4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2011, 2012, Whamcloud, Inc.
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 struct thandle *mdt_trans_create(const struct lu_env *env,
81 struct mdt_device *mdt)
83 return mdt->mdt_bottom->dd_ops->dt_trans_create(env, mdt->mdt_bottom);
86 int mdt_trans_start(const struct lu_env *env, struct mdt_device *mdt,
89 return dt_trans_start(env, mdt->mdt_bottom, th);
92 void mdt_trans_stop(const struct lu_env *env,
93 struct mdt_device *mdt, struct thandle *th)
95 struct mdt_thread_info *mti;
97 mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
98 /* export can require sync operations */
99 if (mti->mti_exp != NULL)
100 th->th_sync |= mti->mti_exp->exp_need_sync;
101 dt_trans_stop(env, mdt->mdt_bottom, th);
104 static inline int mdt_last_rcvd_header_read(const struct lu_env *env,
105 struct mdt_device *mdt)
107 struct mdt_thread_info *mti;
110 mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
113 rc = dt_record_read(env, mdt->mdt_lut.lut_last_rcvd,
114 mdt_buf(env, &mti->mti_lsd, sizeof(mti->mti_lsd)),
117 lsd_le_to_cpu(&mti->mti_lsd, &mdt->mdt_lut.lut_lsd);
119 CDEBUG(D_INFO, "read last_rcvd header rc = %d, uuid = %s, "
120 "last_transno = "LPU64"\n", rc, mdt->mdt_lut.lut_lsd.lsd_uuid,
121 mdt->mdt_lut.lut_lsd.lsd_last_transno);
125 static int mdt_declare_last_rcvd_header_write(const struct lu_env *env,
126 struct mdt_device *mdt,
129 struct mdt_thread_info *mti;
131 mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
133 return dt_declare_record_write(env, mdt->mdt_lut.lut_last_rcvd,
134 sizeof(mti->mti_lsd), 0, th);
137 static int mdt_last_rcvd_header_write(const struct lu_env *env,
138 struct mdt_device *mdt,
141 struct mdt_thread_info *mti;
145 mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
148 lsd_cpu_to_le(&mdt->mdt_lut.lut_lsd, &mti->mti_lsd);
150 rc = dt_record_write(env, mdt->mdt_lut.lut_last_rcvd,
151 mdt_buf_const(env, &mti->mti_lsd,
152 sizeof(mti->mti_lsd)),
155 CDEBUG(D_INFO, "write last_rcvd header rc = %d, uuid = %s, "
156 "last_transno = "LPU64"\n", rc, mdt->mdt_lut.lut_lsd.lsd_uuid,
157 mdt->mdt_lut.lut_lsd.lsd_last_transno);
162 static int mdt_last_rcvd_read(const struct lu_env *env, struct mdt_device *mdt,
163 struct lsd_client_data *lcd, loff_t *off,
166 struct mdt_thread_info *mti;
167 struct lsd_client_data *tmp;
170 mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
172 rc = dt_record_read(env, mdt->mdt_lut.lut_last_rcvd,
173 mdt_buf(env, tmp, sizeof(*tmp)), off);
175 check_lcd(mdt2obd_dev(mdt)->obd_name, index, tmp);
176 lcd_le_to_cpu(tmp, lcd);
179 CDEBUG(D_INFO, "read lcd @%d rc = %d, uuid = %s, last_transno = "LPU64
180 ", last_xid = "LPU64", last_result = %u, last_data = %u, "
181 "last_close_transno = "LPU64", last_close_xid = "LPU64", "
182 "last_close_result = %u\n", (int)(*off - sizeof(*tmp)),
183 rc, lcd->lcd_uuid, lcd->lcd_last_transno, lcd->lcd_last_xid,
184 lcd->lcd_last_result, lcd->lcd_last_data,
185 lcd->lcd_last_close_transno, lcd->lcd_last_close_xid,
186 lcd->lcd_last_close_result);
190 static int mdt_declare_last_rcvd_write(const struct lu_env *env,
191 struct mdt_device *mdt,
192 loff_t off, struct thandle *th)
194 return dt_declare_record_write(env, mdt->mdt_lut.lut_last_rcvd,
195 sizeof(struct lsd_client_data), off, th);
198 static int mdt_last_rcvd_write(const struct lu_env *env,
199 struct mdt_device *mdt,
200 struct lsd_client_data *lcd,
201 loff_t *off, struct thandle *th)
203 struct mdt_thread_info *mti;
204 struct lsd_client_data *tmp;
208 mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
211 lcd_cpu_to_le(lcd, tmp);
213 rc = dt_record_write(env, mdt->mdt_lut.lut_last_rcvd,
214 mdt_buf_const(env, tmp, sizeof(*tmp)), off, th);
216 CDEBUG(D_INFO, "write lcd @%d rc = %d, uuid = %s, last_transno = "LPU64
217 ", last_xid = "LPU64", last_result = %u, last_data = %u, "
218 "last_close_transno = "LPU64", last_close_xid = "LPU64" ,"
219 "last_close_result = %u\n", (int)(*off - sizeof(*tmp)),
220 rc, lcd->lcd_uuid, lcd->lcd_last_transno, lcd->lcd_last_xid,
221 lcd->lcd_last_result, lcd->lcd_last_data,
222 lcd->lcd_last_close_transno, lcd->lcd_last_close_xid,
223 lcd->lcd_last_close_result);
227 static int mdt_clients_data_init(const struct lu_env *env,
228 struct mdt_device *mdt,
229 unsigned long last_size)
231 struct lr_server_data *lsd = &mdt->mdt_lut.lut_lsd;
232 struct lsd_client_data *lcd;
233 struct obd_device *obd = mdt2obd_dev(mdt);
243 /* When we do a clean MDS shutdown, we save the last_transno into
244 * the header. If we find clients with higher last_transno values
245 * then those clients may need recovery done. */
246 LASSERT(cfs_atomic_read(&obd->obd_req_replay_clients) == 0);
247 for (cl_idx = 0, off = lsd->lsd_client_start;
248 off < last_size; cl_idx++) {
250 struct obd_export *exp;
251 struct mdt_thread_info *mti;
253 off = lsd->lsd_client_start +
254 cl_idx * lsd->lsd_client_size;
256 rc = mdt_last_rcvd_read(env, mdt, lcd, &off, cl_idx);
258 CERROR("error reading MDS %s idx %d, off %llu: rc %d\n",
259 LAST_RCVD, cl_idx, off, rc);
261 break; /* read error shouldn't cause startup to fail */
264 if (lcd->lcd_uuid[0] == '\0') {
265 CDEBUG(D_INFO, "skipping zeroed client at offset %d\n",
270 last_transno = lcd_last_transno(lcd);
272 /* These exports are cleaned up by mdt_obd_disconnect(), so
273 * they need to be set up like real exports as
274 * mdt_obd_connect() does.
276 CDEBUG(D_HA, "RCVRNG CLIENT uuid: %s idx: %d lr: "LPU64
277 " srv lr: "LPU64" lx: "LPU64"\n", lcd->lcd_uuid, cl_idx,
278 last_transno, lsd->lsd_last_transno,
281 exp = class_new_export(obd, (struct obd_uuid *)lcd->lcd_uuid);
283 if (PTR_ERR(exp) == -EALREADY) {
284 /* export already exists, zero out this one */
285 CERROR("Duplicate export %s!\n", lcd->lcd_uuid);
288 GOTO(err_client, rc = PTR_ERR(exp));
291 mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
292 LASSERT(mti != NULL);
294 /* copy on-disk lcd to the export */
295 *exp->exp_target_data.ted_lcd = *lcd;
296 rc = mdt_client_add(env, mdt, cl_idx);
297 /* can't fail existing */
298 LASSERTF(rc == 0, "rc = %d\n", rc);
299 /* VBR: set export last committed version */
300 exp->exp_last_committed = last_transno;
301 cfs_spin_lock(&exp->exp_lock);
302 exp->exp_connecting = 0;
303 exp->exp_in_recovery = 0;
304 cfs_spin_unlock(&exp->exp_lock);
305 obd->obd_max_recoverable_clients++;
306 class_export_put(exp);
308 CDEBUG(D_OTHER, "client at idx %d has last_transno="LPU64"\n",
309 cl_idx, last_transno);
310 /* protect __u64 value update */
311 cfs_spin_lock(&mdt->mdt_lut.lut_translock);
312 mdt->mdt_lut.lut_last_transno = max(last_transno,
313 mdt->mdt_lut.lut_last_transno);
314 cfs_spin_unlock(&mdt->mdt_lut.lut_translock);
322 static int mdt_server_data_init(const struct lu_env *env,
323 struct mdt_device *mdt,
324 struct lustre_sb_info *lsi)
326 struct lr_server_data *lsd = &mdt->mdt_lut.lut_lsd;
327 struct lsd_client_data *lcd = NULL;
328 struct obd_device *obd = mdt2obd_dev(mdt);
329 struct mdt_thread_info *mti;
330 struct dt_object *obj;
332 struct lustre_disk_data *ldd;
333 unsigned long last_rcvd_size;
338 /* ensure padding in the struct is the correct size */
339 CLASSERT(offsetof(struct lr_server_data, lsd_padding) +
340 sizeof(lsd->lsd_padding) == LR_SERVER_SIZE);
341 CLASSERT(offsetof(struct lsd_client_data, lcd_padding) +
342 sizeof(lcd->lcd_padding) == LR_CLIENT_SIZE);
344 mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
345 LASSERT(mti != NULL);
346 la = &mti->mti_attr.ma_attr;
348 obj = mdt->mdt_lut.lut_last_rcvd;
349 rc = obj->do_ops->do_attr_get(env, obj, la, BYPASS_CAPA);
353 last_rcvd_size = (unsigned long)la->la_size;
355 if (last_rcvd_size == 0) {
356 LCONSOLE_WARN("%s: new disk, initializing\n", obd->obd_name);
358 memcpy(lsd->lsd_uuid, obd->obd_uuid.uuid,
359 sizeof(lsd->lsd_uuid));
360 lsd->lsd_last_transno = 0;
361 lsd->lsd_mount_count = 0;
362 lsd->lsd_server_size = LR_SERVER_SIZE;
363 lsd->lsd_client_start = LR_CLIENT_START;
364 lsd->lsd_client_size = LR_CLIENT_SIZE;
365 lsd->lsd_feature_compat = OBD_COMPAT_MDT;
366 lsd->lsd_feature_rocompat = OBD_ROCOMPAT_LOVOBJID;
367 lsd->lsd_feature_incompat = OBD_INCOMPAT_MDT |
368 OBD_INCOMPAT_COMMON_LR |
369 OBD_INCOMPAT_MULTI_OI;
371 LCONSOLE_WARN("%s: used disk, loading\n", obd->obd_name);
372 rc = mdt_last_rcvd_header_read(env, mdt);
374 CERROR("error reading MDS %s: rc %d\n", LAST_RCVD, rc);
377 if (strcmp(lsd->lsd_uuid, obd->obd_uuid.uuid) != 0) {
378 LCONSOLE_ERROR_MSG(0x157, "Trying to start OBD %s using"
379 "the wrong disk %s. Were the /dev/ "
380 "assignments rearranged?\n",
381 obd->obd_uuid.uuid, lsd->lsd_uuid);
382 GOTO(out, rc = -EINVAL);
384 lsd->lsd_feature_compat |= OBD_COMPAT_MDT;
385 lsd->lsd_feature_incompat |= OBD_INCOMPAT_MDT |
386 OBD_INCOMPAT_COMMON_LR;
388 mount_count = lsd->lsd_mount_count;
392 if (lsd->lsd_feature_incompat & ~MDT_INCOMPAT_SUPP) {
393 CERROR("%s: unsupported incompat filesystem feature(s) %x\n",
395 lsd->lsd_feature_incompat & ~MDT_INCOMPAT_SUPP);
396 GOTO(out, rc = -EINVAL);
398 if (lsd->lsd_feature_rocompat & ~MDT_ROCOMPAT_SUPP) {
399 CERROR("%s: unsupported read-only filesystem feature(s) %x\n",
401 lsd->lsd_feature_rocompat & ~MDT_ROCOMPAT_SUPP);
402 /* XXX: Do something like remount filesystem read-only */
403 GOTO(out, rc = -EINVAL);
405 /** Interop: evict all clients at first boot with 1.8 last_rcvd */
406 if (!(lsd->lsd_feature_compat & OBD_COMPAT_20)) {
407 if (last_rcvd_size > lsd->lsd_client_start) {
408 LCONSOLE_WARN("Mounting %s at first time on 1.8 FS, "
409 "remove all clients for interop needs\n",
411 simple_truncate(lsi->lsi_srv_mnt->mnt_sb->s_root,
412 lsi->lsi_srv_mnt, LAST_RCVD,
413 lsd->lsd_client_start);
414 last_rcvd_size = lsd->lsd_client_start;
416 /** set 2.0 flag to upgrade/downgrade between 1.8 and 2.0 */
417 lsd->lsd_feature_compat |= OBD_COMPAT_20;
420 if (ldd->ldd_flags & LDD_F_IAM_DIR)
421 lsd->lsd_feature_incompat |= OBD_INCOMPAT_IAM_DIR;
423 lsd->lsd_feature_incompat |= OBD_INCOMPAT_FID;
425 cfs_spin_lock(&mdt->mdt_lut.lut_translock);
426 mdt->mdt_lut.lut_last_transno = lsd->lsd_last_transno;
427 cfs_spin_unlock(&mdt->mdt_lut.lut_translock);
429 CDEBUG(D_INODE, "========BEGIN DUMPING LAST_RCVD========\n");
430 CDEBUG(D_INODE, "%s: server last_transno: "LPU64"\n",
431 obd->obd_name, mdt->mdt_lut.lut_last_transno);
432 CDEBUG(D_INODE, "%s: server mount_count: "LPU64"\n",
433 obd->obd_name, mount_count + 1);
434 CDEBUG(D_INODE, "%s: server data size: %u\n",
435 obd->obd_name, lsd->lsd_server_size);
436 CDEBUG(D_INODE, "%s: per-client data start: %u\n",
437 obd->obd_name, lsd->lsd_client_start);
438 CDEBUG(D_INODE, "%s: per-client data size: %u\n",
439 obd->obd_name, lsd->lsd_client_size);
440 CDEBUG(D_INODE, "%s: last_rcvd size: %lu\n",
441 obd->obd_name, last_rcvd_size);
442 CDEBUG(D_INODE, "%s: last_rcvd clients: %lu\n", obd->obd_name,
443 last_rcvd_size <= lsd->lsd_client_start ? 0 :
444 (last_rcvd_size - lsd->lsd_client_start) /
445 lsd->lsd_client_size);
446 CDEBUG(D_INODE, "========END DUMPING LAST_RCVD========\n");
448 if (!lsd->lsd_server_size || !lsd->lsd_client_start ||
449 !lsd->lsd_client_size) {
450 CERROR("Bad last_rcvd contents!\n");
451 GOTO(out, rc = -EINVAL);
454 rc = mdt_clients_data_init(env, mdt, last_rcvd_size);
456 GOTO(err_client, rc);
458 cfs_spin_lock(&mdt->mdt_lut.lut_translock);
459 /* obd_last_committed is used for compatibility
460 * with other lustre recovery code */
461 obd->obd_last_committed = mdt->mdt_lut.lut_last_transno;
462 cfs_spin_unlock(&mdt->mdt_lut.lut_translock);
464 obd->u.obt.obt_mount_count = mount_count + 1;
465 obd->u.obt.obt_instance = (__u32)obd->u.obt.obt_mount_count;
466 lsd->lsd_mount_count = obd->u.obt.obt_mount_count;
468 /* save it, so mount count and last_transno is current */
469 rc = mdt_server_data_update(env, mdt);
471 GOTO(err_client, rc);
476 class_disconnect_exports(obd);
481 static int mdt_server_data_update(const struct lu_env *env,
482 struct mdt_device *mdt)
484 struct mdt_thread_info *mti;
488 mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
490 th = mdt_trans_create(env, mdt);
494 rc = mdt_declare_last_rcvd_header_write(env, mdt, th);
498 rc = mdt_trans_start(env, mdt, th);
502 CDEBUG(D_SUPER, "MDS mount_count is "LPU64", last_transno is "LPU64"\n",
503 mdt->mdt_lut.lut_obd->u.obt.obt_mount_count,
504 mdt->mdt_lut.lut_last_transno);
506 cfs_spin_lock(&mdt->mdt_lut.lut_translock);
507 mdt->mdt_lut.lut_lsd.lsd_last_transno = mdt->mdt_lut.lut_last_transno;
508 cfs_spin_unlock(&mdt->mdt_lut.lut_translock);
510 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_lut.lut_client_bitmap;
521 struct mdt_thread_info *mti;
522 struct tg_export_data *ted;
523 struct lr_server_data *lsd = &mdt->mdt_lut.lut_lsd;
524 struct obd_device *obd = mdt2obd_dev(mdt);
531 mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
532 LASSERT(mti != NULL);
534 ted = &mti->mti_exp->exp_target_data;
536 LASSERT(bitmap != NULL);
537 if (!strcmp(ted->ted_lcd->lcd_uuid, obd->obd_uuid.uuid))
540 /* the bitmap operations can handle cl_idx > sizeof(long) * 8, so
541 * there's no need for extra complication here
543 cfs_spin_lock(&mdt->mdt_lut.lut_client_bitmap_lock);
544 cl_idx = cfs_find_first_zero_bit(bitmap, LR_MAX_CLIENTS);
545 if (cl_idx >= LR_MAX_CLIENTS ||
546 OBD_FAIL_CHECK(OBD_FAIL_MDS_CLIENT_ADD)) {
547 CERROR("no room for %u clients - fix LR_MAX_CLIENTS\n",
549 cfs_spin_unlock(&mdt->mdt_lut.lut_client_bitmap_lock);
552 cfs_set_bit(cl_idx, bitmap);
553 cfs_spin_unlock(&mdt->mdt_lut.lut_client_bitmap_lock);
555 CDEBUG(D_INFO, "client at idx %d with UUID '%s' added\n",
556 cl_idx, ted->ted_lcd->lcd_uuid);
558 ted->ted_lr_idx = cl_idx;
559 ted->ted_lr_off = lsd->lsd_client_start +
560 (cl_idx * lsd->lsd_client_size);
561 cfs_mutex_init(&ted->ted_lcd_lock);
563 LASSERTF(ted->ted_lr_off > 0, "ted_lr_off = %llu\n", ted->ted_lr_off);
565 /* Write new client data. */
566 off = ted->ted_lr_off;
568 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_CLIENT_ADD))
571 th = mdt_trans_create(env, mdt);
575 rc = mdt_declare_last_rcvd_write(env, mdt, off, th);
579 rc = mdt_trans_start(env, mdt, th);
584 * Until this operations will be committed the sync is needed
585 * for this export. This should be done _after_ starting the
586 * transaction so that many connecting clients will not bring
587 * server down with lots of sync writes.
589 rc = lut_new_client_cb_add(th, mti->mti_exp);
591 /* can't add callback, do sync now */
594 cfs_spin_lock(&mti->mti_exp->exp_lock);
595 mti->mti_exp->exp_need_sync = 1;
596 cfs_spin_unlock(&mti->mti_exp->exp_lock);
599 rc = mdt_last_rcvd_write(env, mdt, ted->ted_lcd, &off, th);
600 CDEBUG(D_INFO, "wrote client lcd at idx %u off %llu (len %u)\n",
601 cl_idx, ted->ted_lr_off, (int)sizeof(*(ted->ted_lcd)));
604 mdt_trans_stop(env, mdt, th);
609 /* Add client data to the MDS. We use a bitmap to locate a free space
610 * in the last_rcvd file if cl_off is -1 (i.e. a new client).
611 * Otherwise, we just have to read the data from the last_rcvd file and
612 * we know its offset.
614 * It should not be possible to fail adding an existing client - otherwise
615 * mdt_init_server_data() callsite needs to be fixed.
617 int mdt_client_add(const struct lu_env *env,
618 struct mdt_device *mdt, int cl_idx)
620 struct mdt_thread_info *mti;
621 struct tg_export_data *ted;
622 unsigned long *bitmap = mdt->mdt_lut.lut_client_bitmap;
623 struct obd_device *obd = mdt2obd_dev(mdt);
624 struct lr_server_data *lsd = &mdt->mdt_lut.lut_lsd;
628 mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
629 LASSERT(mti != NULL);
631 ted = &mti->mti_exp->exp_target_data;
633 LASSERT(bitmap != NULL);
634 LASSERTF(cl_idx >= 0, "%d\n", cl_idx);
636 if (!strcmp(ted->ted_lcd->lcd_uuid, obd->obd_uuid.uuid))
639 cfs_spin_lock(&mdt->mdt_lut.lut_client_bitmap_lock);
640 if (cfs_test_and_set_bit(cl_idx, bitmap)) {
641 CERROR("MDS client %d: bit already set in bitmap!!\n",
645 cfs_spin_unlock(&mdt->mdt_lut.lut_client_bitmap_lock);
647 CDEBUG(D_INFO, "client at idx %d with UUID '%s' added\n",
648 cl_idx, ted->ted_lcd->lcd_uuid);
650 ted->ted_lr_idx = cl_idx;
651 ted->ted_lr_off = lsd->lsd_client_start +
652 (cl_idx * lsd->lsd_client_size);
653 cfs_mutex_init(&ted->ted_lcd_lock);
655 LASSERTF(ted->ted_lr_off > 0, "ted_lr_off = %llu\n", ted->ted_lr_off);
660 int mdt_client_del(const struct lu_env *env, struct mdt_device *mdt)
662 struct mdt_thread_info *mti;
663 struct tg_export_data *ted;
664 struct obd_device *obd = mdt2obd_dev(mdt);
665 struct obd_export *exp;
671 mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
672 LASSERT(mti != NULL);
675 ted = &exp->exp_target_data;
679 /* XXX: If lcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
680 if (!strcmp(ted->ted_lcd->lcd_uuid, obd->obd_uuid.uuid))
683 CDEBUG(D_INFO, "freeing client at idx %u, offset %lld\n",
684 ted->ted_lr_idx, ted->ted_lr_off);
686 off = ted->ted_lr_off;
689 * Don't clear ted_lr_idx here as it is likely also unset. At worst we
690 * leak a client slot that will be cleaned on the next recovery.
693 CERROR("client idx %d has offset %lld\n",
694 ted->ted_lr_idx, off);
695 GOTO(free, rc = -EINVAL);
699 * Clear the bit _after_ zeroing out the client so we don't race with
700 * mdt_client_add and zero out new clients.
702 if (!cfs_test_bit(ted->ted_lr_idx, mdt->mdt_lut.lut_client_bitmap)) {
703 CERROR("MDT client %u: bit already clear in bitmap!!\n",
708 /* Make sure the server's last_transno is up to date.
709 * This should be done before zeroing client slot so last_transno will
710 * be in server data or in client data in case of failure */
711 mdt_server_data_update(env, mdt);
713 th = mdt_trans_create(env, mdt);
715 GOTO(free, rc = PTR_ERR(th));
717 rc = mdt_declare_last_rcvd_write(env, mdt, off, th);
721 rc = mdt_trans_start(env, mdt, th);
725 cfs_mutex_lock(&ted->ted_lcd_lock);
726 memset(ted->ted_lcd->lcd_uuid, 0, sizeof ted->ted_lcd->lcd_uuid);
727 rc = mdt_last_rcvd_write(env, mdt, ted->ted_lcd, &off, th);
728 cfs_mutex_unlock(&ted->ted_lcd_lock);
731 mdt_trans_stop(env, mdt, th);
733 CDEBUG(rc == 0 ? D_INFO : D_ERROR, "Zeroing out client idx %u in "
734 "%s, rc %d\n", ted->ted_lr_idx, LAST_RCVD, rc);
741 * last_rcvd & last_committed update callbacks
743 static int mdt_last_rcvd_update(struct mdt_thread_info *mti,
746 struct mdt_device *mdt = mti->mti_mdt;
747 struct ptlrpc_request *req = mdt_info_req(mti);
748 struct tg_export_data *ted;
749 struct lsd_client_data *lcd;
752 __s32 rc = th->th_result;
756 LASSERT(req->rq_export);
758 ted = &req->rq_export->exp_target_data;
761 cfs_mutex_lock(&ted->ted_lcd_lock);
763 /* if the export has already been disconnected, we have no last_rcvd slot,
764 * update server data with latest transno then */
766 cfs_mutex_unlock(&ted->ted_lcd_lock);
767 CWARN("commit transaction for disconnected client %s: rc %d\n",
768 req->rq_export->exp_client_uuid.uuid, rc);
769 err = mdt_last_rcvd_header_write(mti->mti_env, mdt, th);
773 off = ted->ted_lr_off;
774 LASSERT(ergo(mti->mti_transno == 0, rc != 0));
775 if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE ||
776 lustre_msg_get_opc(req->rq_reqmsg) == MDS_DONE_WRITING) {
777 if (mti->mti_transno != 0) {
778 if (lcd->lcd_last_close_transno > mti->mti_transno) {
779 CERROR("Trying to overwrite bigger transno:"
780 "on-disk: "LPU64", new: "LPU64" "
781 "replay: %d. see LU-617.\n",
782 lcd->lcd_last_close_transno,
783 mti->mti_transno, req_is_replay(req));
784 if (req_is_replay(req)) {
785 cfs_spin_lock(&req->rq_export->exp_lock);
786 req->rq_export->exp_vbr_failed = 1;
787 cfs_spin_unlock(&req->rq_export->exp_lock);
789 cfs_mutex_unlock(&ted->ted_lcd_lock);
790 RETURN(req_is_replay(req) ? -EOVERFLOW : 0);
792 lcd->lcd_last_close_transno = mti->mti_transno;
794 lcd->lcd_last_close_xid = req->rq_xid;
795 lcd->lcd_last_close_result = rc;
797 /* VBR: save versions in last_rcvd for reconstruct. */
798 __u64 *pre_versions = lustre_msg_get_versions(req->rq_repmsg);
800 lcd->lcd_pre_versions[0] = pre_versions[0];
801 lcd->lcd_pre_versions[1] = pre_versions[1];
802 lcd->lcd_pre_versions[2] = pre_versions[2];
803 lcd->lcd_pre_versions[3] = pre_versions[3];
805 if (mti->mti_transno != 0) {
806 if (lcd->lcd_last_transno > mti->mti_transno) {
807 CERROR("Trying to overwrite bigger transno:"
808 "on-disk: "LPU64", new: "LPU64" "
809 "replay: %d. see LU-617.\n",
810 lcd->lcd_last_transno,
811 mti->mti_transno, req_is_replay(req));
812 if (req_is_replay(req)) {
813 cfs_spin_lock(&req->rq_export->exp_lock);
814 req->rq_export->exp_vbr_failed = 1;
815 cfs_spin_unlock(&req->rq_export->exp_lock);
817 cfs_mutex_unlock(&ted->ted_lcd_lock);
818 RETURN(req_is_replay(req) ? -EOVERFLOW : 0);
820 lcd->lcd_last_transno = mti->mti_transno;
822 lcd->lcd_last_xid = req->rq_xid;
823 lcd->lcd_last_result = rc;
824 /*XXX: save intent_disposition in mdt_thread_info?
825 * also there is bug - intent_dispostion is __u64,
826 * see struct ldlm_reply->lock_policy_res1; */
827 lcd->lcd_last_data = mti->mti_opdata;
831 CERROR("client idx %d has offset %lld\n", ted->ted_lr_idx, off);
834 err = mdt_last_rcvd_write(mti->mti_env, mdt, lcd, &off, th);
836 cfs_mutex_unlock(&ted->ted_lcd_lock);
840 extern struct lu_context_key mdt_thread_key;
842 /* add credits for last_rcvd update */
843 static int mdt_txn_start_cb(const struct lu_env *env,
844 struct thandle *th, void *cookie)
846 struct mdt_device *mdt = cookie;
848 /* XXX: later we'll be declaring this at specific offset */
849 return dt_declare_record_write(env, mdt->mdt_lut.lut_last_rcvd,
850 sizeof(struct lsd_client_data), 0, th);
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 && mti->mti_mos != NULL) {
897 dt_version_set(env, mdt_obj2dt(mti->mti_mos),
898 mti->mti_transno, txn);
902 /* filling reply data */
903 CDEBUG(D_INODE, "transno = "LPU64", last_committed = "LPU64"\n",
904 mti->mti_transno, req->rq_export->exp_obd->obd_last_committed);
906 req->rq_transno = mti->mti_transno;
907 lustre_msg_set_transno(req->rq_repmsg, mti->mti_transno);
908 /* if can't add callback, do sync write */
909 txn->th_sync |= !!lut_last_commit_cb_add(txn, &mdt->mdt_lut,
912 return mdt_last_rcvd_update(mti, txn);
915 int mdt_fs_setup(const struct lu_env *env, struct mdt_device *mdt,
916 struct obd_device *obd,
917 struct lustre_sb_info *lsi)
924 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_FS_SETUP))
927 /* prepare transactions callbacks */
928 mdt->mdt_txn_cb.dtc_txn_start = mdt_txn_start_cb;
929 mdt->mdt_txn_cb.dtc_txn_stop = mdt_txn_stop_cb;
930 mdt->mdt_txn_cb.dtc_txn_commit = NULL;
931 mdt->mdt_txn_cb.dtc_cookie = mdt;
932 mdt->mdt_txn_cb.dtc_tag = LCT_MD_THREAD;
933 CFS_INIT_LIST_HEAD(&mdt->mdt_txn_cb.dtc_linkage);
935 dt_txn_callback_add(mdt->mdt_bottom, &mdt->mdt_txn_cb);
937 rc = mdt_server_data_init(env, mdt, lsi);
941 o = dt_store_open(env, mdt->mdt_bottom, "", CAPA_KEYS, &fid);
944 rc = mdt_capa_keys_init(env, mdt);
946 GOTO(put_ck_object, rc);
949 CERROR("cannot open %s: rc = %d\n", CAPA_KEYS, rc);
950 GOTO(disconnect_exports, rc);
955 lu_object_put(env, &o->do_lu);
956 mdt->mdt_ck_obj = NULL;
958 class_disconnect_exports(obd);
962 void mdt_fs_cleanup(const struct lu_env *env, struct mdt_device *mdt)
966 /* Remove transaction callback */
967 dt_txn_callback_del(mdt->mdt_bottom, &mdt->mdt_txn_cb);
969 lu_object_put(env, &mdt->mdt_ck_obj->do_lu);
970 mdt->mdt_ck_obj = NULL;
974 /* reconstruction code */
975 static void mdt_steal_ack_locks(struct ptlrpc_request *req)
977 struct obd_export *exp = req->rq_export;
979 struct ptlrpc_reply_state *oldrep;
980 struct ptlrpc_service *svc;
983 /* CAVEAT EMPTOR: spinlock order */
984 cfs_spin_lock(&exp->exp_lock);
985 cfs_list_for_each (tmp, &exp->exp_outstanding_replies) {
986 oldrep = cfs_list_entry(tmp, struct ptlrpc_reply_state,
989 if (oldrep->rs_xid != req->rq_xid)
992 if (oldrep->rs_opc != lustre_msg_get_opc(req->rq_reqmsg))
993 CERROR ("Resent req xid "LPU64" has mismatched opc: "
994 "new %d old %d\n", req->rq_xid,
995 lustre_msg_get_opc(req->rq_reqmsg),
998 svc = oldrep->rs_service;
999 cfs_spin_lock (&svc->srv_rs_lock);
1001 cfs_list_del_init (&oldrep->rs_exp_list);
1003 CWARN("Stealing %d locks from rs %p x"LPD64".t"LPD64
1005 oldrep->rs_nlocks, oldrep,
1006 oldrep->rs_xid, oldrep->rs_transno, oldrep->rs_opc,
1007 libcfs_nid2str(exp->exp_connection->c_peer.nid));
1009 for (i = 0; i < oldrep->rs_nlocks; i++)
1010 ptlrpc_save_lock(req, &oldrep->rs_locks[i],
1011 oldrep->rs_modes[i], 0);
1012 oldrep->rs_nlocks = 0;
1014 DEBUG_REQ(D_HA, req, "stole locks for");
1015 cfs_spin_lock(&oldrep->rs_lock);
1016 ptlrpc_schedule_difficult_reply (oldrep);
1017 cfs_spin_unlock(&oldrep->rs_lock);
1019 cfs_spin_unlock (&svc->srv_rs_lock);
1022 cfs_spin_unlock(&exp->exp_lock);
1026 * VBR: restore versions
1028 void mdt_vbr_reconstruct(struct ptlrpc_request *req,
1029 struct lsd_client_data *lcd)
1031 __u64 pre_versions[4] = {0};
1032 pre_versions[0] = lcd->lcd_pre_versions[0];
1033 pre_versions[1] = lcd->lcd_pre_versions[1];
1034 pre_versions[2] = lcd->lcd_pre_versions[2];
1035 pre_versions[3] = lcd->lcd_pre_versions[3];
1036 lustre_msg_set_versions(req->rq_repmsg, pre_versions);
1039 void mdt_req_from_lcd(struct ptlrpc_request *req,
1040 struct lsd_client_data *lcd)
1042 DEBUG_REQ(D_HA, req, "restoring transno "LPD64"/status %d",
1043 lcd->lcd_last_transno, lcd->lcd_last_result);
1045 if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE ||
1046 lustre_msg_get_opc(req->rq_repmsg) == MDS_DONE_WRITING) {
1047 req->rq_transno = lcd->lcd_last_close_transno;
1048 req->rq_status = lcd->lcd_last_close_result;
1050 req->rq_transno = lcd->lcd_last_transno;
1051 req->rq_status = lcd->lcd_last_result;
1052 mdt_vbr_reconstruct(req, lcd);
1054 if (req->rq_status != 0)
1055 req->rq_transno = 0;
1056 lustre_msg_set_transno(req->rq_repmsg, req->rq_transno);
1057 lustre_msg_set_status(req->rq_repmsg, req->rq_status);
1058 DEBUG_REQ(D_RPCTRACE, req, "restoring transno "LPD64"/status %d",
1059 req->rq_transno, req->rq_status);
1061 mdt_steal_ack_locks(req);
1064 void mdt_reconstruct_generic(struct mdt_thread_info *mti,
1065 struct mdt_lock_handle *lhc)
1067 struct ptlrpc_request *req = mdt_info_req(mti);
1068 struct tg_export_data *ted = &req->rq_export->exp_target_data;
1070 return mdt_req_from_lcd(req, ted->ted_lcd);
1073 static void mdt_reconstruct_create(struct mdt_thread_info *mti,
1074 struct mdt_lock_handle *lhc)
1076 struct ptlrpc_request *req = mdt_info_req(mti);
1077 struct obd_export *exp = req->rq_export;
1078 struct tg_export_data *ted = &exp->exp_target_data;
1079 struct mdt_device *mdt = mti->mti_mdt;
1080 struct mdt_object *child;
1081 struct mdt_body *body;
1084 mdt_req_from_lcd(req, ted->ted_lcd);
1088 /* if no error, so child was created with requested fid */
1089 child = mdt_object_find(mti->mti_env, mdt, mti->mti_rr.rr_fid2);
1090 if (IS_ERR(child)) {
1091 rc = PTR_ERR(child);
1092 LCONSOLE_WARN("Child "DFID" lookup error %d."
1093 " Evicting client %s with export %s.\n",
1094 PFID(mdt_object_fid(child)), rc,
1095 obd_uuid2str(&exp->exp_client_uuid),
1096 obd_export_nid2str(exp));
1097 mdt_export_evict(exp);
1102 body = req_capsule_server_get(mti->mti_pill, &RMF_MDT_BODY);
1103 mti->mti_attr.ma_need = MA_INODE;
1104 mti->mti_attr.ma_valid = 0;
1105 rc = mo_attr_get(mti->mti_env, mdt_object_child(child), &mti->mti_attr);
1106 if (rc == -EREMOTE) {
1107 /* object was created on remote server */
1108 req->rq_status = rc;
1109 body->valid |= OBD_MD_MDS;
1111 mdt_pack_attr2body(mti, body, &mti->mti_attr.ma_attr,
1112 mdt_object_fid(child));
1113 mdt_object_put(mti->mti_env, child);
1116 static void mdt_reconstruct_setattr(struct mdt_thread_info *mti,
1117 struct mdt_lock_handle *lhc)
1119 struct ptlrpc_request *req = mdt_info_req(mti);
1120 struct obd_export *exp = req->rq_export;
1121 struct mdt_export_data *med = &exp->exp_mdt_data;
1122 struct mdt_device *mdt = mti->mti_mdt;
1123 struct mdt_object *obj;
1124 struct mdt_body *body;
1126 mdt_req_from_lcd(req, med->med_ted.ted_lcd);
1130 body = req_capsule_server_get(mti->mti_pill, &RMF_MDT_BODY);
1131 obj = mdt_object_find(mti->mti_env, mdt, mti->mti_rr.rr_fid1);
1133 int rc = PTR_ERR(obj);
1134 LCONSOLE_WARN(""DFID" lookup error %d."
1135 " Evicting client %s with export %s.\n",
1136 PFID(mdt_object_fid(obj)), rc,
1137 obd_uuid2str(&exp->exp_client_uuid),
1138 obd_export_nid2str(exp));
1139 mdt_export_evict(exp);
1143 mti->mti_attr.ma_need = MA_INODE;
1144 mti->mti_attr.ma_valid = 0;
1145 mo_attr_get(mti->mti_env, mdt_object_child(obj), &mti->mti_attr);
1146 mdt_pack_attr2body(mti, body, &mti->mti_attr.ma_attr,
1147 mdt_object_fid(obj));
1148 if (mti->mti_ioepoch && (mti->mti_ioepoch->flags & MF_EPOCH_OPEN)) {
1149 struct mdt_file_data *mfd;
1150 struct mdt_body *repbody;
1152 repbody = req_capsule_server_get(mti->mti_pill, &RMF_MDT_BODY);
1153 repbody->ioepoch = obj->mot_ioepoch;
1154 cfs_spin_lock(&med->med_open_lock);
1155 cfs_list_for_each_entry(mfd, &med->med_open_head, mfd_list) {
1156 if (mfd->mfd_xid == req->rq_xid)
1159 LASSERT(&mfd->mfd_list != &med->med_open_head);
1160 cfs_spin_unlock(&med->med_open_lock);
1161 repbody->handle.cookie = mfd->mfd_handle.h_cookie;
1164 mdt_object_put(mti->mti_env, obj);
1167 typedef void (*mdt_reconstructor)(struct mdt_thread_info *mti,
1168 struct mdt_lock_handle *lhc);
1170 static mdt_reconstructor reconstructors[REINT_MAX] = {
1171 [REINT_SETATTR] = mdt_reconstruct_setattr,
1172 [REINT_CREATE] = mdt_reconstruct_create,
1173 [REINT_LINK] = mdt_reconstruct_generic,
1174 [REINT_UNLINK] = mdt_reconstruct_generic,
1175 [REINT_RENAME] = mdt_reconstruct_generic,
1176 [REINT_OPEN] = mdt_reconstruct_open,
1177 [REINT_SETXATTR] = mdt_reconstruct_generic
1180 void mdt_reconstruct(struct mdt_thread_info *mti,
1181 struct mdt_lock_handle *lhc)
1184 reconstructors[mti->mti_rr.rr_opcode](mti, lhc);