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.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2009, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2012, 2013, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
38 * Author: Alex Tomas <bzzz@whamcloud.com>
41 #define DEBUG_SUBSYSTEM S_FILTER
43 #include "ofd_internal.h"
45 static int ofd_preprw_read(const struct lu_env *env, struct obd_export *exp,
46 struct ofd_device *ofd, const struct lu_fid *fid,
47 struct lu_attr *la, int niocount,
48 struct niobuf_remote *rnb, int *nr_local,
49 struct niobuf_local *lnb, char *jobid)
51 struct ofd_object *fo;
52 int i, j, rc, tot_bytes = 0;
57 fo = ofd_object_find(env, ofd, fid);
62 ofd_read_lock(env, fo);
63 if (!ofd_object_exists(fo))
64 GOTO(unlock, rc = -ENOENT);
66 /* parse remote buffers to local buffers and prepare the latter */
68 for (i = 0, j = 0; i < niocount; i++) {
69 rc = dt_bufs_get(env, ofd_object_child(fo), rnb + i,
70 lnb + j, 0, ofd_object_capa(env, fo));
73 LASSERT(rc <= PTLRPC_MAX_BRW_PAGES);
74 /* correct index for local buffers to continue with */
77 LASSERT(j <= PTLRPC_MAX_BRW_PAGES);
78 tot_bytes += rnb[i].rnb_len;
81 LASSERT(*nr_local > 0 && *nr_local <= PTLRPC_MAX_BRW_PAGES);
82 rc = dt_attr_get(env, ofd_object_child(fo), la,
83 ofd_object_capa(env, fo));
87 rc = dt_read_prep(env, ofd_object_child(fo), lnb, *nr_local);
91 ofd_counter_incr(exp, LPROC_OFD_STATS_READ, jobid, tot_bytes);
95 dt_bufs_put(env, ofd_object_child(fo), lnb, *nr_local);
97 ofd_read_unlock(env, fo);
98 ofd_object_put(env, fo);
102 static int ofd_preprw_write(const struct lu_env *env, struct obd_export *exp,
103 struct ofd_device *ofd, const struct lu_fid *fid,
104 struct lu_attr *la, struct obdo *oa,
105 int objcount, struct obd_ioobj *obj,
106 struct niobuf_remote *rnb, int *nr_local,
107 struct niobuf_local *lnb, char *jobid)
109 struct ofd_object *fo;
110 int i, j, k, rc = 0, tot_bytes = 0;
113 LASSERT(env != NULL);
114 LASSERT(objcount == 1);
116 if (unlikely(exp->exp_obd->obd_recovering)) {
117 struct ofd_thread_info *info = ofd_info(env);
119 /* copied from ofd_precreate_object */
120 /* XXX this should be consolidated to use the same code
121 * instead of a copy, due to the ongoing risk of bugs. */
122 memset(&info->fti_attr, 0, sizeof(info->fti_attr));
123 info->fti_attr.la_valid = LA_TYPE | LA_MODE;
124 info->fti_attr.la_mode = S_IFREG | S_ISUID | S_ISGID | 0666;
125 info->fti_attr.la_valid |= LA_ATIME | LA_MTIME | LA_CTIME;
126 /* Initialize a/c/m time so any client timestamp will always
127 * be newer and update the inode. ctime = 0 is also handled
128 * specially in osd_inode_setattr(). See LU-221, LU-1042 */
129 info->fti_attr.la_atime = 0;
130 info->fti_attr.la_mtime = 0;
131 info->fti_attr.la_ctime = 0;
133 fo = ofd_object_find_or_create(env, ofd, fid, &info->fti_attr);
135 fo = ofd_object_find(env, ofd, fid);
139 GOTO(out, rc = PTR_ERR(fo));
142 ofd_read_lock(env, fo);
143 if (!ofd_object_exists(fo)) {
144 CERROR("%s: BRW to missing obj "DOSTID"\n",
145 exp->exp_obd->obd_name, POSTID(&obj->ioo_oid));
146 ofd_read_unlock(env, fo);
147 ofd_object_put(env, fo);
148 GOTO(out, rc = -ENOENT);
151 /* Process incoming grant info, set OBD_BRW_GRANTED flag and grant some
152 * space back if possible */
153 ofd_grant_prepare_write(env, exp, oa, rnb, obj->ioo_bufcnt);
155 /* parse remote buffers to local buffers and prepare the latter */
157 for (i = 0, j = 0; i < obj->ioo_bufcnt; i++) {
158 rc = dt_bufs_get(env, ofd_object_child(fo),
160 ofd_object_capa(env, fo));
161 if (unlikely(rc < 0))
163 LASSERT(rc <= PTLRPC_MAX_BRW_PAGES);
164 /* correct index for local buffers to continue with */
165 for (k = 0; k < rc; k++) {
166 lnb[j+k].lnb_flags = rnb[i].rnb_flags;
167 if (!(rnb[i].rnb_flags & OBD_BRW_GRANTED))
168 lnb[j+k].lnb_rc = -ENOSPC;
170 /* remote client can't break through quota */
171 if (exp_connect_rmtclient(exp))
172 lnb[j+k].lnb_flags &= ~OBD_BRW_NOQUOTA;
176 LASSERT(j <= PTLRPC_MAX_BRW_PAGES);
177 tot_bytes += rnb[i].rnb_len;
179 LASSERT(*nr_local > 0 && *nr_local <= PTLRPC_MAX_BRW_PAGES);
181 rc = dt_write_prep(env, ofd_object_child(fo), lnb, *nr_local);
182 if (unlikely(rc != 0))
185 ofd_counter_incr(exp, LPROC_OFD_STATS_WRITE, jobid, tot_bytes);
188 dt_bufs_put(env, ofd_object_child(fo), lnb, *nr_local);
189 ofd_read_unlock(env, fo);
190 /* ofd_grant_prepare_write() was called, so we must commit */
191 ofd_grant_commit(env, exp, rc);
193 /* let's still process incoming grant information packed in the oa,
194 * but without enforcing grant since we won't proceed with the write.
195 * Just like a read request actually. */
196 ofd_grant_prepare_read(env, exp, oa);
200 int ofd_preprw(const struct lu_env *env, int cmd, struct obd_export *exp,
201 struct obdo *oa, int objcount, struct obd_ioobj *obj,
202 struct niobuf_remote *rnb, int *nr_local,
203 struct niobuf_local *lnb, struct obd_trans_info *oti,
204 struct lustre_capa *capa)
206 struct tgt_session_info *tsi = tgt_ses_info(env);
207 struct ofd_device *ofd = ofd_exp(exp);
208 struct ofd_thread_info *info;
210 const struct lu_fid *fid = &oa->o_oi.oi_fid;
213 if (*nr_local > PTLRPC_MAX_BRW_PAGES) {
214 CERROR("%s: bulk has too many pages %d, which exceeds the"
215 "maximum pages per RPC of %d\n",
216 exp->exp_obd->obd_name, *nr_local, PTLRPC_MAX_BRW_PAGES);
220 if (tgt_ses_req(tsi) == NULL) { /* echo client case */
221 LASSERT(oti != NULL);
222 info = ofd_info_init(env, exp);
223 ofd_oti2info(info, oti);
224 jobid = oti->oti_jobid;
226 info = tsi2ofd_info(tsi);
227 jobid = tsi->tsi_jobid;
232 if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOENT)) {
233 struct ofd_seq *oseq;
235 oseq = ofd_seq_load(env, ofd, ostid_seq(&oa->o_oi));
237 CERROR("%s: Can not find seq for "DOSTID
238 ": rc = %ld\n", ofd_name(ofd), POSTID(&oa->o_oi),
243 if (oseq->os_destroys_in_progress == 0) {
244 /* don't fail lookups for orphan recovery, it causes
245 * later LBUGs when objects still exist during
247 ofd_seq_put(env, oseq);
250 ofd_seq_put(env, oseq);
253 LASSERT(objcount == 1);
254 LASSERT(obj->ioo_bufcnt > 0);
256 if (cmd == OBD_BRW_WRITE) {
257 rc = ofd_auth_capa(exp, fid, ostid_seq(&oa->o_oi),
258 capa, CAPA_OPC_OSS_WRITE);
260 la_from_obdo(&info->fti_attr, oa, OBD_MD_FLGETATTR);
261 rc = ofd_preprw_write(env, exp, ofd, fid,
262 &info->fti_attr, oa, objcount,
263 obj, rnb, nr_local, lnb, jobid);
265 } else if (cmd == OBD_BRW_READ) {
266 rc = ofd_auth_capa(exp, fid, ostid_seq(&oa->o_oi),
267 capa, CAPA_OPC_OSS_READ);
269 ofd_grant_prepare_read(env, exp, oa);
270 rc = ofd_preprw_read(env, exp, ofd, fid,
271 &info->fti_attr, obj->ioo_bufcnt,
272 rnb, nr_local, lnb, jobid);
273 obdo_from_la(oa, &info->fti_attr, LA_ATIME);
276 CERROR("%s: wrong cmd %d received!\n",
277 exp->exp_obd->obd_name, cmd);
284 ofd_commitrw_read(const struct lu_env *env, struct ofd_device *ofd,
285 const struct lu_fid *fid, int objcount, int niocount,
286 struct niobuf_local *lnb)
288 struct ofd_object *fo;
292 LASSERT(niocount > 0);
294 fo = ofd_object_find(env, ofd, fid);
298 LASSERT(ofd_object_exists(fo));
299 dt_bufs_put(env, ofd_object_child(fo), lnb, niocount);
301 ofd_read_unlock(env, fo);
302 ofd_object_put(env, fo);
303 /* second put is pair to object_get in ofd_preprw_read */
304 ofd_object_put(env, fo);
310 ofd_write_attr_set(const struct lu_env *env, struct ofd_device *ofd,
311 struct ofd_object *ofd_obj, struct lu_attr *la,
312 struct filter_fid *ff)
314 struct ofd_thread_info *info = ofd_info(env);
315 __u64 valid = la->la_valid;
318 struct dt_object *dt_obj;
325 dt_obj = ofd_object_child(ofd_obj);
326 LASSERT(dt_obj != NULL);
328 la->la_valid &= LA_UID | LA_GID;
330 rc = ofd_attr_handle_ugid(env, ofd_obj, la, 0 /* !is_setattr */);
335 rc = ofd_object_ff_check(env, ofd_obj);
342 if (!la->la_valid && !ff_needed)
343 /* no attributes to set */
346 th = ofd_trans_create(env, ofd);
348 GOTO(out, rc = PTR_ERR(th));
351 rc = dt_declare_attr_set(env, dt_obj, la, th);
357 info->fti_buf.lb_buf = ff;
358 info->fti_buf.lb_len = sizeof(*ff);
359 rc = dt_declare_xattr_set(env, dt_obj, &info->fti_buf,
360 XATTR_NAME_FID, 0, th);
365 /* We don't need a transno for this operation which will be re-executed
366 * anyway when the OST_WRITE (with a transno assigned) is replayed */
367 rc = dt_trans_start_local(env, ofd->ofd_osd , th);
373 rc = dt_attr_set(env, dt_obj, la, th,
374 ofd_object_capa(env, ofd_obj));
379 /* set filter fid EA */
381 rc = dt_xattr_set(env, dt_obj, &info->fti_buf, XATTR_NAME_FID,
389 dt_trans_stop(env, ofd->ofd_osd, th);
391 la->la_valid = valid;
395 struct ofd_soft_sync_callback {
396 struct dt_txn_commit_cb ossc_cb;
397 struct obd_export *ossc_exp;
400 static void ofd_cb_soft_sync(struct lu_env *env, struct thandle *th,
401 struct dt_txn_commit_cb *cb, int err)
403 struct ofd_soft_sync_callback *ossc;
405 ossc = container_of(cb, struct ofd_soft_sync_callback, ossc_cb);
407 CDEBUG(D_INODE, "export %p soft sync count is reset\n", ossc->ossc_exp);
408 atomic_set(&ossc->ossc_exp->exp_filter_data.fed_soft_sync_count, 0);
410 class_export_cb_put(ossc->ossc_exp);
414 static int ofd_soft_sync_cb_add(struct thandle *th, struct obd_export *exp)
416 struct ofd_soft_sync_callback *ossc;
417 struct dt_txn_commit_cb *dcb;
424 ossc->ossc_exp = class_export_cb_get(exp);
426 dcb = &ossc->ossc_cb;
427 dcb->dcb_func = ofd_cb_soft_sync;
428 CFS_INIT_LIST_HEAD(&dcb->dcb_linkage);
429 strncpy(dcb->dcb_name, "ofd_cb_soft_sync", MAX_COMMIT_CB_STR_LEN);
430 dcb->dcb_name[MAX_COMMIT_CB_STR_LEN - 1] = '\0';
432 rc = dt_trans_cb_add(th, dcb);
434 class_export_cb_put(exp);
442 ofd_commitrw_write(const struct lu_env *env, struct obd_export *exp,
443 struct ofd_device *ofd, const struct lu_fid *fid,
444 struct lu_attr *la, struct filter_fid *ff, int objcount,
445 int niocount, struct niobuf_local *lnb, int old_rc)
447 struct ofd_thread_info *info = ofd_info(env);
448 struct ofd_object *fo;
454 struct filter_export_data *fed = &exp->exp_filter_data;
455 bool soft_sync = false;
456 bool cb_registered = false;
460 LASSERT(objcount == 1);
462 fo = ofd_object_find(env, ofd, fid);
464 LASSERT(ofd_object_exists(fo));
466 o = ofd_object_child(fo);
470 GOTO(out, rc = old_rc);
473 * The first write to each object must set some attributes. It is
474 * important to set the uid/gid before calling
475 * dt_declare_write_commit() since quota enforcement is now handled in
478 rc = ofd_write_attr_set(env, ofd, fo, la, ff);
482 la->la_valid &= LA_ATIME | LA_MTIME | LA_CTIME;
485 th = ofd_trans_create(env, ofd);
487 GOTO(out, rc = PTR_ERR(th));
489 th->th_sync |= ofd->ofd_syncjournal;
490 if (th->th_sync == 0) {
491 for (i = 0; i < niocount; i++) {
492 if (!(lnb[i].lnb_flags & OBD_BRW_ASYNC)) {
496 if (lnb[i].lnb_flags & OBD_BRW_SOFT_SYNC)
501 if (OBD_FAIL_CHECK(OBD_FAIL_OST_DQACQ_NET))
502 GOTO(out_stop, rc = -EINPROGRESS);
504 rc = dt_declare_write_commit(env, o, lnb, niocount, th);
509 /* update [mac]time if needed */
510 rc = dt_declare_attr_set(env, o, la, th);
515 rc = ofd_trans_start(env, ofd, fo, th);
519 rc = dt_write_commit(env, o, lnb, niocount, th);
524 rc = dt_attr_set(env, o, la, th, ofd_object_capa(env, fo));
529 /* get attr to return */
530 rc = dt_attr_get(env, o, la, ofd_object_capa(env, fo));
533 /* Force commit to make the just-deleted blocks
534 * reusable. LU-456 */
538 /* do this before trans stop in case commit has finished */
539 if (!th->th_sync && soft_sync && !cb_registered) {
540 ofd_soft_sync_cb_add(th, exp);
541 cb_registered = true;
544 ofd_trans_stop(env, ofd, th, rc);
545 if (rc == -ENOSPC && retries++ < 3) {
546 CDEBUG(D_INODE, "retry after force commit, retries:%d\n",
552 /* reset fed_soft_sync_count upon non-SOFT_SYNC RPC */
553 atomic_set(&fed->fed_soft_sync_count, 0);
554 else if (atomic_inc_return(&fed->fed_soft_sync_count) ==
555 ofd->ofd_soft_sync_limit)
556 dt_commit_async(env, ofd->ofd_osd);
559 dt_bufs_put(env, o, lnb, niocount);
560 ofd_read_unlock(env, fo);
561 ofd_object_put(env, fo);
562 /* second put is pair to object_get in ofd_preprw_write */
563 ofd_object_put(env, fo);
564 ofd_grant_commit(env, info->fti_exp, old_rc);
568 int ofd_commitrw(const struct lu_env *env, int cmd, struct obd_export *exp,
569 struct obdo *oa, int objcount, struct obd_ioobj *obj,
570 struct niobuf_remote *rnb, int npages,
571 struct niobuf_local *lnb, struct obd_trans_info *oti,
574 struct ofd_thread_info *info = ofd_info(env);
575 struct ofd_mod_data *fmd;
577 struct ofd_device *ofd = ofd_exp(exp);
578 struct filter_fid *ff = NULL;
579 const struct lu_fid *fid = &oa->o_oi.oi_fid;
584 if (cmd == OBD_BRW_WRITE) {
585 /* Don't update timestamps if this write is older than a
586 * setattr which modifies the timestamps. b=10150 */
588 /* XXX when we start having persistent reservations this needs
589 * to be changed to ofd_fmd_get() to create the fmd if it
590 * doesn't already exist so we can store the reservation handle
592 valid = OBD_MD_FLUID | OBD_MD_FLGID;
593 fmd = ofd_fmd_find(exp, fid);
594 if (!fmd || fmd->fmd_mactime_xid < info->fti_xid)
595 valid |= OBD_MD_FLATIME | OBD_MD_FLMTIME |
597 ofd_fmd_put(exp, fmd);
598 la_from_obdo(&info->fti_attr, oa, valid);
600 if (oa->o_valid & OBD_MD_FLFID) {
601 ff = &info->fti_mds_fid;
602 ofd_prepare_fidea(ff, oa);
605 rc = ofd_commitrw_write(env, exp, ofd, fid, &info->fti_attr,
606 ff, objcount, npages, lnb, old_rc);
608 obdo_from_la(oa, &info->fti_attr,
609 OFD_VALID_FLAGS | LA_GID | LA_UID);
611 obdo_from_la(oa, &info->fti_attr, LA_GID | LA_UID);
613 /* don't report overquota flag if we failed before reaching
615 if (old_rc == 0 && (rc == 0 || rc == -EDQUOT)) {
616 /* return the overquota flags to client */
617 if (lnb[0].lnb_flags & OBD_BRW_OVER_USRQUOTA) {
618 if (oa->o_valid & OBD_MD_FLFLAGS)
619 oa->o_flags |= OBD_FL_NO_USRQUOTA;
621 oa->o_flags = OBD_FL_NO_USRQUOTA;
624 if (lnb[0].lnb_flags & OBD_BRW_OVER_GRPQUOTA) {
625 if (oa->o_valid & OBD_MD_FLFLAGS)
626 oa->o_flags |= OBD_FL_NO_GRPQUOTA;
628 oa->o_flags = OBD_FL_NO_GRPQUOTA;
631 oa->o_valid |= OBD_MD_FLFLAGS;
632 oa->o_valid |= OBD_MD_FLUSRQUOTA | OBD_MD_FLGRPQUOTA;
634 } else if (cmd == OBD_BRW_READ) {
635 struct ldlm_namespace *ns = ofd->ofd_namespace;
637 /* If oa != NULL then ofd_preprw_read updated the inode
638 * atime and we should update the lvb so that other glimpses
639 * will also get the updated value. bug 5972 */
640 if (oa && ns && ns->ns_lvbo && ns->ns_lvbo->lvbo_update) {
641 struct ldlm_resource *rs = NULL;
643 ost_fid_build_resid(fid, &info->fti_resid);
644 rs = ldlm_resource_get(ns, NULL, &info->fti_resid,
647 ns->ns_lvbo->lvbo_update(rs, NULL, 1);
648 ldlm_resource_putref(rs);
651 rc = ofd_commitrw_read(env, ofd, fid, objcount,
661 ofd_info2oti(info, oti);