4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2012, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lustre/osp/osp_object.c
38 * Lustre OST Proxy Device
40 * Author: Alex Zhuravlev <alexey.zhuravlev@intel.com>
41 * Author: Mikhail Pershin <mike.tappro@intel.com>
45 # define EXPORT_SYMTAB
47 #define DEBUG_SUBSYSTEM S_MDS
49 #include "osp_internal.h"
51 static void osp_object_assign_fid(const struct lu_env *env,
52 struct osp_device *d, struct osp_object *o)
54 struct osp_thread_info *osi = osp_env_info(env);
56 LASSERT(fid_is_zero(lu_object_fid(&o->opo_obj.do_lu)));
57 LASSERT(o->opo_reserved);
60 osp_precreate_get_fid(env, d, &osi->osi_fid);
62 lu_object_assign_fid(env, &o->opo_obj.do_lu, &osi->osi_fid);
65 static int osp_declare_attr_set(const struct lu_env *env, struct dt_object *dt,
66 const struct lu_attr *attr, struct thandle *th)
68 struct osp_device *d = lu2osp_dev(dt->do_lu.lo_dev);
69 struct osp_object *o = dt2osp_obj(dt);
75 * Usually we don't allow server stack to manipulate size
76 * but there is a special case when striping is created
77 * late, after stripless file got truncated to non-zero.
79 * In this case we do the following:
81 * 1) grab id in declare - this can lead to leaked OST objects
82 * but we don't currently have proper mechanism and the only
83 * options we have are to do truncate RPC holding transaction
84 * open (very bad) or to grab id in declare at cost of leaked
85 * OST object in same very rare unfortunate case (just bad)
86 * notice 1.6-2.0 do assignment outside of running transaction
87 * all the time, meaning many more chances for leaked objects.
89 * 2) send synchronous truncate RPC with just assigned id
92 /* there are few places in MDD code still passing NULL
93 * XXX: to be fixed soon */
97 if (attr->la_valid & LA_SIZE && attr->la_size > 0) {
98 LASSERT(!dt_object_exists(dt));
99 osp_object_assign_fid(env, d, o);
100 rc = osp_object_truncate(env, dt, attr->la_size);
106 /* no need in logging for new objects being created */
110 if (!(attr->la_valid & (LA_UID | LA_GID)))
114 * track all UID/GID changes via llog
116 rc = osp_sync_declare_add(env, o, MDS_SETATTR64_REC, th);
121 static int osp_attr_set(const struct lu_env *env, struct dt_object *dt,
122 const struct lu_attr *attr, struct thandle *th,
123 struct lustre_capa *capa)
125 struct osp_object *o = dt2osp_obj(dt);
130 /* we're interested in uid/gid changes only */
131 if (!(attr->la_valid & (LA_UID | LA_GID)))
134 /* new object, the very first ->attr_set()
135 * initializing attributes needs no logging
136 * all subsequent one are subject to the
137 * logging and synchronization with OST */
144 * once transaction is committed put proper command on
145 * the queue going to our OST
147 rc = osp_sync_add(env, o, MDS_SETATTR64_REC, th, attr);
149 /* XXX: send new uid/gid to OST ASAP? */
154 static int osp_declare_object_create(const struct lu_env *env,
155 struct dt_object *dt,
156 struct lu_attr *attr,
157 struct dt_allocation_hint *hint,
158 struct dt_object_format *dof,
161 struct osp_thread_info *osi = osp_env_info(env);
162 struct osp_device *d = lu2osp_dev(dt->do_lu.lo_dev);
163 struct osp_object *o = dt2osp_obj(dt);
164 const struct lu_fid *fid;
169 /* should happen to non-0 OSP only so that at least one object
170 * has been already declared in the scenario and LOD should
172 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_CREATE_FAIL) && d->opd_index == 1)
175 LASSERT(d->opd_last_used_oid_file);
176 fid = lu_object_fid(&dt->do_lu);
179 * There can be gaps in precreated ids and record to unlink llog
180 * XXX: we do not handle gaps yet, implemented before solution
181 * was found to be racy, so we disabled that. there is no
182 * point in making useless but expensive llog declaration.
184 /* rc = osp_sync_declare_add(env, o, MDS_UNLINK64_REC, th); */
186 if (unlikely(!fid_is_zero(fid))) {
187 /* replay case: caller knows fid */
188 osi->osi_off = sizeof(osi->osi_id) * d->opd_index;
189 rc = dt_declare_record_write(env, d->opd_last_used_oid_file,
190 sizeof(osi->osi_id), osi->osi_off,
196 * in declaration we need to reserve object so that we don't block
197 * awaiting precreation RPC to complete
199 rc = osp_precreate_reserve(env, d);
201 * we also need to declare update to local "last used id" file for
202 * recovery if object isn't used for a reason, we need to release
203 * reservation, this can be made in osd_object_release()
206 /* mark id is reserved: in create we don't want to talk
208 LASSERT(o->opo_reserved == 0);
211 /* common for all OSPs file hystorically */
212 osi->osi_off = sizeof(osi->osi_id) * d->opd_index;
213 rc = dt_declare_record_write(env, d->opd_last_used_oid_file,
214 sizeof(osi->osi_id), osi->osi_off,
217 /* not needed in the cache anymore */
218 set_bit(LU_OBJECT_HEARD_BANSHEE,
219 &dt->do_lu.lo_header->loh_flags);
224 static int osp_object_create(const struct lu_env *env, struct dt_object *dt,
225 struct lu_attr *attr,
226 struct dt_allocation_hint *hint,
227 struct dt_object_format *dof, struct thandle *th)
229 struct osp_thread_info *osi = osp_env_info(env);
230 struct osp_device *d = lu2osp_dev(dt->do_lu.lo_dev);
231 struct osp_object *o = dt2osp_obj(dt);
233 struct lu_fid *fid = &osi->osi_fid;
236 if (o->opo_reserved) {
237 /* regular case, fid is assigned holding trunsaction open */
238 osp_object_assign_fid(env, d, o);
241 memcpy(fid, lu_object_fid(&dt->do_lu), sizeof(*fid));
243 LASSERTF(fid_is_sane(fid), "fid for osp_obj %p is insane"DFID"!\n",
246 if (!o->opo_reserved) {
247 /* special case, id was assigned outside of transaction
248 * see comments in osp_declare_attr_set */
249 spin_lock(&d->opd_pre_lock);
250 osp_update_last_fid(d, fid);
251 spin_unlock(&d->opd_pre_lock);
254 CDEBUG(D_INODE, "fid for osp_obj %p is "DFID"!\n", osp_obj, PFID(fid));
256 /* If the precreate ends, it means it will be ready to rollover to
257 * the new sequence soon, all the creation should be synchronized,
258 * otherwise during replay, the replay fid will be inconsistent with
259 * last_used/create fid */
260 if (osp_precreate_end_seq(env, d) && osp_is_fid_client(d))
264 * it's OK if the import is inactive by this moment - id was created
265 * by OST earlier, we just need to maintain it consistently on the disk
266 * once import is reconnected, OSP will claim this and other objects
267 * used and OST either keep them, if they exist or recreate
270 /* we might have lost precreated objects */
271 if (unlikely(d->opd_gap_count) > 0) {
272 spin_lock(&d->opd_pre_lock);
273 if (d->opd_gap_count > 0) {
274 int count = d->opd_gap_count;
275 osi->osi_oi.oi_id = fid_oid(&d->opd_gap_start_fid);
276 d->opd_gap_count = 0;
277 spin_unlock(&d->opd_pre_lock);
279 CDEBUG(D_HA, "Writting gap "DFID"+%d in llog\n",
280 PFID(&d->opd_gap_start_fid), count);
281 /* real gap handling is disabled intil ORI-692 will be
282 * fixed, now we only report gaps */
284 spin_unlock(&d->opd_pre_lock);
288 /* new object, the very first ->attr_set()
289 * initializing attributes needs no logging */
292 /* Only need update last_used oid file, seq file will only be update
293 * during seq rollover */
294 osp_objid_buf_prep(&osi->osi_lb, &osi->osi_off,
295 &d->opd_last_used_fid.f_oid, d->opd_index);
297 rc = dt_record_write(env, d->opd_last_used_oid_file, &osi->osi_lb,
300 CDEBUG(D_HA, "%s: Wrote last used FID: "DFID", index %d: %d\n",
301 d->opd_obd->obd_name, PFID(fid), d->opd_index, rc);
306 static int osp_declare_object_destroy(const struct lu_env *env,
307 struct dt_object *dt, struct thandle *th)
309 struct osp_object *o = dt2osp_obj(dt);
315 * track objects to be destroyed via llog
317 rc = osp_sync_declare_add(env, o, MDS_UNLINK64_REC, th);
322 static int osp_object_destroy(const struct lu_env *env, struct dt_object *dt,
325 struct osp_object *o = dt2osp_obj(dt);
331 * once transaction is committed put proper command on
332 * the queue going to our OST
334 rc = osp_sync_add(env, o, MDS_UNLINK64_REC, th, NULL);
336 /* not needed in cache any more */
337 set_bit(LU_OBJECT_HEARD_BANSHEE, &dt->do_lu.lo_header->loh_flags);
342 struct dt_object_operations osp_obj_ops = {
343 .do_declare_attr_set = osp_declare_attr_set,
344 .do_attr_set = osp_attr_set,
345 .do_declare_create = osp_declare_object_create,
346 .do_create = osp_object_create,
347 .do_declare_destroy = osp_declare_object_destroy,
348 .do_destroy = osp_object_destroy,
351 static int is_ost_obj(struct lu_object *lo)
353 struct osp_device *osp = lu2osp_dev(lo->lo_dev);
355 return !osp->opd_connect_mdt;
358 static int osp_object_init(const struct lu_env *env, struct lu_object *o,
359 const struct lu_object_conf *conf)
361 struct osp_object *po = lu2osp_obj(o);
366 po->opo_obj.do_ops = &osp_obj_ops;
368 struct lu_attr *la = &osp_env_info(env)->osi_attr;
370 po->opo_obj.do_ops = &osp_md_obj_ops;
371 o->lo_header->loh_attr |= LOHA_REMOTE;
372 po->opo_obj.do_lock_ops = &osp_md_lock_ops;
373 /* Do not need get attr for new object */
374 if (!(conf != NULL && (conf->loc_flags & LOC_F_NEW) != 0)) {
375 rc = po->opo_obj.do_ops->do_attr_get(env, lu2dt_obj(o),
378 o->lo_header->loh_attr |=
379 LOHA_EXISTS | (la->la_mode & S_IFMT);
387 static void osp_object_free(const struct lu_env *env, struct lu_object *o)
389 struct osp_object *obj = lu2osp_obj(o);
390 struct lu_object_header *h = o->lo_header;
392 dt_object_fini(&obj->opo_obj);
393 lu_object_header_fini(h);
394 OBD_SLAB_FREE_PTR(obj, osp_object_kmem);
397 static void osp_object_release(const struct lu_env *env, struct lu_object *o)
399 struct osp_object *po = lu2osp_obj(o);
400 struct osp_device *d = lu2osp_dev(o->lo_dev);
405 * release reservation if object was declared but not created
406 * this may require lu_object_put() in LOD
408 if (unlikely(po->opo_reserved)) {
409 LASSERT(d->opd_pre_reserved > 0);
410 spin_lock(&d->opd_pre_lock);
411 d->opd_pre_reserved--;
412 spin_unlock(&d->opd_pre_lock);
414 /* not needed in cache any more */
415 set_bit(LU_OBJECT_HEARD_BANSHEE, &o->lo_header->loh_flags);
420 static int osp_object_print(const struct lu_env *env, void *cookie,
421 lu_printer_t p, const struct lu_object *l)
423 const struct osp_object *o = lu2osp_obj((struct lu_object *)l);
425 return (*p)(env, cookie, LUSTRE_OSP_NAME"-object@%p", o);
428 static int osp_object_invariant(const struct lu_object *o)
433 struct lu_object_operations osp_lu_obj_ops = {
434 .loo_object_init = osp_object_init,
435 .loo_object_free = osp_object_free,
436 .loo_object_release = osp_object_release,
437 .loo_object_print = osp_object_print,
438 .loo_object_invariant = osp_object_invariant