preadv2 / pwritev2(2)

Synopsis

#include <sys/uio.h>

ssize_t preadv2(int fd, const struct iovec *iov, int iovcnt,
                off_t offset, int flags);
ssize_t pwritev2(int fd, const struct iovec *iov, int iovcnt,
                 off_t offset, int flags);

Rod's take

preadv2/pwritev2 extend preadv/pwritev with a flags argument that applies to a single call. RWF_NOWAIT is the standout: it returns EAGAIN instead of blocking on a page cache miss, enabling a polling fast-path without io_uring. RWF_DSYNC/RWF_SYNC give per-call durability without setting O_DSYNC globally on the fd. RWF_HIPRI requests polled NVMe delivery but silently falls back on unsupported devices. Pass flags=0 and behavior is identical to preadv/pwritev.

Flags and arguments that matter

Errors you will actually hit

Example

#include <sys/uio.h>

struct iovec iov = { .iov_base = buf, .iov_len = len };

/* non-blocking read: fast-path from page cache only */
ssize_t n = preadv2(fd, &iov, 1, offset, RWF_NOWAIT);
if (n < 0 && errno == EAGAIN) {
    /* cache miss: queue blocking read or io_uring sqe */
}

/* single durable write without O_DSYNC on the fd */
n = pwritev2(fd, &iov, 1, offset, RWF_DSYNC);

Trimmed from the full walkthrough in preadv2(2): per-call I/O flags without reopening the fd.

BSD equivalent

FreeBSD has preadv/pwritev (FreeBSD 6, 2005) but no preadv2/pwritev2 and no RWF_* flags. Per-call durability requires explicit fdatasync(2)/fsync(2). Polled I/O is handled at the GEOM/cam_iosched layer, not via fd flags.