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.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2012, 2017 Intel Corporation.
28 * Author: Mikhail Pershin <mike.pershin@intel.com>
31 #define DEBUG_SUBSYSTEM S_FILTER
33 #include <dt_object.h>
34 #include "mdt_internal.h"
36 /* --------------- MDT grant code ---------------- */
38 long mdt_grant_connect(const struct lu_env *env,
39 struct obd_export *exp,
40 u64 want, bool conservative)
42 struct mdt_device *mdt = mdt_exp2dev(exp);
48 dt_statfs(env, mdt->mdt_bottom, &mdt->mdt_osfs);
50 left = (mdt->mdt_osfs.os_bavail * mdt->mdt_osfs.os_bsize) / 2;
54 CDEBUG(D_CACHE, "%s: cli %s/%p ocd_grant: %ld want: %llu left: %llu\n",
55 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid,
56 exp, grant, want, left);
61 void mdt_grant_prepare_write(const struct lu_env *env,
62 struct obd_export *exp, struct obdo *oa,
63 struct niobuf_remote *rnb, int niocount)
65 struct mdt_device *mdt = mdt_exp2dev(exp);
70 left = (mdt->mdt_osfs.os_bavail * mdt->mdt_osfs.os_bsize) / 2;
72 /* grant more space back to the client if possible */
75 /* ---------------- end of MDT grant code ---------------- */
77 /* functions below are stubs for now, they will be implemented with
78 * grant support on MDT */
79 static inline void mdt_io_counter_incr(struct obd_export *exp, int opcode,
80 char *jobid, long amount)
85 void mdt_grant_prepare_read(const struct lu_env *env,
86 struct obd_export *exp, struct obdo *oa)
91 void mdt_grant_commit(struct obd_export *exp, unsigned long pending,
98 static inline void mdt_dom_read_lock(struct mdt_object *mo)
100 down_read(&mo->mot_dom_sem);
103 static inline void mdt_dom_read_unlock(struct mdt_object *mo)
105 up_read(&mo->mot_dom_sem);
108 static inline void mdt_dom_write_lock(struct mdt_object *mo)
110 down_write(&mo->mot_dom_sem);
113 static inline void mdt_dom_write_unlock(struct mdt_object *mo)
115 up_write(&mo->mot_dom_sem);
118 static int mdt_preprw_read(const struct lu_env *env, struct obd_export *exp,
119 struct mdt_device *mdt, struct mdt_object *mo,
120 struct lu_attr *la, int niocount,
121 struct niobuf_remote *rnb, int *nr_local,
122 struct niobuf_local *lnb, char *jobid)
124 struct dt_object *dob;
125 int i, j, rc, tot_bytes = 0;
129 mdt_dom_read_lock(mo);
130 if (!mdt_object_exists(mo))
131 GOTO(unlock, rc = -ENOENT);
133 dob = mdt_obj2dt(mo);
134 /* parse remote buffers to local buffers and prepare the latter */
136 for (i = 0, j = 0; i < niocount; i++) {
137 rc = dt_bufs_get(env, dob, rnb + i, lnb + j, 0);
138 if (unlikely(rc < 0))
140 /* correct index for local buffers to continue with */
143 tot_bytes += rnb[i].rnb_len;
146 rc = dt_attr_get(env, dob, la);
150 rc = dt_read_prep(env, dob, lnb, *nr_local);
154 mdt_io_counter_incr(exp, LPROC_MDT_IO_READ, jobid, tot_bytes);
157 dt_bufs_put(env, dob, lnb, *nr_local);
159 mdt_dom_read_unlock(mo);
163 static int mdt_preprw_write(const struct lu_env *env, struct obd_export *exp,
164 struct mdt_device *mdt, struct mdt_object *mo,
165 struct lu_attr *la, struct obdo *oa,
166 int objcount, struct obd_ioobj *obj,
167 struct niobuf_remote *rnb, int *nr_local,
168 struct niobuf_local *lnb, char *jobid)
170 struct dt_object *dob;
171 int i, j, k, rc = 0, tot_bytes = 0;
175 /* Process incoming grant info, set OBD_BRW_GRANTED flag and grant some
176 * space back if possible */
177 mdt_grant_prepare_write(env, exp, oa, rnb, obj->ioo_bufcnt);
179 mdt_dom_read_lock(mo);
180 if (!mdt_object_exists(mo)) {
181 CDEBUG(D_ERROR, "%s: BRW to missing obj "DFID"\n",
182 exp->exp_obd->obd_name, PFID(mdt_object_fid(mo)));
183 GOTO(unlock, rc = -ENOENT);
186 dob = mdt_obj2dt(mo);
187 /* parse remote buffers to local buffers and prepare the latter */
189 for (i = 0, j = 0; i < obj->ioo_bufcnt; i++) {
190 rc = dt_bufs_get(env, dob, rnb + i, lnb + j, 1);
191 if (unlikely(rc < 0))
193 /* correct index for local buffers to continue with */
194 for (k = 0; k < rc; k++)
195 lnb[j+k].lnb_flags = rnb[i].rnb_flags;
198 tot_bytes += rnb[i].rnb_len;
201 rc = dt_write_prep(env, dob, lnb, *nr_local);
205 mdt_io_counter_incr(exp, LPROC_MDT_IO_WRITE, jobid, tot_bytes);
208 dt_bufs_put(env, dob, lnb, *nr_local);
210 mdt_dom_read_unlock(mo);
211 /* tgt_grant_prepare_write() was called, so we must commit */
212 mdt_grant_commit(exp, oa->o_grant_used, rc);
213 /* let's still process incoming grant information packed in the oa,
214 * but without enforcing grant since we won't proceed with the write.
215 * Just like a read request actually. */
216 mdt_grant_prepare_read(env, exp, oa);
220 int mdt_obd_preprw(const struct lu_env *env, int cmd, struct obd_export *exp,
221 struct obdo *oa, int objcount, struct obd_ioobj *obj,
222 struct niobuf_remote *rnb, int *nr_local,
223 struct niobuf_local *lnb)
225 struct tgt_session_info *tsi = tgt_ses_info(env);
226 struct mdt_thread_info *info = tsi2mdt_info(tsi);
227 struct lu_attr *la = &info->mti_attr.ma_attr;
228 struct mdt_device *mdt = mdt_dev(exp->exp_obd->obd_lu_dev);
229 struct mdt_object *mo;
233 /* The default value PTLRPC_MAX_BRW_PAGES is set in tgt_brw_write()
234 * but for MDT it is different, correct it here. */
235 if (*nr_local > MD_MAX_BRW_PAGES)
236 *nr_local = MD_MAX_BRW_PAGES;
238 jobid = tsi->tsi_jobid;
240 if (!oa || objcount != 1 || obj->ioo_bufcnt == 0) {
241 CERROR("%s: bad parameters %p/%i/%i\n",
242 exp->exp_obd->obd_name, oa, objcount, obj->ioo_bufcnt);
246 mo = mdt_object_find(env, mdt, &tsi->tsi_fid);
248 GOTO(out, rc = PTR_ERR(mo));
250 LASSERT(info->mti_object == NULL);
251 info->mti_object = mo;
253 if (cmd == OBD_BRW_WRITE) {
254 la_from_obdo(la, oa, OBD_MD_FLGETATTR);
255 rc = mdt_preprw_write(env, exp, mdt, mo, la, oa,
256 objcount, obj, rnb, nr_local, lnb,
258 } else if (cmd == OBD_BRW_READ) {
259 mdt_grant_prepare_read(env, exp, oa);
260 rc = mdt_preprw_read(env, exp, mdt, mo, la,
261 obj->ioo_bufcnt, rnb, nr_local, lnb,
263 obdo_from_la(oa, la, LA_ATIME);
265 CERROR("%s: wrong cmd %d received!\n",
266 exp->exp_obd->obd_name, cmd);
270 lu_object_put(env, &mo->mot_obj);
271 info->mti_object = NULL;
277 static int mdt_commitrw_read(const struct lu_env *env, struct mdt_device *mdt,
278 struct mdt_object *mo, int objcount, int niocount,
279 struct niobuf_local *lnb)
281 struct dt_object *dob;
286 LASSERT(niocount > 0);
288 dob = mdt_obj2dt(mo);
290 dt_bufs_put(env, dob, lnb, niocount);
292 mdt_dom_read_unlock(mo);
296 static int mdt_commitrw_write(const struct lu_env *env, struct obd_export *exp,
297 struct mdt_device *mdt, struct mdt_object *mo,
298 struct lu_attr *la, int objcount, int niocount,
299 struct niobuf_local *lnb, unsigned long granted,
302 struct dt_device *dt = mdt->mdt_bottom;
303 struct dt_object *dob;
311 dob = mdt_obj2dt(mo);
314 GOTO(out, rc = old_rc);
316 la->la_valid &= LA_ATIME | LA_MTIME | LA_CTIME;
318 if (!dt_object_exists(dob))
319 GOTO(out, rc = -ENOENT);
321 th = dt_trans_create(env, dt);
323 GOTO(out, rc = PTR_ERR(th));
325 for (i = 0; i < niocount; i++) {
326 if (!(lnb[i].lnb_flags & OBD_BRW_ASYNC)) {
332 if (OBD_FAIL_CHECK(OBD_FAIL_OST_DQACQ_NET))
333 GOTO(out_stop, rc = -EINPROGRESS);
335 rc = dt_declare_write_commit(env, dob, lnb, niocount, th);
340 /* update [mac]time if needed */
341 rc = dt_declare_attr_set(env, dob, la, th);
346 rc = dt_trans_start(env, dt, th);
350 dt_write_lock(env, dob, 0);
351 rc = dt_write_commit(env, dob, lnb, niocount, th);
356 rc = dt_attr_set(env, dob, la, th);
360 /* get attr to return */
361 rc = dt_attr_get(env, dob, la);
363 dt_write_unlock(env, dob);
366 /* Force commit to make the just-deleted blocks
367 * reusable. LU-456 */
372 dt_trans_stop(env, dt, th);
373 if (rc == -ENOSPC && retries++ < 3) {
374 CDEBUG(D_INODE, "retry after force commit, retries:%d\n",
380 dt_bufs_put(env, dob, lnb, niocount);
381 mdt_dom_read_unlock(mo);
382 mdt_grant_commit(exp, granted, old_rc);
386 void mdt_dom_obj_lvb_update(const struct lu_env *env, struct mdt_object *mo,
389 struct mdt_device *mdt = mdt_dev(mo->mot_obj.lo_dev);
390 struct ldlm_res_id resid;
391 struct ldlm_resource *res;
393 fid_build_reg_res_name(mdt_object_fid(mo), &resid);
394 res = ldlm_resource_get(mdt->mdt_namespace, NULL, &resid,
399 /* Update lvbo data if exists. */
400 if (mdt_dom_lvb_is_valid(res))
401 mdt_dom_disk_lvbo_update(env, mo, res, increase_only);
402 ldlm_resource_putref(res);
405 int mdt_obd_commitrw(const struct lu_env *env, int cmd, struct obd_export *exp,
406 struct obdo *oa, int objcount, struct obd_ioobj *obj,
407 struct niobuf_remote *rnb, int npages,
408 struct niobuf_local *lnb, int old_rc)
410 struct mdt_thread_info *info = mdt_th_info(env);
411 struct mdt_device *mdt = mdt_dev(exp->exp_obd->obd_lu_dev);
412 struct mdt_object *mo = info->mti_object;
413 struct lu_attr *la = &info->mti_attr.ma_attr;
418 CERROR("%s: no pages to commit\n",
419 exp->exp_obd->obd_name);
425 if (cmd == OBD_BRW_WRITE) {
426 /* Don't update timestamps if this write is older than a
427 * setattr which modifies the timestamps. b=10150 */
429 /* XXX when we start having persistent reservations this needs
430 * to be changed to ofd_fmd_get() to create the fmd if it
431 * doesn't already exist so we can store the reservation handle
433 valid = OBD_MD_FLUID | OBD_MD_FLGID;
434 valid |= OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME;
436 la_from_obdo(la, oa, valid);
438 rc = mdt_commitrw_write(env, exp, mdt, mo, la, objcount,
439 npages, lnb, oa->o_grant_used, old_rc);
441 obdo_from_la(oa, la, VALID_FLAGS | LA_GID | LA_UID);
443 obdo_from_la(oa, la, LA_GID | LA_UID);
445 mdt_dom_obj_lvb_update(env, mo, false);
446 /* don't report overquota flag if we failed before reaching
448 if (old_rc == 0 && (rc == 0 || rc == -EDQUOT)) {
449 /* return the overquota flags to client */
450 if (lnb[0].lnb_flags & OBD_BRW_OVER_USRQUOTA) {
451 if (oa->o_valid & OBD_MD_FLFLAGS)
452 oa->o_flags |= OBD_FL_NO_USRQUOTA;
454 oa->o_flags = OBD_FL_NO_USRQUOTA;
457 if (lnb[0].lnb_flags & OBD_BRW_OVER_GRPQUOTA) {
458 if (oa->o_valid & OBD_MD_FLFLAGS)
459 oa->o_flags |= OBD_FL_NO_GRPQUOTA;
461 oa->o_flags = OBD_FL_NO_GRPQUOTA;
464 oa->o_valid |= OBD_MD_FLFLAGS | OBD_MD_FLUSRQUOTA |
467 } else if (cmd == OBD_BRW_READ) {
468 /* If oa != NULL then mdt_preprw_read updated the inode
469 * atime and we should update the lvb so that other glimpses
470 * will also get the updated value. bug 5972 */
472 mdt_dom_obj_lvb_update(env, mo, true);
473 rc = mdt_commitrw_read(env, mdt, mo, objcount, npages, lnb);
479 /* this put is pair to object_get in ofd_preprw_write */
480 mdt_thread_info_fini(info);
484 int mdt_object_punch(const struct lu_env *env, struct dt_device *dt,
485 struct dt_object *dob, __u64 start, __u64 end,
493 /* we support truncate, not punch yet */
494 LASSERT(end == OBD_OBJECT_EOF);
496 if (!dt_object_exists(dob))
499 th = dt_trans_create(env, dt);
503 rc = dt_declare_attr_set(env, dob, la, th);
507 rc = dt_declare_punch(env, dob, start, OBD_OBJECT_EOF, th);
511 tgt_vbr_obj_set(env, dob);
512 rc = dt_trans_start(env, dt, th);
516 dt_write_lock(env, dob, 0);
517 rc = dt_punch(env, dob, start, OBD_OBJECT_EOF, th);
520 rc = dt_attr_set(env, dob, la, th);
524 dt_write_unlock(env, dob);
527 dt_trans_stop(env, dt, th);
531 int mdt_punch_hdl(struct tgt_session_info *tsi)
533 const struct obdo *oa = &tsi->tsi_ost_body->oa;
534 struct ost_body *repbody;
535 struct mdt_thread_info *info;
537 struct ldlm_namespace *ns = tsi->tsi_tgt->lut_obd->obd_namespace;
538 struct obd_export *exp = tsi->tsi_exp;
539 struct mdt_device *mdt = mdt_dev(exp->exp_obd->obd_lu_dev);
540 struct mdt_object *mo;
541 struct dt_object *dob;
543 struct lustre_handle lh = { 0, };
550 /* check that we do support OBD_CONNECT_TRUNCLOCK. */
551 CLASSERT(OST_CONNECT_SUPPORTED & OBD_CONNECT_TRUNCLOCK);
553 if ((oa->o_valid & (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS)) !=
554 (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS))
555 RETURN(err_serious(-EPROTO));
557 repbody = req_capsule_server_get(tsi->tsi_pill, &RMF_OST_BODY);
559 RETURN(err_serious(-ENOMEM));
561 /* punch start,end are passed in o_size,o_blocks throught wire */
565 if (end != OBD_OBJECT_EOF) /* Only truncate is supported */
568 info = tsi2mdt_info(tsi);
569 la = &info->mti_attr.ma_attr;
570 /* standard truncate optimization: if file body is completely
571 * destroyed, don't send data back to the server. */
573 flags |= LDLM_FL_AST_DISCARD_DATA;
575 repbody->oa.o_oi = oa->o_oi;
576 repbody->oa.o_valid = OBD_MD_FLID;
578 srvlock = (exp_connect_flags(exp) & OBD_CONNECT_SRVLOCK) &&
579 oa->o_valid & OBD_MD_FLFLAGS &&
580 oa->o_flags & OBD_FL_SRVLOCK;
583 rc = tgt_mdt_data_lock(ns, &tsi->tsi_resid, &lh, LCK_PW,
589 CDEBUG(D_INODE, "calling punch for object "DFID", valid = %#llx"
590 ", start = %lld, end = %lld\n", PFID(&tsi->tsi_fid),
591 oa->o_valid, start, end);
593 mo = mdt_object_find(tsi->tsi_env, mdt, &tsi->tsi_fid);
595 GOTO(out_unlock, rc = PTR_ERR(mo));
597 mdt_dom_write_lock(mo);
598 if (!mdt_object_exists(mo))
599 GOTO(out_put, rc = -ENOENT);
600 dob = mdt_obj2dt(mo);
602 la_from_obdo(la, oa, OBD_MD_FLMTIME | OBD_MD_FLATIME | OBD_MD_FLCTIME);
604 la->la_valid |= LA_SIZE;
606 rc = mdt_object_punch(tsi->tsi_env, mdt->mdt_bottom, dob,
608 mdt_dom_write_unlock(mo);
612 mdt_dom_obj_lvb_update(tsi->tsi_env, mo, false);
613 mdt_io_counter_incr(tsi->tsi_exp, LPROC_MDT_IO_PUNCH,
617 lu_object_put(tsi->tsi_env, &mo->mot_obj);
620 mdt_save_lock(info, &lh, LCK_PW, rc);
622 mdt_thread_info_fini(info);
627 * MDT glimpse for Data-on-MDT
629 * If there is write lock on client then function issues glimpse_ast to get
630 * an actual size from that client.
633 int mdt_do_glimpse(const struct lu_env *env, struct ldlm_namespace *ns,
634 struct ldlm_resource *res)
636 union ldlm_policy_data policy;
637 struct lustre_handle lockh;
639 struct ldlm_lock *lock;
640 struct ldlm_glimpse_work *gl_work;
641 struct list_head gl_list;
646 /* There can be only one write lock covering data, try to match it. */
647 policy.l_inodebits.bits = MDS_INODELOCK_DOM;
648 mode = ldlm_lock_match(ns, LDLM_FL_BLOCK_GRANTED | LDLM_FL_TEST_LOCK,
649 &res->lr_name, LDLM_IBITS, &policy,
652 /* There is no PW lock on this object; finished. */
656 lock = ldlm_handle2lock(&lockh);
661 * This check is for lock taken in mdt_reint_unlink() that does
662 * not have l_glimpse_ast set. So the logic is: if there is a lock
663 * with no l_glimpse_ast set, this object is being destroyed already.
664 * Hence, if you are grabbing DLM locks on the server, always set
665 * non-NULL glimpse_ast (e.g., ldlm_request.c::ldlm_glimpse_ast()).
667 if (lock->l_glimpse_ast == NULL) {
668 LDLM_DEBUG(lock, "no l_glimpse_ast");
669 GOTO(out, rc = -ENOENT);
672 OBD_SLAB_ALLOC_PTR_GFP(gl_work, ldlm_glimpse_work_kmem, GFP_ATOMIC);
674 GOTO(out, rc = -ENOMEM);
676 /* Populate the gl_work structure.
677 * Grab additional reference on the lock which will be released in
678 * ldlm_work_gl_ast_lock() */
679 gl_work->gl_lock = LDLM_LOCK_GET(lock);
680 /* The glimpse callback is sent to one single IO lock. As a result,
681 * the gl_work list is just composed of one element */
682 INIT_LIST_HEAD(&gl_list);
683 list_add_tail(&gl_work->gl_list, &gl_list);
684 /* There is actually no need for a glimpse descriptor when glimpsing
686 gl_work->gl_desc = NULL;
687 /* the ldlm_glimpse_work structure is allocated on the stack */
688 gl_work->gl_flags = LDLM_GL_WORK_SLAB_ALLOCATED;
690 ldlm_glimpse_locks(res, &gl_list); /* this will update the LVB */
692 /* If the list is not empty, we failed to glimpse a lock and
693 * must clean it up. Usually due to a race with unlink.*/
694 if (!list_empty(&gl_list)) {
695 LDLM_LOCK_RELEASE(lock);
696 OBD_SLAB_FREE_PTR(gl_work, ldlm_glimpse_work_kmem);
705 static void mdt_lvb2body(struct ldlm_resource *res, struct mdt_body *mb)
707 struct ost_lvb *res_lvb;
710 res_lvb = res->lr_lvb_data;
711 mb->mbo_size = res_lvb->lvb_size;
712 mb->mbo_blocks = res_lvb->lvb_blocks;
713 mb->mbo_mtime = res_lvb->lvb_mtime;
714 mb->mbo_ctime = res_lvb->lvb_ctime;
715 mb->mbo_atime = res_lvb->lvb_atime;
717 CDEBUG(D_DLMTRACE, "size %llu\n", res_lvb->lvb_size);
719 mb->mbo_valid |= OBD_MD_FLATIME | OBD_MD_FLCTIME | OBD_MD_FLMTIME |
720 OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
725 * MDT glimpse for Data-on-MDT
727 * This function is called when MDT get attributes for the DoM object.
728 * If there is write lock on client then function issues glimpse_ast to get
729 * an actual size from that client.
731 int mdt_dom_object_size(const struct lu_env *env, struct mdt_device *mdt,
732 const struct lu_fid *fid, struct mdt_body *mb,
735 struct ldlm_res_id resid;
736 struct ldlm_resource *res;
741 fid_build_reg_res_name(fid, &resid);
742 res = ldlm_resource_get(mdt->mdt_namespace, NULL, &resid,
744 if (IS_ERR(res) || res->lr_lvb_data == NULL)
747 /* if there is no DOM bit in the lock then glimpse is needed
748 * to return valid size */
750 rc = mdt_do_glimpse(env, mdt->mdt_namespace, res);
755 /* Update lvbo data if DoM lock returned or if LVB is not yet valid. */
756 if (dom_lock || !mdt_dom_lvb_is_valid(res))
757 mdt_dom_lvbo_update(res, NULL, NULL, false);
759 mdt_lvb2body(res, mb);
761 ldlm_resource_putref(res);
766 * MDT DoM lock intent policy (glimpse)
768 * Intent policy is called when lock has an intent, for DoM file that
769 * means glimpse lock and policy fills Lock Value Block (LVB).
771 * If already granted lock is found it will be placed in \a lockp and
772 * returned back to caller function.
774 * \param[in] tsi session info
775 * \param[in,out] lockp pointer to the lock
776 * \param[in] flags LDLM flags
778 * \retval ELDLM_LOCK_REPLACED if already granted lock was found
779 * and placed in \a lockp
780 * \retval ELDLM_LOCK_ABORTED in other cases except error
781 * \retval negative value on error
783 int mdt_glimpse_enqueue(struct mdt_thread_info *mti, struct ldlm_namespace *ns,
784 struct ldlm_lock **lockp, __u64 flags)
786 struct ldlm_lock *lock = *lockp;
787 struct ldlm_resource *res = lock->l_resource;
788 ldlm_processing_policy policy;
789 struct ldlm_reply *rep;
790 struct mdt_body *mbo;
795 policy = ldlm_get_processing_policy(res);
796 LASSERT(policy != NULL);
798 req_capsule_set_size(mti->mti_pill, &RMF_MDT_MD, RCL_SERVER, 0);
799 req_capsule_set_size(mti->mti_pill, &RMF_ACL, RCL_SERVER, 0);
800 rc = req_capsule_server_pack(mti->mti_pill);
802 RETURN(err_serious(rc));
804 rep = req_capsule_server_get(mti->mti_pill, &RMF_DLM_REP);
808 mbo = req_capsule_server_get(mti->mti_pill, &RMF_MDT_BODY);
813 /* Check if this is a resend case (MSG_RESENT is set on RPC) and a
814 * lock was found by ldlm_handle_enqueue(); if so no need to grant
816 if (flags & LDLM_FL_RESENT) {
817 rc = LDLM_ITER_CONTINUE;
822 rc = policy(lock, &tmpflags, 0, &err, NULL);
823 check_res_locked(res);
827 /* The lock met with no resistance; we're finished. */
828 if (rc == LDLM_ITER_CONTINUE) {
829 GOTO(fill_mbo, rc = ELDLM_LOCK_REPLACED);
830 } else if (flags & LDLM_FL_BLOCK_NOWAIT) {
831 /* LDLM_FL_BLOCK_NOWAIT means it is for AGL. Do not send glimpse
832 * callback for glimpse size. The real size user will trigger
833 * the glimpse callback when necessary. */
834 GOTO(fill_mbo, rc = ELDLM_LOCK_ABORTED);
837 rc = mdt_do_glimpse(mti->mti_env, ns, res);
839 /* We are racing with unlink(); just return -ENOENT */
840 rep->lock_policy_res2 = ptlrpc_status_hton(-ENOENT);
842 } else if (rc == -EINVAL) {
843 /* this is possible is client lock has been cancelled but
844 * still exists on server. If that lock was found on server
845 * as only conflicting lock then the client has already
846 * size authority and glimpse is not needed. */
847 CDEBUG(D_DLMTRACE, "Glimpse from the client owning lock\n");
852 rc = ELDLM_LOCK_ABORTED;
854 /* LVB can be without valid data in case of DOM */
855 if (!mdt_dom_lvb_is_valid(res))
856 mdt_dom_lvbo_update(res, lock, NULL, false);
857 mdt_lvb2body(res, mbo);
861 int mdt_brw_enqueue(struct mdt_thread_info *mti, struct ldlm_namespace *ns,
862 struct ldlm_lock **lockp, __u64 flags)
864 struct tgt_session_info *tsi = tgt_ses_info(mti->mti_env);
865 struct lu_fid *fid = &tsi->tsi_fid;
866 struct ldlm_lock *lock = *lockp;
867 struct ldlm_resource *res = lock->l_resource;
868 struct ldlm_reply *rep;
869 struct mdt_body *mbo;
870 struct mdt_lock_handle *lhc = &mti->mti_lh[MDT_LH_RMT];
871 struct mdt_object *mo;
876 /* Get lock from request for possible resent case. */
877 mdt_intent_fixup_resent(mti, *lockp, lhc, flags);
878 req_capsule_set_size(mti->mti_pill, &RMF_MDT_MD, RCL_SERVER, 0);
879 req_capsule_set_size(mti->mti_pill, &RMF_ACL, RCL_SERVER, 0);
880 rc = req_capsule_server_pack(mti->mti_pill);
882 RETURN(err_serious(rc));
884 rep = req_capsule_server_get(mti->mti_pill, &RMF_DLM_REP);
888 mbo = req_capsule_server_get(mti->mti_pill, &RMF_MDT_BODY);
892 fid_extract_from_res_name(fid, &res->lr_name);
893 mo = mdt_object_find(mti->mti_env, mti->mti_mdt, fid);
894 if (unlikely(IS_ERR(mo)))
897 if (!mdt_object_exists(mo))
898 GOTO(out, rc = -ENOENT);
900 if (mdt_object_remote(mo))
901 GOTO(out, rc = -EPROTO);
904 if (!lustre_handle_is_used(&lhc->mlh_reg_lh)) {
905 mdt_lock_handle_init(lhc);
906 mdt_lock_reg_init(lhc, (*lockp)->l_req_mode);
907 /* This will block MDT thread but it should be fine until
908 * client caches small amount of data for DoM, which should be
909 * smaller than one BRW RPC and should be able to be
910 * piggybacked by lock cancel RPC.
911 * If the client could hold the lock too long, this code can be
912 * revised to call mdt_object_lock_try(). And if fails, it will
913 * return ELDLM_OK here and fall back into normal lock enqueue
916 rc = mdt_object_lock(mti, mo, lhc, MDS_INODELOCK_DOM);
921 if (!mdt_dom_lvb_is_valid(res)) {
922 rc = mdt_dom_lvb_alloc(res);
925 mdt_dom_disk_lvbo_update(mti->mti_env, mo, res, false);
927 mdt_lvb2body(res, mbo);
929 rep->lock_policy_res2 = clear_serious(rc);
930 if (rep->lock_policy_res2) {
931 lhc->mlh_reg_lh.cookie = 0ull;
932 GOTO(out, rc = ELDLM_LOCK_ABORTED);
935 rc = mdt_intent_lock_replace(mti, lockp, lhc, flags, rc);
937 mdt_object_put(mti->mti_env, mo);
941 void mdt_dom_discard_data(struct mdt_thread_info *info,
942 const struct lu_fid *fid)
944 struct mdt_device *mdt = info->mti_mdt;
945 union ldlm_policy_data *policy = &info->mti_policy;
946 struct ldlm_res_id *res_id = &info->mti_res_id;
947 struct lustre_handle dom_lh;
948 __u64 flags = LDLM_FL_AST_DISCARD_DATA;
951 policy->l_inodebits.bits = MDS_INODELOCK_DOM;
952 policy->l_inodebits.try_bits = 0;
953 fid_build_reg_res_name(fid, res_id);
955 /* Tell the clients that the object is gone now and that they should
956 * throw away any cached pages. */
957 rc = ldlm_cli_enqueue_local(mdt->mdt_namespace, res_id, LDLM_IBITS,
958 policy, LCK_PW, &flags, ldlm_blocking_ast,
959 ldlm_completion_ast, NULL, NULL, 0,
960 LVB_T_NONE, NULL, &dom_lh);
962 /* We only care about the side-effects, just drop the lock. */
964 ldlm_lock_decref(&dom_lh, LCK_PW);