FAN_FS_ERROR allows events without inodes - i.e. for file system-wide
errors. Even though fsnotify_handle_inode_event is not currently used
by fanotify, this patch protects other backends from cases where neither
inode or dir are provided. Also document the constraints of the
interface (inode and dir cannot be both NULL).
Link: https://lore.kernel.org/r/20211025192746.66445-12-krisman@collabora.com
Suggested-by: Amir Goldstein <amir73il@gmail.com>
Signed-off-by: Gabriel Krisman Bertazi <krisman@collabora.com>
Reviewed-by: Amir Goldstein <amir73il@gmail.com>
Reviewed-by: Jan Kara <jack@suse.cz>
Signed-off-by: Jan Kara <jack@suse.cz>
Some file system events (i.e. FS_ERROR) might not be associated with an
inode or directory. For these, we can retrieve the super block from the
data field. But, since the super_block is available in the data field
on every event type, simplify the code to always retrieve it from there,
through a new helper.
Link: https://lore.kernel.org/r/20211025192746.66445-11-krisman@collabora.com
Suggested-by: Jan Kara <jack@suse.cz>
Reviewed-by: Amir Goldstein <amir73il@gmail.com>
Reviewed-by: Jan Kara <jack@suse.cz>
Signed-off-by: Gabriel Krisman Bertazi <krisman@collabora.com>
Signed-off-by: Jan Kara <jack@suse.cz>
Define a new data type to pass for event - FSNOTIFY_EVENT_DENTRY.
Use it to pass the dentry instead of it's ->d_inode where available.
This is needed in preparation to the refactor to retrieve the super
block from the data field. In some cases (i.e. mkdir in kernfs), the
data inode comes from a negative dentry, such that no super block
information would be available. By receiving the dentry itself, instead
of the inode, fsnotify can derive the super block even on these cases.
Link: https://lore.kernel.org/r/20211025192746.66445-3-krisman@collabora.com
Reviewed-by: Jan Kara <jack@suse.cz>
Signed-off-by: Amir Goldstein <amir73il@gmail.com>
[Expand explanation in commit message]
Signed-off-by: Gabriel Krisman Bertazi <krisman@collabora.com>
Signed-off-by: Jan Kara <jack@suse.cz>
fanotify has some hardcoded limits. The only APIs to escape those limits
are FAN_UNLIMITED_QUEUE and FAN_UNLIMITED_MARKS.
Allow finer grained tuning of the system limits via sysfs tunables under
/proc/sys/fs/fanotify, similar to tunables under /proc/sys/fs/inotify,
with some minor differences.
- max_queued_events - global system tunable for group queue size limit.
Like the inotify tunable with the same name, it defaults to 16384 and
applies on initialization of a new group.
- max_user_marks - user ns tunable for marks limit per user.
Like the inotify tunable named max_user_watches, on a machine with
sufficient RAM and it defaults to 1048576 in init userns and can be
further limited per containing user ns.
- max_user_groups - user ns tunable for number of groups per user.
Like the inotify tunable named max_user_instances, it defaults to 128
in init userns and can be further limited per containing user ns.
The slightly different tunable names used for fanotify are derived from
the "group" and "mark" terminology used in the fanotify man pages and
throughout the code.
Considering the fact that the default value for max_user_instances was
increased in kernel v5.10 from 8192 to 1048576, leaving the legacy
fanotify limit of 8192 marks per group in addition to the max_user_marks
limit makes little sense, so the per group marks limit has been removed.
Note that when a group is initialized with FAN_UNLIMITED_MARKS, its own
marks are not accounted in the per user marks account, so in effect the
limit of max_user_marks is only for the collection of groups that are
not initialized with FAN_UNLIMITED_MARKS.
Link: https://lore.kernel.org/r/20210304112921.3996419-2-amir73il@gmail.com
Suggested-by: Jan Kara <jack@suse.cz>
Signed-off-by: Amir Goldstein <amir73il@gmail.com>
Signed-off-by: Jan Kara <jack@suse.cz>
In order to improve event merge performance, hash events in a 128 size
hash table by the event merge key.
The fanotify_event size grows by two pointers, but we just reduced its
size by removing the objectid member, so overall its size is increased
by one pointer.
Permission events and overflow event are not merged so they are also
not hashed.
Link: https://lore.kernel.org/r/20210304104826.3993892-5-amir73il@gmail.com
Signed-off-by: Amir Goldstein <amir73il@gmail.com>
Signed-off-by: Jan Kara <jack@suse.cz>
objectid is only used by fanotify backend and it is just an optimization
for event merge before comparing all fields in event.
Move the objectid member from common struct fsnotify_event into struct
fanotify_event and reduce it to 29-bit hash to cram it together with the
3-bit event type.
Events of different types are never merged, so the combination of event
type and hash form a 32-bit key for fast compare of events.
This reduces the size of events by one pointer and paves the way for
adding hashed queue support for fanotify.
Link: https://lore.kernel.org/r/20210304104826.3993892-3-amir73il@gmail.com
Signed-off-by: Amir Goldstein <amir73il@gmail.com>
Signed-off-by: Jan Kara <jack@suse.cz>
Current code has an assumtion that fsnotify_notify_queue_is_empty() is
called to verify that queue is not empty before trying to peek or remove
an event from queue.
Remove this assumption by moving the fsnotify_notify_queue_is_empty()
into the functions, allow them to return NULL value and check return
value by all callers.
This is a prep patch for multi event queues.
Link: https://lore.kernel.org/r/20210304104826.3993892-2-amir73il@gmail.com
Signed-off-by: Amir Goldstein <amir73il@gmail.com>
Signed-off-by: Jan Kara <jack@suse.cz>
Currently the fs sysctl inotify/max_user_instances is used to limit the
number of inotify instances on the system. For systems running multiple
workloads, the per-user namespace sysctl max_inotify_instances can be
used to further partition inotify instances. However there is no easy
way to set a sensible system level max limit on inotify instances and
further partition it between the workloads. It is much easier to charge
the underlying resource (i.e. memory) behind the inotify instances to
the memcg of the workload and let their memory limits limit the number
of inotify instances they can create.
With inotify instances charged to memcg, the admin can simply set
max_user_instances to INT_MAX and let the memcg limits of the jobs limit
their inotify instances.
Link: https://lore.kernel.org/r/20201220044608.1258123-1-shakeelb@google.com
Reviewed-by: Amir Goldstein <amir73il@gmail.com>
Signed-off-by: Shakeel Butt <shakeelb@google.com>
Signed-off-by: Jan Kara <jack@suse.cz>
fsnotify_parent() used to send two separate events to backends when a
parent inode is watching children and the child inode is also watching.
In an attempt to avoid duplicate events in fanotify, we unified the two
backend callbacks to a single callback and handled the reporting of the
two separate events for the relevant backends (inotify and dnotify).
However the handling is buggy and can result in inotify and dnotify
listeners receiving events of the type they never asked for or spurious
events.
The problem is the unified event callback with two inode marks (parent and
child) is called when any of the parent and child inodes are watched and
interested in the event, but the parent inode's mark that is interested
in the event on the child is not necessarily the one we are currently
reporting to (it could belong to a different group).
So before reporting the parent or child event flavor to backend we need
to check that the mark is really interested in that event flavor.
The semantics of INODE and CHILD marks were hard to follow and made the
logic more complicated than it should have been. Replace it with INODE
and PARENT marks semantics to hopefully make the logic more clear.
Thanks to Hugh Dickins for spotting a bug in the earlier version of this
patch.
Fixes: 497b0c5a7c ("fsnotify: send event to parent and child with single callback")
CC: stable@vger.kernel.org
Link: https://lore.kernel.org/r/20201202120713.702387-4-amir73il@gmail.com
Reported-by: Hugh Dickins <hughd@google.com>
Signed-off-by: Amir Goldstein <amir73il@gmail.com>
Signed-off-by: Jan Kara <jack@suse.cz>
The handle_inode_event() interface was added as (quoting comment):
"a simple variant of handle_event() for groups that only have inode
marks and don't have ignore mask".
In other words, all backends except fanotify. The inotify backend
also falls under this category, but because it required extra arguments
it was left out of the initial pass of backends conversion to the
simple interface.
This results in code duplication between the generic helper
fsnotify_handle_event() and the inotify_handle_event() callback
which also happen to be buggy code.
Generalize the handle_inode_event() arguments and add the check for
FS_EXCL_UNLINK flag to the generic helper, so inotify backend could
be converted to use the simple interface.
Link: https://lore.kernel.org/r/20201202120713.702387-2-amir73il@gmail.com
CC: stable@vger.kernel.org
Fixes: b9a1b97725 ("fsnotify: create method handle_inode_event() in fsnotify_operations")
Signed-off-by: Amir Goldstein <amir73il@gmail.com>
Signed-off-by: Jan Kara <jack@suse.cz>
The method handle_event() grew a lot of complexity due to the design of
fanotify and merging of ignore masks.
Most backends do not care about this complex functionality, so we can hide
this complexity from them.
Introduce a method handle_inode_event() that serves those backends and
passes a single inode mark and less arguments.
This change converts all backends except fanotify and inotify to use the
simplified handle_inode_event() method. In pricipal, inotify could have
also used the new method, but that would require passing more arguments
on the simple helper (data, data_type, cookie), so we leave it with the
handle_event() method.
Link: https://lore.kernel.org/r/20200722125849.17418-9-amir73il@gmail.com
Suggested-by: Jan Kara <jack@suse.cz>
Signed-off-by: Amir Goldstein <amir73il@gmail.com>
Signed-off-by: Jan Kara <jack@suse.cz>
Similar to events "on child" to watching directory, send event
with parent/name info if sb/mount/non-dir marks are interested in
parent/name info.
The FS_EVENT_ON_CHILD flag can be set on sb/mount/non-dir marks to specify
interest in parent/name info for events on non-directory inodes.
Events on "orphan" children (disconnected dentries) are sent without
parent/name info.
Events on directories are sent with parent/name info only if the parent
directory is watching.
After this change, even groups that do not subscribe to events on
children could get an event with mark iterator type TYPE_CHILD and
without mark iterator type TYPE_INODE if fanotify has marks on the same
objects.
dnotify and inotify event handlers can already cope with that situation.
audit does not subscribe to events that are possible on child, so won't
get to this situation. nfsd does not access the marks iterator from its
event handler at the moment, so it is not affected.
This is a bit too fragile, so we should prepare all groups to cope with
mark type TYPE_CHILD preferably using a generic helper.
Link: https://lore.kernel.org/r/20200716084230.30611-16-amir73il@gmail.com
Signed-off-by: Amir Goldstein <amir73il@gmail.com>
Signed-off-by: Jan Kara <jack@suse.cz>
The arguments of fsnotify() are overloaded and mean different things
for different event types.
Replace the to_tell argument with separate arguments @dir and @inode,
because we may be sending to both dir and child. Using the @data
argument to pass the child is not enough, because dirent events pass
this argument (for audit), but we do not report to child.
Document the new fsnotify() function argumenets.
Link: https://lore.kernel.org/r/20200722125849.17418-7-amir73il@gmail.com
Signed-off-by: Amir Goldstein <amir73il@gmail.com>
Signed-off-by: Jan Kara <jack@suse.cz>
The object type iterator is used to collect all the marks of
a specific group that have interest in an event.
It is used by fanotify to get a single handle_event callback
when an event has a match to either of inode/sb/mount marks
of the group.
The nature of fsnotify events is that they are associated with
at most one sb at most one mount and at most one inode.
When a parent and child are both watching, two events are sent
to backend, one associated to parent inode and one associated
to the child inode.
This results in duplicate events in fanotify, which usually
get merged before user reads them, but this is sub-optimal.
It would be better if the same event is sent to backend with
an object type iterator that has both the child inode and its
parent, and let the backend decide if the event should be reported
once (fanotify) or twice (inotify).
Link: https://lore.kernel.org/r/20200716084230.30611-9-amir73il@gmail.com
Signed-off-by: Amir Goldstein <amir73il@gmail.com>
Signed-off-by: Jan Kara <jack@suse.cz>
It was never enabled in uapi and its functionality is about to be
superseded by events FAN_CREATE, FAN_DELETE, FAN_MOVE with group
flag FAN_REPORT_NAME.
Keep a place holder variable name_event instead of removing the
name recording code since it will be used by the new events.
Link: https://lore.kernel.org/r/20200708111156.24659-17-amir73il@gmail.com
Signed-off-by: Amir Goldstein <amir73il@gmail.com>
Signed-off-by: Jan Kara <jack@suse.cz>
The 'inode' argument to handle_event(), sometimes referred to as
'to_tell' is somewhat obsolete.
It is a remnant from the times when a group could only have an inode mark
associated with an event.
We now pass an iter_info array to the callback, with all marks associated
with an event.
Most backends ignore this argument, with two exceptions:
1. dnotify uses it for sanity check that event is on directory
2. fanotify uses it to report fid of directory on directory entry
modification events
Remove the 'inode' argument and add a 'dir' argument.
The callback function signature is deliberately changed, because
the meaning of the argument has changed and the arguments have
been documented.
The 'dir' argument is set to when 'file_name' is specified and it is
referring to the directory that the 'file_name' entry belongs to.
Signed-off-by: Amir Goldstein <amir73il@gmail.com>
Signed-off-by: Jan Kara <jack@suse.cz>
The fsnotify paths are trivial to hit even when there are no watchers and
they are surprisingly expensive. For example, every successful vfs_write()
hits fsnotify_modify which calls both fsnotify_parent and fsnotify unless
FMODE_NONOTIFY is set which is an internal flag invisible to userspace.
As it stands, fsnotify_parent is a guaranteed functional call even if there
are no watchers and fsnotify() does a substantial amount of unnecessary
work before it checks if there are any watchers. A perf profile showed
that applying mnt->mnt_fsnotify_mask in fnotify() was almost half of the
total samples taken in that function during a test. This patch rearranges
the fast paths to reduce the amount of work done when there are no
watchers.
The test motivating this was "perf bench sched messaging --pipe". Despite
the fact the pipes are anonymous, fsnotify is still called a lot and
the overhead is noticeable even though it's completely pointless. It's
likely the overhead is negligible for real IO so this is an extreme
example. This is a comparison of hackbench using processes and pipes on
a 1-socket machine with 8 CPU threads without fanotify watchers.
5.7.0 5.7.0
vanilla fastfsnotify-v1r1
Amean 1 0.4837 ( 0.00%) 0.4630 * 4.27%*
Amean 3 1.5447 ( 0.00%) 1.4557 ( 5.76%)
Amean 5 2.6037 ( 0.00%) 2.4363 ( 6.43%)
Amean 7 3.5987 ( 0.00%) 3.4757 ( 3.42%)
Amean 12 5.8267 ( 0.00%) 5.6983 ( 2.20%)
Amean 18 8.4400 ( 0.00%) 8.1327 ( 3.64%)
Amean 24 11.0187 ( 0.00%) 10.0290 * 8.98%*
Amean 30 13.1013 ( 0.00%) 12.8510 ( 1.91%)
Amean 32 13.9190 ( 0.00%) 13.2410 ( 4.87%)
5.7.0 5.7.0
vanilla fastfsnotify-v1r1
Duration User 157.05 152.79
Duration System 1279.98 1219.32
Duration Elapsed 182.81 174.52
This is showing that the latencies are improved by roughly 2-9%. The
variability is not shown but some of these results are within the noise
as this workload heavily overloads the machine. That said, the system CPU
usage is reduced by quite a bit so it makes sense to avoid the overhead
even if it is a bit tricky to detect at times. A perf profile of just 1
group of tasks showed that 5.14% of samples taken were in either fsnotify()
or fsnotify_parent(). With the patch, 2.8% of samples were in fsnotify,
mostly function entry and the initial check for watchers. The check for
watchers is complicated enough that inlining it may be controversial.
[Amir] Slightly simplify with mnt_or_sb_mask => marks_mask
Link: https://lore.kernel.org/r/20200708111156.24659-1-amir73il@gmail.com
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Signed-off-by: Amir Goldstein <amir73il@gmail.com>
Signed-off-by: Jan Kara <jack@suse.cz>
Dirent events are going to be supported in two flavors:
1. Directory fid info + mask that includes the specific event types
(e.g. FAN_CREATE) and an optional FAN_ONDIR flag.
2. Directory fid info + name + mask that includes only FAN_DIR_MODIFY.
To request the second event flavor, user needs to set the event type
FAN_DIR_MODIFY in the mark mask.
The first flavor is supported since kernel v5.1 for groups initialized
with flag FAN_REPORT_FID. It is intended to be used for watching
directories in "batch mode" - the watcher is notified when directory is
changed and re-scans the directory content in response. This event
flavor is stored more compactly in the event queue, so it is optimal
for workloads with frequent directory changes.
The second event flavor is intended to be used for watching large
directories, where the cost of re-scan of the directory on every change
is considered too high. The watcher getting the event with the directory
fid and entry name is expected to call fstatat(2) to query the content of
the entry after the change.
Legacy inotify events are reported with name and event mask (e.g. "foo",
FAN_CREATE | FAN_ONDIR). That can lead users to the conclusion that
there is *currently* an entry "foo" that is a sub-directory, when in fact
"foo" may be negative or non-dir by the time user gets the event.
To make it clear that the current state of the named entry is unknown,
when reporting an event with name info, fanotify obfuscates the specific
event types (e.g. create,delete,rename) and uses a common event type -
FAN_DIR_MODIFY to describe the change. This should make it harder for
users to make wrong assumptions and write buggy filesystem monitors.
At this point, name info reporting is not yet implemented, so trying to
set FAN_DIR_MODIFY in mark mask will return -EINVAL.
Link: https://lore.kernel.org/r/20200319151022.31456-12-amir73il@gmail.com
Signed-off-by: Amir Goldstein <amir73il@gmail.com>
Signed-off-by: Jan Kara <jack@suse.cz>
