1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright 2008 Sun Microsystems, Inc. All rights reserved
30 * Use is subject to license terms.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lustre/mdt/mdt_recovery.c
38 * Lustre Metadata Target (mdt) recovery-related methods
40 * Author: Huang Hua <huanghua@clusterfs.com>
41 * Author: Pershin Mike <tappro@clusterfs.com>
45 # define EXPORT_SYMTAB
47 #define DEBUG_SUBSYSTEM S_MDS
49 #include "mdt_internal.h"
51 static int mdt_server_data_update(const struct lu_env *env,
52 struct mdt_device *mdt,
55 struct lu_buf *mdt_buf(const struct lu_env *env, void *area, ssize_t len)
58 struct mdt_thread_info *mti;
60 mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
67 const struct lu_buf *mdt_buf_const(const struct lu_env *env,
68 const void *area, ssize_t len)
71 struct mdt_thread_info *mti;
73 mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
76 buf->lb_buf = (void *)area;
81 static inline int mdt_trans_credit_get(const struct lu_env *env,
82 struct mdt_device *mdt,
85 struct dt_device *dev = mdt->mdt_bottom;
88 case MDT_TXN_CAPA_KEYS_WRITE_OP:
89 case MDT_TXN_LAST_RCVD_WRITE_OP:
90 cr = dev->dd_ops->dt_credit_get(env,
100 void mdt_trans_credit_init(const struct lu_env *env,
101 struct mdt_device *mdt,
104 struct mdt_thread_info *mti;
108 mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
109 p = &mti->mti_txn_param;
111 cr = mdt_trans_credit_get(env, mdt, op);
112 txn_param_init(p, cr);
115 struct thandle* mdt_trans_start(const struct lu_env *env,
116 struct mdt_device *mdt)
118 struct mdt_thread_info *mti;
121 mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
122 p = &mti->mti_txn_param;
124 /* export can require sync operations */
125 if (mti->mti_exp != NULL)
126 p->tp_sync = mti->mti_exp->exp_need_sync;
128 return mdt->mdt_bottom->dd_ops->dt_trans_start(env, mdt->mdt_bottom, p);
131 void mdt_trans_stop(const struct lu_env *env,
132 struct mdt_device *mdt, struct thandle *th)
134 mdt->mdt_bottom->dd_ops->dt_trans_stop(env, th);
137 static inline int mdt_last_rcvd_header_read(const struct lu_env *env,
138 struct mdt_device *mdt)
140 struct mdt_thread_info *mti;
143 mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
146 rc = dt_record_read(env, mdt->mdt_last_rcvd,
147 mdt_buf(env, &mti->mti_lsd, sizeof(mti->mti_lsd)),
150 lsd_le_to_cpu(&mti->mti_lsd, &mdt->mdt_lsd);
152 CDEBUG(D_INFO, "read last_rcvd header rc = %d:\n"
154 "last_transno = "LPU64"\n",
155 rc, mdt->mdt_lsd.lsd_uuid,
156 mdt->mdt_lsd.lsd_last_transno);
160 static inline int mdt_last_rcvd_header_write(const struct lu_env *env,
161 struct mdt_device *mdt,
164 struct mdt_thread_info *mti;
169 mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
172 spin_lock(&mti->mti_exp->exp_lock);
173 mti->mti_exp->exp_need_sync = need_sync;
174 spin_unlock(&mti->mti_exp->exp_lock);
176 mdt_trans_credit_init(env, mdt, MDT_TXN_LAST_RCVD_WRITE_OP);
177 th = mdt_trans_start(env, mdt);
182 lsd_cpu_to_le(&mdt->mdt_lsd, &mti->mti_lsd);
184 if (need_sync && mti->mti_exp)
185 mdt_trans_add_cb(th, lut_cb_client, mti->mti_exp);
187 rc = dt_record_write(env, mdt->mdt_last_rcvd,
188 mdt_buf_const(env, &mti->mti_lsd,
189 sizeof(mti->mti_lsd)),
192 mdt_trans_stop(env, mdt, th);
194 CDEBUG(D_INFO, "write last_rcvd header rc = %d:\n"
195 "uuid = %s\nlast_transno = "LPU64"\n",
196 rc, mdt->mdt_lsd.lsd_uuid, mdt->mdt_lsd.lsd_last_transno);
201 static int mdt_last_rcvd_read(const struct lu_env *env,
202 struct mdt_device *mdt,
203 struct lsd_client_data *lcd, loff_t *off)
205 struct mdt_thread_info *mti;
206 struct lsd_client_data *tmp;
209 mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
211 rc = dt_record_read(env, mdt->mdt_last_rcvd,
212 mdt_buf(env, tmp, sizeof(*tmp)), off);
214 lcd_le_to_cpu(tmp, lcd);
216 CDEBUG(D_INFO, "read lcd @%d rc = %d:\n"
218 "last_transno = "LPU64"\n"
219 "last_xid = "LPU64"\n"
222 "last_close_transno = "LPU64"\n"
223 "last_close_xid = "LPU64"\n"
224 "last_close_result = %u\n",
225 (int)(*off - sizeof(*tmp)),
228 lcd->lcd_last_transno,
230 lcd->lcd_last_result,
232 lcd->lcd_last_close_transno,
233 lcd->lcd_last_close_xid,
234 lcd->lcd_last_close_result);
238 static int mdt_last_rcvd_write(const struct lu_env *env,
239 struct mdt_device *mdt,
240 struct lsd_client_data *lcd,
241 loff_t *off, struct thandle *th)
243 struct mdt_thread_info *mti;
244 struct lsd_client_data *tmp;
248 mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
251 lcd_cpu_to_le(lcd, tmp);
253 rc = dt_record_write(env, mdt->mdt_last_rcvd,
254 mdt_buf_const(env, tmp, sizeof(*tmp)), off, th);
256 CDEBUG(D_INFO, "write lcd @%d rc = %d:\n"
258 "last_transno = "LPU64"\n"
259 "last_xid = "LPU64"\n"
262 "last_close_transno = "LPU64"\n"
263 "last_close_xid = "LPU64"\n"
264 "last_close_result = %u\n",
265 (int)(*off - sizeof(*tmp)),
268 lcd->lcd_last_transno,
270 lcd->lcd_last_result,
272 lcd->lcd_last_close_transno,
273 lcd->lcd_last_close_xid,
274 lcd->lcd_last_close_result);
278 static int mdt_clients_data_init(const struct lu_env *env,
279 struct mdt_device *mdt,
280 unsigned long last_size)
282 struct lr_server_data *lsd = &mdt->mdt_lsd;
283 struct lsd_client_data *lcd = NULL;
284 struct obd_device *obd = mdt2obd_dev(mdt);
290 /* When we do a clean MDS shutdown, we save the last_transno into
291 * the header. If we find clients with higher last_transno values
292 * then those clients may need recovery done. */
293 LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0);
294 for (cl_idx = 0, off = lsd->lsd_client_start;
295 off < last_size; cl_idx++) {
297 struct obd_export *exp;
305 off = lsd->lsd_client_start +
306 cl_idx * lsd->lsd_client_size;
308 rc = mdt_last_rcvd_read(env, mdt, lcd, &off);
310 CERROR("error reading MDS %s idx %d, off %llu: rc %d\n",
311 LAST_RCVD, cl_idx, off, rc);
313 break; /* read error shouldn't cause startup to fail */
316 if (lcd->lcd_uuid[0] == '\0') {
317 CDEBUG(D_INFO, "skipping zeroed client at offset %d\n",
322 last_transno = lcd_last_transno(lcd);
324 /* These exports are cleaned up by mdt_obd_disconnect(), so
325 * they need to be set up like real exports as
326 * mdt_obd_connect() does.
328 CDEBUG(D_HA, "RCVRNG CLIENT uuid: %s idx: %d lr: "LPU64
329 " srv lr: "LPU64" lx: "LPU64"\n", lcd->lcd_uuid, cl_idx,
330 last_transno, lsd->lsd_last_transno,
333 exp = class_new_export(obd, (struct obd_uuid *)lcd->lcd_uuid);
335 if (PTR_ERR(exp) == -EALREADY) {
336 /* export already exists, zero out this one */
337 lcd->lcd_uuid[0] = '\0';
339 GOTO(err_client, rc = PTR_ERR(exp));
342 struct mdt_thread_info *mti;
343 mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
344 LASSERT(mti != NULL);
346 exp->exp_mdt_data.med_lcd = lcd;
347 rc = mdt_client_add(env, mdt, cl_idx);
348 /* can't fail existing */
349 LASSERTF(rc == 0, "rc = %d\n", rc);
350 /* VBR: set export last committed version */
351 exp->exp_last_committed = last_transno;
353 spin_lock(&exp->exp_lock);
354 exp->exp_connecting = 0;
355 exp->exp_in_recovery = 0;
356 spin_unlock(&exp->exp_lock);
357 obd->obd_max_recoverable_clients++;
358 class_export_put(exp);
361 CDEBUG(D_OTHER, "client at idx %d has last_transno="LPU64"\n",
362 cl_idx, last_transno);
363 /* protect __u64 value update */
364 spin_lock(&mdt->mdt_transno_lock);
365 mdt->mdt_last_transno = max(last_transno,
366 mdt->mdt_last_transno);
367 spin_unlock(&mdt->mdt_transno_lock);
376 static int mdt_server_data_init(const struct lu_env *env,
377 struct mdt_device *mdt,
378 struct lustre_sb_info *lsi)
380 struct lr_server_data *lsd = &mdt->mdt_lsd;
381 struct lsd_client_data *lcd = NULL;
382 struct obd_device *obd = mdt2obd_dev(mdt);
383 struct mdt_thread_info *mti;
384 struct dt_object *obj;
386 struct lustre_disk_data *ldd;
387 unsigned long last_rcvd_size;
392 /* ensure padding in the struct is the correct size */
393 CLASSERT(offsetof(struct lr_server_data, lsd_padding) +
394 sizeof(lsd->lsd_padding) == LR_SERVER_SIZE);
395 CLASSERT(offsetof(struct lsd_client_data, lcd_padding) +
396 sizeof(lcd->lcd_padding) == LR_CLIENT_SIZE);
398 mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
399 LASSERT(mti != NULL);
400 la = &mti->mti_attr.ma_attr;
402 obj = mdt->mdt_last_rcvd;
403 rc = obj->do_ops->do_attr_get(env, mdt->mdt_last_rcvd, la, BYPASS_CAPA);
407 last_rcvd_size = (unsigned long)la->la_size;
409 if (last_rcvd_size == 0) {
410 LCONSOLE_WARN("%s: new disk, initializing\n", obd->obd_name);
412 memcpy(lsd->lsd_uuid, obd->obd_uuid.uuid,
413 sizeof(lsd->lsd_uuid));
414 lsd->lsd_last_transno = 0;
415 lsd->lsd_mount_count = 0;
416 lsd->lsd_server_size = LR_SERVER_SIZE;
417 lsd->lsd_client_start = LR_CLIENT_START;
418 lsd->lsd_client_size = LR_CLIENT_SIZE;
419 lsd->lsd_feature_compat = OBD_COMPAT_MDT;
420 lsd->lsd_feature_rocompat = OBD_ROCOMPAT_LOVOBJID;
421 lsd->lsd_feature_incompat = OBD_INCOMPAT_MDT |
422 OBD_INCOMPAT_COMMON_LR;
424 LCONSOLE_WARN("%s: used disk, loading\n", obd->obd_name);
425 rc = mdt_last_rcvd_header_read(env, mdt);
427 CERROR("error reading MDS %s: rc %d\n", LAST_RCVD, rc);
430 if (strcmp(lsd->lsd_uuid, obd->obd_uuid.uuid) != 0) {
431 LCONSOLE_ERROR_MSG(0x157, "Trying to start OBD %s using"
432 "the wrong disk %s. Were the /dev/ "
433 "assignments rearranged?\n",
434 obd->obd_uuid.uuid, lsd->lsd_uuid);
435 GOTO(out, rc = -EINVAL);
437 lsd->lsd_feature_compat |= OBD_COMPAT_MDT;
438 lsd->lsd_feature_incompat |= OBD_INCOMPAT_MDT |
439 OBD_INCOMPAT_COMMON_LR;
441 mount_count = lsd->lsd_mount_count;
445 if (lsd->lsd_feature_incompat & ~MDT_INCOMPAT_SUPP) {
446 CERROR("%s: unsupported incompat filesystem feature(s) %x\n",
448 lsd->lsd_feature_incompat & ~MDT_INCOMPAT_SUPP);
449 GOTO(out, rc = -EINVAL);
451 if (lsd->lsd_feature_rocompat & ~MDT_ROCOMPAT_SUPP) {
452 CERROR("%s: unsupported read-only filesystem feature(s) %x\n",
454 lsd->lsd_feature_rocompat & ~MDT_ROCOMPAT_SUPP);
455 /* XXX: Do something like remount filesystem read-only */
456 GOTO(out, rc = -EINVAL);
458 /** Interop: evict all clients at first boot with 1.8 last_rcvd */
459 if (!(lsd->lsd_feature_compat & OBD_COMPAT_20)) {
460 LCONSOLE_WARN("Mounting %s at first time on 1.8 FS, remove all"
461 " clients for interop needs\n", obd->obd_name);
462 simple_truncate(lsi->lsi_srv_mnt->mnt_sb->s_root,
463 lsi->lsi_srv_mnt, LAST_RCVD,
464 lsd->lsd_client_start);
465 last_rcvd_size = lsd->lsd_client_start;
466 /** set 2.0 flag to upgrade/downgrade between 1.8 and 2.0 */
467 lsd->lsd_feature_compat |= OBD_COMPAT_20;
470 if (ldd->ldd_flags & LDD_F_IAM_DIR)
471 lsd->lsd_feature_incompat |= OBD_INCOMPAT_IAM_DIR;
473 lsd->lsd_feature_incompat |= OBD_INCOMPAT_FID;
475 spin_lock(&mdt->mdt_transno_lock);
476 mdt->mdt_last_transno = lsd->lsd_last_transno;
477 spin_unlock(&mdt->mdt_transno_lock);
479 CDEBUG(D_INODE, "========BEGIN DUMPING LAST_RCVD========\n");
480 CDEBUG(D_INODE, "%s: server last_transno: "LPU64"\n",
481 obd->obd_name, mdt->mdt_last_transno);
482 CDEBUG(D_INODE, "%s: server mount_count: "LPU64"\n",
483 obd->obd_name, mount_count + 1);
484 CDEBUG(D_INODE, "%s: server data size: %u\n",
485 obd->obd_name, lsd->lsd_server_size);
486 CDEBUG(D_INODE, "%s: per-client data start: %u\n",
487 obd->obd_name, lsd->lsd_client_start);
488 CDEBUG(D_INODE, "%s: per-client data size: %u\n",
489 obd->obd_name, lsd->lsd_client_size);
490 CDEBUG(D_INODE, "%s: last_rcvd size: %lu\n",
491 obd->obd_name, last_rcvd_size);
492 CDEBUG(D_INODE, "%s: last_rcvd clients: %lu\n", obd->obd_name,
493 last_rcvd_size <= lsd->lsd_client_start ? 0 :
494 (last_rcvd_size - lsd->lsd_client_start) /
495 lsd->lsd_client_size);
496 CDEBUG(D_INODE, "========END DUMPING LAST_RCVD========\n");
498 if (!lsd->lsd_server_size || !lsd->lsd_client_start ||
499 !lsd->lsd_client_size) {
500 CERROR("Bad last_rcvd contents!\n");
501 GOTO(out, rc = -EINVAL);
504 rc = mdt_clients_data_init(env, mdt, last_rcvd_size);
506 GOTO(err_client, rc);
508 spin_lock(&mdt->mdt_transno_lock);
509 /* obd_last_committed is used for compatibility
510 * with other lustre recovery code */
511 obd->obd_last_committed = mdt->mdt_last_transno;
512 spin_unlock(&mdt->mdt_transno_lock);
514 mdt->mdt_mount_count = mount_count + 1;
515 lsd->lsd_mount_count = mdt->mdt_mount_count;
517 /* save it, so mount count and last_transno is current */
518 rc = mdt_server_data_update(env, mdt, (mti->mti_exp &&
519 mti->mti_exp->exp_need_sync));
521 GOTO(err_client, rc);
526 class_disconnect_exports(obd);
531 static int mdt_server_data_update(const struct lu_env *env,
532 struct mdt_device *mdt,
538 CDEBUG(D_SUPER, "MDS mount_count is "LPU64", last_transno is "LPU64"\n",
539 mdt->mdt_mount_count, mdt->mdt_last_transno);
541 spin_lock(&mdt->mdt_transno_lock);
542 mdt->mdt_lsd.lsd_last_transno = mdt->mdt_last_transno;
543 spin_unlock(&mdt->mdt_transno_lock);
546 * This may be called from difficult reply handler and
547 * mdt->mdt_last_rcvd may be NULL that time.
549 if (mdt->mdt_last_rcvd != NULL)
550 rc = mdt_last_rcvd_header_write(env, mdt, need_sync);
554 int mdt_client_new(const struct lu_env *env, struct mdt_device *mdt)
556 unsigned long *bitmap = mdt->mdt_client_bitmap;
557 struct mdt_thread_info *mti;
558 struct mdt_export_data *med;
559 struct lsd_client_data *lcd;
560 struct lr_server_data *lsd = &mdt->mdt_lsd;
561 struct obd_device *obd = mdt2obd_dev(mdt);
568 mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
569 LASSERT(mti != NULL);
571 med = &mti->mti_exp->exp_mdt_data;
574 LASSERT(bitmap != NULL);
575 if (!strcmp(med->med_lcd->lcd_uuid, obd->obd_uuid.uuid))
578 /* the bitmap operations can handle cl_idx > sizeof(long) * 8, so
579 * there's no need for extra complication here
581 spin_lock(&mdt->mdt_client_bitmap_lock);
582 cl_idx = find_first_zero_bit(bitmap, LR_MAX_CLIENTS);
583 if (cl_idx >= LR_MAX_CLIENTS ||
584 OBD_FAIL_CHECK(OBD_FAIL_MDS_CLIENT_ADD)) {
585 CERROR("no room for %u clients - fix LR_MAX_CLIENTS\n",
587 spin_unlock(&mdt->mdt_client_bitmap_lock);
590 set_bit(cl_idx, bitmap);
591 spin_unlock(&mdt->mdt_client_bitmap_lock);
593 CDEBUG(D_INFO, "client at idx %d with UUID '%s' added\n",
594 cl_idx, med->med_lcd->lcd_uuid);
596 med->med_lr_idx = cl_idx;
597 med->med_lr_off = lsd->lsd_client_start +
598 (cl_idx * lsd->lsd_client_size);
599 init_mutex(&med->med_lcd_lock);
601 LASSERTF(med->med_lr_off > 0, "med_lr_off = %llu\n", med->med_lr_off);
603 /* Write new client data. */
604 off = med->med_lr_off;
605 mdt_trans_credit_init(env, mdt, MDT_TXN_LAST_RCVD_WRITE_OP);
607 th = mdt_trans_start(env, mdt);
612 * Until this operations will be committed the sync is needed
613 * for this export. This should be done _after_ starting the
614 * transaction so that many connecting clients will not bring
615 * server down with lots of sync writes.
617 mdt_trans_add_cb(th, lut_cb_client, mti->mti_exp);
618 spin_lock(&mti->mti_exp->exp_lock);
619 mti->mti_exp->exp_need_sync = 1;
620 spin_unlock(&mti->mti_exp->exp_lock);
622 rc = mdt_last_rcvd_write(env, mdt, lcd, &off, th);
623 CDEBUG(D_INFO, "wrote client lcd at idx %u off %llu (len %u)\n",
624 cl_idx, med->med_lr_off, (int)sizeof(*lcd));
625 mdt_trans_stop(env, mdt, th);
630 /* Add client data to the MDS. We use a bitmap to locate a free space
631 * in the last_rcvd file if cl_off is -1 (i.e. a new client).
632 * Otherwise, we just have to read the data from the last_rcvd file and
633 * we know its offset.
635 * It should not be possible to fail adding an existing client - otherwise
636 * mdt_init_server_data() callsite needs to be fixed.
638 int mdt_client_add(const struct lu_env *env,
639 struct mdt_device *mdt, int cl_idx)
641 struct mdt_thread_info *mti;
642 struct mdt_export_data *med;
643 unsigned long *bitmap = mdt->mdt_client_bitmap;
644 struct obd_device *obd = mdt2obd_dev(mdt);
645 struct lr_server_data *lsd = &mdt->mdt_lsd;
649 mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
650 LASSERT(mti != NULL);
652 med = &mti->mti_exp->exp_mdt_data;
654 LASSERT(bitmap != NULL);
655 LASSERTF(cl_idx >= 0, "%d\n", cl_idx);
657 if (!strcmp(med->med_lcd->lcd_uuid, obd->obd_uuid.uuid))
660 spin_lock(&mdt->mdt_client_bitmap_lock);
661 if (test_and_set_bit(cl_idx, bitmap)) {
662 CERROR("MDS client %d: bit already set in bitmap!!\n",
666 spin_unlock(&mdt->mdt_client_bitmap_lock);
668 CDEBUG(D_INFO, "client at idx %d with UUID '%s' added\n",
669 cl_idx, med->med_lcd->lcd_uuid);
671 med->med_lr_idx = cl_idx;
672 med->med_lr_off = lsd->lsd_client_start +
673 (cl_idx * lsd->lsd_client_size);
674 init_mutex(&med->med_lcd_lock);
676 LASSERTF(med->med_lr_off > 0, "med_lr_off = %llu\n", med->med_lr_off);
681 int mdt_client_del(const struct lu_env *env, struct mdt_device *mdt)
683 struct mdt_thread_info *mti;
684 struct mdt_export_data *med;
685 struct lsd_client_data *lcd;
686 struct obd_device *obd = mdt2obd_dev(mdt);
687 struct obd_export *exp;
694 mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
695 LASSERT(mti != NULL);
698 med = &exp->exp_mdt_data;
703 /* XXX: If lcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
704 if (!strcmp(lcd->lcd_uuid, obd->obd_uuid.uuid))
707 CDEBUG(D_INFO, "freeing client at idx %u, offset %lld\n",
708 med->med_lr_idx, med->med_lr_off);
710 off = med->med_lr_off;
713 * Don't clear med_lr_idx here as it is likely also unset. At worst we
714 * leak a client slot that will be cleaned on the next recovery.
717 CERROR("client idx %d has offset %lld\n",
718 med->med_lr_idx, off);
719 GOTO(free, rc = -EINVAL);
723 * Clear the bit _after_ zeroing out the client so we don't race with
724 * mdt_client_add and zero out new clients.
726 if (!test_bit(med->med_lr_idx, mdt->mdt_client_bitmap)) {
727 CERROR("MDT client %u: bit already clear in bitmap!!\n",
732 /* Don't force sync on disconnect if aborting recovery,
733 * or it does num_clients * num_osts. b=17194 */
734 need_sync = (!exp->exp_libclient || exp->exp_need_sync) &&
735 !(exp->exp_flags & OBD_OPT_ABORT_RECOV);
738 * This may be called from difficult reply handler path and
739 * mdt->mdt_last_rcvd may be NULL that time.
741 if (mdt->mdt_last_rcvd != NULL) {
742 mdt_trans_credit_init(env, mdt, MDT_TXN_LAST_RCVD_WRITE_OP);
744 spin_lock(&exp->exp_lock);
745 exp->exp_need_sync = need_sync;
746 spin_unlock(&exp->exp_lock);
748 th = mdt_trans_start(env, mdt);
750 GOTO(free, rc = PTR_ERR(th));
754 * Until this operations will be committed the sync
755 * is needed for this export.
757 mdt_trans_add_cb(th, lut_cb_client, exp);
760 mutex_down(&med->med_lcd_lock);
761 memset(lcd, 0, sizeof *lcd);
763 rc = mdt_last_rcvd_write(env, mdt, lcd, &off, th);
764 mutex_up(&med->med_lcd_lock);
765 mdt_trans_stop(env, mdt, th);
768 CDEBUG(rc == 0 ? D_INFO : D_ERROR, "Zeroing out client idx %u in "
769 "%s %ssync rc %d\n", med->med_lr_idx, LAST_RCVD,
770 need_sync ? "" : "a", rc);
772 spin_lock(&mdt->mdt_client_bitmap_lock);
773 clear_bit(med->med_lr_idx, mdt->mdt_client_bitmap);
774 spin_unlock(&mdt->mdt_client_bitmap_lock);
777 * Make sure the server's last_transno is up to date. Do this
778 * after the client is freed so we know all the client's
779 * transactions have been committed.
781 mdt_server_data_update(env, mdt, need_sync);
791 * last_rcvd & last_committed update callbacks
793 static int mdt_last_rcvd_update(struct mdt_thread_info *mti,
796 struct mdt_device *mdt = mti->mti_mdt;
797 struct ptlrpc_request *req = mdt_info_req(mti);
798 struct mdt_export_data *med;
799 struct lsd_client_data *lcd;
802 __s32 rc = th->th_result;
806 LASSERT(req->rq_export);
808 med = &req->rq_export->exp_mdt_data;
811 /* if the export has already been failed, we have no last_rcvd slot */
812 if (req->rq_export->exp_failed) {
813 CWARN("commit transaction for disconnected client %s: rc %d\n",
814 req->rq_export->exp_client_uuid.uuid, rc);
820 off = med->med_lr_off;
821 LASSERT(ergo(mti->mti_transno == 0, rc != 0));
822 mutex_down(&med->med_lcd_lock);
823 if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE ||
824 lustre_msg_get_opc(req->rq_reqmsg) == MDS_DONE_WRITING) {
825 if (mti->mti_transno != 0)
826 lcd->lcd_last_close_transno = mti->mti_transno;
827 lcd->lcd_last_close_xid = req->rq_xid;
828 lcd->lcd_last_close_result = rc;
830 /* VBR: save versions in last_rcvd for reconstruct. */
831 __u64 *pre_versions = lustre_msg_get_versions(req->rq_repmsg);
833 lcd->lcd_pre_versions[0] = pre_versions[0];
834 lcd->lcd_pre_versions[1] = pre_versions[1];
835 lcd->lcd_pre_versions[2] = pre_versions[2];
836 lcd->lcd_pre_versions[3] = pre_versions[3];
838 if (mti->mti_transno != 0)
839 lcd->lcd_last_transno = mti->mti_transno;
840 lcd->lcd_last_xid = req->rq_xid;
841 lcd->lcd_last_result = rc;
842 /*XXX: save intent_disposition in mdt_thread_info?
843 * also there is bug - intent_dispostion is __u64,
844 * see struct ldlm_reply->lock_policy_res1; */
845 lcd->lcd_last_data = mti->mti_opdata;
849 CERROR("client idx %d has offset %lld\n", med->med_lr_idx, off);
852 err = mdt_last_rcvd_write(mti->mti_env, mdt, lcd, &off, th);
854 mutex_up(&med->med_lcd_lock);
858 extern struct lu_context_key mdt_thread_key;
860 /* add credits for last_rcvd update */
861 static int mdt_txn_start_cb(const struct lu_env *env,
862 struct txn_param *param, void *cookie)
864 struct mdt_device *mdt = cookie;
866 param->tp_credits += mdt_trans_credit_get(env, mdt,
867 MDT_TXN_LAST_RCVD_WRITE_OP);
871 /* Set new object versions */
872 static void mdt_versions_set(struct mdt_thread_info *info)
875 for (i = 0; i < PTLRPC_NUM_VERSIONS; i++)
876 if (info->mti_mos[i] != NULL)
877 mo_version_set(info->mti_env,
878 mdt_object_child(info->mti_mos[i]),
882 /* Update last_rcvd records with latests transaction data */
883 static int mdt_txn_stop_cb(const struct lu_env *env,
884 struct thandle *txn, void *cookie)
886 struct mdt_device *mdt = cookie;
887 struct mdt_txn_info *txi;
888 struct mdt_thread_info *mti;
889 struct ptlrpc_request *req;
891 /* transno in two contexts - for commit_cb and for thread */
892 txi = lu_context_key_get(&txn->th_ctx, &mdt_txn_key);
893 mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
894 req = mdt_info_req(mti);
896 if (mti->mti_mdt == NULL || req == NULL || mti->mti_no_need_trans) {
897 txi->txi_transno = 0;
898 mti->mti_no_need_trans = 0;
902 if (mti->mti_has_trans) {
903 /* XXX: currently there are allowed cases, but the wrong cases
904 * are also possible, so better check is needed here */
905 CDEBUG(D_INFO, "More than one transaction "LPU64"\n",
910 mti->mti_has_trans = 1;
911 spin_lock(&mdt->mdt_transno_lock);
912 if (txn->th_result != 0) {
913 if (mti->mti_transno != 0) {
914 CERROR("Replay transno "LPU64" failed: rc %i\n",
915 mti->mti_transno, txn->th_result);
917 } else if (mti->mti_transno == 0) {
918 mti->mti_transno = ++ mdt->mdt_last_transno;
920 /* should be replay */
921 if (mti->mti_transno > mdt->mdt_last_transno)
922 mdt->mdt_last_transno = mti->mti_transno;
924 spin_unlock(&mdt->mdt_transno_lock);
925 /* sometimes the reply message has not been successfully packed */
926 LASSERT(req != NULL && req->rq_repmsg != NULL);
928 /** VBR: set new versions */
929 if (txn->th_result == 0)
930 mdt_versions_set(mti);
932 /* filling reply data */
933 CDEBUG(D_INODE, "transno = %llu, last_committed = %llu\n",
934 mti->mti_transno, req->rq_export->exp_obd->obd_last_committed);
936 req->rq_transno = mti->mti_transno;
937 lustre_msg_set_transno(req->rq_repmsg, mti->mti_transno);
938 lustre_msg_set_last_xid(req->rq_repmsg,
939 lcd_last_xid(req->rq_export->exp_mdt_data.med_lcd));
940 /* save transno for the commit callback */
941 txi->txi_transno = mti->mti_transno;
943 /* add separate commit callback for transaction handling because we need
944 * export as parameter */
945 mdt_trans_add_cb(txn, lut_cb_last_committed,
946 class_export_cb_get(mti->mti_exp));
948 return mdt_last_rcvd_update(mti, txn);
951 /* commit callback, need to update last_committed value */
952 static int mdt_txn_commit_cb(const struct lu_env *env,
953 struct thandle *txn, void *cookie)
955 struct mdt_device *mdt = cookie;
956 struct mdt_txn_info *txi;
959 txi = lu_context_key_get(&txn->th_ctx, &mdt_txn_key);
961 /* iterate through all additional callbacks */
962 for (i = 0; i < txi->txi_cb_count; i++) {
963 txi->txi_cb[i].lut_cb_func(&mdt->mdt_lut, txi->txi_transno,
964 txi->txi_cb[i].lut_cb_data, 0);
969 int mdt_fs_setup(const struct lu_env *env, struct mdt_device *mdt,
970 struct obd_device *obd,
971 struct lustre_sb_info *lsi)
978 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_FS_SETUP))
981 /* prepare transactions callbacks */
982 mdt->mdt_txn_cb.dtc_txn_start = mdt_txn_start_cb;
983 mdt->mdt_txn_cb.dtc_txn_stop = mdt_txn_stop_cb;
984 mdt->mdt_txn_cb.dtc_txn_commit = mdt_txn_commit_cb;
985 mdt->mdt_txn_cb.dtc_cookie = mdt;
986 mdt->mdt_txn_cb.dtc_tag = LCT_MD_THREAD;
987 CFS_INIT_LIST_HEAD(&mdt->mdt_txn_cb.dtc_linkage);
989 dt_txn_callback_add(mdt->mdt_bottom, &mdt->mdt_txn_cb);
991 rc = mdt_server_data_init(env, mdt, lsi);
995 o = dt_store_open(env, mdt->mdt_bottom, "", CAPA_KEYS, &fid);
998 rc = mdt_capa_keys_init(env, mdt);
1000 GOTO(put_ck_object, rc);
1003 CERROR("cannot open %s: rc = %d\n", CAPA_KEYS, rc);
1004 GOTO(disconnect_exports, rc);
1009 lu_object_put(env, &o->do_lu);
1010 mdt->mdt_ck_obj = NULL;
1012 class_disconnect_exports(obd);
1016 void mdt_fs_cleanup(const struct lu_env *env, struct mdt_device *mdt)
1020 /* Remove transaction callback */
1021 dt_txn_callback_del(mdt->mdt_bottom, &mdt->mdt_txn_cb);
1022 if (mdt->mdt_ck_obj)
1023 lu_object_put(env, &mdt->mdt_ck_obj->do_lu);
1024 mdt->mdt_ck_obj = NULL;
1028 /* reconstruction code */
1029 static void mdt_steal_ack_locks(struct ptlrpc_request *req)
1031 struct obd_export *exp = req->rq_export;
1032 struct list_head *tmp;
1033 struct ptlrpc_reply_state *oldrep;
1034 struct ptlrpc_service *svc;
1037 /* CAVEAT EMPTOR: spinlock order */
1038 spin_lock(&exp->exp_lock);
1039 list_for_each (tmp, &exp->exp_outstanding_replies) {
1040 oldrep = list_entry(tmp, struct ptlrpc_reply_state,rs_exp_list);
1042 if (oldrep->rs_xid != req->rq_xid)
1045 if (oldrep->rs_opc != lustre_msg_get_opc(req->rq_reqmsg))
1046 CERROR ("Resent req xid "LPU64" has mismatched opc: "
1047 "new %d old %d\n", req->rq_xid,
1048 lustre_msg_get_opc(req->rq_reqmsg),
1051 svc = oldrep->rs_service;
1052 spin_lock (&svc->srv_lock);
1054 list_del_init (&oldrep->rs_exp_list);
1056 CWARN("Stealing %d locks from rs %p x"LPD64".t"LPD64
1058 oldrep->rs_nlocks, oldrep,
1059 oldrep->rs_xid, oldrep->rs_transno, oldrep->rs_opc,
1060 libcfs_nid2str(exp->exp_connection->c_peer.nid));
1062 for (i = 0; i < oldrep->rs_nlocks; i++)
1063 ptlrpc_save_lock(req, &oldrep->rs_locks[i],
1064 oldrep->rs_modes[i], 0);
1065 oldrep->rs_nlocks = 0;
1067 DEBUG_REQ(D_HA, req, "stole locks for");
1068 spin_lock(&oldrep->rs_lock);
1069 ptlrpc_schedule_difficult_reply (oldrep);
1070 spin_unlock(&oldrep->rs_lock);
1072 spin_unlock (&svc->srv_lock);
1075 spin_unlock(&exp->exp_lock);
1079 * VBR: restore versions
1081 void mdt_vbr_reconstruct(struct ptlrpc_request *req,
1082 struct lsd_client_data *lcd)
1084 __u64 pre_versions[4] = {0};
1085 pre_versions[0] = lcd->lcd_pre_versions[0];
1086 pre_versions[1] = lcd->lcd_pre_versions[1];
1087 pre_versions[2] = lcd->lcd_pre_versions[2];
1088 pre_versions[3] = lcd->lcd_pre_versions[3];
1089 lustre_msg_set_versions(req->rq_repmsg, pre_versions);
1092 void mdt_req_from_lcd(struct ptlrpc_request *req,
1093 struct lsd_client_data *lcd)
1095 DEBUG_REQ(D_HA, req, "restoring transno "LPD64"/status %d",
1096 lcd->lcd_last_transno, lcd->lcd_last_result);
1098 if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE ||
1099 lustre_msg_get_opc(req->rq_repmsg) == MDS_DONE_WRITING) {
1100 req->rq_transno = lcd->lcd_last_close_transno;
1101 req->rq_status = lcd->lcd_last_close_result;
1103 req->rq_transno = lcd->lcd_last_transno;
1104 req->rq_status = lcd->lcd_last_result;
1105 mdt_vbr_reconstruct(req, lcd);
1107 if (req->rq_status != 0)
1108 req->rq_transno = 0;
1109 lustre_msg_set_transno(req->rq_repmsg, req->rq_transno);
1110 lustre_msg_set_status(req->rq_repmsg, req->rq_status);
1111 DEBUG_REQ(D_RPCTRACE, req, "restoring transno "LPD64"/status %d",
1112 req->rq_transno, req->rq_status);
1114 mdt_steal_ack_locks(req);
1117 void mdt_reconstruct_generic(struct mdt_thread_info *mti,
1118 struct mdt_lock_handle *lhc)
1120 struct ptlrpc_request *req = mdt_info_req(mti);
1121 struct mdt_export_data *med = &req->rq_export->exp_mdt_data;
1123 return mdt_req_from_lcd(req, med->med_lcd);
1126 static void mdt_reconstruct_create(struct mdt_thread_info *mti,
1127 struct mdt_lock_handle *lhc)
1129 struct ptlrpc_request *req = mdt_info_req(mti);
1130 struct obd_export *exp = req->rq_export;
1131 struct mdt_export_data *med = &exp->exp_mdt_data;
1132 struct mdt_device *mdt = mti->mti_mdt;
1133 struct mdt_object *child;
1134 struct mdt_body *body;
1137 mdt_req_from_lcd(req, med->med_lcd);
1141 /* if no error, so child was created with requested fid */
1142 child = mdt_object_find(mti->mti_env, mdt, mti->mti_rr.rr_fid2);
1143 if (IS_ERR(child)) {
1144 rc = PTR_ERR(child);
1145 LCONSOLE_WARN("Child "DFID" lookup error %d."
1146 " Evicting client %s with export %s.\n",
1147 PFID(mdt_object_fid(child)), rc,
1148 obd_uuid2str(&exp->exp_client_uuid),
1149 obd_export_nid2str(exp));
1150 mdt_export_evict(exp);
1155 body = req_capsule_server_get(mti->mti_pill, &RMF_MDT_BODY);
1156 mti->mti_attr.ma_need = MA_INODE;
1157 mti->mti_attr.ma_valid = 0;
1158 rc = mo_attr_get(mti->mti_env, mdt_object_child(child), &mti->mti_attr);
1159 if (rc == -EREMOTE) {
1160 /* object was created on remote server */
1161 req->rq_status = rc;
1162 body->valid |= OBD_MD_MDS;
1164 mdt_pack_attr2body(mti, body, &mti->mti_attr.ma_attr,
1165 mdt_object_fid(child));
1166 mdt_object_put(mti->mti_env, child);
1169 static void mdt_reconstruct_setattr(struct mdt_thread_info *mti,
1170 struct mdt_lock_handle *lhc)
1172 struct ptlrpc_request *req = mdt_info_req(mti);
1173 struct obd_export *exp = req->rq_export;
1174 struct mdt_export_data *med = &exp->exp_mdt_data;
1175 struct mdt_device *mdt = mti->mti_mdt;
1176 struct mdt_object *obj;
1177 struct mdt_body *body;
1179 mdt_req_from_lcd(req, med->med_lcd);
1183 body = req_capsule_server_get(mti->mti_pill, &RMF_MDT_BODY);
1184 obj = mdt_object_find(mti->mti_env, mdt, mti->mti_rr.rr_fid1);
1186 int rc = PTR_ERR(obj);
1187 LCONSOLE_WARN(""DFID" lookup error %d."
1188 " Evicting client %s with export %s.\n",
1189 PFID(mdt_object_fid(obj)), rc,
1190 obd_uuid2str(&exp->exp_client_uuid),
1191 obd_export_nid2str(exp));
1192 mdt_export_evict(exp);
1196 mti->mti_attr.ma_need = MA_INODE;
1197 mti->mti_attr.ma_valid = 0;
1198 mo_attr_get(mti->mti_env, mdt_object_child(obj), &mti->mti_attr);
1199 mdt_pack_attr2body(mti, body, &mti->mti_attr.ma_attr,
1200 mdt_object_fid(obj));
1201 if (mti->mti_epoch && (mti->mti_epoch->flags & MF_EPOCH_OPEN)) {
1202 struct mdt_file_data *mfd;
1203 struct mdt_body *repbody;
1205 repbody = req_capsule_server_get(mti->mti_pill, &RMF_MDT_BODY);
1206 repbody->ioepoch = obj->mot_ioepoch;
1207 spin_lock(&med->med_open_lock);
1208 list_for_each_entry(mfd, &med->med_open_head, mfd_list) {
1209 if (mfd->mfd_xid == req->rq_xid)
1212 LASSERT(&mfd->mfd_list != &med->med_open_head);
1213 spin_unlock(&med->med_open_lock);
1214 repbody->handle.cookie = mfd->mfd_handle.h_cookie;
1217 mdt_object_put(mti->mti_env, obj);
1220 typedef void (*mdt_reconstructor)(struct mdt_thread_info *mti,
1221 struct mdt_lock_handle *lhc);
1223 static mdt_reconstructor reconstructors[REINT_MAX] = {
1224 [REINT_SETATTR] = mdt_reconstruct_setattr,
1225 [REINT_CREATE] = mdt_reconstruct_create,
1226 [REINT_LINK] = mdt_reconstruct_generic,
1227 [REINT_UNLINK] = mdt_reconstruct_generic,
1228 [REINT_RENAME] = mdt_reconstruct_generic,
1229 [REINT_OPEN] = mdt_reconstruct_open,
1230 [REINT_SETXATTR] = mdt_reconstruct_generic
1233 void mdt_reconstruct(struct mdt_thread_info *mti,
1234 struct mdt_lock_handle *lhc)
1237 reconstructors[mti->mti_rr.rr_opcode](mti, lhc);