Fix the AFS file locking whereby the use of the big kernel lock (which
could be slept with) was replaced by a spinlock (which couldn't). The
problem is that the AFS code was doing stuff inside the critical section
that might call schedule(), so this is a broken transformation.
Fix this by the following means:
(1) Use a state machine with a proper state that can only be changed under
the spinlock rather than using a collection of bit flags.
(2) Cache the key used for the lock and the lock type in the afs_vnode
struct so that the manager work function doesn't have to refer to a
file_lock struct that's been dequeued. This makes signal handling
safer.
(4) Move the unlock from afs_do_unlk() to afs_fl_release_private() which
means that unlock is achieved in other circumstances too.
(5) Unlock the file on the server before taking the next conflicting lock.
Also change:
(1) Check the permits on a file before actually trying the lock.
(2) fsync the file before effecting an explicit unlock operation. We
don't fsync if the lock is erased otherwise as we might not be in a
context where we can actually do that.
Further fixes:
(1) Fixed-fileserver address rotation is made to work. It's only used by
the locking functions, so couldn't be tested before.
Fixes: 72f98e7255 ("locks: turn lock_flocks into a spinlock")
Signed-off-by: David Howells <dhowells@redhat.com>
cc: jlayton@redhat.com
Introduce a file-private data record for kAFS and put the key into it
rather than storing the key in file->private_data.
Signed-off-by: David Howells <dhowells@redhat.com>
The current code assumes that volumes and servers are per-cell and are
never shared, but this is not enforced, and, indeed, public cells do exist
that are aliases of each other. Further, an organisation can, say, set up
a public cell and a private cell with overlapping, but not identical, sets
of servers. The difference is purely in the database attached to the VL
servers.
The current code will malfunction if it sees a server in two cells as it
assumes global address -> server record mappings and that each server is in
just one cell.
Further, each server may have multiple addresses - and may have addresses
of different families (IPv4 and IPv6, say).
To this end, the following structural changes are made:
(1) Server record management is overhauled:
(a) Server records are made independent of cell. The namespace keeps
track of them, volume records have lists of them and each vnode
has a server on which its callback interest currently resides.
(b) The cell record no longer keeps a list of servers known to be in
that cell.
(c) The server records are now kept in a flat list because there's no
single address to sort on.
(d) Server records are now keyed by their UUID within the namespace.
(e) The addresses for a server are obtained with the VL.GetAddrsU
rather than with VL.GetEntryByName, using the server's UUID as a
parameter.
(f) Cached server records are garbage collected after a period of
non-use and are counted out of existence before purging is allowed
to complete. This protects the work functions against rmmod.
(g) The servers list is now in /proc/fs/afs/servers.
(2) Volume record management is overhauled:
(a) An RCU-replaceable server list is introduced. This tracks both
servers and their coresponding callback interests.
(b) The superblock is now keyed on cell record and numeric volume ID.
(c) The volume record is now tied to the superblock which mounts it,
and is activated when mounted and deactivated when unmounted.
This makes it easier to handle the cache cookie without causing a
double-use in fscache.
(d) The volume record is loaded from the VLDB using VL.GetEntryByNameU
to get the server UUID list.
(e) The volume name is updated if it is seen to have changed when the
volume is updated (the update is keyed on the volume ID).
(3) The vlocation record is got rid of and VLDB records are no longer
cached. Sufficient information is stored in the volume record, though
an update to a volume record is now no longer shared between related
volumes (volumes come in bundles of three: R/W, R/O and backup).
and the following procedural changes are made:
(1) The fileserver cursor introduced previously is now fleshed out and
used to iterate over fileservers and their addresses.
(2) Volume status is checked during iteration, and the server list is
replaced if a change is detected.
(3) Server status is checked during iteration, and the address list is
replaced if a change is detected.
(4) The abort code is saved into the address list cursor and -ECONNABORTED
returned in afs_make_call() if a remote abort happened rather than
translating the abort into an error message. This allows actions to
be taken depending on the abort code more easily.
(a) If a VMOVED abort is seen then this is handled by rechecking the
volume and restarting the iteration.
(b) If a VBUSY, VRESTARTING or VSALVAGING abort is seen then this is
handled by sleeping for a short period and retrying and/or trying
other servers that might serve that volume. A message is also
displayed once until the condition has cleared.
(c) If a VOFFLINE abort is seen, then this is handled as VBUSY for the
moment.
(d) If a VNOVOL abort is seen, the volume is rechecked in the VLDB to
see if it has been deleted; if not, the fileserver is probably
indicating that the volume couldn't be attached and needs
salvaging.
(e) If statfs() sees one of these aborts, it does not sleep, but
rather returns an error, so as not to block the umount program.
(5) The fileserver iteration functions in vnode.c are now merged into
their callers and more heavily macroised around the cursor. vnode.c
is removed.
(6) Operations on a particular vnode are serialised on that vnode because
the server will lock that vnode whilst it operates on it, so a second
op sent will just have to wait.
(7) Fileservers are probed with FS.GetCapabilities before being used.
This is where service upgrade will be done.
(8) A callback interest on a fileserver is set up before an FS operation
is performed and passed through to afs_make_call() so that it can be
set on the vnode if the operation returns a callback. The callback
interest is passed through to afs_iget() also so that it can be set
there too.
In general, record updating is done on an as-needed basis when we try to
access servers, volumes or vnodes rather than offloading it to work items
and special threads.
Notes:
(1) Pre AFS-3.4 servers are no longer supported, though this can be added
back if necessary (AFS-3.4 was released in 1998).
(2) VBUSY is retried forever for the moment at intervals of 1s.
(3) /proc/fs/afs/<cell>/servers no longer exists.
Signed-off-by: David Howells <dhowells@redhat.com>
Overhaul permit caching in AFS by making it per-vnode and sharing permit
lists where possible.
When most of the fileserver operations are called, they return a status
structure indicating the (revised) details of the vnode or vnodes involved
in the operation. This includes the access mark derived from the ACL
(named CallerAccess in the protocol definition file). This is cacheable
and if the ACL changes, the server will tell us that it is breaking the
callback promise, at which point we can discard the currently cached
permits.
With this patch, the afs_permits structure has, at the end, an array of
{ key, CallerAccess } elements, sorted by key pointer. This is then cached
in a hash table so that it can be shared between vnodes with the same
access permits.
Permit lists can only be shared if they contain the exact same set of
key->CallerAccess mappings.
Note that that table is global rather than being per-net_ns. If the keys
in a permit list cross net_ns boundaries, there is no problem sharing the
cached permits, since the permits are just integer masks.
Since permit lists pin keys, the permit cache also makes it easier for a
future patch to find all occurrences of a key and remove them by means of
setting the afs_permits::invalidated flag and then clearing the appropriate
key pointer. In such an event, memory barriers will need adding.
Lastly, the permit caching is skipped if the server has sent either a
vnode-specific or an entire-server callback since the start of the
operation.
Signed-off-by: David Howells <dhowells@redhat.com>
Overhaul the AFS callback handling by the following means:
(1) Don't give up callback promises on vnodes that we are no longer using,
rather let them just expire on the server or let the server break
them. This is actually more efficient for the server as the callback
lookup is expensive if there are lots of extant callbacks.
(2) Only give up the callback promises we have from a server when the
server record is destroyed. Then we can just give up *all* the
callback promises on it in one go.
(3) Servers can end up being shared between cells if cells are aliased, so
don't add all the vnodes being backed by a particular server into a
big FID-indexed tree on that server as there may be duplicates.
Instead have each volume instance (~= superblock) register an interest
in a server as it starts to make use of it and use this to allow the
processor for callbacks from the server to find the superblock and
thence the inode corresponding to the FID being broken by means of
ilookup_nowait().
(4) Rather than iterating over the entire callback list when a mass-break
comes in from the server, maintain a counter of mass-breaks in
afs_server (cb_seq) and make afs_validate() check it against the copy
in afs_vnode.
It would be nice not to have to take a read_lock whilst doing this,
but that's tricky without using RCU.
(5) Save a ref on the fileserver we're using for a call in the afs_call
struct so that we can access its cb_s_break during call decoding.
(6) Write-lock around callback and status storage in a vnode and read-lock
around getattr so that we don't see the status mid-update.
This has the following consequences:
(1) Data invalidation isn't seen until someone calls afs_validate() on a
vnode. Unfortunately, we need to use a key to query the server, but
getting one from a background thread is tricky without caching loads
of keys all over the place.
(2) Mass invalidation isn't seen until someone calls afs_validate().
(3) Callback breaking is going to hit the inode_hash_lock quite a bit.
Could this be replaced with rcu_read_lock() since inodes are destroyed
under RCU conditions.
Signed-off-by: David Howells <dhowells@redhat.com>
Lay the groundwork for supporting network namespaces (netns) to the AFS
filesystem by moving various global features to a network-namespace struct
(afs_net) and providing an instance of this as a temporary global variable
that everything uses via accessor functions for the moment.
The following changes have been made:
(1) Store the netns in the superblock info. This will be obtained from
the mounter's nsproxy on a manual mount and inherited from the parent
superblock on an automount.
(2) The cell list is made per-netns. It can be viewed through
/proc/net/afs/cells and also be modified by writing commands to that
file.
(3) The local workstation cell is set per-ns in /proc/net/afs/rootcell.
This is unset by default.
(4) The 'rootcell' module parameter, which sets a cell and VL server list
modifies the init net namespace, thereby allowing an AFS root fs to be
theoretically used.
(5) The volume location lists and the file lock manager are made
per-netns.
(6) The AF_RXRPC socket and associated I/O bits are made per-ns.
The various workqueues remain global for the moment.
Changes still to be made:
(1) /proc/fs/afs/ should be moved to /proc/net/afs/ and a symlink emplaced
from the old name.
(2) A per-netns subsys needs to be registered for AFS into which it can
store its per-netns data.
(3) Rather than the AF_RXRPC socket being opened on module init, it needs
to be opened on the creation of a superblock in that netns.
(4) The socket needs to be closed when the last superblock using it is
destroyed and all outstanding client calls on it have been completed.
This prevents a reference loop on the namespace.
(5) It is possible that several namespaces will want to use AFS, in which
case each one will need its own UDP port. These can either be set
through /proc/net/afs/cm_port or the kernel can pick one at random.
The init_ns gets 7001 by default.
Other issues that need resolving:
(1) The DNS keyring needs net-namespacing.
(2) Where do upcalls go (eg. DNS request-key upcall)?
(3) Need something like open_socket_in_file_ns() syscall so that AFS
command line tools attempting to operate on an AFS file/volume have
their RPC calls go to the right place.
Signed-off-by: David Howells <dhowells@redhat.com>
The workqueue "afs_lock_manager" queues work item &vnode->lock_work,
per vnode. Since there can be multiple vnodes and since their work items
can be executed concurrently, alloc_workqueue has been used to replace
the deprecated create_singlethread_workqueue instance.
The WQ_MEM_RECLAIM flag has been set to ensure forward progress under
memory pressure because the workqueue is being used on a memory reclaim
path.
Since there are fixed number of work items, explicit concurrency
limit is unnecessary here.
Signed-off-by: Bhaktipriya Shridhar <bhaktipriya96@gmail.com>
Signed-off-by: David Howells <dhowells@redhat.com>
parallel to mutex_{lock,unlock,trylock,is_locked,lock_nested},
inode_foo(inode) being mutex_foo(&inode->i_mutex).
Please, use those for access to ->i_mutex; over the coming cycle
->i_mutex will become rwsem, with ->lookup() done with it held
only shared.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
Currently, the fl_owner isn't set for flock locks. Some filesystems use
byte-range locks to simulate flock locks and there is a common idiom in
those that does:
fl->fl_owner = (fl_owner_t)filp;
fl->fl_start = 0;
fl->fl_end = OFFSET_MAX;
Since flock locks are generally "owned" by the open file description,
move this into the common flock lock setup code. The fl_start and fl_end
fields are already set appropriately, so remove the unneeded setting of
that in flock ops in those filesystems as well.
Finally, the lease code also sets the fl_owner as if they were owned by
the process and not the open file description. This is incorrect as
leases have the same ownership semantics as flock locks. Set them the
same way. The lease code doesn't actually use the fl_owner value for
anything, so this is more for consistency's sake than a bugfix.
Reported-by: Trond Myklebust <trond.myklebust@primarydata.com>
Signed-off-by: Jeff Layton <jlayton@poochiereds.net>
Acked-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org> (Staging portion)
Acked-by: J. Bruce Fields <bfields@fieldses.org>
Having a global lock that protects all of this code is a clear
scalability problem. Instead of doing that, move most of the code to be
protected by the i_lock instead. The exceptions are the global lists
that the ->fl_link sits on, and the ->fl_block list.
->fl_link is what connects these structures to the
global lists, so we must ensure that we hold those locks when iterating
over or updating these lists.
Furthermore, sound deadlock detection requires that we hold the
blocked_list state steady while checking for loops. We also must ensure
that the search and update to the list are atomic.
For the checking and insertion side of the blocked_list, push the
acquisition of the global lock into __posix_lock_file and ensure that
checking and update of the blocked_list is done without dropping the
lock in between.
On the removal side, when waking up blocked lock waiters, take the
global lock before walking the blocked list and dequeue the waiters from
the global list prior to removal from the fl_block list.
With this, deadlock detection should be race free while we minimize
excessive file_lock_lock thrashing.
Finally, in order to avoid a lock inversion problem when handling
/proc/locks output we must ensure that manipulations of the fl_block
list are also protected by the file_lock_lock.
Signed-off-by: Jeff Layton <jlayton@redhat.com>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
This prepares the removal of the big kernel lock from the
file locking code. We still use the BKL as long as fs/lockd
uses it and ceph might sleep, but we can flip the definition
to a private spinlock as soon as that's done.
All users outside of fs/lockd get converted to use
lock_flocks() instead of lock_kernel() where appropriate.
Based on an earlier patch to use a spinlock from Matthew
Wilcox, who has attempted this a few times before, the
earliest patch from over 10 years ago turned it into
a semaphore, which ended up being slower than the BKL
and was subsequently reverted.
Someone should do some serious performance testing when
this becomes a spinlock, since this has caused problems
before. Using a spinlock should be at least as good
as the BKL in theory, but who knows...
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Acked-by: Matthew Wilcox <willy@linux.intel.com>
Cc: Christoph Hellwig <hch@lst.de>
Cc: Trond Myklebust <trond.myklebust@fys.uio.no>
Cc: "J. Bruce Fields" <bfields@fieldses.org>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Miklos Szeredi <mszeredi@suse.cz>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: John Kacur <jkacur@redhat.com>
Cc: Sage Weil <sage@newdream.net>
Cc: linux-kernel@vger.kernel.org
Cc: linux-fsdevel@vger.kernel.org
Don't unlock on vfs_rejected_lock path in afs_do_setlk, since the lock
is unlocked after abort_attempt label.
Signed-off-by: Jiri Slaby <jirislaby@gmail.com>
Signed-off-by: David Howells <dhowells@redhat.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
The __mandatory_lock(inode) macro makes the same check, but makes the code
more readable.
Signed-off-by: Pavel Emelyanov <xemul@openvz.org>
Cc: David Howells <dhowells@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Fix file locking for AFS:
(*) Start the lock manager thread under a mutex to avoid a race.
(*) Made the locking non-fair: New readlocks will jump pending writelocks if
there's a readlock currently granted on a file. This makes the behaviour
similar to Linux's VFS locking.
Signed-off-by: David Howells <dhowells@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Bruce and David's patches clashed.
fs/afs/flock.c: In function 'afs_do_getlk':
fs/afs/flock.c:459: error: void value not ignored as it ought to be
Cc: "J. Bruce Fields" <bfields@fieldses.org>
Acked-by: David Howells <dhowells@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
David Howells