I squashed the commits that I originally marked for squashing, and everything is pushed into wip-rgw-versioning-2. Following is a rough breakdown of the different development phases, and what I (think I) achieved in each one, hopefully it would give you some hint of the direction when reviewing. Note that I'm still missing some multi-zone related changes, and there are a couple of known regressions, but that should stop you. See http://wiki.ceph.com/Development/RGW_Object_Versioning for design doc. 1. Initial work In this initial phase, the idea was to add a version (or instance) identifier to objects. This allows creating multiple objects with the same name, but different instance ids. Objects can be accessed directly by using their names + version. The bucket index holds entries by name + instance. Versioned objects here are just regular objects that are named differently. 5a95240 - rgw: add versioning_enabled field to bucket info d4aa2ae - rgw: get bucket versioning status op 4d7ffbd - rgw: restful op to set bucket versioning 38275cb - rgw: enable s3 get/set versioning ops 655ae55 - rgw, cls_rgw: add accounted_size for object metadata entry ae13500 - rgw: remove unused code 727b9e8 - rgw: decouple object name from index representation ef3982f - rgw: rename rgw_obj::key to rgw_obj::loc d603ddf - rgw, cls_rgw: various datastructures use new rgw_obj_key c44e574 - rgw: rename cls_rgw_obj::key to cls_rgw_obj::loc e3dc560 - cls_rgw: change data structures to keep single object key structure 3e9177e - rgw: adapt to new objclass interface ad9f460 - radosgw-admin: adapt to new interfaces e81ef5f - test: cls_rgw fixes 3bfa127 - rgw: clean up some locator use d242339 - radosgw-admin: some commands use object_version param c97a05a - rgw: generate random instance id 162c53b - rgw: interface adjustment following a rebase 8ecef67 - rgw: remove old unused code 2. OLH Started olh logic implementation. Adding new cls/rgw calls: - link olh - unlink instance - read olh log - trim olh log New nucket index representation took a couple of iterations to settle. Major issue that was discovered mid-process was that versioned objects need to be sorted from newest to oldest when listing bucket. We now have 3 different kind of entries in the bucket index: - plain - instance - olh Plain entries are entries that repesent the objects' listing order. These entries are named as follows (for versioned objects, non versioned objects are treated as before): <name> \0 <decreasing_str(olh_epoch)> \0 <instance id> The instance entries reside in a different namespace, and objects are indexed there by their name and their version id. Thus, in order to get to the listing entries, we need to first read the instance entry. The olh entry contains the olh log, and the current olh epoch. Extra complexity: handling null-versioned objects (objects that are created on buckets with suspended versioning). Main changes: added a new entry to the bucket index for each versioned object to mark it as versioned (indexed only by its name). Every regular object that is overwritten is converted to a versioned object. Also, needed to co-locate olh and data objects, since objects can be versioned, but also have a 'null' version that needs to match the non-versioned case. So, an olh can point at itself, and when removing an object, we make sure that we don't remove the olh. 27a9408 - rgw: initial olh implementation 105c4e0 - rgw: gen rand lower alphanumberic string f6a21cd - rgw: adjust return code when generating random strings e65a1fa - rgw: gen rand lowercase string (stl string version) cc55d9e - rgw: init olh tag d01b4c7 - rgw: some code cleanup 8ae3c3c - rgw: obj_stat() follows on olh 932e3b1 - cls_rgw: prepare groundwork for olh 4211f28 - cls_rgw: encode / decode obj and list index keys 3ed6627 - cls_rgw: bucket index link olh fc2f127 - cls_rgw: object instance olh linking 4ac79e5 - rgw: bucket index link olh interface 56185d6 - rgw: new api to retrieve olh log 9f8f158 - rgw: implement rgw_bucket_olh_log_entry::dump() c92a332 - rgw-admin: add olh readlog command 717e7ec - cls_rgw: olh init op 946611c - rgw: apply olh log functionality d6d2f58 - rgw, cls_rgw: trim olh log functionality 4f9afa7 - rgw: olh atomicity groundwork 701f65c - rgw: guard against racing writes c1d57e3 - rgw: more atomicity fixes, set_olh() 4ffa2dd - rgw: tie set_olh() to object completion 019c226 - cls_rgw: olh trim op is read/write fc35420 - rgw: follow olh if needed b78481b - rgw: update json encoding for rgw_obj 88c2f1f - rgw: object manifest should reflect instance a5fea1d - rgw: add 'versioning', and 'versions' to handled subresources eadb243 - rgw: add get_type() to rgw ops 28ee25c - cls_rgw: revise the data model f57cc4a - rgw: bucket listing gets extra param for versioning 7a010c5 - rgw, cls_rgw: list object versions is optional 1892aaf - cls_rgw: deletion marker needs to keep instance entry 677c6f9 - rgw: propagate dirent flags to rgw (from cls), other fixes d123836 - rgw: restful api now dumps versions 64a66b5 - rgw: cleanup, get rid of req_state::object 77cdb69 - rgw: request state and various op functionality use rgw_obj_key f9ae1e8 - rgw: fix rgw_obj initialization c7cc445 - cls_rgw: update the appropriate prev key entry b18d36e - rgw, cls_rgw: cls_bucket_list returns raw key in map 704425b - rgw: add support for version-id-marker d5d4347 - rgw: bucket versioning status is tri-state 0fd49fe - rgw: initial versioned object removal implementation 3fb2177 - rgw, cls_rgw: don't remove olh objects 5962d5e - cls_rgw: allow olh linking to null instance objects fed201f - rgw: set olh if object has been versioned 3. Cleanup! At this point it was obvious that the code was in dire need of a cleanup. Trying to limit the amount of different states. Moving certain object operations to RGWRados subclasses. Separating data objects and system objects. 0998856 - rgw: move RGWRadosCtx into RGWRados f10469e - rgw: s/RGWRadosCtx/ObjectCtx b0bcedf - rgw: start reorganizing RGWRados 46774c2 - rgw: remove plain object processor 900c89a - rgw: pass around object context refrences, remove unused code bff2d83 - rgw: don't use put_system_obj() for data objects 2620bbe - rgw: get rid of put_obj_meta(), replace with put_system_obj() 172cceb - rgw: remove old index update calls 853c937 - rgw: remove unused code ed0076f - rgw: switch RGWRados::delete_obj() to new interface 32142fd - rgw: fix missing state initalization 1dd190b - rgw: remove more unused code 54a426b - rgw: rework prepare_get_obj(), get_obj() d485428 - rgw: change RGWRados::get_attr() e1041ed - rgw: clean up system obj interfaces 8114787 - rgw: s/RGWRados::ObjectCtx/RGWObjectCtx f0fa071 - rgw: adjust to new interfaces 0799f62 - rgw: purge intent log d291c08 - rgw: remove unused code ae33ad7 - rgw: switch get_obj_iterate() to new interface 8af2bcf - rgw: convert RGWRados::get_attr() to new interface and some cls_rgw cleanups: dd87374 - cls_rgw: reorganize rgw_bucket_link_olh() 956108d - cls_rgw: more cleanup 4302041 - cls_rgw: more cleanup 4. Back to versioning work More internal work, as described in (2). Also, we now have new radosgw-admin commands to list and set raw bucket index entries. This is really helpful in debugging issues related to bucket index versioning. 28e43ca - cls_rgw: update olh log when unlinking entry 7232a92 - cls_rgw: unlink object instance 518493b - rgw: unlink obj instance 136740c - rgw: follow olh where needed 56b0e6b - cls_rgw: keep null-versioned object as versioned object 6751908 - rgw-admin, cls_rgw: add bi_get objclass operation 061313d - common, rgw: json escaping gets input buf size 358fc98 - cls_rgw: add missing flags encoding to rgw_bucket_dir_entry::dump() 5d41b86 - cls_rgw, rgw-admin: move bi_get() entry encoding to cls 72fdef2 - cls_rgw, rgw-admin: create bi list operation efa541f - rgw, cls_rgw: add bi put 5. Fixes, adjustments, complete missing implementation misc stuff. Fixes, and other missing implementation. 16d5e06 - osd: fix filter_prefix scoping in omap_get_vals 04eeb7c - formatter: no need for dynamic allocation 662d805 - rgw: send "null" version id if needed d03c562 - rgw, cls_rgw: multiple changes related to obj removal 3638bdb - rgw: propagate object owner and mtime for deletion marker 16f2bd3 - rgw: adjust versioning enable/suspend api 4d3b6e3 - rgw: fix access to object through the null instance 520b0c7 - cls_rgw: inc olh epoch when updating log 9c329cc - rgw, cls_rgw: fix update of olh to reflect non existing object b6c0c12 - cls_rgw: add missing cls_cxx_create() c127068 - rgw: add dump_string_header() 0eacb86 - rgw: send x-amz-version-id and x-amz-delete_marker header fields 2481439 - cls_rgw: remove instance entry when removing delete marker 21dd843 - rgw: encode timestamp in pending olh info acec1c8 - rgw, cls_rgw: improve olh atomicity 2dae922 - cls_rgw: guard certain operations using olh tag 5d423c8 - cls_rgw: implement dump() and generate test instances dff4cae - cls_rgw: clean up compilation warnings 82766fa - rgw: remove unused code 8acd45b - rgw: remove warning 2fed1f5 - rgw: read bucket owner when following olh if pending entries 3d0b506 - cls_rgw: revise null object instance handling, versioned epoch 6f4d924 - cls_rgw: don't write list entry when converting when deleting f319b93 - rgw: time out pending olh entries 9e0f7a1 - rgw: Object::Read::read() returns total bytes read 20c61a8 - rgw: Object::Read operations should use state->obj 6be07f4 - rgw: reduce use of Object::get_obj() 14e1ec6 - rgw: parse copy location version id 750f4d7 - cls_rgw, rgw: pending_log can hold multiple entries per epoch a3a45cb - cls_rgw: link, unlink olh ops can get epoch 1266c59 - rgw, cls_rgw: provide optional version id, versioned epoch to olh ops 31695db - rgw: cleaup RGWRados::copy_obj() 4e790b8 - rgw: propagate version id when putting obj bd3738a - rgw: copy obj does versioning too 12dc4e1 - rgw: move versioning handling to Object::Write::write_meta() d2e9d4e - rgw: fix a few regressions -- To unsubscribe from this list: send the line "unsubscribe ceph-devel" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
