We are in the situation that we have to avoid recursive cluster locking,
but there is no way to check if a cluster lock has been taken by a precess
already.
Mostly, we can avoid recursive locking by writing code carefully.
However, we found that it's very hard to handle the routines that are
invoked directly by vfs code. For instance:
const struct inode_operations ocfs2_file_iops = {
.permission = ocfs2_permission,
.get_acl = ocfs2_iop_get_acl,
.set_acl = ocfs2_iop_set_acl,
};
Both ocfs2_permission() and ocfs2_iop_get_acl() call ocfs2_inode_lock(PR):
do_sys_open
may_open
inode_permission
ocfs2_permission
ocfs2_inode_lock() <=== first time
generic_permission
get_acl
ocfs2_iop_get_acl
ocfs2_inode_lock() <=== recursive one
A deadlock will occur if a remote EX request comes in between two of
ocfs2_inode_lock(). Briefly describe how the deadlock is formed:
On one hand, OCFS2_LOCK_BLOCKED flag of this lockres is set in
BAST(ocfs2_generic_handle_bast) when downconvert is started on behalf of
the remote EX lock request. Another hand, the recursive cluster lock
(the second one) will be blocked in in __ocfs2_cluster_lock() because of
OCFS2_LOCK_BLOCKED. But, the downconvert never complete, why? because
there is no chance for the first cluster lock on this node to be
unlocked - we block ourselves in the code path.
The idea to fix this issue is mostly taken from gfs2 code.
1. introduce a new field: struct ocfs2_lock_res.l_holders, to keep track
of the processes' pid who has taken the cluster lock of this lock
resource;
2. introduce a new flag for ocfs2_inode_lock_full:
OCFS2_META_LOCK_GETBH; it means just getting back disk inode bh for
us if we've got cluster lock.
3. export a helper: ocfs2_is_locked_by_me() is used to check if we have
got the cluster lock in the upper code path.
The tracking logic should be used by some of the ocfs2 vfs's callbacks,
to solve the recursive locking issue cuased by the fact that vfs
routines can call into each other.
The performance penalty of processing the holder list should only be
seen at a few cases where the tracking logic is used, such as get/set
acl.
You may ask what if the first time we got a PR lock, and the second time
we want a EX lock? fortunately, this case never happens in the real
world, as far as I can see, including permission check,
(get|set)_(acl|attr), and the gfs2 code also do so.
[sfr@canb.auug.org.au remove some inlines]
Link: http://lkml.kernel.org/r/20170117100948.11657-2-zren@suse.com
Signed-off-by: Eric Ren <zren@suse.com>
Reviewed-by: Junxiao Bi <junxiao.bi@oracle.com>
Reviewed-by: Joseph Qi <jiangqi903@gmail.com>
Cc: Stephen Rothwell <sfr@canb.auug.org.au>
Cc: Mark Fasheh <mfasheh@versity.com>
Cc: Joel Becker <jlbec@evilplan.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
If we are dropping last inode reference from downconvert thread, we will
end up calling ocfs2_mark_lockres_freeing() which can block if the lock
we are freeing is queued thus creating an A-A deadlock. Luckily, since
we are the downconvert thread, we can immediately dequeue the lock and
thus avoid waiting in this case.
Signed-off-by: Jan Kara <jack@suse.cz>
Reviewed-by: Mark Fasheh <mfasheh@suse.de>
Reviewed-by: Srinivas Eeda <srinivas.eeda@oracle.com>
Cc: Joel Becker <jlbec@evilplan.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Add lockdep support to OCFS2. The support also covers all of the cluster
locks except for open locks, journal locks, and local quotafile locks. These
are special because they are acquired for a node, not for a particular process
and lockdep cannot deal with such type of locking.
Signed-off-by: Jan Kara <jack@suse.cz>
Signed-off-by: Joel Becker <joel.becker@oracle.com>
Local and Hard-RO mounts do not need orphan scanning.
Signed-off-by: Sunil Mushran <sunil.mushran@oracle.com>
Signed-off-by: Joel Becker <joel.becker@oracle.com>
When a dentry is unlinked, the unlinking node takes an EX on the dentry lock
before moving the dentry to the orphan directory. Other nodes that have
this dentry in cache have a PR on the same dentry lock. When the EX is
requested, the other nodes flag the corresponding inode as MAYBE_ORPHANED
during downconvert. The inode is finally deleted when the last node to iput
the inode sees that i_nlink==0 and the MAYBE_ORPHANED flag is set.
A problem arises if a node is forced to free dentry locks because of memory
pressure. If this happens, the node will no longer get downconvert
notifications for the dentries that have been unlinked on another node.
If it also happens that node is actively using the corresponding inode and
happens to be the one performing the last iput on that inode, it will fail
to delete the inode as it will not have the MAYBE_ORPHANED flag set.
This patch fixes this shortcoming by introducing a periodic scan of the
orphan directories to delete such inodes. Care has been taken to distribute
the workload across the cluster so that no one node has to perform the task
all the time.
Signed-off-by: Srinivas Eeda <srinivas.eeda@oracle.com>
Signed-off-by: Joel Becker <joel.becker@oracle.com>
For nfs exporting, ocfs2_get_dentry() returns the dentry for fh.
ocfs2_get_dentry() may read from disk when the inode is not in memory,
without any cross cluster lock. this leads to the file system loading a
stale inode.
This patch fixes above problem.
Solution is that in case of inode is not in memory, we get the cluster
lock(PR) of alloc inode where the inode in question is allocated from (this
causes node on which deletion is done sync the alloc inode) before reading
out the inode itsself. then we check the bitmap in the group (the inode in
question allcated from) to see if the bit is clear. if it's clear then it's
stale. if the bit is set, we then check generation as the existing code
does.
We have to read out the inode in question from disk first to know its alloc
slot and allot bit. And if its not stale we read it out using ocfs2_iget().
The second read should then be from cache.
And also we have to add a per superblock nfs_sync_lock to cover the lock for
alloc inode and that for inode in question. this is because ocfs2_get_dentry()
and ocfs2_delete_inode() lock on them in reverse order. nfs_sync_lock is locked
in EX mode in ocfs2_get_dentry() and in PR mode in ocfs2_delete_inode(). so
that mutliple ocfs2_delete_inode() can run concurrently in normal case.
[mfasheh@suse.com: build warning fixes and comment cleanups]
Signed-off-by: Wengang Wang <wen.gang.wang@oracle.com>
Acked-by: Joel Becker <joel.becker@oracle.com>
Signed-off-by: Mark Fasheh <mfasheh@suse.com>
For each quota type each node has local quota file. In this file it stores
changes users have made to disk usage via this node. Once in a while this
information is synced to global file (and thus with other nodes) so that
limits enforcement at least aproximately works.
Global quota files contain all the information about usage and limits. It's
mostly handled by the generic VFS code (which implements a trie of structures
inside a quota file). We only have to provide functions to convert structures
from on-disk format to in-memory one. We also have to provide wrappers for
various quota functions starting transactions and acquiring necessary cluster
locks before the actual IO is really started.
Signed-off-by: Jan Kara <jack@suse.cz>
Signed-off-by: Mark Fasheh <mfasheh@suse.com>
We define the ocfs2_stack_plugin structure to represent a stack driver.
The o2cb stack code is split into stack_o2cb.c. This becomes the
ocfs2_stack_o2cb.ko module.
The stackglue generic functions are similarly split into the
ocfs2_stackglue.ko module. This module now provides an interface to
register drivers. The ocfs2_stack_o2cb driver registers itself. As
part of this interface, ocfs2_stackglue can load drivers on demand.
This is accomplished in ocfs2_cluster_connect().
ocfs2_cluster_disconnect() is now notified when a _hangup() is pending.
If a hangup is pending, it will not release the driver module and will
let _hangup() do that.
Signed-off-by: Joel Becker <joel.becker@oracle.com>
The stack glue initialization function needs a better name so that it can be
used cleanly when stackglue becomes a module.
Signed-off-by: Joel Becker <joel.becker@oracle.com>
Signed-off-by: Mark Fasheh <mfasheh@suse.com>
This step introduces a cluster stack agnostic API for initializing and
exiting. fs/ocfs2/dlmglue.c no longer uses o2cb/o2dlm knowledge to
connect to the stack. It is all handled in stackglue.c.
heartbeat.c no longer needs to know how it gets called.
ocfs2_do_node_down() is now a clean recovery trigger.
The big gotcha is the ordering of initializations and de-initializations done
underneath ocfs2_cluster_connect(). ocfs2_dlm_init() used to do all
o2dlm initialization in one block. Thus, the o2dlm functionality of
ocfs2_cluster_connect() is very straightforward. ocfs2_dlm_shutdown(),
however, did a few things between de-registration of the eviction
callback and actually shutting down the domain. Now de-registration and
shutdown of the domain are wrapped within the single
ocfs2_cluster_disconnect() call. I've checked the code paths to make
sure we can safely tear down things in ocfs2_dlm_shutdown() before
calling ocfs2_cluster_disconnect(). The filesystem has already set
itself to ignore the callback.
Signed-off-by: Joel Becker <joel.becker@oracle.com>
Signed-off-by: Mark Fasheh <mfasheh@suse.com>
This is the first in a series of patches to isolate ocfs2 from the
underlying cluster stack. Here we wrap the dlm locking functions with
ocfs2-specific calls. Because ocfs2 always uses the same dlm lock status
callbacks, we can eliminate the callbacks from the filesystem visible
functions.
Signed-off-by: Joel Becker <joel.becker@oracle.com>
Signed-off-by: Mark Fasheh <mfasheh@suse.com>
This patch contains the following cleanups that are now possible:
- make the following needlessly global functions static:
- dlmglue.c:ocfs2_process_blocked_lock()
- heartbeat.c:ocfs2_node_map_init()
- #if 0 the following unused global function plus support functions:
- heartbeat.c:ocfs2_node_map_is_only()
Signed-off-by: Adrian Bunk <bunk@kernel.org>
Signed-off-by: Mark Fasheh <mark.fasheh@oracle.com>
Currently, when ocfs2 nodes connect via TCP, they advertise their
compatibility level. If the versions do not match, two nodes cannot speak
to each other and they disconnect. As a result, this provides no forward or
backwards compatibility.
This patch implements a simple protocol negotiation at the dlm level by
introducing a major/minor version number scheme for entities that
communicate. Specifically, o2dlm has a major/minor version for interaction
with o2dlm on other nodes, and ocfs2 itself has a major/minor version for
interacting with the filesystem on other nodes.
This will allow rolling upgrades of ocfs2 clusters when changes to the
locking or network protocols can be done in a backwards compatible manner.
In those cases, only the minor number is changed and the negotatied protocol
minor is returned from dlm join. In the far less likely event that a
required protocol change makes backwards compatibility impossible, we simply
bump the major number.
Signed-off-by: Joel Becker <joel.becker@oracle.com>
Signed-off-by: Mark Fasheh <mark.fasheh@oracle.com>
This adds a new dlmglue lock type which is intended to back flock()
requests.
Since these locks are driven from userspace, usage rules are much more
liberal than the typical Ocfs2 internal cluster lock. As a result, we can't
make use of most dlmglue features - lock caching and lock level
optimizations in particular. Additionally, userspace is free to deadlock
itself, so we have to deal with that in the same way as the rest of the
kernel - by allowing a signal to abort a lock request.
In order to keep ocfs2_cluster_lock() complexity down, ocfs2_file_lock()
does it's own dlm coordination. We still use the same helper functions
though, so duplicated code is kept to a minimum.
Signed-off-by: Mark Fasheh <mark.fasheh@oracle.com>
Call this the "inode_lock" now, since it covers both data and meta data.
This patch makes no functional changes.
Signed-off-by: Mark Fasheh <mark.fasheh@oracle.com>
The meta lock now covers both meta data and data, so this just removes the
now-redundant data lock.
Combining locks saves us a round of lock mastery per inode and one less lock
to ping between nodes during read/write.
We don't lose much - since meta locks were always held before a data lock
(and at the same level) ordered writeout mode (the default) ensured that
flushing for the meta data lock also pushed out data anyways.
Signed-off-by: Mark Fasheh <mark.fasheh@oracle.com>
The node maps that are set/unset by these votes are no longer relevant, thus
we can remove the mount and umount votes. Since those are the last two
remaining votes, we can also remove the entire vote infrastructure.
The vote thread has been renamed to the downconvert thread, and the small
amount of functionality related to managing it has been moved into
fs/ocfs2/dlmglue.c. All references to votes have been removed or updated.
Signed-off-by: Mark Fasheh <mark.fasheh@oracle.com>
Add the disk, network and memory structures needed to support data in inode.
Struct ocfs2_inline_data is defined and embedded in ocfs2_dinode for storing
inline data.
A new inode field, i_dyn_features, is added to facilitate tracking of
dynamic inode state. Since it will be used often, we want to mirror it on
ocfs2_inode_info, and transfer it via the meta data lvb.
Signed-off-by: Mark Fasheh <mark.fasheh@oracle.com>
Reviewed-by: Joel Becker <joel.becker@oracle.com>
This patch makes the following needlessly global functions static:
- aops.c: ocfs2_write_data_page()
- dlmglue.c: ocfs2_dump_meta_lvb_info()
- file.c: ocfs2_set_inode_size()
Signed-off-by: Adrian Bunk <bunk@stusta.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Mark Fasheh <mark.fasheh@oracle.com>
Ocfs2 currently does cluster-wide node messaging to check the open state of
an inode during delete. This patch removes that mechanism in favor of an
inode cluster lock which is taken at shared read when an inode is first read
and dropped in clear_inode(). This allows a deleting node to test the
liveness of an inode by attempting to take an exclusive lock.
Signed-off-by: Tiger Yang <tiger.yang@oracle.com>
Signed-off-by: Mark Fasheh <mark.fasheh@oracle.com>
This patch adds the core routines for updating atime in ocfs2.
Signed-off-by: Tiger Yang <tiger.yang@oracle.com>
Signed-off-by: Mark Fasheh <mark.fasheh@oracle.com>
This patch makes the needlessly global ocfs2_create_new_lock() static.
Signed-off-by: Adrian Bunk <bunk@stusta.de>
Signed-off-by: Mark Fasheh <mark.fasheh@oracle.com>
OCFS2 puts inode meta data in the "lock value block" provided by the DLM.
Typically, i_generation is encoded in the lock name so that a deleted inode
on and a new one in the same block don't share the same lvb.
Unfortunately, that scheme means that the read in ocfs2_read_locked_inode()
is potentially thrown away as soon as the meta data lock is taken - we
cannot encode the lock name without first knowing i_generation, which
requires a disk read.
This patch encodes i_generation in the inode meta data lvb, and removes the
value from the inode meta data lock name. This way, the read can be covered
by a lock, and at the same time we can distinguish between an up to date and
a stale LVB.
This will help cold-cache stat(2) performance in particular.
Since this patch changes the protocol version, we take the opportunity to do
a minor re-organization of two of the LVB fields.
Signed-off-by: Mark Fasheh <mark.fasheh@oracle.com>
When i_generation is removed from the lockname, this will help us determine
whether a meta data lvb has information that is in sync with the local
struct inode.
Signed-off-by: Mark Fasheh <mark.fasheh@oracle.com>
lvb_version doesn't need to be a whole 32 bits. Make it an 8 bit field to
free up some space. This should be backwards compatible until we use one of
the fields, in which case we'd bump the lvb version anyway.
Signed-off-by: Mark Fasheh <mark.fasheh@oracle.com>
Replace the dentry vote mechanism with a cluster lock which covers a set
of dentries. This allows us to force d_delete() only on nodes which actually
care about an unlink.
Every node that does a ->lookup() gets a read only lock on the dentry, until
an unlink during which the unlinking node, will request an exclusive lock,
forcing the other nodes who care about that dentry to d_delete() it. The
effect is that we retain a very lightweight ->d_revalidate(), and at the
same time get to make large improvements to the average case performance of
the ocfs2 unlink and rename operations.
This patch adds the cluster lock type which OCFS2 can attach to
dentries. A small number of fs/ocfs2/dcache.c functions are stubbed
out so that this change can compile.
Signed-off-by: Mark Fasheh <mark.fasheh@oracle.com>
Support immutable, and other attributes.
Some renaming and other minor fixes done by myself.
Signed-off-by: Herbert Poetzl <herbert@13thfloor.at>
Signed-off-by: Mark Fasheh <mark.fasheh@oracle.com>
Eric Ren