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 = 0;
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 = data;
189 struct llog_changelog_rec *rec = (struct llog_changelog_rec *)hdr;
191 if (unlikely(!(llh->lgh_hdr->llh_flags & LLOG_F_IS_PLAIN))) {
192 CWARN("%s: changelog "DFID" must be a plain llog\n",
193 mdd2obd_dev(mdd)->obd_name,
194 PFID(&llh->lgh_id.lgl_oi.oi_fid));
198 if (rec->cr_hdr.lrh_type != CHANGELOG_REC) {
199 CWARN("%s: invalid record at index %#x in log "DFID"\n",
200 mdd2obd_dev(mdd)->obd_name, hdr->lrh_index,
201 PFID(&llh->lgh_id.lgl_oi.oi_fid));
206 "seeing record at index %d/%d/%llu t=%x %.*s in log"
207 DFID"\n", hdr->lrh_index, rec->cr_hdr.lrh_index,
208 rec->cr.cr_index, rec->cr.cr_type, rec->cr.cr_namelen,
209 changelog_rec_name(&rec->cr), PFID(&llh->lgh_id.lgl_oi.oi_fid));
211 mdd->mdd_cl.mc_index = rec->cr.cr_index;
212 return LLOG_PROC_BREAK;
215 static int changelog_user_init_cb(const struct lu_env *env,
216 struct llog_handle *llh,
217 struct llog_rec_hdr *hdr, void *data)
219 struct mdd_device *mdd = (struct mdd_device *)data;
220 struct llog_changelog_user_rec *rec =
221 (struct llog_changelog_user_rec *)hdr;
223 if (unlikely(!(llh->lgh_hdr->llh_flags & LLOG_F_IS_PLAIN))) {
224 CWARN("%s: changelog "DFID" must be a plain llog\n",
225 mdd2obd_dev(mdd)->obd_name,
226 PFID(&llh->lgh_id.lgl_oi.oi_fid));
230 if (rec->cur_hdr.lrh_type != CHANGELOG_USER_REC &&
231 rec->cur_hdr.lrh_type != CHANGELOG_USER_REC2) {
232 CWARN("%s: unknown user type %x at index %u in log "DFID"\n",
233 mdd2obd_dev(mdd)->obd_name, hdr->lrh_index,
234 rec->cur_hdr.lrh_type, PFID(&llh->lgh_id.lgl_oi.oi_fid));
238 CDEBUG(D_INFO, "seeing user at index %d/%d id=%d endrec=%llu"
239 " in log "DFID"\n", hdr->lrh_index, rec->cur_hdr.lrh_index,
240 rec->cur_id, rec->cur_endrec, PFID(&llh->lgh_id.lgl_oi.oi_fid));
242 spin_lock(&mdd->mdd_cl.mc_user_lock);
243 mdd->mdd_cl.mc_lastuser = rec->cur_id;
244 mdd->mdd_cl.mc_users++;
245 if (rec->cur_endrec > mdd->mdd_cl.mc_index)
246 mdd->mdd_cl.mc_index = rec->cur_endrec;
247 spin_unlock(&mdd->mdd_cl.mc_user_lock);
249 return LLOG_PROC_BREAK;
252 struct changelog_orphan_data {
254 struct mdd_device *mdd;
257 /* find oldest changelog record index */
258 static int changelog_detect_orphan_cb(const struct lu_env *env,
259 struct llog_handle *llh,
260 struct llog_rec_hdr *hdr, void *data)
262 struct mdd_device *mdd = ((struct changelog_orphan_data *)data)->mdd;
263 struct llog_changelog_rec *rec = container_of(hdr,
264 struct llog_changelog_rec,
267 if (unlikely(!(llh->lgh_hdr->llh_flags & LLOG_F_IS_PLAIN))) {
268 CWARN("%s: changelog "DFID" must be a plain llog\n",
269 mdd2obd_dev(mdd)->obd_name,
270 PFID(&llh->lgh_id.lgl_oi.oi_fid));
274 if (rec->cr_hdr.lrh_type != CHANGELOG_REC) {
275 CWARN("%s: invalid record at index %d in log "DFID"\n",
276 mdd2obd_dev(mdd)->obd_name, hdr->lrh_index,
277 PFID(&llh->lgh_id.lgl_oi.oi_fid));
278 /* try to find some next valid record and thus allow to recover
279 * from a corrupted LLOG, instead to assert and force a crash
284 CDEBUG(D_INFO, "%s: seeing record at index %d/%d/%llu t=%x %.*s in log "
285 DFID"\n", mdd2obd_dev(mdd)->obd_name, hdr->lrh_index,
286 rec->cr_hdr.lrh_index, rec->cr.cr_index, rec->cr.cr_type,
287 rec->cr.cr_namelen, changelog_rec_name(&rec->cr),
288 PFID(&llh->lgh_id.lgl_oi.oi_fid));
290 ((struct changelog_orphan_data *)data)->index = rec->cr.cr_index;
291 return LLOG_PROC_BREAK;
294 /* find oldest changelog user index */
295 static int changelog_user_detect_orphan_cb(const struct lu_env *env,
296 struct llog_handle *llh,
297 struct llog_rec_hdr *hdr, void *data)
299 struct mdd_device *mdd = ((struct changelog_orphan_data *)data)->mdd;
300 struct llog_changelog_user_rec *rec = container_of(hdr,
301 struct llog_changelog_user_rec,
304 if (unlikely(!(llh->lgh_hdr->llh_flags & LLOG_F_IS_PLAIN))) {
305 CWARN("%s: changelog "DFID" must be a plain llog\n",
306 mdd2obd_dev(mdd)->obd_name,
307 PFID(&llh->lgh_id.lgl_oi.oi_fid));
311 if (rec->cur_hdr.lrh_type != CHANGELOG_USER_REC &&
312 rec->cur_hdr.lrh_type != CHANGELOG_USER_REC2) {
313 CWARN("%s: invalid user at index %d in log "DFID"\n",
314 mdd2obd_dev(mdd)->obd_name, hdr->lrh_index,
315 PFID(&llh->lgh_id.lgl_oi.oi_fid));
316 /* try to find some next valid record and thus allow to recover
317 * from a corrupted LLOG, instead to assert and force a crash
322 CDEBUG(D_INFO, "%s: seeing user at index %d/%d id=%d endrec=%llu in "
323 "log "DFID"\n", mdd2obd_dev(mdd)->obd_name, hdr->lrh_index,
324 rec->cur_hdr.lrh_index, rec->cur_id, rec->cur_endrec,
325 PFID(&llh->lgh_id.lgl_oi.oi_fid));
327 if (((struct changelog_orphan_data *)data)->index == 0 ||
328 rec->cur_endrec < ((struct changelog_orphan_data *)data)->index)
329 ((struct changelog_orphan_data *)data)->index = rec->cur_endrec;
334 struct changelog_cancel_cookie {
336 struct mdd_device *mdd;
339 static int llog_changelog_cancel_cb(const struct lu_env *env,
340 struct llog_handle *llh,
341 struct llog_rec_hdr *hdr, void *data)
343 struct llog_changelog_rec *rec = (struct llog_changelog_rec *)hdr;
344 struct changelog_cancel_cookie *cl_cookie =
345 (struct changelog_cancel_cookie *)data;
349 /* This is always a (sub)log, not the catalog */
350 if (unlikely(!(llh->lgh_hdr->llh_flags & LLOG_F_IS_PLAIN))) {
351 CWARN("%s: changelog "DFID" must be a plain llog\n",
352 mdd2obd_dev(cl_cookie->mdd)->obd_name,
353 PFID(&llh->lgh_id.lgl_oi.oi_fid));
357 /* if current context is GC-thread allow it to stop upon umount
358 * remaining records cleanup will occur upon next mount
360 * also during testing, wait for GC-thread to be released
362 * XXX this requires the GC-thread to not fork a sub-thread via
363 * llog[_cat]_process_or_fork() and we may think to also implement
364 * this shutdown mechanism for manually started user unregister which
365 * can also take a long time if huge backlog of records
367 if (unlikely(cl_cookie->mdd->mdd_cl.mc_gc_task == current)) {
368 /* wait to be released */
369 while (CFS_FAIL_CHECK_QUIET(OBD_FAIL_FORCE_GC_THREAD))
372 if (kthread_should_stop())
373 RETURN(LLOG_PROC_BREAK);
376 if (rec->cr.cr_index > cl_cookie->endrec)
377 /* records are in order, so we're done */
378 RETURN(LLOG_PROC_BREAK);
380 if (unlikely(OBD_FAIL_PRECHECK(OBD_FAIL_MDS_CHANGELOG_RACE))) {
381 if (cfs_fail_val == 0)
382 cfs_fail_val = hdr->lrh_index;
383 if (cfs_fail_val == hdr->lrh_index)
384 OBD_RACE(OBD_FAIL_MDS_CHANGELOG_RACE);
387 /* Records folow one by one, cr_index++. We could calculate the
388 * last cr_index at this plain llog. And if it less then cookie endrec
389 * cancel the whole file.
391 if ((LLOG_HDR_BITMAP_SIZE(llh->lgh_hdr) - hdr->lrh_index +
392 rec->cr.cr_index) < cl_cookie->endrec) {
395 if (unlikely(OBD_FAIL_PRECHECK(OBD_FAIL_MDS_CHANGELOG_DEL))) {
396 if (cfs_fail_val == 0) {
397 cfs_fail_val = (unsigned long)llh & 0xFFFFFFFF;
398 OBD_RACE(OBD_FAIL_MDS_CHANGELOG_DEL);
401 rc = llog_destroy(env, llh);
403 CDEBUG(D_HA, "Changelog destroyed plain "DFID"\n",
404 PFID(&llh->lgh_id.lgl_oi.oi_fid));
405 RETURN(LLOG_DEL_PLAIN);
409 /* cancel them one at a time. I suppose we could store up the cookies
410 * and cancel them all at once; probably more efficient, but this is
411 * done as a user call, so who cares... */
413 RETURN(LLOG_DEL_RECORD);
416 static int llog_changelog_cancel(const struct lu_env *env,
417 struct llog_ctxt *ctxt,
418 struct changelog_cancel_cookie *cookie)
420 struct llog_handle *cathandle = ctxt->loc_handle;
425 /* This should only be called with the catalog handle */
426 if (unlikely(!(cathandle->lgh_hdr->llh_flags & LLOG_F_IS_CAT))) {
427 CWARN("%s: changelog "DFID" must be a catalog llog\n",
428 mdd2obd_dev(cookie->mdd)->obd_name,
429 PFID(&cathandle->lgh_id.lgl_oi.oi_fid));
434 rc = llog_cat_process(env, cathandle, llog_changelog_cancel_cb,
437 /* 0 or 1 means we're done */
440 CERROR("%s: cancel idx %u of catalog "DFID": rc = %d\n",
441 ctxt->loc_obd->obd_name, cathandle->lgh_last_idx,
442 PFID(&cathandle->lgh_id.lgl_oi.oi_fid), rc);
447 static struct llog_operations changelog_orig_logops;
450 mdd_changelog_write_header(const struct lu_env *env, struct mdd_device *mdd,
454 mdd_changelog_on(const struct lu_env *env, struct mdd_device *mdd)
458 if ((mdd->mdd_cl.mc_flags & CLM_ON) != 0)
461 LCONSOLE_INFO("%s: changelog on\n", mdd2obd_dev(mdd)->obd_name);
462 if (mdd->mdd_cl.mc_flags & CLM_ERR) {
463 CERROR("Changelogs cannot be enabled due to error "
464 "condition (see %s log).\n",
465 mdd2obd_dev(mdd)->obd_name);
468 spin_lock(&mdd->mdd_cl.mc_lock);
469 mdd->mdd_cl.mc_flags |= CLM_ON;
470 spin_unlock(&mdd->mdd_cl.mc_lock);
471 rc = mdd_changelog_write_header(env, mdd, CLM_START);
477 mdd_changelog_off(const struct lu_env *env, struct mdd_device *mdd)
481 if ((mdd->mdd_cl.mc_flags & CLM_ON) != CLM_ON)
484 LCONSOLE_INFO("%s: changelog off\n", mdd2obd_dev(mdd)->obd_name);
485 rc = mdd_changelog_write_header(env, mdd, CLM_FINI);
486 spin_lock(&mdd->mdd_cl.mc_lock);
487 mdd->mdd_cl.mc_flags &= ~CLM_ON;
488 spin_unlock(&mdd->mdd_cl.mc_lock);
493 static int mdd_changelog_llog_init(const struct lu_env *env,
494 struct mdd_device *mdd)
496 struct obd_device *obd = mdd2obd_dev(mdd);
497 struct llog_ctxt *ctxt = NULL, *uctxt = NULL;
498 struct changelog_orphan_data changelog_orphan = { .index = 0,
500 user_orphan = { .index = 0,
506 /* LU-2844 mdd setup failure should not cause umount oops */
507 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_CHANGELOG_INIT))
510 OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
511 obd->obd_lvfs_ctxt.dt = mdd->mdd_bottom;
512 rc = llog_setup(env, obd, &obd->obd_olg, LLOG_CHANGELOG_ORIG_CTXT,
513 obd, &changelog_orig_logops);
515 CERROR("%s: changelog llog setup failed: rc = %d\n",
520 ctxt = llog_get_context(obd, LLOG_CHANGELOG_ORIG_CTXT);
523 rc = llog_open_create(env, ctxt, &ctxt->loc_handle, NULL,
526 GOTO(out_cleanup, rc);
528 rc = llog_init_handle(env, ctxt->loc_handle, LLOG_F_IS_CAT, NULL);
532 rc = llog_cat_reverse_process(env, ctxt->loc_handle,
533 changelog_init_cb, mdd);
536 CERROR("%s: changelog init failed: rc = %d\n", obd->obd_name,
541 CDEBUG(D_IOCTL, "changelog starting index=%llu\n",
542 mdd->mdd_cl.mc_index);
544 /* setup user changelog */
545 rc = llog_setup(env, obd, &obd->obd_olg, LLOG_CHANGELOG_USER_ORIG_CTXT,
546 obd, &changelog_orig_logops);
548 CERROR("%s: changelog users llog setup failed: rc = %d\n",
553 uctxt = llog_get_context(obd, LLOG_CHANGELOG_USER_ORIG_CTXT);
556 rc = llog_open_create(env, uctxt, &uctxt->loc_handle, NULL,
559 GOTO(out_ucleanup, rc);
561 rc = llog_init_handle(env, uctxt->loc_handle, LLOG_F_IS_CAT, NULL);
563 GOTO(out_uclose, rc);
565 rc = llog_cat_reverse_process(env, uctxt->loc_handle,
566 changelog_user_init_cb, mdd);
568 CERROR("%s: changelog user init failed: rc = %d\n",
570 GOTO(out_uclose, rc);
573 /* If we have registered users, assume we want changelogs on */
574 if (mdd->mdd_cl.mc_lastuser > 0) {
575 rc = mdd_changelog_on(env, mdd);
577 GOTO(out_uclose, rc);
580 /* find and clear any orphan changelog records (1st record index <
581 * smallest of all users current index), likely to come from an
582 * interrupted manual or GC-thread purge, as its user record had
584 * XXX we may wait for a still registered user clear operation to
585 * do the job, but it may then take a long time to reach the user's
586 * real targetted records if a huge purge backlog is still to be
587 * processed as a long time idle user record could have been deleted
588 * XXX we may need to run end of purge as a separate thread
590 rc = llog_cat_process(env, ctxt->loc_handle, changelog_detect_orphan_cb,
591 &changelog_orphan, 0, 0);
593 CERROR("%s: changelog detect orphan failed: rc = %d\n",
595 GOTO(out_uclose, rc);
597 rc = llog_cat_process(env, uctxt->loc_handle,
598 changelog_user_detect_orphan_cb,
601 CERROR("%s: changelog user detect orphan failed: rc = %d\n",
603 GOTO(out_uclose, rc);
605 if (unlikely(changelog_orphan.index < user_orphan.index)) {
606 struct changelog_cancel_cookie cl_cookie = {
607 .endrec = user_orphan.index,
611 CWARN("%s : orphan changelog records found, starting from "
612 "index %llu to index %llu, being cleared now\n",
613 obd->obd_name, changelog_orphan.index, user_orphan.index);
615 /* XXX we may need to run end of purge as a separate thread */
616 rc = llog_changelog_cancel(env, ctxt, &cl_cookie);
618 CERROR("%s: purge of changelog orphan records failed: "
619 "rc = %d\n", obd->obd_name, rc);
620 GOTO(out_uclose, rc);
625 llog_ctxt_put(uctxt);
628 llog_cat_close(env, uctxt->loc_handle);
630 llog_cleanup(env, uctxt);
632 llog_cat_close(env, ctxt->loc_handle);
634 llog_cleanup(env, ctxt);
638 static int mdd_changelog_init(const struct lu_env *env, struct mdd_device *mdd)
640 struct obd_device *obd = mdd2obd_dev(mdd);
643 mdd->mdd_cl.mc_index = 0;
644 spin_lock_init(&mdd->mdd_cl.mc_lock);
645 mdd->mdd_cl.mc_starttime = ktime_get();
646 spin_lock_init(&mdd->mdd_cl.mc_user_lock);
647 mdd->mdd_cl.mc_lastuser = 0;
649 /* ensure a GC check will, and a thread run may, occur upon start */
650 mdd->mdd_cl.mc_gc_time = 0;
651 mdd->mdd_cl.mc_gc_task = MDD_CHLG_GC_NONE;
653 rc = mdd_changelog_llog_init(env, mdd);
655 CERROR("%s: changelog setup during init failed: rc = %d\n",
657 mdd->mdd_cl.mc_flags |= CLM_ERR;
663 static void mdd_changelog_fini(const struct lu_env *env,
664 struct mdd_device *mdd)
666 struct obd_device *obd = mdd2obd_dev(mdd);
667 struct llog_ctxt *ctxt;
669 mdd->mdd_cl.mc_flags = 0;
672 /* stop GC-thread if running */
673 spin_lock(&mdd->mdd_cl.mc_lock);
674 if (likely(mdd->mdd_cl.mc_gc_task == MDD_CHLG_GC_NONE)) {
675 /* avoid any attempt to run a GC-thread */
676 mdd->mdd_cl.mc_gc_task = current;
677 spin_unlock(&mdd->mdd_cl.mc_lock);
679 struct task_struct *gc_task;
681 if (unlikely(mdd->mdd_cl.mc_gc_task == MDD_CHLG_GC_NEED ||
682 mdd->mdd_cl.mc_gc_task == MDD_CHLG_GC_START)) {
683 /* need to wait for birthing GC-thread to be started
684 * and to have set mc_gc_task to itself
686 spin_unlock(&mdd->mdd_cl.mc_lock);
687 schedule_timeout(usecs_to_jiffies(10));
688 /* go back to fully check if GC-thread has started or
689 * even already exited or if a new one is starting...
693 /* take a reference on task_struct to avoid it to be freed
696 gc_task = mdd->mdd_cl.mc_gc_task;
697 get_task_struct(gc_task);
698 spin_unlock(&mdd->mdd_cl.mc_lock);
699 kthread_stop(gc_task);
700 put_task_struct(gc_task);
703 ctxt = llog_get_context(obd, LLOG_CHANGELOG_ORIG_CTXT);
705 llog_cat_close(env, ctxt->loc_handle);
706 llog_cleanup(env, ctxt);
708 ctxt = llog_get_context(obd, LLOG_CHANGELOG_USER_ORIG_CTXT);
710 llog_cat_close(env, ctxt->loc_handle);
711 llog_cleanup(env, ctxt);
715 /** Remove entries with indicies up to and including \a endrec from the
722 mdd_changelog_llog_cancel(const struct lu_env *env, struct mdd_device *mdd,
725 struct obd_device *obd = mdd2obd_dev(mdd);
726 struct llog_ctxt *ctxt;
727 long long unsigned cur;
728 struct changelog_cancel_cookie cookie;
731 ctxt = llog_get_context(obd, LLOG_CHANGELOG_ORIG_CTXT);
735 spin_lock(&mdd->mdd_cl.mc_lock);
736 cur = (long long)mdd->mdd_cl.mc_index;
737 spin_unlock(&mdd->mdd_cl.mc_lock);
741 /* purge to "0" is shorthand for everything */
745 /* If purging all records, write a header entry so we don't have an
746 empty catalog and we're sure to have a valid starting index next
747 time. In case of crash, we just restart with old log so we're
750 /* XXX: transaction is started by llog itself */
751 rc = mdd_changelog_write_header(env, mdd, CLM_PURGE);
756 /* Some records were purged, so reset repeat-access time (so we
757 record new mtime update records, so users can see a file has been
758 changed since the last purge) */
759 mdd->mdd_cl.mc_starttime = ktime_get();
761 cookie.endrec = endrec;
763 rc = llog_changelog_cancel(env, ctxt, &cookie);
769 /** Add a CL_MARK record to the changelog
771 * \param markerflags - CLM_*
774 int mdd_changelog_write_header(const struct lu_env *env,
775 struct mdd_device *mdd, int markerflags)
777 struct obd_device *obd = mdd2obd_dev(mdd);
778 struct llog_changelog_rec *rec;
780 struct llog_ctxt *ctxt;
782 int len = strlen(obd->obd_name);
787 if (mdd->mdd_cl.mc_mask & (1 << CL_MARK)) {
788 mdd->mdd_cl.mc_starttime = ktime_get();
792 reclen = llog_data_len(sizeof(*rec) + len);
793 buf = lu_buf_check_and_alloc(&mdd_env_info(env)->mti_big_buf, reclen);
794 if (buf->lb_buf == NULL)
798 rec->cr.cr_flags = CLF_VERSION;
799 rec->cr.cr_type = CL_MARK;
800 rec->cr.cr_namelen = len;
801 memcpy(changelog_rec_name(&rec->cr), obd->obd_name, rec->cr.cr_namelen);
802 /* Status and action flags */
803 rec->cr.cr_markerflags = mdd->mdd_cl.mc_flags | markerflags;
804 rec->cr_hdr.lrh_len = llog_data_len(changelog_rec_size(&rec->cr) +
806 rec->cr_hdr.lrh_type = CHANGELOG_REC;
807 rec->cr.cr_time = cl_time();
809 ctxt = llog_get_context(obd, LLOG_CHANGELOG_ORIG_CTXT);
812 rc = llog_cat_add(env, ctxt->loc_handle, &rec->cr_hdr, NULL);
817 /* assume on or off event; reset repeat-access time */
818 mdd->mdd_cl.mc_starttime = ktime_get();
823 * Lookup method for "fid" object. Only filenames with correct SEQ:OID format
824 * are valid. We also check if object with passed fid exists or not.
826 static int obf_lookup(const struct lu_env *env, struct md_object *p,
827 const struct lu_name *lname, struct lu_fid *f,
828 struct md_op_spec *spec)
830 char *name = (char *)lname->ln_name;
831 struct mdd_device *mdd = mdo2mdd(p);
832 struct mdd_object *child;
838 sscanf(name, SFID, RFID(f));
839 if (!fid_is_sane(f)) {
840 CWARN("%s: Trying to lookup invalid FID [%s] in %s/%s, FID "
841 "format should be "DFID"\n", mdd2obd_dev(mdd)->obd_name,
842 lname->ln_name, dot_lustre_name, mdd_obf_dir_name,
843 (__u64)FID_SEQ_NORMAL, 1, 0);
844 GOTO(out, rc = -EINVAL);
847 if (!fid_is_norm(f) && !fid_is_igif(f) && !fid_is_root(f) &&
848 !fid_seq_is_dot(f->f_seq)) {
849 CWARN("%s: Trying to lookup invalid FID "DFID" in %s/%s, "
850 "sequence should be >= %#llx or within [%#llx,"
851 "%#llx].\n", mdd2obd_dev(mdd)->obd_name, PFID(f),
852 dot_lustre_name, mdd_obf_dir_name, (__u64)FID_SEQ_NORMAL,
853 (__u64)FID_SEQ_IGIF, (__u64)FID_SEQ_IGIF_MAX);
854 GOTO(out, rc = -EINVAL);
857 /* Check if object with this fid exists */
858 child = mdd_object_find(env, mdd, f);
860 GOTO(out, rc = PTR_ERR(child));
862 if (mdd_object_exists(child) == 0)
865 mdd_object_put(env, child);
871 static int mdd_dummy_create(const struct lu_env *env,
872 struct md_object *pobj,
873 const struct lu_name *lname,
874 struct md_object *child,
875 struct md_op_spec *spec,
881 static int mdd_dummy_rename(const struct lu_env *env,
882 struct md_object *src_pobj,
883 struct md_object *tgt_pobj,
884 const struct lu_fid *lf,
885 const struct lu_name *lsname,
886 struct md_object *tobj,
887 const struct lu_name *ltname,
893 static int mdd_dummy_link(const struct lu_env *env,
894 struct md_object *tgt_obj,
895 struct md_object *src_obj,
896 const struct lu_name *lname,
902 static int mdd_dummy_unlink(const struct lu_env *env,
903 struct md_object *pobj,
904 struct md_object *cobj,
905 const struct lu_name *lname,
912 int mdd_create(const struct lu_env *env, struct md_object *pobj,
913 const struct lu_name *lname, struct md_object *child,
914 struct md_op_spec *spec, struct md_attr *ma);
915 static int mdd_obf_create(const struct lu_env *env, struct md_object *pobj,
916 const struct lu_name *lname, struct md_object *child,
917 struct md_op_spec *spec, struct md_attr *ma)
919 if (spec->sp_cr_flags & MDS_OPEN_VOLATILE)
920 return mdd_create(env, pobj, lname, child, spec, ma);
924 static struct md_dir_operations mdd_obf_dir_ops = {
925 .mdo_lookup = obf_lookup,
926 .mdo_create = mdd_obf_create,
927 .mdo_rename = mdd_dummy_rename,
928 .mdo_link = mdd_dummy_link,
929 .mdo_unlink = mdd_dummy_unlink
932 static struct md_dir_operations mdd_lpf_dir_ops = {
933 .mdo_lookup = mdd_lookup,
934 .mdo_create = mdd_dummy_create,
935 .mdo_rename = mdd_dummy_rename,
936 .mdo_link = mdd_dummy_link,
937 .mdo_unlink = mdd_dummy_unlink
940 static struct md_object *mdo_locate(const struct lu_env *env,
941 struct md_device *md,
942 const struct lu_fid *fid)
944 struct lu_object *obj;
945 struct md_object *mdo;
947 obj = lu_object_find(env, &md->md_lu_dev, fid, NULL);
949 obj = lu_object_locate(obj->lo_header, md->md_lu_dev.ld_type);
950 LASSERT(obj != NULL);
953 mdo = ERR_PTR(PTR_ERR(obj));
958 static int mdd_lpf_setup(const struct lu_env *env, struct mdd_device *m)
960 struct md_object *mdo;
961 struct mdd_object *mdd_lpf;
962 struct lu_fid fid = LU_LPF_FID;
966 rc = mdd_local_file_create(env, m, mdd_object_fid(m->mdd_dot_lustre),
967 mdd_lpf_dir_name, S_IFDIR | S_IRUSR | S_IXUSR,
972 mdo = mdo_locate(env, &m->mdd_md_dev, &fid);
974 RETURN(PTR_ERR(mdo));
976 LASSERT(lu_object_exists(&mdo->mo_lu));
978 mdd_lpf = md2mdd_obj(mdo);
979 mdd_lpf->mod_obj.mo_dir_ops = &mdd_lpf_dir_ops;
980 m->mdd_dot_lustre_objs.mdd_lpf = mdd_lpf;
986 * Create special in-memory "fid" object for open-by-fid.
988 static int mdd_obf_setup(const struct lu_env *env, struct mdd_device *m)
990 struct md_object *mdo;
991 struct mdd_object *mdd_obf;
992 struct lu_fid fid = LU_OBF_FID;
995 rc = mdd_local_file_create(env, m, mdd_object_fid(m->mdd_dot_lustre),
996 mdd_obf_dir_name, S_IFDIR | S_IXUSR, &fid);
1000 mdo = mdo_locate(env, &m->mdd_md_dev, &fid);
1002 RETURN(PTR_ERR(mdo));
1004 LASSERT(lu_object_exists(&mdo->mo_lu));
1006 mdd_obf = md2mdd_obj(mdo);
1007 mdd_obf->mod_obj.mo_dir_ops = &mdd_obf_dir_ops;
1008 m->mdd_dot_lustre_objs.mdd_obf = mdd_obf;
1013 static void mdd_dot_lustre_cleanup(const struct lu_env *env,
1014 struct mdd_device *m)
1016 if (m->mdd_dot_lustre_objs.mdd_lpf != NULL) {
1017 mdd_object_put(env, m->mdd_dot_lustre_objs.mdd_lpf);
1018 m->mdd_dot_lustre_objs.mdd_lpf = NULL;
1020 if (m->mdd_dot_lustre_objs.mdd_obf != NULL) {
1021 mdd_object_put(env, m->mdd_dot_lustre_objs.mdd_obf);
1022 m->mdd_dot_lustre_objs.mdd_obf = NULL;
1024 if (m->mdd_dot_lustre != NULL) {
1025 mdd_object_put(env, m->mdd_dot_lustre);
1026 m->mdd_dot_lustre = NULL;
1030 /** Setup ".lustre" directory object */
1031 static int mdd_dot_lustre_setup(const struct lu_env *env, struct mdd_device *m)
1033 struct md_object *mdo;
1038 /* Create ".lustre" directory in ROOT. */
1039 fid = LU_DOT_LUSTRE_FID;
1040 rc = mdd_local_file_create(env, m, &m->mdd_root_fid,
1042 S_IFDIR | S_IRUGO | S_IWUSR | S_IXUGO,
1046 mdo = mdo_locate(env, &m->mdd_md_dev, &fid);
1048 RETURN(PTR_ERR(mdo));
1049 LASSERT(lu_object_exists(&mdo->mo_lu));
1051 m->mdd_dot_lustre = md2mdd_obj(mdo);
1053 rc = mdd_obf_setup(env, m);
1055 CERROR("%s: error initializing \"fid\" object: rc = %d.\n",
1056 mdd2obd_dev(m)->obd_name, rc);
1060 rc = mdd_lpf_setup(env, m);
1062 CERROR("%s: error initializing \"lost+found\": rc = %d.\n",
1063 mdd2obd_dev(m)->obd_name, rc);
1070 mdd_dot_lustre_cleanup(env, m);
1076 * set llog methods and create LLOG_AGENT_ORIG_CTXT llog
1077 * object in obd_device
1079 static int mdd_hsm_actions_llog_init(const struct lu_env *env,
1080 struct mdd_device *m)
1082 struct obd_device *obd = mdd2obd_dev(m);
1083 struct llog_ctxt *ctxt = NULL;
1087 OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
1088 obd->obd_lvfs_ctxt.dt = m->mdd_bottom;
1090 rc = llog_setup(env, obd, &obd->obd_olg, LLOG_AGENT_ORIG_CTXT,
1091 obd, &llog_common_cat_ops);
1093 CERROR("%s: hsm actions llog setup failed: rc = %d\n",
1098 ctxt = llog_get_context(obd, LLOG_AGENT_ORIG_CTXT);
1101 rc = llog_open_create(env, ctxt, &ctxt->loc_handle, NULL,
1104 CERROR("%s: hsm actions llog open_create failed: rc = %d\n",
1106 GOTO(out_cleanup, rc);
1109 rc = llog_init_handle(env, ctxt->loc_handle, LLOG_F_IS_CAT, NULL);
1111 GOTO(out_close, rc);
1113 llog_ctxt_put(ctxt);
1117 llog_cat_close(env, ctxt->loc_handle);
1118 ctxt->loc_handle = NULL;
1120 llog_cleanup(env, ctxt);
1126 * cleanup the context created by llog_setup_named()
1128 static int mdd_hsm_actions_llog_fini(const struct lu_env *env,
1129 struct mdd_device *m)
1131 struct obd_device *obd = mdd2obd_dev(m);
1132 struct llog_ctxt *lctxt;
1135 lctxt = llog_get_context(obd, LLOG_AGENT_ORIG_CTXT);
1137 llog_cat_close(env, lctxt->loc_handle);
1138 lctxt->loc_handle = NULL;
1139 llog_cleanup(env, lctxt);
1145 static void mdd_device_shutdown(const struct lu_env *env, struct mdd_device *m,
1146 struct lustre_cfg *cfg)
1148 barrier_deregister(m->mdd_bottom);
1149 lfsck_degister(env, m->mdd_bottom);
1150 mdd_hsm_actions_llog_fini(env, m);
1151 mdd_changelog_fini(env, m);
1152 mdd_orphan_index_fini(env, m);
1153 mdd_dot_lustre_cleanup(env, m);
1154 if (mdd2obd_dev(m)->u.obt.obt_nodemap_config_file) {
1155 nm_config_file_deregister_tgt(env,
1156 mdd2obd_dev(m)->u.obt.obt_nodemap_config_file);
1157 mdd2obd_dev(m)->u.obt.obt_nodemap_config_file = NULL;
1159 if (m->mdd_los != NULL) {
1160 local_oid_storage_fini(env, m->mdd_los);
1163 lu_site_purge(env, mdd2lu_dev(m)->ld_site, ~0);
1165 if (m->mdd_child_exp)
1166 obd_disconnect(m->mdd_child_exp);
1169 static int mdd_process_config(const struct lu_env *env,
1170 struct lu_device *d, struct lustre_cfg *cfg)
1172 struct mdd_device *m = lu2mdd_dev(d);
1173 struct dt_device *dt = m->mdd_child;
1174 struct lu_device *next = &dt->dd_lu_dev;
1178 switch (cfg->lcfg_command) {
1182 count = class_modify_config(cfg, PARAM_MDD, &m->mdd_kobj);
1183 rc = count > 0 ? 0 : count;
1185 /* we don't understand; pass it on */
1186 rc = next->ld_ops->ldo_process_config(env, next, cfg);
1190 rc = next->ld_ops->ldo_process_config(env, next, cfg);
1193 dt_conf_get(env, dt, &m->mdd_dt_conf);
1195 case LCFG_PRE_CLEANUP:
1196 rc = next->ld_ops->ldo_process_config(env, next, cfg);
1197 mdd_generic_thread_stop(&m->mdd_orphan_cleanup_thread);
1200 rc = next->ld_ops->ldo_process_config(env, next, cfg);
1201 lu_dev_del_linkage(d->ld_site, d);
1202 mdd_device_shutdown(env, m, cfg);
1205 rc = next->ld_ops->ldo_process_config(env, next, cfg);
1212 static int mdd_recovery_complete(const struct lu_env *env,
1213 struct lu_device *d)
1215 struct mdd_device *mdd = lu2mdd_dev(d);
1216 struct lu_device *next;
1220 LASSERT(mdd != NULL);
1221 next = &mdd->mdd_child->dd_lu_dev;
1223 if (!mdd->mdd_bottom->dd_rdonly)
1224 mdd_orphan_cleanup(env, mdd);
1225 rc = next->ld_ops->ldo_recovery_complete(env, next);
1230 int mdd_local_file_create(const struct lu_env *env, struct mdd_device *mdd,
1231 const struct lu_fid *pfid, const char *name,
1232 __u32 mode, struct lu_fid *fid)
1234 struct dt_object *parent, *dto;
1239 LASSERT(!fid_is_zero(pfid));
1240 parent = dt_locate(env, mdd->mdd_bottom, pfid);
1241 if (unlikely(IS_ERR(parent)))
1242 RETURN(PTR_ERR(parent));
1244 /* create local file/dir, if @fid is passed then try to use it */
1245 if (fid_is_zero(fid))
1246 dto = local_file_find_or_create(env, mdd->mdd_los, parent,
1249 dto = local_file_find_or_create_with_fid(env, mdd->mdd_bottom,
1253 GOTO(out_put, rc = PTR_ERR(dto));
1254 *fid = *lu_object_fid(&dto->do_lu);
1255 /* since stack is not fully set up the local_storage uses own stack
1256 * and we should drop its object from cache */
1257 dt_object_put_nocache(env, dto);
1260 dt_object_put(env, parent);
1264 static int mdd_lfsck_out_notify(const struct lu_env *env, void *data,
1265 enum lfsck_events event)
1270 static int mdd_prepare(const struct lu_env *env,
1271 struct lu_device *pdev,
1272 struct lu_device *cdev)
1274 struct mdd_device *mdd = lu2mdd_dev(cdev);
1275 struct lu_device *next = &mdd->mdd_child->dd_lu_dev;
1276 struct nm_config_file *nodemap_config;
1277 struct obd_device_target *obt = &mdd2obd_dev(mdd)->u.obt;
1283 rc = next->ld_ops->ldo_prepare(env, cdev, next);
1287 /* Setup local dirs */
1288 fid.f_seq = FID_SEQ_LOCAL_NAME;
1291 rc = local_oid_storage_init(env, mdd->mdd_bottom, &fid,
1296 rc = dt_root_get(env, mdd->mdd_child, &mdd->mdd_local_root_fid);
1301 if (mdd_seq_site(mdd)->ss_node_id == 0) {
1302 rc = mdd_local_file_create(env, mdd, &mdd->mdd_local_root_fid,
1303 mdd_root_dir_name, S_IFDIR |
1304 S_IRUGO | S_IWUSR | S_IXUGO, &fid);
1306 CERROR("%s: create root fid failed: rc = %d\n",
1307 mdd2obd_dev(mdd)->obd_name, rc);
1311 mdd->mdd_root_fid = fid;
1312 rc = mdd_dot_lustre_setup(env, mdd);
1314 CERROR("%s: initializing .lustre failed: rc = %d\n",
1315 mdd2obd_dev(mdd)->obd_name, rc);
1319 /* Normal client usually send root access to MDT0 directly,
1320 * the root FID on non-MDT0 will only be used by echo client. */
1321 mdd->mdd_root_fid = fid;
1324 rc = mdd_orphan_index_init(env, mdd);
1328 rc = mdd_changelog_init(env, mdd);
1330 CERROR("%s: failed to initialize changelog: rc = %d\n",
1331 mdd2obd_dev(mdd)->obd_name, rc);
1335 rc = mdd_hsm_actions_llog_init(env, mdd);
1337 GOTO(out_changelog, rc);
1339 nodemap_config = nm_config_file_register_tgt(env, mdd->mdd_bottom,
1341 if (IS_ERR(nodemap_config)) {
1342 rc = PTR_ERR(nodemap_config);
1346 obt->obt_nodemap_config_file = nodemap_config;
1349 rc = lfsck_register(env, mdd->mdd_bottom, mdd->mdd_child,
1350 mdd2obd_dev(mdd), mdd_lfsck_out_notify,
1353 CERROR("%s: failed to initialize lfsck: rc = %d\n",
1354 mdd2obd_dev(mdd)->obd_name, rc);
1355 GOTO(out_nodemap, rc);
1358 rc = barrier_register(mdd->mdd_bottom, mdd->mdd_child);
1360 CERROR("%s: failed to register to barrier: rc = %d\n",
1361 mdd2obd_dev(mdd)->obd_name, rc);
1362 GOTO(out_lfsck, rc);
1368 lfsck_degister(env, mdd->mdd_bottom);
1370 nm_config_file_deregister_tgt(env, obt->obt_nodemap_config_file);
1371 obt->obt_nodemap_config_file = NULL;
1373 mdd_hsm_actions_llog_fini(env, mdd);
1375 mdd_changelog_fini(env, mdd);
1377 mdd_orphan_index_fini(env, mdd);
1379 if (mdd_seq_site(mdd)->ss_node_id == 0)
1380 mdd_dot_lustre_cleanup(env, mdd);
1382 local_oid_storage_fini(env, mdd->mdd_los);
1383 mdd->mdd_los = NULL;
1388 const struct lu_device_operations mdd_lu_ops = {
1389 .ldo_object_alloc = mdd_object_alloc,
1390 .ldo_process_config = mdd_process_config,
1391 .ldo_recovery_complete = mdd_recovery_complete,
1392 .ldo_prepare = mdd_prepare,
1395 static int mdd_root_get(const struct lu_env *env,
1396 struct md_device *m, struct lu_fid *f)
1398 struct mdd_device *mdd = lu2mdd_dev(&m->md_lu_dev);
1401 *f = mdd->mdd_root_fid;
1406 * No permission check is needed.
1408 static int mdd_statfs(const struct lu_env *env, struct md_device *m,
1409 struct obd_statfs *sfs)
1411 struct mdd_device *mdd = lu2mdd_dev(&m->md_lu_dev);
1416 rc = mdd_child_ops(mdd)->dt_statfs(env, mdd->mdd_child, sfs);
1418 sfs->os_namelen = min_t(__u32, sfs->os_namelen, NAME_MAX);
1423 static const struct dt_device_param *mdd_dtconf_get(const struct lu_env *env,
1424 struct md_device *m)
1426 struct mdd_device *mdd = lu2mdd_dev(&m->md_lu_dev);
1428 return &mdd->mdd_dt_conf;
1431 static int mdd_llog_ctxt_get(const struct lu_env *env, struct md_device *m,
1434 struct mdd_device *mdd = lu2mdd_dev(&m->md_lu_dev);
1436 *h = llog_group_get_ctxt(&mdd2obd_dev(mdd)->obd_olg, idx);
1437 return (*h == NULL ? -ENOENT : 0);
1440 static struct lu_device *mdd_device_free(const struct lu_env *env,
1441 struct lu_device *lu)
1443 struct mdd_device *m = lu2mdd_dev(lu);
1446 LASSERT(atomic_read(&lu->ld_ref) == 0);
1447 md_device_fini(&m->mdd_md_dev);
1452 static struct lu_device *mdd_device_alloc(const struct lu_env *env,
1453 struct lu_device_type *t,
1454 struct lustre_cfg *lcfg)
1456 struct lu_device *l;
1457 struct mdd_device *m;
1461 l = ERR_PTR(-ENOMEM);
1466 md_device_init(&m->mdd_md_dev, t);
1467 rc = mdd_init0(env, m, t, lcfg);
1469 mdd_device_free(env, l);
1478 * we use exports to track all mdd users
1480 static int mdd_obd_connect(const struct lu_env *env, struct obd_export **exp,
1481 struct obd_device *obd, struct obd_uuid *cluuid,
1482 struct obd_connect_data *data, void *localdata)
1484 struct mdd_device *mdd = lu2mdd_dev(obd->obd_lu_dev);
1485 struct lustre_handle conn;
1489 CDEBUG(D_CONFIG, "connect #%d\n", mdd->mdd_connects);
1491 rc = class_connect(&conn, obd, cluuid);
1495 *exp = class_conn2export(&conn);
1497 /* Why should there ever be more than 1 connect? */
1498 LASSERT(mdd->mdd_connects == 0);
1499 mdd->mdd_connects++;
1505 * once last export (we don't count self-export) disappeared
1506 * mdd can be released
1508 static int mdd_obd_disconnect(struct obd_export *exp)
1510 struct obd_device *obd = exp->exp_obd;
1511 struct mdd_device *mdd = lu2mdd_dev(obd->obd_lu_dev);
1512 int rc, release = 0;
1515 mdd->mdd_connects--;
1516 if (mdd->mdd_connects == 0)
1519 rc = class_disconnect(exp);
1521 if (rc == 0 && release)
1522 class_manual_cleanup(obd);
1526 static int mdd_obd_get_info(const struct lu_env *env, struct obd_export *exp,
1527 __u32 keylen, void *key, __u32 *vallen, void *val)
1531 if (KEY_IS(KEY_OSP_CONNECTED)) {
1532 struct obd_device *obd = exp->exp_obd;
1533 struct mdd_device *mdd;
1535 if (!obd->obd_set_up || obd->obd_stopping)
1538 mdd = lu2mdd_dev(obd->obd_lu_dev);
1540 rc = obd_get_info(env, mdd->mdd_child_exp, keylen, key, vallen,
1548 static int mdd_obd_set_info_async(const struct lu_env *env,
1549 struct obd_export *exp,
1550 __u32 keylen, void *key,
1551 __u32 vallen, void *val,
1552 struct ptlrpc_request_set *set)
1554 struct obd_device *obd = exp->exp_obd;
1555 struct mdd_device *mdd;
1558 if (!obd->obd_set_up || obd->obd_stopping)
1561 mdd = lu2mdd_dev(obd->obd_lu_dev);
1563 rc = obd_set_info_async(env, mdd->mdd_child_exp, keylen, key,
1568 static struct obd_ops mdd_obd_device_ops = {
1569 .o_owner = THIS_MODULE,
1570 .o_connect = mdd_obd_connect,
1571 .o_disconnect = mdd_obd_disconnect,
1572 .o_get_info = mdd_obd_get_info,
1573 .o_set_info_async = mdd_obd_set_info_async,
1576 static int mdd_changelog_user_register(const struct lu_env *env,
1577 struct mdd_device *mdd, int *id)
1579 struct llog_ctxt *ctxt;
1580 struct llog_changelog_user_rec *rec;
1584 ctxt = llog_get_context(mdd2obd_dev(mdd),
1585 LLOG_CHANGELOG_USER_ORIG_CTXT);
1591 llog_ctxt_put(ctxt);
1595 CFS_RACE(CFS_FAIL_CHLOG_USER_REG_UNREG_RACE);
1597 rec->cur_hdr.lrh_len = sizeof(*rec);
1598 rec->cur_hdr.lrh_type = CHANGELOG_USER_REC;
1599 spin_lock(&mdd->mdd_cl.mc_user_lock);
1600 if (mdd->mdd_cl.mc_lastuser == (unsigned int)(-1)) {
1601 spin_unlock(&mdd->mdd_cl.mc_user_lock);
1602 CERROR("Maximum number of changelog users exceeded!\n");
1603 GOTO(out, rc = -EOVERFLOW);
1605 *id = rec->cur_id = ++mdd->mdd_cl.mc_lastuser;
1606 mdd->mdd_cl.mc_users++;
1607 rec->cur_endrec = mdd->mdd_cl.mc_index;
1609 rec->cur_time = (__u32)ktime_get_real_seconds();
1610 if (OBD_FAIL_CHECK(OBD_FAIL_TIME_IN_CHLOG_USER))
1613 spin_unlock(&mdd->mdd_cl.mc_user_lock);
1615 rc = llog_cat_add(env, ctxt->loc_handle, &rec->cur_hdr, NULL);
1617 CWARN("%s: Failed to register changelog user %d: rc=%d\n",
1618 mdd2obd_dev(mdd)->obd_name, *id, rc);
1619 spin_lock(&mdd->mdd_cl.mc_user_lock);
1620 mdd->mdd_cl.mc_users--;
1621 spin_unlock(&mdd->mdd_cl.mc_user_lock);
1625 CDEBUG(D_IOCTL, "Registered changelog user %d\n", *id);
1627 /* Assume we want it on since somebody registered */
1628 rc = mdd_changelog_on(env, mdd);
1634 llog_ctxt_put(ctxt);
1638 struct mdd_changelog_user_purge {
1639 struct mdd_device *mcup_mdd;
1641 __u32 mcup_usercount;
1647 * changelog_user_purge callback
1649 * Is called once per user.
1651 * Check to see the user requested is available from rec.
1652 * Truncate the changelog.
1653 * Keep track of the total number of users (calls).
1655 static int mdd_changelog_user_purge_cb(const struct lu_env *env,
1656 struct llog_handle *llh,
1657 struct llog_rec_hdr *hdr, void *data)
1659 struct llog_changelog_user_rec *rec;
1660 struct mdd_changelog_user_purge *mcup = data;
1661 struct llog_cookie cookie;
1666 if ((llh->lgh_hdr->llh_flags & LLOG_F_IS_PLAIN) == 0)
1669 rec = container_of(hdr, struct llog_changelog_user_rec, cur_hdr);
1671 mcup->mcup_usercount++;
1673 if (rec->cur_id != mcup->mcup_id) {
1674 /* truncate to the lowest endrec that is not this user */
1675 mcup->mcup_minrec = min(mcup->mcup_minrec,
1680 /* Unregister this user */
1681 cookie.lgc_lgl = llh->lgh_id;
1682 cookie.lgc_index = hdr->lrh_index;
1684 rc = llog_cat_cancel_records(env, llh->u.phd.phd_cat_handle,
1687 mcup->mcup_found = true;
1688 mcup->mcup_usercount--;
1689 spin_lock(&mcup->mcup_mdd->mdd_cl.mc_user_lock);
1690 mcup->mcup_mdd->mdd_cl.mc_users--;
1691 spin_unlock(&mcup->mcup_mdd->mdd_cl.mc_user_lock);
1697 int mdd_changelog_user_purge(const struct lu_env *env,
1698 struct mdd_device *mdd, __u32 id)
1700 struct mdd_changelog_user_purge mcup = {
1703 .mcup_found = false,
1704 .mcup_usercount = 0,
1705 .mcup_minrec = ULLONG_MAX,
1707 struct llog_ctxt *ctxt;
1712 CDEBUG(D_IOCTL, "%s: Purge request: id=%u\n",
1713 mdd2obd_dev(mdd)->obd_name, id);
1715 ctxt = llog_get_context(mdd2obd_dev(mdd),
1716 LLOG_CHANGELOG_USER_ORIG_CTXT);
1718 (ctxt->loc_handle->lgh_hdr->llh_flags & LLOG_F_IS_CAT) == 0)
1719 GOTO(out, rc = -ENXIO);
1721 rc = llog_cat_process(env, ctxt->loc_handle,
1722 mdd_changelog_user_purge_cb, &mcup,
1725 OBD_FAIL_TIMEOUT(OBD_FAIL_LLOG_PURGE_DELAY, cfs_fail_val);
1727 if ((rc == 0) && (mcup.mcup_usercount == 0)) {
1728 spin_lock(&mdd->mdd_cl.mc_user_lock);
1729 if (mdd->mdd_cl.mc_users == 0) {
1730 /* No more users; turn changelogs off */
1731 CDEBUG(D_IOCTL, "turning off changelogs\n");
1732 rc = mdd_changelog_off(env, mdd);
1734 spin_unlock(&mdd->mdd_cl.mc_user_lock);
1737 if ((rc == 0) && mcup.mcup_found) {
1738 CDEBUG(D_IOCTL, "%s: Purging changelog entries for user %d "
1740 mdd2obd_dev(mdd)->obd_name, id, mcup.mcup_minrec);
1741 /* Cancelling record 0 destroys the entire changelog, make sure
1742 we don't do that unless we mean it. */
1743 if (mcup.mcup_minrec != 0 || mcup.mcup_usercount == 0) {
1744 rc = mdd_changelog_llog_cancel(env, mdd,
1748 CWARN("%s: No changelog for user %u; rc=%d\n",
1749 mdd2obd_dev(mdd)->obd_name, id, rc);
1750 GOTO(out, rc = -ENOENT);
1753 CFS_RACE(CFS_FAIL_CHLOG_USER_REG_UNREG_RACE);
1758 llog_ctxt_put(ctxt);
1763 struct mdd_changelog_user_clear {
1771 * changelog_clear callback
1773 * Is called once per user.
1775 * Check to see the user requested is available from rec.
1776 * Check the oldest (smallest) record for boundary conditions.
1777 * Truncate the changelog.
1779 static int mdd_changelog_clear_cb(const struct lu_env *env,
1780 struct llog_handle *llh,
1781 struct llog_rec_hdr *hdr,
1784 struct llog_changelog_user_rec *rec;
1785 struct mdd_changelog_user_clear *mcuc = data;
1790 if ((llh->lgh_hdr->llh_flags & LLOG_F_IS_PLAIN) == 0)
1793 rec = container_of(hdr, struct llog_changelog_user_rec, cur_hdr);
1795 /* Does the changelog id match the requested id? */
1796 if (rec->cur_id != mcuc->mcuc_id) {
1797 mcuc->mcuc_minrec = min(mcuc->mcuc_minrec,
1802 /* cur_endrec is the oldest purgeable record, make sure we're newer */
1803 if (rec->cur_endrec > mcuc->mcuc_endrec) {
1804 CDEBUG(D_IOCTL, "Request %llu out of range: %llu\n",
1805 mcuc->mcuc_endrec, rec->cur_endrec);
1809 /* Flag that we've met all the range and user checks.
1810 * We now know the record to flush.
1812 rec->cur_endrec = mcuc->mcuc_endrec;
1814 rec->cur_time = (__u32)ktime_get_real_seconds();
1815 if (OBD_FAIL_CHECK(OBD_FAIL_TIME_IN_CHLOG_USER))
1818 mcuc->mcuc_flush = true;
1820 CDEBUG(D_IOCTL, "Rewriting changelog user %u endrec to %llu\n",
1821 mcuc->mcuc_id, rec->cur_endrec);
1823 /* Update the endrec */
1824 rc = llog_write(env, llh, hdr, hdr->lrh_index);
1830 * Clear a changelog up to entry specified by endrec for user id.
1832 static int mdd_changelog_clear(const struct lu_env *env,
1833 struct mdd_device *mdd, __u32 id,
1836 struct mdd_changelog_user_clear mcuc = {
1838 .mcuc_minrec = endrec,
1839 .mcuc_flush = false,
1841 struct llog_ctxt *ctxt;
1847 CDEBUG(D_IOCTL, "%s: Purge request: id=%u, endrec=%llu\n",
1848 mdd2obd_dev(mdd)->obd_name, id, endrec);
1849 /* start_rec is the newest (largest value) entry in the changelogs*/
1850 spin_lock(&mdd->mdd_cl.mc_lock);
1851 start_rec = mdd->mdd_cl.mc_index;
1852 spin_unlock(&mdd->mdd_cl.mc_lock);
1854 if (start_rec < endrec) {
1855 CDEBUG(D_IOCTL, "%s: Could not clear changelog, requested "\
1856 "address out of range\n", mdd2obd_dev(mdd)->obd_name);
1861 mcuc.mcuc_endrec = start_rec;
1862 mcuc.mcuc_minrec = ULLONG_MAX;
1864 mcuc.mcuc_endrec = endrec;
1867 ctxt = llog_get_context(mdd2obd_dev(mdd),
1868 LLOG_CHANGELOG_USER_ORIG_CTXT);
1870 (ctxt->loc_handle->lgh_hdr->llh_flags & LLOG_F_IS_CAT) == 0)
1871 GOTO(out, rc = -ENXIO);
1873 rc = llog_cat_process(env, ctxt->loc_handle,
1874 mdd_changelog_clear_cb, (void *)&mcuc,
1877 if (rc == -EINVAL) {
1878 CDEBUG(D_IOCTL, "%s: No changelog recnum <= %llu to clear\n",
1879 mdd2obd_dev(mdd)->obd_name, (unsigned long long) endrec);
1881 } else if (rc < 0) {
1882 CWARN("%s: Failure to clear the changelog for user %d: %d\n",
1883 mdd2obd_dev(mdd)->obd_name, id, rc);
1884 } else if (mcuc.mcuc_flush) {
1885 /* Cancelling record 0 destroys the entire changelog, make sure
1886 we don't do that unless we mean it. */
1887 if (mcuc.mcuc_minrec != 0) {
1888 CDEBUG(D_IOCTL, "%s: Purging changelog entries up "\
1889 "to %llu\n", mdd2obd_dev(mdd)->obd_name,
1892 rc = mdd_changelog_llog_cancel(env, mdd,
1896 CDEBUG(D_IOCTL, "%s: No entry for user %d\n",
1897 mdd2obd_dev(mdd)->obd_name, id);
1904 llog_ctxt_put(ctxt);
1910 * May be called remotely from mdt_iocontrol_handle or locally from
1911 * mdt_iocontrol. Data may be freeform - remote handling doesn't enforce
1912 * an obd_ioctl_data format (but local ioctl handler does).
1914 * \param len - data len
1915 * \param karg - ioctl data, in kernel space
1917 static int mdd_iocontrol(const struct lu_env *env, struct md_device *m,
1918 unsigned int cmd, int len, void *karg)
1920 struct mdd_device *mdd = lu2mdd_dev(&m->md_lu_dev);
1921 struct obd_ioctl_data *data = karg;
1925 /* Doesn't use obd_ioctl_data */
1927 case OBD_IOC_CHANGELOG_CLEAR: {
1928 struct changelog_setinfo *cs = karg;
1930 if (unlikely(!barrier_entry(mdd->mdd_bottom)))
1931 RETURN(-EINPROGRESS);
1933 rc = mdd_changelog_clear(env, mdd, cs->cs_id,
1935 barrier_exit(mdd->mdd_bottom);
1938 case OBD_IOC_START_LFSCK: {
1939 rc = lfsck_start(env, mdd->mdd_bottom,
1940 (struct lfsck_start_param *)karg);
1943 case OBD_IOC_STOP_LFSCK: {
1944 rc = lfsck_stop(env, mdd->mdd_bottom,
1945 (struct lfsck_stop *)karg);
1948 case OBD_IOC_QUERY_LFSCK: {
1949 rc = lfsck_query(env, mdd->mdd_bottom, NULL, NULL,
1950 (struct lfsck_query *)karg);
1955 /* Below ioctls use obd_ioctl_data */
1956 if (len != sizeof(*data)) {
1957 CERROR("Bad ioctl size %d\n", len);
1960 if (data->ioc_version != OBD_IOCTL_VERSION) {
1961 CERROR("Bad magic %x != %x\n", data->ioc_version,
1967 case OBD_IOC_CHANGELOG_REG:
1968 if (unlikely(!barrier_entry(mdd->mdd_bottom)))
1969 RETURN(-EINPROGRESS);
1971 rc = mdd_changelog_user_register(env, mdd, &data->ioc_u32_1);
1972 barrier_exit(mdd->mdd_bottom);
1974 case OBD_IOC_CHANGELOG_DEREG:
1975 if (unlikely(!barrier_entry(mdd->mdd_bottom)))
1976 RETURN(-EINPROGRESS);
1978 /* explicitly clear changelog first, to protect from crash in
1979 * the middle of purge that would lead to unregistered consumer
1980 * but pending changelog entries
1982 rc = mdd_changelog_clear(env, mdd, data->ioc_u32_1, 0);
1984 rc = mdd_changelog_user_purge(env,
1985 mdd, data->ioc_u32_1);
1987 barrier_exit(mdd->mdd_bottom);
1996 /* type constructor/destructor: mdd_type_init, mdd_type_fini */
1997 LU_TYPE_INIT_FINI(mdd, &mdd_thread_key);
1999 static const struct md_device_operations mdd_ops = {
2000 .mdo_statfs = mdd_statfs,
2001 .mdo_root_get = mdd_root_get,
2002 .mdo_llog_ctxt_get = mdd_llog_ctxt_get,
2003 .mdo_iocontrol = mdd_iocontrol,
2004 .mdo_dtconf_get = mdd_dtconf_get,
2007 static struct lu_device_type_operations mdd_device_type_ops = {
2008 .ldto_init = mdd_type_init,
2009 .ldto_fini = mdd_type_fini,
2011 .ldto_start = mdd_type_start,
2012 .ldto_stop = mdd_type_stop,
2014 .ldto_device_alloc = mdd_device_alloc,
2015 .ldto_device_free = mdd_device_free,
2017 .ldto_device_fini = mdd_device_fini
2020 static struct lu_device_type mdd_device_type = {
2021 .ldt_tags = LU_DEVICE_MD,
2022 .ldt_name = LUSTRE_MDD_NAME,
2023 .ldt_ops = &mdd_device_type_ops,
2024 .ldt_ctx_tags = LCT_MD_THREAD
2027 /* context key constructor: mdd_key_init */
2028 LU_KEY_INIT(mdd, struct mdd_thread_info);
2030 static void mdd_key_fini(const struct lu_context *ctx,
2031 struct lu_context_key *key, void *data)
2033 struct mdd_thread_info *info = data;
2035 lu_buf_free(&info->mti_big_buf);
2036 lu_buf_free(&info->mti_link_buf);
2037 lu_buf_free(&info->mti_xattr_buf);
2042 /* context key: mdd_thread_key */
2043 LU_CONTEXT_KEY_DEFINE(mdd, LCT_MD_THREAD);
2045 int mdd_generic_thread_start(struct mdd_generic_thread *thread,
2046 int (*func)(void *), void *data, char *name)
2048 struct task_struct *task;
2050 LASSERT(thread->mgt_init == false);
2051 init_completion(&thread->mgt_started);
2052 init_completion(&thread->mgt_finished);
2053 thread->mgt_data = data;
2054 thread->mgt_abort = false;
2055 thread->mgt_init = true;
2057 task = kthread_run(func, thread, name);
2059 complete(&thread->mgt_finished);
2060 return PTR_ERR(task);
2062 wait_for_completion(&thread->mgt_started);
2066 void mdd_generic_thread_stop(struct mdd_generic_thread *thread)
2068 if (thread->mgt_init == true) {
2069 thread->mgt_abort = true;
2070 wait_for_completion(&thread->mgt_finished);
2074 static int __init mdd_init(void)
2078 rc = lu_kmem_init(mdd_caches);
2082 changelog_orig_logops = llog_common_cat_ops;
2083 changelog_orig_logops.lop_write_rec = mdd_changelog_write_rec;
2085 rc = class_register_type(&mdd_obd_device_ops, NULL, false, NULL,
2086 LUSTRE_MDD_NAME, &mdd_device_type);
2088 lu_kmem_fini(mdd_caches);
2092 static void __exit mdd_exit(void)
2094 class_unregister_type(LUSTRE_MDD_NAME);
2095 lu_kmem_fini(mdd_caches);
2098 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
2099 MODULE_DESCRIPTION("Lustre Meta-data Device Driver ("LUSTRE_MDD_NAME")");
2100 MODULE_VERSION(LUSTRE_VERSION_STRING);
2101 MODULE_LICENSE("GPL");
2103 module_init(mdd_init);
2104 module_exit(mdd_exit);