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 return dt_trans_start(env, mdt->mdt_bottom, p);
129 void mdt_trans_stop(const struct lu_env *env,
130 struct mdt_device *mdt, struct thandle *th)
132 struct mdt_thread_info *mti;
134 mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
135 /* export can require sync operations */
136 if (mti->mti_exp != NULL)
137 th->th_sync |= mti->mti_exp->exp_need_sync;
138 dt_trans_stop(env, mdt->mdt_bottom, th);
141 static inline int mdt_last_rcvd_header_read(const struct lu_env *env,
142 struct mdt_device *mdt)
144 struct mdt_thread_info *mti;
147 mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
150 rc = dt_record_read(env, mdt->mdt_lut.lut_last_rcvd,
151 mdt_buf(env, &mti->mti_lsd, sizeof(mti->mti_lsd)),
154 lsd_le_to_cpu(&mti->mti_lsd, &mdt->mdt_lut.lut_lsd);
156 CDEBUG(D_INFO, "read last_rcvd header rc = %d, uuid = %s, "
157 "last_transno = "LPU64"\n", rc, mdt->mdt_lut.lut_lsd.lsd_uuid,
158 mdt->mdt_lut.lut_lsd.lsd_last_transno);
162 static int mdt_last_rcvd_header_write(const struct lu_env *env,
163 struct mdt_device *mdt,
166 struct mdt_thread_info *mti;
170 mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
173 lsd_cpu_to_le(&mdt->mdt_lut.lut_lsd, &mti->mti_lsd);
175 rc = dt_record_write(env, mdt->mdt_lut.lut_last_rcvd,
176 mdt_buf_const(env, &mti->mti_lsd,
177 sizeof(mti->mti_lsd)),
180 CDEBUG(D_INFO, "write last_rcvd header rc = %d, uuid = %s, "
181 "last_transno = "LPU64"\n", rc, mdt->mdt_lut.lut_lsd.lsd_uuid,
182 mdt->mdt_lut.lut_lsd.lsd_last_transno);
187 static int mdt_last_rcvd_read(const struct lu_env *env, struct mdt_device *mdt,
188 struct lsd_client_data *lcd, loff_t *off,
191 struct mdt_thread_info *mti;
192 struct lsd_client_data *tmp;
195 mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
197 rc = dt_record_read(env, mdt->mdt_lut.lut_last_rcvd,
198 mdt_buf(env, tmp, sizeof(*tmp)), off);
200 check_lcd(mdt2obd_dev(mdt)->obd_name, index, tmp);
201 lcd_le_to_cpu(tmp, lcd);
204 CDEBUG(D_INFO, "read lcd @%d rc = %d, uuid = %s, last_transno = "LPU64
205 ", last_xid = "LPU64", last_result = %u, last_data = %u, "
206 "last_close_transno = "LPU64", last_close_xid = "LPU64", "
207 "last_close_result = %u\n", (int)(*off - sizeof(*tmp)),
208 rc, lcd->lcd_uuid, lcd->lcd_last_transno, lcd->lcd_last_xid,
209 lcd->lcd_last_result, lcd->lcd_last_data,
210 lcd->lcd_last_close_transno, lcd->lcd_last_close_xid,
211 lcd->lcd_last_close_result);
215 static int mdt_last_rcvd_write(const struct lu_env *env,
216 struct mdt_device *mdt,
217 struct lsd_client_data *lcd,
218 loff_t *off, struct thandle *th)
220 struct mdt_thread_info *mti;
221 struct lsd_client_data *tmp;
225 mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
228 lcd_cpu_to_le(lcd, tmp);
230 rc = dt_record_write(env, mdt->mdt_lut.lut_last_rcvd,
231 mdt_buf_const(env, tmp, sizeof(*tmp)), off, th);
233 CDEBUG(D_INFO, "write lcd @%d rc = %d, uuid = %s, last_transno = "LPU64
234 ", last_xid = "LPU64", last_result = %u, last_data = %u, "
235 "last_close_transno = "LPU64", last_close_xid = "LPU64" ,"
236 "last_close_result = %u\n", (int)(*off - sizeof(*tmp)),
237 rc, lcd->lcd_uuid, lcd->lcd_last_transno, lcd->lcd_last_xid,
238 lcd->lcd_last_result, lcd->lcd_last_data,
239 lcd->lcd_last_close_transno, lcd->lcd_last_close_xid,
240 lcd->lcd_last_close_result);
244 static int mdt_clients_data_init(const struct lu_env *env,
245 struct mdt_device *mdt,
246 unsigned long last_size)
248 struct lr_server_data *lsd = &mdt->mdt_lut.lut_lsd;
249 struct lsd_client_data *lcd;
250 struct obd_device *obd = mdt2obd_dev(mdt);
260 /* When we do a clean MDS shutdown, we save the last_transno into
261 * the header. If we find clients with higher last_transno values
262 * then those clients may need recovery done. */
263 LASSERT(cfs_atomic_read(&obd->obd_req_replay_clients) == 0);
264 for (cl_idx = 0, off = lsd->lsd_client_start;
265 off < last_size; cl_idx++) {
267 struct obd_export *exp;
268 struct mdt_thread_info *mti;
270 off = lsd->lsd_client_start +
271 cl_idx * lsd->lsd_client_size;
273 rc = mdt_last_rcvd_read(env, mdt, lcd, &off, cl_idx);
275 CERROR("error reading MDS %s idx %d, off %llu: rc %d\n",
276 LAST_RCVD, cl_idx, off, rc);
278 break; /* read error shouldn't cause startup to fail */
281 if (lcd->lcd_uuid[0] == '\0') {
282 CDEBUG(D_INFO, "skipping zeroed client at offset %d\n",
287 last_transno = lcd_last_transno(lcd);
289 /* These exports are cleaned up by mdt_obd_disconnect(), so
290 * they need to be set up like real exports as
291 * mdt_obd_connect() does.
293 CDEBUG(D_HA, "RCVRNG CLIENT uuid: %s idx: %d lr: "LPU64
294 " srv lr: "LPU64" lx: "LPU64"\n", lcd->lcd_uuid, cl_idx,
295 last_transno, lsd->lsd_last_transno,
298 exp = class_new_export(obd, (struct obd_uuid *)lcd->lcd_uuid);
300 if (PTR_ERR(exp) == -EALREADY) {
301 /* export already exists, zero out this one */
302 CERROR("Duplicate export %s!\n", lcd->lcd_uuid);
305 GOTO(err_client, rc = PTR_ERR(exp));
308 mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
309 LASSERT(mti != NULL);
311 /* copy on-disk lcd to the export */
312 *exp->exp_target_data.ted_lcd = *lcd;
313 rc = mdt_client_add(env, mdt, cl_idx);
314 /* can't fail existing */
315 LASSERTF(rc == 0, "rc = %d\n", rc);
316 /* VBR: set export last committed version */
317 exp->exp_last_committed = last_transno;
318 cfs_spin_lock(&exp->exp_lock);
319 exp->exp_connecting = 0;
320 exp->exp_in_recovery = 0;
321 cfs_spin_unlock(&exp->exp_lock);
322 obd->obd_max_recoverable_clients++;
323 class_export_put(exp);
325 CDEBUG(D_OTHER, "client at idx %d has last_transno="LPU64"\n",
326 cl_idx, last_transno);
327 /* protect __u64 value update */
328 cfs_spin_lock(&mdt->mdt_lut.lut_translock);
329 mdt->mdt_lut.lut_last_transno = max(last_transno,
330 mdt->mdt_lut.lut_last_transno);
331 cfs_spin_unlock(&mdt->mdt_lut.lut_translock);
339 static int mdt_server_data_init(const struct lu_env *env,
340 struct mdt_device *mdt,
341 struct lustre_sb_info *lsi)
343 struct lr_server_data *lsd = &mdt->mdt_lut.lut_lsd;
344 struct lsd_client_data *lcd = NULL;
345 struct obd_device *obd = mdt2obd_dev(mdt);
346 struct mdt_thread_info *mti;
347 struct dt_object *obj;
349 struct lustre_disk_data *ldd;
350 unsigned long last_rcvd_size;
355 /* ensure padding in the struct is the correct size */
356 CLASSERT(offsetof(struct lr_server_data, lsd_padding) +
357 sizeof(lsd->lsd_padding) == LR_SERVER_SIZE);
358 CLASSERT(offsetof(struct lsd_client_data, lcd_padding) +
359 sizeof(lcd->lcd_padding) == LR_CLIENT_SIZE);
361 mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
362 LASSERT(mti != NULL);
363 la = &mti->mti_attr.ma_attr;
365 obj = mdt->mdt_lut.lut_last_rcvd;
366 rc = obj->do_ops->do_attr_get(env, obj, la, BYPASS_CAPA);
370 last_rcvd_size = (unsigned long)la->la_size;
372 if (last_rcvd_size == 0) {
373 LCONSOLE_WARN("%s: new disk, initializing\n", obd->obd_name);
375 memcpy(lsd->lsd_uuid, obd->obd_uuid.uuid,
376 sizeof(lsd->lsd_uuid));
377 lsd->lsd_last_transno = 0;
378 lsd->lsd_mount_count = 0;
379 lsd->lsd_server_size = LR_SERVER_SIZE;
380 lsd->lsd_client_start = LR_CLIENT_START;
381 lsd->lsd_client_size = LR_CLIENT_SIZE;
382 lsd->lsd_feature_compat = OBD_COMPAT_MDT;
383 lsd->lsd_feature_rocompat = OBD_ROCOMPAT_LOVOBJID;
384 lsd->lsd_feature_incompat = OBD_INCOMPAT_MDT |
385 OBD_INCOMPAT_COMMON_LR;
387 LCONSOLE_WARN("%s: used disk, loading\n", obd->obd_name);
388 rc = mdt_last_rcvd_header_read(env, mdt);
390 CERROR("error reading MDS %s: rc %d\n", LAST_RCVD, rc);
393 if (strcmp(lsd->lsd_uuid, obd->obd_uuid.uuid) != 0) {
394 LCONSOLE_ERROR_MSG(0x157, "Trying to start OBD %s using"
395 "the wrong disk %s. Were the /dev/ "
396 "assignments rearranged?\n",
397 obd->obd_uuid.uuid, lsd->lsd_uuid);
398 GOTO(out, rc = -EINVAL);
400 lsd->lsd_feature_compat |= OBD_COMPAT_MDT;
401 lsd->lsd_feature_incompat |= OBD_INCOMPAT_MDT |
402 OBD_INCOMPAT_COMMON_LR;
404 mount_count = lsd->lsd_mount_count;
408 if (lsd->lsd_feature_incompat & ~MDT_INCOMPAT_SUPP) {
409 CERROR("%s: unsupported incompat filesystem feature(s) %x\n",
411 lsd->lsd_feature_incompat & ~MDT_INCOMPAT_SUPP);
412 GOTO(out, rc = -EINVAL);
414 if (lsd->lsd_feature_rocompat & ~MDT_ROCOMPAT_SUPP) {
415 CERROR("%s: unsupported read-only filesystem feature(s) %x\n",
417 lsd->lsd_feature_rocompat & ~MDT_ROCOMPAT_SUPP);
418 /* XXX: Do something like remount filesystem read-only */
419 GOTO(out, rc = -EINVAL);
421 /** Interop: evict all clients at first boot with 1.8 last_rcvd */
422 if (!(lsd->lsd_feature_compat & OBD_COMPAT_20)) {
423 if (last_rcvd_size > lsd->lsd_client_start) {
424 LCONSOLE_WARN("Mounting %s at first time on 1.8 FS, "
425 "remove all clients for interop needs\n",
427 simple_truncate(lsi->lsi_srv_mnt->mnt_sb->s_root,
428 lsi->lsi_srv_mnt, LAST_RCVD,
429 lsd->lsd_client_start);
430 last_rcvd_size = lsd->lsd_client_start;
432 /** set 2.0 flag to upgrade/downgrade between 1.8 and 2.0 */
433 lsd->lsd_feature_compat |= OBD_COMPAT_20;
436 if (ldd->ldd_flags & LDD_F_IAM_DIR)
437 lsd->lsd_feature_incompat |= OBD_INCOMPAT_IAM_DIR;
439 lsd->lsd_feature_incompat |= OBD_INCOMPAT_FID;
441 cfs_spin_lock(&mdt->mdt_lut.lut_translock);
442 mdt->mdt_lut.lut_last_transno = lsd->lsd_last_transno;
443 cfs_spin_unlock(&mdt->mdt_lut.lut_translock);
445 CDEBUG(D_INODE, "========BEGIN DUMPING LAST_RCVD========\n");
446 CDEBUG(D_INODE, "%s: server last_transno: "LPU64"\n",
447 obd->obd_name, mdt->mdt_lut.lut_last_transno);
448 CDEBUG(D_INODE, "%s: server mount_count: "LPU64"\n",
449 obd->obd_name, mount_count + 1);
450 CDEBUG(D_INODE, "%s: server data size: %u\n",
451 obd->obd_name, lsd->lsd_server_size);
452 CDEBUG(D_INODE, "%s: per-client data start: %u\n",
453 obd->obd_name, lsd->lsd_client_start);
454 CDEBUG(D_INODE, "%s: per-client data size: %u\n",
455 obd->obd_name, lsd->lsd_client_size);
456 CDEBUG(D_INODE, "%s: last_rcvd size: %lu\n",
457 obd->obd_name, last_rcvd_size);
458 CDEBUG(D_INODE, "%s: last_rcvd clients: %lu\n", obd->obd_name,
459 last_rcvd_size <= lsd->lsd_client_start ? 0 :
460 (last_rcvd_size - lsd->lsd_client_start) /
461 lsd->lsd_client_size);
462 CDEBUG(D_INODE, "========END DUMPING LAST_RCVD========\n");
464 if (!lsd->lsd_server_size || !lsd->lsd_client_start ||
465 !lsd->lsd_client_size) {
466 CERROR("Bad last_rcvd contents!\n");
467 GOTO(out, rc = -EINVAL);
470 rc = mdt_clients_data_init(env, mdt, last_rcvd_size);
472 GOTO(err_client, rc);
474 cfs_spin_lock(&mdt->mdt_lut.lut_translock);
475 /* obd_last_committed is used for compatibility
476 * with other lustre recovery code */
477 obd->obd_last_committed = mdt->mdt_lut.lut_last_transno;
478 cfs_spin_unlock(&mdt->mdt_lut.lut_translock);
480 obd->u.obt.obt_mount_count = mount_count + 1;
481 obd->u.obt.obt_instance = (__u32)obd->u.obt.obt_mount_count;
482 lsd->lsd_mount_count = obd->u.obt.obt_mount_count;
484 /* save it, so mount count and last_transno is current */
485 rc = mdt_server_data_update(env, mdt);
487 GOTO(err_client, rc);
492 class_disconnect_exports(obd);
497 static int mdt_server_data_update(const struct lu_env *env,
498 struct mdt_device *mdt)
500 struct mdt_thread_info *mti;
504 mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
506 mdt_trans_credit_init(env, mdt, MDT_TXN_LAST_RCVD_WRITE_OP);
507 th = mdt_trans_start(env, mdt);
511 CDEBUG(D_SUPER, "MDS mount_count is "LPU64", last_transno is "LPU64"\n",
512 mdt->mdt_lut.lut_obd->u.obt.obt_mount_count,
513 mdt->mdt_lut.lut_last_transno);
515 cfs_spin_lock(&mdt->mdt_lut.lut_translock);
516 mdt->mdt_lut.lut_lsd.lsd_last_transno = mdt->mdt_lut.lut_last_transno;
517 cfs_spin_unlock(&mdt->mdt_lut.lut_translock);
519 rc = mdt_last_rcvd_header_write(env, mdt, th);
520 mdt_trans_stop(env, mdt, th);
525 int mdt_client_new(const struct lu_env *env, struct mdt_device *mdt)
527 unsigned long *bitmap = mdt->mdt_lut.lut_client_bitmap;
528 struct mdt_thread_info *mti;
529 struct tg_export_data *ted;
530 struct lr_server_data *lsd = &mdt->mdt_lut.lut_lsd;
531 struct obd_device *obd = mdt2obd_dev(mdt);
538 mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
539 LASSERT(mti != NULL);
541 ted = &mti->mti_exp->exp_target_data;
543 LASSERT(bitmap != NULL);
544 if (!strcmp(ted->ted_lcd->lcd_uuid, obd->obd_uuid.uuid))
547 /* the bitmap operations can handle cl_idx > sizeof(long) * 8, so
548 * there's no need for extra complication here
550 cfs_spin_lock(&mdt->mdt_lut.lut_client_bitmap_lock);
551 cl_idx = cfs_find_first_zero_bit(bitmap, LR_MAX_CLIENTS);
552 if (cl_idx >= LR_MAX_CLIENTS ||
553 OBD_FAIL_CHECK(OBD_FAIL_MDS_CLIENT_ADD)) {
554 CERROR("no room for %u clients - fix LR_MAX_CLIENTS\n",
556 cfs_spin_unlock(&mdt->mdt_lut.lut_client_bitmap_lock);
559 cfs_set_bit(cl_idx, bitmap);
560 cfs_spin_unlock(&mdt->mdt_lut.lut_client_bitmap_lock);
562 CDEBUG(D_INFO, "client at idx %d with UUID '%s' added\n",
563 cl_idx, ted->ted_lcd->lcd_uuid);
565 ted->ted_lr_idx = cl_idx;
566 ted->ted_lr_off = lsd->lsd_client_start +
567 (cl_idx * lsd->lsd_client_size);
568 cfs_init_mutex(&ted->ted_lcd_lock);
570 LASSERTF(ted->ted_lr_off > 0, "ted_lr_off = %llu\n", ted->ted_lr_off);
572 /* Write new client data. */
573 off = ted->ted_lr_off;
575 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_CLIENT_ADD))
578 mdt_trans_credit_init(env, mdt, MDT_TXN_LAST_RCVD_WRITE_OP);
580 th = mdt_trans_start(env, mdt);
585 * Until this operations will be committed the sync is needed
586 * for this export. This should be done _after_ starting the
587 * transaction so that many connecting clients will not bring
588 * server down with lots of sync writes.
590 rc = lut_new_client_cb_add(th, mti->mti_exp);
592 /* can't add callback, do sync now */
595 cfs_spin_lock(&mti->mti_exp->exp_lock);
596 mti->mti_exp->exp_need_sync = 1;
597 cfs_spin_unlock(&mti->mti_exp->exp_lock);
600 rc = mdt_last_rcvd_write(env, mdt, ted->ted_lcd, &off, th);
601 CDEBUG(D_INFO, "wrote client lcd at idx %u off %llu (len %u)\n",
602 cl_idx, ted->ted_lr_off, (int)sizeof(*(ted->ted_lcd)));
603 mdt_trans_stop(env, mdt, th);
608 /* Add client data to the MDS. We use a bitmap to locate a free space
609 * in the last_rcvd file if cl_off is -1 (i.e. a new client).
610 * Otherwise, we just have to read the data from the last_rcvd file and
611 * we know its offset.
613 * It should not be possible to fail adding an existing client - otherwise
614 * mdt_init_server_data() callsite needs to be fixed.
616 int mdt_client_add(const struct lu_env *env,
617 struct mdt_device *mdt, int cl_idx)
619 struct mdt_thread_info *mti;
620 struct tg_export_data *ted;
621 unsigned long *bitmap = mdt->mdt_lut.lut_client_bitmap;
622 struct obd_device *obd = mdt2obd_dev(mdt);
623 struct lr_server_data *lsd = &mdt->mdt_lut.lut_lsd;
627 mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
628 LASSERT(mti != NULL);
630 ted = &mti->mti_exp->exp_target_data;
632 LASSERT(bitmap != NULL);
633 LASSERTF(cl_idx >= 0, "%d\n", cl_idx);
635 if (!strcmp(ted->ted_lcd->lcd_uuid, obd->obd_uuid.uuid))
638 cfs_spin_lock(&mdt->mdt_lut.lut_client_bitmap_lock);
639 if (cfs_test_and_set_bit(cl_idx, bitmap)) {
640 CERROR("MDS client %d: bit already set in bitmap!!\n",
644 cfs_spin_unlock(&mdt->mdt_lut.lut_client_bitmap_lock);
646 CDEBUG(D_INFO, "client at idx %d with UUID '%s' added\n",
647 cl_idx, ted->ted_lcd->lcd_uuid);
649 ted->ted_lr_idx = cl_idx;
650 ted->ted_lr_off = lsd->lsd_client_start +
651 (cl_idx * lsd->lsd_client_size);
652 cfs_init_mutex(&ted->ted_lcd_lock);
654 LASSERTF(ted->ted_lr_off > 0, "ted_lr_off = %llu\n", ted->ted_lr_off);
659 int mdt_client_del(const struct lu_env *env, struct mdt_device *mdt)
661 struct mdt_thread_info *mti;
662 struct tg_export_data *ted;
663 struct obd_device *obd = mdt2obd_dev(mdt);
664 struct obd_export *exp;
670 mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
671 LASSERT(mti != NULL);
674 ted = &exp->exp_target_data;
678 /* XXX: If lcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
679 if (!strcmp(ted->ted_lcd->lcd_uuid, obd->obd_uuid.uuid))
682 CDEBUG(D_INFO, "freeing client at idx %u, offset %lld\n",
683 ted->ted_lr_idx, ted->ted_lr_off);
685 off = ted->ted_lr_off;
688 * Don't clear ted_lr_idx here as it is likely also unset. At worst we
689 * leak a client slot that will be cleaned on the next recovery.
692 CERROR("client idx %d has offset %lld\n",
693 ted->ted_lr_idx, off);
694 GOTO(free, rc = -EINVAL);
698 * Clear the bit _after_ zeroing out the client so we don't race with
699 * mdt_client_add and zero out new clients.
701 if (!cfs_test_bit(ted->ted_lr_idx, mdt->mdt_lut.lut_client_bitmap)) {
702 CERROR("MDT client %u: bit already clear in bitmap!!\n",
707 /* Make sure the server's last_transno is up to date.
708 * This should be done before zeroing client slot so last_transno will
709 * be in server data or in client data in case of failure */
710 mdt_server_data_update(env, mdt);
712 mdt_trans_credit_init(env, mdt, MDT_TXN_LAST_RCVD_WRITE_OP);
713 th = mdt_trans_start(env, mdt);
715 GOTO(free, rc = PTR_ERR(th));
717 cfs_mutex_down(&ted->ted_lcd_lock);
718 memset(ted->ted_lcd->lcd_uuid, 0, sizeof ted->ted_lcd->lcd_uuid);
719 rc = mdt_last_rcvd_write(env, mdt, ted->ted_lcd, &off, th);
720 cfs_mutex_up(&ted->ted_lcd_lock);
721 mdt_trans_stop(env, mdt, th);
723 CDEBUG(rc == 0 ? D_INFO : D_ERROR, "Zeroing out client idx %u in "
724 "%s, rc %d\n", ted->ted_lr_idx, LAST_RCVD, rc);
731 * last_rcvd & last_committed update callbacks
733 static int mdt_last_rcvd_update(struct mdt_thread_info *mti,
736 struct mdt_device *mdt = mti->mti_mdt;
737 struct ptlrpc_request *req = mdt_info_req(mti);
738 struct tg_export_data *ted;
739 struct lsd_client_data *lcd;
742 __s32 rc = th->th_result;
746 LASSERT(req->rq_export);
748 ted = &req->rq_export->exp_target_data;
751 cfs_mutex_down(&ted->ted_lcd_lock);
753 /* if the export has already been disconnected, we have no last_rcvd slot,
754 * update server data with latest transno then */
756 cfs_mutex_up(&ted->ted_lcd_lock);
757 CWARN("commit transaction for disconnected client %s: rc %d\n",
758 req->rq_export->exp_client_uuid.uuid, rc);
759 err = mdt_last_rcvd_header_write(mti->mti_env, mdt, th);
763 off = ted->ted_lr_off;
764 LASSERT(ergo(mti->mti_transno == 0, rc != 0));
765 if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE ||
766 lustre_msg_get_opc(req->rq_reqmsg) == MDS_DONE_WRITING) {
767 if (mti->mti_transno != 0) {
768 if (lcd->lcd_last_close_transno > mti->mti_transno) {
769 CERROR("Trying to overwrite bigger transno:"
770 "on-disk: "LPU64", new: "LPU64" "
771 "replay: %d. see LU-617.\n",
772 lcd->lcd_last_close_transno,
773 mti->mti_transno, req_is_replay(req));
774 if (req_is_replay(req)) {
775 cfs_spin_lock(&req->rq_export->exp_lock);
776 req->rq_export->exp_vbr_failed = 1;
777 cfs_spin_unlock(&req->rq_export->exp_lock);
779 cfs_mutex_up(&ted->ted_lcd_lock);
780 RETURN(req_is_replay(req) ? -EOVERFLOW : 0);
782 lcd->lcd_last_close_transno = mti->mti_transno;
784 lcd->lcd_last_close_xid = req->rq_xid;
785 lcd->lcd_last_close_result = rc;
787 /* VBR: save versions in last_rcvd for reconstruct. */
788 __u64 *pre_versions = lustre_msg_get_versions(req->rq_repmsg);
790 lcd->lcd_pre_versions[0] = pre_versions[0];
791 lcd->lcd_pre_versions[1] = pre_versions[1];
792 lcd->lcd_pre_versions[2] = pre_versions[2];
793 lcd->lcd_pre_versions[3] = pre_versions[3];
795 if (mti->mti_transno != 0) {
796 if (lcd->lcd_last_transno > mti->mti_transno) {
797 CERROR("Trying to overwrite bigger transno:"
798 "on-disk: "LPU64", new: "LPU64" "
799 "replay: %d. see LU-617.\n",
800 lcd->lcd_last_transno,
801 mti->mti_transno, req_is_replay(req));
802 if (req_is_replay(req)) {
803 cfs_spin_lock(&req->rq_export->exp_lock);
804 req->rq_export->exp_vbr_failed = 1;
805 cfs_spin_unlock(&req->rq_export->exp_lock);
807 cfs_mutex_up(&ted->ted_lcd_lock);
808 RETURN(req_is_replay(req) ? -EOVERFLOW : 0);
810 lcd->lcd_last_transno = mti->mti_transno;
812 lcd->lcd_last_xid = req->rq_xid;
813 lcd->lcd_last_result = rc;
814 /*XXX: save intent_disposition in mdt_thread_info?
815 * also there is bug - intent_dispostion is __u64,
816 * see struct ldlm_reply->lock_policy_res1; */
817 lcd->lcd_last_data = mti->mti_opdata;
821 CERROR("client idx %d has offset %lld\n", ted->ted_lr_idx, off);
824 err = mdt_last_rcvd_write(mti->mti_env, mdt, lcd, &off, th);
826 cfs_mutex_up(&ted->ted_lcd_lock);
830 extern struct lu_context_key mdt_thread_key;
832 /* add credits for last_rcvd update */
833 static int mdt_txn_start_cb(const struct lu_env *env,
834 struct txn_param *param, void *cookie)
836 struct mdt_device *mdt = cookie;
838 param->tp_credits += mdt_trans_credit_get(env, mdt,
839 MDT_TXN_LAST_RCVD_WRITE_OP);
843 /* Set new object versions */
844 static void mdt_version_set(struct mdt_thread_info *info)
846 if (info->mti_mos != NULL) {
847 mo_version_set(info->mti_env, mdt_object_child(info->mti_mos),
849 info->mti_mos = NULL;
853 /* Update last_rcvd records with latests transaction data */
854 static int mdt_txn_stop_cb(const struct lu_env *env,
855 struct thandle *txn, void *cookie)
857 struct mdt_device *mdt = cookie;
858 struct mdt_thread_info *mti;
859 struct ptlrpc_request *req;
861 mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
862 req = mdt_info_req(mti);
864 if (mti->mti_mdt == NULL || req == NULL || mti->mti_no_need_trans) {
865 mti->mti_no_need_trans = 0;
869 if (mti->mti_has_trans) {
870 /* XXX: currently there are allowed cases, but the wrong cases
871 * are also possible, so better check is needed here */
872 CDEBUG(D_INFO, "More than one transaction "LPU64"\n",
877 mti->mti_has_trans = 1;
878 cfs_spin_lock(&mdt->mdt_lut.lut_translock);
879 if (txn->th_result != 0) {
880 if (mti->mti_transno != 0) {
881 CERROR("Replay transno "LPU64" failed: rc %d\n",
882 mti->mti_transno, txn->th_result);
884 } else if (mti->mti_transno == 0) {
885 mti->mti_transno = ++ mdt->mdt_lut.lut_last_transno;
887 /* should be replay */
888 if (mti->mti_transno > mdt->mdt_lut.lut_last_transno)
889 mdt->mdt_lut.lut_last_transno = mti->mti_transno;
891 cfs_spin_unlock(&mdt->mdt_lut.lut_translock);
892 /* sometimes the reply message has not been successfully packed */
893 LASSERT(req != NULL && req->rq_repmsg != NULL);
895 /** VBR: set new versions */
896 if (txn->th_result == 0)
897 mdt_version_set(mti);
899 /* filling reply data */
900 CDEBUG(D_INODE, "transno = "LPU64", last_committed = "LPU64"\n",
901 mti->mti_transno, req->rq_export->exp_obd->obd_last_committed);
903 req->rq_transno = mti->mti_transno;
904 lustre_msg_set_transno(req->rq_repmsg, mti->mti_transno);
905 /* if can't add callback, do sync write */
906 txn->th_sync = !!lut_last_commit_cb_add(txn, &mdt->mdt_lut,
909 return mdt_last_rcvd_update(mti, txn);
912 int mdt_fs_setup(const struct lu_env *env, struct mdt_device *mdt,
913 struct obd_device *obd,
914 struct lustre_sb_info *lsi)
921 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_FS_SETUP))
924 /* prepare transactions callbacks */
925 mdt->mdt_txn_cb.dtc_txn_start = mdt_txn_start_cb;
926 mdt->mdt_txn_cb.dtc_txn_stop = mdt_txn_stop_cb;
927 mdt->mdt_txn_cb.dtc_txn_commit = NULL;
928 mdt->mdt_txn_cb.dtc_cookie = mdt;
929 mdt->mdt_txn_cb.dtc_tag = LCT_MD_THREAD;
930 CFS_INIT_LIST_HEAD(&mdt->mdt_txn_cb.dtc_linkage);
932 dt_txn_callback_add(mdt->mdt_bottom, &mdt->mdt_txn_cb);
934 rc = mdt_server_data_init(env, mdt, lsi);
938 o = dt_store_open(env, mdt->mdt_bottom, "", CAPA_KEYS, &fid);
941 rc = mdt_capa_keys_init(env, mdt);
943 GOTO(put_ck_object, rc);
946 CERROR("cannot open %s: rc = %d\n", CAPA_KEYS, rc);
947 GOTO(disconnect_exports, rc);
952 lu_object_put(env, &o->do_lu);
953 mdt->mdt_ck_obj = NULL;
955 class_disconnect_exports(obd);
959 void mdt_fs_cleanup(const struct lu_env *env, struct mdt_device *mdt)
963 /* Remove transaction callback */
964 dt_txn_callback_del(mdt->mdt_bottom, &mdt->mdt_txn_cb);
966 lu_object_put(env, &mdt->mdt_ck_obj->do_lu);
967 mdt->mdt_ck_obj = NULL;
971 /* reconstruction code */
972 static void mdt_steal_ack_locks(struct ptlrpc_request *req)
974 struct obd_export *exp = req->rq_export;
976 struct ptlrpc_reply_state *oldrep;
977 struct ptlrpc_service *svc;
980 /* CAVEAT EMPTOR: spinlock order */
981 cfs_spin_lock(&exp->exp_lock);
982 cfs_list_for_each (tmp, &exp->exp_outstanding_replies) {
983 oldrep = cfs_list_entry(tmp, struct ptlrpc_reply_state,
986 if (oldrep->rs_xid != req->rq_xid)
989 if (oldrep->rs_opc != lustre_msg_get_opc(req->rq_reqmsg))
990 CERROR ("Resent req xid "LPU64" has mismatched opc: "
991 "new %d old %d\n", req->rq_xid,
992 lustre_msg_get_opc(req->rq_reqmsg),
995 svc = oldrep->rs_service;
996 cfs_spin_lock (&svc->srv_rs_lock);
998 cfs_list_del_init (&oldrep->rs_exp_list);
1000 CWARN("Stealing %d locks from rs %p x"LPD64".t"LPD64
1002 oldrep->rs_nlocks, oldrep,
1003 oldrep->rs_xid, oldrep->rs_transno, oldrep->rs_opc,
1004 libcfs_nid2str(exp->exp_connection->c_peer.nid));
1006 for (i = 0; i < oldrep->rs_nlocks; i++)
1007 ptlrpc_save_lock(req, &oldrep->rs_locks[i],
1008 oldrep->rs_modes[i], 0);
1009 oldrep->rs_nlocks = 0;
1011 DEBUG_REQ(D_HA, req, "stole locks for");
1012 cfs_spin_lock(&oldrep->rs_lock);
1013 ptlrpc_schedule_difficult_reply (oldrep);
1014 cfs_spin_unlock(&oldrep->rs_lock);
1016 cfs_spin_unlock (&svc->srv_rs_lock);
1019 cfs_spin_unlock(&exp->exp_lock);
1023 * VBR: restore versions
1025 void mdt_vbr_reconstruct(struct ptlrpc_request *req,
1026 struct lsd_client_data *lcd)
1028 __u64 pre_versions[4] = {0};
1029 pre_versions[0] = lcd->lcd_pre_versions[0];
1030 pre_versions[1] = lcd->lcd_pre_versions[1];
1031 pre_versions[2] = lcd->lcd_pre_versions[2];
1032 pre_versions[3] = lcd->lcd_pre_versions[3];
1033 lustre_msg_set_versions(req->rq_repmsg, pre_versions);
1036 void mdt_req_from_lcd(struct ptlrpc_request *req,
1037 struct lsd_client_data *lcd)
1039 DEBUG_REQ(D_HA, req, "restoring transno "LPD64"/status %d",
1040 lcd->lcd_last_transno, lcd->lcd_last_result);
1042 if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE ||
1043 lustre_msg_get_opc(req->rq_repmsg) == MDS_DONE_WRITING) {
1044 req->rq_transno = lcd->lcd_last_close_transno;
1045 req->rq_status = lcd->lcd_last_close_result;
1047 req->rq_transno = lcd->lcd_last_transno;
1048 req->rq_status = lcd->lcd_last_result;
1049 mdt_vbr_reconstruct(req, lcd);
1051 if (req->rq_status != 0)
1052 req->rq_transno = 0;
1053 lustre_msg_set_transno(req->rq_repmsg, req->rq_transno);
1054 lustre_msg_set_status(req->rq_repmsg, req->rq_status);
1055 DEBUG_REQ(D_RPCTRACE, req, "restoring transno "LPD64"/status %d",
1056 req->rq_transno, req->rq_status);
1058 mdt_steal_ack_locks(req);
1061 void mdt_reconstruct_generic(struct mdt_thread_info *mti,
1062 struct mdt_lock_handle *lhc)
1064 struct ptlrpc_request *req = mdt_info_req(mti);
1065 struct tg_export_data *ted = &req->rq_export->exp_target_data;
1067 return mdt_req_from_lcd(req, ted->ted_lcd);
1070 static void mdt_reconstruct_create(struct mdt_thread_info *mti,
1071 struct mdt_lock_handle *lhc)
1073 struct ptlrpc_request *req = mdt_info_req(mti);
1074 struct obd_export *exp = req->rq_export;
1075 struct tg_export_data *ted = &exp->exp_target_data;
1076 struct mdt_device *mdt = mti->mti_mdt;
1077 struct mdt_object *child;
1078 struct mdt_body *body;
1081 mdt_req_from_lcd(req, ted->ted_lcd);
1085 /* if no error, so child was created with requested fid */
1086 child = mdt_object_find(mti->mti_env, mdt, mti->mti_rr.rr_fid2);
1087 if (IS_ERR(child)) {
1088 rc = PTR_ERR(child);
1089 LCONSOLE_WARN("Child "DFID" lookup error %d."
1090 " Evicting client %s with export %s.\n",
1091 PFID(mdt_object_fid(child)), rc,
1092 obd_uuid2str(&exp->exp_client_uuid),
1093 obd_export_nid2str(exp));
1094 mdt_export_evict(exp);
1099 body = req_capsule_server_get(mti->mti_pill, &RMF_MDT_BODY);
1100 mti->mti_attr.ma_need = MA_INODE;
1101 mti->mti_attr.ma_valid = 0;
1102 rc = mo_attr_get(mti->mti_env, mdt_object_child(child), &mti->mti_attr);
1103 if (rc == -EREMOTE) {
1104 /* object was created on remote server */
1105 req->rq_status = rc;
1106 body->valid |= OBD_MD_MDS;
1108 mdt_pack_attr2body(mti, body, &mti->mti_attr.ma_attr,
1109 mdt_object_fid(child));
1110 mdt_object_put(mti->mti_env, child);
1113 static void mdt_reconstruct_setattr(struct mdt_thread_info *mti,
1114 struct mdt_lock_handle *lhc)
1116 struct ptlrpc_request *req = mdt_info_req(mti);
1117 struct obd_export *exp = req->rq_export;
1118 struct mdt_export_data *med = &exp->exp_mdt_data;
1119 struct mdt_device *mdt = mti->mti_mdt;
1120 struct mdt_object *obj;
1121 struct mdt_body *body;
1123 mdt_req_from_lcd(req, med->med_ted.ted_lcd);
1127 body = req_capsule_server_get(mti->mti_pill, &RMF_MDT_BODY);
1128 obj = mdt_object_find(mti->mti_env, mdt, mti->mti_rr.rr_fid1);
1130 int rc = PTR_ERR(obj);
1131 LCONSOLE_WARN(""DFID" lookup error %d."
1132 " Evicting client %s with export %s.\n",
1133 PFID(mdt_object_fid(obj)), rc,
1134 obd_uuid2str(&exp->exp_client_uuid),
1135 obd_export_nid2str(exp));
1136 mdt_export_evict(exp);
1140 mti->mti_attr.ma_need = MA_INODE;
1141 mti->mti_attr.ma_valid = 0;
1142 mo_attr_get(mti->mti_env, mdt_object_child(obj), &mti->mti_attr);
1143 mdt_pack_attr2body(mti, body, &mti->mti_attr.ma_attr,
1144 mdt_object_fid(obj));
1145 if (mti->mti_ioepoch && (mti->mti_ioepoch->flags & MF_EPOCH_OPEN)) {
1146 struct mdt_file_data *mfd;
1147 struct mdt_body *repbody;
1149 repbody = req_capsule_server_get(mti->mti_pill, &RMF_MDT_BODY);
1150 repbody->ioepoch = obj->mot_ioepoch;
1151 cfs_spin_lock(&med->med_open_lock);
1152 cfs_list_for_each_entry(mfd, &med->med_open_head, mfd_list) {
1153 if (mfd->mfd_xid == req->rq_xid)
1156 LASSERT(&mfd->mfd_list != &med->med_open_head);
1157 cfs_spin_unlock(&med->med_open_lock);
1158 repbody->handle.cookie = mfd->mfd_handle.h_cookie;
1161 mdt_object_put(mti->mti_env, obj);
1164 typedef void (*mdt_reconstructor)(struct mdt_thread_info *mti,
1165 struct mdt_lock_handle *lhc);
1167 static mdt_reconstructor reconstructors[REINT_MAX] = {
1168 [REINT_SETATTR] = mdt_reconstruct_setattr,
1169 [REINT_CREATE] = mdt_reconstruct_create,
1170 [REINT_LINK] = mdt_reconstruct_generic,
1171 [REINT_UNLINK] = mdt_reconstruct_generic,
1172 [REINT_RENAME] = mdt_reconstruct_generic,
1173 [REINT_OPEN] = mdt_reconstruct_open,
1174 [REINT_SETXATTR] = mdt_reconstruct_generic
1177 void mdt_reconstruct(struct mdt_thread_info *mti,
1178 struct mdt_lock_handle *lhc)
1181 reconstructors[mti->mti_rr.rr_opcode](mti, lhc);