linux_old1/include/linux/fsnotify_backend.h

393 lines
16 KiB
C
Raw Normal View History

fsnotify: unified filesystem notification backend fsnotify is a backend for filesystem notification. fsnotify does not provide any userspace interface but does provide the basis needed for other notification schemes such as dnotify. fsnotify can be extended to be the backend for inotify or the upcoming fanotify. fsnotify provides a mechanism for "groups" to register for some set of filesystem events and to then deliver those events to those groups for processing. fsnotify has a number of benefits, the first being actually shrinking the size of an inode. Before fsnotify to support both dnotify and inotify an inode had unsigned long i_dnotify_mask; /* Directory notify events */ struct dnotify_struct *i_dnotify; /* for directory notifications */ struct list_head inotify_watches; /* watches on this inode */ struct mutex inotify_mutex; /* protects the watches list But with fsnotify this same functionallity (and more) is done with just __u32 i_fsnotify_mask; /* all events for this inode */ struct hlist_head i_fsnotify_mark_entries; /* marks on this inode */ That's right, inotify, dnotify, and fanotify all in 64 bits. We used that much space just in inotify_watches alone, before this patch set. fsnotify object lifetime and locking is MUCH better than what we have today. inotify locking is incredibly complex. See 8f7b0ba1c8539 as an example of what's been busted since inception. inotify needs to know internal semantics of superblock destruction and unmounting to function. The inode pinning and vfs contortions are horrible. no fsnotify implementers do allocation under locks. This means things like f04b30de3 which (due to an overabundance of caution) changes GFP_KERNEL to GFP_NOFS can be reverted. There are no longer any allocation rules when using or implementing your own fsnotify listener. fsnotify paves the way for fanotify. In brief fanotify is a notification mechanism that delivers the lisener both an 'event' and an open file descriptor to the object in question. This means that fanotify is pathname agnostic. Some on lkml may not care for the original companies or users that pushed for TALPA, but fanotify was designed with flexibility and input for other users in mind. The readahead group expressed interest in fanotify as it could be used to profile disk access on boot without breaking the audit system. The desktop search groups have also expressed interest in fanotify as it solves a number of the race conditions and problems present with managing inotify when more than a limited number of specific files are of interest. fanotify can provide for a userspace access control system which makes it a clean interface for AV vendors to hook without trying to do binary patching on the syscall table, LSM, and everywhere else they do their things today. With this patch series fanotify can be implemented in less than 1200 lines of easy to review code. Almost all of which is the socket based user interface. This patch series builds fsnotify to the point that it can implement dnotify and inotify_user. Patches exist and will be sent soon after acceptance to finish the in kernel inotify conversion (audit) and implement fanotify. Signed-off-by: Eric Paris <eparis@redhat.com> Acked-by: Al Viro <viro@zeniv.linux.org.uk> Cc: Christoph Hellwig <hch@lst.de>
2009-05-22 05:01:20 +08:00
/*
* Filesystem access notification for Linux
*
* Copyright (C) 2008 Red Hat, Inc., Eric Paris <eparis@redhat.com>
*/
#ifndef __LINUX_FSNOTIFY_BACKEND_H
#define __LINUX_FSNOTIFY_BACKEND_H
#ifdef __KERNEL__
#include <linux/idr.h> /* inotify uses this */
fsnotify: unified filesystem notification backend fsnotify is a backend for filesystem notification. fsnotify does not provide any userspace interface but does provide the basis needed for other notification schemes such as dnotify. fsnotify can be extended to be the backend for inotify or the upcoming fanotify. fsnotify provides a mechanism for "groups" to register for some set of filesystem events and to then deliver those events to those groups for processing. fsnotify has a number of benefits, the first being actually shrinking the size of an inode. Before fsnotify to support both dnotify and inotify an inode had unsigned long i_dnotify_mask; /* Directory notify events */ struct dnotify_struct *i_dnotify; /* for directory notifications */ struct list_head inotify_watches; /* watches on this inode */ struct mutex inotify_mutex; /* protects the watches list But with fsnotify this same functionallity (and more) is done with just __u32 i_fsnotify_mask; /* all events for this inode */ struct hlist_head i_fsnotify_mark_entries; /* marks on this inode */ That's right, inotify, dnotify, and fanotify all in 64 bits. We used that much space just in inotify_watches alone, before this patch set. fsnotify object lifetime and locking is MUCH better than what we have today. inotify locking is incredibly complex. See 8f7b0ba1c8539 as an example of what's been busted since inception. inotify needs to know internal semantics of superblock destruction and unmounting to function. The inode pinning and vfs contortions are horrible. no fsnotify implementers do allocation under locks. This means things like f04b30de3 which (due to an overabundance of caution) changes GFP_KERNEL to GFP_NOFS can be reverted. There are no longer any allocation rules when using or implementing your own fsnotify listener. fsnotify paves the way for fanotify. In brief fanotify is a notification mechanism that delivers the lisener both an 'event' and an open file descriptor to the object in question. This means that fanotify is pathname agnostic. Some on lkml may not care for the original companies or users that pushed for TALPA, but fanotify was designed with flexibility and input for other users in mind. The readahead group expressed interest in fanotify as it could be used to profile disk access on boot without breaking the audit system. The desktop search groups have also expressed interest in fanotify as it solves a number of the race conditions and problems present with managing inotify when more than a limited number of specific files are of interest. fanotify can provide for a userspace access control system which makes it a clean interface for AV vendors to hook without trying to do binary patching on the syscall table, LSM, and everywhere else they do their things today. With this patch series fanotify can be implemented in less than 1200 lines of easy to review code. Almost all of which is the socket based user interface. This patch series builds fsnotify to the point that it can implement dnotify and inotify_user. Patches exist and will be sent soon after acceptance to finish the in kernel inotify conversion (audit) and implement fanotify. Signed-off-by: Eric Paris <eparis@redhat.com> Acked-by: Al Viro <viro@zeniv.linux.org.uk> Cc: Christoph Hellwig <hch@lst.de>
2009-05-22 05:01:20 +08:00
#include <linux/fs.h> /* struct inode */
#include <linux/list.h>
#include <linux/path.h> /* struct path */
#include <linux/spinlock.h>
#include <linux/types.h>
#include <asm/atomic.h>
/*
* IN_* from inotfy.h lines up EXACTLY with FS_*, this is so we can easily
* convert between them. dnotify only needs conversion at watch creation
* so no perf loss there. fanotify isn't defined yet, so it can use the
* wholes if it needs more events.
*/
#define FS_ACCESS 0x00000001 /* File was accessed */
#define FS_MODIFY 0x00000002 /* File was modified */
#define FS_ATTRIB 0x00000004 /* Metadata changed */
#define FS_CLOSE_WRITE 0x00000008 /* Writtable file was closed */
#define FS_CLOSE_NOWRITE 0x00000010 /* Unwrittable file closed */
#define FS_OPEN 0x00000020 /* File was opened */
#define FS_MOVED_FROM 0x00000040 /* File was moved from X */
#define FS_MOVED_TO 0x00000080 /* File was moved to Y */
#define FS_CREATE 0x00000100 /* Subfile was created */
#define FS_DELETE 0x00000200 /* Subfile was deleted */
#define FS_DELETE_SELF 0x00000400 /* Self was deleted */
#define FS_MOVE_SELF 0x00000800 /* Self was moved */
#define FS_UNMOUNT 0x00002000 /* inode on umount fs */
#define FS_Q_OVERFLOW 0x00004000 /* Event queued overflowed */
#define FS_IN_IGNORED 0x00008000 /* last inotify event here */
#define FS_IN_ISDIR 0x40000000 /* event occurred against dir */
#define FS_IN_ONESHOT 0x80000000 /* only send event once */
#define FS_DN_RENAME 0x10000000 /* file renamed */
#define FS_DN_MULTISHOT 0x20000000 /* dnotify multishot */
/* This inode cares about things that happen to its children. Always set for
* dnotify and inotify. */
#define FS_EVENT_ON_CHILD 0x08000000
/* This is a list of all events that may get sent to a parernt based on fs event
* happening to inodes inside that directory */
#define FS_EVENTS_POSS_ON_CHILD (FS_ACCESS | FS_MODIFY | FS_ATTRIB |\
FS_CLOSE_WRITE | FS_CLOSE_NOWRITE | FS_OPEN |\
FS_MOVED_FROM | FS_MOVED_TO | FS_CREATE |\
FS_DELETE)
#define FS_MOVE (FS_MOVED_FROM | FS_MOVED_TO)
fsnotify: unified filesystem notification backend fsnotify is a backend for filesystem notification. fsnotify does not provide any userspace interface but does provide the basis needed for other notification schemes such as dnotify. fsnotify can be extended to be the backend for inotify or the upcoming fanotify. fsnotify provides a mechanism for "groups" to register for some set of filesystem events and to then deliver those events to those groups for processing. fsnotify has a number of benefits, the first being actually shrinking the size of an inode. Before fsnotify to support both dnotify and inotify an inode had unsigned long i_dnotify_mask; /* Directory notify events */ struct dnotify_struct *i_dnotify; /* for directory notifications */ struct list_head inotify_watches; /* watches on this inode */ struct mutex inotify_mutex; /* protects the watches list But with fsnotify this same functionallity (and more) is done with just __u32 i_fsnotify_mask; /* all events for this inode */ struct hlist_head i_fsnotify_mark_entries; /* marks on this inode */ That's right, inotify, dnotify, and fanotify all in 64 bits. We used that much space just in inotify_watches alone, before this patch set. fsnotify object lifetime and locking is MUCH better than what we have today. inotify locking is incredibly complex. See 8f7b0ba1c8539 as an example of what's been busted since inception. inotify needs to know internal semantics of superblock destruction and unmounting to function. The inode pinning and vfs contortions are horrible. no fsnotify implementers do allocation under locks. This means things like f04b30de3 which (due to an overabundance of caution) changes GFP_KERNEL to GFP_NOFS can be reverted. There are no longer any allocation rules when using or implementing your own fsnotify listener. fsnotify paves the way for fanotify. In brief fanotify is a notification mechanism that delivers the lisener both an 'event' and an open file descriptor to the object in question. This means that fanotify is pathname agnostic. Some on lkml may not care for the original companies or users that pushed for TALPA, but fanotify was designed with flexibility and input for other users in mind. The readahead group expressed interest in fanotify as it could be used to profile disk access on boot without breaking the audit system. The desktop search groups have also expressed interest in fanotify as it solves a number of the race conditions and problems present with managing inotify when more than a limited number of specific files are of interest. fanotify can provide for a userspace access control system which makes it a clean interface for AV vendors to hook without trying to do binary patching on the syscall table, LSM, and everywhere else they do their things today. With this patch series fanotify can be implemented in less than 1200 lines of easy to review code. Almost all of which is the socket based user interface. This patch series builds fsnotify to the point that it can implement dnotify and inotify_user. Patches exist and will be sent soon after acceptance to finish the in kernel inotify conversion (audit) and implement fanotify. Signed-off-by: Eric Paris <eparis@redhat.com> Acked-by: Al Viro <viro@zeniv.linux.org.uk> Cc: Christoph Hellwig <hch@lst.de>
2009-05-22 05:01:20 +08:00
struct fsnotify_group;
struct fsnotify_event;
struct fsnotify_mark_entry;
struct fsnotify_event_private_data;
fsnotify: unified filesystem notification backend fsnotify is a backend for filesystem notification. fsnotify does not provide any userspace interface but does provide the basis needed for other notification schemes such as dnotify. fsnotify can be extended to be the backend for inotify or the upcoming fanotify. fsnotify provides a mechanism for "groups" to register for some set of filesystem events and to then deliver those events to those groups for processing. fsnotify has a number of benefits, the first being actually shrinking the size of an inode. Before fsnotify to support both dnotify and inotify an inode had unsigned long i_dnotify_mask; /* Directory notify events */ struct dnotify_struct *i_dnotify; /* for directory notifications */ struct list_head inotify_watches; /* watches on this inode */ struct mutex inotify_mutex; /* protects the watches list But with fsnotify this same functionallity (and more) is done with just __u32 i_fsnotify_mask; /* all events for this inode */ struct hlist_head i_fsnotify_mark_entries; /* marks on this inode */ That's right, inotify, dnotify, and fanotify all in 64 bits. We used that much space just in inotify_watches alone, before this patch set. fsnotify object lifetime and locking is MUCH better than what we have today. inotify locking is incredibly complex. See 8f7b0ba1c8539 as an example of what's been busted since inception. inotify needs to know internal semantics of superblock destruction and unmounting to function. The inode pinning and vfs contortions are horrible. no fsnotify implementers do allocation under locks. This means things like f04b30de3 which (due to an overabundance of caution) changes GFP_KERNEL to GFP_NOFS can be reverted. There are no longer any allocation rules when using or implementing your own fsnotify listener. fsnotify paves the way for fanotify. In brief fanotify is a notification mechanism that delivers the lisener both an 'event' and an open file descriptor to the object in question. This means that fanotify is pathname agnostic. Some on lkml may not care for the original companies or users that pushed for TALPA, but fanotify was designed with flexibility and input for other users in mind. The readahead group expressed interest in fanotify as it could be used to profile disk access on boot without breaking the audit system. The desktop search groups have also expressed interest in fanotify as it solves a number of the race conditions and problems present with managing inotify when more than a limited number of specific files are of interest. fanotify can provide for a userspace access control system which makes it a clean interface for AV vendors to hook without trying to do binary patching on the syscall table, LSM, and everywhere else they do their things today. With this patch series fanotify can be implemented in less than 1200 lines of easy to review code. Almost all of which is the socket based user interface. This patch series builds fsnotify to the point that it can implement dnotify and inotify_user. Patches exist and will be sent soon after acceptance to finish the in kernel inotify conversion (audit) and implement fanotify. Signed-off-by: Eric Paris <eparis@redhat.com> Acked-by: Al Viro <viro@zeniv.linux.org.uk> Cc: Christoph Hellwig <hch@lst.de>
2009-05-22 05:01:20 +08:00
/*
* Each group much define these ops. The fsnotify infrastructure will call
* these operations for each relevant group.
*
* should_send_event - given a group, inode, and mask this function determines
* if the group is interested in this event.
fsnotify: unified filesystem notification backend fsnotify is a backend for filesystem notification. fsnotify does not provide any userspace interface but does provide the basis needed for other notification schemes such as dnotify. fsnotify can be extended to be the backend for inotify or the upcoming fanotify. fsnotify provides a mechanism for "groups" to register for some set of filesystem events and to then deliver those events to those groups for processing. fsnotify has a number of benefits, the first being actually shrinking the size of an inode. Before fsnotify to support both dnotify and inotify an inode had unsigned long i_dnotify_mask; /* Directory notify events */ struct dnotify_struct *i_dnotify; /* for directory notifications */ struct list_head inotify_watches; /* watches on this inode */ struct mutex inotify_mutex; /* protects the watches list But with fsnotify this same functionallity (and more) is done with just __u32 i_fsnotify_mask; /* all events for this inode */ struct hlist_head i_fsnotify_mark_entries; /* marks on this inode */ That's right, inotify, dnotify, and fanotify all in 64 bits. We used that much space just in inotify_watches alone, before this patch set. fsnotify object lifetime and locking is MUCH better than what we have today. inotify locking is incredibly complex. See 8f7b0ba1c8539 as an example of what's been busted since inception. inotify needs to know internal semantics of superblock destruction and unmounting to function. The inode pinning and vfs contortions are horrible. no fsnotify implementers do allocation under locks. This means things like f04b30de3 which (due to an overabundance of caution) changes GFP_KERNEL to GFP_NOFS can be reverted. There are no longer any allocation rules when using or implementing your own fsnotify listener. fsnotify paves the way for fanotify. In brief fanotify is a notification mechanism that delivers the lisener both an 'event' and an open file descriptor to the object in question. This means that fanotify is pathname agnostic. Some on lkml may not care for the original companies or users that pushed for TALPA, but fanotify was designed with flexibility and input for other users in mind. The readahead group expressed interest in fanotify as it could be used to profile disk access on boot without breaking the audit system. The desktop search groups have also expressed interest in fanotify as it solves a number of the race conditions and problems present with managing inotify when more than a limited number of specific files are of interest. fanotify can provide for a userspace access control system which makes it a clean interface for AV vendors to hook without trying to do binary patching on the syscall table, LSM, and everywhere else they do their things today. With this patch series fanotify can be implemented in less than 1200 lines of easy to review code. Almost all of which is the socket based user interface. This patch series builds fsnotify to the point that it can implement dnotify and inotify_user. Patches exist and will be sent soon after acceptance to finish the in kernel inotify conversion (audit) and implement fanotify. Signed-off-by: Eric Paris <eparis@redhat.com> Acked-by: Al Viro <viro@zeniv.linux.org.uk> Cc: Christoph Hellwig <hch@lst.de>
2009-05-22 05:01:20 +08:00
* handle_event - main call for a group to handle an fs event
* free_group_priv - called when a group refcnt hits 0 to clean up the private union
* freeing-mark - this means that a mark has been flagged to die when everything
* finishes using it. The function is supplied with what must be a
* valid group and inode to use to clean up.
fsnotify: unified filesystem notification backend fsnotify is a backend for filesystem notification. fsnotify does not provide any userspace interface but does provide the basis needed for other notification schemes such as dnotify. fsnotify can be extended to be the backend for inotify or the upcoming fanotify. fsnotify provides a mechanism for "groups" to register for some set of filesystem events and to then deliver those events to those groups for processing. fsnotify has a number of benefits, the first being actually shrinking the size of an inode. Before fsnotify to support both dnotify and inotify an inode had unsigned long i_dnotify_mask; /* Directory notify events */ struct dnotify_struct *i_dnotify; /* for directory notifications */ struct list_head inotify_watches; /* watches on this inode */ struct mutex inotify_mutex; /* protects the watches list But with fsnotify this same functionallity (and more) is done with just __u32 i_fsnotify_mask; /* all events for this inode */ struct hlist_head i_fsnotify_mark_entries; /* marks on this inode */ That's right, inotify, dnotify, and fanotify all in 64 bits. We used that much space just in inotify_watches alone, before this patch set. fsnotify object lifetime and locking is MUCH better than what we have today. inotify locking is incredibly complex. See 8f7b0ba1c8539 as an example of what's been busted since inception. inotify needs to know internal semantics of superblock destruction and unmounting to function. The inode pinning and vfs contortions are horrible. no fsnotify implementers do allocation under locks. This means things like f04b30de3 which (due to an overabundance of caution) changes GFP_KERNEL to GFP_NOFS can be reverted. There are no longer any allocation rules when using or implementing your own fsnotify listener. fsnotify paves the way for fanotify. In brief fanotify is a notification mechanism that delivers the lisener both an 'event' and an open file descriptor to the object in question. This means that fanotify is pathname agnostic. Some on lkml may not care for the original companies or users that pushed for TALPA, but fanotify was designed with flexibility and input for other users in mind. The readahead group expressed interest in fanotify as it could be used to profile disk access on boot without breaking the audit system. The desktop search groups have also expressed interest in fanotify as it solves a number of the race conditions and problems present with managing inotify when more than a limited number of specific files are of interest. fanotify can provide for a userspace access control system which makes it a clean interface for AV vendors to hook without trying to do binary patching on the syscall table, LSM, and everywhere else they do their things today. With this patch series fanotify can be implemented in less than 1200 lines of easy to review code. Almost all of which is the socket based user interface. This patch series builds fsnotify to the point that it can implement dnotify and inotify_user. Patches exist and will be sent soon after acceptance to finish the in kernel inotify conversion (audit) and implement fanotify. Signed-off-by: Eric Paris <eparis@redhat.com> Acked-by: Al Viro <viro@zeniv.linux.org.uk> Cc: Christoph Hellwig <hch@lst.de>
2009-05-22 05:01:20 +08:00
*/
struct fsnotify_ops {
bool (*should_send_event)(struct fsnotify_group *group, struct inode *inode,
__u32 mask, void *data, int data_type);
fsnotify: unified filesystem notification backend fsnotify is a backend for filesystem notification. fsnotify does not provide any userspace interface but does provide the basis needed for other notification schemes such as dnotify. fsnotify can be extended to be the backend for inotify or the upcoming fanotify. fsnotify provides a mechanism for "groups" to register for some set of filesystem events and to then deliver those events to those groups for processing. fsnotify has a number of benefits, the first being actually shrinking the size of an inode. Before fsnotify to support both dnotify and inotify an inode had unsigned long i_dnotify_mask; /* Directory notify events */ struct dnotify_struct *i_dnotify; /* for directory notifications */ struct list_head inotify_watches; /* watches on this inode */ struct mutex inotify_mutex; /* protects the watches list But with fsnotify this same functionallity (and more) is done with just __u32 i_fsnotify_mask; /* all events for this inode */ struct hlist_head i_fsnotify_mark_entries; /* marks on this inode */ That's right, inotify, dnotify, and fanotify all in 64 bits. We used that much space just in inotify_watches alone, before this patch set. fsnotify object lifetime and locking is MUCH better than what we have today. inotify locking is incredibly complex. See 8f7b0ba1c8539 as an example of what's been busted since inception. inotify needs to know internal semantics of superblock destruction and unmounting to function. The inode pinning and vfs contortions are horrible. no fsnotify implementers do allocation under locks. This means things like f04b30de3 which (due to an overabundance of caution) changes GFP_KERNEL to GFP_NOFS can be reverted. There are no longer any allocation rules when using or implementing your own fsnotify listener. fsnotify paves the way for fanotify. In brief fanotify is a notification mechanism that delivers the lisener both an 'event' and an open file descriptor to the object in question. This means that fanotify is pathname agnostic. Some on lkml may not care for the original companies or users that pushed for TALPA, but fanotify was designed with flexibility and input for other users in mind. The readahead group expressed interest in fanotify as it could be used to profile disk access on boot without breaking the audit system. The desktop search groups have also expressed interest in fanotify as it solves a number of the race conditions and problems present with managing inotify when more than a limited number of specific files are of interest. fanotify can provide for a userspace access control system which makes it a clean interface for AV vendors to hook without trying to do binary patching on the syscall table, LSM, and everywhere else they do their things today. With this patch series fanotify can be implemented in less than 1200 lines of easy to review code. Almost all of which is the socket based user interface. This patch series builds fsnotify to the point that it can implement dnotify and inotify_user. Patches exist and will be sent soon after acceptance to finish the in kernel inotify conversion (audit) and implement fanotify. Signed-off-by: Eric Paris <eparis@redhat.com> Acked-by: Al Viro <viro@zeniv.linux.org.uk> Cc: Christoph Hellwig <hch@lst.de>
2009-05-22 05:01:20 +08:00
int (*handle_event)(struct fsnotify_group *group, struct fsnotify_event *event);
void (*free_group_priv)(struct fsnotify_group *group);
void (*freeing_mark)(struct fsnotify_mark_entry *entry, struct fsnotify_group *group);
void (*free_event_priv)(struct fsnotify_event_private_data *priv);
fsnotify: unified filesystem notification backend fsnotify is a backend for filesystem notification. fsnotify does not provide any userspace interface but does provide the basis needed for other notification schemes such as dnotify. fsnotify can be extended to be the backend for inotify or the upcoming fanotify. fsnotify provides a mechanism for "groups" to register for some set of filesystem events and to then deliver those events to those groups for processing. fsnotify has a number of benefits, the first being actually shrinking the size of an inode. Before fsnotify to support both dnotify and inotify an inode had unsigned long i_dnotify_mask; /* Directory notify events */ struct dnotify_struct *i_dnotify; /* for directory notifications */ struct list_head inotify_watches; /* watches on this inode */ struct mutex inotify_mutex; /* protects the watches list But with fsnotify this same functionallity (and more) is done with just __u32 i_fsnotify_mask; /* all events for this inode */ struct hlist_head i_fsnotify_mark_entries; /* marks on this inode */ That's right, inotify, dnotify, and fanotify all in 64 bits. We used that much space just in inotify_watches alone, before this patch set. fsnotify object lifetime and locking is MUCH better than what we have today. inotify locking is incredibly complex. See 8f7b0ba1c8539 as an example of what's been busted since inception. inotify needs to know internal semantics of superblock destruction and unmounting to function. The inode pinning and vfs contortions are horrible. no fsnotify implementers do allocation under locks. This means things like f04b30de3 which (due to an overabundance of caution) changes GFP_KERNEL to GFP_NOFS can be reverted. There are no longer any allocation rules when using or implementing your own fsnotify listener. fsnotify paves the way for fanotify. In brief fanotify is a notification mechanism that delivers the lisener both an 'event' and an open file descriptor to the object in question. This means that fanotify is pathname agnostic. Some on lkml may not care for the original companies or users that pushed for TALPA, but fanotify was designed with flexibility and input for other users in mind. The readahead group expressed interest in fanotify as it could be used to profile disk access on boot without breaking the audit system. The desktop search groups have also expressed interest in fanotify as it solves a number of the race conditions and problems present with managing inotify when more than a limited number of specific files are of interest. fanotify can provide for a userspace access control system which makes it a clean interface for AV vendors to hook without trying to do binary patching on the syscall table, LSM, and everywhere else they do their things today. With this patch series fanotify can be implemented in less than 1200 lines of easy to review code. Almost all of which is the socket based user interface. This patch series builds fsnotify to the point that it can implement dnotify and inotify_user. Patches exist and will be sent soon after acceptance to finish the in kernel inotify conversion (audit) and implement fanotify. Signed-off-by: Eric Paris <eparis@redhat.com> Acked-by: Al Viro <viro@zeniv.linux.org.uk> Cc: Christoph Hellwig <hch@lst.de>
2009-05-22 05:01:20 +08:00
};
/*
* A group is a "thing" that wants to receive notification about filesystem
* events. The mask holds the subset of event types this group cares about.
* refcnt on a group is up to the implementor and at any moment if it goes 0
* everything will be cleaned up.
*/
struct fsnotify_group {
/*
* global list of all groups receiving events from fsnotify.
* anchored by fsnotify_groups and protected by either fsnotify_grp_mutex
* or fsnotify_grp_srcu depending on write vs read.
*/
struct list_head group_list;
/*
* Defines all of the event types in which this group is interested.
* This mask is a bitwise OR of the FS_* events from above. Each time
* this mask changes for a group (if it changes) the correct functions
* must be called to update the global structures which indicate global
* interest in event types.
*/
__u32 mask;
/*
* How the refcnt is used is up to each group. When the refcnt hits 0
* fsnotify will clean up all of the resources associated with this group.
* As an example, the dnotify group will always have a refcnt=1 and that
* will never change. Inotify, on the other hand, has a group per
* inotify_init() and the refcnt will hit 0 only when that fd has been
* closed.
*/
atomic_t refcnt; /* things with interest in this group */
const struct fsnotify_ops *ops; /* how this group handles things */
/* needed to send notification to userspace */
struct mutex notification_mutex; /* protect the notification_list */
struct list_head notification_list; /* list of event_holder this group needs to send to userspace */
wait_queue_head_t notification_waitq; /* read() on the notification file blocks on this waitq */
unsigned int q_len; /* events on the queue */
unsigned int max_events; /* maximum events allowed on the list */
/* stores all fastapth entries assoc with this group so they can be cleaned on unregister */
spinlock_t mark_lock; /* protect mark_entries list */
atomic_t num_marks; /* 1 for each mark entry and 1 for not being
* past the point of no return when freeing
* a group */
struct list_head mark_entries; /* all inode mark entries for this group */
/* prevents double list_del of group_list. protected by global fsnotify_grp_mutex */
fsnotify: unified filesystem notification backend fsnotify is a backend for filesystem notification. fsnotify does not provide any userspace interface but does provide the basis needed for other notification schemes such as dnotify. fsnotify can be extended to be the backend for inotify or the upcoming fanotify. fsnotify provides a mechanism for "groups" to register for some set of filesystem events and to then deliver those events to those groups for processing. fsnotify has a number of benefits, the first being actually shrinking the size of an inode. Before fsnotify to support both dnotify and inotify an inode had unsigned long i_dnotify_mask; /* Directory notify events */ struct dnotify_struct *i_dnotify; /* for directory notifications */ struct list_head inotify_watches; /* watches on this inode */ struct mutex inotify_mutex; /* protects the watches list But with fsnotify this same functionallity (and more) is done with just __u32 i_fsnotify_mask; /* all events for this inode */ struct hlist_head i_fsnotify_mark_entries; /* marks on this inode */ That's right, inotify, dnotify, and fanotify all in 64 bits. We used that much space just in inotify_watches alone, before this patch set. fsnotify object lifetime and locking is MUCH better than what we have today. inotify locking is incredibly complex. See 8f7b0ba1c8539 as an example of what's been busted since inception. inotify needs to know internal semantics of superblock destruction and unmounting to function. The inode pinning and vfs contortions are horrible. no fsnotify implementers do allocation under locks. This means things like f04b30de3 which (due to an overabundance of caution) changes GFP_KERNEL to GFP_NOFS can be reverted. There are no longer any allocation rules when using or implementing your own fsnotify listener. fsnotify paves the way for fanotify. In brief fanotify is a notification mechanism that delivers the lisener both an 'event' and an open file descriptor to the object in question. This means that fanotify is pathname agnostic. Some on lkml may not care for the original companies or users that pushed for TALPA, but fanotify was designed with flexibility and input for other users in mind. The readahead group expressed interest in fanotify as it could be used to profile disk access on boot without breaking the audit system. The desktop search groups have also expressed interest in fanotify as it solves a number of the race conditions and problems present with managing inotify when more than a limited number of specific files are of interest. fanotify can provide for a userspace access control system which makes it a clean interface for AV vendors to hook without trying to do binary patching on the syscall table, LSM, and everywhere else they do their things today. With this patch series fanotify can be implemented in less than 1200 lines of easy to review code. Almost all of which is the socket based user interface. This patch series builds fsnotify to the point that it can implement dnotify and inotify_user. Patches exist and will be sent soon after acceptance to finish the in kernel inotify conversion (audit) and implement fanotify. Signed-off-by: Eric Paris <eparis@redhat.com> Acked-by: Al Viro <viro@zeniv.linux.org.uk> Cc: Christoph Hellwig <hch@lst.de>
2009-05-22 05:01:20 +08:00
bool on_group_list;
/* groups can define private fields here or use the void *private */
union {
void *private;
#ifdef CONFIG_INOTIFY_USER
struct inotify_group_private_data {
spinlock_t idr_lock;
struct idr idr;
u32 last_wd;
struct fasync_struct *fa; /* async notification */
struct user_struct *user;
} inotify_data;
#endif
fsnotify: unified filesystem notification backend fsnotify is a backend for filesystem notification. fsnotify does not provide any userspace interface but does provide the basis needed for other notification schemes such as dnotify. fsnotify can be extended to be the backend for inotify or the upcoming fanotify. fsnotify provides a mechanism for "groups" to register for some set of filesystem events and to then deliver those events to those groups for processing. fsnotify has a number of benefits, the first being actually shrinking the size of an inode. Before fsnotify to support both dnotify and inotify an inode had unsigned long i_dnotify_mask; /* Directory notify events */ struct dnotify_struct *i_dnotify; /* for directory notifications */ struct list_head inotify_watches; /* watches on this inode */ struct mutex inotify_mutex; /* protects the watches list But with fsnotify this same functionallity (and more) is done with just __u32 i_fsnotify_mask; /* all events for this inode */ struct hlist_head i_fsnotify_mark_entries; /* marks on this inode */ That's right, inotify, dnotify, and fanotify all in 64 bits. We used that much space just in inotify_watches alone, before this patch set. fsnotify object lifetime and locking is MUCH better than what we have today. inotify locking is incredibly complex. See 8f7b0ba1c8539 as an example of what's been busted since inception. inotify needs to know internal semantics of superblock destruction and unmounting to function. The inode pinning and vfs contortions are horrible. no fsnotify implementers do allocation under locks. This means things like f04b30de3 which (due to an overabundance of caution) changes GFP_KERNEL to GFP_NOFS can be reverted. There are no longer any allocation rules when using or implementing your own fsnotify listener. fsnotify paves the way for fanotify. In brief fanotify is a notification mechanism that delivers the lisener both an 'event' and an open file descriptor to the object in question. This means that fanotify is pathname agnostic. Some on lkml may not care for the original companies or users that pushed for TALPA, but fanotify was designed with flexibility and input for other users in mind. The readahead group expressed interest in fanotify as it could be used to profile disk access on boot without breaking the audit system. The desktop search groups have also expressed interest in fanotify as it solves a number of the race conditions and problems present with managing inotify when more than a limited number of specific files are of interest. fanotify can provide for a userspace access control system which makes it a clean interface for AV vendors to hook without trying to do binary patching on the syscall table, LSM, and everywhere else they do their things today. With this patch series fanotify can be implemented in less than 1200 lines of easy to review code. Almost all of which is the socket based user interface. This patch series builds fsnotify to the point that it can implement dnotify and inotify_user. Patches exist and will be sent soon after acceptance to finish the in kernel inotify conversion (audit) and implement fanotify. Signed-off-by: Eric Paris <eparis@redhat.com> Acked-by: Al Viro <viro@zeniv.linux.org.uk> Cc: Christoph Hellwig <hch@lst.de>
2009-05-22 05:01:20 +08:00
};
};
/*
* A single event can be queued in multiple group->notification_lists.
*
* each group->notification_list will point to an event_holder which in turns points
* to the actual event that needs to be sent to userspace.
*
* Seemed cheaper to create a refcnt'd event and a small holder for every group
* than create a different event for every group
*
*/
struct fsnotify_event_holder {
struct fsnotify_event *event;
struct list_head event_list;
};
/*
* Inotify needs to tack data onto an event. This struct lets us later find the
* correct private data of the correct group.
*/
struct fsnotify_event_private_data {
struct fsnotify_group *group;
struct list_head event_list;
};
fsnotify: unified filesystem notification backend fsnotify is a backend for filesystem notification. fsnotify does not provide any userspace interface but does provide the basis needed for other notification schemes such as dnotify. fsnotify can be extended to be the backend for inotify or the upcoming fanotify. fsnotify provides a mechanism for "groups" to register for some set of filesystem events and to then deliver those events to those groups for processing. fsnotify has a number of benefits, the first being actually shrinking the size of an inode. Before fsnotify to support both dnotify and inotify an inode had unsigned long i_dnotify_mask; /* Directory notify events */ struct dnotify_struct *i_dnotify; /* for directory notifications */ struct list_head inotify_watches; /* watches on this inode */ struct mutex inotify_mutex; /* protects the watches list But with fsnotify this same functionallity (and more) is done with just __u32 i_fsnotify_mask; /* all events for this inode */ struct hlist_head i_fsnotify_mark_entries; /* marks on this inode */ That's right, inotify, dnotify, and fanotify all in 64 bits. We used that much space just in inotify_watches alone, before this patch set. fsnotify object lifetime and locking is MUCH better than what we have today. inotify locking is incredibly complex. See 8f7b0ba1c8539 as an example of what's been busted since inception. inotify needs to know internal semantics of superblock destruction and unmounting to function. The inode pinning and vfs contortions are horrible. no fsnotify implementers do allocation under locks. This means things like f04b30de3 which (due to an overabundance of caution) changes GFP_KERNEL to GFP_NOFS can be reverted. There are no longer any allocation rules when using or implementing your own fsnotify listener. fsnotify paves the way for fanotify. In brief fanotify is a notification mechanism that delivers the lisener both an 'event' and an open file descriptor to the object in question. This means that fanotify is pathname agnostic. Some on lkml may not care for the original companies or users that pushed for TALPA, but fanotify was designed with flexibility and input for other users in mind. The readahead group expressed interest in fanotify as it could be used to profile disk access on boot without breaking the audit system. The desktop search groups have also expressed interest in fanotify as it solves a number of the race conditions and problems present with managing inotify when more than a limited number of specific files are of interest. fanotify can provide for a userspace access control system which makes it a clean interface for AV vendors to hook without trying to do binary patching on the syscall table, LSM, and everywhere else they do their things today. With this patch series fanotify can be implemented in less than 1200 lines of easy to review code. Almost all of which is the socket based user interface. This patch series builds fsnotify to the point that it can implement dnotify and inotify_user. Patches exist and will be sent soon after acceptance to finish the in kernel inotify conversion (audit) and implement fanotify. Signed-off-by: Eric Paris <eparis@redhat.com> Acked-by: Al Viro <viro@zeniv.linux.org.uk> Cc: Christoph Hellwig <hch@lst.de>
2009-05-22 05:01:20 +08:00
/*
* all of the information about the original object we want to now send to
* a group. If you want to carry more info from the accessing task to the
* listener this structure is where you need to be adding fields.
*/
struct fsnotify_event {
/*
* If we create an event we are also likely going to need a holder
* to link to a group. So embed one holder in the event. Means only
* one allocation for the common case where we only have one group
*/
struct fsnotify_event_holder holder;
fsnotify: unified filesystem notification backend fsnotify is a backend for filesystem notification. fsnotify does not provide any userspace interface but does provide the basis needed for other notification schemes such as dnotify. fsnotify can be extended to be the backend for inotify or the upcoming fanotify. fsnotify provides a mechanism for "groups" to register for some set of filesystem events and to then deliver those events to those groups for processing. fsnotify has a number of benefits, the first being actually shrinking the size of an inode. Before fsnotify to support both dnotify and inotify an inode had unsigned long i_dnotify_mask; /* Directory notify events */ struct dnotify_struct *i_dnotify; /* for directory notifications */ struct list_head inotify_watches; /* watches on this inode */ struct mutex inotify_mutex; /* protects the watches list But with fsnotify this same functionallity (and more) is done with just __u32 i_fsnotify_mask; /* all events for this inode */ struct hlist_head i_fsnotify_mark_entries; /* marks on this inode */ That's right, inotify, dnotify, and fanotify all in 64 bits. We used that much space just in inotify_watches alone, before this patch set. fsnotify object lifetime and locking is MUCH better than what we have today. inotify locking is incredibly complex. See 8f7b0ba1c8539 as an example of what's been busted since inception. inotify needs to know internal semantics of superblock destruction and unmounting to function. The inode pinning and vfs contortions are horrible. no fsnotify implementers do allocation under locks. This means things like f04b30de3 which (due to an overabundance of caution) changes GFP_KERNEL to GFP_NOFS can be reverted. There are no longer any allocation rules when using or implementing your own fsnotify listener. fsnotify paves the way for fanotify. In brief fanotify is a notification mechanism that delivers the lisener both an 'event' and an open file descriptor to the object in question. This means that fanotify is pathname agnostic. Some on lkml may not care for the original companies or users that pushed for TALPA, but fanotify was designed with flexibility and input for other users in mind. The readahead group expressed interest in fanotify as it could be used to profile disk access on boot without breaking the audit system. The desktop search groups have also expressed interest in fanotify as it solves a number of the race conditions and problems present with managing inotify when more than a limited number of specific files are of interest. fanotify can provide for a userspace access control system which makes it a clean interface for AV vendors to hook without trying to do binary patching on the syscall table, LSM, and everywhere else they do their things today. With this patch series fanotify can be implemented in less than 1200 lines of easy to review code. Almost all of which is the socket based user interface. This patch series builds fsnotify to the point that it can implement dnotify and inotify_user. Patches exist and will be sent soon after acceptance to finish the in kernel inotify conversion (audit) and implement fanotify. Signed-off-by: Eric Paris <eparis@redhat.com> Acked-by: Al Viro <viro@zeniv.linux.org.uk> Cc: Christoph Hellwig <hch@lst.de>
2009-05-22 05:01:20 +08:00
spinlock_t lock; /* protection for the associated event_holder and private_list */
/* to_tell may ONLY be dereferenced during handle_event(). */
struct inode *to_tell; /* either the inode the event happened to or its parent */
/*
* depending on the event type we should have either a path or inode
* We hold a reference on path, but NOT on inode. Since we have the ref on
* the path, it may be dereferenced at any point during this object's
* lifetime. That reference is dropped when this object's refcnt hits
* 0. If this event contains an inode instead of a path, the inode may
* ONLY be used during handle_event().
*/
union {
struct path path;
struct inode *inode;
};
/* when calling fsnotify tell it if the data is a path or inode */
#define FSNOTIFY_EVENT_NONE 0
#define FSNOTIFY_EVENT_PATH 1
#define FSNOTIFY_EVENT_INODE 2
#define FSNOTIFY_EVENT_FILE 3
int data_type; /* which of the above union we have */
atomic_t refcnt; /* how many groups still are using/need to send this event */
__u32 mask; /* the type of access, bitwise OR for FS_* event types */
u32 sync_cookie; /* used to corrolate events, namely inotify mv events */
char *file_name;
size_t name_len;
struct list_head private_data_list; /* groups can store private data here */
fsnotify: unified filesystem notification backend fsnotify is a backend for filesystem notification. fsnotify does not provide any userspace interface but does provide the basis needed for other notification schemes such as dnotify. fsnotify can be extended to be the backend for inotify or the upcoming fanotify. fsnotify provides a mechanism for "groups" to register for some set of filesystem events and to then deliver those events to those groups for processing. fsnotify has a number of benefits, the first being actually shrinking the size of an inode. Before fsnotify to support both dnotify and inotify an inode had unsigned long i_dnotify_mask; /* Directory notify events */ struct dnotify_struct *i_dnotify; /* for directory notifications */ struct list_head inotify_watches; /* watches on this inode */ struct mutex inotify_mutex; /* protects the watches list But with fsnotify this same functionallity (and more) is done with just __u32 i_fsnotify_mask; /* all events for this inode */ struct hlist_head i_fsnotify_mark_entries; /* marks on this inode */ That's right, inotify, dnotify, and fanotify all in 64 bits. We used that much space just in inotify_watches alone, before this patch set. fsnotify object lifetime and locking is MUCH better than what we have today. inotify locking is incredibly complex. See 8f7b0ba1c8539 as an example of what's been busted since inception. inotify needs to know internal semantics of superblock destruction and unmounting to function. The inode pinning and vfs contortions are horrible. no fsnotify implementers do allocation under locks. This means things like f04b30de3 which (due to an overabundance of caution) changes GFP_KERNEL to GFP_NOFS can be reverted. There are no longer any allocation rules when using or implementing your own fsnotify listener. fsnotify paves the way for fanotify. In brief fanotify is a notification mechanism that delivers the lisener both an 'event' and an open file descriptor to the object in question. This means that fanotify is pathname agnostic. Some on lkml may not care for the original companies or users that pushed for TALPA, but fanotify was designed with flexibility and input for other users in mind. The readahead group expressed interest in fanotify as it could be used to profile disk access on boot without breaking the audit system. The desktop search groups have also expressed interest in fanotify as it solves a number of the race conditions and problems present with managing inotify when more than a limited number of specific files are of interest. fanotify can provide for a userspace access control system which makes it a clean interface for AV vendors to hook without trying to do binary patching on the syscall table, LSM, and everywhere else they do their things today. With this patch series fanotify can be implemented in less than 1200 lines of easy to review code. Almost all of which is the socket based user interface. This patch series builds fsnotify to the point that it can implement dnotify and inotify_user. Patches exist and will be sent soon after acceptance to finish the in kernel inotify conversion (audit) and implement fanotify. Signed-off-by: Eric Paris <eparis@redhat.com> Acked-by: Al Viro <viro@zeniv.linux.org.uk> Cc: Christoph Hellwig <hch@lst.de>
2009-05-22 05:01:20 +08:00
};
/*
* a mark is simply an entry attached to an in core inode which allows an
* fsnotify listener to indicate they are either no longer interested in events
* of a type matching mask or only interested in those events.
*
* these are flushed when an inode is evicted from core and may be flushed
* when the inode is modified (as seen by fsnotify_access). Some fsnotify users
* (such as dnotify) will flush these when the open fd is closed and not at
* inode eviction or modification.
*/
struct fsnotify_mark_entry {
__u32 mask; /* mask this mark entry is for */
/* we hold ref for each i_list and g_list. also one ref for each 'thing'
* in kernel that found and may be using this mark. */
atomic_t refcnt; /* active things looking at this mark */
struct inode *inode; /* inode this entry is associated with */
struct fsnotify_group *group; /* group this mark entry is for */
struct hlist_node i_list; /* list of mark_entries by inode->i_fsnotify_mark_entries */
struct list_head g_list; /* list of mark_entries by group->i_fsnotify_mark_entries */
spinlock_t lock; /* protect group, inode, and killme */
struct list_head free_i_list; /* tmp list used when freeing this mark */
struct list_head free_g_list; /* tmp list used when freeing this mark */
void (*free_mark)(struct fsnotify_mark_entry *entry); /* called on final put+free */
};
fsnotify: unified filesystem notification backend fsnotify is a backend for filesystem notification. fsnotify does not provide any userspace interface but does provide the basis needed for other notification schemes such as dnotify. fsnotify can be extended to be the backend for inotify or the upcoming fanotify. fsnotify provides a mechanism for "groups" to register for some set of filesystem events and to then deliver those events to those groups for processing. fsnotify has a number of benefits, the first being actually shrinking the size of an inode. Before fsnotify to support both dnotify and inotify an inode had unsigned long i_dnotify_mask; /* Directory notify events */ struct dnotify_struct *i_dnotify; /* for directory notifications */ struct list_head inotify_watches; /* watches on this inode */ struct mutex inotify_mutex; /* protects the watches list But with fsnotify this same functionallity (and more) is done with just __u32 i_fsnotify_mask; /* all events for this inode */ struct hlist_head i_fsnotify_mark_entries; /* marks on this inode */ That's right, inotify, dnotify, and fanotify all in 64 bits. We used that much space just in inotify_watches alone, before this patch set. fsnotify object lifetime and locking is MUCH better than what we have today. inotify locking is incredibly complex. See 8f7b0ba1c8539 as an example of what's been busted since inception. inotify needs to know internal semantics of superblock destruction and unmounting to function. The inode pinning and vfs contortions are horrible. no fsnotify implementers do allocation under locks. This means things like f04b30de3 which (due to an overabundance of caution) changes GFP_KERNEL to GFP_NOFS can be reverted. There are no longer any allocation rules when using or implementing your own fsnotify listener. fsnotify paves the way for fanotify. In brief fanotify is a notification mechanism that delivers the lisener both an 'event' and an open file descriptor to the object in question. This means that fanotify is pathname agnostic. Some on lkml may not care for the original companies or users that pushed for TALPA, but fanotify was designed with flexibility and input for other users in mind. The readahead group expressed interest in fanotify as it could be used to profile disk access on boot without breaking the audit system. The desktop search groups have also expressed interest in fanotify as it solves a number of the race conditions and problems present with managing inotify when more than a limited number of specific files are of interest. fanotify can provide for a userspace access control system which makes it a clean interface for AV vendors to hook without trying to do binary patching on the syscall table, LSM, and everywhere else they do their things today. With this patch series fanotify can be implemented in less than 1200 lines of easy to review code. Almost all of which is the socket based user interface. This patch series builds fsnotify to the point that it can implement dnotify and inotify_user. Patches exist and will be sent soon after acceptance to finish the in kernel inotify conversion (audit) and implement fanotify. Signed-off-by: Eric Paris <eparis@redhat.com> Acked-by: Al Viro <viro@zeniv.linux.org.uk> Cc: Christoph Hellwig <hch@lst.de>
2009-05-22 05:01:20 +08:00
#ifdef CONFIG_FSNOTIFY
/* called from the vfs helpers */
/* main fsnotify call to send events */
extern void fsnotify(struct inode *to_tell, __u32 mask, void *data, int data_is,
const char *name, u32 cookie);
extern void __fsnotify_parent(struct file *file, struct dentry *dentry, __u32 mask);
extern void __fsnotify_inode_delete(struct inode *inode);
extern u32 fsnotify_get_cookie(void);
fsnotify: unified filesystem notification backend fsnotify is a backend for filesystem notification. fsnotify does not provide any userspace interface but does provide the basis needed for other notification schemes such as dnotify. fsnotify can be extended to be the backend for inotify or the upcoming fanotify. fsnotify provides a mechanism for "groups" to register for some set of filesystem events and to then deliver those events to those groups for processing. fsnotify has a number of benefits, the first being actually shrinking the size of an inode. Before fsnotify to support both dnotify and inotify an inode had unsigned long i_dnotify_mask; /* Directory notify events */ struct dnotify_struct *i_dnotify; /* for directory notifications */ struct list_head inotify_watches; /* watches on this inode */ struct mutex inotify_mutex; /* protects the watches list But with fsnotify this same functionallity (and more) is done with just __u32 i_fsnotify_mask; /* all events for this inode */ struct hlist_head i_fsnotify_mark_entries; /* marks on this inode */ That's right, inotify, dnotify, and fanotify all in 64 bits. We used that much space just in inotify_watches alone, before this patch set. fsnotify object lifetime and locking is MUCH better than what we have today. inotify locking is incredibly complex. See 8f7b0ba1c8539 as an example of what's been busted since inception. inotify needs to know internal semantics of superblock destruction and unmounting to function. The inode pinning and vfs contortions are horrible. no fsnotify implementers do allocation under locks. This means things like f04b30de3 which (due to an overabundance of caution) changes GFP_KERNEL to GFP_NOFS can be reverted. There are no longer any allocation rules when using or implementing your own fsnotify listener. fsnotify paves the way for fanotify. In brief fanotify is a notification mechanism that delivers the lisener both an 'event' and an open file descriptor to the object in question. This means that fanotify is pathname agnostic. Some on lkml may not care for the original companies or users that pushed for TALPA, but fanotify was designed with flexibility and input for other users in mind. The readahead group expressed interest in fanotify as it could be used to profile disk access on boot without breaking the audit system. The desktop search groups have also expressed interest in fanotify as it solves a number of the race conditions and problems present with managing inotify when more than a limited number of specific files are of interest. fanotify can provide for a userspace access control system which makes it a clean interface for AV vendors to hook without trying to do binary patching on the syscall table, LSM, and everywhere else they do their things today. With this patch series fanotify can be implemented in less than 1200 lines of easy to review code. Almost all of which is the socket based user interface. This patch series builds fsnotify to the point that it can implement dnotify and inotify_user. Patches exist and will be sent soon after acceptance to finish the in kernel inotify conversion (audit) and implement fanotify. Signed-off-by: Eric Paris <eparis@redhat.com> Acked-by: Al Viro <viro@zeniv.linux.org.uk> Cc: Christoph Hellwig <hch@lst.de>
2009-05-22 05:01:20 +08:00
static inline int fsnotify_inode_watches_children(struct inode *inode)
{
/* FS_EVENT_ON_CHILD is set if the inode may care */
if (!(inode->i_fsnotify_mask & FS_EVENT_ON_CHILD))
return 0;
/* this inode might care about child events, does it care about the
* specific set of events that can happen on a child? */
return inode->i_fsnotify_mask & FS_EVENTS_POSS_ON_CHILD;
}
/*
* Update the dentry with a flag indicating the interest of its parent to receive
* filesystem events when those events happens to this dentry->d_inode.
*/
static inline void __fsnotify_update_dcache_flags(struct dentry *dentry)
{
struct dentry *parent;
assert_spin_locked(&dcache_lock);
assert_spin_locked(&dentry->d_lock);
parent = dentry->d_parent;
if (parent->d_inode && fsnotify_inode_watches_children(parent->d_inode))
dentry->d_flags |= DCACHE_FSNOTIFY_PARENT_WATCHED;
else
dentry->d_flags &= ~DCACHE_FSNOTIFY_PARENT_WATCHED;
}
/*
* fsnotify_d_instantiate - instantiate a dentry for inode
* Called with dcache_lock held.
*/
static inline void __fsnotify_d_instantiate(struct dentry *dentry, struct inode *inode)
{
if (!inode)
return;
assert_spin_locked(&dcache_lock);
spin_lock(&dentry->d_lock);
__fsnotify_update_dcache_flags(dentry);
spin_unlock(&dentry->d_lock);
}
fsnotify: unified filesystem notification backend fsnotify is a backend for filesystem notification. fsnotify does not provide any userspace interface but does provide the basis needed for other notification schemes such as dnotify. fsnotify can be extended to be the backend for inotify or the upcoming fanotify. fsnotify provides a mechanism for "groups" to register for some set of filesystem events and to then deliver those events to those groups for processing. fsnotify has a number of benefits, the first being actually shrinking the size of an inode. Before fsnotify to support both dnotify and inotify an inode had unsigned long i_dnotify_mask; /* Directory notify events */ struct dnotify_struct *i_dnotify; /* for directory notifications */ struct list_head inotify_watches; /* watches on this inode */ struct mutex inotify_mutex; /* protects the watches list But with fsnotify this same functionallity (and more) is done with just __u32 i_fsnotify_mask; /* all events for this inode */ struct hlist_head i_fsnotify_mark_entries; /* marks on this inode */ That's right, inotify, dnotify, and fanotify all in 64 bits. We used that much space just in inotify_watches alone, before this patch set. fsnotify object lifetime and locking is MUCH better than what we have today. inotify locking is incredibly complex. See 8f7b0ba1c8539 as an example of what's been busted since inception. inotify needs to know internal semantics of superblock destruction and unmounting to function. The inode pinning and vfs contortions are horrible. no fsnotify implementers do allocation under locks. This means things like f04b30de3 which (due to an overabundance of caution) changes GFP_KERNEL to GFP_NOFS can be reverted. There are no longer any allocation rules when using or implementing your own fsnotify listener. fsnotify paves the way for fanotify. In brief fanotify is a notification mechanism that delivers the lisener both an 'event' and an open file descriptor to the object in question. This means that fanotify is pathname agnostic. Some on lkml may not care for the original companies or users that pushed for TALPA, but fanotify was designed with flexibility and input for other users in mind. The readahead group expressed interest in fanotify as it could be used to profile disk access on boot without breaking the audit system. The desktop search groups have also expressed interest in fanotify as it solves a number of the race conditions and problems present with managing inotify when more than a limited number of specific files are of interest. fanotify can provide for a userspace access control system which makes it a clean interface for AV vendors to hook without trying to do binary patching on the syscall table, LSM, and everywhere else they do their things today. With this patch series fanotify can be implemented in less than 1200 lines of easy to review code. Almost all of which is the socket based user interface. This patch series builds fsnotify to the point that it can implement dnotify and inotify_user. Patches exist and will be sent soon after acceptance to finish the in kernel inotify conversion (audit) and implement fanotify. Signed-off-by: Eric Paris <eparis@redhat.com> Acked-by: Al Viro <viro@zeniv.linux.org.uk> Cc: Christoph Hellwig <hch@lst.de>
2009-05-22 05:01:20 +08:00
/* called from fsnotify listeners, such as fanotify or dnotify */
/* must call when a group changes its ->mask */
extern void fsnotify_recalc_global_mask(void);
/* get a reference to an existing or create a new group */
extern struct fsnotify_group *fsnotify_alloc_group(const struct fsnotify_ops *ops);
/* run all marks associated with this group and update group->mask */
extern void fsnotify_recalc_group_mask(struct fsnotify_group *group);
/* drop reference on a group from fsnotify_alloc_group */
fsnotify: unified filesystem notification backend fsnotify is a backend for filesystem notification. fsnotify does not provide any userspace interface but does provide the basis needed for other notification schemes such as dnotify. fsnotify can be extended to be the backend for inotify or the upcoming fanotify. fsnotify provides a mechanism for "groups" to register for some set of filesystem events and to then deliver those events to those groups for processing. fsnotify has a number of benefits, the first being actually shrinking the size of an inode. Before fsnotify to support both dnotify and inotify an inode had unsigned long i_dnotify_mask; /* Directory notify events */ struct dnotify_struct *i_dnotify; /* for directory notifications */ struct list_head inotify_watches; /* watches on this inode */ struct mutex inotify_mutex; /* protects the watches list But with fsnotify this same functionallity (and more) is done with just __u32 i_fsnotify_mask; /* all events for this inode */ struct hlist_head i_fsnotify_mark_entries; /* marks on this inode */ That's right, inotify, dnotify, and fanotify all in 64 bits. We used that much space just in inotify_watches alone, before this patch set. fsnotify object lifetime and locking is MUCH better than what we have today. inotify locking is incredibly complex. See 8f7b0ba1c8539 as an example of what's been busted since inception. inotify needs to know internal semantics of superblock destruction and unmounting to function. The inode pinning and vfs contortions are horrible. no fsnotify implementers do allocation under locks. This means things like f04b30de3 which (due to an overabundance of caution) changes GFP_KERNEL to GFP_NOFS can be reverted. There are no longer any allocation rules when using or implementing your own fsnotify listener. fsnotify paves the way for fanotify. In brief fanotify is a notification mechanism that delivers the lisener both an 'event' and an open file descriptor to the object in question. This means that fanotify is pathname agnostic. Some on lkml may not care for the original companies or users that pushed for TALPA, but fanotify was designed with flexibility and input for other users in mind. The readahead group expressed interest in fanotify as it could be used to profile disk access on boot without breaking the audit system. The desktop search groups have also expressed interest in fanotify as it solves a number of the race conditions and problems present with managing inotify when more than a limited number of specific files are of interest. fanotify can provide for a userspace access control system which makes it a clean interface for AV vendors to hook without trying to do binary patching on the syscall table, LSM, and everywhere else they do their things today. With this patch series fanotify can be implemented in less than 1200 lines of easy to review code. Almost all of which is the socket based user interface. This patch series builds fsnotify to the point that it can implement dnotify and inotify_user. Patches exist and will be sent soon after acceptance to finish the in kernel inotify conversion (audit) and implement fanotify. Signed-off-by: Eric Paris <eparis@redhat.com> Acked-by: Al Viro <viro@zeniv.linux.org.uk> Cc: Christoph Hellwig <hch@lst.de>
2009-05-22 05:01:20 +08:00
extern void fsnotify_put_group(struct fsnotify_group *group);
/* take a reference to an event */
extern void fsnotify_get_event(struct fsnotify_event *event);
extern void fsnotify_put_event(struct fsnotify_event *event);
/* find private data previously attached to an event and unlink it */
extern struct fsnotify_event_private_data *fsnotify_remove_priv_from_event(struct fsnotify_group *group,
struct fsnotify_event *event);
fsnotify: unified filesystem notification backend fsnotify is a backend for filesystem notification. fsnotify does not provide any userspace interface but does provide the basis needed for other notification schemes such as dnotify. fsnotify can be extended to be the backend for inotify or the upcoming fanotify. fsnotify provides a mechanism for "groups" to register for some set of filesystem events and to then deliver those events to those groups for processing. fsnotify has a number of benefits, the first being actually shrinking the size of an inode. Before fsnotify to support both dnotify and inotify an inode had unsigned long i_dnotify_mask; /* Directory notify events */ struct dnotify_struct *i_dnotify; /* for directory notifications */ struct list_head inotify_watches; /* watches on this inode */ struct mutex inotify_mutex; /* protects the watches list But with fsnotify this same functionallity (and more) is done with just __u32 i_fsnotify_mask; /* all events for this inode */ struct hlist_head i_fsnotify_mark_entries; /* marks on this inode */ That's right, inotify, dnotify, and fanotify all in 64 bits. We used that much space just in inotify_watches alone, before this patch set. fsnotify object lifetime and locking is MUCH better than what we have today. inotify locking is incredibly complex. See 8f7b0ba1c8539 as an example of what's been busted since inception. inotify needs to know internal semantics of superblock destruction and unmounting to function. The inode pinning and vfs contortions are horrible. no fsnotify implementers do allocation under locks. This means things like f04b30de3 which (due to an overabundance of caution) changes GFP_KERNEL to GFP_NOFS can be reverted. There are no longer any allocation rules when using or implementing your own fsnotify listener. fsnotify paves the way for fanotify. In brief fanotify is a notification mechanism that delivers the lisener both an 'event' and an open file descriptor to the object in question. This means that fanotify is pathname agnostic. Some on lkml may not care for the original companies or users that pushed for TALPA, but fanotify was designed with flexibility and input for other users in mind. The readahead group expressed interest in fanotify as it could be used to profile disk access on boot without breaking the audit system. The desktop search groups have also expressed interest in fanotify as it solves a number of the race conditions and problems present with managing inotify when more than a limited number of specific files are of interest. fanotify can provide for a userspace access control system which makes it a clean interface for AV vendors to hook without trying to do binary patching on the syscall table, LSM, and everywhere else they do their things today. With this patch series fanotify can be implemented in less than 1200 lines of easy to review code. Almost all of which is the socket based user interface. This patch series builds fsnotify to the point that it can implement dnotify and inotify_user. Patches exist and will be sent soon after acceptance to finish the in kernel inotify conversion (audit) and implement fanotify. Signed-off-by: Eric Paris <eparis@redhat.com> Acked-by: Al Viro <viro@zeniv.linux.org.uk> Cc: Christoph Hellwig <hch@lst.de>
2009-05-22 05:01:20 +08:00
/* attach the event to the group notification queue */
extern int fsnotify_add_notify_event(struct fsnotify_group *group,
struct fsnotify_event *event,
struct fsnotify_event_private_data *priv,
int (*merge)(struct list_head *, struct fsnotify_event *));
/* true if the group notification queue is empty */
extern bool fsnotify_notify_queue_is_empty(struct fsnotify_group *group);
/* return, but do not dequeue the first event on the notification queue */
extern struct fsnotify_event *fsnotify_peek_notify_event(struct fsnotify_group *group);
/* return AND dequeue the first event on the notification queue */
extern struct fsnotify_event *fsnotify_remove_notify_event(struct fsnotify_group *group);
/* functions used to manipulate the marks attached to inodes */
/* run all marks associated with an inode and update inode->i_fsnotify_mask */
extern void fsnotify_recalc_inode_mask(struct inode *inode);
extern void fsnotify_init_mark(struct fsnotify_mark_entry *entry, void (*free_mark)(struct fsnotify_mark_entry *entry));
/* find (and take a reference) to a mark associated with group and inode */
extern struct fsnotify_mark_entry *fsnotify_find_mark_entry(struct fsnotify_group *group, struct inode *inode);
/* copy the values from old into new */
extern void fsnotify_duplicate_mark(struct fsnotify_mark_entry *new, struct fsnotify_mark_entry *old);
/* attach the mark to both the group and the inode */
extern int fsnotify_add_mark(struct fsnotify_mark_entry *entry, struct fsnotify_group *group, struct inode *inode, int allow_dups);
/* given a mark, flag it to be freed when all references are dropped */
extern void fsnotify_destroy_mark_by_entry(struct fsnotify_mark_entry *entry);
/* run all the marks in a group, and flag them to be freed */
extern void fsnotify_clear_marks_by_group(struct fsnotify_group *group);
extern void fsnotify_get_mark(struct fsnotify_mark_entry *entry);
extern void fsnotify_put_mark(struct fsnotify_mark_entry *entry);
extern void fsnotify_unmount_inodes(struct list_head *list);
fsnotify: unified filesystem notification backend fsnotify is a backend for filesystem notification. fsnotify does not provide any userspace interface but does provide the basis needed for other notification schemes such as dnotify. fsnotify can be extended to be the backend for inotify or the upcoming fanotify. fsnotify provides a mechanism for "groups" to register for some set of filesystem events and to then deliver those events to those groups for processing. fsnotify has a number of benefits, the first being actually shrinking the size of an inode. Before fsnotify to support both dnotify and inotify an inode had unsigned long i_dnotify_mask; /* Directory notify events */ struct dnotify_struct *i_dnotify; /* for directory notifications */ struct list_head inotify_watches; /* watches on this inode */ struct mutex inotify_mutex; /* protects the watches list But with fsnotify this same functionallity (and more) is done with just __u32 i_fsnotify_mask; /* all events for this inode */ struct hlist_head i_fsnotify_mark_entries; /* marks on this inode */ That's right, inotify, dnotify, and fanotify all in 64 bits. We used that much space just in inotify_watches alone, before this patch set. fsnotify object lifetime and locking is MUCH better than what we have today. inotify locking is incredibly complex. See 8f7b0ba1c8539 as an example of what's been busted since inception. inotify needs to know internal semantics of superblock destruction and unmounting to function. The inode pinning and vfs contortions are horrible. no fsnotify implementers do allocation under locks. This means things like f04b30de3 which (due to an overabundance of caution) changes GFP_KERNEL to GFP_NOFS can be reverted. There are no longer any allocation rules when using or implementing your own fsnotify listener. fsnotify paves the way for fanotify. In brief fanotify is a notification mechanism that delivers the lisener both an 'event' and an open file descriptor to the object in question. This means that fanotify is pathname agnostic. Some on lkml may not care for the original companies or users that pushed for TALPA, but fanotify was designed with flexibility and input for other users in mind. The readahead group expressed interest in fanotify as it could be used to profile disk access on boot without breaking the audit system. The desktop search groups have also expressed interest in fanotify as it solves a number of the race conditions and problems present with managing inotify when more than a limited number of specific files are of interest. fanotify can provide for a userspace access control system which makes it a clean interface for AV vendors to hook without trying to do binary patching on the syscall table, LSM, and everywhere else they do their things today. With this patch series fanotify can be implemented in less than 1200 lines of easy to review code. Almost all of which is the socket based user interface. This patch series builds fsnotify to the point that it can implement dnotify and inotify_user. Patches exist and will be sent soon after acceptance to finish the in kernel inotify conversion (audit) and implement fanotify. Signed-off-by: Eric Paris <eparis@redhat.com> Acked-by: Al Viro <viro@zeniv.linux.org.uk> Cc: Christoph Hellwig <hch@lst.de>
2009-05-22 05:01:20 +08:00
/* put here because inotify does some weird stuff when destroying watches */
extern struct fsnotify_event *fsnotify_create_event(struct inode *to_tell, __u32 mask,
void *data, int data_is, const char *name,
u32 cookie, gfp_t gfp);
/* fanotify likes to change events after they are on lists... */
extern struct fsnotify_event *fsnotify_clone_event(struct fsnotify_event *old_event);
extern int fsnotify_replace_event(struct fsnotify_event_holder *old_holder,
struct fsnotify_event *new_event);
fsnotify: unified filesystem notification backend fsnotify is a backend for filesystem notification. fsnotify does not provide any userspace interface but does provide the basis needed for other notification schemes such as dnotify. fsnotify can be extended to be the backend for inotify or the upcoming fanotify. fsnotify provides a mechanism for "groups" to register for some set of filesystem events and to then deliver those events to those groups for processing. fsnotify has a number of benefits, the first being actually shrinking the size of an inode. Before fsnotify to support both dnotify and inotify an inode had unsigned long i_dnotify_mask; /* Directory notify events */ struct dnotify_struct *i_dnotify; /* for directory notifications */ struct list_head inotify_watches; /* watches on this inode */ struct mutex inotify_mutex; /* protects the watches list But with fsnotify this same functionallity (and more) is done with just __u32 i_fsnotify_mask; /* all events for this inode */ struct hlist_head i_fsnotify_mark_entries; /* marks on this inode */ That's right, inotify, dnotify, and fanotify all in 64 bits. We used that much space just in inotify_watches alone, before this patch set. fsnotify object lifetime and locking is MUCH better than what we have today. inotify locking is incredibly complex. See 8f7b0ba1c8539 as an example of what's been busted since inception. inotify needs to know internal semantics of superblock destruction and unmounting to function. The inode pinning and vfs contortions are horrible. no fsnotify implementers do allocation under locks. This means things like f04b30de3 which (due to an overabundance of caution) changes GFP_KERNEL to GFP_NOFS can be reverted. There are no longer any allocation rules when using or implementing your own fsnotify listener. fsnotify paves the way for fanotify. In brief fanotify is a notification mechanism that delivers the lisener both an 'event' and an open file descriptor to the object in question. This means that fanotify is pathname agnostic. Some on lkml may not care for the original companies or users that pushed for TALPA, but fanotify was designed with flexibility and input for other users in mind. The readahead group expressed interest in fanotify as it could be used to profile disk access on boot without breaking the audit system. The desktop search groups have also expressed interest in fanotify as it solves a number of the race conditions and problems present with managing inotify when more than a limited number of specific files are of interest. fanotify can provide for a userspace access control system which makes it a clean interface for AV vendors to hook without trying to do binary patching on the syscall table, LSM, and everywhere else they do their things today. With this patch series fanotify can be implemented in less than 1200 lines of easy to review code. Almost all of which is the socket based user interface. This patch series builds fsnotify to the point that it can implement dnotify and inotify_user. Patches exist and will be sent soon after acceptance to finish the in kernel inotify conversion (audit) and implement fanotify. Signed-off-by: Eric Paris <eparis@redhat.com> Acked-by: Al Viro <viro@zeniv.linux.org.uk> Cc: Christoph Hellwig <hch@lst.de>
2009-05-22 05:01:20 +08:00
#else
static inline void fsnotify(struct inode *to_tell, __u32 mask, void *data, int data_is,
const char *name, u32 cookie)
fsnotify: unified filesystem notification backend fsnotify is a backend for filesystem notification. fsnotify does not provide any userspace interface but does provide the basis needed for other notification schemes such as dnotify. fsnotify can be extended to be the backend for inotify or the upcoming fanotify. fsnotify provides a mechanism for "groups" to register for some set of filesystem events and to then deliver those events to those groups for processing. fsnotify has a number of benefits, the first being actually shrinking the size of an inode. Before fsnotify to support both dnotify and inotify an inode had unsigned long i_dnotify_mask; /* Directory notify events */ struct dnotify_struct *i_dnotify; /* for directory notifications */ struct list_head inotify_watches; /* watches on this inode */ struct mutex inotify_mutex; /* protects the watches list But with fsnotify this same functionallity (and more) is done with just __u32 i_fsnotify_mask; /* all events for this inode */ struct hlist_head i_fsnotify_mark_entries; /* marks on this inode */ That's right, inotify, dnotify, and fanotify all in 64 bits. We used that much space just in inotify_watches alone, before this patch set. fsnotify object lifetime and locking is MUCH better than what we have today. inotify locking is incredibly complex. See 8f7b0ba1c8539 as an example of what's been busted since inception. inotify needs to know internal semantics of superblock destruction and unmounting to function. The inode pinning and vfs contortions are horrible. no fsnotify implementers do allocation under locks. This means things like f04b30de3 which (due to an overabundance of caution) changes GFP_KERNEL to GFP_NOFS can be reverted. There are no longer any allocation rules when using or implementing your own fsnotify listener. fsnotify paves the way for fanotify. In brief fanotify is a notification mechanism that delivers the lisener both an 'event' and an open file descriptor to the object in question. This means that fanotify is pathname agnostic. Some on lkml may not care for the original companies or users that pushed for TALPA, but fanotify was designed with flexibility and input for other users in mind. The readahead group expressed interest in fanotify as it could be used to profile disk access on boot without breaking the audit system. The desktop search groups have also expressed interest in fanotify as it solves a number of the race conditions and problems present with managing inotify when more than a limited number of specific files are of interest. fanotify can provide for a userspace access control system which makes it a clean interface for AV vendors to hook without trying to do binary patching on the syscall table, LSM, and everywhere else they do their things today. With this patch series fanotify can be implemented in less than 1200 lines of easy to review code. Almost all of which is the socket based user interface. This patch series builds fsnotify to the point that it can implement dnotify and inotify_user. Patches exist and will be sent soon after acceptance to finish the in kernel inotify conversion (audit) and implement fanotify. Signed-off-by: Eric Paris <eparis@redhat.com> Acked-by: Al Viro <viro@zeniv.linux.org.uk> Cc: Christoph Hellwig <hch@lst.de>
2009-05-22 05:01:20 +08:00
{}
static inline void __fsnotify_parent(struct file *file, struct dentry *dentry, __u32 mask)
{}
static inline void __fsnotify_inode_delete(struct inode *inode)
{}
static inline void __fsnotify_update_dcache_flags(struct dentry *dentry)
{}
static inline void __fsnotify_d_instantiate(struct dentry *dentry, struct inode *inode)
{}
static inline u32 fsnotify_get_cookie(void)
{
return 0;
}
static inline void fsnotify_unmount_inodes(struct list_head *list)
{}
fsnotify: unified filesystem notification backend fsnotify is a backend for filesystem notification. fsnotify does not provide any userspace interface but does provide the basis needed for other notification schemes such as dnotify. fsnotify can be extended to be the backend for inotify or the upcoming fanotify. fsnotify provides a mechanism for "groups" to register for some set of filesystem events and to then deliver those events to those groups for processing. fsnotify has a number of benefits, the first being actually shrinking the size of an inode. Before fsnotify to support both dnotify and inotify an inode had unsigned long i_dnotify_mask; /* Directory notify events */ struct dnotify_struct *i_dnotify; /* for directory notifications */ struct list_head inotify_watches; /* watches on this inode */ struct mutex inotify_mutex; /* protects the watches list But with fsnotify this same functionallity (and more) is done with just __u32 i_fsnotify_mask; /* all events for this inode */ struct hlist_head i_fsnotify_mark_entries; /* marks on this inode */ That's right, inotify, dnotify, and fanotify all in 64 bits. We used that much space just in inotify_watches alone, before this patch set. fsnotify object lifetime and locking is MUCH better than what we have today. inotify locking is incredibly complex. See 8f7b0ba1c8539 as an example of what's been busted since inception. inotify needs to know internal semantics of superblock destruction and unmounting to function. The inode pinning and vfs contortions are horrible. no fsnotify implementers do allocation under locks. This means things like f04b30de3 which (due to an overabundance of caution) changes GFP_KERNEL to GFP_NOFS can be reverted. There are no longer any allocation rules when using or implementing your own fsnotify listener. fsnotify paves the way for fanotify. In brief fanotify is a notification mechanism that delivers the lisener both an 'event' and an open file descriptor to the object in question. This means that fanotify is pathname agnostic. Some on lkml may not care for the original companies or users that pushed for TALPA, but fanotify was designed with flexibility and input for other users in mind. The readahead group expressed interest in fanotify as it could be used to profile disk access on boot without breaking the audit system. The desktop search groups have also expressed interest in fanotify as it solves a number of the race conditions and problems present with managing inotify when more than a limited number of specific files are of interest. fanotify can provide for a userspace access control system which makes it a clean interface for AV vendors to hook without trying to do binary patching on the syscall table, LSM, and everywhere else they do their things today. With this patch series fanotify can be implemented in less than 1200 lines of easy to review code. Almost all of which is the socket based user interface. This patch series builds fsnotify to the point that it can implement dnotify and inotify_user. Patches exist and will be sent soon after acceptance to finish the in kernel inotify conversion (audit) and implement fanotify. Signed-off-by: Eric Paris <eparis@redhat.com> Acked-by: Al Viro <viro@zeniv.linux.org.uk> Cc: Christoph Hellwig <hch@lst.de>
2009-05-22 05:01:20 +08:00
#endif /* CONFIG_FSNOTIFY */
#endif /* __KERNEL __ */
#endif /* __LINUX_FSNOTIFY_BACKEND_H */