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: "DFID"\n",
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.\n",
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;
389 mdt_dom_read_lock(mo);
390 if (!mdt_object_exists(mo))
391 GOTO(unlock, rc = -ENOENT);
393 dob = mdt_obj2dt(mo);
394 /* parse remote buffers to local buffers and prepare the latter */
396 for (i = 0, j = 0; i < niocount; i++) {
397 rc = dt_bufs_get(env, dob, rnb + i, lnb + j, 0);
398 if (unlikely(rc < 0))
400 /* correct index for local buffers to continue with */
403 tot_bytes += rnb[i].rnb_len;
406 rc = dt_attr_get(env, dob, la);
410 rc = dt_read_prep(env, dob, lnb, *nr_local);
414 mdt_io_counter_incr(exp, LPROC_MDT_IO_READ, jobid, tot_bytes);
417 dt_bufs_put(env, dob, lnb, *nr_local);
419 mdt_dom_read_unlock(mo);
423 static int mdt_preprw_write(const struct lu_env *env, struct obd_export *exp,
424 struct mdt_device *mdt, struct mdt_object *mo,
425 struct lu_attr *la, struct obdo *oa,
426 int objcount, struct obd_ioobj *obj,
427 struct niobuf_remote *rnb, int *nr_local,
428 struct niobuf_local *lnb, char *jobid)
430 struct dt_object *dob;
431 int i, j, k, rc = 0, tot_bytes = 0;
435 /* Process incoming grant info, set OBD_BRW_GRANTED flag and grant some
436 * space back if possible */
437 tgt_grant_prepare_write(env, exp, oa, rnb, obj->ioo_bufcnt);
439 mdt_dom_read_lock(mo);
440 if (!mdt_object_exists(mo)) {
441 CDEBUG(D_ERROR, "%s: BRW to missing obj "DFID"\n",
442 exp->exp_obd->obd_name, PFID(mdt_object_fid(mo)));
443 GOTO(unlock, rc = -ENOENT);
446 dob = mdt_obj2dt(mo);
447 /* parse remote buffers to local buffers and prepare the latter */
449 for (i = 0, j = 0; i < obj->ioo_bufcnt; i++) {
450 rc = dt_bufs_get(env, dob, rnb + i, lnb + j, 1);
451 if (unlikely(rc < 0))
453 /* correct index for local buffers to continue with */
454 for (k = 0; k < rc; k++) {
455 lnb[j + k].lnb_flags = rnb[i].rnb_flags;
456 if (!(rnb[i].rnb_flags & OBD_BRW_GRANTED))
457 lnb[j + k].lnb_rc = -ENOSPC;
461 tot_bytes += rnb[i].rnb_len;
464 rc = dt_write_prep(env, dob, lnb, *nr_local);
468 mdt_io_counter_incr(exp, LPROC_MDT_IO_WRITE, jobid, tot_bytes);
471 dt_bufs_put(env, dob, lnb, *nr_local);
473 mdt_dom_read_unlock(mo);
474 /* tgt_grant_prepare_write() was called, so we must commit */
475 tgt_grant_commit(exp, oa->o_grant_used, rc);
476 /* let's still process incoming grant information packed in the oa,
477 * but without enforcing grant since we won't proceed with the write.
478 * Just like a read request actually. */
479 tgt_grant_prepare_read(env, exp, oa);
483 int mdt_obd_preprw(const struct lu_env *env, int cmd, struct obd_export *exp,
484 struct obdo *oa, int objcount, struct obd_ioobj *obj,
485 struct niobuf_remote *rnb, int *nr_local,
486 struct niobuf_local *lnb)
488 struct tgt_session_info *tsi = tgt_ses_info(env);
489 struct mdt_thread_info *info = tsi2mdt_info(tsi);
490 struct lu_attr *la = &info->mti_attr.ma_attr;
491 struct mdt_device *mdt = mdt_dev(exp->exp_obd->obd_lu_dev);
492 struct mdt_object *mo;
496 /* The default value PTLRPC_MAX_BRW_PAGES is set in tgt_brw_write()
497 * but for MDT it is different, correct it here. */
498 if (*nr_local > MD_MAX_BRW_PAGES)
499 *nr_local = MD_MAX_BRW_PAGES;
501 jobid = tsi->tsi_jobid;
503 if (!oa || objcount != 1 || obj->ioo_bufcnt == 0) {
504 CERROR("%s: bad parameters %p/%i/%i\n",
505 exp->exp_obd->obd_name, oa, objcount, obj->ioo_bufcnt);
509 mo = mdt_object_find(env, mdt, &tsi->tsi_fid);
511 GOTO(out, rc = PTR_ERR(mo));
513 LASSERT(info->mti_object == NULL);
514 info->mti_object = mo;
516 if (cmd == OBD_BRW_WRITE) {
517 la_from_obdo(la, oa, OBD_MD_FLGETATTR);
518 rc = mdt_preprw_write(env, exp, mdt, mo, la, oa,
519 objcount, obj, rnb, nr_local, lnb,
521 } else if (cmd == OBD_BRW_READ) {
522 tgt_grant_prepare_read(env, exp, oa);
523 rc = mdt_preprw_read(env, exp, mdt, mo, la,
524 obj->ioo_bufcnt, rnb, nr_local, lnb,
526 obdo_from_la(oa, la, LA_ATIME);
528 CERROR("%s: wrong cmd %d received!\n",
529 exp->exp_obd->obd_name, cmd);
533 lu_object_put(env, &mo->mot_obj);
534 info->mti_object = NULL;
540 static int mdt_commitrw_read(const struct lu_env *env, struct mdt_device *mdt,
541 struct mdt_object *mo, int objcount, int niocount,
542 struct niobuf_local *lnb)
544 struct dt_object *dob;
549 LASSERT(niocount > 0);
551 dob = mdt_obj2dt(mo);
553 dt_bufs_put(env, dob, lnb, niocount);
555 mdt_dom_read_unlock(mo);
559 static int mdt_commitrw_write(const struct lu_env *env, struct obd_export *exp,
560 struct mdt_device *mdt, struct mdt_object *mo,
561 struct lu_attr *la, int objcount, int niocount,
562 struct niobuf_local *lnb, unsigned long granted,
565 struct dt_device *dt = mdt->mdt_bottom;
566 struct dt_object *dob;
574 dob = mdt_obj2dt(mo);
577 GOTO(out, rc = old_rc);
579 la->la_valid &= LA_ATIME | LA_MTIME | LA_CTIME;
581 if (!dt_object_exists(dob))
582 GOTO(out, rc = -ENOENT);
584 th = dt_trans_create(env, dt);
586 GOTO(out, rc = PTR_ERR(th));
588 for (i = 0; i < niocount; i++) {
589 if (!(lnb[i].lnb_flags & OBD_BRW_ASYNC)) {
595 if (OBD_FAIL_CHECK(OBD_FAIL_OST_DQACQ_NET))
596 GOTO(out_stop, rc = -EINPROGRESS);
598 rc = dt_declare_write_commit(env, dob, lnb, niocount, th);
603 /* update [mac]time if needed */
604 rc = dt_declare_attr_set(env, dob, la, th);
609 tgt_vbr_obj_set(env, dob);
610 rc = dt_trans_start(env, dt, th);
614 dt_write_lock(env, dob, 0);
615 rc = dt_write_commit(env, dob, lnb, niocount, th);
620 rc = dt_attr_set(env, dob, la, th);
624 /* get attr to return */
625 rc = dt_attr_get(env, dob, la);
627 dt_write_unlock(env, dob);
630 /* Force commit to make the just-deleted blocks
631 * reusable. LU-456 */
636 if (rc == 0 && granted > 0) {
637 if (tgt_grant_commit_cb_add(th, exp, granted) == 0)
642 dt_trans_stop(env, dt, th);
643 if (rc == -ENOSPC && retries++ < 3) {
644 CDEBUG(D_INODE, "retry after force commit, retries:%d\n",
650 dt_bufs_put(env, dob, lnb, niocount);
651 mdt_dom_read_unlock(mo);
653 tgt_grant_commit(exp, granted, old_rc);
657 void mdt_dom_obj_lvb_update(const struct lu_env *env, struct mdt_object *mo,
660 struct mdt_device *mdt = mdt_dev(mo->mot_obj.lo_dev);
661 struct ldlm_res_id resid;
662 struct ldlm_resource *res;
664 fid_build_reg_res_name(mdt_object_fid(mo), &resid);
665 res = ldlm_resource_get(mdt->mdt_namespace, NULL, &resid,
670 /* Update lvbo data if exists. */
671 if (mdt_dom_lvb_is_valid(res))
672 mdt_dom_disk_lvbo_update(env, mo, res, increase_only);
673 ldlm_resource_putref(res);
676 int mdt_obd_commitrw(const struct lu_env *env, int cmd, struct obd_export *exp,
677 struct obdo *oa, int objcount, struct obd_ioobj *obj,
678 struct niobuf_remote *rnb, int npages,
679 struct niobuf_local *lnb, int old_rc)
681 struct mdt_thread_info *info = mdt_th_info(env);
682 struct mdt_device *mdt = mdt_dev(exp->exp_obd->obd_lu_dev);
683 struct mdt_object *mo = info->mti_object;
684 struct lu_attr *la = &info->mti_attr.ma_attr;
689 CERROR("%s: no pages to commit\n",
690 exp->exp_obd->obd_name);
696 if (cmd == OBD_BRW_WRITE) {
697 /* Don't update timestamps if this write is older than a
698 * setattr which modifies the timestamps. b=10150 */
700 /* XXX when we start having persistent reservations this needs
701 * to be changed to ofd_fmd_get() to create the fmd if it
702 * doesn't already exist so we can store the reservation handle
704 valid = OBD_MD_FLUID | OBD_MD_FLGID;
705 valid |= OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME;
707 la_from_obdo(la, oa, valid);
709 rc = mdt_commitrw_write(env, exp, mdt, mo, la, objcount,
710 npages, lnb, oa->o_grant_used, old_rc);
712 obdo_from_la(oa, la, VALID_FLAGS | LA_GID | LA_UID);
714 obdo_from_la(oa, la, LA_GID | LA_UID);
716 mdt_dom_obj_lvb_update(env, mo, false);
717 /* don't report overquota flag if we failed before reaching
719 if (old_rc == 0 && (rc == 0 || rc == -EDQUOT)) {
720 /* return the overquota flags to client */
721 if (lnb[0].lnb_flags & OBD_BRW_OVER_USRQUOTA) {
722 if (oa->o_valid & OBD_MD_FLFLAGS)
723 oa->o_flags |= OBD_FL_NO_USRQUOTA;
725 oa->o_flags = OBD_FL_NO_USRQUOTA;
728 if (lnb[0].lnb_flags & OBD_BRW_OVER_GRPQUOTA) {
729 if (oa->o_valid & OBD_MD_FLFLAGS)
730 oa->o_flags |= OBD_FL_NO_GRPQUOTA;
732 oa->o_flags = OBD_FL_NO_GRPQUOTA;
735 if (lnb[0].lnb_flags & OBD_BRW_OVER_PRJQUOTA) {
736 if (oa->o_valid & OBD_MD_FLFLAGS)
737 oa->o_flags |= OBD_FL_NO_PRJQUOTA;
739 oa->o_flags = OBD_FL_NO_PRJQUOTA;
742 oa->o_valid |= OBD_MD_FLFLAGS | OBD_MD_FLUSRQUOTA |
743 OBD_MD_FLGRPQUOTA | OBD_MD_FLPRJQUOTA;
745 } else if (cmd == OBD_BRW_READ) {
746 /* If oa != NULL then mdt_preprw_read updated the inode
747 * atime and we should update the lvb so that other glimpses
748 * will also get the updated value. bug 5972 */
750 mdt_dom_obj_lvb_update(env, mo, true);
751 rc = mdt_commitrw_read(env, mdt, mo, objcount, npages, lnb);
757 /* this put is pair to object_get in ofd_preprw_write */
758 mdt_thread_info_fini(info);
762 int mdt_object_punch(const struct lu_env *env, struct dt_device *dt,
763 struct dt_object *dob, __u64 start, __u64 end,
771 /* we support truncate, not punch yet */
772 LASSERT(end == OBD_OBJECT_EOF);
774 if (!dt_object_exists(dob))
777 th = dt_trans_create(env, dt);
781 rc = dt_declare_attr_set(env, dob, la, th);
785 rc = dt_declare_punch(env, dob, start, OBD_OBJECT_EOF, th);
789 tgt_vbr_obj_set(env, dob);
790 rc = dt_trans_start(env, dt, th);
794 dt_write_lock(env, dob, 0);
795 rc = dt_punch(env, dob, start, OBD_OBJECT_EOF, th);
798 rc = dt_attr_set(env, dob, la, th);
802 dt_write_unlock(env, dob);
805 dt_trans_stop(env, dt, th);
809 int mdt_punch_hdl(struct tgt_session_info *tsi)
811 const struct obdo *oa = &tsi->tsi_ost_body->oa;
812 struct ost_body *repbody;
813 struct mdt_thread_info *info;
815 struct ldlm_namespace *ns = tsi->tsi_tgt->lut_obd->obd_namespace;
816 struct obd_export *exp = tsi->tsi_exp;
817 struct mdt_device *mdt = mdt_dev(exp->exp_obd->obd_lu_dev);
818 struct mdt_object *mo;
819 struct dt_object *dob;
821 struct lustre_handle lh = { 0, };
828 /* check that we do support OBD_CONNECT_TRUNCLOCK. */
829 CLASSERT(OST_CONNECT_SUPPORTED & OBD_CONNECT_TRUNCLOCK);
831 if ((oa->o_valid & (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS)) !=
832 (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS))
833 RETURN(err_serious(-EPROTO));
835 repbody = req_capsule_server_get(tsi->tsi_pill, &RMF_OST_BODY);
837 RETURN(err_serious(-ENOMEM));
839 /* punch start,end are passed in o_size,o_blocks throught wire */
843 if (end != OBD_OBJECT_EOF) /* Only truncate is supported */
846 info = tsi2mdt_info(tsi);
847 la = &info->mti_attr.ma_attr;
848 /* standard truncate optimization: if file body is completely
849 * destroyed, don't send data back to the server. */
851 flags |= LDLM_FL_AST_DISCARD_DATA;
853 repbody->oa.o_oi = oa->o_oi;
854 repbody->oa.o_valid = OBD_MD_FLID;
856 srvlock = (exp_connect_flags(exp) & OBD_CONNECT_SRVLOCK) &&
857 oa->o_valid & OBD_MD_FLFLAGS &&
858 oa->o_flags & OBD_FL_SRVLOCK;
861 rc = tgt_mdt_data_lock(ns, &tsi->tsi_resid, &lh, LCK_PW,
867 CDEBUG(D_INODE, "calling punch for object "DFID", valid = %#llx"
868 ", start = %lld, end = %lld\n", PFID(&tsi->tsi_fid),
869 oa->o_valid, start, end);
871 mo = mdt_object_find(tsi->tsi_env, mdt, &tsi->tsi_fid);
873 GOTO(out_unlock, rc = PTR_ERR(mo));
875 if (!mdt_object_exists(mo))
876 GOTO(out_put, rc = -ENOENT);
878 /* Shouldn't happen on dirs */
879 if (S_ISDIR(lu_object_attr(&mo->mot_obj))) {
881 CERROR("%s: Truncate on dir "DFID": rc = %d\n",
882 exp->exp_obd->obd_name, PFID(&tsi->tsi_fid), rc);
886 mdt_dom_write_lock(mo);
887 dob = mdt_obj2dt(mo);
889 la_from_obdo(la, oa, OBD_MD_FLMTIME | OBD_MD_FLATIME | OBD_MD_FLCTIME);
891 la->la_valid |= LA_SIZE;
893 rc = mdt_object_punch(tsi->tsi_env, mdt->mdt_bottom, dob,
895 mdt_dom_write_unlock(mo);
899 mdt_dom_obj_lvb_update(tsi->tsi_env, mo, false);
900 mdt_io_counter_incr(tsi->tsi_exp, LPROC_MDT_IO_PUNCH,
904 lu_object_put(tsi->tsi_env, &mo->mot_obj);
907 tgt_extent_unlock(&lh, LCK_PW);
909 mdt_thread_info_fini(info);
914 * MDT glimpse for Data-on-MDT
916 * If there is write lock on client then function issues glimpse_ast to get
917 * an actual size from that client.
920 int mdt_do_glimpse(const struct lu_env *env, struct ldlm_namespace *ns,
921 struct ldlm_resource *res)
923 union ldlm_policy_data policy;
924 struct lustre_handle lockh;
926 struct ldlm_lock *lock;
927 struct ldlm_glimpse_work *gl_work;
928 struct list_head gl_list;
933 /* There can be only one write lock covering data, try to match it. */
934 policy.l_inodebits.bits = MDS_INODELOCK_DOM;
935 mode = ldlm_lock_match(ns, LDLM_FL_TEST_LOCK,
936 &res->lr_name, LDLM_IBITS, &policy,
939 /* There is no PW lock on this object; finished. */
943 lock = ldlm_handle2lock(&lockh);
948 * This check is for lock taken in mdt_reint_unlink() that does
949 * not have l_glimpse_ast set. So the logic is: if there is a lock
950 * with no l_glimpse_ast set, this object is being destroyed already.
951 * Hence, if you are grabbing DLM locks on the server, always set
952 * non-NULL glimpse_ast (e.g., ldlm_request.c::ldlm_glimpse_ast()).
954 if (lock->l_glimpse_ast == NULL) {
955 LDLM_DEBUG(lock, "no l_glimpse_ast");
956 GOTO(out, rc = -ENOENT);
959 OBD_SLAB_ALLOC_PTR_GFP(gl_work, ldlm_glimpse_work_kmem, GFP_ATOMIC);
961 GOTO(out, rc = -ENOMEM);
963 /* Populate the gl_work structure.
964 * Grab additional reference on the lock which will be released in
965 * ldlm_work_gl_ast_lock() */
966 gl_work->gl_lock = LDLM_LOCK_GET(lock);
967 /* The glimpse callback is sent to one single IO lock. As a result,
968 * the gl_work list is just composed of one element */
969 INIT_LIST_HEAD(&gl_list);
970 list_add_tail(&gl_work->gl_list, &gl_list);
971 /* There is actually no need for a glimpse descriptor when glimpsing
973 gl_work->gl_desc = NULL;
974 /* the ldlm_glimpse_work structure is allocated on the stack */
975 gl_work->gl_flags = LDLM_GL_WORK_SLAB_ALLOCATED;
977 ldlm_glimpse_locks(res, &gl_list); /* this will update the LVB */
979 /* If the list is not empty, we failed to glimpse a lock and
980 * must clean it up. Usually due to a race with unlink.*/
981 if (!list_empty(&gl_list)) {
982 LDLM_LOCK_RELEASE(lock);
983 OBD_SLAB_FREE_PTR(gl_work, ldlm_glimpse_work_kmem);
992 static void mdt_lvb2body(struct ldlm_resource *res, struct mdt_body *mb)
994 struct ost_lvb *res_lvb;
997 res_lvb = res->lr_lvb_data;
998 mb->mbo_dom_size = res_lvb->lvb_size;
999 mb->mbo_dom_blocks = res_lvb->lvb_blocks;
1000 mb->mbo_mtime = res_lvb->lvb_mtime;
1001 mb->mbo_ctime = res_lvb->lvb_ctime;
1002 mb->mbo_atime = res_lvb->lvb_atime;
1004 CDEBUG(D_DLMTRACE, "size %llu\n", res_lvb->lvb_size);
1006 mb->mbo_valid |= OBD_MD_FLATIME | OBD_MD_FLCTIME | OBD_MD_FLMTIME |
1012 * MDT glimpse for Data-on-MDT
1014 * This function is called when MDT get attributes for the DoM object.
1015 * If there is write lock on client then function issues glimpse_ast to get
1016 * an actual size from that client.
1018 int mdt_dom_object_size(const struct lu_env *env, struct mdt_device *mdt,
1019 const struct lu_fid *fid, struct mdt_body *mb,
1022 struct ldlm_res_id resid;
1023 struct ldlm_resource *res;
1028 fid_build_reg_res_name(fid, &resid);
1029 res = ldlm_resource_get(mdt->mdt_namespace, NULL, &resid,
1034 /* Update lvbo data if DoM lock returned or if LVB is not yet valid. */
1035 if (dom_lock || !mdt_dom_lvb_is_valid(res))
1036 mdt_dom_lvbo_update(env, res, NULL, NULL, false);
1038 mdt_lvb2body(res, mb);
1039 ldlm_resource_putref(res);
1044 * MDT DoM lock intent policy (glimpse)
1046 * Intent policy is called when lock has an intent, for DoM file that
1047 * means glimpse lock and policy fills Lock Value Block (LVB).
1049 * If already granted lock is found it will be placed in \a lockp and
1050 * returned back to caller function.
1052 * \param[in] tsi session info
1053 * \param[in,out] lockp pointer to the lock
1054 * \param[in] flags LDLM flags
1056 * \retval ELDLM_LOCK_REPLACED if already granted lock was found
1057 * and placed in \a lockp
1058 * \retval ELDLM_LOCK_ABORTED in other cases except error
1059 * \retval negative value on error
1061 int mdt_glimpse_enqueue(struct mdt_thread_info *mti, struct ldlm_namespace *ns,
1062 struct ldlm_lock **lockp, __u64 flags)
1064 struct ldlm_lock *lock = *lockp;
1065 struct ldlm_resource *res = lock->l_resource;
1066 ldlm_processing_policy policy;
1067 struct ldlm_reply *rep;
1068 struct mdt_body *mbo;
1073 policy = ldlm_get_processing_policy(res);
1074 LASSERT(policy != NULL);
1076 req_capsule_set_size(mti->mti_pill, &RMF_MDT_MD, RCL_SERVER, 0);
1077 req_capsule_set_size(mti->mti_pill, &RMF_ACL, RCL_SERVER, 0);
1078 rc = req_capsule_server_pack(mti->mti_pill);
1080 RETURN(err_serious(rc));
1082 rep = req_capsule_server_get(mti->mti_pill, &RMF_DLM_REP);
1086 mbo = req_capsule_server_get(mti->mti_pill, &RMF_MDT_BODY);
1091 /* Check if this is a resend case (MSG_RESENT is set on RPC) and a
1092 * lock was found by ldlm_handle_enqueue(); if so no need to grant
1094 if (flags & LDLM_FL_RESENT) {
1095 rc = LDLM_ITER_CONTINUE;
1097 __u64 tmpflags = LDLM_FL_BLOCK_NOWAIT;
1098 enum ldlm_error err;
1100 rc = policy(lock, &tmpflags, LDLM_PROCESS_RESCAN, &err, NULL);
1101 check_res_locked(res);
1105 /* The lock met with no resistance; we're finished. */
1106 if (rc == LDLM_ITER_CONTINUE) {
1107 GOTO(fill_mbo, rc = ELDLM_LOCK_REPLACED);
1108 } else if (flags & LDLM_FL_BLOCK_NOWAIT) {
1109 /* LDLM_FL_BLOCK_NOWAIT means it is for AGL. Do not send glimpse
1110 * callback for glimpse size. The real size user will trigger
1111 * the glimpse callback when necessary. */
1112 GOTO(fill_mbo, rc = ELDLM_LOCK_ABORTED);
1115 rc = mdt_do_glimpse(mti->mti_env, ns, res);
1116 if (rc == -ENOENT) {
1117 /* We are racing with unlink(); just return -ENOENT */
1118 rep->lock_policy_res2 = ptlrpc_status_hton(-ENOENT);
1120 } else if (rc == -EINVAL) {
1121 /* this is possible is client lock has been cancelled but
1122 * still exists on server. If that lock was found on server
1123 * as only conflicting lock then the client has already
1124 * size authority and glimpse is not needed. */
1125 CDEBUG(D_DLMTRACE, "Glimpse from the client owning lock\n");
1127 } else if (rc < 0) {
1130 rc = ELDLM_LOCK_ABORTED;
1132 /* LVB can be without valid data in case of DOM */
1133 if (!mdt_dom_lvb_is_valid(res))
1134 mdt_dom_lvbo_update(mti->mti_env, res, lock, NULL, false);
1135 mdt_lvb2body(res, mbo);
1139 int mdt_brw_enqueue(struct mdt_thread_info *mti, struct ldlm_namespace *ns,
1140 struct ldlm_lock **lockp, __u64 flags)
1142 struct tgt_session_info *tsi = tgt_ses_info(mti->mti_env);
1143 struct lu_fid *fid = &tsi->tsi_fid;
1144 struct ldlm_lock *lock = *lockp;
1145 struct ldlm_resource *res = lock->l_resource;
1146 struct ldlm_reply *rep;
1147 struct mdt_body *mbo;
1148 struct mdt_lock_handle *lhc = &mti->mti_lh[MDT_LH_RMT];
1149 struct mdt_object *mo;
1154 /* Get lock from request for possible resent case. */
1155 mdt_intent_fixup_resent(mti, *lockp, lhc, flags);
1156 req_capsule_set_size(mti->mti_pill, &RMF_MDT_MD, RCL_SERVER, 0);
1157 req_capsule_set_size(mti->mti_pill, &RMF_ACL, RCL_SERVER, 0);
1158 rc = req_capsule_server_pack(mti->mti_pill);
1160 RETURN(err_serious(rc));
1162 rep = req_capsule_server_get(mti->mti_pill, &RMF_DLM_REP);
1166 mbo = req_capsule_server_get(mti->mti_pill, &RMF_MDT_BODY);
1170 fid_extract_from_res_name(fid, &res->lr_name);
1171 mo = mdt_object_find(mti->mti_env, mti->mti_mdt, fid);
1172 if (unlikely(IS_ERR(mo)))
1173 RETURN(PTR_ERR(mo));
1175 if (!mdt_object_exists(mo))
1176 GOTO(out, rc = -ENOENT);
1178 if (mdt_object_remote(mo))
1179 GOTO(out, rc = -EPROTO);
1182 if (!lustre_handle_is_used(&lhc->mlh_reg_lh)) {
1183 mdt_lock_handle_init(lhc);
1184 mdt_lock_reg_init(lhc, (*lockp)->l_req_mode);
1185 /* This will block MDT thread but it should be fine until
1186 * client caches small amount of data for DoM, which should be
1187 * smaller than one BRW RPC and should be able to be
1188 * piggybacked by lock cancel RPC.
1189 * If the client could hold the lock too long, this code can be
1190 * revised to call mdt_object_lock_try(). And if fails, it will
1191 * return ELDLM_OK here and fall back into normal lock enqueue
1194 rc = mdt_object_lock(mti, mo, lhc, MDS_INODELOCK_DOM);
1199 if (!mdt_dom_lvb_is_valid(res)) {
1200 rc = mdt_dom_lvb_alloc(res);
1203 mdt_dom_disk_lvbo_update(mti->mti_env, mo, res, false);
1205 mdt_lvb2body(res, mbo);
1207 rep->lock_policy_res2 = clear_serious(rc);
1208 if (rep->lock_policy_res2) {
1209 lhc->mlh_reg_lh.cookie = 0ull;
1210 GOTO(out, rc = ELDLM_LOCK_ABORTED);
1213 rc = mdt_intent_lock_replace(mti, lockp, lhc, flags, rc);
1215 mdt_object_put(mti->mti_env, mo);
1219 void mdt_dom_discard_data(struct mdt_thread_info *info,
1220 const struct lu_fid *fid)
1222 struct mdt_device *mdt = info->mti_mdt;
1223 union ldlm_policy_data *policy = &info->mti_policy;
1224 struct ldlm_res_id *res_id = &info->mti_res_id;
1225 struct lustre_handle dom_lh;
1226 __u64 flags = LDLM_FL_AST_DISCARD_DATA;
1229 policy->l_inodebits.bits = MDS_INODELOCK_DOM;
1230 policy->l_inodebits.try_bits = 0;
1231 fid_build_reg_res_name(fid, res_id);
1233 /* Tell the clients that the object is gone now and that they should
1234 * throw away any cached pages. */
1235 rc = ldlm_cli_enqueue_local(info->mti_env, mdt->mdt_namespace, res_id,
1236 LDLM_IBITS, policy, LCK_PW, &flags,
1237 ldlm_blocking_ast, ldlm_completion_ast,
1238 NULL, NULL, 0, LVB_T_NONE, NULL, &dom_lh);
1240 /* We only care about the side-effects, just drop the lock. */
1242 ldlm_lock_decref_and_cancel(&dom_lh, LCK_PW);
1245 /* check if client has already DoM lock for given resource */
1246 bool mdt_dom_client_has_lock(struct mdt_thread_info *info,
1247 const struct lu_fid *fid)
1249 struct mdt_device *mdt = info->mti_mdt;
1250 union ldlm_policy_data *policy = &info->mti_policy;
1251 struct ldlm_res_id *res_id = &info->mti_res_id;
1252 struct lustre_handle lockh;
1253 enum ldlm_mode mode;
1254 struct ldlm_lock *lock;
1257 policy->l_inodebits.bits = MDS_INODELOCK_DOM;
1258 fid_build_reg_res_name(fid, res_id);
1260 mode = ldlm_lock_match(mdt->mdt_namespace, LDLM_FL_BLOCK_GRANTED |
1261 LDLM_FL_TEST_LOCK, res_id, LDLM_IBITS, policy,
1264 /* There is no other PW lock on this object; finished. */
1268 lock = ldlm_handle2lock(&lockh);
1272 /* check if lock from the same client */
1273 rc = (lock->l_export->exp_handle.h_cookie ==
1274 info->mti_exp->exp_handle.h_cookie);
1275 LDLM_LOCK_PUT(lock);
1280 * MDT request handler for OST_GETATTR RPC.
1282 * This is data-specific request to get object and layout versions under
1283 * IO lock. It is reliable only for Data-on-MDT files.
1285 * \param[in] tsi target session environment for this request
1287 * \retval 0 if successful
1288 * \retval negative value on error
1290 int mdt_data_version_get(struct tgt_session_info *tsi)
1292 struct mdt_thread_info *mti = mdt_th_info(tsi->tsi_env);
1293 struct mdt_device *mdt = mti->mti_mdt;
1294 struct mdt_body *repbody;
1295 struct mdt_object *mo = mti->mti_object;
1296 struct lov_comp_md_v1 *comp;
1297 struct lustre_handle lh = { 0 };
1300 enum ldlm_mode lock_mode = LCK_PR;
1306 req_capsule_set_size(tsi->tsi_pill, &RMF_MDT_MD, RCL_SERVER, 0);
1307 req_capsule_set_size(tsi->tsi_pill, &RMF_ACL, RCL_SERVER, 0);
1308 rc = req_capsule_server_pack(tsi->tsi_pill);
1309 if (unlikely(rc != 0))
1310 RETURN(err_serious(rc));
1312 repbody = req_capsule_server_get(tsi->tsi_pill, &RMF_MDT_BODY);
1313 if (repbody == NULL)
1316 srvlock = tsi->tsi_mdt_body->mbo_valid & OBD_MD_FLFLAGS &&
1317 tsi->tsi_mdt_body->mbo_flags & OBD_FL_SRVLOCK;
1320 if (unlikely(tsi->tsi_mdt_body->mbo_flags & OBD_FL_FLUSH))
1323 fid_build_reg_res_name(&tsi->tsi_fid, &tsi->tsi_resid);
1324 rc = tgt_mdt_data_lock(mdt->mdt_namespace, &tsi->tsi_resid,
1325 &lh, lock_mode, &flags);
1330 if (!mdt_object_exists(mo))
1331 GOTO(out, rc = -ENOENT);
1332 if (mdt_object_remote(mo))
1333 GOTO(out, rc = -EREMOTE);
1334 if (!S_ISREG(lu_object_attr(&mo->mot_obj)))
1335 GOTO(out, rc = -EBADF);
1337 /* Get version first */
1338 version = dt_version_get(tsi->tsi_env, mdt_obj2dt(mo));
1339 if (version && version != -EOPNOTSUPP) {
1340 repbody->mbo_valid |= OBD_MD_FLDATAVERSION;
1341 /* re-use mbo_ioepoch to transfer version */
1342 repbody->mbo_version = version;
1345 /* Read layout to get its version */
1346 rc = mdt_big_xattr_get(mti, mo, XATTR_NAME_LOV);
1347 if (rc == -ENODATA) /* File has no layout yet */
1352 comp = mti->mti_buf.lb_buf;
1353 if (le32_to_cpu(comp->lcm_magic) != LOV_MAGIC_COMP_V1) {
1354 CDEBUG(D_INFO, DFID" has no composite layout",
1355 PFID(&tsi->tsi_fid));
1356 GOTO(out, rc = -ESTALE);
1359 CDEBUG(D_INODE, DFID": layout version: %u\n",
1360 PFID(&tsi->tsi_fid), le32_to_cpu(comp->lcm_layout_gen));
1362 repbody->mbo_valid |= OBD_MD_LAYOUT_VERSION;
1363 /* re-use mbo_rdev for that */
1364 repbody->mbo_layout_gen = le32_to_cpu(comp->lcm_layout_gen);
1368 tgt_mdt_data_unlock(&lh, lock_mode);
1370 repbody->mbo_valid |= OBD_MD_FLFLAGS;
1371 repbody->mbo_flags = OBD_FL_FLUSH;
1375 /* read file data to the buffer */
1376 int mdt_dom_read_on_open(struct mdt_thread_info *mti, struct mdt_device *mdt,
1377 struct lustre_handle *lh)
1379 const struct lu_env *env = mti->mti_env;
1380 struct tgt_session_info *tsi = tgt_ses_info(env);
1381 struct req_capsule *pill = tsi->tsi_pill;
1382 const struct lu_fid *fid;
1383 struct ptlrpc_request *req = tgt_ses_req(tsi);
1384 struct mdt_body *mbo;
1385 struct dt_device *dt = mdt->mdt_bottom;
1386 struct dt_object *mo;
1388 struct niobuf_remote *rnb = NULL;
1389 struct niobuf_local *lnb;
1393 unsigned int len, copied = 0;
1394 int lnbs, nr_local, i;
1395 bool dom_lock = false;
1399 if (!req_capsule_field_present(pill, &RMF_NIOBUF_INLINE, RCL_SERVER)) {
1400 /* There is no reply buffers for this field, this means that
1401 * client has no support for data in reply.
1406 mbo = req_capsule_server_get(pill, &RMF_MDT_BODY);
1408 if (lustre_handle_is_used(lh)) {
1409 struct ldlm_lock *lock;
1411 lock = ldlm_handle2lock(lh);
1413 dom_lock = ldlm_has_dom(lock) && ldlm_has_layout(lock);
1414 LDLM_LOCK_PUT(lock);
1418 /* return data along with open only along with DoM lock */
1419 if (!dom_lock || !mdt->mdt_opts.mo_dom_read_open)
1422 if (!(mbo->mbo_valid & OBD_MD_DOM_SIZE))
1425 if (mbo->mbo_dom_size == 0)
1428 /* check the maximum size available in reply */
1430 req->rq_rqbd->rqbd_svcpt->scp_service->srv_max_reply_size;
1432 CDEBUG(D_INFO, "File size %llu, reply sizes %d/%d/%d\n",
1433 mbo->mbo_dom_size, max_reply_len, req->rq_reqmsg->lm_repsize,
1435 len = req->rq_reqmsg->lm_repsize - req->rq_replen;
1436 max_reply_len -= req->rq_replen;
1438 /* NB: at this moment we have the following sizes:
1439 * - req->rq_replen: used data in reply
1440 * - req->rq_reqmsg->lm_repsize: total allocated reply buffer at client
1441 * - max_reply_len: maximum reply size allowed by protocol
1443 * Ideal case when file size fits in allocated reply buffer,
1444 * that mean we can return whole data in reply. We can also fit more
1445 * data up to max_reply_size in total reply size, but this will cause
1446 * re-allocation on client and resend with larger buffer. This is still
1447 * faster than separate READ IO.
1448 * Third case if file is too big to fit even in maximum size, in that
1449 * case we return just tail to optimize possible append.
1451 * At the moment the following strategy is used:
1452 * 1) try to fit into the buffer we have
1453 * 2) respond with bigger buffer so client will re-allocate it and
1454 * resend (up to srv_max_reply_size value).
1455 * 3) return just file tail otherwise.
1457 if (mbo->mbo_dom_size <= len) {
1458 /* can fit whole data */
1459 len = mbo->mbo_dom_size;
1461 } else if (mbo->mbo_dom_size <= max_reply_len) {
1462 /* It is worth to make this tunable ON/OFF because this will
1463 * cause buffer re-allocation and resend
1465 len = mbo->mbo_dom_size;
1470 /* File tail offset must be aligned with larger page size
1471 * between client and server, so the maximum page size is
1472 * used here to align offset.
1474 * NB: DOM feature was introduced when server supports pagebits
1475 * already, so it should be always non-zero value. Report error
1476 * if it is not for some reason.
1478 if (!req->rq_export->exp_target_data.ted_pagebits) {
1479 CERROR("%s: client page bits are not saved on server\n",
1483 pgbits = max_t(int, PAGE_SHIFT,
1484 req->rq_export->exp_target_data.ted_pagebits);
1485 tail = mbo->mbo_dom_size % (1 << pgbits);
1487 /* no partial tail or tail can't fit in reply */
1488 if (tail == 0 || len < tail)
1492 offset = mbo->mbo_dom_size - len;
1494 LASSERT((offset % PAGE_SIZE) == 0);
1495 rc = req_capsule_server_grow(pill, &RMF_NIOBUF_INLINE,
1496 sizeof(*rnb) + len);
1498 /* failed to grow data buffer, just exit */
1499 GOTO(out, rc = -E2BIG);
1502 /* re-take MDT_BODY and NIOBUF_INLINE buffers after the buffer grow */
1503 mbo = req_capsule_server_get(pill, &RMF_MDT_BODY);
1504 fid = &mbo->mbo_fid1;
1505 if (!fid_is_sane(fid))
1506 GOTO(out, rc = -EINVAL);
1508 rnb = req_capsule_server_get(tsi->tsi_pill, &RMF_NIOBUF_INLINE);
1510 GOTO(out, rc = -EPROTO);
1512 buf = (char *)rnb + sizeof(*rnb);
1514 rnb->rnb_offset = offset;
1516 mo = dt_locate(env, dt, fid);
1518 GOTO(out_rnb, rc = PTR_ERR(mo));
1519 LASSERT(mo != NULL);
1521 dt_read_lock(env, mo, 0);
1522 if (!dt_object_exists(mo))
1523 GOTO(unlock, rc = -ENOENT);
1525 /* parse remote buffers to local buffers and prepare the latter */
1526 lnbs = (len >> PAGE_SHIFT) + 1;
1527 OBD_ALLOC(lnb, sizeof(*lnb) * lnbs);
1529 GOTO(unlock, rc = -ENOMEM);
1531 rc = dt_bufs_get(env, mo, rnb, lnb, 0);
1532 if (unlikely(rc < 0))
1534 LASSERT(rc <= lnbs);
1536 rc = dt_read_prep(env, mo, lnb, nr_local);
1539 /* copy data to the buffer finally */
1540 for (i = 0; i < nr_local; i++) {
1541 char *p = kmap(lnb[i].lnb_page);
1544 LASSERT(lnb[i].lnb_page_offset == 0);
1545 off = lnb[i].lnb_len & ~PAGE_MASK;
1547 memset(p + off, 0, PAGE_SIZE - off);
1549 memcpy(buf + (i << PAGE_SHIFT), p, lnb[i].lnb_len);
1550 kunmap(lnb[i].lnb_page);
1551 copied += lnb[i].lnb_len;
1554 CDEBUG(D_INFO, "Read %i (wanted %u) bytes from %llu\n", copied,
1557 CWARN("%s: read %i bytes for "DFID
1558 " but wanted %u, is size wrong?\n",
1559 tsi->tsi_exp->exp_obd->obd_name, copied,
1560 PFID(&tsi->tsi_fid), len);
1561 /* Ignore partially copied data */
1566 dt_bufs_put(env, mo, lnb, nr_local);
1568 OBD_FREE(lnb, sizeof(*lnb) * lnbs);
1570 dt_read_unlock(env, mo);
1571 lu_object_put(env, &mo->do_lu);
1573 rnb->rnb_len = copied;
1575 /* Don't fail OPEN request if read-on-open is failed, but drop
1576 * a message in log about the error.
1579 CDEBUG(D_INFO, "Read-on-open is failed, rc = %d", rc);