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>
35 #include <lustre_nodemap.h>
37 #include "mdt_internal.h"
39 /* functions below are stubs for now, they will be implemented with
40 * grant support on MDT */
41 static inline void mdt_dom_read_lock(struct mdt_object *mo)
43 down_read(&mo->mot_dom_sem);
46 static inline void mdt_dom_read_unlock(struct mdt_object *mo)
48 up_read(&mo->mot_dom_sem);
51 static inline void mdt_dom_write_lock(struct mdt_object *mo)
53 down_write(&mo->mot_dom_sem);
56 static inline void mdt_dom_write_unlock(struct mdt_object *mo)
58 up_write(&mo->mot_dom_sem);
61 static void mdt_dom_resource_prolong(struct ldlm_prolong_args *arg)
63 struct ldlm_resource *res;
64 struct ldlm_lock *lock;
68 res = ldlm_resource_get(arg->lpa_export->exp_obd->obd_namespace, NULL,
69 &arg->lpa_resid, LDLM_IBITS, 0);
72 "Failed to get resource for resid %llu/%llu\n",
73 arg->lpa_resid.name[0], arg->lpa_resid.name[1]);
78 list_for_each_entry(lock, &res->lr_granted, l_res_link) {
79 if (ldlm_has_dom(lock)) {
80 LDLM_DEBUG(lock, "DOM lock to prolong ");
81 ldlm_lock_prolong_one(lock, arg);
82 /* only one PW or EX lock can be granted,
83 * no need to continue search
85 if (lock->l_granted_mode & (LCK_PW | LCK_EX))
90 ldlm_resource_putref(res);
95 static void mdt_prolong_dom_lock(struct tgt_session_info *tsi,
96 struct ldlm_prolong_args *data)
98 struct obdo *oa = &tsi->tsi_ost_body->oa;
99 struct ldlm_lock *lock;
103 data->lpa_timeout = prolong_timeout(tgt_ses_req(tsi));
104 data->lpa_export = tsi->tsi_exp;
105 data->lpa_resid = tsi->tsi_resid;
107 CDEBUG(D_RPCTRACE, "Prolong DOM lock for req %p with x%llu\n",
108 tgt_ses_req(tsi), tgt_ses_req(tsi)->rq_xid);
110 if (oa->o_valid & OBD_MD_FLHANDLE) {
111 /* mostly a request should be covered by only one lock, try
113 lock = ldlm_handle2lock(&oa->o_handle);
115 LASSERT(lock->l_export == data->lpa_export);
116 ldlm_lock_prolong_one(lock, data);
117 lock->l_last_used = ktime_get();
119 if (data->lpa_locks_cnt > 0)
123 mdt_dom_resource_prolong(data);
127 static int mdt_rw_hpreq_lock_match(struct ptlrpc_request *req,
128 struct ldlm_lock *lock)
130 struct obd_ioobj *ioo;
132 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
136 if (!(lock->l_policy_data.l_inodebits.bits & MDS_INODELOCK_DOM))
139 ioo = req_capsule_client_get(&req->rq_pill, &RMF_OBD_IOOBJ);
140 LASSERT(ioo != NULL);
142 LASSERT(lock->l_resource != NULL);
143 if (!fid_res_name_eq(&ioo->ioo_oid.oi_fid, &lock->l_resource->lr_name))
146 /* a bulk write can only hold a reference on a PW extent lock. */
147 mode = LCK_PW | LCK_GROUP;
149 /* whereas a bulk read can be protected by either a PR or PW
153 if (!(lock->l_granted_mode & mode))
159 static int mdt_rw_hpreq_check(struct ptlrpc_request *req)
161 struct tgt_session_info *tsi;
162 struct obd_ioobj *ioo;
163 struct niobuf_remote *rnb;
165 struct ldlm_prolong_args pa = { 0 };
169 /* Don't use tgt_ses_info() to get session info, because lock_match()
170 * can be called while request has no processing thread yet. */
171 tsi = lu_context_key_get(&req->rq_session, &tgt_session_key);
174 * Use LASSERT below because malformed RPCs should have
175 * been filtered out in tgt_hpreq_handler().
177 opc = lustre_msg_get_opc(req->rq_reqmsg);
178 LASSERT(opc == OST_READ || opc == OST_WRITE);
180 ioo = req_capsule_client_get(&req->rq_pill, &RMF_OBD_IOOBJ);
181 LASSERT(ioo != NULL);
183 rnb = req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE);
184 LASSERT(rnb != NULL);
185 LASSERT(!(rnb->rnb_flags & OBD_BRW_SRVLOCK));
187 pa.lpa_mode = LCK_PW | LCK_GROUP;
189 pa.lpa_mode |= LCK_PR;
191 DEBUG_REQ(D_RPCTRACE, req, "%s %s: refresh rw locks for " DFID,
192 tgt_name(tsi->tsi_tgt), current->comm, PFID(&tsi->tsi_fid));
194 mdt_prolong_dom_lock(tsi, &pa);
196 if (pa.lpa_blocks_cnt > 0) {
198 "%s: refreshed %u locks timeout for req %p\n",
199 tgt_name(tsi->tsi_tgt), pa.lpa_blocks_cnt, req);
203 RETURN(pa.lpa_locks_cnt > 0 ? 0 : -ESTALE);
206 static void mdt_rw_hpreq_fini(struct ptlrpc_request *req)
208 mdt_rw_hpreq_check(req);
211 static struct ptlrpc_hpreq_ops mdt_hpreq_rw = {
212 .hpreq_lock_match = mdt_rw_hpreq_lock_match,
213 .hpreq_check = mdt_rw_hpreq_check,
214 .hpreq_fini = mdt_rw_hpreq_fini
218 * Assign high priority operations to an IO request.
220 * Check if the incoming request is a candidate for
221 * high-priority processing. If it is, assign it a high
222 * priority operations table.
224 * \param[in] tsi target session environment for this request
226 void mdt_hp_brw(struct tgt_session_info *tsi)
228 struct niobuf_remote *rnb;
229 struct obd_ioobj *ioo;
233 ioo = req_capsule_client_get(tsi->tsi_pill, &RMF_OBD_IOOBJ);
234 LASSERT(ioo != NULL); /* must exist after request preprocessing */
235 if (ioo->ioo_bufcnt > 0) {
236 rnb = req_capsule_client_get(tsi->tsi_pill, &RMF_NIOBUF_REMOTE);
237 LASSERT(rnb != NULL); /* must exist after preprocessing */
239 /* no high priority if server lock is needed */
240 if (rnb->rnb_flags & OBD_BRW_SRVLOCK ||
241 (lustre_msg_get_flags(tgt_ses_req(tsi)->rq_reqmsg) &
245 tgt_ses_req(tsi)->rq_ops = &mdt_hpreq_rw;
248 static int mdt_punch_hpreq_lock_match(struct ptlrpc_request *req,
249 struct ldlm_lock *lock)
251 struct tgt_session_info *tsi;
256 /* Don't use tgt_ses_info() to get session info, because lock_match()
257 * can be called while request has no processing thread yet. */
258 tsi = lu_context_key_get(&req->rq_session, &tgt_session_key);
261 * Use LASSERT below because malformed RPCs should have
262 * been filtered out in tgt_hpreq_handler().
264 LASSERT(tsi->tsi_ost_body != NULL);
265 if (tsi->tsi_ost_body->oa.o_valid & OBD_MD_FLHANDLE &&
266 tsi->tsi_ost_body->oa.o_handle.cookie == lock->l_handle.h_cookie)
269 oa = &tsi->tsi_ost_body->oa;
271 LASSERT(lock->l_resource != NULL);
272 if (!fid_res_name_eq(&oa->o_oi.oi_fid, &lock->l_resource->lr_name))
275 if (!(lock->l_granted_mode & (LCK_PW | LCK_GROUP)))
282 * Implementation of ptlrpc_hpreq_ops::hpreq_lock_check for OST_PUNCH request.
284 * High-priority queue request check for whether the given punch request
285 * (\a req) is blocking an LDLM lock cancel. Also checks whether the request is
286 * covered by an LDLM lock.
290 * \param[in] req the incoming request
292 * \retval 1 if \a req is blocking an LDLM lock cancel
293 * \retval 0 if it is not
294 * \retval -ESTALE if lock is not found
296 static int mdt_punch_hpreq_check(struct ptlrpc_request *req)
298 struct tgt_session_info *tsi;
300 struct ldlm_prolong_args pa = { 0 };
304 /* Don't use tgt_ses_info() to get session info, because lock_match()
305 * can be called while request has no processing thread yet. */
306 tsi = lu_context_key_get(&req->rq_session, &tgt_session_key);
307 LASSERT(tsi != NULL);
308 oa = &tsi->tsi_ost_body->oa;
310 LASSERT(!(oa->o_valid & OBD_MD_FLFLAGS &&
311 oa->o_flags & OBD_FL_SRVLOCK));
313 pa.lpa_mode = LCK_PW | LCK_GROUP;
315 CDEBUG(D_DLMTRACE, "%s: refresh DOM lock for "DFID"\n",
316 tgt_name(tsi->tsi_tgt), PFID(&tsi->tsi_fid));
318 mdt_prolong_dom_lock(tsi, &pa);
320 if (pa.lpa_blocks_cnt > 0) {
322 "%s: refreshed %u locks timeout for req %p.\n",
323 tgt_name(tsi->tsi_tgt), pa.lpa_blocks_cnt, req);
327 RETURN(pa.lpa_locks_cnt > 0 ? 0 : -ESTALE);
331 * Implementation of ptlrpc_hpreq_ops::hpreq_lock_fini for OST_PUNCH request.
333 * Called after the request has been handled. It refreshes lock timeout again
334 * so that client has more time to send lock cancel RPC.
336 * \param[in] req request which is being processed.
338 static void mdt_punch_hpreq_fini(struct ptlrpc_request *req)
340 mdt_punch_hpreq_check(req);
343 static struct ptlrpc_hpreq_ops mdt_hpreq_punch = {
344 .hpreq_lock_match = mdt_punch_hpreq_lock_match,
345 .hpreq_check = mdt_punch_hpreq_check,
346 .hpreq_fini = mdt_punch_hpreq_fini
349 void mdt_hp_punch(struct tgt_session_info *tsi)
351 LASSERT(tsi->tsi_ost_body != NULL); /* must exists if we are here */
352 /* no high-priority if server lock is needed */
353 if ((tsi->tsi_ost_body->oa.o_valid & OBD_MD_FLFLAGS &&
354 tsi->tsi_ost_body->oa.o_flags & OBD_FL_SRVLOCK) ||
355 tgt_conn_flags(tsi) & OBD_CONNECT_MDS ||
356 lustre_msg_get_flags(tgt_ses_req(tsi)->rq_reqmsg) & MSG_REPLAY)
358 tgt_ses_req(tsi)->rq_ops = &mdt_hpreq_punch;
361 static int mdt_preprw_read(const struct lu_env *env, struct obd_export *exp,
362 struct mdt_device *mdt, struct mdt_object *mo,
363 struct lu_attr *la, int niocount,
364 struct niobuf_remote *rnb, int *nr_local,
365 struct niobuf_local *lnb, char *jobid)
367 struct tgt_session_info *tsi = tgt_ses_info(env);
368 struct ptlrpc_request *req = tgt_ses_req(tsi);
369 struct dt_object *dob;
370 int i, j, rc, tot_bytes = 0;
371 int maxlnb = *nr_local;
376 mdt_dom_read_lock(mo);
378 /* the only valid case when READ can find object is missing or stale
379 * when export is just evicted and open files are closed forcefully
380 * on server while client's READ can be in progress.
381 * This should not happen on healthy export, object can't be missing
382 * or dying because both states means it was finally destroyed.
384 level = exp->exp_failed ? D_INFO : D_ERROR;
385 if (!mdt_object_exists(mo)) {
387 "%s: READ IO to missing obj "DFID": rc = %d\n",
388 exp->exp_obd->obd_name, PFID(mdt_object_fid(mo)),
390 /* return 0 and continue with empty commit to skip such READ
391 * without more BRW errors.
395 if (lu_object_is_dying(&mo->mot_header)) {
397 "%s: READ IO to stale obj "DFID": rc = %d\n",
398 exp->exp_obd->obd_name, PFID(mdt_object_fid(mo)),
400 /* return 0 and continue with empty commit to skip such READ
401 * without more BRW errors.
406 dob = mdt_obj2dt(mo);
407 /* parse remote buffers to local buffers and prepare the latter */
408 for (i = 0, j = 0; i < niocount; i++) {
409 rc = dt_bufs_get(env, dob, rnb + i, lnb + j, maxlnb, 0);
410 if (unlikely(rc < 0))
412 /* correct index for local buffers to continue with */
416 tot_bytes += rnb[i].rnb_len;
419 rc = dt_attr_get(env, dob, la);
423 rc = dt_read_prep(env, dob, lnb, *nr_local);
427 mdt_counter_incr(req, LPROC_MDT_IO_READ, tot_bytes);
430 dt_bufs_put(env, dob, lnb, *nr_local);
431 mdt_dom_read_unlock(mo);
435 static int mdt_preprw_write(const struct lu_env *env, struct obd_export *exp,
436 struct mdt_device *mdt, struct mdt_object *mo,
437 struct lu_attr *la, struct obdo *oa,
438 int objcount, struct obd_ioobj *obj,
439 struct niobuf_remote *rnb, int *nr_local,
440 struct niobuf_local *lnb, char *jobid)
442 struct tgt_session_info *tsi = tgt_ses_info(env);
443 struct ptlrpc_request *req = tgt_ses_req(tsi);
444 struct dt_object *dob;
445 int i, j, k, rc = 0, tot_bytes = 0;
446 int maxlnb = *nr_local;
450 /* Process incoming grant info, set OBD_BRW_GRANTED flag and grant some
451 * space back if possible */
452 tgt_grant_prepare_write(env, exp, oa, rnb, obj->ioo_bufcnt);
454 mdt_dom_read_lock(mo);
456 /* don't report error in cases with failed export */
457 if (!mdt_object_exists(mo)) {
458 int level = exp->exp_failed ? D_INFO : D_ERROR;
462 "%s: WRITE IO to missing obj "DFID": rc = %d\n",
463 exp->exp_obd->obd_name, PFID(mdt_object_fid(mo)),
465 /* exit with no data written, note nr_local = 0 above */
468 if (lu_object_is_dying(&mo->mot_header)) {
469 /* This is possible race between object destroy followed by
470 * discard BL AST and client cache flushing. Object is
471 * referenced until discard finish.
473 CDEBUG(D_INODE, "WRITE IO to stale object "DFID"\n",
474 PFID(mdt_object_fid(mo)));
475 /* Note: continue with no error here to don't cause BRW errors
476 * but skip transaction in commitrw silently so no data is
481 dob = mdt_obj2dt(mo);
482 /* parse remote buffers to local buffers and prepare the latter */
483 for (i = 0, j = 0; i < obj->ioo_bufcnt; i++) {
484 rc = dt_bufs_get(env, dob, rnb + i, lnb + j, maxlnb, 1);
485 if (unlikely(rc < 0))
487 /* correct index for local buffers to continue with */
488 for (k = 0; k < rc; k++) {
489 lnb[j + k].lnb_flags = rnb[i].rnb_flags;
490 if (!(rnb[i].rnb_flags & OBD_BRW_GRANTED))
491 lnb[j + k].lnb_rc = -ENOSPC;
496 tot_bytes += rnb[i].rnb_len;
499 rc = dt_write_prep(env, dob, lnb, *nr_local);
503 mdt_counter_incr(req, LPROC_MDT_IO_WRITE, tot_bytes);
506 dt_bufs_put(env, dob, lnb, *nr_local);
508 mdt_dom_read_unlock(mo);
509 /* tgt_grant_prepare_write() was called, so we must commit */
510 tgt_grant_commit(exp, oa->o_grant_used, rc);
511 /* let's still process incoming grant information packed in the oa,
512 * but without enforcing grant since we won't proceed with the write.
513 * Just like a read request actually. */
514 tgt_grant_prepare_read(env, exp, oa);
518 int mdt_obd_preprw(const struct lu_env *env, int cmd, struct obd_export *exp,
519 struct obdo *oa, int objcount, struct obd_ioobj *obj,
520 struct niobuf_remote *rnb, int *nr_local,
521 struct niobuf_local *lnb)
523 struct tgt_session_info *tsi = tgt_ses_info(env);
524 struct mdt_thread_info *info = tsi2mdt_info(tsi);
525 struct lu_attr *la = &info->mti_attr.ma_attr;
526 struct mdt_device *mdt = mdt_dev(exp->exp_obd->obd_lu_dev);
527 struct mdt_object *mo;
531 /* The default value PTLRPC_MAX_BRW_PAGES is set in tgt_brw_write()
532 * but for MDT it is different, correct it here. */
533 if (*nr_local > MD_MAX_BRW_PAGES)
534 *nr_local = MD_MAX_BRW_PAGES;
536 jobid = tsi->tsi_jobid;
538 if (!oa || objcount != 1 || obj->ioo_bufcnt == 0) {
539 CERROR("%s: bad parameters %p/%i/%i\n",
540 exp->exp_obd->obd_name, oa, objcount, obj->ioo_bufcnt);
544 mo = mdt_object_find(env, mdt, &tsi->tsi_fid);
546 GOTO(out, rc = PTR_ERR(mo));
548 LASSERT(info->mti_object == NULL);
549 info->mti_object = mo;
551 if (cmd == OBD_BRW_WRITE) {
552 la_from_obdo(la, oa, OBD_MD_FLGETATTR);
553 rc = mdt_preprw_write(env, exp, mdt, mo, la, oa,
554 objcount, obj, rnb, nr_local, lnb,
556 } else if (cmd == OBD_BRW_READ) {
557 tgt_grant_prepare_read(env, exp, oa);
558 rc = mdt_preprw_read(env, exp, mdt, mo, la,
559 obj->ioo_bufcnt, rnb, nr_local, lnb,
561 obdo_from_la(oa, la, LA_ATIME);
563 CERROR("%s: wrong cmd %d received!\n",
564 exp->exp_obd->obd_name, cmd);
568 lu_object_put(env, &mo->mot_obj);
569 info->mti_object = NULL;
575 static int mdt_commitrw_read(const struct lu_env *env, struct mdt_device *mdt,
576 struct mdt_object *mo, int objcount, int niocount,
577 struct niobuf_local *lnb)
579 struct dt_object *dob;
584 dob = mdt_obj2dt(mo);
587 dt_bufs_put(env, dob, lnb, niocount);
589 mdt_dom_read_unlock(mo);
593 static int mdt_commitrw_write(const struct lu_env *env, struct obd_export *exp,
594 struct mdt_device *mdt, struct mdt_object *mo,
595 struct lu_attr *la, struct obdo *oa, int objcount,
596 int niocount, struct niobuf_local *lnb,
597 unsigned long granted, int old_rc)
599 struct dt_device *dt = mdt->mdt_bottom;
600 struct dt_object *dob;
608 dob = mdt_obj2dt(mo);
611 GOTO(out, rc = old_rc);
613 la->la_valid &= LA_ATIME | LA_MTIME | LA_CTIME;
615 if (!dt_object_exists(dob))
616 GOTO(out, rc = -ENOENT);
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 CFS_FAIL_TIMEOUT(OBD_FAIL_MDS_COMMITRW_DELAY, cfs_fail_val);
628 th = dt_trans_create(env, dt);
630 GOTO(out, rc = PTR_ERR(th));
632 for (i = 0; i < niocount; i++) {
633 if (!(lnb[i].lnb_flags & OBD_BRW_ASYNC)) {
639 if (OBD_FAIL_CHECK(OBD_FAIL_OST_DQACQ_NET))
640 GOTO(out_stop, rc = -EINPROGRESS);
642 rc = dt_declare_write_commit(env, dob, lnb, niocount, th);
647 /* update [mac]time if needed */
648 rc = dt_declare_attr_set(env, dob, la, th);
653 tgt_vbr_obj_set(env, dob);
654 rc = dt_trans_start(env, dt, th);
658 dt_write_lock(env, dob, 0);
659 if (lu_object_is_dying(&mo->mot_header)) {
660 /* Commit to stale object can be just skipped silently. */
661 CDEBUG(D_INODE, "skip commit to stale object "DFID"\n",
662 PFID(mdt_object_fid(mo)));
663 GOTO(unlock, rc = 0);
665 rc = dt_write_commit(env, dob, lnb, niocount, th, oa->o_size);
667 restart = th->th_restart_tran;
672 rc = dt_attr_set(env, dob, la, th);
676 /* get attr to return */
677 rc = dt_attr_get(env, dob, la);
679 dt_write_unlock(env, dob);
682 /* Force commit to make the just-deleted blocks
683 * reusable. LU-456 */
688 if (rc == 0 && granted > 0) {
689 if (tgt_grant_commit_cb_add(th, exp, granted) == 0)
693 th->th_result = restart ? 0 : rc;
694 dt_trans_stop(env, dt, th);
695 if (rc == -ENOSPC && retries++ < 3) {
696 CDEBUG(D_INODE, "retry after force commit, retries:%d\n",
703 if (retries % 10000 == 0)
704 CERROR("%s: restart IO write too many times: %d\n",
705 exp->exp_obd->obd_name, retries);
706 CDEBUG(D_INODE, "retry transaction, retries:%d\n",
712 dt_bufs_put(env, dob, lnb, niocount);
713 mdt_dom_read_unlock(mo);
715 tgt_grant_commit(exp, granted, old_rc);
719 void mdt_dom_obj_lvb_update(const struct lu_env *env, struct mdt_object *mo,
720 struct obdo *oa, bool increase_only)
722 struct mdt_device *mdt = mdt_dev(mo->mot_obj.lo_dev);
723 struct ldlm_res_id resid;
724 struct ldlm_resource *res;
726 fid_build_reg_res_name(mdt_object_fid(mo), &resid);
727 res = ldlm_resource_get(mdt->mdt_namespace, NULL, &resid,
732 /* Update lvbo data if exists. */
733 if (mdt_dom_lvb_is_valid(res)) {
734 mdt_dom_disk_lvbo_update(env, mo, res, increase_only);
736 struct ost_lvb *res_lvb = res->lr_lvb_data;
739 oa->o_valid |= OBD_MD_FLBLOCKS | OBD_MD_FLSIZE |
740 OBD_MD_FLMTIME | OBD_MD_FLATIME |
742 oa->o_blocks = res_lvb->lvb_blocks;
743 oa->o_size = res_lvb->lvb_size;
744 oa->o_atime = res_lvb->lvb_atime;
745 oa->o_mtime = res_lvb->lvb_mtime;
746 oa->o_ctime = res_lvb->lvb_ctime;
750 ldlm_resource_putref(res);
753 int mdt_obd_commitrw(const struct lu_env *env, int cmd, struct obd_export *exp,
754 struct obdo *oa, int objcount, struct obd_ioobj *obj,
755 struct niobuf_remote *rnb, int npages,
756 struct niobuf_local *lnb, int old_rc)
758 struct mdt_thread_info *info = mdt_th_info(env);
759 struct mdt_device *mdt = mdt_dev(exp->exp_obd->obd_lu_dev);
760 struct mdt_object *mo = info->mti_object;
761 struct lu_attr *la = &info->mti_attr.ma_attr;
768 if (cmd == OBD_BRW_WRITE) {
769 struct lu_nodemap *nodemap;
770 __u32 mapped_uid, mapped_gid, mapped_projid;
772 nodemap = nodemap_get_from_exp(exp);
774 RETURN(PTR_ERR(nodemap));
775 mapped_uid = nodemap_map_id(nodemap, NODEMAP_UID,
776 NODEMAP_FS_TO_CLIENT,
778 mapped_gid = nodemap_map_id(nodemap, NODEMAP_GID,
779 NODEMAP_FS_TO_CLIENT,
781 mapped_projid = nodemap_map_id(nodemap, NODEMAP_PROJID,
782 NODEMAP_FS_TO_CLIENT,
784 if (!IS_ERR_OR_NULL(nodemap)) {
785 /* do not bypass quota enforcement if squashed uid */
786 if (unlikely(mapped_uid == nodemap->nm_squash_uid)) {
789 for (idx = 0; idx < npages; idx++)
790 lnb[idx].lnb_flags &=
791 ~OBD_BRW_SYS_RESOURCE;
794 nodemap_putref(nodemap);
796 /* Don't update timestamps if this write is older than a
797 * setattr which modifies the timestamps. b=10150 */
799 /* XXX when we start having persistent reservations this needs
800 * to be changed to ofd_fmd_get() to create the fmd if it
801 * doesn't already exist so we can store the reservation handle
803 valid = OBD_MD_FLUID | OBD_MD_FLGID;
804 if (tgt_fmd_check(exp, mdt_object_fid(mo),
805 mdt_info_req(info)->rq_xid))
806 valid |= OBD_MD_FLATIME | OBD_MD_FLMTIME |
809 la_from_obdo(la, oa, valid);
811 rc = mdt_commitrw_write(env, exp, mdt, mo, la, oa, objcount,
812 npages, lnb, oa->o_grant_used, old_rc);
814 obdo_from_la(oa, la, VALID_FLAGS | LA_GID | LA_UID);
816 obdo_from_la(oa, la, LA_GID | LA_UID);
818 mdt_dom_obj_lvb_update(env, mo, NULL, false);
819 /* don't report overquota flag if we failed before reaching
821 if (old_rc == 0 && (rc == 0 || rc == -EDQUOT)) {
822 /* return the overquota flags to client */
823 if (lnb[0].lnb_flags & OBD_BRW_OVER_USRQUOTA) {
824 if (oa->o_valid & OBD_MD_FLFLAGS)
825 oa->o_flags |= OBD_FL_NO_USRQUOTA;
827 oa->o_flags = OBD_FL_NO_USRQUOTA;
830 if (lnb[0].lnb_flags & OBD_BRW_OVER_GRPQUOTA) {
831 if (oa->o_valid & OBD_MD_FLFLAGS)
832 oa->o_flags |= OBD_FL_NO_GRPQUOTA;
834 oa->o_flags = OBD_FL_NO_GRPQUOTA;
837 if (lnb[0].lnb_flags & OBD_BRW_OVER_PRJQUOTA) {
838 if (oa->o_valid & OBD_MD_FLFLAGS)
839 oa->o_flags |= OBD_FL_NO_PRJQUOTA;
841 oa->o_flags = OBD_FL_NO_PRJQUOTA;
845 oa->o_flags |= OBD_FL_ROOT_SQUASH;
847 oa->o_valid |= OBD_MD_FLFLAGS | OBD_MD_FLUSRQUOTA |
848 OBD_MD_FLGRPQUOTA | OBD_MD_FLPRJQUOTA;
850 /* Convert back to client IDs. LU-9671.
851 * nodemap_get_from_exp() may fail due to nodemap deactivated,
852 * server ID will be returned back to client in that case.
854 oa->o_uid = mapped_uid;
855 oa->o_gid = mapped_gid;
856 oa->o_projid = mapped_projid;
857 } else if (cmd == OBD_BRW_READ) {
858 /* If oa != NULL then mdt_preprw_read updated the inode
859 * atime and we should update the lvb so that other glimpses
860 * will also get the updated value. bug 5972 */
862 mdt_dom_obj_lvb_update(env, mo, NULL, true);
863 rc = mdt_commitrw_read(env, mdt, mo, objcount, npages, lnb);
869 mdt_thread_info_fini(info);
873 int mdt_object_fallocate(const struct lu_env *env, struct dt_device *dt,
874 struct dt_object *dob, __u64 start, __u64 end,
875 int mode, struct lu_attr *la)
882 if (!dt_object_exists(dob))
885 th = dt_trans_create(env, dt);
889 rc = dt_declare_attr_set(env, dob, la, th);
893 rc = dt_declare_fallocate(env, dob, start, end, mode, th);
897 tgt_vbr_obj_set(env, dob);
898 rc = dt_trans_start(env, dt, th);
902 dt_write_lock(env, dob, 0);
903 rc = dt_falloc(env, dob, start, end, mode, th);
906 rc = dt_attr_set(env, dob, la, th);
910 dt_write_unlock(env, dob);
913 dt_trans_stop(env, dt, th);
918 * MDT request handler for OST_FALLOCATE RPC.
920 * This is part of request processing. Validate request fields,
921 * preallocate the given MDT object and pack reply.
923 * \param[in] tsi target session environment for this request
925 * \retval 0 if successful
926 * \retval negative value on error
928 int mdt_fallocate_hdl(struct tgt_session_info *tsi)
930 struct obdo *oa = &tsi->tsi_ost_body->oa;
931 struct ptlrpc_request *req = tgt_ses_req(tsi);
932 struct ost_body *repbody;
933 struct mdt_thread_info *info;
934 struct ldlm_namespace *ns = tsi->tsi_tgt->lut_obd->obd_namespace;
935 struct obd_export *exp = tsi->tsi_exp;
936 struct mdt_device *mdt = mdt_dev(exp->exp_obd->obd_lu_dev);
937 struct mdt_object *mo;
938 struct dt_object *dob;
941 struct lustre_handle lh = { 0, };
945 ktime_t kstart = ktime_get();
947 repbody = req_capsule_server_get(tsi->tsi_pill, &RMF_OST_BODY);
949 RETURN(err_serious(-ENOMEM));
952 * fallocate start and end are passed in o_size, o_blocks
955 if ((oa->o_valid & (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS)) !=
956 (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS))
957 RETURN(err_serious(-EPROTO));
961 mode = oa->o_falloc_mode;
964 "fallocate: "DFID", mode = %#x, start = %lld, end = %lld\n",
965 PFID(&tsi->tsi_fid), mode, start, end);
968 * mode == 0 (which is standard prealloc) and PUNCH is supported
969 * Rest of mode options are not supported yet.
971 if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
974 if (mode & FALLOC_FL_PUNCH_HOLE && !(mode & FALLOC_FL_KEEP_SIZE)) {
975 CWARN("%s: PUNCH mode misses KEEP_SIZE flag, setting it\n",
976 tsi->tsi_tgt->lut_obd->obd_name);
977 mode |= FALLOC_FL_KEEP_SIZE;
980 info = tsi2mdt_info(tsi);
981 la = &info->mti_attr.ma_attr;
983 repbody->oa.o_oi = oa->o_oi;
984 repbody->oa.o_valid = OBD_MD_FLID;
986 srvlock = oa->o_valid & OBD_MD_FLFLAGS &&
987 oa->o_flags & OBD_FL_SRVLOCK;
990 rc = tgt_mdt_data_lock(ns, &tsi->tsi_resid, &lh, LCK_PW,
996 mo = mdt_object_find(tsi->tsi_env, mdt, &tsi->tsi_fid);
998 GOTO(out_unlock, rc = PTR_ERR(mo));
1000 if (!mdt_object_exists(mo))
1001 GOTO(out_put, rc = -ENOENT);
1003 /* Shouldn't happen on dirs */
1004 if (S_ISDIR(lu_object_attr(&mo->mot_obj))) {
1006 CERROR("%s: fallocate on dir "DFID": rc = %d\n",
1007 exp->exp_obd->obd_name, PFID(&tsi->tsi_fid), rc);
1011 la_from_obdo(la, oa, OBD_MD_FLMTIME | OBD_MD_FLATIME | OBD_MD_FLCTIME);
1013 mdt_dom_write_lock(mo);
1014 dob = mdt_obj2dt(mo);
1016 if (la->la_valid & (LA_ATIME | LA_MTIME | LA_CTIME))
1017 tgt_fmd_update(tsi->tsi_exp, &tsi->tsi_fid,
1018 tgt_ses_req(tsi)->rq_xid);
1020 rc = mdt_object_fallocate(tsi->tsi_env, mdt->mdt_bottom, dob, start,
1022 mdt_dom_write_unlock(mo);
1026 mdt_dom_obj_lvb_update(tsi->tsi_env, mo, &repbody->oa, false);
1028 mdt_counter_incr(req, LPROC_MDT_FALLOCATE,
1029 ktime_us_delta(ktime_get(), kstart));
1033 lu_object_put(tsi->tsi_env, &mo->mot_obj);
1036 tgt_data_unlock(&lh, LCK_PW);
1038 mdt_thread_info_fini(info);
1042 int mdt_object_punch(const struct lu_env *env, struct dt_device *dt,
1043 struct dt_object *dob, __u64 start, __u64 end,
1051 /* we support truncate, not punch yet */
1052 LASSERT(end == OBD_OBJECT_EOF);
1054 if (!dt_object_exists(dob))
1057 th = dt_trans_create(env, dt);
1059 RETURN(PTR_ERR(th));
1061 rc = dt_declare_attr_set(env, dob, la, th);
1065 rc = dt_declare_punch(env, dob, start, OBD_OBJECT_EOF, th);
1069 tgt_vbr_obj_set(env, dob);
1070 rc = dt_trans_start(env, dt, th);
1074 dt_write_lock(env, dob, 0);
1075 rc = dt_punch(env, dob, start, OBD_OBJECT_EOF, th);
1078 rc = dt_attr_set(env, dob, la, th);
1082 dt_write_unlock(env, dob);
1085 dt_trans_stop(env, dt, th);
1089 int mdt_punch_hdl(struct tgt_session_info *tsi)
1091 const struct obdo *oa = &tsi->tsi_ost_body->oa;
1092 struct ptlrpc_request *req = tgt_ses_req(tsi);
1093 struct ost_body *repbody;
1094 struct mdt_thread_info *info;
1096 struct ldlm_namespace *ns = tsi->tsi_tgt->lut_obd->obd_namespace;
1097 struct obd_export *exp = tsi->tsi_exp;
1098 struct mdt_device *mdt = mdt_dev(exp->exp_obd->obd_lu_dev);
1099 struct mdt_object *mo;
1100 struct dt_object *dob;
1102 struct lustre_handle lh = { 0, };
1103 ktime_t kstart = ktime_get();
1110 if ((oa->o_valid & (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS)) !=
1111 (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS))
1112 RETURN(err_serious(-EPROTO));
1114 repbody = req_capsule_server_get(tsi->tsi_pill, &RMF_OST_BODY);
1115 if (repbody == NULL)
1116 RETURN(err_serious(-ENOMEM));
1118 /* punch start,end are passed in o_size,o_blocks throught wire */
1122 if (end != OBD_OBJECT_EOF) /* Only truncate is supported */
1125 info = tsi2mdt_info(tsi);
1126 la = &info->mti_attr.ma_attr;
1127 /* standard truncate optimization: if file body is completely
1128 * destroyed, don't send data back to the server. */
1130 flags |= LDLM_FL_AST_DISCARD_DATA;
1132 repbody->oa.o_oi = oa->o_oi;
1133 repbody->oa.o_valid = OBD_MD_FLID;
1135 srvlock = (exp_connect_flags(exp) & OBD_CONNECT_SRVLOCK) &&
1136 oa->o_valid & OBD_MD_FLFLAGS &&
1137 oa->o_flags & OBD_FL_SRVLOCK;
1140 rc = tgt_mdt_data_lock(ns, &tsi->tsi_resid, &lh, LCK_PW,
1146 CDEBUG(D_INODE, "calling punch for object "DFID", valid = %#llx"
1147 ", start = %lld, end = %lld\n", PFID(&tsi->tsi_fid),
1148 oa->o_valid, start, end);
1150 mo = mdt_object_find(tsi->tsi_env, mdt, &tsi->tsi_fid);
1152 GOTO(out_unlock, rc = PTR_ERR(mo));
1154 if (!mdt_object_exists(mo))
1155 GOTO(out_put, rc = -ENOENT);
1157 /* Shouldn't happen on dirs */
1158 if (S_ISDIR(lu_object_attr(&mo->mot_obj))) {
1160 CERROR("%s: Truncate on dir "DFID": rc = %d\n",
1161 exp->exp_obd->obd_name, PFID(&tsi->tsi_fid), rc);
1165 mdt_dom_write_lock(mo);
1166 dob = mdt_obj2dt(mo);
1168 la_from_obdo(la, oa, OBD_MD_FLMTIME | OBD_MD_FLATIME | OBD_MD_FLCTIME);
1169 la->la_size = start;
1170 la->la_valid |= LA_SIZE;
1172 /* MDT supports FMD for Data-on-MDT needs */
1173 if (la->la_valid & (LA_ATIME | LA_MTIME | LA_CTIME))
1174 tgt_fmd_update(tsi->tsi_exp, &tsi->tsi_fid,
1175 tgt_ses_req(tsi)->rq_xid);
1177 rc = mdt_object_punch(tsi->tsi_env, mdt->mdt_bottom, dob,
1179 mdt_dom_write_unlock(mo);
1183 mdt_dom_obj_lvb_update(tsi->tsi_env, mo, &repbody->oa, false);
1185 mdt_counter_incr(req, LPROC_MDT_IO_PUNCH,
1186 ktime_us_delta(ktime_get(), kstart));
1189 lu_object_put(tsi->tsi_env, &mo->mot_obj);
1192 tgt_data_unlock(&lh, LCK_PW);
1194 mdt_thread_info_fini(info);
1199 * MDT glimpse for Data-on-MDT
1201 * If there is write lock on client then function issues glimpse_ast to get
1202 * an actual size from that client.
1205 int mdt_do_glimpse(const struct lu_env *env, struct ldlm_namespace *ns,
1206 struct ldlm_resource *res)
1208 union ldlm_policy_data policy;
1209 struct lustre_handle lockh;
1210 enum ldlm_mode mode;
1211 struct ldlm_lock *lock;
1212 struct ldlm_glimpse_work *gl_work;
1218 /* There can be only one write lock covering data, try to match it. */
1219 policy.l_inodebits.bits = MDS_INODELOCK_DOM;
1220 mode = ldlm_lock_match(ns, LDLM_FL_TEST_LOCK,
1221 &res->lr_name, LDLM_IBITS, &policy,
1224 /* There is no PW lock on this object; finished. */
1228 lock = ldlm_handle2lock(&lockh);
1233 * This check is for lock taken in mdt_reint_unlink() that does
1234 * not have l_glimpse_ast set. So the logic is: if there is a lock
1235 * with no l_glimpse_ast set, this object is being destroyed already.
1236 * Hence, if you are grabbing DLM locks on the server, always set
1237 * non-NULL glimpse_ast (e.g., ldlm_request.c::ldlm_glimpse_ast()).
1239 if (lock->l_glimpse_ast == NULL) {
1240 LDLM_DEBUG(lock, "no l_glimpse_ast");
1241 GOTO(out, rc = -ENOENT);
1244 OBD_SLAB_ALLOC_PTR_GFP(gl_work, ldlm_glimpse_work_kmem, GFP_ATOMIC);
1246 GOTO(out, rc = -ENOMEM);
1248 /* Populate the gl_work structure.
1249 * Grab additional reference on the lock which will be released in
1250 * ldlm_work_gl_ast_lock() */
1251 gl_work->gl_lock = LDLM_LOCK_GET(lock);
1252 /* The glimpse callback is sent to one single IO lock. As a result,
1253 * the gl_work list is just composed of one element */
1254 list_add_tail(&gl_work->gl_list, &gl_list);
1255 /* There is actually no need for a glimpse descriptor when glimpsing
1257 gl_work->gl_desc = NULL;
1258 /* the ldlm_glimpse_work structure is allocated on the stack */
1259 gl_work->gl_flags = LDLM_GL_WORK_SLAB_ALLOCATED;
1261 ldlm_glimpse_locks(res, &gl_list); /* this will update the LVB */
1263 /* If the list is not empty, we failed to glimpse a lock and
1264 * must clean it up. Usually due to a race with unlink.*/
1265 if (!list_empty(&gl_list)) {
1266 LDLM_LOCK_RELEASE(lock);
1267 OBD_SLAB_FREE_PTR(gl_work, ldlm_glimpse_work_kmem);
1272 LDLM_LOCK_PUT(lock);
1276 static void mdt_lvb2reply(struct ldlm_resource *res, struct mdt_body *mb,
1277 struct ost_lvb *lvb)
1279 struct ost_lvb *res_lvb;
1282 res_lvb = res->lr_lvb_data;
1288 mb->mbo_dom_size = res_lvb->lvb_size;
1289 mb->mbo_dom_blocks = res_lvb->lvb_blocks;
1290 mb->mbo_mtime = res_lvb->lvb_mtime;
1291 mb->mbo_ctime = res_lvb->lvb_ctime;
1292 mb->mbo_atime = res_lvb->lvb_atime;
1293 mb->mbo_valid |= OBD_MD_FLATIME | OBD_MD_FLCTIME |
1294 OBD_MD_FLMTIME | OBD_MD_DOM_SIZE;
1296 CDEBUG(D_DLMTRACE, "size %llu\n", res_lvb->lvb_size);
1302 * MDT glimpse for Data-on-MDT
1304 * This function is called when MDT get attributes for the DoM object.
1305 * If there is write lock on client then function issues glimpse_ast to get
1306 * an actual size from that client.
1308 int mdt_dom_object_size(const struct lu_env *env, struct mdt_device *mdt,
1309 const struct lu_fid *fid, struct mdt_body *mb,
1312 struct ldlm_res_id resid;
1313 struct ldlm_resource *res;
1318 fid_build_reg_res_name(fid, &resid);
1319 res = ldlm_resource_get(mdt->mdt_namespace, NULL, &resid,
1324 /* Update lvbo data if DoM lock returned or if LVB is not yet valid. */
1325 if (dom_lock || !mdt_dom_lvb_is_valid(res))
1326 mdt_dom_lvbo_update(res, NULL, NULL, false);
1328 mdt_lvb2reply(res, mb, NULL);
1329 ldlm_resource_putref(res);
1334 * MDT DoM lock intent policy (glimpse)
1336 * Intent policy is called when lock has an intent, for DoM file that
1337 * means glimpse lock and policy fills Lock Value Block (LVB).
1339 * If already granted lock is found it will be placed in \a lockp and
1340 * returned back to caller function.
1342 * \param[in] tsi session info
1343 * \param[in,out] lockp pointer to the lock
1344 * \param[in] flags LDLM flags
1346 * \retval ELDLM_LOCK_REPLACED if already granted lock was found
1347 * and placed in \a lockp
1348 * \retval ELDLM_LOCK_ABORTED in other cases except error
1349 * \retval negative value on error
1351 int mdt_glimpse_enqueue(struct mdt_thread_info *mti, struct ldlm_namespace *ns,
1352 struct ldlm_lock **lockp, __u64 flags)
1354 struct ldlm_lock *lock = *lockp;
1355 struct ldlm_resource *res = lock->l_resource;
1356 ldlm_processing_policy policy;
1357 struct ldlm_reply *rep;
1358 struct mdt_body *mbo;
1359 struct ost_lvb *lvb;
1360 bool old_client = !exp_connect_dom_lvb(mti->mti_exp);
1365 policy = ldlm_get_processing_policy(res);
1366 LASSERT(policy != NULL);
1368 if (unlikely(old_client)) {
1369 req_capsule_set_size(mti->mti_pill, &RMF_MDT_MD, RCL_SERVER, 0);
1370 req_capsule_set_size(mti->mti_pill, &RMF_ACL, RCL_SERVER, 0);
1372 req_capsule_set_size(mti->mti_pill, &RMF_DLM_LVB, RCL_SERVER,
1375 rc = req_capsule_server_pack(mti->mti_pill);
1377 RETURN(err_serious(rc));
1379 rep = req_capsule_server_get(mti->mti_pill, &RMF_DLM_REP);
1381 if (unlikely(old_client)) {
1382 mbo = req_capsule_server_get(mti->mti_pill, &RMF_MDT_BODY);
1386 lvb = req_capsule_server_get(mti->mti_pill, &RMF_DLM_LVB);
1392 /* Check if this is a resend case (MSG_RESENT is set on RPC) and a
1393 * lock was found by ldlm_handle_enqueue(); if so no need to grant
1395 if (flags & LDLM_FL_RESENT) {
1396 rc = LDLM_ITER_CONTINUE;
1398 __u64 tmpflags = LDLM_FL_BLOCK_NOWAIT;
1399 enum ldlm_error err;
1401 rc = policy(lock, &tmpflags, LDLM_PROCESS_RESCAN, &err, NULL);
1402 check_res_locked(res);
1406 /* The lock met with no resistance; we're finished. */
1407 if (rc == LDLM_ITER_CONTINUE) {
1408 GOTO(fill_mbo, rc = ELDLM_LOCK_REPLACED);
1409 } else if (flags & LDLM_FL_BLOCK_NOWAIT) {
1410 /* LDLM_FL_BLOCK_NOWAIT means it is for AGL. Do not send glimpse
1411 * callback for glimpse size. The real size user will trigger
1412 * the glimpse callback when necessary. */
1413 GOTO(fill_mbo, rc = ELDLM_LOCK_ABORTED);
1416 rc = mdt_do_glimpse(mti->mti_env, ns, res);
1417 if (rc == -ENOENT) {
1418 /* We are racing with unlink(); just return -ENOENT */
1419 rep->lock_policy_res2 = ptlrpc_status_hton(-ENOENT);
1420 } else if (rc == -EINVAL) {
1421 /* this is possible is client lock has been cancelled but
1422 * still exists on server. If that lock was found on server
1423 * as only conflicting lock then the client has already
1424 * size authority and glimpse is not needed. */
1425 CDEBUG(D_DLMTRACE, "Glimpse from the client owning lock\n");
1426 } else if (rc < 0) {
1429 rc = ELDLM_LOCK_ABORTED;
1431 /* LVB can be without valid data in case of DOM */
1432 if (!mdt_dom_lvb_is_valid(res))
1433 mdt_dom_lvbo_update(res, lock, NULL, false);
1434 mdt_lvb2reply(res, mbo, lvb);
1439 int mdt_brw_enqueue(struct mdt_thread_info *mti, struct ldlm_namespace *ns,
1440 struct ldlm_lock **lockp, __u64 flags)
1442 struct tgt_session_info *tsi = tgt_ses_info(mti->mti_env);
1443 struct lu_fid *fid = &tsi->tsi_fid;
1444 struct ldlm_lock *lock = *lockp;
1445 struct ldlm_resource *res = lock->l_resource;
1446 struct ldlm_reply *rep;
1447 struct mdt_body *mbo;
1448 struct mdt_lock_handle *lhc = &mti->mti_lh[MDT_LH_RMT];
1449 struct mdt_object *mo;
1454 req_capsule_set_size(mti->mti_pill, &RMF_MDT_MD, RCL_SERVER, 0);
1455 req_capsule_set_size(mti->mti_pill, &RMF_ACL, RCL_SERVER, 0);
1456 rc = req_capsule_server_pack(mti->mti_pill);
1458 RETURN(err_serious(rc));
1460 rep = req_capsule_server_get(mti->mti_pill, &RMF_DLM_REP);
1464 mbo = req_capsule_server_get(mti->mti_pill, &RMF_MDT_BODY);
1468 fid_extract_from_res_name(fid, &res->lr_name);
1469 mo = mdt_object_find(mti->mti_env, mti->mti_mdt, fid);
1470 if (unlikely(IS_ERR(mo)))
1471 RETURN(PTR_ERR(mo));
1473 if (!mdt_object_exists(mo))
1474 GOTO(out, rc = -ENOENT);
1476 if (mdt_object_remote(mo))
1477 GOTO(out, rc = -EPROTO);
1479 /* Get lock from request for possible resent case. */
1480 mdt_intent_fixup_resent(mti, *lockp, lhc, flags);
1482 if (!lustre_handle_is_used(&lhc->mlh_reg_lh)) {
1483 mdt_lock_handle_init(lhc);
1484 mdt_lh_reg_init(lhc, *lockp);
1486 /* This will block MDT thread but it should be fine until
1487 * client caches small amount of data for DoM, which should be
1488 * smaller than one BRW RPC and should be able to be
1489 * piggybacked by lock cancel RPC.
1490 * If the client could hold the lock too long, this code can be
1491 * revised to call mdt_object_lock_try(). And if fails, it will
1492 * return ELDLM_OK here and fall back into normal lock enqueue
1495 rc = mdt_object_lock(mti, mo, lhc, MDS_INODELOCK_DOM);
1500 if (!mdt_dom_lvb_is_valid(res)) {
1501 rc = mdt_dom_lvb_alloc(res);
1504 mdt_dom_disk_lvbo_update(mti->mti_env, mo, res, false);
1506 mdt_lvb2reply(res, mbo, NULL);
1508 rep->lock_policy_res2 = clear_serious(rc);
1509 if (rep->lock_policy_res2) {
1510 lhc->mlh_reg_lh.cookie = 0ull;
1511 GOTO(out, rc = ELDLM_LOCK_ABORTED);
1514 rc = mdt_intent_lock_replace(mti, lockp, lhc, flags, rc);
1517 lhc->mlh_reg_lh.cookie = 0ull;
1518 mdt_object_put(mti->mti_env, mo);
1522 /* check if client has already DoM lock for given resource */
1523 bool mdt_dom_client_has_lock(struct mdt_thread_info *info,
1524 const struct lu_fid *fid)
1526 struct mdt_device *mdt = info->mti_mdt;
1527 union ldlm_policy_data *policy = &info->mti_policy;
1528 struct ldlm_res_id *res_id = &info->mti_res_id;
1529 __u64 open_flags = info->mti_spec.sp_cr_flags;
1530 struct lustre_handle lockh;
1531 enum ldlm_mode mode;
1532 struct ldlm_lock *lock;
1536 policy->l_inodebits.bits = MDS_INODELOCK_DOM;
1537 fid_build_reg_res_name(fid, res_id);
1540 lm = (open_flags & MDS_FMODE_WRITE) ? LCK_PW : LCK_PR | LCK_PW;
1541 mode = ldlm_lock_match(mdt->mdt_namespace, LDLM_FL_BLOCK_GRANTED |
1542 LDLM_FL_TEST_LOCK, res_id, LDLM_IBITS, policy,
1545 /* There is no other PW lock on this object; finished. */
1549 lock = ldlm_handle2lock(&lockh);
1553 /* check if lock from the same client */
1554 rc = (lock->l_export->exp_handle.h_cookie ==
1555 info->mti_exp->exp_handle.h_cookie);
1556 LDLM_LOCK_PUT(lock);
1561 * MDT request handler for OST_GETATTR RPC.
1563 * This is data-specific request to get object and layout versions under
1564 * IO lock. It is reliable only for Data-on-MDT files.
1566 * \param[in] tsi target session environment for this request
1568 * \retval 0 if successful
1569 * \retval negative value on error
1571 int mdt_data_version_get(struct tgt_session_info *tsi)
1573 struct mdt_thread_info *mti = mdt_th_info(tsi->tsi_env);
1574 struct mdt_device *mdt = mti->mti_mdt;
1575 struct mdt_body *repbody;
1576 struct mdt_object *mo = mti->mti_object;
1577 struct lov_comp_md_v1 *comp;
1578 struct lustre_handle lh = { 0 };
1581 enum ldlm_mode lock_mode = LCK_PR;
1587 req_capsule_set_size(tsi->tsi_pill, &RMF_MDT_MD, RCL_SERVER, 0);
1588 req_capsule_set_size(tsi->tsi_pill, &RMF_ACL, RCL_SERVER, 0);
1589 rc = req_capsule_server_pack(tsi->tsi_pill);
1590 if (unlikely(rc != 0))
1591 RETURN(err_serious(rc));
1593 repbody = req_capsule_server_get(tsi->tsi_pill, &RMF_MDT_BODY);
1594 if (repbody == NULL)
1597 srvlock = tsi->tsi_mdt_body->mbo_valid & OBD_MD_FLFLAGS &&
1598 tsi->tsi_mdt_body->mbo_flags & OBD_FL_SRVLOCK;
1601 if (unlikely(tsi->tsi_mdt_body->mbo_flags & OBD_FL_FLUSH))
1604 fid_build_reg_res_name(&tsi->tsi_fid, &tsi->tsi_resid);
1605 rc = tgt_mdt_data_lock(mdt->mdt_namespace, &tsi->tsi_resid,
1606 &lh, lock_mode, &flags);
1611 if (!mdt_object_exists(mo))
1612 GOTO(out, rc = -ENOENT);
1613 if (mdt_object_remote(mo))
1614 GOTO(out, rc = -EREMOTE);
1615 if (!S_ISREG(lu_object_attr(&mo->mot_obj)))
1616 GOTO(out, rc = -EBADF);
1618 /* Get version first */
1619 version = dt_version_get(tsi->tsi_env, mdt_obj2dt(mo));
1620 if (version && version != -EOPNOTSUPP) {
1621 repbody->mbo_valid |= OBD_MD_FLDATAVERSION;
1622 /* re-use mbo_ioepoch to transfer version */
1623 repbody->mbo_version = version;
1626 /* Read layout to get its version */
1627 rc = mdt_big_xattr_get(mti, mo, XATTR_NAME_LOV);
1628 if (rc == -ENODATA) /* File has no layout yet */
1633 comp = mti->mti_buf.lb_buf;
1634 if (le32_to_cpu(comp->lcm_magic) != LOV_MAGIC_COMP_V1) {
1635 CDEBUG(D_INFO, DFID" has no composite layout",
1636 PFID(&tsi->tsi_fid));
1637 GOTO(out, rc = -ESTALE);
1640 CDEBUG(D_INODE, DFID": layout version: %u\n",
1641 PFID(&tsi->tsi_fid), le32_to_cpu(comp->lcm_layout_gen));
1643 repbody->mbo_valid |= OBD_MD_LAYOUT_VERSION;
1644 /* re-use mbo_rdev for that */
1645 repbody->mbo_layout_gen = le32_to_cpu(comp->lcm_layout_gen);
1649 tgt_data_unlock(&lh, lock_mode);
1651 repbody->mbo_valid |= OBD_MD_FLFLAGS;
1652 repbody->mbo_flags = OBD_FL_FLUSH;
1656 /* read file data to the buffer */
1657 int mdt_dom_read_on_open(struct mdt_thread_info *mti, struct mdt_device *mdt,
1658 struct lustre_handle *lh)
1660 const struct lu_env *env = mti->mti_env;
1661 struct tgt_session_info *tsi = tgt_ses_info(env);
1662 struct req_capsule *pill = tsi->tsi_pill;
1663 const struct lu_fid *fid;
1664 struct ptlrpc_request *req = tgt_ses_req(tsi);
1665 struct mdt_body *mbo;
1666 struct dt_device *dt = mdt->mdt_bottom;
1667 struct dt_object *mo;
1669 struct niobuf_remote *rnb = NULL;
1670 struct niobuf_local *lnb;
1673 unsigned int len, copied = 0;
1674 __u64 real_dom_size;
1675 int lnbs, nr_local, i;
1676 bool dom_lock = false;
1680 if (!req_capsule_field_present(pill, &RMF_NIOBUF_INLINE, RCL_SERVER)) {
1681 /* There is no reply buffers for this field, this means that
1682 * client has no support for data in reply.
1687 mbo = req_capsule_server_get(pill, &RMF_MDT_BODY);
1688 if (!(mbo->mbo_valid & OBD_MD_DOM_SIZE))
1691 if (!mbo->mbo_dom_size)
1694 if (lustre_handle_is_used(lh)) {
1695 struct ldlm_lock *lock;
1697 lock = ldlm_handle2lock(lh);
1699 dom_lock = ldlm_has_dom(lock) && ldlm_has_layout(lock);
1700 LDLM_LOCK_PUT(lock);
1704 /* return data along with open only along with DoM lock */
1705 if (!dom_lock || !mdt->mdt_opts.mo_dom_read_open)
1708 /* if DoM object holds encrypted content, we need to make sure we
1709 * send whole encryption units, or client will read corrupted content
1711 if (mbo->mbo_valid & LA_FLAGS && mbo->mbo_flags & LUSTRE_ENCRYPT_FL &&
1712 mbo->mbo_dom_size & ~LUSTRE_ENCRYPTION_MASK)
1713 real_dom_size = (mbo->mbo_dom_size & LUSTRE_ENCRYPTION_MASK) +
1714 LUSTRE_ENCRYPTION_UNIT_SIZE;
1716 real_dom_size = mbo->mbo_dom_size;
1718 CDEBUG(D_INFO, "File size %llu, reply sizes %d/%d\n",
1719 real_dom_size, req->rq_reqmsg->lm_repsize, req->rq_replen);
1720 len = req->rq_reqmsg->lm_repsize - req->rq_replen;
1722 /* NB: at this moment we have the following sizes:
1723 * - req->rq_replen: used data in reply
1724 * - req->rq_reqmsg->lm_repsize: total allocated reply buffer at client
1726 * Ideal case when file size fits in allocated reply buffer,
1727 * that mean we can return whole data in reply. We can also fit more
1728 * data up to max_reply_size in total reply size, but this will cause
1729 * re-allocation on client and resend with larger buffer. This is still
1730 * faster than separate READ IO.
1731 * Third case if file is too big to fit even in maximum size, in that
1732 * case we return just tail to optimize possible append.
1734 * At the moment the following strategy is used:
1735 * 1) try to fit into the buffer we have
1736 * 2) return just file tail otherwise.
1738 if (real_dom_size <= len) {
1739 /* can fit whole data */
1740 len = real_dom_size;
1742 } else if (real_dom_size <
1743 mdt_lmm_dom_stripesize(mti->mti_attr.ma_lmm)) {
1746 /* File tail offset must be aligned with larger page size
1747 * between client and server, so the maximum page size is
1748 * used here to align offset.
1750 * NB: DOM feature was introduced when server supports pagebits
1751 * already, so it should be always non-zero value. Report error
1752 * if it is not for some reason.
1754 if (!req->rq_export->exp_target_data.ted_pagebits) {
1755 CERROR("%s: client page bits are not saved on server\n",
1759 pgbits = max_t(int, PAGE_SHIFT,
1760 req->rq_export->exp_target_data.ted_pagebits);
1761 tail = real_dom_size % (1 << pgbits);
1763 /* no partial tail or tail can't fit in reply */
1764 if (tail == 0 || len < tail)
1768 offset = real_dom_size - len;
1770 /* DOM stripe is fully written, so don't expect its tail
1771 * will be used by append.
1776 LASSERT((offset & ~PAGE_MASK) == 0);
1777 rc = req_capsule_server_grow(pill, &RMF_NIOBUF_INLINE,
1778 sizeof(*rnb) + len);
1780 /* failed to grow data buffer, just exit */
1781 GOTO(out, rc = -E2BIG);
1784 /* re-take MDT_BODY and NIOBUF_INLINE buffers after the buffer grow */
1785 mbo = req_capsule_server_get(pill, &RMF_MDT_BODY);
1786 fid = &mbo->mbo_fid1;
1787 if (!fid_is_sane(fid))
1788 GOTO(out, rc = -EINVAL);
1790 rnb = req_capsule_server_get(tsi->tsi_pill, &RMF_NIOBUF_INLINE);
1792 GOTO(out, rc = -EPROTO);
1794 buf = (char *)rnb + sizeof(*rnb);
1796 rnb->rnb_offset = offset;
1798 mo = dt_locate(env, dt, fid);
1800 GOTO(out_rnb, rc = PTR_ERR(mo));
1801 LASSERT(mo != NULL);
1803 dt_read_lock(env, mo, 0);
1804 if (!dt_object_exists(mo))
1805 GOTO(unlock, rc = -ENOENT);
1807 /* parse remote buffers to local buffers and prepare the latter */
1808 lnbs = (len >> PAGE_SHIFT) + 1;
1809 OBD_ALLOC_PTR_ARRAY(lnb, lnbs);
1811 GOTO(unlock, rc = -ENOMEM);
1813 rc = dt_bufs_get(env, mo, rnb, lnb, lnbs, 0);
1814 if (unlikely(rc < 0))
1816 LASSERT(rc <= lnbs);
1818 rc = dt_read_prep(env, mo, lnb, nr_local);
1821 /* copy data to the buffer finally */
1822 for (i = 0; i < nr_local; i++) {
1823 char *p = kmap(lnb[i].lnb_page);
1826 LASSERT(lnb[i].lnb_page_offset == 0);
1827 off = lnb[i].lnb_len & ~PAGE_MASK;
1829 memset(p + off, 0, PAGE_SIZE - off);
1831 memcpy(buf + (i << PAGE_SHIFT), p, lnb[i].lnb_len);
1832 kunmap(lnb[i].lnb_page);
1833 copied += lnb[i].lnb_len;
1836 CDEBUG(D_INFO, "Read %i (wanted %u) bytes from %llu\n", copied,
1839 CWARN("%s: read %i bytes for "DFID
1840 " but wanted %u, is size wrong?\n",
1841 tsi->tsi_exp->exp_obd->obd_name, copied,
1842 PFID(&tsi->tsi_fid), len);
1843 /* Ignore partially copied data */
1848 dt_bufs_put(env, mo, lnb, nr_local);
1850 OBD_FREE_PTR_ARRAY(lnb, lnbs);
1852 dt_read_unlock(env, mo);
1853 lu_object_put(env, &mo->do_lu);
1855 rnb->rnb_len = copied;
1857 /* Don't fail OPEN request if read-on-open is failed, but drop
1858 * a message in log about the error.
1861 CDEBUG(D_INFO, "Read-on-open is failed, rc = %d", rc);
1867 * Completion AST for DOM discard locks:
1869 * CP AST an DOM discard lock is called always right after enqueue or from
1870 * reprocess if lock was blocked, in the latest case l_ast_data is set to
1871 * the mdt_object which is kept while there are pending locks on it.
1873 int ldlm_dom_discard_cp_ast(struct ldlm_lock *lock, __u64 flags, void *data)
1875 struct mdt_object *mo;
1876 struct lustre_handle dom_lh;
1881 /* l_ast_data is set when lock was not granted immediately
1882 * in mdt_dom_discard_data() below but put into waiting list,
1883 * so this CP callback means we are finished and corresponding
1884 * MDT object should be released finally as well as lock itself.
1886 lock_res_and_lock(lock);
1887 if (!lock->l_ast_data) {
1888 unlock_res_and_lock(lock);
1892 mo = lock->l_ast_data;
1893 lock->l_ast_data = NULL;
1894 unlock_res_and_lock(lock);
1896 ldlm_lock2handle(lock, &dom_lh);
1897 ldlm_lock_decref(&dom_lh, LCK_PW);
1899 env = lu_env_find();
1901 mdt_object_put(env, mo);
1906 void mdt_dom_discard_data(struct mdt_thread_info *info,
1907 struct mdt_object *mo)
1909 struct ptlrpc_request *req = mdt_info_req(info);
1910 struct mdt_device *mdt = mdt_dev(mo->mot_obj.lo_dev);
1911 union ldlm_policy_data policy;
1912 struct ldlm_res_id res_id;
1913 struct lustre_handle dom_lh;
1914 struct ldlm_lock *lock;
1915 __u64 flags = LDLM_FL_AST_DISCARD_DATA;
1921 if (req && req_is_replay(req))
1924 policy.l_inodebits.bits = MDS_INODELOCK_DOM;
1925 policy.l_inodebits.try_bits = 0;
1926 fid_build_reg_res_name(mdt_object_fid(mo), &res_id);
1928 /* Keep blocking version of discard for an old client to avoid
1929 * crashes on non-patched clients. LU-11359.
1931 old_client = req && !(exp_connect_flags2(req->rq_export) &
1932 OBD_CONNECT2_ASYNC_DISCARD);
1934 /* Tell the clients that the object is gone now and that they should
1935 * throw away any cached pages. */
1936 rc = ldlm_cli_enqueue_local(info->mti_env, mdt->mdt_namespace, &res_id,
1937 LDLM_IBITS, &policy, LCK_PW, &flags,
1938 ldlm_blocking_ast, old_client ?
1939 ldlm_completion_ast :
1940 ldlm_dom_discard_cp_ast,
1941 NULL, NULL, 0, LVB_T_NONE, NULL, &dom_lh);
1942 if (rc != ELDLM_OK) {
1944 "Failed to issue discard lock, rc = %d\n", rc);
1948 lock = ldlm_handle2lock(&dom_lh);
1949 lock_res_and_lock(lock);
1950 /* if lock is not granted then there are BL ASTs in progress and
1951 * lock will be granted in result of reprocessing with CP callback
1952 * notifying about that. The mdt object has to be kept until that and
1953 * it is saved in l_ast_data of the lock. Lock reference is kept too
1954 * until that to prevent it from canceling.
1956 if (!is_granted_or_cancelled_nolock(lock)) {
1957 mdt_object_get(info->mti_env, mo);
1958 lock->l_ast_data = mo;
1959 unlock_res_and_lock(lock);
1961 unlock_res_and_lock(lock);
1962 ldlm_lock_decref_and_cancel(&dom_lh, LCK_PW);
1964 LDLM_LOCK_PUT(lock);