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 if (tgt_fmd_check(exp, mdt_object_fid(mo),
706 mdt_info_req(info)->rq_xid))
707 valid |= OBD_MD_FLATIME | OBD_MD_FLMTIME |
710 la_from_obdo(la, oa, valid);
712 rc = mdt_commitrw_write(env, exp, mdt, mo, la, objcount,
713 npages, lnb, oa->o_grant_used, old_rc);
715 obdo_from_la(oa, la, VALID_FLAGS | LA_GID | LA_UID);
717 obdo_from_la(oa, la, LA_GID | LA_UID);
719 mdt_dom_obj_lvb_update(env, mo, false);
720 /* don't report overquota flag if we failed before reaching
722 if (old_rc == 0 && (rc == 0 || rc == -EDQUOT)) {
723 /* return the overquota flags to client */
724 if (lnb[0].lnb_flags & OBD_BRW_OVER_USRQUOTA) {
725 if (oa->o_valid & OBD_MD_FLFLAGS)
726 oa->o_flags |= OBD_FL_NO_USRQUOTA;
728 oa->o_flags = OBD_FL_NO_USRQUOTA;
731 if (lnb[0].lnb_flags & OBD_BRW_OVER_GRPQUOTA) {
732 if (oa->o_valid & OBD_MD_FLFLAGS)
733 oa->o_flags |= OBD_FL_NO_GRPQUOTA;
735 oa->o_flags = OBD_FL_NO_GRPQUOTA;
738 if (lnb[0].lnb_flags & OBD_BRW_OVER_PRJQUOTA) {
739 if (oa->o_valid & OBD_MD_FLFLAGS)
740 oa->o_flags |= OBD_FL_NO_PRJQUOTA;
742 oa->o_flags = OBD_FL_NO_PRJQUOTA;
745 oa->o_valid |= OBD_MD_FLFLAGS | OBD_MD_FLUSRQUOTA |
746 OBD_MD_FLGRPQUOTA | OBD_MD_FLPRJQUOTA;
748 } else if (cmd == OBD_BRW_READ) {
749 /* If oa != NULL then mdt_preprw_read updated the inode
750 * atime and we should update the lvb so that other glimpses
751 * will also get the updated value. bug 5972 */
753 mdt_dom_obj_lvb_update(env, mo, true);
754 rc = mdt_commitrw_read(env, mdt, mo, objcount, npages, lnb);
760 /* this put is pair to object_get in ofd_preprw_write */
761 mdt_thread_info_fini(info);
765 int mdt_object_punch(const struct lu_env *env, struct dt_device *dt,
766 struct dt_object *dob, __u64 start, __u64 end,
774 /* we support truncate, not punch yet */
775 LASSERT(end == OBD_OBJECT_EOF);
777 if (!dt_object_exists(dob))
780 th = dt_trans_create(env, dt);
784 rc = dt_declare_attr_set(env, dob, la, th);
788 rc = dt_declare_punch(env, dob, start, OBD_OBJECT_EOF, th);
792 tgt_vbr_obj_set(env, dob);
793 rc = dt_trans_start(env, dt, th);
797 dt_write_lock(env, dob, 0);
798 rc = dt_punch(env, dob, start, OBD_OBJECT_EOF, th);
801 rc = dt_attr_set(env, dob, la, th);
805 dt_write_unlock(env, dob);
808 dt_trans_stop(env, dt, th);
812 int mdt_punch_hdl(struct tgt_session_info *tsi)
814 const struct obdo *oa = &tsi->tsi_ost_body->oa;
815 struct ost_body *repbody;
816 struct mdt_thread_info *info;
818 struct ldlm_namespace *ns = tsi->tsi_tgt->lut_obd->obd_namespace;
819 struct obd_export *exp = tsi->tsi_exp;
820 struct mdt_device *mdt = mdt_dev(exp->exp_obd->obd_lu_dev);
821 struct mdt_object *mo;
822 struct dt_object *dob;
824 struct lustre_handle lh = { 0, };
831 /* check that we do support OBD_CONNECT_TRUNCLOCK. */
832 CLASSERT(OST_CONNECT_SUPPORTED & OBD_CONNECT_TRUNCLOCK);
834 if ((oa->o_valid & (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS)) !=
835 (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS))
836 RETURN(err_serious(-EPROTO));
838 repbody = req_capsule_server_get(tsi->tsi_pill, &RMF_OST_BODY);
840 RETURN(err_serious(-ENOMEM));
842 /* punch start,end are passed in o_size,o_blocks throught wire */
846 if (end != OBD_OBJECT_EOF) /* Only truncate is supported */
849 info = tsi2mdt_info(tsi);
850 la = &info->mti_attr.ma_attr;
851 /* standard truncate optimization: if file body is completely
852 * destroyed, don't send data back to the server. */
854 flags |= LDLM_FL_AST_DISCARD_DATA;
856 repbody->oa.o_oi = oa->o_oi;
857 repbody->oa.o_valid = OBD_MD_FLID;
859 srvlock = (exp_connect_flags(exp) & OBD_CONNECT_SRVLOCK) &&
860 oa->o_valid & OBD_MD_FLFLAGS &&
861 oa->o_flags & OBD_FL_SRVLOCK;
864 rc = tgt_mdt_data_lock(ns, &tsi->tsi_resid, &lh, LCK_PW,
870 CDEBUG(D_INODE, "calling punch for object "DFID", valid = %#llx"
871 ", start = %lld, end = %lld\n", PFID(&tsi->tsi_fid),
872 oa->o_valid, start, end);
874 mo = mdt_object_find(tsi->tsi_env, mdt, &tsi->tsi_fid);
876 GOTO(out_unlock, rc = PTR_ERR(mo));
878 if (!mdt_object_exists(mo))
879 GOTO(out_put, rc = -ENOENT);
881 /* Shouldn't happen on dirs */
882 if (S_ISDIR(lu_object_attr(&mo->mot_obj))) {
884 CERROR("%s: Truncate on dir "DFID": rc = %d\n",
885 exp->exp_obd->obd_name, PFID(&tsi->tsi_fid), rc);
889 mdt_dom_write_lock(mo);
890 dob = mdt_obj2dt(mo);
892 la_from_obdo(la, oa, OBD_MD_FLMTIME | OBD_MD_FLATIME | OBD_MD_FLCTIME);
894 la->la_valid |= LA_SIZE;
896 /* MDT supports FMD for Data-on-MDT needs */
897 if (la->la_valid & (LA_ATIME | LA_MTIME | LA_CTIME))
898 tgt_fmd_update(tsi->tsi_exp, &tsi->tsi_fid,
899 tgt_ses_req(tsi)->rq_xid);
901 rc = mdt_object_punch(tsi->tsi_env, mdt->mdt_bottom, dob,
903 mdt_dom_write_unlock(mo);
907 mdt_dom_obj_lvb_update(tsi->tsi_env, mo, false);
908 mdt_io_counter_incr(tsi->tsi_exp, LPROC_MDT_IO_PUNCH,
912 lu_object_put(tsi->tsi_env, &mo->mot_obj);
915 tgt_extent_unlock(&lh, LCK_PW);
917 mdt_thread_info_fini(info);
922 * MDT glimpse for Data-on-MDT
924 * If there is write lock on client then function issues glimpse_ast to get
925 * an actual size from that client.
928 int mdt_do_glimpse(const struct lu_env *env, struct ldlm_namespace *ns,
929 struct ldlm_resource *res)
931 union ldlm_policy_data policy;
932 struct lustre_handle lockh;
934 struct ldlm_lock *lock;
935 struct ldlm_glimpse_work *gl_work;
936 struct list_head gl_list;
941 /* There can be only one write lock covering data, try to match it. */
942 policy.l_inodebits.bits = MDS_INODELOCK_DOM;
943 mode = ldlm_lock_match(ns, LDLM_FL_TEST_LOCK,
944 &res->lr_name, LDLM_IBITS, &policy,
947 /* There is no PW lock on this object; finished. */
951 lock = ldlm_handle2lock(&lockh);
956 * This check is for lock taken in mdt_reint_unlink() that does
957 * not have l_glimpse_ast set. So the logic is: if there is a lock
958 * with no l_glimpse_ast set, this object is being destroyed already.
959 * Hence, if you are grabbing DLM locks on the server, always set
960 * non-NULL glimpse_ast (e.g., ldlm_request.c::ldlm_glimpse_ast()).
962 if (lock->l_glimpse_ast == NULL) {
963 LDLM_DEBUG(lock, "no l_glimpse_ast");
964 GOTO(out, rc = -ENOENT);
967 OBD_SLAB_ALLOC_PTR_GFP(gl_work, ldlm_glimpse_work_kmem, GFP_ATOMIC);
969 GOTO(out, rc = -ENOMEM);
971 /* Populate the gl_work structure.
972 * Grab additional reference on the lock which will be released in
973 * ldlm_work_gl_ast_lock() */
974 gl_work->gl_lock = LDLM_LOCK_GET(lock);
975 /* The glimpse callback is sent to one single IO lock. As a result,
976 * the gl_work list is just composed of one element */
977 INIT_LIST_HEAD(&gl_list);
978 list_add_tail(&gl_work->gl_list, &gl_list);
979 /* There is actually no need for a glimpse descriptor when glimpsing
981 gl_work->gl_desc = NULL;
982 /* the ldlm_glimpse_work structure is allocated on the stack */
983 gl_work->gl_flags = LDLM_GL_WORK_SLAB_ALLOCATED;
985 ldlm_glimpse_locks(res, &gl_list); /* this will update the LVB */
987 /* If the list is not empty, we failed to glimpse a lock and
988 * must clean it up. Usually due to a race with unlink.*/
989 if (!list_empty(&gl_list)) {
990 LDLM_LOCK_RELEASE(lock);
991 OBD_SLAB_FREE_PTR(gl_work, ldlm_glimpse_work_kmem);
1000 static void mdt_lvb2body(struct ldlm_resource *res, struct mdt_body *mb)
1002 struct ost_lvb *res_lvb;
1005 res_lvb = res->lr_lvb_data;
1006 mb->mbo_dom_size = res_lvb->lvb_size;
1007 mb->mbo_dom_blocks = res_lvb->lvb_blocks;
1008 mb->mbo_mtime = res_lvb->lvb_mtime;
1009 mb->mbo_ctime = res_lvb->lvb_ctime;
1010 mb->mbo_atime = res_lvb->lvb_atime;
1012 CDEBUG(D_DLMTRACE, "size %llu\n", res_lvb->lvb_size);
1014 mb->mbo_valid |= OBD_MD_FLATIME | OBD_MD_FLCTIME | OBD_MD_FLMTIME |
1020 * MDT glimpse for Data-on-MDT
1022 * This function is called when MDT get attributes for the DoM object.
1023 * If there is write lock on client then function issues glimpse_ast to get
1024 * an actual size from that client.
1026 int mdt_dom_object_size(const struct lu_env *env, struct mdt_device *mdt,
1027 const struct lu_fid *fid, struct mdt_body *mb,
1030 struct ldlm_res_id resid;
1031 struct ldlm_resource *res;
1036 fid_build_reg_res_name(fid, &resid);
1037 res = ldlm_resource_get(mdt->mdt_namespace, NULL, &resid,
1042 /* Update lvbo data if DoM lock returned or if LVB is not yet valid. */
1043 if (dom_lock || !mdt_dom_lvb_is_valid(res))
1044 mdt_dom_lvbo_update(env, res, NULL, NULL, false);
1046 mdt_lvb2body(res, mb);
1047 ldlm_resource_putref(res);
1052 * MDT DoM lock intent policy (glimpse)
1054 * Intent policy is called when lock has an intent, for DoM file that
1055 * means glimpse lock and policy fills Lock Value Block (LVB).
1057 * If already granted lock is found it will be placed in \a lockp and
1058 * returned back to caller function.
1060 * \param[in] tsi session info
1061 * \param[in,out] lockp pointer to the lock
1062 * \param[in] flags LDLM flags
1064 * \retval ELDLM_LOCK_REPLACED if already granted lock was found
1065 * and placed in \a lockp
1066 * \retval ELDLM_LOCK_ABORTED in other cases except error
1067 * \retval negative value on error
1069 int mdt_glimpse_enqueue(struct mdt_thread_info *mti, struct ldlm_namespace *ns,
1070 struct ldlm_lock **lockp, __u64 flags)
1072 struct ldlm_lock *lock = *lockp;
1073 struct ldlm_resource *res = lock->l_resource;
1074 ldlm_processing_policy policy;
1075 struct ldlm_reply *rep;
1076 struct mdt_body *mbo;
1081 policy = ldlm_get_processing_policy(res);
1082 LASSERT(policy != NULL);
1084 req_capsule_set_size(mti->mti_pill, &RMF_MDT_MD, RCL_SERVER, 0);
1085 req_capsule_set_size(mti->mti_pill, &RMF_ACL, RCL_SERVER, 0);
1086 rc = req_capsule_server_pack(mti->mti_pill);
1088 RETURN(err_serious(rc));
1090 rep = req_capsule_server_get(mti->mti_pill, &RMF_DLM_REP);
1094 mbo = req_capsule_server_get(mti->mti_pill, &RMF_MDT_BODY);
1099 /* Check if this is a resend case (MSG_RESENT is set on RPC) and a
1100 * lock was found by ldlm_handle_enqueue(); if so no need to grant
1102 if (flags & LDLM_FL_RESENT) {
1103 rc = LDLM_ITER_CONTINUE;
1105 __u64 tmpflags = LDLM_FL_BLOCK_NOWAIT;
1106 enum ldlm_error err;
1108 rc = policy(lock, &tmpflags, LDLM_PROCESS_RESCAN, &err, NULL);
1109 check_res_locked(res);
1113 /* The lock met with no resistance; we're finished. */
1114 if (rc == LDLM_ITER_CONTINUE) {
1115 GOTO(fill_mbo, rc = ELDLM_LOCK_REPLACED);
1116 } else if (flags & LDLM_FL_BLOCK_NOWAIT) {
1117 /* LDLM_FL_BLOCK_NOWAIT means it is for AGL. Do not send glimpse
1118 * callback for glimpse size. The real size user will trigger
1119 * the glimpse callback when necessary. */
1120 GOTO(fill_mbo, rc = ELDLM_LOCK_ABORTED);
1123 rc = mdt_do_glimpse(mti->mti_env, ns, res);
1124 if (rc == -ENOENT) {
1125 /* We are racing with unlink(); just return -ENOENT */
1126 rep->lock_policy_res2 = ptlrpc_status_hton(-ENOENT);
1128 } else if (rc == -EINVAL) {
1129 /* this is possible is client lock has been cancelled but
1130 * still exists on server. If that lock was found on server
1131 * as only conflicting lock then the client has already
1132 * size authority and glimpse is not needed. */
1133 CDEBUG(D_DLMTRACE, "Glimpse from the client owning lock\n");
1135 } else if (rc < 0) {
1138 rc = ELDLM_LOCK_ABORTED;
1140 /* LVB can be without valid data in case of DOM */
1141 if (!mdt_dom_lvb_is_valid(res))
1142 mdt_dom_lvbo_update(mti->mti_env, res, lock, NULL, false);
1143 mdt_lvb2body(res, mbo);
1147 int mdt_brw_enqueue(struct mdt_thread_info *mti, struct ldlm_namespace *ns,
1148 struct ldlm_lock **lockp, __u64 flags)
1150 struct tgt_session_info *tsi = tgt_ses_info(mti->mti_env);
1151 struct lu_fid *fid = &tsi->tsi_fid;
1152 struct ldlm_lock *lock = *lockp;
1153 struct ldlm_resource *res = lock->l_resource;
1154 struct ldlm_reply *rep;
1155 struct mdt_body *mbo;
1156 struct mdt_lock_handle *lhc = &mti->mti_lh[MDT_LH_RMT];
1157 struct mdt_object *mo;
1162 /* Get lock from request for possible resent case. */
1163 mdt_intent_fixup_resent(mti, *lockp, lhc, flags);
1164 req_capsule_set_size(mti->mti_pill, &RMF_MDT_MD, RCL_SERVER, 0);
1165 req_capsule_set_size(mti->mti_pill, &RMF_ACL, RCL_SERVER, 0);
1166 rc = req_capsule_server_pack(mti->mti_pill);
1168 RETURN(err_serious(rc));
1170 rep = req_capsule_server_get(mti->mti_pill, &RMF_DLM_REP);
1174 mbo = req_capsule_server_get(mti->mti_pill, &RMF_MDT_BODY);
1178 fid_extract_from_res_name(fid, &res->lr_name);
1179 mo = mdt_object_find(mti->mti_env, mti->mti_mdt, fid);
1180 if (unlikely(IS_ERR(mo)))
1181 RETURN(PTR_ERR(mo));
1183 if (!mdt_object_exists(mo))
1184 GOTO(out, rc = -ENOENT);
1186 if (mdt_object_remote(mo))
1187 GOTO(out, rc = -EPROTO);
1190 if (!lustre_handle_is_used(&lhc->mlh_reg_lh)) {
1191 mdt_lock_handle_init(lhc);
1192 mdt_lock_reg_init(lhc, (*lockp)->l_req_mode);
1193 /* This will block MDT thread but it should be fine until
1194 * client caches small amount of data for DoM, which should be
1195 * smaller than one BRW RPC and should be able to be
1196 * piggybacked by lock cancel RPC.
1197 * If the client could hold the lock too long, this code can be
1198 * revised to call mdt_object_lock_try(). And if fails, it will
1199 * return ELDLM_OK here and fall back into normal lock enqueue
1202 rc = mdt_object_lock(mti, mo, lhc, MDS_INODELOCK_DOM);
1207 if (!mdt_dom_lvb_is_valid(res)) {
1208 rc = mdt_dom_lvb_alloc(res);
1211 mdt_dom_disk_lvbo_update(mti->mti_env, mo, res, false);
1213 mdt_lvb2body(res, mbo);
1215 rep->lock_policy_res2 = clear_serious(rc);
1216 if (rep->lock_policy_res2) {
1217 lhc->mlh_reg_lh.cookie = 0ull;
1218 GOTO(out, rc = ELDLM_LOCK_ABORTED);
1221 rc = mdt_intent_lock_replace(mti, lockp, lhc, flags, rc);
1223 mdt_object_put(mti->mti_env, mo);
1227 void mdt_dom_discard_data(struct mdt_thread_info *info,
1228 const struct lu_fid *fid)
1230 struct mdt_device *mdt = info->mti_mdt;
1231 union ldlm_policy_data *policy = &info->mti_policy;
1232 struct ldlm_res_id *res_id = &info->mti_res_id;
1233 struct lustre_handle dom_lh;
1234 __u64 flags = LDLM_FL_AST_DISCARD_DATA;
1237 policy->l_inodebits.bits = MDS_INODELOCK_DOM;
1238 policy->l_inodebits.try_bits = 0;
1239 fid_build_reg_res_name(fid, res_id);
1241 /* Tell the clients that the object is gone now and that they should
1242 * throw away any cached pages. */
1243 rc = ldlm_cli_enqueue_local(info->mti_env, mdt->mdt_namespace, res_id,
1244 LDLM_IBITS, policy, LCK_PW, &flags,
1245 ldlm_blocking_ast, ldlm_completion_ast,
1246 NULL, NULL, 0, LVB_T_NONE, NULL, &dom_lh);
1248 /* We only care about the side-effects, just drop the lock. */
1250 ldlm_lock_decref_and_cancel(&dom_lh, LCK_PW);
1253 /* check if client has already DoM lock for given resource */
1254 bool mdt_dom_client_has_lock(struct mdt_thread_info *info,
1255 const struct lu_fid *fid)
1257 struct mdt_device *mdt = info->mti_mdt;
1258 union ldlm_policy_data *policy = &info->mti_policy;
1259 struct ldlm_res_id *res_id = &info->mti_res_id;
1260 struct lustre_handle lockh;
1261 enum ldlm_mode mode;
1262 struct ldlm_lock *lock;
1265 policy->l_inodebits.bits = MDS_INODELOCK_DOM;
1266 fid_build_reg_res_name(fid, res_id);
1268 mode = ldlm_lock_match(mdt->mdt_namespace, LDLM_FL_BLOCK_GRANTED |
1269 LDLM_FL_TEST_LOCK, res_id, LDLM_IBITS, policy,
1272 /* There is no other PW lock on this object; finished. */
1276 lock = ldlm_handle2lock(&lockh);
1280 /* check if lock from the same client */
1281 rc = (lock->l_export->exp_handle.h_cookie ==
1282 info->mti_exp->exp_handle.h_cookie);
1283 LDLM_LOCK_PUT(lock);
1288 * MDT request handler for OST_GETATTR RPC.
1290 * This is data-specific request to get object and layout versions under
1291 * IO lock. It is reliable only for Data-on-MDT files.
1293 * \param[in] tsi target session environment for this request
1295 * \retval 0 if successful
1296 * \retval negative value on error
1298 int mdt_data_version_get(struct tgt_session_info *tsi)
1300 struct mdt_thread_info *mti = mdt_th_info(tsi->tsi_env);
1301 struct mdt_device *mdt = mti->mti_mdt;
1302 struct mdt_body *repbody;
1303 struct mdt_object *mo = mti->mti_object;
1304 struct lov_comp_md_v1 *comp;
1305 struct lustre_handle lh = { 0 };
1308 enum ldlm_mode lock_mode = LCK_PR;
1314 req_capsule_set_size(tsi->tsi_pill, &RMF_MDT_MD, RCL_SERVER, 0);
1315 req_capsule_set_size(tsi->tsi_pill, &RMF_ACL, RCL_SERVER, 0);
1316 rc = req_capsule_server_pack(tsi->tsi_pill);
1317 if (unlikely(rc != 0))
1318 RETURN(err_serious(rc));
1320 repbody = req_capsule_server_get(tsi->tsi_pill, &RMF_MDT_BODY);
1321 if (repbody == NULL)
1324 srvlock = tsi->tsi_mdt_body->mbo_valid & OBD_MD_FLFLAGS &&
1325 tsi->tsi_mdt_body->mbo_flags & OBD_FL_SRVLOCK;
1328 if (unlikely(tsi->tsi_mdt_body->mbo_flags & OBD_FL_FLUSH))
1331 fid_build_reg_res_name(&tsi->tsi_fid, &tsi->tsi_resid);
1332 rc = tgt_mdt_data_lock(mdt->mdt_namespace, &tsi->tsi_resid,
1333 &lh, lock_mode, &flags);
1338 if (!mdt_object_exists(mo))
1339 GOTO(out, rc = -ENOENT);
1340 if (mdt_object_remote(mo))
1341 GOTO(out, rc = -EREMOTE);
1342 if (!S_ISREG(lu_object_attr(&mo->mot_obj)))
1343 GOTO(out, rc = -EBADF);
1345 /* Get version first */
1346 version = dt_version_get(tsi->tsi_env, mdt_obj2dt(mo));
1347 if (version && version != -EOPNOTSUPP) {
1348 repbody->mbo_valid |= OBD_MD_FLDATAVERSION;
1349 /* re-use mbo_ioepoch to transfer version */
1350 repbody->mbo_version = version;
1353 /* Read layout to get its version */
1354 rc = mdt_big_xattr_get(mti, mo, XATTR_NAME_LOV);
1355 if (rc == -ENODATA) /* File has no layout yet */
1360 comp = mti->mti_buf.lb_buf;
1361 if (le32_to_cpu(comp->lcm_magic) != LOV_MAGIC_COMP_V1) {
1362 CDEBUG(D_INFO, DFID" has no composite layout",
1363 PFID(&tsi->tsi_fid));
1364 GOTO(out, rc = -ESTALE);
1367 CDEBUG(D_INODE, DFID": layout version: %u\n",
1368 PFID(&tsi->tsi_fid), le32_to_cpu(comp->lcm_layout_gen));
1370 repbody->mbo_valid |= OBD_MD_LAYOUT_VERSION;
1371 /* re-use mbo_rdev for that */
1372 repbody->mbo_layout_gen = le32_to_cpu(comp->lcm_layout_gen);
1376 tgt_mdt_data_unlock(&lh, lock_mode);
1378 repbody->mbo_valid |= OBD_MD_FLFLAGS;
1379 repbody->mbo_flags = OBD_FL_FLUSH;
1383 /* read file data to the buffer */
1384 int mdt_dom_read_on_open(struct mdt_thread_info *mti, struct mdt_device *mdt,
1385 struct lustre_handle *lh)
1387 const struct lu_env *env = mti->mti_env;
1388 struct tgt_session_info *tsi = tgt_ses_info(env);
1389 struct req_capsule *pill = tsi->tsi_pill;
1390 const struct lu_fid *fid;
1391 struct ptlrpc_request *req = tgt_ses_req(tsi);
1392 struct mdt_body *mbo;
1393 struct dt_device *dt = mdt->mdt_bottom;
1394 struct dt_object *mo;
1396 struct niobuf_remote *rnb = NULL;
1397 struct niobuf_local *lnb;
1400 unsigned int len, copied = 0;
1401 int lnbs, nr_local, i;
1402 bool dom_lock = false;
1406 if (!req_capsule_field_present(pill, &RMF_NIOBUF_INLINE, RCL_SERVER)) {
1407 /* There is no reply buffers for this field, this means that
1408 * client has no support for data in reply.
1413 mbo = req_capsule_server_get(pill, &RMF_MDT_BODY);
1415 if (lustre_handle_is_used(lh)) {
1416 struct ldlm_lock *lock;
1418 lock = ldlm_handle2lock(lh);
1420 dom_lock = ldlm_has_dom(lock) && ldlm_has_layout(lock);
1421 LDLM_LOCK_PUT(lock);
1425 /* return data along with open only along with DoM lock */
1426 if (!dom_lock || !mdt->mdt_opts.mo_dom_read_open)
1429 if (!(mbo->mbo_valid & OBD_MD_DOM_SIZE))
1432 if (mbo->mbo_dom_size == 0)
1435 CDEBUG(D_INFO, "File size %llu, reply sizes %d/%d\n",
1436 mbo->mbo_dom_size, req->rq_reqmsg->lm_repsize, req->rq_replen);
1437 len = req->rq_reqmsg->lm_repsize - req->rq_replen;
1439 /* NB: at this moment we have the following sizes:
1440 * - req->rq_replen: used data in reply
1441 * - req->rq_reqmsg->lm_repsize: total allocated reply buffer at client
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) return just file tail otherwise.
1455 if (mbo->mbo_dom_size <= len) {
1456 /* can fit whole data */
1457 len = mbo->mbo_dom_size;
1462 /* File tail offset must be aligned with larger page size
1463 * between client and server, so the maximum page size is
1464 * used here to align offset.
1466 * NB: DOM feature was introduced when server supports pagebits
1467 * already, so it should be always non-zero value. Report error
1468 * if it is not for some reason.
1470 if (!req->rq_export->exp_target_data.ted_pagebits) {
1471 CERROR("%s: client page bits are not saved on server\n",
1475 pgbits = max_t(int, PAGE_SHIFT,
1476 req->rq_export->exp_target_data.ted_pagebits);
1477 tail = mbo->mbo_dom_size % (1 << pgbits);
1479 /* no partial tail or tail can't fit in reply */
1480 if (tail == 0 || len < tail)
1484 offset = mbo->mbo_dom_size - len;
1486 LASSERT((offset % PAGE_SIZE) == 0);
1487 rc = req_capsule_server_grow(pill, &RMF_NIOBUF_INLINE,
1488 sizeof(*rnb) + len);
1490 /* failed to grow data buffer, just exit */
1491 GOTO(out, rc = -E2BIG);
1494 /* re-take MDT_BODY and NIOBUF_INLINE buffers after the buffer grow */
1495 mbo = req_capsule_server_get(pill, &RMF_MDT_BODY);
1496 fid = &mbo->mbo_fid1;
1497 if (!fid_is_sane(fid))
1498 GOTO(out, rc = -EINVAL);
1500 rnb = req_capsule_server_get(tsi->tsi_pill, &RMF_NIOBUF_INLINE);
1502 GOTO(out, rc = -EPROTO);
1504 buf = (char *)rnb + sizeof(*rnb);
1506 rnb->rnb_offset = offset;
1508 mo = dt_locate(env, dt, fid);
1510 GOTO(out_rnb, rc = PTR_ERR(mo));
1511 LASSERT(mo != NULL);
1513 dt_read_lock(env, mo, 0);
1514 if (!dt_object_exists(mo))
1515 GOTO(unlock, rc = -ENOENT);
1517 /* parse remote buffers to local buffers and prepare the latter */
1518 lnbs = (len >> PAGE_SHIFT) + 1;
1519 OBD_ALLOC(lnb, sizeof(*lnb) * lnbs);
1521 GOTO(unlock, rc = -ENOMEM);
1523 rc = dt_bufs_get(env, mo, rnb, lnb, 0);
1524 if (unlikely(rc < 0))
1526 LASSERT(rc <= lnbs);
1528 rc = dt_read_prep(env, mo, lnb, nr_local);
1531 /* copy data to the buffer finally */
1532 for (i = 0; i < nr_local; i++) {
1533 char *p = kmap(lnb[i].lnb_page);
1536 LASSERT(lnb[i].lnb_page_offset == 0);
1537 off = lnb[i].lnb_len & ~PAGE_MASK;
1539 memset(p + off, 0, PAGE_SIZE - off);
1541 memcpy(buf + (i << PAGE_SHIFT), p, lnb[i].lnb_len);
1542 kunmap(lnb[i].lnb_page);
1543 copied += lnb[i].lnb_len;
1546 CDEBUG(D_INFO, "Read %i (wanted %u) bytes from %llu\n", copied,
1549 CWARN("%s: read %i bytes for "DFID
1550 " but wanted %u, is size wrong?\n",
1551 tsi->tsi_exp->exp_obd->obd_name, copied,
1552 PFID(&tsi->tsi_fid), len);
1553 /* Ignore partially copied data */
1558 dt_bufs_put(env, mo, lnb, nr_local);
1560 OBD_FREE(lnb, sizeof(*lnb) * lnbs);
1562 dt_read_unlock(env, mo);
1563 lu_object_put(env, &mo->do_lu);
1565 rnb->rnb_len = copied;
1567 /* Don't fail OPEN request if read-on-open is failed, but drop
1568 * a message in log about the error.
1571 CDEBUG(D_INFO, "Read-on-open is failed, rc = %d", rc);