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) 2009, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2012, 2013, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lustre/obdclass/llog_osd.c - low level llog routines on top of OSD API
38 * Author: Alexey Zhuravlev <alexey.zhuravlev@intel.com>
39 * Author: Mikhail Pershin <mike.pershin@intel.com>
42 #define DEBUG_SUBSYSTEM S_LOG
45 #include <obd_class.h>
46 #include <lustre_fid.h>
47 #include <dt_object.h>
49 #include "llog_internal.h"
50 #include "local_storage.h"
53 * - multi-chunks or big-declaration approach
54 * - use unique sequence instead of llog sb tracking unique ids
55 * - re-use existing environment
56 * - named llog support (can be used for testing only at the present)
57 * - llog_origin_connect() work with OSD API
60 static int llog_osd_declare_new_object(const struct lu_env *env,
61 struct local_oid_storage *los,
65 struct llog_thread_info *lgi = llog_info(env);
67 lgi->lgi_attr.la_valid = LA_MODE;
68 lgi->lgi_attr.la_mode = S_IFREG | S_IRUGO | S_IWUSR;
69 lgi->lgi_dof.dof_type = dt_mode_to_dft(S_IFREG);
71 return local_object_declare_create(env, los, o, &lgi->lgi_attr,
75 static int llog_osd_create_new_object(const struct lu_env *env,
76 struct local_oid_storage *los,
80 struct llog_thread_info *lgi = llog_info(env);
82 lgi->lgi_attr.la_valid = LA_MODE;
83 lgi->lgi_attr.la_mode = S_IFREG | S_IRUGO | S_IWUSR;
84 lgi->lgi_dof.dof_type = dt_mode_to_dft(S_IFREG);
86 return local_object_create(env, los, o, &lgi->lgi_attr,
90 static int llog_osd_pad(const struct lu_env *env, struct dt_object *o,
91 loff_t *off, int len, int index, struct thandle *th)
93 struct llog_thread_info *lgi = llog_info(env);
100 LASSERT(len >= LLOG_MIN_REC_SIZE && (len & 0x7) == 0);
102 lgi->lgi_tail.lrt_len = lgi->lgi_lrh.lrh_len = len;
103 lgi->lgi_tail.lrt_index = lgi->lgi_lrh.lrh_index = index;
104 lgi->lgi_lrh.lrh_type = LLOG_PAD_MAGIC;
106 lgi->lgi_buf.lb_buf = &lgi->lgi_lrh;
107 lgi->lgi_buf.lb_len = sizeof(lgi->lgi_lrh);
108 dt_write_lock(env, o, 0);
109 rc = dt_record_write(env, o, &lgi->lgi_buf, off, th);
111 CERROR("%s: error writing padding record: rc = %d\n",
112 o->do_lu.lo_dev->ld_obd->obd_name, rc);
116 lgi->lgi_buf.lb_buf = &lgi->lgi_tail;
117 lgi->lgi_buf.lb_len = sizeof(lgi->lgi_tail);
118 *off += len - sizeof(lgi->lgi_lrh) - sizeof(lgi->lgi_tail);
119 rc = dt_record_write(env, o, &lgi->lgi_buf, off, th);
121 CERROR("%s: error writing padding record: rc = %d\n",
122 o->do_lu.lo_dev->ld_obd->obd_name, rc);
124 dt_write_unlock(env, o);
128 static int llog_osd_write_blob(const struct lu_env *env, struct dt_object *o,
129 struct llog_rec_hdr *rec, void *buf,
130 loff_t *off, struct thandle *th)
132 struct llog_thread_info *lgi = llog_info(env);
133 int buflen = rec->lrh_len;
142 CWARN("0-length record\n");
144 CDEBUG(D_OTHER, "write blob with type %x, buf %p/%u at off %llu\n",
145 rec->lrh_type, buf, buflen, *off);
147 lgi->lgi_attr.la_valid = LA_SIZE;
148 lgi->lgi_attr.la_size = *off;
151 lgi->lgi_buf.lb_len = buflen;
152 lgi->lgi_buf.lb_buf = rec;
153 rc = dt_record_write(env, o, &lgi->lgi_buf, off, th);
155 CERROR("%s: error writing log record: rc = %d\n",
156 o->do_lu.lo_dev->ld_obd->obd_name, rc);
161 /* protect the following 3 writes from concurrent read */
162 dt_write_lock(env, o, 0);
163 rec->lrh_len = sizeof(*rec) + buflen + sizeof(lgi->lgi_tail);
164 lgi->lgi_buf.lb_len = sizeof(*rec);
165 lgi->lgi_buf.lb_buf = rec;
166 rc = dt_record_write(env, o, &lgi->lgi_buf, off, th);
168 CERROR("%s: error writing log hdr: rc = %d\n",
169 o->do_lu.lo_dev->ld_obd->obd_name, rc);
170 GOTO(out_unlock, rc);
173 lgi->lgi_buf.lb_len = buflen;
174 lgi->lgi_buf.lb_buf = buf;
175 rc = dt_record_write(env, o, &lgi->lgi_buf, off, th);
177 CERROR("%s: error writing log buffer: rc = %d\n",
178 o->do_lu.lo_dev->ld_obd->obd_name, rc);
179 GOTO(out_unlock, rc);
182 lgi->lgi_tail.lrt_len = rec->lrh_len;
183 lgi->lgi_tail.lrt_index = rec->lrh_index;
184 lgi->lgi_buf.lb_len = sizeof(lgi->lgi_tail);
185 lgi->lgi_buf.lb_buf = &lgi->lgi_tail;
186 rc = dt_record_write(env, o, &lgi->lgi_buf, off, th);
188 CERROR("%s: error writing log tail: rc = %d\n",
189 o->do_lu.lo_dev->ld_obd->obd_name, rc);
192 dt_write_unlock(env, o);
195 /* cleanup the content written above */
197 dt_punch(env, o, lgi->lgi_attr.la_size, OBD_OBJECT_EOF, th,
199 dt_attr_set(env, o, &lgi->lgi_attr, th, BYPASS_CAPA);
205 static int llog_osd_read_header(const struct lu_env *env,
206 struct llog_handle *handle)
208 struct llog_rec_hdr *llh_hdr;
210 struct llog_thread_info *lgi;
215 LASSERT(sizeof(*handle->lgh_hdr) == LLOG_CHUNK_SIZE);
220 lgi = llog_info(env);
222 rc = dt_attr_get(env, o, &lgi->lgi_attr, NULL);
226 LASSERT(lgi->lgi_attr.la_valid & LA_SIZE);
228 if (lgi->lgi_attr.la_size == 0) {
229 CDEBUG(D_HA, "not reading header from 0-byte log\n");
234 lgi->lgi_buf.lb_buf = handle->lgh_hdr;
235 lgi->lgi_buf.lb_len = LLOG_CHUNK_SIZE;
237 rc = dt_record_read(env, o, &lgi->lgi_buf, &lgi->lgi_off);
239 CERROR("%s: error reading log header from "DFID": rc = %d\n",
240 o->do_lu.lo_dev->ld_obd->obd_name,
241 PFID(lu_object_fid(&o->do_lu)), rc);
245 llh_hdr = &handle->lgh_hdr->llh_hdr;
246 if (LLOG_REC_HDR_NEEDS_SWABBING(llh_hdr))
247 lustre_swab_llog_hdr(handle->lgh_hdr);
249 if (llh_hdr->lrh_type != LLOG_HDR_MAGIC) {
250 CERROR("%s: bad log %s "DFID" header magic: %#x "
251 "(expected %#x)\n", o->do_lu.lo_dev->ld_obd->obd_name,
252 handle->lgh_name ? handle->lgh_name : "",
253 PFID(lu_object_fid(&o->do_lu)),
254 llh_hdr->lrh_type, LLOG_HDR_MAGIC);
256 } else if (llh_hdr->lrh_len != LLOG_CHUNK_SIZE) {
257 CERROR("%s: incorrectly sized log %s "DFID" header: "
258 "%#x (expected %#x)\n"
259 "you may need to re-run lconf --write_conf.\n",
260 o->do_lu.lo_dev->ld_obd->obd_name,
261 handle->lgh_name ? handle->lgh_name : "",
262 PFID(lu_object_fid(&o->do_lu)),
263 llh_hdr->lrh_len, LLOG_CHUNK_SIZE);
267 handle->lgh_last_idx = handle->lgh_hdr->llh_tail.lrt_index;
272 static int llog_osd_declare_write_rec(const struct lu_env *env,
273 struct llog_handle *loghandle,
274 struct llog_rec_hdr *rec,
275 int idx, struct thandle *th)
277 struct llog_thread_info *lgi = llog_info(env);
287 LASSERT(rec->lrh_len <= LLOG_CHUNK_SIZE);
289 o = loghandle->lgh_obj;
292 lgi->lgi_buf.lb_len = sizeof(struct llog_log_hdr);
293 lgi->lgi_buf.lb_buf = NULL;
294 /* each time we update header */
295 rc = dt_declare_record_write(env, o, &lgi->lgi_buf, 0,
297 if (rc || idx == 0) /* if error or just header */
300 if (dt_object_exists(o)) {
301 rc = dt_attr_get(env, o, &lgi->lgi_attr, BYPASS_CAPA);
302 lgi->lgi_off = lgi->lgi_attr.la_size;
303 LASSERT(ergo(rc == 0, lgi->lgi_attr.la_valid & LA_SIZE));
307 rc = dt_declare_punch(env, o, lgi->lgi_off, OBD_OBJECT_EOF, th);
314 lgi->lgi_buf.lb_len = rec->lrh_len;
315 lgi->lgi_buf.lb_buf = NULL;
316 /* XXX: implement declared window or multi-chunks approach */
317 rc = dt_declare_record_write(env, o, &lgi->lgi_buf, -1, th);
322 /* returns negative in on error; 0 if success && reccookie == 0; 1 otherwise */
323 /* appends if idx == -1, otherwise overwrites record idx. */
324 static int llog_osd_write_rec(const struct lu_env *env,
325 struct llog_handle *loghandle,
326 struct llog_rec_hdr *rec,
327 struct llog_cookie *reccookie, int cookiecount,
328 void *buf, int idx, struct thandle *th)
330 struct llog_thread_info *lgi = llog_info(env);
331 struct llog_log_hdr *llh;
332 int reclen = rec->lrh_len;
333 int index, rc, old_tail_idx;
334 struct llog_rec_tail *lrt;
341 llh = loghandle->lgh_hdr;
343 o = loghandle->lgh_obj;
347 CDEBUG(D_OTHER, "new record %x to "DFID"\n",
348 rec->lrh_type, PFID(lu_object_fid(&o->do_lu)));
350 /* record length should not bigger than LLOG_CHUNK_SIZE */
352 rc = (reclen > LLOG_CHUNK_SIZE - sizeof(struct llog_rec_hdr) -
353 sizeof(struct llog_rec_tail)) ? -E2BIG : 0;
355 rc = (reclen > LLOG_CHUNK_SIZE) ? -E2BIG : 0;
359 rc = dt_attr_get(env, o, &lgi->lgi_attr, NULL);
364 /* write_blob adds header and tail to lrh_len. */
365 reclen = sizeof(*rec) + rec->lrh_len +
366 sizeof(struct llog_rec_tail);
369 /* no header: only allowed to insert record 1 */
370 if (idx != 1 && lgi->lgi_attr.la_size == 0)
373 if (idx && llh->llh_size && llh->llh_size != rec->lrh_len)
376 if (!ext2_test_bit(idx, llh->llh_bitmap))
377 CERROR("%s: modify unset record %u\n",
378 o->do_lu.lo_dev->ld_obd->obd_name, idx);
379 if (idx != rec->lrh_index)
380 CERROR("%s: index mismatch %d %u\n",
381 o->do_lu.lo_dev->ld_obd->obd_name, idx,
385 rc = llog_osd_write_blob(env, o, &llh->llh_hdr, NULL,
387 /* we are done if we only write the header or on error */
392 /* We assume that caller has set lgh_cur_* */
393 lgi->lgi_off = loghandle->lgh_cur_offset;
395 "modify record "DOSTID": idx:%d/%u/%d, len:%u "
397 POSTID(&loghandle->lgh_id.lgl_oi), idx,
399 loghandle->lgh_cur_idx, rec->lrh_len,
400 (long long)(lgi->lgi_off - sizeof(*llh)));
401 if (rec->lrh_index != loghandle->lgh_cur_idx) {
402 CERROR("%s: modify idx mismatch %u/%d\n",
403 o->do_lu.lo_dev->ld_obd->obd_name, idx,
404 loghandle->lgh_cur_idx);
408 /* Assumes constant lrh_len */
409 lgi->lgi_off = sizeof(*llh) + (idx - 1) * reclen;
412 rc = llog_osd_write_blob(env, o, rec, buf, &lgi->lgi_off, th);
413 if (rc == 0 && reccookie) {
414 reccookie->lgc_lgl = loghandle->lgh_id;
415 reccookie->lgc_index = idx;
421 /* Make sure that records don't cross a chunk boundary, so we can
422 * process them page-at-a-time if needed. If it will cross a chunk
423 * boundary, write in a fake (but referenced) entry to pad the chunk.
425 * We know that llog_current_log() will return a loghandle that is
426 * big enough to hold reclen, so all we care about is padding here.
428 LASSERT(lgi->lgi_attr.la_valid & LA_SIZE);
429 lgi->lgi_off = lgi->lgi_attr.la_size;
430 left = LLOG_CHUNK_SIZE - (lgi->lgi_off & (LLOG_CHUNK_SIZE - 1));
431 /* NOTE: padding is a record, but no bit is set */
432 if (left != 0 && left != reclen &&
433 left < (reclen + LLOG_MIN_REC_SIZE)) {
434 index = loghandle->lgh_last_idx + 1;
435 rc = llog_osd_pad(env, o, &lgi->lgi_off, left, index, th);
438 loghandle->lgh_last_idx++; /*for pad rec*/
440 /* if it's the last idx in log file, then return -ENOSPC */
441 if (loghandle->lgh_last_idx >= LLOG_BITMAP_SIZE(llh) - 1)
444 loghandle->lgh_last_idx++;
445 index = loghandle->lgh_last_idx;
446 LASSERT(index < LLOG_BITMAP_SIZE(llh));
447 rec->lrh_index = index;
449 lrt = (struct llog_rec_tail *)((char *)rec + rec->lrh_len -
451 lrt->lrt_len = rec->lrh_len;
452 lrt->lrt_index = rec->lrh_index;
454 /* The caller should make sure only 1 process access the lgh_last_idx,
455 * Otherwise it might hit the assert.*/
456 LASSERT(index < LLOG_BITMAP_SIZE(llh));
457 spin_lock(&loghandle->lgh_hdr_lock);
458 if (ext2_set_bit(index, llh->llh_bitmap)) {
459 CERROR("%s: index %u already set in log bitmap\n",
460 o->do_lu.lo_dev->ld_obd->obd_name, index);
461 spin_unlock(&loghandle->lgh_hdr_lock);
462 LBUG(); /* should never happen */
465 spin_unlock(&loghandle->lgh_hdr_lock);
466 old_tail_idx = llh->llh_tail.lrt_index;
467 llh->llh_tail.lrt_index = index;
470 rc = llog_osd_write_blob(env, o, &llh->llh_hdr, NULL, &lgi->lgi_off,
475 rc = dt_attr_get(env, o, &lgi->lgi_attr, NULL);
479 LASSERT(lgi->lgi_attr.la_valid & LA_SIZE);
480 lgi->lgi_off = lgi->lgi_attr.la_size;
482 rc = llog_osd_write_blob(env, o, rec, buf, &lgi->lgi_off, th);
485 /* cleanup llog for error case */
487 spin_lock(&loghandle->lgh_hdr_lock);
488 ext2_clear_bit(index, llh->llh_bitmap);
490 spin_unlock(&loghandle->lgh_hdr_lock);
492 /* restore the header */
493 loghandle->lgh_last_idx--;
494 llh->llh_tail.lrt_index = old_tail_idx;
496 llog_osd_write_blob(env, o, &llh->llh_hdr, NULL,
500 CDEBUG(D_RPCTRACE, "added record "DOSTID": idx: %u, %u\n",
501 POSTID(&loghandle->lgh_id.lgl_oi), index, rec->lrh_len);
502 if (rc == 0 && reccookie) {
503 reccookie->lgc_lgl = loghandle->lgh_id;
504 reccookie->lgc_index = index;
505 if ((rec->lrh_type == MDS_UNLINK_REC) ||
506 (rec->lrh_type == MDS_SETATTR64_REC))
507 reccookie->lgc_subsys = LLOG_MDS_OST_ORIG_CTXT;
508 else if (rec->lrh_type == OST_SZ_REC)
509 reccookie->lgc_subsys = LLOG_SIZE_ORIG_CTXT;
511 reccookie->lgc_subsys = -1;
517 /* We can skip reading at least as many log blocks as the number of
518 * minimum sized log records we are skipping. If it turns out
519 * that we are not far enough along the log (because the
520 * actual records are larger than minimum size) we just skip
523 static void llog_skip_over(__u64 *off, int curr, int goal)
527 *off = (*off + (goal - curr - 1) * LLOG_MIN_REC_SIZE) &
528 ~(LLOG_CHUNK_SIZE - 1);
532 * - cur_offset to the furthest point read in the log file
533 * - cur_idx to the log index preceeding cur_offset
534 * returns -EIO/-EINVAL on error
536 static int llog_osd_next_block(const struct lu_env *env,
537 struct llog_handle *loghandle, int *cur_idx,
538 int next_idx, __u64 *cur_offset, void *buf,
541 struct llog_thread_info *lgi = llog_info(env);
543 struct dt_device *dt;
551 if (len == 0 || len & (LLOG_CHUNK_SIZE - 1))
554 CDEBUG(D_OTHER, "looking for log index %u (cur idx %u off "LPU64")\n",
555 next_idx, *cur_idx, *cur_offset);
558 LASSERT(loghandle->lgh_ctxt);
560 o = loghandle->lgh_obj;
562 LASSERT(dt_object_exists(o));
563 dt = lu2dt_dev(o->do_lu.lo_dev);
566 rc = dt_attr_get(env, o, &lgi->lgi_attr, BYPASS_CAPA);
570 while (*cur_offset < lgi->lgi_attr.la_size) {
571 struct llog_rec_hdr *rec, *last_rec;
572 struct llog_rec_tail *tail;
574 llog_skip_over(cur_offset, *cur_idx, next_idx);
576 /* read up to next LLOG_CHUNK_SIZE block */
577 lgi->lgi_buf.lb_len = LLOG_CHUNK_SIZE -
578 (*cur_offset & (LLOG_CHUNK_SIZE - 1));
579 lgi->lgi_buf.lb_buf = buf;
581 /* Note: read lock is not needed around la_size get above at
582 * the time of dt_attr_get(). There are only two cases that
583 * matter. Either la_size == cur_offset, in which case the
584 * entire read is skipped, or la_size > cur_offset and the loop
585 * is entered and this thread is blocked at dt_read_lock()
586 * until the write is completed. When the write completes, then
587 * the dt_read() will be done with the full length, and will
590 dt_read_lock(env, o, 0);
591 rc = dt_read(env, o, &lgi->lgi_buf, cur_offset);
592 dt_read_unlock(env, o);
594 CERROR("%s: can't read llog block from log "DFID
595 " offset "LPU64": rc = %d\n",
596 o->do_lu.lo_dev->ld_obd->obd_name,
597 PFID(lu_object_fid(&o->do_lu)), *cur_offset,
603 /* signal the end of the valid buffer to
605 memset(buf + rc, 0, len - rc);
608 if (rc == 0) /* end of file, nothing to do */
611 if (rc < sizeof(*tail)) {
612 CERROR("%s: invalid llog block at log id "DOSTID"/%u "
614 o->do_lu.lo_dev->ld_obd->obd_name,
615 POSTID(&loghandle->lgh_id.lgl_oi),
616 loghandle->lgh_id.lgl_ogen, *cur_offset);
617 GOTO(out, rc = -EINVAL);
621 if (LLOG_REC_HDR_NEEDS_SWABBING(rec))
622 lustre_swab_llog_rec(rec);
624 tail = (struct llog_rec_tail *)((char *)buf + rc -
625 sizeof(struct llog_rec_tail));
626 /* get the last record in block */
627 last_rec = (struct llog_rec_hdr *)((char *)buf + rc -
628 le32_to_cpu(tail->lrt_len));
630 if (LLOG_REC_HDR_NEEDS_SWABBING(last_rec))
631 lustre_swab_llog_rec(last_rec);
632 LASSERT(last_rec->lrh_index == tail->lrt_index);
634 *cur_idx = tail->lrt_index;
636 /* this shouldn't happen */
637 if (tail->lrt_index == 0) {
638 CERROR("%s: invalid llog tail at log id "DOSTID"/%u "
640 o->do_lu.lo_dev->ld_obd->obd_name,
641 POSTID(&loghandle->lgh_id.lgl_oi),
642 loghandle->lgh_id.lgl_ogen, *cur_offset);
643 GOTO(out, rc = -EINVAL);
645 if (tail->lrt_index < next_idx)
648 /* sanity check that the start of the new buffer is no farther
649 * than the record that we wanted. This shouldn't happen. */
650 if (rec->lrh_index > next_idx) {
651 CERROR("%s: missed desired record? %u > %u\n",
652 o->do_lu.lo_dev->ld_obd->obd_name,
653 rec->lrh_index, next_idx);
654 GOTO(out, rc = -ENOENT);
658 GOTO(out, rc = -EIO);
663 static int llog_osd_prev_block(const struct lu_env *env,
664 struct llog_handle *loghandle,
665 int prev_idx, void *buf, int len)
667 struct llog_thread_info *lgi = llog_info(env);
669 struct dt_device *dt;
675 if (len == 0 || len & (LLOG_CHUNK_SIZE - 1))
678 CDEBUG(D_OTHER, "looking for log index %u\n", prev_idx);
681 LASSERT(loghandle->lgh_ctxt);
683 o = loghandle->lgh_obj;
685 LASSERT(dt_object_exists(o));
686 dt = lu2dt_dev(o->do_lu.lo_dev);
689 cur_offset = LLOG_CHUNK_SIZE;
690 llog_skip_over(&cur_offset, 0, prev_idx);
692 rc = dt_attr_get(env, o, &lgi->lgi_attr, BYPASS_CAPA);
696 while (cur_offset < lgi->lgi_attr.la_size) {
697 struct llog_rec_hdr *rec, *last_rec;
698 struct llog_rec_tail *tail;
700 lgi->lgi_buf.lb_len = len;
701 lgi->lgi_buf.lb_buf = buf;
702 /* It is OK to have locking around dt_read() only, see
703 * comment in llog_osd_next_block for details
705 dt_read_lock(env, o, 0);
706 rc = dt_read(env, o, &lgi->lgi_buf, &cur_offset);
707 dt_read_unlock(env, o);
709 CERROR("%s: can't read llog block from log "DFID
710 " offset "LPU64": rc = %d\n",
711 o->do_lu.lo_dev->ld_obd->obd_name,
712 PFID(lu_object_fid(&o->do_lu)), cur_offset, rc);
716 if (rc == 0) /* end of file, nothing to do */
719 if (rc < sizeof(*tail)) {
720 CERROR("%s: invalid llog block at log id "DOSTID"/%u "
722 o->do_lu.lo_dev->ld_obd->obd_name,
723 POSTID(&loghandle->lgh_id.lgl_oi),
724 loghandle->lgh_id.lgl_ogen, cur_offset);
725 GOTO(out, rc = -EINVAL);
729 if (LLOG_REC_HDR_NEEDS_SWABBING(rec))
730 lustre_swab_llog_rec(rec);
732 tail = (struct llog_rec_tail *)((char *)buf + rc -
733 sizeof(struct llog_rec_tail));
734 /* get the last record in block */
735 last_rec = (struct llog_rec_hdr *)((char *)buf + rc -
736 le32_to_cpu(tail->lrt_len));
738 if (LLOG_REC_HDR_NEEDS_SWABBING(last_rec))
739 lustre_swab_llog_rec(last_rec);
740 LASSERT(last_rec->lrh_index == tail->lrt_index);
742 /* this shouldn't happen */
743 if (tail->lrt_index == 0) {
744 CERROR("%s: invalid llog tail at log id "DOSTID"/%u "
746 o->do_lu.lo_dev->ld_obd->obd_name,
747 POSTID(&loghandle->lgh_id.lgl_oi),
748 loghandle->lgh_id.lgl_ogen, cur_offset);
749 GOTO(out, rc = -EINVAL);
751 if (tail->lrt_index < prev_idx)
754 /* sanity check that the start of the new buffer is no farther
755 * than the record that we wanted. This shouldn't happen. */
756 if (rec->lrh_index > prev_idx) {
757 CERROR("%s: missed desired record? %u > %u\n",
758 o->do_lu.lo_dev->ld_obd->obd_name,
759 rec->lrh_index, prev_idx);
760 GOTO(out, rc = -ENOENT);
764 GOTO(out, rc = -EIO);
769 struct dt_object *llog_osd_dir_get(const struct lu_env *env,
770 struct llog_ctxt *ctxt)
772 struct dt_device *dt;
773 struct dt_thread_info *dti = dt_info(env);
774 struct dt_object *dir;
777 dt = ctxt->loc_exp->exp_obd->obd_lvfs_ctxt.dt;
778 if (ctxt->loc_dir == NULL) {
779 rc = dt_root_get(env, dt, &dti->dti_fid);
782 dir = dt_locate(env, dt, &dti->dti_fid);
784 if (!IS_ERR(dir) && !dt_try_as_dir(env, dir)) {
785 lu_object_put(env, &dir->do_lu);
786 return ERR_PTR(-ENOTDIR);
789 lu_object_get(&ctxt->loc_dir->do_lu);
796 static int llog_osd_open(const struct lu_env *env, struct llog_handle *handle,
797 struct llog_logid *logid, char *name,
798 enum llog_open_param open_param)
800 struct llog_thread_info *lgi = llog_info(env);
801 struct llog_ctxt *ctxt = handle->lgh_ctxt;
803 struct dt_device *dt;
804 struct ls_device *ls;
805 struct local_oid_storage *los;
812 LASSERT(ctxt->loc_exp);
813 LASSERT(ctxt->loc_exp->exp_obd);
814 dt = ctxt->loc_exp->exp_obd->obd_lvfs_ctxt.dt;
817 ls = ls_device_get(dt);
821 mutex_lock(&ls->ls_los_mutex);
822 los = dt_los_find(ls, name != NULL ? FID_SEQ_LLOG_NAME : FID_SEQ_LLOG);
823 mutex_unlock(&ls->ls_los_mutex);
825 ls_device_put(env, ls);
830 logid_to_fid(logid, &lgi->lgi_fid);
832 struct dt_object *llog_dir;
834 llog_dir = llog_osd_dir_get(env, ctxt);
835 if (IS_ERR(llog_dir))
836 GOTO(out, rc = PTR_ERR(llog_dir));
837 dt_read_lock(env, llog_dir, 0);
838 rc = dt_lookup_dir(env, llog_dir, name, &lgi->lgi_fid);
839 dt_read_unlock(env, llog_dir);
840 lu_object_put(env, &llog_dir->do_lu);
841 if (rc == -ENOENT && open_param == LLOG_OPEN_NEW) {
842 /* generate fid for new llog */
843 rc = local_object_fid_generate(env, los,
848 OBD_ALLOC(handle->lgh_name, strlen(name) + 1);
849 if (handle->lgh_name)
850 strcpy(handle->lgh_name, name);
852 GOTO(out, rc = -ENOMEM);
854 LASSERTF(open_param & LLOG_OPEN_NEW, "%#x\n", open_param);
855 /* generate fid for new llog */
856 rc = local_object_fid_generate(env, los, &lgi->lgi_fid);
861 o = ls_locate(env, ls, &lgi->lgi_fid, NULL);
863 GOTO(out_name, rc = PTR_ERR(o));
865 /* No new llog is expected but doesn't exist */
866 if (open_param != LLOG_OPEN_NEW && !dt_object_exists(o))
867 GOTO(out_put, rc = -ENOENT);
869 fid_to_logid(&lgi->lgi_fid, &handle->lgh_id);
871 handle->private_data = los;
872 LASSERT(handle->lgh_ctxt);
877 lu_object_put(env, &o->do_lu);
879 if (handle->lgh_name != NULL)
880 OBD_FREE(handle->lgh_name, strlen(name) + 1);
886 static int llog_osd_exist(struct llog_handle *handle)
888 LASSERT(handle->lgh_obj);
889 return (dt_object_exists(handle->lgh_obj) &&
890 !lu_object_is_dying(handle->lgh_obj->do_lu.lo_header));
893 static int llog_osd_declare_create(const struct lu_env *env,
894 struct llog_handle *res, struct thandle *th)
896 struct llog_thread_info *lgi = llog_info(env);
897 struct local_oid_storage *los;
903 LASSERT(res->lgh_obj);
906 /* object can be created by another thread */
908 if (dt_object_exists(o))
911 los = res->private_data;
914 rc = llog_osd_declare_new_object(env, los, o, th);
918 /* do not declare header initialization here as it's declared
919 * in llog_osd_declare_write_rec() which is always called */
922 struct dt_object *llog_dir;
924 llog_dir = llog_osd_dir_get(env, res->lgh_ctxt);
925 if (IS_ERR(llog_dir))
926 RETURN(PTR_ERR(llog_dir));
927 logid_to_fid(&res->lgh_id, &lgi->lgi_fid);
928 rc = dt_declare_insert(env, llog_dir,
929 (struct dt_rec *)&lgi->lgi_fid,
930 (struct dt_key *)res->lgh_name, th);
931 lu_object_put(env, &llog_dir->do_lu);
933 CERROR("%s: can't declare named llog %s: rc = %d\n",
934 o->do_lu.lo_dev->ld_obd->obd_name,
940 /* This is a callback from the llog_* functions.
941 * Assumes caller has already pushed us into the kernel context. */
942 static int llog_osd_create(const struct lu_env *env, struct llog_handle *res,
945 struct llog_thread_info *lgi = llog_info(env);
946 struct local_oid_storage *los;
956 /* llog can be already created */
957 if (dt_object_exists(o))
960 los = res->private_data;
963 dt_write_lock(env, o, 0);
964 if (!dt_object_exists(o))
965 rc = llog_osd_create_new_object(env, los, o, th);
969 dt_write_unlock(env, o);
974 struct dt_object *llog_dir;
976 llog_dir = llog_osd_dir_get(env, res->lgh_ctxt);
977 if (IS_ERR(llog_dir))
978 RETURN(PTR_ERR(llog_dir));
980 logid_to_fid(&res->lgh_id, &lgi->lgi_fid);
981 dt_read_lock(env, llog_dir, 0);
982 rc = dt_insert(env, llog_dir,
983 (struct dt_rec *)&lgi->lgi_fid,
984 (struct dt_key *)res->lgh_name,
986 dt_read_unlock(env, llog_dir);
987 lu_object_put(env, &llog_dir->do_lu);
989 CERROR("%s: can't create named llog %s: rc = %d\n",
990 o->do_lu.lo_dev->ld_obd->obd_name,
996 static int llog_osd_close(const struct lu_env *env, struct llog_handle *handle)
998 struct local_oid_storage *los;
1003 LASSERT(handle->lgh_obj);
1005 lu_object_put(env, &handle->lgh_obj->do_lu);
1007 los = handle->private_data;
1011 if (handle->lgh_name)
1012 OBD_FREE(handle->lgh_name, strlen(handle->lgh_name) + 1);
1017 static int llog_osd_destroy(const struct lu_env *env,
1018 struct llog_handle *loghandle)
1020 struct llog_ctxt *ctxt;
1021 struct dt_object *o, *llog_dir = NULL;
1022 struct dt_device *d;
1029 ctxt = loghandle->lgh_ctxt;
1032 o = loghandle->lgh_obj;
1035 d = lu2dt_dev(o->do_lu.lo_dev);
1037 LASSERT(d == ctxt->loc_exp->exp_obd->obd_lvfs_ctxt.dt);
1039 th = dt_trans_create(env, d);
1041 RETURN(PTR_ERR(th));
1043 if (loghandle->lgh_name) {
1044 llog_dir = llog_osd_dir_get(env, ctxt);
1045 if (IS_ERR(llog_dir))
1046 GOTO(out_trans, rc = PTR_ERR(llog_dir));
1048 name = loghandle->lgh_name;
1049 rc = dt_declare_delete(env, llog_dir,
1050 (struct dt_key *)name, th);
1052 GOTO(out_trans, rc);
1055 dt_declare_ref_del(env, o, th);
1057 rc = dt_declare_destroy(env, o, th);
1059 GOTO(out_trans, rc);
1061 rc = dt_trans_start_local(env, d, th);
1063 GOTO(out_trans, rc);
1065 dt_write_lock(env, o, 0);
1066 if (dt_object_exists(o)) {
1068 dt_read_lock(env, llog_dir, 0);
1069 rc = dt_delete(env, llog_dir,
1070 (struct dt_key *) name,
1072 dt_read_unlock(env, llog_dir);
1074 CERROR("%s: can't remove llog %s: rc = %d\n",
1075 o->do_lu.lo_dev->ld_obd->obd_name,
1077 GOTO(out_unlock, rc);
1080 dt_ref_del(env, o, th);
1081 rc = dt_destroy(env, o, th);
1083 GOTO(out_unlock, rc);
1086 dt_write_unlock(env, o);
1088 dt_trans_stop(env, d, th);
1089 if (llog_dir != NULL)
1090 lu_object_put(env, &llog_dir->do_lu);
1094 static int llog_osd_setup(const struct lu_env *env, struct obd_device *obd,
1095 struct obd_llog_group *olg, int ctxt_idx,
1096 struct obd_device *disk_obd)
1098 struct llog_thread_info *lgi = llog_info(env);
1099 struct llog_ctxt *ctxt;
1104 LASSERT(olg->olg_ctxts[ctxt_idx]);
1106 ctxt = llog_ctxt_get(olg->olg_ctxts[ctxt_idx]);
1109 /* initialize data allowing to generate new fids,
1110 * literally we need a sequece */
1111 lgi->lgi_fid.f_seq = FID_SEQ_LLOG;
1112 lgi->lgi_fid.f_oid = 1;
1113 lgi->lgi_fid.f_ver = 0;
1114 rc = local_oid_storage_init(env, disk_obd->obd_lvfs_ctxt.dt,
1116 &ctxt->loc_los_nameless);
1120 lgi->lgi_fid.f_seq = FID_SEQ_LLOG_NAME;
1121 lgi->lgi_fid.f_oid = 1;
1122 lgi->lgi_fid.f_ver = 0;
1123 rc = local_oid_storage_init(env, disk_obd->obd_lvfs_ctxt.dt,
1125 &ctxt->loc_los_named);
1127 local_oid_storage_fini(env, ctxt->loc_los_nameless);
1128 ctxt->loc_los_nameless = NULL;
1134 llog_ctxt_put(ctxt);
1138 static int llog_osd_cleanup(const struct lu_env *env, struct llog_ctxt *ctxt)
1140 if (ctxt->loc_los_nameless != NULL) {
1141 local_oid_storage_fini(env, ctxt->loc_los_nameless);
1142 ctxt->loc_los_nameless = NULL;
1145 if (ctxt->loc_los_named != NULL) {
1146 local_oid_storage_fini(env, ctxt->loc_los_named);
1147 ctxt->loc_los_named = NULL;
1153 struct llog_operations llog_osd_ops = {
1154 .lop_next_block = llog_osd_next_block,
1155 .lop_prev_block = llog_osd_prev_block,
1156 .lop_read_header = llog_osd_read_header,
1157 .lop_destroy = llog_osd_destroy,
1158 .lop_setup = llog_osd_setup,
1159 .lop_cleanup = llog_osd_cleanup,
1160 .lop_open = llog_osd_open,
1161 .lop_exist = llog_osd_exist,
1162 .lop_declare_create = llog_osd_declare_create,
1163 .lop_create = llog_osd_create,
1164 .lop_declare_write_rec = llog_osd_declare_write_rec,
1165 .lop_write_rec = llog_osd_write_rec,
1166 .lop_close = llog_osd_close,
1168 EXPORT_SYMBOL(llog_osd_ops);
1170 /* reads the catalog list */
1171 int llog_osd_get_cat_list(const struct lu_env *env, struct dt_device *d,
1172 int idx, int count, struct llog_catid *idarray,
1173 const struct lu_fid *fid)
1175 struct llog_thread_info *lgi = llog_info(env);
1176 struct dt_object *o = NULL;
1184 size = sizeof(*idarray) * count;
1185 lgi->lgi_off = idx * sizeof(*idarray);
1187 lgi->lgi_fid = *fid;
1188 o = dt_locate(env, d, &lgi->lgi_fid);
1192 if (!dt_object_exists(o)) {
1193 th = dt_trans_create(env, d);
1195 GOTO(out, rc = PTR_ERR(th));
1197 lgi->lgi_attr.la_valid = LA_MODE;
1198 lgi->lgi_attr.la_mode = S_IFREG | S_IRUGO | S_IWUSR;
1199 lgi->lgi_dof.dof_type = dt_mode_to_dft(S_IFREG);
1201 rc = dt_declare_create(env, o, &lgi->lgi_attr, NULL,
1204 GOTO(out_trans, rc);
1206 rc = dt_trans_start_local(env, d, th);
1208 GOTO(out_trans, rc);
1210 dt_write_lock(env, o, 0);
1211 if (!dt_object_exists(o))
1212 rc = dt_create(env, o, &lgi->lgi_attr, NULL,
1214 dt_write_unlock(env, o);
1216 dt_trans_stop(env, d, th);
1221 rc = dt_attr_get(env, o, &lgi->lgi_attr, BYPASS_CAPA);
1225 if (!S_ISREG(lgi->lgi_attr.la_mode)) {
1226 CERROR("%s: CATALOGS is not a regular file!: mode = %o\n",
1227 o->do_lu.lo_dev->ld_obd->obd_name,
1228 lgi->lgi_attr.la_mode);
1229 GOTO(out, rc = -ENOENT);
1232 CDEBUG(D_CONFIG, "cat list: disk size=%d, read=%d\n",
1233 (int)lgi->lgi_attr.la_size, size);
1235 /* return just number of llogs */
1236 if (idarray == NULL) {
1237 rc = lgi->lgi_attr.la_size / sizeof(*idarray);
1241 /* read for new ost index or for empty file */
1242 memset(idarray, 0, size);
1243 if (lgi->lgi_attr.la_size < lgi->lgi_off + size)
1245 if (lgi->lgi_attr.la_size < lgi->lgi_off + size)
1246 size = lgi->lgi_attr.la_size - lgi->lgi_off;
1248 lgi->lgi_buf.lb_buf = idarray;
1249 lgi->lgi_buf.lb_len = size;
1250 rc = dt_record_read(env, o, &lgi->lgi_buf, &lgi->lgi_off);
1252 CERROR("%s: error reading CATALOGS: rc = %d\n",
1253 o->do_lu.lo_dev->ld_obd->obd_name, rc);
1259 lu_object_put(env, &o->do_lu);
1262 EXPORT_SYMBOL(llog_osd_get_cat_list);
1264 /* writes the cat list */
1265 int llog_osd_put_cat_list(const struct lu_env *env, struct dt_device *d,
1266 int idx, int count, struct llog_catid *idarray,
1267 const struct lu_fid *fid)
1269 struct llog_thread_info *lgi = llog_info(env);
1270 struct dt_object *o = NULL;
1279 size = sizeof(*idarray) * count;
1280 lgi->lgi_off = idx * sizeof(*idarray);
1281 lgi->lgi_fid = *fid;
1283 o = dt_locate(env, d, &lgi->lgi_fid);
1287 if (!dt_object_exists(o))
1288 GOTO(out, rc = -ENOENT);
1290 rc = dt_attr_get(env, o, &lgi->lgi_attr, BYPASS_CAPA);
1294 if (!S_ISREG(lgi->lgi_attr.la_mode)) {
1295 CERROR("%s: CATALOGS is not a regular file!: mode = %o\n",
1296 o->do_lu.lo_dev->ld_obd->obd_name,
1297 lgi->lgi_attr.la_mode);
1298 GOTO(out, rc = -ENOENT);
1301 th = dt_trans_create(env, d);
1303 GOTO(out, rc = PTR_ERR(th));
1305 lgi->lgi_buf.lb_len = size;
1306 lgi->lgi_buf.lb_buf = idarray;
1307 rc = dt_declare_record_write(env, o, &lgi->lgi_buf, lgi->lgi_off, th);
1311 rc = dt_trans_start_local(env, d, th);
1313 GOTO(out_trans, rc);
1315 rc = dt_record_write(env, o, &lgi->lgi_buf, &lgi->lgi_off, th);
1317 CDEBUG(D_INODE, "error writeing CATALOGS: rc = %d\n", rc);
1319 dt_trans_stop(env, d, th);
1321 lu_object_put(env, &o->do_lu);
1324 EXPORT_SYMBOL(llog_osd_put_cat_list);