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_dom_read_lock(struct mdt_object *mo)
40 down_read(&mo->mot_dom_sem);
43 static inline void mdt_dom_read_unlock(struct mdt_object *mo)
45 up_read(&mo->mot_dom_sem);
48 static inline void mdt_dom_write_lock(struct mdt_object *mo)
50 down_write(&mo->mot_dom_sem);
53 static inline void mdt_dom_write_unlock(struct mdt_object *mo)
55 up_write(&mo->mot_dom_sem);
58 static void mdt_dom_resource_prolong(struct ldlm_prolong_args *arg)
60 struct ldlm_resource *res;
61 struct ldlm_lock *lock;
65 res = ldlm_resource_get(arg->lpa_export->exp_obd->obd_namespace, NULL,
66 &arg->lpa_resid, LDLM_EXTENT, 0);
69 "Failed to get resource for resid %llu/%llu\n",
70 arg->lpa_resid.name[0], arg->lpa_resid.name[1]);
75 list_for_each_entry(lock, &res->lr_granted, l_res_link) {
76 if (ldlm_has_dom(lock)) {
77 LDLM_DEBUG(lock, "DOM lock to prolong ");
78 ldlm_lock_prolong_one(lock, arg);
83 ldlm_resource_putref(res);
88 static void mdt_prolong_dom_lock(struct tgt_session_info *tsi,
89 struct ldlm_prolong_args *data)
91 struct obdo *oa = &tsi->tsi_ost_body->oa;
92 struct ldlm_lock *lock;
96 data->lpa_timeout = prolong_timeout(tgt_ses_req(tsi));
97 data->lpa_export = tsi->tsi_exp;
98 data->lpa_resid = tsi->tsi_resid;
100 CDEBUG(D_RPCTRACE, "Prolong DOM lock for req %p with x%llu\n",
101 tgt_ses_req(tsi), tgt_ses_req(tsi)->rq_xid);
103 if (oa->o_valid & OBD_MD_FLHANDLE) {
104 /* mostly a request should be covered by only one lock, try
106 lock = ldlm_handle2lock(&oa->o_handle);
108 LASSERT(lock->l_export == data->lpa_export);
109 ldlm_lock_prolong_one(lock, data);
110 lock->l_last_used = ktime_get();
112 if (data->lpa_locks_cnt > 0)
116 mdt_dom_resource_prolong(data);
120 static int mdt_rw_hpreq_lock_match(struct ptlrpc_request *req,
121 struct ldlm_lock *lock)
123 struct obd_ioobj *ioo;
125 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
129 if (!(lock->l_policy_data.l_inodebits.bits & MDS_INODELOCK_DOM))
132 ioo = req_capsule_client_get(&req->rq_pill, &RMF_OBD_IOOBJ);
133 LASSERT(ioo != NULL);
135 LASSERT(lock->l_resource != NULL);
136 if (!fid_res_name_eq(&ioo->ioo_oid.oi_fid, &lock->l_resource->lr_name))
139 /* a bulk write can only hold a reference on a PW extent lock. */
142 /* whereas a bulk read can be protected by either a PR or PW
146 if (!(lock->l_granted_mode & mode))
152 static int mdt_rw_hpreq_check(struct ptlrpc_request *req)
154 struct tgt_session_info *tsi;
155 struct obd_ioobj *ioo;
156 struct niobuf_remote *rnb;
158 struct ldlm_prolong_args pa = { 0 };
162 /* Don't use tgt_ses_info() to get session info, because lock_match()
163 * can be called while request has no processing thread yet. */
164 tsi = lu_context_key_get(&req->rq_session, &tgt_session_key);
167 * Use LASSERT below because malformed RPCs should have
168 * been filtered out in tgt_hpreq_handler().
170 opc = lustre_msg_get_opc(req->rq_reqmsg);
171 LASSERT(opc == OST_READ || opc == OST_WRITE);
173 ioo = req_capsule_client_get(&req->rq_pill, &RMF_OBD_IOOBJ);
174 LASSERT(ioo != NULL);
176 rnb = req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE);
177 LASSERT(rnb != NULL);
178 LASSERT(!(rnb->rnb_flags & OBD_BRW_SRVLOCK));
180 pa.lpa_mode = LCK_PW;
182 pa.lpa_mode |= LCK_PR;
184 DEBUG_REQ(D_RPCTRACE, req, "%s %s: refresh rw locks for " DFID,
185 tgt_name(tsi->tsi_tgt), current->comm, PFID(&tsi->tsi_fid));
187 mdt_prolong_dom_lock(tsi, &pa);
189 if (pa.lpa_blocks_cnt > 0) {
191 "%s: refreshed %u locks timeout for req %p\n",
192 tgt_name(tsi->tsi_tgt), pa.lpa_blocks_cnt, req);
196 RETURN(pa.lpa_locks_cnt > 0 ? 0 : -ESTALE);
199 static void mdt_rw_hpreq_fini(struct ptlrpc_request *req)
201 mdt_rw_hpreq_check(req);
204 static struct ptlrpc_hpreq_ops mdt_hpreq_rw = {
205 .hpreq_lock_match = mdt_rw_hpreq_lock_match,
206 .hpreq_check = mdt_rw_hpreq_check,
207 .hpreq_fini = mdt_rw_hpreq_fini
211 * Assign high priority operations to an IO request.
213 * Check if the incoming request is a candidate for
214 * high-priority processing. If it is, assign it a high
215 * priority operations table.
217 * \param[in] tsi target session environment for this request
219 void mdt_hp_brw(struct tgt_session_info *tsi)
221 struct niobuf_remote *rnb;
222 struct obd_ioobj *ioo;
226 ioo = req_capsule_client_get(tsi->tsi_pill, &RMF_OBD_IOOBJ);
227 LASSERT(ioo != NULL); /* must exist after request preprocessing */
228 if (ioo->ioo_bufcnt > 0) {
229 rnb = req_capsule_client_get(tsi->tsi_pill, &RMF_NIOBUF_REMOTE);
230 LASSERT(rnb != NULL); /* must exist after preprocessing */
232 /* no high priority if server lock is needed */
233 if (rnb->rnb_flags & OBD_BRW_SRVLOCK ||
234 (lustre_msg_get_flags(tgt_ses_req(tsi)->rq_reqmsg) &
238 tgt_ses_req(tsi)->rq_ops = &mdt_hpreq_rw;
241 static int mdt_punch_hpreq_lock_match(struct ptlrpc_request *req,
242 struct ldlm_lock *lock)
244 struct tgt_session_info *tsi;
249 /* Don't use tgt_ses_info() to get session info, because lock_match()
250 * can be called while request has no processing thread yet. */
251 tsi = lu_context_key_get(&req->rq_session, &tgt_session_key);
254 * Use LASSERT below because malformed RPCs should have
255 * been filtered out in tgt_hpreq_handler().
257 LASSERT(tsi->tsi_ost_body != NULL);
258 if (tsi->tsi_ost_body->oa.o_valid & OBD_MD_FLHANDLE &&
259 tsi->tsi_ost_body->oa.o_handle.cookie == lock->l_handle.h_cookie)
262 oa = &tsi->tsi_ost_body->oa;
264 LASSERT(lock->l_resource != NULL);
265 if (!fid_res_name_eq(&oa->o_oi.oi_fid, &lock->l_resource->lr_name))
268 if (!(lock->l_granted_mode & LCK_PW))
275 * Implementation of ptlrpc_hpreq_ops::hpreq_lock_check for OST_PUNCH request.
277 * High-priority queue request check for whether the given punch request
278 * (\a req) is blocking an LDLM lock cancel. Also checks whether the request is
279 * covered by an LDLM lock.
283 * \param[in] req the incoming request
285 * \retval 1 if \a req is blocking an LDLM lock cancel
286 * \retval 0 if it is not
287 * \retval -ESTALE if lock is not found
289 static int mdt_punch_hpreq_check(struct ptlrpc_request *req)
291 struct tgt_session_info *tsi;
293 struct ldlm_prolong_args pa = { 0 };
297 /* Don't use tgt_ses_info() to get session info, because lock_match()
298 * can be called while request has no processing thread yet. */
299 tsi = lu_context_key_get(&req->rq_session, &tgt_session_key);
300 LASSERT(tsi != NULL);
301 oa = &tsi->tsi_ost_body->oa;
303 LASSERT(!(oa->o_valid & OBD_MD_FLFLAGS &&
304 oa->o_flags & OBD_FL_SRVLOCK));
306 pa.lpa_mode = LCK_PW;
308 CDEBUG(D_DLMTRACE, "%s: refresh DOM lock for "DFID"\n",
309 tgt_name(tsi->tsi_tgt), PFID(&tsi->tsi_fid));
311 mdt_prolong_dom_lock(tsi, &pa);
314 if (pa.lpa_blocks_cnt > 0) {
316 "%s: refreshed %u locks timeout for req %p.\n",
317 tgt_name(tsi->tsi_tgt), pa.lpa_blocks_cnt, req);
321 RETURN(pa.lpa_locks_cnt > 0 ? 0 : -ESTALE);
325 * Implementation of ptlrpc_hpreq_ops::hpreq_lock_fini for OST_PUNCH request.
327 * Called after the request has been handled. It refreshes lock timeout again
328 * so that client has more time to send lock cancel RPC.
330 * \param[in] req request which is being processed.
332 static void mdt_punch_hpreq_fini(struct ptlrpc_request *req)
334 mdt_punch_hpreq_check(req);
337 static struct ptlrpc_hpreq_ops mdt_hpreq_punch = {
338 .hpreq_lock_match = mdt_punch_hpreq_lock_match,
339 .hpreq_check = mdt_punch_hpreq_check,
340 .hpreq_fini = mdt_punch_hpreq_fini
343 void mdt_hp_punch(struct tgt_session_info *tsi)
345 LASSERT(tsi->tsi_ost_body != NULL); /* must exists if we are here */
346 /* no high-priority if server lock is needed */
347 if ((tsi->tsi_ost_body->oa.o_valid & OBD_MD_FLFLAGS &&
348 tsi->tsi_ost_body->oa.o_flags & OBD_FL_SRVLOCK) ||
349 tgt_conn_flags(tsi) & OBD_CONNECT_MDS ||
350 lustre_msg_get_flags(tgt_ses_req(tsi)->rq_reqmsg) & MSG_REPLAY)
352 tgt_ses_req(tsi)->rq_ops = &mdt_hpreq_punch;
355 static int mdt_preprw_read(const struct lu_env *env, struct obd_export *exp,
356 struct mdt_device *mdt, struct mdt_object *mo,
357 struct lu_attr *la, int niocount,
358 struct niobuf_remote *rnb, int *nr_local,
359 struct niobuf_local *lnb, char *jobid)
361 struct tgt_session_info *tsi = tgt_ses_info(env);
362 struct ptlrpc_request *req = tgt_ses_req(tsi);
363 struct dt_object *dob;
364 int i, j, rc, tot_bytes = 0;
365 int maxlnb = *nr_local;
370 mdt_dom_read_lock(mo);
372 /* the only valid case when READ can find object is missing or stale
373 * when export is just evicted and open files are closed forcefully
374 * on server while client's READ can be in progress.
375 * This should not happen on healthy export, object can't be missing
376 * or dying because both states means it was finally destroyed.
378 level = exp->exp_failed ? D_INFO : D_ERROR;
379 if (!mdt_object_exists(mo)) {
381 "%s: READ IO to missing obj "DFID": rc = %d\n",
382 exp->exp_obd->obd_name, PFID(mdt_object_fid(mo)),
384 /* return 0 and continue with empty commit to skip such READ
385 * without more BRW errors.
389 if (lu_object_is_dying(&mo->mot_header)) {
391 "%s: READ IO to stale obj "DFID": rc = %d\n",
392 exp->exp_obd->obd_name, PFID(mdt_object_fid(mo)),
394 /* return 0 and continue with empty commit to skip such READ
395 * without more BRW errors.
400 dob = mdt_obj2dt(mo);
401 /* parse remote buffers to local buffers and prepare the latter */
402 for (i = 0, j = 0; i < niocount; i++) {
403 rc = dt_bufs_get(env, dob, rnb + i, lnb + j, maxlnb, 0);
404 if (unlikely(rc < 0))
406 /* correct index for local buffers to continue with */
410 tot_bytes += rnb[i].rnb_len;
413 rc = dt_attr_get(env, dob, la);
417 rc = dt_read_prep(env, dob, lnb, *nr_local);
421 mdt_counter_incr(req, LPROC_MDT_IO_READ, tot_bytes);
424 dt_bufs_put(env, dob, lnb, *nr_local);
425 mdt_dom_read_unlock(mo);
429 static int mdt_preprw_write(const struct lu_env *env, struct obd_export *exp,
430 struct mdt_device *mdt, struct mdt_object *mo,
431 struct lu_attr *la, struct obdo *oa,
432 int objcount, struct obd_ioobj *obj,
433 struct niobuf_remote *rnb, int *nr_local,
434 struct niobuf_local *lnb, char *jobid)
436 struct tgt_session_info *tsi = tgt_ses_info(env);
437 struct ptlrpc_request *req = tgt_ses_req(tsi);
438 struct dt_object *dob;
439 int i, j, k, rc = 0, tot_bytes = 0;
440 int maxlnb = *nr_local;
444 /* Process incoming grant info, set OBD_BRW_GRANTED flag and grant some
445 * space back if possible */
446 tgt_grant_prepare_write(env, exp, oa, rnb, obj->ioo_bufcnt);
448 mdt_dom_read_lock(mo);
450 /* don't report error in cases with failed export */
451 if (!mdt_object_exists(mo)) {
452 int level = exp->exp_failed ? D_INFO : D_ERROR;
456 "%s: WRITE IO to missing obj "DFID": rc = %d\n",
457 exp->exp_obd->obd_name, PFID(mdt_object_fid(mo)),
459 /* exit with no data written, note nr_local = 0 above */
462 if (lu_object_is_dying(&mo->mot_header)) {
463 /* This is possible race between object destroy followed by
464 * discard BL AST and client cache flushing. Object is
465 * referenced until discard finish.
467 CDEBUG(D_INODE, "WRITE IO to stale object "DFID"\n",
468 PFID(mdt_object_fid(mo)));
469 /* Note: continue with no error here to don't cause BRW errors
470 * but skip transaction in commitrw silently so no data is
475 dob = mdt_obj2dt(mo);
476 /* parse remote buffers to local buffers and prepare the latter */
477 for (i = 0, j = 0; i < obj->ioo_bufcnt; i++) {
478 rc = dt_bufs_get(env, dob, rnb + i, lnb + j, maxlnb, 1);
479 if (unlikely(rc < 0))
481 /* correct index for local buffers to continue with */
482 for (k = 0; k < rc; k++) {
483 lnb[j + k].lnb_flags = rnb[i].rnb_flags;
484 if (!(rnb[i].rnb_flags & OBD_BRW_GRANTED))
485 lnb[j + k].lnb_rc = -ENOSPC;
490 tot_bytes += rnb[i].rnb_len;
493 rc = dt_write_prep(env, dob, lnb, *nr_local);
497 mdt_counter_incr(req, LPROC_MDT_IO_WRITE, tot_bytes);
500 dt_bufs_put(env, dob, lnb, *nr_local);
502 mdt_dom_read_unlock(mo);
503 /* tgt_grant_prepare_write() was called, so we must commit */
504 tgt_grant_commit(exp, oa->o_grant_used, rc);
505 /* let's still process incoming grant information packed in the oa,
506 * but without enforcing grant since we won't proceed with the write.
507 * Just like a read request actually. */
508 tgt_grant_prepare_read(env, exp, oa);
512 int mdt_obd_preprw(const struct lu_env *env, int cmd, struct obd_export *exp,
513 struct obdo *oa, int objcount, struct obd_ioobj *obj,
514 struct niobuf_remote *rnb, int *nr_local,
515 struct niobuf_local *lnb)
517 struct tgt_session_info *tsi = tgt_ses_info(env);
518 struct mdt_thread_info *info = tsi2mdt_info(tsi);
519 struct lu_attr *la = &info->mti_attr.ma_attr;
520 struct mdt_device *mdt = mdt_dev(exp->exp_obd->obd_lu_dev);
521 struct mdt_object *mo;
525 /* The default value PTLRPC_MAX_BRW_PAGES is set in tgt_brw_write()
526 * but for MDT it is different, correct it here. */
527 if (*nr_local > MD_MAX_BRW_PAGES)
528 *nr_local = MD_MAX_BRW_PAGES;
530 jobid = tsi->tsi_jobid;
532 if (!oa || objcount != 1 || obj->ioo_bufcnt == 0) {
533 CERROR("%s: bad parameters %p/%i/%i\n",
534 exp->exp_obd->obd_name, oa, objcount, obj->ioo_bufcnt);
538 mo = mdt_object_find(env, mdt, &tsi->tsi_fid);
540 GOTO(out, rc = PTR_ERR(mo));
542 LASSERT(info->mti_object == NULL);
543 info->mti_object = mo;
545 if (cmd == OBD_BRW_WRITE) {
546 la_from_obdo(la, oa, OBD_MD_FLGETATTR);
547 rc = mdt_preprw_write(env, exp, mdt, mo, la, oa,
548 objcount, obj, rnb, nr_local, lnb,
550 } else if (cmd == OBD_BRW_READ) {
551 tgt_grant_prepare_read(env, exp, oa);
552 rc = mdt_preprw_read(env, exp, mdt, mo, la,
553 obj->ioo_bufcnt, rnb, nr_local, lnb,
555 obdo_from_la(oa, la, LA_ATIME);
557 CERROR("%s: wrong cmd %d received!\n",
558 exp->exp_obd->obd_name, cmd);
562 lu_object_put(env, &mo->mot_obj);
563 info->mti_object = NULL;
569 static int mdt_commitrw_read(const struct lu_env *env, struct mdt_device *mdt,
570 struct mdt_object *mo, int objcount, int niocount,
571 struct niobuf_local *lnb)
573 struct dt_object *dob;
578 dob = mdt_obj2dt(mo);
581 dt_bufs_put(env, dob, lnb, niocount);
583 mdt_dom_read_unlock(mo);
587 static int mdt_commitrw_write(const struct lu_env *env, struct obd_export *exp,
588 struct mdt_device *mdt, struct mdt_object *mo,
589 struct lu_attr *la, struct obdo *oa, int objcount,
590 int niocount, struct niobuf_local *lnb,
591 unsigned long granted, int old_rc)
593 struct dt_device *dt = mdt->mdt_bottom;
594 struct dt_object *dob;
602 dob = mdt_obj2dt(mo);
605 GOTO(out, rc = old_rc);
607 la->la_valid &= LA_ATIME | LA_MTIME | LA_CTIME;
609 if (!dt_object_exists(dob))
610 GOTO(out, rc = -ENOENT);
611 if (lu_object_is_dying(&mo->mot_header)) {
612 /* Commit to stale object can be just skipped silently. */
613 CDEBUG(D_INODE, "skip commit to stale object "DFID"\n",
614 PFID(mdt_object_fid(mo)));
620 DEBUG_REQ(D_WARNING, tgt_ses_req(tgt_ses_info(env)),
621 "%s: commit with no pages for "DFID": rc = %d\n",
622 exp->exp_obd->obd_name, PFID(mdt_object_fid(mo)), rc);
626 th = dt_trans_create(env, dt);
628 GOTO(out, rc = PTR_ERR(th));
630 for (i = 0; i < niocount; i++) {
631 if (!(lnb[i].lnb_flags & OBD_BRW_ASYNC)) {
637 if (OBD_FAIL_CHECK(OBD_FAIL_OST_DQACQ_NET))
638 GOTO(out_stop, rc = -EINPROGRESS);
640 rc = dt_declare_write_commit(env, dob, lnb, niocount, th);
645 /* update [mac]time if needed */
646 rc = dt_declare_attr_set(env, dob, la, th);
651 tgt_vbr_obj_set(env, dob);
652 rc = dt_trans_start(env, dt, th);
656 dt_write_lock(env, dob, 0);
657 rc = dt_write_commit(env, dob, lnb, niocount, th, oa->o_size);
662 rc = dt_attr_set(env, dob, la, th);
666 /* get attr to return */
667 rc = dt_attr_get(env, dob, la);
669 dt_write_unlock(env, dob);
672 /* Force commit to make the just-deleted blocks
673 * reusable. LU-456 */
678 if (rc == 0 && granted > 0) {
679 if (tgt_grant_commit_cb_add(th, exp, granted) == 0)
684 dt_trans_stop(env, dt, th);
685 if (rc == -ENOSPC && retries++ < 3) {
686 CDEBUG(D_INODE, "retry after force commit, retries:%d\n",
692 dt_bufs_put(env, dob, lnb, niocount);
693 mdt_dom_read_unlock(mo);
695 tgt_grant_commit(exp, granted, old_rc);
699 void mdt_dom_obj_lvb_update(const struct lu_env *env, struct mdt_object *mo,
702 struct mdt_device *mdt = mdt_dev(mo->mot_obj.lo_dev);
703 struct ldlm_res_id resid;
704 struct ldlm_resource *res;
706 fid_build_reg_res_name(mdt_object_fid(mo), &resid);
707 res = ldlm_resource_get(mdt->mdt_namespace, NULL, &resid,
712 /* Update lvbo data if exists. */
713 if (mdt_dom_lvb_is_valid(res))
714 mdt_dom_disk_lvbo_update(env, mo, res, increase_only);
715 ldlm_resource_putref(res);
718 int mdt_obd_commitrw(const struct lu_env *env, int cmd, struct obd_export *exp,
719 struct obdo *oa, int objcount, struct obd_ioobj *obj,
720 struct niobuf_remote *rnb, int npages,
721 struct niobuf_local *lnb, int old_rc)
723 struct mdt_thread_info *info = mdt_th_info(env);
724 struct mdt_device *mdt = mdt_dev(exp->exp_obd->obd_lu_dev);
725 struct mdt_object *mo = info->mti_object;
726 struct lu_attr *la = &info->mti_attr.ma_attr;
732 if (cmd == OBD_BRW_WRITE) {
733 /* Don't update timestamps if this write is older than a
734 * setattr which modifies the timestamps. b=10150 */
736 /* XXX when we start having persistent reservations this needs
737 * to be changed to ofd_fmd_get() to create the fmd if it
738 * doesn't already exist so we can store the reservation handle
740 valid = OBD_MD_FLUID | OBD_MD_FLGID;
741 if (tgt_fmd_check(exp, mdt_object_fid(mo),
742 mdt_info_req(info)->rq_xid))
743 valid |= OBD_MD_FLATIME | OBD_MD_FLMTIME |
746 la_from_obdo(la, oa, valid);
748 rc = mdt_commitrw_write(env, exp, mdt, mo, la, oa, objcount,
749 npages, lnb, oa->o_grant_used, old_rc);
751 obdo_from_la(oa, la, VALID_FLAGS | LA_GID | LA_UID);
753 obdo_from_la(oa, la, LA_GID | LA_UID);
755 mdt_dom_obj_lvb_update(env, mo, false);
756 /* don't report overquota flag if we failed before reaching
758 if (old_rc == 0 && (rc == 0 || rc == -EDQUOT)) {
759 /* return the overquota flags to client */
760 if (lnb[0].lnb_flags & OBD_BRW_OVER_USRQUOTA) {
761 if (oa->o_valid & OBD_MD_FLFLAGS)
762 oa->o_flags |= OBD_FL_NO_USRQUOTA;
764 oa->o_flags = OBD_FL_NO_USRQUOTA;
767 if (lnb[0].lnb_flags & OBD_BRW_OVER_GRPQUOTA) {
768 if (oa->o_valid & OBD_MD_FLFLAGS)
769 oa->o_flags |= OBD_FL_NO_GRPQUOTA;
771 oa->o_flags = OBD_FL_NO_GRPQUOTA;
774 if (lnb[0].lnb_flags & OBD_BRW_OVER_PRJQUOTA) {
775 if (oa->o_valid & OBD_MD_FLFLAGS)
776 oa->o_flags |= OBD_FL_NO_PRJQUOTA;
778 oa->o_flags = OBD_FL_NO_PRJQUOTA;
781 oa->o_valid |= OBD_MD_FLFLAGS | OBD_MD_FLUSRQUOTA |
782 OBD_MD_FLGRPQUOTA | OBD_MD_FLPRJQUOTA;
784 } else if (cmd == OBD_BRW_READ) {
785 /* If oa != NULL then mdt_preprw_read updated the inode
786 * atime and we should update the lvb so that other glimpses
787 * will also get the updated value. bug 5972 */
789 mdt_dom_obj_lvb_update(env, mo, true);
790 rc = mdt_commitrw_read(env, mdt, mo, objcount, npages, lnb);
796 mdt_thread_info_fini(info);
800 int mdt_object_punch(const struct lu_env *env, struct dt_device *dt,
801 struct dt_object *dob, __u64 start, __u64 end,
809 /* we support truncate, not punch yet */
810 LASSERT(end == OBD_OBJECT_EOF);
812 if (!dt_object_exists(dob))
815 th = dt_trans_create(env, dt);
819 rc = dt_declare_attr_set(env, dob, la, th);
823 rc = dt_declare_punch(env, dob, start, OBD_OBJECT_EOF, th);
827 tgt_vbr_obj_set(env, dob);
828 rc = dt_trans_start(env, dt, th);
832 dt_write_lock(env, dob, 0);
833 rc = dt_punch(env, dob, start, OBD_OBJECT_EOF, th);
836 rc = dt_attr_set(env, dob, la, th);
840 dt_write_unlock(env, dob);
843 dt_trans_stop(env, dt, th);
847 int mdt_punch_hdl(struct tgt_session_info *tsi)
849 const struct obdo *oa = &tsi->tsi_ost_body->oa;
850 struct ptlrpc_request *req = tgt_ses_req(tsi);
851 struct ost_body *repbody;
852 struct mdt_thread_info *info;
854 struct ldlm_namespace *ns = tsi->tsi_tgt->lut_obd->obd_namespace;
855 struct obd_export *exp = tsi->tsi_exp;
856 struct mdt_device *mdt = mdt_dev(exp->exp_obd->obd_lu_dev);
857 struct mdt_object *mo;
858 struct dt_object *dob;
860 struct lustre_handle lh = { 0, };
861 ktime_t kstart = ktime_get();
868 /* check that we do support OBD_CONNECT_TRUNCLOCK. */
869 BUILD_BUG_ON(!(OST_CONNECT_SUPPORTED & OBD_CONNECT_TRUNCLOCK));
871 if ((oa->o_valid & (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS)) !=
872 (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS))
873 RETURN(err_serious(-EPROTO));
875 repbody = req_capsule_server_get(tsi->tsi_pill, &RMF_OST_BODY);
877 RETURN(err_serious(-ENOMEM));
879 /* punch start,end are passed in o_size,o_blocks throught wire */
883 if (end != OBD_OBJECT_EOF) /* Only truncate is supported */
886 info = tsi2mdt_info(tsi);
887 la = &info->mti_attr.ma_attr;
888 /* standard truncate optimization: if file body is completely
889 * destroyed, don't send data back to the server. */
891 flags |= LDLM_FL_AST_DISCARD_DATA;
893 repbody->oa.o_oi = oa->o_oi;
894 repbody->oa.o_valid = OBD_MD_FLID;
896 srvlock = (exp_connect_flags(exp) & OBD_CONNECT_SRVLOCK) &&
897 oa->o_valid & OBD_MD_FLFLAGS &&
898 oa->o_flags & OBD_FL_SRVLOCK;
901 rc = tgt_mdt_data_lock(ns, &tsi->tsi_resid, &lh, LCK_PW,
907 CDEBUG(D_INODE, "calling punch for object "DFID", valid = %#llx"
908 ", start = %lld, end = %lld\n", PFID(&tsi->tsi_fid),
909 oa->o_valid, start, end);
911 mo = mdt_object_find(tsi->tsi_env, mdt, &tsi->tsi_fid);
913 GOTO(out_unlock, rc = PTR_ERR(mo));
915 if (!mdt_object_exists(mo))
916 GOTO(out_put, rc = -ENOENT);
918 /* Shouldn't happen on dirs */
919 if (S_ISDIR(lu_object_attr(&mo->mot_obj))) {
921 CERROR("%s: Truncate on dir "DFID": rc = %d\n",
922 exp->exp_obd->obd_name, PFID(&tsi->tsi_fid), rc);
926 mdt_dom_write_lock(mo);
927 dob = mdt_obj2dt(mo);
929 la_from_obdo(la, oa, OBD_MD_FLMTIME | OBD_MD_FLATIME | OBD_MD_FLCTIME);
931 la->la_valid |= LA_SIZE;
933 /* MDT supports FMD for Data-on-MDT needs */
934 if (la->la_valid & (LA_ATIME | LA_MTIME | LA_CTIME))
935 tgt_fmd_update(tsi->tsi_exp, &tsi->tsi_fid,
936 tgt_ses_req(tsi)->rq_xid);
938 rc = mdt_object_punch(tsi->tsi_env, mdt->mdt_bottom, dob,
940 mdt_dom_write_unlock(mo);
944 mdt_dom_obj_lvb_update(tsi->tsi_env, mo, false);
945 mdt_counter_incr(req, LPROC_MDT_IO_PUNCH,
946 ktime_us_delta(ktime_get(), kstart));
949 lu_object_put(tsi->tsi_env, &mo->mot_obj);
952 tgt_data_unlock(&lh, LCK_PW);
954 mdt_thread_info_fini(info);
959 * MDT glimpse for Data-on-MDT
961 * If there is write lock on client then function issues glimpse_ast to get
962 * an actual size from that client.
965 int mdt_do_glimpse(const struct lu_env *env, struct ldlm_namespace *ns,
966 struct ldlm_resource *res)
968 union ldlm_policy_data policy;
969 struct lustre_handle lockh;
971 struct ldlm_lock *lock;
972 struct ldlm_glimpse_work *gl_work;
978 /* There can be only one write lock covering data, try to match it. */
979 policy.l_inodebits.bits = MDS_INODELOCK_DOM;
980 mode = ldlm_lock_match(ns, LDLM_FL_TEST_LOCK,
981 &res->lr_name, LDLM_IBITS, &policy,
984 /* There is no PW lock on this object; finished. */
988 lock = ldlm_handle2lock(&lockh);
993 * This check is for lock taken in mdt_reint_unlink() that does
994 * not have l_glimpse_ast set. So the logic is: if there is a lock
995 * with no l_glimpse_ast set, this object is being destroyed already.
996 * Hence, if you are grabbing DLM locks on the server, always set
997 * non-NULL glimpse_ast (e.g., ldlm_request.c::ldlm_glimpse_ast()).
999 if (lock->l_glimpse_ast == NULL) {
1000 LDLM_DEBUG(lock, "no l_glimpse_ast");
1001 GOTO(out, rc = -ENOENT);
1004 OBD_SLAB_ALLOC_PTR_GFP(gl_work, ldlm_glimpse_work_kmem, GFP_ATOMIC);
1006 GOTO(out, rc = -ENOMEM);
1008 /* Populate the gl_work structure.
1009 * Grab additional reference on the lock which will be released in
1010 * ldlm_work_gl_ast_lock() */
1011 gl_work->gl_lock = LDLM_LOCK_GET(lock);
1012 /* The glimpse callback is sent to one single IO lock. As a result,
1013 * the gl_work list is just composed of one element */
1014 list_add_tail(&gl_work->gl_list, &gl_list);
1015 /* There is actually no need for a glimpse descriptor when glimpsing
1017 gl_work->gl_desc = NULL;
1018 /* the ldlm_glimpse_work structure is allocated on the stack */
1019 gl_work->gl_flags = LDLM_GL_WORK_SLAB_ALLOCATED;
1021 ldlm_glimpse_locks(res, &gl_list); /* this will update the LVB */
1023 /* If the list is not empty, we failed to glimpse a lock and
1024 * must clean it up. Usually due to a race with unlink.*/
1025 if (!list_empty(&gl_list)) {
1026 LDLM_LOCK_RELEASE(lock);
1027 OBD_SLAB_FREE_PTR(gl_work, ldlm_glimpse_work_kmem);
1032 LDLM_LOCK_PUT(lock);
1036 static void mdt_lvb2body(struct ldlm_resource *res, struct mdt_body *mb)
1038 struct ost_lvb *res_lvb;
1041 res_lvb = res->lr_lvb_data;
1042 mb->mbo_dom_size = res_lvb->lvb_size;
1043 mb->mbo_dom_blocks = res_lvb->lvb_blocks;
1044 mb->mbo_mtime = res_lvb->lvb_mtime;
1045 mb->mbo_ctime = res_lvb->lvb_ctime;
1046 mb->mbo_atime = res_lvb->lvb_atime;
1048 CDEBUG(D_DLMTRACE, "size %llu\n", res_lvb->lvb_size);
1050 mb->mbo_valid |= OBD_MD_FLATIME | OBD_MD_FLCTIME | OBD_MD_FLMTIME |
1056 * MDT glimpse for Data-on-MDT
1058 * This function is called when MDT get attributes for the DoM object.
1059 * If there is write lock on client then function issues glimpse_ast to get
1060 * an actual size from that client.
1062 int mdt_dom_object_size(const struct lu_env *env, struct mdt_device *mdt,
1063 const struct lu_fid *fid, struct mdt_body *mb,
1066 struct ldlm_res_id resid;
1067 struct ldlm_resource *res;
1072 fid_build_reg_res_name(fid, &resid);
1073 res = ldlm_resource_get(mdt->mdt_namespace, NULL, &resid,
1078 /* Update lvbo data if DoM lock returned or if LVB is not yet valid. */
1079 if (dom_lock || !mdt_dom_lvb_is_valid(res))
1080 mdt_dom_lvbo_update(res, NULL, NULL, false);
1082 mdt_lvb2body(res, mb);
1083 ldlm_resource_putref(res);
1088 * MDT DoM lock intent policy (glimpse)
1090 * Intent policy is called when lock has an intent, for DoM file that
1091 * means glimpse lock and policy fills Lock Value Block (LVB).
1093 * If already granted lock is found it will be placed in \a lockp and
1094 * returned back to caller function.
1096 * \param[in] tsi session info
1097 * \param[in,out] lockp pointer to the lock
1098 * \param[in] flags LDLM flags
1100 * \retval ELDLM_LOCK_REPLACED if already granted lock was found
1101 * and placed in \a lockp
1102 * \retval ELDLM_LOCK_ABORTED in other cases except error
1103 * \retval negative value on error
1105 int mdt_glimpse_enqueue(struct mdt_thread_info *mti, struct ldlm_namespace *ns,
1106 struct ldlm_lock **lockp, __u64 flags)
1108 struct ldlm_lock *lock = *lockp;
1109 struct ldlm_resource *res = lock->l_resource;
1110 ldlm_processing_policy policy;
1111 struct ldlm_reply *rep;
1112 struct mdt_body *mbo;
1117 policy = ldlm_get_processing_policy(res);
1118 LASSERT(policy != NULL);
1120 req_capsule_set_size(mti->mti_pill, &RMF_MDT_MD, RCL_SERVER, 0);
1121 req_capsule_set_size(mti->mti_pill, &RMF_ACL, RCL_SERVER, 0);
1122 rc = req_capsule_server_pack(mti->mti_pill);
1124 RETURN(err_serious(rc));
1126 rep = req_capsule_server_get(mti->mti_pill, &RMF_DLM_REP);
1130 mbo = req_capsule_server_get(mti->mti_pill, &RMF_MDT_BODY);
1135 /* Check if this is a resend case (MSG_RESENT is set on RPC) and a
1136 * lock was found by ldlm_handle_enqueue(); if so no need to grant
1138 if (flags & LDLM_FL_RESENT) {
1139 rc = LDLM_ITER_CONTINUE;
1141 __u64 tmpflags = LDLM_FL_BLOCK_NOWAIT;
1142 enum ldlm_error err;
1144 rc = policy(lock, &tmpflags, LDLM_PROCESS_RESCAN, &err, NULL);
1145 check_res_locked(res);
1149 /* The lock met with no resistance; we're finished. */
1150 if (rc == LDLM_ITER_CONTINUE) {
1151 GOTO(fill_mbo, rc = ELDLM_LOCK_REPLACED);
1152 } else if (flags & LDLM_FL_BLOCK_NOWAIT) {
1153 /* LDLM_FL_BLOCK_NOWAIT means it is for AGL. Do not send glimpse
1154 * callback for glimpse size. The real size user will trigger
1155 * the glimpse callback when necessary. */
1156 GOTO(fill_mbo, rc = ELDLM_LOCK_ABORTED);
1159 rc = mdt_do_glimpse(mti->mti_env, ns, res);
1160 if (rc == -ENOENT) {
1161 /* We are racing with unlink(); just return -ENOENT */
1162 rep->lock_policy_res2 = ptlrpc_status_hton(-ENOENT);
1164 } else if (rc == -EINVAL) {
1165 /* this is possible is client lock has been cancelled but
1166 * still exists on server. If that lock was found on server
1167 * as only conflicting lock then the client has already
1168 * size authority and glimpse is not needed. */
1169 CDEBUG(D_DLMTRACE, "Glimpse from the client owning lock\n");
1171 } else if (rc < 0) {
1174 rc = ELDLM_LOCK_ABORTED;
1176 /* LVB can be without valid data in case of DOM */
1177 if (!mdt_dom_lvb_is_valid(res))
1178 mdt_dom_lvbo_update(res, lock, NULL, false);
1179 mdt_lvb2body(res, mbo);
1183 int mdt_brw_enqueue(struct mdt_thread_info *mti, struct ldlm_namespace *ns,
1184 struct ldlm_lock **lockp, __u64 flags)
1186 struct tgt_session_info *tsi = tgt_ses_info(mti->mti_env);
1187 struct lu_fid *fid = &tsi->tsi_fid;
1188 struct ldlm_lock *lock = *lockp;
1189 struct ldlm_resource *res = lock->l_resource;
1190 struct ldlm_reply *rep;
1191 struct mdt_body *mbo;
1192 struct mdt_lock_handle *lhc = &mti->mti_lh[MDT_LH_RMT];
1193 struct mdt_object *mo;
1198 /* Get lock from request for possible resent case. */
1199 mdt_intent_fixup_resent(mti, *lockp, lhc, flags);
1200 req_capsule_set_size(mti->mti_pill, &RMF_MDT_MD, RCL_SERVER, 0);
1201 req_capsule_set_size(mti->mti_pill, &RMF_ACL, RCL_SERVER, 0);
1202 rc = req_capsule_server_pack(mti->mti_pill);
1204 RETURN(err_serious(rc));
1206 rep = req_capsule_server_get(mti->mti_pill, &RMF_DLM_REP);
1210 mbo = req_capsule_server_get(mti->mti_pill, &RMF_MDT_BODY);
1214 fid_extract_from_res_name(fid, &res->lr_name);
1215 mo = mdt_object_find(mti->mti_env, mti->mti_mdt, fid);
1216 if (unlikely(IS_ERR(mo)))
1217 RETURN(PTR_ERR(mo));
1219 if (!mdt_object_exists(mo))
1220 GOTO(out, rc = -ENOENT);
1222 if (mdt_object_remote(mo))
1223 GOTO(out, rc = -EPROTO);
1226 if (!lustre_handle_is_used(&lhc->mlh_reg_lh)) {
1227 mdt_lock_handle_init(lhc);
1228 mdt_lh_reg_init(lhc, *lockp);
1230 /* This will block MDT thread but it should be fine until
1231 * client caches small amount of data for DoM, which should be
1232 * smaller than one BRW RPC and should be able to be
1233 * piggybacked by lock cancel RPC.
1234 * If the client could hold the lock too long, this code can be
1235 * revised to call mdt_object_lock_try(). And if fails, it will
1236 * return ELDLM_OK here and fall back into normal lock enqueue
1239 rc = mdt_object_lock(mti, mo, lhc, MDS_INODELOCK_DOM);
1244 if (!mdt_dom_lvb_is_valid(res)) {
1245 rc = mdt_dom_lvb_alloc(res);
1248 mdt_dom_disk_lvbo_update(mti->mti_env, mo, res, false);
1250 mdt_lvb2body(res, mbo);
1252 rep->lock_policy_res2 = clear_serious(rc);
1253 if (rep->lock_policy_res2) {
1254 lhc->mlh_reg_lh.cookie = 0ull;
1255 GOTO(out, rc = ELDLM_LOCK_ABORTED);
1258 rc = mdt_intent_lock_replace(mti, lockp, lhc, flags, rc);
1260 mdt_object_put(mti->mti_env, mo);
1264 /* check if client has already DoM lock for given resource */
1265 bool mdt_dom_client_has_lock(struct mdt_thread_info *info,
1266 const struct lu_fid *fid)
1268 struct mdt_device *mdt = info->mti_mdt;
1269 union ldlm_policy_data *policy = &info->mti_policy;
1270 struct ldlm_res_id *res_id = &info->mti_res_id;
1271 __u64 open_flags = info->mti_spec.sp_cr_flags;
1272 struct lustre_handle lockh;
1273 enum ldlm_mode mode;
1274 struct ldlm_lock *lock;
1278 policy->l_inodebits.bits = MDS_INODELOCK_DOM;
1279 fid_build_reg_res_name(fid, res_id);
1282 lm = (open_flags & MDS_FMODE_WRITE) ? LCK_PW : LCK_PR | LCK_PW;
1283 mode = ldlm_lock_match(mdt->mdt_namespace, LDLM_FL_BLOCK_GRANTED |
1284 LDLM_FL_TEST_LOCK, res_id, LDLM_IBITS, policy,
1287 /* There is no other PW lock on this object; finished. */
1291 lock = ldlm_handle2lock(&lockh);
1295 /* check if lock from the same client */
1296 rc = (lock->l_export->exp_handle.h_cookie ==
1297 info->mti_exp->exp_handle.h_cookie);
1298 LDLM_LOCK_PUT(lock);
1303 * MDT request handler for OST_GETATTR RPC.
1305 * This is data-specific request to get object and layout versions under
1306 * IO lock. It is reliable only for Data-on-MDT files.
1308 * \param[in] tsi target session environment for this request
1310 * \retval 0 if successful
1311 * \retval negative value on error
1313 int mdt_data_version_get(struct tgt_session_info *tsi)
1315 struct mdt_thread_info *mti = mdt_th_info(tsi->tsi_env);
1316 struct mdt_device *mdt = mti->mti_mdt;
1317 struct mdt_body *repbody;
1318 struct mdt_object *mo = mti->mti_object;
1319 struct lov_comp_md_v1 *comp;
1320 struct lustre_handle lh = { 0 };
1323 enum ldlm_mode lock_mode = LCK_PR;
1329 req_capsule_set_size(tsi->tsi_pill, &RMF_MDT_MD, RCL_SERVER, 0);
1330 req_capsule_set_size(tsi->tsi_pill, &RMF_ACL, RCL_SERVER, 0);
1331 rc = req_capsule_server_pack(tsi->tsi_pill);
1332 if (unlikely(rc != 0))
1333 RETURN(err_serious(rc));
1335 repbody = req_capsule_server_get(tsi->tsi_pill, &RMF_MDT_BODY);
1336 if (repbody == NULL)
1339 srvlock = tsi->tsi_mdt_body->mbo_valid & OBD_MD_FLFLAGS &&
1340 tsi->tsi_mdt_body->mbo_flags & OBD_FL_SRVLOCK;
1343 if (unlikely(tsi->tsi_mdt_body->mbo_flags & OBD_FL_FLUSH))
1346 fid_build_reg_res_name(&tsi->tsi_fid, &tsi->tsi_resid);
1347 rc = tgt_mdt_data_lock(mdt->mdt_namespace, &tsi->tsi_resid,
1348 &lh, lock_mode, &flags);
1353 if (!mdt_object_exists(mo))
1354 GOTO(out, rc = -ENOENT);
1355 if (mdt_object_remote(mo))
1356 GOTO(out, rc = -EREMOTE);
1357 if (!S_ISREG(lu_object_attr(&mo->mot_obj)))
1358 GOTO(out, rc = -EBADF);
1360 /* Get version first */
1361 version = dt_version_get(tsi->tsi_env, mdt_obj2dt(mo));
1362 if (version && version != -EOPNOTSUPP) {
1363 repbody->mbo_valid |= OBD_MD_FLDATAVERSION;
1364 /* re-use mbo_ioepoch to transfer version */
1365 repbody->mbo_version = version;
1368 /* Read layout to get its version */
1369 rc = mdt_big_xattr_get(mti, mo, XATTR_NAME_LOV);
1370 if (rc == -ENODATA) /* File has no layout yet */
1375 comp = mti->mti_buf.lb_buf;
1376 if (le32_to_cpu(comp->lcm_magic) != LOV_MAGIC_COMP_V1) {
1377 CDEBUG(D_INFO, DFID" has no composite layout",
1378 PFID(&tsi->tsi_fid));
1379 GOTO(out, rc = -ESTALE);
1382 CDEBUG(D_INODE, DFID": layout version: %u\n",
1383 PFID(&tsi->tsi_fid), le32_to_cpu(comp->lcm_layout_gen));
1385 repbody->mbo_valid |= OBD_MD_LAYOUT_VERSION;
1386 /* re-use mbo_rdev for that */
1387 repbody->mbo_layout_gen = le32_to_cpu(comp->lcm_layout_gen);
1391 tgt_data_unlock(&lh, lock_mode);
1393 repbody->mbo_valid |= OBD_MD_FLFLAGS;
1394 repbody->mbo_flags = OBD_FL_FLUSH;
1398 /* read file data to the buffer */
1399 int mdt_dom_read_on_open(struct mdt_thread_info *mti, struct mdt_device *mdt,
1400 struct lustre_handle *lh)
1402 const struct lu_env *env = mti->mti_env;
1403 struct tgt_session_info *tsi = tgt_ses_info(env);
1404 struct req_capsule *pill = tsi->tsi_pill;
1405 const struct lu_fid *fid;
1406 struct ptlrpc_request *req = tgt_ses_req(tsi);
1407 struct mdt_body *mbo;
1408 struct dt_device *dt = mdt->mdt_bottom;
1409 struct dt_object *mo;
1411 struct niobuf_remote *rnb = NULL;
1412 struct niobuf_local *lnb;
1415 unsigned int len, copied = 0;
1416 __u64 real_dom_size;
1417 int lnbs, nr_local, i;
1418 bool dom_lock = false;
1422 if (!req_capsule_field_present(pill, &RMF_NIOBUF_INLINE, RCL_SERVER)) {
1423 /* There is no reply buffers for this field, this means that
1424 * client has no support for data in reply.
1429 mbo = req_capsule_server_get(pill, &RMF_MDT_BODY);
1430 if (!(mbo->mbo_valid & OBD_MD_DOM_SIZE))
1433 if (!mbo->mbo_dom_size)
1436 if (lustre_handle_is_used(lh)) {
1437 struct ldlm_lock *lock;
1439 lock = ldlm_handle2lock(lh);
1441 dom_lock = ldlm_has_dom(lock) && ldlm_has_layout(lock);
1442 LDLM_LOCK_PUT(lock);
1446 /* return data along with open only along with DoM lock */
1447 if (!dom_lock || !mdt->mdt_opts.mo_dom_read_open)
1450 /* if DoM object holds encrypted content, we need to make sure we
1451 * send whole encryption units, or client will read corrupted content
1453 if (mbo->mbo_valid & LA_FLAGS && mbo->mbo_flags & LUSTRE_ENCRYPT_FL &&
1454 mbo->mbo_dom_size & ~LUSTRE_ENCRYPTION_MASK)
1455 real_dom_size = (mbo->mbo_dom_size & LUSTRE_ENCRYPTION_MASK) +
1456 LUSTRE_ENCRYPTION_UNIT_SIZE;
1458 real_dom_size = mbo->mbo_dom_size;
1460 CDEBUG(D_INFO, "File size %llu, reply sizes %d/%d\n",
1461 real_dom_size, req->rq_reqmsg->lm_repsize, req->rq_replen);
1462 len = req->rq_reqmsg->lm_repsize - req->rq_replen;
1464 /* NB: at this moment we have the following sizes:
1465 * - req->rq_replen: used data in reply
1466 * - req->rq_reqmsg->lm_repsize: total allocated reply buffer at client
1468 * Ideal case when file size fits in allocated reply buffer,
1469 * that mean we can return whole data in reply. We can also fit more
1470 * data up to max_reply_size in total reply size, but this will cause
1471 * re-allocation on client and resend with larger buffer. This is still
1472 * faster than separate READ IO.
1473 * Third case if file is too big to fit even in maximum size, in that
1474 * case we return just tail to optimize possible append.
1476 * At the moment the following strategy is used:
1477 * 1) try to fit into the buffer we have
1478 * 2) return just file tail otherwise.
1480 if (real_dom_size <= len) {
1481 /* can fit whole data */
1482 len = real_dom_size;
1484 } else if (real_dom_size <
1485 mdt_lmm_dom_stripesize(mti->mti_attr.ma_lmm)) {
1488 /* File tail offset must be aligned with larger page size
1489 * between client and server, so the maximum page size is
1490 * used here to align offset.
1492 * NB: DOM feature was introduced when server supports pagebits
1493 * already, so it should be always non-zero value. Report error
1494 * if it is not for some reason.
1496 if (!req->rq_export->exp_target_data.ted_pagebits) {
1497 CERROR("%s: client page bits are not saved on server\n",
1501 pgbits = max_t(int, PAGE_SHIFT,
1502 req->rq_export->exp_target_data.ted_pagebits);
1503 tail = real_dom_size % (1 << pgbits);
1505 /* no partial tail or tail can't fit in reply */
1506 if (tail == 0 || len < tail)
1510 offset = real_dom_size - len;
1512 /* DOM stripe is fully written, so don't expect its tail
1513 * will be used by append.
1518 LASSERT((offset & ~PAGE_MASK) == 0);
1519 rc = req_capsule_server_grow(pill, &RMF_NIOBUF_INLINE,
1520 sizeof(*rnb) + len);
1522 /* failed to grow data buffer, just exit */
1523 GOTO(out, rc = -E2BIG);
1526 /* re-take MDT_BODY and NIOBUF_INLINE buffers after the buffer grow */
1527 mbo = req_capsule_server_get(pill, &RMF_MDT_BODY);
1528 fid = &mbo->mbo_fid1;
1529 if (!fid_is_sane(fid))
1530 GOTO(out, rc = -EINVAL);
1532 rnb = req_capsule_server_get(tsi->tsi_pill, &RMF_NIOBUF_INLINE);
1534 GOTO(out, rc = -EPROTO);
1536 buf = (char *)rnb + sizeof(*rnb);
1538 rnb->rnb_offset = offset;
1540 mo = dt_locate(env, dt, fid);
1542 GOTO(out_rnb, rc = PTR_ERR(mo));
1543 LASSERT(mo != NULL);
1545 dt_read_lock(env, mo, 0);
1546 if (!dt_object_exists(mo))
1547 GOTO(unlock, rc = -ENOENT);
1549 /* parse remote buffers to local buffers and prepare the latter */
1550 lnbs = (len >> PAGE_SHIFT) + 1;
1551 OBD_ALLOC_PTR_ARRAY(lnb, lnbs);
1553 GOTO(unlock, rc = -ENOMEM);
1555 rc = dt_bufs_get(env, mo, rnb, lnb, lnbs, 0);
1556 if (unlikely(rc < 0))
1558 LASSERT(rc <= lnbs);
1560 rc = dt_read_prep(env, mo, lnb, nr_local);
1563 /* copy data to the buffer finally */
1564 for (i = 0; i < nr_local; i++) {
1565 char *p = kmap(lnb[i].lnb_page);
1568 LASSERT(lnb[i].lnb_page_offset == 0);
1569 off = lnb[i].lnb_len & ~PAGE_MASK;
1571 memset(p + off, 0, PAGE_SIZE - off);
1573 memcpy(buf + (i << PAGE_SHIFT), p, lnb[i].lnb_len);
1574 kunmap(lnb[i].lnb_page);
1575 copied += lnb[i].lnb_len;
1578 CDEBUG(D_INFO, "Read %i (wanted %u) bytes from %llu\n", copied,
1581 CWARN("%s: read %i bytes for "DFID
1582 " but wanted %u, is size wrong?\n",
1583 tsi->tsi_exp->exp_obd->obd_name, copied,
1584 PFID(&tsi->tsi_fid), len);
1585 /* Ignore partially copied data */
1590 dt_bufs_put(env, mo, lnb, nr_local);
1592 OBD_FREE_PTR_ARRAY(lnb, lnbs);
1594 dt_read_unlock(env, mo);
1595 lu_object_put(env, &mo->do_lu);
1597 rnb->rnb_len = copied;
1599 /* Don't fail OPEN request if read-on-open is failed, but drop
1600 * a message in log about the error.
1603 CDEBUG(D_INFO, "Read-on-open is failed, rc = %d", rc);
1609 * Completion AST for DOM discard locks:
1611 * CP AST an DOM discard lock is called always right after enqueue or from
1612 * reprocess if lock was blocked, in the latest case l_ast_data is set to
1613 * the mdt_object which is kept while there are pending locks on it.
1615 int ldlm_dom_discard_cp_ast(struct ldlm_lock *lock, __u64 flags, void *data)
1617 struct mdt_object *mo;
1618 struct lustre_handle dom_lh;
1623 /* l_ast_data is set when lock was not granted immediately
1624 * in mdt_dom_discard_data() below but put into waiting list,
1625 * so this CP callback means we are finished and corresponding
1626 * MDT object should be released finally as well as lock itself.
1628 lock_res_and_lock(lock);
1629 if (!lock->l_ast_data) {
1630 unlock_res_and_lock(lock);
1634 mo = lock->l_ast_data;
1635 lock->l_ast_data = NULL;
1636 unlock_res_and_lock(lock);
1638 ldlm_lock2handle(lock, &dom_lh);
1639 ldlm_lock_decref(&dom_lh, LCK_PW);
1641 env = lu_env_find();
1643 mdt_object_put(env, mo);
1648 void mdt_dom_discard_data(struct mdt_thread_info *info,
1649 struct mdt_object *mo)
1651 struct ptlrpc_request *req = mdt_info_req(info);
1652 struct mdt_device *mdt = mdt_dev(mo->mot_obj.lo_dev);
1653 union ldlm_policy_data policy;
1654 struct ldlm_res_id res_id;
1655 struct lustre_handle dom_lh;
1656 struct ldlm_lock *lock;
1657 __u64 flags = LDLM_FL_AST_DISCARD_DATA;
1663 if (req && req_is_replay(req))
1666 policy.l_inodebits.bits = MDS_INODELOCK_DOM;
1667 policy.l_inodebits.try_bits = 0;
1668 fid_build_reg_res_name(mdt_object_fid(mo), &res_id);
1670 /* Keep blocking version of discard for an old client to avoid
1671 * crashes on non-patched clients. LU-11359.
1673 old_client = req && !(exp_connect_flags2(req->rq_export) &
1674 OBD_CONNECT2_ASYNC_DISCARD);
1676 /* Tell the clients that the object is gone now and that they should
1677 * throw away any cached pages. */
1678 rc = ldlm_cli_enqueue_local(info->mti_env, mdt->mdt_namespace, &res_id,
1679 LDLM_IBITS, &policy, LCK_PW, &flags,
1680 ldlm_blocking_ast, old_client ?
1681 ldlm_completion_ast :
1682 ldlm_dom_discard_cp_ast,
1683 NULL, NULL, 0, LVB_T_NONE, NULL, &dom_lh);
1684 if (rc != ELDLM_OK) {
1686 "Failed to issue discard lock, rc = %d\n", rc);
1690 lock = ldlm_handle2lock(&dom_lh);
1691 lock_res_and_lock(lock);
1692 /* if lock is not granted then there are BL ASTs in progress and
1693 * lock will be granted in result of reprocessing with CP callback
1694 * notifying about that. The mdt object has to be kept until that and
1695 * it is saved in l_ast_data of the lock. Lock reference is kept too
1696 * until that to prevent it from canceling.
1698 if (!is_granted_or_cancelled_nolock(lock)) {
1699 mdt_object_get(info->mti_env, mo);
1700 lock->l_ast_data = mo;
1701 unlock_res_and_lock(lock);
1703 unlock_res_and_lock(lock);
1704 ldlm_lock_decref_and_cancel(&dom_lh, LCK_PW);
1706 LDLM_LOCK_PUT(lock);