fanotify_init(2)

🌸 NAME 🌸

okay bestie, meet fanotify_init — the syscall from section 2 that creates and initializes a fanotify group. it's giving "I watch files for a living" energy. 💖

---

📚 LIBRARY 📚

Standard C library (libc, -lc). no cap.

---

💻 SYNOPSIS 💻

#include <fcntl.h>            /* Definition of O_* constants */
#include <sys/fanotify.h>

int fanotify_init(unsigned int flags, unsigned int event_f_flags);

---

✨ DESCRIPTION ✨

First things first, bestie: for the whole fanotify API tea, go read fanotify(7). 🫖

So fanotify_init() initializes a brand new fanotify group and returns a file descriptor for the event queue that's associated with that group. main character energy fr fr. 👑

That file descriptor is used in calls to fanotify_mark(2) to specify the files, directories, mounts, or filesystems for which fanotify events shall be created. Those events get received by reading from the file descriptor. Some events are only informative — like "hey, this file got accessed." Other events can be used to determine whether another application is permitted to access a file or directory. Permission to access filesystem objects is granted by writing to the file descriptor. periodt. ✍️

Multiple programs may be using the fanotify interface at the same time to monitor the same files. sharing is caring, lowkey. 🤝

The number of fanotify groups per user is limited. See fanotify(7) for details about this limit. don't test it, bestie. 🚫

The flags argument contains a multi-bit field defining the notification class of the listening application and further single bit fields specifying the behavior of the file descriptor. so it's like a little combo platter of vibes. 🍱

If multiple listeners for permission events exist, the notification class is used to establish the sequence in which the listeners receive the events. order matters, iykyk. 🔢

Only one of the following notification classes may be specified in flags:

• FAN_CLASS_PRE_CONTENT — This value allows the receipt of events notifying that a file has been accessed and events for permission decisions if a file may be accessed. It is intended for event listeners that may need to write data to files before their final data can be accessed. This notification class might be used by hierarchical storage managers, for example. Use of this flag requires the CAP_SYS_ADMIN capability. so if you're not admin-capable, this is not your moment. 🙅‍♀️
• FAN_CLASS_CONTENT — This value allows the receipt of events notifying that a file has been accessed and events for permission decisions if a file may be accessed. It is intended for event listeners that need to access files when they already contain their final content. This notification class might be used by malware detection programs, for example. Use of this flag requires the CAP_SYS_ADMIN capability. malware detection is giving superhero energy. 🦸‍♀️
• FAN_CLASS_NOTIF — This is the default value. It does not need to be specified. This value only allows the receipt of events notifying that a file has been accessed. Permission decisions before the file is accessed are not possible. so it's just the chill notif vibe. 🔔

Listeners with different notification classes will receive events in the order FAN_CLASS_PRE_CONTENT, FAN_CLASS_CONTENT, FAN_CLASS_NOTIF. The order of notification for listeners in the same notification class is undefined. so don't @ me about fairness. 🤷‍♀️

The following bits can additionally be set in flags:

• FAN_CLOEXEC — Set the close-on-exec flag (FD_CLOEXEC) on the new file descriptor. See the description of the O_CLOEXEC flag in open(2). classic fd hygiene, bestie. 🧼
• FAN_NONBLOCK — Enable the nonblocking flag (O_NONBLOCK) for the file descriptor. Reading from the file descriptor will not block. Instead, if no data is available, read(2) fails with the error EAGAIN. so it's like, "I'm not waiting around." 💅
• FAN_UNLIMITED_QUEUE — Remove the limit on the number of events in the event queue. See fanotify(7) for details about this limit. Use of this flag requires the CAP_SYS_ADMIN capability. queue go brrr. 🚀
• FAN_UNLIMITED_MARKS — Remove the limit on the number of fanotify marks per user. See fanotify(7) for details about this limit. Use of this flag requires the CAP_SYS_ADMIN capability. marks for days. ✨
• FAN_REPORT_TID (since Linux 4.20) — Report thread ID (TID) instead of process ID (PID) in the pid field of the struct fanotify_event_metadata supplied to read(2) (see fanotify(7)). Use of this flag requires the CAP_SYS_ADMIN capability. threads gotta thread. 🧵
• FAN_ENABLE_AUDIT (since Linux 4.15) — Enable generation of audit log records about access mediation performed by permission events. The permission event response has to be marked with the FAN_AUDIT flag for an audit log record to be generated. Use of this flag requires the CAP_AUDIT_WRITE capability. audit tea, spilled for the logs. 🍵
• FAN_REPORT_FID (since Linux 5.1) — This value allows the receipt of events which contain additional information about the underlying filesystem object correlated to an event. An additional record of type FAN_EVENT_INFO_TYPE_FID encapsulates the information about the object and is included alongside the generic event metadata structure. The file descriptor that is used to represent the object correlated to an event is instead substituted with a file handle. It is intended for applications that may find the use of a file handle to identify an object more suitable than a file descriptor. Additionally, it may be used for applications monitoring a directory or a filesystem that are interested in the directory entry modification events FAN_CREATE, FAN_DELETE, FAN_MOVE, and FAN_RENAME, or in events such as FAN_ATTRIB, FAN_DELETE_SELF, and FAN_MOVE_SELF. All the events above require an fanotify group that identifies filesystem objects by file handles. Note that without the flag FAN_REPORT_TARGET_FID, for the directory entry modification events, there is an information record that identifies the modified directory and not the created/deleted/moved child object. The use of FAN_CLASS_CONTENT or FAN_CLASS_PRE_CONTENT is not permitted with this flag and will result in the error EINVAL. See fanotify(7) for additional details. FID is the vibes. 🆔
• FAN_REPORT_DIR_FID (since Linux 5.9) — Events for fanotify groups initialized with this flag will contain (see exceptions below) additional information about a directory object correlated to an event. An additional record of type FAN_EVENT_INFO_TYPE_DFID encapsulates the information about the directory object and is included alongside the generic event metadata structure. For events that occur on a non-directory object, the additional structure includes a file handle that identifies the parent directory filesystem object. Note that there is no guarantee that the directory filesystem object will be found at the location described by the file handle information at the time the event is received. When combined with the flag FAN_REPORT_FID, two records may be reported with events that occur on a non-directory object, one to identify the non-directory object itself and one to identify the parent directory object. Note that in some cases, a filesystem object does not have a parent, for example, when an event occurs on an unlinked but open file. In that case, with the FAN_REPORT_FID flag, the event will be reported with only one record to identify the non-directory object itself, because there is no directory associated with the event. Without the FAN_REPORT_FID flag, no event will be reported. See fanotify(7) for additional details. directory handle drama, but make it informative. 📂
• FAN_REPORT_NAME (since Linux 5.9) — Events for fanotify groups initialized with this flag will contain additional information about the name of the directory entry correlated to an event. This flag must be provided in conjunction with the flag FAN_REPORT_DIR_FID. Providing this flag value without FAN_REPORT_DIR_FID will result in the error EINVAL. This flag may be combined with the flag FAN_REPORT_FID. An additional record of type FAN_EVENT_INFO_TYPE_DFID_NAME, which encapsulates the information about the directory entry, is included alongside the generic event metadata structure and substitutes the additional information record of type FAN_EVENT_INFO_TYPE_DFID. The additional record includes a file handle that identifies a directory filesystem object followed by a name that identifies an entry in that directory. For the directory entry modification events FAN_CREATE, FAN_DELETE, and FAN_MOVE, the reported name is that of the created/deleted/moved directory entry. The event FAN_RENAME may contain two information records. One of type FAN_EVENT_INFO_TYPE_OLD_DFID_NAME identifying the old directory entry, and another of type FAN_EVENT_INFO_TYPE_NEW_DFID_NAME identifying the new directory entry. For other events that occur on a directory object, the reported file handle is that of the directory object itself and the reported name is '.'. For other events that occur on a non-directory object, the reported file handle is that of the parent directory object and the reported name is the name of a directory entry where the object was located at the time of the event. The rationale behind this logic is that the reported directory file handle can be passed to open_by_handle_at(2) to get an open directory file descriptor and that file descriptor along with the reported name can be used to call fstatat(2). The same rule that applies to record type FAN_EVENT_INFO_TYPE_DFID also applies to record type FAN_EVENT_INFO_TYPE_DFID_NAME: if a non-directory object has no parent, either the event will not be reported or it will be reported without the directory entry information. Note that there is no guarantee that the filesystem object will be found at the location described by the directory entry information at the time the event is received. See fanotify(7) for additional details. name drop done right. 📛
• FAN_REPORT_DFID_NAME — This is a synonym for (FAN_REPORT_DIR_FID|FAN_REPORT_NAME). short and sweet, iconic. 💅
• FAN_REPORT_TARGET_FID (since Linux 5.17, 5.15.154, and 5.10.220) — bestie, if you init your fanotify group with this flag, the events are gonna come with extra info about the child that's correlated with directory-entry-modification events. ✨ But you cannot run this solo: it must be provided together with FAN_REPORT_FID, FAN_REPORT_DIR_FID and FAN_REPORT_NAME — or else the error EINVAL will be returned, no cap. For the directory entry modification events FAN_CREATE, FAN_DELETE, FAN_MOVE, and FAN_RENAME, an additional record of type FAN_EVENT_INFO_TYPE_FID gets reported on top of the info records of type FAN_EVENT_INFO_TYPE_DFID, FAN_EVENT_INFO_TYPE_DFID_NAME, FAN_EVENT_INFO_TYPE_OLD_DFID_NAME, and FAN_EVENT_INFO_TYPE_NEW_DFID_NAME. That extra record carries a file handle that identifies the filesystem child object the directory entry is pointing at. 🫶
• FAN_REPORT_DFID_NAME_TARGET — this one is literally just a synonym for (FAN_REPORT_DFID_NAME|FAN_REPORT_FID|FAN_REPORT_TARGET_FID). Periodt. 💅
• FAN_REPORT_MNT (since Linux 6.14) — this value lets you receive events that carry extra info about the underlying mount correlated to an event. 🌈 An extra record of type FAN_EVENT_INFO_TYPE_MNT wraps up the mount info and comes along for the ride right alongside the generic event metadata structure. But heads up, because using FAN_CLASS_CONTENT, FAN_CLASS_PRE_CONTENT, or any of the FAN_REPORT_DFID_NAME_TARGET flags together with this flag is not permitted and will result in the error EINVAL being returned. 🚫 See fanotify(7) for the extra deets.
• FAN_REPORT_PIDFD (since Linux 5.15 and 5.10.220) — groups initialized with this flag get an additional info record sitting pretty alongside the generic fanotify_event_metadata structure. That record is of type FAN_EVENT_INFO_TYPE_PIDFD and it contains a pidfd for the process that was responsible for generating the event. 👑 And a pidfd returned in this info record object is no different from the pidfd you get when you call pidfd_open(2) — same energy. This info record's whole thing is for apps that wanna reliably figure out whether the process responsible for generating an event has been recycled or terminated. 💀 Now the downside: using the FAN_REPORT_TID flag together with FAN_REPORT_PIDFD is currently not supported, and attempting it will result in the error EINVAL being returned. This limitation is currently imposed by the pidfd API, since it only supports the creation of pidfds for thread-group leaders rn. Creating pidfds for non-thread-group leaders may be supported at some point in the future, so this restriction may eventually be lifted. 🙏 For more deets on info records, see fanotify(7).
• FAN_REPORT_FD_ERROR (since Linux 6.13 and 6.12.4 and 6.6.66) — events for groups initialized with this flag may contain an error code that explains the reason for the failure to open a file descriptor. The .fd member of the fanotify_event_metadata structure normally contains an open file descriptor associated with the object of the event, or FAN_NOFD in case a file descriptor could not be opened. But for a group initialized with this flag, instead of FAN_NOFD, the .fd member of the fanotify_event_metadata structure will contain a negative error value. And when the group is also initialized with the flag FAN_REPORT_PIDFD, in case a process file descriptor could not be opened, the .pidfd member of the fanotify_event_info_pidfd structure will also contain a negative error value. For more details, see fanotify(7). 🔍

---

Okay so now we're moving on to event_f_flags. 🧚 The event_f_flags argument defines the file status flags that get set on the open file descriptions that are created for fanotify events. For the details of these flags, go peek at the flag values description in open(2). And event_f_flags includes a multi-bit field for the access mode. 👀 That field can take the following values:

• O_RDONLY — this value allows only read access. 📖
• O_WRONLY — this value allows only write access. ✍️
• O_RDWR — this value allows read and write access. 💫

Additional bits can also be set in event_f_flags. The most useful values are:

• O_LARGEFILE — enable support for files exceeding 2 GB. 🐘 Failing to set this flag will result in an EOVERFLOW error when trying to open a large file which is monitored by an fanotify group on a 32-bit system. Not-a-vibe, so set it.
• O_CLOEXEC (since Linux 3.18) — enable the close-on-exec flag for the file descriptor. See the description of the O_CLOEXEC flag in open(2) for the reasons why this may be useful. 💅

The following are also allowable: O_APPEND, O_DSYNC, O_NOATIME, O_NONBLOCK, and O_SYNC. ✨ Specifying any other flag in event_f_flags yields the error EINVAL (but see BUGS, bestie, there's lore there).

---

🎀 RETURN VALUE 🎀

On success, fanotify_init() returns a new file descriptor. 🥳 On error, -1 is returned, and errno is set to indicate the error. 😭

---

💖 ERRORS 💖

• EINVAL — an invalid value was passed in flags or event_f_flags. FAN_ALL_INIT_FLAGS (deprecated since Linux 4.20) defines all allowable bits for flags. 🙈
• EMFILE — the number of fanotify groups for this user exceeds the limit. See fanotify(7) for details about this limit. 😬
• EMFILE — the per-process limit on the number of open file descriptors has been reached. (Yes, it's listed twice, that's just how it is, we don't make the rules. 💀)
• ENOMEM — the allocation of memory for the notification group failed. 🫠
• ENOSYS — this kernel does not implement fanotify_init(). The fanotify API is available only if the kernel was configured with CONFIG_FANOTIFY. 🔧
• EPERM — the operation is not permitted because the caller lacks a required capability. 🚫

---

🌸 VERSIONS 🌸

Prior to Linux 5.13 (and 5.10.220), calling fanotify_init() required the CAP_SYS_ADMIN capability. Since Linux 5.13 (and 5.10.220), users may call fanotify_init() without the CAP_SYS_ADMIN capability to create and initialize an fanotify group with limited functionality. 👑

The limitations imposed on an event listener created by a user without the CAP_SYS_ADMIN capability are as follows:

• The user cannot request for an unlimited event queue by using FAN_UNLIMITED_QUEUE. 🙅
• The user cannot request for an unlimited number of marks by using FAN_UNLIMITED_MARKS. 🙅
• The user cannot request to use either notification classes FAN_CLASS_CONTENT or FAN_CLASS_PRE_CONTENT. This means that user cannot request permission events. 🔒
• The user is required to create a group that identifies filesystem objects by file handles, for example, by providing the FAN_REPORT_FID flag. 🪪
• The user is limited to only mark inodes. The ability to mark a mount or filesystem via fanotify_mark() through the use of FAN_MARK_MOUNT or FAN_MARK_FILESYSTEM is not permitted. 📍
• The event object in the event queue is limited in terms of the information that is made available to the unprivileged user. A user will also not receive the pid that generated the event, unless the listening process itself generated the event. 🕵️

---

⭐ STANDARDS ⭐

Linux. 🐧 That's it, that's the standard.

---

🦋 HISTORY 🦋

Linux 2.6.37. 🗓️ Ancient. Iconic. The origin story.

---

💀 BUGS 💀

The following bug was present before Linux 3.18:

• The O_CLOEXEC is ignored when passed in event_f_flags. 😤

The following bug was present before Linux 3.14:

• The event_f_flags argument is not checked for invalid flags. Flags that are intended only for internal use, such as FMODE_EXEC, can be set, and will consequently be set for the file descriptors returned when reading from the fanotify file descriptor. 🚨

Okay bestie real talk — that second one is giving "leaked internals" energy. FMODE_EXEC was never meant for you to set, and on old kernels nothing stops you. Keep your kernels fresh, iykyk. 🫶

---

🔥 SEE ALSO 🔥

• fanotify_mark(2)
• fanotify(7)

---

And that's the whole page, bestie!! 💖 You made it to the end, main character energy fr fr. 🌈 This one came out of Linux man-pages 6.20, dated 2025-05-17 — go forth and fanotify responsibly, and remember: the vibe changes, the facts never do. ✨💅🎀🦋