• New name_to_handle_at() and open_by_handle_at() system calls. These system calls provide functionality that is useful for file-system servers that run in user space. Some details here and here.
• A new O_PATH flag is added for open(2). Some details here. O_PATH descriptors can be obtained for symbolic links, and can be passed via SCM_RIGHTS datagrams.
• A new AT_EMPTY_PATH flag allows empty relative pathnames for linkat(2), fchownat(2), fstatat(2), and name_to_handle_at(), in which case the calls operate on their directory file descriptor argument. In addition, an empty pathname can now be supplied to readlinkat(2), to produce the same behavior for that call.
• A new clock_adjtime() system call, analogous to adjtimex(2), permits adjustments to POSIX clocks.
• A new syncfs() system call, which is similar to sync(2), but flushes only the file system containing the file referred to by its file-descriptor argument. Details in the syncfs(2) manual page.
• A new POSIX clock, CLOCK_BOOTTIME, is identical to CLOCK_MONOTONIC, but includes time that the system has been suspended. This clock is intended for applications that want a monotonically increasing clock and also want to be aware of time the system has been suspended. Some background here.
• A thread operating under the SCHED_IDLE policy is now allowed to upgrade itself to the SCHED_BATCH or SCHED_OTHER policy if its nice value falls within the range permitted by its RLIMIT_NICE resource limit.

See also: LWN articles on the kernel 2.6.39 merge window (1, 2, 3) and the Kernel Newbies kernel 2.6.39 summary.

Linux 2.6.38 (14 Mar 2011)

API changes include the following:

• A new AT_NO_AUTOMOUNT flag for fstatat(2), which can be used to suppress automounting of the terminal component of the pathname argument. Further information can be found in the fstatat(2) manual page.
• A new CAP_SYSLOG capability, used (instead of CAP_SYS_ADMIN) to govern privileged syslog(2) operations. Documentation can be found in the manual pages.
• A new FALLOC_FL_PUNCH_HOLE operation for fallocate(2). This operation creates a hole (see page 83 of TLPI) in the file in the byte range indicated by the offset and len arguments. (The file data in the specified range is lost.) File system support is required for the FALLOC_FL_PUNCH_HOLE operation. Among the file systems that support FALLOC_FL_PUNCH_HOLE are XFS and (since Linux 3.0) ext4. Btrfs is capable of supporting the operation, and support is likely to be added in the future. As currently implemented, FALLOC_FL_PUNCH_HOLE must be specified with FALLOC_FL_KEEP_SIZE, which means that the size of a file can't change, even if a hole is punched at the end of the file. Further information can be found in the fallocate(2) manual page.
• New MADV_HUGEPAGE and MADV_NOHUGEPAGE flags for madvise(2). These flags enable and disable an attribute on the memory region that indicates that it is important that the region be backed by huge pages, when this is possible. Further information on this feature can be found in the Kernel source file Documentation/vm/transhuge.txt as well as here, here, and in the madvise(2) manual page.
• The new /proc/sys/kernel/kptr_restrict file can be used to prevent exposure of kernel pointers via /proc files and other interfaces. (This affects how pointers are printed when using the new %pK specifier for the kernel-internal printf() function.) See the proc(5) man page for further details.

See also: LWN articles on the kernel 2.6.38 merge window (1, 2) and the Kernel Newbies kernel 2.6.38 summary.

Linux 2.6.37 (4 Jan 2011)

API changes include the following:

• The permissions on /proc/PID/limits changed from readable for the owner only to readable for all users on the system.
• The fanotify_init() and fanotify_mark() system calls were added. These system calls are designed for use in virus-scanning tools, but may also serve other more general uses. They provide functionality that is in some ways similar to inotify(7). Note, however, that the fanotify interface is not a superset of inotify. (The existence of two APIs with heavily overlapping functionality, rather than a new API that is a superset of the earlier API, is unfortunate.) These two system calls were added in Linux 2.6.36, but disabled while concerns about the API were resolved. In Linux 2.6.37, the system calls have been enabled. As yet, there are no manual pages for fanotify; in the meantime, see the Kernel Newbies 2.6.36 changes page and this LWN.net article for further information.

See also: LWN articles on the kernel 2.6.37 merge window (1, 2) and the Kernel Newbies kernel 2.6.37 summary.

Linux 2.6.36 (20 Oct 2010)

API changes include the following:

• The new prlimit() system call is an enhancement of setrlimit() and getrlimit(). It allows the caller to both set and retrieve its own resource limits (including retrieving the old limit at the same time as a new limit is set), and (with suitable permissions) perform the same task for other processes. This system call does not suffer this kernel bug, which affects getrlimit()/setrlimit(). (See pages 759 and 760 of the book.) Indeed, starting with version 2.13, glibc provides library implementations for setrlimit() and getrlimit() that employ prlimit() to work around the kernel bug. I've added documentation of this system call to the getrlimit(2) man page.
• The inotify API adds a new flag, IN_EXCL_UNLINK, that prevents children of a watched directory from generating events for a directory after they have been unlinked from that directory. I've added documentation of this flag to the inotify(7) man page.
• The OOM killer was been rewritten (again). In the process, the /proc/PID/oom_adj file became obsolete, in favor of the new /proc/PID/oom_score_adj file. For further information, see the proc(5) manual page.

See also: LWN articles on the kernel 2.6.36 merge window (1, 2) and the Kernel Newbies kernel 2.6.36 summary.

glibc API changes

glibc 2.18 (Not yet released)

API changes include the following:

• …

glibc 2.17 (25 Dec 2012)

Note: the minimum Linux kernel version to run with this glibc version is Linux 2.6.16.

API changes include the following:

• A new secure_getenv() function allows secure access to the environment. It is similar to getenv(3), but returns NULL if running in a set-user-ID/set-group-ID process. Documentation can be found in the secure_getenv(3) manual page.

glibc 2.16 (30 Jun 2012)

Note: this and subsequent glibc versions are not expected to work with any Linux kernel less than version 2.6.

API changes include the following:

• The glibc header files now handle the _ISOC11_SOURCE feature test macro, as a mechanism for exposing declarations conforming to the C11 standard.
• A new getauxval(3) function allows retrieval of auxiliary vector (AT_*) key-value pairs passed from the Linux kernel. Further information can be found in my LWN.net article "getauxval() and the auxiliary vector" and in the getauxval(3) manual page that I wrote.

glibc 2.15 (tagged 25 Dec 2011)

API changes include the following:

• A new scandirat() function, which is to scandir() as openat(2) is to open(). Documentation can be found in the scandirat(3) manual page.

glibc 2.14 (tagged 31 May 2011)

No API changes (other than simple wrappers for recently added Linux system calls).

glibc 2.13 (tagged 17 Jan 2011)

API changes include the following:

• Newly added library implementation of setrlimit() and getrlimit() bypass the system calls of the same name, instead using the prlimit() system call to bypass the bug described above in the API changes for Linux 2.6.39.

(C) 2013 Michael Kerrisk,   mtk AT man7.org