We could support zero-length tuples, but contiguous appender doesn't
like zero-length writes, so I've ruled them out for now.
Signed-off-by: Adam C. Emerson <aemerson@redhat.com>
- Option nocapture is not really required so no problem
with the fact that FreeBSD env does not work for params
for nosetests
Signed-off-by: Willem Jan Withagen <wjw@digiware.nl>
Currently if user perform bench-write with io-size > 4G
then its crashing because currently during memory allocation
bufferptr taking size of buffer as a unsigned and io-size > 4G
will overflow with unsigned. so during memset operation it will
try to set io_size size of memory area pointed by bufferptr,
(bufferptr area is: (4G - io_size)), so it will cause
segmentation fault.
Fix is to return error if io-size >= 4G
Fixes: http://tracker.ceph.com/issues/18422
Reported-by: Jason Dillaman <dillaman@redhat.com>
Signed-off-by: Gaurav Kumar Garg <garg.gaurav52@gmail.com>
When the read or flush command fails, we need to
release segs. Or we can abort it like what we do
for write.
Signed-off-by: optimistyzy <optimistyzy@gmail.com>
prior to this change, libcommon is a convenient library which gets
linked into librados, librbd and libcephfs and all ceph executables.
this incurs some problems:
- double dose of libcommon in memory space and HDD: waste of memory
and disk space.
- if an application links to two libraries including libcommon at the
same time. take librados and libcephfs as an example, they could
interfere with each other by changing the other guy's status.
after this change, libcommon is tuned into a shared library and
renamed to libceph-common. it will be installed into $prefix/lib/ceph,
and packaged in librados2.
ceph.spec.in,debian/librados2.install: package libceph-common in
librados2.
CMakeLists.txt:
- do not link against libboost-* if not necessary.
- s/common/ceph-common/g
- install libceph-common into $prefix/lib/ceph
- set rpath to $prefix/lib/ceph
- link against ceph-common if an executable needs access to non public
symbols in ceph.
Signed-off-by: Kefu Chai <kchai@redhat.com>
Struct CapSnap holds a reference to its parent inode. So erasing
struct CapSnap from Inode::cap_snaps may drop inode's last reference.
The inode gets freed in the middle of erasing struct CapSnap
Fixes: http://tracker.ceph.com/issues/18460
Signed-off-by: Yan, Zheng <zyan@redhat.com>
{rmdir/rename}_prepare_witness() use full path to compose slave
requests. But they do not lock all dentries in the path. So someone
else changes the unlocked dentry and causes path travsese of slave
request to fail.
Signed-off-by: Yan, Zheng <zyan@redhat.com>
Subtree map can be very large when thrash exports is enabled.
Printing subtree map causes size log to increase quickly.
Signed-off-by: Yan, Zheng <zyan@redhat.com>
When doing subtree migration or cross subtree rename, we rdlock
fragtreelock of subtree boundaries' inodes. rdlock fragtreelock
has already journaled dirfragtree updates. I can't see why we
need to journal it again when merging subtrees.
The resaon I want to remove this code is that it can journal
unconnected dirfrag in the Migrator::export_reverse() case.
Signed-off-by: Yan, Zheng <zyan@redhat.com>
Server::journal_close_session() modifies session->info before
corresponding ESession get journaled. It's possible the updated
session gets saved to sessionmap before Server::_session_logged()
increases sessionmap's version. This confuses the journal replay
code.
Signed-off-by: Yan, Zheng <zyan@redhat.com>
When a variable is not being observed we currently mark it
"unchangable". This can be misleading so try something hopefully a
little more informative.
Fixes: http://tracker.ceph.com/issues/18424
Signed-off-by: Brad Hubbard <bhubbard@redhat.com>
All clientreplay requests' finish contexts should be executed
before MDCache::export_remaining_imported_caps(). Otherwise
MDCache::try_reconnect_cap() may fail to reconnect client caps.
Fixes: http://tracker.ceph.com/issues/18461
Signed-off-by: Yan, Zheng <zyan@redhat.com>
prior to this change:
a peon may forward the pgstats to leader, and record it locally, but leader will
check if osd has the latest map before process, if not, will use a route op to
indicate peon to send it, then poen will delete routed op when fininaly send
out which make peon cannot send pgstatack when leader has processed the
pgstat update. so osd will always track it util reach a threshold block pgstats
sending, at worst, reopen mon session.
also, both leader and peon will send out the osdmap message to the osd.
after this change:
only the peon will send out the osdmap message. and the pgstatack message
will be routed to the osd as expected. so the osd will not keep track of the
"acked" pg stats in its queue forever before times out.
Fixes: http://tracker.ceph.com/issues/18458
Signed-off-by: Mingxin Liu <mingxin@xsky.com>
