Inode radix tree tagging for reclaim passes a lot of unnecessary
variables around. Over time the xfs-perag has grown a xfs_mount
backpointer, and an internal agno so we don't need to pass other
variables into the tagging functions to supply this information.
Rework the functions to pass the minimal variable set required
and simplify the internal logic and flow.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
The cluster inode variable uses unconventional naming - iq - which
makes it hard to distinguish it between the inode passed into the
function - ip - and that is a vector for mistakes to be made.
Rename all the cluster inode variables to use a more conventional
prefixes to reduce potential future confusion (cilist, cilist_size,
cip).
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Dave Chinner <david@fromorbit.com>
xfs_iflush_cluster() does a gang lookup on the radix tree, meaning
it can find inodes beyond the current cluster if there is sparse
cache population. gang lookups return results in ascending index
order, so stop trying to cluster inodes once the first inode outside
the cluster mask is detected.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Dave Chinner <david@fromorbit.com>
The last thing we do before using call_rcu() on an xfs_inode to be
freed is mark it as invalid. This means there is a window between
when we know for certain that the inode is going to be freed and
when we do actually mark it as "freed".
This is important in the context of RCU lookups - we can look up the
inode, find that it is valid, and then use it as such not realising
that it is in the final stages of being freed.
As such, mark the inode as being invalid the moment we know it is
going to be reclaimed. This can be done while we still hold the
XFS_ILOCK_EXCL and the flush lock in xfs_inode_reclaim, meaning that
it occurs well before we remove it from the radix tree, and that
the i_flags_lock, the XFS_ILOCK and the inode flush lock all act as
synchronisation points for detecting that an inode is about to go
away.
For defensive purposes, this allows us to add a further check to
xfs_iflush_cluster to ensure we skip inodes that are being freed
after we grab the XFS_ILOCK_SHARED and the flush lock - we know that
if the inode number if valid while we have these locks held we know
that it has not progressed through reclaim to the point where it is
clean and is about to be freed.
[bfoster: fixed __xfs_inode_clear_reclaim() using ip->i_ino after it
had already been zeroed.]
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
The xfs_inode freed in xfs_inode_free() has multiple allocated
structures attached to it. We free these in xfs_inode_free() before
we mark the inode as invalid, and before we run call_rcu() to queue
the structure for freeing.
Unfortunately, this freeing can race with other accesses that are in
the RCU current grace period that have found the inode in the radix
tree with a valid state. This includes xfs_iflush_cluster(), which
calls xfs_inode_clean(), and that accesses the inode log item on the
xfs_inode.
The log item structure is freed in xfs_inode_free(), so there is the
possibility we can be accessing freed memory in xfs_iflush_cluster()
after validating the xfs_inode structure as being valid for this RCU
context. Hence we can get spuriously incorrect clean state returned
from such checks. This can lead to use thinking the inode is dirty
when it is, in fact, clean, and so incorrectly attaching it to the
buffer for IO and completion processing.
This then leads to use-after-free situations on the xfs_inode itself
if the IO completes after the current RCU grace period expires. The
buffer callbacks will access the xfs_inode and try to do all sorts
of things it shouldn't with freed memory.
IOWs, xfs_iflush_cluster() only works correctly when racing with
inode reclaim if the inode log item is present and correctly stating
the inode is clean. If the inode is being freed, then reclaim has
already made sure the inode is clean, and hence xfs_iflush_cluster
can skip it. However, we are accessing the inode inode under RCU
read lock protection and so also must ensure that all dynamically
allocated memory we reference in this context is not freed until the
RCU grace period expires.
To fix this, move all the potential memory freeing into
xfs_inode_free_callback() so that we are guarantee RCU protected
lookup code will always have the memory structures it needs
available during the RCU grace period that lookup races can occur
in.
Discovered-by: Brain Foster <bfoster@redhat.com>
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Dave Chinner <david@fromorbit.com>
When unmounting XFS, we call:
xfs_inode_free => xfs_idestroy_fork => xfs_iext_destroy
This goes over the whole indirection array and calls
xfs_iext_irec_remove for each one of the erps (from the last one to
the first one). As a result, we keep shrinking (reallocating
actually) the indirection array until we shrink out all of its
elements. When we have files with huge numbers of extents, umount
takes 30-80 sec, depending on the amount of files that XFS loaded
and the amount of indirection entries of each file. The unmount
stack looks like:
[<ffffffffc0b6d200>] xfs_iext_realloc_indirect+0x40/0x60 [xfs]
[<ffffffffc0b6cd8e>] xfs_iext_irec_remove+0xee/0xf0 [xfs]
[<ffffffffc0b6cdcd>] xfs_iext_destroy+0x3d/0xb0 [xfs]
[<ffffffffc0b6cef6>] xfs_idestroy_fork+0xb6/0xf0 [xfs]
[<ffffffffc0b87002>] xfs_inode_free+0xb2/0xc0 [xfs]
[<ffffffffc0b87260>] xfs_reclaim_inode+0x250/0x340 [xfs]
[<ffffffffc0b87583>] xfs_reclaim_inodes_ag+0x233/0x370 [xfs]
[<ffffffffc0b8823d>] xfs_reclaim_inodes+0x1d/0x20 [xfs]
[<ffffffffc0b96feb>] xfs_unmountfs+0x7b/0x1a0 [xfs]
[<ffffffffc0b98e4d>] xfs_fs_put_super+0x2d/0x70 [xfs]
[<ffffffff811e9e36>] generic_shutdown_super+0x76/0x100
[<ffffffff811ea207>] kill_block_super+0x27/0x70
[<ffffffff811ea519>] deactivate_locked_super+0x49/0x60
[<ffffffff811eaaee>] deactivate_super+0x4e/0x70
[<ffffffff81207593>] cleanup_mnt+0x43/0x90
[<ffffffff81207632>] __cleanup_mnt+0x12/0x20
[<ffffffff8108f8e7>] task_work_run+0xa7/0xe0
[<ffffffff81014ff7>] do_notify_resume+0x97/0xb0
[<ffffffff81717c6f>] int_signal+0x12/0x17
Further, this reallocation prevents us from freeing the extent list
from a RCU callback as allocation can block. Hence if the extent
list is in indirect format, optimise the freeing of the extent list
to only use kmem_free calls by freeing entire extent buffer pages at
a time, rather than extent by extent.
[dchinner: simplified freeing loop based on Christoph's suggestion]
Signed-off-by: Alex Lyakas <alex@zadarastorage.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
We don't write back stale inodes so we should skip them in
xfs_iflush_cluster, too.
cc: <stable@vger.kernel.org> # 3.10.x-
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Dave Chinner <david@fromorbit.com>
Some careless idiot(*) wrote crap code in commit 1a3e8f3 ("xfs:
convert inode cache lookups to use RCU locking") back in late 2010,
and so xfs_iflush_cluster checks the wrong inode for whether it is
still valid under RCU protection. Fix it to lock and check the
correct inode.
(*) Careless-idiot: Dave Chinner <dchinner@redhat.com>
cc: <stable@vger.kernel.org> # 3.10.x-
Discovered-by: Brain Foster <bfoster@redhat.com>
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Dave Chinner <david@fromorbit.com>
When a failure due to an inode buffer occurs, the error handling
fails to abort the inode writeback correctly. This can result in the
inode being reclaimed whilst still in the AIL, leading to
use-after-free situations as well as filesystems that cannot be
unmounted as the inode log items left in the AIL never get removed.
Fix this by ensuring fatal errors from xfs_imap_to_bp() result in
the inode flush being aborted correctly.
cc: <stable@vger.kernel.org> # 3.10.x-
Reported-by: Shyam Kaushik <shyam@zadarastorage.com>
Diagnosed-by: Shyam Kaushik <shyam@zadarastorage.com>
Tested-by: Shyam Kaushik <shyam@zadarastorage.com>
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Dave Chinner <david@fromorbit.com>
Joe Lawrence reported a list_add corruption with 4.6-rc1 when
testing some custom md administration code that made it's own
block device nodes for the md array. The simple test loop of:
for i in {0..100}; do
mknod --mode=0600 $tmp/tmp_node b $MAJOR $MINOR
mdadm --detail --export $tmp/tmp_node > /dev/null
rm -f $tmp/tmp_node
done
Would produce this warning in bd_acquire() when mdadm opened the
device node:
list_add double add: new=ffff88043831c7b8, prev=ffff8804380287d8, next=ffff88043831c7b8.
And then produce this from bd_forget from kdevtmpfs evicting a block
dev inode:
list_del corruption. prev->next should be ffff8800bb83eb10, but was ffff88043831c7b8
This is a regression caused by commit c19b3b05 ("xfs: mode di_mode
to vfs inode"). The issue is that xfs_inactive() frees the
unlinked inode, and the above commit meant that this freeing zeroed
the mode in the struct inode. The problem is that after evict() has
called ->evict_inode, it expects the i_mode to be intact so that it
can call bd_forget() or cd_forget() to drop the reference to the
block device inode attached to the XFS inode.
In reality, the only thing we do in xfs_fs_evict_inode() that is not
generic is call xfs_inactive(). We can move the xfs_inactive() call
to xfs_fs_destroy_inode() without any problems at all, and this
will leave the VFS inode intact until it is completely done with it.
So, remove xfs_fs_evict_inode(), and do the work it used to do in
->destroy_inode instead.
cc: <stable@vger.kernel.org> # 4.6
Reported-by: Joe Lawrence <joe.lawrence@stratus.com>
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
Pull Ceph updates from Sage Weil:
"There is quite a bit here, including some overdue refactoring and
cleanup on the mon_client and osd_client code from Ilya, scattered
writeback support for CephFS and a pile of bug fixes from Zheng, and a
few random cleanups and fixes from others"
[ I already decided not to pull this because of it having been rebased
recently, but ended up changing my mind after all. Next time I'll
really hold people to it. Oh well. - Linus ]
* 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/sage/ceph-client: (34 commits)
libceph: use KMEM_CACHE macro
ceph: use kmem_cache_zalloc
rbd: use KMEM_CACHE macro
ceph: use lookup request to revalidate dentry
ceph: kill ceph_get_dentry_parent_inode()
ceph: fix security xattr deadlock
ceph: don't request vxattrs from MDS
ceph: fix mounting same fs multiple times
ceph: remove unnecessary NULL check
ceph: avoid updating directory inode's i_size accidentally
ceph: fix race during filling readdir cache
libceph: use sizeof_footer() more
ceph: kill ceph_empty_snapc
ceph: fix a wrong comparison
ceph: replace CURRENT_TIME by current_fs_time()
ceph: scattered page writeback
libceph: add helper that duplicates last extent operation
libceph: enable large, variable-sized OSD requests
libceph: osdc->req_mempool should be backed by a slab pool
libceph: make r_request msg_size calculation clearer
...
Pull orangefs filesystem from Mike Marshall.
This finally merges the long-pending orangefs filesystem, which has been
much cleaned up with input from Al Viro over the last six months. From
the documentation file:
"OrangeFS is an LGPL userspace scale-out parallel storage system. It
is ideal for large storage problems faced by HPC, BigData, Streaming
Video, Genomics, Bioinformatics.
Orangefs, originally called PVFS, was first developed in 1993 by Walt
Ligon and Eric Blumer as a parallel file system for Parallel Virtual
Machine (PVM) as part of a NASA grant to study the I/O patterns of
parallel programs.
Orangefs features include:
- Distributes file data among multiple file servers
- Supports simultaneous access by multiple clients
- Stores file data and metadata on servers using local file system
and access methods
- Userspace implementation is easy to install and maintain
- Direct MPI support
- Stateless"
see Documentation/filesystems/orangefs.txt for more in-depth details.
* tag 'ofs-pull-tag-1' of git://git.kernel.org/pub/scm/linux/kernel/git/hubcap/linux: (174 commits)
orangefs: fix orangefs_superblock locking
orangefs: fix do_readv_writev() handling of error halfway through
orangefs: have ->kill_sb() evict the VFS side of things first
orangefs: sanitize ->llseek()
orangefs-bufmap.h: trim unused junk
orangefs: saner calling conventions for getting a slot
orangefs_copy_{to,from}_bufmap(): don't pass bufmap pointer
orangefs: get rid of readdir_handle_s
ornagefs: ensure that truncate has an up to date inode size
orangefs: move code which sets i_link to orangefs_inode_getattr
orangefs: remove needless wrapper around GFP_KERNEL
orangefs: remove wrapper around mutex_lock(&inode->i_mutex)
orangefs: refactor inode type or link_target change detection
orangefs: use new getattr for revalidate and remove old getattr
orangefs: use new getattr in inode getattr and permission
orangefs: use new orangefs_inode_getattr to get size in write and llseek
orangefs: use new orangefs_inode_getattr to create new inodes
orangefs: rename orangefs_inode_getattr to orangefs_inode_old_getattr
orangefs: remove inode->i_lock wrapper
orangefs: put register_chrdev immediately before register_filesystem
...
Commit 0b81d07790 ("fs crypto: move per-file encryption from f2fs
tree to fs/crypto") moved the f2fs crypto files to fs/crypto/ and
renamed the symbol prefixes from "f2fs_" to "fscrypt_" (and from "F2FS_"
to just "FS" for preprocessor symbols).
Because of the symbol renaming, it's a bit hard to see it as a file
move: use
git show -M30 0b81d07790
to lower the rename detection to just 30% similarity and make git show
the files as renamed (the header file won't be shown as a rename even
then - since all it contains is symbol definitions, it looks almost
completely different).
Even with the renames showing as renames, the diffs are not all that
easy to read, since so much is just the renames. But Eric Biggers
noticed that it's not just all renames: the initialization of the
xts_tweak had been broken too, using the inode number rather than the
page offset.
That's not right - it makes the xfs_tweak the same for all pages of each
inode. It _might_ make sense to make the xfs_tweak contain both the
offset _and_ the inode number, but not just the inode number.
Reported-by: Eric Biggers <ebiggers3@gmail.com>
Cc: Jaegeuk Kim <jaegeuk@kernel.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
* switch orangefs_remount() to taking ORANGEFS_SB(sb) instead of sb
* remove from the list _before_ orangefs_unmount() - request_mutex
in the latter will make sure that nothing observed in the loop in
ORANGEFS_DEV_REMOUNT_ALL handling will get freed until the end
of loop
* on removal, keep the forward pointer and zero the back one. That
way we can drop and regain the spinlock in the loop body (again,
ORANGEFS_DEV_REMOUNT_ALL one) and still be able to get to the
rest of the list.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
Signed-off-by: Mike Marshall <hubcap@omnibond.com>
Error should only be returned if nothing had been read/written.
Otherwise we need to report a short read/write instead.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
Signed-off-by: Mike Marshall <hubcap@omnibond.com>
a) open files can't have NULL inodes
b) it's SEEK_END, not ORANGEFS_SEEK_END; no need to get cute.
c) make_bad_inode() on lseek()?
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
Signed-off-by: Mike Marshall <hubcap@omnibond.com>
just have it return the slot number or -E... - the caller checks
the sign anyway
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
Signed-off-by: Mike Marshall <hubcap@omnibond.com>
no point, really - we couldn't keep those across the calls of
getdents(); it would be too easy to DoS, having all slots exhausted.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
Signed-off-by: Mike Marshall <hubcap@omnibond.com>
Now function ocfs2_replay_truncate_records() first modifies tl_used,
then calls ocfs2_extend_trans() to extend transactions for gd and alloc
inode used for freeing clusters. jbd2_journal_restart() may be called
and it may happen that tl_used in truncate log is decreased but the
clusters are not freed, which means these clusters are lost. So we
should avoid extending transactions in these two operations.
Signed-off-by: joyce.xue <xuejiufei@huawei.com>
Reviewed-by: Mark Fasheh <mfasheh@suse.de>
Acked-by: Joseph Qi <joseph.qi@huawei.com>
Cc: Joel Becker <jlbec@evilplan.org>
Cc: Junxiao Bi <junxiao.bi@oracle.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
I found that jbd2_journal_restart() is called in some places without
keeping things consistently before. However, jbd2_journal_restart() may
commit the handle's transaction and restart another one. If the first
transaction is committed successfully while another not, it may cause
filesystem inconsistency or read only. This is an effort to fix this
kind of problems.
This patch (of 3):
The following functions will be called while truncating an extent:
ocfs2_remove_btree_range
-> ocfs2_start_trans
-> ocfs2_remove_extent
-> ocfs2_truncate_rec
-> ocfs2_extend_rotate_transaction
-> jbd2_journal_restart if jbd2_journal_extend fail
-> ocfs2_rotate_tree_left
-> ocfs2_remove_rightmost_path
-> ocfs2_extend_rotate_transaction
-> ocfs2_unlink_subtree
-> ocfs2_update_edge_lengths
-> ocfs2_extend_trans
-> jbd2_journal_restart if jbd2_journal_extend fail
-> ocfs2_et_update_clusters
-> ocfs2_commit_trans
jbd2_journal_restart() may be called and it may happened that the buffers
dirtied in ocfs2_truncate_rec() are committed while buffers dirtied in
ocfs2_et_update_clusters() are not, the total clusters on extent tree and
i_clusters in ocfs2_dinode is inconsistency. So the clusters got from
ocfs2_dinode is incorrect, and it also cause read-only problem when call
ocfs2_commit_truncate() with the error message: "Inode %llu has empty
extent block at %llu".
We should extend enough credits for function ocfs2_remove_rightmost_path
and ocfs2_update_edge_lengths to avoid this inconsistency.
Signed-off-by: joyce.xue <xuejiufei@huawei.com>
Acked-by: Joseph Qi <joseph.qi@huawei.com>
Cc: Mark Fasheh <mfasheh@suse.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>
We have found a bug when two nodes doing umount one after another.
1) Node 1 migrate a lockres that has 3 locks in grant queue such as
N2(PR)<->N3(NL)<->N4(PR) to N2. After migration, lvb of the lock
N3(NL) and N4(PR) are empty on node 2 because migration target do not
copy lvb to these two lock.
2) Node 3 want to convert to PR, it can be granted in
__dlmconvert_master(), and the order of these locks is unchanged. The
lvb of the lock N3(PR) on node 2 is copyed from lockres in function
dlm_update_lvb() while the lvb of lock N4(PR) is still empty.
3) Node 2 want to leave domain, it will migrate this lockres to node 3.
Then node 2 will trigger the BUG in dlm_prepare_lvb_for_migration()
when adding the lock N4(PR) to mres with the following message because
the lvb of mres is already copied from lock N3(PR), but the lvb of lock
N4(PR) is empty.
"Mismatched lvb in lock cookie=%u:%llu, name=%.*s, node=%u"
[akpm@linux-foundation.org: tweak comment]
Signed-off-by: xuejiufei <xuejiufei@huawei.com>
Acked-by: Joseph Qi <joseph.qi@huawei.com>
Cc: Mark Fasheh <mfasheh@suse.de>
Cc: Joel Becker <jlbec@evilplan.org>
Cc: Junxiao Bi <junxiao.bi@oracle.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
In update_backups() there exists a problem of crossing the boundary as
follows:
we assume that lun will be resized to 1TB(cluster_size is 32kb), it will
include 0~33554431 cluster, in update_backups func, it will backup super
block in location of 1TB which is the 33554432th cluster, so the
phenomenon of crossing the boundary happens.
Signed-off-by: Yiwen Jiang <jiangyiwen@huawei.com>
Reviewed-by: Joseph Qi <joseph.qi@huawei.com>
Cc: Xue jiufei <xuejiufei@huawei.com>
Cc: Mark Fasheh <mfasheh@suse.de>
Cc: Joel Becker <jlbec@evilplan.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
This patch fixes a deadlock, as follows:
Node 1 Node 2 Node 3
1)volume a and b are only mount vol a only mount vol b
mounted
2) start to mount b start to mount a
3) check hb of Node 3 check hb of Node 2
in vol a, qs_holds++ in vol b, qs_holds++
4) -------------------- all nodes' network down --------------------
5) progress of mount b the same situation as
failed, and then call Node 2
ocfs2_dismount_volume.
but the process is hung,
since there is a work
in ocfs2_wq cannot beo
completed. This work is
about vol a, because
ocfs2_wq is global wq.
BTW, this work which is
scheduled in ocfs2_wq is
ocfs2_orphan_scan_work,
and the context in this work
needs to take inode lock
of orphan_dir, because
lockres owner are Node 1 and
all nodes' nework has been down
at the same time, so it can't
get the inode lock.
6) Why can't this node be fenced
when network disconnected?
Because the process of
mount is hung what caused qs_holds
is not equal 0.
Because all works in the ocfs2_wq are relative to the super block.
The solution is to change the ocfs2_wq from global to local. In other
words, move it into struct ocfs2_super.
Signed-off-by: Yiwen Jiang <jiangyiwen@huawei.com>
Reviewed-by: Joseph Qi <joseph.qi@huawei.com>
Cc: Xue jiufei <xuejiufei@huawei.com>
Cc: Mark Fasheh <mfasheh@suse.de>
Cc: Joel Becker <jlbec@evilplan.org>
Cc: Cc: Junxiao Bi <junxiao.bi@oracle.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
When master handles convert request, it queues ast first and then
returns status. This may happen that the ast is sent before the request
status because the above two messages are sent by two threads. And
right after the ast is sent, if master down, it may trigger BUG in
dlm_move_lockres_to_recovery_list in the requested node because ast
handler moves it to grant list without clear lock->convert_pending. So
remove BUG_ON statement and check if the ast is processed in
dlmconvert_remote.
Signed-off-by: Joseph Qi <joseph.qi@huawei.com>
Reported-by: Yiwen Jiang <jiangyiwen@huawei.com>
Cc: Junxiao Bi <junxiao.bi@oracle.com>
Cc: Mark Fasheh <mfasheh@suse.de>
Cc: Joel Becker <jlbec@evilplan.org>
Cc: Tariq Saeed <tariq.x.saeed@oracle.com>
Cc: Junxiao Bi <junxiao.bi@oracle.com>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
There is a race window between dlmconvert_remote and
dlm_move_lockres_to_recovery_list, which will cause a lock with
OCFS2_LOCK_BUSY in grant list, thus system hangs.
dlmconvert_remote
{
spin_lock(&res->spinlock);
list_move_tail(&lock->list, &res->converting);
lock->convert_pending = 1;
spin_unlock(&res->spinlock);
status = dlm_send_remote_convert_request();
>>>>>> race window, master has queued ast and return DLM_NORMAL,
and then down before sending ast.
this node detects master down and calls
dlm_move_lockres_to_recovery_list, which will revert the
lock to grant list.
Then OCFS2_LOCK_BUSY won't be cleared as new master won't
send ast any more because it thinks already be authorized.
spin_lock(&res->spinlock);
lock->convert_pending = 0;
if (status != DLM_NORMAL)
dlm_revert_pending_convert(res, lock);
spin_unlock(&res->spinlock);
}
In this case, check if res->state has DLM_LOCK_RES_RECOVERING bit set
(res is still in recovering) or res master changed (new master has
finished recovery), reset the status to DLM_RECOVERING, then it will
retry convert.
Signed-off-by: Joseph Qi <joseph.qi@huawei.com>
Reported-by: Yiwen Jiang <jiangyiwen@huawei.com>
Reviewed-by: Junxiao Bi <junxiao.bi@oracle.com>
Cc: Mark Fasheh <mfasheh@suse.de>
Cc: Joel Becker <jlbec@evilplan.org>
Cc: Tariq Saeed <tariq.x.saeed@oracle.com>
Cc: Junxiao Bi <junxiao.bi@oracle.com>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
The code should call ocfs2_free_alloc_context() to free meta_ac &
data_ac before calling ocfs2_run_deallocs(). Because
ocfs2_run_deallocs() will acquire the system inode's i_mutex hold by
meta_ac. So try to release the lock before ocfs2_run_deallocs().
Fixes: af1310367f41 ("ocfs2: fix sparse file & data ordering issue in direct io.")
Signed-off-by: Ryan Ding <ryan.ding@oracle.com>
Acked-by: Junxiao Bi <junxiao.bi@oracle.com>
Cc: Joseph Qi <joseph.qi@huawei.com>
Cc: Mark Fasheh <mfasheh@suse.de>
Cc: Joel Becker <jlbec@evilplan.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
When doing append direct write in an already allocated cluster, and fast
path in ocfs2_dio_get_block() is triggered, function
ocfs2_dio_end_io_write() will be skipped as there is no context
allocated.
As a result, the disk file size will not be changed as it should be.
The solution is to skip fast path when we are about to change file size.
Fixes: af1310367f41 ("ocfs2: fix sparse file & data ordering issue in direct io.")
Signed-off-by: Ryan Ding <ryan.ding@oracle.com>
Acked-by: Junxiao Bi <junxiao.bi@oracle.com>
Cc: Joseph Qi <joseph.qi@huawei.com>
Cc: Mark Fasheh <mfasheh@suse.de>
Cc: Joel Becker <jlbec@evilplan.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Take ip_alloc_sem to prevent concurrent access to extent tree, which may
cause the extent tree in an unstable state.
Signed-off-by: Ryan Ding <ryan.ding@oracle.com>
Reviewed-by: Junxiao Bi <junxiao.bi@oracle.com>
Cc: Joseph Qi <joseph.qi@huawei.com>
Cc: Mark Fasheh <mfasheh@suse.de>
Cc: Joel Becker <jlbec@evilplan.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
In the current implementation of unaligned aio+dio, lock order behave as
follow:
in user process context:
-> call io_submit()
-> get i_mutex
<== window1
-> get ip_unaligned_aio
-> submit direct io to block device
-> release i_mutex
-> io_submit() return
in dio work queue context(the work queue is created in __blockdev_direct_IO):
-> release ip_unaligned_aio
<== window2
-> get i_mutex
-> clear unwritten flag & change i_size
-> release i_mutex
There is a limitation to the thread number of dio work queue. 256 at
default. If all 256 thread are in the above 'window2' stage, and there
is a user process in the 'window1' stage, the system will became
deadlock. Since the user process hold i_mutex to wait ip_unaligned_aio
lock, while there is a direct bio hold ip_unaligned_aio mutex who is
waiting for a dio work queue thread to be schedule. But all the dio
work queue thread is waiting for i_mutex lock in 'window2'.
This case only happened in a test which send a large number(more than
256) of aio at one io_submit() call.
My design is to remove ip_unaligned_aio lock. Change it to a sync io
instead. Just like ip_unaligned_aio lock, serialize the unaligned aio
dio.
[akpm@linux-foundation.org: remove OCFS2_IOCB_UNALIGNED_IO, per Junxiao Bi]
Signed-off-by: Ryan Ding <ryan.ding@oracle.com>
Reviewed-by: Junxiao Bi <junxiao.bi@oracle.com>
Cc: Joseph Qi <joseph.qi@huawei.com>
Cc: Mark Fasheh <mfasheh@suse.de>
Cc: Joel Becker <jlbec@evilplan.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Clean up ocfs2_file_write_iter & ocfs2_prepare_inode_for_write:
* remove append dio check: it will be checked in ocfs2_direct_IO()
* remove file hole check: file hole is supported for now
* remove inline data check: it will be checked in ocfs2_direct_IO()
* remove the full_coherence check when append dio: we will get the
inode_lock in ocfs2_dio_get_block, there is no need to fall back to
buffer io to ensure the coherence semantics.
Now the drop dio procedure is gone. :)
[akpm@linux-foundation.org: remove unused label]
Signed-off-by: Ryan Ding <ryan.ding@oracle.com>
Reviewed-by: Junxiao Bi <junxiao.bi@oracle.com>
Cc: Joseph Qi <joseph.qi@huawei.com>
Cc: Mark Fasheh <mfasheh@suse.de>
Cc: Joel Becker <jlbec@evilplan.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
There are mainly three issues in the direct io code path after commit
24c40b329e ("ocfs2: implement ocfs2_direct_IO_write"):
* Does not support sparse file.
* Does not support data ordering. eg: when write to a file hole, it
will alloc extent first. If system crashed before io finished, data
will corrupt.
* Potential risk when doing aio+dio. The -EIOCBQUEUED return value is
likely to be ignored by ocfs2_direct_IO_write().
To resolve above problems, re-design direct io code with following ideas:
* Use buffer io to fill in holes. And this will make better
performance also.
* Clear unwritten after direct write finished. So we can make sure
meta data changes after data write to disk. (Unwritten extent is
invisible to user, from user's view, meta data is not changed when
allocate an unwritten extent.)
* Clear ocfs2_direct_IO_write(). Do all ending work in end_io.
This patch has passed fs,dio,ltp-aiodio.part1,ltp-aiodio.part2,ltp-aiodio.part4
test cases of ltp.
For performance improvement, see following test result:
ocfs2 cluster size 1MB, ocfs2 volume is mounted on /mnt/.
The original way:
+ rm /mnt/test.img -f
+ dd if=/dev/zero of=/mnt/test.img bs=4K count=1048576 oflag=direct
1048576+0 records in
1048576+0 records out
4294967296 bytes (4.3 GB) copied, 1707.83 s, 2.5 MB/s
+ rm /mnt/test.img -f
+ dd if=/dev/zero of=/mnt/test.img bs=256K count=16384 oflag=direct
16384+0 records in
16384+0 records out
4294967296 bytes (4.3 GB) copied, 582.705 s, 7.4 MB/s
After this patch:
+ rm /mnt/test.img -f
+ dd if=/dev/zero of=/mnt/test.img bs=4K count=1048576 oflag=direct
1048576+0 records in
1048576+0 records out
4294967296 bytes (4.3 GB) copied, 64.6412 s, 66.4 MB/s
+ rm /mnt/test.img -f
+ dd if=/dev/zero of=/mnt/test.img bs=256K count=16384 oflag=direct
16384+0 records in
16384+0 records out
4294967296 bytes (4.3 GB) copied, 34.7611 s, 124 MB/s
Signed-off-by: Ryan Ding <ryan.ding@oracle.com>
Reviewed-by: Junxiao Bi <junxiao.bi@oracle.com>
Cc: Joseph Qi <joseph.qi@huawei.com>
Cc: Mark Fasheh <mfasheh@suse.de>
Cc: Joel Becker <jlbec@evilplan.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
To support direct io in ocfs2_write_begin_nolock & ocfs2_write_end_nolock.
There is still one issue in the direct write procedure.
phase 1: alloc extent with UNWRITTEN flag
phase 2: submit direct data to disk, add zero page to page cache
phase 3: clear UNWRITTEN flag when data has been written to disk
When there are 2 direct write A(0~3KB),B(4~7KB) writing to the same
cluster 0~7KB (cluster size 8KB). Write request A arrive phase 2 first,
it will zero the region (4~7KB). Before request A enter to phase 3,
request B arrive phase 2, it will zero region (0~3KB). This is just like
request B steps request A.
To resolve this issue, we should let request B knows this cluster is already
under zero, to prevent it from steps the previous write request.
This patch will add function ocfs2_unwritten_check() to do this job. It
will record all clusters that are under direct write(it will be recorded
in the 'ip_unwritten_list' member of inode info), and prevent the later
direct write writing to the same cluster to do the zero work again.
Signed-off-by: Ryan Ding <ryan.ding@oracle.com>
Reviewed-by: Junxiao Bi <junxiao.bi@oracle.com>
Cc: Joseph Qi <joseph.qi@huawei.com>
Cc: Mark Fasheh <mfasheh@suse.de>
Cc: Joel Becker <jlbec@evilplan.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
To support direct io in ocfs2_write_begin_nolock & ocfs2_write_end_nolock.
Direct io needs to get the physical address from write_begin, to map the
user page. This patch is to change the arg 'phys' of
ocfs2_write_cluster to a pointer, so it can be retrieved to write_begin.
And we can retrieve it to the direct io procedure.
Signed-off-by: Ryan Ding <ryan.ding@oracle.com>
Reviewed-by: Junxiao Bi <junxiao.bi@oracle.com>
Cc: Joseph Qi <joseph.qi@huawei.com>
Cc: Mark Fasheh <mfasheh@suse.de>
Cc: Joel Becker <jlbec@evilplan.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
To support direct io in ocfs2_write_begin_nolock & ocfs2_write_end_nolock.
Append direct io do not change i_size in get block phase. It only move
to orphan when starting write. After data is written to disk, it will
delete itself from orphan and update i_size. So skip i_size change
section in write_begin for direct io.
And when there is no extents alloc, no meta data changes needed for
direct io (since write_begin start trans for 2 reason: alloc extents &
change i_size. Now none of them needed). So we can skip start trans
procedure.
Signed-off-by: Ryan Ding <ryan.ding@oracle.com>
Reviewed-by: Junxiao Bi <junxiao.bi@oracle.com>
Cc: Joseph Qi <joseph.qi@huawei.com>
Cc: Mark Fasheh <mfasheh@suse.de>
Cc: Joel Becker <jlbec@evilplan.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
To support direct io in ocfs2_write_begin_nolock & ocfs2_write_end_nolock.
Direct io data will not appear in buffer. The w_target_page member will
not be filled by direct io. So avoid to use it when it's NULL. Unlinke
buffer io and mmap, direct io will call write_begin with more than 1
page a time. So the target_index is not sufficient to describe the
actual data. change it to a range start at target_index, end in
end_index.
Signed-off-by: Ryan Ding <ryan.ding@oracle.com>
Reviewed-by: Junxiao Bi <junxiao.bi@oracle.com>
Cc: Joseph Qi <joseph.qi@huawei.com>
Cc: Mark Fasheh <mfasheh@suse.de>
Cc: Joel Becker <jlbec@evilplan.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
To support direct io in ocfs2_write_begin_nolock & ocfs2_write_end_nolock.
There is a problem in ocfs2's direct io implement: if system crashed
after extents allocated, and before data return, we will get a extent
with dirty data on disk. This problem violate the journal=order
semantics, which means meta changes take effect after data written to
disk. To resolve this issue, direct write can use the UNWRITTEN flag to
describe a extent during direct data writeback. The direct write
procedure should act in the following order:
phase 1: alloc extent with UNWRITTEN flag
phase 2: submit direct data to disk, add zero page to page cache
phase 3: clear UNWRITTEN flag when data has been written to disk
This patch is to change the 'c_unwritten' member of
ocfs2_write_cluster_desc to 'c_clear_unwritten'. Means whether to clear
the unwritten flag. It do not care if a extent is allocated or not.
And use 'c_new' to specify a newly allocated extent. So the direct io
procedure can use c_clear_unwritten to control the UNWRITTEN bit on
extent.
Signed-off-by: Ryan Ding <ryan.ding@oracle.com>
Reviewed-by: Junxiao Bi <junxiao.bi@oracle.com>
Cc: Joseph Qi <joseph.qi@huawei.com>
Cc: Mark Fasheh <mfasheh@suse.de>
Cc: Joel Becker <jlbec@evilplan.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Patchset: fix ocfs2 direct io code patch to support sparse file and data
ordering semantics
The idea is to use buffer io(more precisely use the interface
ocfs2_write_begin_nolock & ocfs2_write_end_nolock) to do the zero work
beyond block size. And clear UNWRITTEN flag until direct io data has
been written to disk, which can prevent data corruption when system
crashed during direct write.
And we will also archive a better performance: eg. dd direct write new
file with block size 4KB: before this patchset:
2.5 MB/s
after this patchset:
66.4 MB/s
This patch (of 8):
To support direct io in ocfs2_write_begin_nolock &
ocfs2_write_end_nolock.
Remove unused args filp & flags. Add new arg type. The type is one of
buffer/direct/mmap. Indicate 3 way to perform write. buffer/mmap type
has implemented. direct type will be implemented later.
Signed-off-by: Ryan Ding <ryan.ding@oracle.com>
Reviewed-by: Junxiao Bi <junxiao.bi@oracle.com>
Cc: Joseph Qi <joseph.qi@huawei.com>
Cc: Mark Fasheh <mfasheh@suse.de>
Cc: Joel Becker <jlbec@evilplan.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Use kmem_cache_zalloc() instead of kmem_cache_alloc() with flag GFP_ZERO.
Signed-off-by: Geliang Tang <geliangtang@163.com>
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
If dentry has no lease, ceph_d_revalidate() previously return 0.
This causes VFS to invalidate the dentry and create a new dentry
for later lookup. Invalidating a dentry also detach any underneath
mount points. So mount point inside cephfs can disapear mystically
(even the mount point is not modified by other hosts).
The fix is using lookup request to revalidate dentry without lease.
This can partly solve the mount points disapear issue (as long as
the mount point is not modified by other hosts)
Signed-off-by: Yan, Zheng <zyan@redhat.com>
When security is enabled, security module can call filesystem's
getxattr/setxattr callbacks during d_instantiate(). For cephfs,
d_instantiate() is usually called by MDS' dispatch thread, while
handling MDS reply. If the MDS reply does not include xattrs and
corresponding caps, getxattr/setxattr need to send a new request
to MDS and waits for the reply. This makes MDS' dispatch sleep,
nobody handles later MDS replies.
The fix is make sure lookup/atomic_open reply include xattrs and
corresponding caps. So getxattr can be handled by cached xattrs.
This requires some modification to both MDS and request message.
(Client tells MDS what caps it wants; MDS encodes proper caps in
the reply)
Smack security module may call setxattr during d_instantiate().
Unlike getxattr, we can't force MDS to issue CEPH_CAP_XATTR_EXCL
to us. So just make setxattr return error when called by MDS'
dispatch thread.
Signed-off-by: Yan, Zheng <zyan@redhat.com>
If page->mapping is NULL, releasepage() callback does not get called.
Remove the unnecessary NULL check to make static code analysis tool
happy
Signed-off-by: Yan, Zheng <zyan@redhat.com>
Readdir cache uses page cache to save dentry pointers. When adding
dentry pointers to middle of a page, we need to make sure the page
already exists. Otherwise the beginning part of the page will be
invalid pointers.
Signed-off-by: Yan, Zheng <zyan@redhat.com>
