inotify: Add man pages for the new syscalls and for VOP_INOTIFY

Reviewed by:	kib
MFC after:	3 months
Sponsored by:	Klara, Inc.
Differential Revision:	https://reviews.freebsd.org/D50315
This commit is contained in:
Mark Johnston 2025-07-03 20:02:21 +00:00
parent e02d4fab54
commit 5202cfcf93
4 changed files with 447 additions and 0 deletions

View file

@ -224,6 +224,7 @@ MAN+= abort2.2 \
getsockopt.2 \
gettimeofday.2 \
getuid.2 \
inotify.2 \
intro.2 \
ioctl.2 \
issetugid.2 \
@ -448,6 +449,11 @@ MLINKS+=getrlimit.2 setrlimit.2
MLINKS+=getsockopt.2 setsockopt.2
MLINKS+=gettimeofday.2 settimeofday.2
MLINKS+=getuid.2 geteuid.2
MLINKS+=inotify.2 inotify_init.2 \
inotify.2 inotify_init1.2 \
inotify.2 inotify_add_watch.2 \
inotify.2 inotify_add_watch_at.2 \
inotify.2 inotify_rm_watch.2
MLINKS+=intro.2 errno.2
MLINKS+=jail.2 jail_attach.2 \
jail.2 jail_get.2 \

379
lib/libsys/inotify.2 Normal file
View file

@ -0,0 +1,379 @@
.\"
.\" SPDX-License-Identifier: BSD-2-Clause
.\"
.\" Copyright (c) 2025 Klara, Inc.
.\"
.Dd May 19, 2025
.Dt INOTIFY 2
.Os
.Sh NAME
.Nm inotify_init ,
.Nm inotify_init1 ,
.Nm inotify_add_watch ,
.Nm inotify_add_watch_at ,
.Nm inotify_rm_watch
.Nd monitor file system events
.Sh LIBRARY
.Lb libc
.Sh SYNOPSIS
.In sys/inotify.h
.Ft int
.Fo inotify_init
.Fc
.Ft int
.Fo inotify_init1
.Fa "int flags"
.Fc
.Ft int
.Fo inotify_add_watch
.Fa "int fd"
.Fa "const char *pathname"
.Fa "uint32_t mask"
.Fc
.Ft int
.Fo inotify_add_watch_at
.Fa "int fd"
.Fa "int dfd"
.Fa "const char *pathname"
.Fa "uint32_t mask"
.Fc
.Ft int
.Fo inotify_rm_watch
.Fa "int fd"
.Fa "uint32_t wd"
.Fc
.Bd -literal
struct inotify_event {
int wd; /* Watch descriptor */
uint32_t mask; /* Event and flags */
uint32_t cookie; /* Unique ID which links rename events */
uint32_t len; /* Name field size, including nul bytes */
char name[0]; /* Filename (nul-terminated) */
};
.Ed
.Sh DESCRIPTION
The inotify system calls provide an interface to monitor file system events.
They aim to be compatible with the Linux inotify interface.
The provided functionality is similar to the
.Dv EVFILT_VNODE
filter of the
.Xr kevent 2
system call, but further allows monitoring of a directory without needing to
open each object in that directory.
This avoids races and reduces the number of file descriptors needed to monitor
a large file hierarchy.
.Pp
inotify allows one or more file system objects, generally files or directories,
to be watched for events, such as file open or close.
Watched objects are associated with a file descriptor returned
by
.Fn inotify_init
or
.Fn inotify_init1 .
When an event occurs, a record describing the event becomes available for
reading from the inotify file descriptor.
Each inotify descriptor thus refers to a queue of events waiting to be read.
inotify descriptors are inherited across
.Xr fork 2
calls and may be passed to other processes via
.Xr unix 4
sockets.
.Pp
The
.Fn inotify_init1
system call accepts two flags.
The
.Dv IN_NONBLOCK
flag causes the inotify descriptor to be opened in non-blocking mode, such that
.Xr read 2
calls will not block if no records are available to consume, and will instead
return
.Er EWOULDBLOCK .
The
.Dv IN_CLOEXEC
flag causes the inotify descriptor to be closed automatically when
.Xr execve 2
is called.
.Pp
To watch a file or directory, the
.Fn inotify_add_watch
or
.Fn inotify_add_watch_at
system calls must be used.
They take a path and a mask of events to watch for, and return a
.Dq watch descriptor ,
a non-negative integer which uniquely identifies the watched object within the
inotify descriptor.
.Pp
The
.Fn inotify_rm_watch
system call removes a watch from an inotify descriptor.
.Pp
When watching a directory, objects within the directory are monitored for events
as well as the directory itself.
A record describing an inotify event consists of a
.Dq struct inotify_event
followed by the name of the object in the directory being watched.
If the watched object itself generates an event, no name is present.
Extra nul bytes may follow the file name in order to provide alignment for a
subsequent record.
.Pp
The following events are defined:
.Bl -tag -width IN_CLOSE_NOWRITE
.It Dv IN_ACCESS
A file's contents were accessed, e.g., by
.Xr read 2
.Xr copy_file_range 2 ,
.Xr sendfile 2 ,
or
.Xr getdirentries 2 .
.It Dv IN_ATTRIB
A file's metadata was changed, e.g., by
.Xr chmod 2
or
.Xr unlink 2 .
.It Dv IN_CLOSE_WRITE
A file that was previously opened for writing was closed.
.It Dv IN_CLOSE_NOWRITE
A file that was previously opened read-only was closed.
.It Dv IN_CREATE
A file within a watched directory was created, e.g., by
.Xr open 2 ,
.Xr mkdir 2 ,
.Xr symlink 2 ,
.Xr mknod 2 ,
or
.Xr bind 2 .
.It Dv IN_DELETE
A file or directory within a watched directory was removed.
.It Dv IN_DELETE_SELF
The watched file or directory itself was deleted.
This event is generated only when the link count of the file drops
to zero.
.It Dv IN_MODIFY
A file's contents were modified, e.g., by
.Xr write 2
or
.Xr copy_file_range 2 .
.It Dv IN_MOVE_SELF
The watched file or directory itself was renamed.
.It Dv IN_MOVED_FROM
A file or directory was moved from a watched directory.
.It Dv IN_MOVED_TO
A file or directory was moved into a watched directory.
A
.Xr rename 2
call thus may generate two events, one for the old name and one for the new
name.
These are linked together by the
.Ar cookie
field in the inotify record, which can be compared to link the two records
to the same event.
.It Dv IN_OPEN
A file was opened.
.El
.Pp
Some additional flags may be set in inotify event records:
.Bl -tag -width IN_Q_OVERFLOW
.It Dv IN_IGNORED
When a watch is removed from a file, for example because it was created with the
.Dv IN_ONESHOT
flag, the file was deleted, or the watch was explicitly removed with
.Xr inotify_rm_watch 2 ,
an event with this mask is generated to indicate that the watch will not
generate any more events.
Once this event is generated, the watch is automatically removed, and in
particular should not be removed manually with
.Xr inotify_rm_watch 2 .
.It Dv IN_ISDIR
When the subject of an event is a directory, this flag is set in the
.Ar mask
.It Dv IN_Q_OVERFLOW
One or more events were dropped, for example because of a kernel memory allocation
failure or because the event queue size hit a limit.
.It Dv IN_UNMOUNT
The filesystem containing the watched object was unmounted.
.El
.Pp
A number of flags may also be specified in the
.Ar mask
given to
.Fn inotify_add_watch
and
.Fn inotify_add_watch_at :
.Bl -tag -width IN_DONT_FOLLOW
.It Dv IN_DONT_FOLLOW
If
.Ar pathname
is a symbolic link, do not follow it.
.It Dv IN_EXCL_UNLINK
This currently has no effect, see the
.Sx BUGS
section.
.In Dv IN_MASK_ADD
When adding a watch to an object, and that object is already watched by the
same inotify descriptor, by default the mask of the existing watch is
overwritten.
When
.Dv IN_MASK_ADD
is specified, the mask of the existing watch is instead logically ORed with
the new mask.
.In Dv IN_MASK_CREATE
When
.Fn inotify_add watch
is used to add a watch to an object,
.Dv IN_MASK_CREATE
is specified, and that object is already watched by the same inotify descriptor,
return an error instead of updating the existing watch.
.In Dv IN_ONESHOT
Monitor the object for a single event, after which the watch is automatically
removed.
As part of removal, a
.Dv IN_IGNORED
event is generated.
.In Dv IN_ONLYDIR
When creating a watch, fail with
.Er ENOTDIR
if the path does not refer to a directory.
.El
.Sh SYSCTL VARIABLES
The following variables are available as both
.Xr sysctl 8
variables and
.Xr loader 8
tunables:
.Bl -tag -width 15
.It Va vfs.inotify.max_events
The maximum number of inotify records that can be queued for a single
inotify descriptor.
Records in excess of this limit are discarded, and a single event with
mask equal to
.Dv IN_Q_OVERFLOW
will be present in the queue.
.It Va vfs.inotify.max_user_instances
The maximum number of inotify descriptors that can be created by a single
user.
.It Va vfs.inotify.max_user_watches
The maximum number of inotify watches per user.
.El
.Sh EXAMPLES
See the example program in
.Pa /usr/share/examples/inotify/inotify.c .
.Sh ERRORS
The
.Fn inotify_init
and
.Fn inotify_init1
functions will fail if:
.Bl -tag -width Er
.It Bq Er ENFILE
The system limit on the total number of open files has been reached.
.It Bq Er EMFILE
A per-process limit on the number of open files has been reached.
.It Bq Er EMFILE
The system limit on the number of inotify descriptors has been reached.
.It Bq Er EINVAL
An unrecognized flag was passed to
.Fn inotify_init1 .
.El
.Pp
The
.Fn inotify_add_watch
and
.Fn inotify_add_watch_at
system calls will fail if:
.Bl -tag -width Er
.It Bq Er EBADF
The
.Ar fd
parameter is not a valid file descriptor.
.It Bq Er EINVAL
The
.Ar fd
parameter is not an inotify descriptor.
.It Bq Er EINVAL
The
.Ar mask
parameter does not specify an event, or
the
.Dv IN_MASK_CREATE
and
.Dv IN_MASK_ADD
flags are both set, or an unrecognized flag was passed.
.It Bq Er ENOTDIR
The
.Ar pathname
parameter refers to a file that is not a directory, and the
.Dv IN_ONLYDIR
flag was specified.
.It Bq Er ENOSPC
The per-user limit on the total number of inotify watches has been reached.
.It Bq Er ECAPMODE
The process is in capability mode and
.Fn inotify_add_watch
was called, or
.Fn inotify_add_watch_at
was called with
.Dv AT_FDCWD
as the directory file descriptor
.Ar dfd .
.It Bq Er ENOTCAPABLE
The process is in capability mode and
.Ar pathname
contains a
.Dq ..
component leading to a directory outside the directory hierarchy specified
by
.Ar dfd .
.El
.Pp
The
.Fn inotify_rm_watch
system call will fail if:
.Bl -tag -width Er
.It Bq Er EBADF
The
.Ar fd
parameter is not a valid file descriptor.
.It Bq Er EINVAL
The
.Ar fd
parameter is not an inotify descriptor.
.It Bq Er EINVAL
The
.Ar wd
parameter is not a valid watch descriptor.
.El
.Sh SEE ALSO
.Xr kevent 2 ,
.Xr capsicum 4
.Sh STANDARDS
The
.Nm
interface originates from Linux and is non-standard.
This implementation aims to be compatible with that of Linux and is based
on the documentation available at
.Pa https://man7.org/linux/man-pages/man7/inotify.7.html .
.Sh HISTORY
The inotify system calls first appeared in
.Fx 15.0 .
.Sh BUGS
If a file in a watched directory has multiple hard links,
an access via any hard link for that file will generate an event, even
if the accessed link belongs to an unwatched directory.
This is not the case for the Linux implementation, where only accesses
via the hard link in the watched directory will generate an event.
.Pp
If a watched directory contains multiple hard links of a file, an event
on one of the hard links will generate an inotify record for each link
in the directory.
.Pp
When a file is unlinked, no more events will be generated for that file,
even if it continues to be accessed.
By default, the Linux implementation will continue to generate events in
this case.
Thus, the
.Fx
implementation behaves as though
.Dv IN_EXCL_UNLINK
is always set.

View file

@ -434,6 +434,7 @@ MAN= accept_filter.9 \
VOP_GETEXTATTR.9 \
VOP_GETPAGES.9 \
VOP_INACTIVE.9 \
VOP_INOTIFY.9 \
VOP_IOCTL.9 \
VOP_LINK.9 \
VOP_LISTEXTATTR.9 \
@ -2460,6 +2461,7 @@ MLINKS+=VOP_CREATE.9 VOP_MKDIR.9 \
MLINKS+=VOP_FSYNC.9 VOP_FDATASYNC.9
MLINKS+=VOP_GETPAGES.9 VOP_PUTPAGES.9
MLINKS+=VOP_INACTIVE.9 VOP_RECLAIM.9
MLINKS+=VOP_INOTIFY.9 VOP_INOTIFY_ADD_WATCH.9
MLINKS+=VOP_LOCK.9 vn_lock.9 \
VOP_LOCK.9 VOP_ISLOCKED.9 \
VOP_LOCK.9 VOP_UNLOCK.9

View file

@ -0,0 +1,60 @@
.\"-
.\" SPDX-License-Identifier: BSD-2-Clause
.\"
.\" Copyright (c) 2025 Klara, Inc.
.\"
.Dd May 27, 2025
.Dt VOP_INOTIFY 9
.Os
.Sh NAME
.Nm VOP_INOTIFY
.Nd vnode inotify interface
.Sh SYNOPSIS
.In sys/param.h
.In sys/vnode.h
.Ft int
.Fo VOP_INOTIFY
.Fa struct vnode *vp
.Fa struct vnode *dvp
.Fa struct componentname *cnp
.Fa int event
.Fa uint32_t cookie
.Fc
.Ft int
.Fo VOP_INOTIFY_ADD_WATCH
.Fa struct vnode *vp
.Fa struct inotify_softc *sc
.Fa uint32_t mask
.Fa uint32_t *wdp
.Fa struct thread *td
.Fc
.Sh DESCRIPTION
The
.Fn VOP_INOTIFY
operation notifies the
.Xr inotify 2
subsystem of a file system event on a vnode.
The
.Fa dvp
and
.Fa cnp
arguments are optional and are only used to obtain a file name for the event.
If they are omitted, the inotify subsystem will use the file name cache to
find a name for the vnode, but this is more expensive.
.Pp
The
.Fn VOP_INOTIFY_ADD_WATCH
operation is for internal use by the inotify subsystem to add a watch to a
vnode.
.Sh LOCKS
The
.Fn VOP_INOTIFY
operation does not assume any particular vnode lock state.
The
.Fn VOP_INOTIFY_ADD_WATCH
operation should be called with the vnode locked.
.Sh RETURN VALUES
Zero is returned on success, otherwise an error code is returned.
.Sh SEE ALSO
.Xr inotify 2 ,
.Xr vnode 9