When attaching default groups (subdirs) of a new group (in mkdir() or
in configfs_register()), configfs recursively takes inode's mutexes
along the path from the parent of the new group to the default
subdirs. This is needed to ensure that the VFS will not race with
operations on these sub-dirs. This is safe for the following reasons:
- the VFS allows one to lock first an inode and second one of its
children (The lock subclasses for this pattern are respectively
I_MUTEX_PARENT and I_MUTEX_CHILD);
- from this rule any inode path can be recursively locked in
descending order as long as it stays under a single mountpoint and
does not follow symlinks.
Unfortunately lockdep does not know (yet?) how to handle such
recursion.
I've tried to use Peter Zijlstra's lock_set_subclass() helper to
upgrade i_mutexes from I_MUTEX_CHILD to I_MUTEX_PARENT when we know
that we might recursively lock some of their descendant, but this
usage does not seem to fit the purpose of lock_set_subclass() because
it leads to several i_mutex locked with subclass I_MUTEX_PARENT by
the same task.
>From inside configfs it is not possible to serialize those recursive
locking with a top-level one, because mkdir() and rmdir() are already
called with inodes locked by the VFS. So using some
mutex_lock_nest_lock() is not an option.
I am proposing two solutions:
1) one that wraps recursive mutex_lock()s with
lockdep_off()/lockdep_on().
2) (as suggested earlier by Peter Zijlstra) one that puts the
i_mutexes recursively locked in different classes based on their
depth from the top-level config_group created. This
induces an arbitrary limit (MAX_LOCK_DEPTH - 2 == 46) on the
nesting of configfs default groups whenever lockdep is activated
but this limit looks reasonably high. Unfortunately, this also
isolates VFS operations on configfs default groups from the others
and thus lowers the chances to detect locking issues.
Nobody likes solution 1), which I can understand.
This patch implements solution 2). However lockdep is still not happy with
configfs_depend_item(). Next patch reworks the locking of
configfs_depend_item() and finally makes lockdep happy.
[ Note: This hides a few locking interactions with the VFS from lockdep.
That was my big concern, because we like lockdep's protection. However,
the current state always dumps a spurious warning. The locking is
correct, so I tell people to ignore the warning and that we'll keep
our eyes on the locking to make sure it stays correct. With this patch,
we eliminate the warning. We do lose some of the lockdep protections,
but this only means that we still have to keep our eyes on the locking.
We're going to do that anyway. -- Joel ]
Signed-off-by: Louis Rilling <louis.rilling@kerlabs.com>
Signed-off-by: Joel Becker <joel.becker@oracle.com>
process 1: process 2:
configfs_mkdir("A")
attach_group("A")
attach_item("A")
d_instantiate("A")
populate_groups("A")
mutex_lock("A")
attach_group("A/B")
attach_item("A")
d_instantiate("A/B")
mkdir("A/B/C")
do_path_lookup("A/B/C", LOOKUP_PARENT)
ok
lookup_create("A/B/C")
mutex_lock("A/B")
ok
configfs_mkdir("A/B/C")
ok
attach_group("A/C")
attach_item("A/C")
d_instantiate("A/C")
populate_groups("A/C")
mutex_lock("A/C")
attach_group("A/C/D")
attach_item("A/C/D")
failure
mutex_unlock("A/C")
detach_groups("A/C")
nothing to do
mkdir("A/C/E")
do_path_lookup("A/C/E", LOOKUP_PARENT)
ok
lookup_create("A/C/E")
mutex_lock("A/C")
ok
configfs_mkdir("A/C/E")
ok
detach_item("A/C")
d_delete("A/C")
mutex_unlock("A")
detach_groups("A")
mutex_lock("A/B")
detach_group("A/B")
detach_groups("A/B")
nothing since no _default_ group
detach_item("A/B")
mutex_unlock("A/B")
d_delete("A/B")
detach_item("A")
d_delete("A")
Two bugs:
1/ "A/B/C" and "A/C/E" are created, but never removed while their parent are
removed in the end. The same could happen with symlink() instead of mkdir().
2/ "A" and "A/C" inodes are not locked while detach_item() is called on them,
which may probably confuse VFS.
This commit fixes 1/, tagging new directories with CONFIGFS_USET_CREATING before
building the inode and instantiating the dentry, and validating the whole
group+default groups hierarchy in a second pass by clearing
CONFIGFS_USET_CREATING.
mkdir(), symlink(), lookup(), and dir_open() simply return -ENOENT if
called in (or linking to) a directory tagged with CONFIGFS_USET_CREATING. This
does not prevent userspace from calling stat() successfuly on such directories,
but this prevents userspace from adding (children to | symlinking from/to |
read/write attributes of | listing the contents of) not validated items. In
other words, userspace will not interact with the subsystem on a new item until
the new item creation completes correctly.
It was first proposed to re-use CONFIGFS_USET_IN_MKDIR instead of a new
flag CONFIGFS_USET_CREATING, but this generated conflicts when checking the
target of a new symlink: a valid target directory in the middle of attaching
a new user-created child item could be wrongly detected as being attached.
2/ is fixed by next commit.
Signed-off-by: Louis Rilling <louis.rilling@kerlabs.com>
Signed-off-by: Joel Becker <joel.becker@oracle.com>
Signed-off-by: Mark Fasheh <mfasheh@suse.com>
On a similar pattern as mkdir() vs rmdir(), a failing symlink() may make rmdir()
fail for the symlink's parent and the symlink's target as well.
failing symlink() making target's rmdir() fail:
process 1: process 2:
symlink("A/S" -> "B")
allow_link()
create_link()
attach to "B" links list
rmdir("B")
detach_prep("B")
error because of new link
configfs_create_link("A", "S")
error (eg -ENOMEM)
failing symlink() making parent's rmdir() fail:
process 1: process 2:
symlink("A/D/S" -> "B")
allow_link()
create_link()
attach to "B" links list
configfs_create_link("A/D", "S")
make_dirent("A/D", "S")
rmdir("A")
detach_prep("A")
detach_prep("A/D")
error because of "S"
create("S")
error (eg -ENOMEM)
We cannot use the same solution as for mkdir() vs rmdir(), since rmdir() on the
target cannot wait on the i_mutex of the new symlink's parent without risking a
deadlock (with other symlink() or sys_rename()). Instead we define a global
mutex protecting all configfs symlinks attachment, so that rmdir() can avoid the
races above.
Signed-off-by: Louis Rilling <louis.rilling@kerlabs.com>
Signed-off-by: Joel Becker <joel.becker@oracle.com>
Signed-off-by: Mark Fasheh <mfasheh@suse.com>
When fixing the rename() vs rmdir() deadlock, we stopped locking default groups'
inodes in configfs_detach_prep(), letting racing mkdir() in default groups
proceed concurrently. This enables races like below happen, which leads to a
failing mkdir() making rmdir() fail, despite the group to remove having no
user-created directory under it in the end.
process A: process B:
/* PWD=A/B */
mkdir("C")
make_item("C")
attach_group("C")
rmdir("A")
detach_prep("A")
detach_prep("B")
error because of "C"
return -ENOTEMPTY
attach_group("C/D")
error (eg -ENOMEM)
return -ENOMEM
This patch prevents such scenarii by making rmdir() wait as long as
detach_prep() fails because a racing mkdir() is in the middle of attach_group().
To achieve this, mkdir() sets a flag CONFIGFS_USET_IN_MKDIR in parent's
configfs_dirent before calling attach_group(), and clears the flag once
attach_group() is done. detach_prep() fails with -EAGAIN whenever the flag is
hit and returns the guilty inode's mutex so that rmdir() can wait on it.
Signed-off-by: Louis Rilling <Louis.Rilling@kerlabs.com>
Signed-off-by: Joel Becker <joel.becker@oracle.com>
This patch introduces configfs_dirent_lock spinlock to protect configfs_dirent
traversals against linkage mutations (add/del/move). This will allow
configfs_detach_prep() to avoid locking i_mutexes.
Locking rules for configfs_dirent linkage mutations are the same plus the
requirement of taking configfs_dirent_lock. For configfs_dirent walking, one can
either take appropriate i_mutex as before, or take configfs_dirent_lock.
The spinlock could actually be a mutex, but the critical sections are either
O(1) or should not be too long (default groups walking in last patch).
ChangeLog:
- Clarify the comment on configfs_dirent_lock usage
- Move sd->s_element init before linking the new dirent
- In lseek(), do not release configfs_dirent_lock before the dirent is
relinked.
Signed-off-by: Louis Rilling <Louis.Rilling@kerlabs.com>
Signed-off-by: Joel Becker <joel.becker@oracle.com>
Sometimes other drivers depend on particular configfs items. For
example, ocfs2 mounts depend on a heartbeat region item. If that
region item is removed with rmdir(2), the ocfs2 mount must BUG or go
readonly. Not happy.
This provides two additional API calls: configfs_depend_item() and
configfs_undepend_item(). A client driver can call
configfs_depend_item() on an existing item to tell configfs that it is
depended on. configfs will then return -EBUSY from rmdir(2) for that
item. When the item is no longer depended on, the client driver calls
configfs_undepend_item() on it.
These API cannot be called underneath any configfs callbacks, as
they will conflict. They can block and allocate. A client driver
probably shouldn't calling them of its own gumption. Rather it should
be providing an API that external subsystems call.
How does this work? Imagine the ocfs2 mount process. When it mounts,
it asks for a heart region item. This is done via a call into the
heartbeat code. Inside the heartbeat code, the region item is looked
up. Here, the heartbeat code calls configfs_depend_item(). If it
succeeds, then heartbeat knows the region is safe to give to ocfs2.
If it fails, it was being torn down anyway, and heartbeat can gracefully
pass up an error.
[ Fixed some bad whitespace in configfs.txt. --Mark ]
Signed-off-by: Joel Becker <joel.becker@oracle.com>
Signed-off-by: Mark Fasheh <mark.fasheh@oracle.com>
Many struct inode_operations in the kernel can be "const". Marking them const
moves these to the .rodata section, which avoids false sharing with potential
dirty data. In addition it'll catch accidental writes at compile time to
these shared resources.
Signed-off-by: Arjan van de Ven <arjan@linux.intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Replace all uses of kmem_cache_t with struct kmem_cache.
The patch was generated using the following script:
#!/bin/sh
#
# Replace one string by another in all the kernel sources.
#
set -e
for file in `find * -name "*.c" -o -name "*.h"|xargs grep -l $1`; do
quilt add $file
sed -e "1,\$s/$1/$2/g" $file >/tmp/$$
mv /tmp/$$ $file
quilt refresh
done
The script was run like this
sh replace kmem_cache_t "struct kmem_cache"
Signed-off-by: Christoph Lameter <clameter@sgi.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
This is a conversion to make the various file_operations structs in fs/
const. Basically a regexp job, with a few manual fixups
The goal is both to increase correctness (harder to accidentally write to
shared datastructures) and reducing the false sharing of cachelines with
things that get dirty in .data (while .rodata is nicely read only and thus
cache clean)
Signed-off-by: Arjan van de Ven <arjan@infradead.org>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
configfs always made item and attribute ownership root.root and
permissions based on a umask of 022. Add ->setattr() to allow
chown(2)/chmod(2), and persist the changes for the lifetime of the
items and attributes.
Signed-off-by: Joel Becker <joel.becker@oracle.com>
Signed-off-by: Mark Fasheh <mark.fasheh@oracle.com>
Configfs, a file system for userspace-driven kernel object configuration.
The OCFS2 stack makes extensive use of this for propagation of cluster
configuration information into kernel.
Signed-off-by: Joel Becker <joel.becker@oracle.com>
Louis Rilling