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) 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 /* functions below are stubs for now, they will be implemented with
37 * grant support on MDT */
38 static inline void mdt_io_counter_incr(struct obd_export *exp, int opcode,
39 char *jobid, long amount)
44 static inline void mdt_dom_read_lock(struct mdt_object *mo)
46 down_read(&mo->mot_dom_sem);
49 static inline void mdt_dom_read_unlock(struct mdt_object *mo)
51 up_read(&mo->mot_dom_sem);
54 static inline void mdt_dom_write_lock(struct mdt_object *mo)
56 down_write(&mo->mot_dom_sem);
59 static inline void mdt_dom_write_unlock(struct mdt_object *mo)
61 up_write(&mo->mot_dom_sem);
65 * Lock prolongation for Data-on-MDT.
66 * This is similar to OFD code but for DOM ibits lock.
68 static inline time64_t prolong_timeout(struct ptlrpc_request *req)
70 struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
74 return obd_timeout / 2;
76 req_timeout = req->rq_deadline - req->rq_arrival_time.tv_sec;
77 return max_t(time64_t, at_est2timeout(at_get(&svcpt->scp_at_estimate)),
81 static void mdt_dom_resource_prolong(struct ldlm_prolong_args *arg)
83 struct ldlm_resource *res;
84 struct ldlm_lock *lock;
88 res = ldlm_resource_get(arg->lpa_export->exp_obd->obd_namespace, NULL,
89 &arg->lpa_resid, LDLM_EXTENT, 0);
92 "Failed to get resource for resid %llu/%llu\n",
93 arg->lpa_resid.name[0], arg->lpa_resid.name[1]);
98 list_for_each_entry(lock, &res->lr_granted, l_res_link) {
99 if (ldlm_has_dom(lock)) {
100 LDLM_DEBUG(lock, "DOM lock to prolong ");
101 ldlm_lock_prolong_one(lock, arg);
106 ldlm_resource_putref(res);
111 static void mdt_prolong_dom_lock(struct tgt_session_info *tsi,
112 struct ldlm_prolong_args *data)
114 struct obdo *oa = &tsi->tsi_ost_body->oa;
115 struct ldlm_lock *lock;
119 data->lpa_timeout = prolong_timeout(tgt_ses_req(tsi));
120 data->lpa_export = tsi->tsi_exp;
121 data->lpa_resid = tsi->tsi_resid;
123 CDEBUG(D_RPCTRACE, "Prolong DOM lock for req %p with x%llu\n",
124 tgt_ses_req(tsi), tgt_ses_req(tsi)->rq_xid);
126 if (oa->o_valid & OBD_MD_FLHANDLE) {
127 /* mostly a request should be covered by only one lock, try
129 lock = ldlm_handle2lock(&oa->o_handle);
131 LASSERT(lock->l_export == data->lpa_export);
132 ldlm_lock_prolong_one(lock, data);
133 lock->l_last_used = ktime_get();
135 if (data->lpa_locks_cnt > 0)
139 mdt_dom_resource_prolong(data);
143 static int mdt_rw_hpreq_lock_match(struct ptlrpc_request *req,
144 struct ldlm_lock *lock)
146 struct obd_ioobj *ioo;
148 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
152 if (!(lock->l_policy_data.l_inodebits.bits & MDS_INODELOCK_DOM))
155 ioo = req_capsule_client_get(&req->rq_pill, &RMF_OBD_IOOBJ);
156 LASSERT(ioo != NULL);
158 LASSERT(lock->l_resource != NULL);
159 if (!fid_res_name_eq(&ioo->ioo_oid.oi_fid, &lock->l_resource->lr_name))
162 /* a bulk write can only hold a reference on a PW extent lock. */
165 /* whereas a bulk read can be protected by either a PR or PW
169 if (!(lock->l_granted_mode & mode))
175 static int mdt_rw_hpreq_check(struct ptlrpc_request *req)
177 struct tgt_session_info *tsi;
178 struct obd_ioobj *ioo;
179 struct niobuf_remote *rnb;
181 struct ldlm_prolong_args pa = { 0 };
185 /* Don't use tgt_ses_info() to get session info, because lock_match()
186 * can be called while request has no processing thread yet. */
187 tsi = lu_context_key_get(&req->rq_session, &tgt_session_key);
190 * Use LASSERT below because malformed RPCs should have
191 * been filtered out in tgt_hpreq_handler().
193 opc = lustre_msg_get_opc(req->rq_reqmsg);
194 LASSERT(opc == OST_READ || opc == OST_WRITE);
196 ioo = req_capsule_client_get(&req->rq_pill, &RMF_OBD_IOOBJ);
197 LASSERT(ioo != NULL);
199 rnb = req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE);
200 LASSERT(rnb != NULL);
201 LASSERT(!(rnb->rnb_flags & OBD_BRW_SRVLOCK));
203 pa.lpa_mode = LCK_PW;
205 pa.lpa_mode |= LCK_PR;
207 DEBUG_REQ(D_RPCTRACE, req, "%s %s: refresh rw locks for " DFID,
208 tgt_name(tsi->tsi_tgt), current->comm, PFID(&tsi->tsi_fid));
210 mdt_prolong_dom_lock(tsi, &pa);
212 if (pa.lpa_blocks_cnt > 0) {
214 "%s: refreshed %u locks timeout for req %p",
215 tgt_name(tsi->tsi_tgt), pa.lpa_blocks_cnt, req);
219 RETURN(pa.lpa_locks_cnt > 0 ? 0 : -ESTALE);
222 static void mdt_rw_hpreq_fini(struct ptlrpc_request *req)
224 mdt_rw_hpreq_check(req);
227 static struct ptlrpc_hpreq_ops mdt_hpreq_rw = {
228 .hpreq_lock_match = mdt_rw_hpreq_lock_match,
229 .hpreq_check = mdt_rw_hpreq_check,
230 .hpreq_fini = mdt_rw_hpreq_fini
234 * Assign high priority operations to an IO request.
236 * Check if the incoming request is a candidate for
237 * high-priority processing. If it is, assign it a high
238 * priority operations table.
240 * \param[in] tsi target session environment for this request
242 void mdt_hp_brw(struct tgt_session_info *tsi)
244 struct niobuf_remote *rnb;
245 struct obd_ioobj *ioo;
249 ioo = req_capsule_client_get(tsi->tsi_pill, &RMF_OBD_IOOBJ);
250 LASSERT(ioo != NULL); /* must exist after request preprocessing */
251 if (ioo->ioo_bufcnt > 0) {
252 rnb = req_capsule_client_get(tsi->tsi_pill, &RMF_NIOBUF_REMOTE);
253 LASSERT(rnb != NULL); /* must exist after preprocessing */
255 /* no high priority if server lock is needed */
256 if (rnb->rnb_flags & OBD_BRW_SRVLOCK ||
257 (lustre_msg_get_flags(tgt_ses_req(tsi)->rq_reqmsg) &
261 tgt_ses_req(tsi)->rq_ops = &mdt_hpreq_rw;
264 static int mdt_punch_hpreq_lock_match(struct ptlrpc_request *req,
265 struct ldlm_lock *lock)
267 struct tgt_session_info *tsi;
272 /* Don't use tgt_ses_info() to get session info, because lock_match()
273 * can be called while request has no processing thread yet. */
274 tsi = lu_context_key_get(&req->rq_session, &tgt_session_key);
277 * Use LASSERT below because malformed RPCs should have
278 * been filtered out in tgt_hpreq_handler().
280 LASSERT(tsi->tsi_ost_body != NULL);
281 if (tsi->tsi_ost_body->oa.o_valid & OBD_MD_FLHANDLE &&
282 tsi->tsi_ost_body->oa.o_handle.cookie == lock->l_handle.h_cookie)
285 oa = &tsi->tsi_ost_body->oa;
287 LASSERT(lock->l_resource != NULL);
288 if (!fid_res_name_eq(&oa->o_oi.oi_fid, &lock->l_resource->lr_name))
291 if (!(lock->l_granted_mode & LCK_PW))
298 * Implementation of ptlrpc_hpreq_ops::hpreq_lock_check for OST_PUNCH request.
300 * High-priority queue request check for whether the given punch request
301 * (\a req) is blocking an LDLM lock cancel. Also checks whether the request is
302 * covered by an LDLM lock.
306 * \param[in] req the incoming request
308 * \retval 1 if \a req is blocking an LDLM lock cancel
309 * \retval 0 if it is not
310 * \retval -ESTALE if lock is not found
312 static int mdt_punch_hpreq_check(struct ptlrpc_request *req)
314 struct tgt_session_info *tsi;
316 struct ldlm_prolong_args pa = { 0 };
320 /* Don't use tgt_ses_info() to get session info, because lock_match()
321 * can be called while request has no processing thread yet. */
322 tsi = lu_context_key_get(&req->rq_session, &tgt_session_key);
323 LASSERT(tsi != NULL);
324 oa = &tsi->tsi_ost_body->oa;
326 LASSERT(!(oa->o_valid & OBD_MD_FLFLAGS &&
327 oa->o_flags & OBD_FL_SRVLOCK));
329 pa.lpa_mode = LCK_PW;
331 CDEBUG(D_DLMTRACE, "%s: refresh DOM lock for "DFID"\n",
332 tgt_name(tsi->tsi_tgt), PFID(&tsi->tsi_fid));
334 mdt_prolong_dom_lock(tsi, &pa);
337 if (pa.lpa_blocks_cnt > 0) {
339 "%s: refreshed %u locks timeout for req %p.\n",
340 tgt_name(tsi->tsi_tgt), pa.lpa_blocks_cnt, req);
344 RETURN(pa.lpa_locks_cnt > 0 ? 0 : -ESTALE);
348 * Implementation of ptlrpc_hpreq_ops::hpreq_lock_fini for OST_PUNCH request.
350 * Called after the request has been handled. It refreshes lock timeout again
351 * so that client has more time to send lock cancel RPC.
353 * \param[in] req request which is being processed.
355 static void mdt_punch_hpreq_fini(struct ptlrpc_request *req)
357 mdt_punch_hpreq_check(req);
360 static struct ptlrpc_hpreq_ops mdt_hpreq_punch = {
361 .hpreq_lock_match = mdt_punch_hpreq_lock_match,
362 .hpreq_check = mdt_punch_hpreq_check,
363 .hpreq_fini = mdt_punch_hpreq_fini
366 void mdt_hp_punch(struct tgt_session_info *tsi)
368 LASSERT(tsi->tsi_ost_body != NULL); /* must exists if we are here */
369 /* no high-priority if server lock is needed */
370 if ((tsi->tsi_ost_body->oa.o_valid & OBD_MD_FLFLAGS &&
371 tsi->tsi_ost_body->oa.o_flags & OBD_FL_SRVLOCK) ||
372 tgt_conn_flags(tsi) & OBD_CONNECT_MDS ||
373 lustre_msg_get_flags(tgt_ses_req(tsi)->rq_reqmsg) & MSG_REPLAY)
375 tgt_ses_req(tsi)->rq_ops = &mdt_hpreq_punch;
378 static int mdt_preprw_read(const struct lu_env *env, struct obd_export *exp,
379 struct mdt_device *mdt, struct mdt_object *mo,
380 struct lu_attr *la, int niocount,
381 struct niobuf_remote *rnb, int *nr_local,
382 struct niobuf_local *lnb, char *jobid)
384 struct dt_object *dob;
385 int i, j, rc, tot_bytes = 0;
386 int maxlnb = *nr_local;
391 mdt_dom_read_lock(mo);
393 /* the only valid case when READ can find object is missing or stale
394 * when export is just evicted and open files are closed forcefully
395 * on server while client's READ can be in progress.
396 * This should not happen on healthy export, object can't be missing
397 * or dying because both states means it was finally destroyed.
399 level = exp->exp_failed ? D_INFO : D_ERROR;
400 if (!mdt_object_exists(mo)) {
402 "%s: READ IO to missing obj "DFID": rc = %d\n",
403 exp->exp_obd->obd_name, PFID(mdt_object_fid(mo)),
405 /* return 0 and continue with empty commit to skip such READ
406 * without more BRW errors.
410 if (lu_object_is_dying(&mo->mot_header)) {
412 "%s: READ IO to stale obj "DFID": rc = %d\n",
413 exp->exp_obd->obd_name, PFID(mdt_object_fid(mo)),
415 /* return 0 and continue with empty commit to skip such READ
416 * without more BRW errors.
421 dob = mdt_obj2dt(mo);
422 /* parse remote buffers to local buffers and prepare the latter */
423 for (i = 0, j = 0; i < niocount; i++) {
424 rc = dt_bufs_get(env, dob, rnb + i, lnb + j, maxlnb, 0);
425 if (unlikely(rc < 0))
427 /* correct index for local buffers to continue with */
431 tot_bytes += rnb[i].rnb_len;
434 rc = dt_attr_get(env, dob, la);
438 rc = dt_read_prep(env, dob, lnb, *nr_local);
442 mdt_io_counter_incr(exp, LPROC_MDT_IO_READ, jobid, tot_bytes);
445 dt_bufs_put(env, dob, lnb, *nr_local);
446 mdt_dom_read_unlock(mo);
450 static int mdt_preprw_write(const struct lu_env *env, struct obd_export *exp,
451 struct mdt_device *mdt, struct mdt_object *mo,
452 struct lu_attr *la, struct obdo *oa,
453 int objcount, struct obd_ioobj *obj,
454 struct niobuf_remote *rnb, int *nr_local,
455 struct niobuf_local *lnb, char *jobid)
457 struct dt_object *dob;
458 int i, j, k, rc = 0, tot_bytes = 0;
459 int maxlnb = *nr_local;
463 /* Process incoming grant info, set OBD_BRW_GRANTED flag and grant some
464 * space back if possible */
465 tgt_grant_prepare_write(env, exp, oa, rnb, obj->ioo_bufcnt);
467 mdt_dom_read_lock(mo);
469 /* don't report error in cases with failed export */
470 if (!mdt_object_exists(mo)) {
471 int level = exp->exp_failed ? D_INFO : D_ERROR;
475 "%s: WRITE IO to missing obj "DFID": rc = %d\n",
476 exp->exp_obd->obd_name, PFID(mdt_object_fid(mo)),
478 /* exit with no data written, note nr_local = 0 above */
481 if (lu_object_is_dying(&mo->mot_header)) {
482 /* This is possible race between object destroy followed by
483 * discard BL AST and client cache flushing. Object is
484 * referenced until discard finish.
486 CDEBUG(D_INODE, "WRITE IO to stale object "DFID"\n",
487 PFID(mdt_object_fid(mo)));
488 /* Note: continue with no error here to don't cause BRW errors
489 * but skip transaction in commitrw silently so no data is
494 dob = mdt_obj2dt(mo);
495 /* parse remote buffers to local buffers and prepare the latter */
496 for (i = 0, j = 0; i < obj->ioo_bufcnt; i++) {
497 rc = dt_bufs_get(env, dob, rnb + i, lnb + j, maxlnb, 1);
498 if (unlikely(rc < 0))
500 /* correct index for local buffers to continue with */
501 for (k = 0; k < rc; k++) {
502 lnb[j + k].lnb_flags = rnb[i].rnb_flags;
503 if (!(rnb[i].rnb_flags & OBD_BRW_GRANTED))
504 lnb[j + k].lnb_rc = -ENOSPC;
509 tot_bytes += rnb[i].rnb_len;
512 rc = dt_write_prep(env, dob, lnb, *nr_local);
516 mdt_io_counter_incr(exp, LPROC_MDT_IO_WRITE, jobid, tot_bytes);
519 dt_bufs_put(env, dob, lnb, *nr_local);
521 mdt_dom_read_unlock(mo);
522 /* tgt_grant_prepare_write() was called, so we must commit */
523 tgt_grant_commit(exp, oa->o_grant_used, rc);
524 /* let's still process incoming grant information packed in the oa,
525 * but without enforcing grant since we won't proceed with the write.
526 * Just like a read request actually. */
527 tgt_grant_prepare_read(env, exp, oa);
531 int mdt_obd_preprw(const struct lu_env *env, int cmd, struct obd_export *exp,
532 struct obdo *oa, int objcount, struct obd_ioobj *obj,
533 struct niobuf_remote *rnb, int *nr_local,
534 struct niobuf_local *lnb)
536 struct tgt_session_info *tsi = tgt_ses_info(env);
537 struct mdt_thread_info *info = tsi2mdt_info(tsi);
538 struct lu_attr *la = &info->mti_attr.ma_attr;
539 struct mdt_device *mdt = mdt_dev(exp->exp_obd->obd_lu_dev);
540 struct mdt_object *mo;
544 /* The default value PTLRPC_MAX_BRW_PAGES is set in tgt_brw_write()
545 * but for MDT it is different, correct it here. */
546 if (*nr_local > MD_MAX_BRW_PAGES)
547 *nr_local = MD_MAX_BRW_PAGES;
549 jobid = tsi->tsi_jobid;
551 if (!oa || objcount != 1 || obj->ioo_bufcnt == 0) {
552 CERROR("%s: bad parameters %p/%i/%i\n",
553 exp->exp_obd->obd_name, oa, objcount, obj->ioo_bufcnt);
557 mo = mdt_object_find(env, mdt, &tsi->tsi_fid);
559 GOTO(out, rc = PTR_ERR(mo));
561 LASSERT(info->mti_object == NULL);
562 info->mti_object = mo;
564 if (cmd == OBD_BRW_WRITE) {
565 la_from_obdo(la, oa, OBD_MD_FLGETATTR);
566 rc = mdt_preprw_write(env, exp, mdt, mo, la, oa,
567 objcount, obj, rnb, nr_local, lnb,
569 } else if (cmd == OBD_BRW_READ) {
570 tgt_grant_prepare_read(env, exp, oa);
571 rc = mdt_preprw_read(env, exp, mdt, mo, la,
572 obj->ioo_bufcnt, rnb, nr_local, lnb,
574 obdo_from_la(oa, la, LA_ATIME);
576 CERROR("%s: wrong cmd %d received!\n",
577 exp->exp_obd->obd_name, cmd);
581 lu_object_put(env, &mo->mot_obj);
582 info->mti_object = NULL;
588 static int mdt_commitrw_read(const struct lu_env *env, struct mdt_device *mdt,
589 struct mdt_object *mo, int objcount, int niocount,
590 struct niobuf_local *lnb)
592 struct dt_object *dob;
597 dob = mdt_obj2dt(mo);
600 dt_bufs_put(env, dob, lnb, niocount);
602 mdt_dom_read_unlock(mo);
606 static int mdt_commitrw_write(const struct lu_env *env, struct obd_export *exp,
607 struct mdt_device *mdt, struct mdt_object *mo,
608 struct lu_attr *la, int objcount, int niocount,
609 struct niobuf_local *lnb, unsigned long granted,
612 struct dt_device *dt = mdt->mdt_bottom;
613 struct dt_object *dob;
621 dob = mdt_obj2dt(mo);
624 GOTO(out, rc = old_rc);
626 la->la_valid &= LA_ATIME | LA_MTIME | LA_CTIME;
628 if (!dt_object_exists(dob))
629 GOTO(out, rc = -ENOENT);
630 if (lu_object_is_dying(&mo->mot_header)) {
631 /* Commit to stale object can be just skipped silently. */
632 CDEBUG(D_INODE, "skip commit to stale object "DFID"\n",
633 PFID(mdt_object_fid(mo)));
639 DEBUG_REQ(D_WARNING, tgt_ses_req(tgt_ses_info(env)),
640 "%s: commit with no pages for "DFID": rc = %d\n",
641 exp->exp_obd->obd_name, PFID(mdt_object_fid(mo)), rc);
645 th = dt_trans_create(env, dt);
647 GOTO(out, rc = PTR_ERR(th));
649 for (i = 0; i < niocount; i++) {
650 if (!(lnb[i].lnb_flags & OBD_BRW_ASYNC)) {
656 if (OBD_FAIL_CHECK(OBD_FAIL_OST_DQACQ_NET))
657 GOTO(out_stop, rc = -EINPROGRESS);
659 rc = dt_declare_write_commit(env, dob, lnb, niocount, th);
664 /* update [mac]time if needed */
665 rc = dt_declare_attr_set(env, dob, la, th);
670 tgt_vbr_obj_set(env, dob);
671 rc = dt_trans_start(env, dt, th);
675 dt_write_lock(env, dob, 0);
676 rc = dt_write_commit(env, dob, lnb, niocount, th);
681 rc = dt_attr_set(env, dob, la, th);
685 /* get attr to return */
686 rc = dt_attr_get(env, dob, la);
688 dt_write_unlock(env, dob);
691 /* Force commit to make the just-deleted blocks
692 * reusable. LU-456 */
697 if (rc == 0 && granted > 0) {
698 if (tgt_grant_commit_cb_add(th, exp, granted) == 0)
703 dt_trans_stop(env, dt, th);
704 if (rc == -ENOSPC && retries++ < 3) {
705 CDEBUG(D_INODE, "retry after force commit, retries:%d\n",
711 dt_bufs_put(env, dob, lnb, niocount);
712 mdt_dom_read_unlock(mo);
714 tgt_grant_commit(exp, granted, old_rc);
718 void mdt_dom_obj_lvb_update(const struct lu_env *env, struct mdt_object *mo,
721 struct mdt_device *mdt = mdt_dev(mo->mot_obj.lo_dev);
722 struct ldlm_res_id resid;
723 struct ldlm_resource *res;
725 fid_build_reg_res_name(mdt_object_fid(mo), &resid);
726 res = ldlm_resource_get(mdt->mdt_namespace, NULL, &resid,
731 /* Update lvbo data if exists. */
732 if (mdt_dom_lvb_is_valid(res))
733 mdt_dom_disk_lvbo_update(env, mo, res, increase_only);
734 ldlm_resource_putref(res);
737 int mdt_obd_commitrw(const struct lu_env *env, int cmd, struct obd_export *exp,
738 struct obdo *oa, int objcount, struct obd_ioobj *obj,
739 struct niobuf_remote *rnb, int npages,
740 struct niobuf_local *lnb, int old_rc)
742 struct mdt_thread_info *info = mdt_th_info(env);
743 struct mdt_device *mdt = mdt_dev(exp->exp_obd->obd_lu_dev);
744 struct mdt_object *mo = info->mti_object;
745 struct lu_attr *la = &info->mti_attr.ma_attr;
751 if (cmd == OBD_BRW_WRITE) {
752 /* Don't update timestamps if this write is older than a
753 * setattr which modifies the timestamps. b=10150 */
755 /* XXX when we start having persistent reservations this needs
756 * to be changed to ofd_fmd_get() to create the fmd if it
757 * doesn't already exist so we can store the reservation handle
759 valid = OBD_MD_FLUID | OBD_MD_FLGID;
760 if (tgt_fmd_check(exp, mdt_object_fid(mo),
761 mdt_info_req(info)->rq_xid))
762 valid |= OBD_MD_FLATIME | OBD_MD_FLMTIME |
765 la_from_obdo(la, oa, valid);
767 rc = mdt_commitrw_write(env, exp, mdt, mo, la, objcount,
768 npages, lnb, oa->o_grant_used, old_rc);
770 obdo_from_la(oa, la, VALID_FLAGS | LA_GID | LA_UID);
772 obdo_from_la(oa, la, LA_GID | LA_UID);
774 mdt_dom_obj_lvb_update(env, mo, false);
775 /* don't report overquota flag if we failed before reaching
777 if (old_rc == 0 && (rc == 0 || rc == -EDQUOT)) {
778 /* return the overquota flags to client */
779 if (lnb[0].lnb_flags & OBD_BRW_OVER_USRQUOTA) {
780 if (oa->o_valid & OBD_MD_FLFLAGS)
781 oa->o_flags |= OBD_FL_NO_USRQUOTA;
783 oa->o_flags = OBD_FL_NO_USRQUOTA;
786 if (lnb[0].lnb_flags & OBD_BRW_OVER_GRPQUOTA) {
787 if (oa->o_valid & OBD_MD_FLFLAGS)
788 oa->o_flags |= OBD_FL_NO_GRPQUOTA;
790 oa->o_flags = OBD_FL_NO_GRPQUOTA;
793 if (lnb[0].lnb_flags & OBD_BRW_OVER_PRJQUOTA) {
794 if (oa->o_valid & OBD_MD_FLFLAGS)
795 oa->o_flags |= OBD_FL_NO_PRJQUOTA;
797 oa->o_flags = OBD_FL_NO_PRJQUOTA;
800 oa->o_valid |= OBD_MD_FLFLAGS | OBD_MD_FLUSRQUOTA |
801 OBD_MD_FLGRPQUOTA | OBD_MD_FLPRJQUOTA;
803 } else if (cmd == OBD_BRW_READ) {
804 /* If oa != NULL then mdt_preprw_read updated the inode
805 * atime and we should update the lvb so that other glimpses
806 * will also get the updated value. bug 5972 */
808 mdt_dom_obj_lvb_update(env, mo, true);
809 rc = mdt_commitrw_read(env, mdt, mo, objcount, npages, lnb);
815 mdt_thread_info_fini(info);
819 int mdt_object_punch(const struct lu_env *env, struct dt_device *dt,
820 struct dt_object *dob, __u64 start, __u64 end,
828 /* we support truncate, not punch yet */
829 LASSERT(end == OBD_OBJECT_EOF);
831 if (!dt_object_exists(dob))
834 th = dt_trans_create(env, dt);
838 rc = dt_declare_attr_set(env, dob, la, th);
842 rc = dt_declare_punch(env, dob, start, OBD_OBJECT_EOF, th);
846 tgt_vbr_obj_set(env, dob);
847 rc = dt_trans_start(env, dt, th);
851 dt_write_lock(env, dob, 0);
852 rc = dt_punch(env, dob, start, OBD_OBJECT_EOF, th);
855 rc = dt_attr_set(env, dob, la, th);
859 dt_write_unlock(env, dob);
862 dt_trans_stop(env, dt, th);
866 int mdt_punch_hdl(struct tgt_session_info *tsi)
868 const struct obdo *oa = &tsi->tsi_ost_body->oa;
869 struct ost_body *repbody;
870 struct mdt_thread_info *info;
872 struct ldlm_namespace *ns = tsi->tsi_tgt->lut_obd->obd_namespace;
873 struct obd_export *exp = tsi->tsi_exp;
874 struct mdt_device *mdt = mdt_dev(exp->exp_obd->obd_lu_dev);
875 struct mdt_object *mo;
876 struct dt_object *dob;
878 struct lustre_handle lh = { 0, };
885 /* check that we do support OBD_CONNECT_TRUNCLOCK. */
886 BUILD_BUG_ON(!(OST_CONNECT_SUPPORTED & OBD_CONNECT_TRUNCLOCK));
888 if ((oa->o_valid & (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS)) !=
889 (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS))
890 RETURN(err_serious(-EPROTO));
892 repbody = req_capsule_server_get(tsi->tsi_pill, &RMF_OST_BODY);
894 RETURN(err_serious(-ENOMEM));
896 /* punch start,end are passed in o_size,o_blocks throught wire */
900 if (end != OBD_OBJECT_EOF) /* Only truncate is supported */
903 info = tsi2mdt_info(tsi);
904 la = &info->mti_attr.ma_attr;
905 /* standard truncate optimization: if file body is completely
906 * destroyed, don't send data back to the server. */
908 flags |= LDLM_FL_AST_DISCARD_DATA;
910 repbody->oa.o_oi = oa->o_oi;
911 repbody->oa.o_valid = OBD_MD_FLID;
913 srvlock = (exp_connect_flags(exp) & OBD_CONNECT_SRVLOCK) &&
914 oa->o_valid & OBD_MD_FLFLAGS &&
915 oa->o_flags & OBD_FL_SRVLOCK;
918 rc = tgt_mdt_data_lock(ns, &tsi->tsi_resid, &lh, LCK_PW,
924 CDEBUG(D_INODE, "calling punch for object "DFID", valid = %#llx"
925 ", start = %lld, end = %lld\n", PFID(&tsi->tsi_fid),
926 oa->o_valid, start, end);
928 mo = mdt_object_find(tsi->tsi_env, mdt, &tsi->tsi_fid);
930 GOTO(out_unlock, rc = PTR_ERR(mo));
932 if (!mdt_object_exists(mo))
933 GOTO(out_put, rc = -ENOENT);
935 /* Shouldn't happen on dirs */
936 if (S_ISDIR(lu_object_attr(&mo->mot_obj))) {
938 CERROR("%s: Truncate on dir "DFID": rc = %d\n",
939 exp->exp_obd->obd_name, PFID(&tsi->tsi_fid), rc);
943 mdt_dom_write_lock(mo);
944 dob = mdt_obj2dt(mo);
946 la_from_obdo(la, oa, OBD_MD_FLMTIME | OBD_MD_FLATIME | OBD_MD_FLCTIME);
948 la->la_valid |= LA_SIZE;
950 /* MDT supports FMD for Data-on-MDT needs */
951 if (la->la_valid & (LA_ATIME | LA_MTIME | LA_CTIME))
952 tgt_fmd_update(tsi->tsi_exp, &tsi->tsi_fid,
953 tgt_ses_req(tsi)->rq_xid);
955 rc = mdt_object_punch(tsi->tsi_env, mdt->mdt_bottom, dob,
957 mdt_dom_write_unlock(mo);
961 mdt_dom_obj_lvb_update(tsi->tsi_env, mo, false);
962 mdt_io_counter_incr(tsi->tsi_exp, LPROC_MDT_IO_PUNCH,
966 lu_object_put(tsi->tsi_env, &mo->mot_obj);
969 tgt_extent_unlock(&lh, LCK_PW);
971 mdt_thread_info_fini(info);
976 * MDT glimpse for Data-on-MDT
978 * If there is write lock on client then function issues glimpse_ast to get
979 * an actual size from that client.
982 int mdt_do_glimpse(const struct lu_env *env, struct ldlm_namespace *ns,
983 struct ldlm_resource *res)
985 union ldlm_policy_data policy;
986 struct lustre_handle lockh;
988 struct ldlm_lock *lock;
989 struct ldlm_glimpse_work *gl_work;
990 struct list_head gl_list;
995 /* There can be only one write lock covering data, try to match it. */
996 policy.l_inodebits.bits = MDS_INODELOCK_DOM;
997 mode = ldlm_lock_match(ns, LDLM_FL_TEST_LOCK,
998 &res->lr_name, LDLM_IBITS, &policy,
1001 /* There is no PW lock on this object; finished. */
1005 lock = ldlm_handle2lock(&lockh);
1010 * This check is for lock taken in mdt_reint_unlink() that does
1011 * not have l_glimpse_ast set. So the logic is: if there is a lock
1012 * with no l_glimpse_ast set, this object is being destroyed already.
1013 * Hence, if you are grabbing DLM locks on the server, always set
1014 * non-NULL glimpse_ast (e.g., ldlm_request.c::ldlm_glimpse_ast()).
1016 if (lock->l_glimpse_ast == NULL) {
1017 LDLM_DEBUG(lock, "no l_glimpse_ast");
1018 GOTO(out, rc = -ENOENT);
1021 OBD_SLAB_ALLOC_PTR_GFP(gl_work, ldlm_glimpse_work_kmem, GFP_ATOMIC);
1023 GOTO(out, rc = -ENOMEM);
1025 /* Populate the gl_work structure.
1026 * Grab additional reference on the lock which will be released in
1027 * ldlm_work_gl_ast_lock() */
1028 gl_work->gl_lock = LDLM_LOCK_GET(lock);
1029 /* The glimpse callback is sent to one single IO lock. As a result,
1030 * the gl_work list is just composed of one element */
1031 INIT_LIST_HEAD(&gl_list);
1032 list_add_tail(&gl_work->gl_list, &gl_list);
1033 /* There is actually no need for a glimpse descriptor when glimpsing
1035 gl_work->gl_desc = NULL;
1036 /* the ldlm_glimpse_work structure is allocated on the stack */
1037 gl_work->gl_flags = LDLM_GL_WORK_SLAB_ALLOCATED;
1039 ldlm_glimpse_locks(res, &gl_list); /* this will update the LVB */
1041 /* If the list is not empty, we failed to glimpse a lock and
1042 * must clean it up. Usually due to a race with unlink.*/
1043 if (!list_empty(&gl_list)) {
1044 LDLM_LOCK_RELEASE(lock);
1045 OBD_SLAB_FREE_PTR(gl_work, ldlm_glimpse_work_kmem);
1050 LDLM_LOCK_PUT(lock);
1054 static void mdt_lvb2body(struct ldlm_resource *res, struct mdt_body *mb)
1056 struct ost_lvb *res_lvb;
1059 res_lvb = res->lr_lvb_data;
1060 mb->mbo_dom_size = res_lvb->lvb_size;
1061 mb->mbo_dom_blocks = res_lvb->lvb_blocks;
1062 mb->mbo_mtime = res_lvb->lvb_mtime;
1063 mb->mbo_ctime = res_lvb->lvb_ctime;
1064 mb->mbo_atime = res_lvb->lvb_atime;
1066 CDEBUG(D_DLMTRACE, "size %llu\n", res_lvb->lvb_size);
1068 mb->mbo_valid |= OBD_MD_FLATIME | OBD_MD_FLCTIME | OBD_MD_FLMTIME |
1074 * MDT glimpse for Data-on-MDT
1076 * This function is called when MDT get attributes for the DoM object.
1077 * If there is write lock on client then function issues glimpse_ast to get
1078 * an actual size from that client.
1080 int mdt_dom_object_size(const struct lu_env *env, struct mdt_device *mdt,
1081 const struct lu_fid *fid, struct mdt_body *mb,
1084 struct ldlm_res_id resid;
1085 struct ldlm_resource *res;
1090 fid_build_reg_res_name(fid, &resid);
1091 res = ldlm_resource_get(mdt->mdt_namespace, NULL, &resid,
1096 /* Update lvbo data if DoM lock returned or if LVB is not yet valid. */
1097 if (dom_lock || !mdt_dom_lvb_is_valid(res))
1098 mdt_dom_lvbo_update(res, NULL, NULL, false);
1100 mdt_lvb2body(res, mb);
1101 ldlm_resource_putref(res);
1106 * MDT DoM lock intent policy (glimpse)
1108 * Intent policy is called when lock has an intent, for DoM file that
1109 * means glimpse lock and policy fills Lock Value Block (LVB).
1111 * If already granted lock is found it will be placed in \a lockp and
1112 * returned back to caller function.
1114 * \param[in] tsi session info
1115 * \param[in,out] lockp pointer to the lock
1116 * \param[in] flags LDLM flags
1118 * \retval ELDLM_LOCK_REPLACED if already granted lock was found
1119 * and placed in \a lockp
1120 * \retval ELDLM_LOCK_ABORTED in other cases except error
1121 * \retval negative value on error
1123 int mdt_glimpse_enqueue(struct mdt_thread_info *mti, struct ldlm_namespace *ns,
1124 struct ldlm_lock **lockp, __u64 flags)
1126 struct ldlm_lock *lock = *lockp;
1127 struct ldlm_resource *res = lock->l_resource;
1128 ldlm_processing_policy policy;
1129 struct ldlm_reply *rep;
1130 struct mdt_body *mbo;
1135 policy = ldlm_get_processing_policy(res);
1136 LASSERT(policy != NULL);
1138 req_capsule_set_size(mti->mti_pill, &RMF_MDT_MD, RCL_SERVER, 0);
1139 req_capsule_set_size(mti->mti_pill, &RMF_ACL, RCL_SERVER, 0);
1140 rc = req_capsule_server_pack(mti->mti_pill);
1142 RETURN(err_serious(rc));
1144 rep = req_capsule_server_get(mti->mti_pill, &RMF_DLM_REP);
1148 mbo = req_capsule_server_get(mti->mti_pill, &RMF_MDT_BODY);
1153 /* Check if this is a resend case (MSG_RESENT is set on RPC) and a
1154 * lock was found by ldlm_handle_enqueue(); if so no need to grant
1156 if (flags & LDLM_FL_RESENT) {
1157 rc = LDLM_ITER_CONTINUE;
1159 __u64 tmpflags = LDLM_FL_BLOCK_NOWAIT;
1160 enum ldlm_error err;
1162 rc = policy(lock, &tmpflags, LDLM_PROCESS_RESCAN, &err, NULL);
1163 check_res_locked(res);
1167 /* The lock met with no resistance; we're finished. */
1168 if (rc == LDLM_ITER_CONTINUE) {
1169 GOTO(fill_mbo, rc = ELDLM_LOCK_REPLACED);
1170 } else if (flags & LDLM_FL_BLOCK_NOWAIT) {
1171 /* LDLM_FL_BLOCK_NOWAIT means it is for AGL. Do not send glimpse
1172 * callback for glimpse size. The real size user will trigger
1173 * the glimpse callback when necessary. */
1174 GOTO(fill_mbo, rc = ELDLM_LOCK_ABORTED);
1177 rc = mdt_do_glimpse(mti->mti_env, ns, res);
1178 if (rc == -ENOENT) {
1179 /* We are racing with unlink(); just return -ENOENT */
1180 rep->lock_policy_res2 = ptlrpc_status_hton(-ENOENT);
1182 } else if (rc == -EINVAL) {
1183 /* this is possible is client lock has been cancelled but
1184 * still exists on server. If that lock was found on server
1185 * as only conflicting lock then the client has already
1186 * size authority and glimpse is not needed. */
1187 CDEBUG(D_DLMTRACE, "Glimpse from the client owning lock\n");
1189 } else if (rc < 0) {
1192 rc = ELDLM_LOCK_ABORTED;
1194 /* LVB can be without valid data in case of DOM */
1195 if (!mdt_dom_lvb_is_valid(res))
1196 mdt_dom_lvbo_update(res, lock, NULL, false);
1197 mdt_lvb2body(res, mbo);
1201 int mdt_brw_enqueue(struct mdt_thread_info *mti, struct ldlm_namespace *ns,
1202 struct ldlm_lock **lockp, __u64 flags)
1204 struct tgt_session_info *tsi = tgt_ses_info(mti->mti_env);
1205 struct lu_fid *fid = &tsi->tsi_fid;
1206 struct ldlm_lock *lock = *lockp;
1207 struct ldlm_resource *res = lock->l_resource;
1208 struct ldlm_reply *rep;
1209 struct mdt_body *mbo;
1210 struct mdt_lock_handle *lhc = &mti->mti_lh[MDT_LH_RMT];
1211 struct mdt_object *mo;
1216 /* Get lock from request for possible resent case. */
1217 mdt_intent_fixup_resent(mti, *lockp, lhc, flags);
1218 req_capsule_set_size(mti->mti_pill, &RMF_MDT_MD, RCL_SERVER, 0);
1219 req_capsule_set_size(mti->mti_pill, &RMF_ACL, RCL_SERVER, 0);
1220 rc = req_capsule_server_pack(mti->mti_pill);
1222 RETURN(err_serious(rc));
1224 rep = req_capsule_server_get(mti->mti_pill, &RMF_DLM_REP);
1228 mbo = req_capsule_server_get(mti->mti_pill, &RMF_MDT_BODY);
1232 fid_extract_from_res_name(fid, &res->lr_name);
1233 mo = mdt_object_find(mti->mti_env, mti->mti_mdt, fid);
1234 if (unlikely(IS_ERR(mo)))
1235 RETURN(PTR_ERR(mo));
1237 if (!mdt_object_exists(mo))
1238 GOTO(out, rc = -ENOENT);
1240 if (mdt_object_remote(mo))
1241 GOTO(out, rc = -EPROTO);
1244 if (!lustre_handle_is_used(&lhc->mlh_reg_lh)) {
1245 mdt_lock_handle_init(lhc);
1246 mdt_lock_reg_init(lhc, (*lockp)->l_req_mode);
1247 /* This will block MDT thread but it should be fine until
1248 * client caches small amount of data for DoM, which should be
1249 * smaller than one BRW RPC and should be able to be
1250 * piggybacked by lock cancel RPC.
1251 * If the client could hold the lock too long, this code can be
1252 * revised to call mdt_object_lock_try(). And if fails, it will
1253 * return ELDLM_OK here and fall back into normal lock enqueue
1256 rc = mdt_object_lock(mti, mo, lhc, MDS_INODELOCK_DOM);
1261 if (!mdt_dom_lvb_is_valid(res)) {
1262 rc = mdt_dom_lvb_alloc(res);
1265 mdt_dom_disk_lvbo_update(mti->mti_env, mo, res, false);
1267 mdt_lvb2body(res, mbo);
1269 rep->lock_policy_res2 = clear_serious(rc);
1270 if (rep->lock_policy_res2) {
1271 lhc->mlh_reg_lh.cookie = 0ull;
1272 GOTO(out, rc = ELDLM_LOCK_ABORTED);
1275 rc = mdt_intent_lock_replace(mti, lockp, lhc, flags, rc);
1277 mdt_object_put(mti->mti_env, mo);
1281 /* check if client has already DoM lock for given resource */
1282 bool mdt_dom_client_has_lock(struct mdt_thread_info *info,
1283 const struct lu_fid *fid)
1285 struct mdt_device *mdt = info->mti_mdt;
1286 union ldlm_policy_data *policy = &info->mti_policy;
1287 struct ldlm_res_id *res_id = &info->mti_res_id;
1288 struct lustre_handle lockh;
1289 enum ldlm_mode mode;
1290 struct ldlm_lock *lock;
1293 policy->l_inodebits.bits = MDS_INODELOCK_DOM;
1294 fid_build_reg_res_name(fid, res_id);
1296 mode = ldlm_lock_match(mdt->mdt_namespace, LDLM_FL_BLOCK_GRANTED |
1297 LDLM_FL_TEST_LOCK, res_id, LDLM_IBITS, policy,
1300 /* There is no other PW lock on this object; finished. */
1304 lock = ldlm_handle2lock(&lockh);
1308 /* check if lock from the same client */
1309 rc = (lock->l_export->exp_handle.h_cookie ==
1310 info->mti_exp->exp_handle.h_cookie);
1311 LDLM_LOCK_PUT(lock);
1316 * MDT request handler for OST_GETATTR RPC.
1318 * This is data-specific request to get object and layout versions under
1319 * IO lock. It is reliable only for Data-on-MDT files.
1321 * \param[in] tsi target session environment for this request
1323 * \retval 0 if successful
1324 * \retval negative value on error
1326 int mdt_data_version_get(struct tgt_session_info *tsi)
1328 struct mdt_thread_info *mti = mdt_th_info(tsi->tsi_env);
1329 struct mdt_device *mdt = mti->mti_mdt;
1330 struct mdt_body *repbody;
1331 struct mdt_object *mo = mti->mti_object;
1332 struct lov_comp_md_v1 *comp;
1333 struct lustre_handle lh = { 0 };
1336 enum ldlm_mode lock_mode = LCK_PR;
1342 req_capsule_set_size(tsi->tsi_pill, &RMF_MDT_MD, RCL_SERVER, 0);
1343 req_capsule_set_size(tsi->tsi_pill, &RMF_ACL, RCL_SERVER, 0);
1344 rc = req_capsule_server_pack(tsi->tsi_pill);
1345 if (unlikely(rc != 0))
1346 RETURN(err_serious(rc));
1348 repbody = req_capsule_server_get(tsi->tsi_pill, &RMF_MDT_BODY);
1349 if (repbody == NULL)
1352 srvlock = tsi->tsi_mdt_body->mbo_valid & OBD_MD_FLFLAGS &&
1353 tsi->tsi_mdt_body->mbo_flags & OBD_FL_SRVLOCK;
1356 if (unlikely(tsi->tsi_mdt_body->mbo_flags & OBD_FL_FLUSH))
1359 fid_build_reg_res_name(&tsi->tsi_fid, &tsi->tsi_resid);
1360 rc = tgt_mdt_data_lock(mdt->mdt_namespace, &tsi->tsi_resid,
1361 &lh, lock_mode, &flags);
1366 if (!mdt_object_exists(mo))
1367 GOTO(out, rc = -ENOENT);
1368 if (mdt_object_remote(mo))
1369 GOTO(out, rc = -EREMOTE);
1370 if (!S_ISREG(lu_object_attr(&mo->mot_obj)))
1371 GOTO(out, rc = -EBADF);
1373 /* Get version first */
1374 version = dt_version_get(tsi->tsi_env, mdt_obj2dt(mo));
1375 if (version && version != -EOPNOTSUPP) {
1376 repbody->mbo_valid |= OBD_MD_FLDATAVERSION;
1377 /* re-use mbo_ioepoch to transfer version */
1378 repbody->mbo_version = version;
1381 /* Read layout to get its version */
1382 rc = mdt_big_xattr_get(mti, mo, XATTR_NAME_LOV);
1383 if (rc == -ENODATA) /* File has no layout yet */
1388 comp = mti->mti_buf.lb_buf;
1389 if (le32_to_cpu(comp->lcm_magic) != LOV_MAGIC_COMP_V1) {
1390 CDEBUG(D_INFO, DFID" has no composite layout",
1391 PFID(&tsi->tsi_fid));
1392 GOTO(out, rc = -ESTALE);
1395 CDEBUG(D_INODE, DFID": layout version: %u\n",
1396 PFID(&tsi->tsi_fid), le32_to_cpu(comp->lcm_layout_gen));
1398 repbody->mbo_valid |= OBD_MD_LAYOUT_VERSION;
1399 /* re-use mbo_rdev for that */
1400 repbody->mbo_layout_gen = le32_to_cpu(comp->lcm_layout_gen);
1404 tgt_mdt_data_unlock(&lh, lock_mode);
1406 repbody->mbo_valid |= OBD_MD_FLFLAGS;
1407 repbody->mbo_flags = OBD_FL_FLUSH;
1411 /* read file data to the buffer */
1412 int mdt_dom_read_on_open(struct mdt_thread_info *mti, struct mdt_device *mdt,
1413 struct lustre_handle *lh)
1415 const struct lu_env *env = mti->mti_env;
1416 struct tgt_session_info *tsi = tgt_ses_info(env);
1417 struct req_capsule *pill = tsi->tsi_pill;
1418 const struct lu_fid *fid;
1419 struct ptlrpc_request *req = tgt_ses_req(tsi);
1420 struct mdt_body *mbo;
1421 struct dt_device *dt = mdt->mdt_bottom;
1422 struct dt_object *mo;
1424 struct niobuf_remote *rnb = NULL;
1425 struct niobuf_local *lnb;
1428 unsigned int len, copied = 0;
1429 int lnbs, nr_local, i;
1430 bool dom_lock = false;
1434 if (!req_capsule_field_present(pill, &RMF_NIOBUF_INLINE, RCL_SERVER)) {
1435 /* There is no reply buffers for this field, this means that
1436 * client has no support for data in reply.
1441 mbo = req_capsule_server_get(pill, &RMF_MDT_BODY);
1443 if (lustre_handle_is_used(lh)) {
1444 struct ldlm_lock *lock;
1446 lock = ldlm_handle2lock(lh);
1448 dom_lock = ldlm_has_dom(lock) && ldlm_has_layout(lock);
1449 LDLM_LOCK_PUT(lock);
1453 /* return data along with open only along with DoM lock */
1454 if (!dom_lock || !mdt->mdt_opts.mo_dom_read_open)
1457 if (!(mbo->mbo_valid & OBD_MD_DOM_SIZE))
1460 if (mbo->mbo_dom_size == 0)
1463 CDEBUG(D_INFO, "File size %llu, reply sizes %d/%d\n",
1464 mbo->mbo_dom_size, req->rq_reqmsg->lm_repsize, req->rq_replen);
1465 len = req->rq_reqmsg->lm_repsize - req->rq_replen;
1467 /* NB: at this moment we have the following sizes:
1468 * - req->rq_replen: used data in reply
1469 * - req->rq_reqmsg->lm_repsize: total allocated reply buffer at client
1471 * Ideal case when file size fits in allocated reply buffer,
1472 * that mean we can return whole data in reply. We can also fit more
1473 * data up to max_reply_size in total reply size, but this will cause
1474 * re-allocation on client and resend with larger buffer. This is still
1475 * faster than separate READ IO.
1476 * Third case if file is too big to fit even in maximum size, in that
1477 * case we return just tail to optimize possible append.
1479 * At the moment the following strategy is used:
1480 * 1) try to fit into the buffer we have
1481 * 2) return just file tail otherwise.
1483 if (mbo->mbo_dom_size <= len) {
1484 /* can fit whole data */
1485 len = mbo->mbo_dom_size;
1490 /* File tail offset must be aligned with larger page size
1491 * between client and server, so the maximum page size is
1492 * used here to align offset.
1494 * NB: DOM feature was introduced when server supports pagebits
1495 * already, so it should be always non-zero value. Report error
1496 * if it is not for some reason.
1498 if (!req->rq_export->exp_target_data.ted_pagebits) {
1499 CERROR("%s: client page bits are not saved on server\n",
1503 pgbits = max_t(int, PAGE_SHIFT,
1504 req->rq_export->exp_target_data.ted_pagebits);
1505 tail = mbo->mbo_dom_size % (1 << pgbits);
1507 /* no partial tail or tail can't fit in reply */
1508 if (tail == 0 || len < tail)
1512 offset = mbo->mbo_dom_size - len;
1514 LASSERT((offset % PAGE_SIZE) == 0);
1515 rc = req_capsule_server_grow(pill, &RMF_NIOBUF_INLINE,
1516 sizeof(*rnb) + len);
1518 /* failed to grow data buffer, just exit */
1519 GOTO(out, rc = -E2BIG);
1522 /* re-take MDT_BODY and NIOBUF_INLINE buffers after the buffer grow */
1523 mbo = req_capsule_server_get(pill, &RMF_MDT_BODY);
1524 fid = &mbo->mbo_fid1;
1525 if (!fid_is_sane(fid))
1526 GOTO(out, rc = -EINVAL);
1528 rnb = req_capsule_server_get(tsi->tsi_pill, &RMF_NIOBUF_INLINE);
1530 GOTO(out, rc = -EPROTO);
1532 buf = (char *)rnb + sizeof(*rnb);
1534 rnb->rnb_offset = offset;
1536 mo = dt_locate(env, dt, fid);
1538 GOTO(out_rnb, rc = PTR_ERR(mo));
1539 LASSERT(mo != NULL);
1541 dt_read_lock(env, mo, 0);
1542 if (!dt_object_exists(mo))
1543 GOTO(unlock, rc = -ENOENT);
1545 /* parse remote buffers to local buffers and prepare the latter */
1546 lnbs = (len >> PAGE_SHIFT) + 1;
1547 OBD_ALLOC(lnb, sizeof(*lnb) * lnbs);
1549 GOTO(unlock, rc = -ENOMEM);
1551 rc = dt_bufs_get(env, mo, rnb, lnb, lnbs, 0);
1552 if (unlikely(rc < 0))
1554 LASSERT(rc <= lnbs);
1556 rc = dt_read_prep(env, mo, lnb, nr_local);
1559 /* copy data to the buffer finally */
1560 for (i = 0; i < nr_local; i++) {
1561 char *p = kmap(lnb[i].lnb_page);
1564 LASSERT(lnb[i].lnb_page_offset == 0);
1565 off = lnb[i].lnb_len & ~PAGE_MASK;
1567 memset(p + off, 0, PAGE_SIZE - off);
1569 memcpy(buf + (i << PAGE_SHIFT), p, lnb[i].lnb_len);
1570 kunmap(lnb[i].lnb_page);
1571 copied += lnb[i].lnb_len;
1574 CDEBUG(D_INFO, "Read %i (wanted %u) bytes from %llu\n", copied,
1577 CWARN("%s: read %i bytes for "DFID
1578 " but wanted %u, is size wrong?\n",
1579 tsi->tsi_exp->exp_obd->obd_name, copied,
1580 PFID(&tsi->tsi_fid), len);
1581 /* Ignore partially copied data */
1586 dt_bufs_put(env, mo, lnb, nr_local);
1588 OBD_FREE(lnb, sizeof(*lnb) * lnbs);
1590 dt_read_unlock(env, mo);
1591 lu_object_put(env, &mo->do_lu);
1593 rnb->rnb_len = copied;
1595 /* Don't fail OPEN request if read-on-open is failed, but drop
1596 * a message in log about the error.
1599 CDEBUG(D_INFO, "Read-on-open is failed, rc = %d", rc);
1605 * Completion AST for DOM discard locks:
1607 * CP AST an DOM discard lock is called always right after enqueue or from
1608 * reprocess if lock was blocked, in the latest case l_ast_data is set to
1609 * the mdt_object which is kept while there are pending locks on it.
1611 int ldlm_dom_discard_cp_ast(struct ldlm_lock *lock, __u64 flags, void *data)
1613 struct mdt_object *mo;
1614 struct lustre_handle dom_lh;
1619 /* l_ast_data is set when lock was not granted immediately
1620 * in mdt_dom_discard_data() below but put into waiting list,
1621 * so this CP callback means we are finished and corresponding
1622 * MDT object should be released finally as well as lock itself.
1624 lock_res_and_lock(lock);
1625 if (!lock->l_ast_data) {
1626 unlock_res_and_lock(lock);
1630 mo = lock->l_ast_data;
1631 lock->l_ast_data = NULL;
1632 unlock_res_and_lock(lock);
1634 ldlm_lock2handle(lock, &dom_lh);
1635 ldlm_lock_decref(&dom_lh, LCK_PW);
1637 env = lu_env_find();
1639 mdt_object_put(env, mo);
1644 void mdt_dom_discard_data(struct mdt_thread_info *info,
1645 struct mdt_object *mo)
1647 struct ptlrpc_request *req = mdt_info_req(info);
1648 struct mdt_device *mdt = mdt_dev(mo->mot_obj.lo_dev);
1649 union ldlm_policy_data policy;
1650 struct ldlm_res_id res_id;
1651 struct lustre_handle dom_lh;
1652 struct ldlm_lock *lock;
1653 __u64 flags = LDLM_FL_AST_DISCARD_DATA;
1659 if (req && req_is_replay(req))
1662 policy.l_inodebits.bits = MDS_INODELOCK_DOM;
1663 policy.l_inodebits.try_bits = 0;
1664 fid_build_reg_res_name(mdt_object_fid(mo), &res_id);
1666 /* Keep blocking version of discard for an old client to avoid
1667 * crashes on non-patched clients. LU-11359.
1669 old_client = req && !(exp_connect_flags2(req->rq_export) &
1670 OBD_CONNECT2_ASYNC_DISCARD);
1672 /* Tell the clients that the object is gone now and that they should
1673 * throw away any cached pages. */
1674 rc = ldlm_cli_enqueue_local(info->mti_env, mdt->mdt_namespace, &res_id,
1675 LDLM_IBITS, &policy, LCK_PW, &flags,
1676 ldlm_blocking_ast, old_client ?
1677 ldlm_completion_ast :
1678 ldlm_dom_discard_cp_ast,
1679 NULL, NULL, 0, LVB_T_NONE, NULL, &dom_lh);
1680 if (rc != ELDLM_OK) {
1682 "Failed to issue discard lock, rc = %d\n", rc);
1686 lock = ldlm_handle2lock(&dom_lh);
1687 lock_res_and_lock(lock);
1688 /* if lock is not granted then there are BL ASTs in progress and
1689 * lock will be granted in result of reprocessing with CP callback
1690 * notifying about that. The mdt object has to be kept until that and
1691 * it is saved in l_ast_data of the lock. Lock reference is kept too
1692 * until that to prevent it from canceling.
1694 if (!is_granted_or_cancelled_nolock(lock)) {
1695 mdt_object_get(info->mti_env, mo);
1696 lock->l_ast_data = mo;
1697 unlock_res_and_lock(lock);
1699 unlock_res_and_lock(lock);
1700 ldlm_lock_decref_and_cancel(&dom_lh, LCK_PW);
1702 LDLM_LOCK_PUT(lock);