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 (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
30 * Use is subject to license terms.
32 * Copyright (c) 2011 Whamcloud, Inc.
36 * This file is part of Lustre, http://www.lustre.org/
37 * Lustre is a trademark of Sun Microsystems, Inc.
39 * lustre/mdt/mdt_recovery.c
41 * Lustre Metadata Target (mdt) recovery-related methods
43 * Author: Huang Hua <huanghua@clusterfs.com>
44 * Author: Pershin Mike <tappro@clusterfs.com>
48 # define EXPORT_SYMTAB
50 #define DEBUG_SUBSYSTEM S_MDS
52 #include "mdt_internal.h"
54 static int mdt_server_data_update(const struct lu_env *env,
55 struct mdt_device *mdt);
57 struct lu_buf *mdt_buf(const struct lu_env *env, void *area, ssize_t len)
60 struct mdt_thread_info *mti;
62 mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
69 const struct lu_buf *mdt_buf_const(const struct lu_env *env,
70 const void *area, ssize_t len)
73 struct mdt_thread_info *mti;
75 mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
78 buf->lb_buf = (void *)area;
83 static inline int mdt_trans_credit_get(const struct lu_env *env,
84 struct mdt_device *mdt,
87 struct dt_device *dev = mdt->mdt_bottom;
90 case MDT_TXN_CAPA_KEYS_WRITE_OP:
91 case MDT_TXN_LAST_RCVD_WRITE_OP:
92 cr = dev->dd_ops->dt_credit_get(env,
102 void mdt_trans_credit_init(const struct lu_env *env,
103 struct mdt_device *mdt,
106 struct mdt_thread_info *mti;
110 mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
111 p = &mti->mti_txn_param;
113 cr = mdt_trans_credit_get(env, mdt, op);
114 txn_param_init(p, cr);
117 struct thandle* mdt_trans_start(const struct lu_env *env,
118 struct mdt_device *mdt)
120 struct mdt_thread_info *mti;
123 mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
124 p = &mti->mti_txn_param;
126 /* export can require sync operations */
127 if (mti->mti_exp != NULL)
128 p->tp_sync = mti->mti_exp->exp_need_sync;
130 return mdt->mdt_bottom->dd_ops->dt_trans_start(env, mdt->mdt_bottom, p);
133 void mdt_trans_stop(const struct lu_env *env,
134 struct mdt_device *mdt, struct thandle *th)
136 mdt->mdt_bottom->dd_ops->dt_trans_stop(env, th);
139 static inline int mdt_last_rcvd_header_read(const struct lu_env *env,
140 struct mdt_device *mdt)
142 struct mdt_thread_info *mti;
145 mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
148 rc = dt_record_read(env, mdt->mdt_lut.lut_last_rcvd,
149 mdt_buf(env, &mti->mti_lsd, sizeof(mti->mti_lsd)),
152 lsd_le_to_cpu(&mti->mti_lsd, &mdt->mdt_lut.lut_lsd);
154 CDEBUG(D_INFO, "read last_rcvd header rc = %d, uuid = %s, "
155 "last_transno = "LPU64"\n", rc, mdt->mdt_lut.lut_lsd.lsd_uuid,
156 mdt->mdt_lut.lut_lsd.lsd_last_transno);
160 static int mdt_last_rcvd_header_write(const struct lu_env *env,
161 struct mdt_device *mdt,
164 struct mdt_thread_info *mti;
168 mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
171 lsd_cpu_to_le(&mdt->mdt_lut.lut_lsd, &mti->mti_lsd);
173 rc = dt_record_write(env, mdt->mdt_lut.lut_last_rcvd,
174 mdt_buf_const(env, &mti->mti_lsd,
175 sizeof(mti->mti_lsd)),
178 CDEBUG(D_INFO, "write last_rcvd header rc = %d, uuid = %s, "
179 "last_transno = "LPU64"\n", rc, mdt->mdt_lut.lut_lsd.lsd_uuid,
180 mdt->mdt_lut.lut_lsd.lsd_last_transno);
185 static int mdt_last_rcvd_read(const struct lu_env *env, struct mdt_device *mdt,
186 struct lsd_client_data *lcd, loff_t *off,
189 struct mdt_thread_info *mti;
190 struct lsd_client_data *tmp;
193 mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
195 rc = dt_record_read(env, mdt->mdt_lut.lut_last_rcvd,
196 mdt_buf(env, tmp, sizeof(*tmp)), off);
198 check_lcd(mdt2obd_dev(mdt)->obd_name, index, tmp);
199 lcd_le_to_cpu(tmp, lcd);
202 CDEBUG(D_INFO, "read lcd @%d rc = %d, uuid = %s, last_transno = "LPU64
203 ", last_xid = "LPU64", last_result = %u, last_data = %u, "
204 "last_close_transno = "LPU64", last_close_xid = "LPU64", "
205 "last_close_result = %u\n", (int)(*off - sizeof(*tmp)),
206 rc, lcd->lcd_uuid, lcd->lcd_last_transno, lcd->lcd_last_xid,
207 lcd->lcd_last_result, lcd->lcd_last_data,
208 lcd->lcd_last_close_transno, lcd->lcd_last_close_xid,
209 lcd->lcd_last_close_result);
213 static int mdt_last_rcvd_write(const struct lu_env *env,
214 struct mdt_device *mdt,
215 struct lsd_client_data *lcd,
216 loff_t *off, struct thandle *th)
218 struct mdt_thread_info *mti;
219 struct lsd_client_data *tmp;
223 mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
226 lcd_cpu_to_le(lcd, tmp);
228 rc = dt_record_write(env, mdt->mdt_lut.lut_last_rcvd,
229 mdt_buf_const(env, tmp, sizeof(*tmp)), off, th);
231 CDEBUG(D_INFO, "write lcd @%d rc = %d, uuid = %s, last_transno = "LPU64
232 ", last_xid = "LPU64", last_result = %u, last_data = %u, "
233 "last_close_transno = "LPU64", last_close_xid = "LPU64" ,"
234 "last_close_result = %u\n", (int)(*off - sizeof(*tmp)),
235 rc, lcd->lcd_uuid, lcd->lcd_last_transno, lcd->lcd_last_xid,
236 lcd->lcd_last_result, lcd->lcd_last_data,
237 lcd->lcd_last_close_transno, lcd->lcd_last_close_xid,
238 lcd->lcd_last_close_result);
242 static int mdt_clients_data_init(const struct lu_env *env,
243 struct mdt_device *mdt,
244 unsigned long last_size)
246 struct lr_server_data *lsd = &mdt->mdt_lut.lut_lsd;
247 struct lsd_client_data *lcd;
248 struct obd_device *obd = mdt2obd_dev(mdt);
258 /* When we do a clean MDS shutdown, we save the last_transno into
259 * the header. If we find clients with higher last_transno values
260 * then those clients may need recovery done. */
261 LASSERT(cfs_atomic_read(&obd->obd_req_replay_clients) == 0);
262 for (cl_idx = 0, off = lsd->lsd_client_start;
263 off < last_size; cl_idx++) {
265 struct obd_export *exp;
266 struct mdt_thread_info *mti;
268 off = lsd->lsd_client_start +
269 cl_idx * lsd->lsd_client_size;
271 rc = mdt_last_rcvd_read(env, mdt, lcd, &off, cl_idx);
273 CERROR("error reading MDS %s idx %d, off %llu: rc %d\n",
274 LAST_RCVD, cl_idx, off, rc);
276 break; /* read error shouldn't cause startup to fail */
279 if (lcd->lcd_uuid[0] == '\0') {
280 CDEBUG(D_INFO, "skipping zeroed client at offset %d\n",
285 last_transno = lcd_last_transno(lcd);
287 /* These exports are cleaned up by mdt_obd_disconnect(), so
288 * they need to be set up like real exports as
289 * mdt_obd_connect() does.
291 CDEBUG(D_HA, "RCVRNG CLIENT uuid: %s idx: %d lr: "LPU64
292 " srv lr: "LPU64" lx: "LPU64"\n", lcd->lcd_uuid, cl_idx,
293 last_transno, lsd->lsd_last_transno,
296 exp = class_new_export(obd, (struct obd_uuid *)lcd->lcd_uuid);
298 if (PTR_ERR(exp) == -EALREADY) {
299 /* export already exists, zero out this one */
300 CERROR("Duplicate export %s!\n", lcd->lcd_uuid);
303 GOTO(err_client, rc = PTR_ERR(exp));
306 mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
307 LASSERT(mti != NULL);
309 /* copy on-disk lcd to the export */
310 *exp->exp_target_data.ted_lcd = *lcd;
311 rc = mdt_client_add(env, mdt, cl_idx);
312 /* can't fail existing */
313 LASSERTF(rc == 0, "rc = %d\n", rc);
314 /* VBR: set export last committed version */
315 exp->exp_last_committed = last_transno;
316 cfs_spin_lock(&exp->exp_lock);
317 exp->exp_connecting = 0;
318 exp->exp_in_recovery = 0;
319 cfs_spin_unlock(&exp->exp_lock);
320 obd->obd_max_recoverable_clients++;
321 class_export_put(exp);
323 CDEBUG(D_OTHER, "client at idx %d has last_transno="LPU64"\n",
324 cl_idx, last_transno);
325 /* protect __u64 value update */
326 cfs_spin_lock(&mdt->mdt_lut.lut_translock);
327 mdt->mdt_lut.lut_last_transno = max(last_transno,
328 mdt->mdt_lut.lut_last_transno);
329 cfs_spin_unlock(&mdt->mdt_lut.lut_translock);
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_lut.lut_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_lut.lut_last_rcvd;
364 rc = obj->do_ops->do_attr_get(env, obj, 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 if (last_rcvd_size > lsd->lsd_client_start) {
422 LCONSOLE_WARN("Mounting %s at first time on 1.8 FS, "
423 "remove all clients for interop needs\n",
425 simple_truncate(lsi->lsi_srv_mnt->mnt_sb->s_root,
426 lsi->lsi_srv_mnt, LAST_RCVD,
427 lsd->lsd_client_start);
428 last_rcvd_size = lsd->lsd_client_start;
430 /** set 2.0 flag to upgrade/downgrade between 1.8 and 2.0 */
431 lsd->lsd_feature_compat |= OBD_COMPAT_20;
434 if (ldd->ldd_flags & LDD_F_IAM_DIR)
435 lsd->lsd_feature_incompat |= OBD_INCOMPAT_IAM_DIR;
437 lsd->lsd_feature_incompat |= OBD_INCOMPAT_FID;
439 cfs_spin_lock(&mdt->mdt_lut.lut_translock);
440 mdt->mdt_lut.lut_last_transno = lsd->lsd_last_transno;
441 cfs_spin_unlock(&mdt->mdt_lut.lut_translock);
443 CDEBUG(D_INODE, "========BEGIN DUMPING LAST_RCVD========\n");
444 CDEBUG(D_INODE, "%s: server last_transno: "LPU64"\n",
445 obd->obd_name, mdt->mdt_lut.lut_last_transno);
446 CDEBUG(D_INODE, "%s: server mount_count: "LPU64"\n",
447 obd->obd_name, mount_count + 1);
448 CDEBUG(D_INODE, "%s: server data size: %u\n",
449 obd->obd_name, lsd->lsd_server_size);
450 CDEBUG(D_INODE, "%s: per-client data start: %u\n",
451 obd->obd_name, lsd->lsd_client_start);
452 CDEBUG(D_INODE, "%s: per-client data size: %u\n",
453 obd->obd_name, lsd->lsd_client_size);
454 CDEBUG(D_INODE, "%s: last_rcvd size: %lu\n",
455 obd->obd_name, last_rcvd_size);
456 CDEBUG(D_INODE, "%s: last_rcvd clients: %lu\n", obd->obd_name,
457 last_rcvd_size <= lsd->lsd_client_start ? 0 :
458 (last_rcvd_size - lsd->lsd_client_start) /
459 lsd->lsd_client_size);
460 CDEBUG(D_INODE, "========END DUMPING LAST_RCVD========\n");
462 if (!lsd->lsd_server_size || !lsd->lsd_client_start ||
463 !lsd->lsd_client_size) {
464 CERROR("Bad last_rcvd contents!\n");
465 GOTO(out, rc = -EINVAL);
468 rc = mdt_clients_data_init(env, mdt, last_rcvd_size);
470 GOTO(err_client, rc);
472 cfs_spin_lock(&mdt->mdt_lut.lut_translock);
473 /* obd_last_committed is used for compatibility
474 * with other lustre recovery code */
475 obd->obd_last_committed = mdt->mdt_lut.lut_last_transno;
476 cfs_spin_unlock(&mdt->mdt_lut.lut_translock);
478 obd->u.obt.obt_mount_count = mount_count + 1;
479 lsd->lsd_mount_count = obd->u.obt.obt_mount_count;
481 /* save it, so mount count and last_transno is current */
482 rc = mdt_server_data_update(env, mdt);
484 GOTO(err_client, rc);
489 class_disconnect_exports(obd);
494 static int mdt_server_data_update(const struct lu_env *env,
495 struct mdt_device *mdt)
497 struct mdt_thread_info *mti;
501 mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
503 mdt_trans_credit_init(env, mdt, MDT_TXN_LAST_RCVD_WRITE_OP);
504 th = mdt_trans_start(env, mdt);
508 CDEBUG(D_SUPER, "MDS mount_count is "LPU64", last_transno is "LPU64"\n",
509 mdt->mdt_lut.lut_obd->u.obt.obt_mount_count,
510 mdt->mdt_lut.lut_last_transno);
512 cfs_spin_lock(&mdt->mdt_lut.lut_translock);
513 mdt->mdt_lut.lut_lsd.lsd_last_transno = mdt->mdt_lut.lut_last_transno;
514 cfs_spin_unlock(&mdt->mdt_lut.lut_translock);
516 rc = mdt_last_rcvd_header_write(env, mdt, th);
517 mdt_trans_stop(env, mdt, th);
522 int mdt_client_new(const struct lu_env *env, struct mdt_device *mdt)
524 unsigned long *bitmap = mdt->mdt_lut.lut_client_bitmap;
525 struct mdt_thread_info *mti;
526 struct tg_export_data *ted;
527 struct lr_server_data *lsd = &mdt->mdt_lut.lut_lsd;
528 struct obd_device *obd = mdt2obd_dev(mdt);
535 mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
536 LASSERT(mti != NULL);
538 ted = &mti->mti_exp->exp_target_data;
540 LASSERT(bitmap != NULL);
541 if (!strcmp(ted->ted_lcd->lcd_uuid, obd->obd_uuid.uuid))
544 /* the bitmap operations can handle cl_idx > sizeof(long) * 8, so
545 * there's no need for extra complication here
547 cfs_spin_lock(&mdt->mdt_lut.lut_client_bitmap_lock);
548 cl_idx = cfs_find_first_zero_bit(bitmap, LR_MAX_CLIENTS);
549 if (cl_idx >= LR_MAX_CLIENTS ||
550 OBD_FAIL_CHECK(OBD_FAIL_MDS_CLIENT_ADD)) {
551 CERROR("no room for %u clients - fix LR_MAX_CLIENTS\n",
553 cfs_spin_unlock(&mdt->mdt_lut.lut_client_bitmap_lock);
556 cfs_set_bit(cl_idx, bitmap);
557 cfs_spin_unlock(&mdt->mdt_lut.lut_client_bitmap_lock);
559 CDEBUG(D_INFO, "client at idx %d with UUID '%s' added\n",
560 cl_idx, ted->ted_lcd->lcd_uuid);
562 ted->ted_lr_idx = cl_idx;
563 ted->ted_lr_off = lsd->lsd_client_start +
564 (cl_idx * lsd->lsd_client_size);
565 cfs_init_mutex(&ted->ted_lcd_lock);
567 LASSERTF(ted->ted_lr_off > 0, "ted_lr_off = %llu\n", ted->ted_lr_off);
569 /* Write new client data. */
570 off = ted->ted_lr_off;
572 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_CLIENT_ADD))
575 mdt_trans_credit_init(env, mdt, MDT_TXN_LAST_RCVD_WRITE_OP);
577 th = mdt_trans_start(env, mdt);
582 * Until this operations will be committed the sync is needed
583 * for this export. This should be done _after_ starting the
584 * transaction so that many connecting clients will not bring
585 * server down with lots of sync writes.
587 mdt_trans_add_cb(th, lut_cb_client, class_export_cb_get(mti->mti_exp));
588 cfs_spin_lock(&mti->mti_exp->exp_lock);
589 mti->mti_exp->exp_need_sync = 1;
590 cfs_spin_unlock(&mti->mti_exp->exp_lock);
592 rc = mdt_last_rcvd_write(env, mdt, ted->ted_lcd, &off, th);
593 CDEBUG(D_INFO, "wrote client lcd at idx %u off %llu (len %u)\n",
594 cl_idx, ted->ted_lr_off, (int)sizeof(*(ted->ted_lcd)));
595 mdt_trans_stop(env, mdt, th);
600 /* Add client data to the MDS. We use a bitmap to locate a free space
601 * in the last_rcvd file if cl_off is -1 (i.e. a new client).
602 * Otherwise, we just have to read the data from the last_rcvd file and
603 * we know its offset.
605 * It should not be possible to fail adding an existing client - otherwise
606 * mdt_init_server_data() callsite needs to be fixed.
608 int mdt_client_add(const struct lu_env *env,
609 struct mdt_device *mdt, int cl_idx)
611 struct mdt_thread_info *mti;
612 struct tg_export_data *ted;
613 unsigned long *bitmap = mdt->mdt_lut.lut_client_bitmap;
614 struct obd_device *obd = mdt2obd_dev(mdt);
615 struct lr_server_data *lsd = &mdt->mdt_lut.lut_lsd;
619 mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
620 LASSERT(mti != NULL);
622 ted = &mti->mti_exp->exp_target_data;
624 LASSERT(bitmap != NULL);
625 LASSERTF(cl_idx >= 0, "%d\n", cl_idx);
627 if (!strcmp(ted->ted_lcd->lcd_uuid, obd->obd_uuid.uuid))
630 cfs_spin_lock(&mdt->mdt_lut.lut_client_bitmap_lock);
631 if (cfs_test_and_set_bit(cl_idx, bitmap)) {
632 CERROR("MDS client %d: bit already set in bitmap!!\n",
636 cfs_spin_unlock(&mdt->mdt_lut.lut_client_bitmap_lock);
638 CDEBUG(D_INFO, "client at idx %d with UUID '%s' added\n",
639 cl_idx, ted->ted_lcd->lcd_uuid);
641 ted->ted_lr_idx = cl_idx;
642 ted->ted_lr_off = lsd->lsd_client_start +
643 (cl_idx * lsd->lsd_client_size);
644 cfs_init_mutex(&ted->ted_lcd_lock);
646 LASSERTF(ted->ted_lr_off > 0, "ted_lr_off = %llu\n", ted->ted_lr_off);
651 int mdt_client_del(const struct lu_env *env, struct mdt_device *mdt)
653 struct mdt_thread_info *mti;
654 struct tg_export_data *ted;
655 struct obd_device *obd = mdt2obd_dev(mdt);
656 struct obd_export *exp;
662 mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
663 LASSERT(mti != NULL);
666 ted = &exp->exp_target_data;
670 /* XXX: If lcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
671 if (!strcmp(ted->ted_lcd->lcd_uuid, obd->obd_uuid.uuid))
674 CDEBUG(D_INFO, "freeing client at idx %u, offset %lld\n",
675 ted->ted_lr_idx, ted->ted_lr_off);
677 off = ted->ted_lr_off;
680 * Don't clear ted_lr_idx here as it is likely also unset. At worst we
681 * leak a client slot that will be cleaned on the next recovery.
684 CERROR("client idx %d has offset %lld\n",
685 ted->ted_lr_idx, off);
686 GOTO(free, rc = -EINVAL);
690 * Clear the bit _after_ zeroing out the client so we don't race with
691 * mdt_client_add and zero out new clients.
693 if (!cfs_test_bit(ted->ted_lr_idx, mdt->mdt_lut.lut_client_bitmap)) {
694 CERROR("MDT client %u: bit already clear in bitmap!!\n",
699 /* Make sure the server's last_transno is up to date.
700 * This should be done before zeroing client slot so last_transno will
701 * be in server data or in client data in case of failure */
702 mdt_server_data_update(env, mdt);
704 mdt_trans_credit_init(env, mdt, MDT_TXN_LAST_RCVD_WRITE_OP);
705 th = mdt_trans_start(env, mdt);
707 GOTO(free, rc = PTR_ERR(th));
709 cfs_mutex_down(&ted->ted_lcd_lock);
710 memset(ted->ted_lcd->lcd_uuid, 0, sizeof ted->ted_lcd->lcd_uuid);
711 rc = mdt_last_rcvd_write(env, mdt, ted->ted_lcd, &off, th);
712 cfs_mutex_up(&ted->ted_lcd_lock);
713 mdt_trans_stop(env, mdt, th);
715 CDEBUG(rc == 0 ? D_INFO : D_ERROR, "Zeroing out client idx %u in "
716 "%s, rc %d\n", ted->ted_lr_idx, LAST_RCVD, rc);
723 * last_rcvd & last_committed update callbacks
725 static int mdt_last_rcvd_update(struct mdt_thread_info *mti,
728 struct mdt_device *mdt = mti->mti_mdt;
729 struct ptlrpc_request *req = mdt_info_req(mti);
730 struct tg_export_data *ted;
731 struct lsd_client_data *lcd;
734 __s32 rc = th->th_result;
738 LASSERT(req->rq_export);
740 ted = &req->rq_export->exp_target_data;
743 cfs_mutex_down(&ted->ted_lcd_lock);
745 /* if the export has already been disconnected, we have no last_rcvd slot,
746 * update server data with latest transno then */
748 cfs_mutex_up(&ted->ted_lcd_lock);
749 CWARN("commit transaction for disconnected client %s: rc %d\n",
750 req->rq_export->exp_client_uuid.uuid, rc);
751 err = mdt_last_rcvd_header_write(mti->mti_env, mdt, th);
755 off = ted->ted_lr_off;
756 LASSERT(ergo(mti->mti_transno == 0, rc != 0));
757 if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE ||
758 lustre_msg_get_opc(req->rq_reqmsg) == MDS_DONE_WRITING) {
759 if (mti->mti_transno != 0) {
760 if (lcd->lcd_last_close_transno > mti->mti_transno) {
761 CERROR("Trying to overwrite bigger transno:"
762 "on-disk: "LPU64", new: "LPU64" "
763 "replay: %d. see LU-617.\n",
764 lcd->lcd_last_close_transno,
765 mti->mti_transno, req_is_replay(req));
766 if (req_is_replay(req)) {
767 cfs_spin_lock(&req->rq_export->exp_lock);
768 req->rq_export->exp_vbr_failed = 1;
769 cfs_spin_unlock(&req->rq_export->exp_lock);
771 cfs_mutex_up(&ted->ted_lcd_lock);
772 RETURN(req_is_replay(req) ? -EOVERFLOW : 0);
774 lcd->lcd_last_close_transno = mti->mti_transno;
776 lcd->lcd_last_close_xid = req->rq_xid;
777 lcd->lcd_last_close_result = rc;
779 /* VBR: save versions in last_rcvd for reconstruct. */
780 __u64 *pre_versions = lustre_msg_get_versions(req->rq_repmsg);
782 lcd->lcd_pre_versions[0] = pre_versions[0];
783 lcd->lcd_pre_versions[1] = pre_versions[1];
784 lcd->lcd_pre_versions[2] = pre_versions[2];
785 lcd->lcd_pre_versions[3] = pre_versions[3];
787 if (mti->mti_transno != 0) {
788 if (lcd->lcd_last_transno > mti->mti_transno) {
789 CERROR("Trying to overwrite bigger transno:"
790 "on-disk: "LPU64", new: "LPU64" "
791 "replay: %d. see LU-617.\n",
792 lcd->lcd_last_transno,
793 mti->mti_transno, req_is_replay(req));
794 if (req_is_replay(req)) {
795 cfs_spin_lock(&req->rq_export->exp_lock);
796 req->rq_export->exp_vbr_failed = 1;
797 cfs_spin_unlock(&req->rq_export->exp_lock);
799 cfs_mutex_up(&ted->ted_lcd_lock);
800 RETURN(req_is_replay(req) ? -EOVERFLOW : 0);
802 lcd->lcd_last_transno = mti->mti_transno;
804 lcd->lcd_last_xid = req->rq_xid;
805 lcd->lcd_last_result = rc;
806 /*XXX: save intent_disposition in mdt_thread_info?
807 * also there is bug - intent_dispostion is __u64,
808 * see struct ldlm_reply->lock_policy_res1; */
809 lcd->lcd_last_data = mti->mti_opdata;
813 CERROR("client idx %d has offset %lld\n", ted->ted_lr_idx, off);
816 err = mdt_last_rcvd_write(mti->mti_env, mdt, lcd, &off, th);
818 cfs_mutex_up(&ted->ted_lcd_lock);
822 extern struct lu_context_key mdt_thread_key;
824 /* add credits for last_rcvd update */
825 static int mdt_txn_start_cb(const struct lu_env *env,
826 struct txn_param *param, void *cookie)
828 struct mdt_device *mdt = cookie;
830 param->tp_credits += mdt_trans_credit_get(env, mdt,
831 MDT_TXN_LAST_RCVD_WRITE_OP);
835 /* Set new object versions */
836 static void mdt_version_set(struct mdt_thread_info *info)
838 if (info->mti_mos != NULL) {
839 mo_version_set(info->mti_env, mdt_object_child(info->mti_mos),
841 info->mti_mos = NULL;
845 /* Update last_rcvd records with latests transaction data */
846 static int mdt_txn_stop_cb(const struct lu_env *env,
847 struct thandle *txn, void *cookie)
849 struct mdt_device *mdt = cookie;
850 struct mdt_txn_info *txi;
851 struct mdt_thread_info *mti;
852 struct ptlrpc_request *req;
854 /* transno in two contexts - for commit_cb and for thread */
855 txi = lu_context_key_get(&txn->th_ctx, &mdt_txn_key);
856 mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
857 req = mdt_info_req(mti);
859 if (mti->mti_mdt == NULL || req == NULL || mti->mti_no_need_trans) {
860 txi->txi_transno = 0;
861 mti->mti_no_need_trans = 0;
865 if (mti->mti_has_trans) {
866 /* XXX: currently there are allowed cases, but the wrong cases
867 * are also possible, so better check is needed here */
868 CDEBUG(D_INFO, "More than one transaction "LPU64"\n",
873 mti->mti_has_trans = 1;
874 cfs_spin_lock(&mdt->mdt_lut.lut_translock);
875 if (txn->th_result != 0) {
876 if (mti->mti_transno != 0) {
877 CERROR("Replay transno "LPU64" failed: rc %d\n",
878 mti->mti_transno, txn->th_result);
880 } else if (mti->mti_transno == 0) {
881 mti->mti_transno = ++ mdt->mdt_lut.lut_last_transno;
883 /* should be replay */
884 if (mti->mti_transno > mdt->mdt_lut.lut_last_transno)
885 mdt->mdt_lut.lut_last_transno = mti->mti_transno;
887 cfs_spin_unlock(&mdt->mdt_lut.lut_translock);
888 /* sometimes the reply message has not been successfully packed */
889 LASSERT(req != NULL && req->rq_repmsg != NULL);
891 /** VBR: set new versions */
892 if (txn->th_result == 0)
893 mdt_version_set(mti);
895 /* filling reply data */
896 CDEBUG(D_INODE, "transno = "LPU64", last_committed = "LPU64"\n",
897 mti->mti_transno, req->rq_export->exp_obd->obd_last_committed);
899 req->rq_transno = mti->mti_transno;
900 lustre_msg_set_transno(req->rq_repmsg, mti->mti_transno);
901 /* save transno for the commit callback */
902 txi->txi_transno = mti->mti_transno;
904 /* add separate commit callback for transaction handling because we need
905 * export as parameter */
906 mdt_trans_add_cb(txn, lut_cb_last_committed,
907 class_export_cb_get(mti->mti_exp));
909 return mdt_last_rcvd_update(mti, txn);
912 /* commit callback, need to update last_committed value */
913 static int mdt_txn_commit_cb(const struct lu_env *env,
914 struct thandle *txn, void *cookie)
916 struct mdt_device *mdt = cookie;
917 struct mdt_txn_info *txi;
920 txi = lu_context_key_get(&txn->th_ctx, &mdt_txn_key);
922 /* iterate through all additional callbacks */
923 for (i = 0; i < txi->txi_cb_count; i++) {
924 txi->txi_cb[i].lut_cb_func(&mdt->mdt_lut, txi->txi_transno,
925 txi->txi_cb[i].lut_cb_data, 0);
930 int mdt_fs_setup(const struct lu_env *env, struct mdt_device *mdt,
931 struct obd_device *obd,
932 struct lustre_sb_info *lsi)
939 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_FS_SETUP))
942 /* prepare transactions callbacks */
943 mdt->mdt_txn_cb.dtc_txn_start = mdt_txn_start_cb;
944 mdt->mdt_txn_cb.dtc_txn_stop = mdt_txn_stop_cb;
945 mdt->mdt_txn_cb.dtc_txn_commit = mdt_txn_commit_cb;
946 mdt->mdt_txn_cb.dtc_cookie = mdt;
947 mdt->mdt_txn_cb.dtc_tag = LCT_MD_THREAD;
948 CFS_INIT_LIST_HEAD(&mdt->mdt_txn_cb.dtc_linkage);
950 dt_txn_callback_add(mdt->mdt_bottom, &mdt->mdt_txn_cb);
952 rc = mdt_server_data_init(env, mdt, lsi);
956 o = dt_store_open(env, mdt->mdt_bottom, "", CAPA_KEYS, &fid);
959 rc = mdt_capa_keys_init(env, mdt);
961 GOTO(put_ck_object, rc);
964 CERROR("cannot open %s: rc = %d\n", CAPA_KEYS, rc);
965 GOTO(disconnect_exports, rc);
970 lu_object_put(env, &o->do_lu);
971 mdt->mdt_ck_obj = NULL;
973 class_disconnect_exports(obd);
977 void mdt_fs_cleanup(const struct lu_env *env, struct mdt_device *mdt)
981 /* Remove transaction callback */
982 dt_txn_callback_del(mdt->mdt_bottom, &mdt->mdt_txn_cb);
984 lu_object_put(env, &mdt->mdt_ck_obj->do_lu);
985 mdt->mdt_ck_obj = NULL;
989 /* reconstruction code */
990 static void mdt_steal_ack_locks(struct ptlrpc_request *req)
992 struct obd_export *exp = req->rq_export;
994 struct ptlrpc_reply_state *oldrep;
995 struct ptlrpc_service *svc;
998 /* CAVEAT EMPTOR: spinlock order */
999 cfs_spin_lock(&exp->exp_lock);
1000 cfs_list_for_each (tmp, &exp->exp_outstanding_replies) {
1001 oldrep = cfs_list_entry(tmp, struct ptlrpc_reply_state,
1004 if (oldrep->rs_xid != req->rq_xid)
1007 if (oldrep->rs_opc != lustre_msg_get_opc(req->rq_reqmsg))
1008 CERROR ("Resent req xid "LPU64" has mismatched opc: "
1009 "new %d old %d\n", req->rq_xid,
1010 lustre_msg_get_opc(req->rq_reqmsg),
1013 svc = oldrep->rs_service;
1014 cfs_spin_lock (&svc->srv_rs_lock);
1016 cfs_list_del_init (&oldrep->rs_exp_list);
1018 CWARN("Stealing %d locks from rs %p x"LPD64".t"LPD64
1020 oldrep->rs_nlocks, oldrep,
1021 oldrep->rs_xid, oldrep->rs_transno, oldrep->rs_opc,
1022 libcfs_nid2str(exp->exp_connection->c_peer.nid));
1024 for (i = 0; i < oldrep->rs_nlocks; i++)
1025 ptlrpc_save_lock(req, &oldrep->rs_locks[i],
1026 oldrep->rs_modes[i], 0);
1027 oldrep->rs_nlocks = 0;
1029 DEBUG_REQ(D_HA, req, "stole locks for");
1030 cfs_spin_lock(&oldrep->rs_lock);
1031 ptlrpc_schedule_difficult_reply (oldrep);
1032 cfs_spin_unlock(&oldrep->rs_lock);
1034 cfs_spin_unlock (&svc->srv_rs_lock);
1037 cfs_spin_unlock(&exp->exp_lock);
1041 * VBR: restore versions
1043 void mdt_vbr_reconstruct(struct ptlrpc_request *req,
1044 struct lsd_client_data *lcd)
1046 __u64 pre_versions[4] = {0};
1047 pre_versions[0] = lcd->lcd_pre_versions[0];
1048 pre_versions[1] = lcd->lcd_pre_versions[1];
1049 pre_versions[2] = lcd->lcd_pre_versions[2];
1050 pre_versions[3] = lcd->lcd_pre_versions[3];
1051 lustre_msg_set_versions(req->rq_repmsg, pre_versions);
1054 void mdt_req_from_lcd(struct ptlrpc_request *req,
1055 struct lsd_client_data *lcd)
1057 DEBUG_REQ(D_HA, req, "restoring transno "LPD64"/status %d",
1058 lcd->lcd_last_transno, lcd->lcd_last_result);
1060 if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE ||
1061 lustre_msg_get_opc(req->rq_repmsg) == MDS_DONE_WRITING) {
1062 req->rq_transno = lcd->lcd_last_close_transno;
1063 req->rq_status = lcd->lcd_last_close_result;
1065 req->rq_transno = lcd->lcd_last_transno;
1066 req->rq_status = lcd->lcd_last_result;
1067 mdt_vbr_reconstruct(req, lcd);
1069 if (req->rq_status != 0)
1070 req->rq_transno = 0;
1071 lustre_msg_set_transno(req->rq_repmsg, req->rq_transno);
1072 lustre_msg_set_status(req->rq_repmsg, req->rq_status);
1073 DEBUG_REQ(D_RPCTRACE, req, "restoring transno "LPD64"/status %d",
1074 req->rq_transno, req->rq_status);
1076 mdt_steal_ack_locks(req);
1079 void mdt_reconstruct_generic(struct mdt_thread_info *mti,
1080 struct mdt_lock_handle *lhc)
1082 struct ptlrpc_request *req = mdt_info_req(mti);
1083 struct tg_export_data *ted = &req->rq_export->exp_target_data;
1085 return mdt_req_from_lcd(req, ted->ted_lcd);
1088 static void mdt_reconstruct_create(struct mdt_thread_info *mti,
1089 struct mdt_lock_handle *lhc)
1091 struct ptlrpc_request *req = mdt_info_req(mti);
1092 struct obd_export *exp = req->rq_export;
1093 struct tg_export_data *ted = &exp->exp_target_data;
1094 struct mdt_device *mdt = mti->mti_mdt;
1095 struct mdt_object *child;
1096 struct mdt_body *body;
1099 mdt_req_from_lcd(req, ted->ted_lcd);
1103 /* if no error, so child was created with requested fid */
1104 child = mdt_object_find(mti->mti_env, mdt, mti->mti_rr.rr_fid2);
1105 if (IS_ERR(child)) {
1106 rc = PTR_ERR(child);
1107 LCONSOLE_WARN("Child "DFID" lookup error %d."
1108 " Evicting client %s with export %s.\n",
1109 PFID(mdt_object_fid(child)), rc,
1110 obd_uuid2str(&exp->exp_client_uuid),
1111 obd_export_nid2str(exp));
1112 mdt_export_evict(exp);
1117 body = req_capsule_server_get(mti->mti_pill, &RMF_MDT_BODY);
1118 mti->mti_attr.ma_need = MA_INODE;
1119 mti->mti_attr.ma_valid = 0;
1120 rc = mo_attr_get(mti->mti_env, mdt_object_child(child), &mti->mti_attr);
1121 if (rc == -EREMOTE) {
1122 /* object was created on remote server */
1123 req->rq_status = rc;
1124 body->valid |= OBD_MD_MDS;
1126 mdt_pack_attr2body(mti, body, &mti->mti_attr.ma_attr,
1127 mdt_object_fid(child));
1128 mdt_object_put(mti->mti_env, child);
1131 static void mdt_reconstruct_setattr(struct mdt_thread_info *mti,
1132 struct mdt_lock_handle *lhc)
1134 struct ptlrpc_request *req = mdt_info_req(mti);
1135 struct obd_export *exp = req->rq_export;
1136 struct mdt_export_data *med = &exp->exp_mdt_data;
1137 struct mdt_device *mdt = mti->mti_mdt;
1138 struct mdt_object *obj;
1139 struct mdt_body *body;
1141 mdt_req_from_lcd(req, med->med_ted.ted_lcd);
1145 body = req_capsule_server_get(mti->mti_pill, &RMF_MDT_BODY);
1146 obj = mdt_object_find(mti->mti_env, mdt, mti->mti_rr.rr_fid1);
1148 int rc = PTR_ERR(obj);
1149 LCONSOLE_WARN(""DFID" lookup error %d."
1150 " Evicting client %s with export %s.\n",
1151 PFID(mdt_object_fid(obj)), rc,
1152 obd_uuid2str(&exp->exp_client_uuid),
1153 obd_export_nid2str(exp));
1154 mdt_export_evict(exp);
1158 mti->mti_attr.ma_need = MA_INODE;
1159 mti->mti_attr.ma_valid = 0;
1160 mo_attr_get(mti->mti_env, mdt_object_child(obj), &mti->mti_attr);
1161 mdt_pack_attr2body(mti, body, &mti->mti_attr.ma_attr,
1162 mdt_object_fid(obj));
1163 if (mti->mti_ioepoch && (mti->mti_ioepoch->flags & MF_EPOCH_OPEN)) {
1164 struct mdt_file_data *mfd;
1165 struct mdt_body *repbody;
1167 repbody = req_capsule_server_get(mti->mti_pill, &RMF_MDT_BODY);
1168 repbody->ioepoch = obj->mot_ioepoch;
1169 cfs_spin_lock(&med->med_open_lock);
1170 cfs_list_for_each_entry(mfd, &med->med_open_head, mfd_list) {
1171 if (mfd->mfd_xid == req->rq_xid)
1174 LASSERT(&mfd->mfd_list != &med->med_open_head);
1175 cfs_spin_unlock(&med->med_open_lock);
1176 repbody->handle.cookie = mfd->mfd_handle.h_cookie;
1179 mdt_object_put(mti->mti_env, obj);
1182 typedef void (*mdt_reconstructor)(struct mdt_thread_info *mti,
1183 struct mdt_lock_handle *lhc);
1185 static mdt_reconstructor reconstructors[REINT_MAX] = {
1186 [REINT_SETATTR] = mdt_reconstruct_setattr,
1187 [REINT_CREATE] = mdt_reconstruct_create,
1188 [REINT_LINK] = mdt_reconstruct_generic,
1189 [REINT_UNLINK] = mdt_reconstruct_generic,
1190 [REINT_RENAME] = mdt_reconstruct_generic,
1191 [REINT_OPEN] = mdt_reconstruct_open,
1192 [REINT_SETXATTR] = mdt_reconstruct_generic
1195 void mdt_reconstruct(struct mdt_thread_info *mti,
1196 struct mdt_lock_handle *lhc)
1199 reconstructors[mti->mti_rr.rr_opcode](mti, lhc);