librgw: dir traversal state GC added

To minimize overhead, traversal events are modeled as a linear
sequence of fh_key, timestamp tuples.  GC operates out-of-line in
LIFO order, and includes a ref cycle.  If the matching directory
object can't be found by handle, it's already gone, and nothing
happens.

To avoid lock contention, scanning of the event sequence is done
in groups of 24 by default, but in groups of sequence/4 if the
event sequence grows larger than 500 events.

Push READDIR events on successful component readdir (bucket or
object listing partial result, as appropriate).

Revise locking, timestamp and update link count in all cases.  Link
counts are approximate (metadata to represent it is not stored).

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

librgw: restore readdir prints (move to dout)

They are useful for checking the exact strings being produced,
and in which sequence (object or common prefix).

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

librgw: move readdir operations into RGWFileHandle::readdir

The motivation here is to localize state updates to the class,
e.g., to hide consistency logic.

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

librgw: implement small-directory name cache

Augment the existing marker cache with cache of name->type
mappings for up to 128 dirents (appx. 4x the common upper-bound
of 32).

The motivation here is to use the name and type information from
readdir, rather than discarding it.  Boosts perf. at least 60%
on unit test corpus.

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

librgw: remote redundnat prints in RGWReaddirRequest

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

librgw: don't increment offset at rcb (it is no longer a counter)

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

librgw: fill in dirent cache

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

librgw: implement framework for fs periodic work

This can be used for, e.g., gc of stale handles and directory
traversals.

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

librgw: add intrusive refcnt to RGWLibFS

Ensures that file system instances cannot be disposed while (e.g.)
background processing is ongoing.

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

librgw: header and namespace reorganization

Move RGWLibProcess and RGWLibFrontend into new rgw_lib_frontend.h
header, which can include rgw_lib.h and rgw_file.h.

This allows the already-defined RGWLibProcess to call methods on
RGWLibFS objects (e.g., to perform GC, which follows in a separate
commit).

In addition, move all of the library and file interface into
namespace rgw.

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

librgw: use string_ref remove_prefix

Save a few cycles over re-assigning a string_ref suffix substring
to itself.

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

librgw: fix initialization of string_ref from temporary

Found by valgrind.  The good news:  amounts to proof that the
sref notation works.

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

cohort_lru: use correct new/delete operator pair

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

librgw: move ObjUnref trace print ahead of...unref

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

librgw: cleanup add_marker and handlers

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

librgw: rename RGWGetObjRequest, RGWListBucketRequest

The behavior of these requests is basically specific to the
needs of NFS READDIR and READ.  (RGWWriteRequest already updated.)

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

common: add (tweaked) sstring.h

Add sstring.h from Seastar, slightly adapted for c++2011 and
boost::string_ref.

An sstring resembles std::string, but the size of its small-string
optimization (static allocation) is controlled by a template
argument (cf. small_vector).

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

librgw: fix RGWLibFS::stat_leaf ex-ref

Found verifying refcounts at RGWLibFS::close.

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

lirgw: call RGWLibFS::close in rgw_umount

Also remove an unused warning in currently-disabled rgw_readv.

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

librgw: implement LRU lifecycle for handles

This re-uses intrusive lane LRU and co-resident intrusive b-tree
for lookups (formerly, we were using just the dictionary).

Machinery for cache drain on unmount/release of an active filesystem
instance has been added.  Currently, this code won't be reached.

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

librgw: remove dead code (rgw_readdir)

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

librgw: use correct finish path for RGWWriteRequest (logging and stats)

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

librgw: update posix-style read path

Update RGWGetObj to take ulp buffer rather than transferring the
elements of a buffer::list and re-traversing it.

This breaks the readv variant, but that was not final.  What it
fixes is offset handling.

Also, skip CREATE_BUCKET if not requested, in librgw_file_aw.

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

librgw: set iov->iov_len to nread (get)

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

librgw: sequential rgw_write works

This commit finishes hooking up RGWWriteRequest, and its supporting
RGWLibContinuedRequest.

The rgw_read call now fails--trivially, because it never handled
starting offset correctly.

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

librgw: update mtime on write finish

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

librgw: -EISDIR

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

librgw: check variant type, fix reversed check for write trans

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

librgw: finish write transactions in rgw_close path

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

librgw: block in RGWWriteRequest::exec_continue

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

librgw: block in RGWWriteRequest::exec_start

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

librgw: wire up more of continuation

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

librgw: dont forget to call exec_start

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

librgw: partial implementation of continued request framework

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

librgw_file_aw: re-enable write sequence

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

librgw: fix some incorrect uses of string temporary

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

librgw: add initial atomic write unit test

Tests are focused on open-write-close, open-read-close sequences.

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

librgw: don't allow open on directories (externally)

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

librgw: restrict objects to single-open

this is the simplest consistent model for an open-write-close
transaction (and it reads stable)

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

librgw: fix last enumeration bug, move tracing to debug

The last issue with enumeration was caused by inserting objects
into handle cache with their key-ified full_object_name()--this
should be the base name.

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

librgw: fix and unify full-name assembly and hashing

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

librgw: incremental listing fixes

* remove path in RGWStatLeafRequest--use bool is_dir instead
* remove unused uri in RGWListBucketRequest
* more dbg prints; problem w/final segment in obj_rec

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

librgw: incremental enumeration progress

As of this change, enumeration is -almost- correct, but the
rules for name, object_name(), and full_object_name() need to be
checked.

The rationale for storing just the last component name in dirents
means that name can't ever have the form "foo/bar" (make sure it
doesn't).  But there may still be code which assumes name ==
object_name() OR assumes object_name() is hashable (it isn't, that
has to be the full object name).

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

librgw: remove unused string uri

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

librgw: namespace bugfixes

1. stat_leaf() and stat_bucket() require RGWFileHandle::FLAG_CREATE
2. refactor RGWStatLeafRequest on the model of RGWListBucketRequest
3. fix uses of is_object() which meant is_file()
4. fix parent_fh passing

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

librgw_file_nfsns: stage namespace traversal (fails)

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

librgw: initial librgw_file_nfsns.cc test

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

librgw: add missing include deps (rgw_file.h)

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

librgw: fix parent chasing in make_fhk

Avoid an...infinite loop.

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

librgw: fix last-of marker detection

Add the count of common_prefixes to the existing listing size, rather
than overwriting it, else size cannot be compared w/ix.

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

librgw: unify result cases in rgw_mkdir

In both paths, instantiate a file handle iff the remote op
succeeded.

If this operation lost a race to create an object with the same
name, we cannot correctly return a valid result and handle in all
cases, so don't attempt it.

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

librgw: RGWListBucketRequest and related bugfixes

1. instantiate handles as required in RGWLibFS::stat_bucket
2. fix bucket_name() when bucket
3. fix full_object_name(), which should not return paths w/buckets
   by default (allow doing so by passing non-default depth)
4. fix uri, prefix, and delimiter logic
5. add logic to parse and forward common_prefixes

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

librgw: implement stat_bucket, call from rgw_lookup

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

librgw: annotate unlink's rgw_lookup

An rgw_lookup w/RGW_LOOKUP_FLAG_NONE will return an existing (hence
validated, modulo CREATE races) handle, but fail with ENOENT if none
exists (which is the desired behavior).

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

librgw: add lookup step in rgw_unlink

Implement the required check for underlying leaf object, since with
the current signature it is not known.

(Is that a fixed requirement? If we have previously done a lookup, we
already know the underlying object by name and type.)

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

librgw: hide uri in RGWListBucketRequest, use full_objectname()

Really would like to remove the requirement to format a URI for these
requests, since it is merely taken apart in parsing before execution.

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

librgw: add marker processing to RGWListBucket(s)Request

Ignores object instances, for now.  Not verified.

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

librgw: reimplement RGWLibFS::stat_leaf

Now the code attempts to read an object header (which may be cached),
and if this fails, uses RGWStatLeafRequest to find a potential matching
file or directory leaf.

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

librgw: variant handle subtype and marker cache

RGWFileHandle gets a type-specific variant part.  For the directory
variant, define a flat_map of cookie to marker.  Hash dirent names and
update the marker_cache in readdirs.

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

librgw: create dir file handles w/correct type

RGW objects which happen to represent directories will always end
in '/', but the trailing slash will not be present in cached versions
of the name.

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

librgw: implement RGWLibFS::stat_leaf

Stats one of <object>, <object/>.   Additional control over handle
creation has been added to RGWLibFS::lookup_fh.

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

librgw: incr rework stat

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

librgw: fix uninitialized rgw_fh

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

cohort_lru: fix a maybe uninitialized warning (that seems invalid)

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

librgw: create explicit directory leaf objects in rgw_mkdir

There is some asymmetry here, since directories are also inferred
from object names containing embedded "/."

The obvious question seems to be, "when can a leaf directory be
garbage collected?"  The answer seems to be, "only explicitly."

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

librgw: consolidate RGWLib RGWOps into RGWRequest

The intermediate derivation mirrored that in RGWRest, but we
don't seem to need it.

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

librgw: don't discard lookup result, fix signedness of min_size

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

librgw: make_fhk put base name, object_name in wrong order

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

librgw: add pseudo() accessor

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

librgw: add pseudofs awareness to lookup_fh

And then allow rgw_mkdir() to create pseudofs vnodes, when parent
is not the root.

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

librgw: make_fhk, compute object hashes path-wise

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

librgw: add depth and make_path to RGWFileHandle

This just (mostly) efficiently builds a full path from its segments
on demand.

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

librgw: compress RGWFileHandle stat data

Just save the state we actually mutate, and fill stat structures
on demand.

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

librgw: fix abuse of std::string& in RGWDeleteObjRequest

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

librgw: fix rgw_unlink

Fix handling of bucket namespace, remove bogus RGWLibFS::is_root()
logic.  Also remove a stray debug check.

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

librgw: ok, make RGW vnodes cache Unix attrs

Ganesa ULP needs to see a sane progression of attrs in the CREATE
path.

With this change shell actions like

$ echo "hi" > /nfs41/mybucket/mom

work as expected.

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

librgw: forge Unix attrs in rgw_create

Possibly RGW vnodes should cache Unix attrs, it would reduce the
latency of (frequent) getattr calls, and eventually lines up with
more materialized Unix attributes.

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

librgw: fix Unix file modes

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

librgw: implement rgw_create + bugfixes

Incremental work to support CREATE,OPEN,WRITE*,CLOSE transactions.
Adds a new tuple-based signature for RGWFileHandle::lookup_fh, and
fixes an inconsisten initialization of fhk and fh.fh_hk.

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

librgw: fix size computation in RGWStatObjRequest

Overrides RGWGetObj::execute(), calls the parent method and then
saves obj_size from the manifest.

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

librgw: fix many rgw_file return codes

This should all be -(POSIX), or at any rate negative.

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

librgw: don't abuse fs_private

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

librgw: introduce RGW_LOOKUP_FLAG_CREATE

The current API relies on rgw_lookup and rgw_lookup_filehandle to
return file handle instances, and like POSIX, needs to deal in some
way with the NFS/POSIX open-for-create concept.

For now, allow rgw_lookup to return the canonical instance for an
object being created, i.e., if the option RGW_LOOKUP_FLAG_CREATE
is passed.  (I added an RGW_OPEN_FLAG_CREATE also, but it doesn't
do anything.)

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

librgw: add getattr unit tests, restructure

The fids1, fids2 vectors were confusing, as fids2 mapped the
objects in the union of all buckets.

Now there's a matrix of <bucket,object>, and it's trivial to
scan the objects in any one bucket.

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

librgw: handle a 3 cases in rgw_getattr

Only the object case is mostly correct.

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

librgw: handle rgw_lookup in the root (minimally)

In addition, check real return code from RGWStatRequest.

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

librgw: change validity check for rgw_lookup

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

librgw: add RGWStatObjRequest and use it

* implement a stat operation over RGWGetObj (returning early)
* use stat to validate rgw_lookup
* use stat to implement rgw_getattr

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

librgw: add diagnostic dump_buckets() call.

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

librgw: fail RGWLibFS::authorize() on user-by-id lookup failure

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

librgw: implement rgw_vfsstat

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

librgw: rgw_mkdir and rgw_create return addresses

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

librgw: pass struct rgw_fh_hk by reference

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

librgw: always add struct before rgw_file_handle in rgw_file.h

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

librgw: bring back lookup_handle()

Now as a member of RGWLibFS, and exposed via rgw_lookup_handle()
for use by library clients.

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

librgw_file_gp: check test state before closing null fh handle

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

librgw: implement intrusive filehandle cache

The cache is keyed on a hash of bucket name, plus a hash of the
object name.

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

librgw: add common/cohort_lru.h

This a cache-line-friendly set of primitives for

a) shareded, intrusive maps based on boost::intrusive
b) an LRU system based on them

The system uses per-lane (shard) locks.  Lane count, front cache
size, and lock type are template params.

Initially, use this in preference to an unshareded map or hashtable,
in new RGW code.

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

librgw_file: release file handles

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>

librgw_file_gp: don't leak a buffer::list

Signed-off-by: Matt Benjamin <mbenjamin@redhat.com>