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);
54 struct lu_buf *mdt_buf(const struct lu_env *env, void *area, ssize_t len)
57 struct mdt_thread_info *mti;
59 mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
66 const struct lu_buf *mdt_buf_const(const struct lu_env *env,
67 const void *area, ssize_t len)
70 struct mdt_thread_info *mti;
72 mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
75 buf->lb_buf = (void *)area;
80 static inline int mdt_trans_credit_get(const struct lu_env *env,
81 struct mdt_device *mdt,
84 struct dt_device *dev = mdt->mdt_bottom;
87 case MDT_TXN_CAPA_KEYS_WRITE_OP:
88 case MDT_TXN_LAST_RCVD_WRITE_OP:
89 cr = dev->dd_ops->dt_credit_get(env,
99 void mdt_trans_credit_init(const struct lu_env *env,
100 struct mdt_device *mdt,
103 struct mdt_thread_info *mti;
107 mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
108 p = &mti->mti_txn_param;
110 cr = mdt_trans_credit_get(env, mdt, op);
111 txn_param_init(p, cr);
114 struct thandle* mdt_trans_start(const struct lu_env *env,
115 struct mdt_device *mdt)
117 struct mdt_thread_info *mti;
120 mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
121 p = &mti->mti_txn_param;
123 /* export can require sync operations */
124 if (mti->mti_exp != NULL)
125 p->tp_sync = mti->mti_exp->exp_need_sync;
127 return mdt->mdt_bottom->dd_ops->dt_trans_start(env, mdt->mdt_bottom, p);
130 void mdt_trans_stop(const struct lu_env *env,
131 struct mdt_device *mdt, struct thandle *th)
133 mdt->mdt_bottom->dd_ops->dt_trans_stop(env, th);
136 static inline int mdt_last_rcvd_header_read(const struct lu_env *env,
137 struct mdt_device *mdt)
139 struct mdt_thread_info *mti;
142 mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
145 rc = dt_record_read(env, mdt->mdt_last_rcvd,
146 mdt_buf(env, &mti->mti_lsd, sizeof(mti->mti_lsd)),
149 lsd_le_to_cpu(&mti->mti_lsd, &mdt->mdt_lsd);
151 CDEBUG(D_INFO, "read last_rcvd header rc = %d, uuid = %s, "
152 "last_transno = "LPU64"\n", rc, mdt->mdt_lsd.lsd_uuid,
153 mdt->mdt_lsd.lsd_last_transno);
157 static int mdt_last_rcvd_header_write(const struct lu_env *env,
158 struct mdt_device *mdt,
161 struct mdt_thread_info *mti;
165 mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
168 lsd_cpu_to_le(&mdt->mdt_lsd, &mti->mti_lsd);
170 rc = dt_record_write(env, mdt->mdt_last_rcvd,
171 mdt_buf_const(env, &mti->mti_lsd,
172 sizeof(mti->mti_lsd)),
175 CDEBUG(D_INFO, "write last_rcvd header rc = %d, uuid = %s, "
176 "last_transno = "LPU64"\n", rc, mdt->mdt_lsd.lsd_uuid,
177 mdt->mdt_lsd.lsd_last_transno);
182 static int mdt_last_rcvd_read(const struct lu_env *env, struct mdt_device *mdt,
183 struct lsd_client_data *lcd, loff_t *off,
186 struct mdt_thread_info *mti;
187 struct lsd_client_data *tmp;
190 mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
192 rc = dt_record_read(env, mdt->mdt_last_rcvd,
193 mdt_buf(env, tmp, sizeof(*tmp)), off);
195 check_lcd(mdt2obd_dev(mdt)->obd_name, index, tmp);
196 lcd_le_to_cpu(tmp, lcd);
199 CDEBUG(D_INFO, "read lcd @%d rc = %d, uuid = %s, last_transno = "LPU64
200 ", last_xid = "LPU64", last_result = %u, last_data = %u, "
201 "last_close_transno = "LPU64", last_close_xid = "LPU64", "
202 "last_close_result = %u\n", (int)(*off - sizeof(*tmp)),
203 rc, lcd->lcd_uuid, lcd->lcd_last_transno, lcd->lcd_last_xid,
204 lcd->lcd_last_result, lcd->lcd_last_data,
205 lcd->lcd_last_close_transno, lcd->lcd_last_close_xid,
206 lcd->lcd_last_close_result);
210 static int mdt_last_rcvd_write(const struct lu_env *env,
211 struct mdt_device *mdt,
212 struct lsd_client_data *lcd,
213 loff_t *off, struct thandle *th)
215 struct mdt_thread_info *mti;
216 struct lsd_client_data *tmp;
220 mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
223 lcd_cpu_to_le(lcd, tmp);
225 rc = dt_record_write(env, mdt->mdt_last_rcvd,
226 mdt_buf_const(env, tmp, sizeof(*tmp)), off, th);
228 CDEBUG(D_INFO, "write lcd @%d rc = %d, uuid = %s, last_transno = "LPU64
229 ", last_xid = "LPU64", last_result = %u, last_data = %u, "
230 "last_close_transno = "LPU64", last_close_xid = "LPU64" ,"
231 "last_close_result = %u\n", (int)(*off - sizeof(*tmp)),
232 rc, lcd->lcd_uuid, lcd->lcd_last_transno, lcd->lcd_last_xid,
233 lcd->lcd_last_result, lcd->lcd_last_data,
234 lcd->lcd_last_close_transno, lcd->lcd_last_close_xid,
235 lcd->lcd_last_close_result);
239 static int mdt_clients_data_init(const struct lu_env *env,
240 struct mdt_device *mdt,
241 unsigned long last_size)
243 struct lr_server_data *lsd = &mdt->mdt_lsd;
244 struct lsd_client_data *lcd = NULL;
245 struct obd_device *obd = mdt2obd_dev(mdt);
251 /* When we do a clean MDS shutdown, we save the last_transno into
252 * the header. If we find clients with higher last_transno values
253 * then those clients may need recovery done. */
254 LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0);
255 for (cl_idx = 0, off = lsd->lsd_client_start;
256 off < last_size; cl_idx++) {
258 struct obd_export *exp;
266 off = lsd->lsd_client_start +
267 cl_idx * lsd->lsd_client_size;
269 rc = mdt_last_rcvd_read(env, mdt, lcd, &off, cl_idx);
271 CERROR("error reading MDS %s idx %d, off %llu: rc %d\n",
272 LAST_RCVD, cl_idx, off, rc);
274 break; /* read error shouldn't cause startup to fail */
277 if (lcd->lcd_uuid[0] == '\0') {
278 CDEBUG(D_INFO, "skipping zeroed client at offset %d\n",
283 last_transno = lcd_last_transno(lcd);
285 /* These exports are cleaned up by mdt_obd_disconnect(), so
286 * they need to be set up like real exports as
287 * mdt_obd_connect() does.
289 CDEBUG(D_HA, "RCVRNG CLIENT uuid: %s idx: %d lr: "LPU64
290 " srv lr: "LPU64" lx: "LPU64"\n", lcd->lcd_uuid, cl_idx,
291 last_transno, lsd->lsd_last_transno,
294 exp = class_new_export(obd, (struct obd_uuid *)lcd->lcd_uuid);
296 if (PTR_ERR(exp) == -EALREADY) {
297 /* export already exists, zero out this one */
298 lcd->lcd_uuid[0] = '\0';
300 GOTO(err_client, rc = PTR_ERR(exp));
303 struct mdt_thread_info *mti;
304 mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
305 LASSERT(mti != NULL);
307 exp->exp_mdt_data.med_lcd = lcd;
308 rc = mdt_client_add(env, mdt, cl_idx);
309 /* can't fail existing */
310 LASSERTF(rc == 0, "rc = %d\n", rc);
311 /* VBR: set export last committed version */
312 exp->exp_last_committed = last_transno;
314 spin_lock(&exp->exp_lock);
315 exp->exp_connecting = 0;
316 exp->exp_in_recovery = 0;
317 spin_unlock(&exp->exp_lock);
318 obd->obd_max_recoverable_clients++;
319 class_export_put(exp);
322 CDEBUG(D_OTHER, "client at idx %d has last_transno="LPU64"\n",
323 cl_idx, last_transno);
324 /* protect __u64 value update */
325 spin_lock(&mdt->mdt_transno_lock);
326 mdt->mdt_last_transno = max(last_transno,
327 mdt->mdt_last_transno);
328 spin_unlock(&mdt->mdt_transno_lock);
337 static int mdt_server_data_init(const struct lu_env *env,
338 struct mdt_device *mdt,
339 struct lustre_sb_info *lsi)
341 struct lr_server_data *lsd = &mdt->mdt_lsd;
342 struct lsd_client_data *lcd = NULL;
343 struct obd_device *obd = mdt2obd_dev(mdt);
344 struct mdt_thread_info *mti;
345 struct dt_object *obj;
347 struct lustre_disk_data *ldd;
348 unsigned long last_rcvd_size;
353 /* ensure padding in the struct is the correct size */
354 CLASSERT(offsetof(struct lr_server_data, lsd_padding) +
355 sizeof(lsd->lsd_padding) == LR_SERVER_SIZE);
356 CLASSERT(offsetof(struct lsd_client_data, lcd_padding) +
357 sizeof(lcd->lcd_padding) == LR_CLIENT_SIZE);
359 mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
360 LASSERT(mti != NULL);
361 la = &mti->mti_attr.ma_attr;
363 obj = mdt->mdt_last_rcvd;
364 rc = obj->do_ops->do_attr_get(env, mdt->mdt_last_rcvd, la, BYPASS_CAPA);
368 last_rcvd_size = (unsigned long)la->la_size;
370 if (last_rcvd_size == 0) {
371 LCONSOLE_WARN("%s: new disk, initializing\n", obd->obd_name);
373 memcpy(lsd->lsd_uuid, obd->obd_uuid.uuid,
374 sizeof(lsd->lsd_uuid));
375 lsd->lsd_last_transno = 0;
376 lsd->lsd_mount_count = 0;
377 lsd->lsd_server_size = LR_SERVER_SIZE;
378 lsd->lsd_client_start = LR_CLIENT_START;
379 lsd->lsd_client_size = LR_CLIENT_SIZE;
380 lsd->lsd_feature_compat = OBD_COMPAT_MDT;
381 lsd->lsd_feature_rocompat = OBD_ROCOMPAT_LOVOBJID;
382 lsd->lsd_feature_incompat = OBD_INCOMPAT_MDT |
383 OBD_INCOMPAT_COMMON_LR;
385 LCONSOLE_WARN("%s: used disk, loading\n", obd->obd_name);
386 rc = mdt_last_rcvd_header_read(env, mdt);
388 CERROR("error reading MDS %s: rc %d\n", LAST_RCVD, rc);
391 if (strcmp(lsd->lsd_uuid, obd->obd_uuid.uuid) != 0) {
392 LCONSOLE_ERROR_MSG(0x157, "Trying to start OBD %s using"
393 "the wrong disk %s. Were the /dev/ "
394 "assignments rearranged?\n",
395 obd->obd_uuid.uuid, lsd->lsd_uuid);
396 GOTO(out, rc = -EINVAL);
398 lsd->lsd_feature_compat |= OBD_COMPAT_MDT;
399 lsd->lsd_feature_incompat |= OBD_INCOMPAT_MDT |
400 OBD_INCOMPAT_COMMON_LR;
402 mount_count = lsd->lsd_mount_count;
406 if (lsd->lsd_feature_incompat & ~MDT_INCOMPAT_SUPP) {
407 CERROR("%s: unsupported incompat filesystem feature(s) %x\n",
409 lsd->lsd_feature_incompat & ~MDT_INCOMPAT_SUPP);
410 GOTO(out, rc = -EINVAL);
412 if (lsd->lsd_feature_rocompat & ~MDT_ROCOMPAT_SUPP) {
413 CERROR("%s: unsupported read-only filesystem feature(s) %x\n",
415 lsd->lsd_feature_rocompat & ~MDT_ROCOMPAT_SUPP);
416 /* XXX: Do something like remount filesystem read-only */
417 GOTO(out, rc = -EINVAL);
419 /** Interop: evict all clients at first boot with 1.8 last_rcvd */
420 if (!(lsd->lsd_feature_compat & OBD_COMPAT_20)) {
421 LCONSOLE_WARN("Mounting %s at first time on 1.8 FS, remove all"
422 " clients for interop needs\n", obd->obd_name);
423 simple_truncate(lsi->lsi_srv_mnt->mnt_sb->s_root,
424 lsi->lsi_srv_mnt, LAST_RCVD,
425 lsd->lsd_client_start);
426 last_rcvd_size = lsd->lsd_client_start;
427 /** set 2.0 flag to upgrade/downgrade between 1.8 and 2.0 */
428 lsd->lsd_feature_compat |= OBD_COMPAT_20;
431 if (ldd->ldd_flags & LDD_F_IAM_DIR)
432 lsd->lsd_feature_incompat |= OBD_INCOMPAT_IAM_DIR;
434 lsd->lsd_feature_incompat |= OBD_INCOMPAT_FID;
436 spin_lock(&mdt->mdt_transno_lock);
437 mdt->mdt_last_transno = lsd->lsd_last_transno;
438 spin_unlock(&mdt->mdt_transno_lock);
440 CDEBUG(D_INODE, "========BEGIN DUMPING LAST_RCVD========\n");
441 CDEBUG(D_INODE, "%s: server last_transno: "LPU64"\n",
442 obd->obd_name, mdt->mdt_last_transno);
443 CDEBUG(D_INODE, "%s: server mount_count: "LPU64"\n",
444 obd->obd_name, mount_count + 1);
445 CDEBUG(D_INODE, "%s: server data size: %u\n",
446 obd->obd_name, lsd->lsd_server_size);
447 CDEBUG(D_INODE, "%s: per-client data start: %u\n",
448 obd->obd_name, lsd->lsd_client_start);
449 CDEBUG(D_INODE, "%s: per-client data size: %u\n",
450 obd->obd_name, lsd->lsd_client_size);
451 CDEBUG(D_INODE, "%s: last_rcvd size: %lu\n",
452 obd->obd_name, last_rcvd_size);
453 CDEBUG(D_INODE, "%s: last_rcvd clients: %lu\n", obd->obd_name,
454 last_rcvd_size <= lsd->lsd_client_start ? 0 :
455 (last_rcvd_size - lsd->lsd_client_start) /
456 lsd->lsd_client_size);
457 CDEBUG(D_INODE, "========END DUMPING LAST_RCVD========\n");
459 if (!lsd->lsd_server_size || !lsd->lsd_client_start ||
460 !lsd->lsd_client_size) {
461 CERROR("Bad last_rcvd contents!\n");
462 GOTO(out, rc = -EINVAL);
465 rc = mdt_clients_data_init(env, mdt, last_rcvd_size);
467 GOTO(err_client, rc);
469 spin_lock(&mdt->mdt_transno_lock);
470 /* obd_last_committed is used for compatibility
471 * with other lustre recovery code */
472 obd->obd_last_committed = mdt->mdt_last_transno;
473 spin_unlock(&mdt->mdt_transno_lock);
475 mdt->mdt_mount_count = mount_count + 1;
476 lsd->lsd_mount_count = mdt->mdt_mount_count;
478 /* save it, so mount count and last_transno is current */
479 rc = mdt_server_data_update(env, mdt);
481 GOTO(err_client, rc);
486 class_disconnect_exports(obd);
491 static int mdt_server_data_update(const struct lu_env *env,
492 struct mdt_device *mdt)
494 struct mdt_thread_info *mti;
498 mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
500 mdt_trans_credit_init(env, mdt, MDT_TXN_LAST_RCVD_WRITE_OP);
501 th = mdt_trans_start(env, mdt);
505 CDEBUG(D_SUPER, "MDS mount_count is "LPU64", last_transno is "LPU64"\n",
506 mdt->mdt_mount_count, mdt->mdt_last_transno);
508 spin_lock(&mdt->mdt_transno_lock);
509 mdt->mdt_lsd.lsd_last_transno = mdt->mdt_last_transno;
510 spin_unlock(&mdt->mdt_transno_lock);
512 rc = mdt_last_rcvd_header_write(env, mdt, th);
513 mdt_trans_stop(env, mdt, th);
518 int mdt_client_new(const struct lu_env *env, struct mdt_device *mdt)
520 unsigned long *bitmap = mdt->mdt_client_bitmap;
521 struct mdt_thread_info *mti;
522 struct mdt_export_data *med;
523 struct lsd_client_data *lcd;
524 struct lr_server_data *lsd = &mdt->mdt_lsd;
525 struct obd_device *obd = mdt2obd_dev(mdt);
532 mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
533 LASSERT(mti != NULL);
535 med = &mti->mti_exp->exp_mdt_data;
538 LASSERT(bitmap != NULL);
539 if (!strcmp(med->med_lcd->lcd_uuid, obd->obd_uuid.uuid))
542 /* the bitmap operations can handle cl_idx > sizeof(long) * 8, so
543 * there's no need for extra complication here
545 spin_lock(&mdt->mdt_client_bitmap_lock);
546 cl_idx = find_first_zero_bit(bitmap, LR_MAX_CLIENTS);
547 if (cl_idx >= LR_MAX_CLIENTS ||
548 OBD_FAIL_CHECK(OBD_FAIL_MDS_CLIENT_ADD)) {
549 CERROR("no room for %u clients - fix LR_MAX_CLIENTS\n",
551 spin_unlock(&mdt->mdt_client_bitmap_lock);
554 set_bit(cl_idx, bitmap);
555 spin_unlock(&mdt->mdt_client_bitmap_lock);
557 CDEBUG(D_INFO, "client at idx %d with UUID '%s' added\n",
558 cl_idx, med->med_lcd->lcd_uuid);
560 med->med_lr_idx = cl_idx;
561 med->med_lr_off = lsd->lsd_client_start +
562 (cl_idx * lsd->lsd_client_size);
563 init_mutex(&med->med_lcd_lock);
565 LASSERTF(med->med_lr_off > 0, "med_lr_off = %llu\n", med->med_lr_off);
567 /* Write new client data. */
568 off = med->med_lr_off;
569 mdt_trans_credit_init(env, mdt, MDT_TXN_LAST_RCVD_WRITE_OP);
571 th = mdt_trans_start(env, mdt);
576 * Until this operations will be committed the sync is needed
577 * for this export. This should be done _after_ starting the
578 * transaction so that many connecting clients will not bring
579 * server down with lots of sync writes.
581 mdt_trans_add_cb(th, lut_cb_client, class_export_cb_get(mti->mti_exp));
582 spin_lock(&mti->mti_exp->exp_lock);
583 mti->mti_exp->exp_need_sync = 1;
584 spin_unlock(&mti->mti_exp->exp_lock);
586 rc = mdt_last_rcvd_write(env, mdt, lcd, &off, th);
587 CDEBUG(D_INFO, "wrote client lcd at idx %u off %llu (len %u)\n",
588 cl_idx, med->med_lr_off, (int)sizeof(*lcd));
589 mdt_trans_stop(env, mdt, th);
594 /* Add client data to the MDS. We use a bitmap to locate a free space
595 * in the last_rcvd file if cl_off is -1 (i.e. a new client).
596 * Otherwise, we just have to read the data from the last_rcvd file and
597 * we know its offset.
599 * It should not be possible to fail adding an existing client - otherwise
600 * mdt_init_server_data() callsite needs to be fixed.
602 int mdt_client_add(const struct lu_env *env,
603 struct mdt_device *mdt, int cl_idx)
605 struct mdt_thread_info *mti;
606 struct mdt_export_data *med;
607 unsigned long *bitmap = mdt->mdt_client_bitmap;
608 struct obd_device *obd = mdt2obd_dev(mdt);
609 struct lr_server_data *lsd = &mdt->mdt_lsd;
613 mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
614 LASSERT(mti != NULL);
616 med = &mti->mti_exp->exp_mdt_data;
618 LASSERT(bitmap != NULL);
619 LASSERTF(cl_idx >= 0, "%d\n", cl_idx);
621 if (!strcmp(med->med_lcd->lcd_uuid, obd->obd_uuid.uuid))
624 spin_lock(&mdt->mdt_client_bitmap_lock);
625 if (test_and_set_bit(cl_idx, bitmap)) {
626 CERROR("MDS client %d: bit already set in bitmap!!\n",
630 spin_unlock(&mdt->mdt_client_bitmap_lock);
632 CDEBUG(D_INFO, "client at idx %d with UUID '%s' added\n",
633 cl_idx, med->med_lcd->lcd_uuid);
635 med->med_lr_idx = cl_idx;
636 med->med_lr_off = lsd->lsd_client_start +
637 (cl_idx * lsd->lsd_client_size);
638 init_mutex(&med->med_lcd_lock);
640 LASSERTF(med->med_lr_off > 0, "med_lr_off = %llu\n", med->med_lr_off);
645 int mdt_client_del(const struct lu_env *env, struct mdt_device *mdt)
647 struct mdt_thread_info *mti;
648 struct mdt_export_data *med;
649 struct lsd_client_data *lcd;
650 struct obd_device *obd = mdt2obd_dev(mdt);
651 struct obd_export *exp;
657 mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
658 LASSERT(mti != NULL);
661 med = &exp->exp_mdt_data;
666 /* XXX: If lcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
667 if (!strcmp(lcd->lcd_uuid, obd->obd_uuid.uuid))
670 CDEBUG(D_INFO, "freeing client at idx %u, offset %lld\n",
671 med->med_lr_idx, med->med_lr_off);
673 off = med->med_lr_off;
676 * Don't clear med_lr_idx here as it is likely also unset. At worst we
677 * leak a client slot that will be cleaned on the next recovery.
680 CERROR("client idx %d has offset %lld\n",
681 med->med_lr_idx, off);
682 GOTO(free, rc = -EINVAL);
686 * Clear the bit _after_ zeroing out the client so we don't race with
687 * mdt_client_add and zero out new clients.
689 if (!test_bit(med->med_lr_idx, mdt->mdt_client_bitmap)) {
690 CERROR("MDT client %u: bit already clear in bitmap!!\n",
695 /* Make sure the server's last_transno is up to date.
696 * This should be done before zeroing client slot so last_transno will
697 * be in server data or in client data in case of failure */
698 mdt_server_data_update(env, mdt);
700 mdt_trans_credit_init(env, mdt, MDT_TXN_LAST_RCVD_WRITE_OP);
701 th = mdt_trans_start(env, mdt);
703 GOTO(free, rc = PTR_ERR(th));
705 mutex_down(&med->med_lcd_lock);
706 memset(lcd, 0, sizeof *lcd);
707 rc = mdt_last_rcvd_write(env, mdt, lcd, &off, th);
709 mutex_up(&med->med_lcd_lock);
710 mdt_trans_stop(env, mdt, th);
712 spin_lock(&mdt->mdt_client_bitmap_lock);
713 clear_bit(med->med_lr_idx, mdt->mdt_client_bitmap);
714 spin_unlock(&mdt->mdt_client_bitmap_lock);
716 CDEBUG(rc == 0 ? D_INFO : D_ERROR, "Zeroing out client idx %u in "
717 "%s, rc %d\n", med->med_lr_idx, LAST_RCVD, rc);
721 mutex_down(&med->med_lcd_lock);
723 mutex_up(&med->med_lcd_lock);
729 * last_rcvd & last_committed update callbacks
731 static int mdt_last_rcvd_update(struct mdt_thread_info *mti,
734 struct mdt_device *mdt = mti->mti_mdt;
735 struct ptlrpc_request *req = mdt_info_req(mti);
736 struct mdt_export_data *med;
737 struct lsd_client_data *lcd;
740 __s32 rc = th->th_result;
744 LASSERT(req->rq_export);
746 med = &req->rq_export->exp_mdt_data;
749 mutex_down(&med->med_lcd_lock);
751 /* if the export has already been disconnected, we have no last_rcvd slot,
752 * update server data with latest transno then */
754 mutex_up(&med->med_lcd_lock);
755 CWARN("commit transaction for disconnected client %s: rc %d\n",
756 req->rq_export->exp_client_uuid.uuid, rc);
757 err = mdt_last_rcvd_header_write(mti->mti_env, mdt, th);
761 off = med->med_lr_off;
762 LASSERT(ergo(mti->mti_transno == 0, rc != 0));
763 if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE ||
764 lustre_msg_get_opc(req->rq_reqmsg) == MDS_DONE_WRITING) {
765 if (mti->mti_transno != 0)
766 lcd->lcd_last_close_transno = mti->mti_transno;
767 lcd->lcd_last_close_xid = req->rq_xid;
768 lcd->lcd_last_close_result = rc;
770 /* VBR: save versions in last_rcvd for reconstruct. */
771 __u64 *pre_versions = lustre_msg_get_versions(req->rq_repmsg);
773 lcd->lcd_pre_versions[0] = pre_versions[0];
774 lcd->lcd_pre_versions[1] = pre_versions[1];
775 lcd->lcd_pre_versions[2] = pre_versions[2];
776 lcd->lcd_pre_versions[3] = pre_versions[3];
778 if (mti->mti_transno != 0)
779 lcd->lcd_last_transno = mti->mti_transno;
780 lcd->lcd_last_xid = req->rq_xid;
781 lcd->lcd_last_result = rc;
782 /*XXX: save intent_disposition in mdt_thread_info?
783 * also there is bug - intent_dispostion is __u64,
784 * see struct ldlm_reply->lock_policy_res1; */
785 lcd->lcd_last_data = mti->mti_opdata;
789 CERROR("client idx %d has offset %lld\n", med->med_lr_idx, off);
792 err = mdt_last_rcvd_write(mti->mti_env, mdt, lcd, &off, th);
794 mutex_up(&med->med_lcd_lock);
798 extern struct lu_context_key mdt_thread_key;
800 /* add credits for last_rcvd update */
801 static int mdt_txn_start_cb(const struct lu_env *env,
802 struct txn_param *param, void *cookie)
804 struct mdt_device *mdt = cookie;
806 param->tp_credits += mdt_trans_credit_get(env, mdt,
807 MDT_TXN_LAST_RCVD_WRITE_OP);
811 /* Set new object versions */
812 static void mdt_versions_set(struct mdt_thread_info *info)
815 for (i = 0; i < PTLRPC_NUM_VERSIONS; i++)
816 if (info->mti_mos[i] != NULL)
817 mo_version_set(info->mti_env,
818 mdt_object_child(info->mti_mos[i]),
822 /* Update last_rcvd records with latests transaction data */
823 static int mdt_txn_stop_cb(const struct lu_env *env,
824 struct thandle *txn, void *cookie)
826 struct mdt_device *mdt = cookie;
827 struct mdt_txn_info *txi;
828 struct mdt_thread_info *mti;
829 struct ptlrpc_request *req;
831 /* transno in two contexts - for commit_cb and for thread */
832 txi = lu_context_key_get(&txn->th_ctx, &mdt_txn_key);
833 mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
834 req = mdt_info_req(mti);
836 if (mti->mti_mdt == NULL || req == NULL || mti->mti_no_need_trans) {
837 txi->txi_transno = 0;
838 mti->mti_no_need_trans = 0;
842 if (mti->mti_has_trans) {
843 /* XXX: currently there are allowed cases, but the wrong cases
844 * are also possible, so better check is needed here */
845 CDEBUG(D_INFO, "More than one transaction "LPU64"\n",
850 mti->mti_has_trans = 1;
851 spin_lock(&mdt->mdt_transno_lock);
852 if (txn->th_result != 0) {
853 if (mti->mti_transno != 0) {
854 CERROR("Replay transno "LPU64" failed: rc %i\n",
855 mti->mti_transno, txn->th_result);
857 } else if (mti->mti_transno == 0) {
858 mti->mti_transno = ++ mdt->mdt_last_transno;
860 /* should be replay */
861 if (mti->mti_transno > mdt->mdt_last_transno)
862 mdt->mdt_last_transno = mti->mti_transno;
864 spin_unlock(&mdt->mdt_transno_lock);
865 /* sometimes the reply message has not been successfully packed */
866 LASSERT(req != NULL && req->rq_repmsg != NULL);
868 /** VBR: set new versions */
869 if (txn->th_result == 0)
870 mdt_versions_set(mti);
872 /* filling reply data */
873 CDEBUG(D_INODE, "transno = %llu, last_committed = %llu\n",
874 mti->mti_transno, req->rq_export->exp_obd->obd_last_committed);
876 req->rq_transno = mti->mti_transno;
877 lustre_msg_set_transno(req->rq_repmsg, mti->mti_transno);
878 /* save transno for the commit callback */
879 txi->txi_transno = mti->mti_transno;
881 /* add separate commit callback for transaction handling because we need
882 * export as parameter */
883 mdt_trans_add_cb(txn, lut_cb_last_committed,
884 class_export_cb_get(mti->mti_exp));
886 return mdt_last_rcvd_update(mti, txn);
889 /* commit callback, need to update last_committed value */
890 static int mdt_txn_commit_cb(const struct lu_env *env,
891 struct thandle *txn, void *cookie)
893 struct mdt_device *mdt = cookie;
894 struct mdt_txn_info *txi;
897 txi = lu_context_key_get(&txn->th_ctx, &mdt_txn_key);
899 /* iterate through all additional callbacks */
900 for (i = 0; i < txi->txi_cb_count; i++) {
901 txi->txi_cb[i].lut_cb_func(&mdt->mdt_lut, txi->txi_transno,
902 txi->txi_cb[i].lut_cb_data, 0);
907 int mdt_fs_setup(const struct lu_env *env, struct mdt_device *mdt,
908 struct obd_device *obd,
909 struct lustre_sb_info *lsi)
916 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_FS_SETUP))
919 /* prepare transactions callbacks */
920 mdt->mdt_txn_cb.dtc_txn_start = mdt_txn_start_cb;
921 mdt->mdt_txn_cb.dtc_txn_stop = mdt_txn_stop_cb;
922 mdt->mdt_txn_cb.dtc_txn_commit = mdt_txn_commit_cb;
923 mdt->mdt_txn_cb.dtc_cookie = mdt;
924 mdt->mdt_txn_cb.dtc_tag = LCT_MD_THREAD;
925 CFS_INIT_LIST_HEAD(&mdt->mdt_txn_cb.dtc_linkage);
927 dt_txn_callback_add(mdt->mdt_bottom, &mdt->mdt_txn_cb);
929 rc = mdt_server_data_init(env, mdt, lsi);
933 o = dt_store_open(env, mdt->mdt_bottom, "", CAPA_KEYS, &fid);
936 rc = mdt_capa_keys_init(env, mdt);
938 GOTO(put_ck_object, rc);
941 CERROR("cannot open %s: rc = %d\n", CAPA_KEYS, rc);
942 GOTO(disconnect_exports, rc);
947 lu_object_put(env, &o->do_lu);
948 mdt->mdt_ck_obj = NULL;
950 class_disconnect_exports(obd);
954 void mdt_fs_cleanup(const struct lu_env *env, struct mdt_device *mdt)
958 /* Remove transaction callback */
959 dt_txn_callback_del(mdt->mdt_bottom, &mdt->mdt_txn_cb);
961 lu_object_put(env, &mdt->mdt_ck_obj->do_lu);
962 mdt->mdt_ck_obj = NULL;
966 /* reconstruction code */
967 static void mdt_steal_ack_locks(struct ptlrpc_request *req)
969 struct obd_export *exp = req->rq_export;
970 struct list_head *tmp;
971 struct ptlrpc_reply_state *oldrep;
972 struct ptlrpc_service *svc;
975 /* CAVEAT EMPTOR: spinlock order */
976 spin_lock(&exp->exp_lock);
977 list_for_each (tmp, &exp->exp_outstanding_replies) {
978 oldrep = list_entry(tmp, struct ptlrpc_reply_state,rs_exp_list);
980 if (oldrep->rs_xid != req->rq_xid)
983 if (oldrep->rs_opc != lustre_msg_get_opc(req->rq_reqmsg))
984 CERROR ("Resent req xid "LPU64" has mismatched opc: "
985 "new %d old %d\n", req->rq_xid,
986 lustre_msg_get_opc(req->rq_reqmsg),
989 svc = oldrep->rs_service;
990 spin_lock (&svc->srv_lock);
992 list_del_init (&oldrep->rs_exp_list);
994 CWARN("Stealing %d locks from rs %p x"LPD64".t"LPD64
996 oldrep->rs_nlocks, oldrep,
997 oldrep->rs_xid, oldrep->rs_transno, oldrep->rs_opc,
998 libcfs_nid2str(exp->exp_connection->c_peer.nid));
1000 for (i = 0; i < oldrep->rs_nlocks; i++)
1001 ptlrpc_save_lock(req, &oldrep->rs_locks[i],
1002 oldrep->rs_modes[i], 0);
1003 oldrep->rs_nlocks = 0;
1005 DEBUG_REQ(D_HA, req, "stole locks for");
1006 spin_lock(&oldrep->rs_lock);
1007 ptlrpc_schedule_difficult_reply (oldrep);
1008 spin_unlock(&oldrep->rs_lock);
1010 spin_unlock (&svc->srv_lock);
1013 spin_unlock(&exp->exp_lock);
1017 * VBR: restore versions
1019 void mdt_vbr_reconstruct(struct ptlrpc_request *req,
1020 struct lsd_client_data *lcd)
1022 __u64 pre_versions[4] = {0};
1023 pre_versions[0] = lcd->lcd_pre_versions[0];
1024 pre_versions[1] = lcd->lcd_pre_versions[1];
1025 pre_versions[2] = lcd->lcd_pre_versions[2];
1026 pre_versions[3] = lcd->lcd_pre_versions[3];
1027 lustre_msg_set_versions(req->rq_repmsg, pre_versions);
1030 void mdt_req_from_lcd(struct ptlrpc_request *req,
1031 struct lsd_client_data *lcd)
1033 DEBUG_REQ(D_HA, req, "restoring transno "LPD64"/status %d",
1034 lcd->lcd_last_transno, lcd->lcd_last_result);
1036 if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE ||
1037 lustre_msg_get_opc(req->rq_repmsg) == MDS_DONE_WRITING) {
1038 req->rq_transno = lcd->lcd_last_close_transno;
1039 req->rq_status = lcd->lcd_last_close_result;
1041 req->rq_transno = lcd->lcd_last_transno;
1042 req->rq_status = lcd->lcd_last_result;
1043 mdt_vbr_reconstruct(req, lcd);
1045 if (req->rq_status != 0)
1046 req->rq_transno = 0;
1047 lustre_msg_set_transno(req->rq_repmsg, req->rq_transno);
1048 lustre_msg_set_status(req->rq_repmsg, req->rq_status);
1049 DEBUG_REQ(D_RPCTRACE, req, "restoring transno "LPD64"/status %d",
1050 req->rq_transno, req->rq_status);
1052 mdt_steal_ack_locks(req);
1055 void mdt_reconstruct_generic(struct mdt_thread_info *mti,
1056 struct mdt_lock_handle *lhc)
1058 struct ptlrpc_request *req = mdt_info_req(mti);
1059 struct mdt_export_data *med = &req->rq_export->exp_mdt_data;
1061 return mdt_req_from_lcd(req, med->med_lcd);
1064 static void mdt_reconstruct_create(struct mdt_thread_info *mti,
1065 struct mdt_lock_handle *lhc)
1067 struct ptlrpc_request *req = mdt_info_req(mti);
1068 struct obd_export *exp = req->rq_export;
1069 struct mdt_export_data *med = &exp->exp_mdt_data;
1070 struct mdt_device *mdt = mti->mti_mdt;
1071 struct mdt_object *child;
1072 struct mdt_body *body;
1075 mdt_req_from_lcd(req, med->med_lcd);
1079 /* if no error, so child was created with requested fid */
1080 child = mdt_object_find(mti->mti_env, mdt, mti->mti_rr.rr_fid2);
1081 if (IS_ERR(child)) {
1082 rc = PTR_ERR(child);
1083 LCONSOLE_WARN("Child "DFID" lookup error %d."
1084 " Evicting client %s with export %s.\n",
1085 PFID(mdt_object_fid(child)), rc,
1086 obd_uuid2str(&exp->exp_client_uuid),
1087 obd_export_nid2str(exp));
1088 mdt_export_evict(exp);
1093 body = req_capsule_server_get(mti->mti_pill, &RMF_MDT_BODY);
1094 mti->mti_attr.ma_need = MA_INODE;
1095 mti->mti_attr.ma_valid = 0;
1096 rc = mo_attr_get(mti->mti_env, mdt_object_child(child), &mti->mti_attr);
1097 if (rc == -EREMOTE) {
1098 /* object was created on remote server */
1099 req->rq_status = rc;
1100 body->valid |= OBD_MD_MDS;
1102 mdt_pack_attr2body(mti, body, &mti->mti_attr.ma_attr,
1103 mdt_object_fid(child));
1104 mdt_object_put(mti->mti_env, child);
1107 static void mdt_reconstruct_setattr(struct mdt_thread_info *mti,
1108 struct mdt_lock_handle *lhc)
1110 struct ptlrpc_request *req = mdt_info_req(mti);
1111 struct obd_export *exp = req->rq_export;
1112 struct mdt_export_data *med = &exp->exp_mdt_data;
1113 struct mdt_device *mdt = mti->mti_mdt;
1114 struct mdt_object *obj;
1115 struct mdt_body *body;
1117 mdt_req_from_lcd(req, med->med_lcd);
1121 body = req_capsule_server_get(mti->mti_pill, &RMF_MDT_BODY);
1122 obj = mdt_object_find(mti->mti_env, mdt, mti->mti_rr.rr_fid1);
1124 int rc = PTR_ERR(obj);
1125 LCONSOLE_WARN(""DFID" lookup error %d."
1126 " Evicting client %s with export %s.\n",
1127 PFID(mdt_object_fid(obj)), rc,
1128 obd_uuid2str(&exp->exp_client_uuid),
1129 obd_export_nid2str(exp));
1130 mdt_export_evict(exp);
1134 mti->mti_attr.ma_need = MA_INODE;
1135 mti->mti_attr.ma_valid = 0;
1136 mo_attr_get(mti->mti_env, mdt_object_child(obj), &mti->mti_attr);
1137 mdt_pack_attr2body(mti, body, &mti->mti_attr.ma_attr,
1138 mdt_object_fid(obj));
1139 if (mti->mti_epoch && (mti->mti_epoch->flags & MF_EPOCH_OPEN)) {
1140 struct mdt_file_data *mfd;
1141 struct mdt_body *repbody;
1143 repbody = req_capsule_server_get(mti->mti_pill, &RMF_MDT_BODY);
1144 repbody->ioepoch = obj->mot_ioepoch;
1145 spin_lock(&med->med_open_lock);
1146 list_for_each_entry(mfd, &med->med_open_head, mfd_list) {
1147 if (mfd->mfd_xid == req->rq_xid)
1150 LASSERT(&mfd->mfd_list != &med->med_open_head);
1151 spin_unlock(&med->med_open_lock);
1152 repbody->handle.cookie = mfd->mfd_handle.h_cookie;
1155 mdt_object_put(mti->mti_env, obj);
1158 typedef void (*mdt_reconstructor)(struct mdt_thread_info *mti,
1159 struct mdt_lock_handle *lhc);
1161 static mdt_reconstructor reconstructors[REINT_MAX] = {
1162 [REINT_SETATTR] = mdt_reconstruct_setattr,
1163 [REINT_CREATE] = mdt_reconstruct_create,
1164 [REINT_LINK] = mdt_reconstruct_generic,
1165 [REINT_UNLINK] = mdt_reconstruct_generic,
1166 [REINT_RENAME] = mdt_reconstruct_generic,
1167 [REINT_OPEN] = mdt_reconstruct_open,
1168 [REINT_SETXATTR] = mdt_reconstruct_generic
1171 void mdt_reconstruct(struct mdt_thread_info *mti,
1172 struct mdt_lock_handle *lhc)
1175 reconstructors[mti->mti_rr.rr_opcode](mti, lhc);