AsyncConnection: Don't use unsafe feature as message encode feature

Fix #12908
Signed-off-by: Haomai Wang <haomaiwang@gmail.com>

ceph-osd-prestart.sh: fix osd data dir ownership check

Signed-off-by: Sage Weil <sage@redhat.com>

Merge pull request #5749 from liewegas/wip-async-ms

msg/async: log tx/rx at level 1

Reviewed-by: Haomai Wang <haomaiwang@gmail.com>

vstart.sh: enable all experimental features for vstart

Signed-off-by: Sage Weil <sage@redhat.com>

ms/async: log message tx/rx at level 1

Signed-off-by: Sage Weil <sage@redhat.com>

Merge pull request #5722 from cxwshawn/vs-fix

vstart.sh: add --mon_num --osd_num --mds_num --rgw_port option

Reviewed-by: Loic Dachary <ldachary@redhat.com>

Merge pull request #5693 from tchaikov/wip-12730

common/SubProcess: silence compiler warnings

Reviewed-by: Loic Dachary <ldachary@redhat.com>

Merge pull request #5643 from dreamhost/wip-make-check-makeopt

make-check: support MAKEOPTS overrides.

Reviewed-by: Loic Dachary <ldachary@redhat.com>

Merge pull request #5299 from hjwsm1989/pgmonitor-const

  mon: added const to dump_* functions in PGMonitor

Reviewed-by: Loic Dachary <ldachary@redhat.com>

Merge pull request #5156 from rubenk/fix-indentation

Fix indentation

Reviewed-by: Loic Dachary <ldachary@redhat.com>

Merge pull request #5275 from tchaikov/wip-12287

pybind/ceph_argparse: do not choke on non-ascii prefix

Reviewed-by: Loic Dachary <ldachary@redhat.com>

Merge pull request #5702 from Sandy4999/wip-doc-sandy

doc:radosgw: correct typos of the command removing a subuser

Reviewed-by: Abhishek Lekshmanan <abhishek.lekshmanan@ril.com>

Merge pull request #5747 from ceph/wip-user

fix ceph-disk

Reviewed-by: Loic Dachary <ldachary@redhat.com>

Merge pull request #5742 from dachary/wip-user

tests: ceph-disk: dmcrypt simplification

Reviewed-by: Sage Weil <sage@redhat.com>

Merge pull request #5746 from ceph/wip-fix-doc-build

doc: fix the code-block in ruby.rst

Reviewed-by: Loic Dachary <ldachary@redhat.com>

doc: fix the code-block in ruby.rst

* and add the link to library homepage in the section titles

Signed-off-by: Kefu Chai <kchai@redhat.com>

Merge pull request #5578 from ceph/wip-newstore

osd: newstore (experimental)

ceph_test_keyvaluedb: add simple commit latency benchmark

Signed-off-by: Sage Weil <sage@redhat.com>

os/newstore: update todo

Signed-off-by: Sage Weil <sage@redhat.com>

do_autogen.sh: build static rocksdb by default

Signed-off-by: Sage Weil <sage@redhat.com>

rocksdb: update alt dist rule

Signed-off-by: Sage Weil <sage@redhat.com>

ceph_test_objectstore: make OMapIterator test work with FileStore

Signed-off-by: Sage Weil <sage@redhat.com>

ceph_test_objectstore: enable newstore tests

Signed-off-by: Sage Weil <sage@redhat.com>

rocksdb: update to 3.11.2

Signed-off-by: Sage Weil <sage@redhat.com>

os/RocksDBStore: make other rmkey match

No need for Slice() here; it can take a string.

Signed-off-by: Sage Weil <sage@redhat.com>

os/RocksDBStore: fix rmkey()

This took way too long to debug!

Signed-off-by: Sage Weil <sage@redhat.com>

ceph_test_keyvaluedb: some simple KeyValueDB unit tests

Signed-off-by: Sage Weil <sage@redhat.com>

os/newstore: fix end bound on collection_list

Signed-off-by: Sage Weil <sage@redhat.com>

os/newstore: flush object before doing omap reads

Signed-off-by: Sage Weil <sage@redhat.com>

os/newstore: add 'newstore backend options' to pass options to e.g. rocksdb

Signed-off-by: Sage Weil <sage@redhat.com>

os/newstore: change escaping chars

# is lowest besides space and !, except for " (which would be too
confusing).

Signed-off-by: Sage Weil <sage@redhat.com>

os/newstore: trim overlay when zeroing extent

Signed-off-by: Sage Weil <sage@redhat.com>

os/newstore: tolerate null pnext to collection_list()

Signed-off-by: Sage Weil <sage@redhat.com>

os/newstore: fix collection range for temp objects

Signed-off-by: Sage Weil <sage@redhat.com>

os/newstore: Implement fiemap

For simplicity we ignore holes inside an fragment now.

Signed-off-by: Xiaoxi Chen <xiaoxi.chen@intel.com>

os/newstore: make sync/async submit_transaction optional

It seems doing this synchronously may be better for SSDs?

Signed-off-by: Sage Weil <sage@redhat.com>

os/newstore: renamed TransContext::fds -> sync_items

Signed-off-by: Sage Weil <sage@redhat.com>

os/newstore: queue kv transactions in kv_sync_thread

It appears that db->submit_transaction() will block if there is a sync
commit that is in progress instead of simply queueing the new txn for
later.  To work around this, submit these to the backend in the
kv_sync_thread prior to the synchronous submit_transaction_sync().

Signed-off-by: Sage Weil <sage@redhat.com>

os/newstore: process multiple aio completions at a time

This isn't affecting things for a slow disk, but it will matter for faster
backends.

Signed-off-by: Sage Weil <sage@redhat.com>

os/newstore: clean up kv commit debug output

Signed-off-by: Sage Weil <sage@redhat.com>

os/newstore: only ftruncate if i_size is incorrect

Even a no-op ftruncate can block in the kernel.  Prior to this change I
could frequently see ftruncate wait for an aio completion on the same
file.

Signed-off-by: Sage Weil <sage@redhat.com>

Revert "os/newstore: avoid sync append for small ios"

This reverts commit 69baab2f7eaca7688ce1d45802a82fc3539cd906.

This is slower.  :(

os/newstore: avoid sync append for small ios

An append is expensive in terms of latency (write, fdatasync, kv commit),
while a wal write is just the kv commit and the write and fdatasync are
async.  For small IOs doing the wal may improve performance.

Signed-off-by: Sage Weil <sage@redhat.com>

rocksdb: fallocate_with_keep_size = false

This improves my 4k random writes on hdd by about 25%.

Signed-off-by: Sage Weil <sage@redhat.com>

Revert "os/NewStore: data_map shouldn't be empty when writing all overlays"

This reverts commit 0d9cce462fec61f754ddcd17cf9a3cf69581d7c5.

We may want to write an overlay if hte object is new and the write is small to defer the cost
of the fsync.

os/NewStore: delay the read of all the overlays until wal applying

The read of all the overlays can be delayed until applying the wal. If
we are doing async wal apply, this can reduce write op latency by
eliminating unnecessary reads in the write code path.

Signed-off-by: Zhiqiang Wang <zhiqiang.wang@intel.com>

os/newstore: fix deadlock when newstore_sync_transaction=true

There is a deadlock issue in Newstore when newstore_sync_transaction = true.
With sync_transaction to true, the txc state machine will go all the way down
from STATE_IO_DONE to STATE_FINISHING in the same thread, while holding the osr->qlock().
The deadlock is caused in _txc_finish and _osr_reap_done, when trying to
lock osr->qlock again.

Since the _txc_finish can be called with(in sync transaction mode) or without
(in async transaction mode) holding the qlock, so fix this by setting the qlock
to PTHREAD_MUTEX_RECURSIVE, thus we can recursive acquire the qlock.

Signed-off-by: Xiaoxi Chen <xiaoxi.chen@intel.com>

os/NewStore: fix the append of the later overlays when doing combination

The data of the later contiguous overlays should be claim_append to
'op->data', instead of 'bl'.

Signed-off-by: Zhiqiang Wang <zhiqiang.wang@intel.com>

os/Newstore: flush_commit return true on STATE_KV_DONE

There is a racing condition here, if the flush_commit() call
happened after _txc_finish_kv and before next state, the context
was pushed to on_commits but no one will handle the context since
we already pass _txc_finish_kv. This bug can be easily reproduce
by putting a sleep(5) after _txc_finish_kv, and trigger the bug by
ceph-osd -i 0 --mkfs.

Fix this bug by return true directly when state >= STATE_KV_DONE(instead
of > in previous code). We already persist the data in STATE_KV_DONE so
it's safe for us to do this.

Signed-off-by: Xiaoxi Chen <xiaoxi.chen@intel.com>

os/NewStore: avoid dup the data of the overlays in the WAL

When writing all the overlays, there is no need to dup the data in WAL.
Instead, we can reference the overlays in the WAL, and remove these
overlays after commiting them to the fs. When replaying, we can get
these data from the referenced overlays. Doing this way, we can save a
write and a deletion for each of the overlay data in the db.

Signed-off-by: Zhiqiang Wang <zhiqiang.wang@intel.com>

os/newstore: fix multiple aio case

Signed-off-by: Sage Weil <sage@redhat.com>

os/newstore: more conservative default for aio queue depth

There appears to be a kernel aio bug when the queue depth is small.

Signed-off-by: Sage Weil <sage@redhat.com>

os/newstore:close fd after writting with O_DIRECT

fix bug in 2b4c60e0a521ad10b94bbc82865b49f2d28c2ac9

Signed-off-by: Xiaoxi Chen <xiaoxi.chen@intel.com>

os/NewStore: need to increase the wal op length when combining overlays

Need to add the length of the combining overlays to the length of the
wal op.

Signed-off-by: Zhiqiang Wang <zhiqiang.wang@intel.com>

os/Newstore:Fix collection_list_range

We need to rule out hobject_t::max before calling get_object_key
(in which will call get_filestore_key_u32 and get an assert failure)

Signed-off-by: Xiaoxi Chen <xiaoxi.chen@intel.com>

os/newstore: fix race in _txc_aio_submit

We cannot rely on the iterator pointers being valid after we submit the
aio because we are racing with the completion.  Make our loop decision
before submitting and avoid dereferencing txc after that point.

Signed-off-by: Sage Weil <sage@redhat.com>

os/newstore : Do not need to call fdatasync if using direct.

skip ::fdatasync if in direct mode.

Signed-off-by: Xiaoxi Chen <xiaoxi.chen@intel.com>

osd/NewStore: fix for skipping the overlay in _do_overlay_trim

When the offset of the write starts at the end of the overlay, that is,
p->first + p->second.length == offset, the overlay could be skipped as
well.

Signed-off-by: Zhiqiang Wang <zhiqiang.wang@intel.com>

os/NewStore: combine contiguous overlays when writing all the overlays

Combine contiguous overlay writes to reduce the numbers of WAL writes
and fs writes.

Signed-off-by: Zhiqiang Wang <zhiqiang.wang@intel.com>

os/Nestore: batch cleanup

batch cleanup wal.

Signed-off-by: Xiaoxi Chen <xiaoxi.chen@intel.com>

os/newstore: fix _txc_aio_submit

The aios may complete before _txc_aio_submit completes.  In fact, the aio
may complete, commit to the kv store, and then queue more wal aio's before
we finish the loop.  Move aios to a separate list to ensure we only submit
them once and do not right another CPU adjusting the list.

Signed-off-by: Sage Weil <sage@redhat.com>

os/newstore: throttle over entire write lifecycle

Take a global throttle when we submit ops and release when they complete.
The first throttles cover the period from submit to commit, while the wal
ones also cover the async post-commit wal work.  The configs are additive
since the wal ones cover both periods; this should make them reasonably
idiot-proof.

Signed-off-by: Sage Weil <sage@redhat.com>

os/NewStore: data_map shouldn't be empty when writing all overlays

This should be an assert instead of creating new data_map.

Signed-off-by: Zhiqiang Wang <zhiqiang.wang@intel.com>

os/NewStore: clear the shared_overlays after writing all the overlays

Signed-off-by: Zhiqiang Wang <zhiqiang.wang@intel.com>

os/NewStore: don't clear overlay in the create/append case of write

Shouldn't clear the overlay in the create/append case of write.
Otherwise, this removes the overlay data and leads to data loss.

Signed-off-by: Zhiqiang Wang <zhiqiang.wang@intel.com>

os/newstore: debug io_submit EAGAIN

Signed-off-by: Sage Weil <sage@redhat.com>

os/newstore: release wal throttle when wal completes, not when queued

If we take the aio path, the io is queued immediately and the resources
are released back to the pool.  Instead release them when wal completes.

Signed-off-by: Sage Weil <sage@redhat.com>

os/newstore: todo

Signed-off-by: Sage Weil <sage@redhat.com>

os/newstore: move toward state-machine

Signed-off-by: Sage Weil <sage@redhat.com>

os/newstore: use aio for wal writes, too

Signed-off-by: Sage Weil <sage@redhat.com>

os/newstore: a few comments about wal

Signed-off-by: Sage Weil <sage@redhat.com>

os/newstore: combined O_DSYNC with O_DIRECT

This avoids the need for an explicit fdatasync when doing O_DIRECT.

Signed-off-by: Sage Weil <sage@redhat.com>

os/newstore: basic aio support

Signed-off-by: Sage Weil <sage@redhat.com>

os/Newstore: add newstore_db_path option

The load of Keyvalue DB is heavy, allow user to put
DB to a seperate(fast) device.

Signed-off-by: Xiaoxi Chen <xiaoxi.chen@intel.com>

os/newstore: throttle wal work

Signed-off-by: Sage Weil <sage@redhat.com>

os/newstore: show # o_direct buffers in debug output

Signed-off-by: Sage Weil <sage@redhat.com>

os/newstore: use a threadpool for applying wal events

Signed-off-by: Sage Weil <sage@redhat.com>

os/newstore: rebuild buffers to be page-aligned for O_DIRECT

Signed-off-by: Sage Weil <sage@redhat.com>

ceph_test_objectstore: fix omap test cleanup

Signed-off-by: Sage Weil <sage@redhat.com>

os/newstore: use fdatasync instead of fsync

On XFS at least, fdatasync is sufficient to make data readable.

Signed-off-by: Sage Weil <sage@redhat.com>

os/newstore: update todo

Signed-off-by: Sage Weil <sage@redhat.com>

os/Newstore: Check onode.omap_head in valid() and next()

The db iter will be set to KeyValueDB::Iterator() if onode.omap_head
not present. In that case if we touch the db iter we will get a segmentation
fault.

Prevent to touch the db iter when onode.omap_head is invalid(equals to 0).

Signed-off-by: Xiaoxi Chen <xiaoxi.chen@intel.com>

Use .str() to output a stringstream.

a nit.

Signed-off-by: Xiaoxi Chen <xiaoxi.chen@intel.com>

os/Newstore: Allow gap in _do_write append mode

We can allow some gap so we only need to ensure
onode.size <= offset.

Signed-off-by: Xiaoxi Chen <xiaoxi.chen@intel.com>

Implement get_omap_iterator

implemented get_omap_iterator

Signed-off-by: Xiaoxi Chen <xiaoxi.chen@intel.com>

os/KeyValueDB: Add raw_key() interface for IteratorImpl

raw_key() is useful to split out the prefix.

Signed-off-by: Xiaoxi Chen <xiaoxi.chen@intel.com>

test/store_test Add get_omap_iterator test cases

omap iterator test cases include:
  iter aganist omap
  lower_bound
  upper_bound

Signed-off-by: Xiaoxi Chen <xiaoxi.chen@intel.com>

os/newstore: drop sync()

Signed-off-by: Sage Weil <sage@redhat.com>

os/newstore: drop sync()

Signed-off-by: Sage Weil <sage@redhat.com>

os/newstore: drop flush

Signed-off-by: Sage Weil <sage@redhat.com>

os/newstore: drop sync_and_flush

Signed-off-by: Sage Weil <sage@redhat.com>

os/newstore: use FS::zero()

Signed-off-by: Sage Weil <sage@redhat.com>

os/newstore: use O_DIRECT is write is page-aligned

Signed-off-by: Sage Weil <sage@redhat.com>

os/newstore: pass flags to _{open,create}_fid

Signed-off-by: Sage Weil <sage@redhat.com>

os/newstore: drop unused FragmentHandle

Signed-off-by: Sage Weil <sage@redhat.com>

os/newstore: do not call completions from kv thread

Reads may call wait_wal() holding user locks, and so we cannot block
progress on WAL completion/flushing by calling callbacks that may take
user locks.

Signed-off-by: Sage Weil <sage@redhat.com>

os/newstore: let wal cleanup kv txn get batched

No need to trigger another sync kv commit here; just let the next KV
commit catch it.

We could possibly do a bit better here by not waking up the kv thread at
all...

Signed-off-by: Sage Weil <sage@redhat.com>

os/newstore: fix off-by-one on overlay_max_length

Signed-off-by: Sage Weil <sage@redhat.com>

os/newstore: use lower_bound for finding overlay extents in map

Signed-off-by: Sage Weil <sage@redhat.com>

os/newstore: use overlay even if it is a new object or append

This avoids the fsync for small writes.

Signed-off-by: Sage Weil <sage@redhat.com>