If MDS sorts dentries in dirfrag in hash order, we use hash value to
compose dentry offset. dentry offset is:

  (0xff << 52) | ((24 bits hash) << 28) |
  (the nth entry hash hash collision)

This offset is stable across directory fragmentation. This alos means
there is no need to reset readdir offset if directory get fragmented
in the middle of readdir.
Signed-off-by: NYan, Zheng <zyan@redhat.com>

Forward seek within same frag does not update fi->last_name, it will
not affect contents of later readdir reply. So there is no need to
forbid marking directory complete
Signed-off-by: NYan, Zheng <zyan@redhat.com>

This is preparation for using hash value as dentry 'offset'
Signed-off-by: NYan, Zheng <zyan@redhat.com>

Set a flag in readdir request, which indicates that client interprets
'end/complete' as bit flags. So that mds can reply additional flags in
readdir reply.
Signed-off-by: NYan, Zheng <zyan@redhat.com>

This avoids defining multiple arrays for entries in readdir reply
Signed-off-by: NYan, Zheng <zyan@redhat.com>

don't distinguish leftmost frag from other frags. always use 2 as
first entry's offset.
Signed-off-by: NYan, Zheng <zyan@redhat.com>

we never add snapdir and the hidden .ceph dir into readdir cache
Signed-off-by: NYan, Zheng <zyan@redhat.com>

use binary search to find cache index that corresponds to readdir
postion.
Signed-off-by: NYan, Zheng <zyan@redhat.com>

Setxattr with NULL value and XATTR_REPLACE flag should be equivalent
to removexattr. But current MDS does not support deleting vxattrs through
MDS_OP_SETXATTR request. The workaround is sending MDS_OP_RMXATTR request
if setxattr actually removs xattr.
Signed-off-by: NYan, Zheng <zyan@redhat.com>

Signed-off-by: NYan, Zheng <zyan@redhat.com>

symlink target is useless for debug and can be very long. It's annoying
to show it in debugfs/mdsc.
Signed-off-by: NYan, Zheng <zyan@redhat.com>

truncate_pagecache() may decrease inode's reference. This can cause
deadlock if inode's last reference is dropped and iput_final() wants
to evict the inode. (evict() calls inode_wait_for_writeback(), which
waits for ceph_writepages_start() to return).

The fix is use work thead to truncate dirty pages. Also add 'forced
umount' check to ceph_update_writeable_page(), which prevents new
pages getting dirty.
Signed-off-by: NYan, Zheng <zyan@redhat.com>

When mds session gets killed, read/write operation may hang.
Client waits for Frw caps, but mds does not know what caps client
wants. To recover this, client sends an open request to mds. The
request will tell mds what caps client wants.
Signed-off-by: NYan, Zheng <zyan@redhat.com>

Signed-off-by: NYan, Zheng <zyan@redhat.com>

To access non-default filesystem, we just need to subscribe to
mdsmap.<MDS_NAMESPACE_ID> and add a new mount option for mds
namespace id.
Signed-off-by: NYan, Zheng <zyan@redhat.com>
[idryomov@gmail.com: switch to a new libceph API]
Signed-off-by: NIlya Dryomov <idryomov@gmail.com>

Signed-off-by: NIlya Dryomov <idryomov@gmail.com>

... with a wrapper around maybe_request_map() - no need for two
osdmap-specific functions.
Signed-off-by: NIlya Dryomov <idryomov@gmail.com>

There is now about a dozen CEPH_OSDMAP_* flags.  This is a debugging
interface, so just dump in hex instead of spelling each flag out.
Signed-off-by: NIlya Dryomov <idryomov@gmail.com>

This adds the "map check" infrastructure for sending osdmap version
checks on CALC_TARGET_POOL_DNE and completing in-flight requests with
-ENOENT if the target pool doesn't exist or has just been deleted.
Signed-off-by: NIlya Dryomov <idryomov@gmail.com>

For map check, we are going to need to send CEPH_MSG_MON_GET_VERSION
messages asynchronously and get a callback on completion.  Refactor MON
client to allow firing off generic requests asynchronously and add an
async variant of ceph_monc_get_version().  ceph_monc_do_statfs() is
switched over and remains sync.
Signed-off-by: NIlya Dryomov <idryomov@gmail.com>

Implement ceph_osdc_watch_check() to be able to check on status of
watch.  Note that the time it takes for a watch/notify event to get
delivered through the notify_wq is taken into account.
Signed-off-by: NIlya Dryomov <idryomov@gmail.com>

Implement ceph_osdc_notify() for sending notifies.

Due to the fact that the current messenger can't do read-in into
pagelists (it can only do write-out from them), I had to go with a page
vector for a NOTIFY_COMPLETE payload, for now.
Signed-off-by: NIlya Dryomov <idryomov@gmail.com>

This adds support and switches rbd to a new, more reliable version of
watch/notify protocol.  As with the OSD client update, this is mostly
about getting the right structures linked into the right places so that
reconnects are properly sent when needed.  watch/notify v2 also
requires sending regular pings to the OSDs - send_linger_ping().

A major change from the old watch/notify implementation is the
introduction of ceph_osd_linger_request - linger requests no longer
piggy back on ceph_osd_request.  ceph_osd_event has been merged into
ceph_osd_linger_request.

All the details are now hidden within libceph, the interface consists
of a simple pair of watch/unwatch functions and ceph_osdc_notify_ack().
ceph_osdc_watch() does return ceph_osd_linger_request, but only to keep
the lifetime management simple.

ceph_osdc_notify_ack() accepts an optional data payload, which is
relayed back to the notifier.

Portions of this patch are loosely based on work by Douglas Fuller
<dfuller@redhat.com> and Mike Christie <michaelc@cs.wisc.edu>.
Signed-off-by: NIlya Dryomov <idryomov@gmail.com>

Introduce __rbd_dev_header_unwatch_sync(), which doesn't flush notify
callbacks.  This is for the new rados_watcherrcb_t, which would be
called from a notify callback.
Signed-off-by: NIlya Dryomov <idryomov@gmail.com>

The unwatch timeout is currently implemented in rbd.  With
watch/unwatch code moving into libceph, we are going to need
a ceph_osdc_wait_request() variant with a timeout.
Signed-off-by: NIlya Dryomov <idryomov@gmail.com>

These are going to be used by request_reinit() code.
Signed-off-by: NIlya Dryomov <idryomov@gmail.com>

This is a major sync up, up to ~Jewel.  The highlights are:

- per-session request trees (vs a global per-client tree)
- per-session locking (vs a global per-client rwlock)
- homeless OSD session
- no ad-hoc global per-client lists
- support for pool quotas
- foundation for watch/notify v2 support
- foundation for map check (pool deletion detection) support

The switchover is incomplete: lingering requests can be setup and
teared down but aren't ever reestablished.  This functionality is
restored with the introduction of the new lingering infrastructure
(ceph_osd_linger_request, linger_work, etc) in a later commit.
Signed-off-by: NIlya Dryomov <idryomov@gmail.com>

OSD client is getting moved from the big per-client lock to a set of
per-session locks.  The big rwlock would only be held for read most of
the time, so a global osdc->osd_lru needs additional protection.
Signed-off-by: NIlya Dryomov <idryomov@gmail.com>

create_osd() is called way too deep in the stack to be able to error
out in a sane way; a failing create_osd() just messes everything up.
The current req_notarget list solution is broken - the list is never
traversed as it's not entirely clear when to do it, I guess.

If we were to start traversing it at regular intervals and retrying
each request, we wouldn't be far off from what __GFP_NOFAIL is doing,
so allocate OSD sessions with __GFP_NOFAIL, at least until we come up
with a better fix.
Signed-off-by: NIlya Dryomov <idryomov@gmail.com>

These are going to be used by homeless OSD sessions code.
Signed-off-by: NIlya Dryomov <idryomov@gmail.com>

Separate osdmap handling from decoding and iterating over a bag of maps
in a fresh MOSDMap message.  This sets up the scene for the updated OSD
client.

Of particular importance here is the addition of pi->was_full, which
can be used to answer "did this pool go full -> not-full in this map?".
This is the key bit for supporting pool quotas.

We won't be able to downgrade map_sem for much longer, so drop
downgrade_write().
Signed-off-by: NIlya Dryomov <idryomov@gmail.com>

This leads to a simpler osdmap handling code, particularly when dealing
with pi->was_full, which is introduced in a later commit.
Signed-off-by: NIlya Dryomov <idryomov@gmail.com>

Both homeless OSD sessions and watch/notify v2, introduced in later
commits, require periodic ticks which don't depend on ->num_requests.
Schedule the initial tick from ceph_osdc_init() and reschedule from
handle_timeout() unconditionally.
Signed-off-by: NIlya Dryomov <idryomov@gmail.com>

ceph_osdc_stop() isn't called if ceph_osdc_init() fails, so we end up
with handle_osds_timeout() running on invalid memory if any one of the
allocations fails.  Call schedule_delayed_work() after everything is
setup, just before returning.
Signed-off-by: NIlya Dryomov <idryomov@gmail.com>

If you specify ACK | ONDISK and set ->r_unsafe_callback, both
->r_callback and ->r_unsafe_callback(true) are called on ack.  This is
very confusing.  Redo this so that only one of them is called:

    ->r_unsafe_callback(true), on ack
    ->r_unsafe_callback(false), on commit

or

    ->r_callback, on ack|commit

Decode everything in decode_MOSDOpReply() to reduce clutter.
Signed-off-by: NIlya Dryomov <idryomov@gmail.com>

finish_read(), its only user, uses it to get to hdr.data_len, which is
what ->r_result is set to on success.  This gains us the ability to
safely call callbacks from contexts other than reply, e.g. map check.
Signed-off-by: NIlya Dryomov <idryomov@gmail.com>

The crux of this is getting rid of ceph_osdc_build_request(), so that
MOSDOp can be encoded not before but after calc_target() calculates the
actual target.  Encoding now happens within ceph_osdc_start_request().

Also nuked is the accompanying bunch of pointers into the encoded
buffer that was used to update fields on each send - instead, the
entire front is re-encoded.  If we want to support target->name_len !=
base->name_len in the future, there is no other way, because oid is
surrounded by other fields in the encoded buffer.

Encoding OSD ops and adding data items to the request message were
mixed together in osd_req_encode_op().  While we want to re-encode OSD
ops, we don't want to add duplicate data items to the message when
resending, so all call to ceph_osdc_msg_data_add() are factored out
into a new setup_request_data().
Signed-off-by: NIlya Dryomov <idryomov@gmail.com>

Replace __calc_request_pg() and most of __map_request() with
calc_target() and start using req->r_t.

ceph_osdc_build_request() however still encodes base_oid, because it's
called before calc_target() is and target_oid is empty at that point in
time; a printf in osdc_show() also shows base_oid.  This is fixed in
"libceph: switch to calc_target(), part 2".
Signed-off-by: NIlya Dryomov <idryomov@gmail.com>

Introduce ceph_osd_request_target, containing all mapping-related
fields of ceph_osd_request and calc_target() for calculating mappings
and populating it.
Signed-off-by: NIlya Dryomov <idryomov@gmail.com>

Add and decode pi->min_size and pi->last_force_request_resend.  These
are going to be used by calc_target().
Signed-off-by: NIlya Dryomov <idryomov@gmail.com>