open(2)

Numbers by call, from the ledger and cross-checked against the cached kernel tables (syscall_64.tbl, syscall_32.tbl, syscall_arm.tbl, syscall_generic.tbl):

Hub row note: the domain hub’s grouped row shows creat’s number set (85 / 8 / - / 8) because one row cannot hold three calls; the per-call values above are the ledger values. There is no open slot on arm64-generic: glibc’s open() there resolves to openat (56), consistent with the glibc 2.26+ wrapper behavior (man2/open.2 C library/kernel differences; worklog 1-c numbering notes).

Purpose

open() resolves pathname and returns a new file descriptor referring to a new open file description; the descriptor is the lowest-numbered free one and the file offset starts at 0. The open file description carries the file offset and the file status flags; duplicated descriptors and fork children share it, while separate open() calls create separate descriptions. openat() is the dirfd-relative form of the same operation; creat() is exactly open(pathname, O_CREAT | O_WRONLY | O_TRUNC, mode). The flags argument splits into one access mode plus file creation flags (which affect the open itself) and file status flags (which affect later I/O and are partly readable/changeable through fcntl(2)).

Synopsis

#include <fcntl.h>
int open(const char *pathname, int flags);
int open(const char *pathname, int flags, mode_t mode);
int openat(int dirfd, const char *pathname, int flags, ...);
int creat(const char *pathname, mode_t mode);

Arguments and behavior

O_* constants (values from /usr/include/asm-generic/fcntl.h and bits/fcntl-linux.h, x86-64 octal; “since” from man2/open.2):

AT_* constants for this family: openat() itself takes no flags argument; the only AT_* constant it uses is the dirfd special value AT_FDCWD = -100 (/usr/include/linux/fcntl.h). The other AT_* bits (AT_SYMLINK_NOFOLLOW, AT_EMPTY_PATH, and so on) belong to the other *at() calls and are documented on their pages (fstatat on stat.md, statx on statx.md).

Return value and errors

On success all three calls return the new nonnegative file descriptor. On error, -1 and errno: EACCES (search/permission denied, including the protected_fifos and protected_regular sysctl cases), EBADF (openat: relative path with invalid dirfd), EBUSY (O_EXCL on a mounted block device), EDQUOT, EEXIST (O_CREAT|O_EXCL), EFAULT, EINTR (blocked open of a slow device interrupted), EINVAL (O_DIRECT unsupported, invalid flag, O_TMPFILE misuse), EISDIR, ELOOP, EMFILE/ENFILE, ENAMETOOLONG, ENOENT, ENOMEM, ENOSPC, ENOTDIR, ENXIO, EOPNOTSUPP (O_TMPFILE unsupported), EOVERFLOW (32-bit open of a >2 GiB file; EFBIG before 2.6.24), EPERM (O_NOATIME without ownership; write blocked by a file seal, see fcntl.md), EROFS, ETXTBSY (writing a running executable, swap file, or module being loaded), EWOULDBLOCK (incompatible lease).

Security and notes

• O_CLOEXEC is the fd-leak defense: setting FD_CLOEXEC via a separate fcntl(F_SETFD) after open() races with a concurrent fork+execve in multithreaded programs and can leak the fd into the child (man2/open.2, and the same warning on fcntl.md F_SETFD).
• O_NOATIME requires the caller’s effective UID to match the file owner, or CAP_FOWNER in the caller’s user namespace; otherwise EPERM.
• Writing can be refused with EPERM by F_SEAL_SHRINK/GROW/WRITE seals on a memfd (fcntl.md).
• O_PATH descriptors are the /proc/self/fd/N idiom for exec/fexecve without read permission; O_PATH needs no permission on the object itself, only search permission on the path prefix.
• openat2 (kernel 5.6; x64 437 / i386 437 / g64 437 / arm32 437) extends openat with struct open_how and RESOLVE_* restrictions; it has its own hub row and page (openat2.md), not covered here.

References

• man2/open.2 in the cached man-pages tree (dated 2021-08-27); .so aliases creat.2, openat.2.
• Constants: /usr/include/asm-generic/fcntl.h, /usr/include/x86_64-linux-gnu/bits/fcntl-linux.h, /usr/include/linux/fcntl.h.
• Syscalls ledger rows SC-open, SC-openat, SC-creat; kernel tables in source-cache/kernel.
• Domain hub for status and sibling pages.