1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright 2008 Sun Microsystems, Inc. All rights reserved
30 * Use is subject to license terms.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lustre/include/lustre_log.h
38 * Generic infrastructure for managing a collection of logs.
39 * These logs are used for:
41 * - orphan recovery: OST adds record on create
42 * - mtime/size consistency: the OST adds a record on first write
43 * - open/unlinked objects: OST adds a record on destroy
45 * - mds unlink log: the MDS adds an entry upon delete
47 * - raid1 replication log between OST's
48 * - MDS replication logs
54 #if defined(__linux__)
55 #include <linux/lustre_log.h>
56 #elif defined(__APPLE__)
57 #include <darwin/lustre_log.h>
58 #elif defined(__WINNT__)
59 #include <winnt/lustre_log.h>
61 #error Unsupported operating system.
66 #include <lustre/lustre_idl.h>
68 #define LOG_NAME_LIMIT(logname, name) \
69 snprintf(logname, sizeof(logname), "LOGS/%s", name)
70 #define LLOG_EEMPTY 4711
72 struct plain_handle_data {
73 struct list_head phd_entry;
74 struct llog_handle *phd_cat_handle;
75 struct llog_cookie phd_cookie; /* cookie of this log in its cat */
79 struct cat_handle_data {
80 struct list_head chd_head;
81 struct llog_handle *chd_current_log; /* currently open log */
84 /* In-memory descriptor for a log object or log catalog */
86 struct rw_semaphore lgh_lock;
87 struct llog_logid lgh_id; /* id of this log */
88 struct llog_log_hdr *lgh_hdr;
89 struct file *lgh_file;
91 int lgh_cur_idx; /* used during llog_process */
92 __u64 lgh_cur_offset; /* used during llog_process */
93 struct llog_ctxt *lgh_ctxt;
95 struct plain_handle_data phd;
96 struct cat_handle_data chd;
100 /* llog.c - general API */
101 typedef int (*llog_cb_t)(struct llog_handle *, struct llog_rec_hdr *, void *);
102 typedef int (*llog_fill_rec_cb_t)(struct llog_rec_hdr *rec, void *data);
103 extern struct llog_handle *llog_alloc_handle(void);
104 int llog_init_handle(struct llog_handle *handle, int flags,
105 struct obd_uuid *uuid);
106 extern void llog_free_handle(struct llog_handle *handle);
107 int llog_process(struct llog_handle *loghandle, llog_cb_t cb,
108 void *data, void *catdata);
109 int llog_reverse_process(struct llog_handle *loghandle, llog_cb_t cb,
110 void *data, void *catdata);
111 extern int llog_cancel_rec(struct llog_handle *loghandle, int index);
112 extern int llog_close(struct llog_handle *cathandle);
113 extern int llog_get_size(struct llog_handle *loghandle);
115 /* llog_cat.c - catalog api */
116 struct llog_process_data {
118 * Any useful data needed while processing catalog. This is
119 * passed later to process callback.
123 * Catalog process callback function, called for each record
129 struct llog_process_cat_data {
131 * Temporary stored first_idx while scanning log.
135 * Temporary stored last_idx while scanning log.
140 struct llog_process_cat_args {
142 * Llog context used in recovery thread on OST (recov_thread.c)
144 struct llog_ctxt *lpca_ctxt;
146 * Llog callback used in recovery thread on OST (recov_thread.c)
150 * Data pointer for llog callback.
155 int llog_cat_put(struct llog_handle *cathandle);
156 int llog_cat_add_rec(struct llog_handle *cathandle, struct llog_rec_hdr *rec,
157 struct llog_cookie *reccookie, void *buf);
158 int llog_cat_cancel_records(struct llog_handle *cathandle, int count,
159 struct llog_cookie *cookies);
160 int llog_cat_process(struct llog_handle *cat_llh, llog_cb_t cb, void *data);
161 int llog_cat_process_thread(void *data);
162 int llog_cat_reverse_process(struct llog_handle *cat_llh, llog_cb_t cb, void *data);
163 int llog_cat_set_first_idx(struct llog_handle *cathandle, int index);
166 int llog_setup(struct obd_device *obd, struct obd_llog_group *olg, int index,
167 struct obd_device *disk_obd, int count, struct llog_logid *logid,
168 struct llog_operations *op);
169 int __llog_ctxt_put(struct llog_ctxt *ctxt);
170 int llog_cleanup(struct llog_ctxt *);
171 int llog_sync(struct llog_ctxt *ctxt, struct obd_export *exp);
172 int llog_add(struct llog_ctxt *ctxt, struct llog_rec_hdr *rec,
173 struct lov_stripe_md *lsm, struct llog_cookie *logcookies,
175 int llog_cancel(struct llog_ctxt *, struct lov_stripe_md *lsm,
176 int count, struct llog_cookie *cookies, int flags);
178 int llog_obd_origin_setup(struct obd_device *obd, struct obd_llog_group *olg,
179 int index, struct obd_device *disk_obd, int count,
180 struct llog_logid *logid);
181 int llog_obd_origin_cleanup(struct llog_ctxt *ctxt);
182 int llog_obd_origin_add(struct llog_ctxt *ctxt,
183 struct llog_rec_hdr *rec, struct lov_stripe_md *lsm,
184 struct llog_cookie *logcookies, int numcookies);
186 int llog_cat_initialize(struct obd_device *obd, struct obd_llog_group *olg,
187 int count, struct obd_uuid *uuid);
188 int obd_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
189 struct obd_device *disk_obd, int count,
190 struct llog_catid *logid, struct obd_uuid *uuid);
192 int obd_llog_finish(struct obd_device *obd, int count);
195 int llog_ioctl(struct llog_ctxt *ctxt, int cmd, struct obd_ioctl_data *data);
196 int llog_catalog_list(struct obd_device *obd, int count,
197 struct obd_ioctl_data *data);
200 int llog_initiator_connect(struct llog_ctxt *ctxt);
201 int llog_receptor_accept(struct llog_ctxt *ctxt, struct obd_import *imp);
202 int llog_origin_connect(struct llog_ctxt *ctxt, int count,
203 struct llog_logid *logid, struct llog_gen *gen,
204 struct obd_uuid *uuid);
205 int llog_handle_connect(struct ptlrpc_request *req);
208 int llog_obd_repl_cancel(struct llog_ctxt *ctxt,
209 struct lov_stripe_md *lsm, int count,
210 struct llog_cookie *cookies, int flags);
211 int llog_obd_repl_sync(struct llog_ctxt *ctxt, struct obd_export *exp);
212 int llog_obd_repl_connect(struct llog_ctxt *ctxt, int count,
213 struct llog_logid *logid, struct llog_gen *gen,
214 struct obd_uuid *uuid);
216 struct llog_operations {
217 int (*lop_write_rec)(struct llog_handle *loghandle,
218 struct llog_rec_hdr *rec,
219 struct llog_cookie *logcookies, int numcookies,
221 int (*lop_destroy)(struct llog_handle *handle);
222 int (*lop_next_block)(struct llog_handle *h, int *curr_idx,
223 int next_idx, __u64 *offset, void *buf, int len);
224 int (*lop_prev_block)(struct llog_handle *h,
225 int prev_idx, void *buf, int len);
226 int (*lop_create)(struct llog_ctxt *ctxt, struct llog_handle **,
227 struct llog_logid *logid, char *name);
228 int (*lop_close)(struct llog_handle *handle);
229 int (*lop_read_header)(struct llog_handle *handle);
231 int (*lop_setup)(struct obd_device *obd, struct obd_llog_group *olg,
232 int ctxt_idx, struct obd_device *disk_obd, int count,
233 struct llog_logid *logid);
234 int (*lop_sync)(struct llog_ctxt *ctxt, struct obd_export *exp);
235 int (*lop_cleanup)(struct llog_ctxt *ctxt);
236 int (*lop_add)(struct llog_ctxt *ctxt, struct llog_rec_hdr *rec,
237 struct lov_stripe_md *lsm,
238 struct llog_cookie *logcookies, int numcookies);
239 int (*lop_cancel)(struct llog_ctxt *ctxt, struct lov_stripe_md *lsm,
240 int count, struct llog_cookie *cookies, int flags);
241 int (*lop_connect)(struct llog_ctxt *ctxt, int count,
242 struct llog_logid *logid, struct llog_gen *gen,
243 struct obd_uuid *uuid);
244 /* XXX add 2 more: commit callbacks and llog recovery functions */
248 extern struct llog_operations llog_lvfs_ops;
249 int llog_get_cat_list(struct obd_device *obd, struct obd_device *disk_obd,
250 char *name, int count, struct llog_catid *idarray);
253 int loc_idx; /* my index the obd array of ctxt's */
254 struct llog_gen loc_gen;
255 struct obd_device *loc_obd; /* points back to the containing obd*/
256 struct obd_llog_group *loc_olg; /* group containing that ctxt */
257 struct obd_export *loc_exp; /* parent "disk" export (e.g. MDS) */
258 struct obd_import *loc_imp; /* to use in RPC's: can be backward
260 struct llog_operations *loc_logops;
261 struct llog_handle *loc_handle;
262 struct llog_commit_master *loc_lcm;
263 struct llog_canceld_ctxt *loc_llcd;
264 struct semaphore loc_sem; /* protects loc_llcd and loc_imp */
265 atomic_t loc_refcount;
269 #define LCM_NAME_SIZE 64
271 struct llog_commit_master {
273 * Thread control flags (start, stop, etc.)
277 * Number of llcds onthis lcm.
281 * Ptlrpc requests set. All cancel rpcs go via this request set.
283 struct ptlrpc_request_set *lcm_set;
285 * Thread control structure. Used for control commit thread.
287 struct ptlrpcd_ctl lcm_pc;
289 * Busy resources waitq
291 cfs_waitq_t lcm_waitq;
293 * Commit thread name buffer. Only used for thread start.
295 char lcm_name[LCM_NAME_SIZE];
298 struct llog_canceld_ctxt {
300 * Llog context this llcd is attached to. Used for accessing
301 * ->loc_import and others in process of canceling cookies
302 * gathered in this llcd.
304 struct llog_ctxt *llcd_ctxt;
306 * Cancel thread control stucture pointer. Used for accessing
307 * it to see if should stop processing and other needs.
309 struct llog_commit_master *llcd_lcm;
311 * Maximal llcd size. Used in calculations on how much of room
312 * left in llcd to cookie comming cookies.
316 * Current llcd size while gathering cookies. This should not be
317 * more than ->llcd_size. Used for determining if we need to
318 * send this llcd (if full) and allocate new one. This is also
319 * used for copying new cookie at the end of buffer.
321 int llcd_cookiebytes;
323 * Pointer to the start of cookies buffer.
325 struct llog_cookie llcd_cookies[0];
328 /* ptlrpc/recov_thread.c */
329 extern struct llog_commit_master *llog_recov_thread_init(char *name);
330 extern void llog_recov_thread_fini(struct llog_commit_master *lcm,
332 extern int llog_recov_thread_start(struct llog_commit_master *lcm);
333 extern void llog_recov_thread_stop(struct llog_commit_master *lcm,
338 #define cap_raise(c, flag) do {} while(0)
340 #define CAP_SYS_RESOURCE 24
342 #endif /* !__KERNEL__ */
344 static inline void llog_gen_init(struct llog_ctxt *ctxt)
346 struct obd_device *obd = ctxt->loc_exp->exp_obd;
348 if (!strcmp(obd->obd_type->typ_name, LUSTRE_MDS_NAME))
349 ctxt->loc_gen.mnt_cnt = obd->u.mds.mds_mount_count;
350 else if (!strstr(obd->obd_type->typ_name, LUSTRE_OST_NAME))
351 ctxt->loc_gen.mnt_cnt = obd->u.filter.fo_mount_count;
353 ctxt->loc_gen.mnt_cnt = 0;
356 static inline int llog_gen_lt(struct llog_gen a, struct llog_gen b)
358 if (a.mnt_cnt < b.mnt_cnt)
360 if (a.mnt_cnt > b.mnt_cnt)
362 return(a.conn_cnt < b.conn_cnt ? 1 : 0);
365 #define LLOG_GEN_INC(gen) ((gen).conn_cnt ++)
366 #define LLOG_PROC_BREAK 0x0001
367 #define LLOG_DEL_RECORD 0x0002
369 static inline int llog_obd2ops(struct llog_ctxt *ctxt,
370 struct llog_operations **lop)
375 *lop = ctxt->loc_logops;
382 static inline int llog_handle2ops(struct llog_handle *loghandle,
383 struct llog_operations **lop)
385 if (loghandle == NULL)
388 return llog_obd2ops(loghandle->lgh_ctxt, lop);
391 static inline int llog_data_len(int len)
393 return size_round(len);
396 static inline struct llog_ctxt *llog_ctxt_get(struct llog_ctxt *ctxt)
398 LASSERT(atomic_read(&ctxt->loc_refcount) > 0);
399 atomic_inc(&ctxt->loc_refcount);
400 CDEBUG(D_INFO, "GETting ctxt %p : new refcount %d\n", ctxt,
401 atomic_read(&ctxt->loc_refcount));
405 static inline void llog_ctxt_put(struct llog_ctxt *ctxt)
409 LASSERT(atomic_read(&ctxt->loc_refcount) > 0);
410 LASSERT(atomic_read(&ctxt->loc_refcount) < 0x5a5a5a);
411 CDEBUG(D_INFO, "PUTting ctxt %p : new refcount %d\n", ctxt,
412 atomic_read(&ctxt->loc_refcount) - 1);
413 __llog_ctxt_put(ctxt);
416 static inline void llog_group_init(struct obd_llog_group *olg, int group)
418 cfs_waitq_init(&olg->olg_waitq);
419 spin_lock_init(&olg->olg_lock);
420 olg->olg_group = group;
423 static inline void llog_group_set_export(struct obd_llog_group *olg,
424 struct obd_export *exp)
426 LASSERT(exp != NULL);
428 spin_lock(&olg->olg_lock);
429 if (olg->olg_exp != NULL && olg->olg_exp != exp)
430 CWARN("%s: export for group %d is changed: 0x%p -> 0x%p\n",
431 exp->exp_obd->obd_name, olg->olg_group,
434 spin_unlock(&olg->olg_lock);
437 static inline int llog_group_set_ctxt(struct obd_llog_group *olg,
438 struct llog_ctxt *ctxt, int index)
440 LASSERT(index >= 0 && index < LLOG_MAX_CTXTS);
442 spin_lock(&olg->olg_lock);
443 if (olg->olg_ctxts[index] != NULL) {
444 spin_unlock(&olg->olg_lock);
447 olg->olg_ctxts[index] = ctxt;
448 spin_unlock(&olg->olg_lock);
452 static inline struct llog_ctxt *llog_group_get_ctxt(struct obd_llog_group *olg,
455 struct llog_ctxt *ctxt;
457 LASSERT(index >= 0 && index < LLOG_MAX_CTXTS);
459 spin_lock(&olg->olg_lock);
460 if (olg->olg_ctxts[index] == NULL) {
463 ctxt = llog_ctxt_get(olg->olg_ctxts[index]);
465 spin_unlock(&olg->olg_lock);
469 static inline struct llog_ctxt *llog_get_context(struct obd_device *obd,
472 return llog_group_get_ctxt(&obd->obd_olg, index);
475 static inline int llog_group_ctxt_null(struct obd_llog_group *olg, int index)
477 return (olg->olg_ctxts[index] == NULL);
480 static inline int llog_ctxt_null(struct obd_device *obd, int index)
482 return (llog_group_ctxt_null(&obd->obd_olg, index));
485 static inline int llog_write_rec(struct llog_handle *handle,
486 struct llog_rec_hdr *rec,
487 struct llog_cookie *logcookies,
488 int numcookies, void *buf, int idx)
490 struct llog_operations *lop;
495 rc = llog_handle2ops(handle, &lop);
499 if (lop->lop_write_rec == NULL)
503 buflen = rec->lrh_len + sizeof(struct llog_rec_hdr)
504 + sizeof(struct llog_rec_tail);
506 buflen = rec->lrh_len;
507 LASSERT(size_round(buflen) == buflen);
509 cap = current->cap_effective;
510 cap_raise(current->cap_effective, CAP_SYS_RESOURCE);
511 rc = lop->lop_write_rec(handle, rec, logcookies, numcookies, buf, idx);
512 current->cap_effective = cap;
516 static inline int llog_read_header(struct llog_handle *handle)
518 struct llog_operations *lop;
522 rc = llog_handle2ops(handle, &lop);
525 if (lop->lop_read_header == NULL)
528 rc = lop->lop_read_header(handle);
532 static inline int llog_destroy(struct llog_handle *handle)
534 struct llog_operations *lop;
538 rc = llog_handle2ops(handle, &lop);
541 if (lop->lop_destroy == NULL)
544 rc = lop->lop_destroy(handle);
549 static inline int llog_cancel(struct obd_export *exp,
550 struct lov_stripe_md *lsm, int count,
551 struct llog_cookie *cookies, int flags)
553 struct llog_operations *lop;
557 rc = llog_handle2ops(loghandle, &lop);
560 if (lop->lop_cancel == NULL)
563 rc = lop->lop_cancel(exp, lsm, count, cookies, flags);
568 static inline int llog_next_block(struct llog_handle *loghandle, int *cur_idx,
569 int next_idx, __u64 *cur_offset, void *buf,
572 struct llog_operations *lop;
576 rc = llog_handle2ops(loghandle, &lop);
579 if (lop->lop_next_block == NULL)
582 rc = lop->lop_next_block(loghandle, cur_idx, next_idx, cur_offset, buf,
587 static inline int llog_prev_block(struct llog_handle *loghandle,
588 int prev_idx, void *buf, int len)
590 struct llog_operations *lop;
594 rc = llog_handle2ops(loghandle, &lop);
597 if (lop->lop_prev_block == NULL)
600 rc = lop->lop_prev_block(loghandle, prev_idx, buf, len);
604 static inline int llog_create(struct llog_ctxt *ctxt, struct llog_handle **res,
605 struct llog_logid *logid, char *name)
607 struct llog_operations *lop;
612 rc = llog_obd2ops(ctxt, &lop);
615 if (lop->lop_create == NULL)
618 cap = current->cap_effective;
619 cap_raise(current->cap_effective, CAP_SYS_RESOURCE);
620 rc = lop->lop_create(ctxt, res, logid, name);
621 current->cap_effective = cap;
625 static inline int llog_connect(struct llog_ctxt *ctxt, int count,
626 struct llog_logid *logid, struct llog_gen *gen,
627 struct obd_uuid *uuid)
629 struct llog_operations *lop;
633 rc = llog_obd2ops(ctxt, &lop);
636 if (lop->lop_connect == NULL)
639 rc = lop->lop_connect(ctxt, count, logid, gen, uuid);