4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2017, Intel Corporation.
28 * Author: Mikhail Pershin <mike.pershin@intel.com>
31 #define DEBUG_SUBSYSTEM S_FILTER
33 #include <dt_object.h>
34 #include "mdt_internal.h"
36 /* functions below are stubs for now, they will be implemented with
37 * grant support on MDT */
38 static inline void mdt_io_counter_incr(struct obd_export *exp, int opcode,
39 char *jobid, long amount)
44 static inline void mdt_dom_read_lock(struct mdt_object *mo)
46 down_read(&mo->mot_dom_sem);
49 static inline void mdt_dom_read_unlock(struct mdt_object *mo)
51 up_read(&mo->mot_dom_sem);
54 static inline void mdt_dom_write_lock(struct mdt_object *mo)
56 down_write(&mo->mot_dom_sem);
59 static inline void mdt_dom_write_unlock(struct mdt_object *mo)
61 up_write(&mo->mot_dom_sem);
65 * Lock prolongation for Data-on-MDT.
66 * This is similar to OFD code but for DOM ibits lock.
68 static inline time64_t prolong_timeout(struct ptlrpc_request *req)
70 struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
74 return obd_timeout / 2;
76 req_timeout = req->rq_deadline - req->rq_arrival_time.tv_sec;
77 return max_t(time64_t, at_est2timeout(at_get(&svcpt->scp_at_estimate)),
81 static void mdt_dom_resource_prolong(struct ldlm_prolong_args *arg)
83 struct ldlm_resource *res;
84 struct ldlm_lock *lock;
88 res = ldlm_resource_get(arg->lpa_export->exp_obd->obd_namespace, NULL,
89 &arg->lpa_resid, LDLM_EXTENT, 0);
92 "Failed to get resource for resid %llu/%llu\n",
93 arg->lpa_resid.name[0], arg->lpa_resid.name[1]);
98 list_for_each_entry(lock, &res->lr_granted, l_res_link) {
99 if (ldlm_has_dom(lock)) {
100 LDLM_DEBUG(lock, "DOM lock to prolong ");
101 ldlm_lock_prolong_one(lock, arg);
106 ldlm_resource_putref(res);
111 static void mdt_prolong_dom_lock(struct tgt_session_info *tsi,
112 struct ldlm_prolong_args *data)
114 struct obdo *oa = &tsi->tsi_ost_body->oa;
115 struct ldlm_lock *lock;
119 data->lpa_timeout = prolong_timeout(tgt_ses_req(tsi));
120 data->lpa_export = tsi->tsi_exp;
121 data->lpa_resid = tsi->tsi_resid;
123 CDEBUG(D_RPCTRACE, "Prolong DOM lock for req %p with x%llu\n",
124 tgt_ses_req(tsi), tgt_ses_req(tsi)->rq_xid);
126 if (oa->o_valid & OBD_MD_FLHANDLE) {
127 /* mostly a request should be covered by only one lock, try
129 lock = ldlm_handle2lock(&oa->o_handle);
131 LASSERT(lock->l_export == data->lpa_export);
132 ldlm_lock_prolong_one(lock, data);
133 lock->l_last_used = ktime_get();
135 if (data->lpa_locks_cnt > 0)
139 mdt_dom_resource_prolong(data);
143 static int mdt_rw_hpreq_lock_match(struct ptlrpc_request *req,
144 struct ldlm_lock *lock)
146 struct obd_ioobj *ioo;
148 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
152 if (!(lock->l_policy_data.l_inodebits.bits & MDS_INODELOCK_DOM))
155 ioo = req_capsule_client_get(&req->rq_pill, &RMF_OBD_IOOBJ);
156 LASSERT(ioo != NULL);
158 LASSERT(lock->l_resource != NULL);
159 if (!fid_res_name_eq(&ioo->ioo_oid.oi_fid, &lock->l_resource->lr_name))
162 /* a bulk write can only hold a reference on a PW extent lock. */
165 /* whereas a bulk read can be protected by either a PR or PW
169 if (!(lock->l_granted_mode & mode))
175 static int mdt_rw_hpreq_check(struct ptlrpc_request *req)
177 struct tgt_session_info *tsi;
178 struct obd_ioobj *ioo;
179 struct niobuf_remote *rnb;
181 struct ldlm_prolong_args pa = { 0 };
185 /* Don't use tgt_ses_info() to get session info, because lock_match()
186 * can be called while request has no processing thread yet. */
187 tsi = lu_context_key_get(&req->rq_session, &tgt_session_key);
190 * Use LASSERT below because malformed RPCs should have
191 * been filtered out in tgt_hpreq_handler().
193 opc = lustre_msg_get_opc(req->rq_reqmsg);
194 LASSERT(opc == OST_READ || opc == OST_WRITE);
196 ioo = req_capsule_client_get(&req->rq_pill, &RMF_OBD_IOOBJ);
197 LASSERT(ioo != NULL);
199 rnb = req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE);
200 LASSERT(rnb != NULL);
201 LASSERT(!(rnb->rnb_flags & OBD_BRW_SRVLOCK));
203 pa.lpa_mode = LCK_PW;
205 pa.lpa_mode |= LCK_PR;
207 DEBUG_REQ(D_RPCTRACE, req, "%s %s: refresh rw locks: "DFID"\n",
208 tgt_name(tsi->tsi_tgt), current->comm, PFID(&tsi->tsi_fid));
210 mdt_prolong_dom_lock(tsi, &pa);
212 if (pa.lpa_blocks_cnt > 0) {
214 "%s: refreshed %u locks timeout for req %p.\n",
215 tgt_name(tsi->tsi_tgt), pa.lpa_blocks_cnt, req);
219 RETURN(pa.lpa_locks_cnt > 0 ? 0 : -ESTALE);
222 static void mdt_rw_hpreq_fini(struct ptlrpc_request *req)
224 mdt_rw_hpreq_check(req);
227 static struct ptlrpc_hpreq_ops mdt_hpreq_rw = {
228 .hpreq_lock_match = mdt_rw_hpreq_lock_match,
229 .hpreq_check = mdt_rw_hpreq_check,
230 .hpreq_fini = mdt_rw_hpreq_fini
234 * Assign high priority operations to an IO request.
236 * Check if the incoming request is a candidate for
237 * high-priority processing. If it is, assign it a high
238 * priority operations table.
240 * \param[in] tsi target session environment for this request
242 void mdt_hp_brw(struct tgt_session_info *tsi)
244 struct niobuf_remote *rnb;
245 struct obd_ioobj *ioo;
249 ioo = req_capsule_client_get(tsi->tsi_pill, &RMF_OBD_IOOBJ);
250 LASSERT(ioo != NULL); /* must exist after request preprocessing */
251 if (ioo->ioo_bufcnt > 0) {
252 rnb = req_capsule_client_get(tsi->tsi_pill, &RMF_NIOBUF_REMOTE);
253 LASSERT(rnb != NULL); /* must exist after preprocessing */
255 /* no high priority if server lock is needed */
256 if (rnb->rnb_flags & OBD_BRW_SRVLOCK ||
257 (lustre_msg_get_flags(tgt_ses_req(tsi)->rq_reqmsg) &
261 tgt_ses_req(tsi)->rq_ops = &mdt_hpreq_rw;
264 static int mdt_punch_hpreq_lock_match(struct ptlrpc_request *req,
265 struct ldlm_lock *lock)
267 struct tgt_session_info *tsi;
272 /* Don't use tgt_ses_info() to get session info, because lock_match()
273 * can be called while request has no processing thread yet. */
274 tsi = lu_context_key_get(&req->rq_session, &tgt_session_key);
277 * Use LASSERT below because malformed RPCs should have
278 * been filtered out in tgt_hpreq_handler().
280 LASSERT(tsi->tsi_ost_body != NULL);
281 if (tsi->tsi_ost_body->oa.o_valid & OBD_MD_FLHANDLE &&
282 tsi->tsi_ost_body->oa.o_handle.cookie == lock->l_handle.h_cookie)
285 oa = &tsi->tsi_ost_body->oa;
287 LASSERT(lock->l_resource != NULL);
288 if (!fid_res_name_eq(&oa->o_oi.oi_fid, &lock->l_resource->lr_name))
291 if (!(lock->l_granted_mode & LCK_PW))
298 * Implementation of ptlrpc_hpreq_ops::hpreq_lock_check for OST_PUNCH request.
300 * High-priority queue request check for whether the given punch request
301 * (\a req) is blocking an LDLM lock cancel. Also checks whether the request is
302 * covered by an LDLM lock.
306 * \param[in] req the incoming request
308 * \retval 1 if \a req is blocking an LDLM lock cancel
309 * \retval 0 if it is not
310 * \retval -ESTALE if lock is not found
312 static int mdt_punch_hpreq_check(struct ptlrpc_request *req)
314 struct tgt_session_info *tsi;
316 struct ldlm_prolong_args pa = { 0 };
320 /* Don't use tgt_ses_info() to get session info, because lock_match()
321 * can be called while request has no processing thread yet. */
322 tsi = lu_context_key_get(&req->rq_session, &tgt_session_key);
323 LASSERT(tsi != NULL);
324 oa = &tsi->tsi_ost_body->oa;
326 LASSERT(!(oa->o_valid & OBD_MD_FLFLAGS &&
327 oa->o_flags & OBD_FL_SRVLOCK));
329 pa.lpa_mode = LCK_PW;
331 CDEBUG(D_DLMTRACE, "%s: refresh DOM lock for "DFID"\n",
332 tgt_name(tsi->tsi_tgt), PFID(&tsi->tsi_fid));
334 mdt_prolong_dom_lock(tsi, &pa);
337 if (pa.lpa_blocks_cnt > 0) {
339 "%s: refreshed %u locks timeout for req %p.\n",
340 tgt_name(tsi->tsi_tgt), pa.lpa_blocks_cnt, req);
344 RETURN(pa.lpa_locks_cnt > 0 ? 0 : -ESTALE);
348 * Implementation of ptlrpc_hpreq_ops::hpreq_lock_fini for OST_PUNCH request.
350 * Called after the request has been handled. It refreshes lock timeout again
351 * so that client has more time to send lock cancel RPC.
353 * \param[in] req request which is being processed.
355 static void mdt_punch_hpreq_fini(struct ptlrpc_request *req)
357 mdt_punch_hpreq_check(req);
360 static struct ptlrpc_hpreq_ops mdt_hpreq_punch = {
361 .hpreq_lock_match = mdt_punch_hpreq_lock_match,
362 .hpreq_check = mdt_punch_hpreq_check,
363 .hpreq_fini = mdt_punch_hpreq_fini
366 void mdt_hp_punch(struct tgt_session_info *tsi)
368 LASSERT(tsi->tsi_ost_body != NULL); /* must exists if we are here */
369 /* no high-priority if server lock is needed */
370 if ((tsi->tsi_ost_body->oa.o_valid & OBD_MD_FLFLAGS &&
371 tsi->tsi_ost_body->oa.o_flags & OBD_FL_SRVLOCK) ||
372 tgt_conn_flags(tsi) & OBD_CONNECT_MDS ||
373 lustre_msg_get_flags(tgt_ses_req(tsi)->rq_reqmsg) & MSG_REPLAY)
375 tgt_ses_req(tsi)->rq_ops = &mdt_hpreq_punch;
378 static int mdt_preprw_read(const struct lu_env *env, struct obd_export *exp,
379 struct mdt_device *mdt, struct mdt_object *mo,
380 struct lu_attr *la, int niocount,
381 struct niobuf_remote *rnb, int *nr_local,
382 struct niobuf_local *lnb, char *jobid)
384 struct dt_object *dob;
385 int i, j, rc, tot_bytes = 0;
390 mdt_dom_read_lock(mo);
392 /* the only valid case when READ can find object is missing or stale
393 * when export is just evicted and open files are closed forcefully
394 * on server while client's READ can be in progress.
395 * This should not happen on healthy export, object can't be missing
396 * or dying because both states means it was finally destroyed.
398 level = exp->exp_failed ? D_INFO : D_ERROR;
399 if (!mdt_object_exists(mo)) {
401 "%s: READ IO to missing obj "DFID": rc = %d\n",
402 exp->exp_obd->obd_name, PFID(mdt_object_fid(mo)),
404 /* return 0 and continue with empty commit to skip such READ
405 * without more BRW errors.
409 if (lu_object_is_dying(&mo->mot_header)) {
411 "%s: READ IO to stale obj "DFID": rc = %d\n",
412 exp->exp_obd->obd_name, PFID(mdt_object_fid(mo)),
414 /* return 0 and continue with empty commit to skip such READ
415 * without more BRW errors.
420 dob = mdt_obj2dt(mo);
421 /* parse remote buffers to local buffers and prepare the latter */
422 for (i = 0, j = 0; i < niocount; i++) {
423 rc = dt_bufs_get(env, dob, rnb + i, lnb + j, 0);
424 if (unlikely(rc < 0))
426 /* correct index for local buffers to continue with */
429 tot_bytes += rnb[i].rnb_len;
432 rc = dt_attr_get(env, dob, la);
436 rc = dt_read_prep(env, dob, lnb, *nr_local);
440 mdt_io_counter_incr(exp, LPROC_MDT_IO_READ, jobid, tot_bytes);
443 dt_bufs_put(env, dob, lnb, *nr_local);
444 mdt_dom_read_unlock(mo);
448 static int mdt_preprw_write(const struct lu_env *env, struct obd_export *exp,
449 struct mdt_device *mdt, struct mdt_object *mo,
450 struct lu_attr *la, struct obdo *oa,
451 int objcount, struct obd_ioobj *obj,
452 struct niobuf_remote *rnb, int *nr_local,
453 struct niobuf_local *lnb, char *jobid)
455 struct dt_object *dob;
456 int i, j, k, rc = 0, tot_bytes = 0;
460 /* Process incoming grant info, set OBD_BRW_GRANTED flag and grant some
461 * space back if possible */
462 tgt_grant_prepare_write(env, exp, oa, rnb, obj->ioo_bufcnt);
464 mdt_dom_read_lock(mo);
466 /* don't report error in cases with failed export */
467 if (!mdt_object_exists(mo)) {
468 int level = exp->exp_failed ? D_INFO : D_ERROR;
472 "%s: WRITE IO to missing obj "DFID": rc = %d\n",
473 exp->exp_obd->obd_name, PFID(mdt_object_fid(mo)),
475 /* exit with no data written, note nr_local = 0 above */
478 if (lu_object_is_dying(&mo->mot_header)) {
479 /* This is possible race between object destroy followed by
480 * discard BL AST and client cache flushing. Object is
481 * referenced until discard finish.
483 CDEBUG(D_INODE, "WRITE IO to stale object "DFID"\n",
484 PFID(mdt_object_fid(mo)));
485 /* Note: continue with no error here to don't cause BRW errors
486 * but skip transaction in commitrw silently so no data is
491 dob = mdt_obj2dt(mo);
492 /* parse remote buffers to local buffers and prepare the latter */
493 for (i = 0, j = 0; i < obj->ioo_bufcnt; i++) {
494 rc = dt_bufs_get(env, dob, rnb + i, lnb + j, 1);
495 if (unlikely(rc < 0))
497 /* correct index for local buffers to continue with */
498 for (k = 0; k < rc; k++) {
499 lnb[j + k].lnb_flags = rnb[i].rnb_flags;
500 if (!(rnb[i].rnb_flags & OBD_BRW_GRANTED))
501 lnb[j + k].lnb_rc = -ENOSPC;
505 tot_bytes += rnb[i].rnb_len;
508 rc = dt_write_prep(env, dob, lnb, *nr_local);
512 mdt_io_counter_incr(exp, LPROC_MDT_IO_WRITE, jobid, tot_bytes);
515 dt_bufs_put(env, dob, lnb, *nr_local);
517 mdt_dom_read_unlock(mo);
518 /* tgt_grant_prepare_write() was called, so we must commit */
519 tgt_grant_commit(exp, oa->o_grant_used, rc);
520 /* let's still process incoming grant information packed in the oa,
521 * but without enforcing grant since we won't proceed with the write.
522 * Just like a read request actually. */
523 tgt_grant_prepare_read(env, exp, oa);
527 int mdt_obd_preprw(const struct lu_env *env, int cmd, struct obd_export *exp,
528 struct obdo *oa, int objcount, struct obd_ioobj *obj,
529 struct niobuf_remote *rnb, int *nr_local,
530 struct niobuf_local *lnb)
532 struct tgt_session_info *tsi = tgt_ses_info(env);
533 struct mdt_thread_info *info = tsi2mdt_info(tsi);
534 struct lu_attr *la = &info->mti_attr.ma_attr;
535 struct mdt_device *mdt = mdt_dev(exp->exp_obd->obd_lu_dev);
536 struct mdt_object *mo;
540 /* The default value PTLRPC_MAX_BRW_PAGES is set in tgt_brw_write()
541 * but for MDT it is different, correct it here. */
542 if (*nr_local > MD_MAX_BRW_PAGES)
543 *nr_local = MD_MAX_BRW_PAGES;
545 jobid = tsi->tsi_jobid;
547 if (!oa || objcount != 1 || obj->ioo_bufcnt == 0) {
548 CERROR("%s: bad parameters %p/%i/%i\n",
549 exp->exp_obd->obd_name, oa, objcount, obj->ioo_bufcnt);
553 mo = mdt_object_find(env, mdt, &tsi->tsi_fid);
555 GOTO(out, rc = PTR_ERR(mo));
557 LASSERT(info->mti_object == NULL);
558 info->mti_object = mo;
560 if (cmd == OBD_BRW_WRITE) {
561 la_from_obdo(la, oa, OBD_MD_FLGETATTR);
562 rc = mdt_preprw_write(env, exp, mdt, mo, la, oa,
563 objcount, obj, rnb, nr_local, lnb,
565 } else if (cmd == OBD_BRW_READ) {
566 tgt_grant_prepare_read(env, exp, oa);
567 rc = mdt_preprw_read(env, exp, mdt, mo, la,
568 obj->ioo_bufcnt, rnb, nr_local, lnb,
570 obdo_from_la(oa, la, LA_ATIME);
572 CERROR("%s: wrong cmd %d received!\n",
573 exp->exp_obd->obd_name, cmd);
577 lu_object_put(env, &mo->mot_obj);
578 info->mti_object = NULL;
584 static int mdt_commitrw_read(const struct lu_env *env, struct mdt_device *mdt,
585 struct mdt_object *mo, int objcount, int niocount,
586 struct niobuf_local *lnb)
588 struct dt_object *dob;
593 dob = mdt_obj2dt(mo);
596 dt_bufs_put(env, dob, lnb, niocount);
598 mdt_dom_read_unlock(mo);
602 static int mdt_commitrw_write(const struct lu_env *env, struct obd_export *exp,
603 struct mdt_device *mdt, struct mdt_object *mo,
604 struct lu_attr *la, int objcount, int niocount,
605 struct niobuf_local *lnb, unsigned long granted,
608 struct dt_device *dt = mdt->mdt_bottom;
609 struct dt_object *dob;
617 dob = mdt_obj2dt(mo);
620 GOTO(out, rc = old_rc);
622 la->la_valid &= LA_ATIME | LA_MTIME | LA_CTIME;
624 if (!dt_object_exists(dob))
625 GOTO(out, rc = -ENOENT);
626 if (lu_object_is_dying(&mo->mot_header)) {
627 /* Commit to stale object can be just skipped silently. */
628 CDEBUG(D_INODE, "skip commit to stale object "DFID"\n",
629 PFID(mdt_object_fid(mo)));
635 DEBUG_REQ(D_WARNING, tgt_ses_req(tgt_ses_info(env)),
636 "%s: commit with no pages for "DFID": rc = %d\n",
637 exp->exp_obd->obd_name, PFID(mdt_object_fid(mo)), rc);
641 th = dt_trans_create(env, dt);
643 GOTO(out, rc = PTR_ERR(th));
645 for (i = 0; i < niocount; i++) {
646 if (!(lnb[i].lnb_flags & OBD_BRW_ASYNC)) {
652 if (OBD_FAIL_CHECK(OBD_FAIL_OST_DQACQ_NET))
653 GOTO(out_stop, rc = -EINPROGRESS);
655 rc = dt_declare_write_commit(env, dob, lnb, niocount, th);
660 /* update [mac]time if needed */
661 rc = dt_declare_attr_set(env, dob, la, th);
666 tgt_vbr_obj_set(env, dob);
667 rc = dt_trans_start(env, dt, th);
671 dt_write_lock(env, dob, 0);
672 rc = dt_write_commit(env, dob, lnb, niocount, th);
677 rc = dt_attr_set(env, dob, la, th);
681 /* get attr to return */
682 rc = dt_attr_get(env, dob, la);
684 dt_write_unlock(env, dob);
687 /* Force commit to make the just-deleted blocks
688 * reusable. LU-456 */
693 if (rc == 0 && granted > 0) {
694 if (tgt_grant_commit_cb_add(th, exp, granted) == 0)
699 dt_trans_stop(env, dt, th);
700 if (rc == -ENOSPC && retries++ < 3) {
701 CDEBUG(D_INODE, "retry after force commit, retries:%d\n",
707 dt_bufs_put(env, dob, lnb, niocount);
708 mdt_dom_read_unlock(mo);
710 tgt_grant_commit(exp, granted, old_rc);
714 void mdt_dom_obj_lvb_update(const struct lu_env *env, struct mdt_object *mo,
717 struct mdt_device *mdt = mdt_dev(mo->mot_obj.lo_dev);
718 struct ldlm_res_id resid;
719 struct ldlm_resource *res;
721 fid_build_reg_res_name(mdt_object_fid(mo), &resid);
722 res = ldlm_resource_get(mdt->mdt_namespace, NULL, &resid,
727 /* Update lvbo data if exists. */
728 if (mdt_dom_lvb_is_valid(res))
729 mdt_dom_disk_lvbo_update(env, mo, res, increase_only);
730 ldlm_resource_putref(res);
733 int mdt_obd_commitrw(const struct lu_env *env, int cmd, struct obd_export *exp,
734 struct obdo *oa, int objcount, struct obd_ioobj *obj,
735 struct niobuf_remote *rnb, int npages,
736 struct niobuf_local *lnb, int old_rc)
738 struct mdt_thread_info *info = mdt_th_info(env);
739 struct mdt_device *mdt = mdt_dev(exp->exp_obd->obd_lu_dev);
740 struct mdt_object *mo = info->mti_object;
741 struct lu_attr *la = &info->mti_attr.ma_attr;
747 if (cmd == OBD_BRW_WRITE) {
748 /* Don't update timestamps if this write is older than a
749 * setattr which modifies the timestamps. b=10150 */
751 /* XXX when we start having persistent reservations this needs
752 * to be changed to ofd_fmd_get() to create the fmd if it
753 * doesn't already exist so we can store the reservation handle
755 valid = OBD_MD_FLUID | OBD_MD_FLGID;
756 if (tgt_fmd_check(exp, mdt_object_fid(mo),
757 mdt_info_req(info)->rq_xid))
758 valid |= OBD_MD_FLATIME | OBD_MD_FLMTIME |
761 la_from_obdo(la, oa, valid);
763 rc = mdt_commitrw_write(env, exp, mdt, mo, la, objcount,
764 npages, lnb, oa->o_grant_used, old_rc);
766 obdo_from_la(oa, la, VALID_FLAGS | LA_GID | LA_UID);
768 obdo_from_la(oa, la, LA_GID | LA_UID);
770 mdt_dom_obj_lvb_update(env, mo, false);
771 /* don't report overquota flag if we failed before reaching
773 if (old_rc == 0 && (rc == 0 || rc == -EDQUOT)) {
774 /* return the overquota flags to client */
775 if (lnb[0].lnb_flags & OBD_BRW_OVER_USRQUOTA) {
776 if (oa->o_valid & OBD_MD_FLFLAGS)
777 oa->o_flags |= OBD_FL_NO_USRQUOTA;
779 oa->o_flags = OBD_FL_NO_USRQUOTA;
782 if (lnb[0].lnb_flags & OBD_BRW_OVER_GRPQUOTA) {
783 if (oa->o_valid & OBD_MD_FLFLAGS)
784 oa->o_flags |= OBD_FL_NO_GRPQUOTA;
786 oa->o_flags = OBD_FL_NO_GRPQUOTA;
789 if (lnb[0].lnb_flags & OBD_BRW_OVER_PRJQUOTA) {
790 if (oa->o_valid & OBD_MD_FLFLAGS)
791 oa->o_flags |= OBD_FL_NO_PRJQUOTA;
793 oa->o_flags = OBD_FL_NO_PRJQUOTA;
796 oa->o_valid |= OBD_MD_FLFLAGS | OBD_MD_FLUSRQUOTA |
797 OBD_MD_FLGRPQUOTA | OBD_MD_FLPRJQUOTA;
799 } else if (cmd == OBD_BRW_READ) {
800 /* If oa != NULL then mdt_preprw_read updated the inode
801 * atime and we should update the lvb so that other glimpses
802 * will also get the updated value. bug 5972 */
804 mdt_dom_obj_lvb_update(env, mo, true);
805 rc = mdt_commitrw_read(env, mdt, mo, objcount, npages, lnb);
811 mdt_thread_info_fini(info);
815 int mdt_object_punch(const struct lu_env *env, struct dt_device *dt,
816 struct dt_object *dob, __u64 start, __u64 end,
824 /* we support truncate, not punch yet */
825 LASSERT(end == OBD_OBJECT_EOF);
827 if (!dt_object_exists(dob))
830 th = dt_trans_create(env, dt);
834 rc = dt_declare_attr_set(env, dob, la, th);
838 rc = dt_declare_punch(env, dob, start, OBD_OBJECT_EOF, th);
842 tgt_vbr_obj_set(env, dob);
843 rc = dt_trans_start(env, dt, th);
847 dt_write_lock(env, dob, 0);
848 rc = dt_punch(env, dob, start, OBD_OBJECT_EOF, th);
851 rc = dt_attr_set(env, dob, la, th);
855 dt_write_unlock(env, dob);
858 dt_trans_stop(env, dt, th);
862 int mdt_punch_hdl(struct tgt_session_info *tsi)
864 const struct obdo *oa = &tsi->tsi_ost_body->oa;
865 struct ost_body *repbody;
866 struct mdt_thread_info *info;
868 struct ldlm_namespace *ns = tsi->tsi_tgt->lut_obd->obd_namespace;
869 struct obd_export *exp = tsi->tsi_exp;
870 struct mdt_device *mdt = mdt_dev(exp->exp_obd->obd_lu_dev);
871 struct mdt_object *mo;
872 struct dt_object *dob;
874 struct lustre_handle lh = { 0, };
881 /* check that we do support OBD_CONNECT_TRUNCLOCK. */
882 CLASSERT(OST_CONNECT_SUPPORTED & OBD_CONNECT_TRUNCLOCK);
884 if ((oa->o_valid & (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS)) !=
885 (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS))
886 RETURN(err_serious(-EPROTO));
888 repbody = req_capsule_server_get(tsi->tsi_pill, &RMF_OST_BODY);
890 RETURN(err_serious(-ENOMEM));
892 /* punch start,end are passed in o_size,o_blocks throught wire */
896 if (end != OBD_OBJECT_EOF) /* Only truncate is supported */
899 info = tsi2mdt_info(tsi);
900 la = &info->mti_attr.ma_attr;
901 /* standard truncate optimization: if file body is completely
902 * destroyed, don't send data back to the server. */
904 flags |= LDLM_FL_AST_DISCARD_DATA;
906 repbody->oa.o_oi = oa->o_oi;
907 repbody->oa.o_valid = OBD_MD_FLID;
909 srvlock = (exp_connect_flags(exp) & OBD_CONNECT_SRVLOCK) &&
910 oa->o_valid & OBD_MD_FLFLAGS &&
911 oa->o_flags & OBD_FL_SRVLOCK;
914 rc = tgt_mdt_data_lock(ns, &tsi->tsi_resid, &lh, LCK_PW,
920 CDEBUG(D_INODE, "calling punch for object "DFID", valid = %#llx"
921 ", start = %lld, end = %lld\n", PFID(&tsi->tsi_fid),
922 oa->o_valid, start, end);
924 mo = mdt_object_find(tsi->tsi_env, mdt, &tsi->tsi_fid);
926 GOTO(out_unlock, rc = PTR_ERR(mo));
928 if (!mdt_object_exists(mo))
929 GOTO(out_put, rc = -ENOENT);
931 /* Shouldn't happen on dirs */
932 if (S_ISDIR(lu_object_attr(&mo->mot_obj))) {
934 CERROR("%s: Truncate on dir "DFID": rc = %d\n",
935 exp->exp_obd->obd_name, PFID(&tsi->tsi_fid), rc);
939 mdt_dom_write_lock(mo);
940 dob = mdt_obj2dt(mo);
942 la_from_obdo(la, oa, OBD_MD_FLMTIME | OBD_MD_FLATIME | OBD_MD_FLCTIME);
944 la->la_valid |= LA_SIZE;
946 /* MDT supports FMD for Data-on-MDT needs */
947 if (la->la_valid & (LA_ATIME | LA_MTIME | LA_CTIME))
948 tgt_fmd_update(tsi->tsi_exp, &tsi->tsi_fid,
949 tgt_ses_req(tsi)->rq_xid);
951 rc = mdt_object_punch(tsi->tsi_env, mdt->mdt_bottom, dob,
953 mdt_dom_write_unlock(mo);
957 mdt_dom_obj_lvb_update(tsi->tsi_env, mo, false);
958 mdt_io_counter_incr(tsi->tsi_exp, LPROC_MDT_IO_PUNCH,
962 lu_object_put(tsi->tsi_env, &mo->mot_obj);
965 tgt_extent_unlock(&lh, LCK_PW);
967 mdt_thread_info_fini(info);
972 * MDT glimpse for Data-on-MDT
974 * If there is write lock on client then function issues glimpse_ast to get
975 * an actual size from that client.
978 int mdt_do_glimpse(const struct lu_env *env, struct ldlm_namespace *ns,
979 struct ldlm_resource *res)
981 union ldlm_policy_data policy;
982 struct lustre_handle lockh;
984 struct ldlm_lock *lock;
985 struct ldlm_glimpse_work *gl_work;
986 struct list_head gl_list;
991 /* There can be only one write lock covering data, try to match it. */
992 policy.l_inodebits.bits = MDS_INODELOCK_DOM;
993 mode = ldlm_lock_match(ns, LDLM_FL_TEST_LOCK,
994 &res->lr_name, LDLM_IBITS, &policy,
997 /* There is no PW lock on this object; finished. */
1001 lock = ldlm_handle2lock(&lockh);
1006 * This check is for lock taken in mdt_reint_unlink() that does
1007 * not have l_glimpse_ast set. So the logic is: if there is a lock
1008 * with no l_glimpse_ast set, this object is being destroyed already.
1009 * Hence, if you are grabbing DLM locks on the server, always set
1010 * non-NULL glimpse_ast (e.g., ldlm_request.c::ldlm_glimpse_ast()).
1012 if (lock->l_glimpse_ast == NULL) {
1013 LDLM_DEBUG(lock, "no l_glimpse_ast");
1014 GOTO(out, rc = -ENOENT);
1017 OBD_SLAB_ALLOC_PTR_GFP(gl_work, ldlm_glimpse_work_kmem, GFP_ATOMIC);
1019 GOTO(out, rc = -ENOMEM);
1021 /* Populate the gl_work structure.
1022 * Grab additional reference on the lock which will be released in
1023 * ldlm_work_gl_ast_lock() */
1024 gl_work->gl_lock = LDLM_LOCK_GET(lock);
1025 /* The glimpse callback is sent to one single IO lock. As a result,
1026 * the gl_work list is just composed of one element */
1027 INIT_LIST_HEAD(&gl_list);
1028 list_add_tail(&gl_work->gl_list, &gl_list);
1029 /* There is actually no need for a glimpse descriptor when glimpsing
1031 gl_work->gl_desc = NULL;
1032 /* the ldlm_glimpse_work structure is allocated on the stack */
1033 gl_work->gl_flags = LDLM_GL_WORK_SLAB_ALLOCATED;
1035 ldlm_glimpse_locks(res, &gl_list); /* this will update the LVB */
1037 /* If the list is not empty, we failed to glimpse a lock and
1038 * must clean it up. Usually due to a race with unlink.*/
1039 if (!list_empty(&gl_list)) {
1040 LDLM_LOCK_RELEASE(lock);
1041 OBD_SLAB_FREE_PTR(gl_work, ldlm_glimpse_work_kmem);
1046 LDLM_LOCK_PUT(lock);
1050 static void mdt_lvb2body(struct ldlm_resource *res, struct mdt_body *mb)
1052 struct ost_lvb *res_lvb;
1055 res_lvb = res->lr_lvb_data;
1056 mb->mbo_dom_size = res_lvb->lvb_size;
1057 mb->mbo_dom_blocks = res_lvb->lvb_blocks;
1058 mb->mbo_mtime = res_lvb->lvb_mtime;
1059 mb->mbo_ctime = res_lvb->lvb_ctime;
1060 mb->mbo_atime = res_lvb->lvb_atime;
1062 CDEBUG(D_DLMTRACE, "size %llu\n", res_lvb->lvb_size);
1064 mb->mbo_valid |= OBD_MD_FLATIME | OBD_MD_FLCTIME | OBD_MD_FLMTIME |
1070 * MDT glimpse for Data-on-MDT
1072 * This function is called when MDT get attributes for the DoM object.
1073 * If there is write lock on client then function issues glimpse_ast to get
1074 * an actual size from that client.
1076 int mdt_dom_object_size(const struct lu_env *env, struct mdt_device *mdt,
1077 const struct lu_fid *fid, struct mdt_body *mb,
1080 struct ldlm_res_id resid;
1081 struct ldlm_resource *res;
1086 fid_build_reg_res_name(fid, &resid);
1087 res = ldlm_resource_get(mdt->mdt_namespace, NULL, &resid,
1092 /* Update lvbo data if DoM lock returned or if LVB is not yet valid. */
1093 if (dom_lock || !mdt_dom_lvb_is_valid(res))
1094 mdt_dom_lvbo_update(res, NULL, NULL, false);
1096 mdt_lvb2body(res, mb);
1097 ldlm_resource_putref(res);
1102 * MDT DoM lock intent policy (glimpse)
1104 * Intent policy is called when lock has an intent, for DoM file that
1105 * means glimpse lock and policy fills Lock Value Block (LVB).
1107 * If already granted lock is found it will be placed in \a lockp and
1108 * returned back to caller function.
1110 * \param[in] tsi session info
1111 * \param[in,out] lockp pointer to the lock
1112 * \param[in] flags LDLM flags
1114 * \retval ELDLM_LOCK_REPLACED if already granted lock was found
1115 * and placed in \a lockp
1116 * \retval ELDLM_LOCK_ABORTED in other cases except error
1117 * \retval negative value on error
1119 int mdt_glimpse_enqueue(struct mdt_thread_info *mti, struct ldlm_namespace *ns,
1120 struct ldlm_lock **lockp, __u64 flags)
1122 struct ldlm_lock *lock = *lockp;
1123 struct ldlm_resource *res = lock->l_resource;
1124 ldlm_processing_policy policy;
1125 struct ldlm_reply *rep;
1126 struct mdt_body *mbo;
1131 policy = ldlm_get_processing_policy(res);
1132 LASSERT(policy != NULL);
1134 req_capsule_set_size(mti->mti_pill, &RMF_MDT_MD, RCL_SERVER, 0);
1135 req_capsule_set_size(mti->mti_pill, &RMF_ACL, RCL_SERVER, 0);
1136 rc = req_capsule_server_pack(mti->mti_pill);
1138 RETURN(err_serious(rc));
1140 rep = req_capsule_server_get(mti->mti_pill, &RMF_DLM_REP);
1144 mbo = req_capsule_server_get(mti->mti_pill, &RMF_MDT_BODY);
1149 /* Check if this is a resend case (MSG_RESENT is set on RPC) and a
1150 * lock was found by ldlm_handle_enqueue(); if so no need to grant
1152 if (flags & LDLM_FL_RESENT) {
1153 rc = LDLM_ITER_CONTINUE;
1155 __u64 tmpflags = LDLM_FL_BLOCK_NOWAIT;
1156 enum ldlm_error err;
1158 rc = policy(lock, &tmpflags, LDLM_PROCESS_RESCAN, &err, NULL);
1159 check_res_locked(res);
1163 /* The lock met with no resistance; we're finished. */
1164 if (rc == LDLM_ITER_CONTINUE) {
1165 GOTO(fill_mbo, rc = ELDLM_LOCK_REPLACED);
1166 } else if (flags & LDLM_FL_BLOCK_NOWAIT) {
1167 /* LDLM_FL_BLOCK_NOWAIT means it is for AGL. Do not send glimpse
1168 * callback for glimpse size. The real size user will trigger
1169 * the glimpse callback when necessary. */
1170 GOTO(fill_mbo, rc = ELDLM_LOCK_ABORTED);
1173 rc = mdt_do_glimpse(mti->mti_env, ns, res);
1174 if (rc == -ENOENT) {
1175 /* We are racing with unlink(); just return -ENOENT */
1176 rep->lock_policy_res2 = ptlrpc_status_hton(-ENOENT);
1178 } else if (rc == -EINVAL) {
1179 /* this is possible is client lock has been cancelled but
1180 * still exists on server. If that lock was found on server
1181 * as only conflicting lock then the client has already
1182 * size authority and glimpse is not needed. */
1183 CDEBUG(D_DLMTRACE, "Glimpse from the client owning lock\n");
1185 } else if (rc < 0) {
1188 rc = ELDLM_LOCK_ABORTED;
1190 /* LVB can be without valid data in case of DOM */
1191 if (!mdt_dom_lvb_is_valid(res))
1192 mdt_dom_lvbo_update(res, lock, NULL, false);
1193 mdt_lvb2body(res, mbo);
1197 int mdt_brw_enqueue(struct mdt_thread_info *mti, struct ldlm_namespace *ns,
1198 struct ldlm_lock **lockp, __u64 flags)
1200 struct tgt_session_info *tsi = tgt_ses_info(mti->mti_env);
1201 struct lu_fid *fid = &tsi->tsi_fid;
1202 struct ldlm_lock *lock = *lockp;
1203 struct ldlm_resource *res = lock->l_resource;
1204 struct ldlm_reply *rep;
1205 struct mdt_body *mbo;
1206 struct mdt_lock_handle *lhc = &mti->mti_lh[MDT_LH_RMT];
1207 struct mdt_object *mo;
1212 /* Get lock from request for possible resent case. */
1213 mdt_intent_fixup_resent(mti, *lockp, lhc, flags);
1214 req_capsule_set_size(mti->mti_pill, &RMF_MDT_MD, RCL_SERVER, 0);
1215 req_capsule_set_size(mti->mti_pill, &RMF_ACL, RCL_SERVER, 0);
1216 rc = req_capsule_server_pack(mti->mti_pill);
1218 RETURN(err_serious(rc));
1220 rep = req_capsule_server_get(mti->mti_pill, &RMF_DLM_REP);
1224 mbo = req_capsule_server_get(mti->mti_pill, &RMF_MDT_BODY);
1228 fid_extract_from_res_name(fid, &res->lr_name);
1229 mo = mdt_object_find(mti->mti_env, mti->mti_mdt, fid);
1230 if (unlikely(IS_ERR(mo)))
1231 RETURN(PTR_ERR(mo));
1233 if (!mdt_object_exists(mo))
1234 GOTO(out, rc = -ENOENT);
1236 if (mdt_object_remote(mo))
1237 GOTO(out, rc = -EPROTO);
1240 if (!lustre_handle_is_used(&lhc->mlh_reg_lh)) {
1241 mdt_lock_handle_init(lhc);
1242 mdt_lock_reg_init(lhc, (*lockp)->l_req_mode);
1243 /* This will block MDT thread but it should be fine until
1244 * client caches small amount of data for DoM, which should be
1245 * smaller than one BRW RPC and should be able to be
1246 * piggybacked by lock cancel RPC.
1247 * If the client could hold the lock too long, this code can be
1248 * revised to call mdt_object_lock_try(). And if fails, it will
1249 * return ELDLM_OK here and fall back into normal lock enqueue
1252 rc = mdt_object_lock(mti, mo, lhc, MDS_INODELOCK_DOM);
1257 if (!mdt_dom_lvb_is_valid(res)) {
1258 rc = mdt_dom_lvb_alloc(res);
1261 mdt_dom_disk_lvbo_update(mti->mti_env, mo, res, false);
1263 mdt_lvb2body(res, mbo);
1265 rep->lock_policy_res2 = clear_serious(rc);
1266 if (rep->lock_policy_res2) {
1267 lhc->mlh_reg_lh.cookie = 0ull;
1268 GOTO(out, rc = ELDLM_LOCK_ABORTED);
1271 rc = mdt_intent_lock_replace(mti, lockp, lhc, flags, rc);
1273 mdt_object_put(mti->mti_env, mo);
1277 /* check if client has already DoM lock for given resource */
1278 bool mdt_dom_client_has_lock(struct mdt_thread_info *info,
1279 const struct lu_fid *fid)
1281 struct mdt_device *mdt = info->mti_mdt;
1282 union ldlm_policy_data *policy = &info->mti_policy;
1283 struct ldlm_res_id *res_id = &info->mti_res_id;
1284 struct lustre_handle lockh;
1285 enum ldlm_mode mode;
1286 struct ldlm_lock *lock;
1289 policy->l_inodebits.bits = MDS_INODELOCK_DOM;
1290 fid_build_reg_res_name(fid, res_id);
1292 mode = ldlm_lock_match(mdt->mdt_namespace, LDLM_FL_BLOCK_GRANTED |
1293 LDLM_FL_TEST_LOCK, res_id, LDLM_IBITS, policy,
1296 /* There is no other PW lock on this object; finished. */
1300 lock = ldlm_handle2lock(&lockh);
1304 /* check if lock from the same client */
1305 rc = (lock->l_export->exp_handle.h_cookie ==
1306 info->mti_exp->exp_handle.h_cookie);
1307 LDLM_LOCK_PUT(lock);
1312 * MDT request handler for OST_GETATTR RPC.
1314 * This is data-specific request to get object and layout versions under
1315 * IO lock. It is reliable only for Data-on-MDT files.
1317 * \param[in] tsi target session environment for this request
1319 * \retval 0 if successful
1320 * \retval negative value on error
1322 int mdt_data_version_get(struct tgt_session_info *tsi)
1324 struct mdt_thread_info *mti = mdt_th_info(tsi->tsi_env);
1325 struct mdt_device *mdt = mti->mti_mdt;
1326 struct mdt_body *repbody;
1327 struct mdt_object *mo = mti->mti_object;
1328 struct lov_comp_md_v1 *comp;
1329 struct lustre_handle lh = { 0 };
1332 enum ldlm_mode lock_mode = LCK_PR;
1338 req_capsule_set_size(tsi->tsi_pill, &RMF_MDT_MD, RCL_SERVER, 0);
1339 req_capsule_set_size(tsi->tsi_pill, &RMF_ACL, RCL_SERVER, 0);
1340 rc = req_capsule_server_pack(tsi->tsi_pill);
1341 if (unlikely(rc != 0))
1342 RETURN(err_serious(rc));
1344 repbody = req_capsule_server_get(tsi->tsi_pill, &RMF_MDT_BODY);
1345 if (repbody == NULL)
1348 srvlock = tsi->tsi_mdt_body->mbo_valid & OBD_MD_FLFLAGS &&
1349 tsi->tsi_mdt_body->mbo_flags & OBD_FL_SRVLOCK;
1352 if (unlikely(tsi->tsi_mdt_body->mbo_flags & OBD_FL_FLUSH))
1355 fid_build_reg_res_name(&tsi->tsi_fid, &tsi->tsi_resid);
1356 rc = tgt_mdt_data_lock(mdt->mdt_namespace, &tsi->tsi_resid,
1357 &lh, lock_mode, &flags);
1362 if (!mdt_object_exists(mo))
1363 GOTO(out, rc = -ENOENT);
1364 if (mdt_object_remote(mo))
1365 GOTO(out, rc = -EREMOTE);
1366 if (!S_ISREG(lu_object_attr(&mo->mot_obj)))
1367 GOTO(out, rc = -EBADF);
1369 /* Get version first */
1370 version = dt_version_get(tsi->tsi_env, mdt_obj2dt(mo));
1371 if (version && version != -EOPNOTSUPP) {
1372 repbody->mbo_valid |= OBD_MD_FLDATAVERSION;
1373 /* re-use mbo_ioepoch to transfer version */
1374 repbody->mbo_version = version;
1377 /* Read layout to get its version */
1378 rc = mdt_big_xattr_get(mti, mo, XATTR_NAME_LOV);
1379 if (rc == -ENODATA) /* File has no layout yet */
1384 comp = mti->mti_buf.lb_buf;
1385 if (le32_to_cpu(comp->lcm_magic) != LOV_MAGIC_COMP_V1) {
1386 CDEBUG(D_INFO, DFID" has no composite layout",
1387 PFID(&tsi->tsi_fid));
1388 GOTO(out, rc = -ESTALE);
1391 CDEBUG(D_INODE, DFID": layout version: %u\n",
1392 PFID(&tsi->tsi_fid), le32_to_cpu(comp->lcm_layout_gen));
1394 repbody->mbo_valid |= OBD_MD_LAYOUT_VERSION;
1395 /* re-use mbo_rdev for that */
1396 repbody->mbo_layout_gen = le32_to_cpu(comp->lcm_layout_gen);
1400 tgt_mdt_data_unlock(&lh, lock_mode);
1402 repbody->mbo_valid |= OBD_MD_FLFLAGS;
1403 repbody->mbo_flags = OBD_FL_FLUSH;
1407 /* read file data to the buffer */
1408 int mdt_dom_read_on_open(struct mdt_thread_info *mti, struct mdt_device *mdt,
1409 struct lustre_handle *lh)
1411 const struct lu_env *env = mti->mti_env;
1412 struct tgt_session_info *tsi = tgt_ses_info(env);
1413 struct req_capsule *pill = tsi->tsi_pill;
1414 const struct lu_fid *fid;
1415 struct ptlrpc_request *req = tgt_ses_req(tsi);
1416 struct mdt_body *mbo;
1417 struct dt_device *dt = mdt->mdt_bottom;
1418 struct dt_object *mo;
1420 struct niobuf_remote *rnb = NULL;
1421 struct niobuf_local *lnb;
1424 unsigned int len, copied = 0;
1425 int lnbs, nr_local, i;
1426 bool dom_lock = false;
1430 if (!req_capsule_field_present(pill, &RMF_NIOBUF_INLINE, RCL_SERVER)) {
1431 /* There is no reply buffers for this field, this means that
1432 * client has no support for data in reply.
1437 mbo = req_capsule_server_get(pill, &RMF_MDT_BODY);
1439 if (lustre_handle_is_used(lh)) {
1440 struct ldlm_lock *lock;
1442 lock = ldlm_handle2lock(lh);
1444 dom_lock = ldlm_has_dom(lock) && ldlm_has_layout(lock);
1445 LDLM_LOCK_PUT(lock);
1449 /* return data along with open only along with DoM lock */
1450 if (!dom_lock || !mdt->mdt_opts.mo_dom_read_open)
1453 if (!(mbo->mbo_valid & OBD_MD_DOM_SIZE))
1456 if (mbo->mbo_dom_size == 0)
1459 CDEBUG(D_INFO, "File size %llu, reply sizes %d/%d\n",
1460 mbo->mbo_dom_size, req->rq_reqmsg->lm_repsize, req->rq_replen);
1461 len = req->rq_reqmsg->lm_repsize - req->rq_replen;
1463 /* NB: at this moment we have the following sizes:
1464 * - req->rq_replen: used data in reply
1465 * - req->rq_reqmsg->lm_repsize: total allocated reply buffer at client
1467 * Ideal case when file size fits in allocated reply buffer,
1468 * that mean we can return whole data in reply. We can also fit more
1469 * data up to max_reply_size in total reply size, but this will cause
1470 * re-allocation on client and resend with larger buffer. This is still
1471 * faster than separate READ IO.
1472 * Third case if file is too big to fit even in maximum size, in that
1473 * case we return just tail to optimize possible append.
1475 * At the moment the following strategy is used:
1476 * 1) try to fit into the buffer we have
1477 * 2) return just file tail otherwise.
1479 if (mbo->mbo_dom_size <= len) {
1480 /* can fit whole data */
1481 len = mbo->mbo_dom_size;
1486 /* File tail offset must be aligned with larger page size
1487 * between client and server, so the maximum page size is
1488 * used here to align offset.
1490 * NB: DOM feature was introduced when server supports pagebits
1491 * already, so it should be always non-zero value. Report error
1492 * if it is not for some reason.
1494 if (!req->rq_export->exp_target_data.ted_pagebits) {
1495 CERROR("%s: client page bits are not saved on server\n",
1499 pgbits = max_t(int, PAGE_SHIFT,
1500 req->rq_export->exp_target_data.ted_pagebits);
1501 tail = mbo->mbo_dom_size % (1 << pgbits);
1503 /* no partial tail or tail can't fit in reply */
1504 if (tail == 0 || len < tail)
1508 offset = mbo->mbo_dom_size - len;
1510 LASSERT((offset % PAGE_SIZE) == 0);
1511 rc = req_capsule_server_grow(pill, &RMF_NIOBUF_INLINE,
1512 sizeof(*rnb) + len);
1514 /* failed to grow data buffer, just exit */
1515 GOTO(out, rc = -E2BIG);
1518 /* re-take MDT_BODY and NIOBUF_INLINE buffers after the buffer grow */
1519 mbo = req_capsule_server_get(pill, &RMF_MDT_BODY);
1520 fid = &mbo->mbo_fid1;
1521 if (!fid_is_sane(fid))
1522 GOTO(out, rc = -EINVAL);
1524 rnb = req_capsule_server_get(tsi->tsi_pill, &RMF_NIOBUF_INLINE);
1526 GOTO(out, rc = -EPROTO);
1528 buf = (char *)rnb + sizeof(*rnb);
1530 rnb->rnb_offset = offset;
1532 mo = dt_locate(env, dt, fid);
1534 GOTO(out_rnb, rc = PTR_ERR(mo));
1535 LASSERT(mo != NULL);
1537 dt_read_lock(env, mo, 0);
1538 if (!dt_object_exists(mo))
1539 GOTO(unlock, rc = -ENOENT);
1541 /* parse remote buffers to local buffers and prepare the latter */
1542 lnbs = (len >> PAGE_SHIFT) + 1;
1543 OBD_ALLOC(lnb, sizeof(*lnb) * lnbs);
1545 GOTO(unlock, rc = -ENOMEM);
1547 rc = dt_bufs_get(env, mo, rnb, lnb, 0);
1548 if (unlikely(rc < 0))
1550 LASSERT(rc <= lnbs);
1552 rc = dt_read_prep(env, mo, lnb, nr_local);
1555 /* copy data to the buffer finally */
1556 for (i = 0; i < nr_local; i++) {
1557 char *p = kmap(lnb[i].lnb_page);
1560 LASSERT(lnb[i].lnb_page_offset == 0);
1561 off = lnb[i].lnb_len & ~PAGE_MASK;
1563 memset(p + off, 0, PAGE_SIZE - off);
1565 memcpy(buf + (i << PAGE_SHIFT), p, lnb[i].lnb_len);
1566 kunmap(lnb[i].lnb_page);
1567 copied += lnb[i].lnb_len;
1570 CDEBUG(D_INFO, "Read %i (wanted %u) bytes from %llu\n", copied,
1573 CWARN("%s: read %i bytes for "DFID
1574 " but wanted %u, is size wrong?\n",
1575 tsi->tsi_exp->exp_obd->obd_name, copied,
1576 PFID(&tsi->tsi_fid), len);
1577 /* Ignore partially copied data */
1582 dt_bufs_put(env, mo, lnb, nr_local);
1584 OBD_FREE(lnb, sizeof(*lnb) * lnbs);
1586 dt_read_unlock(env, mo);
1587 lu_object_put(env, &mo->do_lu);
1589 rnb->rnb_len = copied;
1591 /* Don't fail OPEN request if read-on-open is failed, but drop
1592 * a message in log about the error.
1595 CDEBUG(D_INFO, "Read-on-open is failed, rc = %d", rc);
1601 * Completion AST for DOM discard locks:
1603 * CP AST an DOM discard lock is called always right after enqueue or from
1604 * reprocess if lock was blocked, in the latest case l_ast_data is set to
1605 * the mdt_object which is kept while there are pending locks on it.
1607 int ldlm_dom_discard_cp_ast(struct ldlm_lock *lock, __u64 flags, void *data)
1609 struct mdt_object *mo;
1610 struct lustre_handle dom_lh;
1615 /* l_ast_data is set when lock was not granted immediately
1616 * in mdt_dom_discard_data() below but put into waiting list,
1617 * so this CP callback means we are finished and corresponding
1618 * MDT object should be released finally as well as lock itself.
1620 lock_res_and_lock(lock);
1621 if (!lock->l_ast_data) {
1622 unlock_res_and_lock(lock);
1626 mo = lock->l_ast_data;
1627 lock->l_ast_data = NULL;
1628 unlock_res_and_lock(lock);
1630 ldlm_lock2handle(lock, &dom_lh);
1631 ldlm_lock_decref(&dom_lh, LCK_PW);
1633 env = lu_env_find();
1635 mdt_object_put(env, mo);
1640 void mdt_dom_discard_data(struct mdt_thread_info *info,
1641 struct mdt_object *mo)
1643 struct ptlrpc_request *req = mdt_info_req(info);
1644 struct mdt_device *mdt = mdt_dev(mo->mot_obj.lo_dev);
1645 union ldlm_policy_data policy;
1646 struct ldlm_res_id res_id;
1647 struct lustre_handle dom_lh;
1648 struct ldlm_lock *lock;
1649 __u64 flags = LDLM_FL_AST_DISCARD_DATA;
1655 if (req && req_is_replay(req))
1658 policy.l_inodebits.bits = MDS_INODELOCK_DOM;
1659 policy.l_inodebits.try_bits = 0;
1660 fid_build_reg_res_name(mdt_object_fid(mo), &res_id);
1662 /* Keep blocking version of discard for an old client to avoid
1663 * crashes on non-patched clients. LU-11359.
1665 old_client = req && !(exp_connect_flags2(req->rq_export) &
1666 OBD_CONNECT2_ASYNC_DISCARD);
1668 /* Tell the clients that the object is gone now and that they should
1669 * throw away any cached pages. */
1670 rc = ldlm_cli_enqueue_local(info->mti_env, mdt->mdt_namespace, &res_id,
1671 LDLM_IBITS, &policy, LCK_PW, &flags,
1672 ldlm_blocking_ast, old_client ?
1673 ldlm_completion_ast :
1674 ldlm_dom_discard_cp_ast,
1675 NULL, NULL, 0, LVB_T_NONE, NULL, &dom_lh);
1676 if (rc != ELDLM_OK) {
1678 "Failed to issue discard lock, rc = %d\n", rc);
1682 lock = ldlm_handle2lock(&dom_lh);
1683 lock_res_and_lock(lock);
1684 /* if lock is not granted then there are BL ASTs in progress and
1685 * lock will be granted in result of reprocessing with CP callback
1686 * notifying about that. The mdt object has to be kept until that and
1687 * it is saved in l_ast_data of the lock. Lock reference is kept too
1688 * until that to prevent it from canceling.
1690 if (!is_granted_or_cancelled_nolock(lock)) {
1691 mdt_object_get(info->mti_env, mo);
1692 lock->l_ast_data = mo;
1693 unlock_res_and_lock(lock);
1695 unlock_res_and_lock(lock);
1696 ldlm_lock_decref_and_cancel(&dom_lh, LCK_PW);
1698 LDLM_LOCK_PUT(lock);