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) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2011, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * lustre/mdd/mdd_device.c
34 * Lustre Metadata Server (mdd) routines
36 * Author: Wang Di <wangdi@clusterfs.com>
39 #define DEBUG_SUBSYSTEM S_MDS
41 #include <linux/module.h>
42 #include <linux/kthread.h>
43 #include <obd_class.h>
44 #include <uapi/linux/lustre/lustre_ioctl.h>
45 #include <lustre_mds.h>
46 #include <obd_support.h>
47 #include <lu_object.h>
48 #include <uapi/linux/lustre/lustre_param.h>
49 #include <lustre_fid.h>
50 #include <lustre_nodemap.h>
51 #include <lustre_barrier.h>
53 #include "mdd_internal.h"
55 static const struct md_device_operations mdd_ops;
56 static struct lu_device_type mdd_device_type;
58 static const char mdd_root_dir_name[] = "ROOT";
59 static const char mdd_obf_dir_name[] = "fid";
60 static const char mdd_lpf_dir_name[] = "lost+found";
62 /* Slab for MDD object allocation */
63 struct kmem_cache *mdd_object_kmem;
65 static struct lu_kmem_descr mdd_caches[] = {
67 .ckd_cache = &mdd_object_kmem,
68 .ckd_name = "mdd_obj",
69 .ckd_size = sizeof(struct mdd_object)
76 static int mdd_connect_to_next(const struct lu_env *env, struct mdd_device *m,
79 struct obd_connect_data *data = NULL;
80 struct lu_device *lud = mdd2lu_dev(m);
81 struct obd_device *obd;
85 LASSERT(m->mdd_child_exp == NULL);
87 OBD_ALLOC(data, sizeof(*data));
89 GOTO(out, rc = -ENOMEM);
91 obd = class_name2obd(nextdev);
93 CERROR("can't locate next device: %s\n", nextdev);
94 GOTO(out, rc = -ENOTCONN);
97 data->ocd_connect_flags = OBD_CONNECT_VERSION;
98 data->ocd_version = LUSTRE_VERSION_CODE;
100 rc = obd_connect(NULL, &m->mdd_child_exp, obd, &obd->obd_uuid, data, NULL);
102 CERROR("cannot connect to next dev %s (%d)\n", nextdev, rc);
106 lud->ld_site = m->mdd_child_exp->exp_obd->obd_lu_dev->ld_site;
107 LASSERT(lud->ld_site);
108 m->mdd_child = lu2dt_dev(m->mdd_child_exp->exp_obd->obd_lu_dev);
109 m->mdd_bottom = lu2dt_dev(lud->ld_site->ls_bottom_dev);
110 lu_dev_add_linkage(lud->ld_site, lud);
114 OBD_FREE(data, sizeof(*data));
118 static int mdd_init0(const struct lu_env *env, struct mdd_device *mdd,
119 struct lu_device_type *t, struct lustre_cfg *lcfg)
125 /* LU-8040 Set defaults here, before values configs */
126 mdd->mdd_cl.mc_flags = 0; /* off by default */
127 mdd->mdd_cl.mc_mask = CHANGELOG_DEFMASK;
128 mdd->mdd_cl.mc_deniednext = 60; /* 60 secs by default */
130 dev = lustre_cfg_string(lcfg, 0);
134 mdd->mdd_md_dev.md_lu_dev.ld_obd = class_name2obd(dev);
135 if (mdd->mdd_md_dev.md_lu_dev.ld_obd == NULL)
137 mdd->mdd_md_dev.md_lu_dev.ld_ops = &mdd_lu_ops;
138 mdd->mdd_md_dev.md_ops = &mdd_ops;
140 rc = mdd_connect_to_next(env, mdd, lustre_cfg_string(lcfg, 3));
144 mdd->mdd_atime_diff = MAX_ATIME_DIFF;
145 /* sync permission changes */
146 mdd->mdd_sync_permission = 1;
147 /* enable changelog garbage collection */
148 mdd->mdd_changelog_gc = 1;
149 /* with a significant amount of idle time */
150 mdd->mdd_changelog_max_idle_time = CHLOG_MAX_IDLE_TIME;
151 /* or a significant amount of late indexes */
152 mdd->mdd_changelog_max_idle_indexes = CHLOG_MAX_IDLE_INDEXES;
153 /* with a reasonable interval between each check */
154 mdd->mdd_changelog_min_gc_interval = CHLOG_MIN_GC_INTERVAL;
155 /* with a very few number of free catalog entries */
156 mdd->mdd_changelog_min_free_cat_entries = CHLOG_MIN_FREE_CAT_ENTRIES;
157 /* special default striping for files created with O_APPEND */
158 mdd->mdd_append_stripe_count = 1;
159 mdd->mdd_append_pool[0] = '\0';
161 dt_conf_get(env, mdd->mdd_child, &mdd->mdd_dt_conf);
163 /* we are using service name but not mdd obd name
164 * for compatibility reasons.
165 * It is passed from MDT in lustre_cfg[2] buffer */
166 rc = mdd_procfs_init(mdd, lustre_cfg_string(lcfg, 2));
168 obd_disconnect(mdd->mdd_child_exp);
173 static struct lu_device *mdd_device_fini(const struct lu_env *env,
176 struct mdd_device *mdd = lu2mdd_dev(d);
179 lu_dev_del_linkage(d->ld_site, d);
181 mdd_procfs_fini(mdd);
185 static int changelog_init_cb(const struct lu_env *env, struct llog_handle *llh,
186 struct llog_rec_hdr *hdr, void *data)
188 struct mdd_device *mdd = (struct mdd_device *)data;
189 struct llog_changelog_rec *rec = (struct llog_changelog_rec *)hdr;
191 LASSERT(llh->lgh_hdr->llh_flags & LLOG_F_IS_PLAIN);
192 LASSERT(rec->cr_hdr.lrh_type == CHANGELOG_REC);
195 "seeing record at index %d/%d/%llu t=%x %.*s in log"
196 DFID"\n", hdr->lrh_index, rec->cr_hdr.lrh_index,
197 rec->cr.cr_index, rec->cr.cr_type, rec->cr.cr_namelen,
198 changelog_rec_name(&rec->cr), PFID(&llh->lgh_id.lgl_oi.oi_fid));
200 mdd->mdd_cl.mc_index = rec->cr.cr_index;
201 return LLOG_PROC_BREAK;
204 static int changelog_user_init_cb(const struct lu_env *env,
205 struct llog_handle *llh,
206 struct llog_rec_hdr *hdr, void *data)
208 struct mdd_device *mdd = (struct mdd_device *)data;
209 struct llog_changelog_user_rec *rec =
210 (struct llog_changelog_user_rec *)hdr;
212 LASSERT(llh->lgh_hdr->llh_flags & LLOG_F_IS_PLAIN);
213 LASSERT(rec->cur_hdr.lrh_type == CHANGELOG_USER_REC);
215 CDEBUG(D_INFO, "seeing user at index %d/%d id=%d endrec=%llu"
216 " in log "DFID"\n", hdr->lrh_index, rec->cur_hdr.lrh_index,
217 rec->cur_id, rec->cur_endrec, PFID(&llh->lgh_id.lgl_oi.oi_fid));
219 spin_lock(&mdd->mdd_cl.mc_user_lock);
220 mdd->mdd_cl.mc_lastuser = rec->cur_id;
221 mdd->mdd_cl.mc_users++;
222 if (rec->cur_endrec > mdd->mdd_cl.mc_index)
223 mdd->mdd_cl.mc_index = rec->cur_endrec;
224 spin_unlock(&mdd->mdd_cl.mc_user_lock);
226 return LLOG_PROC_BREAK;
229 struct changelog_orphan_data {
231 struct mdd_device *mdd;
234 /* find oldest changelog record index */
235 static int changelog_detect_orphan_cb(const struct lu_env *env,
236 struct llog_handle *llh,
237 struct llog_rec_hdr *hdr, void *data)
239 struct mdd_device *mdd = ((struct changelog_orphan_data *)data)->mdd;
240 struct llog_changelog_rec *rec = container_of(hdr,
241 struct llog_changelog_rec,
244 LASSERT(llh->lgh_hdr->llh_flags & LLOG_F_IS_PLAIN);
246 if (rec->cr_hdr.lrh_type != CHANGELOG_REC) {
247 CWARN("%s: invalid record at index %d in log "DFID"\n",
248 mdd2obd_dev(mdd)->obd_name, hdr->lrh_index,
249 PFID(&llh->lgh_id.lgl_oi.oi_fid));
250 /* try to find some next valid record and thus allow to recover
251 * from a corrupted LLOG, instead to assert and force a crash
256 CDEBUG(D_INFO, "%s: seeing record at index %d/%d/%llu t=%x %.*s in log "
257 DFID"\n", mdd2obd_dev(mdd)->obd_name, hdr->lrh_index,
258 rec->cr_hdr.lrh_index, rec->cr.cr_index, rec->cr.cr_type,
259 rec->cr.cr_namelen, changelog_rec_name(&rec->cr),
260 PFID(&llh->lgh_id.lgl_oi.oi_fid));
262 ((struct changelog_orphan_data *)data)->index = rec->cr.cr_index;
263 return LLOG_PROC_BREAK;
266 /* find oldest changelog user index */
267 static int changelog_user_detect_orphan_cb(const struct lu_env *env,
268 struct llog_handle *llh,
269 struct llog_rec_hdr *hdr, void *data)
271 struct mdd_device *mdd = ((struct changelog_orphan_data *)data)->mdd;
272 struct llog_changelog_user_rec *rec = container_of(hdr,
273 struct llog_changelog_user_rec,
276 LASSERT(llh->lgh_hdr->llh_flags & LLOG_F_IS_PLAIN);
278 if (rec->cur_hdr.lrh_type != CHANGELOG_USER_REC) {
279 CWARN("%s: invalid user at index %d in log "DFID"\n",
280 mdd2obd_dev(mdd)->obd_name, hdr->lrh_index,
281 PFID(&llh->lgh_id.lgl_oi.oi_fid));
282 /* try to find some next valid record and thus allow to recover
283 * from a corrupted LLOG, instead to assert and force a crash
288 CDEBUG(D_INFO, "%s: seeing user at index %d/%d id=%d endrec=%llu in "
289 "log "DFID"\n", mdd2obd_dev(mdd)->obd_name, hdr->lrh_index,
290 rec->cur_hdr.lrh_index, rec->cur_id, rec->cur_endrec,
291 PFID(&llh->lgh_id.lgl_oi.oi_fid));
293 if (((struct changelog_orphan_data *)data)->index == 0 ||
294 rec->cur_endrec < ((struct changelog_orphan_data *)data)->index)
295 ((struct changelog_orphan_data *)data)->index = rec->cur_endrec;
300 struct changelog_cancel_cookie {
302 struct mdd_device *mdd;
305 static int llog_changelog_cancel_cb(const struct lu_env *env,
306 struct llog_handle *llh,
307 struct llog_rec_hdr *hdr, void *data)
309 struct llog_changelog_rec *rec = (struct llog_changelog_rec *)hdr;
310 struct llog_cookie cookie;
311 struct changelog_cancel_cookie *cl_cookie =
312 (struct changelog_cancel_cookie *)data;
317 /* This is always a (sub)log, not the catalog */
318 LASSERT(llh->lgh_hdr->llh_flags & LLOG_F_IS_PLAIN);
320 /* if current context is GC-thread allow it to stop upon umount
321 * remaining records cleanup will occur upon next mount
323 * also during testing, wait for GC-thread to be released
325 * XXX this requires the GC-thread to not fork a sub-thread via
326 * llog[_cat]_process_or_fork() and we may think to also implement
327 * this shutdown mechanism for manually started user unregister which
328 * can also take a long time if huge backlog of records
330 if (unlikely(cl_cookie->mdd->mdd_cl.mc_gc_task == current)) {
331 /* wait to be released */
332 while (CFS_FAIL_CHECK_QUIET(OBD_FAIL_FORCE_GC_THREAD))
335 if (kthread_should_stop())
336 RETURN(LLOG_PROC_BREAK);
339 if (rec->cr.cr_index > cl_cookie->endrec)
340 /* records are in order, so we're done */
341 RETURN(LLOG_PROC_BREAK);
343 cookie.lgc_lgl = llh->lgh_id;
344 cookie.lgc_index = hdr->lrh_index;
346 /* cancel them one at a time. I suppose we could store up the cookies
347 * and cancel them all at once; probably more efficient, but this is
348 * done as a user call, so who cares... */
349 rc = llog_cat_cancel_records(env, llh->u.phd.phd_cat_handle, 1,
351 RETURN(rc < 0 ? rc : 0);
354 static int llog_changelog_cancel(const struct lu_env *env,
355 struct llog_ctxt *ctxt,
356 struct changelog_cancel_cookie *cookie)
358 struct llog_handle *cathandle = ctxt->loc_handle;
363 /* This should only be called with the catalog handle */
364 LASSERT(cathandle->lgh_hdr->llh_flags & LLOG_F_IS_CAT);
366 rc = llog_cat_process(env, cathandle, llog_changelog_cancel_cb,
369 /* 0 or 1 means we're done */
372 CERROR("%s: cancel idx %u of catalog "DFID": rc = %d\n",
373 ctxt->loc_obd->obd_name, cathandle->lgh_last_idx,
374 PFID(&cathandle->lgh_id.lgl_oi.oi_fid), rc);
379 static struct llog_operations changelog_orig_logops;
382 mdd_changelog_write_header(const struct lu_env *env, struct mdd_device *mdd,
386 mdd_changelog_on(const struct lu_env *env, struct mdd_device *mdd)
390 if ((mdd->mdd_cl.mc_flags & CLM_ON) != 0)
393 LCONSOLE_INFO("%s: changelog on\n", mdd2obd_dev(mdd)->obd_name);
394 if (mdd->mdd_cl.mc_flags & CLM_ERR) {
395 CERROR("Changelogs cannot be enabled due to error "
396 "condition (see %s log).\n",
397 mdd2obd_dev(mdd)->obd_name);
400 spin_lock(&mdd->mdd_cl.mc_lock);
401 mdd->mdd_cl.mc_flags |= CLM_ON;
402 spin_unlock(&mdd->mdd_cl.mc_lock);
403 rc = mdd_changelog_write_header(env, mdd, CLM_START);
409 mdd_changelog_off(const struct lu_env *env, struct mdd_device *mdd)
413 if ((mdd->mdd_cl.mc_flags & CLM_ON) != CLM_ON)
416 LCONSOLE_INFO("%s: changelog off\n", mdd2obd_dev(mdd)->obd_name);
417 rc = mdd_changelog_write_header(env, mdd, CLM_FINI);
418 spin_lock(&mdd->mdd_cl.mc_lock);
419 mdd->mdd_cl.mc_flags &= ~CLM_ON;
420 spin_unlock(&mdd->mdd_cl.mc_lock);
425 static int mdd_changelog_llog_init(const struct lu_env *env,
426 struct mdd_device *mdd)
428 struct obd_device *obd = mdd2obd_dev(mdd);
429 struct llog_ctxt *ctxt = NULL, *uctxt = NULL;
430 struct changelog_orphan_data changelog_orphan = { .index = 0,
432 user_orphan = { .index = 0,
438 /* LU-2844 mdd setup failure should not cause umount oops */
439 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_CHANGELOG_INIT))
442 OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
443 obd->obd_lvfs_ctxt.dt = mdd->mdd_bottom;
444 rc = llog_setup(env, obd, &obd->obd_olg, LLOG_CHANGELOG_ORIG_CTXT,
445 obd, &changelog_orig_logops);
447 CERROR("%s: changelog llog setup failed: rc = %d\n",
452 ctxt = llog_get_context(obd, LLOG_CHANGELOG_ORIG_CTXT);
455 rc = llog_open_create(env, ctxt, &ctxt->loc_handle, NULL,
458 GOTO(out_cleanup, rc);
460 rc = llog_init_handle(env, ctxt->loc_handle, LLOG_F_IS_CAT, NULL);
464 rc = llog_cat_reverse_process(env, ctxt->loc_handle,
465 changelog_init_cb, mdd);
468 CERROR("%s: changelog init failed: rc = %d\n", obd->obd_name,
473 CDEBUG(D_IOCTL, "changelog starting index=%llu\n",
474 mdd->mdd_cl.mc_index);
476 /* setup user changelog */
477 rc = llog_setup(env, obd, &obd->obd_olg, LLOG_CHANGELOG_USER_ORIG_CTXT,
478 obd, &changelog_orig_logops);
480 CERROR("%s: changelog users llog setup failed: rc = %d\n",
485 uctxt = llog_get_context(obd, LLOG_CHANGELOG_USER_ORIG_CTXT);
488 rc = llog_open_create(env, uctxt, &uctxt->loc_handle, NULL,
491 GOTO(out_ucleanup, rc);
493 rc = llog_init_handle(env, uctxt->loc_handle, LLOG_F_IS_CAT, NULL);
495 GOTO(out_uclose, rc);
497 rc = llog_cat_reverse_process(env, uctxt->loc_handle,
498 changelog_user_init_cb, mdd);
500 CERROR("%s: changelog user init failed: rc = %d\n",
502 GOTO(out_uclose, rc);
505 /* If we have registered users, assume we want changelogs on */
506 if (mdd->mdd_cl.mc_lastuser > 0) {
507 rc = mdd_changelog_on(env, mdd);
509 GOTO(out_uclose, rc);
512 /* find and clear any orphan changelog records (1st record index <
513 * smallest of all users current index), likely to come from an
514 * interrupted manual or GC-thread purge, as its user record had
516 * XXX we may wait for a still registered user clear operation to
517 * do the job, but it may then take a long time to reach the user's
518 * real targetted records if a huge purge backlog is still to be
519 * processed as a long time idle user record could have been deleted
520 * XXX we may need to run end of purge as a separate thread
522 rc = llog_cat_process(env, ctxt->loc_handle, changelog_detect_orphan_cb,
523 &changelog_orphan, 0, 0);
525 CERROR("%s: changelog detect orphan failed: rc = %d\n",
527 GOTO(out_uclose, rc);
529 rc = llog_cat_process(env, uctxt->loc_handle,
530 changelog_user_detect_orphan_cb,
533 CERROR("%s: changelog user detect orphan failed: rc = %d\n",
535 GOTO(out_uclose, rc);
537 if (unlikely(changelog_orphan.index < user_orphan.index)) {
538 struct changelog_cancel_cookie cl_cookie = {
539 .endrec = user_orphan.index,
543 CWARN("%s : orphan changelog records found, starting from "
544 "index %llu to index %llu, being cleared now\n",
545 obd->obd_name, changelog_orphan.index, user_orphan.index);
547 /* XXX we may need to run end of purge as a separate thread */
548 rc = llog_changelog_cancel(env, ctxt, &cl_cookie);
550 CERROR("%s: purge of changelog orphan records failed: "
551 "rc = %d\n", obd->obd_name, rc);
552 GOTO(out_uclose, rc);
557 llog_ctxt_put(uctxt);
560 llog_cat_close(env, uctxt->loc_handle);
562 llog_cleanup(env, uctxt);
564 llog_cat_close(env, ctxt->loc_handle);
566 llog_cleanup(env, ctxt);
570 static int mdd_changelog_init(const struct lu_env *env, struct mdd_device *mdd)
572 struct obd_device *obd = mdd2obd_dev(mdd);
575 mdd->mdd_cl.mc_index = 0;
576 spin_lock_init(&mdd->mdd_cl.mc_lock);
577 mdd->mdd_cl.mc_starttime = ktime_get();
578 spin_lock_init(&mdd->mdd_cl.mc_user_lock);
579 mdd->mdd_cl.mc_lastuser = 0;
581 /* ensure a GC check will, and a thread run may, occur upon start */
582 mdd->mdd_cl.mc_gc_time = 0;
583 mdd->mdd_cl.mc_gc_task = MDD_CHLG_GC_NONE;
585 rc = mdd_changelog_llog_init(env, mdd);
587 CERROR("%s: changelog setup during init failed: rc = %d\n",
589 mdd->mdd_cl.mc_flags |= CLM_ERR;
595 static void mdd_changelog_fini(const struct lu_env *env,
596 struct mdd_device *mdd)
598 struct obd_device *obd = mdd2obd_dev(mdd);
599 struct llog_ctxt *ctxt;
601 mdd->mdd_cl.mc_flags = 0;
604 /* stop GC-thread if running */
605 spin_lock(&mdd->mdd_cl.mc_lock);
606 if (likely(mdd->mdd_cl.mc_gc_task == MDD_CHLG_GC_NONE)) {
607 /* avoid any attempt to run a GC-thread */
608 mdd->mdd_cl.mc_gc_task = current;
609 spin_unlock(&mdd->mdd_cl.mc_lock);
611 struct task_struct *gc_task;
613 if (unlikely(mdd->mdd_cl.mc_gc_task == MDD_CHLG_GC_NEED ||
614 mdd->mdd_cl.mc_gc_task == MDD_CHLG_GC_START)) {
615 /* need to wait for birthing GC-thread to be started
616 * and to have set mc_gc_task to itself
618 spin_unlock(&mdd->mdd_cl.mc_lock);
619 /* Add a tiny sleep */
620 schedule_timeout_uninterruptible(1);
621 /* go back to fully check if GC-thread has started or
622 * even already exited or if a new one is starting...
626 /* take a reference on task_struct to avoid it to be freed
629 gc_task = mdd->mdd_cl.mc_gc_task;
630 get_task_struct(gc_task);
631 spin_unlock(&mdd->mdd_cl.mc_lock);
632 kthread_stop(gc_task);
633 put_task_struct(gc_task);
636 ctxt = llog_get_context(obd, LLOG_CHANGELOG_ORIG_CTXT);
638 llog_cat_close(env, ctxt->loc_handle);
639 llog_cleanup(env, ctxt);
641 ctxt = llog_get_context(obd, LLOG_CHANGELOG_USER_ORIG_CTXT);
643 llog_cat_close(env, ctxt->loc_handle);
644 llog_cleanup(env, ctxt);
648 /** Remove entries with indicies up to and including \a endrec from the
655 mdd_changelog_llog_cancel(const struct lu_env *env, struct mdd_device *mdd,
658 struct obd_device *obd = mdd2obd_dev(mdd);
659 struct llog_ctxt *ctxt;
660 long long unsigned cur;
661 struct changelog_cancel_cookie cookie;
664 ctxt = llog_get_context(obd, LLOG_CHANGELOG_ORIG_CTXT);
668 spin_lock(&mdd->mdd_cl.mc_lock);
669 cur = (long long)mdd->mdd_cl.mc_index;
670 spin_unlock(&mdd->mdd_cl.mc_lock);
674 /* purge to "0" is shorthand for everything */
678 /* If purging all records, write a header entry so we don't have an
679 empty catalog and we're sure to have a valid starting index next
680 time. In case of crash, we just restart with old log so we're
683 /* XXX: transaction is started by llog itself */
684 rc = mdd_changelog_write_header(env, mdd, CLM_PURGE);
689 /* Some records were purged, so reset repeat-access time (so we
690 record new mtime update records, so users can see a file has been
691 changed since the last purge) */
692 mdd->mdd_cl.mc_starttime = ktime_get();
694 cookie.endrec = endrec;
696 rc = llog_changelog_cancel(env, ctxt, &cookie);
702 /** Add a CL_MARK record to the changelog
704 * \param markerflags - CLM_*
707 int mdd_changelog_write_header(const struct lu_env *env,
708 struct mdd_device *mdd, int markerflags)
710 struct obd_device *obd = mdd2obd_dev(mdd);
711 struct llog_changelog_rec *rec;
713 struct llog_ctxt *ctxt;
715 int len = strlen(obd->obd_name);
720 if (mdd->mdd_cl.mc_mask & BIT(CL_MARK)) {
721 mdd->mdd_cl.mc_starttime = ktime_get();
725 reclen = llog_data_len(sizeof(*rec) + len);
726 buf = lu_buf_check_and_alloc(&mdd_env_info(env)->mti_big_buf, reclen);
727 if (buf->lb_buf == NULL)
731 rec->cr.cr_flags = CLF_VERSION;
732 rec->cr.cr_type = CL_MARK;
733 rec->cr.cr_namelen = len;
734 memcpy(changelog_rec_name(&rec->cr), obd->obd_name, rec->cr.cr_namelen);
735 /* Status and action flags */
736 rec->cr.cr_markerflags = mdd->mdd_cl.mc_flags | markerflags;
737 rec->cr_hdr.lrh_len = llog_data_len(changelog_rec_size(&rec->cr) +
739 rec->cr_hdr.lrh_type = CHANGELOG_REC;
740 rec->cr.cr_time = cl_time();
742 ctxt = llog_get_context(obd, LLOG_CHANGELOG_ORIG_CTXT);
745 rc = llog_cat_add(env, ctxt->loc_handle, &rec->cr_hdr, NULL);
750 /* assume on or off event; reset repeat-access time */
751 mdd->mdd_cl.mc_starttime = ktime_get();
756 * Lookup method for "fid" object. Only filenames with correct SEQ:OID format
757 * are valid. We also check if object with passed fid exists or not.
759 static int obf_lookup(const struct lu_env *env, struct md_object *p,
760 const struct lu_name *lname, struct lu_fid *f,
761 struct md_op_spec *spec)
763 char *name = (char *)lname->ln_name;
764 struct mdd_device *mdd = mdo2mdd(p);
765 struct mdd_object *child;
771 sscanf(name, SFID, RFID(f));
773 GOTO(out, rc = -ENOENT);
775 if (!fid_is_norm(f) && !fid_is_igif(f) && !fid_is_root(f) &&
776 !fid_seq_is_dot(f->f_seq))
777 GOTO(out, rc = -ENOENT);
779 /* Check if object with this fid exists */
780 child = mdd_object_find(env, mdd, f);
782 GOTO(out, rc = PTR_ERR(child));
784 if (mdd_object_exists(child) == 0)
787 mdd_object_put(env, child);
793 static int mdd_dummy_create(const struct lu_env *env,
794 struct md_object *pobj,
795 const struct lu_name *lname,
796 struct md_object *child,
797 struct md_op_spec *spec,
803 static int mdd_dummy_rename(const struct lu_env *env,
804 struct md_object *src_pobj,
805 struct md_object *tgt_pobj,
806 const struct lu_fid *lf,
807 const struct lu_name *lsname,
808 struct md_object *tobj,
809 const struct lu_name *ltname,
815 static int mdd_dummy_link(const struct lu_env *env,
816 struct md_object *tgt_obj,
817 struct md_object *src_obj,
818 const struct lu_name *lname,
824 static int mdd_dummy_unlink(const struct lu_env *env,
825 struct md_object *pobj,
826 struct md_object *cobj,
827 const struct lu_name *lname,
834 int mdd_create(const struct lu_env *env, struct md_object *pobj,
835 const struct lu_name *lname, struct md_object *child,
836 struct md_op_spec *spec, struct md_attr *ma);
837 static int mdd_obf_create(const struct lu_env *env, struct md_object *pobj,
838 const struct lu_name *lname, struct md_object *child,
839 struct md_op_spec *spec, struct md_attr *ma)
841 if (spec->sp_cr_flags & MDS_OPEN_VOLATILE)
842 return mdd_create(env, pobj, lname, child, spec, ma);
846 static struct md_dir_operations mdd_obf_dir_ops = {
847 .mdo_lookup = obf_lookup,
848 .mdo_create = mdd_obf_create,
849 .mdo_rename = mdd_dummy_rename,
850 .mdo_link = mdd_dummy_link,
851 .mdo_unlink = mdd_dummy_unlink
854 static struct md_dir_operations mdd_lpf_dir_ops = {
855 .mdo_lookup = mdd_lookup,
856 .mdo_create = mdd_dummy_create,
857 .mdo_rename = mdd_dummy_rename,
858 .mdo_link = mdd_dummy_link,
859 .mdo_unlink = mdd_dummy_unlink
862 static struct md_object *mdo_locate(const struct lu_env *env,
863 struct md_device *md,
864 const struct lu_fid *fid)
866 struct lu_object *obj;
867 struct md_object *mdo;
869 obj = lu_object_find(env, &md->md_lu_dev, fid, NULL);
871 obj = lu_object_locate(obj->lo_header, md->md_lu_dev.ld_type);
872 LASSERT(obj != NULL);
880 static int mdd_lpf_setup(const struct lu_env *env, struct mdd_device *m)
882 struct md_object *mdo;
883 struct mdd_object *mdd_lpf;
884 struct lu_fid fid = LU_LPF_FID;
888 rc = mdd_local_file_create(env, m, mdd_object_fid(m->mdd_dot_lustre),
889 mdd_lpf_dir_name, S_IFDIR | S_IRUSR | S_IXUSR,
894 mdo = mdo_locate(env, &m->mdd_md_dev, &fid);
896 RETURN(PTR_ERR(mdo));
898 LASSERT(lu_object_exists(&mdo->mo_lu));
900 mdd_lpf = md2mdd_obj(mdo);
901 mdd_lpf->mod_obj.mo_dir_ops = &mdd_lpf_dir_ops;
902 m->mdd_dot_lustre_objs.mdd_lpf = mdd_lpf;
908 * Create special in-memory "fid" object for open-by-fid.
910 static int mdd_obf_setup(const struct lu_env *env, struct mdd_device *m)
912 struct md_object *mdo;
913 struct mdd_object *mdd_obf;
914 struct lu_fid fid = LU_OBF_FID;
917 rc = mdd_local_file_create(env, m, mdd_object_fid(m->mdd_dot_lustre),
918 mdd_obf_dir_name, S_IFDIR | S_IXUSR, &fid);
922 mdo = mdo_locate(env, &m->mdd_md_dev, &fid);
924 RETURN(PTR_ERR(mdo));
926 LASSERT(lu_object_exists(&mdo->mo_lu));
928 mdd_obf = md2mdd_obj(mdo);
929 mdd_obf->mod_obj.mo_dir_ops = &mdd_obf_dir_ops;
930 m->mdd_dot_lustre_objs.mdd_obf = mdd_obf;
935 static void mdd_dot_lustre_cleanup(const struct lu_env *env,
936 struct mdd_device *m)
938 if (m->mdd_dot_lustre_objs.mdd_lpf != NULL) {
939 mdd_object_put(env, m->mdd_dot_lustre_objs.mdd_lpf);
940 m->mdd_dot_lustre_objs.mdd_lpf = NULL;
942 if (m->mdd_dot_lustre_objs.mdd_obf != NULL) {
943 mdd_object_put(env, m->mdd_dot_lustre_objs.mdd_obf);
944 m->mdd_dot_lustre_objs.mdd_obf = NULL;
946 if (m->mdd_dot_lustre != NULL) {
947 mdd_object_put(env, m->mdd_dot_lustre);
948 m->mdd_dot_lustre = NULL;
952 /** Setup ".lustre" directory object */
953 static int mdd_dot_lustre_setup(const struct lu_env *env, struct mdd_device *m)
955 struct md_object *mdo;
960 /* Create ".lustre" directory in ROOT. */
961 fid = LU_DOT_LUSTRE_FID;
962 rc = mdd_local_file_create(env, m, &m->mdd_root_fid,
964 S_IFDIR | S_IRUGO | S_IWUSR | S_IXUGO,
968 mdo = mdo_locate(env, &m->mdd_md_dev, &fid);
970 RETURN(PTR_ERR(mdo));
971 LASSERT(lu_object_exists(&mdo->mo_lu));
973 m->mdd_dot_lustre = md2mdd_obj(mdo);
975 rc = mdd_obf_setup(env, m);
977 CERROR("%s: error initializing \"fid\" object: rc = %d.\n",
978 mdd2obd_dev(m)->obd_name, rc);
982 rc = mdd_lpf_setup(env, m);
984 CERROR("%s: error initializing \"lost+found\": rc = %d.\n",
985 mdd2obd_dev(m)->obd_name, rc);
992 mdd_dot_lustre_cleanup(env, m);
998 * set llog methods and create LLOG_AGENT_ORIG_CTXT llog
999 * object in obd_device
1001 static int mdd_hsm_actions_llog_init(const struct lu_env *env,
1002 struct mdd_device *m)
1004 struct obd_device *obd = mdd2obd_dev(m);
1005 struct llog_ctxt *ctxt = NULL;
1009 OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
1010 obd->obd_lvfs_ctxt.dt = m->mdd_bottom;
1012 rc = llog_setup(env, obd, &obd->obd_olg, LLOG_AGENT_ORIG_CTXT,
1013 obd, &llog_common_cat_ops);
1015 CERROR("%s: hsm actions llog setup failed: rc = %d\n",
1020 ctxt = llog_get_context(obd, LLOG_AGENT_ORIG_CTXT);
1023 rc = llog_open_create(env, ctxt, &ctxt->loc_handle, NULL,
1026 CERROR("%s: hsm actions llog open_create failed: rc = %d\n",
1028 GOTO(out_cleanup, rc);
1031 rc = llog_init_handle(env, ctxt->loc_handle, LLOG_F_IS_CAT, NULL);
1033 GOTO(out_close, rc);
1035 llog_ctxt_put(ctxt);
1039 llog_cat_close(env, ctxt->loc_handle);
1040 ctxt->loc_handle = NULL;
1042 llog_cleanup(env, ctxt);
1048 * cleanup the context created by llog_setup_named()
1050 static int mdd_hsm_actions_llog_fini(const struct lu_env *env,
1051 struct mdd_device *m)
1053 struct obd_device *obd = mdd2obd_dev(m);
1054 struct llog_ctxt *lctxt;
1057 lctxt = llog_get_context(obd, LLOG_AGENT_ORIG_CTXT);
1059 llog_cat_close(env, lctxt->loc_handle);
1060 lctxt->loc_handle = NULL;
1061 llog_cleanup(env, lctxt);
1067 static void mdd_device_shutdown(const struct lu_env *env, struct mdd_device *m,
1068 struct lustre_cfg *cfg)
1070 barrier_deregister(m->mdd_bottom);
1071 lfsck_degister(env, m->mdd_bottom);
1072 mdd_hsm_actions_llog_fini(env, m);
1073 mdd_changelog_fini(env, m);
1074 mdd_orphan_index_fini(env, m);
1075 mdd_dot_lustre_cleanup(env, m);
1076 if (mdd2obd_dev(m)->u.obt.obt_nodemap_config_file) {
1077 nm_config_file_deregister_tgt(env,
1078 mdd2obd_dev(m)->u.obt.obt_nodemap_config_file);
1079 mdd2obd_dev(m)->u.obt.obt_nodemap_config_file = NULL;
1081 if (m->mdd_los != NULL) {
1082 local_oid_storage_fini(env, m->mdd_los);
1085 lu_site_purge(env, mdd2lu_dev(m)->ld_site, ~0);
1087 if (m->mdd_child_exp)
1088 obd_disconnect(m->mdd_child_exp);
1091 static int mdd_process_config(const struct lu_env *env,
1092 struct lu_device *d, struct lustre_cfg *cfg)
1094 struct mdd_device *m = lu2mdd_dev(d);
1095 struct dt_device *dt = m->mdd_child;
1096 struct lu_device *next = &dt->dd_lu_dev;
1100 switch (cfg->lcfg_command) {
1104 count = class_modify_config(cfg, PARAM_MDD, &m->mdd_kobj);
1105 rc = count > 0 ? 0 : count;
1107 /* we don't understand; pass it on */
1108 rc = next->ld_ops->ldo_process_config(env, next, cfg);
1112 rc = next->ld_ops->ldo_process_config(env, next, cfg);
1115 dt_conf_get(env, dt, &m->mdd_dt_conf);
1117 case LCFG_PRE_CLEANUP:
1118 rc = next->ld_ops->ldo_process_config(env, next, cfg);
1119 mdd_generic_thread_stop(&m->mdd_orphan_cleanup_thread);
1122 rc = next->ld_ops->ldo_process_config(env, next, cfg);
1123 lu_dev_del_linkage(d->ld_site, d);
1124 mdd_device_shutdown(env, m, cfg);
1127 rc = next->ld_ops->ldo_process_config(env, next, cfg);
1134 static int mdd_recovery_complete(const struct lu_env *env,
1135 struct lu_device *d)
1137 struct mdd_device *mdd = lu2mdd_dev(d);
1138 struct lu_device *next;
1142 LASSERT(mdd != NULL);
1143 next = &mdd->mdd_child->dd_lu_dev;
1145 if (!mdd->mdd_bottom->dd_rdonly)
1146 mdd_orphan_cleanup(env, mdd);
1147 rc = next->ld_ops->ldo_recovery_complete(env, next);
1152 int mdd_local_file_create(const struct lu_env *env, struct mdd_device *mdd,
1153 const struct lu_fid *pfid, const char *name,
1154 __u32 mode, struct lu_fid *fid)
1156 struct dt_object *parent, *dto;
1161 LASSERT(!fid_is_zero(pfid));
1162 parent = dt_locate(env, mdd->mdd_bottom, pfid);
1163 if (unlikely(IS_ERR(parent)))
1164 RETURN(PTR_ERR(parent));
1166 /* create local file/dir, if @fid is passed then try to use it */
1167 if (fid_is_zero(fid))
1168 dto = local_file_find_or_create(env, mdd->mdd_los, parent,
1171 dto = local_file_find_or_create_with_fid(env, mdd->mdd_bottom,
1175 GOTO(out_put, rc = PTR_ERR(dto));
1176 *fid = *lu_object_fid(&dto->do_lu);
1177 /* since stack is not fully set up the local_storage uses own stack
1178 * and we should drop its object from cache */
1179 dt_object_put_nocache(env, dto);
1182 dt_object_put(env, parent);
1186 static int mdd_lfsck_out_notify(const struct lu_env *env, void *data,
1187 enum lfsck_events event)
1192 static int mdd_prepare(const struct lu_env *env,
1193 struct lu_device *pdev,
1194 struct lu_device *cdev)
1196 struct mdd_device *mdd = lu2mdd_dev(cdev);
1197 struct lu_device *next = &mdd->mdd_child->dd_lu_dev;
1198 struct nm_config_file *nodemap_config;
1199 struct obd_device_target *obt = &mdd2obd_dev(mdd)->u.obt;
1205 rc = next->ld_ops->ldo_prepare(env, cdev, next);
1209 /* Setup local dirs */
1210 fid.f_seq = FID_SEQ_LOCAL_NAME;
1213 rc = local_oid_storage_init(env, mdd->mdd_bottom, &fid,
1218 rc = dt_root_get(env, mdd->mdd_child, &mdd->mdd_local_root_fid);
1223 if (mdd_seq_site(mdd)->ss_node_id == 0) {
1224 rc = mdd_local_file_create(env, mdd, &mdd->mdd_local_root_fid,
1225 mdd_root_dir_name, S_IFDIR |
1226 S_IRUGO | S_IWUSR | S_IXUGO, &fid);
1228 CERROR("%s: create root fid failed: rc = %d\n",
1229 mdd2obd_dev(mdd)->obd_name, rc);
1233 mdd->mdd_root_fid = fid;
1234 rc = mdd_dot_lustre_setup(env, mdd);
1236 CERROR("%s: initializing .lustre failed: rc = %d\n",
1237 mdd2obd_dev(mdd)->obd_name, rc);
1241 /* Normal client usually send root access to MDT0 directly,
1242 * the root FID on non-MDT0 will only be used by echo client. */
1243 mdd->mdd_root_fid = fid;
1246 rc = mdd_orphan_index_init(env, mdd);
1250 rc = mdd_changelog_init(env, mdd);
1252 CERROR("%s: failed to initialize changelog: rc = %d\n",
1253 mdd2obd_dev(mdd)->obd_name, rc);
1257 rc = mdd_hsm_actions_llog_init(env, mdd);
1259 GOTO(out_changelog, rc);
1261 nodemap_config = nm_config_file_register_tgt(env, mdd->mdd_bottom,
1263 if (IS_ERR(nodemap_config)) {
1264 rc = PTR_ERR(nodemap_config);
1268 obt->obt_nodemap_config_file = nodemap_config;
1271 rc = lfsck_register(env, mdd->mdd_bottom, mdd->mdd_child,
1272 mdd2obd_dev(mdd), mdd_lfsck_out_notify,
1275 CERROR("%s: failed to initialize lfsck: rc = %d\n",
1276 mdd2obd_dev(mdd)->obd_name, rc);
1277 GOTO(out_nodemap, rc);
1280 rc = barrier_register(mdd->mdd_bottom, mdd->mdd_child);
1282 CERROR("%s: failed to register to barrier: rc = %d\n",
1283 mdd2obd_dev(mdd)->obd_name, rc);
1284 GOTO(out_lfsck, rc);
1290 lfsck_degister(env, mdd->mdd_bottom);
1292 nm_config_file_deregister_tgt(env, obt->obt_nodemap_config_file);
1293 obt->obt_nodemap_config_file = NULL;
1295 mdd_hsm_actions_llog_fini(env, mdd);
1297 mdd_changelog_fini(env, mdd);
1299 mdd_orphan_index_fini(env, mdd);
1301 if (mdd_seq_site(mdd)->ss_node_id == 0)
1302 mdd_dot_lustre_cleanup(env, mdd);
1304 local_oid_storage_fini(env, mdd->mdd_los);
1305 mdd->mdd_los = NULL;
1311 * Implementation of lu_device_operations::ldo_fid_alloc() for MDD.
1313 * Find corresponding device by passed parent and name, and allocate FID from
1316 * see include/lu_object.h for the details.
1318 static int mdd_fid_alloc(const struct lu_env *env, struct lu_device *d,
1319 struct lu_fid *fid, struct lu_object *parent,
1320 const struct lu_name *name)
1322 struct mdd_device *mdd = lu2mdd_dev(d);
1323 struct lu_object *o = lu_object_next(parent);
1325 return dt_fid_alloc(env, mdd->mdd_child, fid, o, name);
1328 const struct lu_device_operations mdd_lu_ops = {
1329 .ldo_object_alloc = mdd_object_alloc,
1330 .ldo_process_config = mdd_process_config,
1331 .ldo_recovery_complete = mdd_recovery_complete,
1332 .ldo_prepare = mdd_prepare,
1333 .ldo_fid_alloc = mdd_fid_alloc,
1336 static int mdd_root_get(const struct lu_env *env,
1337 struct md_device *m, struct lu_fid *f)
1339 struct mdd_device *mdd = lu2mdd_dev(&m->md_lu_dev);
1342 *f = mdd->mdd_root_fid;
1347 * No permission check is needed.
1349 static int mdd_statfs(const struct lu_env *env, struct md_device *m,
1350 struct obd_statfs *sfs)
1352 struct mdd_device *mdd = lu2mdd_dev(&m->md_lu_dev);
1357 rc = mdd_child_ops(mdd)->dt_statfs(env, mdd->mdd_child, sfs, NULL);
1359 sfs->os_namelen = min_t(__u32, sfs->os_namelen, NAME_MAX);
1364 static const struct dt_device_param *mdd_dtconf_get(const struct lu_env *env,
1365 struct md_device *m)
1367 struct mdd_device *mdd = lu2mdd_dev(&m->md_lu_dev);
1369 return &mdd->mdd_dt_conf;
1372 static int mdd_llog_ctxt_get(const struct lu_env *env, struct md_device *m,
1375 struct mdd_device *mdd = lu2mdd_dev(&m->md_lu_dev);
1377 *h = llog_group_get_ctxt(&mdd2obd_dev(mdd)->obd_olg, idx);
1378 return (*h == NULL ? -ENOENT : 0);
1381 static struct lu_device *mdd_device_free(const struct lu_env *env,
1382 struct lu_device *lu)
1384 struct mdd_device *m = lu2mdd_dev(lu);
1387 LASSERT(atomic_read(&lu->ld_ref) == 0);
1388 md_device_fini(&m->mdd_md_dev);
1393 static struct lu_device *mdd_device_alloc(const struct lu_env *env,
1394 struct lu_device_type *t,
1395 struct lustre_cfg *lcfg)
1397 struct lu_device *l;
1398 struct mdd_device *m;
1402 l = ERR_PTR(-ENOMEM);
1407 md_device_init(&m->mdd_md_dev, t);
1408 rc = mdd_init0(env, m, t, lcfg);
1410 mdd_device_free(env, l);
1419 * we use exports to track all mdd users
1421 static int mdd_obd_connect(const struct lu_env *env, struct obd_export **exp,
1422 struct obd_device *obd, struct obd_uuid *cluuid,
1423 struct obd_connect_data *data, void *localdata)
1425 struct mdd_device *mdd = lu2mdd_dev(obd->obd_lu_dev);
1426 struct lustre_handle conn;
1430 CDEBUG(D_CONFIG, "connect #%d\n", mdd->mdd_connects);
1432 rc = class_connect(&conn, obd, cluuid);
1436 *exp = class_conn2export(&conn);
1438 /* Why should there ever be more than 1 connect? */
1439 LASSERT(mdd->mdd_connects == 0);
1440 mdd->mdd_connects++;
1446 * once last export (we don't count self-export) disappeared
1447 * mdd can be released
1449 static int mdd_obd_disconnect(struct obd_export *exp)
1451 struct obd_device *obd = exp->exp_obd;
1452 struct mdd_device *mdd = lu2mdd_dev(obd->obd_lu_dev);
1453 int rc, release = 0;
1456 mdd->mdd_connects--;
1457 if (mdd->mdd_connects == 0)
1460 rc = class_disconnect(exp);
1462 if (rc == 0 && release)
1463 class_manual_cleanup(obd);
1467 static int mdd_obd_get_info(const struct lu_env *env, struct obd_export *exp,
1468 __u32 keylen, void *key, __u32 *vallen, void *val)
1472 if (KEY_IS(KEY_OSP_CONNECTED)) {
1473 struct obd_device *obd = exp->exp_obd;
1474 struct mdd_device *mdd;
1476 if (!obd->obd_set_up || obd->obd_stopping)
1479 mdd = lu2mdd_dev(obd->obd_lu_dev);
1481 rc = obd_get_info(env, mdd->mdd_child_exp, keylen, key, vallen,
1489 static int mdd_obd_set_info_async(const struct lu_env *env,
1490 struct obd_export *exp,
1491 __u32 keylen, void *key,
1492 __u32 vallen, void *val,
1493 struct ptlrpc_request_set *set)
1495 struct obd_device *obd = exp->exp_obd;
1496 struct mdd_device *mdd;
1499 if (!obd->obd_set_up || obd->obd_stopping)
1502 mdd = lu2mdd_dev(obd->obd_lu_dev);
1504 rc = obd_set_info_async(env, mdd->mdd_child_exp, keylen, key,
1509 static const struct obd_ops mdd_obd_device_ops = {
1510 .o_owner = THIS_MODULE,
1511 .o_connect = mdd_obd_connect,
1512 .o_disconnect = mdd_obd_disconnect,
1513 .o_get_info = mdd_obd_get_info,
1514 .o_set_info_async = mdd_obd_set_info_async,
1517 static int mdd_changelog_user_register(const struct lu_env *env,
1518 struct mdd_device *mdd, int *id)
1520 struct llog_ctxt *ctxt;
1521 struct llog_changelog_user_rec *rec;
1525 ctxt = llog_get_context(mdd2obd_dev(mdd),
1526 LLOG_CHANGELOG_USER_ORIG_CTXT);
1532 llog_ctxt_put(ctxt);
1536 CFS_RACE(CFS_FAIL_CHLOG_USER_REG_UNREG_RACE);
1538 rec->cur_hdr.lrh_len = sizeof(*rec);
1539 rec->cur_hdr.lrh_type = CHANGELOG_USER_REC;
1540 spin_lock(&mdd->mdd_cl.mc_user_lock);
1541 if (mdd->mdd_cl.mc_lastuser == (unsigned int)(-1)) {
1542 spin_unlock(&mdd->mdd_cl.mc_user_lock);
1543 CERROR("Maximum number of changelog users exceeded!\n");
1544 GOTO(out, rc = -EOVERFLOW);
1546 *id = rec->cur_id = ++mdd->mdd_cl.mc_lastuser;
1547 mdd->mdd_cl.mc_users++;
1548 rec->cur_endrec = mdd->mdd_cl.mc_index;
1550 rec->cur_time = (__u32)ktime_get_real_seconds();
1551 if (OBD_FAIL_CHECK(OBD_FAIL_TIME_IN_CHLOG_USER))
1554 spin_unlock(&mdd->mdd_cl.mc_user_lock);
1556 rc = llog_cat_add(env, ctxt->loc_handle, &rec->cur_hdr, NULL);
1558 CWARN("%s: Failed to register changelog user %d: rc=%d\n",
1559 mdd2obd_dev(mdd)->obd_name, *id, rc);
1560 spin_lock(&mdd->mdd_cl.mc_user_lock);
1561 mdd->mdd_cl.mc_users--;
1562 spin_unlock(&mdd->mdd_cl.mc_user_lock);
1566 CDEBUG(D_IOCTL, "Registered changelog user %d\n", *id);
1568 /* Assume we want it on since somebody registered */
1569 rc = mdd_changelog_on(env, mdd);
1575 llog_ctxt_put(ctxt);
1579 struct mdd_changelog_user_purge {
1580 struct mdd_device *mcup_mdd;
1582 __u32 mcup_usercount;
1588 * changelog_user_purge callback
1590 * Is called once per user.
1592 * Check to see the user requested is available from rec.
1593 * Truncate the changelog.
1594 * Keep track of the total number of users (calls).
1596 static int mdd_changelog_user_purge_cb(const struct lu_env *env,
1597 struct llog_handle *llh,
1598 struct llog_rec_hdr *hdr, void *data)
1600 struct llog_changelog_user_rec *rec;
1601 struct mdd_changelog_user_purge *mcup = data;
1602 struct llog_cookie cookie;
1607 if ((llh->lgh_hdr->llh_flags & LLOG_F_IS_PLAIN) == 0)
1610 rec = container_of(hdr, struct llog_changelog_user_rec, cur_hdr);
1612 mcup->mcup_usercount++;
1614 if (rec->cur_id != mcup->mcup_id) {
1615 /* truncate to the lowest endrec that is not this user */
1616 mcup->mcup_minrec = min(mcup->mcup_minrec,
1621 /* Unregister this user */
1622 cookie.lgc_lgl = llh->lgh_id;
1623 cookie.lgc_index = hdr->lrh_index;
1625 rc = llog_cat_cancel_records(env, llh->u.phd.phd_cat_handle,
1628 mcup->mcup_found = true;
1629 mcup->mcup_usercount--;
1630 spin_lock(&mcup->mcup_mdd->mdd_cl.mc_user_lock);
1631 mcup->mcup_mdd->mdd_cl.mc_users--;
1632 spin_unlock(&mcup->mcup_mdd->mdd_cl.mc_user_lock);
1638 int mdd_changelog_user_purge(const struct lu_env *env,
1639 struct mdd_device *mdd, __u32 id)
1641 struct mdd_changelog_user_purge mcup = {
1644 .mcup_found = false,
1645 .mcup_usercount = 0,
1646 .mcup_minrec = ULLONG_MAX,
1648 struct llog_ctxt *ctxt;
1653 CDEBUG(D_IOCTL, "%s: Purge request: id=%u\n",
1654 mdd2obd_dev(mdd)->obd_name, id);
1656 ctxt = llog_get_context(mdd2obd_dev(mdd),
1657 LLOG_CHANGELOG_USER_ORIG_CTXT);
1659 (ctxt->loc_handle->lgh_hdr->llh_flags & LLOG_F_IS_CAT) == 0)
1660 GOTO(out, rc = -ENXIO);
1662 rc = llog_cat_process(env, ctxt->loc_handle,
1663 mdd_changelog_user_purge_cb, &mcup,
1666 OBD_FAIL_TIMEOUT(OBD_FAIL_LLOG_PURGE_DELAY, cfs_fail_val);
1668 if ((rc == 0) && (mcup.mcup_usercount == 0)) {
1669 spin_lock(&mdd->mdd_cl.mc_user_lock);
1670 if (mdd->mdd_cl.mc_users == 0) {
1671 /* No more users; turn changelogs off */
1672 CDEBUG(D_IOCTL, "turning off changelogs\n");
1673 rc = mdd_changelog_off(env, mdd);
1675 spin_unlock(&mdd->mdd_cl.mc_user_lock);
1678 if ((rc == 0) && mcup.mcup_found) {
1679 CDEBUG(D_IOCTL, "%s: Purging changelog entries for user %d "
1681 mdd2obd_dev(mdd)->obd_name, id, mcup.mcup_minrec);
1682 /* Cancelling record 0 destroys the entire changelog, make sure
1683 we don't do that unless we mean it. */
1684 if (mcup.mcup_minrec != 0 || mcup.mcup_usercount == 0) {
1685 rc = mdd_changelog_llog_cancel(env, mdd,
1689 CWARN("%s: No changelog for user %u; rc=%d\n",
1690 mdd2obd_dev(mdd)->obd_name, id, rc);
1691 GOTO(out, rc = -ENOENT);
1694 CFS_RACE(CFS_FAIL_CHLOG_USER_REG_UNREG_RACE);
1699 llog_ctxt_put(ctxt);
1704 struct mdd_changelog_user_clear {
1712 * changelog_clear callback
1714 * Is called once per user.
1716 * Check to see the user requested is available from rec.
1717 * Check the oldest (smallest) record for boundary conditions.
1718 * Truncate the changelog.
1720 static int mdd_changelog_clear_cb(const struct lu_env *env,
1721 struct llog_handle *llh,
1722 struct llog_rec_hdr *hdr,
1725 struct llog_changelog_user_rec *rec;
1726 struct mdd_changelog_user_clear *mcuc = data;
1731 if ((llh->lgh_hdr->llh_flags & LLOG_F_IS_PLAIN) == 0)
1734 rec = container_of(hdr, struct llog_changelog_user_rec, cur_hdr);
1736 /* Does the changelog id match the requested id? */
1737 if (rec->cur_id != mcuc->mcuc_id) {
1738 mcuc->mcuc_minrec = min(mcuc->mcuc_minrec,
1743 /* cur_endrec is the oldest purgeable record, make sure we're newer */
1744 if (rec->cur_endrec > mcuc->mcuc_endrec) {
1745 CDEBUG(D_IOCTL, "Request %llu out of range: %llu\n",
1746 mcuc->mcuc_endrec, rec->cur_endrec);
1750 /* Flag that we've met all the range and user checks.
1751 * We now know the record to flush.
1753 rec->cur_endrec = mcuc->mcuc_endrec;
1755 rec->cur_time = (__u32)ktime_get_real_seconds();
1756 if (OBD_FAIL_CHECK(OBD_FAIL_TIME_IN_CHLOG_USER))
1759 mcuc->mcuc_flush = true;
1761 CDEBUG(D_IOCTL, "Rewriting changelog user %u endrec to %llu\n",
1762 mcuc->mcuc_id, rec->cur_endrec);
1764 /* Update the endrec */
1765 rc = llog_write(env, llh, hdr, hdr->lrh_index);
1771 * Clear a changelog up to entry specified by endrec for user id.
1773 static int mdd_changelog_clear(const struct lu_env *env,
1774 struct mdd_device *mdd, __u32 id,
1777 struct mdd_changelog_user_clear mcuc = {
1779 .mcuc_minrec = endrec,
1780 .mcuc_flush = false,
1782 struct llog_ctxt *ctxt;
1788 CDEBUG(D_IOCTL, "%s: Purge request: id=%u, endrec=%llu\n",
1789 mdd2obd_dev(mdd)->obd_name, id, endrec);
1790 /* start_rec is the newest (largest value) entry in the changelogs*/
1791 spin_lock(&mdd->mdd_cl.mc_lock);
1792 start_rec = mdd->mdd_cl.mc_index;
1793 spin_unlock(&mdd->mdd_cl.mc_lock);
1795 if (start_rec < endrec) {
1796 CDEBUG(D_IOCTL, "%s: Could not clear changelog, requested "\
1797 "address out of range\n", mdd2obd_dev(mdd)->obd_name);
1802 mcuc.mcuc_endrec = start_rec;
1803 mcuc.mcuc_minrec = ULLONG_MAX;
1805 mcuc.mcuc_endrec = endrec;
1808 ctxt = llog_get_context(mdd2obd_dev(mdd),
1809 LLOG_CHANGELOG_USER_ORIG_CTXT);
1811 (ctxt->loc_handle->lgh_hdr->llh_flags & LLOG_F_IS_CAT) == 0)
1812 GOTO(out, rc = -ENXIO);
1814 rc = llog_cat_process(env, ctxt->loc_handle,
1815 mdd_changelog_clear_cb, (void *)&mcuc,
1819 CWARN("%s: Failure to clear the changelog for user %d: %d\n",
1820 mdd2obd_dev(mdd)->obd_name, id, rc);
1821 } else if (mcuc.mcuc_flush) {
1822 /* Cancelling record 0 destroys the entire changelog, make sure
1823 we don't do that unless we mean it. */
1824 if (mcuc.mcuc_minrec != 0) {
1825 CDEBUG(D_IOCTL, "%s: Purging changelog entries up "\
1826 "to %llu\n", mdd2obd_dev(mdd)->obd_name,
1829 rc = mdd_changelog_llog_cancel(env, mdd,
1833 CWARN("%s: No entry for user %d\n",
1834 mdd2obd_dev(mdd)->obd_name, id);
1841 llog_ctxt_put(ctxt);
1847 * May be called remotely from mdt_iocontrol_handle or locally from
1848 * mdt_iocontrol. Data may be freeform - remote handling doesn't enforce
1849 * an obd_ioctl_data format (but local ioctl handler does).
1851 * \param len - data len
1852 * \param karg - ioctl data, in kernel space
1854 static int mdd_iocontrol(const struct lu_env *env, struct md_device *m,
1855 unsigned int cmd, int len, void *karg)
1857 struct mdd_device *mdd = lu2mdd_dev(&m->md_lu_dev);
1858 struct obd_ioctl_data *data = karg;
1862 /* Doesn't use obd_ioctl_data */
1864 case OBD_IOC_CHANGELOG_CLEAR: {
1865 struct changelog_setinfo *cs = karg;
1867 if (unlikely(!barrier_entry(mdd->mdd_bottom)))
1868 RETURN(-EINPROGRESS);
1870 rc = mdd_changelog_clear(env, mdd, cs->cs_id,
1872 barrier_exit(mdd->mdd_bottom);
1875 case OBD_IOC_START_LFSCK: {
1876 rc = lfsck_start(env, mdd->mdd_bottom,
1877 (struct lfsck_start_param *)karg);
1880 case OBD_IOC_STOP_LFSCK: {
1881 rc = lfsck_stop(env, mdd->mdd_bottom,
1882 (struct lfsck_stop *)karg);
1885 case OBD_IOC_QUERY_LFSCK: {
1886 rc = lfsck_query(env, mdd->mdd_bottom, NULL, NULL,
1887 (struct lfsck_query *)karg);
1892 /* Below ioctls use obd_ioctl_data */
1893 if (len != sizeof(*data)) {
1894 CERROR("Bad ioctl size %d\n", len);
1897 if (data->ioc_version != OBD_IOCTL_VERSION) {
1898 CERROR("Bad magic %x != %x\n", data->ioc_version,
1904 case OBD_IOC_CHANGELOG_REG:
1905 if (unlikely(!barrier_entry(mdd->mdd_bottom)))
1906 RETURN(-EINPROGRESS);
1908 rc = mdd_changelog_user_register(env, mdd, &data->ioc_u32_1);
1909 barrier_exit(mdd->mdd_bottom);
1911 case OBD_IOC_CHANGELOG_DEREG:
1912 if (unlikely(!barrier_entry(mdd->mdd_bottom)))
1913 RETURN(-EINPROGRESS);
1915 /* explicitly clear changelog first, to protect from crash in
1916 * the middle of purge that would lead to unregistered consumer
1917 * but pending changelog entries
1919 rc = mdd_changelog_clear(env, mdd, data->ioc_u32_1, 0);
1921 rc = mdd_changelog_user_purge(env,
1922 mdd, data->ioc_u32_1);
1924 barrier_exit(mdd->mdd_bottom);
1933 /* type constructor/destructor: mdd_type_init, mdd_type_fini */
1934 LU_TYPE_INIT_FINI(mdd, &mdd_thread_key);
1936 static const struct md_device_operations mdd_ops = {
1937 .mdo_statfs = mdd_statfs,
1938 .mdo_root_get = mdd_root_get,
1939 .mdo_llog_ctxt_get = mdd_llog_ctxt_get,
1940 .mdo_iocontrol = mdd_iocontrol,
1941 .mdo_dtconf_get = mdd_dtconf_get,
1944 static struct lu_device_type_operations mdd_device_type_ops = {
1945 .ldto_init = mdd_type_init,
1946 .ldto_fini = mdd_type_fini,
1948 .ldto_start = mdd_type_start,
1949 .ldto_stop = mdd_type_stop,
1951 .ldto_device_alloc = mdd_device_alloc,
1952 .ldto_device_free = mdd_device_free,
1954 .ldto_device_fini = mdd_device_fini
1957 static struct lu_device_type mdd_device_type = {
1958 .ldt_tags = LU_DEVICE_MD,
1959 .ldt_name = LUSTRE_MDD_NAME,
1960 .ldt_ops = &mdd_device_type_ops,
1961 .ldt_ctx_tags = LCT_MD_THREAD
1964 /* context key constructor: mdd_key_init */
1965 LU_KEY_INIT(mdd, struct mdd_thread_info);
1967 static void mdd_key_fini(const struct lu_context *ctx,
1968 struct lu_context_key *key, void *data)
1970 struct mdd_thread_info *info = data;
1972 lu_buf_free(&info->mti_big_buf);
1973 lu_buf_free(&info->mti_link_buf);
1974 lu_buf_free(&info->mti_xattr_buf);
1979 /* context key: mdd_thread_key */
1980 LU_CONTEXT_KEY_DEFINE(mdd, LCT_MD_THREAD);
1982 int mdd_generic_thread_start(struct mdd_generic_thread *thread,
1983 int (*func)(void *), void *data, char *name)
1985 struct task_struct *task;
1987 LASSERT(thread->mgt_init == false);
1988 init_completion(&thread->mgt_started);
1989 init_completion(&thread->mgt_finished);
1990 thread->mgt_data = data;
1991 thread->mgt_abort = false;
1992 thread->mgt_init = true;
1994 task = kthread_run(func, thread, name);
1996 complete(&thread->mgt_finished);
1997 return PTR_ERR(task);
1999 wait_for_completion(&thread->mgt_started);
2003 void mdd_generic_thread_stop(struct mdd_generic_thread *thread)
2005 if (thread->mgt_init == true) {
2006 thread->mgt_abort = true;
2007 wait_for_completion(&thread->mgt_finished);
2011 static int __init mdd_init(void)
2015 rc = lu_kmem_init(mdd_caches);
2019 changelog_orig_logops = llog_common_cat_ops;
2020 changelog_orig_logops.lop_write_rec = mdd_changelog_write_rec;
2022 rc = class_register_type(&mdd_obd_device_ops, NULL, false, NULL,
2023 LUSTRE_MDD_NAME, &mdd_device_type);
2025 lu_kmem_fini(mdd_caches);
2029 static void __exit mdd_exit(void)
2031 class_unregister_type(LUSTRE_MDD_NAME);
2032 lu_kmem_fini(mdd_caches);
2035 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
2036 MODULE_DESCRIPTION("Lustre Meta-data Device Driver ("LUSTRE_MDD_NAME")");
2037 MODULE_VERSION(LUSTRE_VERSION_STRING);
2038 MODULE_LICENSE("GPL");
2040 module_init(mdd_init);
2041 module_exit(mdd_exit);