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 <linux/falloc.h>
36 #include "mdt_internal.h"
38 /* functions below are stubs for now, they will be implemented with
39 * grant support on MDT */
40 static inline void mdt_dom_read_lock(struct mdt_object *mo)
42 down_read(&mo->mot_dom_sem);
45 static inline void mdt_dom_read_unlock(struct mdt_object *mo)
47 up_read(&mo->mot_dom_sem);
50 static inline void mdt_dom_write_lock(struct mdt_object *mo)
52 down_write(&mo->mot_dom_sem);
55 static inline void mdt_dom_write_unlock(struct mdt_object *mo)
57 up_write(&mo->mot_dom_sem);
60 static void mdt_dom_resource_prolong(struct ldlm_prolong_args *arg)
62 struct ldlm_resource *res;
63 struct ldlm_lock *lock;
67 res = ldlm_resource_get(arg->lpa_export->exp_obd->obd_namespace, NULL,
68 &arg->lpa_resid, LDLM_IBITS, 0);
71 "Failed to get resource for resid %llu/%llu\n",
72 arg->lpa_resid.name[0], arg->lpa_resid.name[1]);
77 list_for_each_entry(lock, &res->lr_granted, l_res_link) {
78 if (ldlm_has_dom(lock)) {
79 LDLM_DEBUG(lock, "DOM lock to prolong ");
80 ldlm_lock_prolong_one(lock, arg);
81 /* only one PW or EX lock can be granted,
82 * no need to continue search
84 if (lock->l_granted_mode & (LCK_PW | LCK_EX))
89 ldlm_resource_putref(res);
94 static void mdt_prolong_dom_lock(struct tgt_session_info *tsi,
95 struct ldlm_prolong_args *data)
97 struct obdo *oa = &tsi->tsi_ost_body->oa;
98 struct ldlm_lock *lock;
102 data->lpa_timeout = prolong_timeout(tgt_ses_req(tsi));
103 data->lpa_export = tsi->tsi_exp;
104 data->lpa_resid = tsi->tsi_resid;
106 CDEBUG(D_RPCTRACE, "Prolong DOM lock for req %p with x%llu\n",
107 tgt_ses_req(tsi), tgt_ses_req(tsi)->rq_xid);
109 if (oa->o_valid & OBD_MD_FLHANDLE) {
110 /* mostly a request should be covered by only one lock, try
112 lock = ldlm_handle2lock(&oa->o_handle);
114 LASSERT(lock->l_export == data->lpa_export);
115 ldlm_lock_prolong_one(lock, data);
116 lock->l_last_used = ktime_get();
118 if (data->lpa_locks_cnt > 0)
122 mdt_dom_resource_prolong(data);
126 static int mdt_rw_hpreq_lock_match(struct ptlrpc_request *req,
127 struct ldlm_lock *lock)
129 struct obd_ioobj *ioo;
131 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
135 if (!(lock->l_policy_data.l_inodebits.bits & MDS_INODELOCK_DOM))
138 ioo = req_capsule_client_get(&req->rq_pill, &RMF_OBD_IOOBJ);
139 LASSERT(ioo != NULL);
141 LASSERT(lock->l_resource != NULL);
142 if (!fid_res_name_eq(&ioo->ioo_oid.oi_fid, &lock->l_resource->lr_name))
145 /* a bulk write can only hold a reference on a PW extent lock. */
146 mode = LCK_PW | LCK_GROUP;
148 /* whereas a bulk read can be protected by either a PR or PW
152 if (!(lock->l_granted_mode & mode))
158 static int mdt_rw_hpreq_check(struct ptlrpc_request *req)
160 struct tgt_session_info *tsi;
161 struct obd_ioobj *ioo;
162 struct niobuf_remote *rnb;
164 struct ldlm_prolong_args pa = { 0 };
168 /* Don't use tgt_ses_info() to get session info, because lock_match()
169 * can be called while request has no processing thread yet. */
170 tsi = lu_context_key_get(&req->rq_session, &tgt_session_key);
173 * Use LASSERT below because malformed RPCs should have
174 * been filtered out in tgt_hpreq_handler().
176 opc = lustre_msg_get_opc(req->rq_reqmsg);
177 LASSERT(opc == OST_READ || opc == OST_WRITE);
179 ioo = req_capsule_client_get(&req->rq_pill, &RMF_OBD_IOOBJ);
180 LASSERT(ioo != NULL);
182 rnb = req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE);
183 LASSERT(rnb != NULL);
184 LASSERT(!(rnb->rnb_flags & OBD_BRW_SRVLOCK));
186 pa.lpa_mode = LCK_PW | LCK_GROUP;
188 pa.lpa_mode |= LCK_PR;
190 DEBUG_REQ(D_RPCTRACE, req, "%s %s: refresh rw locks for " DFID,
191 tgt_name(tsi->tsi_tgt), current->comm, PFID(&tsi->tsi_fid));
193 mdt_prolong_dom_lock(tsi, &pa);
195 if (pa.lpa_blocks_cnt > 0) {
197 "%s: refreshed %u locks timeout for req %p\n",
198 tgt_name(tsi->tsi_tgt), pa.lpa_blocks_cnt, req);
202 RETURN(pa.lpa_locks_cnt > 0 ? 0 : -ESTALE);
205 static void mdt_rw_hpreq_fini(struct ptlrpc_request *req)
207 mdt_rw_hpreq_check(req);
210 static struct ptlrpc_hpreq_ops mdt_hpreq_rw = {
211 .hpreq_lock_match = mdt_rw_hpreq_lock_match,
212 .hpreq_check = mdt_rw_hpreq_check,
213 .hpreq_fini = mdt_rw_hpreq_fini
217 * Assign high priority operations to an IO request.
219 * Check if the incoming request is a candidate for
220 * high-priority processing. If it is, assign it a high
221 * priority operations table.
223 * \param[in] tsi target session environment for this request
225 void mdt_hp_brw(struct tgt_session_info *tsi)
227 struct niobuf_remote *rnb;
228 struct obd_ioobj *ioo;
232 ioo = req_capsule_client_get(tsi->tsi_pill, &RMF_OBD_IOOBJ);
233 LASSERT(ioo != NULL); /* must exist after request preprocessing */
234 if (ioo->ioo_bufcnt > 0) {
235 rnb = req_capsule_client_get(tsi->tsi_pill, &RMF_NIOBUF_REMOTE);
236 LASSERT(rnb != NULL); /* must exist after preprocessing */
238 /* no high priority if server lock is needed */
239 if (rnb->rnb_flags & OBD_BRW_SRVLOCK ||
240 (lustre_msg_get_flags(tgt_ses_req(tsi)->rq_reqmsg) &
244 tgt_ses_req(tsi)->rq_ops = &mdt_hpreq_rw;
247 static int mdt_punch_hpreq_lock_match(struct ptlrpc_request *req,
248 struct ldlm_lock *lock)
250 struct tgt_session_info *tsi;
255 /* Don't use tgt_ses_info() to get session info, because lock_match()
256 * can be called while request has no processing thread yet. */
257 tsi = lu_context_key_get(&req->rq_session, &tgt_session_key);
260 * Use LASSERT below because malformed RPCs should have
261 * been filtered out in tgt_hpreq_handler().
263 LASSERT(tsi->tsi_ost_body != NULL);
264 if (tsi->tsi_ost_body->oa.o_valid & OBD_MD_FLHANDLE &&
265 tsi->tsi_ost_body->oa.o_handle.cookie == lock->l_handle.h_cookie)
268 oa = &tsi->tsi_ost_body->oa;
270 LASSERT(lock->l_resource != NULL);
271 if (!fid_res_name_eq(&oa->o_oi.oi_fid, &lock->l_resource->lr_name))
274 if (!(lock->l_granted_mode & (LCK_PW | LCK_GROUP)))
281 * Implementation of ptlrpc_hpreq_ops::hpreq_lock_check for OST_PUNCH request.
283 * High-priority queue request check for whether the given punch request
284 * (\a req) is blocking an LDLM lock cancel. Also checks whether the request is
285 * covered by an LDLM lock.
289 * \param[in] req the incoming request
291 * \retval 1 if \a req is blocking an LDLM lock cancel
292 * \retval 0 if it is not
293 * \retval -ESTALE if lock is not found
295 static int mdt_punch_hpreq_check(struct ptlrpc_request *req)
297 struct tgt_session_info *tsi;
299 struct ldlm_prolong_args pa = { 0 };
303 /* Don't use tgt_ses_info() to get session info, because lock_match()
304 * can be called while request has no processing thread yet. */
305 tsi = lu_context_key_get(&req->rq_session, &tgt_session_key);
306 LASSERT(tsi != NULL);
307 oa = &tsi->tsi_ost_body->oa;
309 LASSERT(!(oa->o_valid & OBD_MD_FLFLAGS &&
310 oa->o_flags & OBD_FL_SRVLOCK));
312 pa.lpa_mode = LCK_PW | LCK_GROUP;
314 CDEBUG(D_DLMTRACE, "%s: refresh DOM lock for "DFID"\n",
315 tgt_name(tsi->tsi_tgt), PFID(&tsi->tsi_fid));
317 mdt_prolong_dom_lock(tsi, &pa);
319 if (pa.lpa_blocks_cnt > 0) {
321 "%s: refreshed %u locks timeout for req %p.\n",
322 tgt_name(tsi->tsi_tgt), pa.lpa_blocks_cnt, req);
326 RETURN(pa.lpa_locks_cnt > 0 ? 0 : -ESTALE);
330 * Implementation of ptlrpc_hpreq_ops::hpreq_lock_fini for OST_PUNCH request.
332 * Called after the request has been handled. It refreshes lock timeout again
333 * so that client has more time to send lock cancel RPC.
335 * \param[in] req request which is being processed.
337 static void mdt_punch_hpreq_fini(struct ptlrpc_request *req)
339 mdt_punch_hpreq_check(req);
342 static struct ptlrpc_hpreq_ops mdt_hpreq_punch = {
343 .hpreq_lock_match = mdt_punch_hpreq_lock_match,
344 .hpreq_check = mdt_punch_hpreq_check,
345 .hpreq_fini = mdt_punch_hpreq_fini
348 void mdt_hp_punch(struct tgt_session_info *tsi)
350 LASSERT(tsi->tsi_ost_body != NULL); /* must exists if we are here */
351 /* no high-priority if server lock is needed */
352 if ((tsi->tsi_ost_body->oa.o_valid & OBD_MD_FLFLAGS &&
353 tsi->tsi_ost_body->oa.o_flags & OBD_FL_SRVLOCK) ||
354 tgt_conn_flags(tsi) & OBD_CONNECT_MDS ||
355 lustre_msg_get_flags(tgt_ses_req(tsi)->rq_reqmsg) & MSG_REPLAY)
357 tgt_ses_req(tsi)->rq_ops = &mdt_hpreq_punch;
360 static int mdt_preprw_read(const struct lu_env *env, struct obd_export *exp,
361 struct mdt_device *mdt, struct mdt_object *mo,
362 struct lu_attr *la, int niocount,
363 struct niobuf_remote *rnb, int *nr_local,
364 struct niobuf_local *lnb, char *jobid)
366 struct tgt_session_info *tsi = tgt_ses_info(env);
367 struct ptlrpc_request *req = tgt_ses_req(tsi);
368 struct dt_object *dob;
369 int i, j, rc, tot_bytes = 0;
370 int maxlnb = *nr_local;
375 mdt_dom_read_lock(mo);
377 /* the only valid case when READ can find object is missing or stale
378 * when export is just evicted and open files are closed forcefully
379 * on server while client's READ can be in progress.
380 * This should not happen on healthy export, object can't be missing
381 * or dying because both states means it was finally destroyed.
383 level = exp->exp_failed ? D_INFO : D_ERROR;
384 if (!mdt_object_exists(mo)) {
386 "%s: READ IO to missing obj "DFID": rc = %d\n",
387 exp->exp_obd->obd_name, PFID(mdt_object_fid(mo)),
389 /* return 0 and continue with empty commit to skip such READ
390 * without more BRW errors.
394 if (lu_object_is_dying(&mo->mot_header)) {
396 "%s: READ IO to stale obj "DFID": rc = %d\n",
397 exp->exp_obd->obd_name, PFID(mdt_object_fid(mo)),
399 /* return 0 and continue with empty commit to skip such READ
400 * without more BRW errors.
405 dob = mdt_obj2dt(mo);
406 /* parse remote buffers to local buffers and prepare the latter */
407 for (i = 0, j = 0; i < niocount; i++) {
408 rc = dt_bufs_get(env, dob, rnb + i, lnb + j, maxlnb, 0);
409 if (unlikely(rc < 0))
411 /* correct index for local buffers to continue with */
415 tot_bytes += rnb[i].rnb_len;
418 rc = dt_attr_get(env, dob, la);
422 rc = dt_read_prep(env, dob, lnb, *nr_local);
426 mdt_counter_incr(req, LPROC_MDT_IO_READ, tot_bytes);
429 dt_bufs_put(env, dob, lnb, *nr_local);
430 mdt_dom_read_unlock(mo);
434 static int mdt_preprw_write(const struct lu_env *env, struct obd_export *exp,
435 struct mdt_device *mdt, struct mdt_object *mo,
436 struct lu_attr *la, struct obdo *oa,
437 int objcount, struct obd_ioobj *obj,
438 struct niobuf_remote *rnb, int *nr_local,
439 struct niobuf_local *lnb, char *jobid)
441 struct tgt_session_info *tsi = tgt_ses_info(env);
442 struct ptlrpc_request *req = tgt_ses_req(tsi);
443 struct dt_object *dob;
444 int i, j, k, rc = 0, tot_bytes = 0;
445 int maxlnb = *nr_local;
449 /* Process incoming grant info, set OBD_BRW_GRANTED flag and grant some
450 * space back if possible */
451 tgt_grant_prepare_write(env, exp, oa, rnb, obj->ioo_bufcnt);
453 mdt_dom_read_lock(mo);
455 /* don't report error in cases with failed export */
456 if (!mdt_object_exists(mo)) {
457 int level = exp->exp_failed ? D_INFO : D_ERROR;
461 "%s: WRITE IO to missing obj "DFID": rc = %d\n",
462 exp->exp_obd->obd_name, PFID(mdt_object_fid(mo)),
464 /* exit with no data written, note nr_local = 0 above */
467 if (lu_object_is_dying(&mo->mot_header)) {
468 /* This is possible race between object destroy followed by
469 * discard BL AST and client cache flushing. Object is
470 * referenced until discard finish.
472 CDEBUG(D_INODE, "WRITE IO to stale object "DFID"\n",
473 PFID(mdt_object_fid(mo)));
474 /* Note: continue with no error here to don't cause BRW errors
475 * but skip transaction in commitrw silently so no data is
480 dob = mdt_obj2dt(mo);
481 /* parse remote buffers to local buffers and prepare the latter */
482 for (i = 0, j = 0; i < obj->ioo_bufcnt; i++) {
483 rc = dt_bufs_get(env, dob, rnb + i, lnb + j, maxlnb, 1);
484 if (unlikely(rc < 0))
486 /* correct index for local buffers to continue with */
487 for (k = 0; k < rc; k++) {
488 lnb[j + k].lnb_flags = rnb[i].rnb_flags;
489 if (!(rnb[i].rnb_flags & OBD_BRW_GRANTED))
490 lnb[j + k].lnb_rc = -ENOSPC;
495 tot_bytes += rnb[i].rnb_len;
498 rc = dt_write_prep(env, dob, lnb, *nr_local);
502 mdt_counter_incr(req, LPROC_MDT_IO_WRITE, tot_bytes);
505 dt_bufs_put(env, dob, lnb, *nr_local);
507 mdt_dom_read_unlock(mo);
508 /* tgt_grant_prepare_write() was called, so we must commit */
509 tgt_grant_commit(exp, oa->o_grant_used, rc);
510 /* let's still process incoming grant information packed in the oa,
511 * but without enforcing grant since we won't proceed with the write.
512 * Just like a read request actually. */
513 tgt_grant_prepare_read(env, exp, oa);
517 int mdt_obd_preprw(const struct lu_env *env, int cmd, struct obd_export *exp,
518 struct obdo *oa, int objcount, struct obd_ioobj *obj,
519 struct niobuf_remote *rnb, int *nr_local,
520 struct niobuf_local *lnb)
522 struct tgt_session_info *tsi = tgt_ses_info(env);
523 struct mdt_thread_info *info = tsi2mdt_info(tsi);
524 struct lu_attr *la = &info->mti_attr.ma_attr;
525 struct mdt_device *mdt = mdt_dev(exp->exp_obd->obd_lu_dev);
526 struct mdt_object *mo;
530 /* The default value PTLRPC_MAX_BRW_PAGES is set in tgt_brw_write()
531 * but for MDT it is different, correct it here. */
532 if (*nr_local > MD_MAX_BRW_PAGES)
533 *nr_local = MD_MAX_BRW_PAGES;
535 jobid = tsi->tsi_jobid;
537 if (!oa || objcount != 1 || obj->ioo_bufcnt == 0) {
538 CERROR("%s: bad parameters %p/%i/%i\n",
539 exp->exp_obd->obd_name, oa, objcount, obj->ioo_bufcnt);
543 mo = mdt_object_find(env, mdt, &tsi->tsi_fid);
545 GOTO(out, rc = PTR_ERR(mo));
547 LASSERT(info->mti_object == NULL);
548 info->mti_object = mo;
550 if (cmd == OBD_BRW_WRITE) {
551 la_from_obdo(la, oa, OBD_MD_FLGETATTR);
552 rc = mdt_preprw_write(env, exp, mdt, mo, la, oa,
553 objcount, obj, rnb, nr_local, lnb,
555 } else if (cmd == OBD_BRW_READ) {
556 tgt_grant_prepare_read(env, exp, oa);
557 rc = mdt_preprw_read(env, exp, mdt, mo, la,
558 obj->ioo_bufcnt, rnb, nr_local, lnb,
560 obdo_from_la(oa, la, LA_ATIME);
562 CERROR("%s: wrong cmd %d received!\n",
563 exp->exp_obd->obd_name, cmd);
567 lu_object_put(env, &mo->mot_obj);
568 info->mti_object = NULL;
574 static int mdt_commitrw_read(const struct lu_env *env, struct mdt_device *mdt,
575 struct mdt_object *mo, int objcount, int niocount,
576 struct niobuf_local *lnb)
578 struct dt_object *dob;
583 dob = mdt_obj2dt(mo);
586 dt_bufs_put(env, dob, lnb, niocount);
588 mdt_dom_read_unlock(mo);
592 static int mdt_commitrw_write(const struct lu_env *env, struct obd_export *exp,
593 struct mdt_device *mdt, struct mdt_object *mo,
594 struct lu_attr *la, struct obdo *oa, int objcount,
595 int niocount, struct niobuf_local *lnb,
596 unsigned long granted, int old_rc)
598 struct dt_device *dt = mdt->mdt_bottom;
599 struct dt_object *dob;
607 dob = mdt_obj2dt(mo);
610 GOTO(out, rc = old_rc);
612 la->la_valid &= LA_ATIME | LA_MTIME | LA_CTIME;
614 if (!dt_object_exists(dob))
615 GOTO(out, rc = -ENOENT);
619 DEBUG_REQ(D_WARNING, tgt_ses_req(tgt_ses_info(env)),
620 "%s: commit with no pages for "DFID": rc = %d\n",
621 exp->exp_obd->obd_name, PFID(mdt_object_fid(mo)), rc);
625 CFS_FAIL_TIMEOUT(OBD_FAIL_MDS_COMMITRW_DELAY, cfs_fail_val);
627 th = dt_trans_create(env, dt);
629 GOTO(out, rc = PTR_ERR(th));
631 for (i = 0; i < niocount; i++) {
632 if (!(lnb[i].lnb_flags & OBD_BRW_ASYNC)) {
638 if (OBD_FAIL_CHECK(OBD_FAIL_OST_DQACQ_NET))
639 GOTO(out_stop, rc = -EINPROGRESS);
641 rc = dt_declare_write_commit(env, dob, lnb, niocount, th);
646 /* update [mac]time if needed */
647 rc = dt_declare_attr_set(env, dob, la, th);
652 tgt_vbr_obj_set(env, dob);
653 rc = dt_trans_start(env, dt, th);
657 dt_write_lock(env, dob, 0);
658 if (lu_object_is_dying(&mo->mot_header)) {
659 /* Commit to stale object can be just skipped silently. */
660 CDEBUG(D_INODE, "skip commit to stale object "DFID"\n",
661 PFID(mdt_object_fid(mo)));
662 GOTO(unlock, rc = 0);
664 rc = dt_write_commit(env, dob, lnb, niocount, th, oa->o_size);
666 restart = th->th_restart_tran;
671 rc = dt_attr_set(env, dob, la, th);
675 /* get attr to return */
676 rc = dt_attr_get(env, dob, la);
678 dt_write_unlock(env, dob);
681 /* Force commit to make the just-deleted blocks
682 * reusable. LU-456 */
687 if (rc == 0 && granted > 0) {
688 if (tgt_grant_commit_cb_add(th, exp, granted) == 0)
692 th->th_result = restart ? 0 : rc;
693 dt_trans_stop(env, dt, th);
694 if (rc == -ENOSPC && retries++ < 3) {
695 CDEBUG(D_INODE, "retry after force commit, retries:%d\n",
702 if (retries % 10000 == 0)
703 CERROR("%s: restart IO write too many times: %d\n",
704 exp->exp_obd->obd_name, retries);
705 CDEBUG(D_INODE, "retry transaction, retries:%d\n",
711 dt_bufs_put(env, dob, lnb, niocount);
712 mdt_dom_read_unlock(mo);
714 tgt_grant_commit(exp, granted, old_rc);
718 void mdt_dom_obj_lvb_update(const struct lu_env *env, struct mdt_object *mo,
719 struct obdo *oa, bool increase_only)
721 struct mdt_device *mdt = mdt_dev(mo->mot_obj.lo_dev);
722 struct ldlm_res_id resid;
723 struct ldlm_resource *res;
725 fid_build_reg_res_name(mdt_object_fid(mo), &resid);
726 res = ldlm_resource_get(mdt->mdt_namespace, NULL, &resid,
731 /* Update lvbo data if exists. */
732 if (mdt_dom_lvb_is_valid(res)) {
733 mdt_dom_disk_lvbo_update(env, mo, res, increase_only);
735 struct ost_lvb *res_lvb = res->lr_lvb_data;
738 oa->o_valid |= OBD_MD_FLBLOCKS | OBD_MD_FLSIZE |
739 OBD_MD_FLMTIME | OBD_MD_FLATIME |
741 oa->o_blocks = res_lvb->lvb_blocks;
742 oa->o_size = res_lvb->lvb_size;
743 oa->o_atime = res_lvb->lvb_atime;
744 oa->o_mtime = res_lvb->lvb_mtime;
745 oa->o_ctime = res_lvb->lvb_ctime;
749 ldlm_resource_putref(res);
752 int mdt_obd_commitrw(const struct lu_env *env, int cmd, struct obd_export *exp,
753 struct obdo *oa, int objcount, struct obd_ioobj *obj,
754 struct niobuf_remote *rnb, int npages,
755 struct niobuf_local *lnb, int old_rc)
757 struct mdt_thread_info *info = mdt_th_info(env);
758 struct mdt_device *mdt = mdt_dev(exp->exp_obd->obd_lu_dev);
759 struct mdt_object *mo = info->mti_object;
760 struct lu_attr *la = &info->mti_attr.ma_attr;
766 if (cmd == OBD_BRW_WRITE) {
767 /* Don't update timestamps if this write is older than a
768 * setattr which modifies the timestamps. b=10150 */
770 /* XXX when we start having persistent reservations this needs
771 * to be changed to ofd_fmd_get() to create the fmd if it
772 * doesn't already exist so we can store the reservation handle
774 valid = OBD_MD_FLUID | OBD_MD_FLGID;
775 if (tgt_fmd_check(exp, mdt_object_fid(mo),
776 mdt_info_req(info)->rq_xid))
777 valid |= OBD_MD_FLATIME | OBD_MD_FLMTIME |
780 la_from_obdo(la, oa, valid);
782 rc = mdt_commitrw_write(env, exp, mdt, mo, la, oa, objcount,
783 npages, lnb, oa->o_grant_used, old_rc);
785 obdo_from_la(oa, la, VALID_FLAGS | LA_GID | LA_UID);
787 obdo_from_la(oa, la, LA_GID | LA_UID);
789 mdt_dom_obj_lvb_update(env, mo, NULL, false);
790 /* don't report overquota flag if we failed before reaching
792 if (old_rc == 0 && (rc == 0 || rc == -EDQUOT)) {
793 /* return the overquota flags to client */
794 if (lnb[0].lnb_flags & OBD_BRW_OVER_USRQUOTA) {
795 if (oa->o_valid & OBD_MD_FLFLAGS)
796 oa->o_flags |= OBD_FL_NO_USRQUOTA;
798 oa->o_flags = OBD_FL_NO_USRQUOTA;
801 if (lnb[0].lnb_flags & OBD_BRW_OVER_GRPQUOTA) {
802 if (oa->o_valid & OBD_MD_FLFLAGS)
803 oa->o_flags |= OBD_FL_NO_GRPQUOTA;
805 oa->o_flags = OBD_FL_NO_GRPQUOTA;
808 if (lnb[0].lnb_flags & OBD_BRW_OVER_PRJQUOTA) {
809 if (oa->o_valid & OBD_MD_FLFLAGS)
810 oa->o_flags |= OBD_FL_NO_PRJQUOTA;
812 oa->o_flags = OBD_FL_NO_PRJQUOTA;
815 oa->o_valid |= OBD_MD_FLFLAGS | OBD_MD_FLUSRQUOTA |
816 OBD_MD_FLGRPQUOTA | OBD_MD_FLPRJQUOTA;
818 } else if (cmd == OBD_BRW_READ) {
819 /* If oa != NULL then mdt_preprw_read updated the inode
820 * atime and we should update the lvb so that other glimpses
821 * will also get the updated value. bug 5972 */
823 mdt_dom_obj_lvb_update(env, mo, NULL, true);
824 rc = mdt_commitrw_read(env, mdt, mo, objcount, npages, lnb);
830 mdt_thread_info_fini(info);
834 int mdt_object_fallocate(const struct lu_env *env, struct dt_device *dt,
835 struct dt_object *dob, __u64 start, __u64 end,
836 int mode, struct lu_attr *la)
843 if (!dt_object_exists(dob))
846 th = dt_trans_create(env, dt);
850 rc = dt_declare_attr_set(env, dob, la, th);
854 rc = dt_declare_fallocate(env, dob, start, end, mode, th);
858 tgt_vbr_obj_set(env, dob);
859 rc = dt_trans_start(env, dt, th);
863 dt_write_lock(env, dob, 0);
864 rc = dt_falloc(env, dob, start, end, mode, th);
867 rc = dt_attr_set(env, dob, la, th);
871 dt_write_unlock(env, dob);
874 dt_trans_stop(env, dt, th);
879 * MDT request handler for OST_FALLOCATE RPC.
881 * This is part of request processing. Validate request fields,
882 * preallocate the given MDT object and pack reply.
884 * \param[in] tsi target session environment for this request
886 * \retval 0 if successful
887 * \retval negative value on error
889 int mdt_fallocate_hdl(struct tgt_session_info *tsi)
891 struct obdo *oa = &tsi->tsi_ost_body->oa;
892 struct ptlrpc_request *req = tgt_ses_req(tsi);
893 struct ost_body *repbody;
894 struct mdt_thread_info *info;
895 struct ldlm_namespace *ns = tsi->tsi_tgt->lut_obd->obd_namespace;
896 struct obd_export *exp = tsi->tsi_exp;
897 struct mdt_device *mdt = mdt_dev(exp->exp_obd->obd_lu_dev);
898 struct mdt_object *mo;
899 struct dt_object *dob;
902 struct lustre_handle lh = { 0, };
906 ktime_t kstart = ktime_get();
908 repbody = req_capsule_server_get(tsi->tsi_pill, &RMF_OST_BODY);
910 RETURN(err_serious(-ENOMEM));
913 * fallocate start and end are passed in o_size, o_blocks
916 if ((oa->o_valid & (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS)) !=
917 (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS))
918 RETURN(err_serious(-EPROTO));
922 mode = oa->o_falloc_mode;
925 "fallocate: "DFID", mode = %#x, start = %lld, end = %lld\n",
926 PFID(&tsi->tsi_fid), mode, start, end);
929 * mode == 0 (which is standard prealloc) and PUNCH is supported
930 * Rest of mode options are not supported yet.
932 if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
935 if (mode & FALLOC_FL_PUNCH_HOLE && !(mode & FALLOC_FL_KEEP_SIZE)) {
936 CWARN("%s: PUNCH mode misses KEEP_SIZE flag, setting it\n",
937 tsi->tsi_tgt->lut_obd->obd_name);
938 mode |= FALLOC_FL_KEEP_SIZE;
941 info = tsi2mdt_info(tsi);
942 la = &info->mti_attr.ma_attr;
944 repbody->oa.o_oi = oa->o_oi;
945 repbody->oa.o_valid = OBD_MD_FLID;
947 srvlock = oa->o_valid & OBD_MD_FLFLAGS &&
948 oa->o_flags & OBD_FL_SRVLOCK;
951 rc = tgt_mdt_data_lock(ns, &tsi->tsi_resid, &lh, LCK_PW,
957 mo = mdt_object_find(tsi->tsi_env, mdt, &tsi->tsi_fid);
959 GOTO(out_unlock, rc = PTR_ERR(mo));
961 if (!mdt_object_exists(mo))
962 GOTO(out_put, rc = -ENOENT);
964 /* Shouldn't happen on dirs */
965 if (S_ISDIR(lu_object_attr(&mo->mot_obj))) {
967 CERROR("%s: fallocate on dir "DFID": rc = %d\n",
968 exp->exp_obd->obd_name, PFID(&tsi->tsi_fid), rc);
972 la_from_obdo(la, oa, OBD_MD_FLMTIME | OBD_MD_FLATIME | OBD_MD_FLCTIME);
974 mdt_dom_write_lock(mo);
975 dob = mdt_obj2dt(mo);
977 if (la->la_valid & (LA_ATIME | LA_MTIME | LA_CTIME))
978 tgt_fmd_update(tsi->tsi_exp, &tsi->tsi_fid,
979 tgt_ses_req(tsi)->rq_xid);
981 rc = mdt_object_fallocate(tsi->tsi_env, mdt->mdt_bottom, dob, start,
983 mdt_dom_write_unlock(mo);
987 mdt_dom_obj_lvb_update(tsi->tsi_env, mo, &repbody->oa, false);
989 mdt_counter_incr(req, LPROC_MDT_FALLOCATE,
990 ktime_us_delta(ktime_get(), kstart));
994 lu_object_put(tsi->tsi_env, &mo->mot_obj);
997 tgt_data_unlock(&lh, LCK_PW);
999 mdt_thread_info_fini(info);
1003 int mdt_object_punch(const struct lu_env *env, struct dt_device *dt,
1004 struct dt_object *dob, __u64 start, __u64 end,
1012 /* we support truncate, not punch yet */
1013 LASSERT(end == OBD_OBJECT_EOF);
1015 if (!dt_object_exists(dob))
1018 th = dt_trans_create(env, dt);
1020 RETURN(PTR_ERR(th));
1022 rc = dt_declare_attr_set(env, dob, la, th);
1026 rc = dt_declare_punch(env, dob, start, OBD_OBJECT_EOF, th);
1030 tgt_vbr_obj_set(env, dob);
1031 rc = dt_trans_start(env, dt, th);
1035 dt_write_lock(env, dob, 0);
1036 rc = dt_punch(env, dob, start, OBD_OBJECT_EOF, th);
1039 rc = dt_attr_set(env, dob, la, th);
1043 dt_write_unlock(env, dob);
1046 dt_trans_stop(env, dt, th);
1050 int mdt_punch_hdl(struct tgt_session_info *tsi)
1052 const struct obdo *oa = &tsi->tsi_ost_body->oa;
1053 struct ptlrpc_request *req = tgt_ses_req(tsi);
1054 struct ost_body *repbody;
1055 struct mdt_thread_info *info;
1057 struct ldlm_namespace *ns = tsi->tsi_tgt->lut_obd->obd_namespace;
1058 struct obd_export *exp = tsi->tsi_exp;
1059 struct mdt_device *mdt = mdt_dev(exp->exp_obd->obd_lu_dev);
1060 struct mdt_object *mo;
1061 struct dt_object *dob;
1063 struct lustre_handle lh = { 0, };
1064 ktime_t kstart = ktime_get();
1071 if ((oa->o_valid & (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS)) !=
1072 (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS))
1073 RETURN(err_serious(-EPROTO));
1075 repbody = req_capsule_server_get(tsi->tsi_pill, &RMF_OST_BODY);
1076 if (repbody == NULL)
1077 RETURN(err_serious(-ENOMEM));
1079 /* punch start,end are passed in o_size,o_blocks throught wire */
1083 if (end != OBD_OBJECT_EOF) /* Only truncate is supported */
1086 info = tsi2mdt_info(tsi);
1087 la = &info->mti_attr.ma_attr;
1088 /* standard truncate optimization: if file body is completely
1089 * destroyed, don't send data back to the server. */
1091 flags |= LDLM_FL_AST_DISCARD_DATA;
1093 repbody->oa.o_oi = oa->o_oi;
1094 repbody->oa.o_valid = OBD_MD_FLID;
1096 srvlock = (exp_connect_flags(exp) & OBD_CONNECT_SRVLOCK) &&
1097 oa->o_valid & OBD_MD_FLFLAGS &&
1098 oa->o_flags & OBD_FL_SRVLOCK;
1101 rc = tgt_mdt_data_lock(ns, &tsi->tsi_resid, &lh, LCK_PW,
1107 CDEBUG(D_INODE, "calling punch for object "DFID", valid = %#llx"
1108 ", start = %lld, end = %lld\n", PFID(&tsi->tsi_fid),
1109 oa->o_valid, start, end);
1111 mo = mdt_object_find(tsi->tsi_env, mdt, &tsi->tsi_fid);
1113 GOTO(out_unlock, rc = PTR_ERR(mo));
1115 if (!mdt_object_exists(mo))
1116 GOTO(out_put, rc = -ENOENT);
1118 /* Shouldn't happen on dirs */
1119 if (S_ISDIR(lu_object_attr(&mo->mot_obj))) {
1121 CERROR("%s: Truncate on dir "DFID": rc = %d\n",
1122 exp->exp_obd->obd_name, PFID(&tsi->tsi_fid), rc);
1126 mdt_dom_write_lock(mo);
1127 dob = mdt_obj2dt(mo);
1129 la_from_obdo(la, oa, OBD_MD_FLMTIME | OBD_MD_FLATIME | OBD_MD_FLCTIME);
1130 la->la_size = start;
1131 la->la_valid |= LA_SIZE;
1133 /* MDT supports FMD for Data-on-MDT needs */
1134 if (la->la_valid & (LA_ATIME | LA_MTIME | LA_CTIME))
1135 tgt_fmd_update(tsi->tsi_exp, &tsi->tsi_fid,
1136 tgt_ses_req(tsi)->rq_xid);
1138 rc = mdt_object_punch(tsi->tsi_env, mdt->mdt_bottom, dob,
1140 mdt_dom_write_unlock(mo);
1144 mdt_dom_obj_lvb_update(tsi->tsi_env, mo, &repbody->oa, false);
1146 mdt_counter_incr(req, LPROC_MDT_IO_PUNCH,
1147 ktime_us_delta(ktime_get(), kstart));
1150 lu_object_put(tsi->tsi_env, &mo->mot_obj);
1153 tgt_data_unlock(&lh, LCK_PW);
1155 mdt_thread_info_fini(info);
1160 * MDT glimpse for Data-on-MDT
1162 * If there is write lock on client then function issues glimpse_ast to get
1163 * an actual size from that client.
1166 int mdt_do_glimpse(const struct lu_env *env, struct ldlm_namespace *ns,
1167 struct ldlm_resource *res)
1169 union ldlm_policy_data policy;
1170 struct lustre_handle lockh;
1171 enum ldlm_mode mode;
1172 struct ldlm_lock *lock;
1173 struct ldlm_glimpse_work *gl_work;
1179 /* There can be only one write lock covering data, try to match it. */
1180 policy.l_inodebits.bits = MDS_INODELOCK_DOM;
1181 mode = ldlm_lock_match(ns, LDLM_FL_TEST_LOCK,
1182 &res->lr_name, LDLM_IBITS, &policy,
1185 /* There is no PW lock on this object; finished. */
1189 lock = ldlm_handle2lock(&lockh);
1194 * This check is for lock taken in mdt_reint_unlink() that does
1195 * not have l_glimpse_ast set. So the logic is: if there is a lock
1196 * with no l_glimpse_ast set, this object is being destroyed already.
1197 * Hence, if you are grabbing DLM locks on the server, always set
1198 * non-NULL glimpse_ast (e.g., ldlm_request.c::ldlm_glimpse_ast()).
1200 if (lock->l_glimpse_ast == NULL) {
1201 LDLM_DEBUG(lock, "no l_glimpse_ast");
1202 GOTO(out, rc = -ENOENT);
1205 OBD_SLAB_ALLOC_PTR_GFP(gl_work, ldlm_glimpse_work_kmem, GFP_ATOMIC);
1207 GOTO(out, rc = -ENOMEM);
1209 /* Populate the gl_work structure.
1210 * Grab additional reference on the lock which will be released in
1211 * ldlm_work_gl_ast_lock() */
1212 gl_work->gl_lock = LDLM_LOCK_GET(lock);
1213 /* The glimpse callback is sent to one single IO lock. As a result,
1214 * the gl_work list is just composed of one element */
1215 list_add_tail(&gl_work->gl_list, &gl_list);
1216 /* There is actually no need for a glimpse descriptor when glimpsing
1218 gl_work->gl_desc = NULL;
1219 /* the ldlm_glimpse_work structure is allocated on the stack */
1220 gl_work->gl_flags = LDLM_GL_WORK_SLAB_ALLOCATED;
1222 ldlm_glimpse_locks(res, &gl_list); /* this will update the LVB */
1224 /* If the list is not empty, we failed to glimpse a lock and
1225 * must clean it up. Usually due to a race with unlink.*/
1226 if (!list_empty(&gl_list)) {
1227 LDLM_LOCK_RELEASE(lock);
1228 OBD_SLAB_FREE_PTR(gl_work, ldlm_glimpse_work_kmem);
1233 LDLM_LOCK_PUT(lock);
1237 static void mdt_lvb2reply(struct ldlm_resource *res, struct mdt_body *mb,
1238 struct ost_lvb *lvb)
1240 struct ost_lvb *res_lvb;
1243 res_lvb = res->lr_lvb_data;
1248 mb->mbo_dom_size = res_lvb->lvb_size;
1249 mb->mbo_dom_blocks = res_lvb->lvb_blocks;
1250 mb->mbo_mtime = res_lvb->lvb_mtime;
1251 mb->mbo_ctime = res_lvb->lvb_ctime;
1252 mb->mbo_atime = res_lvb->lvb_atime;
1253 mb->mbo_valid |= OBD_MD_FLATIME | OBD_MD_FLCTIME |
1254 OBD_MD_FLMTIME | OBD_MD_DOM_SIZE;
1256 CDEBUG(D_DLMTRACE, "size %llu\n", res_lvb->lvb_size);
1261 * MDT glimpse for Data-on-MDT
1263 * This function is called when MDT get attributes for the DoM object.
1264 * If there is write lock on client then function issues glimpse_ast to get
1265 * an actual size from that client.
1267 int mdt_dom_object_size(const struct lu_env *env, struct mdt_device *mdt,
1268 const struct lu_fid *fid, struct mdt_body *mb,
1271 struct ldlm_res_id resid;
1272 struct ldlm_resource *res;
1277 fid_build_reg_res_name(fid, &resid);
1278 res = ldlm_resource_get(mdt->mdt_namespace, NULL, &resid,
1283 /* Update lvbo data if DoM lock returned or if LVB is not yet valid. */
1284 if (dom_lock || !mdt_dom_lvb_is_valid(res))
1285 mdt_dom_lvbo_update(res, NULL, NULL, false);
1287 mdt_lvb2reply(res, mb, NULL);
1288 ldlm_resource_putref(res);
1293 * MDT DoM lock intent policy (glimpse)
1295 * Intent policy is called when lock has an intent, for DoM file that
1296 * means glimpse lock and policy fills Lock Value Block (LVB).
1298 * If already granted lock is found it will be placed in \a lockp and
1299 * returned back to caller function.
1301 * \param[in] tsi session info
1302 * \param[in,out] lockp pointer to the lock
1303 * \param[in] flags LDLM flags
1305 * \retval ELDLM_LOCK_REPLACED if already granted lock was found
1306 * and placed in \a lockp
1307 * \retval ELDLM_LOCK_ABORTED in other cases except error
1308 * \retval negative value on error
1310 int mdt_glimpse_enqueue(struct mdt_thread_info *mti, struct ldlm_namespace *ns,
1311 struct ldlm_lock **lockp, __u64 flags)
1313 struct ldlm_lock *lock = *lockp;
1314 struct ldlm_resource *res = lock->l_resource;
1315 ldlm_processing_policy policy;
1316 struct ldlm_reply *rep;
1317 struct mdt_body *mbo;
1318 struct ost_lvb *lvb;
1319 bool old_client = !exp_connect_dom_lvb(mti->mti_exp);
1324 policy = ldlm_get_processing_policy(res);
1325 LASSERT(policy != NULL);
1327 if (unlikely(old_client)) {
1328 req_capsule_set_size(mti->mti_pill, &RMF_MDT_MD, RCL_SERVER, 0);
1329 req_capsule_set_size(mti->mti_pill, &RMF_ACL, RCL_SERVER, 0);
1331 req_capsule_set_size(mti->mti_pill, &RMF_DLM_LVB, RCL_SERVER,
1334 rc = req_capsule_server_pack(mti->mti_pill);
1336 RETURN(err_serious(rc));
1338 rep = req_capsule_server_get(mti->mti_pill, &RMF_DLM_REP);
1340 if (unlikely(old_client)) {
1341 mbo = req_capsule_server_get(mti->mti_pill, &RMF_MDT_BODY);
1345 lvb = req_capsule_server_get(mti->mti_pill, &RMF_DLM_LVB);
1351 /* Check if this is a resend case (MSG_RESENT is set on RPC) and a
1352 * lock was found by ldlm_handle_enqueue(); if so no need to grant
1354 if (flags & LDLM_FL_RESENT) {
1355 rc = LDLM_ITER_CONTINUE;
1357 __u64 tmpflags = LDLM_FL_BLOCK_NOWAIT;
1358 enum ldlm_error err;
1360 rc = policy(lock, &tmpflags, LDLM_PROCESS_RESCAN, &err, NULL);
1361 check_res_locked(res);
1365 /* The lock met with no resistance; we're finished. */
1366 if (rc == LDLM_ITER_CONTINUE) {
1367 GOTO(fill_mbo, rc = ELDLM_LOCK_REPLACED);
1368 } else if (flags & LDLM_FL_BLOCK_NOWAIT) {
1369 /* LDLM_FL_BLOCK_NOWAIT means it is for AGL. Do not send glimpse
1370 * callback for glimpse size. The real size user will trigger
1371 * the glimpse callback when necessary. */
1372 GOTO(fill_mbo, rc = ELDLM_LOCK_ABORTED);
1375 rc = mdt_do_glimpse(mti->mti_env, ns, res);
1376 if (rc == -ENOENT) {
1377 /* We are racing with unlink(); just return -ENOENT */
1378 rep->lock_policy_res2 = ptlrpc_status_hton(-ENOENT);
1379 } else if (rc == -EINVAL) {
1380 /* this is possible is client lock has been cancelled but
1381 * still exists on server. If that lock was found on server
1382 * as only conflicting lock then the client has already
1383 * size authority and glimpse is not needed. */
1384 CDEBUG(D_DLMTRACE, "Glimpse from the client owning lock\n");
1385 } else if (rc < 0) {
1388 rc = ELDLM_LOCK_ABORTED;
1390 /* LVB can be without valid data in case of DOM */
1391 if (!mdt_dom_lvb_is_valid(res))
1392 mdt_dom_lvbo_update(res, lock, NULL, false);
1393 mdt_lvb2reply(res, mbo, lvb);
1398 int mdt_brw_enqueue(struct mdt_thread_info *mti, struct ldlm_namespace *ns,
1399 struct ldlm_lock **lockp, __u64 flags)
1401 struct tgt_session_info *tsi = tgt_ses_info(mti->mti_env);
1402 struct lu_fid *fid = &tsi->tsi_fid;
1403 struct ldlm_lock *lock = *lockp;
1404 struct ldlm_resource *res = lock->l_resource;
1405 struct ldlm_reply *rep;
1406 struct mdt_body *mbo;
1407 struct mdt_lock_handle *lhc = &mti->mti_lh[MDT_LH_RMT];
1408 struct mdt_object *mo;
1413 req_capsule_set_size(mti->mti_pill, &RMF_MDT_MD, RCL_SERVER, 0);
1414 req_capsule_set_size(mti->mti_pill, &RMF_ACL, RCL_SERVER, 0);
1415 rc = req_capsule_server_pack(mti->mti_pill);
1417 RETURN(err_serious(rc));
1419 rep = req_capsule_server_get(mti->mti_pill, &RMF_DLM_REP);
1423 mbo = req_capsule_server_get(mti->mti_pill, &RMF_MDT_BODY);
1427 fid_extract_from_res_name(fid, &res->lr_name);
1428 mo = mdt_object_find(mti->mti_env, mti->mti_mdt, fid);
1429 if (unlikely(IS_ERR(mo)))
1430 RETURN(PTR_ERR(mo));
1432 if (!mdt_object_exists(mo))
1433 GOTO(out, rc = -ENOENT);
1435 if (mdt_object_remote(mo))
1436 GOTO(out, rc = -EPROTO);
1438 /* Get lock from request for possible resent case. */
1439 mdt_intent_fixup_resent(mti, *lockp, lhc, flags);
1441 if (!lustre_handle_is_used(&lhc->mlh_reg_lh)) {
1442 mdt_lock_handle_init(lhc);
1443 mdt_lh_reg_init(lhc, *lockp);
1445 /* This will block MDT thread but it should be fine until
1446 * client caches small amount of data for DoM, which should be
1447 * smaller than one BRW RPC and should be able to be
1448 * piggybacked by lock cancel RPC.
1449 * If the client could hold the lock too long, this code can be
1450 * revised to call mdt_object_lock_try(). And if fails, it will
1451 * return ELDLM_OK here and fall back into normal lock enqueue
1454 rc = mdt_object_lock(mti, mo, lhc, MDS_INODELOCK_DOM);
1459 if (!mdt_dom_lvb_is_valid(res)) {
1460 rc = mdt_dom_lvb_alloc(res);
1463 mdt_dom_disk_lvbo_update(mti->mti_env, mo, res, false);
1465 mdt_lvb2reply(res, mbo, NULL);
1467 rep->lock_policy_res2 = clear_serious(rc);
1468 if (rep->lock_policy_res2) {
1469 lhc->mlh_reg_lh.cookie = 0ull;
1470 GOTO(out, rc = ELDLM_LOCK_ABORTED);
1473 rc = mdt_intent_lock_replace(mti, lockp, lhc, flags, rc);
1476 lhc->mlh_reg_lh.cookie = 0ull;
1477 mdt_object_put(mti->mti_env, mo);
1481 /* check if client has already DoM lock for given resource */
1482 bool mdt_dom_client_has_lock(struct mdt_thread_info *info,
1483 const struct lu_fid *fid)
1485 struct mdt_device *mdt = info->mti_mdt;
1486 union ldlm_policy_data *policy = &info->mti_policy;
1487 struct ldlm_res_id *res_id = &info->mti_res_id;
1488 __u64 open_flags = info->mti_spec.sp_cr_flags;
1489 struct lustre_handle lockh;
1490 enum ldlm_mode mode;
1491 struct ldlm_lock *lock;
1495 policy->l_inodebits.bits = MDS_INODELOCK_DOM;
1496 fid_build_reg_res_name(fid, res_id);
1499 lm = (open_flags & MDS_FMODE_WRITE) ? LCK_PW : LCK_PR | LCK_PW;
1500 mode = ldlm_lock_match(mdt->mdt_namespace, LDLM_FL_BLOCK_GRANTED |
1501 LDLM_FL_TEST_LOCK, res_id, LDLM_IBITS, policy,
1504 /* There is no other PW lock on this object; finished. */
1508 lock = ldlm_handle2lock(&lockh);
1512 /* check if lock from the same client */
1513 rc = (lock->l_export->exp_handle.h_cookie ==
1514 info->mti_exp->exp_handle.h_cookie);
1515 LDLM_LOCK_PUT(lock);
1520 * MDT request handler for OST_GETATTR RPC.
1522 * This is data-specific request to get object and layout versions under
1523 * IO lock. It is reliable only for Data-on-MDT files.
1525 * \param[in] tsi target session environment for this request
1527 * \retval 0 if successful
1528 * \retval negative value on error
1530 int mdt_data_version_get(struct tgt_session_info *tsi)
1532 struct mdt_thread_info *mti = mdt_th_info(tsi->tsi_env);
1533 struct mdt_device *mdt = mti->mti_mdt;
1534 struct mdt_body *repbody;
1535 struct mdt_object *mo = mti->mti_object;
1536 struct lov_comp_md_v1 *comp;
1537 struct lustre_handle lh = { 0 };
1540 enum ldlm_mode lock_mode = LCK_PR;
1546 req_capsule_set_size(tsi->tsi_pill, &RMF_MDT_MD, RCL_SERVER, 0);
1547 req_capsule_set_size(tsi->tsi_pill, &RMF_ACL, RCL_SERVER, 0);
1548 rc = req_capsule_server_pack(tsi->tsi_pill);
1549 if (unlikely(rc != 0))
1550 RETURN(err_serious(rc));
1552 repbody = req_capsule_server_get(tsi->tsi_pill, &RMF_MDT_BODY);
1553 if (repbody == NULL)
1556 srvlock = tsi->tsi_mdt_body->mbo_valid & OBD_MD_FLFLAGS &&
1557 tsi->tsi_mdt_body->mbo_flags & OBD_FL_SRVLOCK;
1560 if (unlikely(tsi->tsi_mdt_body->mbo_flags & OBD_FL_FLUSH))
1563 fid_build_reg_res_name(&tsi->tsi_fid, &tsi->tsi_resid);
1564 rc = tgt_mdt_data_lock(mdt->mdt_namespace, &tsi->tsi_resid,
1565 &lh, lock_mode, &flags);
1570 if (!mdt_object_exists(mo))
1571 GOTO(out, rc = -ENOENT);
1572 if (mdt_object_remote(mo))
1573 GOTO(out, rc = -EREMOTE);
1574 if (!S_ISREG(lu_object_attr(&mo->mot_obj)))
1575 GOTO(out, rc = -EBADF);
1577 /* Get version first */
1578 version = dt_version_get(tsi->tsi_env, mdt_obj2dt(mo));
1579 if (version && version != -EOPNOTSUPP) {
1580 repbody->mbo_valid |= OBD_MD_FLDATAVERSION;
1581 /* re-use mbo_ioepoch to transfer version */
1582 repbody->mbo_version = version;
1585 /* Read layout to get its version */
1586 rc = mdt_big_xattr_get(mti, mo, XATTR_NAME_LOV);
1587 if (rc == -ENODATA) /* File has no layout yet */
1592 comp = mti->mti_buf.lb_buf;
1593 if (le32_to_cpu(comp->lcm_magic) != LOV_MAGIC_COMP_V1) {
1594 CDEBUG(D_INFO, DFID" has no composite layout",
1595 PFID(&tsi->tsi_fid));
1596 GOTO(out, rc = -ESTALE);
1599 CDEBUG(D_INODE, DFID": layout version: %u\n",
1600 PFID(&tsi->tsi_fid), le32_to_cpu(comp->lcm_layout_gen));
1602 repbody->mbo_valid |= OBD_MD_LAYOUT_VERSION;
1603 /* re-use mbo_rdev for that */
1604 repbody->mbo_layout_gen = le32_to_cpu(comp->lcm_layout_gen);
1608 tgt_data_unlock(&lh, lock_mode);
1610 repbody->mbo_valid |= OBD_MD_FLFLAGS;
1611 repbody->mbo_flags = OBD_FL_FLUSH;
1615 /* read file data to the buffer */
1616 int mdt_dom_read_on_open(struct mdt_thread_info *mti, struct mdt_device *mdt,
1617 struct lustre_handle *lh)
1619 const struct lu_env *env = mti->mti_env;
1620 struct tgt_session_info *tsi = tgt_ses_info(env);
1621 struct req_capsule *pill = tsi->tsi_pill;
1622 const struct lu_fid *fid;
1623 struct ptlrpc_request *req = tgt_ses_req(tsi);
1624 struct mdt_body *mbo;
1625 struct dt_device *dt = mdt->mdt_bottom;
1626 struct dt_object *mo;
1628 struct niobuf_remote *rnb = NULL;
1629 struct niobuf_local *lnb;
1632 unsigned int len, copied = 0;
1633 __u64 real_dom_size;
1634 int lnbs, nr_local, i;
1635 bool dom_lock = false;
1639 if (!req_capsule_field_present(pill, &RMF_NIOBUF_INLINE, RCL_SERVER)) {
1640 /* There is no reply buffers for this field, this means that
1641 * client has no support for data in reply.
1646 mbo = req_capsule_server_get(pill, &RMF_MDT_BODY);
1647 if (!(mbo->mbo_valid & OBD_MD_DOM_SIZE))
1650 if (!mbo->mbo_dom_size)
1653 if (lustre_handle_is_used(lh)) {
1654 struct ldlm_lock *lock;
1656 lock = ldlm_handle2lock(lh);
1658 dom_lock = ldlm_has_dom(lock) && ldlm_has_layout(lock);
1659 LDLM_LOCK_PUT(lock);
1663 /* return data along with open only along with DoM lock */
1664 if (!dom_lock || !mdt->mdt_opts.mo_dom_read_open)
1667 /* if DoM object holds encrypted content, we need to make sure we
1668 * send whole encryption units, or client will read corrupted content
1670 if (mbo->mbo_valid & LA_FLAGS && mbo->mbo_flags & LUSTRE_ENCRYPT_FL &&
1671 mbo->mbo_dom_size & ~LUSTRE_ENCRYPTION_MASK)
1672 real_dom_size = (mbo->mbo_dom_size & LUSTRE_ENCRYPTION_MASK) +
1673 LUSTRE_ENCRYPTION_UNIT_SIZE;
1675 real_dom_size = mbo->mbo_dom_size;
1677 CDEBUG(D_INFO, "File size %llu, reply sizes %d/%d\n",
1678 real_dom_size, req->rq_reqmsg->lm_repsize, req->rq_replen);
1679 len = req->rq_reqmsg->lm_repsize - req->rq_replen;
1681 /* NB: at this moment we have the following sizes:
1682 * - req->rq_replen: used data in reply
1683 * - req->rq_reqmsg->lm_repsize: total allocated reply buffer at client
1685 * Ideal case when file size fits in allocated reply buffer,
1686 * that mean we can return whole data in reply. We can also fit more
1687 * data up to max_reply_size in total reply size, but this will cause
1688 * re-allocation on client and resend with larger buffer. This is still
1689 * faster than separate READ IO.
1690 * Third case if file is too big to fit even in maximum size, in that
1691 * case we return just tail to optimize possible append.
1693 * At the moment the following strategy is used:
1694 * 1) try to fit into the buffer we have
1695 * 2) return just file tail otherwise.
1697 if (real_dom_size <= len) {
1698 /* can fit whole data */
1699 len = real_dom_size;
1701 } else if (real_dom_size <
1702 mdt_lmm_dom_stripesize(mti->mti_attr.ma_lmm)) {
1705 /* File tail offset must be aligned with larger page size
1706 * between client and server, so the maximum page size is
1707 * used here to align offset.
1709 * NB: DOM feature was introduced when server supports pagebits
1710 * already, so it should be always non-zero value. Report error
1711 * if it is not for some reason.
1713 if (!req->rq_export->exp_target_data.ted_pagebits) {
1714 CERROR("%s: client page bits are not saved on server\n",
1718 pgbits = max_t(int, PAGE_SHIFT,
1719 req->rq_export->exp_target_data.ted_pagebits);
1720 tail = real_dom_size % (1 << pgbits);
1722 /* no partial tail or tail can't fit in reply */
1723 if (tail == 0 || len < tail)
1727 offset = real_dom_size - len;
1729 /* DOM stripe is fully written, so don't expect its tail
1730 * will be used by append.
1735 LASSERT((offset & ~PAGE_MASK) == 0);
1736 rc = req_capsule_server_grow(pill, &RMF_NIOBUF_INLINE,
1737 sizeof(*rnb) + len);
1739 /* failed to grow data buffer, just exit */
1740 GOTO(out, rc = -E2BIG);
1743 /* re-take MDT_BODY and NIOBUF_INLINE buffers after the buffer grow */
1744 mbo = req_capsule_server_get(pill, &RMF_MDT_BODY);
1745 fid = &mbo->mbo_fid1;
1746 if (!fid_is_sane(fid))
1747 GOTO(out, rc = -EINVAL);
1749 rnb = req_capsule_server_get(tsi->tsi_pill, &RMF_NIOBUF_INLINE);
1751 GOTO(out, rc = -EPROTO);
1753 buf = (char *)rnb + sizeof(*rnb);
1755 rnb->rnb_offset = offset;
1757 mo = dt_locate(env, dt, fid);
1759 GOTO(out_rnb, rc = PTR_ERR(mo));
1760 LASSERT(mo != NULL);
1762 dt_read_lock(env, mo, 0);
1763 if (!dt_object_exists(mo))
1764 GOTO(unlock, rc = -ENOENT);
1766 /* parse remote buffers to local buffers and prepare the latter */
1767 lnbs = (len >> PAGE_SHIFT) + 1;
1768 OBD_ALLOC_PTR_ARRAY(lnb, lnbs);
1770 GOTO(unlock, rc = -ENOMEM);
1772 rc = dt_bufs_get(env, mo, rnb, lnb, lnbs, 0);
1773 if (unlikely(rc < 0))
1775 LASSERT(rc <= lnbs);
1777 rc = dt_read_prep(env, mo, lnb, nr_local);
1780 /* copy data to the buffer finally */
1781 for (i = 0; i < nr_local; i++) {
1782 char *p = kmap(lnb[i].lnb_page);
1785 LASSERT(lnb[i].lnb_page_offset == 0);
1786 off = lnb[i].lnb_len & ~PAGE_MASK;
1788 memset(p + off, 0, PAGE_SIZE - off);
1790 memcpy(buf + (i << PAGE_SHIFT), p, lnb[i].lnb_len);
1791 kunmap(lnb[i].lnb_page);
1792 copied += lnb[i].lnb_len;
1795 CDEBUG(D_INFO, "Read %i (wanted %u) bytes from %llu\n", copied,
1798 CWARN("%s: read %i bytes for "DFID
1799 " but wanted %u, is size wrong?\n",
1800 tsi->tsi_exp->exp_obd->obd_name, copied,
1801 PFID(&tsi->tsi_fid), len);
1802 /* Ignore partially copied data */
1807 dt_bufs_put(env, mo, lnb, nr_local);
1809 OBD_FREE_PTR_ARRAY(lnb, lnbs);
1811 dt_read_unlock(env, mo);
1812 lu_object_put(env, &mo->do_lu);
1814 rnb->rnb_len = copied;
1816 /* Don't fail OPEN request if read-on-open is failed, but drop
1817 * a message in log about the error.
1820 CDEBUG(D_INFO, "Read-on-open is failed, rc = %d", rc);
1826 * Completion AST for DOM discard locks:
1828 * CP AST an DOM discard lock is called always right after enqueue or from
1829 * reprocess if lock was blocked, in the latest case l_ast_data is set to
1830 * the mdt_object which is kept while there are pending locks on it.
1832 int ldlm_dom_discard_cp_ast(struct ldlm_lock *lock, __u64 flags, void *data)
1834 struct mdt_object *mo;
1835 struct lustre_handle dom_lh;
1840 /* l_ast_data is set when lock was not granted immediately
1841 * in mdt_dom_discard_data() below but put into waiting list,
1842 * so this CP callback means we are finished and corresponding
1843 * MDT object should be released finally as well as lock itself.
1845 lock_res_and_lock(lock);
1846 if (!lock->l_ast_data) {
1847 unlock_res_and_lock(lock);
1851 mo = lock->l_ast_data;
1852 lock->l_ast_data = NULL;
1853 unlock_res_and_lock(lock);
1855 ldlm_lock2handle(lock, &dom_lh);
1856 ldlm_lock_decref(&dom_lh, LCK_PW);
1858 env = lu_env_find();
1860 mdt_object_put(env, mo);
1865 void mdt_dom_discard_data(struct mdt_thread_info *info,
1866 struct mdt_object *mo)
1868 struct ptlrpc_request *req = mdt_info_req(info);
1869 struct mdt_device *mdt = mdt_dev(mo->mot_obj.lo_dev);
1870 union ldlm_policy_data policy;
1871 struct ldlm_res_id res_id;
1872 struct lustre_handle dom_lh;
1873 struct ldlm_lock *lock;
1874 __u64 flags = LDLM_FL_AST_DISCARD_DATA;
1880 if (req && req_is_replay(req))
1883 policy.l_inodebits.bits = MDS_INODELOCK_DOM;
1884 policy.l_inodebits.try_bits = 0;
1885 fid_build_reg_res_name(mdt_object_fid(mo), &res_id);
1887 /* Keep blocking version of discard for an old client to avoid
1888 * crashes on non-patched clients. LU-11359.
1890 old_client = req && !(exp_connect_flags2(req->rq_export) &
1891 OBD_CONNECT2_ASYNC_DISCARD);
1893 /* Tell the clients that the object is gone now and that they should
1894 * throw away any cached pages. */
1895 rc = ldlm_cli_enqueue_local(info->mti_env, mdt->mdt_namespace, &res_id,
1896 LDLM_IBITS, &policy, LCK_PW, &flags,
1897 ldlm_blocking_ast, old_client ?
1898 ldlm_completion_ast :
1899 ldlm_dom_discard_cp_ast,
1900 NULL, NULL, 0, LVB_T_NONE, NULL, &dom_lh);
1901 if (rc != ELDLM_OK) {
1903 "Failed to issue discard lock, rc = %d\n", rc);
1907 lock = ldlm_handle2lock(&dom_lh);
1908 lock_res_and_lock(lock);
1909 /* if lock is not granted then there are BL ASTs in progress and
1910 * lock will be granted in result of reprocessing with CP callback
1911 * notifying about that. The mdt object has to be kept until that and
1912 * it is saved in l_ast_data of the lock. Lock reference is kept too
1913 * until that to prevent it from canceling.
1915 if (!is_granted_or_cancelled_nolock(lock)) {
1916 mdt_object_get(info->mti_env, mo);
1917 lock->l_ast_data = mo;
1918 unlock_res_and_lock(lock);
1920 unlock_res_and_lock(lock);
1921 ldlm_lock_decref_and_cancel(&dom_lh, LCK_PW);
1923 LDLM_LOCK_PUT(lock);