All other add functions for lists have the new item as first argument
and the position where it is added as second argument. This was changed
for no good reason in this function and makes using it unnecessary
confusing.
The name was changed to hlist_add_behind() to cause unconverted code to
generate a compile error instead of using the wrong parameter order.
[akpm@linux-foundation.org: coding-style fixes]
Signed-off-by: Ken Helias <kenhelias@firemail.de>
Cc: "Paul E. McKenney" <paulmck@linux.vnet.ibm.com>
Acked-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com> [intel driver bits]
Cc: Hugh Dickins <hughd@google.com>
Cc: Christoph Hellwig <hch@infradead.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Commit 8581679424 ("fanotify: Fix use after free for permission
events") introduced a double free issue for permission events which are
pending in group's notification queue while group is being destroyed.
These events are freed from fanotify_handle_event() but they are not
removed from groups notification queue and thus they get freed again
from fsnotify_flush_notify().
Fix the problem by removing permission events from notification queue
before freeing them if we skip processing access response. Also expand
comments in fanotify_release() to explain group shutdown in detail.
Fixes: 8581679424
Signed-off-by: Jan Kara <jack@suse.cz>
Reported-by: Douglas Leeder <douglas.leeder@sophos.com>
Tested-by: Douglas Leeder <douglas.leeder@sophos.com>
Reported-by: Heinrich Schuchard <xypron.glpk@gmx.de>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Rename fsnotify_add_notify_event() to fsnotify_add_event() since the
"notify" part is duplicit. Rename fsnotify_remove_notify_event() and
fsnotify_peek_notify_event() to fsnotify_remove_first_event() and
fsnotify_peek_first_event() respectively since "notify" part is duplicit
and they really look at the first event in the queue.
[akpm@linux-foundation.org: coding-style fixes]
Signed-off-by: Jan Kara <jack@suse.cz>
Cc: Eric Paris <eparis@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
This typedef is unnecessary and should just be removed.
Signed-off-by: Joe Perches <joe@perches.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Without this patch fanotify_init does not validate the value passed in
event_f_flags.
When a fanotify event is read from the fanotify file descriptor a new
file descriptor is created where file.f_flags = event_f_flags.
Internal and external open flags are stored together in field f_flags of
struct file. Hence, an application might create file descriptors with
internal flags like FMODE_EXEC, FMODE_NOCMTIME set.
Jan Kara and Eric Paris both aggreed that this is a bug and the value of
event_f_flags should be checked:
https://lkml.org/lkml/2014/4/29/522https://lkml.org/lkml/2014/4/29/539
This updated patch version considers the comments by Michael Kerrisk in
https://lkml.org/lkml/2014/5/4/10
With the patch the value of event_f_flags is checked.
When specifying an invalid value error EINVAL is returned.
Internal flags are disallowed.
File creation flags are disallowed:
O_CREAT, O_DIRECTORY, O_EXCL, O_NOCTTY, O_NOFOLLOW, O_TRUNC, and O_TTY_INIT.
Flags which do not make sense with fanotify are disallowed:
__O_TMPFILE, O_PATH, FASYNC, and O_DIRECT.
This leaves us with the following allowed values:
O_RDONLY, O_WRONLY, O_RDWR are basic functionality. The are stored in the
bits given by O_ACCMODE.
O_APPEND is working as expected. The value might be useful in a logging
application which appends the current status each time the log is opened.
O_LARGEFILE is needed for files exceeding 4GB on 32bit systems.
O_NONBLOCK may be useful when monitoring slow devices like tapes.
O_NDELAY is equal to O_NONBLOCK except for platform parisc.
To avoid code breaking on parisc either both flags should be
allowed or none. The patch allows both.
__O_SYNC and O_DSYNC may be used to avoid data loss on power disruption.
O_NOATIME may be useful to reduce disk activity.
O_CLOEXEC may be useful, if separate processes shall be used to scan files.
Once this patch is accepted, the fanotify_init.2 manpage has to be updated.
Signed-off-by: Heinrich Schuchardt <xypron.glpk@gmx.de>
Reviewed-by: Jan Kara <jack@suse.cz>
Cc: Michael Kerrisk <mtk.manpages@gmail.com>
Cc: Valdis Kletnieks <Valdis.Kletnieks@vt.edu>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
If fanotify_mark is called with illegal value of arguments flags and
marks it usually returns EINVAL.
When fanotify_mark is called with FAN_MARK_FLUSH the argument flags is
not checked for irrelevant flags like FAN_MARK_IGNORED_MASK.
The patch removes this inconsistency.
If an irrelevant flag is set error EINVAL is returned.
Signed-off-by: Heinrich Schuchardt <xypron.glpk@gmx.de>
Acked-by: Michael Kerrisk <mtk.manpages@gmail.com>
Acked-by: Jan Kara <jack@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Do not initialize private_destroy_list twice. list_replace_init()
already takes care of initializing private_destroy_list. We don't need
to initialize it with LIST_HEAD() beforehand.
Signed-off-by: David Cohen <david.a.cohen@linux.intel.com>
Cc: Jan Kara <jack@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Originally from Tvrtko Ursulin (https://lkml.org/lkml/2011/1/12/112)
Avoid having to provide a fake/invalid fd and path when flushing marks
Currently for a group to flush marks it has set it needs to provide a
fake or invalid (but resolvable) file descriptor and path when calling
fanotify_mark. This patch pulls the flush handling a bit up so file
descriptor and path are completely ignored when flushing.
I reworked the patch to be applicable again (the signature of
fanotify_mark has changed since Tvrtko's work).
Signed-off-by: Heinrich Schuchardt <xypron.glpk@gmx.de>
Cc: Tvrtko Ursulin <tvrtko.ursulin@onelan.co.uk>
Reviewed-by: Jan Kara <jack@suse.cz>
Acked-by: Eric Paris <eparis@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
On 64-bit systems, O_LARGEFILE is automatically added to flags inside
the open() syscall (also openat(), blkdev_open(), etc). Userspace
therefore defines O_LARGEFILE to be 0 - you can use it, but it's a
no-op. Everything should be O_LARGEFILE by default.
But: when fanotify does create_fd() it uses dentry_open(), which skips
all that. And userspace can't set O_LARGEFILE in fanotify_init()
because it's defined to 0. So if fanotify gets an event regarding a
large file, the read() will just fail with -EOVERFLOW.
This patch adds O_LARGEFILE to fanotify_init()'s event_f_flags on 64-bit
systems, using the same test as open()/openat()/etc.
Addresses https://bugzilla.redhat.com/show_bug.cgi?id=696821
Signed-off-by: Will Woods <wwoods@redhat.com>
Acked-by: Eric Paris <eparis@redhat.com>
Reviewed-by: Jan Kara <jack@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Move code moving event structure to access_list from copy_event_to_user()
to fanotify_read() where it is more logical (so that we can immediately
see in the main loop that we either move the event to a different list
or free it). Also move special error handling for permission events
from copy_event_to_user() to the main loop to have it in one place with
error handling for normal events. This makes copy_event_to_user()
really only copy the event to user without any side effects.
Signed-off-by: Jan Kara <jack@suse.cz>
Cc: Eric Paris <eparis@redhat.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Swap the error / "read ok" branches in the main loop of fanotify_read().
We will grow the "read ok" part in the next patch and this makes the
indentation easier. Also it is more common to have error conditions
inside an 'if' instead of the fast path.
Signed-off-by: Jan Kara <jack@suse.cz>
Cc: Eric Paris <eparis@redhat.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
access_mutex is used only to guard operations on access_list. There's
no need for sleeping within this lock so just make a spinlock out of it.
Signed-off-by: Jan Kara <jack@suse.cz>
Cc: Eric Paris <eparis@redhat.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Currently, fanotify creates new structure to track the fact that
permission event has been reported to userspace and someone is waiting
for a response to it. As event structures are now completely in the
hands of each notification framework, we can use the event structure for
this tracking instead of allocating a new structure.
Since this makes the event structures for normal events and permission
events even more different and the structures have different lifetime
rules, we split them into two separate structures (where permission
event structure contains the structure for a normal event). This makes
normal events 8 bytes smaller and the code a tad bit cleaner.
[akpm@linux-foundation.org: fix build]
Signed-off-by: Jan Kara <jack@suse.cz>
Cc: Eric Paris <eparis@redhat.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Wu Fengguang <fengguang.wu@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
The prepare_for_access_response() function checks whether
group->fanotify_data.bypass_perm is set. However this test can never be
true because prepare_for_access_response() is called only from
fanotify_read() which means fanotify group is alive with an active fd
while bypass_perm is set from fanotify_release() when all file
descriptors pointing to the group are closed and the group is going
away.
Signed-off-by: Jan Kara <jack@suse.cz>
Cc: Eric Paris <eparis@redhat.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Commit 7053aee26a "fsnotify: do not share events between notification
groups" used overflow event statically allocated in a group with the
size of the generic notification event. This causes problems because
some code looks at type specific parts of event structure and gets
confused by a random data it sees there and causes crashes.
Fix the problem by allocating overflow event with type corresponding to
the group type so code cannot get confused.
Signed-off-by: Jan Kara <jack@suse.cz>
If the event queue overflows when we are handling permission event, we
will never get response from userspace. So we must avoid waiting for it.
Change fsnotify_add_notify_event() to return whether overflow has
happened so that we can detect it in fanotify_handle_event() and act
accordingly.
Signed-off-by: Jan Kara <jack@suse.cz>
Currently we didn't initialize event's list head when we removed it from
the event list. Thus a detection whether overflow event is already
queued wasn't working. Fix it by always initializing the list head when
deleting event from a list.
Signed-off-by: Jan Kara <jack@suse.cz>
My rework of handling of notification events (namely commit 7053aee26a
"fsnotify: do not share events between notification groups") broke
sending of cookies with inotify events. We didn't propagate the value
passed to fsnotify() properly and passed 4 uninitialized bytes to
userspace instead (so it is also an information leak). Sadly I didn't
notice this during my testing because inotify cookies aren't used very
much and LTP inotify tests ignore them.
Fix the problem by passing the cookie value properly.
Fixes: 7053aee26a
Reported-by: Vegard Nossum <vegard.nossum@oracle.com>
Signed-off-by: Jan Kara <jack@suse.cz>
Currently struct fanotify_event_info has been destroyed immediately
after reporting its contents to userspace. However that is wrong for
permission events because those need to stay around until userspace
provides response which is filled back in fanotify_event_info. So change
to code to free permission events only after we have got the response
from userspace.
Reported-and-tested-by: Jiri Kosina <jkosina@suse.cz>
Reported-and-tested-by: Dave Jones <davej@fedoraproject.org>
Signed-off-by: Jan Kara <jack@suse.cz>
The event returned from fsnotify_add_notify_event() cannot ever be used
safely as the event may be freed by the time the function returns (after
dropping notification_mutex). So change the prototype to just return
whether the event was added or merged into some existing event.
Reported-and-tested-by: Jiri Kosina <jkosina@suse.cz>
Reported-and-tested-by: Dave Jones <davej@fedoraproject.org>
Signed-off-by: Jan Kara <jack@suse.cz>
We cannot use the event structure returned from
fsnotify_add_notify_event() because that event can be freed by the time
that function returns. Use the mask argument passed into the event
handler directly instead. This also fixes a possible problem when we
could unnecessarily wait for permission response for a normal fanotify
event which got merged with a permission event.
We also disallow merging of permission event with any other event so
that we know the permission event which we just created is the one on
which we should wait for permission response.
Reported-and-tested-by: Jiri Kosina <jkosina@suse.cz>
Reported-and-tested-by: Dave Jones <davej@fedoraproject.org>
Signed-off-by: Jan Kara <jack@suse.cz>
Commit 91c2e0bcae ("unify compat fanotify_mark(2), switch to
COMPAT_SYSCALL_DEFINE") added a new unified compat fanotify_mark syscall
to be used by all architectures.
Unfortunately the unified version merges the split mask parameter in a
wrong way: the lower and higher word got swapped.
This was discovered with glibc's tst-fanotify test case.
Signed-off-by: Heiko Carstens <heiko.carstens@de.ibm.com>
Reported-by: Andreas Krebbel <krebbel@linux.vnet.ibm.com>
Cc: "James E.J. Bottomley" <jejb@parisc-linux.org>
Acked-by: "David S. Miller" <davem@davemloft.net>
Acked-by: Al Viro <viro@ZenIV.linux.org.uk>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Ralf Baechle <ralf@linux-mips.org>
Cc: <stable@vger.kernel.org> [3.10+]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
We usually rely on the fact that struct members not specified in the
initializer are set to NULL. So do that with fsnotify function pointers
as well.
Signed-off-by: Jan Kara <jack@suse.cz>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Cc: Eric Paris <eparis@parisplace.org>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
After removing event structure creation from the generic layer there is
no reason for separate .should_send_event and .handle_event callbacks.
So just remove the first one.
Signed-off-by: Jan Kara <jack@suse.cz>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Cc: Eric Paris <eparis@parisplace.org>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Currently fsnotify framework creates one event structure for each
notification event and links this event into all interested notification
groups. This is done so that we save memory when several notification
groups are interested in the event. However the need for event
structure shared between inotify & fanotify bloats the event structure
so the result is often higher memory consumption.
Another problem is that fsnotify framework keeps path references with
outstanding events so that fanotify can return open file descriptors
with its events. This has the undesirable effect that filesystem cannot
be unmounted while there are outstanding events - a regression for
inotify compared to a situation before it was converted to fsnotify
framework. For fanotify this problem is hard to avoid and users of
fanotify should kind of expect this behavior when they ask for file
descriptors from notified files.
This patch changes fsnotify and its users to create separate event
structure for each group. This allows for much simpler code (~400 lines
removed by this patch) and also smaller event structures. For example
on 64-bit system original struct fsnotify_event consumes 120 bytes, plus
additional space for file name, additional 24 bytes for second and each
subsequent group linking the event, and additional 32 bytes for each
inotify group for private data. After the conversion inotify event
consumes 48 bytes plus space for file name which is considerably less
memory unless file names are long and there are several groups
interested in the events (both of which are uncommon). Fanotify event
fits in 56 bytes after the conversion (fanotify doesn't care about file
names so its events don't have to have it allocated). A win unless
there are four or more fanotify groups interested in the event.
The conversion also solves the problem with unmount when only inotify is
used as we don't have to grab path references for inotify events.
[hughd@google.com: fanotify: fix corruption preventing startup]
Signed-off-by: Jan Kara <jack@suse.cz>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Cc: Eric Paris <eparis@parisplace.org>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Signed-off-by: Hugh Dickins <hughd@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Rounding of name length when passing it to userspace was done in several
places. Provide a function to do it and use it in all places.
Signed-off-by: Jan Kara <jack@suse.cz>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Cc: Eric Paris <eparis@parisplace.org>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
There have been changes in the locking scheme of fsnotify but the
comments in the source code have not been updated yet. This patch
corrects this.
Signed-off-by: Lino Sanfilippo <LinoSanfilippo@gmx.de>
Cc: Eric Paris <eparis@redhat.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
In inotify_new_watch() the number of watches for a group is compared
against the max number of allowed watches and increased afterwards. The
check and incrementation is not done atomically, so it is possible for
multiple concurrent threads to pass the check and increment the number
of marks above the allowed max.
This patch uses an inotify groups mark_lock to ensure that both check
and incrementation are done atomic. Furthermore we dont have to worry
about the race that allows a concurrent thread to add a watch just after
inotify_update_existing_watch() returned with -ENOENT anymore, since
this is also synchronized by the groups mark mutex now.
Signed-off-by: Lino Sanfilippo <LinoSanfilippo@gmx.de>
Cc: Eric Paris <eparis@redhat.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
There is no need to use a special mutex to protect against the
fcntl/close race (see dnotify.c for a description of this race).
Instead the dnotify_groups mark mutex can be used.
Signed-off-by: Lino Sanfilippo <LinoSanfilippo@gmx.de>
Cc: Eric Paris <eparis@redhat.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
The code under the groups mark_mutex in fanotify_add_inode_mark() and
fanotify_add_vfsmount_mark() is almost identical. So put it into a
seperate function.
Signed-off-by: Lino Sanfilippo <LinoSanfilippo@gmx.de>
Cc: Eric Paris <eparis@redhat.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
For both adding an event to an existing mark and destroying a mark we
first have to find it via fsnotify_find_[inode|vfsmount]_mark(). But
getting the mark and adding an event (or destroying it) is not done
atomically. This opens a race where a thread is about to destroy a mark
while another thread still finds the same mark and adds an event to its
mask although it will be destroyed.
Another race exists concerning the excess of a groups number of marks
limit: When a mark is added the number of group marks is checked against
the max number of marks per group and increased afterwards. Since check
and increment is also not done atomically, this may result in 2 or more
processes passing the check at the same time and increasing the number
of group marks above the allowed limit.
With this patch both races are avoided by doing the concerning
operations with the groups mark mutex locked.
Signed-off-by: Lino Sanfilippo <LinoSanfilippo@gmx.de>
Cc: Eric Paris <eparis@redhat.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
The ->reserved field isn't cleared so we leak one byte of stack
information to userspace.
Signed-off-by: Dan Carpenter <dan.carpenter@oracle.com>
Cc: Eric Paris <eparis@redhat.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
... especially since there's no way to get that sucker
on the list fsnotify_fasync() works with - the only thing
adding to it is fsnotify_fasync() itself and it's never
called for fanotify files while they are opened.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
Pull VFS updates from Al Viro,
Misc cleanups all over the place, mainly wrt /proc interfaces (switch
create_proc_entry to proc_create(), get rid of the deprecated
create_proc_read_entry() in favor of using proc_create_data() and
seq_file etc).
7kloc removed.
* 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/viro/vfs: (204 commits)
don't bother with deferred freeing of fdtables
proc: Move non-public stuff from linux/proc_fs.h to fs/proc/internal.h
proc: Make the PROC_I() and PDE() macros internal to procfs
proc: Supply a function to remove a proc entry by PDE
take cgroup_open() and cpuset_open() to fs/proc/base.c
ppc: Clean up scanlog
ppc: Clean up rtas_flash driver somewhat
hostap: proc: Use remove_proc_subtree()
drm: proc: Use remove_proc_subtree()
drm: proc: Use minor->index to label things, not PDE->name
drm: Constify drm_proc_list[]
zoran: Don't print proc_dir_entry data in debug
reiserfs: Don't access the proc_dir_entry in r_open(), r_start() r_show()
proc: Supply an accessor for getting the data from a PDE's parent
airo: Use remove_proc_subtree()
rtl8192u: Don't need to save device proc dir PDE
rtl8187se: Use a dir under /proc/net/r8180/
proc: Add proc_mkdir_data()
proc: Move some bits from linux/proc_fs.h to linux/{of.h,signal.h,tty.h}
proc: Move PDE_NET() to fs/proc/proc_net.c
...
Pull compat cleanup from Al Viro:
"Mostly about syscall wrappers this time; there will be another pile
with patches in the same general area from various people, but I'd
rather push those after both that and vfs.git pile are in."
* 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/viro/signal:
syscalls.h: slightly reduce the jungles of macros
get rid of union semop in sys_semctl(2) arguments
make do_mremap() static
sparc: no need to sign-extend in sync_file_range() wrapper
ppc compat wrappers for add_key(2) and request_key(2) are pointless
x86: trim sys_ia32.h
x86: sys32_kill and sys32_mprotect are pointless
get rid of compat_sys_semctl() and friends in case of ARCH_WANT_OLD_COMPAT_IPC
merge compat sys_ipc instances
consolidate compat lookup_dcookie()
convert vmsplice to COMPAT_SYSCALL_DEFINE
switch getrusage() to COMPAT_SYSCALL_DEFINE
switch epoll_pwait to COMPAT_SYSCALL_DEFINE
convert sendfile{,64} to COMPAT_SYSCALL_DEFINE
switch signalfd{,4}() to COMPAT_SYSCALL_DEFINE
make SYSCALL_DEFINE<n>-generated wrappers do asmlinkage_protect
make HAVE_SYSCALL_WRAPPERS unconditional
consolidate cond_syscall and SYSCALL_ALIAS declarations
teach SYSCALL_DEFINE<n> how to deal with long long/unsigned long long
get rid of duplicate logics in __SC_....[1-6] definitions
When we run the crackerjack testsuite, the inotify_add_watch test is
stalled.
This is caused by the invalid mask 0 - the task is waiting for the event
but it never comes. inotify_add_watch() should return -EINVAL as it did
before commit 676a0675cf ("inotify: remove broken mask checks causing
unmount to be EINVAL"). That commit removes the invalid mask check, but
that check is needed.
Check the mask's ALL_INOTIFY_BITS before the inotify_arg_to_mask() call.
If none are set, just return -EINVAL.
Because IN_UNMOUNT is in ALL_INOTIFY_BITS, this change will not trigger
the problem that above commit fixed.
[akpm@linux-foundation.org: fix build]
Signed-off-by: Zhao Hongjiang <zhaohongjiang@huawei.com>
Acked-by: Jim Somerville <Jim.Somerville@windriver.com>
Cc: Paul Gortmaker <paul.gortmaker@windriver.com>
Cc: Jerome Marchand <jmarchan@redhat.com>
Cc: Eric Paris <eparis@parisplace.org>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Jeff Layton <jlayton@redhat.com>
Cc: John McCutchan <john@johnmccutchan.com>
Cc: Robert Love <rlove@rlove.org>
Cc: Eric Paris <eparis@parisplace.org>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
... and convert a bunch of SYSCALL_DEFINE ones to SYSCALL_DEFINE<n>,
killing the boilerplate crap around them.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
I'm not sure why, but the hlist for each entry iterators were conceived
list_for_each_entry(pos, head, member)
The hlist ones were greedy and wanted an extra parameter:
hlist_for_each_entry(tpos, pos, head, member)
Why did they need an extra pos parameter? I'm not quite sure. Not only
they don't really need it, it also prevents the iterator from looking
exactly like the list iterator, which is unfortunate.
Besides the semantic patch, there was some manual work required:
- Fix up the actual hlist iterators in linux/list.h
- Fix up the declaration of other iterators based on the hlist ones.
- A very small amount of places were using the 'node' parameter, this
was modified to use 'obj->member' instead.
- Coccinelle didn't handle the hlist_for_each_entry_safe iterator
properly, so those had to be fixed up manually.
The semantic patch which is mostly the work of Peter Senna Tschudin is here:
@@
iterator name hlist_for_each_entry, hlist_for_each_entry_continue, hlist_for_each_entry_from, hlist_for_each_entry_rcu, hlist_for_each_entry_rcu_bh, hlist_for_each_entry_continue_rcu_bh, for_each_busy_worker, ax25_uid_for_each, ax25_for_each, inet_bind_bucket_for_each, sctp_for_each_hentry, sk_for_each, sk_for_each_rcu, sk_for_each_from, sk_for_each_safe, sk_for_each_bound, hlist_for_each_entry_safe, hlist_for_each_entry_continue_rcu, nr_neigh_for_each, nr_neigh_for_each_safe, nr_node_for_each, nr_node_for_each_safe, for_each_gfn_indirect_valid_sp, for_each_gfn_sp, for_each_host;
type T;
expression a,c,d,e;
identifier b;
statement S;
@@
-T b;
<+... when != b
(
hlist_for_each_entry(a,
- b,
c, d) S
|
hlist_for_each_entry_continue(a,
- b,
c) S
|
hlist_for_each_entry_from(a,
- b,
c) S
|
hlist_for_each_entry_rcu(a,
- b,
c, d) S
|
hlist_for_each_entry_rcu_bh(a,
- b,
c, d) S
|
hlist_for_each_entry_continue_rcu_bh(a,
- b,
c) S
|
for_each_busy_worker(a, c,
- b,
d) S
|
ax25_uid_for_each(a,
- b,
c) S
|
ax25_for_each(a,
- b,
c) S
|
inet_bind_bucket_for_each(a,
- b,
c) S
|
sctp_for_each_hentry(a,
- b,
c) S
|
sk_for_each(a,
- b,
c) S
|
sk_for_each_rcu(a,
- b,
c) S
|
sk_for_each_from
-(a, b)
+(a)
S
+ sk_for_each_from(a) S
|
sk_for_each_safe(a,
- b,
c, d) S
|
sk_for_each_bound(a,
- b,
c) S
|
hlist_for_each_entry_safe(a,
- b,
c, d, e) S
|
hlist_for_each_entry_continue_rcu(a,
- b,
c) S
|
nr_neigh_for_each(a,
- b,
c) S
|
nr_neigh_for_each_safe(a,
- b,
c, d) S
|
nr_node_for_each(a,
- b,
c) S
|
nr_node_for_each_safe(a,
- b,
c, d) S
|
- for_each_gfn_sp(a, c, d, b) S
+ for_each_gfn_sp(a, c, d) S
|
- for_each_gfn_indirect_valid_sp(a, c, d, b) S
+ for_each_gfn_indirect_valid_sp(a, c, d) S
|
for_each_host(a,
- b,
c) S
|
for_each_host_safe(a,
- b,
c, d) S
|
for_each_mesh_entry(a,
- b,
c, d) S
)
...+>
[akpm@linux-foundation.org: drop bogus change from net/ipv4/raw.c]
[akpm@linux-foundation.org: drop bogus hunk from net/ipv6/raw.c]
[akpm@linux-foundation.org: checkpatch fixes]
[akpm@linux-foundation.org: fix warnings]
[akpm@linux-foudnation.org: redo intrusive kvm changes]
Tested-by: Peter Senna Tschudin <peter.senna@gmail.com>
Acked-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Signed-off-by: Sasha Levin <sasha.levin@oracle.com>
Cc: Wu Fengguang <fengguang.wu@intel.com>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Cc: Gleb Natapov <gleb@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Convert to the much saner new idr interface.
Note that the adhoc cyclic id allocation is buggy. If wraparound
happens, the previous code with idr_get_new_above() may segfault and
the converted code will trigger WARN and return -EINVAL. Even if it's
fixed to wrap to zero, the code will be prone to unnecessary -ENOSPC
failures after the first wraparound. We probably need to implement
proper cyclic support in idr.
Signed-off-by: Tejun Heo <tj@kernel.org>
Cc: John McCutchan <john@johnmccutchan.com>
Cc: Robert Love <rlove@rlove.org>
Cc: Eric Paris <eparis@parisplace.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
idr_destroy() can destroy idr by itself and idr_remove_all() is being
deprecated. Drop its usage.
Signed-off-by: Tejun Heo <tj@kernel.org>
Cc: John McCutchan <john@johnmccutchan.com>
Cc: Robert Love <rlove@rlove.org>
Cc: Eric Paris <eparis@parisplace.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Pull vfs pile (part one) from Al Viro:
"Assorted stuff - cleaning namei.c up a bit, fixing ->d_name/->d_parent
locking violations, etc.
The most visible changes here are death of FS_REVAL_DOT (replaced with
"has ->d_weak_revalidate()") and a new helper getting from struct file
to inode. Some bits of preparation to xattr method interface changes.
Misc patches by various people sent this cycle *and* ocfs2 fixes from
several cycles ago that should've been upstream right then.
PS: the next vfs pile will be xattr stuff."
* 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/viro/vfs: (46 commits)
saner proc_get_inode() calling conventions
proc: avoid extra pde_put() in proc_fill_super()
fs: change return values from -EACCES to -EPERM
fs/exec.c: make bprm_mm_init() static
ocfs2/dlm: use GFP_ATOMIC inside a spin_lock
ocfs2: fix possible use-after-free with AIO
ocfs2: Fix oops in ocfs2_fast_symlink_readpage() code path
get_empty_filp()/alloc_file() leave both ->f_pos and ->f_version zero
target: writev() on single-element vector is pointless
export kernel_write(), convert open-coded instances
fs: encode_fh: return FILEID_INVALID if invalid fid_type
kill f_vfsmnt
vfs: kill FS_REVAL_DOT by adding a d_weak_revalidate dentry op
nfsd: handle vfs_getattr errors in acl protocol
switch vfs_getattr() to struct path
default SET_PERSONALITY() in linux/elf.h
ceph: prepopulate inodes only when request is aborted
d_hash_and_lookup(): export, switch open-coded instances
9p: switch v9fs_set_create_acl() to inode+fid, do it before d_instantiate()
9p: split dropping the acls from v9fs_set_create_acl()
...
Running the command:
inotifywait -e unmount /mnt/disk
immediately aborts with a -EINVAL return code. This is however a valid
parameter. This abort occurs only if unmount is the sole event
parameter. If other event parameters are supplied, then the unmount
event wait will work.
The problem was introduced by commit 44b350fc23 ("inotify: Fix mask
checks"). In that commit, it states:
The mask checks in inotify_update_existing_watch() and
inotify_new_watch() are useless because inotify_arg_to_mask()
sets FS_IN_IGNORED and FS_EVENT_ON_CHILD bits anyway.
But instead of removing the useless checks, it did this:
mask = inotify_arg_to_mask(arg);
- if (unlikely(!mask))
+ if (unlikely(!(mask & IN_ALL_EVENTS)))
return -EINVAL;
The problem is that IN_ALL_EVENTS doesn't include IN_UNMOUNT, and other
parts of the code keep IN_UNMOUNT separate from IN_ALL_EVENTS. So the
check should be:
if (unlikely(!(mask & (IN_ALL_EVENTS | IN_UNMOUNT))))
But inotify_arg_to_mask(arg) always sets the IN_UNMOUNT bit in the mask
anyway, so the check is always going to pass and thus should simply be
removed. Also note that inotify_arg_to_mask completely controls what
mask bits get set from arg, there's no way for invalid bits to get
enabled there.
Lets fix it by simply removing the useless broken checks.
Signed-off-by: Jim Somerville <Jim.Somerville@windriver.com>
Signed-off-by: Paul Gortmaker <paul.gortmaker@windriver.com>
Cc: Jerome Marchand <jmarchan@redhat.com>
Cc: John McCutchan <john@johnmccutchan.com>
Cc: Robert Love <rlove@rlove.org>
Cc: Eric Paris <eparis@parisplace.org>
Cc: <stable@vger.kernel.org> [2.6.37+]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Pull filesystem notification updates from Eric Paris:
"This pull mostly is about locking changes in the fsnotify system. By
switching the group lock from a spin_lock() to a mutex() we can now
hold the lock across things like iput(). This fixes a problem
involving unmounting a fs and having inodes be busy, first pointed out
by FAT, but reproducible with tmpfs.
This also restores signal driven I/O for inotify, which has been
broken since about 2.6.32."
Ugh. I *hate* the timing of this. It was rebased after the merge
window opened, and then left to sit with the pull request coming the day
before the merge window closes. That's just crap. But apparently the
patches themselves have been around for over a year, just gathering
dust, so now it's suddenly critical.
Fixed up semantic conflict in fs/notify/fdinfo.c as per Stephen
Rothwell's fixes from -next.
* 'for-next' of git://git.infradead.org/users/eparis/notify:
inotify: automatically restart syscalls
inotify: dont skip removal of watch descriptor if creation of ignored event failed
fanotify: dont merge permission events
fsnotify: make fasync generic for both inotify and fanotify
fsnotify: change locking order
fsnotify: dont put marks on temporary list when clearing marks by group
fsnotify: introduce locked versions of fsnotify_add_mark() and fsnotify_remove_mark()
fsnotify: pass group to fsnotify_destroy_mark()
fsnotify: use a mutex instead of a spinlock to protect a groups mark list
fanotify: add an extra flag to mark_remove_from_mask that indicates wheather a mark should be destroyed
fsnotify: take groups mark_lock before mark lock
fsnotify: use reference counting for groups
fsnotify: introduce fsnotify_get_group()
inotify, fanotify: replace fsnotify_put_group() with fsnotify_destroy_group()
The kernel keeps FAN_MARK_IGNORED_SURV_MODIFY bit separately from
fsnotify_mark::mask|ignored_mask thus put it in @mflags (mark flags)
field so the user-space reader will be able to detect if such bit were
used on mark creation procedure.
| pos: 0
| flags: 04002
| fanotify flags:10 event-flags:0
| fanotify mnt_id:12 mflags:40 mask:38 ignored_mask:40000003
| fanotify ino:4f969 sdev:800013 mflags:0 mask:3b ignored_mask:40000000 fhandle-bytes:8 fhandle-type:1 f_handle:69f90400c275b5b4
Signed-off-by: Cyrill Gorcunov <gorcunov@openvz.org>
Cc: Pavel Emelyanov <xemul@parallels.com>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Andrey Vagin <avagin@openvz.org>
Cc: Al Viro <viro@ZenIV.linux.org.uk>
Cc: Alexey Dobriyan <adobriyan@gmail.com>
Cc: James Bottomley <jbottomley@parallels.com>
Cc: "Aneesh Kumar K.V" <aneesh.kumar@linux.vnet.ibm.com>
Cc: Matthew Helsley <matt.helsley@gmail.com>
Cc: "J. Bruce Fields" <bfields@fieldses.org>
Cc: Tvrtko Ursulin <tvrtko.ursulin@onelan.co.uk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Fixes following sparse warning:
fs/notify/inode_mark.c:127:22: warning: symbol 'fsnotify_find_inode_mark_locked' was not declared. Should it be static?
Signed-off-by: Tushar Behera <tushar.behera@linaro.org>
Cc: Eric Paris <eparis@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Pull trivial branch from Jiri Kosina:
"Usual stuff -- comment/printk typo fixes, documentation updates, dead
code elimination."
* 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/jikos/trivial: (39 commits)
HOWTO: fix double words typo
x86 mtrr: fix comment typo in mtrr_bp_init
propagate name change to comments in kernel source
doc: Update the name of profiling based on sysfs
treewide: Fix typos in various drivers
treewide: Fix typos in various Kconfig
wireless: mwifiex: Fix typo in wireless/mwifiex driver
messages: i2o: Fix typo in messages/i2o
scripts/kernel-doc: check that non-void fcts describe their return value
Kernel-doc: Convention: Use a "Return" section to describe return values
radeon: Fix typo and copy/paste error in comments
doc: Remove unnecessary declarations from Documentation/accounting/getdelays.c
various: Fix spelling of "asynchronous" in comments.
Fix misspellings of "whether" in comments.
eisa: Fix spelling of "asynchronous".
various: Fix spelling of "registered" in comments.
doc: fix quite a few typos within Documentation
target: iscsi: fix comment typos in target/iscsi drivers
treewide: fix typo of "suport" in various comments and Kconfig
treewide: fix typo of "suppport" in various comments
...
We were mistakenly returning EINTR when we found an outstanding signal.
Instead we should returen ERESTARTSYS and allow the kernel to handle
things the right way.
Patch-from: Oleg Nesterov <oleg@redhat.com>
Signed-off-by: Eric Paris <eparis@redhat.com>
In inotify_ignored_and_remove_idr() the removal of a watch descriptor is skipped
if the allocation of an ignored event failed and we are leaking memory (the
watch descriptor and the mark linked to it).
This patch ensures that the watch descriptor is removed regardless of whether
event creation failed or not.
Signed-off-by: Lino Sanfilippo <LinoSanfilippo@gmx.de>
Signed-off-by: Eric Paris <eparis@redhat.com>
Boyd Yang reported a problem for the case that multiple threads of the same
thread group are waiting for a reponse for a permission event.
In this case it is possible that some of the threads are never woken up, even
if the response for the event has been received
(see http://marc.info/?l=linux-kernel&m=131822913806350&w=2).
The reason is that we are currently merging permission events if they belong to
the same thread group. But we are not prepared to wake up more than one waiter
for each event. We do
wait_event(group->fanotify_data.access_waitq, event->response ||
atomic_read(&group->fanotify_data.bypass_perm));
and after that
event->response = 0;
which is the reason that even if we woke up all waiters for the same event
some of them may see event->response being already set 0 again, then go back to
sleep and block forever.
With this patch we avoid that more than one thread is waiting for a response
by not merging permission events for the same thread group any more.
Reported-by: Boyd Yang <boyd.yang@gmail.com>
Signed-off-by: Lino Sanfilippo <LinoSanfilipp@gmx.de>
Signed-off-by: Eric Paris <eparis@redhat.com>
inotify is supposed to support async signal notification when information
is available on the inotify fd. This patch moves that support to generic
fsnotify functions so it can be used by all notification mechanisms.
Signed-off-by: Eric Paris <eparis@redhat.com>
In clear_marks_by_group_flags() the mark list of a group is iterated and the
marks are put on a temporary list.
Since we introduced fsnotify_destroy_mark_locked() we dont need the temp list
any more and are able to remove the marks while the mark list is iterated and
the mark list mutex is held.
Signed-off-by: Lino Sanfilippo <LinoSanfilippo@gmx.de>
Signed-off-by: Eric Paris <eparis@redhat.com>
This patch introduces fsnotify_add_mark_locked() and fsnotify_remove_mark_locked()
which are essentially the same as fsnotify_add_mark() and fsnotify_remove_mark() but
assume that the caller has already taken the groups mark mutex.
Signed-off-by: Lino Sanfilippo <LinoSanfilippo@gmx.de>
Signed-off-by: Eric Paris <eparis@redhat.com>
In fsnotify_destroy_mark() dont get the group from the passed mark anymore,
but pass the group itself as an additional parameter to the function.
Signed-off-by: Lino Sanfilippo <LinoSanfilippo@gmx.de>
Signed-off-by: Eric Paris <eparis@redhat.com>
Replaces the groups mark_lock spinlock with a mutex. Using a mutex instead
of a spinlock results in more flexibility (i.e it allows to sleep while the
lock is held).
Signed-off-by: Lino Sanfilippo <LinoSanfilippo@gmx.de>
Signed-off-by: Eric Paris <eparis@redhat.com>
This patch adds an extra flag to mark_remove_from_mask() to inform the caller if
the mark should be destroyed.
With this we dont destroy the mark implicitly in the function itself any more
but let the caller handle it.
Signed-off-by: Lino Sanfilippo <LinoSanfilippo@gmx.de>
Signed-off-by: Eric Paris <eparis@redhat.com>
Race-free addition and removal of a mark to a groups mark list would be easier
if we could lock the mark list of group before we lock the specific mark.
This patch changes the order used to add/remove marks to/from mark lists from
1. mark->lock
2. group->mark_lock
3. inode->i_lock
to
1. group->mark_lock
2. mark->lock
3. inode->i_lock
Signed-off-by: Lino Sanfilippo <LinoSanfilippo@gmx.de>
Signed-off-by: Eric Paris <eparis@redhat.com>
Get a group ref for each mark that is added to the groups list and release that
ref when the mark is freed in fsnotify_put_mark().
We also use get a group reference for duplicated marks and for private event
data.
Now we dont free a group any more when the number of marks becomes 0 but when
the groups ref count does. Since this will only happen when all marks are removed
from a groups mark list, we dont have to set the groups number of marks to 1 at
group creation.
Beside clearing all marks in fsnotify_destroy_group() we do also flush the
groups event queue. This is since events may hold references to groups (due to
private event data) and we have to put those references first before we get a
chance to put the final ref, which will result in a call to
fsnotify_final_destroy_group().
Signed-off-by: Lino Sanfilippo <LinoSanfilippo@gmx.de>
Signed-off-by: Eric Paris <eparis@redhat.com>
Introduce fsnotify_get_group() which increments the reference counter of a group.
Signed-off-by: Lino Sanfilippo <LinoSanfilippo@gmx.de>
Signed-off-by: Eric Paris <eparis@redhat.com>
Currently in fsnotify_put_group() the ref count of a group is decremented and if
it becomes 0 fsnotify_destroy_group() is called. Since a groups ref count is only
at group creation set to 1 and never increased after that a call to fsnotify_put_group()
always results in a call to fsnotify_destroy_group().
With this patch fsnotify_destroy_group() is called directly.
Signed-off-by: Lino Sanfilippo <LinoSanfilippo@gmx.de>
Signed-off-by: Eric Paris <eparis@redhat.com>
"Asynchronous" is misspelled in some comments. No code changes.
Signed-off-by: Adam Buchbinder <adam.buchbinder@gmail.com>
Signed-off-by: Jiri Kosina <jkosina@suse.cz>
If the FAN_Q_OVERFLOW bit set in event->mask, the fanotify event
metadata will not contain a valid file descriptor, but
copy_event_to_user() didn't check for that, and unconditionally does a
fd_install() on the file descriptor.
Which in turn will cause a BUG_ON() in __fd_install().
Introduced by commit 352e3b2492 ("fanotify: sanitize failure exits in
copy_event_to_user()")
Mea culpa - missed that path ;-/
Reported-by: Alex Shi <lkml.alex@gmail.com>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Anders Blomdell noted in 2010 that Fanotify lost events and provided a
test case. Eric Paris confirmed it was a bug and posted a fix to the
list
https://groups.google.com/forum/?fromgroups=#!topic/linux.kernel/RrJfTfyW2BE
but never applied it. Repeated attempts over time to actually get him
to apply it have never had a reply from anyone who has raised it
So apply it anyway
Signed-off-by: Alan Cox <alan@linux.intel.com>
Reported-by: Anders Blomdell <anders.blomdell@control.lth.se>
Cc: Eric Paris <eparis@redhat.com>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
* do copy_to_user() before prepare_for_access_response(); that kills
the need in remove_access_response().
* don't do fd_install() until we are past the last possible failure
exit. Don't use sys_close() on cleanup side - just put_unused_fd()
and fput(). Less racy that way...
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
We don't use "mnt" anymore in send_to_group() after 1968f5eed5 ("fanotify:
use both marks when possible") was applied.
Signed-off-by: Dan Carpenter <dan.carpenter@oracle.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
Removing the parent of a watched file results in "kernel BUG at
fs/notify/mark.c:139".
To reproduce
add "-w /tmp/audit/dir/watched_file" to audit.rules
rm -rf /tmp/audit/dir
This is caused by fsnotify_destroy_mark() being called without an
extra reference taken by the caller.
Reported by Francesco Cosoleto here:
https://bugzilla.novell.com/show_bug.cgi?id=689860
Fix by removing the BUG_ON and adding a comment about not accessing mark after
the iput.
Signed-off-by: Miklos Szeredi <mszeredi@suse.cz>
CC: stable@vger.kernel.org
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
This allows us to move duplicated code in <asm/atomic.h>
(atomic_inc_not_zero() for now) to <linux/atomic.h>
Signed-off-by: Arun Sharma <asharma@fb.com>
Reviewed-by: Eric Dumazet <eric.dumazet@gmail.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: David Miller <davem@davemloft.net>
Cc: Eric Dumazet <eric.dumazet@gmail.com>
Acked-by: Mike Frysinger <vapier@gentoo.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
On an error path in inotify_init1 a normal user can trigger a double
free of struct user. This is a regression introduced by a2ae4cc9a1
("inotify: stop kernel memory leak on file creation failure").
We fix this by making sure that if a group exists the user reference is
dropped when the group is cleaned up. We should not explictly drop the
reference on error and also drop the reference when the group is cleaned
up.
The new lifetime rules are that an inotify group lives from
inotify_new_group to the last fsnotify_put_group. Since the struct user
and inotify_devs are directly tied to this lifetime they are only
changed/updated in those two locations. We get rid of all special
casing of struct user or user->inotify_devs.
Signed-off-by: Eric Paris <eparis@redhat.com>
Cc: stable@kernel.org (2.6.37 and up)
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
All that remains of the inode_lock is protecting the inode hash list
manipulation and traversals. Rename the inode_lock to
inode_hash_lock to reflect it's actual function.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
Protect the per-sb inode list with a new global lock
inode_sb_list_lock and use it to protect the list manipulations and
traversals. This lock replaces the inode_lock as the inodes on the
list can be validity checked while holding the inode->i_lock and
hence the inode_lock is no longer needed to protect the list.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
Protect inode state transitions and validity checks with the
inode->i_lock. This enables us to make inode state transitions
independently of the inode_lock and is the first step to peeling
away the inode_lock from the code.
This requires that __iget() is done atomically with i_state checks
during list traversals so that we don't race with another thread
marking the inode I_FREEING between the state check and grabbing the
reference.
Also remove the unlock_new_inode() memory barrier optimisation
required to avoid taking the inode_lock when clearing I_NEW.
Simplify the code by simply taking the inode->i_lock around the
state change and wakeup. Because the wakeup is no longer tricky,
remove the wake_up_inode() function and open code the wakeup where
necessary.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
dcache_inode_lock can be replaced with per-inode locking. Use existing
inode->i_lock for this. This is slightly non-trivial because we sometimes
need to find the inode from the dentry, which requires d_inode to be
stabilised (either with refcount or d_lock).
Signed-off-by: Nick Piggin <npiggin@kernel.dk>
Add a new lock, dcache_inode_lock, to protect the inode's i_dentry list
from concurrent modification. d_alias is also protected by d_lock.
Signed-off-by: Nick Piggin <npiggin@kernel.dk>
Protect d_subdirs and d_child with d_lock, except in filesystems that aren't
using dcache_lock for these anyway (eg. using i_mutex).
Note: if we change the locking rule in future so that ->d_child protection is
provided only with ->d_parent->d_lock, it may allow us to reduce some locking.
But it would be an exception to an otherwise regular locking scheme, so we'd
have to see some good results. Probably not worthwhile.
Signed-off-by: Nick Piggin <npiggin@kernel.dk>
Conflicts:
MAINTAINERS
arch/arm/mach-omap2/pm24xx.c
drivers/scsi/bfa/bfa_fcpim.c
Needed to update to apply fixes for which the old branch was too
outdated.
The fanotify_event_metadata now has a field which is supposed to
indicate the length of the metadata portion of the event. Fill in that
field as well.
Based-in-part-on-patch-by: Alexey Zaytsev <alexey.zaytsev@gmail.com>
Signed-off-by: Eric Paris <eparis@redhat.com>
We should not try to open a file descriptor for the overflow event since this
will always fail.
Signed-off-by: Lino Sanfilippo <LinoSanfilippo@gmx.de>
Signed-off-by: Eric Paris <eparis@redhat.com>
If fanotify_init is unable to allocate a new fsnotify group it will
return but will not drop its reference on the associated user struct.
Drop that reference on error.
Reported-by: Vegard Nossum <vegard.nossum@gmail.com>
Signed-off-by: Eric Paris <eparis@redhat.com>
If inotify_init is unable to allocate a new file for the new inotify
group we leak the new group. This patch drops the reference on the
group on file allocation failure.
Reported-by: Vegard Nossum <vegard.nossum@gmail.com>
cc: stable@kernel.org
Signed-off-by: Eric Paris <eparis@redhat.com>
When fanotify_release() is called, there may still be processes waiting for
access permission. Currently only processes for which an event has already been
queued into the groups access list will be woken up. Processes for which no
event has been queued will continue to sleep and thus cause a deadlock when
fsnotify_put_group() is called.
Furthermore there is a race allowing further processes to be waiting on the
access wait queue after wake_up (if they arrive before clear_marks_by_group()
is called).
This patch corrects this by setting a flag to inform processes that the group
is about to be destroyed and thus not to wait for access permission.
[additional changelog from eparis]
Lets think about the 4 relevant code paths from the PoV of the
'operator' 'listener' 'responder' and 'closer'. Where operator is the
process doing an action (like open/read) which could require permission.
Listener is the task (or in this case thread) slated with reading from
the fanotify file descriptor. The 'responder' is the thread responsible
for responding to access requests. 'Closer' is the thread attempting to
close the fanotify file descriptor.
The 'operator' is going to end up in:
fanotify_handle_event()
get_response_from_access()
(THIS BLOCKS WAITING ON USERSPACE)
The 'listener' interesting code path
fanotify_read()
copy_event_to_user()
prepare_for_access_response()
(THIS CREATES AN fanotify_response_event)
The 'responder' code path:
fanotify_write()
process_access_response()
(REMOVE A fanotify_response_event, SET RESPONSE, WAKE UP 'operator')
The 'closer':
fanotify_release()
(SUPPOSED TO CLEAN UP THE REST OF THIS MESS)
What we have today is that in the closer we remove all of the
fanotify_response_events and set a bit so no more response events are
ever created in prepare_for_access_response().
The bug is that we never wake all of the operators up and tell them to
move along. You fix that in fanotify_get_response_from_access(). You
also fix other operators which haven't gotten there yet. So I agree
that's a good fix.
[/additional changelog from eparis]
[remove additional changes to minimize patch size]
[move initialization so it was inside CONFIG_FANOTIFY_PERMISSION]
Signed-off-by: Lino Sanfilippo <LinoSanfilippo@gmx.de>
Signed-off-by: Eric Paris <eparis@redhat.com>
In mark_remove_from_mask() we destroy marks that have their event mask cleared.
Thus we should not allow the creation of those marks in the first place.
With this patch we check if the mask given from user is 0 in case of FAN_MARK_ADD.
If so we return an error. Same for FAN_MARK_REMOVE since this does not have any
effect.
Signed-off-by: Lino Sanfilippo <LinoSanfilippo@gmx.de>
Signed-off-by: Eric Paris <eparis@redhat.com>
If adding a mount or inode mark failed fanotify_free_mark() is called explicitly.
But at this time the mark has already been put into the destroy list of the
fsnotify_mark kernel thread. If the thread is too slow it will try to decrease
the reference of a mark, that has already been freed by fanotify_free_mark().
(If its fast enough it will only decrease the marks ref counter from 2 to 1 - note
that the counter has been increased to 2 in add_mark() - which has practically no
effect.)
This patch fixes the ref counting by not calling free_mark() explicitly, but
decreasing the ref counter and rely on the fsnotify_mark thread to cleanup in
case adding the mark has failed.
Signed-off-by: Lino Sanfilippo <LinoSanfilippo@gmx.de>
Signed-off-by: Eric Paris <eparis@redhat.com>
If no event was sent to userspace we cannot expect userspace to respond to
permissions requests. Today such requests just hang forever. This patch will
deny any permissions event which was unable to be sent to userspace.
Reported-by: Tvrtko Ursulin <tvrtko.ursulin@sophos.com>
Signed-off-by: Eric Paris <eparis@redhat.com>
In fanotify_read() return -ERESTARTSYS instead of -EINTR to
make read() restartable across signals (BSD semantic).
Signed-off-by: Eric Paris <eparis@redhat.com>
fs/notify/fanotify/fanotify_user.c: In function 'fanotify_release':
fs/notify/fanotify/fanotify_user.c:375: warning: unused variable 'lre'
fs/notify/fanotify/fanotify_user.c:375: warning: unused variable 're'
this is really ugly.
Cc: Eric Paris <eparis@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Eric Paris <eparis@redhat.com>
If fanotify sets a new bit in the ignored mask it will cause the generic
fsnotify layer to recalculate the real mask. This is stupid since we
didn't change that part.
Signed-off-by: Eric Paris <eparis@redhat.com>
fanotify has a very limited number of events it sends on directories. The
usefulness of these events is yet to be seen and still we send them. This
is particularly painful for mount marks where one might receive many of
these useless events. As such this patch will drop events on IS_DIR()
inodes unless they were explictly requested with FAN_ON_DIR.
This means that a mark on a directory without FAN_EVENT_ON_CHILD or
FAN_ON_DIR is meaningless and will result in no events ever (although it
will still be allowed since detecting it is hard)
Signed-off-by: Eric Paris <eparis@redhat.com>
The _IN_ in the naming is reserved for flags only used by inotify. Since I
am about to use this flag for fanotify rename it to be generic like the
rest.
Signed-off-by: Eric Paris <eparis@redhat.com>
fanotify_should_send_event has a test to see if an object is a file or
directory and does not send an event otherwise. The problem is that the
test is actually checking if the object with a mark is a file or directory,
not if the object the event happened on is a file or directory. We should
check the latter.
Signed-off-by: Eric Paris <eparis@redhat.com>
fanotify currently has no limit on the number of listeners a given user can
have open. This patch limits the total number of listeners per user to
128. This is the same as the inotify default limit.
Signed-off-by: Eric Paris <eparis@redhat.com>
Some fanotify groups, especially those like AV scanners, will need to place
lots of marks, particularly ignore marks. Since ignore marks do not pin
inodes in cache and are cleared if the inode is removed from core (usually
under memory pressure) we expose an interface for listeners, with
CAP_SYS_ADMIN, to override the maximum number of marks and be allowed to
set and 'unlimited' number of marks. Programs which make use of this
feature will be able to OOM a machine.
Signed-off-by: Eric Paris <eparis@redhat.com>
There is currently no limit on the number of marks a given fanotify group
can have. Since fanotify is gated on CAP_SYS_ADMIN this was not seen as
a serious DoS threat. This patch implements a default of 8192, the same as
inotify to work towards removing the CAP_SYS_ADMIN gating and eliminating
the default DoS'able status.
Signed-off-by: Eric Paris <eparis@redhat.com>
fanotify has a defualt max queue depth. This patch allows processes which
explicitly request it to have an 'unlimited' queue depth. These processes
need to be very careful to make sure they cannot fall far enough behind
that they OOM the box. Thus this flag is gated on CAP_SYS_ADMIN.
Signed-off-by: Eric Paris <eparis@redhat.com>
Currently fanotify has no maximum queue depth. Since fanotify is
CAP_SYS_ADMIN only this does not pose a normal user DoS issue, but it
certianly is possible that an fanotify listener which can't keep up could
OOM the box. This patch implements a default 16k depth. This is the same
default depth used by inotify, but given fanotify's better queue merging in
many situations this queue will contain many additional useful events by
comparison.
Signed-off-by: Eric Paris <eparis@redhat.com>
fanotify will clear ignore marks if a task changes the contents of an
inode. The problem is with the races around when userspace finishes
checking a file and when that result is actually attached to the inode.
This race was described as such:
Consider the following scenario with hostile processes A and B, and
victim process C:
1. Process A opens new file for writing. File check request is generated.
2. File check is performed in userspace. Check result is "file has no malware".
3. The "permit" response is delivered to kernel space.
4. File ignored mark set.
5. Process A writes dummy bytes to the file. File ignored flags are cleared.
6. Process B opens the same file for reading. File check request is generated.
7. File check is performed in userspace. Check result is "file has no malware".
8. Process A writes malware bytes to the file. There is no cached response yet.
9. The "permit" response is delivered to kernel space and is cached in fanotify.
10. File ignored mark set.
11. Now any process C will be permitted to open the malware file.
There is a race between steps 8 and 10
While fanotify makes no strong guarantees about systems with hostile
processes there is no reason we cannot harden against this race. We do
that by simply ignoring any ignore marks if the inode has open writers (aka
i_writecount > 0). (We actually do not ignore ignore marks if the
FAN_MARK_SURV_MODIFY flag is set)
Reported-by: Vasily Novikov <vasily.novikov@kaspersky.com>
Signed-off-by: Eric Paris <eparis@redhat.com>
fsnotify perm events do not call fsnotify parent. That means you cannot
register a perm event on a directory and enforce permissions on all inodes in
that directory. This patch fixes that situation.
Signed-off-by: Eric Paris <eparis@redhat.com>
When fsnotify groups return errors they are ignored. For permissions
events these should be passed back up the stack, but for most events these
should continue to be ignored.
Signed-off-by: Eric Paris <eparis@redhat.com>
The fanotify listeners needs to be able to specify what types of operations
they are going to perform so they can be ordered appropriately between other
listeners doing other types of operations. They need this to be able to make
sure that things like hierarchichal storage managers will get access to inodes
before processes which need the data. This patch defines 3 possible uses
which groups must indicate in the fanotify_init() flags.
FAN_CLASS_PRE_CONTENT
FAN_CLASS_CONTENT
FAN_CLASS_NOTIF
Groups will receive notification in that order. The order between 2 groups in
the same class is undeterministic.
FAN_CLASS_PRE_CONTENT is intended to be used by listeners which need access to
the inode before they are certain that the inode contains it's final data. A
hierarchical storage manager should choose to use this class.
FAN_CLASS_CONTENT is intended to be used by listeners which need access to the
inode after it contains its intended contents. This would be the appropriate
level for an AV solution or document control system.
FAN_CLASS_NOTIF is intended for normal async notification about access, much the
same as inotify and dnotify. Syncronous permissions events are not permitted
at this class.
Signed-off-by: Eric Paris <eparis@redhat.com>
fanotify needs to be able to specify that some groups get events before
others. They use this idea to make sure that a hierarchical storage
manager gets access to files before programs which actually use them. This
is purely infrastructure. Everything will have a priority of 0, but the
infrastructure will exist for it to be non-zero.
Signed-off-by: Eric Paris <eparis@redhat.com>
We disabled the ability to build fanotify in commit 7c5347733d.
This reverts that commit and allows people to build fanotify.
Signed-off-by: Eric Paris <eparis@redhat.com>
Pull removal of fsnotify marks into generic_shutdown_super().
Split umount-time work into a new function - evict_inodes().
Make sure that invalidate_inodes() will be able to cope with
I_FREEING once we change locking in iput().
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
Use dget_parent instead of opencoding it. This simplifies the code, but
more importanly prepares for the more complicated locking for a parent
dget in the dcache scale patch series.
It means we do grab a reference to the parent now if need to be watched,
but not with the specified mask. If this turns out to be a problem
we'll have to revisit it, but for now let's keep as much as possible
dcache internals inside dcache.[ch].
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
* 'llseek' of git://git.kernel.org/pub/scm/linux/kernel/git/arnd/bkl:
vfs: make no_llseek the default
vfs: don't use BKL in default_llseek
llseek: automatically add .llseek fop
libfs: use generic_file_llseek for simple_attr
mac80211: disallow seeks in minstrel debug code
lirc: make chardev nonseekable
viotape: use noop_llseek
raw: use explicit llseek file operations
ibmasmfs: use generic_file_llseek
spufs: use llseek in all file operations
arm/omap: use generic_file_llseek in iommu_debug
lkdtm: use generic_file_llseek in debugfs
net/wireless: use generic_file_llseek in debugfs
drm: use noop_llseek
All file_operations should get a .llseek operation so we can make
nonseekable_open the default for future file operations without a
.llseek pointer.
The three cases that we can automatically detect are no_llseek, seq_lseek
and default_llseek. For cases where we can we can automatically prove that
the file offset is always ignored, we use noop_llseek, which maintains
the current behavior of not returning an error from a seek.
New drivers should normally not use noop_llseek but instead use no_llseek
and call nonseekable_open at open time. Existing drivers can be converted
to do the same when the maintainer knows for certain that no user code
relies on calling seek on the device file.
The generated code is often incorrectly indented and right now contains
comments that clarify for each added line why a specific variant was
chosen. In the version that gets submitted upstream, the comments will
be gone and I will manually fix the indentation, because there does not
seem to be a way to do that using coccinelle.
Some amount of new code is currently sitting in linux-next that should get
the same modifications, which I will do at the end of the merge window.
Many thanks to Julia Lawall for helping me learn to write a semantic
patch that does all this.
===== begin semantic patch =====
// This adds an llseek= method to all file operations,
// as a preparation for making no_llseek the default.
//
// The rules are
// - use no_llseek explicitly if we do nonseekable_open
// - use seq_lseek for sequential files
// - use default_llseek if we know we access f_pos
// - use noop_llseek if we know we don't access f_pos,
// but we still want to allow users to call lseek
//
@ open1 exists @
identifier nested_open;
@@
nested_open(...)
{
<+...
nonseekable_open(...)
...+>
}
@ open exists@
identifier open_f;
identifier i, f;
identifier open1.nested_open;
@@
int open_f(struct inode *i, struct file *f)
{
<+...
(
nonseekable_open(...)
|
nested_open(...)
)
...+>
}
@ read disable optional_qualifier exists @
identifier read_f;
identifier f, p, s, off;
type ssize_t, size_t, loff_t;
expression E;
identifier func;
@@
ssize_t read_f(struct file *f, char *p, size_t s, loff_t *off)
{
<+...
(
*off = E
|
*off += E
|
func(..., off, ...)
|
E = *off
)
...+>
}
@ read_no_fpos disable optional_qualifier exists @
identifier read_f;
identifier f, p, s, off;
type ssize_t, size_t, loff_t;
@@
ssize_t read_f(struct file *f, char *p, size_t s, loff_t *off)
{
... when != off
}
@ write @
identifier write_f;
identifier f, p, s, off;
type ssize_t, size_t, loff_t;
expression E;
identifier func;
@@
ssize_t write_f(struct file *f, const char *p, size_t s, loff_t *off)
{
<+...
(
*off = E
|
*off += E
|
func(..., off, ...)
|
E = *off
)
...+>
}
@ write_no_fpos @
identifier write_f;
identifier f, p, s, off;
type ssize_t, size_t, loff_t;
@@
ssize_t write_f(struct file *f, const char *p, size_t s, loff_t *off)
{
... when != off
}
@ fops0 @
identifier fops;
@@
struct file_operations fops = {
...
};
@ has_llseek depends on fops0 @
identifier fops0.fops;
identifier llseek_f;
@@
struct file_operations fops = {
...
.llseek = llseek_f,
...
};
@ has_read depends on fops0 @
identifier fops0.fops;
identifier read_f;
@@
struct file_operations fops = {
...
.read = read_f,
...
};
@ has_write depends on fops0 @
identifier fops0.fops;
identifier write_f;
@@
struct file_operations fops = {
...
.write = write_f,
...
};
@ has_open depends on fops0 @
identifier fops0.fops;
identifier open_f;
@@
struct file_operations fops = {
...
.open = open_f,
...
};
// use no_llseek if we call nonseekable_open
////////////////////////////////////////////
@ nonseekable1 depends on !has_llseek && has_open @
identifier fops0.fops;
identifier nso ~= "nonseekable_open";
@@
struct file_operations fops = {
... .open = nso, ...
+.llseek = no_llseek, /* nonseekable */
};
@ nonseekable2 depends on !has_llseek @
identifier fops0.fops;
identifier open.open_f;
@@
struct file_operations fops = {
... .open = open_f, ...
+.llseek = no_llseek, /* open uses nonseekable */
};
// use seq_lseek for sequential files
/////////////////////////////////////
@ seq depends on !has_llseek @
identifier fops0.fops;
identifier sr ~= "seq_read";
@@
struct file_operations fops = {
... .read = sr, ...
+.llseek = seq_lseek, /* we have seq_read */
};
// use default_llseek if there is a readdir
///////////////////////////////////////////
@ fops1 depends on !has_llseek && !nonseekable1 && !nonseekable2 && !seq @
identifier fops0.fops;
identifier readdir_e;
@@
// any other fop is used that changes pos
struct file_operations fops = {
... .readdir = readdir_e, ...
+.llseek = default_llseek, /* readdir is present */
};
// use default_llseek if at least one of read/write touches f_pos
/////////////////////////////////////////////////////////////////
@ fops2 depends on !fops1 && !has_llseek && !nonseekable1 && !nonseekable2 && !seq @
identifier fops0.fops;
identifier read.read_f;
@@
// read fops use offset
struct file_operations fops = {
... .read = read_f, ...
+.llseek = default_llseek, /* read accesses f_pos */
};
@ fops3 depends on !fops1 && !fops2 && !has_llseek && !nonseekable1 && !nonseekable2 && !seq @
identifier fops0.fops;
identifier write.write_f;
@@
// write fops use offset
struct file_operations fops = {
... .write = write_f, ...
+ .llseek = default_llseek, /* write accesses f_pos */
};
// Use noop_llseek if neither read nor write accesses f_pos
///////////////////////////////////////////////////////////
@ fops4 depends on !fops1 && !fops2 && !fops3 && !has_llseek && !nonseekable1 && !nonseekable2 && !seq @
identifier fops0.fops;
identifier read_no_fpos.read_f;
identifier write_no_fpos.write_f;
@@
// write fops use offset
struct file_operations fops = {
...
.write = write_f,
.read = read_f,
...
+.llseek = noop_llseek, /* read and write both use no f_pos */
};
@ depends on has_write && !has_read && !fops1 && !fops2 && !has_llseek && !nonseekable1 && !nonseekable2 && !seq @
identifier fops0.fops;
identifier write_no_fpos.write_f;
@@
struct file_operations fops = {
... .write = write_f, ...
+.llseek = noop_llseek, /* write uses no f_pos */
};
@ depends on has_read && !has_write && !fops1 && !fops2 && !has_llseek && !nonseekable1 && !nonseekable2 && !seq @
identifier fops0.fops;
identifier read_no_fpos.read_f;
@@
struct file_operations fops = {
... .read = read_f, ...
+.llseek = noop_llseek, /* read uses no f_pos */
};
@ depends on !has_read && !has_write && !fops1 && !fops2 && !has_llseek && !nonseekable1 && !nonseekable2 && !seq @
identifier fops0.fops;
@@
struct file_operations fops = {
...
+.llseek = noop_llseek, /* no read or write fn */
};
===== End semantic patch =====
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Cc: Julia Lawall <julia@diku.dk>
Cc: Christoph Hellwig <hch@infradead.org>
This patch disables the fanotify syscalls by just not building them and
letting the cond_syscall() statements in kernel/sys_ni.c redirect them
to sys_ni_syscall().
It was pointed out by Tvrtko Ursulin that the fanotify interface did not
include an explicit prioritization between groups. This is necessary
for fanotify to be usable for hierarchical storage management software,
as they must get first access to the file, before inotify-like notifiers
see the file.
This feature can be added in an ABI compatible way in the next release
(by using a number of bits in the flags field to carry the info) but it
was suggested by Alan that maybe we should just hold off and do it in
the next cycle, likely with an (new) explicit argument to the syscall.
I don't like this approach best as I know people are already starting to
use the current interface, but Alan is all wise and noone on list backed
me up with just using what we have. I feel this is needlessly ripping
the rug out from under people at the last minute, but if others think it
needs to be a new argument it might be the best way forward.
Three choices:
Go with what we got (and implement the new feature next cycle). Add a
new field right now (and implement the new feature next cycle). Wait
till next cycle to release the ABI (and implement the new feature next
cycle). This is number 3.
Signed-off-by: Eric Paris <eparis@redhat.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
The fsnotify main loop has 2 bools which indicated if we processed the
inode or vfsmount mark in that particular pass through the loop. These
bool can we replaced with the inode_group and vfsmount_group variables
and actually make the code a little easier to understand.
Signed-off-by: Eric Paris <eparis@redhat.com>
Marks were stored on the inode and vfsmonut mark list in order from
highest memory address to lowest memory address. The code to walk those
lists thought they were in order from lowest to highest with
unpredictable results when trying to match up marks from each. It was
possible that extra events would be sent to userspace when inode
marks ignoring events wouldn't get matched with the vfsmount marks.
This problem only affected fanotify when using both vfsmount and inode
marks simultaneously.
Signed-off-by: Eric Paris <eparis@redhat.com>
The appropriate error code when privileged operations are denied is
EPERM, not EACCES.
Signed-off-by: Andreas Gruenbacher <agruen@suse.de>
Signed-off-by: Eric Paris <paris@paris.rdu.redhat.com>
This reminded me... you have two pr_debugs in fanotify_should_send_event
which output redundant information. Maybe you intended it like that so
it is selectable how much log spam you want, or if not you may want to
apply this patch.
Signed-off-by: Tvrtko Ursulin <tvrtko.ursulin@sophos.com>
Signed-off-by: Eric Paris <eparis@redhat.com>
When an fanotify listener is closing it may cause a deadlock between the
listener and the original task doing an fs operation. If the original task
is waiting for a permissions response it will be holding the srcu lock. The
listener cannot clean up and exit until after that srcu lock is syncronized.
Thus deadlock. The fix introduced here is to stop accepting new permissions
events when a listener is shutting down and to grant permission for all
outstanding events. Thus the original task will eventually release the srcu
lock and the listener can complete shutdown.
Reported-by: Andreas Gruenbacher <agruen@suse.de>
Cc: Andreas Gruenbacher <agruen@suse.de>
Signed-off-by: Eric Paris <eparis@redhat.com>
The interesting 2 list lockstep walking didn't quite work out if the inode
marks only had ignores and the vfsmount list requested events. The code to
shortcut list traversal would not run the inode list since it didn't have real
event requests. This code forces inode list traversal when a vfsmount mark
matches the event type. Maybe we could add an i_fsnotify_ignored_mask field
to struct inode to get the shortcut back, but it doesn't seem worth it to grow
struct inode again.
I bet with the recent changes to lock the way we do now it would actually not
be a major perf hit to just drop i_fsnotify_mark_mask altogether. But that is
for another day.
Signed-off-by: Eric Paris <eparis@redhat.com>
The fsnotify main loop has 2 booleans which tell if a particular mark was
sent to the listeners or if it should be processed in the next pass. The
problem is that the booleans were not reset on each traversal of the loop.
So marks could get skipped even when they were not sent to the notifiers.
Reported-by: Tvrtko Ursulin <tvrtko.ursulin@sophos.com>
Signed-off-by: Eric Paris <eparis@redhat.com>
The fanotify code is supposed to get the group from the mark. It accidentally
only used the inode_mark. If the vfsmount_mark was set but not the inode_mark
it would deref the NULL inode_mark. Get the group from the correct place.
Reported-by: Tvrtko Ursulin <tvrtko.ursulin@sophos.com>
Signed-off-by: Eric Paris <eparis@redhat.com>
This reverts commit 3bcf3860a4 (and the
accompanying commit c1e5c95402 "vfs/fsnotify: fsnotify_close can delay
the final work in fput" that was a horribly ugly hack to make it work at
all).
The 'struct file' approach not only causes that disgusting hack, it
somehow breaks pulseaudio, probably due to some other subtlety with
f_count handling.
Fix up various conflicts due to later fsnotify work.
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
* 'for-linus' of git://git.infradead.org/users/eparis/notify: (132 commits)
fanotify: use both marks when possible
fsnotify: pass both the vfsmount mark and inode mark
fsnotify: walk the inode and vfsmount lists simultaneously
fsnotify: rework ignored mark flushing
fsnotify: remove global fsnotify groups lists
fsnotify: remove group->mask
fsnotify: remove the global masks
fsnotify: cleanup should_send_event
fanotify: use the mark in handler functions
audit: use the mark in handler functions
dnotify: use the mark in handler functions
inotify: use the mark in handler functions
fsnotify: send fsnotify_mark to groups in event handling functions
fsnotify: Exchange list heads instead of moving elements
fsnotify: srcu to protect read side of inode and vfsmount locks
fsnotify: use an explicit flag to indicate fsnotify_destroy_mark has been called
fsnotify: use _rcu functions for mark list traversal
fsnotify: place marks on object in order of group memory address
vfs/fsnotify: fsnotify_close can delay the final work in fput
fsnotify: store struct file not struct path
...
Fix up trivial delete/modify conflict in fs/notify/inotify/inotify.c.
add I_CLEAR instead of replacing I_FREEING with it. I_CLEAR is
equivalent to I_FREEING for almost all code looking at either;
it's there to keep track of having called clear_inode() exactly
once per inode lifetime, at some point after having set I_FREEING.
I_CLEAR and I_FREEING never get set at the same time with the
current code, so we can switch to setting i_flags to I_FREEING | I_CLEAR
instead of I_CLEAR without loss of information. As the result of
such change, checks become simpler and the amount of code that needs
to know about I_CLEAR shrinks a lot.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
fanotify currently, when given a vfsmount_mark will look up (if it exists)
the corresponding inode mark. This patch drops that lookup and uses the
mark provided.
Signed-off-by: Eric Paris <eparis@redhat.com>
should_send_event() and handle_event() will both need to look up the inode
event if they get a vfsmount event. Lets just pass both at the same time
since we have them both after walking the lists in lockstep.
Signed-off-by: Eric Paris <eparis@redhat.com>
We currently walk the list of marks on an inode followed by the list of
marks on the vfsmount. These are in order (by the memory address of the
group) so lets walk them both together. Eventually we can pass both the
inode mark and the vfsmount mark to helpers simultaneously.
Signed-off-by: Eric Paris <eparis@redhat.com>
currently ignored_mark clearing is done in a seperate list traversal
before the actual list traversal to send events. There is no need for
this. Do them at the same time.
Signed-off-by: Eric Paris <eparis@redhat.com>
The global fsnotify groups lists were invented as a way to increase the
performance of fsnotify by shortcutting events which were not interesting.
With the changes to walk the object lists rather than global groups lists
these shortcuts are not useful.
Signed-off-by: Eric Paris <eparis@redhat.com>
group->mask is now useless. It was originally a shortcut for fsnotify to
save on performance. These checks are now redundant, so we remove them.
Signed-off-by: Eric Paris <eparis@redhat.com>
Because we walk the object->fsnotify_marks list instead of the global
fsnotify groups list we don't need the fsnotify_inode_mask and
fsnotify_vfsmount_mask as these were simply shortcuts in fsnotify() for
performance. They are now extra checks, rip them out.
Signed-off-by: Eric Paris <eparis@redhat.com>
The change to use srcu and walk the object list rather than the global
fsnotify_group list means that should_send_event is no longer needed for a
number of groups and can be simplified for others. Do that.
Signed-off-by: Eric Paris <eparis@redhat.com>
fanotify now gets a mark in the should_send_event and handle_event
functions. Rather than look up the mark themselves fanotify should just use
the mark it was handed.
Signed-off-by: Eric Paris <eparis@redhat.com>
dnotify now gets a mark in the should_send_event and handle_event
functions. Rather than look up the mark themselves dnotify should just use
the mark it was handed.
Signed-off-by: Eric Paris <eparis@redhat.com>
inotify now gets a mark in the should_send_event and handle_event
functions. Rather than look up the mark themselves inotify should just use
the mark it was handed.
Signed-off-by: Eric Paris <eparis@redhat.com>
With the change of fsnotify to use srcu walking the marks list instead of
walking the global groups list we now know the mark in question. The code can
send the mark to the group's handling functions and the groups won't have to
find those marks themselves.
Signed-off-by: Eric Paris <eparis@redhat.com>
Instead of moving list elements from destroy_list to &private_destroy_list,
exchange the list heads.
Signed-off-by: Andreas Gruenbacher <agruen@suse.de>
Signed-off-by: Eric Paris <eparis@redhat.com>
Currently reading the inode->i_fsnotify_marks or
vfsmount->mnt_fsnotify_marks lists are protected by a spinlock on both the
read and the write side. This patch protects the read side of those lists
with a new single srcu.
Signed-off-by: Eric Paris <eparis@redhat.com>
Currently fsnotify check is mark->group is NULL to decide if
fsnotify_destroy_mark() has already been called or not. With the upcoming
rcu work it is a heck of a lot easier to use an explicit flag than worry
about group being set to NULL.
Signed-off-by: Eric Paris <eparis@redhat.com>
In preparation for srcu locking use all _rcu appropiete functions for mark
list addition, removal, and traversal. The operations are still done under a
spinlock at the end of this patch.
Signed-off-by: Eric Paris <eparis@redhat.com>
fsnotify_marks currently are placed on objects (inodes or vfsmounts) in
arbitrary order. This patch places them in order of the group memory address.
Signed-off-by: Eric Paris <eparis@redhat.com>