I encountered below deadlock when running fsstress
wmtruncate work truncate MDS
--------------- ------------------ --------------------------
lock i_mutex
<- truncate file
lock i_mutex (blocked)
<- revoking Fcb (filelock to MIX)
send request ->
handle request (xlock filelock)
At the initial time, there are some dirty pages in the page cache.
When the kclient receives the truncate message, it reduces inode size
and creates some 'out of i_size' dirty pages. wmtruncate work can't
truncate these dirty pages because it's blocked by the i_mutex. Later
when the kclient receives the cap message that revokes Fcb caps, It
can't flush all dirty pages because writepages() only flushes dirty
pages within the inode size.
When the MDS handles the 'truncate' request from kclient, it waits
for the filelock to become stable. But the filelock is stuck in
unstable state because it can't finish revoking kclient's Fcb caps.
The truncate pagecache locking has already caused lots of trouble
for use. I think it's time simplify it by introducing a new mutex.
We use the new mutex to prevent concurrent truncate_inode_pages().
There is no need to worry about race between buffered write and
truncate_inode_pages(), because our "get caps" mechanism prevents
them from concurrent execution.
Reviewed-by: Sage Weil <sage@inktank.com>
Signed-off-by: Yan, Zheng <zheng.z.yan@intel.com>
The MDS uses caps message to notify clients about deleted inode.
when receiving a such message, invalidate any alias of the inode.
This makes the kernel release the inode ASAP.
Signed-off-by: Yan, Zheng <zheng.z.yan@intel.com>
Reviewed-by: Sage Weil <sage@inktank.com>
If we receive new caps from the auth MDS and the non-auth MDS is
revoking the newly issued caps, we should release the caps from
the non-auth MDS. The scenario is filelock's state changes from
SYNC to LOCK. Non-auth MDS revokes Fc cap, the client gets Fc cap
from the auth MDS at the same time.
Signed-off-by: Yan, Zheng <zheng.z.yan@intel.com>
Reviewed-by: Sage Weil <sage@inktank.com>
If caps are been revoking by the auth MDS, don't consider them as
issued even they are still issued by non-auth MDS. The non-auth
MDS should also be revoking/exporting these caps, the client just
hasn't received the cap revoke/export message.
The race I encountered is: When caps are exporting to new MDS, the
client receives cap import message and cap revoke message from the
new MDS, then receives cap export message from the old MDS. When
the client receives cap revoke message from the new MDS, the revoking
caps are still issued by the old MDS, so the client does nothing.
Later when the cap export message is received, the client removes
the caps issued by the old MDS. (Another way to fix the race is
calling ceph_check_caps() in handle_cap_export())
Signed-off-by: Yan, Zheng <zheng.z.yan@intel.com>
Reviewed-by: Sage Weil <sage@inktank.com>
The locking order for pending vmtruncate is wrong, it can lead to
following race:
write wmtruncate work
------------------------ ----------------------
lock i_mutex
check i_truncate_pending check i_truncate_pending
truncate_inode_pages() lock i_mutex (blocked)
copy data to page cache
unlock i_mutex
truncate_inode_pages()
The fix is take i_mutex before calling __ceph_do_pending_vmtruncate()
Fixes: http://tracker.ceph.com/issues/5453
Signed-off-by: Yan, Zheng <zheng.z.yan@intel.com>
Reviewed-by: Sage Weil <sage@inktank.com>
Drop ignored return value. Fix allocation failure case to not leak.
Signed-off-by: Jianpeng Ma <majianpeng@gmail.com>
Reviewed-by: Sage Weil <sage@inktank.com>
We may receive old request reply from the exporter MDS after receiving
the importer MDS' cap import message.
Signed-off-by: Yan, Zheng <zheng.z.yan@intel.com>
Reviewed-by: Sage Weil <sage@inktank.com>
ceph_encode_inode_release() can race with ceph_open() and release
caps wanted by open files. So it should call __ceph_caps_wanted()
to get the wanted caps.
Signed-off-by: Yan, Zheng <zheng.z.yan@intel.com>
Reviewed-by: Sage Weil <sage@inktank.com>
There is deadlock as illustrated bellow. The fix is taking i_mutex
before getting Fw cap reference.
write truncate MDS
--------------------- -------------------- --------------
get Fw cap
lock i_mutex
lock i_mutex (blocked)
request setattr.size ->
<- revoke Fw cap
Signed-off-by: Yan, Zheng <zheng.z.yan@intel.com>
Reviewed-by: Alex Elder <elder@inktank.com>
Reviewed-by: Sage Weil <sage@inktank.com>
Current ceph code tracks directory's completeness in two places.
ceph_readdir() checks i_release_count to decide if it can set the
I_COMPLETE flag in i_ceph_flags. All other places check the I_COMPLETE
flag. This indirection introduces locking complexity.
This patch adds a new variable i_complete_count to ceph_inode_info.
Set i_release_count's value to it when marking a directory complete.
By comparing the two variables, we know if a directory is complete
Signed-off-by: Yan, Zheng <zheng.z.yan@intel.com>
commit c6ffe10015 moved the flag that tracks if the dcache contents
for a directory are complete to dentry. The problem is there are
lots of places that use ceph_dir_{set,clear,test}_complete() while
holding i_ceph_lock. but ceph_dir_{set,clear,test}_complete() may
sleep because they call dput().
This patch basically reverts that commit. For ceph_d_prune(), it's
called with both the dentry to prune and the parent dentry are
locked. So it's safe to access the parent dentry's d_inode and
clear I_COMPLETE flag.
Signed-off-by: Yan, Zheng <zheng.z.yan@intel.com>
Reviewed-by: Greg Farnum <greg@inktank.com>
Reviewed-by: Sage Weil <sage@inktank.com>
MDS ignores cap update message if migrate_seq mismatch, so when
receiving a cap import message with higher migrate_seq, set mds_want
according to the cap import message.
Signed-off-by: Yan, Zheng <zheng.z.yan@intel.com>
Reviewed-by: Greg Farnum <greg@inktank.com>
Pull Ceph updates from Sage Weil:
"A few groups of patches here. Alex has been hard at work improving
the RBD code, layout groundwork for understanding the new formats and
doing layering. Most of the infrastructure is now in place for the
final bits that will come with the next window.
There are a few changes to the data layout. Jim Schutt's patch fixes
some non-ideal CRUSH behavior, and a set of patches from me updates
the client to speak a newer version of the protocol and implement an
improved hashing strategy across storage nodes (when the server side
supports it too).
A pair of patches from Sam Lang fix the atomicity of open+create
operations. Several patches from Yan, Zheng fix various mds/client
issues that turned up during multi-mds torture tests.
A final set of patches expose file layouts via virtual xattrs, and
allow the policies to be set on directories via xattrs as well
(avoiding the awkward ioctl interface and providing a consistent
interface for both kernel mount and ceph-fuse users)."
* 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/sage/ceph-client: (143 commits)
libceph: add support for HASHPSPOOL pool flag
libceph: update osd request/reply encoding
libceph: calculate placement based on the internal data types
ceph: update support for PGID64, PGPOOL3, OSDENC protocol features
ceph: update "ceph_features.h"
libceph: decode into cpu-native ceph_pg type
libceph: rename ceph_pg -> ceph_pg_v1
rbd: pass length, not op for osd completions
rbd: move rbd_osd_trivial_callback()
libceph: use a do..while loop in con_work()
libceph: use a flag to indicate a fault has occurred
libceph: separate non-locked fault handling
libceph: encapsulate connection backoff
libceph: eliminate sparse warnings
ceph: eliminate sparse warnings in fs code
rbd: eliminate sparse warnings
libceph: define connection flag helpers
rbd: normalize dout() calls
rbd: barriers are hard
rbd: ignore zero-length requests
...
Before printing kuid and kgids values convert them into
the initial user namespace.
Cc: Sage Weil <sage@inktank.com>
Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
- Make the uid and gid arguments of send_cap_msg() used to compose
ceph_mds_caps messages of type kuid_t and kgid_t.
- Pass inode->i_uid and inode->i_gid in __send_cap to send_cap_msg()
through variables of type kuid_t and kgid_t.
- Modify struct ceph_cap_snap to store uids and gids in types kuid_t
and kgid_t. This allows capturing inode->i_uid and inode->i_gid in
ceph_queue_cap_snap() without loss and pssing them to
__ceph_flush_snaps() where they are removed from struct
ceph_cap_snap and passed to send_cap_msg().
- In handle_cap_grant translate uid and gids in the initial user
namespace stored in struct ceph_mds_cap into kuids and kgids
before setting inode->i_uid and inode->i_gid.
Cc: Sage Weil <sage@inktank.com>
Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
The MDS may have incorrect wanted caps after importing caps. So the
client should check the value mds has and send cap update if necessary.
Signed-off-by: Yan, Zheng <zheng.z.yan@intel.com>
Reviewed-by: Sage Weil <sage@inktank.com>
When client wants to release an imported cap, it's possible there
is no reserved cap_release message in corresponding mds session.
so __queue_cap_release causes kernel panic.
Signed-off-by: Yan, Zheng <zheng.z.yan@intel.com>
Reviewed-by: Sage Weil <sage@inktank.com>
Allow revoking duplicated caps issued by non-auth MDS if these caps
are also issued by auth MDS.
Signed-off-by: Yan, Zheng <zheng.z.yan@intel.com>
Reviewed-by: Sage Weil <sage@inktank.com>
If client sends cap message that requests new max size during
exporting caps, the exporting MDS will drop the message quietly.
So the client may wait for the reply that updates the max size
forever. call handle_cap_grant() for cap import message can
avoid this issue.
Signed-off-by: Yan, Zheng <zheng.z.yan@intel.com>
Signed-off-by: Sage Weil <sage@inktank.com>
Add dirty inode to cap_dirty_migrating list instead, this can avoid
ceph_flush_dirty_caps() entering infinite loop.
Signed-off-by: Yan, Zheng <zheng.z.yan@intel.com>
Signed-off-by: Sage Weil <sage@inktank.com>
The cap from non-auth mds doesn't have a meaningful max_size value.
Signed-off-by: Yan, Zheng <zheng.z.yan@intel.com>
Signed-off-by: Sage Weil <sage@inktank.com>
Convert cpu_to_le32(le32_to_cpu(E1) + E2) to use le32_add_cpu().
dpatch engine is used to auto generate this patch.
(https://github.com/weiyj/dpatch)
Signed-off-by: Wei Yongjun <yongjun_wei@trendmicro.com.cn>
Signed-off-by: Sage Weil <sage@inktank.com>
* 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/sage/ceph-client:
rbd: fix safety of rbd_put_client()
rbd: fix a memory leak in rbd_get_client()
ceph: create a new session lock to avoid lock inversion
ceph: fix length validation in parse_reply_info()
ceph: initialize client debugfs outside of monc->mutex
ceph: change "ceph.layout" xattr to be "ceph.file.layout"
Lockdep was reporting a possible circular lock dependency in
dentry_lease_is_valid(). That function needs to sample the
session's s_cap_gen and and s_cap_ttl fields coherently, but needs
to do so while holding a dentry lock. The s_cap_lock field was
being used to protect the two fields, but that can't be taken while
holding a lock on a dentry within the session.
In most cases, the s_cap_gen and s_cap_ttl fields only get operated
on separately. But in three cases they need to be updated together.
Implement a new lock to protect the spots updating both fields
atomically is required.
Signed-off-by: Alex Elder <elder@dreamhost.com>
Reviewed-by: Sage Weil <sage@newdream.net>
We have been using i_lock to protect all kinds of data structures in the
ceph_inode_info struct, including lists of inodes that we need to iterate
over while avoiding races with inode destruction. That requires grabbing
a reference to the inode with the list lock protected, but igrab() now
takes i_lock to check the inode flags.
Changing the list lock ordering would be a painful process.
However, using a ceph-specific i_ceph_lock in the ceph inode instead of
i_lock is a simple mechanical change and avoids the ordering constraints
imposed by igrab().
Reported-by: Amon Ott <a.ott@m-privacy.de>
Signed-off-by: Sage Weil <sage@newdream.net>
We used to use a flag on the directory inode to track whether the dcache
contents for a directory were a complete cached copy. Switch to a dentry
flag CEPH_D_COMPLETE that is safely updated by ->d_prune().
Signed-off-by: Sage Weil <sage@newdream.net>
Replace remaining direct i_nlink updates with a new set_nlink()
updater function.
Signed-off-by: Miklos Szeredi <mszeredi@suse.cz>
Tested-by: Toshiyuki Okajima <toshi.okajima@jp.fujitsu.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
The pool allocation failures are masked by the pool; there is no need to
spam the console about them. (That's the whole point of having the pool
in the first place.)
Mark msg allocations whose failure is safely handled as such.
Signed-off-by: Sage Weil <sage@newdream.net>
Btrfs needs to be able to control how filemap_write_and_wait_range() is called
in fsync to make it less of a painful operation, so push down taking i_mutex and
the calling of filemap_write_and_wait() down into the ->fsync() handlers. Some
file systems can drop taking the i_mutex altogether it seems, like ext3 and
ocfs2. For correctness sake I just pushed everything down in all cases to make
sure that we keep the current behavior the same for everybody, and then each
individual fs maintainer can make up their mind about what to do from there.
Thanks,
Acked-by: Jan Kara <jack@suse.cz>
Signed-off-by: Josef Bacik <josef@redhat.com>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
We should use ihold whenever we already have a stable inode ref, even
when we aren't holding i_lock. This avoids adding new and unnecessary
locking dependencies.
Signed-off-by: Sage Weil <sage@newdream.net>
In e9964c10 we change cap flushing to do a delicate dance because some
inodes on the cap_dirty list could be in a migrating state (got EXPORT but
not IMPORT) in which we couldn't actually flush and move from
dirty->flushing, breaking the while (!empty) { process first } loop
structure. It worked for a single sync thread, but was not reentrant and
triggered infinite loops when multiple syncers came along.
Instead, move inodes with dirty to a separate cap_dirty_migrating list
when in the limbo export-but-no-import state, allowing us to go back to
the simple loop structure (which was reentrant). This is cleaner and more
robust.
Audited the cap_dirty users and this looks fine:
list_empty(&ci->i_dirty_item) is still a reliable indicator of whether we
have dirty caps (which list we're on is irrelevant) and list_del_init()
calls still do the right thing.
Signed-off-by: Sage Weil <sage@newdream.net>
If we grab new_cap, retake the lock, and find we already have a cap now
for the given mds, release new_cap.
Signed-off-by: Sage Weil <sage@newdream.net>
We increments i_wrbuffer_ref when taking the Fb cap. This breaks
the dirty page accounting and causes looping in
__ceph_do_pending_vmtruncate, and ceph client hangs.
This bug can be reproduced occasionally by running blogbench.
Add a new field i_wb_ref to inode and dedicate it to Fb reference
counting.
Signed-off-by: Henry C Chang <henry.cy.chang@gmail.com>
Signed-off-by: Sage Weil <sage@newdream.net>
The __mark_dirty_inode helper now takes i_lock as of 250df6ed. Fix the
one ceph callers that held i_lock (__ceph_mark_dirty_caps) to return the
flags value so that the callers can do it outside of i_lock.
Signed-off-by: Sage Weil <sage@newdream.net>
The NODELAY flag avoids the heuristics that delay cap (issued/wanted)
release. There's no reason for that after we import a cap, and it kills
whatever benefit we get from those delays.
Signed-off-by: Sage Weil <sage@newdream.net>
If we are mid-flush and a cap is migrated to another node, we need to
resend the cap flush message to the new MDS, and do so with the original
flush_seq to avoid leaking across a sync boundary. Previously we didn't
redo the flush (we only flushed newly dirty data), which would cause a
later sync to hang forever.
Signed-off-by: Sage Weil <sage@newdream.net>
The int flushing is global and not clear on each iteration of the loop,
which can cause a second flush of caps to any MDSs with ids greater than
the auth.
Signed-off-by: Sage Weil <sage@newdream.net>
We used to use rdcache_gen to indicate whether we "might" have cached
pages. Now we just look at the mapping to determine that. However, some
old behavior remains from that transition.
First, rdcache_gen == 0 no longer means we have no pages. That can happen
at any time (presumably when we carry FILE_CACHE). We should not reset it
to zero, and we should not check that it is zero.
That means that the only purpose for rdcache_revoking is to resolve races
between new issues of FILE_CACHE and an async invalidate. If they are
equal, we should invalidate. On success, we decrement rdcache_revoking,
so that it is no longer equal to rdcache_gen. Similarly, if we success
in doing a sync invalidate, set revoking = gen - 1. (This is a small
optimization to avoid doing unnecessary invalidate work and does not
affect correctness.)
Signed-off-by: Sage Weil <sage@newdream.net>
If the auth cap migrates to another MDS, clear requested_max_size so that
we resend any pending max_size increase requests. This fixes potential
hangs on writes that extend a file and race with an cap migration between
MDSs.
Signed-off-by: Sage Weil <sage@newdream.net>
This reverts commit d91f2438d8.
The intent of issue_seq is to distinguish between mds->client messages that
(re)create the cap and those that do not, which means we should _only_ be
updating that value in the create paths. By updating it in handle_cap_grant,
we reset it to zero, which then breaks release.
The larger question is what workload/problem made me think it should be
updated here...
Signed-off-by: Sage Weil <sage@newdream.net>
The i_rdcache_gen value only implies we MAY have cached pages; actually
check the mapping to see if it's worth bothering with an invalidate.
Signed-off-by: Sage Weil <sage@newdream.net>
This factors out protocol and low-level storage parts of ceph into a
separate libceph module living in net/ceph and include/linux/ceph. This
is mostly a matter of moving files around. However, a few key pieces
of the interface change as well:
- ceph_client becomes ceph_fs_client and ceph_client, where the latter
captures the mon and osd clients, and the fs_client gets the mds client
and file system specific pieces.
- Mount option parsing and debugfs setup is correspondingly broken into
two pieces.
- The mon client gets a generic handler callback for otherwise unknown
messages (mds map, in this case).
- The basic supported/required feature bits can be expanded (and are by
ceph_fs_client).
No functional change, aside from some subtle error handling cases that got
cleaned up in the refactoring process.
Signed-off-by: Sage Weil <sage@newdream.net>
Yan, Zheng
majianpeng
Linus Torvalds
Eric W. Biederman
Wei Yongjun
Alex Elder
Al Viro
Sage Weil
Miklos Szeredi
Josef Bacik
Henry C Chang
Lucas De Marchi
Yehuda Sadeh